// // Generated by NVIDIA NVVM Compiler // // Compiler Build ID: CL-30188945 // Cuda compilation tools, release 11.4, V11.4.100 // Based on NVVM 7.0.1 // .version 7.4 .target sm_75 .address_size 64 .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE ( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) ; .func (.param .align 4 .b8 func_retval0[12]) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE ( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_1[12] ) ; .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE ( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) ; .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE ( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) ; .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) ; .func _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) ; .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i ( .param .b64 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) ; .func _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) ; .func (.param .b32 func_retval0) _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .b32 _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0 ) ; .func _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj ( .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4 ) ; .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji ( .param .b64 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) ; .func _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) ; .func _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b64 _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) ; .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .b64 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) ; .func _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i ( .param .b64 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) ; .func _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj ( .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5 ) ; .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji ( .param .b64 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) ; .func _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) ; .func _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b64 _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) ; .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .b64 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) ; .func _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i ( .param .b64 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) ; .func _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj ( .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) ; .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji ( .param .b64 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) ; .func _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b64 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b64 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .align 4 .b8 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3[16] ) ; .func _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b64 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b64 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .align 4 .b8 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3[16] ) ; .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .b64 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) ; .func _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i ( .param .b64 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) ; .func _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj ( .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) ; .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji ( .param .b64 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) ; .func _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) ; .func _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b64 _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) ; .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji ( .param .b64 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) ; .func _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i ( .param .b64 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) ; .func _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4 ( .param .b64 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) ; .func _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj ( .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) ; .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji ( .param .b64 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) ; .func _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) ; .func _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4 ( .param .b64 _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) ; .const .align 16 .b8 kernel_params[4400]; .const .align 4 .b8 blackbody_table_r[84] = {29, 102, 202, 68, 32, 91, 6, 187, 106, 245, 160, 64, 70, 135, 27, 69, 116, 236, 145, 186, 80, 122, 78, 64, 233, 214, 80, 69, 204, 23, 255, 185, 47, 231, 225, 63, 225, 233, 127, 69, 46, 163, 5, 185, 140, 201, 57, 63, 62, 242, 145, 69, 57, 83, 244, 55, 140, 190, 1, 62, 188, 152, 143, 69, 56, 43, 241, 55, 12, 244, 25, 62, 203, 178, 108, 69, 239, 4, 29, 55, 172, 83, 204, 62}; .const .align 4 .b8 blackbody_table_g[84] = {248, 127, 244, 195, 243, 107, 30, 58, 21, 202, 154, 189, 162, 255, 60, 196, 203, 14, 166, 57, 139, 228, 244, 62, 242, 232, 127, 196, 124, 32, 252, 56, 26, 201, 111, 63, 210, 54, 158, 196, 68, 134, 163, 54, 56, 161, 162, 63, 99, 41, 178, 196, 51, 65, 40, 184, 0, 73, 184, 63, 139, 241, 146, 196, 213, 130, 181, 183, 26, 236, 166, 63, 88, 102, 250, 195, 69, 75, 154, 182, 124, 165, 139, 63}; .const .align 4 .b8 blackbody_table_b[112] = {9, 69, 131, 46, 255, 159, 80, 179, 227, 141, 203, 184, 42, 172, 133, 187, 27, 124, 211, 45, 54, 97, 110, 51, 111, 37, 80, 185, 78, 1, 237, 60, 30, 246, 119, 173, 137, 225, 75, 52, 115, 250, 186, 185, 118, 134, 186, 61, 224, 248, 30, 174, 164, 233, 152, 52, 218, 74, 1, 186, 28, 124, 32, 62, 131, 89, 173, 173, 105, 74, 60, 52, 74, 91, 131, 185, 10, 129, 182, 188, 187, 100, 54, 170, 123, 15, 141, 178, 18, 125, 202, 57, 164, 242, 60, 191, 144, 85, 61, 43, 170, 146, 234, 178, 129, 89, 222, 57, 18, 153, 64, 191}; .const .align 4 .b8 cie_colour_match[972] = {52, 128, 183, 58, 0, 0, 0, 0, 244, 253, 212, 59, 224, 45, 16, 59, 23, 183, 209, 56, 49, 8, 44, 60, 39, 160, 137, 59, 23, 183, 209, 56, 193, 168, 164, 60, 108, 9, 249, 59, 23, 183, 81, 57, 116, 70, 20, 61, 140, 74, 106, 60, 23, 183, 209, 57, 40, 15, 139, 61, 237, 13, 190, 60, 82, 73, 29, 58, 138, 176, 225, 61, 14, 45, 50, 61, 82, 73, 157, 58, 170, 96, 84, 62, 192, 236, 158, 61, 224, 45, 16, 59, 9, 27, 190, 62, 39, 160, 9, 62, 111, 18, 131, 59, 11, 70, 37, 63, 136, 244, 91, 62, 215, 52, 239, 59, 59, 1, 133, 63, 87, 91, 145, 62, 237, 13, 62, 60, 87, 91, 177, 63, 39, 49, 168, 62, 39, 160, 137, 60, 119, 190, 207, 63, 97, 84, 178, 62, 127, 106, 188, 60, 249, 160, 223, 63, 42, 58, 178, 62, 33, 31, 244, 60, 61, 44, 228, 63, 104, 34, 172, 62, 227, 165, 27, 61, 44, 212, 226, 63, 165, 44, 163, 62, 166, 155, 68, 61, 171, 62, 223, 63, 189, 227, 148, 62, 143, 194, 117, 61, 88, 168, 213, 63, 46, 144, 128, 62, 226, 88, 151, 61, 200, 152, 195, 63, 240, 22, 72, 62, 53, 94, 186, 61, 20, 208, 164, 63, 170, 130, 17, 62, 212, 154, 230, 61, 251, 92, 133, 63, 239, 201, 195, 61, 4, 86, 14, 62, 197, 32, 80, 63, 104, 145, 109, 61, 251, 92, 45, 62, 72, 191, 29, 63, 111, 18, 3, 61, 244, 253, 84, 62, 178, 46, 238, 62, 69, 216, 112, 60, 56, 103, 132, 62, 189, 227, 180, 62, 46, 144, 160, 59, 66, 96, 165, 62, 150, 67, 139, 62, 82, 73, 29, 59, 160, 137, 208, 62, 44, 101, 89, 62, 7, 95, 24, 60, 156, 196, 0, 63, 46, 255, 33, 62, 32, 99, 238, 60, 255, 178, 27, 63, 248, 194, 228, 61, 110, 163, 129, 61, 143, 194, 53, 63, 82, 39, 160, 61, 247, 117, 224, 61, 40, 15, 75, 63, 104, 179, 106, 61, 213, 120, 41, 62, 8, 172, 92, 63, 232, 217, 44, 61, 231, 29, 103, 62, 227, 54, 106, 63, 33, 31, 244, 60, 79, 175, 148, 62, 88, 57, 116, 63, 48, 76, 166, 60, 153, 42, 184, 62, 241, 244, 122, 63, 172, 139, 91, 60, 155, 230, 221, 62, 82, 184, 126, 63, 114, 138, 14, 60, 252, 24, 3, 63, 0, 0, 128, 63, 17, 199, 186, 59, 39, 49, 24, 63, 82, 184, 126, 63, 36, 151, 127, 59, 159, 171, 45, 63, 136, 133, 122, 63, 124, 242, 48, 59, 252, 24, 67, 63, 70, 182, 115, 63, 39, 160, 9, 59, 20, 174, 87, 63, 168, 87, 106, 63, 250, 237, 235, 58, 163, 146, 106, 63, 82, 184, 94, 63, 137, 210, 222, 58, 136, 133, 122, 63, 9, 249, 80, 63, 52, 128, 183, 58, 204, 93, 131, 63, 193, 202, 65, 63, 224, 45, 144, 58, 242, 65, 135, 63, 247, 228, 49, 63, 111, 18, 131, 58, 43, 246, 135, 63, 55, 137, 33, 63, 23, 183, 81, 58, 57, 214, 133, 63, 206, 25, 17, 63, 82, 73, 29, 58, 50, 85, 128, 63, 156, 196, 0, 63, 82, 73, 157, 57, 251, 58, 112, 63, 247, 228, 225, 62, 23, 183, 81, 57, 245, 185, 90, 63, 111, 18, 195, 62, 23, 183, 81, 57, 192, 91, 64, 63, 29, 90, 164, 62, 23, 183, 209, 56, 84, 116, 36, 63, 20, 174, 135, 62, 0, 0, 0, 0, 245, 185, 10, 63, 63, 53, 94, 62, 0, 0, 0, 0, 38, 83, 229, 62, 51, 51, 51, 62, 0, 0, 0, 0, 199, 186, 184, 62, 77, 132, 13, 62, 0, 0, 0, 0, 233, 38, 145, 62, 209, 34, 219, 61, 0, 0, 0, 0, 229, 242, 95, 62, 231, 29, 167, 61, 0, 0, 0, 0, 140, 219, 40, 62, 35, 219, 121, 61, 0, 0, 0, 0, 181, 55, 248, 61, 125, 174, 54, 61, 0, 0, 0, 0, 197, 254, 178, 61, 111, 18, 3, 61, 0, 0, 0, 0, 184, 64, 130, 61, 237, 13, 190, 60, 0, 0, 0, 0, 91, 177, 63, 61, 150, 67, 139, 60, 0, 0, 0, 0, 39, 194, 6, 61, 56, 248, 66, 60, 0, 0, 0, 0, 90, 245, 185, 60, 75, 89, 6, 60, 0, 0, 0, 0, 0, 111, 129, 60, 17, 199, 186, 59, 0, 0, 0, 0, 17, 199, 58, 60, 75, 89, 134, 59, 0, 0, 0, 0, 221, 181, 4, 60, 237, 13, 62, 59, 0, 0, 0, 0, 237, 13, 190, 59, 39, 160, 9, 59, 0, 0, 0, 0, 75, 89, 134, 59, 166, 155, 196, 58, 0, 0, 0, 0, 237, 13, 62, 59, 111, 18, 131, 58, 0, 0, 0, 0, 111, 18, 3, 59, 52, 128, 55, 58, 0, 0, 0, 0, 52, 128, 183, 58, 111, 18, 3, 58, 0, 0, 0, 0, 111, 18, 131, 58, 23, 183, 209, 57, 0, 0, 0, 0, 52, 128, 55, 58, 23, 183, 81, 57, 0, 0, 0, 0, 111, 18, 3, 58, 23, 183, 81, 57, 0, 0, 0, 0, 82, 73, 157, 57, 23, 183, 209, 56, 0, 0, 0, 0, 23, 183, 81, 57, 23, 183, 209, 56, 0, 0, 0, 0, 23, 183, 81, 57, 23, 183, 209, 56, 0, 0, 0, 0, 23, 183, 209, 56, 0, 0, 0, 0, 0, 0, 0, 0, 23, 183, 209, 56, 0, 0, 0, 0, 0, 0, 0, 0, 23, 183, 209, 56, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}; .const .align 4 .b8 sobol_burley_table[512] = {1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 8, 0, 0, 0, 16, 0, 0, 0, 32, 0, 0, 0, 64, 0, 0, 0, 128, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 8, 0, 0, 0, 16, 0, 0, 0, 32, 0, 0, 0, 64, 0, 0, 0, 128, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 8, 0, 0, 0, 16, 0, 0, 0, 32, 0, 0, 0, 64, 0, 0, 0, 128, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 8, 0, 0, 0, 16, 0, 0, 0, 32, 0, 0, 0, 64, 0, 0, 0, 128, 1, 0, 0, 0, 3, 0, 0, 0, 5, 0, 0, 0, 15, 0, 0, 0, 17, 0, 0, 0, 51, 0, 0, 0, 85, 0, 0, 0, 255, 0, 0, 0, 1, 1, 0, 0, 3, 3, 0, 0, 5, 5, 0, 0, 15, 15, 0, 0, 17, 17, 0, 0, 51, 51, 0, 0, 85, 85, 0, 0, 255, 255, 0, 0, 1, 0, 1, 0, 3, 0, 3, 0, 5, 0, 5, 0, 15, 0, 15, 0, 17, 0, 17, 0, 51, 0, 51, 0, 85, 0, 85, 0, 255, 0, 255, 0, 1, 1, 1, 1, 3, 3, 3, 3, 5, 5, 5, 5, 15, 15, 15, 15, 17, 17, 17, 17, 51, 51, 51, 51, 85, 85, 85, 85, 255, 255, 255, 255, 1, 0, 0, 0, 3, 0, 0, 0, 6, 0, 0, 0, 9, 0, 0, 0, 23, 0, 0, 0, 58, 0, 0, 0, 113, 0, 0, 0, 163, 0, 0, 0, 22, 1, 0, 0, 57, 3, 0, 0, 119, 6, 0, 0, 170, 9, 0, 0, 1, 22, 0, 0, 3, 57, 0, 0, 6, 119, 0, 0, 9, 170, 0, 0, 23, 1, 1, 0, 58, 3, 3, 0, 113, 6, 6, 0, 163, 9, 9, 0, 22, 22, 23, 0, 57, 57, 58, 0, 119, 119, 113, 0, 170, 170, 163, 0, 1, 0, 23, 1, 3, 0, 58, 3, 6, 0, 113, 6, 9, 0, 163, 9, 23, 0, 22, 22, 58, 0, 57, 57, 113, 0, 119, 119, 163, 0, 170, 170, 1, 0, 0, 0, 3, 0, 0, 0, 4, 0, 0, 0, 10, 0, 0, 0, 31, 0, 0, 0, 46, 0, 0, 0, 69, 0, 0, 0, 201, 0, 0, 0, 27, 1, 0, 0, 164, 2, 0, 0, 154, 7, 0, 0, 103, 11, 0, 0, 30, 16, 0, 0, 45, 48, 0, 0, 65, 64, 0, 0, 195, 160, 0, 0, 4, 241, 1, 0, 138, 226, 2, 0, 223, 87, 4, 0, 174, 155, 12, 0, 5, 161, 17, 0, 137, 114, 42, 0, 219, 231, 121, 0, 164, 219, 182, 0, 26, 1, 0, 1, 167, 2, 0, 3, 158, 7, 0, 4, 109, 11, 0, 10, 1, 16, 0, 31, 3, 48, 0, 46, 4, 64, 0, 69, 10, 160, 0, 201}; .extern .shared .align 16 .b8 warp_offset[]; .func (.param .b32 func_retval0) _Z4betaff( .param .b32 _Z4betaff_param_0, .param .b32 _Z4betaff_param_1 ) { .reg .pred %p<70>; .reg .f32 %f<824>; .reg .b32 %r<58>; ld.param.f32 %f102, [_Z4betaff_param_0]; ld.param.f32 %f103, [_Z4betaff_param_1]; abs.ftz.f32 %f1, %f102; setp.ltu.ftz.f32 %p1, %f1, 0f40400000; @%p1 bra $L__BB0_6; bra.uni $L__BB0_1; $L__BB0_6: setp.ltu.ftz.f32 %p7, %f1, 0f3FC00000; @%p7 bra $L__BB0_8; bra.uni $L__BB0_7; $L__BB0_8: setp.ltu.ftz.f32 %p8, %f1, 0f3F333333; @%p8 bra $L__BB0_10; bra.uni $L__BB0_9; $L__BB0_10: mov.f32 %f209, 0fBBB34878; mov.f32 %f210, 0f3B6B1C86; fma.rn.ftz.f32 %f211, %f210, %f1, %f209; mov.f32 %f212, 0fBD36CAEF; fma.rn.ftz.f32 %f213, %f211, %f1, %f212; mov.f32 %f214, 0f3E2B5555; fma.rn.ftz.f32 %f215, %f213, %f1, %f214; mov.f32 %f216, 0fBD2C96C7; fma.rn.ftz.f32 %f217, %f215, %f1, %f216; mov.f32 %f218, 0fBF27E6EB; fma.rn.ftz.f32 %f219, %f217, %f1, %f218; mov.f32 %f220, 0f3F13C463; fma.rn.ftz.f32 %f221, %f219, %f1, %f220; mul.ftz.f32 %f222, %f1, %f221; fma.rn.ftz.f32 %f223, %f222, %f1, %f1; setp.lt.ftz.f32 %p9, %f223, 0f00800000; mul.ftz.f32 %f224, %f223, 0f4B000000; selp.f32 %f10, %f224, %f223, %p9; selp.f32 %f225, 0fC1B80000, 0f00000000, %p9; mov.b32 %r11, %f10; add.s32 %r12, %r11, -1059760811; and.b32 %r13, %r12, -8388608; sub.s32 %r14, %r11, %r13; mov.b32 %f226, %r14; cvt.rn.f32.s32 %f227, %r13; mov.f32 %f228, 0f34000000; fma.rn.ftz.f32 %f229, %f227, %f228, %f225; add.ftz.f32 %f230, %f226, 0fBF800000; mov.f32 %f231, 0f3E1039F6; mov.f32 %f232, 0fBE055027; fma.rn.ftz.f32 %f233, %f232, %f230, %f231; mov.f32 %f234, 0fBDF8CDCC; fma.rn.ftz.f32 %f235, %f233, %f230, %f234; mov.f32 %f236, 0f3E0F2955; fma.rn.ftz.f32 %f237, %f235, %f230, %f236; mov.f32 %f238, 0fBE2AD8B9; fma.rn.ftz.f32 %f239, %f237, %f230, %f238; mov.f32 %f240, 0f3E4CED0B; fma.rn.ftz.f32 %f241, %f239, %f230, %f240; mov.f32 %f242, 0fBE7FFF22; fma.rn.ftz.f32 %f243, %f241, %f230, %f242; mov.f32 %f244, 0f3EAAAA78; fma.rn.ftz.f32 %f245, %f243, %f230, %f244; mov.f32 %f246, 0fBF000000; fma.rn.ftz.f32 %f247, %f245, %f230, %f246; mul.ftz.f32 %f248, %f230, %f247; fma.rn.ftz.f32 %f249, %f248, %f230, %f230; mov.f32 %f250, 0f3F317218; fma.rn.ftz.f32 %f801, %f229, %f250, %f249; setp.lt.u32 %p10, %r11, 2139095040; @%p10 bra $L__BB0_12; mov.f32 %f251, 0f7F800000; fma.rn.ftz.f32 %f801, %f10, %f251, %f251; $L__BB0_12: neg.ftz.f32 %f252, %f801; setp.eq.ftz.f32 %p11, %f10, 0f00000000; selp.f32 %f802, 0f7F800000, %f252, %p11; bra.uni $L__BB0_13; $L__BB0_1: setp.ltu.ftz.f32 %p2, %f1, 0f40F9999A; @%p2 bra $L__BB0_5; bra.uni $L__BB0_2; $L__BB0_5: add.ftz.f32 %f148, %f1, 0fC0400000; mov.f32 %f149, 0fC640F6F8; mov.f32 %f150, 0fC43B38FB; fma.rn.ftz.f32 %f151, %f150, %f148, %f149; mov.f32 %f152, 0fC7206560; fma.rn.ftz.f32 %f153, %f151, %f148, %f152; mov.f32 %f154, 0fC73CB6AA; fma.rn.ftz.f32 %f155, %f153, %f148, %f154; mov.f32 %f156, 0fC80BAE5A; fma.rn.ftz.f32 %f157, %f155, %f148, %f156; add.ftz.f32 %f158, %f148, 0fC381A020; mov.f32 %f159, 0fC62864B8; fma.rn.ftz.f32 %f160, %f158, %f148, %f159; mov.f32 %f161, 0fC7B50686; fma.rn.ftz.f32 %f162, %f160, %f148, %f161; mov.f32 %f163, 0fC8498465; fma.rn.ftz.f32 %f164, %f162, %f148, %f163; rcp.approx.ftz.f32 %f165, %f164; fma.rn.ftz.f32 %f802, %f157, %f165, %f148; bra.uni $L__BB0_13; $L__BB0_7: add.ftz.f32 %f166, %f1, 0fC0000000; mov.f32 %f167, 0fB967A002; mov.f32 %f168, 0f385007FA; fma.rn.ftz.f32 %f169, %f168, %f166, %f167; mov.f32 %f170, 0f3A0DE6FC; fma.rn.ftz.f32 %f171, %f169, %f166, %f170; mov.f32 %f172, 0fBA9DE0E2; fma.rn.ftz.f32 %f173, %f171, %f166, %f172; mov.f32 %f174, 0f3B3D05B7; fma.rn.ftz.f32 %f175, %f173, %f166, %f174; mov.f32 %f176, 0fBBF1EB10; fma.rn.ftz.f32 %f177, %f175, %f166, %f176; mov.f32 %f178, 0f3CA89A28; fma.rn.ftz.f32 %f179, %f177, %f166, %f178; mov.f32 %f180, 0fBD89F01A; fma.rn.ftz.f32 %f181, %f179, %f166, %f180; mov.f32 %f182, 0f3EA51A66; fma.rn.ftz.f32 %f183, %f181, %f166, %f182; mov.f32 %f184, 0f3ED87730; fma.rn.ftz.f32 %f185, %f183, %f166, %f184; mul.ftz.f32 %f802, %f166, %f185; bra.uni $L__BB0_13; $L__BB0_2: mul.ftz.f32 %f104, %f1, 0f4B000000; setp.lt.ftz.f32 %p3, %f1, 0f00800000; selp.f32 %f2, %f104, %f1, %p3; selp.f32 %f105, 0fC1B80000, 0f00000000, %p3; mov.b32 %r7, %f2; add.s32 %r8, %r7, -1059760811; and.b32 %r9, %r8, -8388608; sub.s32 %r10, %r7, %r9; mov.b32 %f106, %r10; cvt.rn.f32.s32 %f107, %r9; mov.f32 %f108, 0f34000000; fma.rn.ftz.f32 %f109, %f107, %f108, %f105; add.ftz.f32 %f110, %f106, 0fBF800000; mov.f32 %f111, 0f3E1039F6; mov.f32 %f112, 0fBE055027; fma.rn.ftz.f32 %f113, %f112, %f110, %f111; mov.f32 %f114, 0fBDF8CDCC; fma.rn.ftz.f32 %f115, %f113, %f110, %f114; mov.f32 %f116, 0f3E0F2955; fma.rn.ftz.f32 %f117, %f115, %f110, %f116; mov.f32 %f118, 0fBE2AD8B9; fma.rn.ftz.f32 %f119, %f117, %f110, %f118; mov.f32 %f120, 0f3E4CED0B; fma.rn.ftz.f32 %f121, %f119, %f110, %f120; mov.f32 %f122, 0fBE7FFF22; fma.rn.ftz.f32 %f123, %f121, %f110, %f122; mov.f32 %f124, 0f3EAAAA78; fma.rn.ftz.f32 %f125, %f123, %f110, %f124; mov.f32 %f126, 0fBF000000; fma.rn.ftz.f32 %f127, %f125, %f110, %f126; mul.ftz.f32 %f128, %f110, %f127; fma.rn.ftz.f32 %f129, %f128, %f110, %f110; mov.f32 %f130, 0f3F317218; fma.rn.ftz.f32 %f800, %f109, %f130, %f129; setp.lt.u32 %p4, %r7, 2139095040; @%p4 bra $L__BB0_4; mov.f32 %f131, 0f7F800000; fma.rn.ftz.f32 %f800, %f2, %f131, %f131; $L__BB0_4: rcp.approx.ftz.f32 %f132, %f1; mul.ftz.f32 %f133, %f132, %f132; mul.ftz.f32 %f134, %f800, 0f3F000000; setp.eq.ftz.f32 %p5, %f2, 0f00000000; selp.f32 %f135, 0fFF800000, %f134, %p5; add.ftz.f32 %f136, %f1, 0fBF000000; mul.rn.ftz.f32 %f137, %f135, %f136; sub.ftz.f32 %f138, %f137, %f1; mov.f32 %f139, 0fBB360953; mov.f32 %f140, 0f3A4BE755; fma.rn.ftz.f32 %f141, %f140, %f133, %f139; mov.f32 %f142, 0f3DAAAAA3; fma.rn.ftz.f32 %f143, %f141, %f133, %f142; mov.f32 %f144, 0f3F6B3F8E; fma.rn.ftz.f32 %f145, %f143, %f132, %f144; add.rn.ftz.f32 %f146, %f137, %f145; add.ftz.f32 %f147, %f138, %f146; setp.eq.ftz.f32 %p6, %f1, 0f7F800000; selp.f32 %f802, %f1, %f147, %p6; bra.uni $L__BB0_13; $L__BB0_9: mov.f32 %f186, 0f3F800000; sub.ftz.f32 %f187, %f186, %f1; mov.f32 %f188, 0f3DD47577; mov.f32 %f189, 0f3D3BEF76; fma.rn.ftz.f32 %f190, %f189, %f187, %f188; mov.f32 %f191, 0f3DFB8079; fma.rn.ftz.f32 %f192, %f190, %f187, %f191; mov.f32 %f193, 0f3E0295B5; fma.rn.ftz.f32 %f194, %f192, %f187, %f193; mov.f32 %f195, 0f3E12A765; fma.rn.ftz.f32 %f196, %f194, %f187, %f195; mov.f32 %f197, 0f3E2D6867; fma.rn.ftz.f32 %f198, %f196, %f187, %f197; mov.f32 %f199, 0f3E5462BF; fma.rn.ftz.f32 %f200, %f198, %f187, %f199; mov.f32 %f201, 0f3E8A8A72; fma.rn.ftz.f32 %f202, %f200, %f187, %f201; mov.f32 %f203, 0f3ECD26A4; fma.rn.ftz.f32 %f204, %f202, %f187, %f203; mov.f32 %f205, 0f3F528D32; fma.rn.ftz.f32 %f206, %f204, %f187, %f205; mov.f32 %f207, 0f3F13C468; fma.rn.ftz.f32 %f208, %f206, %f187, %f207; mul.ftz.f32 %f802, %f187, %f208; $L__BB0_13: setp.ge.ftz.f32 %p12, %f102, 0f00000000; mov.f32 %f807, %f802; @%p12 bra $L__BB0_26; cvt.rmi.ftz.f32.f32 %f254, %f1; setp.eq.ftz.f32 %p13, %f1, %f254; mov.f32 %f807, 0f7F800000; @%p13 bra $L__BB0_26; setp.lt.ftz.f32 %p14, %f1, 0f1FEC1E4A; @%p14 bra $L__BB0_23; bra.uni $L__BB0_16; $L__BB0_23: setp.lt.ftz.f32 %p21, %f1, 0f00800000; mul.ftz.f32 %f306, %f1, 0f4B000000; selp.f32 %f28, %f306, %f1, %p21; selp.f32 %f307, 0fC1B80000, 0f00000000, %p21; mov.b32 %r20, %f28; add.s32 %r21, %r20, -1059760811; and.b32 %r22, %r21, -8388608; sub.s32 %r23, %r20, %r22; mov.b32 %f308, %r23; cvt.rn.f32.s32 %f309, %r22; mov.f32 %f310, 0f34000000; fma.rn.ftz.f32 %f311, %f309, %f310, %f307; add.ftz.f32 %f312, %f308, 0fBF800000; mov.f32 %f313, 0f3E1039F6; mov.f32 %f314, 0fBE055027; fma.rn.ftz.f32 %f315, %f314, %f312, %f313; mov.f32 %f316, 0fBDF8CDCC; fma.rn.ftz.f32 %f317, %f315, %f312, %f316; mov.f32 %f318, 0f3E0F2955; fma.rn.ftz.f32 %f319, %f317, %f312, %f318; mov.f32 %f320, 0fBE2AD8B9; fma.rn.ftz.f32 %f321, %f319, %f312, %f320; mov.f32 %f322, 0f3E4CED0B; fma.rn.ftz.f32 %f323, %f321, %f312, %f322; mov.f32 %f324, 0fBE7FFF22; fma.rn.ftz.f32 %f325, %f323, %f312, %f324; mov.f32 %f326, 0f3EAAAA78; fma.rn.ftz.f32 %f327, %f325, %f312, %f326; mov.f32 %f328, 0fBF000000; fma.rn.ftz.f32 %f329, %f327, %f312, %f328; mul.ftz.f32 %f330, %f312, %f329; fma.rn.ftz.f32 %f331, %f330, %f312, %f312; mov.f32 %f332, 0f3F317218; fma.rn.ftz.f32 %f806, %f311, %f332, %f331; setp.lt.u32 %p22, %r20, 2139095040; @%p22 bra $L__BB0_25; mov.f32 %f333, 0f7F800000; fma.rn.ftz.f32 %f806, %f28, %f333, %f333; $L__BB0_25: neg.ftz.f32 %f334, %f806; setp.eq.ftz.f32 %p23, %f28, 0f00000000; selp.f32 %f807, 0f7F800000, %f334, %p23; bra.uni $L__BB0_26; $L__BB0_16: add.ftz.f32 %f256, %f1, %f1; cvt.rni.ftz.f32.f32 %f257, %f256; cvt.rzi.ftz.s32.f32 %r1, %f257; neg.ftz.f32 %f258, %f257; mov.f32 %f259, 0f3F000000; fma.rn.ftz.f32 %f260, %f258, %f259, %f1; mul.ftz.f32 %f261, %f260, 0f40490FDB; mul.rn.ftz.f32 %f16, %f261, %f261; and.b32 %r2, %r1, 1; setp.eq.s32 %p15, %r2, 0; selp.f32 %f17, %f261, 0f3F800000, %p15; mov.f32 %f803, 0fB94D4153; @%p15 bra $L__BB0_18; mov.f32 %f262, 0fBAB607ED; mov.f32 %f263, 0f37CBAC00; fma.rn.ftz.f32 %f803, %f263, %f16, %f262; $L__BB0_18: selp.f32 %f264, 0f3C0885E4, 0f3D2AAABB, %p15; fma.rn.ftz.f32 %f265, %f803, %f16, %f264; selp.f32 %f266, 0fBE2AAAA8, 0fBEFFFFFF, %p15; fma.rn.ftz.f32 %f267, %f265, %f16, %f266; mov.f32 %f268, 0f00000000; fma.rn.ftz.f32 %f269, %f16, %f17, %f268; fma.rn.ftz.f32 %f804, %f267, %f269, %f17; and.b32 %r15, %r1, 2; setp.eq.s32 %p17, %r15, 0; @%p17 bra $L__BB0_20; mov.f32 %f271, 0fBF800000; fma.rn.ftz.f32 %f804, %f804, %f271, %f268; $L__BB0_20: abs.ftz.f32 %f272, %f804; mul.ftz.f32 %f273, %f1, %f272; setp.lt.ftz.f32 %p18, %f273, 0f00800000; mul.ftz.f32 %f274, %f273, 0f4B000000; selp.f32 %f275, 0fC1B80000, 0f00000000, %p18; selp.f32 %f23, %f274, %f273, %p18; mov.b32 %r16, %f23; add.s32 %r17, %r16, -1059760811; and.b32 %r18, %r17, -8388608; sub.s32 %r19, %r16, %r18; mov.b32 %f276, %r19; cvt.rn.f32.s32 %f277, %r18; mov.f32 %f278, 0f34000000; fma.rn.ftz.f32 %f279, %f277, %f278, %f275; add.ftz.f32 %f280, %f276, 0fBF800000; mov.f32 %f281, 0f3E1039F6; mov.f32 %f282, 0fBE055027; fma.rn.ftz.f32 %f283, %f282, %f280, %f281; mov.f32 %f284, 0fBDF8CDCC; fma.rn.ftz.f32 %f285, %f283, %f280, %f284; mov.f32 %f286, 0f3E0F2955; fma.rn.ftz.f32 %f287, %f285, %f280, %f286; mov.f32 %f288, 0fBE2AD8B9; fma.rn.ftz.f32 %f289, %f287, %f280, %f288; mov.f32 %f290, 0f3E4CED0B; fma.rn.ftz.f32 %f291, %f289, %f280, %f290; mov.f32 %f292, 0fBE7FFF22; fma.rn.ftz.f32 %f293, %f291, %f280, %f292; mov.f32 %f294, 0f3EAAAA78; fma.rn.ftz.f32 %f295, %f293, %f280, %f294; mov.f32 %f296, 0fBF000000; fma.rn.ftz.f32 %f297, %f295, %f280, %f296; mul.ftz.f32 %f298, %f297, %f280; fma.rn.ftz.f32 %f299, %f298, %f280, %f280; mov.f32 %f300, 0f3F317218; fma.rn.ftz.f32 %f805, %f279, %f300, %f299; setp.lt.u32 %p19, %r16, 2139095040; @%p19 bra $L__BB0_22; mov.f32 %f301, 0f7F800000; fma.rn.ftz.f32 %f805, %f23, %f301, %f301; $L__BB0_22: mov.f32 %f302, 0f3F928682; sub.ftz.f32 %f303, %f302, %f805; setp.eq.ftz.f32 %p20, %f23, 0f00000000; sub.ftz.f32 %f304, %f302, 0fFF800000; selp.f32 %f305, %f304, %f303, %p20; sub.ftz.f32 %f807, %f305, %f802; $L__BB0_26: abs.ftz.f32 %f34, %f103; setp.ltu.ftz.f32 %p24, %f34, 0f40400000; @%p24 bra $L__BB0_32; bra.uni $L__BB0_27; $L__BB0_32: setp.ltu.ftz.f32 %p30, %f34, 0f3FC00000; @%p30 bra $L__BB0_34; bra.uni $L__BB0_33; $L__BB0_34: setp.ltu.ftz.f32 %p31, %f34, 0f3F333333; @%p31 bra $L__BB0_36; bra.uni $L__BB0_35; $L__BB0_36: mov.f32 %f440, 0fBBB34878; mov.f32 %f441, 0f3B6B1C86; fma.rn.ftz.f32 %f442, %f441, %f34, %f440; mov.f32 %f443, 0fBD36CAEF; fma.rn.ftz.f32 %f444, %f442, %f34, %f443; mov.f32 %f445, 0f3E2B5555; fma.rn.ftz.f32 %f446, %f444, %f34, %f445; mov.f32 %f447, 0fBD2C96C7; fma.rn.ftz.f32 %f448, %f446, %f34, %f447; mov.f32 %f449, 0fBF27E6EB; fma.rn.ftz.f32 %f450, %f448, %f34, %f449; mov.f32 %f451, 0f3F13C463; fma.rn.ftz.f32 %f452, %f450, %f34, %f451; mul.ftz.f32 %f453, %f34, %f452; fma.rn.ftz.f32 %f454, %f453, %f34, %f34; setp.lt.ftz.f32 %p32, %f454, 0f00800000; mul.ftz.f32 %f455, %f454, 0f4B000000; selp.f32 %f43, %f455, %f454, %p32; selp.f32 %f456, 0fC1B80000, 0f00000000, %p32; mov.b32 %r28, %f43; add.s32 %r29, %r28, -1059760811; and.b32 %r30, %r29, -8388608; sub.s32 %r31, %r28, %r30; mov.b32 %f457, %r31; cvt.rn.f32.s32 %f458, %r30; mov.f32 %f459, 0f34000000; fma.rn.ftz.f32 %f460, %f458, %f459, %f456; add.ftz.f32 %f461, %f457, 0fBF800000; mov.f32 %f462, 0f3E1039F6; mov.f32 %f463, 0fBE055027; fma.rn.ftz.f32 %f464, %f463, %f461, %f462; mov.f32 %f465, 0fBDF8CDCC; fma.rn.ftz.f32 %f466, %f464, %f461, %f465; mov.f32 %f467, 0f3E0F2955; fma.rn.ftz.f32 %f468, %f466, %f461, %f467; mov.f32 %f469, 0fBE2AD8B9; fma.rn.ftz.f32 %f470, %f468, %f461, %f469; mov.f32 %f471, 0f3E4CED0B; fma.rn.ftz.f32 %f472, %f470, %f461, %f471; mov.f32 %f473, 0fBE7FFF22; fma.rn.ftz.f32 %f474, %f472, %f461, %f473; mov.f32 %f475, 0f3EAAAA78; fma.rn.ftz.f32 %f476, %f474, %f461, %f475; mov.f32 %f477, 0fBF000000; fma.rn.ftz.f32 %f478, %f476, %f461, %f477; mul.ftz.f32 %f479, %f461, %f478; fma.rn.ftz.f32 %f480, %f479, %f461, %f461; mov.f32 %f481, 0f3F317218; fma.rn.ftz.f32 %f809, %f460, %f481, %f480; setp.lt.u32 %p33, %r28, 2139095040; @%p33 bra $L__BB0_38; mov.f32 %f482, 0f7F800000; fma.rn.ftz.f32 %f809, %f43, %f482, %f482; $L__BB0_38: neg.ftz.f32 %f483, %f809; setp.eq.ftz.f32 %p34, %f43, 0f00000000; selp.f32 %f810, 0f7F800000, %f483, %p34; bra.uni $L__BB0_39; $L__BB0_27: setp.ltu.ftz.f32 %p25, %f34, 0f40F9999A; @%p25 bra $L__BB0_31; bra.uni $L__BB0_28; $L__BB0_31: add.ftz.f32 %f379, %f34, 0fC0400000; mov.f32 %f380, 0fC640F6F8; mov.f32 %f381, 0fC43B38FB; fma.rn.ftz.f32 %f382, %f381, %f379, %f380; mov.f32 %f383, 0fC7206560; fma.rn.ftz.f32 %f384, %f382, %f379, %f383; mov.f32 %f385, 0fC73CB6AA; fma.rn.ftz.f32 %f386, %f384, %f379, %f385; mov.f32 %f387, 0fC80BAE5A; fma.rn.ftz.f32 %f388, %f386, %f379, %f387; add.ftz.f32 %f389, %f379, 0fC381A020; mov.f32 %f390, 0fC62864B8; fma.rn.ftz.f32 %f391, %f389, %f379, %f390; mov.f32 %f392, 0fC7B50686; fma.rn.ftz.f32 %f393, %f391, %f379, %f392; mov.f32 %f394, 0fC8498465; fma.rn.ftz.f32 %f395, %f393, %f379, %f394; rcp.approx.ftz.f32 %f396, %f395; fma.rn.ftz.f32 %f810, %f388, %f396, %f379; bra.uni $L__BB0_39; $L__BB0_33: add.ftz.f32 %f397, %f34, 0fC0000000; mov.f32 %f398, 0fB967A002; mov.f32 %f399, 0f385007FA; fma.rn.ftz.f32 %f400, %f399, %f397, %f398; mov.f32 %f401, 0f3A0DE6FC; fma.rn.ftz.f32 %f402, %f400, %f397, %f401; mov.f32 %f403, 0fBA9DE0E2; fma.rn.ftz.f32 %f404, %f402, %f397, %f403; mov.f32 %f405, 0f3B3D05B7; fma.rn.ftz.f32 %f406, %f404, %f397, %f405; mov.f32 %f407, 0fBBF1EB10; fma.rn.ftz.f32 %f408, %f406, %f397, %f407; mov.f32 %f409, 0f3CA89A28; fma.rn.ftz.f32 %f410, %f408, %f397, %f409; mov.f32 %f411, 0fBD89F01A; fma.rn.ftz.f32 %f412, %f410, %f397, %f411; mov.f32 %f413, 0f3EA51A66; fma.rn.ftz.f32 %f414, %f412, %f397, %f413; mov.f32 %f415, 0f3ED87730; fma.rn.ftz.f32 %f416, %f414, %f397, %f415; mul.ftz.f32 %f810, %f397, %f416; bra.uni $L__BB0_39; $L__BB0_28: mul.ftz.f32 %f335, %f34, 0f4B000000; setp.lt.ftz.f32 %p26, %f34, 0f00800000; selp.f32 %f35, %f335, %f34, %p26; selp.f32 %f336, 0fC1B80000, 0f00000000, %p26; mov.b32 %r24, %f35; add.s32 %r25, %r24, -1059760811; and.b32 %r26, %r25, -8388608; sub.s32 %r27, %r24, %r26; mov.b32 %f337, %r27; cvt.rn.f32.s32 %f338, %r26; mov.f32 %f339, 0f34000000; fma.rn.ftz.f32 %f340, %f338, %f339, %f336; add.ftz.f32 %f341, %f337, 0fBF800000; mov.f32 %f342, 0f3E1039F6; mov.f32 %f343, 0fBE055027; fma.rn.ftz.f32 %f344, %f343, %f341, %f342; mov.f32 %f345, 0fBDF8CDCC; fma.rn.ftz.f32 %f346, %f344, %f341, %f345; mov.f32 %f347, 0f3E0F2955; fma.rn.ftz.f32 %f348, %f346, %f341, %f347; mov.f32 %f349, 0fBE2AD8B9; fma.rn.ftz.f32 %f350, %f348, %f341, %f349; mov.f32 %f351, 0f3E4CED0B; fma.rn.ftz.f32 %f352, %f350, %f341, %f351; mov.f32 %f353, 0fBE7FFF22; fma.rn.ftz.f32 %f354, %f352, %f341, %f353; mov.f32 %f355, 0f3EAAAA78; fma.rn.ftz.f32 %f356, %f354, %f341, %f355; mov.f32 %f357, 0fBF000000; fma.rn.ftz.f32 %f358, %f356, %f341, %f357; mul.ftz.f32 %f359, %f341, %f358; fma.rn.ftz.f32 %f360, %f359, %f341, %f341; mov.f32 %f361, 0f3F317218; fma.rn.ftz.f32 %f808, %f340, %f361, %f360; setp.lt.u32 %p27, %r24, 2139095040; @%p27 bra $L__BB0_30; mov.f32 %f362, 0f7F800000; fma.rn.ftz.f32 %f808, %f35, %f362, %f362; $L__BB0_30: rcp.approx.ftz.f32 %f363, %f34; mul.ftz.f32 %f364, %f363, %f363; mul.ftz.f32 %f365, %f808, 0f3F000000; setp.eq.ftz.f32 %p28, %f35, 0f00000000; selp.f32 %f366, 0fFF800000, %f365, %p28; add.ftz.f32 %f367, %f34, 0fBF000000; mul.rn.ftz.f32 %f368, %f366, %f367; sub.ftz.f32 %f369, %f368, %f34; mov.f32 %f370, 0fBB360953; mov.f32 %f371, 0f3A4BE755; fma.rn.ftz.f32 %f372, %f371, %f364, %f370; mov.f32 %f373, 0f3DAAAAA3; fma.rn.ftz.f32 %f374, %f372, %f364, %f373; mov.f32 %f375, 0f3F6B3F8E; fma.rn.ftz.f32 %f376, %f374, %f363, %f375; add.rn.ftz.f32 %f377, %f368, %f376; add.ftz.f32 %f378, %f369, %f377; setp.eq.ftz.f32 %p29, %f34, 0f7F800000; selp.f32 %f810, %f34, %f378, %p29; bra.uni $L__BB0_39; $L__BB0_35: mov.f32 %f417, 0f3F800000; sub.ftz.f32 %f418, %f417, %f34; mov.f32 %f419, 0f3DD47577; mov.f32 %f420, 0f3D3BEF76; fma.rn.ftz.f32 %f421, %f420, %f418, %f419; mov.f32 %f422, 0f3DFB8079; fma.rn.ftz.f32 %f423, %f421, %f418, %f422; mov.f32 %f424, 0f3E0295B5; fma.rn.ftz.f32 %f425, %f423, %f418, %f424; mov.f32 %f426, 0f3E12A765; fma.rn.ftz.f32 %f427, %f425, %f418, %f426; mov.f32 %f428, 0f3E2D6867; fma.rn.ftz.f32 %f429, %f427, %f418, %f428; mov.f32 %f430, 0f3E5462BF; fma.rn.ftz.f32 %f431, %f429, %f418, %f430; mov.f32 %f432, 0f3E8A8A72; fma.rn.ftz.f32 %f433, %f431, %f418, %f432; mov.f32 %f434, 0f3ECD26A4; fma.rn.ftz.f32 %f435, %f433, %f418, %f434; mov.f32 %f436, 0f3F528D32; fma.rn.ftz.f32 %f437, %f435, %f418, %f436; mov.f32 %f438, 0f3F13C468; fma.rn.ftz.f32 %f439, %f437, %f418, %f438; mul.ftz.f32 %f810, %f418, %f439; $L__BB0_39: setp.ge.ftz.f32 %p35, %f103, 0f00000000; mov.f32 %f815, %f810; @%p35 bra $L__BB0_52; cvt.rmi.ftz.f32.f32 %f485, %f34; setp.eq.ftz.f32 %p36, %f34, %f485; mov.f32 %f815, 0f7F800000; @%p36 bra $L__BB0_52; setp.lt.ftz.f32 %p37, %f34, 0f1FEC1E4A; @%p37 bra $L__BB0_49; bra.uni $L__BB0_42; $L__BB0_49: setp.lt.ftz.f32 %p44, %f34, 0f00800000; mul.ftz.f32 %f537, %f34, 0f4B000000; selp.f32 %f61, %f537, %f34, %p44; selp.f32 %f538, 0fC1B80000, 0f00000000, %p44; mov.b32 %r37, %f61; add.s32 %r38, %r37, -1059760811; and.b32 %r39, %r38, -8388608; sub.s32 %r40, %r37, %r39; mov.b32 %f539, %r40; cvt.rn.f32.s32 %f540, %r39; mov.f32 %f541, 0f34000000; fma.rn.ftz.f32 %f542, %f540, %f541, %f538; add.ftz.f32 %f543, %f539, 0fBF800000; mov.f32 %f544, 0f3E1039F6; mov.f32 %f545, 0fBE055027; fma.rn.ftz.f32 %f546, %f545, %f543, %f544; mov.f32 %f547, 0fBDF8CDCC; fma.rn.ftz.f32 %f548, %f546, %f543, %f547; mov.f32 %f549, 0f3E0F2955; fma.rn.ftz.f32 %f550, %f548, %f543, %f549; mov.f32 %f551, 0fBE2AD8B9; fma.rn.ftz.f32 %f552, %f550, %f543, %f551; mov.f32 %f553, 0f3E4CED0B; fma.rn.ftz.f32 %f554, %f552, %f543, %f553; mov.f32 %f555, 0fBE7FFF22; fma.rn.ftz.f32 %f556, %f554, %f543, %f555; mov.f32 %f557, 0f3EAAAA78; fma.rn.ftz.f32 %f558, %f556, %f543, %f557; mov.f32 %f559, 0fBF000000; fma.rn.ftz.f32 %f560, %f558, %f543, %f559; mul.ftz.f32 %f561, %f543, %f560; fma.rn.ftz.f32 %f562, %f561, %f543, %f543; mov.f32 %f563, 0f3F317218; fma.rn.ftz.f32 %f814, %f542, %f563, %f562; setp.lt.u32 %p45, %r37, 2139095040; @%p45 bra $L__BB0_51; mov.f32 %f564, 0f7F800000; fma.rn.ftz.f32 %f814, %f61, %f564, %f564; $L__BB0_51: neg.ftz.f32 %f565, %f814; setp.eq.ftz.f32 %p46, %f61, 0f00000000; selp.f32 %f815, 0f7F800000, %f565, %p46; bra.uni $L__BB0_52; $L__BB0_42: add.ftz.f32 %f487, %f34, %f34; cvt.rni.ftz.f32.f32 %f488, %f487; cvt.rzi.ftz.s32.f32 %r3, %f488; neg.ftz.f32 %f489, %f488; mov.f32 %f490, 0f3F000000; fma.rn.ftz.f32 %f491, %f489, %f490, %f34; mul.ftz.f32 %f492, %f491, 0f40490FDB; mul.rn.ftz.f32 %f49, %f492, %f492; and.b32 %r4, %r3, 1; setp.eq.s32 %p38, %r4, 0; selp.f32 %f50, %f492, 0f3F800000, %p38; mov.f32 %f811, 0fB94D4153; @%p38 bra $L__BB0_44; mov.f32 %f493, 0fBAB607ED; mov.f32 %f494, 0f37CBAC00; fma.rn.ftz.f32 %f811, %f494, %f49, %f493; $L__BB0_44: selp.f32 %f495, 0f3C0885E4, 0f3D2AAABB, %p38; fma.rn.ftz.f32 %f496, %f811, %f49, %f495; selp.f32 %f497, 0fBE2AAAA8, 0fBEFFFFFF, %p38; fma.rn.ftz.f32 %f498, %f496, %f49, %f497; mov.f32 %f499, 0f00000000; fma.rn.ftz.f32 %f500, %f49, %f50, %f499; fma.rn.ftz.f32 %f812, %f498, %f500, %f50; and.b32 %r32, %r3, 2; setp.eq.s32 %p40, %r32, 0; @%p40 bra $L__BB0_46; mov.f32 %f502, 0fBF800000; fma.rn.ftz.f32 %f812, %f812, %f502, %f499; $L__BB0_46: abs.ftz.f32 %f503, %f812; mul.ftz.f32 %f504, %f34, %f503; setp.lt.ftz.f32 %p41, %f504, 0f00800000; mul.ftz.f32 %f505, %f504, 0f4B000000; selp.f32 %f506, 0fC1B80000, 0f00000000, %p41; selp.f32 %f56, %f505, %f504, %p41; mov.b32 %r33, %f56; add.s32 %r34, %r33, -1059760811; and.b32 %r35, %r34, -8388608; sub.s32 %r36, %r33, %r35; mov.b32 %f507, %r36; cvt.rn.f32.s32 %f508, %r35; mov.f32 %f509, 0f34000000; fma.rn.ftz.f32 %f510, %f508, %f509, %f506; add.ftz.f32 %f511, %f507, 0fBF800000; mov.f32 %f512, 0f3E1039F6; mov.f32 %f513, 0fBE055027; fma.rn.ftz.f32 %f514, %f513, %f511, %f512; mov.f32 %f515, 0fBDF8CDCC; fma.rn.ftz.f32 %f516, %f514, %f511, %f515; mov.f32 %f517, 0f3E0F2955; fma.rn.ftz.f32 %f518, %f516, %f511, %f517; mov.f32 %f519, 0fBE2AD8B9; fma.rn.ftz.f32 %f520, %f518, %f511, %f519; mov.f32 %f521, 0f3E4CED0B; fma.rn.ftz.f32 %f522, %f520, %f511, %f521; mov.f32 %f523, 0fBE7FFF22; fma.rn.ftz.f32 %f524, %f522, %f511, %f523; mov.f32 %f525, 0f3EAAAA78; fma.rn.ftz.f32 %f526, %f524, %f511, %f525; mov.f32 %f527, 0fBF000000; fma.rn.ftz.f32 %f528, %f526, %f511, %f527; mul.ftz.f32 %f529, %f528, %f511; fma.rn.ftz.f32 %f530, %f529, %f511, %f511; mov.f32 %f531, 0f3F317218; fma.rn.ftz.f32 %f813, %f510, %f531, %f530; setp.lt.u32 %p42, %r33, 2139095040; @%p42 bra $L__BB0_48; mov.f32 %f532, 0f7F800000; fma.rn.ftz.f32 %f813, %f56, %f532, %f532; $L__BB0_48: mov.f32 %f533, 0f3F928682; sub.ftz.f32 %f534, %f533, %f813; setp.eq.ftz.f32 %p43, %f56, 0f00000000; sub.ftz.f32 %f535, %f533, 0fFF800000; selp.f32 %f536, %f535, %f534, %p43; sub.ftz.f32 %f815, %f536, %f810; $L__BB0_52: add.ftz.f32 %f67, %f807, %f815; add.ftz.f32 %f68, %f102, %f103; abs.ftz.f32 %f69, %f68; setp.ltu.ftz.f32 %p47, %f69, 0f40400000; @%p47 bra $L__BB0_58; bra.uni $L__BB0_53; $L__BB0_58: setp.ltu.ftz.f32 %p53, %f69, 0f3FC00000; @%p53 bra $L__BB0_60; bra.uni $L__BB0_59; $L__BB0_60: setp.ltu.ftz.f32 %p54, %f69, 0f3F333333; @%p54 bra $L__BB0_62; bra.uni $L__BB0_61; $L__BB0_62: mov.f32 %f671, 0fBBB34878; mov.f32 %f672, 0f3B6B1C86; fma.rn.ftz.f32 %f673, %f672, %f69, %f671; mov.f32 %f674, 0fBD36CAEF; fma.rn.ftz.f32 %f675, %f673, %f69, %f674; mov.f32 %f676, 0f3E2B5555; fma.rn.ftz.f32 %f677, %f675, %f69, %f676; mov.f32 %f678, 0fBD2C96C7; fma.rn.ftz.f32 %f679, %f677, %f69, %f678; mov.f32 %f680, 0fBF27E6EB; fma.rn.ftz.f32 %f681, %f679, %f69, %f680; mov.f32 %f682, 0f3F13C463; fma.rn.ftz.f32 %f683, %f681, %f69, %f682; mul.ftz.f32 %f684, %f69, %f683; fma.rn.ftz.f32 %f685, %f684, %f69, %f69; setp.lt.ftz.f32 %p55, %f685, 0f00800000; mul.ftz.f32 %f686, %f685, 0f4B000000; selp.f32 %f78, %f686, %f685, %p55; selp.f32 %f687, 0fC1B80000, 0f00000000, %p55; mov.b32 %r45, %f78; add.s32 %r46, %r45, -1059760811; and.b32 %r47, %r46, -8388608; sub.s32 %r48, %r45, %r47; mov.b32 %f688, %r48; cvt.rn.f32.s32 %f689, %r47; mov.f32 %f690, 0f34000000; fma.rn.ftz.f32 %f691, %f689, %f690, %f687; add.ftz.f32 %f692, %f688, 0fBF800000; mov.f32 %f693, 0f3E1039F6; mov.f32 %f694, 0fBE055027; fma.rn.ftz.f32 %f695, %f694, %f692, %f693; mov.f32 %f696, 0fBDF8CDCC; fma.rn.ftz.f32 %f697, %f695, %f692, %f696; mov.f32 %f698, 0f3E0F2955; fma.rn.ftz.f32 %f699, %f697, %f692, %f698; mov.f32 %f700, 0fBE2AD8B9; fma.rn.ftz.f32 %f701, %f699, %f692, %f700; mov.f32 %f702, 0f3E4CED0B; fma.rn.ftz.f32 %f703, %f701, %f692, %f702; mov.f32 %f704, 0fBE7FFF22; fma.rn.ftz.f32 %f705, %f703, %f692, %f704; mov.f32 %f706, 0f3EAAAA78; fma.rn.ftz.f32 %f707, %f705, %f692, %f706; mov.f32 %f708, 0fBF000000; fma.rn.ftz.f32 %f709, %f707, %f692, %f708; mul.ftz.f32 %f710, %f692, %f709; fma.rn.ftz.f32 %f711, %f710, %f692, %f692; mov.f32 %f712, 0f3F317218; fma.rn.ftz.f32 %f817, %f691, %f712, %f711; setp.lt.u32 %p56, %r45, 2139095040; @%p56 bra $L__BB0_64; mov.f32 %f713, 0f7F800000; fma.rn.ftz.f32 %f817, %f78, %f713, %f713; $L__BB0_64: neg.ftz.f32 %f714, %f817; setp.eq.ftz.f32 %p57, %f78, 0f00000000; selp.f32 %f818, 0f7F800000, %f714, %p57; bra.uni $L__BB0_65; $L__BB0_53: setp.ltu.ftz.f32 %p48, %f69, 0f40F9999A; @%p48 bra $L__BB0_57; bra.uni $L__BB0_54; $L__BB0_57: add.ftz.f32 %f610, %f69, 0fC0400000; mov.f32 %f611, 0fC640F6F8; mov.f32 %f612, 0fC43B38FB; fma.rn.ftz.f32 %f613, %f612, %f610, %f611; mov.f32 %f614, 0fC7206560; fma.rn.ftz.f32 %f615, %f613, %f610, %f614; mov.f32 %f616, 0fC73CB6AA; fma.rn.ftz.f32 %f617, %f615, %f610, %f616; mov.f32 %f618, 0fC80BAE5A; fma.rn.ftz.f32 %f619, %f617, %f610, %f618; add.ftz.f32 %f620, %f610, 0fC381A020; mov.f32 %f621, 0fC62864B8; fma.rn.ftz.f32 %f622, %f620, %f610, %f621; mov.f32 %f623, 0fC7B50686; fma.rn.ftz.f32 %f624, %f622, %f610, %f623; mov.f32 %f625, 0fC8498465; fma.rn.ftz.f32 %f626, %f624, %f610, %f625; rcp.approx.ftz.f32 %f627, %f626; fma.rn.ftz.f32 %f818, %f619, %f627, %f610; bra.uni $L__BB0_65; $L__BB0_59: add.ftz.f32 %f628, %f69, 0fC0000000; mov.f32 %f629, 0fB967A002; mov.f32 %f630, 0f385007FA; fma.rn.ftz.f32 %f631, %f630, %f628, %f629; mov.f32 %f632, 0f3A0DE6FC; fma.rn.ftz.f32 %f633, %f631, %f628, %f632; mov.f32 %f634, 0fBA9DE0E2; fma.rn.ftz.f32 %f635, %f633, %f628, %f634; mov.f32 %f636, 0f3B3D05B7; fma.rn.ftz.f32 %f637, %f635, %f628, %f636; mov.f32 %f638, 0fBBF1EB10; fma.rn.ftz.f32 %f639, %f637, %f628, %f638; mov.f32 %f640, 0f3CA89A28; fma.rn.ftz.f32 %f641, %f639, %f628, %f640; mov.f32 %f642, 0fBD89F01A; fma.rn.ftz.f32 %f643, %f641, %f628, %f642; mov.f32 %f644, 0f3EA51A66; fma.rn.ftz.f32 %f645, %f643, %f628, %f644; mov.f32 %f646, 0f3ED87730; fma.rn.ftz.f32 %f647, %f645, %f628, %f646; mul.ftz.f32 %f818, %f628, %f647; bra.uni $L__BB0_65; $L__BB0_54: mul.ftz.f32 %f566, %f69, 0f4B000000; setp.lt.ftz.f32 %p49, %f69, 0f00800000; selp.f32 %f70, %f566, %f69, %p49; selp.f32 %f567, 0fC1B80000, 0f00000000, %p49; mov.b32 %r41, %f70; add.s32 %r42, %r41, -1059760811; and.b32 %r43, %r42, -8388608; sub.s32 %r44, %r41, %r43; mov.b32 %f568, %r44; cvt.rn.f32.s32 %f569, %r43; mov.f32 %f570, 0f34000000; fma.rn.ftz.f32 %f571, %f569, %f570, %f567; add.ftz.f32 %f572, %f568, 0fBF800000; mov.f32 %f573, 0f3E1039F6; mov.f32 %f574, 0fBE055027; fma.rn.ftz.f32 %f575, %f574, %f572, %f573; mov.f32 %f576, 0fBDF8CDCC; fma.rn.ftz.f32 %f577, %f575, %f572, %f576; mov.f32 %f578, 0f3E0F2955; fma.rn.ftz.f32 %f579, %f577, %f572, %f578; mov.f32 %f580, 0fBE2AD8B9; fma.rn.ftz.f32 %f581, %f579, %f572, %f580; mov.f32 %f582, 0f3E4CED0B; fma.rn.ftz.f32 %f583, %f581, %f572, %f582; mov.f32 %f584, 0fBE7FFF22; fma.rn.ftz.f32 %f585, %f583, %f572, %f584; mov.f32 %f586, 0f3EAAAA78; fma.rn.ftz.f32 %f587, %f585, %f572, %f586; mov.f32 %f588, 0fBF000000; fma.rn.ftz.f32 %f589, %f587, %f572, %f588; mul.ftz.f32 %f590, %f572, %f589; fma.rn.ftz.f32 %f591, %f590, %f572, %f572; mov.f32 %f592, 0f3F317218; fma.rn.ftz.f32 %f816, %f571, %f592, %f591; setp.lt.u32 %p50, %r41, 2139095040; @%p50 bra $L__BB0_56; mov.f32 %f593, 0f7F800000; fma.rn.ftz.f32 %f816, %f70, %f593, %f593; $L__BB0_56: rcp.approx.ftz.f32 %f594, %f69; mul.ftz.f32 %f595, %f594, %f594; mul.ftz.f32 %f596, %f816, 0f3F000000; setp.eq.ftz.f32 %p51, %f70, 0f00000000; selp.f32 %f597, 0fFF800000, %f596, %p51; add.ftz.f32 %f598, %f69, 0fBF000000; mul.rn.ftz.f32 %f599, %f597, %f598; sub.ftz.f32 %f600, %f599, %f69; mov.f32 %f601, 0fBB360953; mov.f32 %f602, 0f3A4BE755; fma.rn.ftz.f32 %f603, %f602, %f595, %f601; mov.f32 %f604, 0f3DAAAAA3; fma.rn.ftz.f32 %f605, %f603, %f595, %f604; mov.f32 %f606, 0f3F6B3F8E; fma.rn.ftz.f32 %f607, %f605, %f594, %f606; add.rn.ftz.f32 %f608, %f599, %f607; add.ftz.f32 %f609, %f600, %f608; setp.eq.ftz.f32 %p52, %f69, 0f7F800000; selp.f32 %f818, %f69, %f609, %p52; bra.uni $L__BB0_65; $L__BB0_61: mov.f32 %f648, 0f3F800000; sub.ftz.f32 %f649, %f648, %f69; mov.f32 %f650, 0f3DD47577; mov.f32 %f651, 0f3D3BEF76; fma.rn.ftz.f32 %f652, %f651, %f649, %f650; mov.f32 %f653, 0f3DFB8079; fma.rn.ftz.f32 %f654, %f652, %f649, %f653; mov.f32 %f655, 0f3E0295B5; fma.rn.ftz.f32 %f656, %f654, %f649, %f655; mov.f32 %f657, 0f3E12A765; fma.rn.ftz.f32 %f658, %f656, %f649, %f657; mov.f32 %f659, 0f3E2D6867; fma.rn.ftz.f32 %f660, %f658, %f649, %f659; mov.f32 %f661, 0f3E5462BF; fma.rn.ftz.f32 %f662, %f660, %f649, %f661; mov.f32 %f663, 0f3E8A8A72; fma.rn.ftz.f32 %f664, %f662, %f649, %f663; mov.f32 %f665, 0f3ECD26A4; fma.rn.ftz.f32 %f666, %f664, %f649, %f665; mov.f32 %f667, 0f3F528D32; fma.rn.ftz.f32 %f668, %f666, %f649, %f667; mov.f32 %f669, 0f3F13C468; fma.rn.ftz.f32 %f670, %f668, %f649, %f669; mul.ftz.f32 %f818, %f649, %f670; $L__BB0_65: setp.ge.ftz.f32 %p58, %f68, 0f00000000; mov.f32 %f823, %f818; @%p58 bra $L__BB0_78; cvt.rmi.ftz.f32.f32 %f716, %f69; setp.eq.ftz.f32 %p59, %f69, %f716; mov.f32 %f823, 0f7F800000; @%p59 bra $L__BB0_78; setp.lt.ftz.f32 %p60, %f69, 0f1FEC1E4A; @%p60 bra $L__BB0_75; bra.uni $L__BB0_68; $L__BB0_75: setp.lt.ftz.f32 %p67, %f69, 0f00800000; mul.ftz.f32 %f768, %f69, 0f4B000000; selp.f32 %f96, %f768, %f69, %p67; selp.f32 %f769, 0fC1B80000, 0f00000000, %p67; mov.b32 %r54, %f96; add.s32 %r55, %r54, -1059760811; and.b32 %r56, %r55, -8388608; sub.s32 %r57, %r54, %r56; mov.b32 %f770, %r57; cvt.rn.f32.s32 %f771, %r56; mov.f32 %f772, 0f34000000; fma.rn.ftz.f32 %f773, %f771, %f772, %f769; add.ftz.f32 %f774, %f770, 0fBF800000; mov.f32 %f775, 0f3E1039F6; mov.f32 %f776, 0fBE055027; fma.rn.ftz.f32 %f777, %f776, %f774, %f775; mov.f32 %f778, 0fBDF8CDCC; fma.rn.ftz.f32 %f779, %f777, %f774, %f778; mov.f32 %f780, 0f3E0F2955; fma.rn.ftz.f32 %f781, %f779, %f774, %f780; mov.f32 %f782, 0fBE2AD8B9; fma.rn.ftz.f32 %f783, %f781, %f774, %f782; mov.f32 %f784, 0f3E4CED0B; fma.rn.ftz.f32 %f785, %f783, %f774, %f784; mov.f32 %f786, 0fBE7FFF22; fma.rn.ftz.f32 %f787, %f785, %f774, %f786; mov.f32 %f788, 0f3EAAAA78; fma.rn.ftz.f32 %f789, %f787, %f774, %f788; mov.f32 %f790, 0fBF000000; fma.rn.ftz.f32 %f791, %f789, %f774, %f790; mul.ftz.f32 %f792, %f774, %f791; fma.rn.ftz.f32 %f793, %f792, %f774, %f774; mov.f32 %f794, 0f3F317218; fma.rn.ftz.f32 %f822, %f773, %f794, %f793; setp.lt.u32 %p68, %r54, 2139095040; @%p68 bra $L__BB0_77; mov.f32 %f795, 0f7F800000; fma.rn.ftz.f32 %f822, %f96, %f795, %f795; $L__BB0_77: neg.ftz.f32 %f796, %f822; setp.eq.ftz.f32 %p69, %f96, 0f00000000; selp.f32 %f823, 0f7F800000, %f796, %p69; bra.uni $L__BB0_78; $L__BB0_68: add.ftz.f32 %f718, %f69, %f69; cvt.rni.ftz.f32.f32 %f719, %f718; cvt.rzi.ftz.s32.f32 %r5, %f719; neg.ftz.f32 %f720, %f719; mov.f32 %f721, 0f3F000000; fma.rn.ftz.f32 %f722, %f720, %f721, %f69; mul.ftz.f32 %f723, %f722, 0f40490FDB; mul.rn.ftz.f32 %f84, %f723, %f723; and.b32 %r6, %r5, 1; setp.eq.s32 %p61, %r6, 0; selp.f32 %f85, %f723, 0f3F800000, %p61; mov.f32 %f819, 0fB94D4153; @%p61 bra $L__BB0_70; mov.f32 %f724, 0fBAB607ED; mov.f32 %f725, 0f37CBAC00; fma.rn.ftz.f32 %f819, %f725, %f84, %f724; $L__BB0_70: selp.f32 %f726, 0f3C0885E4, 0f3D2AAABB, %p61; fma.rn.ftz.f32 %f727, %f819, %f84, %f726; selp.f32 %f728, 0fBE2AAAA8, 0fBEFFFFFF, %p61; fma.rn.ftz.f32 %f729, %f727, %f84, %f728; mov.f32 %f730, 0f00000000; fma.rn.ftz.f32 %f731, %f84, %f85, %f730; fma.rn.ftz.f32 %f820, %f729, %f731, %f85; and.b32 %r49, %r5, 2; setp.eq.s32 %p63, %r49, 0; @%p63 bra $L__BB0_72; mov.f32 %f733, 0fBF800000; fma.rn.ftz.f32 %f820, %f820, %f733, %f730; $L__BB0_72: abs.ftz.f32 %f734, %f820; mul.ftz.f32 %f735, %f69, %f734; setp.lt.ftz.f32 %p64, %f735, 0f00800000; mul.ftz.f32 %f736, %f735, 0f4B000000; selp.f32 %f737, 0fC1B80000, 0f00000000, %p64; selp.f32 %f91, %f736, %f735, %p64; mov.b32 %r50, %f91; add.s32 %r51, %r50, -1059760811; and.b32 %r52, %r51, -8388608; sub.s32 %r53, %r50, %r52; mov.b32 %f738, %r53; cvt.rn.f32.s32 %f739, %r52; mov.f32 %f740, 0f34000000; fma.rn.ftz.f32 %f741, %f739, %f740, %f737; add.ftz.f32 %f742, %f738, 0fBF800000; mov.f32 %f743, 0f3E1039F6; mov.f32 %f744, 0fBE055027; fma.rn.ftz.f32 %f745, %f744, %f742, %f743; mov.f32 %f746, 0fBDF8CDCC; fma.rn.ftz.f32 %f747, %f745, %f742, %f746; mov.f32 %f748, 0f3E0F2955; fma.rn.ftz.f32 %f749, %f747, %f742, %f748; mov.f32 %f750, 0fBE2AD8B9; fma.rn.ftz.f32 %f751, %f749, %f742, %f750; mov.f32 %f752, 0f3E4CED0B; fma.rn.ftz.f32 %f753, %f751, %f742, %f752; mov.f32 %f754, 0fBE7FFF22; fma.rn.ftz.f32 %f755, %f753, %f742, %f754; mov.f32 %f756, 0f3EAAAA78; fma.rn.ftz.f32 %f757, %f755, %f742, %f756; mov.f32 %f758, 0fBF000000; fma.rn.ftz.f32 %f759, %f757, %f742, %f758; mul.ftz.f32 %f760, %f759, %f742; fma.rn.ftz.f32 %f761, %f760, %f742, %f742; mov.f32 %f762, 0f3F317218; fma.rn.ftz.f32 %f821, %f741, %f762, %f761; setp.lt.u32 %p65, %r50, 2139095040; @%p65 bra $L__BB0_74; mov.f32 %f763, 0f7F800000; fma.rn.ftz.f32 %f821, %f91, %f763, %f763; $L__BB0_74: mov.f32 %f764, 0f3F928682; sub.ftz.f32 %f765, %f764, %f821; setp.eq.ftz.f32 %p66, %f91, 0f00000000; sub.ftz.f32 %f766, %f764, 0fFF800000; selp.f32 %f767, %f766, %f765, %p66; sub.ftz.f32 %f823, %f767, %f818; $L__BB0_78: sub.ftz.f32 %f797, %f67, %f823; mul.ftz.f32 %f798, %f797, 0f3FB8AA3B; ex2.approx.ftz.f32 %f799, %f798; st.param.f32 [func_retval0+0], %f799; ret; } .func (.param .align 4 .b8 func_retval0[16]) _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff( .param .b64 _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_0, .param .b32 _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_1, .param .b32 _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_2, .param .b32 _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_3, .param .b32 _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_4 ) { .reg .f32 %f<102>; .reg .b32 %r<3>; .reg .b64 %rd<2>; ld.param.u64 %rd1, [_Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_0]; ld.param.u32 %r1, [_Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_1]; ld.param.u32 %r2, [_Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_2]; ld.param.f32 %f1, [_Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_3]; ld.param.f32 %f2, [_Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff_param_4]; cvt.rn.f32.u32 %f3, %r1; fma.rn.ftz.f32 %f4, %f3, %f1, 0fBF000000; cvt.rn.f32.u32 %f5, %r2; fma.rn.ftz.f32 %f6, %f5, %f2, 0fBF000000; cvt.rmi.ftz.f32.f32 %f7, %f4; cvt.rmi.ftz.f32.f32 %f8, %f6; sub.ftz.f32 %f9, %f4, %f7; sub.ftz.f32 %f10, %f6, %f8; mov.f32 %f11, 0f40400000; sub.ftz.f32 %f12, %f11, %f9; fma.rn.ftz.f32 %f13, %f9, %f12, 0fC0400000; fma.rn.ftz.f32 %f14, %f9, %f13, 0f3F800000; mul.ftz.f32 %f15, %f9, %f9; mul.ftz.f32 %f16, %f9, 0f40400000; add.ftz.f32 %f17, %f16, 0fC0C00000; fma.rn.ftz.f32 %f18, %f15, %f17, 0f40800000; mul.ftz.f32 %f19, %f18, 0f3E2AAAAB; fma.rn.ftz.f32 %f20, %f14, 0f3E2AAAAB, %f19; sub.ftz.f32 %f21, %f11, %f16; fma.rn.ftz.f32 %f22, %f9, %f21, 0f40400000; fma.rn.ftz.f32 %f23, %f9, %f22, 0f3F800000; mul.ftz.f32 %f24, %f9, %f15; mul.ftz.f32 %f25, %f24, 0f3E2AAAAB; fma.rn.ftz.f32 %f26, %f23, 0f3E2AAAAB, %f25; div.approx.ftz.f32 %f27, %f19, %f20; add.ftz.f32 %f28, %f27, 0fBF800000; add.ftz.f32 %f29, %f7, %f28; add.ftz.f32 %f30, %f29, 0f3F000000; div.approx.ftz.f32 %f31, %f30, %f3; div.approx.ftz.f32 %f32, %f25, %f26; add.ftz.f32 %f33, %f32, 0f3F800000; add.ftz.f32 %f34, %f7, %f33; add.ftz.f32 %f35, %f34, 0f3F000000; div.approx.ftz.f32 %f36, %f35, %f3; mul.ftz.f32 %f37, %f10, %f10; mul.ftz.f32 %f38, %f10, 0f40400000; add.ftz.f32 %f39, %f38, 0fC0C00000; fma.rn.ftz.f32 %f40, %f37, %f39, 0f40800000; mul.ftz.f32 %f41, %f40, 0f3E2AAAAB; sub.ftz.f32 %f42, %f11, %f10; fma.rn.ftz.f32 %f43, %f10, %f42, 0fC0400000; fma.rn.ftz.f32 %f44, %f10, %f43, 0f3F800000; fma.rn.ftz.f32 %f45, %f44, 0f3E2AAAAB, %f41; div.approx.ftz.f32 %f46, %f41, %f45; add.ftz.f32 %f47, %f46, 0fBF800000; add.ftz.f32 %f48, %f8, %f47; add.ftz.f32 %f49, %f48, 0f3F000000; div.approx.ftz.f32 %f50, %f49, %f5; mul.ftz.f32 %f51, %f10, %f37; mul.ftz.f32 %f52, %f51, 0f3E2AAAAB; sub.ftz.f32 %f53, %f11, %f38; fma.rn.ftz.f32 %f54, %f10, %f53, 0f40400000; fma.rn.ftz.f32 %f55, %f10, %f54, 0f3F800000; fma.rn.ftz.f32 %f56, %f55, 0f3E2AAAAB, %f52; div.approx.ftz.f32 %f57, %f52, %f56; add.ftz.f32 %f58, %f57, 0f3F800000; add.ftz.f32 %f59, %f8, %f58; add.ftz.f32 %f60, %f59, 0f3F000000; div.approx.ftz.f32 %f61, %f60, %f5; tex.2d.v4.f32.f32 {%f62, %f63, %f64, %f65}, [%rd1, {%f31, %f50}]; tex.2d.v4.f32.f32 {%f66, %f67, %f68, %f69}, [%rd1, {%f36, %f50}]; mul.ftz.f32 %f70, %f26, %f66; mul.ftz.f32 %f71, %f26, %f67; mul.ftz.f32 %f72, %f26, %f68; mul.ftz.f32 %f73, %f26, %f69; fma.rn.ftz.f32 %f74, %f20, %f62, %f70; fma.rn.ftz.f32 %f75, %f20, %f63, %f71; fma.rn.ftz.f32 %f76, %f20, %f64, %f72; fma.rn.ftz.f32 %f77, %f20, %f65, %f73; tex.2d.v4.f32.f32 {%f78, %f79, %f80, %f81}, [%rd1, {%f31, %f61}]; tex.2d.v4.f32.f32 {%f82, %f83, %f84, %f85}, [%rd1, {%f36, %f61}]; mul.ftz.f32 %f86, %f26, %f82; mul.ftz.f32 %f87, %f26, %f83; mul.ftz.f32 %f88, %f26, %f84; mul.ftz.f32 %f89, %f26, %f85; fma.rn.ftz.f32 %f90, %f20, %f78, %f86; fma.rn.ftz.f32 %f91, %f20, %f79, %f87; fma.rn.ftz.f32 %f92, %f20, %f80, %f88; fma.rn.ftz.f32 %f93, %f20, %f81, %f89; mul.ftz.f32 %f94, %f56, %f90; mul.ftz.f32 %f95, %f56, %f91; mul.ftz.f32 %f96, %f56, %f92; mul.ftz.f32 %f97, %f56, %f93; fma.rn.ftz.f32 %f98, %f45, %f77, %f97; fma.rn.ftz.f32 %f99, %f45, %f76, %f96; fma.rn.ftz.f32 %f100, %f45, %f75, %f95; fma.rn.ftz.f32 %f101, %f45, %f74, %f94; st.param.f32 [func_retval0+0], %f101; st.param.f32 [func_retval0+4], %f100; st.param.f32 [func_retval0+8], %f99; st.param.f32 [func_retval0+12], %f98; ret; } .func (.param .b32 func_retval0) _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff( .param .b64 _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_0, .param .b32 _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_1, .param .b32 _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_2, .param .b32 _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_3, .param .b32 _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_4 ) { .reg .f32 %f<84>; .reg .b32 %r<3>; .reg .b64 %rd<2>; ld.param.u64 %rd1, [_Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_0]; ld.param.u32 %r1, [_Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_1]; ld.param.u32 %r2, [_Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_2]; ld.param.f32 %f1, [_Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_3]; ld.param.f32 %f2, [_Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff_param_4]; cvt.rn.f32.u32 %f3, %r1; fma.rn.ftz.f32 %f4, %f3, %f1, 0fBF000000; cvt.rn.f32.u32 %f5, %r2; fma.rn.ftz.f32 %f6, %f5, %f2, 0fBF000000; cvt.rmi.ftz.f32.f32 %f7, %f4; cvt.rmi.ftz.f32.f32 %f8, %f6; sub.ftz.f32 %f9, %f4, %f7; sub.ftz.f32 %f10, %f6, %f8; mov.f32 %f11, 0f40400000; sub.ftz.f32 %f12, %f11, %f9; fma.rn.ftz.f32 %f13, %f9, %f12, 0fC0400000; fma.rn.ftz.f32 %f14, %f9, %f13, 0f3F800000; mul.ftz.f32 %f15, %f9, %f9; mul.ftz.f32 %f16, %f9, 0f40400000; add.ftz.f32 %f17, %f16, 0fC0C00000; fma.rn.ftz.f32 %f18, %f15, %f17, 0f40800000; mul.ftz.f32 %f19, %f18, 0f3E2AAAAB; fma.rn.ftz.f32 %f20, %f14, 0f3E2AAAAB, %f19; sub.ftz.f32 %f21, %f11, %f16; fma.rn.ftz.f32 %f22, %f9, %f21, 0f40400000; fma.rn.ftz.f32 %f23, %f9, %f22, 0f3F800000; mul.ftz.f32 %f24, %f9, %f15; mul.ftz.f32 %f25, %f24, 0f3E2AAAAB; fma.rn.ftz.f32 %f26, %f23, 0f3E2AAAAB, %f25; div.approx.ftz.f32 %f27, %f19, %f20; add.ftz.f32 %f28, %f27, 0fBF800000; add.ftz.f32 %f29, %f7, %f28; add.ftz.f32 %f30, %f29, 0f3F000000; div.approx.ftz.f32 %f31, %f30, %f3; div.approx.ftz.f32 %f32, %f25, %f26; add.ftz.f32 %f33, %f32, 0f3F800000; add.ftz.f32 %f34, %f7, %f33; add.ftz.f32 %f35, %f34, 0f3F000000; div.approx.ftz.f32 %f36, %f35, %f3; mul.ftz.f32 %f37, %f10, %f10; mul.ftz.f32 %f38, %f10, 0f40400000; add.ftz.f32 %f39, %f38, 0fC0C00000; fma.rn.ftz.f32 %f40, %f37, %f39, 0f40800000; mul.ftz.f32 %f41, %f40, 0f3E2AAAAB; sub.ftz.f32 %f42, %f11, %f10; fma.rn.ftz.f32 %f43, %f10, %f42, 0fC0400000; fma.rn.ftz.f32 %f44, %f10, %f43, 0f3F800000; fma.rn.ftz.f32 %f45, %f44, 0f3E2AAAAB, %f41; div.approx.ftz.f32 %f46, %f41, %f45; add.ftz.f32 %f47, %f46, 0fBF800000; add.ftz.f32 %f48, %f8, %f47; add.ftz.f32 %f49, %f48, 0f3F000000; div.approx.ftz.f32 %f50, %f49, %f5; mul.ftz.f32 %f51, %f10, %f37; mul.ftz.f32 %f52, %f51, 0f3E2AAAAB; sub.ftz.f32 %f53, %f11, %f38; fma.rn.ftz.f32 %f54, %f10, %f53, 0f40400000; fma.rn.ftz.f32 %f55, %f10, %f54, 0f3F800000; fma.rn.ftz.f32 %f56, %f55, 0f3E2AAAAB, %f52; div.approx.ftz.f32 %f57, %f52, %f56; add.ftz.f32 %f58, %f57, 0f3F800000; add.ftz.f32 %f59, %f8, %f58; add.ftz.f32 %f60, %f59, 0f3F000000; div.approx.ftz.f32 %f61, %f60, %f5; tex.2d.v4.f32.f32 {%f62, %f63, %f64, %f65}, [%rd1, {%f31, %f50}]; tex.2d.v4.f32.f32 {%f66, %f67, %f68, %f69}, [%rd1, {%f36, %f50}]; mul.ftz.f32 %f70, %f26, %f66; fma.rn.ftz.f32 %f71, %f20, %f62, %f70; tex.2d.v4.f32.f32 {%f72, %f73, %f74, %f75}, [%rd1, {%f31, %f61}]; tex.2d.v4.f32.f32 {%f76, %f77, %f78, %f79}, [%rd1, {%f36, %f61}]; mul.ftz.f32 %f80, %f26, %f76; fma.rn.ftz.f32 %f81, %f20, %f72, %f80; mul.ftz.f32 %f82, %f56, %f81; fma.rn.ftz.f32 %f83, %f45, %f71, %f82; st.param.f32 [func_retval0+0], %f83; ret; } .func (.param .b32 func_retval0) _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj( .param .b64 _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0, .param .b32 _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1, .param .b32 _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2, .param .b32 _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3, .param .b32 _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4 ) { .local .align 8 .b8 __local_depot3[88]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<520>; .reg .b16 %rs<331>; .reg .f32 %f<484>; .reg .b32 %r<3610>; .reg .b64 %rd<2395>; mov.u64 %SPL, __local_depot3; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd608, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0]; ld.param.f32 %f322, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1]; ld.param.f32 %f323, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2]; ld.param.f32 %f324, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; ld.param.u32 %r784, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4]; cvta.to.global.u64 %rd609, %rd608; ld.global.u64 %rd610, [%rd609]; mov.u64 %rd2298, 0; ld.u64 %rd612, [%rd610+696]; add.s64 %rd613, %rd610, %rd612; add.s64 %rd1, %rd613, 672; add.u64 %rd614, %SP, 0; add.u64 %rd2, %SPL, 0; mov.u32 %r3184, 2147483647; st.local.v2.u32 [%rd2], {%r3184, %r3184}; st.local.v2.u32 [%rd2+8], {%r3184, %r3184}; st.local.v2.u32 [%rd2+16], {%r3184, %r3184}; st.local.v2.u32 [%rd2+24], {%r3184, %r3184}; st.local.u32 [%rd2+32], %r3184; st.local.u64 [%rd2+40], %rd1; st.local.u64 [%rd2+48], %rd2298; st.local.u64 [%rd2+56], %rd2298; st.local.u64 [%rd2+64], %rd2298; add.s64 %rd5, %rd613, 696; setp.eq.s32 %p1, %r784, 0; @%p1 bra $L__BB3_16; setp.ne.s32 %p2, %r784, 1; @%p2 bra $L__BB3_17; cvt.rni.ftz.f32.f32 %f325, %f322; cvt.rzi.ftz.s32.f32 %r1, %f325; cvt.rni.ftz.f32.f32 %f326, %f323; cvt.rzi.ftz.s32.f32 %r2, %f326; cvt.rni.ftz.f32.f32 %f327, %f324; cvt.rzi.ftz.s32.f32 %r3, %f327; ld.u32 %r4, [%rd5]; mov.u64 %rd2297, 0; setp.eq.s32 %p3, %r4, 0; @%p3 bra $L__BB3_7; add.s64 %rd6, %rd1, 64; shr.u32 %r787, %r1, 12; cvt.u64.u32 %rd616, %r787; shl.b64 %rd617, %rd616, 42; shr.u32 %r788, %r2, 12; mul.wide.u32 %rd618, %r788, 2097152; or.b64 %rd619, %rd618, %rd617; shr.u32 %r789, %r3, 12; cvt.u64.u32 %rd620, %r789; or.b64 %rd7, %rd619, %rd620; mov.u32 %r3170, 0; $L__BB3_4: cvt.u64.u32 %rd8, %r3170; mul.wide.u32 %rd621, %r3170, 32; add.s64 %rd622, %rd6, %rd621; ld.u64 %rd623, [%rd622]; setp.eq.s64 %p4, %rd623, %rd7; add.s32 %r3170, %r3170, 1; @%p4 bra $L__BB3_6; setp.lt.u32 %p5, %r3170, %r4; @%p5 bra $L__BB3_4; bra.uni $L__BB3_7; $L__BB3_16: add.u64 %rd647, %SP, 72; add.u64 %rd648, %SPL, 72; st.local.u64 [%rd648], %rd614; add.ftz.f32 %f328, %f322, 0fBF000000; add.ftz.f32 %f329, %f323, 0fBF000000; add.ftz.f32 %f330, %f324, 0fBF000000; { // callseq 0, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd647; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f328; st.param.f32 [param1+4], %f329; st.param.f32 [param1+8], %f330; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f483, [retval0+0]; } // callseq 0 bra.uni $L__BB3_834; $L__BB3_17: add.u64 %rd652, %SPL, 80; st.local.u64 [%rd652], %rd614; add.ftz.f32 %f331, %f322, 0fBF000000; cvt.rmi.ftz.f32.f32 %f332, %f331; add.ftz.f32 %f333, %f323, 0fBF000000; cvt.rmi.ftz.f32.f32 %f334, %f333; add.ftz.f32 %f335, %f324, 0fBF000000; cvt.rmi.ftz.f32.f32 %f336, %f335; sub.ftz.f32 %f337, %f331, %f332; sub.ftz.f32 %f338, %f333, %f334; sub.ftz.f32 %f339, %f335, %f336; mov.f32 %f340, 0f40400000; sub.ftz.f32 %f341, %f340, %f337; fma.rn.ftz.f32 %f342, %f337, %f341, 0fC0400000; fma.rn.ftz.f32 %f343, %f337, %f342, 0f3F800000; mul.ftz.f32 %f344, %f337, %f337; mul.ftz.f32 %f345, %f337, 0f40400000; add.ftz.f32 %f346, %f345, 0fC0C00000; fma.rn.ftz.f32 %f347, %f344, %f346, 0f40800000; mul.ftz.f32 %f348, %f347, 0f3E2AAAAB; fma.rn.ftz.f32 %f7, %f343, 0f3E2AAAAB, %f348; sub.ftz.f32 %f349, %f340, %f345; fma.rn.ftz.f32 %f350, %f337, %f349, 0f40400000; fma.rn.ftz.f32 %f351, %f337, %f350, 0f3F800000; mul.ftz.f32 %f352, %f337, %f344; mul.ftz.f32 %f353, %f352, 0f3E2AAAAB; fma.rn.ftz.f32 %f8, %f351, 0f3E2AAAAB, %f353; sub.ftz.f32 %f354, %f340, %f338; fma.rn.ftz.f32 %f355, %f338, %f354, 0fC0400000; fma.rn.ftz.f32 %f356, %f338, %f355, 0f3F800000; mul.ftz.f32 %f357, %f338, %f338; mul.ftz.f32 %f358, %f338, 0f40400000; add.ftz.f32 %f359, %f358, 0fC0C00000; fma.rn.ftz.f32 %f360, %f357, %f359, 0f40800000; mul.ftz.f32 %f361, %f360, 0f3E2AAAAB; fma.rn.ftz.f32 %f9, %f356, 0f3E2AAAAB, %f361; sub.ftz.f32 %f362, %f340, %f358; fma.rn.ftz.f32 %f363, %f338, %f362, 0f40400000; fma.rn.ftz.f32 %f364, %f338, %f363, 0f3F800000; mul.ftz.f32 %f365, %f338, %f357; mul.ftz.f32 %f366, %f365, 0f3E2AAAAB; fma.rn.ftz.f32 %f10, %f364, 0f3E2AAAAB, %f366; sub.ftz.f32 %f367, %f340, %f339; fma.rn.ftz.f32 %f368, %f339, %f367, 0fC0400000; fma.rn.ftz.f32 %f369, %f339, %f368, 0f3F800000; mul.ftz.f32 %f370, %f339, %f339; mul.ftz.f32 %f371, %f339, 0f40400000; add.ftz.f32 %f372, %f371, 0fC0C00000; fma.rn.ftz.f32 %f373, %f370, %f372, 0f40800000; mul.ftz.f32 %f374, %f373, 0f3E2AAAAB; fma.rn.ftz.f32 %f11, %f369, 0f3E2AAAAB, %f374; sub.ftz.f32 %f375, %f340, %f371; fma.rn.ftz.f32 %f376, %f339, %f375, 0f40400000; fma.rn.ftz.f32 %f377, %f339, %f376, 0f3F800000; mul.ftz.f32 %f378, %f339, %f370; mul.ftz.f32 %f379, %f378, 0f3E2AAAAB; fma.rn.ftz.f32 %f12, %f377, 0f3E2AAAAB, %f379; div.approx.ftz.f32 %f380, %f348, %f7; add.ftz.f32 %f381, %f380, 0fBF800000; add.ftz.f32 %f13, %f332, %f381; div.approx.ftz.f32 %f382, %f353, %f8; add.ftz.f32 %f383, %f382, 0f3F800000; add.ftz.f32 %f14, %f332, %f383; div.approx.ftz.f32 %f384, %f361, %f9; add.ftz.f32 %f385, %f384, 0fBF800000; add.ftz.f32 %f15, %f334, %f385; div.approx.ftz.f32 %f386, %f366, %f10; add.ftz.f32 %f387, %f386, 0f3F800000; add.ftz.f32 %f16, %f334, %f387; div.approx.ftz.f32 %f388, %f374, %f11; add.ftz.f32 %f389, %f388, 0fBF800000; add.ftz.f32 %f17, %f336, %f389; div.approx.ftz.f32 %f390, %f379, %f12; add.ftz.f32 %f391, %f390, 0f3F800000; add.ftz.f32 %f18, %f336, %f391; cvt.rmi.ftz.f32.f32 %f19, %f13; cvt.rmi.ftz.f32.f32 %f392, %f15; cvt.rmi.ftz.f32.f32 %f393, %f17; sub.ftz.f32 %f20, %f15, %f392; sub.ftz.f32 %f21, %f17, %f393; cvt.rzi.ftz.s32.f32 %r7, %f19; cvt.rzi.ftz.s32.f32 %r8, %f392; cvt.rzi.ftz.s32.f32 %r9, %f393; and.b32 %r10, %r7, -8; and.b32 %r3272, %r7, -128; and.b32 %r12, %r7, -4096; add.s64 %rd17, %rd1, 64; shr.u32 %r824, %r9, 12; cvt.u64.u32 %rd18, %r824; shr.u32 %r825, %r8, 12; mul.wide.u32 %rd19, %r825, 2097152; shr.u32 %r826, %r7, 12; cvt.u64.u32 %rd654, %r826; bfi.b64 %rd21, %rd654, %rd18, 42, 32; ld.u32 %r13, [%rd5]; setp.eq.s32 %p10, %r13, 0; @%p10 bra $L__BB3_22; or.b64 %rd22, %rd21, %rd19; mov.u32 %r3171, 0; $L__BB3_19: cvt.u64.u32 %rd23, %r3171; mul.wide.u32 %rd655, %r3171, 32; add.s64 %rd656, %rd17, %rd655; ld.u64 %rd657, [%rd656]; setp.eq.s64 %p11, %rd657, %rd22; add.s32 %r3171, %r3171, 1; @%p11 bra $L__BB3_21; setp.lt.u32 %p12, %r3171, %r13; @%p12 bra $L__BB3_19; bra.uni $L__BB3_22; $L__BB3_6: shl.b64 %rd625, %rd8, 5; add.s64 %rd2297, %rd6, %rd625; $L__BB3_7: setp.eq.s64 %p6, %rd2297, 0; @%p6 bra $L__BB3_15; add.s64 %rd11, %rd2297, 8; ld.u64 %rd12, [%rd2297+8]; setp.eq.s64 %p7, %rd12, 0; @%p7 bra $L__BB3_14; add.s64 %rd13, %rd1, %rd12; and.b32 %r790, %r2, -4096; and.b32 %r791, %r1, -4096; mov.u64 %rd626, 1; st.local.v2.u32 [%rd2+24], {%r791, %r790}; and.b32 %r792, %r3, -4096; st.local.u32 [%rd2+32], %r792; st.local.u64 [%rd2+64], %rd13; shl.b32 %r793, %r1, 3; and.b32 %r794, %r793, 31744; and.b32 %r795, %r2, 3968; shr.u32 %r796, %r795, 2; or.b32 %r797, %r796, %r794; bfe.u32 %r798, %r3, 7, 5; or.b32 %r799, %r798, %r797; shr.u32 %r800, %r797, 6; mul.wide.u32 %rd627, %r800, 8; add.s64 %rd628, %rd13, %rd627; and.b32 %r801, %r799, 63; shl.b64 %rd629, %rd626, %r801; ld.u64 %rd630, [%rd628+4128]; and.b64 %rd631, %rd630, %rd629; setp.eq.s64 %p8, %rd631, 0; mul.wide.u32 %rd632, %r799, 8; add.s64 %rd633, %rd13, %rd632; add.s64 %rd14, %rd633, 8256; @%p8 bra $L__BB3_13; ld.u64 %rd634, [%rd14]; add.s64 %rd15, %rd13, %rd634; and.b32 %r802, %r1, -128; st.local.u32 [%rd2+12], %r802; and.b32 %r803, %r2, -128; st.local.u32 [%rd2+16], %r803; and.b32 %r804, %r3, -128; st.local.u32 [%rd2+20], %r804; st.local.u64 [%rd2+56], %rd15; shl.b32 %r805, %r1, 5; and.b32 %r806, %r805, 3840; shl.b32 %r807, %r2, 1; and.b32 %r808, %r807, 240; or.b32 %r809, %r808, %r806; bfe.u32 %r810, %r3, 3, 4; or.b32 %r811, %r810, %r809; shr.u32 %r812, %r809, 6; mul.wide.u32 %rd635, %r812, 8; add.s64 %rd636, %rd15, %rd635; and.b32 %r813, %r811, 63; shl.b64 %rd638, %rd626, %r813; ld.u64 %rd639, [%rd636+544]; and.b64 %rd640, %rd639, %rd638; setp.eq.s64 %p9, %rd640, 0; mul.wide.u32 %rd641, %r811, 8; add.s64 %rd642, %rd15, %rd641; add.s64 %rd16, %rd642, 1088; @%p9 bra $L__BB3_12; ld.u64 %rd643, [%rd16]; add.s64 %rd644, %rd15, %rd643; and.b32 %r814, %r2, -8; and.b32 %r815, %r1, -8; st.local.v2.u32 [%rd2], {%r815, %r814}; and.b32 %r816, %r3, -8; st.local.u32 [%rd2+8], %r816; st.local.u64 [%rd2+48], %rd644; shl.b32 %r817, %r1, 6; and.b32 %r818, %r817, 448; shl.b32 %r819, %r2, 3; and.b32 %r820, %r819, 56; or.b32 %r821, %r820, %r818; and.b32 %r822, %r3, 7; or.b32 %r823, %r821, %r822; mul.wide.u32 %rd645, %r823, 4; add.s64 %rd646, %rd644, %rd645; ld.f32 %f483, [%rd646+96]; bra.uni $L__BB3_834; $L__BB3_15: ld.f32 %f483, [%rd5+4]; bra.uni $L__BB3_834; $L__BB3_21: shl.b64 %rd659, %rd23, 5; add.s64 %rd2298, %rd17, %rd659; $L__BB3_22: setp.eq.s64 %p13, %rd2298, 0; @%p13 bra $L__BB3_30; add.s64 %rd26, %rd2298, 8; ld.u64 %rd27, [%rd2298+8]; setp.eq.s64 %p14, %rd27, 0; @%p14 bra $L__BB3_29; add.s64 %rd2315, %rd1, %rd27; and.b32 %r828, %r8, -4096; mov.u64 %rd660, 1; st.local.v2.u32 [%rd2+24], {%r12, %r828}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r829, %r7, 3; and.b32 %r830, %r829, 31744; and.b32 %r831, %r8, 3968; shr.u32 %r832, %r831, 2; or.b32 %r833, %r832, %r830; bfe.u32 %r834, %r9, 7, 5; or.b32 %r835, %r833, %r834; shr.u32 %r836, %r833, 6; mul.wide.u32 %rd661, %r836, 8; add.s64 %rd662, %rd2315, %rd661; and.b32 %r837, %r835, 63; shl.b64 %rd663, %rd660, %r837; ld.u64 %rd664, [%rd662+4128]; and.b64 %rd665, %rd664, %rd663; setp.eq.s64 %p15, %rd665, 0; mul.wide.u32 %rd666, %r835, 8; add.s64 %rd667, %rd2315, %rd666; add.s64 %rd29, %rd667, 8256; mov.b64 %rd668, {%r12, %r828}; shr.u64 %rd669, %rd668, 32; cvt.u32.u64 %r3277, %rd669; @%p15 bra $L__BB3_28; ld.u64 %rd670, [%rd29]; add.s64 %rd2311, %rd2315, %rd670; st.local.u32 [%rd2+12], %r3272; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r838, %r7, 5; and.b32 %r839, %r838, 3840; shl.b32 %r840, %r8, 1; and.b32 %r841, %r840, 240; or.b32 %r842, %r841, %r839; bfe.u32 %r843, %r9, 3, 4; or.b32 %r844, %r842, %r843; shr.u32 %r845, %r842, 6; mul.wide.u32 %rd671, %r845, 8; add.s64 %rd672, %rd2311, %rd671; and.b32 %r846, %r844, 63; shl.b64 %rd674, %rd660, %r846; ld.u64 %rd675, [%rd672+544]; and.b64 %rd676, %rd675, %rd674; setp.eq.s64 %p16, %rd676, 0; mul.wide.u32 %rd677, %r844, 8; add.s64 %rd678, %rd2311, %rd677; add.s64 %rd31, %rd678, 1088; @%p16 bra $L__BB3_27; ld.u64 %rd679, [%rd31]; add.s64 %rd2307, %rd2311, %rd679; and.b32 %r847, %r8, -8; st.local.v2.u32 [%rd2], {%r10, %r847}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r848, %r7, 6; and.b32 %r849, %r848, 448; shl.b32 %r850, %r8, 3; and.b32 %r851, %r850, 56; or.b32 %r852, %r851, %r849; and.b32 %r853, %r9, 7; or.b32 %r854, %r852, %r853; mul.wide.u32 %rd680, %r854, 4; add.s64 %rd681, %rd2307, %rd680; ld.f32 %f459, [%rd681+96]; mov.b64 %rd682, {%r10, %r847}; shr.u64 %rd683, %rd682, 32; cvt.u32.u64 %r3184, %rd683; mov.u32 %r3232, %r12; mov.u32 %r3229, %r10; bra.uni $L__BB3_31; $L__BB3_30: ld.f32 %f459, [%rd5+4]; mov.u32 %r3277, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r3232, %r3277; mov.u32 %r3184, %r3277; mov.u32 %r3310, %r3277; mov.u32 %r3270, %r3277; mov.u32 %r3271, %r3277; mov.u32 %r3272, %r3277; mov.u32 %r3227, %r3277; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r3229, %r3277; bra.uni $L__BB3_31; $L__BB3_14: ld.f32 %f483, [%rd11+12]; bra.uni $L__BB3_834; $L__BB3_29: ld.f32 %f459, [%rd26+12]; mov.u32 %r3277, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r3232, %r3277; mov.u32 %r3184, %r3277; mov.u32 %r3310, %r3277; mov.u32 %r3270, %r3277; mov.u32 %r3271, %r3277; mov.u32 %r3272, %r3277; mov.u32 %r3227, %r3277; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r3229, %r3277; bra.uni $L__BB3_31; $L__BB3_13: ld.f32 %f483, [%rd14]; bra.uni $L__BB3_834; $L__BB3_28: ld.f32 %f459, [%rd29]; mov.u32 %r3184, 2147483647; mov.u64 %rd2311, 0; mov.u32 %r3232, %r12; mov.u32 %r3270, %r3184; mov.u32 %r3271, %r3184; mov.u32 %r3272, %r3184; mov.u32 %r3227, %r3184; mov.u64 %rd2307, %rd2311; mov.u32 %r3229, %r3184; bra.uni $L__BB3_31; $L__BB3_12: ld.f32 %f483, [%rd16]; bra.uni $L__BB3_834; $L__BB3_27: ld.f32 %f459, [%rd31]; mov.u64 %rd2307, 0; mov.u32 %r3232, %r12; mov.u32 %r3227, %r3184; mov.u32 %r3229, %r3184; $L__BB3_31: add.s32 %r74, %r9, 1; setp.ne.s32 %p17, %r10, %r3229; mov.u16 %rs265, 0; @%p17 bra $L__BB3_34; and.b32 %r994, %r8, -8; setp.ne.s32 %p18, %r994, %r3184; @%p18 bra $L__BB3_34; and.b32 %r995, %r74, -8; setp.eq.s32 %p19, %r995, %r3227; selp.u16 %rs265, 1, 0, %p19; $L__BB3_34: setp.eq.s16 %p20, %rs265, 0; @%p20 bra $L__BB3_36; shl.b32 %r996, %r7, 6; and.b32 %r997, %r996, 448; shl.b32 %r998, %r8, 3; and.b32 %r999, %r998, 56; or.b32 %r1000, %r999, %r997; and.b32 %r1001, %r74, 7; or.b32 %r1002, %r1000, %r1001; mul.wide.u32 %rd698, %r1002, 4; add.s64 %rd699, %rd2307, %rd698; ld.f32 %f460, [%rd699+96]; $L__BB3_66: and.b32 %r3124, %r7, -8; sub.ftz.f32 %f394, %f460, %f459; fma.rn.ftz.f32 %f40, %f21, %f394, %f459; add.s32 %r139, %r8, 1; setp.ne.s32 %p39, %r3124, %r3229; mov.u16 %rs268, 0; @%p39 bra $L__BB3_69; and.b32 %r1079, %r139, -8; setp.ne.s32 %p40, %r1079, %r3184; @%p40 bra $L__BB3_69; and.b32 %r1080, %r74, -8; setp.eq.s32 %p41, %r1080, %r3227; selp.u16 %rs268, 1, 0, %p41; $L__BB3_69: setp.eq.s16 %p42, %rs268, 0; @%p42 bra $L__BB3_71; shl.b32 %r1081, %r7, 6; and.b32 %r1082, %r1081, 448; shl.b32 %r1083, %r139, 3; and.b32 %r1084, %r1083, 56; or.b32 %r1085, %r1084, %r1082; and.b32 %r1086, %r74, 7; or.b32 %r1087, %r1085, %r1086; mul.wide.u32 %rd767, %r1087, 4; add.s64 %rd768, %rd2307, %rd767; ld.f32 %f461, [%rd768+96]; $L__BB3_101: and.b32 %r3094, %r7, -8; setp.ne.s32 %p61, %r3094, %r3229; mov.u16 %rs271, 0; @%p61 bra $L__BB3_104; and.b32 %r1165, %r139, -8; setp.ne.s32 %p62, %r1165, %r3184; @%p62 bra $L__BB3_104; and.b32 %r1166, %r9, -8; setp.eq.s32 %p63, %r1166, %r3227; selp.u16 %rs271, 1, 0, %p63; $L__BB3_104: setp.eq.s16 %p64, %rs271, 0; @%p64 bra $L__BB3_106; and.b32 %r1167, %r7, 7; shl.b32 %r1168, %r139, 3; and.b32 %r1169, %r1168, 56; and.b32 %r1170, %r9, 7; bfi.b32 %r1171, %r1167, %r1170, 6, 3; or.b32 %r1172, %r1171, %r1169; mul.wide.u32 %rd837, %r1172, 4; add.s64 %rd838, %rd2307, %rd837; ld.f32 %f462, [%rd838+96]; $L__BB3_136: sub.ftz.f32 %f395, %f461, %f462; fma.rn.ftz.f32 %f396, %f21, %f395, %f462; sub.ftz.f32 %f397, %f396, %f40; fma.rn.ftz.f32 %f65, %f20, %f397, %f40; add.s32 %r259, %r7, 1; and.b32 %r260, %r259, -8; setp.ne.s32 %p83, %r260, %r3229; mov.u16 %rs274, 0; @%p83 bra $L__BB3_139; and.b32 %r1246, %r139, -8; setp.ne.s32 %p84, %r1246, %r3184; @%p84 bra $L__BB3_139; and.b32 %r1247, %r9, -8; setp.eq.s32 %p85, %r1247, %r3227; selp.u16 %rs274, 1, 0, %p85; $L__BB3_139: setp.eq.s16 %p86, %rs274, 0; @%p86 bra $L__BB3_141; and.b32 %r1248, %r259, 7; shl.b32 %r1249, %r139, 3; and.b32 %r1250, %r1249, 56; and.b32 %r1251, %r9, 7; bfi.b32 %r1252, %r1248, %r1251, 6, 3; or.b32 %r1253, %r1252, %r1250; mul.wide.u32 %rd905, %r1253, 4; add.s64 %rd906, %rd2307, %rd905; ld.f32 %f463, [%rd906+96]; $L__BB3_171: and.b32 %r3165, %r259, -8; setp.ne.s32 %p105, %r3165, %r3229; mov.u16 %rs277, 0; @%p105 bra $L__BB3_174; and.b32 %r1328, %r139, -8; setp.ne.s32 %p106, %r1328, %r3184; @%p106 bra $L__BB3_174; and.b32 %r1329, %r74, -8; setp.eq.s32 %p107, %r1329, %r3227; selp.u16 %rs277, 1, 0, %p107; $L__BB3_174: setp.eq.s16 %p108, %rs277, 0; @%p108 bra $L__BB3_176; shl.b32 %r1330, %r259, 6; and.b32 %r1331, %r1330, 448; shl.b32 %r1332, %r139, 3; and.b32 %r1333, %r1332, 56; or.b32 %r1334, %r1333, %r1331; and.b32 %r1335, %r74, 7; or.b32 %r1336, %r1334, %r1335; mul.wide.u32 %rd976, %r1336, 4; add.s64 %rd977, %rd2307, %rd976; ld.f32 %f464, [%rd977+96]; $L__BB3_206: and.b32 %r3161, %r259, -8; sub.ftz.f32 %f398, %f464, %f463; fma.rn.ftz.f32 %f90, %f21, %f398, %f463; setp.ne.s32 %p127, %r3161, %r3229; mov.u16 %rs280, 0; @%p127 bra $L__BB3_209; and.b32 %r1415, %r8, -8; setp.ne.s32 %p128, %r1415, %r3184; @%p128 bra $L__BB3_209; and.b32 %r1416, %r74, -8; setp.eq.s32 %p129, %r1416, %r3227; selp.u16 %rs280, 1, 0, %p129; $L__BB3_209: setp.eq.s16 %p130, %rs280, 0; @%p130 bra $L__BB3_211; shl.b32 %r1417, %r259, 6; and.b32 %r1418, %r1417, 448; shl.b32 %r1419, %r8, 3; and.b32 %r1420, %r1419, 56; or.b32 %r1421, %r1418, %r1420; and.b32 %r1422, %r74, 7; or.b32 %r1423, %r1421, %r1422; mul.wide.u32 %rd1047, %r1423, 4; add.s64 %rd1048, %rd2307, %rd1047; ld.f32 %f465, [%rd1048+96]; $L__BB3_241: and.b32 %r3145, %r259, -8; setp.ne.s32 %p149, %r3145, %r3229; mov.u16 %rs283, 0; @%p149 bra $L__BB3_244; and.b32 %r1501, %r8, -8; setp.ne.s32 %p150, %r1501, %r3184; @%p150 bra $L__BB3_244; and.b32 %r1502, %r9, -8; setp.eq.s32 %p151, %r1502, %r3227; selp.u16 %rs283, 1, 0, %p151; $L__BB3_244: setp.eq.s16 %p152, %rs283, 0; @%p152 bra $L__BB3_246; shl.b32 %r1503, %r259, 6; and.b32 %r1504, %r1503, 448; shl.b32 %r1505, %r8, 3; and.b32 %r1506, %r1505, 56; or.b32 %r1507, %r1504, %r1506; and.b32 %r1508, %r9, 7; or.b32 %r1509, %r1507, %r1508; mul.wide.u32 %rd1118, %r1509, 4; add.s64 %rd1119, %rd2307, %rd1118; ld.f32 %f466, [%rd1119+96]; $L__BB3_276: cvt.rmi.ftz.f32.f32 %f445, %f13; sub.ftz.f32 %f399, %f465, %f466; fma.rn.ftz.f32 %f400, %f21, %f399, %f466; sub.ftz.f32 %f401, %f90, %f400; fma.rn.ftz.f32 %f402, %f20, %f401, %f400; sub.ftz.f32 %f403, %f402, %f65; sub.ftz.f32 %f404, %f13, %f445; fma.rn.ftz.f32 %f115, %f404, %f403, %f65; cvt.rmi.ftz.f32.f32 %f405, %f14; sub.ftz.f32 %f116, %f14, %f405; cvt.rzi.ftz.s32.f32 %r471, %f405; and.b32 %r472, %r471, -8; setp.ne.s32 %p171, %r472, %r3229; mov.u16 %rs286, 0; @%p171 bra $L__BB3_279; and.b32 %r1586, %r8, -8; setp.ne.s32 %p172, %r1586, %r3184; @%p172 bra $L__BB3_279; and.b32 %r1587, %r9, -8; setp.eq.s32 %p173, %r1587, %r3227; selp.u16 %rs286, 1, 0, %p173; $L__BB3_279: setp.eq.s16 %p174, %rs286, 0; @%p174 bra $L__BB3_281; shl.b32 %r1588, %r471, 6; and.b32 %r1589, %r1588, 448; shl.b32 %r1590, %r8, 3; and.b32 %r1591, %r1590, 56; or.b32 %r1592, %r1591, %r1589; and.b32 %r1593, %r9, 7; or.b32 %r1594, %r1592, %r1593; mul.wide.u32 %rd1188, %r1594, 4; add.s64 %rd1189, %rd2307, %rd1188; ld.f32 %f467, [%rd1189+96]; $L__BB3_311: setp.ne.s32 %p193, %r472, %r3229; mov.u16 %rs289, 0; @%p193 bra $L__BB3_314; and.b32 %r1671, %r8, -8; setp.ne.s32 %p194, %r1671, %r3184; @%p194 bra $L__BB3_314; and.b32 %r1672, %r74, -8; setp.eq.s32 %p195, %r1672, %r3227; selp.u16 %rs289, 1, 0, %p195; $L__BB3_314: setp.eq.s16 %p196, %rs289, 0; @%p196 bra $L__BB3_316; shl.b32 %r1673, %r471, 6; and.b32 %r1674, %r1673, 448; shl.b32 %r1675, %r8, 3; and.b32 %r1676, %r1675, 56; or.b32 %r1677, %r1676, %r1674; and.b32 %r1678, %r74, 7; or.b32 %r1679, %r1677, %r1678; mul.wide.u32 %rd1258, %r1679, 4; add.s64 %rd1259, %rd2307, %rd1258; ld.f32 %f468, [%rd1259+96]; $L__BB3_346: sub.ftz.f32 %f406, %f468, %f467; fma.rn.ftz.f32 %f141, %f21, %f406, %f467; setp.ne.s32 %p215, %r472, %r3229; mov.u16 %rs292, 0; @%p215 bra $L__BB3_349; and.b32 %r1757, %r139, -8; setp.ne.s32 %p216, %r1757, %r3184; @%p216 bra $L__BB3_349; and.b32 %r1758, %r74, -8; setp.eq.s32 %p217, %r1758, %r3227; selp.u16 %rs292, 1, 0, %p217; $L__BB3_349: setp.eq.s16 %p218, %rs292, 0; @%p218 bra $L__BB3_351; shl.b32 %r1759, %r471, 6; and.b32 %r1760, %r1759, 448; shl.b32 %r1761, %r139, 3; and.b32 %r1762, %r1761, 56; or.b32 %r1763, %r1762, %r1760; and.b32 %r1764, %r74, 7; or.b32 %r1765, %r1763, %r1764; mul.wide.u32 %rd1329, %r1765, 4; add.s64 %rd1330, %rd2307, %rd1329; ld.f32 %f469, [%rd1330+96]; $L__BB3_381: setp.ne.s32 %p237, %r472, %r3229; mov.u16 %rs295, 0; @%p237 bra $L__BB3_384; and.b32 %r1844, %r139, -8; setp.ne.s32 %p238, %r1844, %r3184; @%p238 bra $L__BB3_384; and.b32 %r1845, %r9, -8; setp.eq.s32 %p239, %r1845, %r3227; selp.u16 %rs295, 1, 0, %p239; $L__BB3_384: setp.eq.s16 %p240, %rs295, 0; @%p240 bra $L__BB3_386; and.b32 %r1846, %r471, 7; shl.b32 %r1847, %r139, 3; and.b32 %r1848, %r1847, 56; and.b32 %r1849, %r9, 7; bfi.b32 %r1850, %r1846, %r1849, 6, 3; or.b32 %r1851, %r1850, %r1848; mul.wide.u32 %rd1400, %r1851, 4; add.s64 %rd1401, %rd2307, %rd1400; ld.f32 %f470, [%rd1401+96]; $L__BB3_416: sub.ftz.f32 %f407, %f469, %f470; fma.rn.ftz.f32 %f408, %f21, %f407, %f470; sub.ftz.f32 %f409, %f408, %f141; fma.rn.ftz.f32 %f166, %f20, %f409, %f141; add.s32 %r635, %r471, 1; and.b32 %r636, %r635, -8; setp.ne.s32 %p259, %r636, %r3229; mov.u16 %rs298, 0; @%p259 bra $L__BB3_419; and.b32 %r1926, %r139, -8; setp.ne.s32 %p260, %r1926, %r3184; @%p260 bra $L__BB3_419; and.b32 %r1927, %r9, -8; setp.eq.s32 %p261, %r1927, %r3227; selp.u16 %rs298, 1, 0, %p261; $L__BB3_419: setp.eq.s16 %p262, %rs298, 0; @%p262 bra $L__BB3_421; and.b32 %r1928, %r635, 7; shl.b32 %r1929, %r139, 3; and.b32 %r1930, %r1929, 56; and.b32 %r1931, %r9, 7; bfi.b32 %r1932, %r1928, %r1931, 6, 3; or.b32 %r1933, %r1932, %r1930; mul.wide.u32 %rd1471, %r1933, 4; add.s64 %rd1472, %rd2307, %rd1471; ld.f32 %f471, [%rd1472+96]; $L__BB3_451: setp.ne.s32 %p281, %r636, %r3229; mov.u16 %rs301, 0; @%p281 bra $L__BB3_454; and.b32 %r2008, %r139, -8; setp.ne.s32 %p282, %r2008, %r3184; @%p282 bra $L__BB3_454; and.b32 %r2009, %r74, -8; setp.eq.s32 %p283, %r2009, %r3227; selp.u16 %rs301, 1, 0, %p283; $L__BB3_454: setp.eq.s16 %p284, %rs301, 0; @%p284 bra $L__BB3_456; shl.b32 %r2010, %r635, 6; and.b32 %r2011, %r2010, 448; shl.b32 %r2012, %r139, 3; and.b32 %r2013, %r2012, 56; or.b32 %r2014, %r2013, %r2011; and.b32 %r2015, %r74, 7; or.b32 %r2016, %r2014, %r2015; mul.wide.u32 %rd1542, %r2016, 4; add.s64 %rd1543, %rd2307, %rd1542; ld.f32 %f472, [%rd1543+96]; $L__BB3_486: sub.ftz.f32 %f410, %f472, %f471; fma.rn.ftz.f32 %f191, %f21, %f410, %f471; setp.ne.s32 %p303, %r636, %r3229; mov.u16 %rs304, 0; @%p303 bra $L__BB3_489; and.b32 %r2095, %r8, -8; setp.ne.s32 %p304, %r2095, %r3184; @%p304 bra $L__BB3_489; and.b32 %r2096, %r74, -8; setp.eq.s32 %p305, %r2096, %r3227; selp.u16 %rs304, 1, 0, %p305; $L__BB3_489: setp.eq.s16 %p306, %rs304, 0; @%p306 bra $L__BB3_491; shl.b32 %r2097, %r635, 6; and.b32 %r2098, %r2097, 448; shl.b32 %r2099, %r8, 3; and.b32 %r2100, %r2099, 56; or.b32 %r2101, %r2098, %r2100; and.b32 %r2102, %r74, 7; or.b32 %r2103, %r2101, %r2102; mul.wide.u32 %rd1613, %r2103, 4; add.s64 %rd1614, %rd2307, %rd1613; ld.f32 %f473, [%rd1614+96]; $L__BB3_521: setp.ne.s32 %p325, %r636, %r3229; mov.u16 %rs307, 0; @%p325 bra $L__BB3_524; and.b32 %r2181, %r8, -8; setp.ne.s32 %p326, %r2181, %r3184; @%p326 bra $L__BB3_524; and.b32 %r2182, %r9, -8; setp.eq.s32 %p327, %r2182, %r3227; selp.u16 %rs307, 1, 0, %p327; $L__BB3_524: setp.eq.s16 %p328, %rs307, 0; @%p328 bra $L__BB3_526; shl.b32 %r2183, %r635, 6; and.b32 %r2184, %r2183, 448; shl.b32 %r2185, %r8, 3; and.b32 %r2186, %r2185, 56; or.b32 %r2187, %r2184, %r2186; and.b32 %r2188, %r9, 7; or.b32 %r2189, %r2187, %r2188; mul.wide.u32 %rd1684, %r2189, 4; add.s64 %rd1685, %rd2307, %rd1684; ld.f32 %f474, [%rd1685+96]; $L__BB3_556: add.u64 %rd2288, %SP, 80; sub.ftz.f32 %f411, %f473, %f474; fma.rn.ftz.f32 %f412, %f21, %f411, %f474; sub.ftz.f32 %f413, %f191, %f412; fma.rn.ftz.f32 %f414, %f20, %f413, %f412; sub.ftz.f32 %f415, %f414, %f166; fma.rn.ftz.f32 %f416, %f116, %f415, %f166; mul.ftz.f32 %f417, %f8, %f416; fma.rn.ftz.f32 %f418, %f7, %f115, %f417; mul.ftz.f32 %f419, %f9, %f418; { // callseq 1, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f420, [retval0+0]; } // callseq 1 { // callseq 2, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f421, [retval0+0]; } // callseq 2 mul.ftz.f32 %f422, %f8, %f421; fma.rn.ftz.f32 %f423, %f7, %f420, %f422; fma.rn.ftz.f32 %f424, %f10, %f423, %f419; mul.ftz.f32 %f216, %f11, %f424; { // callseq 3, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f425, [retval0+0]; } // callseq 3 { // callseq 4, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f426, [retval0+0]; } // callseq 4 mul.ftz.f32 %f427, %f8, %f426; fma.rn.ftz.f32 %f428, %f7, %f425, %f427; mul.ftz.f32 %f217, %f9, %f428; { // callseq 5, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f429, [retval0+0]; } // callseq 5 mul.ftz.f32 %f218, %f7, %f429; cvt.rmi.ftz.f32.f32 %f430, %f16; sub.ftz.f32 %f219, %f16, %f430; cvt.rmi.ftz.f32.f32 %f431, %f18; sub.ftz.f32 %f220, %f18, %f431; cvt.rzi.ftz.s32.f32 %r732, %f430; cvt.rzi.ftz.s32.f32 %r733, %f431; ld.local.u32 %r3598, [%rd2]; setp.ne.s32 %p347, %r472, %r3598; mov.u16 %rs310, 0; @%p347 bra $L__BB3_559; and.b32 %r2274, %r732, -8; ld.local.u32 %r2275, [%rd2+4]; setp.ne.s32 %p348, %r2274, %r2275; @%p348 bra $L__BB3_559; and.b32 %r2276, %r733, -8; ld.local.u32 %r2277, [%rd2+8]; setp.eq.s32 %p349, %r2276, %r2277; selp.u16 %rs310, 1, 0, %p349; $L__BB3_559: setp.eq.s16 %p350, %rs310, 0; @%p350 bra $L__BB3_561; shl.b32 %r2278, %r471, 6; and.b32 %r2279, %r2278, 448; shl.b32 %r2280, %r732, 3; and.b32 %r2281, %r2280, 56; or.b32 %r2282, %r2281, %r2279; and.b32 %r2283, %r733, 7; or.b32 %r2284, %r2282, %r2283; ld.local.u64 %rd1749, [%rd2+48]; mul.wide.u32 %rd1750, %r2284, 4; add.s64 %rd1751, %rd1749, %rd1750; ld.f32 %f475, [%rd1751+96]; $L__BB3_591: add.s32 %r741, %r733, 1; setp.ne.s32 %p369, %r472, %r3598; mov.u16 %rs313, 0; @%p369 bra $L__BB3_594; and.b32 %r2377, %r732, -8; ld.local.u32 %r2378, [%rd2+4]; setp.ne.s32 %p370, %r2377, %r2378; @%p370 bra $L__BB3_594; and.b32 %r2379, %r741, -8; ld.local.u32 %r2380, [%rd2+8]; setp.eq.s32 %p371, %r2379, %r2380; selp.u16 %rs313, 1, 0, %p371; $L__BB3_594: setp.eq.s16 %p372, %rs313, 0; @%p372 bra $L__BB3_596; shl.b32 %r2381, %r471, 6; and.b32 %r2382, %r2381, 448; shl.b32 %r2383, %r732, 3; and.b32 %r2384, %r2383, 56; or.b32 %r2385, %r2384, %r2382; and.b32 %r2386, %r741, 7; or.b32 %r2387, %r2385, %r2386; ld.local.u64 %rd1816, [%rd2+48]; mul.wide.u32 %rd1817, %r2387, 4; add.s64 %rd1818, %rd1816, %rd1817; ld.f32 %f476, [%rd1818+96]; $L__BB3_626: sub.ftz.f32 %f432, %f476, %f475; fma.rn.ftz.f32 %f245, %f220, %f432, %f475; add.s32 %r748, %r732, 1; setp.ne.s32 %p391, %r472, %r3598; mov.u16 %rs316, 0; @%p391 bra $L__BB3_629; and.b32 %r2480, %r748, -8; ld.local.u32 %r2481, [%rd2+4]; setp.ne.s32 %p392, %r2480, %r2481; @%p392 bra $L__BB3_629; and.b32 %r2482, %r741, -8; ld.local.u32 %r2483, [%rd2+8]; setp.eq.s32 %p393, %r2482, %r2483; selp.u16 %rs316, 1, 0, %p393; $L__BB3_629: setp.eq.s16 %p394, %rs316, 0; @%p394 bra $L__BB3_631; shl.b32 %r2484, %r471, 6; and.b32 %r2485, %r2484, 448; shl.b32 %r2486, %r748, 3; and.b32 %r2487, %r2486, 56; or.b32 %r2488, %r2487, %r2485; and.b32 %r2489, %r741, 7; or.b32 %r2490, %r2488, %r2489; ld.local.u64 %rd1884, [%rd2+48]; mul.wide.u32 %rd1885, %r2490, 4; add.s64 %rd1886, %rd1884, %rd1885; ld.f32 %f477, [%rd1886+96]; $L__BB3_661: setp.ne.s32 %p413, %r472, %r3598; mov.u16 %rs319, 0; @%p413 bra $L__BB3_664; and.b32 %r2583, %r748, -8; ld.local.u32 %r2584, [%rd2+4]; setp.ne.s32 %p414, %r2583, %r2584; @%p414 bra $L__BB3_664; and.b32 %r2585, %r733, -8; ld.local.u32 %r2586, [%rd2+8]; setp.eq.s32 %p415, %r2585, %r2586; selp.u16 %rs319, 1, 0, %p415; $L__BB3_664: setp.eq.s16 %p416, %rs319, 0; @%p416 bra $L__BB3_666; and.b32 %r2587, %r471, 7; shl.b32 %r2588, %r748, 3; and.b32 %r2589, %r2588, 56; and.b32 %r2590, %r733, 7; bfi.b32 %r2591, %r2587, %r2590, 6, 3; or.b32 %r2592, %r2591, %r2589; ld.local.u64 %rd1951, [%rd2+48]; mul.wide.u32 %rd1952, %r2592, 4; add.s64 %rd1953, %rd1951, %rd1952; ld.f32 %f478, [%rd1953+96]; $L__BB3_696: sub.ftz.f32 %f433, %f477, %f478; fma.rn.ftz.f32 %f434, %f220, %f433, %f478; sub.ftz.f32 %f435, %f434, %f245; fma.rn.ftz.f32 %f270, %f219, %f435, %f245; setp.ne.s32 %p435, %r636, %r3598; mov.u16 %rs322, 0; @%p435 bra $L__BB3_699; and.b32 %r2682, %r748, -8; ld.local.u32 %r2683, [%rd2+4]; setp.ne.s32 %p436, %r2682, %r2683; @%p436 bra $L__BB3_699; and.b32 %r2684, %r733, -8; ld.local.u32 %r2685, [%rd2+8]; setp.eq.s32 %p437, %r2684, %r2685; selp.u16 %rs322, 1, 0, %p437; $L__BB3_699: setp.eq.s16 %p438, %rs322, 0; @%p438 bra $L__BB3_701; and.b32 %r2686, %r635, 7; shl.b32 %r2687, %r748, 3; and.b32 %r2688, %r2687, 56; and.b32 %r2689, %r733, 7; bfi.b32 %r2690, %r2686, %r2689, 6, 3; or.b32 %r2691, %r2690, %r2688; ld.local.u64 %rd2018, [%rd2+48]; mul.wide.u32 %rd2019, %r2691, 4; add.s64 %rd2020, %rd2018, %rd2019; ld.f32 %f479, [%rd2020+96]; $L__BB3_731: setp.ne.s32 %p457, %r636, %r3598; mov.u16 %rs325, 0; @%p457 bra $L__BB3_734; and.b32 %r2781, %r748, -8; ld.local.u32 %r2782, [%rd2+4]; setp.ne.s32 %p458, %r2781, %r2782; @%p458 bra $L__BB3_734; and.b32 %r2783, %r741, -8; ld.local.u32 %r2784, [%rd2+8]; setp.eq.s32 %p459, %r2783, %r2784; selp.u16 %rs325, 1, 0, %p459; $L__BB3_734: setp.eq.s16 %p460, %rs325, 0; @%p460 bra $L__BB3_736; shl.b32 %r2785, %r635, 6; and.b32 %r2786, %r2785, 448; shl.b32 %r2787, %r748, 3; and.b32 %r2788, %r2787, 56; or.b32 %r2789, %r2788, %r2786; and.b32 %r2790, %r741, 7; or.b32 %r2791, %r2789, %r2790; ld.local.u64 %rd2085, [%rd2+48]; mul.wide.u32 %rd2086, %r2791, 4; add.s64 %rd2087, %rd2085, %rd2086; ld.f32 %f480, [%rd2087+96]; $L__BB3_766: sub.ftz.f32 %f436, %f480, %f479; fma.rn.ftz.f32 %f295, %f220, %f436, %f479; setp.ne.s32 %p479, %r636, %r3598; mov.u16 %rs328, 0; @%p479 bra $L__BB3_769; and.b32 %r2884, %r732, -8; ld.local.u32 %r2885, [%rd2+4]; setp.ne.s32 %p480, %r2884, %r2885; @%p480 bra $L__BB3_769; and.b32 %r2886, %r741, -8; ld.local.u32 %r2887, [%rd2+8]; setp.eq.s32 %p481, %r2886, %r2887; selp.u16 %rs328, 1, 0, %p481; $L__BB3_769: setp.eq.s16 %p482, %rs328, 0; @%p482 bra $L__BB3_771; shl.b32 %r2888, %r635, 6; and.b32 %r2889, %r2888, 448; shl.b32 %r2890, %r732, 3; and.b32 %r2891, %r2890, 56; or.b32 %r2892, %r2889, %r2891; and.b32 %r2893, %r741, 7; or.b32 %r2894, %r2892, %r2893; ld.local.u64 %rd2152, [%rd2+48]; mul.wide.u32 %rd2153, %r2894, 4; add.s64 %rd2154, %rd2152, %rd2153; ld.f32 %f481, [%rd2154+96]; $L__BB3_801: setp.ne.s32 %p501, %r636, %r3598; @%p501 bra $L__BB3_804; and.b32 %r2987, %r732, -8; ld.local.u32 %r2988, [%rd2+4]; setp.ne.s32 %p502, %r2987, %r2988; @%p502 bra $L__BB3_804; and.b32 %r2989, %r733, -8; ld.local.u32 %r2990, [%rd2+8]; setp.eq.s32 %p503, %r2989, %r2990; @%p503 bra $L__BB3_832; bra.uni $L__BB3_804; $L__BB3_832: shl.b32 %r3083, %r635, 6; and.b32 %r3084, %r3083, 448; shl.b32 %r3085, %r732, 3; and.b32 %r3086, %r3085, 56; or.b32 %r3087, %r3084, %r3086; and.b32 %r3088, %r733, 7; or.b32 %r3089, %r3087, %r3088; ld.local.u64 %rd2285, [%rd2+48]; mul.wide.u32 %rd2286, %r3089, 4; add.s64 %rd2287, %rd2285, %rd2286; ld.f32 %f482, [%rd2287+96]; bra.uni $L__BB3_833; $L__BB3_804: and.b32 %r779, %r635, -128; ld.local.u32 %r2991, [%rd2+12]; setp.ne.s32 %p504, %r779, %r2991; @%p504 bra $L__BB3_807; and.b32 %r2992, %r732, -128; ld.local.u32 %r2993, [%rd2+16]; setp.ne.s32 %p505, %r2992, %r2993; @%p505 bra $L__BB3_807; and.b32 %r2994, %r733, -128; ld.local.u32 %r2995, [%rd2+20]; setp.eq.s32 %p506, %r2994, %r2995; @%p506 bra $L__BB3_829; bra.uni $L__BB3_807; $L__BB3_829: shl.b32 %r3065, %r635, 5; and.b32 %r3066, %r3065, 3840; shl.b32 %r3067, %r732, 1; and.b32 %r3068, %r3067, 240; or.b32 %r3069, %r3066, %r3068; bfe.u32 %r3070, %r733, 3, 4; or.b32 %r3071, %r3069, %r3070; shr.u32 %r3072, %r3069, 6; ld.local.u64 %rd606, [%rd2+56]; mul.wide.u32 %rd2273, %r3072, 8; add.s64 %rd2274, %rd606, %rd2273; and.b32 %r3073, %r3071, 63; mov.u64 %rd2275, 1; shl.b64 %rd2276, %rd2275, %r3073; ld.u64 %rd2277, [%rd2274+544]; and.b64 %rd2278, %rd2277, %rd2276; setp.eq.s64 %p519, %rd2278, 0; mul.wide.u32 %rd2279, %r3071, 8; add.s64 %rd2280, %rd606, %rd2279; add.s64 %rd607, %rd2280, 1088; @%p519 bra $L__BB3_831; ld.u64 %rd2281, [%rd607]; add.s64 %rd2282, %rd606, %rd2281; and.b32 %r3074, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3074}; and.b32 %r3075, %r733, -8; st.local.u32 [%rd2+8], %r3075; st.local.u64 [%rd2+48], %rd2282; shl.b32 %r3076, %r635, 6; and.b32 %r3077, %r3076, 448; shl.b32 %r3078, %r732, 3; and.b32 %r3079, %r3078, 56; or.b32 %r3080, %r3077, %r3079; and.b32 %r3081, %r733, 7; or.b32 %r3082, %r3080, %r3081; mul.wide.u32 %rd2283, %r3082, 4; add.s64 %rd2284, %rd2282, %rd2283; ld.f32 %f482, [%rd2284+96]; bra.uni $L__BB3_833; $L__BB3_807: and.b32 %r780, %r635, -4096; ld.local.u32 %r2996, [%rd2+24]; setp.ne.s32 %p507, %r780, %r2996; @%p507 bra $L__BB3_810; and.b32 %r2997, %r732, -4096; ld.local.u32 %r2998, [%rd2+28]; setp.ne.s32 %p508, %r2997, %r2998; @%p508 bra $L__BB3_810; and.b32 %r2999, %r733, -4096; ld.local.u32 %r3000, [%rd2+32]; setp.eq.s32 %p509, %r2999, %r3000; @%p509 bra $L__BB3_824; bra.uni $L__BB3_810; $L__BB3_824: shl.b32 %r3036, %r635, 3; and.b32 %r3037, %r3036, 31744; and.b32 %r3038, %r732, 3968; shr.u32 %r3039, %r3038, 2; or.b32 %r3040, %r3037, %r3039; bfe.u32 %r3041, %r733, 7, 5; or.b32 %r3042, %r3040, %r3041; shr.u32 %r3043, %r3040, 6; ld.local.u64 %rd602, [%rd2+64]; mul.wide.u32 %rd2252, %r3043, 8; add.s64 %rd2253, %rd602, %rd2252; and.b32 %r3044, %r3042, 63; mov.u64 %rd2254, 1; shl.b64 %rd2255, %rd2254, %r3044; ld.u64 %rd2256, [%rd2253+4128]; and.b64 %rd2257, %rd2256, %rd2255; setp.eq.s64 %p517, %rd2257, 0; mul.wide.u32 %rd2258, %r3042, 8; add.s64 %rd2259, %rd602, %rd2258; add.s64 %rd603, %rd2259, 8256; @%p517 bra $L__BB3_828; ld.u64 %rd2260, [%rd603]; add.s64 %rd604, %rd602, %rd2260; st.local.u32 [%rd2+12], %r779; and.b32 %r3045, %r732, -128; st.local.u32 [%rd2+16], %r3045; and.b32 %r3046, %r733, -128; st.local.u32 [%rd2+20], %r3046; st.local.u64 [%rd2+56], %rd604; shl.b32 %r3047, %r635, 5; and.b32 %r3048, %r3047, 3840; shl.b32 %r3049, %r732, 1; and.b32 %r3050, %r3049, 240; or.b32 %r3051, %r3048, %r3050; bfe.u32 %r3052, %r733, 3, 4; or.b32 %r3053, %r3051, %r3052; shr.u32 %r3054, %r3051, 6; mul.wide.u32 %rd2261, %r3054, 8; add.s64 %rd2262, %rd604, %rd2261; and.b32 %r3055, %r3053, 63; shl.b64 %rd2264, %rd2254, %r3055; ld.u64 %rd2265, [%rd2262+544]; and.b64 %rd2266, %rd2265, %rd2264; setp.eq.s64 %p518, %rd2266, 0; mul.wide.u32 %rd2267, %r3053, 8; add.s64 %rd2268, %rd604, %rd2267; add.s64 %rd605, %rd2268, 1088; @%p518 bra $L__BB3_827; ld.u64 %rd2269, [%rd605]; add.s64 %rd2270, %rd604, %rd2269; and.b32 %r3056, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3056}; and.b32 %r3057, %r733, -8; st.local.u32 [%rd2+8], %r3057; st.local.u64 [%rd2+48], %rd2270; shl.b32 %r3058, %r635, 6; and.b32 %r3059, %r3058, 448; shl.b32 %r3060, %r732, 3; and.b32 %r3061, %r3060, 56; or.b32 %r3062, %r3059, %r3061; and.b32 %r3063, %r733, 7; or.b32 %r3064, %r3062, %r3063; mul.wide.u32 %rd2271, %r3064, 4; add.s64 %rd2272, %rd2270, %rd2271; ld.f32 %f482, [%rd2272+96]; bra.uni $L__BB3_833; $L__BB3_810: ld.local.u64 %rd590, [%rd2+40]; mov.u64 %rd2394, 0; ld.u32 %r781, [%rd590+24]; setp.eq.s32 %p510, %r781, 0; @%p510 bra $L__BB3_815; add.s64 %rd591, %rd590, 64; shr.u32 %r3002, %r635, 12; cvt.u64.u32 %rd2221, %r3002; shl.b64 %rd2222, %rd2221, 42; shr.u32 %r3003, %r732, 12; mul.wide.u32 %rd2223, %r3003, 2097152; shr.u32 %r3004, %r733, 12; cvt.u64.u32 %rd2224, %r3004; or.b64 %rd2225, %rd2223, %rd2224; or.b64 %rd592, %rd2225, %rd2222; mov.u32 %r3609, 0; $L__BB3_812: cvt.u64.u32 %rd593, %r3609; mul.wide.u32 %rd2226, %r3609, 32; add.s64 %rd2227, %rd591, %rd2226; ld.u64 %rd2228, [%rd2227]; setp.eq.s64 %p511, %rd2228, %rd592; add.s32 %r3609, %r3609, 1; @%p511 bra $L__BB3_814; setp.lt.u32 %p512, %r3609, %r781; @%p512 bra $L__BB3_812; bra.uni $L__BB3_815; $L__BB3_36: and.b32 %r3156, %r7, -128; setp.ne.s32 %p21, %r3156, %r3272; mov.u16 %rs266, 0; @%p21 bra $L__BB3_39; and.b32 %r1003, %r8, -128; setp.ne.s32 %p22, %r1003, %r3271; @%p22 bra $L__BB3_39; and.b32 %r1004, %r74, -128; setp.eq.s32 %p23, %r1004, %r3270; selp.u16 %rs266, 1, 0, %p23; $L__BB3_39: setp.eq.s16 %p24, %rs266, 0; @%p24 bra $L__BB3_43; shl.b32 %r1005, %r7, 5; and.b32 %r1006, %r1005, 3840; shl.b32 %r1007, %r8, 1; and.b32 %r1008, %r1007, 240; or.b32 %r1009, %r1008, %r1006; bfe.u32 %r1010, %r74, 3, 4; or.b32 %r1011, %r1010, %r1009; shr.u32 %r1012, %r1009, 6; mul.wide.u32 %rd700, %r1012, 8; add.s64 %rd701, %rd2311, %rd700; and.b32 %r1013, %r1011, 63; mov.u64 %rd702, 1; shl.b64 %rd703, %rd702, %r1013; ld.u64 %rd704, [%rd701+544]; and.b64 %rd705, %rd704, %rd703; setp.eq.s64 %p25, %rd705, 0; mul.wide.u32 %rd706, %r1011, 8; add.s64 %rd707, %rd2311, %rd706; add.s64 %rd43, %rd707, 1088; @%p25 bra $L__BB3_42; and.b32 %r3229, %r7, -8; ld.u64 %rd708, [%rd43]; add.s64 %rd2307, %rd2311, %rd708; and.b32 %r1014, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1014}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1015, %r7, 6; and.b32 %r1016, %r1015, 448; shl.b32 %r1017, %r8, 3; and.b32 %r1018, %r1017, 56; or.b32 %r1019, %r1018, %r1016; and.b32 %r1020, %r74, 7; or.b32 %r1021, %r1019, %r1020; mul.wide.u32 %rd709, %r1021, 4; add.s64 %rd710, %rd2307, %rd709; ld.f32 %f460, [%rd710+96]; mov.b64 %rd711, {%r3229, %r1014}; shr.u64 %rd712, %rd711, 32; cvt.u32.u64 %r3184, %rd712; bra.uni $L__BB3_66; $L__BB3_71: and.b32 %r3154, %r7, -128; setp.ne.s32 %p43, %r3154, %r3272; mov.u16 %rs269, 0; @%p43 bra $L__BB3_74; and.b32 %r1088, %r139, -128; setp.ne.s32 %p44, %r1088, %r3271; @%p44 bra $L__BB3_74; and.b32 %r1089, %r74, -128; setp.eq.s32 %p45, %r1089, %r3270; selp.u16 %rs269, 1, 0, %p45; $L__BB3_74: setp.eq.s16 %p46, %rs269, 0; @%p46 bra $L__BB3_78; shl.b32 %r1090, %r7, 5; and.b32 %r1091, %r1090, 3840; shl.b32 %r1092, %r139, 1; and.b32 %r1093, %r1092, 240; or.b32 %r1094, %r1093, %r1091; bfe.u32 %r1095, %r74, 3, 4; or.b32 %r1096, %r1094, %r1095; shr.u32 %r1097, %r1094, 6; mul.wide.u32 %rd769, %r1097, 8; add.s64 %rd770, %rd2311, %rd769; and.b32 %r1098, %r1096, 63; mov.u64 %rd771, 1; shl.b64 %rd772, %rd771, %r1098; ld.u64 %rd773, [%rd770+544]; and.b64 %rd774, %rd773, %rd772; setp.eq.s64 %p47, %rd774, 0; mul.wide.u32 %rd775, %r1096, 8; add.s64 %rd776, %rd2311, %rd775; add.s64 %rd72, %rd776, 1088; @%p47 bra $L__BB3_77; and.b32 %r3229, %r7, -8; ld.u64 %rd777, [%rd72]; add.s64 %rd2307, %rd2311, %rd777; and.b32 %r1099, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1099}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1100, %r7, 6; and.b32 %r1101, %r1100, 448; shl.b32 %r1102, %r139, 3; and.b32 %r1103, %r1102, 56; or.b32 %r1104, %r1103, %r1101; and.b32 %r1105, %r74, 7; or.b32 %r1106, %r1104, %r1105; mul.wide.u32 %rd778, %r1106, 4; add.s64 %rd779, %rd2307, %rd778; ld.f32 %f461, [%rd779+96]; mov.b64 %rd780, {%r3229, %r1099}; shr.u64 %rd781, %rd780, 32; cvt.u32.u64 %r3184, %rd781; bra.uni $L__BB3_101; $L__BB3_106: and.b32 %r3127, %r7, -128; setp.ne.s32 %p65, %r3127, %r3272; mov.u16 %rs272, 0; @%p65 bra $L__BB3_109; and.b32 %r1173, %r139, -128; setp.ne.s32 %p66, %r1173, %r3271; @%p66 bra $L__BB3_109; and.b32 %r1174, %r9, -128; setp.eq.s32 %p67, %r1174, %r3270; selp.u16 %rs272, 1, 0, %p67; $L__BB3_109: setp.eq.s16 %p68, %rs272, 0; @%p68 bra $L__BB3_113; shl.b32 %r1175, %r7, 5; and.b32 %r1176, %r1175, 3840; shl.b32 %r1177, %r139, 1; and.b32 %r1178, %r1177, 240; or.b32 %r1179, %r1178, %r1176; bfe.u32 %r1180, %r9, 3, 4; or.b32 %r1181, %r1179, %r1180; shr.u32 %r1182, %r1179, 6; mul.wide.u32 %rd839, %r1182, 8; add.s64 %rd840, %rd2311, %rd839; and.b32 %r1183, %r1181, 63; mov.u64 %rd841, 1; shl.b64 %rd842, %rd841, %r1183; ld.u64 %rd843, [%rd840+544]; and.b64 %rd844, %rd843, %rd842; setp.eq.s64 %p69, %rd844, 0; mul.wide.u32 %rd845, %r1181, 8; add.s64 %rd846, %rd2311, %rd845; add.s64 %rd101, %rd846, 1088; @%p69 bra $L__BB3_112; and.b32 %r3229, %r7, -8; ld.u64 %rd847, [%rd101]; add.s64 %rd2307, %rd2311, %rd847; and.b32 %r1184, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1184}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1185, %r7, 7; shl.b32 %r1186, %r139, 3; and.b32 %r1187, %r1186, 56; and.b32 %r1188, %r9, 7; bfi.b32 %r1189, %r1185, %r1188, 6, 3; or.b32 %r1190, %r1189, %r1187; mul.wide.u32 %rd848, %r1190, 4; add.s64 %rd849, %rd2307, %rd848; ld.f32 %f462, [%rd849+96]; mov.b64 %rd850, {%r3229, %r1184}; shr.u64 %rd851, %rd850, 32; cvt.u32.u64 %r3184, %rd851; bra.uni $L__BB3_136; $L__BB3_141: and.b32 %r261, %r259, -128; setp.ne.s32 %p87, %r261, %r3272; mov.u16 %rs275, 0; @%p87 bra $L__BB3_144; and.b32 %r1254, %r139, -128; setp.ne.s32 %p88, %r1254, %r3271; @%p88 bra $L__BB3_144; and.b32 %r1255, %r9, -128; setp.eq.s32 %p89, %r1255, %r3270; selp.u16 %rs275, 1, 0, %p89; $L__BB3_144: setp.eq.s16 %p90, %rs275, 0; @%p90 bra $L__BB3_148; shl.b32 %r1256, %r259, 5; and.b32 %r1257, %r1256, 3840; shl.b32 %r1258, %r139, 1; and.b32 %r1259, %r1258, 240; or.b32 %r1260, %r1259, %r1257; bfe.u32 %r1261, %r9, 3, 4; or.b32 %r1262, %r1260, %r1261; shr.u32 %r1263, %r1260, 6; mul.wide.u32 %rd907, %r1263, 8; add.s64 %rd908, %rd2311, %rd907; and.b32 %r1264, %r1262, 63; mov.u64 %rd909, 1; shl.b64 %rd910, %rd909, %r1264; ld.u64 %rd911, [%rd908+544]; and.b64 %rd912, %rd911, %rd910; setp.eq.s64 %p91, %rd912, 0; mul.wide.u32 %rd913, %r1262, 8; add.s64 %rd914, %rd2311, %rd913; add.s64 %rd130, %rd914, 1088; @%p91 bra $L__BB3_147; and.b32 %r3229, %r259, -8; ld.u64 %rd915, [%rd130]; add.s64 %rd2307, %rd2311, %rd915; and.b32 %r1265, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1265}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1266, %r259, 7; shl.b32 %r1267, %r139, 3; and.b32 %r1268, %r1267, 56; and.b32 %r1269, %r9, 7; bfi.b32 %r1270, %r1266, %r1269, 6, 3; or.b32 %r1271, %r1270, %r1268; mul.wide.u32 %rd916, %r1271, 4; add.s64 %rd917, %rd2307, %rd916; ld.f32 %f463, [%rd917+96]; mov.b64 %rd918, {%r3229, %r1265}; shr.u64 %rd919, %rd918, 32; cvt.u32.u64 %r3184, %rd919; bra.uni $L__BB3_171; $L__BB3_176: and.b32 %r318, %r259, -128; setp.ne.s32 %p109, %r318, %r3272; mov.u16 %rs278, 0; @%p109 bra $L__BB3_179; and.b32 %r1337, %r139, -128; setp.ne.s32 %p110, %r1337, %r3271; @%p110 bra $L__BB3_179; and.b32 %r1338, %r74, -128; setp.eq.s32 %p111, %r1338, %r3270; selp.u16 %rs278, 1, 0, %p111; $L__BB3_179: setp.eq.s16 %p112, %rs278, 0; @%p112 bra $L__BB3_183; shl.b32 %r1339, %r259, 5; and.b32 %r1340, %r1339, 3840; shl.b32 %r1341, %r139, 1; and.b32 %r1342, %r1341, 240; or.b32 %r1343, %r1342, %r1340; bfe.u32 %r1344, %r74, 3, 4; or.b32 %r1345, %r1343, %r1344; shr.u32 %r1346, %r1343, 6; mul.wide.u32 %rd978, %r1346, 8; add.s64 %rd979, %rd2311, %rd978; and.b32 %r1347, %r1345, 63; mov.u64 %rd980, 1; shl.b64 %rd981, %rd980, %r1347; ld.u64 %rd982, [%rd979+544]; and.b64 %rd983, %rd982, %rd981; setp.eq.s64 %p113, %rd983, 0; mul.wide.u32 %rd984, %r1345, 8; add.s64 %rd985, %rd2311, %rd984; add.s64 %rd159, %rd985, 1088; @%p113 bra $L__BB3_182; and.b32 %r3229, %r259, -8; ld.u64 %rd986, [%rd159]; add.s64 %rd2307, %rd2311, %rd986; and.b32 %r1348, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1348}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1349, %r259, 6; and.b32 %r1350, %r1349, 448; shl.b32 %r1351, %r139, 3; and.b32 %r1352, %r1351, 56; or.b32 %r1353, %r1352, %r1350; and.b32 %r1354, %r74, 7; or.b32 %r1355, %r1353, %r1354; mul.wide.u32 %rd987, %r1355, 4; add.s64 %rd988, %rd2307, %rd987; ld.f32 %f464, [%rd988+96]; mov.b64 %rd989, {%r3229, %r1348}; shr.u64 %rd990, %rd989, 32; cvt.u32.u64 %r3184, %rd990; bra.uni $L__BB3_206; $L__BB3_211: and.b32 %r372, %r259, -128; setp.ne.s32 %p131, %r372, %r3272; mov.u16 %rs281, 0; @%p131 bra $L__BB3_214; and.b32 %r1424, %r8, -128; setp.ne.s32 %p132, %r1424, %r3271; @%p132 bra $L__BB3_214; and.b32 %r1425, %r74, -128; setp.eq.s32 %p133, %r1425, %r3270; selp.u16 %rs281, 1, 0, %p133; $L__BB3_214: setp.eq.s16 %p134, %rs281, 0; @%p134 bra $L__BB3_218; shl.b32 %r1426, %r259, 5; and.b32 %r1427, %r1426, 3840; shl.b32 %r1428, %r8, 1; and.b32 %r1429, %r1428, 240; or.b32 %r1430, %r1427, %r1429; bfe.u32 %r1431, %r74, 3, 4; or.b32 %r1432, %r1431, %r1430; shr.u32 %r1433, %r1430, 6; mul.wide.u32 %rd1049, %r1433, 8; add.s64 %rd1050, %rd2311, %rd1049; and.b32 %r1434, %r1432, 63; mov.u64 %rd1051, 1; shl.b64 %rd1052, %rd1051, %r1434; ld.u64 %rd1053, [%rd1050+544]; and.b64 %rd1054, %rd1053, %rd1052; setp.eq.s64 %p135, %rd1054, 0; mul.wide.u32 %rd1055, %r1432, 8; add.s64 %rd1056, %rd2311, %rd1055; add.s64 %rd188, %rd1056, 1088; @%p135 bra $L__BB3_217; and.b32 %r3229, %r259, -8; ld.u64 %rd1057, [%rd188]; add.s64 %rd2307, %rd2311, %rd1057; and.b32 %r1435, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1435}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1436, %r259, 6; and.b32 %r1437, %r1436, 448; shl.b32 %r1438, %r8, 3; and.b32 %r1439, %r1438, 56; or.b32 %r1440, %r1437, %r1439; and.b32 %r1441, %r74, 7; or.b32 %r1442, %r1440, %r1441; mul.wide.u32 %rd1058, %r1442, 4; add.s64 %rd1059, %rd2307, %rd1058; ld.f32 %f465, [%rd1059+96]; mov.b64 %rd1060, {%r3229, %r1435}; shr.u64 %rd1061, %rd1060, 32; cvt.u32.u64 %r3184, %rd1061; bra.uni $L__BB3_241; $L__BB3_246: and.b32 %r423, %r259, -128; setp.ne.s32 %p153, %r423, %r3272; mov.u16 %rs284, 0; @%p153 bra $L__BB3_249; and.b32 %r1510, %r8, -128; setp.ne.s32 %p154, %r1510, %r3271; @%p154 bra $L__BB3_249; and.b32 %r1511, %r9, -128; setp.eq.s32 %p155, %r1511, %r3270; selp.u16 %rs284, 1, 0, %p155; $L__BB3_249: setp.eq.s16 %p156, %rs284, 0; @%p156 bra $L__BB3_253; shl.b32 %r1512, %r259, 5; and.b32 %r1513, %r1512, 3840; shl.b32 %r1514, %r8, 1; and.b32 %r1515, %r1514, 240; or.b32 %r1516, %r1513, %r1515; bfe.u32 %r1517, %r9, 3, 4; or.b32 %r1518, %r1516, %r1517; shr.u32 %r1519, %r1516, 6; mul.wide.u32 %rd1120, %r1519, 8; add.s64 %rd1121, %rd2311, %rd1120; and.b32 %r1520, %r1518, 63; mov.u64 %rd1122, 1; shl.b64 %rd1123, %rd1122, %r1520; ld.u64 %rd1124, [%rd1121+544]; and.b64 %rd1125, %rd1124, %rd1123; setp.eq.s64 %p157, %rd1125, 0; mul.wide.u32 %rd1126, %r1518, 8; add.s64 %rd1127, %rd2311, %rd1126; add.s64 %rd217, %rd1127, 1088; @%p157 bra $L__BB3_252; and.b32 %r3229, %r259, -8; ld.u64 %rd1128, [%rd217]; add.s64 %rd2307, %rd2311, %rd1128; and.b32 %r1521, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1521}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1522, %r259, 6; and.b32 %r1523, %r1522, 448; shl.b32 %r1524, %r8, 3; and.b32 %r1525, %r1524, 56; or.b32 %r1526, %r1523, %r1525; and.b32 %r1527, %r9, 7; or.b32 %r1528, %r1526, %r1527; mul.wide.u32 %rd1129, %r1528, 4; add.s64 %rd1130, %rd2307, %rd1129; ld.f32 %f466, [%rd1130+96]; mov.b64 %rd1131, {%r3229, %r1521}; shr.u64 %rd1132, %rd1131, 32; cvt.u32.u64 %r3184, %rd1132; bra.uni $L__BB3_276; $L__BB3_281: and.b32 %r473, %r471, -128; setp.ne.s32 %p175, %r473, %r3272; mov.u16 %rs287, 0; @%p175 bra $L__BB3_284; and.b32 %r1595, %r8, -128; setp.ne.s32 %p176, %r1595, %r3271; @%p176 bra $L__BB3_284; and.b32 %r1596, %r9, -128; setp.eq.s32 %p177, %r1596, %r3270; selp.u16 %rs287, 1, 0, %p177; $L__BB3_284: setp.eq.s16 %p178, %rs287, 0; @%p178 bra $L__BB3_288; shl.b32 %r1597, %r471, 5; and.b32 %r1598, %r1597, 3840; shl.b32 %r1599, %r8, 1; and.b32 %r1600, %r1599, 240; or.b32 %r1601, %r1600, %r1598; bfe.u32 %r1602, %r9, 3, 4; or.b32 %r1603, %r1601, %r1602; shr.u32 %r1604, %r1601, 6; mul.wide.u32 %rd1190, %r1604, 8; add.s64 %rd1191, %rd2311, %rd1190; and.b32 %r1605, %r1603, 63; mov.u64 %rd1192, 1; shl.b64 %rd1193, %rd1192, %r1605; ld.u64 %rd1194, [%rd1191+544]; and.b64 %rd1195, %rd1194, %rd1193; setp.eq.s64 %p179, %rd1195, 0; mul.wide.u32 %rd1196, %r1603, 8; add.s64 %rd1197, %rd2311, %rd1196; add.s64 %rd246, %rd1197, 1088; @%p179 bra $L__BB3_287; ld.u64 %rd1198, [%rd246]; add.s64 %rd2307, %rd2311, %rd1198; and.b32 %r1606, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1606}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1607, %r471, 6; and.b32 %r1608, %r1607, 448; shl.b32 %r1609, %r8, 3; and.b32 %r1610, %r1609, 56; or.b32 %r1611, %r1610, %r1608; and.b32 %r1612, %r9, 7; or.b32 %r1613, %r1611, %r1612; mul.wide.u32 %rd1199, %r1613, 4; add.s64 %rd1200, %rd2307, %rd1199; ld.f32 %f467, [%rd1200+96]; mov.b64 %rd1201, {%r472, %r1606}; shr.u64 %rd1202, %rd1201, 32; cvt.u32.u64 %r3184, %rd1202; mov.u32 %r3229, %r472; bra.uni $L__BB3_311; $L__BB3_316: and.b32 %r518, %r471, -128; setp.ne.s32 %p197, %r518, %r3272; mov.u16 %rs290, 0; @%p197 bra $L__BB3_319; and.b32 %r1680, %r8, -128; setp.ne.s32 %p198, %r1680, %r3271; @%p198 bra $L__BB3_319; and.b32 %r1681, %r74, -128; setp.eq.s32 %p199, %r1681, %r3270; selp.u16 %rs290, 1, 0, %p199; $L__BB3_319: setp.eq.s16 %p200, %rs290, 0; @%p200 bra $L__BB3_323; shl.b32 %r1682, %r471, 5; and.b32 %r1683, %r1682, 3840; shl.b32 %r1684, %r8, 1; and.b32 %r1685, %r1684, 240; or.b32 %r1686, %r1685, %r1683; bfe.u32 %r1687, %r74, 3, 4; or.b32 %r1688, %r1687, %r1686; shr.u32 %r1689, %r1686, 6; mul.wide.u32 %rd1260, %r1689, 8; add.s64 %rd1261, %rd2311, %rd1260; and.b32 %r1690, %r1688, 63; mov.u64 %rd1262, 1; shl.b64 %rd1263, %rd1262, %r1690; ld.u64 %rd1264, [%rd1261+544]; and.b64 %rd1265, %rd1264, %rd1263; setp.eq.s64 %p201, %rd1265, 0; mul.wide.u32 %rd1266, %r1688, 8; add.s64 %rd1267, %rd2311, %rd1266; add.s64 %rd275, %rd1267, 1088; @%p201 bra $L__BB3_322; ld.u64 %rd1268, [%rd275]; add.s64 %rd2307, %rd2311, %rd1268; and.b32 %r1691, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1691}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1692, %r471, 6; and.b32 %r1693, %r1692, 448; shl.b32 %r1694, %r8, 3; and.b32 %r1695, %r1694, 56; or.b32 %r1696, %r1695, %r1693; and.b32 %r1697, %r74, 7; or.b32 %r1698, %r1696, %r1697; mul.wide.u32 %rd1269, %r1698, 4; add.s64 %rd1270, %rd2307, %rd1269; ld.f32 %f468, [%rd1270+96]; mov.b64 %rd1271, {%r472, %r1691}; shr.u64 %rd1272, %rd1271, 32; cvt.u32.u64 %r3184, %rd1272; mov.u32 %r3229, %r472; bra.uni $L__BB3_346; $L__BB3_351: and.b32 %r560, %r471, -128; setp.ne.s32 %p219, %r560, %r3272; mov.u16 %rs293, 0; @%p219 bra $L__BB3_354; and.b32 %r1766, %r139, -128; setp.ne.s32 %p220, %r1766, %r3271; @%p220 bra $L__BB3_354; and.b32 %r1767, %r74, -128; setp.eq.s32 %p221, %r1767, %r3270; selp.u16 %rs293, 1, 0, %p221; $L__BB3_354: setp.eq.s16 %p222, %rs293, 0; @%p222 bra $L__BB3_358; shl.b32 %r1768, %r471, 5; and.b32 %r1769, %r1768, 3840; shl.b32 %r1770, %r139, 1; and.b32 %r1771, %r1770, 240; or.b32 %r1772, %r1771, %r1769; bfe.u32 %r1773, %r74, 3, 4; or.b32 %r1774, %r1772, %r1773; shr.u32 %r1775, %r1772, 6; mul.wide.u32 %rd1331, %r1775, 8; add.s64 %rd1332, %rd2311, %rd1331; and.b32 %r1776, %r1774, 63; mov.u64 %rd1333, 1; shl.b64 %rd1334, %rd1333, %r1776; ld.u64 %rd1335, [%rd1332+544]; and.b64 %rd1336, %rd1335, %rd1334; setp.eq.s64 %p223, %rd1336, 0; mul.wide.u32 %rd1337, %r1774, 8; add.s64 %rd1338, %rd2311, %rd1337; add.s64 %rd304, %rd1338, 1088; @%p223 bra $L__BB3_357; ld.u64 %rd1339, [%rd304]; add.s64 %rd2307, %rd2311, %rd1339; and.b32 %r1777, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1777}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1778, %r471, 6; and.b32 %r1779, %r1778, 448; shl.b32 %r1780, %r139, 3; and.b32 %r1781, %r1780, 56; or.b32 %r1782, %r1781, %r1779; and.b32 %r1783, %r74, 7; or.b32 %r1784, %r1782, %r1783; mul.wide.u32 %rd1340, %r1784, 4; add.s64 %rd1341, %rd2307, %rd1340; ld.f32 %f469, [%rd1341+96]; mov.b64 %rd1342, {%r472, %r1777}; shr.u64 %rd1343, %rd1342, 32; cvt.u32.u64 %r3184, %rd1343; mov.u32 %r3229, %r472; bra.uni $L__BB3_381; $L__BB3_386: and.b32 %r599, %r471, -128; setp.ne.s32 %p241, %r599, %r3272; mov.u16 %rs296, 0; @%p241 bra $L__BB3_389; and.b32 %r1852, %r139, -128; setp.ne.s32 %p242, %r1852, %r3271; @%p242 bra $L__BB3_389; and.b32 %r1853, %r9, -128; setp.eq.s32 %p243, %r1853, %r3270; selp.u16 %rs296, 1, 0, %p243; $L__BB3_389: setp.eq.s16 %p244, %rs296, 0; @%p244 bra $L__BB3_393; shl.b32 %r1854, %r471, 5; and.b32 %r1855, %r1854, 3840; shl.b32 %r1856, %r139, 1; and.b32 %r1857, %r1856, 240; or.b32 %r1858, %r1857, %r1855; bfe.u32 %r1859, %r9, 3, 4; or.b32 %r1860, %r1858, %r1859; shr.u32 %r1861, %r1858, 6; mul.wide.u32 %rd1402, %r1861, 8; add.s64 %rd1403, %rd2311, %rd1402; and.b32 %r1862, %r1860, 63; mov.u64 %rd1404, 1; shl.b64 %rd1405, %rd1404, %r1862; ld.u64 %rd1406, [%rd1403+544]; and.b64 %rd1407, %rd1406, %rd1405; setp.eq.s64 %p245, %rd1407, 0; mul.wide.u32 %rd1408, %r1860, 8; add.s64 %rd1409, %rd2311, %rd1408; add.s64 %rd333, %rd1409, 1088; @%p245 bra $L__BB3_392; ld.u64 %rd1410, [%rd333]; add.s64 %rd2307, %rd2311, %rd1410; and.b32 %r1863, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1863}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1864, %r471, 7; shl.b32 %r1865, %r139, 3; and.b32 %r1866, %r1865, 56; and.b32 %r1867, %r9, 7; bfi.b32 %r1868, %r1864, %r1867, 6, 3; or.b32 %r1869, %r1868, %r1866; mul.wide.u32 %rd1411, %r1869, 4; add.s64 %rd1412, %rd2307, %rd1411; ld.f32 %f470, [%rd1412+96]; mov.b64 %rd1413, {%r472, %r1863}; shr.u64 %rd1414, %rd1413, 32; cvt.u32.u64 %r3184, %rd1414; mov.u32 %r3229, %r472; bra.uni $L__BB3_416; $L__BB3_421: and.b32 %r637, %r635, -128; setp.ne.s32 %p263, %r637, %r3272; mov.u16 %rs299, 0; @%p263 bra $L__BB3_424; and.b32 %r1934, %r139, -128; setp.ne.s32 %p264, %r1934, %r3271; @%p264 bra $L__BB3_424; and.b32 %r1935, %r9, -128; setp.eq.s32 %p265, %r1935, %r3270; selp.u16 %rs299, 1, 0, %p265; $L__BB3_424: setp.eq.s16 %p266, %rs299, 0; @%p266 bra $L__BB3_428; shl.b32 %r1936, %r635, 5; and.b32 %r1937, %r1936, 3840; shl.b32 %r1938, %r139, 1; and.b32 %r1939, %r1938, 240; or.b32 %r1940, %r1939, %r1937; bfe.u32 %r1941, %r9, 3, 4; or.b32 %r1942, %r1940, %r1941; shr.u32 %r1943, %r1940, 6; mul.wide.u32 %rd1473, %r1943, 8; add.s64 %rd1474, %rd2311, %rd1473; and.b32 %r1944, %r1942, 63; mov.u64 %rd1475, 1; shl.b64 %rd1476, %rd1475, %r1944; ld.u64 %rd1477, [%rd1474+544]; and.b64 %rd1478, %rd1477, %rd1476; setp.eq.s64 %p267, %rd1478, 0; mul.wide.u32 %rd1479, %r1942, 8; add.s64 %rd1480, %rd2311, %rd1479; add.s64 %rd362, %rd1480, 1088; @%p267 bra $L__BB3_427; ld.u64 %rd1481, [%rd362]; add.s64 %rd2307, %rd2311, %rd1481; and.b32 %r1945, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1945}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1946, %r635, 7; shl.b32 %r1947, %r139, 3; and.b32 %r1948, %r1947, 56; and.b32 %r1949, %r9, 7; bfi.b32 %r1950, %r1946, %r1949, 6, 3; or.b32 %r1951, %r1950, %r1948; mul.wide.u32 %rd1482, %r1951, 4; add.s64 %rd1483, %rd2307, %rd1482; ld.f32 %f471, [%rd1483+96]; mov.b64 %rd1484, {%r636, %r1945}; shr.u64 %rd1485, %rd1484, 32; cvt.u32.u64 %r3184, %rd1485; mov.u32 %r3229, %r636; bra.uni $L__BB3_451; $L__BB3_456: and.b32 %r670, %r635, -128; setp.ne.s32 %p285, %r670, %r3272; mov.u16 %rs302, 0; @%p285 bra $L__BB3_459; and.b32 %r2017, %r139, -128; setp.ne.s32 %p286, %r2017, %r3271; @%p286 bra $L__BB3_459; and.b32 %r2018, %r74, -128; setp.eq.s32 %p287, %r2018, %r3270; selp.u16 %rs302, 1, 0, %p287; $L__BB3_459: setp.eq.s16 %p288, %rs302, 0; @%p288 bra $L__BB3_463; shl.b32 %r2019, %r635, 5; and.b32 %r2020, %r2019, 3840; shl.b32 %r2021, %r139, 1; and.b32 %r2022, %r2021, 240; or.b32 %r2023, %r2022, %r2020; bfe.u32 %r2024, %r74, 3, 4; or.b32 %r2025, %r2023, %r2024; shr.u32 %r2026, %r2023, 6; mul.wide.u32 %rd1544, %r2026, 8; add.s64 %rd1545, %rd2311, %rd1544; and.b32 %r2027, %r2025, 63; mov.u64 %rd1546, 1; shl.b64 %rd1547, %rd1546, %r2027; ld.u64 %rd1548, [%rd1545+544]; and.b64 %rd1549, %rd1548, %rd1547; setp.eq.s64 %p289, %rd1549, 0; mul.wide.u32 %rd1550, %r2025, 8; add.s64 %rd1551, %rd2311, %rd1550; add.s64 %rd391, %rd1551, 1088; @%p289 bra $L__BB3_462; ld.u64 %rd1552, [%rd391]; add.s64 %rd2307, %rd2311, %rd1552; and.b32 %r2028, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2028}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2029, %r635, 6; and.b32 %r2030, %r2029, 448; shl.b32 %r2031, %r139, 3; and.b32 %r2032, %r2031, 56; or.b32 %r2033, %r2032, %r2030; and.b32 %r2034, %r74, 7; or.b32 %r2035, %r2033, %r2034; mul.wide.u32 %rd1553, %r2035, 4; add.s64 %rd1554, %rd2307, %rd1553; ld.f32 %f472, [%rd1554+96]; mov.b64 %rd1555, {%r636, %r2028}; shr.u64 %rd1556, %rd1555, 32; cvt.u32.u64 %r3184, %rd1556; mov.u32 %r3229, %r636; bra.uni $L__BB3_486; $L__BB3_491: and.b32 %r700, %r635, -128; setp.ne.s32 %p307, %r700, %r3272; mov.u16 %rs305, 0; @%p307 bra $L__BB3_494; and.b32 %r2104, %r8, -128; setp.ne.s32 %p308, %r2104, %r3271; @%p308 bra $L__BB3_494; and.b32 %r2105, %r74, -128; setp.eq.s32 %p309, %r2105, %r3270; selp.u16 %rs305, 1, 0, %p309; $L__BB3_494: setp.eq.s16 %p310, %rs305, 0; @%p310 bra $L__BB3_498; shl.b32 %r2106, %r635, 5; and.b32 %r2107, %r2106, 3840; shl.b32 %r2108, %r8, 1; and.b32 %r2109, %r2108, 240; or.b32 %r2110, %r2107, %r2109; bfe.u32 %r2111, %r74, 3, 4; or.b32 %r2112, %r2111, %r2110; shr.u32 %r2113, %r2110, 6; mul.wide.u32 %rd1615, %r2113, 8; add.s64 %rd1616, %rd2311, %rd1615; and.b32 %r2114, %r2112, 63; mov.u64 %rd1617, 1; shl.b64 %rd1618, %rd1617, %r2114; ld.u64 %rd1619, [%rd1616+544]; and.b64 %rd1620, %rd1619, %rd1618; setp.eq.s64 %p311, %rd1620, 0; mul.wide.u32 %rd1621, %r2112, 8; add.s64 %rd1622, %rd2311, %rd1621; add.s64 %rd420, %rd1622, 1088; @%p311 bra $L__BB3_497; ld.u64 %rd1623, [%rd420]; add.s64 %rd2307, %rd2311, %rd1623; and.b32 %r2115, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2115}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2116, %r635, 6; and.b32 %r2117, %r2116, 448; shl.b32 %r2118, %r8, 3; and.b32 %r2119, %r2118, 56; or.b32 %r2120, %r2117, %r2119; and.b32 %r2121, %r74, 7; or.b32 %r2122, %r2120, %r2121; mul.wide.u32 %rd1624, %r2122, 4; add.s64 %rd1625, %rd2307, %rd1624; ld.f32 %f473, [%rd1625+96]; mov.b64 %rd1626, {%r636, %r2115}; shr.u64 %rd1627, %rd1626, 32; cvt.u32.u64 %r3184, %rd1627; mov.u32 %r3229, %r636; bra.uni $L__BB3_521; $L__BB3_526: and.b32 %r727, %r635, -128; setp.ne.s32 %p329, %r727, %r3272; mov.u16 %rs308, 0; @%p329 bra $L__BB3_529; and.b32 %r2190, %r8, -128; setp.ne.s32 %p330, %r2190, %r3271; @%p330 bra $L__BB3_529; and.b32 %r2191, %r9, -128; setp.eq.s32 %p331, %r2191, %r3270; selp.u16 %rs308, 1, 0, %p331; $L__BB3_529: setp.eq.s16 %p332, %rs308, 0; @%p332 bra $L__BB3_533; shl.b32 %r2192, %r635, 5; and.b32 %r2193, %r2192, 3840; shl.b32 %r2194, %r8, 1; and.b32 %r2195, %r2194, 240; or.b32 %r2196, %r2193, %r2195; bfe.u32 %r2197, %r9, 3, 4; or.b32 %r2198, %r2196, %r2197; shr.u32 %r2199, %r2196, 6; mul.wide.u32 %rd1686, %r2199, 8; add.s64 %rd1687, %rd2311, %rd1686; and.b32 %r2200, %r2198, 63; mov.u64 %rd1688, 1; shl.b64 %rd1689, %rd1688, %r2200; ld.u64 %rd1690, [%rd1687+544]; and.b64 %rd1691, %rd1690, %rd1689; setp.eq.s64 %p333, %rd1691, 0; mul.wide.u32 %rd1692, %r2198, 8; add.s64 %rd1693, %rd2311, %rd1692; add.s64 %rd449, %rd1693, 1088; @%p333 bra $L__BB3_532; ld.u64 %rd1694, [%rd449]; add.s64 %rd1695, %rd2311, %rd1694; and.b32 %r2201, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2201}; and.b32 %r2202, %r9, -8; st.local.u32 [%rd2+8], %r2202; st.local.u64 [%rd2+48], %rd1695; shl.b32 %r2203, %r635, 6; and.b32 %r2204, %r2203, 448; shl.b32 %r2205, %r8, 3; and.b32 %r2206, %r2205, 56; or.b32 %r2207, %r2204, %r2206; and.b32 %r2208, %r9, 7; or.b32 %r2209, %r2207, %r2208; mul.wide.u32 %rd1696, %r2209, 4; add.s64 %rd1697, %rd1695, %rd1696; ld.f32 %f474, [%rd1697+96]; bra.uni $L__BB3_556; $L__BB3_561: and.b32 %r735, %r471, -128; ld.local.u32 %r2285, [%rd2+12]; setp.ne.s32 %p351, %r735, %r2285; mov.u16 %rs311, 0; @%p351 bra $L__BB3_564; and.b32 %r2286, %r732, -128; ld.local.u32 %r2287, [%rd2+16]; setp.ne.s32 %p352, %r2286, %r2287; @%p352 bra $L__BB3_564; and.b32 %r2288, %r733, -128; ld.local.u32 %r2289, [%rd2+20]; setp.eq.s32 %p353, %r2288, %r2289; selp.u16 %rs311, 1, 0, %p353; $L__BB3_564: setp.eq.s16 %p354, %rs311, 0; @%p354 bra $L__BB3_568; shl.b32 %r2290, %r471, 5; and.b32 %r2291, %r2290, 3840; shl.b32 %r2292, %r732, 1; and.b32 %r2293, %r2292, 240; or.b32 %r2294, %r2293, %r2291; bfe.u32 %r2295, %r733, 3, 4; or.b32 %r2296, %r2294, %r2295; shr.u32 %r2297, %r2294, 6; ld.local.u64 %rd464, [%rd2+56]; mul.wide.u32 %rd1752, %r2297, 8; add.s64 %rd1753, %rd464, %rd1752; and.b32 %r2298, %r2296, 63; mov.u64 %rd1754, 1; shl.b64 %rd1755, %rd1754, %r2298; ld.u64 %rd1756, [%rd1753+544]; and.b64 %rd1757, %rd1756, %rd1755; setp.eq.s64 %p355, %rd1757, 0; mul.wide.u32 %rd1758, %r2296, 8; add.s64 %rd1759, %rd464, %rd1758; add.s64 %rd465, %rd1759, 1088; @%p355 bra $L__BB3_567; ld.u64 %rd1760, [%rd465]; add.s64 %rd1761, %rd464, %rd1760; and.b32 %r2299, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2299}; and.b32 %r2300, %r733, -8; st.local.u32 [%rd2+8], %r2300; st.local.u64 [%rd2+48], %rd1761; shl.b32 %r2301, %r471, 6; and.b32 %r2302, %r2301, 448; shl.b32 %r2303, %r732, 3; and.b32 %r2304, %r2303, 56; or.b32 %r2305, %r2304, %r2302; and.b32 %r2306, %r733, 7; or.b32 %r2307, %r2305, %r2306; mul.wide.u32 %rd1762, %r2307, 4; add.s64 %rd1763, %rd1761, %rd1762; ld.f32 %f475, [%rd1763+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_591; $L__BB3_596: and.b32 %r742, %r471, -128; ld.local.u32 %r2388, [%rd2+12]; setp.ne.s32 %p373, %r742, %r2388; mov.u16 %rs314, 0; @%p373 bra $L__BB3_599; and.b32 %r2389, %r732, -128; ld.local.u32 %r2390, [%rd2+16]; setp.ne.s32 %p374, %r2389, %r2390; @%p374 bra $L__BB3_599; and.b32 %r2391, %r741, -128; ld.local.u32 %r2392, [%rd2+20]; setp.eq.s32 %p375, %r2391, %r2392; selp.u16 %rs314, 1, 0, %p375; $L__BB3_599: setp.eq.s16 %p376, %rs314, 0; @%p376 bra $L__BB3_603; shl.b32 %r2393, %r471, 5; and.b32 %r2394, %r2393, 3840; shl.b32 %r2395, %r732, 1; and.b32 %r2396, %r2395, 240; or.b32 %r2397, %r2396, %r2394; bfe.u32 %r2398, %r741, 3, 4; or.b32 %r2399, %r2398, %r2397; shr.u32 %r2400, %r2397, 6; ld.local.u64 %rd482, [%rd2+56]; mul.wide.u32 %rd1819, %r2400, 8; add.s64 %rd1820, %rd482, %rd1819; and.b32 %r2401, %r2399, 63; mov.u64 %rd1821, 1; shl.b64 %rd1822, %rd1821, %r2401; ld.u64 %rd1823, [%rd1820+544]; and.b64 %rd1824, %rd1823, %rd1822; setp.eq.s64 %p377, %rd1824, 0; mul.wide.u32 %rd1825, %r2399, 8; add.s64 %rd1826, %rd482, %rd1825; add.s64 %rd483, %rd1826, 1088; @%p377 bra $L__BB3_602; ld.u64 %rd1827, [%rd483]; add.s64 %rd1828, %rd482, %rd1827; and.b32 %r2402, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2402}; and.b32 %r2403, %r741, -8; st.local.u32 [%rd2+8], %r2403; st.local.u64 [%rd2+48], %rd1828; shl.b32 %r2404, %r471, 6; and.b32 %r2405, %r2404, 448; shl.b32 %r2406, %r732, 3; and.b32 %r2407, %r2406, 56; or.b32 %r2408, %r2407, %r2405; and.b32 %r2409, %r741, 7; or.b32 %r2410, %r2408, %r2409; mul.wide.u32 %rd1829, %r2410, 4; add.s64 %rd1830, %rd1828, %rd1829; ld.f32 %f476, [%rd1830+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_626; $L__BB3_631: and.b32 %r749, %r471, -128; ld.local.u32 %r2491, [%rd2+12]; setp.ne.s32 %p395, %r749, %r2491; mov.u16 %rs317, 0; @%p395 bra $L__BB3_634; and.b32 %r2492, %r748, -128; ld.local.u32 %r2493, [%rd2+16]; setp.ne.s32 %p396, %r2492, %r2493; @%p396 bra $L__BB3_634; and.b32 %r2494, %r741, -128; ld.local.u32 %r2495, [%rd2+20]; setp.eq.s32 %p397, %r2494, %r2495; selp.u16 %rs317, 1, 0, %p397; $L__BB3_634: setp.eq.s16 %p398, %rs317, 0; @%p398 bra $L__BB3_638; shl.b32 %r2496, %r471, 5; and.b32 %r2497, %r2496, 3840; shl.b32 %r2498, %r748, 1; and.b32 %r2499, %r2498, 240; or.b32 %r2500, %r2499, %r2497; bfe.u32 %r2501, %r741, 3, 4; or.b32 %r2502, %r2500, %r2501; shr.u32 %r2503, %r2500, 6; ld.local.u64 %rd500, [%rd2+56]; mul.wide.u32 %rd1887, %r2503, 8; add.s64 %rd1888, %rd500, %rd1887; and.b32 %r2504, %r2502, 63; mov.u64 %rd1889, 1; shl.b64 %rd1890, %rd1889, %r2504; ld.u64 %rd1891, [%rd1888+544]; and.b64 %rd1892, %rd1891, %rd1890; setp.eq.s64 %p399, %rd1892, 0; mul.wide.u32 %rd1893, %r2502, 8; add.s64 %rd1894, %rd500, %rd1893; add.s64 %rd501, %rd1894, 1088; @%p399 bra $L__BB3_637; ld.u64 %rd1895, [%rd501]; add.s64 %rd1896, %rd500, %rd1895; and.b32 %r2505, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2505}; and.b32 %r2506, %r741, -8; st.local.u32 [%rd2+8], %r2506; st.local.u64 [%rd2+48], %rd1896; shl.b32 %r2507, %r471, 6; and.b32 %r2508, %r2507, 448; shl.b32 %r2509, %r748, 3; and.b32 %r2510, %r2509, 56; or.b32 %r2511, %r2510, %r2508; and.b32 %r2512, %r741, 7; or.b32 %r2513, %r2511, %r2512; mul.wide.u32 %rd1897, %r2513, 4; add.s64 %rd1898, %rd1896, %rd1897; ld.f32 %f477, [%rd1898+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_661; $L__BB3_666: and.b32 %r755, %r471, -128; ld.local.u32 %r2593, [%rd2+12]; setp.ne.s32 %p417, %r755, %r2593; mov.u16 %rs320, 0; @%p417 bra $L__BB3_669; and.b32 %r2594, %r748, -128; ld.local.u32 %r2595, [%rd2+16]; setp.ne.s32 %p418, %r2594, %r2595; @%p418 bra $L__BB3_669; and.b32 %r2596, %r733, -128; ld.local.u32 %r2597, [%rd2+20]; setp.eq.s32 %p419, %r2596, %r2597; selp.u16 %rs320, 1, 0, %p419; $L__BB3_669: setp.eq.s16 %p420, %rs320, 0; @%p420 bra $L__BB3_673; shl.b32 %r2598, %r471, 5; and.b32 %r2599, %r2598, 3840; shl.b32 %r2600, %r748, 1; and.b32 %r2601, %r2600, 240; or.b32 %r2602, %r2601, %r2599; bfe.u32 %r2603, %r733, 3, 4; or.b32 %r2604, %r2602, %r2603; shr.u32 %r2605, %r2602, 6; ld.local.u64 %rd518, [%rd2+56]; mul.wide.u32 %rd1954, %r2605, 8; add.s64 %rd1955, %rd518, %rd1954; and.b32 %r2606, %r2604, 63; mov.u64 %rd1956, 1; shl.b64 %rd1957, %rd1956, %r2606; ld.u64 %rd1958, [%rd1955+544]; and.b64 %rd1959, %rd1958, %rd1957; setp.eq.s64 %p421, %rd1959, 0; mul.wide.u32 %rd1960, %r2604, 8; add.s64 %rd1961, %rd518, %rd1960; add.s64 %rd519, %rd1961, 1088; @%p421 bra $L__BB3_672; ld.u64 %rd1962, [%rd519]; add.s64 %rd1963, %rd518, %rd1962; and.b32 %r2607, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2607}; and.b32 %r2608, %r733, -8; st.local.u32 [%rd2+8], %r2608; st.local.u64 [%rd2+48], %rd1963; and.b32 %r2609, %r471, 7; shl.b32 %r2610, %r748, 3; and.b32 %r2611, %r2610, 56; and.b32 %r2612, %r733, 7; bfi.b32 %r2613, %r2609, %r2612, 6, 3; or.b32 %r2614, %r2613, %r2611; mul.wide.u32 %rd1964, %r2614, 4; add.s64 %rd1965, %rd1963, %rd1964; ld.f32 %f478, [%rd1965+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_696; $L__BB3_701: and.b32 %r761, %r635, -128; ld.local.u32 %r2692, [%rd2+12]; setp.ne.s32 %p439, %r761, %r2692; mov.u16 %rs323, 0; @%p439 bra $L__BB3_704; and.b32 %r2693, %r748, -128; ld.local.u32 %r2694, [%rd2+16]; setp.ne.s32 %p440, %r2693, %r2694; @%p440 bra $L__BB3_704; and.b32 %r2695, %r733, -128; ld.local.u32 %r2696, [%rd2+20]; setp.eq.s32 %p441, %r2695, %r2696; selp.u16 %rs323, 1, 0, %p441; $L__BB3_704: setp.eq.s16 %p442, %rs323, 0; @%p442 bra $L__BB3_708; shl.b32 %r2697, %r635, 5; and.b32 %r2698, %r2697, 3840; shl.b32 %r2699, %r748, 1; and.b32 %r2700, %r2699, 240; or.b32 %r2701, %r2700, %r2698; bfe.u32 %r2702, %r733, 3, 4; or.b32 %r2703, %r2701, %r2702; shr.u32 %r2704, %r2701, 6; ld.local.u64 %rd536, [%rd2+56]; mul.wide.u32 %rd2021, %r2704, 8; add.s64 %rd2022, %rd536, %rd2021; and.b32 %r2705, %r2703, 63; mov.u64 %rd2023, 1; shl.b64 %rd2024, %rd2023, %r2705; ld.u64 %rd2025, [%rd2022+544]; and.b64 %rd2026, %rd2025, %rd2024; setp.eq.s64 %p443, %rd2026, 0; mul.wide.u32 %rd2027, %r2703, 8; add.s64 %rd2028, %rd536, %rd2027; add.s64 %rd537, %rd2028, 1088; @%p443 bra $L__BB3_707; ld.u64 %rd2029, [%rd537]; add.s64 %rd2030, %rd536, %rd2029; and.b32 %r2706, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2706}; and.b32 %r2707, %r733, -8; st.local.u32 [%rd2+8], %r2707; st.local.u64 [%rd2+48], %rd2030; and.b32 %r2708, %r635, 7; shl.b32 %r2709, %r748, 3; and.b32 %r2710, %r2709, 56; and.b32 %r2711, %r733, 7; bfi.b32 %r2712, %r2708, %r2711, 6, 3; or.b32 %r2713, %r2712, %r2710; mul.wide.u32 %rd2031, %r2713, 4; add.s64 %rd2032, %rd2030, %rd2031; ld.f32 %f479, [%rd2032+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_731; $L__BB3_736: and.b32 %r767, %r635, -128; ld.local.u32 %r2792, [%rd2+12]; setp.ne.s32 %p461, %r767, %r2792; mov.u16 %rs326, 0; @%p461 bra $L__BB3_739; and.b32 %r2793, %r748, -128; ld.local.u32 %r2794, [%rd2+16]; setp.ne.s32 %p462, %r2793, %r2794; @%p462 bra $L__BB3_739; and.b32 %r2795, %r741, -128; ld.local.u32 %r2796, [%rd2+20]; setp.eq.s32 %p463, %r2795, %r2796; selp.u16 %rs326, 1, 0, %p463; $L__BB3_739: setp.eq.s16 %p464, %rs326, 0; @%p464 bra $L__BB3_743; shl.b32 %r2797, %r635, 5; and.b32 %r2798, %r2797, 3840; shl.b32 %r2799, %r748, 1; and.b32 %r2800, %r2799, 240; or.b32 %r2801, %r2800, %r2798; bfe.u32 %r2802, %r741, 3, 4; or.b32 %r2803, %r2801, %r2802; shr.u32 %r2804, %r2801, 6; ld.local.u64 %rd554, [%rd2+56]; mul.wide.u32 %rd2088, %r2804, 8; add.s64 %rd2089, %rd554, %rd2088; and.b32 %r2805, %r2803, 63; mov.u64 %rd2090, 1; shl.b64 %rd2091, %rd2090, %r2805; ld.u64 %rd2092, [%rd2089+544]; and.b64 %rd2093, %rd2092, %rd2091; setp.eq.s64 %p465, %rd2093, 0; mul.wide.u32 %rd2094, %r2803, 8; add.s64 %rd2095, %rd554, %rd2094; add.s64 %rd555, %rd2095, 1088; @%p465 bra $L__BB3_742; ld.u64 %rd2096, [%rd555]; add.s64 %rd2097, %rd554, %rd2096; and.b32 %r2806, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2806}; and.b32 %r2807, %r741, -8; st.local.u32 [%rd2+8], %r2807; st.local.u64 [%rd2+48], %rd2097; shl.b32 %r2808, %r635, 6; and.b32 %r2809, %r2808, 448; shl.b32 %r2810, %r748, 3; and.b32 %r2811, %r2810, 56; or.b32 %r2812, %r2811, %r2809; and.b32 %r2813, %r741, 7; or.b32 %r2814, %r2812, %r2813; mul.wide.u32 %rd2098, %r2814, 4; add.s64 %rd2099, %rd2097, %rd2098; ld.f32 %f480, [%rd2099+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_766; $L__BB3_771: and.b32 %r773, %r635, -128; ld.local.u32 %r2895, [%rd2+12]; setp.ne.s32 %p483, %r773, %r2895; mov.u16 %rs329, 0; @%p483 bra $L__BB3_774; and.b32 %r2896, %r732, -128; ld.local.u32 %r2897, [%rd2+16]; setp.ne.s32 %p484, %r2896, %r2897; @%p484 bra $L__BB3_774; and.b32 %r2898, %r741, -128; ld.local.u32 %r2899, [%rd2+20]; setp.eq.s32 %p485, %r2898, %r2899; selp.u16 %rs329, 1, 0, %p485; $L__BB3_774: setp.eq.s16 %p486, %rs329, 0; @%p486 bra $L__BB3_778; shl.b32 %r2900, %r635, 5; and.b32 %r2901, %r2900, 3840; shl.b32 %r2902, %r732, 1; and.b32 %r2903, %r2902, 240; or.b32 %r2904, %r2901, %r2903; bfe.u32 %r2905, %r741, 3, 4; or.b32 %r2906, %r2905, %r2904; shr.u32 %r2907, %r2904, 6; ld.local.u64 %rd572, [%rd2+56]; mul.wide.u32 %rd2155, %r2907, 8; add.s64 %rd2156, %rd572, %rd2155; and.b32 %r2908, %r2906, 63; mov.u64 %rd2157, 1; shl.b64 %rd2158, %rd2157, %r2908; ld.u64 %rd2159, [%rd2156+544]; and.b64 %rd2160, %rd2159, %rd2158; setp.eq.s64 %p487, %rd2160, 0; mul.wide.u32 %rd2161, %r2906, 8; add.s64 %rd2162, %rd572, %rd2161; add.s64 %rd573, %rd2162, 1088; @%p487 bra $L__BB3_777; ld.u64 %rd2163, [%rd573]; add.s64 %rd2164, %rd572, %rd2163; and.b32 %r2909, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2909}; and.b32 %r2910, %r741, -8; st.local.u32 [%rd2+8], %r2910; st.local.u64 [%rd2+48], %rd2164; shl.b32 %r2911, %r635, 6; and.b32 %r2912, %r2911, 448; shl.b32 %r2913, %r732, 3; and.b32 %r2914, %r2913, 56; or.b32 %r2915, %r2912, %r2914; and.b32 %r2916, %r741, 7; or.b32 %r2917, %r2915, %r2916; mul.wide.u32 %rd2165, %r2917, 4; add.s64 %rd2166, %rd2164, %rd2165; ld.f32 %f481, [%rd2166+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_801; $L__BB3_814: shl.b64 %rd2230, %rd593, 5; add.s64 %rd2394, %rd591, %rd2230; $L__BB3_815: setp.eq.s64 %p513, %rd2394, 0; @%p513 bra $L__BB3_823; add.s64 %rd596, %rd2394, 8; ld.u64 %rd597, [%rd2394+8]; setp.eq.s64 %p514, %rd597, 0; @%p514 bra $L__BB3_822; add.s64 %rd598, %rd590, %rd597; and.b32 %r3005, %r732, -4096; mov.u64 %rd2231, 1; st.local.v2.u32 [%rd2+24], {%r780, %r3005}; and.b32 %r3006, %r733, -4096; st.local.u32 [%rd2+32], %r3006; st.local.u64 [%rd2+64], %rd598; shl.b32 %r3007, %r635, 3; and.b32 %r3008, %r3007, 31744; and.b32 %r3009, %r732, 3968; shr.u32 %r3010, %r3009, 2; or.b32 %r3011, %r3008, %r3010; bfe.u32 %r3012, %r733, 7, 5; or.b32 %r3013, %r3011, %r3012; shr.u32 %r3014, %r3011, 6; mul.wide.u32 %rd2232, %r3014, 8; add.s64 %rd2233, %rd598, %rd2232; and.b32 %r3015, %r3013, 63; shl.b64 %rd2234, %rd2231, %r3015; ld.u64 %rd2235, [%rd2233+4128]; and.b64 %rd2236, %rd2235, %rd2234; setp.eq.s64 %p515, %rd2236, 0; mul.wide.u32 %rd2237, %r3013, 8; add.s64 %rd2238, %rd598, %rd2237; add.s64 %rd599, %rd2238, 8256; @%p515 bra $L__BB3_821; ld.u64 %rd2239, [%rd599]; add.s64 %rd600, %rd598, %rd2239; st.local.u32 [%rd2+12], %r779; and.b32 %r3016, %r732, -128; st.local.u32 [%rd2+16], %r3016; and.b32 %r3017, %r733, -128; st.local.u32 [%rd2+20], %r3017; st.local.u64 [%rd2+56], %rd600; shl.b32 %r3018, %r635, 5; and.b32 %r3019, %r3018, 3840; shl.b32 %r3020, %r732, 1; and.b32 %r3021, %r3020, 240; or.b32 %r3022, %r3019, %r3021; bfe.u32 %r3023, %r733, 3, 4; or.b32 %r3024, %r3022, %r3023; shr.u32 %r3025, %r3022, 6; mul.wide.u32 %rd2240, %r3025, 8; add.s64 %rd2241, %rd600, %rd2240; and.b32 %r3026, %r3024, 63; shl.b64 %rd2243, %rd2231, %r3026; ld.u64 %rd2244, [%rd2241+544]; and.b64 %rd2245, %rd2244, %rd2243; setp.eq.s64 %p516, %rd2245, 0; mul.wide.u32 %rd2246, %r3024, 8; add.s64 %rd2247, %rd600, %rd2246; add.s64 %rd601, %rd2247, 1088; @%p516 bra $L__BB3_820; ld.u64 %rd2248, [%rd601]; add.s64 %rd2249, %rd600, %rd2248; and.b32 %r3027, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3027}; and.b32 %r3028, %r733, -8; st.local.u32 [%rd2+8], %r3028; st.local.u64 [%rd2+48], %rd2249; shl.b32 %r3029, %r635, 6; and.b32 %r3030, %r3029, 448; shl.b32 %r3031, %r732, 3; and.b32 %r3032, %r3031, 56; or.b32 %r3033, %r3030, %r3032; and.b32 %r3034, %r733, 7; or.b32 %r3035, %r3033, %r3034; mul.wide.u32 %rd2250, %r3035, 4; add.s64 %rd2251, %rd2249, %rd2250; ld.f32 %f482, [%rd2251+96]; bra.uni $L__BB3_833; $L__BB3_823: ld.f32 %f482, [%rd590+28]; bra.uni $L__BB3_833; $L__BB3_822: ld.f32 %f482, [%rd596+12]; bra.uni $L__BB3_833; $L__BB3_43: setp.ne.s32 %p26, %r12, %r3232; mov.u16 %rs267, 0; @%p26 bra $L__BB3_46; and.b32 %r1022, %r8, -4096; setp.ne.s32 %p27, %r1022, %r3277; @%p27 bra $L__BB3_46; and.b32 %r1023, %r74, -4096; setp.eq.s32 %p28, %r1023, %r3310; selp.u16 %rs267, 1, 0, %p28; $L__BB3_46: setp.eq.s16 %p29, %rs267, 0; @%p29 bra $L__BB3_52; shl.b32 %r1024, %r7, 3; and.b32 %r1025, %r1024, 31744; and.b32 %r1026, %r8, 3968; shr.u32 %r1027, %r1026, 2; or.b32 %r1028, %r1027, %r1025; bfe.u32 %r1029, %r74, 7, 5; or.b32 %r1030, %r1029, %r1028; shr.u32 %r1031, %r1028, 6; mul.wide.u32 %rd713, %r1031, 8; add.s64 %rd714, %rd2315, %rd713; and.b32 %r1032, %r1030, 63; mov.u64 %rd715, 1; shl.b64 %rd716, %rd715, %r1032; ld.u64 %rd717, [%rd714+4128]; and.b64 %rd718, %rd717, %rd716; setp.eq.s64 %p30, %rd718, 0; mul.wide.u32 %rd719, %r1030, 8; add.s64 %rd720, %rd2315, %rd719; add.s64 %rd45, %rd720, 8256; @%p30 bra $L__BB3_51; and.b32 %r3152, %r7, -128; ld.u64 %rd721, [%rd45]; add.s64 %rd2311, %rd2315, %rd721; st.local.u32 [%rd2+12], %r3152; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1033, %r7, 5; and.b32 %r1034, %r1033, 3840; shl.b32 %r1035, %r8, 1; and.b32 %r1036, %r1035, 240; or.b32 %r1037, %r1036, %r1034; bfe.u32 %r1038, %r74, 3, 4; or.b32 %r1039, %r1038, %r1037; shr.u32 %r1040, %r1037, 6; mul.wide.u32 %rd722, %r1040, 8; add.s64 %rd723, %rd2311, %rd722; and.b32 %r1041, %r1039, 63; shl.b64 %rd725, %rd715, %r1041; ld.u64 %rd726, [%rd723+544]; and.b64 %rd727, %rd726, %rd725; setp.eq.s64 %p31, %rd727, 0; mul.wide.u32 %rd728, %r1039, 8; add.s64 %rd729, %rd2311, %rd728; add.s64 %rd47, %rd729, 1088; and.b32 %r3272, %r7, -128; @%p31 bra $L__BB3_50; and.b32 %r3229, %r7, -8; ld.u64 %rd730, [%rd47]; add.s64 %rd2307, %rd2311, %rd730; and.b32 %r1042, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1042}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1043, %r7, 6; and.b32 %r1044, %r1043, 448; shl.b32 %r1045, %r8, 3; and.b32 %r1046, %r1045, 56; or.b32 %r1047, %r1046, %r1044; and.b32 %r1048, %r74, 7; or.b32 %r1049, %r1047, %r1048; mul.wide.u32 %rd731, %r1049, 4; add.s64 %rd732, %rd2307, %rd731; ld.f32 %f460, [%rd732+96]; mov.b64 %rd733, {%r3229, %r1042}; shr.u64 %rd734, %rd733, 32; cvt.u32.u64 %r3184, %rd734; bra.uni $L__BB3_66; $L__BB3_78: setp.ne.s32 %p48, %r12, %r3232; mov.u16 %rs270, 0; @%p48 bra $L__BB3_81; and.b32 %r1107, %r139, -4096; setp.ne.s32 %p49, %r1107, %r3277; @%p49 bra $L__BB3_81; and.b32 %r1108, %r74, -4096; setp.eq.s32 %p50, %r1108, %r3310; selp.u16 %rs270, 1, 0, %p50; $L__BB3_81: setp.eq.s16 %p51, %rs270, 0; @%p51 bra $L__BB3_87; shl.b32 %r1109, %r7, 3; and.b32 %r1110, %r1109, 31744; and.b32 %r1111, %r139, 3968; shr.u32 %r1112, %r1111, 2; or.b32 %r1113, %r1112, %r1110; bfe.u32 %r1114, %r74, 7, 5; or.b32 %r1115, %r1113, %r1114; shr.u32 %r1116, %r1113, 6; mul.wide.u32 %rd782, %r1116, 8; add.s64 %rd783, %rd2315, %rd782; and.b32 %r1117, %r1115, 63; mov.u64 %rd784, 1; shl.b64 %rd785, %rd784, %r1117; ld.u64 %rd786, [%rd783+4128]; and.b64 %rd787, %rd786, %rd785; setp.eq.s64 %p52, %rd787, 0; mul.wide.u32 %rd788, %r1115, 8; add.s64 %rd789, %rd2315, %rd788; add.s64 %rd74, %rd789, 8256; @%p52 bra $L__BB3_86; and.b32 %r3128, %r7, -128; ld.u64 %rd790, [%rd74]; add.s64 %rd2311, %rd2315, %rd790; st.local.u32 [%rd2+12], %r3128; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1118, %r7, 5; and.b32 %r1119, %r1118, 3840; shl.b32 %r1120, %r139, 1; and.b32 %r1121, %r1120, 240; or.b32 %r1122, %r1121, %r1119; bfe.u32 %r1123, %r74, 3, 4; or.b32 %r1124, %r1122, %r1123; shr.u32 %r1125, %r1122, 6; mul.wide.u32 %rd791, %r1125, 8; add.s64 %rd792, %rd2311, %rd791; and.b32 %r1126, %r1124, 63; shl.b64 %rd794, %rd784, %r1126; ld.u64 %rd795, [%rd792+544]; and.b64 %rd796, %rd795, %rd794; setp.eq.s64 %p53, %rd796, 0; mul.wide.u32 %rd797, %r1124, 8; add.s64 %rd798, %rd2311, %rd797; add.s64 %rd76, %rd798, 1088; and.b32 %r3272, %r7, -128; @%p53 bra $L__BB3_85; and.b32 %r3229, %r7, -8; ld.u64 %rd799, [%rd76]; add.s64 %rd2307, %rd2311, %rd799; and.b32 %r1127, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1127}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1128, %r7, 6; and.b32 %r1129, %r1128, 448; shl.b32 %r1130, %r139, 3; and.b32 %r1131, %r1130, 56; or.b32 %r1132, %r1131, %r1129; and.b32 %r1133, %r74, 7; or.b32 %r1134, %r1132, %r1133; mul.wide.u32 %rd800, %r1134, 4; add.s64 %rd801, %rd2307, %rd800; ld.f32 %f461, [%rd801+96]; mov.b64 %rd802, {%r3229, %r1127}; shr.u64 %rd803, %rd802, 32; cvt.u32.u64 %r3184, %rd803; bra.uni $L__BB3_101; $L__BB3_113: setp.ne.s32 %p70, %r12, %r3232; mov.u16 %rs273, 0; @%p70 bra $L__BB3_116; and.b32 %r1191, %r139, -4096; setp.ne.s32 %p71, %r1191, %r3277; @%p71 bra $L__BB3_116; and.b32 %r1192, %r9, -4096; setp.eq.s32 %p72, %r1192, %r3310; selp.u16 %rs273, 1, 0, %p72; $L__BB3_116: setp.eq.s16 %p73, %rs273, 0; @%p73 bra $L__BB3_122; shl.b32 %r1193, %r7, 3; and.b32 %r1194, %r1193, 31744; and.b32 %r1195, %r139, 3968; shr.u32 %r1196, %r1195, 2; or.b32 %r1197, %r1196, %r1194; bfe.u32 %r1198, %r9, 7, 5; or.b32 %r1199, %r1197, %r1198; shr.u32 %r1200, %r1197, 6; mul.wide.u32 %rd852, %r1200, 8; add.s64 %rd853, %rd2315, %rd852; and.b32 %r1201, %r1199, 63; mov.u64 %rd854, 1; shl.b64 %rd855, %rd854, %r1201; ld.u64 %rd856, [%rd853+4128]; and.b64 %rd857, %rd856, %rd855; setp.eq.s64 %p74, %rd857, 0; mul.wide.u32 %rd858, %r1199, 8; add.s64 %rd859, %rd2315, %rd858; add.s64 %rd103, %rd859, 8256; @%p74 bra $L__BB3_121; and.b32 %r3101, %r7, -128; ld.u64 %rd860, [%rd103]; add.s64 %rd2311, %rd2315, %rd860; st.local.u32 [%rd2+12], %r3101; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1202, %r7, 5; and.b32 %r1203, %r1202, 3840; shl.b32 %r1204, %r139, 1; and.b32 %r1205, %r1204, 240; or.b32 %r1206, %r1205, %r1203; bfe.u32 %r1207, %r9, 3, 4; or.b32 %r1208, %r1206, %r1207; shr.u32 %r1209, %r1206, 6; mul.wide.u32 %rd861, %r1209, 8; add.s64 %rd862, %rd2311, %rd861; and.b32 %r1210, %r1208, 63; shl.b64 %rd864, %rd854, %r1210; ld.u64 %rd865, [%rd862+544]; and.b64 %rd866, %rd865, %rd864; setp.eq.s64 %p75, %rd866, 0; mul.wide.u32 %rd867, %r1208, 8; add.s64 %rd868, %rd2311, %rd867; add.s64 %rd105, %rd868, 1088; and.b32 %r3272, %r7, -128; @%p75 bra $L__BB3_120; and.b32 %r3229, %r7, -8; ld.u64 %rd869, [%rd105]; add.s64 %rd2307, %rd2311, %rd869; and.b32 %r1211, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1211}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1212, %r7, 7; shl.b32 %r1213, %r139, 3; and.b32 %r1214, %r1213, 56; and.b32 %r1215, %r9, 7; bfi.b32 %r1216, %r1212, %r1215, 6, 3; or.b32 %r1217, %r1216, %r1214; mul.wide.u32 %rd870, %r1217, 4; add.s64 %rd871, %rd2307, %rd870; ld.f32 %f462, [%rd871+96]; mov.b64 %rd872, {%r3229, %r1211}; shr.u64 %rd873, %rd872, 32; cvt.u32.u64 %r3184, %rd873; bra.uni $L__BB3_136; $L__BB3_148: and.b32 %r264, %r259, -4096; setp.ne.s32 %p92, %r264, %r3232; mov.u16 %rs276, 0; @%p92 bra $L__BB3_151; and.b32 %r1272, %r139, -4096; setp.ne.s32 %p93, %r1272, %r3277; @%p93 bra $L__BB3_151; and.b32 %r1273, %r9, -4096; setp.eq.s32 %p94, %r1273, %r3310; selp.u16 %rs276, 1, 0, %p94; $L__BB3_151: setp.eq.s16 %p95, %rs276, 0; @%p95 bra $L__BB3_157; shl.b32 %r1274, %r259, 3; and.b32 %r1275, %r1274, 31744; and.b32 %r1276, %r139, 3968; shr.u32 %r1277, %r1276, 2; or.b32 %r1278, %r1277, %r1275; bfe.u32 %r1279, %r9, 7, 5; or.b32 %r1280, %r1278, %r1279; shr.u32 %r1281, %r1278, 6; mul.wide.u32 %rd920, %r1281, 8; add.s64 %rd921, %rd2315, %rd920; and.b32 %r1282, %r1280, 63; mov.u64 %rd922, 1; shl.b64 %rd923, %rd922, %r1282; ld.u64 %rd924, [%rd921+4128]; and.b64 %rd925, %rd924, %rd923; setp.eq.s64 %p96, %rd925, 0; mul.wide.u32 %rd926, %r1280, 8; add.s64 %rd927, %rd2315, %rd926; add.s64 %rd132, %rd927, 8256; @%p96 bra $L__BB3_156; and.b32 %r3134, %r259, -128; ld.u64 %rd928, [%rd132]; add.s64 %rd2311, %rd2315, %rd928; st.local.u32 [%rd2+12], %r3134; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1283, %r259, 5; and.b32 %r1284, %r1283, 3840; shl.b32 %r1285, %r139, 1; and.b32 %r1286, %r1285, 240; or.b32 %r1287, %r1286, %r1284; bfe.u32 %r1288, %r9, 3, 4; or.b32 %r1289, %r1287, %r1288; shr.u32 %r1290, %r1287, 6; mul.wide.u32 %rd929, %r1290, 8; add.s64 %rd930, %rd2311, %rd929; and.b32 %r1291, %r1289, 63; shl.b64 %rd932, %rd922, %r1291; ld.u64 %rd933, [%rd930+544]; and.b64 %rd934, %rd933, %rd932; setp.eq.s64 %p97, %rd934, 0; mul.wide.u32 %rd935, %r1289, 8; add.s64 %rd936, %rd2311, %rd935; add.s64 %rd134, %rd936, 1088; @%p97 bra $L__BB3_155; and.b32 %r3229, %r259, -8; and.b32 %r3272, %r259, -128; ld.u64 %rd937, [%rd134]; add.s64 %rd2307, %rd2311, %rd937; and.b32 %r1292, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1292}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1293, %r259, 7; shl.b32 %r1294, %r139, 3; and.b32 %r1295, %r1294, 56; and.b32 %r1296, %r9, 7; bfi.b32 %r1297, %r1293, %r1296, 6, 3; or.b32 %r1298, %r1297, %r1295; mul.wide.u32 %rd938, %r1298, 4; add.s64 %rd939, %rd2307, %rd938; ld.f32 %f463, [%rd939+96]; mov.b64 %rd940, {%r3229, %r1292}; shr.u64 %rd941, %rd940, 32; cvt.u32.u64 %r3184, %rd941; bra.uni $L__BB3_171; $L__BB3_183: and.b32 %r321, %r259, -4096; setp.ne.s32 %p114, %r321, %r3232; mov.u16 %rs279, 0; @%p114 bra $L__BB3_186; and.b32 %r1356, %r139, -4096; setp.ne.s32 %p115, %r1356, %r3277; @%p115 bra $L__BB3_186; and.b32 %r1357, %r74, -4096; setp.eq.s32 %p116, %r1357, %r3310; selp.u16 %rs279, 1, 0, %p116; $L__BB3_186: setp.eq.s16 %p117, %rs279, 0; @%p117 bra $L__BB3_192; shl.b32 %r1358, %r259, 3; and.b32 %r1359, %r1358, 31744; and.b32 %r1360, %r139, 3968; shr.u32 %r1361, %r1360, 2; or.b32 %r1362, %r1361, %r1359; bfe.u32 %r1363, %r74, 7, 5; or.b32 %r1364, %r1362, %r1363; shr.u32 %r1365, %r1362, 6; mul.wide.u32 %rd991, %r1365, 8; add.s64 %rd992, %rd2315, %rd991; and.b32 %r1366, %r1364, 63; mov.u64 %rd993, 1; shl.b64 %rd994, %rd993, %r1366; ld.u64 %rd995, [%rd992+4128]; and.b64 %rd996, %rd995, %rd994; setp.eq.s64 %p118, %rd996, 0; mul.wide.u32 %rd997, %r1364, 8; add.s64 %rd998, %rd2315, %rd997; add.s64 %rd161, %rd998, 8256; @%p118 bra $L__BB3_191; and.b32 %r3137, %r259, -128; ld.u64 %rd999, [%rd161]; add.s64 %rd2311, %rd2315, %rd999; st.local.u32 [%rd2+12], %r3137; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1367, %r259, 5; and.b32 %r1368, %r1367, 3840; shl.b32 %r1369, %r139, 1; and.b32 %r1370, %r1369, 240; or.b32 %r1371, %r1370, %r1368; bfe.u32 %r1372, %r74, 3, 4; or.b32 %r1373, %r1371, %r1372; shr.u32 %r1374, %r1371, 6; mul.wide.u32 %rd1000, %r1374, 8; add.s64 %rd1001, %rd2311, %rd1000; and.b32 %r1375, %r1373, 63; shl.b64 %rd1003, %rd993, %r1375; ld.u64 %rd1004, [%rd1001+544]; and.b64 %rd1005, %rd1004, %rd1003; setp.eq.s64 %p119, %rd1005, 0; mul.wide.u32 %rd1006, %r1373, 8; add.s64 %rd1007, %rd2311, %rd1006; add.s64 %rd163, %rd1007, 1088; @%p119 bra $L__BB3_190; and.b32 %r3229, %r259, -8; and.b32 %r3272, %r259, -128; ld.u64 %rd1008, [%rd163]; add.s64 %rd2307, %rd2311, %rd1008; and.b32 %r1376, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1376}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1377, %r259, 6; and.b32 %r1378, %r1377, 448; shl.b32 %r1379, %r139, 3; and.b32 %r1380, %r1379, 56; or.b32 %r1381, %r1380, %r1378; and.b32 %r1382, %r74, 7; or.b32 %r1383, %r1381, %r1382; mul.wide.u32 %rd1009, %r1383, 4; add.s64 %rd1010, %rd2307, %rd1009; ld.f32 %f464, [%rd1010+96]; mov.b64 %rd1011, {%r3229, %r1376}; shr.u64 %rd1012, %rd1011, 32; cvt.u32.u64 %r3184, %rd1012; bra.uni $L__BB3_206; $L__BB3_218: and.b32 %r375, %r259, -4096; setp.ne.s32 %p136, %r375, %r3232; mov.u16 %rs282, 0; @%p136 bra $L__BB3_221; and.b32 %r1443, %r8, -4096; setp.ne.s32 %p137, %r1443, %r3277; @%p137 bra $L__BB3_221; and.b32 %r1444, %r74, -4096; setp.eq.s32 %p138, %r1444, %r3310; selp.u16 %rs282, 1, 0, %p138; $L__BB3_221: setp.eq.s16 %p139, %rs282, 0; @%p139 bra $L__BB3_227; shl.b32 %r1445, %r259, 3; and.b32 %r1446, %r1445, 31744; and.b32 %r1447, %r8, 3968; shr.u32 %r1448, %r1447, 2; or.b32 %r1449, %r1446, %r1448; bfe.u32 %r1450, %r74, 7, 5; or.b32 %r1451, %r1450, %r1449; shr.u32 %r1452, %r1449, 6; mul.wide.u32 %rd1062, %r1452, 8; add.s64 %rd1063, %rd2315, %rd1062; and.b32 %r1453, %r1451, 63; mov.u64 %rd1064, 1; shl.b64 %rd1065, %rd1064, %r1453; ld.u64 %rd1066, [%rd1063+4128]; and.b64 %rd1067, %rd1066, %rd1065; setp.eq.s64 %p140, %rd1067, 0; mul.wide.u32 %rd1068, %r1451, 8; add.s64 %rd1069, %rd2315, %rd1068; add.s64 %rd190, %rd1069, 8256; @%p140 bra $L__BB3_226; ld.u64 %rd1070, [%rd190]; add.s64 %rd2311, %rd2315, %rd1070; st.local.u32 [%rd2+12], %r372; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1454, %r259, 5; and.b32 %r1455, %r1454, 3840; shl.b32 %r1456, %r8, 1; and.b32 %r1457, %r1456, 240; or.b32 %r1458, %r1455, %r1457; bfe.u32 %r1459, %r74, 3, 4; or.b32 %r1460, %r1459, %r1458; shr.u32 %r1461, %r1458, 6; mul.wide.u32 %rd1071, %r1461, 8; add.s64 %rd1072, %rd2311, %rd1071; and.b32 %r1462, %r1460, 63; shl.b64 %rd1074, %rd1064, %r1462; ld.u64 %rd1075, [%rd1072+544]; and.b64 %rd1076, %rd1075, %rd1074; setp.eq.s64 %p141, %rd1076, 0; mul.wide.u32 %rd1077, %r1460, 8; add.s64 %rd1078, %rd2311, %rd1077; add.s64 %rd192, %rd1078, 1088; @%p141 bra $L__BB3_225; and.b32 %r3229, %r259, -8; ld.u64 %rd1079, [%rd192]; add.s64 %rd2307, %rd2311, %rd1079; and.b32 %r1463, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1463}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1464, %r259, 6; and.b32 %r1465, %r1464, 448; shl.b32 %r1466, %r8, 3; and.b32 %r1467, %r1466, 56; or.b32 %r1468, %r1465, %r1467; and.b32 %r1469, %r74, 7; or.b32 %r1470, %r1468, %r1469; mul.wide.u32 %rd1080, %r1470, 4; add.s64 %rd1081, %rd2307, %rd1080; ld.f32 %f465, [%rd1081+96]; mov.b64 %rd1082, {%r3229, %r1463}; shr.u64 %rd1083, %rd1082, 32; cvt.u32.u64 %r3184, %rd1083; mov.u32 %r3272, %r372; bra.uni $L__BB3_241; $L__BB3_253: and.b32 %r426, %r259, -4096; setp.ne.s32 %p158, %r426, %r3232; mov.u16 %rs285, 0; @%p158 bra $L__BB3_256; and.b32 %r1529, %r8, -4096; setp.ne.s32 %p159, %r1529, %r3277; @%p159 bra $L__BB3_256; and.b32 %r1530, %r9, -4096; setp.eq.s32 %p160, %r1530, %r3310; selp.u16 %rs285, 1, 0, %p160; $L__BB3_256: setp.eq.s16 %p161, %rs285, 0; @%p161 bra $L__BB3_262; shl.b32 %r1531, %r259, 3; and.b32 %r1532, %r1531, 31744; and.b32 %r1533, %r8, 3968; shr.u32 %r1534, %r1533, 2; or.b32 %r1535, %r1532, %r1534; bfe.u32 %r1536, %r9, 7, 5; or.b32 %r1537, %r1535, %r1536; shr.u32 %r1538, %r1535, 6; mul.wide.u32 %rd1133, %r1538, 8; add.s64 %rd1134, %rd2315, %rd1133; and.b32 %r1539, %r1537, 63; mov.u64 %rd1135, 1; shl.b64 %rd1136, %rd1135, %r1539; ld.u64 %rd1137, [%rd1134+4128]; and.b64 %rd1138, %rd1137, %rd1136; setp.eq.s64 %p162, %rd1138, 0; mul.wide.u32 %rd1139, %r1537, 8; add.s64 %rd1140, %rd2315, %rd1139; add.s64 %rd219, %rd1140, 8256; @%p162 bra $L__BB3_261; ld.u64 %rd1141, [%rd219]; add.s64 %rd2311, %rd2315, %rd1141; st.local.u32 [%rd2+12], %r423; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1540, %r259, 5; and.b32 %r1541, %r1540, 3840; shl.b32 %r1542, %r8, 1; and.b32 %r1543, %r1542, 240; or.b32 %r1544, %r1541, %r1543; bfe.u32 %r1545, %r9, 3, 4; or.b32 %r1546, %r1544, %r1545; shr.u32 %r1547, %r1544, 6; mul.wide.u32 %rd1142, %r1547, 8; add.s64 %rd1143, %rd2311, %rd1142; and.b32 %r1548, %r1546, 63; shl.b64 %rd1145, %rd1135, %r1548; ld.u64 %rd1146, [%rd1143+544]; and.b64 %rd1147, %rd1146, %rd1145; setp.eq.s64 %p163, %rd1147, 0; mul.wide.u32 %rd1148, %r1546, 8; add.s64 %rd1149, %rd2311, %rd1148; add.s64 %rd221, %rd1149, 1088; @%p163 bra $L__BB3_260; and.b32 %r3229, %r259, -8; ld.u64 %rd1150, [%rd221]; add.s64 %rd2307, %rd2311, %rd1150; and.b32 %r1549, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1549}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1550, %r259, 6; and.b32 %r1551, %r1550, 448; shl.b32 %r1552, %r8, 3; and.b32 %r1553, %r1552, 56; or.b32 %r1554, %r1551, %r1553; and.b32 %r1555, %r9, 7; or.b32 %r1556, %r1554, %r1555; mul.wide.u32 %rd1151, %r1556, 4; add.s64 %rd1152, %rd2307, %rd1151; ld.f32 %f466, [%rd1152+96]; mov.b64 %rd1153, {%r3229, %r1549}; shr.u64 %rd1154, %rd1153, 32; cvt.u32.u64 %r3184, %rd1154; mov.u32 %r3272, %r423; bra.uni $L__BB3_276; $L__BB3_288: and.b32 %r476, %r471, -4096; setp.ne.s32 %p180, %r476, %r3232; mov.u16 %rs288, 0; @%p180 bra $L__BB3_291; and.b32 %r1614, %r8, -4096; setp.ne.s32 %p181, %r1614, %r3277; @%p181 bra $L__BB3_291; and.b32 %r1615, %r9, -4096; setp.eq.s32 %p182, %r1615, %r3310; selp.u16 %rs288, 1, 0, %p182; $L__BB3_291: setp.eq.s16 %p183, %rs288, 0; @%p183 bra $L__BB3_297; shl.b32 %r1616, %r471, 3; and.b32 %r1617, %r1616, 31744; and.b32 %r1618, %r8, 3968; shr.u32 %r1619, %r1618, 2; or.b32 %r1620, %r1619, %r1617; bfe.u32 %r1621, %r9, 7, 5; or.b32 %r1622, %r1620, %r1621; shr.u32 %r1623, %r1620, 6; mul.wide.u32 %rd1203, %r1623, 8; add.s64 %rd1204, %rd2315, %rd1203; and.b32 %r1624, %r1622, 63; mov.u64 %rd1205, 1; shl.b64 %rd1206, %rd1205, %r1624; ld.u64 %rd1207, [%rd1204+4128]; and.b64 %rd1208, %rd1207, %rd1206; setp.eq.s64 %p184, %rd1208, 0; mul.wide.u32 %rd1209, %r1622, 8; add.s64 %rd1210, %rd2315, %rd1209; add.s64 %rd248, %rd1210, 8256; @%p184 bra $L__BB3_296; ld.u64 %rd1211, [%rd248]; add.s64 %rd2311, %rd2315, %rd1211; st.local.u32 [%rd2+12], %r473; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1625, %r471, 5; and.b32 %r1626, %r1625, 3840; shl.b32 %r1627, %r8, 1; and.b32 %r1628, %r1627, 240; or.b32 %r1629, %r1628, %r1626; bfe.u32 %r1630, %r9, 3, 4; or.b32 %r1631, %r1629, %r1630; shr.u32 %r1632, %r1629, 6; mul.wide.u32 %rd1212, %r1632, 8; add.s64 %rd1213, %rd2311, %rd1212; and.b32 %r1633, %r1631, 63; shl.b64 %rd1215, %rd1205, %r1633; ld.u64 %rd1216, [%rd1213+544]; and.b64 %rd1217, %rd1216, %rd1215; setp.eq.s64 %p185, %rd1217, 0; mul.wide.u32 %rd1218, %r1631, 8; add.s64 %rd1219, %rd2311, %rd1218; add.s64 %rd250, %rd1219, 1088; @%p185 bra $L__BB3_295; ld.u64 %rd1220, [%rd250]; add.s64 %rd2307, %rd2311, %rd1220; and.b32 %r1634, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1634}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1635, %r471, 6; and.b32 %r1636, %r1635, 448; shl.b32 %r1637, %r8, 3; and.b32 %r1638, %r1637, 56; or.b32 %r1639, %r1638, %r1636; and.b32 %r1640, %r9, 7; or.b32 %r1641, %r1639, %r1640; mul.wide.u32 %rd1221, %r1641, 4; add.s64 %rd1222, %rd2307, %rd1221; ld.f32 %f467, [%rd1222+96]; mov.b64 %rd1223, {%r472, %r1634}; shr.u64 %rd1224, %rd1223, 32; cvt.u32.u64 %r3184, %rd1224; mov.u32 %r3272, %r473; mov.u32 %r3229, %r472; bra.uni $L__BB3_311; $L__BB3_323: and.b32 %r521, %r471, -4096; setp.ne.s32 %p202, %r521, %r3232; mov.u16 %rs291, 0; @%p202 bra $L__BB3_326; and.b32 %r1699, %r8, -4096; setp.ne.s32 %p203, %r1699, %r3277; @%p203 bra $L__BB3_326; and.b32 %r1700, %r74, -4096; setp.eq.s32 %p204, %r1700, %r3310; selp.u16 %rs291, 1, 0, %p204; $L__BB3_326: setp.eq.s16 %p205, %rs291, 0; @%p205 bra $L__BB3_332; shl.b32 %r1701, %r471, 3; and.b32 %r1702, %r1701, 31744; and.b32 %r1703, %r8, 3968; shr.u32 %r1704, %r1703, 2; or.b32 %r1705, %r1704, %r1702; bfe.u32 %r1706, %r74, 7, 5; or.b32 %r1707, %r1706, %r1705; shr.u32 %r1708, %r1705, 6; mul.wide.u32 %rd1273, %r1708, 8; add.s64 %rd1274, %rd2315, %rd1273; and.b32 %r1709, %r1707, 63; mov.u64 %rd1275, 1; shl.b64 %rd1276, %rd1275, %r1709; ld.u64 %rd1277, [%rd1274+4128]; and.b64 %rd1278, %rd1277, %rd1276; setp.eq.s64 %p206, %rd1278, 0; mul.wide.u32 %rd1279, %r1707, 8; add.s64 %rd1280, %rd2315, %rd1279; add.s64 %rd277, %rd1280, 8256; @%p206 bra $L__BB3_331; ld.u64 %rd1281, [%rd277]; add.s64 %rd2311, %rd2315, %rd1281; st.local.u32 [%rd2+12], %r518; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1710, %r471, 5; and.b32 %r1711, %r1710, 3840; shl.b32 %r1712, %r8, 1; and.b32 %r1713, %r1712, 240; or.b32 %r1714, %r1713, %r1711; bfe.u32 %r1715, %r74, 3, 4; or.b32 %r1716, %r1715, %r1714; shr.u32 %r1717, %r1714, 6; mul.wide.u32 %rd1282, %r1717, 8; add.s64 %rd1283, %rd2311, %rd1282; and.b32 %r1718, %r1716, 63; shl.b64 %rd1285, %rd1275, %r1718; ld.u64 %rd1286, [%rd1283+544]; and.b64 %rd1287, %rd1286, %rd1285; setp.eq.s64 %p207, %rd1287, 0; mul.wide.u32 %rd1288, %r1716, 8; add.s64 %rd1289, %rd2311, %rd1288; add.s64 %rd279, %rd1289, 1088; @%p207 bra $L__BB3_330; ld.u64 %rd1290, [%rd279]; add.s64 %rd2307, %rd2311, %rd1290; and.b32 %r1719, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1719}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1720, %r471, 6; and.b32 %r1721, %r1720, 448; shl.b32 %r1722, %r8, 3; and.b32 %r1723, %r1722, 56; or.b32 %r1724, %r1723, %r1721; and.b32 %r1725, %r74, 7; or.b32 %r1726, %r1724, %r1725; mul.wide.u32 %rd1291, %r1726, 4; add.s64 %rd1292, %rd2307, %rd1291; ld.f32 %f468, [%rd1292+96]; mov.b64 %rd1293, {%r472, %r1719}; shr.u64 %rd1294, %rd1293, 32; cvt.u32.u64 %r3184, %rd1294; mov.u32 %r3272, %r518; mov.u32 %r3229, %r472; bra.uni $L__BB3_346; $L__BB3_358: and.b32 %r563, %r471, -4096; setp.ne.s32 %p224, %r563, %r3232; mov.u16 %rs294, 0; @%p224 bra $L__BB3_361; and.b32 %r1785, %r139, -4096; setp.ne.s32 %p225, %r1785, %r3277; @%p225 bra $L__BB3_361; and.b32 %r1786, %r74, -4096; setp.eq.s32 %p226, %r1786, %r3310; selp.u16 %rs294, 1, 0, %p226; $L__BB3_361: setp.eq.s16 %p227, %rs294, 0; @%p227 bra $L__BB3_367; shl.b32 %r1787, %r471, 3; and.b32 %r1788, %r1787, 31744; and.b32 %r1789, %r139, 3968; shr.u32 %r1790, %r1789, 2; or.b32 %r1791, %r1790, %r1788; bfe.u32 %r1792, %r74, 7, 5; or.b32 %r1793, %r1791, %r1792; shr.u32 %r1794, %r1791, 6; mul.wide.u32 %rd1344, %r1794, 8; add.s64 %rd1345, %rd2315, %rd1344; and.b32 %r1795, %r1793, 63; mov.u64 %rd1346, 1; shl.b64 %rd1347, %rd1346, %r1795; ld.u64 %rd1348, [%rd1345+4128]; and.b64 %rd1349, %rd1348, %rd1347; setp.eq.s64 %p228, %rd1349, 0; mul.wide.u32 %rd1350, %r1793, 8; add.s64 %rd1351, %rd2315, %rd1350; add.s64 %rd306, %rd1351, 8256; @%p228 bra $L__BB3_366; ld.u64 %rd1352, [%rd306]; add.s64 %rd2311, %rd2315, %rd1352; st.local.u32 [%rd2+12], %r560; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1796, %r471, 5; and.b32 %r1797, %r1796, 3840; shl.b32 %r1798, %r139, 1; and.b32 %r1799, %r1798, 240; or.b32 %r1800, %r1799, %r1797; bfe.u32 %r1801, %r74, 3, 4; or.b32 %r1802, %r1800, %r1801; shr.u32 %r1803, %r1800, 6; mul.wide.u32 %rd1353, %r1803, 8; add.s64 %rd1354, %rd2311, %rd1353; and.b32 %r1804, %r1802, 63; shl.b64 %rd1356, %rd1346, %r1804; ld.u64 %rd1357, [%rd1354+544]; and.b64 %rd1358, %rd1357, %rd1356; setp.eq.s64 %p229, %rd1358, 0; mul.wide.u32 %rd1359, %r1802, 8; add.s64 %rd1360, %rd2311, %rd1359; add.s64 %rd308, %rd1360, 1088; @%p229 bra $L__BB3_365; ld.u64 %rd1361, [%rd308]; add.s64 %rd2307, %rd2311, %rd1361; and.b32 %r1805, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1805}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1806, %r471, 6; and.b32 %r1807, %r1806, 448; shl.b32 %r1808, %r139, 3; and.b32 %r1809, %r1808, 56; or.b32 %r1810, %r1809, %r1807; and.b32 %r1811, %r74, 7; or.b32 %r1812, %r1810, %r1811; mul.wide.u32 %rd1362, %r1812, 4; add.s64 %rd1363, %rd2307, %rd1362; ld.f32 %f469, [%rd1363+96]; mov.b64 %rd1364, {%r472, %r1805}; shr.u64 %rd1365, %rd1364, 32; cvt.u32.u64 %r3184, %rd1365; mov.u32 %r3272, %r560; mov.u32 %r3229, %r472; bra.uni $L__BB3_381; $L__BB3_393: and.b32 %r602, %r471, -4096; setp.ne.s32 %p246, %r602, %r3232; mov.u16 %rs297, 0; @%p246 bra $L__BB3_396; and.b32 %r1870, %r139, -4096; setp.ne.s32 %p247, %r1870, %r3277; @%p247 bra $L__BB3_396; and.b32 %r1871, %r9, -4096; setp.eq.s32 %p248, %r1871, %r3310; selp.u16 %rs297, 1, 0, %p248; $L__BB3_396: setp.eq.s16 %p249, %rs297, 0; @%p249 bra $L__BB3_402; shl.b32 %r1872, %r471, 3; and.b32 %r1873, %r1872, 31744; and.b32 %r1874, %r139, 3968; shr.u32 %r1875, %r1874, 2; or.b32 %r1876, %r1875, %r1873; bfe.u32 %r1877, %r9, 7, 5; or.b32 %r1878, %r1876, %r1877; shr.u32 %r1879, %r1876, 6; mul.wide.u32 %rd1415, %r1879, 8; add.s64 %rd1416, %rd2315, %rd1415; and.b32 %r1880, %r1878, 63; mov.u64 %rd1417, 1; shl.b64 %rd1418, %rd1417, %r1880; ld.u64 %rd1419, [%rd1416+4128]; and.b64 %rd1420, %rd1419, %rd1418; setp.eq.s64 %p250, %rd1420, 0; mul.wide.u32 %rd1421, %r1878, 8; add.s64 %rd1422, %rd2315, %rd1421; add.s64 %rd335, %rd1422, 8256; @%p250 bra $L__BB3_401; ld.u64 %rd1423, [%rd335]; add.s64 %rd2311, %rd2315, %rd1423; st.local.u32 [%rd2+12], %r599; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1881, %r471, 5; and.b32 %r1882, %r1881, 3840; shl.b32 %r1883, %r139, 1; and.b32 %r1884, %r1883, 240; or.b32 %r1885, %r1884, %r1882; bfe.u32 %r1886, %r9, 3, 4; or.b32 %r1887, %r1885, %r1886; shr.u32 %r1888, %r1885, 6; mul.wide.u32 %rd1424, %r1888, 8; add.s64 %rd1425, %rd2311, %rd1424; and.b32 %r1889, %r1887, 63; shl.b64 %rd1427, %rd1417, %r1889; ld.u64 %rd1428, [%rd1425+544]; and.b64 %rd1429, %rd1428, %rd1427; setp.eq.s64 %p251, %rd1429, 0; mul.wide.u32 %rd1430, %r1887, 8; add.s64 %rd1431, %rd2311, %rd1430; add.s64 %rd337, %rd1431, 1088; @%p251 bra $L__BB3_400; ld.u64 %rd1432, [%rd337]; add.s64 %rd2307, %rd2311, %rd1432; and.b32 %r1890, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1890}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1891, %r471, 7; shl.b32 %r1892, %r139, 3; and.b32 %r1893, %r1892, 56; and.b32 %r1894, %r9, 7; bfi.b32 %r1895, %r1891, %r1894, 6, 3; or.b32 %r1896, %r1895, %r1893; mul.wide.u32 %rd1433, %r1896, 4; add.s64 %rd1434, %rd2307, %rd1433; ld.f32 %f470, [%rd1434+96]; mov.b64 %rd1435, {%r472, %r1890}; shr.u64 %rd1436, %rd1435, 32; cvt.u32.u64 %r3184, %rd1436; mov.u32 %r3272, %r599; mov.u32 %r3229, %r472; bra.uni $L__BB3_416; $L__BB3_428: and.b32 %r640, %r635, -4096; setp.ne.s32 %p268, %r640, %r3232; mov.u16 %rs300, 0; @%p268 bra $L__BB3_431; and.b32 %r1952, %r139, -4096; setp.ne.s32 %p269, %r1952, %r3277; @%p269 bra $L__BB3_431; and.b32 %r1953, %r9, -4096; setp.eq.s32 %p270, %r1953, %r3310; selp.u16 %rs300, 1, 0, %p270; $L__BB3_431: setp.eq.s16 %p271, %rs300, 0; @%p271 bra $L__BB3_437; shl.b32 %r1954, %r635, 3; and.b32 %r1955, %r1954, 31744; and.b32 %r1956, %r139, 3968; shr.u32 %r1957, %r1956, 2; or.b32 %r1958, %r1957, %r1955; bfe.u32 %r1959, %r9, 7, 5; or.b32 %r1960, %r1958, %r1959; shr.u32 %r1961, %r1958, 6; mul.wide.u32 %rd1486, %r1961, 8; add.s64 %rd1487, %rd2315, %rd1486; and.b32 %r1962, %r1960, 63; mov.u64 %rd1488, 1; shl.b64 %rd1489, %rd1488, %r1962; ld.u64 %rd1490, [%rd1487+4128]; and.b64 %rd1491, %rd1490, %rd1489; setp.eq.s64 %p272, %rd1491, 0; mul.wide.u32 %rd1492, %r1960, 8; add.s64 %rd1493, %rd2315, %rd1492; add.s64 %rd364, %rd1493, 8256; @%p272 bra $L__BB3_436; ld.u64 %rd1494, [%rd364]; add.s64 %rd2311, %rd2315, %rd1494; st.local.u32 [%rd2+12], %r637; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1963, %r635, 5; and.b32 %r1964, %r1963, 3840; shl.b32 %r1965, %r139, 1; and.b32 %r1966, %r1965, 240; or.b32 %r1967, %r1966, %r1964; bfe.u32 %r1968, %r9, 3, 4; or.b32 %r1969, %r1967, %r1968; shr.u32 %r1970, %r1967, 6; mul.wide.u32 %rd1495, %r1970, 8; add.s64 %rd1496, %rd2311, %rd1495; and.b32 %r1971, %r1969, 63; shl.b64 %rd1498, %rd1488, %r1971; ld.u64 %rd1499, [%rd1496+544]; and.b64 %rd1500, %rd1499, %rd1498; setp.eq.s64 %p273, %rd1500, 0; mul.wide.u32 %rd1501, %r1969, 8; add.s64 %rd1502, %rd2311, %rd1501; add.s64 %rd366, %rd1502, 1088; @%p273 bra $L__BB3_435; ld.u64 %rd1503, [%rd366]; add.s64 %rd2307, %rd2311, %rd1503; and.b32 %r1972, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1972}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1973, %r635, 7; shl.b32 %r1974, %r139, 3; and.b32 %r1975, %r1974, 56; and.b32 %r1976, %r9, 7; bfi.b32 %r1977, %r1973, %r1976, 6, 3; or.b32 %r1978, %r1977, %r1975; mul.wide.u32 %rd1504, %r1978, 4; add.s64 %rd1505, %rd2307, %rd1504; ld.f32 %f471, [%rd1505+96]; mov.b64 %rd1506, {%r636, %r1972}; shr.u64 %rd1507, %rd1506, 32; cvt.u32.u64 %r3184, %rd1507; mov.u32 %r3272, %r637; mov.u32 %r3229, %r636; bra.uni $L__BB3_451; $L__BB3_463: and.b32 %r673, %r635, -4096; setp.ne.s32 %p290, %r673, %r3232; mov.u16 %rs303, 0; @%p290 bra $L__BB3_466; and.b32 %r2036, %r139, -4096; setp.ne.s32 %p291, %r2036, %r3277; @%p291 bra $L__BB3_466; and.b32 %r2037, %r74, -4096; setp.eq.s32 %p292, %r2037, %r3310; selp.u16 %rs303, 1, 0, %p292; $L__BB3_466: setp.eq.s16 %p293, %rs303, 0; @%p293 bra $L__BB3_472; shl.b32 %r2038, %r635, 3; and.b32 %r2039, %r2038, 31744; and.b32 %r2040, %r139, 3968; shr.u32 %r2041, %r2040, 2; or.b32 %r2042, %r2041, %r2039; bfe.u32 %r2043, %r74, 7, 5; or.b32 %r2044, %r2042, %r2043; shr.u32 %r2045, %r2042, 6; mul.wide.u32 %rd1557, %r2045, 8; add.s64 %rd1558, %rd2315, %rd1557; and.b32 %r2046, %r2044, 63; mov.u64 %rd1559, 1; shl.b64 %rd1560, %rd1559, %r2046; ld.u64 %rd1561, [%rd1558+4128]; and.b64 %rd1562, %rd1561, %rd1560; setp.eq.s64 %p294, %rd1562, 0; mul.wide.u32 %rd1563, %r2044, 8; add.s64 %rd1564, %rd2315, %rd1563; add.s64 %rd393, %rd1564, 8256; @%p294 bra $L__BB3_471; ld.u64 %rd1565, [%rd393]; add.s64 %rd2311, %rd2315, %rd1565; st.local.u32 [%rd2+12], %r670; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2047, %r635, 5; and.b32 %r2048, %r2047, 3840; shl.b32 %r2049, %r139, 1; and.b32 %r2050, %r2049, 240; or.b32 %r2051, %r2050, %r2048; bfe.u32 %r2052, %r74, 3, 4; or.b32 %r2053, %r2051, %r2052; shr.u32 %r2054, %r2051, 6; mul.wide.u32 %rd1566, %r2054, 8; add.s64 %rd1567, %rd2311, %rd1566; and.b32 %r2055, %r2053, 63; shl.b64 %rd1569, %rd1559, %r2055; ld.u64 %rd1570, [%rd1567+544]; and.b64 %rd1571, %rd1570, %rd1569; setp.eq.s64 %p295, %rd1571, 0; mul.wide.u32 %rd1572, %r2053, 8; add.s64 %rd1573, %rd2311, %rd1572; add.s64 %rd395, %rd1573, 1088; @%p295 bra $L__BB3_470; ld.u64 %rd1574, [%rd395]; add.s64 %rd2307, %rd2311, %rd1574; and.b32 %r2056, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2056}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2057, %r635, 6; and.b32 %r2058, %r2057, 448; shl.b32 %r2059, %r139, 3; and.b32 %r2060, %r2059, 56; or.b32 %r2061, %r2060, %r2058; and.b32 %r2062, %r74, 7; or.b32 %r2063, %r2061, %r2062; mul.wide.u32 %rd1575, %r2063, 4; add.s64 %rd1576, %rd2307, %rd1575; ld.f32 %f472, [%rd1576+96]; mov.b64 %rd1577, {%r636, %r2056}; shr.u64 %rd1578, %rd1577, 32; cvt.u32.u64 %r3184, %rd1578; mov.u32 %r3272, %r670; mov.u32 %r3229, %r636; bra.uni $L__BB3_486; $L__BB3_498: and.b32 %r703, %r635, -4096; setp.ne.s32 %p312, %r703, %r3232; mov.u16 %rs306, 0; @%p312 bra $L__BB3_501; and.b32 %r2123, %r8, -4096; setp.ne.s32 %p313, %r2123, %r3277; @%p313 bra $L__BB3_501; and.b32 %r2124, %r74, -4096; setp.eq.s32 %p314, %r2124, %r3310; selp.u16 %rs306, 1, 0, %p314; $L__BB3_501: setp.eq.s16 %p315, %rs306, 0; @%p315 bra $L__BB3_507; shl.b32 %r2125, %r635, 3; and.b32 %r2126, %r2125, 31744; and.b32 %r2127, %r8, 3968; shr.u32 %r2128, %r2127, 2; or.b32 %r2129, %r2126, %r2128; bfe.u32 %r2130, %r74, 7, 5; or.b32 %r2131, %r2130, %r2129; shr.u32 %r2132, %r2129, 6; mul.wide.u32 %rd1628, %r2132, 8; add.s64 %rd1629, %rd2315, %rd1628; and.b32 %r2133, %r2131, 63; mov.u64 %rd1630, 1; shl.b64 %rd1631, %rd1630, %r2133; ld.u64 %rd1632, [%rd1629+4128]; and.b64 %rd1633, %rd1632, %rd1631; setp.eq.s64 %p316, %rd1633, 0; mul.wide.u32 %rd1634, %r2131, 8; add.s64 %rd1635, %rd2315, %rd1634; add.s64 %rd422, %rd1635, 8256; @%p316 bra $L__BB3_506; ld.u64 %rd1636, [%rd422]; add.s64 %rd2311, %rd2315, %rd1636; st.local.u32 [%rd2+12], %r700; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2134, %r635, 5; and.b32 %r2135, %r2134, 3840; shl.b32 %r2136, %r8, 1; and.b32 %r2137, %r2136, 240; or.b32 %r2138, %r2135, %r2137; bfe.u32 %r2139, %r74, 3, 4; or.b32 %r2140, %r2139, %r2138; shr.u32 %r2141, %r2138, 6; mul.wide.u32 %rd1637, %r2141, 8; add.s64 %rd1638, %rd2311, %rd1637; and.b32 %r2142, %r2140, 63; shl.b64 %rd1640, %rd1630, %r2142; ld.u64 %rd1641, [%rd1638+544]; and.b64 %rd1642, %rd1641, %rd1640; setp.eq.s64 %p317, %rd1642, 0; mul.wide.u32 %rd1643, %r2140, 8; add.s64 %rd1644, %rd2311, %rd1643; add.s64 %rd424, %rd1644, 1088; @%p317 bra $L__BB3_505; ld.u64 %rd1645, [%rd424]; add.s64 %rd2307, %rd2311, %rd1645; and.b32 %r2143, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2143}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2144, %r635, 6; and.b32 %r2145, %r2144, 448; shl.b32 %r2146, %r8, 3; and.b32 %r2147, %r2146, 56; or.b32 %r2148, %r2145, %r2147; and.b32 %r2149, %r74, 7; or.b32 %r2150, %r2148, %r2149; mul.wide.u32 %rd1646, %r2150, 4; add.s64 %rd1647, %rd2307, %rd1646; ld.f32 %f473, [%rd1647+96]; mov.b64 %rd1648, {%r636, %r2143}; shr.u64 %rd1649, %rd1648, 32; cvt.u32.u64 %r3184, %rd1649; mov.u32 %r3272, %r700; mov.u32 %r3229, %r636; bra.uni $L__BB3_521; $L__BB3_533: and.b32 %r728, %r635, -4096; setp.ne.s32 %p334, %r728, %r3232; mov.u16 %rs309, 0; @%p334 bra $L__BB3_536; and.b32 %r2210, %r8, -4096; setp.ne.s32 %p335, %r2210, %r3277; @%p335 bra $L__BB3_536; and.b32 %r2211, %r9, -4096; setp.eq.s32 %p336, %r2211, %r3310; selp.u16 %rs309, 1, 0, %p336; $L__BB3_536: setp.eq.s16 %p337, %rs309, 0; @%p337 bra $L__BB3_542; shl.b32 %r2212, %r635, 3; and.b32 %r2213, %r2212, 31744; and.b32 %r2214, %r8, 3968; shr.u32 %r2215, %r2214, 2; or.b32 %r2216, %r2213, %r2215; bfe.u32 %r2217, %r9, 7, 5; or.b32 %r2218, %r2216, %r2217; shr.u32 %r2219, %r2216, 6; mul.wide.u32 %rd1698, %r2219, 8; add.s64 %rd1699, %rd2315, %rd1698; and.b32 %r2220, %r2218, 63; mov.u64 %rd1700, 1; shl.b64 %rd1701, %rd1700, %r2220; ld.u64 %rd1702, [%rd1699+4128]; and.b64 %rd1703, %rd1702, %rd1701; setp.eq.s64 %p338, %rd1703, 0; mul.wide.u32 %rd1704, %r2218, 8; add.s64 %rd1705, %rd2315, %rd1704; add.s64 %rd450, %rd1705, 8256; @%p338 bra $L__BB3_541; ld.u64 %rd1706, [%rd450]; add.s64 %rd451, %rd2315, %rd1706; st.local.u32 [%rd2+12], %r727; and.b32 %r2221, %r8, -128; st.local.u32 [%rd2+16], %r2221; and.b32 %r2222, %r9, -128; st.local.u32 [%rd2+20], %r2222; st.local.u64 [%rd2+56], %rd451; shl.b32 %r2223, %r635, 5; and.b32 %r2224, %r2223, 3840; shl.b32 %r2225, %r8, 1; and.b32 %r2226, %r2225, 240; or.b32 %r2227, %r2224, %r2226; bfe.u32 %r2228, %r9, 3, 4; or.b32 %r2229, %r2227, %r2228; shr.u32 %r2230, %r2227, 6; mul.wide.u32 %rd1707, %r2230, 8; add.s64 %rd1708, %rd451, %rd1707; and.b32 %r2231, %r2229, 63; shl.b64 %rd1710, %rd1700, %r2231; ld.u64 %rd1711, [%rd1708+544]; and.b64 %rd1712, %rd1711, %rd1710; setp.eq.s64 %p339, %rd1712, 0; mul.wide.u32 %rd1713, %r2229, 8; add.s64 %rd1714, %rd451, %rd1713; add.s64 %rd452, %rd1714, 1088; @%p339 bra $L__BB3_540; ld.u64 %rd1715, [%rd452]; add.s64 %rd1716, %rd451, %rd1715; and.b32 %r2232, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2232}; and.b32 %r2233, %r9, -8; st.local.u32 [%rd2+8], %r2233; st.local.u64 [%rd2+48], %rd1716; shl.b32 %r2234, %r635, 6; and.b32 %r2235, %r2234, 448; shl.b32 %r2236, %r8, 3; and.b32 %r2237, %r2236, 56; or.b32 %r2238, %r2235, %r2237; and.b32 %r2239, %r9, 7; or.b32 %r2240, %r2238, %r2239; mul.wide.u32 %rd1717, %r2240, 4; add.s64 %rd1718, %rd1716, %rd1717; ld.f32 %f474, [%rd1718+96]; bra.uni $L__BB3_556; $L__BB3_568: and.b32 %r736, %r471, -4096; ld.local.u32 %r2308, [%rd2+24]; setp.ne.s32 %p356, %r736, %r2308; mov.u16 %rs312, 0; @%p356 bra $L__BB3_571; and.b32 %r2309, %r732, -4096; ld.local.u32 %r2310, [%rd2+28]; setp.ne.s32 %p357, %r2309, %r2310; @%p357 bra $L__BB3_571; and.b32 %r2311, %r733, -4096; ld.local.u32 %r2312, [%rd2+32]; setp.eq.s32 %p358, %r2311, %r2312; selp.u16 %rs312, 1, 0, %p358; $L__BB3_571: setp.eq.s16 %p359, %rs312, 0; @%p359 bra $L__BB3_577; shl.b32 %r2313, %r471, 3; and.b32 %r2314, %r2313, 31744; and.b32 %r2315, %r732, 3968; shr.u32 %r2316, %r2315, 2; or.b32 %r2317, %r2316, %r2314; bfe.u32 %r2318, %r733, 7, 5; or.b32 %r2319, %r2317, %r2318; shr.u32 %r2320, %r2317, 6; ld.local.u64 %rd466, [%rd2+64]; mul.wide.u32 %rd1764, %r2320, 8; add.s64 %rd1765, %rd466, %rd1764; and.b32 %r2321, %r2319, 63; mov.u64 %rd1766, 1; shl.b64 %rd1767, %rd1766, %r2321; ld.u64 %rd1768, [%rd1765+4128]; and.b64 %rd1769, %rd1768, %rd1767; setp.eq.s64 %p360, %rd1769, 0; mul.wide.u32 %rd1770, %r2319, 8; add.s64 %rd1771, %rd466, %rd1770; add.s64 %rd467, %rd1771, 8256; @%p360 bra $L__BB3_576; ld.u64 %rd1772, [%rd467]; add.s64 %rd468, %rd466, %rd1772; st.local.u32 [%rd2+12], %r735; and.b32 %r2322, %r732, -128; st.local.u32 [%rd2+16], %r2322; and.b32 %r2323, %r733, -128; st.local.u32 [%rd2+20], %r2323; st.local.u64 [%rd2+56], %rd468; shl.b32 %r2324, %r471, 5; and.b32 %r2325, %r2324, 3840; shl.b32 %r2326, %r732, 1; and.b32 %r2327, %r2326, 240; or.b32 %r2328, %r2327, %r2325; bfe.u32 %r2329, %r733, 3, 4; or.b32 %r2330, %r2328, %r2329; shr.u32 %r2331, %r2328, 6; mul.wide.u32 %rd1773, %r2331, 8; add.s64 %rd1774, %rd468, %rd1773; and.b32 %r2332, %r2330, 63; shl.b64 %rd1776, %rd1766, %r2332; ld.u64 %rd1777, [%rd1774+544]; and.b64 %rd1778, %rd1777, %rd1776; setp.eq.s64 %p361, %rd1778, 0; mul.wide.u32 %rd1779, %r2330, 8; add.s64 %rd1780, %rd468, %rd1779; add.s64 %rd469, %rd1780, 1088; @%p361 bra $L__BB3_575; ld.u64 %rd1781, [%rd469]; add.s64 %rd1782, %rd468, %rd1781; and.b32 %r2333, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2333}; and.b32 %r2334, %r733, -8; st.local.u32 [%rd2+8], %r2334; st.local.u64 [%rd2+48], %rd1782; shl.b32 %r2335, %r471, 6; and.b32 %r2336, %r2335, 448; shl.b32 %r2337, %r732, 3; and.b32 %r2338, %r2337, 56; or.b32 %r2339, %r2338, %r2336; and.b32 %r2340, %r733, 7; or.b32 %r2341, %r2339, %r2340; mul.wide.u32 %rd1783, %r2341, 4; add.s64 %rd1784, %rd1782, %rd1783; ld.f32 %f475, [%rd1784+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_591; $L__BB3_603: and.b32 %r743, %r471, -4096; ld.local.u32 %r2411, [%rd2+24]; setp.ne.s32 %p378, %r743, %r2411; mov.u16 %rs315, 0; @%p378 bra $L__BB3_606; and.b32 %r2412, %r732, -4096; ld.local.u32 %r2413, [%rd2+28]; setp.ne.s32 %p379, %r2412, %r2413; @%p379 bra $L__BB3_606; and.b32 %r2414, %r741, -4096; ld.local.u32 %r2415, [%rd2+32]; setp.eq.s32 %p380, %r2414, %r2415; selp.u16 %rs315, 1, 0, %p380; $L__BB3_606: setp.eq.s16 %p381, %rs315, 0; @%p381 bra $L__BB3_612; shl.b32 %r2416, %r471, 3; and.b32 %r2417, %r2416, 31744; and.b32 %r2418, %r732, 3968; shr.u32 %r2419, %r2418, 2; or.b32 %r2420, %r2419, %r2417; bfe.u32 %r2421, %r741, 7, 5; or.b32 %r2422, %r2421, %r2420; shr.u32 %r2423, %r2420, 6; ld.local.u64 %rd484, [%rd2+64]; mul.wide.u32 %rd1831, %r2423, 8; add.s64 %rd1832, %rd484, %rd1831; and.b32 %r2424, %r2422, 63; mov.u64 %rd1833, 1; shl.b64 %rd1834, %rd1833, %r2424; ld.u64 %rd1835, [%rd1832+4128]; and.b64 %rd1836, %rd1835, %rd1834; setp.eq.s64 %p382, %rd1836, 0; mul.wide.u32 %rd1837, %r2422, 8; add.s64 %rd1838, %rd484, %rd1837; add.s64 %rd485, %rd1838, 8256; @%p382 bra $L__BB3_611; ld.u64 %rd1839, [%rd485]; add.s64 %rd486, %rd484, %rd1839; st.local.u32 [%rd2+12], %r742; and.b32 %r2425, %r732, -128; st.local.u32 [%rd2+16], %r2425; and.b32 %r2426, %r741, -128; st.local.u32 [%rd2+20], %r2426; st.local.u64 [%rd2+56], %rd486; shl.b32 %r2427, %r471, 5; and.b32 %r2428, %r2427, 3840; shl.b32 %r2429, %r732, 1; and.b32 %r2430, %r2429, 240; or.b32 %r2431, %r2430, %r2428; bfe.u32 %r2432, %r741, 3, 4; or.b32 %r2433, %r2432, %r2431; shr.u32 %r2434, %r2431, 6; mul.wide.u32 %rd1840, %r2434, 8; add.s64 %rd1841, %rd486, %rd1840; and.b32 %r2435, %r2433, 63; shl.b64 %rd1843, %rd1833, %r2435; ld.u64 %rd1844, [%rd1841+544]; and.b64 %rd1845, %rd1844, %rd1843; setp.eq.s64 %p383, %rd1845, 0; mul.wide.u32 %rd1846, %r2433, 8; add.s64 %rd1847, %rd486, %rd1846; add.s64 %rd487, %rd1847, 1088; @%p383 bra $L__BB3_610; ld.u64 %rd1848, [%rd487]; add.s64 %rd1849, %rd486, %rd1848; and.b32 %r2436, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2436}; and.b32 %r2437, %r741, -8; st.local.u32 [%rd2+8], %r2437; st.local.u64 [%rd2+48], %rd1849; shl.b32 %r2438, %r471, 6; and.b32 %r2439, %r2438, 448; shl.b32 %r2440, %r732, 3; and.b32 %r2441, %r2440, 56; or.b32 %r2442, %r2441, %r2439; and.b32 %r2443, %r741, 7; or.b32 %r2444, %r2442, %r2443; mul.wide.u32 %rd1850, %r2444, 4; add.s64 %rd1851, %rd1849, %rd1850; ld.f32 %f476, [%rd1851+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_626; $L__BB3_638: and.b32 %r750, %r471, -4096; ld.local.u32 %r2514, [%rd2+24]; setp.ne.s32 %p400, %r750, %r2514; mov.u16 %rs318, 0; @%p400 bra $L__BB3_641; and.b32 %r2515, %r748, -4096; ld.local.u32 %r2516, [%rd2+28]; setp.ne.s32 %p401, %r2515, %r2516; @%p401 bra $L__BB3_641; and.b32 %r2517, %r741, -4096; ld.local.u32 %r2518, [%rd2+32]; setp.eq.s32 %p402, %r2517, %r2518; selp.u16 %rs318, 1, 0, %p402; $L__BB3_641: setp.eq.s16 %p403, %rs318, 0; @%p403 bra $L__BB3_647; shl.b32 %r2519, %r471, 3; and.b32 %r2520, %r2519, 31744; and.b32 %r2521, %r748, 3968; shr.u32 %r2522, %r2521, 2; or.b32 %r2523, %r2522, %r2520; bfe.u32 %r2524, %r741, 7, 5; or.b32 %r2525, %r2523, %r2524; shr.u32 %r2526, %r2523, 6; ld.local.u64 %rd502, [%rd2+64]; mul.wide.u32 %rd1899, %r2526, 8; add.s64 %rd1900, %rd502, %rd1899; and.b32 %r2527, %r2525, 63; mov.u64 %rd1901, 1; shl.b64 %rd1902, %rd1901, %r2527; ld.u64 %rd1903, [%rd1900+4128]; and.b64 %rd1904, %rd1903, %rd1902; setp.eq.s64 %p404, %rd1904, 0; mul.wide.u32 %rd1905, %r2525, 8; add.s64 %rd1906, %rd502, %rd1905; add.s64 %rd503, %rd1906, 8256; @%p404 bra $L__BB3_646; ld.u64 %rd1907, [%rd503]; add.s64 %rd504, %rd502, %rd1907; st.local.u32 [%rd2+12], %r749; and.b32 %r2528, %r748, -128; st.local.u32 [%rd2+16], %r2528; and.b32 %r2529, %r741, -128; st.local.u32 [%rd2+20], %r2529; st.local.u64 [%rd2+56], %rd504; shl.b32 %r2530, %r471, 5; and.b32 %r2531, %r2530, 3840; shl.b32 %r2532, %r748, 1; and.b32 %r2533, %r2532, 240; or.b32 %r2534, %r2533, %r2531; bfe.u32 %r2535, %r741, 3, 4; or.b32 %r2536, %r2534, %r2535; shr.u32 %r2537, %r2534, 6; mul.wide.u32 %rd1908, %r2537, 8; add.s64 %rd1909, %rd504, %rd1908; and.b32 %r2538, %r2536, 63; shl.b64 %rd1911, %rd1901, %r2538; ld.u64 %rd1912, [%rd1909+544]; and.b64 %rd1913, %rd1912, %rd1911; setp.eq.s64 %p405, %rd1913, 0; mul.wide.u32 %rd1914, %r2536, 8; add.s64 %rd1915, %rd504, %rd1914; add.s64 %rd505, %rd1915, 1088; @%p405 bra $L__BB3_645; ld.u64 %rd1916, [%rd505]; add.s64 %rd1917, %rd504, %rd1916; and.b32 %r2539, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2539}; and.b32 %r2540, %r741, -8; st.local.u32 [%rd2+8], %r2540; st.local.u64 [%rd2+48], %rd1917; shl.b32 %r2541, %r471, 6; and.b32 %r2542, %r2541, 448; shl.b32 %r2543, %r748, 3; and.b32 %r2544, %r2543, 56; or.b32 %r2545, %r2544, %r2542; and.b32 %r2546, %r741, 7; or.b32 %r2547, %r2545, %r2546; mul.wide.u32 %rd1918, %r2547, 4; add.s64 %rd1919, %rd1917, %rd1918; ld.f32 %f477, [%rd1919+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_661; $L__BB3_673: and.b32 %r756, %r471, -4096; ld.local.u32 %r2615, [%rd2+24]; setp.ne.s32 %p422, %r756, %r2615; mov.u16 %rs321, 0; @%p422 bra $L__BB3_676; and.b32 %r2616, %r748, -4096; ld.local.u32 %r2617, [%rd2+28]; setp.ne.s32 %p423, %r2616, %r2617; @%p423 bra $L__BB3_676; and.b32 %r2618, %r733, -4096; ld.local.u32 %r2619, [%rd2+32]; setp.eq.s32 %p424, %r2618, %r2619; selp.u16 %rs321, 1, 0, %p424; $L__BB3_676: setp.eq.s16 %p425, %rs321, 0; @%p425 bra $L__BB3_682; shl.b32 %r2620, %r471, 3; and.b32 %r2621, %r2620, 31744; and.b32 %r2622, %r748, 3968; shr.u32 %r2623, %r2622, 2; or.b32 %r2624, %r2623, %r2621; bfe.u32 %r2625, %r733, 7, 5; or.b32 %r2626, %r2624, %r2625; shr.u32 %r2627, %r2624, 6; ld.local.u64 %rd520, [%rd2+64]; mul.wide.u32 %rd1966, %r2627, 8; add.s64 %rd1967, %rd520, %rd1966; and.b32 %r2628, %r2626, 63; mov.u64 %rd1968, 1; shl.b64 %rd1969, %rd1968, %r2628; ld.u64 %rd1970, [%rd1967+4128]; and.b64 %rd1971, %rd1970, %rd1969; setp.eq.s64 %p426, %rd1971, 0; mul.wide.u32 %rd1972, %r2626, 8; add.s64 %rd1973, %rd520, %rd1972; add.s64 %rd521, %rd1973, 8256; @%p426 bra $L__BB3_681; ld.u64 %rd1974, [%rd521]; add.s64 %rd522, %rd520, %rd1974; st.local.u32 [%rd2+12], %r755; and.b32 %r2629, %r748, -128; st.local.u32 [%rd2+16], %r2629; and.b32 %r2630, %r733, -128; st.local.u32 [%rd2+20], %r2630; st.local.u64 [%rd2+56], %rd522; shl.b32 %r2631, %r471, 5; and.b32 %r2632, %r2631, 3840; shl.b32 %r2633, %r748, 1; and.b32 %r2634, %r2633, 240; or.b32 %r2635, %r2634, %r2632; bfe.u32 %r2636, %r733, 3, 4; or.b32 %r2637, %r2635, %r2636; shr.u32 %r2638, %r2635, 6; mul.wide.u32 %rd1975, %r2638, 8; add.s64 %rd1976, %rd522, %rd1975; and.b32 %r2639, %r2637, 63; shl.b64 %rd1978, %rd1968, %r2639; ld.u64 %rd1979, [%rd1976+544]; and.b64 %rd1980, %rd1979, %rd1978; setp.eq.s64 %p427, %rd1980, 0; mul.wide.u32 %rd1981, %r2637, 8; add.s64 %rd1982, %rd522, %rd1981; add.s64 %rd523, %rd1982, 1088; @%p427 bra $L__BB3_680; ld.u64 %rd1983, [%rd523]; add.s64 %rd1984, %rd522, %rd1983; and.b32 %r2640, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2640}; and.b32 %r2641, %r733, -8; st.local.u32 [%rd2+8], %r2641; st.local.u64 [%rd2+48], %rd1984; and.b32 %r2642, %r471, 7; shl.b32 %r2643, %r748, 3; and.b32 %r2644, %r2643, 56; and.b32 %r2645, %r733, 7; bfi.b32 %r2646, %r2642, %r2645, 6, 3; or.b32 %r2647, %r2646, %r2644; mul.wide.u32 %rd1985, %r2647, 4; add.s64 %rd1986, %rd1984, %rd1985; ld.f32 %f478, [%rd1986+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_696; $L__BB3_708: and.b32 %r762, %r635, -4096; ld.local.u32 %r2714, [%rd2+24]; setp.ne.s32 %p444, %r762, %r2714; mov.u16 %rs324, 0; @%p444 bra $L__BB3_711; and.b32 %r2715, %r748, -4096; ld.local.u32 %r2716, [%rd2+28]; setp.ne.s32 %p445, %r2715, %r2716; @%p445 bra $L__BB3_711; and.b32 %r2717, %r733, -4096; ld.local.u32 %r2718, [%rd2+32]; setp.eq.s32 %p446, %r2717, %r2718; selp.u16 %rs324, 1, 0, %p446; $L__BB3_711: setp.eq.s16 %p447, %rs324, 0; @%p447 bra $L__BB3_717; shl.b32 %r2719, %r635, 3; and.b32 %r2720, %r2719, 31744; and.b32 %r2721, %r748, 3968; shr.u32 %r2722, %r2721, 2; or.b32 %r2723, %r2722, %r2720; bfe.u32 %r2724, %r733, 7, 5; or.b32 %r2725, %r2723, %r2724; shr.u32 %r2726, %r2723, 6; ld.local.u64 %rd538, [%rd2+64]; mul.wide.u32 %rd2033, %r2726, 8; add.s64 %rd2034, %rd538, %rd2033; and.b32 %r2727, %r2725, 63; mov.u64 %rd2035, 1; shl.b64 %rd2036, %rd2035, %r2727; ld.u64 %rd2037, [%rd2034+4128]; and.b64 %rd2038, %rd2037, %rd2036; setp.eq.s64 %p448, %rd2038, 0; mul.wide.u32 %rd2039, %r2725, 8; add.s64 %rd2040, %rd538, %rd2039; add.s64 %rd539, %rd2040, 8256; @%p448 bra $L__BB3_716; ld.u64 %rd2041, [%rd539]; add.s64 %rd540, %rd538, %rd2041; st.local.u32 [%rd2+12], %r761; and.b32 %r2728, %r748, -128; st.local.u32 [%rd2+16], %r2728; and.b32 %r2729, %r733, -128; st.local.u32 [%rd2+20], %r2729; st.local.u64 [%rd2+56], %rd540; shl.b32 %r2730, %r635, 5; and.b32 %r2731, %r2730, 3840; shl.b32 %r2732, %r748, 1; and.b32 %r2733, %r2732, 240; or.b32 %r2734, %r2733, %r2731; bfe.u32 %r2735, %r733, 3, 4; or.b32 %r2736, %r2734, %r2735; shr.u32 %r2737, %r2734, 6; mul.wide.u32 %rd2042, %r2737, 8; add.s64 %rd2043, %rd540, %rd2042; and.b32 %r2738, %r2736, 63; shl.b64 %rd2045, %rd2035, %r2738; ld.u64 %rd2046, [%rd2043+544]; and.b64 %rd2047, %rd2046, %rd2045; setp.eq.s64 %p449, %rd2047, 0; mul.wide.u32 %rd2048, %r2736, 8; add.s64 %rd2049, %rd540, %rd2048; add.s64 %rd541, %rd2049, 1088; @%p449 bra $L__BB3_715; ld.u64 %rd2050, [%rd541]; add.s64 %rd2051, %rd540, %rd2050; and.b32 %r2739, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2739}; and.b32 %r2740, %r733, -8; st.local.u32 [%rd2+8], %r2740; st.local.u64 [%rd2+48], %rd2051; and.b32 %r2741, %r635, 7; shl.b32 %r2742, %r748, 3; and.b32 %r2743, %r2742, 56; and.b32 %r2744, %r733, 7; bfi.b32 %r2745, %r2741, %r2744, 6, 3; or.b32 %r2746, %r2745, %r2743; mul.wide.u32 %rd2052, %r2746, 4; add.s64 %rd2053, %rd2051, %rd2052; ld.f32 %f479, [%rd2053+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_731; $L__BB3_743: and.b32 %r768, %r635, -4096; ld.local.u32 %r2815, [%rd2+24]; setp.ne.s32 %p466, %r768, %r2815; mov.u16 %rs327, 0; @%p466 bra $L__BB3_746; and.b32 %r2816, %r748, -4096; ld.local.u32 %r2817, [%rd2+28]; setp.ne.s32 %p467, %r2816, %r2817; @%p467 bra $L__BB3_746; and.b32 %r2818, %r741, -4096; ld.local.u32 %r2819, [%rd2+32]; setp.eq.s32 %p468, %r2818, %r2819; selp.u16 %rs327, 1, 0, %p468; $L__BB3_746: setp.eq.s16 %p469, %rs327, 0; @%p469 bra $L__BB3_752; shl.b32 %r2820, %r635, 3; and.b32 %r2821, %r2820, 31744; and.b32 %r2822, %r748, 3968; shr.u32 %r2823, %r2822, 2; or.b32 %r2824, %r2823, %r2821; bfe.u32 %r2825, %r741, 7, 5; or.b32 %r2826, %r2824, %r2825; shr.u32 %r2827, %r2824, 6; ld.local.u64 %rd556, [%rd2+64]; mul.wide.u32 %rd2100, %r2827, 8; add.s64 %rd2101, %rd556, %rd2100; and.b32 %r2828, %r2826, 63; mov.u64 %rd2102, 1; shl.b64 %rd2103, %rd2102, %r2828; ld.u64 %rd2104, [%rd2101+4128]; and.b64 %rd2105, %rd2104, %rd2103; setp.eq.s64 %p470, %rd2105, 0; mul.wide.u32 %rd2106, %r2826, 8; add.s64 %rd2107, %rd556, %rd2106; add.s64 %rd557, %rd2107, 8256; @%p470 bra $L__BB3_751; ld.u64 %rd2108, [%rd557]; add.s64 %rd558, %rd556, %rd2108; st.local.u32 [%rd2+12], %r767; and.b32 %r2829, %r748, -128; st.local.u32 [%rd2+16], %r2829; and.b32 %r2830, %r741, -128; st.local.u32 [%rd2+20], %r2830; st.local.u64 [%rd2+56], %rd558; shl.b32 %r2831, %r635, 5; and.b32 %r2832, %r2831, 3840; shl.b32 %r2833, %r748, 1; and.b32 %r2834, %r2833, 240; or.b32 %r2835, %r2834, %r2832; bfe.u32 %r2836, %r741, 3, 4; or.b32 %r2837, %r2835, %r2836; shr.u32 %r2838, %r2835, 6; mul.wide.u32 %rd2109, %r2838, 8; add.s64 %rd2110, %rd558, %rd2109; and.b32 %r2839, %r2837, 63; shl.b64 %rd2112, %rd2102, %r2839; ld.u64 %rd2113, [%rd2110+544]; and.b64 %rd2114, %rd2113, %rd2112; setp.eq.s64 %p471, %rd2114, 0; mul.wide.u32 %rd2115, %r2837, 8; add.s64 %rd2116, %rd558, %rd2115; add.s64 %rd559, %rd2116, 1088; @%p471 bra $L__BB3_750; ld.u64 %rd2117, [%rd559]; add.s64 %rd2118, %rd558, %rd2117; and.b32 %r2840, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2840}; and.b32 %r2841, %r741, -8; st.local.u32 [%rd2+8], %r2841; st.local.u64 [%rd2+48], %rd2118; shl.b32 %r2842, %r635, 6; and.b32 %r2843, %r2842, 448; shl.b32 %r2844, %r748, 3; and.b32 %r2845, %r2844, 56; or.b32 %r2846, %r2845, %r2843; and.b32 %r2847, %r741, 7; or.b32 %r2848, %r2846, %r2847; mul.wide.u32 %rd2119, %r2848, 4; add.s64 %rd2120, %rd2118, %rd2119; ld.f32 %f480, [%rd2120+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_766; $L__BB3_778: and.b32 %r774, %r635, -4096; ld.local.u32 %r2918, [%rd2+24]; setp.ne.s32 %p488, %r774, %r2918; mov.u16 %rs330, 0; @%p488 bra $L__BB3_781; and.b32 %r2919, %r732, -4096; ld.local.u32 %r2920, [%rd2+28]; setp.ne.s32 %p489, %r2919, %r2920; @%p489 bra $L__BB3_781; and.b32 %r2921, %r741, -4096; ld.local.u32 %r2922, [%rd2+32]; setp.eq.s32 %p490, %r2921, %r2922; selp.u16 %rs330, 1, 0, %p490; $L__BB3_781: setp.eq.s16 %p491, %rs330, 0; @%p491 bra $L__BB3_787; shl.b32 %r2923, %r635, 3; and.b32 %r2924, %r2923, 31744; and.b32 %r2925, %r732, 3968; shr.u32 %r2926, %r2925, 2; or.b32 %r2927, %r2924, %r2926; bfe.u32 %r2928, %r741, 7, 5; or.b32 %r2929, %r2928, %r2927; shr.u32 %r2930, %r2927, 6; ld.local.u64 %rd574, [%rd2+64]; mul.wide.u32 %rd2167, %r2930, 8; add.s64 %rd2168, %rd574, %rd2167; and.b32 %r2931, %r2929, 63; mov.u64 %rd2169, 1; shl.b64 %rd2170, %rd2169, %r2931; ld.u64 %rd2171, [%rd2168+4128]; and.b64 %rd2172, %rd2171, %rd2170; setp.eq.s64 %p492, %rd2172, 0; mul.wide.u32 %rd2173, %r2929, 8; add.s64 %rd2174, %rd574, %rd2173; add.s64 %rd575, %rd2174, 8256; @%p492 bra $L__BB3_786; ld.u64 %rd2175, [%rd575]; add.s64 %rd576, %rd574, %rd2175; st.local.u32 [%rd2+12], %r773; and.b32 %r2932, %r732, -128; st.local.u32 [%rd2+16], %r2932; and.b32 %r2933, %r741, -128; st.local.u32 [%rd2+20], %r2933; st.local.u64 [%rd2+56], %rd576; shl.b32 %r2934, %r635, 5; and.b32 %r2935, %r2934, 3840; shl.b32 %r2936, %r732, 1; and.b32 %r2937, %r2936, 240; or.b32 %r2938, %r2935, %r2937; bfe.u32 %r2939, %r741, 3, 4; or.b32 %r2940, %r2939, %r2938; shr.u32 %r2941, %r2938, 6; mul.wide.u32 %rd2176, %r2941, 8; add.s64 %rd2177, %rd576, %rd2176; and.b32 %r2942, %r2940, 63; shl.b64 %rd2179, %rd2169, %r2942; ld.u64 %rd2180, [%rd2177+544]; and.b64 %rd2181, %rd2180, %rd2179; setp.eq.s64 %p493, %rd2181, 0; mul.wide.u32 %rd2182, %r2940, 8; add.s64 %rd2183, %rd576, %rd2182; add.s64 %rd577, %rd2183, 1088; @%p493 bra $L__BB3_785; ld.u64 %rd2184, [%rd577]; add.s64 %rd2185, %rd576, %rd2184; and.b32 %r2943, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2943}; and.b32 %r2944, %r741, -8; st.local.u32 [%rd2+8], %r2944; st.local.u64 [%rd2+48], %rd2185; shl.b32 %r2945, %r635, 6; and.b32 %r2946, %r2945, 448; shl.b32 %r2947, %r732, 3; and.b32 %r2948, %r2947, 56; or.b32 %r2949, %r2946, %r2948; and.b32 %r2950, %r741, 7; or.b32 %r2951, %r2949, %r2950; mul.wide.u32 %rd2186, %r2951, 4; add.s64 %rd2187, %rd2185, %rd2186; ld.f32 %f481, [%rd2187+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_801; $L__BB3_821: ld.f32 %f482, [%rd599]; bra.uni $L__BB3_833; $L__BB3_42: ld.f32 %f460, [%rd43]; bra.uni $L__BB3_66; $L__BB3_77: ld.f32 %f461, [%rd72]; bra.uni $L__BB3_101; $L__BB3_112: ld.f32 %f462, [%rd101]; bra.uni $L__BB3_136; $L__BB3_147: ld.f32 %f463, [%rd130]; bra.uni $L__BB3_171; $L__BB3_182: ld.f32 %f464, [%rd159]; bra.uni $L__BB3_206; $L__BB3_217: ld.f32 %f465, [%rd188]; bra.uni $L__BB3_241; $L__BB3_252: ld.f32 %f466, [%rd217]; bra.uni $L__BB3_276; $L__BB3_287: ld.f32 %f467, [%rd246]; bra.uni $L__BB3_311; $L__BB3_322: ld.f32 %f468, [%rd275]; bra.uni $L__BB3_346; $L__BB3_357: ld.f32 %f469, [%rd304]; bra.uni $L__BB3_381; $L__BB3_392: ld.f32 %f470, [%rd333]; bra.uni $L__BB3_416; $L__BB3_427: ld.f32 %f471, [%rd362]; bra.uni $L__BB3_451; $L__BB3_462: ld.f32 %f472, [%rd391]; bra.uni $L__BB3_486; $L__BB3_497: ld.f32 %f473, [%rd420]; bra.uni $L__BB3_521; $L__BB3_532: ld.f32 %f474, [%rd449]; bra.uni $L__BB3_556; $L__BB3_567: ld.f32 %f475, [%rd465]; bra.uni $L__BB3_591; $L__BB3_602: ld.f32 %f476, [%rd483]; bra.uni $L__BB3_626; $L__BB3_637: ld.f32 %f477, [%rd501]; bra.uni $L__BB3_661; $L__BB3_672: ld.f32 %f478, [%rd519]; bra.uni $L__BB3_696; $L__BB3_707: ld.f32 %f479, [%rd537]; bra.uni $L__BB3_731; $L__BB3_742: ld.f32 %f480, [%rd555]; bra.uni $L__BB3_766; $L__BB3_777: ld.f32 %f481, [%rd573]; bra.uni $L__BB3_801; $L__BB3_820: ld.f32 %f482, [%rd601]; bra.uni $L__BB3_833; $L__BB3_52: ld.u32 %r81, [%rd5]; mov.u64 %rd2302, 0; setp.eq.s32 %p32, %r81, 0; @%p32 bra $L__BB3_57; shr.u32 %r3092, %r7, 12; cvt.u64.u32 %rd2293, %r3092; shl.b64 %rd2292, %rd2293, 42; or.b64 %rd736, %rd19, %rd2292; shr.u32 %r1051, %r74, 12; cvt.u64.u32 %rd737, %r1051; or.b64 %rd51, %rd736, %rd737; mov.u32 %r3181, 0; $L__BB3_54: mul.wide.u32 %rd738, %r3181, 32; add.s64 %rd739, %rd17, %rd738; ld.u64 %rd740, [%rd739]; setp.eq.s64 %p33, %rd740, %rd51; @%p33 bra $L__BB3_56; add.s32 %r3181, %r3181, 1; setp.lt.u32 %p34, %r3181, %r81; @%p34 bra $L__BB3_54; bra.uni $L__BB3_57; $L__BB3_87: ld.u32 %r146, [%rd5]; mov.u64 %rd2308, 0; setp.eq.s32 %p54, %r146, 0; @%p54 bra $L__BB3_92; shr.u32 %r3091, %r7, 12; cvt.u64.u32 %rd2290, %r3091; shl.b64 %rd2289, %rd2290, 42; shr.u32 %r1136, %r74, 12; cvt.u64.u32 %rd805, %r1136; or.b64 %rd806, %rd2289, %rd805; shr.u32 %r1137, %r139, 12; mul.wide.u32 %rd807, %r1137, 2097152; or.b64 %rd80, %rd806, %rd807; mov.u32 %r3230, 0; $L__BB3_89: mul.wide.u32 %rd808, %r3230, 32; add.s64 %rd809, %rd17, %rd808; ld.u64 %rd810, [%rd809]; setp.eq.s64 %p55, %rd810, %rd80; @%p55 bra $L__BB3_91; add.s32 %r3230, %r3230, 1; setp.lt.u32 %p56, %r3230, %r146; @%p56 bra $L__BB3_89; bra.uni $L__BB3_92; $L__BB3_122: ld.u32 %r207, [%rd5]; mov.u64 %rd2314, 0; setp.eq.s32 %p76, %r207, 0; @%p76 bra $L__BB3_127; shr.u32 %r1219, %r139, 12; mul.wide.u32 %rd875, %r1219, 2097152; or.b64 %rd109, %rd21, %rd875; mov.u32 %r3276, 0; $L__BB3_124: mul.wide.u32 %rd876, %r3276, 32; add.s64 %rd877, %rd17, %rd876; ld.u64 %rd878, [%rd877]; setp.eq.s64 %p77, %rd878, %rd109; @%p77 bra $L__BB3_126; add.s32 %r3276, %r3276, 1; setp.lt.u32 %p78, %r3276, %r207; @%p78 bra $L__BB3_124; bra.uni $L__BB3_127; $L__BB3_157: ld.u32 %r269, [%rd5]; mov.u64 %rd2320, 0; setp.eq.s32 %p98, %r269, 0; @%p98 bra $L__BB3_162; shr.u32 %r1300, %r259, 12; cvt.u64.u32 %rd943, %r1300; shl.b64 %rd944, %rd943, 42; or.b64 %rd945, %rd944, %rd18; shr.u32 %r1301, %r139, 12; mul.wide.u32 %rd946, %r1301, 2097152; or.b64 %rd138, %rd945, %rd946; mov.u32 %r3319, 0; $L__BB3_159: mul.wide.u32 %rd947, %r3319, 32; add.s64 %rd948, %rd17, %rd947; ld.u64 %rd949, [%rd948]; setp.eq.s64 %p99, %rd949, %rd138; @%p99 bra $L__BB3_161; add.s32 %r3319, %r3319, 1; setp.lt.u32 %p100, %r3319, %r269; @%p100 bra $L__BB3_159; bra.uni $L__BB3_162; $L__BB3_192: ld.u32 %r326, [%rd5]; mov.u64 %rd2326, 0; setp.eq.s32 %p120, %r326, 0; @%p120 bra $L__BB3_197; shr.u32 %r1385, %r259, 12; cvt.u64.u32 %rd1014, %r1385; shr.u32 %r1386, %r74, 12; cvt.u64.u32 %rd1015, %r1386; bfi.b64 %rd1016, %rd1014, %rd1015, 42, 22; shr.u32 %r1387, %r139, 12; mul.wide.u32 %rd1017, %r1387, 2097152; or.b64 %rd167, %rd1016, %rd1017; mov.u32 %r3359, 0; $L__BB3_194: cvt.u64.u32 %rd168, %r3359; mul.wide.u32 %rd1018, %r3359, 32; add.s64 %rd1019, %rd17, %rd1018; ld.u64 %rd1020, [%rd1019]; setp.eq.s64 %p121, %rd1020, %rd167; add.s32 %r3359, %r3359, 1; @%p121 bra $L__BB3_196; setp.lt.u32 %p122, %r3359, %r326; @%p122 bra $L__BB3_194; bra.uni $L__BB3_197; $L__BB3_227: ld.u32 %r380, [%rd5]; mov.u64 %rd2332, 0; setp.eq.s32 %p142, %r380, 0; @%p142 bra $L__BB3_232; shr.u32 %r1472, %r259, 12; cvt.u64.u32 %rd1085, %r1472; shl.b64 %rd1086, %rd1085, 42; or.b64 %rd1087, %rd1086, %rd19; shr.u32 %r1473, %r74, 12; cvt.u64.u32 %rd1088, %r1473; or.b64 %rd196, %rd1087, %rd1088; mov.u32 %r3396, 0; $L__BB3_229: cvt.u64.u32 %rd197, %r3396; mul.wide.u32 %rd1089, %r3396, 32; add.s64 %rd1090, %rd17, %rd1089; ld.u64 %rd1091, [%rd1090]; setp.eq.s64 %p143, %rd1091, %rd196; add.s32 %r3396, %r3396, 1; @%p143 bra $L__BB3_231; setp.lt.u32 %p144, %r3396, %r380; @%p144 bra $L__BB3_229; bra.uni $L__BB3_232; $L__BB3_262: ld.u32 %r431, [%rd5]; mov.u64 %rd2338, 0; setp.eq.s32 %p164, %r431, 0; @%p164 bra $L__BB3_267; shr.u32 %r1558, %r259, 12; cvt.u64.u32 %rd1156, %r1558; shl.b64 %rd1157, %rd1156, 42; or.b64 %rd1158, %rd19, %rd18; or.b64 %rd225, %rd1158, %rd1157; mov.u32 %r3430, 0; $L__BB3_264: cvt.u64.u32 %rd226, %r3430; mul.wide.u32 %rd1159, %r3430, 32; add.s64 %rd1160, %rd17, %rd1159; ld.u64 %rd1161, [%rd1160]; setp.eq.s64 %p165, %rd1161, %rd225; add.s32 %r3430, %r3430, 1; @%p165 bra $L__BB3_266; setp.lt.u32 %p166, %r3430, %r431; @%p166 bra $L__BB3_264; bra.uni $L__BB3_267; $L__BB3_297: ld.u32 %r481, [%rd5]; mov.u64 %rd2344, 0; setp.eq.s32 %p186, %r481, 0; @%p186 bra $L__BB3_302; shr.u32 %r1643, %r471, 12; cvt.u64.u32 %rd1226, %r1643; shl.b64 %rd1227, %rd1226, 42; or.b64 %rd1228, %rd1227, %rd18; or.b64 %rd254, %rd1228, %rd19; mov.u32 %r3461, 0; $L__BB3_299: cvt.u64.u32 %rd255, %r3461; mul.wide.u32 %rd1229, %r3461, 32; add.s64 %rd1230, %rd17, %rd1229; ld.u64 %rd1231, [%rd1230]; setp.eq.s64 %p187, %rd1231, %rd254; add.s32 %r3461, %r3461, 1; @%p187 bra $L__BB3_301; setp.lt.u32 %p188, %r3461, %r481; @%p188 bra $L__BB3_299; bra.uni $L__BB3_302; $L__BB3_332: ld.u32 %r526, [%rd5]; mov.u64 %rd2350, 0; setp.eq.s32 %p208, %r526, 0; @%p208 bra $L__BB3_337; shr.u32 %r1728, %r471, 12; cvt.u64.u32 %rd1296, %r1728; shl.b64 %rd1297, %rd1296, 42; or.b64 %rd1298, %rd19, %rd1297; shr.u32 %r1729, %r74, 12; cvt.u64.u32 %rd1299, %r1729; or.b64 %rd283, %rd1298, %rd1299; mov.u32 %r3489, 0; $L__BB3_334: cvt.u64.u32 %rd284, %r3489; mul.wide.u32 %rd1300, %r3489, 32; add.s64 %rd1301, %rd17, %rd1300; ld.u64 %rd1302, [%rd1301]; setp.eq.s64 %p209, %rd1302, %rd283; add.s32 %r3489, %r3489, 1; @%p209 bra $L__BB3_336; setp.lt.u32 %p210, %r3489, %r526; @%p210 bra $L__BB3_334; bra.uni $L__BB3_337; $L__BB3_367: ld.u32 %r568, [%rd5]; mov.u64 %rd2356, 0; setp.eq.s32 %p230, %r568, 0; @%p230 bra $L__BB3_372; shr.u32 %r1814, %r471, 12; cvt.u64.u32 %rd1367, %r1814; shr.u32 %r1815, %r74, 12; cvt.u64.u32 %rd1368, %r1815; bfi.b64 %rd1369, %rd1367, %rd1368, 42, 22; shr.u32 %r1816, %r139, 12; mul.wide.u32 %rd1370, %r1816, 2097152; or.b64 %rd312, %rd1369, %rd1370; mov.u32 %r3514, 0; $L__BB3_369: cvt.u64.u32 %rd313, %r3514; mul.wide.u32 %rd1371, %r3514, 32; add.s64 %rd1372, %rd17, %rd1371; ld.u64 %rd1373, [%rd1372]; setp.eq.s64 %p231, %rd1373, %rd312; add.s32 %r3514, %r3514, 1; @%p231 bra $L__BB3_371; setp.lt.u32 %p232, %r3514, %r568; @%p232 bra $L__BB3_369; bra.uni $L__BB3_372; $L__BB3_402: ld.u32 %r607, [%rd5]; mov.u64 %rd2362, 0; setp.eq.s32 %p252, %r607, 0; @%p252 bra $L__BB3_407; shr.u32 %r1898, %r471, 12; cvt.u64.u32 %rd1438, %r1898; shl.b64 %rd1439, %rd1438, 42; or.b64 %rd1440, %rd1439, %rd18; shr.u32 %r1899, %r139, 12; mul.wide.u32 %rd1441, %r1899, 2097152; or.b64 %rd341, %rd1440, %rd1441; mov.u32 %r3536, 0; $L__BB3_404: cvt.u64.u32 %rd342, %r3536; mul.wide.u32 %rd1442, %r3536, 32; add.s64 %rd1443, %rd17, %rd1442; ld.u64 %rd1444, [%rd1443]; setp.eq.s64 %p253, %rd1444, %rd341; add.s32 %r3536, %r3536, 1; @%p253 bra $L__BB3_406; setp.lt.u32 %p254, %r3536, %r607; @%p254 bra $L__BB3_404; bra.uni $L__BB3_407; $L__BB3_437: ld.u32 %r645, [%rd5]; mov.u64 %rd2368, 0; setp.eq.s32 %p274, %r645, 0; @%p274 bra $L__BB3_442; shr.u32 %r1980, %r635, 12; cvt.u64.u32 %rd1509, %r1980; shl.b64 %rd1510, %rd1509, 42; or.b64 %rd1511, %rd1510, %rd18; shr.u32 %r1981, %r139, 12; mul.wide.u32 %rd1512, %r1981, 2097152; or.b64 %rd370, %rd1511, %rd1512; mov.u32 %r3555, 0; $L__BB3_439: cvt.u64.u32 %rd371, %r3555; mul.wide.u32 %rd1513, %r3555, 32; add.s64 %rd1514, %rd17, %rd1513; ld.u64 %rd1515, [%rd1514]; setp.eq.s64 %p275, %rd1515, %rd370; add.s32 %r3555, %r3555, 1; @%p275 bra $L__BB3_441; setp.lt.u32 %p276, %r3555, %r645; @%p276 bra $L__BB3_439; bra.uni $L__BB3_442; $L__BB3_472: ld.u32 %r678, [%rd5]; mov.u64 %rd2374, 0; setp.eq.s32 %p296, %r678, 0; @%p296 bra $L__BB3_477; shr.u32 %r2065, %r635, 12; cvt.u64.u32 %rd1580, %r2065; shr.u32 %r2066, %r74, 12; cvt.u64.u32 %rd1581, %r2066; bfi.b64 %rd1582, %rd1580, %rd1581, 42, 22; shr.u32 %r2067, %r139, 12; mul.wide.u32 %rd1583, %r2067, 2097152; or.b64 %rd399, %rd1582, %rd1583; mov.u32 %r3571, 0; $L__BB3_474: cvt.u64.u32 %rd400, %r3571; mul.wide.u32 %rd1584, %r3571, 32; add.s64 %rd1585, %rd17, %rd1584; ld.u64 %rd1586, [%rd1585]; setp.eq.s64 %p297, %rd1586, %rd399; add.s32 %r3571, %r3571, 1; @%p297 bra $L__BB3_476; setp.lt.u32 %p298, %r3571, %r678; @%p298 bra $L__BB3_474; bra.uni $L__BB3_477; $L__BB3_507: ld.u32 %r708, [%rd5]; mov.u64 %rd2380, 0; setp.eq.s32 %p318, %r708, 0; @%p318 bra $L__BB3_512; shr.u32 %r2152, %r635, 12; cvt.u64.u32 %rd1651, %r2152; shl.b64 %rd1652, %rd1651, 42; or.b64 %rd1653, %rd1652, %rd19; shr.u32 %r2153, %r74, 12; cvt.u64.u32 %rd1654, %r2153; or.b64 %rd428, %rd1653, %rd1654; mov.u32 %r3584, 0; $L__BB3_509: cvt.u64.u32 %rd429, %r3584; mul.wide.u32 %rd1655, %r3584, 32; add.s64 %rd1656, %rd17, %rd1655; ld.u64 %rd1657, [%rd1656]; setp.eq.s64 %p319, %rd1657, %rd428; add.s32 %r3584, %r3584, 1; @%p319 bra $L__BB3_511; setp.lt.u32 %p320, %r3584, %r708; @%p320 bra $L__BB3_509; bra.uni $L__BB3_512; $L__BB3_542: or.b64 %rd453, %rd19, %rd18; ld.u32 %r729, [%rd5]; mov.u64 %rd2386, 0; setp.eq.s32 %p340, %r729, 0; @%p340 bra $L__BB3_547; shr.u32 %r2242, %r635, 12; cvt.u64.u32 %rd1720, %r2242; shl.b64 %rd1721, %rd1720, 42; or.b64 %rd454, %rd453, %rd1721; mov.u32 %r3594, 0; $L__BB3_544: cvt.u64.u32 %rd455, %r3594; mul.wide.u32 %rd1722, %r3594, 32; add.s64 %rd1723, %rd17, %rd1722; ld.u64 %rd1724, [%rd1723]; setp.eq.s64 %p341, %rd1724, %rd454; add.s32 %r3594, %r3594, 1; @%p341 bra $L__BB3_546; setp.lt.u32 %p342, %r3594, %r729; @%p342 bra $L__BB3_544; bra.uni $L__BB3_547; $L__BB3_577: ld.local.u64 %rd470, [%rd2+40]; mov.u64 %rd2387, 0; ld.u32 %r737, [%rd470+24]; setp.eq.s32 %p362, %r737, 0; @%p362 bra $L__BB3_582; add.s64 %rd471, %rd470, 64; shr.u32 %r2343, %r471, 12; cvt.u64.u32 %rd1786, %r2343; shr.u32 %r2344, %r733, 12; cvt.u64.u32 %rd1787, %r2344; bfi.b64 %rd1788, %rd1786, %rd1787, 42, 22; shr.u32 %r2345, %r732, 12; mul.wide.u32 %rd1789, %r2345, 2097152; or.b64 %rd472, %rd1788, %rd1789; mov.u32 %r3595, 0; $L__BB3_579: cvt.u64.u32 %rd473, %r3595; mul.wide.u32 %rd1790, %r3595, 32; add.s64 %rd1791, %rd471, %rd1790; ld.u64 %rd1792, [%rd1791]; setp.eq.s64 %p363, %rd1792, %rd472; add.s32 %r3595, %r3595, 1; @%p363 bra $L__BB3_581; setp.lt.u32 %p364, %r3595, %r737; @%p364 bra $L__BB3_579; bra.uni $L__BB3_582; $L__BB3_612: ld.local.u64 %rd488, [%rd2+40]; mov.u64 %rd2388, 0; ld.u32 %r744, [%rd488+24]; setp.eq.s32 %p384, %r744, 0; @%p384 bra $L__BB3_617; add.s64 %rd489, %rd488, 64; shr.u32 %r2446, %r471, 12; cvt.u64.u32 %rd1853, %r2446; shl.b64 %rd1854, %rd1853, 42; shr.u32 %r2447, %r732, 12; mul.wide.u32 %rd1855, %r2447, 2097152; or.b64 %rd1856, %rd1855, %rd1854; shr.u32 %r2448, %r741, 12; cvt.u64.u32 %rd1857, %r2448; or.b64 %rd490, %rd1856, %rd1857; mov.u32 %r3597, 0; $L__BB3_614: cvt.u64.u32 %rd491, %r3597; mul.wide.u32 %rd1858, %r3597, 32; add.s64 %rd1859, %rd489, %rd1858; ld.u64 %rd1860, [%rd1859]; setp.eq.s64 %p385, %rd1860, %rd490; add.s32 %r3597, %r3597, 1; @%p385 bra $L__BB3_616; setp.lt.u32 %p386, %r3597, %r744; @%p386 bra $L__BB3_614; bra.uni $L__BB3_617; $L__BB3_647: ld.local.u64 %rd506, [%rd2+40]; mov.u64 %rd2389, 0; ld.u32 %r751, [%rd506+24]; setp.eq.s32 %p406, %r751, 0; @%p406 bra $L__BB3_652; add.s64 %rd507, %rd506, 64; shr.u32 %r2549, %r471, 12; cvt.u64.u32 %rd1921, %r2549; shr.u32 %r2550, %r741, 12; cvt.u64.u32 %rd1922, %r2550; bfi.b64 %rd1923, %rd1921, %rd1922, 42, 22; shr.u32 %r2551, %r748, 12; mul.wide.u32 %rd1924, %r2551, 2097152; or.b64 %rd508, %rd1923, %rd1924; mov.u32 %r3599, 0; $L__BB3_649: cvt.u64.u32 %rd509, %r3599; mul.wide.u32 %rd1925, %r3599, 32; add.s64 %rd1926, %rd507, %rd1925; ld.u64 %rd1927, [%rd1926]; setp.eq.s64 %p407, %rd1927, %rd508; add.s32 %r3599, %r3599, 1; @%p407 bra $L__BB3_651; setp.lt.u32 %p408, %r3599, %r751; @%p408 bra $L__BB3_649; bra.uni $L__BB3_652; $L__BB3_682: ld.local.u64 %rd524, [%rd2+40]; mov.u64 %rd2390, 0; ld.u32 %r757, [%rd524+24]; setp.eq.s32 %p428, %r757, 0; @%p428 bra $L__BB3_687; add.s64 %rd525, %rd524, 64; shr.u32 %r2649, %r471, 12; cvt.u64.u32 %rd1988, %r2649; shr.u32 %r2650, %r733, 12; cvt.u64.u32 %rd1989, %r2650; bfi.b64 %rd1990, %rd1988, %rd1989, 42, 22; shr.u32 %r2651, %r748, 12; mul.wide.u32 %rd1991, %r2651, 2097152; or.b64 %rd526, %rd1990, %rd1991; mov.u32 %r3601, 0; $L__BB3_684: cvt.u64.u32 %rd527, %r3601; mul.wide.u32 %rd1992, %r3601, 32; add.s64 %rd1993, %rd525, %rd1992; ld.u64 %rd1994, [%rd1993]; setp.eq.s64 %p429, %rd1994, %rd526; add.s32 %r3601, %r3601, 1; @%p429 bra $L__BB3_686; setp.lt.u32 %p430, %r3601, %r757; @%p430 bra $L__BB3_684; bra.uni $L__BB3_687; $L__BB3_717: ld.local.u64 %rd542, [%rd2+40]; mov.u64 %rd2391, 0; ld.u32 %r763, [%rd542+24]; setp.eq.s32 %p450, %r763, 0; @%p450 bra $L__BB3_722; add.s64 %rd543, %rd542, 64; shr.u32 %r2748, %r635, 12; cvt.u64.u32 %rd2055, %r2748; shr.u32 %r2749, %r733, 12; cvt.u64.u32 %rd2056, %r2749; bfi.b64 %rd2057, %rd2055, %rd2056, 42, 22; shr.u32 %r2750, %r748, 12; mul.wide.u32 %rd2058, %r2750, 2097152; or.b64 %rd544, %rd2057, %rd2058; mov.u32 %r3603, 0; $L__BB3_719: cvt.u64.u32 %rd545, %r3603; mul.wide.u32 %rd2059, %r3603, 32; add.s64 %rd2060, %rd543, %rd2059; ld.u64 %rd2061, [%rd2060]; setp.eq.s64 %p451, %rd2061, %rd544; add.s32 %r3603, %r3603, 1; @%p451 bra $L__BB3_721; setp.lt.u32 %p452, %r3603, %r763; @%p452 bra $L__BB3_719; bra.uni $L__BB3_722; $L__BB3_752: ld.local.u64 %rd560, [%rd2+40]; mov.u64 %rd2392, 0; ld.u32 %r769, [%rd560+24]; setp.eq.s32 %p472, %r769, 0; @%p472 bra $L__BB3_757; add.s64 %rd561, %rd560, 64; shr.u32 %r2850, %r635, 12; cvt.u64.u32 %rd2122, %r2850; shr.u32 %r2851, %r741, 12; cvt.u64.u32 %rd2123, %r2851; bfi.b64 %rd2124, %rd2122, %rd2123, 42, 22; shr.u32 %r2852, %r748, 12; mul.wide.u32 %rd2125, %r2852, 2097152; or.b64 %rd562, %rd2124, %rd2125; mov.u32 %r3605, 0; $L__BB3_754: cvt.u64.u32 %rd563, %r3605; mul.wide.u32 %rd2126, %r3605, 32; add.s64 %rd2127, %rd561, %rd2126; ld.u64 %rd2128, [%rd2127]; setp.eq.s64 %p473, %rd2128, %rd562; add.s32 %r3605, %r3605, 1; @%p473 bra $L__BB3_756; setp.lt.u32 %p474, %r3605, %r769; @%p474 bra $L__BB3_754; bra.uni $L__BB3_757; $L__BB3_787: ld.local.u64 %rd578, [%rd2+40]; mov.u64 %rd2393, 0; ld.u32 %r775, [%rd578+24]; setp.eq.s32 %p494, %r775, 0; @%p494 bra $L__BB3_792; add.s64 %rd579, %rd578, 64; shr.u32 %r2953, %r635, 12; cvt.u64.u32 %rd2189, %r2953; shl.b64 %rd2190, %rd2189, 42; shr.u32 %r2954, %r732, 12; mul.wide.u32 %rd2191, %r2954, 2097152; or.b64 %rd2192, %rd2190, %rd2191; shr.u32 %r2955, %r741, 12; cvt.u64.u32 %rd2193, %r2955; or.b64 %rd580, %rd2192, %rd2193; mov.u32 %r3607, 0; $L__BB3_789: cvt.u64.u32 %rd581, %r3607; mul.wide.u32 %rd2194, %r3607, 32; add.s64 %rd2195, %rd579, %rd2194; ld.u64 %rd2196, [%rd2195]; setp.eq.s64 %p495, %rd2196, %rd580; add.s32 %r3607, %r3607, 1; @%p495 bra $L__BB3_791; setp.lt.u32 %p496, %r3607, %r775; @%p496 bra $L__BB3_789; bra.uni $L__BB3_792; $L__BB3_831: ld.f32 %f482, [%rd607]; bra.uni $L__BB3_833; $L__BB3_828: ld.f32 %f482, [%rd603]; bra.uni $L__BB3_833; $L__BB3_51: ld.f32 %f460, [%rd45]; bra.uni $L__BB3_66; $L__BB3_86: ld.f32 %f461, [%rd74]; bra.uni $L__BB3_101; $L__BB3_121: ld.f32 %f462, [%rd103]; bra.uni $L__BB3_136; $L__BB3_156: ld.f32 %f463, [%rd132]; bra.uni $L__BB3_171; $L__BB3_191: ld.f32 %f464, [%rd161]; bra.uni $L__BB3_206; $L__BB3_226: ld.f32 %f465, [%rd190]; bra.uni $L__BB3_241; $L__BB3_261: ld.f32 %f466, [%rd219]; bra.uni $L__BB3_276; $L__BB3_296: ld.f32 %f467, [%rd248]; bra.uni $L__BB3_311; $L__BB3_331: ld.f32 %f468, [%rd277]; bra.uni $L__BB3_346; $L__BB3_366: ld.f32 %f469, [%rd306]; bra.uni $L__BB3_381; $L__BB3_401: ld.f32 %f470, [%rd335]; bra.uni $L__BB3_416; $L__BB3_436: ld.f32 %f471, [%rd364]; bra.uni $L__BB3_451; $L__BB3_471: ld.f32 %f472, [%rd393]; bra.uni $L__BB3_486; $L__BB3_506: ld.f32 %f473, [%rd422]; bra.uni $L__BB3_521; $L__BB3_541: ld.f32 %f474, [%rd450]; bra.uni $L__BB3_556; $L__BB3_576: ld.f32 %f475, [%rd467]; bra.uni $L__BB3_591; $L__BB3_611: ld.f32 %f476, [%rd485]; bra.uni $L__BB3_626; $L__BB3_646: ld.f32 %f477, [%rd503]; bra.uni $L__BB3_661; $L__BB3_681: ld.f32 %f478, [%rd521]; bra.uni $L__BB3_696; $L__BB3_716: ld.f32 %f479, [%rd539]; bra.uni $L__BB3_731; $L__BB3_751: ld.f32 %f480, [%rd557]; bra.uni $L__BB3_766; $L__BB3_786: ld.f32 %f481, [%rd575]; bra.uni $L__BB3_801; $L__BB3_827: ld.f32 %f482, [%rd605]; $L__BB3_833: ld.param.f32 %f458, [_Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; add.ftz.f32 %f457, %f458, 0fBF000000; cvt.rmi.ftz.f32.f32 %f456, %f457; sub.ftz.f32 %f455, %f457, %f456; mul.ftz.f32 %f454, %f455, 0f40400000; mov.f32 %f453, 0f40400000; sub.ftz.f32 %f452, %f453, %f454; fma.rn.ftz.f32 %f451, %f455, %f452, 0f40400000; mul.ftz.f32 %f450, %f455, %f455; mul.ftz.f32 %f449, %f455, %f450; mul.ftz.f32 %f448, %f449, 0f3E2AAAAB; fma.rn.ftz.f32 %f447, %f455, %f451, 0f3F800000; fma.rn.ftz.f32 %f446, %f447, 0f3E2AAAAB, %f448; sub.ftz.f32 %f437, %f481, %f482; fma.rn.ftz.f32 %f438, %f220, %f437, %f482; sub.ftz.f32 %f439, %f295, %f438; fma.rn.ftz.f32 %f440, %f219, %f439, %f438; sub.ftz.f32 %f441, %f440, %f270; fma.rn.ftz.f32 %f442, %f116, %f441, %f270; fma.rn.ftz.f32 %f443, %f8, %f442, %f218; fma.rn.ftz.f32 %f444, %f10, %f443, %f217; fma.rn.ftz.f32 %f483, %f446, %f444, %f216; $L__BB3_834: st.param.f32 [func_retval0+0], %f483; ret; $L__BB3_50: ld.f32 %f460, [%rd47]; bra.uni $L__BB3_66; $L__BB3_85: ld.f32 %f461, [%rd76]; bra.uni $L__BB3_101; $L__BB3_120: ld.f32 %f462, [%rd105]; bra.uni $L__BB3_136; $L__BB3_155: and.b32 %r3272, %r259, -128; ld.f32 %f463, [%rd134]; bra.uni $L__BB3_171; $L__BB3_190: and.b32 %r3272, %r259, -128; ld.f32 %f464, [%rd163]; bra.uni $L__BB3_206; $L__BB3_225: ld.f32 %f465, [%rd192]; mov.u32 %r3272, %r372; bra.uni $L__BB3_241; $L__BB3_260: ld.f32 %f466, [%rd221]; mov.u32 %r3272, %r423; bra.uni $L__BB3_276; $L__BB3_295: ld.f32 %f467, [%rd250]; mov.u32 %r3272, %r473; bra.uni $L__BB3_311; $L__BB3_330: ld.f32 %f468, [%rd279]; mov.u32 %r3272, %r518; bra.uni $L__BB3_346; $L__BB3_365: ld.f32 %f469, [%rd308]; mov.u32 %r3272, %r560; bra.uni $L__BB3_381; $L__BB3_400: ld.f32 %f470, [%rd337]; mov.u32 %r3272, %r599; bra.uni $L__BB3_416; $L__BB3_435: ld.f32 %f471, [%rd366]; mov.u32 %r3272, %r637; bra.uni $L__BB3_451; $L__BB3_470: ld.f32 %f472, [%rd395]; mov.u32 %r3272, %r670; bra.uni $L__BB3_486; $L__BB3_505: ld.f32 %f473, [%rd424]; mov.u32 %r3272, %r700; bra.uni $L__BB3_521; $L__BB3_540: ld.f32 %f474, [%rd452]; bra.uni $L__BB3_556; $L__BB3_575: ld.f32 %f475, [%rd469]; bra.uni $L__BB3_591; $L__BB3_610: ld.f32 %f476, [%rd487]; bra.uni $L__BB3_626; $L__BB3_645: ld.f32 %f477, [%rd505]; bra.uni $L__BB3_661; $L__BB3_680: ld.f32 %f478, [%rd523]; bra.uni $L__BB3_696; $L__BB3_715: ld.f32 %f479, [%rd541]; bra.uni $L__BB3_731; $L__BB3_750: ld.f32 %f480, [%rd559]; bra.uni $L__BB3_766; $L__BB3_785: ld.f32 %f481, [%rd577]; bra.uni $L__BB3_801; $L__BB3_56: cvt.u64.u32 %rd2291, %r3181; shl.b64 %rd742, %rd2291, 5; add.s64 %rd2302, %rd17, %rd742; $L__BB3_57: setp.eq.s64 %p35, %rd2302, 0; @%p35 bra $L__BB3_65; add.s64 %rd55, %rd2302, 8; ld.u64 %rd56, [%rd2302+8]; setp.eq.s64 %p36, %rd56, 0; @%p36 bra $L__BB3_64; add.s64 %rd2315, %rd1, %rd56; and.b32 %r1052, %r8, -4096; mov.u64 %rd743, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1052}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1053, %r7, 3; and.b32 %r1054, %r1053, 31744; and.b32 %r1055, %r8, 3968; shr.u32 %r1056, %r1055, 2; or.b32 %r1057, %r1056, %r1054; bfe.u32 %r1058, %r74, 7, 5; or.b32 %r1059, %r1058, %r1057; shr.u32 %r1060, %r1057, 6; mul.wide.u32 %rd744, %r1060, 8; add.s64 %rd745, %rd2315, %rd744; and.b32 %r1061, %r1059, 63; shl.b64 %rd746, %rd743, %r1061; ld.u64 %rd747, [%rd745+4128]; and.b64 %rd748, %rd747, %rd746; setp.eq.s64 %p37, %rd748, 0; mul.wide.u32 %rd749, %r1059, 8; add.s64 %rd750, %rd2315, %rd749; add.s64 %rd58, %rd750, 8256; mov.b64 %rd751, {%r12, %r1052}; shr.u64 %rd752, %rd751, 32; cvt.u32.u64 %r3277, %rd752; @%p37 bra $L__BB3_63; and.b32 %r3157, %r7, -128; ld.u64 %rd753, [%rd58]; add.s64 %rd2311, %rd2315, %rd753; st.local.u32 [%rd2+12], %r3157; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1062, %r7, 5; and.b32 %r1063, %r1062, 3840; shl.b32 %r1064, %r8, 1; and.b32 %r1065, %r1064, 240; or.b32 %r1066, %r1065, %r1063; bfe.u32 %r1067, %r74, 3, 4; or.b32 %r1068, %r1067, %r1066; shr.u32 %r1069, %r1066, 6; mul.wide.u32 %rd754, %r1069, 8; add.s64 %rd755, %rd2311, %rd754; and.b32 %r1070, %r1068, 63; shl.b64 %rd757, %rd743, %r1070; ld.u64 %rd758, [%rd755+544]; and.b64 %rd759, %rd758, %rd757; setp.eq.s64 %p38, %rd759, 0; mul.wide.u32 %rd760, %r1068, 8; add.s64 %rd761, %rd2311, %rd760; add.s64 %rd60, %rd761, 1088; and.b32 %r3272, %r7, -128; @%p38 bra $L__BB3_62; and.b32 %r3229, %r7, -8; ld.u64 %rd762, [%rd60]; add.s64 %rd2307, %rd2311, %rd762; and.b32 %r1071, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1071}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1072, %r7, 6; and.b32 %r1073, %r1072, 448; shl.b32 %r1074, %r8, 3; and.b32 %r1075, %r1074, 56; or.b32 %r1076, %r1075, %r1073; and.b32 %r1077, %r74, 7; or.b32 %r1078, %r1076, %r1077; mul.wide.u32 %rd763, %r1078, 4; add.s64 %rd764, %rd2307, %rd763; ld.f32 %f460, [%rd764+96]; mov.b64 %rd765, {%r3229, %r1071}; shr.u64 %rd766, %rd765, 32; cvt.u32.u64 %r3184, %rd766; mov.u32 %r3232, %r12; bra.uni $L__BB3_66; $L__BB3_65: ld.f32 %f460, [%rd5+4]; bra.uni $L__BB3_66; $L__BB3_91: cvt.u64.u32 %rd2294, %r3230; shl.b64 %rd812, %rd2294, 5; add.s64 %rd2308, %rd17, %rd812; $L__BB3_92: setp.eq.s64 %p57, %rd2308, 0; @%p57 bra $L__BB3_100; add.s64 %rd84, %rd2308, 8; ld.u64 %rd85, [%rd2308+8]; setp.eq.s64 %p58, %rd85, 0; @%p58 bra $L__BB3_99; add.s64 %rd2315, %rd1, %rd85; and.b32 %r1138, %r139, -4096; mov.u64 %rd813, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1138}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1139, %r7, 3; and.b32 %r1140, %r1139, 31744; and.b32 %r1141, %r139, 3968; shr.u32 %r1142, %r1141, 2; or.b32 %r1143, %r1142, %r1140; bfe.u32 %r1144, %r74, 7, 5; or.b32 %r1145, %r1143, %r1144; shr.u32 %r1146, %r1143, 6; mul.wide.u32 %rd814, %r1146, 8; add.s64 %rd815, %rd2315, %rd814; and.b32 %r1147, %r1145, 63; shl.b64 %rd816, %rd813, %r1147; ld.u64 %rd817, [%rd815+4128]; and.b64 %rd818, %rd817, %rd816; setp.eq.s64 %p59, %rd818, 0; mul.wide.u32 %rd819, %r1145, 8; add.s64 %rd820, %rd2315, %rd819; add.s64 %rd87, %rd820, 8256; mov.b64 %rd821, {%r12, %r1138}; shr.u64 %rd822, %rd821, 32; cvt.u32.u64 %r3277, %rd822; @%p59 bra $L__BB3_98; and.b32 %r3131, %r7, -128; ld.u64 %rd823, [%rd87]; add.s64 %rd2311, %rd2315, %rd823; st.local.u32 [%rd2+12], %r3131; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1148, %r7, 5; and.b32 %r1149, %r1148, 3840; shl.b32 %r1150, %r139, 1; and.b32 %r1151, %r1150, 240; or.b32 %r1152, %r1151, %r1149; bfe.u32 %r1153, %r74, 3, 4; or.b32 %r1154, %r1152, %r1153; shr.u32 %r1155, %r1152, 6; mul.wide.u32 %rd824, %r1155, 8; add.s64 %rd825, %rd2311, %rd824; and.b32 %r1156, %r1154, 63; shl.b64 %rd827, %rd813, %r1156; ld.u64 %rd828, [%rd825+544]; and.b64 %rd829, %rd828, %rd827; setp.eq.s64 %p60, %rd829, 0; mul.wide.u32 %rd830, %r1154, 8; add.s64 %rd831, %rd2311, %rd830; add.s64 %rd89, %rd831, 1088; and.b32 %r3272, %r7, -128; @%p60 bra $L__BB3_97; and.b32 %r3229, %r7, -8; ld.u64 %rd832, [%rd89]; add.s64 %rd2307, %rd2311, %rd832; and.b32 %r1157, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1157}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1158, %r7, 6; and.b32 %r1159, %r1158, 448; shl.b32 %r1160, %r139, 3; and.b32 %r1161, %r1160, 56; or.b32 %r1162, %r1161, %r1159; and.b32 %r1163, %r74, 7; or.b32 %r1164, %r1162, %r1163; mul.wide.u32 %rd833, %r1164, 4; add.s64 %rd834, %rd2307, %rd833; ld.f32 %f461, [%rd834+96]; mov.b64 %rd835, {%r3229, %r1157}; shr.u64 %rd836, %rd835, 32; cvt.u32.u64 %r3184, %rd836; mov.u32 %r3232, %r12; bra.uni $L__BB3_101; $L__BB3_100: ld.f32 %f461, [%rd5+4]; bra.uni $L__BB3_101; $L__BB3_126: cvt.u64.u32 %rd2295, %r3276; shl.b64 %rd880, %rd2295, 5; add.s64 %rd2314, %rd17, %rd880; $L__BB3_127: setp.eq.s64 %p79, %rd2314, 0; @%p79 bra $L__BB3_135; add.s64 %rd113, %rd2314, 8; ld.u64 %rd114, [%rd2314+8]; setp.eq.s64 %p80, %rd114, 0; @%p80 bra $L__BB3_134; add.s64 %rd2315, %rd1, %rd114; and.b32 %r1220, %r139, -4096; mov.u64 %rd881, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1220}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1221, %r7, 3; and.b32 %r1222, %r1221, 31744; and.b32 %r1223, %r139, 3968; shr.u32 %r1224, %r1223, 2; or.b32 %r1225, %r1224, %r1222; bfe.u32 %r1226, %r9, 7, 5; or.b32 %r1227, %r1225, %r1226; shr.u32 %r1228, %r1225, 6; mul.wide.u32 %rd882, %r1228, 8; add.s64 %rd883, %rd2315, %rd882; and.b32 %r1229, %r1227, 63; shl.b64 %rd884, %rd881, %r1229; ld.u64 %rd885, [%rd883+4128]; and.b64 %rd886, %rd885, %rd884; setp.eq.s64 %p81, %rd886, 0; mul.wide.u32 %rd887, %r1227, 8; add.s64 %rd888, %rd2315, %rd887; add.s64 %rd116, %rd888, 8256; mov.b64 %rd889, {%r12, %r1220}; shr.u64 %rd890, %rd889, 32; cvt.u32.u64 %r3277, %rd890; @%p81 bra $L__BB3_133; and.b32 %r3104, %r7, -128; ld.u64 %rd891, [%rd116]; add.s64 %rd2311, %rd2315, %rd891; st.local.u32 [%rd2+12], %r3104; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1230, %r7, 5; and.b32 %r1231, %r1230, 3840; shl.b32 %r1232, %r139, 1; and.b32 %r1233, %r1232, 240; or.b32 %r1234, %r1233, %r1231; bfe.u32 %r1235, %r9, 3, 4; or.b32 %r1236, %r1234, %r1235; shr.u32 %r1237, %r1234, 6; mul.wide.u32 %rd892, %r1237, 8; add.s64 %rd893, %rd2311, %rd892; and.b32 %r1238, %r1236, 63; shl.b64 %rd895, %rd881, %r1238; ld.u64 %rd896, [%rd893+544]; and.b64 %rd897, %rd896, %rd895; setp.eq.s64 %p82, %rd897, 0; mul.wide.u32 %rd898, %r1236, 8; add.s64 %rd899, %rd2311, %rd898; add.s64 %rd118, %rd899, 1088; and.b32 %r3272, %r7, -128; @%p82 bra $L__BB3_132; and.b32 %r3229, %r7, -8; ld.u64 %rd900, [%rd118]; add.s64 %rd2307, %rd2311, %rd900; and.b32 %r1239, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1239}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1240, %r7, 7; shl.b32 %r1241, %r139, 3; and.b32 %r1242, %r1241, 56; and.b32 %r1243, %r9, 7; bfi.b32 %r1244, %r1240, %r1243, 6, 3; or.b32 %r1245, %r1244, %r1242; mul.wide.u32 %rd901, %r1245, 4; add.s64 %rd902, %rd2307, %rd901; ld.f32 %f462, [%rd902+96]; mov.b64 %rd903, {%r3229, %r1239}; shr.u64 %rd904, %rd903, 32; cvt.u32.u64 %r3184, %rd904; mov.u32 %r3232, %r12; bra.uni $L__BB3_136; $L__BB3_135: ld.f32 %f462, [%rd5+4]; bra.uni $L__BB3_136; $L__BB3_161: cvt.u64.u32 %rd2296, %r3319; shl.b64 %rd951, %rd2296, 5; add.s64 %rd2320, %rd17, %rd951; $L__BB3_162: setp.eq.s64 %p101, %rd2320, 0; @%p101 bra $L__BB3_170; add.s64 %rd142, %rd2320, 8; ld.u64 %rd143, [%rd2320+8]; setp.eq.s64 %p102, %rd143, 0; @%p102 bra $L__BB3_169; and.b32 %r3109, %r259, -4096; add.s64 %rd2315, %rd1, %rd143; and.b32 %r1302, %r139, -4096; mov.u64 %rd952, 1; st.local.v2.u32 [%rd2+24], {%r3109, %r1302}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1303, %r259, 3; and.b32 %r1304, %r1303, 31744; and.b32 %r1305, %r139, 3968; shr.u32 %r1306, %r1305, 2; or.b32 %r1307, %r1306, %r1304; bfe.u32 %r1308, %r9, 7, 5; or.b32 %r1309, %r1307, %r1308; shr.u32 %r1310, %r1307, 6; mul.wide.u32 %rd953, %r1310, 8; add.s64 %rd954, %rd2315, %rd953; and.b32 %r1311, %r1309, 63; shl.b64 %rd955, %rd952, %r1311; ld.u64 %rd956, [%rd954+4128]; and.b64 %rd957, %rd956, %rd955; setp.eq.s64 %p103, %rd957, 0; mul.wide.u32 %rd958, %r1309, 8; add.s64 %rd959, %rd2315, %rd958; add.s64 %rd145, %rd959, 8256; mov.b64 %rd960, {%r3109, %r1302}; shr.u64 %rd961, %rd960, 32; cvt.u32.u64 %r3277, %rd961; @%p103 bra $L__BB3_168; and.b32 %r3110, %r259, -128; ld.u64 %rd962, [%rd145]; add.s64 %rd2311, %rd2315, %rd962; st.local.u32 [%rd2+12], %r3110; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1312, %r259, 5; and.b32 %r1313, %r1312, 3840; shl.b32 %r1314, %r139, 1; and.b32 %r1315, %r1314, 240; or.b32 %r1316, %r1315, %r1313; bfe.u32 %r1317, %r9, 3, 4; or.b32 %r1318, %r1316, %r1317; shr.u32 %r1319, %r1316, 6; mul.wide.u32 %rd963, %r1319, 8; add.s64 %rd964, %rd2311, %rd963; and.b32 %r1320, %r1318, 63; shl.b64 %rd966, %rd952, %r1320; ld.u64 %rd967, [%rd964+544]; and.b64 %rd968, %rd967, %rd966; setp.eq.s64 %p104, %rd968, 0; mul.wide.u32 %rd969, %r1318, 8; add.s64 %rd970, %rd2311, %rd969; add.s64 %rd147, %rd970, 1088; @%p104 bra $L__BB3_167; and.b32 %r3229, %r259, -8; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.u64 %rd971, [%rd147]; add.s64 %rd2307, %rd2311, %rd971; and.b32 %r1321, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1321}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1322, %r259, 7; shl.b32 %r1323, %r139, 3; and.b32 %r1324, %r1323, 56; and.b32 %r1325, %r9, 7; bfi.b32 %r1326, %r1322, %r1325, 6, 3; or.b32 %r1327, %r1326, %r1324; mul.wide.u32 %rd972, %r1327, 4; add.s64 %rd973, %rd2307, %rd972; ld.f32 %f463, [%rd973+96]; mov.b64 %rd974, {%r3229, %r1321}; shr.u64 %rd975, %rd974, 32; cvt.u32.u64 %r3184, %rd975; bra.uni $L__BB3_171; $L__BB3_170: ld.f32 %f463, [%rd5+4]; bra.uni $L__BB3_171; $L__BB3_196: shl.b64 %rd1022, %rd168, 5; add.s64 %rd2326, %rd17, %rd1022; $L__BB3_197: setp.eq.s64 %p123, %rd2326, 0; @%p123 bra $L__BB3_205; add.s64 %rd171, %rd2326, 8; ld.u64 %rd172, [%rd2326+8]; setp.eq.s64 %p124, %rd172, 0; @%p124 bra $L__BB3_204; and.b32 %r3116, %r259, -4096; add.s64 %rd2315, %rd1, %rd172; and.b32 %r1388, %r139, -4096; mov.u64 %rd1023, 1; st.local.v2.u32 [%rd2+24], {%r3116, %r1388}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1389, %r259, 3; and.b32 %r1390, %r1389, 31744; and.b32 %r1391, %r139, 3968; shr.u32 %r1392, %r1391, 2; or.b32 %r1393, %r1392, %r1390; bfe.u32 %r1394, %r74, 7, 5; or.b32 %r1395, %r1393, %r1394; shr.u32 %r1396, %r1393, 6; mul.wide.u32 %rd1024, %r1396, 8; add.s64 %rd1025, %rd2315, %rd1024; and.b32 %r1397, %r1395, 63; shl.b64 %rd1026, %rd1023, %r1397; ld.u64 %rd1027, [%rd1025+4128]; and.b64 %rd1028, %rd1027, %rd1026; setp.eq.s64 %p125, %rd1028, 0; mul.wide.u32 %rd1029, %r1395, 8; add.s64 %rd1030, %rd2315, %rd1029; add.s64 %rd174, %rd1030, 8256; mov.b64 %rd1031, {%r3116, %r1388}; shr.u64 %rd1032, %rd1031, 32; cvt.u32.u64 %r3277, %rd1032; @%p125 bra $L__BB3_203; and.b32 %r3117, %r259, -128; ld.u64 %rd1033, [%rd174]; add.s64 %rd2311, %rd2315, %rd1033; st.local.u32 [%rd2+12], %r3117; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1398, %r259, 5; and.b32 %r1399, %r1398, 3840; shl.b32 %r1400, %r139, 1; and.b32 %r1401, %r1400, 240; or.b32 %r1402, %r1401, %r1399; bfe.u32 %r1403, %r74, 3, 4; or.b32 %r1404, %r1402, %r1403; shr.u32 %r1405, %r1402, 6; mul.wide.u32 %rd1034, %r1405, 8; add.s64 %rd1035, %rd2311, %rd1034; and.b32 %r1406, %r1404, 63; shl.b64 %rd1037, %rd1023, %r1406; ld.u64 %rd1038, [%rd1035+544]; and.b64 %rd1039, %rd1038, %rd1037; setp.eq.s64 %p126, %rd1039, 0; mul.wide.u32 %rd1040, %r1404, 8; add.s64 %rd1041, %rd2311, %rd1040; add.s64 %rd176, %rd1041, 1088; @%p126 bra $L__BB3_202; and.b32 %r3229, %r259, -8; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.u64 %rd1042, [%rd176]; add.s64 %rd2307, %rd2311, %rd1042; and.b32 %r1407, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1407}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1408, %r259, 6; and.b32 %r1409, %r1408, 448; shl.b32 %r1410, %r139, 3; and.b32 %r1411, %r1410, 56; or.b32 %r1412, %r1411, %r1409; and.b32 %r1413, %r74, 7; or.b32 %r1414, %r1412, %r1413; mul.wide.u32 %rd1043, %r1414, 4; add.s64 %rd1044, %rd2307, %rd1043; ld.f32 %f464, [%rd1044+96]; mov.b64 %rd1045, {%r3229, %r1407}; shr.u64 %rd1046, %rd1045, 32; cvt.u32.u64 %r3184, %rd1046; bra.uni $L__BB3_206; $L__BB3_205: ld.f32 %f464, [%rd5+4]; bra.uni $L__BB3_206; $L__BB3_231: shl.b64 %rd1093, %rd197, 5; add.s64 %rd2332, %rd17, %rd1093; $L__BB3_232: setp.eq.s64 %p145, %rd2332, 0; @%p145 bra $L__BB3_240; add.s64 %rd200, %rd2332, 8; ld.u64 %rd201, [%rd2332+8]; setp.eq.s64 %p146, %rd201, 0; @%p146 bra $L__BB3_239; and.b32 %r3140, %r259, -4096; add.s64 %rd2315, %rd1, %rd201; and.b32 %r1474, %r8, -4096; mov.u64 %rd1094, 1; st.local.v2.u32 [%rd2+24], {%r3140, %r1474}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1475, %r259, 3; and.b32 %r1476, %r1475, 31744; and.b32 %r1477, %r8, 3968; shr.u32 %r1478, %r1477, 2; or.b32 %r1479, %r1476, %r1478; bfe.u32 %r1480, %r74, 7, 5; or.b32 %r1481, %r1480, %r1479; shr.u32 %r1482, %r1479, 6; mul.wide.u32 %rd1095, %r1482, 8; add.s64 %rd1096, %rd2315, %rd1095; and.b32 %r1483, %r1481, 63; shl.b64 %rd1097, %rd1094, %r1483; ld.u64 %rd1098, [%rd1096+4128]; and.b64 %rd1099, %rd1098, %rd1097; setp.eq.s64 %p147, %rd1099, 0; mul.wide.u32 %rd1100, %r1481, 8; add.s64 %rd1101, %rd2315, %rd1100; add.s64 %rd203, %rd1101, 8256; mov.b64 %rd1102, {%r3140, %r1474}; shr.u64 %rd1103, %rd1102, 32; cvt.u32.u64 %r3277, %rd1103; @%p147 bra $L__BB3_238; and.b32 %r3141, %r259, -128; ld.u64 %rd1104, [%rd203]; add.s64 %rd2311, %rd2315, %rd1104; st.local.u32 [%rd2+12], %r3141; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1484, %r259, 5; and.b32 %r1485, %r1484, 3840; shl.b32 %r1486, %r8, 1; and.b32 %r1487, %r1486, 240; or.b32 %r1488, %r1485, %r1487; bfe.u32 %r1489, %r74, 3, 4; or.b32 %r1490, %r1489, %r1488; shr.u32 %r1491, %r1488, 6; mul.wide.u32 %rd1105, %r1491, 8; add.s64 %rd1106, %rd2311, %rd1105; and.b32 %r1492, %r1490, 63; shl.b64 %rd1108, %rd1094, %r1492; ld.u64 %rd1109, [%rd1106+544]; and.b64 %rd1110, %rd1109, %rd1108; setp.eq.s64 %p148, %rd1110, 0; mul.wide.u32 %rd1111, %r1490, 8; add.s64 %rd1112, %rd2311, %rd1111; add.s64 %rd205, %rd1112, 1088; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; @%p148 bra $L__BB3_237; and.b32 %r3229, %r259, -8; ld.u64 %rd1113, [%rd205]; add.s64 %rd2307, %rd2311, %rd1113; and.b32 %r1493, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1493}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1494, %r259, 6; and.b32 %r1495, %r1494, 448; shl.b32 %r1496, %r8, 3; and.b32 %r1497, %r1496, 56; or.b32 %r1498, %r1495, %r1497; and.b32 %r1499, %r74, 7; or.b32 %r1500, %r1498, %r1499; mul.wide.u32 %rd1114, %r1500, 4; add.s64 %rd1115, %rd2307, %rd1114; ld.f32 %f465, [%rd1115+96]; mov.b64 %rd1116, {%r3229, %r1493}; shr.u64 %rd1117, %rd1116, 32; cvt.u32.u64 %r3184, %rd1117; bra.uni $L__BB3_241; $L__BB3_240: ld.f32 %f465, [%rd5+4]; bra.uni $L__BB3_241; $L__BB3_266: shl.b64 %rd1163, %rd226, 5; add.s64 %rd2338, %rd17, %rd1163; $L__BB3_267: setp.eq.s64 %p167, %rd2338, 0; @%p167 bra $L__BB3_275; add.s64 %rd229, %rd2338, 8; ld.u64 %rd230, [%rd2338+8]; setp.eq.s64 %p168, %rd230, 0; @%p168 bra $L__BB3_274; add.s64 %rd2315, %rd1, %rd230; and.b32 %r1559, %r8, -4096; mov.u64 %rd1164, 1; st.local.v2.u32 [%rd2+24], {%r426, %r1559}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1560, %r259, 3; and.b32 %r1561, %r1560, 31744; and.b32 %r1562, %r8, 3968; shr.u32 %r1563, %r1562, 2; or.b32 %r1564, %r1561, %r1563; bfe.u32 %r1565, %r9, 7, 5; or.b32 %r1566, %r1564, %r1565; shr.u32 %r1567, %r1564, 6; mul.wide.u32 %rd1165, %r1567, 8; add.s64 %rd1166, %rd2315, %rd1165; and.b32 %r1568, %r1566, 63; shl.b64 %rd1167, %rd1164, %r1568; ld.u64 %rd1168, [%rd1166+4128]; and.b64 %rd1169, %rd1168, %rd1167; setp.eq.s64 %p169, %rd1169, 0; mul.wide.u32 %rd1170, %r1566, 8; add.s64 %rd1171, %rd2315, %rd1170; add.s64 %rd232, %rd1171, 8256; mov.b64 %rd1172, {%r426, %r1559}; shr.u64 %rd1173, %rd1172, 32; cvt.u32.u64 %r3277, %rd1173; @%p169 bra $L__BB3_273; ld.u64 %rd1174, [%rd232]; add.s64 %rd2311, %rd2315, %rd1174; st.local.u32 [%rd2+12], %r423; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1569, %r259, 5; and.b32 %r1570, %r1569, 3840; shl.b32 %r1571, %r8, 1; and.b32 %r1572, %r1571, 240; or.b32 %r1573, %r1570, %r1572; bfe.u32 %r1574, %r9, 3, 4; or.b32 %r1575, %r1573, %r1574; shr.u32 %r1576, %r1573, 6; mul.wide.u32 %rd1175, %r1576, 8; add.s64 %rd1176, %rd2311, %rd1175; and.b32 %r1577, %r1575, 63; shl.b64 %rd1178, %rd1164, %r1577; ld.u64 %rd1179, [%rd1176+544]; and.b64 %rd1180, %rd1179, %rd1178; setp.eq.s64 %p170, %rd1180, 0; mul.wide.u32 %rd1181, %r1575, 8; add.s64 %rd1182, %rd2311, %rd1181; add.s64 %rd234, %rd1182, 1088; @%p170 bra $L__BB3_272; and.b32 %r3229, %r259, -8; ld.u64 %rd1183, [%rd234]; add.s64 %rd2307, %rd2311, %rd1183; and.b32 %r1578, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1578}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1579, %r259, 6; and.b32 %r1580, %r1579, 448; shl.b32 %r1581, %r8, 3; and.b32 %r1582, %r1581, 56; or.b32 %r1583, %r1580, %r1582; and.b32 %r1584, %r9, 7; or.b32 %r1585, %r1583, %r1584; mul.wide.u32 %rd1184, %r1585, 4; add.s64 %rd1185, %rd2307, %rd1184; ld.f32 %f466, [%rd1185+96]; mov.b64 %rd1186, {%r3229, %r1578}; shr.u64 %rd1187, %rd1186, 32; cvt.u32.u64 %r3184, %rd1187; mov.u32 %r3232, %r426; mov.u32 %r3272, %r423; bra.uni $L__BB3_276; $L__BB3_275: ld.f32 %f466, [%rd5+4]; bra.uni $L__BB3_276; $L__BB3_301: shl.b64 %rd1233, %rd255, 5; add.s64 %rd2344, %rd17, %rd1233; $L__BB3_302: setp.eq.s64 %p189, %rd2344, 0; @%p189 bra $L__BB3_310; add.s64 %rd258, %rd2344, 8; ld.u64 %rd259, [%rd2344+8]; setp.eq.s64 %p190, %rd259, 0; @%p190 bra $L__BB3_309; add.s64 %rd2315, %rd1, %rd259; and.b32 %r1644, %r8, -4096; mov.u64 %rd1234, 1; st.local.v2.u32 [%rd2+24], {%r476, %r1644}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1645, %r471, 3; and.b32 %r1646, %r1645, 31744; and.b32 %r1647, %r8, 3968; shr.u32 %r1648, %r1647, 2; or.b32 %r1649, %r1648, %r1646; bfe.u32 %r1650, %r9, 7, 5; or.b32 %r1651, %r1649, %r1650; shr.u32 %r1652, %r1649, 6; mul.wide.u32 %rd1235, %r1652, 8; add.s64 %rd1236, %rd2315, %rd1235; and.b32 %r1653, %r1651, 63; shl.b64 %rd1237, %rd1234, %r1653; ld.u64 %rd1238, [%rd1236+4128]; and.b64 %rd1239, %rd1238, %rd1237; setp.eq.s64 %p191, %rd1239, 0; mul.wide.u32 %rd1240, %r1651, 8; add.s64 %rd1241, %rd2315, %rd1240; add.s64 %rd261, %rd1241, 8256; mov.b64 %rd1242, {%r476, %r1644}; shr.u64 %rd1243, %rd1242, 32; cvt.u32.u64 %r3277, %rd1243; @%p191 bra $L__BB3_308; ld.u64 %rd1244, [%rd261]; add.s64 %rd2311, %rd2315, %rd1244; st.local.u32 [%rd2+12], %r473; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1654, %r471, 5; and.b32 %r1655, %r1654, 3840; shl.b32 %r1656, %r8, 1; and.b32 %r1657, %r1656, 240; or.b32 %r1658, %r1657, %r1655; bfe.u32 %r1659, %r9, 3, 4; or.b32 %r1660, %r1658, %r1659; shr.u32 %r1661, %r1658, 6; mul.wide.u32 %rd1245, %r1661, 8; add.s64 %rd1246, %rd2311, %rd1245; and.b32 %r1662, %r1660, 63; shl.b64 %rd1248, %rd1234, %r1662; ld.u64 %rd1249, [%rd1246+544]; and.b64 %rd1250, %rd1249, %rd1248; setp.eq.s64 %p192, %rd1250, 0; mul.wide.u32 %rd1251, %r1660, 8; add.s64 %rd1252, %rd2311, %rd1251; add.s64 %rd263, %rd1252, 1088; @%p192 bra $L__BB3_307; ld.u64 %rd1253, [%rd263]; add.s64 %rd2307, %rd2311, %rd1253; and.b32 %r1663, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1663}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1664, %r471, 6; and.b32 %r1665, %r1664, 448; shl.b32 %r1666, %r8, 3; and.b32 %r1667, %r1666, 56; or.b32 %r1668, %r1667, %r1665; and.b32 %r1669, %r9, 7; or.b32 %r1670, %r1668, %r1669; mul.wide.u32 %rd1254, %r1670, 4; add.s64 %rd1255, %rd2307, %rd1254; ld.f32 %f467, [%rd1255+96]; mov.b64 %rd1256, {%r472, %r1663}; shr.u64 %rd1257, %rd1256, 32; cvt.u32.u64 %r3184, %rd1257; mov.u32 %r3232, %r476; mov.u32 %r3272, %r473; mov.u32 %r3229, %r472; bra.uni $L__BB3_311; $L__BB3_310: ld.f32 %f467, [%rd5+4]; bra.uni $L__BB3_311; $L__BB3_336: shl.b64 %rd1304, %rd284, 5; add.s64 %rd2350, %rd17, %rd1304; $L__BB3_337: setp.eq.s64 %p211, %rd2350, 0; @%p211 bra $L__BB3_345; add.s64 %rd287, %rd2350, 8; ld.u64 %rd288, [%rd2350+8]; setp.eq.s64 %p212, %rd288, 0; @%p212 bra $L__BB3_344; add.s64 %rd2315, %rd1, %rd288; and.b32 %r1730, %r8, -4096; mov.u64 %rd1305, 1; st.local.v2.u32 [%rd2+24], {%r521, %r1730}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1731, %r471, 3; and.b32 %r1732, %r1731, 31744; and.b32 %r1733, %r8, 3968; shr.u32 %r1734, %r1733, 2; or.b32 %r1735, %r1734, %r1732; bfe.u32 %r1736, %r74, 7, 5; or.b32 %r1737, %r1736, %r1735; shr.u32 %r1738, %r1735, 6; mul.wide.u32 %rd1306, %r1738, 8; add.s64 %rd1307, %rd2315, %rd1306; and.b32 %r1739, %r1737, 63; shl.b64 %rd1308, %rd1305, %r1739; ld.u64 %rd1309, [%rd1307+4128]; and.b64 %rd1310, %rd1309, %rd1308; setp.eq.s64 %p213, %rd1310, 0; mul.wide.u32 %rd1311, %r1737, 8; add.s64 %rd1312, %rd2315, %rd1311; add.s64 %rd290, %rd1312, 8256; mov.b64 %rd1313, {%r521, %r1730}; shr.u64 %rd1314, %rd1313, 32; cvt.u32.u64 %r3277, %rd1314; @%p213 bra $L__BB3_343; ld.u64 %rd1315, [%rd290]; add.s64 %rd2311, %rd2315, %rd1315; st.local.u32 [%rd2+12], %r518; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1740, %r471, 5; and.b32 %r1741, %r1740, 3840; shl.b32 %r1742, %r8, 1; and.b32 %r1743, %r1742, 240; or.b32 %r1744, %r1743, %r1741; bfe.u32 %r1745, %r74, 3, 4; or.b32 %r1746, %r1745, %r1744; shr.u32 %r1747, %r1744, 6; mul.wide.u32 %rd1316, %r1747, 8; add.s64 %rd1317, %rd2311, %rd1316; and.b32 %r1748, %r1746, 63; shl.b64 %rd1319, %rd1305, %r1748; ld.u64 %rd1320, [%rd1317+544]; and.b64 %rd1321, %rd1320, %rd1319; setp.eq.s64 %p214, %rd1321, 0; mul.wide.u32 %rd1322, %r1746, 8; add.s64 %rd1323, %rd2311, %rd1322; add.s64 %rd292, %rd1323, 1088; @%p214 bra $L__BB3_342; ld.u64 %rd1324, [%rd292]; add.s64 %rd2307, %rd2311, %rd1324; and.b32 %r1749, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1749}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1750, %r471, 6; and.b32 %r1751, %r1750, 448; shl.b32 %r1752, %r8, 3; and.b32 %r1753, %r1752, 56; or.b32 %r1754, %r1753, %r1751; and.b32 %r1755, %r74, 7; or.b32 %r1756, %r1754, %r1755; mul.wide.u32 %rd1325, %r1756, 4; add.s64 %rd1326, %rd2307, %rd1325; ld.f32 %f468, [%rd1326+96]; mov.b64 %rd1327, {%r472, %r1749}; shr.u64 %rd1328, %rd1327, 32; cvt.u32.u64 %r3184, %rd1328; mov.u32 %r3232, %r521; mov.u32 %r3272, %r518; mov.u32 %r3229, %r472; bra.uni $L__BB3_346; $L__BB3_345: ld.f32 %f468, [%rd5+4]; bra.uni $L__BB3_346; $L__BB3_371: shl.b64 %rd1375, %rd313, 5; add.s64 %rd2356, %rd17, %rd1375; $L__BB3_372: setp.eq.s64 %p233, %rd2356, 0; @%p233 bra $L__BB3_380; add.s64 %rd316, %rd2356, 8; ld.u64 %rd317, [%rd2356+8]; setp.eq.s64 %p234, %rd317, 0; @%p234 bra $L__BB3_379; add.s64 %rd2315, %rd1, %rd317; and.b32 %r1817, %r139, -4096; mov.u64 %rd1376, 1; st.local.v2.u32 [%rd2+24], {%r563, %r1817}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1818, %r471, 3; and.b32 %r1819, %r1818, 31744; and.b32 %r1820, %r139, 3968; shr.u32 %r1821, %r1820, 2; or.b32 %r1822, %r1821, %r1819; bfe.u32 %r1823, %r74, 7, 5; or.b32 %r1824, %r1822, %r1823; shr.u32 %r1825, %r1822, 6; mul.wide.u32 %rd1377, %r1825, 8; add.s64 %rd1378, %rd2315, %rd1377; and.b32 %r1826, %r1824, 63; shl.b64 %rd1379, %rd1376, %r1826; ld.u64 %rd1380, [%rd1378+4128]; and.b64 %rd1381, %rd1380, %rd1379; setp.eq.s64 %p235, %rd1381, 0; mul.wide.u32 %rd1382, %r1824, 8; add.s64 %rd1383, %rd2315, %rd1382; add.s64 %rd319, %rd1383, 8256; mov.b64 %rd1384, {%r563, %r1817}; shr.u64 %rd1385, %rd1384, 32; cvt.u32.u64 %r3277, %rd1385; @%p235 bra $L__BB3_378; ld.u64 %rd1386, [%rd319]; add.s64 %rd2311, %rd2315, %rd1386; st.local.u32 [%rd2+12], %r560; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1827, %r471, 5; and.b32 %r1828, %r1827, 3840; shl.b32 %r1829, %r139, 1; and.b32 %r1830, %r1829, 240; or.b32 %r1831, %r1830, %r1828; bfe.u32 %r1832, %r74, 3, 4; or.b32 %r1833, %r1831, %r1832; shr.u32 %r1834, %r1831, 6; mul.wide.u32 %rd1387, %r1834, 8; add.s64 %rd1388, %rd2311, %rd1387; and.b32 %r1835, %r1833, 63; shl.b64 %rd1390, %rd1376, %r1835; ld.u64 %rd1391, [%rd1388+544]; and.b64 %rd1392, %rd1391, %rd1390; setp.eq.s64 %p236, %rd1392, 0; mul.wide.u32 %rd1393, %r1833, 8; add.s64 %rd1394, %rd2311, %rd1393; add.s64 %rd321, %rd1394, 1088; @%p236 bra $L__BB3_377; ld.u64 %rd1395, [%rd321]; add.s64 %rd2307, %rd2311, %rd1395; and.b32 %r1836, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1836}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1837, %r471, 6; and.b32 %r1838, %r1837, 448; shl.b32 %r1839, %r139, 3; and.b32 %r1840, %r1839, 56; or.b32 %r1841, %r1840, %r1838; and.b32 %r1842, %r74, 7; or.b32 %r1843, %r1841, %r1842; mul.wide.u32 %rd1396, %r1843, 4; add.s64 %rd1397, %rd2307, %rd1396; ld.f32 %f469, [%rd1397+96]; mov.b64 %rd1398, {%r472, %r1836}; shr.u64 %rd1399, %rd1398, 32; cvt.u32.u64 %r3184, %rd1399; mov.u32 %r3232, %r563; mov.u32 %r3272, %r560; mov.u32 %r3229, %r472; bra.uni $L__BB3_381; $L__BB3_380: ld.f32 %f469, [%rd5+4]; bra.uni $L__BB3_381; $L__BB3_406: shl.b64 %rd1446, %rd342, 5; add.s64 %rd2362, %rd17, %rd1446; $L__BB3_407: setp.eq.s64 %p255, %rd2362, 0; @%p255 bra $L__BB3_415; add.s64 %rd345, %rd2362, 8; ld.u64 %rd346, [%rd2362+8]; setp.eq.s64 %p256, %rd346, 0; @%p256 bra $L__BB3_414; add.s64 %rd2315, %rd1, %rd346; and.b32 %r1900, %r139, -4096; mov.u64 %rd1447, 1; st.local.v2.u32 [%rd2+24], {%r602, %r1900}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1901, %r471, 3; and.b32 %r1902, %r1901, 31744; and.b32 %r1903, %r139, 3968; shr.u32 %r1904, %r1903, 2; or.b32 %r1905, %r1904, %r1902; bfe.u32 %r1906, %r9, 7, 5; or.b32 %r1907, %r1905, %r1906; shr.u32 %r1908, %r1905, 6; mul.wide.u32 %rd1448, %r1908, 8; add.s64 %rd1449, %rd2315, %rd1448; and.b32 %r1909, %r1907, 63; shl.b64 %rd1450, %rd1447, %r1909; ld.u64 %rd1451, [%rd1449+4128]; and.b64 %rd1452, %rd1451, %rd1450; setp.eq.s64 %p257, %rd1452, 0; mul.wide.u32 %rd1453, %r1907, 8; add.s64 %rd1454, %rd2315, %rd1453; add.s64 %rd348, %rd1454, 8256; mov.b64 %rd1455, {%r602, %r1900}; shr.u64 %rd1456, %rd1455, 32; cvt.u32.u64 %r3277, %rd1456; @%p257 bra $L__BB3_413; ld.u64 %rd1457, [%rd348]; add.s64 %rd2311, %rd2315, %rd1457; st.local.u32 [%rd2+12], %r599; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1910, %r471, 5; and.b32 %r1911, %r1910, 3840; shl.b32 %r1912, %r139, 1; and.b32 %r1913, %r1912, 240; or.b32 %r1914, %r1913, %r1911; bfe.u32 %r1915, %r9, 3, 4; or.b32 %r1916, %r1914, %r1915; shr.u32 %r1917, %r1914, 6; mul.wide.u32 %rd1458, %r1917, 8; add.s64 %rd1459, %rd2311, %rd1458; and.b32 %r1918, %r1916, 63; shl.b64 %rd1461, %rd1447, %r1918; ld.u64 %rd1462, [%rd1459+544]; and.b64 %rd1463, %rd1462, %rd1461; setp.eq.s64 %p258, %rd1463, 0; mul.wide.u32 %rd1464, %r1916, 8; add.s64 %rd1465, %rd2311, %rd1464; add.s64 %rd350, %rd1465, 1088; @%p258 bra $L__BB3_412; ld.u64 %rd1466, [%rd350]; add.s64 %rd2307, %rd2311, %rd1466; and.b32 %r1919, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1919}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1920, %r471, 7; shl.b32 %r1921, %r139, 3; and.b32 %r1922, %r1921, 56; and.b32 %r1923, %r9, 7; bfi.b32 %r1924, %r1920, %r1923, 6, 3; or.b32 %r1925, %r1924, %r1922; mul.wide.u32 %rd1467, %r1925, 4; add.s64 %rd1468, %rd2307, %rd1467; ld.f32 %f470, [%rd1468+96]; mov.b64 %rd1469, {%r472, %r1919}; shr.u64 %rd1470, %rd1469, 32; cvt.u32.u64 %r3184, %rd1470; mov.u32 %r3232, %r602; mov.u32 %r3272, %r599; mov.u32 %r3229, %r472; bra.uni $L__BB3_416; $L__BB3_415: ld.f32 %f470, [%rd5+4]; bra.uni $L__BB3_416; $L__BB3_441: shl.b64 %rd1517, %rd371, 5; add.s64 %rd2368, %rd17, %rd1517; $L__BB3_442: setp.eq.s64 %p277, %rd2368, 0; @%p277 bra $L__BB3_450; add.s64 %rd374, %rd2368, 8; ld.u64 %rd375, [%rd2368+8]; setp.eq.s64 %p278, %rd375, 0; @%p278 bra $L__BB3_449; add.s64 %rd2315, %rd1, %rd375; and.b32 %r1982, %r139, -4096; mov.u64 %rd1518, 1; st.local.v2.u32 [%rd2+24], {%r640, %r1982}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1983, %r635, 3; and.b32 %r1984, %r1983, 31744; and.b32 %r1985, %r139, 3968; shr.u32 %r1986, %r1985, 2; or.b32 %r1987, %r1986, %r1984; bfe.u32 %r1988, %r9, 7, 5; or.b32 %r1989, %r1987, %r1988; shr.u32 %r1990, %r1987, 6; mul.wide.u32 %rd1519, %r1990, 8; add.s64 %rd1520, %rd2315, %rd1519; and.b32 %r1991, %r1989, 63; shl.b64 %rd1521, %rd1518, %r1991; ld.u64 %rd1522, [%rd1520+4128]; and.b64 %rd1523, %rd1522, %rd1521; setp.eq.s64 %p279, %rd1523, 0; mul.wide.u32 %rd1524, %r1989, 8; add.s64 %rd1525, %rd2315, %rd1524; add.s64 %rd377, %rd1525, 8256; mov.b64 %rd1526, {%r640, %r1982}; shr.u64 %rd1527, %rd1526, 32; cvt.u32.u64 %r3277, %rd1527; @%p279 bra $L__BB3_448; ld.u64 %rd1528, [%rd377]; add.s64 %rd2311, %rd2315, %rd1528; st.local.u32 [%rd2+12], %r637; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1992, %r635, 5; and.b32 %r1993, %r1992, 3840; shl.b32 %r1994, %r139, 1; and.b32 %r1995, %r1994, 240; or.b32 %r1996, %r1995, %r1993; bfe.u32 %r1997, %r9, 3, 4; or.b32 %r1998, %r1996, %r1997; shr.u32 %r1999, %r1996, 6; mul.wide.u32 %rd1529, %r1999, 8; add.s64 %rd1530, %rd2311, %rd1529; and.b32 %r2000, %r1998, 63; shl.b64 %rd1532, %rd1518, %r2000; ld.u64 %rd1533, [%rd1530+544]; and.b64 %rd1534, %rd1533, %rd1532; setp.eq.s64 %p280, %rd1534, 0; mul.wide.u32 %rd1535, %r1998, 8; add.s64 %rd1536, %rd2311, %rd1535; add.s64 %rd379, %rd1536, 1088; @%p280 bra $L__BB3_447; ld.u64 %rd1537, [%rd379]; add.s64 %rd2307, %rd2311, %rd1537; and.b32 %r2001, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2001}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2002, %r635, 7; shl.b32 %r2003, %r139, 3; and.b32 %r2004, %r2003, 56; and.b32 %r2005, %r9, 7; bfi.b32 %r2006, %r2002, %r2005, 6, 3; or.b32 %r2007, %r2006, %r2004; mul.wide.u32 %rd1538, %r2007, 4; add.s64 %rd1539, %rd2307, %rd1538; ld.f32 %f471, [%rd1539+96]; mov.b64 %rd1540, {%r636, %r2001}; shr.u64 %rd1541, %rd1540, 32; cvt.u32.u64 %r3184, %rd1541; mov.u32 %r3232, %r640; mov.u32 %r3272, %r637; mov.u32 %r3229, %r636; bra.uni $L__BB3_451; $L__BB3_450: ld.f32 %f471, [%rd5+4]; bra.uni $L__BB3_451; $L__BB3_476: shl.b64 %rd1588, %rd400, 5; add.s64 %rd2374, %rd17, %rd1588; $L__BB3_477: setp.eq.s64 %p299, %rd2374, 0; @%p299 bra $L__BB3_485; add.s64 %rd403, %rd2374, 8; ld.u64 %rd404, [%rd2374+8]; setp.eq.s64 %p300, %rd404, 0; @%p300 bra $L__BB3_484; add.s64 %rd2315, %rd1, %rd404; and.b32 %r2068, %r139, -4096; mov.u64 %rd1589, 1; st.local.v2.u32 [%rd2+24], {%r673, %r2068}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2069, %r635, 3; and.b32 %r2070, %r2069, 31744; and.b32 %r2071, %r139, 3968; shr.u32 %r2072, %r2071, 2; or.b32 %r2073, %r2072, %r2070; bfe.u32 %r2074, %r74, 7, 5; or.b32 %r2075, %r2073, %r2074; shr.u32 %r2076, %r2073, 6; mul.wide.u32 %rd1590, %r2076, 8; add.s64 %rd1591, %rd2315, %rd1590; and.b32 %r2077, %r2075, 63; shl.b64 %rd1592, %rd1589, %r2077; ld.u64 %rd1593, [%rd1591+4128]; and.b64 %rd1594, %rd1593, %rd1592; setp.eq.s64 %p301, %rd1594, 0; mul.wide.u32 %rd1595, %r2075, 8; add.s64 %rd1596, %rd2315, %rd1595; add.s64 %rd406, %rd1596, 8256; mov.b64 %rd1597, {%r673, %r2068}; shr.u64 %rd1598, %rd1597, 32; cvt.u32.u64 %r3277, %rd1598; @%p301 bra $L__BB3_483; ld.u64 %rd1599, [%rd406]; add.s64 %rd2311, %rd2315, %rd1599; st.local.u32 [%rd2+12], %r670; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2078, %r635, 5; and.b32 %r2079, %r2078, 3840; shl.b32 %r2080, %r139, 1; and.b32 %r2081, %r2080, 240; or.b32 %r2082, %r2081, %r2079; bfe.u32 %r2083, %r74, 3, 4; or.b32 %r2084, %r2082, %r2083; shr.u32 %r2085, %r2082, 6; mul.wide.u32 %rd1600, %r2085, 8; add.s64 %rd1601, %rd2311, %rd1600; and.b32 %r2086, %r2084, 63; shl.b64 %rd1603, %rd1589, %r2086; ld.u64 %rd1604, [%rd1601+544]; and.b64 %rd1605, %rd1604, %rd1603; setp.eq.s64 %p302, %rd1605, 0; mul.wide.u32 %rd1606, %r2084, 8; add.s64 %rd1607, %rd2311, %rd1606; add.s64 %rd408, %rd1607, 1088; @%p302 bra $L__BB3_482; ld.u64 %rd1608, [%rd408]; add.s64 %rd2307, %rd2311, %rd1608; and.b32 %r2087, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2087}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2088, %r635, 6; and.b32 %r2089, %r2088, 448; shl.b32 %r2090, %r139, 3; and.b32 %r2091, %r2090, 56; or.b32 %r2092, %r2091, %r2089; and.b32 %r2093, %r74, 7; or.b32 %r2094, %r2092, %r2093; mul.wide.u32 %rd1609, %r2094, 4; add.s64 %rd1610, %rd2307, %rd1609; ld.f32 %f472, [%rd1610+96]; mov.b64 %rd1611, {%r636, %r2087}; shr.u64 %rd1612, %rd1611, 32; cvt.u32.u64 %r3184, %rd1612; mov.u32 %r3232, %r673; mov.u32 %r3272, %r670; mov.u32 %r3229, %r636; bra.uni $L__BB3_486; $L__BB3_485: ld.f32 %f472, [%rd5+4]; bra.uni $L__BB3_486; $L__BB3_511: shl.b64 %rd1659, %rd429, 5; add.s64 %rd2380, %rd17, %rd1659; $L__BB3_512: setp.eq.s64 %p321, %rd2380, 0; @%p321 bra $L__BB3_520; add.s64 %rd432, %rd2380, 8; ld.u64 %rd433, [%rd2380+8]; setp.eq.s64 %p322, %rd433, 0; @%p322 bra $L__BB3_519; add.s64 %rd2315, %rd1, %rd433; and.b32 %r2154, %r8, -4096; mov.u64 %rd1660, 1; st.local.v2.u32 [%rd2+24], {%r703, %r2154}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2155, %r635, 3; and.b32 %r2156, %r2155, 31744; and.b32 %r2157, %r8, 3968; shr.u32 %r2158, %r2157, 2; or.b32 %r2159, %r2156, %r2158; bfe.u32 %r2160, %r74, 7, 5; or.b32 %r2161, %r2160, %r2159; shr.u32 %r2162, %r2159, 6; mul.wide.u32 %rd1661, %r2162, 8; add.s64 %rd1662, %rd2315, %rd1661; and.b32 %r2163, %r2161, 63; shl.b64 %rd1663, %rd1660, %r2163; ld.u64 %rd1664, [%rd1662+4128]; and.b64 %rd1665, %rd1664, %rd1663; setp.eq.s64 %p323, %rd1665, 0; mul.wide.u32 %rd1666, %r2161, 8; add.s64 %rd1667, %rd2315, %rd1666; add.s64 %rd435, %rd1667, 8256; mov.b64 %rd1668, {%r703, %r2154}; shr.u64 %rd1669, %rd1668, 32; cvt.u32.u64 %r3277, %rd1669; @%p323 bra $L__BB3_518; ld.u64 %rd1670, [%rd435]; add.s64 %rd2311, %rd2315, %rd1670; st.local.u32 [%rd2+12], %r700; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2164, %r635, 5; and.b32 %r2165, %r2164, 3840; shl.b32 %r2166, %r8, 1; and.b32 %r2167, %r2166, 240; or.b32 %r2168, %r2165, %r2167; bfe.u32 %r2169, %r74, 3, 4; or.b32 %r2170, %r2169, %r2168; shr.u32 %r2171, %r2168, 6; mul.wide.u32 %rd1671, %r2171, 8; add.s64 %rd1672, %rd2311, %rd1671; and.b32 %r2172, %r2170, 63; shl.b64 %rd1674, %rd1660, %r2172; ld.u64 %rd1675, [%rd1672+544]; and.b64 %rd1676, %rd1675, %rd1674; setp.eq.s64 %p324, %rd1676, 0; mul.wide.u32 %rd1677, %r2170, 8; add.s64 %rd1678, %rd2311, %rd1677; add.s64 %rd437, %rd1678, 1088; @%p324 bra $L__BB3_517; ld.u64 %rd1679, [%rd437]; add.s64 %rd2307, %rd2311, %rd1679; and.b32 %r2173, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2173}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2174, %r635, 6; and.b32 %r2175, %r2174, 448; shl.b32 %r2176, %r8, 3; and.b32 %r2177, %r2176, 56; or.b32 %r2178, %r2175, %r2177; and.b32 %r2179, %r74, 7; or.b32 %r2180, %r2178, %r2179; mul.wide.u32 %rd1680, %r2180, 4; add.s64 %rd1681, %rd2307, %rd1680; ld.f32 %f473, [%rd1681+96]; mov.b64 %rd1682, {%r636, %r2173}; shr.u64 %rd1683, %rd1682, 32; cvt.u32.u64 %r3184, %rd1683; mov.u32 %r3232, %r703; mov.u32 %r3272, %r700; mov.u32 %r3229, %r636; bra.uni $L__BB3_521; $L__BB3_520: ld.f32 %f473, [%rd5+4]; bra.uni $L__BB3_521; $L__BB3_546: shl.b64 %rd1726, %rd455, 5; add.s64 %rd2386, %rd17, %rd1726; $L__BB3_547: setp.eq.s64 %p343, %rd2386, 0; @%p343 bra $L__BB3_555; add.s64 %rd458, %rd2386, 8; ld.u64 %rd459, [%rd2386+8]; setp.eq.s64 %p344, %rd459, 0; @%p344 bra $L__BB3_554; add.s64 %rd460, %rd1, %rd459; and.b32 %r2243, %r8, -4096; mov.u64 %rd1727, 1; st.local.v2.u32 [%rd2+24], {%r728, %r2243}; and.b32 %r2244, %r9, -4096; st.local.u32 [%rd2+32], %r2244; st.local.u64 [%rd2+64], %rd460; shl.b32 %r2245, %r635, 3; and.b32 %r2246, %r2245, 31744; and.b32 %r2247, %r8, 3968; shr.u32 %r2248, %r2247, 2; or.b32 %r2249, %r2246, %r2248; bfe.u32 %r2250, %r9, 7, 5; or.b32 %r2251, %r2249, %r2250; shr.u32 %r2252, %r2249, 6; mul.wide.u32 %rd1728, %r2252, 8; add.s64 %rd1729, %rd460, %rd1728; and.b32 %r2253, %r2251, 63; shl.b64 %rd1730, %rd1727, %r2253; ld.u64 %rd1731, [%rd1729+4128]; and.b64 %rd1732, %rd1731, %rd1730; setp.eq.s64 %p345, %rd1732, 0; mul.wide.u32 %rd1733, %r2251, 8; add.s64 %rd1734, %rd460, %rd1733; add.s64 %rd461, %rd1734, 8256; @%p345 bra $L__BB3_553; ld.u64 %rd1735, [%rd461]; add.s64 %rd462, %rd460, %rd1735; st.local.u32 [%rd2+12], %r727; and.b32 %r2254, %r8, -128; st.local.u32 [%rd2+16], %r2254; and.b32 %r2255, %r9, -128; st.local.u32 [%rd2+20], %r2255; st.local.u64 [%rd2+56], %rd462; shl.b32 %r2256, %r635, 5; and.b32 %r2257, %r2256, 3840; shl.b32 %r2258, %r8, 1; and.b32 %r2259, %r2258, 240; or.b32 %r2260, %r2257, %r2259; bfe.u32 %r2261, %r9, 3, 4; or.b32 %r2262, %r2260, %r2261; shr.u32 %r2263, %r2260, 6; mul.wide.u32 %rd1736, %r2263, 8; add.s64 %rd1737, %rd462, %rd1736; and.b32 %r2264, %r2262, 63; shl.b64 %rd1739, %rd1727, %r2264; ld.u64 %rd1740, [%rd1737+544]; and.b64 %rd1741, %rd1740, %rd1739; setp.eq.s64 %p346, %rd1741, 0; mul.wide.u32 %rd1742, %r2262, 8; add.s64 %rd1743, %rd462, %rd1742; add.s64 %rd463, %rd1743, 1088; @%p346 bra $L__BB3_552; ld.u64 %rd1744, [%rd463]; add.s64 %rd1745, %rd462, %rd1744; and.b32 %r2265, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2265}; and.b32 %r2266, %r9, -8; st.local.u32 [%rd2+8], %r2266; st.local.u64 [%rd2+48], %rd1745; shl.b32 %r2267, %r635, 6; and.b32 %r2268, %r2267, 448; shl.b32 %r2269, %r8, 3; and.b32 %r2270, %r2269, 56; or.b32 %r2271, %r2268, %r2270; and.b32 %r2272, %r9, 7; or.b32 %r2273, %r2271, %r2272; mul.wide.u32 %rd1746, %r2273, 4; add.s64 %rd1747, %rd1745, %rd1746; ld.f32 %f474, [%rd1747+96]; bra.uni $L__BB3_556; $L__BB3_555: ld.f32 %f474, [%rd5+4]; bra.uni $L__BB3_556; $L__BB3_581: shl.b64 %rd1794, %rd473, 5; add.s64 %rd2387, %rd471, %rd1794; $L__BB3_582: setp.eq.s64 %p365, %rd2387, 0; @%p365 bra $L__BB3_590; add.s64 %rd476, %rd2387, 8; ld.u64 %rd477, [%rd2387+8]; setp.eq.s64 %p366, %rd477, 0; @%p366 bra $L__BB3_589; add.s64 %rd478, %rd470, %rd477; and.b32 %r2346, %r732, -4096; mov.u64 %rd1795, 1; st.local.v2.u32 [%rd2+24], {%r736, %r2346}; and.b32 %r2347, %r733, -4096; st.local.u32 [%rd2+32], %r2347; st.local.u64 [%rd2+64], %rd478; shl.b32 %r2348, %r471, 3; and.b32 %r2349, %r2348, 31744; and.b32 %r2350, %r732, 3968; shr.u32 %r2351, %r2350, 2; or.b32 %r2352, %r2351, %r2349; bfe.u32 %r2353, %r733, 7, 5; or.b32 %r2354, %r2352, %r2353; shr.u32 %r2355, %r2352, 6; mul.wide.u32 %rd1796, %r2355, 8; add.s64 %rd1797, %rd478, %rd1796; and.b32 %r2356, %r2354, 63; shl.b64 %rd1798, %rd1795, %r2356; ld.u64 %rd1799, [%rd1797+4128]; and.b64 %rd1800, %rd1799, %rd1798; setp.eq.s64 %p367, %rd1800, 0; mul.wide.u32 %rd1801, %r2354, 8; add.s64 %rd1802, %rd478, %rd1801; add.s64 %rd479, %rd1802, 8256; @%p367 bra $L__BB3_588; ld.u64 %rd1803, [%rd479]; add.s64 %rd480, %rd478, %rd1803; st.local.u32 [%rd2+12], %r735; and.b32 %r2357, %r732, -128; st.local.u32 [%rd2+16], %r2357; and.b32 %r2358, %r733, -128; st.local.u32 [%rd2+20], %r2358; st.local.u64 [%rd2+56], %rd480; shl.b32 %r2359, %r471, 5; and.b32 %r2360, %r2359, 3840; shl.b32 %r2361, %r732, 1; and.b32 %r2362, %r2361, 240; or.b32 %r2363, %r2362, %r2360; bfe.u32 %r2364, %r733, 3, 4; or.b32 %r2365, %r2363, %r2364; shr.u32 %r2366, %r2363, 6; mul.wide.u32 %rd1804, %r2366, 8; add.s64 %rd1805, %rd480, %rd1804; and.b32 %r2367, %r2365, 63; shl.b64 %rd1807, %rd1795, %r2367; ld.u64 %rd1808, [%rd1805+544]; and.b64 %rd1809, %rd1808, %rd1807; setp.eq.s64 %p368, %rd1809, 0; mul.wide.u32 %rd1810, %r2365, 8; add.s64 %rd1811, %rd480, %rd1810; add.s64 %rd481, %rd1811, 1088; @%p368 bra $L__BB3_587; ld.u64 %rd1812, [%rd481]; add.s64 %rd1813, %rd480, %rd1812; and.b32 %r2368, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2368}; and.b32 %r2369, %r733, -8; st.local.u32 [%rd2+8], %r2369; st.local.u64 [%rd2+48], %rd1813; shl.b32 %r2370, %r471, 6; and.b32 %r2371, %r2370, 448; shl.b32 %r2372, %r732, 3; and.b32 %r2373, %r2372, 56; or.b32 %r2374, %r2373, %r2371; and.b32 %r2375, %r733, 7; or.b32 %r2376, %r2374, %r2375; mul.wide.u32 %rd1814, %r2376, 4; add.s64 %rd1815, %rd1813, %rd1814; ld.f32 %f475, [%rd1815+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_591; $L__BB3_590: ld.f32 %f475, [%rd470+28]; bra.uni $L__BB3_591; $L__BB3_616: shl.b64 %rd1862, %rd491, 5; add.s64 %rd2388, %rd489, %rd1862; $L__BB3_617: setp.eq.s64 %p387, %rd2388, 0; @%p387 bra $L__BB3_625; add.s64 %rd494, %rd2388, 8; ld.u64 %rd495, [%rd2388+8]; setp.eq.s64 %p388, %rd495, 0; @%p388 bra $L__BB3_624; add.s64 %rd496, %rd488, %rd495; and.b32 %r2449, %r732, -4096; mov.u64 %rd1863, 1; st.local.v2.u32 [%rd2+24], {%r743, %r2449}; and.b32 %r2450, %r741, -4096; st.local.u32 [%rd2+32], %r2450; st.local.u64 [%rd2+64], %rd496; shl.b32 %r2451, %r471, 3; and.b32 %r2452, %r2451, 31744; and.b32 %r2453, %r732, 3968; shr.u32 %r2454, %r2453, 2; or.b32 %r2455, %r2454, %r2452; bfe.u32 %r2456, %r741, 7, 5; or.b32 %r2457, %r2456, %r2455; shr.u32 %r2458, %r2455, 6; mul.wide.u32 %rd1864, %r2458, 8; add.s64 %rd1865, %rd496, %rd1864; and.b32 %r2459, %r2457, 63; shl.b64 %rd1866, %rd1863, %r2459; ld.u64 %rd1867, [%rd1865+4128]; and.b64 %rd1868, %rd1867, %rd1866; setp.eq.s64 %p389, %rd1868, 0; mul.wide.u32 %rd1869, %r2457, 8; add.s64 %rd1870, %rd496, %rd1869; add.s64 %rd497, %rd1870, 8256; @%p389 bra $L__BB3_623; ld.u64 %rd1871, [%rd497]; add.s64 %rd498, %rd496, %rd1871; st.local.u32 [%rd2+12], %r742; and.b32 %r2460, %r732, -128; st.local.u32 [%rd2+16], %r2460; and.b32 %r2461, %r741, -128; st.local.u32 [%rd2+20], %r2461; st.local.u64 [%rd2+56], %rd498; shl.b32 %r2462, %r471, 5; and.b32 %r2463, %r2462, 3840; shl.b32 %r2464, %r732, 1; and.b32 %r2465, %r2464, 240; or.b32 %r2466, %r2465, %r2463; bfe.u32 %r2467, %r741, 3, 4; or.b32 %r2468, %r2467, %r2466; shr.u32 %r2469, %r2466, 6; mul.wide.u32 %rd1872, %r2469, 8; add.s64 %rd1873, %rd498, %rd1872; and.b32 %r2470, %r2468, 63; shl.b64 %rd1875, %rd1863, %r2470; ld.u64 %rd1876, [%rd1873+544]; and.b64 %rd1877, %rd1876, %rd1875; setp.eq.s64 %p390, %rd1877, 0; mul.wide.u32 %rd1878, %r2468, 8; add.s64 %rd1879, %rd498, %rd1878; add.s64 %rd499, %rd1879, 1088; @%p390 bra $L__BB3_622; ld.u64 %rd1880, [%rd499]; add.s64 %rd1881, %rd498, %rd1880; and.b32 %r2471, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2471}; and.b32 %r2472, %r741, -8; st.local.u32 [%rd2+8], %r2472; st.local.u64 [%rd2+48], %rd1881; shl.b32 %r2473, %r471, 6; and.b32 %r2474, %r2473, 448; shl.b32 %r2475, %r732, 3; and.b32 %r2476, %r2475, 56; or.b32 %r2477, %r2476, %r2474; and.b32 %r2478, %r741, 7; or.b32 %r2479, %r2477, %r2478; mul.wide.u32 %rd1882, %r2479, 4; add.s64 %rd1883, %rd1881, %rd1882; ld.f32 %f476, [%rd1883+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_626; $L__BB3_625: ld.f32 %f476, [%rd488+28]; bra.uni $L__BB3_626; $L__BB3_651: shl.b64 %rd1929, %rd509, 5; add.s64 %rd2389, %rd507, %rd1929; $L__BB3_652: setp.eq.s64 %p409, %rd2389, 0; @%p409 bra $L__BB3_660; add.s64 %rd512, %rd2389, 8; ld.u64 %rd513, [%rd2389+8]; setp.eq.s64 %p410, %rd513, 0; @%p410 bra $L__BB3_659; add.s64 %rd514, %rd506, %rd513; and.b32 %r2552, %r748, -4096; mov.u64 %rd1930, 1; st.local.v2.u32 [%rd2+24], {%r750, %r2552}; and.b32 %r2553, %r741, -4096; st.local.u32 [%rd2+32], %r2553; st.local.u64 [%rd2+64], %rd514; shl.b32 %r2554, %r471, 3; and.b32 %r2555, %r2554, 31744; and.b32 %r2556, %r748, 3968; shr.u32 %r2557, %r2556, 2; or.b32 %r2558, %r2557, %r2555; bfe.u32 %r2559, %r741, 7, 5; or.b32 %r2560, %r2558, %r2559; shr.u32 %r2561, %r2558, 6; mul.wide.u32 %rd1931, %r2561, 8; add.s64 %rd1932, %rd514, %rd1931; and.b32 %r2562, %r2560, 63; shl.b64 %rd1933, %rd1930, %r2562; ld.u64 %rd1934, [%rd1932+4128]; and.b64 %rd1935, %rd1934, %rd1933; setp.eq.s64 %p411, %rd1935, 0; mul.wide.u32 %rd1936, %r2560, 8; add.s64 %rd1937, %rd514, %rd1936; add.s64 %rd515, %rd1937, 8256; @%p411 bra $L__BB3_658; ld.u64 %rd1938, [%rd515]; add.s64 %rd516, %rd514, %rd1938; st.local.u32 [%rd2+12], %r749; and.b32 %r2563, %r748, -128; st.local.u32 [%rd2+16], %r2563; and.b32 %r2564, %r741, -128; st.local.u32 [%rd2+20], %r2564; st.local.u64 [%rd2+56], %rd516; shl.b32 %r2565, %r471, 5; and.b32 %r2566, %r2565, 3840; shl.b32 %r2567, %r748, 1; and.b32 %r2568, %r2567, 240; or.b32 %r2569, %r2568, %r2566; bfe.u32 %r2570, %r741, 3, 4; or.b32 %r2571, %r2569, %r2570; shr.u32 %r2572, %r2569, 6; mul.wide.u32 %rd1939, %r2572, 8; add.s64 %rd1940, %rd516, %rd1939; and.b32 %r2573, %r2571, 63; shl.b64 %rd1942, %rd1930, %r2573; ld.u64 %rd1943, [%rd1940+544]; and.b64 %rd1944, %rd1943, %rd1942; setp.eq.s64 %p412, %rd1944, 0; mul.wide.u32 %rd1945, %r2571, 8; add.s64 %rd1946, %rd516, %rd1945; add.s64 %rd517, %rd1946, 1088; @%p412 bra $L__BB3_657; ld.u64 %rd1947, [%rd517]; add.s64 %rd1948, %rd516, %rd1947; and.b32 %r2574, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2574}; and.b32 %r2575, %r741, -8; st.local.u32 [%rd2+8], %r2575; st.local.u64 [%rd2+48], %rd1948; shl.b32 %r2576, %r471, 6; and.b32 %r2577, %r2576, 448; shl.b32 %r2578, %r748, 3; and.b32 %r2579, %r2578, 56; or.b32 %r2580, %r2579, %r2577; and.b32 %r2581, %r741, 7; or.b32 %r2582, %r2580, %r2581; mul.wide.u32 %rd1949, %r2582, 4; add.s64 %rd1950, %rd1948, %rd1949; ld.f32 %f477, [%rd1950+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_661; $L__BB3_660: ld.f32 %f477, [%rd506+28]; bra.uni $L__BB3_661; $L__BB3_686: shl.b64 %rd1996, %rd527, 5; add.s64 %rd2390, %rd525, %rd1996; $L__BB3_687: setp.eq.s64 %p431, %rd2390, 0; @%p431 bra $L__BB3_695; add.s64 %rd530, %rd2390, 8; ld.u64 %rd531, [%rd2390+8]; setp.eq.s64 %p432, %rd531, 0; @%p432 bra $L__BB3_694; add.s64 %rd532, %rd524, %rd531; and.b32 %r2652, %r748, -4096; mov.u64 %rd1997, 1; st.local.v2.u32 [%rd2+24], {%r756, %r2652}; and.b32 %r2653, %r733, -4096; st.local.u32 [%rd2+32], %r2653; st.local.u64 [%rd2+64], %rd532; shl.b32 %r2654, %r471, 3; and.b32 %r2655, %r2654, 31744; and.b32 %r2656, %r748, 3968; shr.u32 %r2657, %r2656, 2; or.b32 %r2658, %r2657, %r2655; bfe.u32 %r2659, %r733, 7, 5; or.b32 %r2660, %r2658, %r2659; shr.u32 %r2661, %r2658, 6; mul.wide.u32 %rd1998, %r2661, 8; add.s64 %rd1999, %rd532, %rd1998; and.b32 %r2662, %r2660, 63; shl.b64 %rd2000, %rd1997, %r2662; ld.u64 %rd2001, [%rd1999+4128]; and.b64 %rd2002, %rd2001, %rd2000; setp.eq.s64 %p433, %rd2002, 0; mul.wide.u32 %rd2003, %r2660, 8; add.s64 %rd2004, %rd532, %rd2003; add.s64 %rd533, %rd2004, 8256; @%p433 bra $L__BB3_693; ld.u64 %rd2005, [%rd533]; add.s64 %rd534, %rd532, %rd2005; st.local.u32 [%rd2+12], %r755; and.b32 %r2663, %r748, -128; st.local.u32 [%rd2+16], %r2663; and.b32 %r2664, %r733, -128; st.local.u32 [%rd2+20], %r2664; st.local.u64 [%rd2+56], %rd534; shl.b32 %r2665, %r471, 5; and.b32 %r2666, %r2665, 3840; shl.b32 %r2667, %r748, 1; and.b32 %r2668, %r2667, 240; or.b32 %r2669, %r2668, %r2666; bfe.u32 %r2670, %r733, 3, 4; or.b32 %r2671, %r2669, %r2670; shr.u32 %r2672, %r2669, 6; mul.wide.u32 %rd2006, %r2672, 8; add.s64 %rd2007, %rd534, %rd2006; and.b32 %r2673, %r2671, 63; shl.b64 %rd2009, %rd1997, %r2673; ld.u64 %rd2010, [%rd2007+544]; and.b64 %rd2011, %rd2010, %rd2009; setp.eq.s64 %p434, %rd2011, 0; mul.wide.u32 %rd2012, %r2671, 8; add.s64 %rd2013, %rd534, %rd2012; add.s64 %rd535, %rd2013, 1088; @%p434 bra $L__BB3_692; ld.u64 %rd2014, [%rd535]; add.s64 %rd2015, %rd534, %rd2014; and.b32 %r2674, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2674}; and.b32 %r2675, %r733, -8; st.local.u32 [%rd2+8], %r2675; st.local.u64 [%rd2+48], %rd2015; and.b32 %r2676, %r471, 7; shl.b32 %r2677, %r748, 3; and.b32 %r2678, %r2677, 56; and.b32 %r2679, %r733, 7; bfi.b32 %r2680, %r2676, %r2679, 6, 3; or.b32 %r2681, %r2680, %r2678; mul.wide.u32 %rd2016, %r2681, 4; add.s64 %rd2017, %rd2015, %rd2016; ld.f32 %f478, [%rd2017+96]; mov.u32 %r3598, %r472; bra.uni $L__BB3_696; $L__BB3_695: ld.f32 %f478, [%rd524+28]; bra.uni $L__BB3_696; $L__BB3_721: shl.b64 %rd2063, %rd545, 5; add.s64 %rd2391, %rd543, %rd2063; $L__BB3_722: setp.eq.s64 %p453, %rd2391, 0; @%p453 bra $L__BB3_730; add.s64 %rd548, %rd2391, 8; ld.u64 %rd549, [%rd2391+8]; setp.eq.s64 %p454, %rd549, 0; @%p454 bra $L__BB3_729; add.s64 %rd550, %rd542, %rd549; and.b32 %r2751, %r748, -4096; mov.u64 %rd2064, 1; st.local.v2.u32 [%rd2+24], {%r762, %r2751}; and.b32 %r2752, %r733, -4096; st.local.u32 [%rd2+32], %r2752; st.local.u64 [%rd2+64], %rd550; shl.b32 %r2753, %r635, 3; and.b32 %r2754, %r2753, 31744; and.b32 %r2755, %r748, 3968; shr.u32 %r2756, %r2755, 2; or.b32 %r2757, %r2756, %r2754; bfe.u32 %r2758, %r733, 7, 5; or.b32 %r2759, %r2757, %r2758; shr.u32 %r2760, %r2757, 6; mul.wide.u32 %rd2065, %r2760, 8; add.s64 %rd2066, %rd550, %rd2065; and.b32 %r2761, %r2759, 63; shl.b64 %rd2067, %rd2064, %r2761; ld.u64 %rd2068, [%rd2066+4128]; and.b64 %rd2069, %rd2068, %rd2067; setp.eq.s64 %p455, %rd2069, 0; mul.wide.u32 %rd2070, %r2759, 8; add.s64 %rd2071, %rd550, %rd2070; add.s64 %rd551, %rd2071, 8256; @%p455 bra $L__BB3_728; ld.u64 %rd2072, [%rd551]; add.s64 %rd552, %rd550, %rd2072; st.local.u32 [%rd2+12], %r761; and.b32 %r2762, %r748, -128; st.local.u32 [%rd2+16], %r2762; and.b32 %r2763, %r733, -128; st.local.u32 [%rd2+20], %r2763; st.local.u64 [%rd2+56], %rd552; shl.b32 %r2764, %r635, 5; and.b32 %r2765, %r2764, 3840; shl.b32 %r2766, %r748, 1; and.b32 %r2767, %r2766, 240; or.b32 %r2768, %r2767, %r2765; bfe.u32 %r2769, %r733, 3, 4; or.b32 %r2770, %r2768, %r2769; shr.u32 %r2771, %r2768, 6; mul.wide.u32 %rd2073, %r2771, 8; add.s64 %rd2074, %rd552, %rd2073; and.b32 %r2772, %r2770, 63; shl.b64 %rd2076, %rd2064, %r2772; ld.u64 %rd2077, [%rd2074+544]; and.b64 %rd2078, %rd2077, %rd2076; setp.eq.s64 %p456, %rd2078, 0; mul.wide.u32 %rd2079, %r2770, 8; add.s64 %rd2080, %rd552, %rd2079; add.s64 %rd553, %rd2080, 1088; @%p456 bra $L__BB3_727; ld.u64 %rd2081, [%rd553]; add.s64 %rd2082, %rd552, %rd2081; and.b32 %r2773, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2773}; and.b32 %r2774, %r733, -8; st.local.u32 [%rd2+8], %r2774; st.local.u64 [%rd2+48], %rd2082; and.b32 %r2775, %r635, 7; shl.b32 %r2776, %r748, 3; and.b32 %r2777, %r2776, 56; and.b32 %r2778, %r733, 7; bfi.b32 %r2779, %r2775, %r2778, 6, 3; or.b32 %r2780, %r2779, %r2777; mul.wide.u32 %rd2083, %r2780, 4; add.s64 %rd2084, %rd2082, %rd2083; ld.f32 %f479, [%rd2084+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_731; $L__BB3_730: ld.f32 %f479, [%rd542+28]; bra.uni $L__BB3_731; $L__BB3_756: shl.b64 %rd2130, %rd563, 5; add.s64 %rd2392, %rd561, %rd2130; $L__BB3_757: setp.eq.s64 %p475, %rd2392, 0; @%p475 bra $L__BB3_765; add.s64 %rd566, %rd2392, 8; ld.u64 %rd567, [%rd2392+8]; setp.eq.s64 %p476, %rd567, 0; @%p476 bra $L__BB3_764; add.s64 %rd568, %rd560, %rd567; and.b32 %r2853, %r748, -4096; mov.u64 %rd2131, 1; st.local.v2.u32 [%rd2+24], {%r768, %r2853}; and.b32 %r2854, %r741, -4096; st.local.u32 [%rd2+32], %r2854; st.local.u64 [%rd2+64], %rd568; shl.b32 %r2855, %r635, 3; and.b32 %r2856, %r2855, 31744; and.b32 %r2857, %r748, 3968; shr.u32 %r2858, %r2857, 2; or.b32 %r2859, %r2858, %r2856; bfe.u32 %r2860, %r741, 7, 5; or.b32 %r2861, %r2859, %r2860; shr.u32 %r2862, %r2859, 6; mul.wide.u32 %rd2132, %r2862, 8; add.s64 %rd2133, %rd568, %rd2132; and.b32 %r2863, %r2861, 63; shl.b64 %rd2134, %rd2131, %r2863; ld.u64 %rd2135, [%rd2133+4128]; and.b64 %rd2136, %rd2135, %rd2134; setp.eq.s64 %p477, %rd2136, 0; mul.wide.u32 %rd2137, %r2861, 8; add.s64 %rd2138, %rd568, %rd2137; add.s64 %rd569, %rd2138, 8256; @%p477 bra $L__BB3_763; ld.u64 %rd2139, [%rd569]; add.s64 %rd570, %rd568, %rd2139; st.local.u32 [%rd2+12], %r767; and.b32 %r2864, %r748, -128; st.local.u32 [%rd2+16], %r2864; and.b32 %r2865, %r741, -128; st.local.u32 [%rd2+20], %r2865; st.local.u64 [%rd2+56], %rd570; shl.b32 %r2866, %r635, 5; and.b32 %r2867, %r2866, 3840; shl.b32 %r2868, %r748, 1; and.b32 %r2869, %r2868, 240; or.b32 %r2870, %r2869, %r2867; bfe.u32 %r2871, %r741, 3, 4; or.b32 %r2872, %r2870, %r2871; shr.u32 %r2873, %r2870, 6; mul.wide.u32 %rd2140, %r2873, 8; add.s64 %rd2141, %rd570, %rd2140; and.b32 %r2874, %r2872, 63; shl.b64 %rd2143, %rd2131, %r2874; ld.u64 %rd2144, [%rd2141+544]; and.b64 %rd2145, %rd2144, %rd2143; setp.eq.s64 %p478, %rd2145, 0; mul.wide.u32 %rd2146, %r2872, 8; add.s64 %rd2147, %rd570, %rd2146; add.s64 %rd571, %rd2147, 1088; @%p478 bra $L__BB3_762; ld.u64 %rd2148, [%rd571]; add.s64 %rd2149, %rd570, %rd2148; and.b32 %r2875, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2875}; and.b32 %r2876, %r741, -8; st.local.u32 [%rd2+8], %r2876; st.local.u64 [%rd2+48], %rd2149; shl.b32 %r2877, %r635, 6; and.b32 %r2878, %r2877, 448; shl.b32 %r2879, %r748, 3; and.b32 %r2880, %r2879, 56; or.b32 %r2881, %r2880, %r2878; and.b32 %r2882, %r741, 7; or.b32 %r2883, %r2881, %r2882; mul.wide.u32 %rd2150, %r2883, 4; add.s64 %rd2151, %rd2149, %rd2150; ld.f32 %f480, [%rd2151+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_766; $L__BB3_765: ld.f32 %f480, [%rd560+28]; bra.uni $L__BB3_766; $L__BB3_791: shl.b64 %rd2198, %rd581, 5; add.s64 %rd2393, %rd579, %rd2198; $L__BB3_792: setp.eq.s64 %p497, %rd2393, 0; @%p497 bra $L__BB3_800; add.s64 %rd584, %rd2393, 8; ld.u64 %rd585, [%rd2393+8]; setp.eq.s64 %p498, %rd585, 0; @%p498 bra $L__BB3_799; add.s64 %rd586, %rd578, %rd585; and.b32 %r2956, %r732, -4096; mov.u64 %rd2199, 1; st.local.v2.u32 [%rd2+24], {%r774, %r2956}; and.b32 %r2957, %r741, -4096; st.local.u32 [%rd2+32], %r2957; st.local.u64 [%rd2+64], %rd586; shl.b32 %r2958, %r635, 3; and.b32 %r2959, %r2958, 31744; and.b32 %r2960, %r732, 3968; shr.u32 %r2961, %r2960, 2; or.b32 %r2962, %r2959, %r2961; bfe.u32 %r2963, %r741, 7, 5; or.b32 %r2964, %r2963, %r2962; shr.u32 %r2965, %r2962, 6; mul.wide.u32 %rd2200, %r2965, 8; add.s64 %rd2201, %rd586, %rd2200; and.b32 %r2966, %r2964, 63; shl.b64 %rd2202, %rd2199, %r2966; ld.u64 %rd2203, [%rd2201+4128]; and.b64 %rd2204, %rd2203, %rd2202; setp.eq.s64 %p499, %rd2204, 0; mul.wide.u32 %rd2205, %r2964, 8; add.s64 %rd2206, %rd586, %rd2205; add.s64 %rd587, %rd2206, 8256; @%p499 bra $L__BB3_798; ld.u64 %rd2207, [%rd587]; add.s64 %rd588, %rd586, %rd2207; st.local.u32 [%rd2+12], %r773; and.b32 %r2967, %r732, -128; st.local.u32 [%rd2+16], %r2967; and.b32 %r2968, %r741, -128; st.local.u32 [%rd2+20], %r2968; st.local.u64 [%rd2+56], %rd588; shl.b32 %r2969, %r635, 5; and.b32 %r2970, %r2969, 3840; shl.b32 %r2971, %r732, 1; and.b32 %r2972, %r2971, 240; or.b32 %r2973, %r2970, %r2972; bfe.u32 %r2974, %r741, 3, 4; or.b32 %r2975, %r2974, %r2973; shr.u32 %r2976, %r2973, 6; mul.wide.u32 %rd2208, %r2976, 8; add.s64 %rd2209, %rd588, %rd2208; and.b32 %r2977, %r2975, 63; shl.b64 %rd2211, %rd2199, %r2977; ld.u64 %rd2212, [%rd2209+544]; and.b64 %rd2213, %rd2212, %rd2211; setp.eq.s64 %p500, %rd2213, 0; mul.wide.u32 %rd2214, %r2975, 8; add.s64 %rd2215, %rd588, %rd2214; add.s64 %rd589, %rd2215, 1088; @%p500 bra $L__BB3_797; ld.u64 %rd2216, [%rd589]; add.s64 %rd2217, %rd588, %rd2216; and.b32 %r2978, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2978}; and.b32 %r2979, %r741, -8; st.local.u32 [%rd2+8], %r2979; st.local.u64 [%rd2+48], %rd2217; shl.b32 %r2980, %r635, 6; and.b32 %r2981, %r2980, 448; shl.b32 %r2982, %r732, 3; and.b32 %r2983, %r2982, 56; or.b32 %r2984, %r2981, %r2983; and.b32 %r2985, %r741, 7; or.b32 %r2986, %r2984, %r2985; mul.wide.u32 %rd2218, %r2986, 4; add.s64 %rd2219, %rd2217, %rd2218; ld.f32 %f481, [%rd2219+96]; mov.u32 %r3598, %r636; bra.uni $L__BB3_801; $L__BB3_800: ld.f32 %f481, [%rd578+28]; bra.uni $L__BB3_801; $L__BB3_64: ld.f32 %f460, [%rd55+12]; bra.uni $L__BB3_66; $L__BB3_99: ld.f32 %f461, [%rd84+12]; bra.uni $L__BB3_101; $L__BB3_134: ld.f32 %f462, [%rd113+12]; bra.uni $L__BB3_136; $L__BB3_169: ld.f32 %f463, [%rd142+12]; bra.uni $L__BB3_171; $L__BB3_204: ld.f32 %f464, [%rd171+12]; bra.uni $L__BB3_206; $L__BB3_239: ld.f32 %f465, [%rd200+12]; bra.uni $L__BB3_241; $L__BB3_274: ld.f32 %f466, [%rd229+12]; bra.uni $L__BB3_276; $L__BB3_309: ld.f32 %f467, [%rd258+12]; bra.uni $L__BB3_311; $L__BB3_344: ld.f32 %f468, [%rd287+12]; bra.uni $L__BB3_346; $L__BB3_379: ld.f32 %f469, [%rd316+12]; bra.uni $L__BB3_381; $L__BB3_414: ld.f32 %f470, [%rd345+12]; bra.uni $L__BB3_416; $L__BB3_449: ld.f32 %f471, [%rd374+12]; bra.uni $L__BB3_451; $L__BB3_484: ld.f32 %f472, [%rd403+12]; bra.uni $L__BB3_486; $L__BB3_519: ld.f32 %f473, [%rd432+12]; bra.uni $L__BB3_521; $L__BB3_554: ld.f32 %f474, [%rd458+12]; bra.uni $L__BB3_556; $L__BB3_589: ld.f32 %f475, [%rd476+12]; bra.uni $L__BB3_591; $L__BB3_624: ld.f32 %f476, [%rd494+12]; bra.uni $L__BB3_626; $L__BB3_659: ld.f32 %f477, [%rd512+12]; bra.uni $L__BB3_661; $L__BB3_694: ld.f32 %f478, [%rd530+12]; bra.uni $L__BB3_696; $L__BB3_729: ld.f32 %f479, [%rd548+12]; bra.uni $L__BB3_731; $L__BB3_764: ld.f32 %f480, [%rd566+12]; bra.uni $L__BB3_766; $L__BB3_799: ld.f32 %f481, [%rd584+12]; bra.uni $L__BB3_801; $L__BB3_63: ld.f32 %f460, [%rd58]; mov.u32 %r3232, %r12; bra.uni $L__BB3_66; $L__BB3_98: ld.f32 %f461, [%rd87]; mov.u32 %r3232, %r12; bra.uni $L__BB3_101; $L__BB3_133: ld.f32 %f462, [%rd116]; mov.u32 %r3232, %r12; bra.uni $L__BB3_136; $L__BB3_168: and.b32 %r3232, %r259, -4096; ld.f32 %f463, [%rd145]; bra.uni $L__BB3_171; $L__BB3_203: and.b32 %r3232, %r259, -4096; ld.f32 %f464, [%rd174]; bra.uni $L__BB3_206; $L__BB3_238: and.b32 %r3232, %r259, -4096; ld.f32 %f465, [%rd203]; bra.uni $L__BB3_241; $L__BB3_273: ld.f32 %f466, [%rd232]; mov.u32 %r3232, %r426; bra.uni $L__BB3_276; $L__BB3_308: ld.f32 %f467, [%rd261]; mov.u32 %r3232, %r476; bra.uni $L__BB3_311; $L__BB3_343: ld.f32 %f468, [%rd290]; mov.u32 %r3232, %r521; bra.uni $L__BB3_346; $L__BB3_378: ld.f32 %f469, [%rd319]; mov.u32 %r3232, %r563; bra.uni $L__BB3_381; $L__BB3_413: ld.f32 %f470, [%rd348]; mov.u32 %r3232, %r602; bra.uni $L__BB3_416; $L__BB3_448: ld.f32 %f471, [%rd377]; mov.u32 %r3232, %r640; bra.uni $L__BB3_451; $L__BB3_483: ld.f32 %f472, [%rd406]; mov.u32 %r3232, %r673; bra.uni $L__BB3_486; $L__BB3_518: ld.f32 %f473, [%rd435]; mov.u32 %r3232, %r703; bra.uni $L__BB3_521; $L__BB3_553: ld.f32 %f474, [%rd461]; bra.uni $L__BB3_556; $L__BB3_588: ld.f32 %f475, [%rd479]; bra.uni $L__BB3_591; $L__BB3_623: ld.f32 %f476, [%rd497]; bra.uni $L__BB3_626; $L__BB3_658: ld.f32 %f477, [%rd515]; bra.uni $L__BB3_661; $L__BB3_693: ld.f32 %f478, [%rd533]; bra.uni $L__BB3_696; $L__BB3_728: ld.f32 %f479, [%rd551]; bra.uni $L__BB3_731; $L__BB3_763: ld.f32 %f480, [%rd569]; bra.uni $L__BB3_766; $L__BB3_798: ld.f32 %f481, [%rd587]; bra.uni $L__BB3_801; $L__BB3_62: ld.f32 %f460, [%rd60]; mov.u32 %r3232, %r12; bra.uni $L__BB3_66; $L__BB3_97: ld.f32 %f461, [%rd89]; mov.u32 %r3232, %r12; bra.uni $L__BB3_101; $L__BB3_132: ld.f32 %f462, [%rd118]; mov.u32 %r3232, %r12; bra.uni $L__BB3_136; $L__BB3_167: and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.f32 %f463, [%rd147]; bra.uni $L__BB3_171; $L__BB3_202: and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.f32 %f464, [%rd176]; bra.uni $L__BB3_206; $L__BB3_237: ld.f32 %f465, [%rd205]; bra.uni $L__BB3_241; $L__BB3_272: ld.f32 %f466, [%rd234]; mov.u32 %r3232, %r426; mov.u32 %r3272, %r423; bra.uni $L__BB3_276; $L__BB3_307: ld.f32 %f467, [%rd263]; mov.u32 %r3232, %r476; mov.u32 %r3272, %r473; bra.uni $L__BB3_311; $L__BB3_342: ld.f32 %f468, [%rd292]; mov.u32 %r3232, %r521; mov.u32 %r3272, %r518; bra.uni $L__BB3_346; $L__BB3_377: ld.f32 %f469, [%rd321]; mov.u32 %r3232, %r563; mov.u32 %r3272, %r560; bra.uni $L__BB3_381; $L__BB3_412: ld.f32 %f470, [%rd350]; mov.u32 %r3232, %r602; mov.u32 %r3272, %r599; bra.uni $L__BB3_416; $L__BB3_447: ld.f32 %f471, [%rd379]; mov.u32 %r3232, %r640; mov.u32 %r3272, %r637; bra.uni $L__BB3_451; $L__BB3_482: ld.f32 %f472, [%rd408]; mov.u32 %r3232, %r673; mov.u32 %r3272, %r670; bra.uni $L__BB3_486; $L__BB3_517: ld.f32 %f473, [%rd437]; mov.u32 %r3232, %r703; mov.u32 %r3272, %r700; bra.uni $L__BB3_521; $L__BB3_552: ld.f32 %f474, [%rd463]; bra.uni $L__BB3_556; $L__BB3_587: ld.f32 %f475, [%rd481]; bra.uni $L__BB3_591; $L__BB3_622: ld.f32 %f476, [%rd499]; bra.uni $L__BB3_626; $L__BB3_657: ld.f32 %f477, [%rd517]; bra.uni $L__BB3_661; $L__BB3_692: ld.f32 %f478, [%rd535]; bra.uni $L__BB3_696; $L__BB3_727: ld.f32 %f479, [%rd553]; bra.uni $L__BB3_731; $L__BB3_762: ld.f32 %f480, [%rd571]; bra.uni $L__BB3_766; $L__BB3_797: ld.f32 %f481, [%rd589]; bra.uni $L__BB3_801; } .func (.param .align 4 .b8 func_retval0[12]) _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj( .param .b64 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4 ) { .local .align 8 .b8 __local_depot4[88]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<344>; .reg .b16 %rs<211>; .reg .f32 %f<960>; .reg .b32 %r<2700>; .reg .b64 %rd<1705>; mov.u64 %SPL, __local_depot4; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd464, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0]; ld.param.f32 %f626, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1]; ld.param.f32 %f627, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2]; ld.param.f32 %f628, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; ld.param.u32 %r732, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4]; cvta.to.global.u64 %rd465, %rd464; ld.global.u64 %rd466, [%rd465]; mov.u64 %rd1616, 0; ld.u64 %rd468, [%rd466+696]; add.s64 %rd469, %rd466, %rd468; add.s64 %rd1, %rd469, 672; add.u64 %rd470, %SP, 0; add.u64 %rd2, %SPL, 0; mov.u32 %r2289, 2147483647; st.local.v2.u32 [%rd2], {%r2289, %r2289}; add.s64 %rd4, %rd2, 8; st.local.v2.u32 [%rd2+8], {%r2289, %r2289}; st.local.v2.u32 [%rd2+16], {%r2289, %r2289}; st.local.v2.u32 [%rd2+24], {%r2289, %r2289}; st.local.u32 [%rd2+32], %r2289; st.local.u64 [%rd2+40], %rd1; st.local.u64 [%rd2+48], %rd1616; st.local.u64 [%rd2+56], %rd1616; st.local.u64 [%rd2+64], %rd1616; add.s64 %rd5, %rd469, 696; setp.eq.s32 %p1, %r732, 0; @%p1 bra $L__BB4_16; setp.ne.s32 %p2, %r732, 1; @%p2 bra $L__BB4_17; cvt.rni.ftz.f32.f32 %f629, %f626; cvt.rzi.ftz.s32.f32 %r1, %f629; cvt.rni.ftz.f32.f32 %f630, %f627; cvt.rzi.ftz.s32.f32 %r2, %f630; cvt.rni.ftz.f32.f32 %f631, %f628; cvt.rzi.ftz.s32.f32 %r3, %f631; ld.u32 %r4, [%rd5]; mov.u64 %rd1615, 0; setp.eq.s32 %p3, %r4, 0; @%p3 bra $L__BB4_7; add.s64 %rd6, %rd1, 96; shr.u32 %r735, %r1, 12; cvt.u64.u32 %rd472, %r735; shl.b64 %rd473, %rd472, 42; shr.u32 %r736, %r2, 12; mul.wide.u32 %rd474, %r736, 2097152; or.b64 %rd475, %rd474, %rd473; shr.u32 %r737, %r3, 12; cvt.u64.u32 %rd476, %r737; or.b64 %rd7, %rd475, %rd476; mov.u32 %r2275, 0; $L__BB4_4: cvt.u64.u32 %rd8, %r2275; mul.wide.u32 %rd477, %r2275, 32; add.s64 %rd478, %rd6, %rd477; ld.u64 %rd479, [%rd478]; setp.eq.s64 %p4, %rd479, %rd7; add.s32 %r2275, %r2275, 1; @%p4 bra $L__BB4_6; setp.lt.u32 %p5, %r2275, %r4; @%p5 bra $L__BB4_4; bra.uni $L__BB4_7; $L__BB4_16: add.u64 %rd503, %SP, 72; add.u64 %rd504, %SPL, 72; st.local.u64 [%rd504], %rd470; add.ftz.f32 %f632, %f626, 0fBF000000; add.ftz.f32 %f633, %f627, 0fBF000000; add.ftz.f32 %f634, %f628, 0fBF000000; { // callseq 6, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd503; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f632; st.param.f32 [param1+4], %f633; st.param.f32 [param1+8], %f634; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f957, [retval0+0]; ld.param.f32 %f958, [retval0+4]; ld.param.f32 %f959, [retval0+8]; } // callseq 6 bra.uni $L__BB4_554; $L__BB4_17: add.u64 %rd508, %SPL, 80; st.local.u64 [%rd508], %rd470; add.ftz.f32 %f635, %f626, 0fBF000000; cvt.rmi.ftz.f32.f32 %f636, %f635; add.ftz.f32 %f637, %f627, 0fBF000000; cvt.rmi.ftz.f32.f32 %f638, %f637; add.ftz.f32 %f639, %f628, 0fBF000000; cvt.rmi.ftz.f32.f32 %f640, %f639; sub.ftz.f32 %f641, %f635, %f636; sub.ftz.f32 %f642, %f637, %f638; sub.ftz.f32 %f643, %f639, %f640; mov.f32 %f644, 0f40400000; sub.ftz.f32 %f645, %f644, %f641; fma.rn.ftz.f32 %f646, %f641, %f645, 0fC0400000; fma.rn.ftz.f32 %f647, %f641, %f646, 0f3F800000; mul.ftz.f32 %f648, %f641, %f641; mul.ftz.f32 %f649, %f641, 0f40400000; add.ftz.f32 %f650, %f649, 0fC0C00000; fma.rn.ftz.f32 %f651, %f648, %f650, 0f40800000; mul.ftz.f32 %f652, %f651, 0f3E2AAAAB; fma.rn.ftz.f32 %f25, %f647, 0f3E2AAAAB, %f652; sub.ftz.f32 %f653, %f644, %f649; fma.rn.ftz.f32 %f654, %f641, %f653, 0f40400000; fma.rn.ftz.f32 %f655, %f641, %f654, 0f3F800000; mul.ftz.f32 %f656, %f641, %f648; mul.ftz.f32 %f657, %f656, 0f3E2AAAAB; fma.rn.ftz.f32 %f26, %f655, 0f3E2AAAAB, %f657; sub.ftz.f32 %f658, %f644, %f642; fma.rn.ftz.f32 %f659, %f642, %f658, 0fC0400000; fma.rn.ftz.f32 %f660, %f642, %f659, 0f3F800000; mul.ftz.f32 %f661, %f642, %f642; mul.ftz.f32 %f662, %f642, 0f40400000; add.ftz.f32 %f663, %f662, 0fC0C00000; fma.rn.ftz.f32 %f664, %f661, %f663, 0f40800000; mul.ftz.f32 %f665, %f664, 0f3E2AAAAB; fma.rn.ftz.f32 %f27, %f660, 0f3E2AAAAB, %f665; sub.ftz.f32 %f666, %f644, %f662; fma.rn.ftz.f32 %f667, %f642, %f666, 0f40400000; fma.rn.ftz.f32 %f668, %f642, %f667, 0f3F800000; sub.ftz.f32 %f671, %f644, %f643; fma.rn.ftz.f32 %f672, %f643, %f671, 0fC0400000; fma.rn.ftz.f32 %f673, %f643, %f672, 0f3F800000; mul.ftz.f32 %f674, %f643, %f643; mul.ftz.f32 %f675, %f643, 0f40400000; add.ftz.f32 %f676, %f675, 0fC0C00000; fma.rn.ftz.f32 %f677, %f674, %f676, 0f40800000; mul.ftz.f32 %f678, %f677, 0f3E2AAAAB; fma.rn.ftz.f32 %f29, %f673, 0f3E2AAAAB, %f678; sub.ftz.f32 %f679, %f644, %f675; fma.rn.ftz.f32 %f680, %f643, %f679, 0f40400000; fma.rn.ftz.f32 %f681, %f643, %f680, 0f3F800000; div.approx.ftz.f32 %f684, %f652, %f25; add.ftz.f32 %f685, %f684, 0fBF800000; add.ftz.f32 %f31, %f636, %f685; div.approx.ftz.f32 %f686, %f657, %f26; add.ftz.f32 %f687, %f686, 0f3F800000; add.ftz.f32 %f32, %f636, %f687; div.approx.ftz.f32 %f688, %f665, %f27; add.ftz.f32 %f689, %f688, 0fBF800000; add.ftz.f32 %f33, %f638, %f689; div.approx.ftz.f32 %f692, %f678, %f29; add.ftz.f32 %f693, %f692, 0fBF800000; add.ftz.f32 %f35, %f640, %f693; cvt.rmi.ftz.f32.f32 %f37, %f31; cvt.rmi.ftz.f32.f32 %f696, %f33; cvt.rmi.ftz.f32.f32 %f697, %f35; sub.ftz.f32 %f38, %f33, %f696; sub.ftz.f32 %f39, %f35, %f697; cvt.rzi.ftz.s32.f32 %r7, %f37; cvt.rzi.ftz.s32.f32 %r8, %f696; cvt.rzi.ftz.s32.f32 %r9, %f697; and.b32 %r10, %r7, -8; and.b32 %r2377, %r7, -128; and.b32 %r12, %r7, -4096; add.s64 %rd17, %rd1, 96; shr.u32 %r772, %r9, 12; cvt.u64.u32 %rd18, %r772; shr.u32 %r773, %r8, 12; mul.wide.u32 %rd19, %r773, 2097152; shr.u32 %r774, %r7, 12; cvt.u64.u32 %rd510, %r774; bfi.b64 %rd21, %rd510, %rd18, 42, 32; ld.u32 %r13, [%rd5]; setp.eq.s32 %p10, %r13, 0; @%p10 bra $L__BB4_22; or.b64 %rd22, %rd21, %rd19; mov.u32 %r2276, 0; $L__BB4_19: cvt.u64.u32 %rd23, %r2276; mul.wide.u32 %rd511, %r2276, 32; add.s64 %rd512, %rd17, %rd511; ld.u64 %rd513, [%rd512]; setp.eq.s64 %p11, %rd513, %rd22; add.s32 %r2276, %r2276, 1; @%p11 bra $L__BB4_21; setp.lt.u32 %p12, %r2276, %r13; @%p12 bra $L__BB4_19; bra.uni $L__BB4_22; $L__BB4_6: shl.b64 %rd481, %rd8, 5; add.s64 %rd1615, %rd6, %rd481; $L__BB4_7: setp.eq.s64 %p6, %rd1615, 0; @%p6 bra $L__BB4_15; add.s64 %rd11, %rd1615, 8; ld.u64 %rd12, [%rd1615+8]; setp.eq.s64 %p7, %rd12, 0; @%p7 bra $L__BB4_14; add.s64 %rd13, %rd1, %rd12; and.b32 %r738, %r2, -4096; and.b32 %r739, %r1, -4096; mov.u64 %rd482, 1; st.local.v2.u32 [%rd4+16], {%r739, %r738}; and.b32 %r740, %r3, -4096; st.local.u32 [%rd4+24], %r740; st.local.u64 [%rd4+56], %rd13; shl.b32 %r741, %r1, 3; and.b32 %r742, %r741, 31744; and.b32 %r743, %r2, 3968; shr.u32 %r744, %r743, 2; or.b32 %r745, %r744, %r742; bfe.u32 %r746, %r3, 7, 5; or.b32 %r747, %r746, %r745; shr.u32 %r748, %r745, 6; mul.wide.u32 %rd483, %r748, 8; add.s64 %rd484, %rd13, %rd483; and.b32 %r749, %r747, 63; shl.b64 %rd485, %rd482, %r749; ld.u64 %rd486, [%rd484+4128]; and.b64 %rd487, %rd486, %rd485; setp.eq.s64 %p8, %rd487, 0; mul.wide.u32 %rd488, %r747, 16; add.s64 %rd489, %rd13, %rd488; add.s64 %rd14, %rd489, 8256; @%p8 bra $L__BB4_13; ld.u64 %rd490, [%rd14]; add.s64 %rd15, %rd13, %rd490; and.b32 %r750, %r1, -128; st.local.u32 [%rd4+4], %r750; and.b32 %r751, %r2, -128; st.local.u32 [%rd4+8], %r751; and.b32 %r752, %r3, -128; st.local.u32 [%rd4+12], %r752; st.local.u64 [%rd4+48], %rd15; shl.b32 %r753, %r1, 5; and.b32 %r754, %r753, 3840; shl.b32 %r755, %r2, 1; and.b32 %r756, %r755, 240; or.b32 %r757, %r756, %r754; bfe.u32 %r758, %r3, 3, 4; or.b32 %r759, %r758, %r757; shr.u32 %r760, %r757, 6; mul.wide.u32 %rd491, %r760, 8; add.s64 %rd492, %rd15, %rd491; and.b32 %r761, %r759, 63; shl.b64 %rd494, %rd482, %r761; ld.u64 %rd495, [%rd492+544]; and.b64 %rd496, %rd495, %rd494; setp.eq.s64 %p9, %rd496, 0; mul.wide.u32 %rd497, %r759, 16; add.s64 %rd498, %rd15, %rd497; add.s64 %rd16, %rd498, 1088; @%p9 bra $L__BB4_12; ld.u64 %rd499, [%rd16]; add.s64 %rd500, %rd15, %rd499; and.b32 %r762, %r2, -8; and.b32 %r763, %r1, -8; st.local.v2.u32 [%rd2], {%r763, %r762}; and.b32 %r764, %r3, -8; st.local.u32 [%rd4], %r764; st.local.u64 [%rd4+40], %rd500; shl.b32 %r765, %r1, 6; and.b32 %r766, %r765, 448; shl.b32 %r767, %r2, 3; and.b32 %r768, %r767, 56; or.b32 %r769, %r768, %r766; and.b32 %r770, %r3, 7; or.b32 %r771, %r769, %r770; mul.wide.u32 %rd501, %r771, 12; add.s64 %rd502, %rd500, %rd501; ld.f32 %f957, [%rd502+128]; ld.f32 %f958, [%rd502+132]; ld.f32 %f959, [%rd502+136]; bra.uni $L__BB4_554; $L__BB4_15: ld.f32 %f957, [%rd5+4]; ld.f32 %f958, [%rd5+8]; ld.f32 %f959, [%rd5+12]; bra.uni $L__BB4_554; $L__BB4_21: shl.b64 %rd515, %rd23, 5; add.s64 %rd1616, %rd17, %rd515; $L__BB4_22: setp.eq.s64 %p13, %rd1616, 0; @%p13 bra $L__BB4_30; add.s64 %rd26, %rd1616, 8; ld.u64 %rd27, [%rd1616+8]; setp.eq.s64 %p14, %rd27, 0; @%p14 bra $L__BB4_29; add.s64 %rd1636, %rd1, %rd27; and.b32 %r776, %r8, -4096; mov.u64 %rd516, 1; st.local.v2.u32 [%rd4+16], {%r12, %r776}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r777, %r7, 3; and.b32 %r778, %r777, 31744; and.b32 %r779, %r8, 3968; shr.u32 %r780, %r779, 2; or.b32 %r781, %r780, %r778; bfe.u32 %r782, %r9, 7, 5; or.b32 %r783, %r781, %r782; shr.u32 %r784, %r781, 6; mul.wide.u32 %rd517, %r784, 8; add.s64 %rd518, %rd1636, %rd517; and.b32 %r785, %r783, 63; shl.b64 %rd519, %rd516, %r785; ld.u64 %rd520, [%rd518+4128]; and.b64 %rd521, %rd520, %rd519; setp.eq.s64 %p15, %rd521, 0; mov.b64 %rd522, {%r12, %r776}; shr.u64 %rd523, %rd522, 32; cvt.u32.u64 %r2382, %rd523; mul.wide.u32 %rd524, %r783, 16; add.s64 %rd525, %rd1636, %rd524; add.s64 %rd29, %rd525, 8256; @%p15 bra $L__BB4_28; ld.u64 %rd526, [%rd29]; add.s64 %rd1628, %rd1636, %rd526; st.local.u32 [%rd4+4], %r2377; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r786, %r7, 5; and.b32 %r787, %r786, 3840; shl.b32 %r788, %r8, 1; and.b32 %r789, %r788, 240; or.b32 %r790, %r789, %r787; bfe.u32 %r791, %r9, 3, 4; or.b32 %r792, %r790, %r791; shr.u32 %r793, %r790, 6; mul.wide.u32 %rd527, %r793, 8; add.s64 %rd528, %rd1628, %rd527; and.b32 %r794, %r792, 63; shl.b64 %rd530, %rd516, %r794; ld.u64 %rd531, [%rd528+544]; and.b64 %rd532, %rd531, %rd530; setp.eq.s64 %p16, %rd532, 0; mul.wide.u32 %rd533, %r792, 16; add.s64 %rd534, %rd1628, %rd533; add.s64 %rd31, %rd534, 1088; @%p16 bra $L__BB4_27; ld.u64 %rd535, [%rd31]; add.s64 %rd1621, %rd1628, %rd535; and.b32 %r795, %r8, -8; st.local.v2.u32 [%rd2], {%r10, %r795}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r796, %r7, 6; and.b32 %r797, %r796, 448; shl.b32 %r798, %r8, 3; and.b32 %r799, %r798, 56; or.b32 %r800, %r799, %r797; and.b32 %r801, %r9, 7; or.b32 %r802, %r800, %r801; mul.wide.u32 %rd536, %r802, 12; add.s64 %rd537, %rd1621, %rd536; ld.f32 %f911, [%rd537+128]; ld.f32 %f910, [%rd537+132]; ld.f32 %f909, [%rd537+136]; mov.b64 %rd538, {%r10, %r795}; shr.u64 %rd539, %rd538, 32; cvt.u32.u64 %r2289, %rd539; mov.u32 %r2337, %r12; mov.u32 %r2334, %r10; bra.uni $L__BB4_31; $L__BB4_30: ld.f32 %f911, [%rd5+4]; mov.u64 %rd1621, 0; ld.f32 %f910, [%rd5+8]; ld.f32 %f909, [%rd5+12]; mov.u32 %r2382, 2147483647; mov.u32 %r2337, %r2382; mov.u32 %r2289, %r2382; mov.u32 %r2415, %r2382; mov.u32 %r2375, %r2382; mov.u32 %r2376, %r2382; mov.u32 %r2377, %r2382; mov.u32 %r2332, %r2382; mov.u64 %rd1628, %rd1621; mov.u64 %rd1636, %rd1621; mov.u32 %r2334, %r2382; bra.uni $L__BB4_31; $L__BB4_14: ld.f32 %f957, [%rd11+12]; ld.f32 %f958, [%rd11+16]; ld.f32 %f959, [%rd11+20]; bra.uni $L__BB4_554; $L__BB4_29: ld.f32 %f911, [%rd26+12]; mov.u64 %rd1621, 0; ld.f32 %f910, [%rd26+16]; ld.f32 %f909, [%rd26+20]; mov.u32 %r2382, 2147483647; mov.u32 %r2337, %r2382; mov.u32 %r2289, %r2382; mov.u32 %r2415, %r2382; mov.u32 %r2375, %r2382; mov.u32 %r2376, %r2382; mov.u32 %r2377, %r2382; mov.u32 %r2332, %r2382; mov.u64 %rd1628, %rd1621; mov.u64 %rd1636, %rd1621; mov.u32 %r2334, %r2382; bra.uni $L__BB4_31; $L__BB4_13: ld.f32 %f957, [%rd14]; ld.f32 %f958, [%rd14+4]; ld.f32 %f959, [%rd14+8]; bra.uni $L__BB4_554; $L__BB4_28: ld.f32 %f911, [%rd29]; mov.u64 %rd1621, 0; ld.f32 %f910, [%rd29+4]; ld.f32 %f909, [%rd29+8]; mov.u32 %r2289, 2147483647; mov.u32 %r2337, %r12; mov.u32 %r2375, %r2289; mov.u32 %r2376, %r2289; mov.u32 %r2377, %r2289; mov.u32 %r2332, %r2289; mov.u64 %rd1628, %rd1621; mov.u32 %r2334, %r2289; bra.uni $L__BB4_31; $L__BB4_12: ld.f32 %f957, [%rd16]; ld.f32 %f958, [%rd16+4]; ld.f32 %f959, [%rd16+8]; bra.uni $L__BB4_554; $L__BB4_27: ld.f32 %f911, [%rd31]; mov.u64 %rd1621, 0; ld.f32 %f910, [%rd31+4]; ld.f32 %f909, [%rd31+8]; mov.u32 %r2337, %r12; mov.u32 %r2332, %r2289; mov.u32 %r2334, %r2289; $L__BB4_31: add.s32 %r74, %r9, 1; setp.ne.s32 %p17, %r10, %r2334; mov.u16 %rs169, 0; @%p17 bra $L__BB4_34; and.b32 %r942, %r8, -8; setp.ne.s32 %p18, %r942, %r2289; @%p18 bra $L__BB4_34; and.b32 %r943, %r74, -8; setp.eq.s32 %p19, %r943, %r2332; selp.u16 %rs169, 1, 0, %p19; $L__BB4_34: setp.eq.s16 %p20, %rs169, 0; @%p20 bra $L__BB4_36; shl.b32 %r944, %r7, 6; and.b32 %r945, %r944, 448; shl.b32 %r946, %r8, 3; and.b32 %r947, %r946, 56; or.b32 %r948, %r947, %r945; and.b32 %r949, %r74, 7; or.b32 %r950, %r948, %r949; mul.wide.u32 %rd554, %r950, 12; add.s64 %rd555, %rd1621, %rd554; ld.f32 %f914, [%rd555+128]; ld.f32 %f913, [%rd555+132]; ld.f32 %f912, [%rd555+136]; $L__BB4_66: and.b32 %r2257, %r7, -8; sub.ftz.f32 %f698, %f914, %f911; sub.ftz.f32 %f699, %f913, %f910; sub.ftz.f32 %f700, %f912, %f909; fma.rn.ftz.f32 %f94, %f39, %f698, %f911; fma.rn.ftz.f32 %f95, %f39, %f699, %f910; fma.rn.ftz.f32 %f96, %f39, %f700, %f909; add.s32 %r139, %r8, 1; setp.ne.s32 %p39, %r2257, %r2334; mov.u16 %rs172, 0; @%p39 bra $L__BB4_69; and.b32 %r1027, %r139, -8; setp.ne.s32 %p40, %r1027, %r2289; @%p40 bra $L__BB4_69; and.b32 %r1028, %r74, -8; setp.eq.s32 %p41, %r1028, %r2332; selp.u16 %rs172, 1, 0, %p41; $L__BB4_69: setp.eq.s16 %p42, %rs172, 0; @%p42 bra $L__BB4_71; shl.b32 %r1029, %r7, 6; and.b32 %r1030, %r1029, 448; shl.b32 %r1031, %r139, 3; and.b32 %r1032, %r1031, 56; or.b32 %r1033, %r1032, %r1030; and.b32 %r1034, %r74, 7; or.b32 %r1035, %r1033, %r1034; mul.wide.u32 %rd623, %r1035, 12; add.s64 %rd624, %rd1621, %rd623; ld.f32 %f917, [%rd624+128]; ld.f32 %f916, [%rd624+132]; ld.f32 %f915, [%rd624+136]; $L__BB4_101: and.b32 %r2226, %r7, -8; setp.ne.s32 %p61, %r2226, %r2334; mov.u16 %rs175, 0; @%p61 bra $L__BB4_104; and.b32 %r1113, %r139, -8; setp.ne.s32 %p62, %r1113, %r2289; @%p62 bra $L__BB4_104; and.b32 %r1114, %r9, -8; setp.eq.s32 %p63, %r1114, %r2332; selp.u16 %rs175, 1, 0, %p63; $L__BB4_104: setp.eq.s16 %p64, %rs175, 0; @%p64 bra $L__BB4_106; and.b32 %r1115, %r7, 7; shl.b32 %r1116, %r139, 3; and.b32 %r1117, %r1116, 56; and.b32 %r1118, %r9, 7; bfi.b32 %r1119, %r1115, %r1118, 6, 3; or.b32 %r1120, %r1119, %r1117; mul.wide.u32 %rd693, %r1120, 12; add.s64 %rd694, %rd1621, %rd693; ld.f32 %f920, [%rd694+128]; ld.f32 %f919, [%rd694+132]; ld.f32 %f918, [%rd694+136]; $L__BB4_136: sub.ftz.f32 %f701, %f917, %f920; sub.ftz.f32 %f702, %f916, %f919; sub.ftz.f32 %f703, %f915, %f918; fma.rn.ftz.f32 %f704, %f39, %f701, %f920; fma.rn.ftz.f32 %f705, %f39, %f702, %f919; fma.rn.ftz.f32 %f706, %f39, %f703, %f918; sub.ftz.f32 %f707, %f704, %f94; sub.ftz.f32 %f708, %f705, %f95; sub.ftz.f32 %f709, %f706, %f96; fma.rn.ftz.f32 %f169, %f38, %f707, %f94; fma.rn.ftz.f32 %f170, %f38, %f708, %f95; fma.rn.ftz.f32 %f171, %f38, %f709, %f96; add.s32 %r259, %r7, 1; and.b32 %r260, %r259, -8; setp.ne.s32 %p83, %r260, %r2334; mov.u16 %rs178, 0; @%p83 bra $L__BB4_139; and.b32 %r1194, %r139, -8; setp.ne.s32 %p84, %r1194, %r2289; @%p84 bra $L__BB4_139; and.b32 %r1195, %r9, -8; setp.eq.s32 %p85, %r1195, %r2332; selp.u16 %rs178, 1, 0, %p85; $L__BB4_139: setp.eq.s16 %p86, %rs178, 0; @%p86 bra $L__BB4_141; and.b32 %r1196, %r259, 7; shl.b32 %r1197, %r139, 3; and.b32 %r1198, %r1197, 56; and.b32 %r1199, %r9, 7; bfi.b32 %r1200, %r1196, %r1199, 6, 3; or.b32 %r1201, %r1200, %r1198; mul.wide.u32 %rd761, %r1201, 12; add.s64 %rd762, %rd1621, %rd761; ld.f32 %f923, [%rd762+128]; ld.f32 %f922, [%rd762+132]; ld.f32 %f921, [%rd762+136]; $L__BB4_171: setp.ne.s32 %p105, %r260, %r2334; mov.u16 %rs181, 0; @%p105 bra $L__BB4_174; and.b32 %r1276, %r139, -8; setp.ne.s32 %p106, %r1276, %r2289; @%p106 bra $L__BB4_174; and.b32 %r1277, %r74, -8; setp.eq.s32 %p107, %r1277, %r2332; selp.u16 %rs181, 1, 0, %p107; $L__BB4_174: setp.eq.s16 %p108, %rs181, 0; @%p108 bra $L__BB4_176; shl.b32 %r1278, %r259, 6; and.b32 %r1279, %r1278, 448; shl.b32 %r1280, %r139, 3; and.b32 %r1281, %r1280, 56; or.b32 %r1282, %r1281, %r1279; and.b32 %r1283, %r74, 7; or.b32 %r1284, %r1282, %r1283; mul.wide.u32 %rd832, %r1284, 12; add.s64 %rd833, %rd1621, %rd832; ld.f32 %f926, [%rd833+128]; ld.f32 %f925, [%rd833+132]; ld.f32 %f924, [%rd833+136]; $L__BB4_206: sub.ftz.f32 %f710, %f926, %f923; sub.ftz.f32 %f711, %f925, %f922; sub.ftz.f32 %f712, %f924, %f921; fma.rn.ftz.f32 %f244, %f39, %f710, %f923; fma.rn.ftz.f32 %f245, %f39, %f711, %f922; fma.rn.ftz.f32 %f246, %f39, %f712, %f921; setp.ne.s32 %p127, %r260, %r2334; mov.u16 %rs184, 0; @%p127 bra $L__BB4_209; and.b32 %r1363, %r8, -8; setp.ne.s32 %p128, %r1363, %r2289; @%p128 bra $L__BB4_209; and.b32 %r1364, %r74, -8; setp.eq.s32 %p129, %r1364, %r2332; selp.u16 %rs184, 1, 0, %p129; $L__BB4_209: setp.eq.s16 %p130, %rs184, 0; @%p130 bra $L__BB4_211; shl.b32 %r1365, %r259, 6; and.b32 %r1366, %r1365, 448; shl.b32 %r1367, %r8, 3; and.b32 %r1368, %r1367, 56; or.b32 %r1369, %r1366, %r1368; and.b32 %r1370, %r74, 7; or.b32 %r1371, %r1369, %r1370; mul.wide.u32 %rd903, %r1371, 12; add.s64 %rd904, %rd1621, %rd903; ld.f32 %f929, [%rd904+128]; ld.f32 %f928, [%rd904+132]; ld.f32 %f927, [%rd904+136]; $L__BB4_241: setp.ne.s32 %p149, %r260, %r2334; mov.u16 %rs187, 0; @%p149 bra $L__BB4_244; and.b32 %r1449, %r8, -8; setp.ne.s32 %p150, %r1449, %r2289; @%p150 bra $L__BB4_244; and.b32 %r1450, %r9, -8; setp.eq.s32 %p151, %r1450, %r2332; selp.u16 %rs187, 1, 0, %p151; $L__BB4_244: setp.eq.s16 %p152, %rs187, 0; @%p152 bra $L__BB4_246; shl.b32 %r1451, %r259, 6; and.b32 %r1452, %r1451, 448; shl.b32 %r1453, %r8, 3; and.b32 %r1454, %r1453, 56; or.b32 %r1455, %r1452, %r1454; and.b32 %r1456, %r9, 7; or.b32 %r1457, %r1455, %r1456; mul.wide.u32 %rd974, %r1457, 12; add.s64 %rd975, %rd1621, %rd974; ld.f32 %f932, [%rd975+128]; ld.f32 %f931, [%rd975+132]; ld.f32 %f930, [%rd975+136]; $L__BB4_276: cvt.rmi.ftz.f32.f32 %f835, %f31; sub.ftz.f32 %f713, %f31, %f835; sub.ftz.f32 %f714, %f929, %f932; sub.ftz.f32 %f715, %f928, %f931; sub.ftz.f32 %f716, %f927, %f930; fma.rn.ftz.f32 %f717, %f39, %f714, %f932; fma.rn.ftz.f32 %f718, %f39, %f715, %f931; fma.rn.ftz.f32 %f719, %f39, %f716, %f930; sub.ftz.f32 %f720, %f244, %f717; sub.ftz.f32 %f721, %f245, %f718; sub.ftz.f32 %f722, %f246, %f719; fma.rn.ftz.f32 %f723, %f38, %f720, %f717; fma.rn.ftz.f32 %f724, %f38, %f721, %f718; fma.rn.ftz.f32 %f725, %f38, %f722, %f719; sub.ftz.f32 %f726, %f723, %f169; sub.ftz.f32 %f727, %f724, %f170; sub.ftz.f32 %f728, %f725, %f171; fma.rn.ftz.f32 %f319, %f713, %f726, %f169; fma.rn.ftz.f32 %f320, %f713, %f727, %f170; fma.rn.ftz.f32 %f321, %f713, %f728, %f171; cvt.rmi.ftz.f32.f32 %f322, %f32; cvt.rzi.ftz.s32.f32 %r471, %f322; and.b32 %r472, %r471, -8; setp.ne.s32 %p171, %r472, %r2334; mov.u16 %rs190, 0; @%p171 bra $L__BB4_279; and.b32 %r1534, %r8, -8; setp.ne.s32 %p172, %r1534, %r2289; @%p172 bra $L__BB4_279; and.b32 %r1535, %r9, -8; setp.eq.s32 %p173, %r1535, %r2332; selp.u16 %rs190, 1, 0, %p173; $L__BB4_279: setp.eq.s16 %p174, %rs190, 0; @%p174 bra $L__BB4_281; shl.b32 %r1536, %r471, 6; and.b32 %r1537, %r1536, 448; shl.b32 %r1538, %r8, 3; and.b32 %r1539, %r1538, 56; or.b32 %r1540, %r1539, %r1537; and.b32 %r1541, %r9, 7; or.b32 %r1542, %r1540, %r1541; mul.wide.u32 %rd1044, %r1542, 12; add.s64 %rd1045, %rd1621, %rd1044; ld.f32 %f935, [%rd1045+128]; ld.f32 %f934, [%rd1045+132]; ld.f32 %f933, [%rd1045+136]; $L__BB4_311: setp.ne.s32 %p193, %r472, %r2334; mov.u16 %rs193, 0; @%p193 bra $L__BB4_314; and.b32 %r1619, %r8, -8; setp.ne.s32 %p194, %r1619, %r2289; @%p194 bra $L__BB4_314; and.b32 %r1620, %r74, -8; setp.eq.s32 %p195, %r1620, %r2332; selp.u16 %rs193, 1, 0, %p195; $L__BB4_314: setp.eq.s16 %p196, %rs193, 0; @%p196 bra $L__BB4_316; shl.b32 %r1621, %r471, 6; and.b32 %r1622, %r1621, 448; shl.b32 %r1623, %r8, 3; and.b32 %r1624, %r1623, 56; or.b32 %r1625, %r1624, %r1622; and.b32 %r1626, %r74, 7; or.b32 %r1627, %r1625, %r1626; mul.wide.u32 %rd1114, %r1627, 12; add.s64 %rd1115, %rd1621, %rd1114; ld.f32 %f938, [%rd1115+128]; ld.f32 %f937, [%rd1115+132]; ld.f32 %f936, [%rd1115+136]; $L__BB4_346: sub.ftz.f32 %f729, %f938, %f935; sub.ftz.f32 %f730, %f937, %f934; sub.ftz.f32 %f731, %f936, %f933; fma.rn.ftz.f32 %f395, %f39, %f729, %f935; fma.rn.ftz.f32 %f396, %f39, %f730, %f934; fma.rn.ftz.f32 %f397, %f39, %f731, %f933; setp.ne.s32 %p215, %r472, %r2334; mov.u16 %rs196, 0; @%p215 bra $L__BB4_349; and.b32 %r1705, %r139, -8; setp.ne.s32 %p216, %r1705, %r2289; @%p216 bra $L__BB4_349; and.b32 %r1706, %r74, -8; setp.eq.s32 %p217, %r1706, %r2332; selp.u16 %rs196, 1, 0, %p217; $L__BB4_349: setp.eq.s16 %p218, %rs196, 0; @%p218 bra $L__BB4_351; shl.b32 %r1707, %r471, 6; and.b32 %r1708, %r1707, 448; shl.b32 %r1709, %r139, 3; and.b32 %r1710, %r1709, 56; or.b32 %r1711, %r1710, %r1708; and.b32 %r1712, %r74, 7; or.b32 %r1713, %r1711, %r1712; mul.wide.u32 %rd1185, %r1713, 12; add.s64 %rd1186, %rd1621, %rd1185; ld.f32 %f941, [%rd1186+128]; ld.f32 %f940, [%rd1186+132]; ld.f32 %f939, [%rd1186+136]; $L__BB4_381: setp.ne.s32 %p237, %r472, %r2334; mov.u16 %rs199, 0; @%p237 bra $L__BB4_384; and.b32 %r1792, %r139, -8; setp.ne.s32 %p238, %r1792, %r2289; @%p238 bra $L__BB4_384; and.b32 %r1793, %r9, -8; setp.eq.s32 %p239, %r1793, %r2332; selp.u16 %rs199, 1, 0, %p239; $L__BB4_384: setp.eq.s16 %p240, %rs199, 0; @%p240 bra $L__BB4_386; and.b32 %r1794, %r471, 7; shl.b32 %r1795, %r139, 3; and.b32 %r1796, %r1795, 56; and.b32 %r1797, %r9, 7; bfi.b32 %r1798, %r1794, %r1797, 6, 3; or.b32 %r1799, %r1798, %r1796; mul.wide.u32 %rd1256, %r1799, 12; add.s64 %rd1257, %rd1621, %rd1256; ld.f32 %f944, [%rd1257+128]; ld.f32 %f943, [%rd1257+132]; ld.f32 %f942, [%rd1257+136]; $L__BB4_416: sub.ftz.f32 %f732, %f941, %f944; sub.ftz.f32 %f733, %f940, %f943; sub.ftz.f32 %f734, %f939, %f942; fma.rn.ftz.f32 %f735, %f39, %f732, %f944; fma.rn.ftz.f32 %f736, %f39, %f733, %f943; fma.rn.ftz.f32 %f737, %f39, %f734, %f942; sub.ftz.f32 %f738, %f735, %f395; sub.ftz.f32 %f739, %f736, %f396; sub.ftz.f32 %f740, %f737, %f397; fma.rn.ftz.f32 %f470, %f38, %f738, %f395; fma.rn.ftz.f32 %f471, %f38, %f739, %f396; fma.rn.ftz.f32 %f472, %f38, %f740, %f397; add.s32 %r635, %r471, 1; and.b32 %r636, %r635, -8; setp.ne.s32 %p259, %r636, %r2334; mov.u16 %rs202, 0; @%p259 bra $L__BB4_419; and.b32 %r1874, %r139, -8; setp.ne.s32 %p260, %r1874, %r2289; @%p260 bra $L__BB4_419; and.b32 %r1875, %r9, -8; setp.eq.s32 %p261, %r1875, %r2332; selp.u16 %rs202, 1, 0, %p261; $L__BB4_419: setp.eq.s16 %p262, %rs202, 0; @%p262 bra $L__BB4_421; and.b32 %r1876, %r635, 7; shl.b32 %r1877, %r139, 3; and.b32 %r1878, %r1877, 56; and.b32 %r1879, %r9, 7; bfi.b32 %r1880, %r1876, %r1879, 6, 3; or.b32 %r1881, %r1880, %r1878; mul.wide.u32 %rd1327, %r1881, 12; add.s64 %rd1328, %rd1621, %rd1327; ld.f32 %f947, [%rd1328+128]; ld.f32 %f946, [%rd1328+132]; ld.f32 %f945, [%rd1328+136]; $L__BB4_451: setp.ne.s32 %p281, %r636, %r2334; mov.u16 %rs205, 0; @%p281 bra $L__BB4_454; and.b32 %r1956, %r139, -8; setp.ne.s32 %p282, %r1956, %r2289; @%p282 bra $L__BB4_454; and.b32 %r1957, %r74, -8; setp.eq.s32 %p283, %r1957, %r2332; selp.u16 %rs205, 1, 0, %p283; $L__BB4_454: setp.eq.s16 %p284, %rs205, 0; @%p284 bra $L__BB4_456; shl.b32 %r1958, %r635, 6; and.b32 %r1959, %r1958, 448; shl.b32 %r1960, %r139, 3; and.b32 %r1961, %r1960, 56; or.b32 %r1962, %r1961, %r1959; and.b32 %r1963, %r74, 7; or.b32 %r1964, %r1962, %r1963; mul.wide.u32 %rd1398, %r1964, 12; add.s64 %rd1399, %rd1621, %rd1398; ld.f32 %f950, [%rd1399+128]; ld.f32 %f949, [%rd1399+132]; ld.f32 %f948, [%rd1399+136]; $L__BB4_486: sub.ftz.f32 %f741, %f950, %f947; sub.ftz.f32 %f742, %f949, %f946; sub.ftz.f32 %f743, %f948, %f945; fma.rn.ftz.f32 %f545, %f39, %f741, %f947; fma.rn.ftz.f32 %f546, %f39, %f742, %f946; fma.rn.ftz.f32 %f547, %f39, %f743, %f945; setp.ne.s32 %p303, %r636, %r2334; mov.u16 %rs208, 0; @%p303 bra $L__BB4_489; and.b32 %r2043, %r8, -8; setp.ne.s32 %p304, %r2043, %r2289; @%p304 bra $L__BB4_489; and.b32 %r2044, %r74, -8; setp.eq.s32 %p305, %r2044, %r2332; selp.u16 %rs208, 1, 0, %p305; $L__BB4_489: setp.eq.s16 %p306, %rs208, 0; @%p306 bra $L__BB4_491; shl.b32 %r2045, %r635, 6; and.b32 %r2046, %r2045, 448; shl.b32 %r2047, %r8, 3; and.b32 %r2048, %r2047, 56; or.b32 %r2049, %r2046, %r2048; and.b32 %r2050, %r74, 7; or.b32 %r2051, %r2049, %r2050; mul.wide.u32 %rd1469, %r2051, 12; add.s64 %rd1470, %rd1621, %rd1469; ld.f32 %f953, [%rd1470+128]; ld.f32 %f952, [%rd1470+132]; ld.f32 %f951, [%rd1470+136]; $L__BB4_521: setp.ne.s32 %p325, %r636, %r2334; @%p325 bra $L__BB4_524; and.b32 %r2129, %r8, -8; setp.ne.s32 %p326, %r2129, %r2289; @%p326 bra $L__BB4_524; and.b32 %r2130, %r9, -8; setp.eq.s32 %p327, %r2130, %r2332; @%p327 bra $L__BB4_552; bra.uni $L__BB4_524; $L__BB4_552: shl.b32 %r2215, %r635, 6; and.b32 %r2216, %r2215, 448; shl.b32 %r2217, %r8, 3; and.b32 %r2218, %r2217, 56; or.b32 %r2219, %r2216, %r2218; and.b32 %r2220, %r9, 7; or.b32 %r2221, %r2219, %r2220; mul.wide.u32 %rd1602, %r2221, 12; add.s64 %rd1603, %rd1621, %rd1602; ld.f32 %f956, [%rd1603+128]; ld.f32 %f955, [%rd1603+132]; ld.f32 %f954, [%rd1603+136]; bra.uni $L__BB4_553; $L__BB4_524: and.b32 %r727, %r635, -128; setp.ne.s32 %p328, %r727, %r2377; @%p328 bra $L__BB4_527; and.b32 %r2131, %r8, -128; setp.ne.s32 %p329, %r2131, %r2376; @%p329 bra $L__BB4_527; and.b32 %r2132, %r9, -128; setp.eq.s32 %p330, %r2132, %r2375; @%p330 bra $L__BB4_549; bra.uni $L__BB4_527; $L__BB4_549: shl.b32 %r2197, %r635, 5; and.b32 %r2198, %r2197, 3840; shl.b32 %r2199, %r8, 1; and.b32 %r2200, %r2199, 240; or.b32 %r2201, %r2198, %r2200; bfe.u32 %r2202, %r9, 3, 4; or.b32 %r2203, %r2201, %r2202; shr.u32 %r2204, %r2201, 6; mul.wide.u32 %rd1590, %r2204, 8; add.s64 %rd1591, %rd1628, %rd1590; and.b32 %r2205, %r2203, 63; mov.u64 %rd1592, 1; shl.b64 %rd1593, %rd1592, %r2205; ld.u64 %rd1594, [%rd1591+544]; and.b64 %rd1595, %rd1594, %rd1593; setp.eq.s64 %p343, %rd1595, 0; mul.wide.u32 %rd1596, %r2203, 16; add.s64 %rd1597, %rd1628, %rd1596; add.s64 %rd463, %rd1597, 1088; @%p343 bra $L__BB4_551; ld.u64 %rd1598, [%rd463]; add.s64 %rd1599, %rd1628, %rd1598; and.b32 %r2206, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2206}; and.b32 %r2207, %r9, -8; st.local.u32 [%rd4], %r2207; st.local.u64 [%rd4+40], %rd1599; shl.b32 %r2208, %r635, 6; and.b32 %r2209, %r2208, 448; shl.b32 %r2210, %r8, 3; and.b32 %r2211, %r2210, 56; or.b32 %r2212, %r2209, %r2211; and.b32 %r2213, %r9, 7; or.b32 %r2214, %r2212, %r2213; mul.wide.u32 %rd1600, %r2214, 12; add.s64 %rd1601, %rd1599, %rd1600; ld.f32 %f956, [%rd1601+128]; ld.f32 %f955, [%rd1601+132]; ld.f32 %f954, [%rd1601+136]; bra.uni $L__BB4_553; $L__BB4_527: and.b32 %r728, %r635, -4096; setp.ne.s32 %p331, %r728, %r2337; @%p331 bra $L__BB4_530; and.b32 %r2133, %r8, -4096; setp.ne.s32 %p332, %r2133, %r2382; @%p332 bra $L__BB4_530; and.b32 %r2134, %r9, -4096; setp.eq.s32 %p333, %r2134, %r2415; @%p333 bra $L__BB4_544; bra.uni $L__BB4_530; $L__BB4_544: shl.b32 %r2168, %r635, 3; and.b32 %r2169, %r2168, 31744; and.b32 %r2170, %r8, 3968; shr.u32 %r2171, %r2170, 2; or.b32 %r2172, %r2169, %r2171; bfe.u32 %r2173, %r9, 7, 5; or.b32 %r2174, %r2172, %r2173; shr.u32 %r2175, %r2172, 6; mul.wide.u32 %rd1569, %r2175, 8; add.s64 %rd1570, %rd1636, %rd1569; and.b32 %r2176, %r2174, 63; mov.u64 %rd1571, 1; shl.b64 %rd1572, %rd1571, %r2176; ld.u64 %rd1573, [%rd1570+4128]; and.b64 %rd1574, %rd1573, %rd1572; setp.eq.s64 %p341, %rd1574, 0; mul.wide.u32 %rd1575, %r2174, 16; add.s64 %rd1576, %rd1636, %rd1575; add.s64 %rd460, %rd1576, 8256; @%p341 bra $L__BB4_548; ld.u64 %rd1577, [%rd460]; add.s64 %rd461, %rd1636, %rd1577; st.local.u32 [%rd4+4], %r727; and.b32 %r2177, %r8, -128; st.local.u32 [%rd4+8], %r2177; and.b32 %r2178, %r9, -128; st.local.u32 [%rd4+12], %r2178; st.local.u64 [%rd4+48], %rd461; shl.b32 %r2179, %r635, 5; and.b32 %r2180, %r2179, 3840; shl.b32 %r2181, %r8, 1; and.b32 %r2182, %r2181, 240; or.b32 %r2183, %r2180, %r2182; bfe.u32 %r2184, %r9, 3, 4; or.b32 %r2185, %r2183, %r2184; shr.u32 %r2186, %r2183, 6; mul.wide.u32 %rd1578, %r2186, 8; add.s64 %rd1579, %rd461, %rd1578; and.b32 %r2187, %r2185, 63; shl.b64 %rd1581, %rd1571, %r2187; ld.u64 %rd1582, [%rd1579+544]; and.b64 %rd1583, %rd1582, %rd1581; setp.eq.s64 %p342, %rd1583, 0; mul.wide.u32 %rd1584, %r2185, 16; add.s64 %rd1585, %rd461, %rd1584; add.s64 %rd462, %rd1585, 1088; @%p342 bra $L__BB4_547; ld.u64 %rd1586, [%rd462]; add.s64 %rd1587, %rd461, %rd1586; and.b32 %r2188, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2188}; and.b32 %r2189, %r9, -8; st.local.u32 [%rd4], %r2189; st.local.u64 [%rd4+40], %rd1587; shl.b32 %r2190, %r635, 6; and.b32 %r2191, %r2190, 448; shl.b32 %r2192, %r8, 3; and.b32 %r2193, %r2192, 56; or.b32 %r2194, %r2191, %r2193; and.b32 %r2195, %r9, 7; or.b32 %r2196, %r2194, %r2195; mul.wide.u32 %rd1588, %r2196, 12; add.s64 %rd1589, %rd1587, %rd1588; ld.f32 %f956, [%rd1589+128]; ld.f32 %f955, [%rd1589+132]; ld.f32 %f954, [%rd1589+136]; bra.uni $L__BB4_553; $L__BB4_530: or.b64 %rd449, %rd19, %rd18; ld.u32 %r729, [%rd5]; mov.u64 %rd1704, 0; setp.eq.s32 %p334, %r729, 0; @%p334 bra $L__BB4_535; shr.u32 %r2136, %r635, 12; cvt.u64.u32 %rd1541, %r2136; shl.b64 %rd1542, %rd1541, 42; or.b64 %rd450, %rd449, %rd1542; mov.u32 %r2699, 0; $L__BB4_532: cvt.u64.u32 %rd451, %r2699; mul.wide.u32 %rd1543, %r2699, 32; add.s64 %rd1544, %rd17, %rd1543; ld.u64 %rd1545, [%rd1544]; setp.eq.s64 %p335, %rd1545, %rd450; add.s32 %r2699, %r2699, 1; @%p335 bra $L__BB4_534; setp.lt.u32 %p336, %r2699, %r729; @%p336 bra $L__BB4_532; bra.uni $L__BB4_535; $L__BB4_36: and.b32 %r2271, %r7, -128; setp.ne.s32 %p21, %r2271, %r2377; mov.u16 %rs170, 0; @%p21 bra $L__BB4_39; and.b32 %r951, %r8, -128; setp.ne.s32 %p22, %r951, %r2376; @%p22 bra $L__BB4_39; and.b32 %r952, %r74, -128; setp.eq.s32 %p23, %r952, %r2375; selp.u16 %rs170, 1, 0, %p23; $L__BB4_39: setp.eq.s16 %p24, %rs170, 0; @%p24 bra $L__BB4_43; shl.b32 %r953, %r7, 5; and.b32 %r954, %r953, 3840; shl.b32 %r955, %r8, 1; and.b32 %r956, %r955, 240; or.b32 %r957, %r956, %r954; bfe.u32 %r958, %r74, 3, 4; or.b32 %r959, %r958, %r957; shr.u32 %r960, %r957, 6; mul.wide.u32 %rd556, %r960, 8; add.s64 %rd557, %rd1628, %rd556; and.b32 %r961, %r959, 63; mov.u64 %rd558, 1; shl.b64 %rd559, %rd558, %r961; ld.u64 %rd560, [%rd557+544]; and.b64 %rd561, %rd560, %rd559; setp.eq.s64 %p25, %rd561, 0; mul.wide.u32 %rd562, %r959, 16; add.s64 %rd563, %rd1628, %rd562; add.s64 %rd43, %rd563, 1088; @%p25 bra $L__BB4_42; and.b32 %r2334, %r7, -8; ld.u64 %rd564, [%rd43]; add.s64 %rd1621, %rd1628, %rd564; and.b32 %r962, %r8, -8; st.local.v2.u32 [%rd2], {%r2334, %r962}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r963, %r7, 6; and.b32 %r964, %r963, 448; shl.b32 %r965, %r8, 3; and.b32 %r966, %r965, 56; or.b32 %r967, %r966, %r964; and.b32 %r968, %r74, 7; or.b32 %r969, %r967, %r968; mul.wide.u32 %rd565, %r969, 12; add.s64 %rd566, %rd1621, %rd565; ld.f32 %f914, [%rd566+128]; ld.f32 %f913, [%rd566+132]; ld.f32 %f912, [%rd566+136]; mov.b64 %rd567, {%r2334, %r962}; shr.u64 %rd568, %rd567, 32; cvt.u32.u64 %r2289, %rd568; bra.uni $L__BB4_66; $L__BB4_71: and.b32 %r2269, %r7, -128; setp.ne.s32 %p43, %r2269, %r2377; mov.u16 %rs173, 0; @%p43 bra $L__BB4_74; and.b32 %r1036, %r139, -128; setp.ne.s32 %p44, %r1036, %r2376; @%p44 bra $L__BB4_74; and.b32 %r1037, %r74, -128; setp.eq.s32 %p45, %r1037, %r2375; selp.u16 %rs173, 1, 0, %p45; $L__BB4_74: setp.eq.s16 %p46, %rs173, 0; @%p46 bra $L__BB4_78; shl.b32 %r1038, %r7, 5; and.b32 %r1039, %r1038, 3840; shl.b32 %r1040, %r139, 1; and.b32 %r1041, %r1040, 240; or.b32 %r1042, %r1041, %r1039; bfe.u32 %r1043, %r74, 3, 4; or.b32 %r1044, %r1042, %r1043; shr.u32 %r1045, %r1042, 6; mul.wide.u32 %rd625, %r1045, 8; add.s64 %rd626, %rd1628, %rd625; and.b32 %r1046, %r1044, 63; mov.u64 %rd627, 1; shl.b64 %rd628, %rd627, %r1046; ld.u64 %rd629, [%rd626+544]; and.b64 %rd630, %rd629, %rd628; setp.eq.s64 %p47, %rd630, 0; mul.wide.u32 %rd631, %r1044, 16; add.s64 %rd632, %rd1628, %rd631; add.s64 %rd72, %rd632, 1088; @%p47 bra $L__BB4_77; and.b32 %r2334, %r7, -8; ld.u64 %rd633, [%rd72]; add.s64 %rd1621, %rd1628, %rd633; and.b32 %r1047, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1047}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1048, %r7, 6; and.b32 %r1049, %r1048, 448; shl.b32 %r1050, %r139, 3; and.b32 %r1051, %r1050, 56; or.b32 %r1052, %r1051, %r1049; and.b32 %r1053, %r74, 7; or.b32 %r1054, %r1052, %r1053; mul.wide.u32 %rd634, %r1054, 12; add.s64 %rd635, %rd1621, %rd634; ld.f32 %f917, [%rd635+128]; ld.f32 %f916, [%rd635+132]; ld.f32 %f915, [%rd635+136]; mov.b64 %rd636, {%r2334, %r1047}; shr.u64 %rd637, %rd636, 32; cvt.u32.u64 %r2289, %rd637; bra.uni $L__BB4_101; $L__BB4_106: and.b32 %r2262, %r7, -128; setp.ne.s32 %p65, %r2262, %r2377; mov.u16 %rs176, 0; @%p65 bra $L__BB4_109; and.b32 %r1121, %r139, -128; setp.ne.s32 %p66, %r1121, %r2376; @%p66 bra $L__BB4_109; and.b32 %r1122, %r9, -128; setp.eq.s32 %p67, %r1122, %r2375; selp.u16 %rs176, 1, 0, %p67; $L__BB4_109: setp.eq.s16 %p68, %rs176, 0; @%p68 bra $L__BB4_113; shl.b32 %r1123, %r7, 5; and.b32 %r1124, %r1123, 3840; shl.b32 %r1125, %r139, 1; and.b32 %r1126, %r1125, 240; or.b32 %r1127, %r1126, %r1124; bfe.u32 %r1128, %r9, 3, 4; or.b32 %r1129, %r1127, %r1128; shr.u32 %r1130, %r1127, 6; mul.wide.u32 %rd695, %r1130, 8; add.s64 %rd696, %rd1628, %rd695; and.b32 %r1131, %r1129, 63; mov.u64 %rd697, 1; shl.b64 %rd698, %rd697, %r1131; ld.u64 %rd699, [%rd696+544]; and.b64 %rd700, %rd699, %rd698; setp.eq.s64 %p69, %rd700, 0; mul.wide.u32 %rd701, %r1129, 16; add.s64 %rd702, %rd1628, %rd701; add.s64 %rd101, %rd702, 1088; @%p69 bra $L__BB4_112; and.b32 %r2334, %r7, -8; ld.u64 %rd703, [%rd101]; add.s64 %rd1621, %rd1628, %rd703; and.b32 %r1132, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1132}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1133, %r7, 7; shl.b32 %r1134, %r139, 3; and.b32 %r1135, %r1134, 56; and.b32 %r1136, %r9, 7; bfi.b32 %r1137, %r1133, %r1136, 6, 3; or.b32 %r1138, %r1137, %r1135; mul.wide.u32 %rd704, %r1138, 12; add.s64 %rd705, %rd1621, %rd704; ld.f32 %f920, [%rd705+128]; ld.f32 %f919, [%rd705+132]; ld.f32 %f918, [%rd705+136]; mov.b64 %rd706, {%r2334, %r1132}; shr.u64 %rd707, %rd706, 32; cvt.u32.u64 %r2289, %rd707; bra.uni $L__BB4_136; $L__BB4_141: and.b32 %r261, %r259, -128; setp.ne.s32 %p87, %r261, %r2377; mov.u16 %rs179, 0; @%p87 bra $L__BB4_144; and.b32 %r1202, %r139, -128; setp.ne.s32 %p88, %r1202, %r2376; @%p88 bra $L__BB4_144; and.b32 %r1203, %r9, -128; setp.eq.s32 %p89, %r1203, %r2375; selp.u16 %rs179, 1, 0, %p89; $L__BB4_144: setp.eq.s16 %p90, %rs179, 0; @%p90 bra $L__BB4_148; shl.b32 %r1204, %r259, 5; and.b32 %r1205, %r1204, 3840; shl.b32 %r1206, %r139, 1; and.b32 %r1207, %r1206, 240; or.b32 %r1208, %r1207, %r1205; bfe.u32 %r1209, %r9, 3, 4; or.b32 %r1210, %r1208, %r1209; shr.u32 %r1211, %r1208, 6; mul.wide.u32 %rd763, %r1211, 8; add.s64 %rd764, %rd1628, %rd763; and.b32 %r1212, %r1210, 63; mov.u64 %rd765, 1; shl.b64 %rd766, %rd765, %r1212; ld.u64 %rd767, [%rd764+544]; and.b64 %rd768, %rd767, %rd766; setp.eq.s64 %p91, %rd768, 0; mul.wide.u32 %rd769, %r1210, 16; add.s64 %rd770, %rd1628, %rd769; add.s64 %rd130, %rd770, 1088; @%p91 bra $L__BB4_147; ld.u64 %rd771, [%rd130]; add.s64 %rd1621, %rd1628, %rd771; and.b32 %r1213, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1213}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1214, %r259, 7; shl.b32 %r1215, %r139, 3; and.b32 %r1216, %r1215, 56; and.b32 %r1217, %r9, 7; bfi.b32 %r1218, %r1214, %r1217, 6, 3; or.b32 %r1219, %r1218, %r1216; mul.wide.u32 %rd772, %r1219, 12; add.s64 %rd773, %rd1621, %rd772; ld.f32 %f923, [%rd773+128]; ld.f32 %f922, [%rd773+132]; ld.f32 %f921, [%rd773+136]; mov.b64 %rd774, {%r260, %r1213}; shr.u64 %rd775, %rd774, 32; cvt.u32.u64 %r2289, %rd775; mov.u32 %r2334, %r260; bra.uni $L__BB4_171; $L__BB4_176: and.b32 %r318, %r259, -128; setp.ne.s32 %p109, %r318, %r2377; mov.u16 %rs182, 0; @%p109 bra $L__BB4_179; and.b32 %r1285, %r139, -128; setp.ne.s32 %p110, %r1285, %r2376; @%p110 bra $L__BB4_179; and.b32 %r1286, %r74, -128; setp.eq.s32 %p111, %r1286, %r2375; selp.u16 %rs182, 1, 0, %p111; $L__BB4_179: setp.eq.s16 %p112, %rs182, 0; @%p112 bra $L__BB4_183; shl.b32 %r1287, %r259, 5; and.b32 %r1288, %r1287, 3840; shl.b32 %r1289, %r139, 1; and.b32 %r1290, %r1289, 240; or.b32 %r1291, %r1290, %r1288; bfe.u32 %r1292, %r74, 3, 4; or.b32 %r1293, %r1291, %r1292; shr.u32 %r1294, %r1291, 6; mul.wide.u32 %rd834, %r1294, 8; add.s64 %rd835, %rd1628, %rd834; and.b32 %r1295, %r1293, 63; mov.u64 %rd836, 1; shl.b64 %rd837, %rd836, %r1295; ld.u64 %rd838, [%rd835+544]; and.b64 %rd839, %rd838, %rd837; setp.eq.s64 %p113, %rd839, 0; mul.wide.u32 %rd840, %r1293, 16; add.s64 %rd841, %rd1628, %rd840; add.s64 %rd159, %rd841, 1088; @%p113 bra $L__BB4_182; ld.u64 %rd842, [%rd159]; add.s64 %rd1621, %rd1628, %rd842; and.b32 %r1296, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1296}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1297, %r259, 6; and.b32 %r1298, %r1297, 448; shl.b32 %r1299, %r139, 3; and.b32 %r1300, %r1299, 56; or.b32 %r1301, %r1300, %r1298; and.b32 %r1302, %r74, 7; or.b32 %r1303, %r1301, %r1302; mul.wide.u32 %rd843, %r1303, 12; add.s64 %rd844, %rd1621, %rd843; ld.f32 %f926, [%rd844+128]; ld.f32 %f925, [%rd844+132]; ld.f32 %f924, [%rd844+136]; mov.b64 %rd845, {%r260, %r1296}; shr.u64 %rd846, %rd845, 32; cvt.u32.u64 %r2289, %rd846; mov.u32 %r2334, %r260; bra.uni $L__BB4_206; $L__BB4_211: and.b32 %r372, %r259, -128; setp.ne.s32 %p131, %r372, %r2377; mov.u16 %rs185, 0; @%p131 bra $L__BB4_214; and.b32 %r1372, %r8, -128; setp.ne.s32 %p132, %r1372, %r2376; @%p132 bra $L__BB4_214; and.b32 %r1373, %r74, -128; setp.eq.s32 %p133, %r1373, %r2375; selp.u16 %rs185, 1, 0, %p133; $L__BB4_214: setp.eq.s16 %p134, %rs185, 0; @%p134 bra $L__BB4_218; shl.b32 %r1374, %r259, 5; and.b32 %r1375, %r1374, 3840; shl.b32 %r1376, %r8, 1; and.b32 %r1377, %r1376, 240; or.b32 %r1378, %r1375, %r1377; bfe.u32 %r1379, %r74, 3, 4; or.b32 %r1380, %r1379, %r1378; shr.u32 %r1381, %r1378, 6; mul.wide.u32 %rd905, %r1381, 8; add.s64 %rd906, %rd1628, %rd905; and.b32 %r1382, %r1380, 63; mov.u64 %rd907, 1; shl.b64 %rd908, %rd907, %r1382; ld.u64 %rd909, [%rd906+544]; and.b64 %rd910, %rd909, %rd908; setp.eq.s64 %p135, %rd910, 0; mul.wide.u32 %rd911, %r1380, 16; add.s64 %rd912, %rd1628, %rd911; add.s64 %rd188, %rd912, 1088; @%p135 bra $L__BB4_217; ld.u64 %rd913, [%rd188]; add.s64 %rd1621, %rd1628, %rd913; and.b32 %r1383, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1383}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1384, %r259, 6; and.b32 %r1385, %r1384, 448; shl.b32 %r1386, %r8, 3; and.b32 %r1387, %r1386, 56; or.b32 %r1388, %r1385, %r1387; and.b32 %r1389, %r74, 7; or.b32 %r1390, %r1388, %r1389; mul.wide.u32 %rd914, %r1390, 12; add.s64 %rd915, %rd1621, %rd914; ld.f32 %f929, [%rd915+128]; ld.f32 %f928, [%rd915+132]; ld.f32 %f927, [%rd915+136]; mov.b64 %rd916, {%r260, %r1383}; shr.u64 %rd917, %rd916, 32; cvt.u32.u64 %r2289, %rd917; mov.u32 %r2334, %r260; bra.uni $L__BB4_241; $L__BB4_246: and.b32 %r423, %r259, -128; setp.ne.s32 %p153, %r423, %r2377; mov.u16 %rs188, 0; @%p153 bra $L__BB4_249; and.b32 %r1458, %r8, -128; setp.ne.s32 %p154, %r1458, %r2376; @%p154 bra $L__BB4_249; and.b32 %r1459, %r9, -128; setp.eq.s32 %p155, %r1459, %r2375; selp.u16 %rs188, 1, 0, %p155; $L__BB4_249: setp.eq.s16 %p156, %rs188, 0; @%p156 bra $L__BB4_253; shl.b32 %r1460, %r259, 5; and.b32 %r1461, %r1460, 3840; shl.b32 %r1462, %r8, 1; and.b32 %r1463, %r1462, 240; or.b32 %r1464, %r1461, %r1463; bfe.u32 %r1465, %r9, 3, 4; or.b32 %r1466, %r1464, %r1465; shr.u32 %r1467, %r1464, 6; mul.wide.u32 %rd976, %r1467, 8; add.s64 %rd977, %rd1628, %rd976; and.b32 %r1468, %r1466, 63; mov.u64 %rd978, 1; shl.b64 %rd979, %rd978, %r1468; ld.u64 %rd980, [%rd977+544]; and.b64 %rd981, %rd980, %rd979; setp.eq.s64 %p157, %rd981, 0; mul.wide.u32 %rd982, %r1466, 16; add.s64 %rd983, %rd1628, %rd982; add.s64 %rd217, %rd983, 1088; @%p157 bra $L__BB4_252; ld.u64 %rd984, [%rd217]; add.s64 %rd1621, %rd1628, %rd984; and.b32 %r1469, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1469}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1470, %r259, 6; and.b32 %r1471, %r1470, 448; shl.b32 %r1472, %r8, 3; and.b32 %r1473, %r1472, 56; or.b32 %r1474, %r1471, %r1473; and.b32 %r1475, %r9, 7; or.b32 %r1476, %r1474, %r1475; mul.wide.u32 %rd985, %r1476, 12; add.s64 %rd986, %rd1621, %rd985; ld.f32 %f932, [%rd986+128]; ld.f32 %f931, [%rd986+132]; ld.f32 %f930, [%rd986+136]; mov.b64 %rd987, {%r260, %r1469}; shr.u64 %rd988, %rd987, 32; cvt.u32.u64 %r2289, %rd988; mov.u32 %r2334, %r260; bra.uni $L__BB4_276; $L__BB4_281: and.b32 %r473, %r471, -128; setp.ne.s32 %p175, %r473, %r2377; mov.u16 %rs191, 0; @%p175 bra $L__BB4_284; and.b32 %r1543, %r8, -128; setp.ne.s32 %p176, %r1543, %r2376; @%p176 bra $L__BB4_284; and.b32 %r1544, %r9, -128; setp.eq.s32 %p177, %r1544, %r2375; selp.u16 %rs191, 1, 0, %p177; $L__BB4_284: setp.eq.s16 %p178, %rs191, 0; @%p178 bra $L__BB4_288; shl.b32 %r1545, %r471, 5; and.b32 %r1546, %r1545, 3840; shl.b32 %r1547, %r8, 1; and.b32 %r1548, %r1547, 240; or.b32 %r1549, %r1548, %r1546; bfe.u32 %r1550, %r9, 3, 4; or.b32 %r1551, %r1549, %r1550; shr.u32 %r1552, %r1549, 6; mul.wide.u32 %rd1046, %r1552, 8; add.s64 %rd1047, %rd1628, %rd1046; and.b32 %r1553, %r1551, 63; mov.u64 %rd1048, 1; shl.b64 %rd1049, %rd1048, %r1553; ld.u64 %rd1050, [%rd1047+544]; and.b64 %rd1051, %rd1050, %rd1049; setp.eq.s64 %p179, %rd1051, 0; mul.wide.u32 %rd1052, %r1551, 16; add.s64 %rd1053, %rd1628, %rd1052; add.s64 %rd246, %rd1053, 1088; @%p179 bra $L__BB4_287; ld.u64 %rd1054, [%rd246]; add.s64 %rd1621, %rd1628, %rd1054; and.b32 %r1554, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1554}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1555, %r471, 6; and.b32 %r1556, %r1555, 448; shl.b32 %r1557, %r8, 3; and.b32 %r1558, %r1557, 56; or.b32 %r1559, %r1558, %r1556; and.b32 %r1560, %r9, 7; or.b32 %r1561, %r1559, %r1560; mul.wide.u32 %rd1055, %r1561, 12; add.s64 %rd1056, %rd1621, %rd1055; ld.f32 %f935, [%rd1056+128]; ld.f32 %f934, [%rd1056+132]; ld.f32 %f933, [%rd1056+136]; mov.b64 %rd1057, {%r472, %r1554}; shr.u64 %rd1058, %rd1057, 32; cvt.u32.u64 %r2289, %rd1058; mov.u32 %r2334, %r472; bra.uni $L__BB4_311; $L__BB4_316: and.b32 %r518, %r471, -128; setp.ne.s32 %p197, %r518, %r2377; mov.u16 %rs194, 0; @%p197 bra $L__BB4_319; and.b32 %r1628, %r8, -128; setp.ne.s32 %p198, %r1628, %r2376; @%p198 bra $L__BB4_319; and.b32 %r1629, %r74, -128; setp.eq.s32 %p199, %r1629, %r2375; selp.u16 %rs194, 1, 0, %p199; $L__BB4_319: setp.eq.s16 %p200, %rs194, 0; @%p200 bra $L__BB4_323; shl.b32 %r1630, %r471, 5; and.b32 %r1631, %r1630, 3840; shl.b32 %r1632, %r8, 1; and.b32 %r1633, %r1632, 240; or.b32 %r1634, %r1633, %r1631; bfe.u32 %r1635, %r74, 3, 4; or.b32 %r1636, %r1635, %r1634; shr.u32 %r1637, %r1634, 6; mul.wide.u32 %rd1116, %r1637, 8; add.s64 %rd1117, %rd1628, %rd1116; and.b32 %r1638, %r1636, 63; mov.u64 %rd1118, 1; shl.b64 %rd1119, %rd1118, %r1638; ld.u64 %rd1120, [%rd1117+544]; and.b64 %rd1121, %rd1120, %rd1119; setp.eq.s64 %p201, %rd1121, 0; mul.wide.u32 %rd1122, %r1636, 16; add.s64 %rd1123, %rd1628, %rd1122; add.s64 %rd275, %rd1123, 1088; @%p201 bra $L__BB4_322; ld.u64 %rd1124, [%rd275]; add.s64 %rd1621, %rd1628, %rd1124; and.b32 %r1639, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1639}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1640, %r471, 6; and.b32 %r1641, %r1640, 448; shl.b32 %r1642, %r8, 3; and.b32 %r1643, %r1642, 56; or.b32 %r1644, %r1643, %r1641; and.b32 %r1645, %r74, 7; or.b32 %r1646, %r1644, %r1645; mul.wide.u32 %rd1125, %r1646, 12; add.s64 %rd1126, %rd1621, %rd1125; ld.f32 %f938, [%rd1126+128]; ld.f32 %f937, [%rd1126+132]; ld.f32 %f936, [%rd1126+136]; mov.b64 %rd1127, {%r472, %r1639}; shr.u64 %rd1128, %rd1127, 32; cvt.u32.u64 %r2289, %rd1128; mov.u32 %r2334, %r472; bra.uni $L__BB4_346; $L__BB4_351: and.b32 %r560, %r471, -128; setp.ne.s32 %p219, %r560, %r2377; mov.u16 %rs197, 0; @%p219 bra $L__BB4_354; and.b32 %r1714, %r139, -128; setp.ne.s32 %p220, %r1714, %r2376; @%p220 bra $L__BB4_354; and.b32 %r1715, %r74, -128; setp.eq.s32 %p221, %r1715, %r2375; selp.u16 %rs197, 1, 0, %p221; $L__BB4_354: setp.eq.s16 %p222, %rs197, 0; @%p222 bra $L__BB4_358; shl.b32 %r1716, %r471, 5; and.b32 %r1717, %r1716, 3840; shl.b32 %r1718, %r139, 1; and.b32 %r1719, %r1718, 240; or.b32 %r1720, %r1719, %r1717; bfe.u32 %r1721, %r74, 3, 4; or.b32 %r1722, %r1720, %r1721; shr.u32 %r1723, %r1720, 6; mul.wide.u32 %rd1187, %r1723, 8; add.s64 %rd1188, %rd1628, %rd1187; and.b32 %r1724, %r1722, 63; mov.u64 %rd1189, 1; shl.b64 %rd1190, %rd1189, %r1724; ld.u64 %rd1191, [%rd1188+544]; and.b64 %rd1192, %rd1191, %rd1190; setp.eq.s64 %p223, %rd1192, 0; mul.wide.u32 %rd1193, %r1722, 16; add.s64 %rd1194, %rd1628, %rd1193; add.s64 %rd304, %rd1194, 1088; @%p223 bra $L__BB4_357; ld.u64 %rd1195, [%rd304]; add.s64 %rd1621, %rd1628, %rd1195; and.b32 %r1725, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1725}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1726, %r471, 6; and.b32 %r1727, %r1726, 448; shl.b32 %r1728, %r139, 3; and.b32 %r1729, %r1728, 56; or.b32 %r1730, %r1729, %r1727; and.b32 %r1731, %r74, 7; or.b32 %r1732, %r1730, %r1731; mul.wide.u32 %rd1196, %r1732, 12; add.s64 %rd1197, %rd1621, %rd1196; ld.f32 %f941, [%rd1197+128]; ld.f32 %f940, [%rd1197+132]; ld.f32 %f939, [%rd1197+136]; mov.b64 %rd1198, {%r472, %r1725}; shr.u64 %rd1199, %rd1198, 32; cvt.u32.u64 %r2289, %rd1199; mov.u32 %r2334, %r472; bra.uni $L__BB4_381; $L__BB4_386: and.b32 %r599, %r471, -128; setp.ne.s32 %p241, %r599, %r2377; mov.u16 %rs200, 0; @%p241 bra $L__BB4_389; and.b32 %r1800, %r139, -128; setp.ne.s32 %p242, %r1800, %r2376; @%p242 bra $L__BB4_389; and.b32 %r1801, %r9, -128; setp.eq.s32 %p243, %r1801, %r2375; selp.u16 %rs200, 1, 0, %p243; $L__BB4_389: setp.eq.s16 %p244, %rs200, 0; @%p244 bra $L__BB4_393; shl.b32 %r1802, %r471, 5; and.b32 %r1803, %r1802, 3840; shl.b32 %r1804, %r139, 1; and.b32 %r1805, %r1804, 240; or.b32 %r1806, %r1805, %r1803; bfe.u32 %r1807, %r9, 3, 4; or.b32 %r1808, %r1806, %r1807; shr.u32 %r1809, %r1806, 6; mul.wide.u32 %rd1258, %r1809, 8; add.s64 %rd1259, %rd1628, %rd1258; and.b32 %r1810, %r1808, 63; mov.u64 %rd1260, 1; shl.b64 %rd1261, %rd1260, %r1810; ld.u64 %rd1262, [%rd1259+544]; and.b64 %rd1263, %rd1262, %rd1261; setp.eq.s64 %p245, %rd1263, 0; mul.wide.u32 %rd1264, %r1808, 16; add.s64 %rd1265, %rd1628, %rd1264; add.s64 %rd333, %rd1265, 1088; @%p245 bra $L__BB4_392; ld.u64 %rd1266, [%rd333]; add.s64 %rd1621, %rd1628, %rd1266; and.b32 %r1811, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1811}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1812, %r471, 7; shl.b32 %r1813, %r139, 3; and.b32 %r1814, %r1813, 56; and.b32 %r1815, %r9, 7; bfi.b32 %r1816, %r1812, %r1815, 6, 3; or.b32 %r1817, %r1816, %r1814; mul.wide.u32 %rd1267, %r1817, 12; add.s64 %rd1268, %rd1621, %rd1267; ld.f32 %f944, [%rd1268+128]; ld.f32 %f943, [%rd1268+132]; ld.f32 %f942, [%rd1268+136]; mov.b64 %rd1269, {%r472, %r1811}; shr.u64 %rd1270, %rd1269, 32; cvt.u32.u64 %r2289, %rd1270; mov.u32 %r2334, %r472; bra.uni $L__BB4_416; $L__BB4_421: and.b32 %r637, %r635, -128; setp.ne.s32 %p263, %r637, %r2377; mov.u16 %rs203, 0; @%p263 bra $L__BB4_424; and.b32 %r1882, %r139, -128; setp.ne.s32 %p264, %r1882, %r2376; @%p264 bra $L__BB4_424; and.b32 %r1883, %r9, -128; setp.eq.s32 %p265, %r1883, %r2375; selp.u16 %rs203, 1, 0, %p265; $L__BB4_424: setp.eq.s16 %p266, %rs203, 0; @%p266 bra $L__BB4_428; shl.b32 %r1884, %r635, 5; and.b32 %r1885, %r1884, 3840; shl.b32 %r1886, %r139, 1; and.b32 %r1887, %r1886, 240; or.b32 %r1888, %r1887, %r1885; bfe.u32 %r1889, %r9, 3, 4; or.b32 %r1890, %r1888, %r1889; shr.u32 %r1891, %r1888, 6; mul.wide.u32 %rd1329, %r1891, 8; add.s64 %rd1330, %rd1628, %rd1329; and.b32 %r1892, %r1890, 63; mov.u64 %rd1331, 1; shl.b64 %rd1332, %rd1331, %r1892; ld.u64 %rd1333, [%rd1330+544]; and.b64 %rd1334, %rd1333, %rd1332; setp.eq.s64 %p267, %rd1334, 0; mul.wide.u32 %rd1335, %r1890, 16; add.s64 %rd1336, %rd1628, %rd1335; add.s64 %rd362, %rd1336, 1088; @%p267 bra $L__BB4_427; ld.u64 %rd1337, [%rd362]; add.s64 %rd1621, %rd1628, %rd1337; and.b32 %r1893, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1893}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1894, %r635, 7; shl.b32 %r1895, %r139, 3; and.b32 %r1896, %r1895, 56; and.b32 %r1897, %r9, 7; bfi.b32 %r1898, %r1894, %r1897, 6, 3; or.b32 %r1899, %r1898, %r1896; mul.wide.u32 %rd1338, %r1899, 12; add.s64 %rd1339, %rd1621, %rd1338; ld.f32 %f947, [%rd1339+128]; ld.f32 %f946, [%rd1339+132]; ld.f32 %f945, [%rd1339+136]; mov.b64 %rd1340, {%r636, %r1893}; shr.u64 %rd1341, %rd1340, 32; cvt.u32.u64 %r2289, %rd1341; mov.u32 %r2334, %r636; bra.uni $L__BB4_451; $L__BB4_456: and.b32 %r670, %r635, -128; setp.ne.s32 %p285, %r670, %r2377; mov.u16 %rs206, 0; @%p285 bra $L__BB4_459; and.b32 %r1965, %r139, -128; setp.ne.s32 %p286, %r1965, %r2376; @%p286 bra $L__BB4_459; and.b32 %r1966, %r74, -128; setp.eq.s32 %p287, %r1966, %r2375; selp.u16 %rs206, 1, 0, %p287; $L__BB4_459: setp.eq.s16 %p288, %rs206, 0; @%p288 bra $L__BB4_463; shl.b32 %r1967, %r635, 5; and.b32 %r1968, %r1967, 3840; shl.b32 %r1969, %r139, 1; and.b32 %r1970, %r1969, 240; or.b32 %r1971, %r1970, %r1968; bfe.u32 %r1972, %r74, 3, 4; or.b32 %r1973, %r1971, %r1972; shr.u32 %r1974, %r1971, 6; mul.wide.u32 %rd1400, %r1974, 8; add.s64 %rd1401, %rd1628, %rd1400; and.b32 %r1975, %r1973, 63; mov.u64 %rd1402, 1; shl.b64 %rd1403, %rd1402, %r1975; ld.u64 %rd1404, [%rd1401+544]; and.b64 %rd1405, %rd1404, %rd1403; setp.eq.s64 %p289, %rd1405, 0; mul.wide.u32 %rd1406, %r1973, 16; add.s64 %rd1407, %rd1628, %rd1406; add.s64 %rd391, %rd1407, 1088; @%p289 bra $L__BB4_462; ld.u64 %rd1408, [%rd391]; add.s64 %rd1621, %rd1628, %rd1408; and.b32 %r1976, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1976}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1977, %r635, 6; and.b32 %r1978, %r1977, 448; shl.b32 %r1979, %r139, 3; and.b32 %r1980, %r1979, 56; or.b32 %r1981, %r1980, %r1978; and.b32 %r1982, %r74, 7; or.b32 %r1983, %r1981, %r1982; mul.wide.u32 %rd1409, %r1983, 12; add.s64 %rd1410, %rd1621, %rd1409; ld.f32 %f950, [%rd1410+128]; ld.f32 %f949, [%rd1410+132]; ld.f32 %f948, [%rd1410+136]; mov.b64 %rd1411, {%r636, %r1976}; shr.u64 %rd1412, %rd1411, 32; cvt.u32.u64 %r2289, %rd1412; mov.u32 %r2334, %r636; bra.uni $L__BB4_486; $L__BB4_491: and.b32 %r700, %r635, -128; setp.ne.s32 %p307, %r700, %r2377; mov.u16 %rs209, 0; @%p307 bra $L__BB4_494; and.b32 %r2052, %r8, -128; setp.ne.s32 %p308, %r2052, %r2376; @%p308 bra $L__BB4_494; and.b32 %r2053, %r74, -128; setp.eq.s32 %p309, %r2053, %r2375; selp.u16 %rs209, 1, 0, %p309; $L__BB4_494: setp.eq.s16 %p310, %rs209, 0; @%p310 bra $L__BB4_498; shl.b32 %r2054, %r635, 5; and.b32 %r2055, %r2054, 3840; shl.b32 %r2056, %r8, 1; and.b32 %r2057, %r2056, 240; or.b32 %r2058, %r2055, %r2057; bfe.u32 %r2059, %r74, 3, 4; or.b32 %r2060, %r2059, %r2058; shr.u32 %r2061, %r2058, 6; mul.wide.u32 %rd1471, %r2061, 8; add.s64 %rd1472, %rd1628, %rd1471; and.b32 %r2062, %r2060, 63; mov.u64 %rd1473, 1; shl.b64 %rd1474, %rd1473, %r2062; ld.u64 %rd1475, [%rd1472+544]; and.b64 %rd1476, %rd1475, %rd1474; setp.eq.s64 %p311, %rd1476, 0; mul.wide.u32 %rd1477, %r2060, 16; add.s64 %rd1478, %rd1628, %rd1477; add.s64 %rd420, %rd1478, 1088; @%p311 bra $L__BB4_497; ld.u64 %rd1479, [%rd420]; add.s64 %rd1621, %rd1628, %rd1479; and.b32 %r2063, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2063}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r2064, %r635, 6; and.b32 %r2065, %r2064, 448; shl.b32 %r2066, %r8, 3; and.b32 %r2067, %r2066, 56; or.b32 %r2068, %r2065, %r2067; and.b32 %r2069, %r74, 7; or.b32 %r2070, %r2068, %r2069; mul.wide.u32 %rd1480, %r2070, 12; add.s64 %rd1481, %rd1621, %rd1480; ld.f32 %f953, [%rd1481+128]; ld.f32 %f952, [%rd1481+132]; ld.f32 %f951, [%rd1481+136]; mov.b64 %rd1482, {%r636, %r2063}; shr.u64 %rd1483, %rd1482, 32; cvt.u32.u64 %r2289, %rd1483; mov.u32 %r2334, %r636; bra.uni $L__BB4_521; $L__BB4_534: shl.b64 %rd1547, %rd451, 5; add.s64 %rd1704, %rd17, %rd1547; $L__BB4_535: setp.eq.s64 %p337, %rd1704, 0; @%p337 bra $L__BB4_543; add.s64 %rd454, %rd1704, 8; ld.u64 %rd455, [%rd1704+8]; setp.eq.s64 %p338, %rd455, 0; @%p338 bra $L__BB4_542; add.s64 %rd456, %rd1, %rd455; and.b32 %r2137, %r8, -4096; mov.u64 %rd1548, 1; st.local.v2.u32 [%rd4+16], {%r728, %r2137}; and.b32 %r2138, %r9, -4096; st.local.u32 [%rd4+24], %r2138; st.local.u64 [%rd4+56], %rd456; shl.b32 %r2139, %r635, 3; and.b32 %r2140, %r2139, 31744; and.b32 %r2141, %r8, 3968; shr.u32 %r2142, %r2141, 2; or.b32 %r2143, %r2140, %r2142; bfe.u32 %r2144, %r9, 7, 5; or.b32 %r2145, %r2143, %r2144; shr.u32 %r2146, %r2143, 6; mul.wide.u32 %rd1549, %r2146, 8; add.s64 %rd1550, %rd456, %rd1549; and.b32 %r2147, %r2145, 63; shl.b64 %rd1551, %rd1548, %r2147; ld.u64 %rd1552, [%rd1550+4128]; and.b64 %rd1553, %rd1552, %rd1551; setp.eq.s64 %p339, %rd1553, 0; mul.wide.u32 %rd1554, %r2145, 16; add.s64 %rd1555, %rd456, %rd1554; add.s64 %rd457, %rd1555, 8256; @%p339 bra $L__BB4_541; ld.u64 %rd1556, [%rd457]; add.s64 %rd458, %rd456, %rd1556; st.local.u32 [%rd4+4], %r727; and.b32 %r2148, %r8, -128; st.local.u32 [%rd4+8], %r2148; and.b32 %r2149, %r9, -128; st.local.u32 [%rd4+12], %r2149; st.local.u64 [%rd4+48], %rd458; shl.b32 %r2150, %r635, 5; and.b32 %r2151, %r2150, 3840; shl.b32 %r2152, %r8, 1; and.b32 %r2153, %r2152, 240; or.b32 %r2154, %r2151, %r2153; bfe.u32 %r2155, %r9, 3, 4; or.b32 %r2156, %r2154, %r2155; shr.u32 %r2157, %r2154, 6; mul.wide.u32 %rd1557, %r2157, 8; add.s64 %rd1558, %rd458, %rd1557; and.b32 %r2158, %r2156, 63; shl.b64 %rd1560, %rd1548, %r2158; ld.u64 %rd1561, [%rd1558+544]; and.b64 %rd1562, %rd1561, %rd1560; setp.eq.s64 %p340, %rd1562, 0; mul.wide.u32 %rd1563, %r2156, 16; add.s64 %rd1564, %rd458, %rd1563; add.s64 %rd459, %rd1564, 1088; @%p340 bra $L__BB4_540; ld.u64 %rd1565, [%rd459]; add.s64 %rd1566, %rd458, %rd1565; and.b32 %r2159, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2159}; and.b32 %r2160, %r9, -8; st.local.u32 [%rd4], %r2160; st.local.u64 [%rd4+40], %rd1566; shl.b32 %r2161, %r635, 6; and.b32 %r2162, %r2161, 448; shl.b32 %r2163, %r8, 3; and.b32 %r2164, %r2163, 56; or.b32 %r2165, %r2162, %r2164; and.b32 %r2166, %r9, 7; or.b32 %r2167, %r2165, %r2166; mul.wide.u32 %rd1567, %r2167, 12; add.s64 %rd1568, %rd1566, %rd1567; ld.f32 %f956, [%rd1568+128]; ld.f32 %f955, [%rd1568+132]; ld.f32 %f954, [%rd1568+136]; bra.uni $L__BB4_553; $L__BB4_543: ld.f32 %f956, [%rd5+4]; ld.f32 %f955, [%rd5+8]; ld.f32 %f954, [%rd5+12]; bra.uni $L__BB4_553; $L__BB4_542: ld.f32 %f956, [%rd454+12]; ld.f32 %f955, [%rd454+16]; ld.f32 %f954, [%rd454+20]; bra.uni $L__BB4_553; $L__BB4_43: setp.ne.s32 %p26, %r12, %r2337; mov.u16 %rs171, 0; @%p26 bra $L__BB4_46; and.b32 %r970, %r8, -4096; setp.ne.s32 %p27, %r970, %r2382; @%p27 bra $L__BB4_46; and.b32 %r971, %r74, -4096; setp.eq.s32 %p28, %r971, %r2415; selp.u16 %rs171, 1, 0, %p28; $L__BB4_46: setp.eq.s16 %p29, %rs171, 0; @%p29 bra $L__BB4_52; shl.b32 %r972, %r7, 3; and.b32 %r973, %r972, 31744; and.b32 %r974, %r8, 3968; shr.u32 %r975, %r974, 2; or.b32 %r976, %r975, %r973; bfe.u32 %r977, %r74, 7, 5; or.b32 %r978, %r977, %r976; shr.u32 %r979, %r976, 6; mul.wide.u32 %rd569, %r979, 8; add.s64 %rd570, %rd1636, %rd569; and.b32 %r980, %r978, 63; mov.u64 %rd571, 1; shl.b64 %rd572, %rd571, %r980; ld.u64 %rd573, [%rd570+4128]; and.b64 %rd574, %rd573, %rd572; setp.eq.s64 %p30, %rd574, 0; mul.wide.u32 %rd575, %r978, 16; add.s64 %rd576, %rd1636, %rd575; add.s64 %rd45, %rd576, 8256; @%p30 bra $L__BB4_51; and.b32 %r2267, %r7, -128; ld.u64 %rd577, [%rd45]; add.s64 %rd1628, %rd1636, %rd577; st.local.u32 [%rd4+4], %r2267; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r981, %r7, 5; and.b32 %r982, %r981, 3840; shl.b32 %r983, %r8, 1; and.b32 %r984, %r983, 240; or.b32 %r985, %r984, %r982; bfe.u32 %r986, %r74, 3, 4; or.b32 %r987, %r986, %r985; shr.u32 %r988, %r985, 6; mul.wide.u32 %rd578, %r988, 8; add.s64 %rd579, %rd1628, %rd578; and.b32 %r989, %r987, 63; shl.b64 %rd581, %rd571, %r989; ld.u64 %rd582, [%rd579+544]; and.b64 %rd583, %rd582, %rd581; setp.eq.s64 %p31, %rd583, 0; mul.wide.u32 %rd584, %r987, 16; add.s64 %rd585, %rd1628, %rd584; add.s64 %rd47, %rd585, 1088; and.b32 %r2377, %r7, -128; @%p31 bra $L__BB4_50; and.b32 %r2334, %r7, -8; ld.u64 %rd586, [%rd47]; add.s64 %rd1621, %rd1628, %rd586; and.b32 %r990, %r8, -8; st.local.v2.u32 [%rd2], {%r2334, %r990}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r991, %r7, 6; and.b32 %r992, %r991, 448; shl.b32 %r993, %r8, 3; and.b32 %r994, %r993, 56; or.b32 %r995, %r994, %r992; and.b32 %r996, %r74, 7; or.b32 %r997, %r995, %r996; mul.wide.u32 %rd587, %r997, 12; add.s64 %rd588, %rd1621, %rd587; ld.f32 %f914, [%rd588+128]; ld.f32 %f913, [%rd588+132]; ld.f32 %f912, [%rd588+136]; mov.b64 %rd589, {%r2334, %r990}; shr.u64 %rd590, %rd589, 32; cvt.u32.u64 %r2289, %rd590; bra.uni $L__BB4_66; $L__BB4_78: setp.ne.s32 %p48, %r12, %r2337; mov.u16 %rs174, 0; @%p48 bra $L__BB4_81; and.b32 %r1055, %r139, -4096; setp.ne.s32 %p49, %r1055, %r2382; @%p49 bra $L__BB4_81; and.b32 %r1056, %r74, -4096; setp.eq.s32 %p50, %r1056, %r2415; selp.u16 %rs174, 1, 0, %p50; $L__BB4_81: setp.eq.s16 %p51, %rs174, 0; @%p51 bra $L__BB4_87; shl.b32 %r1057, %r7, 3; and.b32 %r1058, %r1057, 31744; and.b32 %r1059, %r139, 3968; shr.u32 %r1060, %r1059, 2; or.b32 %r1061, %r1060, %r1058; bfe.u32 %r1062, %r74, 7, 5; or.b32 %r1063, %r1061, %r1062; shr.u32 %r1064, %r1061, 6; mul.wide.u32 %rd638, %r1064, 8; add.s64 %rd639, %rd1636, %rd638; and.b32 %r1065, %r1063, 63; mov.u64 %rd640, 1; shl.b64 %rd641, %rd640, %r1065; ld.u64 %rd642, [%rd639+4128]; and.b64 %rd643, %rd642, %rd641; setp.eq.s64 %p52, %rd643, 0; mul.wide.u32 %rd644, %r1063, 16; add.s64 %rd645, %rd1636, %rd644; add.s64 %rd74, %rd645, 8256; @%p52 bra $L__BB4_86; and.b32 %r2260, %r7, -128; ld.u64 %rd646, [%rd74]; add.s64 %rd1628, %rd1636, %rd646; st.local.u32 [%rd4+4], %r2260; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1066, %r7, 5; and.b32 %r1067, %r1066, 3840; shl.b32 %r1068, %r139, 1; and.b32 %r1069, %r1068, 240; or.b32 %r1070, %r1069, %r1067; bfe.u32 %r1071, %r74, 3, 4; or.b32 %r1072, %r1070, %r1071; shr.u32 %r1073, %r1070, 6; mul.wide.u32 %rd647, %r1073, 8; add.s64 %rd648, %rd1628, %rd647; and.b32 %r1074, %r1072, 63; shl.b64 %rd650, %rd640, %r1074; ld.u64 %rd651, [%rd648+544]; and.b64 %rd652, %rd651, %rd650; setp.eq.s64 %p53, %rd652, 0; mul.wide.u32 %rd653, %r1072, 16; add.s64 %rd654, %rd1628, %rd653; add.s64 %rd76, %rd654, 1088; and.b32 %r2377, %r7, -128; @%p53 bra $L__BB4_85; and.b32 %r2334, %r7, -8; ld.u64 %rd655, [%rd76]; add.s64 %rd1621, %rd1628, %rd655; and.b32 %r1075, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1075}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1076, %r7, 6; and.b32 %r1077, %r1076, 448; shl.b32 %r1078, %r139, 3; and.b32 %r1079, %r1078, 56; or.b32 %r1080, %r1079, %r1077; and.b32 %r1081, %r74, 7; or.b32 %r1082, %r1080, %r1081; mul.wide.u32 %rd656, %r1082, 12; add.s64 %rd657, %rd1621, %rd656; ld.f32 %f917, [%rd657+128]; ld.f32 %f916, [%rd657+132]; ld.f32 %f915, [%rd657+136]; mov.b64 %rd658, {%r2334, %r1075}; shr.u64 %rd659, %rd658, 32; cvt.u32.u64 %r2289, %rd659; bra.uni $L__BB4_101; $L__BB4_113: setp.ne.s32 %p70, %r12, %r2337; mov.u16 %rs177, 0; @%p70 bra $L__BB4_116; and.b32 %r1139, %r139, -4096; setp.ne.s32 %p71, %r1139, %r2382; @%p71 bra $L__BB4_116; and.b32 %r1140, %r9, -4096; setp.eq.s32 %p72, %r1140, %r2415; selp.u16 %rs177, 1, 0, %p72; $L__BB4_116: setp.eq.s16 %p73, %rs177, 0; @%p73 bra $L__BB4_122; shl.b32 %r1141, %r7, 3; and.b32 %r1142, %r1141, 31744; and.b32 %r1143, %r139, 3968; shr.u32 %r1144, %r1143, 2; or.b32 %r1145, %r1144, %r1142; bfe.u32 %r1146, %r9, 7, 5; or.b32 %r1147, %r1145, %r1146; shr.u32 %r1148, %r1145, 6; mul.wide.u32 %rd708, %r1148, 8; add.s64 %rd709, %rd1636, %rd708; and.b32 %r1149, %r1147, 63; mov.u64 %rd710, 1; shl.b64 %rd711, %rd710, %r1149; ld.u64 %rd712, [%rd709+4128]; and.b64 %rd713, %rd712, %rd711; setp.eq.s64 %p74, %rd713, 0; mul.wide.u32 %rd714, %r1147, 16; add.s64 %rd715, %rd1636, %rd714; add.s64 %rd103, %rd715, 8256; @%p74 bra $L__BB4_121; and.b32 %r2233, %r7, -128; ld.u64 %rd716, [%rd103]; add.s64 %rd1628, %rd1636, %rd716; st.local.u32 [%rd4+4], %r2233; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1150, %r7, 5; and.b32 %r1151, %r1150, 3840; shl.b32 %r1152, %r139, 1; and.b32 %r1153, %r1152, 240; or.b32 %r1154, %r1153, %r1151; bfe.u32 %r1155, %r9, 3, 4; or.b32 %r1156, %r1154, %r1155; shr.u32 %r1157, %r1154, 6; mul.wide.u32 %rd717, %r1157, 8; add.s64 %rd718, %rd1628, %rd717; and.b32 %r1158, %r1156, 63; shl.b64 %rd720, %rd710, %r1158; ld.u64 %rd721, [%rd718+544]; and.b64 %rd722, %rd721, %rd720; setp.eq.s64 %p75, %rd722, 0; mul.wide.u32 %rd723, %r1156, 16; add.s64 %rd724, %rd1628, %rd723; add.s64 %rd105, %rd724, 1088; and.b32 %r2377, %r7, -128; @%p75 bra $L__BB4_120; and.b32 %r2334, %r7, -8; ld.u64 %rd725, [%rd105]; add.s64 %rd1621, %rd1628, %rd725; and.b32 %r1159, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1159}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1160, %r7, 7; shl.b32 %r1161, %r139, 3; and.b32 %r1162, %r1161, 56; and.b32 %r1163, %r9, 7; bfi.b32 %r1164, %r1160, %r1163, 6, 3; or.b32 %r1165, %r1164, %r1162; mul.wide.u32 %rd726, %r1165, 12; add.s64 %rd727, %rd1621, %rd726; ld.f32 %f920, [%rd727+128]; ld.f32 %f919, [%rd727+132]; ld.f32 %f918, [%rd727+136]; mov.b64 %rd728, {%r2334, %r1159}; shr.u64 %rd729, %rd728, 32; cvt.u32.u64 %r2289, %rd729; bra.uni $L__BB4_136; $L__BB4_148: and.b32 %r264, %r259, -4096; setp.ne.s32 %p92, %r264, %r2337; mov.u16 %rs180, 0; @%p92 bra $L__BB4_151; and.b32 %r1220, %r139, -4096; setp.ne.s32 %p93, %r1220, %r2382; @%p93 bra $L__BB4_151; and.b32 %r1221, %r9, -4096; setp.eq.s32 %p94, %r1221, %r2415; selp.u16 %rs180, 1, 0, %p94; $L__BB4_151: setp.eq.s16 %p95, %rs180, 0; @%p95 bra $L__BB4_157; shl.b32 %r1222, %r259, 3; and.b32 %r1223, %r1222, 31744; and.b32 %r1224, %r139, 3968; shr.u32 %r1225, %r1224, 2; or.b32 %r1226, %r1225, %r1223; bfe.u32 %r1227, %r9, 7, 5; or.b32 %r1228, %r1226, %r1227; shr.u32 %r1229, %r1226, 6; mul.wide.u32 %rd776, %r1229, 8; add.s64 %rd777, %rd1636, %rd776; and.b32 %r1230, %r1228, 63; mov.u64 %rd778, 1; shl.b64 %rd779, %rd778, %r1230; ld.u64 %rd780, [%rd777+4128]; and.b64 %rd781, %rd780, %rd779; setp.eq.s64 %p96, %rd781, 0; mul.wide.u32 %rd782, %r1228, 16; add.s64 %rd783, %rd1636, %rd782; add.s64 %rd132, %rd783, 8256; @%p96 bra $L__BB4_156; ld.u64 %rd784, [%rd132]; add.s64 %rd1628, %rd1636, %rd784; st.local.u32 [%rd4+4], %r261; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1231, %r259, 5; and.b32 %r1232, %r1231, 3840; shl.b32 %r1233, %r139, 1; and.b32 %r1234, %r1233, 240; or.b32 %r1235, %r1234, %r1232; bfe.u32 %r1236, %r9, 3, 4; or.b32 %r1237, %r1235, %r1236; shr.u32 %r1238, %r1235, 6; mul.wide.u32 %rd785, %r1238, 8; add.s64 %rd786, %rd1628, %rd785; and.b32 %r1239, %r1237, 63; shl.b64 %rd788, %rd778, %r1239; ld.u64 %rd789, [%rd786+544]; and.b64 %rd790, %rd789, %rd788; setp.eq.s64 %p97, %rd790, 0; mul.wide.u32 %rd791, %r1237, 16; add.s64 %rd792, %rd1628, %rd791; add.s64 %rd134, %rd792, 1088; @%p97 bra $L__BB4_155; ld.u64 %rd793, [%rd134]; add.s64 %rd1621, %rd1628, %rd793; and.b32 %r1240, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1240}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1241, %r259, 7; shl.b32 %r1242, %r139, 3; and.b32 %r1243, %r1242, 56; and.b32 %r1244, %r9, 7; bfi.b32 %r1245, %r1241, %r1244, 6, 3; or.b32 %r1246, %r1245, %r1243; mul.wide.u32 %rd794, %r1246, 12; add.s64 %rd795, %rd1621, %rd794; ld.f32 %f923, [%rd795+128]; ld.f32 %f922, [%rd795+132]; ld.f32 %f921, [%rd795+136]; mov.b64 %rd796, {%r260, %r1240}; shr.u64 %rd797, %rd796, 32; cvt.u32.u64 %r2289, %rd797; mov.u32 %r2377, %r261; mov.u32 %r2334, %r260; bra.uni $L__BB4_171; $L__BB4_183: and.b32 %r321, %r259, -4096; setp.ne.s32 %p114, %r321, %r2337; mov.u16 %rs183, 0; @%p114 bra $L__BB4_186; and.b32 %r1304, %r139, -4096; setp.ne.s32 %p115, %r1304, %r2382; @%p115 bra $L__BB4_186; and.b32 %r1305, %r74, -4096; setp.eq.s32 %p116, %r1305, %r2415; selp.u16 %rs183, 1, 0, %p116; $L__BB4_186: setp.eq.s16 %p117, %rs183, 0; @%p117 bra $L__BB4_192; shl.b32 %r1306, %r259, 3; and.b32 %r1307, %r1306, 31744; and.b32 %r1308, %r139, 3968; shr.u32 %r1309, %r1308, 2; or.b32 %r1310, %r1309, %r1307; bfe.u32 %r1311, %r74, 7, 5; or.b32 %r1312, %r1310, %r1311; shr.u32 %r1313, %r1310, 6; mul.wide.u32 %rd847, %r1313, 8; add.s64 %rd848, %rd1636, %rd847; and.b32 %r1314, %r1312, 63; mov.u64 %rd849, 1; shl.b64 %rd850, %rd849, %r1314; ld.u64 %rd851, [%rd848+4128]; and.b64 %rd852, %rd851, %rd850; setp.eq.s64 %p118, %rd852, 0; mul.wide.u32 %rd853, %r1312, 16; add.s64 %rd854, %rd1636, %rd853; add.s64 %rd161, %rd854, 8256; @%p118 bra $L__BB4_191; ld.u64 %rd855, [%rd161]; add.s64 %rd1628, %rd1636, %rd855; st.local.u32 [%rd4+4], %r318; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1315, %r259, 5; and.b32 %r1316, %r1315, 3840; shl.b32 %r1317, %r139, 1; and.b32 %r1318, %r1317, 240; or.b32 %r1319, %r1318, %r1316; bfe.u32 %r1320, %r74, 3, 4; or.b32 %r1321, %r1319, %r1320; shr.u32 %r1322, %r1319, 6; mul.wide.u32 %rd856, %r1322, 8; add.s64 %rd857, %rd1628, %rd856; and.b32 %r1323, %r1321, 63; shl.b64 %rd859, %rd849, %r1323; ld.u64 %rd860, [%rd857+544]; and.b64 %rd861, %rd860, %rd859; setp.eq.s64 %p119, %rd861, 0; mul.wide.u32 %rd862, %r1321, 16; add.s64 %rd863, %rd1628, %rd862; add.s64 %rd163, %rd863, 1088; @%p119 bra $L__BB4_190; ld.u64 %rd864, [%rd163]; add.s64 %rd1621, %rd1628, %rd864; and.b32 %r1324, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1324}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1325, %r259, 6; and.b32 %r1326, %r1325, 448; shl.b32 %r1327, %r139, 3; and.b32 %r1328, %r1327, 56; or.b32 %r1329, %r1328, %r1326; and.b32 %r1330, %r74, 7; or.b32 %r1331, %r1329, %r1330; mul.wide.u32 %rd865, %r1331, 12; add.s64 %rd866, %rd1621, %rd865; ld.f32 %f926, [%rd866+128]; ld.f32 %f925, [%rd866+132]; ld.f32 %f924, [%rd866+136]; mov.b64 %rd867, {%r260, %r1324}; shr.u64 %rd868, %rd867, 32; cvt.u32.u64 %r2289, %rd868; mov.u32 %r2377, %r318; mov.u32 %r2334, %r260; bra.uni $L__BB4_206; $L__BB4_218: and.b32 %r375, %r259, -4096; setp.ne.s32 %p136, %r375, %r2337; mov.u16 %rs186, 0; @%p136 bra $L__BB4_221; and.b32 %r1391, %r8, -4096; setp.ne.s32 %p137, %r1391, %r2382; @%p137 bra $L__BB4_221; and.b32 %r1392, %r74, -4096; setp.eq.s32 %p138, %r1392, %r2415; selp.u16 %rs186, 1, 0, %p138; $L__BB4_221: setp.eq.s16 %p139, %rs186, 0; @%p139 bra $L__BB4_227; shl.b32 %r1393, %r259, 3; and.b32 %r1394, %r1393, 31744; and.b32 %r1395, %r8, 3968; shr.u32 %r1396, %r1395, 2; or.b32 %r1397, %r1394, %r1396; bfe.u32 %r1398, %r74, 7, 5; or.b32 %r1399, %r1398, %r1397; shr.u32 %r1400, %r1397, 6; mul.wide.u32 %rd918, %r1400, 8; add.s64 %rd919, %rd1636, %rd918; and.b32 %r1401, %r1399, 63; mov.u64 %rd920, 1; shl.b64 %rd921, %rd920, %r1401; ld.u64 %rd922, [%rd919+4128]; and.b64 %rd923, %rd922, %rd921; setp.eq.s64 %p140, %rd923, 0; mul.wide.u32 %rd924, %r1399, 16; add.s64 %rd925, %rd1636, %rd924; add.s64 %rd190, %rd925, 8256; @%p140 bra $L__BB4_226; ld.u64 %rd926, [%rd190]; add.s64 %rd1628, %rd1636, %rd926; st.local.u32 [%rd4+4], %r372; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1402, %r259, 5; and.b32 %r1403, %r1402, 3840; shl.b32 %r1404, %r8, 1; and.b32 %r1405, %r1404, 240; or.b32 %r1406, %r1403, %r1405; bfe.u32 %r1407, %r74, 3, 4; or.b32 %r1408, %r1407, %r1406; shr.u32 %r1409, %r1406, 6; mul.wide.u32 %rd927, %r1409, 8; add.s64 %rd928, %rd1628, %rd927; and.b32 %r1410, %r1408, 63; shl.b64 %rd930, %rd920, %r1410; ld.u64 %rd931, [%rd928+544]; and.b64 %rd932, %rd931, %rd930; setp.eq.s64 %p141, %rd932, 0; mul.wide.u32 %rd933, %r1408, 16; add.s64 %rd934, %rd1628, %rd933; add.s64 %rd192, %rd934, 1088; @%p141 bra $L__BB4_225; ld.u64 %rd935, [%rd192]; add.s64 %rd1621, %rd1628, %rd935; and.b32 %r1411, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1411}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1412, %r259, 6; and.b32 %r1413, %r1412, 448; shl.b32 %r1414, %r8, 3; and.b32 %r1415, %r1414, 56; or.b32 %r1416, %r1413, %r1415; and.b32 %r1417, %r74, 7; or.b32 %r1418, %r1416, %r1417; mul.wide.u32 %rd936, %r1418, 12; add.s64 %rd937, %rd1621, %rd936; ld.f32 %f929, [%rd937+128]; ld.f32 %f928, [%rd937+132]; ld.f32 %f927, [%rd937+136]; mov.b64 %rd938, {%r260, %r1411}; shr.u64 %rd939, %rd938, 32; cvt.u32.u64 %r2289, %rd939; mov.u32 %r2377, %r372; mov.u32 %r2334, %r260; bra.uni $L__BB4_241; $L__BB4_253: and.b32 %r426, %r259, -4096; setp.ne.s32 %p158, %r426, %r2337; mov.u16 %rs189, 0; @%p158 bra $L__BB4_256; and.b32 %r1477, %r8, -4096; setp.ne.s32 %p159, %r1477, %r2382; @%p159 bra $L__BB4_256; and.b32 %r1478, %r9, -4096; setp.eq.s32 %p160, %r1478, %r2415; selp.u16 %rs189, 1, 0, %p160; $L__BB4_256: setp.eq.s16 %p161, %rs189, 0; @%p161 bra $L__BB4_262; shl.b32 %r1479, %r259, 3; and.b32 %r1480, %r1479, 31744; and.b32 %r1481, %r8, 3968; shr.u32 %r1482, %r1481, 2; or.b32 %r1483, %r1480, %r1482; bfe.u32 %r1484, %r9, 7, 5; or.b32 %r1485, %r1483, %r1484; shr.u32 %r1486, %r1483, 6; mul.wide.u32 %rd989, %r1486, 8; add.s64 %rd990, %rd1636, %rd989; and.b32 %r1487, %r1485, 63; mov.u64 %rd991, 1; shl.b64 %rd992, %rd991, %r1487; ld.u64 %rd993, [%rd990+4128]; and.b64 %rd994, %rd993, %rd992; setp.eq.s64 %p162, %rd994, 0; mul.wide.u32 %rd995, %r1485, 16; add.s64 %rd996, %rd1636, %rd995; add.s64 %rd219, %rd996, 8256; @%p162 bra $L__BB4_261; ld.u64 %rd997, [%rd219]; add.s64 %rd1628, %rd1636, %rd997; st.local.u32 [%rd4+4], %r423; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1488, %r259, 5; and.b32 %r1489, %r1488, 3840; shl.b32 %r1490, %r8, 1; and.b32 %r1491, %r1490, 240; or.b32 %r1492, %r1489, %r1491; bfe.u32 %r1493, %r9, 3, 4; or.b32 %r1494, %r1492, %r1493; shr.u32 %r1495, %r1492, 6; mul.wide.u32 %rd998, %r1495, 8; add.s64 %rd999, %rd1628, %rd998; and.b32 %r1496, %r1494, 63; shl.b64 %rd1001, %rd991, %r1496; ld.u64 %rd1002, [%rd999+544]; and.b64 %rd1003, %rd1002, %rd1001; setp.eq.s64 %p163, %rd1003, 0; mul.wide.u32 %rd1004, %r1494, 16; add.s64 %rd1005, %rd1628, %rd1004; add.s64 %rd221, %rd1005, 1088; @%p163 bra $L__BB4_260; ld.u64 %rd1006, [%rd221]; add.s64 %rd1621, %rd1628, %rd1006; and.b32 %r1497, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1497}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1498, %r259, 6; and.b32 %r1499, %r1498, 448; shl.b32 %r1500, %r8, 3; and.b32 %r1501, %r1500, 56; or.b32 %r1502, %r1499, %r1501; and.b32 %r1503, %r9, 7; or.b32 %r1504, %r1502, %r1503; mul.wide.u32 %rd1007, %r1504, 12; add.s64 %rd1008, %rd1621, %rd1007; ld.f32 %f932, [%rd1008+128]; ld.f32 %f931, [%rd1008+132]; ld.f32 %f930, [%rd1008+136]; mov.b64 %rd1009, {%r260, %r1497}; shr.u64 %rd1010, %rd1009, 32; cvt.u32.u64 %r2289, %rd1010; mov.u32 %r2377, %r423; mov.u32 %r2334, %r260; bra.uni $L__BB4_276; $L__BB4_288: and.b32 %r476, %r471, -4096; setp.ne.s32 %p180, %r476, %r2337; mov.u16 %rs192, 0; @%p180 bra $L__BB4_291; and.b32 %r1562, %r8, -4096; setp.ne.s32 %p181, %r1562, %r2382; @%p181 bra $L__BB4_291; and.b32 %r1563, %r9, -4096; setp.eq.s32 %p182, %r1563, %r2415; selp.u16 %rs192, 1, 0, %p182; $L__BB4_291: setp.eq.s16 %p183, %rs192, 0; @%p183 bra $L__BB4_297; shl.b32 %r1564, %r471, 3; and.b32 %r1565, %r1564, 31744; and.b32 %r1566, %r8, 3968; shr.u32 %r1567, %r1566, 2; or.b32 %r1568, %r1567, %r1565; bfe.u32 %r1569, %r9, 7, 5; or.b32 %r1570, %r1568, %r1569; shr.u32 %r1571, %r1568, 6; mul.wide.u32 %rd1059, %r1571, 8; add.s64 %rd1060, %rd1636, %rd1059; and.b32 %r1572, %r1570, 63; mov.u64 %rd1061, 1; shl.b64 %rd1062, %rd1061, %r1572; ld.u64 %rd1063, [%rd1060+4128]; and.b64 %rd1064, %rd1063, %rd1062; setp.eq.s64 %p184, %rd1064, 0; mul.wide.u32 %rd1065, %r1570, 16; add.s64 %rd1066, %rd1636, %rd1065; add.s64 %rd248, %rd1066, 8256; @%p184 bra $L__BB4_296; ld.u64 %rd1067, [%rd248]; add.s64 %rd1628, %rd1636, %rd1067; st.local.u32 [%rd4+4], %r473; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1573, %r471, 5; and.b32 %r1574, %r1573, 3840; shl.b32 %r1575, %r8, 1; and.b32 %r1576, %r1575, 240; or.b32 %r1577, %r1576, %r1574; bfe.u32 %r1578, %r9, 3, 4; or.b32 %r1579, %r1577, %r1578; shr.u32 %r1580, %r1577, 6; mul.wide.u32 %rd1068, %r1580, 8; add.s64 %rd1069, %rd1628, %rd1068; and.b32 %r1581, %r1579, 63; shl.b64 %rd1071, %rd1061, %r1581; ld.u64 %rd1072, [%rd1069+544]; and.b64 %rd1073, %rd1072, %rd1071; setp.eq.s64 %p185, %rd1073, 0; mul.wide.u32 %rd1074, %r1579, 16; add.s64 %rd1075, %rd1628, %rd1074; add.s64 %rd250, %rd1075, 1088; @%p185 bra $L__BB4_295; ld.u64 %rd1076, [%rd250]; add.s64 %rd1621, %rd1628, %rd1076; and.b32 %r1582, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1582}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1583, %r471, 6; and.b32 %r1584, %r1583, 448; shl.b32 %r1585, %r8, 3; and.b32 %r1586, %r1585, 56; or.b32 %r1587, %r1586, %r1584; and.b32 %r1588, %r9, 7; or.b32 %r1589, %r1587, %r1588; mul.wide.u32 %rd1077, %r1589, 12; add.s64 %rd1078, %rd1621, %rd1077; ld.f32 %f935, [%rd1078+128]; ld.f32 %f934, [%rd1078+132]; ld.f32 %f933, [%rd1078+136]; mov.b64 %rd1079, {%r472, %r1582}; shr.u64 %rd1080, %rd1079, 32; cvt.u32.u64 %r2289, %rd1080; mov.u32 %r2377, %r473; mov.u32 %r2334, %r472; bra.uni $L__BB4_311; $L__BB4_323: and.b32 %r521, %r471, -4096; setp.ne.s32 %p202, %r521, %r2337; mov.u16 %rs195, 0; @%p202 bra $L__BB4_326; and.b32 %r1647, %r8, -4096; setp.ne.s32 %p203, %r1647, %r2382; @%p203 bra $L__BB4_326; and.b32 %r1648, %r74, -4096; setp.eq.s32 %p204, %r1648, %r2415; selp.u16 %rs195, 1, 0, %p204; $L__BB4_326: setp.eq.s16 %p205, %rs195, 0; @%p205 bra $L__BB4_332; shl.b32 %r1649, %r471, 3; and.b32 %r1650, %r1649, 31744; and.b32 %r1651, %r8, 3968; shr.u32 %r1652, %r1651, 2; or.b32 %r1653, %r1652, %r1650; bfe.u32 %r1654, %r74, 7, 5; or.b32 %r1655, %r1654, %r1653; shr.u32 %r1656, %r1653, 6; mul.wide.u32 %rd1129, %r1656, 8; add.s64 %rd1130, %rd1636, %rd1129; and.b32 %r1657, %r1655, 63; mov.u64 %rd1131, 1; shl.b64 %rd1132, %rd1131, %r1657; ld.u64 %rd1133, [%rd1130+4128]; and.b64 %rd1134, %rd1133, %rd1132; setp.eq.s64 %p206, %rd1134, 0; mul.wide.u32 %rd1135, %r1655, 16; add.s64 %rd1136, %rd1636, %rd1135; add.s64 %rd277, %rd1136, 8256; @%p206 bra $L__BB4_331; ld.u64 %rd1137, [%rd277]; add.s64 %rd1628, %rd1636, %rd1137; st.local.u32 [%rd4+4], %r518; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1658, %r471, 5; and.b32 %r1659, %r1658, 3840; shl.b32 %r1660, %r8, 1; and.b32 %r1661, %r1660, 240; or.b32 %r1662, %r1661, %r1659; bfe.u32 %r1663, %r74, 3, 4; or.b32 %r1664, %r1663, %r1662; shr.u32 %r1665, %r1662, 6; mul.wide.u32 %rd1138, %r1665, 8; add.s64 %rd1139, %rd1628, %rd1138; and.b32 %r1666, %r1664, 63; shl.b64 %rd1141, %rd1131, %r1666; ld.u64 %rd1142, [%rd1139+544]; and.b64 %rd1143, %rd1142, %rd1141; setp.eq.s64 %p207, %rd1143, 0; mul.wide.u32 %rd1144, %r1664, 16; add.s64 %rd1145, %rd1628, %rd1144; add.s64 %rd279, %rd1145, 1088; @%p207 bra $L__BB4_330; ld.u64 %rd1146, [%rd279]; add.s64 %rd1621, %rd1628, %rd1146; and.b32 %r1667, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1667}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1668, %r471, 6; and.b32 %r1669, %r1668, 448; shl.b32 %r1670, %r8, 3; and.b32 %r1671, %r1670, 56; or.b32 %r1672, %r1671, %r1669; and.b32 %r1673, %r74, 7; or.b32 %r1674, %r1672, %r1673; mul.wide.u32 %rd1147, %r1674, 12; add.s64 %rd1148, %rd1621, %rd1147; ld.f32 %f938, [%rd1148+128]; ld.f32 %f937, [%rd1148+132]; ld.f32 %f936, [%rd1148+136]; mov.b64 %rd1149, {%r472, %r1667}; shr.u64 %rd1150, %rd1149, 32; cvt.u32.u64 %r2289, %rd1150; mov.u32 %r2377, %r518; mov.u32 %r2334, %r472; bra.uni $L__BB4_346; $L__BB4_358: and.b32 %r563, %r471, -4096; setp.ne.s32 %p224, %r563, %r2337; mov.u16 %rs198, 0; @%p224 bra $L__BB4_361; and.b32 %r1733, %r139, -4096; setp.ne.s32 %p225, %r1733, %r2382; @%p225 bra $L__BB4_361; and.b32 %r1734, %r74, -4096; setp.eq.s32 %p226, %r1734, %r2415; selp.u16 %rs198, 1, 0, %p226; $L__BB4_361: setp.eq.s16 %p227, %rs198, 0; @%p227 bra $L__BB4_367; shl.b32 %r1735, %r471, 3; and.b32 %r1736, %r1735, 31744; and.b32 %r1737, %r139, 3968; shr.u32 %r1738, %r1737, 2; or.b32 %r1739, %r1738, %r1736; bfe.u32 %r1740, %r74, 7, 5; or.b32 %r1741, %r1739, %r1740; shr.u32 %r1742, %r1739, 6; mul.wide.u32 %rd1200, %r1742, 8; add.s64 %rd1201, %rd1636, %rd1200; and.b32 %r1743, %r1741, 63; mov.u64 %rd1202, 1; shl.b64 %rd1203, %rd1202, %r1743; ld.u64 %rd1204, [%rd1201+4128]; and.b64 %rd1205, %rd1204, %rd1203; setp.eq.s64 %p228, %rd1205, 0; mul.wide.u32 %rd1206, %r1741, 16; add.s64 %rd1207, %rd1636, %rd1206; add.s64 %rd306, %rd1207, 8256; @%p228 bra $L__BB4_366; ld.u64 %rd1208, [%rd306]; add.s64 %rd1628, %rd1636, %rd1208; st.local.u32 [%rd4+4], %r560; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1744, %r471, 5; and.b32 %r1745, %r1744, 3840; shl.b32 %r1746, %r139, 1; and.b32 %r1747, %r1746, 240; or.b32 %r1748, %r1747, %r1745; bfe.u32 %r1749, %r74, 3, 4; or.b32 %r1750, %r1748, %r1749; shr.u32 %r1751, %r1748, 6; mul.wide.u32 %rd1209, %r1751, 8; add.s64 %rd1210, %rd1628, %rd1209; and.b32 %r1752, %r1750, 63; shl.b64 %rd1212, %rd1202, %r1752; ld.u64 %rd1213, [%rd1210+544]; and.b64 %rd1214, %rd1213, %rd1212; setp.eq.s64 %p229, %rd1214, 0; mul.wide.u32 %rd1215, %r1750, 16; add.s64 %rd1216, %rd1628, %rd1215; add.s64 %rd308, %rd1216, 1088; @%p229 bra $L__BB4_365; ld.u64 %rd1217, [%rd308]; add.s64 %rd1621, %rd1628, %rd1217; and.b32 %r1753, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1753}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1754, %r471, 6; and.b32 %r1755, %r1754, 448; shl.b32 %r1756, %r139, 3; and.b32 %r1757, %r1756, 56; or.b32 %r1758, %r1757, %r1755; and.b32 %r1759, %r74, 7; or.b32 %r1760, %r1758, %r1759; mul.wide.u32 %rd1218, %r1760, 12; add.s64 %rd1219, %rd1621, %rd1218; ld.f32 %f941, [%rd1219+128]; ld.f32 %f940, [%rd1219+132]; ld.f32 %f939, [%rd1219+136]; mov.b64 %rd1220, {%r472, %r1753}; shr.u64 %rd1221, %rd1220, 32; cvt.u32.u64 %r2289, %rd1221; mov.u32 %r2377, %r560; mov.u32 %r2334, %r472; bra.uni $L__BB4_381; $L__BB4_393: and.b32 %r602, %r471, -4096; setp.ne.s32 %p246, %r602, %r2337; mov.u16 %rs201, 0; @%p246 bra $L__BB4_396; and.b32 %r1818, %r139, -4096; setp.ne.s32 %p247, %r1818, %r2382; @%p247 bra $L__BB4_396; and.b32 %r1819, %r9, -4096; setp.eq.s32 %p248, %r1819, %r2415; selp.u16 %rs201, 1, 0, %p248; $L__BB4_396: setp.eq.s16 %p249, %rs201, 0; @%p249 bra $L__BB4_402; shl.b32 %r1820, %r471, 3; and.b32 %r1821, %r1820, 31744; and.b32 %r1822, %r139, 3968; shr.u32 %r1823, %r1822, 2; or.b32 %r1824, %r1823, %r1821; bfe.u32 %r1825, %r9, 7, 5; or.b32 %r1826, %r1824, %r1825; shr.u32 %r1827, %r1824, 6; mul.wide.u32 %rd1271, %r1827, 8; add.s64 %rd1272, %rd1636, %rd1271; and.b32 %r1828, %r1826, 63; mov.u64 %rd1273, 1; shl.b64 %rd1274, %rd1273, %r1828; ld.u64 %rd1275, [%rd1272+4128]; and.b64 %rd1276, %rd1275, %rd1274; setp.eq.s64 %p250, %rd1276, 0; mul.wide.u32 %rd1277, %r1826, 16; add.s64 %rd1278, %rd1636, %rd1277; add.s64 %rd335, %rd1278, 8256; @%p250 bra $L__BB4_401; ld.u64 %rd1279, [%rd335]; add.s64 %rd1628, %rd1636, %rd1279; st.local.u32 [%rd4+4], %r599; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1829, %r471, 5; and.b32 %r1830, %r1829, 3840; shl.b32 %r1831, %r139, 1; and.b32 %r1832, %r1831, 240; or.b32 %r1833, %r1832, %r1830; bfe.u32 %r1834, %r9, 3, 4; or.b32 %r1835, %r1833, %r1834; shr.u32 %r1836, %r1833, 6; mul.wide.u32 %rd1280, %r1836, 8; add.s64 %rd1281, %rd1628, %rd1280; and.b32 %r1837, %r1835, 63; shl.b64 %rd1283, %rd1273, %r1837; ld.u64 %rd1284, [%rd1281+544]; and.b64 %rd1285, %rd1284, %rd1283; setp.eq.s64 %p251, %rd1285, 0; mul.wide.u32 %rd1286, %r1835, 16; add.s64 %rd1287, %rd1628, %rd1286; add.s64 %rd337, %rd1287, 1088; @%p251 bra $L__BB4_400; ld.u64 %rd1288, [%rd337]; add.s64 %rd1621, %rd1628, %rd1288; and.b32 %r1838, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1838}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1839, %r471, 7; shl.b32 %r1840, %r139, 3; and.b32 %r1841, %r1840, 56; and.b32 %r1842, %r9, 7; bfi.b32 %r1843, %r1839, %r1842, 6, 3; or.b32 %r1844, %r1843, %r1841; mul.wide.u32 %rd1289, %r1844, 12; add.s64 %rd1290, %rd1621, %rd1289; ld.f32 %f944, [%rd1290+128]; ld.f32 %f943, [%rd1290+132]; ld.f32 %f942, [%rd1290+136]; mov.b64 %rd1291, {%r472, %r1838}; shr.u64 %rd1292, %rd1291, 32; cvt.u32.u64 %r2289, %rd1292; mov.u32 %r2377, %r599; mov.u32 %r2334, %r472; bra.uni $L__BB4_416; $L__BB4_428: and.b32 %r640, %r635, -4096; setp.ne.s32 %p268, %r640, %r2337; mov.u16 %rs204, 0; @%p268 bra $L__BB4_431; and.b32 %r1900, %r139, -4096; setp.ne.s32 %p269, %r1900, %r2382; @%p269 bra $L__BB4_431; and.b32 %r1901, %r9, -4096; setp.eq.s32 %p270, %r1901, %r2415; selp.u16 %rs204, 1, 0, %p270; $L__BB4_431: setp.eq.s16 %p271, %rs204, 0; @%p271 bra $L__BB4_437; shl.b32 %r1902, %r635, 3; and.b32 %r1903, %r1902, 31744; and.b32 %r1904, %r139, 3968; shr.u32 %r1905, %r1904, 2; or.b32 %r1906, %r1905, %r1903; bfe.u32 %r1907, %r9, 7, 5; or.b32 %r1908, %r1906, %r1907; shr.u32 %r1909, %r1906, 6; mul.wide.u32 %rd1342, %r1909, 8; add.s64 %rd1343, %rd1636, %rd1342; and.b32 %r1910, %r1908, 63; mov.u64 %rd1344, 1; shl.b64 %rd1345, %rd1344, %r1910; ld.u64 %rd1346, [%rd1343+4128]; and.b64 %rd1347, %rd1346, %rd1345; setp.eq.s64 %p272, %rd1347, 0; mul.wide.u32 %rd1348, %r1908, 16; add.s64 %rd1349, %rd1636, %rd1348; add.s64 %rd364, %rd1349, 8256; @%p272 bra $L__BB4_436; ld.u64 %rd1350, [%rd364]; add.s64 %rd1628, %rd1636, %rd1350; st.local.u32 [%rd4+4], %r637; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1911, %r635, 5; and.b32 %r1912, %r1911, 3840; shl.b32 %r1913, %r139, 1; and.b32 %r1914, %r1913, 240; or.b32 %r1915, %r1914, %r1912; bfe.u32 %r1916, %r9, 3, 4; or.b32 %r1917, %r1915, %r1916; shr.u32 %r1918, %r1915, 6; mul.wide.u32 %rd1351, %r1918, 8; add.s64 %rd1352, %rd1628, %rd1351; and.b32 %r1919, %r1917, 63; shl.b64 %rd1354, %rd1344, %r1919; ld.u64 %rd1355, [%rd1352+544]; and.b64 %rd1356, %rd1355, %rd1354; setp.eq.s64 %p273, %rd1356, 0; mul.wide.u32 %rd1357, %r1917, 16; add.s64 %rd1358, %rd1628, %rd1357; add.s64 %rd366, %rd1358, 1088; @%p273 bra $L__BB4_435; ld.u64 %rd1359, [%rd366]; add.s64 %rd1621, %rd1628, %rd1359; and.b32 %r1920, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1920}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1921, %r635, 7; shl.b32 %r1922, %r139, 3; and.b32 %r1923, %r1922, 56; and.b32 %r1924, %r9, 7; bfi.b32 %r1925, %r1921, %r1924, 6, 3; or.b32 %r1926, %r1925, %r1923; mul.wide.u32 %rd1360, %r1926, 12; add.s64 %rd1361, %rd1621, %rd1360; ld.f32 %f947, [%rd1361+128]; ld.f32 %f946, [%rd1361+132]; ld.f32 %f945, [%rd1361+136]; mov.b64 %rd1362, {%r636, %r1920}; shr.u64 %rd1363, %rd1362, 32; cvt.u32.u64 %r2289, %rd1363; mov.u32 %r2377, %r637; mov.u32 %r2334, %r636; bra.uni $L__BB4_451; $L__BB4_463: and.b32 %r673, %r635, -4096; setp.ne.s32 %p290, %r673, %r2337; mov.u16 %rs207, 0; @%p290 bra $L__BB4_466; and.b32 %r1984, %r139, -4096; setp.ne.s32 %p291, %r1984, %r2382; @%p291 bra $L__BB4_466; and.b32 %r1985, %r74, -4096; setp.eq.s32 %p292, %r1985, %r2415; selp.u16 %rs207, 1, 0, %p292; $L__BB4_466: setp.eq.s16 %p293, %rs207, 0; @%p293 bra $L__BB4_472; shl.b32 %r1986, %r635, 3; and.b32 %r1987, %r1986, 31744; and.b32 %r1988, %r139, 3968; shr.u32 %r1989, %r1988, 2; or.b32 %r1990, %r1989, %r1987; bfe.u32 %r1991, %r74, 7, 5; or.b32 %r1992, %r1990, %r1991; shr.u32 %r1993, %r1990, 6; mul.wide.u32 %rd1413, %r1993, 8; add.s64 %rd1414, %rd1636, %rd1413; and.b32 %r1994, %r1992, 63; mov.u64 %rd1415, 1; shl.b64 %rd1416, %rd1415, %r1994; ld.u64 %rd1417, [%rd1414+4128]; and.b64 %rd1418, %rd1417, %rd1416; setp.eq.s64 %p294, %rd1418, 0; mul.wide.u32 %rd1419, %r1992, 16; add.s64 %rd1420, %rd1636, %rd1419; add.s64 %rd393, %rd1420, 8256; @%p294 bra $L__BB4_471; ld.u64 %rd1421, [%rd393]; add.s64 %rd1628, %rd1636, %rd1421; st.local.u32 [%rd4+4], %r670; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1995, %r635, 5; and.b32 %r1996, %r1995, 3840; shl.b32 %r1997, %r139, 1; and.b32 %r1998, %r1997, 240; or.b32 %r1999, %r1998, %r1996; bfe.u32 %r2000, %r74, 3, 4; or.b32 %r2001, %r1999, %r2000; shr.u32 %r2002, %r1999, 6; mul.wide.u32 %rd1422, %r2002, 8; add.s64 %rd1423, %rd1628, %rd1422; and.b32 %r2003, %r2001, 63; shl.b64 %rd1425, %rd1415, %r2003; ld.u64 %rd1426, [%rd1423+544]; and.b64 %rd1427, %rd1426, %rd1425; setp.eq.s64 %p295, %rd1427, 0; mul.wide.u32 %rd1428, %r2001, 16; add.s64 %rd1429, %rd1628, %rd1428; add.s64 %rd395, %rd1429, 1088; @%p295 bra $L__BB4_470; ld.u64 %rd1430, [%rd395]; add.s64 %rd1621, %rd1628, %rd1430; and.b32 %r2004, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2004}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r2005, %r635, 6; and.b32 %r2006, %r2005, 448; shl.b32 %r2007, %r139, 3; and.b32 %r2008, %r2007, 56; or.b32 %r2009, %r2008, %r2006; and.b32 %r2010, %r74, 7; or.b32 %r2011, %r2009, %r2010; mul.wide.u32 %rd1431, %r2011, 12; add.s64 %rd1432, %rd1621, %rd1431; ld.f32 %f950, [%rd1432+128]; ld.f32 %f949, [%rd1432+132]; ld.f32 %f948, [%rd1432+136]; mov.b64 %rd1433, {%r636, %r2004}; shr.u64 %rd1434, %rd1433, 32; cvt.u32.u64 %r2289, %rd1434; mov.u32 %r2377, %r670; mov.u32 %r2334, %r636; bra.uni $L__BB4_486; $L__BB4_498: and.b32 %r703, %r635, -4096; setp.ne.s32 %p312, %r703, %r2337; mov.u16 %rs210, 0; @%p312 bra $L__BB4_501; and.b32 %r2071, %r8, -4096; setp.ne.s32 %p313, %r2071, %r2382; @%p313 bra $L__BB4_501; and.b32 %r2072, %r74, -4096; setp.eq.s32 %p314, %r2072, %r2415; selp.u16 %rs210, 1, 0, %p314; $L__BB4_501: setp.eq.s16 %p315, %rs210, 0; @%p315 bra $L__BB4_507; shl.b32 %r2073, %r635, 3; and.b32 %r2074, %r2073, 31744; and.b32 %r2075, %r8, 3968; shr.u32 %r2076, %r2075, 2; or.b32 %r2077, %r2074, %r2076; bfe.u32 %r2078, %r74, 7, 5; or.b32 %r2079, %r2078, %r2077; shr.u32 %r2080, %r2077, 6; mul.wide.u32 %rd1484, %r2080, 8; add.s64 %rd1485, %rd1636, %rd1484; and.b32 %r2081, %r2079, 63; mov.u64 %rd1486, 1; shl.b64 %rd1487, %rd1486, %r2081; ld.u64 %rd1488, [%rd1485+4128]; and.b64 %rd1489, %rd1488, %rd1487; setp.eq.s64 %p316, %rd1489, 0; mul.wide.u32 %rd1490, %r2079, 16; add.s64 %rd1491, %rd1636, %rd1490; add.s64 %rd422, %rd1491, 8256; @%p316 bra $L__BB4_506; ld.u64 %rd1492, [%rd422]; add.s64 %rd1628, %rd1636, %rd1492; st.local.u32 [%rd4+4], %r700; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r2082, %r635, 5; and.b32 %r2083, %r2082, 3840; shl.b32 %r2084, %r8, 1; and.b32 %r2085, %r2084, 240; or.b32 %r2086, %r2083, %r2085; bfe.u32 %r2087, %r74, 3, 4; or.b32 %r2088, %r2087, %r2086; shr.u32 %r2089, %r2086, 6; mul.wide.u32 %rd1493, %r2089, 8; add.s64 %rd1494, %rd1628, %rd1493; and.b32 %r2090, %r2088, 63; shl.b64 %rd1496, %rd1486, %r2090; ld.u64 %rd1497, [%rd1494+544]; and.b64 %rd1498, %rd1497, %rd1496; setp.eq.s64 %p317, %rd1498, 0; mul.wide.u32 %rd1499, %r2088, 16; add.s64 %rd1500, %rd1628, %rd1499; add.s64 %rd424, %rd1500, 1088; @%p317 bra $L__BB4_505; ld.u64 %rd1501, [%rd424]; add.s64 %rd1621, %rd1628, %rd1501; and.b32 %r2091, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2091}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r2092, %r635, 6; and.b32 %r2093, %r2092, 448; shl.b32 %r2094, %r8, 3; and.b32 %r2095, %r2094, 56; or.b32 %r2096, %r2093, %r2095; and.b32 %r2097, %r74, 7; or.b32 %r2098, %r2096, %r2097; mul.wide.u32 %rd1502, %r2098, 12; add.s64 %rd1503, %rd1621, %rd1502; ld.f32 %f953, [%rd1503+128]; ld.f32 %f952, [%rd1503+132]; ld.f32 %f951, [%rd1503+136]; mov.b64 %rd1504, {%r636, %r2091}; shr.u64 %rd1505, %rd1504, 32; cvt.u32.u64 %r2289, %rd1505; mov.u32 %r2377, %r700; mov.u32 %r2334, %r636; bra.uni $L__BB4_521; $L__BB4_541: ld.f32 %f956, [%rd457]; ld.f32 %f955, [%rd457+4]; ld.f32 %f954, [%rd457+8]; bra.uni $L__BB4_553; $L__BB4_42: ld.f32 %f914, [%rd43]; ld.f32 %f913, [%rd43+4]; ld.f32 %f912, [%rd43+8]; bra.uni $L__BB4_66; $L__BB4_77: ld.f32 %f917, [%rd72]; ld.f32 %f916, [%rd72+4]; ld.f32 %f915, [%rd72+8]; bra.uni $L__BB4_101; $L__BB4_112: ld.f32 %f920, [%rd101]; ld.f32 %f919, [%rd101+4]; ld.f32 %f918, [%rd101+8]; bra.uni $L__BB4_136; $L__BB4_147: ld.f32 %f923, [%rd130]; ld.f32 %f922, [%rd130+4]; ld.f32 %f921, [%rd130+8]; bra.uni $L__BB4_171; $L__BB4_182: ld.f32 %f926, [%rd159]; ld.f32 %f925, [%rd159+4]; ld.f32 %f924, [%rd159+8]; bra.uni $L__BB4_206; $L__BB4_217: ld.f32 %f929, [%rd188]; ld.f32 %f928, [%rd188+4]; ld.f32 %f927, [%rd188+8]; bra.uni $L__BB4_241; $L__BB4_252: ld.f32 %f932, [%rd217]; ld.f32 %f931, [%rd217+4]; ld.f32 %f930, [%rd217+8]; bra.uni $L__BB4_276; $L__BB4_287: ld.f32 %f935, [%rd246]; ld.f32 %f934, [%rd246+4]; ld.f32 %f933, [%rd246+8]; bra.uni $L__BB4_311; $L__BB4_322: ld.f32 %f938, [%rd275]; ld.f32 %f937, [%rd275+4]; ld.f32 %f936, [%rd275+8]; bra.uni $L__BB4_346; $L__BB4_357: ld.f32 %f941, [%rd304]; ld.f32 %f940, [%rd304+4]; ld.f32 %f939, [%rd304+8]; bra.uni $L__BB4_381; $L__BB4_392: ld.f32 %f944, [%rd333]; ld.f32 %f943, [%rd333+4]; ld.f32 %f942, [%rd333+8]; bra.uni $L__BB4_416; $L__BB4_427: ld.f32 %f947, [%rd362]; ld.f32 %f946, [%rd362+4]; ld.f32 %f945, [%rd362+8]; bra.uni $L__BB4_451; $L__BB4_462: ld.f32 %f950, [%rd391]; ld.f32 %f949, [%rd391+4]; ld.f32 %f948, [%rd391+8]; bra.uni $L__BB4_486; $L__BB4_497: ld.f32 %f953, [%rd420]; ld.f32 %f952, [%rd420+4]; ld.f32 %f951, [%rd420+8]; bra.uni $L__BB4_521; $L__BB4_540: ld.f32 %f956, [%rd459]; ld.f32 %f955, [%rd459+4]; ld.f32 %f954, [%rd459+8]; bra.uni $L__BB4_553; $L__BB4_52: ld.u32 %r81, [%rd5]; mov.u64 %rd1620, 0; setp.eq.s32 %p32, %r81, 0; @%p32 bra $L__BB4_57; shr.u32 %r2224, %r7, 12; cvt.u64.u32 %rd1610, %r2224; shl.b64 %rd1609, %rd1610, 42; or.b64 %rd592, %rd19, %rd1609; shr.u32 %r999, %r74, 12; cvt.u64.u32 %rd593, %r999; or.b64 %rd51, %rd592, %rd593; mov.u32 %r2286, 0; $L__BB4_54: mul.wide.u32 %rd594, %r2286, 32; add.s64 %rd595, %rd17, %rd594; ld.u64 %rd596, [%rd595]; setp.eq.s64 %p33, %rd596, %rd51; @%p33 bra $L__BB4_56; add.s32 %r2286, %r2286, 1; setp.lt.u32 %p34, %r2286, %r81; @%p34 bra $L__BB4_54; bra.uni $L__BB4_57; $L__BB4_87: ld.u32 %r146, [%rd5]; mov.u64 %rd1626, 0; setp.eq.s32 %p54, %r146, 0; @%p54 bra $L__BB4_92; shr.u32 %r2223, %r7, 12; cvt.u64.u32 %rd1607, %r2223; shl.b64 %rd1606, %rd1607, 42; shr.u32 %r1084, %r74, 12; cvt.u64.u32 %rd661, %r1084; or.b64 %rd662, %rd1606, %rd661; shr.u32 %r1085, %r139, 12; mul.wide.u32 %rd663, %r1085, 2097152; or.b64 %rd80, %rd662, %rd663; mov.u32 %r2335, 0; $L__BB4_89: mul.wide.u32 %rd664, %r2335, 32; add.s64 %rd665, %rd17, %rd664; ld.u64 %rd666, [%rd665]; setp.eq.s64 %p55, %rd666, %rd80; @%p55 bra $L__BB4_91; add.s32 %r2335, %r2335, 1; setp.lt.u32 %p56, %r2335, %r146; @%p56 bra $L__BB4_89; bra.uni $L__BB4_92; $L__BB4_122: ld.u32 %r207, [%rd5]; mov.u64 %rd1632, 0; setp.eq.s32 %p76, %r207, 0; @%p76 bra $L__BB4_127; shr.u32 %r1167, %r139, 12; mul.wide.u32 %rd731, %r1167, 2097152; or.b64 %rd109, %rd21, %rd731; mov.u32 %r2381, 0; $L__BB4_124: mul.wide.u32 %rd732, %r2381, 32; add.s64 %rd733, %rd17, %rd732; ld.u64 %rd734, [%rd733]; setp.eq.s64 %p77, %rd734, %rd109; @%p77 bra $L__BB4_126; add.s32 %r2381, %r2381, 1; setp.lt.u32 %p78, %r2381, %r207; @%p78 bra $L__BB4_124; bra.uni $L__BB4_127; $L__BB4_157: ld.u32 %r269, [%rd5]; mov.u64 %rd1638, 0; setp.eq.s32 %p98, %r269, 0; @%p98 bra $L__BB4_162; shr.u32 %r1248, %r259, 12; cvt.u64.u32 %rd799, %r1248; shl.b64 %rd800, %rd799, 42; or.b64 %rd801, %rd800, %rd18; shr.u32 %r1249, %r139, 12; mul.wide.u32 %rd802, %r1249, 2097152; or.b64 %rd138, %rd801, %rd802; mov.u32 %r2424, 0; $L__BB4_159: mul.wide.u32 %rd803, %r2424, 32; add.s64 %rd804, %rd17, %rd803; ld.u64 %rd805, [%rd804]; setp.eq.s64 %p99, %rd805, %rd138; @%p99 bra $L__BB4_161; add.s32 %r2424, %r2424, 1; setp.lt.u32 %p100, %r2424, %r269; @%p100 bra $L__BB4_159; bra.uni $L__BB4_162; $L__BB4_192: ld.u32 %r326, [%rd5]; mov.u64 %rd1644, 0; setp.eq.s32 %p120, %r326, 0; @%p120 bra $L__BB4_197; shr.u32 %r1333, %r259, 12; cvt.u64.u32 %rd870, %r1333; shr.u32 %r1334, %r74, 12; cvt.u64.u32 %rd871, %r1334; bfi.b64 %rd872, %rd870, %rd871, 42, 22; shr.u32 %r1335, %r139, 12; mul.wide.u32 %rd873, %r1335, 2097152; or.b64 %rd167, %rd872, %rd873; mov.u32 %r2464, 0; $L__BB4_194: mul.wide.u32 %rd874, %r2464, 32; add.s64 %rd875, %rd17, %rd874; ld.u64 %rd876, [%rd875]; setp.eq.s64 %p121, %rd876, %rd167; @%p121 bra $L__BB4_196; add.s32 %r2464, %r2464, 1; setp.lt.u32 %p122, %r2464, %r326; @%p122 bra $L__BB4_194; bra.uni $L__BB4_197; $L__BB4_227: ld.u32 %r380, [%rd5]; mov.u64 %rd1650, 0; setp.eq.s32 %p142, %r380, 0; @%p142 bra $L__BB4_232; shr.u32 %r1420, %r259, 12; cvt.u64.u32 %rd941, %r1420; shl.b64 %rd942, %rd941, 42; or.b64 %rd943, %rd942, %rd19; shr.u32 %r1421, %r74, 12; cvt.u64.u32 %rd944, %r1421; or.b64 %rd196, %rd943, %rd944; mov.u32 %r2501, 0; $L__BB4_229: cvt.u64.u32 %rd197, %r2501; mul.wide.u32 %rd945, %r2501, 32; add.s64 %rd946, %rd17, %rd945; ld.u64 %rd947, [%rd946]; setp.eq.s64 %p143, %rd947, %rd196; add.s32 %r2501, %r2501, 1; @%p143 bra $L__BB4_231; setp.lt.u32 %p144, %r2501, %r380; @%p144 bra $L__BB4_229; bra.uni $L__BB4_232; $L__BB4_262: ld.u32 %r431, [%rd5]; mov.u64 %rd1656, 0; setp.eq.s32 %p164, %r431, 0; @%p164 bra $L__BB4_267; shr.u32 %r1506, %r259, 12; cvt.u64.u32 %rd1012, %r1506; shl.b64 %rd1013, %rd1012, 42; or.b64 %rd1014, %rd19, %rd18; or.b64 %rd225, %rd1014, %rd1013; mov.u32 %r2535, 0; $L__BB4_264: cvt.u64.u32 %rd226, %r2535; mul.wide.u32 %rd1015, %r2535, 32; add.s64 %rd1016, %rd17, %rd1015; ld.u64 %rd1017, [%rd1016]; setp.eq.s64 %p165, %rd1017, %rd225; add.s32 %r2535, %r2535, 1; @%p165 bra $L__BB4_266; setp.lt.u32 %p166, %r2535, %r431; @%p166 bra $L__BB4_264; bra.uni $L__BB4_267; $L__BB4_297: ld.u32 %r481, [%rd5]; mov.u64 %rd1662, 0; setp.eq.s32 %p186, %r481, 0; @%p186 bra $L__BB4_302; shr.u32 %r1591, %r471, 12; cvt.u64.u32 %rd1082, %r1591; shl.b64 %rd1083, %rd1082, 42; or.b64 %rd1084, %rd1083, %rd18; or.b64 %rd254, %rd1084, %rd19; mov.u32 %r2566, 0; $L__BB4_299: cvt.u64.u32 %rd255, %r2566; mul.wide.u32 %rd1085, %r2566, 32; add.s64 %rd1086, %rd17, %rd1085; ld.u64 %rd1087, [%rd1086]; setp.eq.s64 %p187, %rd1087, %rd254; add.s32 %r2566, %r2566, 1; @%p187 bra $L__BB4_301; setp.lt.u32 %p188, %r2566, %r481; @%p188 bra $L__BB4_299; bra.uni $L__BB4_302; $L__BB4_332: ld.u32 %r526, [%rd5]; mov.u64 %rd1668, 0; setp.eq.s32 %p208, %r526, 0; @%p208 bra $L__BB4_337; shr.u32 %r1676, %r471, 12; cvt.u64.u32 %rd1152, %r1676; shl.b64 %rd1153, %rd1152, 42; or.b64 %rd1154, %rd19, %rd1153; shr.u32 %r1677, %r74, 12; cvt.u64.u32 %rd1155, %r1677; or.b64 %rd283, %rd1154, %rd1155; mov.u32 %r2594, 0; $L__BB4_334: cvt.u64.u32 %rd284, %r2594; mul.wide.u32 %rd1156, %r2594, 32; add.s64 %rd1157, %rd17, %rd1156; ld.u64 %rd1158, [%rd1157]; setp.eq.s64 %p209, %rd1158, %rd283; add.s32 %r2594, %r2594, 1; @%p209 bra $L__BB4_336; setp.lt.u32 %p210, %r2594, %r526; @%p210 bra $L__BB4_334; bra.uni $L__BB4_337; $L__BB4_367: ld.u32 %r568, [%rd5]; mov.u64 %rd1674, 0; setp.eq.s32 %p230, %r568, 0; @%p230 bra $L__BB4_372; shr.u32 %r1762, %r471, 12; cvt.u64.u32 %rd1223, %r1762; shr.u32 %r1763, %r74, 12; cvt.u64.u32 %rd1224, %r1763; bfi.b64 %rd1225, %rd1223, %rd1224, 42, 22; shr.u32 %r1764, %r139, 12; mul.wide.u32 %rd1226, %r1764, 2097152; or.b64 %rd312, %rd1225, %rd1226; mov.u32 %r2619, 0; $L__BB4_369: cvt.u64.u32 %rd313, %r2619; mul.wide.u32 %rd1227, %r2619, 32; add.s64 %rd1228, %rd17, %rd1227; ld.u64 %rd1229, [%rd1228]; setp.eq.s64 %p231, %rd1229, %rd312; add.s32 %r2619, %r2619, 1; @%p231 bra $L__BB4_371; setp.lt.u32 %p232, %r2619, %r568; @%p232 bra $L__BB4_369; bra.uni $L__BB4_372; $L__BB4_402: ld.u32 %r607, [%rd5]; mov.u64 %rd1680, 0; setp.eq.s32 %p252, %r607, 0; @%p252 bra $L__BB4_407; shr.u32 %r1846, %r471, 12; cvt.u64.u32 %rd1294, %r1846; shl.b64 %rd1295, %rd1294, 42; or.b64 %rd1296, %rd1295, %rd18; shr.u32 %r1847, %r139, 12; mul.wide.u32 %rd1297, %r1847, 2097152; or.b64 %rd341, %rd1296, %rd1297; mov.u32 %r2641, 0; $L__BB4_404: cvt.u64.u32 %rd342, %r2641; mul.wide.u32 %rd1298, %r2641, 32; add.s64 %rd1299, %rd17, %rd1298; ld.u64 %rd1300, [%rd1299]; setp.eq.s64 %p253, %rd1300, %rd341; add.s32 %r2641, %r2641, 1; @%p253 bra $L__BB4_406; setp.lt.u32 %p254, %r2641, %r607; @%p254 bra $L__BB4_404; bra.uni $L__BB4_407; $L__BB4_437: ld.u32 %r645, [%rd5]; mov.u64 %rd1686, 0; setp.eq.s32 %p274, %r645, 0; @%p274 bra $L__BB4_442; shr.u32 %r1928, %r635, 12; cvt.u64.u32 %rd1365, %r1928; shl.b64 %rd1366, %rd1365, 42; or.b64 %rd1367, %rd1366, %rd18; shr.u32 %r1929, %r139, 12; mul.wide.u32 %rd1368, %r1929, 2097152; or.b64 %rd370, %rd1367, %rd1368; mov.u32 %r2660, 0; $L__BB4_439: cvt.u64.u32 %rd371, %r2660; mul.wide.u32 %rd1369, %r2660, 32; add.s64 %rd1370, %rd17, %rd1369; ld.u64 %rd1371, [%rd1370]; setp.eq.s64 %p275, %rd1371, %rd370; add.s32 %r2660, %r2660, 1; @%p275 bra $L__BB4_441; setp.lt.u32 %p276, %r2660, %r645; @%p276 bra $L__BB4_439; bra.uni $L__BB4_442; $L__BB4_472: ld.u32 %r678, [%rd5]; mov.u64 %rd1692, 0; setp.eq.s32 %p296, %r678, 0; @%p296 bra $L__BB4_477; shr.u32 %r2013, %r635, 12; cvt.u64.u32 %rd1436, %r2013; shr.u32 %r2014, %r74, 12; cvt.u64.u32 %rd1437, %r2014; bfi.b64 %rd1438, %rd1436, %rd1437, 42, 22; shr.u32 %r2015, %r139, 12; mul.wide.u32 %rd1439, %r2015, 2097152; or.b64 %rd399, %rd1438, %rd1439; mov.u32 %r2676, 0; $L__BB4_474: cvt.u64.u32 %rd400, %r2676; mul.wide.u32 %rd1440, %r2676, 32; add.s64 %rd1441, %rd17, %rd1440; ld.u64 %rd1442, [%rd1441]; setp.eq.s64 %p297, %rd1442, %rd399; add.s32 %r2676, %r2676, 1; @%p297 bra $L__BB4_476; setp.lt.u32 %p298, %r2676, %r678; @%p298 bra $L__BB4_474; bra.uni $L__BB4_477; $L__BB4_507: ld.u32 %r708, [%rd5]; mov.u64 %rd1698, 0; setp.eq.s32 %p318, %r708, 0; @%p318 bra $L__BB4_512; shr.u32 %r2100, %r635, 12; cvt.u64.u32 %rd1507, %r2100; shl.b64 %rd1508, %rd1507, 42; or.b64 %rd1509, %rd1508, %rd19; shr.u32 %r2101, %r74, 12; cvt.u64.u32 %rd1510, %r2101; or.b64 %rd428, %rd1509, %rd1510; mov.u32 %r2689, 0; $L__BB4_509: cvt.u64.u32 %rd429, %r2689; mul.wide.u32 %rd1511, %r2689, 32; add.s64 %rd1512, %rd17, %rd1511; ld.u64 %rd1513, [%rd1512]; setp.eq.s64 %p319, %rd1513, %rd428; add.s32 %r2689, %r2689, 1; @%p319 bra $L__BB4_511; setp.lt.u32 %p320, %r2689, %r708; @%p320 bra $L__BB4_509; bra.uni $L__BB4_512; $L__BB4_551: ld.f32 %f956, [%rd463]; ld.f32 %f955, [%rd463+4]; ld.f32 %f954, [%rd463+8]; bra.uni $L__BB4_553; $L__BB4_548: ld.f32 %f956, [%rd460]; ld.f32 %f955, [%rd460+4]; ld.f32 %f954, [%rd460+8]; bra.uni $L__BB4_553; $L__BB4_51: ld.f32 %f914, [%rd45]; ld.f32 %f913, [%rd45+4]; ld.f32 %f912, [%rd45+8]; bra.uni $L__BB4_66; $L__BB4_86: ld.f32 %f917, [%rd74]; ld.f32 %f916, [%rd74+4]; ld.f32 %f915, [%rd74+8]; bra.uni $L__BB4_101; $L__BB4_121: ld.f32 %f920, [%rd103]; ld.f32 %f919, [%rd103+4]; ld.f32 %f918, [%rd103+8]; bra.uni $L__BB4_136; $L__BB4_156: ld.f32 %f923, [%rd132]; ld.f32 %f922, [%rd132+4]; ld.f32 %f921, [%rd132+8]; bra.uni $L__BB4_171; $L__BB4_191: ld.f32 %f926, [%rd161]; ld.f32 %f925, [%rd161+4]; ld.f32 %f924, [%rd161+8]; bra.uni $L__BB4_206; $L__BB4_226: ld.f32 %f929, [%rd190]; ld.f32 %f928, [%rd190+4]; ld.f32 %f927, [%rd190+8]; bra.uni $L__BB4_241; $L__BB4_261: ld.f32 %f932, [%rd219]; ld.f32 %f931, [%rd219+4]; ld.f32 %f930, [%rd219+8]; bra.uni $L__BB4_276; $L__BB4_296: ld.f32 %f935, [%rd248]; ld.f32 %f934, [%rd248+4]; ld.f32 %f933, [%rd248+8]; bra.uni $L__BB4_311; $L__BB4_331: ld.f32 %f938, [%rd277]; ld.f32 %f937, [%rd277+4]; ld.f32 %f936, [%rd277+8]; bra.uni $L__BB4_346; $L__BB4_366: ld.f32 %f941, [%rd306]; ld.f32 %f940, [%rd306+4]; ld.f32 %f939, [%rd306+8]; bra.uni $L__BB4_381; $L__BB4_401: ld.f32 %f944, [%rd335]; ld.f32 %f943, [%rd335+4]; ld.f32 %f942, [%rd335+8]; bra.uni $L__BB4_416; $L__BB4_436: ld.f32 %f947, [%rd364]; ld.f32 %f946, [%rd364+4]; ld.f32 %f945, [%rd364+8]; bra.uni $L__BB4_451; $L__BB4_471: ld.f32 %f950, [%rd393]; ld.f32 %f949, [%rd393+4]; ld.f32 %f948, [%rd393+8]; bra.uni $L__BB4_486; $L__BB4_506: ld.f32 %f953, [%rd422]; ld.f32 %f952, [%rd422+4]; ld.f32 %f951, [%rd422+8]; bra.uni $L__BB4_521; $L__BB4_547: ld.f32 %f956, [%rd462]; ld.f32 %f955, [%rd462+4]; ld.f32 %f954, [%rd462+8]; $L__BB4_553: ld.param.f32 %f908, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2]; ld.param.f32 %f907, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1]; add.ftz.f32 %f906, %f907, 0fBF000000; add.ftz.f32 %f905, %f908, 0fBF000000; ld.param.f32 %f904, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; cvt.rmi.ftz.f32.f32 %f903, %f905; add.ftz.f32 %f902, %f904, 0fBF000000; sub.ftz.f32 %f901, %f905, %f903; cvt.rmi.ftz.f32.f32 %f900, %f902; cvt.rmi.ftz.f32.f32 %f899, %f906; sub.ftz.f32 %f898, %f902, %f900; mul.ftz.f32 %f897, %f898, %f898; mul.ftz.f32 %f896, %f901, %f901; sub.ftz.f32 %f895, %f906, %f899; mul.ftz.f32 %f894, %f901, %f896; mul.ftz.f32 %f893, %f895, 0f40400000; mul.ftz.f32 %f892, %f898, %f897; mov.f32 %f891, 0f40400000; mul.ftz.f32 %f890, %f895, %f895; mul.ftz.f32 %f889, %f894, 0f3E2AAAAB; mul.ftz.f32 %f888, %f892, 0f3E2AAAAB; sub.ftz.f32 %f887, %f891, %f901; fma.rn.ftz.f32 %f886, %f901, %f887, 0fC0400000; mul.ftz.f32 %f885, %f898, 0f40400000; add.ftz.f32 %f884, %f885, 0fC0C00000; fma.rn.ftz.f32 %f883, %f897, %f884, 0f40800000; mul.ftz.f32 %f882, %f901, 0f40400000; add.ftz.f32 %f881, %f882, 0fC0C00000; mul.ftz.f32 %f880, %f883, 0f3E2AAAAB; fma.rn.ftz.f32 %f879, %f896, %f881, 0f40800000; mul.ftz.f32 %f878, %f879, 0f3E2AAAAB; fma.rn.ftz.f32 %f877, %f901, %f886, 0f3F800000; fma.rn.ftz.f32 %f876, %f877, 0f3E2AAAAB, %f878; fma.rn.ftz.f32 %f875, %f681, 0f3E2AAAAB, %f888; sub.ftz.f32 %f874, %f891, %f885; fma.rn.ftz.f32 %f873, %f898, %f874, 0f40400000; fma.rn.ftz.f32 %f872, %f898, %f873, 0f3F800000; fma.rn.ftz.f32 %f871, %f872, 0f3E2AAAAB, %f888; div.approx.ftz.f32 %f870, %f888, %f871; add.ftz.f32 %f869, %f870, 0f3F800000; add.ftz.f32 %f868, %f900, %f869; div.approx.ftz.f32 %f867, %f878, %f876; add.ftz.f32 %f866, %f867, 0fBF800000; add.ftz.f32 %f865, %f903, %f866; fma.rn.ftz.f32 %f864, %f673, 0f3E2AAAAB, %f880; fma.rn.ftz.f32 %f863, %f668, 0f3E2AAAAB, %f889; sub.ftz.f32 %f862, %f891, %f898; fma.rn.ftz.f32 %f861, %f898, %f862, 0fC0400000; fma.rn.ftz.f32 %f860, %f898, %f861, 0f3F800000; fma.rn.ftz.f32 %f859, %f860, 0f3E2AAAAB, %f880; div.approx.ftz.f32 %f858, %f880, %f859; add.ftz.f32 %f857, %f858, 0fBF800000; add.ftz.f32 %f856, %f900, %f857; sub.ftz.f32 %f855, %f891, %f882; fma.rn.ftz.f32 %f854, %f901, %f855, 0f40400000; fma.rn.ftz.f32 %f853, %f901, %f854, 0f3F800000; fma.rn.ftz.f32 %f852, %f853, 0f3E2AAAAB, %f889; div.approx.ftz.f32 %f851, %f889, %f852; add.ftz.f32 %f850, %f851, 0f3F800000; add.ftz.f32 %f849, %f903, %f850; add.u64 %rd1605, %SP, 80; sub.ftz.f32 %f848, %f891, %f895; fma.rn.ftz.f32 %f847, %f895, %f848, 0fC0400000; add.ftz.f32 %f846, %f893, 0fC0C00000; fma.rn.ftz.f32 %f845, %f890, %f846, 0f40800000; mul.ftz.f32 %f844, %f845, 0f3E2AAAAB; fma.rn.ftz.f32 %f843, %f895, %f847, 0f3F800000; fma.rn.ftz.f32 %f842, %f843, 0f3E2AAAAB, %f844; sub.ftz.f32 %f841, %f891, %f893; fma.rn.ftz.f32 %f840, %f895, %f841, 0f40400000; mul.ftz.f32 %f839, %f895, %f890; mul.ftz.f32 %f838, %f839, 0f3E2AAAAB; fma.rn.ftz.f32 %f837, %f895, %f840, 0f3F800000; fma.rn.ftz.f32 %f836, %f837, 0f3E2AAAAB, %f838; sub.ftz.f32 %f744, %f32, %f322; sub.ftz.f32 %f745, %f953, %f956; sub.ftz.f32 %f746, %f952, %f955; sub.ftz.f32 %f747, %f951, %f954; fma.rn.ftz.f32 %f748, %f39, %f745, %f956; fma.rn.ftz.f32 %f749, %f39, %f746, %f955; fma.rn.ftz.f32 %f750, %f39, %f747, %f954; sub.ftz.f32 %f751, %f545, %f748; sub.ftz.f32 %f752, %f546, %f749; sub.ftz.f32 %f753, %f547, %f750; fma.rn.ftz.f32 %f754, %f38, %f751, %f748; fma.rn.ftz.f32 %f755, %f38, %f752, %f749; fma.rn.ftz.f32 %f756, %f38, %f753, %f750; sub.ftz.f32 %f757, %f754, %f470; sub.ftz.f32 %f758, %f755, %f471; sub.ftz.f32 %f759, %f756, %f472; fma.rn.ftz.f32 %f760, %f744, %f757, %f470; fma.rn.ftz.f32 %f761, %f744, %f758, %f471; fma.rn.ftz.f32 %f762, %f744, %f759, %f472; mul.ftz.f32 %f763, %f836, %f760; mul.ftz.f32 %f764, %f836, %f761; mul.ftz.f32 %f765, %f836, %f762; fma.rn.ftz.f32 %f766, %f842, %f319, %f763; fma.rn.ftz.f32 %f767, %f842, %f320, %f764; fma.rn.ftz.f32 %f768, %f842, %f321, %f765; mul.ftz.f32 %f769, %f876, %f766; mul.ftz.f32 %f770, %f876, %f767; mul.ftz.f32 %f771, %f876, %f768; { // callseq 7, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f31; st.param.f32 [param1+4], %f849; st.param.f32 [param1+8], %f856; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f772, [retval0+0]; ld.param.f32 %f773, [retval0+4]; ld.param.f32 %f774, [retval0+8]; } // callseq 7 { // callseq 8, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f32; st.param.f32 [param1+4], %f849; st.param.f32 [param1+8], %f856; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f775, [retval0+0]; ld.param.f32 %f776, [retval0+4]; ld.param.f32 %f777, [retval0+8]; } // callseq 8 mul.ftz.f32 %f778, %f836, %f775; mul.ftz.f32 %f779, %f836, %f776; mul.ftz.f32 %f780, %f836, %f777; fma.rn.ftz.f32 %f781, %f842, %f772, %f778; fma.rn.ftz.f32 %f782, %f842, %f773, %f779; fma.rn.ftz.f32 %f783, %f842, %f774, %f780; fma.rn.ftz.f32 %f784, %f852, %f781, %f769; fma.rn.ftz.f32 %f785, %f852, %f782, %f770; fma.rn.ftz.f32 %f786, %f852, %f783, %f771; mul.ftz.f32 %f787, %f859, %f784; mul.ftz.f32 %f788, %f859, %f785; { // callseq 9, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f31; st.param.f32 [param1+4], %f865; st.param.f32 [param1+8], %f868; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f789, [retval0+0]; ld.param.f32 %f790, [retval0+4]; ld.param.f32 %f791, [retval0+8]; } // callseq 9 { // callseq 10, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f32; st.param.f32 [param1+4], %f865; st.param.f32 [param1+8], %f868; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f792, [retval0+0]; ld.param.f32 %f793, [retval0+4]; ld.param.f32 %f794, [retval0+8]; } // callseq 10 mul.ftz.f32 %f795, %f836, %f792; mul.ftz.f32 %f796, %f836, %f793; mul.ftz.f32 %f797, %f836, %f794; fma.rn.ftz.f32 %f798, %f842, %f789, %f795; fma.rn.ftz.f32 %f799, %f842, %f790, %f796; fma.rn.ftz.f32 %f800, %f842, %f791, %f797; { // callseq 11, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f31; st.param.f32 [param1+4], %f849; st.param.f32 [param1+8], %f868; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f801, [retval0+0]; ld.param.f32 %f802, [retval0+4]; ld.param.f32 %f803, [retval0+8]; } // callseq 11 { // callseq 12, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1605; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f32; st.param.f32 [param1+4], %f849; st.param.f32 [param1+8], %f868; .param .align 4 .b8 retval0[12]; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE, ( param0, param1 ); ld.param.f32 %f804, [retval0+0]; ld.param.f32 %f805, [retval0+4]; ld.param.f32 %f806, [retval0+8]; } // callseq 12 mul.ftz.f32 %f807, %f836, %f804; mul.ftz.f32 %f808, %f836, %f805; mul.ftz.f32 %f809, %f836, %f806; fma.rn.ftz.f32 %f810, %f842, %f801, %f807; fma.rn.ftz.f32 %f811, %f842, %f802, %f808; fma.rn.ftz.f32 %f812, %f842, %f803, %f809; mul.ftz.f32 %f813, %f852, %f810; mul.ftz.f32 %f814, %f852, %f811; mul.ftz.f32 %f815, %f852, %f812; fma.rn.ftz.f32 %f816, %f876, %f798, %f813; fma.rn.ftz.f32 %f817, %f876, %f799, %f814; fma.rn.ftz.f32 %f818, %f876, %f800, %f815; mul.ftz.f32 %f819, %f871, %f818; fma.rn.ftz.f32 %f959, %f859, %f786, %f819; fma.rn.ftz.f32 %f958, %f871, %f817, %f788; fma.rn.ftz.f32 %f957, %f871, %f816, %f787; $L__BB4_554: st.param.f32 [func_retval0+0], %f957; st.param.f32 [func_retval0+4], %f958; st.param.f32 [func_retval0+8], %f959; ret; $L__BB4_50: ld.f32 %f914, [%rd47]; ld.f32 %f913, [%rd47+4]; ld.f32 %f912, [%rd47+8]; bra.uni $L__BB4_66; $L__BB4_85: ld.f32 %f917, [%rd76]; ld.f32 %f916, [%rd76+4]; ld.f32 %f915, [%rd76+8]; bra.uni $L__BB4_101; $L__BB4_120: ld.f32 %f920, [%rd105]; ld.f32 %f919, [%rd105+4]; ld.f32 %f918, [%rd105+8]; bra.uni $L__BB4_136; $L__BB4_155: ld.f32 %f923, [%rd134]; ld.f32 %f922, [%rd134+4]; ld.f32 %f921, [%rd134+8]; mov.u32 %r2377, %r261; bra.uni $L__BB4_171; $L__BB4_190: ld.f32 %f926, [%rd163]; ld.f32 %f925, [%rd163+4]; ld.f32 %f924, [%rd163+8]; mov.u32 %r2377, %r318; bra.uni $L__BB4_206; $L__BB4_225: ld.f32 %f929, [%rd192]; ld.f32 %f928, [%rd192+4]; ld.f32 %f927, [%rd192+8]; mov.u32 %r2377, %r372; bra.uni $L__BB4_241; $L__BB4_260: ld.f32 %f932, [%rd221]; ld.f32 %f931, [%rd221+4]; ld.f32 %f930, [%rd221+8]; mov.u32 %r2377, %r423; bra.uni $L__BB4_276; $L__BB4_295: ld.f32 %f935, [%rd250]; ld.f32 %f934, [%rd250+4]; ld.f32 %f933, [%rd250+8]; mov.u32 %r2377, %r473; bra.uni $L__BB4_311; $L__BB4_330: ld.f32 %f938, [%rd279]; ld.f32 %f937, [%rd279+4]; ld.f32 %f936, [%rd279+8]; mov.u32 %r2377, %r518; bra.uni $L__BB4_346; $L__BB4_365: ld.f32 %f941, [%rd308]; ld.f32 %f940, [%rd308+4]; ld.f32 %f939, [%rd308+8]; mov.u32 %r2377, %r560; bra.uni $L__BB4_381; $L__BB4_400: ld.f32 %f944, [%rd337]; ld.f32 %f943, [%rd337+4]; ld.f32 %f942, [%rd337+8]; mov.u32 %r2377, %r599; bra.uni $L__BB4_416; $L__BB4_435: ld.f32 %f947, [%rd366]; ld.f32 %f946, [%rd366+4]; ld.f32 %f945, [%rd366+8]; mov.u32 %r2377, %r637; bra.uni $L__BB4_451; $L__BB4_470: ld.f32 %f950, [%rd395]; ld.f32 %f949, [%rd395+4]; ld.f32 %f948, [%rd395+8]; mov.u32 %r2377, %r670; bra.uni $L__BB4_486; $L__BB4_505: ld.f32 %f953, [%rd424]; ld.f32 %f952, [%rd424+4]; ld.f32 %f951, [%rd424+8]; mov.u32 %r2377, %r700; bra.uni $L__BB4_521; $L__BB4_56: cvt.u64.u32 %rd1608, %r2286; shl.b64 %rd598, %rd1608, 5; add.s64 %rd1620, %rd17, %rd598; $L__BB4_57: setp.eq.s64 %p35, %rd1620, 0; @%p35 bra $L__BB4_65; add.s64 %rd55, %rd1620, 8; ld.u64 %rd56, [%rd1620+8]; setp.eq.s64 %p36, %rd56, 0; @%p36 bra $L__BB4_64; add.s64 %rd1636, %rd1, %rd56; and.b32 %r1000, %r8, -4096; mov.u64 %rd599, 1; st.local.v2.u32 [%rd4+16], {%r12, %r1000}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1001, %r7, 3; and.b32 %r1002, %r1001, 31744; and.b32 %r1003, %r8, 3968; shr.u32 %r1004, %r1003, 2; or.b32 %r1005, %r1004, %r1002; bfe.u32 %r1006, %r74, 7, 5; or.b32 %r1007, %r1006, %r1005; shr.u32 %r1008, %r1005, 6; mul.wide.u32 %rd600, %r1008, 8; add.s64 %rd601, %rd1636, %rd600; and.b32 %r1009, %r1007, 63; shl.b64 %rd602, %rd599, %r1009; ld.u64 %rd603, [%rd601+4128]; and.b64 %rd604, %rd603, %rd602; setp.eq.s64 %p37, %rd604, 0; mov.b64 %rd605, {%r12, %r1000}; shr.u64 %rd606, %rd605, 32; cvt.u32.u64 %r2382, %rd606; mul.wide.u32 %rd607, %r1007, 16; add.s64 %rd608, %rd1636, %rd607; add.s64 %rd58, %rd608, 8256; @%p37 bra $L__BB4_63; and.b32 %r2272, %r7, -128; ld.u64 %rd609, [%rd58]; add.s64 %rd1628, %rd1636, %rd609; st.local.u32 [%rd4+4], %r2272; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1010, %r7, 5; and.b32 %r1011, %r1010, 3840; shl.b32 %r1012, %r8, 1; and.b32 %r1013, %r1012, 240; or.b32 %r1014, %r1013, %r1011; bfe.u32 %r1015, %r74, 3, 4; or.b32 %r1016, %r1015, %r1014; shr.u32 %r1017, %r1014, 6; mul.wide.u32 %rd610, %r1017, 8; add.s64 %rd611, %rd1628, %rd610; and.b32 %r1018, %r1016, 63; shl.b64 %rd613, %rd599, %r1018; ld.u64 %rd614, [%rd611+544]; and.b64 %rd615, %rd614, %rd613; setp.eq.s64 %p38, %rd615, 0; mul.wide.u32 %rd616, %r1016, 16; add.s64 %rd617, %rd1628, %rd616; add.s64 %rd60, %rd617, 1088; and.b32 %r2377, %r7, -128; @%p38 bra $L__BB4_62; and.b32 %r2334, %r7, -8; ld.u64 %rd618, [%rd60]; add.s64 %rd1621, %rd1628, %rd618; and.b32 %r1019, %r8, -8; st.local.v2.u32 [%rd2], {%r2334, %r1019}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1020, %r7, 6; and.b32 %r1021, %r1020, 448; shl.b32 %r1022, %r8, 3; and.b32 %r1023, %r1022, 56; or.b32 %r1024, %r1023, %r1021; and.b32 %r1025, %r74, 7; or.b32 %r1026, %r1024, %r1025; mul.wide.u32 %rd619, %r1026, 12; add.s64 %rd620, %rd1621, %rd619; ld.f32 %f914, [%rd620+128]; ld.f32 %f913, [%rd620+132]; ld.f32 %f912, [%rd620+136]; mov.b64 %rd621, {%r2334, %r1019}; shr.u64 %rd622, %rd621, 32; cvt.u32.u64 %r2289, %rd622; mov.u32 %r2337, %r12; bra.uni $L__BB4_66; $L__BB4_65: ld.f32 %f914, [%rd5+4]; ld.f32 %f913, [%rd5+8]; ld.f32 %f912, [%rd5+12]; bra.uni $L__BB4_66; $L__BB4_91: cvt.u64.u32 %rd1611, %r2335; shl.b64 %rd668, %rd1611, 5; add.s64 %rd1626, %rd17, %rd668; $L__BB4_92: setp.eq.s64 %p57, %rd1626, 0; @%p57 bra $L__BB4_100; add.s64 %rd84, %rd1626, 8; ld.u64 %rd85, [%rd1626+8]; setp.eq.s64 %p58, %rd85, 0; @%p58 bra $L__BB4_99; add.s64 %rd1636, %rd1, %rd85; and.b32 %r1086, %r139, -4096; mov.u64 %rd669, 1; st.local.v2.u32 [%rd4+16], {%r12, %r1086}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1087, %r7, 3; and.b32 %r1088, %r1087, 31744; and.b32 %r1089, %r139, 3968; shr.u32 %r1090, %r1089, 2; or.b32 %r1091, %r1090, %r1088; bfe.u32 %r1092, %r74, 7, 5; or.b32 %r1093, %r1091, %r1092; shr.u32 %r1094, %r1091, 6; mul.wide.u32 %rd670, %r1094, 8; add.s64 %rd671, %rd1636, %rd670; and.b32 %r1095, %r1093, 63; shl.b64 %rd672, %rd669, %r1095; ld.u64 %rd673, [%rd671+4128]; and.b64 %rd674, %rd673, %rd672; setp.eq.s64 %p59, %rd674, 0; mov.b64 %rd675, {%r12, %r1086}; shr.u64 %rd676, %rd675, 32; cvt.u32.u64 %r2382, %rd676; mul.wide.u32 %rd677, %r1093, 16; add.s64 %rd678, %rd1636, %rd677; add.s64 %rd87, %rd678, 8256; @%p59 bra $L__BB4_98; and.b32 %r2264, %r7, -128; ld.u64 %rd679, [%rd87]; add.s64 %rd1628, %rd1636, %rd679; st.local.u32 [%rd4+4], %r2264; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1096, %r7, 5; and.b32 %r1097, %r1096, 3840; shl.b32 %r1098, %r139, 1; and.b32 %r1099, %r1098, 240; or.b32 %r1100, %r1099, %r1097; bfe.u32 %r1101, %r74, 3, 4; or.b32 %r1102, %r1100, %r1101; shr.u32 %r1103, %r1100, 6; mul.wide.u32 %rd680, %r1103, 8; add.s64 %rd681, %rd1628, %rd680; and.b32 %r1104, %r1102, 63; shl.b64 %rd683, %rd669, %r1104; ld.u64 %rd684, [%rd681+544]; and.b64 %rd685, %rd684, %rd683; setp.eq.s64 %p60, %rd685, 0; mul.wide.u32 %rd686, %r1102, 16; add.s64 %rd687, %rd1628, %rd686; add.s64 %rd89, %rd687, 1088; and.b32 %r2377, %r7, -128; @%p60 bra $L__BB4_97; and.b32 %r2334, %r7, -8; ld.u64 %rd688, [%rd89]; add.s64 %rd1621, %rd1628, %rd688; and.b32 %r1105, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1105}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1106, %r7, 6; and.b32 %r1107, %r1106, 448; shl.b32 %r1108, %r139, 3; and.b32 %r1109, %r1108, 56; or.b32 %r1110, %r1109, %r1107; and.b32 %r1111, %r74, 7; or.b32 %r1112, %r1110, %r1111; mul.wide.u32 %rd689, %r1112, 12; add.s64 %rd690, %rd1621, %rd689; ld.f32 %f917, [%rd690+128]; ld.f32 %f916, [%rd690+132]; ld.f32 %f915, [%rd690+136]; mov.b64 %rd691, {%r2334, %r1105}; shr.u64 %rd692, %rd691, 32; cvt.u32.u64 %r2289, %rd692; mov.u32 %r2337, %r12; bra.uni $L__BB4_101; $L__BB4_100: ld.f32 %f917, [%rd5+4]; ld.f32 %f916, [%rd5+8]; ld.f32 %f915, [%rd5+12]; bra.uni $L__BB4_101; $L__BB4_126: cvt.u64.u32 %rd1612, %r2381; shl.b64 %rd736, %rd1612, 5; add.s64 %rd1632, %rd17, %rd736; $L__BB4_127: setp.eq.s64 %p79, %rd1632, 0; @%p79 bra $L__BB4_135; add.s64 %rd113, %rd1632, 8; ld.u64 %rd114, [%rd1632+8]; setp.eq.s64 %p80, %rd114, 0; @%p80 bra $L__BB4_134; add.s64 %rd1636, %rd1, %rd114; and.b32 %r1168, %r139, -4096; mov.u64 %rd737, 1; st.local.v2.u32 [%rd4+16], {%r12, %r1168}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1169, %r7, 3; and.b32 %r1170, %r1169, 31744; and.b32 %r1171, %r139, 3968; shr.u32 %r1172, %r1171, 2; or.b32 %r1173, %r1172, %r1170; bfe.u32 %r1174, %r9, 7, 5; or.b32 %r1175, %r1173, %r1174; shr.u32 %r1176, %r1173, 6; mul.wide.u32 %rd738, %r1176, 8; add.s64 %rd739, %rd1636, %rd738; and.b32 %r1177, %r1175, 63; shl.b64 %rd740, %rd737, %r1177; ld.u64 %rd741, [%rd739+4128]; and.b64 %rd742, %rd741, %rd740; setp.eq.s64 %p81, %rd742, 0; mov.b64 %rd743, {%r12, %r1168}; shr.u64 %rd744, %rd743, 32; cvt.u32.u64 %r2382, %rd744; mul.wide.u32 %rd745, %r1175, 16; add.s64 %rd746, %rd1636, %rd745; add.s64 %rd116, %rd746, 8256; @%p81 bra $L__BB4_133; and.b32 %r2236, %r7, -128; ld.u64 %rd747, [%rd116]; add.s64 %rd1628, %rd1636, %rd747; st.local.u32 [%rd4+4], %r2236; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1178, %r7, 5; and.b32 %r1179, %r1178, 3840; shl.b32 %r1180, %r139, 1; and.b32 %r1181, %r1180, 240; or.b32 %r1182, %r1181, %r1179; bfe.u32 %r1183, %r9, 3, 4; or.b32 %r1184, %r1182, %r1183; shr.u32 %r1185, %r1182, 6; mul.wide.u32 %rd748, %r1185, 8; add.s64 %rd749, %rd1628, %rd748; and.b32 %r1186, %r1184, 63; shl.b64 %rd751, %rd737, %r1186; ld.u64 %rd752, [%rd749+544]; and.b64 %rd753, %rd752, %rd751; setp.eq.s64 %p82, %rd753, 0; mul.wide.u32 %rd754, %r1184, 16; add.s64 %rd755, %rd1628, %rd754; add.s64 %rd118, %rd755, 1088; and.b32 %r2377, %r7, -128; @%p82 bra $L__BB4_132; and.b32 %r2334, %r7, -8; ld.u64 %rd756, [%rd118]; add.s64 %rd1621, %rd1628, %rd756; and.b32 %r1187, %r139, -8; st.local.v2.u32 [%rd2], {%r2334, %r1187}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1188, %r7, 7; shl.b32 %r1189, %r139, 3; and.b32 %r1190, %r1189, 56; and.b32 %r1191, %r9, 7; bfi.b32 %r1192, %r1188, %r1191, 6, 3; or.b32 %r1193, %r1192, %r1190; mul.wide.u32 %rd757, %r1193, 12; add.s64 %rd758, %rd1621, %rd757; ld.f32 %f920, [%rd758+128]; ld.f32 %f919, [%rd758+132]; ld.f32 %f918, [%rd758+136]; mov.b64 %rd759, {%r2334, %r1187}; shr.u64 %rd760, %rd759, 32; cvt.u32.u64 %r2289, %rd760; mov.u32 %r2337, %r12; bra.uni $L__BB4_136; $L__BB4_135: ld.f32 %f920, [%rd5+4]; ld.f32 %f919, [%rd5+8]; ld.f32 %f918, [%rd5+12]; bra.uni $L__BB4_136; $L__BB4_161: cvt.u64.u32 %rd1613, %r2424; shl.b64 %rd807, %rd1613, 5; add.s64 %rd1638, %rd17, %rd807; $L__BB4_162: setp.eq.s64 %p101, %rd1638, 0; @%p101 bra $L__BB4_170; add.s64 %rd142, %rd1638, 8; ld.u64 %rd143, [%rd1638+8]; setp.eq.s64 %p102, %rd143, 0; @%p102 bra $L__BB4_169; and.b32 %r2241, %r259, -4096; add.s64 %rd1636, %rd1, %rd143; and.b32 %r1250, %r139, -4096; mov.u64 %rd808, 1; st.local.v2.u32 [%rd4+16], {%r2241, %r1250}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1251, %r259, 3; and.b32 %r1252, %r1251, 31744; and.b32 %r1253, %r139, 3968; shr.u32 %r1254, %r1253, 2; or.b32 %r1255, %r1254, %r1252; bfe.u32 %r1256, %r9, 7, 5; or.b32 %r1257, %r1255, %r1256; shr.u32 %r1258, %r1255, 6; mul.wide.u32 %rd809, %r1258, 8; add.s64 %rd810, %rd1636, %rd809; and.b32 %r1259, %r1257, 63; shl.b64 %rd811, %rd808, %r1259; ld.u64 %rd812, [%rd810+4128]; and.b64 %rd813, %rd812, %rd811; setp.eq.s64 %p103, %rd813, 0; mov.b64 %rd814, {%r2241, %r1250}; shr.u64 %rd815, %rd814, 32; cvt.u32.u64 %r2382, %rd815; mul.wide.u32 %rd816, %r1257, 16; add.s64 %rd817, %rd1636, %rd816; add.s64 %rd145, %rd817, 8256; @%p103 bra $L__BB4_168; and.b32 %r2242, %r259, -128; ld.u64 %rd818, [%rd145]; add.s64 %rd1628, %rd1636, %rd818; st.local.u32 [%rd4+4], %r2242; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1260, %r259, 5; and.b32 %r1261, %r1260, 3840; shl.b32 %r1262, %r139, 1; and.b32 %r1263, %r1262, 240; or.b32 %r1264, %r1263, %r1261; bfe.u32 %r1265, %r9, 3, 4; or.b32 %r1266, %r1264, %r1265; shr.u32 %r1267, %r1264, 6; mul.wide.u32 %rd819, %r1267, 8; add.s64 %rd820, %rd1628, %rd819; and.b32 %r1268, %r1266, 63; shl.b64 %rd822, %rd808, %r1268; ld.u64 %rd823, [%rd820+544]; and.b64 %rd824, %rd823, %rd822; setp.eq.s64 %p104, %rd824, 0; mul.wide.u32 %rd825, %r1266, 16; add.s64 %rd826, %rd1628, %rd825; add.s64 %rd147, %rd826, 1088; and.b32 %r2337, %r259, -4096; and.b32 %r2377, %r259, -128; @%p104 bra $L__BB4_167; ld.u64 %rd827, [%rd147]; add.s64 %rd1621, %rd1628, %rd827; and.b32 %r1269, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1269}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1270, %r259, 7; shl.b32 %r1271, %r139, 3; and.b32 %r1272, %r1271, 56; and.b32 %r1273, %r9, 7; bfi.b32 %r1274, %r1270, %r1273, 6, 3; or.b32 %r1275, %r1274, %r1272; mul.wide.u32 %rd828, %r1275, 12; add.s64 %rd829, %rd1621, %rd828; ld.f32 %f923, [%rd829+128]; ld.f32 %f922, [%rd829+132]; ld.f32 %f921, [%rd829+136]; mov.b64 %rd830, {%r260, %r1269}; shr.u64 %rd831, %rd830, 32; cvt.u32.u64 %r2289, %rd831; mov.u32 %r2334, %r260; bra.uni $L__BB4_171; $L__BB4_170: ld.f32 %f923, [%rd5+4]; ld.f32 %f922, [%rd5+8]; ld.f32 %f921, [%rd5+12]; bra.uni $L__BB4_171; $L__BB4_196: cvt.u64.u32 %rd1614, %r2464; shl.b64 %rd878, %rd1614, 5; add.s64 %rd1644, %rd17, %rd878; $L__BB4_197: setp.eq.s64 %p123, %rd1644, 0; @%p123 bra $L__BB4_205; add.s64 %rd171, %rd1644, 8; ld.u64 %rd172, [%rd1644+8]; setp.eq.s64 %p124, %rd172, 0; @%p124 bra $L__BB4_204; and.b32 %r2249, %r259, -4096; add.s64 %rd1636, %rd1, %rd172; and.b32 %r1336, %r139, -4096; mov.u64 %rd879, 1; st.local.v2.u32 [%rd4+16], {%r2249, %r1336}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1337, %r259, 3; and.b32 %r1338, %r1337, 31744; and.b32 %r1339, %r139, 3968; shr.u32 %r1340, %r1339, 2; or.b32 %r1341, %r1340, %r1338; bfe.u32 %r1342, %r74, 7, 5; or.b32 %r1343, %r1341, %r1342; shr.u32 %r1344, %r1341, 6; mul.wide.u32 %rd880, %r1344, 8; add.s64 %rd881, %rd1636, %rd880; and.b32 %r1345, %r1343, 63; shl.b64 %rd882, %rd879, %r1345; ld.u64 %rd883, [%rd881+4128]; and.b64 %rd884, %rd883, %rd882; setp.eq.s64 %p125, %rd884, 0; mov.b64 %rd885, {%r2249, %r1336}; shr.u64 %rd886, %rd885, 32; cvt.u32.u64 %r2382, %rd886; mul.wide.u32 %rd887, %r1343, 16; add.s64 %rd888, %rd1636, %rd887; add.s64 %rd174, %rd888, 8256; @%p125 bra $L__BB4_203; and.b32 %r2250, %r259, -128; ld.u64 %rd889, [%rd174]; add.s64 %rd1628, %rd1636, %rd889; st.local.u32 [%rd4+4], %r2250; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1346, %r259, 5; and.b32 %r1347, %r1346, 3840; shl.b32 %r1348, %r139, 1; and.b32 %r1349, %r1348, 240; or.b32 %r1350, %r1349, %r1347; bfe.u32 %r1351, %r74, 3, 4; or.b32 %r1352, %r1350, %r1351; shr.u32 %r1353, %r1350, 6; mul.wide.u32 %rd890, %r1353, 8; add.s64 %rd891, %rd1628, %rd890; and.b32 %r1354, %r1352, 63; shl.b64 %rd893, %rd879, %r1354; ld.u64 %rd894, [%rd891+544]; and.b64 %rd895, %rd894, %rd893; setp.eq.s64 %p126, %rd895, 0; mul.wide.u32 %rd896, %r1352, 16; add.s64 %rd897, %rd1628, %rd896; add.s64 %rd176, %rd897, 1088; and.b32 %r2337, %r259, -4096; and.b32 %r2377, %r259, -128; @%p126 bra $L__BB4_202; ld.u64 %rd898, [%rd176]; add.s64 %rd1621, %rd1628, %rd898; and.b32 %r1355, %r139, -8; st.local.v2.u32 [%rd2], {%r260, %r1355}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1356, %r259, 6; and.b32 %r1357, %r1356, 448; shl.b32 %r1358, %r139, 3; and.b32 %r1359, %r1358, 56; or.b32 %r1360, %r1359, %r1357; and.b32 %r1361, %r74, 7; or.b32 %r1362, %r1360, %r1361; mul.wide.u32 %rd899, %r1362, 12; add.s64 %rd900, %rd1621, %rd899; ld.f32 %f926, [%rd900+128]; ld.f32 %f925, [%rd900+132]; ld.f32 %f924, [%rd900+136]; mov.b64 %rd901, {%r260, %r1355}; shr.u64 %rd902, %rd901, 32; cvt.u32.u64 %r2289, %rd902; mov.u32 %r2334, %r260; bra.uni $L__BB4_206; $L__BB4_205: ld.f32 %f926, [%rd5+4]; ld.f32 %f925, [%rd5+8]; ld.f32 %f924, [%rd5+12]; bra.uni $L__BB4_206; $L__BB4_231: shl.b64 %rd949, %rd197, 5; add.s64 %rd1650, %rd17, %rd949; $L__BB4_232: setp.eq.s64 %p145, %rd1650, 0; @%p145 bra $L__BB4_240; add.s64 %rd200, %rd1650, 8; ld.u64 %rd201, [%rd1650+8]; setp.eq.s64 %p146, %rd201, 0; @%p146 bra $L__BB4_239; add.s64 %rd1636, %rd1, %rd201; and.b32 %r1422, %r8, -4096; mov.u64 %rd950, 1; st.local.v2.u32 [%rd4+16], {%r375, %r1422}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1423, %r259, 3; and.b32 %r1424, %r1423, 31744; and.b32 %r1425, %r8, 3968; shr.u32 %r1426, %r1425, 2; or.b32 %r1427, %r1424, %r1426; bfe.u32 %r1428, %r74, 7, 5; or.b32 %r1429, %r1428, %r1427; shr.u32 %r1430, %r1427, 6; mul.wide.u32 %rd951, %r1430, 8; add.s64 %rd952, %rd1636, %rd951; and.b32 %r1431, %r1429, 63; shl.b64 %rd953, %rd950, %r1431; ld.u64 %rd954, [%rd952+4128]; and.b64 %rd955, %rd954, %rd953; setp.eq.s64 %p147, %rd955, 0; mov.b64 %rd956, {%r375, %r1422}; shr.u64 %rd957, %rd956, 32; cvt.u32.u64 %r2382, %rd957; mul.wide.u32 %rd958, %r1429, 16; add.s64 %rd959, %rd1636, %rd958; add.s64 %rd203, %rd959, 8256; @%p147 bra $L__BB4_238; ld.u64 %rd960, [%rd203]; add.s64 %rd1628, %rd1636, %rd960; st.local.u32 [%rd4+4], %r372; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1432, %r259, 5; and.b32 %r1433, %r1432, 3840; shl.b32 %r1434, %r8, 1; and.b32 %r1435, %r1434, 240; or.b32 %r1436, %r1433, %r1435; bfe.u32 %r1437, %r74, 3, 4; or.b32 %r1438, %r1437, %r1436; shr.u32 %r1439, %r1436, 6; mul.wide.u32 %rd961, %r1439, 8; add.s64 %rd962, %rd1628, %rd961; and.b32 %r1440, %r1438, 63; shl.b64 %rd964, %rd950, %r1440; ld.u64 %rd965, [%rd962+544]; and.b64 %rd966, %rd965, %rd964; setp.eq.s64 %p148, %rd966, 0; mul.wide.u32 %rd967, %r1438, 16; add.s64 %rd968, %rd1628, %rd967; add.s64 %rd205, %rd968, 1088; @%p148 bra $L__BB4_237; ld.u64 %rd969, [%rd205]; add.s64 %rd1621, %rd1628, %rd969; and.b32 %r1441, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1441}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1442, %r259, 6; and.b32 %r1443, %r1442, 448; shl.b32 %r1444, %r8, 3; and.b32 %r1445, %r1444, 56; or.b32 %r1446, %r1443, %r1445; and.b32 %r1447, %r74, 7; or.b32 %r1448, %r1446, %r1447; mul.wide.u32 %rd970, %r1448, 12; add.s64 %rd971, %rd1621, %rd970; ld.f32 %f929, [%rd971+128]; ld.f32 %f928, [%rd971+132]; ld.f32 %f927, [%rd971+136]; mov.b64 %rd972, {%r260, %r1441}; shr.u64 %rd973, %rd972, 32; cvt.u32.u64 %r2289, %rd973; mov.u32 %r2337, %r375; mov.u32 %r2377, %r372; mov.u32 %r2334, %r260; bra.uni $L__BB4_241; $L__BB4_240: ld.f32 %f929, [%rd5+4]; ld.f32 %f928, [%rd5+8]; ld.f32 %f927, [%rd5+12]; bra.uni $L__BB4_241; $L__BB4_266: shl.b64 %rd1019, %rd226, 5; add.s64 %rd1656, %rd17, %rd1019; $L__BB4_267: setp.eq.s64 %p167, %rd1656, 0; @%p167 bra $L__BB4_275; add.s64 %rd229, %rd1656, 8; ld.u64 %rd230, [%rd1656+8]; setp.eq.s64 %p168, %rd230, 0; @%p168 bra $L__BB4_274; add.s64 %rd1636, %rd1, %rd230; and.b32 %r1507, %r8, -4096; mov.u64 %rd1020, 1; st.local.v2.u32 [%rd4+16], {%r426, %r1507}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1508, %r259, 3; and.b32 %r1509, %r1508, 31744; and.b32 %r1510, %r8, 3968; shr.u32 %r1511, %r1510, 2; or.b32 %r1512, %r1509, %r1511; bfe.u32 %r1513, %r9, 7, 5; or.b32 %r1514, %r1512, %r1513; shr.u32 %r1515, %r1512, 6; mul.wide.u32 %rd1021, %r1515, 8; add.s64 %rd1022, %rd1636, %rd1021; and.b32 %r1516, %r1514, 63; shl.b64 %rd1023, %rd1020, %r1516; ld.u64 %rd1024, [%rd1022+4128]; and.b64 %rd1025, %rd1024, %rd1023; setp.eq.s64 %p169, %rd1025, 0; mov.b64 %rd1026, {%r426, %r1507}; shr.u64 %rd1027, %rd1026, 32; cvt.u32.u64 %r2382, %rd1027; mul.wide.u32 %rd1028, %r1514, 16; add.s64 %rd1029, %rd1636, %rd1028; add.s64 %rd232, %rd1029, 8256; @%p169 bra $L__BB4_273; ld.u64 %rd1030, [%rd232]; add.s64 %rd1628, %rd1636, %rd1030; st.local.u32 [%rd4+4], %r423; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1517, %r259, 5; and.b32 %r1518, %r1517, 3840; shl.b32 %r1519, %r8, 1; and.b32 %r1520, %r1519, 240; or.b32 %r1521, %r1518, %r1520; bfe.u32 %r1522, %r9, 3, 4; or.b32 %r1523, %r1521, %r1522; shr.u32 %r1524, %r1521, 6; mul.wide.u32 %rd1031, %r1524, 8; add.s64 %rd1032, %rd1628, %rd1031; and.b32 %r1525, %r1523, 63; shl.b64 %rd1034, %rd1020, %r1525; ld.u64 %rd1035, [%rd1032+544]; and.b64 %rd1036, %rd1035, %rd1034; setp.eq.s64 %p170, %rd1036, 0; mul.wide.u32 %rd1037, %r1523, 16; add.s64 %rd1038, %rd1628, %rd1037; add.s64 %rd234, %rd1038, 1088; @%p170 bra $L__BB4_272; ld.u64 %rd1039, [%rd234]; add.s64 %rd1621, %rd1628, %rd1039; and.b32 %r1526, %r8, -8; st.local.v2.u32 [%rd2], {%r260, %r1526}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1527, %r259, 6; and.b32 %r1528, %r1527, 448; shl.b32 %r1529, %r8, 3; and.b32 %r1530, %r1529, 56; or.b32 %r1531, %r1528, %r1530; and.b32 %r1532, %r9, 7; or.b32 %r1533, %r1531, %r1532; mul.wide.u32 %rd1040, %r1533, 12; add.s64 %rd1041, %rd1621, %rd1040; ld.f32 %f932, [%rd1041+128]; ld.f32 %f931, [%rd1041+132]; ld.f32 %f930, [%rd1041+136]; mov.b64 %rd1042, {%r260, %r1526}; shr.u64 %rd1043, %rd1042, 32; cvt.u32.u64 %r2289, %rd1043; mov.u32 %r2337, %r426; mov.u32 %r2377, %r423; mov.u32 %r2334, %r260; bra.uni $L__BB4_276; $L__BB4_275: ld.f32 %f932, [%rd5+4]; ld.f32 %f931, [%rd5+8]; ld.f32 %f930, [%rd5+12]; bra.uni $L__BB4_276; $L__BB4_301: shl.b64 %rd1089, %rd255, 5; add.s64 %rd1662, %rd17, %rd1089; $L__BB4_302: setp.eq.s64 %p189, %rd1662, 0; @%p189 bra $L__BB4_310; add.s64 %rd258, %rd1662, 8; ld.u64 %rd259, [%rd1662+8]; setp.eq.s64 %p190, %rd259, 0; @%p190 bra $L__BB4_309; add.s64 %rd1636, %rd1, %rd259; and.b32 %r1592, %r8, -4096; mov.u64 %rd1090, 1; st.local.v2.u32 [%rd4+16], {%r476, %r1592}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1593, %r471, 3; and.b32 %r1594, %r1593, 31744; and.b32 %r1595, %r8, 3968; shr.u32 %r1596, %r1595, 2; or.b32 %r1597, %r1596, %r1594; bfe.u32 %r1598, %r9, 7, 5; or.b32 %r1599, %r1597, %r1598; shr.u32 %r1600, %r1597, 6; mul.wide.u32 %rd1091, %r1600, 8; add.s64 %rd1092, %rd1636, %rd1091; and.b32 %r1601, %r1599, 63; shl.b64 %rd1093, %rd1090, %r1601; ld.u64 %rd1094, [%rd1092+4128]; and.b64 %rd1095, %rd1094, %rd1093; setp.eq.s64 %p191, %rd1095, 0; mov.b64 %rd1096, {%r476, %r1592}; shr.u64 %rd1097, %rd1096, 32; cvt.u32.u64 %r2382, %rd1097; mul.wide.u32 %rd1098, %r1599, 16; add.s64 %rd1099, %rd1636, %rd1098; add.s64 %rd261, %rd1099, 8256; @%p191 bra $L__BB4_308; ld.u64 %rd1100, [%rd261]; add.s64 %rd1628, %rd1636, %rd1100; st.local.u32 [%rd4+4], %r473; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1602, %r471, 5; and.b32 %r1603, %r1602, 3840; shl.b32 %r1604, %r8, 1; and.b32 %r1605, %r1604, 240; or.b32 %r1606, %r1605, %r1603; bfe.u32 %r1607, %r9, 3, 4; or.b32 %r1608, %r1606, %r1607; shr.u32 %r1609, %r1606, 6; mul.wide.u32 %rd1101, %r1609, 8; add.s64 %rd1102, %rd1628, %rd1101; and.b32 %r1610, %r1608, 63; shl.b64 %rd1104, %rd1090, %r1610; ld.u64 %rd1105, [%rd1102+544]; and.b64 %rd1106, %rd1105, %rd1104; setp.eq.s64 %p192, %rd1106, 0; mul.wide.u32 %rd1107, %r1608, 16; add.s64 %rd1108, %rd1628, %rd1107; add.s64 %rd263, %rd1108, 1088; @%p192 bra $L__BB4_307; ld.u64 %rd1109, [%rd263]; add.s64 %rd1621, %rd1628, %rd1109; and.b32 %r1611, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1611}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1612, %r471, 6; and.b32 %r1613, %r1612, 448; shl.b32 %r1614, %r8, 3; and.b32 %r1615, %r1614, 56; or.b32 %r1616, %r1615, %r1613; and.b32 %r1617, %r9, 7; or.b32 %r1618, %r1616, %r1617; mul.wide.u32 %rd1110, %r1618, 12; add.s64 %rd1111, %rd1621, %rd1110; ld.f32 %f935, [%rd1111+128]; ld.f32 %f934, [%rd1111+132]; ld.f32 %f933, [%rd1111+136]; mov.b64 %rd1112, {%r472, %r1611}; shr.u64 %rd1113, %rd1112, 32; cvt.u32.u64 %r2289, %rd1113; mov.u32 %r2337, %r476; mov.u32 %r2377, %r473; mov.u32 %r2334, %r472; bra.uni $L__BB4_311; $L__BB4_310: ld.f32 %f935, [%rd5+4]; ld.f32 %f934, [%rd5+8]; ld.f32 %f933, [%rd5+12]; bra.uni $L__BB4_311; $L__BB4_336: shl.b64 %rd1160, %rd284, 5; add.s64 %rd1668, %rd17, %rd1160; $L__BB4_337: setp.eq.s64 %p211, %rd1668, 0; @%p211 bra $L__BB4_345; add.s64 %rd287, %rd1668, 8; ld.u64 %rd288, [%rd1668+8]; setp.eq.s64 %p212, %rd288, 0; @%p212 bra $L__BB4_344; add.s64 %rd1636, %rd1, %rd288; and.b32 %r1678, %r8, -4096; mov.u64 %rd1161, 1; st.local.v2.u32 [%rd4+16], {%r521, %r1678}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1679, %r471, 3; and.b32 %r1680, %r1679, 31744; and.b32 %r1681, %r8, 3968; shr.u32 %r1682, %r1681, 2; or.b32 %r1683, %r1682, %r1680; bfe.u32 %r1684, %r74, 7, 5; or.b32 %r1685, %r1684, %r1683; shr.u32 %r1686, %r1683, 6; mul.wide.u32 %rd1162, %r1686, 8; add.s64 %rd1163, %rd1636, %rd1162; and.b32 %r1687, %r1685, 63; shl.b64 %rd1164, %rd1161, %r1687; ld.u64 %rd1165, [%rd1163+4128]; and.b64 %rd1166, %rd1165, %rd1164; setp.eq.s64 %p213, %rd1166, 0; mov.b64 %rd1167, {%r521, %r1678}; shr.u64 %rd1168, %rd1167, 32; cvt.u32.u64 %r2382, %rd1168; mul.wide.u32 %rd1169, %r1685, 16; add.s64 %rd1170, %rd1636, %rd1169; add.s64 %rd290, %rd1170, 8256; @%p213 bra $L__BB4_343; ld.u64 %rd1171, [%rd290]; add.s64 %rd1628, %rd1636, %rd1171; st.local.u32 [%rd4+4], %r518; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1688, %r471, 5; and.b32 %r1689, %r1688, 3840; shl.b32 %r1690, %r8, 1; and.b32 %r1691, %r1690, 240; or.b32 %r1692, %r1691, %r1689; bfe.u32 %r1693, %r74, 3, 4; or.b32 %r1694, %r1693, %r1692; shr.u32 %r1695, %r1692, 6; mul.wide.u32 %rd1172, %r1695, 8; add.s64 %rd1173, %rd1628, %rd1172; and.b32 %r1696, %r1694, 63; shl.b64 %rd1175, %rd1161, %r1696; ld.u64 %rd1176, [%rd1173+544]; and.b64 %rd1177, %rd1176, %rd1175; setp.eq.s64 %p214, %rd1177, 0; mul.wide.u32 %rd1178, %r1694, 16; add.s64 %rd1179, %rd1628, %rd1178; add.s64 %rd292, %rd1179, 1088; @%p214 bra $L__BB4_342; ld.u64 %rd1180, [%rd292]; add.s64 %rd1621, %rd1628, %rd1180; and.b32 %r1697, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1697}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1698, %r471, 6; and.b32 %r1699, %r1698, 448; shl.b32 %r1700, %r8, 3; and.b32 %r1701, %r1700, 56; or.b32 %r1702, %r1701, %r1699; and.b32 %r1703, %r74, 7; or.b32 %r1704, %r1702, %r1703; mul.wide.u32 %rd1181, %r1704, 12; add.s64 %rd1182, %rd1621, %rd1181; ld.f32 %f938, [%rd1182+128]; ld.f32 %f937, [%rd1182+132]; ld.f32 %f936, [%rd1182+136]; mov.b64 %rd1183, {%r472, %r1697}; shr.u64 %rd1184, %rd1183, 32; cvt.u32.u64 %r2289, %rd1184; mov.u32 %r2337, %r521; mov.u32 %r2377, %r518; mov.u32 %r2334, %r472; bra.uni $L__BB4_346; $L__BB4_345: ld.f32 %f938, [%rd5+4]; ld.f32 %f937, [%rd5+8]; ld.f32 %f936, [%rd5+12]; bra.uni $L__BB4_346; $L__BB4_371: shl.b64 %rd1231, %rd313, 5; add.s64 %rd1674, %rd17, %rd1231; $L__BB4_372: setp.eq.s64 %p233, %rd1674, 0; @%p233 bra $L__BB4_380; add.s64 %rd316, %rd1674, 8; ld.u64 %rd317, [%rd1674+8]; setp.eq.s64 %p234, %rd317, 0; @%p234 bra $L__BB4_379; add.s64 %rd1636, %rd1, %rd317; and.b32 %r1765, %r139, -4096; mov.u64 %rd1232, 1; st.local.v2.u32 [%rd4+16], {%r563, %r1765}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1766, %r471, 3; and.b32 %r1767, %r1766, 31744; and.b32 %r1768, %r139, 3968; shr.u32 %r1769, %r1768, 2; or.b32 %r1770, %r1769, %r1767; bfe.u32 %r1771, %r74, 7, 5; or.b32 %r1772, %r1770, %r1771; shr.u32 %r1773, %r1770, 6; mul.wide.u32 %rd1233, %r1773, 8; add.s64 %rd1234, %rd1636, %rd1233; and.b32 %r1774, %r1772, 63; shl.b64 %rd1235, %rd1232, %r1774; ld.u64 %rd1236, [%rd1234+4128]; and.b64 %rd1237, %rd1236, %rd1235; setp.eq.s64 %p235, %rd1237, 0; mov.b64 %rd1238, {%r563, %r1765}; shr.u64 %rd1239, %rd1238, 32; cvt.u32.u64 %r2382, %rd1239; mul.wide.u32 %rd1240, %r1772, 16; add.s64 %rd1241, %rd1636, %rd1240; add.s64 %rd319, %rd1241, 8256; @%p235 bra $L__BB4_378; ld.u64 %rd1242, [%rd319]; add.s64 %rd1628, %rd1636, %rd1242; st.local.u32 [%rd4+4], %r560; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1775, %r471, 5; and.b32 %r1776, %r1775, 3840; shl.b32 %r1777, %r139, 1; and.b32 %r1778, %r1777, 240; or.b32 %r1779, %r1778, %r1776; bfe.u32 %r1780, %r74, 3, 4; or.b32 %r1781, %r1779, %r1780; shr.u32 %r1782, %r1779, 6; mul.wide.u32 %rd1243, %r1782, 8; add.s64 %rd1244, %rd1628, %rd1243; and.b32 %r1783, %r1781, 63; shl.b64 %rd1246, %rd1232, %r1783; ld.u64 %rd1247, [%rd1244+544]; and.b64 %rd1248, %rd1247, %rd1246; setp.eq.s64 %p236, %rd1248, 0; mul.wide.u32 %rd1249, %r1781, 16; add.s64 %rd1250, %rd1628, %rd1249; add.s64 %rd321, %rd1250, 1088; @%p236 bra $L__BB4_377; ld.u64 %rd1251, [%rd321]; add.s64 %rd1621, %rd1628, %rd1251; and.b32 %r1784, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1784}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r1785, %r471, 6; and.b32 %r1786, %r1785, 448; shl.b32 %r1787, %r139, 3; and.b32 %r1788, %r1787, 56; or.b32 %r1789, %r1788, %r1786; and.b32 %r1790, %r74, 7; or.b32 %r1791, %r1789, %r1790; mul.wide.u32 %rd1252, %r1791, 12; add.s64 %rd1253, %rd1621, %rd1252; ld.f32 %f941, [%rd1253+128]; ld.f32 %f940, [%rd1253+132]; ld.f32 %f939, [%rd1253+136]; mov.b64 %rd1254, {%r472, %r1784}; shr.u64 %rd1255, %rd1254, 32; cvt.u32.u64 %r2289, %rd1255; mov.u32 %r2337, %r563; mov.u32 %r2377, %r560; mov.u32 %r2334, %r472; bra.uni $L__BB4_381; $L__BB4_380: ld.f32 %f941, [%rd5+4]; ld.f32 %f940, [%rd5+8]; ld.f32 %f939, [%rd5+12]; bra.uni $L__BB4_381; $L__BB4_406: shl.b64 %rd1302, %rd342, 5; add.s64 %rd1680, %rd17, %rd1302; $L__BB4_407: setp.eq.s64 %p255, %rd1680, 0; @%p255 bra $L__BB4_415; add.s64 %rd345, %rd1680, 8; ld.u64 %rd346, [%rd1680+8]; setp.eq.s64 %p256, %rd346, 0; @%p256 bra $L__BB4_414; add.s64 %rd1636, %rd1, %rd346; and.b32 %r1848, %r139, -4096; mov.u64 %rd1303, 1; st.local.v2.u32 [%rd4+16], {%r602, %r1848}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1849, %r471, 3; and.b32 %r1850, %r1849, 31744; and.b32 %r1851, %r139, 3968; shr.u32 %r1852, %r1851, 2; or.b32 %r1853, %r1852, %r1850; bfe.u32 %r1854, %r9, 7, 5; or.b32 %r1855, %r1853, %r1854; shr.u32 %r1856, %r1853, 6; mul.wide.u32 %rd1304, %r1856, 8; add.s64 %rd1305, %rd1636, %rd1304; and.b32 %r1857, %r1855, 63; shl.b64 %rd1306, %rd1303, %r1857; ld.u64 %rd1307, [%rd1305+4128]; and.b64 %rd1308, %rd1307, %rd1306; setp.eq.s64 %p257, %rd1308, 0; mov.b64 %rd1309, {%r602, %r1848}; shr.u64 %rd1310, %rd1309, 32; cvt.u32.u64 %r2382, %rd1310; mul.wide.u32 %rd1311, %r1855, 16; add.s64 %rd1312, %rd1636, %rd1311; add.s64 %rd348, %rd1312, 8256; @%p257 bra $L__BB4_413; ld.u64 %rd1313, [%rd348]; add.s64 %rd1628, %rd1636, %rd1313; st.local.u32 [%rd4+4], %r599; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1858, %r471, 5; and.b32 %r1859, %r1858, 3840; shl.b32 %r1860, %r139, 1; and.b32 %r1861, %r1860, 240; or.b32 %r1862, %r1861, %r1859; bfe.u32 %r1863, %r9, 3, 4; or.b32 %r1864, %r1862, %r1863; shr.u32 %r1865, %r1862, 6; mul.wide.u32 %rd1314, %r1865, 8; add.s64 %rd1315, %rd1628, %rd1314; and.b32 %r1866, %r1864, 63; shl.b64 %rd1317, %rd1303, %r1866; ld.u64 %rd1318, [%rd1315+544]; and.b64 %rd1319, %rd1318, %rd1317; setp.eq.s64 %p258, %rd1319, 0; mul.wide.u32 %rd1320, %r1864, 16; add.s64 %rd1321, %rd1628, %rd1320; add.s64 %rd350, %rd1321, 1088; @%p258 bra $L__BB4_412; ld.u64 %rd1322, [%rd350]; add.s64 %rd1621, %rd1628, %rd1322; and.b32 %r1867, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1867}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1868, %r471, 7; shl.b32 %r1869, %r139, 3; and.b32 %r1870, %r1869, 56; and.b32 %r1871, %r9, 7; bfi.b32 %r1872, %r1868, %r1871, 6, 3; or.b32 %r1873, %r1872, %r1870; mul.wide.u32 %rd1323, %r1873, 12; add.s64 %rd1324, %rd1621, %rd1323; ld.f32 %f944, [%rd1324+128]; ld.f32 %f943, [%rd1324+132]; ld.f32 %f942, [%rd1324+136]; mov.b64 %rd1325, {%r472, %r1867}; shr.u64 %rd1326, %rd1325, 32; cvt.u32.u64 %r2289, %rd1326; mov.u32 %r2337, %r602; mov.u32 %r2377, %r599; mov.u32 %r2334, %r472; bra.uni $L__BB4_416; $L__BB4_415: ld.f32 %f944, [%rd5+4]; ld.f32 %f943, [%rd5+8]; ld.f32 %f942, [%rd5+12]; bra.uni $L__BB4_416; $L__BB4_441: shl.b64 %rd1373, %rd371, 5; add.s64 %rd1686, %rd17, %rd1373; $L__BB4_442: setp.eq.s64 %p277, %rd1686, 0; @%p277 bra $L__BB4_450; add.s64 %rd374, %rd1686, 8; ld.u64 %rd375, [%rd1686+8]; setp.eq.s64 %p278, %rd375, 0; @%p278 bra $L__BB4_449; add.s64 %rd1636, %rd1, %rd375; and.b32 %r1930, %r139, -4096; mov.u64 %rd1374, 1; st.local.v2.u32 [%rd4+16], {%r640, %r1930}; and.b32 %r2415, %r9, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r1931, %r635, 3; and.b32 %r1932, %r1931, 31744; and.b32 %r1933, %r139, 3968; shr.u32 %r1934, %r1933, 2; or.b32 %r1935, %r1934, %r1932; bfe.u32 %r1936, %r9, 7, 5; or.b32 %r1937, %r1935, %r1936; shr.u32 %r1938, %r1935, 6; mul.wide.u32 %rd1375, %r1938, 8; add.s64 %rd1376, %rd1636, %rd1375; and.b32 %r1939, %r1937, 63; shl.b64 %rd1377, %rd1374, %r1939; ld.u64 %rd1378, [%rd1376+4128]; and.b64 %rd1379, %rd1378, %rd1377; setp.eq.s64 %p279, %rd1379, 0; mov.b64 %rd1380, {%r640, %r1930}; shr.u64 %rd1381, %rd1380, 32; cvt.u32.u64 %r2382, %rd1381; mul.wide.u32 %rd1382, %r1937, 16; add.s64 %rd1383, %rd1636, %rd1382; add.s64 %rd377, %rd1383, 8256; @%p279 bra $L__BB4_448; ld.u64 %rd1384, [%rd377]; add.s64 %rd1628, %rd1636, %rd1384; st.local.u32 [%rd4+4], %r637; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r9, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r1940, %r635, 5; and.b32 %r1941, %r1940, 3840; shl.b32 %r1942, %r139, 1; and.b32 %r1943, %r1942, 240; or.b32 %r1944, %r1943, %r1941; bfe.u32 %r1945, %r9, 3, 4; or.b32 %r1946, %r1944, %r1945; shr.u32 %r1947, %r1944, 6; mul.wide.u32 %rd1385, %r1947, 8; add.s64 %rd1386, %rd1628, %rd1385; and.b32 %r1948, %r1946, 63; shl.b64 %rd1388, %rd1374, %r1948; ld.u64 %rd1389, [%rd1386+544]; and.b64 %rd1390, %rd1389, %rd1388; setp.eq.s64 %p280, %rd1390, 0; mul.wide.u32 %rd1391, %r1946, 16; add.s64 %rd1392, %rd1628, %rd1391; add.s64 %rd379, %rd1392, 1088; @%p280 bra $L__BB4_447; ld.u64 %rd1393, [%rd379]; add.s64 %rd1621, %rd1628, %rd1393; and.b32 %r1949, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1949}; and.b32 %r2332, %r9, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; and.b32 %r1950, %r635, 7; shl.b32 %r1951, %r139, 3; and.b32 %r1952, %r1951, 56; and.b32 %r1953, %r9, 7; bfi.b32 %r1954, %r1950, %r1953, 6, 3; or.b32 %r1955, %r1954, %r1952; mul.wide.u32 %rd1394, %r1955, 12; add.s64 %rd1395, %rd1621, %rd1394; ld.f32 %f947, [%rd1395+128]; ld.f32 %f946, [%rd1395+132]; ld.f32 %f945, [%rd1395+136]; mov.b64 %rd1396, {%r636, %r1949}; shr.u64 %rd1397, %rd1396, 32; cvt.u32.u64 %r2289, %rd1397; mov.u32 %r2337, %r640; mov.u32 %r2377, %r637; mov.u32 %r2334, %r636; bra.uni $L__BB4_451; $L__BB4_450: ld.f32 %f947, [%rd5+4]; ld.f32 %f946, [%rd5+8]; ld.f32 %f945, [%rd5+12]; bra.uni $L__BB4_451; $L__BB4_476: shl.b64 %rd1444, %rd400, 5; add.s64 %rd1692, %rd17, %rd1444; $L__BB4_477: setp.eq.s64 %p299, %rd1692, 0; @%p299 bra $L__BB4_485; add.s64 %rd403, %rd1692, 8; ld.u64 %rd404, [%rd1692+8]; setp.eq.s64 %p300, %rd404, 0; @%p300 bra $L__BB4_484; add.s64 %rd1636, %rd1, %rd404; and.b32 %r2016, %r139, -4096; mov.u64 %rd1445, 1; st.local.v2.u32 [%rd4+16], {%r673, %r2016}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r2017, %r635, 3; and.b32 %r2018, %r2017, 31744; and.b32 %r2019, %r139, 3968; shr.u32 %r2020, %r2019, 2; or.b32 %r2021, %r2020, %r2018; bfe.u32 %r2022, %r74, 7, 5; or.b32 %r2023, %r2021, %r2022; shr.u32 %r2024, %r2021, 6; mul.wide.u32 %rd1446, %r2024, 8; add.s64 %rd1447, %rd1636, %rd1446; and.b32 %r2025, %r2023, 63; shl.b64 %rd1448, %rd1445, %r2025; ld.u64 %rd1449, [%rd1447+4128]; and.b64 %rd1450, %rd1449, %rd1448; setp.eq.s64 %p301, %rd1450, 0; mov.b64 %rd1451, {%r673, %r2016}; shr.u64 %rd1452, %rd1451, 32; cvt.u32.u64 %r2382, %rd1452; mul.wide.u32 %rd1453, %r2023, 16; add.s64 %rd1454, %rd1636, %rd1453; add.s64 %rd406, %rd1454, 8256; @%p301 bra $L__BB4_483; ld.u64 %rd1455, [%rd406]; add.s64 %rd1628, %rd1636, %rd1455; st.local.u32 [%rd4+4], %r670; and.b32 %r2376, %r139, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r2026, %r635, 5; and.b32 %r2027, %r2026, 3840; shl.b32 %r2028, %r139, 1; and.b32 %r2029, %r2028, 240; or.b32 %r2030, %r2029, %r2027; bfe.u32 %r2031, %r74, 3, 4; or.b32 %r2032, %r2030, %r2031; shr.u32 %r2033, %r2030, 6; mul.wide.u32 %rd1456, %r2033, 8; add.s64 %rd1457, %rd1628, %rd1456; and.b32 %r2034, %r2032, 63; shl.b64 %rd1459, %rd1445, %r2034; ld.u64 %rd1460, [%rd1457+544]; and.b64 %rd1461, %rd1460, %rd1459; setp.eq.s64 %p302, %rd1461, 0; mul.wide.u32 %rd1462, %r2032, 16; add.s64 %rd1463, %rd1628, %rd1462; add.s64 %rd408, %rd1463, 1088; @%p302 bra $L__BB4_482; ld.u64 %rd1464, [%rd408]; add.s64 %rd1621, %rd1628, %rd1464; and.b32 %r2035, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2035}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r2036, %r635, 6; and.b32 %r2037, %r2036, 448; shl.b32 %r2038, %r139, 3; and.b32 %r2039, %r2038, 56; or.b32 %r2040, %r2039, %r2037; and.b32 %r2041, %r74, 7; or.b32 %r2042, %r2040, %r2041; mul.wide.u32 %rd1465, %r2042, 12; add.s64 %rd1466, %rd1621, %rd1465; ld.f32 %f950, [%rd1466+128]; ld.f32 %f949, [%rd1466+132]; ld.f32 %f948, [%rd1466+136]; mov.b64 %rd1467, {%r636, %r2035}; shr.u64 %rd1468, %rd1467, 32; cvt.u32.u64 %r2289, %rd1468; mov.u32 %r2337, %r673; mov.u32 %r2377, %r670; mov.u32 %r2334, %r636; bra.uni $L__BB4_486; $L__BB4_485: ld.f32 %f950, [%rd5+4]; ld.f32 %f949, [%rd5+8]; ld.f32 %f948, [%rd5+12]; bra.uni $L__BB4_486; $L__BB4_511: shl.b64 %rd1515, %rd429, 5; add.s64 %rd1698, %rd17, %rd1515; $L__BB4_512: setp.eq.s64 %p321, %rd1698, 0; @%p321 bra $L__BB4_520; add.s64 %rd432, %rd1698, 8; ld.u64 %rd433, [%rd1698+8]; setp.eq.s64 %p322, %rd433, 0; @%p322 bra $L__BB4_519; add.s64 %rd1636, %rd1, %rd433; and.b32 %r2102, %r8, -4096; mov.u64 %rd1516, 1; st.local.v2.u32 [%rd4+16], {%r703, %r2102}; and.b32 %r2415, %r74, -4096; st.local.u32 [%rd4+24], %r2415; st.local.u64 [%rd4+56], %rd1636; shl.b32 %r2103, %r635, 3; and.b32 %r2104, %r2103, 31744; and.b32 %r2105, %r8, 3968; shr.u32 %r2106, %r2105, 2; or.b32 %r2107, %r2104, %r2106; bfe.u32 %r2108, %r74, 7, 5; or.b32 %r2109, %r2108, %r2107; shr.u32 %r2110, %r2107, 6; mul.wide.u32 %rd1517, %r2110, 8; add.s64 %rd1518, %rd1636, %rd1517; and.b32 %r2111, %r2109, 63; shl.b64 %rd1519, %rd1516, %r2111; ld.u64 %rd1520, [%rd1518+4128]; and.b64 %rd1521, %rd1520, %rd1519; setp.eq.s64 %p323, %rd1521, 0; mov.b64 %rd1522, {%r703, %r2102}; shr.u64 %rd1523, %rd1522, 32; cvt.u32.u64 %r2382, %rd1523; mul.wide.u32 %rd1524, %r2109, 16; add.s64 %rd1525, %rd1636, %rd1524; add.s64 %rd435, %rd1525, 8256; @%p323 bra $L__BB4_518; ld.u64 %rd1526, [%rd435]; add.s64 %rd1628, %rd1636, %rd1526; st.local.u32 [%rd4+4], %r700; and.b32 %r2376, %r8, -128; st.local.u32 [%rd4+8], %r2376; and.b32 %r2375, %r74, -128; st.local.u32 [%rd4+12], %r2375; st.local.u64 [%rd4+48], %rd1628; shl.b32 %r2112, %r635, 5; and.b32 %r2113, %r2112, 3840; shl.b32 %r2114, %r8, 1; and.b32 %r2115, %r2114, 240; or.b32 %r2116, %r2113, %r2115; bfe.u32 %r2117, %r74, 3, 4; or.b32 %r2118, %r2117, %r2116; shr.u32 %r2119, %r2116, 6; mul.wide.u32 %rd1527, %r2119, 8; add.s64 %rd1528, %rd1628, %rd1527; and.b32 %r2120, %r2118, 63; shl.b64 %rd1530, %rd1516, %r2120; ld.u64 %rd1531, [%rd1528+544]; and.b64 %rd1532, %rd1531, %rd1530; setp.eq.s64 %p324, %rd1532, 0; mul.wide.u32 %rd1533, %r2118, 16; add.s64 %rd1534, %rd1628, %rd1533; add.s64 %rd437, %rd1534, 1088; @%p324 bra $L__BB4_517; ld.u64 %rd1535, [%rd437]; add.s64 %rd1621, %rd1628, %rd1535; and.b32 %r2121, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2121}; and.b32 %r2332, %r74, -8; st.local.u32 [%rd4], %r2332; st.local.u64 [%rd4+40], %rd1621; shl.b32 %r2122, %r635, 6; and.b32 %r2123, %r2122, 448; shl.b32 %r2124, %r8, 3; and.b32 %r2125, %r2124, 56; or.b32 %r2126, %r2123, %r2125; and.b32 %r2127, %r74, 7; or.b32 %r2128, %r2126, %r2127; mul.wide.u32 %rd1536, %r2128, 12; add.s64 %rd1537, %rd1621, %rd1536; ld.f32 %f953, [%rd1537+128]; ld.f32 %f952, [%rd1537+132]; ld.f32 %f951, [%rd1537+136]; mov.b64 %rd1538, {%r636, %r2121}; shr.u64 %rd1539, %rd1538, 32; cvt.u32.u64 %r2289, %rd1539; mov.u32 %r2337, %r703; mov.u32 %r2377, %r700; mov.u32 %r2334, %r636; bra.uni $L__BB4_521; $L__BB4_520: ld.f32 %f953, [%rd5+4]; ld.f32 %f952, [%rd5+8]; ld.f32 %f951, [%rd5+12]; bra.uni $L__BB4_521; $L__BB4_64: ld.f32 %f914, [%rd55+12]; ld.f32 %f913, [%rd55+16]; ld.f32 %f912, [%rd55+20]; bra.uni $L__BB4_66; $L__BB4_99: ld.f32 %f917, [%rd84+12]; ld.f32 %f916, [%rd84+16]; ld.f32 %f915, [%rd84+20]; bra.uni $L__BB4_101; $L__BB4_134: ld.f32 %f920, [%rd113+12]; ld.f32 %f919, [%rd113+16]; ld.f32 %f918, [%rd113+20]; bra.uni $L__BB4_136; $L__BB4_169: ld.f32 %f923, [%rd142+12]; ld.f32 %f922, [%rd142+16]; ld.f32 %f921, [%rd142+20]; bra.uni $L__BB4_171; $L__BB4_204: ld.f32 %f926, [%rd171+12]; ld.f32 %f925, [%rd171+16]; ld.f32 %f924, [%rd171+20]; bra.uni $L__BB4_206; $L__BB4_239: ld.f32 %f929, [%rd200+12]; ld.f32 %f928, [%rd200+16]; ld.f32 %f927, [%rd200+20]; bra.uni $L__BB4_241; $L__BB4_274: ld.f32 %f932, [%rd229+12]; ld.f32 %f931, [%rd229+16]; ld.f32 %f930, [%rd229+20]; bra.uni $L__BB4_276; $L__BB4_309: ld.f32 %f935, [%rd258+12]; ld.f32 %f934, [%rd258+16]; ld.f32 %f933, [%rd258+20]; bra.uni $L__BB4_311; $L__BB4_344: ld.f32 %f938, [%rd287+12]; ld.f32 %f937, [%rd287+16]; ld.f32 %f936, [%rd287+20]; bra.uni $L__BB4_346; $L__BB4_379: ld.f32 %f941, [%rd316+12]; ld.f32 %f940, [%rd316+16]; ld.f32 %f939, [%rd316+20]; bra.uni $L__BB4_381; $L__BB4_414: ld.f32 %f944, [%rd345+12]; ld.f32 %f943, [%rd345+16]; ld.f32 %f942, [%rd345+20]; bra.uni $L__BB4_416; $L__BB4_449: ld.f32 %f947, [%rd374+12]; ld.f32 %f946, [%rd374+16]; ld.f32 %f945, [%rd374+20]; bra.uni $L__BB4_451; $L__BB4_484: ld.f32 %f950, [%rd403+12]; ld.f32 %f949, [%rd403+16]; ld.f32 %f948, [%rd403+20]; bra.uni $L__BB4_486; $L__BB4_519: ld.f32 %f953, [%rd432+12]; ld.f32 %f952, [%rd432+16]; ld.f32 %f951, [%rd432+20]; bra.uni $L__BB4_521; $L__BB4_63: ld.f32 %f914, [%rd58]; ld.f32 %f913, [%rd58+4]; ld.f32 %f912, [%rd58+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_66; $L__BB4_98: ld.f32 %f917, [%rd87]; ld.f32 %f916, [%rd87+4]; ld.f32 %f915, [%rd87+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_101; $L__BB4_133: ld.f32 %f920, [%rd116]; ld.f32 %f919, [%rd116+4]; ld.f32 %f918, [%rd116+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_136; $L__BB4_168: and.b32 %r2337, %r259, -4096; ld.f32 %f923, [%rd145]; ld.f32 %f922, [%rd145+4]; ld.f32 %f921, [%rd145+8]; bra.uni $L__BB4_171; $L__BB4_203: and.b32 %r2337, %r259, -4096; ld.f32 %f926, [%rd174]; ld.f32 %f925, [%rd174+4]; ld.f32 %f924, [%rd174+8]; bra.uni $L__BB4_206; $L__BB4_238: ld.f32 %f929, [%rd203]; ld.f32 %f928, [%rd203+4]; ld.f32 %f927, [%rd203+8]; mov.u32 %r2337, %r375; bra.uni $L__BB4_241; $L__BB4_273: ld.f32 %f932, [%rd232]; ld.f32 %f931, [%rd232+4]; ld.f32 %f930, [%rd232+8]; mov.u32 %r2337, %r426; bra.uni $L__BB4_276; $L__BB4_308: ld.f32 %f935, [%rd261]; ld.f32 %f934, [%rd261+4]; ld.f32 %f933, [%rd261+8]; mov.u32 %r2337, %r476; bra.uni $L__BB4_311; $L__BB4_343: ld.f32 %f938, [%rd290]; ld.f32 %f937, [%rd290+4]; ld.f32 %f936, [%rd290+8]; mov.u32 %r2337, %r521; bra.uni $L__BB4_346; $L__BB4_378: ld.f32 %f941, [%rd319]; ld.f32 %f940, [%rd319+4]; ld.f32 %f939, [%rd319+8]; mov.u32 %r2337, %r563; bra.uni $L__BB4_381; $L__BB4_413: ld.f32 %f944, [%rd348]; ld.f32 %f943, [%rd348+4]; ld.f32 %f942, [%rd348+8]; mov.u32 %r2337, %r602; bra.uni $L__BB4_416; $L__BB4_448: ld.f32 %f947, [%rd377]; ld.f32 %f946, [%rd377+4]; ld.f32 %f945, [%rd377+8]; mov.u32 %r2337, %r640; bra.uni $L__BB4_451; $L__BB4_483: ld.f32 %f950, [%rd406]; ld.f32 %f949, [%rd406+4]; ld.f32 %f948, [%rd406+8]; mov.u32 %r2337, %r673; bra.uni $L__BB4_486; $L__BB4_518: ld.f32 %f953, [%rd435]; ld.f32 %f952, [%rd435+4]; ld.f32 %f951, [%rd435+8]; mov.u32 %r2337, %r703; bra.uni $L__BB4_521; $L__BB4_62: ld.f32 %f914, [%rd60]; ld.f32 %f913, [%rd60+4]; ld.f32 %f912, [%rd60+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_66; $L__BB4_97: ld.f32 %f917, [%rd89]; ld.f32 %f916, [%rd89+4]; ld.f32 %f915, [%rd89+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_101; $L__BB4_132: ld.f32 %f920, [%rd118]; ld.f32 %f919, [%rd118+4]; ld.f32 %f918, [%rd118+8]; mov.u32 %r2337, %r12; bra.uni $L__BB4_136; $L__BB4_167: ld.f32 %f923, [%rd147]; ld.f32 %f922, [%rd147+4]; ld.f32 %f921, [%rd147+8]; bra.uni $L__BB4_171; $L__BB4_202: ld.f32 %f926, [%rd176]; ld.f32 %f925, [%rd176+4]; ld.f32 %f924, [%rd176+8]; bra.uni $L__BB4_206; $L__BB4_237: ld.f32 %f929, [%rd205]; ld.f32 %f928, [%rd205+4]; ld.f32 %f927, [%rd205+8]; mov.u32 %r2337, %r375; mov.u32 %r2377, %r372; bra.uni $L__BB4_241; $L__BB4_272: ld.f32 %f932, [%rd234]; ld.f32 %f931, [%rd234+4]; ld.f32 %f930, [%rd234+8]; mov.u32 %r2337, %r426; mov.u32 %r2377, %r423; bra.uni $L__BB4_276; $L__BB4_307: ld.f32 %f935, [%rd263]; ld.f32 %f934, [%rd263+4]; ld.f32 %f933, [%rd263+8]; mov.u32 %r2337, %r476; mov.u32 %r2377, %r473; bra.uni $L__BB4_311; $L__BB4_342: ld.f32 %f938, [%rd292]; ld.f32 %f937, [%rd292+4]; ld.f32 %f936, [%rd292+8]; mov.u32 %r2337, %r521; mov.u32 %r2377, %r518; bra.uni $L__BB4_346; $L__BB4_377: ld.f32 %f941, [%rd321]; ld.f32 %f940, [%rd321+4]; ld.f32 %f939, [%rd321+8]; mov.u32 %r2337, %r563; mov.u32 %r2377, %r560; bra.uni $L__BB4_381; $L__BB4_412: ld.f32 %f944, [%rd350]; ld.f32 %f943, [%rd350+4]; ld.f32 %f942, [%rd350+8]; mov.u32 %r2337, %r602; mov.u32 %r2377, %r599; bra.uni $L__BB4_416; $L__BB4_447: ld.f32 %f947, [%rd379]; ld.f32 %f946, [%rd379+4]; ld.f32 %f945, [%rd379+8]; mov.u32 %r2337, %r640; mov.u32 %r2377, %r637; bra.uni $L__BB4_451; $L__BB4_482: ld.f32 %f950, [%rd408]; ld.f32 %f949, [%rd408+4]; ld.f32 %f948, [%rd408+8]; mov.u32 %r2337, %r673; mov.u32 %r2377, %r670; bra.uni $L__BB4_486; $L__BB4_517: ld.f32 %f953, [%rd437]; ld.f32 %f952, [%rd437+4]; ld.f32 %f951, [%rd437+8]; mov.u32 %r2337, %r703; mov.u32 %r2377, %r700; bra.uni $L__BB4_521; } .func (.param .b32 func_retval0) _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj( .param .b64 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4 ) { .local .align 8 .b8 __local_depot5[88]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<520>; .reg .b16 %rs<519>; .reg .f32 %f<954>; .reg .b32 %r<5208>; .reg .b64 %rd<2395>; mov.u64 %SPL, __local_depot5; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd608, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0]; ld.param.f32 %f322, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1]; ld.param.f32 %f323, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2]; ld.param.f32 %f324, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; ld.param.u32 %r784, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4]; cvta.to.global.u64 %rd609, %rd608; ld.global.u64 %rd610, [%rd609]; mov.u64 %rd2298, 0; ld.u64 %rd612, [%rd610+696]; add.s64 %rd613, %rd610, %rd612; add.s64 %rd1, %rd613, 672; add.u64 %rd614, %SP, 0; add.u64 %rd2, %SPL, 0; mov.u32 %r4782, 2147483647; st.local.v2.u32 [%rd2], {%r4782, %r4782}; st.local.v2.u32 [%rd2+8], {%r4782, %r4782}; st.local.v2.u32 [%rd2+16], {%r4782, %r4782}; st.local.v2.u32 [%rd2+24], {%r4782, %r4782}; st.local.u32 [%rd2+32], %r4782; st.local.u64 [%rd2+40], %rd1; st.local.u64 [%rd2+48], %rd2298; st.local.u64 [%rd2+56], %rd2298; st.local.u64 [%rd2+64], %rd2298; add.s64 %rd5, %rd613, 696; setp.eq.s32 %p1, %r784, 0; @%p1 bra $L__BB5_16; setp.ne.s32 %p2, %r784, 1; @%p2 bra $L__BB5_17; cvt.rni.ftz.f32.f32 %f325, %f322; cvt.rzi.ftz.s32.f32 %r1, %f325; cvt.rni.ftz.f32.f32 %f326, %f323; cvt.rzi.ftz.s32.f32 %r2, %f326; cvt.rni.ftz.f32.f32 %f327, %f324; cvt.rzi.ftz.s32.f32 %r3, %f327; ld.u32 %r4, [%rd5]; mov.u64 %rd2297, 0; setp.eq.s32 %p3, %r4, 0; @%p3 bra $L__BB5_7; add.s64 %rd6, %rd1, 64; shr.u32 %r787, %r1, 12; cvt.u64.u32 %rd616, %r787; shl.b64 %rd617, %rd616, 42; shr.u32 %r788, %r2, 12; mul.wide.u32 %rd618, %r788, 2097152; or.b64 %rd619, %rd618, %rd617; shr.u32 %r789, %r3, 12; cvt.u64.u32 %rd620, %r789; or.b64 %rd7, %rd619, %rd620; mov.u32 %r4768, 0; $L__BB5_4: cvt.u64.u32 %rd8, %r4768; mul.wide.u32 %rd621, %r4768, 32; add.s64 %rd622, %rd6, %rd621; ld.u64 %rd623, [%rd622]; setp.eq.s64 %p4, %rd623, %rd7; add.s32 %r4768, %r4768, 1; @%p4 bra $L__BB5_6; setp.lt.u32 %p5, %r4768, %r4; @%p5 bra $L__BB5_4; bra.uni $L__BB5_7; $L__BB5_16: add.u64 %rd647, %SP, 72; add.u64 %rd648, %SPL, 72; st.local.u64 [%rd648], %rd614; add.ftz.f32 %f333, %f322, 0fBF000000; add.ftz.f32 %f334, %f323, 0fBF000000; add.ftz.f32 %f335, %f324, 0fBF000000; { // callseq 13, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd647; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f333; st.param.f32 [param1+4], %f334; st.param.f32 [param1+8], %f335; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f953, [retval0+0]; } // callseq 13 bra.uni $L__BB5_834; $L__BB5_17: add.u64 %rd652, %SPL, 80; st.local.u64 [%rd652], %rd614; add.ftz.f32 %f336, %f322, 0fBF000000; cvt.rmi.ftz.f32.f32 %f337, %f336; add.ftz.f32 %f338, %f323, 0fBF000000; cvt.rmi.ftz.f32.f32 %f339, %f338; add.ftz.f32 %f340, %f324, 0fBF000000; cvt.rmi.ftz.f32.f32 %f341, %f340; sub.ftz.f32 %f342, %f336, %f337; sub.ftz.f32 %f343, %f338, %f339; sub.ftz.f32 %f344, %f340, %f341; mov.f32 %f345, 0f40400000; sub.ftz.f32 %f346, %f345, %f342; fma.rn.ftz.f32 %f347, %f342, %f346, 0fC0400000; fma.rn.ftz.f32 %f348, %f342, %f347, 0f3F800000; mul.ftz.f32 %f349, %f342, %f342; mul.ftz.f32 %f350, %f342, 0f40400000; add.ftz.f32 %f351, %f350, 0fC0C00000; fma.rn.ftz.f32 %f352, %f349, %f351, 0f40800000; mul.ftz.f32 %f353, %f352, 0f3E2AAAAB; fma.rn.ftz.f32 %f7, %f348, 0f3E2AAAAB, %f353; sub.ftz.f32 %f354, %f345, %f350; fma.rn.ftz.f32 %f355, %f342, %f354, 0f40400000; fma.rn.ftz.f32 %f356, %f342, %f355, 0f3F800000; mul.ftz.f32 %f357, %f342, %f349; mul.ftz.f32 %f358, %f357, 0f3E2AAAAB; fma.rn.ftz.f32 %f8, %f356, 0f3E2AAAAB, %f358; sub.ftz.f32 %f359, %f345, %f343; fma.rn.ftz.f32 %f360, %f343, %f359, 0fC0400000; fma.rn.ftz.f32 %f361, %f343, %f360, 0f3F800000; mul.ftz.f32 %f362, %f343, %f343; mul.ftz.f32 %f363, %f343, 0f40400000; add.ftz.f32 %f364, %f363, 0fC0C00000; fma.rn.ftz.f32 %f365, %f362, %f364, 0f40800000; mul.ftz.f32 %f366, %f365, 0f3E2AAAAB; fma.rn.ftz.f32 %f9, %f361, 0f3E2AAAAB, %f366; sub.ftz.f32 %f367, %f345, %f363; fma.rn.ftz.f32 %f368, %f343, %f367, 0f40400000; fma.rn.ftz.f32 %f369, %f343, %f368, 0f3F800000; mul.ftz.f32 %f370, %f343, %f362; mul.ftz.f32 %f371, %f370, 0f3E2AAAAB; fma.rn.ftz.f32 %f10, %f369, 0f3E2AAAAB, %f371; sub.ftz.f32 %f372, %f345, %f344; fma.rn.ftz.f32 %f373, %f344, %f372, 0fC0400000; fma.rn.ftz.f32 %f374, %f344, %f373, 0f3F800000; mul.ftz.f32 %f375, %f344, %f344; mul.ftz.f32 %f376, %f344, 0f40400000; add.ftz.f32 %f377, %f376, 0fC0C00000; fma.rn.ftz.f32 %f378, %f375, %f377, 0f40800000; mul.ftz.f32 %f379, %f378, 0f3E2AAAAB; fma.rn.ftz.f32 %f11, %f374, 0f3E2AAAAB, %f379; sub.ftz.f32 %f380, %f345, %f376; fma.rn.ftz.f32 %f381, %f344, %f380, 0f40400000; fma.rn.ftz.f32 %f382, %f344, %f381, 0f3F800000; mul.ftz.f32 %f383, %f344, %f375; mul.ftz.f32 %f384, %f383, 0f3E2AAAAB; fma.rn.ftz.f32 %f12, %f382, 0f3E2AAAAB, %f384; div.approx.ftz.f32 %f385, %f353, %f7; add.ftz.f32 %f386, %f385, 0fBF800000; add.ftz.f32 %f13, %f337, %f386; div.approx.ftz.f32 %f387, %f358, %f8; add.ftz.f32 %f388, %f387, 0f3F800000; add.ftz.f32 %f14, %f337, %f388; div.approx.ftz.f32 %f389, %f366, %f9; add.ftz.f32 %f390, %f389, 0fBF800000; add.ftz.f32 %f15, %f339, %f390; div.approx.ftz.f32 %f391, %f371, %f10; add.ftz.f32 %f392, %f391, 0f3F800000; add.ftz.f32 %f16, %f339, %f392; div.approx.ftz.f32 %f393, %f379, %f11; add.ftz.f32 %f394, %f393, 0fBF800000; add.ftz.f32 %f17, %f341, %f394; div.approx.ftz.f32 %f395, %f384, %f12; add.ftz.f32 %f396, %f395, 0f3F800000; add.ftz.f32 %f18, %f341, %f396; cvt.rmi.ftz.f32.f32 %f19, %f13; cvt.rmi.ftz.f32.f32 %f397, %f15; cvt.rmi.ftz.f32.f32 %f398, %f17; sub.ftz.f32 %f20, %f15, %f397; sub.ftz.f32 %f21, %f17, %f398; cvt.rzi.ftz.s32.f32 %r7, %f19; cvt.rzi.ftz.s32.f32 %r8, %f397; cvt.rzi.ftz.s32.f32 %r9, %f398; and.b32 %r10, %r7, -8; and.b32 %r4870, %r7, -128; and.b32 %r12, %r7, -4096; add.s64 %rd17, %rd1, 64; shr.u32 %r841, %r9, 12; cvt.u64.u32 %rd18, %r841; shr.u32 %r842, %r8, 12; mul.wide.u32 %rd19, %r842, 2097152; shr.u32 %r843, %r7, 12; cvt.u64.u32 %rd654, %r843; bfi.b64 %rd21, %rd654, %rd18, 42, 32; ld.u32 %r13, [%rd5]; setp.eq.s32 %p10, %r13, 0; @%p10 bra $L__BB5_22; or.b64 %rd22, %rd21, %rd19; mov.u32 %r4769, 0; $L__BB5_19: cvt.u64.u32 %rd23, %r4769; mul.wide.u32 %rd655, %r4769, 32; add.s64 %rd656, %rd17, %rd655; ld.u64 %rd657, [%rd656]; setp.eq.s64 %p11, %rd657, %rd22; add.s32 %r4769, %r4769, 1; @%p11 bra $L__BB5_21; setp.lt.u32 %p12, %r4769, %r13; @%p12 bra $L__BB5_19; bra.uni $L__BB5_22; $L__BB5_6: shl.b64 %rd625, %rd8, 5; add.s64 %rd2297, %rd6, %rd625; $L__BB5_7: setp.eq.s64 %p6, %rd2297, 0; @%p6 bra $L__BB5_15; add.s64 %rd11, %rd2297, 8; ld.u64 %rd12, [%rd2297+8]; setp.eq.s64 %p7, %rd12, 0; @%p7 bra $L__BB5_14; add.s64 %rd13, %rd1, %rd12; and.b32 %r790, %r2, -4096; and.b32 %r791, %r1, -4096; mov.u64 %rd626, 1; st.local.v2.u32 [%rd2+24], {%r791, %r790}; and.b32 %r792, %r3, -4096; st.local.u32 [%rd2+32], %r792; st.local.u64 [%rd2+64], %rd13; shl.b32 %r793, %r1, 3; and.b32 %r794, %r793, 31744; and.b32 %r795, %r2, 3968; shr.u32 %r796, %r795, 2; or.b32 %r797, %r796, %r794; bfe.u32 %r798, %r3, 7, 5; or.b32 %r799, %r798, %r797; shr.u32 %r800, %r797, 6; mul.wide.u32 %rd627, %r800, 8; add.s64 %rd628, %rd13, %rd627; and.b32 %r801, %r799, 63; shl.b64 %rd629, %rd626, %r801; ld.u64 %rd630, [%rd628+4128]; and.b64 %rd631, %rd630, %rd629; setp.eq.s64 %p8, %rd631, 0; mul.wide.u32 %rd632, %r799, 8; add.s64 %rd633, %rd13, %rd632; add.s64 %rd14, %rd633, 8256; @%p8 bra $L__BB5_13; ld.u64 %rd634, [%rd14]; add.s64 %rd15, %rd13, %rd634; and.b32 %r802, %r1, -128; st.local.u32 [%rd2+12], %r802; and.b32 %r803, %r2, -128; st.local.u32 [%rd2+16], %r803; and.b32 %r804, %r3, -128; st.local.u32 [%rd2+20], %r804; st.local.u64 [%rd2+56], %rd15; shl.b32 %r805, %r1, 5; and.b32 %r806, %r805, 3840; shl.b32 %r807, %r2, 1; and.b32 %r808, %r807, 240; or.b32 %r809, %r808, %r806; bfe.u32 %r810, %r3, 3, 4; or.b32 %r811, %r810, %r809; shr.u32 %r812, %r809, 6; mul.wide.u32 %rd635, %r812, 8; add.s64 %rd636, %rd15, %rd635; and.b32 %r813, %r811, 63; shl.b64 %rd638, %rd626, %r813; ld.u64 %rd639, [%rd636+544]; and.b64 %rd640, %rd639, %rd638; setp.eq.s64 %p9, %rd640, 0; mul.wide.u32 %rd641, %r811, 8; add.s64 %rd642, %rd15, %rd641; add.s64 %rd16, %rd642, 1088; @%p9 bra $L__BB5_12; ld.u64 %rd643, [%rd16]; add.s64 %rd644, %rd15, %rd643; and.b32 %r814, %r2, -8; and.b32 %r815, %r1, -8; mov.u32 %r816, 1; st.local.v2.u32 [%rd2], {%r815, %r814}; and.b32 %r817, %r3, -8; st.local.u32 [%rd2+8], %r817; st.local.u64 [%rd2+48], %rd644; shl.b32 %r818, %r1, 6; and.b32 %r819, %r818, 448; shl.b32 %r820, %r2, 3; and.b32 %r821, %r820, 56; or.b32 %r822, %r821, %r819; and.b32 %r823, %r3, 7; or.b32 %r824, %r822, %r823; ld.u8 %rs133, [%rd644+15]; shr.u16 %rs134, %rs133, 5; mov.u32 %r825, 5; cvt.u32.u16 %r826, %rs134; sub.s32 %r827, %r825, %r826; shr.u32 %r828, %r824, %r827; mul.wide.u32 %rd645, %r828, 4; add.s64 %rd646, %rd644, %rd645; mov.u32 %r829, 32; shr.u32 %r830, %r829, %r826; add.s32 %r831, %r830, -1; and.b32 %r832, %r831, %r824; shl.b32 %r833, %r832, %r826; ld.u32 %r834, [%rd646+96]; shr.u32 %r835, %r834, %r833; shl.b32 %r836, %r816, %r826; mov.u32 %r837, -1; shl.b32 %r838, %r837, %r836; not.b32 %r839, %r838; and.b32 %r840, %r835, %r839; cvt.rn.f32.u32 %f328, %r840; ld.v2.f32 {%f329, %f330}, [%rd644+80]; fma.rn.ftz.f32 %f953, %f330, %f328, %f329; bra.uni $L__BB5_834; $L__BB5_15: ld.f32 %f953, [%rd5+4]; bra.uni $L__BB5_834; $L__BB5_21: shl.b64 %rd659, %rd23, 5; add.s64 %rd2298, %rd17, %rd659; $L__BB5_22: setp.eq.s64 %p13, %rd2298, 0; @%p13 bra $L__BB5_30; add.s64 %rd26, %rd2298, 8; ld.u64 %rd27, [%rd2298+8]; setp.eq.s64 %p14, %rd27, 0; @%p14 bra $L__BB5_29; add.s64 %rd2315, %rd1, %rd27; and.b32 %r845, %r8, -4096; mov.u64 %rd660, 1; st.local.v2.u32 [%rd2+24], {%r12, %r845}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r846, %r7, 3; and.b32 %r847, %r846, 31744; and.b32 %r848, %r8, 3968; shr.u32 %r849, %r848, 2; or.b32 %r850, %r849, %r847; bfe.u32 %r851, %r9, 7, 5; or.b32 %r852, %r850, %r851; shr.u32 %r853, %r850, 6; mul.wide.u32 %rd661, %r853, 8; add.s64 %rd662, %rd2315, %rd661; and.b32 %r854, %r852, 63; shl.b64 %rd663, %rd660, %r854; ld.u64 %rd664, [%rd662+4128]; and.b64 %rd665, %rd664, %rd663; setp.eq.s64 %p15, %rd665, 0; mul.wide.u32 %rd666, %r852, 8; add.s64 %rd667, %rd2315, %rd666; add.s64 %rd29, %rd667, 8256; mov.b64 %rd668, {%r12, %r845}; shr.u64 %rd669, %rd668, 32; cvt.u32.u64 %r4875, %rd669; @%p15 bra $L__BB5_28; ld.u64 %rd670, [%rd29]; add.s64 %rd2311, %rd2315, %rd670; st.local.u32 [%rd2+12], %r4870; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r855, %r7, 5; and.b32 %r856, %r855, 3840; shl.b32 %r857, %r8, 1; and.b32 %r858, %r857, 240; or.b32 %r859, %r858, %r856; bfe.u32 %r860, %r9, 3, 4; or.b32 %r861, %r859, %r860; shr.u32 %r862, %r859, 6; mul.wide.u32 %rd671, %r862, 8; add.s64 %rd672, %rd2311, %rd671; and.b32 %r863, %r861, 63; shl.b64 %rd674, %rd660, %r863; ld.u64 %rd675, [%rd672+544]; and.b64 %rd676, %rd675, %rd674; setp.eq.s64 %p16, %rd676, 0; mul.wide.u32 %rd677, %r861, 8; add.s64 %rd678, %rd2311, %rd677; add.s64 %rd31, %rd678, 1088; @%p16 bra $L__BB5_27; ld.u64 %rd679, [%rd31]; add.s64 %rd2307, %rd2311, %rd679; and.b32 %r864, %r8, -8; mov.u32 %r865, 1; st.local.v2.u32 [%rd2], {%r10, %r864}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r866, %r7, 6; and.b32 %r867, %r866, 448; shl.b32 %r868, %r8, 3; and.b32 %r869, %r868, 56; or.b32 %r870, %r869, %r867; and.b32 %r871, %r9, 7; or.b32 %r872, %r870, %r871; ld.u8 %rs135, [%rd2307+15]; shr.u16 %rs136, %rs135, 5; mov.u32 %r873, 5; cvt.u32.u16 %r874, %rs136; sub.s32 %r875, %r873, %r874; shr.u32 %r876, %r872, %r875; mul.wide.u32 %rd680, %r876, 4; add.s64 %rd681, %rd2307, %rd680; mov.u32 %r877, 32; shr.u32 %r878, %r877, %r874; add.s32 %r879, %r878, -1; and.b32 %r880, %r879, %r872; shl.b32 %r881, %r880, %r874; ld.u32 %r882, [%rd681+96]; shr.u32 %r883, %r882, %r881; shl.b32 %r884, %r865, %r874; mov.u32 %r885, -1; shl.b32 %r886, %r885, %r884; not.b32 %r887, %r886; and.b32 %r888, %r883, %r887; cvt.rn.f32.u32 %f399, %r888; ld.v2.f32 {%f400, %f401}, [%rd2307+80]; fma.rn.ftz.f32 %f929, %f401, %f399, %f400; mov.b64 %rd682, {%r10, %r864}; shr.u64 %rd683, %rd682, 32; cvt.u32.u64 %r4782, %rd683; mov.u32 %r4830, %r12; mov.u32 %r4827, %r10; bra.uni $L__BB5_31; $L__BB5_30: ld.f32 %f929, [%rd5+4]; mov.u32 %r4875, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r4830, %r4875; mov.u32 %r4782, %r4875; mov.u32 %r4908, %r4875; mov.u32 %r4868, %r4875; mov.u32 %r4869, %r4875; mov.u32 %r4870, %r4875; mov.u32 %r4825, %r4875; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r4827, %r4875; bra.uni $L__BB5_31; $L__BB5_14: ld.f32 %f953, [%rd11+12]; bra.uni $L__BB5_834; $L__BB5_29: ld.f32 %f929, [%rd26+12]; mov.u32 %r4875, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r4830, %r4875; mov.u32 %r4782, %r4875; mov.u32 %r4908, %r4875; mov.u32 %r4868, %r4875; mov.u32 %r4869, %r4875; mov.u32 %r4870, %r4875; mov.u32 %r4825, %r4875; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r4827, %r4875; bra.uni $L__BB5_31; $L__BB5_13: ld.f32 %f953, [%rd14]; bra.uni $L__BB5_834; $L__BB5_28: ld.f32 %f929, [%rd29]; mov.u32 %r4782, 2147483647; mov.u64 %rd2311, 0; mov.u32 %r4830, %r12; mov.u32 %r4868, %r4782; mov.u32 %r4869, %r4782; mov.u32 %r4870, %r4782; mov.u32 %r4825, %r4782; mov.u64 %rd2307, %rd2311; mov.u32 %r4827, %r4782; bra.uni $L__BB5_31; $L__BB5_12: ld.f32 %f953, [%rd16]; bra.uni $L__BB5_834; $L__BB5_27: ld.f32 %f929, [%rd31]; mov.u64 %rd2307, 0; mov.u32 %r4830, %r12; mov.u32 %r4825, %r4782; mov.u32 %r4827, %r4782; $L__BB5_31: add.s32 %r74, %r9, 1; setp.ne.s32 %p17, %r10, %r4827; mov.u16 %rs453, 0; @%p17 bra $L__BB5_34; and.b32 %r1028, %r8, -8; setp.ne.s32 %p18, %r1028, %r4782; @%p18 bra $L__BB5_34; and.b32 %r1029, %r74, -8; setp.eq.s32 %p19, %r1029, %r4825; selp.u16 %rs453, 1, 0, %p19; $L__BB5_34: setp.eq.s16 %p20, %rs453, 0; @%p20 bra $L__BB5_36; shl.b32 %r1030, %r7, 6; and.b32 %r1031, %r1030, 448; shl.b32 %r1032, %r8, 3; and.b32 %r1033, %r1032, 56; or.b32 %r1034, %r1033, %r1031; and.b32 %r1035, %r74, 7; or.b32 %r1036, %r1034, %r1035; ld.u8 %rs139, [%rd2307+15]; shr.u16 %rs140, %rs139, 5; mov.u32 %r1037, 5; cvt.u32.u16 %r1038, %rs140; sub.s32 %r1039, %r1037, %r1038; shr.u32 %r1040, %r1036, %r1039; mul.wide.u32 %rd698, %r1040, 4; add.s64 %rd699, %rd2307, %rd698; mov.u32 %r1041, 32; shr.u32 %r1042, %r1041, %r1038; add.s32 %r1043, %r1042, -1; and.b32 %r1044, %r1043, %r1036; shl.b32 %r1045, %r1044, %r1038; ld.u32 %r1046, [%rd699+96]; shr.u32 %r1047, %r1046, %r1045; mov.u32 %r1048, 1; shl.b32 %r1049, %r1048, %r1038; mov.u32 %r1050, -1; shl.b32 %r1051, %r1050, %r1049; not.b32 %r1052, %r1051; and.b32 %r1053, %r1047, %r1052; cvt.rn.f32.u32 %f404, %r1053; ld.v2.f32 {%f405, %f406}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f406, %f404, %f405; $L__BB5_66: and.b32 %r4722, %r7, -8; sub.ftz.f32 %f424, %f930, %f929; fma.rn.ftz.f32 %f40, %f21, %f424, %f929; add.s32 %r139, %r8, 1; setp.ne.s32 %p39, %r4722, %r4827; mov.u16 %rs456, 0; @%p39 bra $L__BB5_69; and.b32 %r1181, %r139, -8; setp.ne.s32 %p40, %r1181, %r4782; @%p40 bra $L__BB5_69; and.b32 %r1182, %r74, -8; setp.eq.s32 %p41, %r1182, %r4825; selp.u16 %rs456, 1, 0, %p41; $L__BB5_69: setp.eq.s16 %p42, %rs456, 0; @%p42 bra $L__BB5_71; shl.b32 %r1183, %r7, 6; and.b32 %r1184, %r1183, 448; shl.b32 %r1185, %r139, 3; and.b32 %r1186, %r1185, 56; or.b32 %r1187, %r1186, %r1184; and.b32 %r1188, %r74, 7; or.b32 %r1189, %r1187, %r1188; ld.u8 %rs153, [%rd2307+15]; shr.u16 %rs154, %rs153, 5; mov.u32 %r1190, 5; cvt.u32.u16 %r1191, %rs154; sub.s32 %r1192, %r1190, %r1191; shr.u32 %r1193, %r1189, %r1192; mul.wide.u32 %rd767, %r1193, 4; add.s64 %rd768, %rd2307, %rd767; mov.u32 %r1194, 32; shr.u32 %r1195, %r1194, %r1191; add.s32 %r1196, %r1195, -1; and.b32 %r1197, %r1196, %r1189; shl.b32 %r1198, %r1197, %r1191; ld.u32 %r1199, [%rd768+96]; shr.u32 %r1200, %r1199, %r1198; mov.u32 %r1201, 1; shl.b32 %r1202, %r1201, %r1191; mov.u32 %r1203, -1; shl.b32 %r1204, %r1203, %r1202; not.b32 %r1205, %r1204; and.b32 %r1206, %r1200, %r1205; cvt.rn.f32.u32 %f425, %r1206; ld.v2.f32 {%f426, %f427}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f427, %f425, %f426; $L__BB5_101: and.b32 %r4692, %r7, -8; setp.ne.s32 %p61, %r4692, %r4827; mov.u16 %rs459, 0; @%p61 bra $L__BB5_104; and.b32 %r1335, %r139, -8; setp.ne.s32 %p62, %r1335, %r4782; @%p62 bra $L__BB5_104; and.b32 %r1336, %r9, -8; setp.eq.s32 %p63, %r1336, %r4825; selp.u16 %rs459, 1, 0, %p63; $L__BB5_104: setp.eq.s16 %p64, %rs459, 0; @%p64 bra $L__BB5_106; and.b32 %r1337, %r7, 7; shl.b32 %r1338, %r139, 3; and.b32 %r1339, %r1338, 56; and.b32 %r1340, %r9, 7; bfi.b32 %r1341, %r1337, %r1340, 6, 3; or.b32 %r1342, %r1341, %r1339; ld.u8 %rs167, [%rd2307+15]; shr.u16 %rs168, %rs167, 5; mov.u32 %r1343, 5; cvt.u32.u16 %r1344, %rs168; sub.s32 %r1345, %r1343, %r1344; shr.u32 %r1346, %r1342, %r1345; mul.wide.u32 %rd837, %r1346, 4; add.s64 %rd838, %rd2307, %rd837; mov.u32 %r1347, 32; shr.u32 %r1348, %r1347, %r1344; add.s32 %r1349, %r1348, -1; and.b32 %r1350, %r1349, %r1342; shl.b32 %r1351, %r1350, %r1344; ld.u32 %r1352, [%rd838+96]; shr.u32 %r1353, %r1352, %r1351; mov.u32 %r1354, 1; shl.b32 %r1355, %r1354, %r1344; mov.u32 %r1356, -1; shl.b32 %r1357, %r1356, %r1355; not.b32 %r1358, %r1357; and.b32 %r1359, %r1353, %r1358; cvt.rn.f32.u32 %f445, %r1359; ld.v2.f32 {%f446, %f447}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f447, %f445, %f446; $L__BB5_136: sub.ftz.f32 %f465, %f931, %f932; fma.rn.ftz.f32 %f466, %f21, %f465, %f932; sub.ftz.f32 %f467, %f466, %f40; fma.rn.ftz.f32 %f65, %f20, %f467, %f40; add.s32 %r259, %r7, 1; and.b32 %r260, %r259, -8; setp.ne.s32 %p83, %r260, %r4827; mov.u16 %rs462, 0; @%p83 bra $L__BB5_139; and.b32 %r1484, %r139, -8; setp.ne.s32 %p84, %r1484, %r4782; @%p84 bra $L__BB5_139; and.b32 %r1485, %r9, -8; setp.eq.s32 %p85, %r1485, %r4825; selp.u16 %rs462, 1, 0, %p85; $L__BB5_139: setp.eq.s16 %p86, %rs462, 0; @%p86 bra $L__BB5_141; and.b32 %r1486, %r259, 7; shl.b32 %r1487, %r139, 3; and.b32 %r1488, %r1487, 56; and.b32 %r1489, %r9, 7; bfi.b32 %r1490, %r1486, %r1489, 6, 3; or.b32 %r1491, %r1490, %r1488; ld.u8 %rs181, [%rd2307+15]; shr.u16 %rs182, %rs181, 5; mov.u32 %r1492, 5; cvt.u32.u16 %r1493, %rs182; sub.s32 %r1494, %r1492, %r1493; shr.u32 %r1495, %r1491, %r1494; mul.wide.u32 %rd905, %r1495, 4; add.s64 %rd906, %rd2307, %rd905; mov.u32 %r1496, 32; shr.u32 %r1497, %r1496, %r1493; add.s32 %r1498, %r1497, -1; and.b32 %r1499, %r1498, %r1491; shl.b32 %r1500, %r1499, %r1493; ld.u32 %r1501, [%rd906+96]; shr.u32 %r1502, %r1501, %r1500; mov.u32 %r1503, 1; shl.b32 %r1504, %r1503, %r1493; mov.u32 %r1505, -1; shl.b32 %r1506, %r1505, %r1504; not.b32 %r1507, %r1506; and.b32 %r1508, %r1502, %r1507; cvt.rn.f32.u32 %f468, %r1508; ld.v2.f32 {%f469, %f470}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f470, %f468, %f469; $L__BB5_171: and.b32 %r4763, %r259, -8; setp.ne.s32 %p105, %r4763, %r4827; mov.u16 %rs465, 0; @%p105 bra $L__BB5_174; and.b32 %r1634, %r139, -8; setp.ne.s32 %p106, %r1634, %r4782; @%p106 bra $L__BB5_174; and.b32 %r1635, %r74, -8; setp.eq.s32 %p107, %r1635, %r4825; selp.u16 %rs465, 1, 0, %p107; $L__BB5_174: setp.eq.s16 %p108, %rs465, 0; @%p108 bra $L__BB5_176; shl.b32 %r1636, %r259, 6; and.b32 %r1637, %r1636, 448; shl.b32 %r1638, %r139, 3; and.b32 %r1639, %r1638, 56; or.b32 %r1640, %r1639, %r1637; and.b32 %r1641, %r74, 7; or.b32 %r1642, %r1640, %r1641; ld.u8 %rs195, [%rd2307+15]; shr.u16 %rs196, %rs195, 5; mov.u32 %r1643, 5; cvt.u32.u16 %r1644, %rs196; sub.s32 %r1645, %r1643, %r1644; shr.u32 %r1646, %r1642, %r1645; mul.wide.u32 %rd976, %r1646, 4; add.s64 %rd977, %rd2307, %rd976; mov.u32 %r1647, 32; shr.u32 %r1648, %r1647, %r1644; add.s32 %r1649, %r1648, -1; and.b32 %r1650, %r1649, %r1642; shl.b32 %r1651, %r1650, %r1644; ld.u32 %r1652, [%rd977+96]; shr.u32 %r1653, %r1652, %r1651; mov.u32 %r1654, 1; shl.b32 %r1655, %r1654, %r1644; mov.u32 %r1656, -1; shl.b32 %r1657, %r1656, %r1655; not.b32 %r1658, %r1657; and.b32 %r1659, %r1653, %r1658; cvt.rn.f32.u32 %f488, %r1659; ld.v2.f32 {%f489, %f490}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f490, %f488, %f489; $L__BB5_206: and.b32 %r4759, %r259, -8; sub.ftz.f32 %f508, %f934, %f933; fma.rn.ftz.f32 %f90, %f21, %f508, %f933; setp.ne.s32 %p127, %r4759, %r4827; mov.u16 %rs468, 0; @%p127 bra $L__BB5_209; and.b32 %r1789, %r8, -8; setp.ne.s32 %p128, %r1789, %r4782; @%p128 bra $L__BB5_209; and.b32 %r1790, %r74, -8; setp.eq.s32 %p129, %r1790, %r4825; selp.u16 %rs468, 1, 0, %p129; $L__BB5_209: setp.eq.s16 %p130, %rs468, 0; @%p130 bra $L__BB5_211; shl.b32 %r1791, %r259, 6; and.b32 %r1792, %r1791, 448; shl.b32 %r1793, %r8, 3; and.b32 %r1794, %r1793, 56; or.b32 %r1795, %r1792, %r1794; and.b32 %r1796, %r74, 7; or.b32 %r1797, %r1795, %r1796; ld.u8 %rs209, [%rd2307+15]; shr.u16 %rs210, %rs209, 5; mov.u32 %r1798, 5; cvt.u32.u16 %r1799, %rs210; sub.s32 %r1800, %r1798, %r1799; shr.u32 %r1801, %r1797, %r1800; mul.wide.u32 %rd1047, %r1801, 4; add.s64 %rd1048, %rd2307, %rd1047; mov.u32 %r1802, 32; shr.u32 %r1803, %r1802, %r1799; add.s32 %r1804, %r1803, -1; and.b32 %r1805, %r1804, %r1797; shl.b32 %r1806, %r1805, %r1799; ld.u32 %r1807, [%rd1048+96]; shr.u32 %r1808, %r1807, %r1806; mov.u32 %r1809, 1; shl.b32 %r1810, %r1809, %r1799; mov.u32 %r1811, -1; shl.b32 %r1812, %r1811, %r1810; not.b32 %r1813, %r1812; and.b32 %r1814, %r1808, %r1813; cvt.rn.f32.u32 %f509, %r1814; ld.v2.f32 {%f510, %f511}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f511, %f509, %f510; $L__BB5_241: and.b32 %r4743, %r259, -8; setp.ne.s32 %p149, %r4743, %r4827; mov.u16 %rs471, 0; @%p149 bra $L__BB5_244; and.b32 %r1943, %r8, -8; setp.ne.s32 %p150, %r1943, %r4782; @%p150 bra $L__BB5_244; and.b32 %r1944, %r9, -8; setp.eq.s32 %p151, %r1944, %r4825; selp.u16 %rs471, 1, 0, %p151; $L__BB5_244: setp.eq.s16 %p152, %rs471, 0; @%p152 bra $L__BB5_246; shl.b32 %r1945, %r259, 6; and.b32 %r1946, %r1945, 448; shl.b32 %r1947, %r8, 3; and.b32 %r1948, %r1947, 56; or.b32 %r1949, %r1946, %r1948; and.b32 %r1950, %r9, 7; or.b32 %r1951, %r1949, %r1950; ld.u8 %rs223, [%rd2307+15]; shr.u16 %rs224, %rs223, 5; mov.u32 %r1952, 5; cvt.u32.u16 %r1953, %rs224; sub.s32 %r1954, %r1952, %r1953; shr.u32 %r1955, %r1951, %r1954; mul.wide.u32 %rd1118, %r1955, 4; add.s64 %rd1119, %rd2307, %rd1118; mov.u32 %r1956, 32; shr.u32 %r1957, %r1956, %r1953; add.s32 %r1958, %r1957, -1; and.b32 %r1959, %r1958, %r1951; shl.b32 %r1960, %r1959, %r1953; ld.u32 %r1961, [%rd1119+96]; shr.u32 %r1962, %r1961, %r1960; mov.u32 %r1963, 1; shl.b32 %r1964, %r1963, %r1953; mov.u32 %r1965, -1; shl.b32 %r1966, %r1965, %r1964; not.b32 %r1967, %r1966; and.b32 %r1968, %r1962, %r1967; cvt.rn.f32.u32 %f529, %r1968; ld.v2.f32 {%f530, %f531}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f531, %f529, %f530; $L__BB5_276: cvt.rmi.ftz.f32.f32 %f915, %f13; sub.ftz.f32 %f549, %f935, %f936; fma.rn.ftz.f32 %f550, %f21, %f549, %f936; sub.ftz.f32 %f551, %f90, %f550; fma.rn.ftz.f32 %f552, %f20, %f551, %f550; sub.ftz.f32 %f553, %f552, %f65; sub.ftz.f32 %f554, %f13, %f915; fma.rn.ftz.f32 %f115, %f554, %f553, %f65; cvt.rmi.ftz.f32.f32 %f555, %f14; sub.ftz.f32 %f116, %f14, %f555; cvt.rzi.ftz.s32.f32 %r471, %f555; and.b32 %r472, %r471, -8; setp.ne.s32 %p171, %r472, %r4827; mov.u16 %rs474, 0; @%p171 bra $L__BB5_279; and.b32 %r2096, %r8, -8; setp.ne.s32 %p172, %r2096, %r4782; @%p172 bra $L__BB5_279; and.b32 %r2097, %r9, -8; setp.eq.s32 %p173, %r2097, %r4825; selp.u16 %rs474, 1, 0, %p173; $L__BB5_279: setp.eq.s16 %p174, %rs474, 0; @%p174 bra $L__BB5_281; shl.b32 %r2098, %r471, 6; and.b32 %r2099, %r2098, 448; shl.b32 %r2100, %r8, 3; and.b32 %r2101, %r2100, 56; or.b32 %r2102, %r2101, %r2099; and.b32 %r2103, %r9, 7; or.b32 %r2104, %r2102, %r2103; ld.u8 %rs237, [%rd2307+15]; shr.u16 %rs238, %rs237, 5; mov.u32 %r2105, 5; cvt.u32.u16 %r2106, %rs238; sub.s32 %r2107, %r2105, %r2106; shr.u32 %r2108, %r2104, %r2107; mul.wide.u32 %rd1188, %r2108, 4; add.s64 %rd1189, %rd2307, %rd1188; mov.u32 %r2109, 32; shr.u32 %r2110, %r2109, %r2106; add.s32 %r2111, %r2110, -1; and.b32 %r2112, %r2111, %r2104; shl.b32 %r2113, %r2112, %r2106; ld.u32 %r2114, [%rd1189+96]; shr.u32 %r2115, %r2114, %r2113; mov.u32 %r2116, 1; shl.b32 %r2117, %r2116, %r2106; mov.u32 %r2118, -1; shl.b32 %r2119, %r2118, %r2117; not.b32 %r2120, %r2119; and.b32 %r2121, %r2115, %r2120; cvt.rn.f32.u32 %f556, %r2121; ld.v2.f32 {%f557, %f558}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f558, %f556, %f557; $L__BB5_311: setp.ne.s32 %p193, %r472, %r4827; mov.u16 %rs477, 0; @%p193 bra $L__BB5_314; and.b32 %r2249, %r8, -8; setp.ne.s32 %p194, %r2249, %r4782; @%p194 bra $L__BB5_314; and.b32 %r2250, %r74, -8; setp.eq.s32 %p195, %r2250, %r4825; selp.u16 %rs477, 1, 0, %p195; $L__BB5_314: setp.eq.s16 %p196, %rs477, 0; @%p196 bra $L__BB5_316; shl.b32 %r2251, %r471, 6; and.b32 %r2252, %r2251, 448; shl.b32 %r2253, %r8, 3; and.b32 %r2254, %r2253, 56; or.b32 %r2255, %r2254, %r2252; and.b32 %r2256, %r74, 7; or.b32 %r2257, %r2255, %r2256; ld.u8 %rs251, [%rd2307+15]; shr.u16 %rs252, %rs251, 5; mov.u32 %r2258, 5; cvt.u32.u16 %r2259, %rs252; sub.s32 %r2260, %r2258, %r2259; shr.u32 %r2261, %r2257, %r2260; mul.wide.u32 %rd1258, %r2261, 4; add.s64 %rd1259, %rd2307, %rd1258; mov.u32 %r2262, 32; shr.u32 %r2263, %r2262, %r2259; add.s32 %r2264, %r2263, -1; and.b32 %r2265, %r2264, %r2257; shl.b32 %r2266, %r2265, %r2259; ld.u32 %r2267, [%rd1259+96]; shr.u32 %r2268, %r2267, %r2266; mov.u32 %r2269, 1; shl.b32 %r2270, %r2269, %r2259; mov.u32 %r2271, -1; shl.b32 %r2272, %r2271, %r2270; not.b32 %r2273, %r2272; and.b32 %r2274, %r2268, %r2273; cvt.rn.f32.u32 %f576, %r2274; ld.v2.f32 {%f577, %f578}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f578, %f576, %f577; $L__BB5_346: sub.ftz.f32 %f596, %f938, %f937; fma.rn.ftz.f32 %f141, %f21, %f596, %f937; setp.ne.s32 %p215, %r472, %r4827; mov.u16 %rs480, 0; @%p215 bra $L__BB5_349; and.b32 %r2403, %r139, -8; setp.ne.s32 %p216, %r2403, %r4782; @%p216 bra $L__BB5_349; and.b32 %r2404, %r74, -8; setp.eq.s32 %p217, %r2404, %r4825; selp.u16 %rs480, 1, 0, %p217; $L__BB5_349: setp.eq.s16 %p218, %rs480, 0; @%p218 bra $L__BB5_351; shl.b32 %r2405, %r471, 6; and.b32 %r2406, %r2405, 448; shl.b32 %r2407, %r139, 3; and.b32 %r2408, %r2407, 56; or.b32 %r2409, %r2408, %r2406; and.b32 %r2410, %r74, 7; or.b32 %r2411, %r2409, %r2410; ld.u8 %rs265, [%rd2307+15]; shr.u16 %rs266, %rs265, 5; mov.u32 %r2412, 5; cvt.u32.u16 %r2413, %rs266; sub.s32 %r2414, %r2412, %r2413; shr.u32 %r2415, %r2411, %r2414; mul.wide.u32 %rd1329, %r2415, 4; add.s64 %rd1330, %rd2307, %rd1329; mov.u32 %r2416, 32; shr.u32 %r2417, %r2416, %r2413; add.s32 %r2418, %r2417, -1; and.b32 %r2419, %r2418, %r2411; shl.b32 %r2420, %r2419, %r2413; ld.u32 %r2421, [%rd1330+96]; shr.u32 %r2422, %r2421, %r2420; mov.u32 %r2423, 1; shl.b32 %r2424, %r2423, %r2413; mov.u32 %r2425, -1; shl.b32 %r2426, %r2425, %r2424; not.b32 %r2427, %r2426; and.b32 %r2428, %r2422, %r2427; cvt.rn.f32.u32 %f597, %r2428; ld.v2.f32 {%f598, %f599}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f599, %f597, %f598; $L__BB5_381: setp.ne.s32 %p237, %r472, %r4827; mov.u16 %rs483, 0; @%p237 bra $L__BB5_384; and.b32 %r2558, %r139, -8; setp.ne.s32 %p238, %r2558, %r4782; @%p238 bra $L__BB5_384; and.b32 %r2559, %r9, -8; setp.eq.s32 %p239, %r2559, %r4825; selp.u16 %rs483, 1, 0, %p239; $L__BB5_384: setp.eq.s16 %p240, %rs483, 0; @%p240 bra $L__BB5_386; and.b32 %r2560, %r471, 7; shl.b32 %r2561, %r139, 3; and.b32 %r2562, %r2561, 56; and.b32 %r2563, %r9, 7; bfi.b32 %r2564, %r2560, %r2563, 6, 3; or.b32 %r2565, %r2564, %r2562; ld.u8 %rs279, [%rd2307+15]; shr.u16 %rs280, %rs279, 5; mov.u32 %r2566, 5; cvt.u32.u16 %r2567, %rs280; sub.s32 %r2568, %r2566, %r2567; shr.u32 %r2569, %r2565, %r2568; mul.wide.u32 %rd1400, %r2569, 4; add.s64 %rd1401, %rd2307, %rd1400; mov.u32 %r2570, 32; shr.u32 %r2571, %r2570, %r2567; add.s32 %r2572, %r2571, -1; and.b32 %r2573, %r2572, %r2565; shl.b32 %r2574, %r2573, %r2567; ld.u32 %r2575, [%rd1401+96]; shr.u32 %r2576, %r2575, %r2574; mov.u32 %r2577, 1; shl.b32 %r2578, %r2577, %r2567; mov.u32 %r2579, -1; shl.b32 %r2580, %r2579, %r2578; not.b32 %r2581, %r2580; and.b32 %r2582, %r2576, %r2581; cvt.rn.f32.u32 %f617, %r2582; ld.v2.f32 {%f618, %f619}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f619, %f617, %f618; $L__BB5_416: sub.ftz.f32 %f637, %f939, %f940; fma.rn.ftz.f32 %f638, %f21, %f637, %f940; sub.ftz.f32 %f639, %f638, %f141; fma.rn.ftz.f32 %f166, %f20, %f639, %f141; add.s32 %r635, %r471, 1; and.b32 %r636, %r635, -8; setp.ne.s32 %p259, %r636, %r4827; mov.u16 %rs486, 0; @%p259 bra $L__BB5_419; and.b32 %r2708, %r139, -8; setp.ne.s32 %p260, %r2708, %r4782; @%p260 bra $L__BB5_419; and.b32 %r2709, %r9, -8; setp.eq.s32 %p261, %r2709, %r4825; selp.u16 %rs486, 1, 0, %p261; $L__BB5_419: setp.eq.s16 %p262, %rs486, 0; @%p262 bra $L__BB5_421; and.b32 %r2710, %r635, 7; shl.b32 %r2711, %r139, 3; and.b32 %r2712, %r2711, 56; and.b32 %r2713, %r9, 7; bfi.b32 %r2714, %r2710, %r2713, 6, 3; or.b32 %r2715, %r2714, %r2712; ld.u8 %rs293, [%rd2307+15]; shr.u16 %rs294, %rs293, 5; mov.u32 %r2716, 5; cvt.u32.u16 %r2717, %rs294; sub.s32 %r2718, %r2716, %r2717; shr.u32 %r2719, %r2715, %r2718; mul.wide.u32 %rd1471, %r2719, 4; add.s64 %rd1472, %rd2307, %rd1471; mov.u32 %r2720, 32; shr.u32 %r2721, %r2720, %r2717; add.s32 %r2722, %r2721, -1; and.b32 %r2723, %r2722, %r2715; shl.b32 %r2724, %r2723, %r2717; ld.u32 %r2725, [%rd1472+96]; shr.u32 %r2726, %r2725, %r2724; mov.u32 %r2727, 1; shl.b32 %r2728, %r2727, %r2717; mov.u32 %r2729, -1; shl.b32 %r2730, %r2729, %r2728; not.b32 %r2731, %r2730; and.b32 %r2732, %r2726, %r2731; cvt.rn.f32.u32 %f640, %r2732; ld.v2.f32 {%f641, %f642}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f642, %f640, %f641; $L__BB5_451: setp.ne.s32 %p281, %r636, %r4827; mov.u16 %rs489, 0; @%p281 bra $L__BB5_454; and.b32 %r2858, %r139, -8; setp.ne.s32 %p282, %r2858, %r4782; @%p282 bra $L__BB5_454; and.b32 %r2859, %r74, -8; setp.eq.s32 %p283, %r2859, %r4825; selp.u16 %rs489, 1, 0, %p283; $L__BB5_454: setp.eq.s16 %p284, %rs489, 0; @%p284 bra $L__BB5_456; shl.b32 %r2860, %r635, 6; and.b32 %r2861, %r2860, 448; shl.b32 %r2862, %r139, 3; and.b32 %r2863, %r2862, 56; or.b32 %r2864, %r2863, %r2861; and.b32 %r2865, %r74, 7; or.b32 %r2866, %r2864, %r2865; ld.u8 %rs307, [%rd2307+15]; shr.u16 %rs308, %rs307, 5; mov.u32 %r2867, 5; cvt.u32.u16 %r2868, %rs308; sub.s32 %r2869, %r2867, %r2868; shr.u32 %r2870, %r2866, %r2869; mul.wide.u32 %rd1542, %r2870, 4; add.s64 %rd1543, %rd2307, %rd1542; mov.u32 %r2871, 32; shr.u32 %r2872, %r2871, %r2868; add.s32 %r2873, %r2872, -1; and.b32 %r2874, %r2873, %r2866; shl.b32 %r2875, %r2874, %r2868; ld.u32 %r2876, [%rd1543+96]; shr.u32 %r2877, %r2876, %r2875; mov.u32 %r2878, 1; shl.b32 %r2879, %r2878, %r2868; mov.u32 %r2880, -1; shl.b32 %r2881, %r2880, %r2879; not.b32 %r2882, %r2881; and.b32 %r2883, %r2877, %r2882; cvt.rn.f32.u32 %f660, %r2883; ld.v2.f32 {%f661, %f662}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f662, %f660, %f661; $L__BB5_486: sub.ftz.f32 %f680, %f942, %f941; fma.rn.ftz.f32 %f191, %f21, %f680, %f941; setp.ne.s32 %p303, %r636, %r4827; mov.u16 %rs492, 0; @%p303 bra $L__BB5_489; and.b32 %r3013, %r8, -8; setp.ne.s32 %p304, %r3013, %r4782; @%p304 bra $L__BB5_489; and.b32 %r3014, %r74, -8; setp.eq.s32 %p305, %r3014, %r4825; selp.u16 %rs492, 1, 0, %p305; $L__BB5_489: setp.eq.s16 %p306, %rs492, 0; @%p306 bra $L__BB5_491; shl.b32 %r3015, %r635, 6; and.b32 %r3016, %r3015, 448; shl.b32 %r3017, %r8, 3; and.b32 %r3018, %r3017, 56; or.b32 %r3019, %r3016, %r3018; and.b32 %r3020, %r74, 7; or.b32 %r3021, %r3019, %r3020; ld.u8 %rs321, [%rd2307+15]; shr.u16 %rs322, %rs321, 5; mov.u32 %r3022, 5; cvt.u32.u16 %r3023, %rs322; sub.s32 %r3024, %r3022, %r3023; shr.u32 %r3025, %r3021, %r3024; mul.wide.u32 %rd1613, %r3025, 4; add.s64 %rd1614, %rd2307, %rd1613; mov.u32 %r3026, 32; shr.u32 %r3027, %r3026, %r3023; add.s32 %r3028, %r3027, -1; and.b32 %r3029, %r3028, %r3021; shl.b32 %r3030, %r3029, %r3023; ld.u32 %r3031, [%rd1614+96]; shr.u32 %r3032, %r3031, %r3030; mov.u32 %r3033, 1; shl.b32 %r3034, %r3033, %r3023; mov.u32 %r3035, -1; shl.b32 %r3036, %r3035, %r3034; not.b32 %r3037, %r3036; and.b32 %r3038, %r3032, %r3037; cvt.rn.f32.u32 %f681, %r3038; ld.v2.f32 {%f682, %f683}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f683, %f681, %f682; $L__BB5_521: setp.ne.s32 %p325, %r636, %r4827; mov.u16 %rs495, 0; @%p325 bra $L__BB5_524; and.b32 %r3167, %r8, -8; setp.ne.s32 %p326, %r3167, %r4782; @%p326 bra $L__BB5_524; and.b32 %r3168, %r9, -8; setp.eq.s32 %p327, %r3168, %r4825; selp.u16 %rs495, 1, 0, %p327; $L__BB5_524: setp.eq.s16 %p328, %rs495, 0; @%p328 bra $L__BB5_526; shl.b32 %r3169, %r635, 6; and.b32 %r3170, %r3169, 448; shl.b32 %r3171, %r8, 3; and.b32 %r3172, %r3171, 56; or.b32 %r3173, %r3170, %r3172; and.b32 %r3174, %r9, 7; or.b32 %r3175, %r3173, %r3174; ld.u8 %rs335, [%rd2307+15]; shr.u16 %rs336, %rs335, 5; mov.u32 %r3176, 5; cvt.u32.u16 %r3177, %rs336; sub.s32 %r3178, %r3176, %r3177; shr.u32 %r3179, %r3175, %r3178; mul.wide.u32 %rd1684, %r3179, 4; add.s64 %rd1685, %rd2307, %rd1684; mov.u32 %r3180, 32; shr.u32 %r3181, %r3180, %r3177; add.s32 %r3182, %r3181, -1; and.b32 %r3183, %r3182, %r3175; shl.b32 %r3184, %r3183, %r3177; ld.u32 %r3185, [%rd1685+96]; shr.u32 %r3186, %r3185, %r3184; mov.u32 %r3187, 1; shl.b32 %r3188, %r3187, %r3177; mov.u32 %r3189, -1; shl.b32 %r3190, %r3189, %r3188; not.b32 %r3191, %r3190; and.b32 %r3192, %r3186, %r3191; cvt.rn.f32.u32 %f701, %r3192; ld.v2.f32 {%f702, %f703}, [%rd2307+80]; fma.rn.ftz.f32 %f944, %f703, %f701, %f702; $L__BB5_556: add.u64 %rd2288, %SP, 80; sub.ftz.f32 %f721, %f943, %f944; fma.rn.ftz.f32 %f722, %f21, %f721, %f944; sub.ftz.f32 %f723, %f191, %f722; fma.rn.ftz.f32 %f724, %f20, %f723, %f722; sub.ftz.f32 %f725, %f724, %f166; fma.rn.ftz.f32 %f726, %f116, %f725, %f166; mul.ftz.f32 %f727, %f8, %f726; fma.rn.ftz.f32 %f728, %f7, %f115, %f727; mul.ftz.f32 %f729, %f9, %f728; { // callseq 14, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f730, [retval0+0]; } // callseq 14 { // callseq 15, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f731, [retval0+0]; } // callseq 15 mul.ftz.f32 %f732, %f8, %f731; fma.rn.ftz.f32 %f733, %f7, %f730, %f732; fma.rn.ftz.f32 %f734, %f10, %f733, %f729; mul.ftz.f32 %f216, %f11, %f734; { // callseq 16, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f735, [retval0+0]; } // callseq 16 { // callseq 17, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f736, [retval0+0]; } // callseq 17 mul.ftz.f32 %f737, %f8, %f736; fma.rn.ftz.f32 %f738, %f7, %f735, %f737; mul.ftz.f32 %f217, %f9, %f738; { // callseq 18, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f739, [retval0+0]; } // callseq 18 mul.ftz.f32 %f218, %f7, %f739; cvt.rmi.ftz.f32.f32 %f740, %f16; sub.ftz.f32 %f219, %f16, %f740; cvt.rmi.ftz.f32.f32 %f741, %f18; sub.ftz.f32 %f220, %f18, %f741; cvt.rzi.ftz.s32.f32 %r732, %f740; cvt.rzi.ftz.s32.f32 %r733, %f741; ld.local.u32 %r5196, [%rd2]; setp.ne.s32 %p347, %r472, %r5196; mov.u16 %rs498, 0; @%p347 bra $L__BB5_559; and.b32 %r3328, %r732, -8; ld.local.u32 %r3329, [%rd2+4]; setp.ne.s32 %p348, %r3328, %r3329; @%p348 bra $L__BB5_559; and.b32 %r3330, %r733, -8; ld.local.u32 %r3331, [%rd2+8]; setp.eq.s32 %p349, %r3330, %r3331; selp.u16 %rs498, 1, 0, %p349; $L__BB5_559: setp.eq.s16 %p350, %rs498, 0; @%p350 bra $L__BB5_561; ld.local.u64 %rd1749, [%rd2+48]; shl.b32 %r3332, %r471, 6; and.b32 %r3333, %r3332, 448; shl.b32 %r3334, %r732, 3; and.b32 %r3335, %r3334, 56; or.b32 %r3336, %r3335, %r3333; and.b32 %r3337, %r733, 7; or.b32 %r3338, %r3336, %r3337; ld.u8 %rs349, [%rd1749+15]; shr.u16 %rs350, %rs349, 5; mov.u32 %r3339, 5; cvt.u32.u16 %r3340, %rs350; sub.s32 %r3341, %r3339, %r3340; shr.u32 %r3342, %r3338, %r3341; mul.wide.u32 %rd1750, %r3342, 4; add.s64 %rd1751, %rd1749, %rd1750; mov.u32 %r3343, 32; shr.u32 %r3344, %r3343, %r3340; add.s32 %r3345, %r3344, -1; and.b32 %r3346, %r3345, %r3338; shl.b32 %r3347, %r3346, %r3340; ld.u32 %r3348, [%rd1751+96]; shr.u32 %r3349, %r3348, %r3347; mov.u32 %r3350, 1; shl.b32 %r3351, %r3350, %r3340; mov.u32 %r3352, -1; shl.b32 %r3353, %r3352, %r3351; not.b32 %r3354, %r3353; and.b32 %r3355, %r3349, %r3354; cvt.rn.f32.u32 %f742, %r3355; ld.v2.f32 {%f743, %f744}, [%rd1749+80]; fma.rn.ftz.f32 %f945, %f744, %f742, %f743; $L__BB5_591: add.s32 %r741, %r733, 1; setp.ne.s32 %p369, %r472, %r5196; mov.u16 %rs501, 0; @%p369 bra $L__BB5_594; and.b32 %r3499, %r732, -8; ld.local.u32 %r3500, [%rd2+4]; setp.ne.s32 %p370, %r3499, %r3500; @%p370 bra $L__BB5_594; and.b32 %r3501, %r741, -8; ld.local.u32 %r3502, [%rd2+8]; setp.eq.s32 %p371, %r3501, %r3502; selp.u16 %rs501, 1, 0, %p371; $L__BB5_594: setp.eq.s16 %p372, %rs501, 0; @%p372 bra $L__BB5_596; ld.local.u64 %rd1816, [%rd2+48]; shl.b32 %r3503, %r471, 6; and.b32 %r3504, %r3503, 448; shl.b32 %r3505, %r732, 3; and.b32 %r3506, %r3505, 56; or.b32 %r3507, %r3506, %r3504; and.b32 %r3508, %r741, 7; or.b32 %r3509, %r3507, %r3508; ld.u8 %rs363, [%rd1816+15]; shr.u16 %rs364, %rs363, 5; mov.u32 %r3510, 5; cvt.u32.u16 %r3511, %rs364; sub.s32 %r3512, %r3510, %r3511; shr.u32 %r3513, %r3509, %r3512; mul.wide.u32 %rd1817, %r3513, 4; add.s64 %rd1818, %rd1816, %rd1817; mov.u32 %r3514, 32; shr.u32 %r3515, %r3514, %r3511; add.s32 %r3516, %r3515, -1; and.b32 %r3517, %r3516, %r3509; shl.b32 %r3518, %r3517, %r3511; ld.u32 %r3519, [%rd1818+96]; shr.u32 %r3520, %r3519, %r3518; mov.u32 %r3521, 1; shl.b32 %r3522, %r3521, %r3511; mov.u32 %r3523, -1; shl.b32 %r3524, %r3523, %r3522; not.b32 %r3525, %r3524; and.b32 %r3526, %r3520, %r3525; cvt.rn.f32.u32 %f762, %r3526; ld.v2.f32 {%f763, %f764}, [%rd1816+80]; fma.rn.ftz.f32 %f946, %f764, %f762, %f763; $L__BB5_626: sub.ftz.f32 %f782, %f946, %f945; fma.rn.ftz.f32 %f245, %f220, %f782, %f945; add.s32 %r748, %r732, 1; setp.ne.s32 %p391, %r472, %r5196; mov.u16 %rs504, 0; @%p391 bra $L__BB5_629; and.b32 %r3670, %r748, -8; ld.local.u32 %r3671, [%rd2+4]; setp.ne.s32 %p392, %r3670, %r3671; @%p392 bra $L__BB5_629; and.b32 %r3672, %r741, -8; ld.local.u32 %r3673, [%rd2+8]; setp.eq.s32 %p393, %r3672, %r3673; selp.u16 %rs504, 1, 0, %p393; $L__BB5_629: setp.eq.s16 %p394, %rs504, 0; @%p394 bra $L__BB5_631; ld.local.u64 %rd1884, [%rd2+48]; shl.b32 %r3674, %r471, 6; and.b32 %r3675, %r3674, 448; shl.b32 %r3676, %r748, 3; and.b32 %r3677, %r3676, 56; or.b32 %r3678, %r3677, %r3675; and.b32 %r3679, %r741, 7; or.b32 %r3680, %r3678, %r3679; ld.u8 %rs377, [%rd1884+15]; shr.u16 %rs378, %rs377, 5; mov.u32 %r3681, 5; cvt.u32.u16 %r3682, %rs378; sub.s32 %r3683, %r3681, %r3682; shr.u32 %r3684, %r3680, %r3683; mul.wide.u32 %rd1885, %r3684, 4; add.s64 %rd1886, %rd1884, %rd1885; mov.u32 %r3685, 32; shr.u32 %r3686, %r3685, %r3682; add.s32 %r3687, %r3686, -1; and.b32 %r3688, %r3687, %r3680; shl.b32 %r3689, %r3688, %r3682; ld.u32 %r3690, [%rd1886+96]; shr.u32 %r3691, %r3690, %r3689; mov.u32 %r3692, 1; shl.b32 %r3693, %r3692, %r3682; mov.u32 %r3694, -1; shl.b32 %r3695, %r3694, %r3693; not.b32 %r3696, %r3695; and.b32 %r3697, %r3691, %r3696; cvt.rn.f32.u32 %f783, %r3697; ld.v2.f32 {%f784, %f785}, [%rd1884+80]; fma.rn.ftz.f32 %f947, %f785, %f783, %f784; $L__BB5_661: setp.ne.s32 %p413, %r472, %r5196; mov.u16 %rs507, 0; @%p413 bra $L__BB5_664; and.b32 %r3841, %r748, -8; ld.local.u32 %r3842, [%rd2+4]; setp.ne.s32 %p414, %r3841, %r3842; @%p414 bra $L__BB5_664; and.b32 %r3843, %r733, -8; ld.local.u32 %r3844, [%rd2+8]; setp.eq.s32 %p415, %r3843, %r3844; selp.u16 %rs507, 1, 0, %p415; $L__BB5_664: setp.eq.s16 %p416, %rs507, 0; @%p416 bra $L__BB5_666; ld.local.u64 %rd1951, [%rd2+48]; and.b32 %r3845, %r471, 7; shl.b32 %r3846, %r748, 3; and.b32 %r3847, %r3846, 56; and.b32 %r3848, %r733, 7; bfi.b32 %r3849, %r3845, %r3848, 6, 3; or.b32 %r3850, %r3849, %r3847; ld.u8 %rs391, [%rd1951+15]; shr.u16 %rs392, %rs391, 5; mov.u32 %r3851, 5; cvt.u32.u16 %r3852, %rs392; sub.s32 %r3853, %r3851, %r3852; shr.u32 %r3854, %r3850, %r3853; mul.wide.u32 %rd1952, %r3854, 4; add.s64 %rd1953, %rd1951, %rd1952; mov.u32 %r3855, 32; shr.u32 %r3856, %r3855, %r3852; add.s32 %r3857, %r3856, -1; and.b32 %r3858, %r3857, %r3850; shl.b32 %r3859, %r3858, %r3852; ld.u32 %r3860, [%rd1953+96]; shr.u32 %r3861, %r3860, %r3859; mov.u32 %r3862, 1; shl.b32 %r3863, %r3862, %r3852; mov.u32 %r3864, -1; shl.b32 %r3865, %r3864, %r3863; not.b32 %r3866, %r3865; and.b32 %r3867, %r3861, %r3866; cvt.rn.f32.u32 %f803, %r3867; ld.v2.f32 {%f804, %f805}, [%rd1951+80]; fma.rn.ftz.f32 %f948, %f805, %f803, %f804; $L__BB5_696: sub.ftz.f32 %f823, %f947, %f948; fma.rn.ftz.f32 %f824, %f220, %f823, %f948; sub.ftz.f32 %f825, %f824, %f245; fma.rn.ftz.f32 %f270, %f219, %f825, %f245; setp.ne.s32 %p435, %r636, %r5196; mov.u16 %rs510, 0; @%p435 bra $L__BB5_699; and.b32 %r4008, %r748, -8; ld.local.u32 %r4009, [%rd2+4]; setp.ne.s32 %p436, %r4008, %r4009; @%p436 bra $L__BB5_699; and.b32 %r4010, %r733, -8; ld.local.u32 %r4011, [%rd2+8]; setp.eq.s32 %p437, %r4010, %r4011; selp.u16 %rs510, 1, 0, %p437; $L__BB5_699: setp.eq.s16 %p438, %rs510, 0; @%p438 bra $L__BB5_701; ld.local.u64 %rd2018, [%rd2+48]; and.b32 %r4012, %r635, 7; shl.b32 %r4013, %r748, 3; and.b32 %r4014, %r4013, 56; and.b32 %r4015, %r733, 7; bfi.b32 %r4016, %r4012, %r4015, 6, 3; or.b32 %r4017, %r4016, %r4014; ld.u8 %rs405, [%rd2018+15]; shr.u16 %rs406, %rs405, 5; mov.u32 %r4018, 5; cvt.u32.u16 %r4019, %rs406; sub.s32 %r4020, %r4018, %r4019; shr.u32 %r4021, %r4017, %r4020; mul.wide.u32 %rd2019, %r4021, 4; add.s64 %rd2020, %rd2018, %rd2019; mov.u32 %r4022, 32; shr.u32 %r4023, %r4022, %r4019; add.s32 %r4024, %r4023, -1; and.b32 %r4025, %r4024, %r4017; shl.b32 %r4026, %r4025, %r4019; ld.u32 %r4027, [%rd2020+96]; shr.u32 %r4028, %r4027, %r4026; mov.u32 %r4029, 1; shl.b32 %r4030, %r4029, %r4019; mov.u32 %r4031, -1; shl.b32 %r4032, %r4031, %r4030; not.b32 %r4033, %r4032; and.b32 %r4034, %r4028, %r4033; cvt.rn.f32.u32 %f826, %r4034; ld.v2.f32 {%f827, %f828}, [%rd2018+80]; fma.rn.ftz.f32 %f949, %f828, %f826, %f827; $L__BB5_731: setp.ne.s32 %p457, %r636, %r5196; mov.u16 %rs513, 0; @%p457 bra $L__BB5_734; and.b32 %r4175, %r748, -8; ld.local.u32 %r4176, [%rd2+4]; setp.ne.s32 %p458, %r4175, %r4176; @%p458 bra $L__BB5_734; and.b32 %r4177, %r741, -8; ld.local.u32 %r4178, [%rd2+8]; setp.eq.s32 %p459, %r4177, %r4178; selp.u16 %rs513, 1, 0, %p459; $L__BB5_734: setp.eq.s16 %p460, %rs513, 0; @%p460 bra $L__BB5_736; ld.local.u64 %rd2085, [%rd2+48]; shl.b32 %r4179, %r635, 6; and.b32 %r4180, %r4179, 448; shl.b32 %r4181, %r748, 3; and.b32 %r4182, %r4181, 56; or.b32 %r4183, %r4182, %r4180; and.b32 %r4184, %r741, 7; or.b32 %r4185, %r4183, %r4184; ld.u8 %rs419, [%rd2085+15]; shr.u16 %rs420, %rs419, 5; mov.u32 %r4186, 5; cvt.u32.u16 %r4187, %rs420; sub.s32 %r4188, %r4186, %r4187; shr.u32 %r4189, %r4185, %r4188; mul.wide.u32 %rd2086, %r4189, 4; add.s64 %rd2087, %rd2085, %rd2086; mov.u32 %r4190, 32; shr.u32 %r4191, %r4190, %r4187; add.s32 %r4192, %r4191, -1; and.b32 %r4193, %r4192, %r4185; shl.b32 %r4194, %r4193, %r4187; ld.u32 %r4195, [%rd2087+96]; shr.u32 %r4196, %r4195, %r4194; mov.u32 %r4197, 1; shl.b32 %r4198, %r4197, %r4187; mov.u32 %r4199, -1; shl.b32 %r4200, %r4199, %r4198; not.b32 %r4201, %r4200; and.b32 %r4202, %r4196, %r4201; cvt.rn.f32.u32 %f846, %r4202; ld.v2.f32 {%f847, %f848}, [%rd2085+80]; fma.rn.ftz.f32 %f950, %f848, %f846, %f847; $L__BB5_766: sub.ftz.f32 %f866, %f950, %f949; fma.rn.ftz.f32 %f295, %f220, %f866, %f949; setp.ne.s32 %p479, %r636, %r5196; mov.u16 %rs516, 0; @%p479 bra $L__BB5_769; and.b32 %r4346, %r732, -8; ld.local.u32 %r4347, [%rd2+4]; setp.ne.s32 %p480, %r4346, %r4347; @%p480 bra $L__BB5_769; and.b32 %r4348, %r741, -8; ld.local.u32 %r4349, [%rd2+8]; setp.eq.s32 %p481, %r4348, %r4349; selp.u16 %rs516, 1, 0, %p481; $L__BB5_769: setp.eq.s16 %p482, %rs516, 0; @%p482 bra $L__BB5_771; ld.local.u64 %rd2152, [%rd2+48]; shl.b32 %r4350, %r635, 6; and.b32 %r4351, %r4350, 448; shl.b32 %r4352, %r732, 3; and.b32 %r4353, %r4352, 56; or.b32 %r4354, %r4351, %r4353; and.b32 %r4355, %r741, 7; or.b32 %r4356, %r4354, %r4355; ld.u8 %rs433, [%rd2152+15]; shr.u16 %rs434, %rs433, 5; mov.u32 %r4357, 5; cvt.u32.u16 %r4358, %rs434; sub.s32 %r4359, %r4357, %r4358; shr.u32 %r4360, %r4356, %r4359; mul.wide.u32 %rd2153, %r4360, 4; add.s64 %rd2154, %rd2152, %rd2153; mov.u32 %r4361, 32; shr.u32 %r4362, %r4361, %r4358; add.s32 %r4363, %r4362, -1; and.b32 %r4364, %r4363, %r4356; shl.b32 %r4365, %r4364, %r4358; ld.u32 %r4366, [%rd2154+96]; shr.u32 %r4367, %r4366, %r4365; mov.u32 %r4368, 1; shl.b32 %r4369, %r4368, %r4358; mov.u32 %r4370, -1; shl.b32 %r4371, %r4370, %r4369; not.b32 %r4372, %r4371; and.b32 %r4373, %r4367, %r4372; cvt.rn.f32.u32 %f867, %r4373; ld.v2.f32 {%f868, %f869}, [%rd2152+80]; fma.rn.ftz.f32 %f951, %f869, %f867, %f868; $L__BB5_801: setp.ne.s32 %p501, %r636, %r5196; @%p501 bra $L__BB5_804; and.b32 %r4517, %r732, -8; ld.local.u32 %r4518, [%rd2+4]; setp.ne.s32 %p502, %r4517, %r4518; @%p502 bra $L__BB5_804; and.b32 %r4519, %r733, -8; ld.local.u32 %r4520, [%rd2+8]; setp.eq.s32 %p503, %r4519, %r4520; @%p503 bra $L__BB5_832; bra.uni $L__BB5_804; $L__BB5_832: ld.local.u64 %rd2285, [%rd2+48]; shl.b32 %r4664, %r635, 6; and.b32 %r4665, %r4664, 448; shl.b32 %r4666, %r732, 3; and.b32 %r4667, %r4666, 56; or.b32 %r4668, %r4665, %r4667; and.b32 %r4669, %r733, 7; or.b32 %r4670, %r4668, %r4669; ld.u8 %rs451, [%rd2285+15]; shr.u16 %rs452, %rs451, 5; mov.u32 %r4671, 5; cvt.u32.u16 %r4672, %rs452; sub.s32 %r4673, %r4671, %r4672; shr.u32 %r4674, %r4670, %r4673; mul.wide.u32 %rd2286, %r4674, 4; add.s64 %rd2287, %rd2285, %rd2286; mov.u32 %r4675, 32; shr.u32 %r4676, %r4675, %r4672; add.s32 %r4677, %r4676, -1; and.b32 %r4678, %r4677, %r4670; shl.b32 %r4679, %r4678, %r4672; ld.u32 %r4680, [%rd2287+96]; shr.u32 %r4681, %r4680, %r4679; mov.u32 %r4682, 1; shl.b32 %r4683, %r4682, %r4672; mov.u32 %r4684, -1; shl.b32 %r4685, %r4684, %r4683; not.b32 %r4686, %r4685; and.b32 %r4687, %r4681, %r4686; cvt.rn.f32.u32 %f902, %r4687; ld.v2.f32 {%f903, %f904}, [%rd2285+80]; fma.rn.ftz.f32 %f952, %f904, %f902, %f903; bra.uni $L__BB5_833; $L__BB5_804: and.b32 %r779, %r635, -128; ld.local.u32 %r4521, [%rd2+12]; setp.ne.s32 %p504, %r779, %r4521; @%p504 bra $L__BB5_807; and.b32 %r4522, %r732, -128; ld.local.u32 %r4523, [%rd2+16]; setp.ne.s32 %p505, %r4522, %r4523; @%p505 bra $L__BB5_807; and.b32 %r4524, %r733, -128; ld.local.u32 %r4525, [%rd2+20]; setp.eq.s32 %p506, %r4524, %r4525; @%p506 bra $L__BB5_829; bra.uni $L__BB5_807; $L__BB5_829: shl.b32 %r4629, %r635, 5; and.b32 %r4630, %r4629, 3840; shl.b32 %r4631, %r732, 1; and.b32 %r4632, %r4631, 240; or.b32 %r4633, %r4630, %r4632; bfe.u32 %r4634, %r733, 3, 4; or.b32 %r4635, %r4633, %r4634; shr.u32 %r4636, %r4633, 6; ld.local.u64 %rd606, [%rd2+56]; mul.wide.u32 %rd2273, %r4636, 8; add.s64 %rd2274, %rd606, %rd2273; and.b32 %r4637, %r4635, 63; mov.u64 %rd2275, 1; shl.b64 %rd2276, %rd2275, %r4637; ld.u64 %rd2277, [%rd2274+544]; and.b64 %rd2278, %rd2277, %rd2276; setp.eq.s64 %p519, %rd2278, 0; mul.wide.u32 %rd2279, %r4635, 8; add.s64 %rd2280, %rd606, %rd2279; add.s64 %rd607, %rd2280, 1088; @%p519 bra $L__BB5_831; ld.u64 %rd2281, [%rd607]; add.s64 %rd2282, %rd606, %rd2281; and.b32 %r4638, %r732, -8; mov.u32 %r4639, 1; st.local.v2.u32 [%rd2], {%r636, %r4638}; and.b32 %r4640, %r733, -8; st.local.u32 [%rd2+8], %r4640; st.local.u64 [%rd2+48], %rd2282; shl.b32 %r4641, %r635, 6; and.b32 %r4642, %r4641, 448; shl.b32 %r4643, %r732, 3; and.b32 %r4644, %r4643, 56; or.b32 %r4645, %r4642, %r4644; and.b32 %r4646, %r733, 7; or.b32 %r4647, %r4645, %r4646; ld.u8 %rs449, [%rd2282+15]; shr.u16 %rs450, %rs449, 5; mov.u32 %r4648, 5; cvt.u32.u16 %r4649, %rs450; sub.s32 %r4650, %r4648, %r4649; shr.u32 %r4651, %r4647, %r4650; mul.wide.u32 %rd2283, %r4651, 4; add.s64 %rd2284, %rd2282, %rd2283; mov.u32 %r4652, 32; shr.u32 %r4653, %r4652, %r4649; add.s32 %r4654, %r4653, -1; and.b32 %r4655, %r4654, %r4647; shl.b32 %r4656, %r4655, %r4649; ld.u32 %r4657, [%rd2284+96]; shr.u32 %r4658, %r4657, %r4656; shl.b32 %r4659, %r4639, %r4649; mov.u32 %r4660, -1; shl.b32 %r4661, %r4660, %r4659; not.b32 %r4662, %r4661; and.b32 %r4663, %r4658, %r4662; cvt.rn.f32.u32 %f897, %r4663; ld.v2.f32 {%f898, %f899}, [%rd2282+80]; fma.rn.ftz.f32 %f952, %f899, %f897, %f898; bra.uni $L__BB5_833; $L__BB5_807: and.b32 %r780, %r635, -4096; ld.local.u32 %r4526, [%rd2+24]; setp.ne.s32 %p507, %r780, %r4526; @%p507 bra $L__BB5_810; and.b32 %r4527, %r732, -4096; ld.local.u32 %r4528, [%rd2+28]; setp.ne.s32 %p508, %r4527, %r4528; @%p508 bra $L__BB5_810; and.b32 %r4529, %r733, -4096; ld.local.u32 %r4530, [%rd2+32]; setp.eq.s32 %p509, %r4529, %r4530; @%p509 bra $L__BB5_824; bra.uni $L__BB5_810; $L__BB5_824: shl.b32 %r4583, %r635, 3; and.b32 %r4584, %r4583, 31744; and.b32 %r4585, %r732, 3968; shr.u32 %r4586, %r4585, 2; or.b32 %r4587, %r4584, %r4586; bfe.u32 %r4588, %r733, 7, 5; or.b32 %r4589, %r4587, %r4588; shr.u32 %r4590, %r4587, 6; ld.local.u64 %rd602, [%rd2+64]; mul.wide.u32 %rd2252, %r4590, 8; add.s64 %rd2253, %rd602, %rd2252; and.b32 %r4591, %r4589, 63; mov.u64 %rd2254, 1; shl.b64 %rd2255, %rd2254, %r4591; ld.u64 %rd2256, [%rd2253+4128]; and.b64 %rd2257, %rd2256, %rd2255; setp.eq.s64 %p517, %rd2257, 0; mul.wide.u32 %rd2258, %r4589, 8; add.s64 %rd2259, %rd602, %rd2258; add.s64 %rd603, %rd2259, 8256; @%p517 bra $L__BB5_828; ld.u64 %rd2260, [%rd603]; add.s64 %rd604, %rd602, %rd2260; st.local.u32 [%rd2+12], %r779; and.b32 %r4592, %r732, -128; st.local.u32 [%rd2+16], %r4592; and.b32 %r4593, %r733, -128; st.local.u32 [%rd2+20], %r4593; st.local.u64 [%rd2+56], %rd604; shl.b32 %r4594, %r635, 5; and.b32 %r4595, %r4594, 3840; shl.b32 %r4596, %r732, 1; and.b32 %r4597, %r4596, 240; or.b32 %r4598, %r4595, %r4597; bfe.u32 %r4599, %r733, 3, 4; or.b32 %r4600, %r4598, %r4599; shr.u32 %r4601, %r4598, 6; mul.wide.u32 %rd2261, %r4601, 8; add.s64 %rd2262, %rd604, %rd2261; and.b32 %r4602, %r4600, 63; shl.b64 %rd2264, %rd2254, %r4602; ld.u64 %rd2265, [%rd2262+544]; and.b64 %rd2266, %rd2265, %rd2264; setp.eq.s64 %p518, %rd2266, 0; mul.wide.u32 %rd2267, %r4600, 8; add.s64 %rd2268, %rd604, %rd2267; add.s64 %rd605, %rd2268, 1088; @%p518 bra $L__BB5_827; ld.u64 %rd2269, [%rd605]; add.s64 %rd2270, %rd604, %rd2269; and.b32 %r4603, %r732, -8; mov.u32 %r4604, 1; st.local.v2.u32 [%rd2], {%r636, %r4603}; and.b32 %r4605, %r733, -8; st.local.u32 [%rd2+8], %r4605; st.local.u64 [%rd2+48], %rd2270; shl.b32 %r4606, %r635, 6; and.b32 %r4607, %r4606, 448; shl.b32 %r4608, %r732, 3; and.b32 %r4609, %r4608, 56; or.b32 %r4610, %r4607, %r4609; and.b32 %r4611, %r733, 7; or.b32 %r4612, %r4610, %r4611; ld.u8 %rs447, [%rd2270+15]; shr.u16 %rs448, %rs447, 5; mov.u32 %r4613, 5; cvt.u32.u16 %r4614, %rs448; sub.s32 %r4615, %r4613, %r4614; shr.u32 %r4616, %r4612, %r4615; mul.wide.u32 %rd2271, %r4616, 4; add.s64 %rd2272, %rd2270, %rd2271; mov.u32 %r4617, 32; shr.u32 %r4618, %r4617, %r4614; add.s32 %r4619, %r4618, -1; and.b32 %r4620, %r4619, %r4612; shl.b32 %r4621, %r4620, %r4614; ld.u32 %r4622, [%rd2272+96]; shr.u32 %r4623, %r4622, %r4621; shl.b32 %r4624, %r4604, %r4614; mov.u32 %r4625, -1; shl.b32 %r4626, %r4625, %r4624; not.b32 %r4627, %r4626; and.b32 %r4628, %r4623, %r4627; cvt.rn.f32.u32 %f892, %r4628; ld.v2.f32 {%f893, %f894}, [%rd2270+80]; fma.rn.ftz.f32 %f952, %f894, %f892, %f893; bra.uni $L__BB5_833; $L__BB5_810: ld.local.u64 %rd590, [%rd2+40]; mov.u64 %rd2394, 0; ld.u32 %r781, [%rd590+24]; setp.eq.s32 %p510, %r781, 0; @%p510 bra $L__BB5_815; add.s64 %rd591, %rd590, 64; shr.u32 %r4532, %r635, 12; cvt.u64.u32 %rd2221, %r4532; shl.b64 %rd2222, %rd2221, 42; shr.u32 %r4533, %r732, 12; mul.wide.u32 %rd2223, %r4533, 2097152; shr.u32 %r4534, %r733, 12; cvt.u64.u32 %rd2224, %r4534; or.b64 %rd2225, %rd2223, %rd2224; or.b64 %rd592, %rd2225, %rd2222; mov.u32 %r5207, 0; $L__BB5_812: cvt.u64.u32 %rd593, %r5207; mul.wide.u32 %rd2226, %r5207, 32; add.s64 %rd2227, %rd591, %rd2226; ld.u64 %rd2228, [%rd2227]; setp.eq.s64 %p511, %rd2228, %rd592; add.s32 %r5207, %r5207, 1; @%p511 bra $L__BB5_814; setp.lt.u32 %p512, %r5207, %r781; @%p512 bra $L__BB5_812; bra.uni $L__BB5_815; $L__BB5_36: and.b32 %r4754, %r7, -128; setp.ne.s32 %p21, %r4754, %r4870; mov.u16 %rs454, 0; @%p21 bra $L__BB5_39; and.b32 %r1054, %r8, -128; setp.ne.s32 %p22, %r1054, %r4869; @%p22 bra $L__BB5_39; and.b32 %r1055, %r74, -128; setp.eq.s32 %p23, %r1055, %r4868; selp.u16 %rs454, 1, 0, %p23; $L__BB5_39: setp.eq.s16 %p24, %rs454, 0; @%p24 bra $L__BB5_43; shl.b32 %r1056, %r7, 5; and.b32 %r1057, %r1056, 3840; shl.b32 %r1058, %r8, 1; and.b32 %r1059, %r1058, 240; or.b32 %r1060, %r1059, %r1057; bfe.u32 %r1061, %r74, 3, 4; or.b32 %r1062, %r1061, %r1060; shr.u32 %r1063, %r1060, 6; mul.wide.u32 %rd700, %r1063, 8; add.s64 %rd701, %rd2311, %rd700; and.b32 %r1064, %r1062, 63; mov.u64 %rd702, 1; shl.b64 %rd703, %rd702, %r1064; ld.u64 %rd704, [%rd701+544]; and.b64 %rd705, %rd704, %rd703; setp.eq.s64 %p25, %rd705, 0; mul.wide.u32 %rd706, %r1062, 8; add.s64 %rd707, %rd2311, %rd706; add.s64 %rd43, %rd707, 1088; @%p25 bra $L__BB5_42; and.b32 %r4827, %r7, -8; ld.u64 %rd708, [%rd43]; add.s64 %rd2307, %rd2311, %rd708; and.b32 %r1065, %r8, -8; mov.u32 %r1066, 1; st.local.v2.u32 [%rd2], {%r4827, %r1065}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1067, %r7, 6; and.b32 %r1068, %r1067, 448; shl.b32 %r1069, %r8, 3; and.b32 %r1070, %r1069, 56; or.b32 %r1071, %r1070, %r1068; and.b32 %r1072, %r74, 7; or.b32 %r1073, %r1071, %r1072; ld.u8 %rs143, [%rd2307+15]; shr.u16 %rs144, %rs143, 5; mov.u32 %r1074, 5; cvt.u32.u16 %r1075, %rs144; sub.s32 %r1076, %r1074, %r1075; shr.u32 %r1077, %r1073, %r1076; mul.wide.u32 %rd709, %r1077, 4; add.s64 %rd710, %rd2307, %rd709; mov.u32 %r1078, 32; shr.u32 %r1079, %r1078, %r1075; add.s32 %r1080, %r1079, -1; and.b32 %r1081, %r1080, %r1073; shl.b32 %r1082, %r1081, %r1075; ld.u32 %r1083, [%rd710+96]; shr.u32 %r1084, %r1083, %r1082; shl.b32 %r1085, %r1066, %r1075; mov.u32 %r1086, -1; shl.b32 %r1087, %r1086, %r1085; not.b32 %r1088, %r1087; and.b32 %r1089, %r1084, %r1088; cvt.rn.f32.u32 %f409, %r1089; ld.v2.f32 {%f410, %f411}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f411, %f409, %f410; mov.b64 %rd711, {%r4827, %r1065}; shr.u64 %rd712, %rd711, 32; cvt.u32.u64 %r4782, %rd712; bra.uni $L__BB5_66; $L__BB5_71: and.b32 %r4752, %r7, -128; setp.ne.s32 %p43, %r4752, %r4870; mov.u16 %rs457, 0; @%p43 bra $L__BB5_74; and.b32 %r1207, %r139, -128; setp.ne.s32 %p44, %r1207, %r4869; @%p44 bra $L__BB5_74; and.b32 %r1208, %r74, -128; setp.eq.s32 %p45, %r1208, %r4868; selp.u16 %rs457, 1, 0, %p45; $L__BB5_74: setp.eq.s16 %p46, %rs457, 0; @%p46 bra $L__BB5_78; shl.b32 %r1209, %r7, 5; and.b32 %r1210, %r1209, 3840; shl.b32 %r1211, %r139, 1; and.b32 %r1212, %r1211, 240; or.b32 %r1213, %r1212, %r1210; bfe.u32 %r1214, %r74, 3, 4; or.b32 %r1215, %r1213, %r1214; shr.u32 %r1216, %r1213, 6; mul.wide.u32 %rd769, %r1216, 8; add.s64 %rd770, %rd2311, %rd769; and.b32 %r1217, %r1215, 63; mov.u64 %rd771, 1; shl.b64 %rd772, %rd771, %r1217; ld.u64 %rd773, [%rd770+544]; and.b64 %rd774, %rd773, %rd772; setp.eq.s64 %p47, %rd774, 0; mul.wide.u32 %rd775, %r1215, 8; add.s64 %rd776, %rd2311, %rd775; add.s64 %rd72, %rd776, 1088; @%p47 bra $L__BB5_77; and.b32 %r4827, %r7, -8; ld.u64 %rd777, [%rd72]; add.s64 %rd2307, %rd2311, %rd777; and.b32 %r1218, %r139, -8; mov.u32 %r1219, 1; st.local.v2.u32 [%rd2], {%r4827, %r1218}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1220, %r7, 6; and.b32 %r1221, %r1220, 448; shl.b32 %r1222, %r139, 3; and.b32 %r1223, %r1222, 56; or.b32 %r1224, %r1223, %r1221; and.b32 %r1225, %r74, 7; or.b32 %r1226, %r1224, %r1225; ld.u8 %rs157, [%rd2307+15]; shr.u16 %rs158, %rs157, 5; mov.u32 %r1227, 5; cvt.u32.u16 %r1228, %rs158; sub.s32 %r1229, %r1227, %r1228; shr.u32 %r1230, %r1226, %r1229; mul.wide.u32 %rd778, %r1230, 4; add.s64 %rd779, %rd2307, %rd778; mov.u32 %r1231, 32; shr.u32 %r1232, %r1231, %r1228; add.s32 %r1233, %r1232, -1; and.b32 %r1234, %r1233, %r1226; shl.b32 %r1235, %r1234, %r1228; ld.u32 %r1236, [%rd779+96]; shr.u32 %r1237, %r1236, %r1235; shl.b32 %r1238, %r1219, %r1228; mov.u32 %r1239, -1; shl.b32 %r1240, %r1239, %r1238; not.b32 %r1241, %r1240; and.b32 %r1242, %r1237, %r1241; cvt.rn.f32.u32 %f430, %r1242; ld.v2.f32 {%f431, %f432}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f432, %f430, %f431; mov.b64 %rd780, {%r4827, %r1218}; shr.u64 %rd781, %rd780, 32; cvt.u32.u64 %r4782, %rd781; bra.uni $L__BB5_101; $L__BB5_106: and.b32 %r4725, %r7, -128; setp.ne.s32 %p65, %r4725, %r4870; mov.u16 %rs460, 0; @%p65 bra $L__BB5_109; and.b32 %r1360, %r139, -128; setp.ne.s32 %p66, %r1360, %r4869; @%p66 bra $L__BB5_109; and.b32 %r1361, %r9, -128; setp.eq.s32 %p67, %r1361, %r4868; selp.u16 %rs460, 1, 0, %p67; $L__BB5_109: setp.eq.s16 %p68, %rs460, 0; @%p68 bra $L__BB5_113; shl.b32 %r1362, %r7, 5; and.b32 %r1363, %r1362, 3840; shl.b32 %r1364, %r139, 1; and.b32 %r1365, %r1364, 240; or.b32 %r1366, %r1365, %r1363; bfe.u32 %r1367, %r9, 3, 4; or.b32 %r1368, %r1366, %r1367; shr.u32 %r1369, %r1366, 6; mul.wide.u32 %rd839, %r1369, 8; add.s64 %rd840, %rd2311, %rd839; and.b32 %r1370, %r1368, 63; mov.u64 %rd841, 1; shl.b64 %rd842, %rd841, %r1370; ld.u64 %rd843, [%rd840+544]; and.b64 %rd844, %rd843, %rd842; setp.eq.s64 %p69, %rd844, 0; mul.wide.u32 %rd845, %r1368, 8; add.s64 %rd846, %rd2311, %rd845; add.s64 %rd101, %rd846, 1088; @%p69 bra $L__BB5_112; and.b32 %r4827, %r7, -8; ld.u64 %rd847, [%rd101]; add.s64 %rd2307, %rd2311, %rd847; and.b32 %r1371, %r139, -8; mov.u32 %r1372, 1; st.local.v2.u32 [%rd2], {%r4827, %r1371}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1373, %r7, 7; shl.b32 %r1374, %r139, 3; and.b32 %r1375, %r1374, 56; and.b32 %r1376, %r9, 7; bfi.b32 %r1377, %r1373, %r1376, 6, 3; or.b32 %r1378, %r1377, %r1375; ld.u8 %rs171, [%rd2307+15]; shr.u16 %rs172, %rs171, 5; mov.u32 %r1379, 5; cvt.u32.u16 %r1380, %rs172; sub.s32 %r1381, %r1379, %r1380; shr.u32 %r1382, %r1378, %r1381; mul.wide.u32 %rd848, %r1382, 4; add.s64 %rd849, %rd2307, %rd848; mov.u32 %r1383, 32; shr.u32 %r1384, %r1383, %r1380; add.s32 %r1385, %r1384, -1; and.b32 %r1386, %r1385, %r1378; shl.b32 %r1387, %r1386, %r1380; ld.u32 %r1388, [%rd849+96]; shr.u32 %r1389, %r1388, %r1387; shl.b32 %r1390, %r1372, %r1380; mov.u32 %r1391, -1; shl.b32 %r1392, %r1391, %r1390; not.b32 %r1393, %r1392; and.b32 %r1394, %r1389, %r1393; cvt.rn.f32.u32 %f450, %r1394; ld.v2.f32 {%f451, %f452}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f452, %f450, %f451; mov.b64 %rd850, {%r4827, %r1371}; shr.u64 %rd851, %rd850, 32; cvt.u32.u64 %r4782, %rd851; bra.uni $L__BB5_136; $L__BB5_141: and.b32 %r261, %r259, -128; setp.ne.s32 %p87, %r261, %r4870; mov.u16 %rs463, 0; @%p87 bra $L__BB5_144; and.b32 %r1509, %r139, -128; setp.ne.s32 %p88, %r1509, %r4869; @%p88 bra $L__BB5_144; and.b32 %r1510, %r9, -128; setp.eq.s32 %p89, %r1510, %r4868; selp.u16 %rs463, 1, 0, %p89; $L__BB5_144: setp.eq.s16 %p90, %rs463, 0; @%p90 bra $L__BB5_148; shl.b32 %r1511, %r259, 5; and.b32 %r1512, %r1511, 3840; shl.b32 %r1513, %r139, 1; and.b32 %r1514, %r1513, 240; or.b32 %r1515, %r1514, %r1512; bfe.u32 %r1516, %r9, 3, 4; or.b32 %r1517, %r1515, %r1516; shr.u32 %r1518, %r1515, 6; mul.wide.u32 %rd907, %r1518, 8; add.s64 %rd908, %rd2311, %rd907; and.b32 %r1519, %r1517, 63; mov.u64 %rd909, 1; shl.b64 %rd910, %rd909, %r1519; ld.u64 %rd911, [%rd908+544]; and.b64 %rd912, %rd911, %rd910; setp.eq.s64 %p91, %rd912, 0; mul.wide.u32 %rd913, %r1517, 8; add.s64 %rd914, %rd2311, %rd913; add.s64 %rd130, %rd914, 1088; @%p91 bra $L__BB5_147; and.b32 %r4827, %r259, -8; ld.u64 %rd915, [%rd130]; add.s64 %rd2307, %rd2311, %rd915; and.b32 %r1520, %r139, -8; mov.u32 %r1521, 1; st.local.v2.u32 [%rd2], {%r4827, %r1520}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1522, %r259, 7; shl.b32 %r1523, %r139, 3; and.b32 %r1524, %r1523, 56; and.b32 %r1525, %r9, 7; bfi.b32 %r1526, %r1522, %r1525, 6, 3; or.b32 %r1527, %r1526, %r1524; ld.u8 %rs185, [%rd2307+15]; shr.u16 %rs186, %rs185, 5; mov.u32 %r1528, 5; cvt.u32.u16 %r1529, %rs186; sub.s32 %r1530, %r1528, %r1529; shr.u32 %r1531, %r1527, %r1530; mul.wide.u32 %rd916, %r1531, 4; add.s64 %rd917, %rd2307, %rd916; mov.u32 %r1532, 32; shr.u32 %r1533, %r1532, %r1529; add.s32 %r1534, %r1533, -1; and.b32 %r1535, %r1534, %r1527; shl.b32 %r1536, %r1535, %r1529; ld.u32 %r1537, [%rd917+96]; shr.u32 %r1538, %r1537, %r1536; shl.b32 %r1539, %r1521, %r1529; mov.u32 %r1540, -1; shl.b32 %r1541, %r1540, %r1539; not.b32 %r1542, %r1541; and.b32 %r1543, %r1538, %r1542; cvt.rn.f32.u32 %f473, %r1543; ld.v2.f32 {%f474, %f475}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f475, %f473, %f474; mov.b64 %rd918, {%r4827, %r1520}; shr.u64 %rd919, %rd918, 32; cvt.u32.u64 %r4782, %rd919; bra.uni $L__BB5_171; $L__BB5_176: and.b32 %r318, %r259, -128; setp.ne.s32 %p109, %r318, %r4870; mov.u16 %rs466, 0; @%p109 bra $L__BB5_179; and.b32 %r1660, %r139, -128; setp.ne.s32 %p110, %r1660, %r4869; @%p110 bra $L__BB5_179; and.b32 %r1661, %r74, -128; setp.eq.s32 %p111, %r1661, %r4868; selp.u16 %rs466, 1, 0, %p111; $L__BB5_179: setp.eq.s16 %p112, %rs466, 0; @%p112 bra $L__BB5_183; shl.b32 %r1662, %r259, 5; and.b32 %r1663, %r1662, 3840; shl.b32 %r1664, %r139, 1; and.b32 %r1665, %r1664, 240; or.b32 %r1666, %r1665, %r1663; bfe.u32 %r1667, %r74, 3, 4; or.b32 %r1668, %r1666, %r1667; shr.u32 %r1669, %r1666, 6; mul.wide.u32 %rd978, %r1669, 8; add.s64 %rd979, %rd2311, %rd978; and.b32 %r1670, %r1668, 63; mov.u64 %rd980, 1; shl.b64 %rd981, %rd980, %r1670; ld.u64 %rd982, [%rd979+544]; and.b64 %rd983, %rd982, %rd981; setp.eq.s64 %p113, %rd983, 0; mul.wide.u32 %rd984, %r1668, 8; add.s64 %rd985, %rd2311, %rd984; add.s64 %rd159, %rd985, 1088; @%p113 bra $L__BB5_182; and.b32 %r4827, %r259, -8; ld.u64 %rd986, [%rd159]; add.s64 %rd2307, %rd2311, %rd986; and.b32 %r1671, %r139, -8; mov.u32 %r1672, 1; st.local.v2.u32 [%rd2], {%r4827, %r1671}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1673, %r259, 6; and.b32 %r1674, %r1673, 448; shl.b32 %r1675, %r139, 3; and.b32 %r1676, %r1675, 56; or.b32 %r1677, %r1676, %r1674; and.b32 %r1678, %r74, 7; or.b32 %r1679, %r1677, %r1678; ld.u8 %rs199, [%rd2307+15]; shr.u16 %rs200, %rs199, 5; mov.u32 %r1680, 5; cvt.u32.u16 %r1681, %rs200; sub.s32 %r1682, %r1680, %r1681; shr.u32 %r1683, %r1679, %r1682; mul.wide.u32 %rd987, %r1683, 4; add.s64 %rd988, %rd2307, %rd987; mov.u32 %r1684, 32; shr.u32 %r1685, %r1684, %r1681; add.s32 %r1686, %r1685, -1; and.b32 %r1687, %r1686, %r1679; shl.b32 %r1688, %r1687, %r1681; ld.u32 %r1689, [%rd988+96]; shr.u32 %r1690, %r1689, %r1688; shl.b32 %r1691, %r1672, %r1681; mov.u32 %r1692, -1; shl.b32 %r1693, %r1692, %r1691; not.b32 %r1694, %r1693; and.b32 %r1695, %r1690, %r1694; cvt.rn.f32.u32 %f493, %r1695; ld.v2.f32 {%f494, %f495}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f495, %f493, %f494; mov.b64 %rd989, {%r4827, %r1671}; shr.u64 %rd990, %rd989, 32; cvt.u32.u64 %r4782, %rd990; bra.uni $L__BB5_206; $L__BB5_211: and.b32 %r372, %r259, -128; setp.ne.s32 %p131, %r372, %r4870; mov.u16 %rs469, 0; @%p131 bra $L__BB5_214; and.b32 %r1815, %r8, -128; setp.ne.s32 %p132, %r1815, %r4869; @%p132 bra $L__BB5_214; and.b32 %r1816, %r74, -128; setp.eq.s32 %p133, %r1816, %r4868; selp.u16 %rs469, 1, 0, %p133; $L__BB5_214: setp.eq.s16 %p134, %rs469, 0; @%p134 bra $L__BB5_218; shl.b32 %r1817, %r259, 5; and.b32 %r1818, %r1817, 3840; shl.b32 %r1819, %r8, 1; and.b32 %r1820, %r1819, 240; or.b32 %r1821, %r1818, %r1820; bfe.u32 %r1822, %r74, 3, 4; or.b32 %r1823, %r1822, %r1821; shr.u32 %r1824, %r1821, 6; mul.wide.u32 %rd1049, %r1824, 8; add.s64 %rd1050, %rd2311, %rd1049; and.b32 %r1825, %r1823, 63; mov.u64 %rd1051, 1; shl.b64 %rd1052, %rd1051, %r1825; ld.u64 %rd1053, [%rd1050+544]; and.b64 %rd1054, %rd1053, %rd1052; setp.eq.s64 %p135, %rd1054, 0; mul.wide.u32 %rd1055, %r1823, 8; add.s64 %rd1056, %rd2311, %rd1055; add.s64 %rd188, %rd1056, 1088; @%p135 bra $L__BB5_217; and.b32 %r4827, %r259, -8; ld.u64 %rd1057, [%rd188]; add.s64 %rd2307, %rd2311, %rd1057; and.b32 %r1826, %r8, -8; mov.u32 %r1827, 1; st.local.v2.u32 [%rd2], {%r4827, %r1826}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1828, %r259, 6; and.b32 %r1829, %r1828, 448; shl.b32 %r1830, %r8, 3; and.b32 %r1831, %r1830, 56; or.b32 %r1832, %r1829, %r1831; and.b32 %r1833, %r74, 7; or.b32 %r1834, %r1832, %r1833; ld.u8 %rs213, [%rd2307+15]; shr.u16 %rs214, %rs213, 5; mov.u32 %r1835, 5; cvt.u32.u16 %r1836, %rs214; sub.s32 %r1837, %r1835, %r1836; shr.u32 %r1838, %r1834, %r1837; mul.wide.u32 %rd1058, %r1838, 4; add.s64 %rd1059, %rd2307, %rd1058; mov.u32 %r1839, 32; shr.u32 %r1840, %r1839, %r1836; add.s32 %r1841, %r1840, -1; and.b32 %r1842, %r1841, %r1834; shl.b32 %r1843, %r1842, %r1836; ld.u32 %r1844, [%rd1059+96]; shr.u32 %r1845, %r1844, %r1843; shl.b32 %r1846, %r1827, %r1836; mov.u32 %r1847, -1; shl.b32 %r1848, %r1847, %r1846; not.b32 %r1849, %r1848; and.b32 %r1850, %r1845, %r1849; cvt.rn.f32.u32 %f514, %r1850; ld.v2.f32 {%f515, %f516}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f516, %f514, %f515; mov.b64 %rd1060, {%r4827, %r1826}; shr.u64 %rd1061, %rd1060, 32; cvt.u32.u64 %r4782, %rd1061; bra.uni $L__BB5_241; $L__BB5_246: and.b32 %r423, %r259, -128; setp.ne.s32 %p153, %r423, %r4870; mov.u16 %rs472, 0; @%p153 bra $L__BB5_249; and.b32 %r1969, %r8, -128; setp.ne.s32 %p154, %r1969, %r4869; @%p154 bra $L__BB5_249; and.b32 %r1970, %r9, -128; setp.eq.s32 %p155, %r1970, %r4868; selp.u16 %rs472, 1, 0, %p155; $L__BB5_249: setp.eq.s16 %p156, %rs472, 0; @%p156 bra $L__BB5_253; shl.b32 %r1971, %r259, 5; and.b32 %r1972, %r1971, 3840; shl.b32 %r1973, %r8, 1; and.b32 %r1974, %r1973, 240; or.b32 %r1975, %r1972, %r1974; bfe.u32 %r1976, %r9, 3, 4; or.b32 %r1977, %r1975, %r1976; shr.u32 %r1978, %r1975, 6; mul.wide.u32 %rd1120, %r1978, 8; add.s64 %rd1121, %rd2311, %rd1120; and.b32 %r1979, %r1977, 63; mov.u64 %rd1122, 1; shl.b64 %rd1123, %rd1122, %r1979; ld.u64 %rd1124, [%rd1121+544]; and.b64 %rd1125, %rd1124, %rd1123; setp.eq.s64 %p157, %rd1125, 0; mul.wide.u32 %rd1126, %r1977, 8; add.s64 %rd1127, %rd2311, %rd1126; add.s64 %rd217, %rd1127, 1088; @%p157 bra $L__BB5_252; and.b32 %r4827, %r259, -8; ld.u64 %rd1128, [%rd217]; add.s64 %rd2307, %rd2311, %rd1128; and.b32 %r1980, %r8, -8; mov.u32 %r1981, 1; st.local.v2.u32 [%rd2], {%r4827, %r1980}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1982, %r259, 6; and.b32 %r1983, %r1982, 448; shl.b32 %r1984, %r8, 3; and.b32 %r1985, %r1984, 56; or.b32 %r1986, %r1983, %r1985; and.b32 %r1987, %r9, 7; or.b32 %r1988, %r1986, %r1987; ld.u8 %rs227, [%rd2307+15]; shr.u16 %rs228, %rs227, 5; mov.u32 %r1989, 5; cvt.u32.u16 %r1990, %rs228; sub.s32 %r1991, %r1989, %r1990; shr.u32 %r1992, %r1988, %r1991; mul.wide.u32 %rd1129, %r1992, 4; add.s64 %rd1130, %rd2307, %rd1129; mov.u32 %r1993, 32; shr.u32 %r1994, %r1993, %r1990; add.s32 %r1995, %r1994, -1; and.b32 %r1996, %r1995, %r1988; shl.b32 %r1997, %r1996, %r1990; ld.u32 %r1998, [%rd1130+96]; shr.u32 %r1999, %r1998, %r1997; shl.b32 %r2000, %r1981, %r1990; mov.u32 %r2001, -1; shl.b32 %r2002, %r2001, %r2000; not.b32 %r2003, %r2002; and.b32 %r2004, %r1999, %r2003; cvt.rn.f32.u32 %f534, %r2004; ld.v2.f32 {%f535, %f536}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f536, %f534, %f535; mov.b64 %rd1131, {%r4827, %r1980}; shr.u64 %rd1132, %rd1131, 32; cvt.u32.u64 %r4782, %rd1132; bra.uni $L__BB5_276; $L__BB5_281: and.b32 %r473, %r471, -128; setp.ne.s32 %p175, %r473, %r4870; mov.u16 %rs475, 0; @%p175 bra $L__BB5_284; and.b32 %r2122, %r8, -128; setp.ne.s32 %p176, %r2122, %r4869; @%p176 bra $L__BB5_284; and.b32 %r2123, %r9, -128; setp.eq.s32 %p177, %r2123, %r4868; selp.u16 %rs475, 1, 0, %p177; $L__BB5_284: setp.eq.s16 %p178, %rs475, 0; @%p178 bra $L__BB5_288; shl.b32 %r2124, %r471, 5; and.b32 %r2125, %r2124, 3840; shl.b32 %r2126, %r8, 1; and.b32 %r2127, %r2126, 240; or.b32 %r2128, %r2127, %r2125; bfe.u32 %r2129, %r9, 3, 4; or.b32 %r2130, %r2128, %r2129; shr.u32 %r2131, %r2128, 6; mul.wide.u32 %rd1190, %r2131, 8; add.s64 %rd1191, %rd2311, %rd1190; and.b32 %r2132, %r2130, 63; mov.u64 %rd1192, 1; shl.b64 %rd1193, %rd1192, %r2132; ld.u64 %rd1194, [%rd1191+544]; and.b64 %rd1195, %rd1194, %rd1193; setp.eq.s64 %p179, %rd1195, 0; mul.wide.u32 %rd1196, %r2130, 8; add.s64 %rd1197, %rd2311, %rd1196; add.s64 %rd246, %rd1197, 1088; @%p179 bra $L__BB5_287; ld.u64 %rd1198, [%rd246]; add.s64 %rd2307, %rd2311, %rd1198; and.b32 %r2133, %r8, -8; mov.u32 %r2134, 1; st.local.v2.u32 [%rd2], {%r472, %r2133}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2135, %r471, 6; and.b32 %r2136, %r2135, 448; shl.b32 %r2137, %r8, 3; and.b32 %r2138, %r2137, 56; or.b32 %r2139, %r2138, %r2136; and.b32 %r2140, %r9, 7; or.b32 %r2141, %r2139, %r2140; ld.u8 %rs241, [%rd2307+15]; shr.u16 %rs242, %rs241, 5; mov.u32 %r2142, 5; cvt.u32.u16 %r2143, %rs242; sub.s32 %r2144, %r2142, %r2143; shr.u32 %r2145, %r2141, %r2144; mul.wide.u32 %rd1199, %r2145, 4; add.s64 %rd1200, %rd2307, %rd1199; mov.u32 %r2146, 32; shr.u32 %r2147, %r2146, %r2143; add.s32 %r2148, %r2147, -1; and.b32 %r2149, %r2148, %r2141; shl.b32 %r2150, %r2149, %r2143; ld.u32 %r2151, [%rd1200+96]; shr.u32 %r2152, %r2151, %r2150; shl.b32 %r2153, %r2134, %r2143; mov.u32 %r2154, -1; shl.b32 %r2155, %r2154, %r2153; not.b32 %r2156, %r2155; and.b32 %r2157, %r2152, %r2156; cvt.rn.f32.u32 %f561, %r2157; ld.v2.f32 {%f562, %f563}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f563, %f561, %f562; mov.b64 %rd1201, {%r472, %r2133}; shr.u64 %rd1202, %rd1201, 32; cvt.u32.u64 %r4782, %rd1202; mov.u32 %r4827, %r472; bra.uni $L__BB5_311; $L__BB5_316: and.b32 %r518, %r471, -128; setp.ne.s32 %p197, %r518, %r4870; mov.u16 %rs478, 0; @%p197 bra $L__BB5_319; and.b32 %r2275, %r8, -128; setp.ne.s32 %p198, %r2275, %r4869; @%p198 bra $L__BB5_319; and.b32 %r2276, %r74, -128; setp.eq.s32 %p199, %r2276, %r4868; selp.u16 %rs478, 1, 0, %p199; $L__BB5_319: setp.eq.s16 %p200, %rs478, 0; @%p200 bra $L__BB5_323; shl.b32 %r2277, %r471, 5; and.b32 %r2278, %r2277, 3840; shl.b32 %r2279, %r8, 1; and.b32 %r2280, %r2279, 240; or.b32 %r2281, %r2280, %r2278; bfe.u32 %r2282, %r74, 3, 4; or.b32 %r2283, %r2282, %r2281; shr.u32 %r2284, %r2281, 6; mul.wide.u32 %rd1260, %r2284, 8; add.s64 %rd1261, %rd2311, %rd1260; and.b32 %r2285, %r2283, 63; mov.u64 %rd1262, 1; shl.b64 %rd1263, %rd1262, %r2285; ld.u64 %rd1264, [%rd1261+544]; and.b64 %rd1265, %rd1264, %rd1263; setp.eq.s64 %p201, %rd1265, 0; mul.wide.u32 %rd1266, %r2283, 8; add.s64 %rd1267, %rd2311, %rd1266; add.s64 %rd275, %rd1267, 1088; @%p201 bra $L__BB5_322; ld.u64 %rd1268, [%rd275]; add.s64 %rd2307, %rd2311, %rd1268; and.b32 %r2286, %r8, -8; mov.u32 %r2287, 1; st.local.v2.u32 [%rd2], {%r472, %r2286}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2288, %r471, 6; and.b32 %r2289, %r2288, 448; shl.b32 %r2290, %r8, 3; and.b32 %r2291, %r2290, 56; or.b32 %r2292, %r2291, %r2289; and.b32 %r2293, %r74, 7; or.b32 %r2294, %r2292, %r2293; ld.u8 %rs255, [%rd2307+15]; shr.u16 %rs256, %rs255, 5; mov.u32 %r2295, 5; cvt.u32.u16 %r2296, %rs256; sub.s32 %r2297, %r2295, %r2296; shr.u32 %r2298, %r2294, %r2297; mul.wide.u32 %rd1269, %r2298, 4; add.s64 %rd1270, %rd2307, %rd1269; mov.u32 %r2299, 32; shr.u32 %r2300, %r2299, %r2296; add.s32 %r2301, %r2300, -1; and.b32 %r2302, %r2301, %r2294; shl.b32 %r2303, %r2302, %r2296; ld.u32 %r2304, [%rd1270+96]; shr.u32 %r2305, %r2304, %r2303; shl.b32 %r2306, %r2287, %r2296; mov.u32 %r2307, -1; shl.b32 %r2308, %r2307, %r2306; not.b32 %r2309, %r2308; and.b32 %r2310, %r2305, %r2309; cvt.rn.f32.u32 %f581, %r2310; ld.v2.f32 {%f582, %f583}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f583, %f581, %f582; mov.b64 %rd1271, {%r472, %r2286}; shr.u64 %rd1272, %rd1271, 32; cvt.u32.u64 %r4782, %rd1272; mov.u32 %r4827, %r472; bra.uni $L__BB5_346; $L__BB5_351: and.b32 %r560, %r471, -128; setp.ne.s32 %p219, %r560, %r4870; mov.u16 %rs481, 0; @%p219 bra $L__BB5_354; and.b32 %r2429, %r139, -128; setp.ne.s32 %p220, %r2429, %r4869; @%p220 bra $L__BB5_354; and.b32 %r2430, %r74, -128; setp.eq.s32 %p221, %r2430, %r4868; selp.u16 %rs481, 1, 0, %p221; $L__BB5_354: setp.eq.s16 %p222, %rs481, 0; @%p222 bra $L__BB5_358; shl.b32 %r2431, %r471, 5; and.b32 %r2432, %r2431, 3840; shl.b32 %r2433, %r139, 1; and.b32 %r2434, %r2433, 240; or.b32 %r2435, %r2434, %r2432; bfe.u32 %r2436, %r74, 3, 4; or.b32 %r2437, %r2435, %r2436; shr.u32 %r2438, %r2435, 6; mul.wide.u32 %rd1331, %r2438, 8; add.s64 %rd1332, %rd2311, %rd1331; and.b32 %r2439, %r2437, 63; mov.u64 %rd1333, 1; shl.b64 %rd1334, %rd1333, %r2439; ld.u64 %rd1335, [%rd1332+544]; and.b64 %rd1336, %rd1335, %rd1334; setp.eq.s64 %p223, %rd1336, 0; mul.wide.u32 %rd1337, %r2437, 8; add.s64 %rd1338, %rd2311, %rd1337; add.s64 %rd304, %rd1338, 1088; @%p223 bra $L__BB5_357; ld.u64 %rd1339, [%rd304]; add.s64 %rd2307, %rd2311, %rd1339; and.b32 %r2440, %r139, -8; mov.u32 %r2441, 1; st.local.v2.u32 [%rd2], {%r472, %r2440}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2442, %r471, 6; and.b32 %r2443, %r2442, 448; shl.b32 %r2444, %r139, 3; and.b32 %r2445, %r2444, 56; or.b32 %r2446, %r2445, %r2443; and.b32 %r2447, %r74, 7; or.b32 %r2448, %r2446, %r2447; ld.u8 %rs269, [%rd2307+15]; shr.u16 %rs270, %rs269, 5; mov.u32 %r2449, 5; cvt.u32.u16 %r2450, %rs270; sub.s32 %r2451, %r2449, %r2450; shr.u32 %r2452, %r2448, %r2451; mul.wide.u32 %rd1340, %r2452, 4; add.s64 %rd1341, %rd2307, %rd1340; mov.u32 %r2453, 32; shr.u32 %r2454, %r2453, %r2450; add.s32 %r2455, %r2454, -1; and.b32 %r2456, %r2455, %r2448; shl.b32 %r2457, %r2456, %r2450; ld.u32 %r2458, [%rd1341+96]; shr.u32 %r2459, %r2458, %r2457; shl.b32 %r2460, %r2441, %r2450; mov.u32 %r2461, -1; shl.b32 %r2462, %r2461, %r2460; not.b32 %r2463, %r2462; and.b32 %r2464, %r2459, %r2463; cvt.rn.f32.u32 %f602, %r2464; ld.v2.f32 {%f603, %f604}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f604, %f602, %f603; mov.b64 %rd1342, {%r472, %r2440}; shr.u64 %rd1343, %rd1342, 32; cvt.u32.u64 %r4782, %rd1343; mov.u32 %r4827, %r472; bra.uni $L__BB5_381; $L__BB5_386: and.b32 %r599, %r471, -128; setp.ne.s32 %p241, %r599, %r4870; mov.u16 %rs484, 0; @%p241 bra $L__BB5_389; and.b32 %r2583, %r139, -128; setp.ne.s32 %p242, %r2583, %r4869; @%p242 bra $L__BB5_389; and.b32 %r2584, %r9, -128; setp.eq.s32 %p243, %r2584, %r4868; selp.u16 %rs484, 1, 0, %p243; $L__BB5_389: setp.eq.s16 %p244, %rs484, 0; @%p244 bra $L__BB5_393; shl.b32 %r2585, %r471, 5; and.b32 %r2586, %r2585, 3840; shl.b32 %r2587, %r139, 1; and.b32 %r2588, %r2587, 240; or.b32 %r2589, %r2588, %r2586; bfe.u32 %r2590, %r9, 3, 4; or.b32 %r2591, %r2589, %r2590; shr.u32 %r2592, %r2589, 6; mul.wide.u32 %rd1402, %r2592, 8; add.s64 %rd1403, %rd2311, %rd1402; and.b32 %r2593, %r2591, 63; mov.u64 %rd1404, 1; shl.b64 %rd1405, %rd1404, %r2593; ld.u64 %rd1406, [%rd1403+544]; and.b64 %rd1407, %rd1406, %rd1405; setp.eq.s64 %p245, %rd1407, 0; mul.wide.u32 %rd1408, %r2591, 8; add.s64 %rd1409, %rd2311, %rd1408; add.s64 %rd333, %rd1409, 1088; @%p245 bra $L__BB5_392; ld.u64 %rd1410, [%rd333]; add.s64 %rd2307, %rd2311, %rd1410; and.b32 %r2594, %r139, -8; mov.u32 %r2595, 1; st.local.v2.u32 [%rd2], {%r472, %r2594}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2596, %r471, 7; shl.b32 %r2597, %r139, 3; and.b32 %r2598, %r2597, 56; and.b32 %r2599, %r9, 7; bfi.b32 %r2600, %r2596, %r2599, 6, 3; or.b32 %r2601, %r2600, %r2598; ld.u8 %rs283, [%rd2307+15]; shr.u16 %rs284, %rs283, 5; mov.u32 %r2602, 5; cvt.u32.u16 %r2603, %rs284; sub.s32 %r2604, %r2602, %r2603; shr.u32 %r2605, %r2601, %r2604; mul.wide.u32 %rd1411, %r2605, 4; add.s64 %rd1412, %rd2307, %rd1411; mov.u32 %r2606, 32; shr.u32 %r2607, %r2606, %r2603; add.s32 %r2608, %r2607, -1; and.b32 %r2609, %r2608, %r2601; shl.b32 %r2610, %r2609, %r2603; ld.u32 %r2611, [%rd1412+96]; shr.u32 %r2612, %r2611, %r2610; shl.b32 %r2613, %r2595, %r2603; mov.u32 %r2614, -1; shl.b32 %r2615, %r2614, %r2613; not.b32 %r2616, %r2615; and.b32 %r2617, %r2612, %r2616; cvt.rn.f32.u32 %f622, %r2617; ld.v2.f32 {%f623, %f624}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f624, %f622, %f623; mov.b64 %rd1413, {%r472, %r2594}; shr.u64 %rd1414, %rd1413, 32; cvt.u32.u64 %r4782, %rd1414; mov.u32 %r4827, %r472; bra.uni $L__BB5_416; $L__BB5_421: and.b32 %r637, %r635, -128; setp.ne.s32 %p263, %r637, %r4870; mov.u16 %rs487, 0; @%p263 bra $L__BB5_424; and.b32 %r2733, %r139, -128; setp.ne.s32 %p264, %r2733, %r4869; @%p264 bra $L__BB5_424; and.b32 %r2734, %r9, -128; setp.eq.s32 %p265, %r2734, %r4868; selp.u16 %rs487, 1, 0, %p265; $L__BB5_424: setp.eq.s16 %p266, %rs487, 0; @%p266 bra $L__BB5_428; shl.b32 %r2735, %r635, 5; and.b32 %r2736, %r2735, 3840; shl.b32 %r2737, %r139, 1; and.b32 %r2738, %r2737, 240; or.b32 %r2739, %r2738, %r2736; bfe.u32 %r2740, %r9, 3, 4; or.b32 %r2741, %r2739, %r2740; shr.u32 %r2742, %r2739, 6; mul.wide.u32 %rd1473, %r2742, 8; add.s64 %rd1474, %rd2311, %rd1473; and.b32 %r2743, %r2741, 63; mov.u64 %rd1475, 1; shl.b64 %rd1476, %rd1475, %r2743; ld.u64 %rd1477, [%rd1474+544]; and.b64 %rd1478, %rd1477, %rd1476; setp.eq.s64 %p267, %rd1478, 0; mul.wide.u32 %rd1479, %r2741, 8; add.s64 %rd1480, %rd2311, %rd1479; add.s64 %rd362, %rd1480, 1088; @%p267 bra $L__BB5_427; ld.u64 %rd1481, [%rd362]; add.s64 %rd2307, %rd2311, %rd1481; and.b32 %r2744, %r139, -8; mov.u32 %r2745, 1; st.local.v2.u32 [%rd2], {%r636, %r2744}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2746, %r635, 7; shl.b32 %r2747, %r139, 3; and.b32 %r2748, %r2747, 56; and.b32 %r2749, %r9, 7; bfi.b32 %r2750, %r2746, %r2749, 6, 3; or.b32 %r2751, %r2750, %r2748; ld.u8 %rs297, [%rd2307+15]; shr.u16 %rs298, %rs297, 5; mov.u32 %r2752, 5; cvt.u32.u16 %r2753, %rs298; sub.s32 %r2754, %r2752, %r2753; shr.u32 %r2755, %r2751, %r2754; mul.wide.u32 %rd1482, %r2755, 4; add.s64 %rd1483, %rd2307, %rd1482; mov.u32 %r2756, 32; shr.u32 %r2757, %r2756, %r2753; add.s32 %r2758, %r2757, -1; and.b32 %r2759, %r2758, %r2751; shl.b32 %r2760, %r2759, %r2753; ld.u32 %r2761, [%rd1483+96]; shr.u32 %r2762, %r2761, %r2760; shl.b32 %r2763, %r2745, %r2753; mov.u32 %r2764, -1; shl.b32 %r2765, %r2764, %r2763; not.b32 %r2766, %r2765; and.b32 %r2767, %r2762, %r2766; cvt.rn.f32.u32 %f645, %r2767; ld.v2.f32 {%f646, %f647}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f647, %f645, %f646; mov.b64 %rd1484, {%r636, %r2744}; shr.u64 %rd1485, %rd1484, 32; cvt.u32.u64 %r4782, %rd1485; mov.u32 %r4827, %r636; bra.uni $L__BB5_451; $L__BB5_456: and.b32 %r670, %r635, -128; setp.ne.s32 %p285, %r670, %r4870; mov.u16 %rs490, 0; @%p285 bra $L__BB5_459; and.b32 %r2884, %r139, -128; setp.ne.s32 %p286, %r2884, %r4869; @%p286 bra $L__BB5_459; and.b32 %r2885, %r74, -128; setp.eq.s32 %p287, %r2885, %r4868; selp.u16 %rs490, 1, 0, %p287; $L__BB5_459: setp.eq.s16 %p288, %rs490, 0; @%p288 bra $L__BB5_463; shl.b32 %r2886, %r635, 5; and.b32 %r2887, %r2886, 3840; shl.b32 %r2888, %r139, 1; and.b32 %r2889, %r2888, 240; or.b32 %r2890, %r2889, %r2887; bfe.u32 %r2891, %r74, 3, 4; or.b32 %r2892, %r2890, %r2891; shr.u32 %r2893, %r2890, 6; mul.wide.u32 %rd1544, %r2893, 8; add.s64 %rd1545, %rd2311, %rd1544; and.b32 %r2894, %r2892, 63; mov.u64 %rd1546, 1; shl.b64 %rd1547, %rd1546, %r2894; ld.u64 %rd1548, [%rd1545+544]; and.b64 %rd1549, %rd1548, %rd1547; setp.eq.s64 %p289, %rd1549, 0; mul.wide.u32 %rd1550, %r2892, 8; add.s64 %rd1551, %rd2311, %rd1550; add.s64 %rd391, %rd1551, 1088; @%p289 bra $L__BB5_462; ld.u64 %rd1552, [%rd391]; add.s64 %rd2307, %rd2311, %rd1552; and.b32 %r2895, %r139, -8; mov.u32 %r2896, 1; st.local.v2.u32 [%rd2], {%r636, %r2895}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2897, %r635, 6; and.b32 %r2898, %r2897, 448; shl.b32 %r2899, %r139, 3; and.b32 %r2900, %r2899, 56; or.b32 %r2901, %r2900, %r2898; and.b32 %r2902, %r74, 7; or.b32 %r2903, %r2901, %r2902; ld.u8 %rs311, [%rd2307+15]; shr.u16 %rs312, %rs311, 5; mov.u32 %r2904, 5; cvt.u32.u16 %r2905, %rs312; sub.s32 %r2906, %r2904, %r2905; shr.u32 %r2907, %r2903, %r2906; mul.wide.u32 %rd1553, %r2907, 4; add.s64 %rd1554, %rd2307, %rd1553; mov.u32 %r2908, 32; shr.u32 %r2909, %r2908, %r2905; add.s32 %r2910, %r2909, -1; and.b32 %r2911, %r2910, %r2903; shl.b32 %r2912, %r2911, %r2905; ld.u32 %r2913, [%rd1554+96]; shr.u32 %r2914, %r2913, %r2912; shl.b32 %r2915, %r2896, %r2905; mov.u32 %r2916, -1; shl.b32 %r2917, %r2916, %r2915; not.b32 %r2918, %r2917; and.b32 %r2919, %r2914, %r2918; cvt.rn.f32.u32 %f665, %r2919; ld.v2.f32 {%f666, %f667}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f667, %f665, %f666; mov.b64 %rd1555, {%r636, %r2895}; shr.u64 %rd1556, %rd1555, 32; cvt.u32.u64 %r4782, %rd1556; mov.u32 %r4827, %r636; bra.uni $L__BB5_486; $L__BB5_491: and.b32 %r700, %r635, -128; setp.ne.s32 %p307, %r700, %r4870; mov.u16 %rs493, 0; @%p307 bra $L__BB5_494; and.b32 %r3039, %r8, -128; setp.ne.s32 %p308, %r3039, %r4869; @%p308 bra $L__BB5_494; and.b32 %r3040, %r74, -128; setp.eq.s32 %p309, %r3040, %r4868; selp.u16 %rs493, 1, 0, %p309; $L__BB5_494: setp.eq.s16 %p310, %rs493, 0; @%p310 bra $L__BB5_498; shl.b32 %r3041, %r635, 5; and.b32 %r3042, %r3041, 3840; shl.b32 %r3043, %r8, 1; and.b32 %r3044, %r3043, 240; or.b32 %r3045, %r3042, %r3044; bfe.u32 %r3046, %r74, 3, 4; or.b32 %r3047, %r3046, %r3045; shr.u32 %r3048, %r3045, 6; mul.wide.u32 %rd1615, %r3048, 8; add.s64 %rd1616, %rd2311, %rd1615; and.b32 %r3049, %r3047, 63; mov.u64 %rd1617, 1; shl.b64 %rd1618, %rd1617, %r3049; ld.u64 %rd1619, [%rd1616+544]; and.b64 %rd1620, %rd1619, %rd1618; setp.eq.s64 %p311, %rd1620, 0; mul.wide.u32 %rd1621, %r3047, 8; add.s64 %rd1622, %rd2311, %rd1621; add.s64 %rd420, %rd1622, 1088; @%p311 bra $L__BB5_497; ld.u64 %rd1623, [%rd420]; add.s64 %rd2307, %rd2311, %rd1623; and.b32 %r3050, %r8, -8; mov.u32 %r3051, 1; st.local.v2.u32 [%rd2], {%r636, %r3050}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r3052, %r635, 6; and.b32 %r3053, %r3052, 448; shl.b32 %r3054, %r8, 3; and.b32 %r3055, %r3054, 56; or.b32 %r3056, %r3053, %r3055; and.b32 %r3057, %r74, 7; or.b32 %r3058, %r3056, %r3057; ld.u8 %rs325, [%rd2307+15]; shr.u16 %rs326, %rs325, 5; mov.u32 %r3059, 5; cvt.u32.u16 %r3060, %rs326; sub.s32 %r3061, %r3059, %r3060; shr.u32 %r3062, %r3058, %r3061; mul.wide.u32 %rd1624, %r3062, 4; add.s64 %rd1625, %rd2307, %rd1624; mov.u32 %r3063, 32; shr.u32 %r3064, %r3063, %r3060; add.s32 %r3065, %r3064, -1; and.b32 %r3066, %r3065, %r3058; shl.b32 %r3067, %r3066, %r3060; ld.u32 %r3068, [%rd1625+96]; shr.u32 %r3069, %r3068, %r3067; shl.b32 %r3070, %r3051, %r3060; mov.u32 %r3071, -1; shl.b32 %r3072, %r3071, %r3070; not.b32 %r3073, %r3072; and.b32 %r3074, %r3069, %r3073; cvt.rn.f32.u32 %f686, %r3074; ld.v2.f32 {%f687, %f688}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f688, %f686, %f687; mov.b64 %rd1626, {%r636, %r3050}; shr.u64 %rd1627, %rd1626, 32; cvt.u32.u64 %r4782, %rd1627; mov.u32 %r4827, %r636; bra.uni $L__BB5_521; $L__BB5_526: and.b32 %r727, %r635, -128; setp.ne.s32 %p329, %r727, %r4870; mov.u16 %rs496, 0; @%p329 bra $L__BB5_529; and.b32 %r3193, %r8, -128; setp.ne.s32 %p330, %r3193, %r4869; @%p330 bra $L__BB5_529; and.b32 %r3194, %r9, -128; setp.eq.s32 %p331, %r3194, %r4868; selp.u16 %rs496, 1, 0, %p331; $L__BB5_529: setp.eq.s16 %p332, %rs496, 0; @%p332 bra $L__BB5_533; shl.b32 %r3195, %r635, 5; and.b32 %r3196, %r3195, 3840; shl.b32 %r3197, %r8, 1; and.b32 %r3198, %r3197, 240; or.b32 %r3199, %r3196, %r3198; bfe.u32 %r3200, %r9, 3, 4; or.b32 %r3201, %r3199, %r3200; shr.u32 %r3202, %r3199, 6; mul.wide.u32 %rd1686, %r3202, 8; add.s64 %rd1687, %rd2311, %rd1686; and.b32 %r3203, %r3201, 63; mov.u64 %rd1688, 1; shl.b64 %rd1689, %rd1688, %r3203; ld.u64 %rd1690, [%rd1687+544]; and.b64 %rd1691, %rd1690, %rd1689; setp.eq.s64 %p333, %rd1691, 0; mul.wide.u32 %rd1692, %r3201, 8; add.s64 %rd1693, %rd2311, %rd1692; add.s64 %rd449, %rd1693, 1088; @%p333 bra $L__BB5_532; ld.u64 %rd1694, [%rd449]; add.s64 %rd1695, %rd2311, %rd1694; and.b32 %r3204, %r8, -8; mov.u32 %r3205, 1; st.local.v2.u32 [%rd2], {%r636, %r3204}; and.b32 %r3206, %r9, -8; st.local.u32 [%rd2+8], %r3206; st.local.u64 [%rd2+48], %rd1695; shl.b32 %r3207, %r635, 6; and.b32 %r3208, %r3207, 448; shl.b32 %r3209, %r8, 3; and.b32 %r3210, %r3209, 56; or.b32 %r3211, %r3208, %r3210; and.b32 %r3212, %r9, 7; or.b32 %r3213, %r3211, %r3212; ld.u8 %rs339, [%rd1695+15]; shr.u16 %rs340, %rs339, 5; mov.u32 %r3214, 5; cvt.u32.u16 %r3215, %rs340; sub.s32 %r3216, %r3214, %r3215; shr.u32 %r3217, %r3213, %r3216; mul.wide.u32 %rd1696, %r3217, 4; add.s64 %rd1697, %rd1695, %rd1696; mov.u32 %r3218, 32; shr.u32 %r3219, %r3218, %r3215; add.s32 %r3220, %r3219, -1; and.b32 %r3221, %r3220, %r3213; shl.b32 %r3222, %r3221, %r3215; ld.u32 %r3223, [%rd1697+96]; shr.u32 %r3224, %r3223, %r3222; shl.b32 %r3225, %r3205, %r3215; mov.u32 %r3226, -1; shl.b32 %r3227, %r3226, %r3225; not.b32 %r3228, %r3227; and.b32 %r3229, %r3224, %r3228; cvt.rn.f32.u32 %f706, %r3229; ld.v2.f32 {%f707, %f708}, [%rd1695+80]; fma.rn.ftz.f32 %f944, %f708, %f706, %f707; bra.uni $L__BB5_556; $L__BB5_561: and.b32 %r735, %r471, -128; ld.local.u32 %r3356, [%rd2+12]; setp.ne.s32 %p351, %r735, %r3356; mov.u16 %rs499, 0; @%p351 bra $L__BB5_564; and.b32 %r3357, %r732, -128; ld.local.u32 %r3358, [%rd2+16]; setp.ne.s32 %p352, %r3357, %r3358; @%p352 bra $L__BB5_564; and.b32 %r3359, %r733, -128; ld.local.u32 %r3360, [%rd2+20]; setp.eq.s32 %p353, %r3359, %r3360; selp.u16 %rs499, 1, 0, %p353; $L__BB5_564: setp.eq.s16 %p354, %rs499, 0; @%p354 bra $L__BB5_568; shl.b32 %r3361, %r471, 5; and.b32 %r3362, %r3361, 3840; shl.b32 %r3363, %r732, 1; and.b32 %r3364, %r3363, 240; or.b32 %r3365, %r3364, %r3362; bfe.u32 %r3366, %r733, 3, 4; or.b32 %r3367, %r3365, %r3366; shr.u32 %r3368, %r3365, 6; ld.local.u64 %rd464, [%rd2+56]; mul.wide.u32 %rd1752, %r3368, 8; add.s64 %rd1753, %rd464, %rd1752; and.b32 %r3369, %r3367, 63; mov.u64 %rd1754, 1; shl.b64 %rd1755, %rd1754, %r3369; ld.u64 %rd1756, [%rd1753+544]; and.b64 %rd1757, %rd1756, %rd1755; setp.eq.s64 %p355, %rd1757, 0; mul.wide.u32 %rd1758, %r3367, 8; add.s64 %rd1759, %rd464, %rd1758; add.s64 %rd465, %rd1759, 1088; @%p355 bra $L__BB5_567; ld.u64 %rd1760, [%rd465]; add.s64 %rd1761, %rd464, %rd1760; and.b32 %r3370, %r732, -8; mov.u32 %r3371, 1; st.local.v2.u32 [%rd2], {%r472, %r3370}; and.b32 %r3372, %r733, -8; st.local.u32 [%rd2+8], %r3372; st.local.u64 [%rd2+48], %rd1761; shl.b32 %r3373, %r471, 6; and.b32 %r3374, %r3373, 448; shl.b32 %r3375, %r732, 3; and.b32 %r3376, %r3375, 56; or.b32 %r3377, %r3376, %r3374; and.b32 %r3378, %r733, 7; or.b32 %r3379, %r3377, %r3378; ld.u8 %rs353, [%rd1761+15]; shr.u16 %rs354, %rs353, 5; mov.u32 %r3380, 5; cvt.u32.u16 %r3381, %rs354; sub.s32 %r3382, %r3380, %r3381; shr.u32 %r3383, %r3379, %r3382; mul.wide.u32 %rd1762, %r3383, 4; add.s64 %rd1763, %rd1761, %rd1762; mov.u32 %r3384, 32; shr.u32 %r3385, %r3384, %r3381; add.s32 %r3386, %r3385, -1; and.b32 %r3387, %r3386, %r3379; shl.b32 %r3388, %r3387, %r3381; ld.u32 %r3389, [%rd1763+96]; shr.u32 %r3390, %r3389, %r3388; shl.b32 %r3391, %r3371, %r3381; mov.u32 %r3392, -1; shl.b32 %r3393, %r3392, %r3391; not.b32 %r3394, %r3393; and.b32 %r3395, %r3390, %r3394; cvt.rn.f32.u32 %f747, %r3395; ld.v2.f32 {%f748, %f749}, [%rd1761+80]; fma.rn.ftz.f32 %f945, %f749, %f747, %f748; mov.u32 %r5196, %r472; bra.uni $L__BB5_591; $L__BB5_596: and.b32 %r742, %r471, -128; ld.local.u32 %r3527, [%rd2+12]; setp.ne.s32 %p373, %r742, %r3527; mov.u16 %rs502, 0; @%p373 bra $L__BB5_599; and.b32 %r3528, %r732, -128; ld.local.u32 %r3529, [%rd2+16]; setp.ne.s32 %p374, %r3528, %r3529; @%p374 bra $L__BB5_599; and.b32 %r3530, %r741, -128; ld.local.u32 %r3531, [%rd2+20]; setp.eq.s32 %p375, %r3530, %r3531; selp.u16 %rs502, 1, 0, %p375; $L__BB5_599: setp.eq.s16 %p376, %rs502, 0; @%p376 bra $L__BB5_603; shl.b32 %r3532, %r471, 5; and.b32 %r3533, %r3532, 3840; shl.b32 %r3534, %r732, 1; and.b32 %r3535, %r3534, 240; or.b32 %r3536, %r3535, %r3533; bfe.u32 %r3537, %r741, 3, 4; or.b32 %r3538, %r3537, %r3536; shr.u32 %r3539, %r3536, 6; ld.local.u64 %rd482, [%rd2+56]; mul.wide.u32 %rd1819, %r3539, 8; add.s64 %rd1820, %rd482, %rd1819; and.b32 %r3540, %r3538, 63; mov.u64 %rd1821, 1; shl.b64 %rd1822, %rd1821, %r3540; ld.u64 %rd1823, [%rd1820+544]; and.b64 %rd1824, %rd1823, %rd1822; setp.eq.s64 %p377, %rd1824, 0; mul.wide.u32 %rd1825, %r3538, 8; add.s64 %rd1826, %rd482, %rd1825; add.s64 %rd483, %rd1826, 1088; @%p377 bra $L__BB5_602; ld.u64 %rd1827, [%rd483]; add.s64 %rd1828, %rd482, %rd1827; and.b32 %r3541, %r732, -8; mov.u32 %r3542, 1; st.local.v2.u32 [%rd2], {%r472, %r3541}; and.b32 %r3543, %r741, -8; st.local.u32 [%rd2+8], %r3543; st.local.u64 [%rd2+48], %rd1828; shl.b32 %r3544, %r471, 6; and.b32 %r3545, %r3544, 448; shl.b32 %r3546, %r732, 3; and.b32 %r3547, %r3546, 56; or.b32 %r3548, %r3547, %r3545; and.b32 %r3549, %r741, 7; or.b32 %r3550, %r3548, %r3549; ld.u8 %rs367, [%rd1828+15]; shr.u16 %rs368, %rs367, 5; mov.u32 %r3551, 5; cvt.u32.u16 %r3552, %rs368; sub.s32 %r3553, %r3551, %r3552; shr.u32 %r3554, %r3550, %r3553; mul.wide.u32 %rd1829, %r3554, 4; add.s64 %rd1830, %rd1828, %rd1829; mov.u32 %r3555, 32; shr.u32 %r3556, %r3555, %r3552; add.s32 %r3557, %r3556, -1; and.b32 %r3558, %r3557, %r3550; shl.b32 %r3559, %r3558, %r3552; ld.u32 %r3560, [%rd1830+96]; shr.u32 %r3561, %r3560, %r3559; shl.b32 %r3562, %r3542, %r3552; mov.u32 %r3563, -1; shl.b32 %r3564, %r3563, %r3562; not.b32 %r3565, %r3564; and.b32 %r3566, %r3561, %r3565; cvt.rn.f32.u32 %f767, %r3566; ld.v2.f32 {%f768, %f769}, [%rd1828+80]; fma.rn.ftz.f32 %f946, %f769, %f767, %f768; mov.u32 %r5196, %r472; bra.uni $L__BB5_626; $L__BB5_631: and.b32 %r749, %r471, -128; ld.local.u32 %r3698, [%rd2+12]; setp.ne.s32 %p395, %r749, %r3698; mov.u16 %rs505, 0; @%p395 bra $L__BB5_634; and.b32 %r3699, %r748, -128; ld.local.u32 %r3700, [%rd2+16]; setp.ne.s32 %p396, %r3699, %r3700; @%p396 bra $L__BB5_634; and.b32 %r3701, %r741, -128; ld.local.u32 %r3702, [%rd2+20]; setp.eq.s32 %p397, %r3701, %r3702; selp.u16 %rs505, 1, 0, %p397; $L__BB5_634: setp.eq.s16 %p398, %rs505, 0; @%p398 bra $L__BB5_638; shl.b32 %r3703, %r471, 5; and.b32 %r3704, %r3703, 3840; shl.b32 %r3705, %r748, 1; and.b32 %r3706, %r3705, 240; or.b32 %r3707, %r3706, %r3704; bfe.u32 %r3708, %r741, 3, 4; or.b32 %r3709, %r3707, %r3708; shr.u32 %r3710, %r3707, 6; ld.local.u64 %rd500, [%rd2+56]; mul.wide.u32 %rd1887, %r3710, 8; add.s64 %rd1888, %rd500, %rd1887; and.b32 %r3711, %r3709, 63; mov.u64 %rd1889, 1; shl.b64 %rd1890, %rd1889, %r3711; ld.u64 %rd1891, [%rd1888+544]; and.b64 %rd1892, %rd1891, %rd1890; setp.eq.s64 %p399, %rd1892, 0; mul.wide.u32 %rd1893, %r3709, 8; add.s64 %rd1894, %rd500, %rd1893; add.s64 %rd501, %rd1894, 1088; @%p399 bra $L__BB5_637; ld.u64 %rd1895, [%rd501]; add.s64 %rd1896, %rd500, %rd1895; and.b32 %r3712, %r748, -8; mov.u32 %r3713, 1; st.local.v2.u32 [%rd2], {%r472, %r3712}; and.b32 %r3714, %r741, -8; st.local.u32 [%rd2+8], %r3714; st.local.u64 [%rd2+48], %rd1896; shl.b32 %r3715, %r471, 6; and.b32 %r3716, %r3715, 448; shl.b32 %r3717, %r748, 3; and.b32 %r3718, %r3717, 56; or.b32 %r3719, %r3718, %r3716; and.b32 %r3720, %r741, 7; or.b32 %r3721, %r3719, %r3720; ld.u8 %rs381, [%rd1896+15]; shr.u16 %rs382, %rs381, 5; mov.u32 %r3722, 5; cvt.u32.u16 %r3723, %rs382; sub.s32 %r3724, %r3722, %r3723; shr.u32 %r3725, %r3721, %r3724; mul.wide.u32 %rd1897, %r3725, 4; add.s64 %rd1898, %rd1896, %rd1897; mov.u32 %r3726, 32; shr.u32 %r3727, %r3726, %r3723; add.s32 %r3728, %r3727, -1; and.b32 %r3729, %r3728, %r3721; shl.b32 %r3730, %r3729, %r3723; ld.u32 %r3731, [%rd1898+96]; shr.u32 %r3732, %r3731, %r3730; shl.b32 %r3733, %r3713, %r3723; mov.u32 %r3734, -1; shl.b32 %r3735, %r3734, %r3733; not.b32 %r3736, %r3735; and.b32 %r3737, %r3732, %r3736; cvt.rn.f32.u32 %f788, %r3737; ld.v2.f32 {%f789, %f790}, [%rd1896+80]; fma.rn.ftz.f32 %f947, %f790, %f788, %f789; mov.u32 %r5196, %r472; bra.uni $L__BB5_661; $L__BB5_666: and.b32 %r755, %r471, -128; ld.local.u32 %r3868, [%rd2+12]; setp.ne.s32 %p417, %r755, %r3868; mov.u16 %rs508, 0; @%p417 bra $L__BB5_669; and.b32 %r3869, %r748, -128; ld.local.u32 %r3870, [%rd2+16]; setp.ne.s32 %p418, %r3869, %r3870; @%p418 bra $L__BB5_669; and.b32 %r3871, %r733, -128; ld.local.u32 %r3872, [%rd2+20]; setp.eq.s32 %p419, %r3871, %r3872; selp.u16 %rs508, 1, 0, %p419; $L__BB5_669: setp.eq.s16 %p420, %rs508, 0; @%p420 bra $L__BB5_673; shl.b32 %r3873, %r471, 5; and.b32 %r3874, %r3873, 3840; shl.b32 %r3875, %r748, 1; and.b32 %r3876, %r3875, 240; or.b32 %r3877, %r3876, %r3874; bfe.u32 %r3878, %r733, 3, 4; or.b32 %r3879, %r3877, %r3878; shr.u32 %r3880, %r3877, 6; ld.local.u64 %rd518, [%rd2+56]; mul.wide.u32 %rd1954, %r3880, 8; add.s64 %rd1955, %rd518, %rd1954; and.b32 %r3881, %r3879, 63; mov.u64 %rd1956, 1; shl.b64 %rd1957, %rd1956, %r3881; ld.u64 %rd1958, [%rd1955+544]; and.b64 %rd1959, %rd1958, %rd1957; setp.eq.s64 %p421, %rd1959, 0; mul.wide.u32 %rd1960, %r3879, 8; add.s64 %rd1961, %rd518, %rd1960; add.s64 %rd519, %rd1961, 1088; @%p421 bra $L__BB5_672; ld.u64 %rd1962, [%rd519]; add.s64 %rd1963, %rd518, %rd1962; and.b32 %r3882, %r748, -8; mov.u32 %r3883, 1; st.local.v2.u32 [%rd2], {%r472, %r3882}; and.b32 %r3884, %r733, -8; st.local.u32 [%rd2+8], %r3884; st.local.u64 [%rd2+48], %rd1963; and.b32 %r3885, %r471, 7; shl.b32 %r3886, %r748, 3; and.b32 %r3887, %r3886, 56; and.b32 %r3888, %r733, 7; bfi.b32 %r3889, %r3885, %r3888, 6, 3; or.b32 %r3890, %r3889, %r3887; ld.u8 %rs395, [%rd1963+15]; shr.u16 %rs396, %rs395, 5; mov.u32 %r3891, 5; cvt.u32.u16 %r3892, %rs396; sub.s32 %r3893, %r3891, %r3892; shr.u32 %r3894, %r3890, %r3893; mul.wide.u32 %rd1964, %r3894, 4; add.s64 %rd1965, %rd1963, %rd1964; mov.u32 %r3895, 32; shr.u32 %r3896, %r3895, %r3892; add.s32 %r3897, %r3896, -1; and.b32 %r3898, %r3897, %r3890; shl.b32 %r3899, %r3898, %r3892; ld.u32 %r3900, [%rd1965+96]; shr.u32 %r3901, %r3900, %r3899; shl.b32 %r3902, %r3883, %r3892; mov.u32 %r3903, -1; shl.b32 %r3904, %r3903, %r3902; not.b32 %r3905, %r3904; and.b32 %r3906, %r3901, %r3905; cvt.rn.f32.u32 %f808, %r3906; ld.v2.f32 {%f809, %f810}, [%rd1963+80]; fma.rn.ftz.f32 %f948, %f810, %f808, %f809; mov.u32 %r5196, %r472; bra.uni $L__BB5_696; $L__BB5_701: and.b32 %r761, %r635, -128; ld.local.u32 %r4035, [%rd2+12]; setp.ne.s32 %p439, %r761, %r4035; mov.u16 %rs511, 0; @%p439 bra $L__BB5_704; and.b32 %r4036, %r748, -128; ld.local.u32 %r4037, [%rd2+16]; setp.ne.s32 %p440, %r4036, %r4037; @%p440 bra $L__BB5_704; and.b32 %r4038, %r733, -128; ld.local.u32 %r4039, [%rd2+20]; setp.eq.s32 %p441, %r4038, %r4039; selp.u16 %rs511, 1, 0, %p441; $L__BB5_704: setp.eq.s16 %p442, %rs511, 0; @%p442 bra $L__BB5_708; shl.b32 %r4040, %r635, 5; and.b32 %r4041, %r4040, 3840; shl.b32 %r4042, %r748, 1; and.b32 %r4043, %r4042, 240; or.b32 %r4044, %r4043, %r4041; bfe.u32 %r4045, %r733, 3, 4; or.b32 %r4046, %r4044, %r4045; shr.u32 %r4047, %r4044, 6; ld.local.u64 %rd536, [%rd2+56]; mul.wide.u32 %rd2021, %r4047, 8; add.s64 %rd2022, %rd536, %rd2021; and.b32 %r4048, %r4046, 63; mov.u64 %rd2023, 1; shl.b64 %rd2024, %rd2023, %r4048; ld.u64 %rd2025, [%rd2022+544]; and.b64 %rd2026, %rd2025, %rd2024; setp.eq.s64 %p443, %rd2026, 0; mul.wide.u32 %rd2027, %r4046, 8; add.s64 %rd2028, %rd536, %rd2027; add.s64 %rd537, %rd2028, 1088; @%p443 bra $L__BB5_707; ld.u64 %rd2029, [%rd537]; add.s64 %rd2030, %rd536, %rd2029; and.b32 %r4049, %r748, -8; mov.u32 %r4050, 1; st.local.v2.u32 [%rd2], {%r636, %r4049}; and.b32 %r4051, %r733, -8; st.local.u32 [%rd2+8], %r4051; st.local.u64 [%rd2+48], %rd2030; and.b32 %r4052, %r635, 7; shl.b32 %r4053, %r748, 3; and.b32 %r4054, %r4053, 56; and.b32 %r4055, %r733, 7; bfi.b32 %r4056, %r4052, %r4055, 6, 3; or.b32 %r4057, %r4056, %r4054; ld.u8 %rs409, [%rd2030+15]; shr.u16 %rs410, %rs409, 5; mov.u32 %r4058, 5; cvt.u32.u16 %r4059, %rs410; sub.s32 %r4060, %r4058, %r4059; shr.u32 %r4061, %r4057, %r4060; mul.wide.u32 %rd2031, %r4061, 4; add.s64 %rd2032, %rd2030, %rd2031; mov.u32 %r4062, 32; shr.u32 %r4063, %r4062, %r4059; add.s32 %r4064, %r4063, -1; and.b32 %r4065, %r4064, %r4057; shl.b32 %r4066, %r4065, %r4059; ld.u32 %r4067, [%rd2032+96]; shr.u32 %r4068, %r4067, %r4066; shl.b32 %r4069, %r4050, %r4059; mov.u32 %r4070, -1; shl.b32 %r4071, %r4070, %r4069; not.b32 %r4072, %r4071; and.b32 %r4073, %r4068, %r4072; cvt.rn.f32.u32 %f831, %r4073; ld.v2.f32 {%f832, %f833}, [%rd2030+80]; fma.rn.ftz.f32 %f949, %f833, %f831, %f832; mov.u32 %r5196, %r636; bra.uni $L__BB5_731; $L__BB5_736: and.b32 %r767, %r635, -128; ld.local.u32 %r4203, [%rd2+12]; setp.ne.s32 %p461, %r767, %r4203; mov.u16 %rs514, 0; @%p461 bra $L__BB5_739; and.b32 %r4204, %r748, -128; ld.local.u32 %r4205, [%rd2+16]; setp.ne.s32 %p462, %r4204, %r4205; @%p462 bra $L__BB5_739; and.b32 %r4206, %r741, -128; ld.local.u32 %r4207, [%rd2+20]; setp.eq.s32 %p463, %r4206, %r4207; selp.u16 %rs514, 1, 0, %p463; $L__BB5_739: setp.eq.s16 %p464, %rs514, 0; @%p464 bra $L__BB5_743; shl.b32 %r4208, %r635, 5; and.b32 %r4209, %r4208, 3840; shl.b32 %r4210, %r748, 1; and.b32 %r4211, %r4210, 240; or.b32 %r4212, %r4211, %r4209; bfe.u32 %r4213, %r741, 3, 4; or.b32 %r4214, %r4212, %r4213; shr.u32 %r4215, %r4212, 6; ld.local.u64 %rd554, [%rd2+56]; mul.wide.u32 %rd2088, %r4215, 8; add.s64 %rd2089, %rd554, %rd2088; and.b32 %r4216, %r4214, 63; mov.u64 %rd2090, 1; shl.b64 %rd2091, %rd2090, %r4216; ld.u64 %rd2092, [%rd2089+544]; and.b64 %rd2093, %rd2092, %rd2091; setp.eq.s64 %p465, %rd2093, 0; mul.wide.u32 %rd2094, %r4214, 8; add.s64 %rd2095, %rd554, %rd2094; add.s64 %rd555, %rd2095, 1088; @%p465 bra $L__BB5_742; ld.u64 %rd2096, [%rd555]; add.s64 %rd2097, %rd554, %rd2096; and.b32 %r4217, %r748, -8; mov.u32 %r4218, 1; st.local.v2.u32 [%rd2], {%r636, %r4217}; and.b32 %r4219, %r741, -8; st.local.u32 [%rd2+8], %r4219; st.local.u64 [%rd2+48], %rd2097; shl.b32 %r4220, %r635, 6; and.b32 %r4221, %r4220, 448; shl.b32 %r4222, %r748, 3; and.b32 %r4223, %r4222, 56; or.b32 %r4224, %r4223, %r4221; and.b32 %r4225, %r741, 7; or.b32 %r4226, %r4224, %r4225; ld.u8 %rs423, [%rd2097+15]; shr.u16 %rs424, %rs423, 5; mov.u32 %r4227, 5; cvt.u32.u16 %r4228, %rs424; sub.s32 %r4229, %r4227, %r4228; shr.u32 %r4230, %r4226, %r4229; mul.wide.u32 %rd2098, %r4230, 4; add.s64 %rd2099, %rd2097, %rd2098; mov.u32 %r4231, 32; shr.u32 %r4232, %r4231, %r4228; add.s32 %r4233, %r4232, -1; and.b32 %r4234, %r4233, %r4226; shl.b32 %r4235, %r4234, %r4228; ld.u32 %r4236, [%rd2099+96]; shr.u32 %r4237, %r4236, %r4235; shl.b32 %r4238, %r4218, %r4228; mov.u32 %r4239, -1; shl.b32 %r4240, %r4239, %r4238; not.b32 %r4241, %r4240; and.b32 %r4242, %r4237, %r4241; cvt.rn.f32.u32 %f851, %r4242; ld.v2.f32 {%f852, %f853}, [%rd2097+80]; fma.rn.ftz.f32 %f950, %f853, %f851, %f852; mov.u32 %r5196, %r636; bra.uni $L__BB5_766; $L__BB5_771: and.b32 %r773, %r635, -128; ld.local.u32 %r4374, [%rd2+12]; setp.ne.s32 %p483, %r773, %r4374; mov.u16 %rs517, 0; @%p483 bra $L__BB5_774; and.b32 %r4375, %r732, -128; ld.local.u32 %r4376, [%rd2+16]; setp.ne.s32 %p484, %r4375, %r4376; @%p484 bra $L__BB5_774; and.b32 %r4377, %r741, -128; ld.local.u32 %r4378, [%rd2+20]; setp.eq.s32 %p485, %r4377, %r4378; selp.u16 %rs517, 1, 0, %p485; $L__BB5_774: setp.eq.s16 %p486, %rs517, 0; @%p486 bra $L__BB5_778; shl.b32 %r4379, %r635, 5; and.b32 %r4380, %r4379, 3840; shl.b32 %r4381, %r732, 1; and.b32 %r4382, %r4381, 240; or.b32 %r4383, %r4380, %r4382; bfe.u32 %r4384, %r741, 3, 4; or.b32 %r4385, %r4384, %r4383; shr.u32 %r4386, %r4383, 6; ld.local.u64 %rd572, [%rd2+56]; mul.wide.u32 %rd2155, %r4386, 8; add.s64 %rd2156, %rd572, %rd2155; and.b32 %r4387, %r4385, 63; mov.u64 %rd2157, 1; shl.b64 %rd2158, %rd2157, %r4387; ld.u64 %rd2159, [%rd2156+544]; and.b64 %rd2160, %rd2159, %rd2158; setp.eq.s64 %p487, %rd2160, 0; mul.wide.u32 %rd2161, %r4385, 8; add.s64 %rd2162, %rd572, %rd2161; add.s64 %rd573, %rd2162, 1088; @%p487 bra $L__BB5_777; ld.u64 %rd2163, [%rd573]; add.s64 %rd2164, %rd572, %rd2163; and.b32 %r4388, %r732, -8; mov.u32 %r4389, 1; st.local.v2.u32 [%rd2], {%r636, %r4388}; and.b32 %r4390, %r741, -8; st.local.u32 [%rd2+8], %r4390; st.local.u64 [%rd2+48], %rd2164; shl.b32 %r4391, %r635, 6; and.b32 %r4392, %r4391, 448; shl.b32 %r4393, %r732, 3; and.b32 %r4394, %r4393, 56; or.b32 %r4395, %r4392, %r4394; and.b32 %r4396, %r741, 7; or.b32 %r4397, %r4395, %r4396; ld.u8 %rs437, [%rd2164+15]; shr.u16 %rs438, %rs437, 5; mov.u32 %r4398, 5; cvt.u32.u16 %r4399, %rs438; sub.s32 %r4400, %r4398, %r4399; shr.u32 %r4401, %r4397, %r4400; mul.wide.u32 %rd2165, %r4401, 4; add.s64 %rd2166, %rd2164, %rd2165; mov.u32 %r4402, 32; shr.u32 %r4403, %r4402, %r4399; add.s32 %r4404, %r4403, -1; and.b32 %r4405, %r4404, %r4397; shl.b32 %r4406, %r4405, %r4399; ld.u32 %r4407, [%rd2166+96]; shr.u32 %r4408, %r4407, %r4406; shl.b32 %r4409, %r4389, %r4399; mov.u32 %r4410, -1; shl.b32 %r4411, %r4410, %r4409; not.b32 %r4412, %r4411; and.b32 %r4413, %r4408, %r4412; cvt.rn.f32.u32 %f872, %r4413; ld.v2.f32 {%f873, %f874}, [%rd2164+80]; fma.rn.ftz.f32 %f951, %f874, %f872, %f873; mov.u32 %r5196, %r636; bra.uni $L__BB5_801; $L__BB5_814: shl.b64 %rd2230, %rd593, 5; add.s64 %rd2394, %rd591, %rd2230; $L__BB5_815: setp.eq.s64 %p513, %rd2394, 0; @%p513 bra $L__BB5_823; add.s64 %rd596, %rd2394, 8; ld.u64 %rd597, [%rd2394+8]; setp.eq.s64 %p514, %rd597, 0; @%p514 bra $L__BB5_822; add.s64 %rd598, %rd590, %rd597; and.b32 %r4535, %r732, -4096; mov.u64 %rd2231, 1; st.local.v2.u32 [%rd2+24], {%r780, %r4535}; and.b32 %r4536, %r733, -4096; st.local.u32 [%rd2+32], %r4536; st.local.u64 [%rd2+64], %rd598; shl.b32 %r4537, %r635, 3; and.b32 %r4538, %r4537, 31744; and.b32 %r4539, %r732, 3968; shr.u32 %r4540, %r4539, 2; or.b32 %r4541, %r4538, %r4540; bfe.u32 %r4542, %r733, 7, 5; or.b32 %r4543, %r4541, %r4542; shr.u32 %r4544, %r4541, 6; mul.wide.u32 %rd2232, %r4544, 8; add.s64 %rd2233, %rd598, %rd2232; and.b32 %r4545, %r4543, 63; shl.b64 %rd2234, %rd2231, %r4545; ld.u64 %rd2235, [%rd2233+4128]; and.b64 %rd2236, %rd2235, %rd2234; setp.eq.s64 %p515, %rd2236, 0; mul.wide.u32 %rd2237, %r4543, 8; add.s64 %rd2238, %rd598, %rd2237; add.s64 %rd599, %rd2238, 8256; @%p515 bra $L__BB5_821; ld.u64 %rd2239, [%rd599]; add.s64 %rd600, %rd598, %rd2239; st.local.u32 [%rd2+12], %r779; and.b32 %r4546, %r732, -128; st.local.u32 [%rd2+16], %r4546; and.b32 %r4547, %r733, -128; st.local.u32 [%rd2+20], %r4547; st.local.u64 [%rd2+56], %rd600; shl.b32 %r4548, %r635, 5; and.b32 %r4549, %r4548, 3840; shl.b32 %r4550, %r732, 1; and.b32 %r4551, %r4550, 240; or.b32 %r4552, %r4549, %r4551; bfe.u32 %r4553, %r733, 3, 4; or.b32 %r4554, %r4552, %r4553; shr.u32 %r4555, %r4552, 6; mul.wide.u32 %rd2240, %r4555, 8; add.s64 %rd2241, %rd600, %rd2240; and.b32 %r4556, %r4554, 63; shl.b64 %rd2243, %rd2231, %r4556; ld.u64 %rd2244, [%rd2241+544]; and.b64 %rd2245, %rd2244, %rd2243; setp.eq.s64 %p516, %rd2245, 0; mul.wide.u32 %rd2246, %r4554, 8; add.s64 %rd2247, %rd600, %rd2246; add.s64 %rd601, %rd2247, 1088; @%p516 bra $L__BB5_820; ld.u64 %rd2248, [%rd601]; add.s64 %rd2249, %rd600, %rd2248; and.b32 %r4557, %r732, -8; mov.u32 %r4558, 1; st.local.v2.u32 [%rd2], {%r636, %r4557}; and.b32 %r4559, %r733, -8; st.local.u32 [%rd2+8], %r4559; st.local.u64 [%rd2+48], %rd2249; shl.b32 %r4560, %r635, 6; and.b32 %r4561, %r4560, 448; shl.b32 %r4562, %r732, 3; and.b32 %r4563, %r4562, 56; or.b32 %r4564, %r4561, %r4563; and.b32 %r4565, %r733, 7; or.b32 %r4566, %r4564, %r4565; ld.u8 %rs445, [%rd2249+15]; shr.u16 %rs446, %rs445, 5; mov.u32 %r4567, 5; cvt.u32.u16 %r4568, %rs446; sub.s32 %r4569, %r4567, %r4568; shr.u32 %r4570, %r4566, %r4569; mul.wide.u32 %rd2250, %r4570, 4; add.s64 %rd2251, %rd2249, %rd2250; mov.u32 %r4571, 32; shr.u32 %r4572, %r4571, %r4568; add.s32 %r4573, %r4572, -1; and.b32 %r4574, %r4573, %r4566; shl.b32 %r4575, %r4574, %r4568; ld.u32 %r4576, [%rd2251+96]; shr.u32 %r4577, %r4576, %r4575; shl.b32 %r4578, %r4558, %r4568; mov.u32 %r4579, -1; shl.b32 %r4580, %r4579, %r4578; not.b32 %r4581, %r4580; and.b32 %r4582, %r4577, %r4581; cvt.rn.f32.u32 %f887, %r4582; ld.v2.f32 {%f888, %f889}, [%rd2249+80]; fma.rn.ftz.f32 %f952, %f889, %f887, %f888; bra.uni $L__BB5_833; $L__BB5_823: ld.f32 %f952, [%rd590+28]; bra.uni $L__BB5_833; $L__BB5_822: ld.f32 %f952, [%rd596+12]; bra.uni $L__BB5_833; $L__BB5_43: setp.ne.s32 %p26, %r12, %r4830; mov.u16 %rs455, 0; @%p26 bra $L__BB5_46; and.b32 %r1090, %r8, -4096; setp.ne.s32 %p27, %r1090, %r4875; @%p27 bra $L__BB5_46; and.b32 %r1091, %r74, -4096; setp.eq.s32 %p28, %r1091, %r4908; selp.u16 %rs455, 1, 0, %p28; $L__BB5_46: setp.eq.s16 %p29, %rs455, 0; @%p29 bra $L__BB5_52; shl.b32 %r1092, %r7, 3; and.b32 %r1093, %r1092, 31744; and.b32 %r1094, %r8, 3968; shr.u32 %r1095, %r1094, 2; or.b32 %r1096, %r1095, %r1093; bfe.u32 %r1097, %r74, 7, 5; or.b32 %r1098, %r1097, %r1096; shr.u32 %r1099, %r1096, 6; mul.wide.u32 %rd713, %r1099, 8; add.s64 %rd714, %rd2315, %rd713; and.b32 %r1100, %r1098, 63; mov.u64 %rd715, 1; shl.b64 %rd716, %rd715, %r1100; ld.u64 %rd717, [%rd714+4128]; and.b64 %rd718, %rd717, %rd716; setp.eq.s64 %p30, %rd718, 0; mul.wide.u32 %rd719, %r1098, 8; add.s64 %rd720, %rd2315, %rd719; add.s64 %rd45, %rd720, 8256; @%p30 bra $L__BB5_51; and.b32 %r4750, %r7, -128; ld.u64 %rd721, [%rd45]; add.s64 %rd2311, %rd2315, %rd721; st.local.u32 [%rd2+12], %r4750; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1101, %r7, 5; and.b32 %r1102, %r1101, 3840; shl.b32 %r1103, %r8, 1; and.b32 %r1104, %r1103, 240; or.b32 %r1105, %r1104, %r1102; bfe.u32 %r1106, %r74, 3, 4; or.b32 %r1107, %r1106, %r1105; shr.u32 %r1108, %r1105, 6; mul.wide.u32 %rd722, %r1108, 8; add.s64 %rd723, %rd2311, %rd722; and.b32 %r1109, %r1107, 63; shl.b64 %rd725, %rd715, %r1109; ld.u64 %rd726, [%rd723+544]; and.b64 %rd727, %rd726, %rd725; setp.eq.s64 %p31, %rd727, 0; mul.wide.u32 %rd728, %r1107, 8; add.s64 %rd729, %rd2311, %rd728; add.s64 %rd47, %rd729, 1088; and.b32 %r4870, %r7, -128; @%p31 bra $L__BB5_50; and.b32 %r4827, %r7, -8; ld.u64 %rd730, [%rd47]; add.s64 %rd2307, %rd2311, %rd730; and.b32 %r1110, %r8, -8; mov.u32 %r1111, 1; st.local.v2.u32 [%rd2], {%r4827, %r1110}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1112, %r7, 6; and.b32 %r1113, %r1112, 448; shl.b32 %r1114, %r8, 3; and.b32 %r1115, %r1114, 56; or.b32 %r1116, %r1115, %r1113; and.b32 %r1117, %r74, 7; or.b32 %r1118, %r1116, %r1117; ld.u8 %rs147, [%rd2307+15]; shr.u16 %rs148, %rs147, 5; mov.u32 %r1119, 5; cvt.u32.u16 %r1120, %rs148; sub.s32 %r1121, %r1119, %r1120; shr.u32 %r1122, %r1118, %r1121; mul.wide.u32 %rd731, %r1122, 4; add.s64 %rd732, %rd2307, %rd731; mov.u32 %r1123, 32; shr.u32 %r1124, %r1123, %r1120; add.s32 %r1125, %r1124, -1; and.b32 %r1126, %r1125, %r1118; shl.b32 %r1127, %r1126, %r1120; ld.u32 %r1128, [%rd732+96]; shr.u32 %r1129, %r1128, %r1127; shl.b32 %r1130, %r1111, %r1120; mov.u32 %r1131, -1; shl.b32 %r1132, %r1131, %r1130; not.b32 %r1133, %r1132; and.b32 %r1134, %r1129, %r1133; cvt.rn.f32.u32 %f414, %r1134; ld.v2.f32 {%f415, %f416}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f416, %f414, %f415; mov.b64 %rd733, {%r4827, %r1110}; shr.u64 %rd734, %rd733, 32; cvt.u32.u64 %r4782, %rd734; bra.uni $L__BB5_66; $L__BB5_78: setp.ne.s32 %p48, %r12, %r4830; mov.u16 %rs458, 0; @%p48 bra $L__BB5_81; and.b32 %r1243, %r139, -4096; setp.ne.s32 %p49, %r1243, %r4875; @%p49 bra $L__BB5_81; and.b32 %r1244, %r74, -4096; setp.eq.s32 %p50, %r1244, %r4908; selp.u16 %rs458, 1, 0, %p50; $L__BB5_81: setp.eq.s16 %p51, %rs458, 0; @%p51 bra $L__BB5_87; shl.b32 %r1245, %r7, 3; and.b32 %r1246, %r1245, 31744; and.b32 %r1247, %r139, 3968; shr.u32 %r1248, %r1247, 2; or.b32 %r1249, %r1248, %r1246; bfe.u32 %r1250, %r74, 7, 5; or.b32 %r1251, %r1249, %r1250; shr.u32 %r1252, %r1249, 6; mul.wide.u32 %rd782, %r1252, 8; add.s64 %rd783, %rd2315, %rd782; and.b32 %r1253, %r1251, 63; mov.u64 %rd784, 1; shl.b64 %rd785, %rd784, %r1253; ld.u64 %rd786, [%rd783+4128]; and.b64 %rd787, %rd786, %rd785; setp.eq.s64 %p52, %rd787, 0; mul.wide.u32 %rd788, %r1251, 8; add.s64 %rd789, %rd2315, %rd788; add.s64 %rd74, %rd789, 8256; @%p52 bra $L__BB5_86; and.b32 %r4726, %r7, -128; ld.u64 %rd790, [%rd74]; add.s64 %rd2311, %rd2315, %rd790; st.local.u32 [%rd2+12], %r4726; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1254, %r7, 5; and.b32 %r1255, %r1254, 3840; shl.b32 %r1256, %r139, 1; and.b32 %r1257, %r1256, 240; or.b32 %r1258, %r1257, %r1255; bfe.u32 %r1259, %r74, 3, 4; or.b32 %r1260, %r1258, %r1259; shr.u32 %r1261, %r1258, 6; mul.wide.u32 %rd791, %r1261, 8; add.s64 %rd792, %rd2311, %rd791; and.b32 %r1262, %r1260, 63; shl.b64 %rd794, %rd784, %r1262; ld.u64 %rd795, [%rd792+544]; and.b64 %rd796, %rd795, %rd794; setp.eq.s64 %p53, %rd796, 0; mul.wide.u32 %rd797, %r1260, 8; add.s64 %rd798, %rd2311, %rd797; add.s64 %rd76, %rd798, 1088; and.b32 %r4870, %r7, -128; @%p53 bra $L__BB5_85; and.b32 %r4827, %r7, -8; ld.u64 %rd799, [%rd76]; add.s64 %rd2307, %rd2311, %rd799; and.b32 %r1263, %r139, -8; mov.u32 %r1264, 1; st.local.v2.u32 [%rd2], {%r4827, %r1263}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1265, %r7, 6; and.b32 %r1266, %r1265, 448; shl.b32 %r1267, %r139, 3; and.b32 %r1268, %r1267, 56; or.b32 %r1269, %r1268, %r1266; and.b32 %r1270, %r74, 7; or.b32 %r1271, %r1269, %r1270; ld.u8 %rs161, [%rd2307+15]; shr.u16 %rs162, %rs161, 5; mov.u32 %r1272, 5; cvt.u32.u16 %r1273, %rs162; sub.s32 %r1274, %r1272, %r1273; shr.u32 %r1275, %r1271, %r1274; mul.wide.u32 %rd800, %r1275, 4; add.s64 %rd801, %rd2307, %rd800; mov.u32 %r1276, 32; shr.u32 %r1277, %r1276, %r1273; add.s32 %r1278, %r1277, -1; and.b32 %r1279, %r1278, %r1271; shl.b32 %r1280, %r1279, %r1273; ld.u32 %r1281, [%rd801+96]; shr.u32 %r1282, %r1281, %r1280; shl.b32 %r1283, %r1264, %r1273; mov.u32 %r1284, -1; shl.b32 %r1285, %r1284, %r1283; not.b32 %r1286, %r1285; and.b32 %r1287, %r1282, %r1286; cvt.rn.f32.u32 %f435, %r1287; ld.v2.f32 {%f436, %f437}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f437, %f435, %f436; mov.b64 %rd802, {%r4827, %r1263}; shr.u64 %rd803, %rd802, 32; cvt.u32.u64 %r4782, %rd803; bra.uni $L__BB5_101; $L__BB5_113: setp.ne.s32 %p70, %r12, %r4830; mov.u16 %rs461, 0; @%p70 bra $L__BB5_116; and.b32 %r1395, %r139, -4096; setp.ne.s32 %p71, %r1395, %r4875; @%p71 bra $L__BB5_116; and.b32 %r1396, %r9, -4096; setp.eq.s32 %p72, %r1396, %r4908; selp.u16 %rs461, 1, 0, %p72; $L__BB5_116: setp.eq.s16 %p73, %rs461, 0; @%p73 bra $L__BB5_122; shl.b32 %r1397, %r7, 3; and.b32 %r1398, %r1397, 31744; and.b32 %r1399, %r139, 3968; shr.u32 %r1400, %r1399, 2; or.b32 %r1401, %r1400, %r1398; bfe.u32 %r1402, %r9, 7, 5; or.b32 %r1403, %r1401, %r1402; shr.u32 %r1404, %r1401, 6; mul.wide.u32 %rd852, %r1404, 8; add.s64 %rd853, %rd2315, %rd852; and.b32 %r1405, %r1403, 63; mov.u64 %rd854, 1; shl.b64 %rd855, %rd854, %r1405; ld.u64 %rd856, [%rd853+4128]; and.b64 %rd857, %rd856, %rd855; setp.eq.s64 %p74, %rd857, 0; mul.wide.u32 %rd858, %r1403, 8; add.s64 %rd859, %rd2315, %rd858; add.s64 %rd103, %rd859, 8256; @%p74 bra $L__BB5_121; and.b32 %r4699, %r7, -128; ld.u64 %rd860, [%rd103]; add.s64 %rd2311, %rd2315, %rd860; st.local.u32 [%rd2+12], %r4699; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1406, %r7, 5; and.b32 %r1407, %r1406, 3840; shl.b32 %r1408, %r139, 1; and.b32 %r1409, %r1408, 240; or.b32 %r1410, %r1409, %r1407; bfe.u32 %r1411, %r9, 3, 4; or.b32 %r1412, %r1410, %r1411; shr.u32 %r1413, %r1410, 6; mul.wide.u32 %rd861, %r1413, 8; add.s64 %rd862, %rd2311, %rd861; and.b32 %r1414, %r1412, 63; shl.b64 %rd864, %rd854, %r1414; ld.u64 %rd865, [%rd862+544]; and.b64 %rd866, %rd865, %rd864; setp.eq.s64 %p75, %rd866, 0; mul.wide.u32 %rd867, %r1412, 8; add.s64 %rd868, %rd2311, %rd867; add.s64 %rd105, %rd868, 1088; and.b32 %r4870, %r7, -128; @%p75 bra $L__BB5_120; and.b32 %r4827, %r7, -8; ld.u64 %rd869, [%rd105]; add.s64 %rd2307, %rd2311, %rd869; and.b32 %r1415, %r139, -8; mov.u32 %r1416, 1; st.local.v2.u32 [%rd2], {%r4827, %r1415}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1417, %r7, 7; shl.b32 %r1418, %r139, 3; and.b32 %r1419, %r1418, 56; and.b32 %r1420, %r9, 7; bfi.b32 %r1421, %r1417, %r1420, 6, 3; or.b32 %r1422, %r1421, %r1419; ld.u8 %rs175, [%rd2307+15]; shr.u16 %rs176, %rs175, 5; mov.u32 %r1423, 5; cvt.u32.u16 %r1424, %rs176; sub.s32 %r1425, %r1423, %r1424; shr.u32 %r1426, %r1422, %r1425; mul.wide.u32 %rd870, %r1426, 4; add.s64 %rd871, %rd2307, %rd870; mov.u32 %r1427, 32; shr.u32 %r1428, %r1427, %r1424; add.s32 %r1429, %r1428, -1; and.b32 %r1430, %r1429, %r1422; shl.b32 %r1431, %r1430, %r1424; ld.u32 %r1432, [%rd871+96]; shr.u32 %r1433, %r1432, %r1431; shl.b32 %r1434, %r1416, %r1424; mov.u32 %r1435, -1; shl.b32 %r1436, %r1435, %r1434; not.b32 %r1437, %r1436; and.b32 %r1438, %r1433, %r1437; cvt.rn.f32.u32 %f455, %r1438; ld.v2.f32 {%f456, %f457}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f457, %f455, %f456; mov.b64 %rd872, {%r4827, %r1415}; shr.u64 %rd873, %rd872, 32; cvt.u32.u64 %r4782, %rd873; bra.uni $L__BB5_136; $L__BB5_148: and.b32 %r264, %r259, -4096; setp.ne.s32 %p92, %r264, %r4830; mov.u16 %rs464, 0; @%p92 bra $L__BB5_151; and.b32 %r1544, %r139, -4096; setp.ne.s32 %p93, %r1544, %r4875; @%p93 bra $L__BB5_151; and.b32 %r1545, %r9, -4096; setp.eq.s32 %p94, %r1545, %r4908; selp.u16 %rs464, 1, 0, %p94; $L__BB5_151: setp.eq.s16 %p95, %rs464, 0; @%p95 bra $L__BB5_157; shl.b32 %r1546, %r259, 3; and.b32 %r1547, %r1546, 31744; and.b32 %r1548, %r139, 3968; shr.u32 %r1549, %r1548, 2; or.b32 %r1550, %r1549, %r1547; bfe.u32 %r1551, %r9, 7, 5; or.b32 %r1552, %r1550, %r1551; shr.u32 %r1553, %r1550, 6; mul.wide.u32 %rd920, %r1553, 8; add.s64 %rd921, %rd2315, %rd920; and.b32 %r1554, %r1552, 63; mov.u64 %rd922, 1; shl.b64 %rd923, %rd922, %r1554; ld.u64 %rd924, [%rd921+4128]; and.b64 %rd925, %rd924, %rd923; setp.eq.s64 %p96, %rd925, 0; mul.wide.u32 %rd926, %r1552, 8; add.s64 %rd927, %rd2315, %rd926; add.s64 %rd132, %rd927, 8256; @%p96 bra $L__BB5_156; and.b32 %r4732, %r259, -128; ld.u64 %rd928, [%rd132]; add.s64 %rd2311, %rd2315, %rd928; st.local.u32 [%rd2+12], %r4732; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1555, %r259, 5; and.b32 %r1556, %r1555, 3840; shl.b32 %r1557, %r139, 1; and.b32 %r1558, %r1557, 240; or.b32 %r1559, %r1558, %r1556; bfe.u32 %r1560, %r9, 3, 4; or.b32 %r1561, %r1559, %r1560; shr.u32 %r1562, %r1559, 6; mul.wide.u32 %rd929, %r1562, 8; add.s64 %rd930, %rd2311, %rd929; and.b32 %r1563, %r1561, 63; shl.b64 %rd932, %rd922, %r1563; ld.u64 %rd933, [%rd930+544]; and.b64 %rd934, %rd933, %rd932; setp.eq.s64 %p97, %rd934, 0; mul.wide.u32 %rd935, %r1561, 8; add.s64 %rd936, %rd2311, %rd935; add.s64 %rd134, %rd936, 1088; @%p97 bra $L__BB5_155; and.b32 %r4827, %r259, -8; and.b32 %r4870, %r259, -128; ld.u64 %rd937, [%rd134]; add.s64 %rd2307, %rd2311, %rd937; and.b32 %r1564, %r139, -8; mov.u32 %r1565, 1; st.local.v2.u32 [%rd2], {%r4827, %r1564}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1566, %r259, 7; shl.b32 %r1567, %r139, 3; and.b32 %r1568, %r1567, 56; and.b32 %r1569, %r9, 7; bfi.b32 %r1570, %r1566, %r1569, 6, 3; or.b32 %r1571, %r1570, %r1568; ld.u8 %rs189, [%rd2307+15]; shr.u16 %rs190, %rs189, 5; mov.u32 %r1572, 5; cvt.u32.u16 %r1573, %rs190; sub.s32 %r1574, %r1572, %r1573; shr.u32 %r1575, %r1571, %r1574; mul.wide.u32 %rd938, %r1575, 4; add.s64 %rd939, %rd2307, %rd938; mov.u32 %r1576, 32; shr.u32 %r1577, %r1576, %r1573; add.s32 %r1578, %r1577, -1; and.b32 %r1579, %r1578, %r1571; shl.b32 %r1580, %r1579, %r1573; ld.u32 %r1581, [%rd939+96]; shr.u32 %r1582, %r1581, %r1580; shl.b32 %r1583, %r1565, %r1573; mov.u32 %r1584, -1; shl.b32 %r1585, %r1584, %r1583; not.b32 %r1586, %r1585; and.b32 %r1587, %r1582, %r1586; cvt.rn.f32.u32 %f478, %r1587; ld.v2.f32 {%f479, %f480}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f480, %f478, %f479; mov.b64 %rd940, {%r4827, %r1564}; shr.u64 %rd941, %rd940, 32; cvt.u32.u64 %r4782, %rd941; bra.uni $L__BB5_171; $L__BB5_183: and.b32 %r321, %r259, -4096; setp.ne.s32 %p114, %r321, %r4830; mov.u16 %rs467, 0; @%p114 bra $L__BB5_186; and.b32 %r1696, %r139, -4096; setp.ne.s32 %p115, %r1696, %r4875; @%p115 bra $L__BB5_186; and.b32 %r1697, %r74, -4096; setp.eq.s32 %p116, %r1697, %r4908; selp.u16 %rs467, 1, 0, %p116; $L__BB5_186: setp.eq.s16 %p117, %rs467, 0; @%p117 bra $L__BB5_192; shl.b32 %r1698, %r259, 3; and.b32 %r1699, %r1698, 31744; and.b32 %r1700, %r139, 3968; shr.u32 %r1701, %r1700, 2; or.b32 %r1702, %r1701, %r1699; bfe.u32 %r1703, %r74, 7, 5; or.b32 %r1704, %r1702, %r1703; shr.u32 %r1705, %r1702, 6; mul.wide.u32 %rd991, %r1705, 8; add.s64 %rd992, %rd2315, %rd991; and.b32 %r1706, %r1704, 63; mov.u64 %rd993, 1; shl.b64 %rd994, %rd993, %r1706; ld.u64 %rd995, [%rd992+4128]; and.b64 %rd996, %rd995, %rd994; setp.eq.s64 %p118, %rd996, 0; mul.wide.u32 %rd997, %r1704, 8; add.s64 %rd998, %rd2315, %rd997; add.s64 %rd161, %rd998, 8256; @%p118 bra $L__BB5_191; and.b32 %r4735, %r259, -128; ld.u64 %rd999, [%rd161]; add.s64 %rd2311, %rd2315, %rd999; st.local.u32 [%rd2+12], %r4735; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1707, %r259, 5; and.b32 %r1708, %r1707, 3840; shl.b32 %r1709, %r139, 1; and.b32 %r1710, %r1709, 240; or.b32 %r1711, %r1710, %r1708; bfe.u32 %r1712, %r74, 3, 4; or.b32 %r1713, %r1711, %r1712; shr.u32 %r1714, %r1711, 6; mul.wide.u32 %rd1000, %r1714, 8; add.s64 %rd1001, %rd2311, %rd1000; and.b32 %r1715, %r1713, 63; shl.b64 %rd1003, %rd993, %r1715; ld.u64 %rd1004, [%rd1001+544]; and.b64 %rd1005, %rd1004, %rd1003; setp.eq.s64 %p119, %rd1005, 0; mul.wide.u32 %rd1006, %r1713, 8; add.s64 %rd1007, %rd2311, %rd1006; add.s64 %rd163, %rd1007, 1088; @%p119 bra $L__BB5_190; and.b32 %r4827, %r259, -8; and.b32 %r4870, %r259, -128; ld.u64 %rd1008, [%rd163]; add.s64 %rd2307, %rd2311, %rd1008; and.b32 %r1716, %r139, -8; mov.u32 %r1717, 1; st.local.v2.u32 [%rd2], {%r4827, %r1716}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1718, %r259, 6; and.b32 %r1719, %r1718, 448; shl.b32 %r1720, %r139, 3; and.b32 %r1721, %r1720, 56; or.b32 %r1722, %r1721, %r1719; and.b32 %r1723, %r74, 7; or.b32 %r1724, %r1722, %r1723; ld.u8 %rs203, [%rd2307+15]; shr.u16 %rs204, %rs203, 5; mov.u32 %r1725, 5; cvt.u32.u16 %r1726, %rs204; sub.s32 %r1727, %r1725, %r1726; shr.u32 %r1728, %r1724, %r1727; mul.wide.u32 %rd1009, %r1728, 4; add.s64 %rd1010, %rd2307, %rd1009; mov.u32 %r1729, 32; shr.u32 %r1730, %r1729, %r1726; add.s32 %r1731, %r1730, -1; and.b32 %r1732, %r1731, %r1724; shl.b32 %r1733, %r1732, %r1726; ld.u32 %r1734, [%rd1010+96]; shr.u32 %r1735, %r1734, %r1733; shl.b32 %r1736, %r1717, %r1726; mov.u32 %r1737, -1; shl.b32 %r1738, %r1737, %r1736; not.b32 %r1739, %r1738; and.b32 %r1740, %r1735, %r1739; cvt.rn.f32.u32 %f498, %r1740; ld.v2.f32 {%f499, %f500}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f500, %f498, %f499; mov.b64 %rd1011, {%r4827, %r1716}; shr.u64 %rd1012, %rd1011, 32; cvt.u32.u64 %r4782, %rd1012; bra.uni $L__BB5_206; $L__BB5_218: and.b32 %r375, %r259, -4096; setp.ne.s32 %p136, %r375, %r4830; mov.u16 %rs470, 0; @%p136 bra $L__BB5_221; and.b32 %r1851, %r8, -4096; setp.ne.s32 %p137, %r1851, %r4875; @%p137 bra $L__BB5_221; and.b32 %r1852, %r74, -4096; setp.eq.s32 %p138, %r1852, %r4908; selp.u16 %rs470, 1, 0, %p138; $L__BB5_221: setp.eq.s16 %p139, %rs470, 0; @%p139 bra $L__BB5_227; shl.b32 %r1853, %r259, 3; and.b32 %r1854, %r1853, 31744; and.b32 %r1855, %r8, 3968; shr.u32 %r1856, %r1855, 2; or.b32 %r1857, %r1854, %r1856; bfe.u32 %r1858, %r74, 7, 5; or.b32 %r1859, %r1858, %r1857; shr.u32 %r1860, %r1857, 6; mul.wide.u32 %rd1062, %r1860, 8; add.s64 %rd1063, %rd2315, %rd1062; and.b32 %r1861, %r1859, 63; mov.u64 %rd1064, 1; shl.b64 %rd1065, %rd1064, %r1861; ld.u64 %rd1066, [%rd1063+4128]; and.b64 %rd1067, %rd1066, %rd1065; setp.eq.s64 %p140, %rd1067, 0; mul.wide.u32 %rd1068, %r1859, 8; add.s64 %rd1069, %rd2315, %rd1068; add.s64 %rd190, %rd1069, 8256; @%p140 bra $L__BB5_226; ld.u64 %rd1070, [%rd190]; add.s64 %rd2311, %rd2315, %rd1070; st.local.u32 [%rd2+12], %r372; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1862, %r259, 5; and.b32 %r1863, %r1862, 3840; shl.b32 %r1864, %r8, 1; and.b32 %r1865, %r1864, 240; or.b32 %r1866, %r1863, %r1865; bfe.u32 %r1867, %r74, 3, 4; or.b32 %r1868, %r1867, %r1866; shr.u32 %r1869, %r1866, 6; mul.wide.u32 %rd1071, %r1869, 8; add.s64 %rd1072, %rd2311, %rd1071; and.b32 %r1870, %r1868, 63; shl.b64 %rd1074, %rd1064, %r1870; ld.u64 %rd1075, [%rd1072+544]; and.b64 %rd1076, %rd1075, %rd1074; setp.eq.s64 %p141, %rd1076, 0; mul.wide.u32 %rd1077, %r1868, 8; add.s64 %rd1078, %rd2311, %rd1077; add.s64 %rd192, %rd1078, 1088; @%p141 bra $L__BB5_225; and.b32 %r4827, %r259, -8; ld.u64 %rd1079, [%rd192]; add.s64 %rd2307, %rd2311, %rd1079; and.b32 %r1871, %r8, -8; mov.u32 %r1872, 1; st.local.v2.u32 [%rd2], {%r4827, %r1871}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1873, %r259, 6; and.b32 %r1874, %r1873, 448; shl.b32 %r1875, %r8, 3; and.b32 %r1876, %r1875, 56; or.b32 %r1877, %r1874, %r1876; and.b32 %r1878, %r74, 7; or.b32 %r1879, %r1877, %r1878; ld.u8 %rs217, [%rd2307+15]; shr.u16 %rs218, %rs217, 5; mov.u32 %r1880, 5; cvt.u32.u16 %r1881, %rs218; sub.s32 %r1882, %r1880, %r1881; shr.u32 %r1883, %r1879, %r1882; mul.wide.u32 %rd1080, %r1883, 4; add.s64 %rd1081, %rd2307, %rd1080; mov.u32 %r1884, 32; shr.u32 %r1885, %r1884, %r1881; add.s32 %r1886, %r1885, -1; and.b32 %r1887, %r1886, %r1879; shl.b32 %r1888, %r1887, %r1881; ld.u32 %r1889, [%rd1081+96]; shr.u32 %r1890, %r1889, %r1888; shl.b32 %r1891, %r1872, %r1881; mov.u32 %r1892, -1; shl.b32 %r1893, %r1892, %r1891; not.b32 %r1894, %r1893; and.b32 %r1895, %r1890, %r1894; cvt.rn.f32.u32 %f519, %r1895; ld.v2.f32 {%f520, %f521}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f521, %f519, %f520; mov.b64 %rd1082, {%r4827, %r1871}; shr.u64 %rd1083, %rd1082, 32; cvt.u32.u64 %r4782, %rd1083; mov.u32 %r4870, %r372; bra.uni $L__BB5_241; $L__BB5_253: and.b32 %r426, %r259, -4096; setp.ne.s32 %p158, %r426, %r4830; mov.u16 %rs473, 0; @%p158 bra $L__BB5_256; and.b32 %r2005, %r8, -4096; setp.ne.s32 %p159, %r2005, %r4875; @%p159 bra $L__BB5_256; and.b32 %r2006, %r9, -4096; setp.eq.s32 %p160, %r2006, %r4908; selp.u16 %rs473, 1, 0, %p160; $L__BB5_256: setp.eq.s16 %p161, %rs473, 0; @%p161 bra $L__BB5_262; shl.b32 %r2007, %r259, 3; and.b32 %r2008, %r2007, 31744; and.b32 %r2009, %r8, 3968; shr.u32 %r2010, %r2009, 2; or.b32 %r2011, %r2008, %r2010; bfe.u32 %r2012, %r9, 7, 5; or.b32 %r2013, %r2011, %r2012; shr.u32 %r2014, %r2011, 6; mul.wide.u32 %rd1133, %r2014, 8; add.s64 %rd1134, %rd2315, %rd1133; and.b32 %r2015, %r2013, 63; mov.u64 %rd1135, 1; shl.b64 %rd1136, %rd1135, %r2015; ld.u64 %rd1137, [%rd1134+4128]; and.b64 %rd1138, %rd1137, %rd1136; setp.eq.s64 %p162, %rd1138, 0; mul.wide.u32 %rd1139, %r2013, 8; add.s64 %rd1140, %rd2315, %rd1139; add.s64 %rd219, %rd1140, 8256; @%p162 bra $L__BB5_261; ld.u64 %rd1141, [%rd219]; add.s64 %rd2311, %rd2315, %rd1141; st.local.u32 [%rd2+12], %r423; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2016, %r259, 5; and.b32 %r2017, %r2016, 3840; shl.b32 %r2018, %r8, 1; and.b32 %r2019, %r2018, 240; or.b32 %r2020, %r2017, %r2019; bfe.u32 %r2021, %r9, 3, 4; or.b32 %r2022, %r2020, %r2021; shr.u32 %r2023, %r2020, 6; mul.wide.u32 %rd1142, %r2023, 8; add.s64 %rd1143, %rd2311, %rd1142; and.b32 %r2024, %r2022, 63; shl.b64 %rd1145, %rd1135, %r2024; ld.u64 %rd1146, [%rd1143+544]; and.b64 %rd1147, %rd1146, %rd1145; setp.eq.s64 %p163, %rd1147, 0; mul.wide.u32 %rd1148, %r2022, 8; add.s64 %rd1149, %rd2311, %rd1148; add.s64 %rd221, %rd1149, 1088; @%p163 bra $L__BB5_260; and.b32 %r4827, %r259, -8; ld.u64 %rd1150, [%rd221]; add.s64 %rd2307, %rd2311, %rd1150; and.b32 %r2025, %r8, -8; mov.u32 %r2026, 1; st.local.v2.u32 [%rd2], {%r4827, %r2025}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2027, %r259, 6; and.b32 %r2028, %r2027, 448; shl.b32 %r2029, %r8, 3; and.b32 %r2030, %r2029, 56; or.b32 %r2031, %r2028, %r2030; and.b32 %r2032, %r9, 7; or.b32 %r2033, %r2031, %r2032; ld.u8 %rs231, [%rd2307+15]; shr.u16 %rs232, %rs231, 5; mov.u32 %r2034, 5; cvt.u32.u16 %r2035, %rs232; sub.s32 %r2036, %r2034, %r2035; shr.u32 %r2037, %r2033, %r2036; mul.wide.u32 %rd1151, %r2037, 4; add.s64 %rd1152, %rd2307, %rd1151; mov.u32 %r2038, 32; shr.u32 %r2039, %r2038, %r2035; add.s32 %r2040, %r2039, -1; and.b32 %r2041, %r2040, %r2033; shl.b32 %r2042, %r2041, %r2035; ld.u32 %r2043, [%rd1152+96]; shr.u32 %r2044, %r2043, %r2042; shl.b32 %r2045, %r2026, %r2035; mov.u32 %r2046, -1; shl.b32 %r2047, %r2046, %r2045; not.b32 %r2048, %r2047; and.b32 %r2049, %r2044, %r2048; cvt.rn.f32.u32 %f539, %r2049; ld.v2.f32 {%f540, %f541}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f541, %f539, %f540; mov.b64 %rd1153, {%r4827, %r2025}; shr.u64 %rd1154, %rd1153, 32; cvt.u32.u64 %r4782, %rd1154; mov.u32 %r4870, %r423; bra.uni $L__BB5_276; $L__BB5_288: and.b32 %r476, %r471, -4096; setp.ne.s32 %p180, %r476, %r4830; mov.u16 %rs476, 0; @%p180 bra $L__BB5_291; and.b32 %r2158, %r8, -4096; setp.ne.s32 %p181, %r2158, %r4875; @%p181 bra $L__BB5_291; and.b32 %r2159, %r9, -4096; setp.eq.s32 %p182, %r2159, %r4908; selp.u16 %rs476, 1, 0, %p182; $L__BB5_291: setp.eq.s16 %p183, %rs476, 0; @%p183 bra $L__BB5_297; shl.b32 %r2160, %r471, 3; and.b32 %r2161, %r2160, 31744; and.b32 %r2162, %r8, 3968; shr.u32 %r2163, %r2162, 2; or.b32 %r2164, %r2163, %r2161; bfe.u32 %r2165, %r9, 7, 5; or.b32 %r2166, %r2164, %r2165; shr.u32 %r2167, %r2164, 6; mul.wide.u32 %rd1203, %r2167, 8; add.s64 %rd1204, %rd2315, %rd1203; and.b32 %r2168, %r2166, 63; mov.u64 %rd1205, 1; shl.b64 %rd1206, %rd1205, %r2168; ld.u64 %rd1207, [%rd1204+4128]; and.b64 %rd1208, %rd1207, %rd1206; setp.eq.s64 %p184, %rd1208, 0; mul.wide.u32 %rd1209, %r2166, 8; add.s64 %rd1210, %rd2315, %rd1209; add.s64 %rd248, %rd1210, 8256; @%p184 bra $L__BB5_296; ld.u64 %rd1211, [%rd248]; add.s64 %rd2311, %rd2315, %rd1211; st.local.u32 [%rd2+12], %r473; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2169, %r471, 5; and.b32 %r2170, %r2169, 3840; shl.b32 %r2171, %r8, 1; and.b32 %r2172, %r2171, 240; or.b32 %r2173, %r2172, %r2170; bfe.u32 %r2174, %r9, 3, 4; or.b32 %r2175, %r2173, %r2174; shr.u32 %r2176, %r2173, 6; mul.wide.u32 %rd1212, %r2176, 8; add.s64 %rd1213, %rd2311, %rd1212; and.b32 %r2177, %r2175, 63; shl.b64 %rd1215, %rd1205, %r2177; ld.u64 %rd1216, [%rd1213+544]; and.b64 %rd1217, %rd1216, %rd1215; setp.eq.s64 %p185, %rd1217, 0; mul.wide.u32 %rd1218, %r2175, 8; add.s64 %rd1219, %rd2311, %rd1218; add.s64 %rd250, %rd1219, 1088; @%p185 bra $L__BB5_295; ld.u64 %rd1220, [%rd250]; add.s64 %rd2307, %rd2311, %rd1220; and.b32 %r2178, %r8, -8; mov.u32 %r2179, 1; st.local.v2.u32 [%rd2], {%r472, %r2178}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2180, %r471, 6; and.b32 %r2181, %r2180, 448; shl.b32 %r2182, %r8, 3; and.b32 %r2183, %r2182, 56; or.b32 %r2184, %r2183, %r2181; and.b32 %r2185, %r9, 7; or.b32 %r2186, %r2184, %r2185; ld.u8 %rs245, [%rd2307+15]; shr.u16 %rs246, %rs245, 5; mov.u32 %r2187, 5; cvt.u32.u16 %r2188, %rs246; sub.s32 %r2189, %r2187, %r2188; shr.u32 %r2190, %r2186, %r2189; mul.wide.u32 %rd1221, %r2190, 4; add.s64 %rd1222, %rd2307, %rd1221; mov.u32 %r2191, 32; shr.u32 %r2192, %r2191, %r2188; add.s32 %r2193, %r2192, -1; and.b32 %r2194, %r2193, %r2186; shl.b32 %r2195, %r2194, %r2188; ld.u32 %r2196, [%rd1222+96]; shr.u32 %r2197, %r2196, %r2195; shl.b32 %r2198, %r2179, %r2188; mov.u32 %r2199, -1; shl.b32 %r2200, %r2199, %r2198; not.b32 %r2201, %r2200; and.b32 %r2202, %r2197, %r2201; cvt.rn.f32.u32 %f566, %r2202; ld.v2.f32 {%f567, %f568}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f568, %f566, %f567; mov.b64 %rd1223, {%r472, %r2178}; shr.u64 %rd1224, %rd1223, 32; cvt.u32.u64 %r4782, %rd1224; mov.u32 %r4870, %r473; mov.u32 %r4827, %r472; bra.uni $L__BB5_311; $L__BB5_323: and.b32 %r521, %r471, -4096; setp.ne.s32 %p202, %r521, %r4830; mov.u16 %rs479, 0; @%p202 bra $L__BB5_326; and.b32 %r2311, %r8, -4096; setp.ne.s32 %p203, %r2311, %r4875; @%p203 bra $L__BB5_326; and.b32 %r2312, %r74, -4096; setp.eq.s32 %p204, %r2312, %r4908; selp.u16 %rs479, 1, 0, %p204; $L__BB5_326: setp.eq.s16 %p205, %rs479, 0; @%p205 bra $L__BB5_332; shl.b32 %r2313, %r471, 3; and.b32 %r2314, %r2313, 31744; and.b32 %r2315, %r8, 3968; shr.u32 %r2316, %r2315, 2; or.b32 %r2317, %r2316, %r2314; bfe.u32 %r2318, %r74, 7, 5; or.b32 %r2319, %r2318, %r2317; shr.u32 %r2320, %r2317, 6; mul.wide.u32 %rd1273, %r2320, 8; add.s64 %rd1274, %rd2315, %rd1273; and.b32 %r2321, %r2319, 63; mov.u64 %rd1275, 1; shl.b64 %rd1276, %rd1275, %r2321; ld.u64 %rd1277, [%rd1274+4128]; and.b64 %rd1278, %rd1277, %rd1276; setp.eq.s64 %p206, %rd1278, 0; mul.wide.u32 %rd1279, %r2319, 8; add.s64 %rd1280, %rd2315, %rd1279; add.s64 %rd277, %rd1280, 8256; @%p206 bra $L__BB5_331; ld.u64 %rd1281, [%rd277]; add.s64 %rd2311, %rd2315, %rd1281; st.local.u32 [%rd2+12], %r518; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2322, %r471, 5; and.b32 %r2323, %r2322, 3840; shl.b32 %r2324, %r8, 1; and.b32 %r2325, %r2324, 240; or.b32 %r2326, %r2325, %r2323; bfe.u32 %r2327, %r74, 3, 4; or.b32 %r2328, %r2327, %r2326; shr.u32 %r2329, %r2326, 6; mul.wide.u32 %rd1282, %r2329, 8; add.s64 %rd1283, %rd2311, %rd1282; and.b32 %r2330, %r2328, 63; shl.b64 %rd1285, %rd1275, %r2330; ld.u64 %rd1286, [%rd1283+544]; and.b64 %rd1287, %rd1286, %rd1285; setp.eq.s64 %p207, %rd1287, 0; mul.wide.u32 %rd1288, %r2328, 8; add.s64 %rd1289, %rd2311, %rd1288; add.s64 %rd279, %rd1289, 1088; @%p207 bra $L__BB5_330; ld.u64 %rd1290, [%rd279]; add.s64 %rd2307, %rd2311, %rd1290; and.b32 %r2331, %r8, -8; mov.u32 %r2332, 1; st.local.v2.u32 [%rd2], {%r472, %r2331}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2333, %r471, 6; and.b32 %r2334, %r2333, 448; shl.b32 %r2335, %r8, 3; and.b32 %r2336, %r2335, 56; or.b32 %r2337, %r2336, %r2334; and.b32 %r2338, %r74, 7; or.b32 %r2339, %r2337, %r2338; ld.u8 %rs259, [%rd2307+15]; shr.u16 %rs260, %rs259, 5; mov.u32 %r2340, 5; cvt.u32.u16 %r2341, %rs260; sub.s32 %r2342, %r2340, %r2341; shr.u32 %r2343, %r2339, %r2342; mul.wide.u32 %rd1291, %r2343, 4; add.s64 %rd1292, %rd2307, %rd1291; mov.u32 %r2344, 32; shr.u32 %r2345, %r2344, %r2341; add.s32 %r2346, %r2345, -1; and.b32 %r2347, %r2346, %r2339; shl.b32 %r2348, %r2347, %r2341; ld.u32 %r2349, [%rd1292+96]; shr.u32 %r2350, %r2349, %r2348; shl.b32 %r2351, %r2332, %r2341; mov.u32 %r2352, -1; shl.b32 %r2353, %r2352, %r2351; not.b32 %r2354, %r2353; and.b32 %r2355, %r2350, %r2354; cvt.rn.f32.u32 %f586, %r2355; ld.v2.f32 {%f587, %f588}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f588, %f586, %f587; mov.b64 %rd1293, {%r472, %r2331}; shr.u64 %rd1294, %rd1293, 32; cvt.u32.u64 %r4782, %rd1294; mov.u32 %r4870, %r518; mov.u32 %r4827, %r472; bra.uni $L__BB5_346; $L__BB5_358: and.b32 %r563, %r471, -4096; setp.ne.s32 %p224, %r563, %r4830; mov.u16 %rs482, 0; @%p224 bra $L__BB5_361; and.b32 %r2465, %r139, -4096; setp.ne.s32 %p225, %r2465, %r4875; @%p225 bra $L__BB5_361; and.b32 %r2466, %r74, -4096; setp.eq.s32 %p226, %r2466, %r4908; selp.u16 %rs482, 1, 0, %p226; $L__BB5_361: setp.eq.s16 %p227, %rs482, 0; @%p227 bra $L__BB5_367; shl.b32 %r2467, %r471, 3; and.b32 %r2468, %r2467, 31744; and.b32 %r2469, %r139, 3968; shr.u32 %r2470, %r2469, 2; or.b32 %r2471, %r2470, %r2468; bfe.u32 %r2472, %r74, 7, 5; or.b32 %r2473, %r2471, %r2472; shr.u32 %r2474, %r2471, 6; mul.wide.u32 %rd1344, %r2474, 8; add.s64 %rd1345, %rd2315, %rd1344; and.b32 %r2475, %r2473, 63; mov.u64 %rd1346, 1; shl.b64 %rd1347, %rd1346, %r2475; ld.u64 %rd1348, [%rd1345+4128]; and.b64 %rd1349, %rd1348, %rd1347; setp.eq.s64 %p228, %rd1349, 0; mul.wide.u32 %rd1350, %r2473, 8; add.s64 %rd1351, %rd2315, %rd1350; add.s64 %rd306, %rd1351, 8256; @%p228 bra $L__BB5_366; ld.u64 %rd1352, [%rd306]; add.s64 %rd2311, %rd2315, %rd1352; st.local.u32 [%rd2+12], %r560; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2476, %r471, 5; and.b32 %r2477, %r2476, 3840; shl.b32 %r2478, %r139, 1; and.b32 %r2479, %r2478, 240; or.b32 %r2480, %r2479, %r2477; bfe.u32 %r2481, %r74, 3, 4; or.b32 %r2482, %r2480, %r2481; shr.u32 %r2483, %r2480, 6; mul.wide.u32 %rd1353, %r2483, 8; add.s64 %rd1354, %rd2311, %rd1353; and.b32 %r2484, %r2482, 63; shl.b64 %rd1356, %rd1346, %r2484; ld.u64 %rd1357, [%rd1354+544]; and.b64 %rd1358, %rd1357, %rd1356; setp.eq.s64 %p229, %rd1358, 0; mul.wide.u32 %rd1359, %r2482, 8; add.s64 %rd1360, %rd2311, %rd1359; add.s64 %rd308, %rd1360, 1088; @%p229 bra $L__BB5_365; ld.u64 %rd1361, [%rd308]; add.s64 %rd2307, %rd2311, %rd1361; and.b32 %r2485, %r139, -8; mov.u32 %r2486, 1; st.local.v2.u32 [%rd2], {%r472, %r2485}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2487, %r471, 6; and.b32 %r2488, %r2487, 448; shl.b32 %r2489, %r139, 3; and.b32 %r2490, %r2489, 56; or.b32 %r2491, %r2490, %r2488; and.b32 %r2492, %r74, 7; or.b32 %r2493, %r2491, %r2492; ld.u8 %rs273, [%rd2307+15]; shr.u16 %rs274, %rs273, 5; mov.u32 %r2494, 5; cvt.u32.u16 %r2495, %rs274; sub.s32 %r2496, %r2494, %r2495; shr.u32 %r2497, %r2493, %r2496; mul.wide.u32 %rd1362, %r2497, 4; add.s64 %rd1363, %rd2307, %rd1362; mov.u32 %r2498, 32; shr.u32 %r2499, %r2498, %r2495; add.s32 %r2500, %r2499, -1; and.b32 %r2501, %r2500, %r2493; shl.b32 %r2502, %r2501, %r2495; ld.u32 %r2503, [%rd1363+96]; shr.u32 %r2504, %r2503, %r2502; shl.b32 %r2505, %r2486, %r2495; mov.u32 %r2506, -1; shl.b32 %r2507, %r2506, %r2505; not.b32 %r2508, %r2507; and.b32 %r2509, %r2504, %r2508; cvt.rn.f32.u32 %f607, %r2509; ld.v2.f32 {%f608, %f609}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f609, %f607, %f608; mov.b64 %rd1364, {%r472, %r2485}; shr.u64 %rd1365, %rd1364, 32; cvt.u32.u64 %r4782, %rd1365; mov.u32 %r4870, %r560; mov.u32 %r4827, %r472; bra.uni $L__BB5_381; $L__BB5_393: and.b32 %r602, %r471, -4096; setp.ne.s32 %p246, %r602, %r4830; mov.u16 %rs485, 0; @%p246 bra $L__BB5_396; and.b32 %r2618, %r139, -4096; setp.ne.s32 %p247, %r2618, %r4875; @%p247 bra $L__BB5_396; and.b32 %r2619, %r9, -4096; setp.eq.s32 %p248, %r2619, %r4908; selp.u16 %rs485, 1, 0, %p248; $L__BB5_396: setp.eq.s16 %p249, %rs485, 0; @%p249 bra $L__BB5_402; shl.b32 %r2620, %r471, 3; and.b32 %r2621, %r2620, 31744; and.b32 %r2622, %r139, 3968; shr.u32 %r2623, %r2622, 2; or.b32 %r2624, %r2623, %r2621; bfe.u32 %r2625, %r9, 7, 5; or.b32 %r2626, %r2624, %r2625; shr.u32 %r2627, %r2624, 6; mul.wide.u32 %rd1415, %r2627, 8; add.s64 %rd1416, %rd2315, %rd1415; and.b32 %r2628, %r2626, 63; mov.u64 %rd1417, 1; shl.b64 %rd1418, %rd1417, %r2628; ld.u64 %rd1419, [%rd1416+4128]; and.b64 %rd1420, %rd1419, %rd1418; setp.eq.s64 %p250, %rd1420, 0; mul.wide.u32 %rd1421, %r2626, 8; add.s64 %rd1422, %rd2315, %rd1421; add.s64 %rd335, %rd1422, 8256; @%p250 bra $L__BB5_401; ld.u64 %rd1423, [%rd335]; add.s64 %rd2311, %rd2315, %rd1423; st.local.u32 [%rd2+12], %r599; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2629, %r471, 5; and.b32 %r2630, %r2629, 3840; shl.b32 %r2631, %r139, 1; and.b32 %r2632, %r2631, 240; or.b32 %r2633, %r2632, %r2630; bfe.u32 %r2634, %r9, 3, 4; or.b32 %r2635, %r2633, %r2634; shr.u32 %r2636, %r2633, 6; mul.wide.u32 %rd1424, %r2636, 8; add.s64 %rd1425, %rd2311, %rd1424; and.b32 %r2637, %r2635, 63; shl.b64 %rd1427, %rd1417, %r2637; ld.u64 %rd1428, [%rd1425+544]; and.b64 %rd1429, %rd1428, %rd1427; setp.eq.s64 %p251, %rd1429, 0; mul.wide.u32 %rd1430, %r2635, 8; add.s64 %rd1431, %rd2311, %rd1430; add.s64 %rd337, %rd1431, 1088; @%p251 bra $L__BB5_400; ld.u64 %rd1432, [%rd337]; add.s64 %rd2307, %rd2311, %rd1432; and.b32 %r2638, %r139, -8; mov.u32 %r2639, 1; st.local.v2.u32 [%rd2], {%r472, %r2638}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2640, %r471, 7; shl.b32 %r2641, %r139, 3; and.b32 %r2642, %r2641, 56; and.b32 %r2643, %r9, 7; bfi.b32 %r2644, %r2640, %r2643, 6, 3; or.b32 %r2645, %r2644, %r2642; ld.u8 %rs287, [%rd2307+15]; shr.u16 %rs288, %rs287, 5; mov.u32 %r2646, 5; cvt.u32.u16 %r2647, %rs288; sub.s32 %r2648, %r2646, %r2647; shr.u32 %r2649, %r2645, %r2648; mul.wide.u32 %rd1433, %r2649, 4; add.s64 %rd1434, %rd2307, %rd1433; mov.u32 %r2650, 32; shr.u32 %r2651, %r2650, %r2647; add.s32 %r2652, %r2651, -1; and.b32 %r2653, %r2652, %r2645; shl.b32 %r2654, %r2653, %r2647; ld.u32 %r2655, [%rd1434+96]; shr.u32 %r2656, %r2655, %r2654; shl.b32 %r2657, %r2639, %r2647; mov.u32 %r2658, -1; shl.b32 %r2659, %r2658, %r2657; not.b32 %r2660, %r2659; and.b32 %r2661, %r2656, %r2660; cvt.rn.f32.u32 %f627, %r2661; ld.v2.f32 {%f628, %f629}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f629, %f627, %f628; mov.b64 %rd1435, {%r472, %r2638}; shr.u64 %rd1436, %rd1435, 32; cvt.u32.u64 %r4782, %rd1436; mov.u32 %r4870, %r599; mov.u32 %r4827, %r472; bra.uni $L__BB5_416; $L__BB5_428: and.b32 %r640, %r635, -4096; setp.ne.s32 %p268, %r640, %r4830; mov.u16 %rs488, 0; @%p268 bra $L__BB5_431; and.b32 %r2768, %r139, -4096; setp.ne.s32 %p269, %r2768, %r4875; @%p269 bra $L__BB5_431; and.b32 %r2769, %r9, -4096; setp.eq.s32 %p270, %r2769, %r4908; selp.u16 %rs488, 1, 0, %p270; $L__BB5_431: setp.eq.s16 %p271, %rs488, 0; @%p271 bra $L__BB5_437; shl.b32 %r2770, %r635, 3; and.b32 %r2771, %r2770, 31744; and.b32 %r2772, %r139, 3968; shr.u32 %r2773, %r2772, 2; or.b32 %r2774, %r2773, %r2771; bfe.u32 %r2775, %r9, 7, 5; or.b32 %r2776, %r2774, %r2775; shr.u32 %r2777, %r2774, 6; mul.wide.u32 %rd1486, %r2777, 8; add.s64 %rd1487, %rd2315, %rd1486; and.b32 %r2778, %r2776, 63; mov.u64 %rd1488, 1; shl.b64 %rd1489, %rd1488, %r2778; ld.u64 %rd1490, [%rd1487+4128]; and.b64 %rd1491, %rd1490, %rd1489; setp.eq.s64 %p272, %rd1491, 0; mul.wide.u32 %rd1492, %r2776, 8; add.s64 %rd1493, %rd2315, %rd1492; add.s64 %rd364, %rd1493, 8256; @%p272 bra $L__BB5_436; ld.u64 %rd1494, [%rd364]; add.s64 %rd2311, %rd2315, %rd1494; st.local.u32 [%rd2+12], %r637; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2779, %r635, 5; and.b32 %r2780, %r2779, 3840; shl.b32 %r2781, %r139, 1; and.b32 %r2782, %r2781, 240; or.b32 %r2783, %r2782, %r2780; bfe.u32 %r2784, %r9, 3, 4; or.b32 %r2785, %r2783, %r2784; shr.u32 %r2786, %r2783, 6; mul.wide.u32 %rd1495, %r2786, 8; add.s64 %rd1496, %rd2311, %rd1495; and.b32 %r2787, %r2785, 63; shl.b64 %rd1498, %rd1488, %r2787; ld.u64 %rd1499, [%rd1496+544]; and.b64 %rd1500, %rd1499, %rd1498; setp.eq.s64 %p273, %rd1500, 0; mul.wide.u32 %rd1501, %r2785, 8; add.s64 %rd1502, %rd2311, %rd1501; add.s64 %rd366, %rd1502, 1088; @%p273 bra $L__BB5_435; ld.u64 %rd1503, [%rd366]; add.s64 %rd2307, %rd2311, %rd1503; and.b32 %r2788, %r139, -8; mov.u32 %r2789, 1; st.local.v2.u32 [%rd2], {%r636, %r2788}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2790, %r635, 7; shl.b32 %r2791, %r139, 3; and.b32 %r2792, %r2791, 56; and.b32 %r2793, %r9, 7; bfi.b32 %r2794, %r2790, %r2793, 6, 3; or.b32 %r2795, %r2794, %r2792; ld.u8 %rs301, [%rd2307+15]; shr.u16 %rs302, %rs301, 5; mov.u32 %r2796, 5; cvt.u32.u16 %r2797, %rs302; sub.s32 %r2798, %r2796, %r2797; shr.u32 %r2799, %r2795, %r2798; mul.wide.u32 %rd1504, %r2799, 4; add.s64 %rd1505, %rd2307, %rd1504; mov.u32 %r2800, 32; shr.u32 %r2801, %r2800, %r2797; add.s32 %r2802, %r2801, -1; and.b32 %r2803, %r2802, %r2795; shl.b32 %r2804, %r2803, %r2797; ld.u32 %r2805, [%rd1505+96]; shr.u32 %r2806, %r2805, %r2804; shl.b32 %r2807, %r2789, %r2797; mov.u32 %r2808, -1; shl.b32 %r2809, %r2808, %r2807; not.b32 %r2810, %r2809; and.b32 %r2811, %r2806, %r2810; cvt.rn.f32.u32 %f650, %r2811; ld.v2.f32 {%f651, %f652}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f652, %f650, %f651; mov.b64 %rd1506, {%r636, %r2788}; shr.u64 %rd1507, %rd1506, 32; cvt.u32.u64 %r4782, %rd1507; mov.u32 %r4870, %r637; mov.u32 %r4827, %r636; bra.uni $L__BB5_451; $L__BB5_463: and.b32 %r673, %r635, -4096; setp.ne.s32 %p290, %r673, %r4830; mov.u16 %rs491, 0; @%p290 bra $L__BB5_466; and.b32 %r2920, %r139, -4096; setp.ne.s32 %p291, %r2920, %r4875; @%p291 bra $L__BB5_466; and.b32 %r2921, %r74, -4096; setp.eq.s32 %p292, %r2921, %r4908; selp.u16 %rs491, 1, 0, %p292; $L__BB5_466: setp.eq.s16 %p293, %rs491, 0; @%p293 bra $L__BB5_472; shl.b32 %r2922, %r635, 3; and.b32 %r2923, %r2922, 31744; and.b32 %r2924, %r139, 3968; shr.u32 %r2925, %r2924, 2; or.b32 %r2926, %r2925, %r2923; bfe.u32 %r2927, %r74, 7, 5; or.b32 %r2928, %r2926, %r2927; shr.u32 %r2929, %r2926, 6; mul.wide.u32 %rd1557, %r2929, 8; add.s64 %rd1558, %rd2315, %rd1557; and.b32 %r2930, %r2928, 63; mov.u64 %rd1559, 1; shl.b64 %rd1560, %rd1559, %r2930; ld.u64 %rd1561, [%rd1558+4128]; and.b64 %rd1562, %rd1561, %rd1560; setp.eq.s64 %p294, %rd1562, 0; mul.wide.u32 %rd1563, %r2928, 8; add.s64 %rd1564, %rd2315, %rd1563; add.s64 %rd393, %rd1564, 8256; @%p294 bra $L__BB5_471; ld.u64 %rd1565, [%rd393]; add.s64 %rd2311, %rd2315, %rd1565; st.local.u32 [%rd2+12], %r670; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2931, %r635, 5; and.b32 %r2932, %r2931, 3840; shl.b32 %r2933, %r139, 1; and.b32 %r2934, %r2933, 240; or.b32 %r2935, %r2934, %r2932; bfe.u32 %r2936, %r74, 3, 4; or.b32 %r2937, %r2935, %r2936; shr.u32 %r2938, %r2935, 6; mul.wide.u32 %rd1566, %r2938, 8; add.s64 %rd1567, %rd2311, %rd1566; and.b32 %r2939, %r2937, 63; shl.b64 %rd1569, %rd1559, %r2939; ld.u64 %rd1570, [%rd1567+544]; and.b64 %rd1571, %rd1570, %rd1569; setp.eq.s64 %p295, %rd1571, 0; mul.wide.u32 %rd1572, %r2937, 8; add.s64 %rd1573, %rd2311, %rd1572; add.s64 %rd395, %rd1573, 1088; @%p295 bra $L__BB5_470; ld.u64 %rd1574, [%rd395]; add.s64 %rd2307, %rd2311, %rd1574; and.b32 %r2940, %r139, -8; mov.u32 %r2941, 1; st.local.v2.u32 [%rd2], {%r636, %r2940}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2942, %r635, 6; and.b32 %r2943, %r2942, 448; shl.b32 %r2944, %r139, 3; and.b32 %r2945, %r2944, 56; or.b32 %r2946, %r2945, %r2943; and.b32 %r2947, %r74, 7; or.b32 %r2948, %r2946, %r2947; ld.u8 %rs315, [%rd2307+15]; shr.u16 %rs316, %rs315, 5; mov.u32 %r2949, 5; cvt.u32.u16 %r2950, %rs316; sub.s32 %r2951, %r2949, %r2950; shr.u32 %r2952, %r2948, %r2951; mul.wide.u32 %rd1575, %r2952, 4; add.s64 %rd1576, %rd2307, %rd1575; mov.u32 %r2953, 32; shr.u32 %r2954, %r2953, %r2950; add.s32 %r2955, %r2954, -1; and.b32 %r2956, %r2955, %r2948; shl.b32 %r2957, %r2956, %r2950; ld.u32 %r2958, [%rd1576+96]; shr.u32 %r2959, %r2958, %r2957; shl.b32 %r2960, %r2941, %r2950; mov.u32 %r2961, -1; shl.b32 %r2962, %r2961, %r2960; not.b32 %r2963, %r2962; and.b32 %r2964, %r2959, %r2963; cvt.rn.f32.u32 %f670, %r2964; ld.v2.f32 {%f671, %f672}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f672, %f670, %f671; mov.b64 %rd1577, {%r636, %r2940}; shr.u64 %rd1578, %rd1577, 32; cvt.u32.u64 %r4782, %rd1578; mov.u32 %r4870, %r670; mov.u32 %r4827, %r636; bra.uni $L__BB5_486; $L__BB5_498: and.b32 %r703, %r635, -4096; setp.ne.s32 %p312, %r703, %r4830; mov.u16 %rs494, 0; @%p312 bra $L__BB5_501; and.b32 %r3075, %r8, -4096; setp.ne.s32 %p313, %r3075, %r4875; @%p313 bra $L__BB5_501; and.b32 %r3076, %r74, -4096; setp.eq.s32 %p314, %r3076, %r4908; selp.u16 %rs494, 1, 0, %p314; $L__BB5_501: setp.eq.s16 %p315, %rs494, 0; @%p315 bra $L__BB5_507; shl.b32 %r3077, %r635, 3; and.b32 %r3078, %r3077, 31744; and.b32 %r3079, %r8, 3968; shr.u32 %r3080, %r3079, 2; or.b32 %r3081, %r3078, %r3080; bfe.u32 %r3082, %r74, 7, 5; or.b32 %r3083, %r3082, %r3081; shr.u32 %r3084, %r3081, 6; mul.wide.u32 %rd1628, %r3084, 8; add.s64 %rd1629, %rd2315, %rd1628; and.b32 %r3085, %r3083, 63; mov.u64 %rd1630, 1; shl.b64 %rd1631, %rd1630, %r3085; ld.u64 %rd1632, [%rd1629+4128]; and.b64 %rd1633, %rd1632, %rd1631; setp.eq.s64 %p316, %rd1633, 0; mul.wide.u32 %rd1634, %r3083, 8; add.s64 %rd1635, %rd2315, %rd1634; add.s64 %rd422, %rd1635, 8256; @%p316 bra $L__BB5_506; ld.u64 %rd1636, [%rd422]; add.s64 %rd2311, %rd2315, %rd1636; st.local.u32 [%rd2+12], %r700; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r3086, %r635, 5; and.b32 %r3087, %r3086, 3840; shl.b32 %r3088, %r8, 1; and.b32 %r3089, %r3088, 240; or.b32 %r3090, %r3087, %r3089; bfe.u32 %r3091, %r74, 3, 4; or.b32 %r3092, %r3091, %r3090; shr.u32 %r3093, %r3090, 6; mul.wide.u32 %rd1637, %r3093, 8; add.s64 %rd1638, %rd2311, %rd1637; and.b32 %r3094, %r3092, 63; shl.b64 %rd1640, %rd1630, %r3094; ld.u64 %rd1641, [%rd1638+544]; and.b64 %rd1642, %rd1641, %rd1640; setp.eq.s64 %p317, %rd1642, 0; mul.wide.u32 %rd1643, %r3092, 8; add.s64 %rd1644, %rd2311, %rd1643; add.s64 %rd424, %rd1644, 1088; @%p317 bra $L__BB5_505; ld.u64 %rd1645, [%rd424]; add.s64 %rd2307, %rd2311, %rd1645; and.b32 %r3095, %r8, -8; mov.u32 %r3096, 1; st.local.v2.u32 [%rd2], {%r636, %r3095}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r3097, %r635, 6; and.b32 %r3098, %r3097, 448; shl.b32 %r3099, %r8, 3; and.b32 %r3100, %r3099, 56; or.b32 %r3101, %r3098, %r3100; and.b32 %r3102, %r74, 7; or.b32 %r3103, %r3101, %r3102; ld.u8 %rs329, [%rd2307+15]; shr.u16 %rs330, %rs329, 5; mov.u32 %r3104, 5; cvt.u32.u16 %r3105, %rs330; sub.s32 %r3106, %r3104, %r3105; shr.u32 %r3107, %r3103, %r3106; mul.wide.u32 %rd1646, %r3107, 4; add.s64 %rd1647, %rd2307, %rd1646; mov.u32 %r3108, 32; shr.u32 %r3109, %r3108, %r3105; add.s32 %r3110, %r3109, -1; and.b32 %r3111, %r3110, %r3103; shl.b32 %r3112, %r3111, %r3105; ld.u32 %r3113, [%rd1647+96]; shr.u32 %r3114, %r3113, %r3112; shl.b32 %r3115, %r3096, %r3105; mov.u32 %r3116, -1; shl.b32 %r3117, %r3116, %r3115; not.b32 %r3118, %r3117; and.b32 %r3119, %r3114, %r3118; cvt.rn.f32.u32 %f691, %r3119; ld.v2.f32 {%f692, %f693}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f693, %f691, %f692; mov.b64 %rd1648, {%r636, %r3095}; shr.u64 %rd1649, %rd1648, 32; cvt.u32.u64 %r4782, %rd1649; mov.u32 %r4870, %r700; mov.u32 %r4827, %r636; bra.uni $L__BB5_521; $L__BB5_533: and.b32 %r728, %r635, -4096; setp.ne.s32 %p334, %r728, %r4830; mov.u16 %rs497, 0; @%p334 bra $L__BB5_536; and.b32 %r3230, %r8, -4096; setp.ne.s32 %p335, %r3230, %r4875; @%p335 bra $L__BB5_536; and.b32 %r3231, %r9, -4096; setp.eq.s32 %p336, %r3231, %r4908; selp.u16 %rs497, 1, 0, %p336; $L__BB5_536: setp.eq.s16 %p337, %rs497, 0; @%p337 bra $L__BB5_542; shl.b32 %r3232, %r635, 3; and.b32 %r3233, %r3232, 31744; and.b32 %r3234, %r8, 3968; shr.u32 %r3235, %r3234, 2; or.b32 %r3236, %r3233, %r3235; bfe.u32 %r3237, %r9, 7, 5; or.b32 %r3238, %r3236, %r3237; shr.u32 %r3239, %r3236, 6; mul.wide.u32 %rd1698, %r3239, 8; add.s64 %rd1699, %rd2315, %rd1698; and.b32 %r3240, %r3238, 63; mov.u64 %rd1700, 1; shl.b64 %rd1701, %rd1700, %r3240; ld.u64 %rd1702, [%rd1699+4128]; and.b64 %rd1703, %rd1702, %rd1701; setp.eq.s64 %p338, %rd1703, 0; mul.wide.u32 %rd1704, %r3238, 8; add.s64 %rd1705, %rd2315, %rd1704; add.s64 %rd450, %rd1705, 8256; @%p338 bra $L__BB5_541; ld.u64 %rd1706, [%rd450]; add.s64 %rd451, %rd2315, %rd1706; st.local.u32 [%rd2+12], %r727; and.b32 %r3241, %r8, -128; st.local.u32 [%rd2+16], %r3241; and.b32 %r3242, %r9, -128; st.local.u32 [%rd2+20], %r3242; st.local.u64 [%rd2+56], %rd451; shl.b32 %r3243, %r635, 5; and.b32 %r3244, %r3243, 3840; shl.b32 %r3245, %r8, 1; and.b32 %r3246, %r3245, 240; or.b32 %r3247, %r3244, %r3246; bfe.u32 %r3248, %r9, 3, 4; or.b32 %r3249, %r3247, %r3248; shr.u32 %r3250, %r3247, 6; mul.wide.u32 %rd1707, %r3250, 8; add.s64 %rd1708, %rd451, %rd1707; and.b32 %r3251, %r3249, 63; shl.b64 %rd1710, %rd1700, %r3251; ld.u64 %rd1711, [%rd1708+544]; and.b64 %rd1712, %rd1711, %rd1710; setp.eq.s64 %p339, %rd1712, 0; mul.wide.u32 %rd1713, %r3249, 8; add.s64 %rd1714, %rd451, %rd1713; add.s64 %rd452, %rd1714, 1088; @%p339 bra $L__BB5_540; ld.u64 %rd1715, [%rd452]; add.s64 %rd1716, %rd451, %rd1715; and.b32 %r3252, %r8, -8; mov.u32 %r3253, 1; st.local.v2.u32 [%rd2], {%r636, %r3252}; and.b32 %r3254, %r9, -8; st.local.u32 [%rd2+8], %r3254; st.local.u64 [%rd2+48], %rd1716; shl.b32 %r3255, %r635, 6; and.b32 %r3256, %r3255, 448; shl.b32 %r3257, %r8, 3; and.b32 %r3258, %r3257, 56; or.b32 %r3259, %r3256, %r3258; and.b32 %r3260, %r9, 7; or.b32 %r3261, %r3259, %r3260; ld.u8 %rs343, [%rd1716+15]; shr.u16 %rs344, %rs343, 5; mov.u32 %r3262, 5; cvt.u32.u16 %r3263, %rs344; sub.s32 %r3264, %r3262, %r3263; shr.u32 %r3265, %r3261, %r3264; mul.wide.u32 %rd1717, %r3265, 4; add.s64 %rd1718, %rd1716, %rd1717; mov.u32 %r3266, 32; shr.u32 %r3267, %r3266, %r3263; add.s32 %r3268, %r3267, -1; and.b32 %r3269, %r3268, %r3261; shl.b32 %r3270, %r3269, %r3263; ld.u32 %r3271, [%rd1718+96]; shr.u32 %r3272, %r3271, %r3270; shl.b32 %r3273, %r3253, %r3263; mov.u32 %r3274, -1; shl.b32 %r3275, %r3274, %r3273; not.b32 %r3276, %r3275; and.b32 %r3277, %r3272, %r3276; cvt.rn.f32.u32 %f711, %r3277; ld.v2.f32 {%f712, %f713}, [%rd1716+80]; fma.rn.ftz.f32 %f944, %f713, %f711, %f712; bra.uni $L__BB5_556; $L__BB5_568: and.b32 %r736, %r471, -4096; ld.local.u32 %r3396, [%rd2+24]; setp.ne.s32 %p356, %r736, %r3396; mov.u16 %rs500, 0; @%p356 bra $L__BB5_571; and.b32 %r3397, %r732, -4096; ld.local.u32 %r3398, [%rd2+28]; setp.ne.s32 %p357, %r3397, %r3398; @%p357 bra $L__BB5_571; and.b32 %r3399, %r733, -4096; ld.local.u32 %r3400, [%rd2+32]; setp.eq.s32 %p358, %r3399, %r3400; selp.u16 %rs500, 1, 0, %p358; $L__BB5_571: setp.eq.s16 %p359, %rs500, 0; @%p359 bra $L__BB5_577; shl.b32 %r3401, %r471, 3; and.b32 %r3402, %r3401, 31744; and.b32 %r3403, %r732, 3968; shr.u32 %r3404, %r3403, 2; or.b32 %r3405, %r3404, %r3402; bfe.u32 %r3406, %r733, 7, 5; or.b32 %r3407, %r3405, %r3406; shr.u32 %r3408, %r3405, 6; ld.local.u64 %rd466, [%rd2+64]; mul.wide.u32 %rd1764, %r3408, 8; add.s64 %rd1765, %rd466, %rd1764; and.b32 %r3409, %r3407, 63; mov.u64 %rd1766, 1; shl.b64 %rd1767, %rd1766, %r3409; ld.u64 %rd1768, [%rd1765+4128]; and.b64 %rd1769, %rd1768, %rd1767; setp.eq.s64 %p360, %rd1769, 0; mul.wide.u32 %rd1770, %r3407, 8; add.s64 %rd1771, %rd466, %rd1770; add.s64 %rd467, %rd1771, 8256; @%p360 bra $L__BB5_576; ld.u64 %rd1772, [%rd467]; add.s64 %rd468, %rd466, %rd1772; st.local.u32 [%rd2+12], %r735; and.b32 %r3410, %r732, -128; st.local.u32 [%rd2+16], %r3410; and.b32 %r3411, %r733, -128; st.local.u32 [%rd2+20], %r3411; st.local.u64 [%rd2+56], %rd468; shl.b32 %r3412, %r471, 5; and.b32 %r3413, %r3412, 3840; shl.b32 %r3414, %r732, 1; and.b32 %r3415, %r3414, 240; or.b32 %r3416, %r3415, %r3413; bfe.u32 %r3417, %r733, 3, 4; or.b32 %r3418, %r3416, %r3417; shr.u32 %r3419, %r3416, 6; mul.wide.u32 %rd1773, %r3419, 8; add.s64 %rd1774, %rd468, %rd1773; and.b32 %r3420, %r3418, 63; shl.b64 %rd1776, %rd1766, %r3420; ld.u64 %rd1777, [%rd1774+544]; and.b64 %rd1778, %rd1777, %rd1776; setp.eq.s64 %p361, %rd1778, 0; mul.wide.u32 %rd1779, %r3418, 8; add.s64 %rd1780, %rd468, %rd1779; add.s64 %rd469, %rd1780, 1088; @%p361 bra $L__BB5_575; ld.u64 %rd1781, [%rd469]; add.s64 %rd1782, %rd468, %rd1781; and.b32 %r3421, %r732, -8; mov.u32 %r3422, 1; st.local.v2.u32 [%rd2], {%r472, %r3421}; and.b32 %r3423, %r733, -8; st.local.u32 [%rd2+8], %r3423; st.local.u64 [%rd2+48], %rd1782; shl.b32 %r3424, %r471, 6; and.b32 %r3425, %r3424, 448; shl.b32 %r3426, %r732, 3; and.b32 %r3427, %r3426, 56; or.b32 %r3428, %r3427, %r3425; and.b32 %r3429, %r733, 7; or.b32 %r3430, %r3428, %r3429; ld.u8 %rs357, [%rd1782+15]; shr.u16 %rs358, %rs357, 5; mov.u32 %r3431, 5; cvt.u32.u16 %r3432, %rs358; sub.s32 %r3433, %r3431, %r3432; shr.u32 %r3434, %r3430, %r3433; mul.wide.u32 %rd1783, %r3434, 4; add.s64 %rd1784, %rd1782, %rd1783; mov.u32 %r3435, 32; shr.u32 %r3436, %r3435, %r3432; add.s32 %r3437, %r3436, -1; and.b32 %r3438, %r3437, %r3430; shl.b32 %r3439, %r3438, %r3432; ld.u32 %r3440, [%rd1784+96]; shr.u32 %r3441, %r3440, %r3439; shl.b32 %r3442, %r3422, %r3432; mov.u32 %r3443, -1; shl.b32 %r3444, %r3443, %r3442; not.b32 %r3445, %r3444; and.b32 %r3446, %r3441, %r3445; cvt.rn.f32.u32 %f752, %r3446; ld.v2.f32 {%f753, %f754}, [%rd1782+80]; fma.rn.ftz.f32 %f945, %f754, %f752, %f753; mov.u32 %r5196, %r472; bra.uni $L__BB5_591; $L__BB5_603: and.b32 %r743, %r471, -4096; ld.local.u32 %r3567, [%rd2+24]; setp.ne.s32 %p378, %r743, %r3567; mov.u16 %rs503, 0; @%p378 bra $L__BB5_606; and.b32 %r3568, %r732, -4096; ld.local.u32 %r3569, [%rd2+28]; setp.ne.s32 %p379, %r3568, %r3569; @%p379 bra $L__BB5_606; and.b32 %r3570, %r741, -4096; ld.local.u32 %r3571, [%rd2+32]; setp.eq.s32 %p380, %r3570, %r3571; selp.u16 %rs503, 1, 0, %p380; $L__BB5_606: setp.eq.s16 %p381, %rs503, 0; @%p381 bra $L__BB5_612; shl.b32 %r3572, %r471, 3; and.b32 %r3573, %r3572, 31744; and.b32 %r3574, %r732, 3968; shr.u32 %r3575, %r3574, 2; or.b32 %r3576, %r3575, %r3573; bfe.u32 %r3577, %r741, 7, 5; or.b32 %r3578, %r3577, %r3576; shr.u32 %r3579, %r3576, 6; ld.local.u64 %rd484, [%rd2+64]; mul.wide.u32 %rd1831, %r3579, 8; add.s64 %rd1832, %rd484, %rd1831; and.b32 %r3580, %r3578, 63; mov.u64 %rd1833, 1; shl.b64 %rd1834, %rd1833, %r3580; ld.u64 %rd1835, [%rd1832+4128]; and.b64 %rd1836, %rd1835, %rd1834; setp.eq.s64 %p382, %rd1836, 0; mul.wide.u32 %rd1837, %r3578, 8; add.s64 %rd1838, %rd484, %rd1837; add.s64 %rd485, %rd1838, 8256; @%p382 bra $L__BB5_611; ld.u64 %rd1839, [%rd485]; add.s64 %rd486, %rd484, %rd1839; st.local.u32 [%rd2+12], %r742; and.b32 %r3581, %r732, -128; st.local.u32 [%rd2+16], %r3581; and.b32 %r3582, %r741, -128; st.local.u32 [%rd2+20], %r3582; st.local.u64 [%rd2+56], %rd486; shl.b32 %r3583, %r471, 5; and.b32 %r3584, %r3583, 3840; shl.b32 %r3585, %r732, 1; and.b32 %r3586, %r3585, 240; or.b32 %r3587, %r3586, %r3584; bfe.u32 %r3588, %r741, 3, 4; or.b32 %r3589, %r3588, %r3587; shr.u32 %r3590, %r3587, 6; mul.wide.u32 %rd1840, %r3590, 8; add.s64 %rd1841, %rd486, %rd1840; and.b32 %r3591, %r3589, 63; shl.b64 %rd1843, %rd1833, %r3591; ld.u64 %rd1844, [%rd1841+544]; and.b64 %rd1845, %rd1844, %rd1843; setp.eq.s64 %p383, %rd1845, 0; mul.wide.u32 %rd1846, %r3589, 8; add.s64 %rd1847, %rd486, %rd1846; add.s64 %rd487, %rd1847, 1088; @%p383 bra $L__BB5_610; ld.u64 %rd1848, [%rd487]; add.s64 %rd1849, %rd486, %rd1848; and.b32 %r3592, %r732, -8; mov.u32 %r3593, 1; st.local.v2.u32 [%rd2], {%r472, %r3592}; and.b32 %r3594, %r741, -8; st.local.u32 [%rd2+8], %r3594; st.local.u64 [%rd2+48], %rd1849; shl.b32 %r3595, %r471, 6; and.b32 %r3596, %r3595, 448; shl.b32 %r3597, %r732, 3; and.b32 %r3598, %r3597, 56; or.b32 %r3599, %r3598, %r3596; and.b32 %r3600, %r741, 7; or.b32 %r3601, %r3599, %r3600; ld.u8 %rs371, [%rd1849+15]; shr.u16 %rs372, %rs371, 5; mov.u32 %r3602, 5; cvt.u32.u16 %r3603, %rs372; sub.s32 %r3604, %r3602, %r3603; shr.u32 %r3605, %r3601, %r3604; mul.wide.u32 %rd1850, %r3605, 4; add.s64 %rd1851, %rd1849, %rd1850; mov.u32 %r3606, 32; shr.u32 %r3607, %r3606, %r3603; add.s32 %r3608, %r3607, -1; and.b32 %r3609, %r3608, %r3601; shl.b32 %r3610, %r3609, %r3603; ld.u32 %r3611, [%rd1851+96]; shr.u32 %r3612, %r3611, %r3610; shl.b32 %r3613, %r3593, %r3603; mov.u32 %r3614, -1; shl.b32 %r3615, %r3614, %r3613; not.b32 %r3616, %r3615; and.b32 %r3617, %r3612, %r3616; cvt.rn.f32.u32 %f772, %r3617; ld.v2.f32 {%f773, %f774}, [%rd1849+80]; fma.rn.ftz.f32 %f946, %f774, %f772, %f773; mov.u32 %r5196, %r472; bra.uni $L__BB5_626; $L__BB5_638: and.b32 %r750, %r471, -4096; ld.local.u32 %r3738, [%rd2+24]; setp.ne.s32 %p400, %r750, %r3738; mov.u16 %rs506, 0; @%p400 bra $L__BB5_641; and.b32 %r3739, %r748, -4096; ld.local.u32 %r3740, [%rd2+28]; setp.ne.s32 %p401, %r3739, %r3740; @%p401 bra $L__BB5_641; and.b32 %r3741, %r741, -4096; ld.local.u32 %r3742, [%rd2+32]; setp.eq.s32 %p402, %r3741, %r3742; selp.u16 %rs506, 1, 0, %p402; $L__BB5_641: setp.eq.s16 %p403, %rs506, 0; @%p403 bra $L__BB5_647; shl.b32 %r3743, %r471, 3; and.b32 %r3744, %r3743, 31744; and.b32 %r3745, %r748, 3968; shr.u32 %r3746, %r3745, 2; or.b32 %r3747, %r3746, %r3744; bfe.u32 %r3748, %r741, 7, 5; or.b32 %r3749, %r3747, %r3748; shr.u32 %r3750, %r3747, 6; ld.local.u64 %rd502, [%rd2+64]; mul.wide.u32 %rd1899, %r3750, 8; add.s64 %rd1900, %rd502, %rd1899; and.b32 %r3751, %r3749, 63; mov.u64 %rd1901, 1; shl.b64 %rd1902, %rd1901, %r3751; ld.u64 %rd1903, [%rd1900+4128]; and.b64 %rd1904, %rd1903, %rd1902; setp.eq.s64 %p404, %rd1904, 0; mul.wide.u32 %rd1905, %r3749, 8; add.s64 %rd1906, %rd502, %rd1905; add.s64 %rd503, %rd1906, 8256; @%p404 bra $L__BB5_646; ld.u64 %rd1907, [%rd503]; add.s64 %rd504, %rd502, %rd1907; st.local.u32 [%rd2+12], %r749; and.b32 %r3752, %r748, -128; st.local.u32 [%rd2+16], %r3752; and.b32 %r3753, %r741, -128; st.local.u32 [%rd2+20], %r3753; st.local.u64 [%rd2+56], %rd504; shl.b32 %r3754, %r471, 5; and.b32 %r3755, %r3754, 3840; shl.b32 %r3756, %r748, 1; and.b32 %r3757, %r3756, 240; or.b32 %r3758, %r3757, %r3755; bfe.u32 %r3759, %r741, 3, 4; or.b32 %r3760, %r3758, %r3759; shr.u32 %r3761, %r3758, 6; mul.wide.u32 %rd1908, %r3761, 8; add.s64 %rd1909, %rd504, %rd1908; and.b32 %r3762, %r3760, 63; shl.b64 %rd1911, %rd1901, %r3762; ld.u64 %rd1912, [%rd1909+544]; and.b64 %rd1913, %rd1912, %rd1911; setp.eq.s64 %p405, %rd1913, 0; mul.wide.u32 %rd1914, %r3760, 8; add.s64 %rd1915, %rd504, %rd1914; add.s64 %rd505, %rd1915, 1088; @%p405 bra $L__BB5_645; ld.u64 %rd1916, [%rd505]; add.s64 %rd1917, %rd504, %rd1916; and.b32 %r3763, %r748, -8; mov.u32 %r3764, 1; st.local.v2.u32 [%rd2], {%r472, %r3763}; and.b32 %r3765, %r741, -8; st.local.u32 [%rd2+8], %r3765; st.local.u64 [%rd2+48], %rd1917; shl.b32 %r3766, %r471, 6; and.b32 %r3767, %r3766, 448; shl.b32 %r3768, %r748, 3; and.b32 %r3769, %r3768, 56; or.b32 %r3770, %r3769, %r3767; and.b32 %r3771, %r741, 7; or.b32 %r3772, %r3770, %r3771; ld.u8 %rs385, [%rd1917+15]; shr.u16 %rs386, %rs385, 5; mov.u32 %r3773, 5; cvt.u32.u16 %r3774, %rs386; sub.s32 %r3775, %r3773, %r3774; shr.u32 %r3776, %r3772, %r3775; mul.wide.u32 %rd1918, %r3776, 4; add.s64 %rd1919, %rd1917, %rd1918; mov.u32 %r3777, 32; shr.u32 %r3778, %r3777, %r3774; add.s32 %r3779, %r3778, -1; and.b32 %r3780, %r3779, %r3772; shl.b32 %r3781, %r3780, %r3774; ld.u32 %r3782, [%rd1919+96]; shr.u32 %r3783, %r3782, %r3781; shl.b32 %r3784, %r3764, %r3774; mov.u32 %r3785, -1; shl.b32 %r3786, %r3785, %r3784; not.b32 %r3787, %r3786; and.b32 %r3788, %r3783, %r3787; cvt.rn.f32.u32 %f793, %r3788; ld.v2.f32 {%f794, %f795}, [%rd1917+80]; fma.rn.ftz.f32 %f947, %f795, %f793, %f794; mov.u32 %r5196, %r472; bra.uni $L__BB5_661; $L__BB5_673: and.b32 %r756, %r471, -4096; ld.local.u32 %r3907, [%rd2+24]; setp.ne.s32 %p422, %r756, %r3907; mov.u16 %rs509, 0; @%p422 bra $L__BB5_676; and.b32 %r3908, %r748, -4096; ld.local.u32 %r3909, [%rd2+28]; setp.ne.s32 %p423, %r3908, %r3909; @%p423 bra $L__BB5_676; and.b32 %r3910, %r733, -4096; ld.local.u32 %r3911, [%rd2+32]; setp.eq.s32 %p424, %r3910, %r3911; selp.u16 %rs509, 1, 0, %p424; $L__BB5_676: setp.eq.s16 %p425, %rs509, 0; @%p425 bra $L__BB5_682; shl.b32 %r3912, %r471, 3; and.b32 %r3913, %r3912, 31744; and.b32 %r3914, %r748, 3968; shr.u32 %r3915, %r3914, 2; or.b32 %r3916, %r3915, %r3913; bfe.u32 %r3917, %r733, 7, 5; or.b32 %r3918, %r3916, %r3917; shr.u32 %r3919, %r3916, 6; ld.local.u64 %rd520, [%rd2+64]; mul.wide.u32 %rd1966, %r3919, 8; add.s64 %rd1967, %rd520, %rd1966; and.b32 %r3920, %r3918, 63; mov.u64 %rd1968, 1; shl.b64 %rd1969, %rd1968, %r3920; ld.u64 %rd1970, [%rd1967+4128]; and.b64 %rd1971, %rd1970, %rd1969; setp.eq.s64 %p426, %rd1971, 0; mul.wide.u32 %rd1972, %r3918, 8; add.s64 %rd1973, %rd520, %rd1972; add.s64 %rd521, %rd1973, 8256; @%p426 bra $L__BB5_681; ld.u64 %rd1974, [%rd521]; add.s64 %rd522, %rd520, %rd1974; st.local.u32 [%rd2+12], %r755; and.b32 %r3921, %r748, -128; st.local.u32 [%rd2+16], %r3921; and.b32 %r3922, %r733, -128; st.local.u32 [%rd2+20], %r3922; st.local.u64 [%rd2+56], %rd522; shl.b32 %r3923, %r471, 5; and.b32 %r3924, %r3923, 3840; shl.b32 %r3925, %r748, 1; and.b32 %r3926, %r3925, 240; or.b32 %r3927, %r3926, %r3924; bfe.u32 %r3928, %r733, 3, 4; or.b32 %r3929, %r3927, %r3928; shr.u32 %r3930, %r3927, 6; mul.wide.u32 %rd1975, %r3930, 8; add.s64 %rd1976, %rd522, %rd1975; and.b32 %r3931, %r3929, 63; shl.b64 %rd1978, %rd1968, %r3931; ld.u64 %rd1979, [%rd1976+544]; and.b64 %rd1980, %rd1979, %rd1978; setp.eq.s64 %p427, %rd1980, 0; mul.wide.u32 %rd1981, %r3929, 8; add.s64 %rd1982, %rd522, %rd1981; add.s64 %rd523, %rd1982, 1088; @%p427 bra $L__BB5_680; ld.u64 %rd1983, [%rd523]; add.s64 %rd1984, %rd522, %rd1983; and.b32 %r3932, %r748, -8; mov.u32 %r3933, 1; st.local.v2.u32 [%rd2], {%r472, %r3932}; and.b32 %r3934, %r733, -8; st.local.u32 [%rd2+8], %r3934; st.local.u64 [%rd2+48], %rd1984; and.b32 %r3935, %r471, 7; shl.b32 %r3936, %r748, 3; and.b32 %r3937, %r3936, 56; and.b32 %r3938, %r733, 7; bfi.b32 %r3939, %r3935, %r3938, 6, 3; or.b32 %r3940, %r3939, %r3937; ld.u8 %rs399, [%rd1984+15]; shr.u16 %rs400, %rs399, 5; mov.u32 %r3941, 5; cvt.u32.u16 %r3942, %rs400; sub.s32 %r3943, %r3941, %r3942; shr.u32 %r3944, %r3940, %r3943; mul.wide.u32 %rd1985, %r3944, 4; add.s64 %rd1986, %rd1984, %rd1985; mov.u32 %r3945, 32; shr.u32 %r3946, %r3945, %r3942; add.s32 %r3947, %r3946, -1; and.b32 %r3948, %r3947, %r3940; shl.b32 %r3949, %r3948, %r3942; ld.u32 %r3950, [%rd1986+96]; shr.u32 %r3951, %r3950, %r3949; shl.b32 %r3952, %r3933, %r3942; mov.u32 %r3953, -1; shl.b32 %r3954, %r3953, %r3952; not.b32 %r3955, %r3954; and.b32 %r3956, %r3951, %r3955; cvt.rn.f32.u32 %f813, %r3956; ld.v2.f32 {%f814, %f815}, [%rd1984+80]; fma.rn.ftz.f32 %f948, %f815, %f813, %f814; mov.u32 %r5196, %r472; bra.uni $L__BB5_696; $L__BB5_708: and.b32 %r762, %r635, -4096; ld.local.u32 %r4074, [%rd2+24]; setp.ne.s32 %p444, %r762, %r4074; mov.u16 %rs512, 0; @%p444 bra $L__BB5_711; and.b32 %r4075, %r748, -4096; ld.local.u32 %r4076, [%rd2+28]; setp.ne.s32 %p445, %r4075, %r4076; @%p445 bra $L__BB5_711; and.b32 %r4077, %r733, -4096; ld.local.u32 %r4078, [%rd2+32]; setp.eq.s32 %p446, %r4077, %r4078; selp.u16 %rs512, 1, 0, %p446; $L__BB5_711: setp.eq.s16 %p447, %rs512, 0; @%p447 bra $L__BB5_717; shl.b32 %r4079, %r635, 3; and.b32 %r4080, %r4079, 31744; and.b32 %r4081, %r748, 3968; shr.u32 %r4082, %r4081, 2; or.b32 %r4083, %r4082, %r4080; bfe.u32 %r4084, %r733, 7, 5; or.b32 %r4085, %r4083, %r4084; shr.u32 %r4086, %r4083, 6; ld.local.u64 %rd538, [%rd2+64]; mul.wide.u32 %rd2033, %r4086, 8; add.s64 %rd2034, %rd538, %rd2033; and.b32 %r4087, %r4085, 63; mov.u64 %rd2035, 1; shl.b64 %rd2036, %rd2035, %r4087; ld.u64 %rd2037, [%rd2034+4128]; and.b64 %rd2038, %rd2037, %rd2036; setp.eq.s64 %p448, %rd2038, 0; mul.wide.u32 %rd2039, %r4085, 8; add.s64 %rd2040, %rd538, %rd2039; add.s64 %rd539, %rd2040, 8256; @%p448 bra $L__BB5_716; ld.u64 %rd2041, [%rd539]; add.s64 %rd540, %rd538, %rd2041; st.local.u32 [%rd2+12], %r761; and.b32 %r4088, %r748, -128; st.local.u32 [%rd2+16], %r4088; and.b32 %r4089, %r733, -128; st.local.u32 [%rd2+20], %r4089; st.local.u64 [%rd2+56], %rd540; shl.b32 %r4090, %r635, 5; and.b32 %r4091, %r4090, 3840; shl.b32 %r4092, %r748, 1; and.b32 %r4093, %r4092, 240; or.b32 %r4094, %r4093, %r4091; bfe.u32 %r4095, %r733, 3, 4; or.b32 %r4096, %r4094, %r4095; shr.u32 %r4097, %r4094, 6; mul.wide.u32 %rd2042, %r4097, 8; add.s64 %rd2043, %rd540, %rd2042; and.b32 %r4098, %r4096, 63; shl.b64 %rd2045, %rd2035, %r4098; ld.u64 %rd2046, [%rd2043+544]; and.b64 %rd2047, %rd2046, %rd2045; setp.eq.s64 %p449, %rd2047, 0; mul.wide.u32 %rd2048, %r4096, 8; add.s64 %rd2049, %rd540, %rd2048; add.s64 %rd541, %rd2049, 1088; @%p449 bra $L__BB5_715; ld.u64 %rd2050, [%rd541]; add.s64 %rd2051, %rd540, %rd2050; and.b32 %r4099, %r748, -8; mov.u32 %r4100, 1; st.local.v2.u32 [%rd2], {%r636, %r4099}; and.b32 %r4101, %r733, -8; st.local.u32 [%rd2+8], %r4101; st.local.u64 [%rd2+48], %rd2051; and.b32 %r4102, %r635, 7; shl.b32 %r4103, %r748, 3; and.b32 %r4104, %r4103, 56; and.b32 %r4105, %r733, 7; bfi.b32 %r4106, %r4102, %r4105, 6, 3; or.b32 %r4107, %r4106, %r4104; ld.u8 %rs413, [%rd2051+15]; shr.u16 %rs414, %rs413, 5; mov.u32 %r4108, 5; cvt.u32.u16 %r4109, %rs414; sub.s32 %r4110, %r4108, %r4109; shr.u32 %r4111, %r4107, %r4110; mul.wide.u32 %rd2052, %r4111, 4; add.s64 %rd2053, %rd2051, %rd2052; mov.u32 %r4112, 32; shr.u32 %r4113, %r4112, %r4109; add.s32 %r4114, %r4113, -1; and.b32 %r4115, %r4114, %r4107; shl.b32 %r4116, %r4115, %r4109; ld.u32 %r4117, [%rd2053+96]; shr.u32 %r4118, %r4117, %r4116; shl.b32 %r4119, %r4100, %r4109; mov.u32 %r4120, -1; shl.b32 %r4121, %r4120, %r4119; not.b32 %r4122, %r4121; and.b32 %r4123, %r4118, %r4122; cvt.rn.f32.u32 %f836, %r4123; ld.v2.f32 {%f837, %f838}, [%rd2051+80]; fma.rn.ftz.f32 %f949, %f838, %f836, %f837; mov.u32 %r5196, %r636; bra.uni $L__BB5_731; $L__BB5_743: and.b32 %r768, %r635, -4096; ld.local.u32 %r4243, [%rd2+24]; setp.ne.s32 %p466, %r768, %r4243; mov.u16 %rs515, 0; @%p466 bra $L__BB5_746; and.b32 %r4244, %r748, -4096; ld.local.u32 %r4245, [%rd2+28]; setp.ne.s32 %p467, %r4244, %r4245; @%p467 bra $L__BB5_746; and.b32 %r4246, %r741, -4096; ld.local.u32 %r4247, [%rd2+32]; setp.eq.s32 %p468, %r4246, %r4247; selp.u16 %rs515, 1, 0, %p468; $L__BB5_746: setp.eq.s16 %p469, %rs515, 0; @%p469 bra $L__BB5_752; shl.b32 %r4248, %r635, 3; and.b32 %r4249, %r4248, 31744; and.b32 %r4250, %r748, 3968; shr.u32 %r4251, %r4250, 2; or.b32 %r4252, %r4251, %r4249; bfe.u32 %r4253, %r741, 7, 5; or.b32 %r4254, %r4252, %r4253; shr.u32 %r4255, %r4252, 6; ld.local.u64 %rd556, [%rd2+64]; mul.wide.u32 %rd2100, %r4255, 8; add.s64 %rd2101, %rd556, %rd2100; and.b32 %r4256, %r4254, 63; mov.u64 %rd2102, 1; shl.b64 %rd2103, %rd2102, %r4256; ld.u64 %rd2104, [%rd2101+4128]; and.b64 %rd2105, %rd2104, %rd2103; setp.eq.s64 %p470, %rd2105, 0; mul.wide.u32 %rd2106, %r4254, 8; add.s64 %rd2107, %rd556, %rd2106; add.s64 %rd557, %rd2107, 8256; @%p470 bra $L__BB5_751; ld.u64 %rd2108, [%rd557]; add.s64 %rd558, %rd556, %rd2108; st.local.u32 [%rd2+12], %r767; and.b32 %r4257, %r748, -128; st.local.u32 [%rd2+16], %r4257; and.b32 %r4258, %r741, -128; st.local.u32 [%rd2+20], %r4258; st.local.u64 [%rd2+56], %rd558; shl.b32 %r4259, %r635, 5; and.b32 %r4260, %r4259, 3840; shl.b32 %r4261, %r748, 1; and.b32 %r4262, %r4261, 240; or.b32 %r4263, %r4262, %r4260; bfe.u32 %r4264, %r741, 3, 4; or.b32 %r4265, %r4263, %r4264; shr.u32 %r4266, %r4263, 6; mul.wide.u32 %rd2109, %r4266, 8; add.s64 %rd2110, %rd558, %rd2109; and.b32 %r4267, %r4265, 63; shl.b64 %rd2112, %rd2102, %r4267; ld.u64 %rd2113, [%rd2110+544]; and.b64 %rd2114, %rd2113, %rd2112; setp.eq.s64 %p471, %rd2114, 0; mul.wide.u32 %rd2115, %r4265, 8; add.s64 %rd2116, %rd558, %rd2115; add.s64 %rd559, %rd2116, 1088; @%p471 bra $L__BB5_750; ld.u64 %rd2117, [%rd559]; add.s64 %rd2118, %rd558, %rd2117; and.b32 %r4268, %r748, -8; mov.u32 %r4269, 1; st.local.v2.u32 [%rd2], {%r636, %r4268}; and.b32 %r4270, %r741, -8; st.local.u32 [%rd2+8], %r4270; st.local.u64 [%rd2+48], %rd2118; shl.b32 %r4271, %r635, 6; and.b32 %r4272, %r4271, 448; shl.b32 %r4273, %r748, 3; and.b32 %r4274, %r4273, 56; or.b32 %r4275, %r4274, %r4272; and.b32 %r4276, %r741, 7; or.b32 %r4277, %r4275, %r4276; ld.u8 %rs427, [%rd2118+15]; shr.u16 %rs428, %rs427, 5; mov.u32 %r4278, 5; cvt.u32.u16 %r4279, %rs428; sub.s32 %r4280, %r4278, %r4279; shr.u32 %r4281, %r4277, %r4280; mul.wide.u32 %rd2119, %r4281, 4; add.s64 %rd2120, %rd2118, %rd2119; mov.u32 %r4282, 32; shr.u32 %r4283, %r4282, %r4279; add.s32 %r4284, %r4283, -1; and.b32 %r4285, %r4284, %r4277; shl.b32 %r4286, %r4285, %r4279; ld.u32 %r4287, [%rd2120+96]; shr.u32 %r4288, %r4287, %r4286; shl.b32 %r4289, %r4269, %r4279; mov.u32 %r4290, -1; shl.b32 %r4291, %r4290, %r4289; not.b32 %r4292, %r4291; and.b32 %r4293, %r4288, %r4292; cvt.rn.f32.u32 %f856, %r4293; ld.v2.f32 {%f857, %f858}, [%rd2118+80]; fma.rn.ftz.f32 %f950, %f858, %f856, %f857; mov.u32 %r5196, %r636; bra.uni $L__BB5_766; $L__BB5_778: and.b32 %r774, %r635, -4096; ld.local.u32 %r4414, [%rd2+24]; setp.ne.s32 %p488, %r774, %r4414; mov.u16 %rs518, 0; @%p488 bra $L__BB5_781; and.b32 %r4415, %r732, -4096; ld.local.u32 %r4416, [%rd2+28]; setp.ne.s32 %p489, %r4415, %r4416; @%p489 bra $L__BB5_781; and.b32 %r4417, %r741, -4096; ld.local.u32 %r4418, [%rd2+32]; setp.eq.s32 %p490, %r4417, %r4418; selp.u16 %rs518, 1, 0, %p490; $L__BB5_781: setp.eq.s16 %p491, %rs518, 0; @%p491 bra $L__BB5_787; shl.b32 %r4419, %r635, 3; and.b32 %r4420, %r4419, 31744; and.b32 %r4421, %r732, 3968; shr.u32 %r4422, %r4421, 2; or.b32 %r4423, %r4420, %r4422; bfe.u32 %r4424, %r741, 7, 5; or.b32 %r4425, %r4424, %r4423; shr.u32 %r4426, %r4423, 6; ld.local.u64 %rd574, [%rd2+64]; mul.wide.u32 %rd2167, %r4426, 8; add.s64 %rd2168, %rd574, %rd2167; and.b32 %r4427, %r4425, 63; mov.u64 %rd2169, 1; shl.b64 %rd2170, %rd2169, %r4427; ld.u64 %rd2171, [%rd2168+4128]; and.b64 %rd2172, %rd2171, %rd2170; setp.eq.s64 %p492, %rd2172, 0; mul.wide.u32 %rd2173, %r4425, 8; add.s64 %rd2174, %rd574, %rd2173; add.s64 %rd575, %rd2174, 8256; @%p492 bra $L__BB5_786; ld.u64 %rd2175, [%rd575]; add.s64 %rd576, %rd574, %rd2175; st.local.u32 [%rd2+12], %r773; and.b32 %r4428, %r732, -128; st.local.u32 [%rd2+16], %r4428; and.b32 %r4429, %r741, -128; st.local.u32 [%rd2+20], %r4429; st.local.u64 [%rd2+56], %rd576; shl.b32 %r4430, %r635, 5; and.b32 %r4431, %r4430, 3840; shl.b32 %r4432, %r732, 1; and.b32 %r4433, %r4432, 240; or.b32 %r4434, %r4431, %r4433; bfe.u32 %r4435, %r741, 3, 4; or.b32 %r4436, %r4435, %r4434; shr.u32 %r4437, %r4434, 6; mul.wide.u32 %rd2176, %r4437, 8; add.s64 %rd2177, %rd576, %rd2176; and.b32 %r4438, %r4436, 63; shl.b64 %rd2179, %rd2169, %r4438; ld.u64 %rd2180, [%rd2177+544]; and.b64 %rd2181, %rd2180, %rd2179; setp.eq.s64 %p493, %rd2181, 0; mul.wide.u32 %rd2182, %r4436, 8; add.s64 %rd2183, %rd576, %rd2182; add.s64 %rd577, %rd2183, 1088; @%p493 bra $L__BB5_785; ld.u64 %rd2184, [%rd577]; add.s64 %rd2185, %rd576, %rd2184; and.b32 %r4439, %r732, -8; mov.u32 %r4440, 1; st.local.v2.u32 [%rd2], {%r636, %r4439}; and.b32 %r4441, %r741, -8; st.local.u32 [%rd2+8], %r4441; st.local.u64 [%rd2+48], %rd2185; shl.b32 %r4442, %r635, 6; and.b32 %r4443, %r4442, 448; shl.b32 %r4444, %r732, 3; and.b32 %r4445, %r4444, 56; or.b32 %r4446, %r4443, %r4445; and.b32 %r4447, %r741, 7; or.b32 %r4448, %r4446, %r4447; ld.u8 %rs441, [%rd2185+15]; shr.u16 %rs442, %rs441, 5; mov.u32 %r4449, 5; cvt.u32.u16 %r4450, %rs442; sub.s32 %r4451, %r4449, %r4450; shr.u32 %r4452, %r4448, %r4451; mul.wide.u32 %rd2186, %r4452, 4; add.s64 %rd2187, %rd2185, %rd2186; mov.u32 %r4453, 32; shr.u32 %r4454, %r4453, %r4450; add.s32 %r4455, %r4454, -1; and.b32 %r4456, %r4455, %r4448; shl.b32 %r4457, %r4456, %r4450; ld.u32 %r4458, [%rd2187+96]; shr.u32 %r4459, %r4458, %r4457; shl.b32 %r4460, %r4440, %r4450; mov.u32 %r4461, -1; shl.b32 %r4462, %r4461, %r4460; not.b32 %r4463, %r4462; and.b32 %r4464, %r4459, %r4463; cvt.rn.f32.u32 %f877, %r4464; ld.v2.f32 {%f878, %f879}, [%rd2185+80]; fma.rn.ftz.f32 %f951, %f879, %f877, %f878; mov.u32 %r5196, %r636; bra.uni $L__BB5_801; $L__BB5_821: ld.f32 %f952, [%rd599]; bra.uni $L__BB5_833; $L__BB5_42: ld.f32 %f930, [%rd43]; bra.uni $L__BB5_66; $L__BB5_77: ld.f32 %f931, [%rd72]; bra.uni $L__BB5_101; $L__BB5_112: ld.f32 %f932, [%rd101]; bra.uni $L__BB5_136; $L__BB5_147: ld.f32 %f933, [%rd130]; bra.uni $L__BB5_171; $L__BB5_182: ld.f32 %f934, [%rd159]; bra.uni $L__BB5_206; $L__BB5_217: ld.f32 %f935, [%rd188]; bra.uni $L__BB5_241; $L__BB5_252: ld.f32 %f936, [%rd217]; bra.uni $L__BB5_276; $L__BB5_287: ld.f32 %f937, [%rd246]; bra.uni $L__BB5_311; $L__BB5_322: ld.f32 %f938, [%rd275]; bra.uni $L__BB5_346; $L__BB5_357: ld.f32 %f939, [%rd304]; bra.uni $L__BB5_381; $L__BB5_392: ld.f32 %f940, [%rd333]; bra.uni $L__BB5_416; $L__BB5_427: ld.f32 %f941, [%rd362]; bra.uni $L__BB5_451; $L__BB5_462: ld.f32 %f942, [%rd391]; bra.uni $L__BB5_486; $L__BB5_497: ld.f32 %f943, [%rd420]; bra.uni $L__BB5_521; $L__BB5_532: ld.f32 %f944, [%rd449]; bra.uni $L__BB5_556; $L__BB5_567: ld.f32 %f945, [%rd465]; bra.uni $L__BB5_591; $L__BB5_602: ld.f32 %f946, [%rd483]; bra.uni $L__BB5_626; $L__BB5_637: ld.f32 %f947, [%rd501]; bra.uni $L__BB5_661; $L__BB5_672: ld.f32 %f948, [%rd519]; bra.uni $L__BB5_696; $L__BB5_707: ld.f32 %f949, [%rd537]; bra.uni $L__BB5_731; $L__BB5_742: ld.f32 %f950, [%rd555]; bra.uni $L__BB5_766; $L__BB5_777: ld.f32 %f951, [%rd573]; bra.uni $L__BB5_801; $L__BB5_820: ld.f32 %f952, [%rd601]; bra.uni $L__BB5_833; $L__BB5_52: ld.u32 %r81, [%rd5]; mov.u64 %rd2302, 0; setp.eq.s32 %p32, %r81, 0; @%p32 bra $L__BB5_57; shr.u32 %r4690, %r7, 12; cvt.u64.u32 %rd2293, %r4690; shl.b64 %rd2292, %rd2293, 42; or.b64 %rd736, %rd19, %rd2292; shr.u32 %r1136, %r74, 12; cvt.u64.u32 %rd737, %r1136; or.b64 %rd51, %rd736, %rd737; mov.u32 %r4779, 0; $L__BB5_54: mul.wide.u32 %rd738, %r4779, 32; add.s64 %rd739, %rd17, %rd738; ld.u64 %rd740, [%rd739]; setp.eq.s64 %p33, %rd740, %rd51; @%p33 bra $L__BB5_56; add.s32 %r4779, %r4779, 1; setp.lt.u32 %p34, %r4779, %r81; @%p34 bra $L__BB5_54; bra.uni $L__BB5_57; $L__BB5_87: ld.u32 %r146, [%rd5]; mov.u64 %rd2308, 0; setp.eq.s32 %p54, %r146, 0; @%p54 bra $L__BB5_92; shr.u32 %r4689, %r7, 12; cvt.u64.u32 %rd2290, %r4689; shl.b64 %rd2289, %rd2290, 42; shr.u32 %r1289, %r74, 12; cvt.u64.u32 %rd805, %r1289; or.b64 %rd806, %rd2289, %rd805; shr.u32 %r1290, %r139, 12; mul.wide.u32 %rd807, %r1290, 2097152; or.b64 %rd80, %rd806, %rd807; mov.u32 %r4828, 0; $L__BB5_89: mul.wide.u32 %rd808, %r4828, 32; add.s64 %rd809, %rd17, %rd808; ld.u64 %rd810, [%rd809]; setp.eq.s64 %p55, %rd810, %rd80; @%p55 bra $L__BB5_91; add.s32 %r4828, %r4828, 1; setp.lt.u32 %p56, %r4828, %r146; @%p56 bra $L__BB5_89; bra.uni $L__BB5_92; $L__BB5_122: ld.u32 %r207, [%rd5]; mov.u64 %rd2314, 0; setp.eq.s32 %p76, %r207, 0; @%p76 bra $L__BB5_127; shr.u32 %r1440, %r139, 12; mul.wide.u32 %rd875, %r1440, 2097152; or.b64 %rd109, %rd21, %rd875; mov.u32 %r4874, 0; $L__BB5_124: mul.wide.u32 %rd876, %r4874, 32; add.s64 %rd877, %rd17, %rd876; ld.u64 %rd878, [%rd877]; setp.eq.s64 %p77, %rd878, %rd109; @%p77 bra $L__BB5_126; add.s32 %r4874, %r4874, 1; setp.lt.u32 %p78, %r4874, %r207; @%p78 bra $L__BB5_124; bra.uni $L__BB5_127; $L__BB5_157: ld.u32 %r269, [%rd5]; mov.u64 %rd2320, 0; setp.eq.s32 %p98, %r269, 0; @%p98 bra $L__BB5_162; shr.u32 %r1589, %r259, 12; cvt.u64.u32 %rd943, %r1589; shl.b64 %rd944, %rd943, 42; or.b64 %rd945, %rd944, %rd18; shr.u32 %r1590, %r139, 12; mul.wide.u32 %rd946, %r1590, 2097152; or.b64 %rd138, %rd945, %rd946; mov.u32 %r4917, 0; $L__BB5_159: mul.wide.u32 %rd947, %r4917, 32; add.s64 %rd948, %rd17, %rd947; ld.u64 %rd949, [%rd948]; setp.eq.s64 %p99, %rd949, %rd138; @%p99 bra $L__BB5_161; add.s32 %r4917, %r4917, 1; setp.lt.u32 %p100, %r4917, %r269; @%p100 bra $L__BB5_159; bra.uni $L__BB5_162; $L__BB5_192: ld.u32 %r326, [%rd5]; mov.u64 %rd2326, 0; setp.eq.s32 %p120, %r326, 0; @%p120 bra $L__BB5_197; shr.u32 %r1742, %r259, 12; cvt.u64.u32 %rd1014, %r1742; shr.u32 %r1743, %r74, 12; cvt.u64.u32 %rd1015, %r1743; bfi.b64 %rd1016, %rd1014, %rd1015, 42, 22; shr.u32 %r1744, %r139, 12; mul.wide.u32 %rd1017, %r1744, 2097152; or.b64 %rd167, %rd1016, %rd1017; mov.u32 %r4957, 0; $L__BB5_194: cvt.u64.u32 %rd168, %r4957; mul.wide.u32 %rd1018, %r4957, 32; add.s64 %rd1019, %rd17, %rd1018; ld.u64 %rd1020, [%rd1019]; setp.eq.s64 %p121, %rd1020, %rd167; add.s32 %r4957, %r4957, 1; @%p121 bra $L__BB5_196; setp.lt.u32 %p122, %r4957, %r326; @%p122 bra $L__BB5_194; bra.uni $L__BB5_197; $L__BB5_227: ld.u32 %r380, [%rd5]; mov.u64 %rd2332, 0; setp.eq.s32 %p142, %r380, 0; @%p142 bra $L__BB5_232; shr.u32 %r1897, %r259, 12; cvt.u64.u32 %rd1085, %r1897; shl.b64 %rd1086, %rd1085, 42; or.b64 %rd1087, %rd1086, %rd19; shr.u32 %r1898, %r74, 12; cvt.u64.u32 %rd1088, %r1898; or.b64 %rd196, %rd1087, %rd1088; mov.u32 %r4994, 0; $L__BB5_229: cvt.u64.u32 %rd197, %r4994; mul.wide.u32 %rd1089, %r4994, 32; add.s64 %rd1090, %rd17, %rd1089; ld.u64 %rd1091, [%rd1090]; setp.eq.s64 %p143, %rd1091, %rd196; add.s32 %r4994, %r4994, 1; @%p143 bra $L__BB5_231; setp.lt.u32 %p144, %r4994, %r380; @%p144 bra $L__BB5_229; bra.uni $L__BB5_232; $L__BB5_262: ld.u32 %r431, [%rd5]; mov.u64 %rd2338, 0; setp.eq.s32 %p164, %r431, 0; @%p164 bra $L__BB5_267; shr.u32 %r2051, %r259, 12; cvt.u64.u32 %rd1156, %r2051; shl.b64 %rd1157, %rd1156, 42; or.b64 %rd1158, %rd19, %rd18; or.b64 %rd225, %rd1158, %rd1157; mov.u32 %r5028, 0; $L__BB5_264: cvt.u64.u32 %rd226, %r5028; mul.wide.u32 %rd1159, %r5028, 32; add.s64 %rd1160, %rd17, %rd1159; ld.u64 %rd1161, [%rd1160]; setp.eq.s64 %p165, %rd1161, %rd225; add.s32 %r5028, %r5028, 1; @%p165 bra $L__BB5_266; setp.lt.u32 %p166, %r5028, %r431; @%p166 bra $L__BB5_264; bra.uni $L__BB5_267; $L__BB5_297: ld.u32 %r481, [%rd5]; mov.u64 %rd2344, 0; setp.eq.s32 %p186, %r481, 0; @%p186 bra $L__BB5_302; shr.u32 %r2204, %r471, 12; cvt.u64.u32 %rd1226, %r2204; shl.b64 %rd1227, %rd1226, 42; or.b64 %rd1228, %rd1227, %rd18; or.b64 %rd254, %rd1228, %rd19; mov.u32 %r5059, 0; $L__BB5_299: cvt.u64.u32 %rd255, %r5059; mul.wide.u32 %rd1229, %r5059, 32; add.s64 %rd1230, %rd17, %rd1229; ld.u64 %rd1231, [%rd1230]; setp.eq.s64 %p187, %rd1231, %rd254; add.s32 %r5059, %r5059, 1; @%p187 bra $L__BB5_301; setp.lt.u32 %p188, %r5059, %r481; @%p188 bra $L__BB5_299; bra.uni $L__BB5_302; $L__BB5_332: ld.u32 %r526, [%rd5]; mov.u64 %rd2350, 0; setp.eq.s32 %p208, %r526, 0; @%p208 bra $L__BB5_337; shr.u32 %r2357, %r471, 12; cvt.u64.u32 %rd1296, %r2357; shl.b64 %rd1297, %rd1296, 42; or.b64 %rd1298, %rd19, %rd1297; shr.u32 %r2358, %r74, 12; cvt.u64.u32 %rd1299, %r2358; or.b64 %rd283, %rd1298, %rd1299; mov.u32 %r5087, 0; $L__BB5_334: cvt.u64.u32 %rd284, %r5087; mul.wide.u32 %rd1300, %r5087, 32; add.s64 %rd1301, %rd17, %rd1300; ld.u64 %rd1302, [%rd1301]; setp.eq.s64 %p209, %rd1302, %rd283; add.s32 %r5087, %r5087, 1; @%p209 bra $L__BB5_336; setp.lt.u32 %p210, %r5087, %r526; @%p210 bra $L__BB5_334; bra.uni $L__BB5_337; $L__BB5_367: ld.u32 %r568, [%rd5]; mov.u64 %rd2356, 0; setp.eq.s32 %p230, %r568, 0; @%p230 bra $L__BB5_372; shr.u32 %r2511, %r471, 12; cvt.u64.u32 %rd1367, %r2511; shr.u32 %r2512, %r74, 12; cvt.u64.u32 %rd1368, %r2512; bfi.b64 %rd1369, %rd1367, %rd1368, 42, 22; shr.u32 %r2513, %r139, 12; mul.wide.u32 %rd1370, %r2513, 2097152; or.b64 %rd312, %rd1369, %rd1370; mov.u32 %r5112, 0; $L__BB5_369: cvt.u64.u32 %rd313, %r5112; mul.wide.u32 %rd1371, %r5112, 32; add.s64 %rd1372, %rd17, %rd1371; ld.u64 %rd1373, [%rd1372]; setp.eq.s64 %p231, %rd1373, %rd312; add.s32 %r5112, %r5112, 1; @%p231 bra $L__BB5_371; setp.lt.u32 %p232, %r5112, %r568; @%p232 bra $L__BB5_369; bra.uni $L__BB5_372; $L__BB5_402: ld.u32 %r607, [%rd5]; mov.u64 %rd2362, 0; setp.eq.s32 %p252, %r607, 0; @%p252 bra $L__BB5_407; shr.u32 %r2663, %r471, 12; cvt.u64.u32 %rd1438, %r2663; shl.b64 %rd1439, %rd1438, 42; or.b64 %rd1440, %rd1439, %rd18; shr.u32 %r2664, %r139, 12; mul.wide.u32 %rd1441, %r2664, 2097152; or.b64 %rd341, %rd1440, %rd1441; mov.u32 %r5134, 0; $L__BB5_404: cvt.u64.u32 %rd342, %r5134; mul.wide.u32 %rd1442, %r5134, 32; add.s64 %rd1443, %rd17, %rd1442; ld.u64 %rd1444, [%rd1443]; setp.eq.s64 %p253, %rd1444, %rd341; add.s32 %r5134, %r5134, 1; @%p253 bra $L__BB5_406; setp.lt.u32 %p254, %r5134, %r607; @%p254 bra $L__BB5_404; bra.uni $L__BB5_407; $L__BB5_437: ld.u32 %r645, [%rd5]; mov.u64 %rd2368, 0; setp.eq.s32 %p274, %r645, 0; @%p274 bra $L__BB5_442; shr.u32 %r2813, %r635, 12; cvt.u64.u32 %rd1509, %r2813; shl.b64 %rd1510, %rd1509, 42; or.b64 %rd1511, %rd1510, %rd18; shr.u32 %r2814, %r139, 12; mul.wide.u32 %rd1512, %r2814, 2097152; or.b64 %rd370, %rd1511, %rd1512; mov.u32 %r5153, 0; $L__BB5_439: cvt.u64.u32 %rd371, %r5153; mul.wide.u32 %rd1513, %r5153, 32; add.s64 %rd1514, %rd17, %rd1513; ld.u64 %rd1515, [%rd1514]; setp.eq.s64 %p275, %rd1515, %rd370; add.s32 %r5153, %r5153, 1; @%p275 bra $L__BB5_441; setp.lt.u32 %p276, %r5153, %r645; @%p276 bra $L__BB5_439; bra.uni $L__BB5_442; $L__BB5_472: ld.u32 %r678, [%rd5]; mov.u64 %rd2374, 0; setp.eq.s32 %p296, %r678, 0; @%p296 bra $L__BB5_477; shr.u32 %r2966, %r635, 12; cvt.u64.u32 %rd1580, %r2966; shr.u32 %r2967, %r74, 12; cvt.u64.u32 %rd1581, %r2967; bfi.b64 %rd1582, %rd1580, %rd1581, 42, 22; shr.u32 %r2968, %r139, 12; mul.wide.u32 %rd1583, %r2968, 2097152; or.b64 %rd399, %rd1582, %rd1583; mov.u32 %r5169, 0; $L__BB5_474: cvt.u64.u32 %rd400, %r5169; mul.wide.u32 %rd1584, %r5169, 32; add.s64 %rd1585, %rd17, %rd1584; ld.u64 %rd1586, [%rd1585]; setp.eq.s64 %p297, %rd1586, %rd399; add.s32 %r5169, %r5169, 1; @%p297 bra $L__BB5_476; setp.lt.u32 %p298, %r5169, %r678; @%p298 bra $L__BB5_474; bra.uni $L__BB5_477; $L__BB5_507: ld.u32 %r708, [%rd5]; mov.u64 %rd2380, 0; setp.eq.s32 %p318, %r708, 0; @%p318 bra $L__BB5_512; shr.u32 %r3121, %r635, 12; cvt.u64.u32 %rd1651, %r3121; shl.b64 %rd1652, %rd1651, 42; or.b64 %rd1653, %rd1652, %rd19; shr.u32 %r3122, %r74, 12; cvt.u64.u32 %rd1654, %r3122; or.b64 %rd428, %rd1653, %rd1654; mov.u32 %r5182, 0; $L__BB5_509: cvt.u64.u32 %rd429, %r5182; mul.wide.u32 %rd1655, %r5182, 32; add.s64 %rd1656, %rd17, %rd1655; ld.u64 %rd1657, [%rd1656]; setp.eq.s64 %p319, %rd1657, %rd428; add.s32 %r5182, %r5182, 1; @%p319 bra $L__BB5_511; setp.lt.u32 %p320, %r5182, %r708; @%p320 bra $L__BB5_509; bra.uni $L__BB5_512; $L__BB5_542: or.b64 %rd453, %rd19, %rd18; ld.u32 %r729, [%rd5]; mov.u64 %rd2386, 0; setp.eq.s32 %p340, %r729, 0; @%p340 bra $L__BB5_547; shr.u32 %r3279, %r635, 12; cvt.u64.u32 %rd1720, %r3279; shl.b64 %rd1721, %rd1720, 42; or.b64 %rd454, %rd453, %rd1721; mov.u32 %r5192, 0; $L__BB5_544: cvt.u64.u32 %rd455, %r5192; mul.wide.u32 %rd1722, %r5192, 32; add.s64 %rd1723, %rd17, %rd1722; ld.u64 %rd1724, [%rd1723]; setp.eq.s64 %p341, %rd1724, %rd454; add.s32 %r5192, %r5192, 1; @%p341 bra $L__BB5_546; setp.lt.u32 %p342, %r5192, %r729; @%p342 bra $L__BB5_544; bra.uni $L__BB5_547; $L__BB5_577: ld.local.u64 %rd470, [%rd2+40]; mov.u64 %rd2387, 0; ld.u32 %r737, [%rd470+24]; setp.eq.s32 %p362, %r737, 0; @%p362 bra $L__BB5_582; add.s64 %rd471, %rd470, 64; shr.u32 %r3448, %r471, 12; cvt.u64.u32 %rd1786, %r3448; shr.u32 %r3449, %r733, 12; cvt.u64.u32 %rd1787, %r3449; bfi.b64 %rd1788, %rd1786, %rd1787, 42, 22; shr.u32 %r3450, %r732, 12; mul.wide.u32 %rd1789, %r3450, 2097152; or.b64 %rd472, %rd1788, %rd1789; mov.u32 %r5193, 0; $L__BB5_579: cvt.u64.u32 %rd473, %r5193; mul.wide.u32 %rd1790, %r5193, 32; add.s64 %rd1791, %rd471, %rd1790; ld.u64 %rd1792, [%rd1791]; setp.eq.s64 %p363, %rd1792, %rd472; add.s32 %r5193, %r5193, 1; @%p363 bra $L__BB5_581; setp.lt.u32 %p364, %r5193, %r737; @%p364 bra $L__BB5_579; bra.uni $L__BB5_582; $L__BB5_612: ld.local.u64 %rd488, [%rd2+40]; mov.u64 %rd2388, 0; ld.u32 %r744, [%rd488+24]; setp.eq.s32 %p384, %r744, 0; @%p384 bra $L__BB5_617; add.s64 %rd489, %rd488, 64; shr.u32 %r3619, %r471, 12; cvt.u64.u32 %rd1853, %r3619; shl.b64 %rd1854, %rd1853, 42; shr.u32 %r3620, %r732, 12; mul.wide.u32 %rd1855, %r3620, 2097152; or.b64 %rd1856, %rd1855, %rd1854; shr.u32 %r3621, %r741, 12; cvt.u64.u32 %rd1857, %r3621; or.b64 %rd490, %rd1856, %rd1857; mov.u32 %r5195, 0; $L__BB5_614: cvt.u64.u32 %rd491, %r5195; mul.wide.u32 %rd1858, %r5195, 32; add.s64 %rd1859, %rd489, %rd1858; ld.u64 %rd1860, [%rd1859]; setp.eq.s64 %p385, %rd1860, %rd490; add.s32 %r5195, %r5195, 1; @%p385 bra $L__BB5_616; setp.lt.u32 %p386, %r5195, %r744; @%p386 bra $L__BB5_614; bra.uni $L__BB5_617; $L__BB5_647: ld.local.u64 %rd506, [%rd2+40]; mov.u64 %rd2389, 0; ld.u32 %r751, [%rd506+24]; setp.eq.s32 %p406, %r751, 0; @%p406 bra $L__BB5_652; add.s64 %rd507, %rd506, 64; shr.u32 %r3790, %r471, 12; cvt.u64.u32 %rd1921, %r3790; shr.u32 %r3791, %r741, 12; cvt.u64.u32 %rd1922, %r3791; bfi.b64 %rd1923, %rd1921, %rd1922, 42, 22; shr.u32 %r3792, %r748, 12; mul.wide.u32 %rd1924, %r3792, 2097152; or.b64 %rd508, %rd1923, %rd1924; mov.u32 %r5197, 0; $L__BB5_649: cvt.u64.u32 %rd509, %r5197; mul.wide.u32 %rd1925, %r5197, 32; add.s64 %rd1926, %rd507, %rd1925; ld.u64 %rd1927, [%rd1926]; setp.eq.s64 %p407, %rd1927, %rd508; add.s32 %r5197, %r5197, 1; @%p407 bra $L__BB5_651; setp.lt.u32 %p408, %r5197, %r751; @%p408 bra $L__BB5_649; bra.uni $L__BB5_652; $L__BB5_682: ld.local.u64 %rd524, [%rd2+40]; mov.u64 %rd2390, 0; ld.u32 %r757, [%rd524+24]; setp.eq.s32 %p428, %r757, 0; @%p428 bra $L__BB5_687; add.s64 %rd525, %rd524, 64; shr.u32 %r3958, %r471, 12; cvt.u64.u32 %rd1988, %r3958; shr.u32 %r3959, %r733, 12; cvt.u64.u32 %rd1989, %r3959; bfi.b64 %rd1990, %rd1988, %rd1989, 42, 22; shr.u32 %r3960, %r748, 12; mul.wide.u32 %rd1991, %r3960, 2097152; or.b64 %rd526, %rd1990, %rd1991; mov.u32 %r5199, 0; $L__BB5_684: cvt.u64.u32 %rd527, %r5199; mul.wide.u32 %rd1992, %r5199, 32; add.s64 %rd1993, %rd525, %rd1992; ld.u64 %rd1994, [%rd1993]; setp.eq.s64 %p429, %rd1994, %rd526; add.s32 %r5199, %r5199, 1; @%p429 bra $L__BB5_686; setp.lt.u32 %p430, %r5199, %r757; @%p430 bra $L__BB5_684; bra.uni $L__BB5_687; $L__BB5_717: ld.local.u64 %rd542, [%rd2+40]; mov.u64 %rd2391, 0; ld.u32 %r763, [%rd542+24]; setp.eq.s32 %p450, %r763, 0; @%p450 bra $L__BB5_722; add.s64 %rd543, %rd542, 64; shr.u32 %r4125, %r635, 12; cvt.u64.u32 %rd2055, %r4125; shr.u32 %r4126, %r733, 12; cvt.u64.u32 %rd2056, %r4126; bfi.b64 %rd2057, %rd2055, %rd2056, 42, 22; shr.u32 %r4127, %r748, 12; mul.wide.u32 %rd2058, %r4127, 2097152; or.b64 %rd544, %rd2057, %rd2058; mov.u32 %r5201, 0; $L__BB5_719: cvt.u64.u32 %rd545, %r5201; mul.wide.u32 %rd2059, %r5201, 32; add.s64 %rd2060, %rd543, %rd2059; ld.u64 %rd2061, [%rd2060]; setp.eq.s64 %p451, %rd2061, %rd544; add.s32 %r5201, %r5201, 1; @%p451 bra $L__BB5_721; setp.lt.u32 %p452, %r5201, %r763; @%p452 bra $L__BB5_719; bra.uni $L__BB5_722; $L__BB5_752: ld.local.u64 %rd560, [%rd2+40]; mov.u64 %rd2392, 0; ld.u32 %r769, [%rd560+24]; setp.eq.s32 %p472, %r769, 0; @%p472 bra $L__BB5_757; add.s64 %rd561, %rd560, 64; shr.u32 %r4295, %r635, 12; cvt.u64.u32 %rd2122, %r4295; shr.u32 %r4296, %r741, 12; cvt.u64.u32 %rd2123, %r4296; bfi.b64 %rd2124, %rd2122, %rd2123, 42, 22; shr.u32 %r4297, %r748, 12; mul.wide.u32 %rd2125, %r4297, 2097152; or.b64 %rd562, %rd2124, %rd2125; mov.u32 %r5203, 0; $L__BB5_754: cvt.u64.u32 %rd563, %r5203; mul.wide.u32 %rd2126, %r5203, 32; add.s64 %rd2127, %rd561, %rd2126; ld.u64 %rd2128, [%rd2127]; setp.eq.s64 %p473, %rd2128, %rd562; add.s32 %r5203, %r5203, 1; @%p473 bra $L__BB5_756; setp.lt.u32 %p474, %r5203, %r769; @%p474 bra $L__BB5_754; bra.uni $L__BB5_757; $L__BB5_787: ld.local.u64 %rd578, [%rd2+40]; mov.u64 %rd2393, 0; ld.u32 %r775, [%rd578+24]; setp.eq.s32 %p494, %r775, 0; @%p494 bra $L__BB5_792; add.s64 %rd579, %rd578, 64; shr.u32 %r4466, %r635, 12; cvt.u64.u32 %rd2189, %r4466; shl.b64 %rd2190, %rd2189, 42; shr.u32 %r4467, %r732, 12; mul.wide.u32 %rd2191, %r4467, 2097152; or.b64 %rd2192, %rd2190, %rd2191; shr.u32 %r4468, %r741, 12; cvt.u64.u32 %rd2193, %r4468; or.b64 %rd580, %rd2192, %rd2193; mov.u32 %r5205, 0; $L__BB5_789: cvt.u64.u32 %rd581, %r5205; mul.wide.u32 %rd2194, %r5205, 32; add.s64 %rd2195, %rd579, %rd2194; ld.u64 %rd2196, [%rd2195]; setp.eq.s64 %p495, %rd2196, %rd580; add.s32 %r5205, %r5205, 1; @%p495 bra $L__BB5_791; setp.lt.u32 %p496, %r5205, %r775; @%p496 bra $L__BB5_789; bra.uni $L__BB5_792; $L__BB5_831: ld.f32 %f952, [%rd607]; bra.uni $L__BB5_833; $L__BB5_828: ld.f32 %f952, [%rd603]; bra.uni $L__BB5_833; $L__BB5_51: ld.f32 %f930, [%rd45]; bra.uni $L__BB5_66; $L__BB5_86: ld.f32 %f931, [%rd74]; bra.uni $L__BB5_101; $L__BB5_121: ld.f32 %f932, [%rd103]; bra.uni $L__BB5_136; $L__BB5_156: ld.f32 %f933, [%rd132]; bra.uni $L__BB5_171; $L__BB5_191: ld.f32 %f934, [%rd161]; bra.uni $L__BB5_206; $L__BB5_226: ld.f32 %f935, [%rd190]; bra.uni $L__BB5_241; $L__BB5_261: ld.f32 %f936, [%rd219]; bra.uni $L__BB5_276; $L__BB5_296: ld.f32 %f937, [%rd248]; bra.uni $L__BB5_311; $L__BB5_331: ld.f32 %f938, [%rd277]; bra.uni $L__BB5_346; $L__BB5_366: ld.f32 %f939, [%rd306]; bra.uni $L__BB5_381; $L__BB5_401: ld.f32 %f940, [%rd335]; bra.uni $L__BB5_416; $L__BB5_436: ld.f32 %f941, [%rd364]; bra.uni $L__BB5_451; $L__BB5_471: ld.f32 %f942, [%rd393]; bra.uni $L__BB5_486; $L__BB5_506: ld.f32 %f943, [%rd422]; bra.uni $L__BB5_521; $L__BB5_541: ld.f32 %f944, [%rd450]; bra.uni $L__BB5_556; $L__BB5_576: ld.f32 %f945, [%rd467]; bra.uni $L__BB5_591; $L__BB5_611: ld.f32 %f946, [%rd485]; bra.uni $L__BB5_626; $L__BB5_646: ld.f32 %f947, [%rd503]; bra.uni $L__BB5_661; $L__BB5_681: ld.f32 %f948, [%rd521]; bra.uni $L__BB5_696; $L__BB5_716: ld.f32 %f949, [%rd539]; bra.uni $L__BB5_731; $L__BB5_751: ld.f32 %f950, [%rd557]; bra.uni $L__BB5_766; $L__BB5_786: ld.f32 %f951, [%rd575]; bra.uni $L__BB5_801; $L__BB5_827: ld.f32 %f952, [%rd605]; $L__BB5_833: ld.param.f32 %f928, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; add.ftz.f32 %f927, %f928, 0fBF000000; cvt.rmi.ftz.f32.f32 %f926, %f927; sub.ftz.f32 %f925, %f927, %f926; mul.ftz.f32 %f924, %f925, 0f40400000; mov.f32 %f923, 0f40400000; sub.ftz.f32 %f922, %f923, %f924; fma.rn.ftz.f32 %f921, %f925, %f922, 0f40400000; mul.ftz.f32 %f920, %f925, %f925; mul.ftz.f32 %f919, %f925, %f920; mul.ftz.f32 %f918, %f919, 0f3E2AAAAB; fma.rn.ftz.f32 %f917, %f925, %f921, 0f3F800000; fma.rn.ftz.f32 %f916, %f917, 0f3E2AAAAB, %f918; sub.ftz.f32 %f907, %f951, %f952; fma.rn.ftz.f32 %f908, %f220, %f907, %f952; sub.ftz.f32 %f909, %f295, %f908; fma.rn.ftz.f32 %f910, %f219, %f909, %f908; sub.ftz.f32 %f911, %f910, %f270; fma.rn.ftz.f32 %f912, %f116, %f911, %f270; fma.rn.ftz.f32 %f913, %f8, %f912, %f218; fma.rn.ftz.f32 %f914, %f10, %f913, %f217; fma.rn.ftz.f32 %f953, %f916, %f914, %f216; $L__BB5_834: st.param.f32 [func_retval0+0], %f953; ret; $L__BB5_50: ld.f32 %f930, [%rd47]; bra.uni $L__BB5_66; $L__BB5_85: ld.f32 %f931, [%rd76]; bra.uni $L__BB5_101; $L__BB5_120: ld.f32 %f932, [%rd105]; bra.uni $L__BB5_136; $L__BB5_155: and.b32 %r4870, %r259, -128; ld.f32 %f933, [%rd134]; bra.uni $L__BB5_171; $L__BB5_190: and.b32 %r4870, %r259, -128; ld.f32 %f934, [%rd163]; bra.uni $L__BB5_206; $L__BB5_225: ld.f32 %f935, [%rd192]; mov.u32 %r4870, %r372; bra.uni $L__BB5_241; $L__BB5_260: ld.f32 %f936, [%rd221]; mov.u32 %r4870, %r423; bra.uni $L__BB5_276; $L__BB5_295: ld.f32 %f937, [%rd250]; mov.u32 %r4870, %r473; bra.uni $L__BB5_311; $L__BB5_330: ld.f32 %f938, [%rd279]; mov.u32 %r4870, %r518; bra.uni $L__BB5_346; $L__BB5_365: ld.f32 %f939, [%rd308]; mov.u32 %r4870, %r560; bra.uni $L__BB5_381; $L__BB5_400: ld.f32 %f940, [%rd337]; mov.u32 %r4870, %r599; bra.uni $L__BB5_416; $L__BB5_435: ld.f32 %f941, [%rd366]; mov.u32 %r4870, %r637; bra.uni $L__BB5_451; $L__BB5_470: ld.f32 %f942, [%rd395]; mov.u32 %r4870, %r670; bra.uni $L__BB5_486; $L__BB5_505: ld.f32 %f943, [%rd424]; mov.u32 %r4870, %r700; bra.uni $L__BB5_521; $L__BB5_540: ld.f32 %f944, [%rd452]; bra.uni $L__BB5_556; $L__BB5_575: ld.f32 %f945, [%rd469]; bra.uni $L__BB5_591; $L__BB5_610: ld.f32 %f946, [%rd487]; bra.uni $L__BB5_626; $L__BB5_645: ld.f32 %f947, [%rd505]; bra.uni $L__BB5_661; $L__BB5_680: ld.f32 %f948, [%rd523]; bra.uni $L__BB5_696; $L__BB5_715: ld.f32 %f949, [%rd541]; bra.uni $L__BB5_731; $L__BB5_750: ld.f32 %f950, [%rd559]; bra.uni $L__BB5_766; $L__BB5_785: ld.f32 %f951, [%rd577]; bra.uni $L__BB5_801; $L__BB5_56: cvt.u64.u32 %rd2291, %r4779; shl.b64 %rd742, %rd2291, 5; add.s64 %rd2302, %rd17, %rd742; $L__BB5_57: setp.eq.s64 %p35, %rd2302, 0; @%p35 bra $L__BB5_65; add.s64 %rd55, %rd2302, 8; ld.u64 %rd56, [%rd2302+8]; setp.eq.s64 %p36, %rd56, 0; @%p36 bra $L__BB5_64; add.s64 %rd2315, %rd1, %rd56; and.b32 %r1137, %r8, -4096; mov.u64 %rd743, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1137}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1138, %r7, 3; and.b32 %r1139, %r1138, 31744; and.b32 %r1140, %r8, 3968; shr.u32 %r1141, %r1140, 2; or.b32 %r1142, %r1141, %r1139; bfe.u32 %r1143, %r74, 7, 5; or.b32 %r1144, %r1143, %r1142; shr.u32 %r1145, %r1142, 6; mul.wide.u32 %rd744, %r1145, 8; add.s64 %rd745, %rd2315, %rd744; and.b32 %r1146, %r1144, 63; shl.b64 %rd746, %rd743, %r1146; ld.u64 %rd747, [%rd745+4128]; and.b64 %rd748, %rd747, %rd746; setp.eq.s64 %p37, %rd748, 0; mul.wide.u32 %rd749, %r1144, 8; add.s64 %rd750, %rd2315, %rd749; add.s64 %rd58, %rd750, 8256; mov.b64 %rd751, {%r12, %r1137}; shr.u64 %rd752, %rd751, 32; cvt.u32.u64 %r4875, %rd752; @%p37 bra $L__BB5_63; and.b32 %r4755, %r7, -128; ld.u64 %rd753, [%rd58]; add.s64 %rd2311, %rd2315, %rd753; st.local.u32 [%rd2+12], %r4755; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1147, %r7, 5; and.b32 %r1148, %r1147, 3840; shl.b32 %r1149, %r8, 1; and.b32 %r1150, %r1149, 240; or.b32 %r1151, %r1150, %r1148; bfe.u32 %r1152, %r74, 3, 4; or.b32 %r1153, %r1152, %r1151; shr.u32 %r1154, %r1151, 6; mul.wide.u32 %rd754, %r1154, 8; add.s64 %rd755, %rd2311, %rd754; and.b32 %r1155, %r1153, 63; shl.b64 %rd757, %rd743, %r1155; ld.u64 %rd758, [%rd755+544]; and.b64 %rd759, %rd758, %rd757; setp.eq.s64 %p38, %rd759, 0; mul.wide.u32 %rd760, %r1153, 8; add.s64 %rd761, %rd2311, %rd760; add.s64 %rd60, %rd761, 1088; and.b32 %r4870, %r7, -128; @%p38 bra $L__BB5_62; and.b32 %r4827, %r7, -8; ld.u64 %rd762, [%rd60]; add.s64 %rd2307, %rd2311, %rd762; and.b32 %r1156, %r8, -8; mov.u32 %r1157, 1; st.local.v2.u32 [%rd2], {%r4827, %r1156}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1158, %r7, 6; and.b32 %r1159, %r1158, 448; shl.b32 %r1160, %r8, 3; and.b32 %r1161, %r1160, 56; or.b32 %r1162, %r1161, %r1159; and.b32 %r1163, %r74, 7; or.b32 %r1164, %r1162, %r1163; ld.u8 %rs149, [%rd2307+15]; shr.u16 %rs150, %rs149, 5; mov.u32 %r1165, 5; cvt.u32.u16 %r1166, %rs150; sub.s32 %r1167, %r1165, %r1166; shr.u32 %r1168, %r1164, %r1167; mul.wide.u32 %rd763, %r1168, 4; add.s64 %rd764, %rd2307, %rd763; mov.u32 %r1169, 32; shr.u32 %r1170, %r1169, %r1166; add.s32 %r1171, %r1170, -1; and.b32 %r1172, %r1171, %r1164; shl.b32 %r1173, %r1172, %r1166; ld.u32 %r1174, [%rd764+96]; shr.u32 %r1175, %r1174, %r1173; shl.b32 %r1176, %r1157, %r1166; mov.u32 %r1177, -1; shl.b32 %r1178, %r1177, %r1176; not.b32 %r1179, %r1178; and.b32 %r1180, %r1175, %r1179; cvt.rn.f32.u32 %f419, %r1180; ld.v2.f32 {%f420, %f421}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f421, %f419, %f420; mov.b64 %rd765, {%r4827, %r1156}; shr.u64 %rd766, %rd765, 32; cvt.u32.u64 %r4782, %rd766; mov.u32 %r4830, %r12; bra.uni $L__BB5_66; $L__BB5_65: ld.f32 %f930, [%rd5+4]; bra.uni $L__BB5_66; $L__BB5_91: cvt.u64.u32 %rd2294, %r4828; shl.b64 %rd812, %rd2294, 5; add.s64 %rd2308, %rd17, %rd812; $L__BB5_92: setp.eq.s64 %p57, %rd2308, 0; @%p57 bra $L__BB5_100; add.s64 %rd84, %rd2308, 8; ld.u64 %rd85, [%rd2308+8]; setp.eq.s64 %p58, %rd85, 0; @%p58 bra $L__BB5_99; add.s64 %rd2315, %rd1, %rd85; and.b32 %r1291, %r139, -4096; mov.u64 %rd813, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1291}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1292, %r7, 3; and.b32 %r1293, %r1292, 31744; and.b32 %r1294, %r139, 3968; shr.u32 %r1295, %r1294, 2; or.b32 %r1296, %r1295, %r1293; bfe.u32 %r1297, %r74, 7, 5; or.b32 %r1298, %r1296, %r1297; shr.u32 %r1299, %r1296, 6; mul.wide.u32 %rd814, %r1299, 8; add.s64 %rd815, %rd2315, %rd814; and.b32 %r1300, %r1298, 63; shl.b64 %rd816, %rd813, %r1300; ld.u64 %rd817, [%rd815+4128]; and.b64 %rd818, %rd817, %rd816; setp.eq.s64 %p59, %rd818, 0; mul.wide.u32 %rd819, %r1298, 8; add.s64 %rd820, %rd2315, %rd819; add.s64 %rd87, %rd820, 8256; mov.b64 %rd821, {%r12, %r1291}; shr.u64 %rd822, %rd821, 32; cvt.u32.u64 %r4875, %rd822; @%p59 bra $L__BB5_98; and.b32 %r4729, %r7, -128; ld.u64 %rd823, [%rd87]; add.s64 %rd2311, %rd2315, %rd823; st.local.u32 [%rd2+12], %r4729; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1301, %r7, 5; and.b32 %r1302, %r1301, 3840; shl.b32 %r1303, %r139, 1; and.b32 %r1304, %r1303, 240; or.b32 %r1305, %r1304, %r1302; bfe.u32 %r1306, %r74, 3, 4; or.b32 %r1307, %r1305, %r1306; shr.u32 %r1308, %r1305, 6; mul.wide.u32 %rd824, %r1308, 8; add.s64 %rd825, %rd2311, %rd824; and.b32 %r1309, %r1307, 63; shl.b64 %rd827, %rd813, %r1309; ld.u64 %rd828, [%rd825+544]; and.b64 %rd829, %rd828, %rd827; setp.eq.s64 %p60, %rd829, 0; mul.wide.u32 %rd830, %r1307, 8; add.s64 %rd831, %rd2311, %rd830; add.s64 %rd89, %rd831, 1088; and.b32 %r4870, %r7, -128; @%p60 bra $L__BB5_97; and.b32 %r4827, %r7, -8; ld.u64 %rd832, [%rd89]; add.s64 %rd2307, %rd2311, %rd832; and.b32 %r1310, %r139, -8; mov.u32 %r1311, 1; st.local.v2.u32 [%rd2], {%r4827, %r1310}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1312, %r7, 6; and.b32 %r1313, %r1312, 448; shl.b32 %r1314, %r139, 3; and.b32 %r1315, %r1314, 56; or.b32 %r1316, %r1315, %r1313; and.b32 %r1317, %r74, 7; or.b32 %r1318, %r1316, %r1317; ld.u8 %rs163, [%rd2307+15]; shr.u16 %rs164, %rs163, 5; mov.u32 %r1319, 5; cvt.u32.u16 %r1320, %rs164; sub.s32 %r1321, %r1319, %r1320; shr.u32 %r1322, %r1318, %r1321; mul.wide.u32 %rd833, %r1322, 4; add.s64 %rd834, %rd2307, %rd833; mov.u32 %r1323, 32; shr.u32 %r1324, %r1323, %r1320; add.s32 %r1325, %r1324, -1; and.b32 %r1326, %r1325, %r1318; shl.b32 %r1327, %r1326, %r1320; ld.u32 %r1328, [%rd834+96]; shr.u32 %r1329, %r1328, %r1327; shl.b32 %r1330, %r1311, %r1320; mov.u32 %r1331, -1; shl.b32 %r1332, %r1331, %r1330; not.b32 %r1333, %r1332; and.b32 %r1334, %r1329, %r1333; cvt.rn.f32.u32 %f440, %r1334; ld.v2.f32 {%f441, %f442}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f442, %f440, %f441; mov.b64 %rd835, {%r4827, %r1310}; shr.u64 %rd836, %rd835, 32; cvt.u32.u64 %r4782, %rd836; mov.u32 %r4830, %r12; bra.uni $L__BB5_101; $L__BB5_100: ld.f32 %f931, [%rd5+4]; bra.uni $L__BB5_101; $L__BB5_126: cvt.u64.u32 %rd2295, %r4874; shl.b64 %rd880, %rd2295, 5; add.s64 %rd2314, %rd17, %rd880; $L__BB5_127: setp.eq.s64 %p79, %rd2314, 0; @%p79 bra $L__BB5_135; add.s64 %rd113, %rd2314, 8; ld.u64 %rd114, [%rd2314+8]; setp.eq.s64 %p80, %rd114, 0; @%p80 bra $L__BB5_134; add.s64 %rd2315, %rd1, %rd114; and.b32 %r1441, %r139, -4096; mov.u64 %rd881, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1441}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1442, %r7, 3; and.b32 %r1443, %r1442, 31744; and.b32 %r1444, %r139, 3968; shr.u32 %r1445, %r1444, 2; or.b32 %r1446, %r1445, %r1443; bfe.u32 %r1447, %r9, 7, 5; or.b32 %r1448, %r1446, %r1447; shr.u32 %r1449, %r1446, 6; mul.wide.u32 %rd882, %r1449, 8; add.s64 %rd883, %rd2315, %rd882; and.b32 %r1450, %r1448, 63; shl.b64 %rd884, %rd881, %r1450; ld.u64 %rd885, [%rd883+4128]; and.b64 %rd886, %rd885, %rd884; setp.eq.s64 %p81, %rd886, 0; mul.wide.u32 %rd887, %r1448, 8; add.s64 %rd888, %rd2315, %rd887; add.s64 %rd116, %rd888, 8256; mov.b64 %rd889, {%r12, %r1441}; shr.u64 %rd890, %rd889, 32; cvt.u32.u64 %r4875, %rd890; @%p81 bra $L__BB5_133; and.b32 %r4702, %r7, -128; ld.u64 %rd891, [%rd116]; add.s64 %rd2311, %rd2315, %rd891; st.local.u32 [%rd2+12], %r4702; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1451, %r7, 5; and.b32 %r1452, %r1451, 3840; shl.b32 %r1453, %r139, 1; and.b32 %r1454, %r1453, 240; or.b32 %r1455, %r1454, %r1452; bfe.u32 %r1456, %r9, 3, 4; or.b32 %r1457, %r1455, %r1456; shr.u32 %r1458, %r1455, 6; mul.wide.u32 %rd892, %r1458, 8; add.s64 %rd893, %rd2311, %rd892; and.b32 %r1459, %r1457, 63; shl.b64 %rd895, %rd881, %r1459; ld.u64 %rd896, [%rd893+544]; and.b64 %rd897, %rd896, %rd895; setp.eq.s64 %p82, %rd897, 0; mul.wide.u32 %rd898, %r1457, 8; add.s64 %rd899, %rd2311, %rd898; add.s64 %rd118, %rd899, 1088; and.b32 %r4870, %r7, -128; @%p82 bra $L__BB5_132; and.b32 %r4827, %r7, -8; ld.u64 %rd900, [%rd118]; add.s64 %rd2307, %rd2311, %rd900; and.b32 %r1460, %r139, -8; mov.u32 %r1461, 1; st.local.v2.u32 [%rd2], {%r4827, %r1460}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1462, %r7, 7; shl.b32 %r1463, %r139, 3; and.b32 %r1464, %r1463, 56; and.b32 %r1465, %r9, 7; bfi.b32 %r1466, %r1462, %r1465, 6, 3; or.b32 %r1467, %r1466, %r1464; ld.u8 %rs177, [%rd2307+15]; shr.u16 %rs178, %rs177, 5; mov.u32 %r1468, 5; cvt.u32.u16 %r1469, %rs178; sub.s32 %r1470, %r1468, %r1469; shr.u32 %r1471, %r1467, %r1470; mul.wide.u32 %rd901, %r1471, 4; add.s64 %rd902, %rd2307, %rd901; mov.u32 %r1472, 32; shr.u32 %r1473, %r1472, %r1469; add.s32 %r1474, %r1473, -1; and.b32 %r1475, %r1474, %r1467; shl.b32 %r1476, %r1475, %r1469; ld.u32 %r1477, [%rd902+96]; shr.u32 %r1478, %r1477, %r1476; shl.b32 %r1479, %r1461, %r1469; mov.u32 %r1480, -1; shl.b32 %r1481, %r1480, %r1479; not.b32 %r1482, %r1481; and.b32 %r1483, %r1478, %r1482; cvt.rn.f32.u32 %f460, %r1483; ld.v2.f32 {%f461, %f462}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f462, %f460, %f461; mov.b64 %rd903, {%r4827, %r1460}; shr.u64 %rd904, %rd903, 32; cvt.u32.u64 %r4782, %rd904; mov.u32 %r4830, %r12; bra.uni $L__BB5_136; $L__BB5_135: ld.f32 %f932, [%rd5+4]; bra.uni $L__BB5_136; $L__BB5_161: cvt.u64.u32 %rd2296, %r4917; shl.b64 %rd951, %rd2296, 5; add.s64 %rd2320, %rd17, %rd951; $L__BB5_162: setp.eq.s64 %p101, %rd2320, 0; @%p101 bra $L__BB5_170; add.s64 %rd142, %rd2320, 8; ld.u64 %rd143, [%rd2320+8]; setp.eq.s64 %p102, %rd143, 0; @%p102 bra $L__BB5_169; and.b32 %r4707, %r259, -4096; add.s64 %rd2315, %rd1, %rd143; and.b32 %r1591, %r139, -4096; mov.u64 %rd952, 1; st.local.v2.u32 [%rd2+24], {%r4707, %r1591}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1592, %r259, 3; and.b32 %r1593, %r1592, 31744; and.b32 %r1594, %r139, 3968; shr.u32 %r1595, %r1594, 2; or.b32 %r1596, %r1595, %r1593; bfe.u32 %r1597, %r9, 7, 5; or.b32 %r1598, %r1596, %r1597; shr.u32 %r1599, %r1596, 6; mul.wide.u32 %rd953, %r1599, 8; add.s64 %rd954, %rd2315, %rd953; and.b32 %r1600, %r1598, 63; shl.b64 %rd955, %rd952, %r1600; ld.u64 %rd956, [%rd954+4128]; and.b64 %rd957, %rd956, %rd955; setp.eq.s64 %p103, %rd957, 0; mul.wide.u32 %rd958, %r1598, 8; add.s64 %rd959, %rd2315, %rd958; add.s64 %rd145, %rd959, 8256; mov.b64 %rd960, {%r4707, %r1591}; shr.u64 %rd961, %rd960, 32; cvt.u32.u64 %r4875, %rd961; @%p103 bra $L__BB5_168; and.b32 %r4708, %r259, -128; ld.u64 %rd962, [%rd145]; add.s64 %rd2311, %rd2315, %rd962; st.local.u32 [%rd2+12], %r4708; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1601, %r259, 5; and.b32 %r1602, %r1601, 3840; shl.b32 %r1603, %r139, 1; and.b32 %r1604, %r1603, 240; or.b32 %r1605, %r1604, %r1602; bfe.u32 %r1606, %r9, 3, 4; or.b32 %r1607, %r1605, %r1606; shr.u32 %r1608, %r1605, 6; mul.wide.u32 %rd963, %r1608, 8; add.s64 %rd964, %rd2311, %rd963; and.b32 %r1609, %r1607, 63; shl.b64 %rd966, %rd952, %r1609; ld.u64 %rd967, [%rd964+544]; and.b64 %rd968, %rd967, %rd966; setp.eq.s64 %p104, %rd968, 0; mul.wide.u32 %rd969, %r1607, 8; add.s64 %rd970, %rd2311, %rd969; add.s64 %rd147, %rd970, 1088; @%p104 bra $L__BB5_167; and.b32 %r4827, %r259, -8; and.b32 %r4830, %r259, -4096; and.b32 %r4870, %r259, -128; ld.u64 %rd971, [%rd147]; add.s64 %rd2307, %rd2311, %rd971; and.b32 %r1610, %r139, -8; mov.u32 %r1611, 1; st.local.v2.u32 [%rd2], {%r4827, %r1610}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1612, %r259, 7; shl.b32 %r1613, %r139, 3; and.b32 %r1614, %r1613, 56; and.b32 %r1615, %r9, 7; bfi.b32 %r1616, %r1612, %r1615, 6, 3; or.b32 %r1617, %r1616, %r1614; ld.u8 %rs191, [%rd2307+15]; shr.u16 %rs192, %rs191, 5; mov.u32 %r1618, 5; cvt.u32.u16 %r1619, %rs192; sub.s32 %r1620, %r1618, %r1619; shr.u32 %r1621, %r1617, %r1620; mul.wide.u32 %rd972, %r1621, 4; add.s64 %rd973, %rd2307, %rd972; mov.u32 %r1622, 32; shr.u32 %r1623, %r1622, %r1619; add.s32 %r1624, %r1623, -1; and.b32 %r1625, %r1624, %r1617; shl.b32 %r1626, %r1625, %r1619; ld.u32 %r1627, [%rd973+96]; shr.u32 %r1628, %r1627, %r1626; shl.b32 %r1629, %r1611, %r1619; mov.u32 %r1630, -1; shl.b32 %r1631, %r1630, %r1629; not.b32 %r1632, %r1631; and.b32 %r1633, %r1628, %r1632; cvt.rn.f32.u32 %f483, %r1633; ld.v2.f32 {%f484, %f485}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f485, %f483, %f484; mov.b64 %rd974, {%r4827, %r1610}; shr.u64 %rd975, %rd974, 32; cvt.u32.u64 %r4782, %rd975; bra.uni $L__BB5_171; $L__BB5_170: ld.f32 %f933, [%rd5+4]; bra.uni $L__BB5_171; $L__BB5_196: shl.b64 %rd1022, %rd168, 5; add.s64 %rd2326, %rd17, %rd1022; $L__BB5_197: setp.eq.s64 %p123, %rd2326, 0; @%p123 bra $L__BB5_205; add.s64 %rd171, %rd2326, 8; ld.u64 %rd172, [%rd2326+8]; setp.eq.s64 %p124, %rd172, 0; @%p124 bra $L__BB5_204; and.b32 %r4714, %r259, -4096; add.s64 %rd2315, %rd1, %rd172; and.b32 %r1745, %r139, -4096; mov.u64 %rd1023, 1; st.local.v2.u32 [%rd2+24], {%r4714, %r1745}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1746, %r259, 3; and.b32 %r1747, %r1746, 31744; and.b32 %r1748, %r139, 3968; shr.u32 %r1749, %r1748, 2; or.b32 %r1750, %r1749, %r1747; bfe.u32 %r1751, %r74, 7, 5; or.b32 %r1752, %r1750, %r1751; shr.u32 %r1753, %r1750, 6; mul.wide.u32 %rd1024, %r1753, 8; add.s64 %rd1025, %rd2315, %rd1024; and.b32 %r1754, %r1752, 63; shl.b64 %rd1026, %rd1023, %r1754; ld.u64 %rd1027, [%rd1025+4128]; and.b64 %rd1028, %rd1027, %rd1026; setp.eq.s64 %p125, %rd1028, 0; mul.wide.u32 %rd1029, %r1752, 8; add.s64 %rd1030, %rd2315, %rd1029; add.s64 %rd174, %rd1030, 8256; mov.b64 %rd1031, {%r4714, %r1745}; shr.u64 %rd1032, %rd1031, 32; cvt.u32.u64 %r4875, %rd1032; @%p125 bra $L__BB5_203; and.b32 %r4715, %r259, -128; ld.u64 %rd1033, [%rd174]; add.s64 %rd2311, %rd2315, %rd1033; st.local.u32 [%rd2+12], %r4715; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1755, %r259, 5; and.b32 %r1756, %r1755, 3840; shl.b32 %r1757, %r139, 1; and.b32 %r1758, %r1757, 240; or.b32 %r1759, %r1758, %r1756; bfe.u32 %r1760, %r74, 3, 4; or.b32 %r1761, %r1759, %r1760; shr.u32 %r1762, %r1759, 6; mul.wide.u32 %rd1034, %r1762, 8; add.s64 %rd1035, %rd2311, %rd1034; and.b32 %r1763, %r1761, 63; shl.b64 %rd1037, %rd1023, %r1763; ld.u64 %rd1038, [%rd1035+544]; and.b64 %rd1039, %rd1038, %rd1037; setp.eq.s64 %p126, %rd1039, 0; mul.wide.u32 %rd1040, %r1761, 8; add.s64 %rd1041, %rd2311, %rd1040; add.s64 %rd176, %rd1041, 1088; @%p126 bra $L__BB5_202; and.b32 %r4827, %r259, -8; and.b32 %r4830, %r259, -4096; and.b32 %r4870, %r259, -128; ld.u64 %rd1042, [%rd176]; add.s64 %rd2307, %rd2311, %rd1042; and.b32 %r1764, %r139, -8; mov.u32 %r1765, 1; st.local.v2.u32 [%rd2], {%r4827, %r1764}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1766, %r259, 6; and.b32 %r1767, %r1766, 448; shl.b32 %r1768, %r139, 3; and.b32 %r1769, %r1768, 56; or.b32 %r1770, %r1769, %r1767; and.b32 %r1771, %r74, 7; or.b32 %r1772, %r1770, %r1771; ld.u8 %rs205, [%rd2307+15]; shr.u16 %rs206, %rs205, 5; mov.u32 %r1773, 5; cvt.u32.u16 %r1774, %rs206; sub.s32 %r1775, %r1773, %r1774; shr.u32 %r1776, %r1772, %r1775; mul.wide.u32 %rd1043, %r1776, 4; add.s64 %rd1044, %rd2307, %rd1043; mov.u32 %r1777, 32; shr.u32 %r1778, %r1777, %r1774; add.s32 %r1779, %r1778, -1; and.b32 %r1780, %r1779, %r1772; shl.b32 %r1781, %r1780, %r1774; ld.u32 %r1782, [%rd1044+96]; shr.u32 %r1783, %r1782, %r1781; shl.b32 %r1784, %r1765, %r1774; mov.u32 %r1785, -1; shl.b32 %r1786, %r1785, %r1784; not.b32 %r1787, %r1786; and.b32 %r1788, %r1783, %r1787; cvt.rn.f32.u32 %f503, %r1788; ld.v2.f32 {%f504, %f505}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f505, %f503, %f504; mov.b64 %rd1045, {%r4827, %r1764}; shr.u64 %rd1046, %rd1045, 32; cvt.u32.u64 %r4782, %rd1046; bra.uni $L__BB5_206; $L__BB5_205: ld.f32 %f934, [%rd5+4]; bra.uni $L__BB5_206; $L__BB5_231: shl.b64 %rd1093, %rd197, 5; add.s64 %rd2332, %rd17, %rd1093; $L__BB5_232: setp.eq.s64 %p145, %rd2332, 0; @%p145 bra $L__BB5_240; add.s64 %rd200, %rd2332, 8; ld.u64 %rd201, [%rd2332+8]; setp.eq.s64 %p146, %rd201, 0; @%p146 bra $L__BB5_239; and.b32 %r4738, %r259, -4096; add.s64 %rd2315, %rd1, %rd201; and.b32 %r1899, %r8, -4096; mov.u64 %rd1094, 1; st.local.v2.u32 [%rd2+24], {%r4738, %r1899}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1900, %r259, 3; and.b32 %r1901, %r1900, 31744; and.b32 %r1902, %r8, 3968; shr.u32 %r1903, %r1902, 2; or.b32 %r1904, %r1901, %r1903; bfe.u32 %r1905, %r74, 7, 5; or.b32 %r1906, %r1905, %r1904; shr.u32 %r1907, %r1904, 6; mul.wide.u32 %rd1095, %r1907, 8; add.s64 %rd1096, %rd2315, %rd1095; and.b32 %r1908, %r1906, 63; shl.b64 %rd1097, %rd1094, %r1908; ld.u64 %rd1098, [%rd1096+4128]; and.b64 %rd1099, %rd1098, %rd1097; setp.eq.s64 %p147, %rd1099, 0; mul.wide.u32 %rd1100, %r1906, 8; add.s64 %rd1101, %rd2315, %rd1100; add.s64 %rd203, %rd1101, 8256; mov.b64 %rd1102, {%r4738, %r1899}; shr.u64 %rd1103, %rd1102, 32; cvt.u32.u64 %r4875, %rd1103; @%p147 bra $L__BB5_238; and.b32 %r4739, %r259, -128; ld.u64 %rd1104, [%rd203]; add.s64 %rd2311, %rd2315, %rd1104; st.local.u32 [%rd2+12], %r4739; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1909, %r259, 5; and.b32 %r1910, %r1909, 3840; shl.b32 %r1911, %r8, 1; and.b32 %r1912, %r1911, 240; or.b32 %r1913, %r1910, %r1912; bfe.u32 %r1914, %r74, 3, 4; or.b32 %r1915, %r1914, %r1913; shr.u32 %r1916, %r1913, 6; mul.wide.u32 %rd1105, %r1916, 8; add.s64 %rd1106, %rd2311, %rd1105; and.b32 %r1917, %r1915, 63; shl.b64 %rd1108, %rd1094, %r1917; ld.u64 %rd1109, [%rd1106+544]; and.b64 %rd1110, %rd1109, %rd1108; setp.eq.s64 %p148, %rd1110, 0; mul.wide.u32 %rd1111, %r1915, 8; add.s64 %rd1112, %rd2311, %rd1111; add.s64 %rd205, %rd1112, 1088; and.b32 %r4830, %r259, -4096; and.b32 %r4870, %r259, -128; @%p148 bra $L__BB5_237; and.b32 %r4827, %r259, -8; ld.u64 %rd1113, [%rd205]; add.s64 %rd2307, %rd2311, %rd1113; and.b32 %r1918, %r8, -8; mov.u32 %r1919, 1; st.local.v2.u32 [%rd2], {%r4827, %r1918}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1920, %r259, 6; and.b32 %r1921, %r1920, 448; shl.b32 %r1922, %r8, 3; and.b32 %r1923, %r1922, 56; or.b32 %r1924, %r1921, %r1923; and.b32 %r1925, %r74, 7; or.b32 %r1926, %r1924, %r1925; ld.u8 %rs219, [%rd2307+15]; shr.u16 %rs220, %rs219, 5; mov.u32 %r1927, 5; cvt.u32.u16 %r1928, %rs220; sub.s32 %r1929, %r1927, %r1928; shr.u32 %r1930, %r1926, %r1929; mul.wide.u32 %rd1114, %r1930, 4; add.s64 %rd1115, %rd2307, %rd1114; mov.u32 %r1931, 32; shr.u32 %r1932, %r1931, %r1928; add.s32 %r1933, %r1932, -1; and.b32 %r1934, %r1933, %r1926; shl.b32 %r1935, %r1934, %r1928; ld.u32 %r1936, [%rd1115+96]; shr.u32 %r1937, %r1936, %r1935; shl.b32 %r1938, %r1919, %r1928; mov.u32 %r1939, -1; shl.b32 %r1940, %r1939, %r1938; not.b32 %r1941, %r1940; and.b32 %r1942, %r1937, %r1941; cvt.rn.f32.u32 %f524, %r1942; ld.v2.f32 {%f525, %f526}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f526, %f524, %f525; mov.b64 %rd1116, {%r4827, %r1918}; shr.u64 %rd1117, %rd1116, 32; cvt.u32.u64 %r4782, %rd1117; bra.uni $L__BB5_241; $L__BB5_240: ld.f32 %f935, [%rd5+4]; bra.uni $L__BB5_241; $L__BB5_266: shl.b64 %rd1163, %rd226, 5; add.s64 %rd2338, %rd17, %rd1163; $L__BB5_267: setp.eq.s64 %p167, %rd2338, 0; @%p167 bra $L__BB5_275; add.s64 %rd229, %rd2338, 8; ld.u64 %rd230, [%rd2338+8]; setp.eq.s64 %p168, %rd230, 0; @%p168 bra $L__BB5_274; add.s64 %rd2315, %rd1, %rd230; and.b32 %r2052, %r8, -4096; mov.u64 %rd1164, 1; st.local.v2.u32 [%rd2+24], {%r426, %r2052}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2053, %r259, 3; and.b32 %r2054, %r2053, 31744; and.b32 %r2055, %r8, 3968; shr.u32 %r2056, %r2055, 2; or.b32 %r2057, %r2054, %r2056; bfe.u32 %r2058, %r9, 7, 5; or.b32 %r2059, %r2057, %r2058; shr.u32 %r2060, %r2057, 6; mul.wide.u32 %rd1165, %r2060, 8; add.s64 %rd1166, %rd2315, %rd1165; and.b32 %r2061, %r2059, 63; shl.b64 %rd1167, %rd1164, %r2061; ld.u64 %rd1168, [%rd1166+4128]; and.b64 %rd1169, %rd1168, %rd1167; setp.eq.s64 %p169, %rd1169, 0; mul.wide.u32 %rd1170, %r2059, 8; add.s64 %rd1171, %rd2315, %rd1170; add.s64 %rd232, %rd1171, 8256; mov.b64 %rd1172, {%r426, %r2052}; shr.u64 %rd1173, %rd1172, 32; cvt.u32.u64 %r4875, %rd1173; @%p169 bra $L__BB5_273; ld.u64 %rd1174, [%rd232]; add.s64 %rd2311, %rd2315, %rd1174; st.local.u32 [%rd2+12], %r423; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2062, %r259, 5; and.b32 %r2063, %r2062, 3840; shl.b32 %r2064, %r8, 1; and.b32 %r2065, %r2064, 240; or.b32 %r2066, %r2063, %r2065; bfe.u32 %r2067, %r9, 3, 4; or.b32 %r2068, %r2066, %r2067; shr.u32 %r2069, %r2066, 6; mul.wide.u32 %rd1175, %r2069, 8; add.s64 %rd1176, %rd2311, %rd1175; and.b32 %r2070, %r2068, 63; shl.b64 %rd1178, %rd1164, %r2070; ld.u64 %rd1179, [%rd1176+544]; and.b64 %rd1180, %rd1179, %rd1178; setp.eq.s64 %p170, %rd1180, 0; mul.wide.u32 %rd1181, %r2068, 8; add.s64 %rd1182, %rd2311, %rd1181; add.s64 %rd234, %rd1182, 1088; @%p170 bra $L__BB5_272; and.b32 %r4827, %r259, -8; ld.u64 %rd1183, [%rd234]; add.s64 %rd2307, %rd2311, %rd1183; and.b32 %r2071, %r8, -8; mov.u32 %r2072, 1; st.local.v2.u32 [%rd2], {%r4827, %r2071}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2073, %r259, 6; and.b32 %r2074, %r2073, 448; shl.b32 %r2075, %r8, 3; and.b32 %r2076, %r2075, 56; or.b32 %r2077, %r2074, %r2076; and.b32 %r2078, %r9, 7; or.b32 %r2079, %r2077, %r2078; ld.u8 %rs233, [%rd2307+15]; shr.u16 %rs234, %rs233, 5; mov.u32 %r2080, 5; cvt.u32.u16 %r2081, %rs234; sub.s32 %r2082, %r2080, %r2081; shr.u32 %r2083, %r2079, %r2082; mul.wide.u32 %rd1184, %r2083, 4; add.s64 %rd1185, %rd2307, %rd1184; mov.u32 %r2084, 32; shr.u32 %r2085, %r2084, %r2081; add.s32 %r2086, %r2085, -1; and.b32 %r2087, %r2086, %r2079; shl.b32 %r2088, %r2087, %r2081; ld.u32 %r2089, [%rd1185+96]; shr.u32 %r2090, %r2089, %r2088; shl.b32 %r2091, %r2072, %r2081; mov.u32 %r2092, -1; shl.b32 %r2093, %r2092, %r2091; not.b32 %r2094, %r2093; and.b32 %r2095, %r2090, %r2094; cvt.rn.f32.u32 %f544, %r2095; ld.v2.f32 {%f545, %f546}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f546, %f544, %f545; mov.b64 %rd1186, {%r4827, %r2071}; shr.u64 %rd1187, %rd1186, 32; cvt.u32.u64 %r4782, %rd1187; mov.u32 %r4830, %r426; mov.u32 %r4870, %r423; bra.uni $L__BB5_276; $L__BB5_275: ld.f32 %f936, [%rd5+4]; bra.uni $L__BB5_276; $L__BB5_301: shl.b64 %rd1233, %rd255, 5; add.s64 %rd2344, %rd17, %rd1233; $L__BB5_302: setp.eq.s64 %p189, %rd2344, 0; @%p189 bra $L__BB5_310; add.s64 %rd258, %rd2344, 8; ld.u64 %rd259, [%rd2344+8]; setp.eq.s64 %p190, %rd259, 0; @%p190 bra $L__BB5_309; add.s64 %rd2315, %rd1, %rd259; and.b32 %r2205, %r8, -4096; mov.u64 %rd1234, 1; st.local.v2.u32 [%rd2+24], {%r476, %r2205}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2206, %r471, 3; and.b32 %r2207, %r2206, 31744; and.b32 %r2208, %r8, 3968; shr.u32 %r2209, %r2208, 2; or.b32 %r2210, %r2209, %r2207; bfe.u32 %r2211, %r9, 7, 5; or.b32 %r2212, %r2210, %r2211; shr.u32 %r2213, %r2210, 6; mul.wide.u32 %rd1235, %r2213, 8; add.s64 %rd1236, %rd2315, %rd1235; and.b32 %r2214, %r2212, 63; shl.b64 %rd1237, %rd1234, %r2214; ld.u64 %rd1238, [%rd1236+4128]; and.b64 %rd1239, %rd1238, %rd1237; setp.eq.s64 %p191, %rd1239, 0; mul.wide.u32 %rd1240, %r2212, 8; add.s64 %rd1241, %rd2315, %rd1240; add.s64 %rd261, %rd1241, 8256; mov.b64 %rd1242, {%r476, %r2205}; shr.u64 %rd1243, %rd1242, 32; cvt.u32.u64 %r4875, %rd1243; @%p191 bra $L__BB5_308; ld.u64 %rd1244, [%rd261]; add.s64 %rd2311, %rd2315, %rd1244; st.local.u32 [%rd2+12], %r473; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2215, %r471, 5; and.b32 %r2216, %r2215, 3840; shl.b32 %r2217, %r8, 1; and.b32 %r2218, %r2217, 240; or.b32 %r2219, %r2218, %r2216; bfe.u32 %r2220, %r9, 3, 4; or.b32 %r2221, %r2219, %r2220; shr.u32 %r2222, %r2219, 6; mul.wide.u32 %rd1245, %r2222, 8; add.s64 %rd1246, %rd2311, %rd1245; and.b32 %r2223, %r2221, 63; shl.b64 %rd1248, %rd1234, %r2223; ld.u64 %rd1249, [%rd1246+544]; and.b64 %rd1250, %rd1249, %rd1248; setp.eq.s64 %p192, %rd1250, 0; mul.wide.u32 %rd1251, %r2221, 8; add.s64 %rd1252, %rd2311, %rd1251; add.s64 %rd263, %rd1252, 1088; @%p192 bra $L__BB5_307; ld.u64 %rd1253, [%rd263]; add.s64 %rd2307, %rd2311, %rd1253; and.b32 %r2224, %r8, -8; mov.u32 %r2225, 1; st.local.v2.u32 [%rd2], {%r472, %r2224}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2226, %r471, 6; and.b32 %r2227, %r2226, 448; shl.b32 %r2228, %r8, 3; and.b32 %r2229, %r2228, 56; or.b32 %r2230, %r2229, %r2227; and.b32 %r2231, %r9, 7; or.b32 %r2232, %r2230, %r2231; ld.u8 %rs247, [%rd2307+15]; shr.u16 %rs248, %rs247, 5; mov.u32 %r2233, 5; cvt.u32.u16 %r2234, %rs248; sub.s32 %r2235, %r2233, %r2234; shr.u32 %r2236, %r2232, %r2235; mul.wide.u32 %rd1254, %r2236, 4; add.s64 %rd1255, %rd2307, %rd1254; mov.u32 %r2237, 32; shr.u32 %r2238, %r2237, %r2234; add.s32 %r2239, %r2238, -1; and.b32 %r2240, %r2239, %r2232; shl.b32 %r2241, %r2240, %r2234; ld.u32 %r2242, [%rd1255+96]; shr.u32 %r2243, %r2242, %r2241; shl.b32 %r2244, %r2225, %r2234; mov.u32 %r2245, -1; shl.b32 %r2246, %r2245, %r2244; not.b32 %r2247, %r2246; and.b32 %r2248, %r2243, %r2247; cvt.rn.f32.u32 %f571, %r2248; ld.v2.f32 {%f572, %f573}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f573, %f571, %f572; mov.b64 %rd1256, {%r472, %r2224}; shr.u64 %rd1257, %rd1256, 32; cvt.u32.u64 %r4782, %rd1257; mov.u32 %r4830, %r476; mov.u32 %r4870, %r473; mov.u32 %r4827, %r472; bra.uni $L__BB5_311; $L__BB5_310: ld.f32 %f937, [%rd5+4]; bra.uni $L__BB5_311; $L__BB5_336: shl.b64 %rd1304, %rd284, 5; add.s64 %rd2350, %rd17, %rd1304; $L__BB5_337: setp.eq.s64 %p211, %rd2350, 0; @%p211 bra $L__BB5_345; add.s64 %rd287, %rd2350, 8; ld.u64 %rd288, [%rd2350+8]; setp.eq.s64 %p212, %rd288, 0; @%p212 bra $L__BB5_344; add.s64 %rd2315, %rd1, %rd288; and.b32 %r2359, %r8, -4096; mov.u64 %rd1305, 1; st.local.v2.u32 [%rd2+24], {%r521, %r2359}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2360, %r471, 3; and.b32 %r2361, %r2360, 31744; and.b32 %r2362, %r8, 3968; shr.u32 %r2363, %r2362, 2; or.b32 %r2364, %r2363, %r2361; bfe.u32 %r2365, %r74, 7, 5; or.b32 %r2366, %r2365, %r2364; shr.u32 %r2367, %r2364, 6; mul.wide.u32 %rd1306, %r2367, 8; add.s64 %rd1307, %rd2315, %rd1306; and.b32 %r2368, %r2366, 63; shl.b64 %rd1308, %rd1305, %r2368; ld.u64 %rd1309, [%rd1307+4128]; and.b64 %rd1310, %rd1309, %rd1308; setp.eq.s64 %p213, %rd1310, 0; mul.wide.u32 %rd1311, %r2366, 8; add.s64 %rd1312, %rd2315, %rd1311; add.s64 %rd290, %rd1312, 8256; mov.b64 %rd1313, {%r521, %r2359}; shr.u64 %rd1314, %rd1313, 32; cvt.u32.u64 %r4875, %rd1314; @%p213 bra $L__BB5_343; ld.u64 %rd1315, [%rd290]; add.s64 %rd2311, %rd2315, %rd1315; st.local.u32 [%rd2+12], %r518; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2369, %r471, 5; and.b32 %r2370, %r2369, 3840; shl.b32 %r2371, %r8, 1; and.b32 %r2372, %r2371, 240; or.b32 %r2373, %r2372, %r2370; bfe.u32 %r2374, %r74, 3, 4; or.b32 %r2375, %r2374, %r2373; shr.u32 %r2376, %r2373, 6; mul.wide.u32 %rd1316, %r2376, 8; add.s64 %rd1317, %rd2311, %rd1316; and.b32 %r2377, %r2375, 63; shl.b64 %rd1319, %rd1305, %r2377; ld.u64 %rd1320, [%rd1317+544]; and.b64 %rd1321, %rd1320, %rd1319; setp.eq.s64 %p214, %rd1321, 0; mul.wide.u32 %rd1322, %r2375, 8; add.s64 %rd1323, %rd2311, %rd1322; add.s64 %rd292, %rd1323, 1088; @%p214 bra $L__BB5_342; ld.u64 %rd1324, [%rd292]; add.s64 %rd2307, %rd2311, %rd1324; and.b32 %r2378, %r8, -8; mov.u32 %r2379, 1; st.local.v2.u32 [%rd2], {%r472, %r2378}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2380, %r471, 6; and.b32 %r2381, %r2380, 448; shl.b32 %r2382, %r8, 3; and.b32 %r2383, %r2382, 56; or.b32 %r2384, %r2383, %r2381; and.b32 %r2385, %r74, 7; or.b32 %r2386, %r2384, %r2385; ld.u8 %rs261, [%rd2307+15]; shr.u16 %rs262, %rs261, 5; mov.u32 %r2387, 5; cvt.u32.u16 %r2388, %rs262; sub.s32 %r2389, %r2387, %r2388; shr.u32 %r2390, %r2386, %r2389; mul.wide.u32 %rd1325, %r2390, 4; add.s64 %rd1326, %rd2307, %rd1325; mov.u32 %r2391, 32; shr.u32 %r2392, %r2391, %r2388; add.s32 %r2393, %r2392, -1; and.b32 %r2394, %r2393, %r2386; shl.b32 %r2395, %r2394, %r2388; ld.u32 %r2396, [%rd1326+96]; shr.u32 %r2397, %r2396, %r2395; shl.b32 %r2398, %r2379, %r2388; mov.u32 %r2399, -1; shl.b32 %r2400, %r2399, %r2398; not.b32 %r2401, %r2400; and.b32 %r2402, %r2397, %r2401; cvt.rn.f32.u32 %f591, %r2402; ld.v2.f32 {%f592, %f593}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f593, %f591, %f592; mov.b64 %rd1327, {%r472, %r2378}; shr.u64 %rd1328, %rd1327, 32; cvt.u32.u64 %r4782, %rd1328; mov.u32 %r4830, %r521; mov.u32 %r4870, %r518; mov.u32 %r4827, %r472; bra.uni $L__BB5_346; $L__BB5_345: ld.f32 %f938, [%rd5+4]; bra.uni $L__BB5_346; $L__BB5_371: shl.b64 %rd1375, %rd313, 5; add.s64 %rd2356, %rd17, %rd1375; $L__BB5_372: setp.eq.s64 %p233, %rd2356, 0; @%p233 bra $L__BB5_380; add.s64 %rd316, %rd2356, 8; ld.u64 %rd317, [%rd2356+8]; setp.eq.s64 %p234, %rd317, 0; @%p234 bra $L__BB5_379; add.s64 %rd2315, %rd1, %rd317; and.b32 %r2514, %r139, -4096; mov.u64 %rd1376, 1; st.local.v2.u32 [%rd2+24], {%r563, %r2514}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2515, %r471, 3; and.b32 %r2516, %r2515, 31744; and.b32 %r2517, %r139, 3968; shr.u32 %r2518, %r2517, 2; or.b32 %r2519, %r2518, %r2516; bfe.u32 %r2520, %r74, 7, 5; or.b32 %r2521, %r2519, %r2520; shr.u32 %r2522, %r2519, 6; mul.wide.u32 %rd1377, %r2522, 8; add.s64 %rd1378, %rd2315, %rd1377; and.b32 %r2523, %r2521, 63; shl.b64 %rd1379, %rd1376, %r2523; ld.u64 %rd1380, [%rd1378+4128]; and.b64 %rd1381, %rd1380, %rd1379; setp.eq.s64 %p235, %rd1381, 0; mul.wide.u32 %rd1382, %r2521, 8; add.s64 %rd1383, %rd2315, %rd1382; add.s64 %rd319, %rd1383, 8256; mov.b64 %rd1384, {%r563, %r2514}; shr.u64 %rd1385, %rd1384, 32; cvt.u32.u64 %r4875, %rd1385; @%p235 bra $L__BB5_378; ld.u64 %rd1386, [%rd319]; add.s64 %rd2311, %rd2315, %rd1386; st.local.u32 [%rd2+12], %r560; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2524, %r471, 5; and.b32 %r2525, %r2524, 3840; shl.b32 %r2526, %r139, 1; and.b32 %r2527, %r2526, 240; or.b32 %r2528, %r2527, %r2525; bfe.u32 %r2529, %r74, 3, 4; or.b32 %r2530, %r2528, %r2529; shr.u32 %r2531, %r2528, 6; mul.wide.u32 %rd1387, %r2531, 8; add.s64 %rd1388, %rd2311, %rd1387; and.b32 %r2532, %r2530, 63; shl.b64 %rd1390, %rd1376, %r2532; ld.u64 %rd1391, [%rd1388+544]; and.b64 %rd1392, %rd1391, %rd1390; setp.eq.s64 %p236, %rd1392, 0; mul.wide.u32 %rd1393, %r2530, 8; add.s64 %rd1394, %rd2311, %rd1393; add.s64 %rd321, %rd1394, 1088; @%p236 bra $L__BB5_377; ld.u64 %rd1395, [%rd321]; add.s64 %rd2307, %rd2311, %rd1395; and.b32 %r2533, %r139, -8; mov.u32 %r2534, 1; st.local.v2.u32 [%rd2], {%r472, %r2533}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2535, %r471, 6; and.b32 %r2536, %r2535, 448; shl.b32 %r2537, %r139, 3; and.b32 %r2538, %r2537, 56; or.b32 %r2539, %r2538, %r2536; and.b32 %r2540, %r74, 7; or.b32 %r2541, %r2539, %r2540; ld.u8 %rs275, [%rd2307+15]; shr.u16 %rs276, %rs275, 5; mov.u32 %r2542, 5; cvt.u32.u16 %r2543, %rs276; sub.s32 %r2544, %r2542, %r2543; shr.u32 %r2545, %r2541, %r2544; mul.wide.u32 %rd1396, %r2545, 4; add.s64 %rd1397, %rd2307, %rd1396; mov.u32 %r2546, 32; shr.u32 %r2547, %r2546, %r2543; add.s32 %r2548, %r2547, -1; and.b32 %r2549, %r2548, %r2541; shl.b32 %r2550, %r2549, %r2543; ld.u32 %r2551, [%rd1397+96]; shr.u32 %r2552, %r2551, %r2550; shl.b32 %r2553, %r2534, %r2543; mov.u32 %r2554, -1; shl.b32 %r2555, %r2554, %r2553; not.b32 %r2556, %r2555; and.b32 %r2557, %r2552, %r2556; cvt.rn.f32.u32 %f612, %r2557; ld.v2.f32 {%f613, %f614}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f614, %f612, %f613; mov.b64 %rd1398, {%r472, %r2533}; shr.u64 %rd1399, %rd1398, 32; cvt.u32.u64 %r4782, %rd1399; mov.u32 %r4830, %r563; mov.u32 %r4870, %r560; mov.u32 %r4827, %r472; bra.uni $L__BB5_381; $L__BB5_380: ld.f32 %f939, [%rd5+4]; bra.uni $L__BB5_381; $L__BB5_406: shl.b64 %rd1446, %rd342, 5; add.s64 %rd2362, %rd17, %rd1446; $L__BB5_407: setp.eq.s64 %p255, %rd2362, 0; @%p255 bra $L__BB5_415; add.s64 %rd345, %rd2362, 8; ld.u64 %rd346, [%rd2362+8]; setp.eq.s64 %p256, %rd346, 0; @%p256 bra $L__BB5_414; add.s64 %rd2315, %rd1, %rd346; and.b32 %r2665, %r139, -4096; mov.u64 %rd1447, 1; st.local.v2.u32 [%rd2+24], {%r602, %r2665}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2666, %r471, 3; and.b32 %r2667, %r2666, 31744; and.b32 %r2668, %r139, 3968; shr.u32 %r2669, %r2668, 2; or.b32 %r2670, %r2669, %r2667; bfe.u32 %r2671, %r9, 7, 5; or.b32 %r2672, %r2670, %r2671; shr.u32 %r2673, %r2670, 6; mul.wide.u32 %rd1448, %r2673, 8; add.s64 %rd1449, %rd2315, %rd1448; and.b32 %r2674, %r2672, 63; shl.b64 %rd1450, %rd1447, %r2674; ld.u64 %rd1451, [%rd1449+4128]; and.b64 %rd1452, %rd1451, %rd1450; setp.eq.s64 %p257, %rd1452, 0; mul.wide.u32 %rd1453, %r2672, 8; add.s64 %rd1454, %rd2315, %rd1453; add.s64 %rd348, %rd1454, 8256; mov.b64 %rd1455, {%r602, %r2665}; shr.u64 %rd1456, %rd1455, 32; cvt.u32.u64 %r4875, %rd1456; @%p257 bra $L__BB5_413; ld.u64 %rd1457, [%rd348]; add.s64 %rd2311, %rd2315, %rd1457; st.local.u32 [%rd2+12], %r599; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2675, %r471, 5; and.b32 %r2676, %r2675, 3840; shl.b32 %r2677, %r139, 1; and.b32 %r2678, %r2677, 240; or.b32 %r2679, %r2678, %r2676; bfe.u32 %r2680, %r9, 3, 4; or.b32 %r2681, %r2679, %r2680; shr.u32 %r2682, %r2679, 6; mul.wide.u32 %rd1458, %r2682, 8; add.s64 %rd1459, %rd2311, %rd1458; and.b32 %r2683, %r2681, 63; shl.b64 %rd1461, %rd1447, %r2683; ld.u64 %rd1462, [%rd1459+544]; and.b64 %rd1463, %rd1462, %rd1461; setp.eq.s64 %p258, %rd1463, 0; mul.wide.u32 %rd1464, %r2681, 8; add.s64 %rd1465, %rd2311, %rd1464; add.s64 %rd350, %rd1465, 1088; @%p258 bra $L__BB5_412; ld.u64 %rd1466, [%rd350]; add.s64 %rd2307, %rd2311, %rd1466; and.b32 %r2684, %r139, -8; mov.u32 %r2685, 1; st.local.v2.u32 [%rd2], {%r472, %r2684}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2686, %r471, 7; shl.b32 %r2687, %r139, 3; and.b32 %r2688, %r2687, 56; and.b32 %r2689, %r9, 7; bfi.b32 %r2690, %r2686, %r2689, 6, 3; or.b32 %r2691, %r2690, %r2688; ld.u8 %rs289, [%rd2307+15]; shr.u16 %rs290, %rs289, 5; mov.u32 %r2692, 5; cvt.u32.u16 %r2693, %rs290; sub.s32 %r2694, %r2692, %r2693; shr.u32 %r2695, %r2691, %r2694; mul.wide.u32 %rd1467, %r2695, 4; add.s64 %rd1468, %rd2307, %rd1467; mov.u32 %r2696, 32; shr.u32 %r2697, %r2696, %r2693; add.s32 %r2698, %r2697, -1; and.b32 %r2699, %r2698, %r2691; shl.b32 %r2700, %r2699, %r2693; ld.u32 %r2701, [%rd1468+96]; shr.u32 %r2702, %r2701, %r2700; shl.b32 %r2703, %r2685, %r2693; mov.u32 %r2704, -1; shl.b32 %r2705, %r2704, %r2703; not.b32 %r2706, %r2705; and.b32 %r2707, %r2702, %r2706; cvt.rn.f32.u32 %f632, %r2707; ld.v2.f32 {%f633, %f634}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f634, %f632, %f633; mov.b64 %rd1469, {%r472, %r2684}; shr.u64 %rd1470, %rd1469, 32; cvt.u32.u64 %r4782, %rd1470; mov.u32 %r4830, %r602; mov.u32 %r4870, %r599; mov.u32 %r4827, %r472; bra.uni $L__BB5_416; $L__BB5_415: ld.f32 %f940, [%rd5+4]; bra.uni $L__BB5_416; $L__BB5_441: shl.b64 %rd1517, %rd371, 5; add.s64 %rd2368, %rd17, %rd1517; $L__BB5_442: setp.eq.s64 %p277, %rd2368, 0; @%p277 bra $L__BB5_450; add.s64 %rd374, %rd2368, 8; ld.u64 %rd375, [%rd2368+8]; setp.eq.s64 %p278, %rd375, 0; @%p278 bra $L__BB5_449; add.s64 %rd2315, %rd1, %rd375; and.b32 %r2815, %r139, -4096; mov.u64 %rd1518, 1; st.local.v2.u32 [%rd2+24], {%r640, %r2815}; and.b32 %r4908, %r9, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2816, %r635, 3; and.b32 %r2817, %r2816, 31744; and.b32 %r2818, %r139, 3968; shr.u32 %r2819, %r2818, 2; or.b32 %r2820, %r2819, %r2817; bfe.u32 %r2821, %r9, 7, 5; or.b32 %r2822, %r2820, %r2821; shr.u32 %r2823, %r2820, 6; mul.wide.u32 %rd1519, %r2823, 8; add.s64 %rd1520, %rd2315, %rd1519; and.b32 %r2824, %r2822, 63; shl.b64 %rd1521, %rd1518, %r2824; ld.u64 %rd1522, [%rd1520+4128]; and.b64 %rd1523, %rd1522, %rd1521; setp.eq.s64 %p279, %rd1523, 0; mul.wide.u32 %rd1524, %r2822, 8; add.s64 %rd1525, %rd2315, %rd1524; add.s64 %rd377, %rd1525, 8256; mov.b64 %rd1526, {%r640, %r2815}; shr.u64 %rd1527, %rd1526, 32; cvt.u32.u64 %r4875, %rd1527; @%p279 bra $L__BB5_448; ld.u64 %rd1528, [%rd377]; add.s64 %rd2311, %rd2315, %rd1528; st.local.u32 [%rd2+12], %r637; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r9, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2825, %r635, 5; and.b32 %r2826, %r2825, 3840; shl.b32 %r2827, %r139, 1; and.b32 %r2828, %r2827, 240; or.b32 %r2829, %r2828, %r2826; bfe.u32 %r2830, %r9, 3, 4; or.b32 %r2831, %r2829, %r2830; shr.u32 %r2832, %r2829, 6; mul.wide.u32 %rd1529, %r2832, 8; add.s64 %rd1530, %rd2311, %rd1529; and.b32 %r2833, %r2831, 63; shl.b64 %rd1532, %rd1518, %r2833; ld.u64 %rd1533, [%rd1530+544]; and.b64 %rd1534, %rd1533, %rd1532; setp.eq.s64 %p280, %rd1534, 0; mul.wide.u32 %rd1535, %r2831, 8; add.s64 %rd1536, %rd2311, %rd1535; add.s64 %rd379, %rd1536, 1088; @%p280 bra $L__BB5_447; ld.u64 %rd1537, [%rd379]; add.s64 %rd2307, %rd2311, %rd1537; and.b32 %r2834, %r139, -8; mov.u32 %r2835, 1; st.local.v2.u32 [%rd2], {%r636, %r2834}; and.b32 %r4825, %r9, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2836, %r635, 7; shl.b32 %r2837, %r139, 3; and.b32 %r2838, %r2837, 56; and.b32 %r2839, %r9, 7; bfi.b32 %r2840, %r2836, %r2839, 6, 3; or.b32 %r2841, %r2840, %r2838; ld.u8 %rs303, [%rd2307+15]; shr.u16 %rs304, %rs303, 5; mov.u32 %r2842, 5; cvt.u32.u16 %r2843, %rs304; sub.s32 %r2844, %r2842, %r2843; shr.u32 %r2845, %r2841, %r2844; mul.wide.u32 %rd1538, %r2845, 4; add.s64 %rd1539, %rd2307, %rd1538; mov.u32 %r2846, 32; shr.u32 %r2847, %r2846, %r2843; add.s32 %r2848, %r2847, -1; and.b32 %r2849, %r2848, %r2841; shl.b32 %r2850, %r2849, %r2843; ld.u32 %r2851, [%rd1539+96]; shr.u32 %r2852, %r2851, %r2850; shl.b32 %r2853, %r2835, %r2843; mov.u32 %r2854, -1; shl.b32 %r2855, %r2854, %r2853; not.b32 %r2856, %r2855; and.b32 %r2857, %r2852, %r2856; cvt.rn.f32.u32 %f655, %r2857; ld.v2.f32 {%f656, %f657}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f657, %f655, %f656; mov.b64 %rd1540, {%r636, %r2834}; shr.u64 %rd1541, %rd1540, 32; cvt.u32.u64 %r4782, %rd1541; mov.u32 %r4830, %r640; mov.u32 %r4870, %r637; mov.u32 %r4827, %r636; bra.uni $L__BB5_451; $L__BB5_450: ld.f32 %f941, [%rd5+4]; bra.uni $L__BB5_451; $L__BB5_476: shl.b64 %rd1588, %rd400, 5; add.s64 %rd2374, %rd17, %rd1588; $L__BB5_477: setp.eq.s64 %p299, %rd2374, 0; @%p299 bra $L__BB5_485; add.s64 %rd403, %rd2374, 8; ld.u64 %rd404, [%rd2374+8]; setp.eq.s64 %p300, %rd404, 0; @%p300 bra $L__BB5_484; add.s64 %rd2315, %rd1, %rd404; and.b32 %r2969, %r139, -4096; mov.u64 %rd1589, 1; st.local.v2.u32 [%rd2+24], {%r673, %r2969}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2970, %r635, 3; and.b32 %r2971, %r2970, 31744; and.b32 %r2972, %r139, 3968; shr.u32 %r2973, %r2972, 2; or.b32 %r2974, %r2973, %r2971; bfe.u32 %r2975, %r74, 7, 5; or.b32 %r2976, %r2974, %r2975; shr.u32 %r2977, %r2974, 6; mul.wide.u32 %rd1590, %r2977, 8; add.s64 %rd1591, %rd2315, %rd1590; and.b32 %r2978, %r2976, 63; shl.b64 %rd1592, %rd1589, %r2978; ld.u64 %rd1593, [%rd1591+4128]; and.b64 %rd1594, %rd1593, %rd1592; setp.eq.s64 %p301, %rd1594, 0; mul.wide.u32 %rd1595, %r2976, 8; add.s64 %rd1596, %rd2315, %rd1595; add.s64 %rd406, %rd1596, 8256; mov.b64 %rd1597, {%r673, %r2969}; shr.u64 %rd1598, %rd1597, 32; cvt.u32.u64 %r4875, %rd1598; @%p301 bra $L__BB5_483; ld.u64 %rd1599, [%rd406]; add.s64 %rd2311, %rd2315, %rd1599; st.local.u32 [%rd2+12], %r670; and.b32 %r4869, %r139, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2979, %r635, 5; and.b32 %r2980, %r2979, 3840; shl.b32 %r2981, %r139, 1; and.b32 %r2982, %r2981, 240; or.b32 %r2983, %r2982, %r2980; bfe.u32 %r2984, %r74, 3, 4; or.b32 %r2985, %r2983, %r2984; shr.u32 %r2986, %r2983, 6; mul.wide.u32 %rd1600, %r2986, 8; add.s64 %rd1601, %rd2311, %rd1600; and.b32 %r2987, %r2985, 63; shl.b64 %rd1603, %rd1589, %r2987; ld.u64 %rd1604, [%rd1601+544]; and.b64 %rd1605, %rd1604, %rd1603; setp.eq.s64 %p302, %rd1605, 0; mul.wide.u32 %rd1606, %r2985, 8; add.s64 %rd1607, %rd2311, %rd1606; add.s64 %rd408, %rd1607, 1088; @%p302 bra $L__BB5_482; ld.u64 %rd1608, [%rd408]; add.s64 %rd2307, %rd2311, %rd1608; and.b32 %r2988, %r139, -8; mov.u32 %r2989, 1; st.local.v2.u32 [%rd2], {%r636, %r2988}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2990, %r635, 6; and.b32 %r2991, %r2990, 448; shl.b32 %r2992, %r139, 3; and.b32 %r2993, %r2992, 56; or.b32 %r2994, %r2993, %r2991; and.b32 %r2995, %r74, 7; or.b32 %r2996, %r2994, %r2995; ld.u8 %rs317, [%rd2307+15]; shr.u16 %rs318, %rs317, 5; mov.u32 %r2997, 5; cvt.u32.u16 %r2998, %rs318; sub.s32 %r2999, %r2997, %r2998; shr.u32 %r3000, %r2996, %r2999; mul.wide.u32 %rd1609, %r3000, 4; add.s64 %rd1610, %rd2307, %rd1609; mov.u32 %r3001, 32; shr.u32 %r3002, %r3001, %r2998; add.s32 %r3003, %r3002, -1; and.b32 %r3004, %r3003, %r2996; shl.b32 %r3005, %r3004, %r2998; ld.u32 %r3006, [%rd1610+96]; shr.u32 %r3007, %r3006, %r3005; shl.b32 %r3008, %r2989, %r2998; mov.u32 %r3009, -1; shl.b32 %r3010, %r3009, %r3008; not.b32 %r3011, %r3010; and.b32 %r3012, %r3007, %r3011; cvt.rn.f32.u32 %f675, %r3012; ld.v2.f32 {%f676, %f677}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f677, %f675, %f676; mov.b64 %rd1611, {%r636, %r2988}; shr.u64 %rd1612, %rd1611, 32; cvt.u32.u64 %r4782, %rd1612; mov.u32 %r4830, %r673; mov.u32 %r4870, %r670; mov.u32 %r4827, %r636; bra.uni $L__BB5_486; $L__BB5_485: ld.f32 %f942, [%rd5+4]; bra.uni $L__BB5_486; $L__BB5_511: shl.b64 %rd1659, %rd429, 5; add.s64 %rd2380, %rd17, %rd1659; $L__BB5_512: setp.eq.s64 %p321, %rd2380, 0; @%p321 bra $L__BB5_520; add.s64 %rd432, %rd2380, 8; ld.u64 %rd433, [%rd2380+8]; setp.eq.s64 %p322, %rd433, 0; @%p322 bra $L__BB5_519; add.s64 %rd2315, %rd1, %rd433; and.b32 %r3123, %r8, -4096; mov.u64 %rd1660, 1; st.local.v2.u32 [%rd2+24], {%r703, %r3123}; and.b32 %r4908, %r74, -4096; st.local.u32 [%rd2+32], %r4908; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r3124, %r635, 3; and.b32 %r3125, %r3124, 31744; and.b32 %r3126, %r8, 3968; shr.u32 %r3127, %r3126, 2; or.b32 %r3128, %r3125, %r3127; bfe.u32 %r3129, %r74, 7, 5; or.b32 %r3130, %r3129, %r3128; shr.u32 %r3131, %r3128, 6; mul.wide.u32 %rd1661, %r3131, 8; add.s64 %rd1662, %rd2315, %rd1661; and.b32 %r3132, %r3130, 63; shl.b64 %rd1663, %rd1660, %r3132; ld.u64 %rd1664, [%rd1662+4128]; and.b64 %rd1665, %rd1664, %rd1663; setp.eq.s64 %p323, %rd1665, 0; mul.wide.u32 %rd1666, %r3130, 8; add.s64 %rd1667, %rd2315, %rd1666; add.s64 %rd435, %rd1667, 8256; mov.b64 %rd1668, {%r703, %r3123}; shr.u64 %rd1669, %rd1668, 32; cvt.u32.u64 %r4875, %rd1669; @%p323 bra $L__BB5_518; ld.u64 %rd1670, [%rd435]; add.s64 %rd2311, %rd2315, %rd1670; st.local.u32 [%rd2+12], %r700; and.b32 %r4869, %r8, -128; st.local.u32 [%rd2+16], %r4869; and.b32 %r4868, %r74, -128; st.local.u32 [%rd2+20], %r4868; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r3133, %r635, 5; and.b32 %r3134, %r3133, 3840; shl.b32 %r3135, %r8, 1; and.b32 %r3136, %r3135, 240; or.b32 %r3137, %r3134, %r3136; bfe.u32 %r3138, %r74, 3, 4; or.b32 %r3139, %r3138, %r3137; shr.u32 %r3140, %r3137, 6; mul.wide.u32 %rd1671, %r3140, 8; add.s64 %rd1672, %rd2311, %rd1671; and.b32 %r3141, %r3139, 63; shl.b64 %rd1674, %rd1660, %r3141; ld.u64 %rd1675, [%rd1672+544]; and.b64 %rd1676, %rd1675, %rd1674; setp.eq.s64 %p324, %rd1676, 0; mul.wide.u32 %rd1677, %r3139, 8; add.s64 %rd1678, %rd2311, %rd1677; add.s64 %rd437, %rd1678, 1088; @%p324 bra $L__BB5_517; ld.u64 %rd1679, [%rd437]; add.s64 %rd2307, %rd2311, %rd1679; and.b32 %r3142, %r8, -8; mov.u32 %r3143, 1; st.local.v2.u32 [%rd2], {%r636, %r3142}; and.b32 %r4825, %r74, -8; st.local.u32 [%rd2+8], %r4825; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r3144, %r635, 6; and.b32 %r3145, %r3144, 448; shl.b32 %r3146, %r8, 3; and.b32 %r3147, %r3146, 56; or.b32 %r3148, %r3145, %r3147; and.b32 %r3149, %r74, 7; or.b32 %r3150, %r3148, %r3149; ld.u8 %rs331, [%rd2307+15]; shr.u16 %rs332, %rs331, 5; mov.u32 %r3151, 5; cvt.u32.u16 %r3152, %rs332; sub.s32 %r3153, %r3151, %r3152; shr.u32 %r3154, %r3150, %r3153; mul.wide.u32 %rd1680, %r3154, 4; add.s64 %rd1681, %rd2307, %rd1680; mov.u32 %r3155, 32; shr.u32 %r3156, %r3155, %r3152; add.s32 %r3157, %r3156, -1; and.b32 %r3158, %r3157, %r3150; shl.b32 %r3159, %r3158, %r3152; ld.u32 %r3160, [%rd1681+96]; shr.u32 %r3161, %r3160, %r3159; shl.b32 %r3162, %r3143, %r3152; mov.u32 %r3163, -1; shl.b32 %r3164, %r3163, %r3162; not.b32 %r3165, %r3164; and.b32 %r3166, %r3161, %r3165; cvt.rn.f32.u32 %f696, %r3166; ld.v2.f32 {%f697, %f698}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f698, %f696, %f697; mov.b64 %rd1682, {%r636, %r3142}; shr.u64 %rd1683, %rd1682, 32; cvt.u32.u64 %r4782, %rd1683; mov.u32 %r4830, %r703; mov.u32 %r4870, %r700; mov.u32 %r4827, %r636; bra.uni $L__BB5_521; $L__BB5_520: ld.f32 %f943, [%rd5+4]; bra.uni $L__BB5_521; $L__BB5_546: shl.b64 %rd1726, %rd455, 5; add.s64 %rd2386, %rd17, %rd1726; $L__BB5_547: setp.eq.s64 %p343, %rd2386, 0; @%p343 bra $L__BB5_555; add.s64 %rd458, %rd2386, 8; ld.u64 %rd459, [%rd2386+8]; setp.eq.s64 %p344, %rd459, 0; @%p344 bra $L__BB5_554; add.s64 %rd460, %rd1, %rd459; and.b32 %r3280, %r8, -4096; mov.u64 %rd1727, 1; st.local.v2.u32 [%rd2+24], {%r728, %r3280}; and.b32 %r3281, %r9, -4096; st.local.u32 [%rd2+32], %r3281; st.local.u64 [%rd2+64], %rd460; shl.b32 %r3282, %r635, 3; and.b32 %r3283, %r3282, 31744; and.b32 %r3284, %r8, 3968; shr.u32 %r3285, %r3284, 2; or.b32 %r3286, %r3283, %r3285; bfe.u32 %r3287, %r9, 7, 5; or.b32 %r3288, %r3286, %r3287; shr.u32 %r3289, %r3286, 6; mul.wide.u32 %rd1728, %r3289, 8; add.s64 %rd1729, %rd460, %rd1728; and.b32 %r3290, %r3288, 63; shl.b64 %rd1730, %rd1727, %r3290; ld.u64 %rd1731, [%rd1729+4128]; and.b64 %rd1732, %rd1731, %rd1730; setp.eq.s64 %p345, %rd1732, 0; mul.wide.u32 %rd1733, %r3288, 8; add.s64 %rd1734, %rd460, %rd1733; add.s64 %rd461, %rd1734, 8256; @%p345 bra $L__BB5_553; ld.u64 %rd1735, [%rd461]; add.s64 %rd462, %rd460, %rd1735; st.local.u32 [%rd2+12], %r727; and.b32 %r3291, %r8, -128; st.local.u32 [%rd2+16], %r3291; and.b32 %r3292, %r9, -128; st.local.u32 [%rd2+20], %r3292; st.local.u64 [%rd2+56], %rd462; shl.b32 %r3293, %r635, 5; and.b32 %r3294, %r3293, 3840; shl.b32 %r3295, %r8, 1; and.b32 %r3296, %r3295, 240; or.b32 %r3297, %r3294, %r3296; bfe.u32 %r3298, %r9, 3, 4; or.b32 %r3299, %r3297, %r3298; shr.u32 %r3300, %r3297, 6; mul.wide.u32 %rd1736, %r3300, 8; add.s64 %rd1737, %rd462, %rd1736; and.b32 %r3301, %r3299, 63; shl.b64 %rd1739, %rd1727, %r3301; ld.u64 %rd1740, [%rd1737+544]; and.b64 %rd1741, %rd1740, %rd1739; setp.eq.s64 %p346, %rd1741, 0; mul.wide.u32 %rd1742, %r3299, 8; add.s64 %rd1743, %rd462, %rd1742; add.s64 %rd463, %rd1743, 1088; @%p346 bra $L__BB5_552; ld.u64 %rd1744, [%rd463]; add.s64 %rd1745, %rd462, %rd1744; and.b32 %r3302, %r8, -8; mov.u32 %r3303, 1; st.local.v2.u32 [%rd2], {%r636, %r3302}; and.b32 %r3304, %r9, -8; st.local.u32 [%rd2+8], %r3304; st.local.u64 [%rd2+48], %rd1745; shl.b32 %r3305, %r635, 6; and.b32 %r3306, %r3305, 448; shl.b32 %r3307, %r8, 3; and.b32 %r3308, %r3307, 56; or.b32 %r3309, %r3306, %r3308; and.b32 %r3310, %r9, 7; or.b32 %r3311, %r3309, %r3310; ld.u8 %rs345, [%rd1745+15]; shr.u16 %rs346, %rs345, 5; mov.u32 %r3312, 5; cvt.u32.u16 %r3313, %rs346; sub.s32 %r3314, %r3312, %r3313; shr.u32 %r3315, %r3311, %r3314; mul.wide.u32 %rd1746, %r3315, 4; add.s64 %rd1747, %rd1745, %rd1746; mov.u32 %r3316, 32; shr.u32 %r3317, %r3316, %r3313; add.s32 %r3318, %r3317, -1; and.b32 %r3319, %r3318, %r3311; shl.b32 %r3320, %r3319, %r3313; ld.u32 %r3321, [%rd1747+96]; shr.u32 %r3322, %r3321, %r3320; shl.b32 %r3323, %r3303, %r3313; mov.u32 %r3324, -1; shl.b32 %r3325, %r3324, %r3323; not.b32 %r3326, %r3325; and.b32 %r3327, %r3322, %r3326; cvt.rn.f32.u32 %f716, %r3327; ld.v2.f32 {%f717, %f718}, [%rd1745+80]; fma.rn.ftz.f32 %f944, %f718, %f716, %f717; bra.uni $L__BB5_556; $L__BB5_555: ld.f32 %f944, [%rd5+4]; bra.uni $L__BB5_556; $L__BB5_581: shl.b64 %rd1794, %rd473, 5; add.s64 %rd2387, %rd471, %rd1794; $L__BB5_582: setp.eq.s64 %p365, %rd2387, 0; @%p365 bra $L__BB5_590; add.s64 %rd476, %rd2387, 8; ld.u64 %rd477, [%rd2387+8]; setp.eq.s64 %p366, %rd477, 0; @%p366 bra $L__BB5_589; add.s64 %rd478, %rd470, %rd477; and.b32 %r3451, %r732, -4096; mov.u64 %rd1795, 1; st.local.v2.u32 [%rd2+24], {%r736, %r3451}; and.b32 %r3452, %r733, -4096; st.local.u32 [%rd2+32], %r3452; st.local.u64 [%rd2+64], %rd478; shl.b32 %r3453, %r471, 3; and.b32 %r3454, %r3453, 31744; and.b32 %r3455, %r732, 3968; shr.u32 %r3456, %r3455, 2; or.b32 %r3457, %r3456, %r3454; bfe.u32 %r3458, %r733, 7, 5; or.b32 %r3459, %r3457, %r3458; shr.u32 %r3460, %r3457, 6; mul.wide.u32 %rd1796, %r3460, 8; add.s64 %rd1797, %rd478, %rd1796; and.b32 %r3461, %r3459, 63; shl.b64 %rd1798, %rd1795, %r3461; ld.u64 %rd1799, [%rd1797+4128]; and.b64 %rd1800, %rd1799, %rd1798; setp.eq.s64 %p367, %rd1800, 0; mul.wide.u32 %rd1801, %r3459, 8; add.s64 %rd1802, %rd478, %rd1801; add.s64 %rd479, %rd1802, 8256; @%p367 bra $L__BB5_588; ld.u64 %rd1803, [%rd479]; add.s64 %rd480, %rd478, %rd1803; st.local.u32 [%rd2+12], %r735; and.b32 %r3462, %r732, -128; st.local.u32 [%rd2+16], %r3462; and.b32 %r3463, %r733, -128; st.local.u32 [%rd2+20], %r3463; st.local.u64 [%rd2+56], %rd480; shl.b32 %r3464, %r471, 5; and.b32 %r3465, %r3464, 3840; shl.b32 %r3466, %r732, 1; and.b32 %r3467, %r3466, 240; or.b32 %r3468, %r3467, %r3465; bfe.u32 %r3469, %r733, 3, 4; or.b32 %r3470, %r3468, %r3469; shr.u32 %r3471, %r3468, 6; mul.wide.u32 %rd1804, %r3471, 8; add.s64 %rd1805, %rd480, %rd1804; and.b32 %r3472, %r3470, 63; shl.b64 %rd1807, %rd1795, %r3472; ld.u64 %rd1808, [%rd1805+544]; and.b64 %rd1809, %rd1808, %rd1807; setp.eq.s64 %p368, %rd1809, 0; mul.wide.u32 %rd1810, %r3470, 8; add.s64 %rd1811, %rd480, %rd1810; add.s64 %rd481, %rd1811, 1088; @%p368 bra $L__BB5_587; ld.u64 %rd1812, [%rd481]; add.s64 %rd1813, %rd480, %rd1812; and.b32 %r3473, %r732, -8; mov.u32 %r3474, 1; st.local.v2.u32 [%rd2], {%r472, %r3473}; and.b32 %r3475, %r733, -8; st.local.u32 [%rd2+8], %r3475; st.local.u64 [%rd2+48], %rd1813; shl.b32 %r3476, %r471, 6; and.b32 %r3477, %r3476, 448; shl.b32 %r3478, %r732, 3; and.b32 %r3479, %r3478, 56; or.b32 %r3480, %r3479, %r3477; and.b32 %r3481, %r733, 7; or.b32 %r3482, %r3480, %r3481; ld.u8 %rs359, [%rd1813+15]; shr.u16 %rs360, %rs359, 5; mov.u32 %r3483, 5; cvt.u32.u16 %r3484, %rs360; sub.s32 %r3485, %r3483, %r3484; shr.u32 %r3486, %r3482, %r3485; mul.wide.u32 %rd1814, %r3486, 4; add.s64 %rd1815, %rd1813, %rd1814; mov.u32 %r3487, 32; shr.u32 %r3488, %r3487, %r3484; add.s32 %r3489, %r3488, -1; and.b32 %r3490, %r3489, %r3482; shl.b32 %r3491, %r3490, %r3484; ld.u32 %r3492, [%rd1815+96]; shr.u32 %r3493, %r3492, %r3491; shl.b32 %r3494, %r3474, %r3484; mov.u32 %r3495, -1; shl.b32 %r3496, %r3495, %r3494; not.b32 %r3497, %r3496; and.b32 %r3498, %r3493, %r3497; cvt.rn.f32.u32 %f757, %r3498; ld.v2.f32 {%f758, %f759}, [%rd1813+80]; fma.rn.ftz.f32 %f945, %f759, %f757, %f758; mov.u32 %r5196, %r472; bra.uni $L__BB5_591; $L__BB5_590: ld.f32 %f945, [%rd470+28]; bra.uni $L__BB5_591; $L__BB5_616: shl.b64 %rd1862, %rd491, 5; add.s64 %rd2388, %rd489, %rd1862; $L__BB5_617: setp.eq.s64 %p387, %rd2388, 0; @%p387 bra $L__BB5_625; add.s64 %rd494, %rd2388, 8; ld.u64 %rd495, [%rd2388+8]; setp.eq.s64 %p388, %rd495, 0; @%p388 bra $L__BB5_624; add.s64 %rd496, %rd488, %rd495; and.b32 %r3622, %r732, -4096; mov.u64 %rd1863, 1; st.local.v2.u32 [%rd2+24], {%r743, %r3622}; and.b32 %r3623, %r741, -4096; st.local.u32 [%rd2+32], %r3623; st.local.u64 [%rd2+64], %rd496; shl.b32 %r3624, %r471, 3; and.b32 %r3625, %r3624, 31744; and.b32 %r3626, %r732, 3968; shr.u32 %r3627, %r3626, 2; or.b32 %r3628, %r3627, %r3625; bfe.u32 %r3629, %r741, 7, 5; or.b32 %r3630, %r3629, %r3628; shr.u32 %r3631, %r3628, 6; mul.wide.u32 %rd1864, %r3631, 8; add.s64 %rd1865, %rd496, %rd1864; and.b32 %r3632, %r3630, 63; shl.b64 %rd1866, %rd1863, %r3632; ld.u64 %rd1867, [%rd1865+4128]; and.b64 %rd1868, %rd1867, %rd1866; setp.eq.s64 %p389, %rd1868, 0; mul.wide.u32 %rd1869, %r3630, 8; add.s64 %rd1870, %rd496, %rd1869; add.s64 %rd497, %rd1870, 8256; @%p389 bra $L__BB5_623; ld.u64 %rd1871, [%rd497]; add.s64 %rd498, %rd496, %rd1871; st.local.u32 [%rd2+12], %r742; and.b32 %r3633, %r732, -128; st.local.u32 [%rd2+16], %r3633; and.b32 %r3634, %r741, -128; st.local.u32 [%rd2+20], %r3634; st.local.u64 [%rd2+56], %rd498; shl.b32 %r3635, %r471, 5; and.b32 %r3636, %r3635, 3840; shl.b32 %r3637, %r732, 1; and.b32 %r3638, %r3637, 240; or.b32 %r3639, %r3638, %r3636; bfe.u32 %r3640, %r741, 3, 4; or.b32 %r3641, %r3640, %r3639; shr.u32 %r3642, %r3639, 6; mul.wide.u32 %rd1872, %r3642, 8; add.s64 %rd1873, %rd498, %rd1872; and.b32 %r3643, %r3641, 63; shl.b64 %rd1875, %rd1863, %r3643; ld.u64 %rd1876, [%rd1873+544]; and.b64 %rd1877, %rd1876, %rd1875; setp.eq.s64 %p390, %rd1877, 0; mul.wide.u32 %rd1878, %r3641, 8; add.s64 %rd1879, %rd498, %rd1878; add.s64 %rd499, %rd1879, 1088; @%p390 bra $L__BB5_622; ld.u64 %rd1880, [%rd499]; add.s64 %rd1881, %rd498, %rd1880; and.b32 %r3644, %r732, -8; mov.u32 %r3645, 1; st.local.v2.u32 [%rd2], {%r472, %r3644}; and.b32 %r3646, %r741, -8; st.local.u32 [%rd2+8], %r3646; st.local.u64 [%rd2+48], %rd1881; shl.b32 %r3647, %r471, 6; and.b32 %r3648, %r3647, 448; shl.b32 %r3649, %r732, 3; and.b32 %r3650, %r3649, 56; or.b32 %r3651, %r3650, %r3648; and.b32 %r3652, %r741, 7; or.b32 %r3653, %r3651, %r3652; ld.u8 %rs373, [%rd1881+15]; shr.u16 %rs374, %rs373, 5; mov.u32 %r3654, 5; cvt.u32.u16 %r3655, %rs374; sub.s32 %r3656, %r3654, %r3655; shr.u32 %r3657, %r3653, %r3656; mul.wide.u32 %rd1882, %r3657, 4; add.s64 %rd1883, %rd1881, %rd1882; mov.u32 %r3658, 32; shr.u32 %r3659, %r3658, %r3655; add.s32 %r3660, %r3659, -1; and.b32 %r3661, %r3660, %r3653; shl.b32 %r3662, %r3661, %r3655; ld.u32 %r3663, [%rd1883+96]; shr.u32 %r3664, %r3663, %r3662; shl.b32 %r3665, %r3645, %r3655; mov.u32 %r3666, -1; shl.b32 %r3667, %r3666, %r3665; not.b32 %r3668, %r3667; and.b32 %r3669, %r3664, %r3668; cvt.rn.f32.u32 %f777, %r3669; ld.v2.f32 {%f778, %f779}, [%rd1881+80]; fma.rn.ftz.f32 %f946, %f779, %f777, %f778; mov.u32 %r5196, %r472; bra.uni $L__BB5_626; $L__BB5_625: ld.f32 %f946, [%rd488+28]; bra.uni $L__BB5_626; $L__BB5_651: shl.b64 %rd1929, %rd509, 5; add.s64 %rd2389, %rd507, %rd1929; $L__BB5_652: setp.eq.s64 %p409, %rd2389, 0; @%p409 bra $L__BB5_660; add.s64 %rd512, %rd2389, 8; ld.u64 %rd513, [%rd2389+8]; setp.eq.s64 %p410, %rd513, 0; @%p410 bra $L__BB5_659; add.s64 %rd514, %rd506, %rd513; and.b32 %r3793, %r748, -4096; mov.u64 %rd1930, 1; st.local.v2.u32 [%rd2+24], {%r750, %r3793}; and.b32 %r3794, %r741, -4096; st.local.u32 [%rd2+32], %r3794; st.local.u64 [%rd2+64], %rd514; shl.b32 %r3795, %r471, 3; and.b32 %r3796, %r3795, 31744; and.b32 %r3797, %r748, 3968; shr.u32 %r3798, %r3797, 2; or.b32 %r3799, %r3798, %r3796; bfe.u32 %r3800, %r741, 7, 5; or.b32 %r3801, %r3799, %r3800; shr.u32 %r3802, %r3799, 6; mul.wide.u32 %rd1931, %r3802, 8; add.s64 %rd1932, %rd514, %rd1931; and.b32 %r3803, %r3801, 63; shl.b64 %rd1933, %rd1930, %r3803; ld.u64 %rd1934, [%rd1932+4128]; and.b64 %rd1935, %rd1934, %rd1933; setp.eq.s64 %p411, %rd1935, 0; mul.wide.u32 %rd1936, %r3801, 8; add.s64 %rd1937, %rd514, %rd1936; add.s64 %rd515, %rd1937, 8256; @%p411 bra $L__BB5_658; ld.u64 %rd1938, [%rd515]; add.s64 %rd516, %rd514, %rd1938; st.local.u32 [%rd2+12], %r749; and.b32 %r3804, %r748, -128; st.local.u32 [%rd2+16], %r3804; and.b32 %r3805, %r741, -128; st.local.u32 [%rd2+20], %r3805; st.local.u64 [%rd2+56], %rd516; shl.b32 %r3806, %r471, 5; and.b32 %r3807, %r3806, 3840; shl.b32 %r3808, %r748, 1; and.b32 %r3809, %r3808, 240; or.b32 %r3810, %r3809, %r3807; bfe.u32 %r3811, %r741, 3, 4; or.b32 %r3812, %r3810, %r3811; shr.u32 %r3813, %r3810, 6; mul.wide.u32 %rd1939, %r3813, 8; add.s64 %rd1940, %rd516, %rd1939; and.b32 %r3814, %r3812, 63; shl.b64 %rd1942, %rd1930, %r3814; ld.u64 %rd1943, [%rd1940+544]; and.b64 %rd1944, %rd1943, %rd1942; setp.eq.s64 %p412, %rd1944, 0; mul.wide.u32 %rd1945, %r3812, 8; add.s64 %rd1946, %rd516, %rd1945; add.s64 %rd517, %rd1946, 1088; @%p412 bra $L__BB5_657; ld.u64 %rd1947, [%rd517]; add.s64 %rd1948, %rd516, %rd1947; and.b32 %r3815, %r748, -8; mov.u32 %r3816, 1; st.local.v2.u32 [%rd2], {%r472, %r3815}; and.b32 %r3817, %r741, -8; st.local.u32 [%rd2+8], %r3817; st.local.u64 [%rd2+48], %rd1948; shl.b32 %r3818, %r471, 6; and.b32 %r3819, %r3818, 448; shl.b32 %r3820, %r748, 3; and.b32 %r3821, %r3820, 56; or.b32 %r3822, %r3821, %r3819; and.b32 %r3823, %r741, 7; or.b32 %r3824, %r3822, %r3823; ld.u8 %rs387, [%rd1948+15]; shr.u16 %rs388, %rs387, 5; mov.u32 %r3825, 5; cvt.u32.u16 %r3826, %rs388; sub.s32 %r3827, %r3825, %r3826; shr.u32 %r3828, %r3824, %r3827; mul.wide.u32 %rd1949, %r3828, 4; add.s64 %rd1950, %rd1948, %rd1949; mov.u32 %r3829, 32; shr.u32 %r3830, %r3829, %r3826; add.s32 %r3831, %r3830, -1; and.b32 %r3832, %r3831, %r3824; shl.b32 %r3833, %r3832, %r3826; ld.u32 %r3834, [%rd1950+96]; shr.u32 %r3835, %r3834, %r3833; shl.b32 %r3836, %r3816, %r3826; mov.u32 %r3837, -1; shl.b32 %r3838, %r3837, %r3836; not.b32 %r3839, %r3838; and.b32 %r3840, %r3835, %r3839; cvt.rn.f32.u32 %f798, %r3840; ld.v2.f32 {%f799, %f800}, [%rd1948+80]; fma.rn.ftz.f32 %f947, %f800, %f798, %f799; mov.u32 %r5196, %r472; bra.uni $L__BB5_661; $L__BB5_660: ld.f32 %f947, [%rd506+28]; bra.uni $L__BB5_661; $L__BB5_686: shl.b64 %rd1996, %rd527, 5; add.s64 %rd2390, %rd525, %rd1996; $L__BB5_687: setp.eq.s64 %p431, %rd2390, 0; @%p431 bra $L__BB5_695; add.s64 %rd530, %rd2390, 8; ld.u64 %rd531, [%rd2390+8]; setp.eq.s64 %p432, %rd531, 0; @%p432 bra $L__BB5_694; add.s64 %rd532, %rd524, %rd531; and.b32 %r3961, %r748, -4096; mov.u64 %rd1997, 1; st.local.v2.u32 [%rd2+24], {%r756, %r3961}; and.b32 %r3962, %r733, -4096; st.local.u32 [%rd2+32], %r3962; st.local.u64 [%rd2+64], %rd532; shl.b32 %r3963, %r471, 3; and.b32 %r3964, %r3963, 31744; and.b32 %r3965, %r748, 3968; shr.u32 %r3966, %r3965, 2; or.b32 %r3967, %r3966, %r3964; bfe.u32 %r3968, %r733, 7, 5; or.b32 %r3969, %r3967, %r3968; shr.u32 %r3970, %r3967, 6; mul.wide.u32 %rd1998, %r3970, 8; add.s64 %rd1999, %rd532, %rd1998; and.b32 %r3971, %r3969, 63; shl.b64 %rd2000, %rd1997, %r3971; ld.u64 %rd2001, [%rd1999+4128]; and.b64 %rd2002, %rd2001, %rd2000; setp.eq.s64 %p433, %rd2002, 0; mul.wide.u32 %rd2003, %r3969, 8; add.s64 %rd2004, %rd532, %rd2003; add.s64 %rd533, %rd2004, 8256; @%p433 bra $L__BB5_693; ld.u64 %rd2005, [%rd533]; add.s64 %rd534, %rd532, %rd2005; st.local.u32 [%rd2+12], %r755; and.b32 %r3972, %r748, -128; st.local.u32 [%rd2+16], %r3972; and.b32 %r3973, %r733, -128; st.local.u32 [%rd2+20], %r3973; st.local.u64 [%rd2+56], %rd534; shl.b32 %r3974, %r471, 5; and.b32 %r3975, %r3974, 3840; shl.b32 %r3976, %r748, 1; and.b32 %r3977, %r3976, 240; or.b32 %r3978, %r3977, %r3975; bfe.u32 %r3979, %r733, 3, 4; or.b32 %r3980, %r3978, %r3979; shr.u32 %r3981, %r3978, 6; mul.wide.u32 %rd2006, %r3981, 8; add.s64 %rd2007, %rd534, %rd2006; and.b32 %r3982, %r3980, 63; shl.b64 %rd2009, %rd1997, %r3982; ld.u64 %rd2010, [%rd2007+544]; and.b64 %rd2011, %rd2010, %rd2009; setp.eq.s64 %p434, %rd2011, 0; mul.wide.u32 %rd2012, %r3980, 8; add.s64 %rd2013, %rd534, %rd2012; add.s64 %rd535, %rd2013, 1088; @%p434 bra $L__BB5_692; ld.u64 %rd2014, [%rd535]; add.s64 %rd2015, %rd534, %rd2014; and.b32 %r3983, %r748, -8; mov.u32 %r3984, 1; st.local.v2.u32 [%rd2], {%r472, %r3983}; and.b32 %r3985, %r733, -8; st.local.u32 [%rd2+8], %r3985; st.local.u64 [%rd2+48], %rd2015; and.b32 %r3986, %r471, 7; shl.b32 %r3987, %r748, 3; and.b32 %r3988, %r3987, 56; and.b32 %r3989, %r733, 7; bfi.b32 %r3990, %r3986, %r3989, 6, 3; or.b32 %r3991, %r3990, %r3988; ld.u8 %rs401, [%rd2015+15]; shr.u16 %rs402, %rs401, 5; mov.u32 %r3992, 5; cvt.u32.u16 %r3993, %rs402; sub.s32 %r3994, %r3992, %r3993; shr.u32 %r3995, %r3991, %r3994; mul.wide.u32 %rd2016, %r3995, 4; add.s64 %rd2017, %rd2015, %rd2016; mov.u32 %r3996, 32; shr.u32 %r3997, %r3996, %r3993; add.s32 %r3998, %r3997, -1; and.b32 %r3999, %r3998, %r3991; shl.b32 %r4000, %r3999, %r3993; ld.u32 %r4001, [%rd2017+96]; shr.u32 %r4002, %r4001, %r4000; shl.b32 %r4003, %r3984, %r3993; mov.u32 %r4004, -1; shl.b32 %r4005, %r4004, %r4003; not.b32 %r4006, %r4005; and.b32 %r4007, %r4002, %r4006; cvt.rn.f32.u32 %f818, %r4007; ld.v2.f32 {%f819, %f820}, [%rd2015+80]; fma.rn.ftz.f32 %f948, %f820, %f818, %f819; mov.u32 %r5196, %r472; bra.uni $L__BB5_696; $L__BB5_695: ld.f32 %f948, [%rd524+28]; bra.uni $L__BB5_696; $L__BB5_721: shl.b64 %rd2063, %rd545, 5; add.s64 %rd2391, %rd543, %rd2063; $L__BB5_722: setp.eq.s64 %p453, %rd2391, 0; @%p453 bra $L__BB5_730; add.s64 %rd548, %rd2391, 8; ld.u64 %rd549, [%rd2391+8]; setp.eq.s64 %p454, %rd549, 0; @%p454 bra $L__BB5_729; add.s64 %rd550, %rd542, %rd549; and.b32 %r4128, %r748, -4096; mov.u64 %rd2064, 1; st.local.v2.u32 [%rd2+24], {%r762, %r4128}; and.b32 %r4129, %r733, -4096; st.local.u32 [%rd2+32], %r4129; st.local.u64 [%rd2+64], %rd550; shl.b32 %r4130, %r635, 3; and.b32 %r4131, %r4130, 31744; and.b32 %r4132, %r748, 3968; shr.u32 %r4133, %r4132, 2; or.b32 %r4134, %r4133, %r4131; bfe.u32 %r4135, %r733, 7, 5; or.b32 %r4136, %r4134, %r4135; shr.u32 %r4137, %r4134, 6; mul.wide.u32 %rd2065, %r4137, 8; add.s64 %rd2066, %rd550, %rd2065; and.b32 %r4138, %r4136, 63; shl.b64 %rd2067, %rd2064, %r4138; ld.u64 %rd2068, [%rd2066+4128]; and.b64 %rd2069, %rd2068, %rd2067; setp.eq.s64 %p455, %rd2069, 0; mul.wide.u32 %rd2070, %r4136, 8; add.s64 %rd2071, %rd550, %rd2070; add.s64 %rd551, %rd2071, 8256; @%p455 bra $L__BB5_728; ld.u64 %rd2072, [%rd551]; add.s64 %rd552, %rd550, %rd2072; st.local.u32 [%rd2+12], %r761; and.b32 %r4139, %r748, -128; st.local.u32 [%rd2+16], %r4139; and.b32 %r4140, %r733, -128; st.local.u32 [%rd2+20], %r4140; st.local.u64 [%rd2+56], %rd552; shl.b32 %r4141, %r635, 5; and.b32 %r4142, %r4141, 3840; shl.b32 %r4143, %r748, 1; and.b32 %r4144, %r4143, 240; or.b32 %r4145, %r4144, %r4142; bfe.u32 %r4146, %r733, 3, 4; or.b32 %r4147, %r4145, %r4146; shr.u32 %r4148, %r4145, 6; mul.wide.u32 %rd2073, %r4148, 8; add.s64 %rd2074, %rd552, %rd2073; and.b32 %r4149, %r4147, 63; shl.b64 %rd2076, %rd2064, %r4149; ld.u64 %rd2077, [%rd2074+544]; and.b64 %rd2078, %rd2077, %rd2076; setp.eq.s64 %p456, %rd2078, 0; mul.wide.u32 %rd2079, %r4147, 8; add.s64 %rd2080, %rd552, %rd2079; add.s64 %rd553, %rd2080, 1088; @%p456 bra $L__BB5_727; ld.u64 %rd2081, [%rd553]; add.s64 %rd2082, %rd552, %rd2081; and.b32 %r4150, %r748, -8; mov.u32 %r4151, 1; st.local.v2.u32 [%rd2], {%r636, %r4150}; and.b32 %r4152, %r733, -8; st.local.u32 [%rd2+8], %r4152; st.local.u64 [%rd2+48], %rd2082; and.b32 %r4153, %r635, 7; shl.b32 %r4154, %r748, 3; and.b32 %r4155, %r4154, 56; and.b32 %r4156, %r733, 7; bfi.b32 %r4157, %r4153, %r4156, 6, 3; or.b32 %r4158, %r4157, %r4155; ld.u8 %rs415, [%rd2082+15]; shr.u16 %rs416, %rs415, 5; mov.u32 %r4159, 5; cvt.u32.u16 %r4160, %rs416; sub.s32 %r4161, %r4159, %r4160; shr.u32 %r4162, %r4158, %r4161; mul.wide.u32 %rd2083, %r4162, 4; add.s64 %rd2084, %rd2082, %rd2083; mov.u32 %r4163, 32; shr.u32 %r4164, %r4163, %r4160; add.s32 %r4165, %r4164, -1; and.b32 %r4166, %r4165, %r4158; shl.b32 %r4167, %r4166, %r4160; ld.u32 %r4168, [%rd2084+96]; shr.u32 %r4169, %r4168, %r4167; shl.b32 %r4170, %r4151, %r4160; mov.u32 %r4171, -1; shl.b32 %r4172, %r4171, %r4170; not.b32 %r4173, %r4172; and.b32 %r4174, %r4169, %r4173; cvt.rn.f32.u32 %f841, %r4174; ld.v2.f32 {%f842, %f843}, [%rd2082+80]; fma.rn.ftz.f32 %f949, %f843, %f841, %f842; mov.u32 %r5196, %r636; bra.uni $L__BB5_731; $L__BB5_730: ld.f32 %f949, [%rd542+28]; bra.uni $L__BB5_731; $L__BB5_756: shl.b64 %rd2130, %rd563, 5; add.s64 %rd2392, %rd561, %rd2130; $L__BB5_757: setp.eq.s64 %p475, %rd2392, 0; @%p475 bra $L__BB5_765; add.s64 %rd566, %rd2392, 8; ld.u64 %rd567, [%rd2392+8]; setp.eq.s64 %p476, %rd567, 0; @%p476 bra $L__BB5_764; add.s64 %rd568, %rd560, %rd567; and.b32 %r4298, %r748, -4096; mov.u64 %rd2131, 1; st.local.v2.u32 [%rd2+24], {%r768, %r4298}; and.b32 %r4299, %r741, -4096; st.local.u32 [%rd2+32], %r4299; st.local.u64 [%rd2+64], %rd568; shl.b32 %r4300, %r635, 3; and.b32 %r4301, %r4300, 31744; and.b32 %r4302, %r748, 3968; shr.u32 %r4303, %r4302, 2; or.b32 %r4304, %r4303, %r4301; bfe.u32 %r4305, %r741, 7, 5; or.b32 %r4306, %r4304, %r4305; shr.u32 %r4307, %r4304, 6; mul.wide.u32 %rd2132, %r4307, 8; add.s64 %rd2133, %rd568, %rd2132; and.b32 %r4308, %r4306, 63; shl.b64 %rd2134, %rd2131, %r4308; ld.u64 %rd2135, [%rd2133+4128]; and.b64 %rd2136, %rd2135, %rd2134; setp.eq.s64 %p477, %rd2136, 0; mul.wide.u32 %rd2137, %r4306, 8; add.s64 %rd2138, %rd568, %rd2137; add.s64 %rd569, %rd2138, 8256; @%p477 bra $L__BB5_763; ld.u64 %rd2139, [%rd569]; add.s64 %rd570, %rd568, %rd2139; st.local.u32 [%rd2+12], %r767; and.b32 %r4309, %r748, -128; st.local.u32 [%rd2+16], %r4309; and.b32 %r4310, %r741, -128; st.local.u32 [%rd2+20], %r4310; st.local.u64 [%rd2+56], %rd570; shl.b32 %r4311, %r635, 5; and.b32 %r4312, %r4311, 3840; shl.b32 %r4313, %r748, 1; and.b32 %r4314, %r4313, 240; or.b32 %r4315, %r4314, %r4312; bfe.u32 %r4316, %r741, 3, 4; or.b32 %r4317, %r4315, %r4316; shr.u32 %r4318, %r4315, 6; mul.wide.u32 %rd2140, %r4318, 8; add.s64 %rd2141, %rd570, %rd2140; and.b32 %r4319, %r4317, 63; shl.b64 %rd2143, %rd2131, %r4319; ld.u64 %rd2144, [%rd2141+544]; and.b64 %rd2145, %rd2144, %rd2143; setp.eq.s64 %p478, %rd2145, 0; mul.wide.u32 %rd2146, %r4317, 8; add.s64 %rd2147, %rd570, %rd2146; add.s64 %rd571, %rd2147, 1088; @%p478 bra $L__BB5_762; ld.u64 %rd2148, [%rd571]; add.s64 %rd2149, %rd570, %rd2148; and.b32 %r4320, %r748, -8; mov.u32 %r4321, 1; st.local.v2.u32 [%rd2], {%r636, %r4320}; and.b32 %r4322, %r741, -8; st.local.u32 [%rd2+8], %r4322; st.local.u64 [%rd2+48], %rd2149; shl.b32 %r4323, %r635, 6; and.b32 %r4324, %r4323, 448; shl.b32 %r4325, %r748, 3; and.b32 %r4326, %r4325, 56; or.b32 %r4327, %r4326, %r4324; and.b32 %r4328, %r741, 7; or.b32 %r4329, %r4327, %r4328; ld.u8 %rs429, [%rd2149+15]; shr.u16 %rs430, %rs429, 5; mov.u32 %r4330, 5; cvt.u32.u16 %r4331, %rs430; sub.s32 %r4332, %r4330, %r4331; shr.u32 %r4333, %r4329, %r4332; mul.wide.u32 %rd2150, %r4333, 4; add.s64 %rd2151, %rd2149, %rd2150; mov.u32 %r4334, 32; shr.u32 %r4335, %r4334, %r4331; add.s32 %r4336, %r4335, -1; and.b32 %r4337, %r4336, %r4329; shl.b32 %r4338, %r4337, %r4331; ld.u32 %r4339, [%rd2151+96]; shr.u32 %r4340, %r4339, %r4338; shl.b32 %r4341, %r4321, %r4331; mov.u32 %r4342, -1; shl.b32 %r4343, %r4342, %r4341; not.b32 %r4344, %r4343; and.b32 %r4345, %r4340, %r4344; cvt.rn.f32.u32 %f861, %r4345; ld.v2.f32 {%f862, %f863}, [%rd2149+80]; fma.rn.ftz.f32 %f950, %f863, %f861, %f862; mov.u32 %r5196, %r636; bra.uni $L__BB5_766; $L__BB5_765: ld.f32 %f950, [%rd560+28]; bra.uni $L__BB5_766; $L__BB5_791: shl.b64 %rd2198, %rd581, 5; add.s64 %rd2393, %rd579, %rd2198; $L__BB5_792: setp.eq.s64 %p497, %rd2393, 0; @%p497 bra $L__BB5_800; add.s64 %rd584, %rd2393, 8; ld.u64 %rd585, [%rd2393+8]; setp.eq.s64 %p498, %rd585, 0; @%p498 bra $L__BB5_799; add.s64 %rd586, %rd578, %rd585; and.b32 %r4469, %r732, -4096; mov.u64 %rd2199, 1; st.local.v2.u32 [%rd2+24], {%r774, %r4469}; and.b32 %r4470, %r741, -4096; st.local.u32 [%rd2+32], %r4470; st.local.u64 [%rd2+64], %rd586; shl.b32 %r4471, %r635, 3; and.b32 %r4472, %r4471, 31744; and.b32 %r4473, %r732, 3968; shr.u32 %r4474, %r4473, 2; or.b32 %r4475, %r4472, %r4474; bfe.u32 %r4476, %r741, 7, 5; or.b32 %r4477, %r4476, %r4475; shr.u32 %r4478, %r4475, 6; mul.wide.u32 %rd2200, %r4478, 8; add.s64 %rd2201, %rd586, %rd2200; and.b32 %r4479, %r4477, 63; shl.b64 %rd2202, %rd2199, %r4479; ld.u64 %rd2203, [%rd2201+4128]; and.b64 %rd2204, %rd2203, %rd2202; setp.eq.s64 %p499, %rd2204, 0; mul.wide.u32 %rd2205, %r4477, 8; add.s64 %rd2206, %rd586, %rd2205; add.s64 %rd587, %rd2206, 8256; @%p499 bra $L__BB5_798; ld.u64 %rd2207, [%rd587]; add.s64 %rd588, %rd586, %rd2207; st.local.u32 [%rd2+12], %r773; and.b32 %r4480, %r732, -128; st.local.u32 [%rd2+16], %r4480; and.b32 %r4481, %r741, -128; st.local.u32 [%rd2+20], %r4481; st.local.u64 [%rd2+56], %rd588; shl.b32 %r4482, %r635, 5; and.b32 %r4483, %r4482, 3840; shl.b32 %r4484, %r732, 1; and.b32 %r4485, %r4484, 240; or.b32 %r4486, %r4483, %r4485; bfe.u32 %r4487, %r741, 3, 4; or.b32 %r4488, %r4487, %r4486; shr.u32 %r4489, %r4486, 6; mul.wide.u32 %rd2208, %r4489, 8; add.s64 %rd2209, %rd588, %rd2208; and.b32 %r4490, %r4488, 63; shl.b64 %rd2211, %rd2199, %r4490; ld.u64 %rd2212, [%rd2209+544]; and.b64 %rd2213, %rd2212, %rd2211; setp.eq.s64 %p500, %rd2213, 0; mul.wide.u32 %rd2214, %r4488, 8; add.s64 %rd2215, %rd588, %rd2214; add.s64 %rd589, %rd2215, 1088; @%p500 bra $L__BB5_797; ld.u64 %rd2216, [%rd589]; add.s64 %rd2217, %rd588, %rd2216; and.b32 %r4491, %r732, -8; mov.u32 %r4492, 1; st.local.v2.u32 [%rd2], {%r636, %r4491}; and.b32 %r4493, %r741, -8; st.local.u32 [%rd2+8], %r4493; st.local.u64 [%rd2+48], %rd2217; shl.b32 %r4494, %r635, 6; and.b32 %r4495, %r4494, 448; shl.b32 %r4496, %r732, 3; and.b32 %r4497, %r4496, 56; or.b32 %r4498, %r4495, %r4497; and.b32 %r4499, %r741, 7; or.b32 %r4500, %r4498, %r4499; ld.u8 %rs443, [%rd2217+15]; shr.u16 %rs444, %rs443, 5; mov.u32 %r4501, 5; cvt.u32.u16 %r4502, %rs444; sub.s32 %r4503, %r4501, %r4502; shr.u32 %r4504, %r4500, %r4503; mul.wide.u32 %rd2218, %r4504, 4; add.s64 %rd2219, %rd2217, %rd2218; mov.u32 %r4505, 32; shr.u32 %r4506, %r4505, %r4502; add.s32 %r4507, %r4506, -1; and.b32 %r4508, %r4507, %r4500; shl.b32 %r4509, %r4508, %r4502; ld.u32 %r4510, [%rd2219+96]; shr.u32 %r4511, %r4510, %r4509; shl.b32 %r4512, %r4492, %r4502; mov.u32 %r4513, -1; shl.b32 %r4514, %r4513, %r4512; not.b32 %r4515, %r4514; and.b32 %r4516, %r4511, %r4515; cvt.rn.f32.u32 %f882, %r4516; ld.v2.f32 {%f883, %f884}, [%rd2217+80]; fma.rn.ftz.f32 %f951, %f884, %f882, %f883; mov.u32 %r5196, %r636; bra.uni $L__BB5_801; $L__BB5_800: ld.f32 %f951, [%rd578+28]; bra.uni $L__BB5_801; $L__BB5_64: ld.f32 %f930, [%rd55+12]; bra.uni $L__BB5_66; $L__BB5_99: ld.f32 %f931, [%rd84+12]; bra.uni $L__BB5_101; $L__BB5_134: ld.f32 %f932, [%rd113+12]; bra.uni $L__BB5_136; $L__BB5_169: ld.f32 %f933, [%rd142+12]; bra.uni $L__BB5_171; $L__BB5_204: ld.f32 %f934, [%rd171+12]; bra.uni $L__BB5_206; $L__BB5_239: ld.f32 %f935, [%rd200+12]; bra.uni $L__BB5_241; $L__BB5_274: ld.f32 %f936, [%rd229+12]; bra.uni $L__BB5_276; $L__BB5_309: ld.f32 %f937, [%rd258+12]; bra.uni $L__BB5_311; $L__BB5_344: ld.f32 %f938, [%rd287+12]; bra.uni $L__BB5_346; $L__BB5_379: ld.f32 %f939, [%rd316+12]; bra.uni $L__BB5_381; $L__BB5_414: ld.f32 %f940, [%rd345+12]; bra.uni $L__BB5_416; $L__BB5_449: ld.f32 %f941, [%rd374+12]; bra.uni $L__BB5_451; $L__BB5_484: ld.f32 %f942, [%rd403+12]; bra.uni $L__BB5_486; $L__BB5_519: ld.f32 %f943, [%rd432+12]; bra.uni $L__BB5_521; $L__BB5_554: ld.f32 %f944, [%rd458+12]; bra.uni $L__BB5_556; $L__BB5_589: ld.f32 %f945, [%rd476+12]; bra.uni $L__BB5_591; $L__BB5_624: ld.f32 %f946, [%rd494+12]; bra.uni $L__BB5_626; $L__BB5_659: ld.f32 %f947, [%rd512+12]; bra.uni $L__BB5_661; $L__BB5_694: ld.f32 %f948, [%rd530+12]; bra.uni $L__BB5_696; $L__BB5_729: ld.f32 %f949, [%rd548+12]; bra.uni $L__BB5_731; $L__BB5_764: ld.f32 %f950, [%rd566+12]; bra.uni $L__BB5_766; $L__BB5_799: ld.f32 %f951, [%rd584+12]; bra.uni $L__BB5_801; $L__BB5_63: ld.f32 %f930, [%rd58]; mov.u32 %r4830, %r12; bra.uni $L__BB5_66; $L__BB5_98: ld.f32 %f931, [%rd87]; mov.u32 %r4830, %r12; bra.uni $L__BB5_101; $L__BB5_133: ld.f32 %f932, [%rd116]; mov.u32 %r4830, %r12; bra.uni $L__BB5_136; $L__BB5_168: and.b32 %r4830, %r259, -4096; ld.f32 %f933, [%rd145]; bra.uni $L__BB5_171; $L__BB5_203: and.b32 %r4830, %r259, -4096; ld.f32 %f934, [%rd174]; bra.uni $L__BB5_206; $L__BB5_238: and.b32 %r4830, %r259, -4096; ld.f32 %f935, [%rd203]; bra.uni $L__BB5_241; $L__BB5_273: ld.f32 %f936, [%rd232]; mov.u32 %r4830, %r426; bra.uni $L__BB5_276; $L__BB5_308: ld.f32 %f937, [%rd261]; mov.u32 %r4830, %r476; bra.uni $L__BB5_311; $L__BB5_343: ld.f32 %f938, [%rd290]; mov.u32 %r4830, %r521; bra.uni $L__BB5_346; $L__BB5_378: ld.f32 %f939, [%rd319]; mov.u32 %r4830, %r563; bra.uni $L__BB5_381; $L__BB5_413: ld.f32 %f940, [%rd348]; mov.u32 %r4830, %r602; bra.uni $L__BB5_416; $L__BB5_448: ld.f32 %f941, [%rd377]; mov.u32 %r4830, %r640; bra.uni $L__BB5_451; $L__BB5_483: ld.f32 %f942, [%rd406]; mov.u32 %r4830, %r673; bra.uni $L__BB5_486; $L__BB5_518: ld.f32 %f943, [%rd435]; mov.u32 %r4830, %r703; bra.uni $L__BB5_521; $L__BB5_553: ld.f32 %f944, [%rd461]; bra.uni $L__BB5_556; $L__BB5_588: ld.f32 %f945, [%rd479]; bra.uni $L__BB5_591; $L__BB5_623: ld.f32 %f946, [%rd497]; bra.uni $L__BB5_626; $L__BB5_658: ld.f32 %f947, [%rd515]; bra.uni $L__BB5_661; $L__BB5_693: ld.f32 %f948, [%rd533]; bra.uni $L__BB5_696; $L__BB5_728: ld.f32 %f949, [%rd551]; bra.uni $L__BB5_731; $L__BB5_763: ld.f32 %f950, [%rd569]; bra.uni $L__BB5_766; $L__BB5_798: ld.f32 %f951, [%rd587]; bra.uni $L__BB5_801; $L__BB5_62: ld.f32 %f930, [%rd60]; mov.u32 %r4830, %r12; bra.uni $L__BB5_66; $L__BB5_97: ld.f32 %f931, [%rd89]; mov.u32 %r4830, %r12; bra.uni $L__BB5_101; $L__BB5_132: ld.f32 %f932, [%rd118]; mov.u32 %r4830, %r12; bra.uni $L__BB5_136; $L__BB5_167: and.b32 %r4830, %r259, -4096; and.b32 %r4870, %r259, -128; ld.f32 %f933, [%rd147]; bra.uni $L__BB5_171; $L__BB5_202: and.b32 %r4830, %r259, -4096; and.b32 %r4870, %r259, -128; ld.f32 %f934, [%rd176]; bra.uni $L__BB5_206; $L__BB5_237: ld.f32 %f935, [%rd205]; bra.uni $L__BB5_241; $L__BB5_272: ld.f32 %f936, [%rd234]; mov.u32 %r4830, %r426; mov.u32 %r4870, %r423; bra.uni $L__BB5_276; $L__BB5_307: ld.f32 %f937, [%rd263]; mov.u32 %r4830, %r476; mov.u32 %r4870, %r473; bra.uni $L__BB5_311; $L__BB5_342: ld.f32 %f938, [%rd292]; mov.u32 %r4830, %r521; mov.u32 %r4870, %r518; bra.uni $L__BB5_346; $L__BB5_377: ld.f32 %f939, [%rd321]; mov.u32 %r4830, %r563; mov.u32 %r4870, %r560; bra.uni $L__BB5_381; $L__BB5_412: ld.f32 %f940, [%rd350]; mov.u32 %r4830, %r602; mov.u32 %r4870, %r599; bra.uni $L__BB5_416; $L__BB5_447: ld.f32 %f941, [%rd379]; mov.u32 %r4830, %r640; mov.u32 %r4870, %r637; bra.uni $L__BB5_451; $L__BB5_482: ld.f32 %f942, [%rd408]; mov.u32 %r4830, %r673; mov.u32 %r4870, %r670; bra.uni $L__BB5_486; $L__BB5_517: ld.f32 %f943, [%rd437]; mov.u32 %r4830, %r703; mov.u32 %r4870, %r700; bra.uni $L__BB5_521; $L__BB5_552: ld.f32 %f944, [%rd463]; bra.uni $L__BB5_556; $L__BB5_587: ld.f32 %f945, [%rd481]; bra.uni $L__BB5_591; $L__BB5_622: ld.f32 %f946, [%rd499]; bra.uni $L__BB5_626; $L__BB5_657: ld.f32 %f947, [%rd517]; bra.uni $L__BB5_661; $L__BB5_692: ld.f32 %f948, [%rd535]; bra.uni $L__BB5_696; $L__BB5_727: ld.f32 %f949, [%rd553]; bra.uni $L__BB5_731; $L__BB5_762: ld.f32 %f950, [%rd571]; bra.uni $L__BB5_766; $L__BB5_797: ld.f32 %f951, [%rd589]; bra.uni $L__BB5_801; } .func (.param .b32 func_retval0) _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj( .param .b64 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3, .param .b32 _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4 ) { .local .align 8 .b8 __local_depot6[88]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<520>; .reg .b16 %rs<425>; .reg .f32 %f<954>; .reg .b32 %r<3610>; .reg .b64 %rd<2395>; mov.u64 %SPL, __local_depot6; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd608, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_0]; ld.param.f32 %f322, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_1]; ld.param.f32 %f323, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_2]; ld.param.f32 %f324, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; ld.param.u32 %r784, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_4]; cvta.to.global.u64 %rd609, %rd608; ld.global.u64 %rd610, [%rd609]; mov.u64 %rd2298, 0; ld.u64 %rd612, [%rd610+696]; add.s64 %rd613, %rd610, %rd612; add.s64 %rd1, %rd613, 672; add.u64 %rd614, %SP, 0; add.u64 %rd2, %SPL, 0; mov.u32 %r3184, 2147483647; st.local.v2.u32 [%rd2], {%r3184, %r3184}; st.local.v2.u32 [%rd2+8], {%r3184, %r3184}; st.local.v2.u32 [%rd2+16], {%r3184, %r3184}; st.local.v2.u32 [%rd2+24], {%r3184, %r3184}; st.local.u32 [%rd2+32], %r3184; st.local.u64 [%rd2+40], %rd1; st.local.u64 [%rd2+48], %rd2298; st.local.u64 [%rd2+56], %rd2298; st.local.u64 [%rd2+64], %rd2298; add.s64 %rd5, %rd613, 696; setp.eq.s32 %p1, %r784, 0; @%p1 bra $L__BB6_16; setp.ne.s32 %p2, %r784, 1; @%p2 bra $L__BB6_17; cvt.rni.ftz.f32.f32 %f325, %f322; cvt.rzi.ftz.s32.f32 %r1, %f325; cvt.rni.ftz.f32.f32 %f326, %f323; cvt.rzi.ftz.s32.f32 %r2, %f326; cvt.rni.ftz.f32.f32 %f327, %f324; cvt.rzi.ftz.s32.f32 %r3, %f327; ld.u32 %r4, [%rd5]; mov.u64 %rd2297, 0; setp.eq.s32 %p3, %r4, 0; @%p3 bra $L__BB6_7; add.s64 %rd6, %rd1, 64; shr.u32 %r787, %r1, 12; cvt.u64.u32 %rd616, %r787; shl.b64 %rd617, %rd616, 42; shr.u32 %r788, %r2, 12; mul.wide.u32 %rd618, %r788, 2097152; or.b64 %rd619, %rd618, %rd617; shr.u32 %r789, %r3, 12; cvt.u64.u32 %rd620, %r789; or.b64 %rd7, %rd619, %rd620; mov.u32 %r3170, 0; $L__BB6_4: cvt.u64.u32 %rd8, %r3170; mul.wide.u32 %rd621, %r3170, 32; add.s64 %rd622, %rd6, %rd621; ld.u64 %rd623, [%rd622]; setp.eq.s64 %p4, %rd623, %rd7; add.s32 %r3170, %r3170, 1; @%p4 bra $L__BB6_6; setp.lt.u32 %p5, %r3170, %r4; @%p5 bra $L__BB6_4; bra.uni $L__BB6_7; $L__BB6_16: add.u64 %rd647, %SP, 72; add.u64 %rd648, %SPL, 72; st.local.u64 [%rd648], %rd614; add.ftz.f32 %f333, %f322, 0fBF000000; add.ftz.f32 %f334, %f323, 0fBF000000; add.ftz.f32 %f335, %f324, 0fBF000000; { // callseq 19, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd647; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f333; st.param.f32 [param1+4], %f334; st.param.f32 [param1+8], %f335; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f953, [retval0+0]; } // callseq 19 bra.uni $L__BB6_834; $L__BB6_17: add.u64 %rd652, %SPL, 80; st.local.u64 [%rd652], %rd614; add.ftz.f32 %f336, %f322, 0fBF000000; cvt.rmi.ftz.f32.f32 %f337, %f336; add.ftz.f32 %f338, %f323, 0fBF000000; cvt.rmi.ftz.f32.f32 %f339, %f338; add.ftz.f32 %f340, %f324, 0fBF000000; cvt.rmi.ftz.f32.f32 %f341, %f340; sub.ftz.f32 %f342, %f336, %f337; sub.ftz.f32 %f343, %f338, %f339; sub.ftz.f32 %f344, %f340, %f341; mov.f32 %f345, 0f40400000; sub.ftz.f32 %f346, %f345, %f342; fma.rn.ftz.f32 %f347, %f342, %f346, 0fC0400000; fma.rn.ftz.f32 %f348, %f342, %f347, 0f3F800000; mul.ftz.f32 %f349, %f342, %f342; mul.ftz.f32 %f350, %f342, 0f40400000; add.ftz.f32 %f351, %f350, 0fC0C00000; fma.rn.ftz.f32 %f352, %f349, %f351, 0f40800000; mul.ftz.f32 %f353, %f352, 0f3E2AAAAB; fma.rn.ftz.f32 %f7, %f348, 0f3E2AAAAB, %f353; sub.ftz.f32 %f354, %f345, %f350; fma.rn.ftz.f32 %f355, %f342, %f354, 0f40400000; fma.rn.ftz.f32 %f356, %f342, %f355, 0f3F800000; mul.ftz.f32 %f357, %f342, %f349; mul.ftz.f32 %f358, %f357, 0f3E2AAAAB; fma.rn.ftz.f32 %f8, %f356, 0f3E2AAAAB, %f358; sub.ftz.f32 %f359, %f345, %f343; fma.rn.ftz.f32 %f360, %f343, %f359, 0fC0400000; fma.rn.ftz.f32 %f361, %f343, %f360, 0f3F800000; mul.ftz.f32 %f362, %f343, %f343; mul.ftz.f32 %f363, %f343, 0f40400000; add.ftz.f32 %f364, %f363, 0fC0C00000; fma.rn.ftz.f32 %f365, %f362, %f364, 0f40800000; mul.ftz.f32 %f366, %f365, 0f3E2AAAAB; fma.rn.ftz.f32 %f9, %f361, 0f3E2AAAAB, %f366; sub.ftz.f32 %f367, %f345, %f363; fma.rn.ftz.f32 %f368, %f343, %f367, 0f40400000; fma.rn.ftz.f32 %f369, %f343, %f368, 0f3F800000; mul.ftz.f32 %f370, %f343, %f362; mul.ftz.f32 %f371, %f370, 0f3E2AAAAB; fma.rn.ftz.f32 %f10, %f369, 0f3E2AAAAB, %f371; sub.ftz.f32 %f372, %f345, %f344; fma.rn.ftz.f32 %f373, %f344, %f372, 0fC0400000; fma.rn.ftz.f32 %f374, %f344, %f373, 0f3F800000; mul.ftz.f32 %f375, %f344, %f344; mul.ftz.f32 %f376, %f344, 0f40400000; add.ftz.f32 %f377, %f376, 0fC0C00000; fma.rn.ftz.f32 %f378, %f375, %f377, 0f40800000; mul.ftz.f32 %f379, %f378, 0f3E2AAAAB; fma.rn.ftz.f32 %f11, %f374, 0f3E2AAAAB, %f379; sub.ftz.f32 %f380, %f345, %f376; fma.rn.ftz.f32 %f381, %f344, %f380, 0f40400000; fma.rn.ftz.f32 %f382, %f344, %f381, 0f3F800000; mul.ftz.f32 %f383, %f344, %f375; mul.ftz.f32 %f384, %f383, 0f3E2AAAAB; fma.rn.ftz.f32 %f12, %f382, 0f3E2AAAAB, %f384; div.approx.ftz.f32 %f385, %f353, %f7; add.ftz.f32 %f386, %f385, 0fBF800000; add.ftz.f32 %f13, %f337, %f386; div.approx.ftz.f32 %f387, %f358, %f8; add.ftz.f32 %f388, %f387, 0f3F800000; add.ftz.f32 %f14, %f337, %f388; div.approx.ftz.f32 %f389, %f366, %f9; add.ftz.f32 %f390, %f389, 0fBF800000; add.ftz.f32 %f15, %f339, %f390; div.approx.ftz.f32 %f391, %f371, %f10; add.ftz.f32 %f392, %f391, 0f3F800000; add.ftz.f32 %f16, %f339, %f392; div.approx.ftz.f32 %f393, %f379, %f11; add.ftz.f32 %f394, %f393, 0fBF800000; add.ftz.f32 %f17, %f341, %f394; div.approx.ftz.f32 %f395, %f384, %f12; add.ftz.f32 %f396, %f395, 0f3F800000; add.ftz.f32 %f18, %f341, %f396; cvt.rmi.ftz.f32.f32 %f19, %f13; cvt.rmi.ftz.f32.f32 %f397, %f15; cvt.rmi.ftz.f32.f32 %f398, %f17; sub.ftz.f32 %f20, %f15, %f397; sub.ftz.f32 %f21, %f17, %f398; cvt.rzi.ftz.s32.f32 %r7, %f19; cvt.rzi.ftz.s32.f32 %r8, %f397; cvt.rzi.ftz.s32.f32 %r9, %f398; and.b32 %r10, %r7, -8; and.b32 %r3272, %r7, -128; and.b32 %r12, %r7, -4096; add.s64 %rd17, %rd1, 64; shr.u32 %r824, %r9, 12; cvt.u64.u32 %rd18, %r824; shr.u32 %r825, %r8, 12; mul.wide.u32 %rd19, %r825, 2097152; shr.u32 %r826, %r7, 12; cvt.u64.u32 %rd654, %r826; bfi.b64 %rd21, %rd654, %rd18, 42, 32; ld.u32 %r13, [%rd5]; setp.eq.s32 %p10, %r13, 0; @%p10 bra $L__BB6_22; or.b64 %rd22, %rd21, %rd19; mov.u32 %r3171, 0; $L__BB6_19: cvt.u64.u32 %rd23, %r3171; mul.wide.u32 %rd655, %r3171, 32; add.s64 %rd656, %rd17, %rd655; ld.u64 %rd657, [%rd656]; setp.eq.s64 %p11, %rd657, %rd22; add.s32 %r3171, %r3171, 1; @%p11 bra $L__BB6_21; setp.lt.u32 %p12, %r3171, %r13; @%p12 bra $L__BB6_19; bra.uni $L__BB6_22; $L__BB6_6: shl.b64 %rd625, %rd8, 5; add.s64 %rd2297, %rd6, %rd625; $L__BB6_7: setp.eq.s64 %p6, %rd2297, 0; @%p6 bra $L__BB6_15; add.s64 %rd11, %rd2297, 8; ld.u64 %rd12, [%rd2297+8]; setp.eq.s64 %p7, %rd12, 0; @%p7 bra $L__BB6_14; add.s64 %rd13, %rd1, %rd12; and.b32 %r790, %r2, -4096; and.b32 %r791, %r1, -4096; mov.u64 %rd626, 1; st.local.v2.u32 [%rd2+24], {%r791, %r790}; and.b32 %r792, %r3, -4096; st.local.u32 [%rd2+32], %r792; st.local.u64 [%rd2+64], %rd13; shl.b32 %r793, %r1, 3; and.b32 %r794, %r793, 31744; and.b32 %r795, %r2, 3968; shr.u32 %r796, %r795, 2; or.b32 %r797, %r796, %r794; bfe.u32 %r798, %r3, 7, 5; or.b32 %r799, %r798, %r797; shr.u32 %r800, %r797, 6; mul.wide.u32 %rd627, %r800, 8; add.s64 %rd628, %rd13, %rd627; and.b32 %r801, %r799, 63; shl.b64 %rd629, %rd626, %r801; ld.u64 %rd630, [%rd628+4128]; and.b64 %rd631, %rd630, %rd629; setp.eq.s64 %p8, %rd631, 0; mul.wide.u32 %rd632, %r799, 8; add.s64 %rd633, %rd13, %rd632; add.s64 %rd14, %rd633, 8256; @%p8 bra $L__BB6_13; ld.u64 %rd634, [%rd14]; add.s64 %rd15, %rd13, %rd634; and.b32 %r802, %r1, -128; st.local.u32 [%rd2+12], %r802; and.b32 %r803, %r2, -128; st.local.u32 [%rd2+16], %r803; and.b32 %r804, %r3, -128; st.local.u32 [%rd2+20], %r804; st.local.u64 [%rd2+56], %rd15; shl.b32 %r805, %r1, 5; and.b32 %r806, %r805, 3840; shl.b32 %r807, %r2, 1; and.b32 %r808, %r807, 240; or.b32 %r809, %r808, %r806; bfe.u32 %r810, %r3, 3, 4; or.b32 %r811, %r810, %r809; shr.u32 %r812, %r809, 6; mul.wide.u32 %rd635, %r812, 8; add.s64 %rd636, %rd15, %rd635; and.b32 %r813, %r811, 63; shl.b64 %rd638, %rd626, %r813; ld.u64 %rd639, [%rd636+544]; and.b64 %rd640, %rd639, %rd638; setp.eq.s64 %p9, %rd640, 0; mul.wide.u32 %rd641, %r811, 8; add.s64 %rd642, %rd15, %rd641; add.s64 %rd16, %rd642, 1088; @%p9 bra $L__BB6_12; ld.u64 %rd643, [%rd16]; add.s64 %rd644, %rd15, %rd643; and.b32 %r814, %r2, -8; and.b32 %r815, %r1, -8; st.local.v2.u32 [%rd2], {%r815, %r814}; and.b32 %r816, %r3, -8; st.local.u32 [%rd2+8], %r816; st.local.u64 [%rd2+48], %rd644; shl.b32 %r817, %r1, 6; and.b32 %r818, %r817, 448; shl.b32 %r819, %r2, 3; and.b32 %r820, %r819, 56; or.b32 %r821, %r820, %r818; and.b32 %r822, %r3, 7; or.b32 %r823, %r821, %r822; mul.wide.u32 %rd645, %r823, 2; add.s64 %rd646, %rd644, %rd645; ld.u16 %rs133, [%rd646+96]; cvt.rn.f32.u16 %f328, %rs133; ld.v2.f32 {%f329, %f330}, [%rd644+80]; fma.rn.ftz.f32 %f953, %f330, %f328, %f329; bra.uni $L__BB6_834; $L__BB6_15: ld.f32 %f953, [%rd5+4]; bra.uni $L__BB6_834; $L__BB6_21: shl.b64 %rd659, %rd23, 5; add.s64 %rd2298, %rd17, %rd659; $L__BB6_22: setp.eq.s64 %p13, %rd2298, 0; @%p13 bra $L__BB6_30; add.s64 %rd26, %rd2298, 8; ld.u64 %rd27, [%rd2298+8]; setp.eq.s64 %p14, %rd27, 0; @%p14 bra $L__BB6_29; add.s64 %rd2315, %rd1, %rd27; and.b32 %r828, %r8, -4096; mov.u64 %rd660, 1; st.local.v2.u32 [%rd2+24], {%r12, %r828}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r829, %r7, 3; and.b32 %r830, %r829, 31744; and.b32 %r831, %r8, 3968; shr.u32 %r832, %r831, 2; or.b32 %r833, %r832, %r830; bfe.u32 %r834, %r9, 7, 5; or.b32 %r835, %r833, %r834; shr.u32 %r836, %r833, 6; mul.wide.u32 %rd661, %r836, 8; add.s64 %rd662, %rd2315, %rd661; and.b32 %r837, %r835, 63; shl.b64 %rd663, %rd660, %r837; ld.u64 %rd664, [%rd662+4128]; and.b64 %rd665, %rd664, %rd663; setp.eq.s64 %p15, %rd665, 0; mul.wide.u32 %rd666, %r835, 8; add.s64 %rd667, %rd2315, %rd666; add.s64 %rd29, %rd667, 8256; mov.b64 %rd668, {%r12, %r828}; shr.u64 %rd669, %rd668, 32; cvt.u32.u64 %r3277, %rd669; @%p15 bra $L__BB6_28; ld.u64 %rd670, [%rd29]; add.s64 %rd2311, %rd2315, %rd670; st.local.u32 [%rd2+12], %r3272; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r838, %r7, 5; and.b32 %r839, %r838, 3840; shl.b32 %r840, %r8, 1; and.b32 %r841, %r840, 240; or.b32 %r842, %r841, %r839; bfe.u32 %r843, %r9, 3, 4; or.b32 %r844, %r842, %r843; shr.u32 %r845, %r842, 6; mul.wide.u32 %rd671, %r845, 8; add.s64 %rd672, %rd2311, %rd671; and.b32 %r846, %r844, 63; shl.b64 %rd674, %rd660, %r846; ld.u64 %rd675, [%rd672+544]; and.b64 %rd676, %rd675, %rd674; setp.eq.s64 %p16, %rd676, 0; mul.wide.u32 %rd677, %r844, 8; add.s64 %rd678, %rd2311, %rd677; add.s64 %rd31, %rd678, 1088; @%p16 bra $L__BB6_27; ld.u64 %rd679, [%rd31]; add.s64 %rd2307, %rd2311, %rd679; and.b32 %r847, %r8, -8; st.local.v2.u32 [%rd2], {%r10, %r847}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r848, %r7, 6; and.b32 %r849, %r848, 448; shl.b32 %r850, %r8, 3; and.b32 %r851, %r850, 56; or.b32 %r852, %r851, %r849; and.b32 %r853, %r9, 7; or.b32 %r854, %r852, %r853; mul.wide.u32 %rd680, %r854, 2; add.s64 %rd681, %rd2307, %rd680; ld.u16 %rs134, [%rd681+96]; cvt.rn.f32.u16 %f399, %rs134; ld.v2.f32 {%f400, %f401}, [%rd2307+80]; fma.rn.ftz.f32 %f929, %f401, %f399, %f400; mov.b64 %rd682, {%r10, %r847}; shr.u64 %rd683, %rd682, 32; cvt.u32.u64 %r3184, %rd683; mov.u32 %r3232, %r12; mov.u32 %r3229, %r10; bra.uni $L__BB6_31; $L__BB6_30: ld.f32 %f929, [%rd5+4]; mov.u32 %r3277, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r3232, %r3277; mov.u32 %r3184, %r3277; mov.u32 %r3310, %r3277; mov.u32 %r3270, %r3277; mov.u32 %r3271, %r3277; mov.u32 %r3272, %r3277; mov.u32 %r3227, %r3277; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r3229, %r3277; bra.uni $L__BB6_31; $L__BB6_14: ld.f32 %f953, [%rd11+12]; bra.uni $L__BB6_834; $L__BB6_29: ld.f32 %f929, [%rd26+12]; mov.u32 %r3277, 2147483647; mov.u64 %rd2315, 0; mov.u32 %r3232, %r3277; mov.u32 %r3184, %r3277; mov.u32 %r3310, %r3277; mov.u32 %r3270, %r3277; mov.u32 %r3271, %r3277; mov.u32 %r3272, %r3277; mov.u32 %r3227, %r3277; mov.u64 %rd2311, %rd2315; mov.u64 %rd2307, %rd2315; mov.u32 %r3229, %r3277; bra.uni $L__BB6_31; $L__BB6_13: ld.f32 %f953, [%rd14]; bra.uni $L__BB6_834; $L__BB6_28: ld.f32 %f929, [%rd29]; mov.u32 %r3184, 2147483647; mov.u64 %rd2311, 0; mov.u32 %r3232, %r12; mov.u32 %r3270, %r3184; mov.u32 %r3271, %r3184; mov.u32 %r3272, %r3184; mov.u32 %r3227, %r3184; mov.u64 %rd2307, %rd2311; mov.u32 %r3229, %r3184; bra.uni $L__BB6_31; $L__BB6_12: ld.f32 %f953, [%rd16]; bra.uni $L__BB6_834; $L__BB6_27: ld.f32 %f929, [%rd31]; mov.u64 %rd2307, 0; mov.u32 %r3232, %r12; mov.u32 %r3227, %r3184; mov.u32 %r3229, %r3184; $L__BB6_31: add.s32 %r74, %r9, 1; setp.ne.s32 %p17, %r10, %r3229; mov.u16 %rs359, 0; @%p17 bra $L__BB6_34; and.b32 %r994, %r8, -8; setp.ne.s32 %p18, %r994, %r3184; @%p18 bra $L__BB6_34; and.b32 %r995, %r74, -8; setp.eq.s32 %p19, %r995, %r3227; selp.u16 %rs359, 1, 0, %p19; $L__BB6_34: setp.eq.s16 %p20, %rs359, 0; @%p20 bra $L__BB6_36; shl.b32 %r996, %r7, 6; and.b32 %r997, %r996, 448; shl.b32 %r998, %r8, 3; and.b32 %r999, %r998, 56; or.b32 %r1000, %r999, %r997; and.b32 %r1001, %r74, 7; or.b32 %r1002, %r1000, %r1001; mul.wide.u32 %rd698, %r1002, 2; add.s64 %rd699, %rd2307, %rd698; ld.u16 %rs137, [%rd699+96]; cvt.rn.f32.u16 %f404, %rs137; ld.v2.f32 {%f405, %f406}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f406, %f404, %f405; $L__BB6_66: and.b32 %r3124, %r7, -8; sub.ftz.f32 %f424, %f930, %f929; fma.rn.ftz.f32 %f40, %f21, %f424, %f929; add.s32 %r139, %r8, 1; setp.ne.s32 %p39, %r3124, %r3229; mov.u16 %rs362, 0; @%p39 bra $L__BB6_69; and.b32 %r1079, %r139, -8; setp.ne.s32 %p40, %r1079, %r3184; @%p40 bra $L__BB6_69; and.b32 %r1080, %r74, -8; setp.eq.s32 %p41, %r1080, %r3227; selp.u16 %rs362, 1, 0, %p41; $L__BB6_69: setp.eq.s16 %p42, %rs362, 0; @%p42 bra $L__BB6_71; shl.b32 %r1081, %r7, 6; and.b32 %r1082, %r1081, 448; shl.b32 %r1083, %r139, 3; and.b32 %r1084, %r1083, 56; or.b32 %r1085, %r1084, %r1082; and.b32 %r1086, %r74, 7; or.b32 %r1087, %r1085, %r1086; mul.wide.u32 %rd767, %r1087, 2; add.s64 %rd768, %rd2307, %rd767; ld.u16 %rs147, [%rd768+96]; cvt.rn.f32.u16 %f425, %rs147; ld.v2.f32 {%f426, %f427}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f427, %f425, %f426; $L__BB6_101: and.b32 %r3094, %r7, -8; setp.ne.s32 %p61, %r3094, %r3229; mov.u16 %rs365, 0; @%p61 bra $L__BB6_104; and.b32 %r1165, %r139, -8; setp.ne.s32 %p62, %r1165, %r3184; @%p62 bra $L__BB6_104; and.b32 %r1166, %r9, -8; setp.eq.s32 %p63, %r1166, %r3227; selp.u16 %rs365, 1, 0, %p63; $L__BB6_104: setp.eq.s16 %p64, %rs365, 0; @%p64 bra $L__BB6_106; and.b32 %r1167, %r7, 7; shl.b32 %r1168, %r139, 3; and.b32 %r1169, %r1168, 56; and.b32 %r1170, %r9, 7; bfi.b32 %r1171, %r1167, %r1170, 6, 3; or.b32 %r1172, %r1171, %r1169; mul.wide.u32 %rd837, %r1172, 2; add.s64 %rd838, %rd2307, %rd837; ld.u16 %rs157, [%rd838+96]; cvt.rn.f32.u16 %f445, %rs157; ld.v2.f32 {%f446, %f447}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f447, %f445, %f446; $L__BB6_136: sub.ftz.f32 %f465, %f931, %f932; fma.rn.ftz.f32 %f466, %f21, %f465, %f932; sub.ftz.f32 %f467, %f466, %f40; fma.rn.ftz.f32 %f65, %f20, %f467, %f40; add.s32 %r259, %r7, 1; and.b32 %r260, %r259, -8; setp.ne.s32 %p83, %r260, %r3229; mov.u16 %rs368, 0; @%p83 bra $L__BB6_139; and.b32 %r1246, %r139, -8; setp.ne.s32 %p84, %r1246, %r3184; @%p84 bra $L__BB6_139; and.b32 %r1247, %r9, -8; setp.eq.s32 %p85, %r1247, %r3227; selp.u16 %rs368, 1, 0, %p85; $L__BB6_139: setp.eq.s16 %p86, %rs368, 0; @%p86 bra $L__BB6_141; and.b32 %r1248, %r259, 7; shl.b32 %r1249, %r139, 3; and.b32 %r1250, %r1249, 56; and.b32 %r1251, %r9, 7; bfi.b32 %r1252, %r1248, %r1251, 6, 3; or.b32 %r1253, %r1252, %r1250; mul.wide.u32 %rd905, %r1253, 2; add.s64 %rd906, %rd2307, %rd905; ld.u16 %rs167, [%rd906+96]; cvt.rn.f32.u16 %f468, %rs167; ld.v2.f32 {%f469, %f470}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f470, %f468, %f469; $L__BB6_171: and.b32 %r3165, %r259, -8; setp.ne.s32 %p105, %r3165, %r3229; mov.u16 %rs371, 0; @%p105 bra $L__BB6_174; and.b32 %r1328, %r139, -8; setp.ne.s32 %p106, %r1328, %r3184; @%p106 bra $L__BB6_174; and.b32 %r1329, %r74, -8; setp.eq.s32 %p107, %r1329, %r3227; selp.u16 %rs371, 1, 0, %p107; $L__BB6_174: setp.eq.s16 %p108, %rs371, 0; @%p108 bra $L__BB6_176; shl.b32 %r1330, %r259, 6; and.b32 %r1331, %r1330, 448; shl.b32 %r1332, %r139, 3; and.b32 %r1333, %r1332, 56; or.b32 %r1334, %r1333, %r1331; and.b32 %r1335, %r74, 7; or.b32 %r1336, %r1334, %r1335; mul.wide.u32 %rd976, %r1336, 2; add.s64 %rd977, %rd2307, %rd976; ld.u16 %rs177, [%rd977+96]; cvt.rn.f32.u16 %f488, %rs177; ld.v2.f32 {%f489, %f490}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f490, %f488, %f489; $L__BB6_206: and.b32 %r3161, %r259, -8; sub.ftz.f32 %f508, %f934, %f933; fma.rn.ftz.f32 %f90, %f21, %f508, %f933; setp.ne.s32 %p127, %r3161, %r3229; mov.u16 %rs374, 0; @%p127 bra $L__BB6_209; and.b32 %r1415, %r8, -8; setp.ne.s32 %p128, %r1415, %r3184; @%p128 bra $L__BB6_209; and.b32 %r1416, %r74, -8; setp.eq.s32 %p129, %r1416, %r3227; selp.u16 %rs374, 1, 0, %p129; $L__BB6_209: setp.eq.s16 %p130, %rs374, 0; @%p130 bra $L__BB6_211; shl.b32 %r1417, %r259, 6; and.b32 %r1418, %r1417, 448; shl.b32 %r1419, %r8, 3; and.b32 %r1420, %r1419, 56; or.b32 %r1421, %r1418, %r1420; and.b32 %r1422, %r74, 7; or.b32 %r1423, %r1421, %r1422; mul.wide.u32 %rd1047, %r1423, 2; add.s64 %rd1048, %rd2307, %rd1047; ld.u16 %rs187, [%rd1048+96]; cvt.rn.f32.u16 %f509, %rs187; ld.v2.f32 {%f510, %f511}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f511, %f509, %f510; $L__BB6_241: and.b32 %r3145, %r259, -8; setp.ne.s32 %p149, %r3145, %r3229; mov.u16 %rs377, 0; @%p149 bra $L__BB6_244; and.b32 %r1501, %r8, -8; setp.ne.s32 %p150, %r1501, %r3184; @%p150 bra $L__BB6_244; and.b32 %r1502, %r9, -8; setp.eq.s32 %p151, %r1502, %r3227; selp.u16 %rs377, 1, 0, %p151; $L__BB6_244: setp.eq.s16 %p152, %rs377, 0; @%p152 bra $L__BB6_246; shl.b32 %r1503, %r259, 6; and.b32 %r1504, %r1503, 448; shl.b32 %r1505, %r8, 3; and.b32 %r1506, %r1505, 56; or.b32 %r1507, %r1504, %r1506; and.b32 %r1508, %r9, 7; or.b32 %r1509, %r1507, %r1508; mul.wide.u32 %rd1118, %r1509, 2; add.s64 %rd1119, %rd2307, %rd1118; ld.u16 %rs197, [%rd1119+96]; cvt.rn.f32.u16 %f529, %rs197; ld.v2.f32 {%f530, %f531}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f531, %f529, %f530; $L__BB6_276: cvt.rmi.ftz.f32.f32 %f915, %f13; sub.ftz.f32 %f549, %f935, %f936; fma.rn.ftz.f32 %f550, %f21, %f549, %f936; sub.ftz.f32 %f551, %f90, %f550; fma.rn.ftz.f32 %f552, %f20, %f551, %f550; sub.ftz.f32 %f553, %f552, %f65; sub.ftz.f32 %f554, %f13, %f915; fma.rn.ftz.f32 %f115, %f554, %f553, %f65; cvt.rmi.ftz.f32.f32 %f555, %f14; sub.ftz.f32 %f116, %f14, %f555; cvt.rzi.ftz.s32.f32 %r471, %f555; and.b32 %r472, %r471, -8; setp.ne.s32 %p171, %r472, %r3229; mov.u16 %rs380, 0; @%p171 bra $L__BB6_279; and.b32 %r1586, %r8, -8; setp.ne.s32 %p172, %r1586, %r3184; @%p172 bra $L__BB6_279; and.b32 %r1587, %r9, -8; setp.eq.s32 %p173, %r1587, %r3227; selp.u16 %rs380, 1, 0, %p173; $L__BB6_279: setp.eq.s16 %p174, %rs380, 0; @%p174 bra $L__BB6_281; shl.b32 %r1588, %r471, 6; and.b32 %r1589, %r1588, 448; shl.b32 %r1590, %r8, 3; and.b32 %r1591, %r1590, 56; or.b32 %r1592, %r1591, %r1589; and.b32 %r1593, %r9, 7; or.b32 %r1594, %r1592, %r1593; mul.wide.u32 %rd1188, %r1594, 2; add.s64 %rd1189, %rd2307, %rd1188; ld.u16 %rs207, [%rd1189+96]; cvt.rn.f32.u16 %f556, %rs207; ld.v2.f32 {%f557, %f558}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f558, %f556, %f557; $L__BB6_311: setp.ne.s32 %p193, %r472, %r3229; mov.u16 %rs383, 0; @%p193 bra $L__BB6_314; and.b32 %r1671, %r8, -8; setp.ne.s32 %p194, %r1671, %r3184; @%p194 bra $L__BB6_314; and.b32 %r1672, %r74, -8; setp.eq.s32 %p195, %r1672, %r3227; selp.u16 %rs383, 1, 0, %p195; $L__BB6_314: setp.eq.s16 %p196, %rs383, 0; @%p196 bra $L__BB6_316; shl.b32 %r1673, %r471, 6; and.b32 %r1674, %r1673, 448; shl.b32 %r1675, %r8, 3; and.b32 %r1676, %r1675, 56; or.b32 %r1677, %r1676, %r1674; and.b32 %r1678, %r74, 7; or.b32 %r1679, %r1677, %r1678; mul.wide.u32 %rd1258, %r1679, 2; add.s64 %rd1259, %rd2307, %rd1258; ld.u16 %rs217, [%rd1259+96]; cvt.rn.f32.u16 %f576, %rs217; ld.v2.f32 {%f577, %f578}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f578, %f576, %f577; $L__BB6_346: sub.ftz.f32 %f596, %f938, %f937; fma.rn.ftz.f32 %f141, %f21, %f596, %f937; setp.ne.s32 %p215, %r472, %r3229; mov.u16 %rs386, 0; @%p215 bra $L__BB6_349; and.b32 %r1757, %r139, -8; setp.ne.s32 %p216, %r1757, %r3184; @%p216 bra $L__BB6_349; and.b32 %r1758, %r74, -8; setp.eq.s32 %p217, %r1758, %r3227; selp.u16 %rs386, 1, 0, %p217; $L__BB6_349: setp.eq.s16 %p218, %rs386, 0; @%p218 bra $L__BB6_351; shl.b32 %r1759, %r471, 6; and.b32 %r1760, %r1759, 448; shl.b32 %r1761, %r139, 3; and.b32 %r1762, %r1761, 56; or.b32 %r1763, %r1762, %r1760; and.b32 %r1764, %r74, 7; or.b32 %r1765, %r1763, %r1764; mul.wide.u32 %rd1329, %r1765, 2; add.s64 %rd1330, %rd2307, %rd1329; ld.u16 %rs227, [%rd1330+96]; cvt.rn.f32.u16 %f597, %rs227; ld.v2.f32 {%f598, %f599}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f599, %f597, %f598; $L__BB6_381: setp.ne.s32 %p237, %r472, %r3229; mov.u16 %rs389, 0; @%p237 bra $L__BB6_384; and.b32 %r1844, %r139, -8; setp.ne.s32 %p238, %r1844, %r3184; @%p238 bra $L__BB6_384; and.b32 %r1845, %r9, -8; setp.eq.s32 %p239, %r1845, %r3227; selp.u16 %rs389, 1, 0, %p239; $L__BB6_384: setp.eq.s16 %p240, %rs389, 0; @%p240 bra $L__BB6_386; and.b32 %r1846, %r471, 7; shl.b32 %r1847, %r139, 3; and.b32 %r1848, %r1847, 56; and.b32 %r1849, %r9, 7; bfi.b32 %r1850, %r1846, %r1849, 6, 3; or.b32 %r1851, %r1850, %r1848; mul.wide.u32 %rd1400, %r1851, 2; add.s64 %rd1401, %rd2307, %rd1400; ld.u16 %rs237, [%rd1401+96]; cvt.rn.f32.u16 %f617, %rs237; ld.v2.f32 {%f618, %f619}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f619, %f617, %f618; $L__BB6_416: sub.ftz.f32 %f637, %f939, %f940; fma.rn.ftz.f32 %f638, %f21, %f637, %f940; sub.ftz.f32 %f639, %f638, %f141; fma.rn.ftz.f32 %f166, %f20, %f639, %f141; add.s32 %r635, %r471, 1; and.b32 %r636, %r635, -8; setp.ne.s32 %p259, %r636, %r3229; mov.u16 %rs392, 0; @%p259 bra $L__BB6_419; and.b32 %r1926, %r139, -8; setp.ne.s32 %p260, %r1926, %r3184; @%p260 bra $L__BB6_419; and.b32 %r1927, %r9, -8; setp.eq.s32 %p261, %r1927, %r3227; selp.u16 %rs392, 1, 0, %p261; $L__BB6_419: setp.eq.s16 %p262, %rs392, 0; @%p262 bra $L__BB6_421; and.b32 %r1928, %r635, 7; shl.b32 %r1929, %r139, 3; and.b32 %r1930, %r1929, 56; and.b32 %r1931, %r9, 7; bfi.b32 %r1932, %r1928, %r1931, 6, 3; or.b32 %r1933, %r1932, %r1930; mul.wide.u32 %rd1471, %r1933, 2; add.s64 %rd1472, %rd2307, %rd1471; ld.u16 %rs247, [%rd1472+96]; cvt.rn.f32.u16 %f640, %rs247; ld.v2.f32 {%f641, %f642}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f642, %f640, %f641; $L__BB6_451: setp.ne.s32 %p281, %r636, %r3229; mov.u16 %rs395, 0; @%p281 bra $L__BB6_454; and.b32 %r2008, %r139, -8; setp.ne.s32 %p282, %r2008, %r3184; @%p282 bra $L__BB6_454; and.b32 %r2009, %r74, -8; setp.eq.s32 %p283, %r2009, %r3227; selp.u16 %rs395, 1, 0, %p283; $L__BB6_454: setp.eq.s16 %p284, %rs395, 0; @%p284 bra $L__BB6_456; shl.b32 %r2010, %r635, 6; and.b32 %r2011, %r2010, 448; shl.b32 %r2012, %r139, 3; and.b32 %r2013, %r2012, 56; or.b32 %r2014, %r2013, %r2011; and.b32 %r2015, %r74, 7; or.b32 %r2016, %r2014, %r2015; mul.wide.u32 %rd1542, %r2016, 2; add.s64 %rd1543, %rd2307, %rd1542; ld.u16 %rs257, [%rd1543+96]; cvt.rn.f32.u16 %f660, %rs257; ld.v2.f32 {%f661, %f662}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f662, %f660, %f661; $L__BB6_486: sub.ftz.f32 %f680, %f942, %f941; fma.rn.ftz.f32 %f191, %f21, %f680, %f941; setp.ne.s32 %p303, %r636, %r3229; mov.u16 %rs398, 0; @%p303 bra $L__BB6_489; and.b32 %r2095, %r8, -8; setp.ne.s32 %p304, %r2095, %r3184; @%p304 bra $L__BB6_489; and.b32 %r2096, %r74, -8; setp.eq.s32 %p305, %r2096, %r3227; selp.u16 %rs398, 1, 0, %p305; $L__BB6_489: setp.eq.s16 %p306, %rs398, 0; @%p306 bra $L__BB6_491; shl.b32 %r2097, %r635, 6; and.b32 %r2098, %r2097, 448; shl.b32 %r2099, %r8, 3; and.b32 %r2100, %r2099, 56; or.b32 %r2101, %r2098, %r2100; and.b32 %r2102, %r74, 7; or.b32 %r2103, %r2101, %r2102; mul.wide.u32 %rd1613, %r2103, 2; add.s64 %rd1614, %rd2307, %rd1613; ld.u16 %rs267, [%rd1614+96]; cvt.rn.f32.u16 %f681, %rs267; ld.v2.f32 {%f682, %f683}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f683, %f681, %f682; $L__BB6_521: setp.ne.s32 %p325, %r636, %r3229; mov.u16 %rs401, 0; @%p325 bra $L__BB6_524; and.b32 %r2181, %r8, -8; setp.ne.s32 %p326, %r2181, %r3184; @%p326 bra $L__BB6_524; and.b32 %r2182, %r9, -8; setp.eq.s32 %p327, %r2182, %r3227; selp.u16 %rs401, 1, 0, %p327; $L__BB6_524: setp.eq.s16 %p328, %rs401, 0; @%p328 bra $L__BB6_526; shl.b32 %r2183, %r635, 6; and.b32 %r2184, %r2183, 448; shl.b32 %r2185, %r8, 3; and.b32 %r2186, %r2185, 56; or.b32 %r2187, %r2184, %r2186; and.b32 %r2188, %r9, 7; or.b32 %r2189, %r2187, %r2188; mul.wide.u32 %rd1684, %r2189, 2; add.s64 %rd1685, %rd2307, %rd1684; ld.u16 %rs277, [%rd1685+96]; cvt.rn.f32.u16 %f701, %rs277; ld.v2.f32 {%f702, %f703}, [%rd2307+80]; fma.rn.ftz.f32 %f944, %f703, %f701, %f702; $L__BB6_556: add.u64 %rd2288, %SP, 80; sub.ftz.f32 %f721, %f943, %f944; fma.rn.ftz.f32 %f722, %f21, %f721, %f944; sub.ftz.f32 %f723, %f191, %f722; fma.rn.ftz.f32 %f724, %f20, %f723, %f722; sub.ftz.f32 %f725, %f724, %f166; fma.rn.ftz.f32 %f726, %f116, %f725, %f166; mul.ftz.f32 %f727, %f8, %f726; fma.rn.ftz.f32 %f728, %f7, %f115, %f727; mul.ftz.f32 %f729, %f9, %f728; { // callseq 20, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f730, [retval0+0]; } // callseq 20 { // callseq 21, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f17; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f731, [retval0+0]; } // callseq 21 mul.ftz.f32 %f732, %f8, %f731; fma.rn.ftz.f32 %f733, %f7, %f730, %f732; fma.rn.ftz.f32 %f734, %f10, %f733, %f729; mul.ftz.f32 %f216, %f11, %f734; { // callseq 22, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f735, [retval0+0]; } // callseq 22 { // callseq 23, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f14; st.param.f32 [param1+4], %f15; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f736, [retval0+0]; } // callseq 23 mul.ftz.f32 %f737, %f8, %f736; fma.rn.ftz.f32 %f738, %f7, %f735, %f737; mul.ftz.f32 %f217, %f9, %f738; { // callseq 24, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2288; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f13; st.param.f32 [param1+4], %f16; st.param.f32 [param1+8], %f18; .param .b32 retval0; call.uni (retval0), _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE, ( param0, param1 ); ld.param.f32 %f739, [retval0+0]; } // callseq 24 mul.ftz.f32 %f218, %f7, %f739; cvt.rmi.ftz.f32.f32 %f740, %f16; sub.ftz.f32 %f219, %f16, %f740; cvt.rmi.ftz.f32.f32 %f741, %f18; sub.ftz.f32 %f220, %f18, %f741; cvt.rzi.ftz.s32.f32 %r732, %f740; cvt.rzi.ftz.s32.f32 %r733, %f741; ld.local.u32 %r3598, [%rd2]; setp.ne.s32 %p347, %r472, %r3598; mov.u16 %rs404, 0; @%p347 bra $L__BB6_559; and.b32 %r2274, %r732, -8; ld.local.u32 %r2275, [%rd2+4]; setp.ne.s32 %p348, %r2274, %r2275; @%p348 bra $L__BB6_559; and.b32 %r2276, %r733, -8; ld.local.u32 %r2277, [%rd2+8]; setp.eq.s32 %p349, %r2276, %r2277; selp.u16 %rs404, 1, 0, %p349; $L__BB6_559: setp.eq.s16 %p350, %rs404, 0; @%p350 bra $L__BB6_561; shl.b32 %r2278, %r471, 6; and.b32 %r2279, %r2278, 448; shl.b32 %r2280, %r732, 3; and.b32 %r2281, %r2280, 56; or.b32 %r2282, %r2281, %r2279; and.b32 %r2283, %r733, 7; or.b32 %r2284, %r2282, %r2283; ld.local.u64 %rd1749, [%rd2+48]; mul.wide.u32 %rd1750, %r2284, 2; add.s64 %rd1751, %rd1749, %rd1750; ld.u16 %rs287, [%rd1751+96]; cvt.rn.f32.u16 %f742, %rs287; ld.v2.f32 {%f743, %f744}, [%rd1749+80]; fma.rn.ftz.f32 %f945, %f744, %f742, %f743; $L__BB6_591: add.s32 %r741, %r733, 1; setp.ne.s32 %p369, %r472, %r3598; mov.u16 %rs407, 0; @%p369 bra $L__BB6_594; and.b32 %r2377, %r732, -8; ld.local.u32 %r2378, [%rd2+4]; setp.ne.s32 %p370, %r2377, %r2378; @%p370 bra $L__BB6_594; and.b32 %r2379, %r741, -8; ld.local.u32 %r2380, [%rd2+8]; setp.eq.s32 %p371, %r2379, %r2380; selp.u16 %rs407, 1, 0, %p371; $L__BB6_594: setp.eq.s16 %p372, %rs407, 0; @%p372 bra $L__BB6_596; shl.b32 %r2381, %r471, 6; and.b32 %r2382, %r2381, 448; shl.b32 %r2383, %r732, 3; and.b32 %r2384, %r2383, 56; or.b32 %r2385, %r2384, %r2382; and.b32 %r2386, %r741, 7; or.b32 %r2387, %r2385, %r2386; ld.local.u64 %rd1816, [%rd2+48]; mul.wide.u32 %rd1817, %r2387, 2; add.s64 %rd1818, %rd1816, %rd1817; ld.u16 %rs297, [%rd1818+96]; cvt.rn.f32.u16 %f762, %rs297; ld.v2.f32 {%f763, %f764}, [%rd1816+80]; fma.rn.ftz.f32 %f946, %f764, %f762, %f763; $L__BB6_626: sub.ftz.f32 %f782, %f946, %f945; fma.rn.ftz.f32 %f245, %f220, %f782, %f945; add.s32 %r748, %r732, 1; setp.ne.s32 %p391, %r472, %r3598; mov.u16 %rs410, 0; @%p391 bra $L__BB6_629; and.b32 %r2480, %r748, -8; ld.local.u32 %r2481, [%rd2+4]; setp.ne.s32 %p392, %r2480, %r2481; @%p392 bra $L__BB6_629; and.b32 %r2482, %r741, -8; ld.local.u32 %r2483, [%rd2+8]; setp.eq.s32 %p393, %r2482, %r2483; selp.u16 %rs410, 1, 0, %p393; $L__BB6_629: setp.eq.s16 %p394, %rs410, 0; @%p394 bra $L__BB6_631; shl.b32 %r2484, %r471, 6; and.b32 %r2485, %r2484, 448; shl.b32 %r2486, %r748, 3; and.b32 %r2487, %r2486, 56; or.b32 %r2488, %r2487, %r2485; and.b32 %r2489, %r741, 7; or.b32 %r2490, %r2488, %r2489; ld.local.u64 %rd1884, [%rd2+48]; mul.wide.u32 %rd1885, %r2490, 2; add.s64 %rd1886, %rd1884, %rd1885; ld.u16 %rs307, [%rd1886+96]; cvt.rn.f32.u16 %f783, %rs307; ld.v2.f32 {%f784, %f785}, [%rd1884+80]; fma.rn.ftz.f32 %f947, %f785, %f783, %f784; $L__BB6_661: setp.ne.s32 %p413, %r472, %r3598; mov.u16 %rs413, 0; @%p413 bra $L__BB6_664; and.b32 %r2583, %r748, -8; ld.local.u32 %r2584, [%rd2+4]; setp.ne.s32 %p414, %r2583, %r2584; @%p414 bra $L__BB6_664; and.b32 %r2585, %r733, -8; ld.local.u32 %r2586, [%rd2+8]; setp.eq.s32 %p415, %r2585, %r2586; selp.u16 %rs413, 1, 0, %p415; $L__BB6_664: setp.eq.s16 %p416, %rs413, 0; @%p416 bra $L__BB6_666; and.b32 %r2587, %r471, 7; shl.b32 %r2588, %r748, 3; and.b32 %r2589, %r2588, 56; and.b32 %r2590, %r733, 7; bfi.b32 %r2591, %r2587, %r2590, 6, 3; or.b32 %r2592, %r2591, %r2589; ld.local.u64 %rd1951, [%rd2+48]; mul.wide.u32 %rd1952, %r2592, 2; add.s64 %rd1953, %rd1951, %rd1952; ld.u16 %rs317, [%rd1953+96]; cvt.rn.f32.u16 %f803, %rs317; ld.v2.f32 {%f804, %f805}, [%rd1951+80]; fma.rn.ftz.f32 %f948, %f805, %f803, %f804; $L__BB6_696: sub.ftz.f32 %f823, %f947, %f948; fma.rn.ftz.f32 %f824, %f220, %f823, %f948; sub.ftz.f32 %f825, %f824, %f245; fma.rn.ftz.f32 %f270, %f219, %f825, %f245; setp.ne.s32 %p435, %r636, %r3598; mov.u16 %rs416, 0; @%p435 bra $L__BB6_699; and.b32 %r2682, %r748, -8; ld.local.u32 %r2683, [%rd2+4]; setp.ne.s32 %p436, %r2682, %r2683; @%p436 bra $L__BB6_699; and.b32 %r2684, %r733, -8; ld.local.u32 %r2685, [%rd2+8]; setp.eq.s32 %p437, %r2684, %r2685; selp.u16 %rs416, 1, 0, %p437; $L__BB6_699: setp.eq.s16 %p438, %rs416, 0; @%p438 bra $L__BB6_701; and.b32 %r2686, %r635, 7; shl.b32 %r2687, %r748, 3; and.b32 %r2688, %r2687, 56; and.b32 %r2689, %r733, 7; bfi.b32 %r2690, %r2686, %r2689, 6, 3; or.b32 %r2691, %r2690, %r2688; ld.local.u64 %rd2018, [%rd2+48]; mul.wide.u32 %rd2019, %r2691, 2; add.s64 %rd2020, %rd2018, %rd2019; ld.u16 %rs327, [%rd2020+96]; cvt.rn.f32.u16 %f826, %rs327; ld.v2.f32 {%f827, %f828}, [%rd2018+80]; fma.rn.ftz.f32 %f949, %f828, %f826, %f827; $L__BB6_731: setp.ne.s32 %p457, %r636, %r3598; mov.u16 %rs419, 0; @%p457 bra $L__BB6_734; and.b32 %r2781, %r748, -8; ld.local.u32 %r2782, [%rd2+4]; setp.ne.s32 %p458, %r2781, %r2782; @%p458 bra $L__BB6_734; and.b32 %r2783, %r741, -8; ld.local.u32 %r2784, [%rd2+8]; setp.eq.s32 %p459, %r2783, %r2784; selp.u16 %rs419, 1, 0, %p459; $L__BB6_734: setp.eq.s16 %p460, %rs419, 0; @%p460 bra $L__BB6_736; shl.b32 %r2785, %r635, 6; and.b32 %r2786, %r2785, 448; shl.b32 %r2787, %r748, 3; and.b32 %r2788, %r2787, 56; or.b32 %r2789, %r2788, %r2786; and.b32 %r2790, %r741, 7; or.b32 %r2791, %r2789, %r2790; ld.local.u64 %rd2085, [%rd2+48]; mul.wide.u32 %rd2086, %r2791, 2; add.s64 %rd2087, %rd2085, %rd2086; ld.u16 %rs337, [%rd2087+96]; cvt.rn.f32.u16 %f846, %rs337; ld.v2.f32 {%f847, %f848}, [%rd2085+80]; fma.rn.ftz.f32 %f950, %f848, %f846, %f847; $L__BB6_766: sub.ftz.f32 %f866, %f950, %f949; fma.rn.ftz.f32 %f295, %f220, %f866, %f949; setp.ne.s32 %p479, %r636, %r3598; mov.u16 %rs422, 0; @%p479 bra $L__BB6_769; and.b32 %r2884, %r732, -8; ld.local.u32 %r2885, [%rd2+4]; setp.ne.s32 %p480, %r2884, %r2885; @%p480 bra $L__BB6_769; and.b32 %r2886, %r741, -8; ld.local.u32 %r2887, [%rd2+8]; setp.eq.s32 %p481, %r2886, %r2887; selp.u16 %rs422, 1, 0, %p481; $L__BB6_769: setp.eq.s16 %p482, %rs422, 0; @%p482 bra $L__BB6_771; shl.b32 %r2888, %r635, 6; and.b32 %r2889, %r2888, 448; shl.b32 %r2890, %r732, 3; and.b32 %r2891, %r2890, 56; or.b32 %r2892, %r2889, %r2891; and.b32 %r2893, %r741, 7; or.b32 %r2894, %r2892, %r2893; ld.local.u64 %rd2152, [%rd2+48]; mul.wide.u32 %rd2153, %r2894, 2; add.s64 %rd2154, %rd2152, %rd2153; ld.u16 %rs347, [%rd2154+96]; cvt.rn.f32.u16 %f867, %rs347; ld.v2.f32 {%f868, %f869}, [%rd2152+80]; fma.rn.ftz.f32 %f951, %f869, %f867, %f868; $L__BB6_801: setp.ne.s32 %p501, %r636, %r3598; @%p501 bra $L__BB6_804; and.b32 %r2987, %r732, -8; ld.local.u32 %r2988, [%rd2+4]; setp.ne.s32 %p502, %r2987, %r2988; @%p502 bra $L__BB6_804; and.b32 %r2989, %r733, -8; ld.local.u32 %r2990, [%rd2+8]; setp.eq.s32 %p503, %r2989, %r2990; @%p503 bra $L__BB6_832; bra.uni $L__BB6_804; $L__BB6_832: shl.b32 %r3083, %r635, 6; and.b32 %r3084, %r3083, 448; shl.b32 %r3085, %r732, 3; and.b32 %r3086, %r3085, 56; or.b32 %r3087, %r3084, %r3086; and.b32 %r3088, %r733, 7; or.b32 %r3089, %r3087, %r3088; ld.local.u64 %rd2285, [%rd2+48]; mul.wide.u32 %rd2286, %r3089, 2; add.s64 %rd2287, %rd2285, %rd2286; ld.u16 %rs358, [%rd2287+96]; cvt.rn.f32.u16 %f902, %rs358; ld.v2.f32 {%f903, %f904}, [%rd2285+80]; fma.rn.ftz.f32 %f952, %f904, %f902, %f903; bra.uni $L__BB6_833; $L__BB6_804: and.b32 %r779, %r635, -128; ld.local.u32 %r2991, [%rd2+12]; setp.ne.s32 %p504, %r779, %r2991; @%p504 bra $L__BB6_807; and.b32 %r2992, %r732, -128; ld.local.u32 %r2993, [%rd2+16]; setp.ne.s32 %p505, %r2992, %r2993; @%p505 bra $L__BB6_807; and.b32 %r2994, %r733, -128; ld.local.u32 %r2995, [%rd2+20]; setp.eq.s32 %p506, %r2994, %r2995; @%p506 bra $L__BB6_829; bra.uni $L__BB6_807; $L__BB6_829: shl.b32 %r3065, %r635, 5; and.b32 %r3066, %r3065, 3840; shl.b32 %r3067, %r732, 1; and.b32 %r3068, %r3067, 240; or.b32 %r3069, %r3066, %r3068; bfe.u32 %r3070, %r733, 3, 4; or.b32 %r3071, %r3069, %r3070; shr.u32 %r3072, %r3069, 6; ld.local.u64 %rd606, [%rd2+56]; mul.wide.u32 %rd2273, %r3072, 8; add.s64 %rd2274, %rd606, %rd2273; and.b32 %r3073, %r3071, 63; mov.u64 %rd2275, 1; shl.b64 %rd2276, %rd2275, %r3073; ld.u64 %rd2277, [%rd2274+544]; and.b64 %rd2278, %rd2277, %rd2276; setp.eq.s64 %p519, %rd2278, 0; mul.wide.u32 %rd2279, %r3071, 8; add.s64 %rd2280, %rd606, %rd2279; add.s64 %rd607, %rd2280, 1088; @%p519 bra $L__BB6_831; ld.u64 %rd2281, [%rd607]; add.s64 %rd2282, %rd606, %rd2281; and.b32 %r3074, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3074}; and.b32 %r3075, %r733, -8; st.local.u32 [%rd2+8], %r3075; st.local.u64 [%rd2+48], %rd2282; shl.b32 %r3076, %r635, 6; and.b32 %r3077, %r3076, 448; shl.b32 %r3078, %r732, 3; and.b32 %r3079, %r3078, 56; or.b32 %r3080, %r3077, %r3079; and.b32 %r3081, %r733, 7; or.b32 %r3082, %r3080, %r3081; mul.wide.u32 %rd2283, %r3082, 2; add.s64 %rd2284, %rd2282, %rd2283; ld.u16 %rs357, [%rd2284+96]; cvt.rn.f32.u16 %f897, %rs357; ld.v2.f32 {%f898, %f899}, [%rd2282+80]; fma.rn.ftz.f32 %f952, %f899, %f897, %f898; bra.uni $L__BB6_833; $L__BB6_807: and.b32 %r780, %r635, -4096; ld.local.u32 %r2996, [%rd2+24]; setp.ne.s32 %p507, %r780, %r2996; @%p507 bra $L__BB6_810; and.b32 %r2997, %r732, -4096; ld.local.u32 %r2998, [%rd2+28]; setp.ne.s32 %p508, %r2997, %r2998; @%p508 bra $L__BB6_810; and.b32 %r2999, %r733, -4096; ld.local.u32 %r3000, [%rd2+32]; setp.eq.s32 %p509, %r2999, %r3000; @%p509 bra $L__BB6_824; bra.uni $L__BB6_810; $L__BB6_824: shl.b32 %r3036, %r635, 3; and.b32 %r3037, %r3036, 31744; and.b32 %r3038, %r732, 3968; shr.u32 %r3039, %r3038, 2; or.b32 %r3040, %r3037, %r3039; bfe.u32 %r3041, %r733, 7, 5; or.b32 %r3042, %r3040, %r3041; shr.u32 %r3043, %r3040, 6; ld.local.u64 %rd602, [%rd2+64]; mul.wide.u32 %rd2252, %r3043, 8; add.s64 %rd2253, %rd602, %rd2252; and.b32 %r3044, %r3042, 63; mov.u64 %rd2254, 1; shl.b64 %rd2255, %rd2254, %r3044; ld.u64 %rd2256, [%rd2253+4128]; and.b64 %rd2257, %rd2256, %rd2255; setp.eq.s64 %p517, %rd2257, 0; mul.wide.u32 %rd2258, %r3042, 8; add.s64 %rd2259, %rd602, %rd2258; add.s64 %rd603, %rd2259, 8256; @%p517 bra $L__BB6_828; ld.u64 %rd2260, [%rd603]; add.s64 %rd604, %rd602, %rd2260; st.local.u32 [%rd2+12], %r779; and.b32 %r3045, %r732, -128; st.local.u32 [%rd2+16], %r3045; and.b32 %r3046, %r733, -128; st.local.u32 [%rd2+20], %r3046; st.local.u64 [%rd2+56], %rd604; shl.b32 %r3047, %r635, 5; and.b32 %r3048, %r3047, 3840; shl.b32 %r3049, %r732, 1; and.b32 %r3050, %r3049, 240; or.b32 %r3051, %r3048, %r3050; bfe.u32 %r3052, %r733, 3, 4; or.b32 %r3053, %r3051, %r3052; shr.u32 %r3054, %r3051, 6; mul.wide.u32 %rd2261, %r3054, 8; add.s64 %rd2262, %rd604, %rd2261; and.b32 %r3055, %r3053, 63; shl.b64 %rd2264, %rd2254, %r3055; ld.u64 %rd2265, [%rd2262+544]; and.b64 %rd2266, %rd2265, %rd2264; setp.eq.s64 %p518, %rd2266, 0; mul.wide.u32 %rd2267, %r3053, 8; add.s64 %rd2268, %rd604, %rd2267; add.s64 %rd605, %rd2268, 1088; @%p518 bra $L__BB6_827; ld.u64 %rd2269, [%rd605]; add.s64 %rd2270, %rd604, %rd2269; and.b32 %r3056, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3056}; and.b32 %r3057, %r733, -8; st.local.u32 [%rd2+8], %r3057; st.local.u64 [%rd2+48], %rd2270; shl.b32 %r3058, %r635, 6; and.b32 %r3059, %r3058, 448; shl.b32 %r3060, %r732, 3; and.b32 %r3061, %r3060, 56; or.b32 %r3062, %r3059, %r3061; and.b32 %r3063, %r733, 7; or.b32 %r3064, %r3062, %r3063; mul.wide.u32 %rd2271, %r3064, 2; add.s64 %rd2272, %rd2270, %rd2271; ld.u16 %rs356, [%rd2272+96]; cvt.rn.f32.u16 %f892, %rs356; ld.v2.f32 {%f893, %f894}, [%rd2270+80]; fma.rn.ftz.f32 %f952, %f894, %f892, %f893; bra.uni $L__BB6_833; $L__BB6_810: ld.local.u64 %rd590, [%rd2+40]; mov.u64 %rd2394, 0; ld.u32 %r781, [%rd590+24]; setp.eq.s32 %p510, %r781, 0; @%p510 bra $L__BB6_815; add.s64 %rd591, %rd590, 64; shr.u32 %r3002, %r635, 12; cvt.u64.u32 %rd2221, %r3002; shl.b64 %rd2222, %rd2221, 42; shr.u32 %r3003, %r732, 12; mul.wide.u32 %rd2223, %r3003, 2097152; shr.u32 %r3004, %r733, 12; cvt.u64.u32 %rd2224, %r3004; or.b64 %rd2225, %rd2223, %rd2224; or.b64 %rd592, %rd2225, %rd2222; mov.u32 %r3609, 0; $L__BB6_812: cvt.u64.u32 %rd593, %r3609; mul.wide.u32 %rd2226, %r3609, 32; add.s64 %rd2227, %rd591, %rd2226; ld.u64 %rd2228, [%rd2227]; setp.eq.s64 %p511, %rd2228, %rd592; add.s32 %r3609, %r3609, 1; @%p511 bra $L__BB6_814; setp.lt.u32 %p512, %r3609, %r781; @%p512 bra $L__BB6_812; bra.uni $L__BB6_815; $L__BB6_36: and.b32 %r3156, %r7, -128; setp.ne.s32 %p21, %r3156, %r3272; mov.u16 %rs360, 0; @%p21 bra $L__BB6_39; and.b32 %r1003, %r8, -128; setp.ne.s32 %p22, %r1003, %r3271; @%p22 bra $L__BB6_39; and.b32 %r1004, %r74, -128; setp.eq.s32 %p23, %r1004, %r3270; selp.u16 %rs360, 1, 0, %p23; $L__BB6_39: setp.eq.s16 %p24, %rs360, 0; @%p24 bra $L__BB6_43; shl.b32 %r1005, %r7, 5; and.b32 %r1006, %r1005, 3840; shl.b32 %r1007, %r8, 1; and.b32 %r1008, %r1007, 240; or.b32 %r1009, %r1008, %r1006; bfe.u32 %r1010, %r74, 3, 4; or.b32 %r1011, %r1010, %r1009; shr.u32 %r1012, %r1009, 6; mul.wide.u32 %rd700, %r1012, 8; add.s64 %rd701, %rd2311, %rd700; and.b32 %r1013, %r1011, 63; mov.u64 %rd702, 1; shl.b64 %rd703, %rd702, %r1013; ld.u64 %rd704, [%rd701+544]; and.b64 %rd705, %rd704, %rd703; setp.eq.s64 %p25, %rd705, 0; mul.wide.u32 %rd706, %r1011, 8; add.s64 %rd707, %rd2311, %rd706; add.s64 %rd43, %rd707, 1088; @%p25 bra $L__BB6_42; and.b32 %r3229, %r7, -8; ld.u64 %rd708, [%rd43]; add.s64 %rd2307, %rd2311, %rd708; and.b32 %r1014, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1014}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1015, %r7, 6; and.b32 %r1016, %r1015, 448; shl.b32 %r1017, %r8, 3; and.b32 %r1018, %r1017, 56; or.b32 %r1019, %r1018, %r1016; and.b32 %r1020, %r74, 7; or.b32 %r1021, %r1019, %r1020; mul.wide.u32 %rd709, %r1021, 2; add.s64 %rd710, %rd2307, %rd709; ld.u16 %rs140, [%rd710+96]; cvt.rn.f32.u16 %f409, %rs140; ld.v2.f32 {%f410, %f411}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f411, %f409, %f410; mov.b64 %rd711, {%r3229, %r1014}; shr.u64 %rd712, %rd711, 32; cvt.u32.u64 %r3184, %rd712; bra.uni $L__BB6_66; $L__BB6_71: and.b32 %r3154, %r7, -128; setp.ne.s32 %p43, %r3154, %r3272; mov.u16 %rs363, 0; @%p43 bra $L__BB6_74; and.b32 %r1088, %r139, -128; setp.ne.s32 %p44, %r1088, %r3271; @%p44 bra $L__BB6_74; and.b32 %r1089, %r74, -128; setp.eq.s32 %p45, %r1089, %r3270; selp.u16 %rs363, 1, 0, %p45; $L__BB6_74: setp.eq.s16 %p46, %rs363, 0; @%p46 bra $L__BB6_78; shl.b32 %r1090, %r7, 5; and.b32 %r1091, %r1090, 3840; shl.b32 %r1092, %r139, 1; and.b32 %r1093, %r1092, 240; or.b32 %r1094, %r1093, %r1091; bfe.u32 %r1095, %r74, 3, 4; or.b32 %r1096, %r1094, %r1095; shr.u32 %r1097, %r1094, 6; mul.wide.u32 %rd769, %r1097, 8; add.s64 %rd770, %rd2311, %rd769; and.b32 %r1098, %r1096, 63; mov.u64 %rd771, 1; shl.b64 %rd772, %rd771, %r1098; ld.u64 %rd773, [%rd770+544]; and.b64 %rd774, %rd773, %rd772; setp.eq.s64 %p47, %rd774, 0; mul.wide.u32 %rd775, %r1096, 8; add.s64 %rd776, %rd2311, %rd775; add.s64 %rd72, %rd776, 1088; @%p47 bra $L__BB6_77; and.b32 %r3229, %r7, -8; ld.u64 %rd777, [%rd72]; add.s64 %rd2307, %rd2311, %rd777; and.b32 %r1099, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1099}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1100, %r7, 6; and.b32 %r1101, %r1100, 448; shl.b32 %r1102, %r139, 3; and.b32 %r1103, %r1102, 56; or.b32 %r1104, %r1103, %r1101; and.b32 %r1105, %r74, 7; or.b32 %r1106, %r1104, %r1105; mul.wide.u32 %rd778, %r1106, 2; add.s64 %rd779, %rd2307, %rd778; ld.u16 %rs150, [%rd779+96]; cvt.rn.f32.u16 %f430, %rs150; ld.v2.f32 {%f431, %f432}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f432, %f430, %f431; mov.b64 %rd780, {%r3229, %r1099}; shr.u64 %rd781, %rd780, 32; cvt.u32.u64 %r3184, %rd781; bra.uni $L__BB6_101; $L__BB6_106: and.b32 %r3127, %r7, -128; setp.ne.s32 %p65, %r3127, %r3272; mov.u16 %rs366, 0; @%p65 bra $L__BB6_109; and.b32 %r1173, %r139, -128; setp.ne.s32 %p66, %r1173, %r3271; @%p66 bra $L__BB6_109; and.b32 %r1174, %r9, -128; setp.eq.s32 %p67, %r1174, %r3270; selp.u16 %rs366, 1, 0, %p67; $L__BB6_109: setp.eq.s16 %p68, %rs366, 0; @%p68 bra $L__BB6_113; shl.b32 %r1175, %r7, 5; and.b32 %r1176, %r1175, 3840; shl.b32 %r1177, %r139, 1; and.b32 %r1178, %r1177, 240; or.b32 %r1179, %r1178, %r1176; bfe.u32 %r1180, %r9, 3, 4; or.b32 %r1181, %r1179, %r1180; shr.u32 %r1182, %r1179, 6; mul.wide.u32 %rd839, %r1182, 8; add.s64 %rd840, %rd2311, %rd839; and.b32 %r1183, %r1181, 63; mov.u64 %rd841, 1; shl.b64 %rd842, %rd841, %r1183; ld.u64 %rd843, [%rd840+544]; and.b64 %rd844, %rd843, %rd842; setp.eq.s64 %p69, %rd844, 0; mul.wide.u32 %rd845, %r1181, 8; add.s64 %rd846, %rd2311, %rd845; add.s64 %rd101, %rd846, 1088; @%p69 bra $L__BB6_112; and.b32 %r3229, %r7, -8; ld.u64 %rd847, [%rd101]; add.s64 %rd2307, %rd2311, %rd847; and.b32 %r1184, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1184}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1185, %r7, 7; shl.b32 %r1186, %r139, 3; and.b32 %r1187, %r1186, 56; and.b32 %r1188, %r9, 7; bfi.b32 %r1189, %r1185, %r1188, 6, 3; or.b32 %r1190, %r1189, %r1187; mul.wide.u32 %rd848, %r1190, 2; add.s64 %rd849, %rd2307, %rd848; ld.u16 %rs160, [%rd849+96]; cvt.rn.f32.u16 %f450, %rs160; ld.v2.f32 {%f451, %f452}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f452, %f450, %f451; mov.b64 %rd850, {%r3229, %r1184}; shr.u64 %rd851, %rd850, 32; cvt.u32.u64 %r3184, %rd851; bra.uni $L__BB6_136; $L__BB6_141: and.b32 %r261, %r259, -128; setp.ne.s32 %p87, %r261, %r3272; mov.u16 %rs369, 0; @%p87 bra $L__BB6_144; and.b32 %r1254, %r139, -128; setp.ne.s32 %p88, %r1254, %r3271; @%p88 bra $L__BB6_144; and.b32 %r1255, %r9, -128; setp.eq.s32 %p89, %r1255, %r3270; selp.u16 %rs369, 1, 0, %p89; $L__BB6_144: setp.eq.s16 %p90, %rs369, 0; @%p90 bra $L__BB6_148; shl.b32 %r1256, %r259, 5; and.b32 %r1257, %r1256, 3840; shl.b32 %r1258, %r139, 1; and.b32 %r1259, %r1258, 240; or.b32 %r1260, %r1259, %r1257; bfe.u32 %r1261, %r9, 3, 4; or.b32 %r1262, %r1260, %r1261; shr.u32 %r1263, %r1260, 6; mul.wide.u32 %rd907, %r1263, 8; add.s64 %rd908, %rd2311, %rd907; and.b32 %r1264, %r1262, 63; mov.u64 %rd909, 1; shl.b64 %rd910, %rd909, %r1264; ld.u64 %rd911, [%rd908+544]; and.b64 %rd912, %rd911, %rd910; setp.eq.s64 %p91, %rd912, 0; mul.wide.u32 %rd913, %r1262, 8; add.s64 %rd914, %rd2311, %rd913; add.s64 %rd130, %rd914, 1088; @%p91 bra $L__BB6_147; and.b32 %r3229, %r259, -8; ld.u64 %rd915, [%rd130]; add.s64 %rd2307, %rd2311, %rd915; and.b32 %r1265, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1265}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1266, %r259, 7; shl.b32 %r1267, %r139, 3; and.b32 %r1268, %r1267, 56; and.b32 %r1269, %r9, 7; bfi.b32 %r1270, %r1266, %r1269, 6, 3; or.b32 %r1271, %r1270, %r1268; mul.wide.u32 %rd916, %r1271, 2; add.s64 %rd917, %rd2307, %rd916; ld.u16 %rs170, [%rd917+96]; cvt.rn.f32.u16 %f473, %rs170; ld.v2.f32 {%f474, %f475}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f475, %f473, %f474; mov.b64 %rd918, {%r3229, %r1265}; shr.u64 %rd919, %rd918, 32; cvt.u32.u64 %r3184, %rd919; bra.uni $L__BB6_171; $L__BB6_176: and.b32 %r318, %r259, -128; setp.ne.s32 %p109, %r318, %r3272; mov.u16 %rs372, 0; @%p109 bra $L__BB6_179; and.b32 %r1337, %r139, -128; setp.ne.s32 %p110, %r1337, %r3271; @%p110 bra $L__BB6_179; and.b32 %r1338, %r74, -128; setp.eq.s32 %p111, %r1338, %r3270; selp.u16 %rs372, 1, 0, %p111; $L__BB6_179: setp.eq.s16 %p112, %rs372, 0; @%p112 bra $L__BB6_183; shl.b32 %r1339, %r259, 5; and.b32 %r1340, %r1339, 3840; shl.b32 %r1341, %r139, 1; and.b32 %r1342, %r1341, 240; or.b32 %r1343, %r1342, %r1340; bfe.u32 %r1344, %r74, 3, 4; or.b32 %r1345, %r1343, %r1344; shr.u32 %r1346, %r1343, 6; mul.wide.u32 %rd978, %r1346, 8; add.s64 %rd979, %rd2311, %rd978; and.b32 %r1347, %r1345, 63; mov.u64 %rd980, 1; shl.b64 %rd981, %rd980, %r1347; ld.u64 %rd982, [%rd979+544]; and.b64 %rd983, %rd982, %rd981; setp.eq.s64 %p113, %rd983, 0; mul.wide.u32 %rd984, %r1345, 8; add.s64 %rd985, %rd2311, %rd984; add.s64 %rd159, %rd985, 1088; @%p113 bra $L__BB6_182; and.b32 %r3229, %r259, -8; ld.u64 %rd986, [%rd159]; add.s64 %rd2307, %rd2311, %rd986; and.b32 %r1348, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1348}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1349, %r259, 6; and.b32 %r1350, %r1349, 448; shl.b32 %r1351, %r139, 3; and.b32 %r1352, %r1351, 56; or.b32 %r1353, %r1352, %r1350; and.b32 %r1354, %r74, 7; or.b32 %r1355, %r1353, %r1354; mul.wide.u32 %rd987, %r1355, 2; add.s64 %rd988, %rd2307, %rd987; ld.u16 %rs180, [%rd988+96]; cvt.rn.f32.u16 %f493, %rs180; ld.v2.f32 {%f494, %f495}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f495, %f493, %f494; mov.b64 %rd989, {%r3229, %r1348}; shr.u64 %rd990, %rd989, 32; cvt.u32.u64 %r3184, %rd990; bra.uni $L__BB6_206; $L__BB6_211: and.b32 %r372, %r259, -128; setp.ne.s32 %p131, %r372, %r3272; mov.u16 %rs375, 0; @%p131 bra $L__BB6_214; and.b32 %r1424, %r8, -128; setp.ne.s32 %p132, %r1424, %r3271; @%p132 bra $L__BB6_214; and.b32 %r1425, %r74, -128; setp.eq.s32 %p133, %r1425, %r3270; selp.u16 %rs375, 1, 0, %p133; $L__BB6_214: setp.eq.s16 %p134, %rs375, 0; @%p134 bra $L__BB6_218; shl.b32 %r1426, %r259, 5; and.b32 %r1427, %r1426, 3840; shl.b32 %r1428, %r8, 1; and.b32 %r1429, %r1428, 240; or.b32 %r1430, %r1427, %r1429; bfe.u32 %r1431, %r74, 3, 4; or.b32 %r1432, %r1431, %r1430; shr.u32 %r1433, %r1430, 6; mul.wide.u32 %rd1049, %r1433, 8; add.s64 %rd1050, %rd2311, %rd1049; and.b32 %r1434, %r1432, 63; mov.u64 %rd1051, 1; shl.b64 %rd1052, %rd1051, %r1434; ld.u64 %rd1053, [%rd1050+544]; and.b64 %rd1054, %rd1053, %rd1052; setp.eq.s64 %p135, %rd1054, 0; mul.wide.u32 %rd1055, %r1432, 8; add.s64 %rd1056, %rd2311, %rd1055; add.s64 %rd188, %rd1056, 1088; @%p135 bra $L__BB6_217; and.b32 %r3229, %r259, -8; ld.u64 %rd1057, [%rd188]; add.s64 %rd2307, %rd2311, %rd1057; and.b32 %r1435, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1435}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1436, %r259, 6; and.b32 %r1437, %r1436, 448; shl.b32 %r1438, %r8, 3; and.b32 %r1439, %r1438, 56; or.b32 %r1440, %r1437, %r1439; and.b32 %r1441, %r74, 7; or.b32 %r1442, %r1440, %r1441; mul.wide.u32 %rd1058, %r1442, 2; add.s64 %rd1059, %rd2307, %rd1058; ld.u16 %rs190, [%rd1059+96]; cvt.rn.f32.u16 %f514, %rs190; ld.v2.f32 {%f515, %f516}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f516, %f514, %f515; mov.b64 %rd1060, {%r3229, %r1435}; shr.u64 %rd1061, %rd1060, 32; cvt.u32.u64 %r3184, %rd1061; bra.uni $L__BB6_241; $L__BB6_246: and.b32 %r423, %r259, -128; setp.ne.s32 %p153, %r423, %r3272; mov.u16 %rs378, 0; @%p153 bra $L__BB6_249; and.b32 %r1510, %r8, -128; setp.ne.s32 %p154, %r1510, %r3271; @%p154 bra $L__BB6_249; and.b32 %r1511, %r9, -128; setp.eq.s32 %p155, %r1511, %r3270; selp.u16 %rs378, 1, 0, %p155; $L__BB6_249: setp.eq.s16 %p156, %rs378, 0; @%p156 bra $L__BB6_253; shl.b32 %r1512, %r259, 5; and.b32 %r1513, %r1512, 3840; shl.b32 %r1514, %r8, 1; and.b32 %r1515, %r1514, 240; or.b32 %r1516, %r1513, %r1515; bfe.u32 %r1517, %r9, 3, 4; or.b32 %r1518, %r1516, %r1517; shr.u32 %r1519, %r1516, 6; mul.wide.u32 %rd1120, %r1519, 8; add.s64 %rd1121, %rd2311, %rd1120; and.b32 %r1520, %r1518, 63; mov.u64 %rd1122, 1; shl.b64 %rd1123, %rd1122, %r1520; ld.u64 %rd1124, [%rd1121+544]; and.b64 %rd1125, %rd1124, %rd1123; setp.eq.s64 %p157, %rd1125, 0; mul.wide.u32 %rd1126, %r1518, 8; add.s64 %rd1127, %rd2311, %rd1126; add.s64 %rd217, %rd1127, 1088; @%p157 bra $L__BB6_252; and.b32 %r3229, %r259, -8; ld.u64 %rd1128, [%rd217]; add.s64 %rd2307, %rd2311, %rd1128; and.b32 %r1521, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1521}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1522, %r259, 6; and.b32 %r1523, %r1522, 448; shl.b32 %r1524, %r8, 3; and.b32 %r1525, %r1524, 56; or.b32 %r1526, %r1523, %r1525; and.b32 %r1527, %r9, 7; or.b32 %r1528, %r1526, %r1527; mul.wide.u32 %rd1129, %r1528, 2; add.s64 %rd1130, %rd2307, %rd1129; ld.u16 %rs200, [%rd1130+96]; cvt.rn.f32.u16 %f534, %rs200; ld.v2.f32 {%f535, %f536}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f536, %f534, %f535; mov.b64 %rd1131, {%r3229, %r1521}; shr.u64 %rd1132, %rd1131, 32; cvt.u32.u64 %r3184, %rd1132; bra.uni $L__BB6_276; $L__BB6_281: and.b32 %r473, %r471, -128; setp.ne.s32 %p175, %r473, %r3272; mov.u16 %rs381, 0; @%p175 bra $L__BB6_284; and.b32 %r1595, %r8, -128; setp.ne.s32 %p176, %r1595, %r3271; @%p176 bra $L__BB6_284; and.b32 %r1596, %r9, -128; setp.eq.s32 %p177, %r1596, %r3270; selp.u16 %rs381, 1, 0, %p177; $L__BB6_284: setp.eq.s16 %p178, %rs381, 0; @%p178 bra $L__BB6_288; shl.b32 %r1597, %r471, 5; and.b32 %r1598, %r1597, 3840; shl.b32 %r1599, %r8, 1; and.b32 %r1600, %r1599, 240; or.b32 %r1601, %r1600, %r1598; bfe.u32 %r1602, %r9, 3, 4; or.b32 %r1603, %r1601, %r1602; shr.u32 %r1604, %r1601, 6; mul.wide.u32 %rd1190, %r1604, 8; add.s64 %rd1191, %rd2311, %rd1190; and.b32 %r1605, %r1603, 63; mov.u64 %rd1192, 1; shl.b64 %rd1193, %rd1192, %r1605; ld.u64 %rd1194, [%rd1191+544]; and.b64 %rd1195, %rd1194, %rd1193; setp.eq.s64 %p179, %rd1195, 0; mul.wide.u32 %rd1196, %r1603, 8; add.s64 %rd1197, %rd2311, %rd1196; add.s64 %rd246, %rd1197, 1088; @%p179 bra $L__BB6_287; ld.u64 %rd1198, [%rd246]; add.s64 %rd2307, %rd2311, %rd1198; and.b32 %r1606, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1606}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1607, %r471, 6; and.b32 %r1608, %r1607, 448; shl.b32 %r1609, %r8, 3; and.b32 %r1610, %r1609, 56; or.b32 %r1611, %r1610, %r1608; and.b32 %r1612, %r9, 7; or.b32 %r1613, %r1611, %r1612; mul.wide.u32 %rd1199, %r1613, 2; add.s64 %rd1200, %rd2307, %rd1199; ld.u16 %rs210, [%rd1200+96]; cvt.rn.f32.u16 %f561, %rs210; ld.v2.f32 {%f562, %f563}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f563, %f561, %f562; mov.b64 %rd1201, {%r472, %r1606}; shr.u64 %rd1202, %rd1201, 32; cvt.u32.u64 %r3184, %rd1202; mov.u32 %r3229, %r472; bra.uni $L__BB6_311; $L__BB6_316: and.b32 %r518, %r471, -128; setp.ne.s32 %p197, %r518, %r3272; mov.u16 %rs384, 0; @%p197 bra $L__BB6_319; and.b32 %r1680, %r8, -128; setp.ne.s32 %p198, %r1680, %r3271; @%p198 bra $L__BB6_319; and.b32 %r1681, %r74, -128; setp.eq.s32 %p199, %r1681, %r3270; selp.u16 %rs384, 1, 0, %p199; $L__BB6_319: setp.eq.s16 %p200, %rs384, 0; @%p200 bra $L__BB6_323; shl.b32 %r1682, %r471, 5; and.b32 %r1683, %r1682, 3840; shl.b32 %r1684, %r8, 1; and.b32 %r1685, %r1684, 240; or.b32 %r1686, %r1685, %r1683; bfe.u32 %r1687, %r74, 3, 4; or.b32 %r1688, %r1687, %r1686; shr.u32 %r1689, %r1686, 6; mul.wide.u32 %rd1260, %r1689, 8; add.s64 %rd1261, %rd2311, %rd1260; and.b32 %r1690, %r1688, 63; mov.u64 %rd1262, 1; shl.b64 %rd1263, %rd1262, %r1690; ld.u64 %rd1264, [%rd1261+544]; and.b64 %rd1265, %rd1264, %rd1263; setp.eq.s64 %p201, %rd1265, 0; mul.wide.u32 %rd1266, %r1688, 8; add.s64 %rd1267, %rd2311, %rd1266; add.s64 %rd275, %rd1267, 1088; @%p201 bra $L__BB6_322; ld.u64 %rd1268, [%rd275]; add.s64 %rd2307, %rd2311, %rd1268; and.b32 %r1691, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1691}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1692, %r471, 6; and.b32 %r1693, %r1692, 448; shl.b32 %r1694, %r8, 3; and.b32 %r1695, %r1694, 56; or.b32 %r1696, %r1695, %r1693; and.b32 %r1697, %r74, 7; or.b32 %r1698, %r1696, %r1697; mul.wide.u32 %rd1269, %r1698, 2; add.s64 %rd1270, %rd2307, %rd1269; ld.u16 %rs220, [%rd1270+96]; cvt.rn.f32.u16 %f581, %rs220; ld.v2.f32 {%f582, %f583}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f583, %f581, %f582; mov.b64 %rd1271, {%r472, %r1691}; shr.u64 %rd1272, %rd1271, 32; cvt.u32.u64 %r3184, %rd1272; mov.u32 %r3229, %r472; bra.uni $L__BB6_346; $L__BB6_351: and.b32 %r560, %r471, -128; setp.ne.s32 %p219, %r560, %r3272; mov.u16 %rs387, 0; @%p219 bra $L__BB6_354; and.b32 %r1766, %r139, -128; setp.ne.s32 %p220, %r1766, %r3271; @%p220 bra $L__BB6_354; and.b32 %r1767, %r74, -128; setp.eq.s32 %p221, %r1767, %r3270; selp.u16 %rs387, 1, 0, %p221; $L__BB6_354: setp.eq.s16 %p222, %rs387, 0; @%p222 bra $L__BB6_358; shl.b32 %r1768, %r471, 5; and.b32 %r1769, %r1768, 3840; shl.b32 %r1770, %r139, 1; and.b32 %r1771, %r1770, 240; or.b32 %r1772, %r1771, %r1769; bfe.u32 %r1773, %r74, 3, 4; or.b32 %r1774, %r1772, %r1773; shr.u32 %r1775, %r1772, 6; mul.wide.u32 %rd1331, %r1775, 8; add.s64 %rd1332, %rd2311, %rd1331; and.b32 %r1776, %r1774, 63; mov.u64 %rd1333, 1; shl.b64 %rd1334, %rd1333, %r1776; ld.u64 %rd1335, [%rd1332+544]; and.b64 %rd1336, %rd1335, %rd1334; setp.eq.s64 %p223, %rd1336, 0; mul.wide.u32 %rd1337, %r1774, 8; add.s64 %rd1338, %rd2311, %rd1337; add.s64 %rd304, %rd1338, 1088; @%p223 bra $L__BB6_357; ld.u64 %rd1339, [%rd304]; add.s64 %rd2307, %rd2311, %rd1339; and.b32 %r1777, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1777}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1778, %r471, 6; and.b32 %r1779, %r1778, 448; shl.b32 %r1780, %r139, 3; and.b32 %r1781, %r1780, 56; or.b32 %r1782, %r1781, %r1779; and.b32 %r1783, %r74, 7; or.b32 %r1784, %r1782, %r1783; mul.wide.u32 %rd1340, %r1784, 2; add.s64 %rd1341, %rd2307, %rd1340; ld.u16 %rs230, [%rd1341+96]; cvt.rn.f32.u16 %f602, %rs230; ld.v2.f32 {%f603, %f604}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f604, %f602, %f603; mov.b64 %rd1342, {%r472, %r1777}; shr.u64 %rd1343, %rd1342, 32; cvt.u32.u64 %r3184, %rd1343; mov.u32 %r3229, %r472; bra.uni $L__BB6_381; $L__BB6_386: and.b32 %r599, %r471, -128; setp.ne.s32 %p241, %r599, %r3272; mov.u16 %rs390, 0; @%p241 bra $L__BB6_389; and.b32 %r1852, %r139, -128; setp.ne.s32 %p242, %r1852, %r3271; @%p242 bra $L__BB6_389; and.b32 %r1853, %r9, -128; setp.eq.s32 %p243, %r1853, %r3270; selp.u16 %rs390, 1, 0, %p243; $L__BB6_389: setp.eq.s16 %p244, %rs390, 0; @%p244 bra $L__BB6_393; shl.b32 %r1854, %r471, 5; and.b32 %r1855, %r1854, 3840; shl.b32 %r1856, %r139, 1; and.b32 %r1857, %r1856, 240; or.b32 %r1858, %r1857, %r1855; bfe.u32 %r1859, %r9, 3, 4; or.b32 %r1860, %r1858, %r1859; shr.u32 %r1861, %r1858, 6; mul.wide.u32 %rd1402, %r1861, 8; add.s64 %rd1403, %rd2311, %rd1402; and.b32 %r1862, %r1860, 63; mov.u64 %rd1404, 1; shl.b64 %rd1405, %rd1404, %r1862; ld.u64 %rd1406, [%rd1403+544]; and.b64 %rd1407, %rd1406, %rd1405; setp.eq.s64 %p245, %rd1407, 0; mul.wide.u32 %rd1408, %r1860, 8; add.s64 %rd1409, %rd2311, %rd1408; add.s64 %rd333, %rd1409, 1088; @%p245 bra $L__BB6_392; ld.u64 %rd1410, [%rd333]; add.s64 %rd2307, %rd2311, %rd1410; and.b32 %r1863, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1863}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1864, %r471, 7; shl.b32 %r1865, %r139, 3; and.b32 %r1866, %r1865, 56; and.b32 %r1867, %r9, 7; bfi.b32 %r1868, %r1864, %r1867, 6, 3; or.b32 %r1869, %r1868, %r1866; mul.wide.u32 %rd1411, %r1869, 2; add.s64 %rd1412, %rd2307, %rd1411; ld.u16 %rs240, [%rd1412+96]; cvt.rn.f32.u16 %f622, %rs240; ld.v2.f32 {%f623, %f624}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f624, %f622, %f623; mov.b64 %rd1413, {%r472, %r1863}; shr.u64 %rd1414, %rd1413, 32; cvt.u32.u64 %r3184, %rd1414; mov.u32 %r3229, %r472; bra.uni $L__BB6_416; $L__BB6_421: and.b32 %r637, %r635, -128; setp.ne.s32 %p263, %r637, %r3272; mov.u16 %rs393, 0; @%p263 bra $L__BB6_424; and.b32 %r1934, %r139, -128; setp.ne.s32 %p264, %r1934, %r3271; @%p264 bra $L__BB6_424; and.b32 %r1935, %r9, -128; setp.eq.s32 %p265, %r1935, %r3270; selp.u16 %rs393, 1, 0, %p265; $L__BB6_424: setp.eq.s16 %p266, %rs393, 0; @%p266 bra $L__BB6_428; shl.b32 %r1936, %r635, 5; and.b32 %r1937, %r1936, 3840; shl.b32 %r1938, %r139, 1; and.b32 %r1939, %r1938, 240; or.b32 %r1940, %r1939, %r1937; bfe.u32 %r1941, %r9, 3, 4; or.b32 %r1942, %r1940, %r1941; shr.u32 %r1943, %r1940, 6; mul.wide.u32 %rd1473, %r1943, 8; add.s64 %rd1474, %rd2311, %rd1473; and.b32 %r1944, %r1942, 63; mov.u64 %rd1475, 1; shl.b64 %rd1476, %rd1475, %r1944; ld.u64 %rd1477, [%rd1474+544]; and.b64 %rd1478, %rd1477, %rd1476; setp.eq.s64 %p267, %rd1478, 0; mul.wide.u32 %rd1479, %r1942, 8; add.s64 %rd1480, %rd2311, %rd1479; add.s64 %rd362, %rd1480, 1088; @%p267 bra $L__BB6_427; ld.u64 %rd1481, [%rd362]; add.s64 %rd2307, %rd2311, %rd1481; and.b32 %r1945, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1945}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1946, %r635, 7; shl.b32 %r1947, %r139, 3; and.b32 %r1948, %r1947, 56; and.b32 %r1949, %r9, 7; bfi.b32 %r1950, %r1946, %r1949, 6, 3; or.b32 %r1951, %r1950, %r1948; mul.wide.u32 %rd1482, %r1951, 2; add.s64 %rd1483, %rd2307, %rd1482; ld.u16 %rs250, [%rd1483+96]; cvt.rn.f32.u16 %f645, %rs250; ld.v2.f32 {%f646, %f647}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f647, %f645, %f646; mov.b64 %rd1484, {%r636, %r1945}; shr.u64 %rd1485, %rd1484, 32; cvt.u32.u64 %r3184, %rd1485; mov.u32 %r3229, %r636; bra.uni $L__BB6_451; $L__BB6_456: and.b32 %r670, %r635, -128; setp.ne.s32 %p285, %r670, %r3272; mov.u16 %rs396, 0; @%p285 bra $L__BB6_459; and.b32 %r2017, %r139, -128; setp.ne.s32 %p286, %r2017, %r3271; @%p286 bra $L__BB6_459; and.b32 %r2018, %r74, -128; setp.eq.s32 %p287, %r2018, %r3270; selp.u16 %rs396, 1, 0, %p287; $L__BB6_459: setp.eq.s16 %p288, %rs396, 0; @%p288 bra $L__BB6_463; shl.b32 %r2019, %r635, 5; and.b32 %r2020, %r2019, 3840; shl.b32 %r2021, %r139, 1; and.b32 %r2022, %r2021, 240; or.b32 %r2023, %r2022, %r2020; bfe.u32 %r2024, %r74, 3, 4; or.b32 %r2025, %r2023, %r2024; shr.u32 %r2026, %r2023, 6; mul.wide.u32 %rd1544, %r2026, 8; add.s64 %rd1545, %rd2311, %rd1544; and.b32 %r2027, %r2025, 63; mov.u64 %rd1546, 1; shl.b64 %rd1547, %rd1546, %r2027; ld.u64 %rd1548, [%rd1545+544]; and.b64 %rd1549, %rd1548, %rd1547; setp.eq.s64 %p289, %rd1549, 0; mul.wide.u32 %rd1550, %r2025, 8; add.s64 %rd1551, %rd2311, %rd1550; add.s64 %rd391, %rd1551, 1088; @%p289 bra $L__BB6_462; ld.u64 %rd1552, [%rd391]; add.s64 %rd2307, %rd2311, %rd1552; and.b32 %r2028, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2028}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2029, %r635, 6; and.b32 %r2030, %r2029, 448; shl.b32 %r2031, %r139, 3; and.b32 %r2032, %r2031, 56; or.b32 %r2033, %r2032, %r2030; and.b32 %r2034, %r74, 7; or.b32 %r2035, %r2033, %r2034; mul.wide.u32 %rd1553, %r2035, 2; add.s64 %rd1554, %rd2307, %rd1553; ld.u16 %rs260, [%rd1554+96]; cvt.rn.f32.u16 %f665, %rs260; ld.v2.f32 {%f666, %f667}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f667, %f665, %f666; mov.b64 %rd1555, {%r636, %r2028}; shr.u64 %rd1556, %rd1555, 32; cvt.u32.u64 %r3184, %rd1556; mov.u32 %r3229, %r636; bra.uni $L__BB6_486; $L__BB6_491: and.b32 %r700, %r635, -128; setp.ne.s32 %p307, %r700, %r3272; mov.u16 %rs399, 0; @%p307 bra $L__BB6_494; and.b32 %r2104, %r8, -128; setp.ne.s32 %p308, %r2104, %r3271; @%p308 bra $L__BB6_494; and.b32 %r2105, %r74, -128; setp.eq.s32 %p309, %r2105, %r3270; selp.u16 %rs399, 1, 0, %p309; $L__BB6_494: setp.eq.s16 %p310, %rs399, 0; @%p310 bra $L__BB6_498; shl.b32 %r2106, %r635, 5; and.b32 %r2107, %r2106, 3840; shl.b32 %r2108, %r8, 1; and.b32 %r2109, %r2108, 240; or.b32 %r2110, %r2107, %r2109; bfe.u32 %r2111, %r74, 3, 4; or.b32 %r2112, %r2111, %r2110; shr.u32 %r2113, %r2110, 6; mul.wide.u32 %rd1615, %r2113, 8; add.s64 %rd1616, %rd2311, %rd1615; and.b32 %r2114, %r2112, 63; mov.u64 %rd1617, 1; shl.b64 %rd1618, %rd1617, %r2114; ld.u64 %rd1619, [%rd1616+544]; and.b64 %rd1620, %rd1619, %rd1618; setp.eq.s64 %p311, %rd1620, 0; mul.wide.u32 %rd1621, %r2112, 8; add.s64 %rd1622, %rd2311, %rd1621; add.s64 %rd420, %rd1622, 1088; @%p311 bra $L__BB6_497; ld.u64 %rd1623, [%rd420]; add.s64 %rd2307, %rd2311, %rd1623; and.b32 %r2115, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2115}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2116, %r635, 6; and.b32 %r2117, %r2116, 448; shl.b32 %r2118, %r8, 3; and.b32 %r2119, %r2118, 56; or.b32 %r2120, %r2117, %r2119; and.b32 %r2121, %r74, 7; or.b32 %r2122, %r2120, %r2121; mul.wide.u32 %rd1624, %r2122, 2; add.s64 %rd1625, %rd2307, %rd1624; ld.u16 %rs270, [%rd1625+96]; cvt.rn.f32.u16 %f686, %rs270; ld.v2.f32 {%f687, %f688}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f688, %f686, %f687; mov.b64 %rd1626, {%r636, %r2115}; shr.u64 %rd1627, %rd1626, 32; cvt.u32.u64 %r3184, %rd1627; mov.u32 %r3229, %r636; bra.uni $L__BB6_521; $L__BB6_526: and.b32 %r727, %r635, -128; setp.ne.s32 %p329, %r727, %r3272; mov.u16 %rs402, 0; @%p329 bra $L__BB6_529; and.b32 %r2190, %r8, -128; setp.ne.s32 %p330, %r2190, %r3271; @%p330 bra $L__BB6_529; and.b32 %r2191, %r9, -128; setp.eq.s32 %p331, %r2191, %r3270; selp.u16 %rs402, 1, 0, %p331; $L__BB6_529: setp.eq.s16 %p332, %rs402, 0; @%p332 bra $L__BB6_533; shl.b32 %r2192, %r635, 5; and.b32 %r2193, %r2192, 3840; shl.b32 %r2194, %r8, 1; and.b32 %r2195, %r2194, 240; or.b32 %r2196, %r2193, %r2195; bfe.u32 %r2197, %r9, 3, 4; or.b32 %r2198, %r2196, %r2197; shr.u32 %r2199, %r2196, 6; mul.wide.u32 %rd1686, %r2199, 8; add.s64 %rd1687, %rd2311, %rd1686; and.b32 %r2200, %r2198, 63; mov.u64 %rd1688, 1; shl.b64 %rd1689, %rd1688, %r2200; ld.u64 %rd1690, [%rd1687+544]; and.b64 %rd1691, %rd1690, %rd1689; setp.eq.s64 %p333, %rd1691, 0; mul.wide.u32 %rd1692, %r2198, 8; add.s64 %rd1693, %rd2311, %rd1692; add.s64 %rd449, %rd1693, 1088; @%p333 bra $L__BB6_532; ld.u64 %rd1694, [%rd449]; add.s64 %rd1695, %rd2311, %rd1694; and.b32 %r2201, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2201}; and.b32 %r2202, %r9, -8; st.local.u32 [%rd2+8], %r2202; st.local.u64 [%rd2+48], %rd1695; shl.b32 %r2203, %r635, 6; and.b32 %r2204, %r2203, 448; shl.b32 %r2205, %r8, 3; and.b32 %r2206, %r2205, 56; or.b32 %r2207, %r2204, %r2206; and.b32 %r2208, %r9, 7; or.b32 %r2209, %r2207, %r2208; mul.wide.u32 %rd1696, %r2209, 2; add.s64 %rd1697, %rd1695, %rd1696; ld.u16 %rs280, [%rd1697+96]; cvt.rn.f32.u16 %f706, %rs280; ld.v2.f32 {%f707, %f708}, [%rd1695+80]; fma.rn.ftz.f32 %f944, %f708, %f706, %f707; bra.uni $L__BB6_556; $L__BB6_561: and.b32 %r735, %r471, -128; ld.local.u32 %r2285, [%rd2+12]; setp.ne.s32 %p351, %r735, %r2285; mov.u16 %rs405, 0; @%p351 bra $L__BB6_564; and.b32 %r2286, %r732, -128; ld.local.u32 %r2287, [%rd2+16]; setp.ne.s32 %p352, %r2286, %r2287; @%p352 bra $L__BB6_564; and.b32 %r2288, %r733, -128; ld.local.u32 %r2289, [%rd2+20]; setp.eq.s32 %p353, %r2288, %r2289; selp.u16 %rs405, 1, 0, %p353; $L__BB6_564: setp.eq.s16 %p354, %rs405, 0; @%p354 bra $L__BB6_568; shl.b32 %r2290, %r471, 5; and.b32 %r2291, %r2290, 3840; shl.b32 %r2292, %r732, 1; and.b32 %r2293, %r2292, 240; or.b32 %r2294, %r2293, %r2291; bfe.u32 %r2295, %r733, 3, 4; or.b32 %r2296, %r2294, %r2295; shr.u32 %r2297, %r2294, 6; ld.local.u64 %rd464, [%rd2+56]; mul.wide.u32 %rd1752, %r2297, 8; add.s64 %rd1753, %rd464, %rd1752; and.b32 %r2298, %r2296, 63; mov.u64 %rd1754, 1; shl.b64 %rd1755, %rd1754, %r2298; ld.u64 %rd1756, [%rd1753+544]; and.b64 %rd1757, %rd1756, %rd1755; setp.eq.s64 %p355, %rd1757, 0; mul.wide.u32 %rd1758, %r2296, 8; add.s64 %rd1759, %rd464, %rd1758; add.s64 %rd465, %rd1759, 1088; @%p355 bra $L__BB6_567; ld.u64 %rd1760, [%rd465]; add.s64 %rd1761, %rd464, %rd1760; and.b32 %r2299, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2299}; and.b32 %r2300, %r733, -8; st.local.u32 [%rd2+8], %r2300; st.local.u64 [%rd2+48], %rd1761; shl.b32 %r2301, %r471, 6; and.b32 %r2302, %r2301, 448; shl.b32 %r2303, %r732, 3; and.b32 %r2304, %r2303, 56; or.b32 %r2305, %r2304, %r2302; and.b32 %r2306, %r733, 7; or.b32 %r2307, %r2305, %r2306; mul.wide.u32 %rd1762, %r2307, 2; add.s64 %rd1763, %rd1761, %rd1762; ld.u16 %rs290, [%rd1763+96]; cvt.rn.f32.u16 %f747, %rs290; ld.v2.f32 {%f748, %f749}, [%rd1761+80]; fma.rn.ftz.f32 %f945, %f749, %f747, %f748; mov.u32 %r3598, %r472; bra.uni $L__BB6_591; $L__BB6_596: and.b32 %r742, %r471, -128; ld.local.u32 %r2388, [%rd2+12]; setp.ne.s32 %p373, %r742, %r2388; mov.u16 %rs408, 0; @%p373 bra $L__BB6_599; and.b32 %r2389, %r732, -128; ld.local.u32 %r2390, [%rd2+16]; setp.ne.s32 %p374, %r2389, %r2390; @%p374 bra $L__BB6_599; and.b32 %r2391, %r741, -128; ld.local.u32 %r2392, [%rd2+20]; setp.eq.s32 %p375, %r2391, %r2392; selp.u16 %rs408, 1, 0, %p375; $L__BB6_599: setp.eq.s16 %p376, %rs408, 0; @%p376 bra $L__BB6_603; shl.b32 %r2393, %r471, 5; and.b32 %r2394, %r2393, 3840; shl.b32 %r2395, %r732, 1; and.b32 %r2396, %r2395, 240; or.b32 %r2397, %r2396, %r2394; bfe.u32 %r2398, %r741, 3, 4; or.b32 %r2399, %r2398, %r2397; shr.u32 %r2400, %r2397, 6; ld.local.u64 %rd482, [%rd2+56]; mul.wide.u32 %rd1819, %r2400, 8; add.s64 %rd1820, %rd482, %rd1819; and.b32 %r2401, %r2399, 63; mov.u64 %rd1821, 1; shl.b64 %rd1822, %rd1821, %r2401; ld.u64 %rd1823, [%rd1820+544]; and.b64 %rd1824, %rd1823, %rd1822; setp.eq.s64 %p377, %rd1824, 0; mul.wide.u32 %rd1825, %r2399, 8; add.s64 %rd1826, %rd482, %rd1825; add.s64 %rd483, %rd1826, 1088; @%p377 bra $L__BB6_602; ld.u64 %rd1827, [%rd483]; add.s64 %rd1828, %rd482, %rd1827; and.b32 %r2402, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2402}; and.b32 %r2403, %r741, -8; st.local.u32 [%rd2+8], %r2403; st.local.u64 [%rd2+48], %rd1828; shl.b32 %r2404, %r471, 6; and.b32 %r2405, %r2404, 448; shl.b32 %r2406, %r732, 3; and.b32 %r2407, %r2406, 56; or.b32 %r2408, %r2407, %r2405; and.b32 %r2409, %r741, 7; or.b32 %r2410, %r2408, %r2409; mul.wide.u32 %rd1829, %r2410, 2; add.s64 %rd1830, %rd1828, %rd1829; ld.u16 %rs300, [%rd1830+96]; cvt.rn.f32.u16 %f767, %rs300; ld.v2.f32 {%f768, %f769}, [%rd1828+80]; fma.rn.ftz.f32 %f946, %f769, %f767, %f768; mov.u32 %r3598, %r472; bra.uni $L__BB6_626; $L__BB6_631: and.b32 %r749, %r471, -128; ld.local.u32 %r2491, [%rd2+12]; setp.ne.s32 %p395, %r749, %r2491; mov.u16 %rs411, 0; @%p395 bra $L__BB6_634; and.b32 %r2492, %r748, -128; ld.local.u32 %r2493, [%rd2+16]; setp.ne.s32 %p396, %r2492, %r2493; @%p396 bra $L__BB6_634; and.b32 %r2494, %r741, -128; ld.local.u32 %r2495, [%rd2+20]; setp.eq.s32 %p397, %r2494, %r2495; selp.u16 %rs411, 1, 0, %p397; $L__BB6_634: setp.eq.s16 %p398, %rs411, 0; @%p398 bra $L__BB6_638; shl.b32 %r2496, %r471, 5; and.b32 %r2497, %r2496, 3840; shl.b32 %r2498, %r748, 1; and.b32 %r2499, %r2498, 240; or.b32 %r2500, %r2499, %r2497; bfe.u32 %r2501, %r741, 3, 4; or.b32 %r2502, %r2500, %r2501; shr.u32 %r2503, %r2500, 6; ld.local.u64 %rd500, [%rd2+56]; mul.wide.u32 %rd1887, %r2503, 8; add.s64 %rd1888, %rd500, %rd1887; and.b32 %r2504, %r2502, 63; mov.u64 %rd1889, 1; shl.b64 %rd1890, %rd1889, %r2504; ld.u64 %rd1891, [%rd1888+544]; and.b64 %rd1892, %rd1891, %rd1890; setp.eq.s64 %p399, %rd1892, 0; mul.wide.u32 %rd1893, %r2502, 8; add.s64 %rd1894, %rd500, %rd1893; add.s64 %rd501, %rd1894, 1088; @%p399 bra $L__BB6_637; ld.u64 %rd1895, [%rd501]; add.s64 %rd1896, %rd500, %rd1895; and.b32 %r2505, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2505}; and.b32 %r2506, %r741, -8; st.local.u32 [%rd2+8], %r2506; st.local.u64 [%rd2+48], %rd1896; shl.b32 %r2507, %r471, 6; and.b32 %r2508, %r2507, 448; shl.b32 %r2509, %r748, 3; and.b32 %r2510, %r2509, 56; or.b32 %r2511, %r2510, %r2508; and.b32 %r2512, %r741, 7; or.b32 %r2513, %r2511, %r2512; mul.wide.u32 %rd1897, %r2513, 2; add.s64 %rd1898, %rd1896, %rd1897; ld.u16 %rs310, [%rd1898+96]; cvt.rn.f32.u16 %f788, %rs310; ld.v2.f32 {%f789, %f790}, [%rd1896+80]; fma.rn.ftz.f32 %f947, %f790, %f788, %f789; mov.u32 %r3598, %r472; bra.uni $L__BB6_661; $L__BB6_666: and.b32 %r755, %r471, -128; ld.local.u32 %r2593, [%rd2+12]; setp.ne.s32 %p417, %r755, %r2593; mov.u16 %rs414, 0; @%p417 bra $L__BB6_669; and.b32 %r2594, %r748, -128; ld.local.u32 %r2595, [%rd2+16]; setp.ne.s32 %p418, %r2594, %r2595; @%p418 bra $L__BB6_669; and.b32 %r2596, %r733, -128; ld.local.u32 %r2597, [%rd2+20]; setp.eq.s32 %p419, %r2596, %r2597; selp.u16 %rs414, 1, 0, %p419; $L__BB6_669: setp.eq.s16 %p420, %rs414, 0; @%p420 bra $L__BB6_673; shl.b32 %r2598, %r471, 5; and.b32 %r2599, %r2598, 3840; shl.b32 %r2600, %r748, 1; and.b32 %r2601, %r2600, 240; or.b32 %r2602, %r2601, %r2599; bfe.u32 %r2603, %r733, 3, 4; or.b32 %r2604, %r2602, %r2603; shr.u32 %r2605, %r2602, 6; ld.local.u64 %rd518, [%rd2+56]; mul.wide.u32 %rd1954, %r2605, 8; add.s64 %rd1955, %rd518, %rd1954; and.b32 %r2606, %r2604, 63; mov.u64 %rd1956, 1; shl.b64 %rd1957, %rd1956, %r2606; ld.u64 %rd1958, [%rd1955+544]; and.b64 %rd1959, %rd1958, %rd1957; setp.eq.s64 %p421, %rd1959, 0; mul.wide.u32 %rd1960, %r2604, 8; add.s64 %rd1961, %rd518, %rd1960; add.s64 %rd519, %rd1961, 1088; @%p421 bra $L__BB6_672; ld.u64 %rd1962, [%rd519]; add.s64 %rd1963, %rd518, %rd1962; and.b32 %r2607, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2607}; and.b32 %r2608, %r733, -8; st.local.u32 [%rd2+8], %r2608; st.local.u64 [%rd2+48], %rd1963; and.b32 %r2609, %r471, 7; shl.b32 %r2610, %r748, 3; and.b32 %r2611, %r2610, 56; and.b32 %r2612, %r733, 7; bfi.b32 %r2613, %r2609, %r2612, 6, 3; or.b32 %r2614, %r2613, %r2611; mul.wide.u32 %rd1964, %r2614, 2; add.s64 %rd1965, %rd1963, %rd1964; ld.u16 %rs320, [%rd1965+96]; cvt.rn.f32.u16 %f808, %rs320; ld.v2.f32 {%f809, %f810}, [%rd1963+80]; fma.rn.ftz.f32 %f948, %f810, %f808, %f809; mov.u32 %r3598, %r472; bra.uni $L__BB6_696; $L__BB6_701: and.b32 %r761, %r635, -128; ld.local.u32 %r2692, [%rd2+12]; setp.ne.s32 %p439, %r761, %r2692; mov.u16 %rs417, 0; @%p439 bra $L__BB6_704; and.b32 %r2693, %r748, -128; ld.local.u32 %r2694, [%rd2+16]; setp.ne.s32 %p440, %r2693, %r2694; @%p440 bra $L__BB6_704; and.b32 %r2695, %r733, -128; ld.local.u32 %r2696, [%rd2+20]; setp.eq.s32 %p441, %r2695, %r2696; selp.u16 %rs417, 1, 0, %p441; $L__BB6_704: setp.eq.s16 %p442, %rs417, 0; @%p442 bra $L__BB6_708; shl.b32 %r2697, %r635, 5; and.b32 %r2698, %r2697, 3840; shl.b32 %r2699, %r748, 1; and.b32 %r2700, %r2699, 240; or.b32 %r2701, %r2700, %r2698; bfe.u32 %r2702, %r733, 3, 4; or.b32 %r2703, %r2701, %r2702; shr.u32 %r2704, %r2701, 6; ld.local.u64 %rd536, [%rd2+56]; mul.wide.u32 %rd2021, %r2704, 8; add.s64 %rd2022, %rd536, %rd2021; and.b32 %r2705, %r2703, 63; mov.u64 %rd2023, 1; shl.b64 %rd2024, %rd2023, %r2705; ld.u64 %rd2025, [%rd2022+544]; and.b64 %rd2026, %rd2025, %rd2024; setp.eq.s64 %p443, %rd2026, 0; mul.wide.u32 %rd2027, %r2703, 8; add.s64 %rd2028, %rd536, %rd2027; add.s64 %rd537, %rd2028, 1088; @%p443 bra $L__BB6_707; ld.u64 %rd2029, [%rd537]; add.s64 %rd2030, %rd536, %rd2029; and.b32 %r2706, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2706}; and.b32 %r2707, %r733, -8; st.local.u32 [%rd2+8], %r2707; st.local.u64 [%rd2+48], %rd2030; and.b32 %r2708, %r635, 7; shl.b32 %r2709, %r748, 3; and.b32 %r2710, %r2709, 56; and.b32 %r2711, %r733, 7; bfi.b32 %r2712, %r2708, %r2711, 6, 3; or.b32 %r2713, %r2712, %r2710; mul.wide.u32 %rd2031, %r2713, 2; add.s64 %rd2032, %rd2030, %rd2031; ld.u16 %rs330, [%rd2032+96]; cvt.rn.f32.u16 %f831, %rs330; ld.v2.f32 {%f832, %f833}, [%rd2030+80]; fma.rn.ftz.f32 %f949, %f833, %f831, %f832; mov.u32 %r3598, %r636; bra.uni $L__BB6_731; $L__BB6_736: and.b32 %r767, %r635, -128; ld.local.u32 %r2792, [%rd2+12]; setp.ne.s32 %p461, %r767, %r2792; mov.u16 %rs420, 0; @%p461 bra $L__BB6_739; and.b32 %r2793, %r748, -128; ld.local.u32 %r2794, [%rd2+16]; setp.ne.s32 %p462, %r2793, %r2794; @%p462 bra $L__BB6_739; and.b32 %r2795, %r741, -128; ld.local.u32 %r2796, [%rd2+20]; setp.eq.s32 %p463, %r2795, %r2796; selp.u16 %rs420, 1, 0, %p463; $L__BB6_739: setp.eq.s16 %p464, %rs420, 0; @%p464 bra $L__BB6_743; shl.b32 %r2797, %r635, 5; and.b32 %r2798, %r2797, 3840; shl.b32 %r2799, %r748, 1; and.b32 %r2800, %r2799, 240; or.b32 %r2801, %r2800, %r2798; bfe.u32 %r2802, %r741, 3, 4; or.b32 %r2803, %r2801, %r2802; shr.u32 %r2804, %r2801, 6; ld.local.u64 %rd554, [%rd2+56]; mul.wide.u32 %rd2088, %r2804, 8; add.s64 %rd2089, %rd554, %rd2088; and.b32 %r2805, %r2803, 63; mov.u64 %rd2090, 1; shl.b64 %rd2091, %rd2090, %r2805; ld.u64 %rd2092, [%rd2089+544]; and.b64 %rd2093, %rd2092, %rd2091; setp.eq.s64 %p465, %rd2093, 0; mul.wide.u32 %rd2094, %r2803, 8; add.s64 %rd2095, %rd554, %rd2094; add.s64 %rd555, %rd2095, 1088; @%p465 bra $L__BB6_742; ld.u64 %rd2096, [%rd555]; add.s64 %rd2097, %rd554, %rd2096; and.b32 %r2806, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2806}; and.b32 %r2807, %r741, -8; st.local.u32 [%rd2+8], %r2807; st.local.u64 [%rd2+48], %rd2097; shl.b32 %r2808, %r635, 6; and.b32 %r2809, %r2808, 448; shl.b32 %r2810, %r748, 3; and.b32 %r2811, %r2810, 56; or.b32 %r2812, %r2811, %r2809; and.b32 %r2813, %r741, 7; or.b32 %r2814, %r2812, %r2813; mul.wide.u32 %rd2098, %r2814, 2; add.s64 %rd2099, %rd2097, %rd2098; ld.u16 %rs340, [%rd2099+96]; cvt.rn.f32.u16 %f851, %rs340; ld.v2.f32 {%f852, %f853}, [%rd2097+80]; fma.rn.ftz.f32 %f950, %f853, %f851, %f852; mov.u32 %r3598, %r636; bra.uni $L__BB6_766; $L__BB6_771: and.b32 %r773, %r635, -128; ld.local.u32 %r2895, [%rd2+12]; setp.ne.s32 %p483, %r773, %r2895; mov.u16 %rs423, 0; @%p483 bra $L__BB6_774; and.b32 %r2896, %r732, -128; ld.local.u32 %r2897, [%rd2+16]; setp.ne.s32 %p484, %r2896, %r2897; @%p484 bra $L__BB6_774; and.b32 %r2898, %r741, -128; ld.local.u32 %r2899, [%rd2+20]; setp.eq.s32 %p485, %r2898, %r2899; selp.u16 %rs423, 1, 0, %p485; $L__BB6_774: setp.eq.s16 %p486, %rs423, 0; @%p486 bra $L__BB6_778; shl.b32 %r2900, %r635, 5; and.b32 %r2901, %r2900, 3840; shl.b32 %r2902, %r732, 1; and.b32 %r2903, %r2902, 240; or.b32 %r2904, %r2901, %r2903; bfe.u32 %r2905, %r741, 3, 4; or.b32 %r2906, %r2905, %r2904; shr.u32 %r2907, %r2904, 6; ld.local.u64 %rd572, [%rd2+56]; mul.wide.u32 %rd2155, %r2907, 8; add.s64 %rd2156, %rd572, %rd2155; and.b32 %r2908, %r2906, 63; mov.u64 %rd2157, 1; shl.b64 %rd2158, %rd2157, %r2908; ld.u64 %rd2159, [%rd2156+544]; and.b64 %rd2160, %rd2159, %rd2158; setp.eq.s64 %p487, %rd2160, 0; mul.wide.u32 %rd2161, %r2906, 8; add.s64 %rd2162, %rd572, %rd2161; add.s64 %rd573, %rd2162, 1088; @%p487 bra $L__BB6_777; ld.u64 %rd2163, [%rd573]; add.s64 %rd2164, %rd572, %rd2163; and.b32 %r2909, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2909}; and.b32 %r2910, %r741, -8; st.local.u32 [%rd2+8], %r2910; st.local.u64 [%rd2+48], %rd2164; shl.b32 %r2911, %r635, 6; and.b32 %r2912, %r2911, 448; shl.b32 %r2913, %r732, 3; and.b32 %r2914, %r2913, 56; or.b32 %r2915, %r2912, %r2914; and.b32 %r2916, %r741, 7; or.b32 %r2917, %r2915, %r2916; mul.wide.u32 %rd2165, %r2917, 2; add.s64 %rd2166, %rd2164, %rd2165; ld.u16 %rs350, [%rd2166+96]; cvt.rn.f32.u16 %f872, %rs350; ld.v2.f32 {%f873, %f874}, [%rd2164+80]; fma.rn.ftz.f32 %f951, %f874, %f872, %f873; mov.u32 %r3598, %r636; bra.uni $L__BB6_801; $L__BB6_814: shl.b64 %rd2230, %rd593, 5; add.s64 %rd2394, %rd591, %rd2230; $L__BB6_815: setp.eq.s64 %p513, %rd2394, 0; @%p513 bra $L__BB6_823; add.s64 %rd596, %rd2394, 8; ld.u64 %rd597, [%rd2394+8]; setp.eq.s64 %p514, %rd597, 0; @%p514 bra $L__BB6_822; add.s64 %rd598, %rd590, %rd597; and.b32 %r3005, %r732, -4096; mov.u64 %rd2231, 1; st.local.v2.u32 [%rd2+24], {%r780, %r3005}; and.b32 %r3006, %r733, -4096; st.local.u32 [%rd2+32], %r3006; st.local.u64 [%rd2+64], %rd598; shl.b32 %r3007, %r635, 3; and.b32 %r3008, %r3007, 31744; and.b32 %r3009, %r732, 3968; shr.u32 %r3010, %r3009, 2; or.b32 %r3011, %r3008, %r3010; bfe.u32 %r3012, %r733, 7, 5; or.b32 %r3013, %r3011, %r3012; shr.u32 %r3014, %r3011, 6; mul.wide.u32 %rd2232, %r3014, 8; add.s64 %rd2233, %rd598, %rd2232; and.b32 %r3015, %r3013, 63; shl.b64 %rd2234, %rd2231, %r3015; ld.u64 %rd2235, [%rd2233+4128]; and.b64 %rd2236, %rd2235, %rd2234; setp.eq.s64 %p515, %rd2236, 0; mul.wide.u32 %rd2237, %r3013, 8; add.s64 %rd2238, %rd598, %rd2237; add.s64 %rd599, %rd2238, 8256; @%p515 bra $L__BB6_821; ld.u64 %rd2239, [%rd599]; add.s64 %rd600, %rd598, %rd2239; st.local.u32 [%rd2+12], %r779; and.b32 %r3016, %r732, -128; st.local.u32 [%rd2+16], %r3016; and.b32 %r3017, %r733, -128; st.local.u32 [%rd2+20], %r3017; st.local.u64 [%rd2+56], %rd600; shl.b32 %r3018, %r635, 5; and.b32 %r3019, %r3018, 3840; shl.b32 %r3020, %r732, 1; and.b32 %r3021, %r3020, 240; or.b32 %r3022, %r3019, %r3021; bfe.u32 %r3023, %r733, 3, 4; or.b32 %r3024, %r3022, %r3023; shr.u32 %r3025, %r3022, 6; mul.wide.u32 %rd2240, %r3025, 8; add.s64 %rd2241, %rd600, %rd2240; and.b32 %r3026, %r3024, 63; shl.b64 %rd2243, %rd2231, %r3026; ld.u64 %rd2244, [%rd2241+544]; and.b64 %rd2245, %rd2244, %rd2243; setp.eq.s64 %p516, %rd2245, 0; mul.wide.u32 %rd2246, %r3024, 8; add.s64 %rd2247, %rd600, %rd2246; add.s64 %rd601, %rd2247, 1088; @%p516 bra $L__BB6_820; ld.u64 %rd2248, [%rd601]; add.s64 %rd2249, %rd600, %rd2248; and.b32 %r3027, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r3027}; and.b32 %r3028, %r733, -8; st.local.u32 [%rd2+8], %r3028; st.local.u64 [%rd2+48], %rd2249; shl.b32 %r3029, %r635, 6; and.b32 %r3030, %r3029, 448; shl.b32 %r3031, %r732, 3; and.b32 %r3032, %r3031, 56; or.b32 %r3033, %r3030, %r3032; and.b32 %r3034, %r733, 7; or.b32 %r3035, %r3033, %r3034; mul.wide.u32 %rd2250, %r3035, 2; add.s64 %rd2251, %rd2249, %rd2250; ld.u16 %rs355, [%rd2251+96]; cvt.rn.f32.u16 %f887, %rs355; ld.v2.f32 {%f888, %f889}, [%rd2249+80]; fma.rn.ftz.f32 %f952, %f889, %f887, %f888; bra.uni $L__BB6_833; $L__BB6_823: ld.f32 %f952, [%rd590+28]; bra.uni $L__BB6_833; $L__BB6_822: ld.f32 %f952, [%rd596+12]; bra.uni $L__BB6_833; $L__BB6_43: setp.ne.s32 %p26, %r12, %r3232; mov.u16 %rs361, 0; @%p26 bra $L__BB6_46; and.b32 %r1022, %r8, -4096; setp.ne.s32 %p27, %r1022, %r3277; @%p27 bra $L__BB6_46; and.b32 %r1023, %r74, -4096; setp.eq.s32 %p28, %r1023, %r3310; selp.u16 %rs361, 1, 0, %p28; $L__BB6_46: setp.eq.s16 %p29, %rs361, 0; @%p29 bra $L__BB6_52; shl.b32 %r1024, %r7, 3; and.b32 %r1025, %r1024, 31744; and.b32 %r1026, %r8, 3968; shr.u32 %r1027, %r1026, 2; or.b32 %r1028, %r1027, %r1025; bfe.u32 %r1029, %r74, 7, 5; or.b32 %r1030, %r1029, %r1028; shr.u32 %r1031, %r1028, 6; mul.wide.u32 %rd713, %r1031, 8; add.s64 %rd714, %rd2315, %rd713; and.b32 %r1032, %r1030, 63; mov.u64 %rd715, 1; shl.b64 %rd716, %rd715, %r1032; ld.u64 %rd717, [%rd714+4128]; and.b64 %rd718, %rd717, %rd716; setp.eq.s64 %p30, %rd718, 0; mul.wide.u32 %rd719, %r1030, 8; add.s64 %rd720, %rd2315, %rd719; add.s64 %rd45, %rd720, 8256; @%p30 bra $L__BB6_51; and.b32 %r3152, %r7, -128; ld.u64 %rd721, [%rd45]; add.s64 %rd2311, %rd2315, %rd721; st.local.u32 [%rd2+12], %r3152; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1033, %r7, 5; and.b32 %r1034, %r1033, 3840; shl.b32 %r1035, %r8, 1; and.b32 %r1036, %r1035, 240; or.b32 %r1037, %r1036, %r1034; bfe.u32 %r1038, %r74, 3, 4; or.b32 %r1039, %r1038, %r1037; shr.u32 %r1040, %r1037, 6; mul.wide.u32 %rd722, %r1040, 8; add.s64 %rd723, %rd2311, %rd722; and.b32 %r1041, %r1039, 63; shl.b64 %rd725, %rd715, %r1041; ld.u64 %rd726, [%rd723+544]; and.b64 %rd727, %rd726, %rd725; setp.eq.s64 %p31, %rd727, 0; mul.wide.u32 %rd728, %r1039, 8; add.s64 %rd729, %rd2311, %rd728; add.s64 %rd47, %rd729, 1088; and.b32 %r3272, %r7, -128; @%p31 bra $L__BB6_50; and.b32 %r3229, %r7, -8; ld.u64 %rd730, [%rd47]; add.s64 %rd2307, %rd2311, %rd730; and.b32 %r1042, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1042}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1043, %r7, 6; and.b32 %r1044, %r1043, 448; shl.b32 %r1045, %r8, 3; and.b32 %r1046, %r1045, 56; or.b32 %r1047, %r1046, %r1044; and.b32 %r1048, %r74, 7; or.b32 %r1049, %r1047, %r1048; mul.wide.u32 %rd731, %r1049, 2; add.s64 %rd732, %rd2307, %rd731; ld.u16 %rs143, [%rd732+96]; cvt.rn.f32.u16 %f414, %rs143; ld.v2.f32 {%f415, %f416}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f416, %f414, %f415; mov.b64 %rd733, {%r3229, %r1042}; shr.u64 %rd734, %rd733, 32; cvt.u32.u64 %r3184, %rd734; bra.uni $L__BB6_66; $L__BB6_78: setp.ne.s32 %p48, %r12, %r3232; mov.u16 %rs364, 0; @%p48 bra $L__BB6_81; and.b32 %r1107, %r139, -4096; setp.ne.s32 %p49, %r1107, %r3277; @%p49 bra $L__BB6_81; and.b32 %r1108, %r74, -4096; setp.eq.s32 %p50, %r1108, %r3310; selp.u16 %rs364, 1, 0, %p50; $L__BB6_81: setp.eq.s16 %p51, %rs364, 0; @%p51 bra $L__BB6_87; shl.b32 %r1109, %r7, 3; and.b32 %r1110, %r1109, 31744; and.b32 %r1111, %r139, 3968; shr.u32 %r1112, %r1111, 2; or.b32 %r1113, %r1112, %r1110; bfe.u32 %r1114, %r74, 7, 5; or.b32 %r1115, %r1113, %r1114; shr.u32 %r1116, %r1113, 6; mul.wide.u32 %rd782, %r1116, 8; add.s64 %rd783, %rd2315, %rd782; and.b32 %r1117, %r1115, 63; mov.u64 %rd784, 1; shl.b64 %rd785, %rd784, %r1117; ld.u64 %rd786, [%rd783+4128]; and.b64 %rd787, %rd786, %rd785; setp.eq.s64 %p52, %rd787, 0; mul.wide.u32 %rd788, %r1115, 8; add.s64 %rd789, %rd2315, %rd788; add.s64 %rd74, %rd789, 8256; @%p52 bra $L__BB6_86; and.b32 %r3128, %r7, -128; ld.u64 %rd790, [%rd74]; add.s64 %rd2311, %rd2315, %rd790; st.local.u32 [%rd2+12], %r3128; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1118, %r7, 5; and.b32 %r1119, %r1118, 3840; shl.b32 %r1120, %r139, 1; and.b32 %r1121, %r1120, 240; or.b32 %r1122, %r1121, %r1119; bfe.u32 %r1123, %r74, 3, 4; or.b32 %r1124, %r1122, %r1123; shr.u32 %r1125, %r1122, 6; mul.wide.u32 %rd791, %r1125, 8; add.s64 %rd792, %rd2311, %rd791; and.b32 %r1126, %r1124, 63; shl.b64 %rd794, %rd784, %r1126; ld.u64 %rd795, [%rd792+544]; and.b64 %rd796, %rd795, %rd794; setp.eq.s64 %p53, %rd796, 0; mul.wide.u32 %rd797, %r1124, 8; add.s64 %rd798, %rd2311, %rd797; add.s64 %rd76, %rd798, 1088; and.b32 %r3272, %r7, -128; @%p53 bra $L__BB6_85; and.b32 %r3229, %r7, -8; ld.u64 %rd799, [%rd76]; add.s64 %rd2307, %rd2311, %rd799; and.b32 %r1127, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1127}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1128, %r7, 6; and.b32 %r1129, %r1128, 448; shl.b32 %r1130, %r139, 3; and.b32 %r1131, %r1130, 56; or.b32 %r1132, %r1131, %r1129; and.b32 %r1133, %r74, 7; or.b32 %r1134, %r1132, %r1133; mul.wide.u32 %rd800, %r1134, 2; add.s64 %rd801, %rd2307, %rd800; ld.u16 %rs153, [%rd801+96]; cvt.rn.f32.u16 %f435, %rs153; ld.v2.f32 {%f436, %f437}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f437, %f435, %f436; mov.b64 %rd802, {%r3229, %r1127}; shr.u64 %rd803, %rd802, 32; cvt.u32.u64 %r3184, %rd803; bra.uni $L__BB6_101; $L__BB6_113: setp.ne.s32 %p70, %r12, %r3232; mov.u16 %rs367, 0; @%p70 bra $L__BB6_116; and.b32 %r1191, %r139, -4096; setp.ne.s32 %p71, %r1191, %r3277; @%p71 bra $L__BB6_116; and.b32 %r1192, %r9, -4096; setp.eq.s32 %p72, %r1192, %r3310; selp.u16 %rs367, 1, 0, %p72; $L__BB6_116: setp.eq.s16 %p73, %rs367, 0; @%p73 bra $L__BB6_122; shl.b32 %r1193, %r7, 3; and.b32 %r1194, %r1193, 31744; and.b32 %r1195, %r139, 3968; shr.u32 %r1196, %r1195, 2; or.b32 %r1197, %r1196, %r1194; bfe.u32 %r1198, %r9, 7, 5; or.b32 %r1199, %r1197, %r1198; shr.u32 %r1200, %r1197, 6; mul.wide.u32 %rd852, %r1200, 8; add.s64 %rd853, %rd2315, %rd852; and.b32 %r1201, %r1199, 63; mov.u64 %rd854, 1; shl.b64 %rd855, %rd854, %r1201; ld.u64 %rd856, [%rd853+4128]; and.b64 %rd857, %rd856, %rd855; setp.eq.s64 %p74, %rd857, 0; mul.wide.u32 %rd858, %r1199, 8; add.s64 %rd859, %rd2315, %rd858; add.s64 %rd103, %rd859, 8256; @%p74 bra $L__BB6_121; and.b32 %r3101, %r7, -128; ld.u64 %rd860, [%rd103]; add.s64 %rd2311, %rd2315, %rd860; st.local.u32 [%rd2+12], %r3101; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1202, %r7, 5; and.b32 %r1203, %r1202, 3840; shl.b32 %r1204, %r139, 1; and.b32 %r1205, %r1204, 240; or.b32 %r1206, %r1205, %r1203; bfe.u32 %r1207, %r9, 3, 4; or.b32 %r1208, %r1206, %r1207; shr.u32 %r1209, %r1206, 6; mul.wide.u32 %rd861, %r1209, 8; add.s64 %rd862, %rd2311, %rd861; and.b32 %r1210, %r1208, 63; shl.b64 %rd864, %rd854, %r1210; ld.u64 %rd865, [%rd862+544]; and.b64 %rd866, %rd865, %rd864; setp.eq.s64 %p75, %rd866, 0; mul.wide.u32 %rd867, %r1208, 8; add.s64 %rd868, %rd2311, %rd867; add.s64 %rd105, %rd868, 1088; and.b32 %r3272, %r7, -128; @%p75 bra $L__BB6_120; and.b32 %r3229, %r7, -8; ld.u64 %rd869, [%rd105]; add.s64 %rd2307, %rd2311, %rd869; and.b32 %r1211, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1211}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1212, %r7, 7; shl.b32 %r1213, %r139, 3; and.b32 %r1214, %r1213, 56; and.b32 %r1215, %r9, 7; bfi.b32 %r1216, %r1212, %r1215, 6, 3; or.b32 %r1217, %r1216, %r1214; mul.wide.u32 %rd870, %r1217, 2; add.s64 %rd871, %rd2307, %rd870; ld.u16 %rs163, [%rd871+96]; cvt.rn.f32.u16 %f455, %rs163; ld.v2.f32 {%f456, %f457}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f457, %f455, %f456; mov.b64 %rd872, {%r3229, %r1211}; shr.u64 %rd873, %rd872, 32; cvt.u32.u64 %r3184, %rd873; bra.uni $L__BB6_136; $L__BB6_148: and.b32 %r264, %r259, -4096; setp.ne.s32 %p92, %r264, %r3232; mov.u16 %rs370, 0; @%p92 bra $L__BB6_151; and.b32 %r1272, %r139, -4096; setp.ne.s32 %p93, %r1272, %r3277; @%p93 bra $L__BB6_151; and.b32 %r1273, %r9, -4096; setp.eq.s32 %p94, %r1273, %r3310; selp.u16 %rs370, 1, 0, %p94; $L__BB6_151: setp.eq.s16 %p95, %rs370, 0; @%p95 bra $L__BB6_157; shl.b32 %r1274, %r259, 3; and.b32 %r1275, %r1274, 31744; and.b32 %r1276, %r139, 3968; shr.u32 %r1277, %r1276, 2; or.b32 %r1278, %r1277, %r1275; bfe.u32 %r1279, %r9, 7, 5; or.b32 %r1280, %r1278, %r1279; shr.u32 %r1281, %r1278, 6; mul.wide.u32 %rd920, %r1281, 8; add.s64 %rd921, %rd2315, %rd920; and.b32 %r1282, %r1280, 63; mov.u64 %rd922, 1; shl.b64 %rd923, %rd922, %r1282; ld.u64 %rd924, [%rd921+4128]; and.b64 %rd925, %rd924, %rd923; setp.eq.s64 %p96, %rd925, 0; mul.wide.u32 %rd926, %r1280, 8; add.s64 %rd927, %rd2315, %rd926; add.s64 %rd132, %rd927, 8256; @%p96 bra $L__BB6_156; and.b32 %r3134, %r259, -128; ld.u64 %rd928, [%rd132]; add.s64 %rd2311, %rd2315, %rd928; st.local.u32 [%rd2+12], %r3134; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1283, %r259, 5; and.b32 %r1284, %r1283, 3840; shl.b32 %r1285, %r139, 1; and.b32 %r1286, %r1285, 240; or.b32 %r1287, %r1286, %r1284; bfe.u32 %r1288, %r9, 3, 4; or.b32 %r1289, %r1287, %r1288; shr.u32 %r1290, %r1287, 6; mul.wide.u32 %rd929, %r1290, 8; add.s64 %rd930, %rd2311, %rd929; and.b32 %r1291, %r1289, 63; shl.b64 %rd932, %rd922, %r1291; ld.u64 %rd933, [%rd930+544]; and.b64 %rd934, %rd933, %rd932; setp.eq.s64 %p97, %rd934, 0; mul.wide.u32 %rd935, %r1289, 8; add.s64 %rd936, %rd2311, %rd935; add.s64 %rd134, %rd936, 1088; @%p97 bra $L__BB6_155; and.b32 %r3229, %r259, -8; and.b32 %r3272, %r259, -128; ld.u64 %rd937, [%rd134]; add.s64 %rd2307, %rd2311, %rd937; and.b32 %r1292, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1292}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1293, %r259, 7; shl.b32 %r1294, %r139, 3; and.b32 %r1295, %r1294, 56; and.b32 %r1296, %r9, 7; bfi.b32 %r1297, %r1293, %r1296, 6, 3; or.b32 %r1298, %r1297, %r1295; mul.wide.u32 %rd938, %r1298, 2; add.s64 %rd939, %rd2307, %rd938; ld.u16 %rs173, [%rd939+96]; cvt.rn.f32.u16 %f478, %rs173; ld.v2.f32 {%f479, %f480}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f480, %f478, %f479; mov.b64 %rd940, {%r3229, %r1292}; shr.u64 %rd941, %rd940, 32; cvt.u32.u64 %r3184, %rd941; bra.uni $L__BB6_171; $L__BB6_183: and.b32 %r321, %r259, -4096; setp.ne.s32 %p114, %r321, %r3232; mov.u16 %rs373, 0; @%p114 bra $L__BB6_186; and.b32 %r1356, %r139, -4096; setp.ne.s32 %p115, %r1356, %r3277; @%p115 bra $L__BB6_186; and.b32 %r1357, %r74, -4096; setp.eq.s32 %p116, %r1357, %r3310; selp.u16 %rs373, 1, 0, %p116; $L__BB6_186: setp.eq.s16 %p117, %rs373, 0; @%p117 bra $L__BB6_192; shl.b32 %r1358, %r259, 3; and.b32 %r1359, %r1358, 31744; and.b32 %r1360, %r139, 3968; shr.u32 %r1361, %r1360, 2; or.b32 %r1362, %r1361, %r1359; bfe.u32 %r1363, %r74, 7, 5; or.b32 %r1364, %r1362, %r1363; shr.u32 %r1365, %r1362, 6; mul.wide.u32 %rd991, %r1365, 8; add.s64 %rd992, %rd2315, %rd991; and.b32 %r1366, %r1364, 63; mov.u64 %rd993, 1; shl.b64 %rd994, %rd993, %r1366; ld.u64 %rd995, [%rd992+4128]; and.b64 %rd996, %rd995, %rd994; setp.eq.s64 %p118, %rd996, 0; mul.wide.u32 %rd997, %r1364, 8; add.s64 %rd998, %rd2315, %rd997; add.s64 %rd161, %rd998, 8256; @%p118 bra $L__BB6_191; and.b32 %r3137, %r259, -128; ld.u64 %rd999, [%rd161]; add.s64 %rd2311, %rd2315, %rd999; st.local.u32 [%rd2+12], %r3137; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1367, %r259, 5; and.b32 %r1368, %r1367, 3840; shl.b32 %r1369, %r139, 1; and.b32 %r1370, %r1369, 240; or.b32 %r1371, %r1370, %r1368; bfe.u32 %r1372, %r74, 3, 4; or.b32 %r1373, %r1371, %r1372; shr.u32 %r1374, %r1371, 6; mul.wide.u32 %rd1000, %r1374, 8; add.s64 %rd1001, %rd2311, %rd1000; and.b32 %r1375, %r1373, 63; shl.b64 %rd1003, %rd993, %r1375; ld.u64 %rd1004, [%rd1001+544]; and.b64 %rd1005, %rd1004, %rd1003; setp.eq.s64 %p119, %rd1005, 0; mul.wide.u32 %rd1006, %r1373, 8; add.s64 %rd1007, %rd2311, %rd1006; add.s64 %rd163, %rd1007, 1088; @%p119 bra $L__BB6_190; and.b32 %r3229, %r259, -8; and.b32 %r3272, %r259, -128; ld.u64 %rd1008, [%rd163]; add.s64 %rd2307, %rd2311, %rd1008; and.b32 %r1376, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1376}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1377, %r259, 6; and.b32 %r1378, %r1377, 448; shl.b32 %r1379, %r139, 3; and.b32 %r1380, %r1379, 56; or.b32 %r1381, %r1380, %r1378; and.b32 %r1382, %r74, 7; or.b32 %r1383, %r1381, %r1382; mul.wide.u32 %rd1009, %r1383, 2; add.s64 %rd1010, %rd2307, %rd1009; ld.u16 %rs183, [%rd1010+96]; cvt.rn.f32.u16 %f498, %rs183; ld.v2.f32 {%f499, %f500}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f500, %f498, %f499; mov.b64 %rd1011, {%r3229, %r1376}; shr.u64 %rd1012, %rd1011, 32; cvt.u32.u64 %r3184, %rd1012; bra.uni $L__BB6_206; $L__BB6_218: and.b32 %r375, %r259, -4096; setp.ne.s32 %p136, %r375, %r3232; mov.u16 %rs376, 0; @%p136 bra $L__BB6_221; and.b32 %r1443, %r8, -4096; setp.ne.s32 %p137, %r1443, %r3277; @%p137 bra $L__BB6_221; and.b32 %r1444, %r74, -4096; setp.eq.s32 %p138, %r1444, %r3310; selp.u16 %rs376, 1, 0, %p138; $L__BB6_221: setp.eq.s16 %p139, %rs376, 0; @%p139 bra $L__BB6_227; shl.b32 %r1445, %r259, 3; and.b32 %r1446, %r1445, 31744; and.b32 %r1447, %r8, 3968; shr.u32 %r1448, %r1447, 2; or.b32 %r1449, %r1446, %r1448; bfe.u32 %r1450, %r74, 7, 5; or.b32 %r1451, %r1450, %r1449; shr.u32 %r1452, %r1449, 6; mul.wide.u32 %rd1062, %r1452, 8; add.s64 %rd1063, %rd2315, %rd1062; and.b32 %r1453, %r1451, 63; mov.u64 %rd1064, 1; shl.b64 %rd1065, %rd1064, %r1453; ld.u64 %rd1066, [%rd1063+4128]; and.b64 %rd1067, %rd1066, %rd1065; setp.eq.s64 %p140, %rd1067, 0; mul.wide.u32 %rd1068, %r1451, 8; add.s64 %rd1069, %rd2315, %rd1068; add.s64 %rd190, %rd1069, 8256; @%p140 bra $L__BB6_226; ld.u64 %rd1070, [%rd190]; add.s64 %rd2311, %rd2315, %rd1070; st.local.u32 [%rd2+12], %r372; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1454, %r259, 5; and.b32 %r1455, %r1454, 3840; shl.b32 %r1456, %r8, 1; and.b32 %r1457, %r1456, 240; or.b32 %r1458, %r1455, %r1457; bfe.u32 %r1459, %r74, 3, 4; or.b32 %r1460, %r1459, %r1458; shr.u32 %r1461, %r1458, 6; mul.wide.u32 %rd1071, %r1461, 8; add.s64 %rd1072, %rd2311, %rd1071; and.b32 %r1462, %r1460, 63; shl.b64 %rd1074, %rd1064, %r1462; ld.u64 %rd1075, [%rd1072+544]; and.b64 %rd1076, %rd1075, %rd1074; setp.eq.s64 %p141, %rd1076, 0; mul.wide.u32 %rd1077, %r1460, 8; add.s64 %rd1078, %rd2311, %rd1077; add.s64 %rd192, %rd1078, 1088; @%p141 bra $L__BB6_225; and.b32 %r3229, %r259, -8; ld.u64 %rd1079, [%rd192]; add.s64 %rd2307, %rd2311, %rd1079; and.b32 %r1463, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1463}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1464, %r259, 6; and.b32 %r1465, %r1464, 448; shl.b32 %r1466, %r8, 3; and.b32 %r1467, %r1466, 56; or.b32 %r1468, %r1465, %r1467; and.b32 %r1469, %r74, 7; or.b32 %r1470, %r1468, %r1469; mul.wide.u32 %rd1080, %r1470, 2; add.s64 %rd1081, %rd2307, %rd1080; ld.u16 %rs193, [%rd1081+96]; cvt.rn.f32.u16 %f519, %rs193; ld.v2.f32 {%f520, %f521}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f521, %f519, %f520; mov.b64 %rd1082, {%r3229, %r1463}; shr.u64 %rd1083, %rd1082, 32; cvt.u32.u64 %r3184, %rd1083; mov.u32 %r3272, %r372; bra.uni $L__BB6_241; $L__BB6_253: and.b32 %r426, %r259, -4096; setp.ne.s32 %p158, %r426, %r3232; mov.u16 %rs379, 0; @%p158 bra $L__BB6_256; and.b32 %r1529, %r8, -4096; setp.ne.s32 %p159, %r1529, %r3277; @%p159 bra $L__BB6_256; and.b32 %r1530, %r9, -4096; setp.eq.s32 %p160, %r1530, %r3310; selp.u16 %rs379, 1, 0, %p160; $L__BB6_256: setp.eq.s16 %p161, %rs379, 0; @%p161 bra $L__BB6_262; shl.b32 %r1531, %r259, 3; and.b32 %r1532, %r1531, 31744; and.b32 %r1533, %r8, 3968; shr.u32 %r1534, %r1533, 2; or.b32 %r1535, %r1532, %r1534; bfe.u32 %r1536, %r9, 7, 5; or.b32 %r1537, %r1535, %r1536; shr.u32 %r1538, %r1535, 6; mul.wide.u32 %rd1133, %r1538, 8; add.s64 %rd1134, %rd2315, %rd1133; and.b32 %r1539, %r1537, 63; mov.u64 %rd1135, 1; shl.b64 %rd1136, %rd1135, %r1539; ld.u64 %rd1137, [%rd1134+4128]; and.b64 %rd1138, %rd1137, %rd1136; setp.eq.s64 %p162, %rd1138, 0; mul.wide.u32 %rd1139, %r1537, 8; add.s64 %rd1140, %rd2315, %rd1139; add.s64 %rd219, %rd1140, 8256; @%p162 bra $L__BB6_261; ld.u64 %rd1141, [%rd219]; add.s64 %rd2311, %rd2315, %rd1141; st.local.u32 [%rd2+12], %r423; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1540, %r259, 5; and.b32 %r1541, %r1540, 3840; shl.b32 %r1542, %r8, 1; and.b32 %r1543, %r1542, 240; or.b32 %r1544, %r1541, %r1543; bfe.u32 %r1545, %r9, 3, 4; or.b32 %r1546, %r1544, %r1545; shr.u32 %r1547, %r1544, 6; mul.wide.u32 %rd1142, %r1547, 8; add.s64 %rd1143, %rd2311, %rd1142; and.b32 %r1548, %r1546, 63; shl.b64 %rd1145, %rd1135, %r1548; ld.u64 %rd1146, [%rd1143+544]; and.b64 %rd1147, %rd1146, %rd1145; setp.eq.s64 %p163, %rd1147, 0; mul.wide.u32 %rd1148, %r1546, 8; add.s64 %rd1149, %rd2311, %rd1148; add.s64 %rd221, %rd1149, 1088; @%p163 bra $L__BB6_260; and.b32 %r3229, %r259, -8; ld.u64 %rd1150, [%rd221]; add.s64 %rd2307, %rd2311, %rd1150; and.b32 %r1549, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1549}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1550, %r259, 6; and.b32 %r1551, %r1550, 448; shl.b32 %r1552, %r8, 3; and.b32 %r1553, %r1552, 56; or.b32 %r1554, %r1551, %r1553; and.b32 %r1555, %r9, 7; or.b32 %r1556, %r1554, %r1555; mul.wide.u32 %rd1151, %r1556, 2; add.s64 %rd1152, %rd2307, %rd1151; ld.u16 %rs203, [%rd1152+96]; cvt.rn.f32.u16 %f539, %rs203; ld.v2.f32 {%f540, %f541}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f541, %f539, %f540; mov.b64 %rd1153, {%r3229, %r1549}; shr.u64 %rd1154, %rd1153, 32; cvt.u32.u64 %r3184, %rd1154; mov.u32 %r3272, %r423; bra.uni $L__BB6_276; $L__BB6_288: and.b32 %r476, %r471, -4096; setp.ne.s32 %p180, %r476, %r3232; mov.u16 %rs382, 0; @%p180 bra $L__BB6_291; and.b32 %r1614, %r8, -4096; setp.ne.s32 %p181, %r1614, %r3277; @%p181 bra $L__BB6_291; and.b32 %r1615, %r9, -4096; setp.eq.s32 %p182, %r1615, %r3310; selp.u16 %rs382, 1, 0, %p182; $L__BB6_291: setp.eq.s16 %p183, %rs382, 0; @%p183 bra $L__BB6_297; shl.b32 %r1616, %r471, 3; and.b32 %r1617, %r1616, 31744; and.b32 %r1618, %r8, 3968; shr.u32 %r1619, %r1618, 2; or.b32 %r1620, %r1619, %r1617; bfe.u32 %r1621, %r9, 7, 5; or.b32 %r1622, %r1620, %r1621; shr.u32 %r1623, %r1620, 6; mul.wide.u32 %rd1203, %r1623, 8; add.s64 %rd1204, %rd2315, %rd1203; and.b32 %r1624, %r1622, 63; mov.u64 %rd1205, 1; shl.b64 %rd1206, %rd1205, %r1624; ld.u64 %rd1207, [%rd1204+4128]; and.b64 %rd1208, %rd1207, %rd1206; setp.eq.s64 %p184, %rd1208, 0; mul.wide.u32 %rd1209, %r1622, 8; add.s64 %rd1210, %rd2315, %rd1209; add.s64 %rd248, %rd1210, 8256; @%p184 bra $L__BB6_296; ld.u64 %rd1211, [%rd248]; add.s64 %rd2311, %rd2315, %rd1211; st.local.u32 [%rd2+12], %r473; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1625, %r471, 5; and.b32 %r1626, %r1625, 3840; shl.b32 %r1627, %r8, 1; and.b32 %r1628, %r1627, 240; or.b32 %r1629, %r1628, %r1626; bfe.u32 %r1630, %r9, 3, 4; or.b32 %r1631, %r1629, %r1630; shr.u32 %r1632, %r1629, 6; mul.wide.u32 %rd1212, %r1632, 8; add.s64 %rd1213, %rd2311, %rd1212; and.b32 %r1633, %r1631, 63; shl.b64 %rd1215, %rd1205, %r1633; ld.u64 %rd1216, [%rd1213+544]; and.b64 %rd1217, %rd1216, %rd1215; setp.eq.s64 %p185, %rd1217, 0; mul.wide.u32 %rd1218, %r1631, 8; add.s64 %rd1219, %rd2311, %rd1218; add.s64 %rd250, %rd1219, 1088; @%p185 bra $L__BB6_295; ld.u64 %rd1220, [%rd250]; add.s64 %rd2307, %rd2311, %rd1220; and.b32 %r1634, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1634}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1635, %r471, 6; and.b32 %r1636, %r1635, 448; shl.b32 %r1637, %r8, 3; and.b32 %r1638, %r1637, 56; or.b32 %r1639, %r1638, %r1636; and.b32 %r1640, %r9, 7; or.b32 %r1641, %r1639, %r1640; mul.wide.u32 %rd1221, %r1641, 2; add.s64 %rd1222, %rd2307, %rd1221; ld.u16 %rs213, [%rd1222+96]; cvt.rn.f32.u16 %f566, %rs213; ld.v2.f32 {%f567, %f568}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f568, %f566, %f567; mov.b64 %rd1223, {%r472, %r1634}; shr.u64 %rd1224, %rd1223, 32; cvt.u32.u64 %r3184, %rd1224; mov.u32 %r3272, %r473; mov.u32 %r3229, %r472; bra.uni $L__BB6_311; $L__BB6_323: and.b32 %r521, %r471, -4096; setp.ne.s32 %p202, %r521, %r3232; mov.u16 %rs385, 0; @%p202 bra $L__BB6_326; and.b32 %r1699, %r8, -4096; setp.ne.s32 %p203, %r1699, %r3277; @%p203 bra $L__BB6_326; and.b32 %r1700, %r74, -4096; setp.eq.s32 %p204, %r1700, %r3310; selp.u16 %rs385, 1, 0, %p204; $L__BB6_326: setp.eq.s16 %p205, %rs385, 0; @%p205 bra $L__BB6_332; shl.b32 %r1701, %r471, 3; and.b32 %r1702, %r1701, 31744; and.b32 %r1703, %r8, 3968; shr.u32 %r1704, %r1703, 2; or.b32 %r1705, %r1704, %r1702; bfe.u32 %r1706, %r74, 7, 5; or.b32 %r1707, %r1706, %r1705; shr.u32 %r1708, %r1705, 6; mul.wide.u32 %rd1273, %r1708, 8; add.s64 %rd1274, %rd2315, %rd1273; and.b32 %r1709, %r1707, 63; mov.u64 %rd1275, 1; shl.b64 %rd1276, %rd1275, %r1709; ld.u64 %rd1277, [%rd1274+4128]; and.b64 %rd1278, %rd1277, %rd1276; setp.eq.s64 %p206, %rd1278, 0; mul.wide.u32 %rd1279, %r1707, 8; add.s64 %rd1280, %rd2315, %rd1279; add.s64 %rd277, %rd1280, 8256; @%p206 bra $L__BB6_331; ld.u64 %rd1281, [%rd277]; add.s64 %rd2311, %rd2315, %rd1281; st.local.u32 [%rd2+12], %r518; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1710, %r471, 5; and.b32 %r1711, %r1710, 3840; shl.b32 %r1712, %r8, 1; and.b32 %r1713, %r1712, 240; or.b32 %r1714, %r1713, %r1711; bfe.u32 %r1715, %r74, 3, 4; or.b32 %r1716, %r1715, %r1714; shr.u32 %r1717, %r1714, 6; mul.wide.u32 %rd1282, %r1717, 8; add.s64 %rd1283, %rd2311, %rd1282; and.b32 %r1718, %r1716, 63; shl.b64 %rd1285, %rd1275, %r1718; ld.u64 %rd1286, [%rd1283+544]; and.b64 %rd1287, %rd1286, %rd1285; setp.eq.s64 %p207, %rd1287, 0; mul.wide.u32 %rd1288, %r1716, 8; add.s64 %rd1289, %rd2311, %rd1288; add.s64 %rd279, %rd1289, 1088; @%p207 bra $L__BB6_330; ld.u64 %rd1290, [%rd279]; add.s64 %rd2307, %rd2311, %rd1290; and.b32 %r1719, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1719}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1720, %r471, 6; and.b32 %r1721, %r1720, 448; shl.b32 %r1722, %r8, 3; and.b32 %r1723, %r1722, 56; or.b32 %r1724, %r1723, %r1721; and.b32 %r1725, %r74, 7; or.b32 %r1726, %r1724, %r1725; mul.wide.u32 %rd1291, %r1726, 2; add.s64 %rd1292, %rd2307, %rd1291; ld.u16 %rs223, [%rd1292+96]; cvt.rn.f32.u16 %f586, %rs223; ld.v2.f32 {%f587, %f588}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f588, %f586, %f587; mov.b64 %rd1293, {%r472, %r1719}; shr.u64 %rd1294, %rd1293, 32; cvt.u32.u64 %r3184, %rd1294; mov.u32 %r3272, %r518; mov.u32 %r3229, %r472; bra.uni $L__BB6_346; $L__BB6_358: and.b32 %r563, %r471, -4096; setp.ne.s32 %p224, %r563, %r3232; mov.u16 %rs388, 0; @%p224 bra $L__BB6_361; and.b32 %r1785, %r139, -4096; setp.ne.s32 %p225, %r1785, %r3277; @%p225 bra $L__BB6_361; and.b32 %r1786, %r74, -4096; setp.eq.s32 %p226, %r1786, %r3310; selp.u16 %rs388, 1, 0, %p226; $L__BB6_361: setp.eq.s16 %p227, %rs388, 0; @%p227 bra $L__BB6_367; shl.b32 %r1787, %r471, 3; and.b32 %r1788, %r1787, 31744; and.b32 %r1789, %r139, 3968; shr.u32 %r1790, %r1789, 2; or.b32 %r1791, %r1790, %r1788; bfe.u32 %r1792, %r74, 7, 5; or.b32 %r1793, %r1791, %r1792; shr.u32 %r1794, %r1791, 6; mul.wide.u32 %rd1344, %r1794, 8; add.s64 %rd1345, %rd2315, %rd1344; and.b32 %r1795, %r1793, 63; mov.u64 %rd1346, 1; shl.b64 %rd1347, %rd1346, %r1795; ld.u64 %rd1348, [%rd1345+4128]; and.b64 %rd1349, %rd1348, %rd1347; setp.eq.s64 %p228, %rd1349, 0; mul.wide.u32 %rd1350, %r1793, 8; add.s64 %rd1351, %rd2315, %rd1350; add.s64 %rd306, %rd1351, 8256; @%p228 bra $L__BB6_366; ld.u64 %rd1352, [%rd306]; add.s64 %rd2311, %rd2315, %rd1352; st.local.u32 [%rd2+12], %r560; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1796, %r471, 5; and.b32 %r1797, %r1796, 3840; shl.b32 %r1798, %r139, 1; and.b32 %r1799, %r1798, 240; or.b32 %r1800, %r1799, %r1797; bfe.u32 %r1801, %r74, 3, 4; or.b32 %r1802, %r1800, %r1801; shr.u32 %r1803, %r1800, 6; mul.wide.u32 %rd1353, %r1803, 8; add.s64 %rd1354, %rd2311, %rd1353; and.b32 %r1804, %r1802, 63; shl.b64 %rd1356, %rd1346, %r1804; ld.u64 %rd1357, [%rd1354+544]; and.b64 %rd1358, %rd1357, %rd1356; setp.eq.s64 %p229, %rd1358, 0; mul.wide.u32 %rd1359, %r1802, 8; add.s64 %rd1360, %rd2311, %rd1359; add.s64 %rd308, %rd1360, 1088; @%p229 bra $L__BB6_365; ld.u64 %rd1361, [%rd308]; add.s64 %rd2307, %rd2311, %rd1361; and.b32 %r1805, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1805}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1806, %r471, 6; and.b32 %r1807, %r1806, 448; shl.b32 %r1808, %r139, 3; and.b32 %r1809, %r1808, 56; or.b32 %r1810, %r1809, %r1807; and.b32 %r1811, %r74, 7; or.b32 %r1812, %r1810, %r1811; mul.wide.u32 %rd1362, %r1812, 2; add.s64 %rd1363, %rd2307, %rd1362; ld.u16 %rs233, [%rd1363+96]; cvt.rn.f32.u16 %f607, %rs233; ld.v2.f32 {%f608, %f609}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f609, %f607, %f608; mov.b64 %rd1364, {%r472, %r1805}; shr.u64 %rd1365, %rd1364, 32; cvt.u32.u64 %r3184, %rd1365; mov.u32 %r3272, %r560; mov.u32 %r3229, %r472; bra.uni $L__BB6_381; $L__BB6_393: and.b32 %r602, %r471, -4096; setp.ne.s32 %p246, %r602, %r3232; mov.u16 %rs391, 0; @%p246 bra $L__BB6_396; and.b32 %r1870, %r139, -4096; setp.ne.s32 %p247, %r1870, %r3277; @%p247 bra $L__BB6_396; and.b32 %r1871, %r9, -4096; setp.eq.s32 %p248, %r1871, %r3310; selp.u16 %rs391, 1, 0, %p248; $L__BB6_396: setp.eq.s16 %p249, %rs391, 0; @%p249 bra $L__BB6_402; shl.b32 %r1872, %r471, 3; and.b32 %r1873, %r1872, 31744; and.b32 %r1874, %r139, 3968; shr.u32 %r1875, %r1874, 2; or.b32 %r1876, %r1875, %r1873; bfe.u32 %r1877, %r9, 7, 5; or.b32 %r1878, %r1876, %r1877; shr.u32 %r1879, %r1876, 6; mul.wide.u32 %rd1415, %r1879, 8; add.s64 %rd1416, %rd2315, %rd1415; and.b32 %r1880, %r1878, 63; mov.u64 %rd1417, 1; shl.b64 %rd1418, %rd1417, %r1880; ld.u64 %rd1419, [%rd1416+4128]; and.b64 %rd1420, %rd1419, %rd1418; setp.eq.s64 %p250, %rd1420, 0; mul.wide.u32 %rd1421, %r1878, 8; add.s64 %rd1422, %rd2315, %rd1421; add.s64 %rd335, %rd1422, 8256; @%p250 bra $L__BB6_401; ld.u64 %rd1423, [%rd335]; add.s64 %rd2311, %rd2315, %rd1423; st.local.u32 [%rd2+12], %r599; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1881, %r471, 5; and.b32 %r1882, %r1881, 3840; shl.b32 %r1883, %r139, 1; and.b32 %r1884, %r1883, 240; or.b32 %r1885, %r1884, %r1882; bfe.u32 %r1886, %r9, 3, 4; or.b32 %r1887, %r1885, %r1886; shr.u32 %r1888, %r1885, 6; mul.wide.u32 %rd1424, %r1888, 8; add.s64 %rd1425, %rd2311, %rd1424; and.b32 %r1889, %r1887, 63; shl.b64 %rd1427, %rd1417, %r1889; ld.u64 %rd1428, [%rd1425+544]; and.b64 %rd1429, %rd1428, %rd1427; setp.eq.s64 %p251, %rd1429, 0; mul.wide.u32 %rd1430, %r1887, 8; add.s64 %rd1431, %rd2311, %rd1430; add.s64 %rd337, %rd1431, 1088; @%p251 bra $L__BB6_400; ld.u64 %rd1432, [%rd337]; add.s64 %rd2307, %rd2311, %rd1432; and.b32 %r1890, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1890}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1891, %r471, 7; shl.b32 %r1892, %r139, 3; and.b32 %r1893, %r1892, 56; and.b32 %r1894, %r9, 7; bfi.b32 %r1895, %r1891, %r1894, 6, 3; or.b32 %r1896, %r1895, %r1893; mul.wide.u32 %rd1433, %r1896, 2; add.s64 %rd1434, %rd2307, %rd1433; ld.u16 %rs243, [%rd1434+96]; cvt.rn.f32.u16 %f627, %rs243; ld.v2.f32 {%f628, %f629}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f629, %f627, %f628; mov.b64 %rd1435, {%r472, %r1890}; shr.u64 %rd1436, %rd1435, 32; cvt.u32.u64 %r3184, %rd1436; mov.u32 %r3272, %r599; mov.u32 %r3229, %r472; bra.uni $L__BB6_416; $L__BB6_428: and.b32 %r640, %r635, -4096; setp.ne.s32 %p268, %r640, %r3232; mov.u16 %rs394, 0; @%p268 bra $L__BB6_431; and.b32 %r1952, %r139, -4096; setp.ne.s32 %p269, %r1952, %r3277; @%p269 bra $L__BB6_431; and.b32 %r1953, %r9, -4096; setp.eq.s32 %p270, %r1953, %r3310; selp.u16 %rs394, 1, 0, %p270; $L__BB6_431: setp.eq.s16 %p271, %rs394, 0; @%p271 bra $L__BB6_437; shl.b32 %r1954, %r635, 3; and.b32 %r1955, %r1954, 31744; and.b32 %r1956, %r139, 3968; shr.u32 %r1957, %r1956, 2; or.b32 %r1958, %r1957, %r1955; bfe.u32 %r1959, %r9, 7, 5; or.b32 %r1960, %r1958, %r1959; shr.u32 %r1961, %r1958, 6; mul.wide.u32 %rd1486, %r1961, 8; add.s64 %rd1487, %rd2315, %rd1486; and.b32 %r1962, %r1960, 63; mov.u64 %rd1488, 1; shl.b64 %rd1489, %rd1488, %r1962; ld.u64 %rd1490, [%rd1487+4128]; and.b64 %rd1491, %rd1490, %rd1489; setp.eq.s64 %p272, %rd1491, 0; mul.wide.u32 %rd1492, %r1960, 8; add.s64 %rd1493, %rd2315, %rd1492; add.s64 %rd364, %rd1493, 8256; @%p272 bra $L__BB6_436; ld.u64 %rd1494, [%rd364]; add.s64 %rd2311, %rd2315, %rd1494; st.local.u32 [%rd2+12], %r637; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1963, %r635, 5; and.b32 %r1964, %r1963, 3840; shl.b32 %r1965, %r139, 1; and.b32 %r1966, %r1965, 240; or.b32 %r1967, %r1966, %r1964; bfe.u32 %r1968, %r9, 3, 4; or.b32 %r1969, %r1967, %r1968; shr.u32 %r1970, %r1967, 6; mul.wide.u32 %rd1495, %r1970, 8; add.s64 %rd1496, %rd2311, %rd1495; and.b32 %r1971, %r1969, 63; shl.b64 %rd1498, %rd1488, %r1971; ld.u64 %rd1499, [%rd1496+544]; and.b64 %rd1500, %rd1499, %rd1498; setp.eq.s64 %p273, %rd1500, 0; mul.wide.u32 %rd1501, %r1969, 8; add.s64 %rd1502, %rd2311, %rd1501; add.s64 %rd366, %rd1502, 1088; @%p273 bra $L__BB6_435; ld.u64 %rd1503, [%rd366]; add.s64 %rd2307, %rd2311, %rd1503; and.b32 %r1972, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r1972}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1973, %r635, 7; shl.b32 %r1974, %r139, 3; and.b32 %r1975, %r1974, 56; and.b32 %r1976, %r9, 7; bfi.b32 %r1977, %r1973, %r1976, 6, 3; or.b32 %r1978, %r1977, %r1975; mul.wide.u32 %rd1504, %r1978, 2; add.s64 %rd1505, %rd2307, %rd1504; ld.u16 %rs253, [%rd1505+96]; cvt.rn.f32.u16 %f650, %rs253; ld.v2.f32 {%f651, %f652}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f652, %f650, %f651; mov.b64 %rd1506, {%r636, %r1972}; shr.u64 %rd1507, %rd1506, 32; cvt.u32.u64 %r3184, %rd1507; mov.u32 %r3272, %r637; mov.u32 %r3229, %r636; bra.uni $L__BB6_451; $L__BB6_463: and.b32 %r673, %r635, -4096; setp.ne.s32 %p290, %r673, %r3232; mov.u16 %rs397, 0; @%p290 bra $L__BB6_466; and.b32 %r2036, %r139, -4096; setp.ne.s32 %p291, %r2036, %r3277; @%p291 bra $L__BB6_466; and.b32 %r2037, %r74, -4096; setp.eq.s32 %p292, %r2037, %r3310; selp.u16 %rs397, 1, 0, %p292; $L__BB6_466: setp.eq.s16 %p293, %rs397, 0; @%p293 bra $L__BB6_472; shl.b32 %r2038, %r635, 3; and.b32 %r2039, %r2038, 31744; and.b32 %r2040, %r139, 3968; shr.u32 %r2041, %r2040, 2; or.b32 %r2042, %r2041, %r2039; bfe.u32 %r2043, %r74, 7, 5; or.b32 %r2044, %r2042, %r2043; shr.u32 %r2045, %r2042, 6; mul.wide.u32 %rd1557, %r2045, 8; add.s64 %rd1558, %rd2315, %rd1557; and.b32 %r2046, %r2044, 63; mov.u64 %rd1559, 1; shl.b64 %rd1560, %rd1559, %r2046; ld.u64 %rd1561, [%rd1558+4128]; and.b64 %rd1562, %rd1561, %rd1560; setp.eq.s64 %p294, %rd1562, 0; mul.wide.u32 %rd1563, %r2044, 8; add.s64 %rd1564, %rd2315, %rd1563; add.s64 %rd393, %rd1564, 8256; @%p294 bra $L__BB6_471; ld.u64 %rd1565, [%rd393]; add.s64 %rd2311, %rd2315, %rd1565; st.local.u32 [%rd2+12], %r670; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2047, %r635, 5; and.b32 %r2048, %r2047, 3840; shl.b32 %r2049, %r139, 1; and.b32 %r2050, %r2049, 240; or.b32 %r2051, %r2050, %r2048; bfe.u32 %r2052, %r74, 3, 4; or.b32 %r2053, %r2051, %r2052; shr.u32 %r2054, %r2051, 6; mul.wide.u32 %rd1566, %r2054, 8; add.s64 %rd1567, %rd2311, %rd1566; and.b32 %r2055, %r2053, 63; shl.b64 %rd1569, %rd1559, %r2055; ld.u64 %rd1570, [%rd1567+544]; and.b64 %rd1571, %rd1570, %rd1569; setp.eq.s64 %p295, %rd1571, 0; mul.wide.u32 %rd1572, %r2053, 8; add.s64 %rd1573, %rd2311, %rd1572; add.s64 %rd395, %rd1573, 1088; @%p295 bra $L__BB6_470; ld.u64 %rd1574, [%rd395]; add.s64 %rd2307, %rd2311, %rd1574; and.b32 %r2056, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2056}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2057, %r635, 6; and.b32 %r2058, %r2057, 448; shl.b32 %r2059, %r139, 3; and.b32 %r2060, %r2059, 56; or.b32 %r2061, %r2060, %r2058; and.b32 %r2062, %r74, 7; or.b32 %r2063, %r2061, %r2062; mul.wide.u32 %rd1575, %r2063, 2; add.s64 %rd1576, %rd2307, %rd1575; ld.u16 %rs263, [%rd1576+96]; cvt.rn.f32.u16 %f670, %rs263; ld.v2.f32 {%f671, %f672}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f672, %f670, %f671; mov.b64 %rd1577, {%r636, %r2056}; shr.u64 %rd1578, %rd1577, 32; cvt.u32.u64 %r3184, %rd1578; mov.u32 %r3272, %r670; mov.u32 %r3229, %r636; bra.uni $L__BB6_486; $L__BB6_498: and.b32 %r703, %r635, -4096; setp.ne.s32 %p312, %r703, %r3232; mov.u16 %rs400, 0; @%p312 bra $L__BB6_501; and.b32 %r2123, %r8, -4096; setp.ne.s32 %p313, %r2123, %r3277; @%p313 bra $L__BB6_501; and.b32 %r2124, %r74, -4096; setp.eq.s32 %p314, %r2124, %r3310; selp.u16 %rs400, 1, 0, %p314; $L__BB6_501: setp.eq.s16 %p315, %rs400, 0; @%p315 bra $L__BB6_507; shl.b32 %r2125, %r635, 3; and.b32 %r2126, %r2125, 31744; and.b32 %r2127, %r8, 3968; shr.u32 %r2128, %r2127, 2; or.b32 %r2129, %r2126, %r2128; bfe.u32 %r2130, %r74, 7, 5; or.b32 %r2131, %r2130, %r2129; shr.u32 %r2132, %r2129, 6; mul.wide.u32 %rd1628, %r2132, 8; add.s64 %rd1629, %rd2315, %rd1628; and.b32 %r2133, %r2131, 63; mov.u64 %rd1630, 1; shl.b64 %rd1631, %rd1630, %r2133; ld.u64 %rd1632, [%rd1629+4128]; and.b64 %rd1633, %rd1632, %rd1631; setp.eq.s64 %p316, %rd1633, 0; mul.wide.u32 %rd1634, %r2131, 8; add.s64 %rd1635, %rd2315, %rd1634; add.s64 %rd422, %rd1635, 8256; @%p316 bra $L__BB6_506; ld.u64 %rd1636, [%rd422]; add.s64 %rd2311, %rd2315, %rd1636; st.local.u32 [%rd2+12], %r700; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2134, %r635, 5; and.b32 %r2135, %r2134, 3840; shl.b32 %r2136, %r8, 1; and.b32 %r2137, %r2136, 240; or.b32 %r2138, %r2135, %r2137; bfe.u32 %r2139, %r74, 3, 4; or.b32 %r2140, %r2139, %r2138; shr.u32 %r2141, %r2138, 6; mul.wide.u32 %rd1637, %r2141, 8; add.s64 %rd1638, %rd2311, %rd1637; and.b32 %r2142, %r2140, 63; shl.b64 %rd1640, %rd1630, %r2142; ld.u64 %rd1641, [%rd1638+544]; and.b64 %rd1642, %rd1641, %rd1640; setp.eq.s64 %p317, %rd1642, 0; mul.wide.u32 %rd1643, %r2140, 8; add.s64 %rd1644, %rd2311, %rd1643; add.s64 %rd424, %rd1644, 1088; @%p317 bra $L__BB6_505; ld.u64 %rd1645, [%rd424]; add.s64 %rd2307, %rd2311, %rd1645; and.b32 %r2143, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2143}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2144, %r635, 6; and.b32 %r2145, %r2144, 448; shl.b32 %r2146, %r8, 3; and.b32 %r2147, %r2146, 56; or.b32 %r2148, %r2145, %r2147; and.b32 %r2149, %r74, 7; or.b32 %r2150, %r2148, %r2149; mul.wide.u32 %rd1646, %r2150, 2; add.s64 %rd1647, %rd2307, %rd1646; ld.u16 %rs273, [%rd1647+96]; cvt.rn.f32.u16 %f691, %rs273; ld.v2.f32 {%f692, %f693}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f693, %f691, %f692; mov.b64 %rd1648, {%r636, %r2143}; shr.u64 %rd1649, %rd1648, 32; cvt.u32.u64 %r3184, %rd1649; mov.u32 %r3272, %r700; mov.u32 %r3229, %r636; bra.uni $L__BB6_521; $L__BB6_533: and.b32 %r728, %r635, -4096; setp.ne.s32 %p334, %r728, %r3232; mov.u16 %rs403, 0; @%p334 bra $L__BB6_536; and.b32 %r2210, %r8, -4096; setp.ne.s32 %p335, %r2210, %r3277; @%p335 bra $L__BB6_536; and.b32 %r2211, %r9, -4096; setp.eq.s32 %p336, %r2211, %r3310; selp.u16 %rs403, 1, 0, %p336; $L__BB6_536: setp.eq.s16 %p337, %rs403, 0; @%p337 bra $L__BB6_542; shl.b32 %r2212, %r635, 3; and.b32 %r2213, %r2212, 31744; and.b32 %r2214, %r8, 3968; shr.u32 %r2215, %r2214, 2; or.b32 %r2216, %r2213, %r2215; bfe.u32 %r2217, %r9, 7, 5; or.b32 %r2218, %r2216, %r2217; shr.u32 %r2219, %r2216, 6; mul.wide.u32 %rd1698, %r2219, 8; add.s64 %rd1699, %rd2315, %rd1698; and.b32 %r2220, %r2218, 63; mov.u64 %rd1700, 1; shl.b64 %rd1701, %rd1700, %r2220; ld.u64 %rd1702, [%rd1699+4128]; and.b64 %rd1703, %rd1702, %rd1701; setp.eq.s64 %p338, %rd1703, 0; mul.wide.u32 %rd1704, %r2218, 8; add.s64 %rd1705, %rd2315, %rd1704; add.s64 %rd450, %rd1705, 8256; @%p338 bra $L__BB6_541; ld.u64 %rd1706, [%rd450]; add.s64 %rd451, %rd2315, %rd1706; st.local.u32 [%rd2+12], %r727; and.b32 %r2221, %r8, -128; st.local.u32 [%rd2+16], %r2221; and.b32 %r2222, %r9, -128; st.local.u32 [%rd2+20], %r2222; st.local.u64 [%rd2+56], %rd451; shl.b32 %r2223, %r635, 5; and.b32 %r2224, %r2223, 3840; shl.b32 %r2225, %r8, 1; and.b32 %r2226, %r2225, 240; or.b32 %r2227, %r2224, %r2226; bfe.u32 %r2228, %r9, 3, 4; or.b32 %r2229, %r2227, %r2228; shr.u32 %r2230, %r2227, 6; mul.wide.u32 %rd1707, %r2230, 8; add.s64 %rd1708, %rd451, %rd1707; and.b32 %r2231, %r2229, 63; shl.b64 %rd1710, %rd1700, %r2231; ld.u64 %rd1711, [%rd1708+544]; and.b64 %rd1712, %rd1711, %rd1710; setp.eq.s64 %p339, %rd1712, 0; mul.wide.u32 %rd1713, %r2229, 8; add.s64 %rd1714, %rd451, %rd1713; add.s64 %rd452, %rd1714, 1088; @%p339 bra $L__BB6_540; ld.u64 %rd1715, [%rd452]; add.s64 %rd1716, %rd451, %rd1715; and.b32 %r2232, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2232}; and.b32 %r2233, %r9, -8; st.local.u32 [%rd2+8], %r2233; st.local.u64 [%rd2+48], %rd1716; shl.b32 %r2234, %r635, 6; and.b32 %r2235, %r2234, 448; shl.b32 %r2236, %r8, 3; and.b32 %r2237, %r2236, 56; or.b32 %r2238, %r2235, %r2237; and.b32 %r2239, %r9, 7; or.b32 %r2240, %r2238, %r2239; mul.wide.u32 %rd1717, %r2240, 2; add.s64 %rd1718, %rd1716, %rd1717; ld.u16 %rs283, [%rd1718+96]; cvt.rn.f32.u16 %f711, %rs283; ld.v2.f32 {%f712, %f713}, [%rd1716+80]; fma.rn.ftz.f32 %f944, %f713, %f711, %f712; bra.uni $L__BB6_556; $L__BB6_568: and.b32 %r736, %r471, -4096; ld.local.u32 %r2308, [%rd2+24]; setp.ne.s32 %p356, %r736, %r2308; mov.u16 %rs406, 0; @%p356 bra $L__BB6_571; and.b32 %r2309, %r732, -4096; ld.local.u32 %r2310, [%rd2+28]; setp.ne.s32 %p357, %r2309, %r2310; @%p357 bra $L__BB6_571; and.b32 %r2311, %r733, -4096; ld.local.u32 %r2312, [%rd2+32]; setp.eq.s32 %p358, %r2311, %r2312; selp.u16 %rs406, 1, 0, %p358; $L__BB6_571: setp.eq.s16 %p359, %rs406, 0; @%p359 bra $L__BB6_577; shl.b32 %r2313, %r471, 3; and.b32 %r2314, %r2313, 31744; and.b32 %r2315, %r732, 3968; shr.u32 %r2316, %r2315, 2; or.b32 %r2317, %r2316, %r2314; bfe.u32 %r2318, %r733, 7, 5; or.b32 %r2319, %r2317, %r2318; shr.u32 %r2320, %r2317, 6; ld.local.u64 %rd466, [%rd2+64]; mul.wide.u32 %rd1764, %r2320, 8; add.s64 %rd1765, %rd466, %rd1764; and.b32 %r2321, %r2319, 63; mov.u64 %rd1766, 1; shl.b64 %rd1767, %rd1766, %r2321; ld.u64 %rd1768, [%rd1765+4128]; and.b64 %rd1769, %rd1768, %rd1767; setp.eq.s64 %p360, %rd1769, 0; mul.wide.u32 %rd1770, %r2319, 8; add.s64 %rd1771, %rd466, %rd1770; add.s64 %rd467, %rd1771, 8256; @%p360 bra $L__BB6_576; ld.u64 %rd1772, [%rd467]; add.s64 %rd468, %rd466, %rd1772; st.local.u32 [%rd2+12], %r735; and.b32 %r2322, %r732, -128; st.local.u32 [%rd2+16], %r2322; and.b32 %r2323, %r733, -128; st.local.u32 [%rd2+20], %r2323; st.local.u64 [%rd2+56], %rd468; shl.b32 %r2324, %r471, 5; and.b32 %r2325, %r2324, 3840; shl.b32 %r2326, %r732, 1; and.b32 %r2327, %r2326, 240; or.b32 %r2328, %r2327, %r2325; bfe.u32 %r2329, %r733, 3, 4; or.b32 %r2330, %r2328, %r2329; shr.u32 %r2331, %r2328, 6; mul.wide.u32 %rd1773, %r2331, 8; add.s64 %rd1774, %rd468, %rd1773; and.b32 %r2332, %r2330, 63; shl.b64 %rd1776, %rd1766, %r2332; ld.u64 %rd1777, [%rd1774+544]; and.b64 %rd1778, %rd1777, %rd1776; setp.eq.s64 %p361, %rd1778, 0; mul.wide.u32 %rd1779, %r2330, 8; add.s64 %rd1780, %rd468, %rd1779; add.s64 %rd469, %rd1780, 1088; @%p361 bra $L__BB6_575; ld.u64 %rd1781, [%rd469]; add.s64 %rd1782, %rd468, %rd1781; and.b32 %r2333, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2333}; and.b32 %r2334, %r733, -8; st.local.u32 [%rd2+8], %r2334; st.local.u64 [%rd2+48], %rd1782; shl.b32 %r2335, %r471, 6; and.b32 %r2336, %r2335, 448; shl.b32 %r2337, %r732, 3; and.b32 %r2338, %r2337, 56; or.b32 %r2339, %r2338, %r2336; and.b32 %r2340, %r733, 7; or.b32 %r2341, %r2339, %r2340; mul.wide.u32 %rd1783, %r2341, 2; add.s64 %rd1784, %rd1782, %rd1783; ld.u16 %rs293, [%rd1784+96]; cvt.rn.f32.u16 %f752, %rs293; ld.v2.f32 {%f753, %f754}, [%rd1782+80]; fma.rn.ftz.f32 %f945, %f754, %f752, %f753; mov.u32 %r3598, %r472; bra.uni $L__BB6_591; $L__BB6_603: and.b32 %r743, %r471, -4096; ld.local.u32 %r2411, [%rd2+24]; setp.ne.s32 %p378, %r743, %r2411; mov.u16 %rs409, 0; @%p378 bra $L__BB6_606; and.b32 %r2412, %r732, -4096; ld.local.u32 %r2413, [%rd2+28]; setp.ne.s32 %p379, %r2412, %r2413; @%p379 bra $L__BB6_606; and.b32 %r2414, %r741, -4096; ld.local.u32 %r2415, [%rd2+32]; setp.eq.s32 %p380, %r2414, %r2415; selp.u16 %rs409, 1, 0, %p380; $L__BB6_606: setp.eq.s16 %p381, %rs409, 0; @%p381 bra $L__BB6_612; shl.b32 %r2416, %r471, 3; and.b32 %r2417, %r2416, 31744; and.b32 %r2418, %r732, 3968; shr.u32 %r2419, %r2418, 2; or.b32 %r2420, %r2419, %r2417; bfe.u32 %r2421, %r741, 7, 5; or.b32 %r2422, %r2421, %r2420; shr.u32 %r2423, %r2420, 6; ld.local.u64 %rd484, [%rd2+64]; mul.wide.u32 %rd1831, %r2423, 8; add.s64 %rd1832, %rd484, %rd1831; and.b32 %r2424, %r2422, 63; mov.u64 %rd1833, 1; shl.b64 %rd1834, %rd1833, %r2424; ld.u64 %rd1835, [%rd1832+4128]; and.b64 %rd1836, %rd1835, %rd1834; setp.eq.s64 %p382, %rd1836, 0; mul.wide.u32 %rd1837, %r2422, 8; add.s64 %rd1838, %rd484, %rd1837; add.s64 %rd485, %rd1838, 8256; @%p382 bra $L__BB6_611; ld.u64 %rd1839, [%rd485]; add.s64 %rd486, %rd484, %rd1839; st.local.u32 [%rd2+12], %r742; and.b32 %r2425, %r732, -128; st.local.u32 [%rd2+16], %r2425; and.b32 %r2426, %r741, -128; st.local.u32 [%rd2+20], %r2426; st.local.u64 [%rd2+56], %rd486; shl.b32 %r2427, %r471, 5; and.b32 %r2428, %r2427, 3840; shl.b32 %r2429, %r732, 1; and.b32 %r2430, %r2429, 240; or.b32 %r2431, %r2430, %r2428; bfe.u32 %r2432, %r741, 3, 4; or.b32 %r2433, %r2432, %r2431; shr.u32 %r2434, %r2431, 6; mul.wide.u32 %rd1840, %r2434, 8; add.s64 %rd1841, %rd486, %rd1840; and.b32 %r2435, %r2433, 63; shl.b64 %rd1843, %rd1833, %r2435; ld.u64 %rd1844, [%rd1841+544]; and.b64 %rd1845, %rd1844, %rd1843; setp.eq.s64 %p383, %rd1845, 0; mul.wide.u32 %rd1846, %r2433, 8; add.s64 %rd1847, %rd486, %rd1846; add.s64 %rd487, %rd1847, 1088; @%p383 bra $L__BB6_610; ld.u64 %rd1848, [%rd487]; add.s64 %rd1849, %rd486, %rd1848; and.b32 %r2436, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2436}; and.b32 %r2437, %r741, -8; st.local.u32 [%rd2+8], %r2437; st.local.u64 [%rd2+48], %rd1849; shl.b32 %r2438, %r471, 6; and.b32 %r2439, %r2438, 448; shl.b32 %r2440, %r732, 3; and.b32 %r2441, %r2440, 56; or.b32 %r2442, %r2441, %r2439; and.b32 %r2443, %r741, 7; or.b32 %r2444, %r2442, %r2443; mul.wide.u32 %rd1850, %r2444, 2; add.s64 %rd1851, %rd1849, %rd1850; ld.u16 %rs303, [%rd1851+96]; cvt.rn.f32.u16 %f772, %rs303; ld.v2.f32 {%f773, %f774}, [%rd1849+80]; fma.rn.ftz.f32 %f946, %f774, %f772, %f773; mov.u32 %r3598, %r472; bra.uni $L__BB6_626; $L__BB6_638: and.b32 %r750, %r471, -4096; ld.local.u32 %r2514, [%rd2+24]; setp.ne.s32 %p400, %r750, %r2514; mov.u16 %rs412, 0; @%p400 bra $L__BB6_641; and.b32 %r2515, %r748, -4096; ld.local.u32 %r2516, [%rd2+28]; setp.ne.s32 %p401, %r2515, %r2516; @%p401 bra $L__BB6_641; and.b32 %r2517, %r741, -4096; ld.local.u32 %r2518, [%rd2+32]; setp.eq.s32 %p402, %r2517, %r2518; selp.u16 %rs412, 1, 0, %p402; $L__BB6_641: setp.eq.s16 %p403, %rs412, 0; @%p403 bra $L__BB6_647; shl.b32 %r2519, %r471, 3; and.b32 %r2520, %r2519, 31744; and.b32 %r2521, %r748, 3968; shr.u32 %r2522, %r2521, 2; or.b32 %r2523, %r2522, %r2520; bfe.u32 %r2524, %r741, 7, 5; or.b32 %r2525, %r2523, %r2524; shr.u32 %r2526, %r2523, 6; ld.local.u64 %rd502, [%rd2+64]; mul.wide.u32 %rd1899, %r2526, 8; add.s64 %rd1900, %rd502, %rd1899; and.b32 %r2527, %r2525, 63; mov.u64 %rd1901, 1; shl.b64 %rd1902, %rd1901, %r2527; ld.u64 %rd1903, [%rd1900+4128]; and.b64 %rd1904, %rd1903, %rd1902; setp.eq.s64 %p404, %rd1904, 0; mul.wide.u32 %rd1905, %r2525, 8; add.s64 %rd1906, %rd502, %rd1905; add.s64 %rd503, %rd1906, 8256; @%p404 bra $L__BB6_646; ld.u64 %rd1907, [%rd503]; add.s64 %rd504, %rd502, %rd1907; st.local.u32 [%rd2+12], %r749; and.b32 %r2528, %r748, -128; st.local.u32 [%rd2+16], %r2528; and.b32 %r2529, %r741, -128; st.local.u32 [%rd2+20], %r2529; st.local.u64 [%rd2+56], %rd504; shl.b32 %r2530, %r471, 5; and.b32 %r2531, %r2530, 3840; shl.b32 %r2532, %r748, 1; and.b32 %r2533, %r2532, 240; or.b32 %r2534, %r2533, %r2531; bfe.u32 %r2535, %r741, 3, 4; or.b32 %r2536, %r2534, %r2535; shr.u32 %r2537, %r2534, 6; mul.wide.u32 %rd1908, %r2537, 8; add.s64 %rd1909, %rd504, %rd1908; and.b32 %r2538, %r2536, 63; shl.b64 %rd1911, %rd1901, %r2538; ld.u64 %rd1912, [%rd1909+544]; and.b64 %rd1913, %rd1912, %rd1911; setp.eq.s64 %p405, %rd1913, 0; mul.wide.u32 %rd1914, %r2536, 8; add.s64 %rd1915, %rd504, %rd1914; add.s64 %rd505, %rd1915, 1088; @%p405 bra $L__BB6_645; ld.u64 %rd1916, [%rd505]; add.s64 %rd1917, %rd504, %rd1916; and.b32 %r2539, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2539}; and.b32 %r2540, %r741, -8; st.local.u32 [%rd2+8], %r2540; st.local.u64 [%rd2+48], %rd1917; shl.b32 %r2541, %r471, 6; and.b32 %r2542, %r2541, 448; shl.b32 %r2543, %r748, 3; and.b32 %r2544, %r2543, 56; or.b32 %r2545, %r2544, %r2542; and.b32 %r2546, %r741, 7; or.b32 %r2547, %r2545, %r2546; mul.wide.u32 %rd1918, %r2547, 2; add.s64 %rd1919, %rd1917, %rd1918; ld.u16 %rs313, [%rd1919+96]; cvt.rn.f32.u16 %f793, %rs313; ld.v2.f32 {%f794, %f795}, [%rd1917+80]; fma.rn.ftz.f32 %f947, %f795, %f793, %f794; mov.u32 %r3598, %r472; bra.uni $L__BB6_661; $L__BB6_673: and.b32 %r756, %r471, -4096; ld.local.u32 %r2615, [%rd2+24]; setp.ne.s32 %p422, %r756, %r2615; mov.u16 %rs415, 0; @%p422 bra $L__BB6_676; and.b32 %r2616, %r748, -4096; ld.local.u32 %r2617, [%rd2+28]; setp.ne.s32 %p423, %r2616, %r2617; @%p423 bra $L__BB6_676; and.b32 %r2618, %r733, -4096; ld.local.u32 %r2619, [%rd2+32]; setp.eq.s32 %p424, %r2618, %r2619; selp.u16 %rs415, 1, 0, %p424; $L__BB6_676: setp.eq.s16 %p425, %rs415, 0; @%p425 bra $L__BB6_682; shl.b32 %r2620, %r471, 3; and.b32 %r2621, %r2620, 31744; and.b32 %r2622, %r748, 3968; shr.u32 %r2623, %r2622, 2; or.b32 %r2624, %r2623, %r2621; bfe.u32 %r2625, %r733, 7, 5; or.b32 %r2626, %r2624, %r2625; shr.u32 %r2627, %r2624, 6; ld.local.u64 %rd520, [%rd2+64]; mul.wide.u32 %rd1966, %r2627, 8; add.s64 %rd1967, %rd520, %rd1966; and.b32 %r2628, %r2626, 63; mov.u64 %rd1968, 1; shl.b64 %rd1969, %rd1968, %r2628; ld.u64 %rd1970, [%rd1967+4128]; and.b64 %rd1971, %rd1970, %rd1969; setp.eq.s64 %p426, %rd1971, 0; mul.wide.u32 %rd1972, %r2626, 8; add.s64 %rd1973, %rd520, %rd1972; add.s64 %rd521, %rd1973, 8256; @%p426 bra $L__BB6_681; ld.u64 %rd1974, [%rd521]; add.s64 %rd522, %rd520, %rd1974; st.local.u32 [%rd2+12], %r755; and.b32 %r2629, %r748, -128; st.local.u32 [%rd2+16], %r2629; and.b32 %r2630, %r733, -128; st.local.u32 [%rd2+20], %r2630; st.local.u64 [%rd2+56], %rd522; shl.b32 %r2631, %r471, 5; and.b32 %r2632, %r2631, 3840; shl.b32 %r2633, %r748, 1; and.b32 %r2634, %r2633, 240; or.b32 %r2635, %r2634, %r2632; bfe.u32 %r2636, %r733, 3, 4; or.b32 %r2637, %r2635, %r2636; shr.u32 %r2638, %r2635, 6; mul.wide.u32 %rd1975, %r2638, 8; add.s64 %rd1976, %rd522, %rd1975; and.b32 %r2639, %r2637, 63; shl.b64 %rd1978, %rd1968, %r2639; ld.u64 %rd1979, [%rd1976+544]; and.b64 %rd1980, %rd1979, %rd1978; setp.eq.s64 %p427, %rd1980, 0; mul.wide.u32 %rd1981, %r2637, 8; add.s64 %rd1982, %rd522, %rd1981; add.s64 %rd523, %rd1982, 1088; @%p427 bra $L__BB6_680; ld.u64 %rd1983, [%rd523]; add.s64 %rd1984, %rd522, %rd1983; and.b32 %r2640, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2640}; and.b32 %r2641, %r733, -8; st.local.u32 [%rd2+8], %r2641; st.local.u64 [%rd2+48], %rd1984; and.b32 %r2642, %r471, 7; shl.b32 %r2643, %r748, 3; and.b32 %r2644, %r2643, 56; and.b32 %r2645, %r733, 7; bfi.b32 %r2646, %r2642, %r2645, 6, 3; or.b32 %r2647, %r2646, %r2644; mul.wide.u32 %rd1985, %r2647, 2; add.s64 %rd1986, %rd1984, %rd1985; ld.u16 %rs323, [%rd1986+96]; cvt.rn.f32.u16 %f813, %rs323; ld.v2.f32 {%f814, %f815}, [%rd1984+80]; fma.rn.ftz.f32 %f948, %f815, %f813, %f814; mov.u32 %r3598, %r472; bra.uni $L__BB6_696; $L__BB6_708: and.b32 %r762, %r635, -4096; ld.local.u32 %r2714, [%rd2+24]; setp.ne.s32 %p444, %r762, %r2714; mov.u16 %rs418, 0; @%p444 bra $L__BB6_711; and.b32 %r2715, %r748, -4096; ld.local.u32 %r2716, [%rd2+28]; setp.ne.s32 %p445, %r2715, %r2716; @%p445 bra $L__BB6_711; and.b32 %r2717, %r733, -4096; ld.local.u32 %r2718, [%rd2+32]; setp.eq.s32 %p446, %r2717, %r2718; selp.u16 %rs418, 1, 0, %p446; $L__BB6_711: setp.eq.s16 %p447, %rs418, 0; @%p447 bra $L__BB6_717; shl.b32 %r2719, %r635, 3; and.b32 %r2720, %r2719, 31744; and.b32 %r2721, %r748, 3968; shr.u32 %r2722, %r2721, 2; or.b32 %r2723, %r2722, %r2720; bfe.u32 %r2724, %r733, 7, 5; or.b32 %r2725, %r2723, %r2724; shr.u32 %r2726, %r2723, 6; ld.local.u64 %rd538, [%rd2+64]; mul.wide.u32 %rd2033, %r2726, 8; add.s64 %rd2034, %rd538, %rd2033; and.b32 %r2727, %r2725, 63; mov.u64 %rd2035, 1; shl.b64 %rd2036, %rd2035, %r2727; ld.u64 %rd2037, [%rd2034+4128]; and.b64 %rd2038, %rd2037, %rd2036; setp.eq.s64 %p448, %rd2038, 0; mul.wide.u32 %rd2039, %r2725, 8; add.s64 %rd2040, %rd538, %rd2039; add.s64 %rd539, %rd2040, 8256; @%p448 bra $L__BB6_716; ld.u64 %rd2041, [%rd539]; add.s64 %rd540, %rd538, %rd2041; st.local.u32 [%rd2+12], %r761; and.b32 %r2728, %r748, -128; st.local.u32 [%rd2+16], %r2728; and.b32 %r2729, %r733, -128; st.local.u32 [%rd2+20], %r2729; st.local.u64 [%rd2+56], %rd540; shl.b32 %r2730, %r635, 5; and.b32 %r2731, %r2730, 3840; shl.b32 %r2732, %r748, 1; and.b32 %r2733, %r2732, 240; or.b32 %r2734, %r2733, %r2731; bfe.u32 %r2735, %r733, 3, 4; or.b32 %r2736, %r2734, %r2735; shr.u32 %r2737, %r2734, 6; mul.wide.u32 %rd2042, %r2737, 8; add.s64 %rd2043, %rd540, %rd2042; and.b32 %r2738, %r2736, 63; shl.b64 %rd2045, %rd2035, %r2738; ld.u64 %rd2046, [%rd2043+544]; and.b64 %rd2047, %rd2046, %rd2045; setp.eq.s64 %p449, %rd2047, 0; mul.wide.u32 %rd2048, %r2736, 8; add.s64 %rd2049, %rd540, %rd2048; add.s64 %rd541, %rd2049, 1088; @%p449 bra $L__BB6_715; ld.u64 %rd2050, [%rd541]; add.s64 %rd2051, %rd540, %rd2050; and.b32 %r2739, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2739}; and.b32 %r2740, %r733, -8; st.local.u32 [%rd2+8], %r2740; st.local.u64 [%rd2+48], %rd2051; and.b32 %r2741, %r635, 7; shl.b32 %r2742, %r748, 3; and.b32 %r2743, %r2742, 56; and.b32 %r2744, %r733, 7; bfi.b32 %r2745, %r2741, %r2744, 6, 3; or.b32 %r2746, %r2745, %r2743; mul.wide.u32 %rd2052, %r2746, 2; add.s64 %rd2053, %rd2051, %rd2052; ld.u16 %rs333, [%rd2053+96]; cvt.rn.f32.u16 %f836, %rs333; ld.v2.f32 {%f837, %f838}, [%rd2051+80]; fma.rn.ftz.f32 %f949, %f838, %f836, %f837; mov.u32 %r3598, %r636; bra.uni $L__BB6_731; $L__BB6_743: and.b32 %r768, %r635, -4096; ld.local.u32 %r2815, [%rd2+24]; setp.ne.s32 %p466, %r768, %r2815; mov.u16 %rs421, 0; @%p466 bra $L__BB6_746; and.b32 %r2816, %r748, -4096; ld.local.u32 %r2817, [%rd2+28]; setp.ne.s32 %p467, %r2816, %r2817; @%p467 bra $L__BB6_746; and.b32 %r2818, %r741, -4096; ld.local.u32 %r2819, [%rd2+32]; setp.eq.s32 %p468, %r2818, %r2819; selp.u16 %rs421, 1, 0, %p468; $L__BB6_746: setp.eq.s16 %p469, %rs421, 0; @%p469 bra $L__BB6_752; shl.b32 %r2820, %r635, 3; and.b32 %r2821, %r2820, 31744; and.b32 %r2822, %r748, 3968; shr.u32 %r2823, %r2822, 2; or.b32 %r2824, %r2823, %r2821; bfe.u32 %r2825, %r741, 7, 5; or.b32 %r2826, %r2824, %r2825; shr.u32 %r2827, %r2824, 6; ld.local.u64 %rd556, [%rd2+64]; mul.wide.u32 %rd2100, %r2827, 8; add.s64 %rd2101, %rd556, %rd2100; and.b32 %r2828, %r2826, 63; mov.u64 %rd2102, 1; shl.b64 %rd2103, %rd2102, %r2828; ld.u64 %rd2104, [%rd2101+4128]; and.b64 %rd2105, %rd2104, %rd2103; setp.eq.s64 %p470, %rd2105, 0; mul.wide.u32 %rd2106, %r2826, 8; add.s64 %rd2107, %rd556, %rd2106; add.s64 %rd557, %rd2107, 8256; @%p470 bra $L__BB6_751; ld.u64 %rd2108, [%rd557]; add.s64 %rd558, %rd556, %rd2108; st.local.u32 [%rd2+12], %r767; and.b32 %r2829, %r748, -128; st.local.u32 [%rd2+16], %r2829; and.b32 %r2830, %r741, -128; st.local.u32 [%rd2+20], %r2830; st.local.u64 [%rd2+56], %rd558; shl.b32 %r2831, %r635, 5; and.b32 %r2832, %r2831, 3840; shl.b32 %r2833, %r748, 1; and.b32 %r2834, %r2833, 240; or.b32 %r2835, %r2834, %r2832; bfe.u32 %r2836, %r741, 3, 4; or.b32 %r2837, %r2835, %r2836; shr.u32 %r2838, %r2835, 6; mul.wide.u32 %rd2109, %r2838, 8; add.s64 %rd2110, %rd558, %rd2109; and.b32 %r2839, %r2837, 63; shl.b64 %rd2112, %rd2102, %r2839; ld.u64 %rd2113, [%rd2110+544]; and.b64 %rd2114, %rd2113, %rd2112; setp.eq.s64 %p471, %rd2114, 0; mul.wide.u32 %rd2115, %r2837, 8; add.s64 %rd2116, %rd558, %rd2115; add.s64 %rd559, %rd2116, 1088; @%p471 bra $L__BB6_750; ld.u64 %rd2117, [%rd559]; add.s64 %rd2118, %rd558, %rd2117; and.b32 %r2840, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2840}; and.b32 %r2841, %r741, -8; st.local.u32 [%rd2+8], %r2841; st.local.u64 [%rd2+48], %rd2118; shl.b32 %r2842, %r635, 6; and.b32 %r2843, %r2842, 448; shl.b32 %r2844, %r748, 3; and.b32 %r2845, %r2844, 56; or.b32 %r2846, %r2845, %r2843; and.b32 %r2847, %r741, 7; or.b32 %r2848, %r2846, %r2847; mul.wide.u32 %rd2119, %r2848, 2; add.s64 %rd2120, %rd2118, %rd2119; ld.u16 %rs343, [%rd2120+96]; cvt.rn.f32.u16 %f856, %rs343; ld.v2.f32 {%f857, %f858}, [%rd2118+80]; fma.rn.ftz.f32 %f950, %f858, %f856, %f857; mov.u32 %r3598, %r636; bra.uni $L__BB6_766; $L__BB6_778: and.b32 %r774, %r635, -4096; ld.local.u32 %r2918, [%rd2+24]; setp.ne.s32 %p488, %r774, %r2918; mov.u16 %rs424, 0; @%p488 bra $L__BB6_781; and.b32 %r2919, %r732, -4096; ld.local.u32 %r2920, [%rd2+28]; setp.ne.s32 %p489, %r2919, %r2920; @%p489 bra $L__BB6_781; and.b32 %r2921, %r741, -4096; ld.local.u32 %r2922, [%rd2+32]; setp.eq.s32 %p490, %r2921, %r2922; selp.u16 %rs424, 1, 0, %p490; $L__BB6_781: setp.eq.s16 %p491, %rs424, 0; @%p491 bra $L__BB6_787; shl.b32 %r2923, %r635, 3; and.b32 %r2924, %r2923, 31744; and.b32 %r2925, %r732, 3968; shr.u32 %r2926, %r2925, 2; or.b32 %r2927, %r2924, %r2926; bfe.u32 %r2928, %r741, 7, 5; or.b32 %r2929, %r2928, %r2927; shr.u32 %r2930, %r2927, 6; ld.local.u64 %rd574, [%rd2+64]; mul.wide.u32 %rd2167, %r2930, 8; add.s64 %rd2168, %rd574, %rd2167; and.b32 %r2931, %r2929, 63; mov.u64 %rd2169, 1; shl.b64 %rd2170, %rd2169, %r2931; ld.u64 %rd2171, [%rd2168+4128]; and.b64 %rd2172, %rd2171, %rd2170; setp.eq.s64 %p492, %rd2172, 0; mul.wide.u32 %rd2173, %r2929, 8; add.s64 %rd2174, %rd574, %rd2173; add.s64 %rd575, %rd2174, 8256; @%p492 bra $L__BB6_786; ld.u64 %rd2175, [%rd575]; add.s64 %rd576, %rd574, %rd2175; st.local.u32 [%rd2+12], %r773; and.b32 %r2932, %r732, -128; st.local.u32 [%rd2+16], %r2932; and.b32 %r2933, %r741, -128; st.local.u32 [%rd2+20], %r2933; st.local.u64 [%rd2+56], %rd576; shl.b32 %r2934, %r635, 5; and.b32 %r2935, %r2934, 3840; shl.b32 %r2936, %r732, 1; and.b32 %r2937, %r2936, 240; or.b32 %r2938, %r2935, %r2937; bfe.u32 %r2939, %r741, 3, 4; or.b32 %r2940, %r2939, %r2938; shr.u32 %r2941, %r2938, 6; mul.wide.u32 %rd2176, %r2941, 8; add.s64 %rd2177, %rd576, %rd2176; and.b32 %r2942, %r2940, 63; shl.b64 %rd2179, %rd2169, %r2942; ld.u64 %rd2180, [%rd2177+544]; and.b64 %rd2181, %rd2180, %rd2179; setp.eq.s64 %p493, %rd2181, 0; mul.wide.u32 %rd2182, %r2940, 8; add.s64 %rd2183, %rd576, %rd2182; add.s64 %rd577, %rd2183, 1088; @%p493 bra $L__BB6_785; ld.u64 %rd2184, [%rd577]; add.s64 %rd2185, %rd576, %rd2184; and.b32 %r2943, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2943}; and.b32 %r2944, %r741, -8; st.local.u32 [%rd2+8], %r2944; st.local.u64 [%rd2+48], %rd2185; shl.b32 %r2945, %r635, 6; and.b32 %r2946, %r2945, 448; shl.b32 %r2947, %r732, 3; and.b32 %r2948, %r2947, 56; or.b32 %r2949, %r2946, %r2948; and.b32 %r2950, %r741, 7; or.b32 %r2951, %r2949, %r2950; mul.wide.u32 %rd2186, %r2951, 2; add.s64 %rd2187, %rd2185, %rd2186; ld.u16 %rs353, [%rd2187+96]; cvt.rn.f32.u16 %f877, %rs353; ld.v2.f32 {%f878, %f879}, [%rd2185+80]; fma.rn.ftz.f32 %f951, %f879, %f877, %f878; mov.u32 %r3598, %r636; bra.uni $L__BB6_801; $L__BB6_821: ld.f32 %f952, [%rd599]; bra.uni $L__BB6_833; $L__BB6_42: ld.f32 %f930, [%rd43]; bra.uni $L__BB6_66; $L__BB6_77: ld.f32 %f931, [%rd72]; bra.uni $L__BB6_101; $L__BB6_112: ld.f32 %f932, [%rd101]; bra.uni $L__BB6_136; $L__BB6_147: ld.f32 %f933, [%rd130]; bra.uni $L__BB6_171; $L__BB6_182: ld.f32 %f934, [%rd159]; bra.uni $L__BB6_206; $L__BB6_217: ld.f32 %f935, [%rd188]; bra.uni $L__BB6_241; $L__BB6_252: ld.f32 %f936, [%rd217]; bra.uni $L__BB6_276; $L__BB6_287: ld.f32 %f937, [%rd246]; bra.uni $L__BB6_311; $L__BB6_322: ld.f32 %f938, [%rd275]; bra.uni $L__BB6_346; $L__BB6_357: ld.f32 %f939, [%rd304]; bra.uni $L__BB6_381; $L__BB6_392: ld.f32 %f940, [%rd333]; bra.uni $L__BB6_416; $L__BB6_427: ld.f32 %f941, [%rd362]; bra.uni $L__BB6_451; $L__BB6_462: ld.f32 %f942, [%rd391]; bra.uni $L__BB6_486; $L__BB6_497: ld.f32 %f943, [%rd420]; bra.uni $L__BB6_521; $L__BB6_532: ld.f32 %f944, [%rd449]; bra.uni $L__BB6_556; $L__BB6_567: ld.f32 %f945, [%rd465]; bra.uni $L__BB6_591; $L__BB6_602: ld.f32 %f946, [%rd483]; bra.uni $L__BB6_626; $L__BB6_637: ld.f32 %f947, [%rd501]; bra.uni $L__BB6_661; $L__BB6_672: ld.f32 %f948, [%rd519]; bra.uni $L__BB6_696; $L__BB6_707: ld.f32 %f949, [%rd537]; bra.uni $L__BB6_731; $L__BB6_742: ld.f32 %f950, [%rd555]; bra.uni $L__BB6_766; $L__BB6_777: ld.f32 %f951, [%rd573]; bra.uni $L__BB6_801; $L__BB6_820: ld.f32 %f952, [%rd601]; bra.uni $L__BB6_833; $L__BB6_52: ld.u32 %r81, [%rd5]; mov.u64 %rd2302, 0; setp.eq.s32 %p32, %r81, 0; @%p32 bra $L__BB6_57; shr.u32 %r3092, %r7, 12; cvt.u64.u32 %rd2293, %r3092; shl.b64 %rd2292, %rd2293, 42; or.b64 %rd736, %rd19, %rd2292; shr.u32 %r1051, %r74, 12; cvt.u64.u32 %rd737, %r1051; or.b64 %rd51, %rd736, %rd737; mov.u32 %r3181, 0; $L__BB6_54: mul.wide.u32 %rd738, %r3181, 32; add.s64 %rd739, %rd17, %rd738; ld.u64 %rd740, [%rd739]; setp.eq.s64 %p33, %rd740, %rd51; @%p33 bra $L__BB6_56; add.s32 %r3181, %r3181, 1; setp.lt.u32 %p34, %r3181, %r81; @%p34 bra $L__BB6_54; bra.uni $L__BB6_57; $L__BB6_87: ld.u32 %r146, [%rd5]; mov.u64 %rd2308, 0; setp.eq.s32 %p54, %r146, 0; @%p54 bra $L__BB6_92; shr.u32 %r3091, %r7, 12; cvt.u64.u32 %rd2290, %r3091; shl.b64 %rd2289, %rd2290, 42; shr.u32 %r1136, %r74, 12; cvt.u64.u32 %rd805, %r1136; or.b64 %rd806, %rd2289, %rd805; shr.u32 %r1137, %r139, 12; mul.wide.u32 %rd807, %r1137, 2097152; or.b64 %rd80, %rd806, %rd807; mov.u32 %r3230, 0; $L__BB6_89: mul.wide.u32 %rd808, %r3230, 32; add.s64 %rd809, %rd17, %rd808; ld.u64 %rd810, [%rd809]; setp.eq.s64 %p55, %rd810, %rd80; @%p55 bra $L__BB6_91; add.s32 %r3230, %r3230, 1; setp.lt.u32 %p56, %r3230, %r146; @%p56 bra $L__BB6_89; bra.uni $L__BB6_92; $L__BB6_122: ld.u32 %r207, [%rd5]; mov.u64 %rd2314, 0; setp.eq.s32 %p76, %r207, 0; @%p76 bra $L__BB6_127; shr.u32 %r1219, %r139, 12; mul.wide.u32 %rd875, %r1219, 2097152; or.b64 %rd109, %rd21, %rd875; mov.u32 %r3276, 0; $L__BB6_124: mul.wide.u32 %rd876, %r3276, 32; add.s64 %rd877, %rd17, %rd876; ld.u64 %rd878, [%rd877]; setp.eq.s64 %p77, %rd878, %rd109; @%p77 bra $L__BB6_126; add.s32 %r3276, %r3276, 1; setp.lt.u32 %p78, %r3276, %r207; @%p78 bra $L__BB6_124; bra.uni $L__BB6_127; $L__BB6_157: ld.u32 %r269, [%rd5]; mov.u64 %rd2320, 0; setp.eq.s32 %p98, %r269, 0; @%p98 bra $L__BB6_162; shr.u32 %r1300, %r259, 12; cvt.u64.u32 %rd943, %r1300; shl.b64 %rd944, %rd943, 42; or.b64 %rd945, %rd944, %rd18; shr.u32 %r1301, %r139, 12; mul.wide.u32 %rd946, %r1301, 2097152; or.b64 %rd138, %rd945, %rd946; mov.u32 %r3319, 0; $L__BB6_159: mul.wide.u32 %rd947, %r3319, 32; add.s64 %rd948, %rd17, %rd947; ld.u64 %rd949, [%rd948]; setp.eq.s64 %p99, %rd949, %rd138; @%p99 bra $L__BB6_161; add.s32 %r3319, %r3319, 1; setp.lt.u32 %p100, %r3319, %r269; @%p100 bra $L__BB6_159; bra.uni $L__BB6_162; $L__BB6_192: ld.u32 %r326, [%rd5]; mov.u64 %rd2326, 0; setp.eq.s32 %p120, %r326, 0; @%p120 bra $L__BB6_197; shr.u32 %r1385, %r259, 12; cvt.u64.u32 %rd1014, %r1385; shr.u32 %r1386, %r74, 12; cvt.u64.u32 %rd1015, %r1386; bfi.b64 %rd1016, %rd1014, %rd1015, 42, 22; shr.u32 %r1387, %r139, 12; mul.wide.u32 %rd1017, %r1387, 2097152; or.b64 %rd167, %rd1016, %rd1017; mov.u32 %r3359, 0; $L__BB6_194: cvt.u64.u32 %rd168, %r3359; mul.wide.u32 %rd1018, %r3359, 32; add.s64 %rd1019, %rd17, %rd1018; ld.u64 %rd1020, [%rd1019]; setp.eq.s64 %p121, %rd1020, %rd167; add.s32 %r3359, %r3359, 1; @%p121 bra $L__BB6_196; setp.lt.u32 %p122, %r3359, %r326; @%p122 bra $L__BB6_194; bra.uni $L__BB6_197; $L__BB6_227: ld.u32 %r380, [%rd5]; mov.u64 %rd2332, 0; setp.eq.s32 %p142, %r380, 0; @%p142 bra $L__BB6_232; shr.u32 %r1472, %r259, 12; cvt.u64.u32 %rd1085, %r1472; shl.b64 %rd1086, %rd1085, 42; or.b64 %rd1087, %rd1086, %rd19; shr.u32 %r1473, %r74, 12; cvt.u64.u32 %rd1088, %r1473; or.b64 %rd196, %rd1087, %rd1088; mov.u32 %r3396, 0; $L__BB6_229: cvt.u64.u32 %rd197, %r3396; mul.wide.u32 %rd1089, %r3396, 32; add.s64 %rd1090, %rd17, %rd1089; ld.u64 %rd1091, [%rd1090]; setp.eq.s64 %p143, %rd1091, %rd196; add.s32 %r3396, %r3396, 1; @%p143 bra $L__BB6_231; setp.lt.u32 %p144, %r3396, %r380; @%p144 bra $L__BB6_229; bra.uni $L__BB6_232; $L__BB6_262: ld.u32 %r431, [%rd5]; mov.u64 %rd2338, 0; setp.eq.s32 %p164, %r431, 0; @%p164 bra $L__BB6_267; shr.u32 %r1558, %r259, 12; cvt.u64.u32 %rd1156, %r1558; shl.b64 %rd1157, %rd1156, 42; or.b64 %rd1158, %rd19, %rd18; or.b64 %rd225, %rd1158, %rd1157; mov.u32 %r3430, 0; $L__BB6_264: cvt.u64.u32 %rd226, %r3430; mul.wide.u32 %rd1159, %r3430, 32; add.s64 %rd1160, %rd17, %rd1159; ld.u64 %rd1161, [%rd1160]; setp.eq.s64 %p165, %rd1161, %rd225; add.s32 %r3430, %r3430, 1; @%p165 bra $L__BB6_266; setp.lt.u32 %p166, %r3430, %r431; @%p166 bra $L__BB6_264; bra.uni $L__BB6_267; $L__BB6_297: ld.u32 %r481, [%rd5]; mov.u64 %rd2344, 0; setp.eq.s32 %p186, %r481, 0; @%p186 bra $L__BB6_302; shr.u32 %r1643, %r471, 12; cvt.u64.u32 %rd1226, %r1643; shl.b64 %rd1227, %rd1226, 42; or.b64 %rd1228, %rd1227, %rd18; or.b64 %rd254, %rd1228, %rd19; mov.u32 %r3461, 0; $L__BB6_299: cvt.u64.u32 %rd255, %r3461; mul.wide.u32 %rd1229, %r3461, 32; add.s64 %rd1230, %rd17, %rd1229; ld.u64 %rd1231, [%rd1230]; setp.eq.s64 %p187, %rd1231, %rd254; add.s32 %r3461, %r3461, 1; @%p187 bra $L__BB6_301; setp.lt.u32 %p188, %r3461, %r481; @%p188 bra $L__BB6_299; bra.uni $L__BB6_302; $L__BB6_332: ld.u32 %r526, [%rd5]; mov.u64 %rd2350, 0; setp.eq.s32 %p208, %r526, 0; @%p208 bra $L__BB6_337; shr.u32 %r1728, %r471, 12; cvt.u64.u32 %rd1296, %r1728; shl.b64 %rd1297, %rd1296, 42; or.b64 %rd1298, %rd19, %rd1297; shr.u32 %r1729, %r74, 12; cvt.u64.u32 %rd1299, %r1729; or.b64 %rd283, %rd1298, %rd1299; mov.u32 %r3489, 0; $L__BB6_334: cvt.u64.u32 %rd284, %r3489; mul.wide.u32 %rd1300, %r3489, 32; add.s64 %rd1301, %rd17, %rd1300; ld.u64 %rd1302, [%rd1301]; setp.eq.s64 %p209, %rd1302, %rd283; add.s32 %r3489, %r3489, 1; @%p209 bra $L__BB6_336; setp.lt.u32 %p210, %r3489, %r526; @%p210 bra $L__BB6_334; bra.uni $L__BB6_337; $L__BB6_367: ld.u32 %r568, [%rd5]; mov.u64 %rd2356, 0; setp.eq.s32 %p230, %r568, 0; @%p230 bra $L__BB6_372; shr.u32 %r1814, %r471, 12; cvt.u64.u32 %rd1367, %r1814; shr.u32 %r1815, %r74, 12; cvt.u64.u32 %rd1368, %r1815; bfi.b64 %rd1369, %rd1367, %rd1368, 42, 22; shr.u32 %r1816, %r139, 12; mul.wide.u32 %rd1370, %r1816, 2097152; or.b64 %rd312, %rd1369, %rd1370; mov.u32 %r3514, 0; $L__BB6_369: cvt.u64.u32 %rd313, %r3514; mul.wide.u32 %rd1371, %r3514, 32; add.s64 %rd1372, %rd17, %rd1371; ld.u64 %rd1373, [%rd1372]; setp.eq.s64 %p231, %rd1373, %rd312; add.s32 %r3514, %r3514, 1; @%p231 bra $L__BB6_371; setp.lt.u32 %p232, %r3514, %r568; @%p232 bra $L__BB6_369; bra.uni $L__BB6_372; $L__BB6_402: ld.u32 %r607, [%rd5]; mov.u64 %rd2362, 0; setp.eq.s32 %p252, %r607, 0; @%p252 bra $L__BB6_407; shr.u32 %r1898, %r471, 12; cvt.u64.u32 %rd1438, %r1898; shl.b64 %rd1439, %rd1438, 42; or.b64 %rd1440, %rd1439, %rd18; shr.u32 %r1899, %r139, 12; mul.wide.u32 %rd1441, %r1899, 2097152; or.b64 %rd341, %rd1440, %rd1441; mov.u32 %r3536, 0; $L__BB6_404: cvt.u64.u32 %rd342, %r3536; mul.wide.u32 %rd1442, %r3536, 32; add.s64 %rd1443, %rd17, %rd1442; ld.u64 %rd1444, [%rd1443]; setp.eq.s64 %p253, %rd1444, %rd341; add.s32 %r3536, %r3536, 1; @%p253 bra $L__BB6_406; setp.lt.u32 %p254, %r3536, %r607; @%p254 bra $L__BB6_404; bra.uni $L__BB6_407; $L__BB6_437: ld.u32 %r645, [%rd5]; mov.u64 %rd2368, 0; setp.eq.s32 %p274, %r645, 0; @%p274 bra $L__BB6_442; shr.u32 %r1980, %r635, 12; cvt.u64.u32 %rd1509, %r1980; shl.b64 %rd1510, %rd1509, 42; or.b64 %rd1511, %rd1510, %rd18; shr.u32 %r1981, %r139, 12; mul.wide.u32 %rd1512, %r1981, 2097152; or.b64 %rd370, %rd1511, %rd1512; mov.u32 %r3555, 0; $L__BB6_439: cvt.u64.u32 %rd371, %r3555; mul.wide.u32 %rd1513, %r3555, 32; add.s64 %rd1514, %rd17, %rd1513; ld.u64 %rd1515, [%rd1514]; setp.eq.s64 %p275, %rd1515, %rd370; add.s32 %r3555, %r3555, 1; @%p275 bra $L__BB6_441; setp.lt.u32 %p276, %r3555, %r645; @%p276 bra $L__BB6_439; bra.uni $L__BB6_442; $L__BB6_472: ld.u32 %r678, [%rd5]; mov.u64 %rd2374, 0; setp.eq.s32 %p296, %r678, 0; @%p296 bra $L__BB6_477; shr.u32 %r2065, %r635, 12; cvt.u64.u32 %rd1580, %r2065; shr.u32 %r2066, %r74, 12; cvt.u64.u32 %rd1581, %r2066; bfi.b64 %rd1582, %rd1580, %rd1581, 42, 22; shr.u32 %r2067, %r139, 12; mul.wide.u32 %rd1583, %r2067, 2097152; or.b64 %rd399, %rd1582, %rd1583; mov.u32 %r3571, 0; $L__BB6_474: cvt.u64.u32 %rd400, %r3571; mul.wide.u32 %rd1584, %r3571, 32; add.s64 %rd1585, %rd17, %rd1584; ld.u64 %rd1586, [%rd1585]; setp.eq.s64 %p297, %rd1586, %rd399; add.s32 %r3571, %r3571, 1; @%p297 bra $L__BB6_476; setp.lt.u32 %p298, %r3571, %r678; @%p298 bra $L__BB6_474; bra.uni $L__BB6_477; $L__BB6_507: ld.u32 %r708, [%rd5]; mov.u64 %rd2380, 0; setp.eq.s32 %p318, %r708, 0; @%p318 bra $L__BB6_512; shr.u32 %r2152, %r635, 12; cvt.u64.u32 %rd1651, %r2152; shl.b64 %rd1652, %rd1651, 42; or.b64 %rd1653, %rd1652, %rd19; shr.u32 %r2153, %r74, 12; cvt.u64.u32 %rd1654, %r2153; or.b64 %rd428, %rd1653, %rd1654; mov.u32 %r3584, 0; $L__BB6_509: cvt.u64.u32 %rd429, %r3584; mul.wide.u32 %rd1655, %r3584, 32; add.s64 %rd1656, %rd17, %rd1655; ld.u64 %rd1657, [%rd1656]; setp.eq.s64 %p319, %rd1657, %rd428; add.s32 %r3584, %r3584, 1; @%p319 bra $L__BB6_511; setp.lt.u32 %p320, %r3584, %r708; @%p320 bra $L__BB6_509; bra.uni $L__BB6_512; $L__BB6_542: or.b64 %rd453, %rd19, %rd18; ld.u32 %r729, [%rd5]; mov.u64 %rd2386, 0; setp.eq.s32 %p340, %r729, 0; @%p340 bra $L__BB6_547; shr.u32 %r2242, %r635, 12; cvt.u64.u32 %rd1720, %r2242; shl.b64 %rd1721, %rd1720, 42; or.b64 %rd454, %rd453, %rd1721; mov.u32 %r3594, 0; $L__BB6_544: cvt.u64.u32 %rd455, %r3594; mul.wide.u32 %rd1722, %r3594, 32; add.s64 %rd1723, %rd17, %rd1722; ld.u64 %rd1724, [%rd1723]; setp.eq.s64 %p341, %rd1724, %rd454; add.s32 %r3594, %r3594, 1; @%p341 bra $L__BB6_546; setp.lt.u32 %p342, %r3594, %r729; @%p342 bra $L__BB6_544; bra.uni $L__BB6_547; $L__BB6_577: ld.local.u64 %rd470, [%rd2+40]; mov.u64 %rd2387, 0; ld.u32 %r737, [%rd470+24]; setp.eq.s32 %p362, %r737, 0; @%p362 bra $L__BB6_582; add.s64 %rd471, %rd470, 64; shr.u32 %r2343, %r471, 12; cvt.u64.u32 %rd1786, %r2343; shr.u32 %r2344, %r733, 12; cvt.u64.u32 %rd1787, %r2344; bfi.b64 %rd1788, %rd1786, %rd1787, 42, 22; shr.u32 %r2345, %r732, 12; mul.wide.u32 %rd1789, %r2345, 2097152; or.b64 %rd472, %rd1788, %rd1789; mov.u32 %r3595, 0; $L__BB6_579: cvt.u64.u32 %rd473, %r3595; mul.wide.u32 %rd1790, %r3595, 32; add.s64 %rd1791, %rd471, %rd1790; ld.u64 %rd1792, [%rd1791]; setp.eq.s64 %p363, %rd1792, %rd472; add.s32 %r3595, %r3595, 1; @%p363 bra $L__BB6_581; setp.lt.u32 %p364, %r3595, %r737; @%p364 bra $L__BB6_579; bra.uni $L__BB6_582; $L__BB6_612: ld.local.u64 %rd488, [%rd2+40]; mov.u64 %rd2388, 0; ld.u32 %r744, [%rd488+24]; setp.eq.s32 %p384, %r744, 0; @%p384 bra $L__BB6_617; add.s64 %rd489, %rd488, 64; shr.u32 %r2446, %r471, 12; cvt.u64.u32 %rd1853, %r2446; shl.b64 %rd1854, %rd1853, 42; shr.u32 %r2447, %r732, 12; mul.wide.u32 %rd1855, %r2447, 2097152; or.b64 %rd1856, %rd1855, %rd1854; shr.u32 %r2448, %r741, 12; cvt.u64.u32 %rd1857, %r2448; or.b64 %rd490, %rd1856, %rd1857; mov.u32 %r3597, 0; $L__BB6_614: cvt.u64.u32 %rd491, %r3597; mul.wide.u32 %rd1858, %r3597, 32; add.s64 %rd1859, %rd489, %rd1858; ld.u64 %rd1860, [%rd1859]; setp.eq.s64 %p385, %rd1860, %rd490; add.s32 %r3597, %r3597, 1; @%p385 bra $L__BB6_616; setp.lt.u32 %p386, %r3597, %r744; @%p386 bra $L__BB6_614; bra.uni $L__BB6_617; $L__BB6_647: ld.local.u64 %rd506, [%rd2+40]; mov.u64 %rd2389, 0; ld.u32 %r751, [%rd506+24]; setp.eq.s32 %p406, %r751, 0; @%p406 bra $L__BB6_652; add.s64 %rd507, %rd506, 64; shr.u32 %r2549, %r471, 12; cvt.u64.u32 %rd1921, %r2549; shr.u32 %r2550, %r741, 12; cvt.u64.u32 %rd1922, %r2550; bfi.b64 %rd1923, %rd1921, %rd1922, 42, 22; shr.u32 %r2551, %r748, 12; mul.wide.u32 %rd1924, %r2551, 2097152; or.b64 %rd508, %rd1923, %rd1924; mov.u32 %r3599, 0; $L__BB6_649: cvt.u64.u32 %rd509, %r3599; mul.wide.u32 %rd1925, %r3599, 32; add.s64 %rd1926, %rd507, %rd1925; ld.u64 %rd1927, [%rd1926]; setp.eq.s64 %p407, %rd1927, %rd508; add.s32 %r3599, %r3599, 1; @%p407 bra $L__BB6_651; setp.lt.u32 %p408, %r3599, %r751; @%p408 bra $L__BB6_649; bra.uni $L__BB6_652; $L__BB6_682: ld.local.u64 %rd524, [%rd2+40]; mov.u64 %rd2390, 0; ld.u32 %r757, [%rd524+24]; setp.eq.s32 %p428, %r757, 0; @%p428 bra $L__BB6_687; add.s64 %rd525, %rd524, 64; shr.u32 %r2649, %r471, 12; cvt.u64.u32 %rd1988, %r2649; shr.u32 %r2650, %r733, 12; cvt.u64.u32 %rd1989, %r2650; bfi.b64 %rd1990, %rd1988, %rd1989, 42, 22; shr.u32 %r2651, %r748, 12; mul.wide.u32 %rd1991, %r2651, 2097152; or.b64 %rd526, %rd1990, %rd1991; mov.u32 %r3601, 0; $L__BB6_684: cvt.u64.u32 %rd527, %r3601; mul.wide.u32 %rd1992, %r3601, 32; add.s64 %rd1993, %rd525, %rd1992; ld.u64 %rd1994, [%rd1993]; setp.eq.s64 %p429, %rd1994, %rd526; add.s32 %r3601, %r3601, 1; @%p429 bra $L__BB6_686; setp.lt.u32 %p430, %r3601, %r757; @%p430 bra $L__BB6_684; bra.uni $L__BB6_687; $L__BB6_717: ld.local.u64 %rd542, [%rd2+40]; mov.u64 %rd2391, 0; ld.u32 %r763, [%rd542+24]; setp.eq.s32 %p450, %r763, 0; @%p450 bra $L__BB6_722; add.s64 %rd543, %rd542, 64; shr.u32 %r2748, %r635, 12; cvt.u64.u32 %rd2055, %r2748; shr.u32 %r2749, %r733, 12; cvt.u64.u32 %rd2056, %r2749; bfi.b64 %rd2057, %rd2055, %rd2056, 42, 22; shr.u32 %r2750, %r748, 12; mul.wide.u32 %rd2058, %r2750, 2097152; or.b64 %rd544, %rd2057, %rd2058; mov.u32 %r3603, 0; $L__BB6_719: cvt.u64.u32 %rd545, %r3603; mul.wide.u32 %rd2059, %r3603, 32; add.s64 %rd2060, %rd543, %rd2059; ld.u64 %rd2061, [%rd2060]; setp.eq.s64 %p451, %rd2061, %rd544; add.s32 %r3603, %r3603, 1; @%p451 bra $L__BB6_721; setp.lt.u32 %p452, %r3603, %r763; @%p452 bra $L__BB6_719; bra.uni $L__BB6_722; $L__BB6_752: ld.local.u64 %rd560, [%rd2+40]; mov.u64 %rd2392, 0; ld.u32 %r769, [%rd560+24]; setp.eq.s32 %p472, %r769, 0; @%p472 bra $L__BB6_757; add.s64 %rd561, %rd560, 64; shr.u32 %r2850, %r635, 12; cvt.u64.u32 %rd2122, %r2850; shr.u32 %r2851, %r741, 12; cvt.u64.u32 %rd2123, %r2851; bfi.b64 %rd2124, %rd2122, %rd2123, 42, 22; shr.u32 %r2852, %r748, 12; mul.wide.u32 %rd2125, %r2852, 2097152; or.b64 %rd562, %rd2124, %rd2125; mov.u32 %r3605, 0; $L__BB6_754: cvt.u64.u32 %rd563, %r3605; mul.wide.u32 %rd2126, %r3605, 32; add.s64 %rd2127, %rd561, %rd2126; ld.u64 %rd2128, [%rd2127]; setp.eq.s64 %p473, %rd2128, %rd562; add.s32 %r3605, %r3605, 1; @%p473 bra $L__BB6_756; setp.lt.u32 %p474, %r3605, %r769; @%p474 bra $L__BB6_754; bra.uni $L__BB6_757; $L__BB6_787: ld.local.u64 %rd578, [%rd2+40]; mov.u64 %rd2393, 0; ld.u32 %r775, [%rd578+24]; setp.eq.s32 %p494, %r775, 0; @%p494 bra $L__BB6_792; add.s64 %rd579, %rd578, 64; shr.u32 %r2953, %r635, 12; cvt.u64.u32 %rd2189, %r2953; shl.b64 %rd2190, %rd2189, 42; shr.u32 %r2954, %r732, 12; mul.wide.u32 %rd2191, %r2954, 2097152; or.b64 %rd2192, %rd2190, %rd2191; shr.u32 %r2955, %r741, 12; cvt.u64.u32 %rd2193, %r2955; or.b64 %rd580, %rd2192, %rd2193; mov.u32 %r3607, 0; $L__BB6_789: cvt.u64.u32 %rd581, %r3607; mul.wide.u32 %rd2194, %r3607, 32; add.s64 %rd2195, %rd579, %rd2194; ld.u64 %rd2196, [%rd2195]; setp.eq.s64 %p495, %rd2196, %rd580; add.s32 %r3607, %r3607, 1; @%p495 bra $L__BB6_791; setp.lt.u32 %p496, %r3607, %r775; @%p496 bra $L__BB6_789; bra.uni $L__BB6_792; $L__BB6_831: ld.f32 %f952, [%rd607]; bra.uni $L__BB6_833; $L__BB6_828: ld.f32 %f952, [%rd603]; bra.uni $L__BB6_833; $L__BB6_51: ld.f32 %f930, [%rd45]; bra.uni $L__BB6_66; $L__BB6_86: ld.f32 %f931, [%rd74]; bra.uni $L__BB6_101; $L__BB6_121: ld.f32 %f932, [%rd103]; bra.uni $L__BB6_136; $L__BB6_156: ld.f32 %f933, [%rd132]; bra.uni $L__BB6_171; $L__BB6_191: ld.f32 %f934, [%rd161]; bra.uni $L__BB6_206; $L__BB6_226: ld.f32 %f935, [%rd190]; bra.uni $L__BB6_241; $L__BB6_261: ld.f32 %f936, [%rd219]; bra.uni $L__BB6_276; $L__BB6_296: ld.f32 %f937, [%rd248]; bra.uni $L__BB6_311; $L__BB6_331: ld.f32 %f938, [%rd277]; bra.uni $L__BB6_346; $L__BB6_366: ld.f32 %f939, [%rd306]; bra.uni $L__BB6_381; $L__BB6_401: ld.f32 %f940, [%rd335]; bra.uni $L__BB6_416; $L__BB6_436: ld.f32 %f941, [%rd364]; bra.uni $L__BB6_451; $L__BB6_471: ld.f32 %f942, [%rd393]; bra.uni $L__BB6_486; $L__BB6_506: ld.f32 %f943, [%rd422]; bra.uni $L__BB6_521; $L__BB6_541: ld.f32 %f944, [%rd450]; bra.uni $L__BB6_556; $L__BB6_576: ld.f32 %f945, [%rd467]; bra.uni $L__BB6_591; $L__BB6_611: ld.f32 %f946, [%rd485]; bra.uni $L__BB6_626; $L__BB6_646: ld.f32 %f947, [%rd503]; bra.uni $L__BB6_661; $L__BB6_681: ld.f32 %f948, [%rd521]; bra.uni $L__BB6_696; $L__BB6_716: ld.f32 %f949, [%rd539]; bra.uni $L__BB6_731; $L__BB6_751: ld.f32 %f950, [%rd557]; bra.uni $L__BB6_766; $L__BB6_786: ld.f32 %f951, [%rd575]; bra.uni $L__BB6_801; $L__BB6_827: ld.f32 %f952, [%rd605]; $L__BB6_833: ld.param.f32 %f928, [_Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj_param_3]; add.ftz.f32 %f927, %f928, 0fBF000000; cvt.rmi.ftz.f32.f32 %f926, %f927; sub.ftz.f32 %f925, %f927, %f926; mul.ftz.f32 %f924, %f925, 0f40400000; mov.f32 %f923, 0f40400000; sub.ftz.f32 %f922, %f923, %f924; fma.rn.ftz.f32 %f921, %f925, %f922, 0f40400000; mul.ftz.f32 %f920, %f925, %f925; mul.ftz.f32 %f919, %f925, %f920; mul.ftz.f32 %f918, %f919, 0f3E2AAAAB; fma.rn.ftz.f32 %f917, %f925, %f921, 0f3F800000; fma.rn.ftz.f32 %f916, %f917, 0f3E2AAAAB, %f918; sub.ftz.f32 %f907, %f951, %f952; fma.rn.ftz.f32 %f908, %f220, %f907, %f952; sub.ftz.f32 %f909, %f295, %f908; fma.rn.ftz.f32 %f910, %f219, %f909, %f908; sub.ftz.f32 %f911, %f910, %f270; fma.rn.ftz.f32 %f912, %f116, %f911, %f270; fma.rn.ftz.f32 %f913, %f8, %f912, %f218; fma.rn.ftz.f32 %f914, %f10, %f913, %f217; fma.rn.ftz.f32 %f953, %f916, %f914, %f216; $L__BB6_834: st.param.f32 [func_retval0+0], %f953; ret; $L__BB6_50: ld.f32 %f930, [%rd47]; bra.uni $L__BB6_66; $L__BB6_85: ld.f32 %f931, [%rd76]; bra.uni $L__BB6_101; $L__BB6_120: ld.f32 %f932, [%rd105]; bra.uni $L__BB6_136; $L__BB6_155: and.b32 %r3272, %r259, -128; ld.f32 %f933, [%rd134]; bra.uni $L__BB6_171; $L__BB6_190: and.b32 %r3272, %r259, -128; ld.f32 %f934, [%rd163]; bra.uni $L__BB6_206; $L__BB6_225: ld.f32 %f935, [%rd192]; mov.u32 %r3272, %r372; bra.uni $L__BB6_241; $L__BB6_260: ld.f32 %f936, [%rd221]; mov.u32 %r3272, %r423; bra.uni $L__BB6_276; $L__BB6_295: ld.f32 %f937, [%rd250]; mov.u32 %r3272, %r473; bra.uni $L__BB6_311; $L__BB6_330: ld.f32 %f938, [%rd279]; mov.u32 %r3272, %r518; bra.uni $L__BB6_346; $L__BB6_365: ld.f32 %f939, [%rd308]; mov.u32 %r3272, %r560; bra.uni $L__BB6_381; $L__BB6_400: ld.f32 %f940, [%rd337]; mov.u32 %r3272, %r599; bra.uni $L__BB6_416; $L__BB6_435: ld.f32 %f941, [%rd366]; mov.u32 %r3272, %r637; bra.uni $L__BB6_451; $L__BB6_470: ld.f32 %f942, [%rd395]; mov.u32 %r3272, %r670; bra.uni $L__BB6_486; $L__BB6_505: ld.f32 %f943, [%rd424]; mov.u32 %r3272, %r700; bra.uni $L__BB6_521; $L__BB6_540: ld.f32 %f944, [%rd452]; bra.uni $L__BB6_556; $L__BB6_575: ld.f32 %f945, [%rd469]; bra.uni $L__BB6_591; $L__BB6_610: ld.f32 %f946, [%rd487]; bra.uni $L__BB6_626; $L__BB6_645: ld.f32 %f947, [%rd505]; bra.uni $L__BB6_661; $L__BB6_680: ld.f32 %f948, [%rd523]; bra.uni $L__BB6_696; $L__BB6_715: ld.f32 %f949, [%rd541]; bra.uni $L__BB6_731; $L__BB6_750: ld.f32 %f950, [%rd559]; bra.uni $L__BB6_766; $L__BB6_785: ld.f32 %f951, [%rd577]; bra.uni $L__BB6_801; $L__BB6_56: cvt.u64.u32 %rd2291, %r3181; shl.b64 %rd742, %rd2291, 5; add.s64 %rd2302, %rd17, %rd742; $L__BB6_57: setp.eq.s64 %p35, %rd2302, 0; @%p35 bra $L__BB6_65; add.s64 %rd55, %rd2302, 8; ld.u64 %rd56, [%rd2302+8]; setp.eq.s64 %p36, %rd56, 0; @%p36 bra $L__BB6_64; add.s64 %rd2315, %rd1, %rd56; and.b32 %r1052, %r8, -4096; mov.u64 %rd743, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1052}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1053, %r7, 3; and.b32 %r1054, %r1053, 31744; and.b32 %r1055, %r8, 3968; shr.u32 %r1056, %r1055, 2; or.b32 %r1057, %r1056, %r1054; bfe.u32 %r1058, %r74, 7, 5; or.b32 %r1059, %r1058, %r1057; shr.u32 %r1060, %r1057, 6; mul.wide.u32 %rd744, %r1060, 8; add.s64 %rd745, %rd2315, %rd744; and.b32 %r1061, %r1059, 63; shl.b64 %rd746, %rd743, %r1061; ld.u64 %rd747, [%rd745+4128]; and.b64 %rd748, %rd747, %rd746; setp.eq.s64 %p37, %rd748, 0; mul.wide.u32 %rd749, %r1059, 8; add.s64 %rd750, %rd2315, %rd749; add.s64 %rd58, %rd750, 8256; mov.b64 %rd751, {%r12, %r1052}; shr.u64 %rd752, %rd751, 32; cvt.u32.u64 %r3277, %rd752; @%p37 bra $L__BB6_63; and.b32 %r3157, %r7, -128; ld.u64 %rd753, [%rd58]; add.s64 %rd2311, %rd2315, %rd753; st.local.u32 [%rd2+12], %r3157; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1062, %r7, 5; and.b32 %r1063, %r1062, 3840; shl.b32 %r1064, %r8, 1; and.b32 %r1065, %r1064, 240; or.b32 %r1066, %r1065, %r1063; bfe.u32 %r1067, %r74, 3, 4; or.b32 %r1068, %r1067, %r1066; shr.u32 %r1069, %r1066, 6; mul.wide.u32 %rd754, %r1069, 8; add.s64 %rd755, %rd2311, %rd754; and.b32 %r1070, %r1068, 63; shl.b64 %rd757, %rd743, %r1070; ld.u64 %rd758, [%rd755+544]; and.b64 %rd759, %rd758, %rd757; setp.eq.s64 %p38, %rd759, 0; mul.wide.u32 %rd760, %r1068, 8; add.s64 %rd761, %rd2311, %rd760; add.s64 %rd60, %rd761, 1088; and.b32 %r3272, %r7, -128; @%p38 bra $L__BB6_62; and.b32 %r3229, %r7, -8; ld.u64 %rd762, [%rd60]; add.s64 %rd2307, %rd2311, %rd762; and.b32 %r1071, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1071}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1072, %r7, 6; and.b32 %r1073, %r1072, 448; shl.b32 %r1074, %r8, 3; and.b32 %r1075, %r1074, 56; or.b32 %r1076, %r1075, %r1073; and.b32 %r1077, %r74, 7; or.b32 %r1078, %r1076, %r1077; mul.wide.u32 %rd763, %r1078, 2; add.s64 %rd764, %rd2307, %rd763; ld.u16 %rs144, [%rd764+96]; cvt.rn.f32.u16 %f419, %rs144; ld.v2.f32 {%f420, %f421}, [%rd2307+80]; fma.rn.ftz.f32 %f930, %f421, %f419, %f420; mov.b64 %rd765, {%r3229, %r1071}; shr.u64 %rd766, %rd765, 32; cvt.u32.u64 %r3184, %rd766; mov.u32 %r3232, %r12; bra.uni $L__BB6_66; $L__BB6_65: ld.f32 %f930, [%rd5+4]; bra.uni $L__BB6_66; $L__BB6_91: cvt.u64.u32 %rd2294, %r3230; shl.b64 %rd812, %rd2294, 5; add.s64 %rd2308, %rd17, %rd812; $L__BB6_92: setp.eq.s64 %p57, %rd2308, 0; @%p57 bra $L__BB6_100; add.s64 %rd84, %rd2308, 8; ld.u64 %rd85, [%rd2308+8]; setp.eq.s64 %p58, %rd85, 0; @%p58 bra $L__BB6_99; add.s64 %rd2315, %rd1, %rd85; and.b32 %r1138, %r139, -4096; mov.u64 %rd813, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1138}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1139, %r7, 3; and.b32 %r1140, %r1139, 31744; and.b32 %r1141, %r139, 3968; shr.u32 %r1142, %r1141, 2; or.b32 %r1143, %r1142, %r1140; bfe.u32 %r1144, %r74, 7, 5; or.b32 %r1145, %r1143, %r1144; shr.u32 %r1146, %r1143, 6; mul.wide.u32 %rd814, %r1146, 8; add.s64 %rd815, %rd2315, %rd814; and.b32 %r1147, %r1145, 63; shl.b64 %rd816, %rd813, %r1147; ld.u64 %rd817, [%rd815+4128]; and.b64 %rd818, %rd817, %rd816; setp.eq.s64 %p59, %rd818, 0; mul.wide.u32 %rd819, %r1145, 8; add.s64 %rd820, %rd2315, %rd819; add.s64 %rd87, %rd820, 8256; mov.b64 %rd821, {%r12, %r1138}; shr.u64 %rd822, %rd821, 32; cvt.u32.u64 %r3277, %rd822; @%p59 bra $L__BB6_98; and.b32 %r3131, %r7, -128; ld.u64 %rd823, [%rd87]; add.s64 %rd2311, %rd2315, %rd823; st.local.u32 [%rd2+12], %r3131; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1148, %r7, 5; and.b32 %r1149, %r1148, 3840; shl.b32 %r1150, %r139, 1; and.b32 %r1151, %r1150, 240; or.b32 %r1152, %r1151, %r1149; bfe.u32 %r1153, %r74, 3, 4; or.b32 %r1154, %r1152, %r1153; shr.u32 %r1155, %r1152, 6; mul.wide.u32 %rd824, %r1155, 8; add.s64 %rd825, %rd2311, %rd824; and.b32 %r1156, %r1154, 63; shl.b64 %rd827, %rd813, %r1156; ld.u64 %rd828, [%rd825+544]; and.b64 %rd829, %rd828, %rd827; setp.eq.s64 %p60, %rd829, 0; mul.wide.u32 %rd830, %r1154, 8; add.s64 %rd831, %rd2311, %rd830; add.s64 %rd89, %rd831, 1088; and.b32 %r3272, %r7, -128; @%p60 bra $L__BB6_97; and.b32 %r3229, %r7, -8; ld.u64 %rd832, [%rd89]; add.s64 %rd2307, %rd2311, %rd832; and.b32 %r1157, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1157}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1158, %r7, 6; and.b32 %r1159, %r1158, 448; shl.b32 %r1160, %r139, 3; and.b32 %r1161, %r1160, 56; or.b32 %r1162, %r1161, %r1159; and.b32 %r1163, %r74, 7; or.b32 %r1164, %r1162, %r1163; mul.wide.u32 %rd833, %r1164, 2; add.s64 %rd834, %rd2307, %rd833; ld.u16 %rs154, [%rd834+96]; cvt.rn.f32.u16 %f440, %rs154; ld.v2.f32 {%f441, %f442}, [%rd2307+80]; fma.rn.ftz.f32 %f931, %f442, %f440, %f441; mov.b64 %rd835, {%r3229, %r1157}; shr.u64 %rd836, %rd835, 32; cvt.u32.u64 %r3184, %rd836; mov.u32 %r3232, %r12; bra.uni $L__BB6_101; $L__BB6_100: ld.f32 %f931, [%rd5+4]; bra.uni $L__BB6_101; $L__BB6_126: cvt.u64.u32 %rd2295, %r3276; shl.b64 %rd880, %rd2295, 5; add.s64 %rd2314, %rd17, %rd880; $L__BB6_127: setp.eq.s64 %p79, %rd2314, 0; @%p79 bra $L__BB6_135; add.s64 %rd113, %rd2314, 8; ld.u64 %rd114, [%rd2314+8]; setp.eq.s64 %p80, %rd114, 0; @%p80 bra $L__BB6_134; add.s64 %rd2315, %rd1, %rd114; and.b32 %r1220, %r139, -4096; mov.u64 %rd881, 1; st.local.v2.u32 [%rd2+24], {%r12, %r1220}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1221, %r7, 3; and.b32 %r1222, %r1221, 31744; and.b32 %r1223, %r139, 3968; shr.u32 %r1224, %r1223, 2; or.b32 %r1225, %r1224, %r1222; bfe.u32 %r1226, %r9, 7, 5; or.b32 %r1227, %r1225, %r1226; shr.u32 %r1228, %r1225, 6; mul.wide.u32 %rd882, %r1228, 8; add.s64 %rd883, %rd2315, %rd882; and.b32 %r1229, %r1227, 63; shl.b64 %rd884, %rd881, %r1229; ld.u64 %rd885, [%rd883+4128]; and.b64 %rd886, %rd885, %rd884; setp.eq.s64 %p81, %rd886, 0; mul.wide.u32 %rd887, %r1227, 8; add.s64 %rd888, %rd2315, %rd887; add.s64 %rd116, %rd888, 8256; mov.b64 %rd889, {%r12, %r1220}; shr.u64 %rd890, %rd889, 32; cvt.u32.u64 %r3277, %rd890; @%p81 bra $L__BB6_133; and.b32 %r3104, %r7, -128; ld.u64 %rd891, [%rd116]; add.s64 %rd2311, %rd2315, %rd891; st.local.u32 [%rd2+12], %r3104; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1230, %r7, 5; and.b32 %r1231, %r1230, 3840; shl.b32 %r1232, %r139, 1; and.b32 %r1233, %r1232, 240; or.b32 %r1234, %r1233, %r1231; bfe.u32 %r1235, %r9, 3, 4; or.b32 %r1236, %r1234, %r1235; shr.u32 %r1237, %r1234, 6; mul.wide.u32 %rd892, %r1237, 8; add.s64 %rd893, %rd2311, %rd892; and.b32 %r1238, %r1236, 63; shl.b64 %rd895, %rd881, %r1238; ld.u64 %rd896, [%rd893+544]; and.b64 %rd897, %rd896, %rd895; setp.eq.s64 %p82, %rd897, 0; mul.wide.u32 %rd898, %r1236, 8; add.s64 %rd899, %rd2311, %rd898; add.s64 %rd118, %rd899, 1088; and.b32 %r3272, %r7, -128; @%p82 bra $L__BB6_132; and.b32 %r3229, %r7, -8; ld.u64 %rd900, [%rd118]; add.s64 %rd2307, %rd2311, %rd900; and.b32 %r1239, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1239}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1240, %r7, 7; shl.b32 %r1241, %r139, 3; and.b32 %r1242, %r1241, 56; and.b32 %r1243, %r9, 7; bfi.b32 %r1244, %r1240, %r1243, 6, 3; or.b32 %r1245, %r1244, %r1242; mul.wide.u32 %rd901, %r1245, 2; add.s64 %rd902, %rd2307, %rd901; ld.u16 %rs164, [%rd902+96]; cvt.rn.f32.u16 %f460, %rs164; ld.v2.f32 {%f461, %f462}, [%rd2307+80]; fma.rn.ftz.f32 %f932, %f462, %f460, %f461; mov.b64 %rd903, {%r3229, %r1239}; shr.u64 %rd904, %rd903, 32; cvt.u32.u64 %r3184, %rd904; mov.u32 %r3232, %r12; bra.uni $L__BB6_136; $L__BB6_135: ld.f32 %f932, [%rd5+4]; bra.uni $L__BB6_136; $L__BB6_161: cvt.u64.u32 %rd2296, %r3319; shl.b64 %rd951, %rd2296, 5; add.s64 %rd2320, %rd17, %rd951; $L__BB6_162: setp.eq.s64 %p101, %rd2320, 0; @%p101 bra $L__BB6_170; add.s64 %rd142, %rd2320, 8; ld.u64 %rd143, [%rd2320+8]; setp.eq.s64 %p102, %rd143, 0; @%p102 bra $L__BB6_169; and.b32 %r3109, %r259, -4096; add.s64 %rd2315, %rd1, %rd143; and.b32 %r1302, %r139, -4096; mov.u64 %rd952, 1; st.local.v2.u32 [%rd2+24], {%r3109, %r1302}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1303, %r259, 3; and.b32 %r1304, %r1303, 31744; and.b32 %r1305, %r139, 3968; shr.u32 %r1306, %r1305, 2; or.b32 %r1307, %r1306, %r1304; bfe.u32 %r1308, %r9, 7, 5; or.b32 %r1309, %r1307, %r1308; shr.u32 %r1310, %r1307, 6; mul.wide.u32 %rd953, %r1310, 8; add.s64 %rd954, %rd2315, %rd953; and.b32 %r1311, %r1309, 63; shl.b64 %rd955, %rd952, %r1311; ld.u64 %rd956, [%rd954+4128]; and.b64 %rd957, %rd956, %rd955; setp.eq.s64 %p103, %rd957, 0; mul.wide.u32 %rd958, %r1309, 8; add.s64 %rd959, %rd2315, %rd958; add.s64 %rd145, %rd959, 8256; mov.b64 %rd960, {%r3109, %r1302}; shr.u64 %rd961, %rd960, 32; cvt.u32.u64 %r3277, %rd961; @%p103 bra $L__BB6_168; and.b32 %r3110, %r259, -128; ld.u64 %rd962, [%rd145]; add.s64 %rd2311, %rd2315, %rd962; st.local.u32 [%rd2+12], %r3110; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1312, %r259, 5; and.b32 %r1313, %r1312, 3840; shl.b32 %r1314, %r139, 1; and.b32 %r1315, %r1314, 240; or.b32 %r1316, %r1315, %r1313; bfe.u32 %r1317, %r9, 3, 4; or.b32 %r1318, %r1316, %r1317; shr.u32 %r1319, %r1316, 6; mul.wide.u32 %rd963, %r1319, 8; add.s64 %rd964, %rd2311, %rd963; and.b32 %r1320, %r1318, 63; shl.b64 %rd966, %rd952, %r1320; ld.u64 %rd967, [%rd964+544]; and.b64 %rd968, %rd967, %rd966; setp.eq.s64 %p104, %rd968, 0; mul.wide.u32 %rd969, %r1318, 8; add.s64 %rd970, %rd2311, %rd969; add.s64 %rd147, %rd970, 1088; @%p104 bra $L__BB6_167; and.b32 %r3229, %r259, -8; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.u64 %rd971, [%rd147]; add.s64 %rd2307, %rd2311, %rd971; and.b32 %r1321, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1321}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1322, %r259, 7; shl.b32 %r1323, %r139, 3; and.b32 %r1324, %r1323, 56; and.b32 %r1325, %r9, 7; bfi.b32 %r1326, %r1322, %r1325, 6, 3; or.b32 %r1327, %r1326, %r1324; mul.wide.u32 %rd972, %r1327, 2; add.s64 %rd973, %rd2307, %rd972; ld.u16 %rs174, [%rd973+96]; cvt.rn.f32.u16 %f483, %rs174; ld.v2.f32 {%f484, %f485}, [%rd2307+80]; fma.rn.ftz.f32 %f933, %f485, %f483, %f484; mov.b64 %rd974, {%r3229, %r1321}; shr.u64 %rd975, %rd974, 32; cvt.u32.u64 %r3184, %rd975; bra.uni $L__BB6_171; $L__BB6_170: ld.f32 %f933, [%rd5+4]; bra.uni $L__BB6_171; $L__BB6_196: shl.b64 %rd1022, %rd168, 5; add.s64 %rd2326, %rd17, %rd1022; $L__BB6_197: setp.eq.s64 %p123, %rd2326, 0; @%p123 bra $L__BB6_205; add.s64 %rd171, %rd2326, 8; ld.u64 %rd172, [%rd2326+8]; setp.eq.s64 %p124, %rd172, 0; @%p124 bra $L__BB6_204; and.b32 %r3116, %r259, -4096; add.s64 %rd2315, %rd1, %rd172; and.b32 %r1388, %r139, -4096; mov.u64 %rd1023, 1; st.local.v2.u32 [%rd2+24], {%r3116, %r1388}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1389, %r259, 3; and.b32 %r1390, %r1389, 31744; and.b32 %r1391, %r139, 3968; shr.u32 %r1392, %r1391, 2; or.b32 %r1393, %r1392, %r1390; bfe.u32 %r1394, %r74, 7, 5; or.b32 %r1395, %r1393, %r1394; shr.u32 %r1396, %r1393, 6; mul.wide.u32 %rd1024, %r1396, 8; add.s64 %rd1025, %rd2315, %rd1024; and.b32 %r1397, %r1395, 63; shl.b64 %rd1026, %rd1023, %r1397; ld.u64 %rd1027, [%rd1025+4128]; and.b64 %rd1028, %rd1027, %rd1026; setp.eq.s64 %p125, %rd1028, 0; mul.wide.u32 %rd1029, %r1395, 8; add.s64 %rd1030, %rd2315, %rd1029; add.s64 %rd174, %rd1030, 8256; mov.b64 %rd1031, {%r3116, %r1388}; shr.u64 %rd1032, %rd1031, 32; cvt.u32.u64 %r3277, %rd1032; @%p125 bra $L__BB6_203; and.b32 %r3117, %r259, -128; ld.u64 %rd1033, [%rd174]; add.s64 %rd2311, %rd2315, %rd1033; st.local.u32 [%rd2+12], %r3117; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1398, %r259, 5; and.b32 %r1399, %r1398, 3840; shl.b32 %r1400, %r139, 1; and.b32 %r1401, %r1400, 240; or.b32 %r1402, %r1401, %r1399; bfe.u32 %r1403, %r74, 3, 4; or.b32 %r1404, %r1402, %r1403; shr.u32 %r1405, %r1402, 6; mul.wide.u32 %rd1034, %r1405, 8; add.s64 %rd1035, %rd2311, %rd1034; and.b32 %r1406, %r1404, 63; shl.b64 %rd1037, %rd1023, %r1406; ld.u64 %rd1038, [%rd1035+544]; and.b64 %rd1039, %rd1038, %rd1037; setp.eq.s64 %p126, %rd1039, 0; mul.wide.u32 %rd1040, %r1404, 8; add.s64 %rd1041, %rd2311, %rd1040; add.s64 %rd176, %rd1041, 1088; @%p126 bra $L__BB6_202; and.b32 %r3229, %r259, -8; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.u64 %rd1042, [%rd176]; add.s64 %rd2307, %rd2311, %rd1042; and.b32 %r1407, %r139, -8; st.local.v2.u32 [%rd2], {%r3229, %r1407}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1408, %r259, 6; and.b32 %r1409, %r1408, 448; shl.b32 %r1410, %r139, 3; and.b32 %r1411, %r1410, 56; or.b32 %r1412, %r1411, %r1409; and.b32 %r1413, %r74, 7; or.b32 %r1414, %r1412, %r1413; mul.wide.u32 %rd1043, %r1414, 2; add.s64 %rd1044, %rd2307, %rd1043; ld.u16 %rs184, [%rd1044+96]; cvt.rn.f32.u16 %f503, %rs184; ld.v2.f32 {%f504, %f505}, [%rd2307+80]; fma.rn.ftz.f32 %f934, %f505, %f503, %f504; mov.b64 %rd1045, {%r3229, %r1407}; shr.u64 %rd1046, %rd1045, 32; cvt.u32.u64 %r3184, %rd1046; bra.uni $L__BB6_206; $L__BB6_205: ld.f32 %f934, [%rd5+4]; bra.uni $L__BB6_206; $L__BB6_231: shl.b64 %rd1093, %rd197, 5; add.s64 %rd2332, %rd17, %rd1093; $L__BB6_232: setp.eq.s64 %p145, %rd2332, 0; @%p145 bra $L__BB6_240; add.s64 %rd200, %rd2332, 8; ld.u64 %rd201, [%rd2332+8]; setp.eq.s64 %p146, %rd201, 0; @%p146 bra $L__BB6_239; and.b32 %r3140, %r259, -4096; add.s64 %rd2315, %rd1, %rd201; and.b32 %r1474, %r8, -4096; mov.u64 %rd1094, 1; st.local.v2.u32 [%rd2+24], {%r3140, %r1474}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1475, %r259, 3; and.b32 %r1476, %r1475, 31744; and.b32 %r1477, %r8, 3968; shr.u32 %r1478, %r1477, 2; or.b32 %r1479, %r1476, %r1478; bfe.u32 %r1480, %r74, 7, 5; or.b32 %r1481, %r1480, %r1479; shr.u32 %r1482, %r1479, 6; mul.wide.u32 %rd1095, %r1482, 8; add.s64 %rd1096, %rd2315, %rd1095; and.b32 %r1483, %r1481, 63; shl.b64 %rd1097, %rd1094, %r1483; ld.u64 %rd1098, [%rd1096+4128]; and.b64 %rd1099, %rd1098, %rd1097; setp.eq.s64 %p147, %rd1099, 0; mul.wide.u32 %rd1100, %r1481, 8; add.s64 %rd1101, %rd2315, %rd1100; add.s64 %rd203, %rd1101, 8256; mov.b64 %rd1102, {%r3140, %r1474}; shr.u64 %rd1103, %rd1102, 32; cvt.u32.u64 %r3277, %rd1103; @%p147 bra $L__BB6_238; and.b32 %r3141, %r259, -128; ld.u64 %rd1104, [%rd203]; add.s64 %rd2311, %rd2315, %rd1104; st.local.u32 [%rd2+12], %r3141; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1484, %r259, 5; and.b32 %r1485, %r1484, 3840; shl.b32 %r1486, %r8, 1; and.b32 %r1487, %r1486, 240; or.b32 %r1488, %r1485, %r1487; bfe.u32 %r1489, %r74, 3, 4; or.b32 %r1490, %r1489, %r1488; shr.u32 %r1491, %r1488, 6; mul.wide.u32 %rd1105, %r1491, 8; add.s64 %rd1106, %rd2311, %rd1105; and.b32 %r1492, %r1490, 63; shl.b64 %rd1108, %rd1094, %r1492; ld.u64 %rd1109, [%rd1106+544]; and.b64 %rd1110, %rd1109, %rd1108; setp.eq.s64 %p148, %rd1110, 0; mul.wide.u32 %rd1111, %r1490, 8; add.s64 %rd1112, %rd2311, %rd1111; add.s64 %rd205, %rd1112, 1088; and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; @%p148 bra $L__BB6_237; and.b32 %r3229, %r259, -8; ld.u64 %rd1113, [%rd205]; add.s64 %rd2307, %rd2311, %rd1113; and.b32 %r1493, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1493}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1494, %r259, 6; and.b32 %r1495, %r1494, 448; shl.b32 %r1496, %r8, 3; and.b32 %r1497, %r1496, 56; or.b32 %r1498, %r1495, %r1497; and.b32 %r1499, %r74, 7; or.b32 %r1500, %r1498, %r1499; mul.wide.u32 %rd1114, %r1500, 2; add.s64 %rd1115, %rd2307, %rd1114; ld.u16 %rs194, [%rd1115+96]; cvt.rn.f32.u16 %f524, %rs194; ld.v2.f32 {%f525, %f526}, [%rd2307+80]; fma.rn.ftz.f32 %f935, %f526, %f524, %f525; mov.b64 %rd1116, {%r3229, %r1493}; shr.u64 %rd1117, %rd1116, 32; cvt.u32.u64 %r3184, %rd1117; bra.uni $L__BB6_241; $L__BB6_240: ld.f32 %f935, [%rd5+4]; bra.uni $L__BB6_241; $L__BB6_266: shl.b64 %rd1163, %rd226, 5; add.s64 %rd2338, %rd17, %rd1163; $L__BB6_267: setp.eq.s64 %p167, %rd2338, 0; @%p167 bra $L__BB6_275; add.s64 %rd229, %rd2338, 8; ld.u64 %rd230, [%rd2338+8]; setp.eq.s64 %p168, %rd230, 0; @%p168 bra $L__BB6_274; add.s64 %rd2315, %rd1, %rd230; and.b32 %r1559, %r8, -4096; mov.u64 %rd1164, 1; st.local.v2.u32 [%rd2+24], {%r426, %r1559}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1560, %r259, 3; and.b32 %r1561, %r1560, 31744; and.b32 %r1562, %r8, 3968; shr.u32 %r1563, %r1562, 2; or.b32 %r1564, %r1561, %r1563; bfe.u32 %r1565, %r9, 7, 5; or.b32 %r1566, %r1564, %r1565; shr.u32 %r1567, %r1564, 6; mul.wide.u32 %rd1165, %r1567, 8; add.s64 %rd1166, %rd2315, %rd1165; and.b32 %r1568, %r1566, 63; shl.b64 %rd1167, %rd1164, %r1568; ld.u64 %rd1168, [%rd1166+4128]; and.b64 %rd1169, %rd1168, %rd1167; setp.eq.s64 %p169, %rd1169, 0; mul.wide.u32 %rd1170, %r1566, 8; add.s64 %rd1171, %rd2315, %rd1170; add.s64 %rd232, %rd1171, 8256; mov.b64 %rd1172, {%r426, %r1559}; shr.u64 %rd1173, %rd1172, 32; cvt.u32.u64 %r3277, %rd1173; @%p169 bra $L__BB6_273; ld.u64 %rd1174, [%rd232]; add.s64 %rd2311, %rd2315, %rd1174; st.local.u32 [%rd2+12], %r423; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1569, %r259, 5; and.b32 %r1570, %r1569, 3840; shl.b32 %r1571, %r8, 1; and.b32 %r1572, %r1571, 240; or.b32 %r1573, %r1570, %r1572; bfe.u32 %r1574, %r9, 3, 4; or.b32 %r1575, %r1573, %r1574; shr.u32 %r1576, %r1573, 6; mul.wide.u32 %rd1175, %r1576, 8; add.s64 %rd1176, %rd2311, %rd1175; and.b32 %r1577, %r1575, 63; shl.b64 %rd1178, %rd1164, %r1577; ld.u64 %rd1179, [%rd1176+544]; and.b64 %rd1180, %rd1179, %rd1178; setp.eq.s64 %p170, %rd1180, 0; mul.wide.u32 %rd1181, %r1575, 8; add.s64 %rd1182, %rd2311, %rd1181; add.s64 %rd234, %rd1182, 1088; @%p170 bra $L__BB6_272; and.b32 %r3229, %r259, -8; ld.u64 %rd1183, [%rd234]; add.s64 %rd2307, %rd2311, %rd1183; and.b32 %r1578, %r8, -8; st.local.v2.u32 [%rd2], {%r3229, %r1578}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1579, %r259, 6; and.b32 %r1580, %r1579, 448; shl.b32 %r1581, %r8, 3; and.b32 %r1582, %r1581, 56; or.b32 %r1583, %r1580, %r1582; and.b32 %r1584, %r9, 7; or.b32 %r1585, %r1583, %r1584; mul.wide.u32 %rd1184, %r1585, 2; add.s64 %rd1185, %rd2307, %rd1184; ld.u16 %rs204, [%rd1185+96]; cvt.rn.f32.u16 %f544, %rs204; ld.v2.f32 {%f545, %f546}, [%rd2307+80]; fma.rn.ftz.f32 %f936, %f546, %f544, %f545; mov.b64 %rd1186, {%r3229, %r1578}; shr.u64 %rd1187, %rd1186, 32; cvt.u32.u64 %r3184, %rd1187; mov.u32 %r3232, %r426; mov.u32 %r3272, %r423; bra.uni $L__BB6_276; $L__BB6_275: ld.f32 %f936, [%rd5+4]; bra.uni $L__BB6_276; $L__BB6_301: shl.b64 %rd1233, %rd255, 5; add.s64 %rd2344, %rd17, %rd1233; $L__BB6_302: setp.eq.s64 %p189, %rd2344, 0; @%p189 bra $L__BB6_310; add.s64 %rd258, %rd2344, 8; ld.u64 %rd259, [%rd2344+8]; setp.eq.s64 %p190, %rd259, 0; @%p190 bra $L__BB6_309; add.s64 %rd2315, %rd1, %rd259; and.b32 %r1644, %r8, -4096; mov.u64 %rd1234, 1; st.local.v2.u32 [%rd2+24], {%r476, %r1644}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1645, %r471, 3; and.b32 %r1646, %r1645, 31744; and.b32 %r1647, %r8, 3968; shr.u32 %r1648, %r1647, 2; or.b32 %r1649, %r1648, %r1646; bfe.u32 %r1650, %r9, 7, 5; or.b32 %r1651, %r1649, %r1650; shr.u32 %r1652, %r1649, 6; mul.wide.u32 %rd1235, %r1652, 8; add.s64 %rd1236, %rd2315, %rd1235; and.b32 %r1653, %r1651, 63; shl.b64 %rd1237, %rd1234, %r1653; ld.u64 %rd1238, [%rd1236+4128]; and.b64 %rd1239, %rd1238, %rd1237; setp.eq.s64 %p191, %rd1239, 0; mul.wide.u32 %rd1240, %r1651, 8; add.s64 %rd1241, %rd2315, %rd1240; add.s64 %rd261, %rd1241, 8256; mov.b64 %rd1242, {%r476, %r1644}; shr.u64 %rd1243, %rd1242, 32; cvt.u32.u64 %r3277, %rd1243; @%p191 bra $L__BB6_308; ld.u64 %rd1244, [%rd261]; add.s64 %rd2311, %rd2315, %rd1244; st.local.u32 [%rd2+12], %r473; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1654, %r471, 5; and.b32 %r1655, %r1654, 3840; shl.b32 %r1656, %r8, 1; and.b32 %r1657, %r1656, 240; or.b32 %r1658, %r1657, %r1655; bfe.u32 %r1659, %r9, 3, 4; or.b32 %r1660, %r1658, %r1659; shr.u32 %r1661, %r1658, 6; mul.wide.u32 %rd1245, %r1661, 8; add.s64 %rd1246, %rd2311, %rd1245; and.b32 %r1662, %r1660, 63; shl.b64 %rd1248, %rd1234, %r1662; ld.u64 %rd1249, [%rd1246+544]; and.b64 %rd1250, %rd1249, %rd1248; setp.eq.s64 %p192, %rd1250, 0; mul.wide.u32 %rd1251, %r1660, 8; add.s64 %rd1252, %rd2311, %rd1251; add.s64 %rd263, %rd1252, 1088; @%p192 bra $L__BB6_307; ld.u64 %rd1253, [%rd263]; add.s64 %rd2307, %rd2311, %rd1253; and.b32 %r1663, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1663}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1664, %r471, 6; and.b32 %r1665, %r1664, 448; shl.b32 %r1666, %r8, 3; and.b32 %r1667, %r1666, 56; or.b32 %r1668, %r1667, %r1665; and.b32 %r1669, %r9, 7; or.b32 %r1670, %r1668, %r1669; mul.wide.u32 %rd1254, %r1670, 2; add.s64 %rd1255, %rd2307, %rd1254; ld.u16 %rs214, [%rd1255+96]; cvt.rn.f32.u16 %f571, %rs214; ld.v2.f32 {%f572, %f573}, [%rd2307+80]; fma.rn.ftz.f32 %f937, %f573, %f571, %f572; mov.b64 %rd1256, {%r472, %r1663}; shr.u64 %rd1257, %rd1256, 32; cvt.u32.u64 %r3184, %rd1257; mov.u32 %r3232, %r476; mov.u32 %r3272, %r473; mov.u32 %r3229, %r472; bra.uni $L__BB6_311; $L__BB6_310: ld.f32 %f937, [%rd5+4]; bra.uni $L__BB6_311; $L__BB6_336: shl.b64 %rd1304, %rd284, 5; add.s64 %rd2350, %rd17, %rd1304; $L__BB6_337: setp.eq.s64 %p211, %rd2350, 0; @%p211 bra $L__BB6_345; add.s64 %rd287, %rd2350, 8; ld.u64 %rd288, [%rd2350+8]; setp.eq.s64 %p212, %rd288, 0; @%p212 bra $L__BB6_344; add.s64 %rd2315, %rd1, %rd288; and.b32 %r1730, %r8, -4096; mov.u64 %rd1305, 1; st.local.v2.u32 [%rd2+24], {%r521, %r1730}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1731, %r471, 3; and.b32 %r1732, %r1731, 31744; and.b32 %r1733, %r8, 3968; shr.u32 %r1734, %r1733, 2; or.b32 %r1735, %r1734, %r1732; bfe.u32 %r1736, %r74, 7, 5; or.b32 %r1737, %r1736, %r1735; shr.u32 %r1738, %r1735, 6; mul.wide.u32 %rd1306, %r1738, 8; add.s64 %rd1307, %rd2315, %rd1306; and.b32 %r1739, %r1737, 63; shl.b64 %rd1308, %rd1305, %r1739; ld.u64 %rd1309, [%rd1307+4128]; and.b64 %rd1310, %rd1309, %rd1308; setp.eq.s64 %p213, %rd1310, 0; mul.wide.u32 %rd1311, %r1737, 8; add.s64 %rd1312, %rd2315, %rd1311; add.s64 %rd290, %rd1312, 8256; mov.b64 %rd1313, {%r521, %r1730}; shr.u64 %rd1314, %rd1313, 32; cvt.u32.u64 %r3277, %rd1314; @%p213 bra $L__BB6_343; ld.u64 %rd1315, [%rd290]; add.s64 %rd2311, %rd2315, %rd1315; st.local.u32 [%rd2+12], %r518; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1740, %r471, 5; and.b32 %r1741, %r1740, 3840; shl.b32 %r1742, %r8, 1; and.b32 %r1743, %r1742, 240; or.b32 %r1744, %r1743, %r1741; bfe.u32 %r1745, %r74, 3, 4; or.b32 %r1746, %r1745, %r1744; shr.u32 %r1747, %r1744, 6; mul.wide.u32 %rd1316, %r1747, 8; add.s64 %rd1317, %rd2311, %rd1316; and.b32 %r1748, %r1746, 63; shl.b64 %rd1319, %rd1305, %r1748; ld.u64 %rd1320, [%rd1317+544]; and.b64 %rd1321, %rd1320, %rd1319; setp.eq.s64 %p214, %rd1321, 0; mul.wide.u32 %rd1322, %r1746, 8; add.s64 %rd1323, %rd2311, %rd1322; add.s64 %rd292, %rd1323, 1088; @%p214 bra $L__BB6_342; ld.u64 %rd1324, [%rd292]; add.s64 %rd2307, %rd2311, %rd1324; and.b32 %r1749, %r8, -8; st.local.v2.u32 [%rd2], {%r472, %r1749}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1750, %r471, 6; and.b32 %r1751, %r1750, 448; shl.b32 %r1752, %r8, 3; and.b32 %r1753, %r1752, 56; or.b32 %r1754, %r1753, %r1751; and.b32 %r1755, %r74, 7; or.b32 %r1756, %r1754, %r1755; mul.wide.u32 %rd1325, %r1756, 2; add.s64 %rd1326, %rd2307, %rd1325; ld.u16 %rs224, [%rd1326+96]; cvt.rn.f32.u16 %f591, %rs224; ld.v2.f32 {%f592, %f593}, [%rd2307+80]; fma.rn.ftz.f32 %f938, %f593, %f591, %f592; mov.b64 %rd1327, {%r472, %r1749}; shr.u64 %rd1328, %rd1327, 32; cvt.u32.u64 %r3184, %rd1328; mov.u32 %r3232, %r521; mov.u32 %r3272, %r518; mov.u32 %r3229, %r472; bra.uni $L__BB6_346; $L__BB6_345: ld.f32 %f938, [%rd5+4]; bra.uni $L__BB6_346; $L__BB6_371: shl.b64 %rd1375, %rd313, 5; add.s64 %rd2356, %rd17, %rd1375; $L__BB6_372: setp.eq.s64 %p233, %rd2356, 0; @%p233 bra $L__BB6_380; add.s64 %rd316, %rd2356, 8; ld.u64 %rd317, [%rd2356+8]; setp.eq.s64 %p234, %rd317, 0; @%p234 bra $L__BB6_379; add.s64 %rd2315, %rd1, %rd317; and.b32 %r1817, %r139, -4096; mov.u64 %rd1376, 1; st.local.v2.u32 [%rd2+24], {%r563, %r1817}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1818, %r471, 3; and.b32 %r1819, %r1818, 31744; and.b32 %r1820, %r139, 3968; shr.u32 %r1821, %r1820, 2; or.b32 %r1822, %r1821, %r1819; bfe.u32 %r1823, %r74, 7, 5; or.b32 %r1824, %r1822, %r1823; shr.u32 %r1825, %r1822, 6; mul.wide.u32 %rd1377, %r1825, 8; add.s64 %rd1378, %rd2315, %rd1377; and.b32 %r1826, %r1824, 63; shl.b64 %rd1379, %rd1376, %r1826; ld.u64 %rd1380, [%rd1378+4128]; and.b64 %rd1381, %rd1380, %rd1379; setp.eq.s64 %p235, %rd1381, 0; mul.wide.u32 %rd1382, %r1824, 8; add.s64 %rd1383, %rd2315, %rd1382; add.s64 %rd319, %rd1383, 8256; mov.b64 %rd1384, {%r563, %r1817}; shr.u64 %rd1385, %rd1384, 32; cvt.u32.u64 %r3277, %rd1385; @%p235 bra $L__BB6_378; ld.u64 %rd1386, [%rd319]; add.s64 %rd2311, %rd2315, %rd1386; st.local.u32 [%rd2+12], %r560; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1827, %r471, 5; and.b32 %r1828, %r1827, 3840; shl.b32 %r1829, %r139, 1; and.b32 %r1830, %r1829, 240; or.b32 %r1831, %r1830, %r1828; bfe.u32 %r1832, %r74, 3, 4; or.b32 %r1833, %r1831, %r1832; shr.u32 %r1834, %r1831, 6; mul.wide.u32 %rd1387, %r1834, 8; add.s64 %rd1388, %rd2311, %rd1387; and.b32 %r1835, %r1833, 63; shl.b64 %rd1390, %rd1376, %r1835; ld.u64 %rd1391, [%rd1388+544]; and.b64 %rd1392, %rd1391, %rd1390; setp.eq.s64 %p236, %rd1392, 0; mul.wide.u32 %rd1393, %r1833, 8; add.s64 %rd1394, %rd2311, %rd1393; add.s64 %rd321, %rd1394, 1088; @%p236 bra $L__BB6_377; ld.u64 %rd1395, [%rd321]; add.s64 %rd2307, %rd2311, %rd1395; and.b32 %r1836, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1836}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r1837, %r471, 6; and.b32 %r1838, %r1837, 448; shl.b32 %r1839, %r139, 3; and.b32 %r1840, %r1839, 56; or.b32 %r1841, %r1840, %r1838; and.b32 %r1842, %r74, 7; or.b32 %r1843, %r1841, %r1842; mul.wide.u32 %rd1396, %r1843, 2; add.s64 %rd1397, %rd2307, %rd1396; ld.u16 %rs234, [%rd1397+96]; cvt.rn.f32.u16 %f612, %rs234; ld.v2.f32 {%f613, %f614}, [%rd2307+80]; fma.rn.ftz.f32 %f939, %f614, %f612, %f613; mov.b64 %rd1398, {%r472, %r1836}; shr.u64 %rd1399, %rd1398, 32; cvt.u32.u64 %r3184, %rd1399; mov.u32 %r3232, %r563; mov.u32 %r3272, %r560; mov.u32 %r3229, %r472; bra.uni $L__BB6_381; $L__BB6_380: ld.f32 %f939, [%rd5+4]; bra.uni $L__BB6_381; $L__BB6_406: shl.b64 %rd1446, %rd342, 5; add.s64 %rd2362, %rd17, %rd1446; $L__BB6_407: setp.eq.s64 %p255, %rd2362, 0; @%p255 bra $L__BB6_415; add.s64 %rd345, %rd2362, 8; ld.u64 %rd346, [%rd2362+8]; setp.eq.s64 %p256, %rd346, 0; @%p256 bra $L__BB6_414; add.s64 %rd2315, %rd1, %rd346; and.b32 %r1900, %r139, -4096; mov.u64 %rd1447, 1; st.local.v2.u32 [%rd2+24], {%r602, %r1900}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1901, %r471, 3; and.b32 %r1902, %r1901, 31744; and.b32 %r1903, %r139, 3968; shr.u32 %r1904, %r1903, 2; or.b32 %r1905, %r1904, %r1902; bfe.u32 %r1906, %r9, 7, 5; or.b32 %r1907, %r1905, %r1906; shr.u32 %r1908, %r1905, 6; mul.wide.u32 %rd1448, %r1908, 8; add.s64 %rd1449, %rd2315, %rd1448; and.b32 %r1909, %r1907, 63; shl.b64 %rd1450, %rd1447, %r1909; ld.u64 %rd1451, [%rd1449+4128]; and.b64 %rd1452, %rd1451, %rd1450; setp.eq.s64 %p257, %rd1452, 0; mul.wide.u32 %rd1453, %r1907, 8; add.s64 %rd1454, %rd2315, %rd1453; add.s64 %rd348, %rd1454, 8256; mov.b64 %rd1455, {%r602, %r1900}; shr.u64 %rd1456, %rd1455, 32; cvt.u32.u64 %r3277, %rd1456; @%p257 bra $L__BB6_413; ld.u64 %rd1457, [%rd348]; add.s64 %rd2311, %rd2315, %rd1457; st.local.u32 [%rd2+12], %r599; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1910, %r471, 5; and.b32 %r1911, %r1910, 3840; shl.b32 %r1912, %r139, 1; and.b32 %r1913, %r1912, 240; or.b32 %r1914, %r1913, %r1911; bfe.u32 %r1915, %r9, 3, 4; or.b32 %r1916, %r1914, %r1915; shr.u32 %r1917, %r1914, 6; mul.wide.u32 %rd1458, %r1917, 8; add.s64 %rd1459, %rd2311, %rd1458; and.b32 %r1918, %r1916, 63; shl.b64 %rd1461, %rd1447, %r1918; ld.u64 %rd1462, [%rd1459+544]; and.b64 %rd1463, %rd1462, %rd1461; setp.eq.s64 %p258, %rd1463, 0; mul.wide.u32 %rd1464, %r1916, 8; add.s64 %rd1465, %rd2311, %rd1464; add.s64 %rd350, %rd1465, 1088; @%p258 bra $L__BB6_412; ld.u64 %rd1466, [%rd350]; add.s64 %rd2307, %rd2311, %rd1466; and.b32 %r1919, %r139, -8; st.local.v2.u32 [%rd2], {%r472, %r1919}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r1920, %r471, 7; shl.b32 %r1921, %r139, 3; and.b32 %r1922, %r1921, 56; and.b32 %r1923, %r9, 7; bfi.b32 %r1924, %r1920, %r1923, 6, 3; or.b32 %r1925, %r1924, %r1922; mul.wide.u32 %rd1467, %r1925, 2; add.s64 %rd1468, %rd2307, %rd1467; ld.u16 %rs244, [%rd1468+96]; cvt.rn.f32.u16 %f632, %rs244; ld.v2.f32 {%f633, %f634}, [%rd2307+80]; fma.rn.ftz.f32 %f940, %f634, %f632, %f633; mov.b64 %rd1469, {%r472, %r1919}; shr.u64 %rd1470, %rd1469, 32; cvt.u32.u64 %r3184, %rd1470; mov.u32 %r3232, %r602; mov.u32 %r3272, %r599; mov.u32 %r3229, %r472; bra.uni $L__BB6_416; $L__BB6_415: ld.f32 %f940, [%rd5+4]; bra.uni $L__BB6_416; $L__BB6_441: shl.b64 %rd1517, %rd371, 5; add.s64 %rd2368, %rd17, %rd1517; $L__BB6_442: setp.eq.s64 %p277, %rd2368, 0; @%p277 bra $L__BB6_450; add.s64 %rd374, %rd2368, 8; ld.u64 %rd375, [%rd2368+8]; setp.eq.s64 %p278, %rd375, 0; @%p278 bra $L__BB6_449; add.s64 %rd2315, %rd1, %rd375; and.b32 %r1982, %r139, -4096; mov.u64 %rd1518, 1; st.local.v2.u32 [%rd2+24], {%r640, %r1982}; and.b32 %r3310, %r9, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r1983, %r635, 3; and.b32 %r1984, %r1983, 31744; and.b32 %r1985, %r139, 3968; shr.u32 %r1986, %r1985, 2; or.b32 %r1987, %r1986, %r1984; bfe.u32 %r1988, %r9, 7, 5; or.b32 %r1989, %r1987, %r1988; shr.u32 %r1990, %r1987, 6; mul.wide.u32 %rd1519, %r1990, 8; add.s64 %rd1520, %rd2315, %rd1519; and.b32 %r1991, %r1989, 63; shl.b64 %rd1521, %rd1518, %r1991; ld.u64 %rd1522, [%rd1520+4128]; and.b64 %rd1523, %rd1522, %rd1521; setp.eq.s64 %p279, %rd1523, 0; mul.wide.u32 %rd1524, %r1989, 8; add.s64 %rd1525, %rd2315, %rd1524; add.s64 %rd377, %rd1525, 8256; mov.b64 %rd1526, {%r640, %r1982}; shr.u64 %rd1527, %rd1526, 32; cvt.u32.u64 %r3277, %rd1527; @%p279 bra $L__BB6_448; ld.u64 %rd1528, [%rd377]; add.s64 %rd2311, %rd2315, %rd1528; st.local.u32 [%rd2+12], %r637; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r9, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r1992, %r635, 5; and.b32 %r1993, %r1992, 3840; shl.b32 %r1994, %r139, 1; and.b32 %r1995, %r1994, 240; or.b32 %r1996, %r1995, %r1993; bfe.u32 %r1997, %r9, 3, 4; or.b32 %r1998, %r1996, %r1997; shr.u32 %r1999, %r1996, 6; mul.wide.u32 %rd1529, %r1999, 8; add.s64 %rd1530, %rd2311, %rd1529; and.b32 %r2000, %r1998, 63; shl.b64 %rd1532, %rd1518, %r2000; ld.u64 %rd1533, [%rd1530+544]; and.b64 %rd1534, %rd1533, %rd1532; setp.eq.s64 %p280, %rd1534, 0; mul.wide.u32 %rd1535, %r1998, 8; add.s64 %rd1536, %rd2311, %rd1535; add.s64 %rd379, %rd1536, 1088; @%p280 bra $L__BB6_447; ld.u64 %rd1537, [%rd379]; add.s64 %rd2307, %rd2311, %rd1537; and.b32 %r2001, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2001}; and.b32 %r3227, %r9, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; and.b32 %r2002, %r635, 7; shl.b32 %r2003, %r139, 3; and.b32 %r2004, %r2003, 56; and.b32 %r2005, %r9, 7; bfi.b32 %r2006, %r2002, %r2005, 6, 3; or.b32 %r2007, %r2006, %r2004; mul.wide.u32 %rd1538, %r2007, 2; add.s64 %rd1539, %rd2307, %rd1538; ld.u16 %rs254, [%rd1539+96]; cvt.rn.f32.u16 %f655, %rs254; ld.v2.f32 {%f656, %f657}, [%rd2307+80]; fma.rn.ftz.f32 %f941, %f657, %f655, %f656; mov.b64 %rd1540, {%r636, %r2001}; shr.u64 %rd1541, %rd1540, 32; cvt.u32.u64 %r3184, %rd1541; mov.u32 %r3232, %r640; mov.u32 %r3272, %r637; mov.u32 %r3229, %r636; bra.uni $L__BB6_451; $L__BB6_450: ld.f32 %f941, [%rd5+4]; bra.uni $L__BB6_451; $L__BB6_476: shl.b64 %rd1588, %rd400, 5; add.s64 %rd2374, %rd17, %rd1588; $L__BB6_477: setp.eq.s64 %p299, %rd2374, 0; @%p299 bra $L__BB6_485; add.s64 %rd403, %rd2374, 8; ld.u64 %rd404, [%rd2374+8]; setp.eq.s64 %p300, %rd404, 0; @%p300 bra $L__BB6_484; add.s64 %rd2315, %rd1, %rd404; and.b32 %r2068, %r139, -4096; mov.u64 %rd1589, 1; st.local.v2.u32 [%rd2+24], {%r673, %r2068}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2069, %r635, 3; and.b32 %r2070, %r2069, 31744; and.b32 %r2071, %r139, 3968; shr.u32 %r2072, %r2071, 2; or.b32 %r2073, %r2072, %r2070; bfe.u32 %r2074, %r74, 7, 5; or.b32 %r2075, %r2073, %r2074; shr.u32 %r2076, %r2073, 6; mul.wide.u32 %rd1590, %r2076, 8; add.s64 %rd1591, %rd2315, %rd1590; and.b32 %r2077, %r2075, 63; shl.b64 %rd1592, %rd1589, %r2077; ld.u64 %rd1593, [%rd1591+4128]; and.b64 %rd1594, %rd1593, %rd1592; setp.eq.s64 %p301, %rd1594, 0; mul.wide.u32 %rd1595, %r2075, 8; add.s64 %rd1596, %rd2315, %rd1595; add.s64 %rd406, %rd1596, 8256; mov.b64 %rd1597, {%r673, %r2068}; shr.u64 %rd1598, %rd1597, 32; cvt.u32.u64 %r3277, %rd1598; @%p301 bra $L__BB6_483; ld.u64 %rd1599, [%rd406]; add.s64 %rd2311, %rd2315, %rd1599; st.local.u32 [%rd2+12], %r670; and.b32 %r3271, %r139, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2078, %r635, 5; and.b32 %r2079, %r2078, 3840; shl.b32 %r2080, %r139, 1; and.b32 %r2081, %r2080, 240; or.b32 %r2082, %r2081, %r2079; bfe.u32 %r2083, %r74, 3, 4; or.b32 %r2084, %r2082, %r2083; shr.u32 %r2085, %r2082, 6; mul.wide.u32 %rd1600, %r2085, 8; add.s64 %rd1601, %rd2311, %rd1600; and.b32 %r2086, %r2084, 63; shl.b64 %rd1603, %rd1589, %r2086; ld.u64 %rd1604, [%rd1601+544]; and.b64 %rd1605, %rd1604, %rd1603; setp.eq.s64 %p302, %rd1605, 0; mul.wide.u32 %rd1606, %r2084, 8; add.s64 %rd1607, %rd2311, %rd1606; add.s64 %rd408, %rd1607, 1088; @%p302 bra $L__BB6_482; ld.u64 %rd1608, [%rd408]; add.s64 %rd2307, %rd2311, %rd1608; and.b32 %r2087, %r139, -8; st.local.v2.u32 [%rd2], {%r636, %r2087}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2088, %r635, 6; and.b32 %r2089, %r2088, 448; shl.b32 %r2090, %r139, 3; and.b32 %r2091, %r2090, 56; or.b32 %r2092, %r2091, %r2089; and.b32 %r2093, %r74, 7; or.b32 %r2094, %r2092, %r2093; mul.wide.u32 %rd1609, %r2094, 2; add.s64 %rd1610, %rd2307, %rd1609; ld.u16 %rs264, [%rd1610+96]; cvt.rn.f32.u16 %f675, %rs264; ld.v2.f32 {%f676, %f677}, [%rd2307+80]; fma.rn.ftz.f32 %f942, %f677, %f675, %f676; mov.b64 %rd1611, {%r636, %r2087}; shr.u64 %rd1612, %rd1611, 32; cvt.u32.u64 %r3184, %rd1612; mov.u32 %r3232, %r673; mov.u32 %r3272, %r670; mov.u32 %r3229, %r636; bra.uni $L__BB6_486; $L__BB6_485: ld.f32 %f942, [%rd5+4]; bra.uni $L__BB6_486; $L__BB6_511: shl.b64 %rd1659, %rd429, 5; add.s64 %rd2380, %rd17, %rd1659; $L__BB6_512: setp.eq.s64 %p321, %rd2380, 0; @%p321 bra $L__BB6_520; add.s64 %rd432, %rd2380, 8; ld.u64 %rd433, [%rd2380+8]; setp.eq.s64 %p322, %rd433, 0; @%p322 bra $L__BB6_519; add.s64 %rd2315, %rd1, %rd433; and.b32 %r2154, %r8, -4096; mov.u64 %rd1660, 1; st.local.v2.u32 [%rd2+24], {%r703, %r2154}; and.b32 %r3310, %r74, -4096; st.local.u32 [%rd2+32], %r3310; st.local.u64 [%rd2+64], %rd2315; shl.b32 %r2155, %r635, 3; and.b32 %r2156, %r2155, 31744; and.b32 %r2157, %r8, 3968; shr.u32 %r2158, %r2157, 2; or.b32 %r2159, %r2156, %r2158; bfe.u32 %r2160, %r74, 7, 5; or.b32 %r2161, %r2160, %r2159; shr.u32 %r2162, %r2159, 6; mul.wide.u32 %rd1661, %r2162, 8; add.s64 %rd1662, %rd2315, %rd1661; and.b32 %r2163, %r2161, 63; shl.b64 %rd1663, %rd1660, %r2163; ld.u64 %rd1664, [%rd1662+4128]; and.b64 %rd1665, %rd1664, %rd1663; setp.eq.s64 %p323, %rd1665, 0; mul.wide.u32 %rd1666, %r2161, 8; add.s64 %rd1667, %rd2315, %rd1666; add.s64 %rd435, %rd1667, 8256; mov.b64 %rd1668, {%r703, %r2154}; shr.u64 %rd1669, %rd1668, 32; cvt.u32.u64 %r3277, %rd1669; @%p323 bra $L__BB6_518; ld.u64 %rd1670, [%rd435]; add.s64 %rd2311, %rd2315, %rd1670; st.local.u32 [%rd2+12], %r700; and.b32 %r3271, %r8, -128; st.local.u32 [%rd2+16], %r3271; and.b32 %r3270, %r74, -128; st.local.u32 [%rd2+20], %r3270; st.local.u64 [%rd2+56], %rd2311; shl.b32 %r2164, %r635, 5; and.b32 %r2165, %r2164, 3840; shl.b32 %r2166, %r8, 1; and.b32 %r2167, %r2166, 240; or.b32 %r2168, %r2165, %r2167; bfe.u32 %r2169, %r74, 3, 4; or.b32 %r2170, %r2169, %r2168; shr.u32 %r2171, %r2168, 6; mul.wide.u32 %rd1671, %r2171, 8; add.s64 %rd1672, %rd2311, %rd1671; and.b32 %r2172, %r2170, 63; shl.b64 %rd1674, %rd1660, %r2172; ld.u64 %rd1675, [%rd1672+544]; and.b64 %rd1676, %rd1675, %rd1674; setp.eq.s64 %p324, %rd1676, 0; mul.wide.u32 %rd1677, %r2170, 8; add.s64 %rd1678, %rd2311, %rd1677; add.s64 %rd437, %rd1678, 1088; @%p324 bra $L__BB6_517; ld.u64 %rd1679, [%rd437]; add.s64 %rd2307, %rd2311, %rd1679; and.b32 %r2173, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2173}; and.b32 %r3227, %r74, -8; st.local.u32 [%rd2+8], %r3227; st.local.u64 [%rd2+48], %rd2307; shl.b32 %r2174, %r635, 6; and.b32 %r2175, %r2174, 448; shl.b32 %r2176, %r8, 3; and.b32 %r2177, %r2176, 56; or.b32 %r2178, %r2175, %r2177; and.b32 %r2179, %r74, 7; or.b32 %r2180, %r2178, %r2179; mul.wide.u32 %rd1680, %r2180, 2; add.s64 %rd1681, %rd2307, %rd1680; ld.u16 %rs274, [%rd1681+96]; cvt.rn.f32.u16 %f696, %rs274; ld.v2.f32 {%f697, %f698}, [%rd2307+80]; fma.rn.ftz.f32 %f943, %f698, %f696, %f697; mov.b64 %rd1682, {%r636, %r2173}; shr.u64 %rd1683, %rd1682, 32; cvt.u32.u64 %r3184, %rd1683; mov.u32 %r3232, %r703; mov.u32 %r3272, %r700; mov.u32 %r3229, %r636; bra.uni $L__BB6_521; $L__BB6_520: ld.f32 %f943, [%rd5+4]; bra.uni $L__BB6_521; $L__BB6_546: shl.b64 %rd1726, %rd455, 5; add.s64 %rd2386, %rd17, %rd1726; $L__BB6_547: setp.eq.s64 %p343, %rd2386, 0; @%p343 bra $L__BB6_555; add.s64 %rd458, %rd2386, 8; ld.u64 %rd459, [%rd2386+8]; setp.eq.s64 %p344, %rd459, 0; @%p344 bra $L__BB6_554; add.s64 %rd460, %rd1, %rd459; and.b32 %r2243, %r8, -4096; mov.u64 %rd1727, 1; st.local.v2.u32 [%rd2+24], {%r728, %r2243}; and.b32 %r2244, %r9, -4096; st.local.u32 [%rd2+32], %r2244; st.local.u64 [%rd2+64], %rd460; shl.b32 %r2245, %r635, 3; and.b32 %r2246, %r2245, 31744; and.b32 %r2247, %r8, 3968; shr.u32 %r2248, %r2247, 2; or.b32 %r2249, %r2246, %r2248; bfe.u32 %r2250, %r9, 7, 5; or.b32 %r2251, %r2249, %r2250; shr.u32 %r2252, %r2249, 6; mul.wide.u32 %rd1728, %r2252, 8; add.s64 %rd1729, %rd460, %rd1728; and.b32 %r2253, %r2251, 63; shl.b64 %rd1730, %rd1727, %r2253; ld.u64 %rd1731, [%rd1729+4128]; and.b64 %rd1732, %rd1731, %rd1730; setp.eq.s64 %p345, %rd1732, 0; mul.wide.u32 %rd1733, %r2251, 8; add.s64 %rd1734, %rd460, %rd1733; add.s64 %rd461, %rd1734, 8256; @%p345 bra $L__BB6_553; ld.u64 %rd1735, [%rd461]; add.s64 %rd462, %rd460, %rd1735; st.local.u32 [%rd2+12], %r727; and.b32 %r2254, %r8, -128; st.local.u32 [%rd2+16], %r2254; and.b32 %r2255, %r9, -128; st.local.u32 [%rd2+20], %r2255; st.local.u64 [%rd2+56], %rd462; shl.b32 %r2256, %r635, 5; and.b32 %r2257, %r2256, 3840; shl.b32 %r2258, %r8, 1; and.b32 %r2259, %r2258, 240; or.b32 %r2260, %r2257, %r2259; bfe.u32 %r2261, %r9, 3, 4; or.b32 %r2262, %r2260, %r2261; shr.u32 %r2263, %r2260, 6; mul.wide.u32 %rd1736, %r2263, 8; add.s64 %rd1737, %rd462, %rd1736; and.b32 %r2264, %r2262, 63; shl.b64 %rd1739, %rd1727, %r2264; ld.u64 %rd1740, [%rd1737+544]; and.b64 %rd1741, %rd1740, %rd1739; setp.eq.s64 %p346, %rd1741, 0; mul.wide.u32 %rd1742, %r2262, 8; add.s64 %rd1743, %rd462, %rd1742; add.s64 %rd463, %rd1743, 1088; @%p346 bra $L__BB6_552; ld.u64 %rd1744, [%rd463]; add.s64 %rd1745, %rd462, %rd1744; and.b32 %r2265, %r8, -8; st.local.v2.u32 [%rd2], {%r636, %r2265}; and.b32 %r2266, %r9, -8; st.local.u32 [%rd2+8], %r2266; st.local.u64 [%rd2+48], %rd1745; shl.b32 %r2267, %r635, 6; and.b32 %r2268, %r2267, 448; shl.b32 %r2269, %r8, 3; and.b32 %r2270, %r2269, 56; or.b32 %r2271, %r2268, %r2270; and.b32 %r2272, %r9, 7; or.b32 %r2273, %r2271, %r2272; mul.wide.u32 %rd1746, %r2273, 2; add.s64 %rd1747, %rd1745, %rd1746; ld.u16 %rs284, [%rd1747+96]; cvt.rn.f32.u16 %f716, %rs284; ld.v2.f32 {%f717, %f718}, [%rd1745+80]; fma.rn.ftz.f32 %f944, %f718, %f716, %f717; bra.uni $L__BB6_556; $L__BB6_555: ld.f32 %f944, [%rd5+4]; bra.uni $L__BB6_556; $L__BB6_581: shl.b64 %rd1794, %rd473, 5; add.s64 %rd2387, %rd471, %rd1794; $L__BB6_582: setp.eq.s64 %p365, %rd2387, 0; @%p365 bra $L__BB6_590; add.s64 %rd476, %rd2387, 8; ld.u64 %rd477, [%rd2387+8]; setp.eq.s64 %p366, %rd477, 0; @%p366 bra $L__BB6_589; add.s64 %rd478, %rd470, %rd477; and.b32 %r2346, %r732, -4096; mov.u64 %rd1795, 1; st.local.v2.u32 [%rd2+24], {%r736, %r2346}; and.b32 %r2347, %r733, -4096; st.local.u32 [%rd2+32], %r2347; st.local.u64 [%rd2+64], %rd478; shl.b32 %r2348, %r471, 3; and.b32 %r2349, %r2348, 31744; and.b32 %r2350, %r732, 3968; shr.u32 %r2351, %r2350, 2; or.b32 %r2352, %r2351, %r2349; bfe.u32 %r2353, %r733, 7, 5; or.b32 %r2354, %r2352, %r2353; shr.u32 %r2355, %r2352, 6; mul.wide.u32 %rd1796, %r2355, 8; add.s64 %rd1797, %rd478, %rd1796; and.b32 %r2356, %r2354, 63; shl.b64 %rd1798, %rd1795, %r2356; ld.u64 %rd1799, [%rd1797+4128]; and.b64 %rd1800, %rd1799, %rd1798; setp.eq.s64 %p367, %rd1800, 0; mul.wide.u32 %rd1801, %r2354, 8; add.s64 %rd1802, %rd478, %rd1801; add.s64 %rd479, %rd1802, 8256; @%p367 bra $L__BB6_588; ld.u64 %rd1803, [%rd479]; add.s64 %rd480, %rd478, %rd1803; st.local.u32 [%rd2+12], %r735; and.b32 %r2357, %r732, -128; st.local.u32 [%rd2+16], %r2357; and.b32 %r2358, %r733, -128; st.local.u32 [%rd2+20], %r2358; st.local.u64 [%rd2+56], %rd480; shl.b32 %r2359, %r471, 5; and.b32 %r2360, %r2359, 3840; shl.b32 %r2361, %r732, 1; and.b32 %r2362, %r2361, 240; or.b32 %r2363, %r2362, %r2360; bfe.u32 %r2364, %r733, 3, 4; or.b32 %r2365, %r2363, %r2364; shr.u32 %r2366, %r2363, 6; mul.wide.u32 %rd1804, %r2366, 8; add.s64 %rd1805, %rd480, %rd1804; and.b32 %r2367, %r2365, 63; shl.b64 %rd1807, %rd1795, %r2367; ld.u64 %rd1808, [%rd1805+544]; and.b64 %rd1809, %rd1808, %rd1807; setp.eq.s64 %p368, %rd1809, 0; mul.wide.u32 %rd1810, %r2365, 8; add.s64 %rd1811, %rd480, %rd1810; add.s64 %rd481, %rd1811, 1088; @%p368 bra $L__BB6_587; ld.u64 %rd1812, [%rd481]; add.s64 %rd1813, %rd480, %rd1812; and.b32 %r2368, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2368}; and.b32 %r2369, %r733, -8; st.local.u32 [%rd2+8], %r2369; st.local.u64 [%rd2+48], %rd1813; shl.b32 %r2370, %r471, 6; and.b32 %r2371, %r2370, 448; shl.b32 %r2372, %r732, 3; and.b32 %r2373, %r2372, 56; or.b32 %r2374, %r2373, %r2371; and.b32 %r2375, %r733, 7; or.b32 %r2376, %r2374, %r2375; mul.wide.u32 %rd1814, %r2376, 2; add.s64 %rd1815, %rd1813, %rd1814; ld.u16 %rs294, [%rd1815+96]; cvt.rn.f32.u16 %f757, %rs294; ld.v2.f32 {%f758, %f759}, [%rd1813+80]; fma.rn.ftz.f32 %f945, %f759, %f757, %f758; mov.u32 %r3598, %r472; bra.uni $L__BB6_591; $L__BB6_590: ld.f32 %f945, [%rd470+28]; bra.uni $L__BB6_591; $L__BB6_616: shl.b64 %rd1862, %rd491, 5; add.s64 %rd2388, %rd489, %rd1862; $L__BB6_617: setp.eq.s64 %p387, %rd2388, 0; @%p387 bra $L__BB6_625; add.s64 %rd494, %rd2388, 8; ld.u64 %rd495, [%rd2388+8]; setp.eq.s64 %p388, %rd495, 0; @%p388 bra $L__BB6_624; add.s64 %rd496, %rd488, %rd495; and.b32 %r2449, %r732, -4096; mov.u64 %rd1863, 1; st.local.v2.u32 [%rd2+24], {%r743, %r2449}; and.b32 %r2450, %r741, -4096; st.local.u32 [%rd2+32], %r2450; st.local.u64 [%rd2+64], %rd496; shl.b32 %r2451, %r471, 3; and.b32 %r2452, %r2451, 31744; and.b32 %r2453, %r732, 3968; shr.u32 %r2454, %r2453, 2; or.b32 %r2455, %r2454, %r2452; bfe.u32 %r2456, %r741, 7, 5; or.b32 %r2457, %r2456, %r2455; shr.u32 %r2458, %r2455, 6; mul.wide.u32 %rd1864, %r2458, 8; add.s64 %rd1865, %rd496, %rd1864; and.b32 %r2459, %r2457, 63; shl.b64 %rd1866, %rd1863, %r2459; ld.u64 %rd1867, [%rd1865+4128]; and.b64 %rd1868, %rd1867, %rd1866; setp.eq.s64 %p389, %rd1868, 0; mul.wide.u32 %rd1869, %r2457, 8; add.s64 %rd1870, %rd496, %rd1869; add.s64 %rd497, %rd1870, 8256; @%p389 bra $L__BB6_623; ld.u64 %rd1871, [%rd497]; add.s64 %rd498, %rd496, %rd1871; st.local.u32 [%rd2+12], %r742; and.b32 %r2460, %r732, -128; st.local.u32 [%rd2+16], %r2460; and.b32 %r2461, %r741, -128; st.local.u32 [%rd2+20], %r2461; st.local.u64 [%rd2+56], %rd498; shl.b32 %r2462, %r471, 5; and.b32 %r2463, %r2462, 3840; shl.b32 %r2464, %r732, 1; and.b32 %r2465, %r2464, 240; or.b32 %r2466, %r2465, %r2463; bfe.u32 %r2467, %r741, 3, 4; or.b32 %r2468, %r2467, %r2466; shr.u32 %r2469, %r2466, 6; mul.wide.u32 %rd1872, %r2469, 8; add.s64 %rd1873, %rd498, %rd1872; and.b32 %r2470, %r2468, 63; shl.b64 %rd1875, %rd1863, %r2470; ld.u64 %rd1876, [%rd1873+544]; and.b64 %rd1877, %rd1876, %rd1875; setp.eq.s64 %p390, %rd1877, 0; mul.wide.u32 %rd1878, %r2468, 8; add.s64 %rd1879, %rd498, %rd1878; add.s64 %rd499, %rd1879, 1088; @%p390 bra $L__BB6_622; ld.u64 %rd1880, [%rd499]; add.s64 %rd1881, %rd498, %rd1880; and.b32 %r2471, %r732, -8; st.local.v2.u32 [%rd2], {%r472, %r2471}; and.b32 %r2472, %r741, -8; st.local.u32 [%rd2+8], %r2472; st.local.u64 [%rd2+48], %rd1881; shl.b32 %r2473, %r471, 6; and.b32 %r2474, %r2473, 448; shl.b32 %r2475, %r732, 3; and.b32 %r2476, %r2475, 56; or.b32 %r2477, %r2476, %r2474; and.b32 %r2478, %r741, 7; or.b32 %r2479, %r2477, %r2478; mul.wide.u32 %rd1882, %r2479, 2; add.s64 %rd1883, %rd1881, %rd1882; ld.u16 %rs304, [%rd1883+96]; cvt.rn.f32.u16 %f777, %rs304; ld.v2.f32 {%f778, %f779}, [%rd1881+80]; fma.rn.ftz.f32 %f946, %f779, %f777, %f778; mov.u32 %r3598, %r472; bra.uni $L__BB6_626; $L__BB6_625: ld.f32 %f946, [%rd488+28]; bra.uni $L__BB6_626; $L__BB6_651: shl.b64 %rd1929, %rd509, 5; add.s64 %rd2389, %rd507, %rd1929; $L__BB6_652: setp.eq.s64 %p409, %rd2389, 0; @%p409 bra $L__BB6_660; add.s64 %rd512, %rd2389, 8; ld.u64 %rd513, [%rd2389+8]; setp.eq.s64 %p410, %rd513, 0; @%p410 bra $L__BB6_659; add.s64 %rd514, %rd506, %rd513; and.b32 %r2552, %r748, -4096; mov.u64 %rd1930, 1; st.local.v2.u32 [%rd2+24], {%r750, %r2552}; and.b32 %r2553, %r741, -4096; st.local.u32 [%rd2+32], %r2553; st.local.u64 [%rd2+64], %rd514; shl.b32 %r2554, %r471, 3; and.b32 %r2555, %r2554, 31744; and.b32 %r2556, %r748, 3968; shr.u32 %r2557, %r2556, 2; or.b32 %r2558, %r2557, %r2555; bfe.u32 %r2559, %r741, 7, 5; or.b32 %r2560, %r2558, %r2559; shr.u32 %r2561, %r2558, 6; mul.wide.u32 %rd1931, %r2561, 8; add.s64 %rd1932, %rd514, %rd1931; and.b32 %r2562, %r2560, 63; shl.b64 %rd1933, %rd1930, %r2562; ld.u64 %rd1934, [%rd1932+4128]; and.b64 %rd1935, %rd1934, %rd1933; setp.eq.s64 %p411, %rd1935, 0; mul.wide.u32 %rd1936, %r2560, 8; add.s64 %rd1937, %rd514, %rd1936; add.s64 %rd515, %rd1937, 8256; @%p411 bra $L__BB6_658; ld.u64 %rd1938, [%rd515]; add.s64 %rd516, %rd514, %rd1938; st.local.u32 [%rd2+12], %r749; and.b32 %r2563, %r748, -128; st.local.u32 [%rd2+16], %r2563; and.b32 %r2564, %r741, -128; st.local.u32 [%rd2+20], %r2564; st.local.u64 [%rd2+56], %rd516; shl.b32 %r2565, %r471, 5; and.b32 %r2566, %r2565, 3840; shl.b32 %r2567, %r748, 1; and.b32 %r2568, %r2567, 240; or.b32 %r2569, %r2568, %r2566; bfe.u32 %r2570, %r741, 3, 4; or.b32 %r2571, %r2569, %r2570; shr.u32 %r2572, %r2569, 6; mul.wide.u32 %rd1939, %r2572, 8; add.s64 %rd1940, %rd516, %rd1939; and.b32 %r2573, %r2571, 63; shl.b64 %rd1942, %rd1930, %r2573; ld.u64 %rd1943, [%rd1940+544]; and.b64 %rd1944, %rd1943, %rd1942; setp.eq.s64 %p412, %rd1944, 0; mul.wide.u32 %rd1945, %r2571, 8; add.s64 %rd1946, %rd516, %rd1945; add.s64 %rd517, %rd1946, 1088; @%p412 bra $L__BB6_657; ld.u64 %rd1947, [%rd517]; add.s64 %rd1948, %rd516, %rd1947; and.b32 %r2574, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2574}; and.b32 %r2575, %r741, -8; st.local.u32 [%rd2+8], %r2575; st.local.u64 [%rd2+48], %rd1948; shl.b32 %r2576, %r471, 6; and.b32 %r2577, %r2576, 448; shl.b32 %r2578, %r748, 3; and.b32 %r2579, %r2578, 56; or.b32 %r2580, %r2579, %r2577; and.b32 %r2581, %r741, 7; or.b32 %r2582, %r2580, %r2581; mul.wide.u32 %rd1949, %r2582, 2; add.s64 %rd1950, %rd1948, %rd1949; ld.u16 %rs314, [%rd1950+96]; cvt.rn.f32.u16 %f798, %rs314; ld.v2.f32 {%f799, %f800}, [%rd1948+80]; fma.rn.ftz.f32 %f947, %f800, %f798, %f799; mov.u32 %r3598, %r472; bra.uni $L__BB6_661; $L__BB6_660: ld.f32 %f947, [%rd506+28]; bra.uni $L__BB6_661; $L__BB6_686: shl.b64 %rd1996, %rd527, 5; add.s64 %rd2390, %rd525, %rd1996; $L__BB6_687: setp.eq.s64 %p431, %rd2390, 0; @%p431 bra $L__BB6_695; add.s64 %rd530, %rd2390, 8; ld.u64 %rd531, [%rd2390+8]; setp.eq.s64 %p432, %rd531, 0; @%p432 bra $L__BB6_694; add.s64 %rd532, %rd524, %rd531; and.b32 %r2652, %r748, -4096; mov.u64 %rd1997, 1; st.local.v2.u32 [%rd2+24], {%r756, %r2652}; and.b32 %r2653, %r733, -4096; st.local.u32 [%rd2+32], %r2653; st.local.u64 [%rd2+64], %rd532; shl.b32 %r2654, %r471, 3; and.b32 %r2655, %r2654, 31744; and.b32 %r2656, %r748, 3968; shr.u32 %r2657, %r2656, 2; or.b32 %r2658, %r2657, %r2655; bfe.u32 %r2659, %r733, 7, 5; or.b32 %r2660, %r2658, %r2659; shr.u32 %r2661, %r2658, 6; mul.wide.u32 %rd1998, %r2661, 8; add.s64 %rd1999, %rd532, %rd1998; and.b32 %r2662, %r2660, 63; shl.b64 %rd2000, %rd1997, %r2662; ld.u64 %rd2001, [%rd1999+4128]; and.b64 %rd2002, %rd2001, %rd2000; setp.eq.s64 %p433, %rd2002, 0; mul.wide.u32 %rd2003, %r2660, 8; add.s64 %rd2004, %rd532, %rd2003; add.s64 %rd533, %rd2004, 8256; @%p433 bra $L__BB6_693; ld.u64 %rd2005, [%rd533]; add.s64 %rd534, %rd532, %rd2005; st.local.u32 [%rd2+12], %r755; and.b32 %r2663, %r748, -128; st.local.u32 [%rd2+16], %r2663; and.b32 %r2664, %r733, -128; st.local.u32 [%rd2+20], %r2664; st.local.u64 [%rd2+56], %rd534; shl.b32 %r2665, %r471, 5; and.b32 %r2666, %r2665, 3840; shl.b32 %r2667, %r748, 1; and.b32 %r2668, %r2667, 240; or.b32 %r2669, %r2668, %r2666; bfe.u32 %r2670, %r733, 3, 4; or.b32 %r2671, %r2669, %r2670; shr.u32 %r2672, %r2669, 6; mul.wide.u32 %rd2006, %r2672, 8; add.s64 %rd2007, %rd534, %rd2006; and.b32 %r2673, %r2671, 63; shl.b64 %rd2009, %rd1997, %r2673; ld.u64 %rd2010, [%rd2007+544]; and.b64 %rd2011, %rd2010, %rd2009; setp.eq.s64 %p434, %rd2011, 0; mul.wide.u32 %rd2012, %r2671, 8; add.s64 %rd2013, %rd534, %rd2012; add.s64 %rd535, %rd2013, 1088; @%p434 bra $L__BB6_692; ld.u64 %rd2014, [%rd535]; add.s64 %rd2015, %rd534, %rd2014; and.b32 %r2674, %r748, -8; st.local.v2.u32 [%rd2], {%r472, %r2674}; and.b32 %r2675, %r733, -8; st.local.u32 [%rd2+8], %r2675; st.local.u64 [%rd2+48], %rd2015; and.b32 %r2676, %r471, 7; shl.b32 %r2677, %r748, 3; and.b32 %r2678, %r2677, 56; and.b32 %r2679, %r733, 7; bfi.b32 %r2680, %r2676, %r2679, 6, 3; or.b32 %r2681, %r2680, %r2678; mul.wide.u32 %rd2016, %r2681, 2; add.s64 %rd2017, %rd2015, %rd2016; ld.u16 %rs324, [%rd2017+96]; cvt.rn.f32.u16 %f818, %rs324; ld.v2.f32 {%f819, %f820}, [%rd2015+80]; fma.rn.ftz.f32 %f948, %f820, %f818, %f819; mov.u32 %r3598, %r472; bra.uni $L__BB6_696; $L__BB6_695: ld.f32 %f948, [%rd524+28]; bra.uni $L__BB6_696; $L__BB6_721: shl.b64 %rd2063, %rd545, 5; add.s64 %rd2391, %rd543, %rd2063; $L__BB6_722: setp.eq.s64 %p453, %rd2391, 0; @%p453 bra $L__BB6_730; add.s64 %rd548, %rd2391, 8; ld.u64 %rd549, [%rd2391+8]; setp.eq.s64 %p454, %rd549, 0; @%p454 bra $L__BB6_729; add.s64 %rd550, %rd542, %rd549; and.b32 %r2751, %r748, -4096; mov.u64 %rd2064, 1; st.local.v2.u32 [%rd2+24], {%r762, %r2751}; and.b32 %r2752, %r733, -4096; st.local.u32 [%rd2+32], %r2752; st.local.u64 [%rd2+64], %rd550; shl.b32 %r2753, %r635, 3; and.b32 %r2754, %r2753, 31744; and.b32 %r2755, %r748, 3968; shr.u32 %r2756, %r2755, 2; or.b32 %r2757, %r2756, %r2754; bfe.u32 %r2758, %r733, 7, 5; or.b32 %r2759, %r2757, %r2758; shr.u32 %r2760, %r2757, 6; mul.wide.u32 %rd2065, %r2760, 8; add.s64 %rd2066, %rd550, %rd2065; and.b32 %r2761, %r2759, 63; shl.b64 %rd2067, %rd2064, %r2761; ld.u64 %rd2068, [%rd2066+4128]; and.b64 %rd2069, %rd2068, %rd2067; setp.eq.s64 %p455, %rd2069, 0; mul.wide.u32 %rd2070, %r2759, 8; add.s64 %rd2071, %rd550, %rd2070; add.s64 %rd551, %rd2071, 8256; @%p455 bra $L__BB6_728; ld.u64 %rd2072, [%rd551]; add.s64 %rd552, %rd550, %rd2072; st.local.u32 [%rd2+12], %r761; and.b32 %r2762, %r748, -128; st.local.u32 [%rd2+16], %r2762; and.b32 %r2763, %r733, -128; st.local.u32 [%rd2+20], %r2763; st.local.u64 [%rd2+56], %rd552; shl.b32 %r2764, %r635, 5; and.b32 %r2765, %r2764, 3840; shl.b32 %r2766, %r748, 1; and.b32 %r2767, %r2766, 240; or.b32 %r2768, %r2767, %r2765; bfe.u32 %r2769, %r733, 3, 4; or.b32 %r2770, %r2768, %r2769; shr.u32 %r2771, %r2768, 6; mul.wide.u32 %rd2073, %r2771, 8; add.s64 %rd2074, %rd552, %rd2073; and.b32 %r2772, %r2770, 63; shl.b64 %rd2076, %rd2064, %r2772; ld.u64 %rd2077, [%rd2074+544]; and.b64 %rd2078, %rd2077, %rd2076; setp.eq.s64 %p456, %rd2078, 0; mul.wide.u32 %rd2079, %r2770, 8; add.s64 %rd2080, %rd552, %rd2079; add.s64 %rd553, %rd2080, 1088; @%p456 bra $L__BB6_727; ld.u64 %rd2081, [%rd553]; add.s64 %rd2082, %rd552, %rd2081; and.b32 %r2773, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2773}; and.b32 %r2774, %r733, -8; st.local.u32 [%rd2+8], %r2774; st.local.u64 [%rd2+48], %rd2082; and.b32 %r2775, %r635, 7; shl.b32 %r2776, %r748, 3; and.b32 %r2777, %r2776, 56; and.b32 %r2778, %r733, 7; bfi.b32 %r2779, %r2775, %r2778, 6, 3; or.b32 %r2780, %r2779, %r2777; mul.wide.u32 %rd2083, %r2780, 2; add.s64 %rd2084, %rd2082, %rd2083; ld.u16 %rs334, [%rd2084+96]; cvt.rn.f32.u16 %f841, %rs334; ld.v2.f32 {%f842, %f843}, [%rd2082+80]; fma.rn.ftz.f32 %f949, %f843, %f841, %f842; mov.u32 %r3598, %r636; bra.uni $L__BB6_731; $L__BB6_730: ld.f32 %f949, [%rd542+28]; bra.uni $L__BB6_731; $L__BB6_756: shl.b64 %rd2130, %rd563, 5; add.s64 %rd2392, %rd561, %rd2130; $L__BB6_757: setp.eq.s64 %p475, %rd2392, 0; @%p475 bra $L__BB6_765; add.s64 %rd566, %rd2392, 8; ld.u64 %rd567, [%rd2392+8]; setp.eq.s64 %p476, %rd567, 0; @%p476 bra $L__BB6_764; add.s64 %rd568, %rd560, %rd567; and.b32 %r2853, %r748, -4096; mov.u64 %rd2131, 1; st.local.v2.u32 [%rd2+24], {%r768, %r2853}; and.b32 %r2854, %r741, -4096; st.local.u32 [%rd2+32], %r2854; st.local.u64 [%rd2+64], %rd568; shl.b32 %r2855, %r635, 3; and.b32 %r2856, %r2855, 31744; and.b32 %r2857, %r748, 3968; shr.u32 %r2858, %r2857, 2; or.b32 %r2859, %r2858, %r2856; bfe.u32 %r2860, %r741, 7, 5; or.b32 %r2861, %r2859, %r2860; shr.u32 %r2862, %r2859, 6; mul.wide.u32 %rd2132, %r2862, 8; add.s64 %rd2133, %rd568, %rd2132; and.b32 %r2863, %r2861, 63; shl.b64 %rd2134, %rd2131, %r2863; ld.u64 %rd2135, [%rd2133+4128]; and.b64 %rd2136, %rd2135, %rd2134; setp.eq.s64 %p477, %rd2136, 0; mul.wide.u32 %rd2137, %r2861, 8; add.s64 %rd2138, %rd568, %rd2137; add.s64 %rd569, %rd2138, 8256; @%p477 bra $L__BB6_763; ld.u64 %rd2139, [%rd569]; add.s64 %rd570, %rd568, %rd2139; st.local.u32 [%rd2+12], %r767; and.b32 %r2864, %r748, -128; st.local.u32 [%rd2+16], %r2864; and.b32 %r2865, %r741, -128; st.local.u32 [%rd2+20], %r2865; st.local.u64 [%rd2+56], %rd570; shl.b32 %r2866, %r635, 5; and.b32 %r2867, %r2866, 3840; shl.b32 %r2868, %r748, 1; and.b32 %r2869, %r2868, 240; or.b32 %r2870, %r2869, %r2867; bfe.u32 %r2871, %r741, 3, 4; or.b32 %r2872, %r2870, %r2871; shr.u32 %r2873, %r2870, 6; mul.wide.u32 %rd2140, %r2873, 8; add.s64 %rd2141, %rd570, %rd2140; and.b32 %r2874, %r2872, 63; shl.b64 %rd2143, %rd2131, %r2874; ld.u64 %rd2144, [%rd2141+544]; and.b64 %rd2145, %rd2144, %rd2143; setp.eq.s64 %p478, %rd2145, 0; mul.wide.u32 %rd2146, %r2872, 8; add.s64 %rd2147, %rd570, %rd2146; add.s64 %rd571, %rd2147, 1088; @%p478 bra $L__BB6_762; ld.u64 %rd2148, [%rd571]; add.s64 %rd2149, %rd570, %rd2148; and.b32 %r2875, %r748, -8; st.local.v2.u32 [%rd2], {%r636, %r2875}; and.b32 %r2876, %r741, -8; st.local.u32 [%rd2+8], %r2876; st.local.u64 [%rd2+48], %rd2149; shl.b32 %r2877, %r635, 6; and.b32 %r2878, %r2877, 448; shl.b32 %r2879, %r748, 3; and.b32 %r2880, %r2879, 56; or.b32 %r2881, %r2880, %r2878; and.b32 %r2882, %r741, 7; or.b32 %r2883, %r2881, %r2882; mul.wide.u32 %rd2150, %r2883, 2; add.s64 %rd2151, %rd2149, %rd2150; ld.u16 %rs344, [%rd2151+96]; cvt.rn.f32.u16 %f861, %rs344; ld.v2.f32 {%f862, %f863}, [%rd2149+80]; fma.rn.ftz.f32 %f950, %f863, %f861, %f862; mov.u32 %r3598, %r636; bra.uni $L__BB6_766; $L__BB6_765: ld.f32 %f950, [%rd560+28]; bra.uni $L__BB6_766; $L__BB6_791: shl.b64 %rd2198, %rd581, 5; add.s64 %rd2393, %rd579, %rd2198; $L__BB6_792: setp.eq.s64 %p497, %rd2393, 0; @%p497 bra $L__BB6_800; add.s64 %rd584, %rd2393, 8; ld.u64 %rd585, [%rd2393+8]; setp.eq.s64 %p498, %rd585, 0; @%p498 bra $L__BB6_799; add.s64 %rd586, %rd578, %rd585; and.b32 %r2956, %r732, -4096; mov.u64 %rd2199, 1; st.local.v2.u32 [%rd2+24], {%r774, %r2956}; and.b32 %r2957, %r741, -4096; st.local.u32 [%rd2+32], %r2957; st.local.u64 [%rd2+64], %rd586; shl.b32 %r2958, %r635, 3; and.b32 %r2959, %r2958, 31744; and.b32 %r2960, %r732, 3968; shr.u32 %r2961, %r2960, 2; or.b32 %r2962, %r2959, %r2961; bfe.u32 %r2963, %r741, 7, 5; or.b32 %r2964, %r2963, %r2962; shr.u32 %r2965, %r2962, 6; mul.wide.u32 %rd2200, %r2965, 8; add.s64 %rd2201, %rd586, %rd2200; and.b32 %r2966, %r2964, 63; shl.b64 %rd2202, %rd2199, %r2966; ld.u64 %rd2203, [%rd2201+4128]; and.b64 %rd2204, %rd2203, %rd2202; setp.eq.s64 %p499, %rd2204, 0; mul.wide.u32 %rd2205, %r2964, 8; add.s64 %rd2206, %rd586, %rd2205; add.s64 %rd587, %rd2206, 8256; @%p499 bra $L__BB6_798; ld.u64 %rd2207, [%rd587]; add.s64 %rd588, %rd586, %rd2207; st.local.u32 [%rd2+12], %r773; and.b32 %r2967, %r732, -128; st.local.u32 [%rd2+16], %r2967; and.b32 %r2968, %r741, -128; st.local.u32 [%rd2+20], %r2968; st.local.u64 [%rd2+56], %rd588; shl.b32 %r2969, %r635, 5; and.b32 %r2970, %r2969, 3840; shl.b32 %r2971, %r732, 1; and.b32 %r2972, %r2971, 240; or.b32 %r2973, %r2970, %r2972; bfe.u32 %r2974, %r741, 3, 4; or.b32 %r2975, %r2974, %r2973; shr.u32 %r2976, %r2973, 6; mul.wide.u32 %rd2208, %r2976, 8; add.s64 %rd2209, %rd588, %rd2208; and.b32 %r2977, %r2975, 63; shl.b64 %rd2211, %rd2199, %r2977; ld.u64 %rd2212, [%rd2209+544]; and.b64 %rd2213, %rd2212, %rd2211; setp.eq.s64 %p500, %rd2213, 0; mul.wide.u32 %rd2214, %r2975, 8; add.s64 %rd2215, %rd588, %rd2214; add.s64 %rd589, %rd2215, 1088; @%p500 bra $L__BB6_797; ld.u64 %rd2216, [%rd589]; add.s64 %rd2217, %rd588, %rd2216; and.b32 %r2978, %r732, -8; st.local.v2.u32 [%rd2], {%r636, %r2978}; and.b32 %r2979, %r741, -8; st.local.u32 [%rd2+8], %r2979; st.local.u64 [%rd2+48], %rd2217; shl.b32 %r2980, %r635, 6; and.b32 %r2981, %r2980, 448; shl.b32 %r2982, %r732, 3; and.b32 %r2983, %r2982, 56; or.b32 %r2984, %r2981, %r2983; and.b32 %r2985, %r741, 7; or.b32 %r2986, %r2984, %r2985; mul.wide.u32 %rd2218, %r2986, 2; add.s64 %rd2219, %rd2217, %rd2218; ld.u16 %rs354, [%rd2219+96]; cvt.rn.f32.u16 %f882, %rs354; ld.v2.f32 {%f883, %f884}, [%rd2217+80]; fma.rn.ftz.f32 %f951, %f884, %f882, %f883; mov.u32 %r3598, %r636; bra.uni $L__BB6_801; $L__BB6_800: ld.f32 %f951, [%rd578+28]; bra.uni $L__BB6_801; $L__BB6_64: ld.f32 %f930, [%rd55+12]; bra.uni $L__BB6_66; $L__BB6_99: ld.f32 %f931, [%rd84+12]; bra.uni $L__BB6_101; $L__BB6_134: ld.f32 %f932, [%rd113+12]; bra.uni $L__BB6_136; $L__BB6_169: ld.f32 %f933, [%rd142+12]; bra.uni $L__BB6_171; $L__BB6_204: ld.f32 %f934, [%rd171+12]; bra.uni $L__BB6_206; $L__BB6_239: ld.f32 %f935, [%rd200+12]; bra.uni $L__BB6_241; $L__BB6_274: ld.f32 %f936, [%rd229+12]; bra.uni $L__BB6_276; $L__BB6_309: ld.f32 %f937, [%rd258+12]; bra.uni $L__BB6_311; $L__BB6_344: ld.f32 %f938, [%rd287+12]; bra.uni $L__BB6_346; $L__BB6_379: ld.f32 %f939, [%rd316+12]; bra.uni $L__BB6_381; $L__BB6_414: ld.f32 %f940, [%rd345+12]; bra.uni $L__BB6_416; $L__BB6_449: ld.f32 %f941, [%rd374+12]; bra.uni $L__BB6_451; $L__BB6_484: ld.f32 %f942, [%rd403+12]; bra.uni $L__BB6_486; $L__BB6_519: ld.f32 %f943, [%rd432+12]; bra.uni $L__BB6_521; $L__BB6_554: ld.f32 %f944, [%rd458+12]; bra.uni $L__BB6_556; $L__BB6_589: ld.f32 %f945, [%rd476+12]; bra.uni $L__BB6_591; $L__BB6_624: ld.f32 %f946, [%rd494+12]; bra.uni $L__BB6_626; $L__BB6_659: ld.f32 %f947, [%rd512+12]; bra.uni $L__BB6_661; $L__BB6_694: ld.f32 %f948, [%rd530+12]; bra.uni $L__BB6_696; $L__BB6_729: ld.f32 %f949, [%rd548+12]; bra.uni $L__BB6_731; $L__BB6_764: ld.f32 %f950, [%rd566+12]; bra.uni $L__BB6_766; $L__BB6_799: ld.f32 %f951, [%rd584+12]; bra.uni $L__BB6_801; $L__BB6_63: ld.f32 %f930, [%rd58]; mov.u32 %r3232, %r12; bra.uni $L__BB6_66; $L__BB6_98: ld.f32 %f931, [%rd87]; mov.u32 %r3232, %r12; bra.uni $L__BB6_101; $L__BB6_133: ld.f32 %f932, [%rd116]; mov.u32 %r3232, %r12; bra.uni $L__BB6_136; $L__BB6_168: and.b32 %r3232, %r259, -4096; ld.f32 %f933, [%rd145]; bra.uni $L__BB6_171; $L__BB6_203: and.b32 %r3232, %r259, -4096; ld.f32 %f934, [%rd174]; bra.uni $L__BB6_206; $L__BB6_238: and.b32 %r3232, %r259, -4096; ld.f32 %f935, [%rd203]; bra.uni $L__BB6_241; $L__BB6_273: ld.f32 %f936, [%rd232]; mov.u32 %r3232, %r426; bra.uni $L__BB6_276; $L__BB6_308: ld.f32 %f937, [%rd261]; mov.u32 %r3232, %r476; bra.uni $L__BB6_311; $L__BB6_343: ld.f32 %f938, [%rd290]; mov.u32 %r3232, %r521; bra.uni $L__BB6_346; $L__BB6_378: ld.f32 %f939, [%rd319]; mov.u32 %r3232, %r563; bra.uni $L__BB6_381; $L__BB6_413: ld.f32 %f940, [%rd348]; mov.u32 %r3232, %r602; bra.uni $L__BB6_416; $L__BB6_448: ld.f32 %f941, [%rd377]; mov.u32 %r3232, %r640; bra.uni $L__BB6_451; $L__BB6_483: ld.f32 %f942, [%rd406]; mov.u32 %r3232, %r673; bra.uni $L__BB6_486; $L__BB6_518: ld.f32 %f943, [%rd435]; mov.u32 %r3232, %r703; bra.uni $L__BB6_521; $L__BB6_553: ld.f32 %f944, [%rd461]; bra.uni $L__BB6_556; $L__BB6_588: ld.f32 %f945, [%rd479]; bra.uni $L__BB6_591; $L__BB6_623: ld.f32 %f946, [%rd497]; bra.uni $L__BB6_626; $L__BB6_658: ld.f32 %f947, [%rd515]; bra.uni $L__BB6_661; $L__BB6_693: ld.f32 %f948, [%rd533]; bra.uni $L__BB6_696; $L__BB6_728: ld.f32 %f949, [%rd551]; bra.uni $L__BB6_731; $L__BB6_763: ld.f32 %f950, [%rd569]; bra.uni $L__BB6_766; $L__BB6_798: ld.f32 %f951, [%rd587]; bra.uni $L__BB6_801; $L__BB6_62: ld.f32 %f930, [%rd60]; mov.u32 %r3232, %r12; bra.uni $L__BB6_66; $L__BB6_97: ld.f32 %f931, [%rd89]; mov.u32 %r3232, %r12; bra.uni $L__BB6_101; $L__BB6_132: ld.f32 %f932, [%rd118]; mov.u32 %r3232, %r12; bra.uni $L__BB6_136; $L__BB6_167: and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.f32 %f933, [%rd147]; bra.uni $L__BB6_171; $L__BB6_202: and.b32 %r3232, %r259, -4096; and.b32 %r3272, %r259, -128; ld.f32 %f934, [%rd176]; bra.uni $L__BB6_206; $L__BB6_237: ld.f32 %f935, [%rd205]; bra.uni $L__BB6_241; $L__BB6_272: ld.f32 %f936, [%rd234]; mov.u32 %r3232, %r426; mov.u32 %r3272, %r423; bra.uni $L__BB6_276; $L__BB6_307: ld.f32 %f937, [%rd263]; mov.u32 %r3232, %r476; mov.u32 %r3272, %r473; bra.uni $L__BB6_311; $L__BB6_342: ld.f32 %f938, [%rd292]; mov.u32 %r3232, %r521; mov.u32 %r3272, %r518; bra.uni $L__BB6_346; $L__BB6_377: ld.f32 %f939, [%rd321]; mov.u32 %r3232, %r563; mov.u32 %r3272, %r560; bra.uni $L__BB6_381; $L__BB6_412: ld.f32 %f940, [%rd350]; mov.u32 %r3232, %r602; mov.u32 %r3272, %r599; bra.uni $L__BB6_416; $L__BB6_447: ld.f32 %f941, [%rd379]; mov.u32 %r3232, %r640; mov.u32 %r3272, %r637; bra.uni $L__BB6_451; $L__BB6_482: ld.f32 %f942, [%rd408]; mov.u32 %r3232, %r673; mov.u32 %r3272, %r670; bra.uni $L__BB6_486; $L__BB6_517: ld.f32 %f943, [%rd437]; mov.u32 %r3232, %r703; mov.u32 %r3272, %r700; bra.uni $L__BB6_521; $L__BB6_552: ld.f32 %f944, [%rd463]; bra.uni $L__BB6_556; $L__BB6_587: ld.f32 %f945, [%rd481]; bra.uni $L__BB6_591; $L__BB6_622: ld.f32 %f946, [%rd499]; bra.uni $L__BB6_626; $L__BB6_657: ld.f32 %f947, [%rd517]; bra.uni $L__BB6_661; $L__BB6_692: ld.f32 %f948, [%rd535]; bra.uni $L__BB6_696; $L__BB6_727: ld.f32 %f949, [%rd553]; bra.uni $L__BB6_731; $L__BB6_762: ld.f32 %f950, [%rd571]; bra.uni $L__BB6_766; $L__BB6_797: ld.f32 %f951, [%rd589]; bra.uni $L__BB6_801; } .func (.param .align 4 .b8 func_retval0[16]) _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff( .param .b64 _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_0, .param .b32 _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_1, .param .b32 _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_2, .param .b32 _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_3 ) { .reg .f32 %f<180>; .reg .b32 %r<6>; .reg .b64 %rd<4>; ld.param.u64 %rd1, [_Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_0]; ld.param.f32 %f1, [_Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_1]; ld.param.f32 %f2, [_Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_2]; ld.param.f32 %f3, [_Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff_param_3]; cvta.to.global.u64 %rd2, %rd1; ld.global.u64 %rd3, [%rd2]; ld.global.u32 %r1, [%rd2+20]; cvt.rn.f32.u32 %f4, %r1; fma.rn.ftz.f32 %f5, %f4, %f1, 0fBF000000; ld.global.v2.u32 {%r2, %r3}, [%rd2+24]; cvt.rn.f32.u32 %f6, %r2; fma.rn.ftz.f32 %f7, %f6, %f2, 0fBF000000; cvt.rn.f32.u32 %f8, %r3; fma.rn.ftz.f32 %f9, %f8, %f3, 0fBF000000; cvt.rmi.ftz.f32.f32 %f10, %f5; cvt.rmi.ftz.f32.f32 %f11, %f7; cvt.rmi.ftz.f32.f32 %f12, %f9; sub.ftz.f32 %f13, %f5, %f10; sub.ftz.f32 %f14, %f7, %f11; sub.ftz.f32 %f15, %f9, %f12; mov.f32 %f16, 0f40400000; sub.ftz.f32 %f17, %f16, %f13; fma.rn.ftz.f32 %f18, %f13, %f17, 0fC0400000; fma.rn.ftz.f32 %f19, %f13, %f18, 0f3F800000; mul.ftz.f32 %f20, %f13, %f13; mul.ftz.f32 %f21, %f13, 0f40400000; add.ftz.f32 %f22, %f21, 0fC0C00000; fma.rn.ftz.f32 %f23, %f20, %f22, 0f40800000; mul.ftz.f32 %f24, %f23, 0f3E2AAAAB; fma.rn.ftz.f32 %f25, %f19, 0f3E2AAAAB, %f24; sub.ftz.f32 %f26, %f16, %f21; fma.rn.ftz.f32 %f27, %f13, %f26, 0f40400000; fma.rn.ftz.f32 %f28, %f13, %f27, 0f3F800000; mul.ftz.f32 %f29, %f13, %f20; mul.ftz.f32 %f30, %f29, 0f3E2AAAAB; fma.rn.ftz.f32 %f31, %f28, 0f3E2AAAAB, %f30; sub.ftz.f32 %f32, %f16, %f14; fma.rn.ftz.f32 %f33, %f14, %f32, 0fC0400000; fma.rn.ftz.f32 %f34, %f14, %f33, 0f3F800000; mul.ftz.f32 %f35, %f14, %f14; mul.ftz.f32 %f36, %f14, 0f40400000; add.ftz.f32 %f37, %f36, 0fC0C00000; fma.rn.ftz.f32 %f38, %f35, %f37, 0f40800000; mul.ftz.f32 %f39, %f38, 0f3E2AAAAB; fma.rn.ftz.f32 %f40, %f34, 0f3E2AAAAB, %f39; sub.ftz.f32 %f41, %f16, %f36; fma.rn.ftz.f32 %f42, %f14, %f41, 0f40400000; fma.rn.ftz.f32 %f43, %f14, %f42, 0f3F800000; mul.ftz.f32 %f44, %f14, %f35; mul.ftz.f32 %f45, %f44, 0f3E2AAAAB; fma.rn.ftz.f32 %f46, %f43, 0f3E2AAAAB, %f45; sub.ftz.f32 %f47, %f16, %f15; fma.rn.ftz.f32 %f48, %f15, %f47, 0fC0400000; fma.rn.ftz.f32 %f49, %f15, %f48, 0f3F800000; mul.ftz.f32 %f50, %f15, %f15; mul.ftz.f32 %f51, %f15, 0f40400000; add.ftz.f32 %f52, %f51, 0fC0C00000; fma.rn.ftz.f32 %f53, %f50, %f52, 0f40800000; mul.ftz.f32 %f54, %f53, 0f3E2AAAAB; fma.rn.ftz.f32 %f55, %f49, 0f3E2AAAAB, %f54; sub.ftz.f32 %f56, %f16, %f51; fma.rn.ftz.f32 %f57, %f15, %f56, 0f40400000; fma.rn.ftz.f32 %f58, %f15, %f57, 0f3F800000; mul.ftz.f32 %f59, %f15, %f50; mul.ftz.f32 %f60, %f59, 0f3E2AAAAB; fma.rn.ftz.f32 %f61, %f58, 0f3E2AAAAB, %f60; div.approx.ftz.f32 %f62, %f24, %f25; add.ftz.f32 %f63, %f62, 0fBF800000; add.ftz.f32 %f64, %f10, %f63; add.ftz.f32 %f65, %f64, 0f3F000000; div.approx.ftz.f32 %f66, %f65, %f4; div.approx.ftz.f32 %f67, %f30, %f31; add.ftz.f32 %f68, %f67, 0f3F800000; add.ftz.f32 %f69, %f10, %f68; add.ftz.f32 %f70, %f69, 0f3F000000; div.approx.ftz.f32 %f71, %f70, %f4; div.approx.ftz.f32 %f72, %f39, %f40; add.ftz.f32 %f73, %f72, 0fBF800000; add.ftz.f32 %f74, %f11, %f73; add.ftz.f32 %f75, %f74, 0f3F000000; div.approx.ftz.f32 %f76, %f75, %f6; div.approx.ftz.f32 %f77, %f45, %f46; add.ftz.f32 %f78, %f77, 0f3F800000; add.ftz.f32 %f79, %f11, %f78; add.ftz.f32 %f80, %f79, 0f3F000000; div.approx.ftz.f32 %f81, %f80, %f6; div.approx.ftz.f32 %f82, %f54, %f55; add.ftz.f32 %f83, %f82, 0fBF800000; add.ftz.f32 %f84, %f12, %f83; add.ftz.f32 %f85, %f84, 0f3F000000; div.approx.ftz.f32 %f86, %f85, %f8; div.approx.ftz.f32 %f87, %f60, %f61; add.ftz.f32 %f88, %f87, 0f3F800000; add.ftz.f32 %f89, %f12, %f88; add.ftz.f32 %f90, %f89, 0f3F000000; div.approx.ftz.f32 %f91, %f90, %f8; tex.3d.v4.f32.f32 {%f92, %f93, %f94, %f95}, [%rd3, {%f66, %f76, %f86, %f86}]; tex.3d.v4.f32.f32 {%f96, %f97, %f98, %f99}, [%rd3, {%f71, %f76, %f86, %f86}]; mul.ftz.f32 %f100, %f31, %f96; mul.ftz.f32 %f101, %f31, %f97; mul.ftz.f32 %f102, %f31, %f98; mul.ftz.f32 %f103, %f31, %f99; fma.rn.ftz.f32 %f104, %f25, %f92, %f100; fma.rn.ftz.f32 %f105, %f25, %f93, %f101; fma.rn.ftz.f32 %f106, %f25, %f94, %f102; fma.rn.ftz.f32 %f107, %f25, %f95, %f103; tex.3d.v4.f32.f32 {%f108, %f109, %f110, %f111}, [%rd3, {%f66, %f81, %f86, %f86}]; tex.3d.v4.f32.f32 {%f112, %f113, %f114, %f115}, [%rd3, {%f71, %f81, %f86, %f86}]; mul.ftz.f32 %f116, %f31, %f112; mul.ftz.f32 %f117, %f31, %f113; mul.ftz.f32 %f118, %f31, %f114; mul.ftz.f32 %f119, %f31, %f115; fma.rn.ftz.f32 %f120, %f25, %f108, %f116; fma.rn.ftz.f32 %f121, %f25, %f109, %f117; fma.rn.ftz.f32 %f122, %f25, %f110, %f118; fma.rn.ftz.f32 %f123, %f25, %f111, %f119; mul.ftz.f32 %f124, %f46, %f120; mul.ftz.f32 %f125, %f46, %f121; mul.ftz.f32 %f126, %f46, %f122; mul.ftz.f32 %f127, %f46, %f123; fma.rn.ftz.f32 %f128, %f40, %f104, %f124; fma.rn.ftz.f32 %f129, %f40, %f105, %f125; fma.rn.ftz.f32 %f130, %f40, %f106, %f126; fma.rn.ftz.f32 %f131, %f40, %f107, %f127; tex.3d.v4.f32.f32 {%f132, %f133, %f134, %f135}, [%rd3, {%f66, %f76, %f91, %f91}]; tex.3d.v4.f32.f32 {%f136, %f137, %f138, %f139}, [%rd3, {%f71, %f76, %f91, %f91}]; mul.ftz.f32 %f140, %f31, %f136; mul.ftz.f32 %f141, %f31, %f137; mul.ftz.f32 %f142, %f31, %f138; mul.ftz.f32 %f143, %f31, %f139; fma.rn.ftz.f32 %f144, %f25, %f132, %f140; fma.rn.ftz.f32 %f145, %f25, %f133, %f141; fma.rn.ftz.f32 %f146, %f25, %f134, %f142; fma.rn.ftz.f32 %f147, %f25, %f135, %f143; tex.3d.v4.f32.f32 {%f148, %f149, %f150, %f151}, [%rd3, {%f66, %f81, %f91, %f91}]; tex.3d.v4.f32.f32 {%f152, %f153, %f154, %f155}, [%rd3, {%f71, %f81, %f91, %f91}]; mul.ftz.f32 %f156, %f31, %f152; mul.ftz.f32 %f157, %f31, %f153; mul.ftz.f32 %f158, %f31, %f154; mul.ftz.f32 %f159, %f31, %f155; fma.rn.ftz.f32 %f160, %f25, %f148, %f156; fma.rn.ftz.f32 %f161, %f25, %f149, %f157; fma.rn.ftz.f32 %f162, %f25, %f150, %f158; fma.rn.ftz.f32 %f163, %f25, %f151, %f159; mul.ftz.f32 %f164, %f46, %f160; mul.ftz.f32 %f165, %f46, %f161; mul.ftz.f32 %f166, %f46, %f162; mul.ftz.f32 %f167, %f46, %f163; fma.rn.ftz.f32 %f168, %f40, %f144, %f164; fma.rn.ftz.f32 %f169, %f40, %f145, %f165; fma.rn.ftz.f32 %f170, %f40, %f146, %f166; fma.rn.ftz.f32 %f171, %f40, %f147, %f167; mul.ftz.f32 %f172, %f61, %f168; mul.ftz.f32 %f173, %f61, %f169; mul.ftz.f32 %f174, %f61, %f170; mul.ftz.f32 %f175, %f61, %f171; fma.rn.ftz.f32 %f176, %f55, %f131, %f175; fma.rn.ftz.f32 %f177, %f55, %f130, %f174; fma.rn.ftz.f32 %f178, %f55, %f129, %f173; fma.rn.ftz.f32 %f179, %f55, %f128, %f172; st.param.f32 [func_retval0+0], %f179; st.param.f32 [func_retval0+4], %f178; st.param.f32 [func_retval0+8], %f177; st.param.f32 [func_retval0+12], %f176; ret; } .func (.param .b32 func_retval0) _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff( .param .b64 _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_0, .param .b32 _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_1, .param .b32 _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_2, .param .b32 _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_3 ) { .reg .f32 %f<138>; .reg .b32 %r<6>; .reg .b64 %rd<4>; ld.param.u64 %rd1, [_Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_0]; ld.param.f32 %f1, [_Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_1]; ld.param.f32 %f2, [_Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_2]; ld.param.f32 %f3, [_Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff_param_3]; cvta.to.global.u64 %rd2, %rd1; ld.global.u64 %rd3, [%rd2]; ld.global.u32 %r1, [%rd2+20]; cvt.rn.f32.u32 %f4, %r1; fma.rn.ftz.f32 %f5, %f4, %f1, 0fBF000000; ld.global.v2.u32 {%r2, %r3}, [%rd2+24]; cvt.rn.f32.u32 %f6, %r2; fma.rn.ftz.f32 %f7, %f6, %f2, 0fBF000000; cvt.rn.f32.u32 %f8, %r3; fma.rn.ftz.f32 %f9, %f8, %f3, 0fBF000000; cvt.rmi.ftz.f32.f32 %f10, %f5; cvt.rmi.ftz.f32.f32 %f11, %f7; cvt.rmi.ftz.f32.f32 %f12, %f9; sub.ftz.f32 %f13, %f5, %f10; sub.ftz.f32 %f14, %f7, %f11; sub.ftz.f32 %f15, %f9, %f12; mov.f32 %f16, 0f40400000; sub.ftz.f32 %f17, %f16, %f13; fma.rn.ftz.f32 %f18, %f13, %f17, 0fC0400000; fma.rn.ftz.f32 %f19, %f13, %f18, 0f3F800000; mul.ftz.f32 %f20, %f13, %f13; mul.ftz.f32 %f21, %f13, 0f40400000; add.ftz.f32 %f22, %f21, 0fC0C00000; fma.rn.ftz.f32 %f23, %f20, %f22, 0f40800000; mul.ftz.f32 %f24, %f23, 0f3E2AAAAB; fma.rn.ftz.f32 %f25, %f19, 0f3E2AAAAB, %f24; sub.ftz.f32 %f26, %f16, %f21; fma.rn.ftz.f32 %f27, %f13, %f26, 0f40400000; fma.rn.ftz.f32 %f28, %f13, %f27, 0f3F800000; mul.ftz.f32 %f29, %f13, %f20; mul.ftz.f32 %f30, %f29, 0f3E2AAAAB; fma.rn.ftz.f32 %f31, %f28, 0f3E2AAAAB, %f30; sub.ftz.f32 %f32, %f16, %f14; fma.rn.ftz.f32 %f33, %f14, %f32, 0fC0400000; fma.rn.ftz.f32 %f34, %f14, %f33, 0f3F800000; mul.ftz.f32 %f35, %f14, %f14; mul.ftz.f32 %f36, %f14, 0f40400000; add.ftz.f32 %f37, %f36, 0fC0C00000; fma.rn.ftz.f32 %f38, %f35, %f37, 0f40800000; mul.ftz.f32 %f39, %f38, 0f3E2AAAAB; fma.rn.ftz.f32 %f40, %f34, 0f3E2AAAAB, %f39; sub.ftz.f32 %f41, %f16, %f36; fma.rn.ftz.f32 %f42, %f14, %f41, 0f40400000; fma.rn.ftz.f32 %f43, %f14, %f42, 0f3F800000; mul.ftz.f32 %f44, %f14, %f35; mul.ftz.f32 %f45, %f44, 0f3E2AAAAB; fma.rn.ftz.f32 %f46, %f43, 0f3E2AAAAB, %f45; sub.ftz.f32 %f47, %f16, %f15; fma.rn.ftz.f32 %f48, %f15, %f47, 0fC0400000; fma.rn.ftz.f32 %f49, %f15, %f48, 0f3F800000; mul.ftz.f32 %f50, %f15, %f15; mul.ftz.f32 %f51, %f15, 0f40400000; add.ftz.f32 %f52, %f51, 0fC0C00000; fma.rn.ftz.f32 %f53, %f50, %f52, 0f40800000; mul.ftz.f32 %f54, %f53, 0f3E2AAAAB; fma.rn.ftz.f32 %f55, %f49, 0f3E2AAAAB, %f54; sub.ftz.f32 %f56, %f16, %f51; fma.rn.ftz.f32 %f57, %f15, %f56, 0f40400000; fma.rn.ftz.f32 %f58, %f15, %f57, 0f3F800000; mul.ftz.f32 %f59, %f15, %f50; mul.ftz.f32 %f60, %f59, 0f3E2AAAAB; fma.rn.ftz.f32 %f61, %f58, 0f3E2AAAAB, %f60; div.approx.ftz.f32 %f62, %f24, %f25; add.ftz.f32 %f63, %f62, 0fBF800000; add.ftz.f32 %f64, %f10, %f63; add.ftz.f32 %f65, %f64, 0f3F000000; div.approx.ftz.f32 %f66, %f65, %f4; div.approx.ftz.f32 %f67, %f30, %f31; add.ftz.f32 %f68, %f67, 0f3F800000; add.ftz.f32 %f69, %f10, %f68; add.ftz.f32 %f70, %f69, 0f3F000000; div.approx.ftz.f32 %f71, %f70, %f4; div.approx.ftz.f32 %f72, %f39, %f40; add.ftz.f32 %f73, %f72, 0fBF800000; add.ftz.f32 %f74, %f11, %f73; add.ftz.f32 %f75, %f74, 0f3F000000; div.approx.ftz.f32 %f76, %f75, %f6; div.approx.ftz.f32 %f77, %f45, %f46; add.ftz.f32 %f78, %f77, 0f3F800000; add.ftz.f32 %f79, %f11, %f78; add.ftz.f32 %f80, %f79, 0f3F000000; div.approx.ftz.f32 %f81, %f80, %f6; div.approx.ftz.f32 %f82, %f54, %f55; add.ftz.f32 %f83, %f82, 0fBF800000; add.ftz.f32 %f84, %f12, %f83; add.ftz.f32 %f85, %f84, 0f3F000000; div.approx.ftz.f32 %f86, %f85, %f8; div.approx.ftz.f32 %f87, %f60, %f61; add.ftz.f32 %f88, %f87, 0f3F800000; add.ftz.f32 %f89, %f12, %f88; add.ftz.f32 %f90, %f89, 0f3F000000; div.approx.ftz.f32 %f91, %f90, %f8; tex.3d.v4.f32.f32 {%f92, %f93, %f94, %f95}, [%rd3, {%f66, %f76, %f86, %f86}]; tex.3d.v4.f32.f32 {%f96, %f97, %f98, %f99}, [%rd3, {%f71, %f76, %f86, %f86}]; mul.ftz.f32 %f100, %f31, %f96; fma.rn.ftz.f32 %f101, %f25, %f92, %f100; tex.3d.v4.f32.f32 {%f102, %f103, %f104, %f105}, [%rd3, {%f66, %f81, %f86, %f86}]; tex.3d.v4.f32.f32 {%f106, %f107, %f108, %f109}, [%rd3, {%f71, %f81, %f86, %f86}]; mul.ftz.f32 %f110, %f31, %f106; fma.rn.ftz.f32 %f111, %f25, %f102, %f110; mul.ftz.f32 %f112, %f46, %f111; fma.rn.ftz.f32 %f113, %f40, %f101, %f112; tex.3d.v4.f32.f32 {%f114, %f115, %f116, %f117}, [%rd3, {%f66, %f76, %f91, %f91}]; tex.3d.v4.f32.f32 {%f118, %f119, %f120, %f121}, [%rd3, {%f71, %f76, %f91, %f91}]; mul.ftz.f32 %f122, %f31, %f118; fma.rn.ftz.f32 %f123, %f25, %f114, %f122; tex.3d.v4.f32.f32 {%f124, %f125, %f126, %f127}, [%rd3, {%f66, %f81, %f91, %f91}]; tex.3d.v4.f32.f32 {%f128, %f129, %f130, %f131}, [%rd3, {%f71, %f81, %f91, %f91}]; mul.ftz.f32 %f132, %f31, %f128; fma.rn.ftz.f32 %f133, %f25, %f124, %f132; mul.ftz.f32 %f134, %f46, %f133; fma.rn.ftz.f32 %f135, %f40, %f123, %f134; mul.ftz.f32 %f136, %f61, %f135; fma.rn.ftz.f32 %f137, %f55, %f113, %f136; st.param.f32 [func_retval0+0], %f137; ret; } .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) { .reg .pred %p<174>; .reg .b16 %rs<106>; .reg .f32 %f<128>; .reg .b32 %r<881>; .reg .b64 %rd<708>; ld.param.u64 %rd160, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0]; cvta.to.local.u64 %rd1, %rd160; ld.param.f32 %f103, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+8]; ld.param.f32 %f104, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+4]; ld.param.f32 %f105, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorIfLi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1]; cvt.rmi.ftz.f32.f32 %f106, %f105; cvt.rmi.ftz.f32.f32 %f107, %f104; cvt.rmi.ftz.f32.f32 %f108, %f103; sub.ftz.f32 %f1, %f105, %f106; sub.ftz.f32 %f2, %f104, %f107; sub.ftz.f32 %f3, %f103, %f108; cvt.rzi.ftz.s32.f32 %r1, %f106; cvt.rzi.ftz.s32.f32 %r2, %f107; cvt.rzi.ftz.s32.f32 %r3, %f108; and.b32 %r4, %r1, -8; ld.local.u64 %rd2, [%rd1]; ld.u32 %r49, [%rd2]; setp.ne.s32 %p1, %r4, %r49; mov.u16 %rs85, 0; @%p1 bra $L__BB9_3; and.b32 %r50, %r2, -8; ld.u32 %r51, [%rd2+4]; setp.ne.s32 %p2, %r50, %r51; @%p2 bra $L__BB9_3; and.b32 %r52, %r3, -8; ld.u32 %r53, [%rd2+8]; setp.eq.s32 %p3, %r52, %r53; selp.u16 %rs85, 1, 0, %p3; $L__BB9_3: setp.eq.s16 %p4, %rs85, 0; @%p4 bra $L__BB9_5; shl.b32 %r54, %r1, 6; and.b32 %r55, %r54, 448; shl.b32 %r56, %r2, 3; and.b32 %r57, %r56, 56; or.b32 %r58, %r57, %r55; and.b32 %r59, %r3, 7; or.b32 %r60, %r58, %r59; ld.u64 %rd161, [%rd2+48]; mul.wide.u32 %rd162, %r60, 4; add.s64 %rd163, %rd161, %rd162; ld.f32 %f120, [%rd163+96]; $L__BB9_35: add.s32 %r10, %r3, 1; ld.local.u64 %rd22, [%rd1]; ld.u32 %r153, [%rd22]; setp.ne.s32 %p23, %r4, %r153; mov.u16 %rs88, 0; @%p23 bra $L__BB9_38; and.b32 %r154, %r2, -8; ld.u32 %r155, [%rd22+4]; setp.ne.s32 %p24, %r154, %r155; @%p24 bra $L__BB9_38; and.b32 %r156, %r10, -8; ld.u32 %r157, [%rd22+8]; setp.eq.s32 %p25, %r156, %r157; selp.u16 %rs88, 1, 0, %p25; $L__BB9_38: setp.eq.s16 %p26, %rs88, 0; @%p26 bra $L__BB9_40; shl.b32 %r158, %r1, 6; and.b32 %r159, %r158, 448; shl.b32 %r160, %r2, 3; and.b32 %r161, %r160, 56; or.b32 %r162, %r161, %r159; and.b32 %r163, %r10, 7; or.b32 %r164, %r162, %r163; ld.u64 %rd228, [%rd22+48]; mul.wide.u32 %rd229, %r164, 4; add.s64 %rd230, %rd228, %rd229; ld.f32 %f121, [%rd230+96]; $L__BB9_70: sub.ftz.f32 %f109, %f121, %f120; fma.rn.ftz.f32 %f28, %f3, %f109, %f120; add.s32 %r16, %r2, 1; ld.local.u64 %rd42, [%rd1]; ld.u32 %r257, [%rd42]; setp.ne.s32 %p45, %r4, %r257; mov.u16 %rs91, 0; @%p45 bra $L__BB9_73; and.b32 %r258, %r16, -8; ld.u32 %r259, [%rd42+4]; setp.ne.s32 %p46, %r258, %r259; @%p46 bra $L__BB9_73; and.b32 %r260, %r10, -8; ld.u32 %r261, [%rd42+8]; setp.eq.s32 %p47, %r260, %r261; selp.u16 %rs91, 1, 0, %p47; $L__BB9_73: setp.eq.s16 %p48, %rs91, 0; @%p48 bra $L__BB9_75; shl.b32 %r262, %r1, 6; and.b32 %r263, %r262, 448; shl.b32 %r264, %r16, 3; and.b32 %r265, %r264, 56; or.b32 %r266, %r265, %r263; and.b32 %r267, %r10, 7; or.b32 %r268, %r266, %r267; ld.u64 %rd296, [%rd42+48]; mul.wide.u32 %rd297, %r268, 4; add.s64 %rd298, %rd296, %rd297; ld.f32 %f122, [%rd298+96]; $L__BB9_105: ld.local.u64 %rd62, [%rd1]; ld.u32 %r361, [%rd62]; setp.ne.s32 %p67, %r4, %r361; mov.u16 %rs94, 0; @%p67 bra $L__BB9_108; and.b32 %r362, %r16, -8; ld.u32 %r363, [%rd62+4]; setp.ne.s32 %p68, %r362, %r363; @%p68 bra $L__BB9_108; and.b32 %r364, %r3, -8; ld.u32 %r365, [%rd62+8]; setp.eq.s32 %p69, %r364, %r365; selp.u16 %rs94, 1, 0, %p69; $L__BB9_108: setp.eq.s16 %p70, %rs94, 0; @%p70 bra $L__BB9_110; and.b32 %r366, %r1, 7; shl.b32 %r367, %r16, 3; and.b32 %r368, %r367, 56; and.b32 %r369, %r3, 7; bfi.b32 %r370, %r366, %r369, 6, 3; or.b32 %r371, %r370, %r368; ld.u64 %rd363, [%rd62+48]; mul.wide.u32 %rd364, %r371, 4; add.s64 %rd365, %rd363, %rd364; ld.f32 %f123, [%rd365+96]; $L__BB9_140: sub.ftz.f32 %f110, %f122, %f123; fma.rn.ftz.f32 %f111, %f3, %f110, %f123; sub.ftz.f32 %f112, %f111, %f28; fma.rn.ftz.f32 %f53, %f2, %f112, %f28; add.s32 %r27, %r1, 1; and.b32 %r28, %r27, -8; ld.local.u64 %rd81, [%rd1]; ld.u32 %r461, [%rd81]; setp.ne.s32 %p89, %r28, %r461; mov.u16 %rs97, 0; @%p89 bra $L__BB9_143; and.b32 %r462, %r16, -8; ld.u32 %r463, [%rd81+4]; setp.ne.s32 %p90, %r462, %r463; @%p90 bra $L__BB9_143; and.b32 %r464, %r3, -8; ld.u32 %r465, [%rd81+8]; setp.eq.s32 %p91, %r464, %r465; selp.u16 %rs97, 1, 0, %p91; $L__BB9_143: setp.eq.s16 %p92, %rs97, 0; @%p92 bra $L__BB9_145; and.b32 %r466, %r27, 7; shl.b32 %r467, %r16, 3; and.b32 %r468, %r467, 56; and.b32 %r469, %r3, 7; bfi.b32 %r470, %r466, %r469, 6, 3; or.b32 %r471, %r470, %r468; ld.u64 %rd430, [%rd81+48]; mul.wide.u32 %rd431, %r471, 4; add.s64 %rd432, %rd430, %rd431; ld.f32 %f124, [%rd432+96]; $L__BB9_175: ld.local.u64 %rd101, [%rd1]; ld.u32 %r561, [%rd101]; setp.ne.s32 %p111, %r28, %r561; mov.u16 %rs100, 0; @%p111 bra $L__BB9_178; and.b32 %r562, %r16, -8; ld.u32 %r563, [%rd101+4]; setp.ne.s32 %p112, %r562, %r563; @%p112 bra $L__BB9_178; and.b32 %r564, %r10, -8; ld.u32 %r565, [%rd101+8]; setp.eq.s32 %p113, %r564, %r565; selp.u16 %rs100, 1, 0, %p113; $L__BB9_178: setp.eq.s16 %p114, %rs100, 0; @%p114 bra $L__BB9_180; shl.b32 %r566, %r27, 6; and.b32 %r567, %r566, 448; shl.b32 %r568, %r16, 3; and.b32 %r569, %r568, 56; or.b32 %r570, %r569, %r567; and.b32 %r571, %r10, 7; or.b32 %r572, %r570, %r571; ld.u64 %rd497, [%rd101+48]; mul.wide.u32 %rd498, %r572, 4; add.s64 %rd499, %rd497, %rd498; ld.f32 %f125, [%rd499+96]; $L__BB9_210: sub.ftz.f32 %f113, %f125, %f124; fma.rn.ftz.f32 %f78, %f3, %f113, %f124; ld.local.u64 %rd120, [%rd1]; ld.u32 %r665, [%rd120]; setp.ne.s32 %p133, %r28, %r665; mov.u16 %rs103, 0; @%p133 bra $L__BB9_213; and.b32 %r666, %r2, -8; ld.u32 %r667, [%rd120+4]; setp.ne.s32 %p134, %r666, %r667; @%p134 bra $L__BB9_213; and.b32 %r668, %r10, -8; ld.u32 %r669, [%rd120+8]; setp.eq.s32 %p135, %r668, %r669; selp.u16 %rs103, 1, 0, %p135; $L__BB9_213: setp.eq.s16 %p136, %rs103, 0; @%p136 bra $L__BB9_215; shl.b32 %r670, %r27, 6; and.b32 %r671, %r670, 448; shl.b32 %r672, %r2, 3; and.b32 %r673, %r672, 56; or.b32 %r674, %r671, %r673; and.b32 %r675, %r10, 7; or.b32 %r676, %r674, %r675; ld.u64 %rd564, [%rd120+48]; mul.wide.u32 %rd565, %r676, 4; add.s64 %rd566, %rd564, %rd565; ld.f32 %f126, [%rd566+96]; $L__BB9_245: ld.local.u64 %rd140, [%rd1]; ld.u32 %r769, [%rd140]; setp.ne.s32 %p155, %r28, %r769; @%p155 bra $L__BB9_248; and.b32 %r770, %r2, -8; ld.u32 %r771, [%rd140+4]; setp.ne.s32 %p156, %r770, %r771; @%p156 bra $L__BB9_248; and.b32 %r772, %r3, -8; ld.u32 %r773, [%rd140+8]; setp.eq.s32 %p157, %r772, %r773; @%p157 bra $L__BB9_276; bra.uni $L__BB9_248; $L__BB9_276: shl.b32 %r866, %r27, 6; and.b32 %r867, %r866, 448; shl.b32 %r868, %r2, 3; and.b32 %r869, %r868, 56; or.b32 %r870, %r867, %r869; and.b32 %r871, %r3, 7; or.b32 %r872, %r870, %r871; ld.u64 %rd697, [%rd140+48]; mul.wide.u32 %rd698, %r872, 4; add.s64 %rd699, %rd697, %rd698; ld.f32 %f127, [%rd699+96]; bra.uni $L__BB9_277; $L__BB9_248: and.b32 %r44, %r27, -128; ld.u32 %r774, [%rd140+12]; setp.ne.s32 %p158, %r44, %r774; @%p158 bra $L__BB9_251; and.b32 %r775, %r2, -128; ld.u32 %r776, [%rd140+16]; setp.ne.s32 %p159, %r775, %r776; @%p159 bra $L__BB9_251; and.b32 %r777, %r3, -128; ld.u32 %r778, [%rd140+20]; setp.eq.s32 %p160, %r777, %r778; @%p160 bra $L__BB9_273; bra.uni $L__BB9_251; $L__BB9_273: shl.b32 %r848, %r27, 5; and.b32 %r849, %r848, 3840; shl.b32 %r850, %r2, 1; and.b32 %r851, %r850, 240; or.b32 %r852, %r849, %r851; bfe.u32 %r853, %r3, 3, 4; or.b32 %r854, %r852, %r853; shr.u32 %r855, %r852, 6; ld.u64 %rd158, [%rd140+56]; mul.wide.u32 %rd685, %r855, 8; add.s64 %rd686, %rd158, %rd685; and.b32 %r856, %r854, 63; mov.u64 %rd687, 1; shl.b64 %rd688, %rd687, %r856; ld.u64 %rd689, [%rd686+544]; and.b64 %rd690, %rd689, %rd688; setp.eq.s64 %p173, %rd690, 0; mul.wide.u32 %rd691, %r854, 8; add.s64 %rd692, %rd158, %rd691; add.s64 %rd159, %rd692, 1088; @%p173 bra $L__BB9_275; ld.u64 %rd693, [%rd159]; add.s64 %rd694, %rd158, %rd693; and.b32 %r857, %r2, -8; st.v2.u32 [%rd140], {%r28, %r857}; and.b32 %r858, %r3, -8; st.u32 [%rd140+8], %r858; st.u64 [%rd140+48], %rd694; shl.b32 %r859, %r27, 6; and.b32 %r860, %r859, 448; shl.b32 %r861, %r2, 3; and.b32 %r862, %r861, 56; or.b32 %r863, %r860, %r862; and.b32 %r864, %r3, 7; or.b32 %r865, %r863, %r864; mul.wide.u32 %rd695, %r865, 4; add.s64 %rd696, %rd694, %rd695; ld.f32 %f127, [%rd696+96]; bra.uni $L__BB9_277; $L__BB9_251: and.b32 %r45, %r27, -4096; ld.u32 %r779, [%rd140+24]; setp.ne.s32 %p161, %r45, %r779; @%p161 bra $L__BB9_254; and.b32 %r780, %r2, -4096; ld.u32 %r781, [%rd140+28]; setp.ne.s32 %p162, %r780, %r781; @%p162 bra $L__BB9_254; and.b32 %r782, %r3, -4096; ld.u32 %r783, [%rd140+32]; setp.eq.s32 %p163, %r782, %r783; @%p163 bra $L__BB9_268; bra.uni $L__BB9_254; $L__BB9_268: shl.b32 %r819, %r27, 3; and.b32 %r820, %r819, 31744; and.b32 %r821, %r2, 3968; shr.u32 %r822, %r821, 2; or.b32 %r823, %r820, %r822; bfe.u32 %r824, %r3, 7, 5; or.b32 %r825, %r823, %r824; shr.u32 %r826, %r823, 6; ld.u64 %rd154, [%rd140+64]; mul.wide.u32 %rd664, %r826, 8; add.s64 %rd665, %rd154, %rd664; and.b32 %r827, %r825, 63; mov.u64 %rd666, 1; shl.b64 %rd667, %rd666, %r827; ld.u64 %rd668, [%rd665+4128]; and.b64 %rd669, %rd668, %rd667; setp.eq.s64 %p171, %rd669, 0; mul.wide.u32 %rd670, %r825, 8; add.s64 %rd671, %rd154, %rd670; add.s64 %rd155, %rd671, 8256; @%p171 bra $L__BB9_272; ld.u64 %rd672, [%rd155]; add.s64 %rd156, %rd154, %rd672; st.u32 [%rd140+12], %r44; and.b32 %r828, %r2, -128; st.u32 [%rd140+16], %r828; and.b32 %r829, %r3, -128; st.u32 [%rd140+20], %r829; st.u64 [%rd140+56], %rd156; shl.b32 %r830, %r27, 5; and.b32 %r831, %r830, 3840; shl.b32 %r832, %r2, 1; and.b32 %r833, %r832, 240; or.b32 %r834, %r831, %r833; bfe.u32 %r835, %r3, 3, 4; or.b32 %r836, %r834, %r835; shr.u32 %r837, %r834, 6; mul.wide.u32 %rd673, %r837, 8; add.s64 %rd674, %rd156, %rd673; and.b32 %r838, %r836, 63; shl.b64 %rd676, %rd666, %r838; ld.u64 %rd677, [%rd674+544]; and.b64 %rd678, %rd677, %rd676; setp.eq.s64 %p172, %rd678, 0; mul.wide.u32 %rd679, %r836, 8; add.s64 %rd680, %rd156, %rd679; add.s64 %rd157, %rd680, 1088; @%p172 bra $L__BB9_271; ld.u64 %rd681, [%rd157]; add.s64 %rd682, %rd156, %rd681; and.b32 %r839, %r2, -8; st.v2.u32 [%rd140], {%r28, %r839}; and.b32 %r840, %r3, -8; st.u32 [%rd140+8], %r840; st.u64 [%rd140+48], %rd682; shl.b32 %r841, %r27, 6; and.b32 %r842, %r841, 448; shl.b32 %r843, %r2, 3; and.b32 %r844, %r843, 56; or.b32 %r845, %r842, %r844; and.b32 %r846, %r3, 7; or.b32 %r847, %r845, %r846; mul.wide.u32 %rd683, %r847, 4; add.s64 %rd684, %rd682, %rd683; ld.f32 %f127, [%rd684+96]; bra.uni $L__BB9_277; $L__BB9_254: ld.u64 %rd141, [%rd140+40]; mov.u64 %rd707, 0; add.s64 %rd142, %rd141, 24; ld.u32 %r46, [%rd141+24]; setp.eq.s32 %p164, %r46, 0; @%p164 bra $L__BB9_259; add.s64 %rd143, %rd141, 64; shr.u32 %r785, %r27, 12; cvt.u64.u32 %rd633, %r785; shl.b64 %rd634, %rd633, 42; shr.u32 %r786, %r2, 12; mul.wide.u32 %rd635, %r786, 2097152; shr.u32 %r787, %r3, 12; cvt.u64.u32 %rd636, %r787; or.b64 %rd637, %rd635, %rd636; or.b64 %rd144, %rd637, %rd634; mov.u32 %r880, 0; $L__BB9_256: cvt.u64.u32 %rd145, %r880; mul.wide.u32 %rd638, %r880, 32; add.s64 %rd639, %rd143, %rd638; ld.u64 %rd640, [%rd639]; setp.eq.s64 %p165, %rd640, %rd144; add.s32 %r880, %r880, 1; @%p165 bra $L__BB9_258; setp.lt.u32 %p166, %r880, %r46; @%p166 bra $L__BB9_256; bra.uni $L__BB9_259; $L__BB9_5: and.b32 %r5, %r1, -128; ld.u32 %r61, [%rd2+12]; setp.ne.s32 %p5, %r5, %r61; mov.u16 %rs86, 0; @%p5 bra $L__BB9_8; and.b32 %r62, %r2, -128; ld.u32 %r63, [%rd2+16]; setp.ne.s32 %p6, %r62, %r63; @%p6 bra $L__BB9_8; and.b32 %r64, %r3, -128; ld.u32 %r65, [%rd2+20]; setp.eq.s32 %p7, %r64, %r65; selp.u16 %rs86, 1, 0, %p7; $L__BB9_8: setp.eq.s16 %p8, %rs86, 0; @%p8 bra $L__BB9_12; shl.b32 %r66, %r1, 5; and.b32 %r67, %r66, 3840; shl.b32 %r68, %r2, 1; and.b32 %r69, %r68, 240; or.b32 %r70, %r69, %r67; bfe.u32 %r71, %r3, 3, 4; or.b32 %r72, %r70, %r71; shr.u32 %r73, %r70, 6; ld.u64 %rd3, [%rd2+56]; mul.wide.u32 %rd164, %r73, 8; add.s64 %rd165, %rd3, %rd164; and.b32 %r74, %r72, 63; mov.u64 %rd166, 1; shl.b64 %rd167, %rd166, %r74; ld.u64 %rd168, [%rd165+544]; and.b64 %rd169, %rd168, %rd167; setp.eq.s64 %p9, %rd169, 0; mul.wide.u32 %rd170, %r72, 8; add.s64 %rd171, %rd3, %rd170; add.s64 %rd4, %rd171, 1088; @%p9 bra $L__BB9_11; ld.u64 %rd172, [%rd4]; add.s64 %rd173, %rd3, %rd172; and.b32 %r75, %r2, -8; st.v2.u32 [%rd2], {%r4, %r75}; and.b32 %r76, %r3, -8; st.u32 [%rd2+8], %r76; st.u64 [%rd2+48], %rd173; shl.b32 %r77, %r1, 6; and.b32 %r78, %r77, 448; shl.b32 %r79, %r2, 3; and.b32 %r80, %r79, 56; or.b32 %r81, %r80, %r78; and.b32 %r82, %r3, 7; or.b32 %r83, %r81, %r82; mul.wide.u32 %rd174, %r83, 4; add.s64 %rd175, %rd173, %rd174; ld.f32 %f120, [%rd175+96]; bra.uni $L__BB9_35; $L__BB9_40: and.b32 %r11, %r1, -128; ld.u32 %r165, [%rd22+12]; setp.ne.s32 %p27, %r11, %r165; mov.u16 %rs89, 0; @%p27 bra $L__BB9_43; and.b32 %r166, %r2, -128; ld.u32 %r167, [%rd22+16]; setp.ne.s32 %p28, %r166, %r167; @%p28 bra $L__BB9_43; and.b32 %r168, %r10, -128; ld.u32 %r169, [%rd22+20]; setp.eq.s32 %p29, %r168, %r169; selp.u16 %rs89, 1, 0, %p29; $L__BB9_43: setp.eq.s16 %p30, %rs89, 0; @%p30 bra $L__BB9_47; shl.b32 %r170, %r1, 5; and.b32 %r171, %r170, 3840; shl.b32 %r172, %r2, 1; and.b32 %r173, %r172, 240; or.b32 %r174, %r173, %r171; bfe.u32 %r175, %r10, 3, 4; or.b32 %r176, %r175, %r174; shr.u32 %r177, %r174, 6; ld.u64 %rd23, [%rd22+56]; mul.wide.u32 %rd231, %r177, 8; add.s64 %rd232, %rd23, %rd231; and.b32 %r178, %r176, 63; mov.u64 %rd233, 1; shl.b64 %rd234, %rd233, %r178; ld.u64 %rd235, [%rd232+544]; and.b64 %rd236, %rd235, %rd234; setp.eq.s64 %p31, %rd236, 0; mul.wide.u32 %rd237, %r176, 8; add.s64 %rd238, %rd23, %rd237; add.s64 %rd24, %rd238, 1088; @%p31 bra $L__BB9_46; ld.u64 %rd239, [%rd24]; add.s64 %rd240, %rd23, %rd239; and.b32 %r179, %r2, -8; st.v2.u32 [%rd22], {%r4, %r179}; and.b32 %r180, %r10, -8; st.u32 [%rd22+8], %r180; st.u64 [%rd22+48], %rd240; shl.b32 %r181, %r1, 6; and.b32 %r182, %r181, 448; shl.b32 %r183, %r2, 3; and.b32 %r184, %r183, 56; or.b32 %r185, %r184, %r182; and.b32 %r186, %r10, 7; or.b32 %r187, %r185, %r186; mul.wide.u32 %rd241, %r187, 4; add.s64 %rd242, %rd240, %rd241; ld.f32 %f121, [%rd242+96]; bra.uni $L__BB9_70; $L__BB9_75: and.b32 %r17, %r1, -128; ld.u32 %r269, [%rd42+12]; setp.ne.s32 %p49, %r17, %r269; mov.u16 %rs92, 0; @%p49 bra $L__BB9_78; and.b32 %r270, %r16, -128; ld.u32 %r271, [%rd42+16]; setp.ne.s32 %p50, %r270, %r271; @%p50 bra $L__BB9_78; and.b32 %r272, %r10, -128; ld.u32 %r273, [%rd42+20]; setp.eq.s32 %p51, %r272, %r273; selp.u16 %rs92, 1, 0, %p51; $L__BB9_78: setp.eq.s16 %p52, %rs92, 0; @%p52 bra $L__BB9_82; shl.b32 %r274, %r1, 5; and.b32 %r275, %r274, 3840; shl.b32 %r276, %r16, 1; and.b32 %r277, %r276, 240; or.b32 %r278, %r277, %r275; bfe.u32 %r279, %r10, 3, 4; or.b32 %r280, %r278, %r279; shr.u32 %r281, %r278, 6; ld.u64 %rd43, [%rd42+56]; mul.wide.u32 %rd299, %r281, 8; add.s64 %rd300, %rd43, %rd299; and.b32 %r282, %r280, 63; mov.u64 %rd301, 1; shl.b64 %rd302, %rd301, %r282; ld.u64 %rd303, [%rd300+544]; and.b64 %rd304, %rd303, %rd302; setp.eq.s64 %p53, %rd304, 0; mul.wide.u32 %rd305, %r280, 8; add.s64 %rd306, %rd43, %rd305; add.s64 %rd44, %rd306, 1088; @%p53 bra $L__BB9_81; ld.u64 %rd307, [%rd44]; add.s64 %rd308, %rd43, %rd307; and.b32 %r283, %r16, -8; st.v2.u32 [%rd42], {%r4, %r283}; and.b32 %r284, %r10, -8; st.u32 [%rd42+8], %r284; st.u64 [%rd42+48], %rd308; shl.b32 %r285, %r1, 6; and.b32 %r286, %r285, 448; shl.b32 %r287, %r16, 3; and.b32 %r288, %r287, 56; or.b32 %r289, %r288, %r286; and.b32 %r290, %r10, 7; or.b32 %r291, %r289, %r290; mul.wide.u32 %rd309, %r291, 4; add.s64 %rd310, %rd308, %rd309; ld.f32 %f122, [%rd310+96]; bra.uni $L__BB9_105; $L__BB9_110: and.b32 %r22, %r1, -128; ld.u32 %r372, [%rd62+12]; setp.ne.s32 %p71, %r22, %r372; mov.u16 %rs95, 0; @%p71 bra $L__BB9_113; and.b32 %r373, %r16, -128; ld.u32 %r374, [%rd62+16]; setp.ne.s32 %p72, %r373, %r374; @%p72 bra $L__BB9_113; and.b32 %r375, %r3, -128; ld.u32 %r376, [%rd62+20]; setp.eq.s32 %p73, %r375, %r376; selp.u16 %rs95, 1, 0, %p73; $L__BB9_113: setp.eq.s16 %p74, %rs95, 0; @%p74 bra $L__BB9_117; shl.b32 %r377, %r1, 5; and.b32 %r378, %r377, 3840; shl.b32 %r379, %r16, 1; and.b32 %r380, %r379, 240; or.b32 %r381, %r380, %r378; bfe.u32 %r382, %r3, 3, 4; or.b32 %r383, %r381, %r382; shr.u32 %r384, %r381, 6; ld.u64 %rd63, [%rd62+56]; mul.wide.u32 %rd366, %r384, 8; add.s64 %rd367, %rd63, %rd366; and.b32 %r385, %r383, 63; mov.u64 %rd368, 1; shl.b64 %rd369, %rd368, %r385; ld.u64 %rd370, [%rd367+544]; and.b64 %rd371, %rd370, %rd369; setp.eq.s64 %p75, %rd371, 0; mul.wide.u32 %rd372, %r383, 8; add.s64 %rd373, %rd63, %rd372; add.s64 %rd64, %rd373, 1088; @%p75 bra $L__BB9_116; ld.u64 %rd374, [%rd64]; add.s64 %rd375, %rd63, %rd374; and.b32 %r386, %r16, -8; st.v2.u32 [%rd62], {%r4, %r386}; and.b32 %r387, %r3, -8; st.u32 [%rd62+8], %r387; st.u64 [%rd62+48], %rd375; and.b32 %r388, %r1, 7; shl.b32 %r389, %r16, 3; and.b32 %r390, %r389, 56; and.b32 %r391, %r3, 7; bfi.b32 %r392, %r388, %r391, 6, 3; or.b32 %r393, %r392, %r390; mul.wide.u32 %rd376, %r393, 4; add.s64 %rd377, %rd375, %rd376; ld.f32 %f123, [%rd377+96]; bra.uni $L__BB9_140; $L__BB9_145: and.b32 %r29, %r27, -128; ld.u32 %r472, [%rd81+12]; setp.ne.s32 %p93, %r29, %r472; mov.u16 %rs98, 0; @%p93 bra $L__BB9_148; and.b32 %r473, %r16, -128; ld.u32 %r474, [%rd81+16]; setp.ne.s32 %p94, %r473, %r474; @%p94 bra $L__BB9_148; and.b32 %r475, %r3, -128; ld.u32 %r476, [%rd81+20]; setp.eq.s32 %p95, %r475, %r476; selp.u16 %rs98, 1, 0, %p95; $L__BB9_148: setp.eq.s16 %p96, %rs98, 0; @%p96 bra $L__BB9_152; shl.b32 %r477, %r27, 5; and.b32 %r478, %r477, 3840; shl.b32 %r479, %r16, 1; and.b32 %r480, %r479, 240; or.b32 %r481, %r480, %r478; bfe.u32 %r482, %r3, 3, 4; or.b32 %r483, %r481, %r482; shr.u32 %r484, %r481, 6; ld.u64 %rd82, [%rd81+56]; mul.wide.u32 %rd433, %r484, 8; add.s64 %rd434, %rd82, %rd433; and.b32 %r485, %r483, 63; mov.u64 %rd435, 1; shl.b64 %rd436, %rd435, %r485; ld.u64 %rd437, [%rd434+544]; and.b64 %rd438, %rd437, %rd436; setp.eq.s64 %p97, %rd438, 0; mul.wide.u32 %rd439, %r483, 8; add.s64 %rd440, %rd82, %rd439; add.s64 %rd83, %rd440, 1088; @%p97 bra $L__BB9_151; ld.u64 %rd441, [%rd83]; add.s64 %rd442, %rd82, %rd441; and.b32 %r486, %r16, -8; st.v2.u32 [%rd81], {%r28, %r486}; and.b32 %r487, %r3, -8; st.u32 [%rd81+8], %r487; st.u64 [%rd81+48], %rd442; and.b32 %r488, %r27, 7; shl.b32 %r489, %r16, 3; and.b32 %r490, %r489, 56; and.b32 %r491, %r3, 7; bfi.b32 %r492, %r488, %r491, 6, 3; or.b32 %r493, %r492, %r490; mul.wide.u32 %rd443, %r493, 4; add.s64 %rd444, %rd442, %rd443; ld.f32 %f124, [%rd444+96]; bra.uni $L__BB9_175; $L__BB9_180: and.b32 %r34, %r27, -128; ld.u32 %r573, [%rd101+12]; setp.ne.s32 %p115, %r34, %r573; mov.u16 %rs101, 0; @%p115 bra $L__BB9_183; and.b32 %r574, %r16, -128; ld.u32 %r575, [%rd101+16]; setp.ne.s32 %p116, %r574, %r575; @%p116 bra $L__BB9_183; and.b32 %r576, %r10, -128; ld.u32 %r577, [%rd101+20]; setp.eq.s32 %p117, %r576, %r577; selp.u16 %rs101, 1, 0, %p117; $L__BB9_183: setp.eq.s16 %p118, %rs101, 0; @%p118 bra $L__BB9_187; shl.b32 %r578, %r27, 5; and.b32 %r579, %r578, 3840; shl.b32 %r580, %r16, 1; and.b32 %r581, %r580, 240; or.b32 %r582, %r581, %r579; bfe.u32 %r583, %r10, 3, 4; or.b32 %r584, %r582, %r583; shr.u32 %r585, %r582, 6; ld.u64 %rd102, [%rd101+56]; mul.wide.u32 %rd500, %r585, 8; add.s64 %rd501, %rd102, %rd500; and.b32 %r586, %r584, 63; mov.u64 %rd502, 1; shl.b64 %rd503, %rd502, %r586; ld.u64 %rd504, [%rd501+544]; and.b64 %rd505, %rd504, %rd503; setp.eq.s64 %p119, %rd505, 0; mul.wide.u32 %rd506, %r584, 8; add.s64 %rd507, %rd102, %rd506; add.s64 %rd103, %rd507, 1088; @%p119 bra $L__BB9_186; ld.u64 %rd508, [%rd103]; add.s64 %rd509, %rd102, %rd508; and.b32 %r587, %r16, -8; st.v2.u32 [%rd101], {%r28, %r587}; and.b32 %r588, %r10, -8; st.u32 [%rd101+8], %r588; st.u64 [%rd101+48], %rd509; shl.b32 %r589, %r27, 6; and.b32 %r590, %r589, 448; shl.b32 %r591, %r16, 3; and.b32 %r592, %r591, 56; or.b32 %r593, %r592, %r590; and.b32 %r594, %r10, 7; or.b32 %r595, %r593, %r594; mul.wide.u32 %rd510, %r595, 4; add.s64 %rd511, %rd509, %rd510; ld.f32 %f125, [%rd511+96]; bra.uni $L__BB9_210; $L__BB9_215: and.b32 %r39, %r27, -128; ld.u32 %r677, [%rd120+12]; setp.ne.s32 %p137, %r39, %r677; mov.u16 %rs104, 0; @%p137 bra $L__BB9_218; and.b32 %r678, %r2, -128; ld.u32 %r679, [%rd120+16]; setp.ne.s32 %p138, %r678, %r679; @%p138 bra $L__BB9_218; and.b32 %r680, %r10, -128; ld.u32 %r681, [%rd120+20]; setp.eq.s32 %p139, %r680, %r681; selp.u16 %rs104, 1, 0, %p139; $L__BB9_218: setp.eq.s16 %p140, %rs104, 0; @%p140 bra $L__BB9_222; shl.b32 %r682, %r27, 5; and.b32 %r683, %r682, 3840; shl.b32 %r684, %r2, 1; and.b32 %r685, %r684, 240; or.b32 %r686, %r683, %r685; bfe.u32 %r687, %r10, 3, 4; or.b32 %r688, %r687, %r686; shr.u32 %r689, %r686, 6; ld.u64 %rd121, [%rd120+56]; mul.wide.u32 %rd567, %r689, 8; add.s64 %rd568, %rd121, %rd567; and.b32 %r690, %r688, 63; mov.u64 %rd569, 1; shl.b64 %rd570, %rd569, %r690; ld.u64 %rd571, [%rd568+544]; and.b64 %rd572, %rd571, %rd570; setp.eq.s64 %p141, %rd572, 0; mul.wide.u32 %rd573, %r688, 8; add.s64 %rd574, %rd121, %rd573; add.s64 %rd122, %rd574, 1088; @%p141 bra $L__BB9_221; ld.u64 %rd575, [%rd122]; add.s64 %rd576, %rd121, %rd575; and.b32 %r691, %r2, -8; st.v2.u32 [%rd120], {%r28, %r691}; and.b32 %r692, %r10, -8; st.u32 [%rd120+8], %r692; st.u64 [%rd120+48], %rd576; shl.b32 %r693, %r27, 6; and.b32 %r694, %r693, 448; shl.b32 %r695, %r2, 3; and.b32 %r696, %r695, 56; or.b32 %r697, %r694, %r696; and.b32 %r698, %r10, 7; or.b32 %r699, %r697, %r698; mul.wide.u32 %rd577, %r699, 4; add.s64 %rd578, %rd576, %rd577; ld.f32 %f126, [%rd578+96]; bra.uni $L__BB9_245; $L__BB9_258: shl.b64 %rd642, %rd145, 5; add.s64 %rd707, %rd143, %rd642; $L__BB9_259: setp.eq.s64 %p167, %rd707, 0; @%p167 bra $L__BB9_267; add.s64 %rd148, %rd707, 8; ld.u64 %rd149, [%rd707+8]; setp.eq.s64 %p168, %rd149, 0; @%p168 bra $L__BB9_266; add.s64 %rd150, %rd141, %rd149; and.b32 %r788, %r2, -4096; mov.u64 %rd643, 1; st.v2.u32 [%rd140+24], {%r45, %r788}; and.b32 %r789, %r3, -4096; st.u32 [%rd140+32], %r789; st.u64 [%rd140+64], %rd150; shl.b32 %r790, %r27, 3; and.b32 %r791, %r790, 31744; and.b32 %r792, %r2, 3968; shr.u32 %r793, %r792, 2; or.b32 %r794, %r791, %r793; bfe.u32 %r795, %r3, 7, 5; or.b32 %r796, %r794, %r795; shr.u32 %r797, %r794, 6; mul.wide.u32 %rd644, %r797, 8; add.s64 %rd645, %rd150, %rd644; and.b32 %r798, %r796, 63; shl.b64 %rd646, %rd643, %r798; ld.u64 %rd647, [%rd645+4128]; and.b64 %rd648, %rd647, %rd646; setp.eq.s64 %p169, %rd648, 0; mul.wide.u32 %rd649, %r796, 8; add.s64 %rd650, %rd150, %rd649; add.s64 %rd151, %rd650, 8256; @%p169 bra $L__BB9_265; ld.u64 %rd651, [%rd151]; add.s64 %rd152, %rd150, %rd651; st.u32 [%rd140+12], %r44; and.b32 %r799, %r2, -128; st.u32 [%rd140+16], %r799; and.b32 %r800, %r3, -128; st.u32 [%rd140+20], %r800; st.u64 [%rd140+56], %rd152; shl.b32 %r801, %r27, 5; and.b32 %r802, %r801, 3840; shl.b32 %r803, %r2, 1; and.b32 %r804, %r803, 240; or.b32 %r805, %r802, %r804; bfe.u32 %r806, %r3, 3, 4; or.b32 %r807, %r805, %r806; shr.u32 %r808, %r805, 6; mul.wide.u32 %rd652, %r808, 8; add.s64 %rd653, %rd152, %rd652; and.b32 %r809, %r807, 63; shl.b64 %rd655, %rd643, %r809; ld.u64 %rd656, [%rd653+544]; and.b64 %rd657, %rd656, %rd655; setp.eq.s64 %p170, %rd657, 0; mul.wide.u32 %rd658, %r807, 8; add.s64 %rd659, %rd152, %rd658; add.s64 %rd153, %rd659, 1088; @%p170 bra $L__BB9_264; ld.u64 %rd660, [%rd153]; add.s64 %rd661, %rd152, %rd660; and.b32 %r810, %r2, -8; st.v2.u32 [%rd140], {%r28, %r810}; and.b32 %r811, %r3, -8; st.u32 [%rd140+8], %r811; st.u64 [%rd140+48], %rd661; shl.b32 %r812, %r27, 6; and.b32 %r813, %r812, 448; shl.b32 %r814, %r2, 3; and.b32 %r815, %r814, 56; or.b32 %r816, %r813, %r815; and.b32 %r817, %r3, 7; or.b32 %r818, %r816, %r817; mul.wide.u32 %rd662, %r818, 4; add.s64 %rd663, %rd661, %rd662; ld.f32 %f127, [%rd663+96]; bra.uni $L__BB9_277; $L__BB9_267: ld.f32 %f127, [%rd142+4]; bra.uni $L__BB9_277; $L__BB9_266: ld.f32 %f127, [%rd148+12]; bra.uni $L__BB9_277; $L__BB9_12: and.b32 %r6, %r1, -4096; ld.u32 %r84, [%rd2+24]; setp.ne.s32 %p10, %r6, %r84; mov.u16 %rs87, 0; @%p10 bra $L__BB9_15; and.b32 %r85, %r2, -4096; ld.u32 %r86, [%rd2+28]; setp.ne.s32 %p11, %r85, %r86; @%p11 bra $L__BB9_15; and.b32 %r87, %r3, -4096; ld.u32 %r88, [%rd2+32]; setp.eq.s32 %p12, %r87, %r88; selp.u16 %rs87, 1, 0, %p12; $L__BB9_15: setp.eq.s16 %p13, %rs87, 0; @%p13 bra $L__BB9_21; shl.b32 %r89, %r1, 3; and.b32 %r90, %r89, 31744; and.b32 %r91, %r2, 3968; shr.u32 %r92, %r91, 2; or.b32 %r93, %r92, %r90; bfe.u32 %r94, %r3, 7, 5; or.b32 %r95, %r93, %r94; shr.u32 %r96, %r93, 6; ld.u64 %rd5, [%rd2+64]; mul.wide.u32 %rd176, %r96, 8; add.s64 %rd177, %rd5, %rd176; and.b32 %r97, %r95, 63; mov.u64 %rd178, 1; shl.b64 %rd179, %rd178, %r97; ld.u64 %rd180, [%rd177+4128]; and.b64 %rd181, %rd180, %rd179; setp.eq.s64 %p14, %rd181, 0; mul.wide.u32 %rd182, %r95, 8; add.s64 %rd183, %rd5, %rd182; add.s64 %rd6, %rd183, 8256; @%p14 bra $L__BB9_20; ld.u64 %rd184, [%rd6]; add.s64 %rd7, %rd5, %rd184; st.u32 [%rd2+12], %r5; and.b32 %r98, %r2, -128; st.u32 [%rd2+16], %r98; and.b32 %r99, %r3, -128; st.u32 [%rd2+20], %r99; st.u64 [%rd2+56], %rd7; shl.b32 %r100, %r1, 5; and.b32 %r101, %r100, 3840; shl.b32 %r102, %r2, 1; and.b32 %r103, %r102, 240; or.b32 %r104, %r103, %r101; bfe.u32 %r105, %r3, 3, 4; or.b32 %r106, %r104, %r105; shr.u32 %r107, %r104, 6; mul.wide.u32 %rd185, %r107, 8; add.s64 %rd186, %rd7, %rd185; and.b32 %r108, %r106, 63; shl.b64 %rd188, %rd178, %r108; ld.u64 %rd189, [%rd186+544]; and.b64 %rd190, %rd189, %rd188; setp.eq.s64 %p15, %rd190, 0; mul.wide.u32 %rd191, %r106, 8; add.s64 %rd192, %rd7, %rd191; add.s64 %rd8, %rd192, 1088; @%p15 bra $L__BB9_19; ld.u64 %rd193, [%rd8]; add.s64 %rd194, %rd7, %rd193; and.b32 %r109, %r2, -8; st.v2.u32 [%rd2], {%r4, %r109}; and.b32 %r110, %r3, -8; st.u32 [%rd2+8], %r110; st.u64 [%rd2+48], %rd194; shl.b32 %r111, %r1, 6; and.b32 %r112, %r111, 448; shl.b32 %r113, %r2, 3; and.b32 %r114, %r113, 56; or.b32 %r115, %r114, %r112; and.b32 %r116, %r3, 7; or.b32 %r117, %r115, %r116; mul.wide.u32 %rd195, %r117, 4; add.s64 %rd196, %rd194, %rd195; ld.f32 %f120, [%rd196+96]; bra.uni $L__BB9_35; $L__BB9_47: and.b32 %r12, %r1, -4096; ld.u32 %r188, [%rd22+24]; setp.ne.s32 %p32, %r12, %r188; mov.u16 %rs90, 0; @%p32 bra $L__BB9_50; and.b32 %r189, %r2, -4096; ld.u32 %r190, [%rd22+28]; setp.ne.s32 %p33, %r189, %r190; @%p33 bra $L__BB9_50; and.b32 %r191, %r10, -4096; ld.u32 %r192, [%rd22+32]; setp.eq.s32 %p34, %r191, %r192; selp.u16 %rs90, 1, 0, %p34; $L__BB9_50: setp.eq.s16 %p35, %rs90, 0; @%p35 bra $L__BB9_56; shl.b32 %r193, %r1, 3; and.b32 %r194, %r193, 31744; and.b32 %r195, %r2, 3968; shr.u32 %r196, %r195, 2; or.b32 %r197, %r196, %r194; bfe.u32 %r198, %r10, 7, 5; or.b32 %r199, %r198, %r197; shr.u32 %r200, %r197, 6; ld.u64 %rd25, [%rd22+64]; mul.wide.u32 %rd243, %r200, 8; add.s64 %rd244, %rd25, %rd243; and.b32 %r201, %r199, 63; mov.u64 %rd245, 1; shl.b64 %rd246, %rd245, %r201; ld.u64 %rd247, [%rd244+4128]; and.b64 %rd248, %rd247, %rd246; setp.eq.s64 %p36, %rd248, 0; mul.wide.u32 %rd249, %r199, 8; add.s64 %rd250, %rd25, %rd249; add.s64 %rd26, %rd250, 8256; @%p36 bra $L__BB9_55; ld.u64 %rd251, [%rd26]; add.s64 %rd27, %rd25, %rd251; st.u32 [%rd22+12], %r11; and.b32 %r202, %r2, -128; st.u32 [%rd22+16], %r202; and.b32 %r203, %r10, -128; st.u32 [%rd22+20], %r203; st.u64 [%rd22+56], %rd27; shl.b32 %r204, %r1, 5; and.b32 %r205, %r204, 3840; shl.b32 %r206, %r2, 1; and.b32 %r207, %r206, 240; or.b32 %r208, %r207, %r205; bfe.u32 %r209, %r10, 3, 4; or.b32 %r210, %r209, %r208; shr.u32 %r211, %r208, 6; mul.wide.u32 %rd252, %r211, 8; add.s64 %rd253, %rd27, %rd252; and.b32 %r212, %r210, 63; shl.b64 %rd255, %rd245, %r212; ld.u64 %rd256, [%rd253+544]; and.b64 %rd257, %rd256, %rd255; setp.eq.s64 %p37, %rd257, 0; mul.wide.u32 %rd258, %r210, 8; add.s64 %rd259, %rd27, %rd258; add.s64 %rd28, %rd259, 1088; @%p37 bra $L__BB9_54; ld.u64 %rd260, [%rd28]; add.s64 %rd261, %rd27, %rd260; and.b32 %r213, %r2, -8; st.v2.u32 [%rd22], {%r4, %r213}; and.b32 %r214, %r10, -8; st.u32 [%rd22+8], %r214; st.u64 [%rd22+48], %rd261; shl.b32 %r215, %r1, 6; and.b32 %r216, %r215, 448; shl.b32 %r217, %r2, 3; and.b32 %r218, %r217, 56; or.b32 %r219, %r218, %r216; and.b32 %r220, %r10, 7; or.b32 %r221, %r219, %r220; mul.wide.u32 %rd262, %r221, 4; add.s64 %rd263, %rd261, %rd262; ld.f32 %f121, [%rd263+96]; bra.uni $L__BB9_70; $L__BB9_82: and.b32 %r18, %r1, -4096; ld.u32 %r292, [%rd42+24]; setp.ne.s32 %p54, %r18, %r292; mov.u16 %rs93, 0; @%p54 bra $L__BB9_85; and.b32 %r293, %r16, -4096; ld.u32 %r294, [%rd42+28]; setp.ne.s32 %p55, %r293, %r294; @%p55 bra $L__BB9_85; and.b32 %r295, %r10, -4096; ld.u32 %r296, [%rd42+32]; setp.eq.s32 %p56, %r295, %r296; selp.u16 %rs93, 1, 0, %p56; $L__BB9_85: setp.eq.s16 %p57, %rs93, 0; @%p57 bra $L__BB9_91; shl.b32 %r297, %r1, 3; and.b32 %r298, %r297, 31744; and.b32 %r299, %r16, 3968; shr.u32 %r300, %r299, 2; or.b32 %r301, %r300, %r298; bfe.u32 %r302, %r10, 7, 5; or.b32 %r303, %r301, %r302; shr.u32 %r304, %r301, 6; ld.u64 %rd45, [%rd42+64]; mul.wide.u32 %rd311, %r304, 8; add.s64 %rd312, %rd45, %rd311; and.b32 %r305, %r303, 63; mov.u64 %rd313, 1; shl.b64 %rd314, %rd313, %r305; ld.u64 %rd315, [%rd312+4128]; and.b64 %rd316, %rd315, %rd314; setp.eq.s64 %p58, %rd316, 0; mul.wide.u32 %rd317, %r303, 8; add.s64 %rd318, %rd45, %rd317; add.s64 %rd46, %rd318, 8256; @%p58 bra $L__BB9_90; ld.u64 %rd319, [%rd46]; add.s64 %rd47, %rd45, %rd319; st.u32 [%rd42+12], %r17; and.b32 %r306, %r16, -128; st.u32 [%rd42+16], %r306; and.b32 %r307, %r10, -128; st.u32 [%rd42+20], %r307; st.u64 [%rd42+56], %rd47; shl.b32 %r308, %r1, 5; and.b32 %r309, %r308, 3840; shl.b32 %r310, %r16, 1; and.b32 %r311, %r310, 240; or.b32 %r312, %r311, %r309; bfe.u32 %r313, %r10, 3, 4; or.b32 %r314, %r312, %r313; shr.u32 %r315, %r312, 6; mul.wide.u32 %rd320, %r315, 8; add.s64 %rd321, %rd47, %rd320; and.b32 %r316, %r314, 63; shl.b64 %rd323, %rd313, %r316; ld.u64 %rd324, [%rd321+544]; and.b64 %rd325, %rd324, %rd323; setp.eq.s64 %p59, %rd325, 0; mul.wide.u32 %rd326, %r314, 8; add.s64 %rd327, %rd47, %rd326; add.s64 %rd48, %rd327, 1088; @%p59 bra $L__BB9_89; ld.u64 %rd328, [%rd48]; add.s64 %rd329, %rd47, %rd328; and.b32 %r317, %r16, -8; st.v2.u32 [%rd42], {%r4, %r317}; and.b32 %r318, %r10, -8; st.u32 [%rd42+8], %r318; st.u64 [%rd42+48], %rd329; shl.b32 %r319, %r1, 6; and.b32 %r320, %r319, 448; shl.b32 %r321, %r16, 3; and.b32 %r322, %r321, 56; or.b32 %r323, %r322, %r320; and.b32 %r324, %r10, 7; or.b32 %r325, %r323, %r324; mul.wide.u32 %rd330, %r325, 4; add.s64 %rd331, %rd329, %rd330; ld.f32 %f122, [%rd331+96]; bra.uni $L__BB9_105; $L__BB9_117: and.b32 %r23, %r1, -4096; ld.u32 %r394, [%rd62+24]; setp.ne.s32 %p76, %r23, %r394; mov.u16 %rs96, 0; @%p76 bra $L__BB9_120; and.b32 %r395, %r16, -4096; ld.u32 %r396, [%rd62+28]; setp.ne.s32 %p77, %r395, %r396; @%p77 bra $L__BB9_120; and.b32 %r397, %r3, -4096; ld.u32 %r398, [%rd62+32]; setp.eq.s32 %p78, %r397, %r398; selp.u16 %rs96, 1, 0, %p78; $L__BB9_120: setp.eq.s16 %p79, %rs96, 0; @%p79 bra $L__BB9_126; shl.b32 %r399, %r1, 3; and.b32 %r400, %r399, 31744; and.b32 %r401, %r16, 3968; shr.u32 %r402, %r401, 2; or.b32 %r403, %r402, %r400; bfe.u32 %r404, %r3, 7, 5; or.b32 %r405, %r403, %r404; shr.u32 %r406, %r403, 6; ld.u64 %rd65, [%rd62+64]; mul.wide.u32 %rd378, %r406, 8; add.s64 %rd379, %rd65, %rd378; and.b32 %r407, %r405, 63; mov.u64 %rd380, 1; shl.b64 %rd381, %rd380, %r407; ld.u64 %rd382, [%rd379+4128]; and.b64 %rd383, %rd382, %rd381; setp.eq.s64 %p80, %rd383, 0; mul.wide.u32 %rd384, %r405, 8; add.s64 %rd385, %rd65, %rd384; add.s64 %rd66, %rd385, 8256; @%p80 bra $L__BB9_125; ld.u64 %rd386, [%rd66]; add.s64 %rd67, %rd65, %rd386; st.u32 [%rd62+12], %r22; and.b32 %r408, %r16, -128; st.u32 [%rd62+16], %r408; and.b32 %r409, %r3, -128; st.u32 [%rd62+20], %r409; st.u64 [%rd62+56], %rd67; shl.b32 %r410, %r1, 5; and.b32 %r411, %r410, 3840; shl.b32 %r412, %r16, 1; and.b32 %r413, %r412, 240; or.b32 %r414, %r413, %r411; bfe.u32 %r415, %r3, 3, 4; or.b32 %r416, %r414, %r415; shr.u32 %r417, %r414, 6; mul.wide.u32 %rd387, %r417, 8; add.s64 %rd388, %rd67, %rd387; and.b32 %r418, %r416, 63; shl.b64 %rd390, %rd380, %r418; ld.u64 %rd391, [%rd388+544]; and.b64 %rd392, %rd391, %rd390; setp.eq.s64 %p81, %rd392, 0; mul.wide.u32 %rd393, %r416, 8; add.s64 %rd394, %rd67, %rd393; add.s64 %rd68, %rd394, 1088; @%p81 bra $L__BB9_124; ld.u64 %rd395, [%rd68]; add.s64 %rd396, %rd67, %rd395; and.b32 %r419, %r16, -8; st.v2.u32 [%rd62], {%r4, %r419}; and.b32 %r420, %r3, -8; st.u32 [%rd62+8], %r420; st.u64 [%rd62+48], %rd396; and.b32 %r421, %r1, 7; shl.b32 %r422, %r16, 3; and.b32 %r423, %r422, 56; and.b32 %r424, %r3, 7; bfi.b32 %r425, %r421, %r424, 6, 3; or.b32 %r426, %r425, %r423; mul.wide.u32 %rd397, %r426, 4; add.s64 %rd398, %rd396, %rd397; ld.f32 %f123, [%rd398+96]; bra.uni $L__BB9_140; $L__BB9_152: and.b32 %r30, %r27, -4096; ld.u32 %r494, [%rd81+24]; setp.ne.s32 %p98, %r30, %r494; mov.u16 %rs99, 0; @%p98 bra $L__BB9_155; and.b32 %r495, %r16, -4096; ld.u32 %r496, [%rd81+28]; setp.ne.s32 %p99, %r495, %r496; @%p99 bra $L__BB9_155; and.b32 %r497, %r3, -4096; ld.u32 %r498, [%rd81+32]; setp.eq.s32 %p100, %r497, %r498; selp.u16 %rs99, 1, 0, %p100; $L__BB9_155: setp.eq.s16 %p101, %rs99, 0; @%p101 bra $L__BB9_161; shl.b32 %r499, %r27, 3; and.b32 %r500, %r499, 31744; and.b32 %r501, %r16, 3968; shr.u32 %r502, %r501, 2; or.b32 %r503, %r502, %r500; bfe.u32 %r504, %r3, 7, 5; or.b32 %r505, %r503, %r504; shr.u32 %r506, %r503, 6; ld.u64 %rd84, [%rd81+64]; mul.wide.u32 %rd445, %r506, 8; add.s64 %rd446, %rd84, %rd445; and.b32 %r507, %r505, 63; mov.u64 %rd447, 1; shl.b64 %rd448, %rd447, %r507; ld.u64 %rd449, [%rd446+4128]; and.b64 %rd450, %rd449, %rd448; setp.eq.s64 %p102, %rd450, 0; mul.wide.u32 %rd451, %r505, 8; add.s64 %rd452, %rd84, %rd451; add.s64 %rd85, %rd452, 8256; @%p102 bra $L__BB9_160; ld.u64 %rd453, [%rd85]; add.s64 %rd86, %rd84, %rd453; st.u32 [%rd81+12], %r29; and.b32 %r508, %r16, -128; st.u32 [%rd81+16], %r508; and.b32 %r509, %r3, -128; st.u32 [%rd81+20], %r509; st.u64 [%rd81+56], %rd86; shl.b32 %r510, %r27, 5; and.b32 %r511, %r510, 3840; shl.b32 %r512, %r16, 1; and.b32 %r513, %r512, 240; or.b32 %r514, %r513, %r511; bfe.u32 %r515, %r3, 3, 4; or.b32 %r516, %r514, %r515; shr.u32 %r517, %r514, 6; mul.wide.u32 %rd454, %r517, 8; add.s64 %rd455, %rd86, %rd454; and.b32 %r518, %r516, 63; shl.b64 %rd457, %rd447, %r518; ld.u64 %rd458, [%rd455+544]; and.b64 %rd459, %rd458, %rd457; setp.eq.s64 %p103, %rd459, 0; mul.wide.u32 %rd460, %r516, 8; add.s64 %rd461, %rd86, %rd460; add.s64 %rd87, %rd461, 1088; @%p103 bra $L__BB9_159; ld.u64 %rd462, [%rd87]; add.s64 %rd463, %rd86, %rd462; and.b32 %r519, %r16, -8; st.v2.u32 [%rd81], {%r28, %r519}; and.b32 %r520, %r3, -8; st.u32 [%rd81+8], %r520; st.u64 [%rd81+48], %rd463; and.b32 %r521, %r27, 7; shl.b32 %r522, %r16, 3; and.b32 %r523, %r522, 56; and.b32 %r524, %r3, 7; bfi.b32 %r525, %r521, %r524, 6, 3; or.b32 %r526, %r525, %r523; mul.wide.u32 %rd464, %r526, 4; add.s64 %rd465, %rd463, %rd464; ld.f32 %f124, [%rd465+96]; bra.uni $L__BB9_175; $L__BB9_187: and.b32 %r35, %r27, -4096; ld.u32 %r596, [%rd101+24]; setp.ne.s32 %p120, %r35, %r596; mov.u16 %rs102, 0; @%p120 bra $L__BB9_190; and.b32 %r597, %r16, -4096; ld.u32 %r598, [%rd101+28]; setp.ne.s32 %p121, %r597, %r598; @%p121 bra $L__BB9_190; and.b32 %r599, %r10, -4096; ld.u32 %r600, [%rd101+32]; setp.eq.s32 %p122, %r599, %r600; selp.u16 %rs102, 1, 0, %p122; $L__BB9_190: setp.eq.s16 %p123, %rs102, 0; @%p123 bra $L__BB9_196; shl.b32 %r601, %r27, 3; and.b32 %r602, %r601, 31744; and.b32 %r603, %r16, 3968; shr.u32 %r604, %r603, 2; or.b32 %r605, %r604, %r602; bfe.u32 %r606, %r10, 7, 5; or.b32 %r607, %r605, %r606; shr.u32 %r608, %r605, 6; ld.u64 %rd104, [%rd101+64]; mul.wide.u32 %rd512, %r608, 8; add.s64 %rd513, %rd104, %rd512; and.b32 %r609, %r607, 63; mov.u64 %rd514, 1; shl.b64 %rd515, %rd514, %r609; ld.u64 %rd516, [%rd513+4128]; and.b64 %rd517, %rd516, %rd515; setp.eq.s64 %p124, %rd517, 0; mul.wide.u32 %rd518, %r607, 8; add.s64 %rd519, %rd104, %rd518; add.s64 %rd105, %rd519, 8256; @%p124 bra $L__BB9_195; ld.u64 %rd520, [%rd105]; add.s64 %rd106, %rd104, %rd520; st.u32 [%rd101+12], %r34; and.b32 %r610, %r16, -128; st.u32 [%rd101+16], %r610; and.b32 %r611, %r10, -128; st.u32 [%rd101+20], %r611; st.u64 [%rd101+56], %rd106; shl.b32 %r612, %r27, 5; and.b32 %r613, %r612, 3840; shl.b32 %r614, %r16, 1; and.b32 %r615, %r614, 240; or.b32 %r616, %r615, %r613; bfe.u32 %r617, %r10, 3, 4; or.b32 %r618, %r616, %r617; shr.u32 %r619, %r616, 6; mul.wide.u32 %rd521, %r619, 8; add.s64 %rd522, %rd106, %rd521; and.b32 %r620, %r618, 63; shl.b64 %rd524, %rd514, %r620; ld.u64 %rd525, [%rd522+544]; and.b64 %rd526, %rd525, %rd524; setp.eq.s64 %p125, %rd526, 0; mul.wide.u32 %rd527, %r618, 8; add.s64 %rd528, %rd106, %rd527; add.s64 %rd107, %rd528, 1088; @%p125 bra $L__BB9_194; ld.u64 %rd529, [%rd107]; add.s64 %rd530, %rd106, %rd529; and.b32 %r621, %r16, -8; st.v2.u32 [%rd101], {%r28, %r621}; and.b32 %r622, %r10, -8; st.u32 [%rd101+8], %r622; st.u64 [%rd101+48], %rd530; shl.b32 %r623, %r27, 6; and.b32 %r624, %r623, 448; shl.b32 %r625, %r16, 3; and.b32 %r626, %r625, 56; or.b32 %r627, %r626, %r624; and.b32 %r628, %r10, 7; or.b32 %r629, %r627, %r628; mul.wide.u32 %rd531, %r629, 4; add.s64 %rd532, %rd530, %rd531; ld.f32 %f125, [%rd532+96]; bra.uni $L__BB9_210; $L__BB9_222: and.b32 %r40, %r27, -4096; ld.u32 %r700, [%rd120+24]; setp.ne.s32 %p142, %r40, %r700; mov.u16 %rs105, 0; @%p142 bra $L__BB9_225; and.b32 %r701, %r2, -4096; ld.u32 %r702, [%rd120+28]; setp.ne.s32 %p143, %r701, %r702; @%p143 bra $L__BB9_225; and.b32 %r703, %r10, -4096; ld.u32 %r704, [%rd120+32]; setp.eq.s32 %p144, %r703, %r704; selp.u16 %rs105, 1, 0, %p144; $L__BB9_225: setp.eq.s16 %p145, %rs105, 0; @%p145 bra $L__BB9_231; shl.b32 %r705, %r27, 3; and.b32 %r706, %r705, 31744; and.b32 %r707, %r2, 3968; shr.u32 %r708, %r707, 2; or.b32 %r709, %r706, %r708; bfe.u32 %r710, %r10, 7, 5; or.b32 %r711, %r710, %r709; shr.u32 %r712, %r709, 6; ld.u64 %rd123, [%rd120+64]; mul.wide.u32 %rd579, %r712, 8; add.s64 %rd580, %rd123, %rd579; and.b32 %r713, %r711, 63; mov.u64 %rd581, 1; shl.b64 %rd582, %rd581, %r713; ld.u64 %rd583, [%rd580+4128]; and.b64 %rd584, %rd583, %rd582; setp.eq.s64 %p146, %rd584, 0; mul.wide.u32 %rd585, %r711, 8; add.s64 %rd586, %rd123, %rd585; add.s64 %rd124, %rd586, 8256; @%p146 bra $L__BB9_230; ld.u64 %rd587, [%rd124]; add.s64 %rd125, %rd123, %rd587; st.u32 [%rd120+12], %r39; and.b32 %r714, %r2, -128; st.u32 [%rd120+16], %r714; and.b32 %r715, %r10, -128; st.u32 [%rd120+20], %r715; st.u64 [%rd120+56], %rd125; shl.b32 %r716, %r27, 5; and.b32 %r717, %r716, 3840; shl.b32 %r718, %r2, 1; and.b32 %r719, %r718, 240; or.b32 %r720, %r717, %r719; bfe.u32 %r721, %r10, 3, 4; or.b32 %r722, %r721, %r720; shr.u32 %r723, %r720, 6; mul.wide.u32 %rd588, %r723, 8; add.s64 %rd589, %rd125, %rd588; and.b32 %r724, %r722, 63; shl.b64 %rd591, %rd581, %r724; ld.u64 %rd592, [%rd589+544]; and.b64 %rd593, %rd592, %rd591; setp.eq.s64 %p147, %rd593, 0; mul.wide.u32 %rd594, %r722, 8; add.s64 %rd595, %rd125, %rd594; add.s64 %rd126, %rd595, 1088; @%p147 bra $L__BB9_229; ld.u64 %rd596, [%rd126]; add.s64 %rd597, %rd125, %rd596; and.b32 %r725, %r2, -8; st.v2.u32 [%rd120], {%r28, %r725}; and.b32 %r726, %r10, -8; st.u32 [%rd120+8], %r726; st.u64 [%rd120+48], %rd597; shl.b32 %r727, %r27, 6; and.b32 %r728, %r727, 448; shl.b32 %r729, %r2, 3; and.b32 %r730, %r729, 56; or.b32 %r731, %r728, %r730; and.b32 %r732, %r10, 7; or.b32 %r733, %r731, %r732; mul.wide.u32 %rd598, %r733, 4; add.s64 %rd599, %rd597, %rd598; ld.f32 %f126, [%rd599+96]; bra.uni $L__BB9_245; $L__BB9_265: ld.f32 %f127, [%rd151]; bra.uni $L__BB9_277; $L__BB9_11: ld.f32 %f120, [%rd4]; bra.uni $L__BB9_35; $L__BB9_46: ld.f32 %f121, [%rd24]; bra.uni $L__BB9_70; $L__BB9_81: ld.f32 %f122, [%rd44]; bra.uni $L__BB9_105; $L__BB9_116: ld.f32 %f123, [%rd64]; bra.uni $L__BB9_140; $L__BB9_151: ld.f32 %f124, [%rd83]; bra.uni $L__BB9_175; $L__BB9_186: ld.f32 %f125, [%rd103]; bra.uni $L__BB9_210; $L__BB9_221: ld.f32 %f126, [%rd122]; bra.uni $L__BB9_245; $L__BB9_264: ld.f32 %f127, [%rd153]; bra.uni $L__BB9_277; $L__BB9_21: ld.u64 %rd9, [%rd2+40]; mov.u64 %rd700, 0; add.s64 %rd10, %rd9, 24; ld.u32 %r7, [%rd9+24]; setp.eq.s32 %p16, %r7, 0; @%p16 bra $L__BB9_26; add.s64 %rd11, %rd9, 64; shr.u32 %r119, %r1, 12; cvt.u64.u32 %rd198, %r119; shr.u32 %r120, %r3, 12; cvt.u64.u32 %rd199, %r120; bfi.b64 %rd200, %rd198, %rd199, 42, 22; shr.u32 %r121, %r2, 12; mul.wide.u32 %rd201, %r121, 2097152; or.b64 %rd12, %rd200, %rd201; mov.u32 %r873, 0; $L__BB9_23: cvt.u64.u32 %rd13, %r873; mul.wide.u32 %rd202, %r873, 32; add.s64 %rd203, %rd11, %rd202; ld.u64 %rd204, [%rd203]; setp.eq.s64 %p17, %rd204, %rd12; add.s32 %r873, %r873, 1; @%p17 bra $L__BB9_25; setp.lt.u32 %p18, %r873, %r7; @%p18 bra $L__BB9_23; bra.uni $L__BB9_26; $L__BB9_56: ld.u64 %rd29, [%rd22+40]; mov.u64 %rd701, 0; add.s64 %rd30, %rd29, 24; ld.u32 %r13, [%rd29+24]; setp.eq.s32 %p38, %r13, 0; @%p38 bra $L__BB9_61; add.s64 %rd31, %rd29, 64; shr.u32 %r223, %r1, 12; cvt.u64.u32 %rd265, %r223; shl.b64 %rd266, %rd265, 42; shr.u32 %r224, %r2, 12; mul.wide.u32 %rd267, %r224, 2097152; or.b64 %rd268, %rd267, %rd266; shr.u32 %r225, %r10, 12; cvt.u64.u32 %rd269, %r225; or.b64 %rd32, %rd268, %rd269; mov.u32 %r874, 0; $L__BB9_58: cvt.u64.u32 %rd33, %r874; mul.wide.u32 %rd270, %r874, 32; add.s64 %rd271, %rd31, %rd270; ld.u64 %rd272, [%rd271]; setp.eq.s64 %p39, %rd272, %rd32; add.s32 %r874, %r874, 1; @%p39 bra $L__BB9_60; setp.lt.u32 %p40, %r874, %r13; @%p40 bra $L__BB9_58; bra.uni $L__BB9_61; $L__BB9_91: ld.u64 %rd49, [%rd42+40]; mov.u64 %rd702, 0; add.s64 %rd50, %rd49, 24; ld.u32 %r19, [%rd49+24]; setp.eq.s32 %p60, %r19, 0; @%p60 bra $L__BB9_96; add.s64 %rd51, %rd49, 64; shr.u32 %r327, %r1, 12; cvt.u64.u32 %rd333, %r327; shr.u32 %r328, %r10, 12; cvt.u64.u32 %rd334, %r328; bfi.b64 %rd335, %rd333, %rd334, 42, 22; shr.u32 %r329, %r16, 12; mul.wide.u32 %rd336, %r329, 2097152; or.b64 %rd52, %rd335, %rd336; mov.u32 %r875, 0; $L__BB9_93: cvt.u64.u32 %rd53, %r875; mul.wide.u32 %rd337, %r875, 32; add.s64 %rd338, %rd51, %rd337; ld.u64 %rd339, [%rd338]; setp.eq.s64 %p61, %rd339, %rd52; add.s32 %r875, %r875, 1; @%p61 bra $L__BB9_95; setp.lt.u32 %p62, %r875, %r19; @%p62 bra $L__BB9_93; bra.uni $L__BB9_96; $L__BB9_126: ld.u64 %rd69, [%rd62+40]; mov.u64 %rd703, 0; ld.u32 %r24, [%rd69+24]; setp.eq.s32 %p82, %r24, 0; @%p82 bra $L__BB9_131; add.s64 %rd70, %rd69, 64; shr.u32 %r428, %r1, 12; cvt.u64.u32 %rd400, %r428; shr.u32 %r429, %r3, 12; cvt.u64.u32 %rd401, %r429; bfi.b64 %rd402, %rd400, %rd401, 42, 22; shr.u32 %r430, %r16, 12; mul.wide.u32 %rd403, %r430, 2097152; or.b64 %rd71, %rd402, %rd403; mov.u32 %r876, 0; $L__BB9_128: cvt.u64.u32 %rd72, %r876; mul.wide.u32 %rd404, %r876, 32; add.s64 %rd405, %rd70, %rd404; ld.u64 %rd406, [%rd405]; setp.eq.s64 %p83, %rd406, %rd71; add.s32 %r876, %r876, 1; @%p83 bra $L__BB9_130; setp.lt.u32 %p84, %r876, %r24; @%p84 bra $L__BB9_128; bra.uni $L__BB9_131; $L__BB9_161: ld.u64 %rd88, [%rd81+40]; mov.u64 %rd704, 0; add.s64 %rd89, %rd88, 24; ld.u32 %r31, [%rd88+24]; setp.eq.s32 %p104, %r31, 0; @%p104 bra $L__BB9_166; add.s64 %rd90, %rd88, 64; shr.u32 %r528, %r27, 12; cvt.u64.u32 %rd467, %r528; shr.u32 %r529, %r3, 12; cvt.u64.u32 %rd468, %r529; bfi.b64 %rd469, %rd467, %rd468, 42, 22; shr.u32 %r530, %r16, 12; mul.wide.u32 %rd470, %r530, 2097152; or.b64 %rd91, %rd469, %rd470; mov.u32 %r877, 0; $L__BB9_163: cvt.u64.u32 %rd92, %r877; mul.wide.u32 %rd471, %r877, 32; add.s64 %rd472, %rd90, %rd471; ld.u64 %rd473, [%rd472]; setp.eq.s64 %p105, %rd473, %rd91; add.s32 %r877, %r877, 1; @%p105 bra $L__BB9_165; setp.lt.u32 %p106, %r877, %r31; @%p106 bra $L__BB9_163; bra.uni $L__BB9_166; $L__BB9_196: ld.u64 %rd108, [%rd101+40]; mov.u64 %rd705, 0; ld.u32 %r36, [%rd108+24]; setp.eq.s32 %p126, %r36, 0; @%p126 bra $L__BB9_201; add.s64 %rd109, %rd108, 64; shr.u32 %r631, %r27, 12; cvt.u64.u32 %rd534, %r631; shr.u32 %r632, %r10, 12; cvt.u64.u32 %rd535, %r632; bfi.b64 %rd536, %rd534, %rd535, 42, 22; shr.u32 %r633, %r16, 12; mul.wide.u32 %rd537, %r633, 2097152; or.b64 %rd110, %rd536, %rd537; mov.u32 %r878, 0; $L__BB9_198: cvt.u64.u32 %rd111, %r878; mul.wide.u32 %rd538, %r878, 32; add.s64 %rd539, %rd109, %rd538; ld.u64 %rd540, [%rd539]; setp.eq.s64 %p127, %rd540, %rd110; add.s32 %r878, %r878, 1; @%p127 bra $L__BB9_200; setp.lt.u32 %p128, %r878, %r36; @%p128 bra $L__BB9_198; bra.uni $L__BB9_201; $L__BB9_231: ld.u64 %rd127, [%rd120+40]; mov.u64 %rd706, 0; add.s64 %rd128, %rd127, 24; ld.u32 %r41, [%rd127+24]; setp.eq.s32 %p148, %r41, 0; @%p148 bra $L__BB9_236; add.s64 %rd129, %rd127, 64; shr.u32 %r735, %r27, 12; cvt.u64.u32 %rd601, %r735; shl.b64 %rd602, %rd601, 42; shr.u32 %r736, %r2, 12; mul.wide.u32 %rd603, %r736, 2097152; or.b64 %rd604, %rd602, %rd603; shr.u32 %r737, %r10, 12; cvt.u64.u32 %rd605, %r737; or.b64 %rd130, %rd604, %rd605; mov.u32 %r879, 0; $L__BB9_233: cvt.u64.u32 %rd131, %r879; mul.wide.u32 %rd606, %r879, 32; add.s64 %rd607, %rd129, %rd606; ld.u64 %rd608, [%rd607]; setp.eq.s64 %p149, %rd608, %rd130; add.s32 %r879, %r879, 1; @%p149 bra $L__BB9_235; setp.lt.u32 %p150, %r879, %r41; @%p150 bra $L__BB9_233; bra.uni $L__BB9_236; $L__BB9_275: ld.f32 %f127, [%rd159]; bra.uni $L__BB9_277; $L__BB9_272: ld.f32 %f127, [%rd155]; bra.uni $L__BB9_277; $L__BB9_20: ld.f32 %f120, [%rd6]; bra.uni $L__BB9_35; $L__BB9_55: ld.f32 %f121, [%rd26]; bra.uni $L__BB9_70; $L__BB9_90: ld.f32 %f122, [%rd46]; bra.uni $L__BB9_105; $L__BB9_125: ld.f32 %f123, [%rd66]; bra.uni $L__BB9_140; $L__BB9_160: ld.f32 %f124, [%rd85]; bra.uni $L__BB9_175; $L__BB9_195: ld.f32 %f125, [%rd105]; bra.uni $L__BB9_210; $L__BB9_230: ld.f32 %f126, [%rd124]; bra.uni $L__BB9_245; $L__BB9_271: ld.f32 %f127, [%rd157]; $L__BB9_277: sub.ftz.f32 %f114, %f126, %f127; fma.rn.ftz.f32 %f115, %f3, %f114, %f127; sub.ftz.f32 %f116, %f78, %f115; fma.rn.ftz.f32 %f117, %f2, %f116, %f115; sub.ftz.f32 %f118, %f117, %f53; fma.rn.ftz.f32 %f119, %f1, %f118, %f53; st.param.f32 [func_retval0+0], %f119; ret; $L__BB9_19: ld.f32 %f120, [%rd8]; bra.uni $L__BB9_35; $L__BB9_54: ld.f32 %f121, [%rd28]; bra.uni $L__BB9_70; $L__BB9_89: ld.f32 %f122, [%rd48]; bra.uni $L__BB9_105; $L__BB9_124: ld.f32 %f123, [%rd68]; bra.uni $L__BB9_140; $L__BB9_159: ld.f32 %f124, [%rd87]; bra.uni $L__BB9_175; $L__BB9_194: ld.f32 %f125, [%rd107]; bra.uni $L__BB9_210; $L__BB9_229: ld.f32 %f126, [%rd126]; bra.uni $L__BB9_245; $L__BB9_25: shl.b64 %rd206, %rd13, 5; add.s64 %rd700, %rd11, %rd206; $L__BB9_26: setp.eq.s64 %p19, %rd700, 0; @%p19 bra $L__BB9_34; add.s64 %rd16, %rd700, 8; ld.u64 %rd17, [%rd700+8]; setp.eq.s64 %p20, %rd17, 0; @%p20 bra $L__BB9_33; add.s64 %rd18, %rd9, %rd17; and.b32 %r122, %r2, -4096; mov.u64 %rd207, 1; st.v2.u32 [%rd2+24], {%r6, %r122}; and.b32 %r123, %r3, -4096; st.u32 [%rd2+32], %r123; st.u64 [%rd2+64], %rd18; shl.b32 %r124, %r1, 3; and.b32 %r125, %r124, 31744; and.b32 %r126, %r2, 3968; shr.u32 %r127, %r126, 2; or.b32 %r128, %r127, %r125; bfe.u32 %r129, %r3, 7, 5; or.b32 %r130, %r128, %r129; shr.u32 %r131, %r128, 6; mul.wide.u32 %rd208, %r131, 8; add.s64 %rd209, %rd18, %rd208; and.b32 %r132, %r130, 63; shl.b64 %rd210, %rd207, %r132; ld.u64 %rd211, [%rd209+4128]; and.b64 %rd212, %rd211, %rd210; setp.eq.s64 %p21, %rd212, 0; mul.wide.u32 %rd213, %r130, 8; add.s64 %rd214, %rd18, %rd213; add.s64 %rd19, %rd214, 8256; @%p21 bra $L__BB9_32; ld.u64 %rd215, [%rd19]; add.s64 %rd20, %rd18, %rd215; st.u32 [%rd2+12], %r5; and.b32 %r133, %r2, -128; st.u32 [%rd2+16], %r133; and.b32 %r134, %r3, -128; st.u32 [%rd2+20], %r134; st.u64 [%rd2+56], %rd20; shl.b32 %r135, %r1, 5; and.b32 %r136, %r135, 3840; shl.b32 %r137, %r2, 1; and.b32 %r138, %r137, 240; or.b32 %r139, %r138, %r136; bfe.u32 %r140, %r3, 3, 4; or.b32 %r141, %r139, %r140; shr.u32 %r142, %r139, 6; mul.wide.u32 %rd216, %r142, 8; add.s64 %rd217, %rd20, %rd216; and.b32 %r143, %r141, 63; shl.b64 %rd219, %rd207, %r143; ld.u64 %rd220, [%rd217+544]; and.b64 %rd221, %rd220, %rd219; setp.eq.s64 %p22, %rd221, 0; mul.wide.u32 %rd222, %r141, 8; add.s64 %rd223, %rd20, %rd222; add.s64 %rd21, %rd223, 1088; @%p22 bra $L__BB9_31; ld.u64 %rd224, [%rd21]; add.s64 %rd225, %rd20, %rd224; and.b32 %r144, %r2, -8; st.v2.u32 [%rd2], {%r4, %r144}; and.b32 %r145, %r3, -8; st.u32 [%rd2+8], %r145; st.u64 [%rd2+48], %rd225; shl.b32 %r146, %r1, 6; and.b32 %r147, %r146, 448; shl.b32 %r148, %r2, 3; and.b32 %r149, %r148, 56; or.b32 %r150, %r149, %r147; and.b32 %r151, %r3, 7; or.b32 %r152, %r150, %r151; mul.wide.u32 %rd226, %r152, 4; add.s64 %rd227, %rd225, %rd226; ld.f32 %f120, [%rd227+96]; bra.uni $L__BB9_35; $L__BB9_34: ld.f32 %f120, [%rd10+4]; bra.uni $L__BB9_35; $L__BB9_60: shl.b64 %rd274, %rd33, 5; add.s64 %rd701, %rd31, %rd274; $L__BB9_61: setp.eq.s64 %p41, %rd701, 0; @%p41 bra $L__BB9_69; add.s64 %rd36, %rd701, 8; ld.u64 %rd37, [%rd701+8]; setp.eq.s64 %p42, %rd37, 0; @%p42 bra $L__BB9_68; add.s64 %rd38, %rd29, %rd37; and.b32 %r226, %r2, -4096; mov.u64 %rd275, 1; st.v2.u32 [%rd22+24], {%r12, %r226}; and.b32 %r227, %r10, -4096; st.u32 [%rd22+32], %r227; st.u64 [%rd22+64], %rd38; shl.b32 %r228, %r1, 3; and.b32 %r229, %r228, 31744; and.b32 %r230, %r2, 3968; shr.u32 %r231, %r230, 2; or.b32 %r232, %r231, %r229; bfe.u32 %r233, %r10, 7, 5; or.b32 %r234, %r233, %r232; shr.u32 %r235, %r232, 6; mul.wide.u32 %rd276, %r235, 8; add.s64 %rd277, %rd38, %rd276; and.b32 %r236, %r234, 63; shl.b64 %rd278, %rd275, %r236; ld.u64 %rd279, [%rd277+4128]; and.b64 %rd280, %rd279, %rd278; setp.eq.s64 %p43, %rd280, 0; mul.wide.u32 %rd281, %r234, 8; add.s64 %rd282, %rd38, %rd281; add.s64 %rd39, %rd282, 8256; @%p43 bra $L__BB9_67; ld.u64 %rd283, [%rd39]; add.s64 %rd40, %rd38, %rd283; st.u32 [%rd22+12], %r11; and.b32 %r237, %r2, -128; st.u32 [%rd22+16], %r237; and.b32 %r238, %r10, -128; st.u32 [%rd22+20], %r238; st.u64 [%rd22+56], %rd40; shl.b32 %r239, %r1, 5; and.b32 %r240, %r239, 3840; shl.b32 %r241, %r2, 1; and.b32 %r242, %r241, 240; or.b32 %r243, %r242, %r240; bfe.u32 %r244, %r10, 3, 4; or.b32 %r245, %r244, %r243; shr.u32 %r246, %r243, 6; mul.wide.u32 %rd284, %r246, 8; add.s64 %rd285, %rd40, %rd284; and.b32 %r247, %r245, 63; shl.b64 %rd287, %rd275, %r247; ld.u64 %rd288, [%rd285+544]; and.b64 %rd289, %rd288, %rd287; setp.eq.s64 %p44, %rd289, 0; mul.wide.u32 %rd290, %r245, 8; add.s64 %rd291, %rd40, %rd290; add.s64 %rd41, %rd291, 1088; @%p44 bra $L__BB9_66; ld.u64 %rd292, [%rd41]; add.s64 %rd293, %rd40, %rd292; and.b32 %r248, %r2, -8; st.v2.u32 [%rd22], {%r4, %r248}; and.b32 %r249, %r10, -8; st.u32 [%rd22+8], %r249; st.u64 [%rd22+48], %rd293; shl.b32 %r250, %r1, 6; and.b32 %r251, %r250, 448; shl.b32 %r252, %r2, 3; and.b32 %r253, %r252, 56; or.b32 %r254, %r253, %r251; and.b32 %r255, %r10, 7; or.b32 %r256, %r254, %r255; mul.wide.u32 %rd294, %r256, 4; add.s64 %rd295, %rd293, %rd294; ld.f32 %f121, [%rd295+96]; bra.uni $L__BB9_70; $L__BB9_69: ld.f32 %f121, [%rd30+4]; bra.uni $L__BB9_70; $L__BB9_95: shl.b64 %rd341, %rd53, 5; add.s64 %rd702, %rd51, %rd341; $L__BB9_96: setp.eq.s64 %p63, %rd702, 0; @%p63 bra $L__BB9_104; add.s64 %rd56, %rd702, 8; ld.u64 %rd57, [%rd702+8]; setp.eq.s64 %p64, %rd57, 0; @%p64 bra $L__BB9_103; add.s64 %rd58, %rd49, %rd57; and.b32 %r330, %r16, -4096; mov.u64 %rd342, 1; st.v2.u32 [%rd42+24], {%r18, %r330}; and.b32 %r331, %r10, -4096; st.u32 [%rd42+32], %r331; st.u64 [%rd42+64], %rd58; shl.b32 %r332, %r1, 3; and.b32 %r333, %r332, 31744; and.b32 %r334, %r16, 3968; shr.u32 %r335, %r334, 2; or.b32 %r336, %r335, %r333; bfe.u32 %r337, %r10, 7, 5; or.b32 %r338, %r336, %r337; shr.u32 %r339, %r336, 6; mul.wide.u32 %rd343, %r339, 8; add.s64 %rd344, %rd58, %rd343; and.b32 %r340, %r338, 63; shl.b64 %rd345, %rd342, %r340; ld.u64 %rd346, [%rd344+4128]; and.b64 %rd347, %rd346, %rd345; setp.eq.s64 %p65, %rd347, 0; mul.wide.u32 %rd348, %r338, 8; add.s64 %rd349, %rd58, %rd348; add.s64 %rd59, %rd349, 8256; @%p65 bra $L__BB9_102; ld.u64 %rd350, [%rd59]; add.s64 %rd60, %rd58, %rd350; st.u32 [%rd42+12], %r17; and.b32 %r341, %r16, -128; st.u32 [%rd42+16], %r341; and.b32 %r342, %r10, -128; st.u32 [%rd42+20], %r342; st.u64 [%rd42+56], %rd60; shl.b32 %r343, %r1, 5; and.b32 %r344, %r343, 3840; shl.b32 %r345, %r16, 1; and.b32 %r346, %r345, 240; or.b32 %r347, %r346, %r344; bfe.u32 %r348, %r10, 3, 4; or.b32 %r349, %r347, %r348; shr.u32 %r350, %r347, 6; mul.wide.u32 %rd351, %r350, 8; add.s64 %rd352, %rd60, %rd351; and.b32 %r351, %r349, 63; shl.b64 %rd354, %rd342, %r351; ld.u64 %rd355, [%rd352+544]; and.b64 %rd356, %rd355, %rd354; setp.eq.s64 %p66, %rd356, 0; mul.wide.u32 %rd357, %r349, 8; add.s64 %rd358, %rd60, %rd357; add.s64 %rd61, %rd358, 1088; @%p66 bra $L__BB9_101; ld.u64 %rd359, [%rd61]; add.s64 %rd360, %rd60, %rd359; and.b32 %r352, %r16, -8; st.v2.u32 [%rd42], {%r4, %r352}; and.b32 %r353, %r10, -8; st.u32 [%rd42+8], %r353; st.u64 [%rd42+48], %rd360; shl.b32 %r354, %r1, 6; and.b32 %r355, %r354, 448; shl.b32 %r356, %r16, 3; and.b32 %r357, %r356, 56; or.b32 %r358, %r357, %r355; and.b32 %r359, %r10, 7; or.b32 %r360, %r358, %r359; mul.wide.u32 %rd361, %r360, 4; add.s64 %rd362, %rd360, %rd361; ld.f32 %f122, [%rd362+96]; bra.uni $L__BB9_105; $L__BB9_104: ld.f32 %f122, [%rd50+4]; bra.uni $L__BB9_105; $L__BB9_130: shl.b64 %rd408, %rd72, 5; add.s64 %rd703, %rd70, %rd408; $L__BB9_131: setp.eq.s64 %p85, %rd703, 0; @%p85 bra $L__BB9_139; add.s64 %rd75, %rd703, 8; ld.u64 %rd76, [%rd703+8]; setp.eq.s64 %p86, %rd76, 0; @%p86 bra $L__BB9_138; add.s64 %rd77, %rd69, %rd76; and.b32 %r431, %r16, -4096; mov.u64 %rd409, 1; st.v2.u32 [%rd62+24], {%r23, %r431}; and.b32 %r432, %r3, -4096; st.u32 [%rd62+32], %r432; st.u64 [%rd62+64], %rd77; shl.b32 %r433, %r1, 3; and.b32 %r434, %r433, 31744; and.b32 %r435, %r16, 3968; shr.u32 %r436, %r435, 2; or.b32 %r437, %r436, %r434; bfe.u32 %r438, %r3, 7, 5; or.b32 %r439, %r437, %r438; shr.u32 %r440, %r437, 6; mul.wide.u32 %rd410, %r440, 8; add.s64 %rd411, %rd77, %rd410; and.b32 %r441, %r439, 63; shl.b64 %rd412, %rd409, %r441; ld.u64 %rd413, [%rd411+4128]; and.b64 %rd414, %rd413, %rd412; setp.eq.s64 %p87, %rd414, 0; mul.wide.u32 %rd415, %r439, 8; add.s64 %rd416, %rd77, %rd415; add.s64 %rd78, %rd416, 8256; @%p87 bra $L__BB9_137; ld.u64 %rd417, [%rd78]; add.s64 %rd79, %rd77, %rd417; st.u32 [%rd62+12], %r22; and.b32 %r442, %r16, -128; st.u32 [%rd62+16], %r442; and.b32 %r443, %r3, -128; st.u32 [%rd62+20], %r443; st.u64 [%rd62+56], %rd79; shl.b32 %r444, %r1, 5; and.b32 %r445, %r444, 3840; shl.b32 %r446, %r16, 1; and.b32 %r447, %r446, 240; or.b32 %r448, %r447, %r445; bfe.u32 %r449, %r3, 3, 4; or.b32 %r450, %r448, %r449; shr.u32 %r451, %r448, 6; mul.wide.u32 %rd418, %r451, 8; add.s64 %rd419, %rd79, %rd418; and.b32 %r452, %r450, 63; shl.b64 %rd421, %rd409, %r452; ld.u64 %rd422, [%rd419+544]; and.b64 %rd423, %rd422, %rd421; setp.eq.s64 %p88, %rd423, 0; mul.wide.u32 %rd424, %r450, 8; add.s64 %rd425, %rd79, %rd424; add.s64 %rd80, %rd425, 1088; @%p88 bra $L__BB9_136; ld.u64 %rd426, [%rd80]; add.s64 %rd427, %rd79, %rd426; and.b32 %r453, %r16, -8; st.v2.u32 [%rd62], {%r4, %r453}; and.b32 %r454, %r3, -8; st.u32 [%rd62+8], %r454; st.u64 [%rd62+48], %rd427; and.b32 %r455, %r1, 7; shl.b32 %r456, %r16, 3; and.b32 %r457, %r456, 56; and.b32 %r458, %r3, 7; bfi.b32 %r459, %r455, %r458, 6, 3; or.b32 %r460, %r459, %r457; mul.wide.u32 %rd428, %r460, 4; add.s64 %rd429, %rd427, %rd428; ld.f32 %f123, [%rd429+96]; bra.uni $L__BB9_140; $L__BB9_139: ld.f32 %f123, [%rd69+28]; bra.uni $L__BB9_140; $L__BB9_165: shl.b64 %rd475, %rd92, 5; add.s64 %rd704, %rd90, %rd475; $L__BB9_166: setp.eq.s64 %p107, %rd704, 0; @%p107 bra $L__BB9_174; add.s64 %rd95, %rd704, 8; ld.u64 %rd96, [%rd704+8]; setp.eq.s64 %p108, %rd96, 0; @%p108 bra $L__BB9_173; add.s64 %rd97, %rd88, %rd96; and.b32 %r531, %r16, -4096; mov.u64 %rd476, 1; st.v2.u32 [%rd81+24], {%r30, %r531}; and.b32 %r532, %r3, -4096; st.u32 [%rd81+32], %r532; st.u64 [%rd81+64], %rd97; shl.b32 %r533, %r27, 3; and.b32 %r534, %r533, 31744; and.b32 %r535, %r16, 3968; shr.u32 %r536, %r535, 2; or.b32 %r537, %r536, %r534; bfe.u32 %r538, %r3, 7, 5; or.b32 %r539, %r537, %r538; shr.u32 %r540, %r537, 6; mul.wide.u32 %rd477, %r540, 8; add.s64 %rd478, %rd97, %rd477; and.b32 %r541, %r539, 63; shl.b64 %rd479, %rd476, %r541; ld.u64 %rd480, [%rd478+4128]; and.b64 %rd481, %rd480, %rd479; setp.eq.s64 %p109, %rd481, 0; mul.wide.u32 %rd482, %r539, 8; add.s64 %rd483, %rd97, %rd482; add.s64 %rd98, %rd483, 8256; @%p109 bra $L__BB9_172; ld.u64 %rd484, [%rd98]; add.s64 %rd99, %rd97, %rd484; st.u32 [%rd81+12], %r29; and.b32 %r542, %r16, -128; st.u32 [%rd81+16], %r542; and.b32 %r543, %r3, -128; st.u32 [%rd81+20], %r543; st.u64 [%rd81+56], %rd99; shl.b32 %r544, %r27, 5; and.b32 %r545, %r544, 3840; shl.b32 %r546, %r16, 1; and.b32 %r547, %r546, 240; or.b32 %r548, %r547, %r545; bfe.u32 %r549, %r3, 3, 4; or.b32 %r550, %r548, %r549; shr.u32 %r551, %r548, 6; mul.wide.u32 %rd485, %r551, 8; add.s64 %rd486, %rd99, %rd485; and.b32 %r552, %r550, 63; shl.b64 %rd488, %rd476, %r552; ld.u64 %rd489, [%rd486+544]; and.b64 %rd490, %rd489, %rd488; setp.eq.s64 %p110, %rd490, 0; mul.wide.u32 %rd491, %r550, 8; add.s64 %rd492, %rd99, %rd491; add.s64 %rd100, %rd492, 1088; @%p110 bra $L__BB9_171; ld.u64 %rd493, [%rd100]; add.s64 %rd494, %rd99, %rd493; and.b32 %r553, %r16, -8; st.v2.u32 [%rd81], {%r28, %r553}; and.b32 %r554, %r3, -8; st.u32 [%rd81+8], %r554; st.u64 [%rd81+48], %rd494; and.b32 %r555, %r27, 7; shl.b32 %r556, %r16, 3; and.b32 %r557, %r556, 56; and.b32 %r558, %r3, 7; bfi.b32 %r559, %r555, %r558, 6, 3; or.b32 %r560, %r559, %r557; mul.wide.u32 %rd495, %r560, 4; add.s64 %rd496, %rd494, %rd495; ld.f32 %f124, [%rd496+96]; bra.uni $L__BB9_175; $L__BB9_174: ld.f32 %f124, [%rd89+4]; bra.uni $L__BB9_175; $L__BB9_200: shl.b64 %rd542, %rd111, 5; add.s64 %rd705, %rd109, %rd542; $L__BB9_201: setp.eq.s64 %p129, %rd705, 0; @%p129 bra $L__BB9_209; add.s64 %rd114, %rd705, 8; ld.u64 %rd115, [%rd705+8]; setp.eq.s64 %p130, %rd115, 0; @%p130 bra $L__BB9_208; add.s64 %rd116, %rd108, %rd115; and.b32 %r634, %r16, -4096; mov.u64 %rd543, 1; st.v2.u32 [%rd101+24], {%r35, %r634}; and.b32 %r635, %r10, -4096; st.u32 [%rd101+32], %r635; st.u64 [%rd101+64], %rd116; shl.b32 %r636, %r27, 3; and.b32 %r637, %r636, 31744; and.b32 %r638, %r16, 3968; shr.u32 %r639, %r638, 2; or.b32 %r640, %r639, %r637; bfe.u32 %r641, %r10, 7, 5; or.b32 %r642, %r640, %r641; shr.u32 %r643, %r640, 6; mul.wide.u32 %rd544, %r643, 8; add.s64 %rd545, %rd116, %rd544; and.b32 %r644, %r642, 63; shl.b64 %rd546, %rd543, %r644; ld.u64 %rd547, [%rd545+4128]; and.b64 %rd548, %rd547, %rd546; setp.eq.s64 %p131, %rd548, 0; mul.wide.u32 %rd549, %r642, 8; add.s64 %rd550, %rd116, %rd549; add.s64 %rd117, %rd550, 8256; @%p131 bra $L__BB9_207; ld.u64 %rd551, [%rd117]; add.s64 %rd118, %rd116, %rd551; st.u32 [%rd101+12], %r34; and.b32 %r645, %r16, -128; st.u32 [%rd101+16], %r645; and.b32 %r646, %r10, -128; st.u32 [%rd101+20], %r646; st.u64 [%rd101+56], %rd118; shl.b32 %r647, %r27, 5; and.b32 %r648, %r647, 3840; shl.b32 %r649, %r16, 1; and.b32 %r650, %r649, 240; or.b32 %r651, %r650, %r648; bfe.u32 %r652, %r10, 3, 4; or.b32 %r653, %r651, %r652; shr.u32 %r654, %r651, 6; mul.wide.u32 %rd552, %r654, 8; add.s64 %rd553, %rd118, %rd552; and.b32 %r655, %r653, 63; shl.b64 %rd555, %rd543, %r655; ld.u64 %rd556, [%rd553+544]; and.b64 %rd557, %rd556, %rd555; setp.eq.s64 %p132, %rd557, 0; mul.wide.u32 %rd558, %r653, 8; add.s64 %rd559, %rd118, %rd558; add.s64 %rd119, %rd559, 1088; @%p132 bra $L__BB9_206; ld.u64 %rd560, [%rd119]; add.s64 %rd561, %rd118, %rd560; and.b32 %r656, %r16, -8; st.v2.u32 [%rd101], {%r28, %r656}; and.b32 %r657, %r10, -8; st.u32 [%rd101+8], %r657; st.u64 [%rd101+48], %rd561; shl.b32 %r658, %r27, 6; and.b32 %r659, %r658, 448; shl.b32 %r660, %r16, 3; and.b32 %r661, %r660, 56; or.b32 %r662, %r661, %r659; and.b32 %r663, %r10, 7; or.b32 %r664, %r662, %r663; mul.wide.u32 %rd562, %r664, 4; add.s64 %rd563, %rd561, %rd562; ld.f32 %f125, [%rd563+96]; bra.uni $L__BB9_210; $L__BB9_209: ld.f32 %f125, [%rd108+28]; bra.uni $L__BB9_210; $L__BB9_235: shl.b64 %rd610, %rd131, 5; add.s64 %rd706, %rd129, %rd610; $L__BB9_236: setp.eq.s64 %p151, %rd706, 0; @%p151 bra $L__BB9_244; add.s64 %rd134, %rd706, 8; ld.u64 %rd135, [%rd706+8]; setp.eq.s64 %p152, %rd135, 0; @%p152 bra $L__BB9_243; add.s64 %rd136, %rd127, %rd135; and.b32 %r738, %r2, -4096; mov.u64 %rd611, 1; st.v2.u32 [%rd120+24], {%r40, %r738}; and.b32 %r739, %r10, -4096; st.u32 [%rd120+32], %r739; st.u64 [%rd120+64], %rd136; shl.b32 %r740, %r27, 3; and.b32 %r741, %r740, 31744; and.b32 %r742, %r2, 3968; shr.u32 %r743, %r742, 2; or.b32 %r744, %r741, %r743; bfe.u32 %r745, %r10, 7, 5; or.b32 %r746, %r745, %r744; shr.u32 %r747, %r744, 6; mul.wide.u32 %rd612, %r747, 8; add.s64 %rd613, %rd136, %rd612; and.b32 %r748, %r746, 63; shl.b64 %rd614, %rd611, %r748; ld.u64 %rd615, [%rd613+4128]; and.b64 %rd616, %rd615, %rd614; setp.eq.s64 %p153, %rd616, 0; mul.wide.u32 %rd617, %r746, 8; add.s64 %rd618, %rd136, %rd617; add.s64 %rd137, %rd618, 8256; @%p153 bra $L__BB9_242; ld.u64 %rd619, [%rd137]; add.s64 %rd138, %rd136, %rd619; st.u32 [%rd120+12], %r39; and.b32 %r749, %r2, -128; st.u32 [%rd120+16], %r749; and.b32 %r750, %r10, -128; st.u32 [%rd120+20], %r750; st.u64 [%rd120+56], %rd138; shl.b32 %r751, %r27, 5; and.b32 %r752, %r751, 3840; shl.b32 %r753, %r2, 1; and.b32 %r754, %r753, 240; or.b32 %r755, %r752, %r754; bfe.u32 %r756, %r10, 3, 4; or.b32 %r757, %r756, %r755; shr.u32 %r758, %r755, 6; mul.wide.u32 %rd620, %r758, 8; add.s64 %rd621, %rd138, %rd620; and.b32 %r759, %r757, 63; shl.b64 %rd623, %rd611, %r759; ld.u64 %rd624, [%rd621+544]; and.b64 %rd625, %rd624, %rd623; setp.eq.s64 %p154, %rd625, 0; mul.wide.u32 %rd626, %r757, 8; add.s64 %rd627, %rd138, %rd626; add.s64 %rd139, %rd627, 1088; @%p154 bra $L__BB9_241; ld.u64 %rd628, [%rd139]; add.s64 %rd629, %rd138, %rd628; and.b32 %r760, %r2, -8; st.v2.u32 [%rd120], {%r28, %r760}; and.b32 %r761, %r10, -8; st.u32 [%rd120+8], %r761; st.u64 [%rd120+48], %rd629; shl.b32 %r762, %r27, 6; and.b32 %r763, %r762, 448; shl.b32 %r764, %r2, 3; and.b32 %r765, %r764, 56; or.b32 %r766, %r763, %r765; and.b32 %r767, %r10, 7; or.b32 %r768, %r766, %r767; mul.wide.u32 %rd630, %r768, 4; add.s64 %rd631, %rd629, %rd630; ld.f32 %f126, [%rd631+96]; bra.uni $L__BB9_245; $L__BB9_244: ld.f32 %f126, [%rd128+4]; bra.uni $L__BB9_245; $L__BB9_33: ld.f32 %f120, [%rd16+12]; bra.uni $L__BB9_35; $L__BB9_68: ld.f32 %f121, [%rd36+12]; bra.uni $L__BB9_70; $L__BB9_103: ld.f32 %f122, [%rd56+12]; bra.uni $L__BB9_105; $L__BB9_138: ld.f32 %f123, [%rd75+12]; bra.uni $L__BB9_140; $L__BB9_173: ld.f32 %f124, [%rd95+12]; bra.uni $L__BB9_175; $L__BB9_208: ld.f32 %f125, [%rd114+12]; bra.uni $L__BB9_210; $L__BB9_243: ld.f32 %f126, [%rd134+12]; bra.uni $L__BB9_245; $L__BB9_32: ld.f32 %f120, [%rd19]; bra.uni $L__BB9_35; $L__BB9_67: ld.f32 %f121, [%rd39]; bra.uni $L__BB9_70; $L__BB9_102: ld.f32 %f122, [%rd59]; bra.uni $L__BB9_105; $L__BB9_137: ld.f32 %f123, [%rd78]; bra.uni $L__BB9_140; $L__BB9_172: ld.f32 %f124, [%rd98]; bra.uni $L__BB9_175; $L__BB9_207: ld.f32 %f125, [%rd117]; bra.uni $L__BB9_210; $L__BB9_242: ld.f32 %f126, [%rd137]; bra.uni $L__BB9_245; $L__BB9_31: ld.f32 %f120, [%rd21]; bra.uni $L__BB9_35; $L__BB9_66: ld.f32 %f121, [%rd41]; bra.uni $L__BB9_70; $L__BB9_101: ld.f32 %f122, [%rd61]; bra.uni $L__BB9_105; $L__BB9_136: ld.f32 %f123, [%rd80]; bra.uni $L__BB9_140; $L__BB9_171: ld.f32 %f124, [%rd100]; bra.uni $L__BB9_175; $L__BB9_206: ld.f32 %f125, [%rd119]; bra.uni $L__BB9_210; $L__BB9_241: ld.f32 %f126, [%rd139]; bra.uni $L__BB9_245; } .func (.param .align 4 .b8 func_retval0[12]) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_1[12] ) { .reg .pred %p<174>; .reg .b16 %rs<106>; .reg .f32 %f<364>; .reg .b32 %r<881>; .reg .b64 %rd<703>; ld.param.u64 %rd155, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_0]; cvta.to.local.u64 %rd1, %rd155; ld.param.f32 %f301, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_1+8]; ld.param.f32 %f302, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_1+4]; ld.param.f32 %f303, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Vec3IfEELi0ELi1ELi2EEELi1ELb0EEclIfS2_EES3_T0_IJT_EE_param_1]; cvt.rmi.ftz.f32.f32 %f304, %f303; cvt.rmi.ftz.f32.f32 %f305, %f302; cvt.rmi.ftz.f32.f32 %f306, %f301; sub.ftz.f32 %f1, %f303, %f304; sub.ftz.f32 %f2, %f302, %f305; sub.ftz.f32 %f3, %f301, %f306; cvt.rzi.ftz.s32.f32 %r1, %f304; cvt.rzi.ftz.s32.f32 %r2, %f305; cvt.rzi.ftz.s32.f32 %r3, %f306; and.b32 %r4, %r1, -8; ld.local.u64 %rd2, [%rd1]; ld.u32 %r49, [%rd2]; setp.ne.s32 %p1, %r4, %r49; mov.u16 %rs85, 0; @%p1 bra $L__BB10_3; and.b32 %r50, %r2, -8; ld.u32 %r51, [%rd2+4]; setp.ne.s32 %p2, %r50, %r51; @%p2 bra $L__BB10_3; and.b32 %r52, %r3, -8; ld.u32 %r53, [%rd2+8]; setp.eq.s32 %p3, %r52, %r53; selp.u16 %rs85, 1, 0, %p3; $L__BB10_3: setp.eq.s16 %p4, %rs85, 0; @%p4 bra $L__BB10_5; shl.b32 %r54, %r1, 6; and.b32 %r55, %r54, 448; shl.b32 %r56, %r2, 3; and.b32 %r57, %r56, 56; or.b32 %r58, %r57, %r55; and.b32 %r59, %r3, 7; or.b32 %r60, %r58, %r59; ld.u64 %rd156, [%rd2+48]; mul.wide.u32 %rd157, %r60, 12; add.s64 %rd158, %rd156, %rd157; ld.f32 %f342, [%rd158+128]; ld.f32 %f341, [%rd158+132]; ld.f32 %f340, [%rd158+136]; $L__BB10_35: add.s32 %r10, %r3, 1; ld.local.u64 %rd22, [%rd1]; ld.u32 %r153, [%rd22]; setp.ne.s32 %p23, %r4, %r153; mov.u16 %rs88, 0; @%p23 bra $L__BB10_38; and.b32 %r154, %r2, -8; ld.u32 %r155, [%rd22+4]; setp.ne.s32 %p24, %r154, %r155; @%p24 bra $L__BB10_38; and.b32 %r156, %r10, -8; ld.u32 %r157, [%rd22+8]; setp.eq.s32 %p25, %r156, %r157; selp.u16 %rs88, 1, 0, %p25; $L__BB10_38: setp.eq.s16 %p26, %rs88, 0; @%p26 bra $L__BB10_40; shl.b32 %r158, %r1, 6; and.b32 %r159, %r158, 448; shl.b32 %r160, %r2, 3; and.b32 %r161, %r160, 56; or.b32 %r162, %r161, %r159; and.b32 %r163, %r10, 7; or.b32 %r164, %r162, %r163; ld.u64 %rd223, [%rd22+48]; mul.wide.u32 %rd224, %r164, 12; add.s64 %rd225, %rd223, %rd224; ld.f32 %f345, [%rd225+128]; ld.f32 %f344, [%rd225+132]; ld.f32 %f343, [%rd225+136]; $L__BB10_70: sub.ftz.f32 %f307, %f345, %f342; sub.ftz.f32 %f308, %f344, %f341; sub.ftz.f32 %f309, %f343, %f340; fma.rn.ftz.f32 %f76, %f3, %f307, %f342; fma.rn.ftz.f32 %f77, %f3, %f308, %f341; fma.rn.ftz.f32 %f78, %f3, %f309, %f340; add.s32 %r16, %r2, 1; ld.local.u64 %rd41, [%rd1]; ld.u32 %r257, [%rd41]; setp.ne.s32 %p45, %r4, %r257; mov.u16 %rs91, 0; @%p45 bra $L__BB10_73; and.b32 %r258, %r16, -8; ld.u32 %r259, [%rd41+4]; setp.ne.s32 %p46, %r258, %r259; @%p46 bra $L__BB10_73; and.b32 %r260, %r10, -8; ld.u32 %r261, [%rd41+8]; setp.eq.s32 %p47, %r260, %r261; selp.u16 %rs91, 1, 0, %p47; $L__BB10_73: setp.eq.s16 %p48, %rs91, 0; @%p48 bra $L__BB10_75; shl.b32 %r262, %r1, 6; and.b32 %r263, %r262, 448; shl.b32 %r264, %r16, 3; and.b32 %r265, %r264, 56; or.b32 %r266, %r265, %r263; and.b32 %r267, %r10, 7; or.b32 %r268, %r266, %r267; ld.u64 %rd291, [%rd41+48]; mul.wide.u32 %rd292, %r268, 12; add.s64 %rd293, %rd291, %rd292; ld.f32 %f348, [%rd293+128]; ld.f32 %f347, [%rd293+132]; ld.f32 %f346, [%rd293+136]; $L__BB10_105: ld.local.u64 %rd60, [%rd1]; ld.u32 %r361, [%rd60]; setp.ne.s32 %p67, %r4, %r361; mov.u16 %rs94, 0; @%p67 bra $L__BB10_108; and.b32 %r362, %r16, -8; ld.u32 %r363, [%rd60+4]; setp.ne.s32 %p68, %r362, %r363; @%p68 bra $L__BB10_108; and.b32 %r364, %r3, -8; ld.u32 %r365, [%rd60+8]; setp.eq.s32 %p69, %r364, %r365; selp.u16 %rs94, 1, 0, %p69; $L__BB10_108: setp.eq.s16 %p70, %rs94, 0; @%p70 bra $L__BB10_110; and.b32 %r366, %r1, 7; shl.b32 %r367, %r16, 3; and.b32 %r368, %r367, 56; and.b32 %r369, %r3, 7; bfi.b32 %r370, %r366, %r369, 6, 3; or.b32 %r371, %r370, %r368; ld.u64 %rd358, [%rd60+48]; mul.wide.u32 %rd359, %r371, 12; add.s64 %rd360, %rd358, %rd359; ld.f32 %f351, [%rd360+128]; ld.f32 %f350, [%rd360+132]; ld.f32 %f349, [%rd360+136]; $L__BB10_140: sub.ftz.f32 %f310, %f348, %f351; sub.ftz.f32 %f311, %f347, %f350; sub.ftz.f32 %f312, %f346, %f349; fma.rn.ftz.f32 %f313, %f3, %f310, %f351; fma.rn.ftz.f32 %f314, %f3, %f311, %f350; fma.rn.ftz.f32 %f315, %f3, %f312, %f349; sub.ftz.f32 %f316, %f313, %f76; sub.ftz.f32 %f317, %f314, %f77; sub.ftz.f32 %f318, %f315, %f78; fma.rn.ftz.f32 %f151, %f2, %f316, %f76; fma.rn.ftz.f32 %f152, %f2, %f317, %f77; fma.rn.ftz.f32 %f153, %f2, %f318, %f78; add.s32 %r27, %r1, 1; and.b32 %r28, %r27, -8; ld.local.u64 %rd79, [%rd1]; ld.u32 %r461, [%rd79]; setp.ne.s32 %p89, %r28, %r461; mov.u16 %rs97, 0; @%p89 bra $L__BB10_143; and.b32 %r462, %r16, -8; ld.u32 %r463, [%rd79+4]; setp.ne.s32 %p90, %r462, %r463; @%p90 bra $L__BB10_143; and.b32 %r464, %r3, -8; ld.u32 %r465, [%rd79+8]; setp.eq.s32 %p91, %r464, %r465; selp.u16 %rs97, 1, 0, %p91; $L__BB10_143: setp.eq.s16 %p92, %rs97, 0; @%p92 bra $L__BB10_145; and.b32 %r466, %r27, 7; shl.b32 %r467, %r16, 3; and.b32 %r468, %r467, 56; and.b32 %r469, %r3, 7; bfi.b32 %r470, %r466, %r469, 6, 3; or.b32 %r471, %r470, %r468; ld.u64 %rd425, [%rd79+48]; mul.wide.u32 %rd426, %r471, 12; add.s64 %rd427, %rd425, %rd426; ld.f32 %f354, [%rd427+128]; ld.f32 %f353, [%rd427+132]; ld.f32 %f352, [%rd427+136]; $L__BB10_175: ld.local.u64 %rd98, [%rd1]; ld.u32 %r561, [%rd98]; setp.ne.s32 %p111, %r28, %r561; mov.u16 %rs100, 0; @%p111 bra $L__BB10_178; and.b32 %r562, %r16, -8; ld.u32 %r563, [%rd98+4]; setp.ne.s32 %p112, %r562, %r563; @%p112 bra $L__BB10_178; and.b32 %r564, %r10, -8; ld.u32 %r565, [%rd98+8]; setp.eq.s32 %p113, %r564, %r565; selp.u16 %rs100, 1, 0, %p113; $L__BB10_178: setp.eq.s16 %p114, %rs100, 0; @%p114 bra $L__BB10_180; shl.b32 %r566, %r27, 6; and.b32 %r567, %r566, 448; shl.b32 %r568, %r16, 3; and.b32 %r569, %r568, 56; or.b32 %r570, %r569, %r567; and.b32 %r571, %r10, 7; or.b32 %r572, %r570, %r571; ld.u64 %rd492, [%rd98+48]; mul.wide.u32 %rd493, %r572, 12; add.s64 %rd494, %rd492, %rd493; ld.f32 %f357, [%rd494+128]; ld.f32 %f356, [%rd494+132]; ld.f32 %f355, [%rd494+136]; $L__BB10_210: sub.ftz.f32 %f319, %f357, %f354; sub.ftz.f32 %f320, %f356, %f353; sub.ftz.f32 %f321, %f355, %f352; fma.rn.ftz.f32 %f226, %f3, %f319, %f354; fma.rn.ftz.f32 %f227, %f3, %f320, %f353; fma.rn.ftz.f32 %f228, %f3, %f321, %f352; ld.local.u64 %rd117, [%rd1]; ld.u32 %r665, [%rd117]; setp.ne.s32 %p133, %r28, %r665; mov.u16 %rs103, 0; @%p133 bra $L__BB10_213; and.b32 %r666, %r2, -8; ld.u32 %r667, [%rd117+4]; setp.ne.s32 %p134, %r666, %r667; @%p134 bra $L__BB10_213; and.b32 %r668, %r10, -8; ld.u32 %r669, [%rd117+8]; setp.eq.s32 %p135, %r668, %r669; selp.u16 %rs103, 1, 0, %p135; $L__BB10_213: setp.eq.s16 %p136, %rs103, 0; @%p136 bra $L__BB10_215; shl.b32 %r670, %r27, 6; and.b32 %r671, %r670, 448; shl.b32 %r672, %r2, 3; and.b32 %r673, %r672, 56; or.b32 %r674, %r671, %r673; and.b32 %r675, %r10, 7; or.b32 %r676, %r674, %r675; ld.u64 %rd559, [%rd117+48]; mul.wide.u32 %rd560, %r676, 12; add.s64 %rd561, %rd559, %rd560; ld.f32 %f360, [%rd561+128]; ld.f32 %f359, [%rd561+132]; ld.f32 %f358, [%rd561+136]; $L__BB10_245: ld.local.u64 %rd136, [%rd1]; ld.u32 %r769, [%rd136]; setp.ne.s32 %p155, %r28, %r769; @%p155 bra $L__BB10_248; and.b32 %r770, %r2, -8; ld.u32 %r771, [%rd136+4]; setp.ne.s32 %p156, %r770, %r771; @%p156 bra $L__BB10_248; and.b32 %r772, %r3, -8; ld.u32 %r773, [%rd136+8]; setp.eq.s32 %p157, %r772, %r773; @%p157 bra $L__BB10_276; bra.uni $L__BB10_248; $L__BB10_276: shl.b32 %r866, %r27, 6; and.b32 %r867, %r866, 448; shl.b32 %r868, %r2, 3; and.b32 %r869, %r868, 56; or.b32 %r870, %r867, %r869; and.b32 %r871, %r3, 7; or.b32 %r872, %r870, %r871; ld.u64 %rd692, [%rd136+48]; mul.wide.u32 %rd693, %r872, 12; add.s64 %rd694, %rd692, %rd693; ld.f32 %f363, [%rd694+128]; ld.f32 %f362, [%rd694+132]; ld.f32 %f361, [%rd694+136]; bra.uni $L__BB10_277; $L__BB10_248: and.b32 %r44, %r27, -128; ld.u32 %r774, [%rd136+12]; setp.ne.s32 %p158, %r44, %r774; @%p158 bra $L__BB10_251; and.b32 %r775, %r2, -128; ld.u32 %r776, [%rd136+16]; setp.ne.s32 %p159, %r775, %r776; @%p159 bra $L__BB10_251; and.b32 %r777, %r3, -128; ld.u32 %r778, [%rd136+20]; setp.eq.s32 %p160, %r777, %r778; @%p160 bra $L__BB10_273; bra.uni $L__BB10_251; $L__BB10_273: shl.b32 %r848, %r27, 5; and.b32 %r849, %r848, 3840; shl.b32 %r850, %r2, 1; and.b32 %r851, %r850, 240; or.b32 %r852, %r849, %r851; bfe.u32 %r853, %r3, 3, 4; or.b32 %r854, %r852, %r853; shr.u32 %r855, %r852, 6; ld.u64 %rd153, [%rd136+56]; mul.wide.u32 %rd680, %r855, 8; add.s64 %rd681, %rd153, %rd680; and.b32 %r856, %r854, 63; mov.u64 %rd682, 1; shl.b64 %rd683, %rd682, %r856; ld.u64 %rd684, [%rd681+544]; and.b64 %rd685, %rd684, %rd683; setp.eq.s64 %p173, %rd685, 0; mul.wide.u32 %rd686, %r854, 16; add.s64 %rd687, %rd153, %rd686; add.s64 %rd154, %rd687, 1088; @%p173 bra $L__BB10_275; ld.u64 %rd688, [%rd154]; add.s64 %rd689, %rd153, %rd688; and.b32 %r857, %r2, -8; st.v2.u32 [%rd136], {%r28, %r857}; and.b32 %r858, %r3, -8; st.u32 [%rd136+8], %r858; st.u64 [%rd136+48], %rd689; shl.b32 %r859, %r27, 6; and.b32 %r860, %r859, 448; shl.b32 %r861, %r2, 3; and.b32 %r862, %r861, 56; or.b32 %r863, %r860, %r862; and.b32 %r864, %r3, 7; or.b32 %r865, %r863, %r864; mul.wide.u32 %rd690, %r865, 12; add.s64 %rd691, %rd689, %rd690; ld.f32 %f363, [%rd691+128]; ld.f32 %f362, [%rd691+132]; ld.f32 %f361, [%rd691+136]; bra.uni $L__BB10_277; $L__BB10_251: and.b32 %r45, %r27, -4096; ld.u32 %r779, [%rd136+24]; setp.ne.s32 %p161, %r45, %r779; @%p161 bra $L__BB10_254; and.b32 %r780, %r2, -4096; ld.u32 %r781, [%rd136+28]; setp.ne.s32 %p162, %r780, %r781; @%p162 bra $L__BB10_254; and.b32 %r782, %r3, -4096; ld.u32 %r783, [%rd136+32]; setp.eq.s32 %p163, %r782, %r783; @%p163 bra $L__BB10_268; bra.uni $L__BB10_254; $L__BB10_268: shl.b32 %r819, %r27, 3; and.b32 %r820, %r819, 31744; and.b32 %r821, %r2, 3968; shr.u32 %r822, %r821, 2; or.b32 %r823, %r820, %r822; bfe.u32 %r824, %r3, 7, 5; or.b32 %r825, %r823, %r824; shr.u32 %r826, %r823, 6; ld.u64 %rd149, [%rd136+64]; mul.wide.u32 %rd659, %r826, 8; add.s64 %rd660, %rd149, %rd659; and.b32 %r827, %r825, 63; mov.u64 %rd661, 1; shl.b64 %rd662, %rd661, %r827; ld.u64 %rd663, [%rd660+4128]; and.b64 %rd664, %rd663, %rd662; setp.eq.s64 %p171, %rd664, 0; mul.wide.u32 %rd665, %r825, 16; add.s64 %rd666, %rd149, %rd665; add.s64 %rd150, %rd666, 8256; @%p171 bra $L__BB10_272; ld.u64 %rd667, [%rd150]; add.s64 %rd151, %rd149, %rd667; st.u32 [%rd136+12], %r44; and.b32 %r828, %r2, -128; st.u32 [%rd136+16], %r828; and.b32 %r829, %r3, -128; st.u32 [%rd136+20], %r829; st.u64 [%rd136+56], %rd151; shl.b32 %r830, %r27, 5; and.b32 %r831, %r830, 3840; shl.b32 %r832, %r2, 1; and.b32 %r833, %r832, 240; or.b32 %r834, %r831, %r833; bfe.u32 %r835, %r3, 3, 4; or.b32 %r836, %r834, %r835; shr.u32 %r837, %r834, 6; mul.wide.u32 %rd668, %r837, 8; add.s64 %rd669, %rd151, %rd668; and.b32 %r838, %r836, 63; shl.b64 %rd671, %rd661, %r838; ld.u64 %rd672, [%rd669+544]; and.b64 %rd673, %rd672, %rd671; setp.eq.s64 %p172, %rd673, 0; mul.wide.u32 %rd674, %r836, 16; add.s64 %rd675, %rd151, %rd674; add.s64 %rd152, %rd675, 1088; @%p172 bra $L__BB10_271; ld.u64 %rd676, [%rd152]; add.s64 %rd677, %rd151, %rd676; and.b32 %r839, %r2, -8; st.v2.u32 [%rd136], {%r28, %r839}; and.b32 %r840, %r3, -8; st.u32 [%rd136+8], %r840; st.u64 [%rd136+48], %rd677; shl.b32 %r841, %r27, 6; and.b32 %r842, %r841, 448; shl.b32 %r843, %r2, 3; and.b32 %r844, %r843, 56; or.b32 %r845, %r842, %r844; and.b32 %r846, %r3, 7; or.b32 %r847, %r845, %r846; mul.wide.u32 %rd678, %r847, 12; add.s64 %rd679, %rd677, %rd678; ld.f32 %f363, [%rd679+128]; ld.f32 %f362, [%rd679+132]; ld.f32 %f361, [%rd679+136]; bra.uni $L__BB10_277; $L__BB10_254: ld.u64 %rd137, [%rd136+40]; mov.u64 %rd702, 0; ld.u32 %r46, [%rd137+24]; setp.eq.s32 %p164, %r46, 0; @%p164 bra $L__BB10_259; add.s64 %rd138, %rd137, 96; shr.u32 %r785, %r27, 12; cvt.u64.u32 %rd628, %r785; shl.b64 %rd629, %rd628, 42; shr.u32 %r786, %r2, 12; mul.wide.u32 %rd630, %r786, 2097152; shr.u32 %r787, %r3, 12; cvt.u64.u32 %rd631, %r787; or.b64 %rd632, %rd630, %rd631; or.b64 %rd139, %rd632, %rd629; mov.u32 %r880, 0; $L__BB10_256: cvt.u64.u32 %rd140, %r880; mul.wide.u32 %rd633, %r880, 32; add.s64 %rd634, %rd138, %rd633; ld.u64 %rd635, [%rd634]; setp.eq.s64 %p165, %rd635, %rd139; add.s32 %r880, %r880, 1; @%p165 bra $L__BB10_258; setp.lt.u32 %p166, %r880, %r46; @%p166 bra $L__BB10_256; bra.uni $L__BB10_259; $L__BB10_5: and.b32 %r5, %r1, -128; ld.u32 %r61, [%rd2+12]; setp.ne.s32 %p5, %r5, %r61; mov.u16 %rs86, 0; @%p5 bra $L__BB10_8; and.b32 %r62, %r2, -128; ld.u32 %r63, [%rd2+16]; setp.ne.s32 %p6, %r62, %r63; @%p6 bra $L__BB10_8; and.b32 %r64, %r3, -128; ld.u32 %r65, [%rd2+20]; setp.eq.s32 %p7, %r64, %r65; selp.u16 %rs86, 1, 0, %p7; $L__BB10_8: setp.eq.s16 %p8, %rs86, 0; @%p8 bra $L__BB10_12; shl.b32 %r66, %r1, 5; and.b32 %r67, %r66, 3840; shl.b32 %r68, %r2, 1; and.b32 %r69, %r68, 240; or.b32 %r70, %r69, %r67; bfe.u32 %r71, %r3, 3, 4; or.b32 %r72, %r70, %r71; shr.u32 %r73, %r70, 6; ld.u64 %rd3, [%rd2+56]; mul.wide.u32 %rd159, %r73, 8; add.s64 %rd160, %rd3, %rd159; and.b32 %r74, %r72, 63; mov.u64 %rd161, 1; shl.b64 %rd162, %rd161, %r74; ld.u64 %rd163, [%rd160+544]; and.b64 %rd164, %rd163, %rd162; setp.eq.s64 %p9, %rd164, 0; mul.wide.u32 %rd165, %r72, 16; add.s64 %rd166, %rd3, %rd165; add.s64 %rd4, %rd166, 1088; @%p9 bra $L__BB10_11; ld.u64 %rd167, [%rd4]; add.s64 %rd168, %rd3, %rd167; and.b32 %r75, %r2, -8; st.v2.u32 [%rd2], {%r4, %r75}; and.b32 %r76, %r3, -8; st.u32 [%rd2+8], %r76; st.u64 [%rd2+48], %rd168; shl.b32 %r77, %r1, 6; and.b32 %r78, %r77, 448; shl.b32 %r79, %r2, 3; and.b32 %r80, %r79, 56; or.b32 %r81, %r80, %r78; and.b32 %r82, %r3, 7; or.b32 %r83, %r81, %r82; mul.wide.u32 %rd169, %r83, 12; add.s64 %rd170, %rd168, %rd169; ld.f32 %f342, [%rd170+128]; ld.f32 %f341, [%rd170+132]; ld.f32 %f340, [%rd170+136]; bra.uni $L__BB10_35; $L__BB10_40: and.b32 %r11, %r1, -128; ld.u32 %r165, [%rd22+12]; setp.ne.s32 %p27, %r11, %r165; mov.u16 %rs89, 0; @%p27 bra $L__BB10_43; and.b32 %r166, %r2, -128; ld.u32 %r167, [%rd22+16]; setp.ne.s32 %p28, %r166, %r167; @%p28 bra $L__BB10_43; and.b32 %r168, %r10, -128; ld.u32 %r169, [%rd22+20]; setp.eq.s32 %p29, %r168, %r169; selp.u16 %rs89, 1, 0, %p29; $L__BB10_43: setp.eq.s16 %p30, %rs89, 0; @%p30 bra $L__BB10_47; shl.b32 %r170, %r1, 5; and.b32 %r171, %r170, 3840; shl.b32 %r172, %r2, 1; and.b32 %r173, %r172, 240; or.b32 %r174, %r173, %r171; bfe.u32 %r175, %r10, 3, 4; or.b32 %r176, %r175, %r174; shr.u32 %r177, %r174, 6; ld.u64 %rd23, [%rd22+56]; mul.wide.u32 %rd226, %r177, 8; add.s64 %rd227, %rd23, %rd226; and.b32 %r178, %r176, 63; mov.u64 %rd228, 1; shl.b64 %rd229, %rd228, %r178; ld.u64 %rd230, [%rd227+544]; and.b64 %rd231, %rd230, %rd229; setp.eq.s64 %p31, %rd231, 0; mul.wide.u32 %rd232, %r176, 16; add.s64 %rd233, %rd23, %rd232; add.s64 %rd24, %rd233, 1088; @%p31 bra $L__BB10_46; ld.u64 %rd234, [%rd24]; add.s64 %rd235, %rd23, %rd234; and.b32 %r179, %r2, -8; st.v2.u32 [%rd22], {%r4, %r179}; and.b32 %r180, %r10, -8; st.u32 [%rd22+8], %r180; st.u64 [%rd22+48], %rd235; shl.b32 %r181, %r1, 6; and.b32 %r182, %r181, 448; shl.b32 %r183, %r2, 3; and.b32 %r184, %r183, 56; or.b32 %r185, %r184, %r182; and.b32 %r186, %r10, 7; or.b32 %r187, %r185, %r186; mul.wide.u32 %rd236, %r187, 12; add.s64 %rd237, %rd235, %rd236; ld.f32 %f345, [%rd237+128]; ld.f32 %f344, [%rd237+132]; ld.f32 %f343, [%rd237+136]; bra.uni $L__BB10_70; $L__BB10_75: and.b32 %r17, %r1, -128; ld.u32 %r269, [%rd41+12]; setp.ne.s32 %p49, %r17, %r269; mov.u16 %rs92, 0; @%p49 bra $L__BB10_78; and.b32 %r270, %r16, -128; ld.u32 %r271, [%rd41+16]; setp.ne.s32 %p50, %r270, %r271; @%p50 bra $L__BB10_78; and.b32 %r272, %r10, -128; ld.u32 %r273, [%rd41+20]; setp.eq.s32 %p51, %r272, %r273; selp.u16 %rs92, 1, 0, %p51; $L__BB10_78: setp.eq.s16 %p52, %rs92, 0; @%p52 bra $L__BB10_82; shl.b32 %r274, %r1, 5; and.b32 %r275, %r274, 3840; shl.b32 %r276, %r16, 1; and.b32 %r277, %r276, 240; or.b32 %r278, %r277, %r275; bfe.u32 %r279, %r10, 3, 4; or.b32 %r280, %r278, %r279; shr.u32 %r281, %r278, 6; ld.u64 %rd42, [%rd41+56]; mul.wide.u32 %rd294, %r281, 8; add.s64 %rd295, %rd42, %rd294; and.b32 %r282, %r280, 63; mov.u64 %rd296, 1; shl.b64 %rd297, %rd296, %r282; ld.u64 %rd298, [%rd295+544]; and.b64 %rd299, %rd298, %rd297; setp.eq.s64 %p53, %rd299, 0; mul.wide.u32 %rd300, %r280, 16; add.s64 %rd301, %rd42, %rd300; add.s64 %rd43, %rd301, 1088; @%p53 bra $L__BB10_81; ld.u64 %rd302, [%rd43]; add.s64 %rd303, %rd42, %rd302; and.b32 %r283, %r16, -8; st.v2.u32 [%rd41], {%r4, %r283}; and.b32 %r284, %r10, -8; st.u32 [%rd41+8], %r284; st.u64 [%rd41+48], %rd303; shl.b32 %r285, %r1, 6; and.b32 %r286, %r285, 448; shl.b32 %r287, %r16, 3; and.b32 %r288, %r287, 56; or.b32 %r289, %r288, %r286; and.b32 %r290, %r10, 7; or.b32 %r291, %r289, %r290; mul.wide.u32 %rd304, %r291, 12; add.s64 %rd305, %rd303, %rd304; ld.f32 %f348, [%rd305+128]; ld.f32 %f347, [%rd305+132]; ld.f32 %f346, [%rd305+136]; bra.uni $L__BB10_105; $L__BB10_110: and.b32 %r22, %r1, -128; ld.u32 %r372, [%rd60+12]; setp.ne.s32 %p71, %r22, %r372; mov.u16 %rs95, 0; @%p71 bra $L__BB10_113; and.b32 %r373, %r16, -128; ld.u32 %r374, [%rd60+16]; setp.ne.s32 %p72, %r373, %r374; @%p72 bra $L__BB10_113; and.b32 %r375, %r3, -128; ld.u32 %r376, [%rd60+20]; setp.eq.s32 %p73, %r375, %r376; selp.u16 %rs95, 1, 0, %p73; $L__BB10_113: setp.eq.s16 %p74, %rs95, 0; @%p74 bra $L__BB10_117; shl.b32 %r377, %r1, 5; and.b32 %r378, %r377, 3840; shl.b32 %r379, %r16, 1; and.b32 %r380, %r379, 240; or.b32 %r381, %r380, %r378; bfe.u32 %r382, %r3, 3, 4; or.b32 %r383, %r381, %r382; shr.u32 %r384, %r381, 6; ld.u64 %rd61, [%rd60+56]; mul.wide.u32 %rd361, %r384, 8; add.s64 %rd362, %rd61, %rd361; and.b32 %r385, %r383, 63; mov.u64 %rd363, 1; shl.b64 %rd364, %rd363, %r385; ld.u64 %rd365, [%rd362+544]; and.b64 %rd366, %rd365, %rd364; setp.eq.s64 %p75, %rd366, 0; mul.wide.u32 %rd367, %r383, 16; add.s64 %rd368, %rd61, %rd367; add.s64 %rd62, %rd368, 1088; @%p75 bra $L__BB10_116; ld.u64 %rd369, [%rd62]; add.s64 %rd370, %rd61, %rd369; and.b32 %r386, %r16, -8; st.v2.u32 [%rd60], {%r4, %r386}; and.b32 %r387, %r3, -8; st.u32 [%rd60+8], %r387; st.u64 [%rd60+48], %rd370; and.b32 %r388, %r1, 7; shl.b32 %r389, %r16, 3; and.b32 %r390, %r389, 56; and.b32 %r391, %r3, 7; bfi.b32 %r392, %r388, %r391, 6, 3; or.b32 %r393, %r392, %r390; mul.wide.u32 %rd371, %r393, 12; add.s64 %rd372, %rd370, %rd371; ld.f32 %f351, [%rd372+128]; ld.f32 %f350, [%rd372+132]; ld.f32 %f349, [%rd372+136]; bra.uni $L__BB10_140; $L__BB10_145: and.b32 %r29, %r27, -128; ld.u32 %r472, [%rd79+12]; setp.ne.s32 %p93, %r29, %r472; mov.u16 %rs98, 0; @%p93 bra $L__BB10_148; and.b32 %r473, %r16, -128; ld.u32 %r474, [%rd79+16]; setp.ne.s32 %p94, %r473, %r474; @%p94 bra $L__BB10_148; and.b32 %r475, %r3, -128; ld.u32 %r476, [%rd79+20]; setp.eq.s32 %p95, %r475, %r476; selp.u16 %rs98, 1, 0, %p95; $L__BB10_148: setp.eq.s16 %p96, %rs98, 0; @%p96 bra $L__BB10_152; shl.b32 %r477, %r27, 5; and.b32 %r478, %r477, 3840; shl.b32 %r479, %r16, 1; and.b32 %r480, %r479, 240; or.b32 %r481, %r480, %r478; bfe.u32 %r482, %r3, 3, 4; or.b32 %r483, %r481, %r482; shr.u32 %r484, %r481, 6; ld.u64 %rd80, [%rd79+56]; mul.wide.u32 %rd428, %r484, 8; add.s64 %rd429, %rd80, %rd428; and.b32 %r485, %r483, 63; mov.u64 %rd430, 1; shl.b64 %rd431, %rd430, %r485; ld.u64 %rd432, [%rd429+544]; and.b64 %rd433, %rd432, %rd431; setp.eq.s64 %p97, %rd433, 0; mul.wide.u32 %rd434, %r483, 16; add.s64 %rd435, %rd80, %rd434; add.s64 %rd81, %rd435, 1088; @%p97 bra $L__BB10_151; ld.u64 %rd436, [%rd81]; add.s64 %rd437, %rd80, %rd436; and.b32 %r486, %r16, -8; st.v2.u32 [%rd79], {%r28, %r486}; and.b32 %r487, %r3, -8; st.u32 [%rd79+8], %r487; st.u64 [%rd79+48], %rd437; and.b32 %r488, %r27, 7; shl.b32 %r489, %r16, 3; and.b32 %r490, %r489, 56; and.b32 %r491, %r3, 7; bfi.b32 %r492, %r488, %r491, 6, 3; or.b32 %r493, %r492, %r490; mul.wide.u32 %rd438, %r493, 12; add.s64 %rd439, %rd437, %rd438; ld.f32 %f354, [%rd439+128]; ld.f32 %f353, [%rd439+132]; ld.f32 %f352, [%rd439+136]; bra.uni $L__BB10_175; $L__BB10_180: and.b32 %r34, %r27, -128; ld.u32 %r573, [%rd98+12]; setp.ne.s32 %p115, %r34, %r573; mov.u16 %rs101, 0; @%p115 bra $L__BB10_183; and.b32 %r574, %r16, -128; ld.u32 %r575, [%rd98+16]; setp.ne.s32 %p116, %r574, %r575; @%p116 bra $L__BB10_183; and.b32 %r576, %r10, -128; ld.u32 %r577, [%rd98+20]; setp.eq.s32 %p117, %r576, %r577; selp.u16 %rs101, 1, 0, %p117; $L__BB10_183: setp.eq.s16 %p118, %rs101, 0; @%p118 bra $L__BB10_187; shl.b32 %r578, %r27, 5; and.b32 %r579, %r578, 3840; shl.b32 %r580, %r16, 1; and.b32 %r581, %r580, 240; or.b32 %r582, %r581, %r579; bfe.u32 %r583, %r10, 3, 4; or.b32 %r584, %r582, %r583; shr.u32 %r585, %r582, 6; ld.u64 %rd99, [%rd98+56]; mul.wide.u32 %rd495, %r585, 8; add.s64 %rd496, %rd99, %rd495; and.b32 %r586, %r584, 63; mov.u64 %rd497, 1; shl.b64 %rd498, %rd497, %r586; ld.u64 %rd499, [%rd496+544]; and.b64 %rd500, %rd499, %rd498; setp.eq.s64 %p119, %rd500, 0; mul.wide.u32 %rd501, %r584, 16; add.s64 %rd502, %rd99, %rd501; add.s64 %rd100, %rd502, 1088; @%p119 bra $L__BB10_186; ld.u64 %rd503, [%rd100]; add.s64 %rd504, %rd99, %rd503; and.b32 %r587, %r16, -8; st.v2.u32 [%rd98], {%r28, %r587}; and.b32 %r588, %r10, -8; st.u32 [%rd98+8], %r588; st.u64 [%rd98+48], %rd504; shl.b32 %r589, %r27, 6; and.b32 %r590, %r589, 448; shl.b32 %r591, %r16, 3; and.b32 %r592, %r591, 56; or.b32 %r593, %r592, %r590; and.b32 %r594, %r10, 7; or.b32 %r595, %r593, %r594; mul.wide.u32 %rd505, %r595, 12; add.s64 %rd506, %rd504, %rd505; ld.f32 %f357, [%rd506+128]; ld.f32 %f356, [%rd506+132]; ld.f32 %f355, [%rd506+136]; bra.uni $L__BB10_210; $L__BB10_215: and.b32 %r39, %r27, -128; ld.u32 %r677, [%rd117+12]; setp.ne.s32 %p137, %r39, %r677; mov.u16 %rs104, 0; @%p137 bra $L__BB10_218; and.b32 %r678, %r2, -128; ld.u32 %r679, [%rd117+16]; setp.ne.s32 %p138, %r678, %r679; @%p138 bra $L__BB10_218; and.b32 %r680, %r10, -128; ld.u32 %r681, [%rd117+20]; setp.eq.s32 %p139, %r680, %r681; selp.u16 %rs104, 1, 0, %p139; $L__BB10_218: setp.eq.s16 %p140, %rs104, 0; @%p140 bra $L__BB10_222; shl.b32 %r682, %r27, 5; and.b32 %r683, %r682, 3840; shl.b32 %r684, %r2, 1; and.b32 %r685, %r684, 240; or.b32 %r686, %r683, %r685; bfe.u32 %r687, %r10, 3, 4; or.b32 %r688, %r687, %r686; shr.u32 %r689, %r686, 6; ld.u64 %rd118, [%rd117+56]; mul.wide.u32 %rd562, %r689, 8; add.s64 %rd563, %rd118, %rd562; and.b32 %r690, %r688, 63; mov.u64 %rd564, 1; shl.b64 %rd565, %rd564, %r690; ld.u64 %rd566, [%rd563+544]; and.b64 %rd567, %rd566, %rd565; setp.eq.s64 %p141, %rd567, 0; mul.wide.u32 %rd568, %r688, 16; add.s64 %rd569, %rd118, %rd568; add.s64 %rd119, %rd569, 1088; @%p141 bra $L__BB10_221; ld.u64 %rd570, [%rd119]; add.s64 %rd571, %rd118, %rd570; and.b32 %r691, %r2, -8; st.v2.u32 [%rd117], {%r28, %r691}; and.b32 %r692, %r10, -8; st.u32 [%rd117+8], %r692; st.u64 [%rd117+48], %rd571; shl.b32 %r693, %r27, 6; and.b32 %r694, %r693, 448; shl.b32 %r695, %r2, 3; and.b32 %r696, %r695, 56; or.b32 %r697, %r694, %r696; and.b32 %r698, %r10, 7; or.b32 %r699, %r697, %r698; mul.wide.u32 %rd572, %r699, 12; add.s64 %rd573, %rd571, %rd572; ld.f32 %f360, [%rd573+128]; ld.f32 %f359, [%rd573+132]; ld.f32 %f358, [%rd573+136]; bra.uni $L__BB10_245; $L__BB10_258: shl.b64 %rd637, %rd140, 5; add.s64 %rd702, %rd138, %rd637; $L__BB10_259: setp.eq.s64 %p167, %rd702, 0; @%p167 bra $L__BB10_267; add.s64 %rd143, %rd702, 8; ld.u64 %rd144, [%rd702+8]; setp.eq.s64 %p168, %rd144, 0; @%p168 bra $L__BB10_266; add.s64 %rd145, %rd137, %rd144; and.b32 %r788, %r2, -4096; mov.u64 %rd638, 1; st.v2.u32 [%rd136+24], {%r45, %r788}; and.b32 %r789, %r3, -4096; st.u32 [%rd136+32], %r789; st.u64 [%rd136+64], %rd145; shl.b32 %r790, %r27, 3; and.b32 %r791, %r790, 31744; and.b32 %r792, %r2, 3968; shr.u32 %r793, %r792, 2; or.b32 %r794, %r791, %r793; bfe.u32 %r795, %r3, 7, 5; or.b32 %r796, %r794, %r795; shr.u32 %r797, %r794, 6; mul.wide.u32 %rd639, %r797, 8; add.s64 %rd640, %rd145, %rd639; and.b32 %r798, %r796, 63; shl.b64 %rd641, %rd638, %r798; ld.u64 %rd642, [%rd640+4128]; and.b64 %rd643, %rd642, %rd641; setp.eq.s64 %p169, %rd643, 0; mul.wide.u32 %rd644, %r796, 16; add.s64 %rd645, %rd145, %rd644; add.s64 %rd146, %rd645, 8256; @%p169 bra $L__BB10_265; ld.u64 %rd646, [%rd146]; add.s64 %rd147, %rd145, %rd646; st.u32 [%rd136+12], %r44; and.b32 %r799, %r2, -128; st.u32 [%rd136+16], %r799; and.b32 %r800, %r3, -128; st.u32 [%rd136+20], %r800; st.u64 [%rd136+56], %rd147; shl.b32 %r801, %r27, 5; and.b32 %r802, %r801, 3840; shl.b32 %r803, %r2, 1; and.b32 %r804, %r803, 240; or.b32 %r805, %r802, %r804; bfe.u32 %r806, %r3, 3, 4; or.b32 %r807, %r805, %r806; shr.u32 %r808, %r805, 6; mul.wide.u32 %rd647, %r808, 8; add.s64 %rd648, %rd147, %rd647; and.b32 %r809, %r807, 63; shl.b64 %rd650, %rd638, %r809; ld.u64 %rd651, [%rd648+544]; and.b64 %rd652, %rd651, %rd650; setp.eq.s64 %p170, %rd652, 0; mul.wide.u32 %rd653, %r807, 16; add.s64 %rd654, %rd147, %rd653; add.s64 %rd148, %rd654, 1088; @%p170 bra $L__BB10_264; ld.u64 %rd655, [%rd148]; add.s64 %rd656, %rd147, %rd655; and.b32 %r810, %r2, -8; st.v2.u32 [%rd136], {%r28, %r810}; and.b32 %r811, %r3, -8; st.u32 [%rd136+8], %r811; st.u64 [%rd136+48], %rd656; shl.b32 %r812, %r27, 6; and.b32 %r813, %r812, 448; shl.b32 %r814, %r2, 3; and.b32 %r815, %r814, 56; or.b32 %r816, %r813, %r815; and.b32 %r817, %r3, 7; or.b32 %r818, %r816, %r817; mul.wide.u32 %rd657, %r818, 12; add.s64 %rd658, %rd656, %rd657; ld.f32 %f363, [%rd658+128]; ld.f32 %f362, [%rd658+132]; ld.f32 %f361, [%rd658+136]; bra.uni $L__BB10_277; $L__BB10_267: ld.f32 %f363, [%rd137+28]; ld.f32 %f362, [%rd137+32]; ld.f32 %f361, [%rd137+36]; bra.uni $L__BB10_277; $L__BB10_266: ld.f32 %f363, [%rd143+12]; ld.f32 %f362, [%rd143+16]; ld.f32 %f361, [%rd143+20]; bra.uni $L__BB10_277; $L__BB10_12: and.b32 %r6, %r1, -4096; ld.u32 %r84, [%rd2+24]; setp.ne.s32 %p10, %r6, %r84; mov.u16 %rs87, 0; @%p10 bra $L__BB10_15; and.b32 %r85, %r2, -4096; ld.u32 %r86, [%rd2+28]; setp.ne.s32 %p11, %r85, %r86; @%p11 bra $L__BB10_15; and.b32 %r87, %r3, -4096; ld.u32 %r88, [%rd2+32]; setp.eq.s32 %p12, %r87, %r88; selp.u16 %rs87, 1, 0, %p12; $L__BB10_15: setp.eq.s16 %p13, %rs87, 0; @%p13 bra $L__BB10_21; shl.b32 %r89, %r1, 3; and.b32 %r90, %r89, 31744; and.b32 %r91, %r2, 3968; shr.u32 %r92, %r91, 2; or.b32 %r93, %r92, %r90; bfe.u32 %r94, %r3, 7, 5; or.b32 %r95, %r93, %r94; shr.u32 %r96, %r93, 6; ld.u64 %rd5, [%rd2+64]; mul.wide.u32 %rd171, %r96, 8; add.s64 %rd172, %rd5, %rd171; and.b32 %r97, %r95, 63; mov.u64 %rd173, 1; shl.b64 %rd174, %rd173, %r97; ld.u64 %rd175, [%rd172+4128]; and.b64 %rd176, %rd175, %rd174; setp.eq.s64 %p14, %rd176, 0; mul.wide.u32 %rd177, %r95, 16; add.s64 %rd178, %rd5, %rd177; add.s64 %rd6, %rd178, 8256; @%p14 bra $L__BB10_20; ld.u64 %rd179, [%rd6]; add.s64 %rd7, %rd5, %rd179; st.u32 [%rd2+12], %r5; and.b32 %r98, %r2, -128; st.u32 [%rd2+16], %r98; and.b32 %r99, %r3, -128; st.u32 [%rd2+20], %r99; st.u64 [%rd2+56], %rd7; shl.b32 %r100, %r1, 5; and.b32 %r101, %r100, 3840; shl.b32 %r102, %r2, 1; and.b32 %r103, %r102, 240; or.b32 %r104, %r103, %r101; bfe.u32 %r105, %r3, 3, 4; or.b32 %r106, %r104, %r105; shr.u32 %r107, %r104, 6; mul.wide.u32 %rd180, %r107, 8; add.s64 %rd181, %rd7, %rd180; and.b32 %r108, %r106, 63; shl.b64 %rd183, %rd173, %r108; ld.u64 %rd184, [%rd181+544]; and.b64 %rd185, %rd184, %rd183; setp.eq.s64 %p15, %rd185, 0; mul.wide.u32 %rd186, %r106, 16; add.s64 %rd187, %rd7, %rd186; add.s64 %rd8, %rd187, 1088; @%p15 bra $L__BB10_19; ld.u64 %rd188, [%rd8]; add.s64 %rd189, %rd7, %rd188; and.b32 %r109, %r2, -8; st.v2.u32 [%rd2], {%r4, %r109}; and.b32 %r110, %r3, -8; st.u32 [%rd2+8], %r110; st.u64 [%rd2+48], %rd189; shl.b32 %r111, %r1, 6; and.b32 %r112, %r111, 448; shl.b32 %r113, %r2, 3; and.b32 %r114, %r113, 56; or.b32 %r115, %r114, %r112; and.b32 %r116, %r3, 7; or.b32 %r117, %r115, %r116; mul.wide.u32 %rd190, %r117, 12; add.s64 %rd191, %rd189, %rd190; ld.f32 %f342, [%rd191+128]; ld.f32 %f341, [%rd191+132]; ld.f32 %f340, [%rd191+136]; bra.uni $L__BB10_35; $L__BB10_47: and.b32 %r12, %r1, -4096; ld.u32 %r188, [%rd22+24]; setp.ne.s32 %p32, %r12, %r188; mov.u16 %rs90, 0; @%p32 bra $L__BB10_50; and.b32 %r189, %r2, -4096; ld.u32 %r190, [%rd22+28]; setp.ne.s32 %p33, %r189, %r190; @%p33 bra $L__BB10_50; and.b32 %r191, %r10, -4096; ld.u32 %r192, [%rd22+32]; setp.eq.s32 %p34, %r191, %r192; selp.u16 %rs90, 1, 0, %p34; $L__BB10_50: setp.eq.s16 %p35, %rs90, 0; @%p35 bra $L__BB10_56; shl.b32 %r193, %r1, 3; and.b32 %r194, %r193, 31744; and.b32 %r195, %r2, 3968; shr.u32 %r196, %r195, 2; or.b32 %r197, %r196, %r194; bfe.u32 %r198, %r10, 7, 5; or.b32 %r199, %r198, %r197; shr.u32 %r200, %r197, 6; ld.u64 %rd25, [%rd22+64]; mul.wide.u32 %rd238, %r200, 8; add.s64 %rd239, %rd25, %rd238; and.b32 %r201, %r199, 63; mov.u64 %rd240, 1; shl.b64 %rd241, %rd240, %r201; ld.u64 %rd242, [%rd239+4128]; and.b64 %rd243, %rd242, %rd241; setp.eq.s64 %p36, %rd243, 0; mul.wide.u32 %rd244, %r199, 16; add.s64 %rd245, %rd25, %rd244; add.s64 %rd26, %rd245, 8256; @%p36 bra $L__BB10_55; ld.u64 %rd246, [%rd26]; add.s64 %rd27, %rd25, %rd246; st.u32 [%rd22+12], %r11; and.b32 %r202, %r2, -128; st.u32 [%rd22+16], %r202; and.b32 %r203, %r10, -128; st.u32 [%rd22+20], %r203; st.u64 [%rd22+56], %rd27; shl.b32 %r204, %r1, 5; and.b32 %r205, %r204, 3840; shl.b32 %r206, %r2, 1; and.b32 %r207, %r206, 240; or.b32 %r208, %r207, %r205; bfe.u32 %r209, %r10, 3, 4; or.b32 %r210, %r209, %r208; shr.u32 %r211, %r208, 6; mul.wide.u32 %rd247, %r211, 8; add.s64 %rd248, %rd27, %rd247; and.b32 %r212, %r210, 63; shl.b64 %rd250, %rd240, %r212; ld.u64 %rd251, [%rd248+544]; and.b64 %rd252, %rd251, %rd250; setp.eq.s64 %p37, %rd252, 0; mul.wide.u32 %rd253, %r210, 16; add.s64 %rd254, %rd27, %rd253; add.s64 %rd28, %rd254, 1088; @%p37 bra $L__BB10_54; ld.u64 %rd255, [%rd28]; add.s64 %rd256, %rd27, %rd255; and.b32 %r213, %r2, -8; st.v2.u32 [%rd22], {%r4, %r213}; and.b32 %r214, %r10, -8; st.u32 [%rd22+8], %r214; st.u64 [%rd22+48], %rd256; shl.b32 %r215, %r1, 6; and.b32 %r216, %r215, 448; shl.b32 %r217, %r2, 3; and.b32 %r218, %r217, 56; or.b32 %r219, %r218, %r216; and.b32 %r220, %r10, 7; or.b32 %r221, %r219, %r220; mul.wide.u32 %rd257, %r221, 12; add.s64 %rd258, %rd256, %rd257; ld.f32 %f345, [%rd258+128]; ld.f32 %f344, [%rd258+132]; ld.f32 %f343, [%rd258+136]; bra.uni $L__BB10_70; $L__BB10_82: and.b32 %r18, %r1, -4096; ld.u32 %r292, [%rd41+24]; setp.ne.s32 %p54, %r18, %r292; mov.u16 %rs93, 0; @%p54 bra $L__BB10_85; and.b32 %r293, %r16, -4096; ld.u32 %r294, [%rd41+28]; setp.ne.s32 %p55, %r293, %r294; @%p55 bra $L__BB10_85; and.b32 %r295, %r10, -4096; ld.u32 %r296, [%rd41+32]; setp.eq.s32 %p56, %r295, %r296; selp.u16 %rs93, 1, 0, %p56; $L__BB10_85: setp.eq.s16 %p57, %rs93, 0; @%p57 bra $L__BB10_91; shl.b32 %r297, %r1, 3; and.b32 %r298, %r297, 31744; and.b32 %r299, %r16, 3968; shr.u32 %r300, %r299, 2; or.b32 %r301, %r300, %r298; bfe.u32 %r302, %r10, 7, 5; or.b32 %r303, %r301, %r302; shr.u32 %r304, %r301, 6; ld.u64 %rd44, [%rd41+64]; mul.wide.u32 %rd306, %r304, 8; add.s64 %rd307, %rd44, %rd306; and.b32 %r305, %r303, 63; mov.u64 %rd308, 1; shl.b64 %rd309, %rd308, %r305; ld.u64 %rd310, [%rd307+4128]; and.b64 %rd311, %rd310, %rd309; setp.eq.s64 %p58, %rd311, 0; mul.wide.u32 %rd312, %r303, 16; add.s64 %rd313, %rd44, %rd312; add.s64 %rd45, %rd313, 8256; @%p58 bra $L__BB10_90; ld.u64 %rd314, [%rd45]; add.s64 %rd46, %rd44, %rd314; st.u32 [%rd41+12], %r17; and.b32 %r306, %r16, -128; st.u32 [%rd41+16], %r306; and.b32 %r307, %r10, -128; st.u32 [%rd41+20], %r307; st.u64 [%rd41+56], %rd46; shl.b32 %r308, %r1, 5; and.b32 %r309, %r308, 3840; shl.b32 %r310, %r16, 1; and.b32 %r311, %r310, 240; or.b32 %r312, %r311, %r309; bfe.u32 %r313, %r10, 3, 4; or.b32 %r314, %r312, %r313; shr.u32 %r315, %r312, 6; mul.wide.u32 %rd315, %r315, 8; add.s64 %rd316, %rd46, %rd315; and.b32 %r316, %r314, 63; shl.b64 %rd318, %rd308, %r316; ld.u64 %rd319, [%rd316+544]; and.b64 %rd320, %rd319, %rd318; setp.eq.s64 %p59, %rd320, 0; mul.wide.u32 %rd321, %r314, 16; add.s64 %rd322, %rd46, %rd321; add.s64 %rd47, %rd322, 1088; @%p59 bra $L__BB10_89; ld.u64 %rd323, [%rd47]; add.s64 %rd324, %rd46, %rd323; and.b32 %r317, %r16, -8; st.v2.u32 [%rd41], {%r4, %r317}; and.b32 %r318, %r10, -8; st.u32 [%rd41+8], %r318; st.u64 [%rd41+48], %rd324; shl.b32 %r319, %r1, 6; and.b32 %r320, %r319, 448; shl.b32 %r321, %r16, 3; and.b32 %r322, %r321, 56; or.b32 %r323, %r322, %r320; and.b32 %r324, %r10, 7; or.b32 %r325, %r323, %r324; mul.wide.u32 %rd325, %r325, 12; add.s64 %rd326, %rd324, %rd325; ld.f32 %f348, [%rd326+128]; ld.f32 %f347, [%rd326+132]; ld.f32 %f346, [%rd326+136]; bra.uni $L__BB10_105; $L__BB10_117: and.b32 %r23, %r1, -4096; ld.u32 %r394, [%rd60+24]; setp.ne.s32 %p76, %r23, %r394; mov.u16 %rs96, 0; @%p76 bra $L__BB10_120; and.b32 %r395, %r16, -4096; ld.u32 %r396, [%rd60+28]; setp.ne.s32 %p77, %r395, %r396; @%p77 bra $L__BB10_120; and.b32 %r397, %r3, -4096; ld.u32 %r398, [%rd60+32]; setp.eq.s32 %p78, %r397, %r398; selp.u16 %rs96, 1, 0, %p78; $L__BB10_120: setp.eq.s16 %p79, %rs96, 0; @%p79 bra $L__BB10_126; shl.b32 %r399, %r1, 3; and.b32 %r400, %r399, 31744; and.b32 %r401, %r16, 3968; shr.u32 %r402, %r401, 2; or.b32 %r403, %r402, %r400; bfe.u32 %r404, %r3, 7, 5; or.b32 %r405, %r403, %r404; shr.u32 %r406, %r403, 6; ld.u64 %rd63, [%rd60+64]; mul.wide.u32 %rd373, %r406, 8; add.s64 %rd374, %rd63, %rd373; and.b32 %r407, %r405, 63; mov.u64 %rd375, 1; shl.b64 %rd376, %rd375, %r407; ld.u64 %rd377, [%rd374+4128]; and.b64 %rd378, %rd377, %rd376; setp.eq.s64 %p80, %rd378, 0; mul.wide.u32 %rd379, %r405, 16; add.s64 %rd380, %rd63, %rd379; add.s64 %rd64, %rd380, 8256; @%p80 bra $L__BB10_125; ld.u64 %rd381, [%rd64]; add.s64 %rd65, %rd63, %rd381; st.u32 [%rd60+12], %r22; and.b32 %r408, %r16, -128; st.u32 [%rd60+16], %r408; and.b32 %r409, %r3, -128; st.u32 [%rd60+20], %r409; st.u64 [%rd60+56], %rd65; shl.b32 %r410, %r1, 5; and.b32 %r411, %r410, 3840; shl.b32 %r412, %r16, 1; and.b32 %r413, %r412, 240; or.b32 %r414, %r413, %r411; bfe.u32 %r415, %r3, 3, 4; or.b32 %r416, %r414, %r415; shr.u32 %r417, %r414, 6; mul.wide.u32 %rd382, %r417, 8; add.s64 %rd383, %rd65, %rd382; and.b32 %r418, %r416, 63; shl.b64 %rd385, %rd375, %r418; ld.u64 %rd386, [%rd383+544]; and.b64 %rd387, %rd386, %rd385; setp.eq.s64 %p81, %rd387, 0; mul.wide.u32 %rd388, %r416, 16; add.s64 %rd389, %rd65, %rd388; add.s64 %rd66, %rd389, 1088; @%p81 bra $L__BB10_124; ld.u64 %rd390, [%rd66]; add.s64 %rd391, %rd65, %rd390; and.b32 %r419, %r16, -8; st.v2.u32 [%rd60], {%r4, %r419}; and.b32 %r420, %r3, -8; st.u32 [%rd60+8], %r420; st.u64 [%rd60+48], %rd391; and.b32 %r421, %r1, 7; shl.b32 %r422, %r16, 3; and.b32 %r423, %r422, 56; and.b32 %r424, %r3, 7; bfi.b32 %r425, %r421, %r424, 6, 3; or.b32 %r426, %r425, %r423; mul.wide.u32 %rd392, %r426, 12; add.s64 %rd393, %rd391, %rd392; ld.f32 %f351, [%rd393+128]; ld.f32 %f350, [%rd393+132]; ld.f32 %f349, [%rd393+136]; bra.uni $L__BB10_140; $L__BB10_152: and.b32 %r30, %r27, -4096; ld.u32 %r494, [%rd79+24]; setp.ne.s32 %p98, %r30, %r494; mov.u16 %rs99, 0; @%p98 bra $L__BB10_155; and.b32 %r495, %r16, -4096; ld.u32 %r496, [%rd79+28]; setp.ne.s32 %p99, %r495, %r496; @%p99 bra $L__BB10_155; and.b32 %r497, %r3, -4096; ld.u32 %r498, [%rd79+32]; setp.eq.s32 %p100, %r497, %r498; selp.u16 %rs99, 1, 0, %p100; $L__BB10_155: setp.eq.s16 %p101, %rs99, 0; @%p101 bra $L__BB10_161; shl.b32 %r499, %r27, 3; and.b32 %r500, %r499, 31744; and.b32 %r501, %r16, 3968; shr.u32 %r502, %r501, 2; or.b32 %r503, %r502, %r500; bfe.u32 %r504, %r3, 7, 5; or.b32 %r505, %r503, %r504; shr.u32 %r506, %r503, 6; ld.u64 %rd82, [%rd79+64]; mul.wide.u32 %rd440, %r506, 8; add.s64 %rd441, %rd82, %rd440; and.b32 %r507, %r505, 63; mov.u64 %rd442, 1; shl.b64 %rd443, %rd442, %r507; ld.u64 %rd444, [%rd441+4128]; and.b64 %rd445, %rd444, %rd443; setp.eq.s64 %p102, %rd445, 0; mul.wide.u32 %rd446, %r505, 16; add.s64 %rd447, %rd82, %rd446; add.s64 %rd83, %rd447, 8256; @%p102 bra $L__BB10_160; ld.u64 %rd448, [%rd83]; add.s64 %rd84, %rd82, %rd448; st.u32 [%rd79+12], %r29; and.b32 %r508, %r16, -128; st.u32 [%rd79+16], %r508; and.b32 %r509, %r3, -128; st.u32 [%rd79+20], %r509; st.u64 [%rd79+56], %rd84; shl.b32 %r510, %r27, 5; and.b32 %r511, %r510, 3840; shl.b32 %r512, %r16, 1; and.b32 %r513, %r512, 240; or.b32 %r514, %r513, %r511; bfe.u32 %r515, %r3, 3, 4; or.b32 %r516, %r514, %r515; shr.u32 %r517, %r514, 6; mul.wide.u32 %rd449, %r517, 8; add.s64 %rd450, %rd84, %rd449; and.b32 %r518, %r516, 63; shl.b64 %rd452, %rd442, %r518; ld.u64 %rd453, [%rd450+544]; and.b64 %rd454, %rd453, %rd452; setp.eq.s64 %p103, %rd454, 0; mul.wide.u32 %rd455, %r516, 16; add.s64 %rd456, %rd84, %rd455; add.s64 %rd85, %rd456, 1088; @%p103 bra $L__BB10_159; ld.u64 %rd457, [%rd85]; add.s64 %rd458, %rd84, %rd457; and.b32 %r519, %r16, -8; st.v2.u32 [%rd79], {%r28, %r519}; and.b32 %r520, %r3, -8; st.u32 [%rd79+8], %r520; st.u64 [%rd79+48], %rd458; and.b32 %r521, %r27, 7; shl.b32 %r522, %r16, 3; and.b32 %r523, %r522, 56; and.b32 %r524, %r3, 7; bfi.b32 %r525, %r521, %r524, 6, 3; or.b32 %r526, %r525, %r523; mul.wide.u32 %rd459, %r526, 12; add.s64 %rd460, %rd458, %rd459; ld.f32 %f354, [%rd460+128]; ld.f32 %f353, [%rd460+132]; ld.f32 %f352, [%rd460+136]; bra.uni $L__BB10_175; $L__BB10_187: and.b32 %r35, %r27, -4096; ld.u32 %r596, [%rd98+24]; setp.ne.s32 %p120, %r35, %r596; mov.u16 %rs102, 0; @%p120 bra $L__BB10_190; and.b32 %r597, %r16, -4096; ld.u32 %r598, [%rd98+28]; setp.ne.s32 %p121, %r597, %r598; @%p121 bra $L__BB10_190; and.b32 %r599, %r10, -4096; ld.u32 %r600, [%rd98+32]; setp.eq.s32 %p122, %r599, %r600; selp.u16 %rs102, 1, 0, %p122; $L__BB10_190: setp.eq.s16 %p123, %rs102, 0; @%p123 bra $L__BB10_196; shl.b32 %r601, %r27, 3; and.b32 %r602, %r601, 31744; and.b32 %r603, %r16, 3968; shr.u32 %r604, %r603, 2; or.b32 %r605, %r604, %r602; bfe.u32 %r606, %r10, 7, 5; or.b32 %r607, %r605, %r606; shr.u32 %r608, %r605, 6; ld.u64 %rd101, [%rd98+64]; mul.wide.u32 %rd507, %r608, 8; add.s64 %rd508, %rd101, %rd507; and.b32 %r609, %r607, 63; mov.u64 %rd509, 1; shl.b64 %rd510, %rd509, %r609; ld.u64 %rd511, [%rd508+4128]; and.b64 %rd512, %rd511, %rd510; setp.eq.s64 %p124, %rd512, 0; mul.wide.u32 %rd513, %r607, 16; add.s64 %rd514, %rd101, %rd513; add.s64 %rd102, %rd514, 8256; @%p124 bra $L__BB10_195; ld.u64 %rd515, [%rd102]; add.s64 %rd103, %rd101, %rd515; st.u32 [%rd98+12], %r34; and.b32 %r610, %r16, -128; st.u32 [%rd98+16], %r610; and.b32 %r611, %r10, -128; st.u32 [%rd98+20], %r611; st.u64 [%rd98+56], %rd103; shl.b32 %r612, %r27, 5; and.b32 %r613, %r612, 3840; shl.b32 %r614, %r16, 1; and.b32 %r615, %r614, 240; or.b32 %r616, %r615, %r613; bfe.u32 %r617, %r10, 3, 4; or.b32 %r618, %r616, %r617; shr.u32 %r619, %r616, 6; mul.wide.u32 %rd516, %r619, 8; add.s64 %rd517, %rd103, %rd516; and.b32 %r620, %r618, 63; shl.b64 %rd519, %rd509, %r620; ld.u64 %rd520, [%rd517+544]; and.b64 %rd521, %rd520, %rd519; setp.eq.s64 %p125, %rd521, 0; mul.wide.u32 %rd522, %r618, 16; add.s64 %rd523, %rd103, %rd522; add.s64 %rd104, %rd523, 1088; @%p125 bra $L__BB10_194; ld.u64 %rd524, [%rd104]; add.s64 %rd525, %rd103, %rd524; and.b32 %r621, %r16, -8; st.v2.u32 [%rd98], {%r28, %r621}; and.b32 %r622, %r10, -8; st.u32 [%rd98+8], %r622; st.u64 [%rd98+48], %rd525; shl.b32 %r623, %r27, 6; and.b32 %r624, %r623, 448; shl.b32 %r625, %r16, 3; and.b32 %r626, %r625, 56; or.b32 %r627, %r626, %r624; and.b32 %r628, %r10, 7; or.b32 %r629, %r627, %r628; mul.wide.u32 %rd526, %r629, 12; add.s64 %rd527, %rd525, %rd526; ld.f32 %f357, [%rd527+128]; ld.f32 %f356, [%rd527+132]; ld.f32 %f355, [%rd527+136]; bra.uni $L__BB10_210; $L__BB10_222: and.b32 %r40, %r27, -4096; ld.u32 %r700, [%rd117+24]; setp.ne.s32 %p142, %r40, %r700; mov.u16 %rs105, 0; @%p142 bra $L__BB10_225; and.b32 %r701, %r2, -4096; ld.u32 %r702, [%rd117+28]; setp.ne.s32 %p143, %r701, %r702; @%p143 bra $L__BB10_225; and.b32 %r703, %r10, -4096; ld.u32 %r704, [%rd117+32]; setp.eq.s32 %p144, %r703, %r704; selp.u16 %rs105, 1, 0, %p144; $L__BB10_225: setp.eq.s16 %p145, %rs105, 0; @%p145 bra $L__BB10_231; shl.b32 %r705, %r27, 3; and.b32 %r706, %r705, 31744; and.b32 %r707, %r2, 3968; shr.u32 %r708, %r707, 2; or.b32 %r709, %r706, %r708; bfe.u32 %r710, %r10, 7, 5; or.b32 %r711, %r710, %r709; shr.u32 %r712, %r709, 6; ld.u64 %rd120, [%rd117+64]; mul.wide.u32 %rd574, %r712, 8; add.s64 %rd575, %rd120, %rd574; and.b32 %r713, %r711, 63; mov.u64 %rd576, 1; shl.b64 %rd577, %rd576, %r713; ld.u64 %rd578, [%rd575+4128]; and.b64 %rd579, %rd578, %rd577; setp.eq.s64 %p146, %rd579, 0; mul.wide.u32 %rd580, %r711, 16; add.s64 %rd581, %rd120, %rd580; add.s64 %rd121, %rd581, 8256; @%p146 bra $L__BB10_230; ld.u64 %rd582, [%rd121]; add.s64 %rd122, %rd120, %rd582; st.u32 [%rd117+12], %r39; and.b32 %r714, %r2, -128; st.u32 [%rd117+16], %r714; and.b32 %r715, %r10, -128; st.u32 [%rd117+20], %r715; st.u64 [%rd117+56], %rd122; shl.b32 %r716, %r27, 5; and.b32 %r717, %r716, 3840; shl.b32 %r718, %r2, 1; and.b32 %r719, %r718, 240; or.b32 %r720, %r717, %r719; bfe.u32 %r721, %r10, 3, 4; or.b32 %r722, %r721, %r720; shr.u32 %r723, %r720, 6; mul.wide.u32 %rd583, %r723, 8; add.s64 %rd584, %rd122, %rd583; and.b32 %r724, %r722, 63; shl.b64 %rd586, %rd576, %r724; ld.u64 %rd587, [%rd584+544]; and.b64 %rd588, %rd587, %rd586; setp.eq.s64 %p147, %rd588, 0; mul.wide.u32 %rd589, %r722, 16; add.s64 %rd590, %rd122, %rd589; add.s64 %rd123, %rd590, 1088; @%p147 bra $L__BB10_229; ld.u64 %rd591, [%rd123]; add.s64 %rd592, %rd122, %rd591; and.b32 %r725, %r2, -8; st.v2.u32 [%rd117], {%r28, %r725}; and.b32 %r726, %r10, -8; st.u32 [%rd117+8], %r726; st.u64 [%rd117+48], %rd592; shl.b32 %r727, %r27, 6; and.b32 %r728, %r727, 448; shl.b32 %r729, %r2, 3; and.b32 %r730, %r729, 56; or.b32 %r731, %r728, %r730; and.b32 %r732, %r10, 7; or.b32 %r733, %r731, %r732; mul.wide.u32 %rd593, %r733, 12; add.s64 %rd594, %rd592, %rd593; ld.f32 %f360, [%rd594+128]; ld.f32 %f359, [%rd594+132]; ld.f32 %f358, [%rd594+136]; bra.uni $L__BB10_245; $L__BB10_265: ld.f32 %f363, [%rd146]; ld.f32 %f362, [%rd146+4]; ld.f32 %f361, [%rd146+8]; bra.uni $L__BB10_277; $L__BB10_11: ld.f32 %f342, [%rd4]; ld.f32 %f341, [%rd4+4]; ld.f32 %f340, [%rd4+8]; bra.uni $L__BB10_35; $L__BB10_46: ld.f32 %f345, [%rd24]; ld.f32 %f344, [%rd24+4]; ld.f32 %f343, [%rd24+8]; bra.uni $L__BB10_70; $L__BB10_81: ld.f32 %f348, [%rd43]; ld.f32 %f347, [%rd43+4]; ld.f32 %f346, [%rd43+8]; bra.uni $L__BB10_105; $L__BB10_116: ld.f32 %f351, [%rd62]; ld.f32 %f350, [%rd62+4]; ld.f32 %f349, [%rd62+8]; bra.uni $L__BB10_140; $L__BB10_151: ld.f32 %f354, [%rd81]; ld.f32 %f353, [%rd81+4]; ld.f32 %f352, [%rd81+8]; bra.uni $L__BB10_175; $L__BB10_186: ld.f32 %f357, [%rd100]; ld.f32 %f356, [%rd100+4]; ld.f32 %f355, [%rd100+8]; bra.uni $L__BB10_210; $L__BB10_221: ld.f32 %f360, [%rd119]; ld.f32 %f359, [%rd119+4]; ld.f32 %f358, [%rd119+8]; bra.uni $L__BB10_245; $L__BB10_264: ld.f32 %f363, [%rd148]; ld.f32 %f362, [%rd148+4]; ld.f32 %f361, [%rd148+8]; bra.uni $L__BB10_277; $L__BB10_21: ld.u64 %rd9, [%rd2+40]; mov.u64 %rd695, 0; add.s64 %rd10, %rd9, 24; ld.u32 %r7, [%rd9+24]; setp.eq.s32 %p16, %r7, 0; @%p16 bra $L__BB10_26; add.s64 %rd11, %rd9, 96; shr.u32 %r119, %r1, 12; cvt.u64.u32 %rd193, %r119; shr.u32 %r120, %r3, 12; cvt.u64.u32 %rd194, %r120; bfi.b64 %rd195, %rd193, %rd194, 42, 22; shr.u32 %r121, %r2, 12; mul.wide.u32 %rd196, %r121, 2097152; or.b64 %rd12, %rd195, %rd196; mov.u32 %r873, 0; $L__BB10_23: cvt.u64.u32 %rd13, %r873; mul.wide.u32 %rd197, %r873, 32; add.s64 %rd198, %rd11, %rd197; ld.u64 %rd199, [%rd198]; setp.eq.s64 %p17, %rd199, %rd12; add.s32 %r873, %r873, 1; @%p17 bra $L__BB10_25; setp.lt.u32 %p18, %r873, %r7; @%p18 bra $L__BB10_23; bra.uni $L__BB10_26; $L__BB10_56: ld.u64 %rd29, [%rd22+40]; mov.u64 %rd696, 0; ld.u32 %r13, [%rd29+24]; setp.eq.s32 %p38, %r13, 0; @%p38 bra $L__BB10_61; add.s64 %rd30, %rd29, 96; shr.u32 %r223, %r1, 12; cvt.u64.u32 %rd260, %r223; shl.b64 %rd261, %rd260, 42; shr.u32 %r224, %r2, 12; mul.wide.u32 %rd262, %r224, 2097152; or.b64 %rd263, %rd262, %rd261; shr.u32 %r225, %r10, 12; cvt.u64.u32 %rd264, %r225; or.b64 %rd31, %rd263, %rd264; mov.u32 %r874, 0; $L__BB10_58: cvt.u64.u32 %rd32, %r874; mul.wide.u32 %rd265, %r874, 32; add.s64 %rd266, %rd30, %rd265; ld.u64 %rd267, [%rd266]; setp.eq.s64 %p39, %rd267, %rd31; add.s32 %r874, %r874, 1; @%p39 bra $L__BB10_60; setp.lt.u32 %p40, %r874, %r13; @%p40 bra $L__BB10_58; bra.uni $L__BB10_61; $L__BB10_91: ld.u64 %rd48, [%rd41+40]; mov.u64 %rd697, 0; ld.u32 %r19, [%rd48+24]; setp.eq.s32 %p60, %r19, 0; @%p60 bra $L__BB10_96; add.s64 %rd49, %rd48, 96; shr.u32 %r327, %r1, 12; cvt.u64.u32 %rd328, %r327; shr.u32 %r328, %r10, 12; cvt.u64.u32 %rd329, %r328; bfi.b64 %rd330, %rd328, %rd329, 42, 22; shr.u32 %r329, %r16, 12; mul.wide.u32 %rd331, %r329, 2097152; or.b64 %rd50, %rd330, %rd331; mov.u32 %r875, 0; $L__BB10_93: cvt.u64.u32 %rd51, %r875; mul.wide.u32 %rd332, %r875, 32; add.s64 %rd333, %rd49, %rd332; ld.u64 %rd334, [%rd333]; setp.eq.s64 %p61, %rd334, %rd50; add.s32 %r875, %r875, 1; @%p61 bra $L__BB10_95; setp.lt.u32 %p62, %r875, %r19; @%p62 bra $L__BB10_93; bra.uni $L__BB10_96; $L__BB10_126: ld.u64 %rd67, [%rd60+40]; mov.u64 %rd698, 0; ld.u32 %r24, [%rd67+24]; setp.eq.s32 %p82, %r24, 0; @%p82 bra $L__BB10_131; add.s64 %rd68, %rd67, 96; shr.u32 %r428, %r1, 12; cvt.u64.u32 %rd395, %r428; shr.u32 %r429, %r3, 12; cvt.u64.u32 %rd396, %r429; bfi.b64 %rd397, %rd395, %rd396, 42, 22; shr.u32 %r430, %r16, 12; mul.wide.u32 %rd398, %r430, 2097152; or.b64 %rd69, %rd397, %rd398; mov.u32 %r876, 0; $L__BB10_128: cvt.u64.u32 %rd70, %r876; mul.wide.u32 %rd399, %r876, 32; add.s64 %rd400, %rd68, %rd399; ld.u64 %rd401, [%rd400]; setp.eq.s64 %p83, %rd401, %rd69; add.s32 %r876, %r876, 1; @%p83 bra $L__BB10_130; setp.lt.u32 %p84, %r876, %r24; @%p84 bra $L__BB10_128; bra.uni $L__BB10_131; $L__BB10_161: ld.u64 %rd86, [%rd79+40]; mov.u64 %rd699, 0; ld.u32 %r31, [%rd86+24]; setp.eq.s32 %p104, %r31, 0; @%p104 bra $L__BB10_166; add.s64 %rd87, %rd86, 96; shr.u32 %r528, %r27, 12; cvt.u64.u32 %rd462, %r528; shr.u32 %r529, %r3, 12; cvt.u64.u32 %rd463, %r529; bfi.b64 %rd464, %rd462, %rd463, 42, 22; shr.u32 %r530, %r16, 12; mul.wide.u32 %rd465, %r530, 2097152; or.b64 %rd88, %rd464, %rd465; mov.u32 %r877, 0; $L__BB10_163: cvt.u64.u32 %rd89, %r877; mul.wide.u32 %rd466, %r877, 32; add.s64 %rd467, %rd87, %rd466; ld.u64 %rd468, [%rd467]; setp.eq.s64 %p105, %rd468, %rd88; add.s32 %r877, %r877, 1; @%p105 bra $L__BB10_165; setp.lt.u32 %p106, %r877, %r31; @%p106 bra $L__BB10_163; bra.uni $L__BB10_166; $L__BB10_196: ld.u64 %rd105, [%rd98+40]; mov.u64 %rd700, 0; ld.u32 %r36, [%rd105+24]; setp.eq.s32 %p126, %r36, 0; @%p126 bra $L__BB10_201; add.s64 %rd106, %rd105, 96; shr.u32 %r631, %r27, 12; cvt.u64.u32 %rd529, %r631; shr.u32 %r632, %r10, 12; cvt.u64.u32 %rd530, %r632; bfi.b64 %rd531, %rd529, %rd530, 42, 22; shr.u32 %r633, %r16, 12; mul.wide.u32 %rd532, %r633, 2097152; or.b64 %rd107, %rd531, %rd532; mov.u32 %r878, 0; $L__BB10_198: cvt.u64.u32 %rd108, %r878; mul.wide.u32 %rd533, %r878, 32; add.s64 %rd534, %rd106, %rd533; ld.u64 %rd535, [%rd534]; setp.eq.s64 %p127, %rd535, %rd107; add.s32 %r878, %r878, 1; @%p127 bra $L__BB10_200; setp.lt.u32 %p128, %r878, %r36; @%p128 bra $L__BB10_198; bra.uni $L__BB10_201; $L__BB10_231: ld.u64 %rd124, [%rd117+40]; mov.u64 %rd701, 0; ld.u32 %r41, [%rd124+24]; setp.eq.s32 %p148, %r41, 0; @%p148 bra $L__BB10_236; add.s64 %rd125, %rd124, 96; shr.u32 %r735, %r27, 12; cvt.u64.u32 %rd596, %r735; shl.b64 %rd597, %rd596, 42; shr.u32 %r736, %r2, 12; mul.wide.u32 %rd598, %r736, 2097152; or.b64 %rd599, %rd597, %rd598; shr.u32 %r737, %r10, 12; cvt.u64.u32 %rd600, %r737; or.b64 %rd126, %rd599, %rd600; mov.u32 %r879, 0; $L__BB10_233: cvt.u64.u32 %rd127, %r879; mul.wide.u32 %rd601, %r879, 32; add.s64 %rd602, %rd125, %rd601; ld.u64 %rd603, [%rd602]; setp.eq.s64 %p149, %rd603, %rd126; add.s32 %r879, %r879, 1; @%p149 bra $L__BB10_235; setp.lt.u32 %p150, %r879, %r41; @%p150 bra $L__BB10_233; bra.uni $L__BB10_236; $L__BB10_275: ld.f32 %f363, [%rd154]; ld.f32 %f362, [%rd154+4]; ld.f32 %f361, [%rd154+8]; bra.uni $L__BB10_277; $L__BB10_272: ld.f32 %f363, [%rd150]; ld.f32 %f362, [%rd150+4]; ld.f32 %f361, [%rd150+8]; bra.uni $L__BB10_277; $L__BB10_20: ld.f32 %f342, [%rd6]; ld.f32 %f341, [%rd6+4]; ld.f32 %f340, [%rd6+8]; bra.uni $L__BB10_35; $L__BB10_55: ld.f32 %f345, [%rd26]; ld.f32 %f344, [%rd26+4]; ld.f32 %f343, [%rd26+8]; bra.uni $L__BB10_70; $L__BB10_90: ld.f32 %f348, [%rd45]; ld.f32 %f347, [%rd45+4]; ld.f32 %f346, [%rd45+8]; bra.uni $L__BB10_105; $L__BB10_125: ld.f32 %f351, [%rd64]; ld.f32 %f350, [%rd64+4]; ld.f32 %f349, [%rd64+8]; bra.uni $L__BB10_140; $L__BB10_160: ld.f32 %f354, [%rd83]; ld.f32 %f353, [%rd83+4]; ld.f32 %f352, [%rd83+8]; bra.uni $L__BB10_175; $L__BB10_195: ld.f32 %f357, [%rd102]; ld.f32 %f356, [%rd102+4]; ld.f32 %f355, [%rd102+8]; bra.uni $L__BB10_210; $L__BB10_230: ld.f32 %f360, [%rd121]; ld.f32 %f359, [%rd121+4]; ld.f32 %f358, [%rd121+8]; bra.uni $L__BB10_245; $L__BB10_271: ld.f32 %f363, [%rd152]; ld.f32 %f362, [%rd152+4]; ld.f32 %f361, [%rd152+8]; $L__BB10_277: sub.ftz.f32 %f322, %f360, %f363; sub.ftz.f32 %f323, %f359, %f362; sub.ftz.f32 %f324, %f358, %f361; fma.rn.ftz.f32 %f325, %f3, %f322, %f363; fma.rn.ftz.f32 %f326, %f3, %f323, %f362; fma.rn.ftz.f32 %f327, %f3, %f324, %f361; sub.ftz.f32 %f328, %f226, %f325; sub.ftz.f32 %f329, %f227, %f326; sub.ftz.f32 %f330, %f228, %f327; fma.rn.ftz.f32 %f331, %f2, %f328, %f325; fma.rn.ftz.f32 %f332, %f2, %f329, %f326; fma.rn.ftz.f32 %f333, %f2, %f330, %f327; sub.ftz.f32 %f334, %f331, %f151; sub.ftz.f32 %f335, %f332, %f152; sub.ftz.f32 %f336, %f333, %f153; fma.rn.ftz.f32 %f337, %f1, %f336, %f153; fma.rn.ftz.f32 %f338, %f1, %f335, %f152; fma.rn.ftz.f32 %f339, %f1, %f334, %f151; st.param.f32 [func_retval0+0], %f339; st.param.f32 [func_retval0+4], %f338; st.param.f32 [func_retval0+8], %f337; ret; $L__BB10_19: ld.f32 %f342, [%rd8]; ld.f32 %f341, [%rd8+4]; ld.f32 %f340, [%rd8+8]; bra.uni $L__BB10_35; $L__BB10_54: ld.f32 %f345, [%rd28]; ld.f32 %f344, [%rd28+4]; ld.f32 %f343, [%rd28+8]; bra.uni $L__BB10_70; $L__BB10_89: ld.f32 %f348, [%rd47]; ld.f32 %f347, [%rd47+4]; ld.f32 %f346, [%rd47+8]; bra.uni $L__BB10_105; $L__BB10_124: ld.f32 %f351, [%rd66]; ld.f32 %f350, [%rd66+4]; ld.f32 %f349, [%rd66+8]; bra.uni $L__BB10_140; $L__BB10_159: ld.f32 %f354, [%rd85]; ld.f32 %f353, [%rd85+4]; ld.f32 %f352, [%rd85+8]; bra.uni $L__BB10_175; $L__BB10_194: ld.f32 %f357, [%rd104]; ld.f32 %f356, [%rd104+4]; ld.f32 %f355, [%rd104+8]; bra.uni $L__BB10_210; $L__BB10_229: ld.f32 %f360, [%rd123]; ld.f32 %f359, [%rd123+4]; ld.f32 %f358, [%rd123+8]; bra.uni $L__BB10_245; $L__BB10_25: shl.b64 %rd201, %rd13, 5; add.s64 %rd695, %rd11, %rd201; $L__BB10_26: setp.eq.s64 %p19, %rd695, 0; @%p19 bra $L__BB10_34; add.s64 %rd16, %rd695, 8; ld.u64 %rd17, [%rd695+8]; setp.eq.s64 %p20, %rd17, 0; @%p20 bra $L__BB10_33; add.s64 %rd18, %rd9, %rd17; and.b32 %r122, %r2, -4096; mov.u64 %rd202, 1; st.v2.u32 [%rd2+24], {%r6, %r122}; and.b32 %r123, %r3, -4096; st.u32 [%rd2+32], %r123; st.u64 [%rd2+64], %rd18; shl.b32 %r124, %r1, 3; and.b32 %r125, %r124, 31744; and.b32 %r126, %r2, 3968; shr.u32 %r127, %r126, 2; or.b32 %r128, %r127, %r125; bfe.u32 %r129, %r3, 7, 5; or.b32 %r130, %r128, %r129; shr.u32 %r131, %r128, 6; mul.wide.u32 %rd203, %r131, 8; add.s64 %rd204, %rd18, %rd203; and.b32 %r132, %r130, 63; shl.b64 %rd205, %rd202, %r132; ld.u64 %rd206, [%rd204+4128]; and.b64 %rd207, %rd206, %rd205; setp.eq.s64 %p21, %rd207, 0; mul.wide.u32 %rd208, %r130, 16; add.s64 %rd209, %rd18, %rd208; add.s64 %rd19, %rd209, 8256; @%p21 bra $L__BB10_32; ld.u64 %rd210, [%rd19]; add.s64 %rd20, %rd18, %rd210; st.u32 [%rd2+12], %r5; and.b32 %r133, %r2, -128; st.u32 [%rd2+16], %r133; and.b32 %r134, %r3, -128; st.u32 [%rd2+20], %r134; st.u64 [%rd2+56], %rd20; shl.b32 %r135, %r1, 5; and.b32 %r136, %r135, 3840; shl.b32 %r137, %r2, 1; and.b32 %r138, %r137, 240; or.b32 %r139, %r138, %r136; bfe.u32 %r140, %r3, 3, 4; or.b32 %r141, %r139, %r140; shr.u32 %r142, %r139, 6; mul.wide.u32 %rd211, %r142, 8; add.s64 %rd212, %rd20, %rd211; and.b32 %r143, %r141, 63; shl.b64 %rd214, %rd202, %r143; ld.u64 %rd215, [%rd212+544]; and.b64 %rd216, %rd215, %rd214; setp.eq.s64 %p22, %rd216, 0; mul.wide.u32 %rd217, %r141, 16; add.s64 %rd218, %rd20, %rd217; add.s64 %rd21, %rd218, 1088; @%p22 bra $L__BB10_31; ld.u64 %rd219, [%rd21]; add.s64 %rd220, %rd20, %rd219; and.b32 %r144, %r2, -8; st.v2.u32 [%rd2], {%r4, %r144}; and.b32 %r145, %r3, -8; st.u32 [%rd2+8], %r145; st.u64 [%rd2+48], %rd220; shl.b32 %r146, %r1, 6; and.b32 %r147, %r146, 448; shl.b32 %r148, %r2, 3; and.b32 %r149, %r148, 56; or.b32 %r150, %r149, %r147; and.b32 %r151, %r3, 7; or.b32 %r152, %r150, %r151; mul.wide.u32 %rd221, %r152, 12; add.s64 %rd222, %rd220, %rd221; ld.f32 %f342, [%rd222+128]; ld.f32 %f341, [%rd222+132]; ld.f32 %f340, [%rd222+136]; bra.uni $L__BB10_35; $L__BB10_34: ld.f32 %f342, [%rd10+4]; ld.f32 %f341, [%rd10+8]; ld.f32 %f340, [%rd10+12]; bra.uni $L__BB10_35; $L__BB10_60: shl.b64 %rd269, %rd32, 5; add.s64 %rd696, %rd30, %rd269; $L__BB10_61: setp.eq.s64 %p41, %rd696, 0; @%p41 bra $L__BB10_69; add.s64 %rd35, %rd696, 8; ld.u64 %rd36, [%rd696+8]; setp.eq.s64 %p42, %rd36, 0; @%p42 bra $L__BB10_68; add.s64 %rd37, %rd29, %rd36; and.b32 %r226, %r2, -4096; mov.u64 %rd270, 1; st.v2.u32 [%rd22+24], {%r12, %r226}; and.b32 %r227, %r10, -4096; st.u32 [%rd22+32], %r227; st.u64 [%rd22+64], %rd37; shl.b32 %r228, %r1, 3; and.b32 %r229, %r228, 31744; and.b32 %r230, %r2, 3968; shr.u32 %r231, %r230, 2; or.b32 %r232, %r231, %r229; bfe.u32 %r233, %r10, 7, 5; or.b32 %r234, %r233, %r232; shr.u32 %r235, %r232, 6; mul.wide.u32 %rd271, %r235, 8; add.s64 %rd272, %rd37, %rd271; and.b32 %r236, %r234, 63; shl.b64 %rd273, %rd270, %r236; ld.u64 %rd274, [%rd272+4128]; and.b64 %rd275, %rd274, %rd273; setp.eq.s64 %p43, %rd275, 0; mul.wide.u32 %rd276, %r234, 16; add.s64 %rd277, %rd37, %rd276; add.s64 %rd38, %rd277, 8256; @%p43 bra $L__BB10_67; ld.u64 %rd278, [%rd38]; add.s64 %rd39, %rd37, %rd278; st.u32 [%rd22+12], %r11; and.b32 %r237, %r2, -128; st.u32 [%rd22+16], %r237; and.b32 %r238, %r10, -128; st.u32 [%rd22+20], %r238; st.u64 [%rd22+56], %rd39; shl.b32 %r239, %r1, 5; and.b32 %r240, %r239, 3840; shl.b32 %r241, %r2, 1; and.b32 %r242, %r241, 240; or.b32 %r243, %r242, %r240; bfe.u32 %r244, %r10, 3, 4; or.b32 %r245, %r244, %r243; shr.u32 %r246, %r243, 6; mul.wide.u32 %rd279, %r246, 8; add.s64 %rd280, %rd39, %rd279; and.b32 %r247, %r245, 63; shl.b64 %rd282, %rd270, %r247; ld.u64 %rd283, [%rd280+544]; and.b64 %rd284, %rd283, %rd282; setp.eq.s64 %p44, %rd284, 0; mul.wide.u32 %rd285, %r245, 16; add.s64 %rd286, %rd39, %rd285; add.s64 %rd40, %rd286, 1088; @%p44 bra $L__BB10_66; ld.u64 %rd287, [%rd40]; add.s64 %rd288, %rd39, %rd287; and.b32 %r248, %r2, -8; st.v2.u32 [%rd22], {%r4, %r248}; and.b32 %r249, %r10, -8; st.u32 [%rd22+8], %r249; st.u64 [%rd22+48], %rd288; shl.b32 %r250, %r1, 6; and.b32 %r251, %r250, 448; shl.b32 %r252, %r2, 3; and.b32 %r253, %r252, 56; or.b32 %r254, %r253, %r251; and.b32 %r255, %r10, 7; or.b32 %r256, %r254, %r255; mul.wide.u32 %rd289, %r256, 12; add.s64 %rd290, %rd288, %rd289; ld.f32 %f345, [%rd290+128]; ld.f32 %f344, [%rd290+132]; ld.f32 %f343, [%rd290+136]; bra.uni $L__BB10_70; $L__BB10_69: ld.f32 %f345, [%rd29+28]; ld.f32 %f344, [%rd29+32]; ld.f32 %f343, [%rd29+36]; bra.uni $L__BB10_70; $L__BB10_95: shl.b64 %rd336, %rd51, 5; add.s64 %rd697, %rd49, %rd336; $L__BB10_96: setp.eq.s64 %p63, %rd697, 0; @%p63 bra $L__BB10_104; add.s64 %rd54, %rd697, 8; ld.u64 %rd55, [%rd697+8]; setp.eq.s64 %p64, %rd55, 0; @%p64 bra $L__BB10_103; add.s64 %rd56, %rd48, %rd55; and.b32 %r330, %r16, -4096; mov.u64 %rd337, 1; st.v2.u32 [%rd41+24], {%r18, %r330}; and.b32 %r331, %r10, -4096; st.u32 [%rd41+32], %r331; st.u64 [%rd41+64], %rd56; shl.b32 %r332, %r1, 3; and.b32 %r333, %r332, 31744; and.b32 %r334, %r16, 3968; shr.u32 %r335, %r334, 2; or.b32 %r336, %r335, %r333; bfe.u32 %r337, %r10, 7, 5; or.b32 %r338, %r336, %r337; shr.u32 %r339, %r336, 6; mul.wide.u32 %rd338, %r339, 8; add.s64 %rd339, %rd56, %rd338; and.b32 %r340, %r338, 63; shl.b64 %rd340, %rd337, %r340; ld.u64 %rd341, [%rd339+4128]; and.b64 %rd342, %rd341, %rd340; setp.eq.s64 %p65, %rd342, 0; mul.wide.u32 %rd343, %r338, 16; add.s64 %rd344, %rd56, %rd343; add.s64 %rd57, %rd344, 8256; @%p65 bra $L__BB10_102; ld.u64 %rd345, [%rd57]; add.s64 %rd58, %rd56, %rd345; st.u32 [%rd41+12], %r17; and.b32 %r341, %r16, -128; st.u32 [%rd41+16], %r341; and.b32 %r342, %r10, -128; st.u32 [%rd41+20], %r342; st.u64 [%rd41+56], %rd58; shl.b32 %r343, %r1, 5; and.b32 %r344, %r343, 3840; shl.b32 %r345, %r16, 1; and.b32 %r346, %r345, 240; or.b32 %r347, %r346, %r344; bfe.u32 %r348, %r10, 3, 4; or.b32 %r349, %r347, %r348; shr.u32 %r350, %r347, 6; mul.wide.u32 %rd346, %r350, 8; add.s64 %rd347, %rd58, %rd346; and.b32 %r351, %r349, 63; shl.b64 %rd349, %rd337, %r351; ld.u64 %rd350, [%rd347+544]; and.b64 %rd351, %rd350, %rd349; setp.eq.s64 %p66, %rd351, 0; mul.wide.u32 %rd352, %r349, 16; add.s64 %rd353, %rd58, %rd352; add.s64 %rd59, %rd353, 1088; @%p66 bra $L__BB10_101; ld.u64 %rd354, [%rd59]; add.s64 %rd355, %rd58, %rd354; and.b32 %r352, %r16, -8; st.v2.u32 [%rd41], {%r4, %r352}; and.b32 %r353, %r10, -8; st.u32 [%rd41+8], %r353; st.u64 [%rd41+48], %rd355; shl.b32 %r354, %r1, 6; and.b32 %r355, %r354, 448; shl.b32 %r356, %r16, 3; and.b32 %r357, %r356, 56; or.b32 %r358, %r357, %r355; and.b32 %r359, %r10, 7; or.b32 %r360, %r358, %r359; mul.wide.u32 %rd356, %r360, 12; add.s64 %rd357, %rd355, %rd356; ld.f32 %f348, [%rd357+128]; ld.f32 %f347, [%rd357+132]; ld.f32 %f346, [%rd357+136]; bra.uni $L__BB10_105; $L__BB10_104: ld.f32 %f348, [%rd48+28]; ld.f32 %f347, [%rd48+32]; ld.f32 %f346, [%rd48+36]; bra.uni $L__BB10_105; $L__BB10_130: shl.b64 %rd403, %rd70, 5; add.s64 %rd698, %rd68, %rd403; $L__BB10_131: setp.eq.s64 %p85, %rd698, 0; @%p85 bra $L__BB10_139; add.s64 %rd73, %rd698, 8; ld.u64 %rd74, [%rd698+8]; setp.eq.s64 %p86, %rd74, 0; @%p86 bra $L__BB10_138; add.s64 %rd75, %rd67, %rd74; and.b32 %r431, %r16, -4096; mov.u64 %rd404, 1; st.v2.u32 [%rd60+24], {%r23, %r431}; and.b32 %r432, %r3, -4096; st.u32 [%rd60+32], %r432; st.u64 [%rd60+64], %rd75; shl.b32 %r433, %r1, 3; and.b32 %r434, %r433, 31744; and.b32 %r435, %r16, 3968; shr.u32 %r436, %r435, 2; or.b32 %r437, %r436, %r434; bfe.u32 %r438, %r3, 7, 5; or.b32 %r439, %r437, %r438; shr.u32 %r440, %r437, 6; mul.wide.u32 %rd405, %r440, 8; add.s64 %rd406, %rd75, %rd405; and.b32 %r441, %r439, 63; shl.b64 %rd407, %rd404, %r441; ld.u64 %rd408, [%rd406+4128]; and.b64 %rd409, %rd408, %rd407; setp.eq.s64 %p87, %rd409, 0; mul.wide.u32 %rd410, %r439, 16; add.s64 %rd411, %rd75, %rd410; add.s64 %rd76, %rd411, 8256; @%p87 bra $L__BB10_137; ld.u64 %rd412, [%rd76]; add.s64 %rd77, %rd75, %rd412; st.u32 [%rd60+12], %r22; and.b32 %r442, %r16, -128; st.u32 [%rd60+16], %r442; and.b32 %r443, %r3, -128; st.u32 [%rd60+20], %r443; st.u64 [%rd60+56], %rd77; shl.b32 %r444, %r1, 5; and.b32 %r445, %r444, 3840; shl.b32 %r446, %r16, 1; and.b32 %r447, %r446, 240; or.b32 %r448, %r447, %r445; bfe.u32 %r449, %r3, 3, 4; or.b32 %r450, %r448, %r449; shr.u32 %r451, %r448, 6; mul.wide.u32 %rd413, %r451, 8; add.s64 %rd414, %rd77, %rd413; and.b32 %r452, %r450, 63; shl.b64 %rd416, %rd404, %r452; ld.u64 %rd417, [%rd414+544]; and.b64 %rd418, %rd417, %rd416; setp.eq.s64 %p88, %rd418, 0; mul.wide.u32 %rd419, %r450, 16; add.s64 %rd420, %rd77, %rd419; add.s64 %rd78, %rd420, 1088; @%p88 bra $L__BB10_136; ld.u64 %rd421, [%rd78]; add.s64 %rd422, %rd77, %rd421; and.b32 %r453, %r16, -8; st.v2.u32 [%rd60], {%r4, %r453}; and.b32 %r454, %r3, -8; st.u32 [%rd60+8], %r454; st.u64 [%rd60+48], %rd422; and.b32 %r455, %r1, 7; shl.b32 %r456, %r16, 3; and.b32 %r457, %r456, 56; and.b32 %r458, %r3, 7; bfi.b32 %r459, %r455, %r458, 6, 3; or.b32 %r460, %r459, %r457; mul.wide.u32 %rd423, %r460, 12; add.s64 %rd424, %rd422, %rd423; ld.f32 %f351, [%rd424+128]; ld.f32 %f350, [%rd424+132]; ld.f32 %f349, [%rd424+136]; bra.uni $L__BB10_140; $L__BB10_139: ld.f32 %f351, [%rd67+28]; ld.f32 %f350, [%rd67+32]; ld.f32 %f349, [%rd67+36]; bra.uni $L__BB10_140; $L__BB10_165: shl.b64 %rd470, %rd89, 5; add.s64 %rd699, %rd87, %rd470; $L__BB10_166: setp.eq.s64 %p107, %rd699, 0; @%p107 bra $L__BB10_174; add.s64 %rd92, %rd699, 8; ld.u64 %rd93, [%rd699+8]; setp.eq.s64 %p108, %rd93, 0; @%p108 bra $L__BB10_173; add.s64 %rd94, %rd86, %rd93; and.b32 %r531, %r16, -4096; mov.u64 %rd471, 1; st.v2.u32 [%rd79+24], {%r30, %r531}; and.b32 %r532, %r3, -4096; st.u32 [%rd79+32], %r532; st.u64 [%rd79+64], %rd94; shl.b32 %r533, %r27, 3; and.b32 %r534, %r533, 31744; and.b32 %r535, %r16, 3968; shr.u32 %r536, %r535, 2; or.b32 %r537, %r536, %r534; bfe.u32 %r538, %r3, 7, 5; or.b32 %r539, %r537, %r538; shr.u32 %r540, %r537, 6; mul.wide.u32 %rd472, %r540, 8; add.s64 %rd473, %rd94, %rd472; and.b32 %r541, %r539, 63; shl.b64 %rd474, %rd471, %r541; ld.u64 %rd475, [%rd473+4128]; and.b64 %rd476, %rd475, %rd474; setp.eq.s64 %p109, %rd476, 0; mul.wide.u32 %rd477, %r539, 16; add.s64 %rd478, %rd94, %rd477; add.s64 %rd95, %rd478, 8256; @%p109 bra $L__BB10_172; ld.u64 %rd479, [%rd95]; add.s64 %rd96, %rd94, %rd479; st.u32 [%rd79+12], %r29; and.b32 %r542, %r16, -128; st.u32 [%rd79+16], %r542; and.b32 %r543, %r3, -128; st.u32 [%rd79+20], %r543; st.u64 [%rd79+56], %rd96; shl.b32 %r544, %r27, 5; and.b32 %r545, %r544, 3840; shl.b32 %r546, %r16, 1; and.b32 %r547, %r546, 240; or.b32 %r548, %r547, %r545; bfe.u32 %r549, %r3, 3, 4; or.b32 %r550, %r548, %r549; shr.u32 %r551, %r548, 6; mul.wide.u32 %rd480, %r551, 8; add.s64 %rd481, %rd96, %rd480; and.b32 %r552, %r550, 63; shl.b64 %rd483, %rd471, %r552; ld.u64 %rd484, [%rd481+544]; and.b64 %rd485, %rd484, %rd483; setp.eq.s64 %p110, %rd485, 0; mul.wide.u32 %rd486, %r550, 16; add.s64 %rd487, %rd96, %rd486; add.s64 %rd97, %rd487, 1088; @%p110 bra $L__BB10_171; ld.u64 %rd488, [%rd97]; add.s64 %rd489, %rd96, %rd488; and.b32 %r553, %r16, -8; st.v2.u32 [%rd79], {%r28, %r553}; and.b32 %r554, %r3, -8; st.u32 [%rd79+8], %r554; st.u64 [%rd79+48], %rd489; and.b32 %r555, %r27, 7; shl.b32 %r556, %r16, 3; and.b32 %r557, %r556, 56; and.b32 %r558, %r3, 7; bfi.b32 %r559, %r555, %r558, 6, 3; or.b32 %r560, %r559, %r557; mul.wide.u32 %rd490, %r560, 12; add.s64 %rd491, %rd489, %rd490; ld.f32 %f354, [%rd491+128]; ld.f32 %f353, [%rd491+132]; ld.f32 %f352, [%rd491+136]; bra.uni $L__BB10_175; $L__BB10_174: ld.f32 %f354, [%rd86+28]; ld.f32 %f353, [%rd86+32]; ld.f32 %f352, [%rd86+36]; bra.uni $L__BB10_175; $L__BB10_200: shl.b64 %rd537, %rd108, 5; add.s64 %rd700, %rd106, %rd537; $L__BB10_201: setp.eq.s64 %p129, %rd700, 0; @%p129 bra $L__BB10_209; add.s64 %rd111, %rd700, 8; ld.u64 %rd112, [%rd700+8]; setp.eq.s64 %p130, %rd112, 0; @%p130 bra $L__BB10_208; add.s64 %rd113, %rd105, %rd112; and.b32 %r634, %r16, -4096; mov.u64 %rd538, 1; st.v2.u32 [%rd98+24], {%r35, %r634}; and.b32 %r635, %r10, -4096; st.u32 [%rd98+32], %r635; st.u64 [%rd98+64], %rd113; shl.b32 %r636, %r27, 3; and.b32 %r637, %r636, 31744; and.b32 %r638, %r16, 3968; shr.u32 %r639, %r638, 2; or.b32 %r640, %r639, %r637; bfe.u32 %r641, %r10, 7, 5; or.b32 %r642, %r640, %r641; shr.u32 %r643, %r640, 6; mul.wide.u32 %rd539, %r643, 8; add.s64 %rd540, %rd113, %rd539; and.b32 %r644, %r642, 63; shl.b64 %rd541, %rd538, %r644; ld.u64 %rd542, [%rd540+4128]; and.b64 %rd543, %rd542, %rd541; setp.eq.s64 %p131, %rd543, 0; mul.wide.u32 %rd544, %r642, 16; add.s64 %rd545, %rd113, %rd544; add.s64 %rd114, %rd545, 8256; @%p131 bra $L__BB10_207; ld.u64 %rd546, [%rd114]; add.s64 %rd115, %rd113, %rd546; st.u32 [%rd98+12], %r34; and.b32 %r645, %r16, -128; st.u32 [%rd98+16], %r645; and.b32 %r646, %r10, -128; st.u32 [%rd98+20], %r646; st.u64 [%rd98+56], %rd115; shl.b32 %r647, %r27, 5; and.b32 %r648, %r647, 3840; shl.b32 %r649, %r16, 1; and.b32 %r650, %r649, 240; or.b32 %r651, %r650, %r648; bfe.u32 %r652, %r10, 3, 4; or.b32 %r653, %r651, %r652; shr.u32 %r654, %r651, 6; mul.wide.u32 %rd547, %r654, 8; add.s64 %rd548, %rd115, %rd547; and.b32 %r655, %r653, 63; shl.b64 %rd550, %rd538, %r655; ld.u64 %rd551, [%rd548+544]; and.b64 %rd552, %rd551, %rd550; setp.eq.s64 %p132, %rd552, 0; mul.wide.u32 %rd553, %r653, 16; add.s64 %rd554, %rd115, %rd553; add.s64 %rd116, %rd554, 1088; @%p132 bra $L__BB10_206; ld.u64 %rd555, [%rd116]; add.s64 %rd556, %rd115, %rd555; and.b32 %r656, %r16, -8; st.v2.u32 [%rd98], {%r28, %r656}; and.b32 %r657, %r10, -8; st.u32 [%rd98+8], %r657; st.u64 [%rd98+48], %rd556; shl.b32 %r658, %r27, 6; and.b32 %r659, %r658, 448; shl.b32 %r660, %r16, 3; and.b32 %r661, %r660, 56; or.b32 %r662, %r661, %r659; and.b32 %r663, %r10, 7; or.b32 %r664, %r662, %r663; mul.wide.u32 %rd557, %r664, 12; add.s64 %rd558, %rd556, %rd557; ld.f32 %f357, [%rd558+128]; ld.f32 %f356, [%rd558+132]; ld.f32 %f355, [%rd558+136]; bra.uni $L__BB10_210; $L__BB10_209: ld.f32 %f357, [%rd105+28]; ld.f32 %f356, [%rd105+32]; ld.f32 %f355, [%rd105+36]; bra.uni $L__BB10_210; $L__BB10_235: shl.b64 %rd605, %rd127, 5; add.s64 %rd701, %rd125, %rd605; $L__BB10_236: setp.eq.s64 %p151, %rd701, 0; @%p151 bra $L__BB10_244; add.s64 %rd130, %rd701, 8; ld.u64 %rd131, [%rd701+8]; setp.eq.s64 %p152, %rd131, 0; @%p152 bra $L__BB10_243; add.s64 %rd132, %rd124, %rd131; and.b32 %r738, %r2, -4096; mov.u64 %rd606, 1; st.v2.u32 [%rd117+24], {%r40, %r738}; and.b32 %r739, %r10, -4096; st.u32 [%rd117+32], %r739; st.u64 [%rd117+64], %rd132; shl.b32 %r740, %r27, 3; and.b32 %r741, %r740, 31744; and.b32 %r742, %r2, 3968; shr.u32 %r743, %r742, 2; or.b32 %r744, %r741, %r743; bfe.u32 %r745, %r10, 7, 5; or.b32 %r746, %r745, %r744; shr.u32 %r747, %r744, 6; mul.wide.u32 %rd607, %r747, 8; add.s64 %rd608, %rd132, %rd607; and.b32 %r748, %r746, 63; shl.b64 %rd609, %rd606, %r748; ld.u64 %rd610, [%rd608+4128]; and.b64 %rd611, %rd610, %rd609; setp.eq.s64 %p153, %rd611, 0; mul.wide.u32 %rd612, %r746, 16; add.s64 %rd613, %rd132, %rd612; add.s64 %rd133, %rd613, 8256; @%p153 bra $L__BB10_242; ld.u64 %rd614, [%rd133]; add.s64 %rd134, %rd132, %rd614; st.u32 [%rd117+12], %r39; and.b32 %r749, %r2, -128; st.u32 [%rd117+16], %r749; and.b32 %r750, %r10, -128; st.u32 [%rd117+20], %r750; st.u64 [%rd117+56], %rd134; shl.b32 %r751, %r27, 5; and.b32 %r752, %r751, 3840; shl.b32 %r753, %r2, 1; and.b32 %r754, %r753, 240; or.b32 %r755, %r752, %r754; bfe.u32 %r756, %r10, 3, 4; or.b32 %r757, %r756, %r755; shr.u32 %r758, %r755, 6; mul.wide.u32 %rd615, %r758, 8; add.s64 %rd616, %rd134, %rd615; and.b32 %r759, %r757, 63; shl.b64 %rd618, %rd606, %r759; ld.u64 %rd619, [%rd616+544]; and.b64 %rd620, %rd619, %rd618; setp.eq.s64 %p154, %rd620, 0; mul.wide.u32 %rd621, %r757, 16; add.s64 %rd622, %rd134, %rd621; add.s64 %rd135, %rd622, 1088; @%p154 bra $L__BB10_241; ld.u64 %rd623, [%rd135]; add.s64 %rd624, %rd134, %rd623; and.b32 %r760, %r2, -8; st.v2.u32 [%rd117], {%r28, %r760}; and.b32 %r761, %r10, -8; st.u32 [%rd117+8], %r761; st.u64 [%rd117+48], %rd624; shl.b32 %r762, %r27, 6; and.b32 %r763, %r762, 448; shl.b32 %r764, %r2, 3; and.b32 %r765, %r764, 56; or.b32 %r766, %r763, %r765; and.b32 %r767, %r10, 7; or.b32 %r768, %r766, %r767; mul.wide.u32 %rd625, %r768, 12; add.s64 %rd626, %rd624, %rd625; ld.f32 %f360, [%rd626+128]; ld.f32 %f359, [%rd626+132]; ld.f32 %f358, [%rd626+136]; bra.uni $L__BB10_245; $L__BB10_244: ld.f32 %f360, [%rd124+28]; ld.f32 %f359, [%rd124+32]; ld.f32 %f358, [%rd124+36]; bra.uni $L__BB10_245; $L__BB10_33: ld.f32 %f342, [%rd16+12]; ld.f32 %f341, [%rd16+16]; ld.f32 %f340, [%rd16+20]; bra.uni $L__BB10_35; $L__BB10_68: ld.f32 %f345, [%rd35+12]; ld.f32 %f344, [%rd35+16]; ld.f32 %f343, [%rd35+20]; bra.uni $L__BB10_70; $L__BB10_103: ld.f32 %f348, [%rd54+12]; ld.f32 %f347, [%rd54+16]; ld.f32 %f346, [%rd54+20]; bra.uni $L__BB10_105; $L__BB10_138: ld.f32 %f351, [%rd73+12]; ld.f32 %f350, [%rd73+16]; ld.f32 %f349, [%rd73+20]; bra.uni $L__BB10_140; $L__BB10_173: ld.f32 %f354, [%rd92+12]; ld.f32 %f353, [%rd92+16]; ld.f32 %f352, [%rd92+20]; bra.uni $L__BB10_175; $L__BB10_208: ld.f32 %f357, [%rd111+12]; ld.f32 %f356, [%rd111+16]; ld.f32 %f355, [%rd111+20]; bra.uni $L__BB10_210; $L__BB10_243: ld.f32 %f360, [%rd130+12]; ld.f32 %f359, [%rd130+16]; ld.f32 %f358, [%rd130+20]; bra.uni $L__BB10_245; $L__BB10_32: ld.f32 %f342, [%rd19]; ld.f32 %f341, [%rd19+4]; ld.f32 %f340, [%rd19+8]; bra.uni $L__BB10_35; $L__BB10_67: ld.f32 %f345, [%rd38]; ld.f32 %f344, [%rd38+4]; ld.f32 %f343, [%rd38+8]; bra.uni $L__BB10_70; $L__BB10_102: ld.f32 %f348, [%rd57]; ld.f32 %f347, [%rd57+4]; ld.f32 %f346, [%rd57+8]; bra.uni $L__BB10_105; $L__BB10_137: ld.f32 %f351, [%rd76]; ld.f32 %f350, [%rd76+4]; ld.f32 %f349, [%rd76+8]; bra.uni $L__BB10_140; $L__BB10_172: ld.f32 %f354, [%rd95]; ld.f32 %f353, [%rd95+4]; ld.f32 %f352, [%rd95+8]; bra.uni $L__BB10_175; $L__BB10_207: ld.f32 %f357, [%rd114]; ld.f32 %f356, [%rd114+4]; ld.f32 %f355, [%rd114+8]; bra.uni $L__BB10_210; $L__BB10_242: ld.f32 %f360, [%rd133]; ld.f32 %f359, [%rd133+4]; ld.f32 %f358, [%rd133+8]; bra.uni $L__BB10_245; $L__BB10_31: ld.f32 %f342, [%rd21]; ld.f32 %f341, [%rd21+4]; ld.f32 %f340, [%rd21+8]; bra.uni $L__BB10_35; $L__BB10_66: ld.f32 %f345, [%rd40]; ld.f32 %f344, [%rd40+4]; ld.f32 %f343, [%rd40+8]; bra.uni $L__BB10_70; $L__BB10_101: ld.f32 %f348, [%rd59]; ld.f32 %f347, [%rd59+4]; ld.f32 %f346, [%rd59+8]; bra.uni $L__BB10_105; $L__BB10_136: ld.f32 %f351, [%rd78]; ld.f32 %f350, [%rd78+4]; ld.f32 %f349, [%rd78+8]; bra.uni $L__BB10_140; $L__BB10_171: ld.f32 %f354, [%rd97]; ld.f32 %f353, [%rd97+4]; ld.f32 %f352, [%rd97+8]; bra.uni $L__BB10_175; $L__BB10_206: ld.f32 %f357, [%rd116]; ld.f32 %f356, [%rd116+4]; ld.f32 %f355, [%rd116+8]; bra.uni $L__BB10_210; $L__BB10_241: ld.f32 %f360, [%rd135]; ld.f32 %f359, [%rd135+4]; ld.f32 %f358, [%rd135+8]; bra.uni $L__BB10_245; } .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) { .reg .pred %p<174>; .reg .b16 %rs<170>; .reg .f32 %f<288>; .reg .b32 %r<1425>; .reg .b64 %rd<708>; ld.param.u64 %rd160, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0]; cvta.to.local.u64 %rd1, %rd160; ld.param.f32 %f103, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+8]; ld.param.f32 %f104, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+4]; ld.param.f32 %f105, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_3FpNELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1]; cvt.rmi.ftz.f32.f32 %f106, %f105; cvt.rmi.ftz.f32.f32 %f107, %f104; cvt.rmi.ftz.f32.f32 %f108, %f103; sub.ftz.f32 %f1, %f105, %f106; sub.ftz.f32 %f2, %f104, %f107; sub.ftz.f32 %f3, %f103, %f108; cvt.rzi.ftz.s32.f32 %r1, %f106; cvt.rzi.ftz.s32.f32 %r2, %f107; cvt.rzi.ftz.s32.f32 %r3, %f108; and.b32 %r4, %r1, -8; ld.local.u64 %rd2, [%rd1]; ld.u32 %r49, [%rd2]; setp.ne.s32 %p1, %r4, %r49; mov.u16 %rs149, 0; @%p1 bra $L__BB11_3; and.b32 %r50, %r2, -8; ld.u32 %r51, [%rd2+4]; setp.ne.s32 %p2, %r50, %r51; @%p2 bra $L__BB11_3; and.b32 %r52, %r3, -8; ld.u32 %r53, [%rd2+8]; setp.eq.s32 %p3, %r52, %r53; selp.u16 %rs149, 1, 0, %p3; $L__BB11_3: setp.eq.s16 %p4, %rs149, 0; @%p4 bra $L__BB11_5; ld.u64 %rd161, [%rd2+48]; shl.b32 %r54, %r1, 6; and.b32 %r55, %r54, 448; shl.b32 %r56, %r2, 3; and.b32 %r57, %r56, 56; or.b32 %r58, %r57, %r55; and.b32 %r59, %r3, 7; or.b32 %r60, %r58, %r59; ld.u8 %rs45, [%rd161+15]; shr.u16 %rs46, %rs45, 5; mov.u32 %r61, 5; cvt.u32.u16 %r62, %rs46; sub.s32 %r63, %r61, %r62; shr.u32 %r64, %r60, %r63; mul.wide.u32 %rd162, %r64, 4; add.s64 %rd163, %rd161, %rd162; mov.u32 %r65, 32; shr.u32 %r66, %r65, %r62; add.s32 %r67, %r66, -1; and.b32 %r68, %r67, %r60; shl.b32 %r69, %r68, %r62; ld.u32 %r70, [%rd163+96]; shr.u32 %r71, %r70, %r69; mov.u32 %r72, 1; shl.b32 %r73, %r72, %r62; mov.u32 %r74, -1; shl.b32 %r75, %r74, %r73; not.b32 %r76, %r75; and.b32 %r77, %r71, %r76; cvt.rn.f32.u32 %f109, %r77; ld.v2.f32 {%f110, %f111}, [%rd161+80]; fma.rn.ftz.f32 %f280, %f111, %f109, %f110; $L__BB11_35: add.s32 %r10, %r3, 1; ld.local.u64 %rd22, [%rd1]; ld.u32 %r221, [%rd22]; setp.ne.s32 %p23, %r4, %r221; mov.u16 %rs152, 0; @%p23 bra $L__BB11_38; and.b32 %r222, %r2, -8; ld.u32 %r223, [%rd22+4]; setp.ne.s32 %p24, %r222, %r223; @%p24 bra $L__BB11_38; and.b32 %r224, %r10, -8; ld.u32 %r225, [%rd22+8]; setp.eq.s32 %p25, %r224, %r225; selp.u16 %rs152, 1, 0, %p25; $L__BB11_38: setp.eq.s16 %p26, %rs152, 0; @%p26 bra $L__BB11_40; ld.u64 %rd228, [%rd22+48]; shl.b32 %r226, %r1, 6; and.b32 %r227, %r226, 448; shl.b32 %r228, %r2, 3; and.b32 %r229, %r228, 56; or.b32 %r230, %r229, %r227; and.b32 %r231, %r10, 7; or.b32 %r232, %r230, %r231; ld.u8 %rs59, [%rd228+15]; shr.u16 %rs60, %rs59, 5; mov.u32 %r233, 5; cvt.u32.u16 %r234, %rs60; sub.s32 %r235, %r233, %r234; shr.u32 %r236, %r232, %r235; mul.wide.u32 %rd229, %r236, 4; add.s64 %rd230, %rd228, %rd229; mov.u32 %r237, 32; shr.u32 %r238, %r237, %r234; add.s32 %r239, %r238, -1; and.b32 %r240, %r239, %r232; shl.b32 %r241, %r240, %r234; ld.u32 %r242, [%rd230+96]; shr.u32 %r243, %r242, %r241; mov.u32 %r244, 1; shl.b32 %r245, %r244, %r234; mov.u32 %r246, -1; shl.b32 %r247, %r246, %r245; not.b32 %r248, %r247; and.b32 %r249, %r243, %r248; cvt.rn.f32.u32 %f129, %r249; ld.v2.f32 {%f130, %f131}, [%rd228+80]; fma.rn.ftz.f32 %f281, %f131, %f129, %f130; $L__BB11_70: sub.ftz.f32 %f149, %f281, %f280; fma.rn.ftz.f32 %f28, %f3, %f149, %f280; add.s32 %r16, %r2, 1; ld.local.u64 %rd42, [%rd1]; ld.u32 %r393, [%rd42]; setp.ne.s32 %p45, %r4, %r393; mov.u16 %rs155, 0; @%p45 bra $L__BB11_73; and.b32 %r394, %r16, -8; ld.u32 %r395, [%rd42+4]; setp.ne.s32 %p46, %r394, %r395; @%p46 bra $L__BB11_73; and.b32 %r396, %r10, -8; ld.u32 %r397, [%rd42+8]; setp.eq.s32 %p47, %r396, %r397; selp.u16 %rs155, 1, 0, %p47; $L__BB11_73: setp.eq.s16 %p48, %rs155, 0; @%p48 bra $L__BB11_75; ld.u64 %rd296, [%rd42+48]; shl.b32 %r398, %r1, 6; and.b32 %r399, %r398, 448; shl.b32 %r400, %r16, 3; and.b32 %r401, %r400, 56; or.b32 %r402, %r401, %r399; and.b32 %r403, %r10, 7; or.b32 %r404, %r402, %r403; ld.u8 %rs73, [%rd296+15]; shr.u16 %rs74, %rs73, 5; mov.u32 %r405, 5; cvt.u32.u16 %r406, %rs74; sub.s32 %r407, %r405, %r406; shr.u32 %r408, %r404, %r407; mul.wide.u32 %rd297, %r408, 4; add.s64 %rd298, %rd296, %rd297; mov.u32 %r409, 32; shr.u32 %r410, %r409, %r406; add.s32 %r411, %r410, -1; and.b32 %r412, %r411, %r404; shl.b32 %r413, %r412, %r406; ld.u32 %r414, [%rd298+96]; shr.u32 %r415, %r414, %r413; mov.u32 %r416, 1; shl.b32 %r417, %r416, %r406; mov.u32 %r418, -1; shl.b32 %r419, %r418, %r417; not.b32 %r420, %r419; and.b32 %r421, %r415, %r420; cvt.rn.f32.u32 %f150, %r421; ld.v2.f32 {%f151, %f152}, [%rd296+80]; fma.rn.ftz.f32 %f282, %f152, %f150, %f151; $L__BB11_105: ld.local.u64 %rd62, [%rd1]; ld.u32 %r565, [%rd62]; setp.ne.s32 %p67, %r4, %r565; mov.u16 %rs158, 0; @%p67 bra $L__BB11_108; and.b32 %r566, %r16, -8; ld.u32 %r567, [%rd62+4]; setp.ne.s32 %p68, %r566, %r567; @%p68 bra $L__BB11_108; and.b32 %r568, %r3, -8; ld.u32 %r569, [%rd62+8]; setp.eq.s32 %p69, %r568, %r569; selp.u16 %rs158, 1, 0, %p69; $L__BB11_108: setp.eq.s16 %p70, %rs158, 0; @%p70 bra $L__BB11_110; ld.u64 %rd363, [%rd62+48]; and.b32 %r570, %r1, 7; shl.b32 %r571, %r16, 3; and.b32 %r572, %r571, 56; and.b32 %r573, %r3, 7; bfi.b32 %r574, %r570, %r573, 6, 3; or.b32 %r575, %r574, %r572; ld.u8 %rs87, [%rd363+15]; shr.u16 %rs88, %rs87, 5; mov.u32 %r576, 5; cvt.u32.u16 %r577, %rs88; sub.s32 %r578, %r576, %r577; shr.u32 %r579, %r575, %r578; mul.wide.u32 %rd364, %r579, 4; add.s64 %rd365, %rd363, %rd364; mov.u32 %r580, 32; shr.u32 %r581, %r580, %r577; add.s32 %r582, %r581, -1; and.b32 %r583, %r582, %r575; shl.b32 %r584, %r583, %r577; ld.u32 %r585, [%rd365+96]; shr.u32 %r586, %r585, %r584; mov.u32 %r587, 1; shl.b32 %r588, %r587, %r577; mov.u32 %r589, -1; shl.b32 %r590, %r589, %r588; not.b32 %r591, %r590; and.b32 %r592, %r586, %r591; cvt.rn.f32.u32 %f170, %r592; ld.v2.f32 {%f171, %f172}, [%rd363+80]; fma.rn.ftz.f32 %f283, %f172, %f170, %f171; $L__BB11_140: sub.ftz.f32 %f190, %f282, %f283; fma.rn.ftz.f32 %f191, %f3, %f190, %f283; sub.ftz.f32 %f192, %f191, %f28; fma.rn.ftz.f32 %f53, %f2, %f192, %f28; add.s32 %r27, %r1, 1; and.b32 %r28, %r27, -8; ld.local.u64 %rd81, [%rd1]; ld.u32 %r733, [%rd81]; setp.ne.s32 %p89, %r28, %r733; mov.u16 %rs161, 0; @%p89 bra $L__BB11_143; and.b32 %r734, %r16, -8; ld.u32 %r735, [%rd81+4]; setp.ne.s32 %p90, %r734, %r735; @%p90 bra $L__BB11_143; and.b32 %r736, %r3, -8; ld.u32 %r737, [%rd81+8]; setp.eq.s32 %p91, %r736, %r737; selp.u16 %rs161, 1, 0, %p91; $L__BB11_143: setp.eq.s16 %p92, %rs161, 0; @%p92 bra $L__BB11_145; ld.u64 %rd430, [%rd81+48]; and.b32 %r738, %r27, 7; shl.b32 %r739, %r16, 3; and.b32 %r740, %r739, 56; and.b32 %r741, %r3, 7; bfi.b32 %r742, %r738, %r741, 6, 3; or.b32 %r743, %r742, %r740; ld.u8 %rs101, [%rd430+15]; shr.u16 %rs102, %rs101, 5; mov.u32 %r744, 5; cvt.u32.u16 %r745, %rs102; sub.s32 %r746, %r744, %r745; shr.u32 %r747, %r743, %r746; mul.wide.u32 %rd431, %r747, 4; add.s64 %rd432, %rd430, %rd431; mov.u32 %r748, 32; shr.u32 %r749, %r748, %r745; add.s32 %r750, %r749, -1; and.b32 %r751, %r750, %r743; shl.b32 %r752, %r751, %r745; ld.u32 %r753, [%rd432+96]; shr.u32 %r754, %r753, %r752; mov.u32 %r755, 1; shl.b32 %r756, %r755, %r745; mov.u32 %r757, -1; shl.b32 %r758, %r757, %r756; not.b32 %r759, %r758; and.b32 %r760, %r754, %r759; cvt.rn.f32.u32 %f193, %r760; ld.v2.f32 {%f194, %f195}, [%rd430+80]; fma.rn.ftz.f32 %f284, %f195, %f193, %f194; $L__BB11_175: ld.local.u64 %rd101, [%rd1]; ld.u32 %r901, [%rd101]; setp.ne.s32 %p111, %r28, %r901; mov.u16 %rs164, 0; @%p111 bra $L__BB11_178; and.b32 %r902, %r16, -8; ld.u32 %r903, [%rd101+4]; setp.ne.s32 %p112, %r902, %r903; @%p112 bra $L__BB11_178; and.b32 %r904, %r10, -8; ld.u32 %r905, [%rd101+8]; setp.eq.s32 %p113, %r904, %r905; selp.u16 %rs164, 1, 0, %p113; $L__BB11_178: setp.eq.s16 %p114, %rs164, 0; @%p114 bra $L__BB11_180; ld.u64 %rd497, [%rd101+48]; shl.b32 %r906, %r27, 6; and.b32 %r907, %r906, 448; shl.b32 %r908, %r16, 3; and.b32 %r909, %r908, 56; or.b32 %r910, %r909, %r907; and.b32 %r911, %r10, 7; or.b32 %r912, %r910, %r911; ld.u8 %rs115, [%rd497+15]; shr.u16 %rs116, %rs115, 5; mov.u32 %r913, 5; cvt.u32.u16 %r914, %rs116; sub.s32 %r915, %r913, %r914; shr.u32 %r916, %r912, %r915; mul.wide.u32 %rd498, %r916, 4; add.s64 %rd499, %rd497, %rd498; mov.u32 %r917, 32; shr.u32 %r918, %r917, %r914; add.s32 %r919, %r918, -1; and.b32 %r920, %r919, %r912; shl.b32 %r921, %r920, %r914; ld.u32 %r922, [%rd499+96]; shr.u32 %r923, %r922, %r921; mov.u32 %r924, 1; shl.b32 %r925, %r924, %r914; mov.u32 %r926, -1; shl.b32 %r927, %r926, %r925; not.b32 %r928, %r927; and.b32 %r929, %r923, %r928; cvt.rn.f32.u32 %f213, %r929; ld.v2.f32 {%f214, %f215}, [%rd497+80]; fma.rn.ftz.f32 %f285, %f215, %f213, %f214; $L__BB11_210: sub.ftz.f32 %f233, %f285, %f284; fma.rn.ftz.f32 %f78, %f3, %f233, %f284; ld.local.u64 %rd120, [%rd1]; ld.u32 %r1073, [%rd120]; setp.ne.s32 %p133, %r28, %r1073; mov.u16 %rs167, 0; @%p133 bra $L__BB11_213; and.b32 %r1074, %r2, -8; ld.u32 %r1075, [%rd120+4]; setp.ne.s32 %p134, %r1074, %r1075; @%p134 bra $L__BB11_213; and.b32 %r1076, %r10, -8; ld.u32 %r1077, [%rd120+8]; setp.eq.s32 %p135, %r1076, %r1077; selp.u16 %rs167, 1, 0, %p135; $L__BB11_213: setp.eq.s16 %p136, %rs167, 0; @%p136 bra $L__BB11_215; ld.u64 %rd564, [%rd120+48]; shl.b32 %r1078, %r27, 6; and.b32 %r1079, %r1078, 448; shl.b32 %r1080, %r2, 3; and.b32 %r1081, %r1080, 56; or.b32 %r1082, %r1079, %r1081; and.b32 %r1083, %r10, 7; or.b32 %r1084, %r1082, %r1083; ld.u8 %rs129, [%rd564+15]; shr.u16 %rs130, %rs129, 5; mov.u32 %r1085, 5; cvt.u32.u16 %r1086, %rs130; sub.s32 %r1087, %r1085, %r1086; shr.u32 %r1088, %r1084, %r1087; mul.wide.u32 %rd565, %r1088, 4; add.s64 %rd566, %rd564, %rd565; mov.u32 %r1089, 32; shr.u32 %r1090, %r1089, %r1086; add.s32 %r1091, %r1090, -1; and.b32 %r1092, %r1091, %r1084; shl.b32 %r1093, %r1092, %r1086; ld.u32 %r1094, [%rd566+96]; shr.u32 %r1095, %r1094, %r1093; mov.u32 %r1096, 1; shl.b32 %r1097, %r1096, %r1086; mov.u32 %r1098, -1; shl.b32 %r1099, %r1098, %r1097; not.b32 %r1100, %r1099; and.b32 %r1101, %r1095, %r1100; cvt.rn.f32.u32 %f234, %r1101; ld.v2.f32 {%f235, %f236}, [%rd564+80]; fma.rn.ftz.f32 %f286, %f236, %f234, %f235; $L__BB11_245: ld.local.u64 %rd140, [%rd1]; ld.u32 %r1245, [%rd140]; setp.ne.s32 %p155, %r28, %r1245; @%p155 bra $L__BB11_248; and.b32 %r1246, %r2, -8; ld.u32 %r1247, [%rd140+4]; setp.ne.s32 %p156, %r1246, %r1247; @%p156 bra $L__BB11_248; and.b32 %r1248, %r3, -8; ld.u32 %r1249, [%rd140+8]; setp.eq.s32 %p157, %r1248, %r1249; @%p157 bra $L__BB11_276; bra.uni $L__BB11_248; $L__BB11_276: ld.u64 %rd697, [%rd140+48]; shl.b32 %r1393, %r27, 6; and.b32 %r1394, %r1393, 448; shl.b32 %r1395, %r2, 3; and.b32 %r1396, %r1395, 56; or.b32 %r1397, %r1394, %r1396; and.b32 %r1398, %r3, 7; or.b32 %r1399, %r1397, %r1398; ld.u8 %rs147, [%rd697+15]; shr.u16 %rs148, %rs147, 5; mov.u32 %r1400, 5; cvt.u32.u16 %r1401, %rs148; sub.s32 %r1402, %r1400, %r1401; shr.u32 %r1403, %r1399, %r1402; mul.wide.u32 %rd698, %r1403, 4; add.s64 %rd699, %rd697, %rd698; mov.u32 %r1404, 32; shr.u32 %r1405, %r1404, %r1401; add.s32 %r1406, %r1405, -1; and.b32 %r1407, %r1406, %r1399; shl.b32 %r1408, %r1407, %r1401; ld.u32 %r1409, [%rd699+96]; shr.u32 %r1410, %r1409, %r1408; mov.u32 %r1411, 1; shl.b32 %r1412, %r1411, %r1401; mov.u32 %r1413, -1; shl.b32 %r1414, %r1413, %r1412; not.b32 %r1415, %r1414; and.b32 %r1416, %r1410, %r1415; cvt.rn.f32.u32 %f269, %r1416; ld.v2.f32 {%f270, %f271}, [%rd697+80]; fma.rn.ftz.f32 %f287, %f271, %f269, %f270; bra.uni $L__BB11_277; $L__BB11_248: and.b32 %r44, %r27, -128; ld.u32 %r1250, [%rd140+12]; setp.ne.s32 %p158, %r44, %r1250; @%p158 bra $L__BB11_251; and.b32 %r1251, %r2, -128; ld.u32 %r1252, [%rd140+16]; setp.ne.s32 %p159, %r1251, %r1252; @%p159 bra $L__BB11_251; and.b32 %r1253, %r3, -128; ld.u32 %r1254, [%rd140+20]; setp.eq.s32 %p160, %r1253, %r1254; @%p160 bra $L__BB11_273; bra.uni $L__BB11_251; $L__BB11_273: shl.b32 %r1358, %r27, 5; and.b32 %r1359, %r1358, 3840; shl.b32 %r1360, %r2, 1; and.b32 %r1361, %r1360, 240; or.b32 %r1362, %r1359, %r1361; bfe.u32 %r1363, %r3, 3, 4; or.b32 %r1364, %r1362, %r1363; shr.u32 %r1365, %r1362, 6; ld.u64 %rd158, [%rd140+56]; mul.wide.u32 %rd685, %r1365, 8; add.s64 %rd686, %rd158, %rd685; and.b32 %r1366, %r1364, 63; mov.u64 %rd687, 1; shl.b64 %rd688, %rd687, %r1366; ld.u64 %rd689, [%rd686+544]; and.b64 %rd690, %rd689, %rd688; setp.eq.s64 %p173, %rd690, 0; mul.wide.u32 %rd691, %r1364, 8; add.s64 %rd692, %rd158, %rd691; add.s64 %rd159, %rd692, 1088; @%p173 bra $L__BB11_275; ld.u64 %rd693, [%rd159]; add.s64 %rd694, %rd158, %rd693; and.b32 %r1367, %r2, -8; mov.u32 %r1368, 1; st.v2.u32 [%rd140], {%r28, %r1367}; and.b32 %r1369, %r3, -8; st.u32 [%rd140+8], %r1369; st.u64 [%rd140+48], %rd694; shl.b32 %r1370, %r27, 6; and.b32 %r1371, %r1370, 448; shl.b32 %r1372, %r2, 3; and.b32 %r1373, %r1372, 56; or.b32 %r1374, %r1371, %r1373; and.b32 %r1375, %r3, 7; or.b32 %r1376, %r1374, %r1375; ld.u8 %rs145, [%rd694+15]; shr.u16 %rs146, %rs145, 5; mov.u32 %r1377, 5; cvt.u32.u16 %r1378, %rs146; sub.s32 %r1379, %r1377, %r1378; shr.u32 %r1380, %r1376, %r1379; mul.wide.u32 %rd695, %r1380, 4; add.s64 %rd696, %rd694, %rd695; mov.u32 %r1381, 32; shr.u32 %r1382, %r1381, %r1378; add.s32 %r1383, %r1382, -1; and.b32 %r1384, %r1383, %r1376; shl.b32 %r1385, %r1384, %r1378; ld.u32 %r1386, [%rd696+96]; shr.u32 %r1387, %r1386, %r1385; shl.b32 %r1388, %r1368, %r1378; mov.u32 %r1389, -1; shl.b32 %r1390, %r1389, %r1388; not.b32 %r1391, %r1390; and.b32 %r1392, %r1387, %r1391; cvt.rn.f32.u32 %f264, %r1392; ld.v2.f32 {%f265, %f266}, [%rd694+80]; fma.rn.ftz.f32 %f287, %f266, %f264, %f265; bra.uni $L__BB11_277; $L__BB11_251: and.b32 %r45, %r27, -4096; ld.u32 %r1255, [%rd140+24]; setp.ne.s32 %p161, %r45, %r1255; @%p161 bra $L__BB11_254; and.b32 %r1256, %r2, -4096; ld.u32 %r1257, [%rd140+28]; setp.ne.s32 %p162, %r1256, %r1257; @%p162 bra $L__BB11_254; and.b32 %r1258, %r3, -4096; ld.u32 %r1259, [%rd140+32]; setp.eq.s32 %p163, %r1258, %r1259; @%p163 bra $L__BB11_268; bra.uni $L__BB11_254; $L__BB11_268: shl.b32 %r1312, %r27, 3; and.b32 %r1313, %r1312, 31744; and.b32 %r1314, %r2, 3968; shr.u32 %r1315, %r1314, 2; or.b32 %r1316, %r1313, %r1315; bfe.u32 %r1317, %r3, 7, 5; or.b32 %r1318, %r1316, %r1317; shr.u32 %r1319, %r1316, 6; ld.u64 %rd154, [%rd140+64]; mul.wide.u32 %rd664, %r1319, 8; add.s64 %rd665, %rd154, %rd664; and.b32 %r1320, %r1318, 63; mov.u64 %rd666, 1; shl.b64 %rd667, %rd666, %r1320; ld.u64 %rd668, [%rd665+4128]; and.b64 %rd669, %rd668, %rd667; setp.eq.s64 %p171, %rd669, 0; mul.wide.u32 %rd670, %r1318, 8; add.s64 %rd671, %rd154, %rd670; add.s64 %rd155, %rd671, 8256; @%p171 bra $L__BB11_272; ld.u64 %rd672, [%rd155]; add.s64 %rd156, %rd154, %rd672; st.u32 [%rd140+12], %r44; and.b32 %r1321, %r2, -128; st.u32 [%rd140+16], %r1321; and.b32 %r1322, %r3, -128; st.u32 [%rd140+20], %r1322; st.u64 [%rd140+56], %rd156; shl.b32 %r1323, %r27, 5; and.b32 %r1324, %r1323, 3840; shl.b32 %r1325, %r2, 1; and.b32 %r1326, %r1325, 240; or.b32 %r1327, %r1324, %r1326; bfe.u32 %r1328, %r3, 3, 4; or.b32 %r1329, %r1327, %r1328; shr.u32 %r1330, %r1327, 6; mul.wide.u32 %rd673, %r1330, 8; add.s64 %rd674, %rd156, %rd673; and.b32 %r1331, %r1329, 63; shl.b64 %rd676, %rd666, %r1331; ld.u64 %rd677, [%rd674+544]; and.b64 %rd678, %rd677, %rd676; setp.eq.s64 %p172, %rd678, 0; mul.wide.u32 %rd679, %r1329, 8; add.s64 %rd680, %rd156, %rd679; add.s64 %rd157, %rd680, 1088; @%p172 bra $L__BB11_271; ld.u64 %rd681, [%rd157]; add.s64 %rd682, %rd156, %rd681; and.b32 %r1332, %r2, -8; mov.u32 %r1333, 1; st.v2.u32 [%rd140], {%r28, %r1332}; and.b32 %r1334, %r3, -8; st.u32 [%rd140+8], %r1334; st.u64 [%rd140+48], %rd682; shl.b32 %r1335, %r27, 6; and.b32 %r1336, %r1335, 448; shl.b32 %r1337, %r2, 3; and.b32 %r1338, %r1337, 56; or.b32 %r1339, %r1336, %r1338; and.b32 %r1340, %r3, 7; or.b32 %r1341, %r1339, %r1340; ld.u8 %rs143, [%rd682+15]; shr.u16 %rs144, %rs143, 5; mov.u32 %r1342, 5; cvt.u32.u16 %r1343, %rs144; sub.s32 %r1344, %r1342, %r1343; shr.u32 %r1345, %r1341, %r1344; mul.wide.u32 %rd683, %r1345, 4; add.s64 %rd684, %rd682, %rd683; mov.u32 %r1346, 32; shr.u32 %r1347, %r1346, %r1343; add.s32 %r1348, %r1347, -1; and.b32 %r1349, %r1348, %r1341; shl.b32 %r1350, %r1349, %r1343; ld.u32 %r1351, [%rd684+96]; shr.u32 %r1352, %r1351, %r1350; shl.b32 %r1353, %r1333, %r1343; mov.u32 %r1354, -1; shl.b32 %r1355, %r1354, %r1353; not.b32 %r1356, %r1355; and.b32 %r1357, %r1352, %r1356; cvt.rn.f32.u32 %f259, %r1357; ld.v2.f32 {%f260, %f261}, [%rd682+80]; fma.rn.ftz.f32 %f287, %f261, %f259, %f260; bra.uni $L__BB11_277; $L__BB11_254: ld.u64 %rd141, [%rd140+40]; mov.u64 %rd707, 0; add.s64 %rd142, %rd141, 24; ld.u32 %r46, [%rd141+24]; setp.eq.s32 %p164, %r46, 0; @%p164 bra $L__BB11_259; add.s64 %rd143, %rd141, 64; shr.u32 %r1261, %r27, 12; cvt.u64.u32 %rd633, %r1261; shl.b64 %rd634, %rd633, 42; shr.u32 %r1262, %r2, 12; mul.wide.u32 %rd635, %r1262, 2097152; shr.u32 %r1263, %r3, 12; cvt.u64.u32 %rd636, %r1263; or.b64 %rd637, %rd635, %rd636; or.b64 %rd144, %rd637, %rd634; mov.u32 %r1424, 0; $L__BB11_256: cvt.u64.u32 %rd145, %r1424; mul.wide.u32 %rd638, %r1424, 32; add.s64 %rd639, %rd143, %rd638; ld.u64 %rd640, [%rd639]; setp.eq.s64 %p165, %rd640, %rd144; add.s32 %r1424, %r1424, 1; @%p165 bra $L__BB11_258; setp.lt.u32 %p166, %r1424, %r46; @%p166 bra $L__BB11_256; bra.uni $L__BB11_259; $L__BB11_5: and.b32 %r5, %r1, -128; ld.u32 %r78, [%rd2+12]; setp.ne.s32 %p5, %r5, %r78; mov.u16 %rs150, 0; @%p5 bra $L__BB11_8; and.b32 %r79, %r2, -128; ld.u32 %r80, [%rd2+16]; setp.ne.s32 %p6, %r79, %r80; @%p6 bra $L__BB11_8; and.b32 %r81, %r3, -128; ld.u32 %r82, [%rd2+20]; setp.eq.s32 %p7, %r81, %r82; selp.u16 %rs150, 1, 0, %p7; $L__BB11_8: setp.eq.s16 %p8, %rs150, 0; @%p8 bra $L__BB11_12; shl.b32 %r83, %r1, 5; and.b32 %r84, %r83, 3840; shl.b32 %r85, %r2, 1; and.b32 %r86, %r85, 240; or.b32 %r87, %r86, %r84; bfe.u32 %r88, %r3, 3, 4; or.b32 %r89, %r87, %r88; shr.u32 %r90, %r87, 6; ld.u64 %rd3, [%rd2+56]; mul.wide.u32 %rd164, %r90, 8; add.s64 %rd165, %rd3, %rd164; and.b32 %r91, %r89, 63; mov.u64 %rd166, 1; shl.b64 %rd167, %rd166, %r91; ld.u64 %rd168, [%rd165+544]; and.b64 %rd169, %rd168, %rd167; setp.eq.s64 %p9, %rd169, 0; mul.wide.u32 %rd170, %r89, 8; add.s64 %rd171, %rd3, %rd170; add.s64 %rd4, %rd171, 1088; @%p9 bra $L__BB11_11; ld.u64 %rd172, [%rd4]; add.s64 %rd173, %rd3, %rd172; and.b32 %r92, %r2, -8; mov.u32 %r93, 1; st.v2.u32 [%rd2], {%r4, %r92}; and.b32 %r94, %r3, -8; st.u32 [%rd2+8], %r94; st.u64 [%rd2+48], %rd173; shl.b32 %r95, %r1, 6; and.b32 %r96, %r95, 448; shl.b32 %r97, %r2, 3; and.b32 %r98, %r97, 56; or.b32 %r99, %r98, %r96; and.b32 %r100, %r3, 7; or.b32 %r101, %r99, %r100; ld.u8 %rs49, [%rd173+15]; shr.u16 %rs50, %rs49, 5; mov.u32 %r102, 5; cvt.u32.u16 %r103, %rs50; sub.s32 %r104, %r102, %r103; shr.u32 %r105, %r101, %r104; mul.wide.u32 %rd174, %r105, 4; add.s64 %rd175, %rd173, %rd174; mov.u32 %r106, 32; shr.u32 %r107, %r106, %r103; add.s32 %r108, %r107, -1; and.b32 %r109, %r108, %r101; shl.b32 %r110, %r109, %r103; ld.u32 %r111, [%rd175+96]; shr.u32 %r112, %r111, %r110; shl.b32 %r113, %r93, %r103; mov.u32 %r114, -1; shl.b32 %r115, %r114, %r113; not.b32 %r116, %r115; and.b32 %r117, %r112, %r116; cvt.rn.f32.u32 %f114, %r117; ld.v2.f32 {%f115, %f116}, [%rd173+80]; fma.rn.ftz.f32 %f280, %f116, %f114, %f115; bra.uni $L__BB11_35; $L__BB11_40: and.b32 %r11, %r1, -128; ld.u32 %r250, [%rd22+12]; setp.ne.s32 %p27, %r11, %r250; mov.u16 %rs153, 0; @%p27 bra $L__BB11_43; and.b32 %r251, %r2, -128; ld.u32 %r252, [%rd22+16]; setp.ne.s32 %p28, %r251, %r252; @%p28 bra $L__BB11_43; and.b32 %r253, %r10, -128; ld.u32 %r254, [%rd22+20]; setp.eq.s32 %p29, %r253, %r254; selp.u16 %rs153, 1, 0, %p29; $L__BB11_43: setp.eq.s16 %p30, %rs153, 0; @%p30 bra $L__BB11_47; shl.b32 %r255, %r1, 5; and.b32 %r256, %r255, 3840; shl.b32 %r257, %r2, 1; and.b32 %r258, %r257, 240; or.b32 %r259, %r258, %r256; bfe.u32 %r260, %r10, 3, 4; or.b32 %r261, %r260, %r259; shr.u32 %r262, %r259, 6; ld.u64 %rd23, [%rd22+56]; mul.wide.u32 %rd231, %r262, 8; add.s64 %rd232, %rd23, %rd231; and.b32 %r263, %r261, 63; mov.u64 %rd233, 1; shl.b64 %rd234, %rd233, %r263; ld.u64 %rd235, [%rd232+544]; and.b64 %rd236, %rd235, %rd234; setp.eq.s64 %p31, %rd236, 0; mul.wide.u32 %rd237, %r261, 8; add.s64 %rd238, %rd23, %rd237; add.s64 %rd24, %rd238, 1088; @%p31 bra $L__BB11_46; ld.u64 %rd239, [%rd24]; add.s64 %rd240, %rd23, %rd239; and.b32 %r264, %r2, -8; mov.u32 %r265, 1; st.v2.u32 [%rd22], {%r4, %r264}; and.b32 %r266, %r10, -8; st.u32 [%rd22+8], %r266; st.u64 [%rd22+48], %rd240; shl.b32 %r267, %r1, 6; and.b32 %r268, %r267, 448; shl.b32 %r269, %r2, 3; and.b32 %r270, %r269, 56; or.b32 %r271, %r270, %r268; and.b32 %r272, %r10, 7; or.b32 %r273, %r271, %r272; ld.u8 %rs63, [%rd240+15]; shr.u16 %rs64, %rs63, 5; mov.u32 %r274, 5; cvt.u32.u16 %r275, %rs64; sub.s32 %r276, %r274, %r275; shr.u32 %r277, %r273, %r276; mul.wide.u32 %rd241, %r277, 4; add.s64 %rd242, %rd240, %rd241; mov.u32 %r278, 32; shr.u32 %r279, %r278, %r275; add.s32 %r280, %r279, -1; and.b32 %r281, %r280, %r273; shl.b32 %r282, %r281, %r275; ld.u32 %r283, [%rd242+96]; shr.u32 %r284, %r283, %r282; shl.b32 %r285, %r265, %r275; mov.u32 %r286, -1; shl.b32 %r287, %r286, %r285; not.b32 %r288, %r287; and.b32 %r289, %r284, %r288; cvt.rn.f32.u32 %f134, %r289; ld.v2.f32 {%f135, %f136}, [%rd240+80]; fma.rn.ftz.f32 %f281, %f136, %f134, %f135; bra.uni $L__BB11_70; $L__BB11_75: and.b32 %r17, %r1, -128; ld.u32 %r422, [%rd42+12]; setp.ne.s32 %p49, %r17, %r422; mov.u16 %rs156, 0; @%p49 bra $L__BB11_78; and.b32 %r423, %r16, -128; ld.u32 %r424, [%rd42+16]; setp.ne.s32 %p50, %r423, %r424; @%p50 bra $L__BB11_78; and.b32 %r425, %r10, -128; ld.u32 %r426, [%rd42+20]; setp.eq.s32 %p51, %r425, %r426; selp.u16 %rs156, 1, 0, %p51; $L__BB11_78: setp.eq.s16 %p52, %rs156, 0; @%p52 bra $L__BB11_82; shl.b32 %r427, %r1, 5; and.b32 %r428, %r427, 3840; shl.b32 %r429, %r16, 1; and.b32 %r430, %r429, 240; or.b32 %r431, %r430, %r428; bfe.u32 %r432, %r10, 3, 4; or.b32 %r433, %r431, %r432; shr.u32 %r434, %r431, 6; ld.u64 %rd43, [%rd42+56]; mul.wide.u32 %rd299, %r434, 8; add.s64 %rd300, %rd43, %rd299; and.b32 %r435, %r433, 63; mov.u64 %rd301, 1; shl.b64 %rd302, %rd301, %r435; ld.u64 %rd303, [%rd300+544]; and.b64 %rd304, %rd303, %rd302; setp.eq.s64 %p53, %rd304, 0; mul.wide.u32 %rd305, %r433, 8; add.s64 %rd306, %rd43, %rd305; add.s64 %rd44, %rd306, 1088; @%p53 bra $L__BB11_81; ld.u64 %rd307, [%rd44]; add.s64 %rd308, %rd43, %rd307; and.b32 %r436, %r16, -8; mov.u32 %r437, 1; st.v2.u32 [%rd42], {%r4, %r436}; and.b32 %r438, %r10, -8; st.u32 [%rd42+8], %r438; st.u64 [%rd42+48], %rd308; shl.b32 %r439, %r1, 6; and.b32 %r440, %r439, 448; shl.b32 %r441, %r16, 3; and.b32 %r442, %r441, 56; or.b32 %r443, %r442, %r440; and.b32 %r444, %r10, 7; or.b32 %r445, %r443, %r444; ld.u8 %rs77, [%rd308+15]; shr.u16 %rs78, %rs77, 5; mov.u32 %r446, 5; cvt.u32.u16 %r447, %rs78; sub.s32 %r448, %r446, %r447; shr.u32 %r449, %r445, %r448; mul.wide.u32 %rd309, %r449, 4; add.s64 %rd310, %rd308, %rd309; mov.u32 %r450, 32; shr.u32 %r451, %r450, %r447; add.s32 %r452, %r451, -1; and.b32 %r453, %r452, %r445; shl.b32 %r454, %r453, %r447; ld.u32 %r455, [%rd310+96]; shr.u32 %r456, %r455, %r454; shl.b32 %r457, %r437, %r447; mov.u32 %r458, -1; shl.b32 %r459, %r458, %r457; not.b32 %r460, %r459; and.b32 %r461, %r456, %r460; cvt.rn.f32.u32 %f155, %r461; ld.v2.f32 {%f156, %f157}, [%rd308+80]; fma.rn.ftz.f32 %f282, %f157, %f155, %f156; bra.uni $L__BB11_105; $L__BB11_110: and.b32 %r22, %r1, -128; ld.u32 %r593, [%rd62+12]; setp.ne.s32 %p71, %r22, %r593; mov.u16 %rs159, 0; @%p71 bra $L__BB11_113; and.b32 %r594, %r16, -128; ld.u32 %r595, [%rd62+16]; setp.ne.s32 %p72, %r594, %r595; @%p72 bra $L__BB11_113; and.b32 %r596, %r3, -128; ld.u32 %r597, [%rd62+20]; setp.eq.s32 %p73, %r596, %r597; selp.u16 %rs159, 1, 0, %p73; $L__BB11_113: setp.eq.s16 %p74, %rs159, 0; @%p74 bra $L__BB11_117; shl.b32 %r598, %r1, 5; and.b32 %r599, %r598, 3840; shl.b32 %r600, %r16, 1; and.b32 %r601, %r600, 240; or.b32 %r602, %r601, %r599; bfe.u32 %r603, %r3, 3, 4; or.b32 %r604, %r602, %r603; shr.u32 %r605, %r602, 6; ld.u64 %rd63, [%rd62+56]; mul.wide.u32 %rd366, %r605, 8; add.s64 %rd367, %rd63, %rd366; and.b32 %r606, %r604, 63; mov.u64 %rd368, 1; shl.b64 %rd369, %rd368, %r606; ld.u64 %rd370, [%rd367+544]; and.b64 %rd371, %rd370, %rd369; setp.eq.s64 %p75, %rd371, 0; mul.wide.u32 %rd372, %r604, 8; add.s64 %rd373, %rd63, %rd372; add.s64 %rd64, %rd373, 1088; @%p75 bra $L__BB11_116; ld.u64 %rd374, [%rd64]; add.s64 %rd375, %rd63, %rd374; and.b32 %r607, %r16, -8; mov.u32 %r608, 1; st.v2.u32 [%rd62], {%r4, %r607}; and.b32 %r609, %r3, -8; st.u32 [%rd62+8], %r609; st.u64 [%rd62+48], %rd375; and.b32 %r610, %r1, 7; shl.b32 %r611, %r16, 3; and.b32 %r612, %r611, 56; and.b32 %r613, %r3, 7; bfi.b32 %r614, %r610, %r613, 6, 3; or.b32 %r615, %r614, %r612; ld.u8 %rs91, [%rd375+15]; shr.u16 %rs92, %rs91, 5; mov.u32 %r616, 5; cvt.u32.u16 %r617, %rs92; sub.s32 %r618, %r616, %r617; shr.u32 %r619, %r615, %r618; mul.wide.u32 %rd376, %r619, 4; add.s64 %rd377, %rd375, %rd376; mov.u32 %r620, 32; shr.u32 %r621, %r620, %r617; add.s32 %r622, %r621, -1; and.b32 %r623, %r622, %r615; shl.b32 %r624, %r623, %r617; ld.u32 %r625, [%rd377+96]; shr.u32 %r626, %r625, %r624; shl.b32 %r627, %r608, %r617; mov.u32 %r628, -1; shl.b32 %r629, %r628, %r627; not.b32 %r630, %r629; and.b32 %r631, %r626, %r630; cvt.rn.f32.u32 %f175, %r631; ld.v2.f32 {%f176, %f177}, [%rd375+80]; fma.rn.ftz.f32 %f283, %f177, %f175, %f176; bra.uni $L__BB11_140; $L__BB11_145: and.b32 %r29, %r27, -128; ld.u32 %r761, [%rd81+12]; setp.ne.s32 %p93, %r29, %r761; mov.u16 %rs162, 0; @%p93 bra $L__BB11_148; and.b32 %r762, %r16, -128; ld.u32 %r763, [%rd81+16]; setp.ne.s32 %p94, %r762, %r763; @%p94 bra $L__BB11_148; and.b32 %r764, %r3, -128; ld.u32 %r765, [%rd81+20]; setp.eq.s32 %p95, %r764, %r765; selp.u16 %rs162, 1, 0, %p95; $L__BB11_148: setp.eq.s16 %p96, %rs162, 0; @%p96 bra $L__BB11_152; shl.b32 %r766, %r27, 5; and.b32 %r767, %r766, 3840; shl.b32 %r768, %r16, 1; and.b32 %r769, %r768, 240; or.b32 %r770, %r769, %r767; bfe.u32 %r771, %r3, 3, 4; or.b32 %r772, %r770, %r771; shr.u32 %r773, %r770, 6; ld.u64 %rd82, [%rd81+56]; mul.wide.u32 %rd433, %r773, 8; add.s64 %rd434, %rd82, %rd433; and.b32 %r774, %r772, 63; mov.u64 %rd435, 1; shl.b64 %rd436, %rd435, %r774; ld.u64 %rd437, [%rd434+544]; and.b64 %rd438, %rd437, %rd436; setp.eq.s64 %p97, %rd438, 0; mul.wide.u32 %rd439, %r772, 8; add.s64 %rd440, %rd82, %rd439; add.s64 %rd83, %rd440, 1088; @%p97 bra $L__BB11_151; ld.u64 %rd441, [%rd83]; add.s64 %rd442, %rd82, %rd441; and.b32 %r775, %r16, -8; mov.u32 %r776, 1; st.v2.u32 [%rd81], {%r28, %r775}; and.b32 %r777, %r3, -8; st.u32 [%rd81+8], %r777; st.u64 [%rd81+48], %rd442; and.b32 %r778, %r27, 7; shl.b32 %r779, %r16, 3; and.b32 %r780, %r779, 56; and.b32 %r781, %r3, 7; bfi.b32 %r782, %r778, %r781, 6, 3; or.b32 %r783, %r782, %r780; ld.u8 %rs105, [%rd442+15]; shr.u16 %rs106, %rs105, 5; mov.u32 %r784, 5; cvt.u32.u16 %r785, %rs106; sub.s32 %r786, %r784, %r785; shr.u32 %r787, %r783, %r786; mul.wide.u32 %rd443, %r787, 4; add.s64 %rd444, %rd442, %rd443; mov.u32 %r788, 32; shr.u32 %r789, %r788, %r785; add.s32 %r790, %r789, -1; and.b32 %r791, %r790, %r783; shl.b32 %r792, %r791, %r785; ld.u32 %r793, [%rd444+96]; shr.u32 %r794, %r793, %r792; shl.b32 %r795, %r776, %r785; mov.u32 %r796, -1; shl.b32 %r797, %r796, %r795; not.b32 %r798, %r797; and.b32 %r799, %r794, %r798; cvt.rn.f32.u32 %f198, %r799; ld.v2.f32 {%f199, %f200}, [%rd442+80]; fma.rn.ftz.f32 %f284, %f200, %f198, %f199; bra.uni $L__BB11_175; $L__BB11_180: and.b32 %r34, %r27, -128; ld.u32 %r930, [%rd101+12]; setp.ne.s32 %p115, %r34, %r930; mov.u16 %rs165, 0; @%p115 bra $L__BB11_183; and.b32 %r931, %r16, -128; ld.u32 %r932, [%rd101+16]; setp.ne.s32 %p116, %r931, %r932; @%p116 bra $L__BB11_183; and.b32 %r933, %r10, -128; ld.u32 %r934, [%rd101+20]; setp.eq.s32 %p117, %r933, %r934; selp.u16 %rs165, 1, 0, %p117; $L__BB11_183: setp.eq.s16 %p118, %rs165, 0; @%p118 bra $L__BB11_187; shl.b32 %r935, %r27, 5; and.b32 %r936, %r935, 3840; shl.b32 %r937, %r16, 1; and.b32 %r938, %r937, 240; or.b32 %r939, %r938, %r936; bfe.u32 %r940, %r10, 3, 4; or.b32 %r941, %r939, %r940; shr.u32 %r942, %r939, 6; ld.u64 %rd102, [%rd101+56]; mul.wide.u32 %rd500, %r942, 8; add.s64 %rd501, %rd102, %rd500; and.b32 %r943, %r941, 63; mov.u64 %rd502, 1; shl.b64 %rd503, %rd502, %r943; ld.u64 %rd504, [%rd501+544]; and.b64 %rd505, %rd504, %rd503; setp.eq.s64 %p119, %rd505, 0; mul.wide.u32 %rd506, %r941, 8; add.s64 %rd507, %rd102, %rd506; add.s64 %rd103, %rd507, 1088; @%p119 bra $L__BB11_186; ld.u64 %rd508, [%rd103]; add.s64 %rd509, %rd102, %rd508; and.b32 %r944, %r16, -8; mov.u32 %r945, 1; st.v2.u32 [%rd101], {%r28, %r944}; and.b32 %r946, %r10, -8; st.u32 [%rd101+8], %r946; st.u64 [%rd101+48], %rd509; shl.b32 %r947, %r27, 6; and.b32 %r948, %r947, 448; shl.b32 %r949, %r16, 3; and.b32 %r950, %r949, 56; or.b32 %r951, %r950, %r948; and.b32 %r952, %r10, 7; or.b32 %r953, %r951, %r952; ld.u8 %rs119, [%rd509+15]; shr.u16 %rs120, %rs119, 5; mov.u32 %r954, 5; cvt.u32.u16 %r955, %rs120; sub.s32 %r956, %r954, %r955; shr.u32 %r957, %r953, %r956; mul.wide.u32 %rd510, %r957, 4; add.s64 %rd511, %rd509, %rd510; mov.u32 %r958, 32; shr.u32 %r959, %r958, %r955; add.s32 %r960, %r959, -1; and.b32 %r961, %r960, %r953; shl.b32 %r962, %r961, %r955; ld.u32 %r963, [%rd511+96]; shr.u32 %r964, %r963, %r962; shl.b32 %r965, %r945, %r955; mov.u32 %r966, -1; shl.b32 %r967, %r966, %r965; not.b32 %r968, %r967; and.b32 %r969, %r964, %r968; cvt.rn.f32.u32 %f218, %r969; ld.v2.f32 {%f219, %f220}, [%rd509+80]; fma.rn.ftz.f32 %f285, %f220, %f218, %f219; bra.uni $L__BB11_210; $L__BB11_215: and.b32 %r39, %r27, -128; ld.u32 %r1102, [%rd120+12]; setp.ne.s32 %p137, %r39, %r1102; mov.u16 %rs168, 0; @%p137 bra $L__BB11_218; and.b32 %r1103, %r2, -128; ld.u32 %r1104, [%rd120+16]; setp.ne.s32 %p138, %r1103, %r1104; @%p138 bra $L__BB11_218; and.b32 %r1105, %r10, -128; ld.u32 %r1106, [%rd120+20]; setp.eq.s32 %p139, %r1105, %r1106; selp.u16 %rs168, 1, 0, %p139; $L__BB11_218: setp.eq.s16 %p140, %rs168, 0; @%p140 bra $L__BB11_222; shl.b32 %r1107, %r27, 5; and.b32 %r1108, %r1107, 3840; shl.b32 %r1109, %r2, 1; and.b32 %r1110, %r1109, 240; or.b32 %r1111, %r1108, %r1110; bfe.u32 %r1112, %r10, 3, 4; or.b32 %r1113, %r1112, %r1111; shr.u32 %r1114, %r1111, 6; ld.u64 %rd121, [%rd120+56]; mul.wide.u32 %rd567, %r1114, 8; add.s64 %rd568, %rd121, %rd567; and.b32 %r1115, %r1113, 63; mov.u64 %rd569, 1; shl.b64 %rd570, %rd569, %r1115; ld.u64 %rd571, [%rd568+544]; and.b64 %rd572, %rd571, %rd570; setp.eq.s64 %p141, %rd572, 0; mul.wide.u32 %rd573, %r1113, 8; add.s64 %rd574, %rd121, %rd573; add.s64 %rd122, %rd574, 1088; @%p141 bra $L__BB11_221; ld.u64 %rd575, [%rd122]; add.s64 %rd576, %rd121, %rd575; and.b32 %r1116, %r2, -8; mov.u32 %r1117, 1; st.v2.u32 [%rd120], {%r28, %r1116}; and.b32 %r1118, %r10, -8; st.u32 [%rd120+8], %r1118; st.u64 [%rd120+48], %rd576; shl.b32 %r1119, %r27, 6; and.b32 %r1120, %r1119, 448; shl.b32 %r1121, %r2, 3; and.b32 %r1122, %r1121, 56; or.b32 %r1123, %r1120, %r1122; and.b32 %r1124, %r10, 7; or.b32 %r1125, %r1123, %r1124; ld.u8 %rs133, [%rd576+15]; shr.u16 %rs134, %rs133, 5; mov.u32 %r1126, 5; cvt.u32.u16 %r1127, %rs134; sub.s32 %r1128, %r1126, %r1127; shr.u32 %r1129, %r1125, %r1128; mul.wide.u32 %rd577, %r1129, 4; add.s64 %rd578, %rd576, %rd577; mov.u32 %r1130, 32; shr.u32 %r1131, %r1130, %r1127; add.s32 %r1132, %r1131, -1; and.b32 %r1133, %r1132, %r1125; shl.b32 %r1134, %r1133, %r1127; ld.u32 %r1135, [%rd578+96]; shr.u32 %r1136, %r1135, %r1134; shl.b32 %r1137, %r1117, %r1127; mov.u32 %r1138, -1; shl.b32 %r1139, %r1138, %r1137; not.b32 %r1140, %r1139; and.b32 %r1141, %r1136, %r1140; cvt.rn.f32.u32 %f239, %r1141; ld.v2.f32 {%f240, %f241}, [%rd576+80]; fma.rn.ftz.f32 %f286, %f241, %f239, %f240; bra.uni $L__BB11_245; $L__BB11_258: shl.b64 %rd642, %rd145, 5; add.s64 %rd707, %rd143, %rd642; $L__BB11_259: setp.eq.s64 %p167, %rd707, 0; @%p167 bra $L__BB11_267; add.s64 %rd148, %rd707, 8; ld.u64 %rd149, [%rd707+8]; setp.eq.s64 %p168, %rd149, 0; @%p168 bra $L__BB11_266; add.s64 %rd150, %rd141, %rd149; and.b32 %r1264, %r2, -4096; mov.u64 %rd643, 1; st.v2.u32 [%rd140+24], {%r45, %r1264}; and.b32 %r1265, %r3, -4096; st.u32 [%rd140+32], %r1265; st.u64 [%rd140+64], %rd150; shl.b32 %r1266, %r27, 3; and.b32 %r1267, %r1266, 31744; and.b32 %r1268, %r2, 3968; shr.u32 %r1269, %r1268, 2; or.b32 %r1270, %r1267, %r1269; bfe.u32 %r1271, %r3, 7, 5; or.b32 %r1272, %r1270, %r1271; shr.u32 %r1273, %r1270, 6; mul.wide.u32 %rd644, %r1273, 8; add.s64 %rd645, %rd150, %rd644; and.b32 %r1274, %r1272, 63; shl.b64 %rd646, %rd643, %r1274; ld.u64 %rd647, [%rd645+4128]; and.b64 %rd648, %rd647, %rd646; setp.eq.s64 %p169, %rd648, 0; mul.wide.u32 %rd649, %r1272, 8; add.s64 %rd650, %rd150, %rd649; add.s64 %rd151, %rd650, 8256; @%p169 bra $L__BB11_265; ld.u64 %rd651, [%rd151]; add.s64 %rd152, %rd150, %rd651; st.u32 [%rd140+12], %r44; and.b32 %r1275, %r2, -128; st.u32 [%rd140+16], %r1275; and.b32 %r1276, %r3, -128; st.u32 [%rd140+20], %r1276; st.u64 [%rd140+56], %rd152; shl.b32 %r1277, %r27, 5; and.b32 %r1278, %r1277, 3840; shl.b32 %r1279, %r2, 1; and.b32 %r1280, %r1279, 240; or.b32 %r1281, %r1278, %r1280; bfe.u32 %r1282, %r3, 3, 4; or.b32 %r1283, %r1281, %r1282; shr.u32 %r1284, %r1281, 6; mul.wide.u32 %rd652, %r1284, 8; add.s64 %rd653, %rd152, %rd652; and.b32 %r1285, %r1283, 63; shl.b64 %rd655, %rd643, %r1285; ld.u64 %rd656, [%rd653+544]; and.b64 %rd657, %rd656, %rd655; setp.eq.s64 %p170, %rd657, 0; mul.wide.u32 %rd658, %r1283, 8; add.s64 %rd659, %rd152, %rd658; add.s64 %rd153, %rd659, 1088; @%p170 bra $L__BB11_264; ld.u64 %rd660, [%rd153]; add.s64 %rd661, %rd152, %rd660; and.b32 %r1286, %r2, -8; mov.u32 %r1287, 1; st.v2.u32 [%rd140], {%r28, %r1286}; and.b32 %r1288, %r3, -8; st.u32 [%rd140+8], %r1288; st.u64 [%rd140+48], %rd661; shl.b32 %r1289, %r27, 6; and.b32 %r1290, %r1289, 448; shl.b32 %r1291, %r2, 3; and.b32 %r1292, %r1291, 56; or.b32 %r1293, %r1290, %r1292; and.b32 %r1294, %r3, 7; or.b32 %r1295, %r1293, %r1294; ld.u8 %rs141, [%rd661+15]; shr.u16 %rs142, %rs141, 5; mov.u32 %r1296, 5; cvt.u32.u16 %r1297, %rs142; sub.s32 %r1298, %r1296, %r1297; shr.u32 %r1299, %r1295, %r1298; mul.wide.u32 %rd662, %r1299, 4; add.s64 %rd663, %rd661, %rd662; mov.u32 %r1300, 32; shr.u32 %r1301, %r1300, %r1297; add.s32 %r1302, %r1301, -1; and.b32 %r1303, %r1302, %r1295; shl.b32 %r1304, %r1303, %r1297; ld.u32 %r1305, [%rd663+96]; shr.u32 %r1306, %r1305, %r1304; shl.b32 %r1307, %r1287, %r1297; mov.u32 %r1308, -1; shl.b32 %r1309, %r1308, %r1307; not.b32 %r1310, %r1309; and.b32 %r1311, %r1306, %r1310; cvt.rn.f32.u32 %f254, %r1311; ld.v2.f32 {%f255, %f256}, [%rd661+80]; fma.rn.ftz.f32 %f287, %f256, %f254, %f255; bra.uni $L__BB11_277; $L__BB11_267: ld.f32 %f287, [%rd142+4]; bra.uni $L__BB11_277; $L__BB11_266: ld.f32 %f287, [%rd148+12]; bra.uni $L__BB11_277; $L__BB11_12: and.b32 %r6, %r1, -4096; ld.u32 %r118, [%rd2+24]; setp.ne.s32 %p10, %r6, %r118; mov.u16 %rs151, 0; @%p10 bra $L__BB11_15; and.b32 %r119, %r2, -4096; ld.u32 %r120, [%rd2+28]; setp.ne.s32 %p11, %r119, %r120; @%p11 bra $L__BB11_15; and.b32 %r121, %r3, -4096; ld.u32 %r122, [%rd2+32]; setp.eq.s32 %p12, %r121, %r122; selp.u16 %rs151, 1, 0, %p12; $L__BB11_15: setp.eq.s16 %p13, %rs151, 0; @%p13 bra $L__BB11_21; shl.b32 %r123, %r1, 3; and.b32 %r124, %r123, 31744; and.b32 %r125, %r2, 3968; shr.u32 %r126, %r125, 2; or.b32 %r127, %r126, %r124; bfe.u32 %r128, %r3, 7, 5; or.b32 %r129, %r127, %r128; shr.u32 %r130, %r127, 6; ld.u64 %rd5, [%rd2+64]; mul.wide.u32 %rd176, %r130, 8; add.s64 %rd177, %rd5, %rd176; and.b32 %r131, %r129, 63; mov.u64 %rd178, 1; shl.b64 %rd179, %rd178, %r131; ld.u64 %rd180, [%rd177+4128]; and.b64 %rd181, %rd180, %rd179; setp.eq.s64 %p14, %rd181, 0; mul.wide.u32 %rd182, %r129, 8; add.s64 %rd183, %rd5, %rd182; add.s64 %rd6, %rd183, 8256; @%p14 bra $L__BB11_20; ld.u64 %rd184, [%rd6]; add.s64 %rd7, %rd5, %rd184; st.u32 [%rd2+12], %r5; and.b32 %r132, %r2, -128; st.u32 [%rd2+16], %r132; and.b32 %r133, %r3, -128; st.u32 [%rd2+20], %r133; st.u64 [%rd2+56], %rd7; shl.b32 %r134, %r1, 5; and.b32 %r135, %r134, 3840; shl.b32 %r136, %r2, 1; and.b32 %r137, %r136, 240; or.b32 %r138, %r137, %r135; bfe.u32 %r139, %r3, 3, 4; or.b32 %r140, %r138, %r139; shr.u32 %r141, %r138, 6; mul.wide.u32 %rd185, %r141, 8; add.s64 %rd186, %rd7, %rd185; and.b32 %r142, %r140, 63; shl.b64 %rd188, %rd178, %r142; ld.u64 %rd189, [%rd186+544]; and.b64 %rd190, %rd189, %rd188; setp.eq.s64 %p15, %rd190, 0; mul.wide.u32 %rd191, %r140, 8; add.s64 %rd192, %rd7, %rd191; add.s64 %rd8, %rd192, 1088; @%p15 bra $L__BB11_19; ld.u64 %rd193, [%rd8]; add.s64 %rd194, %rd7, %rd193; and.b32 %r143, %r2, -8; mov.u32 %r144, 1; st.v2.u32 [%rd2], {%r4, %r143}; and.b32 %r145, %r3, -8; st.u32 [%rd2+8], %r145; st.u64 [%rd2+48], %rd194; shl.b32 %r146, %r1, 6; and.b32 %r147, %r146, 448; shl.b32 %r148, %r2, 3; and.b32 %r149, %r148, 56; or.b32 %r150, %r149, %r147; and.b32 %r151, %r3, 7; or.b32 %r152, %r150, %r151; ld.u8 %rs53, [%rd194+15]; shr.u16 %rs54, %rs53, 5; mov.u32 %r153, 5; cvt.u32.u16 %r154, %rs54; sub.s32 %r155, %r153, %r154; shr.u32 %r156, %r152, %r155; mul.wide.u32 %rd195, %r156, 4; add.s64 %rd196, %rd194, %rd195; mov.u32 %r157, 32; shr.u32 %r158, %r157, %r154; add.s32 %r159, %r158, -1; and.b32 %r160, %r159, %r152; shl.b32 %r161, %r160, %r154; ld.u32 %r162, [%rd196+96]; shr.u32 %r163, %r162, %r161; shl.b32 %r164, %r144, %r154; mov.u32 %r165, -1; shl.b32 %r166, %r165, %r164; not.b32 %r167, %r166; and.b32 %r168, %r163, %r167; cvt.rn.f32.u32 %f119, %r168; ld.v2.f32 {%f120, %f121}, [%rd194+80]; fma.rn.ftz.f32 %f280, %f121, %f119, %f120; bra.uni $L__BB11_35; $L__BB11_47: and.b32 %r12, %r1, -4096; ld.u32 %r290, [%rd22+24]; setp.ne.s32 %p32, %r12, %r290; mov.u16 %rs154, 0; @%p32 bra $L__BB11_50; and.b32 %r291, %r2, -4096; ld.u32 %r292, [%rd22+28]; setp.ne.s32 %p33, %r291, %r292; @%p33 bra $L__BB11_50; and.b32 %r293, %r10, -4096; ld.u32 %r294, [%rd22+32]; setp.eq.s32 %p34, %r293, %r294; selp.u16 %rs154, 1, 0, %p34; $L__BB11_50: setp.eq.s16 %p35, %rs154, 0; @%p35 bra $L__BB11_56; shl.b32 %r295, %r1, 3; and.b32 %r296, %r295, 31744; and.b32 %r297, %r2, 3968; shr.u32 %r298, %r297, 2; or.b32 %r299, %r298, %r296; bfe.u32 %r300, %r10, 7, 5; or.b32 %r301, %r300, %r299; shr.u32 %r302, %r299, 6; ld.u64 %rd25, [%rd22+64]; mul.wide.u32 %rd243, %r302, 8; add.s64 %rd244, %rd25, %rd243; and.b32 %r303, %r301, 63; mov.u64 %rd245, 1; shl.b64 %rd246, %rd245, %r303; ld.u64 %rd247, [%rd244+4128]; and.b64 %rd248, %rd247, %rd246; setp.eq.s64 %p36, %rd248, 0; mul.wide.u32 %rd249, %r301, 8; add.s64 %rd250, %rd25, %rd249; add.s64 %rd26, %rd250, 8256; @%p36 bra $L__BB11_55; ld.u64 %rd251, [%rd26]; add.s64 %rd27, %rd25, %rd251; st.u32 [%rd22+12], %r11; and.b32 %r304, %r2, -128; st.u32 [%rd22+16], %r304; and.b32 %r305, %r10, -128; st.u32 [%rd22+20], %r305; st.u64 [%rd22+56], %rd27; shl.b32 %r306, %r1, 5; and.b32 %r307, %r306, 3840; shl.b32 %r308, %r2, 1; and.b32 %r309, %r308, 240; or.b32 %r310, %r309, %r307; bfe.u32 %r311, %r10, 3, 4; or.b32 %r312, %r311, %r310; shr.u32 %r313, %r310, 6; mul.wide.u32 %rd252, %r313, 8; add.s64 %rd253, %rd27, %rd252; and.b32 %r314, %r312, 63; shl.b64 %rd255, %rd245, %r314; ld.u64 %rd256, [%rd253+544]; and.b64 %rd257, %rd256, %rd255; setp.eq.s64 %p37, %rd257, 0; mul.wide.u32 %rd258, %r312, 8; add.s64 %rd259, %rd27, %rd258; add.s64 %rd28, %rd259, 1088; @%p37 bra $L__BB11_54; ld.u64 %rd260, [%rd28]; add.s64 %rd261, %rd27, %rd260; and.b32 %r315, %r2, -8; mov.u32 %r316, 1; st.v2.u32 [%rd22], {%r4, %r315}; and.b32 %r317, %r10, -8; st.u32 [%rd22+8], %r317; st.u64 [%rd22+48], %rd261; shl.b32 %r318, %r1, 6; and.b32 %r319, %r318, 448; shl.b32 %r320, %r2, 3; and.b32 %r321, %r320, 56; or.b32 %r322, %r321, %r319; and.b32 %r323, %r10, 7; or.b32 %r324, %r322, %r323; ld.u8 %rs67, [%rd261+15]; shr.u16 %rs68, %rs67, 5; mov.u32 %r325, 5; cvt.u32.u16 %r326, %rs68; sub.s32 %r327, %r325, %r326; shr.u32 %r328, %r324, %r327; mul.wide.u32 %rd262, %r328, 4; add.s64 %rd263, %rd261, %rd262; mov.u32 %r329, 32; shr.u32 %r330, %r329, %r326; add.s32 %r331, %r330, -1; and.b32 %r332, %r331, %r324; shl.b32 %r333, %r332, %r326; ld.u32 %r334, [%rd263+96]; shr.u32 %r335, %r334, %r333; shl.b32 %r336, %r316, %r326; mov.u32 %r337, -1; shl.b32 %r338, %r337, %r336; not.b32 %r339, %r338; and.b32 %r340, %r335, %r339; cvt.rn.f32.u32 %f139, %r340; ld.v2.f32 {%f140, %f141}, [%rd261+80]; fma.rn.ftz.f32 %f281, %f141, %f139, %f140; bra.uni $L__BB11_70; $L__BB11_82: and.b32 %r18, %r1, -4096; ld.u32 %r462, [%rd42+24]; setp.ne.s32 %p54, %r18, %r462; mov.u16 %rs157, 0; @%p54 bra $L__BB11_85; and.b32 %r463, %r16, -4096; ld.u32 %r464, [%rd42+28]; setp.ne.s32 %p55, %r463, %r464; @%p55 bra $L__BB11_85; and.b32 %r465, %r10, -4096; ld.u32 %r466, [%rd42+32]; setp.eq.s32 %p56, %r465, %r466; selp.u16 %rs157, 1, 0, %p56; $L__BB11_85: setp.eq.s16 %p57, %rs157, 0; @%p57 bra $L__BB11_91; shl.b32 %r467, %r1, 3; and.b32 %r468, %r467, 31744; and.b32 %r469, %r16, 3968; shr.u32 %r470, %r469, 2; or.b32 %r471, %r470, %r468; bfe.u32 %r472, %r10, 7, 5; or.b32 %r473, %r471, %r472; shr.u32 %r474, %r471, 6; ld.u64 %rd45, [%rd42+64]; mul.wide.u32 %rd311, %r474, 8; add.s64 %rd312, %rd45, %rd311; and.b32 %r475, %r473, 63; mov.u64 %rd313, 1; shl.b64 %rd314, %rd313, %r475; ld.u64 %rd315, [%rd312+4128]; and.b64 %rd316, %rd315, %rd314; setp.eq.s64 %p58, %rd316, 0; mul.wide.u32 %rd317, %r473, 8; add.s64 %rd318, %rd45, %rd317; add.s64 %rd46, %rd318, 8256; @%p58 bra $L__BB11_90; ld.u64 %rd319, [%rd46]; add.s64 %rd47, %rd45, %rd319; st.u32 [%rd42+12], %r17; and.b32 %r476, %r16, -128; st.u32 [%rd42+16], %r476; and.b32 %r477, %r10, -128; st.u32 [%rd42+20], %r477; st.u64 [%rd42+56], %rd47; shl.b32 %r478, %r1, 5; and.b32 %r479, %r478, 3840; shl.b32 %r480, %r16, 1; and.b32 %r481, %r480, 240; or.b32 %r482, %r481, %r479; bfe.u32 %r483, %r10, 3, 4; or.b32 %r484, %r482, %r483; shr.u32 %r485, %r482, 6; mul.wide.u32 %rd320, %r485, 8; add.s64 %rd321, %rd47, %rd320; and.b32 %r486, %r484, 63; shl.b64 %rd323, %rd313, %r486; ld.u64 %rd324, [%rd321+544]; and.b64 %rd325, %rd324, %rd323; setp.eq.s64 %p59, %rd325, 0; mul.wide.u32 %rd326, %r484, 8; add.s64 %rd327, %rd47, %rd326; add.s64 %rd48, %rd327, 1088; @%p59 bra $L__BB11_89; ld.u64 %rd328, [%rd48]; add.s64 %rd329, %rd47, %rd328; and.b32 %r487, %r16, -8; mov.u32 %r488, 1; st.v2.u32 [%rd42], {%r4, %r487}; and.b32 %r489, %r10, -8; st.u32 [%rd42+8], %r489; st.u64 [%rd42+48], %rd329; shl.b32 %r490, %r1, 6; and.b32 %r491, %r490, 448; shl.b32 %r492, %r16, 3; and.b32 %r493, %r492, 56; or.b32 %r494, %r493, %r491; and.b32 %r495, %r10, 7; or.b32 %r496, %r494, %r495; ld.u8 %rs81, [%rd329+15]; shr.u16 %rs82, %rs81, 5; mov.u32 %r497, 5; cvt.u32.u16 %r498, %rs82; sub.s32 %r499, %r497, %r498; shr.u32 %r500, %r496, %r499; mul.wide.u32 %rd330, %r500, 4; add.s64 %rd331, %rd329, %rd330; mov.u32 %r501, 32; shr.u32 %r502, %r501, %r498; add.s32 %r503, %r502, -1; and.b32 %r504, %r503, %r496; shl.b32 %r505, %r504, %r498; ld.u32 %r506, [%rd331+96]; shr.u32 %r507, %r506, %r505; shl.b32 %r508, %r488, %r498; mov.u32 %r509, -1; shl.b32 %r510, %r509, %r508; not.b32 %r511, %r510; and.b32 %r512, %r507, %r511; cvt.rn.f32.u32 %f160, %r512; ld.v2.f32 {%f161, %f162}, [%rd329+80]; fma.rn.ftz.f32 %f282, %f162, %f160, %f161; bra.uni $L__BB11_105; $L__BB11_117: and.b32 %r23, %r1, -4096; ld.u32 %r632, [%rd62+24]; setp.ne.s32 %p76, %r23, %r632; mov.u16 %rs160, 0; @%p76 bra $L__BB11_120; and.b32 %r633, %r16, -4096; ld.u32 %r634, [%rd62+28]; setp.ne.s32 %p77, %r633, %r634; @%p77 bra $L__BB11_120; and.b32 %r635, %r3, -4096; ld.u32 %r636, [%rd62+32]; setp.eq.s32 %p78, %r635, %r636; selp.u16 %rs160, 1, 0, %p78; $L__BB11_120: setp.eq.s16 %p79, %rs160, 0; @%p79 bra $L__BB11_126; shl.b32 %r637, %r1, 3; and.b32 %r638, %r637, 31744; and.b32 %r639, %r16, 3968; shr.u32 %r640, %r639, 2; or.b32 %r641, %r640, %r638; bfe.u32 %r642, %r3, 7, 5; or.b32 %r643, %r641, %r642; shr.u32 %r644, %r641, 6; ld.u64 %rd65, [%rd62+64]; mul.wide.u32 %rd378, %r644, 8; add.s64 %rd379, %rd65, %rd378; and.b32 %r645, %r643, 63; mov.u64 %rd380, 1; shl.b64 %rd381, %rd380, %r645; ld.u64 %rd382, [%rd379+4128]; and.b64 %rd383, %rd382, %rd381; setp.eq.s64 %p80, %rd383, 0; mul.wide.u32 %rd384, %r643, 8; add.s64 %rd385, %rd65, %rd384; add.s64 %rd66, %rd385, 8256; @%p80 bra $L__BB11_125; ld.u64 %rd386, [%rd66]; add.s64 %rd67, %rd65, %rd386; st.u32 [%rd62+12], %r22; and.b32 %r646, %r16, -128; st.u32 [%rd62+16], %r646; and.b32 %r647, %r3, -128; st.u32 [%rd62+20], %r647; st.u64 [%rd62+56], %rd67; shl.b32 %r648, %r1, 5; and.b32 %r649, %r648, 3840; shl.b32 %r650, %r16, 1; and.b32 %r651, %r650, 240; or.b32 %r652, %r651, %r649; bfe.u32 %r653, %r3, 3, 4; or.b32 %r654, %r652, %r653; shr.u32 %r655, %r652, 6; mul.wide.u32 %rd387, %r655, 8; add.s64 %rd388, %rd67, %rd387; and.b32 %r656, %r654, 63; shl.b64 %rd390, %rd380, %r656; ld.u64 %rd391, [%rd388+544]; and.b64 %rd392, %rd391, %rd390; setp.eq.s64 %p81, %rd392, 0; mul.wide.u32 %rd393, %r654, 8; add.s64 %rd394, %rd67, %rd393; add.s64 %rd68, %rd394, 1088; @%p81 bra $L__BB11_124; ld.u64 %rd395, [%rd68]; add.s64 %rd396, %rd67, %rd395; and.b32 %r657, %r16, -8; mov.u32 %r658, 1; st.v2.u32 [%rd62], {%r4, %r657}; and.b32 %r659, %r3, -8; st.u32 [%rd62+8], %r659; st.u64 [%rd62+48], %rd396; and.b32 %r660, %r1, 7; shl.b32 %r661, %r16, 3; and.b32 %r662, %r661, 56; and.b32 %r663, %r3, 7; bfi.b32 %r664, %r660, %r663, 6, 3; or.b32 %r665, %r664, %r662; ld.u8 %rs95, [%rd396+15]; shr.u16 %rs96, %rs95, 5; mov.u32 %r666, 5; cvt.u32.u16 %r667, %rs96; sub.s32 %r668, %r666, %r667; shr.u32 %r669, %r665, %r668; mul.wide.u32 %rd397, %r669, 4; add.s64 %rd398, %rd396, %rd397; mov.u32 %r670, 32; shr.u32 %r671, %r670, %r667; add.s32 %r672, %r671, -1; and.b32 %r673, %r672, %r665; shl.b32 %r674, %r673, %r667; ld.u32 %r675, [%rd398+96]; shr.u32 %r676, %r675, %r674; shl.b32 %r677, %r658, %r667; mov.u32 %r678, -1; shl.b32 %r679, %r678, %r677; not.b32 %r680, %r679; and.b32 %r681, %r676, %r680; cvt.rn.f32.u32 %f180, %r681; ld.v2.f32 {%f181, %f182}, [%rd396+80]; fma.rn.ftz.f32 %f283, %f182, %f180, %f181; bra.uni $L__BB11_140; $L__BB11_152: and.b32 %r30, %r27, -4096; ld.u32 %r800, [%rd81+24]; setp.ne.s32 %p98, %r30, %r800; mov.u16 %rs163, 0; @%p98 bra $L__BB11_155; and.b32 %r801, %r16, -4096; ld.u32 %r802, [%rd81+28]; setp.ne.s32 %p99, %r801, %r802; @%p99 bra $L__BB11_155; and.b32 %r803, %r3, -4096; ld.u32 %r804, [%rd81+32]; setp.eq.s32 %p100, %r803, %r804; selp.u16 %rs163, 1, 0, %p100; $L__BB11_155: setp.eq.s16 %p101, %rs163, 0; @%p101 bra $L__BB11_161; shl.b32 %r805, %r27, 3; and.b32 %r806, %r805, 31744; and.b32 %r807, %r16, 3968; shr.u32 %r808, %r807, 2; or.b32 %r809, %r808, %r806; bfe.u32 %r810, %r3, 7, 5; or.b32 %r811, %r809, %r810; shr.u32 %r812, %r809, 6; ld.u64 %rd84, [%rd81+64]; mul.wide.u32 %rd445, %r812, 8; add.s64 %rd446, %rd84, %rd445; and.b32 %r813, %r811, 63; mov.u64 %rd447, 1; shl.b64 %rd448, %rd447, %r813; ld.u64 %rd449, [%rd446+4128]; and.b64 %rd450, %rd449, %rd448; setp.eq.s64 %p102, %rd450, 0; mul.wide.u32 %rd451, %r811, 8; add.s64 %rd452, %rd84, %rd451; add.s64 %rd85, %rd452, 8256; @%p102 bra $L__BB11_160; ld.u64 %rd453, [%rd85]; add.s64 %rd86, %rd84, %rd453; st.u32 [%rd81+12], %r29; and.b32 %r814, %r16, -128; st.u32 [%rd81+16], %r814; and.b32 %r815, %r3, -128; st.u32 [%rd81+20], %r815; st.u64 [%rd81+56], %rd86; shl.b32 %r816, %r27, 5; and.b32 %r817, %r816, 3840; shl.b32 %r818, %r16, 1; and.b32 %r819, %r818, 240; or.b32 %r820, %r819, %r817; bfe.u32 %r821, %r3, 3, 4; or.b32 %r822, %r820, %r821; shr.u32 %r823, %r820, 6; mul.wide.u32 %rd454, %r823, 8; add.s64 %rd455, %rd86, %rd454; and.b32 %r824, %r822, 63; shl.b64 %rd457, %rd447, %r824; ld.u64 %rd458, [%rd455+544]; and.b64 %rd459, %rd458, %rd457; setp.eq.s64 %p103, %rd459, 0; mul.wide.u32 %rd460, %r822, 8; add.s64 %rd461, %rd86, %rd460; add.s64 %rd87, %rd461, 1088; @%p103 bra $L__BB11_159; ld.u64 %rd462, [%rd87]; add.s64 %rd463, %rd86, %rd462; and.b32 %r825, %r16, -8; mov.u32 %r826, 1; st.v2.u32 [%rd81], {%r28, %r825}; and.b32 %r827, %r3, -8; st.u32 [%rd81+8], %r827; st.u64 [%rd81+48], %rd463; and.b32 %r828, %r27, 7; shl.b32 %r829, %r16, 3; and.b32 %r830, %r829, 56; and.b32 %r831, %r3, 7; bfi.b32 %r832, %r828, %r831, 6, 3; or.b32 %r833, %r832, %r830; ld.u8 %rs109, [%rd463+15]; shr.u16 %rs110, %rs109, 5; mov.u32 %r834, 5; cvt.u32.u16 %r835, %rs110; sub.s32 %r836, %r834, %r835; shr.u32 %r837, %r833, %r836; mul.wide.u32 %rd464, %r837, 4; add.s64 %rd465, %rd463, %rd464; mov.u32 %r838, 32; shr.u32 %r839, %r838, %r835; add.s32 %r840, %r839, -1; and.b32 %r841, %r840, %r833; shl.b32 %r842, %r841, %r835; ld.u32 %r843, [%rd465+96]; shr.u32 %r844, %r843, %r842; shl.b32 %r845, %r826, %r835; mov.u32 %r846, -1; shl.b32 %r847, %r846, %r845; not.b32 %r848, %r847; and.b32 %r849, %r844, %r848; cvt.rn.f32.u32 %f203, %r849; ld.v2.f32 {%f204, %f205}, [%rd463+80]; fma.rn.ftz.f32 %f284, %f205, %f203, %f204; bra.uni $L__BB11_175; $L__BB11_187: and.b32 %r35, %r27, -4096; ld.u32 %r970, [%rd101+24]; setp.ne.s32 %p120, %r35, %r970; mov.u16 %rs166, 0; @%p120 bra $L__BB11_190; and.b32 %r971, %r16, -4096; ld.u32 %r972, [%rd101+28]; setp.ne.s32 %p121, %r971, %r972; @%p121 bra $L__BB11_190; and.b32 %r973, %r10, -4096; ld.u32 %r974, [%rd101+32]; setp.eq.s32 %p122, %r973, %r974; selp.u16 %rs166, 1, 0, %p122; $L__BB11_190: setp.eq.s16 %p123, %rs166, 0; @%p123 bra $L__BB11_196; shl.b32 %r975, %r27, 3; and.b32 %r976, %r975, 31744; and.b32 %r977, %r16, 3968; shr.u32 %r978, %r977, 2; or.b32 %r979, %r978, %r976; bfe.u32 %r980, %r10, 7, 5; or.b32 %r981, %r979, %r980; shr.u32 %r982, %r979, 6; ld.u64 %rd104, [%rd101+64]; mul.wide.u32 %rd512, %r982, 8; add.s64 %rd513, %rd104, %rd512; and.b32 %r983, %r981, 63; mov.u64 %rd514, 1; shl.b64 %rd515, %rd514, %r983; ld.u64 %rd516, [%rd513+4128]; and.b64 %rd517, %rd516, %rd515; setp.eq.s64 %p124, %rd517, 0; mul.wide.u32 %rd518, %r981, 8; add.s64 %rd519, %rd104, %rd518; add.s64 %rd105, %rd519, 8256; @%p124 bra $L__BB11_195; ld.u64 %rd520, [%rd105]; add.s64 %rd106, %rd104, %rd520; st.u32 [%rd101+12], %r34; and.b32 %r984, %r16, -128; st.u32 [%rd101+16], %r984; and.b32 %r985, %r10, -128; st.u32 [%rd101+20], %r985; st.u64 [%rd101+56], %rd106; shl.b32 %r986, %r27, 5; and.b32 %r987, %r986, 3840; shl.b32 %r988, %r16, 1; and.b32 %r989, %r988, 240; or.b32 %r990, %r989, %r987; bfe.u32 %r991, %r10, 3, 4; or.b32 %r992, %r990, %r991; shr.u32 %r993, %r990, 6; mul.wide.u32 %rd521, %r993, 8; add.s64 %rd522, %rd106, %rd521; and.b32 %r994, %r992, 63; shl.b64 %rd524, %rd514, %r994; ld.u64 %rd525, [%rd522+544]; and.b64 %rd526, %rd525, %rd524; setp.eq.s64 %p125, %rd526, 0; mul.wide.u32 %rd527, %r992, 8; add.s64 %rd528, %rd106, %rd527; add.s64 %rd107, %rd528, 1088; @%p125 bra $L__BB11_194; ld.u64 %rd529, [%rd107]; add.s64 %rd530, %rd106, %rd529; and.b32 %r995, %r16, -8; mov.u32 %r996, 1; st.v2.u32 [%rd101], {%r28, %r995}; and.b32 %r997, %r10, -8; st.u32 [%rd101+8], %r997; st.u64 [%rd101+48], %rd530; shl.b32 %r998, %r27, 6; and.b32 %r999, %r998, 448; shl.b32 %r1000, %r16, 3; and.b32 %r1001, %r1000, 56; or.b32 %r1002, %r1001, %r999; and.b32 %r1003, %r10, 7; or.b32 %r1004, %r1002, %r1003; ld.u8 %rs123, [%rd530+15]; shr.u16 %rs124, %rs123, 5; mov.u32 %r1005, 5; cvt.u32.u16 %r1006, %rs124; sub.s32 %r1007, %r1005, %r1006; shr.u32 %r1008, %r1004, %r1007; mul.wide.u32 %rd531, %r1008, 4; add.s64 %rd532, %rd530, %rd531; mov.u32 %r1009, 32; shr.u32 %r1010, %r1009, %r1006; add.s32 %r1011, %r1010, -1; and.b32 %r1012, %r1011, %r1004; shl.b32 %r1013, %r1012, %r1006; ld.u32 %r1014, [%rd532+96]; shr.u32 %r1015, %r1014, %r1013; shl.b32 %r1016, %r996, %r1006; mov.u32 %r1017, -1; shl.b32 %r1018, %r1017, %r1016; not.b32 %r1019, %r1018; and.b32 %r1020, %r1015, %r1019; cvt.rn.f32.u32 %f223, %r1020; ld.v2.f32 {%f224, %f225}, [%rd530+80]; fma.rn.ftz.f32 %f285, %f225, %f223, %f224; bra.uni $L__BB11_210; $L__BB11_222: and.b32 %r40, %r27, -4096; ld.u32 %r1142, [%rd120+24]; setp.ne.s32 %p142, %r40, %r1142; mov.u16 %rs169, 0; @%p142 bra $L__BB11_225; and.b32 %r1143, %r2, -4096; ld.u32 %r1144, [%rd120+28]; setp.ne.s32 %p143, %r1143, %r1144; @%p143 bra $L__BB11_225; and.b32 %r1145, %r10, -4096; ld.u32 %r1146, [%rd120+32]; setp.eq.s32 %p144, %r1145, %r1146; selp.u16 %rs169, 1, 0, %p144; $L__BB11_225: setp.eq.s16 %p145, %rs169, 0; @%p145 bra $L__BB11_231; shl.b32 %r1147, %r27, 3; and.b32 %r1148, %r1147, 31744; and.b32 %r1149, %r2, 3968; shr.u32 %r1150, %r1149, 2; or.b32 %r1151, %r1148, %r1150; bfe.u32 %r1152, %r10, 7, 5; or.b32 %r1153, %r1152, %r1151; shr.u32 %r1154, %r1151, 6; ld.u64 %rd123, [%rd120+64]; mul.wide.u32 %rd579, %r1154, 8; add.s64 %rd580, %rd123, %rd579; and.b32 %r1155, %r1153, 63; mov.u64 %rd581, 1; shl.b64 %rd582, %rd581, %r1155; ld.u64 %rd583, [%rd580+4128]; and.b64 %rd584, %rd583, %rd582; setp.eq.s64 %p146, %rd584, 0; mul.wide.u32 %rd585, %r1153, 8; add.s64 %rd586, %rd123, %rd585; add.s64 %rd124, %rd586, 8256; @%p146 bra $L__BB11_230; ld.u64 %rd587, [%rd124]; add.s64 %rd125, %rd123, %rd587; st.u32 [%rd120+12], %r39; and.b32 %r1156, %r2, -128; st.u32 [%rd120+16], %r1156; and.b32 %r1157, %r10, -128; st.u32 [%rd120+20], %r1157; st.u64 [%rd120+56], %rd125; shl.b32 %r1158, %r27, 5; and.b32 %r1159, %r1158, 3840; shl.b32 %r1160, %r2, 1; and.b32 %r1161, %r1160, 240; or.b32 %r1162, %r1159, %r1161; bfe.u32 %r1163, %r10, 3, 4; or.b32 %r1164, %r1163, %r1162; shr.u32 %r1165, %r1162, 6; mul.wide.u32 %rd588, %r1165, 8; add.s64 %rd589, %rd125, %rd588; and.b32 %r1166, %r1164, 63; shl.b64 %rd591, %rd581, %r1166; ld.u64 %rd592, [%rd589+544]; and.b64 %rd593, %rd592, %rd591; setp.eq.s64 %p147, %rd593, 0; mul.wide.u32 %rd594, %r1164, 8; add.s64 %rd595, %rd125, %rd594; add.s64 %rd126, %rd595, 1088; @%p147 bra $L__BB11_229; ld.u64 %rd596, [%rd126]; add.s64 %rd597, %rd125, %rd596; and.b32 %r1167, %r2, -8; mov.u32 %r1168, 1; st.v2.u32 [%rd120], {%r28, %r1167}; and.b32 %r1169, %r10, -8; st.u32 [%rd120+8], %r1169; st.u64 [%rd120+48], %rd597; shl.b32 %r1170, %r27, 6; and.b32 %r1171, %r1170, 448; shl.b32 %r1172, %r2, 3; and.b32 %r1173, %r1172, 56; or.b32 %r1174, %r1171, %r1173; and.b32 %r1175, %r10, 7; or.b32 %r1176, %r1174, %r1175; ld.u8 %rs137, [%rd597+15]; shr.u16 %rs138, %rs137, 5; mov.u32 %r1177, 5; cvt.u32.u16 %r1178, %rs138; sub.s32 %r1179, %r1177, %r1178; shr.u32 %r1180, %r1176, %r1179; mul.wide.u32 %rd598, %r1180, 4; add.s64 %rd599, %rd597, %rd598; mov.u32 %r1181, 32; shr.u32 %r1182, %r1181, %r1178; add.s32 %r1183, %r1182, -1; and.b32 %r1184, %r1183, %r1176; shl.b32 %r1185, %r1184, %r1178; ld.u32 %r1186, [%rd599+96]; shr.u32 %r1187, %r1186, %r1185; shl.b32 %r1188, %r1168, %r1178; mov.u32 %r1189, -1; shl.b32 %r1190, %r1189, %r1188; not.b32 %r1191, %r1190; and.b32 %r1192, %r1187, %r1191; cvt.rn.f32.u32 %f244, %r1192; ld.v2.f32 {%f245, %f246}, [%rd597+80]; fma.rn.ftz.f32 %f286, %f246, %f244, %f245; bra.uni $L__BB11_245; $L__BB11_265: ld.f32 %f287, [%rd151]; bra.uni $L__BB11_277; $L__BB11_11: ld.f32 %f280, [%rd4]; bra.uni $L__BB11_35; $L__BB11_46: ld.f32 %f281, [%rd24]; bra.uni $L__BB11_70; $L__BB11_81: ld.f32 %f282, [%rd44]; bra.uni $L__BB11_105; $L__BB11_116: ld.f32 %f283, [%rd64]; bra.uni $L__BB11_140; $L__BB11_151: ld.f32 %f284, [%rd83]; bra.uni $L__BB11_175; $L__BB11_186: ld.f32 %f285, [%rd103]; bra.uni $L__BB11_210; $L__BB11_221: ld.f32 %f286, [%rd122]; bra.uni $L__BB11_245; $L__BB11_264: ld.f32 %f287, [%rd153]; bra.uni $L__BB11_277; $L__BB11_21: ld.u64 %rd9, [%rd2+40]; mov.u64 %rd700, 0; add.s64 %rd10, %rd9, 24; ld.u32 %r7, [%rd9+24]; setp.eq.s32 %p16, %r7, 0; @%p16 bra $L__BB11_26; add.s64 %rd11, %rd9, 64; shr.u32 %r170, %r1, 12; cvt.u64.u32 %rd198, %r170; shr.u32 %r171, %r3, 12; cvt.u64.u32 %rd199, %r171; bfi.b64 %rd200, %rd198, %rd199, 42, 22; shr.u32 %r172, %r2, 12; mul.wide.u32 %rd201, %r172, 2097152; or.b64 %rd12, %rd200, %rd201; mov.u32 %r1417, 0; $L__BB11_23: cvt.u64.u32 %rd13, %r1417; mul.wide.u32 %rd202, %r1417, 32; add.s64 %rd203, %rd11, %rd202; ld.u64 %rd204, [%rd203]; setp.eq.s64 %p17, %rd204, %rd12; add.s32 %r1417, %r1417, 1; @%p17 bra $L__BB11_25; setp.lt.u32 %p18, %r1417, %r7; @%p18 bra $L__BB11_23; bra.uni $L__BB11_26; $L__BB11_56: ld.u64 %rd29, [%rd22+40]; mov.u64 %rd701, 0; add.s64 %rd30, %rd29, 24; ld.u32 %r13, [%rd29+24]; setp.eq.s32 %p38, %r13, 0; @%p38 bra $L__BB11_61; add.s64 %rd31, %rd29, 64; shr.u32 %r342, %r1, 12; cvt.u64.u32 %rd265, %r342; shl.b64 %rd266, %rd265, 42; shr.u32 %r343, %r2, 12; mul.wide.u32 %rd267, %r343, 2097152; or.b64 %rd268, %rd267, %rd266; shr.u32 %r344, %r10, 12; cvt.u64.u32 %rd269, %r344; or.b64 %rd32, %rd268, %rd269; mov.u32 %r1418, 0; $L__BB11_58: cvt.u64.u32 %rd33, %r1418; mul.wide.u32 %rd270, %r1418, 32; add.s64 %rd271, %rd31, %rd270; ld.u64 %rd272, [%rd271]; setp.eq.s64 %p39, %rd272, %rd32; add.s32 %r1418, %r1418, 1; @%p39 bra $L__BB11_60; setp.lt.u32 %p40, %r1418, %r13; @%p40 bra $L__BB11_58; bra.uni $L__BB11_61; $L__BB11_91: ld.u64 %rd49, [%rd42+40]; mov.u64 %rd702, 0; add.s64 %rd50, %rd49, 24; ld.u32 %r19, [%rd49+24]; setp.eq.s32 %p60, %r19, 0; @%p60 bra $L__BB11_96; add.s64 %rd51, %rd49, 64; shr.u32 %r514, %r1, 12; cvt.u64.u32 %rd333, %r514; shr.u32 %r515, %r10, 12; cvt.u64.u32 %rd334, %r515; bfi.b64 %rd335, %rd333, %rd334, 42, 22; shr.u32 %r516, %r16, 12; mul.wide.u32 %rd336, %r516, 2097152; or.b64 %rd52, %rd335, %rd336; mov.u32 %r1419, 0; $L__BB11_93: cvt.u64.u32 %rd53, %r1419; mul.wide.u32 %rd337, %r1419, 32; add.s64 %rd338, %rd51, %rd337; ld.u64 %rd339, [%rd338]; setp.eq.s64 %p61, %rd339, %rd52; add.s32 %r1419, %r1419, 1; @%p61 bra $L__BB11_95; setp.lt.u32 %p62, %r1419, %r19; @%p62 bra $L__BB11_93; bra.uni $L__BB11_96; $L__BB11_126: ld.u64 %rd69, [%rd62+40]; mov.u64 %rd703, 0; ld.u32 %r24, [%rd69+24]; setp.eq.s32 %p82, %r24, 0; @%p82 bra $L__BB11_131; add.s64 %rd70, %rd69, 64; shr.u32 %r683, %r1, 12; cvt.u64.u32 %rd400, %r683; shr.u32 %r684, %r3, 12; cvt.u64.u32 %rd401, %r684; bfi.b64 %rd402, %rd400, %rd401, 42, 22; shr.u32 %r685, %r16, 12; mul.wide.u32 %rd403, %r685, 2097152; or.b64 %rd71, %rd402, %rd403; mov.u32 %r1420, 0; $L__BB11_128: cvt.u64.u32 %rd72, %r1420; mul.wide.u32 %rd404, %r1420, 32; add.s64 %rd405, %rd70, %rd404; ld.u64 %rd406, [%rd405]; setp.eq.s64 %p83, %rd406, %rd71; add.s32 %r1420, %r1420, 1; @%p83 bra $L__BB11_130; setp.lt.u32 %p84, %r1420, %r24; @%p84 bra $L__BB11_128; bra.uni $L__BB11_131; $L__BB11_161: ld.u64 %rd88, [%rd81+40]; mov.u64 %rd704, 0; add.s64 %rd89, %rd88, 24; ld.u32 %r31, [%rd88+24]; setp.eq.s32 %p104, %r31, 0; @%p104 bra $L__BB11_166; add.s64 %rd90, %rd88, 64; shr.u32 %r851, %r27, 12; cvt.u64.u32 %rd467, %r851; shr.u32 %r852, %r3, 12; cvt.u64.u32 %rd468, %r852; bfi.b64 %rd469, %rd467, %rd468, 42, 22; shr.u32 %r853, %r16, 12; mul.wide.u32 %rd470, %r853, 2097152; or.b64 %rd91, %rd469, %rd470; mov.u32 %r1421, 0; $L__BB11_163: cvt.u64.u32 %rd92, %r1421; mul.wide.u32 %rd471, %r1421, 32; add.s64 %rd472, %rd90, %rd471; ld.u64 %rd473, [%rd472]; setp.eq.s64 %p105, %rd473, %rd91; add.s32 %r1421, %r1421, 1; @%p105 bra $L__BB11_165; setp.lt.u32 %p106, %r1421, %r31; @%p106 bra $L__BB11_163; bra.uni $L__BB11_166; $L__BB11_196: ld.u64 %rd108, [%rd101+40]; mov.u64 %rd705, 0; ld.u32 %r36, [%rd108+24]; setp.eq.s32 %p126, %r36, 0; @%p126 bra $L__BB11_201; add.s64 %rd109, %rd108, 64; shr.u32 %r1022, %r27, 12; cvt.u64.u32 %rd534, %r1022; shr.u32 %r1023, %r10, 12; cvt.u64.u32 %rd535, %r1023; bfi.b64 %rd536, %rd534, %rd535, 42, 22; shr.u32 %r1024, %r16, 12; mul.wide.u32 %rd537, %r1024, 2097152; or.b64 %rd110, %rd536, %rd537; mov.u32 %r1422, 0; $L__BB11_198: cvt.u64.u32 %rd111, %r1422; mul.wide.u32 %rd538, %r1422, 32; add.s64 %rd539, %rd109, %rd538; ld.u64 %rd540, [%rd539]; setp.eq.s64 %p127, %rd540, %rd110; add.s32 %r1422, %r1422, 1; @%p127 bra $L__BB11_200; setp.lt.u32 %p128, %r1422, %r36; @%p128 bra $L__BB11_198; bra.uni $L__BB11_201; $L__BB11_231: ld.u64 %rd127, [%rd120+40]; mov.u64 %rd706, 0; add.s64 %rd128, %rd127, 24; ld.u32 %r41, [%rd127+24]; setp.eq.s32 %p148, %r41, 0; @%p148 bra $L__BB11_236; add.s64 %rd129, %rd127, 64; shr.u32 %r1194, %r27, 12; cvt.u64.u32 %rd601, %r1194; shl.b64 %rd602, %rd601, 42; shr.u32 %r1195, %r2, 12; mul.wide.u32 %rd603, %r1195, 2097152; or.b64 %rd604, %rd602, %rd603; shr.u32 %r1196, %r10, 12; cvt.u64.u32 %rd605, %r1196; or.b64 %rd130, %rd604, %rd605; mov.u32 %r1423, 0; $L__BB11_233: cvt.u64.u32 %rd131, %r1423; mul.wide.u32 %rd606, %r1423, 32; add.s64 %rd607, %rd129, %rd606; ld.u64 %rd608, [%rd607]; setp.eq.s64 %p149, %rd608, %rd130; add.s32 %r1423, %r1423, 1; @%p149 bra $L__BB11_235; setp.lt.u32 %p150, %r1423, %r41; @%p150 bra $L__BB11_233; bra.uni $L__BB11_236; $L__BB11_275: ld.f32 %f287, [%rd159]; bra.uni $L__BB11_277; $L__BB11_272: ld.f32 %f287, [%rd155]; bra.uni $L__BB11_277; $L__BB11_20: ld.f32 %f280, [%rd6]; bra.uni $L__BB11_35; $L__BB11_55: ld.f32 %f281, [%rd26]; bra.uni $L__BB11_70; $L__BB11_90: ld.f32 %f282, [%rd46]; bra.uni $L__BB11_105; $L__BB11_125: ld.f32 %f283, [%rd66]; bra.uni $L__BB11_140; $L__BB11_160: ld.f32 %f284, [%rd85]; bra.uni $L__BB11_175; $L__BB11_195: ld.f32 %f285, [%rd105]; bra.uni $L__BB11_210; $L__BB11_230: ld.f32 %f286, [%rd124]; bra.uni $L__BB11_245; $L__BB11_271: ld.f32 %f287, [%rd157]; $L__BB11_277: sub.ftz.f32 %f274, %f286, %f287; fma.rn.ftz.f32 %f275, %f3, %f274, %f287; sub.ftz.f32 %f276, %f78, %f275; fma.rn.ftz.f32 %f277, %f2, %f276, %f275; sub.ftz.f32 %f278, %f277, %f53; fma.rn.ftz.f32 %f279, %f1, %f278, %f53; st.param.f32 [func_retval0+0], %f279; ret; $L__BB11_19: ld.f32 %f280, [%rd8]; bra.uni $L__BB11_35; $L__BB11_54: ld.f32 %f281, [%rd28]; bra.uni $L__BB11_70; $L__BB11_89: ld.f32 %f282, [%rd48]; bra.uni $L__BB11_105; $L__BB11_124: ld.f32 %f283, [%rd68]; bra.uni $L__BB11_140; $L__BB11_159: ld.f32 %f284, [%rd87]; bra.uni $L__BB11_175; $L__BB11_194: ld.f32 %f285, [%rd107]; bra.uni $L__BB11_210; $L__BB11_229: ld.f32 %f286, [%rd126]; bra.uni $L__BB11_245; $L__BB11_25: shl.b64 %rd206, %rd13, 5; add.s64 %rd700, %rd11, %rd206; $L__BB11_26: setp.eq.s64 %p19, %rd700, 0; @%p19 bra $L__BB11_34; add.s64 %rd16, %rd700, 8; ld.u64 %rd17, [%rd700+8]; setp.eq.s64 %p20, %rd17, 0; @%p20 bra $L__BB11_33; add.s64 %rd18, %rd9, %rd17; and.b32 %r173, %r2, -4096; mov.u64 %rd207, 1; st.v2.u32 [%rd2+24], {%r6, %r173}; and.b32 %r174, %r3, -4096; st.u32 [%rd2+32], %r174; st.u64 [%rd2+64], %rd18; shl.b32 %r175, %r1, 3; and.b32 %r176, %r175, 31744; and.b32 %r177, %r2, 3968; shr.u32 %r178, %r177, 2; or.b32 %r179, %r178, %r176; bfe.u32 %r180, %r3, 7, 5; or.b32 %r181, %r179, %r180; shr.u32 %r182, %r179, 6; mul.wide.u32 %rd208, %r182, 8; add.s64 %rd209, %rd18, %rd208; and.b32 %r183, %r181, 63; shl.b64 %rd210, %rd207, %r183; ld.u64 %rd211, [%rd209+4128]; and.b64 %rd212, %rd211, %rd210; setp.eq.s64 %p21, %rd212, 0; mul.wide.u32 %rd213, %r181, 8; add.s64 %rd214, %rd18, %rd213; add.s64 %rd19, %rd214, 8256; @%p21 bra $L__BB11_32; ld.u64 %rd215, [%rd19]; add.s64 %rd20, %rd18, %rd215; st.u32 [%rd2+12], %r5; and.b32 %r184, %r2, -128; st.u32 [%rd2+16], %r184; and.b32 %r185, %r3, -128; st.u32 [%rd2+20], %r185; st.u64 [%rd2+56], %rd20; shl.b32 %r186, %r1, 5; and.b32 %r187, %r186, 3840; shl.b32 %r188, %r2, 1; and.b32 %r189, %r188, 240; or.b32 %r190, %r189, %r187; bfe.u32 %r191, %r3, 3, 4; or.b32 %r192, %r190, %r191; shr.u32 %r193, %r190, 6; mul.wide.u32 %rd216, %r193, 8; add.s64 %rd217, %rd20, %rd216; and.b32 %r194, %r192, 63; shl.b64 %rd219, %rd207, %r194; ld.u64 %rd220, [%rd217+544]; and.b64 %rd221, %rd220, %rd219; setp.eq.s64 %p22, %rd221, 0; mul.wide.u32 %rd222, %r192, 8; add.s64 %rd223, %rd20, %rd222; add.s64 %rd21, %rd223, 1088; @%p22 bra $L__BB11_31; ld.u64 %rd224, [%rd21]; add.s64 %rd225, %rd20, %rd224; and.b32 %r195, %r2, -8; mov.u32 %r196, 1; st.v2.u32 [%rd2], {%r4, %r195}; and.b32 %r197, %r3, -8; st.u32 [%rd2+8], %r197; st.u64 [%rd2+48], %rd225; shl.b32 %r198, %r1, 6; and.b32 %r199, %r198, 448; shl.b32 %r200, %r2, 3; and.b32 %r201, %r200, 56; or.b32 %r202, %r201, %r199; and.b32 %r203, %r3, 7; or.b32 %r204, %r202, %r203; ld.u8 %rs55, [%rd225+15]; shr.u16 %rs56, %rs55, 5; mov.u32 %r205, 5; cvt.u32.u16 %r206, %rs56; sub.s32 %r207, %r205, %r206; shr.u32 %r208, %r204, %r207; mul.wide.u32 %rd226, %r208, 4; add.s64 %rd227, %rd225, %rd226; mov.u32 %r209, 32; shr.u32 %r210, %r209, %r206; add.s32 %r211, %r210, -1; and.b32 %r212, %r211, %r204; shl.b32 %r213, %r212, %r206; ld.u32 %r214, [%rd227+96]; shr.u32 %r215, %r214, %r213; shl.b32 %r216, %r196, %r206; mov.u32 %r217, -1; shl.b32 %r218, %r217, %r216; not.b32 %r219, %r218; and.b32 %r220, %r215, %r219; cvt.rn.f32.u32 %f124, %r220; ld.v2.f32 {%f125, %f126}, [%rd225+80]; fma.rn.ftz.f32 %f280, %f126, %f124, %f125; bra.uni $L__BB11_35; $L__BB11_34: ld.f32 %f280, [%rd10+4]; bra.uni $L__BB11_35; $L__BB11_60: shl.b64 %rd274, %rd33, 5; add.s64 %rd701, %rd31, %rd274; $L__BB11_61: setp.eq.s64 %p41, %rd701, 0; @%p41 bra $L__BB11_69; add.s64 %rd36, %rd701, 8; ld.u64 %rd37, [%rd701+8]; setp.eq.s64 %p42, %rd37, 0; @%p42 bra $L__BB11_68; add.s64 %rd38, %rd29, %rd37; and.b32 %r345, %r2, -4096; mov.u64 %rd275, 1; st.v2.u32 [%rd22+24], {%r12, %r345}; and.b32 %r346, %r10, -4096; st.u32 [%rd22+32], %r346; st.u64 [%rd22+64], %rd38; shl.b32 %r347, %r1, 3; and.b32 %r348, %r347, 31744; and.b32 %r349, %r2, 3968; shr.u32 %r350, %r349, 2; or.b32 %r351, %r350, %r348; bfe.u32 %r352, %r10, 7, 5; or.b32 %r353, %r352, %r351; shr.u32 %r354, %r351, 6; mul.wide.u32 %rd276, %r354, 8; add.s64 %rd277, %rd38, %rd276; and.b32 %r355, %r353, 63; shl.b64 %rd278, %rd275, %r355; ld.u64 %rd279, [%rd277+4128]; and.b64 %rd280, %rd279, %rd278; setp.eq.s64 %p43, %rd280, 0; mul.wide.u32 %rd281, %r353, 8; add.s64 %rd282, %rd38, %rd281; add.s64 %rd39, %rd282, 8256; @%p43 bra $L__BB11_67; ld.u64 %rd283, [%rd39]; add.s64 %rd40, %rd38, %rd283; st.u32 [%rd22+12], %r11; and.b32 %r356, %r2, -128; st.u32 [%rd22+16], %r356; and.b32 %r357, %r10, -128; st.u32 [%rd22+20], %r357; st.u64 [%rd22+56], %rd40; shl.b32 %r358, %r1, 5; and.b32 %r359, %r358, 3840; shl.b32 %r360, %r2, 1; and.b32 %r361, %r360, 240; or.b32 %r362, %r361, %r359; bfe.u32 %r363, %r10, 3, 4; or.b32 %r364, %r363, %r362; shr.u32 %r365, %r362, 6; mul.wide.u32 %rd284, %r365, 8; add.s64 %rd285, %rd40, %rd284; and.b32 %r366, %r364, 63; shl.b64 %rd287, %rd275, %r366; ld.u64 %rd288, [%rd285+544]; and.b64 %rd289, %rd288, %rd287; setp.eq.s64 %p44, %rd289, 0; mul.wide.u32 %rd290, %r364, 8; add.s64 %rd291, %rd40, %rd290; add.s64 %rd41, %rd291, 1088; @%p44 bra $L__BB11_66; ld.u64 %rd292, [%rd41]; add.s64 %rd293, %rd40, %rd292; and.b32 %r367, %r2, -8; mov.u32 %r368, 1; st.v2.u32 [%rd22], {%r4, %r367}; and.b32 %r369, %r10, -8; st.u32 [%rd22+8], %r369; st.u64 [%rd22+48], %rd293; shl.b32 %r370, %r1, 6; and.b32 %r371, %r370, 448; shl.b32 %r372, %r2, 3; and.b32 %r373, %r372, 56; or.b32 %r374, %r373, %r371; and.b32 %r375, %r10, 7; or.b32 %r376, %r374, %r375; ld.u8 %rs69, [%rd293+15]; shr.u16 %rs70, %rs69, 5; mov.u32 %r377, 5; cvt.u32.u16 %r378, %rs70; sub.s32 %r379, %r377, %r378; shr.u32 %r380, %r376, %r379; mul.wide.u32 %rd294, %r380, 4; add.s64 %rd295, %rd293, %rd294; mov.u32 %r381, 32; shr.u32 %r382, %r381, %r378; add.s32 %r383, %r382, -1; and.b32 %r384, %r383, %r376; shl.b32 %r385, %r384, %r378; ld.u32 %r386, [%rd295+96]; shr.u32 %r387, %r386, %r385; shl.b32 %r388, %r368, %r378; mov.u32 %r389, -1; shl.b32 %r390, %r389, %r388; not.b32 %r391, %r390; and.b32 %r392, %r387, %r391; cvt.rn.f32.u32 %f144, %r392; ld.v2.f32 {%f145, %f146}, [%rd293+80]; fma.rn.ftz.f32 %f281, %f146, %f144, %f145; bra.uni $L__BB11_70; $L__BB11_69: ld.f32 %f281, [%rd30+4]; bra.uni $L__BB11_70; $L__BB11_95: shl.b64 %rd341, %rd53, 5; add.s64 %rd702, %rd51, %rd341; $L__BB11_96: setp.eq.s64 %p63, %rd702, 0; @%p63 bra $L__BB11_104; add.s64 %rd56, %rd702, 8; ld.u64 %rd57, [%rd702+8]; setp.eq.s64 %p64, %rd57, 0; @%p64 bra $L__BB11_103; add.s64 %rd58, %rd49, %rd57; and.b32 %r517, %r16, -4096; mov.u64 %rd342, 1; st.v2.u32 [%rd42+24], {%r18, %r517}; and.b32 %r518, %r10, -4096; st.u32 [%rd42+32], %r518; st.u64 [%rd42+64], %rd58; shl.b32 %r519, %r1, 3; and.b32 %r520, %r519, 31744; and.b32 %r521, %r16, 3968; shr.u32 %r522, %r521, 2; or.b32 %r523, %r522, %r520; bfe.u32 %r524, %r10, 7, 5; or.b32 %r525, %r523, %r524; shr.u32 %r526, %r523, 6; mul.wide.u32 %rd343, %r526, 8; add.s64 %rd344, %rd58, %rd343; and.b32 %r527, %r525, 63; shl.b64 %rd345, %rd342, %r527; ld.u64 %rd346, [%rd344+4128]; and.b64 %rd347, %rd346, %rd345; setp.eq.s64 %p65, %rd347, 0; mul.wide.u32 %rd348, %r525, 8; add.s64 %rd349, %rd58, %rd348; add.s64 %rd59, %rd349, 8256; @%p65 bra $L__BB11_102; ld.u64 %rd350, [%rd59]; add.s64 %rd60, %rd58, %rd350; st.u32 [%rd42+12], %r17; and.b32 %r528, %r16, -128; st.u32 [%rd42+16], %r528; and.b32 %r529, %r10, -128; st.u32 [%rd42+20], %r529; st.u64 [%rd42+56], %rd60; shl.b32 %r530, %r1, 5; and.b32 %r531, %r530, 3840; shl.b32 %r532, %r16, 1; and.b32 %r533, %r532, 240; or.b32 %r534, %r533, %r531; bfe.u32 %r535, %r10, 3, 4; or.b32 %r536, %r534, %r535; shr.u32 %r537, %r534, 6; mul.wide.u32 %rd351, %r537, 8; add.s64 %rd352, %rd60, %rd351; and.b32 %r538, %r536, 63; shl.b64 %rd354, %rd342, %r538; ld.u64 %rd355, [%rd352+544]; and.b64 %rd356, %rd355, %rd354; setp.eq.s64 %p66, %rd356, 0; mul.wide.u32 %rd357, %r536, 8; add.s64 %rd358, %rd60, %rd357; add.s64 %rd61, %rd358, 1088; @%p66 bra $L__BB11_101; ld.u64 %rd359, [%rd61]; add.s64 %rd360, %rd60, %rd359; and.b32 %r539, %r16, -8; mov.u32 %r540, 1; st.v2.u32 [%rd42], {%r4, %r539}; and.b32 %r541, %r10, -8; st.u32 [%rd42+8], %r541; st.u64 [%rd42+48], %rd360; shl.b32 %r542, %r1, 6; and.b32 %r543, %r542, 448; shl.b32 %r544, %r16, 3; and.b32 %r545, %r544, 56; or.b32 %r546, %r545, %r543; and.b32 %r547, %r10, 7; or.b32 %r548, %r546, %r547; ld.u8 %rs83, [%rd360+15]; shr.u16 %rs84, %rs83, 5; mov.u32 %r549, 5; cvt.u32.u16 %r550, %rs84; sub.s32 %r551, %r549, %r550; shr.u32 %r552, %r548, %r551; mul.wide.u32 %rd361, %r552, 4; add.s64 %rd362, %rd360, %rd361; mov.u32 %r553, 32; shr.u32 %r554, %r553, %r550; add.s32 %r555, %r554, -1; and.b32 %r556, %r555, %r548; shl.b32 %r557, %r556, %r550; ld.u32 %r558, [%rd362+96]; shr.u32 %r559, %r558, %r557; shl.b32 %r560, %r540, %r550; mov.u32 %r561, -1; shl.b32 %r562, %r561, %r560; not.b32 %r563, %r562; and.b32 %r564, %r559, %r563; cvt.rn.f32.u32 %f165, %r564; ld.v2.f32 {%f166, %f167}, [%rd360+80]; fma.rn.ftz.f32 %f282, %f167, %f165, %f166; bra.uni $L__BB11_105; $L__BB11_104: ld.f32 %f282, [%rd50+4]; bra.uni $L__BB11_105; $L__BB11_130: shl.b64 %rd408, %rd72, 5; add.s64 %rd703, %rd70, %rd408; $L__BB11_131: setp.eq.s64 %p85, %rd703, 0; @%p85 bra $L__BB11_139; add.s64 %rd75, %rd703, 8; ld.u64 %rd76, [%rd703+8]; setp.eq.s64 %p86, %rd76, 0; @%p86 bra $L__BB11_138; add.s64 %rd77, %rd69, %rd76; and.b32 %r686, %r16, -4096; mov.u64 %rd409, 1; st.v2.u32 [%rd62+24], {%r23, %r686}; and.b32 %r687, %r3, -4096; st.u32 [%rd62+32], %r687; st.u64 [%rd62+64], %rd77; shl.b32 %r688, %r1, 3; and.b32 %r689, %r688, 31744; and.b32 %r690, %r16, 3968; shr.u32 %r691, %r690, 2; or.b32 %r692, %r691, %r689; bfe.u32 %r693, %r3, 7, 5; or.b32 %r694, %r692, %r693; shr.u32 %r695, %r692, 6; mul.wide.u32 %rd410, %r695, 8; add.s64 %rd411, %rd77, %rd410; and.b32 %r696, %r694, 63; shl.b64 %rd412, %rd409, %r696; ld.u64 %rd413, [%rd411+4128]; and.b64 %rd414, %rd413, %rd412; setp.eq.s64 %p87, %rd414, 0; mul.wide.u32 %rd415, %r694, 8; add.s64 %rd416, %rd77, %rd415; add.s64 %rd78, %rd416, 8256; @%p87 bra $L__BB11_137; ld.u64 %rd417, [%rd78]; add.s64 %rd79, %rd77, %rd417; st.u32 [%rd62+12], %r22; and.b32 %r697, %r16, -128; st.u32 [%rd62+16], %r697; and.b32 %r698, %r3, -128; st.u32 [%rd62+20], %r698; st.u64 [%rd62+56], %rd79; shl.b32 %r699, %r1, 5; and.b32 %r700, %r699, 3840; shl.b32 %r701, %r16, 1; and.b32 %r702, %r701, 240; or.b32 %r703, %r702, %r700; bfe.u32 %r704, %r3, 3, 4; or.b32 %r705, %r703, %r704; shr.u32 %r706, %r703, 6; mul.wide.u32 %rd418, %r706, 8; add.s64 %rd419, %rd79, %rd418; and.b32 %r707, %r705, 63; shl.b64 %rd421, %rd409, %r707; ld.u64 %rd422, [%rd419+544]; and.b64 %rd423, %rd422, %rd421; setp.eq.s64 %p88, %rd423, 0; mul.wide.u32 %rd424, %r705, 8; add.s64 %rd425, %rd79, %rd424; add.s64 %rd80, %rd425, 1088; @%p88 bra $L__BB11_136; ld.u64 %rd426, [%rd80]; add.s64 %rd427, %rd79, %rd426; and.b32 %r708, %r16, -8; mov.u32 %r709, 1; st.v2.u32 [%rd62], {%r4, %r708}; and.b32 %r710, %r3, -8; st.u32 [%rd62+8], %r710; st.u64 [%rd62+48], %rd427; and.b32 %r711, %r1, 7; shl.b32 %r712, %r16, 3; and.b32 %r713, %r712, 56; and.b32 %r714, %r3, 7; bfi.b32 %r715, %r711, %r714, 6, 3; or.b32 %r716, %r715, %r713; ld.u8 %rs97, [%rd427+15]; shr.u16 %rs98, %rs97, 5; mov.u32 %r717, 5; cvt.u32.u16 %r718, %rs98; sub.s32 %r719, %r717, %r718; shr.u32 %r720, %r716, %r719; mul.wide.u32 %rd428, %r720, 4; add.s64 %rd429, %rd427, %rd428; mov.u32 %r721, 32; shr.u32 %r722, %r721, %r718; add.s32 %r723, %r722, -1; and.b32 %r724, %r723, %r716; shl.b32 %r725, %r724, %r718; ld.u32 %r726, [%rd429+96]; shr.u32 %r727, %r726, %r725; shl.b32 %r728, %r709, %r718; mov.u32 %r729, -1; shl.b32 %r730, %r729, %r728; not.b32 %r731, %r730; and.b32 %r732, %r727, %r731; cvt.rn.f32.u32 %f185, %r732; ld.v2.f32 {%f186, %f187}, [%rd427+80]; fma.rn.ftz.f32 %f283, %f187, %f185, %f186; bra.uni $L__BB11_140; $L__BB11_139: ld.f32 %f283, [%rd69+28]; bra.uni $L__BB11_140; $L__BB11_165: shl.b64 %rd475, %rd92, 5; add.s64 %rd704, %rd90, %rd475; $L__BB11_166: setp.eq.s64 %p107, %rd704, 0; @%p107 bra $L__BB11_174; add.s64 %rd95, %rd704, 8; ld.u64 %rd96, [%rd704+8]; setp.eq.s64 %p108, %rd96, 0; @%p108 bra $L__BB11_173; add.s64 %rd97, %rd88, %rd96; and.b32 %r854, %r16, -4096; mov.u64 %rd476, 1; st.v2.u32 [%rd81+24], {%r30, %r854}; and.b32 %r855, %r3, -4096; st.u32 [%rd81+32], %r855; st.u64 [%rd81+64], %rd97; shl.b32 %r856, %r27, 3; and.b32 %r857, %r856, 31744; and.b32 %r858, %r16, 3968; shr.u32 %r859, %r858, 2; or.b32 %r860, %r859, %r857; bfe.u32 %r861, %r3, 7, 5; or.b32 %r862, %r860, %r861; shr.u32 %r863, %r860, 6; mul.wide.u32 %rd477, %r863, 8; add.s64 %rd478, %rd97, %rd477; and.b32 %r864, %r862, 63; shl.b64 %rd479, %rd476, %r864; ld.u64 %rd480, [%rd478+4128]; and.b64 %rd481, %rd480, %rd479; setp.eq.s64 %p109, %rd481, 0; mul.wide.u32 %rd482, %r862, 8; add.s64 %rd483, %rd97, %rd482; add.s64 %rd98, %rd483, 8256; @%p109 bra $L__BB11_172; ld.u64 %rd484, [%rd98]; add.s64 %rd99, %rd97, %rd484; st.u32 [%rd81+12], %r29; and.b32 %r865, %r16, -128; st.u32 [%rd81+16], %r865; and.b32 %r866, %r3, -128; st.u32 [%rd81+20], %r866; st.u64 [%rd81+56], %rd99; shl.b32 %r867, %r27, 5; and.b32 %r868, %r867, 3840; shl.b32 %r869, %r16, 1; and.b32 %r870, %r869, 240; or.b32 %r871, %r870, %r868; bfe.u32 %r872, %r3, 3, 4; or.b32 %r873, %r871, %r872; shr.u32 %r874, %r871, 6; mul.wide.u32 %rd485, %r874, 8; add.s64 %rd486, %rd99, %rd485; and.b32 %r875, %r873, 63; shl.b64 %rd488, %rd476, %r875; ld.u64 %rd489, [%rd486+544]; and.b64 %rd490, %rd489, %rd488; setp.eq.s64 %p110, %rd490, 0; mul.wide.u32 %rd491, %r873, 8; add.s64 %rd492, %rd99, %rd491; add.s64 %rd100, %rd492, 1088; @%p110 bra $L__BB11_171; ld.u64 %rd493, [%rd100]; add.s64 %rd494, %rd99, %rd493; and.b32 %r876, %r16, -8; mov.u32 %r877, 1; st.v2.u32 [%rd81], {%r28, %r876}; and.b32 %r878, %r3, -8; st.u32 [%rd81+8], %r878; st.u64 [%rd81+48], %rd494; and.b32 %r879, %r27, 7; shl.b32 %r880, %r16, 3; and.b32 %r881, %r880, 56; and.b32 %r882, %r3, 7; bfi.b32 %r883, %r879, %r882, 6, 3; or.b32 %r884, %r883, %r881; ld.u8 %rs111, [%rd494+15]; shr.u16 %rs112, %rs111, 5; mov.u32 %r885, 5; cvt.u32.u16 %r886, %rs112; sub.s32 %r887, %r885, %r886; shr.u32 %r888, %r884, %r887; mul.wide.u32 %rd495, %r888, 4; add.s64 %rd496, %rd494, %rd495; mov.u32 %r889, 32; shr.u32 %r890, %r889, %r886; add.s32 %r891, %r890, -1; and.b32 %r892, %r891, %r884; shl.b32 %r893, %r892, %r886; ld.u32 %r894, [%rd496+96]; shr.u32 %r895, %r894, %r893; shl.b32 %r896, %r877, %r886; mov.u32 %r897, -1; shl.b32 %r898, %r897, %r896; not.b32 %r899, %r898; and.b32 %r900, %r895, %r899; cvt.rn.f32.u32 %f208, %r900; ld.v2.f32 {%f209, %f210}, [%rd494+80]; fma.rn.ftz.f32 %f284, %f210, %f208, %f209; bra.uni $L__BB11_175; $L__BB11_174: ld.f32 %f284, [%rd89+4]; bra.uni $L__BB11_175; $L__BB11_200: shl.b64 %rd542, %rd111, 5; add.s64 %rd705, %rd109, %rd542; $L__BB11_201: setp.eq.s64 %p129, %rd705, 0; @%p129 bra $L__BB11_209; add.s64 %rd114, %rd705, 8; ld.u64 %rd115, [%rd705+8]; setp.eq.s64 %p130, %rd115, 0; @%p130 bra $L__BB11_208; add.s64 %rd116, %rd108, %rd115; and.b32 %r1025, %r16, -4096; mov.u64 %rd543, 1; st.v2.u32 [%rd101+24], {%r35, %r1025}; and.b32 %r1026, %r10, -4096; st.u32 [%rd101+32], %r1026; st.u64 [%rd101+64], %rd116; shl.b32 %r1027, %r27, 3; and.b32 %r1028, %r1027, 31744; and.b32 %r1029, %r16, 3968; shr.u32 %r1030, %r1029, 2; or.b32 %r1031, %r1030, %r1028; bfe.u32 %r1032, %r10, 7, 5; or.b32 %r1033, %r1031, %r1032; shr.u32 %r1034, %r1031, 6; mul.wide.u32 %rd544, %r1034, 8; add.s64 %rd545, %rd116, %rd544; and.b32 %r1035, %r1033, 63; shl.b64 %rd546, %rd543, %r1035; ld.u64 %rd547, [%rd545+4128]; and.b64 %rd548, %rd547, %rd546; setp.eq.s64 %p131, %rd548, 0; mul.wide.u32 %rd549, %r1033, 8; add.s64 %rd550, %rd116, %rd549; add.s64 %rd117, %rd550, 8256; @%p131 bra $L__BB11_207; ld.u64 %rd551, [%rd117]; add.s64 %rd118, %rd116, %rd551; st.u32 [%rd101+12], %r34; and.b32 %r1036, %r16, -128; st.u32 [%rd101+16], %r1036; and.b32 %r1037, %r10, -128; st.u32 [%rd101+20], %r1037; st.u64 [%rd101+56], %rd118; shl.b32 %r1038, %r27, 5; and.b32 %r1039, %r1038, 3840; shl.b32 %r1040, %r16, 1; and.b32 %r1041, %r1040, 240; or.b32 %r1042, %r1041, %r1039; bfe.u32 %r1043, %r10, 3, 4; or.b32 %r1044, %r1042, %r1043; shr.u32 %r1045, %r1042, 6; mul.wide.u32 %rd552, %r1045, 8; add.s64 %rd553, %rd118, %rd552; and.b32 %r1046, %r1044, 63; shl.b64 %rd555, %rd543, %r1046; ld.u64 %rd556, [%rd553+544]; and.b64 %rd557, %rd556, %rd555; setp.eq.s64 %p132, %rd557, 0; mul.wide.u32 %rd558, %r1044, 8; add.s64 %rd559, %rd118, %rd558; add.s64 %rd119, %rd559, 1088; @%p132 bra $L__BB11_206; ld.u64 %rd560, [%rd119]; add.s64 %rd561, %rd118, %rd560; and.b32 %r1047, %r16, -8; mov.u32 %r1048, 1; st.v2.u32 [%rd101], {%r28, %r1047}; and.b32 %r1049, %r10, -8; st.u32 [%rd101+8], %r1049; st.u64 [%rd101+48], %rd561; shl.b32 %r1050, %r27, 6; and.b32 %r1051, %r1050, 448; shl.b32 %r1052, %r16, 3; and.b32 %r1053, %r1052, 56; or.b32 %r1054, %r1053, %r1051; and.b32 %r1055, %r10, 7; or.b32 %r1056, %r1054, %r1055; ld.u8 %rs125, [%rd561+15]; shr.u16 %rs126, %rs125, 5; mov.u32 %r1057, 5; cvt.u32.u16 %r1058, %rs126; sub.s32 %r1059, %r1057, %r1058; shr.u32 %r1060, %r1056, %r1059; mul.wide.u32 %rd562, %r1060, 4; add.s64 %rd563, %rd561, %rd562; mov.u32 %r1061, 32; shr.u32 %r1062, %r1061, %r1058; add.s32 %r1063, %r1062, -1; and.b32 %r1064, %r1063, %r1056; shl.b32 %r1065, %r1064, %r1058; ld.u32 %r1066, [%rd563+96]; shr.u32 %r1067, %r1066, %r1065; shl.b32 %r1068, %r1048, %r1058; mov.u32 %r1069, -1; shl.b32 %r1070, %r1069, %r1068; not.b32 %r1071, %r1070; and.b32 %r1072, %r1067, %r1071; cvt.rn.f32.u32 %f228, %r1072; ld.v2.f32 {%f229, %f230}, [%rd561+80]; fma.rn.ftz.f32 %f285, %f230, %f228, %f229; bra.uni $L__BB11_210; $L__BB11_209: ld.f32 %f285, [%rd108+28]; bra.uni $L__BB11_210; $L__BB11_235: shl.b64 %rd610, %rd131, 5; add.s64 %rd706, %rd129, %rd610; $L__BB11_236: setp.eq.s64 %p151, %rd706, 0; @%p151 bra $L__BB11_244; add.s64 %rd134, %rd706, 8; ld.u64 %rd135, [%rd706+8]; setp.eq.s64 %p152, %rd135, 0; @%p152 bra $L__BB11_243; add.s64 %rd136, %rd127, %rd135; and.b32 %r1197, %r2, -4096; mov.u64 %rd611, 1; st.v2.u32 [%rd120+24], {%r40, %r1197}; and.b32 %r1198, %r10, -4096; st.u32 [%rd120+32], %r1198; st.u64 [%rd120+64], %rd136; shl.b32 %r1199, %r27, 3; and.b32 %r1200, %r1199, 31744; and.b32 %r1201, %r2, 3968; shr.u32 %r1202, %r1201, 2; or.b32 %r1203, %r1200, %r1202; bfe.u32 %r1204, %r10, 7, 5; or.b32 %r1205, %r1204, %r1203; shr.u32 %r1206, %r1203, 6; mul.wide.u32 %rd612, %r1206, 8; add.s64 %rd613, %rd136, %rd612; and.b32 %r1207, %r1205, 63; shl.b64 %rd614, %rd611, %r1207; ld.u64 %rd615, [%rd613+4128]; and.b64 %rd616, %rd615, %rd614; setp.eq.s64 %p153, %rd616, 0; mul.wide.u32 %rd617, %r1205, 8; add.s64 %rd618, %rd136, %rd617; add.s64 %rd137, %rd618, 8256; @%p153 bra $L__BB11_242; ld.u64 %rd619, [%rd137]; add.s64 %rd138, %rd136, %rd619; st.u32 [%rd120+12], %r39; and.b32 %r1208, %r2, -128; st.u32 [%rd120+16], %r1208; and.b32 %r1209, %r10, -128; st.u32 [%rd120+20], %r1209; st.u64 [%rd120+56], %rd138; shl.b32 %r1210, %r27, 5; and.b32 %r1211, %r1210, 3840; shl.b32 %r1212, %r2, 1; and.b32 %r1213, %r1212, 240; or.b32 %r1214, %r1211, %r1213; bfe.u32 %r1215, %r10, 3, 4; or.b32 %r1216, %r1215, %r1214; shr.u32 %r1217, %r1214, 6; mul.wide.u32 %rd620, %r1217, 8; add.s64 %rd621, %rd138, %rd620; and.b32 %r1218, %r1216, 63; shl.b64 %rd623, %rd611, %r1218; ld.u64 %rd624, [%rd621+544]; and.b64 %rd625, %rd624, %rd623; setp.eq.s64 %p154, %rd625, 0; mul.wide.u32 %rd626, %r1216, 8; add.s64 %rd627, %rd138, %rd626; add.s64 %rd139, %rd627, 1088; @%p154 bra $L__BB11_241; ld.u64 %rd628, [%rd139]; add.s64 %rd629, %rd138, %rd628; and.b32 %r1219, %r2, -8; mov.u32 %r1220, 1; st.v2.u32 [%rd120], {%r28, %r1219}; and.b32 %r1221, %r10, -8; st.u32 [%rd120+8], %r1221; st.u64 [%rd120+48], %rd629; shl.b32 %r1222, %r27, 6; and.b32 %r1223, %r1222, 448; shl.b32 %r1224, %r2, 3; and.b32 %r1225, %r1224, 56; or.b32 %r1226, %r1223, %r1225; and.b32 %r1227, %r10, 7; or.b32 %r1228, %r1226, %r1227; ld.u8 %rs139, [%rd629+15]; shr.u16 %rs140, %rs139, 5; mov.u32 %r1229, 5; cvt.u32.u16 %r1230, %rs140; sub.s32 %r1231, %r1229, %r1230; shr.u32 %r1232, %r1228, %r1231; mul.wide.u32 %rd630, %r1232, 4; add.s64 %rd631, %rd629, %rd630; mov.u32 %r1233, 32; shr.u32 %r1234, %r1233, %r1230; add.s32 %r1235, %r1234, -1; and.b32 %r1236, %r1235, %r1228; shl.b32 %r1237, %r1236, %r1230; ld.u32 %r1238, [%rd631+96]; shr.u32 %r1239, %r1238, %r1237; shl.b32 %r1240, %r1220, %r1230; mov.u32 %r1241, -1; shl.b32 %r1242, %r1241, %r1240; not.b32 %r1243, %r1242; and.b32 %r1244, %r1239, %r1243; cvt.rn.f32.u32 %f249, %r1244; ld.v2.f32 {%f250, %f251}, [%rd629+80]; fma.rn.ftz.f32 %f286, %f251, %f249, %f250; bra.uni $L__BB11_245; $L__BB11_244: ld.f32 %f286, [%rd128+4]; bra.uni $L__BB11_245; $L__BB11_33: ld.f32 %f280, [%rd16+12]; bra.uni $L__BB11_35; $L__BB11_68: ld.f32 %f281, [%rd36+12]; bra.uni $L__BB11_70; $L__BB11_103: ld.f32 %f282, [%rd56+12]; bra.uni $L__BB11_105; $L__BB11_138: ld.f32 %f283, [%rd75+12]; bra.uni $L__BB11_140; $L__BB11_173: ld.f32 %f284, [%rd95+12]; bra.uni $L__BB11_175; $L__BB11_208: ld.f32 %f285, [%rd114+12]; bra.uni $L__BB11_210; $L__BB11_243: ld.f32 %f286, [%rd134+12]; bra.uni $L__BB11_245; $L__BB11_32: ld.f32 %f280, [%rd19]; bra.uni $L__BB11_35; $L__BB11_67: ld.f32 %f281, [%rd39]; bra.uni $L__BB11_70; $L__BB11_102: ld.f32 %f282, [%rd59]; bra.uni $L__BB11_105; $L__BB11_137: ld.f32 %f283, [%rd78]; bra.uni $L__BB11_140; $L__BB11_172: ld.f32 %f284, [%rd98]; bra.uni $L__BB11_175; $L__BB11_207: ld.f32 %f285, [%rd117]; bra.uni $L__BB11_210; $L__BB11_242: ld.f32 %f286, [%rd137]; bra.uni $L__BB11_245; $L__BB11_31: ld.f32 %f280, [%rd21]; bra.uni $L__BB11_35; $L__BB11_66: ld.f32 %f281, [%rd41]; bra.uni $L__BB11_70; $L__BB11_101: ld.f32 %f282, [%rd61]; bra.uni $L__BB11_105; $L__BB11_136: ld.f32 %f283, [%rd80]; bra.uni $L__BB11_140; $L__BB11_171: ld.f32 %f284, [%rd100]; bra.uni $L__BB11_175; $L__BB11_206: ld.f32 %f285, [%rd119]; bra.uni $L__BB11_210; $L__BB11_241: ld.f32 %f286, [%rd139]; bra.uni $L__BB11_245; } .func (.param .b32 func_retval0) _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE( .param .b64 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0, .param .align 4 .b8 _ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1[12] ) { .reg .pred %p<174>; .reg .b16 %rs<138>; .reg .f32 %f<288>; .reg .b32 %r<881>; .reg .b64 %rd<708>; ld.param.u64 %rd160, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_0]; cvta.to.local.u64 %rd1, %rd160; ld.param.f32 %f103, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+8]; ld.param.f32 %f104, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1+4]; ld.param.f32 %f105, [_ZNK7nanovdb16SampleFromVoxelsINS_12ReadAccessorINS_4Fp16ELi0ELi1ELi2EEELi1ELb0EEclIfNS_4Vec3EEEfT0_IJT_EE_param_1]; cvt.rmi.ftz.f32.f32 %f106, %f105; cvt.rmi.ftz.f32.f32 %f107, %f104; cvt.rmi.ftz.f32.f32 %f108, %f103; sub.ftz.f32 %f1, %f105, %f106; sub.ftz.f32 %f2, %f104, %f107; sub.ftz.f32 %f3, %f103, %f108; cvt.rzi.ftz.s32.f32 %r1, %f106; cvt.rzi.ftz.s32.f32 %r2, %f107; cvt.rzi.ftz.s32.f32 %r3, %f108; and.b32 %r4, %r1, -8; ld.local.u64 %rd2, [%rd1]; ld.u32 %r49, [%rd2]; setp.ne.s32 %p1, %r4, %r49; mov.u16 %rs117, 0; @%p1 bra $L__BB12_3; and.b32 %r50, %r2, -8; ld.u32 %r51, [%rd2+4]; setp.ne.s32 %p2, %r50, %r51; @%p2 bra $L__BB12_3; and.b32 %r52, %r3, -8; ld.u32 %r53, [%rd2+8]; setp.eq.s32 %p3, %r52, %r53; selp.u16 %rs117, 1, 0, %p3; $L__BB12_3: setp.eq.s16 %p4, %rs117, 0; @%p4 bra $L__BB12_5; shl.b32 %r54, %r1, 6; and.b32 %r55, %r54, 448; shl.b32 %r56, %r2, 3; and.b32 %r57, %r56, 56; or.b32 %r58, %r57, %r55; and.b32 %r59, %r3, 7; or.b32 %r60, %r58, %r59; ld.u64 %rd161, [%rd2+48]; mul.wide.u32 %rd162, %r60, 2; add.s64 %rd163, %rd161, %rd162; ld.u16 %rs45, [%rd163+96]; cvt.rn.f32.u16 %f109, %rs45; ld.v2.f32 {%f110, %f111}, [%rd161+80]; fma.rn.ftz.f32 %f280, %f111, %f109, %f110; $L__BB12_35: add.s32 %r10, %r3, 1; ld.local.u64 %rd22, [%rd1]; ld.u32 %r153, [%rd22]; setp.ne.s32 %p23, %r4, %r153; mov.u16 %rs120, 0; @%p23 bra $L__BB12_38; and.b32 %r154, %r2, -8; ld.u32 %r155, [%rd22+4]; setp.ne.s32 %p24, %r154, %r155; @%p24 bra $L__BB12_38; and.b32 %r156, %r10, -8; ld.u32 %r157, [%rd22+8]; setp.eq.s32 %p25, %r156, %r157; selp.u16 %rs120, 1, 0, %p25; $L__BB12_38: setp.eq.s16 %p26, %rs120, 0; @%p26 bra $L__BB12_40; shl.b32 %r158, %r1, 6; and.b32 %r159, %r158, 448; shl.b32 %r160, %r2, 3; and.b32 %r161, %r160, 56; or.b32 %r162, %r161, %r159; and.b32 %r163, %r10, 7; or.b32 %r164, %r162, %r163; ld.u64 %rd228, [%rd22+48]; mul.wide.u32 %rd229, %r164, 2; add.s64 %rd230, %rd228, %rd229; ld.u16 %rs55, [%rd230+96]; cvt.rn.f32.u16 %f129, %rs55; ld.v2.f32 {%f130, %f131}, [%rd228+80]; fma.rn.ftz.f32 %f281, %f131, %f129, %f130; $L__BB12_70: sub.ftz.f32 %f149, %f281, %f280; fma.rn.ftz.f32 %f28, %f3, %f149, %f280; add.s32 %r16, %r2, 1; ld.local.u64 %rd42, [%rd1]; ld.u32 %r257, [%rd42]; setp.ne.s32 %p45, %r4, %r257; mov.u16 %rs123, 0; @%p45 bra $L__BB12_73; and.b32 %r258, %r16, -8; ld.u32 %r259, [%rd42+4]; setp.ne.s32 %p46, %r258, %r259; @%p46 bra $L__BB12_73; and.b32 %r260, %r10, -8; ld.u32 %r261, [%rd42+8]; setp.eq.s32 %p47, %r260, %r261; selp.u16 %rs123, 1, 0, %p47; $L__BB12_73: setp.eq.s16 %p48, %rs123, 0; @%p48 bra $L__BB12_75; shl.b32 %r262, %r1, 6; and.b32 %r263, %r262, 448; shl.b32 %r264, %r16, 3; and.b32 %r265, %r264, 56; or.b32 %r266, %r265, %r263; and.b32 %r267, %r10, 7; or.b32 %r268, %r266, %r267; ld.u64 %rd296, [%rd42+48]; mul.wide.u32 %rd297, %r268, 2; add.s64 %rd298, %rd296, %rd297; ld.u16 %rs65, [%rd298+96]; cvt.rn.f32.u16 %f150, %rs65; ld.v2.f32 {%f151, %f152}, [%rd296+80]; fma.rn.ftz.f32 %f282, %f152, %f150, %f151; $L__BB12_105: ld.local.u64 %rd62, [%rd1]; ld.u32 %r361, [%rd62]; setp.ne.s32 %p67, %r4, %r361; mov.u16 %rs126, 0; @%p67 bra $L__BB12_108; and.b32 %r362, %r16, -8; ld.u32 %r363, [%rd62+4]; setp.ne.s32 %p68, %r362, %r363; @%p68 bra $L__BB12_108; and.b32 %r364, %r3, -8; ld.u32 %r365, [%rd62+8]; setp.eq.s32 %p69, %r364, %r365; selp.u16 %rs126, 1, 0, %p69; $L__BB12_108: setp.eq.s16 %p70, %rs126, 0; @%p70 bra $L__BB12_110; and.b32 %r366, %r1, 7; shl.b32 %r367, %r16, 3; and.b32 %r368, %r367, 56; and.b32 %r369, %r3, 7; bfi.b32 %r370, %r366, %r369, 6, 3; or.b32 %r371, %r370, %r368; ld.u64 %rd363, [%rd62+48]; mul.wide.u32 %rd364, %r371, 2; add.s64 %rd365, %rd363, %rd364; ld.u16 %rs75, [%rd365+96]; cvt.rn.f32.u16 %f170, %rs75; ld.v2.f32 {%f171, %f172}, [%rd363+80]; fma.rn.ftz.f32 %f283, %f172, %f170, %f171; $L__BB12_140: sub.ftz.f32 %f190, %f282, %f283; fma.rn.ftz.f32 %f191, %f3, %f190, %f283; sub.ftz.f32 %f192, %f191, %f28; fma.rn.ftz.f32 %f53, %f2, %f192, %f28; add.s32 %r27, %r1, 1; and.b32 %r28, %r27, -8; ld.local.u64 %rd81, [%rd1]; ld.u32 %r461, [%rd81]; setp.ne.s32 %p89, %r28, %r461; mov.u16 %rs129, 0; @%p89 bra $L__BB12_143; and.b32 %r462, %r16, -8; ld.u32 %r463, [%rd81+4]; setp.ne.s32 %p90, %r462, %r463; @%p90 bra $L__BB12_143; and.b32 %r464, %r3, -8; ld.u32 %r465, [%rd81+8]; setp.eq.s32 %p91, %r464, %r465; selp.u16 %rs129, 1, 0, %p91; $L__BB12_143: setp.eq.s16 %p92, %rs129, 0; @%p92 bra $L__BB12_145; and.b32 %r466, %r27, 7; shl.b32 %r467, %r16, 3; and.b32 %r468, %r467, 56; and.b32 %r469, %r3, 7; bfi.b32 %r470, %r466, %r469, 6, 3; or.b32 %r471, %r470, %r468; ld.u64 %rd430, [%rd81+48]; mul.wide.u32 %rd431, %r471, 2; add.s64 %rd432, %rd430, %rd431; ld.u16 %rs85, [%rd432+96]; cvt.rn.f32.u16 %f193, %rs85; ld.v2.f32 {%f194, %f195}, [%rd430+80]; fma.rn.ftz.f32 %f284, %f195, %f193, %f194; $L__BB12_175: ld.local.u64 %rd101, [%rd1]; ld.u32 %r561, [%rd101]; setp.ne.s32 %p111, %r28, %r561; mov.u16 %rs132, 0; @%p111 bra $L__BB12_178; and.b32 %r562, %r16, -8; ld.u32 %r563, [%rd101+4]; setp.ne.s32 %p112, %r562, %r563; @%p112 bra $L__BB12_178; and.b32 %r564, %r10, -8; ld.u32 %r565, [%rd101+8]; setp.eq.s32 %p113, %r564, %r565; selp.u16 %rs132, 1, 0, %p113; $L__BB12_178: setp.eq.s16 %p114, %rs132, 0; @%p114 bra $L__BB12_180; shl.b32 %r566, %r27, 6; and.b32 %r567, %r566, 448; shl.b32 %r568, %r16, 3; and.b32 %r569, %r568, 56; or.b32 %r570, %r569, %r567; and.b32 %r571, %r10, 7; or.b32 %r572, %r570, %r571; ld.u64 %rd497, [%rd101+48]; mul.wide.u32 %rd498, %r572, 2; add.s64 %rd499, %rd497, %rd498; ld.u16 %rs95, [%rd499+96]; cvt.rn.f32.u16 %f213, %rs95; ld.v2.f32 {%f214, %f215}, [%rd497+80]; fma.rn.ftz.f32 %f285, %f215, %f213, %f214; $L__BB12_210: sub.ftz.f32 %f233, %f285, %f284; fma.rn.ftz.f32 %f78, %f3, %f233, %f284; ld.local.u64 %rd120, [%rd1]; ld.u32 %r665, [%rd120]; setp.ne.s32 %p133, %r28, %r665; mov.u16 %rs135, 0; @%p133 bra $L__BB12_213; and.b32 %r666, %r2, -8; ld.u32 %r667, [%rd120+4]; setp.ne.s32 %p134, %r666, %r667; @%p134 bra $L__BB12_213; and.b32 %r668, %r10, -8; ld.u32 %r669, [%rd120+8]; setp.eq.s32 %p135, %r668, %r669; selp.u16 %rs135, 1, 0, %p135; $L__BB12_213: setp.eq.s16 %p136, %rs135, 0; @%p136 bra $L__BB12_215; shl.b32 %r670, %r27, 6; and.b32 %r671, %r670, 448; shl.b32 %r672, %r2, 3; and.b32 %r673, %r672, 56; or.b32 %r674, %r671, %r673; and.b32 %r675, %r10, 7; or.b32 %r676, %r674, %r675; ld.u64 %rd564, [%rd120+48]; mul.wide.u32 %rd565, %r676, 2; add.s64 %rd566, %rd564, %rd565; ld.u16 %rs105, [%rd566+96]; cvt.rn.f32.u16 %f234, %rs105; ld.v2.f32 {%f235, %f236}, [%rd564+80]; fma.rn.ftz.f32 %f286, %f236, %f234, %f235; $L__BB12_245: ld.local.u64 %rd140, [%rd1]; ld.u32 %r769, [%rd140]; setp.ne.s32 %p155, %r28, %r769; @%p155 bra $L__BB12_248; and.b32 %r770, %r2, -8; ld.u32 %r771, [%rd140+4]; setp.ne.s32 %p156, %r770, %r771; @%p156 bra $L__BB12_248; and.b32 %r772, %r3, -8; ld.u32 %r773, [%rd140+8]; setp.eq.s32 %p157, %r772, %r773; @%p157 bra $L__BB12_276; bra.uni $L__BB12_248; $L__BB12_276: shl.b32 %r866, %r27, 6; and.b32 %r867, %r866, 448; shl.b32 %r868, %r2, 3; and.b32 %r869, %r868, 56; or.b32 %r870, %r867, %r869; and.b32 %r871, %r3, 7; or.b32 %r872, %r870, %r871; ld.u64 %rd697, [%rd140+48]; mul.wide.u32 %rd698, %r872, 2; add.s64 %rd699, %rd697, %rd698; ld.u16 %rs116, [%rd699+96]; cvt.rn.f32.u16 %f269, %rs116; ld.v2.f32 {%f270, %f271}, [%rd697+80]; fma.rn.ftz.f32 %f287, %f271, %f269, %f270; bra.uni $L__BB12_277; $L__BB12_248: and.b32 %r44, %r27, -128; ld.u32 %r774, [%rd140+12]; setp.ne.s32 %p158, %r44, %r774; @%p158 bra $L__BB12_251; and.b32 %r775, %r2, -128; ld.u32 %r776, [%rd140+16]; setp.ne.s32 %p159, %r775, %r776; @%p159 bra $L__BB12_251; and.b32 %r777, %r3, -128; ld.u32 %r778, [%rd140+20]; setp.eq.s32 %p160, %r777, %r778; @%p160 bra $L__BB12_273; bra.uni $L__BB12_251; $L__BB12_273: shl.b32 %r848, %r27, 5; and.b32 %r849, %r848, 3840; shl.b32 %r850, %r2, 1; and.b32 %r851, %r850, 240; or.b32 %r852, %r849, %r851; bfe.u32 %r853, %r3, 3, 4; or.b32 %r854, %r852, %r853; shr.u32 %r855, %r852, 6; ld.u64 %rd158, [%rd140+56]; mul.wide.u32 %rd685, %r855, 8; add.s64 %rd686, %rd158, %rd685; and.b32 %r856, %r854, 63; mov.u64 %rd687, 1; shl.b64 %rd688, %rd687, %r856; ld.u64 %rd689, [%rd686+544]; and.b64 %rd690, %rd689, %rd688; setp.eq.s64 %p173, %rd690, 0; mul.wide.u32 %rd691, %r854, 8; add.s64 %rd692, %rd158, %rd691; add.s64 %rd159, %rd692, 1088; @%p173 bra $L__BB12_275; ld.u64 %rd693, [%rd159]; add.s64 %rd694, %rd158, %rd693; and.b32 %r857, %r2, -8; st.v2.u32 [%rd140], {%r28, %r857}; and.b32 %r858, %r3, -8; st.u32 [%rd140+8], %r858; st.u64 [%rd140+48], %rd694; shl.b32 %r859, %r27, 6; and.b32 %r860, %r859, 448; shl.b32 %r861, %r2, 3; and.b32 %r862, %r861, 56; or.b32 %r863, %r860, %r862; and.b32 %r864, %r3, 7; or.b32 %r865, %r863, %r864; mul.wide.u32 %rd695, %r865, 2; add.s64 %rd696, %rd694, %rd695; ld.u16 %rs115, [%rd696+96]; cvt.rn.f32.u16 %f264, %rs115; ld.v2.f32 {%f265, %f266}, [%rd694+80]; fma.rn.ftz.f32 %f287, %f266, %f264, %f265; bra.uni $L__BB12_277; $L__BB12_251: and.b32 %r45, %r27, -4096; ld.u32 %r779, [%rd140+24]; setp.ne.s32 %p161, %r45, %r779; @%p161 bra $L__BB12_254; and.b32 %r780, %r2, -4096; ld.u32 %r781, [%rd140+28]; setp.ne.s32 %p162, %r780, %r781; @%p162 bra $L__BB12_254; and.b32 %r782, %r3, -4096; ld.u32 %r783, [%rd140+32]; setp.eq.s32 %p163, %r782, %r783; @%p163 bra $L__BB12_268; bra.uni $L__BB12_254; $L__BB12_268: shl.b32 %r819, %r27, 3; and.b32 %r820, %r819, 31744; and.b32 %r821, %r2, 3968; shr.u32 %r822, %r821, 2; or.b32 %r823, %r820, %r822; bfe.u32 %r824, %r3, 7, 5; or.b32 %r825, %r823, %r824; shr.u32 %r826, %r823, 6; ld.u64 %rd154, [%rd140+64]; mul.wide.u32 %rd664, %r826, 8; add.s64 %rd665, %rd154, %rd664; and.b32 %r827, %r825, 63; mov.u64 %rd666, 1; shl.b64 %rd667, %rd666, %r827; ld.u64 %rd668, [%rd665+4128]; and.b64 %rd669, %rd668, %rd667; setp.eq.s64 %p171, %rd669, 0; mul.wide.u32 %rd670, %r825, 8; add.s64 %rd671, %rd154, %rd670; add.s64 %rd155, %rd671, 8256; @%p171 bra $L__BB12_272; ld.u64 %rd672, [%rd155]; add.s64 %rd156, %rd154, %rd672; st.u32 [%rd140+12], %r44; and.b32 %r828, %r2, -128; st.u32 [%rd140+16], %r828; and.b32 %r829, %r3, -128; st.u32 [%rd140+20], %r829; st.u64 [%rd140+56], %rd156; shl.b32 %r830, %r27, 5; and.b32 %r831, %r830, 3840; shl.b32 %r832, %r2, 1; and.b32 %r833, %r832, 240; or.b32 %r834, %r831, %r833; bfe.u32 %r835, %r3, 3, 4; or.b32 %r836, %r834, %r835; shr.u32 %r837, %r834, 6; mul.wide.u32 %rd673, %r837, 8; add.s64 %rd674, %rd156, %rd673; and.b32 %r838, %r836, 63; shl.b64 %rd676, %rd666, %r838; ld.u64 %rd677, [%rd674+544]; and.b64 %rd678, %rd677, %rd676; setp.eq.s64 %p172, %rd678, 0; mul.wide.u32 %rd679, %r836, 8; add.s64 %rd680, %rd156, %rd679; add.s64 %rd157, %rd680, 1088; @%p172 bra $L__BB12_271; ld.u64 %rd681, [%rd157]; add.s64 %rd682, %rd156, %rd681; and.b32 %r839, %r2, -8; st.v2.u32 [%rd140], {%r28, %r839}; and.b32 %r840, %r3, -8; st.u32 [%rd140+8], %r840; st.u64 [%rd140+48], %rd682; shl.b32 %r841, %r27, 6; and.b32 %r842, %r841, 448; shl.b32 %r843, %r2, 3; and.b32 %r844, %r843, 56; or.b32 %r845, %r842, %r844; and.b32 %r846, %r3, 7; or.b32 %r847, %r845, %r846; mul.wide.u32 %rd683, %r847, 2; add.s64 %rd684, %rd682, %rd683; ld.u16 %rs114, [%rd684+96]; cvt.rn.f32.u16 %f259, %rs114; ld.v2.f32 {%f260, %f261}, [%rd682+80]; fma.rn.ftz.f32 %f287, %f261, %f259, %f260; bra.uni $L__BB12_277; $L__BB12_254: ld.u64 %rd141, [%rd140+40]; mov.u64 %rd707, 0; add.s64 %rd142, %rd141, 24; ld.u32 %r46, [%rd141+24]; setp.eq.s32 %p164, %r46, 0; @%p164 bra $L__BB12_259; add.s64 %rd143, %rd141, 64; shr.u32 %r785, %r27, 12; cvt.u64.u32 %rd633, %r785; shl.b64 %rd634, %rd633, 42; shr.u32 %r786, %r2, 12; mul.wide.u32 %rd635, %r786, 2097152; shr.u32 %r787, %r3, 12; cvt.u64.u32 %rd636, %r787; or.b64 %rd637, %rd635, %rd636; or.b64 %rd144, %rd637, %rd634; mov.u32 %r880, 0; $L__BB12_256: cvt.u64.u32 %rd145, %r880; mul.wide.u32 %rd638, %r880, 32; add.s64 %rd639, %rd143, %rd638; ld.u64 %rd640, [%rd639]; setp.eq.s64 %p165, %rd640, %rd144; add.s32 %r880, %r880, 1; @%p165 bra $L__BB12_258; setp.lt.u32 %p166, %r880, %r46; @%p166 bra $L__BB12_256; bra.uni $L__BB12_259; $L__BB12_5: and.b32 %r5, %r1, -128; ld.u32 %r61, [%rd2+12]; setp.ne.s32 %p5, %r5, %r61; mov.u16 %rs118, 0; @%p5 bra $L__BB12_8; and.b32 %r62, %r2, -128; ld.u32 %r63, [%rd2+16]; setp.ne.s32 %p6, %r62, %r63; @%p6 bra $L__BB12_8; and.b32 %r64, %r3, -128; ld.u32 %r65, [%rd2+20]; setp.eq.s32 %p7, %r64, %r65; selp.u16 %rs118, 1, 0, %p7; $L__BB12_8: setp.eq.s16 %p8, %rs118, 0; @%p8 bra $L__BB12_12; shl.b32 %r66, %r1, 5; and.b32 %r67, %r66, 3840; shl.b32 %r68, %r2, 1; and.b32 %r69, %r68, 240; or.b32 %r70, %r69, %r67; bfe.u32 %r71, %r3, 3, 4; or.b32 %r72, %r70, %r71; shr.u32 %r73, %r70, 6; ld.u64 %rd3, [%rd2+56]; mul.wide.u32 %rd164, %r73, 8; add.s64 %rd165, %rd3, %rd164; and.b32 %r74, %r72, 63; mov.u64 %rd166, 1; shl.b64 %rd167, %rd166, %r74; ld.u64 %rd168, [%rd165+544]; and.b64 %rd169, %rd168, %rd167; setp.eq.s64 %p9, %rd169, 0; mul.wide.u32 %rd170, %r72, 8; add.s64 %rd171, %rd3, %rd170; add.s64 %rd4, %rd171, 1088; @%p9 bra $L__BB12_11; ld.u64 %rd172, [%rd4]; add.s64 %rd173, %rd3, %rd172; and.b32 %r75, %r2, -8; st.v2.u32 [%rd2], {%r4, %r75}; and.b32 %r76, %r3, -8; st.u32 [%rd2+8], %r76; st.u64 [%rd2+48], %rd173; shl.b32 %r77, %r1, 6; and.b32 %r78, %r77, 448; shl.b32 %r79, %r2, 3; and.b32 %r80, %r79, 56; or.b32 %r81, %r80, %r78; and.b32 %r82, %r3, 7; or.b32 %r83, %r81, %r82; mul.wide.u32 %rd174, %r83, 2; add.s64 %rd175, %rd173, %rd174; ld.u16 %rs48, [%rd175+96]; cvt.rn.f32.u16 %f114, %rs48; ld.v2.f32 {%f115, %f116}, [%rd173+80]; fma.rn.ftz.f32 %f280, %f116, %f114, %f115; bra.uni $L__BB12_35; $L__BB12_40: and.b32 %r11, %r1, -128; ld.u32 %r165, [%rd22+12]; setp.ne.s32 %p27, %r11, %r165; mov.u16 %rs121, 0; @%p27 bra $L__BB12_43; and.b32 %r166, %r2, -128; ld.u32 %r167, [%rd22+16]; setp.ne.s32 %p28, %r166, %r167; @%p28 bra $L__BB12_43; and.b32 %r168, %r10, -128; ld.u32 %r169, [%rd22+20]; setp.eq.s32 %p29, %r168, %r169; selp.u16 %rs121, 1, 0, %p29; $L__BB12_43: setp.eq.s16 %p30, %rs121, 0; @%p30 bra $L__BB12_47; shl.b32 %r170, %r1, 5; and.b32 %r171, %r170, 3840; shl.b32 %r172, %r2, 1; and.b32 %r173, %r172, 240; or.b32 %r174, %r173, %r171; bfe.u32 %r175, %r10, 3, 4; or.b32 %r176, %r175, %r174; shr.u32 %r177, %r174, 6; ld.u64 %rd23, [%rd22+56]; mul.wide.u32 %rd231, %r177, 8; add.s64 %rd232, %rd23, %rd231; and.b32 %r178, %r176, 63; mov.u64 %rd233, 1; shl.b64 %rd234, %rd233, %r178; ld.u64 %rd235, [%rd232+544]; and.b64 %rd236, %rd235, %rd234; setp.eq.s64 %p31, %rd236, 0; mul.wide.u32 %rd237, %r176, 8; add.s64 %rd238, %rd23, %rd237; add.s64 %rd24, %rd238, 1088; @%p31 bra $L__BB12_46; ld.u64 %rd239, [%rd24]; add.s64 %rd240, %rd23, %rd239; and.b32 %r179, %r2, -8; st.v2.u32 [%rd22], {%r4, %r179}; and.b32 %r180, %r10, -8; st.u32 [%rd22+8], %r180; st.u64 [%rd22+48], %rd240; shl.b32 %r181, %r1, 6; and.b32 %r182, %r181, 448; shl.b32 %r183, %r2, 3; and.b32 %r184, %r183, 56; or.b32 %r185, %r184, %r182; and.b32 %r186, %r10, 7; or.b32 %r187, %r185, %r186; mul.wide.u32 %rd241, %r187, 2; add.s64 %rd242, %rd240, %rd241; ld.u16 %rs58, [%rd242+96]; cvt.rn.f32.u16 %f134, %rs58; ld.v2.f32 {%f135, %f136}, [%rd240+80]; fma.rn.ftz.f32 %f281, %f136, %f134, %f135; bra.uni $L__BB12_70; $L__BB12_75: and.b32 %r17, %r1, -128; ld.u32 %r269, [%rd42+12]; setp.ne.s32 %p49, %r17, %r269; mov.u16 %rs124, 0; @%p49 bra $L__BB12_78; and.b32 %r270, %r16, -128; ld.u32 %r271, [%rd42+16]; setp.ne.s32 %p50, %r270, %r271; @%p50 bra $L__BB12_78; and.b32 %r272, %r10, -128; ld.u32 %r273, [%rd42+20]; setp.eq.s32 %p51, %r272, %r273; selp.u16 %rs124, 1, 0, %p51; $L__BB12_78: setp.eq.s16 %p52, %rs124, 0; @%p52 bra $L__BB12_82; shl.b32 %r274, %r1, 5; and.b32 %r275, %r274, 3840; shl.b32 %r276, %r16, 1; and.b32 %r277, %r276, 240; or.b32 %r278, %r277, %r275; bfe.u32 %r279, %r10, 3, 4; or.b32 %r280, %r278, %r279; shr.u32 %r281, %r278, 6; ld.u64 %rd43, [%rd42+56]; mul.wide.u32 %rd299, %r281, 8; add.s64 %rd300, %rd43, %rd299; and.b32 %r282, %r280, 63; mov.u64 %rd301, 1; shl.b64 %rd302, %rd301, %r282; ld.u64 %rd303, [%rd300+544]; and.b64 %rd304, %rd303, %rd302; setp.eq.s64 %p53, %rd304, 0; mul.wide.u32 %rd305, %r280, 8; add.s64 %rd306, %rd43, %rd305; add.s64 %rd44, %rd306, 1088; @%p53 bra $L__BB12_81; ld.u64 %rd307, [%rd44]; add.s64 %rd308, %rd43, %rd307; and.b32 %r283, %r16, -8; st.v2.u32 [%rd42], {%r4, %r283}; and.b32 %r284, %r10, -8; st.u32 [%rd42+8], %r284; st.u64 [%rd42+48], %rd308; shl.b32 %r285, %r1, 6; and.b32 %r286, %r285, 448; shl.b32 %r287, %r16, 3; and.b32 %r288, %r287, 56; or.b32 %r289, %r288, %r286; and.b32 %r290, %r10, 7; or.b32 %r291, %r289, %r290; mul.wide.u32 %rd309, %r291, 2; add.s64 %rd310, %rd308, %rd309; ld.u16 %rs68, [%rd310+96]; cvt.rn.f32.u16 %f155, %rs68; ld.v2.f32 {%f156, %f157}, [%rd308+80]; fma.rn.ftz.f32 %f282, %f157, %f155, %f156; bra.uni $L__BB12_105; $L__BB12_110: and.b32 %r22, %r1, -128; ld.u32 %r372, [%rd62+12]; setp.ne.s32 %p71, %r22, %r372; mov.u16 %rs127, 0; @%p71 bra $L__BB12_113; and.b32 %r373, %r16, -128; ld.u32 %r374, [%rd62+16]; setp.ne.s32 %p72, %r373, %r374; @%p72 bra $L__BB12_113; and.b32 %r375, %r3, -128; ld.u32 %r376, [%rd62+20]; setp.eq.s32 %p73, %r375, %r376; selp.u16 %rs127, 1, 0, %p73; $L__BB12_113: setp.eq.s16 %p74, %rs127, 0; @%p74 bra $L__BB12_117; shl.b32 %r377, %r1, 5; and.b32 %r378, %r377, 3840; shl.b32 %r379, %r16, 1; and.b32 %r380, %r379, 240; or.b32 %r381, %r380, %r378; bfe.u32 %r382, %r3, 3, 4; or.b32 %r383, %r381, %r382; shr.u32 %r384, %r381, 6; ld.u64 %rd63, [%rd62+56]; mul.wide.u32 %rd366, %r384, 8; add.s64 %rd367, %rd63, %rd366; and.b32 %r385, %r383, 63; mov.u64 %rd368, 1; shl.b64 %rd369, %rd368, %r385; ld.u64 %rd370, [%rd367+544]; and.b64 %rd371, %rd370, %rd369; setp.eq.s64 %p75, %rd371, 0; mul.wide.u32 %rd372, %r383, 8; add.s64 %rd373, %rd63, %rd372; add.s64 %rd64, %rd373, 1088; @%p75 bra $L__BB12_116; ld.u64 %rd374, [%rd64]; add.s64 %rd375, %rd63, %rd374; and.b32 %r386, %r16, -8; st.v2.u32 [%rd62], {%r4, %r386}; and.b32 %r387, %r3, -8; st.u32 [%rd62+8], %r387; st.u64 [%rd62+48], %rd375; and.b32 %r388, %r1, 7; shl.b32 %r389, %r16, 3; and.b32 %r390, %r389, 56; and.b32 %r391, %r3, 7; bfi.b32 %r392, %r388, %r391, 6, 3; or.b32 %r393, %r392, %r390; mul.wide.u32 %rd376, %r393, 2; add.s64 %rd377, %rd375, %rd376; ld.u16 %rs78, [%rd377+96]; cvt.rn.f32.u16 %f175, %rs78; ld.v2.f32 {%f176, %f177}, [%rd375+80]; fma.rn.ftz.f32 %f283, %f177, %f175, %f176; bra.uni $L__BB12_140; $L__BB12_145: and.b32 %r29, %r27, -128; ld.u32 %r472, [%rd81+12]; setp.ne.s32 %p93, %r29, %r472; mov.u16 %rs130, 0; @%p93 bra $L__BB12_148; and.b32 %r473, %r16, -128; ld.u32 %r474, [%rd81+16]; setp.ne.s32 %p94, %r473, %r474; @%p94 bra $L__BB12_148; and.b32 %r475, %r3, -128; ld.u32 %r476, [%rd81+20]; setp.eq.s32 %p95, %r475, %r476; selp.u16 %rs130, 1, 0, %p95; $L__BB12_148: setp.eq.s16 %p96, %rs130, 0; @%p96 bra $L__BB12_152; shl.b32 %r477, %r27, 5; and.b32 %r478, %r477, 3840; shl.b32 %r479, %r16, 1; and.b32 %r480, %r479, 240; or.b32 %r481, %r480, %r478; bfe.u32 %r482, %r3, 3, 4; or.b32 %r483, %r481, %r482; shr.u32 %r484, %r481, 6; ld.u64 %rd82, [%rd81+56]; mul.wide.u32 %rd433, %r484, 8; add.s64 %rd434, %rd82, %rd433; and.b32 %r485, %r483, 63; mov.u64 %rd435, 1; shl.b64 %rd436, %rd435, %r485; ld.u64 %rd437, [%rd434+544]; and.b64 %rd438, %rd437, %rd436; setp.eq.s64 %p97, %rd438, 0; mul.wide.u32 %rd439, %r483, 8; add.s64 %rd440, %rd82, %rd439; add.s64 %rd83, %rd440, 1088; @%p97 bra $L__BB12_151; ld.u64 %rd441, [%rd83]; add.s64 %rd442, %rd82, %rd441; and.b32 %r486, %r16, -8; st.v2.u32 [%rd81], {%r28, %r486}; and.b32 %r487, %r3, -8; st.u32 [%rd81+8], %r487; st.u64 [%rd81+48], %rd442; and.b32 %r488, %r27, 7; shl.b32 %r489, %r16, 3; and.b32 %r490, %r489, 56; and.b32 %r491, %r3, 7; bfi.b32 %r492, %r488, %r491, 6, 3; or.b32 %r493, %r492, %r490; mul.wide.u32 %rd443, %r493, 2; add.s64 %rd444, %rd442, %rd443; ld.u16 %rs88, [%rd444+96]; cvt.rn.f32.u16 %f198, %rs88; ld.v2.f32 {%f199, %f200}, [%rd442+80]; fma.rn.ftz.f32 %f284, %f200, %f198, %f199; bra.uni $L__BB12_175; $L__BB12_180: and.b32 %r34, %r27, -128; ld.u32 %r573, [%rd101+12]; setp.ne.s32 %p115, %r34, %r573; mov.u16 %rs133, 0; @%p115 bra $L__BB12_183; and.b32 %r574, %r16, -128; ld.u32 %r575, [%rd101+16]; setp.ne.s32 %p116, %r574, %r575; @%p116 bra $L__BB12_183; and.b32 %r576, %r10, -128; ld.u32 %r577, [%rd101+20]; setp.eq.s32 %p117, %r576, %r577; selp.u16 %rs133, 1, 0, %p117; $L__BB12_183: setp.eq.s16 %p118, %rs133, 0; @%p118 bra $L__BB12_187; shl.b32 %r578, %r27, 5; and.b32 %r579, %r578, 3840; shl.b32 %r580, %r16, 1; and.b32 %r581, %r580, 240; or.b32 %r582, %r581, %r579; bfe.u32 %r583, %r10, 3, 4; or.b32 %r584, %r582, %r583; shr.u32 %r585, %r582, 6; ld.u64 %rd102, [%rd101+56]; mul.wide.u32 %rd500, %r585, 8; add.s64 %rd501, %rd102, %rd500; and.b32 %r586, %r584, 63; mov.u64 %rd502, 1; shl.b64 %rd503, %rd502, %r586; ld.u64 %rd504, [%rd501+544]; and.b64 %rd505, %rd504, %rd503; setp.eq.s64 %p119, %rd505, 0; mul.wide.u32 %rd506, %r584, 8; add.s64 %rd507, %rd102, %rd506; add.s64 %rd103, %rd507, 1088; @%p119 bra $L__BB12_186; ld.u64 %rd508, [%rd103]; add.s64 %rd509, %rd102, %rd508; and.b32 %r587, %r16, -8; st.v2.u32 [%rd101], {%r28, %r587}; and.b32 %r588, %r10, -8; st.u32 [%rd101+8], %r588; st.u64 [%rd101+48], %rd509; shl.b32 %r589, %r27, 6; and.b32 %r590, %r589, 448; shl.b32 %r591, %r16, 3; and.b32 %r592, %r591, 56; or.b32 %r593, %r592, %r590; and.b32 %r594, %r10, 7; or.b32 %r595, %r593, %r594; mul.wide.u32 %rd510, %r595, 2; add.s64 %rd511, %rd509, %rd510; ld.u16 %rs98, [%rd511+96]; cvt.rn.f32.u16 %f218, %rs98; ld.v2.f32 {%f219, %f220}, [%rd509+80]; fma.rn.ftz.f32 %f285, %f220, %f218, %f219; bra.uni $L__BB12_210; $L__BB12_215: and.b32 %r39, %r27, -128; ld.u32 %r677, [%rd120+12]; setp.ne.s32 %p137, %r39, %r677; mov.u16 %rs136, 0; @%p137 bra $L__BB12_218; and.b32 %r678, %r2, -128; ld.u32 %r679, [%rd120+16]; setp.ne.s32 %p138, %r678, %r679; @%p138 bra $L__BB12_218; and.b32 %r680, %r10, -128; ld.u32 %r681, [%rd120+20]; setp.eq.s32 %p139, %r680, %r681; selp.u16 %rs136, 1, 0, %p139; $L__BB12_218: setp.eq.s16 %p140, %rs136, 0; @%p140 bra $L__BB12_222; shl.b32 %r682, %r27, 5; and.b32 %r683, %r682, 3840; shl.b32 %r684, %r2, 1; and.b32 %r685, %r684, 240; or.b32 %r686, %r683, %r685; bfe.u32 %r687, %r10, 3, 4; or.b32 %r688, %r687, %r686; shr.u32 %r689, %r686, 6; ld.u64 %rd121, [%rd120+56]; mul.wide.u32 %rd567, %r689, 8; add.s64 %rd568, %rd121, %rd567; and.b32 %r690, %r688, 63; mov.u64 %rd569, 1; shl.b64 %rd570, %rd569, %r690; ld.u64 %rd571, [%rd568+544]; and.b64 %rd572, %rd571, %rd570; setp.eq.s64 %p141, %rd572, 0; mul.wide.u32 %rd573, %r688, 8; add.s64 %rd574, %rd121, %rd573; add.s64 %rd122, %rd574, 1088; @%p141 bra $L__BB12_221; ld.u64 %rd575, [%rd122]; add.s64 %rd576, %rd121, %rd575; and.b32 %r691, %r2, -8; st.v2.u32 [%rd120], {%r28, %r691}; and.b32 %r692, %r10, -8; st.u32 [%rd120+8], %r692; st.u64 [%rd120+48], %rd576; shl.b32 %r693, %r27, 6; and.b32 %r694, %r693, 448; shl.b32 %r695, %r2, 3; and.b32 %r696, %r695, 56; or.b32 %r697, %r694, %r696; and.b32 %r698, %r10, 7; or.b32 %r699, %r697, %r698; mul.wide.u32 %rd577, %r699, 2; add.s64 %rd578, %rd576, %rd577; ld.u16 %rs108, [%rd578+96]; cvt.rn.f32.u16 %f239, %rs108; ld.v2.f32 {%f240, %f241}, [%rd576+80]; fma.rn.ftz.f32 %f286, %f241, %f239, %f240; bra.uni $L__BB12_245; $L__BB12_258: shl.b64 %rd642, %rd145, 5; add.s64 %rd707, %rd143, %rd642; $L__BB12_259: setp.eq.s64 %p167, %rd707, 0; @%p167 bra $L__BB12_267; add.s64 %rd148, %rd707, 8; ld.u64 %rd149, [%rd707+8]; setp.eq.s64 %p168, %rd149, 0; @%p168 bra $L__BB12_266; add.s64 %rd150, %rd141, %rd149; and.b32 %r788, %r2, -4096; mov.u64 %rd643, 1; st.v2.u32 [%rd140+24], {%r45, %r788}; and.b32 %r789, %r3, -4096; st.u32 [%rd140+32], %r789; st.u64 [%rd140+64], %rd150; shl.b32 %r790, %r27, 3; and.b32 %r791, %r790, 31744; and.b32 %r792, %r2, 3968; shr.u32 %r793, %r792, 2; or.b32 %r794, %r791, %r793; bfe.u32 %r795, %r3, 7, 5; or.b32 %r796, %r794, %r795; shr.u32 %r797, %r794, 6; mul.wide.u32 %rd644, %r797, 8; add.s64 %rd645, %rd150, %rd644; and.b32 %r798, %r796, 63; shl.b64 %rd646, %rd643, %r798; ld.u64 %rd647, [%rd645+4128]; and.b64 %rd648, %rd647, %rd646; setp.eq.s64 %p169, %rd648, 0; mul.wide.u32 %rd649, %r796, 8; add.s64 %rd650, %rd150, %rd649; add.s64 %rd151, %rd650, 8256; @%p169 bra $L__BB12_265; ld.u64 %rd651, [%rd151]; add.s64 %rd152, %rd150, %rd651; st.u32 [%rd140+12], %r44; and.b32 %r799, %r2, -128; st.u32 [%rd140+16], %r799; and.b32 %r800, %r3, -128; st.u32 [%rd140+20], %r800; st.u64 [%rd140+56], %rd152; shl.b32 %r801, %r27, 5; and.b32 %r802, %r801, 3840; shl.b32 %r803, %r2, 1; and.b32 %r804, %r803, 240; or.b32 %r805, %r802, %r804; bfe.u32 %r806, %r3, 3, 4; or.b32 %r807, %r805, %r806; shr.u32 %r808, %r805, 6; mul.wide.u32 %rd652, %r808, 8; add.s64 %rd653, %rd152, %rd652; and.b32 %r809, %r807, 63; shl.b64 %rd655, %rd643, %r809; ld.u64 %rd656, [%rd653+544]; and.b64 %rd657, %rd656, %rd655; setp.eq.s64 %p170, %rd657, 0; mul.wide.u32 %rd658, %r807, 8; add.s64 %rd659, %rd152, %rd658; add.s64 %rd153, %rd659, 1088; @%p170 bra $L__BB12_264; ld.u64 %rd660, [%rd153]; add.s64 %rd661, %rd152, %rd660; and.b32 %r810, %r2, -8; st.v2.u32 [%rd140], {%r28, %r810}; and.b32 %r811, %r3, -8; st.u32 [%rd140+8], %r811; st.u64 [%rd140+48], %rd661; shl.b32 %r812, %r27, 6; and.b32 %r813, %r812, 448; shl.b32 %r814, %r2, 3; and.b32 %r815, %r814, 56; or.b32 %r816, %r813, %r815; and.b32 %r817, %r3, 7; or.b32 %r818, %r816, %r817; mul.wide.u32 %rd662, %r818, 2; add.s64 %rd663, %rd661, %rd662; ld.u16 %rs113, [%rd663+96]; cvt.rn.f32.u16 %f254, %rs113; ld.v2.f32 {%f255, %f256}, [%rd661+80]; fma.rn.ftz.f32 %f287, %f256, %f254, %f255; bra.uni $L__BB12_277; $L__BB12_267: ld.f32 %f287, [%rd142+4]; bra.uni $L__BB12_277; $L__BB12_266: ld.f32 %f287, [%rd148+12]; bra.uni $L__BB12_277; $L__BB12_12: and.b32 %r6, %r1, -4096; ld.u32 %r84, [%rd2+24]; setp.ne.s32 %p10, %r6, %r84; mov.u16 %rs119, 0; @%p10 bra $L__BB12_15; and.b32 %r85, %r2, -4096; ld.u32 %r86, [%rd2+28]; setp.ne.s32 %p11, %r85, %r86; @%p11 bra $L__BB12_15; and.b32 %r87, %r3, -4096; ld.u32 %r88, [%rd2+32]; setp.eq.s32 %p12, %r87, %r88; selp.u16 %rs119, 1, 0, %p12; $L__BB12_15: setp.eq.s16 %p13, %rs119, 0; @%p13 bra $L__BB12_21; shl.b32 %r89, %r1, 3; and.b32 %r90, %r89, 31744; and.b32 %r91, %r2, 3968; shr.u32 %r92, %r91, 2; or.b32 %r93, %r92, %r90; bfe.u32 %r94, %r3, 7, 5; or.b32 %r95, %r93, %r94; shr.u32 %r96, %r93, 6; ld.u64 %rd5, [%rd2+64]; mul.wide.u32 %rd176, %r96, 8; add.s64 %rd177, %rd5, %rd176; and.b32 %r97, %r95, 63; mov.u64 %rd178, 1; shl.b64 %rd179, %rd178, %r97; ld.u64 %rd180, [%rd177+4128]; and.b64 %rd181, %rd180, %rd179; setp.eq.s64 %p14, %rd181, 0; mul.wide.u32 %rd182, %r95, 8; add.s64 %rd183, %rd5, %rd182; add.s64 %rd6, %rd183, 8256; @%p14 bra $L__BB12_20; ld.u64 %rd184, [%rd6]; add.s64 %rd7, %rd5, %rd184; st.u32 [%rd2+12], %r5; and.b32 %r98, %r2, -128; st.u32 [%rd2+16], %r98; and.b32 %r99, %r3, -128; st.u32 [%rd2+20], %r99; st.u64 [%rd2+56], %rd7; shl.b32 %r100, %r1, 5; and.b32 %r101, %r100, 3840; shl.b32 %r102, %r2, 1; and.b32 %r103, %r102, 240; or.b32 %r104, %r103, %r101; bfe.u32 %r105, %r3, 3, 4; or.b32 %r106, %r104, %r105; shr.u32 %r107, %r104, 6; mul.wide.u32 %rd185, %r107, 8; add.s64 %rd186, %rd7, %rd185; and.b32 %r108, %r106, 63; shl.b64 %rd188, %rd178, %r108; ld.u64 %rd189, [%rd186+544]; and.b64 %rd190, %rd189, %rd188; setp.eq.s64 %p15, %rd190, 0; mul.wide.u32 %rd191, %r106, 8; add.s64 %rd192, %rd7, %rd191; add.s64 %rd8, %rd192, 1088; @%p15 bra $L__BB12_19; ld.u64 %rd193, [%rd8]; add.s64 %rd194, %rd7, %rd193; and.b32 %r109, %r2, -8; st.v2.u32 [%rd2], {%r4, %r109}; and.b32 %r110, %r3, -8; st.u32 [%rd2+8], %r110; st.u64 [%rd2+48], %rd194; shl.b32 %r111, %r1, 6; and.b32 %r112, %r111, 448; shl.b32 %r113, %r2, 3; and.b32 %r114, %r113, 56; or.b32 %r115, %r114, %r112; and.b32 %r116, %r3, 7; or.b32 %r117, %r115, %r116; mul.wide.u32 %rd195, %r117, 2; add.s64 %rd196, %rd194, %rd195; ld.u16 %rs51, [%rd196+96]; cvt.rn.f32.u16 %f119, %rs51; ld.v2.f32 {%f120, %f121}, [%rd194+80]; fma.rn.ftz.f32 %f280, %f121, %f119, %f120; bra.uni $L__BB12_35; $L__BB12_47: and.b32 %r12, %r1, -4096; ld.u32 %r188, [%rd22+24]; setp.ne.s32 %p32, %r12, %r188; mov.u16 %rs122, 0; @%p32 bra $L__BB12_50; and.b32 %r189, %r2, -4096; ld.u32 %r190, [%rd22+28]; setp.ne.s32 %p33, %r189, %r190; @%p33 bra $L__BB12_50; and.b32 %r191, %r10, -4096; ld.u32 %r192, [%rd22+32]; setp.eq.s32 %p34, %r191, %r192; selp.u16 %rs122, 1, 0, %p34; $L__BB12_50: setp.eq.s16 %p35, %rs122, 0; @%p35 bra $L__BB12_56; shl.b32 %r193, %r1, 3; and.b32 %r194, %r193, 31744; and.b32 %r195, %r2, 3968; shr.u32 %r196, %r195, 2; or.b32 %r197, %r196, %r194; bfe.u32 %r198, %r10, 7, 5; or.b32 %r199, %r198, %r197; shr.u32 %r200, %r197, 6; ld.u64 %rd25, [%rd22+64]; mul.wide.u32 %rd243, %r200, 8; add.s64 %rd244, %rd25, %rd243; and.b32 %r201, %r199, 63; mov.u64 %rd245, 1; shl.b64 %rd246, %rd245, %r201; ld.u64 %rd247, [%rd244+4128]; and.b64 %rd248, %rd247, %rd246; setp.eq.s64 %p36, %rd248, 0; mul.wide.u32 %rd249, %r199, 8; add.s64 %rd250, %rd25, %rd249; add.s64 %rd26, %rd250, 8256; @%p36 bra $L__BB12_55; ld.u64 %rd251, [%rd26]; add.s64 %rd27, %rd25, %rd251; st.u32 [%rd22+12], %r11; and.b32 %r202, %r2, -128; st.u32 [%rd22+16], %r202; and.b32 %r203, %r10, -128; st.u32 [%rd22+20], %r203; st.u64 [%rd22+56], %rd27; shl.b32 %r204, %r1, 5; and.b32 %r205, %r204, 3840; shl.b32 %r206, %r2, 1; and.b32 %r207, %r206, 240; or.b32 %r208, %r207, %r205; bfe.u32 %r209, %r10, 3, 4; or.b32 %r210, %r209, %r208; shr.u32 %r211, %r208, 6; mul.wide.u32 %rd252, %r211, 8; add.s64 %rd253, %rd27, %rd252; and.b32 %r212, %r210, 63; shl.b64 %rd255, %rd245, %r212; ld.u64 %rd256, [%rd253+544]; and.b64 %rd257, %rd256, %rd255; setp.eq.s64 %p37, %rd257, 0; mul.wide.u32 %rd258, %r210, 8; add.s64 %rd259, %rd27, %rd258; add.s64 %rd28, %rd259, 1088; @%p37 bra $L__BB12_54; ld.u64 %rd260, [%rd28]; add.s64 %rd261, %rd27, %rd260; and.b32 %r213, %r2, -8; st.v2.u32 [%rd22], {%r4, %r213}; and.b32 %r214, %r10, -8; st.u32 [%rd22+8], %r214; st.u64 [%rd22+48], %rd261; shl.b32 %r215, %r1, 6; and.b32 %r216, %r215, 448; shl.b32 %r217, %r2, 3; and.b32 %r218, %r217, 56; or.b32 %r219, %r218, %r216; and.b32 %r220, %r10, 7; or.b32 %r221, %r219, %r220; mul.wide.u32 %rd262, %r221, 2; add.s64 %rd263, %rd261, %rd262; ld.u16 %rs61, [%rd263+96]; cvt.rn.f32.u16 %f139, %rs61; ld.v2.f32 {%f140, %f141}, [%rd261+80]; fma.rn.ftz.f32 %f281, %f141, %f139, %f140; bra.uni $L__BB12_70; $L__BB12_82: and.b32 %r18, %r1, -4096; ld.u32 %r292, [%rd42+24]; setp.ne.s32 %p54, %r18, %r292; mov.u16 %rs125, 0; @%p54 bra $L__BB12_85; and.b32 %r293, %r16, -4096; ld.u32 %r294, [%rd42+28]; setp.ne.s32 %p55, %r293, %r294; @%p55 bra $L__BB12_85; and.b32 %r295, %r10, -4096; ld.u32 %r296, [%rd42+32]; setp.eq.s32 %p56, %r295, %r296; selp.u16 %rs125, 1, 0, %p56; $L__BB12_85: setp.eq.s16 %p57, %rs125, 0; @%p57 bra $L__BB12_91; shl.b32 %r297, %r1, 3; and.b32 %r298, %r297, 31744; and.b32 %r299, %r16, 3968; shr.u32 %r300, %r299, 2; or.b32 %r301, %r300, %r298; bfe.u32 %r302, %r10, 7, 5; or.b32 %r303, %r301, %r302; shr.u32 %r304, %r301, 6; ld.u64 %rd45, [%rd42+64]; mul.wide.u32 %rd311, %r304, 8; add.s64 %rd312, %rd45, %rd311; and.b32 %r305, %r303, 63; mov.u64 %rd313, 1; shl.b64 %rd314, %rd313, %r305; ld.u64 %rd315, [%rd312+4128]; and.b64 %rd316, %rd315, %rd314; setp.eq.s64 %p58, %rd316, 0; mul.wide.u32 %rd317, %r303, 8; add.s64 %rd318, %rd45, %rd317; add.s64 %rd46, %rd318, 8256; @%p58 bra $L__BB12_90; ld.u64 %rd319, [%rd46]; add.s64 %rd47, %rd45, %rd319; st.u32 [%rd42+12], %r17; and.b32 %r306, %r16, -128; st.u32 [%rd42+16], %r306; and.b32 %r307, %r10, -128; st.u32 [%rd42+20], %r307; st.u64 [%rd42+56], %rd47; shl.b32 %r308, %r1, 5; and.b32 %r309, %r308, 3840; shl.b32 %r310, %r16, 1; and.b32 %r311, %r310, 240; or.b32 %r312, %r311, %r309; bfe.u32 %r313, %r10, 3, 4; or.b32 %r314, %r312, %r313; shr.u32 %r315, %r312, 6; mul.wide.u32 %rd320, %r315, 8; add.s64 %rd321, %rd47, %rd320; and.b32 %r316, %r314, 63; shl.b64 %rd323, %rd313, %r316; ld.u64 %rd324, [%rd321+544]; and.b64 %rd325, %rd324, %rd323; setp.eq.s64 %p59, %rd325, 0; mul.wide.u32 %rd326, %r314, 8; add.s64 %rd327, %rd47, %rd326; add.s64 %rd48, %rd327, 1088; @%p59 bra $L__BB12_89; ld.u64 %rd328, [%rd48]; add.s64 %rd329, %rd47, %rd328; and.b32 %r317, %r16, -8; st.v2.u32 [%rd42], {%r4, %r317}; and.b32 %r318, %r10, -8; st.u32 [%rd42+8], %r318; st.u64 [%rd42+48], %rd329; shl.b32 %r319, %r1, 6; and.b32 %r320, %r319, 448; shl.b32 %r321, %r16, 3; and.b32 %r322, %r321, 56; or.b32 %r323, %r322, %r320; and.b32 %r324, %r10, 7; or.b32 %r325, %r323, %r324; mul.wide.u32 %rd330, %r325, 2; add.s64 %rd331, %rd329, %rd330; ld.u16 %rs71, [%rd331+96]; cvt.rn.f32.u16 %f160, %rs71; ld.v2.f32 {%f161, %f162}, [%rd329+80]; fma.rn.ftz.f32 %f282, %f162, %f160, %f161; bra.uni $L__BB12_105; $L__BB12_117: and.b32 %r23, %r1, -4096; ld.u32 %r394, [%rd62+24]; setp.ne.s32 %p76, %r23, %r394; mov.u16 %rs128, 0; @%p76 bra $L__BB12_120; and.b32 %r395, %r16, -4096; ld.u32 %r396, [%rd62+28]; setp.ne.s32 %p77, %r395, %r396; @%p77 bra $L__BB12_120; and.b32 %r397, %r3, -4096; ld.u32 %r398, [%rd62+32]; setp.eq.s32 %p78, %r397, %r398; selp.u16 %rs128, 1, 0, %p78; $L__BB12_120: setp.eq.s16 %p79, %rs128, 0; @%p79 bra $L__BB12_126; shl.b32 %r399, %r1, 3; and.b32 %r400, %r399, 31744; and.b32 %r401, %r16, 3968; shr.u32 %r402, %r401, 2; or.b32 %r403, %r402, %r400; bfe.u32 %r404, %r3, 7, 5; or.b32 %r405, %r403, %r404; shr.u32 %r406, %r403, 6; ld.u64 %rd65, [%rd62+64]; mul.wide.u32 %rd378, %r406, 8; add.s64 %rd379, %rd65, %rd378; and.b32 %r407, %r405, 63; mov.u64 %rd380, 1; shl.b64 %rd381, %rd380, %r407; ld.u64 %rd382, [%rd379+4128]; and.b64 %rd383, %rd382, %rd381; setp.eq.s64 %p80, %rd383, 0; mul.wide.u32 %rd384, %r405, 8; add.s64 %rd385, %rd65, %rd384; add.s64 %rd66, %rd385, 8256; @%p80 bra $L__BB12_125; ld.u64 %rd386, [%rd66]; add.s64 %rd67, %rd65, %rd386; st.u32 [%rd62+12], %r22; and.b32 %r408, %r16, -128; st.u32 [%rd62+16], %r408; and.b32 %r409, %r3, -128; st.u32 [%rd62+20], %r409; st.u64 [%rd62+56], %rd67; shl.b32 %r410, %r1, 5; and.b32 %r411, %r410, 3840; shl.b32 %r412, %r16, 1; and.b32 %r413, %r412, 240; or.b32 %r414, %r413, %r411; bfe.u32 %r415, %r3, 3, 4; or.b32 %r416, %r414, %r415; shr.u32 %r417, %r414, 6; mul.wide.u32 %rd387, %r417, 8; add.s64 %rd388, %rd67, %rd387; and.b32 %r418, %r416, 63; shl.b64 %rd390, %rd380, %r418; ld.u64 %rd391, [%rd388+544]; and.b64 %rd392, %rd391, %rd390; setp.eq.s64 %p81, %rd392, 0; mul.wide.u32 %rd393, %r416, 8; add.s64 %rd394, %rd67, %rd393; add.s64 %rd68, %rd394, 1088; @%p81 bra $L__BB12_124; ld.u64 %rd395, [%rd68]; add.s64 %rd396, %rd67, %rd395; and.b32 %r419, %r16, -8; st.v2.u32 [%rd62], {%r4, %r419}; and.b32 %r420, %r3, -8; st.u32 [%rd62+8], %r420; st.u64 [%rd62+48], %rd396; and.b32 %r421, %r1, 7; shl.b32 %r422, %r16, 3; and.b32 %r423, %r422, 56; and.b32 %r424, %r3, 7; bfi.b32 %r425, %r421, %r424, 6, 3; or.b32 %r426, %r425, %r423; mul.wide.u32 %rd397, %r426, 2; add.s64 %rd398, %rd396, %rd397; ld.u16 %rs81, [%rd398+96]; cvt.rn.f32.u16 %f180, %rs81; ld.v2.f32 {%f181, %f182}, [%rd396+80]; fma.rn.ftz.f32 %f283, %f182, %f180, %f181; bra.uni $L__BB12_140; $L__BB12_152: and.b32 %r30, %r27, -4096; ld.u32 %r494, [%rd81+24]; setp.ne.s32 %p98, %r30, %r494; mov.u16 %rs131, 0; @%p98 bra $L__BB12_155; and.b32 %r495, %r16, -4096; ld.u32 %r496, [%rd81+28]; setp.ne.s32 %p99, %r495, %r496; @%p99 bra $L__BB12_155; and.b32 %r497, %r3, -4096; ld.u32 %r498, [%rd81+32]; setp.eq.s32 %p100, %r497, %r498; selp.u16 %rs131, 1, 0, %p100; $L__BB12_155: setp.eq.s16 %p101, %rs131, 0; @%p101 bra $L__BB12_161; shl.b32 %r499, %r27, 3; and.b32 %r500, %r499, 31744; and.b32 %r501, %r16, 3968; shr.u32 %r502, %r501, 2; or.b32 %r503, %r502, %r500; bfe.u32 %r504, %r3, 7, 5; or.b32 %r505, %r503, %r504; shr.u32 %r506, %r503, 6; ld.u64 %rd84, [%rd81+64]; mul.wide.u32 %rd445, %r506, 8; add.s64 %rd446, %rd84, %rd445; and.b32 %r507, %r505, 63; mov.u64 %rd447, 1; shl.b64 %rd448, %rd447, %r507; ld.u64 %rd449, [%rd446+4128]; and.b64 %rd450, %rd449, %rd448; setp.eq.s64 %p102, %rd450, 0; mul.wide.u32 %rd451, %r505, 8; add.s64 %rd452, %rd84, %rd451; add.s64 %rd85, %rd452, 8256; @%p102 bra $L__BB12_160; ld.u64 %rd453, [%rd85]; add.s64 %rd86, %rd84, %rd453; st.u32 [%rd81+12], %r29; and.b32 %r508, %r16, -128; st.u32 [%rd81+16], %r508; and.b32 %r509, %r3, -128; st.u32 [%rd81+20], %r509; st.u64 [%rd81+56], %rd86; shl.b32 %r510, %r27, 5; and.b32 %r511, %r510, 3840; shl.b32 %r512, %r16, 1; and.b32 %r513, %r512, 240; or.b32 %r514, %r513, %r511; bfe.u32 %r515, %r3, 3, 4; or.b32 %r516, %r514, %r515; shr.u32 %r517, %r514, 6; mul.wide.u32 %rd454, %r517, 8; add.s64 %rd455, %rd86, %rd454; and.b32 %r518, %r516, 63; shl.b64 %rd457, %rd447, %r518; ld.u64 %rd458, [%rd455+544]; and.b64 %rd459, %rd458, %rd457; setp.eq.s64 %p103, %rd459, 0; mul.wide.u32 %rd460, %r516, 8; add.s64 %rd461, %rd86, %rd460; add.s64 %rd87, %rd461, 1088; @%p103 bra $L__BB12_159; ld.u64 %rd462, [%rd87]; add.s64 %rd463, %rd86, %rd462; and.b32 %r519, %r16, -8; st.v2.u32 [%rd81], {%r28, %r519}; and.b32 %r520, %r3, -8; st.u32 [%rd81+8], %r520; st.u64 [%rd81+48], %rd463; and.b32 %r521, %r27, 7; shl.b32 %r522, %r16, 3; and.b32 %r523, %r522, 56; and.b32 %r524, %r3, 7; bfi.b32 %r525, %r521, %r524, 6, 3; or.b32 %r526, %r525, %r523; mul.wide.u32 %rd464, %r526, 2; add.s64 %rd465, %rd463, %rd464; ld.u16 %rs91, [%rd465+96]; cvt.rn.f32.u16 %f203, %rs91; ld.v2.f32 {%f204, %f205}, [%rd463+80]; fma.rn.ftz.f32 %f284, %f205, %f203, %f204; bra.uni $L__BB12_175; $L__BB12_187: and.b32 %r35, %r27, -4096; ld.u32 %r596, [%rd101+24]; setp.ne.s32 %p120, %r35, %r596; mov.u16 %rs134, 0; @%p120 bra $L__BB12_190; and.b32 %r597, %r16, -4096; ld.u32 %r598, [%rd101+28]; setp.ne.s32 %p121, %r597, %r598; @%p121 bra $L__BB12_190; and.b32 %r599, %r10, -4096; ld.u32 %r600, [%rd101+32]; setp.eq.s32 %p122, %r599, %r600; selp.u16 %rs134, 1, 0, %p122; $L__BB12_190: setp.eq.s16 %p123, %rs134, 0; @%p123 bra $L__BB12_196; shl.b32 %r601, %r27, 3; and.b32 %r602, %r601, 31744; and.b32 %r603, %r16, 3968; shr.u32 %r604, %r603, 2; or.b32 %r605, %r604, %r602; bfe.u32 %r606, %r10, 7, 5; or.b32 %r607, %r605, %r606; shr.u32 %r608, %r605, 6; ld.u64 %rd104, [%rd101+64]; mul.wide.u32 %rd512, %r608, 8; add.s64 %rd513, %rd104, %rd512; and.b32 %r609, %r607, 63; mov.u64 %rd514, 1; shl.b64 %rd515, %rd514, %r609; ld.u64 %rd516, [%rd513+4128]; and.b64 %rd517, %rd516, %rd515; setp.eq.s64 %p124, %rd517, 0; mul.wide.u32 %rd518, %r607, 8; add.s64 %rd519, %rd104, %rd518; add.s64 %rd105, %rd519, 8256; @%p124 bra $L__BB12_195; ld.u64 %rd520, [%rd105]; add.s64 %rd106, %rd104, %rd520; st.u32 [%rd101+12], %r34; and.b32 %r610, %r16, -128; st.u32 [%rd101+16], %r610; and.b32 %r611, %r10, -128; st.u32 [%rd101+20], %r611; st.u64 [%rd101+56], %rd106; shl.b32 %r612, %r27, 5; and.b32 %r613, %r612, 3840; shl.b32 %r614, %r16, 1; and.b32 %r615, %r614, 240; or.b32 %r616, %r615, %r613; bfe.u32 %r617, %r10, 3, 4; or.b32 %r618, %r616, %r617; shr.u32 %r619, %r616, 6; mul.wide.u32 %rd521, %r619, 8; add.s64 %rd522, %rd106, %rd521; and.b32 %r620, %r618, 63; shl.b64 %rd524, %rd514, %r620; ld.u64 %rd525, [%rd522+544]; and.b64 %rd526, %rd525, %rd524; setp.eq.s64 %p125, %rd526, 0; mul.wide.u32 %rd527, %r618, 8; add.s64 %rd528, %rd106, %rd527; add.s64 %rd107, %rd528, 1088; @%p125 bra $L__BB12_194; ld.u64 %rd529, [%rd107]; add.s64 %rd530, %rd106, %rd529; and.b32 %r621, %r16, -8; st.v2.u32 [%rd101], {%r28, %r621}; and.b32 %r622, %r10, -8; st.u32 [%rd101+8], %r622; st.u64 [%rd101+48], %rd530; shl.b32 %r623, %r27, 6; and.b32 %r624, %r623, 448; shl.b32 %r625, %r16, 3; and.b32 %r626, %r625, 56; or.b32 %r627, %r626, %r624; and.b32 %r628, %r10, 7; or.b32 %r629, %r627, %r628; mul.wide.u32 %rd531, %r629, 2; add.s64 %rd532, %rd530, %rd531; ld.u16 %rs101, [%rd532+96]; cvt.rn.f32.u16 %f223, %rs101; ld.v2.f32 {%f224, %f225}, [%rd530+80]; fma.rn.ftz.f32 %f285, %f225, %f223, %f224; bra.uni $L__BB12_210; $L__BB12_222: and.b32 %r40, %r27, -4096; ld.u32 %r700, [%rd120+24]; setp.ne.s32 %p142, %r40, %r700; mov.u16 %rs137, 0; @%p142 bra $L__BB12_225; and.b32 %r701, %r2, -4096; ld.u32 %r702, [%rd120+28]; setp.ne.s32 %p143, %r701, %r702; @%p143 bra $L__BB12_225; and.b32 %r703, %r10, -4096; ld.u32 %r704, [%rd120+32]; setp.eq.s32 %p144, %r703, %r704; selp.u16 %rs137, 1, 0, %p144; $L__BB12_225: setp.eq.s16 %p145, %rs137, 0; @%p145 bra $L__BB12_231; shl.b32 %r705, %r27, 3; and.b32 %r706, %r705, 31744; and.b32 %r707, %r2, 3968; shr.u32 %r708, %r707, 2; or.b32 %r709, %r706, %r708; bfe.u32 %r710, %r10, 7, 5; or.b32 %r711, %r710, %r709; shr.u32 %r712, %r709, 6; ld.u64 %rd123, [%rd120+64]; mul.wide.u32 %rd579, %r712, 8; add.s64 %rd580, %rd123, %rd579; and.b32 %r713, %r711, 63; mov.u64 %rd581, 1; shl.b64 %rd582, %rd581, %r713; ld.u64 %rd583, [%rd580+4128]; and.b64 %rd584, %rd583, %rd582; setp.eq.s64 %p146, %rd584, 0; mul.wide.u32 %rd585, %r711, 8; add.s64 %rd586, %rd123, %rd585; add.s64 %rd124, %rd586, 8256; @%p146 bra $L__BB12_230; ld.u64 %rd587, [%rd124]; add.s64 %rd125, %rd123, %rd587; st.u32 [%rd120+12], %r39; and.b32 %r714, %r2, -128; st.u32 [%rd120+16], %r714; and.b32 %r715, %r10, -128; st.u32 [%rd120+20], %r715; st.u64 [%rd120+56], %rd125; shl.b32 %r716, %r27, 5; and.b32 %r717, %r716, 3840; shl.b32 %r718, %r2, 1; and.b32 %r719, %r718, 240; or.b32 %r720, %r717, %r719; bfe.u32 %r721, %r10, 3, 4; or.b32 %r722, %r721, %r720; shr.u32 %r723, %r720, 6; mul.wide.u32 %rd588, %r723, 8; add.s64 %rd589, %rd125, %rd588; and.b32 %r724, %r722, 63; shl.b64 %rd591, %rd581, %r724; ld.u64 %rd592, [%rd589+544]; and.b64 %rd593, %rd592, %rd591; setp.eq.s64 %p147, %rd593, 0; mul.wide.u32 %rd594, %r722, 8; add.s64 %rd595, %rd125, %rd594; add.s64 %rd126, %rd595, 1088; @%p147 bra $L__BB12_229; ld.u64 %rd596, [%rd126]; add.s64 %rd597, %rd125, %rd596; and.b32 %r725, %r2, -8; st.v2.u32 [%rd120], {%r28, %r725}; and.b32 %r726, %r10, -8; st.u32 [%rd120+8], %r726; st.u64 [%rd120+48], %rd597; shl.b32 %r727, %r27, 6; and.b32 %r728, %r727, 448; shl.b32 %r729, %r2, 3; and.b32 %r730, %r729, 56; or.b32 %r731, %r728, %r730; and.b32 %r732, %r10, 7; or.b32 %r733, %r731, %r732; mul.wide.u32 %rd598, %r733, 2; add.s64 %rd599, %rd597, %rd598; ld.u16 %rs111, [%rd599+96]; cvt.rn.f32.u16 %f244, %rs111; ld.v2.f32 {%f245, %f246}, [%rd597+80]; fma.rn.ftz.f32 %f286, %f246, %f244, %f245; bra.uni $L__BB12_245; $L__BB12_265: ld.f32 %f287, [%rd151]; bra.uni $L__BB12_277; $L__BB12_11: ld.f32 %f280, [%rd4]; bra.uni $L__BB12_35; $L__BB12_46: ld.f32 %f281, [%rd24]; bra.uni $L__BB12_70; $L__BB12_81: ld.f32 %f282, [%rd44]; bra.uni $L__BB12_105; $L__BB12_116: ld.f32 %f283, [%rd64]; bra.uni $L__BB12_140; $L__BB12_151: ld.f32 %f284, [%rd83]; bra.uni $L__BB12_175; $L__BB12_186: ld.f32 %f285, [%rd103]; bra.uni $L__BB12_210; $L__BB12_221: ld.f32 %f286, [%rd122]; bra.uni $L__BB12_245; $L__BB12_264: ld.f32 %f287, [%rd153]; bra.uni $L__BB12_277; $L__BB12_21: ld.u64 %rd9, [%rd2+40]; mov.u64 %rd700, 0; add.s64 %rd10, %rd9, 24; ld.u32 %r7, [%rd9+24]; setp.eq.s32 %p16, %r7, 0; @%p16 bra $L__BB12_26; add.s64 %rd11, %rd9, 64; shr.u32 %r119, %r1, 12; cvt.u64.u32 %rd198, %r119; shr.u32 %r120, %r3, 12; cvt.u64.u32 %rd199, %r120; bfi.b64 %rd200, %rd198, %rd199, 42, 22; shr.u32 %r121, %r2, 12; mul.wide.u32 %rd201, %r121, 2097152; or.b64 %rd12, %rd200, %rd201; mov.u32 %r873, 0; $L__BB12_23: cvt.u64.u32 %rd13, %r873; mul.wide.u32 %rd202, %r873, 32; add.s64 %rd203, %rd11, %rd202; ld.u64 %rd204, [%rd203]; setp.eq.s64 %p17, %rd204, %rd12; add.s32 %r873, %r873, 1; @%p17 bra $L__BB12_25; setp.lt.u32 %p18, %r873, %r7; @%p18 bra $L__BB12_23; bra.uni $L__BB12_26; $L__BB12_56: ld.u64 %rd29, [%rd22+40]; mov.u64 %rd701, 0; add.s64 %rd30, %rd29, 24; ld.u32 %r13, [%rd29+24]; setp.eq.s32 %p38, %r13, 0; @%p38 bra $L__BB12_61; add.s64 %rd31, %rd29, 64; shr.u32 %r223, %r1, 12; cvt.u64.u32 %rd265, %r223; shl.b64 %rd266, %rd265, 42; shr.u32 %r224, %r2, 12; mul.wide.u32 %rd267, %r224, 2097152; or.b64 %rd268, %rd267, %rd266; shr.u32 %r225, %r10, 12; cvt.u64.u32 %rd269, %r225; or.b64 %rd32, %rd268, %rd269; mov.u32 %r874, 0; $L__BB12_58: cvt.u64.u32 %rd33, %r874; mul.wide.u32 %rd270, %r874, 32; add.s64 %rd271, %rd31, %rd270; ld.u64 %rd272, [%rd271]; setp.eq.s64 %p39, %rd272, %rd32; add.s32 %r874, %r874, 1; @%p39 bra $L__BB12_60; setp.lt.u32 %p40, %r874, %r13; @%p40 bra $L__BB12_58; bra.uni $L__BB12_61; $L__BB12_91: ld.u64 %rd49, [%rd42+40]; mov.u64 %rd702, 0; add.s64 %rd50, %rd49, 24; ld.u32 %r19, [%rd49+24]; setp.eq.s32 %p60, %r19, 0; @%p60 bra $L__BB12_96; add.s64 %rd51, %rd49, 64; shr.u32 %r327, %r1, 12; cvt.u64.u32 %rd333, %r327; shr.u32 %r328, %r10, 12; cvt.u64.u32 %rd334, %r328; bfi.b64 %rd335, %rd333, %rd334, 42, 22; shr.u32 %r329, %r16, 12; mul.wide.u32 %rd336, %r329, 2097152; or.b64 %rd52, %rd335, %rd336; mov.u32 %r875, 0; $L__BB12_93: cvt.u64.u32 %rd53, %r875; mul.wide.u32 %rd337, %r875, 32; add.s64 %rd338, %rd51, %rd337; ld.u64 %rd339, [%rd338]; setp.eq.s64 %p61, %rd339, %rd52; add.s32 %r875, %r875, 1; @%p61 bra $L__BB12_95; setp.lt.u32 %p62, %r875, %r19; @%p62 bra $L__BB12_93; bra.uni $L__BB12_96; $L__BB12_126: ld.u64 %rd69, [%rd62+40]; mov.u64 %rd703, 0; ld.u32 %r24, [%rd69+24]; setp.eq.s32 %p82, %r24, 0; @%p82 bra $L__BB12_131; add.s64 %rd70, %rd69, 64; shr.u32 %r428, %r1, 12; cvt.u64.u32 %rd400, %r428; shr.u32 %r429, %r3, 12; cvt.u64.u32 %rd401, %r429; bfi.b64 %rd402, %rd400, %rd401, 42, 22; shr.u32 %r430, %r16, 12; mul.wide.u32 %rd403, %r430, 2097152; or.b64 %rd71, %rd402, %rd403; mov.u32 %r876, 0; $L__BB12_128: cvt.u64.u32 %rd72, %r876; mul.wide.u32 %rd404, %r876, 32; add.s64 %rd405, %rd70, %rd404; ld.u64 %rd406, [%rd405]; setp.eq.s64 %p83, %rd406, %rd71; add.s32 %r876, %r876, 1; @%p83 bra $L__BB12_130; setp.lt.u32 %p84, %r876, %r24; @%p84 bra $L__BB12_128; bra.uni $L__BB12_131; $L__BB12_161: ld.u64 %rd88, [%rd81+40]; mov.u64 %rd704, 0; add.s64 %rd89, %rd88, 24; ld.u32 %r31, [%rd88+24]; setp.eq.s32 %p104, %r31, 0; @%p104 bra $L__BB12_166; add.s64 %rd90, %rd88, 64; shr.u32 %r528, %r27, 12; cvt.u64.u32 %rd467, %r528; shr.u32 %r529, %r3, 12; cvt.u64.u32 %rd468, %r529; bfi.b64 %rd469, %rd467, %rd468, 42, 22; shr.u32 %r530, %r16, 12; mul.wide.u32 %rd470, %r530, 2097152; or.b64 %rd91, %rd469, %rd470; mov.u32 %r877, 0; $L__BB12_163: cvt.u64.u32 %rd92, %r877; mul.wide.u32 %rd471, %r877, 32; add.s64 %rd472, %rd90, %rd471; ld.u64 %rd473, [%rd472]; setp.eq.s64 %p105, %rd473, %rd91; add.s32 %r877, %r877, 1; @%p105 bra $L__BB12_165; setp.lt.u32 %p106, %r877, %r31; @%p106 bra $L__BB12_163; bra.uni $L__BB12_166; $L__BB12_196: ld.u64 %rd108, [%rd101+40]; mov.u64 %rd705, 0; ld.u32 %r36, [%rd108+24]; setp.eq.s32 %p126, %r36, 0; @%p126 bra $L__BB12_201; add.s64 %rd109, %rd108, 64; shr.u32 %r631, %r27, 12; cvt.u64.u32 %rd534, %r631; shr.u32 %r632, %r10, 12; cvt.u64.u32 %rd535, %r632; bfi.b64 %rd536, %rd534, %rd535, 42, 22; shr.u32 %r633, %r16, 12; mul.wide.u32 %rd537, %r633, 2097152; or.b64 %rd110, %rd536, %rd537; mov.u32 %r878, 0; $L__BB12_198: cvt.u64.u32 %rd111, %r878; mul.wide.u32 %rd538, %r878, 32; add.s64 %rd539, %rd109, %rd538; ld.u64 %rd540, [%rd539]; setp.eq.s64 %p127, %rd540, %rd110; add.s32 %r878, %r878, 1; @%p127 bra $L__BB12_200; setp.lt.u32 %p128, %r878, %r36; @%p128 bra $L__BB12_198; bra.uni $L__BB12_201; $L__BB12_231: ld.u64 %rd127, [%rd120+40]; mov.u64 %rd706, 0; add.s64 %rd128, %rd127, 24; ld.u32 %r41, [%rd127+24]; setp.eq.s32 %p148, %r41, 0; @%p148 bra $L__BB12_236; add.s64 %rd129, %rd127, 64; shr.u32 %r735, %r27, 12; cvt.u64.u32 %rd601, %r735; shl.b64 %rd602, %rd601, 42; shr.u32 %r736, %r2, 12; mul.wide.u32 %rd603, %r736, 2097152; or.b64 %rd604, %rd602, %rd603; shr.u32 %r737, %r10, 12; cvt.u64.u32 %rd605, %r737; or.b64 %rd130, %rd604, %rd605; mov.u32 %r879, 0; $L__BB12_233: cvt.u64.u32 %rd131, %r879; mul.wide.u32 %rd606, %r879, 32; add.s64 %rd607, %rd129, %rd606; ld.u64 %rd608, [%rd607]; setp.eq.s64 %p149, %rd608, %rd130; add.s32 %r879, %r879, 1; @%p149 bra $L__BB12_235; setp.lt.u32 %p150, %r879, %r41; @%p150 bra $L__BB12_233; bra.uni $L__BB12_236; $L__BB12_275: ld.f32 %f287, [%rd159]; bra.uni $L__BB12_277; $L__BB12_272: ld.f32 %f287, [%rd155]; bra.uni $L__BB12_277; $L__BB12_20: ld.f32 %f280, [%rd6]; bra.uni $L__BB12_35; $L__BB12_55: ld.f32 %f281, [%rd26]; bra.uni $L__BB12_70; $L__BB12_90: ld.f32 %f282, [%rd46]; bra.uni $L__BB12_105; $L__BB12_125: ld.f32 %f283, [%rd66]; bra.uni $L__BB12_140; $L__BB12_160: ld.f32 %f284, [%rd85]; bra.uni $L__BB12_175; $L__BB12_195: ld.f32 %f285, [%rd105]; bra.uni $L__BB12_210; $L__BB12_230: ld.f32 %f286, [%rd124]; bra.uni $L__BB12_245; $L__BB12_271: ld.f32 %f287, [%rd157]; $L__BB12_277: sub.ftz.f32 %f274, %f286, %f287; fma.rn.ftz.f32 %f275, %f3, %f274, %f287; sub.ftz.f32 %f276, %f78, %f275; fma.rn.ftz.f32 %f277, %f2, %f276, %f275; sub.ftz.f32 %f278, %f277, %f53; fma.rn.ftz.f32 %f279, %f1, %f278, %f53; st.param.f32 [func_retval0+0], %f279; ret; $L__BB12_19: ld.f32 %f280, [%rd8]; bra.uni $L__BB12_35; $L__BB12_54: ld.f32 %f281, [%rd28]; bra.uni $L__BB12_70; $L__BB12_89: ld.f32 %f282, [%rd48]; bra.uni $L__BB12_105; $L__BB12_124: ld.f32 %f283, [%rd68]; bra.uni $L__BB12_140; $L__BB12_159: ld.f32 %f284, [%rd87]; bra.uni $L__BB12_175; $L__BB12_194: ld.f32 %f285, [%rd107]; bra.uni $L__BB12_210; $L__BB12_229: ld.f32 %f286, [%rd126]; bra.uni $L__BB12_245; $L__BB12_25: shl.b64 %rd206, %rd13, 5; add.s64 %rd700, %rd11, %rd206; $L__BB12_26: setp.eq.s64 %p19, %rd700, 0; @%p19 bra $L__BB12_34; add.s64 %rd16, %rd700, 8; ld.u64 %rd17, [%rd700+8]; setp.eq.s64 %p20, %rd17, 0; @%p20 bra $L__BB12_33; add.s64 %rd18, %rd9, %rd17; and.b32 %r122, %r2, -4096; mov.u64 %rd207, 1; st.v2.u32 [%rd2+24], {%r6, %r122}; and.b32 %r123, %r3, -4096; st.u32 [%rd2+32], %r123; st.u64 [%rd2+64], %rd18; shl.b32 %r124, %r1, 3; and.b32 %r125, %r124, 31744; and.b32 %r126, %r2, 3968; shr.u32 %r127, %r126, 2; or.b32 %r128, %r127, %r125; bfe.u32 %r129, %r3, 7, 5; or.b32 %r130, %r128, %r129; shr.u32 %r131, %r128, 6; mul.wide.u32 %rd208, %r131, 8; add.s64 %rd209, %rd18, %rd208; and.b32 %r132, %r130, 63; shl.b64 %rd210, %rd207, %r132; ld.u64 %rd211, [%rd209+4128]; and.b64 %rd212, %rd211, %rd210; setp.eq.s64 %p21, %rd212, 0; mul.wide.u32 %rd213, %r130, 8; add.s64 %rd214, %rd18, %rd213; add.s64 %rd19, %rd214, 8256; @%p21 bra $L__BB12_32; ld.u64 %rd215, [%rd19]; add.s64 %rd20, %rd18, %rd215; st.u32 [%rd2+12], %r5; and.b32 %r133, %r2, -128; st.u32 [%rd2+16], %r133; and.b32 %r134, %r3, -128; st.u32 [%rd2+20], %r134; st.u64 [%rd2+56], %rd20; shl.b32 %r135, %r1, 5; and.b32 %r136, %r135, 3840; shl.b32 %r137, %r2, 1; and.b32 %r138, %r137, 240; or.b32 %r139, %r138, %r136; bfe.u32 %r140, %r3, 3, 4; or.b32 %r141, %r139, %r140; shr.u32 %r142, %r139, 6; mul.wide.u32 %rd216, %r142, 8; add.s64 %rd217, %rd20, %rd216; and.b32 %r143, %r141, 63; shl.b64 %rd219, %rd207, %r143; ld.u64 %rd220, [%rd217+544]; and.b64 %rd221, %rd220, %rd219; setp.eq.s64 %p22, %rd221, 0; mul.wide.u32 %rd222, %r141, 8; add.s64 %rd223, %rd20, %rd222; add.s64 %rd21, %rd223, 1088; @%p22 bra $L__BB12_31; ld.u64 %rd224, [%rd21]; add.s64 %rd225, %rd20, %rd224; and.b32 %r144, %r2, -8; st.v2.u32 [%rd2], {%r4, %r144}; and.b32 %r145, %r3, -8; st.u32 [%rd2+8], %r145; st.u64 [%rd2+48], %rd225; shl.b32 %r146, %r1, 6; and.b32 %r147, %r146, 448; shl.b32 %r148, %r2, 3; and.b32 %r149, %r148, 56; or.b32 %r150, %r149, %r147; and.b32 %r151, %r3, 7; or.b32 %r152, %r150, %r151; mul.wide.u32 %rd226, %r152, 2; add.s64 %rd227, %rd225, %rd226; ld.u16 %rs52, [%rd227+96]; cvt.rn.f32.u16 %f124, %rs52; ld.v2.f32 {%f125, %f126}, [%rd225+80]; fma.rn.ftz.f32 %f280, %f126, %f124, %f125; bra.uni $L__BB12_35; $L__BB12_34: ld.f32 %f280, [%rd10+4]; bra.uni $L__BB12_35; $L__BB12_60: shl.b64 %rd274, %rd33, 5; add.s64 %rd701, %rd31, %rd274; $L__BB12_61: setp.eq.s64 %p41, %rd701, 0; @%p41 bra $L__BB12_69; add.s64 %rd36, %rd701, 8; ld.u64 %rd37, [%rd701+8]; setp.eq.s64 %p42, %rd37, 0; @%p42 bra $L__BB12_68; add.s64 %rd38, %rd29, %rd37; and.b32 %r226, %r2, -4096; mov.u64 %rd275, 1; st.v2.u32 [%rd22+24], {%r12, %r226}; and.b32 %r227, %r10, -4096; st.u32 [%rd22+32], %r227; st.u64 [%rd22+64], %rd38; shl.b32 %r228, %r1, 3; and.b32 %r229, %r228, 31744; and.b32 %r230, %r2, 3968; shr.u32 %r231, %r230, 2; or.b32 %r232, %r231, %r229; bfe.u32 %r233, %r10, 7, 5; or.b32 %r234, %r233, %r232; shr.u32 %r235, %r232, 6; mul.wide.u32 %rd276, %r235, 8; add.s64 %rd277, %rd38, %rd276; and.b32 %r236, %r234, 63; shl.b64 %rd278, %rd275, %r236; ld.u64 %rd279, [%rd277+4128]; and.b64 %rd280, %rd279, %rd278; setp.eq.s64 %p43, %rd280, 0; mul.wide.u32 %rd281, %r234, 8; add.s64 %rd282, %rd38, %rd281; add.s64 %rd39, %rd282, 8256; @%p43 bra $L__BB12_67; ld.u64 %rd283, [%rd39]; add.s64 %rd40, %rd38, %rd283; st.u32 [%rd22+12], %r11; and.b32 %r237, %r2, -128; st.u32 [%rd22+16], %r237; and.b32 %r238, %r10, -128; st.u32 [%rd22+20], %r238; st.u64 [%rd22+56], %rd40; shl.b32 %r239, %r1, 5; and.b32 %r240, %r239, 3840; shl.b32 %r241, %r2, 1; and.b32 %r242, %r241, 240; or.b32 %r243, %r242, %r240; bfe.u32 %r244, %r10, 3, 4; or.b32 %r245, %r244, %r243; shr.u32 %r246, %r243, 6; mul.wide.u32 %rd284, %r246, 8; add.s64 %rd285, %rd40, %rd284; and.b32 %r247, %r245, 63; shl.b64 %rd287, %rd275, %r247; ld.u64 %rd288, [%rd285+544]; and.b64 %rd289, %rd288, %rd287; setp.eq.s64 %p44, %rd289, 0; mul.wide.u32 %rd290, %r245, 8; add.s64 %rd291, %rd40, %rd290; add.s64 %rd41, %rd291, 1088; @%p44 bra $L__BB12_66; ld.u64 %rd292, [%rd41]; add.s64 %rd293, %rd40, %rd292; and.b32 %r248, %r2, -8; st.v2.u32 [%rd22], {%r4, %r248}; and.b32 %r249, %r10, -8; st.u32 [%rd22+8], %r249; st.u64 [%rd22+48], %rd293; shl.b32 %r250, %r1, 6; and.b32 %r251, %r250, 448; shl.b32 %r252, %r2, 3; and.b32 %r253, %r252, 56; or.b32 %r254, %r253, %r251; and.b32 %r255, %r10, 7; or.b32 %r256, %r254, %r255; mul.wide.u32 %rd294, %r256, 2; add.s64 %rd295, %rd293, %rd294; ld.u16 %rs62, [%rd295+96]; cvt.rn.f32.u16 %f144, %rs62; ld.v2.f32 {%f145, %f146}, [%rd293+80]; fma.rn.ftz.f32 %f281, %f146, %f144, %f145; bra.uni $L__BB12_70; $L__BB12_69: ld.f32 %f281, [%rd30+4]; bra.uni $L__BB12_70; $L__BB12_95: shl.b64 %rd341, %rd53, 5; add.s64 %rd702, %rd51, %rd341; $L__BB12_96: setp.eq.s64 %p63, %rd702, 0; @%p63 bra $L__BB12_104; add.s64 %rd56, %rd702, 8; ld.u64 %rd57, [%rd702+8]; setp.eq.s64 %p64, %rd57, 0; @%p64 bra $L__BB12_103; add.s64 %rd58, %rd49, %rd57; and.b32 %r330, %r16, -4096; mov.u64 %rd342, 1; st.v2.u32 [%rd42+24], {%r18, %r330}; and.b32 %r331, %r10, -4096; st.u32 [%rd42+32], %r331; st.u64 [%rd42+64], %rd58; shl.b32 %r332, %r1, 3; and.b32 %r333, %r332, 31744; and.b32 %r334, %r16, 3968; shr.u32 %r335, %r334, 2; or.b32 %r336, %r335, %r333; bfe.u32 %r337, %r10, 7, 5; or.b32 %r338, %r336, %r337; shr.u32 %r339, %r336, 6; mul.wide.u32 %rd343, %r339, 8; add.s64 %rd344, %rd58, %rd343; and.b32 %r340, %r338, 63; shl.b64 %rd345, %rd342, %r340; ld.u64 %rd346, [%rd344+4128]; and.b64 %rd347, %rd346, %rd345; setp.eq.s64 %p65, %rd347, 0; mul.wide.u32 %rd348, %r338, 8; add.s64 %rd349, %rd58, %rd348; add.s64 %rd59, %rd349, 8256; @%p65 bra $L__BB12_102; ld.u64 %rd350, [%rd59]; add.s64 %rd60, %rd58, %rd350; st.u32 [%rd42+12], %r17; and.b32 %r341, %r16, -128; st.u32 [%rd42+16], %r341; and.b32 %r342, %r10, -128; st.u32 [%rd42+20], %r342; st.u64 [%rd42+56], %rd60; shl.b32 %r343, %r1, 5; and.b32 %r344, %r343, 3840; shl.b32 %r345, %r16, 1; and.b32 %r346, %r345, 240; or.b32 %r347, %r346, %r344; bfe.u32 %r348, %r10, 3, 4; or.b32 %r349, %r347, %r348; shr.u32 %r350, %r347, 6; mul.wide.u32 %rd351, %r350, 8; add.s64 %rd352, %rd60, %rd351; and.b32 %r351, %r349, 63; shl.b64 %rd354, %rd342, %r351; ld.u64 %rd355, [%rd352+544]; and.b64 %rd356, %rd355, %rd354; setp.eq.s64 %p66, %rd356, 0; mul.wide.u32 %rd357, %r349, 8; add.s64 %rd358, %rd60, %rd357; add.s64 %rd61, %rd358, 1088; @%p66 bra $L__BB12_101; ld.u64 %rd359, [%rd61]; add.s64 %rd360, %rd60, %rd359; and.b32 %r352, %r16, -8; st.v2.u32 [%rd42], {%r4, %r352}; and.b32 %r353, %r10, -8; st.u32 [%rd42+8], %r353; st.u64 [%rd42+48], %rd360; shl.b32 %r354, %r1, 6; and.b32 %r355, %r354, 448; shl.b32 %r356, %r16, 3; and.b32 %r357, %r356, 56; or.b32 %r358, %r357, %r355; and.b32 %r359, %r10, 7; or.b32 %r360, %r358, %r359; mul.wide.u32 %rd361, %r360, 2; add.s64 %rd362, %rd360, %rd361; ld.u16 %rs72, [%rd362+96]; cvt.rn.f32.u16 %f165, %rs72; ld.v2.f32 {%f166, %f167}, [%rd360+80]; fma.rn.ftz.f32 %f282, %f167, %f165, %f166; bra.uni $L__BB12_105; $L__BB12_104: ld.f32 %f282, [%rd50+4]; bra.uni $L__BB12_105; $L__BB12_130: shl.b64 %rd408, %rd72, 5; add.s64 %rd703, %rd70, %rd408; $L__BB12_131: setp.eq.s64 %p85, %rd703, 0; @%p85 bra $L__BB12_139; add.s64 %rd75, %rd703, 8; ld.u64 %rd76, [%rd703+8]; setp.eq.s64 %p86, %rd76, 0; @%p86 bra $L__BB12_138; add.s64 %rd77, %rd69, %rd76; and.b32 %r431, %r16, -4096; mov.u64 %rd409, 1; st.v2.u32 [%rd62+24], {%r23, %r431}; and.b32 %r432, %r3, -4096; st.u32 [%rd62+32], %r432; st.u64 [%rd62+64], %rd77; shl.b32 %r433, %r1, 3; and.b32 %r434, %r433, 31744; and.b32 %r435, %r16, 3968; shr.u32 %r436, %r435, 2; or.b32 %r437, %r436, %r434; bfe.u32 %r438, %r3, 7, 5; or.b32 %r439, %r437, %r438; shr.u32 %r440, %r437, 6; mul.wide.u32 %rd410, %r440, 8; add.s64 %rd411, %rd77, %rd410; and.b32 %r441, %r439, 63; shl.b64 %rd412, %rd409, %r441; ld.u64 %rd413, [%rd411+4128]; and.b64 %rd414, %rd413, %rd412; setp.eq.s64 %p87, %rd414, 0; mul.wide.u32 %rd415, %r439, 8; add.s64 %rd416, %rd77, %rd415; add.s64 %rd78, %rd416, 8256; @%p87 bra $L__BB12_137; ld.u64 %rd417, [%rd78]; add.s64 %rd79, %rd77, %rd417; st.u32 [%rd62+12], %r22; and.b32 %r442, %r16, -128; st.u32 [%rd62+16], %r442; and.b32 %r443, %r3, -128; st.u32 [%rd62+20], %r443; st.u64 [%rd62+56], %rd79; shl.b32 %r444, %r1, 5; and.b32 %r445, %r444, 3840; shl.b32 %r446, %r16, 1; and.b32 %r447, %r446, 240; or.b32 %r448, %r447, %r445; bfe.u32 %r449, %r3, 3, 4; or.b32 %r450, %r448, %r449; shr.u32 %r451, %r448, 6; mul.wide.u32 %rd418, %r451, 8; add.s64 %rd419, %rd79, %rd418; and.b32 %r452, %r450, 63; shl.b64 %rd421, %rd409, %r452; ld.u64 %rd422, [%rd419+544]; and.b64 %rd423, %rd422, %rd421; setp.eq.s64 %p88, %rd423, 0; mul.wide.u32 %rd424, %r450, 8; add.s64 %rd425, %rd79, %rd424; add.s64 %rd80, %rd425, 1088; @%p88 bra $L__BB12_136; ld.u64 %rd426, [%rd80]; add.s64 %rd427, %rd79, %rd426; and.b32 %r453, %r16, -8; st.v2.u32 [%rd62], {%r4, %r453}; and.b32 %r454, %r3, -8; st.u32 [%rd62+8], %r454; st.u64 [%rd62+48], %rd427; and.b32 %r455, %r1, 7; shl.b32 %r456, %r16, 3; and.b32 %r457, %r456, 56; and.b32 %r458, %r3, 7; bfi.b32 %r459, %r455, %r458, 6, 3; or.b32 %r460, %r459, %r457; mul.wide.u32 %rd428, %r460, 2; add.s64 %rd429, %rd427, %rd428; ld.u16 %rs82, [%rd429+96]; cvt.rn.f32.u16 %f185, %rs82; ld.v2.f32 {%f186, %f187}, [%rd427+80]; fma.rn.ftz.f32 %f283, %f187, %f185, %f186; bra.uni $L__BB12_140; $L__BB12_139: ld.f32 %f283, [%rd69+28]; bra.uni $L__BB12_140; $L__BB12_165: shl.b64 %rd475, %rd92, 5; add.s64 %rd704, %rd90, %rd475; $L__BB12_166: setp.eq.s64 %p107, %rd704, 0; @%p107 bra $L__BB12_174; add.s64 %rd95, %rd704, 8; ld.u64 %rd96, [%rd704+8]; setp.eq.s64 %p108, %rd96, 0; @%p108 bra $L__BB12_173; add.s64 %rd97, %rd88, %rd96; and.b32 %r531, %r16, -4096; mov.u64 %rd476, 1; st.v2.u32 [%rd81+24], {%r30, %r531}; and.b32 %r532, %r3, -4096; st.u32 [%rd81+32], %r532; st.u64 [%rd81+64], %rd97; shl.b32 %r533, %r27, 3; and.b32 %r534, %r533, 31744; and.b32 %r535, %r16, 3968; shr.u32 %r536, %r535, 2; or.b32 %r537, %r536, %r534; bfe.u32 %r538, %r3, 7, 5; or.b32 %r539, %r537, %r538; shr.u32 %r540, %r537, 6; mul.wide.u32 %rd477, %r540, 8; add.s64 %rd478, %rd97, %rd477; and.b32 %r541, %r539, 63; shl.b64 %rd479, %rd476, %r541; ld.u64 %rd480, [%rd478+4128]; and.b64 %rd481, %rd480, %rd479; setp.eq.s64 %p109, %rd481, 0; mul.wide.u32 %rd482, %r539, 8; add.s64 %rd483, %rd97, %rd482; add.s64 %rd98, %rd483, 8256; @%p109 bra $L__BB12_172; ld.u64 %rd484, [%rd98]; add.s64 %rd99, %rd97, %rd484; st.u32 [%rd81+12], %r29; and.b32 %r542, %r16, -128; st.u32 [%rd81+16], %r542; and.b32 %r543, %r3, -128; st.u32 [%rd81+20], %r543; st.u64 [%rd81+56], %rd99; shl.b32 %r544, %r27, 5; and.b32 %r545, %r544, 3840; shl.b32 %r546, %r16, 1; and.b32 %r547, %r546, 240; or.b32 %r548, %r547, %r545; bfe.u32 %r549, %r3, 3, 4; or.b32 %r550, %r548, %r549; shr.u32 %r551, %r548, 6; mul.wide.u32 %rd485, %r551, 8; add.s64 %rd486, %rd99, %rd485; and.b32 %r552, %r550, 63; shl.b64 %rd488, %rd476, %r552; ld.u64 %rd489, [%rd486+544]; and.b64 %rd490, %rd489, %rd488; setp.eq.s64 %p110, %rd490, 0; mul.wide.u32 %rd491, %r550, 8; add.s64 %rd492, %rd99, %rd491; add.s64 %rd100, %rd492, 1088; @%p110 bra $L__BB12_171; ld.u64 %rd493, [%rd100]; add.s64 %rd494, %rd99, %rd493; and.b32 %r553, %r16, -8; st.v2.u32 [%rd81], {%r28, %r553}; and.b32 %r554, %r3, -8; st.u32 [%rd81+8], %r554; st.u64 [%rd81+48], %rd494; and.b32 %r555, %r27, 7; shl.b32 %r556, %r16, 3; and.b32 %r557, %r556, 56; and.b32 %r558, %r3, 7; bfi.b32 %r559, %r555, %r558, 6, 3; or.b32 %r560, %r559, %r557; mul.wide.u32 %rd495, %r560, 2; add.s64 %rd496, %rd494, %rd495; ld.u16 %rs92, [%rd496+96]; cvt.rn.f32.u16 %f208, %rs92; ld.v2.f32 {%f209, %f210}, [%rd494+80]; fma.rn.ftz.f32 %f284, %f210, %f208, %f209; bra.uni $L__BB12_175; $L__BB12_174: ld.f32 %f284, [%rd89+4]; bra.uni $L__BB12_175; $L__BB12_200: shl.b64 %rd542, %rd111, 5; add.s64 %rd705, %rd109, %rd542; $L__BB12_201: setp.eq.s64 %p129, %rd705, 0; @%p129 bra $L__BB12_209; add.s64 %rd114, %rd705, 8; ld.u64 %rd115, [%rd705+8]; setp.eq.s64 %p130, %rd115, 0; @%p130 bra $L__BB12_208; add.s64 %rd116, %rd108, %rd115; and.b32 %r634, %r16, -4096; mov.u64 %rd543, 1; st.v2.u32 [%rd101+24], {%r35, %r634}; and.b32 %r635, %r10, -4096; st.u32 [%rd101+32], %r635; st.u64 [%rd101+64], %rd116; shl.b32 %r636, %r27, 3; and.b32 %r637, %r636, 31744; and.b32 %r638, %r16, 3968; shr.u32 %r639, %r638, 2; or.b32 %r640, %r639, %r637; bfe.u32 %r641, %r10, 7, 5; or.b32 %r642, %r640, %r641; shr.u32 %r643, %r640, 6; mul.wide.u32 %rd544, %r643, 8; add.s64 %rd545, %rd116, %rd544; and.b32 %r644, %r642, 63; shl.b64 %rd546, %rd543, %r644; ld.u64 %rd547, [%rd545+4128]; and.b64 %rd548, %rd547, %rd546; setp.eq.s64 %p131, %rd548, 0; mul.wide.u32 %rd549, %r642, 8; add.s64 %rd550, %rd116, %rd549; add.s64 %rd117, %rd550, 8256; @%p131 bra $L__BB12_207; ld.u64 %rd551, [%rd117]; add.s64 %rd118, %rd116, %rd551; st.u32 [%rd101+12], %r34; and.b32 %r645, %r16, -128; st.u32 [%rd101+16], %r645; and.b32 %r646, %r10, -128; st.u32 [%rd101+20], %r646; st.u64 [%rd101+56], %rd118; shl.b32 %r647, %r27, 5; and.b32 %r648, %r647, 3840; shl.b32 %r649, %r16, 1; and.b32 %r650, %r649, 240; or.b32 %r651, %r650, %r648; bfe.u32 %r652, %r10, 3, 4; or.b32 %r653, %r651, %r652; shr.u32 %r654, %r651, 6; mul.wide.u32 %rd552, %r654, 8; add.s64 %rd553, %rd118, %rd552; and.b32 %r655, %r653, 63; shl.b64 %rd555, %rd543, %r655; ld.u64 %rd556, [%rd553+544]; and.b64 %rd557, %rd556, %rd555; setp.eq.s64 %p132, %rd557, 0; mul.wide.u32 %rd558, %r653, 8; add.s64 %rd559, %rd118, %rd558; add.s64 %rd119, %rd559, 1088; @%p132 bra $L__BB12_206; ld.u64 %rd560, [%rd119]; add.s64 %rd561, %rd118, %rd560; and.b32 %r656, %r16, -8; st.v2.u32 [%rd101], {%r28, %r656}; and.b32 %r657, %r10, -8; st.u32 [%rd101+8], %r657; st.u64 [%rd101+48], %rd561; shl.b32 %r658, %r27, 6; and.b32 %r659, %r658, 448; shl.b32 %r660, %r16, 3; and.b32 %r661, %r660, 56; or.b32 %r662, %r661, %r659; and.b32 %r663, %r10, 7; or.b32 %r664, %r662, %r663; mul.wide.u32 %rd562, %r664, 2; add.s64 %rd563, %rd561, %rd562; ld.u16 %rs102, [%rd563+96]; cvt.rn.f32.u16 %f228, %rs102; ld.v2.f32 {%f229, %f230}, [%rd561+80]; fma.rn.ftz.f32 %f285, %f230, %f228, %f229; bra.uni $L__BB12_210; $L__BB12_209: ld.f32 %f285, [%rd108+28]; bra.uni $L__BB12_210; $L__BB12_235: shl.b64 %rd610, %rd131, 5; add.s64 %rd706, %rd129, %rd610; $L__BB12_236: setp.eq.s64 %p151, %rd706, 0; @%p151 bra $L__BB12_244; add.s64 %rd134, %rd706, 8; ld.u64 %rd135, [%rd706+8]; setp.eq.s64 %p152, %rd135, 0; @%p152 bra $L__BB12_243; add.s64 %rd136, %rd127, %rd135; and.b32 %r738, %r2, -4096; mov.u64 %rd611, 1; st.v2.u32 [%rd120+24], {%r40, %r738}; and.b32 %r739, %r10, -4096; st.u32 [%rd120+32], %r739; st.u64 [%rd120+64], %rd136; shl.b32 %r740, %r27, 3; and.b32 %r741, %r740, 31744; and.b32 %r742, %r2, 3968; shr.u32 %r743, %r742, 2; or.b32 %r744, %r741, %r743; bfe.u32 %r745, %r10, 7, 5; or.b32 %r746, %r745, %r744; shr.u32 %r747, %r744, 6; mul.wide.u32 %rd612, %r747, 8; add.s64 %rd613, %rd136, %rd612; and.b32 %r748, %r746, 63; shl.b64 %rd614, %rd611, %r748; ld.u64 %rd615, [%rd613+4128]; and.b64 %rd616, %rd615, %rd614; setp.eq.s64 %p153, %rd616, 0; mul.wide.u32 %rd617, %r746, 8; add.s64 %rd618, %rd136, %rd617; add.s64 %rd137, %rd618, 8256; @%p153 bra $L__BB12_242; ld.u64 %rd619, [%rd137]; add.s64 %rd138, %rd136, %rd619; st.u32 [%rd120+12], %r39; and.b32 %r749, %r2, -128; st.u32 [%rd120+16], %r749; and.b32 %r750, %r10, -128; st.u32 [%rd120+20], %r750; st.u64 [%rd120+56], %rd138; shl.b32 %r751, %r27, 5; and.b32 %r752, %r751, 3840; shl.b32 %r753, %r2, 1; and.b32 %r754, %r753, 240; or.b32 %r755, %r752, %r754; bfe.u32 %r756, %r10, 3, 4; or.b32 %r757, %r756, %r755; shr.u32 %r758, %r755, 6; mul.wide.u32 %rd620, %r758, 8; add.s64 %rd621, %rd138, %rd620; and.b32 %r759, %r757, 63; shl.b64 %rd623, %rd611, %r759; ld.u64 %rd624, [%rd621+544]; and.b64 %rd625, %rd624, %rd623; setp.eq.s64 %p154, %rd625, 0; mul.wide.u32 %rd626, %r757, 8; add.s64 %rd627, %rd138, %rd626; add.s64 %rd139, %rd627, 1088; @%p154 bra $L__BB12_241; ld.u64 %rd628, [%rd139]; add.s64 %rd629, %rd138, %rd628; and.b32 %r760, %r2, -8; st.v2.u32 [%rd120], {%r28, %r760}; and.b32 %r761, %r10, -8; st.u32 [%rd120+8], %r761; st.u64 [%rd120+48], %rd629; shl.b32 %r762, %r27, 6; and.b32 %r763, %r762, 448; shl.b32 %r764, %r2, 3; and.b32 %r765, %r764, 56; or.b32 %r766, %r763, %r765; and.b32 %r767, %r10, 7; or.b32 %r768, %r766, %r767; mul.wide.u32 %rd630, %r768, 2; add.s64 %rd631, %rd629, %rd630; ld.u16 %rs112, [%rd631+96]; cvt.rn.f32.u16 %f249, %rs112; ld.v2.f32 {%f250, %f251}, [%rd629+80]; fma.rn.ftz.f32 %f286, %f251, %f249, %f250; bra.uni $L__BB12_245; $L__BB12_244: ld.f32 %f286, [%rd128+4]; bra.uni $L__BB12_245; $L__BB12_33: ld.f32 %f280, [%rd16+12]; bra.uni $L__BB12_35; $L__BB12_68: ld.f32 %f281, [%rd36+12]; bra.uni $L__BB12_70; $L__BB12_103: ld.f32 %f282, [%rd56+12]; bra.uni $L__BB12_105; $L__BB12_138: ld.f32 %f283, [%rd75+12]; bra.uni $L__BB12_140; $L__BB12_173: ld.f32 %f284, [%rd95+12]; bra.uni $L__BB12_175; $L__BB12_208: ld.f32 %f285, [%rd114+12]; bra.uni $L__BB12_210; $L__BB12_243: ld.f32 %f286, [%rd134+12]; bra.uni $L__BB12_245; $L__BB12_32: ld.f32 %f280, [%rd19]; bra.uni $L__BB12_35; $L__BB12_67: ld.f32 %f281, [%rd39]; bra.uni $L__BB12_70; $L__BB12_102: ld.f32 %f282, [%rd59]; bra.uni $L__BB12_105; $L__BB12_137: ld.f32 %f283, [%rd78]; bra.uni $L__BB12_140; $L__BB12_172: ld.f32 %f284, [%rd98]; bra.uni $L__BB12_175; $L__BB12_207: ld.f32 %f285, [%rd117]; bra.uni $L__BB12_210; $L__BB12_242: ld.f32 %f286, [%rd137]; bra.uni $L__BB12_245; $L__BB12_31: ld.f32 %f280, [%rd21]; bra.uni $L__BB12_35; $L__BB12_66: ld.f32 %f281, [%rd41]; bra.uni $L__BB12_70; $L__BB12_101: ld.f32 %f282, [%rd61]; bra.uni $L__BB12_105; $L__BB12_136: ld.f32 %f283, [%rd80]; bra.uni $L__BB12_140; $L__BB12_171: ld.f32 %f284, [%rd100]; bra.uni $L__BB12_175; $L__BB12_206: ld.f32 %f285, [%rd119]; bra.uni $L__BB12_210; $L__BB12_241: ld.f32 %f286, [%rd139]; bra.uni $L__BB12_245; } .func (.param .b32 func_retval0) _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_( .param .b64 _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_0, .param .align 4 .b8 _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_1[16], .param .b64 _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_2, .param .b64 _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_3 ) { .reg .pred %p<52>; .reg .b16 %rs<2>; .reg .f32 %f<453>; .reg .b32 %r<226>; .reg .b64 %rd<145>; ld.param.u64 %rd8, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_0]; ld.param.u64 %rd6, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_2]; ld.param.u64 %rd7, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_3]; ld.param.u32 %r52, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_1+12]; ld.param.u32 %r51, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_1+8]; ld.param.u32 %r49, [_Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6__param_1]; cvta.to.local.u64 %rd9, %rd8; add.s64 %rd1, %rd9, 60; ld.local.s32 %rd3, [%rd9+60]; setp.eq.s64 %p1, %rd3, -1; mov.u32 %r3, -1; @%p1 bra $L__BB13_2; ld.const.u64 %rd10, [kernel_params+2240]; cvta.to.global.u64 %rd11, %rd10; shl.b64 %rd12, %rd3, 2; add.s64 %rd13, %rd11, %rd12; ld.global.u32 %r3, [%rd13]; $L__BB13_2: and.b32 %r54, %r51, 2; setp.eq.s32 %p2, %r54, 0; @%p2 bra $L__BB13_38; bra.uni $L__BB13_3; $L__BB13_38: setp.eq.s32 %p34, %r49, 4; @%p34 bra $L__BB13_65; bra.uni $L__BB13_39; $L__BB13_65: setp.eq.s64 %p50, %rd6, 0; @%p50 bra $L__BB13_67; mov.u32 %r205, 0; st.u32 [%rd6], %r205; $L__BB13_67: setp.eq.s64 %p51, %rd7, 0; @%p51 bra $L__BB13_69; mov.u32 %r206, 0; st.u32 [%rd7], %r206; $L__BB13_69: ld.const.u64 %rd137, [kernel_params+2320]; cvta.to.global.u64 %rd138, %rd137; ld.const.u64 %rd139, [kernel_params+2288]; cvta.to.global.u64 %rd140, %rd139; add.s32 %r207, %r3, 4; mul.wide.s32 %rd141, %r207, 4; add.s64 %rd142, %rd140, %rd141; ld.global.u32 %r208, [%rd142]; add.s32 %r209, %r208, %r52; mul.wide.u32 %rd143, %r209, 4; add.s64 %rd144, %rd138, %rd143; ld.global.f32 %f452, [%rd144]; bra.uni $L__BB13_70; $L__BB13_3: ld.const.u64 %rd14, [kernel_params+2232]; cvta.to.global.u64 %rd15, %rd14; shl.b64 %rd16, %rd3, 4; add.s64 %rd17, %rd15, %rd16; ld.global.v4.u32 {%r56, %r57, %r58, %r59}, [%rd17]; mov.u32 %r212, 0; ld.const.u64 %rd18, [kernel_params+2248]; cvta.to.global.u64 %rd19, %rd18; mul.wide.u32 %rd20, %r56, 8; add.s64 %rd21, %rd19, %rd20; ld.global.v2.f32 {%f54, %f55}, [%rd21]; mul.wide.u32 %rd22, %r57, 8; add.s64 %rd23, %rd19, %rd22; ld.global.v2.f32 {%f58, %f59}, [%rd23]; mul.wide.u32 %rd24, %r58, 8; add.s64 %rd25, %rd19, %rd24; ld.global.v2.f32 {%f62, %f63}, [%rd25]; sub.ftz.f32 %f1, %f58, %f54; sub.ftz.f32 %f2, %f59, %f55; sub.ftz.f32 %f3, %f62, %f54; sub.ftz.f32 %f4, %f63, %f55; ld.local.f32 %f66, [%rd1+8]; ld.local.v2.u32 {%r63, %r64}, [%rd1+12]; mov.b32 %f67, %r63; mul.ftz.f32 %f68, %f3, %f67; mul.ftz.f32 %f69, %f4, %f67; fma.rn.ftz.f32 %f70, %f1, %f66, %f68; fma.rn.ftz.f32 %f71, %f2, %f66, %f69; add.ftz.f32 %f5, %f54, %f70; add.ftz.f32 %f6, %f55, %f71; setp.eq.s32 %p3, %r64, -1; mov.u32 %r211, %r212; @%p3 bra $L__BB13_5; ld.const.u64 %rd26, [kernel_params+2160]; cvta.to.global.u64 %rd27, %rd26; mul.wide.s32 %rd28, %r64, 208; add.s64 %rd29, %rd27, %rd28; ld.global.u32 %r211, [%rd29+160]; $L__BB13_5: add.s32 %r67, %r211, %r3; shr.u32 %r218, %r67, 1; ld.const.u64 %rd30, [kernel_params+2288]; cvta.to.global.u64 %rd4, %rd30; mov.f32 %f441, 0f3F000000; mov.f32 %f445, %f5; mov.f32 %f444, %f6; bra.uni $L__BB13_6; $L__BB13_31: add.s32 %r212, %r212, 4; mul.ftz.f32 %f441, %f22, 0f3F000000; $L__BB13_6: mul.ftz.f32 %f10, %f441, 0f3F000000; setp.lt.ftz.f32 %p4, %f445, %f441; @%p4 bra $L__BB13_8; bra.uni $L__BB13_7; $L__BB13_8: setp.lt.ftz.f32 %p6, %f444, %f441; mov.u32 %r214, 0; @%p6 bra $L__BB13_10; sub.ftz.f32 %f444, %f444, %f441; mov.u32 %r214, 1; bra.uni $L__BB13_10; $L__BB13_7: setp.lt.ftz.f32 %p5, %f444, %f441; sub.ftz.f32 %f73, %f444, %f441; selp.f32 %f444, %f444, %f73, %p5; selp.b32 %r214, 3, 2, %p5; sub.ftz.f32 %f445, %f445, %f441; $L__BB13_10: add.s32 %r73, %r214, %r218; mul.wide.s32 %rd31, %r73, 4; add.s64 %rd32, %rd4, %rd31; ld.global.u32 %r12, [%rd32]; and.b32 %r74, %r12, 1073741824; setp.eq.s32 %p7, %r74, 0; mov.u32 %r219, -1; @%p7 bra $L__BB13_33; and.b32 %r218, %r12, 1073741823; setp.lt.s32 %p8, %r12, 0; @%p8 bra $L__BB13_32; mul.ftz.f32 %f16, %f10, 0f3F000000; setp.lt.ftz.f32 %p9, %f445, %f10; @%p9 bra $L__BB13_14; bra.uni $L__BB13_13; $L__BB13_14: setp.lt.ftz.f32 %p11, %f444, %f10; mov.u32 %r215, 0; @%p11 bra $L__BB13_16; sub.ftz.f32 %f444, %f444, %f10; mov.u32 %r215, 1; bra.uni $L__BB13_16; $L__BB13_13: setp.lt.ftz.f32 %p10, %f444, %f10; sub.ftz.f32 %f74, %f444, %f10; selp.f32 %f444, %f444, %f74, %p10; selp.b32 %r215, 3, 2, %p10; sub.ftz.f32 %f445, %f445, %f10; $L__BB13_16: add.s32 %r82, %r215, %r218; mul.wide.u32 %rd33, %r82, 4; add.s64 %rd34, %rd4, %rd33; ld.global.u32 %r16, [%rd34]; and.b32 %r83, %r16, 1073741824; setp.eq.s32 %p12, %r83, 0; @%p12 bra $L__BB13_33; and.b32 %r218, %r16, 1073741823; setp.lt.s32 %p13, %r16, 0; @%p13 bra $L__BB13_32; mul.ftz.f32 %f22, %f16, 0f3F000000; setp.lt.ftz.f32 %p14, %f445, %f16; @%p14 bra $L__BB13_20; bra.uni $L__BB13_19; $L__BB13_20: setp.lt.ftz.f32 %p16, %f444, %f16; mov.u32 %r216, 0; @%p16 bra $L__BB13_22; sub.ftz.f32 %f444, %f444, %f16; mov.u32 %r216, 1; bra.uni $L__BB13_22; $L__BB13_19: setp.lt.ftz.f32 %p15, %f444, %f16; sub.ftz.f32 %f75, %f444, %f16; selp.f32 %f444, %f444, %f75, %p15; selp.b32 %r216, 3, 2, %p15; sub.ftz.f32 %f445, %f445, %f16; $L__BB13_22: add.s32 %r91, %r216, %r218; mul.wide.u32 %rd35, %r91, 4; add.s64 %rd36, %rd4, %rd35; ld.global.u32 %r20, [%rd36]; and.b32 %r92, %r20, 1073741824; setp.eq.s32 %p17, %r92, 0; @%p17 bra $L__BB13_33; and.b32 %r218, %r20, 1073741823; setp.lt.s32 %p18, %r20, 0; @%p18 bra $L__BB13_32; add.s32 %r98, %r212, 3; setp.gt.u32 %p19, %r98, 254; @%p19 bra $L__BB13_33; setp.lt.ftz.f32 %p20, %f445, %f22; @%p20 bra $L__BB13_27; bra.uni $L__BB13_26; $L__BB13_27: setp.lt.ftz.f32 %p22, %f444, %f22; mov.u32 %r217, 0; @%p22 bra $L__BB13_29; sub.ftz.f32 %f444, %f444, %f22; mov.u32 %r217, 1; bra.uni $L__BB13_29; $L__BB13_26: setp.lt.ftz.f32 %p21, %f444, %f22; sub.ftz.f32 %f76, %f444, %f22; selp.f32 %f444, %f444, %f76, %p21; selp.b32 %r217, 3, 2, %p21; sub.ftz.f32 %f445, %f445, %f22; $L__BB13_29: add.s32 %r106, %r217, %r218; mul.wide.u32 %rd37, %r106, 4; add.s64 %rd38, %rd4, %rd37; ld.global.u32 %r24, [%rd38]; and.b32 %r107, %r24, 1073741824; setp.eq.s32 %p23, %r107, 0; @%p23 bra $L__BB13_33; and.b32 %r218, %r24, 1073741823; setp.lt.s32 %p24, %r24, 0; @%p24 bra $L__BB13_32; bra.uni $L__BB13_31; $L__BB13_32: mul.wide.u32 %rd39, %r218, 4; add.s64 %rd40, %rd4, %rd39; ld.global.u32 %r221, [%rd40+8]; ld.global.u32 %r220, [%rd40+4]; ld.global.u32 %r219, [%rd40]; $L__BB13_33: add.s32 %r110, %r219, 2; mul.wide.s32 %rd41, %r110, 4; add.s64 %rd42, %rd4, %rd41; ld.global.u32 %r111, [%rd42]; add.s32 %r112, %r111, %r221; mul.wide.s32 %rd43, %r112, 4; add.s64 %rd44, %rd4, %rd43; add.s32 %r113, %r220, 1; mov.u32 %r114, 1; mul.wide.s32 %rd45, %r113, 4; add.s64 %rd46, %rd4, %rd45; ld.global.u32 %r115, [%rd46]; and.b32 %r116, %r115, 15; shl.b32 %r117, %r114, %r116; cvt.rn.f32.s32 %f77, %r117; and.b32 %r118, %r115, 16; setp.eq.s32 %p25, %r118, 0; mul.ftz.f32 %f78, %f77, 0f3F000000; selp.f32 %f79, %f77, %f78, %p25; shr.u32 %r119, %r115, 4; and.b32 %r120, %r119, 1; setp.eq.b32 %p26, %r120, 1; selp.b32 %r121, -1, 0, %p26; add.s32 %r122, %r116, %r121; shl.b32 %r123, %r114, %r122; cvt.rn.f32.s32 %f80, %r123; rcp.approx.ftz.f32 %f81, %f80; mov.f32 %f82, 0f3F800000; shr.u32 %r124, %r115, 22; bfe.u32 %r125, %r115, 12, 10; cvt.rn.f32.s32 %f83, %r124; mul.ftz.f32 %f84, %f81, %f83; cvt.rn.f32.s32 %f85, %r125; mul.ftz.f32 %f86, %f81, %f85; sub.ftz.f32 %f87, %f5, %f84; div.approx.ftz.f32 %f88, %f87, %f81; sub.ftz.f32 %f89, %f6, %f86; div.approx.ftz.f32 %f90, %f89, %f81; mul.ftz.f32 %f91, %f88, %f88; mul.ftz.f32 %f92, %f88, %f91; sub.ftz.f32 %f93, %f88, %f91; mul.ftz.f32 %f94, %f93, 0f40400000; sub.ftz.f32 %f95, %f82, %f94; sub.ftz.f32 %f96, %f95, %f92; mul.ftz.f32 %f97, %f96, 0f3E2AAAAB; mul.ftz.f32 %f98, %f91, 0f40C00000; mov.f32 %f99, 0f40800000; sub.ftz.f32 %f100, %f99, %f98; fma.rn.ftz.f32 %f101, %f92, 0f40400000, %f100; mul.ftz.f32 %f102, %f101, 0f3E2AAAAB; add.ftz.f32 %f103, %f88, %f91; sub.ftz.f32 %f104, %f103, %f92; fma.rn.ftz.f32 %f105, %f104, 0f40400000, 0f3F800000; mul.ftz.f32 %f106, %f105, 0f3E2AAAAB; mul.ftz.f32 %f107, %f92, 0f3E2AAAAB; mul.ftz.f32 %f108, %f91, 0f3F000000; sub.ftz.f32 %f109, %f88, %f108; add.ftz.f32 %f110, %f109, 0fBF000000; mul.ftz.f32 %f111, %f91, 0f3FC00000; add.ftz.f32 %f112, %f88, %f88; sub.ftz.f32 %f113, %f111, %f112; sub.ftz.f32 %f114, %f88, %f111; add.ftz.f32 %f115, %f114, 0f3F000000; mul.ftz.f32 %f116, %f90, %f90; mul.ftz.f32 %f117, %f90, %f116; sub.ftz.f32 %f118, %f90, %f116; mul.ftz.f32 %f119, %f118, 0f40400000; sub.ftz.f32 %f120, %f82, %f119; sub.ftz.f32 %f121, %f120, %f117; mul.ftz.f32 %f122, %f121, 0f3E2AAAAB; mul.ftz.f32 %f123, %f116, 0f40C00000; sub.ftz.f32 %f124, %f99, %f123; fma.rn.ftz.f32 %f125, %f117, 0f40400000, %f124; mul.ftz.f32 %f126, %f125, 0f3E2AAAAB; add.ftz.f32 %f127, %f90, %f116; sub.ftz.f32 %f128, %f127, %f117; fma.rn.ftz.f32 %f129, %f128, 0f40400000, 0f3F800000; mul.ftz.f32 %f130, %f129, 0f3E2AAAAB; mul.ftz.f32 %f131, %f117, 0f3E2AAAAB; mul.ftz.f32 %f132, %f116, 0f3F000000; sub.ftz.f32 %f133, %f90, %f132; add.ftz.f32 %f134, %f133, 0fBF000000; mul.ftz.f32 %f135, %f116, 0f3FC00000; add.ftz.f32 %f136, %f90, %f90; sub.ftz.f32 %f137, %f135, %f136; sub.ftz.f32 %f138, %f90, %f135; add.ftz.f32 %f139, %f138, 0f3F000000; and.b32 %r126, %r115, 256; setp.eq.s32 %p27, %r126, 0; sub.ftz.f32 %f140, %f130, %f122; fma.rn.ftz.f32 %f141, %f122, 0f40000000, %f126; selp.f32 %f142, %f122, 0f00000000, %p27; selp.f32 %f143, %f126, %f141, %p27; selp.f32 %f144, %f130, %f140, %p27; and.b32 %r127, %r115, 512; setp.eq.s32 %p28, %r127, 0; sub.ftz.f32 %f145, %f102, %f107; fma.rn.ftz.f32 %f146, %f107, 0f40000000, %f106; selp.f32 %f147, %f102, %f145, %p28; selp.f32 %f148, %f106, %f146, %p28; selp.f32 %f149, %f107, 0f00000000, %p28; and.b32 %r128, %r115, 1024; setp.eq.s32 %p29, %r128, 0; sub.ftz.f32 %f150, %f143, %f131; fma.rn.ftz.f32 %f151, %f131, 0f40000000, %f144; selp.f32 %f152, %f143, %f150, %p29; selp.f32 %f153, %f144, %f151, %p29; selp.f32 %f154, %f131, 0f00000000, %p29; and.b32 %r129, %r115, 2048; setp.eq.s32 %p30, %r129, 0; sub.ftz.f32 %f155, %f148, %f97; fma.rn.ftz.f32 %f156, %f97, 0f40000000, %f147; selp.f32 %f157, %f97, 0f00000000, %p30; selp.f32 %f158, %f147, %f156, %p30; selp.f32 %f159, %f148, %f155, %p30; sub.ftz.f32 %f160, %f139, %f134; fma.rn.ftz.f32 %f161, %f134, 0f40000000, %f137; selp.f32 %f162, %f134, 0f00000000, %p27; selp.f32 %f163, %f137, %f161, %p27; selp.f32 %f164, %f139, %f160, %p27; sub.ftz.f32 %f165, %f113, %f108; fma.rn.ftz.f32 %f166, %f108, 0f40000000, %f115; selp.f32 %f167, %f113, %f165, %p28; selp.f32 %f168, %f115, %f166, %p28; selp.f32 %f169, %f108, 0f00000000, %p28; sub.ftz.f32 %f170, %f163, %f132; fma.rn.ftz.f32 %f171, %f132, 0f40000000, %f164; selp.f32 %f172, %f163, %f170, %p29; selp.f32 %f173, %f164, %f171, %p29; selp.f32 %f174, %f132, 0f00000000, %p29; sub.ftz.f32 %f175, %f168, %f110; fma.rn.ftz.f32 %f176, %f110, 0f40000000, %f167; selp.f32 %f177, %f110, 0f00000000, %p30; selp.f32 %f178, %f167, %f176, %p30; selp.f32 %f179, %f168, %f175, %p30; mul.ftz.f32 %f180, %f157, %f142; mul.ftz.f32 %f181, %f177, %f142; mul.ftz.f32 %f182, %f79, %f181; mul.ftz.f32 %f183, %f157, %f162; mul.ftz.f32 %f184, %f79, %f183; mul.ftz.f32 %f185, %f158, %f142; mul.ftz.f32 %f186, %f178, %f142; mul.ftz.f32 %f187, %f79, %f186; mul.ftz.f32 %f188, %f158, %f162; mul.ftz.f32 %f189, %f79, %f188; mul.ftz.f32 %f190, %f159, %f142; mul.ftz.f32 %f191, %f179, %f142; mul.ftz.f32 %f192, %f79, %f191; mul.ftz.f32 %f193, %f159, %f162; mul.ftz.f32 %f194, %f79, %f193; mul.ftz.f32 %f195, %f149, %f142; mul.ftz.f32 %f196, %f169, %f142; mul.ftz.f32 %f197, %f79, %f196; mul.ftz.f32 %f198, %f149, %f162; mul.ftz.f32 %f199, %f79, %f198; mul.ftz.f32 %f200, %f157, %f152; mul.ftz.f32 %f201, %f177, %f152; mul.ftz.f32 %f202, %f79, %f201; mul.ftz.f32 %f203, %f157, %f172; mul.ftz.f32 %f204, %f79, %f203; mul.ftz.f32 %f205, %f158, %f152; mul.ftz.f32 %f206, %f178, %f152; mul.ftz.f32 %f207, %f79, %f206; mul.ftz.f32 %f208, %f158, %f172; mul.ftz.f32 %f209, %f79, %f208; mul.ftz.f32 %f210, %f159, %f152; mul.ftz.f32 %f211, %f179, %f152; mul.ftz.f32 %f212, %f79, %f211; mul.ftz.f32 %f213, %f159, %f172; mul.ftz.f32 %f214, %f79, %f213; mul.ftz.f32 %f215, %f149, %f152; mul.ftz.f32 %f216, %f169, %f152; mul.ftz.f32 %f217, %f79, %f216; mul.ftz.f32 %f218, %f149, %f172; mul.ftz.f32 %f219, %f79, %f218; mul.ftz.f32 %f220, %f157, %f153; mul.ftz.f32 %f221, %f177, %f153; mul.ftz.f32 %f222, %f79, %f221; mul.ftz.f32 %f223, %f157, %f173; mul.ftz.f32 %f224, %f79, %f223; mul.ftz.f32 %f225, %f158, %f153; mul.ftz.f32 %f226, %f178, %f153; mul.ftz.f32 %f227, %f79, %f226; mul.ftz.f32 %f228, %f158, %f173; mul.ftz.f32 %f229, %f79, %f228; mul.ftz.f32 %f230, %f159, %f153; mul.ftz.f32 %f231, %f179, %f153; mul.ftz.f32 %f232, %f79, %f231; mul.ftz.f32 %f233, %f159, %f173; mul.ftz.f32 %f234, %f79, %f233; mul.ftz.f32 %f235, %f149, %f153; mul.ftz.f32 %f236, %f169, %f153; mul.ftz.f32 %f237, %f79, %f236; mul.ftz.f32 %f238, %f149, %f173; mul.ftz.f32 %f239, %f79, %f238; mul.ftz.f32 %f240, %f157, %f154; mul.ftz.f32 %f241, %f177, %f154; mul.ftz.f32 %f242, %f79, %f241; mul.ftz.f32 %f243, %f157, %f174; mul.ftz.f32 %f244, %f79, %f243; mul.ftz.f32 %f245, %f158, %f154; mul.ftz.f32 %f246, %f178, %f154; mul.ftz.f32 %f247, %f79, %f246; mul.ftz.f32 %f248, %f158, %f174; mul.ftz.f32 %f249, %f79, %f248; mul.ftz.f32 %f250, %f159, %f154; mul.ftz.f32 %f251, %f179, %f154; mul.ftz.f32 %f252, %f79, %f251; mul.ftz.f32 %f253, %f159, %f174; mul.ftz.f32 %f254, %f79, %f253; mul.ftz.f32 %f255, %f149, %f154; mul.ftz.f32 %f256, %f169, %f154; mul.ftz.f32 %f257, %f79, %f256; mul.ftz.f32 %f258, %f149, %f174; mul.ftz.f32 %f259, %f79, %f258; ld.const.u64 %rd47, [kernel_params+2320]; cvta.to.global.u64 %rd48, %rd47; ld.global.u32 %r130, [%rd44]; add.s32 %r131, %r130, %r52; mul.wide.s32 %rd49, %r131, 4; add.s64 %rd50, %rd48, %rd49; ld.global.f32 %f260, [%rd50]; fma.rn.ftz.f32 %f261, %f260, %f180, 0f00000000; fma.rn.ftz.f32 %f262, %f260, %f182, 0f00000000; fma.rn.ftz.f32 %f263, %f260, %f184, 0f00000000; ld.global.u32 %r132, [%rd44+4]; add.s32 %r133, %r132, %r52; mul.wide.s32 %rd51, %r133, 4; add.s64 %rd52, %rd48, %rd51; ld.global.f32 %f264, [%rd52]; fma.rn.ftz.f32 %f265, %f264, %f185, %f261; fma.rn.ftz.f32 %f266, %f264, %f187, %f262; fma.rn.ftz.f32 %f267, %f264, %f189, %f263; ld.global.u32 %r134, [%rd44+8]; add.s32 %r135, %r134, %r52; mul.wide.s32 %rd53, %r135, 4; add.s64 %rd54, %rd48, %rd53; ld.global.f32 %f268, [%rd54]; fma.rn.ftz.f32 %f269, %f268, %f190, %f265; fma.rn.ftz.f32 %f270, %f268, %f192, %f266; fma.rn.ftz.f32 %f271, %f268, %f194, %f267; ld.global.u32 %r136, [%rd44+12]; add.s32 %r137, %r136, %r52; mul.wide.s32 %rd55, %r137, 4; add.s64 %rd56, %rd48, %rd55; ld.global.f32 %f272, [%rd56]; fma.rn.ftz.f32 %f273, %f272, %f195, %f269; fma.rn.ftz.f32 %f274, %f272, %f197, %f270; fma.rn.ftz.f32 %f275, %f272, %f199, %f271; ld.global.u32 %r138, [%rd44+16]; add.s32 %r139, %r138, %r52; mul.wide.s32 %rd57, %r139, 4; add.s64 %rd58, %rd48, %rd57; ld.global.f32 %f276, [%rd58]; fma.rn.ftz.f32 %f277, %f276, %f200, %f273; fma.rn.ftz.f32 %f278, %f276, %f202, %f274; fma.rn.ftz.f32 %f279, %f276, %f204, %f275; ld.global.u32 %r140, [%rd44+20]; add.s32 %r141, %r140, %r52; mul.wide.s32 %rd59, %r141, 4; add.s64 %rd60, %rd48, %rd59; ld.global.f32 %f280, [%rd60]; fma.rn.ftz.f32 %f281, %f280, %f205, %f277; fma.rn.ftz.f32 %f282, %f280, %f207, %f278; fma.rn.ftz.f32 %f283, %f280, %f209, %f279; ld.global.u32 %r142, [%rd44+24]; add.s32 %r143, %r142, %r52; mul.wide.s32 %rd61, %r143, 4; add.s64 %rd62, %rd48, %rd61; ld.global.f32 %f284, [%rd62]; fma.rn.ftz.f32 %f285, %f284, %f210, %f281; fma.rn.ftz.f32 %f286, %f284, %f212, %f282; fma.rn.ftz.f32 %f287, %f284, %f214, %f283; ld.global.u32 %r144, [%rd44+28]; add.s32 %r145, %r144, %r52; mul.wide.s32 %rd63, %r145, 4; add.s64 %rd64, %rd48, %rd63; ld.global.f32 %f288, [%rd64]; fma.rn.ftz.f32 %f289, %f288, %f215, %f285; fma.rn.ftz.f32 %f290, %f288, %f217, %f286; fma.rn.ftz.f32 %f291, %f288, %f219, %f287; ld.global.u32 %r146, [%rd44+32]; add.s32 %r147, %r146, %r52; mul.wide.s32 %rd65, %r147, 4; add.s64 %rd66, %rd48, %rd65; ld.global.f32 %f292, [%rd66]; fma.rn.ftz.f32 %f293, %f292, %f220, %f289; fma.rn.ftz.f32 %f294, %f292, %f222, %f290; fma.rn.ftz.f32 %f295, %f292, %f224, %f291; ld.global.u32 %r148, [%rd44+36]; add.s32 %r149, %r148, %r52; mul.wide.s32 %rd67, %r149, 4; add.s64 %rd68, %rd48, %rd67; ld.global.f32 %f296, [%rd68]; fma.rn.ftz.f32 %f297, %f296, %f225, %f293; fma.rn.ftz.f32 %f298, %f296, %f227, %f294; fma.rn.ftz.f32 %f299, %f296, %f229, %f295; ld.global.u32 %r150, [%rd44+40]; add.s32 %r151, %r150, %r52; mul.wide.s32 %rd69, %r151, 4; add.s64 %rd70, %rd48, %rd69; ld.global.f32 %f300, [%rd70]; fma.rn.ftz.f32 %f301, %f300, %f230, %f297; fma.rn.ftz.f32 %f302, %f300, %f232, %f298; fma.rn.ftz.f32 %f303, %f300, %f234, %f299; ld.global.u32 %r152, [%rd44+44]; add.s32 %r153, %r152, %r52; mul.wide.s32 %rd71, %r153, 4; add.s64 %rd72, %rd48, %rd71; ld.global.f32 %f304, [%rd72]; fma.rn.ftz.f32 %f305, %f304, %f235, %f301; fma.rn.ftz.f32 %f306, %f304, %f237, %f302; fma.rn.ftz.f32 %f307, %f304, %f239, %f303; ld.global.u32 %r154, [%rd44+48]; add.s32 %r155, %r154, %r52; mul.wide.s32 %rd73, %r155, 4; add.s64 %rd74, %rd48, %rd73; ld.global.f32 %f308, [%rd74]; fma.rn.ftz.f32 %f309, %f308, %f240, %f305; fma.rn.ftz.f32 %f310, %f308, %f242, %f306; fma.rn.ftz.f32 %f311, %f308, %f244, %f307; ld.global.u32 %r156, [%rd44+52]; add.s32 %r157, %r156, %r52; mul.wide.s32 %rd75, %r157, 4; add.s64 %rd76, %rd48, %rd75; ld.global.f32 %f312, [%rd76]; fma.rn.ftz.f32 %f313, %f312, %f245, %f309; fma.rn.ftz.f32 %f314, %f312, %f247, %f310; fma.rn.ftz.f32 %f315, %f312, %f249, %f311; ld.global.u32 %r158, [%rd44+56]; add.s32 %r159, %r158, %r52; mul.wide.s32 %rd77, %r159, 4; add.s64 %rd78, %rd48, %rd77; ld.global.f32 %f316, [%rd78]; fma.rn.ftz.f32 %f317, %f316, %f250, %f313; fma.rn.ftz.f32 %f318, %f316, %f252, %f314; fma.rn.ftz.f32 %f319, %f316, %f254, %f315; ld.global.u32 %r160, [%rd44+60]; add.s32 %r161, %r160, %r52; mul.wide.s32 %rd79, %r161, 4; add.s64 %rd80, %rd48, %rd79; ld.global.f32 %f320, [%rd80]; fma.rn.ftz.f32 %f452, %f320, %f255, %f317; fma.rn.ftz.f32 %f35, %f320, %f257, %f318; fma.rn.ftz.f32 %f36, %f320, %f259, %f319; or.b64 %rd81, %rd6, %rd7; setp.eq.s64 %p31, %rd81, 0; @%p31 bra $L__BB13_70; setp.eq.s64 %p32, %rd6, 0; @%p32 bra $L__BB13_36; ld.local.f32 %f321, [%rd1+40]; ld.local.f32 %f322, [%rd1+48]; mul.ftz.f32 %f323, %f3, %f322; fma.rn.ftz.f32 %f324, %f1, %f321, %f323; mul.ftz.f32 %f325, %f4, %f322; fma.rn.ftz.f32 %f326, %f2, %f321, %f325; mul.ftz.f32 %f327, %f36, %f326; fma.rn.ftz.f32 %f328, %f35, %f324, %f327; st.f32 [%rd6], %f328; $L__BB13_36: setp.eq.s64 %p33, %rd7, 0; @%p33 bra $L__BB13_70; ld.local.f32 %f329, [%rd1+44]; ld.local.f32 %f330, [%rd1+52]; mul.ftz.f32 %f331, %f3, %f330; fma.rn.ftz.f32 %f332, %f1, %f329, %f331; mul.ftz.f32 %f333, %f4, %f330; fma.rn.ftz.f32 %f334, %f2, %f329, %f333; mul.ftz.f32 %f335, %f36, %f334; fma.rn.ftz.f32 %f336, %f35, %f332, %f335; st.f32 [%rd7], %f336; bra.uni $L__BB13_70; $L__BB13_39: setp.eq.s32 %p35, %r49, 16; @%p35 bra $L__BB13_41; setp.ne.s32 %p36, %r49, 8; @%p36 bra $L__BB13_46; $L__BB13_41: cvt.u32.u64 %r162, %rd3; ld.const.u64 %rd82, [kernel_params+2232]; cvta.to.global.u64 %rd83, %rd82; mul.wide.s32 %rd84, %r162, 16; add.s64 %rd85, %rd83, %rd84; ld.global.v4.u32 {%r163, %r164, %r165, %r166}, [%rd85]; ld.const.u64 %rd86, [kernel_params+2248]; cvta.to.global.u64 %rd87, %rd86; mul.wide.u32 %rd88, %r163, 8; add.s64 %rd89, %rd87, %rd88; ld.global.v2.f32 {%f337, %f338}, [%rd89]; mul.wide.u32 %rd90, %r164, 8; add.s64 %rd91, %rd87, %rd90; ld.global.v2.f32 {%f341, %f342}, [%rd91]; mul.wide.u32 %rd92, %r165, 8; add.s64 %rd93, %rd87, %rd92; ld.global.v2.f32 {%f345, %f346}, [%rd93]; ld.const.u64 %rd94, [kernel_params+2288]; cvta.to.global.u64 %rd95, %rd94; mul.wide.s32 %rd96, %r3, 4; add.s64 %rd97, %rd95, %rd96; ld.const.u64 %rd98, [kernel_params+2320]; cvta.to.global.u64 %rd99, %rd98; ld.global.u32 %r170, [%rd97]; add.s32 %r171, %r170, %r52; mul.wide.u32 %rd100, %r171, 4; add.s64 %rd101, %rd99, %rd100; ld.global.u32 %r172, [%rd97+4]; add.s32 %r173, %r172, %r52; mul.wide.u32 %rd102, %r173, 4; add.s64 %rd103, %rd99, %rd102; ld.global.u32 %r174, [%rd97+8]; add.s32 %r175, %r174, %r52; mul.wide.u32 %rd104, %r175, 4; add.s64 %rd105, %rd99, %rd104; ld.global.u32 %r176, [%rd97+12]; add.s32 %r177, %r176, %r52; mul.wide.u32 %rd106, %r177, 4; add.s64 %rd107, %rd99, %rd106; ld.global.u16 %rs1, [%rd97+20]; setp.ne.s16 %p37, %rs1, 4; ld.global.f32 %f349, [%rd103]; ld.global.f32 %f350, [%rd101]; add.ftz.f32 %f351, %f350, %f349; mul.ftz.f32 %f352, %f351, 0f3F000000; ld.global.f32 %f353, [%rd107]; add.ftz.f32 %f354, %f350, %f353; mul.ftz.f32 %f355, %f354, 0f3F000000; selp.f32 %f356, %f352, %f349, %p37; selp.f32 %f357, %f355, %f353, %p37; sub.ftz.f32 %f358, %f356, %f350; fma.rn.ftz.f32 %f359, %f337, %f358, %f350; ld.global.f32 %f360, [%rd105]; sub.ftz.f32 %f361, %f360, %f357; fma.rn.ftz.f32 %f362, %f337, %f361, %f357; sub.ftz.f32 %f363, %f362, %f359; fma.rn.ftz.f32 %f37, %f338, %f363, %f359; fma.rn.ftz.f32 %f364, %f341, %f358, %f350; fma.rn.ftz.f32 %f365, %f341, %f361, %f357; sub.ftz.f32 %f366, %f365, %f364; fma.rn.ftz.f32 %f38, %f342, %f366, %f364; fma.rn.ftz.f32 %f367, %f345, %f358, %f350; fma.rn.ftz.f32 %f368, %f345, %f361, %f357; sub.ftz.f32 %f369, %f368, %f367; fma.rn.ftz.f32 %f39, %f346, %f369, %f367; setp.eq.s64 %p38, %rd6, 0; @%p38 bra $L__BB13_43; ld.local.f32 %f370, [%rd1+40]; ld.local.f32 %f371, [%rd1+48]; mul.ftz.f32 %f372, %f39, %f371; fma.rn.ftz.f32 %f373, %f38, %f370, %f372; add.ftz.f32 %f374, %f370, %f371; mul.ftz.f32 %f375, %f37, %f374; sub.ftz.f32 %f376, %f373, %f375; st.f32 [%rd6], %f376; $L__BB13_43: setp.eq.s64 %p39, %rd7, 0; @%p39 bra $L__BB13_45; ld.local.f32 %f377, [%rd1+44]; ld.local.f32 %f378, [%rd1+52]; mul.ftz.f32 %f379, %f39, %f378; fma.rn.ftz.f32 %f380, %f38, %f377, %f379; add.ftz.f32 %f381, %f377, %f378; mul.ftz.f32 %f382, %f37, %f381; sub.ftz.f32 %f383, %f380, %f382; st.f32 [%rd7], %f383; $L__BB13_45: ld.local.f32 %f384, [%rd1+8]; ld.local.f32 %f385, [%rd1+12]; mul.ftz.f32 %f386, %f39, %f385; fma.rn.ftz.f32 %f387, %f38, %f384, %f386; mov.f32 %f388, 0f3F800000; sub.ftz.f32 %f389, %f388, %f384; sub.ftz.f32 %f390, %f389, %f385; fma.rn.ftz.f32 %f452, %f37, %f390, %f387; $L__BB13_70: st.param.f32 [func_retval0+0], %f452; ret; $L__BB13_46: setp.eq.s32 %p40, %r49, 32; @%p40 bra $L__BB13_57; bra.uni $L__BB13_47; $L__BB13_57: cvt.u32.u64 %r183, %rd3; ld.const.u64 %rd112, [kernel_params+2232]; cvta.to.global.u64 %rd113, %rd112; mul.wide.s32 %rd114, %r183, 16; add.s64 %rd115, %rd113, %rd114; ld.global.v4.u32 {%r184, %r185, %r186, %r187}, [%rd115]; ld.const.u64 %rd116, [kernel_params+2248]; cvta.to.global.u64 %rd117, %rd116; mul.wide.u32 %rd118, %r184, 8; add.s64 %rd119, %rd117, %rd118; ld.global.v2.f32 {%f393, %f394}, [%rd119]; mul.wide.u32 %rd120, %r185, 8; add.s64 %rd121, %rd117, %rd120; ld.global.v2.f32 {%f395, %f396}, [%rd121]; mul.wide.u32 %rd122, %r186, 8; add.s64 %rd123, %rd117, %rd122; ld.global.v2.f32 {%f397, %f398}, [%rd123]; ld.const.u64 %rd124, [kernel_params+2288]; cvta.to.global.u64 %rd125, %rd124; add.s32 %r191, %r3, 5; mul.wide.s32 %rd126, %r191, 4; add.s64 %rd5, %rd125, %rd126; ld.global.u32 %r35, [%rd5+4]; ld.global.u32 %r36, [%rd5]; and.b32 %r37, %r36, 65535; setp.eq.s32 %p46, %r37, 4; @%p46 bra $L__BB13_59; bra.uni $L__BB13_58; $L__BB13_59: add.s32 %r224, %r35, 1; add.s32 %r223, %r35, 2; add.s32 %r222, %r35, 3; mov.u32 %r225, %r35; bra.uni $L__BB13_60; $L__BB13_47: add.s32 %r178, %r49, -1; setp.lt.u32 %p41, %r178, 2; @%p41 bra $L__BB13_52; bra.uni $L__BB13_48; $L__BB13_52: setp.eq.s64 %p44, %rd6, 0; @%p44 bra $L__BB13_54; mov.u32 %r181, 0; st.u32 [%rd6], %r181; $L__BB13_54: setp.eq.s64 %p45, %rd7, 0; @%p45 bra $L__BB13_56; mov.u32 %r182, 0; st.u32 [%rd7], %r182; $L__BB13_56: ld.const.u64 %rd108, [kernel_params+2320]; cvta.to.global.u64 %rd109, %rd108; mul.wide.s32 %rd110, %r52, 4; add.s64 %rd111, %rd109, %rd110; ld.global.f32 %f452, [%rd111]; bra.uni $L__BB13_70; $L__BB13_58: ld.global.u32 %r223, [%rd5+8]; shr.u32 %r192, %r36, 16; add.s32 %r225, %r35, %r192; add.s32 %r193, %r192, 1; rem.u32 %r194, %r193, %r37; add.s32 %r195, %r194, %r37; rem.u32 %r196, %r195, %r37; add.s32 %r224, %r196, %r35; add.s32 %r197, %r192, -1; rem.s32 %r198, %r197, %r37; add.s32 %r199, %r198, %r37; rem.s32 %r200, %r199, %r37; add.s32 %r222, %r200, %r35; $L__BB13_60: ld.const.u64 %rd127, [kernel_params+2320]; cvta.to.global.u64 %rd128, %rd127; add.s32 %r201, %r225, %r52; mul.wide.s32 %rd129, %r201, 4; add.s64 %rd130, %rd128, %rd129; add.s32 %r202, %r224, %r52; mul.wide.s32 %rd131, %r202, 4; add.s64 %rd132, %rd128, %rd131; add.s32 %r203, %r223, %r52; mul.wide.s32 %rd133, %r203, 4; add.s64 %rd134, %rd128, %rd133; add.s32 %r204, %r222, %r52; mul.wide.s32 %rd135, %r204, 4; add.s64 %rd136, %rd128, %rd135; ld.global.f32 %f399, [%rd132]; ld.global.f32 %f400, [%rd130]; add.ftz.f32 %f401, %f400, %f399; mul.ftz.f32 %f402, %f401, 0f3F000000; ld.global.f32 %f403, [%rd136]; add.ftz.f32 %f404, %f400, %f403; mul.ftz.f32 %f405, %f404, 0f3F000000; setp.ne.s32 %p47, %r37, 4; selp.f32 %f406, %f402, %f399, %p47; selp.f32 %f407, %f405, %f403, %p47; sub.ftz.f32 %f408, %f406, %f400; fma.rn.ftz.f32 %f409, %f393, %f408, %f400; ld.global.f32 %f410, [%rd134]; sub.ftz.f32 %f411, %f410, %f407; fma.rn.ftz.f32 %f412, %f393, %f411, %f407; sub.ftz.f32 %f413, %f412, %f409; fma.rn.ftz.f32 %f48, %f394, %f413, %f409; fma.rn.ftz.f32 %f414, %f395, %f408, %f400; fma.rn.ftz.f32 %f415, %f395, %f411, %f407; sub.ftz.f32 %f416, %f415, %f414; fma.rn.ftz.f32 %f49, %f396, %f416, %f414; fma.rn.ftz.f32 %f417, %f397, %f408, %f400; fma.rn.ftz.f32 %f418, %f397, %f411, %f407; sub.ftz.f32 %f419, %f418, %f417; fma.rn.ftz.f32 %f50, %f398, %f419, %f417; setp.eq.s64 %p48, %rd6, 0; @%p48 bra $L__BB13_62; ld.local.f32 %f420, [%rd1+40]; ld.local.f32 %f421, [%rd1+48]; mul.ftz.f32 %f422, %f50, %f421; fma.rn.ftz.f32 %f423, %f49, %f420, %f422; add.ftz.f32 %f424, %f420, %f421; mul.ftz.f32 %f425, %f48, %f424; sub.ftz.f32 %f426, %f423, %f425; st.f32 [%rd6], %f426; $L__BB13_62: setp.eq.s64 %p49, %rd7, 0; @%p49 bra $L__BB13_64; ld.local.f32 %f427, [%rd1+44]; ld.local.f32 %f428, [%rd1+52]; mul.ftz.f32 %f429, %f50, %f428; fma.rn.ftz.f32 %f430, %f49, %f427, %f429; add.ftz.f32 %f431, %f427, %f428; mul.ftz.f32 %f432, %f48, %f431; sub.ftz.f32 %f433, %f430, %f432; st.f32 [%rd7], %f433; $L__BB13_64: ld.local.f32 %f434, [%rd1+8]; ld.local.f32 %f435, [%rd1+12]; mul.ftz.f32 %f436, %f50, %f435; fma.rn.ftz.f32 %f437, %f49, %f434, %f436; mov.f32 %f438, 0f3F800000; sub.ftz.f32 %f439, %f438, %f434; sub.ftz.f32 %f440, %f439, %f435; fma.rn.ftz.f32 %f452, %f48, %f440, %f437; bra.uni $L__BB13_70; $L__BB13_48: setp.eq.s64 %p42, %rd6, 0; @%p42 bra $L__BB13_50; mov.u32 %r179, 0; st.u32 [%rd6], %r179; $L__BB13_50: setp.eq.s64 %p43, %rd7, 0; mov.f32 %f452, 0f00000000; @%p43 bra $L__BB13_70; mov.u32 %r180, 0; st.u32 [%rd7], %r180; bra.uni $L__BB13_70; } .func (.param .align 4 .b8 func_retval0[8]) _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_( .param .b64 _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_0, .param .align 4 .b8 _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_1[16], .param .b64 _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_2, .param .b64 _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_3 ) { .reg .pred %p<52>; .reg .b16 %rs<2>; .reg .f32 %f<662>; .reg .b32 %r<220>; .reg .b64 %rd<151>; ld.param.u64 %rd10, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_0]; ld.param.u64 %rd8, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_2]; ld.param.u64 %rd9, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_3]; ld.param.u32 %r52, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_1+12]; ld.param.u32 %r51, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_1+8]; ld.param.u32 %r49, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_1]; cvta.to.local.u64 %rd11, %rd10; add.s64 %rd1, %rd11, 60; ld.local.s32 %rd3, [%rd11+60]; setp.eq.s64 %p1, %rd3, -1; mov.u32 %r2, -1; @%p1 bra $L__BB14_2; ld.const.u64 %rd12, [kernel_params+2240]; cvta.to.global.u64 %rd13, %rd12; shl.b64 %rd14, %rd3, 2; add.s64 %rd15, %rd13, %rd14; ld.global.u32 %r2, [%rd15]; $L__BB14_2: and.b32 %r54, %r51, 2; setp.eq.s32 %p2, %r54, 0; @%p2 bra $L__BB14_38; bra.uni $L__BB14_3; $L__BB14_38: setp.eq.s32 %p34, %r49, 4; @%p34 bra $L__BB14_69; bra.uni $L__BB14_39; $L__BB14_69: setp.eq.s64 %p50, %rd8, 0; @%p50 bra $L__BB14_71; mov.f32 %f637, 0f00000000; st.v2.f32 [%rd8], {%f637, %f637}; $L__BB14_71: setp.eq.s64 %p51, %rd9, 0; @%p51 bra $L__BB14_73; mov.f32 %f638, 0f00000000; st.v2.f32 [%rd9], {%f638, %f638}; $L__BB14_73: ld.const.u64 %rd139, [kernel_params+2328]; cvta.to.global.u64 %rd140, %rd139; ld.const.u64 %rd141, [kernel_params+2288]; cvta.to.global.u64 %rd142, %rd141; add.s32 %r201, %r2, 4; mul.wide.s32 %rd143, %r201, 4; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r202, [%rd144]; add.s32 %r203, %r202, %r52; mul.wide.u32 %rd145, %r203, 8; add.s64 %rd146, %rd140, %rd145; ld.global.v2.f32 {%f660, %f661}, [%rd146]; bra.uni $L__BB14_74; $L__BB14_3: ld.const.u64 %rd16, [kernel_params+2232]; cvta.to.global.u64 %rd17, %rd16; shl.b64 %rd18, %rd3, 4; add.s64 %rd19, %rd17, %rd18; ld.global.v4.u32 {%r56, %r57, %r58, %r59}, [%rd19]; mov.u32 %r206, 0; ld.const.u64 %rd20, [kernel_params+2248]; cvta.to.global.u64 %rd21, %rd20; mul.wide.u32 %rd22, %r56, 8; add.s64 %rd23, %rd21, %rd22; ld.global.v2.f32 {%f106, %f107}, [%rd23]; mul.wide.u32 %rd24, %r57, 8; add.s64 %rd25, %rd21, %rd24; ld.global.v2.f32 {%f110, %f111}, [%rd25]; mul.wide.u32 %rd26, %r58, 8; add.s64 %rd27, %rd21, %rd26; ld.global.v2.f32 {%f114, %f115}, [%rd27]; sub.ftz.f32 %f1, %f110, %f106; sub.ftz.f32 %f2, %f111, %f107; sub.ftz.f32 %f3, %f114, %f106; sub.ftz.f32 %f4, %f115, %f107; ld.local.f32 %f118, [%rd1+8]; ld.local.v2.u32 {%r63, %r64}, [%rd1+12]; mov.b32 %f119, %r63; mul.ftz.f32 %f120, %f3, %f119; mul.ftz.f32 %f121, %f4, %f119; fma.rn.ftz.f32 %f122, %f1, %f118, %f120; fma.rn.ftz.f32 %f123, %f2, %f118, %f121; add.ftz.f32 %f5, %f106, %f122; add.ftz.f32 %f6, %f107, %f123; setp.eq.s32 %p3, %r64, -1; mov.u32 %r205, %r206; @%p3 bra $L__BB14_5; ld.const.u64 %rd28, [kernel_params+2160]; cvta.to.global.u64 %rd29, %rd28; mul.wide.s32 %rd30, %r64, 208; add.s64 %rd31, %rd29, %rd30; ld.global.u32 %r205, [%rd31+160]; $L__BB14_5: add.s32 %r67, %r205, %r2; shr.u32 %r212, %r67, 1; ld.const.u64 %rd32, [kernel_params+2288]; cvta.to.global.u64 %rd6, %rd32; mov.f32 %f641, 0f3F000000; mov.f32 %f645, %f5; mov.f32 %f644, %f6; bra.uni $L__BB14_6; $L__BB14_31: add.s32 %r206, %r206, 4; mul.ftz.f32 %f641, %f22, 0f3F000000; $L__BB14_6: mul.ftz.f32 %f10, %f641, 0f3F000000; setp.lt.ftz.f32 %p4, %f645, %f641; @%p4 bra $L__BB14_8; bra.uni $L__BB14_7; $L__BB14_8: setp.lt.ftz.f32 %p6, %f644, %f641; mov.u32 %r208, 0; @%p6 bra $L__BB14_10; sub.ftz.f32 %f644, %f644, %f641; mov.u32 %r208, 1; bra.uni $L__BB14_10; $L__BB14_7: setp.lt.ftz.f32 %p5, %f644, %f641; sub.ftz.f32 %f125, %f644, %f641; selp.f32 %f644, %f644, %f125, %p5; selp.b32 %r208, 3, 2, %p5; sub.ftz.f32 %f645, %f645, %f641; $L__BB14_10: add.s32 %r73, %r208, %r212; mul.wide.s32 %rd33, %r73, 4; add.s64 %rd34, %rd6, %rd33; ld.global.u32 %r12, [%rd34]; and.b32 %r74, %r12, 1073741824; setp.eq.s32 %p7, %r74, 0; mov.u32 %r213, -1; @%p7 bra $L__BB14_33; and.b32 %r212, %r12, 1073741823; setp.lt.s32 %p8, %r12, 0; @%p8 bra $L__BB14_32; mul.ftz.f32 %f16, %f10, 0f3F000000; setp.lt.ftz.f32 %p9, %f645, %f10; @%p9 bra $L__BB14_14; bra.uni $L__BB14_13; $L__BB14_14: setp.lt.ftz.f32 %p11, %f644, %f10; mov.u32 %r209, 0; @%p11 bra $L__BB14_16; sub.ftz.f32 %f644, %f644, %f10; mov.u32 %r209, 1; bra.uni $L__BB14_16; $L__BB14_13: setp.lt.ftz.f32 %p10, %f644, %f10; sub.ftz.f32 %f126, %f644, %f10; selp.f32 %f644, %f644, %f126, %p10; selp.b32 %r209, 3, 2, %p10; sub.ftz.f32 %f645, %f645, %f10; $L__BB14_16: add.s32 %r82, %r209, %r212; mul.wide.u32 %rd35, %r82, 4; add.s64 %rd36, %rd6, %rd35; ld.global.u32 %r16, [%rd36]; and.b32 %r83, %r16, 1073741824; setp.eq.s32 %p12, %r83, 0; @%p12 bra $L__BB14_33; and.b32 %r212, %r16, 1073741823; setp.lt.s32 %p13, %r16, 0; @%p13 bra $L__BB14_32; mul.ftz.f32 %f22, %f16, 0f3F000000; setp.lt.ftz.f32 %p14, %f645, %f16; @%p14 bra $L__BB14_20; bra.uni $L__BB14_19; $L__BB14_20: setp.lt.ftz.f32 %p16, %f644, %f16; mov.u32 %r210, 0; @%p16 bra $L__BB14_22; sub.ftz.f32 %f644, %f644, %f16; mov.u32 %r210, 1; bra.uni $L__BB14_22; $L__BB14_19: setp.lt.ftz.f32 %p15, %f644, %f16; sub.ftz.f32 %f127, %f644, %f16; selp.f32 %f644, %f644, %f127, %p15; selp.b32 %r210, 3, 2, %p15; sub.ftz.f32 %f645, %f645, %f16; $L__BB14_22: add.s32 %r91, %r210, %r212; mul.wide.u32 %rd37, %r91, 4; add.s64 %rd38, %rd6, %rd37; ld.global.u32 %r20, [%rd38]; and.b32 %r92, %r20, 1073741824; setp.eq.s32 %p17, %r92, 0; @%p17 bra $L__BB14_33; and.b32 %r212, %r20, 1073741823; setp.lt.s32 %p18, %r20, 0; @%p18 bra $L__BB14_32; add.s32 %r98, %r206, 3; setp.gt.u32 %p19, %r98, 254; @%p19 bra $L__BB14_33; setp.lt.ftz.f32 %p20, %f645, %f22; @%p20 bra $L__BB14_27; bra.uni $L__BB14_26; $L__BB14_27: setp.lt.ftz.f32 %p22, %f644, %f22; mov.u32 %r211, 0; @%p22 bra $L__BB14_29; sub.ftz.f32 %f644, %f644, %f22; mov.u32 %r211, 1; bra.uni $L__BB14_29; $L__BB14_26: setp.lt.ftz.f32 %p21, %f644, %f22; sub.ftz.f32 %f128, %f644, %f22; selp.f32 %f644, %f644, %f128, %p21; selp.b32 %r211, 3, 2, %p21; sub.ftz.f32 %f645, %f645, %f22; $L__BB14_29: add.s32 %r106, %r211, %r212; mul.wide.u32 %rd39, %r106, 4; add.s64 %rd40, %rd6, %rd39; ld.global.u32 %r24, [%rd40]; and.b32 %r107, %r24, 1073741824; setp.eq.s32 %p23, %r107, 0; @%p23 bra $L__BB14_33; and.b32 %r212, %r24, 1073741823; setp.lt.s32 %p24, %r24, 0; @%p24 bra $L__BB14_32; bra.uni $L__BB14_31; $L__BB14_32: mul.wide.u32 %rd41, %r212, 4; add.s64 %rd42, %rd6, %rd41; ld.global.u32 %r215, [%rd42+8]; ld.global.u32 %r214, [%rd42+4]; ld.global.u32 %r213, [%rd42]; $L__BB14_33: add.s32 %r110, %r213, 2; mul.wide.s32 %rd43, %r110, 4; add.s64 %rd44, %rd6, %rd43; ld.global.u32 %r111, [%rd44]; add.s32 %r112, %r111, %r215; mul.wide.s32 %rd45, %r112, 4; add.s64 %rd46, %rd6, %rd45; add.s32 %r113, %r214, 1; mov.u32 %r114, 1; mul.wide.s32 %rd47, %r113, 4; add.s64 %rd48, %rd6, %rd47; ld.global.u32 %r115, [%rd48]; and.b32 %r116, %r115, 15; shl.b32 %r117, %r114, %r116; cvt.rn.f32.s32 %f129, %r117; and.b32 %r118, %r115, 16; setp.eq.s32 %p25, %r118, 0; mul.ftz.f32 %f130, %f129, 0f3F000000; selp.f32 %f131, %f129, %f130, %p25; shr.u32 %r119, %r115, 4; and.b32 %r120, %r119, 1; setp.eq.b32 %p26, %r120, 1; selp.b32 %r121, -1, 0, %p26; add.s32 %r122, %r116, %r121; shl.b32 %r123, %r114, %r122; cvt.rn.f32.s32 %f132, %r123; rcp.approx.ftz.f32 %f133, %f132; mov.f32 %f134, 0f3F800000; shr.u32 %r124, %r115, 22; bfe.u32 %r125, %r115, 12, 10; cvt.rn.f32.s32 %f135, %r124; mul.ftz.f32 %f136, %f133, %f135; cvt.rn.f32.s32 %f137, %r125; mul.ftz.f32 %f138, %f133, %f137; sub.ftz.f32 %f139, %f5, %f136; div.approx.ftz.f32 %f140, %f139, %f133; sub.ftz.f32 %f141, %f6, %f138; div.approx.ftz.f32 %f142, %f141, %f133; mul.ftz.f32 %f143, %f140, %f140; mul.ftz.f32 %f144, %f140, %f143; sub.ftz.f32 %f145, %f140, %f143; mul.ftz.f32 %f146, %f145, 0f40400000; sub.ftz.f32 %f147, %f134, %f146; sub.ftz.f32 %f148, %f147, %f144; mul.ftz.f32 %f149, %f148, 0f3E2AAAAB; mul.ftz.f32 %f150, %f143, 0f40C00000; mov.f32 %f151, 0f40800000; sub.ftz.f32 %f152, %f151, %f150; fma.rn.ftz.f32 %f153, %f144, 0f40400000, %f152; mul.ftz.f32 %f154, %f153, 0f3E2AAAAB; add.ftz.f32 %f155, %f140, %f143; sub.ftz.f32 %f156, %f155, %f144; fma.rn.ftz.f32 %f157, %f156, 0f40400000, 0f3F800000; mul.ftz.f32 %f158, %f157, 0f3E2AAAAB; mul.ftz.f32 %f159, %f144, 0f3E2AAAAB; mul.ftz.f32 %f160, %f143, 0f3F000000; sub.ftz.f32 %f161, %f140, %f160; add.ftz.f32 %f162, %f161, 0fBF000000; mul.ftz.f32 %f163, %f143, 0f3FC00000; add.ftz.f32 %f164, %f140, %f140; sub.ftz.f32 %f165, %f163, %f164; sub.ftz.f32 %f166, %f140, %f163; add.ftz.f32 %f167, %f166, 0f3F000000; mul.ftz.f32 %f168, %f142, %f142; mul.ftz.f32 %f169, %f142, %f168; sub.ftz.f32 %f170, %f142, %f168; mul.ftz.f32 %f171, %f170, 0f40400000; sub.ftz.f32 %f172, %f134, %f171; sub.ftz.f32 %f173, %f172, %f169; mul.ftz.f32 %f174, %f173, 0f3E2AAAAB; mul.ftz.f32 %f175, %f168, 0f40C00000; sub.ftz.f32 %f176, %f151, %f175; fma.rn.ftz.f32 %f177, %f169, 0f40400000, %f176; mul.ftz.f32 %f178, %f177, 0f3E2AAAAB; add.ftz.f32 %f179, %f142, %f168; sub.ftz.f32 %f180, %f179, %f169; fma.rn.ftz.f32 %f181, %f180, 0f40400000, 0f3F800000; mul.ftz.f32 %f182, %f181, 0f3E2AAAAB; mul.ftz.f32 %f183, %f169, 0f3E2AAAAB; mul.ftz.f32 %f184, %f168, 0f3F000000; sub.ftz.f32 %f185, %f142, %f184; add.ftz.f32 %f186, %f185, 0fBF000000; mul.ftz.f32 %f187, %f168, 0f3FC00000; add.ftz.f32 %f188, %f142, %f142; sub.ftz.f32 %f189, %f187, %f188; sub.ftz.f32 %f190, %f142, %f187; add.ftz.f32 %f191, %f190, 0f3F000000; and.b32 %r126, %r115, 256; setp.eq.s32 %p27, %r126, 0; sub.ftz.f32 %f192, %f182, %f174; fma.rn.ftz.f32 %f193, %f174, 0f40000000, %f178; selp.f32 %f194, %f174, 0f00000000, %p27; selp.f32 %f195, %f178, %f193, %p27; selp.f32 %f196, %f182, %f192, %p27; and.b32 %r127, %r115, 512; setp.eq.s32 %p28, %r127, 0; sub.ftz.f32 %f197, %f154, %f159; fma.rn.ftz.f32 %f198, %f159, 0f40000000, %f158; selp.f32 %f199, %f154, %f197, %p28; selp.f32 %f200, %f158, %f198, %p28; selp.f32 %f201, %f159, 0f00000000, %p28; and.b32 %r128, %r115, 1024; setp.eq.s32 %p29, %r128, 0; sub.ftz.f32 %f202, %f195, %f183; fma.rn.ftz.f32 %f203, %f183, 0f40000000, %f196; selp.f32 %f204, %f195, %f202, %p29; selp.f32 %f205, %f196, %f203, %p29; selp.f32 %f206, %f183, 0f00000000, %p29; and.b32 %r129, %r115, 2048; setp.eq.s32 %p30, %r129, 0; sub.ftz.f32 %f207, %f200, %f149; fma.rn.ftz.f32 %f208, %f149, 0f40000000, %f199; selp.f32 %f209, %f149, 0f00000000, %p30; selp.f32 %f210, %f199, %f208, %p30; selp.f32 %f211, %f200, %f207, %p30; sub.ftz.f32 %f212, %f191, %f186; fma.rn.ftz.f32 %f213, %f186, 0f40000000, %f189; selp.f32 %f214, %f186, 0f00000000, %p27; selp.f32 %f215, %f189, %f213, %p27; selp.f32 %f216, %f191, %f212, %p27; sub.ftz.f32 %f217, %f165, %f160; fma.rn.ftz.f32 %f218, %f160, 0f40000000, %f167; selp.f32 %f219, %f165, %f217, %p28; selp.f32 %f220, %f167, %f218, %p28; selp.f32 %f221, %f160, 0f00000000, %p28; sub.ftz.f32 %f222, %f215, %f184; fma.rn.ftz.f32 %f223, %f184, 0f40000000, %f216; selp.f32 %f224, %f215, %f222, %p29; selp.f32 %f225, %f216, %f223, %p29; selp.f32 %f226, %f184, 0f00000000, %p29; sub.ftz.f32 %f227, %f220, %f162; fma.rn.ftz.f32 %f228, %f162, 0f40000000, %f219; selp.f32 %f229, %f162, 0f00000000, %p30; selp.f32 %f230, %f219, %f228, %p30; selp.f32 %f231, %f220, %f227, %p30; mul.ftz.f32 %f232, %f209, %f194; mul.ftz.f32 %f233, %f229, %f194; mul.ftz.f32 %f234, %f131, %f233; mul.ftz.f32 %f235, %f209, %f214; mul.ftz.f32 %f236, %f131, %f235; mul.ftz.f32 %f237, %f210, %f194; mul.ftz.f32 %f238, %f230, %f194; mul.ftz.f32 %f239, %f131, %f238; mul.ftz.f32 %f240, %f210, %f214; mul.ftz.f32 %f241, %f131, %f240; mul.ftz.f32 %f242, %f211, %f194; mul.ftz.f32 %f243, %f231, %f194; mul.ftz.f32 %f244, %f131, %f243; mul.ftz.f32 %f245, %f211, %f214; mul.ftz.f32 %f246, %f131, %f245; mul.ftz.f32 %f247, %f201, %f194; mul.ftz.f32 %f248, %f221, %f194; mul.ftz.f32 %f249, %f131, %f248; mul.ftz.f32 %f250, %f201, %f214; mul.ftz.f32 %f251, %f131, %f250; mul.ftz.f32 %f252, %f209, %f204; mul.ftz.f32 %f253, %f229, %f204; mul.ftz.f32 %f254, %f131, %f253; mul.ftz.f32 %f255, %f209, %f224; mul.ftz.f32 %f256, %f131, %f255; mul.ftz.f32 %f257, %f210, %f204; mul.ftz.f32 %f258, %f230, %f204; mul.ftz.f32 %f259, %f131, %f258; mul.ftz.f32 %f260, %f210, %f224; mul.ftz.f32 %f261, %f131, %f260; mul.ftz.f32 %f262, %f211, %f204; mul.ftz.f32 %f263, %f231, %f204; mul.ftz.f32 %f264, %f131, %f263; mul.ftz.f32 %f265, %f211, %f224; mul.ftz.f32 %f266, %f131, %f265; mul.ftz.f32 %f267, %f201, %f204; mul.ftz.f32 %f268, %f221, %f204; mul.ftz.f32 %f269, %f131, %f268; mul.ftz.f32 %f270, %f201, %f224; mul.ftz.f32 %f271, %f131, %f270; mul.ftz.f32 %f272, %f209, %f205; mul.ftz.f32 %f273, %f229, %f205; mul.ftz.f32 %f274, %f131, %f273; mul.ftz.f32 %f275, %f209, %f225; mul.ftz.f32 %f276, %f131, %f275; mul.ftz.f32 %f277, %f210, %f205; mul.ftz.f32 %f278, %f230, %f205; mul.ftz.f32 %f279, %f131, %f278; mul.ftz.f32 %f280, %f210, %f225; mul.ftz.f32 %f281, %f131, %f280; mul.ftz.f32 %f282, %f211, %f205; mul.ftz.f32 %f283, %f231, %f205; mul.ftz.f32 %f284, %f131, %f283; mul.ftz.f32 %f285, %f211, %f225; mul.ftz.f32 %f286, %f131, %f285; mul.ftz.f32 %f287, %f201, %f205; mul.ftz.f32 %f288, %f221, %f205; mul.ftz.f32 %f289, %f131, %f288; mul.ftz.f32 %f290, %f201, %f225; mul.ftz.f32 %f291, %f131, %f290; mul.ftz.f32 %f292, %f209, %f206; mul.ftz.f32 %f293, %f229, %f206; mul.ftz.f32 %f294, %f131, %f293; mul.ftz.f32 %f295, %f209, %f226; mul.ftz.f32 %f296, %f131, %f295; mul.ftz.f32 %f297, %f210, %f206; mul.ftz.f32 %f298, %f230, %f206; mul.ftz.f32 %f299, %f131, %f298; mul.ftz.f32 %f300, %f210, %f226; mul.ftz.f32 %f301, %f131, %f300; mul.ftz.f32 %f302, %f211, %f206; mul.ftz.f32 %f303, %f231, %f206; mul.ftz.f32 %f304, %f131, %f303; mul.ftz.f32 %f305, %f211, %f226; mul.ftz.f32 %f306, %f131, %f305; mul.ftz.f32 %f307, %f201, %f206; mul.ftz.f32 %f308, %f221, %f206; mul.ftz.f32 %f309, %f131, %f308; mul.ftz.f32 %f310, %f201, %f226; mul.ftz.f32 %f311, %f131, %f310; ld.const.u64 %rd49, [kernel_params+2328]; cvta.to.global.u64 %rd50, %rd49; ld.global.u32 %r130, [%rd46]; add.s32 %r131, %r130, %r52; mul.wide.s32 %rd51, %r131, 8; add.s64 %rd52, %rd50, %rd51; ld.global.v2.f32 {%f312, %f313}, [%rd52]; fma.rn.ftz.f32 %f316, %f312, %f232, 0f00000000; fma.rn.ftz.f32 %f317, %f313, %f232, 0f00000000; fma.rn.ftz.f32 %f318, %f312, %f234, 0f00000000; fma.rn.ftz.f32 %f319, %f313, %f234, 0f00000000; fma.rn.ftz.f32 %f320, %f312, %f236, 0f00000000; fma.rn.ftz.f32 %f321, %f313, %f236, 0f00000000; ld.global.u32 %r132, [%rd46+4]; add.s32 %r133, %r132, %r52; mul.wide.s32 %rd53, %r133, 8; add.s64 %rd54, %rd50, %rd53; ld.global.v2.f32 {%f322, %f323}, [%rd54]; fma.rn.ftz.f32 %f326, %f322, %f237, %f316; fma.rn.ftz.f32 %f327, %f323, %f237, %f317; fma.rn.ftz.f32 %f328, %f322, %f239, %f318; fma.rn.ftz.f32 %f329, %f323, %f239, %f319; fma.rn.ftz.f32 %f330, %f322, %f241, %f320; fma.rn.ftz.f32 %f331, %f323, %f241, %f321; ld.global.u32 %r134, [%rd46+8]; add.s32 %r135, %r134, %r52; mul.wide.s32 %rd55, %r135, 8; add.s64 %rd56, %rd50, %rd55; ld.global.v2.f32 {%f332, %f333}, [%rd56]; fma.rn.ftz.f32 %f336, %f332, %f242, %f326; fma.rn.ftz.f32 %f337, %f333, %f242, %f327; fma.rn.ftz.f32 %f338, %f332, %f244, %f328; fma.rn.ftz.f32 %f339, %f333, %f244, %f329; fma.rn.ftz.f32 %f340, %f332, %f246, %f330; fma.rn.ftz.f32 %f341, %f333, %f246, %f331; ld.global.u32 %r136, [%rd46+12]; add.s32 %r137, %r136, %r52; mul.wide.s32 %rd57, %r137, 8; add.s64 %rd58, %rd50, %rd57; ld.global.v2.f32 {%f342, %f343}, [%rd58]; fma.rn.ftz.f32 %f346, %f342, %f247, %f336; fma.rn.ftz.f32 %f347, %f343, %f247, %f337; fma.rn.ftz.f32 %f348, %f342, %f249, %f338; fma.rn.ftz.f32 %f349, %f343, %f249, %f339; fma.rn.ftz.f32 %f350, %f342, %f251, %f340; fma.rn.ftz.f32 %f351, %f343, %f251, %f341; ld.global.u32 %r138, [%rd46+16]; add.s32 %r139, %r138, %r52; mul.wide.s32 %rd59, %r139, 8; add.s64 %rd60, %rd50, %rd59; ld.global.v2.f32 {%f352, %f353}, [%rd60]; fma.rn.ftz.f32 %f356, %f352, %f252, %f346; fma.rn.ftz.f32 %f357, %f353, %f252, %f347; fma.rn.ftz.f32 %f358, %f352, %f254, %f348; fma.rn.ftz.f32 %f359, %f353, %f254, %f349; fma.rn.ftz.f32 %f360, %f352, %f256, %f350; fma.rn.ftz.f32 %f361, %f353, %f256, %f351; ld.global.u32 %r140, [%rd46+20]; add.s32 %r141, %r140, %r52; mul.wide.s32 %rd61, %r141, 8; add.s64 %rd62, %rd50, %rd61; ld.global.v2.f32 {%f362, %f363}, [%rd62]; fma.rn.ftz.f32 %f366, %f362, %f257, %f356; fma.rn.ftz.f32 %f367, %f363, %f257, %f357; fma.rn.ftz.f32 %f368, %f362, %f259, %f358; fma.rn.ftz.f32 %f369, %f363, %f259, %f359; fma.rn.ftz.f32 %f370, %f362, %f261, %f360; fma.rn.ftz.f32 %f371, %f363, %f261, %f361; ld.global.u32 %r142, [%rd46+24]; add.s32 %r143, %r142, %r52; mul.wide.s32 %rd63, %r143, 8; add.s64 %rd64, %rd50, %rd63; ld.global.v2.f32 {%f372, %f373}, [%rd64]; fma.rn.ftz.f32 %f376, %f372, %f262, %f366; fma.rn.ftz.f32 %f377, %f373, %f262, %f367; fma.rn.ftz.f32 %f378, %f372, %f264, %f368; fma.rn.ftz.f32 %f379, %f373, %f264, %f369; fma.rn.ftz.f32 %f380, %f372, %f266, %f370; fma.rn.ftz.f32 %f381, %f373, %f266, %f371; ld.global.u32 %r144, [%rd46+28]; add.s32 %r145, %r144, %r52; mul.wide.s32 %rd65, %r145, 8; add.s64 %rd66, %rd50, %rd65; ld.global.v2.f32 {%f382, %f383}, [%rd66]; fma.rn.ftz.f32 %f386, %f382, %f267, %f376; fma.rn.ftz.f32 %f387, %f383, %f267, %f377; fma.rn.ftz.f32 %f388, %f382, %f269, %f378; fma.rn.ftz.f32 %f389, %f383, %f269, %f379; fma.rn.ftz.f32 %f390, %f382, %f271, %f380; fma.rn.ftz.f32 %f391, %f383, %f271, %f381; ld.global.u32 %r146, [%rd46+32]; add.s32 %r147, %r146, %r52; mul.wide.s32 %rd67, %r147, 8; add.s64 %rd68, %rd50, %rd67; ld.global.v2.f32 {%f392, %f393}, [%rd68]; fma.rn.ftz.f32 %f396, %f392, %f272, %f386; fma.rn.ftz.f32 %f397, %f393, %f272, %f387; fma.rn.ftz.f32 %f398, %f392, %f274, %f388; fma.rn.ftz.f32 %f399, %f393, %f274, %f389; fma.rn.ftz.f32 %f400, %f392, %f276, %f390; fma.rn.ftz.f32 %f401, %f393, %f276, %f391; ld.global.u32 %r148, [%rd46+36]; add.s32 %r149, %r148, %r52; mul.wide.s32 %rd69, %r149, 8; add.s64 %rd70, %rd50, %rd69; ld.global.v2.f32 {%f402, %f403}, [%rd70]; fma.rn.ftz.f32 %f406, %f402, %f277, %f396; fma.rn.ftz.f32 %f407, %f403, %f277, %f397; fma.rn.ftz.f32 %f408, %f402, %f279, %f398; fma.rn.ftz.f32 %f409, %f403, %f279, %f399; fma.rn.ftz.f32 %f410, %f402, %f281, %f400; fma.rn.ftz.f32 %f411, %f403, %f281, %f401; ld.global.u32 %r150, [%rd46+40]; add.s32 %r151, %r150, %r52; mul.wide.s32 %rd71, %r151, 8; add.s64 %rd72, %rd50, %rd71; ld.global.v2.f32 {%f412, %f413}, [%rd72]; fma.rn.ftz.f32 %f416, %f412, %f282, %f406; fma.rn.ftz.f32 %f417, %f413, %f282, %f407; fma.rn.ftz.f32 %f418, %f412, %f284, %f408; fma.rn.ftz.f32 %f419, %f413, %f284, %f409; fma.rn.ftz.f32 %f420, %f412, %f286, %f410; fma.rn.ftz.f32 %f421, %f413, %f286, %f411; ld.global.u32 %r152, [%rd46+44]; add.s32 %r153, %r152, %r52; mul.wide.s32 %rd73, %r153, 8; add.s64 %rd74, %rd50, %rd73; ld.global.v2.f32 {%f422, %f423}, [%rd74]; fma.rn.ftz.f32 %f426, %f422, %f287, %f416; fma.rn.ftz.f32 %f427, %f423, %f287, %f417; fma.rn.ftz.f32 %f428, %f422, %f289, %f418; fma.rn.ftz.f32 %f429, %f423, %f289, %f419; fma.rn.ftz.f32 %f430, %f422, %f291, %f420; fma.rn.ftz.f32 %f431, %f423, %f291, %f421; ld.global.u32 %r154, [%rd46+48]; add.s32 %r155, %r154, %r52; mul.wide.s32 %rd75, %r155, 8; add.s64 %rd76, %rd50, %rd75; ld.global.v2.f32 {%f432, %f433}, [%rd76]; fma.rn.ftz.f32 %f436, %f432, %f292, %f426; fma.rn.ftz.f32 %f437, %f433, %f292, %f427; fma.rn.ftz.f32 %f438, %f432, %f294, %f428; fma.rn.ftz.f32 %f439, %f433, %f294, %f429; fma.rn.ftz.f32 %f440, %f432, %f296, %f430; fma.rn.ftz.f32 %f441, %f433, %f296, %f431; ld.global.u32 %r156, [%rd46+52]; add.s32 %r157, %r156, %r52; mul.wide.s32 %rd77, %r157, 8; add.s64 %rd78, %rd50, %rd77; ld.global.v2.f32 {%f442, %f443}, [%rd78]; fma.rn.ftz.f32 %f446, %f442, %f297, %f436; fma.rn.ftz.f32 %f447, %f443, %f297, %f437; fma.rn.ftz.f32 %f448, %f442, %f299, %f438; fma.rn.ftz.f32 %f449, %f443, %f299, %f439; fma.rn.ftz.f32 %f450, %f442, %f301, %f440; fma.rn.ftz.f32 %f451, %f443, %f301, %f441; ld.global.u32 %r158, [%rd46+56]; add.s32 %r159, %r158, %r52; mul.wide.s32 %rd79, %r159, 8; add.s64 %rd80, %rd50, %rd79; ld.global.v2.f32 {%f452, %f453}, [%rd80]; fma.rn.ftz.f32 %f456, %f452, %f302, %f446; fma.rn.ftz.f32 %f457, %f453, %f302, %f447; fma.rn.ftz.f32 %f458, %f452, %f304, %f448; fma.rn.ftz.f32 %f459, %f453, %f304, %f449; fma.rn.ftz.f32 %f460, %f452, %f306, %f450; fma.rn.ftz.f32 %f461, %f453, %f306, %f451; ld.global.u32 %r160, [%rd46+60]; add.s32 %r161, %r160, %r52; mul.wide.s32 %rd81, %r161, 8; add.s64 %rd82, %rd50, %rd81; ld.global.v2.f32 {%f462, %f463}, [%rd82]; fma.rn.ftz.f32 %f34, %f462, %f309, %f458; fma.rn.ftz.f32 %f35, %f463, %f309, %f459; fma.rn.ftz.f32 %f36, %f462, %f311, %f460; fma.rn.ftz.f32 %f37, %f463, %f311, %f461; fma.rn.ftz.f32 %f661, %f463, %f307, %f457; fma.rn.ftz.f32 %f660, %f462, %f307, %f456; or.b64 %rd83, %rd8, %rd9; setp.eq.s64 %p31, %rd83, 0; @%p31 bra $L__BB14_74; ld.param.u64 %rd147, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_2]; setp.eq.s64 %p32, %rd147, 0; @%p32 bra $L__BB14_36; ld.param.u64 %rd148, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_2]; ld.local.f32 %f466, [%rd1+40]; ld.local.f32 %f467, [%rd1+48]; mul.ftz.f32 %f468, %f3, %f467; fma.rn.ftz.f32 %f469, %f1, %f466, %f468; mul.ftz.f32 %f470, %f4, %f467; fma.rn.ftz.f32 %f471, %f2, %f466, %f470; mul.ftz.f32 %f472, %f36, %f471; mul.ftz.f32 %f473, %f37, %f471; fma.rn.ftz.f32 %f474, %f35, %f469, %f473; fma.rn.ftz.f32 %f475, %f34, %f469, %f472; st.v2.f32 [%rd148], {%f475, %f474}; $L__BB14_36: ld.param.u64 %rd149, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_3]; setp.eq.s64 %p33, %rd149, 0; @%p33 bra $L__BB14_74; ld.param.u64 %rd150, [_Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7__param_3]; ld.local.f32 %f476, [%rd1+44]; ld.local.f32 %f477, [%rd1+52]; mul.ftz.f32 %f478, %f3, %f477; fma.rn.ftz.f32 %f479, %f1, %f476, %f478; mul.ftz.f32 %f480, %f4, %f477; fma.rn.ftz.f32 %f481, %f2, %f476, %f480; mul.ftz.f32 %f482, %f36, %f481; mul.ftz.f32 %f483, %f37, %f481; fma.rn.ftz.f32 %f484, %f35, %f479, %f483; fma.rn.ftz.f32 %f485, %f34, %f479, %f482; st.v2.f32 [%rd150], {%f485, %f484}; bra.uni $L__BB14_74; $L__BB14_39: setp.eq.s32 %p35, %r49, 16; @%p35 bra $L__BB14_41; setp.ne.s32 %p36, %r49, 8; @%p36 bra $L__BB14_48; $L__BB14_41: cvt.u32.u64 %r162, %rd3; ld.const.u64 %rd84, [kernel_params+2232]; cvta.to.global.u64 %rd85, %rd84; mul.wide.s32 %rd86, %r162, 16; add.s64 %rd87, %rd85, %rd86; ld.global.v4.u32 {%r163, %r164, %r165, %r166}, [%rd87]; ld.const.u64 %rd88, [kernel_params+2248]; cvta.to.global.u64 %rd89, %rd88; mul.wide.u32 %rd90, %r163, 8; add.s64 %rd91, %rd89, %rd90; ld.global.v2.f32 {%f486, %f487}, [%rd91]; mul.wide.u32 %rd92, %r164, 8; add.s64 %rd93, %rd89, %rd92; ld.global.v2.f32 {%f488, %f489}, [%rd93]; mul.wide.u32 %rd94, %r165, 8; add.s64 %rd95, %rd89, %rd94; ld.global.v2.f32 {%f490, %f491}, [%rd95]; ld.const.u64 %rd96, [kernel_params+2288]; cvta.to.global.u64 %rd97, %rd96; mul.wide.s32 %rd98, %r2, 4; add.s64 %rd99, %rd97, %rd98; ld.const.u64 %rd100, [kernel_params+2328]; cvta.to.global.u64 %rd101, %rd100; ld.global.u32 %r170, [%rd99]; add.s32 %r171, %r170, %r52; mul.wide.u32 %rd102, %r171, 8; add.s64 %rd103, %rd101, %rd102; ld.global.v2.f32 {%f492, %f493}, [%rd103]; ld.global.u32 %r172, [%rd99+4]; add.s32 %r173, %r172, %r52; mul.wide.u32 %rd104, %r173, 8; add.s64 %rd105, %rd101, %rd104; ld.global.v2.f32 {%f654, %f655}, [%rd105]; ld.global.u32 %r174, [%rd99+8]; add.s32 %r175, %r174, %r52; mul.wide.u32 %rd106, %r175, 8; add.s64 %rd107, %rd101, %rd106; ld.global.v2.f32 {%f496, %f497}, [%rd107]; ld.global.u32 %r176, [%rd99+12]; add.s32 %r177, %r176, %r52; mul.wide.u32 %rd108, %r177, 8; add.s64 %rd109, %rd101, %rd108; ld.global.v2.f32 {%f652, %f653}, [%rd109]; ld.global.u16 %rs1, [%rd99+20]; setp.eq.s16 %p37, %rs1, 4; @%p37 bra $L__BB14_43; add.ftz.f32 %f500, %f492, %f654; mul.ftz.f32 %f654, %f500, 0f3F000000; add.ftz.f32 %f501, %f493, %f655; mul.ftz.f32 %f655, %f501, 0f3F000000; add.ftz.f32 %f502, %f492, %f652; mul.ftz.f32 %f652, %f502, 0f3F000000; add.ftz.f32 %f503, %f493, %f653; mul.ftz.f32 %f653, %f503, 0f3F000000; $L__BB14_43: sub.ftz.f32 %f504, %f654, %f492; sub.ftz.f32 %f505, %f655, %f493; fma.rn.ftz.f32 %f506, %f486, %f504, %f492; fma.rn.ftz.f32 %f507, %f486, %f505, %f493; sub.ftz.f32 %f508, %f496, %f652; sub.ftz.f32 %f509, %f497, %f653; fma.rn.ftz.f32 %f510, %f486, %f508, %f652; fma.rn.ftz.f32 %f511, %f486, %f509, %f653; sub.ftz.f32 %f512, %f510, %f506; sub.ftz.f32 %f513, %f511, %f507; fma.rn.ftz.f32 %f62, %f487, %f512, %f506; fma.rn.ftz.f32 %f63, %f487, %f513, %f507; fma.rn.ftz.f32 %f514, %f488, %f504, %f492; fma.rn.ftz.f32 %f515, %f488, %f505, %f493; fma.rn.ftz.f32 %f516, %f488, %f508, %f652; fma.rn.ftz.f32 %f517, %f488, %f509, %f653; sub.ftz.f32 %f518, %f516, %f514; sub.ftz.f32 %f519, %f517, %f515; fma.rn.ftz.f32 %f64, %f489, %f518, %f514; fma.rn.ftz.f32 %f65, %f489, %f519, %f515; fma.rn.ftz.f32 %f520, %f490, %f504, %f492; fma.rn.ftz.f32 %f521, %f490, %f505, %f493; fma.rn.ftz.f32 %f522, %f490, %f508, %f652; fma.rn.ftz.f32 %f523, %f490, %f509, %f653; sub.ftz.f32 %f524, %f522, %f520; sub.ftz.f32 %f525, %f523, %f521; fma.rn.ftz.f32 %f66, %f491, %f524, %f520; fma.rn.ftz.f32 %f67, %f491, %f525, %f521; setp.eq.s64 %p38, %rd8, 0; @%p38 bra $L__BB14_45; ld.local.f32 %f526, [%rd1+40]; ld.local.f32 %f527, [%rd1+48]; mul.ftz.f32 %f528, %f66, %f527; mul.ftz.f32 %f529, %f67, %f527; fma.rn.ftz.f32 %f530, %f64, %f526, %f528; fma.rn.ftz.f32 %f531, %f65, %f526, %f529; add.ftz.f32 %f532, %f527, %f526; mul.ftz.f32 %f533, %f62, %f532; mul.ftz.f32 %f534, %f63, %f532; sub.ftz.f32 %f535, %f531, %f534; sub.ftz.f32 %f536, %f530, %f533; st.v2.f32 [%rd8], {%f536, %f535}; $L__BB14_45: setp.eq.s64 %p39, %rd9, 0; @%p39 bra $L__BB14_47; ld.local.f32 %f537, [%rd1+44]; ld.local.f32 %f538, [%rd1+52]; mul.ftz.f32 %f539, %f66, %f538; mul.ftz.f32 %f540, %f67, %f538; fma.rn.ftz.f32 %f541, %f64, %f537, %f539; fma.rn.ftz.f32 %f542, %f65, %f537, %f540; add.ftz.f32 %f543, %f538, %f537; mul.ftz.f32 %f544, %f62, %f543; mul.ftz.f32 %f545, %f63, %f543; sub.ftz.f32 %f546, %f542, %f545; sub.ftz.f32 %f547, %f541, %f544; st.v2.f32 [%rd9], {%f547, %f546}; $L__BB14_47: ld.local.f32 %f548, [%rd1+8]; ld.local.f32 %f549, [%rd1+12]; mul.ftz.f32 %f550, %f66, %f549; mul.ftz.f32 %f551, %f67, %f549; fma.rn.ftz.f32 %f552, %f64, %f548, %f550; fma.rn.ftz.f32 %f553, %f65, %f548, %f551; mov.f32 %f554, 0f3F800000; sub.ftz.f32 %f555, %f554, %f548; sub.ftz.f32 %f556, %f555, %f549; fma.rn.ftz.f32 %f661, %f63, %f556, %f553; fma.rn.ftz.f32 %f660, %f62, %f556, %f552; $L__BB14_74: st.param.f32 [func_retval0+0], %f660; st.param.f32 [func_retval0+4], %f661; ret; $L__BB14_48: setp.eq.s32 %p40, %r49, 32; @%p40 bra $L__BB14_59; bra.uni $L__BB14_49; $L__BB14_59: cvt.u32.u64 %r179, %rd3; ld.const.u64 %rd114, [kernel_params+2232]; cvta.to.global.u64 %rd115, %rd114; mul.wide.s32 %rd116, %r179, 16; add.s64 %rd117, %rd115, %rd116; ld.global.v4.u32 {%r180, %r181, %r182, %r183}, [%rd117]; ld.const.u64 %rd118, [kernel_params+2248]; cvta.to.global.u64 %rd119, %rd118; mul.wide.u32 %rd120, %r180, 8; add.s64 %rd121, %rd119, %rd120; ld.global.v2.f32 {%f566, %f567}, [%rd121]; mul.wide.u32 %rd122, %r181, 8; add.s64 %rd123, %rd119, %rd122; ld.global.v2.f32 {%f568, %f569}, [%rd123]; mul.wide.u32 %rd124, %r182, 8; add.s64 %rd125, %rd119, %rd124; ld.global.v2.f32 {%f570, %f571}, [%rd125]; ld.const.u64 %rd126, [kernel_params+2288]; cvta.to.global.u64 %rd127, %rd126; add.s32 %r187, %r2, 5; mul.wide.s32 %rd128, %r187, 4; add.s64 %rd7, %rd127, %rd128; ld.global.u32 %r35, [%rd7+4]; ld.global.u32 %r36, [%rd7]; and.b32 %r37, %r36, 65535; setp.eq.s32 %p46, %r37, 4; @%p46 bra $L__BB14_61; bra.uni $L__BB14_60; $L__BB14_61: add.s32 %r217, %r35, 1; add.s32 %r218, %r35, 2; add.s32 %r219, %r35, 3; mov.u32 %r216, %r35; bra.uni $L__BB14_62; $L__BB14_49: add.s32 %r178, %r49, -1; setp.lt.u32 %p41, %r178, 2; @%p41 bra $L__BB14_54; bra.uni $L__BB14_50; $L__BB14_54: setp.eq.s64 %p44, %rd8, 0; @%p44 bra $L__BB14_56; mov.f32 %f562, 0f00000000; st.v2.f32 [%rd8], {%f562, %f562}; $L__BB14_56: setp.eq.s64 %p45, %rd9, 0; @%p45 bra $L__BB14_58; mov.f32 %f563, 0f00000000; st.v2.f32 [%rd9], {%f563, %f563}; $L__BB14_58: ld.const.u64 %rd110, [kernel_params+2328]; cvta.to.global.u64 %rd111, %rd110; mul.wide.s32 %rd112, %r52, 8; add.s64 %rd113, %rd111, %rd112; ld.global.v2.f32 {%f660, %f661}, [%rd113]; bra.uni $L__BB14_74; $L__BB14_60: ld.global.u32 %r218, [%rd7+8]; shr.u32 %r188, %r36, 16; add.s32 %r216, %r35, %r188; add.s32 %r189, %r188, 1; rem.u32 %r190, %r189, %r37; add.s32 %r191, %r190, %r37; rem.u32 %r192, %r191, %r37; add.s32 %r217, %r192, %r35; add.s32 %r193, %r188, -1; rem.s32 %r194, %r193, %r37; add.s32 %r195, %r194, %r37; rem.s32 %r196, %r195, %r37; add.s32 %r219, %r196, %r35; $L__BB14_62: ld.const.u64 %rd129, [kernel_params+2328]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r197, %r216, %r52; mul.wide.s32 %rd131, %r197, 8; add.s64 %rd132, %rd130, %rd131; ld.global.v2.f32 {%f572, %f573}, [%rd132]; add.s32 %r198, %r217, %r52; mul.wide.s32 %rd133, %r198, 8; add.s64 %rd134, %rd130, %rd133; ld.global.v2.f32 {%f658, %f659}, [%rd134]; add.s32 %r199, %r218, %r52; mul.wide.s32 %rd135, %r199, 8; add.s64 %rd136, %rd130, %rd135; ld.global.v2.f32 {%f576, %f577}, [%rd136]; add.s32 %r200, %r219, %r52; mul.wide.s32 %rd137, %r200, 8; add.s64 %rd138, %rd130, %rd137; ld.global.v2.f32 {%f656, %f657}, [%rd138]; @%p46 bra $L__BB14_64; add.ftz.f32 %f580, %f572, %f658; mul.ftz.f32 %f658, %f580, 0f3F000000; add.ftz.f32 %f581, %f573, %f659; mul.ftz.f32 %f659, %f581, 0f3F000000; add.ftz.f32 %f582, %f572, %f656; mul.ftz.f32 %f656, %f582, 0f3F000000; add.ftz.f32 %f583, %f573, %f657; mul.ftz.f32 %f657, %f583, 0f3F000000; $L__BB14_64: sub.ftz.f32 %f584, %f658, %f572; sub.ftz.f32 %f585, %f659, %f573; fma.rn.ftz.f32 %f586, %f566, %f584, %f572; fma.rn.ftz.f32 %f587, %f566, %f585, %f573; sub.ftz.f32 %f588, %f576, %f656; sub.ftz.f32 %f589, %f577, %f657; fma.rn.ftz.f32 %f590, %f566, %f588, %f656; fma.rn.ftz.f32 %f591, %f566, %f589, %f657; sub.ftz.f32 %f592, %f590, %f586; sub.ftz.f32 %f593, %f591, %f587; fma.rn.ftz.f32 %f94, %f567, %f592, %f586; fma.rn.ftz.f32 %f95, %f567, %f593, %f587; fma.rn.ftz.f32 %f594, %f568, %f584, %f572; fma.rn.ftz.f32 %f595, %f568, %f585, %f573; fma.rn.ftz.f32 %f596, %f568, %f588, %f656; fma.rn.ftz.f32 %f597, %f568, %f589, %f657; sub.ftz.f32 %f598, %f596, %f594; sub.ftz.f32 %f599, %f597, %f595; fma.rn.ftz.f32 %f96, %f569, %f598, %f594; fma.rn.ftz.f32 %f97, %f569, %f599, %f595; fma.rn.ftz.f32 %f600, %f570, %f584, %f572; fma.rn.ftz.f32 %f601, %f570, %f585, %f573; fma.rn.ftz.f32 %f602, %f570, %f588, %f656; fma.rn.ftz.f32 %f603, %f570, %f589, %f657; sub.ftz.f32 %f604, %f602, %f600; sub.ftz.f32 %f605, %f603, %f601; fma.rn.ftz.f32 %f98, %f571, %f604, %f600; fma.rn.ftz.f32 %f99, %f571, %f605, %f601; setp.eq.s64 %p48, %rd8, 0; @%p48 bra $L__BB14_66; ld.local.f32 %f606, [%rd1+40]; ld.local.f32 %f607, [%rd1+48]; mul.ftz.f32 %f608, %f98, %f607; mul.ftz.f32 %f609, %f99, %f607; fma.rn.ftz.f32 %f610, %f96, %f606, %f608; fma.rn.ftz.f32 %f611, %f97, %f606, %f609; add.ftz.f32 %f612, %f607, %f606; mul.ftz.f32 %f613, %f94, %f612; mul.ftz.f32 %f614, %f95, %f612; sub.ftz.f32 %f615, %f611, %f614; sub.ftz.f32 %f616, %f610, %f613; st.v2.f32 [%rd8], {%f616, %f615}; $L__BB14_66: setp.eq.s64 %p49, %rd9, 0; @%p49 bra $L__BB14_68; ld.local.f32 %f617, [%rd1+44]; ld.local.f32 %f618, [%rd1+52]; mul.ftz.f32 %f619, %f98, %f618; mul.ftz.f32 %f620, %f99, %f618; fma.rn.ftz.f32 %f621, %f96, %f617, %f619; fma.rn.ftz.f32 %f622, %f97, %f617, %f620; add.ftz.f32 %f623, %f618, %f617; mul.ftz.f32 %f624, %f94, %f623; mul.ftz.f32 %f625, %f95, %f623; sub.ftz.f32 %f626, %f622, %f625; sub.ftz.f32 %f627, %f621, %f624; st.v2.f32 [%rd9], {%f627, %f626}; $L__BB14_68: ld.local.f32 %f628, [%rd1+8]; ld.local.f32 %f629, [%rd1+12]; mul.ftz.f32 %f630, %f98, %f629; mul.ftz.f32 %f631, %f99, %f629; fma.rn.ftz.f32 %f632, %f96, %f628, %f630; fma.rn.ftz.f32 %f633, %f97, %f628, %f631; mov.f32 %f634, 0f3F800000; sub.ftz.f32 %f635, %f634, %f628; sub.ftz.f32 %f636, %f635, %f629; fma.rn.ftz.f32 %f661, %f95, %f636, %f633; fma.rn.ftz.f32 %f660, %f94, %f636, %f632; bra.uni $L__BB14_74; $L__BB14_50: setp.eq.s64 %p42, %rd8, 0; @%p42 bra $L__BB14_52; mov.f32 %f557, 0f00000000; st.v2.f32 [%rd8], {%f557, %f557}; $L__BB14_52: setp.eq.s64 %p43, %rd9, 0; mov.f32 %f660, 0f00000000; mov.f32 %f661, %f660; @%p43 bra $L__BB14_74; mov.f32 %f660, 0f00000000; st.v2.f32 [%rd9], {%f660, %f660}; mov.f32 %f661, %f660; bra.uni $L__BB14_74; } .func (.param .align 4 .b8 func_retval0[12]) _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_( .param .b64 _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_0, .param .align 4 .b8 _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_1[16], .param .b64 _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_2, .param .b64 _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_3 ) { .reg .pred %p<52>; .reg .b16 %rs<2>; .reg .f32 %f<753>; .reg .b32 %r<226>; .reg .b64 %rd<157>; ld.param.u64 %rd10, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_0]; ld.param.u64 %rd8, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_2]; ld.param.u64 %rd9, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_3]; ld.param.u32 %r52, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_1+12]; ld.param.u32 %r51, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_1+8]; ld.param.u32 %r49, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_1]; cvta.to.local.u64 %rd11, %rd10; add.s64 %rd1, %rd11, 60; ld.local.s32 %rd3, [%rd11+60]; setp.eq.s64 %p1, %rd3, -1; mov.u32 %r2, -1; @%p1 bra $L__BB15_2; ld.const.u64 %rd12, [kernel_params+2240]; cvta.to.global.u64 %rd13, %rd12; shl.b64 %rd14, %rd3, 2; add.s64 %rd15, %rd13, %rd14; ld.global.u32 %r2, [%rd15]; $L__BB15_2: and.b32 %r54, %r51, 2; setp.eq.s32 %p2, %r54, 0; @%p2 bra $L__BB15_38; bra.uni $L__BB15_3; $L__BB15_38: setp.eq.s32 %p34, %r49, 4; @%p34 bra $L__BB15_69; bra.uni $L__BB15_39; $L__BB15_69: setp.eq.s64 %p50, %rd8, 0; @%p50 bra $L__BB15_71; mov.u32 %r205, 0; st.u32 [%rd8], %r205; st.u32 [%rd8+4], %r205; st.u32 [%rd8+8], %r205; $L__BB15_71: setp.eq.s64 %p51, %rd9, 0; @%p51 bra $L__BB15_73; mov.u32 %r206, 0; st.u32 [%rd9], %r206; st.u32 [%rd9+4], %r206; st.u32 [%rd9+8], %r206; $L__BB15_73: ld.const.u64 %rd139, [kernel_params+2336]; cvta.to.global.u64 %rd140, %rd139; ld.const.u64 %rd141, [kernel_params+2288]; cvta.to.global.u64 %rd142, %rd141; add.s32 %r207, %r2, 4; mul.wide.s32 %rd143, %r207, 4; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r208, [%rd144]; add.s32 %r209, %r208, %r52; mul.wide.u32 %rd145, %r209, 12; add.s64 %rd146, %rd140, %rd145; ld.global.f32 %f752, [%rd146+8]; ld.global.f32 %f751, [%rd146+4]; ld.global.f32 %f750, [%rd146]; bra.uni $L__BB15_74; $L__BB15_3: ld.const.u64 %rd16, [kernel_params+2232]; cvta.to.global.u64 %rd17, %rd16; shl.b64 %rd18, %rd3, 4; add.s64 %rd19, %rd17, %rd18; ld.global.v4.u32 {%r56, %r57, %r58, %r59}, [%rd19]; mov.u32 %r212, 0; ld.const.u64 %rd20, [kernel_params+2248]; cvta.to.global.u64 %rd21, %rd20; mul.wide.u32 %rd22, %r56, 8; add.s64 %rd23, %rd21, %rd22; ld.global.v2.f32 {%f136, %f137}, [%rd23]; mul.wide.u32 %rd24, %r57, 8; add.s64 %rd25, %rd21, %rd24; ld.global.v2.f32 {%f140, %f141}, [%rd25]; mul.wide.u32 %rd26, %r58, 8; add.s64 %rd27, %rd21, %rd26; ld.global.v2.f32 {%f144, %f145}, [%rd27]; sub.ftz.f32 %f1, %f140, %f136; sub.ftz.f32 %f2, %f141, %f137; sub.ftz.f32 %f3, %f144, %f136; sub.ftz.f32 %f4, %f145, %f137; ld.local.f32 %f148, [%rd1+8]; ld.local.v2.u32 {%r63, %r64}, [%rd1+12]; mov.b32 %f149, %r63; mul.ftz.f32 %f150, %f3, %f149; mul.ftz.f32 %f151, %f4, %f149; fma.rn.ftz.f32 %f152, %f1, %f148, %f150; fma.rn.ftz.f32 %f153, %f2, %f148, %f151; add.ftz.f32 %f5, %f136, %f152; add.ftz.f32 %f6, %f137, %f153; setp.eq.s32 %p3, %r64, -1; mov.u32 %r211, %r212; @%p3 bra $L__BB15_5; ld.const.u64 %rd28, [kernel_params+2160]; cvta.to.global.u64 %rd29, %rd28; mul.wide.s32 %rd30, %r64, 208; add.s64 %rd31, %rd29, %rd30; ld.global.u32 %r211, [%rd31+160]; $L__BB15_5: add.s32 %r67, %r211, %r2; shr.u32 %r218, %r67, 1; ld.const.u64 %rd32, [kernel_params+2288]; cvta.to.global.u64 %rd6, %rd32; mov.f32 %f727, 0f3F000000; mov.f32 %f731, %f5; mov.f32 %f730, %f6; bra.uni $L__BB15_6; $L__BB15_31: add.s32 %r212, %r212, 4; mul.ftz.f32 %f727, %f22, 0f3F000000; $L__BB15_6: mul.ftz.f32 %f10, %f727, 0f3F000000; setp.lt.ftz.f32 %p4, %f731, %f727; @%p4 bra $L__BB15_8; bra.uni $L__BB15_7; $L__BB15_8: setp.lt.ftz.f32 %p6, %f730, %f727; mov.u32 %r214, 0; @%p6 bra $L__BB15_10; sub.ftz.f32 %f730, %f730, %f727; mov.u32 %r214, 1; bra.uni $L__BB15_10; $L__BB15_7: setp.lt.ftz.f32 %p5, %f730, %f727; sub.ftz.f32 %f155, %f730, %f727; selp.f32 %f730, %f730, %f155, %p5; selp.b32 %r214, 3, 2, %p5; sub.ftz.f32 %f731, %f731, %f727; $L__BB15_10: add.s32 %r73, %r214, %r218; mul.wide.s32 %rd33, %r73, 4; add.s64 %rd34, %rd6, %rd33; ld.global.u32 %r12, [%rd34]; and.b32 %r74, %r12, 1073741824; setp.eq.s32 %p7, %r74, 0; mov.u32 %r219, -1; @%p7 bra $L__BB15_33; and.b32 %r218, %r12, 1073741823; setp.lt.s32 %p8, %r12, 0; @%p8 bra $L__BB15_32; mul.ftz.f32 %f16, %f10, 0f3F000000; setp.lt.ftz.f32 %p9, %f731, %f10; @%p9 bra $L__BB15_14; bra.uni $L__BB15_13; $L__BB15_14: setp.lt.ftz.f32 %p11, %f730, %f10; mov.u32 %r215, 0; @%p11 bra $L__BB15_16; sub.ftz.f32 %f730, %f730, %f10; mov.u32 %r215, 1; bra.uni $L__BB15_16; $L__BB15_13: setp.lt.ftz.f32 %p10, %f730, %f10; sub.ftz.f32 %f156, %f730, %f10; selp.f32 %f730, %f730, %f156, %p10; selp.b32 %r215, 3, 2, %p10; sub.ftz.f32 %f731, %f731, %f10; $L__BB15_16: add.s32 %r82, %r215, %r218; mul.wide.u32 %rd35, %r82, 4; add.s64 %rd36, %rd6, %rd35; ld.global.u32 %r16, [%rd36]; and.b32 %r83, %r16, 1073741824; setp.eq.s32 %p12, %r83, 0; @%p12 bra $L__BB15_33; and.b32 %r218, %r16, 1073741823; setp.lt.s32 %p13, %r16, 0; @%p13 bra $L__BB15_32; mul.ftz.f32 %f22, %f16, 0f3F000000; setp.lt.ftz.f32 %p14, %f731, %f16; @%p14 bra $L__BB15_20; bra.uni $L__BB15_19; $L__BB15_20: setp.lt.ftz.f32 %p16, %f730, %f16; mov.u32 %r216, 0; @%p16 bra $L__BB15_22; sub.ftz.f32 %f730, %f730, %f16; mov.u32 %r216, 1; bra.uni $L__BB15_22; $L__BB15_19: setp.lt.ftz.f32 %p15, %f730, %f16; sub.ftz.f32 %f157, %f730, %f16; selp.f32 %f730, %f730, %f157, %p15; selp.b32 %r216, 3, 2, %p15; sub.ftz.f32 %f731, %f731, %f16; $L__BB15_22: add.s32 %r91, %r216, %r218; mul.wide.u32 %rd37, %r91, 4; add.s64 %rd38, %rd6, %rd37; ld.global.u32 %r20, [%rd38]; and.b32 %r92, %r20, 1073741824; setp.eq.s32 %p17, %r92, 0; @%p17 bra $L__BB15_33; and.b32 %r218, %r20, 1073741823; setp.lt.s32 %p18, %r20, 0; @%p18 bra $L__BB15_32; add.s32 %r98, %r212, 3; setp.gt.u32 %p19, %r98, 254; @%p19 bra $L__BB15_33; setp.lt.ftz.f32 %p20, %f731, %f22; @%p20 bra $L__BB15_27; bra.uni $L__BB15_26; $L__BB15_27: setp.lt.ftz.f32 %p22, %f730, %f22; mov.u32 %r217, 0; @%p22 bra $L__BB15_29; sub.ftz.f32 %f730, %f730, %f22; mov.u32 %r217, 1; bra.uni $L__BB15_29; $L__BB15_26: setp.lt.ftz.f32 %p21, %f730, %f22; sub.ftz.f32 %f158, %f730, %f22; selp.f32 %f730, %f730, %f158, %p21; selp.b32 %r217, 3, 2, %p21; sub.ftz.f32 %f731, %f731, %f22; $L__BB15_29: add.s32 %r106, %r217, %r218; mul.wide.u32 %rd39, %r106, 4; add.s64 %rd40, %rd6, %rd39; ld.global.u32 %r24, [%rd40]; and.b32 %r107, %r24, 1073741824; setp.eq.s32 %p23, %r107, 0; @%p23 bra $L__BB15_33; and.b32 %r218, %r24, 1073741823; setp.lt.s32 %p24, %r24, 0; @%p24 bra $L__BB15_32; bra.uni $L__BB15_31; $L__BB15_32: mul.wide.u32 %rd41, %r218, 4; add.s64 %rd42, %rd6, %rd41; ld.global.u32 %r221, [%rd42+8]; ld.global.u32 %r220, [%rd42+4]; ld.global.u32 %r219, [%rd42]; $L__BB15_33: add.s32 %r110, %r219, 2; mul.wide.s32 %rd43, %r110, 4; add.s64 %rd44, %rd6, %rd43; ld.global.u32 %r111, [%rd44]; add.s32 %r112, %r111, %r221; mul.wide.s32 %rd45, %r112, 4; add.s64 %rd46, %rd6, %rd45; add.s32 %r113, %r220, 1; mov.u32 %r114, 1; mul.wide.s32 %rd47, %r113, 4; add.s64 %rd48, %rd6, %rd47; ld.global.u32 %r115, [%rd48]; and.b32 %r116, %r115, 15; shl.b32 %r117, %r114, %r116; cvt.rn.f32.s32 %f159, %r117; and.b32 %r118, %r115, 16; setp.eq.s32 %p25, %r118, 0; mul.ftz.f32 %f160, %f159, 0f3F000000; selp.f32 %f161, %f159, %f160, %p25; shr.u32 %r119, %r115, 4; and.b32 %r120, %r119, 1; setp.eq.b32 %p26, %r120, 1; selp.b32 %r121, -1, 0, %p26; add.s32 %r122, %r116, %r121; shl.b32 %r123, %r114, %r122; cvt.rn.f32.s32 %f162, %r123; rcp.approx.ftz.f32 %f163, %f162; mov.f32 %f164, 0f3F800000; shr.u32 %r124, %r115, 22; bfe.u32 %r125, %r115, 12, 10; cvt.rn.f32.s32 %f165, %r124; mul.ftz.f32 %f166, %f163, %f165; cvt.rn.f32.s32 %f167, %r125; mul.ftz.f32 %f168, %f163, %f167; sub.ftz.f32 %f169, %f5, %f166; div.approx.ftz.f32 %f170, %f169, %f163; sub.ftz.f32 %f171, %f6, %f168; div.approx.ftz.f32 %f172, %f171, %f163; mul.ftz.f32 %f173, %f170, %f170; mul.ftz.f32 %f174, %f170, %f173; sub.ftz.f32 %f175, %f170, %f173; mul.ftz.f32 %f176, %f175, 0f40400000; sub.ftz.f32 %f177, %f164, %f176; sub.ftz.f32 %f178, %f177, %f174; mul.ftz.f32 %f179, %f178, 0f3E2AAAAB; mul.ftz.f32 %f180, %f173, 0f40C00000; mov.f32 %f181, 0f40800000; sub.ftz.f32 %f182, %f181, %f180; fma.rn.ftz.f32 %f183, %f174, 0f40400000, %f182; mul.ftz.f32 %f184, %f183, 0f3E2AAAAB; add.ftz.f32 %f185, %f170, %f173; sub.ftz.f32 %f186, %f185, %f174; fma.rn.ftz.f32 %f187, %f186, 0f40400000, 0f3F800000; mul.ftz.f32 %f188, %f187, 0f3E2AAAAB; mul.ftz.f32 %f189, %f174, 0f3E2AAAAB; mul.ftz.f32 %f190, %f173, 0f3F000000; sub.ftz.f32 %f191, %f170, %f190; add.ftz.f32 %f192, %f191, 0fBF000000; mul.ftz.f32 %f193, %f173, 0f3FC00000; add.ftz.f32 %f194, %f170, %f170; sub.ftz.f32 %f195, %f193, %f194; sub.ftz.f32 %f196, %f170, %f193; add.ftz.f32 %f197, %f196, 0f3F000000; mul.ftz.f32 %f198, %f172, %f172; mul.ftz.f32 %f199, %f172, %f198; sub.ftz.f32 %f200, %f172, %f198; mul.ftz.f32 %f201, %f200, 0f40400000; sub.ftz.f32 %f202, %f164, %f201; sub.ftz.f32 %f203, %f202, %f199; mul.ftz.f32 %f204, %f203, 0f3E2AAAAB; mul.ftz.f32 %f205, %f198, 0f40C00000; sub.ftz.f32 %f206, %f181, %f205; fma.rn.ftz.f32 %f207, %f199, 0f40400000, %f206; mul.ftz.f32 %f208, %f207, 0f3E2AAAAB; add.ftz.f32 %f209, %f172, %f198; sub.ftz.f32 %f210, %f209, %f199; fma.rn.ftz.f32 %f211, %f210, 0f40400000, 0f3F800000; mul.ftz.f32 %f212, %f211, 0f3E2AAAAB; mul.ftz.f32 %f213, %f199, 0f3E2AAAAB; mul.ftz.f32 %f214, %f198, 0f3F000000; sub.ftz.f32 %f215, %f172, %f214; add.ftz.f32 %f216, %f215, 0fBF000000; mul.ftz.f32 %f217, %f198, 0f3FC00000; add.ftz.f32 %f218, %f172, %f172; sub.ftz.f32 %f219, %f217, %f218; sub.ftz.f32 %f220, %f172, %f217; add.ftz.f32 %f221, %f220, 0f3F000000; and.b32 %r126, %r115, 256; setp.eq.s32 %p27, %r126, 0; sub.ftz.f32 %f222, %f212, %f204; fma.rn.ftz.f32 %f223, %f204, 0f40000000, %f208; selp.f32 %f224, %f204, 0f00000000, %p27; selp.f32 %f225, %f208, %f223, %p27; selp.f32 %f226, %f212, %f222, %p27; and.b32 %r127, %r115, 512; setp.eq.s32 %p28, %r127, 0; sub.ftz.f32 %f227, %f184, %f189; fma.rn.ftz.f32 %f228, %f189, 0f40000000, %f188; selp.f32 %f229, %f184, %f227, %p28; selp.f32 %f230, %f188, %f228, %p28; selp.f32 %f231, %f189, 0f00000000, %p28; and.b32 %r128, %r115, 1024; setp.eq.s32 %p29, %r128, 0; sub.ftz.f32 %f232, %f225, %f213; fma.rn.ftz.f32 %f233, %f213, 0f40000000, %f226; selp.f32 %f234, %f225, %f232, %p29; selp.f32 %f235, %f226, %f233, %p29; selp.f32 %f236, %f213, 0f00000000, %p29; and.b32 %r129, %r115, 2048; setp.eq.s32 %p30, %r129, 0; sub.ftz.f32 %f237, %f230, %f179; fma.rn.ftz.f32 %f238, %f179, 0f40000000, %f229; selp.f32 %f239, %f179, 0f00000000, %p30; selp.f32 %f240, %f229, %f238, %p30; selp.f32 %f241, %f230, %f237, %p30; sub.ftz.f32 %f242, %f221, %f216; fma.rn.ftz.f32 %f243, %f216, 0f40000000, %f219; selp.f32 %f244, %f216, 0f00000000, %p27; selp.f32 %f245, %f219, %f243, %p27; selp.f32 %f246, %f221, %f242, %p27; sub.ftz.f32 %f247, %f195, %f190; fma.rn.ftz.f32 %f248, %f190, 0f40000000, %f197; selp.f32 %f249, %f195, %f247, %p28; selp.f32 %f250, %f197, %f248, %p28; selp.f32 %f251, %f190, 0f00000000, %p28; sub.ftz.f32 %f252, %f245, %f214; fma.rn.ftz.f32 %f253, %f214, 0f40000000, %f246; selp.f32 %f254, %f245, %f252, %p29; selp.f32 %f255, %f246, %f253, %p29; selp.f32 %f256, %f214, 0f00000000, %p29; sub.ftz.f32 %f257, %f250, %f192; fma.rn.ftz.f32 %f258, %f192, 0f40000000, %f249; selp.f32 %f259, %f192, 0f00000000, %p30; selp.f32 %f260, %f249, %f258, %p30; selp.f32 %f261, %f250, %f257, %p30; mul.ftz.f32 %f262, %f239, %f224; mul.ftz.f32 %f263, %f259, %f224; mul.ftz.f32 %f264, %f161, %f263; mul.ftz.f32 %f265, %f239, %f244; mul.ftz.f32 %f266, %f161, %f265; mul.ftz.f32 %f267, %f240, %f224; mul.ftz.f32 %f268, %f260, %f224; mul.ftz.f32 %f269, %f161, %f268; mul.ftz.f32 %f270, %f240, %f244; mul.ftz.f32 %f271, %f161, %f270; mul.ftz.f32 %f272, %f241, %f224; mul.ftz.f32 %f273, %f261, %f224; mul.ftz.f32 %f274, %f161, %f273; mul.ftz.f32 %f275, %f241, %f244; mul.ftz.f32 %f276, %f161, %f275; mul.ftz.f32 %f277, %f231, %f224; mul.ftz.f32 %f278, %f251, %f224; mul.ftz.f32 %f279, %f161, %f278; mul.ftz.f32 %f280, %f231, %f244; mul.ftz.f32 %f281, %f161, %f280; mul.ftz.f32 %f282, %f239, %f234; mul.ftz.f32 %f283, %f259, %f234; mul.ftz.f32 %f284, %f161, %f283; mul.ftz.f32 %f285, %f239, %f254; mul.ftz.f32 %f286, %f161, %f285; mul.ftz.f32 %f287, %f240, %f234; mul.ftz.f32 %f288, %f260, %f234; mul.ftz.f32 %f289, %f161, %f288; mul.ftz.f32 %f290, %f240, %f254; mul.ftz.f32 %f291, %f161, %f290; mul.ftz.f32 %f292, %f241, %f234; mul.ftz.f32 %f293, %f261, %f234; mul.ftz.f32 %f294, %f161, %f293; mul.ftz.f32 %f295, %f241, %f254; mul.ftz.f32 %f296, %f161, %f295; mul.ftz.f32 %f297, %f231, %f234; mul.ftz.f32 %f298, %f251, %f234; mul.ftz.f32 %f299, %f161, %f298; mul.ftz.f32 %f300, %f231, %f254; mul.ftz.f32 %f301, %f161, %f300; mul.ftz.f32 %f302, %f239, %f235; mul.ftz.f32 %f303, %f259, %f235; mul.ftz.f32 %f304, %f161, %f303; mul.ftz.f32 %f305, %f239, %f255; mul.ftz.f32 %f306, %f161, %f305; mul.ftz.f32 %f307, %f240, %f235; mul.ftz.f32 %f308, %f260, %f235; mul.ftz.f32 %f309, %f161, %f308; mul.ftz.f32 %f310, %f240, %f255; mul.ftz.f32 %f311, %f161, %f310; mul.ftz.f32 %f312, %f241, %f235; mul.ftz.f32 %f313, %f261, %f235; mul.ftz.f32 %f314, %f161, %f313; mul.ftz.f32 %f315, %f241, %f255; mul.ftz.f32 %f316, %f161, %f315; mul.ftz.f32 %f317, %f231, %f235; mul.ftz.f32 %f318, %f251, %f235; mul.ftz.f32 %f319, %f161, %f318; mul.ftz.f32 %f320, %f231, %f255; mul.ftz.f32 %f321, %f161, %f320; mul.ftz.f32 %f322, %f239, %f236; mul.ftz.f32 %f323, %f259, %f236; mul.ftz.f32 %f324, %f161, %f323; mul.ftz.f32 %f325, %f239, %f256; mul.ftz.f32 %f326, %f161, %f325; mul.ftz.f32 %f327, %f240, %f236; mul.ftz.f32 %f328, %f260, %f236; mul.ftz.f32 %f329, %f161, %f328; mul.ftz.f32 %f330, %f240, %f256; mul.ftz.f32 %f331, %f161, %f330; mul.ftz.f32 %f332, %f241, %f236; mul.ftz.f32 %f333, %f261, %f236; mul.ftz.f32 %f334, %f161, %f333; mul.ftz.f32 %f335, %f241, %f256; mul.ftz.f32 %f336, %f161, %f335; mul.ftz.f32 %f337, %f231, %f236; mul.ftz.f32 %f338, %f251, %f236; mul.ftz.f32 %f339, %f161, %f338; mul.ftz.f32 %f340, %f231, %f256; mul.ftz.f32 %f341, %f161, %f340; ld.const.u64 %rd49, [kernel_params+2336]; cvta.to.global.u64 %rd50, %rd49; ld.global.u32 %r130, [%rd46]; add.s32 %r131, %r130, %r52; mul.wide.s32 %rd51, %r131, 12; add.s64 %rd52, %rd50, %rd51; ld.global.f32 %f342, [%rd52]; ld.global.f32 %f343, [%rd52+4]; ld.global.f32 %f344, [%rd52+8]; fma.rn.ftz.f32 %f345, %f342, %f262, 0f00000000; fma.rn.ftz.f32 %f346, %f343, %f262, 0f00000000; fma.rn.ftz.f32 %f347, %f344, %f262, 0f00000000; fma.rn.ftz.f32 %f348, %f342, %f264, 0f00000000; fma.rn.ftz.f32 %f349, %f343, %f264, 0f00000000; fma.rn.ftz.f32 %f350, %f344, %f264, 0f00000000; fma.rn.ftz.f32 %f351, %f342, %f266, 0f00000000; fma.rn.ftz.f32 %f352, %f343, %f266, 0f00000000; fma.rn.ftz.f32 %f353, %f344, %f266, 0f00000000; ld.global.u32 %r132, [%rd46+4]; add.s32 %r133, %r132, %r52; mul.wide.s32 %rd53, %r133, 12; add.s64 %rd54, %rd50, %rd53; ld.global.f32 %f354, [%rd54]; ld.global.f32 %f355, [%rd54+4]; ld.global.f32 %f356, [%rd54+8]; fma.rn.ftz.f32 %f357, %f354, %f267, %f345; fma.rn.ftz.f32 %f358, %f355, %f267, %f346; fma.rn.ftz.f32 %f359, %f356, %f267, %f347; fma.rn.ftz.f32 %f360, %f354, %f269, %f348; fma.rn.ftz.f32 %f361, %f355, %f269, %f349; fma.rn.ftz.f32 %f362, %f356, %f269, %f350; fma.rn.ftz.f32 %f363, %f354, %f271, %f351; fma.rn.ftz.f32 %f364, %f355, %f271, %f352; fma.rn.ftz.f32 %f365, %f356, %f271, %f353; ld.global.u32 %r134, [%rd46+8]; add.s32 %r135, %r134, %r52; mul.wide.s32 %rd55, %r135, 12; add.s64 %rd56, %rd50, %rd55; ld.global.f32 %f366, [%rd56]; ld.global.f32 %f367, [%rd56+4]; ld.global.f32 %f368, [%rd56+8]; fma.rn.ftz.f32 %f369, %f366, %f272, %f357; fma.rn.ftz.f32 %f370, %f367, %f272, %f358; fma.rn.ftz.f32 %f371, %f368, %f272, %f359; fma.rn.ftz.f32 %f372, %f366, %f274, %f360; fma.rn.ftz.f32 %f373, %f367, %f274, %f361; fma.rn.ftz.f32 %f374, %f368, %f274, %f362; fma.rn.ftz.f32 %f375, %f366, %f276, %f363; fma.rn.ftz.f32 %f376, %f367, %f276, %f364; fma.rn.ftz.f32 %f377, %f368, %f276, %f365; ld.global.u32 %r136, [%rd46+12]; add.s32 %r137, %r136, %r52; mul.wide.s32 %rd57, %r137, 12; add.s64 %rd58, %rd50, %rd57; ld.global.f32 %f378, [%rd58]; ld.global.f32 %f379, [%rd58+4]; ld.global.f32 %f380, [%rd58+8]; fma.rn.ftz.f32 %f381, %f378, %f277, %f369; fma.rn.ftz.f32 %f382, %f379, %f277, %f370; fma.rn.ftz.f32 %f383, %f380, %f277, %f371; fma.rn.ftz.f32 %f384, %f378, %f279, %f372; fma.rn.ftz.f32 %f385, %f379, %f279, %f373; fma.rn.ftz.f32 %f386, %f380, %f279, %f374; fma.rn.ftz.f32 %f387, %f378, %f281, %f375; fma.rn.ftz.f32 %f388, %f379, %f281, %f376; fma.rn.ftz.f32 %f389, %f380, %f281, %f377; ld.global.u32 %r138, [%rd46+16]; add.s32 %r139, %r138, %r52; mul.wide.s32 %rd59, %r139, 12; add.s64 %rd60, %rd50, %rd59; ld.global.f32 %f390, [%rd60]; ld.global.f32 %f391, [%rd60+4]; ld.global.f32 %f392, [%rd60+8]; fma.rn.ftz.f32 %f393, %f390, %f282, %f381; fma.rn.ftz.f32 %f394, %f391, %f282, %f382; fma.rn.ftz.f32 %f395, %f392, %f282, %f383; fma.rn.ftz.f32 %f396, %f390, %f284, %f384; fma.rn.ftz.f32 %f397, %f391, %f284, %f385; fma.rn.ftz.f32 %f398, %f392, %f284, %f386; fma.rn.ftz.f32 %f399, %f390, %f286, %f387; fma.rn.ftz.f32 %f400, %f391, %f286, %f388; fma.rn.ftz.f32 %f401, %f392, %f286, %f389; ld.global.u32 %r140, [%rd46+20]; add.s32 %r141, %r140, %r52; mul.wide.s32 %rd61, %r141, 12; add.s64 %rd62, %rd50, %rd61; ld.global.f32 %f402, [%rd62]; ld.global.f32 %f403, [%rd62+4]; ld.global.f32 %f404, [%rd62+8]; fma.rn.ftz.f32 %f405, %f402, %f287, %f393; fma.rn.ftz.f32 %f406, %f403, %f287, %f394; fma.rn.ftz.f32 %f407, %f404, %f287, %f395; fma.rn.ftz.f32 %f408, %f402, %f289, %f396; fma.rn.ftz.f32 %f409, %f403, %f289, %f397; fma.rn.ftz.f32 %f410, %f404, %f289, %f398; fma.rn.ftz.f32 %f411, %f402, %f291, %f399; fma.rn.ftz.f32 %f412, %f403, %f291, %f400; fma.rn.ftz.f32 %f413, %f404, %f291, %f401; ld.global.u32 %r142, [%rd46+24]; add.s32 %r143, %r142, %r52; mul.wide.s32 %rd63, %r143, 12; add.s64 %rd64, %rd50, %rd63; ld.global.f32 %f414, [%rd64]; ld.global.f32 %f415, [%rd64+4]; ld.global.f32 %f416, [%rd64+8]; fma.rn.ftz.f32 %f417, %f414, %f292, %f405; fma.rn.ftz.f32 %f418, %f415, %f292, %f406; fma.rn.ftz.f32 %f419, %f416, %f292, %f407; fma.rn.ftz.f32 %f420, %f414, %f294, %f408; fma.rn.ftz.f32 %f421, %f415, %f294, %f409; fma.rn.ftz.f32 %f422, %f416, %f294, %f410; fma.rn.ftz.f32 %f423, %f414, %f296, %f411; fma.rn.ftz.f32 %f424, %f415, %f296, %f412; fma.rn.ftz.f32 %f425, %f416, %f296, %f413; ld.global.u32 %r144, [%rd46+28]; add.s32 %r145, %r144, %r52; mul.wide.s32 %rd65, %r145, 12; add.s64 %rd66, %rd50, %rd65; ld.global.f32 %f426, [%rd66]; ld.global.f32 %f427, [%rd66+4]; ld.global.f32 %f428, [%rd66+8]; fma.rn.ftz.f32 %f429, %f426, %f297, %f417; fma.rn.ftz.f32 %f430, %f427, %f297, %f418; fma.rn.ftz.f32 %f431, %f428, %f297, %f419; fma.rn.ftz.f32 %f432, %f426, %f299, %f420; fma.rn.ftz.f32 %f433, %f427, %f299, %f421; fma.rn.ftz.f32 %f434, %f428, %f299, %f422; fma.rn.ftz.f32 %f435, %f426, %f301, %f423; fma.rn.ftz.f32 %f436, %f427, %f301, %f424; fma.rn.ftz.f32 %f437, %f428, %f301, %f425; ld.global.u32 %r146, [%rd46+32]; add.s32 %r147, %r146, %r52; mul.wide.s32 %rd67, %r147, 12; add.s64 %rd68, %rd50, %rd67; ld.global.f32 %f438, [%rd68]; ld.global.f32 %f439, [%rd68+4]; ld.global.f32 %f440, [%rd68+8]; fma.rn.ftz.f32 %f441, %f438, %f302, %f429; fma.rn.ftz.f32 %f442, %f439, %f302, %f430; fma.rn.ftz.f32 %f443, %f440, %f302, %f431; fma.rn.ftz.f32 %f444, %f438, %f304, %f432; fma.rn.ftz.f32 %f445, %f439, %f304, %f433; fma.rn.ftz.f32 %f446, %f440, %f304, %f434; fma.rn.ftz.f32 %f447, %f438, %f306, %f435; fma.rn.ftz.f32 %f448, %f439, %f306, %f436; fma.rn.ftz.f32 %f449, %f440, %f306, %f437; ld.global.u32 %r148, [%rd46+36]; add.s32 %r149, %r148, %r52; mul.wide.s32 %rd69, %r149, 12; add.s64 %rd70, %rd50, %rd69; ld.global.f32 %f450, [%rd70]; ld.global.f32 %f451, [%rd70+4]; ld.global.f32 %f452, [%rd70+8]; fma.rn.ftz.f32 %f453, %f450, %f307, %f441; fma.rn.ftz.f32 %f454, %f451, %f307, %f442; fma.rn.ftz.f32 %f455, %f452, %f307, %f443; fma.rn.ftz.f32 %f456, %f450, %f309, %f444; fma.rn.ftz.f32 %f457, %f451, %f309, %f445; fma.rn.ftz.f32 %f458, %f452, %f309, %f446; fma.rn.ftz.f32 %f459, %f450, %f311, %f447; fma.rn.ftz.f32 %f460, %f451, %f311, %f448; fma.rn.ftz.f32 %f461, %f452, %f311, %f449; ld.global.u32 %r150, [%rd46+40]; add.s32 %r151, %r150, %r52; mul.wide.s32 %rd71, %r151, 12; add.s64 %rd72, %rd50, %rd71; ld.global.f32 %f462, [%rd72]; ld.global.f32 %f463, [%rd72+4]; ld.global.f32 %f464, [%rd72+8]; fma.rn.ftz.f32 %f465, %f462, %f312, %f453; fma.rn.ftz.f32 %f466, %f463, %f312, %f454; fma.rn.ftz.f32 %f467, %f464, %f312, %f455; fma.rn.ftz.f32 %f468, %f462, %f314, %f456; fma.rn.ftz.f32 %f469, %f463, %f314, %f457; fma.rn.ftz.f32 %f470, %f464, %f314, %f458; fma.rn.ftz.f32 %f471, %f462, %f316, %f459; fma.rn.ftz.f32 %f472, %f463, %f316, %f460; fma.rn.ftz.f32 %f473, %f464, %f316, %f461; ld.global.u32 %r152, [%rd46+44]; add.s32 %r153, %r152, %r52; mul.wide.s32 %rd73, %r153, 12; add.s64 %rd74, %rd50, %rd73; ld.global.f32 %f474, [%rd74]; ld.global.f32 %f475, [%rd74+4]; ld.global.f32 %f476, [%rd74+8]; fma.rn.ftz.f32 %f477, %f474, %f317, %f465; fma.rn.ftz.f32 %f478, %f475, %f317, %f466; fma.rn.ftz.f32 %f479, %f476, %f317, %f467; fma.rn.ftz.f32 %f480, %f474, %f319, %f468; fma.rn.ftz.f32 %f481, %f475, %f319, %f469; fma.rn.ftz.f32 %f482, %f476, %f319, %f470; fma.rn.ftz.f32 %f483, %f474, %f321, %f471; fma.rn.ftz.f32 %f484, %f475, %f321, %f472; fma.rn.ftz.f32 %f485, %f476, %f321, %f473; ld.global.u32 %r154, [%rd46+48]; add.s32 %r155, %r154, %r52; mul.wide.s32 %rd75, %r155, 12; add.s64 %rd76, %rd50, %rd75; ld.global.f32 %f486, [%rd76]; ld.global.f32 %f487, [%rd76+4]; ld.global.f32 %f488, [%rd76+8]; fma.rn.ftz.f32 %f489, %f486, %f322, %f477; fma.rn.ftz.f32 %f490, %f487, %f322, %f478; fma.rn.ftz.f32 %f491, %f488, %f322, %f479; fma.rn.ftz.f32 %f492, %f486, %f324, %f480; fma.rn.ftz.f32 %f493, %f487, %f324, %f481; fma.rn.ftz.f32 %f494, %f488, %f324, %f482; fma.rn.ftz.f32 %f495, %f486, %f326, %f483; fma.rn.ftz.f32 %f496, %f487, %f326, %f484; fma.rn.ftz.f32 %f497, %f488, %f326, %f485; ld.global.u32 %r156, [%rd46+52]; add.s32 %r157, %r156, %r52; mul.wide.s32 %rd77, %r157, 12; add.s64 %rd78, %rd50, %rd77; ld.global.f32 %f498, [%rd78]; ld.global.f32 %f499, [%rd78+4]; ld.global.f32 %f500, [%rd78+8]; fma.rn.ftz.f32 %f501, %f498, %f327, %f489; fma.rn.ftz.f32 %f502, %f499, %f327, %f490; fma.rn.ftz.f32 %f503, %f500, %f327, %f491; fma.rn.ftz.f32 %f504, %f498, %f329, %f492; fma.rn.ftz.f32 %f505, %f499, %f329, %f493; fma.rn.ftz.f32 %f506, %f500, %f329, %f494; fma.rn.ftz.f32 %f507, %f498, %f331, %f495; fma.rn.ftz.f32 %f508, %f499, %f331, %f496; fma.rn.ftz.f32 %f509, %f500, %f331, %f497; ld.global.u32 %r158, [%rd46+56]; add.s32 %r159, %r158, %r52; mul.wide.s32 %rd79, %r159, 12; add.s64 %rd80, %rd50, %rd79; ld.global.f32 %f510, [%rd80]; ld.global.f32 %f511, [%rd80+4]; ld.global.f32 %f512, [%rd80+8]; fma.rn.ftz.f32 %f513, %f510, %f332, %f501; fma.rn.ftz.f32 %f514, %f511, %f332, %f502; fma.rn.ftz.f32 %f515, %f512, %f332, %f503; fma.rn.ftz.f32 %f516, %f510, %f334, %f504; fma.rn.ftz.f32 %f517, %f511, %f334, %f505; fma.rn.ftz.f32 %f518, %f512, %f334, %f506; fma.rn.ftz.f32 %f519, %f510, %f336, %f507; fma.rn.ftz.f32 %f520, %f511, %f336, %f508; fma.rn.ftz.f32 %f521, %f512, %f336, %f509; ld.global.u32 %r160, [%rd46+60]; add.s32 %r161, %r160, %r52; mul.wide.s32 %rd81, %r161, 12; add.s64 %rd82, %rd50, %rd81; ld.global.f32 %f522, [%rd82]; ld.global.f32 %f523, [%rd82+4]; ld.global.f32 %f524, [%rd82+8]; fma.rn.ftz.f32 %f34, %f522, %f339, %f516; fma.rn.ftz.f32 %f35, %f523, %f339, %f517; fma.rn.ftz.f32 %f36, %f524, %f339, %f518; fma.rn.ftz.f32 %f37, %f522, %f341, %f519; fma.rn.ftz.f32 %f38, %f523, %f341, %f520; fma.rn.ftz.f32 %f39, %f524, %f341, %f521; fma.rn.ftz.f32 %f752, %f524, %f337, %f515; fma.rn.ftz.f32 %f751, %f523, %f337, %f514; fma.rn.ftz.f32 %f750, %f522, %f337, %f513; or.b64 %rd83, %rd8, %rd9; setp.eq.s64 %p31, %rd83, 0; @%p31 bra $L__BB15_74; ld.param.u64 %rd147, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_2]; setp.eq.s64 %p32, %rd147, 0; @%p32 bra $L__BB15_36; ld.param.u64 %rd153, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_0]; cvta.to.local.u64 %rd152, %rd153; add.s64 %rd151, %rd152, 60; ld.param.u64 %rd148, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_2]; ld.local.f32 %f525, [%rd151+40]; ld.local.f32 %f526, [%rd151+48]; mul.ftz.f32 %f527, %f3, %f526; fma.rn.ftz.f32 %f528, %f1, %f525, %f527; mul.ftz.f32 %f529, %f4, %f526; fma.rn.ftz.f32 %f530, %f2, %f525, %f529; mul.ftz.f32 %f531, %f37, %f530; mul.ftz.f32 %f532, %f38, %f530; mul.ftz.f32 %f533, %f39, %f530; fma.rn.ftz.f32 %f534, %f34, %f528, %f531; fma.rn.ftz.f32 %f535, %f35, %f528, %f532; fma.rn.ftz.f32 %f536, %f36, %f528, %f533; st.f32 [%rd148], %f534; st.f32 [%rd148+4], %f535; st.f32 [%rd148+8], %f536; $L__BB15_36: ld.param.u64 %rd149, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_3]; setp.eq.s64 %p33, %rd149, 0; @%p33 bra $L__BB15_74; ld.param.u64 %rd156, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_0]; cvta.to.local.u64 %rd155, %rd156; add.s64 %rd154, %rd155, 60; ld.param.u64 %rd150, [_Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7__param_3]; ld.local.f32 %f537, [%rd154+44]; ld.local.f32 %f538, [%rd154+52]; mul.ftz.f32 %f539, %f3, %f538; fma.rn.ftz.f32 %f540, %f1, %f537, %f539; mul.ftz.f32 %f541, %f4, %f538; fma.rn.ftz.f32 %f542, %f2, %f537, %f541; mul.ftz.f32 %f543, %f37, %f542; mul.ftz.f32 %f544, %f38, %f542; mul.ftz.f32 %f545, %f39, %f542; fma.rn.ftz.f32 %f546, %f34, %f540, %f543; fma.rn.ftz.f32 %f547, %f35, %f540, %f544; fma.rn.ftz.f32 %f548, %f36, %f540, %f545; st.f32 [%rd150], %f546; st.f32 [%rd150+4], %f547; st.f32 [%rd150+8], %f548; bra.uni $L__BB15_74; $L__BB15_39: setp.eq.s32 %p35, %r49, 16; @%p35 bra $L__BB15_41; setp.ne.s32 %p36, %r49, 8; @%p36 bra $L__BB15_48; $L__BB15_41: cvt.u32.u64 %r162, %rd3; ld.const.u64 %rd84, [kernel_params+2232]; cvta.to.global.u64 %rd85, %rd84; mul.wide.s32 %rd86, %r162, 16; add.s64 %rd87, %rd85, %rd86; ld.global.v4.u32 {%r163, %r164, %r165, %r166}, [%rd87]; ld.const.u64 %rd88, [kernel_params+2248]; cvta.to.global.u64 %rd89, %rd88; mul.wide.u32 %rd90, %r163, 8; add.s64 %rd91, %rd89, %rd90; ld.global.v2.f32 {%f549, %f550}, [%rd91]; mul.wide.u32 %rd92, %r164, 8; add.s64 %rd93, %rd89, %rd92; ld.global.v2.f32 {%f551, %f552}, [%rd93]; mul.wide.u32 %rd94, %r165, 8; add.s64 %rd95, %rd89, %rd94; ld.global.v2.f32 {%f553, %f554}, [%rd95]; ld.const.u64 %rd96, [kernel_params+2288]; cvta.to.global.u64 %rd97, %rd96; mul.wide.s32 %rd98, %r2, 4; add.s64 %rd99, %rd97, %rd98; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; ld.global.u32 %r170, [%rd99]; add.s32 %r171, %r170, %r52; mul.wide.u32 %rd102, %r171, 12; add.s64 %rd103, %rd101, %rd102; ld.global.f32 %f49, [%rd103]; ld.global.f32 %f50, [%rd103+4]; ld.global.f32 %f51, [%rd103+8]; ld.global.u32 %r172, [%rd99+4]; add.s32 %r173, %r172, %r52; mul.wide.u32 %rd104, %r173, 12; add.s64 %rd105, %rd101, %rd104; ld.global.f32 %f741, [%rd105]; ld.global.f32 %f742, [%rd105+4]; ld.global.f32 %f743, [%rd105+8]; ld.global.u32 %r174, [%rd99+8]; add.s32 %r175, %r174, %r52; mul.wide.u32 %rd106, %r175, 12; add.s64 %rd107, %rd101, %rd106; ld.global.f32 %f55, [%rd107]; ld.global.f32 %f56, [%rd107+4]; ld.global.f32 %f57, [%rd107+8]; ld.global.u32 %r176, [%rd99+12]; add.s32 %r177, %r176, %r52; mul.wide.u32 %rd108, %r177, 12; add.s64 %rd109, %rd101, %rd108; ld.global.f32 %f738, [%rd109]; ld.global.f32 %f739, [%rd109+4]; ld.global.f32 %f740, [%rd109+8]; ld.global.u16 %rs1, [%rd99+20]; setp.eq.s16 %p37, %rs1, 4; @%p37 bra $L__BB15_43; add.ftz.f32 %f555, %f49, %f741; mul.ftz.f32 %f741, %f555, 0f3F000000; add.ftz.f32 %f556, %f50, %f742; mul.ftz.f32 %f742, %f556, 0f3F000000; add.ftz.f32 %f557, %f51, %f743; mul.ftz.f32 %f743, %f557, 0f3F000000; add.ftz.f32 %f558, %f49, %f738; mul.ftz.f32 %f738, %f558, 0f3F000000; add.ftz.f32 %f559, %f50, %f739; mul.ftz.f32 %f739, %f559, 0f3F000000; add.ftz.f32 %f560, %f51, %f740; mul.ftz.f32 %f740, %f560, 0f3F000000; $L__BB15_43: sub.ftz.f32 %f561, %f741, %f49; sub.ftz.f32 %f562, %f742, %f50; sub.ftz.f32 %f563, %f743, %f51; fma.rn.ftz.f32 %f564, %f549, %f561, %f49; fma.rn.ftz.f32 %f565, %f549, %f562, %f50; fma.rn.ftz.f32 %f566, %f549, %f563, %f51; sub.ftz.f32 %f567, %f55, %f738; sub.ftz.f32 %f568, %f56, %f739; sub.ftz.f32 %f569, %f57, %f740; fma.rn.ftz.f32 %f570, %f549, %f567, %f738; fma.rn.ftz.f32 %f571, %f549, %f568, %f739; fma.rn.ftz.f32 %f572, %f549, %f569, %f740; sub.ftz.f32 %f573, %f570, %f564; sub.ftz.f32 %f574, %f571, %f565; sub.ftz.f32 %f575, %f572, %f566; fma.rn.ftz.f32 %f73, %f550, %f573, %f564; fma.rn.ftz.f32 %f74, %f550, %f574, %f565; fma.rn.ftz.f32 %f75, %f550, %f575, %f566; fma.rn.ftz.f32 %f576, %f551, %f561, %f49; fma.rn.ftz.f32 %f577, %f551, %f562, %f50; fma.rn.ftz.f32 %f578, %f551, %f563, %f51; fma.rn.ftz.f32 %f579, %f551, %f567, %f738; fma.rn.ftz.f32 %f580, %f551, %f568, %f739; fma.rn.ftz.f32 %f581, %f551, %f569, %f740; sub.ftz.f32 %f582, %f579, %f576; sub.ftz.f32 %f583, %f580, %f577; sub.ftz.f32 %f584, %f581, %f578; fma.rn.ftz.f32 %f76, %f552, %f582, %f576; fma.rn.ftz.f32 %f77, %f552, %f583, %f577; fma.rn.ftz.f32 %f78, %f552, %f584, %f578; fma.rn.ftz.f32 %f585, %f553, %f561, %f49; fma.rn.ftz.f32 %f586, %f553, %f562, %f50; fma.rn.ftz.f32 %f587, %f553, %f563, %f51; fma.rn.ftz.f32 %f588, %f553, %f567, %f738; fma.rn.ftz.f32 %f589, %f553, %f568, %f739; fma.rn.ftz.f32 %f590, %f553, %f569, %f740; sub.ftz.f32 %f591, %f588, %f585; sub.ftz.f32 %f592, %f589, %f586; sub.ftz.f32 %f593, %f590, %f587; fma.rn.ftz.f32 %f79, %f554, %f591, %f585; fma.rn.ftz.f32 %f80, %f554, %f592, %f586; fma.rn.ftz.f32 %f81, %f554, %f593, %f587; setp.eq.s64 %p38, %rd8, 0; @%p38 bra $L__BB15_45; ld.local.f32 %f594, [%rd1+40]; ld.local.f32 %f595, [%rd1+48]; mul.ftz.f32 %f596, %f79, %f595; mul.ftz.f32 %f597, %f80, %f595; mul.ftz.f32 %f598, %f81, %f595; fma.rn.ftz.f32 %f599, %f76, %f594, %f596; fma.rn.ftz.f32 %f600, %f77, %f594, %f597; fma.rn.ftz.f32 %f601, %f78, %f594, %f598; add.ftz.f32 %f602, %f594, %f595; mul.ftz.f32 %f603, %f73, %f602; mul.ftz.f32 %f604, %f74, %f602; mul.ftz.f32 %f605, %f75, %f602; sub.ftz.f32 %f606, %f599, %f603; sub.ftz.f32 %f607, %f600, %f604; sub.ftz.f32 %f608, %f601, %f605; st.f32 [%rd8], %f606; st.f32 [%rd8+4], %f607; st.f32 [%rd8+8], %f608; $L__BB15_45: setp.eq.s64 %p39, %rd9, 0; @%p39 bra $L__BB15_47; ld.local.f32 %f609, [%rd1+44]; ld.local.f32 %f610, [%rd1+52]; mul.ftz.f32 %f611, %f79, %f610; mul.ftz.f32 %f612, %f80, %f610; mul.ftz.f32 %f613, %f81, %f610; fma.rn.ftz.f32 %f614, %f76, %f609, %f611; fma.rn.ftz.f32 %f615, %f77, %f609, %f612; fma.rn.ftz.f32 %f616, %f78, %f609, %f613; add.ftz.f32 %f617, %f609, %f610; mul.ftz.f32 %f618, %f73, %f617; mul.ftz.f32 %f619, %f74, %f617; mul.ftz.f32 %f620, %f75, %f617; sub.ftz.f32 %f621, %f614, %f618; sub.ftz.f32 %f622, %f615, %f619; sub.ftz.f32 %f623, %f616, %f620; st.f32 [%rd9], %f621; st.f32 [%rd9+4], %f622; st.f32 [%rd9+8], %f623; $L__BB15_47: ld.local.f32 %f624, [%rd1+8]; ld.local.f32 %f625, [%rd1+12]; mul.ftz.f32 %f626, %f79, %f625; mul.ftz.f32 %f627, %f80, %f625; mul.ftz.f32 %f628, %f81, %f625; fma.rn.ftz.f32 %f629, %f76, %f624, %f626; fma.rn.ftz.f32 %f630, %f77, %f624, %f627; fma.rn.ftz.f32 %f631, %f78, %f624, %f628; mov.f32 %f632, 0f3F800000; sub.ftz.f32 %f633, %f632, %f624; sub.ftz.f32 %f634, %f633, %f625; fma.rn.ftz.f32 %f752, %f75, %f634, %f631; fma.rn.ftz.f32 %f751, %f74, %f634, %f630; fma.rn.ftz.f32 %f750, %f73, %f634, %f629; $L__BB15_74: st.param.f32 [func_retval0+0], %f750; st.param.f32 [func_retval0+4], %f751; st.param.f32 [func_retval0+8], %f752; ret; $L__BB15_48: setp.eq.s32 %p40, %r49, 32; @%p40 bra $L__BB15_59; bra.uni $L__BB15_49; $L__BB15_59: cvt.u32.u64 %r183, %rd3; ld.const.u64 %rd114, [kernel_params+2232]; cvta.to.global.u64 %rd115, %rd114; mul.wide.s32 %rd116, %r183, 16; add.s64 %rd117, %rd115, %rd116; ld.global.v4.u32 {%r184, %r185, %r186, %r187}, [%rd117]; ld.const.u64 %rd118, [kernel_params+2248]; cvta.to.global.u64 %rd119, %rd118; mul.wide.u32 %rd120, %r184, 8; add.s64 %rd121, %rd119, %rd120; ld.global.v2.f32 {%f641, %f642}, [%rd121]; mul.wide.u32 %rd122, %r185, 8; add.s64 %rd123, %rd119, %rd122; ld.global.v2.f32 {%f643, %f644}, [%rd123]; mul.wide.u32 %rd124, %r186, 8; add.s64 %rd125, %rd119, %rd124; ld.global.v2.f32 {%f645, %f646}, [%rd125]; ld.const.u64 %rd126, [kernel_params+2288]; cvta.to.global.u64 %rd127, %rd126; add.s32 %r191, %r2, 5; mul.wide.s32 %rd128, %r191, 4; add.s64 %rd7, %rd127, %rd128; ld.global.u32 %r35, [%rd7+4]; ld.global.u32 %r36, [%rd7]; and.b32 %r37, %r36, 65535; setp.eq.s32 %p46, %r37, 4; @%p46 bra $L__BB15_61; bra.uni $L__BB15_60; $L__BB15_61: add.s32 %r223, %r35, 1; add.s32 %r224, %r35, 2; add.s32 %r225, %r35, 3; mov.u32 %r222, %r35; bra.uni $L__BB15_62; $L__BB15_49: add.s32 %r178, %r49, -1; setp.lt.u32 %p41, %r178, 2; @%p41 bra $L__BB15_54; bra.uni $L__BB15_50; $L__BB15_54: setp.eq.s64 %p44, %rd8, 0; @%p44 bra $L__BB15_56; mov.u32 %r181, 0; st.u32 [%rd8], %r181; st.u32 [%rd8+4], %r181; st.u32 [%rd8+8], %r181; $L__BB15_56: setp.eq.s64 %p45, %rd9, 0; @%p45 bra $L__BB15_58; mov.u32 %r182, 0; st.u32 [%rd9], %r182; st.u32 [%rd9+4], %r182; st.u32 [%rd9+8], %r182; $L__BB15_58: ld.const.u64 %rd110, [kernel_params+2336]; cvta.to.global.u64 %rd111, %rd110; mul.wide.s32 %rd112, %r52, 12; add.s64 %rd113, %rd111, %rd112; ld.global.f32 %f752, [%rd113+8]; ld.global.f32 %f751, [%rd113+4]; ld.global.f32 %f750, [%rd113]; bra.uni $L__BB15_74; $L__BB15_60: ld.global.u32 %r224, [%rd7+8]; shr.u32 %r192, %r36, 16; add.s32 %r222, %r35, %r192; add.s32 %r193, %r192, 1; rem.u32 %r194, %r193, %r37; add.s32 %r195, %r194, %r37; rem.u32 %r196, %r195, %r37; add.s32 %r223, %r196, %r35; add.s32 %r197, %r192, -1; rem.s32 %r198, %r197, %r37; add.s32 %r199, %r198, %r37; rem.s32 %r200, %r199, %r37; add.s32 %r225, %r200, %r35; $L__BB15_62: ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r201, %r222, %r52; mul.wide.s32 %rd131, %r201, 12; add.s64 %rd132, %rd130, %rd131; ld.global.f32 %f94, [%rd132]; ld.global.f32 %f95, [%rd132+4]; ld.global.f32 %f96, [%rd132+8]; add.s32 %r202, %r223, %r52; mul.wide.s32 %rd133, %r202, 12; add.s64 %rd134, %rd130, %rd133; ld.global.f32 %f747, [%rd134]; ld.global.f32 %f748, [%rd134+4]; ld.global.f32 %f749, [%rd134+8]; add.s32 %r203, %r224, %r52; mul.wide.s32 %rd135, %r203, 12; add.s64 %rd136, %rd130, %rd135; ld.global.f32 %f100, [%rd136]; ld.global.f32 %f101, [%rd136+4]; ld.global.f32 %f102, [%rd136+8]; add.s32 %r204, %r225, %r52; mul.wide.s32 %rd137, %r204, 12; add.s64 %rd138, %rd130, %rd137; ld.global.f32 %f744, [%rd138]; ld.global.f32 %f745, [%rd138+4]; ld.global.f32 %f746, [%rd138+8]; @%p46 bra $L__BB15_64; add.ftz.f32 %f647, %f94, %f747; mul.ftz.f32 %f747, %f647, 0f3F000000; add.ftz.f32 %f648, %f95, %f748; mul.ftz.f32 %f748, %f648, 0f3F000000; add.ftz.f32 %f649, %f96, %f749; mul.ftz.f32 %f749, %f649, 0f3F000000; add.ftz.f32 %f650, %f94, %f744; mul.ftz.f32 %f744, %f650, 0f3F000000; add.ftz.f32 %f651, %f95, %f745; mul.ftz.f32 %f745, %f651, 0f3F000000; add.ftz.f32 %f652, %f96, %f746; mul.ftz.f32 %f746, %f652, 0f3F000000; $L__BB15_64: sub.ftz.f32 %f653, %f747, %f94; sub.ftz.f32 %f654, %f748, %f95; sub.ftz.f32 %f655, %f749, %f96; fma.rn.ftz.f32 %f656, %f641, %f653, %f94; fma.rn.ftz.f32 %f657, %f641, %f654, %f95; fma.rn.ftz.f32 %f658, %f641, %f655, %f96; sub.ftz.f32 %f659, %f100, %f744; sub.ftz.f32 %f660, %f101, %f745; sub.ftz.f32 %f661, %f102, %f746; fma.rn.ftz.f32 %f662, %f641, %f659, %f744; fma.rn.ftz.f32 %f663, %f641, %f660, %f745; fma.rn.ftz.f32 %f664, %f641, %f661, %f746; sub.ftz.f32 %f665, %f662, %f656; sub.ftz.f32 %f666, %f663, %f657; sub.ftz.f32 %f667, %f664, %f658; fma.rn.ftz.f32 %f118, %f642, %f665, %f656; fma.rn.ftz.f32 %f119, %f642, %f666, %f657; fma.rn.ftz.f32 %f120, %f642, %f667, %f658; fma.rn.ftz.f32 %f668, %f643, %f653, %f94; fma.rn.ftz.f32 %f669, %f643, %f654, %f95; fma.rn.ftz.f32 %f670, %f643, %f655, %f96; fma.rn.ftz.f32 %f671, %f643, %f659, %f744; fma.rn.ftz.f32 %f672, %f643, %f660, %f745; fma.rn.ftz.f32 %f673, %f643, %f661, %f746; sub.ftz.f32 %f674, %f671, %f668; sub.ftz.f32 %f675, %f672, %f669; sub.ftz.f32 %f676, %f673, %f670; fma.rn.ftz.f32 %f121, %f644, %f674, %f668; fma.rn.ftz.f32 %f122, %f644, %f675, %f669; fma.rn.ftz.f32 %f123, %f644, %f676, %f670; fma.rn.ftz.f32 %f677, %f645, %f653, %f94; fma.rn.ftz.f32 %f678, %f645, %f654, %f95; fma.rn.ftz.f32 %f679, %f645, %f655, %f96; fma.rn.ftz.f32 %f680, %f645, %f659, %f744; fma.rn.ftz.f32 %f681, %f645, %f660, %f745; fma.rn.ftz.f32 %f682, %f645, %f661, %f746; sub.ftz.f32 %f683, %f680, %f677; sub.ftz.f32 %f684, %f681, %f678; sub.ftz.f32 %f685, %f682, %f679; fma.rn.ftz.f32 %f124, %f646, %f683, %f677; fma.rn.ftz.f32 %f125, %f646, %f684, %f678; fma.rn.ftz.f32 %f126, %f646, %f685, %f679; setp.eq.s64 %p48, %rd8, 0; @%p48 bra $L__BB15_66; ld.local.f32 %f686, [%rd1+40]; ld.local.f32 %f687, [%rd1+48]; mul.ftz.f32 %f688, %f124, %f687; mul.ftz.f32 %f689, %f125, %f687; mul.ftz.f32 %f690, %f126, %f687; fma.rn.ftz.f32 %f691, %f121, %f686, %f688; fma.rn.ftz.f32 %f692, %f122, %f686, %f689; fma.rn.ftz.f32 %f693, %f123, %f686, %f690; add.ftz.f32 %f694, %f686, %f687; mul.ftz.f32 %f695, %f118, %f694; mul.ftz.f32 %f696, %f119, %f694; mul.ftz.f32 %f697, %f120, %f694; sub.ftz.f32 %f698, %f691, %f695; sub.ftz.f32 %f699, %f692, %f696; sub.ftz.f32 %f700, %f693, %f697; st.f32 [%rd8], %f698; st.f32 [%rd8+4], %f699; st.f32 [%rd8+8], %f700; $L__BB15_66: setp.eq.s64 %p49, %rd9, 0; @%p49 bra $L__BB15_68; ld.local.f32 %f701, [%rd1+44]; ld.local.f32 %f702, [%rd1+52]; mul.ftz.f32 %f703, %f124, %f702; mul.ftz.f32 %f704, %f125, %f702; mul.ftz.f32 %f705, %f126, %f702; fma.rn.ftz.f32 %f706, %f121, %f701, %f703; fma.rn.ftz.f32 %f707, %f122, %f701, %f704; fma.rn.ftz.f32 %f708, %f123, %f701, %f705; add.ftz.f32 %f709, %f701, %f702; mul.ftz.f32 %f710, %f118, %f709; mul.ftz.f32 %f711, %f119, %f709; mul.ftz.f32 %f712, %f120, %f709; sub.ftz.f32 %f713, %f706, %f710; sub.ftz.f32 %f714, %f707, %f711; sub.ftz.f32 %f715, %f708, %f712; st.f32 [%rd9], %f713; st.f32 [%rd9+4], %f714; st.f32 [%rd9+8], %f715; $L__BB15_68: ld.local.f32 %f716, [%rd1+8]; ld.local.f32 %f717, [%rd1+12]; mul.ftz.f32 %f718, %f124, %f717; mul.ftz.f32 %f719, %f125, %f717; mul.ftz.f32 %f720, %f126, %f717; fma.rn.ftz.f32 %f721, %f121, %f716, %f718; fma.rn.ftz.f32 %f722, %f122, %f716, %f719; fma.rn.ftz.f32 %f723, %f123, %f716, %f720; mov.f32 %f724, 0f3F800000; sub.ftz.f32 %f725, %f724, %f716; sub.ftz.f32 %f726, %f725, %f717; fma.rn.ftz.f32 %f752, %f120, %f726, %f723; fma.rn.ftz.f32 %f751, %f119, %f726, %f722; fma.rn.ftz.f32 %f750, %f118, %f726, %f721; bra.uni $L__BB15_74; $L__BB15_50: setp.eq.s64 %p42, %rd8, 0; @%p42 bra $L__BB15_52; mov.u32 %r179, 0; st.u32 [%rd8], %r179; st.u32 [%rd8+4], %r179; st.u32 [%rd8+8], %r179; $L__BB15_52: setp.eq.s64 %p43, %rd9, 0; mov.f32 %f750, 0f00000000; mov.f32 %f751, %f750; mov.f32 %f752, %f750; @%p43 bra $L__BB15_74; mov.u32 %r180, 0; st.u32 [%rd9], %r180; st.u32 [%rd9+4], %r180; st.u32 [%rd9+8], %r180; mov.f32 %f751, %f750; mov.f32 %f752, %f750; bra.uni $L__BB15_74; } .func (.param .align 4 .b8 func_retval0[16]) _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_( .param .b64 _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_0, .param .align 4 .b8 _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_1[16], .param .b64 _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_2, .param .b64 _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_3 ) { .reg .pred %p<142>; .reg .b16 %rs<182>; .reg .f32 %f<2122>; .reg .b32 %r<359>; .reg .b64 %rd<229>; ld.param.u64 %rd15, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_0]; ld.param.u64 %rd16, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_2]; ld.param.u64 %rd17, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_3]; ld.param.u32 %r96, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_1+12]; ld.param.u32 %r95, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_1+8]; ld.param.u32 %r94, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_1+4]; ld.param.u32 %r93, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_1]; cvta.to.local.u64 %rd18, %rd15; add.s64 %rd1, %rd18, 60; ld.local.s32 %rd3, [%rd18+60]; setp.eq.s64 %p1, %rd3, -1; mov.u32 %r2, -1; @%p1 bra $L__BB16_2; ld.const.u64 %rd19, [kernel_params+2240]; cvta.to.global.u64 %rd20, %rd19; shl.b64 %rd21, %rd3, 2; add.s64 %rd22, %rd20, %rd21; ld.global.u32 %r2, [%rd22]; $L__BB16_2: and.b32 %r98, %r95, 2; setp.eq.s32 %p2, %r98, 0; @%p2 bra $L__BB16_215; bra.uni $L__BB16_3; $L__BB16_215: setp.eq.s32 %p111, %r93, 4; @%p111 bra $L__BB16_284; bra.uni $L__BB16_216; $L__BB16_284: setp.eq.s64 %p140, %rd16, 0; @%p140 bra $L__BB16_286; mov.f32 %f1987, 0f00000000; st.v4.f32 [%rd16], {%f1987, %f1987, %f1987, %f1987}; $L__BB16_286: setp.eq.s64 %p141, %rd17, 0; @%p141 bra $L__BB16_288; mov.f32 %f1988, 0f00000000; st.v4.f32 [%rd17], {%f1988, %f1988, %f1988, %f1988}; $L__BB16_288: ld.const.u64 %rd211, [kernel_params+2344]; cvta.to.global.u64 %rd212, %rd211; ld.const.u64 %rd213, [kernel_params+2288]; cvta.to.global.u64 %rd214, %rd213; add.s32 %r329, %r2, 4; mul.wide.s32 %rd215, %r329, 4; add.s64 %rd216, %rd214, %rd215; ld.global.u32 %r330, [%rd216]; add.s32 %r331, %r330, %r96; mul.wide.u32 %rd217, %r331, 16; add.s64 %rd218, %rd212, %rd217; ld.global.v4.f32 {%f2118, %f2119, %f2120, %f2121}, [%rd218]; bra.uni $L__BB16_289; $L__BB16_3: ld.const.u64 %rd23, [kernel_params+2232]; cvta.to.global.u64 %rd24, %rd23; shl.b64 %rd25, %rd3, 4; add.s64 %rd26, %rd24, %rd25; ld.global.v4.u32 {%r99, %r100, %r101, %r102}, [%rd26]; ld.const.u64 %rd27, [kernel_params+2248]; cvta.to.global.u64 %rd28, %rd27; mul.wide.u32 %rd29, %r99, 8; add.s64 %rd30, %rd28, %rd29; ld.global.v2.f32 {%f721, %f722}, [%rd30]; mul.wide.u32 %rd31, %r100, 8; add.s64 %rd32, %rd28, %rd31; ld.global.v2.f32 {%f725, %f726}, [%rd32]; mul.wide.u32 %rd33, %r101, 8; add.s64 %rd34, %rd28, %rd33; ld.global.v2.f32 {%f729, %f730}, [%rd34]; sub.ftz.f32 %f1, %f725, %f721; sub.ftz.f32 %f2, %f726, %f722; sub.ftz.f32 %f3, %f729, %f721; sub.ftz.f32 %f4, %f730, %f722; ld.local.f32 %f733, [%rd1+8]; ld.local.v2.u32 {%r106, %r107}, [%rd1+12]; mov.b32 %f734, %r106; mul.ftz.f32 %f735, %f3, %f734; mul.ftz.f32 %f736, %f4, %f734; fma.rn.ftz.f32 %f737, %f1, %f733, %f735; fma.rn.ftz.f32 %f738, %f2, %f733, %f736; add.ftz.f32 %f5, %f721, %f737; add.ftz.f32 %f6, %f722, %f738; setp.eq.s32 %p3, %r94, 4; @%p3 bra $L__BB16_35; bra.uni $L__BB16_4; $L__BB16_35: setp.eq.s32 %p32, %r107, -1; mov.u32 %r345, 0; mov.u32 %r344, %r345; @%p32 bra $L__BB16_37; ld.const.u64 %rd90, [kernel_params+2160]; cvta.to.global.u64 %rd91, %rd90; mul.wide.s32 %rd92, %r107, 208; add.s64 %rd93, %rd91, %rd92; ld.global.u32 %r344, [%rd93+160]; $L__BB16_37: add.s32 %r208, %r344, %r2; shr.u32 %r351, %r208, 1; ld.const.u64 %rd94, [kernel_params+2288]; cvta.to.global.u64 %rd7, %rd94; mov.f32 %f2004, 0f3F000000; mov.f32 %f2008, %f5; mov.f32 %f2007, %f6; bra.uni $L__BB16_38; $L__BB16_63: add.s32 %r345, %r345, 4; mul.ftz.f32 %f2004, %f61, 0f3F000000; $L__BB16_38: mul.ftz.f32 %f49, %f2004, 0f3F000000; setp.lt.ftz.f32 %p33, %f2008, %f2004; @%p33 bra $L__BB16_40; bra.uni $L__BB16_39; $L__BB16_40: setp.lt.ftz.f32 %p35, %f2007, %f2004; mov.u32 %r347, 0; @%p35 bra $L__BB16_42; sub.ftz.f32 %f2007, %f2007, %f2004; mov.u32 %r347, 1; bra.uni $L__BB16_42; $L__BB16_39: setp.lt.ftz.f32 %p34, %f2007, %f2004; sub.ftz.f32 %f1236, %f2007, %f2004; selp.f32 %f2007, %f2007, %f1236, %p34; selp.b32 %r347, 3, 2, %p34; sub.ftz.f32 %f2008, %f2008, %f2004; $L__BB16_42: add.s32 %r214, %r347, %r351; mul.wide.s32 %rd95, %r214, 4; add.s64 %rd96, %rd7, %rd95; ld.global.u32 %r41, [%rd96]; and.b32 %r215, %r41, 1073741824; setp.eq.s32 %p36, %r215, 0; mov.u32 %r352, -1; @%p36 bra $L__BB16_65; and.b32 %r351, %r41, 1073741823; setp.lt.s32 %p37, %r41, 0; @%p37 bra $L__BB16_64; mul.ftz.f32 %f55, %f49, 0f3F000000; setp.lt.ftz.f32 %p38, %f2008, %f49; @%p38 bra $L__BB16_46; bra.uni $L__BB16_45; $L__BB16_46: setp.lt.ftz.f32 %p40, %f2007, %f49; mov.u32 %r348, 0; @%p40 bra $L__BB16_48; sub.ftz.f32 %f2007, %f2007, %f49; mov.u32 %r348, 1; bra.uni $L__BB16_48; $L__BB16_45: setp.lt.ftz.f32 %p39, %f2007, %f49; sub.ftz.f32 %f1237, %f2007, %f49; selp.f32 %f2007, %f2007, %f1237, %p39; selp.b32 %r348, 3, 2, %p39; sub.ftz.f32 %f2008, %f2008, %f49; $L__BB16_48: add.s32 %r223, %r348, %r351; mul.wide.u32 %rd97, %r223, 4; add.s64 %rd98, %rd7, %rd97; ld.global.u32 %r45, [%rd98]; and.b32 %r224, %r45, 1073741824; setp.eq.s32 %p41, %r224, 0; @%p41 bra $L__BB16_65; and.b32 %r351, %r45, 1073741823; setp.lt.s32 %p42, %r45, 0; @%p42 bra $L__BB16_64; mul.ftz.f32 %f61, %f55, 0f3F000000; setp.lt.ftz.f32 %p43, %f2008, %f55; @%p43 bra $L__BB16_52; bra.uni $L__BB16_51; $L__BB16_52: setp.lt.ftz.f32 %p45, %f2007, %f55; mov.u32 %r349, 0; @%p45 bra $L__BB16_54; sub.ftz.f32 %f2007, %f2007, %f55; mov.u32 %r349, 1; bra.uni $L__BB16_54; $L__BB16_51: setp.lt.ftz.f32 %p44, %f2007, %f55; sub.ftz.f32 %f1238, %f2007, %f55; selp.f32 %f2007, %f2007, %f1238, %p44; selp.b32 %r349, 3, 2, %p44; sub.ftz.f32 %f2008, %f2008, %f55; $L__BB16_54: add.s32 %r232, %r349, %r351; mul.wide.u32 %rd99, %r232, 4; add.s64 %rd100, %rd7, %rd99; ld.global.u32 %r49, [%rd100]; and.b32 %r233, %r49, 1073741824; setp.eq.s32 %p46, %r233, 0; @%p46 bra $L__BB16_65; and.b32 %r351, %r49, 1073741823; setp.lt.s32 %p47, %r49, 0; @%p47 bra $L__BB16_64; add.s32 %r239, %r345, 3; setp.gt.u32 %p48, %r239, 254; @%p48 bra $L__BB16_65; setp.lt.ftz.f32 %p49, %f2008, %f61; @%p49 bra $L__BB16_59; bra.uni $L__BB16_58; $L__BB16_59: setp.lt.ftz.f32 %p51, %f2007, %f61; mov.u32 %r350, 0; @%p51 bra $L__BB16_61; sub.ftz.f32 %f2007, %f2007, %f61; mov.u32 %r350, 1; bra.uni $L__BB16_61; $L__BB16_58: setp.lt.ftz.f32 %p50, %f2007, %f61; sub.ftz.f32 %f1239, %f2007, %f61; selp.f32 %f2007, %f2007, %f1239, %p50; selp.b32 %r350, 3, 2, %p50; sub.ftz.f32 %f2008, %f2008, %f61; $L__BB16_61: add.s32 %r247, %r350, %r351; mul.wide.u32 %rd101, %r247, 4; add.s64 %rd102, %rd7, %rd101; ld.global.u32 %r53, [%rd102]; and.b32 %r248, %r53, 1073741824; setp.eq.s32 %p52, %r248, 0; @%p52 bra $L__BB16_65; and.b32 %r351, %r53, 1073741823; setp.lt.s32 %p53, %r53, 0; @%p53 bra $L__BB16_64; bra.uni $L__BB16_63; $L__BB16_64: mul.wide.u32 %rd103, %r351, 4; add.s64 %rd104, %rd7, %rd103; ld.global.u32 %r354, [%rd104+8]; ld.global.u32 %r353, [%rd104+4]; ld.global.u32 %r352, [%rd104]; $L__BB16_65: add.s32 %r251, %r352, 2; mul.wide.s32 %rd105, %r251, 4; add.s64 %rd106, %rd7, %rd105; ld.global.u32 %r252, [%rd106]; add.s32 %r253, %r252, %r354; mul.wide.s32 %rd107, %r253, 4; add.s64 %rd108, %rd7, %rd107; ld.global.u32 %r63, [%rd108+4]; ld.global.u32 %r64, [%rd108+8]; ld.global.u32 %r65, [%rd108+12]; ld.global.u32 %r66, [%rd108+16]; ld.global.u32 %r67, [%rd108+20]; ld.global.u32 %r68, [%rd108+24]; ld.global.u32 %r69, [%rd108+28]; ld.global.u32 %r70, [%rd108+32]; ld.global.u32 %r71, [%rd108+36]; ld.global.u32 %r72, [%rd108+40]; ld.global.u32 %r73, [%rd108+44]; ld.global.u32 %r74, [%rd108+48]; ld.global.u32 %r75, [%rd108+52]; ld.global.u32 %r76, [%rd108+56]; ld.global.u32 %r77, [%rd108+60]; add.s32 %r254, %r353, 1; mov.u32 %r255, 1; mul.wide.s32 %rd109, %r254, 4; add.s64 %rd110, %rd7, %rd109; ld.global.u32 %r256, [%rd110]; and.b32 %r257, %r256, 15; shl.b32 %r258, %r255, %r257; cvt.rn.f32.s32 %f1240, %r258; and.b32 %r259, %r256, 16; setp.eq.s32 %p54, %r259, 0; mul.ftz.f32 %f1241, %f1240, 0f3F000000; selp.f32 %f73, %f1240, %f1241, %p54; shr.u32 %r260, %r256, 4; and.b32 %r261, %r260, 1; setp.eq.b32 %p55, %r261, 1; selp.b32 %r262, -1, 0, %p55; add.s32 %r263, %r257, %r262; shl.b32 %r264, %r255, %r263; cvt.rn.f32.s32 %f1242, %r264; rcp.approx.ftz.f32 %f1243, %f1242; mov.f32 %f1244, 0f3F800000; shr.u32 %r265, %r256, 22; bfe.u32 %r266, %r256, 12, 10; cvt.rn.f32.s32 %f1245, %r265; mul.ftz.f32 %f1246, %f1243, %f1245; cvt.rn.f32.s32 %f1247, %r266; mul.ftz.f32 %f1248, %f1243, %f1247; sub.ftz.f32 %f1249, %f5, %f1246; div.approx.ftz.f32 %f1250, %f1249, %f1243; sub.ftz.f32 %f1251, %f6, %f1248; div.approx.ftz.f32 %f1252, %f1251, %f1243; mul.ftz.f32 %f1253, %f1250, %f1250; mul.ftz.f32 %f1254, %f1250, %f1253; sub.ftz.f32 %f1255, %f1250, %f1253; mul.ftz.f32 %f1256, %f1255, 0f40400000; sub.ftz.f32 %f1257, %f1244, %f1256; sub.ftz.f32 %f1258, %f1257, %f1254; mul.ftz.f32 %f1259, %f1258, 0f3E2AAAAB; mul.ftz.f32 %f1260, %f1253, 0f40C00000; mov.f32 %f1261, 0f40800000; sub.ftz.f32 %f1262, %f1261, %f1260; fma.rn.ftz.f32 %f1263, %f1254, 0f40400000, %f1262; mul.ftz.f32 %f1264, %f1263, 0f3E2AAAAB; add.ftz.f32 %f1265, %f1250, %f1253; sub.ftz.f32 %f1266, %f1265, %f1254; fma.rn.ftz.f32 %f1267, %f1266, 0f40400000, 0f3F800000; mul.ftz.f32 %f1268, %f1267, 0f3E2AAAAB; mul.ftz.f32 %f1269, %f1254, 0f3E2AAAAB; mul.ftz.f32 %f1270, %f1253, 0f3F000000; sub.ftz.f32 %f1271, %f1250, %f1270; add.ftz.f32 %f1272, %f1271, 0fBF000000; mul.ftz.f32 %f1273, %f1253, 0f3FC00000; add.ftz.f32 %f1274, %f1250, %f1250; sub.ftz.f32 %f1275, %f1273, %f1274; sub.ftz.f32 %f1276, %f1250, %f1273; add.ftz.f32 %f1277, %f1276, 0f3F000000; mul.ftz.f32 %f1278, %f1252, %f1252; mul.ftz.f32 %f1279, %f1252, %f1278; sub.ftz.f32 %f1280, %f1252, %f1278; mul.ftz.f32 %f1281, %f1280, 0f40400000; sub.ftz.f32 %f1282, %f1244, %f1281; sub.ftz.f32 %f1283, %f1282, %f1279; mul.ftz.f32 %f1284, %f1283, 0f3E2AAAAB; mul.ftz.f32 %f1285, %f1278, 0f40C00000; sub.ftz.f32 %f1286, %f1261, %f1285; fma.rn.ftz.f32 %f1287, %f1279, 0f40400000, %f1286; mul.ftz.f32 %f1288, %f1287, 0f3E2AAAAB; add.ftz.f32 %f1289, %f1252, %f1278; sub.ftz.f32 %f1290, %f1289, %f1279; fma.rn.ftz.f32 %f1291, %f1290, 0f40400000, 0f3F800000; mul.ftz.f32 %f1292, %f1291, 0f3E2AAAAB; mul.ftz.f32 %f1293, %f1279, 0f3E2AAAAB; mul.ftz.f32 %f1294, %f1278, 0f3F000000; sub.ftz.f32 %f1295, %f1252, %f1294; add.ftz.f32 %f1296, %f1295, 0fBF000000; mul.ftz.f32 %f1297, %f1278, 0f3FC00000; add.ftz.f32 %f1298, %f1252, %f1252; sub.ftz.f32 %f1299, %f1297, %f1298; sub.ftz.f32 %f1300, %f1252, %f1297; add.ftz.f32 %f1301, %f1300, 0f3F000000; and.b32 %r267, %r256, 256; setp.eq.s32 %p56, %r267, 0; sub.ftz.f32 %f1302, %f1292, %f1284; fma.rn.ftz.f32 %f1303, %f1284, 0f40000000, %f1288; selp.f32 %f1304, %f1292, %f1302, %p56; selp.f32 %f1305, %f1288, %f1303, %p56; selp.f32 %f74, %f1284, 0f00000000, %p56; and.b32 %r268, %r256, 512; setp.eq.s32 %p57, %r268, 0; sub.ftz.f32 %f1306, %f1264, %f1269; fma.rn.ftz.f32 %f1307, %f1269, 0f40000000, %f1268; selp.f32 %f75, %f1269, 0f00000000, %p57; selp.f32 %f1308, %f1268, %f1307, %p57; selp.f32 %f1309, %f1264, %f1306, %p57; and.b32 %r269, %r256, 1024; setp.eq.s32 %p58, %r269, 0; sub.ftz.f32 %f1310, %f1305, %f1293; fma.rn.ftz.f32 %f1311, %f1293, 0f40000000, %f1304; selp.f32 %f76, %f1293, 0f00000000, %p58; selp.f32 %f77, %f1304, %f1311, %p58; selp.f32 %f78, %f1305, %f1310, %p58; and.b32 %r270, %r256, 2048; setp.eq.s32 %p59, %r270, 0; sub.ftz.f32 %f1312, %f1308, %f1259; fma.rn.ftz.f32 %f1313, %f1259, 0f40000000, %f1309; selp.f32 %f79, %f1308, %f1312, %p59; selp.f32 %f80, %f1309, %f1313, %p59; selp.f32 %f81, %f1259, 0f00000000, %p59; sub.ftz.f32 %f1314, %f1301, %f1296; fma.rn.ftz.f32 %f1315, %f1296, 0f40000000, %f1299; selp.f32 %f1316, %f1301, %f1314, %p56; selp.f32 %f1317, %f1299, %f1315, %p56; selp.f32 %f82, %f1296, 0f00000000, %p56; sub.ftz.f32 %f1318, %f1275, %f1270; fma.rn.ftz.f32 %f1319, %f1270, 0f40000000, %f1277; selp.f32 %f83, %f1270, 0f00000000, %p57; selp.f32 %f1320, %f1277, %f1319, %p57; selp.f32 %f1321, %f1275, %f1318, %p57; sub.ftz.f32 %f1322, %f1317, %f1294; fma.rn.ftz.f32 %f1323, %f1294, 0f40000000, %f1316; selp.f32 %f84, %f1294, 0f00000000, %p58; selp.f32 %f85, %f1316, %f1323, %p58; selp.f32 %f86, %f1317, %f1322, %p58; sub.ftz.f32 %f1324, %f1320, %f1272; fma.rn.ftz.f32 %f1325, %f1272, 0f40000000, %f1321; selp.f32 %f87, %f1320, %f1324, %p59; selp.f32 %f88, %f1321, %f1325, %p59; selp.f32 %f89, %f1272, 0f00000000, %p59; ld.const.u64 %rd111, [kernel_params+2352]; cvta.to.global.u64 %rd8, %rd111; ld.global.u32 %r271, [%rd108]; add.s32 %r272, %r271, %r96; mul.wide.s32 %rd112, %r272, 4; add.s64 %rd113, %rd8, %rd112; ld.global.v4.u8 {%rs81, %rs82, %rs83, %rs84}, [%rd113]; cvt.rn.f32.u16 %f1326, %rs81; mul.ftz.f32 %f90, %f1326, 0f3B808081; cvt.rn.f32.u16 %f1327, %rs82; mul.ftz.f32 %f91, %f1327, 0f3B808081; cvt.rn.f32.u16 %f1328, %rs83; mul.ftz.f32 %f92, %f1328, 0f3B808081; setp.lt.ftz.f32 %p60, %f90, 0f3D25AEE6; @%p60 bra $L__BB16_67; bra.uni $L__BB16_66; $L__BB16_67: mul.ftz.f32 %f2015, %f90, 0f3D9E8391; bra.uni $L__BB16_68; $L__BB16_216: add.s32 %r289, %r93, -8; shf.l.wrap.b32 %r288, %r289, %r289, 29; setp.lt.u32 %p112, %r288, 2; @%p112 bra $L__BB16_267; setp.eq.s32 %p113, %r288, 7; @%p113 bra $L__BB16_219; setp.ne.s32 %p114, %r288, 3; @%p114 bra $L__BB16_274; $L__BB16_219: cvt.u32.u64 %r290, %rd3; ld.const.u64 %rd147, [kernel_params+2232]; cvta.to.global.u64 %rd148, %rd147; mul.wide.s32 %rd149, %r290, 16; add.s64 %rd150, %rd148, %rd149; ld.global.v4.u32 {%r291, %r292, %r293, %r294}, [%rd150]; ld.const.u64 %rd151, [kernel_params+2248]; cvta.to.global.u64 %rd152, %rd151; mul.wide.u32 %rd153, %r291, 8; add.s64 %rd154, %rd152, %rd153; ld.global.v2.f32 {%f1658, %f1659}, [%rd154]; mul.wide.u32 %rd155, %r292, 8; add.s64 %rd156, %rd152, %rd155; ld.global.v2.f32 {%f1660, %f1661}, [%rd156]; mul.wide.u32 %rd157, %r293, 8; add.s64 %rd158, %rd152, %rd157; ld.global.v2.f32 {%f1662, %f1663}, [%rd158]; ld.const.u64 %rd159, [kernel_params+2288]; cvta.to.global.u64 %rd160, %rd159; add.s32 %r298, %r2, 5; mul.wide.s32 %rd161, %r298, 4; add.s64 %rd9, %rd160, %rd161; ld.global.u32 %r79, [%rd9+4]; ld.global.u32 %r80, [%rd9]; and.b32 %r81, %r80, 65535; setp.eq.s32 %p115, %r81, 4; @%p115 bra $L__BB16_221; bra.uni $L__BB16_220; $L__BB16_221: add.s32 %r356, %r79, 1; add.s32 %r357, %r79, 2; add.s32 %r358, %r79, 3; mov.u32 %r355, %r79; bra.uni $L__BB16_222; $L__BB16_4: setp.eq.s32 %p4, %r107, -1; mov.u32 %r334, 0; mov.u32 %r333, %r334; @%p4 bra $L__BB16_6; ld.const.u64 %rd35, [kernel_params+2160]; cvta.to.global.u64 %rd36, %rd35; mul.wide.s32 %rd37, %r107, 208; add.s64 %rd38, %rd36, %rd37; ld.global.u32 %r333, [%rd38+160]; $L__BB16_6: add.s32 %r111, %r333, %r2; shr.u32 %r340, %r111, 1; ld.const.u64 %rd39, [kernel_params+2288]; cvta.to.global.u64 %rd6, %rd39; mov.f32 %f1993, 0f3F000000; mov.f32 %f1997, %f5; mov.f32 %f1996, %f6; bra.uni $L__BB16_7; $L__BB16_32: add.s32 %r334, %r334, 4; mul.ftz.f32 %f1993, %f22, 0f3F000000; $L__BB16_7: mul.ftz.f32 %f10, %f1993, 0f3F000000; setp.lt.ftz.f32 %p5, %f1997, %f1993; @%p5 bra $L__BB16_9; bra.uni $L__BB16_8; $L__BB16_9: setp.lt.ftz.f32 %p7, %f1996, %f1993; mov.u32 %r336, 0; @%p7 bra $L__BB16_11; sub.ftz.f32 %f1996, %f1996, %f1993; mov.u32 %r336, 1; bra.uni $L__BB16_11; $L__BB16_8: setp.lt.ftz.f32 %p6, %f1996, %f1993; sub.ftz.f32 %f740, %f1996, %f1993; selp.f32 %f1996, %f1996, %f740, %p6; selp.b32 %r336, 3, 2, %p6; sub.ftz.f32 %f1997, %f1997, %f1993; $L__BB16_11: add.s32 %r117, %r336, %r340; mul.wide.s32 %rd40, %r117, 4; add.s64 %rd41, %rd6, %rd40; ld.global.u32 %r12, [%rd41]; and.b32 %r118, %r12, 1073741824; setp.eq.s32 %p8, %r118, 0; mov.u32 %r341, -1; @%p8 bra $L__BB16_34; and.b32 %r340, %r12, 1073741823; setp.lt.s32 %p9, %r12, 0; @%p9 bra $L__BB16_33; mul.ftz.f32 %f16, %f10, 0f3F000000; setp.lt.ftz.f32 %p10, %f1997, %f10; @%p10 bra $L__BB16_15; bra.uni $L__BB16_14; $L__BB16_15: setp.lt.ftz.f32 %p12, %f1996, %f10; mov.u32 %r337, 0; @%p12 bra $L__BB16_17; sub.ftz.f32 %f1996, %f1996, %f10; mov.u32 %r337, 1; bra.uni $L__BB16_17; $L__BB16_14: setp.lt.ftz.f32 %p11, %f1996, %f10; sub.ftz.f32 %f741, %f1996, %f10; selp.f32 %f1996, %f1996, %f741, %p11; selp.b32 %r337, 3, 2, %p11; sub.ftz.f32 %f1997, %f1997, %f10; $L__BB16_17: add.s32 %r126, %r337, %r340; mul.wide.u32 %rd42, %r126, 4; add.s64 %rd43, %rd6, %rd42; ld.global.u32 %r16, [%rd43]; and.b32 %r127, %r16, 1073741824; setp.eq.s32 %p13, %r127, 0; @%p13 bra $L__BB16_34; and.b32 %r340, %r16, 1073741823; setp.lt.s32 %p14, %r16, 0; @%p14 bra $L__BB16_33; mul.ftz.f32 %f22, %f16, 0f3F000000; setp.lt.ftz.f32 %p15, %f1997, %f16; @%p15 bra $L__BB16_21; bra.uni $L__BB16_20; $L__BB16_21: setp.lt.ftz.f32 %p17, %f1996, %f16; mov.u32 %r338, 0; @%p17 bra $L__BB16_23; sub.ftz.f32 %f1996, %f1996, %f16; mov.u32 %r338, 1; bra.uni $L__BB16_23; $L__BB16_20: setp.lt.ftz.f32 %p16, %f1996, %f16; sub.ftz.f32 %f742, %f1996, %f16; selp.f32 %f1996, %f1996, %f742, %p16; selp.b32 %r338, 3, 2, %p16; sub.ftz.f32 %f1997, %f1997, %f16; $L__BB16_23: add.s32 %r135, %r338, %r340; mul.wide.u32 %rd44, %r135, 4; add.s64 %rd45, %rd6, %rd44; ld.global.u32 %r20, [%rd45]; and.b32 %r136, %r20, 1073741824; setp.eq.s32 %p18, %r136, 0; @%p18 bra $L__BB16_34; and.b32 %r340, %r20, 1073741823; setp.lt.s32 %p19, %r20, 0; @%p19 bra $L__BB16_33; add.s32 %r142, %r334, 3; setp.gt.u32 %p20, %r142, 254; @%p20 bra $L__BB16_34; setp.lt.ftz.f32 %p21, %f1997, %f22; @%p21 bra $L__BB16_28; bra.uni $L__BB16_27; $L__BB16_28: setp.lt.ftz.f32 %p23, %f1996, %f22; mov.u32 %r339, 0; @%p23 bra $L__BB16_30; sub.ftz.f32 %f1996, %f1996, %f22; mov.u32 %r339, 1; bra.uni $L__BB16_30; $L__BB16_27: setp.lt.ftz.f32 %p22, %f1996, %f22; sub.ftz.f32 %f743, %f1996, %f22; selp.f32 %f1996, %f1996, %f743, %p22; selp.b32 %r339, 3, 2, %p22; sub.ftz.f32 %f1997, %f1997, %f22; $L__BB16_30: add.s32 %r150, %r339, %r340; mul.wide.u32 %rd46, %r150, 4; add.s64 %rd47, %rd6, %rd46; ld.global.u32 %r24, [%rd47]; and.b32 %r151, %r24, 1073741824; setp.eq.s32 %p24, %r151, 0; @%p24 bra $L__BB16_34; and.b32 %r340, %r24, 1073741823; setp.lt.s32 %p25, %r24, 0; @%p25 bra $L__BB16_33; bra.uni $L__BB16_32; $L__BB16_33: mul.wide.u32 %rd48, %r340, 4; add.s64 %rd49, %rd6, %rd48; ld.global.u32 %r343, [%rd49+8]; ld.global.u32 %r342, [%rd49+4]; ld.global.u32 %r341, [%rd49]; $L__BB16_34: add.s32 %r154, %r341, 2; mul.wide.s32 %rd50, %r154, 4; add.s64 %rd51, %rd6, %rd50; ld.global.u32 %r155, [%rd51]; add.s32 %r156, %r155, %r343; mul.wide.s32 %rd52, %r156, 4; add.s64 %rd53, %rd6, %rd52; add.s32 %r157, %r342, 1; mov.u32 %r158, 1; mul.wide.s32 %rd54, %r157, 4; add.s64 %rd55, %rd6, %rd54; ld.global.u32 %r159, [%rd55]; and.b32 %r160, %r159, 15; shl.b32 %r161, %r158, %r160; cvt.rn.f32.s32 %f744, %r161; and.b32 %r162, %r159, 16; setp.eq.s32 %p26, %r162, 0; mul.ftz.f32 %f745, %f744, 0f3F000000; selp.f32 %f746, %f744, %f745, %p26; shr.u32 %r163, %r159, 4; and.b32 %r164, %r163, 1; setp.eq.b32 %p27, %r164, 1; selp.b32 %r165, -1, 0, %p27; add.s32 %r166, %r160, %r165; shl.b32 %r167, %r158, %r166; cvt.rn.f32.s32 %f747, %r167; rcp.approx.ftz.f32 %f748, %f747; mov.f32 %f749, 0f3F800000; shr.u32 %r168, %r159, 22; bfe.u32 %r169, %r159, 12, 10; cvt.rn.f32.s32 %f750, %r168; mul.ftz.f32 %f751, %f748, %f750; cvt.rn.f32.s32 %f752, %r169; mul.ftz.f32 %f753, %f748, %f752; sub.ftz.f32 %f754, %f5, %f751; div.approx.ftz.f32 %f755, %f754, %f748; sub.ftz.f32 %f756, %f6, %f753; div.approx.ftz.f32 %f757, %f756, %f748; mul.ftz.f32 %f758, %f755, %f755; mul.ftz.f32 %f759, %f755, %f758; sub.ftz.f32 %f760, %f755, %f758; mul.ftz.f32 %f761, %f760, 0f40400000; sub.ftz.f32 %f762, %f749, %f761; sub.ftz.f32 %f763, %f762, %f759; mul.ftz.f32 %f764, %f763, 0f3E2AAAAB; mul.ftz.f32 %f765, %f758, 0f40C00000; mov.f32 %f766, 0f40800000; sub.ftz.f32 %f767, %f766, %f765; fma.rn.ftz.f32 %f768, %f759, 0f40400000, %f767; mul.ftz.f32 %f769, %f768, 0f3E2AAAAB; add.ftz.f32 %f770, %f755, %f758; sub.ftz.f32 %f771, %f770, %f759; fma.rn.ftz.f32 %f772, %f771, 0f40400000, 0f3F800000; mul.ftz.f32 %f773, %f772, 0f3E2AAAAB; mul.ftz.f32 %f774, %f759, 0f3E2AAAAB; mul.ftz.f32 %f775, %f758, 0f3F000000; sub.ftz.f32 %f776, %f755, %f775; add.ftz.f32 %f777, %f776, 0fBF000000; mul.ftz.f32 %f778, %f758, 0f3FC00000; add.ftz.f32 %f779, %f755, %f755; sub.ftz.f32 %f780, %f778, %f779; sub.ftz.f32 %f781, %f755, %f778; add.ftz.f32 %f782, %f781, 0f3F000000; mul.ftz.f32 %f783, %f757, %f757; mul.ftz.f32 %f784, %f757, %f783; sub.ftz.f32 %f785, %f757, %f783; mul.ftz.f32 %f786, %f785, 0f40400000; sub.ftz.f32 %f787, %f749, %f786; sub.ftz.f32 %f788, %f787, %f784; mul.ftz.f32 %f789, %f788, 0f3E2AAAAB; mul.ftz.f32 %f790, %f783, 0f40C00000; sub.ftz.f32 %f791, %f766, %f790; fma.rn.ftz.f32 %f792, %f784, 0f40400000, %f791; mul.ftz.f32 %f793, %f792, 0f3E2AAAAB; add.ftz.f32 %f794, %f757, %f783; sub.ftz.f32 %f795, %f794, %f784; fma.rn.ftz.f32 %f796, %f795, 0f40400000, 0f3F800000; mul.ftz.f32 %f797, %f796, 0f3E2AAAAB; mul.ftz.f32 %f798, %f784, 0f3E2AAAAB; mul.ftz.f32 %f799, %f783, 0f3F000000; sub.ftz.f32 %f800, %f757, %f799; add.ftz.f32 %f801, %f800, 0fBF000000; mul.ftz.f32 %f802, %f783, 0f3FC00000; add.ftz.f32 %f803, %f757, %f757; sub.ftz.f32 %f804, %f802, %f803; sub.ftz.f32 %f805, %f757, %f802; add.ftz.f32 %f806, %f805, 0f3F000000; and.b32 %r170, %r159, 256; setp.eq.s32 %p28, %r170, 0; sub.ftz.f32 %f807, %f797, %f789; fma.rn.ftz.f32 %f808, %f789, 0f40000000, %f793; selp.f32 %f809, %f797, %f807, %p28; selp.f32 %f810, %f793, %f808, %p28; selp.f32 %f811, %f789, 0f00000000, %p28; and.b32 %r171, %r159, 512; setp.eq.s32 %p29, %r171, 0; sub.ftz.f32 %f812, %f769, %f774; fma.rn.ftz.f32 %f813, %f774, 0f40000000, %f773; selp.f32 %f814, %f774, 0f00000000, %p29; selp.f32 %f815, %f773, %f813, %p29; selp.f32 %f816, %f769, %f812, %p29; and.b32 %r172, %r159, 1024; setp.eq.s32 %p30, %r172, 0; sub.ftz.f32 %f817, %f810, %f798; fma.rn.ftz.f32 %f818, %f798, 0f40000000, %f809; selp.f32 %f819, %f798, 0f00000000, %p30; selp.f32 %f820, %f809, %f818, %p30; selp.f32 %f821, %f810, %f817, %p30; and.b32 %r173, %r159, 2048; setp.eq.s32 %p31, %r173, 0; sub.ftz.f32 %f822, %f815, %f764; fma.rn.ftz.f32 %f823, %f764, 0f40000000, %f816; selp.f32 %f824, %f815, %f822, %p31; selp.f32 %f825, %f816, %f823, %p31; selp.f32 %f826, %f764, 0f00000000, %p31; sub.ftz.f32 %f827, %f806, %f801; fma.rn.ftz.f32 %f828, %f801, 0f40000000, %f804; selp.f32 %f829, %f806, %f827, %p28; selp.f32 %f830, %f804, %f828, %p28; selp.f32 %f831, %f801, 0f00000000, %p28; sub.ftz.f32 %f832, %f780, %f775; fma.rn.ftz.f32 %f833, %f775, 0f40000000, %f782; selp.f32 %f834, %f775, 0f00000000, %p29; selp.f32 %f835, %f782, %f833, %p29; selp.f32 %f836, %f780, %f832, %p29; sub.ftz.f32 %f837, %f830, %f799; fma.rn.ftz.f32 %f838, %f799, 0f40000000, %f829; selp.f32 %f839, %f799, 0f00000000, %p30; selp.f32 %f840, %f829, %f838, %p30; selp.f32 %f841, %f830, %f837, %p30; sub.ftz.f32 %f842, %f835, %f777; fma.rn.ftz.f32 %f843, %f777, 0f40000000, %f836; selp.f32 %f844, %f835, %f842, %p31; selp.f32 %f845, %f836, %f843, %p31; selp.f32 %f846, %f777, 0f00000000, %p31; mul.ftz.f32 %f847, %f826, %f811; mul.ftz.f32 %f848, %f846, %f811; mul.ftz.f32 %f849, %f746, %f848; mul.ftz.f32 %f850, %f826, %f831; mul.ftz.f32 %f851, %f746, %f850; mul.ftz.f32 %f852, %f825, %f811; mul.ftz.f32 %f853, %f845, %f811; mul.ftz.f32 %f854, %f746, %f853; mul.ftz.f32 %f855, %f825, %f831; mul.ftz.f32 %f856, %f746, %f855; mul.ftz.f32 %f857, %f824, %f811; mul.ftz.f32 %f858, %f844, %f811; mul.ftz.f32 %f859, %f746, %f858; mul.ftz.f32 %f860, %f824, %f831; mul.ftz.f32 %f861, %f746, %f860; mul.ftz.f32 %f862, %f814, %f811; mul.ftz.f32 %f863, %f834, %f811; mul.ftz.f32 %f864, %f746, %f863; mul.ftz.f32 %f865, %f814, %f831; mul.ftz.f32 %f866, %f746, %f865; mul.ftz.f32 %f867, %f826, %f821; mul.ftz.f32 %f868, %f846, %f821; mul.ftz.f32 %f869, %f746, %f868; mul.ftz.f32 %f870, %f826, %f841; mul.ftz.f32 %f871, %f746, %f870; mul.ftz.f32 %f872, %f825, %f821; mul.ftz.f32 %f873, %f845, %f821; mul.ftz.f32 %f874, %f746, %f873; mul.ftz.f32 %f875, %f825, %f841; mul.ftz.f32 %f876, %f746, %f875; mul.ftz.f32 %f877, %f824, %f821; mul.ftz.f32 %f878, %f844, %f821; mul.ftz.f32 %f879, %f746, %f878; mul.ftz.f32 %f880, %f824, %f841; mul.ftz.f32 %f881, %f746, %f880; mul.ftz.f32 %f882, %f814, %f821; mul.ftz.f32 %f883, %f834, %f821; mul.ftz.f32 %f884, %f746, %f883; mul.ftz.f32 %f885, %f814, %f841; mul.ftz.f32 %f886, %f746, %f885; mul.ftz.f32 %f887, %f826, %f820; mul.ftz.f32 %f888, %f846, %f820; mul.ftz.f32 %f889, %f746, %f888; mul.ftz.f32 %f890, %f826, %f840; mul.ftz.f32 %f891, %f746, %f890; mul.ftz.f32 %f892, %f825, %f820; mul.ftz.f32 %f893, %f845, %f820; mul.ftz.f32 %f894, %f746, %f893; mul.ftz.f32 %f895, %f825, %f840; mul.ftz.f32 %f896, %f746, %f895; mul.ftz.f32 %f897, %f824, %f820; mul.ftz.f32 %f898, %f844, %f820; mul.ftz.f32 %f899, %f746, %f898; mul.ftz.f32 %f900, %f824, %f840; mul.ftz.f32 %f901, %f746, %f900; mul.ftz.f32 %f902, %f814, %f820; mul.ftz.f32 %f903, %f834, %f820; mul.ftz.f32 %f904, %f746, %f903; mul.ftz.f32 %f905, %f814, %f840; mul.ftz.f32 %f906, %f746, %f905; mul.ftz.f32 %f907, %f826, %f819; mul.ftz.f32 %f908, %f846, %f819; mul.ftz.f32 %f909, %f746, %f908; mul.ftz.f32 %f910, %f826, %f839; mul.ftz.f32 %f911, %f746, %f910; mul.ftz.f32 %f912, %f825, %f819; mul.ftz.f32 %f913, %f845, %f819; mul.ftz.f32 %f914, %f746, %f913; mul.ftz.f32 %f915, %f825, %f839; mul.ftz.f32 %f916, %f746, %f915; mul.ftz.f32 %f917, %f824, %f819; mul.ftz.f32 %f918, %f844, %f819; mul.ftz.f32 %f919, %f746, %f918; mul.ftz.f32 %f920, %f824, %f839; mul.ftz.f32 %f921, %f746, %f920; mul.ftz.f32 %f922, %f814, %f819; mul.ftz.f32 %f923, %f834, %f819; mul.ftz.f32 %f924, %f746, %f923; mul.ftz.f32 %f925, %f814, %f839; mul.ftz.f32 %f926, %f746, %f925; ld.const.u64 %rd56, [kernel_params+2344]; cvta.to.global.u64 %rd57, %rd56; ld.global.u32 %r174, [%rd53]; add.s32 %r175, %r174, %r96; mul.wide.s32 %rd58, %r175, 16; add.s64 %rd59, %rd57, %rd58; ld.global.v4.f32 {%f927, %f928, %f929, %f930}, [%rd59]; fma.rn.ftz.f32 %f935, %f927, %f847, 0f00000000; fma.rn.ftz.f32 %f936, %f928, %f847, 0f00000000; fma.rn.ftz.f32 %f937, %f929, %f847, 0f00000000; fma.rn.ftz.f32 %f938, %f930, %f847, 0f00000000; fma.rn.ftz.f32 %f939, %f927, %f849, 0f00000000; fma.rn.ftz.f32 %f940, %f928, %f849, 0f00000000; fma.rn.ftz.f32 %f941, %f929, %f849, 0f00000000; fma.rn.ftz.f32 %f942, %f930, %f849, 0f00000000; fma.rn.ftz.f32 %f943, %f927, %f851, 0f00000000; fma.rn.ftz.f32 %f944, %f928, %f851, 0f00000000; fma.rn.ftz.f32 %f945, %f929, %f851, 0f00000000; fma.rn.ftz.f32 %f946, %f930, %f851, 0f00000000; ld.global.u32 %r176, [%rd53+4]; add.s32 %r177, %r176, %r96; mul.wide.s32 %rd60, %r177, 16; add.s64 %rd61, %rd57, %rd60; ld.global.v4.f32 {%f947, %f948, %f949, %f950}, [%rd61]; fma.rn.ftz.f32 %f955, %f947, %f852, %f935; fma.rn.ftz.f32 %f956, %f948, %f852, %f936; fma.rn.ftz.f32 %f957, %f949, %f852, %f937; fma.rn.ftz.f32 %f958, %f950, %f852, %f938; fma.rn.ftz.f32 %f959, %f947, %f854, %f939; fma.rn.ftz.f32 %f960, %f948, %f854, %f940; fma.rn.ftz.f32 %f961, %f949, %f854, %f941; fma.rn.ftz.f32 %f962, %f950, %f854, %f942; fma.rn.ftz.f32 %f963, %f947, %f856, %f943; fma.rn.ftz.f32 %f964, %f948, %f856, %f944; fma.rn.ftz.f32 %f965, %f949, %f856, %f945; fma.rn.ftz.f32 %f966, %f950, %f856, %f946; ld.global.u32 %r178, [%rd53+8]; add.s32 %r179, %r178, %r96; mul.wide.s32 %rd62, %r179, 16; add.s64 %rd63, %rd57, %rd62; ld.global.v4.f32 {%f967, %f968, %f969, %f970}, [%rd63]; fma.rn.ftz.f32 %f975, %f967, %f857, %f955; fma.rn.ftz.f32 %f976, %f968, %f857, %f956; fma.rn.ftz.f32 %f977, %f969, %f857, %f957; fma.rn.ftz.f32 %f978, %f970, %f857, %f958; fma.rn.ftz.f32 %f979, %f967, %f859, %f959; fma.rn.ftz.f32 %f980, %f968, %f859, %f960; fma.rn.ftz.f32 %f981, %f969, %f859, %f961; fma.rn.ftz.f32 %f982, %f970, %f859, %f962; fma.rn.ftz.f32 %f983, %f967, %f861, %f963; fma.rn.ftz.f32 %f984, %f968, %f861, %f964; fma.rn.ftz.f32 %f985, %f969, %f861, %f965; fma.rn.ftz.f32 %f986, %f970, %f861, %f966; ld.global.u32 %r180, [%rd53+12]; add.s32 %r181, %r180, %r96; mul.wide.s32 %rd64, %r181, 16; add.s64 %rd65, %rd57, %rd64; ld.global.v4.f32 {%f987, %f988, %f989, %f990}, [%rd65]; fma.rn.ftz.f32 %f995, %f987, %f862, %f975; fma.rn.ftz.f32 %f996, %f988, %f862, %f976; fma.rn.ftz.f32 %f997, %f989, %f862, %f977; fma.rn.ftz.f32 %f998, %f990, %f862, %f978; fma.rn.ftz.f32 %f999, %f987, %f864, %f979; fma.rn.ftz.f32 %f1000, %f988, %f864, %f980; fma.rn.ftz.f32 %f1001, %f989, %f864, %f981; fma.rn.ftz.f32 %f1002, %f990, %f864, %f982; fma.rn.ftz.f32 %f1003, %f987, %f866, %f983; fma.rn.ftz.f32 %f1004, %f988, %f866, %f984; fma.rn.ftz.f32 %f1005, %f989, %f866, %f985; fma.rn.ftz.f32 %f1006, %f990, %f866, %f986; ld.global.u32 %r182, [%rd53+16]; add.s32 %r183, %r182, %r96; mul.wide.s32 %rd66, %r183, 16; add.s64 %rd67, %rd57, %rd66; ld.global.v4.f32 {%f1007, %f1008, %f1009, %f1010}, [%rd67]; fma.rn.ftz.f32 %f1015, %f1007, %f867, %f995; fma.rn.ftz.f32 %f1016, %f1008, %f867, %f996; fma.rn.ftz.f32 %f1017, %f1009, %f867, %f997; fma.rn.ftz.f32 %f1018, %f1010, %f867, %f998; fma.rn.ftz.f32 %f1019, %f1007, %f869, %f999; fma.rn.ftz.f32 %f1020, %f1008, %f869, %f1000; fma.rn.ftz.f32 %f1021, %f1009, %f869, %f1001; fma.rn.ftz.f32 %f1022, %f1010, %f869, %f1002; fma.rn.ftz.f32 %f1023, %f1007, %f871, %f1003; fma.rn.ftz.f32 %f1024, %f1008, %f871, %f1004; fma.rn.ftz.f32 %f1025, %f1009, %f871, %f1005; fma.rn.ftz.f32 %f1026, %f1010, %f871, %f1006; ld.global.u32 %r184, [%rd53+20]; add.s32 %r185, %r184, %r96; mul.wide.s32 %rd68, %r185, 16; add.s64 %rd69, %rd57, %rd68; ld.global.v4.f32 {%f1027, %f1028, %f1029, %f1030}, [%rd69]; fma.rn.ftz.f32 %f1035, %f1027, %f872, %f1015; fma.rn.ftz.f32 %f1036, %f1028, %f872, %f1016; fma.rn.ftz.f32 %f1037, %f1029, %f872, %f1017; fma.rn.ftz.f32 %f1038, %f1030, %f872, %f1018; fma.rn.ftz.f32 %f1039, %f1027, %f874, %f1019; fma.rn.ftz.f32 %f1040, %f1028, %f874, %f1020; fma.rn.ftz.f32 %f1041, %f1029, %f874, %f1021; fma.rn.ftz.f32 %f1042, %f1030, %f874, %f1022; fma.rn.ftz.f32 %f1043, %f1027, %f876, %f1023; fma.rn.ftz.f32 %f1044, %f1028, %f876, %f1024; fma.rn.ftz.f32 %f1045, %f1029, %f876, %f1025; fma.rn.ftz.f32 %f1046, %f1030, %f876, %f1026; ld.global.u32 %r186, [%rd53+24]; add.s32 %r187, %r186, %r96; mul.wide.s32 %rd70, %r187, 16; add.s64 %rd71, %rd57, %rd70; ld.global.v4.f32 {%f1047, %f1048, %f1049, %f1050}, [%rd71]; fma.rn.ftz.f32 %f1055, %f1047, %f877, %f1035; fma.rn.ftz.f32 %f1056, %f1048, %f877, %f1036; fma.rn.ftz.f32 %f1057, %f1049, %f877, %f1037; fma.rn.ftz.f32 %f1058, %f1050, %f877, %f1038; fma.rn.ftz.f32 %f1059, %f1047, %f879, %f1039; fma.rn.ftz.f32 %f1060, %f1048, %f879, %f1040; fma.rn.ftz.f32 %f1061, %f1049, %f879, %f1041; fma.rn.ftz.f32 %f1062, %f1050, %f879, %f1042; fma.rn.ftz.f32 %f1063, %f1047, %f881, %f1043; fma.rn.ftz.f32 %f1064, %f1048, %f881, %f1044; fma.rn.ftz.f32 %f1065, %f1049, %f881, %f1045; fma.rn.ftz.f32 %f1066, %f1050, %f881, %f1046; ld.global.u32 %r188, [%rd53+28]; add.s32 %r189, %r188, %r96; mul.wide.s32 %rd72, %r189, 16; add.s64 %rd73, %rd57, %rd72; ld.global.v4.f32 {%f1067, %f1068, %f1069, %f1070}, [%rd73]; fma.rn.ftz.f32 %f1075, %f1067, %f882, %f1055; fma.rn.ftz.f32 %f1076, %f1068, %f882, %f1056; fma.rn.ftz.f32 %f1077, %f1069, %f882, %f1057; fma.rn.ftz.f32 %f1078, %f1070, %f882, %f1058; fma.rn.ftz.f32 %f1079, %f1067, %f884, %f1059; fma.rn.ftz.f32 %f1080, %f1068, %f884, %f1060; fma.rn.ftz.f32 %f1081, %f1069, %f884, %f1061; fma.rn.ftz.f32 %f1082, %f1070, %f884, %f1062; fma.rn.ftz.f32 %f1083, %f1067, %f886, %f1063; fma.rn.ftz.f32 %f1084, %f1068, %f886, %f1064; fma.rn.ftz.f32 %f1085, %f1069, %f886, %f1065; fma.rn.ftz.f32 %f1086, %f1070, %f886, %f1066; ld.global.u32 %r190, [%rd53+32]; add.s32 %r191, %r190, %r96; mul.wide.s32 %rd74, %r191, 16; add.s64 %rd75, %rd57, %rd74; ld.global.v4.f32 {%f1087, %f1088, %f1089, %f1090}, [%rd75]; fma.rn.ftz.f32 %f1095, %f1087, %f887, %f1075; fma.rn.ftz.f32 %f1096, %f1088, %f887, %f1076; fma.rn.ftz.f32 %f1097, %f1089, %f887, %f1077; fma.rn.ftz.f32 %f1098, %f1090, %f887, %f1078; fma.rn.ftz.f32 %f1099, %f1087, %f889, %f1079; fma.rn.ftz.f32 %f1100, %f1088, %f889, %f1080; fma.rn.ftz.f32 %f1101, %f1089, %f889, %f1081; fma.rn.ftz.f32 %f1102, %f1090, %f889, %f1082; fma.rn.ftz.f32 %f1103, %f1087, %f891, %f1083; fma.rn.ftz.f32 %f1104, %f1088, %f891, %f1084; fma.rn.ftz.f32 %f1105, %f1089, %f891, %f1085; fma.rn.ftz.f32 %f1106, %f1090, %f891, %f1086; ld.global.u32 %r192, [%rd53+36]; add.s32 %r193, %r192, %r96; mul.wide.s32 %rd76, %r193, 16; add.s64 %rd77, %rd57, %rd76; ld.global.v4.f32 {%f1107, %f1108, %f1109, %f1110}, [%rd77]; fma.rn.ftz.f32 %f1115, %f1107, %f892, %f1095; fma.rn.ftz.f32 %f1116, %f1108, %f892, %f1096; fma.rn.ftz.f32 %f1117, %f1109, %f892, %f1097; fma.rn.ftz.f32 %f1118, %f1110, %f892, %f1098; fma.rn.ftz.f32 %f1119, %f1107, %f894, %f1099; fma.rn.ftz.f32 %f1120, %f1108, %f894, %f1100; fma.rn.ftz.f32 %f1121, %f1109, %f894, %f1101; fma.rn.ftz.f32 %f1122, %f1110, %f894, %f1102; fma.rn.ftz.f32 %f1123, %f1107, %f896, %f1103; fma.rn.ftz.f32 %f1124, %f1108, %f896, %f1104; fma.rn.ftz.f32 %f1125, %f1109, %f896, %f1105; fma.rn.ftz.f32 %f1126, %f1110, %f896, %f1106; ld.global.u32 %r194, [%rd53+40]; add.s32 %r195, %r194, %r96; mul.wide.s32 %rd78, %r195, 16; add.s64 %rd79, %rd57, %rd78; ld.global.v4.f32 {%f1127, %f1128, %f1129, %f1130}, [%rd79]; fma.rn.ftz.f32 %f1135, %f1127, %f897, %f1115; fma.rn.ftz.f32 %f1136, %f1128, %f897, %f1116; fma.rn.ftz.f32 %f1137, %f1129, %f897, %f1117; fma.rn.ftz.f32 %f1138, %f1130, %f897, %f1118; fma.rn.ftz.f32 %f1139, %f1127, %f899, %f1119; fma.rn.ftz.f32 %f1140, %f1128, %f899, %f1120; fma.rn.ftz.f32 %f1141, %f1129, %f899, %f1121; fma.rn.ftz.f32 %f1142, %f1130, %f899, %f1122; fma.rn.ftz.f32 %f1143, %f1127, %f901, %f1123; fma.rn.ftz.f32 %f1144, %f1128, %f901, %f1124; fma.rn.ftz.f32 %f1145, %f1129, %f901, %f1125; fma.rn.ftz.f32 %f1146, %f1130, %f901, %f1126; ld.global.u32 %r196, [%rd53+44]; add.s32 %r197, %r196, %r96; mul.wide.s32 %rd80, %r197, 16; add.s64 %rd81, %rd57, %rd80; ld.global.v4.f32 {%f1147, %f1148, %f1149, %f1150}, [%rd81]; fma.rn.ftz.f32 %f1155, %f1147, %f902, %f1135; fma.rn.ftz.f32 %f1156, %f1148, %f902, %f1136; fma.rn.ftz.f32 %f1157, %f1149, %f902, %f1137; fma.rn.ftz.f32 %f1158, %f1150, %f902, %f1138; fma.rn.ftz.f32 %f1159, %f1147, %f904, %f1139; fma.rn.ftz.f32 %f1160, %f1148, %f904, %f1140; fma.rn.ftz.f32 %f1161, %f1149, %f904, %f1141; fma.rn.ftz.f32 %f1162, %f1150, %f904, %f1142; fma.rn.ftz.f32 %f1163, %f1147, %f906, %f1143; fma.rn.ftz.f32 %f1164, %f1148, %f906, %f1144; fma.rn.ftz.f32 %f1165, %f1149, %f906, %f1145; fma.rn.ftz.f32 %f1166, %f1150, %f906, %f1146; ld.global.u32 %r198, [%rd53+48]; add.s32 %r199, %r198, %r96; mul.wide.s32 %rd82, %r199, 16; add.s64 %rd83, %rd57, %rd82; ld.global.v4.f32 {%f1167, %f1168, %f1169, %f1170}, [%rd83]; fma.rn.ftz.f32 %f1175, %f1167, %f907, %f1155; fma.rn.ftz.f32 %f1176, %f1168, %f907, %f1156; fma.rn.ftz.f32 %f1177, %f1169, %f907, %f1157; fma.rn.ftz.f32 %f1178, %f1170, %f907, %f1158; fma.rn.ftz.f32 %f1179, %f1167, %f909, %f1159; fma.rn.ftz.f32 %f1180, %f1168, %f909, %f1160; fma.rn.ftz.f32 %f1181, %f1169, %f909, %f1161; fma.rn.ftz.f32 %f1182, %f1170, %f909, %f1162; fma.rn.ftz.f32 %f1183, %f1167, %f911, %f1163; fma.rn.ftz.f32 %f1184, %f1168, %f911, %f1164; fma.rn.ftz.f32 %f1185, %f1169, %f911, %f1165; fma.rn.ftz.f32 %f1186, %f1170, %f911, %f1166; ld.global.u32 %r200, [%rd53+52]; add.s32 %r201, %r200, %r96; mul.wide.s32 %rd84, %r201, 16; add.s64 %rd85, %rd57, %rd84; ld.global.v4.f32 {%f1187, %f1188, %f1189, %f1190}, [%rd85]; fma.rn.ftz.f32 %f1195, %f1187, %f912, %f1175; fma.rn.ftz.f32 %f1196, %f1188, %f912, %f1176; fma.rn.ftz.f32 %f1197, %f1189, %f912, %f1177; fma.rn.ftz.f32 %f1198, %f1190, %f912, %f1178; fma.rn.ftz.f32 %f1199, %f1187, %f914, %f1179; fma.rn.ftz.f32 %f1200, %f1188, %f914, %f1180; fma.rn.ftz.f32 %f1201, %f1189, %f914, %f1181; fma.rn.ftz.f32 %f1202, %f1190, %f914, %f1182; fma.rn.ftz.f32 %f1203, %f1187, %f916, %f1183; fma.rn.ftz.f32 %f1204, %f1188, %f916, %f1184; fma.rn.ftz.f32 %f1205, %f1189, %f916, %f1185; fma.rn.ftz.f32 %f1206, %f1190, %f916, %f1186; ld.global.u32 %r202, [%rd53+56]; add.s32 %r203, %r202, %r96; mul.wide.s32 %rd86, %r203, 16; add.s64 %rd87, %rd57, %rd86; ld.global.v4.f32 {%f1207, %f1208, %f1209, %f1210}, [%rd87]; fma.rn.ftz.f32 %f1215, %f1207, %f917, %f1195; fma.rn.ftz.f32 %f1216, %f1208, %f917, %f1196; fma.rn.ftz.f32 %f1217, %f1209, %f917, %f1197; fma.rn.ftz.f32 %f1218, %f1210, %f917, %f1198; fma.rn.ftz.f32 %f1219, %f1207, %f919, %f1199; fma.rn.ftz.f32 %f1220, %f1208, %f919, %f1200; fma.rn.ftz.f32 %f1221, %f1209, %f919, %f1201; fma.rn.ftz.f32 %f1222, %f1210, %f919, %f1202; fma.rn.ftz.f32 %f1223, %f1207, %f921, %f1203; fma.rn.ftz.f32 %f1224, %f1208, %f921, %f1204; fma.rn.ftz.f32 %f1225, %f1209, %f921, %f1205; fma.rn.ftz.f32 %f1226, %f1210, %f921, %f1206; ld.global.u32 %r204, [%rd53+60]; add.s32 %r205, %r204, %r96; mul.wide.s32 %rd88, %r205, 16; add.s64 %rd89, %rd57, %rd88; ld.global.v4.f32 {%f1227, %f1228, %f1229, %f1230}, [%rd89]; fma.rn.ftz.f32 %f2067, %f1227, %f924, %f1219; fma.rn.ftz.f32 %f2068, %f1228, %f924, %f1220; fma.rn.ftz.f32 %f2069, %f1229, %f924, %f1221; fma.rn.ftz.f32 %f2070, %f1230, %f924, %f1222; fma.rn.ftz.f32 %f2063, %f1227, %f926, %f1223; fma.rn.ftz.f32 %f2064, %f1228, %f926, %f1224; fma.rn.ftz.f32 %f2065, %f1229, %f926, %f1225; fma.rn.ftz.f32 %f2066, %f1230, %f926, %f1226; fma.rn.ftz.f32 %f2121, %f1230, %f922, %f1218; fma.rn.ftz.f32 %f2120, %f1229, %f922, %f1217; fma.rn.ftz.f32 %f2119, %f1228, %f922, %f1216; fma.rn.ftz.f32 %f2118, %f1227, %f922, %f1215; bra.uni $L__BB16_210; $L__BB16_66: add.ftz.f32 %f1329, %f90, 0f3D6147AE; mul.ftz.f32 %f1330, %f1329, 0f3F72A76F; lg2.approx.ftz.f32 %f1331, %f1330; mul.ftz.f32 %f1332, %f1331, 0f4019999A; ex2.approx.ftz.f32 %f2015, %f1332; $L__BB16_68: mul.ftz.f32 %f96, %f81, %f74; mul.ftz.f32 %f1333, %f89, %f74; mul.ftz.f32 %f97, %f73, %f1333; mul.ftz.f32 %f1334, %f81, %f82; mul.ftz.f32 %f98, %f73, %f1334; mul.ftz.f32 %f99, %f80, %f74; mul.ftz.f32 %f1335, %f88, %f74; mul.ftz.f32 %f100, %f73, %f1335; mul.ftz.f32 %f1336, %f80, %f82; mul.ftz.f32 %f101, %f73, %f1336; mul.ftz.f32 %f102, %f79, %f74; mul.ftz.f32 %f1337, %f87, %f74; mul.ftz.f32 %f103, %f73, %f1337; mul.ftz.f32 %f1338, %f79, %f82; mul.ftz.f32 %f104, %f73, %f1338; mul.ftz.f32 %f105, %f75, %f74; mul.ftz.f32 %f1339, %f83, %f74; mul.ftz.f32 %f106, %f73, %f1339; mul.ftz.f32 %f1340, %f75, %f82; mul.ftz.f32 %f107, %f73, %f1340; mul.ftz.f32 %f108, %f81, %f78; mul.ftz.f32 %f1341, %f89, %f78; mul.ftz.f32 %f109, %f73, %f1341; mul.ftz.f32 %f1342, %f81, %f86; mul.ftz.f32 %f110, %f73, %f1342; mul.ftz.f32 %f111, %f80, %f78; mul.ftz.f32 %f1343, %f88, %f78; mul.ftz.f32 %f112, %f73, %f1343; mul.ftz.f32 %f1344, %f80, %f86; mul.ftz.f32 %f113, %f73, %f1344; mul.ftz.f32 %f114, %f79, %f78; mul.ftz.f32 %f1345, %f87, %f78; mul.ftz.f32 %f115, %f73, %f1345; mul.ftz.f32 %f1346, %f79, %f86; mul.ftz.f32 %f116, %f73, %f1346; mul.ftz.f32 %f117, %f75, %f78; mul.ftz.f32 %f1347, %f83, %f78; mul.ftz.f32 %f118, %f73, %f1347; mul.ftz.f32 %f1348, %f75, %f86; mul.ftz.f32 %f119, %f73, %f1348; mul.ftz.f32 %f120, %f81, %f77; mul.ftz.f32 %f1349, %f89, %f77; mul.ftz.f32 %f121, %f73, %f1349; mul.ftz.f32 %f1350, %f81, %f85; mul.ftz.f32 %f122, %f73, %f1350; mul.ftz.f32 %f123, %f80, %f77; mul.ftz.f32 %f1351, %f88, %f77; mul.ftz.f32 %f124, %f73, %f1351; mul.ftz.f32 %f1352, %f80, %f85; mul.ftz.f32 %f125, %f73, %f1352; mul.ftz.f32 %f126, %f79, %f77; mul.ftz.f32 %f1353, %f87, %f77; mul.ftz.f32 %f127, %f73, %f1353; mul.ftz.f32 %f1354, %f79, %f85; mul.ftz.f32 %f128, %f73, %f1354; mul.ftz.f32 %f129, %f75, %f77; mul.ftz.f32 %f1355, %f83, %f77; mul.ftz.f32 %f130, %f73, %f1355; mul.ftz.f32 %f1356, %f75, %f85; mul.ftz.f32 %f131, %f73, %f1356; mul.ftz.f32 %f132, %f81, %f76; mul.ftz.f32 %f1357, %f89, %f76; mul.ftz.f32 %f133, %f73, %f1357; mul.ftz.f32 %f1358, %f81, %f84; mul.ftz.f32 %f134, %f73, %f1358; mul.ftz.f32 %f135, %f80, %f76; mul.ftz.f32 %f1359, %f88, %f76; mul.ftz.f32 %f136, %f73, %f1359; mul.ftz.f32 %f1360, %f80, %f84; mul.ftz.f32 %f137, %f73, %f1360; mul.ftz.f32 %f138, %f79, %f76; mul.ftz.f32 %f1361, %f87, %f76; mul.ftz.f32 %f139, %f73, %f1361; mul.ftz.f32 %f1362, %f79, %f84; mul.ftz.f32 %f140, %f73, %f1362; mul.ftz.f32 %f141, %f75, %f76; mul.ftz.f32 %f1363, %f83, %f76; mul.ftz.f32 %f142, %f73, %f1363; mul.ftz.f32 %f1364, %f75, %f84; mul.ftz.f32 %f143, %f73, %f1364; setp.lt.ftz.f32 %p61, %f91, 0f3D25AEE6; @%p61 bra $L__BB16_70; bra.uni $L__BB16_69; $L__BB16_70: mul.ftz.f32 %f2016, %f91, 0f3D9E8391; bra.uni $L__BB16_71; $L__BB16_69: add.ftz.f32 %f1365, %f91, 0f3D6147AE; mul.ftz.f32 %f1366, %f1365, 0f3F72A76F; lg2.approx.ftz.f32 %f1367, %f1366; mul.ftz.f32 %f1368, %f1367, 0f4019999A; ex2.approx.ftz.f32 %f2016, %f1368; $L__BB16_71: setp.lt.ftz.f32 %p62, %f92, 0f3D25AEE6; @%p62 bra $L__BB16_73; bra.uni $L__BB16_72; $L__BB16_73: mul.ftz.f32 %f2017, %f92, 0f3D9E8391; bra.uni $L__BB16_74; $L__BB16_72: add.ftz.f32 %f1369, %f92, 0f3D6147AE; mul.ftz.f32 %f1370, %f1369, 0f3F72A76F; lg2.approx.ftz.f32 %f1371, %f1370; mul.ftz.f32 %f1372, %f1371, 0f4019999A; ex2.approx.ftz.f32 %f2017, %f1372; $L__BB16_74: cvt.rn.f32.u16 %f1373, %rs84; mul.ftz.f32 %f1374, %f1373, 0f3B808081; fma.rn.ftz.f32 %f150, %f2015, %f96, 0f00000000; fma.rn.ftz.f32 %f151, %f2016, %f96, 0f00000000; fma.rn.ftz.f32 %f152, %f2017, %f96, 0f00000000; fma.rn.ftz.f32 %f153, %f1374, %f96, 0f00000000; fma.rn.ftz.f32 %f154, %f2015, %f97, 0f00000000; fma.rn.ftz.f32 %f155, %f2016, %f97, 0f00000000; fma.rn.ftz.f32 %f156, %f2017, %f97, 0f00000000; fma.rn.ftz.f32 %f157, %f1374, %f97, 0f00000000; fma.rn.ftz.f32 %f158, %f2015, %f98, 0f00000000; fma.rn.ftz.f32 %f159, %f2016, %f98, 0f00000000; fma.rn.ftz.f32 %f160, %f2017, %f98, 0f00000000; fma.rn.ftz.f32 %f161, %f1374, %f98, 0f00000000; add.s32 %r273, %r63, %r96; mul.wide.s32 %rd114, %r273, 4; add.s64 %rd115, %rd8, %rd114; ld.global.v4.u8 {%rs86, %rs87, %rs88, %rs89}, [%rd115]; cvt.rn.f32.u16 %f1375, %rs86; mul.ftz.f32 %f162, %f1375, 0f3B808081; cvt.rn.f32.u16 %f1376, %rs87; mul.ftz.f32 %f163, %f1376, 0f3B808081; cvt.rn.f32.u16 %f1377, %rs88; mul.ftz.f32 %f164, %f1377, 0f3B808081; setp.lt.ftz.f32 %p63, %f162, 0f3D25AEE6; @%p63 bra $L__BB16_76; bra.uni $L__BB16_75; $L__BB16_76: mul.ftz.f32 %f2018, %f162, 0f3D9E8391; bra.uni $L__BB16_77; $L__BB16_75: add.ftz.f32 %f1378, %f162, 0f3D6147AE; mul.ftz.f32 %f1379, %f1378, 0f3F72A76F; lg2.approx.ftz.f32 %f1380, %f1379; mul.ftz.f32 %f1381, %f1380, 0f4019999A; ex2.approx.ftz.f32 %f2018, %f1381; $L__BB16_77: setp.lt.ftz.f32 %p64, %f163, 0f3D25AEE6; @%p64 bra $L__BB16_79; bra.uni $L__BB16_78; $L__BB16_79: mul.ftz.f32 %f2019, %f163, 0f3D9E8391; bra.uni $L__BB16_80; $L__BB16_78: add.ftz.f32 %f1382, %f163, 0f3D6147AE; mul.ftz.f32 %f1383, %f1382, 0f3F72A76F; lg2.approx.ftz.f32 %f1384, %f1383; mul.ftz.f32 %f1385, %f1384, 0f4019999A; ex2.approx.ftz.f32 %f2019, %f1385; $L__BB16_80: setp.lt.ftz.f32 %p65, %f164, 0f3D25AEE6; @%p65 bra $L__BB16_82; bra.uni $L__BB16_81; $L__BB16_82: mul.ftz.f32 %f2020, %f164, 0f3D9E8391; bra.uni $L__BB16_83; $L__BB16_81: add.ftz.f32 %f1386, %f164, 0f3D6147AE; mul.ftz.f32 %f1387, %f1386, 0f3F72A76F; lg2.approx.ftz.f32 %f1388, %f1387; mul.ftz.f32 %f1389, %f1388, 0f4019999A; ex2.approx.ftz.f32 %f2020, %f1389; $L__BB16_83: cvt.rn.f32.u16 %f1390, %rs89; mul.ftz.f32 %f1391, %f1390, 0f3B808081; fma.rn.ftz.f32 %f174, %f2018, %f99, %f150; fma.rn.ftz.f32 %f175, %f2019, %f99, %f151; fma.rn.ftz.f32 %f176, %f2020, %f99, %f152; fma.rn.ftz.f32 %f177, %f1391, %f99, %f153; fma.rn.ftz.f32 %f178, %f2018, %f100, %f154; fma.rn.ftz.f32 %f179, %f2019, %f100, %f155; fma.rn.ftz.f32 %f180, %f2020, %f100, %f156; fma.rn.ftz.f32 %f181, %f1391, %f100, %f157; fma.rn.ftz.f32 %f182, %f2018, %f101, %f158; fma.rn.ftz.f32 %f183, %f2019, %f101, %f159; fma.rn.ftz.f32 %f184, %f2020, %f101, %f160; fma.rn.ftz.f32 %f185, %f1391, %f101, %f161; add.s32 %r274, %r64, %r96; mul.wide.s32 %rd116, %r274, 4; add.s64 %rd117, %rd8, %rd116; ld.global.v4.u8 {%rs91, %rs92, %rs93, %rs94}, [%rd117]; cvt.rn.f32.u16 %f1392, %rs91; mul.ftz.f32 %f186, %f1392, 0f3B808081; cvt.rn.f32.u16 %f1393, %rs92; mul.ftz.f32 %f187, %f1393, 0f3B808081; cvt.rn.f32.u16 %f1394, %rs93; mul.ftz.f32 %f188, %f1394, 0f3B808081; setp.lt.ftz.f32 %p66, %f186, 0f3D25AEE6; @%p66 bra $L__BB16_85; bra.uni $L__BB16_84; $L__BB16_85: mul.ftz.f32 %f2021, %f186, 0f3D9E8391; bra.uni $L__BB16_86; $L__BB16_84: add.ftz.f32 %f1395, %f186, 0f3D6147AE; mul.ftz.f32 %f1396, %f1395, 0f3F72A76F; lg2.approx.ftz.f32 %f1397, %f1396; mul.ftz.f32 %f1398, %f1397, 0f4019999A; ex2.approx.ftz.f32 %f2021, %f1398; $L__BB16_86: setp.lt.ftz.f32 %p67, %f187, 0f3D25AEE6; @%p67 bra $L__BB16_88; bra.uni $L__BB16_87; $L__BB16_88: mul.ftz.f32 %f2022, %f187, 0f3D9E8391; bra.uni $L__BB16_89; $L__BB16_87: add.ftz.f32 %f1399, %f187, 0f3D6147AE; mul.ftz.f32 %f1400, %f1399, 0f3F72A76F; lg2.approx.ftz.f32 %f1401, %f1400; mul.ftz.f32 %f1402, %f1401, 0f4019999A; ex2.approx.ftz.f32 %f2022, %f1402; $L__BB16_89: setp.lt.ftz.f32 %p68, %f188, 0f3D25AEE6; @%p68 bra $L__BB16_91; bra.uni $L__BB16_90; $L__BB16_91: mul.ftz.f32 %f2023, %f188, 0f3D9E8391; bra.uni $L__BB16_92; $L__BB16_90: add.ftz.f32 %f1403, %f188, 0f3D6147AE; mul.ftz.f32 %f1404, %f1403, 0f3F72A76F; lg2.approx.ftz.f32 %f1405, %f1404; mul.ftz.f32 %f1406, %f1405, 0f4019999A; ex2.approx.ftz.f32 %f2023, %f1406; $L__BB16_92: cvt.rn.f32.u16 %f1407, %rs94; mul.ftz.f32 %f1408, %f1407, 0f3B808081; fma.rn.ftz.f32 %f198, %f2021, %f102, %f174; fma.rn.ftz.f32 %f199, %f2022, %f102, %f175; fma.rn.ftz.f32 %f200, %f2023, %f102, %f176; fma.rn.ftz.f32 %f201, %f1408, %f102, %f177; fma.rn.ftz.f32 %f202, %f2021, %f103, %f178; fma.rn.ftz.f32 %f203, %f2022, %f103, %f179; fma.rn.ftz.f32 %f204, %f2023, %f103, %f180; fma.rn.ftz.f32 %f205, %f1408, %f103, %f181; fma.rn.ftz.f32 %f206, %f2021, %f104, %f182; fma.rn.ftz.f32 %f207, %f2022, %f104, %f183; fma.rn.ftz.f32 %f208, %f2023, %f104, %f184; fma.rn.ftz.f32 %f209, %f1408, %f104, %f185; add.s32 %r275, %r65, %r96; mul.wide.s32 %rd118, %r275, 4; add.s64 %rd119, %rd8, %rd118; ld.global.v4.u8 {%rs96, %rs97, %rs98, %rs99}, [%rd119]; cvt.rn.f32.u16 %f1409, %rs96; mul.ftz.f32 %f210, %f1409, 0f3B808081; cvt.rn.f32.u16 %f1410, %rs97; mul.ftz.f32 %f211, %f1410, 0f3B808081; cvt.rn.f32.u16 %f1411, %rs98; mul.ftz.f32 %f212, %f1411, 0f3B808081; setp.lt.ftz.f32 %p69, %f210, 0f3D25AEE6; @%p69 bra $L__BB16_94; bra.uni $L__BB16_93; $L__BB16_94: mul.ftz.f32 %f2024, %f210, 0f3D9E8391; bra.uni $L__BB16_95; $L__BB16_93: add.ftz.f32 %f1412, %f210, 0f3D6147AE; mul.ftz.f32 %f1413, %f1412, 0f3F72A76F; lg2.approx.ftz.f32 %f1414, %f1413; mul.ftz.f32 %f1415, %f1414, 0f4019999A; ex2.approx.ftz.f32 %f2024, %f1415; $L__BB16_95: setp.lt.ftz.f32 %p70, %f211, 0f3D25AEE6; @%p70 bra $L__BB16_97; bra.uni $L__BB16_96; $L__BB16_97: mul.ftz.f32 %f2025, %f211, 0f3D9E8391; bra.uni $L__BB16_98; $L__BB16_96: add.ftz.f32 %f1416, %f211, 0f3D6147AE; mul.ftz.f32 %f1417, %f1416, 0f3F72A76F; lg2.approx.ftz.f32 %f1418, %f1417; mul.ftz.f32 %f1419, %f1418, 0f4019999A; ex2.approx.ftz.f32 %f2025, %f1419; $L__BB16_98: setp.lt.ftz.f32 %p71, %f212, 0f3D25AEE6; @%p71 bra $L__BB16_100; bra.uni $L__BB16_99; $L__BB16_100: mul.ftz.f32 %f2026, %f212, 0f3D9E8391; bra.uni $L__BB16_101; $L__BB16_99: add.ftz.f32 %f1420, %f212, 0f3D6147AE; mul.ftz.f32 %f1421, %f1420, 0f3F72A76F; lg2.approx.ftz.f32 %f1422, %f1421; mul.ftz.f32 %f1423, %f1422, 0f4019999A; ex2.approx.ftz.f32 %f2026, %f1423; $L__BB16_101: cvt.rn.f32.u16 %f1424, %rs99; mul.ftz.f32 %f1425, %f1424, 0f3B808081; fma.rn.ftz.f32 %f222, %f2024, %f105, %f198; fma.rn.ftz.f32 %f223, %f2025, %f105, %f199; fma.rn.ftz.f32 %f224, %f2026, %f105, %f200; fma.rn.ftz.f32 %f225, %f1425, %f105, %f201; fma.rn.ftz.f32 %f226, %f2024, %f106, %f202; fma.rn.ftz.f32 %f227, %f2025, %f106, %f203; fma.rn.ftz.f32 %f228, %f2026, %f106, %f204; fma.rn.ftz.f32 %f229, %f1425, %f106, %f205; fma.rn.ftz.f32 %f230, %f2024, %f107, %f206; fma.rn.ftz.f32 %f231, %f2025, %f107, %f207; fma.rn.ftz.f32 %f232, %f2026, %f107, %f208; fma.rn.ftz.f32 %f233, %f1425, %f107, %f209; add.s32 %r276, %r66, %r96; mul.wide.s32 %rd120, %r276, 4; add.s64 %rd121, %rd8, %rd120; ld.global.v4.u8 {%rs101, %rs102, %rs103, %rs104}, [%rd121]; cvt.rn.f32.u16 %f1426, %rs101; mul.ftz.f32 %f234, %f1426, 0f3B808081; cvt.rn.f32.u16 %f1427, %rs102; mul.ftz.f32 %f235, %f1427, 0f3B808081; cvt.rn.f32.u16 %f1428, %rs103; mul.ftz.f32 %f236, %f1428, 0f3B808081; setp.lt.ftz.f32 %p72, %f234, 0f3D25AEE6; @%p72 bra $L__BB16_103; bra.uni $L__BB16_102; $L__BB16_103: mul.ftz.f32 %f2027, %f234, 0f3D9E8391; bra.uni $L__BB16_104; $L__BB16_102: add.ftz.f32 %f1429, %f234, 0f3D6147AE; mul.ftz.f32 %f1430, %f1429, 0f3F72A76F; lg2.approx.ftz.f32 %f1431, %f1430; mul.ftz.f32 %f1432, %f1431, 0f4019999A; ex2.approx.ftz.f32 %f2027, %f1432; $L__BB16_104: setp.lt.ftz.f32 %p73, %f235, 0f3D25AEE6; @%p73 bra $L__BB16_106; bra.uni $L__BB16_105; $L__BB16_106: mul.ftz.f32 %f2028, %f235, 0f3D9E8391; bra.uni $L__BB16_107; $L__BB16_105: add.ftz.f32 %f1433, %f235, 0f3D6147AE; mul.ftz.f32 %f1434, %f1433, 0f3F72A76F; lg2.approx.ftz.f32 %f1435, %f1434; mul.ftz.f32 %f1436, %f1435, 0f4019999A; ex2.approx.ftz.f32 %f2028, %f1436; $L__BB16_107: setp.lt.ftz.f32 %p74, %f236, 0f3D25AEE6; @%p74 bra $L__BB16_109; bra.uni $L__BB16_108; $L__BB16_109: mul.ftz.f32 %f2029, %f236, 0f3D9E8391; bra.uni $L__BB16_110; $L__BB16_108: add.ftz.f32 %f1437, %f236, 0f3D6147AE; mul.ftz.f32 %f1438, %f1437, 0f3F72A76F; lg2.approx.ftz.f32 %f1439, %f1438; mul.ftz.f32 %f1440, %f1439, 0f4019999A; ex2.approx.ftz.f32 %f2029, %f1440; $L__BB16_110: cvt.rn.f32.u16 %f1441, %rs104; mul.ftz.f32 %f1442, %f1441, 0f3B808081; fma.rn.ftz.f32 %f246, %f2027, %f108, %f222; fma.rn.ftz.f32 %f247, %f2028, %f108, %f223; fma.rn.ftz.f32 %f248, %f2029, %f108, %f224; fma.rn.ftz.f32 %f249, %f1442, %f108, %f225; fma.rn.ftz.f32 %f250, %f2027, %f109, %f226; fma.rn.ftz.f32 %f251, %f2028, %f109, %f227; fma.rn.ftz.f32 %f252, %f2029, %f109, %f228; fma.rn.ftz.f32 %f253, %f1442, %f109, %f229; fma.rn.ftz.f32 %f254, %f2027, %f110, %f230; fma.rn.ftz.f32 %f255, %f2028, %f110, %f231; fma.rn.ftz.f32 %f256, %f2029, %f110, %f232; fma.rn.ftz.f32 %f257, %f1442, %f110, %f233; add.s32 %r277, %r67, %r96; mul.wide.s32 %rd122, %r277, 4; add.s64 %rd123, %rd8, %rd122; ld.global.v4.u8 {%rs106, %rs107, %rs108, %rs109}, [%rd123]; cvt.rn.f32.u16 %f1443, %rs106; mul.ftz.f32 %f258, %f1443, 0f3B808081; cvt.rn.f32.u16 %f1444, %rs107; mul.ftz.f32 %f259, %f1444, 0f3B808081; cvt.rn.f32.u16 %f1445, %rs108; mul.ftz.f32 %f260, %f1445, 0f3B808081; setp.lt.ftz.f32 %p75, %f258, 0f3D25AEE6; @%p75 bra $L__BB16_112; bra.uni $L__BB16_111; $L__BB16_112: mul.ftz.f32 %f2030, %f258, 0f3D9E8391; bra.uni $L__BB16_113; $L__BB16_111: add.ftz.f32 %f1446, %f258, 0f3D6147AE; mul.ftz.f32 %f1447, %f1446, 0f3F72A76F; lg2.approx.ftz.f32 %f1448, %f1447; mul.ftz.f32 %f1449, %f1448, 0f4019999A; ex2.approx.ftz.f32 %f2030, %f1449; $L__BB16_113: setp.lt.ftz.f32 %p76, %f259, 0f3D25AEE6; @%p76 bra $L__BB16_115; bra.uni $L__BB16_114; $L__BB16_115: mul.ftz.f32 %f2031, %f259, 0f3D9E8391; bra.uni $L__BB16_116; $L__BB16_114: add.ftz.f32 %f1450, %f259, 0f3D6147AE; mul.ftz.f32 %f1451, %f1450, 0f3F72A76F; lg2.approx.ftz.f32 %f1452, %f1451; mul.ftz.f32 %f1453, %f1452, 0f4019999A; ex2.approx.ftz.f32 %f2031, %f1453; $L__BB16_116: setp.lt.ftz.f32 %p77, %f260, 0f3D25AEE6; @%p77 bra $L__BB16_118; bra.uni $L__BB16_117; $L__BB16_118: mul.ftz.f32 %f2032, %f260, 0f3D9E8391; bra.uni $L__BB16_119; $L__BB16_117: add.ftz.f32 %f1454, %f260, 0f3D6147AE; mul.ftz.f32 %f1455, %f1454, 0f3F72A76F; lg2.approx.ftz.f32 %f1456, %f1455; mul.ftz.f32 %f1457, %f1456, 0f4019999A; ex2.approx.ftz.f32 %f2032, %f1457; $L__BB16_119: cvt.rn.f32.u16 %f1458, %rs109; mul.ftz.f32 %f1459, %f1458, 0f3B808081; fma.rn.ftz.f32 %f270, %f2030, %f111, %f246; fma.rn.ftz.f32 %f271, %f2031, %f111, %f247; fma.rn.ftz.f32 %f272, %f2032, %f111, %f248; fma.rn.ftz.f32 %f273, %f1459, %f111, %f249; fma.rn.ftz.f32 %f274, %f2030, %f112, %f250; fma.rn.ftz.f32 %f275, %f2031, %f112, %f251; fma.rn.ftz.f32 %f276, %f2032, %f112, %f252; fma.rn.ftz.f32 %f277, %f1459, %f112, %f253; fma.rn.ftz.f32 %f278, %f2030, %f113, %f254; fma.rn.ftz.f32 %f279, %f2031, %f113, %f255; fma.rn.ftz.f32 %f280, %f2032, %f113, %f256; fma.rn.ftz.f32 %f281, %f1459, %f113, %f257; add.s32 %r278, %r68, %r96; mul.wide.s32 %rd124, %r278, 4; add.s64 %rd125, %rd8, %rd124; ld.global.v4.u8 {%rs111, %rs112, %rs113, %rs114}, [%rd125]; cvt.rn.f32.u16 %f1460, %rs111; mul.ftz.f32 %f282, %f1460, 0f3B808081; cvt.rn.f32.u16 %f1461, %rs112; mul.ftz.f32 %f283, %f1461, 0f3B808081; cvt.rn.f32.u16 %f1462, %rs113; mul.ftz.f32 %f284, %f1462, 0f3B808081; setp.lt.ftz.f32 %p78, %f282, 0f3D25AEE6; @%p78 bra $L__BB16_121; bra.uni $L__BB16_120; $L__BB16_121: mul.ftz.f32 %f2033, %f282, 0f3D9E8391; bra.uni $L__BB16_122; $L__BB16_120: add.ftz.f32 %f1463, %f282, 0f3D6147AE; mul.ftz.f32 %f1464, %f1463, 0f3F72A76F; lg2.approx.ftz.f32 %f1465, %f1464; mul.ftz.f32 %f1466, %f1465, 0f4019999A; ex2.approx.ftz.f32 %f2033, %f1466; $L__BB16_122: setp.lt.ftz.f32 %p79, %f283, 0f3D25AEE6; @%p79 bra $L__BB16_124; bra.uni $L__BB16_123; $L__BB16_124: mul.ftz.f32 %f2034, %f283, 0f3D9E8391; bra.uni $L__BB16_125; $L__BB16_123: add.ftz.f32 %f1467, %f283, 0f3D6147AE; mul.ftz.f32 %f1468, %f1467, 0f3F72A76F; lg2.approx.ftz.f32 %f1469, %f1468; mul.ftz.f32 %f1470, %f1469, 0f4019999A; ex2.approx.ftz.f32 %f2034, %f1470; $L__BB16_125: setp.lt.ftz.f32 %p80, %f284, 0f3D25AEE6; @%p80 bra $L__BB16_127; bra.uni $L__BB16_126; $L__BB16_127: mul.ftz.f32 %f2035, %f284, 0f3D9E8391; bra.uni $L__BB16_128; $L__BB16_126: add.ftz.f32 %f1471, %f284, 0f3D6147AE; mul.ftz.f32 %f1472, %f1471, 0f3F72A76F; lg2.approx.ftz.f32 %f1473, %f1472; mul.ftz.f32 %f1474, %f1473, 0f4019999A; ex2.approx.ftz.f32 %f2035, %f1474; $L__BB16_128: cvt.rn.f32.u16 %f1475, %rs114; mul.ftz.f32 %f1476, %f1475, 0f3B808081; fma.rn.ftz.f32 %f294, %f2033, %f114, %f270; fma.rn.ftz.f32 %f295, %f2034, %f114, %f271; fma.rn.ftz.f32 %f296, %f2035, %f114, %f272; fma.rn.ftz.f32 %f297, %f1476, %f114, %f273; fma.rn.ftz.f32 %f298, %f2033, %f115, %f274; fma.rn.ftz.f32 %f299, %f2034, %f115, %f275; fma.rn.ftz.f32 %f300, %f2035, %f115, %f276; fma.rn.ftz.f32 %f301, %f1476, %f115, %f277; fma.rn.ftz.f32 %f302, %f2033, %f116, %f278; fma.rn.ftz.f32 %f303, %f2034, %f116, %f279; fma.rn.ftz.f32 %f304, %f2035, %f116, %f280; fma.rn.ftz.f32 %f305, %f1476, %f116, %f281; add.s32 %r279, %r69, %r96; mul.wide.s32 %rd126, %r279, 4; add.s64 %rd127, %rd8, %rd126; ld.global.v4.u8 {%rs116, %rs117, %rs118, %rs119}, [%rd127]; cvt.rn.f32.u16 %f1477, %rs116; mul.ftz.f32 %f306, %f1477, 0f3B808081; cvt.rn.f32.u16 %f1478, %rs117; mul.ftz.f32 %f307, %f1478, 0f3B808081; cvt.rn.f32.u16 %f1479, %rs118; mul.ftz.f32 %f308, %f1479, 0f3B808081; setp.lt.ftz.f32 %p81, %f306, 0f3D25AEE6; @%p81 bra $L__BB16_130; bra.uni $L__BB16_129; $L__BB16_130: mul.ftz.f32 %f2036, %f306, 0f3D9E8391; bra.uni $L__BB16_131; $L__BB16_129: add.ftz.f32 %f1480, %f306, 0f3D6147AE; mul.ftz.f32 %f1481, %f1480, 0f3F72A76F; lg2.approx.ftz.f32 %f1482, %f1481; mul.ftz.f32 %f1483, %f1482, 0f4019999A; ex2.approx.ftz.f32 %f2036, %f1483; $L__BB16_131: setp.lt.ftz.f32 %p82, %f307, 0f3D25AEE6; @%p82 bra $L__BB16_133; bra.uni $L__BB16_132; $L__BB16_133: mul.ftz.f32 %f2037, %f307, 0f3D9E8391; bra.uni $L__BB16_134; $L__BB16_132: add.ftz.f32 %f1484, %f307, 0f3D6147AE; mul.ftz.f32 %f1485, %f1484, 0f3F72A76F; lg2.approx.ftz.f32 %f1486, %f1485; mul.ftz.f32 %f1487, %f1486, 0f4019999A; ex2.approx.ftz.f32 %f2037, %f1487; $L__BB16_134: setp.lt.ftz.f32 %p83, %f308, 0f3D25AEE6; @%p83 bra $L__BB16_136; bra.uni $L__BB16_135; $L__BB16_136: mul.ftz.f32 %f2038, %f308, 0f3D9E8391; bra.uni $L__BB16_137; $L__BB16_135: add.ftz.f32 %f1488, %f308, 0f3D6147AE; mul.ftz.f32 %f1489, %f1488, 0f3F72A76F; lg2.approx.ftz.f32 %f1490, %f1489; mul.ftz.f32 %f1491, %f1490, 0f4019999A; ex2.approx.ftz.f32 %f2038, %f1491; $L__BB16_137: cvt.rn.f32.u16 %f1492, %rs119; mul.ftz.f32 %f1493, %f1492, 0f3B808081; fma.rn.ftz.f32 %f318, %f2036, %f117, %f294; fma.rn.ftz.f32 %f319, %f2037, %f117, %f295; fma.rn.ftz.f32 %f320, %f2038, %f117, %f296; fma.rn.ftz.f32 %f321, %f1493, %f117, %f297; fma.rn.ftz.f32 %f322, %f2036, %f118, %f298; fma.rn.ftz.f32 %f323, %f2037, %f118, %f299; fma.rn.ftz.f32 %f324, %f2038, %f118, %f300; fma.rn.ftz.f32 %f325, %f1493, %f118, %f301; fma.rn.ftz.f32 %f326, %f2036, %f119, %f302; fma.rn.ftz.f32 %f327, %f2037, %f119, %f303; fma.rn.ftz.f32 %f328, %f2038, %f119, %f304; fma.rn.ftz.f32 %f329, %f1493, %f119, %f305; add.s32 %r280, %r70, %r96; mul.wide.s32 %rd128, %r280, 4; add.s64 %rd129, %rd8, %rd128; ld.global.v4.u8 {%rs121, %rs122, %rs123, %rs124}, [%rd129]; cvt.rn.f32.u16 %f1494, %rs121; mul.ftz.f32 %f330, %f1494, 0f3B808081; cvt.rn.f32.u16 %f1495, %rs122; mul.ftz.f32 %f331, %f1495, 0f3B808081; cvt.rn.f32.u16 %f1496, %rs123; mul.ftz.f32 %f332, %f1496, 0f3B808081; setp.lt.ftz.f32 %p84, %f330, 0f3D25AEE6; @%p84 bra $L__BB16_139; bra.uni $L__BB16_138; $L__BB16_139: mul.ftz.f32 %f2039, %f330, 0f3D9E8391; bra.uni $L__BB16_140; $L__BB16_138: add.ftz.f32 %f1497, %f330, 0f3D6147AE; mul.ftz.f32 %f1498, %f1497, 0f3F72A76F; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f1499, 0f4019999A; ex2.approx.ftz.f32 %f2039, %f1500; $L__BB16_140: setp.lt.ftz.f32 %p85, %f331, 0f3D25AEE6; @%p85 bra $L__BB16_142; bra.uni $L__BB16_141; $L__BB16_142: mul.ftz.f32 %f2040, %f331, 0f3D9E8391; bra.uni $L__BB16_143; $L__BB16_141: add.ftz.f32 %f1501, %f331, 0f3D6147AE; mul.ftz.f32 %f1502, %f1501, 0f3F72A76F; lg2.approx.ftz.f32 %f1503, %f1502; mul.ftz.f32 %f1504, %f1503, 0f4019999A; ex2.approx.ftz.f32 %f2040, %f1504; $L__BB16_143: setp.lt.ftz.f32 %p86, %f332, 0f3D25AEE6; @%p86 bra $L__BB16_145; bra.uni $L__BB16_144; $L__BB16_145: mul.ftz.f32 %f2041, %f332, 0f3D9E8391; bra.uni $L__BB16_146; $L__BB16_144: add.ftz.f32 %f1505, %f332, 0f3D6147AE; mul.ftz.f32 %f1506, %f1505, 0f3F72A76F; lg2.approx.ftz.f32 %f1507, %f1506; mul.ftz.f32 %f1508, %f1507, 0f4019999A; ex2.approx.ftz.f32 %f2041, %f1508; $L__BB16_146: cvt.rn.f32.u16 %f1509, %rs124; mul.ftz.f32 %f1510, %f1509, 0f3B808081; fma.rn.ftz.f32 %f342, %f2039, %f120, %f318; fma.rn.ftz.f32 %f343, %f2040, %f120, %f319; fma.rn.ftz.f32 %f344, %f2041, %f120, %f320; fma.rn.ftz.f32 %f345, %f1510, %f120, %f321; fma.rn.ftz.f32 %f346, %f2039, %f121, %f322; fma.rn.ftz.f32 %f347, %f2040, %f121, %f323; fma.rn.ftz.f32 %f348, %f2041, %f121, %f324; fma.rn.ftz.f32 %f349, %f1510, %f121, %f325; fma.rn.ftz.f32 %f350, %f2039, %f122, %f326; fma.rn.ftz.f32 %f351, %f2040, %f122, %f327; fma.rn.ftz.f32 %f352, %f2041, %f122, %f328; fma.rn.ftz.f32 %f353, %f1510, %f122, %f329; add.s32 %r281, %r71, %r96; mul.wide.s32 %rd130, %r281, 4; add.s64 %rd131, %rd8, %rd130; ld.global.v4.u8 {%rs126, %rs127, %rs128, %rs129}, [%rd131]; cvt.rn.f32.u16 %f1511, %rs126; mul.ftz.f32 %f354, %f1511, 0f3B808081; cvt.rn.f32.u16 %f1512, %rs127; mul.ftz.f32 %f355, %f1512, 0f3B808081; cvt.rn.f32.u16 %f1513, %rs128; mul.ftz.f32 %f356, %f1513, 0f3B808081; setp.lt.ftz.f32 %p87, %f354, 0f3D25AEE6; @%p87 bra $L__BB16_148; bra.uni $L__BB16_147; $L__BB16_148: mul.ftz.f32 %f2042, %f354, 0f3D9E8391; bra.uni $L__BB16_149; $L__BB16_147: add.ftz.f32 %f1514, %f354, 0f3D6147AE; mul.ftz.f32 %f1515, %f1514, 0f3F72A76F; lg2.approx.ftz.f32 %f1516, %f1515; mul.ftz.f32 %f1517, %f1516, 0f4019999A; ex2.approx.ftz.f32 %f2042, %f1517; $L__BB16_149: setp.lt.ftz.f32 %p88, %f355, 0f3D25AEE6; @%p88 bra $L__BB16_151; bra.uni $L__BB16_150; $L__BB16_151: mul.ftz.f32 %f2043, %f355, 0f3D9E8391; bra.uni $L__BB16_152; $L__BB16_150: add.ftz.f32 %f1518, %f355, 0f3D6147AE; mul.ftz.f32 %f1519, %f1518, 0f3F72A76F; lg2.approx.ftz.f32 %f1520, %f1519; mul.ftz.f32 %f1521, %f1520, 0f4019999A; ex2.approx.ftz.f32 %f2043, %f1521; $L__BB16_152: setp.lt.ftz.f32 %p89, %f356, 0f3D25AEE6; @%p89 bra $L__BB16_154; bra.uni $L__BB16_153; $L__BB16_154: mul.ftz.f32 %f2044, %f356, 0f3D9E8391; bra.uni $L__BB16_155; $L__BB16_153: add.ftz.f32 %f1522, %f356, 0f3D6147AE; mul.ftz.f32 %f1523, %f1522, 0f3F72A76F; lg2.approx.ftz.f32 %f1524, %f1523; mul.ftz.f32 %f1525, %f1524, 0f4019999A; ex2.approx.ftz.f32 %f2044, %f1525; $L__BB16_155: cvt.rn.f32.u16 %f1526, %rs129; mul.ftz.f32 %f1527, %f1526, 0f3B808081; fma.rn.ftz.f32 %f366, %f2042, %f123, %f342; fma.rn.ftz.f32 %f367, %f2043, %f123, %f343; fma.rn.ftz.f32 %f368, %f2044, %f123, %f344; fma.rn.ftz.f32 %f369, %f1527, %f123, %f345; fma.rn.ftz.f32 %f370, %f2042, %f124, %f346; fma.rn.ftz.f32 %f371, %f2043, %f124, %f347; fma.rn.ftz.f32 %f372, %f2044, %f124, %f348; fma.rn.ftz.f32 %f373, %f1527, %f124, %f349; fma.rn.ftz.f32 %f374, %f2042, %f125, %f350; fma.rn.ftz.f32 %f375, %f2043, %f125, %f351; fma.rn.ftz.f32 %f376, %f2044, %f125, %f352; fma.rn.ftz.f32 %f377, %f1527, %f125, %f353; add.s32 %r282, %r72, %r96; mul.wide.s32 %rd132, %r282, 4; add.s64 %rd133, %rd8, %rd132; ld.global.v4.u8 {%rs131, %rs132, %rs133, %rs134}, [%rd133]; cvt.rn.f32.u16 %f1528, %rs131; mul.ftz.f32 %f378, %f1528, 0f3B808081; cvt.rn.f32.u16 %f1529, %rs132; mul.ftz.f32 %f379, %f1529, 0f3B808081; cvt.rn.f32.u16 %f1530, %rs133; mul.ftz.f32 %f380, %f1530, 0f3B808081; setp.lt.ftz.f32 %p90, %f378, 0f3D25AEE6; @%p90 bra $L__BB16_157; bra.uni $L__BB16_156; $L__BB16_157: mul.ftz.f32 %f2045, %f378, 0f3D9E8391; bra.uni $L__BB16_158; $L__BB16_156: add.ftz.f32 %f1531, %f378, 0f3D6147AE; mul.ftz.f32 %f1532, %f1531, 0f3F72A76F; lg2.approx.ftz.f32 %f1533, %f1532; mul.ftz.f32 %f1534, %f1533, 0f4019999A; ex2.approx.ftz.f32 %f2045, %f1534; $L__BB16_158: setp.lt.ftz.f32 %p91, %f379, 0f3D25AEE6; @%p91 bra $L__BB16_160; bra.uni $L__BB16_159; $L__BB16_160: mul.ftz.f32 %f2046, %f379, 0f3D9E8391; bra.uni $L__BB16_161; $L__BB16_159: add.ftz.f32 %f1535, %f379, 0f3D6147AE; mul.ftz.f32 %f1536, %f1535, 0f3F72A76F; lg2.approx.ftz.f32 %f1537, %f1536; mul.ftz.f32 %f1538, %f1537, 0f4019999A; ex2.approx.ftz.f32 %f2046, %f1538; $L__BB16_161: setp.lt.ftz.f32 %p92, %f380, 0f3D25AEE6; @%p92 bra $L__BB16_163; bra.uni $L__BB16_162; $L__BB16_163: mul.ftz.f32 %f2047, %f380, 0f3D9E8391; bra.uni $L__BB16_164; $L__BB16_162: add.ftz.f32 %f1539, %f380, 0f3D6147AE; mul.ftz.f32 %f1540, %f1539, 0f3F72A76F; lg2.approx.ftz.f32 %f1541, %f1540; mul.ftz.f32 %f1542, %f1541, 0f4019999A; ex2.approx.ftz.f32 %f2047, %f1542; $L__BB16_164: cvt.rn.f32.u16 %f1543, %rs134; mul.ftz.f32 %f1544, %f1543, 0f3B808081; fma.rn.ftz.f32 %f390, %f2045, %f126, %f366; fma.rn.ftz.f32 %f391, %f2046, %f126, %f367; fma.rn.ftz.f32 %f392, %f2047, %f126, %f368; fma.rn.ftz.f32 %f393, %f1544, %f126, %f369; fma.rn.ftz.f32 %f394, %f2045, %f127, %f370; fma.rn.ftz.f32 %f395, %f2046, %f127, %f371; fma.rn.ftz.f32 %f396, %f2047, %f127, %f372; fma.rn.ftz.f32 %f397, %f1544, %f127, %f373; fma.rn.ftz.f32 %f398, %f2045, %f128, %f374; fma.rn.ftz.f32 %f399, %f2046, %f128, %f375; fma.rn.ftz.f32 %f400, %f2047, %f128, %f376; fma.rn.ftz.f32 %f401, %f1544, %f128, %f377; add.s32 %r283, %r73, %r96; mul.wide.s32 %rd134, %r283, 4; add.s64 %rd135, %rd8, %rd134; ld.global.v4.u8 {%rs136, %rs137, %rs138, %rs139}, [%rd135]; cvt.rn.f32.u16 %f1545, %rs136; mul.ftz.f32 %f402, %f1545, 0f3B808081; cvt.rn.f32.u16 %f1546, %rs137; mul.ftz.f32 %f403, %f1546, 0f3B808081; cvt.rn.f32.u16 %f1547, %rs138; mul.ftz.f32 %f404, %f1547, 0f3B808081; setp.lt.ftz.f32 %p93, %f402, 0f3D25AEE6; @%p93 bra $L__BB16_166; bra.uni $L__BB16_165; $L__BB16_166: mul.ftz.f32 %f2048, %f402, 0f3D9E8391; bra.uni $L__BB16_167; $L__BB16_165: add.ftz.f32 %f1548, %f402, 0f3D6147AE; mul.ftz.f32 %f1549, %f1548, 0f3F72A76F; lg2.approx.ftz.f32 %f1550, %f1549; mul.ftz.f32 %f1551, %f1550, 0f4019999A; ex2.approx.ftz.f32 %f2048, %f1551; $L__BB16_167: setp.lt.ftz.f32 %p94, %f403, 0f3D25AEE6; @%p94 bra $L__BB16_169; bra.uni $L__BB16_168; $L__BB16_169: mul.ftz.f32 %f2049, %f403, 0f3D9E8391; bra.uni $L__BB16_170; $L__BB16_168: add.ftz.f32 %f1552, %f403, 0f3D6147AE; mul.ftz.f32 %f1553, %f1552, 0f3F72A76F; lg2.approx.ftz.f32 %f1554, %f1553; mul.ftz.f32 %f1555, %f1554, 0f4019999A; ex2.approx.ftz.f32 %f2049, %f1555; $L__BB16_170: setp.lt.ftz.f32 %p95, %f404, 0f3D25AEE6; @%p95 bra $L__BB16_172; bra.uni $L__BB16_171; $L__BB16_172: mul.ftz.f32 %f2050, %f404, 0f3D9E8391; bra.uni $L__BB16_173; $L__BB16_171: add.ftz.f32 %f1556, %f404, 0f3D6147AE; mul.ftz.f32 %f1557, %f1556, 0f3F72A76F; lg2.approx.ftz.f32 %f1558, %f1557; mul.ftz.f32 %f1559, %f1558, 0f4019999A; ex2.approx.ftz.f32 %f2050, %f1559; $L__BB16_173: cvt.rn.f32.u16 %f1560, %rs139; mul.ftz.f32 %f1561, %f1560, 0f3B808081; fma.rn.ftz.f32 %f414, %f2048, %f129, %f390; fma.rn.ftz.f32 %f415, %f2049, %f129, %f391; fma.rn.ftz.f32 %f416, %f2050, %f129, %f392; fma.rn.ftz.f32 %f417, %f1561, %f129, %f393; fma.rn.ftz.f32 %f418, %f2048, %f130, %f394; fma.rn.ftz.f32 %f419, %f2049, %f130, %f395; fma.rn.ftz.f32 %f420, %f2050, %f130, %f396; fma.rn.ftz.f32 %f421, %f1561, %f130, %f397; fma.rn.ftz.f32 %f422, %f2048, %f131, %f398; fma.rn.ftz.f32 %f423, %f2049, %f131, %f399; fma.rn.ftz.f32 %f424, %f2050, %f131, %f400; fma.rn.ftz.f32 %f425, %f1561, %f131, %f401; add.s32 %r284, %r74, %r96; mul.wide.s32 %rd136, %r284, 4; add.s64 %rd137, %rd8, %rd136; ld.global.v4.u8 {%rs141, %rs142, %rs143, %rs144}, [%rd137]; cvt.rn.f32.u16 %f1562, %rs141; mul.ftz.f32 %f426, %f1562, 0f3B808081; cvt.rn.f32.u16 %f1563, %rs142; mul.ftz.f32 %f427, %f1563, 0f3B808081; cvt.rn.f32.u16 %f1564, %rs143; mul.ftz.f32 %f428, %f1564, 0f3B808081; setp.lt.ftz.f32 %p96, %f426, 0f3D25AEE6; @%p96 bra $L__BB16_175; bra.uni $L__BB16_174; $L__BB16_175: mul.ftz.f32 %f2051, %f426, 0f3D9E8391; bra.uni $L__BB16_176; $L__BB16_174: add.ftz.f32 %f1565, %f426, 0f3D6147AE; mul.ftz.f32 %f1566, %f1565, 0f3F72A76F; lg2.approx.ftz.f32 %f1567, %f1566; mul.ftz.f32 %f1568, %f1567, 0f4019999A; ex2.approx.ftz.f32 %f2051, %f1568; $L__BB16_176: setp.lt.ftz.f32 %p97, %f427, 0f3D25AEE6; @%p97 bra $L__BB16_178; bra.uni $L__BB16_177; $L__BB16_178: mul.ftz.f32 %f2052, %f427, 0f3D9E8391; bra.uni $L__BB16_179; $L__BB16_177: add.ftz.f32 %f1569, %f427, 0f3D6147AE; mul.ftz.f32 %f1570, %f1569, 0f3F72A76F; lg2.approx.ftz.f32 %f1571, %f1570; mul.ftz.f32 %f1572, %f1571, 0f4019999A; ex2.approx.ftz.f32 %f2052, %f1572; $L__BB16_179: setp.lt.ftz.f32 %p98, %f428, 0f3D25AEE6; @%p98 bra $L__BB16_181; bra.uni $L__BB16_180; $L__BB16_181: mul.ftz.f32 %f2053, %f428, 0f3D9E8391; bra.uni $L__BB16_182; $L__BB16_180: add.ftz.f32 %f1573, %f428, 0f3D6147AE; mul.ftz.f32 %f1574, %f1573, 0f3F72A76F; lg2.approx.ftz.f32 %f1575, %f1574; mul.ftz.f32 %f1576, %f1575, 0f4019999A; ex2.approx.ftz.f32 %f2053, %f1576; $L__BB16_182: cvt.rn.f32.u16 %f1577, %rs144; mul.ftz.f32 %f1578, %f1577, 0f3B808081; fma.rn.ftz.f32 %f438, %f2051, %f132, %f414; fma.rn.ftz.f32 %f439, %f2052, %f132, %f415; fma.rn.ftz.f32 %f440, %f2053, %f132, %f416; fma.rn.ftz.f32 %f441, %f1578, %f132, %f417; fma.rn.ftz.f32 %f442, %f2051, %f133, %f418; fma.rn.ftz.f32 %f443, %f2052, %f133, %f419; fma.rn.ftz.f32 %f444, %f2053, %f133, %f420; fma.rn.ftz.f32 %f445, %f1578, %f133, %f421; fma.rn.ftz.f32 %f446, %f2051, %f134, %f422; fma.rn.ftz.f32 %f447, %f2052, %f134, %f423; fma.rn.ftz.f32 %f448, %f2053, %f134, %f424; fma.rn.ftz.f32 %f449, %f1578, %f134, %f425; add.s32 %r285, %r75, %r96; mul.wide.s32 %rd138, %r285, 4; add.s64 %rd139, %rd8, %rd138; ld.global.v4.u8 {%rs146, %rs147, %rs148, %rs149}, [%rd139]; cvt.rn.f32.u16 %f1579, %rs146; mul.ftz.f32 %f450, %f1579, 0f3B808081; cvt.rn.f32.u16 %f1580, %rs147; mul.ftz.f32 %f451, %f1580, 0f3B808081; cvt.rn.f32.u16 %f1581, %rs148; mul.ftz.f32 %f452, %f1581, 0f3B808081; setp.lt.ftz.f32 %p99, %f450, 0f3D25AEE6; @%p99 bra $L__BB16_184; bra.uni $L__BB16_183; $L__BB16_184: mul.ftz.f32 %f2054, %f450, 0f3D9E8391; bra.uni $L__BB16_185; $L__BB16_183: add.ftz.f32 %f1582, %f450, 0f3D6147AE; mul.ftz.f32 %f1583, %f1582, 0f3F72A76F; lg2.approx.ftz.f32 %f1584, %f1583; mul.ftz.f32 %f1585, %f1584, 0f4019999A; ex2.approx.ftz.f32 %f2054, %f1585; $L__BB16_185: setp.lt.ftz.f32 %p100, %f451, 0f3D25AEE6; @%p100 bra $L__BB16_187; bra.uni $L__BB16_186; $L__BB16_187: mul.ftz.f32 %f2055, %f451, 0f3D9E8391; bra.uni $L__BB16_188; $L__BB16_186: add.ftz.f32 %f1586, %f451, 0f3D6147AE; mul.ftz.f32 %f1587, %f1586, 0f3F72A76F; lg2.approx.ftz.f32 %f1588, %f1587; mul.ftz.f32 %f1589, %f1588, 0f4019999A; ex2.approx.ftz.f32 %f2055, %f1589; $L__BB16_188: setp.lt.ftz.f32 %p101, %f452, 0f3D25AEE6; @%p101 bra $L__BB16_190; bra.uni $L__BB16_189; $L__BB16_190: mul.ftz.f32 %f2056, %f452, 0f3D9E8391; bra.uni $L__BB16_191; $L__BB16_189: add.ftz.f32 %f1590, %f452, 0f3D6147AE; mul.ftz.f32 %f1591, %f1590, 0f3F72A76F; lg2.approx.ftz.f32 %f1592, %f1591; mul.ftz.f32 %f1593, %f1592, 0f4019999A; ex2.approx.ftz.f32 %f2056, %f1593; $L__BB16_191: cvt.rn.f32.u16 %f1594, %rs149; mul.ftz.f32 %f1595, %f1594, 0f3B808081; fma.rn.ftz.f32 %f462, %f2054, %f135, %f438; fma.rn.ftz.f32 %f463, %f2055, %f135, %f439; fma.rn.ftz.f32 %f464, %f2056, %f135, %f440; fma.rn.ftz.f32 %f465, %f1595, %f135, %f441; fma.rn.ftz.f32 %f466, %f2054, %f136, %f442; fma.rn.ftz.f32 %f467, %f2055, %f136, %f443; fma.rn.ftz.f32 %f468, %f2056, %f136, %f444; fma.rn.ftz.f32 %f469, %f1595, %f136, %f445; fma.rn.ftz.f32 %f470, %f2054, %f137, %f446; fma.rn.ftz.f32 %f471, %f2055, %f137, %f447; fma.rn.ftz.f32 %f472, %f2056, %f137, %f448; fma.rn.ftz.f32 %f473, %f1595, %f137, %f449; add.s32 %r286, %r76, %r96; mul.wide.s32 %rd140, %r286, 4; add.s64 %rd141, %rd8, %rd140; ld.global.v4.u8 {%rs151, %rs152, %rs153, %rs154}, [%rd141]; cvt.rn.f32.u16 %f1596, %rs151; mul.ftz.f32 %f474, %f1596, 0f3B808081; cvt.rn.f32.u16 %f1597, %rs152; mul.ftz.f32 %f475, %f1597, 0f3B808081; cvt.rn.f32.u16 %f1598, %rs153; mul.ftz.f32 %f476, %f1598, 0f3B808081; setp.lt.ftz.f32 %p102, %f474, 0f3D25AEE6; @%p102 bra $L__BB16_193; bra.uni $L__BB16_192; $L__BB16_193: mul.ftz.f32 %f2057, %f474, 0f3D9E8391; bra.uni $L__BB16_194; $L__BB16_192: add.ftz.f32 %f1599, %f474, 0f3D6147AE; mul.ftz.f32 %f1600, %f1599, 0f3F72A76F; lg2.approx.ftz.f32 %f1601, %f1600; mul.ftz.f32 %f1602, %f1601, 0f4019999A; ex2.approx.ftz.f32 %f2057, %f1602; $L__BB16_194: setp.lt.ftz.f32 %p103, %f475, 0f3D25AEE6; @%p103 bra $L__BB16_196; bra.uni $L__BB16_195; $L__BB16_196: mul.ftz.f32 %f2058, %f475, 0f3D9E8391; bra.uni $L__BB16_197; $L__BB16_195: add.ftz.f32 %f1603, %f475, 0f3D6147AE; mul.ftz.f32 %f1604, %f1603, 0f3F72A76F; lg2.approx.ftz.f32 %f1605, %f1604; mul.ftz.f32 %f1606, %f1605, 0f4019999A; ex2.approx.ftz.f32 %f2058, %f1606; $L__BB16_197: setp.lt.ftz.f32 %p104, %f476, 0f3D25AEE6; @%p104 bra $L__BB16_199; bra.uni $L__BB16_198; $L__BB16_199: mul.ftz.f32 %f2059, %f476, 0f3D9E8391; bra.uni $L__BB16_200; $L__BB16_198: add.ftz.f32 %f1607, %f476, 0f3D6147AE; mul.ftz.f32 %f1608, %f1607, 0f3F72A76F; lg2.approx.ftz.f32 %f1609, %f1608; mul.ftz.f32 %f1610, %f1609, 0f4019999A; ex2.approx.ftz.f32 %f2059, %f1610; $L__BB16_200: cvt.rn.f32.u16 %f1611, %rs154; mul.ftz.f32 %f1612, %f1611, 0f3B808081; fma.rn.ftz.f32 %f486, %f2057, %f138, %f462; fma.rn.ftz.f32 %f487, %f2058, %f138, %f463; fma.rn.ftz.f32 %f488, %f2059, %f138, %f464; fma.rn.ftz.f32 %f489, %f1612, %f138, %f465; fma.rn.ftz.f32 %f490, %f2057, %f139, %f466; fma.rn.ftz.f32 %f491, %f2058, %f139, %f467; fma.rn.ftz.f32 %f492, %f2059, %f139, %f468; fma.rn.ftz.f32 %f493, %f1612, %f139, %f469; fma.rn.ftz.f32 %f494, %f2057, %f140, %f470; fma.rn.ftz.f32 %f495, %f2058, %f140, %f471; fma.rn.ftz.f32 %f496, %f2059, %f140, %f472; fma.rn.ftz.f32 %f497, %f1612, %f140, %f473; add.s32 %r287, %r77, %r96; mul.wide.s32 %rd142, %r287, 4; add.s64 %rd143, %rd8, %rd142; ld.global.v4.u8 {%rs156, %rs157, %rs158, %rs159}, [%rd143]; cvt.rn.f32.u16 %f1613, %rs156; mul.ftz.f32 %f498, %f1613, 0f3B808081; cvt.rn.f32.u16 %f1614, %rs157; mul.ftz.f32 %f499, %f1614, 0f3B808081; cvt.rn.f32.u16 %f1615, %rs158; mul.ftz.f32 %f500, %f1615, 0f3B808081; setp.lt.ftz.f32 %p105, %f498, 0f3D25AEE6; @%p105 bra $L__BB16_202; bra.uni $L__BB16_201; $L__BB16_202: mul.ftz.f32 %f2060, %f498, 0f3D9E8391; bra.uni $L__BB16_203; $L__BB16_201: add.ftz.f32 %f1616, %f498, 0f3D6147AE; mul.ftz.f32 %f1617, %f1616, 0f3F72A76F; lg2.approx.ftz.f32 %f1618, %f1617; mul.ftz.f32 %f1619, %f1618, 0f4019999A; ex2.approx.ftz.f32 %f2060, %f1619; $L__BB16_203: setp.lt.ftz.f32 %p106, %f499, 0f3D25AEE6; @%p106 bra $L__BB16_205; bra.uni $L__BB16_204; $L__BB16_205: mul.ftz.f32 %f2061, %f499, 0f3D9E8391; bra.uni $L__BB16_206; $L__BB16_204: add.ftz.f32 %f1620, %f499, 0f3D6147AE; mul.ftz.f32 %f1621, %f1620, 0f3F72A76F; lg2.approx.ftz.f32 %f1622, %f1621; mul.ftz.f32 %f1623, %f1622, 0f4019999A; ex2.approx.ftz.f32 %f2061, %f1623; $L__BB16_206: setp.lt.ftz.f32 %p107, %f500, 0f3D25AEE6; @%p107 bra $L__BB16_208; bra.uni $L__BB16_207; $L__BB16_208: mul.ftz.f32 %f2062, %f500, 0f3D9E8391; bra.uni $L__BB16_209; $L__BB16_207: add.ftz.f32 %f1624, %f500, 0f3D6147AE; mul.ftz.f32 %f1625, %f1624, 0f3F72A76F; lg2.approx.ftz.f32 %f1626, %f1625; mul.ftz.f32 %f1627, %f1626, 0f4019999A; ex2.approx.ftz.f32 %f2062, %f1627; $L__BB16_209: cvt.rn.f32.u16 %f1628, %rs159; mul.ftz.f32 %f1629, %f1628, 0f3B808081; fma.rn.ftz.f32 %f2067, %f2060, %f142, %f490; fma.rn.ftz.f32 %f2068, %f2061, %f142, %f491; fma.rn.ftz.f32 %f2069, %f2062, %f142, %f492; fma.rn.ftz.f32 %f2070, %f1629, %f142, %f493; fma.rn.ftz.f32 %f2063, %f2060, %f143, %f494; fma.rn.ftz.f32 %f2064, %f2061, %f143, %f495; fma.rn.ftz.f32 %f2065, %f2062, %f143, %f496; fma.rn.ftz.f32 %f2066, %f1629, %f143, %f497; fma.rn.ftz.f32 %f2120, %f2062, %f141, %f488; fma.rn.ftz.f32 %f2119, %f2061, %f141, %f487; fma.rn.ftz.f32 %f2118, %f2060, %f141, %f486; fma.rn.ftz.f32 %f2121, %f1629, %f141, %f489; $L__BB16_210: ld.param.u64 %rd220, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_3]; ld.param.u64 %rd219, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_2]; or.b64 %rd144, %rd219, %rd220; setp.eq.s64 %p108, %rd144, 0; @%p108 bra $L__BB16_289; ld.param.u64 %rd221, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_2]; setp.eq.s64 %p109, %rd221, 0; @%p109 bra $L__BB16_213; ld.param.u64 %rd226, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_0]; cvta.to.local.u64 %rd225, %rd226; ld.param.u64 %rd222, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_2]; ld.local.f32 %f1630, [%rd225+100]; ld.local.f32 %f1631, [%rd225+108]; mul.ftz.f32 %f1632, %f3, %f1631; fma.rn.ftz.f32 %f1633, %f1, %f1630, %f1632; mul.ftz.f32 %f1634, %f4, %f1631; fma.rn.ftz.f32 %f1635, %f2, %f1630, %f1634; mul.ftz.f32 %f1636, %f2063, %f1635; mul.ftz.f32 %f1637, %f2064, %f1635; mul.ftz.f32 %f1638, %f2065, %f1635; mul.ftz.f32 %f1639, %f2066, %f1635; fma.rn.ftz.f32 %f1640, %f2070, %f1633, %f1639; fma.rn.ftz.f32 %f1641, %f2069, %f1633, %f1638; fma.rn.ftz.f32 %f1642, %f2068, %f1633, %f1637; fma.rn.ftz.f32 %f1643, %f2067, %f1633, %f1636; st.v4.f32 [%rd222], {%f1643, %f1642, %f1641, %f1640}; $L__BB16_213: ld.param.u64 %rd223, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_3]; setp.eq.s64 %p110, %rd223, 0; @%p110 bra $L__BB16_289; ld.param.u64 %rd228, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_0]; cvta.to.local.u64 %rd227, %rd228; ld.param.u64 %rd224, [_Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7__param_3]; ld.local.f32 %f1644, [%rd227+104]; ld.local.f32 %f1645, [%rd227+112]; mul.ftz.f32 %f1646, %f3, %f1645; fma.rn.ftz.f32 %f1647, %f1, %f1644, %f1646; mul.ftz.f32 %f1648, %f4, %f1645; fma.rn.ftz.f32 %f1649, %f2, %f1644, %f1648; mul.ftz.f32 %f1650, %f2063, %f1649; mul.ftz.f32 %f1651, %f2064, %f1649; mul.ftz.f32 %f1652, %f2065, %f1649; mul.ftz.f32 %f1653, %f2066, %f1649; fma.rn.ftz.f32 %f1654, %f2070, %f1647, %f1653; fma.rn.ftz.f32 %f1655, %f2069, %f1647, %f1652; fma.rn.ftz.f32 %f1656, %f2068, %f1647, %f1651; fma.rn.ftz.f32 %f1657, %f2067, %f1647, %f1650; st.v4.f32 [%rd224], {%f1657, %f1656, %f1655, %f1654}; bra.uni $L__BB16_289; $L__BB16_267: cvt.u32.u64 %r312, %rd3; ld.const.u64 %rd181, [kernel_params+2232]; cvta.to.global.u64 %rd182, %rd181; mul.wide.s32 %rd183, %r312, 16; add.s64 %rd184, %rd182, %rd183; ld.global.v4.u32 {%r313, %r314, %r315, %r316}, [%rd184]; ld.const.u64 %rd185, [kernel_params+2248]; cvta.to.global.u64 %rd186, %rd185; mul.wide.u32 %rd187, %r313, 8; add.s64 %rd188, %rd186, %rd187; ld.global.v2.f32 {%f1847, %f1848}, [%rd188]; mul.wide.u32 %rd189, %r314, 8; add.s64 %rd190, %rd186, %rd189; ld.global.v2.f32 {%f1849, %f1850}, [%rd190]; mul.wide.u32 %rd191, %r315, 8; add.s64 %rd192, %rd186, %rd191; ld.global.v2.f32 {%f1851, %f1852}, [%rd192]; ld.const.u64 %rd193, [kernel_params+2288]; cvta.to.global.u64 %rd194, %rd193; mul.wide.s32 %rd195, %r2, 4; add.s64 %rd196, %rd194, %rd195; ld.const.u64 %rd197, [kernel_params+2344]; cvta.to.global.u64 %rd198, %rd197; ld.global.u32 %r320, [%rd196]; add.s32 %r321, %r320, %r96; mul.wide.u32 %rd199, %r321, 16; add.s64 %rd200, %rd198, %rd199; ld.global.v4.f32 {%f1853, %f1854, %f1855, %f1856}, [%rd200]; ld.global.u32 %r322, [%rd196+4]; add.s32 %r323, %r322, %r96; mul.wide.u32 %rd201, %r323, 16; add.s64 %rd202, %rd198, %rd201; ld.global.v4.f32 {%f2114, %f2115, %f2116, %f2117}, [%rd202]; ld.global.u32 %r324, [%rd196+8]; add.s32 %r325, %r324, %r96; mul.wide.u32 %rd203, %r325, 16; add.s64 %rd204, %rd198, %rd203; ld.global.v4.f32 {%f1861, %f1862, %f1863, %f1864}, [%rd204]; ld.global.u32 %r326, [%rd196+12]; add.s32 %r327, %r326, %r96; mul.wide.u32 %rd205, %r327, 16; add.s64 %rd206, %rd198, %rd205; ld.global.v4.f32 {%f2110, %f2111, %f2112, %f2113}, [%rd206]; ld.global.u16 %rs181, [%rd196+20]; setp.eq.s16 %p132, %rs181, 4; @%p132 bra $L__BB16_269; add.ftz.f32 %f1869, %f1853, %f2114; mul.ftz.f32 %f2114, %f1869, 0f3F000000; add.ftz.f32 %f1870, %f1854, %f2115; mul.ftz.f32 %f2115, %f1870, 0f3F000000; add.ftz.f32 %f1871, %f1855, %f2116; mul.ftz.f32 %f2116, %f1871, 0f3F000000; add.ftz.f32 %f1872, %f1856, %f2117; mul.ftz.f32 %f2117, %f1872, 0f3F000000; add.ftz.f32 %f1873, %f1853, %f2110; mul.ftz.f32 %f2110, %f1873, 0f3F000000; add.ftz.f32 %f1874, %f1854, %f2111; mul.ftz.f32 %f2111, %f1874, 0f3F000000; add.ftz.f32 %f1875, %f1855, %f2112; mul.ftz.f32 %f2112, %f1875, 0f3F000000; add.ftz.f32 %f1876, %f1856, %f2113; mul.ftz.f32 %f2113, %f1876, 0f3F000000; $L__BB16_269: sub.ftz.f32 %f1877, %f2114, %f1853; sub.ftz.f32 %f1878, %f2115, %f1854; sub.ftz.f32 %f1879, %f2116, %f1855; sub.ftz.f32 %f1880, %f2117, %f1856; fma.rn.ftz.f32 %f1881, %f1847, %f1877, %f1853; fma.rn.ftz.f32 %f1882, %f1847, %f1878, %f1854; fma.rn.ftz.f32 %f1883, %f1847, %f1879, %f1855; fma.rn.ftz.f32 %f1884, %f1847, %f1880, %f1856; sub.ftz.f32 %f1885, %f1861, %f2110; sub.ftz.f32 %f1886, %f1862, %f2111; sub.ftz.f32 %f1887, %f1863, %f2112; sub.ftz.f32 %f1888, %f1864, %f2113; fma.rn.ftz.f32 %f1889, %f1847, %f1885, %f2110; fma.rn.ftz.f32 %f1890, %f1847, %f1886, %f2111; fma.rn.ftz.f32 %f1891, %f1847, %f1887, %f2112; fma.rn.ftz.f32 %f1892, %f1847, %f1888, %f2113; sub.ftz.f32 %f1893, %f1889, %f1881; sub.ftz.f32 %f1894, %f1890, %f1882; sub.ftz.f32 %f1895, %f1891, %f1883; sub.ftz.f32 %f1896, %f1892, %f1884; fma.rn.ftz.f32 %f693, %f1848, %f1893, %f1881; fma.rn.ftz.f32 %f694, %f1848, %f1894, %f1882; fma.rn.ftz.f32 %f695, %f1848, %f1895, %f1883; fma.rn.ftz.f32 %f696, %f1848, %f1896, %f1884; fma.rn.ftz.f32 %f1897, %f1849, %f1877, %f1853; fma.rn.ftz.f32 %f1898, %f1849, %f1878, %f1854; fma.rn.ftz.f32 %f1899, %f1849, %f1879, %f1855; fma.rn.ftz.f32 %f1900, %f1849, %f1880, %f1856; fma.rn.ftz.f32 %f1901, %f1849, %f1885, %f2110; fma.rn.ftz.f32 %f1902, %f1849, %f1886, %f2111; fma.rn.ftz.f32 %f1903, %f1849, %f1887, %f2112; fma.rn.ftz.f32 %f1904, %f1849, %f1888, %f2113; sub.ftz.f32 %f1905, %f1901, %f1897; sub.ftz.f32 %f1906, %f1902, %f1898; sub.ftz.f32 %f1907, %f1903, %f1899; sub.ftz.f32 %f1908, %f1904, %f1900; fma.rn.ftz.f32 %f697, %f1850, %f1905, %f1897; fma.rn.ftz.f32 %f698, %f1850, %f1906, %f1898; fma.rn.ftz.f32 %f699, %f1850, %f1907, %f1899; fma.rn.ftz.f32 %f700, %f1850, %f1908, %f1900; fma.rn.ftz.f32 %f1909, %f1851, %f1877, %f1853; fma.rn.ftz.f32 %f1910, %f1851, %f1878, %f1854; fma.rn.ftz.f32 %f1911, %f1851, %f1879, %f1855; fma.rn.ftz.f32 %f1912, %f1851, %f1880, %f1856; fma.rn.ftz.f32 %f1913, %f1851, %f1885, %f2110; fma.rn.ftz.f32 %f1914, %f1851, %f1886, %f2111; fma.rn.ftz.f32 %f1915, %f1851, %f1887, %f2112; fma.rn.ftz.f32 %f1916, %f1851, %f1888, %f2113; sub.ftz.f32 %f1917, %f1913, %f1909; sub.ftz.f32 %f1918, %f1914, %f1910; sub.ftz.f32 %f1919, %f1915, %f1911; sub.ftz.f32 %f1920, %f1916, %f1912; fma.rn.ftz.f32 %f701, %f1852, %f1917, %f1909; fma.rn.ftz.f32 %f702, %f1852, %f1918, %f1910; fma.rn.ftz.f32 %f703, %f1852, %f1919, %f1911; fma.rn.ftz.f32 %f704, %f1852, %f1920, %f1912; setp.eq.s64 %p133, %rd16, 0; @%p133 bra $L__BB16_271; ld.local.f32 %f1921, [%rd1+40]; ld.local.f32 %f1922, [%rd1+48]; mul.ftz.f32 %f1923, %f701, %f1922; mul.ftz.f32 %f1924, %f702, %f1922; mul.ftz.f32 %f1925, %f703, %f1922; mul.ftz.f32 %f1926, %f704, %f1922; fma.rn.ftz.f32 %f1927, %f697, %f1921, %f1923; fma.rn.ftz.f32 %f1928, %f698, %f1921, %f1924; fma.rn.ftz.f32 %f1929, %f699, %f1921, %f1925; fma.rn.ftz.f32 %f1930, %f700, %f1921, %f1926; add.ftz.f32 %f1931, %f1921, %f1922; mul.ftz.f32 %f1932, %f693, %f1931; mul.ftz.f32 %f1933, %f694, %f1931; mul.ftz.f32 %f1934, %f695, %f1931; mul.ftz.f32 %f1935, %f696, %f1931; sub.ftz.f32 %f1936, %f1930, %f1935; sub.ftz.f32 %f1937, %f1929, %f1934; sub.ftz.f32 %f1938, %f1928, %f1933; sub.ftz.f32 %f1939, %f1927, %f1932; st.v4.f32 [%rd16], {%f1939, %f1938, %f1937, %f1936}; $L__BB16_271: setp.eq.s64 %p134, %rd17, 0; @%p134 bra $L__BB16_273; ld.local.f32 %f1940, [%rd1+44]; ld.local.f32 %f1941, [%rd1+52]; mul.ftz.f32 %f1942, %f701, %f1941; mul.ftz.f32 %f1943, %f702, %f1941; mul.ftz.f32 %f1944, %f703, %f1941; mul.ftz.f32 %f1945, %f704, %f1941; fma.rn.ftz.f32 %f1946, %f697, %f1940, %f1942; fma.rn.ftz.f32 %f1947, %f698, %f1940, %f1943; fma.rn.ftz.f32 %f1948, %f699, %f1940, %f1944; fma.rn.ftz.f32 %f1949, %f700, %f1940, %f1945; add.ftz.f32 %f1950, %f1940, %f1941; mul.ftz.f32 %f1951, %f693, %f1950; mul.ftz.f32 %f1952, %f694, %f1950; mul.ftz.f32 %f1953, %f695, %f1950; mul.ftz.f32 %f1954, %f696, %f1950; sub.ftz.f32 %f1955, %f1949, %f1954; sub.ftz.f32 %f1956, %f1948, %f1953; sub.ftz.f32 %f1957, %f1947, %f1952; sub.ftz.f32 %f1958, %f1946, %f1951; st.v4.f32 [%rd17], {%f1958, %f1957, %f1956, %f1955}; $L__BB16_273: ld.local.f32 %f1959, [%rd1+8]; ld.local.f32 %f1960, [%rd1+12]; mul.ftz.f32 %f1961, %f701, %f1960; mul.ftz.f32 %f1962, %f702, %f1960; mul.ftz.f32 %f1963, %f703, %f1960; mul.ftz.f32 %f1964, %f704, %f1960; fma.rn.ftz.f32 %f1965, %f697, %f1959, %f1961; fma.rn.ftz.f32 %f1966, %f698, %f1959, %f1962; fma.rn.ftz.f32 %f1967, %f699, %f1959, %f1963; fma.rn.ftz.f32 %f1968, %f700, %f1959, %f1964; mov.f32 %f1969, 0f3F800000; sub.ftz.f32 %f1970, %f1969, %f1959; sub.ftz.f32 %f1971, %f1970, %f1960; fma.rn.ftz.f32 %f2121, %f696, %f1971, %f1968; fma.rn.ftz.f32 %f2120, %f695, %f1971, %f1967; fma.rn.ftz.f32 %f2119, %f694, %f1971, %f1966; fma.rn.ftz.f32 %f2118, %f693, %f1971, %f1965; bra.uni $L__BB16_289; $L__BB16_274: add.s32 %r328, %r93, -1; setp.lt.u32 %p135, %r328, 2; @%p135 bra $L__BB16_279; bra.uni $L__BB16_275; $L__BB16_279: setp.eq.s64 %p138, %rd16, 0; @%p138 bra $L__BB16_281; mov.f32 %f1981, 0f00000000; st.v4.f32 [%rd16], {%f1981, %f1981, %f1981, %f1981}; $L__BB16_281: setp.eq.s64 %p139, %rd17, 0; @%p139 bra $L__BB16_283; mov.f32 %f1982, 0f00000000; st.v4.f32 [%rd17], {%f1982, %f1982, %f1982, %f1982}; $L__BB16_283: ld.const.u64 %rd207, [kernel_params+2344]; cvta.to.global.u64 %rd208, %rd207; mul.wide.s32 %rd209, %r96, 16; add.s64 %rd210, %rd208, %rd209; ld.global.v4.f32 {%f2118, %f2119, %f2120, %f2121}, [%rd210]; bra.uni $L__BB16_289; $L__BB16_220: ld.global.u32 %r357, [%rd9+8]; shr.u32 %r299, %r80, 16; add.s32 %r355, %r79, %r299; add.s32 %r300, %r299, 1; rem.u32 %r301, %r300, %r81; add.s32 %r302, %r301, %r81; rem.u32 %r303, %r302, %r81; add.s32 %r356, %r303, %r79; add.s32 %r304, %r299, -1; rem.s32 %r305, %r304, %r81; add.s32 %r306, %r305, %r81; rem.s32 %r307, %r306, %r81; add.s32 %r358, %r307, %r79; $L__BB16_222: add.s32 %r308, %r355, %r96; cvt.s64.s32 %rd10, %r308; add.s32 %r309, %r356, %r96; cvt.s64.s32 %rd11, %r309; add.s32 %r310, %r357, %r96; cvt.s64.s32 %rd12, %r310; add.s32 %r311, %r358, %r96; cvt.s64.s32 %rd13, %r311; setp.eq.s32 %p116, %r93, 64; @%p116 bra $L__BB16_224; bra.uni $L__BB16_223; $L__BB16_224: ld.const.u64 %rd172, [kernel_params+2352]; cvta.to.global.u64 %rd14, %rd172; shl.b64 %rd173, %rd10, 2; add.s64 %rd174, %rd14, %rd173; ld.global.v4.u8 {%rs161, %rs162, %rs163, %rs164}, [%rd174]; cvt.rn.f32.u16 %f1680, %rs161; mul.ftz.f32 %f556, %f1680, 0f3B808081; cvt.rn.f32.u16 %f1681, %rs162; mul.ftz.f32 %f557, %f1681, 0f3B808081; cvt.rn.f32.u16 %f1682, %rs163; mul.ftz.f32 %f558, %f1682, 0f3B808081; setp.lt.ftz.f32 %p117, %f556, 0f3D25AEE6; @%p117 bra $L__BB16_226; bra.uni $L__BB16_225; $L__BB16_226: mul.ftz.f32 %f2098, %f556, 0f3D9E8391; bra.uni $L__BB16_227; $L__BB16_223: ld.const.u64 %rd162, [kernel_params+2344]; cvta.to.global.u64 %rd163, %rd162; shl.b64 %rd164, %rd10, 4; add.s64 %rd165, %rd163, %rd164; ld.global.v4.f32 {%f2098, %f2099, %f2100, %f2101}, [%rd165]; shl.b64 %rd166, %rd11, 4; add.s64 %rd167, %rd163, %rd166; ld.global.v4.f32 {%f2094, %f2095, %f2096, %f2097}, [%rd167]; shl.b64 %rd168, %rd12, 4; add.s64 %rd169, %rd163, %rd168; ld.global.v4.f32 {%f2090, %f2091, %f2092, %f2093}, [%rd169]; shl.b64 %rd170, %rd13, 4; add.s64 %rd171, %rd163, %rd170; ld.global.v4.f32 {%f2087, %f2086, %f2088, %f2089}, [%rd171]; bra.uni $L__BB16_260; $L__BB16_275: setp.eq.s64 %p136, %rd16, 0; @%p136 bra $L__BB16_277; mov.f32 %f1972, 0f00000000; st.v4.f32 [%rd16], {%f1972, %f1972, %f1972, %f1972}; $L__BB16_277: setp.eq.s64 %p137, %rd17, 0; mov.f32 %f2118, 0f00000000; mov.f32 %f2119, %f2118; mov.f32 %f2120, %f2118; mov.f32 %f2121, %f2118; @%p137 bra $L__BB16_289; mov.f32 %f2118, 0f00000000; st.v4.f32 [%rd17], {%f2118, %f2118, %f2118, %f2118}; mov.f32 %f2119, %f2118; mov.f32 %f2120, %f2118; mov.f32 %f2121, %f2118; bra.uni $L__BB16_289; $L__BB16_225: add.ftz.f32 %f1683, %f556, 0f3D6147AE; mul.ftz.f32 %f1684, %f1683, 0f3F72A76F; lg2.approx.ftz.f32 %f1685, %f1684; mul.ftz.f32 %f1686, %f1685, 0f4019999A; ex2.approx.ftz.f32 %f2098, %f1686; $L__BB16_227: setp.lt.ftz.f32 %p118, %f557, 0f3D25AEE6; @%p118 bra $L__BB16_229; bra.uni $L__BB16_228; $L__BB16_229: mul.ftz.f32 %f2099, %f557, 0f3D9E8391; bra.uni $L__BB16_230; $L__BB16_228: add.ftz.f32 %f1687, %f557, 0f3D6147AE; mul.ftz.f32 %f1688, %f1687, 0f3F72A76F; lg2.approx.ftz.f32 %f1689, %f1688; mul.ftz.f32 %f1690, %f1689, 0f4019999A; ex2.approx.ftz.f32 %f2099, %f1690; $L__BB16_230: setp.lt.ftz.f32 %p119, %f558, 0f3D25AEE6; @%p119 bra $L__BB16_232; bra.uni $L__BB16_231; $L__BB16_232: mul.ftz.f32 %f2100, %f558, 0f3D9E8391; bra.uni $L__BB16_233; $L__BB16_231: add.ftz.f32 %f1691, %f558, 0f3D6147AE; mul.ftz.f32 %f1692, %f1691, 0f3F72A76F; lg2.approx.ftz.f32 %f1693, %f1692; mul.ftz.f32 %f1694, %f1693, 0f4019999A; ex2.approx.ftz.f32 %f2100, %f1694; $L__BB16_233: shl.b64 %rd175, %rd11, 2; add.s64 %rd176, %rd14, %rd175; ld.global.v4.u8 {%rs165, %rs166, %rs167, %rs168}, [%rd176]; cvt.rn.f32.u16 %f1695, %rs165; mul.ftz.f32 %f568, %f1695, 0f3B808081; cvt.rn.f32.u16 %f1696, %rs166; mul.ftz.f32 %f569, %f1696, 0f3B808081; cvt.rn.f32.u16 %f1697, %rs167; mul.ftz.f32 %f570, %f1697, 0f3B808081; setp.lt.ftz.f32 %p120, %f568, 0f3D25AEE6; @%p120 bra $L__BB16_235; bra.uni $L__BB16_234; $L__BB16_235: mul.ftz.f32 %f2094, %f568, 0f3D9E8391; bra.uni $L__BB16_236; $L__BB16_234: add.ftz.f32 %f1698, %f568, 0f3D6147AE; mul.ftz.f32 %f1699, %f1698, 0f3F72A76F; lg2.approx.ftz.f32 %f1700, %f1699; mul.ftz.f32 %f1701, %f1700, 0f4019999A; ex2.approx.ftz.f32 %f2094, %f1701; $L__BB16_236: setp.lt.ftz.f32 %p121, %f569, 0f3D25AEE6; @%p121 bra $L__BB16_238; bra.uni $L__BB16_237; $L__BB16_238: mul.ftz.f32 %f2095, %f569, 0f3D9E8391; bra.uni $L__BB16_239; $L__BB16_237: add.ftz.f32 %f1702, %f569, 0f3D6147AE; mul.ftz.f32 %f1703, %f1702, 0f3F72A76F; lg2.approx.ftz.f32 %f1704, %f1703; mul.ftz.f32 %f1705, %f1704, 0f4019999A; ex2.approx.ftz.f32 %f2095, %f1705; $L__BB16_239: setp.lt.ftz.f32 %p122, %f570, 0f3D25AEE6; @%p122 bra $L__BB16_241; bra.uni $L__BB16_240; $L__BB16_241: mul.ftz.f32 %f2096, %f570, 0f3D9E8391; bra.uni $L__BB16_242; $L__BB16_240: add.ftz.f32 %f1706, %f570, 0f3D6147AE; mul.ftz.f32 %f1707, %f1706, 0f3F72A76F; lg2.approx.ftz.f32 %f1708, %f1707; mul.ftz.f32 %f1709, %f1708, 0f4019999A; ex2.approx.ftz.f32 %f2096, %f1709; $L__BB16_242: shl.b64 %rd177, %rd12, 2; add.s64 %rd178, %rd14, %rd177; ld.global.v4.u8 {%rs169, %rs170, %rs171, %rs172}, [%rd178]; cvt.rn.f32.u16 %f1710, %rs169; mul.ftz.f32 %f580, %f1710, 0f3B808081; cvt.rn.f32.u16 %f1711, %rs170; mul.ftz.f32 %f581, %f1711, 0f3B808081; cvt.rn.f32.u16 %f1712, %rs171; mul.ftz.f32 %f582, %f1712, 0f3B808081; setp.lt.ftz.f32 %p123, %f580, 0f3D25AEE6; @%p123 bra $L__BB16_244; bra.uni $L__BB16_243; $L__BB16_244: mul.ftz.f32 %f2090, %f580, 0f3D9E8391; bra.uni $L__BB16_245; $L__BB16_243: add.ftz.f32 %f1713, %f580, 0f3D6147AE; mul.ftz.f32 %f1714, %f1713, 0f3F72A76F; lg2.approx.ftz.f32 %f1715, %f1714; mul.ftz.f32 %f1716, %f1715, 0f4019999A; ex2.approx.ftz.f32 %f2090, %f1716; $L__BB16_245: setp.lt.ftz.f32 %p124, %f581, 0f3D25AEE6; @%p124 bra $L__BB16_247; bra.uni $L__BB16_246; $L__BB16_247: mul.ftz.f32 %f2091, %f581, 0f3D9E8391; bra.uni $L__BB16_248; $L__BB16_246: add.ftz.f32 %f1717, %f581, 0f3D6147AE; mul.ftz.f32 %f1718, %f1717, 0f3F72A76F; lg2.approx.ftz.f32 %f1719, %f1718; mul.ftz.f32 %f1720, %f1719, 0f4019999A; ex2.approx.ftz.f32 %f2091, %f1720; $L__BB16_248: setp.lt.ftz.f32 %p125, %f582, 0f3D25AEE6; @%p125 bra $L__BB16_250; bra.uni $L__BB16_249; $L__BB16_250: mul.ftz.f32 %f2092, %f582, 0f3D9E8391; bra.uni $L__BB16_251; $L__BB16_249: add.ftz.f32 %f1721, %f582, 0f3D6147AE; mul.ftz.f32 %f1722, %f1721, 0f3F72A76F; lg2.approx.ftz.f32 %f1723, %f1722; mul.ftz.f32 %f1724, %f1723, 0f4019999A; ex2.approx.ftz.f32 %f2092, %f1724; $L__BB16_251: shl.b64 %rd179, %rd13, 2; add.s64 %rd180, %rd14, %rd179; ld.global.v4.u8 {%rs173, %rs174, %rs175, %rs176}, [%rd180]; cvt.rn.f32.u16 %f1725, %rs173; mul.ftz.f32 %f592, %f1725, 0f3B808081; cvt.rn.f32.u16 %f1726, %rs174; mul.ftz.f32 %f593, %f1726, 0f3B808081; cvt.rn.f32.u16 %f1727, %rs175; mul.ftz.f32 %f594, %f1727, 0f3B808081; setp.lt.ftz.f32 %p126, %f592, 0f3D25AEE6; @%p126 bra $L__BB16_253; bra.uni $L__BB16_252; $L__BB16_253: mul.ftz.f32 %f2087, %f592, 0f3D9E8391; bra.uni $L__BB16_254; $L__BB16_252: add.ftz.f32 %f1728, %f592, 0f3D6147AE; mul.ftz.f32 %f1729, %f1728, 0f3F72A76F; lg2.approx.ftz.f32 %f1730, %f1729; mul.ftz.f32 %f1731, %f1730, 0f4019999A; ex2.approx.ftz.f32 %f2087, %f1731; $L__BB16_254: setp.lt.ftz.f32 %p127, %f593, 0f3D25AEE6; @%p127 bra $L__BB16_256; bra.uni $L__BB16_255; $L__BB16_256: mul.ftz.f32 %f2086, %f593, 0f3D9E8391; bra.uni $L__BB16_257; $L__BB16_255: add.ftz.f32 %f1732, %f593, 0f3D6147AE; mul.ftz.f32 %f1733, %f1732, 0f3F72A76F; lg2.approx.ftz.f32 %f1734, %f1733; mul.ftz.f32 %f1735, %f1734, 0f4019999A; ex2.approx.ftz.f32 %f2086, %f1735; $L__BB16_257: cvt.rn.f32.u16 %f1736, %rs164; mul.ftz.f32 %f2101, %f1736, 0f3B808081; cvt.rn.f32.u16 %f1737, %rs168; mul.ftz.f32 %f2097, %f1737, 0f3B808081; cvt.rn.f32.u16 %f1738, %rs172; mul.ftz.f32 %f2093, %f1738, 0f3B808081; cvt.rn.f32.u16 %f1739, %rs176; mul.ftz.f32 %f2089, %f1739, 0f3B808081; setp.lt.ftz.f32 %p128, %f594, 0f3D25AEE6; @%p128 bra $L__BB16_259; bra.uni $L__BB16_258; $L__BB16_259: mul.ftz.f32 %f2088, %f594, 0f3D9E8391; bra.uni $L__BB16_260; $L__BB16_258: add.ftz.f32 %f1740, %f594, 0f3D6147AE; mul.ftz.f32 %f1741, %f1740, 0f3F72A76F; lg2.approx.ftz.f32 %f1742, %f1741; mul.ftz.f32 %f1743, %f1742, 0f4019999A; ex2.approx.ftz.f32 %f2088, %f1743; $L__BB16_260: @%p115 bra $L__BB16_262; add.ftz.f32 %f1744, %f2094, %f2098; mul.ftz.f32 %f2094, %f1744, 0f3F000000; add.ftz.f32 %f1745, %f2095, %f2099; mul.ftz.f32 %f2095, %f1745, 0f3F000000; add.ftz.f32 %f1746, %f2096, %f2100; mul.ftz.f32 %f2096, %f1746, 0f3F000000; add.ftz.f32 %f1747, %f2097, %f2101; mul.ftz.f32 %f2097, %f1747, 0f3F000000; add.ftz.f32 %f1748, %f2087, %f2098; mul.ftz.f32 %f2087, %f1748, 0f3F000000; add.ftz.f32 %f1749, %f2086, %f2099; mul.ftz.f32 %f2086, %f1749, 0f3F000000; add.ftz.f32 %f1750, %f2088, %f2100; mul.ftz.f32 %f2088, %f1750, 0f3F000000; add.ftz.f32 %f1751, %f2089, %f2101; mul.ftz.f32 %f2089, %f1751, 0f3F000000; $L__BB16_262: sub.ftz.f32 %f1752, %f2094, %f2098; sub.ftz.f32 %f1753, %f2095, %f2099; sub.ftz.f32 %f1754, %f2096, %f2100; sub.ftz.f32 %f1755, %f2097, %f2101; fma.rn.ftz.f32 %f1756, %f1658, %f1752, %f2098; fma.rn.ftz.f32 %f1757, %f1658, %f1753, %f2099; fma.rn.ftz.f32 %f1758, %f1658, %f1754, %f2100; fma.rn.ftz.f32 %f1759, %f1658, %f1755, %f2101; sub.ftz.f32 %f1760, %f2090, %f2087; sub.ftz.f32 %f1761, %f2091, %f2086; sub.ftz.f32 %f1762, %f2092, %f2088; sub.ftz.f32 %f1763, %f2093, %f2089; fma.rn.ftz.f32 %f1764, %f1658, %f1760, %f2087; fma.rn.ftz.f32 %f1765, %f1658, %f1761, %f2086; fma.rn.ftz.f32 %f1766, %f1658, %f1762, %f2088; fma.rn.ftz.f32 %f1767, %f1658, %f1763, %f2089; sub.ftz.f32 %f1768, %f1764, %f1756; sub.ftz.f32 %f1769, %f1765, %f1757; sub.ftz.f32 %f1770, %f1766, %f1758; sub.ftz.f32 %f1771, %f1767, %f1759; fma.rn.ftz.f32 %f639, %f1659, %f1768, %f1756; fma.rn.ftz.f32 %f640, %f1659, %f1769, %f1757; fma.rn.ftz.f32 %f641, %f1659, %f1770, %f1758; fma.rn.ftz.f32 %f642, %f1659, %f1771, %f1759; fma.rn.ftz.f32 %f1772, %f1660, %f1752, %f2098; fma.rn.ftz.f32 %f1773, %f1660, %f1753, %f2099; fma.rn.ftz.f32 %f1774, %f1660, %f1754, %f2100; fma.rn.ftz.f32 %f1775, %f1660, %f1755, %f2101; fma.rn.ftz.f32 %f1776, %f1660, %f1760, %f2087; fma.rn.ftz.f32 %f1777, %f1660, %f1761, %f2086; fma.rn.ftz.f32 %f1778, %f1660, %f1762, %f2088; fma.rn.ftz.f32 %f1779, %f1660, %f1763, %f2089; sub.ftz.f32 %f1780, %f1776, %f1772; sub.ftz.f32 %f1781, %f1777, %f1773; sub.ftz.f32 %f1782, %f1778, %f1774; sub.ftz.f32 %f1783, %f1779, %f1775; fma.rn.ftz.f32 %f643, %f1661, %f1780, %f1772; fma.rn.ftz.f32 %f644, %f1661, %f1781, %f1773; fma.rn.ftz.f32 %f645, %f1661, %f1782, %f1774; fma.rn.ftz.f32 %f646, %f1661, %f1783, %f1775; fma.rn.ftz.f32 %f1784, %f1662, %f1752, %f2098; fma.rn.ftz.f32 %f1785, %f1662, %f1753, %f2099; fma.rn.ftz.f32 %f1786, %f1662, %f1754, %f2100; fma.rn.ftz.f32 %f1787, %f1662, %f1755, %f2101; fma.rn.ftz.f32 %f1788, %f1662, %f1760, %f2087; fma.rn.ftz.f32 %f1789, %f1662, %f1761, %f2086; fma.rn.ftz.f32 %f1790, %f1662, %f1762, %f2088; fma.rn.ftz.f32 %f1791, %f1662, %f1763, %f2089; sub.ftz.f32 %f1792, %f1788, %f1784; sub.ftz.f32 %f1793, %f1789, %f1785; sub.ftz.f32 %f1794, %f1790, %f1786; sub.ftz.f32 %f1795, %f1791, %f1787; fma.rn.ftz.f32 %f647, %f1663, %f1792, %f1784; fma.rn.ftz.f32 %f648, %f1663, %f1793, %f1785; fma.rn.ftz.f32 %f649, %f1663, %f1794, %f1786; fma.rn.ftz.f32 %f650, %f1663, %f1795, %f1787; setp.eq.s64 %p130, %rd16, 0; @%p130 bra $L__BB16_264; ld.local.f32 %f1796, [%rd1+40]; ld.local.f32 %f1797, [%rd1+48]; mul.ftz.f32 %f1798, %f647, %f1797; mul.ftz.f32 %f1799, %f648, %f1797; mul.ftz.f32 %f1800, %f649, %f1797; mul.ftz.f32 %f1801, %f650, %f1797; fma.rn.ftz.f32 %f1802, %f643, %f1796, %f1798; fma.rn.ftz.f32 %f1803, %f644, %f1796, %f1799; fma.rn.ftz.f32 %f1804, %f645, %f1796, %f1800; fma.rn.ftz.f32 %f1805, %f646, %f1796, %f1801; add.ftz.f32 %f1806, %f1796, %f1797; mul.ftz.f32 %f1807, %f639, %f1806; mul.ftz.f32 %f1808, %f640, %f1806; mul.ftz.f32 %f1809, %f641, %f1806; mul.ftz.f32 %f1810, %f642, %f1806; sub.ftz.f32 %f1811, %f1805, %f1810; sub.ftz.f32 %f1812, %f1804, %f1809; sub.ftz.f32 %f1813, %f1803, %f1808; sub.ftz.f32 %f1814, %f1802, %f1807; st.v4.f32 [%rd16], {%f1814, %f1813, %f1812, %f1811}; $L__BB16_264: setp.eq.s64 %p131, %rd17, 0; @%p131 bra $L__BB16_266; ld.local.f32 %f1815, [%rd1+44]; ld.local.f32 %f1816, [%rd1+52]; mul.ftz.f32 %f1817, %f647, %f1816; mul.ftz.f32 %f1818, %f648, %f1816; mul.ftz.f32 %f1819, %f649, %f1816; mul.ftz.f32 %f1820, %f650, %f1816; fma.rn.ftz.f32 %f1821, %f643, %f1815, %f1817; fma.rn.ftz.f32 %f1822, %f644, %f1815, %f1818; fma.rn.ftz.f32 %f1823, %f645, %f1815, %f1819; fma.rn.ftz.f32 %f1824, %f646, %f1815, %f1820; add.ftz.f32 %f1825, %f1815, %f1816; mul.ftz.f32 %f1826, %f639, %f1825; mul.ftz.f32 %f1827, %f640, %f1825; mul.ftz.f32 %f1828, %f641, %f1825; mul.ftz.f32 %f1829, %f642, %f1825; sub.ftz.f32 %f1830, %f1824, %f1829; sub.ftz.f32 %f1831, %f1823, %f1828; sub.ftz.f32 %f1832, %f1822, %f1827; sub.ftz.f32 %f1833, %f1821, %f1826; st.v4.f32 [%rd17], {%f1833, %f1832, %f1831, %f1830}; $L__BB16_266: ld.local.f32 %f1834, [%rd1+8]; ld.local.f32 %f1835, [%rd1+12]; mul.ftz.f32 %f1836, %f647, %f1835; mul.ftz.f32 %f1837, %f648, %f1835; mul.ftz.f32 %f1838, %f649, %f1835; mul.ftz.f32 %f1839, %f650, %f1835; fma.rn.ftz.f32 %f1840, %f643, %f1834, %f1836; fma.rn.ftz.f32 %f1841, %f644, %f1834, %f1837; fma.rn.ftz.f32 %f1842, %f645, %f1834, %f1838; fma.rn.ftz.f32 %f1843, %f646, %f1834, %f1839; mov.f32 %f1844, 0f3F800000; sub.ftz.f32 %f1845, %f1844, %f1834; sub.ftz.f32 %f1846, %f1845, %f1835; fma.rn.ftz.f32 %f2121, %f642, %f1846, %f1843; fma.rn.ftz.f32 %f2120, %f641, %f1846, %f1842; fma.rn.ftz.f32 %f2119, %f640, %f1846, %f1841; fma.rn.ftz.f32 %f2118, %f639, %f1846, %f1840; $L__BB16_289: st.param.f32 [func_retval0+0], %f2118; st.param.f32 [func_retval0+4], %f2119; st.param.f32 [func_retval0+8], %f2120; st.param.f32 [func_retval0+12], %f2121; ret; } .func _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib( .param .b64 _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib_param_0 ) { .reg .pred %p<23>; .reg .b16 %rs<12>; .reg .f32 %f<231>; .reg .b32 %r<81>; .reg .b64 %rd<218>; ld.param.u64 %rd118, [_Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib_param_0]; cvta.to.local.u64 %rd119, %rd118; ld.const.u64 %rd120, [kernel_params+2216]; cvta.to.global.u64 %rd121, %rd120; add.s64 %rd1, %rd119, 60; ld.local.u32 %r29, [%rd119+60]; cvt.s64.s32 %rd2, %r29; mul.wide.s32 %rd122, %r29, 4; add.s64 %rd123, %rd121, %rd122; ld.global.u32 %r1, [%rd123]; st.local.u32 [%rd119+48], %r1; ld.const.u64 %rd124, [kernel_params+2160]; cvta.to.global.u64 %rd125, %rd124; ld.local.u32 %r30, [%rd119+76]; mul.wide.s32 %rd126, %r30, 208; add.s64 %rd127, %rd125, %rd126; add.s64 %rd3, %rd127, 152; ld.global.u32 %r2, [%rd127+156]; ld.global.u32 %r3, [%rd127+152]; shl.b32 %r31, %r3, 1; cvt.rn.f32.s32 %f92, %r31; ld.local.f32 %f93, [%rd119+84]; mul.ftz.f32 %f94, %f93, %f92; cvt.rzi.ftz.s32.f32 %r32, %f94; add.s32 %r33, %r31, -1; min.s32 %r4, %r32, %r33; cvt.rn.f32.s32 %f95, %r4; sub.ftz.f32 %f1, %f94, %f95; ld.const.u64 %rd128, [kernel_params+2312]; cvta.to.global.u64 %rd4, %rd128; ld.global.u32 %r77, [%rd127+164]; mul.wide.u32 %rd129, %r77, 16; add.s64 %rd130, %rd4, %rd129; ld.global.u16 %rs10, [%rd130+12]; ld.global.u32 %r76, [%rd130+8]; ld.global.u64 %rd199, [%rd130]; setp.eq.s64 %p1, %rd199, 11; mov.u16 %rs8, %rs10; mov.u32 %r71, %r76; @%p1 bra $L__BB17_6; mov.u64 %rd180, %rd199; mov.u32 %r71, %r76; mov.u16 %rs8, %rs10; mov.u32 %r72, %r77; $L__BB17_2: setp.eq.s64 %p2, %rd180, 0; @%p2 bra $L__BB17_4; add.s32 %r72, %r72, 2; bra.uni $L__BB17_5; $L__BB17_4: setp.eq.s16 %p3, %rs8, 0; mov.u32 %r75, -1; mov.u32 %r72, %r71; @%p3 bra $L__BB17_7; $L__BB17_5: mul.wide.u32 %rd131, %r72, 16; add.s64 %rd132, %rd4, %rd131; ld.global.u16 %rs8, [%rd132+12]; ld.global.u32 %r71, [%rd132+8]; ld.global.u64 %rd180, [%rd132]; setp.ne.s64 %p4, %rd180, 11; @%p4 bra $L__BB17_2; $L__BB17_6: setp.eq.s16 %p5, %rs8, 0; selp.b32 %r75, -1, %r71, %p5; $L__BB17_7: ld.const.u64 %rd133, [kernel_params+2232]; cvta.to.global.u64 %rd134, %rd133; shl.b64 %rd135, %rd2, 4; add.s64 %rd136, %rd134, %rd135; ld.global.v4.u32 {%r35, %r36, %r37, %r38}, [%rd136]; setp.eq.s32 %p6, %r4, %r3; @%p6 bra $L__BB17_9; bra.uni $L__BB17_8; $L__BB17_9: ld.const.u64 %rd142, [kernel_params+2256]; cvta.to.global.u64 %rd143, %rd142; mul.wide.u32 %rd144, %r38, 12; add.s64 %rd181, %rd143, %rd144; add.s64 %rd182, %rd181, 4; add.s64 %rd183, %rd181, 8; add.s32 %r45, %r38, 1; mul.wide.u32 %rd145, %r45, 12; add.s64 %rd184, %rd143, %rd145; add.s64 %rd185, %rd184, 4; add.s64 %rd186, %rd184, 8; add.s32 %r46, %r38, 2; mul.wide.u32 %rd146, %r46, 12; add.s64 %rd187, %rd143, %rd146; add.s64 %rd188, %rd187, 4; add.s64 %rd189, %rd187, 8; bra.uni $L__BB17_10; $L__BB17_8: setp.gt.s32 %p7, %r4, %r3; selp.b32 %r39, -1, 0, %p7; add.s32 %r40, %r4, %r39; mad.lo.s32 %r41, %r40, %r2, %r75; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r42, %r35, %r41; mul.wide.u32 %rd139, %r42, 12; add.s64 %rd181, %rd138, %rd139; add.s64 %rd182, %rd181, 4; add.s64 %rd183, %rd181, 8; add.s32 %r43, %r36, %r41; mul.wide.u32 %rd140, %r43, 12; add.s64 %rd184, %rd138, %rd140; add.s64 %rd185, %rd184, 4; add.s64 %rd186, %rd184, 8; add.s32 %r44, %r37, %r41; mul.wide.u32 %rd141, %r44, 12; add.s64 %rd187, %rd138, %rd141; add.s64 %rd188, %rd187, 4; add.s64 %rd189, %rd187, 8; $L__BB17_10: ld.global.f32 %f2, [%rd189]; ld.global.f32 %f3, [%rd188]; ld.global.f32 %f4, [%rd187]; ld.global.f32 %f5, [%rd186]; ld.global.f32 %f6, [%rd185]; ld.global.f32 %f7, [%rd184]; ld.global.f32 %f8, [%rd183]; ld.global.f32 %f9, [%rd182]; ld.global.f32 %f10, [%rd181]; add.s32 %r19, %r4, 1; setp.eq.s32 %p8, %r19, %r3; @%p8 bra $L__BB17_12; bra.uni $L__BB17_11; $L__BB17_12: ld.const.u64 %rd152, [kernel_params+2256]; cvta.to.global.u64 %rd153, %rd152; mul.wide.u32 %rd154, %r38, 12; add.s64 %rd190, %rd153, %rd154; add.s32 %r53, %r38, 1; mul.wide.u32 %rd155, %r53, 12; add.s64 %rd193, %rd153, %rd155; add.s32 %r54, %r38, 2; mul.wide.u32 %rd156, %r54, 12; add.s64 %rd196, %rd153, %rd156; add.s64 %rd191, %rd190, 4; add.s64 %rd192, %rd190, 8; add.s64 %rd194, %rd193, 4; add.s64 %rd195, %rd193, 8; add.s64 %rd197, %rd196, 4; add.s64 %rd198, %rd196, 8; bra.uni $L__BB17_13; $L__BB17_11: setp.ge.s32 %p9, %r4, %r3; selp.b32 %r47, -1, 0, %p9; add.s32 %r48, %r19, %r47; mad.lo.s32 %r49, %r48, %r2, %r75; ld.const.u64 %rd147, [kernel_params+2336]; cvta.to.global.u64 %rd148, %rd147; add.s32 %r50, %r35, %r49; mul.wide.u32 %rd149, %r50, 12; add.s64 %rd190, %rd148, %rd149; add.s32 %r51, %r36, %r49; mul.wide.u32 %rd150, %r51, 12; add.s64 %rd193, %rd148, %rd150; add.s32 %r52, %r37, %r49; mul.wide.u32 %rd151, %r52, 12; add.s64 %rd196, %rd148, %rd151; add.s64 %rd191, %rd190, 4; add.s64 %rd192, %rd190, 8; add.s64 %rd194, %rd193, 4; add.s64 %rd195, %rd193, 8; add.s64 %rd197, %rd196, 4; add.s64 %rd198, %rd196, 8; $L__BB17_13: mov.f32 %f96, 0f3F800000; sub.ftz.f32 %f11, %f96, %f1; ld.global.f32 %f97, [%rd190]; mul.ftz.f32 %f98, %f1, %f97; ld.global.f32 %f99, [%rd191]; mul.ftz.f32 %f100, %f1, %f99; ld.global.f32 %f101, [%rd192]; mul.ftz.f32 %f102, %f1, %f101; fma.rn.ftz.f32 %f103, %f11, %f10, %f98; fma.rn.ftz.f32 %f104, %f11, %f9, %f100; fma.rn.ftz.f32 %f105, %f11, %f8, %f102; ld.global.f32 %f106, [%rd193]; mul.ftz.f32 %f107, %f1, %f106; ld.global.f32 %f108, [%rd194]; mul.ftz.f32 %f109, %f1, %f108; ld.global.f32 %f110, [%rd195]; mul.ftz.f32 %f111, %f1, %f110; fma.rn.ftz.f32 %f112, %f11, %f7, %f107; fma.rn.ftz.f32 %f113, %f11, %f6, %f109; fma.rn.ftz.f32 %f114, %f11, %f5, %f111; ld.global.f32 %f115, [%rd196]; mul.ftz.f32 %f116, %f1, %f115; ld.global.f32 %f117, [%rd197]; mul.ftz.f32 %f118, %f1, %f117; ld.global.f32 %f119, [%rd198]; mul.ftz.f32 %f120, %f1, %f119; fma.rn.ftz.f32 %f121, %f11, %f4, %f116; fma.rn.ftz.f32 %f122, %f11, %f3, %f118; fma.rn.ftz.f32 %f123, %f11, %f2, %f120; sub.ftz.f32 %f12, %f112, %f103; sub.ftz.f32 %f13, %f113, %f104; sub.ftz.f32 %f14, %f114, %f105; ld.local.f32 %f15, [%rd1+8]; fma.rn.ftz.f32 %f124, %f15, %f12, %f103; fma.rn.ftz.f32 %f125, %f15, %f13, %f104; fma.rn.ftz.f32 %f126, %f14, %f15, %f105; sub.ftz.f32 %f16, %f121, %f103; sub.ftz.f32 %f17, %f122, %f104; sub.ftz.f32 %f18, %f123, %f105; ld.local.f32 %f19, [%rd1+12]; fma.rn.ftz.f32 %f225, %f19, %f16, %f124; fma.rn.ftz.f32 %f226, %f17, %f19, %f125; fma.rn.ftz.f32 %f227, %f18, %f19, %f126; ld.local.u32 %r20, [%rd1+-4]; and.b32 %r55, %r20, 4; setp.ne.s32 %p10, %r55, 0; @%p10 bra $L__BB17_18; and.b32 %r56, %r20, 2; setp.eq.s32 %p11, %r56, 0; @%p11 bra $L__BB17_16; bra.uni $L__BB17_15; $L__BB17_16: ld.global.v4.f32 {%f213, %f214, %f215, %f216}, [%rd3+-152]; ld.global.v4.f32 {%f217, %f218, %f219, %f220}, [%rd3+-136]; ld.global.v4.f32 {%f221, %f222, %f223, %f224}, [%rd3+-120]; bra.uni $L__BB17_17; $L__BB17_15: ld.local.v4.f32 {%f213, %f214, %f215, %f216}, [%rd1+84]; ld.local.v4.f32 {%f217, %f218, %f219, %f220}, [%rd1+100]; ld.local.v4.f32 {%f221, %f222, %f223, %f224}, [%rd1+116]; $L__BB17_17: mul.ftz.f32 %f151, %f226, %f214; fma.rn.ftz.f32 %f152, %f225, %f213, %f151; fma.rn.ftz.f32 %f153, %f227, %f215, %f152; mul.ftz.f32 %f154, %f226, %f218; fma.rn.ftz.f32 %f155, %f225, %f217, %f154; fma.rn.ftz.f32 %f156, %f227, %f219, %f155; mul.ftz.f32 %f157, %f226, %f222; fma.rn.ftz.f32 %f158, %f225, %f221, %f157; fma.rn.ftz.f32 %f159, %f227, %f223, %f158; add.ftz.f32 %f225, %f216, %f153; add.ftz.f32 %f226, %f220, %f156; add.ftz.f32 %f227, %f224, %f159; $L__BB17_18: st.local.v2.f32 [%rd119], {%f225, %f226}; st.local.f32 [%rd1+-52], %f227; and.b32 %r57, %r20, 8; setp.eq.s32 %p12, %r57, 0; mul.ftz.f32 %f68, %f14, %f17; mul.ftz.f32 %f69, %f18, %f13; mul.ftz.f32 %f70, %f18, %f12; mul.ftz.f32 %f71, %f14, %f16; mul.ftz.f32 %f72, %f13, %f16; mul.ftz.f32 %f73, %f17, %f12; @%p12 bra $L__BB17_20; bra.uni $L__BB17_19; $L__BB17_20: sub.ftz.f32 %f167, %f69, %f68; sub.ftz.f32 %f168, %f71, %f70; mul.ftz.f32 %f169, %f168, %f168; fma.rn.ftz.f32 %f170, %f167, %f167, %f169; sub.ftz.f32 %f171, %f73, %f72; fma.rn.ftz.f32 %f172, %f171, %f171, %f170; sqrt.approx.ftz.f32 %f173, %f172; div.approx.ftz.f32 %f230, %f171, %f173; div.approx.ftz.f32 %f229, %f168, %f173; div.approx.ftz.f32 %f228, %f167, %f173; bra.uni $L__BB17_21; $L__BB17_19: sub.ftz.f32 %f160, %f68, %f69; sub.ftz.f32 %f161, %f70, %f71; mul.ftz.f32 %f162, %f161, %f161; fma.rn.ftz.f32 %f163, %f160, %f160, %f162; sub.ftz.f32 %f164, %f72, %f73; fma.rn.ftz.f32 %f165, %f164, %f164, %f163; sqrt.approx.ftz.f32 %f166, %f165; div.approx.ftz.f32 %f230, %f164, %f166; div.approx.ftz.f32 %f229, %f161, %f166; div.approx.ftz.f32 %f228, %f160, %f166; $L__BB17_21: st.local.v2.f32 [%rd1+-36], {%f228, %f229}; st.local.f32 [%rd1+-28], %f230; st.local.f32 [%rd1+-48], %f228; st.local.f32 [%rd1+-44], %f229; st.local.f32 [%rd1+-40], %f230; st.local.f32 [%rd1+56], %f12; st.local.f32 [%rd1+60], %f13; st.local.f32 [%rd1+64], %f14; st.local.v2.f32 [%rd1+68], {%f16, %f17}; st.local.f32 [%rd1+76], %f18; setp.gt.s32 %p13, %r1, -1; @%p13 bra $L__BB17_36; setp.eq.s64 %p14, %rd199, 12; @%p14 bra $L__BB17_28; mov.u32 %r80, -1; $L__BB17_24: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB17_26; add.s32 %r77, %r77, 2; bra.uni $L__BB17_27; $L__BB17_26: setp.eq.s16 %p16, %rs10, 0; mov.u32 %r77, %r76; @%p16 bra $L__BB17_29; $L__BB17_27: mul.wide.u32 %rd158, %r77, 16; add.s64 %rd159, %rd4, %rd158; ld.global.u16 %rs10, [%rd159+12]; ld.global.u32 %r76, [%rd159+8]; ld.global.u64 %rd199, [%rd159]; setp.ne.s64 %p17, %rd199, 12; @%p17 bra $L__BB17_24; $L__BB17_28: setp.eq.s16 %p18, %rs10, 0; selp.b32 %r80, -1, %r76, %p18; $L__BB17_29: @%p6 bra $L__BB17_31; bra.uni $L__BB17_30; $L__BB17_31: ld.const.u64 %rd165, [kernel_params+2224]; cvta.to.global.u64 %rd166, %rd165; mul.wide.u32 %rd167, %r35, 12; add.s64 %rd208, %rd166, %rd167; add.s64 %rd207, %rd208, 4; add.s64 %rd206, %rd208, 8; mul.wide.u32 %rd168, %r36, 12; add.s64 %rd205, %rd166, %rd168; add.s64 %rd204, %rd205, 4; add.s64 %rd203, %rd205, 8; mul.wide.u32 %rd169, %r37, 12; add.s64 %rd202, %rd166, %rd169; add.s64 %rd201, %rd202, 4; add.s64 %rd200, %rd202, 8; bra.uni $L__BB17_32; $L__BB17_30: setp.gt.s32 %p20, %r4, %r3; selp.b32 %r59, -1, 0, %p20; add.s32 %r60, %r4, %r59; mad.lo.s32 %r61, %r60, %r2, %r80; ld.const.u64 %rd160, [kernel_params+2336]; cvta.to.global.u64 %rd161, %rd160; add.s32 %r62, %r61, %r35; mul.wide.u32 %rd162, %r62, 12; add.s64 %rd208, %rd161, %rd162; add.s64 %rd207, %rd208, 4; add.s64 %rd206, %rd208, 8; add.s32 %r63, %r61, %r36; mul.wide.u32 %rd163, %r63, 12; add.s64 %rd205, %rd161, %rd163; add.s64 %rd204, %rd205, 4; add.s64 %rd203, %rd205, 8; add.s32 %r64, %r61, %r37; mul.wide.u32 %rd164, %r64, 12; add.s64 %rd202, %rd161, %rd164; add.s64 %rd201, %rd202, 4; add.s64 %rd200, %rd202, 8; $L__BB17_32: ld.global.f32 %f83, [%rd208]; ld.global.f32 %f84, [%rd207]; ld.global.f32 %f85, [%rd206]; ld.global.f32 %f86, [%rd205]; ld.global.f32 %f87, [%rd204]; ld.global.f32 %f88, [%rd203]; ld.global.f32 %f89, [%rd202]; ld.global.f32 %f90, [%rd201]; ld.global.f32 %f91, [%rd200]; @%p8 bra $L__BB17_34; bra.uni $L__BB17_33; $L__BB17_34: ld.const.u64 %rd175, [kernel_params+2224]; cvta.to.global.u64 %rd176, %rd175; mul.wide.u32 %rd177, %r35, 12; add.s64 %rd217, %rd176, %rd177; add.s64 %rd216, %rd217, 4; add.s64 %rd215, %rd217, 8; mul.wide.u32 %rd178, %r36, 12; add.s64 %rd214, %rd176, %rd178; add.s64 %rd213, %rd214, 4; add.s64 %rd212, %rd214, 8; mul.wide.u32 %rd179, %r37, 12; add.s64 %rd211, %rd176, %rd179; add.s64 %rd210, %rd211, 4; add.s64 %rd209, %rd211, 8; bra.uni $L__BB17_35; $L__BB17_33: setp.ge.s32 %p22, %r4, %r3; selp.b32 %r65, -1, 0, %p22; add.s32 %r66, %r19, %r65; mad.lo.s32 %r67, %r66, %r2, %r80; ld.const.u64 %rd170, [kernel_params+2336]; cvta.to.global.u64 %rd171, %rd170; add.s32 %r68, %r67, %r35; mul.wide.u32 %rd172, %r68, 12; add.s64 %rd217, %rd171, %rd172; add.s64 %rd216, %rd217, 4; add.s64 %rd215, %rd217, 8; add.s32 %r69, %r67, %r36; mul.wide.u32 %rd173, %r69, 12; add.s64 %rd214, %rd171, %rd173; add.s64 %rd213, %rd214, 4; add.s64 %rd212, %rd214, 8; add.s32 %r70, %r67, %r37; mul.wide.u32 %rd174, %r70, 12; add.s64 %rd211, %rd171, %rd174; add.s64 %rd210, %rd211, 4; add.s64 %rd209, %rd211, 8; $L__BB17_35: ld.global.f32 %f174, [%rd217]; mul.ftz.f32 %f175, %f1, %f174; ld.global.f32 %f176, [%rd216]; mul.ftz.f32 %f177, %f1, %f176; ld.global.f32 %f178, [%rd215]; mul.ftz.f32 %f179, %f1, %f178; fma.rn.ftz.f32 %f180, %f11, %f83, %f175; fma.rn.ftz.f32 %f181, %f11, %f84, %f177; fma.rn.ftz.f32 %f182, %f11, %f85, %f179; ld.global.f32 %f183, [%rd214]; mul.ftz.f32 %f184, %f1, %f183; ld.global.f32 %f185, [%rd213]; mul.ftz.f32 %f186, %f1, %f185; ld.global.f32 %f187, [%rd212]; mul.ftz.f32 %f188, %f1, %f187; fma.rn.ftz.f32 %f189, %f11, %f86, %f184; fma.rn.ftz.f32 %f190, %f11, %f87, %f186; fma.rn.ftz.f32 %f191, %f11, %f88, %f188; ld.global.f32 %f192, [%rd211]; mul.ftz.f32 %f193, %f1, %f192; ld.global.f32 %f194, [%rd210]; mul.ftz.f32 %f195, %f1, %f194; ld.global.f32 %f196, [%rd209]; mul.ftz.f32 %f197, %f1, %f196; fma.rn.ftz.f32 %f198, %f11, %f89, %f193; fma.rn.ftz.f32 %f199, %f11, %f90, %f195; fma.rn.ftz.f32 %f200, %f11, %f91, %f197; sub.ftz.f32 %f202, %f96, %f15; sub.ftz.f32 %f203, %f202, %f19; mul.ftz.f32 %f204, %f180, %f203; mul.ftz.f32 %f205, %f181, %f203; mul.ftz.f32 %f206, %f182, %f203; fma.rn.ftz.f32 %f207, %f189, %f15, %f204; fma.rn.ftz.f32 %f208, %f190, %f15, %f205; fma.rn.ftz.f32 %f209, %f191, %f15, %f206; fma.rn.ftz.f32 %f210, %f198, %f19, %f207; fma.rn.ftz.f32 %f211, %f199, %f19, %f208; fma.rn.ftz.f32 %f212, %f200, %f19, %f209; st.local.f32 [%rd1+-48], %f210; st.local.f32 [%rd1+-44], %f211; st.local.f32 [%rd1+-40], %f212; $L__BB17_36: ret; } .func (.param .b32 func_retval0) _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample( .param .b32 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_0, .param .b32 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_1, .param .b32 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_2, .param .align 4 .b8 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3[12], .param .b32 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_4, .param .b32 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_5, .param .b64 _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_6 ) { .reg .pred %p<94>; .reg .b16 %rs<24>; .reg .f32 %f<1346>; .reg .b32 %r<141>; .reg .b64 %rd<192>; ld.param.f32 %f341, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_0]; ld.param.f32 %f342, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_1]; ld.param.f32 %f343, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_2]; ld.param.u32 %r44, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_4]; ld.param.u32 %r45, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_5]; ld.param.u64 %rd75, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_6]; ld.param.f32 %f346, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+8]; ld.param.f32 %f345, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+4]; ld.param.f32 %f344, [_Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3]; cvta.to.local.u64 %rd1, %rd75; ld.const.u32 %r1, [kernel_params+1492]; add.s32 %r132, %r1, 1; ld.const.u64 %rd76, [kernel_params+2360]; cvta.to.global.u64 %rd2, %rd76; mov.u32 %r131, 0; $L__BB18_1: shr.s32 %r47, %r132, 1; add.s32 %r48, %r47, %r131; mul.wide.s32 %rd77, %r48, 16; add.s64 %rd78, %rd2, %rd77; ld.global.f32 %f347, [%rd78]; setp.gt.ftz.f32 %p1, %f347, %f341; add.s32 %r49, %r48, 1; not.b32 %r50, %r47; add.s32 %r51, %r132, %r50; selp.b32 %r131, %r131, %r49, %p1; selp.b32 %r132, %r47, %r51, %p1; setp.gt.s32 %p2, %r132, 0; @%p2 bra $L__BB18_1; add.s32 %r52, %r131, -1; max.s32 %r53, %r52, 0; add.s32 %r54, %r1, -1; min.s32 %r55, %r53, %r54; mul.wide.s32 %rd79, %r55, 16; add.s64 %rd3, %rd2, %rd79; ld.global.f32 %f348, [%rd3]; sub.ftz.f32 %f349, %f341, %f348; ld.global.f32 %f350, [%rd3+16]; sub.ftz.f32 %f351, %f350, %f348; div.approx.ftz.f32 %f1, %f349, %f351; ld.global.u32 %r7, [%rd3+4]; cvt.s64.s32 %rd4, %r7; setp.gt.s32 %p3, %r7, -1; @%p3 bra $L__BB18_51; bra.uni $L__BB18_3; $L__BB18_51: and.b32 %r79, %r45, 1073741824; setp.eq.s32 %p60, %r79, 0; ld.global.u32 %r19, [%rd3+12]; cvt.s64.s32 %rd8, %r19; @%p60 bra $L__BB18_53; ld.const.u64 %rd84, [kernel_params+2184]; cvta.to.global.u64 %rd85, %rd84; shl.b64 %rd86, %rd8, 2; add.s64 %rd87, %rd85, %rd86; ld.global.u8 %rs14, [%rd87]; and.b16 %rs15, %rs14, 128; setp.eq.s16 %p61, %rs15, 0; mov.u16 %rs23, 0; @%p61 bra $L__BB18_98; $L__BB18_53: ld.global.u32 %r20, [%rd3+8]; ld.const.u64 %rd88, [kernel_params+2184]; cvta.to.global.u64 %rd89, %rd88; shl.b64 %rd90, %rd8, 2; add.s64 %rd91, %rd89, %rd90; ld.global.u32 %r21, [%rd91]; shr.u32 %r80, %r21, 6; and.b32 %r81, %r80, 1; setp.eq.b32 %p62, %r81, 1; setp.ge.ftz.f32 %p63, %f343, 0f00000000; and.pred %p64, %p63, %p62; @%p64 bra $L__BB18_55; bra.uni $L__BB18_54; $L__BB18_55: ld.const.u64 %rd103, [kernel_params+2160]; cvta.to.global.u64 %rd104, %rd103; mul.lo.s64 %rd105, %rd8, 208; add.s64 %rd106, %rd104, %rd105; ld.global.u32 %r22, [%rd106+156]; ld.global.u32 %r23, [%rd106+152]; shl.b32 %r85, %r23, 1; cvt.rn.f32.s32 %f698, %r85; mul.ftz.f32 %f115, %f698, %f343; cvt.rzi.ftz.s32.f32 %r86, %f115; add.s32 %r87, %r85, -1; min.s32 %r24, %r86, %r87; ld.const.u64 %rd107, [kernel_params+2312]; cvta.to.global.u64 %rd10, %rd107; ld.global.u32 %r136, [%rd106+164]; mul.wide.u32 %rd108, %r136, 16; add.s64 %rd109, %rd10, %rd108; ld.global.u16 %rs19, [%rd109+12]; ld.global.u32 %r135, [%rd109+8]; ld.global.u64 %rd172, [%rd109]; setp.eq.s64 %p65, %rd172, 11; @%p65 bra $L__BB18_60; $L__BB18_56: setp.eq.s64 %p66, %rd172, 0; @%p66 bra $L__BB18_58; add.s32 %r136, %r136, 2; bra.uni $L__BB18_59; $L__BB18_58: setp.eq.s16 %p67, %rs19, 0; mov.u32 %r139, -1; mov.u32 %r136, %r135; @%p67 bra $L__BB18_61; $L__BB18_59: mul.wide.u32 %rd110, %r136, 16; add.s64 %rd111, %rd10, %rd110; ld.global.u16 %rs19, [%rd111+12]; ld.global.u32 %r135, [%rd111+8]; ld.global.u64 %rd172, [%rd111]; setp.ne.s64 %p68, %rd172, 11; @%p68 bra $L__BB18_56; $L__BB18_60: setp.eq.s16 %p69, %rs19, 0; selp.b32 %r139, -1, %r135, %p69; $L__BB18_61: ld.const.u64 %rd191, [kernel_params+2232]; cvta.to.global.u64 %rd112, %rd191; shl.b64 %rd113, %rd4, 4; add.s64 %rd114, %rd112, %rd113; ld.global.v4.u32 {%r89, %r90, %r91, %r92}, [%rd114]; setp.eq.s32 %p70, %r24, %r23; @%p70 bra $L__BB18_63; bra.uni $L__BB18_62; $L__BB18_63: ld.const.u64 %rd120, [kernel_params+2256]; cvta.to.global.u64 %rd121, %rd120; mul.wide.u32 %rd122, %r92, 12; add.s64 %rd180, %rd121, %rd122; add.s32 %r99, %r92, 1; mul.wide.u32 %rd123, %r99, 12; add.s64 %rd177, %rd121, %rd123; add.s32 %r100, %r92, 2; mul.wide.u32 %rd124, %r100, 12; add.s64 %rd174, %rd121, %rd124; add.s64 %rd179, %rd180, 4; add.s64 %rd178, %rd180, 8; add.s64 %rd176, %rd177, 4; add.s64 %rd175, %rd177, 8; add.s64 %rd173, %rd174, 4; add.s64 %rd181, %rd174, 8; bra.uni $L__BB18_64; $L__BB18_3: not.b32 %r8, %r7; ld.const.u64 %rd80, [kernel_params+2368]; cvta.to.global.u64 %rd81, %rd80; mul.wide.s32 %rd82, %r8, 192; add.s64 %rd83, %rd81, %rd82; add.s64 %rd6, %rd83, 20; ld.global.f32 %f352, [%rd83+20]; cvt.rn.f32.s32 %f353, %r44; setp.lt.ftz.f32 %p4, %f352, %f353; mov.u16 %rs23, 0; @%p4 bra $L__BB18_98; and.b32 %r56, %r45, 1073741824; setp.eq.s32 %p5, %r56, 0; @%p5 bra $L__BB18_6; bra.uni $L__BB18_5; $L__BB18_6: ld.global.u32 %r133, [%rd6+-4]; bra.uni $L__BB18_7; $L__BB18_54: ld.const.u64 %rd191, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd191; shl.b64 %rd93, %rd4, 4; add.s64 %rd94, %rd92, %rd93; ld.const.u64 %rd95, [kernel_params+2256]; cvta.to.global.u64 %rd96, %rd95; ld.global.u32 %r82, [%rd94+12]; mul.wide.u32 %rd97, %r82, 12; add.s64 %rd98, %rd96, %rd97; ld.global.f32 %f1309, [%rd98]; ld.global.f32 %f1308, [%rd98+4]; ld.global.f32 %f1307, [%rd98+8]; add.s32 %r83, %r82, 1; mul.wide.u32 %rd99, %r83, 12; add.s64 %rd100, %rd96, %rd99; ld.global.f32 %f1306, [%rd100]; ld.global.f32 %f1305, [%rd100+4]; ld.global.f32 %f1304, [%rd100+8]; add.s32 %r84, %r82, 2; mul.wide.u32 %rd101, %r84, 12; add.s64 %rd102, %rd96, %rd101; ld.global.f32 %f1303, [%rd102]; ld.global.f32 %f1302, [%rd102+4]; ld.global.f32 %f1310, [%rd102+8]; mov.u16 %rs22, 0; bra.uni $L__BB18_68; $L__BB18_62: setp.gt.s32 %p71, %r24, %r23; selp.b32 %r93, -1, 0, %p71; add.s32 %r94, %r24, %r93; mad.lo.s32 %r95, %r94, %r22, %r139; ld.const.u64 %rd115, [kernel_params+2336]; cvta.to.global.u64 %rd116, %rd115; add.s32 %r96, %r89, %r95; mul.wide.u32 %rd117, %r96, 12; add.s64 %rd180, %rd116, %rd117; add.s32 %r97, %r90, %r95; mul.wide.u32 %rd118, %r97, 12; add.s64 %rd177, %rd116, %rd118; add.s32 %r98, %r91, %r95; mul.wide.u32 %rd119, %r98, 12; add.s64 %rd174, %rd116, %rd119; add.s64 %rd179, %rd180, 4; add.s64 %rd178, %rd180, 8; add.s64 %rd176, %rd177, 4; add.s64 %rd175, %rd177, 8; add.s64 %rd173, %rd174, 4; add.s64 %rd181, %rd174, 8; $L__BB18_64: ld.global.f32 %f116, [%rd181]; ld.global.f32 %f117, [%rd180]; ld.global.f32 %f118, [%rd179]; ld.global.f32 %f119, [%rd178]; ld.global.f32 %f120, [%rd177]; ld.global.f32 %f121, [%rd176]; ld.global.f32 %f122, [%rd175]; ld.global.f32 %f123, [%rd174]; ld.global.f32 %f124, [%rd173]; add.s32 %r39, %r24, 1; setp.eq.s32 %p72, %r39, %r23; @%p72 bra $L__BB18_66; bra.uni $L__BB18_65; $L__BB18_66: ld.const.u64 %rd130, [kernel_params+2256]; cvta.to.global.u64 %rd131, %rd130; mul.wide.u32 %rd132, %r92, 12; add.s64 %rd182, %rd131, %rd132; add.s32 %r107, %r92, 1; mul.wide.u32 %rd133, %r107, 12; add.s64 %rd185, %rd131, %rd133; add.s32 %r108, %r92, 2; mul.wide.u32 %rd134, %r108, 12; add.s64 %rd188, %rd131, %rd134; add.s64 %rd183, %rd182, 4; add.s64 %rd184, %rd182, 8; add.s64 %rd186, %rd185, 4; add.s64 %rd187, %rd185, 8; add.s64 %rd189, %rd188, 4; add.s64 %rd190, %rd188, 8; bra.uni $L__BB18_67; $L__BB18_65: setp.ge.s32 %p73, %r24, %r23; selp.b32 %r101, -1, 0, %p73; add.s32 %r102, %r39, %r101; mad.lo.s32 %r103, %r102, %r22, %r139; ld.const.u64 %rd125, [kernel_params+2336]; cvta.to.global.u64 %rd126, %rd125; add.s32 %r104, %r89, %r103; mul.wide.u32 %rd127, %r104, 12; add.s64 %rd182, %rd126, %rd127; add.s32 %r105, %r90, %r103; mul.wide.u32 %rd128, %r105, 12; add.s64 %rd185, %rd126, %rd128; add.s32 %r106, %r91, %r103; mul.wide.u32 %rd129, %r106, 12; add.s64 %rd188, %rd126, %rd129; add.s64 %rd183, %rd182, 4; add.s64 %rd184, %rd182, 8; add.s64 %rd186, %rd185, 4; add.s64 %rd187, %rd185, 8; add.s64 %rd189, %rd188, 4; add.s64 %rd190, %rd188, 8; $L__BB18_67: cvt.rn.f32.s32 %f699, %r24; sub.ftz.f32 %f700, %f115, %f699; mov.f32 %f701, 0f3F800000; sub.ftz.f32 %f702, %f701, %f700; ld.global.f32 %f703, [%rd182]; mul.ftz.f32 %f704, %f700, %f703; ld.global.f32 %f705, [%rd183]; mul.ftz.f32 %f706, %f700, %f705; ld.global.f32 %f707, [%rd184]; mul.ftz.f32 %f708, %f700, %f707; fma.rn.ftz.f32 %f1309, %f702, %f117, %f704; fma.rn.ftz.f32 %f1308, %f702, %f118, %f706; fma.rn.ftz.f32 %f1307, %f702, %f119, %f708; ld.global.f32 %f709, [%rd185]; mul.ftz.f32 %f710, %f700, %f709; ld.global.f32 %f711, [%rd186]; mul.ftz.f32 %f712, %f700, %f711; ld.global.f32 %f713, [%rd187]; mul.ftz.f32 %f714, %f700, %f713; fma.rn.ftz.f32 %f1306, %f702, %f120, %f710; fma.rn.ftz.f32 %f1305, %f702, %f121, %f712; fma.rn.ftz.f32 %f1304, %f702, %f122, %f714; ld.global.f32 %f715, [%rd188]; mul.ftz.f32 %f716, %f700, %f715; ld.global.f32 %f717, [%rd189]; mul.ftz.f32 %f718, %f700, %f717; ld.global.f32 %f719, [%rd190]; mul.ftz.f32 %f720, %f700, %f719; fma.rn.ftz.f32 %f1303, %f702, %f123, %f716; fma.rn.ftz.f32 %f1302, %f702, %f124, %f718; fma.rn.ftz.f32 %f1310, %f702, %f116, %f720; mov.u16 %rs22, 1; $L__BB18_68: and.b32 %r40, %r21, 4; setp.ne.s32 %p74, %r40, 0; @%p74 bra $L__BB18_76; and.b32 %r109, %r21, 2; setp.eq.s32 %p75, %r109, 0; ld.const.u64 %rd135, [kernel_params+2160]; cvta.to.global.u64 %rd136, %rd135; mul.lo.s64 %rd137, %rd8, 208; add.s64 %rd70, %rd136, %rd137; @%p75 bra $L__BB18_74; bra.uni $L__BB18_70; $L__BB18_74: ld.global.v4.f32 {%f1290, %f1291, %f1292, %f1293}, [%rd70]; ld.global.v4.f32 {%f1294, %f1295, %f1296, %f1297}, [%rd70+16]; ld.global.v4.f32 {%f1298, %f1299, %f1300, %f1301}, [%rd70+32]; bra.uni $L__BB18_75; $L__BB18_70: ld.const.u64 %rd138, [kernel_params+2176]; cvta.to.global.u64 %rd139, %rd138; ld.global.u32 %rd140, [%rd70+168]; ld.global.u32 %r110, [%rd70+152]; shl.b32 %r111, %r110, 1; cvt.rn.f32.s32 %f721, %r111; selp.f32 %f722, %f343, 0f3F000000, %p63; mul.ftz.f32 %f723, %f722, %f721; cvt.rzi.ftz.s32.f32 %r112, %f723; add.s32 %r113, %r111, -1; min.s32 %r114, %r112, %r113; cvt.rn.f32.s32 %f724, %r114; sub.ftz.f32 %f143, %f723, %f724; cvt.s64.s32 %rd141, %r114; add.s64 %rd142, %rd141, %rd140; shl.b64 %rd143, %rd142, 6; add.s64 %rd71, %rd139, %rd143; ld.global.v4.f32 {%f725, %f726, %f727, %f728}, [%rd71]; ld.global.v4.f32 {%f729, %f730, %f731, %f732}, [%rd71+64]; mul.ftz.f32 %f733, %f726, %f730; fma.rn.ftz.f32 %f734, %f725, %f729, %f733; mul.ftz.f32 %f735, %f728, %f732; fma.rn.ftz.f32 %f736, %f727, %f731, %f735; add.ftz.f32 %f152, %f734, %f736; setp.gt.ftz.f32 %p77, %f152, 0f3F7FDF3B; mov.f32 %f737, 0f3F800000; sub.ftz.f32 %f153, %f737, %f143; @%p77 bra $L__BB18_72; bra.uni $L__BB18_71; $L__BB18_72: mul.ftz.f32 %f791, %f725, %f153; fma.rn.ftz.f32 %f792, %f729, %f143, %f791; mul.ftz.f32 %f793, %f726, %f153; fma.rn.ftz.f32 %f794, %f730, %f143, %f793; mul.ftz.f32 %f795, %f727, %f153; fma.rn.ftz.f32 %f796, %f731, %f143, %f795; mul.ftz.f32 %f797, %f728, %f153; fma.rn.ftz.f32 %f798, %f732, %f143, %f797; mul.ftz.f32 %f799, %f794, %f794; fma.rn.ftz.f32 %f800, %f792, %f792, %f799; mul.ftz.f32 %f801, %f798, %f798; fma.rn.ftz.f32 %f802, %f796, %f796, %f801; add.ftz.f32 %f803, %f800, %f802; rsqrt.approx.ftz.f32 %f804, %f803; mul.ftz.f32 %f1289, %f798, %f804; mul.ftz.f32 %f1288, %f796, %f804; mul.ftz.f32 %f1287, %f794, %f804; mul.ftz.f32 %f1286, %f792, %f804; bra.uni $L__BB18_73; $L__BB18_5: ld.global.u32 %r133, [%rd6+-4]; and.b32 %r57, %r133, 4194304; setp.eq.s32 %p6, %r57, 0; @%p6 bra $L__BB18_7; bra.uni $L__BB18_98; $L__BB18_7: ld.global.u32 %r12, [%rd6+-20]; st.local.u32 [%rd1+76], %r12; st.local.u32 [%rd1+64], %r133; mov.u32 %r134, -1; st.local.u32 [%rd1+56], %r134; st.local.u32 [%rd1+60], %r134; st.local.u32 [%rd1+68], %r8; st.local.f32 [%rd1+40], %f1; st.local.f32 [%rd1+44], %f342; setp.eq.s32 %p7, %r7, 0; @%p7 bra $L__BB18_9; ld.global.u32 %r134, [%rd6+24]; $L__BB18_9: st.local.u32 [%rd1+72], %r134; add.s32 %r59, %r12, -1; setp.lt.u32 %p8, %r59, 2; @%p8 bra $L__BB18_50; bra.uni $L__BB18_10; $L__BB18_50: mov.u32 %r76, 0; st.local.u32 [%rd1], %r76; st.local.u32 [%rd1+4], %r76; st.local.u32 [%rd1+8], %r76; st.local.u32 [%rd1+12], %r76; st.local.u32 [%rd1+16], %r76; st.local.u32 [%rd1+20], %r76; st.local.u32 [%rd1+24], %r76; st.local.u32 [%rd1+28], %r76; st.local.u32 [%rd1+32], %r76; mov.u32 %r77, 1065353216; st.local.u32 [%rd1+48], %r77; mov.u32 %r78, 2139095039; st.local.u32 [%rd1+36], %r78; mov.u16 %rs23, 1; bra.uni $L__BB18_98; $L__BB18_10: ld.global.f32 %f1261, [%rd6+-16]; ld.global.f32 %f1260, [%rd6+-12]; ld.global.f32 %f1259, [%rd6+-8]; st.local.f32 [%rd1], %f1261; st.local.f32 [%rd1+4], %f1260; st.local.f32 [%rd1+8], %f1259; setp.eq.s32 %p9, %r12, 4; @%p9 bra $L__BB18_32; bra.uni $L__BB18_11; $L__BB18_32: ld.global.f32 %f57, [%rd6+124]; ld.global.v4.f32 {%f561, %f562, %f563, %f564}, [%rd6+140]; sub.ftz.f32 %f565, %f344, %f1261; sub.ftz.f32 %f566, %f345, %f1260; mul.ftz.f32 %f567, %f566, %f566; fma.rn.ftz.f32 %f568, %f565, %f565, %f567; sub.ftz.f32 %f569, %f346, %f1259; fma.rn.ftz.f32 %f570, %f569, %f569, %f568; sqrt.approx.ftz.f32 %f571, %f570; div.approx.ftz.f32 %f62, %f565, %f571; div.approx.ftz.f32 %f63, %f566, %f571; div.approx.ftz.f32 %f64, %f569, %f571; st.local.f32 [%rd1], %f1261; st.local.f32 [%rd1+4], %f1260; st.local.f32 [%rd1+8], %f1259; setp.leu.ftz.f32 %p35, %f57, 0f00000000; @%p35 bra $L__BB18_34; setp.neu.ftz.f32 %p36, %f62, %f63; setp.neu.ftz.f32 %p37, %f62, %f64; or.pred %p38, %p36, %p37; neg.ftz.f32 %f572, %f64; selp.f32 %f573, %f572, %f64, %p38; add.ftz.f32 %f574, %f62, %f573; neg.ftz.f32 %f575, %f63; selp.f32 %f576, %f63, %f575, %p38; sub.ftz.f32 %f577, %f576, %f62; sub.ftz.f32 %f578, %f64, %f63; mul.ftz.f32 %f579, %f574, %f574; fma.rn.ftz.f32 %f580, %f578, %f578, %f579; fma.rn.ftz.f32 %f581, %f577, %f577, %f580; sqrt.approx.ftz.f32 %f582, %f581; div.approx.ftz.f32 %f583, %f578, %f582; div.approx.ftz.f32 %f584, %f574, %f582; div.approx.ftz.f32 %f585, %f577, %f582; mul.ftz.f32 %f586, %f63, %f585; mul.ftz.f32 %f587, %f64, %f584; sub.ftz.f32 %f588, %f586, %f587; mul.ftz.f32 %f589, %f64, %f583; mul.ftz.f32 %f590, %f62, %f585; sub.ftz.f32 %f591, %f589, %f590; mul.ftz.f32 %f592, %f62, %f584; mul.ftz.f32 %f593, %f63, %f583; sub.ftz.f32 %f594, %f592, %f593; mul.ftz.f32 %f595, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f596, %f595; sqrt.approx.ftz.f32 %f597, %f342; mul.ftz.f32 %f598, %f597, %f596; sin.approx.ftz.f32 %f599, %f595; mul.ftz.f32 %f600, %f597, %f599; mul.ftz.f32 %f601, %f588, %f600; mul.ftz.f32 %f602, %f591, %f600; mul.ftz.f32 %f603, %f594, %f600; fma.rn.ftz.f32 %f604, %f583, %f598, %f601; fma.rn.ftz.f32 %f605, %f584, %f598, %f602; fma.rn.ftz.f32 %f606, %f585, %f598, %f603; fma.rn.ftz.f32 %f1261, %f57, %f604, %f1261; fma.rn.ftz.f32 %f1260, %f57, %f605, %f1260; fma.rn.ftz.f32 %f1259, %f57, %f606, %f1259; st.local.f32 [%rd1], %f1261; st.local.f32 [%rd1+4], %f1260; st.local.f32 [%rd1+8], %f1259; $L__BB18_34: ld.global.f32 %f71, [%rd6+128]; st.local.f32 [%rd1+48], %f71; sub.ftz.f32 %f608, %f1261, %f344; sub.ftz.f32 %f609, %f1260, %f345; mul.ftz.f32 %f610, %f609, %f609; fma.rn.ftz.f32 %f611, %f608, %f608, %f610; sub.ftz.f32 %f612, %f1259, %f346; fma.rn.ftz.f32 %f613, %f612, %f612, %f611; sqrt.approx.ftz.f32 %f72, %f613; st.local.f32 [%rd1+36], %f72; rsqrt.approx.ftz.f32 %f614, %f613; mul.ftz.f32 %f73, %f608, %f614; mul.ftz.f32 %f74, %f609, %f614; mul.ftz.f32 %f75, %f612, %f614; st.local.f32 [%rd1+24], %f73; st.local.f32 [%rd1+28], %f74; st.local.f32 [%rd1+32], %f75; neg.ftz.f32 %f76, %f73; neg.ftz.f32 %f77, %f74; neg.ftz.f32 %f78, %f75; st.local.f32 [%rd1+12], %f76; st.local.f32 [%rd1+16], %f77; st.local.f32 [%rd1+20], %f78; mul.ftz.f32 %f79, %f71, 0f3DA2F983; st.local.f32 [%rd1+52], %f79; ld.global.v2.f32 {%f615, %f616}, [%rd6+132]; mul.ftz.f32 %f617, %f561, %f73; mul.ftz.f32 %f618, %f562, %f74; neg.ftz.f32 %f619, %f618; sub.ftz.f32 %f620, %f619, %f617; mul.ftz.f32 %f621, %f563, %f75; sub.ftz.f32 %f82, %f620, %f621; setp.le.ftz.f32 %p39, %f82, %f615; mov.f32 %f1275, 0f00000000; mov.f32 %f1270, %f1275; @%p39 bra $L__BB18_37; sub.ftz.f32 %f83, %f82, %f615; setp.geu.ftz.f32 %p40, %f83, %f616; setp.eq.ftz.f32 %p41, %f616, 0f00000000; or.pred %p42, %p41, %p40; mov.f32 %f1270, %f82; @%p42 bra $L__BB18_37; div.approx.ftz.f32 %f622, %f83, %f616; mul.ftz.f32 %f623, %f622, %f622; mul.ftz.f32 %f624, %f623, 0fC0000000; mul.ftz.f32 %f625, %f622, %f624; fma.rn.ftz.f32 %f626, %f623, 0f40400000, %f625; mul.ftz.f32 %f1270, %f82, %f626; $L__BB18_37: mul.ftz.f32 %f629, %f1270, %f79; st.local.f32 [%rd1+52], %f629; mul.ftz.f32 %f630, %f74, %f74; fma.rn.ftz.f32 %f631, %f73, %f73, %f630; fma.rn.ftz.f32 %f86, %f75, %f75, %f631; setp.leu.ftz.f32 %p43, %f86, 0f00000000; mov.f32 %f1273, %f1275; mov.f32 %f1274, %f1275; @%p43 bra $L__BB18_46; setp.eq.ftz.f32 %p44, %f73, 0f80000000; mov.f32 %f1273, 0f00000000; setp.eq.ftz.f32 %p45, %f74, 0f80000000; and.pred %p46, %p44, %p45; @%p46 bra $L__BB18_45; abs.ftz.f32 %f87, %f77; setp.eq.ftz.f32 %p47, %f87, 0f00000000; abs.ftz.f32 %f88, %f76; setp.eq.ftz.f32 %p48, %f88, 0f00000000; and.pred %p49, %p47, %p48; mov.b32 %r17, %f77; mov.b32 %r68, %f76; and.b32 %r18, %r68, -2147483648; @%p49 bra $L__BB18_43; bra.uni $L__BB18_40; $L__BB18_43: shr.s32 %r73, %r17, 31; and.b32 %r74, %r73, 1078530011; or.b32 %r75, %r74, %r18; mov.b32 %f1271, %r75; bra.uni $L__BB18_44; $L__BB18_71: mov.f32 %f738, 0fBF800000; max.ftz.f32 %f739, %f152, %f738; min.ftz.f32 %f741, %f739, %f737; abs.ftz.f32 %f742, %f741; sub.ftz.f32 %f743, %f737, %f742; mul.ftz.f32 %f744, %f743, 0f3F000000; sqrt.approx.ftz.f32 %f745, %f744; setp.gt.ftz.f32 %p78, %f742, 0f3F11EB85; selp.f32 %f746, %f745, %f742, %p78; mul.ftz.f32 %f747, %f746, %f746; mov.f32 %f748, 0f3C94D2E9; mov.f32 %f749, 0f3D53F941; fma.rn.ftz.f32 %f750, %f749, %f747, %f748; mov.f32 %f751, 0f3D3F841F; fma.rn.ftz.f32 %f752, %f750, %f747, %f751; mov.f32 %f753, 0f3D994929; fma.rn.ftz.f32 %f754, %f752, %f747, %f753; mov.f32 %f755, 0f3E2AAB94; fma.rn.ftz.f32 %f756, %f754, %f747, %f755; mul.ftz.f32 %f757, %f747, %f756; fma.rn.ftz.f32 %f758, %f757, %f746, %f746; add.ftz.f32 %f759, %f758, %f758; mov.f32 %f760, 0f3FC90FDB; sub.ftz.f32 %f761, %f760, %f758; selp.f32 %f762, %f759, %f761, %p78; setp.lt.ftz.f32 %p79, %f741, 0f00000000; mov.f32 %f763, 0f40490FDB; sub.ftz.f32 %f764, %f763, %f762; selp.f32 %f765, %f764, %f762, %p79; mul.ftz.f32 %f766, %f725, %f152; sub.ftz.f32 %f767, %f729, %f766; mul.ftz.f32 %f768, %f726, %f152; sub.ftz.f32 %f769, %f730, %f768; mul.ftz.f32 %f770, %f727, %f152; sub.ftz.f32 %f771, %f731, %f770; mul.ftz.f32 %f772, %f728, %f152; sub.ftz.f32 %f773, %f732, %f772; mul.ftz.f32 %f774, %f769, %f769; fma.rn.ftz.f32 %f775, %f767, %f767, %f774; mul.ftz.f32 %f776, %f773, %f773; fma.rn.ftz.f32 %f777, %f771, %f771, %f776; add.ftz.f32 %f778, %f775, %f777; rsqrt.approx.ftz.f32 %f779, %f778; mul.ftz.f32 %f780, %f767, %f779; mul.ftz.f32 %f781, %f769, %f779; mul.ftz.f32 %f782, %f771, %f779; mul.ftz.f32 %f783, %f773, %f779; mul.ftz.f32 %f784, %f143, %f765; cos.approx.ftz.f32 %f785, %f784; sin.approx.ftz.f32 %f786, %f784; mul.ftz.f32 %f787, %f780, %f786; mul.ftz.f32 %f788, %f781, %f786; mul.ftz.f32 %f789, %f782, %f786; mul.ftz.f32 %f790, %f783, %f786; fma.rn.ftz.f32 %f1289, %f728, %f785, %f790; fma.rn.ftz.f32 %f1288, %f727, %f785, %f789; fma.rn.ftz.f32 %f1287, %f726, %f785, %f788; fma.rn.ftz.f32 %f1286, %f725, %f785, %f787; $L__BB18_73: ld.global.v4.f32 {%f805, %f806, %f807, %f808}, [%rd71+16]; ld.global.v4.f32 {%f813, %f814, %f815, %f816}, [%rd71+80]; mul.ftz.f32 %f821, %f143, %f813; mul.ftz.f32 %f822, %f143, %f814; mul.ftz.f32 %f823, %f143, %f815; mul.ftz.f32 %f824, %f143, %f816; fma.rn.ftz.f32 %f825, %f153, %f808, %f824; ld.global.v4.f32 {%f826, %f827, %f828, %f829}, [%rd71+32]; ld.global.v4.f32 {%f834, %f835, %f836, %f837}, [%rd71+96]; mul.ftz.f32 %f842, %f143, %f834; mul.ftz.f32 %f843, %f143, %f835; mul.ftz.f32 %f844, %f143, %f836; mul.ftz.f32 %f845, %f143, %f837; fma.rn.ftz.f32 %f846, %f153, %f826, %f842; fma.rn.ftz.f32 %f847, %f153, %f827, %f843; fma.rn.ftz.f32 %f848, %f153, %f828, %f844; fma.rn.ftz.f32 %f849, %f153, %f829, %f845; ld.global.v4.f32 {%f850, %f851, %f852, %f853}, [%rd71+48]; ld.global.v4.f32 {%f858, %f859, %f860, %f861}, [%rd71+112]; mul.ftz.f32 %f866, %f143, %f858; mul.ftz.f32 %f867, %f143, %f859; mul.ftz.f32 %f868, %f143, %f860; mul.ftz.f32 %f869, %f143, %f861; fma.rn.ftz.f32 %f870, %f153, %f850, %f866; fma.rn.ftz.f32 %f871, %f153, %f851, %f867; fma.rn.ftz.f32 %f872, %f153, %f852, %f868; fma.rn.ftz.f32 %f873, %f153, %f853, %f869; mul.ftz.f32 %f874, %f1289, 0f3FB504F3; mul.ftz.f32 %f875, %f1286, 0f3FB504F3; mul.ftz.f32 %f876, %f1287, 0f3FB504F3; mul.ftz.f32 %f877, %f1288, 0f3FB504F3; mul.ftz.f32 %f878, %f874, %f875; mul.ftz.f32 %f879, %f874, %f876; mul.ftz.f32 %f880, %f874, %f877; mul.ftz.f32 %f881, %f875, %f875; mul.ftz.f32 %f882, %f875, %f876; mul.ftz.f32 %f883, %f875, %f877; mul.ftz.f32 %f884, %f876, %f876; mul.ftz.f32 %f885, %f876, %f877; mul.ftz.f32 %f886, %f877, %f877; sub.ftz.f32 %f888, %f737, %f884; sub.ftz.f32 %f889, %f888, %f886; sub.ftz.f32 %f890, %f882, %f880; add.ftz.f32 %f891, %f879, %f883; add.ftz.f32 %f892, %f880, %f882; sub.ftz.f32 %f893, %f737, %f881; sub.ftz.f32 %f894, %f893, %f886; sub.ftz.f32 %f895, %f885, %f878; sub.ftz.f32 %f896, %f883, %f879; add.ftz.f32 %f897, %f878, %f885; sub.ftz.f32 %f898, %f893, %f884; mul.ftz.f32 %f899, %f890, %f848; fma.rn.ftz.f32 %f900, %f889, %f825, %f899; mul.ftz.f32 %f901, %f890, %f849; fma.rn.ftz.f32 %f902, %f889, %f846, %f901; mul.ftz.f32 %f903, %f890, %f870; fma.rn.ftz.f32 %f904, %f889, %f847, %f903; mul.ftz.f32 %f905, %f894, %f848; fma.rn.ftz.f32 %f906, %f892, %f825, %f905; mul.ftz.f32 %f907, %f894, %f849; fma.rn.ftz.f32 %f908, %f892, %f846, %f907; mul.ftz.f32 %f909, %f894, %f870; fma.rn.ftz.f32 %f910, %f892, %f847, %f909; mul.ftz.f32 %f911, %f897, %f848; fma.rn.ftz.f32 %f912, %f896, %f825, %f911; mul.ftz.f32 %f913, %f897, %f849; fma.rn.ftz.f32 %f914, %f896, %f846, %f913; mul.ftz.f32 %f915, %f897, %f870; fma.rn.ftz.f32 %f916, %f896, %f847, %f915; fma.rn.ftz.f32 %f1296, %f895, %f873, %f910; fma.rn.ftz.f32 %f1295, %f895, %f872, %f908; fma.rn.ftz.f32 %f1294, %f895, %f871, %f906; fma.rn.ftz.f32 %f1300, %f898, %f873, %f916; fma.rn.ftz.f32 %f1299, %f898, %f872, %f914; fma.rn.ftz.f32 %f1298, %f898, %f871, %f912; fma.rn.ftz.f32 %f1292, %f891, %f873, %f904; fma.rn.ftz.f32 %f1291, %f891, %f872, %f902; fma.rn.ftz.f32 %f1290, %f891, %f871, %f900; fma.rn.ftz.f32 %f1301, %f153, %f807, %f823; fma.rn.ftz.f32 %f1297, %f153, %f806, %f822; fma.rn.ftz.f32 %f1293, %f153, %f805, %f821; $L__BB18_75: mul.ftz.f32 %f929, %f1308, %f1291; fma.rn.ftz.f32 %f930, %f1309, %f1290, %f929; fma.rn.ftz.f32 %f931, %f1307, %f1292, %f930; add.ftz.f32 %f202, %f1293, %f931; mul.ftz.f32 %f932, %f1308, %f1295; fma.rn.ftz.f32 %f933, %f1309, %f1294, %f932; fma.rn.ftz.f32 %f934, %f1307, %f1296, %f933; add.ftz.f32 %f203, %f1297, %f934; mul.ftz.f32 %f935, %f1308, %f1299; fma.rn.ftz.f32 %f936, %f1309, %f1298, %f935; fma.rn.ftz.f32 %f937, %f1307, %f1300, %f936; add.ftz.f32 %f1307, %f1301, %f937; mul.ftz.f32 %f938, %f1305, %f1291; fma.rn.ftz.f32 %f939, %f1306, %f1290, %f938; fma.rn.ftz.f32 %f940, %f1304, %f1292, %f939; add.ftz.f32 %f205, %f1293, %f940; mul.ftz.f32 %f941, %f1305, %f1295; fma.rn.ftz.f32 %f942, %f1306, %f1294, %f941; fma.rn.ftz.f32 %f943, %f1304, %f1296, %f942; add.ftz.f32 %f206, %f1297, %f943; mul.ftz.f32 %f944, %f1305, %f1299; fma.rn.ftz.f32 %f945, %f1306, %f1298, %f944; fma.rn.ftz.f32 %f946, %f1304, %f1300, %f945; add.ftz.f32 %f1304, %f1301, %f946; mul.ftz.f32 %f947, %f1302, %f1291; fma.rn.ftz.f32 %f948, %f1303, %f1290, %f947; fma.rn.ftz.f32 %f949, %f1310, %f1292, %f948; add.ftz.f32 %f208, %f1293, %f949; mul.ftz.f32 %f950, %f1302, %f1295; fma.rn.ftz.f32 %f951, %f1303, %f1294, %f950; fma.rn.ftz.f32 %f952, %f1310, %f1296, %f951; add.ftz.f32 %f209, %f1297, %f952; mul.ftz.f32 %f953, %f1302, %f1299; fma.rn.ftz.f32 %f954, %f1303, %f1298, %f953; fma.rn.ftz.f32 %f955, %f1310, %f1300, %f954; add.ftz.f32 %f1310, %f1301, %f955; mov.u16 %rs22, 1; mov.f32 %f1302, %f209; mov.f32 %f1303, %f208; mov.f32 %f1305, %f206; mov.f32 %f1306, %f205; mov.f32 %f1308, %f203; mov.f32 %f1309, %f202; $L__BB18_76: sub.ftz.f32 %f956, %f1310, %f1307; sub.ftz.f32 %f957, %f1305, %f1308; mul.ftz.f32 %f958, %f957, %f956; sub.ftz.f32 %f959, %f1302, %f1308; sub.ftz.f32 %f960, %f1304, %f1307; mul.ftz.f32 %f961, %f960, %f959; sub.ftz.f32 %f1316, %f958, %f961; sub.ftz.f32 %f962, %f1303, %f1309; mul.ftz.f32 %f963, %f960, %f962; sub.ftz.f32 %f964, %f1306, %f1309; mul.ftz.f32 %f965, %f964, %f956; sub.ftz.f32 %f1315, %f963, %f965; mul.ftz.f32 %f966, %f959, %f964; mul.ftz.f32 %f967, %f957, %f962; sub.ftz.f32 %f1314, %f966, %f967; mul.ftz.f32 %f968, %f1315, %f1315; fma.rn.ftz.f32 %f969, %f1316, %f1316, %f968; fma.rn.ftz.f32 %f970, %f1314, %f1314, %f969; sqrt.approx.ftz.f32 %f223, %f970; setp.neu.ftz.f32 %p80, %f223, 0f00000000; @%p80 bra $L__BB18_77; bra.uni $L__BB18_78; $L__BB18_77: div.approx.ftz.f32 %f1314, %f1314, %f223; div.approx.ftz.f32 %f1315, %f1315, %f223; div.approx.ftz.f32 %f1316, %f1316, %f223; $L__BB18_78: st.local.f32 [%rd1+12], %f1316; st.local.f32 [%rd1+16], %f1315; st.local.f32 [%rd1+20], %f1314; mul.ftz.f32 %f236, %f223, 0f3F000000; and.b32 %r115, %r21, 8; setp.eq.s32 %p81, %r115, 0; @%p81 bra $L__BB18_80; neg.ftz.f32 %f1316, %f1316; st.local.f32 [%rd1+12], %f1316; neg.ftz.f32 %f1315, %f1315; st.local.f32 [%rd1+16], %f1315; neg.ftz.f32 %f1314, %f1314; st.local.f32 [%rd1+20], %f1314; $L__BB18_80: mov.u32 %r117, 1065353216; st.local.u32 [%rd1+52], %r117; ld.const.u64 %rd144, [kernel_params+2216]; cvta.to.global.u64 %rd145, %rd144; shl.b64 %rd146, %rd4, 2; add.s64 %rd147, %rd145, %rd146; ld.global.u32 %r118, [%rd147]; st.local.u32 [%rd1+56], %r19; st.local.u32 [%rd1+60], %r7; mov.u32 %r140, -1; st.local.u32 [%rd1+68], %r140; or.b32 %r41, %r118, 268435456; st.local.u32 [%rd1+64], %r41; mov.u32 %r119, 5; st.local.u32 [%rd1+76], %r119; setp.eq.s32 %p82, %r19, -1; @%p82 bra $L__BB18_82; ld.const.u64 %rd148, [kernel_params+2160]; cvta.to.global.u64 %rd149, %rd148; mul.lo.s64 %rd150, %rd8, 208; add.s64 %rd151, %rd149, %rd150; ld.global.u32 %r140, [%rd151+192]; $L__BB18_82: st.local.u32 [%rd1+72], %r140; setp.lt.ftz.f32 %p83, %f1, %f342; @%p83 bra $L__BB18_84; bra.uni $L__BB18_83; $L__BB18_84: mul.ftz.f32 %f1317, %f1, 0f3F000000; sub.ftz.f32 %f1318, %f342, %f1317; bra.uni $L__BB18_85; $L__BB18_83: mul.ftz.f32 %f1318, %f342, 0f3F000000; sub.ftz.f32 %f1317, %f1, %f1318; $L__BB18_85: mov.f32 %f972, 0f3F800000; sub.ftz.f32 %f973, %f972, %f1317; sub.ftz.f32 %f974, %f973, %f1318; mul.ftz.f32 %f975, %f1306, %f1318; fma.rn.ftz.f32 %f976, %f1309, %f1317, %f975; mul.ftz.f32 %f977, %f1305, %f1318; fma.rn.ftz.f32 %f978, %f1308, %f1317, %f977; mul.ftz.f32 %f979, %f1304, %f1318; fma.rn.ftz.f32 %f980, %f1307, %f1317, %f979; fma.rn.ftz.f32 %f981, %f1303, %f974, %f976; fma.rn.ftz.f32 %f982, %f1302, %f974, %f978; fma.rn.ftz.f32 %f983, %f1310, %f974, %f980; st.local.f32 [%rd1], %f981; st.local.f32 [%rd1+4], %f982; st.local.f32 [%rd1+8], %f983; sub.ftz.f32 %f984, %f981, %f344; sub.ftz.f32 %f985, %f982, %f345; sub.ftz.f32 %f986, %f983, %f346; mul.ftz.f32 %f987, %f985, %f985; fma.rn.ftz.f32 %f988, %f984, %f984, %f987; fma.rn.ftz.f32 %f989, %f986, %f986, %f988; sqrt.approx.ftz.f32 %f249, %f989; st.local.f32 [%rd1+36], %f249; rsqrt.approx.ftz.f32 %f990, %f989; mul.ftz.f32 %f991, %f984, %f990; mul.ftz.f32 %f992, %f985, %f990; mul.ftz.f32 %f993, %f986, %f990; st.local.f32 [%rd1+24], %f991; st.local.f32 [%rd1+28], %f992; st.local.f32 [%rd1+32], %f993; mul.ftz.f32 %f994, %f1316, %f991; mul.ftz.f32 %f995, %f1315, %f992; neg.ftz.f32 %f996, %f995; sub.ftz.f32 %f997, %f996, %f994; mul.ftz.f32 %f998, %f993, %f1314; sub.ftz.f32 %f999, %f997, %f998; abs.ftz.f32 %f250, %f999; setp.eq.ftz.f32 %p84, %f250, 0f00000000; mov.f32 %f1345, 0f00000000; @%p84 bra $L__BB18_87; ld.const.f32 %f1000, [kernel_params+1500]; mul.ftz.f32 %f1001, %f249, %f249; mul.ftz.f32 %f1002, %f1001, %f1000; div.approx.ftz.f32 %f1345, %f1002, %f250; $L__BB18_87: st.local.f32 [%rd1+48], %f1345; setp.eq.ftz.f32 %p85, %f236, 0f00000000; setp.eq.s16 %p86, %rs22, 0; or.pred %p87, %p86, %p85; @%p87 bra $L__BB18_97; cvta.to.global.u64 %rd152, %rd191; shl.b64 %rd153, %rd4, 4; add.s64 %rd154, %rd152, %rd153; ld.const.u64 %rd155, [kernel_params+2256]; cvta.to.global.u64 %rd156, %rd155; ld.global.u32 %r120, [%rd154+12]; mul.wide.u32 %rd157, %r120, 12; add.s64 %rd158, %rd156, %rd157; ld.global.f32 %f1343, [%rd158]; ld.global.f32 %f1342, [%rd158+4]; ld.global.f32 %f1341, [%rd158+8]; add.s32 %r121, %r120, 1; mul.wide.u32 %rd159, %r121, 12; add.s64 %rd160, %rd156, %rd159; ld.global.f32 %f1340, [%rd160]; ld.global.f32 %f1339, [%rd160+4]; ld.global.f32 %f1338, [%rd160+8]; add.s32 %r122, %r120, 2; mul.wide.u32 %rd161, %r122, 12; add.s64 %rd162, %rd156, %rd161; ld.global.f32 %f1337, [%rd162]; ld.global.f32 %f1336, [%rd162+4]; ld.global.f32 %f1344, [%rd162+8]; @%p74 bra $L__BB18_96; and.b32 %r123, %r21, 2; setp.eq.s32 %p89, %r123, 0; ld.const.u64 %rd163, [kernel_params+2160]; cvta.to.global.u64 %rd164, %rd163; mul.lo.s64 %rd165, %rd8, 208; add.s64 %rd73, %rd164, %rd165; @%p89 bra $L__BB18_94; bra.uni $L__BB18_90; $L__BB18_94: ld.global.v4.f32 {%f1324, %f1325, %f1326, %f1327}, [%rd73]; ld.global.v4.f32 {%f1328, %f1329, %f1330, %f1331}, [%rd73+16]; ld.global.v4.f32 {%f1332, %f1333, %f1334, %f1335}, [%rd73+32]; bra.uni $L__BB18_95; $L__BB18_90: ld.const.u64 %rd166, [kernel_params+2176]; cvta.to.global.u64 %rd167, %rd166; ld.global.u32 %rd168, [%rd73+168]; ld.global.u32 %r124, [%rd73+152]; shl.b32 %r125, %r124, 1; cvt.rn.f32.s32 %f1003, %r125; mul.ftz.f32 %f1004, %f1003, 0f3F000000; cvt.rzi.ftz.s32.f32 %r126, %f1004; add.s32 %r127, %r125, -1; min.s32 %r128, %r126, %r127; cvt.rn.f32.s32 %f1005, %r128; sub.ftz.f32 %f262, %f1004, %f1005; cvt.s64.s32 %rd169, %r128; add.s64 %rd170, %rd169, %rd168; shl.b64 %rd171, %rd170, 6; add.s64 %rd74, %rd167, %rd171; ld.global.v4.f32 {%f1006, %f1007, %f1008, %f1009}, [%rd74]; ld.global.v4.f32 {%f1010, %f1011, %f1012, %f1013}, [%rd74+64]; mul.ftz.f32 %f1014, %f1007, %f1011; fma.rn.ftz.f32 %f1015, %f1006, %f1010, %f1014; mul.ftz.f32 %f1016, %f1009, %f1013; fma.rn.ftz.f32 %f1017, %f1008, %f1012, %f1016; add.ftz.f32 %f271, %f1015, %f1017; setp.gt.ftz.f32 %p90, %f271, 0f3F7FDF3B; sub.ftz.f32 %f272, %f972, %f262; @%p90 bra $L__BB18_92; bra.uni $L__BB18_91; $L__BB18_92: mul.ftz.f32 %f1072, %f1006, %f272; fma.rn.ftz.f32 %f1073, %f1010, %f262, %f1072; mul.ftz.f32 %f1074, %f1007, %f272; fma.rn.ftz.f32 %f1075, %f1011, %f262, %f1074; mul.ftz.f32 %f1076, %f1008, %f272; fma.rn.ftz.f32 %f1077, %f1012, %f262, %f1076; mul.ftz.f32 %f1078, %f1009, %f272; fma.rn.ftz.f32 %f1079, %f1013, %f262, %f1078; mul.ftz.f32 %f1080, %f1075, %f1075; fma.rn.ftz.f32 %f1081, %f1073, %f1073, %f1080; mul.ftz.f32 %f1082, %f1079, %f1079; fma.rn.ftz.f32 %f1083, %f1077, %f1077, %f1082; add.ftz.f32 %f1084, %f1081, %f1083; rsqrt.approx.ftz.f32 %f1085, %f1084; mul.ftz.f32 %f1323, %f1079, %f1085; mul.ftz.f32 %f1322, %f1077, %f1085; mul.ftz.f32 %f1321, %f1075, %f1085; mul.ftz.f32 %f1320, %f1073, %f1085; bra.uni $L__BB18_93; $L__BB18_11: setp.eq.s32 %p10, %r12, 0; @%p10 bra $L__BB18_18; ld.global.v4.f32 {%f354, %f355, %f356, %f357}, [%rd6+124]; ld.global.v4.f32 {%f361, %f362, %f363, %f364}, [%rd6+140]; ld.global.v4.f32 {%f368, %f369, %f370, %f371}, [%rd6+156]; abs.ftz.f32 %f1276, %f357; mul.ftz.f32 %f372, %f354, 0f3F000000; mul.ftz.f32 %f373, %f355, 0f3F000000; mul.ftz.f32 %f374, %f356, 0f3F000000; mul.ftz.f32 %f375, %f361, 0f3F000000; mul.ftz.f32 %f376, %f362, 0f3F000000; mul.ftz.f32 %f377, %f363, 0f3F000000; mul.ftz.f32 %f378, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f379, %f378; sqrt.approx.ftz.f32 %f380, %f342; mul.ftz.f32 %f381, %f380, %f379; sin.approx.ftz.f32 %f382, %f378; mul.ftz.f32 %f383, %f380, %f382; mul.ftz.f32 %f384, %f375, %f383; mul.ftz.f32 %f385, %f376, %f383; mul.ftz.f32 %f386, %f377, %f383; fma.rn.ftz.f32 %f387, %f372, %f381, %f384; fma.rn.ftz.f32 %f388, %f373, %f381, %f385; fma.rn.ftz.f32 %f389, %f374, %f381, %f386; add.ftz.f32 %f390, %f1261, %f387; add.ftz.f32 %f391, %f1260, %f388; add.ftz.f32 %f392, %f1259, %f389; st.local.f32 [%rd1], %f390; st.local.f32 [%rd1+4], %f391; st.local.f32 [%rd1+8], %f392; st.local.f32 [%rd1+48], %f1276; mul.ftz.f32 %f393, %f355, %f388; fma.rn.ftz.f32 %f394, %f354, %f387, %f393; fma.rn.ftz.f32 %f395, %f356, %f389, %f394; mul.ftz.f32 %f396, %f355, %f355; fma.rn.ftz.f32 %f397, %f354, %f354, %f396; fma.rn.ftz.f32 %f398, %f356, %f356, %f397; rcp.approx.ftz.f32 %f399, %f398; mul.ftz.f32 %f400, %f399, %f395; mul.ftz.f32 %f401, %f362, %f388; fma.rn.ftz.f32 %f402, %f361, %f387, %f401; fma.rn.ftz.f32 %f403, %f363, %f389, %f402; mul.ftz.f32 %f404, %f362, %f362; fma.rn.ftz.f32 %f405, %f361, %f361, %f404; fma.rn.ftz.f32 %f406, %f363, %f363, %f405; rcp.approx.ftz.f32 %f407, %f406; mul.ftz.f32 %f408, %f407, %f403; add.ftz.f32 %f409, %f408, 0f3F000000; st.local.f32 [%rd1+40], %f409; neg.ftz.f32 %f410, %f400; sub.ftz.f32 %f411, %f410, %f408; st.local.f32 [%rd1+44], %f411; st.local.f32 [%rd1+12], %f368; st.local.f32 [%rd1+16], %f369; st.local.f32 [%rd1+20], %f370; sub.ftz.f32 %f412, %f390, %f344; sub.ftz.f32 %f413, %f391, %f345; sub.ftz.f32 %f414, %f392, %f346; mul.ftz.f32 %f415, %f413, %f413; fma.rn.ftz.f32 %f416, %f412, %f412, %f415; fma.rn.ftz.f32 %f417, %f414, %f414, %f416; sqrt.approx.ftz.f32 %f14, %f417; st.local.f32 [%rd1+36], %f14; rsqrt.approx.ftz.f32 %f418, %f417; mul.ftz.f32 %f15, %f412, %f418; mul.ftz.f32 %f16, %f413, %f418; mul.ftz.f32 %f17, %f414, %f418; st.local.f32 [%rd1+24], %f15; st.local.f32 [%rd1+28], %f16; st.local.f32 [%rd1+32], %f17; mul.ftz.f32 %f18, %f1276, 0f3E800000; st.local.f32 [%rd1+52], %f18; setp.leu.ftz.f32 %p11, %f364, 0f00000000; @%p11 bra $L__BB18_14; mul.ftz.f32 %f419, %f16, %f369; fma.rn.ftz.f32 %f420, %f15, %f368, %f419; fma.rn.ftz.f32 %f421, %f17, %f370, %f420; neg.ftz.f32 %f422, %f421; mul.ftz.f32 %f423, %f421, %f421; mov.f32 %f424, 0f3F800000; sub.ftz.f32 %f425, %f424, %f423; mov.f32 %f426, 0f00000000; max.ftz.f32 %f427, %f425, %f426; sqrt.approx.ftz.f32 %f428, %f427; div.approx.ftz.f32 %f429, %f428, %f422; mul.ftz.f32 %f430, %f364, %f429; sub.ftz.f32 %f431, %f424, %f430; ld.global.f32 %f432, [%rd6+168]; mul.ftz.f32 %f433, %f432, %f431; max.ftz.f32 %f434, %f433, %f426; mul.ftz.f32 %f435, %f434, %f18; st.local.f32 [%rd1+52], %f435; $L__BB18_14: setp.geu.ftz.f32 %p12, %f357, 0f00000000; @%p12 bra $L__BB18_49; mul.ftz.f32 %f437, %f368, %f15; mul.ftz.f32 %f438, %f369, %f16; neg.ftz.f32 %f439, %f438; sub.ftz.f32 %f440, %f439, %f437; mul.ftz.f32 %f441, %f370, %f17; sub.ftz.f32 %f19, %f440, %f441; setp.le.ftz.f32 %p13, %f19, 0f00000000; mov.f32 %f1258, 0f00000000; @%p13 bra $L__BB18_17; mul.ftz.f32 %f442, %f14, %f14; div.approx.ftz.f32 %f1258, %f442, %f19; $L__BB18_17: mul.ftz.f32 %f1276, %f1258, %f1276; st.local.f32 [%rd1+48], %f1276; bra.uni $L__BB18_49; $L__BB18_91: mov.f32 %f1019, 0fBF800000; max.ftz.f32 %f1020, %f271, %f1019; min.ftz.f32 %f1022, %f1020, %f972; abs.ftz.f32 %f1023, %f1022; sub.ftz.f32 %f1024, %f972, %f1023; mul.ftz.f32 %f1025, %f1024, 0f3F000000; sqrt.approx.ftz.f32 %f1026, %f1025; setp.gt.ftz.f32 %p91, %f1023, 0f3F11EB85; selp.f32 %f1027, %f1026, %f1023, %p91; mul.ftz.f32 %f1028, %f1027, %f1027; mov.f32 %f1029, 0f3C94D2E9; mov.f32 %f1030, 0f3D53F941; fma.rn.ftz.f32 %f1031, %f1030, %f1028, %f1029; mov.f32 %f1032, 0f3D3F841F; fma.rn.ftz.f32 %f1033, %f1031, %f1028, %f1032; mov.f32 %f1034, 0f3D994929; fma.rn.ftz.f32 %f1035, %f1033, %f1028, %f1034; mov.f32 %f1036, 0f3E2AAB94; fma.rn.ftz.f32 %f1037, %f1035, %f1028, %f1036; mul.ftz.f32 %f1038, %f1028, %f1037; fma.rn.ftz.f32 %f1039, %f1038, %f1027, %f1027; add.ftz.f32 %f1040, %f1039, %f1039; mov.f32 %f1041, 0f3FC90FDB; sub.ftz.f32 %f1042, %f1041, %f1039; selp.f32 %f1043, %f1040, %f1042, %p91; setp.lt.ftz.f32 %p92, %f1022, 0f00000000; mov.f32 %f1044, 0f40490FDB; sub.ftz.f32 %f1045, %f1044, %f1043; selp.f32 %f1046, %f1045, %f1043, %p92; mul.ftz.f32 %f1047, %f1006, %f271; sub.ftz.f32 %f1048, %f1010, %f1047; mul.ftz.f32 %f1049, %f1007, %f271; sub.ftz.f32 %f1050, %f1011, %f1049; mul.ftz.f32 %f1051, %f1008, %f271; sub.ftz.f32 %f1052, %f1012, %f1051; mul.ftz.f32 %f1053, %f1009, %f271; sub.ftz.f32 %f1054, %f1013, %f1053; mul.ftz.f32 %f1055, %f1050, %f1050; fma.rn.ftz.f32 %f1056, %f1048, %f1048, %f1055; mul.ftz.f32 %f1057, %f1054, %f1054; fma.rn.ftz.f32 %f1058, %f1052, %f1052, %f1057; add.ftz.f32 %f1059, %f1056, %f1058; rsqrt.approx.ftz.f32 %f1060, %f1059; mul.ftz.f32 %f1061, %f1048, %f1060; mul.ftz.f32 %f1062, %f1050, %f1060; mul.ftz.f32 %f1063, %f1052, %f1060; mul.ftz.f32 %f1064, %f1054, %f1060; mul.ftz.f32 %f1065, %f262, %f1046; cos.approx.ftz.f32 %f1066, %f1065; sin.approx.ftz.f32 %f1067, %f1065; mul.ftz.f32 %f1068, %f1061, %f1067; mul.ftz.f32 %f1069, %f1062, %f1067; mul.ftz.f32 %f1070, %f1063, %f1067; mul.ftz.f32 %f1071, %f1064, %f1067; fma.rn.ftz.f32 %f1323, %f1009, %f1066, %f1071; fma.rn.ftz.f32 %f1322, %f1008, %f1066, %f1070; fma.rn.ftz.f32 %f1321, %f1007, %f1066, %f1069; fma.rn.ftz.f32 %f1320, %f1006, %f1066, %f1068; $L__BB18_93: ld.global.v4.f32 {%f1086, %f1087, %f1088, %f1089}, [%rd74+16]; ld.global.v4.f32 {%f1094, %f1095, %f1096, %f1097}, [%rd74+80]; mul.ftz.f32 %f1102, %f262, %f1094; mul.ftz.f32 %f1103, %f262, %f1095; mul.ftz.f32 %f1104, %f262, %f1096; mul.ftz.f32 %f1105, %f262, %f1097; fma.rn.ftz.f32 %f1106, %f272, %f1089, %f1105; ld.global.v4.f32 {%f1107, %f1108, %f1109, %f1110}, [%rd74+32]; ld.global.v4.f32 {%f1115, %f1116, %f1117, %f1118}, [%rd74+96]; mul.ftz.f32 %f1123, %f262, %f1115; mul.ftz.f32 %f1124, %f262, %f1116; mul.ftz.f32 %f1125, %f262, %f1117; mul.ftz.f32 %f1126, %f262, %f1118; fma.rn.ftz.f32 %f1127, %f272, %f1107, %f1123; fma.rn.ftz.f32 %f1128, %f272, %f1108, %f1124; fma.rn.ftz.f32 %f1129, %f272, %f1109, %f1125; fma.rn.ftz.f32 %f1130, %f272, %f1110, %f1126; ld.global.v4.f32 {%f1131, %f1132, %f1133, %f1134}, [%rd74+48]; ld.global.v4.f32 {%f1139, %f1140, %f1141, %f1142}, [%rd74+112]; mul.ftz.f32 %f1147, %f262, %f1139; mul.ftz.f32 %f1148, %f262, %f1140; mul.ftz.f32 %f1149, %f262, %f1141; mul.ftz.f32 %f1150, %f262, %f1142; fma.rn.ftz.f32 %f1151, %f272, %f1131, %f1147; fma.rn.ftz.f32 %f1152, %f272, %f1132, %f1148; fma.rn.ftz.f32 %f1153, %f272, %f1133, %f1149; fma.rn.ftz.f32 %f1154, %f272, %f1134, %f1150; mul.ftz.f32 %f1155, %f1323, 0f3FB504F3; mul.ftz.f32 %f1156, %f1320, 0f3FB504F3; mul.ftz.f32 %f1157, %f1321, 0f3FB504F3; mul.ftz.f32 %f1158, %f1322, 0f3FB504F3; mul.ftz.f32 %f1159, %f1155, %f1156; mul.ftz.f32 %f1160, %f1155, %f1157; mul.ftz.f32 %f1161, %f1155, %f1158; mul.ftz.f32 %f1162, %f1156, %f1156; mul.ftz.f32 %f1163, %f1156, %f1157; mul.ftz.f32 %f1164, %f1156, %f1158; mul.ftz.f32 %f1165, %f1157, %f1157; mul.ftz.f32 %f1166, %f1157, %f1158; mul.ftz.f32 %f1167, %f1158, %f1158; sub.ftz.f32 %f1169, %f972, %f1165; sub.ftz.f32 %f1170, %f1169, %f1167; sub.ftz.f32 %f1171, %f1163, %f1161; add.ftz.f32 %f1172, %f1160, %f1164; add.ftz.f32 %f1173, %f1161, %f1163; sub.ftz.f32 %f1174, %f972, %f1162; sub.ftz.f32 %f1175, %f1174, %f1167; sub.ftz.f32 %f1176, %f1166, %f1159; sub.ftz.f32 %f1177, %f1164, %f1160; add.ftz.f32 %f1178, %f1159, %f1166; sub.ftz.f32 %f1179, %f1174, %f1165; mul.ftz.f32 %f1180, %f1171, %f1129; fma.rn.ftz.f32 %f1181, %f1170, %f1106, %f1180; mul.ftz.f32 %f1182, %f1171, %f1130; fma.rn.ftz.f32 %f1183, %f1170, %f1127, %f1182; mul.ftz.f32 %f1184, %f1171, %f1151; fma.rn.ftz.f32 %f1185, %f1170, %f1128, %f1184; mul.ftz.f32 %f1186, %f1175, %f1129; fma.rn.ftz.f32 %f1187, %f1173, %f1106, %f1186; mul.ftz.f32 %f1188, %f1175, %f1130; fma.rn.ftz.f32 %f1189, %f1173, %f1127, %f1188; mul.ftz.f32 %f1190, %f1175, %f1151; fma.rn.ftz.f32 %f1191, %f1173, %f1128, %f1190; mul.ftz.f32 %f1192, %f1178, %f1129; fma.rn.ftz.f32 %f1193, %f1177, %f1106, %f1192; mul.ftz.f32 %f1194, %f1178, %f1130; fma.rn.ftz.f32 %f1195, %f1177, %f1127, %f1194; mul.ftz.f32 %f1196, %f1178, %f1151; fma.rn.ftz.f32 %f1197, %f1177, %f1128, %f1196; fma.rn.ftz.f32 %f1330, %f1176, %f1154, %f1191; fma.rn.ftz.f32 %f1329, %f1176, %f1153, %f1189; fma.rn.ftz.f32 %f1328, %f1176, %f1152, %f1187; fma.rn.ftz.f32 %f1334, %f1179, %f1154, %f1197; fma.rn.ftz.f32 %f1333, %f1179, %f1153, %f1195; fma.rn.ftz.f32 %f1332, %f1179, %f1152, %f1193; fma.rn.ftz.f32 %f1326, %f1172, %f1154, %f1185; fma.rn.ftz.f32 %f1325, %f1172, %f1153, %f1183; fma.rn.ftz.f32 %f1324, %f1172, %f1152, %f1181; fma.rn.ftz.f32 %f1335, %f272, %f1088, %f1104; fma.rn.ftz.f32 %f1331, %f272, %f1087, %f1103; fma.rn.ftz.f32 %f1327, %f272, %f1086, %f1102; $L__BB18_95: mul.ftz.f32 %f1210, %f1342, %f1325; fma.rn.ftz.f32 %f1211, %f1343, %f1324, %f1210; fma.rn.ftz.f32 %f1212, %f1341, %f1326, %f1211; add.ftz.f32 %f321, %f1327, %f1212; mul.ftz.f32 %f1213, %f1342, %f1329; fma.rn.ftz.f32 %f1214, %f1343, %f1328, %f1213; fma.rn.ftz.f32 %f1215, %f1341, %f1330, %f1214; add.ftz.f32 %f322, %f1331, %f1215; mul.ftz.f32 %f1216, %f1342, %f1333; fma.rn.ftz.f32 %f1217, %f1343, %f1332, %f1216; fma.rn.ftz.f32 %f1218, %f1341, %f1334, %f1217; add.ftz.f32 %f1341, %f1335, %f1218; mul.ftz.f32 %f1219, %f1339, %f1325; fma.rn.ftz.f32 %f1220, %f1340, %f1324, %f1219; fma.rn.ftz.f32 %f1221, %f1338, %f1326, %f1220; add.ftz.f32 %f324, %f1327, %f1221; mul.ftz.f32 %f1222, %f1339, %f1329; fma.rn.ftz.f32 %f1223, %f1340, %f1328, %f1222; fma.rn.ftz.f32 %f1224, %f1338, %f1330, %f1223; add.ftz.f32 %f325, %f1331, %f1224; mul.ftz.f32 %f1225, %f1339, %f1333; fma.rn.ftz.f32 %f1226, %f1340, %f1332, %f1225; fma.rn.ftz.f32 %f1227, %f1338, %f1334, %f1226; add.ftz.f32 %f1338, %f1335, %f1227; mul.ftz.f32 %f1228, %f1336, %f1325; fma.rn.ftz.f32 %f1229, %f1337, %f1324, %f1228; fma.rn.ftz.f32 %f1230, %f1344, %f1326, %f1229; add.ftz.f32 %f327, %f1327, %f1230; mul.ftz.f32 %f1231, %f1336, %f1329; fma.rn.ftz.f32 %f1232, %f1337, %f1328, %f1231; fma.rn.ftz.f32 %f1233, %f1344, %f1330, %f1232; add.ftz.f32 %f328, %f1331, %f1233; mul.ftz.f32 %f1234, %f1336, %f1333; fma.rn.ftz.f32 %f1235, %f1337, %f1332, %f1234; fma.rn.ftz.f32 %f1236, %f1344, %f1334, %f1235; add.ftz.f32 %f1344, %f1335, %f1236; mov.f32 %f1336, %f328; mov.f32 %f1337, %f327; mov.f32 %f1339, %f325; mov.f32 %f1340, %f324; mov.f32 %f1342, %f322; mov.f32 %f1343, %f321; $L__BB18_96: sub.ftz.f32 %f1237, %f1341, %f1338; sub.ftz.f32 %f1238, %f1336, %f1339; mul.ftz.f32 %f1239, %f1238, %f1237; sub.ftz.f32 %f1240, %f1342, %f1339; sub.ftz.f32 %f1241, %f1344, %f1338; mul.ftz.f32 %f1242, %f1240, %f1241; sub.ftz.f32 %f1243, %f1239, %f1242; sub.ftz.f32 %f1244, %f1343, %f1340; mul.ftz.f32 %f1245, %f1244, %f1241; sub.ftz.f32 %f1246, %f1337, %f1340; mul.ftz.f32 %f1247, %f1246, %f1237; sub.ftz.f32 %f1248, %f1245, %f1247; mul.ftz.f32 %f1249, %f1246, %f1240; mul.ftz.f32 %f1250, %f1238, %f1244; sub.ftz.f32 %f1251, %f1249, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1243, %f1243, %f1252; fma.rn.ftz.f32 %f1254, %f1251, %f1251, %f1253; sqrt.approx.ftz.f32 %f1255, %f1254; mul.ftz.f32 %f1256, %f1255, 0f3F000000; mul.ftz.f32 %f1257, %f1345, %f1256; div.approx.ftz.f32 %f1345, %f1257, %f236; st.local.f32 [%rd1+48], %f1345; $L__BB18_97: st.local.f32 [%rd1+40], %f1317; st.local.f32 [%rd1+44], %f1318; or.b32 %r129, %r41, %r20; st.local.u32 [%rd1+64], %r129; setp.gt.ftz.f32 %p93, %f1345, 0f00000000; selp.u16 %rs23, 1, 0, %p93; $L__BB18_98: cvt.u32.u16 %r130, %rs23; st.param.b32 [func_retval0+0], %r130; ret; $L__BB18_18: ld.global.f32 %f23, [%rd6+124]; sub.ftz.f32 %f443, %f344, %f1261; sub.ftz.f32 %f444, %f345, %f1260; mul.ftz.f32 %f445, %f444, %f444; fma.rn.ftz.f32 %f446, %f443, %f443, %f445; sub.ftz.f32 %f447, %f346, %f1259; fma.rn.ftz.f32 %f448, %f447, %f447, %f446; sqrt.approx.ftz.f32 %f449, %f448; div.approx.ftz.f32 %f24, %f443, %f449; div.approx.ftz.f32 %f25, %f444, %f449; div.approx.ftz.f32 %f26, %f447, %f449; st.local.f32 [%rd1], %f1261; st.local.f32 [%rd1+4], %f1260; st.local.f32 [%rd1+8], %f1259; setp.leu.ftz.f32 %p14, %f23, 0f00000000; @%p14 bra $L__BB18_20; setp.neu.ftz.f32 %p15, %f24, %f25; setp.neu.ftz.f32 %p16, %f24, %f26; or.pred %p17, %p15, %p16; neg.ftz.f32 %f450, %f26; selp.f32 %f451, %f450, %f26, %p17; add.ftz.f32 %f452, %f24, %f451; neg.ftz.f32 %f453, %f25; selp.f32 %f454, %f25, %f453, %p17; sub.ftz.f32 %f455, %f454, %f24; sub.ftz.f32 %f456, %f26, %f25; mul.ftz.f32 %f457, %f452, %f452; fma.rn.ftz.f32 %f458, %f456, %f456, %f457; fma.rn.ftz.f32 %f459, %f455, %f455, %f458; sqrt.approx.ftz.f32 %f460, %f459; div.approx.ftz.f32 %f461, %f456, %f460; div.approx.ftz.f32 %f462, %f452, %f460; div.approx.ftz.f32 %f463, %f455, %f460; mul.ftz.f32 %f464, %f25, %f463; mul.ftz.f32 %f465, %f26, %f462; sub.ftz.f32 %f466, %f464, %f465; mul.ftz.f32 %f467, %f26, %f461; mul.ftz.f32 %f468, %f24, %f463; sub.ftz.f32 %f469, %f467, %f468; mul.ftz.f32 %f470, %f24, %f462; mul.ftz.f32 %f471, %f25, %f461; sub.ftz.f32 %f472, %f470, %f471; mul.ftz.f32 %f473, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f474, %f473; sqrt.approx.ftz.f32 %f475, %f342; mul.ftz.f32 %f476, %f475, %f474; sin.approx.ftz.f32 %f477, %f473; mul.ftz.f32 %f478, %f475, %f477; mul.ftz.f32 %f479, %f466, %f478; mul.ftz.f32 %f480, %f469, %f478; mul.ftz.f32 %f481, %f472, %f478; fma.rn.ftz.f32 %f482, %f461, %f476, %f479; fma.rn.ftz.f32 %f483, %f462, %f476, %f480; fma.rn.ftz.f32 %f484, %f463, %f476, %f481; fma.rn.ftz.f32 %f1261, %f23, %f482, %f1261; fma.rn.ftz.f32 %f1260, %f23, %f483, %f1260; fma.rn.ftz.f32 %f1259, %f23, %f484, %f1259; st.local.f32 [%rd1], %f1261; st.local.f32 [%rd1+4], %f1260; st.local.f32 [%rd1+8], %f1259; $L__BB18_20: ld.global.f32 %f33, [%rd6+128]; st.local.f32 [%rd1+48], %f33; sub.ftz.f32 %f487, %f1261, %f344; sub.ftz.f32 %f488, %f1260, %f345; mul.ftz.f32 %f489, %f488, %f488; fma.rn.ftz.f32 %f490, %f487, %f487, %f489; sub.ftz.f32 %f491, %f1259, %f346; fma.rn.ftz.f32 %f492, %f491, %f491, %f490; sqrt.approx.ftz.f32 %f34, %f492; st.local.f32 [%rd1+36], %f34; rsqrt.approx.ftz.f32 %f493, %f492; mul.ftz.f32 %f35, %f487, %f493; mul.ftz.f32 %f36, %f488, %f493; mul.ftz.f32 %f37, %f491, %f493; st.local.f32 [%rd1+24], %f35; st.local.f32 [%rd1+28], %f36; st.local.f32 [%rd1+32], %f37; neg.ftz.f32 %f38, %f35; neg.ftz.f32 %f39, %f36; neg.ftz.f32 %f40, %f37; st.local.f32 [%rd1+12], %f38; st.local.f32 [%rd1+16], %f39; st.local.f32 [%rd1+20], %f40; mul.ftz.f32 %f494, %f33, 0f3DA2F983; st.local.f32 [%rd1+52], %f494; mul.ftz.f32 %f495, %f36, %f36; fma.rn.ftz.f32 %f496, %f35, %f35, %f495; fma.rn.ftz.f32 %f41, %f37, %f37, %f496; setp.leu.ftz.f32 %p18, %f41, 0f00000000; mov.f32 %f1266, 0f00000000; mov.f32 %f1264, %f1266; mov.f32 %f1265, %f1266; @%p18 bra $L__BB18_29; setp.eq.ftz.f32 %p19, %f35, 0f80000000; mov.f32 %f1264, 0f00000000; setp.eq.ftz.f32 %p20, %f36, 0f80000000; and.pred %p21, %p19, %p20; @%p21 bra $L__BB18_28; abs.ftz.f32 %f42, %f39; setp.eq.ftz.f32 %p22, %f42, 0f00000000; abs.ftz.f32 %f43, %f38; setp.eq.ftz.f32 %p23, %f43, 0f00000000; and.pred %p24, %p22, %p23; mov.b32 %r15, %f39; mov.b32 %r60, %f38; and.b32 %r16, %r60, -2147483648; @%p24 bra $L__BB18_26; bra.uni $L__BB18_23; $L__BB18_26: shr.s32 %r65, %r15, 31; and.b32 %r66, %r65, 1078530011; or.b32 %r67, %r66, %r16; mov.b32 %f1262, %r67; bra.uni $L__BB18_27; $L__BB18_40: setp.eq.ftz.f32 %p50, %f87, 0f7F800000; setp.eq.ftz.f32 %p51, %f88, 0f7F800000; and.pred %p52, %p50, %p51; @%p52 bra $L__BB18_42; bra.uni $L__BB18_41; $L__BB18_42: setp.lt.s32 %p56, %r17, 0; selp.b32 %r71, 1075235812, 1061752795, %p56; or.b32 %r72, %r71, %r18; mov.b32 %f1271, %r72; bra.uni $L__BB18_44; $L__BB18_41: setp.lt.s32 %p53, %r17, 0; min.ftz.f32 %f633, %f88, %f87; max.ftz.f32 %f634, %f88, %f87; div.full.ftz.f32 %f635, %f633, %f634; mul.rn.ftz.f32 %f636, %f635, %f635; mov.f32 %f637, 0fC0B59883; mov.f32 %f638, 0fBF52C7EA; fma.rn.ftz.f32 %f639, %f636, %f638, %f637; mov.f32 %f640, 0fC0D21907; fma.rn.ftz.f32 %f641, %f639, %f636, %f640; mul.ftz.f32 %f642, %f636, %f641; mul.ftz.f32 %f643, %f635, %f642; add.ftz.f32 %f644, %f636, 0f41355DC0; mov.f32 %f645, 0f41E6BD60; fma.rn.ftz.f32 %f646, %f644, %f636, %f645; mov.f32 %f647, 0f419D92C8; fma.rn.ftz.f32 %f648, %f646, %f636, %f647; rcp.approx.ftz.f32 %f649, %f648; fma.rn.ftz.f32 %f650, %f643, %f649, %f635; mov.f32 %f651, 0f3FC90FDB; sub.ftz.f32 %f652, %f651, %f650; setp.gt.ftz.f32 %p54, %f88, %f87; selp.f32 %f653, %f652, %f650, %p54; mov.f32 %f654, 0f40490FDB; sub.ftz.f32 %f655, %f654, %f653; selp.f32 %f656, %f655, %f653, %p53; mov.b32 %r69, %f656; or.b32 %r70, %r18, %r69; mov.b32 %f657, %r70; add.ftz.f32 %f658, %f87, %f88; setp.le.ftz.f32 %p55, %f658, 0f7F800000; selp.f32 %f1271, %f657, %f658, %p55; $L__BB18_44: fma.rn.ftz.f32 %f1273, %f1271, 0fBE22F983, 0f3F000000; $L__BB18_45: sqrt.approx.ftz.f32 %f659, %f86; div.approx.ftz.f32 %f660, %f78, %f659; mov.f32 %f661, 0fBF800000; max.ftz.f32 %f662, %f660, %f661; mov.f32 %f663, 0f3F800000; min.ftz.f32 %f664, %f662, %f663; abs.ftz.f32 %f665, %f664; sub.ftz.f32 %f666, %f663, %f665; mul.ftz.f32 %f667, %f666, 0f3F000000; sqrt.approx.ftz.f32 %f668, %f667; setp.gt.ftz.f32 %p57, %f665, 0f3F11EB85; selp.f32 %f669, %f668, %f665, %p57; mul.ftz.f32 %f670, %f669, %f669; mov.f32 %f671, 0f3C94D2E9; mov.f32 %f672, 0f3D53F941; fma.rn.ftz.f32 %f673, %f672, %f670, %f671; mov.f32 %f674, 0f3D3F841F; fma.rn.ftz.f32 %f675, %f673, %f670, %f674; mov.f32 %f676, 0f3D994929; fma.rn.ftz.f32 %f677, %f675, %f670, %f676; mov.f32 %f678, 0f3E2AAB94; fma.rn.ftz.f32 %f679, %f677, %f670, %f678; mul.ftz.f32 %f680, %f670, %f679; fma.rn.ftz.f32 %f681, %f680, %f669, %f669; add.ftz.f32 %f682, %f681, %f681; mov.f32 %f683, 0f3FC90FDB; sub.ftz.f32 %f684, %f683, %f681; selp.f32 %f685, %f682, %f684, %p57; setp.lt.ftz.f32 %p58, %f664, 0f00000000; mov.f32 %f686, 0f40490FDB; sub.ftz.f32 %f687, %f686, %f685; selp.f32 %f688, %f687, %f685, %p58; fma.rn.ftz.f32 %f1274, %f688, 0fBEA2F983, 0f3F800000; $L__BB18_46: st.local.f32 [%rd1+40], %f1273; st.local.f32 [%rd1+44], %f1274; mul.ftz.f32 %f690, %f63, %f74; neg.ftz.f32 %f691, %f690; mul.ftz.f32 %f692, %f62, %f73; sub.ftz.f32 %f693, %f691, %f692; mul.ftz.f32 %f694, %f64, %f75; sub.ftz.f32 %f98, %f693, %f694; setp.le.ftz.f32 %p59, %f98, 0f00000000; @%p59 bra $L__BB18_48; mul.ftz.f32 %f695, %f72, %f72; div.approx.ftz.f32 %f1275, %f695, %f98; $L__BB18_48: mul.ftz.f32 %f1276, %f1275, %f71; st.local.f32 [%rd1+48], %f1276; bra.uni $L__BB18_49; $L__BB18_23: setp.eq.ftz.f32 %p25, %f42, 0f7F800000; setp.eq.ftz.f32 %p26, %f43, 0f7F800000; and.pred %p27, %p25, %p26; @%p27 bra $L__BB18_25; bra.uni $L__BB18_24; $L__BB18_25: setp.lt.s32 %p31, %r15, 0; selp.b32 %r63, 1075235812, 1061752795, %p31; or.b32 %r64, %r63, %r16; mov.b32 %f1262, %r64; bra.uni $L__BB18_27; $L__BB18_24: setp.lt.s32 %p28, %r15, 0; min.ftz.f32 %f498, %f43, %f42; max.ftz.f32 %f499, %f43, %f42; div.full.ftz.f32 %f500, %f498, %f499; mul.rn.ftz.f32 %f501, %f500, %f500; mov.f32 %f502, 0fC0B59883; mov.f32 %f503, 0fBF52C7EA; fma.rn.ftz.f32 %f504, %f501, %f503, %f502; mov.f32 %f505, 0fC0D21907; fma.rn.ftz.f32 %f506, %f504, %f501, %f505; mul.ftz.f32 %f507, %f501, %f506; mul.ftz.f32 %f508, %f500, %f507; add.ftz.f32 %f509, %f501, 0f41355DC0; mov.f32 %f510, 0f41E6BD60; fma.rn.ftz.f32 %f511, %f509, %f501, %f510; mov.f32 %f512, 0f419D92C8; fma.rn.ftz.f32 %f513, %f511, %f501, %f512; rcp.approx.ftz.f32 %f514, %f513; fma.rn.ftz.f32 %f515, %f508, %f514, %f500; mov.f32 %f516, 0f3FC90FDB; sub.ftz.f32 %f517, %f516, %f515; setp.gt.ftz.f32 %p29, %f43, %f42; selp.f32 %f518, %f517, %f515, %p29; mov.f32 %f519, 0f40490FDB; sub.ftz.f32 %f520, %f519, %f518; selp.f32 %f521, %f520, %f518, %p28; mov.b32 %r61, %f521; or.b32 %r62, %r16, %r61; mov.b32 %f522, %r62; add.ftz.f32 %f523, %f42, %f43; setp.le.ftz.f32 %p30, %f523, 0f7F800000; selp.f32 %f1262, %f522, %f523, %p30; $L__BB18_27: fma.rn.ftz.f32 %f1264, %f1262, 0fBE22F983, 0f3F000000; $L__BB18_28: sqrt.approx.ftz.f32 %f524, %f41; div.approx.ftz.f32 %f525, %f40, %f524; mov.f32 %f526, 0fBF800000; max.ftz.f32 %f527, %f525, %f526; mov.f32 %f528, 0f3F800000; min.ftz.f32 %f529, %f527, %f528; abs.ftz.f32 %f530, %f529; sub.ftz.f32 %f531, %f528, %f530; mul.ftz.f32 %f532, %f531, 0f3F000000; sqrt.approx.ftz.f32 %f533, %f532; setp.gt.ftz.f32 %p32, %f530, 0f3F11EB85; selp.f32 %f534, %f533, %f530, %p32; mul.ftz.f32 %f535, %f534, %f534; mov.f32 %f536, 0f3C94D2E9; mov.f32 %f537, 0f3D53F941; fma.rn.ftz.f32 %f538, %f537, %f535, %f536; mov.f32 %f539, 0f3D3F841F; fma.rn.ftz.f32 %f540, %f538, %f535, %f539; mov.f32 %f541, 0f3D994929; fma.rn.ftz.f32 %f542, %f540, %f535, %f541; mov.f32 %f543, 0f3E2AAB94; fma.rn.ftz.f32 %f544, %f542, %f535, %f543; mul.ftz.f32 %f545, %f535, %f544; fma.rn.ftz.f32 %f546, %f545, %f534, %f534; add.ftz.f32 %f547, %f546, %f546; mov.f32 %f548, 0f3FC90FDB; sub.ftz.f32 %f549, %f548, %f546; selp.f32 %f550, %f547, %f549, %p32; setp.lt.ftz.f32 %p33, %f529, 0f00000000; mov.f32 %f551, 0f40490FDB; sub.ftz.f32 %f552, %f551, %f550; selp.f32 %f553, %f552, %f550, %p33; fma.rn.ftz.f32 %f1265, %f553, 0fBEA2F983, 0f3F800000; $L__BB18_29: st.local.f32 [%rd1+40], %f1264; st.local.f32 [%rd1+44], %f1265; mul.ftz.f32 %f555, %f25, %f36; neg.ftz.f32 %f556, %f555; mul.ftz.f32 %f557, %f24, %f35; sub.ftz.f32 %f558, %f556, %f557; mul.ftz.f32 %f559, %f26, %f37; sub.ftz.f32 %f53, %f558, %f559; setp.le.ftz.f32 %p34, %f53, 0f00000000; @%p34 bra $L__BB18_31; mul.ftz.f32 %f560, %f34, %f34; div.approx.ftz.f32 %f1266, %f560, %f53; $L__BB18_31: mul.ftz.f32 %f1276, %f1266, %f33; st.local.f32 [%rd1+48], %f1276; $L__BB18_49: ld.const.f32 %f696, [kernel_params+1504]; mul.ftz.f32 %f697, %f696, %f1276; st.local.f32 [%rd1+48], %f697; mov.u16 %rs23, 1; bra.uni $L__BB18_98; } .func (.param .b32 func_retval0) _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample( .param .b32 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_0, .param .b32 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_1, .param .b32 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_2, .param .align 4 .b8 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3[12], .param .b32 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_4, .param .b32 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_5, .param .b64 _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_6 ) { .reg .pred %p<311>; .reg .b16 %rs<16>; .reg .f32 %f<3962>; .reg .b32 %r<333>; .reg .b64 %rd<272>; ld.param.f32 %f931, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_0]; ld.param.f32 %f932, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_1]; ld.param.f32 %f933, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_2]; ld.param.u32 %r108, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_4]; ld.param.u32 %r109, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_5]; ld.param.u64 %rd96, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_6]; ld.param.f32 %f936, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+8]; ld.param.f32 %f935, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+4]; ld.param.f32 %f934, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3]; cvta.to.local.u64 %rd1, %rd96; ld.const.u32 %r1, [kernel_params+1492]; add.s32 %r294, %r1, 1; ld.const.u64 %rd97, [kernel_params+2360]; cvta.to.global.u64 %rd2, %rd97; mov.u32 %r293, 0; $L__BB19_1: shr.s32 %r111, %r294, 1; add.s32 %r112, %r111, %r293; mul.wide.s32 %rd98, %r112, 16; add.s64 %rd99, %rd2, %rd98; ld.global.f32 %f937, [%rd99]; setp.gt.ftz.f32 %p5, %f937, %f931; add.s32 %r113, %r112, 1; not.b32 %r114, %r111; add.s32 %r115, %r294, %r114; selp.b32 %r293, %r293, %r113, %p5; selp.b32 %r294, %r111, %r115, %p5; setp.gt.s32 %p6, %r294, 0; @%p6 bra $L__BB19_1; add.s32 %r116, %r293, -1; max.s32 %r117, %r116, 0; add.s32 %r118, %r1, -1; min.s32 %r119, %r117, %r118; mul.wide.s32 %rd100, %r119, 16; add.s64 %rd3, %rd2, %rd100; ld.global.f32 %f938, [%rd3]; sub.ftz.f32 %f939, %f931, %f938; ld.global.f32 %f940, [%rd3+16]; sub.ftz.f32 %f941, %f940, %f938; div.approx.ftz.f32 %f1, %f939, %f941; ld.global.u32 %r7, [%rd3+4]; cvt.s64.s32 %rd4, %r7; setp.gt.s32 %p7, %r7, -1; @%p7 bra $L__BB19_176; bra.uni $L__BB19_3; $L__BB19_176: and.b32 %r220, %r109, 1073741824; setp.eq.s32 %p235, %r220, 0; ld.global.u32 %r83, [%rd3+12]; cvt.s64.s32 %rd27, %r83; @%p235 bra $L__BB19_178; ld.const.u64 %rd137, [kernel_params+2184]; cvta.to.global.u64 %rd138, %rd137; shl.b64 %rd139, %rd27, 2; add.s64 %rd140, %rd138, %rd139; ld.global.u8 %rs7, [%rd140]; and.b16 %rs8, %rs7, 128; setp.eq.s16 %p237, %rs8, 0; mov.pred %p310, 0; @%p237 bra $L__BB19_264; $L__BB19_178: ld.global.u32 %r84, [%rd3+8]; ld.const.u64 %rd141, [kernel_params+2184]; cvta.to.global.u64 %rd142, %rd141; shl.b64 %rd143, %rd27, 2; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r85, [%rd144]; shr.u32 %r221, %r85, 6; and.b32 %r222, %r221, 1; setp.eq.b32 %p238, %r222, 1; setp.ge.ftz.f32 %p239, %f933, 0f00000000; and.pred %p240, %p239, %p238; @%p240 bra $L__BB19_180; bra.uni $L__BB19_179; $L__BB19_180: ld.const.u64 %rd156, [kernel_params+2160]; cvta.to.global.u64 %rd157, %rd156; mul.lo.s64 %rd158, %rd27, 208; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r86, [%rd159+156]; ld.global.u32 %r87, [%rd159+152]; shl.b32 %r226, %r87, 1; cvt.rn.f32.s32 %f2612, %r226; mul.ftz.f32 %f450, %f2612, %f933; cvt.rzi.ftz.s32.f32 %r227, %f450; add.s32 %r228, %r226, -1; min.s32 %r88, %r227, %r228; ld.const.u64 %rd160, [kernel_params+2312]; cvta.to.global.u64 %rd29, %rd160; ld.global.u32 %r328, [%rd159+164]; mul.wide.u32 %rd161, %r328, 16; add.s64 %rd162, %rd29, %rd161; ld.global.u16 %rs12, [%rd162+12]; ld.global.u32 %r327, [%rd162+8]; ld.global.u64 %rd252, [%rd162]; setp.eq.s64 %p241, %rd252, 11; @%p241 bra $L__BB19_185; $L__BB19_181: setp.eq.s64 %p242, %rd252, 0; @%p242 bra $L__BB19_183; add.s32 %r328, %r328, 2; bra.uni $L__BB19_184; $L__BB19_183: setp.eq.s16 %p243, %rs12, 0; mov.u32 %r331, -1; mov.u32 %r328, %r327; @%p243 bra $L__BB19_186; $L__BB19_184: mul.wide.u32 %rd163, %r328, 16; add.s64 %rd164, %rd29, %rd163; ld.global.u16 %rs12, [%rd164+12]; ld.global.u32 %r327, [%rd164+8]; ld.global.u64 %rd252, [%rd164]; setp.ne.s64 %p244, %rd252, 11; @%p244 bra $L__BB19_181; $L__BB19_185: setp.eq.s16 %p245, %rs12, 0; selp.b32 %r331, -1, %r327, %p245; $L__BB19_186: ld.const.u64 %rd271, [kernel_params+2232]; cvta.to.global.u64 %rd165, %rd271; shl.b64 %rd166, %rd4, 4; add.s64 %rd167, %rd165, %rd166; ld.global.v4.u32 {%r230, %r231, %r232, %r233}, [%rd167]; setp.eq.s32 %p246, %r88, %r87; @%p246 bra $L__BB19_188; bra.uni $L__BB19_187; $L__BB19_188: ld.const.u64 %rd173, [kernel_params+2256]; cvta.to.global.u64 %rd174, %rd173; mul.wide.u32 %rd175, %r233, 12; add.s64 %rd253, %rd174, %rd175; add.s32 %r240, %r233, 1; mul.wide.u32 %rd176, %r240, 12; add.s64 %rd256, %rd174, %rd176; add.s32 %r241, %r233, 2; mul.wide.u32 %rd177, %r241, 12; add.s64 %rd259, %rd174, %rd177; add.s64 %rd254, %rd253, 4; add.s64 %rd255, %rd253, 8; add.s64 %rd257, %rd256, 4; add.s64 %rd258, %rd256, 8; add.s64 %rd260, %rd259, 4; add.s64 %rd261, %rd259, 8; bra.uni $L__BB19_189; $L__BB19_3: not.b32 %r8, %r7; ld.const.u64 %rd101, [kernel_params+2368]; cvta.to.global.u64 %rd6, %rd101; mul.wide.s32 %rd102, %r8, 192; add.s64 %rd103, %rd6, %rd102; add.s64 %rd7, %rd103, 20; ld.global.f32 %f942, [%rd103+20]; cvt.rn.f32.s32 %f943, %r108; setp.lt.ftz.f32 %p9, %f942, %f943; mov.pred %p310, 0; @%p9 bra $L__BB19_264; and.b32 %r120, %r109, 1073741824; setp.eq.s32 %p10, %r120, 0; @%p10 bra $L__BB19_6; bra.uni $L__BB19_5; $L__BB19_6: ld.global.u32 %r295, [%rd7+-4]; bra.uni $L__BB19_7; $L__BB19_179: ld.const.u64 %rd271, [kernel_params+2232]; cvta.to.global.u64 %rd145, %rd271; shl.b64 %rd146, %rd4, 4; add.s64 %rd147, %rd145, %rd146; ld.const.u64 %rd148, [kernel_params+2256]; cvta.to.global.u64 %rd149, %rd148; ld.global.u32 %r223, [%rd147+12]; mul.wide.u32 %rd150, %r223, 12; add.s64 %rd151, %rd149, %rd150; ld.global.f32 %f3862, [%rd151]; ld.global.f32 %f3863, [%rd151+4]; ld.global.f32 %f3864, [%rd151+8]; add.s32 %r224, %r223, 1; mul.wide.u32 %rd152, %r224, 12; add.s64 %rd153, %rd149, %rd152; ld.global.f32 %f3865, [%rd153]; ld.global.f32 %f3866, [%rd153+4]; ld.global.f32 %f3867, [%rd153+8]; add.s32 %r225, %r223, 2; mul.wide.u32 %rd154, %r225, 12; add.s64 %rd155, %rd149, %rd154; ld.global.f32 %f3868, [%rd155]; ld.global.f32 %f3869, [%rd155+4]; ld.global.f32 %f3870, [%rd155+8]; mov.u16 %rs15, 0; bra.uni $L__BB19_193; $L__BB19_187: setp.gt.s32 %p247, %r88, %r87; selp.b32 %r234, -1, 0, %p247; add.s32 %r235, %r88, %r234; mad.lo.s32 %r236, %r235, %r86, %r331; ld.const.u64 %rd168, [kernel_params+2336]; cvta.to.global.u64 %rd169, %rd168; add.s32 %r237, %r230, %r236; mul.wide.u32 %rd170, %r237, 12; add.s64 %rd253, %rd169, %rd170; add.s32 %r238, %r231, %r236; mul.wide.u32 %rd171, %r238, 12; add.s64 %rd256, %rd169, %rd171; add.s32 %r239, %r232, %r236; mul.wide.u32 %rd172, %r239, 12; add.s64 %rd259, %rd169, %rd172; add.s64 %rd254, %rd253, 4; add.s64 %rd255, %rd253, 8; add.s64 %rd257, %rd256, 4; add.s64 %rd258, %rd256, 8; add.s64 %rd260, %rd259, 4; add.s64 %rd261, %rd259, 8; $L__BB19_189: ld.global.f32 %f451, [%rd261]; ld.global.f32 %f452, [%rd260]; ld.global.f32 %f453, [%rd259]; ld.global.f32 %f454, [%rd258]; ld.global.f32 %f455, [%rd257]; ld.global.f32 %f456, [%rd256]; ld.global.f32 %f457, [%rd255]; ld.global.f32 %f458, [%rd254]; ld.global.f32 %f459, [%rd253]; add.s32 %r103, %r88, 1; setp.eq.s32 %p248, %r103, %r87; @%p248 bra $L__BB19_191; bra.uni $L__BB19_190; $L__BB19_191: ld.const.u64 %rd183, [kernel_params+2256]; cvta.to.global.u64 %rd184, %rd183; mul.wide.u32 %rd185, %r233, 12; add.s64 %rd262, %rd184, %rd185; add.s32 %r248, %r233, 1; mul.wide.u32 %rd186, %r248, 12; add.s64 %rd265, %rd184, %rd186; add.s32 %r249, %r233, 2; mul.wide.u32 %rd187, %r249, 12; add.s64 %rd268, %rd184, %rd187; add.s64 %rd263, %rd262, 4; add.s64 %rd264, %rd262, 8; add.s64 %rd266, %rd265, 4; add.s64 %rd267, %rd265, 8; add.s64 %rd269, %rd268, 4; add.s64 %rd270, %rd268, 8; bra.uni $L__BB19_192; $L__BB19_190: setp.ge.s32 %p249, %r88, %r87; selp.b32 %r242, -1, 0, %p249; add.s32 %r243, %r103, %r242; mad.lo.s32 %r244, %r243, %r86, %r331; ld.const.u64 %rd178, [kernel_params+2336]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r245, %r230, %r244; mul.wide.u32 %rd180, %r245, 12; add.s64 %rd262, %rd179, %rd180; add.s32 %r246, %r231, %r244; mul.wide.u32 %rd181, %r246, 12; add.s64 %rd265, %rd179, %rd181; add.s32 %r247, %r232, %r244; mul.wide.u32 %rd182, %r247, 12; add.s64 %rd268, %rd179, %rd182; add.s64 %rd263, %rd262, 4; add.s64 %rd264, %rd262, 8; add.s64 %rd266, %rd265, 4; add.s64 %rd267, %rd265, 8; add.s64 %rd269, %rd268, 4; add.s64 %rd270, %rd268, 8; $L__BB19_192: cvt.rn.f32.s32 %f2613, %r88; sub.ftz.f32 %f2614, %f450, %f2613; mov.f32 %f2615, 0f3F800000; sub.ftz.f32 %f2616, %f2615, %f2614; ld.global.f32 %f2617, [%rd262]; mul.ftz.f32 %f2618, %f2614, %f2617; ld.global.f32 %f2619, [%rd263]; mul.ftz.f32 %f2620, %f2614, %f2619; ld.global.f32 %f2621, [%rd264]; mul.ftz.f32 %f2622, %f2614, %f2621; fma.rn.ftz.f32 %f3862, %f2616, %f459, %f2618; fma.rn.ftz.f32 %f3863, %f2616, %f458, %f2620; fma.rn.ftz.f32 %f3864, %f2616, %f457, %f2622; ld.global.f32 %f2623, [%rd265]; mul.ftz.f32 %f2624, %f2614, %f2623; ld.global.f32 %f2625, [%rd266]; mul.ftz.f32 %f2626, %f2614, %f2625; ld.global.f32 %f2627, [%rd267]; mul.ftz.f32 %f2628, %f2614, %f2627; fma.rn.ftz.f32 %f3865, %f2616, %f456, %f2624; fma.rn.ftz.f32 %f3866, %f2616, %f455, %f2626; fma.rn.ftz.f32 %f3867, %f2616, %f454, %f2628; ld.global.f32 %f2629, [%rd268]; mul.ftz.f32 %f2630, %f2614, %f2629; ld.global.f32 %f2631, [%rd269]; mul.ftz.f32 %f2632, %f2614, %f2631; ld.global.f32 %f2633, [%rd270]; mul.ftz.f32 %f2634, %f2614, %f2633; fma.rn.ftz.f32 %f3868, %f2616, %f453, %f2630; fma.rn.ftz.f32 %f3869, %f2616, %f452, %f2632; fma.rn.ftz.f32 %f3870, %f2616, %f451, %f2634; mov.u16 %rs15, 1; $L__BB19_193: and.b32 %r104, %r85, 4; setp.ne.s32 %p250, %r104, 0; @%p250 bra $L__BB19_201; and.b32 %r250, %r85, 2; setp.eq.s32 %p251, %r250, 0; ld.const.u64 %rd188, [kernel_params+2160]; cvta.to.global.u64 %rd189, %rd188; mul.lo.s64 %rd190, %rd27, 208; add.s64 %rd89, %rd189, %rd190; @%p251 bra $L__BB19_199; bra.uni $L__BB19_195; $L__BB19_199: ld.global.v4.f32 {%f3850, %f3851, %f3852, %f3853}, [%rd89]; ld.global.v4.f32 {%f3854, %f3855, %f3856, %f3857}, [%rd89+16]; ld.global.v4.f32 {%f3858, %f3859, %f3860, %f3861}, [%rd89+32]; bra.uni $L__BB19_200; $L__BB19_195: ld.const.u64 %rd191, [kernel_params+2176]; cvta.to.global.u64 %rd192, %rd191; ld.global.u32 %rd193, [%rd89+168]; ld.global.u32 %r251, [%rd89+152]; shl.b32 %r252, %r251, 1; cvt.rn.f32.s32 %f2635, %r252; selp.f32 %f2636, %f933, 0f3F000000, %p239; mul.ftz.f32 %f2637, %f2636, %f2635; cvt.rzi.ftz.s32.f32 %r253, %f2637; add.s32 %r254, %r252, -1; min.s32 %r255, %r253, %r254; cvt.rn.f32.s32 %f2638, %r255; sub.ftz.f32 %f478, %f2637, %f2638; cvt.s64.s32 %rd194, %r255; add.s64 %rd195, %rd194, %rd193; shl.b64 %rd196, %rd195, 6; add.s64 %rd90, %rd192, %rd196; ld.global.v4.f32 {%f2639, %f2640, %f2641, %f2642}, [%rd90]; ld.global.v4.f32 {%f2643, %f2644, %f2645, %f2646}, [%rd90+64]; mul.ftz.f32 %f2647, %f2640, %f2644; fma.rn.ftz.f32 %f2648, %f2639, %f2643, %f2647; mul.ftz.f32 %f2649, %f2642, %f2646; fma.rn.ftz.f32 %f2650, %f2641, %f2645, %f2649; add.ftz.f32 %f487, %f2648, %f2650; setp.gt.ftz.f32 %p253, %f487, 0f3F7FDF3B; mov.f32 %f2651, 0f3F800000; sub.ftz.f32 %f488, %f2651, %f478; @%p253 bra $L__BB19_197; bra.uni $L__BB19_196; $L__BB19_197: mul.ftz.f32 %f2705, %f2639, %f488; fma.rn.ftz.f32 %f2706, %f2643, %f478, %f2705; mul.ftz.f32 %f2707, %f2640, %f488; fma.rn.ftz.f32 %f2708, %f2644, %f478, %f2707; mul.ftz.f32 %f2709, %f2641, %f488; fma.rn.ftz.f32 %f2710, %f2645, %f478, %f2709; mul.ftz.f32 %f2711, %f2642, %f488; fma.rn.ftz.f32 %f2712, %f2646, %f478, %f2711; mul.ftz.f32 %f2713, %f2708, %f2708; fma.rn.ftz.f32 %f2714, %f2706, %f2706, %f2713; mul.ftz.f32 %f2715, %f2712, %f2712; fma.rn.ftz.f32 %f2716, %f2710, %f2710, %f2715; add.ftz.f32 %f2717, %f2714, %f2716; rsqrt.approx.ftz.f32 %f2718, %f2717; mul.ftz.f32 %f3849, %f2712, %f2718; mul.ftz.f32 %f3848, %f2710, %f2718; mul.ftz.f32 %f3847, %f2708, %f2718; mul.ftz.f32 %f3846, %f2706, %f2718; bra.uni $L__BB19_198; $L__BB19_5: ld.global.u32 %r295, [%rd7+-4]; and.b32 %r121, %r295, 4194304; setp.eq.s32 %p12, %r121, 0; @%p12 bra $L__BB19_7; bra.uni $L__BB19_264; $L__BB19_7: ld.global.u32 %r12, [%rd7+-20]; st.local.u32 [%rd1+76], %r12; st.local.u32 [%rd1+64], %r295; mov.u32 %r296, -1; st.local.u32 [%rd1+56], %r296; st.local.u32 [%rd1+60], %r296; st.local.u32 [%rd1+68], %r8; st.local.f32 [%rd1+40], %f1; st.local.f32 [%rd1+44], %f932; setp.eq.s32 %p13, %r7, 0; @%p13 bra $L__BB19_9; ld.global.u32 %r296, [%rd7+24]; $L__BB19_9: st.local.u32 [%rd1+72], %r296; setp.eq.s32 %p14, %r12, 1; @%p14 bra $L__BB19_172; bra.uni $L__BB19_10; $L__BB19_172: ld.global.f32 %f425, [%rd7+-16]; ld.global.f32 %f426, [%rd7+-12]; ld.global.f32 %f427, [%rd7+-8]; ld.global.f32 %f428, [%rd7+132]; ld.global.f32 %f429, [%rd7+124]; setp.leu.ftz.f32 %p231, %f429, 0f00000000; mov.f32 %f3833, %f425; mov.f32 %f3834, %f426; mov.f32 %f3835, %f427; @%p231 bra $L__BB19_174; setp.neu.ftz.f32 %p232, %f425, %f426; setp.neu.ftz.f32 %p233, %f425, %f427; or.pred %p234, %p232, %p233; neg.ftz.f32 %f2560, %f427; selp.f32 %f2561, %f2560, %f427, %p234; add.ftz.f32 %f2562, %f425, %f2561; neg.ftz.f32 %f2563, %f426; selp.f32 %f2564, %f426, %f2563, %p234; sub.ftz.f32 %f2565, %f2564, %f425; sub.ftz.f32 %f2566, %f427, %f426; mul.ftz.f32 %f2567, %f2562, %f2562; fma.rn.ftz.f32 %f2568, %f2566, %f2566, %f2567; fma.rn.ftz.f32 %f2569, %f2565, %f2565, %f2568; sqrt.approx.ftz.f32 %f2570, %f2569; div.approx.ftz.f32 %f2571, %f2566, %f2570; div.approx.ftz.f32 %f2572, %f2562, %f2570; div.approx.ftz.f32 %f2573, %f2565, %f2570; mul.ftz.f32 %f2574, %f426, %f2573; mul.ftz.f32 %f2575, %f427, %f2572; sub.ftz.f32 %f2576, %f2574, %f2575; mul.ftz.f32 %f2577, %f427, %f2571; mul.ftz.f32 %f2578, %f425, %f2573; sub.ftz.f32 %f2579, %f2577, %f2578; mul.ftz.f32 %f2580, %f425, %f2572; mul.ftz.f32 %f2581, %f426, %f2571; sub.ftz.f32 %f2582, %f2580, %f2581; mul.ftz.f32 %f2583, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f2584, %f2583; sqrt.approx.ftz.f32 %f2585, %f932; mul.ftz.f32 %f2586, %f2585, %f2584; sin.approx.ftz.f32 %f2587, %f2583; mul.ftz.f32 %f2588, %f2585, %f2587; mul.ftz.f32 %f2589, %f2576, %f2588; mul.ftz.f32 %f2590, %f2579, %f2588; mul.ftz.f32 %f2591, %f2582, %f2588; fma.rn.ftz.f32 %f2592, %f2571, %f2586, %f2589; fma.rn.ftz.f32 %f2593, %f2572, %f2586, %f2590; fma.rn.ftz.f32 %f2594, %f2573, %f2586, %f2591; fma.rn.ftz.f32 %f2595, %f429, %f2592, %f425; fma.rn.ftz.f32 %f2596, %f429, %f2593, %f426; fma.rn.ftz.f32 %f2597, %f429, %f2594, %f427; mul.ftz.f32 %f2598, %f2596, %f2596; fma.rn.ftz.f32 %f2599, %f2595, %f2595, %f2598; fma.rn.ftz.f32 %f2600, %f2597, %f2597, %f2599; sqrt.approx.ftz.f32 %f2601, %f2600; div.approx.ftz.f32 %f3833, %f2595, %f2601; div.approx.ftz.f32 %f3834, %f2596, %f2601; div.approx.ftz.f32 %f3835, %f2597, %f2601; $L__BB19_174: st.local.f32 [%rd1], %f3833; st.local.f32 [%rd1+4], %f3834; st.local.f32 [%rd1+8], %f3835; st.local.f32 [%rd1+12], %f3833; st.local.f32 [%rd1+16], %f3834; st.local.f32 [%rd1+20], %f3835; neg.ftz.f32 %f2602, %f3833; st.local.f32 [%rd1+24], %f2602; neg.ftz.f32 %f2603, %f3834; st.local.f32 [%rd1+28], %f2603; neg.ftz.f32 %f2604, %f3835; st.local.f32 [%rd1+32], %f2604; mov.u32 %r219, 2139095039; st.local.u32 [%rd1+36], %r219; mul.ftz.f32 %f2605, %f426, %f3834; fma.rn.ftz.f32 %f2606, %f425, %f3833, %f2605; fma.rn.ftz.f32 %f2607, %f427, %f3835, %f2606; mul.ftz.f32 %f2608, %f2607, %f2607; mul.ftz.f32 %f2609, %f2607, %f2608; div.approx.ftz.f32 %f3836, %f428, %f2609; st.local.f32 [%rd1+48], %f3836; st.local.f32 [%rd1+52], %f3836; bra.uni $L__BB19_175; $L__BB19_10: setp.eq.s32 %p15, %r12, 2; @%p15 bra $L__BB19_66; bra.uni $L__BB19_11; $L__BB19_66: ld.const.f32 %f3787, [kernel_params+1044]; ld.const.v2.f32 {%f1539, %f1540}, [kernel_params+1056]; mov.u32 %r305, 0; setp.leu.ftz.f32 %p93, %f3787, 0f00000000; @%p93 bra $L__BB19_75; ld.const.u32 %r19, [kernel_params+1048]; setp.lt.s32 %p94, %r19, 1; mov.u32 %r305, 0; @%p94 bra $L__BB19_74; ld.const.u32 %r20, [kernel_params+1052]; and.b32 %r304, %r19, 3; add.s32 %r144, %r19, -1; setp.lt.u32 %p95, %r144, 3; mov.u32 %r301, 0; mov.u32 %r305, %r301; @%p95 bra $L__BB19_71; sub.s32 %r299, %r19, %r304; mov.u32 %r301, 0; mov.u32 %r305, %r301; $L__BB19_70: add.s32 %r147, %r20, %r301; mul.wide.s32 %rd104, %r147, 192; add.s64 %rd105, %rd6, %rd104; ld.global.v4.f32 {%f1541, %f1542, %f1543, %f1544}, [%rd105+176]; ld.global.f32 %f1548, [%rd105+4]; sub.ftz.f32 %f1549, %f934, %f1548; ld.global.f32 %f1550, [%rd105+8]; sub.ftz.f32 %f1551, %f935, %f1550; ld.global.f32 %f1552, [%rd105+12]; sub.ftz.f32 %f1553, %f936, %f1552; mul.ftz.f32 %f1554, %f1551, %f1542; fma.rn.ftz.f32 %f1555, %f1549, %f1541, %f1554; fma.rn.ftz.f32 %f1556, %f1553, %f1543, %f1555; setp.gt.ftz.f32 %p96, %f1556, 0f38D1B717; selp.u32 %r148, 1, 0, %p96; add.s32 %r149, %r305, %r148; ld.global.v4.f32 {%f1557, %f1558, %f1559, %f1560}, [%rd105+368]; ld.global.f32 %f1564, [%rd105+196]; sub.ftz.f32 %f1565, %f934, %f1564; ld.global.f32 %f1566, [%rd105+200]; sub.ftz.f32 %f1567, %f935, %f1566; ld.global.f32 %f1568, [%rd105+204]; sub.ftz.f32 %f1569, %f936, %f1568; mul.ftz.f32 %f1570, %f1567, %f1558; fma.rn.ftz.f32 %f1571, %f1565, %f1557, %f1570; fma.rn.ftz.f32 %f1572, %f1569, %f1559, %f1571; setp.gt.ftz.f32 %p97, %f1572, 0f38D1B717; selp.u32 %r150, 1, 0, %p97; add.s32 %r151, %r149, %r150; ld.global.v4.f32 {%f1573, %f1574, %f1575, %f1576}, [%rd105+560]; ld.global.f32 %f1580, [%rd105+388]; sub.ftz.f32 %f1581, %f934, %f1580; ld.global.f32 %f1582, [%rd105+392]; sub.ftz.f32 %f1583, %f935, %f1582; ld.global.f32 %f1584, [%rd105+396]; sub.ftz.f32 %f1585, %f936, %f1584; mul.ftz.f32 %f1586, %f1583, %f1574; fma.rn.ftz.f32 %f1587, %f1581, %f1573, %f1586; fma.rn.ftz.f32 %f1588, %f1585, %f1575, %f1587; setp.gt.ftz.f32 %p98, %f1588, 0f38D1B717; selp.u32 %r152, 1, 0, %p98; add.s32 %r153, %r151, %r152; ld.global.v4.f32 {%f1589, %f1590, %f1591, %f1592}, [%rd105+752]; ld.global.f32 %f1596, [%rd105+580]; sub.ftz.f32 %f1597, %f934, %f1596; ld.global.f32 %f1598, [%rd105+584]; sub.ftz.f32 %f1599, %f935, %f1598; ld.global.f32 %f1600, [%rd105+588]; sub.ftz.f32 %f1601, %f936, %f1600; mul.ftz.f32 %f1602, %f1599, %f1590; fma.rn.ftz.f32 %f1603, %f1597, %f1589, %f1602; fma.rn.ftz.f32 %f1604, %f1601, %f1591, %f1603; setp.gt.ftz.f32 %p99, %f1604, 0f38D1B717; selp.u32 %r154, 1, 0, %p99; add.s32 %r305, %r153, %r154; add.s32 %r301, %r301, 4; add.s32 %r299, %r299, -4; setp.ne.s32 %p100, %r299, 0; @%p100 bra $L__BB19_70; $L__BB19_71: setp.eq.s32 %p101, %r304, 0; @%p101 bra $L__BB19_74; add.s32 %r155, %r301, %r20; mul.wide.s32 %rd106, %r155, 192; add.s64 %rd249, %rd6, %rd106; $L__BB19_73: .pragma "nounroll"; ld.global.v4.f32 {%f1605, %f1606, %f1607, %f1608}, [%rd249+176]; ld.global.f32 %f1612, [%rd249+4]; sub.ftz.f32 %f1613, %f934, %f1612; ld.global.f32 %f1614, [%rd249+8]; sub.ftz.f32 %f1615, %f935, %f1614; ld.global.f32 %f1616, [%rd249+12]; sub.ftz.f32 %f1617, %f936, %f1616; mul.ftz.f32 %f1618, %f1615, %f1606; fma.rn.ftz.f32 %f1619, %f1613, %f1605, %f1618; fma.rn.ftz.f32 %f1620, %f1617, %f1607, %f1619; setp.gt.ftz.f32 %p102, %f1620, 0f38D1B717; selp.u32 %r156, 1, 0, %p102; add.s32 %r305, %r305, %r156; add.s64 %rd249, %rd249, 192; add.s32 %r304, %r304, -1; setp.ne.s32 %p103, %r304, 0; @%p103 bra $L__BB19_73; $L__BB19_74: setp.eq.s32 %p104, %r305, 0; selp.f32 %f3787, 0f00000000, %f3787, %p104; $L__BB19_75: add.ftz.f32 %f1621, %f1539, %f3787; add.ftz.f32 %f204, %f1540, %f1621; setp.eq.ftz.f32 %p105, %f204, 0f00000000; @%p105 bra $L__BB19_170; bra.uni $L__BB19_76; $L__BB19_170: mov.u32 %r216, 1034090883; st.local.u32 [%rd1+48], %r216; mov.f32 %f2547, 0f3F800000; add.ftz.f32 %f2548, %f1, %f1; sub.ftz.f32 %f3832, %f2547, %f2548; mul.ftz.f32 %f2549, %f3832, %f3832; sub.ftz.f32 %f2550, %f2547, %f2549; mov.f32 %f2551, 0f00000000; max.ftz.f32 %f2552, %f2551, %f2550; sqrt.approx.ftz.f32 %f2553, %f2552; mul.ftz.f32 %f2554, %f932, 0f40C90FDB; cos.approx.ftz.f32 %f2555, %f2554; sin.approx.ftz.f32 %f2556, %f2554; mul.ftz.f32 %f3831, %f2553, %f2556; mul.ftz.f32 %f3830, %f2553, %f2555; $L__BB19_171: neg.ftz.f32 %f2557, %f3830; neg.ftz.f32 %f2558, %f3831; neg.ftz.f32 %f2559, %f3832; st.local.f32 [%rd1], %f2557; st.local.f32 [%rd1+4], %f2558; st.local.f32 [%rd1+8], %f2559; st.local.f32 [%rd1+12], %f2557; st.local.f32 [%rd1+16], %f2558; st.local.f32 [%rd1+20], %f2559; st.local.f32 [%rd1+24], %f3830; st.local.f32 [%rd1+28], %f3831; st.local.f32 [%rd1+32], %f3832; mov.u32 %r217, 2139095039; st.local.u32 [%rd1+36], %r217; mov.u32 %r218, 1065353216; st.local.u32 [%rd1+52], %r218; ld.local.f32 %f3836, [%rd1+48]; $L__BB19_175: ld.const.f32 %f2610, [kernel_params+1504]; mul.ftz.f32 %f2611, %f2610, %f3836; st.local.f32 [%rd1+48], %f2611; setp.gt.ftz.f32 %p310, %f2611, 0f00000000; bra.uni $L__BB19_264; $L__BB19_196: mov.f32 %f2652, 0fBF800000; max.ftz.f32 %f2653, %f487, %f2652; min.ftz.f32 %f2655, %f2653, %f2651; abs.ftz.f32 %f2656, %f2655; sub.ftz.f32 %f2657, %f2651, %f2656; mul.ftz.f32 %f2658, %f2657, 0f3F000000; sqrt.approx.ftz.f32 %f2659, %f2658; setp.gt.ftz.f32 %p254, %f2656, 0f3F11EB85; selp.f32 %f2660, %f2659, %f2656, %p254; mul.ftz.f32 %f2661, %f2660, %f2660; mov.f32 %f2662, 0f3C94D2E9; mov.f32 %f2663, 0f3D53F941; fma.rn.ftz.f32 %f2664, %f2663, %f2661, %f2662; mov.f32 %f2665, 0f3D3F841F; fma.rn.ftz.f32 %f2666, %f2664, %f2661, %f2665; mov.f32 %f2667, 0f3D994929; fma.rn.ftz.f32 %f2668, %f2666, %f2661, %f2667; mov.f32 %f2669, 0f3E2AAB94; fma.rn.ftz.f32 %f2670, %f2668, %f2661, %f2669; mul.ftz.f32 %f2671, %f2661, %f2670; fma.rn.ftz.f32 %f2672, %f2671, %f2660, %f2660; add.ftz.f32 %f2673, %f2672, %f2672; mov.f32 %f2674, 0f3FC90FDB; sub.ftz.f32 %f2675, %f2674, %f2672; selp.f32 %f2676, %f2673, %f2675, %p254; setp.lt.ftz.f32 %p255, %f2655, 0f00000000; mov.f32 %f2677, 0f40490FDB; sub.ftz.f32 %f2678, %f2677, %f2676; selp.f32 %f2679, %f2678, %f2676, %p255; mul.ftz.f32 %f2680, %f2639, %f487; sub.ftz.f32 %f2681, %f2643, %f2680; mul.ftz.f32 %f2682, %f2640, %f487; sub.ftz.f32 %f2683, %f2644, %f2682; mul.ftz.f32 %f2684, %f2641, %f487; sub.ftz.f32 %f2685, %f2645, %f2684; mul.ftz.f32 %f2686, %f2642, %f487; sub.ftz.f32 %f2687, %f2646, %f2686; mul.ftz.f32 %f2688, %f2683, %f2683; fma.rn.ftz.f32 %f2689, %f2681, %f2681, %f2688; mul.ftz.f32 %f2690, %f2687, %f2687; fma.rn.ftz.f32 %f2691, %f2685, %f2685, %f2690; add.ftz.f32 %f2692, %f2689, %f2691; rsqrt.approx.ftz.f32 %f2693, %f2692; mul.ftz.f32 %f2694, %f2681, %f2693; mul.ftz.f32 %f2695, %f2683, %f2693; mul.ftz.f32 %f2696, %f2685, %f2693; mul.ftz.f32 %f2697, %f2687, %f2693; mul.ftz.f32 %f2698, %f478, %f2679; cos.approx.ftz.f32 %f2699, %f2698; sin.approx.ftz.f32 %f2700, %f2698; mul.ftz.f32 %f2701, %f2694, %f2700; mul.ftz.f32 %f2702, %f2695, %f2700; mul.ftz.f32 %f2703, %f2696, %f2700; mul.ftz.f32 %f2704, %f2697, %f2700; fma.rn.ftz.f32 %f3849, %f2642, %f2699, %f2704; fma.rn.ftz.f32 %f3848, %f2641, %f2699, %f2703; fma.rn.ftz.f32 %f3847, %f2640, %f2699, %f2702; fma.rn.ftz.f32 %f3846, %f2639, %f2699, %f2701; $L__BB19_198: ld.global.v4.f32 {%f2719, %f2720, %f2721, %f2722}, [%rd90+16]; ld.global.v4.f32 {%f2727, %f2728, %f2729, %f2730}, [%rd90+80]; mul.ftz.f32 %f2735, %f478, %f2727; mul.ftz.f32 %f2736, %f478, %f2728; mul.ftz.f32 %f2737, %f478, %f2729; mul.ftz.f32 %f2738, %f478, %f2730; fma.rn.ftz.f32 %f2739, %f488, %f2722, %f2738; ld.global.v4.f32 {%f2740, %f2741, %f2742, %f2743}, [%rd90+32]; ld.global.v4.f32 {%f2748, %f2749, %f2750, %f2751}, [%rd90+96]; mul.ftz.f32 %f2756, %f478, %f2748; mul.ftz.f32 %f2757, %f478, %f2749; mul.ftz.f32 %f2758, %f478, %f2750; mul.ftz.f32 %f2759, %f478, %f2751; fma.rn.ftz.f32 %f2760, %f488, %f2740, %f2756; fma.rn.ftz.f32 %f2761, %f488, %f2741, %f2757; fma.rn.ftz.f32 %f2762, %f488, %f2742, %f2758; fma.rn.ftz.f32 %f2763, %f488, %f2743, %f2759; ld.global.v4.f32 {%f2764, %f2765, %f2766, %f2767}, [%rd90+48]; ld.global.v4.f32 {%f2772, %f2773, %f2774, %f2775}, [%rd90+112]; mul.ftz.f32 %f2780, %f478, %f2772; mul.ftz.f32 %f2781, %f478, %f2773; mul.ftz.f32 %f2782, %f478, %f2774; mul.ftz.f32 %f2783, %f478, %f2775; fma.rn.ftz.f32 %f2784, %f488, %f2764, %f2780; fma.rn.ftz.f32 %f2785, %f488, %f2765, %f2781; fma.rn.ftz.f32 %f2786, %f488, %f2766, %f2782; fma.rn.ftz.f32 %f2787, %f488, %f2767, %f2783; mul.ftz.f32 %f2788, %f3849, 0f3FB504F3; mul.ftz.f32 %f2789, %f3846, 0f3FB504F3; mul.ftz.f32 %f2790, %f3847, 0f3FB504F3; mul.ftz.f32 %f2791, %f3848, 0f3FB504F3; mul.ftz.f32 %f2792, %f2788, %f2789; mul.ftz.f32 %f2793, %f2788, %f2790; mul.ftz.f32 %f2794, %f2788, %f2791; mul.ftz.f32 %f2795, %f2789, %f2789; mul.ftz.f32 %f2796, %f2789, %f2790; mul.ftz.f32 %f2797, %f2789, %f2791; mul.ftz.f32 %f2798, %f2790, %f2790; mul.ftz.f32 %f2799, %f2790, %f2791; mul.ftz.f32 %f2800, %f2791, %f2791; sub.ftz.f32 %f2802, %f2651, %f2798; sub.ftz.f32 %f2803, %f2802, %f2800; sub.ftz.f32 %f2804, %f2796, %f2794; add.ftz.f32 %f2805, %f2793, %f2797; add.ftz.f32 %f2806, %f2794, %f2796; sub.ftz.f32 %f2807, %f2651, %f2795; sub.ftz.f32 %f2808, %f2807, %f2800; sub.ftz.f32 %f2809, %f2799, %f2792; sub.ftz.f32 %f2810, %f2797, %f2793; add.ftz.f32 %f2811, %f2792, %f2799; sub.ftz.f32 %f2812, %f2807, %f2798; mul.ftz.f32 %f2813, %f2804, %f2762; fma.rn.ftz.f32 %f2814, %f2803, %f2739, %f2813; mul.ftz.f32 %f2815, %f2804, %f2763; fma.rn.ftz.f32 %f2816, %f2803, %f2760, %f2815; mul.ftz.f32 %f2817, %f2804, %f2784; fma.rn.ftz.f32 %f2818, %f2803, %f2761, %f2817; mul.ftz.f32 %f2819, %f2808, %f2762; fma.rn.ftz.f32 %f2820, %f2806, %f2739, %f2819; mul.ftz.f32 %f2821, %f2808, %f2763; fma.rn.ftz.f32 %f2822, %f2806, %f2760, %f2821; mul.ftz.f32 %f2823, %f2808, %f2784; fma.rn.ftz.f32 %f2824, %f2806, %f2761, %f2823; mul.ftz.f32 %f2825, %f2811, %f2762; fma.rn.ftz.f32 %f2826, %f2810, %f2739, %f2825; mul.ftz.f32 %f2827, %f2811, %f2763; fma.rn.ftz.f32 %f2828, %f2810, %f2760, %f2827; mul.ftz.f32 %f2829, %f2811, %f2784; fma.rn.ftz.f32 %f2830, %f2810, %f2761, %f2829; fma.rn.ftz.f32 %f3856, %f2809, %f2787, %f2824; fma.rn.ftz.f32 %f3855, %f2809, %f2786, %f2822; fma.rn.ftz.f32 %f3854, %f2809, %f2785, %f2820; fma.rn.ftz.f32 %f3860, %f2812, %f2787, %f2830; fma.rn.ftz.f32 %f3859, %f2812, %f2786, %f2828; fma.rn.ftz.f32 %f3858, %f2812, %f2785, %f2826; fma.rn.ftz.f32 %f3852, %f2805, %f2787, %f2818; fma.rn.ftz.f32 %f3851, %f2805, %f2786, %f2816; fma.rn.ftz.f32 %f3850, %f2805, %f2785, %f2814; fma.rn.ftz.f32 %f3861, %f488, %f2721, %f2737; fma.rn.ftz.f32 %f3857, %f488, %f2720, %f2736; fma.rn.ftz.f32 %f3853, %f488, %f2719, %f2735; $L__BB19_200: mul.ftz.f32 %f2843, %f3863, %f3851; fma.rn.ftz.f32 %f2844, %f3862, %f3850, %f2843; fma.rn.ftz.f32 %f2845, %f3864, %f3852, %f2844; add.ftz.f32 %f537, %f3853, %f2845; mul.ftz.f32 %f2846, %f3863, %f3855; fma.rn.ftz.f32 %f2847, %f3862, %f3854, %f2846; fma.rn.ftz.f32 %f2848, %f3864, %f3856, %f2847; add.ftz.f32 %f538, %f3857, %f2848; mul.ftz.f32 %f2849, %f3863, %f3859; fma.rn.ftz.f32 %f2850, %f3862, %f3858, %f2849; fma.rn.ftz.f32 %f2851, %f3864, %f3860, %f2850; add.ftz.f32 %f3864, %f3861, %f2851; mul.ftz.f32 %f2852, %f3866, %f3851; fma.rn.ftz.f32 %f2853, %f3865, %f3850, %f2852; fma.rn.ftz.f32 %f2854, %f3867, %f3852, %f2853; add.ftz.f32 %f540, %f3853, %f2854; mul.ftz.f32 %f2855, %f3866, %f3855; fma.rn.ftz.f32 %f2856, %f3865, %f3854, %f2855; fma.rn.ftz.f32 %f2857, %f3867, %f3856, %f2856; add.ftz.f32 %f541, %f3857, %f2857; mul.ftz.f32 %f2858, %f3866, %f3859; fma.rn.ftz.f32 %f2859, %f3865, %f3858, %f2858; fma.rn.ftz.f32 %f2860, %f3867, %f3860, %f2859; add.ftz.f32 %f3867, %f3861, %f2860; mul.ftz.f32 %f2861, %f3869, %f3851; fma.rn.ftz.f32 %f2862, %f3868, %f3850, %f2861; fma.rn.ftz.f32 %f2863, %f3870, %f3852, %f2862; add.ftz.f32 %f543, %f3853, %f2863; mul.ftz.f32 %f2864, %f3869, %f3855; fma.rn.ftz.f32 %f2865, %f3868, %f3854, %f2864; fma.rn.ftz.f32 %f2866, %f3870, %f3856, %f2865; add.ftz.f32 %f544, %f3857, %f2866; mul.ftz.f32 %f2867, %f3869, %f3859; fma.rn.ftz.f32 %f2868, %f3868, %f3858, %f2867; fma.rn.ftz.f32 %f2869, %f3870, %f3860, %f2868; add.ftz.f32 %f3870, %f3861, %f2869; mov.u16 %rs15, 1; mov.f32 %f3862, %f537; mov.f32 %f3863, %f538; mov.f32 %f3865, %f540; mov.f32 %f3866, %f541; mov.f32 %f3868, %f543; mov.f32 %f3869, %f544; $L__BB19_201: sub.ftz.f32 %f2870, %f3865, %f3862; sub.ftz.f32 %f2871, %f3866, %f3863; mul.ftz.f32 %f2872, %f2871, %f2871; fma.rn.ftz.f32 %f2873, %f2870, %f2870, %f2872; sub.ftz.f32 %f2874, %f3867, %f3864; fma.rn.ftz.f32 %f2875, %f2874, %f2874, %f2873; sub.ftz.f32 %f2876, %f3868, %f3862; sub.ftz.f32 %f2877, %f3869, %f3863; mul.ftz.f32 %f2878, %f2877, %f2877; fma.rn.ftz.f32 %f2879, %f2876, %f2876, %f2878; sub.ftz.f32 %f2880, %f3870, %f3864; fma.rn.ftz.f32 %f2881, %f2880, %f2880, %f2879; sub.ftz.f32 %f2882, %f3868, %f3865; sub.ftz.f32 %f2883, %f3869, %f3866; mul.ftz.f32 %f2884, %f2883, %f2883; fma.rn.ftz.f32 %f2885, %f2882, %f2882, %f2884; sub.ftz.f32 %f2886, %f3870, %f3867; fma.rn.ftz.f32 %f2887, %f2886, %f2886, %f2885; max.ftz.f32 %f2888, %f2881, %f2887; max.ftz.f32 %f555, %f2875, %f2888; mul.ftz.f32 %f2889, %f2871, %f2880; mul.ftz.f32 %f2890, %f2874, %f2877; sub.ftz.f32 %f556, %f2889, %f2890; mul.ftz.f32 %f2891, %f2874, %f2876; mul.ftz.f32 %f2892, %f2870, %f2880; sub.ftz.f32 %f557, %f2891, %f2892; mul.ftz.f32 %f2893, %f2870, %f2877; mul.ftz.f32 %f2894, %f2871, %f2876; sub.ftz.f32 %f558, %f2893, %f2894; mul.ftz.f32 %f2895, %f557, %f557; fma.rn.ftz.f32 %f2896, %f556, %f556, %f2895; fma.rn.ftz.f32 %f559, %f558, %f558, %f2896; sqrt.approx.ftz.f32 %f560, %f559; setp.neu.ftz.f32 %p256, %f560, 0f00000000; mov.f32 %f3876, %f556; mov.f32 %f3875, %f557; mov.f32 %f3874, %f558; @%p256 bra $L__BB19_202; bra.uni $L__BB19_203; $L__BB19_202: div.approx.ftz.f32 %f3874, %f558, %f560; div.approx.ftz.f32 %f3875, %f557, %f560; div.approx.ftz.f32 %f3876, %f556, %f560; $L__BB19_203: st.local.f32 [%rd1+12], %f3876; st.local.f32 [%rd1+16], %f3875; st.local.f32 [%rd1+20], %f3874; mul.ftz.f32 %f3960, %f560, 0f3F000000; and.b32 %r256, %r85, 8; setp.eq.s32 %p257, %r256, 0; @%p257 bra $L__BB19_205; neg.ftz.f32 %f3876, %f3876; st.local.f32 [%rd1+12], %f3876; neg.ftz.f32 %f3875, %f3875; st.local.f32 [%rd1+16], %f3875; neg.ftz.f32 %f3874, %f3874; st.local.f32 [%rd1+20], %f3874; $L__BB19_205: mov.u32 %r258, 1065353216; st.local.u32 [%rd1+52], %r258; ld.const.u64 %rd197, [kernel_params+2216]; cvta.to.global.u64 %rd198, %rd197; shl.b64 %rd199, %rd4, 2; add.s64 %rd200, %rd198, %rd199; ld.global.u32 %r259, [%rd200]; st.local.u32 [%rd1+56], %r83; st.local.u32 [%rd1+60], %r7; mov.u32 %r332, -1; st.local.u32 [%rd1+68], %r332; or.b32 %r105, %r259, 268435456; st.local.u32 [%rd1+64], %r105; mov.u32 %r260, 5; st.local.u32 [%rd1+76], %r260; setp.eq.s32 %p258, %r83, -1; @%p258 bra $L__BB19_207; ld.const.u64 %rd201, [kernel_params+2160]; cvta.to.global.u64 %rd202, %rd201; mul.lo.s64 %rd203, %rd27, 208; add.s64 %rd204, %rd202, %rd203; ld.global.u32 %r332, [%rd204+192]; $L__BB19_207: st.local.u32 [%rd1+72], %r332; sub.ftz.f32 %f580, %f3862, %f934; sub.ftz.f32 %f581, %f3863, %f935; mul.ftz.f32 %f2897, %f581, %f557; fma.rn.ftz.f32 %f2898, %f580, %f556, %f2897; sub.ftz.f32 %f582, %f3864, %f936; fma.rn.ftz.f32 %f2899, %f582, %f558, %f2898; div.approx.ftz.f32 %f2900, %f2899, %f559; abs.ftz.f32 %f2901, %f2900; mul.ftz.f32 %f2902, %f2901, %f2901; setp.gt.ftz.f32 %p259, %f555, %f2902; @%p259 bra $L__BB19_224; bra.uni $L__BB19_208; $L__BB19_224: sub.ftz.f32 %f681, %f3865, %f934; sub.ftz.f32 %f682, %f3868, %f934; sub.ftz.f32 %f683, %f3869, %f935; sub.ftz.f32 %f684, %f3870, %f936; sub.ftz.f32 %f685, %f3867, %f936; mul.ftz.f32 %f3190, %f581, %f685; sub.ftz.f32 %f686, %f3866, %f935; mul.ftz.f32 %f3191, %f582, %f686; sub.ftz.f32 %f3906, %f3190, %f3191; mul.ftz.f32 %f3192, %f582, %f681; mul.ftz.f32 %f3193, %f580, %f685; sub.ftz.f32 %f3907, %f3192, %f3193; mul.ftz.f32 %f3194, %f580, %f686; mul.ftz.f32 %f3195, %f581, %f681; sub.ftz.f32 %f3908, %f3194, %f3195; mul.ftz.f32 %f3196, %f3907, %f3907; fma.rn.ftz.f32 %f3197, %f3906, %f3906, %f3196; fma.rn.ftz.f32 %f3198, %f3908, %f3908, %f3197; sqrt.approx.ftz.f32 %f690, %f3198; setp.neu.ftz.f32 %p270, %f690, 0f00000000; @%p270 bra $L__BB19_225; bra.uni $L__BB19_226; $L__BB19_225: rcp.approx.ftz.f32 %f3199, %f690; mul.ftz.f32 %f3908, %f3908, %f3199; mul.ftz.f32 %f3907, %f3907, %f3199; mul.ftz.f32 %f3906, %f3906, %f3199; $L__BB19_226: mul.ftz.f32 %f3200, %f582, %f683; mul.ftz.f32 %f3201, %f581, %f684; sub.ftz.f32 %f3909, %f3201, %f3200; mul.ftz.f32 %f3202, %f580, %f684; mul.ftz.f32 %f3203, %f582, %f682; sub.ftz.f32 %f3910, %f3203, %f3202; mul.ftz.f32 %f3204, %f581, %f682; mul.ftz.f32 %f3205, %f580, %f683; sub.ftz.f32 %f3911, %f3205, %f3204; mul.ftz.f32 %f3206, %f3910, %f3910; fma.rn.ftz.f32 %f3207, %f3909, %f3909, %f3206; fma.rn.ftz.f32 %f3208, %f3911, %f3911, %f3207; sqrt.approx.ftz.f32 %f706, %f3208; setp.neu.ftz.f32 %p271, %f706, 0f00000000; @%p271 bra $L__BB19_227; bra.uni $L__BB19_228; $L__BB19_227: rcp.approx.ftz.f32 %f3209, %f706; mul.ftz.f32 %f3911, %f3911, %f3209; mul.ftz.f32 %f3910, %f3910, %f3209; mul.ftz.f32 %f3909, %f3909, %f3209; $L__BB19_228: mul.ftz.f32 %f3210, %f685, %f683; mul.ftz.f32 %f3211, %f686, %f684; sub.ftz.f32 %f3912, %f3211, %f3210; mul.ftz.f32 %f3212, %f681, %f684; mul.ftz.f32 %f3213, %f685, %f682; sub.ftz.f32 %f3913, %f3213, %f3212; mul.ftz.f32 %f3214, %f686, %f682; mul.ftz.f32 %f3215, %f681, %f683; sub.ftz.f32 %f3914, %f3215, %f3214; mul.ftz.f32 %f3216, %f3913, %f3913; fma.rn.ftz.f32 %f3217, %f3912, %f3912, %f3216; fma.rn.ftz.f32 %f3218, %f3914, %f3914, %f3217; sqrt.approx.ftz.f32 %f722, %f3218; setp.neu.ftz.f32 %p272, %f722, 0f00000000; @%p272 bra $L__BB19_229; bra.uni $L__BB19_230; $L__BB19_229: rcp.approx.ftz.f32 %f3219, %f722; mul.ftz.f32 %f3914, %f3914, %f3219; mul.ftz.f32 %f3913, %f3913, %f3219; mul.ftz.f32 %f3912, %f3912, %f3219; $L__BB19_230: mul.ftz.f32 %f3220, %f581, %f581; fma.rn.ftz.f32 %f3221, %f580, %f580, %f3220; fma.rn.ftz.f32 %f3222, %f582, %f582, %f3221; sqrt.approx.ftz.f32 %f735, %f3222; setp.neu.ftz.f32 %p273, %f735, 0f00000000; mov.f32 %f3915, %f580; mov.f32 %f3916, %f581; mov.f32 %f3917, %f582; @%p273 bra $L__BB19_231; bra.uni $L__BB19_232; $L__BB19_231: rcp.approx.ftz.f32 %f3223, %f735; mul.ftz.f32 %f3917, %f582, %f3223; mul.ftz.f32 %f3916, %f581, %f3223; mul.ftz.f32 %f3915, %f580, %f3223; $L__BB19_232: mul.ftz.f32 %f3224, %f686, %f686; fma.rn.ftz.f32 %f3225, %f681, %f681, %f3224; fma.rn.ftz.f32 %f3226, %f685, %f685, %f3225; sqrt.approx.ftz.f32 %f745, %f3226; setp.neu.ftz.f32 %p274, %f745, 0f00000000; mov.f32 %f3918, %f681; mov.f32 %f3919, %f686; mov.f32 %f3920, %f685; @%p274 bra $L__BB19_233; bra.uni $L__BB19_234; $L__BB19_233: rcp.approx.ftz.f32 %f3227, %f745; mul.ftz.f32 %f3920, %f685, %f3227; mul.ftz.f32 %f3919, %f686, %f3227; mul.ftz.f32 %f3918, %f681, %f3227; $L__BB19_234: mul.ftz.f32 %f3228, %f683, %f683; fma.rn.ftz.f32 %f3229, %f682, %f682, %f3228; fma.rn.ftz.f32 %f3230, %f684, %f684, %f3229; sqrt.approx.ftz.f32 %f755, %f3230; setp.neu.ftz.f32 %p275, %f755, 0f00000000; mov.f32 %f3921, %f682; mov.f32 %f3922, %f683; mov.f32 %f3923, %f684; @%p275 bra $L__BB19_235; bra.uni $L__BB19_236; $L__BB19_235: rcp.approx.ftz.f32 %f3231, %f755; mul.ftz.f32 %f3923, %f684, %f3231; mul.ftz.f32 %f3922, %f683, %f3231; mul.ftz.f32 %f3921, %f682, %f3231; $L__BB19_236: mul.ftz.f32 %f3232, %f3907, %f3910; fma.rn.ftz.f32 %f3233, %f3906, %f3909, %f3232; fma.rn.ftz.f32 %f771, %f3908, %f3911, %f3233; mul.ftz.f32 %f3234, %f3907, %f3913; fma.rn.ftz.f32 %f3235, %f3906, %f3912, %f3234; fma.rn.ftz.f32 %f3236, %f3908, %f3914, %f3235; neg.ftz.f32 %f3237, %f3236; mul.ftz.f32 %f3238, %f3910, %f3913; fma.rn.ftz.f32 %f3239, %f3909, %f3912, %f3238; fma.rn.ftz.f32 %f3240, %f3911, %f3914, %f3239; abs.ftz.f32 %f3241, %f771; setp.lt.ftz.f32 %p276, %f3241, 0f3F800000; mov.f32 %f3242, 0f3F800000; add.ftz.f32 %f3243, %f3241, 0fBF800000; add.ftz.f32 %f3244, %f3243, 0f3F800000; selp.f32 %f3245, %f3244, 0f3F800000, %p276; sub.ftz.f32 %f3246, %f3242, %f3245; sqrt.approx.ftz.f32 %f3247, %f3246; mul.ftz.f32 %f3248, %f3245, 0f3CB1485D; mov.f32 %f3249, 0f3D9FB43B; sub.ftz.f32 %f3250, %f3249, %f3248; fma.rn.ftz.f32 %f3251, %f3245, %f3250, 0fBE5A6B93; fma.rn.ftz.f32 %f3252, %f3245, %f3251, 0f3FC90FDB; mul.ftz.f32 %f3253, %f3247, %f3252; setp.lt.ftz.f32 %p277, %f771, 0f00000000; mov.f32 %f3254, 0f40490FDB; sub.ftz.f32 %f3255, %f3254, %f3253; selp.f32 %f3256, %f3255, %f3253, %p277; abs.ftz.f32 %f3257, %f3237; setp.lt.ftz.f32 %p278, %f3257, 0f3F800000; add.ftz.f32 %f3258, %f3257, 0fBF800000; add.ftz.f32 %f3259, %f3258, 0f3F800000; selp.f32 %f3260, %f3259, 0f3F800000, %p278; sub.ftz.f32 %f3261, %f3242, %f3260; sqrt.approx.ftz.f32 %f3262, %f3261; mul.ftz.f32 %f3263, %f3260, 0f3CB1485D; sub.ftz.f32 %f3264, %f3249, %f3263; fma.rn.ftz.f32 %f3265, %f3260, %f3264, 0fBE5A6B93; fma.rn.ftz.f32 %f3266, %f3260, %f3265, 0f3FC90FDB; mul.ftz.f32 %f3267, %f3262, %f3266; setp.gt.ftz.f32 %p279, %f3236, 0f80000000; sub.ftz.f32 %f3268, %f3254, %f3267; selp.f32 %f3269, %f3268, %f3267, %p279; abs.ftz.f32 %f3270, %f3240; setp.lt.ftz.f32 %p280, %f3270, 0f3F800000; add.ftz.f32 %f3271, %f3270, 0fBF800000; add.ftz.f32 %f3272, %f3271, 0f3F800000; selp.f32 %f3273, %f3272, 0f3F800000, %p280; sub.ftz.f32 %f3274, %f3242, %f3273; sqrt.approx.ftz.f32 %f3275, %f3274; mul.ftz.f32 %f3276, %f3273, 0f3CB1485D; sub.ftz.f32 %f3277, %f3249, %f3276; fma.rn.ftz.f32 %f3278, %f3273, %f3277, 0fBE5A6B93; fma.rn.ftz.f32 %f3279, %f3273, %f3278, 0f3FC90FDB; mul.ftz.f32 %f3280, %f3275, %f3279; setp.lt.ftz.f32 %p281, %f3240, 0f00000000; sub.ftz.f32 %f3281, %f3254, %f3280; selp.f32 %f3282, %f3281, %f3280, %p281; add.ftz.f32 %f3283, %f3256, %f3269; add.ftz.f32 %f3284, %f3283, %f3282; add.ftz.f32 %f772, %f3284, 0fC0490FDB; mul.ftz.f32 %f3285, %f3916, %f3919; fma.rn.ftz.f32 %f3286, %f3915, %f3918, %f3285; fma.rn.ftz.f32 %f3287, %f3917, %f3920, %f3286; mul.ftz.f32 %f3288, %f3256, 0f3EA2F983; mov.b32 %r270, %f3288; and.b32 %r271, %r270, -2147483648; or.b32 %r272, %r271, 1056964608; mov.b32 %f3289, %r272; add.ftz.f32 %f3290, %f3288, %f3289; cvt.rzi.ftz.s32.f32 %r273, %f3290; cvt.rn.f32.s32 %f3291, %r273; mul.ftz.f32 %f3292, %f3291, 0f40490000; sub.ftz.f32 %f3293, %f3256, %f3292; mul.ftz.f32 %f3294, %f3291, 0f3A7DA000; sub.ftz.f32 %f3295, %f3293, %f3294; mul.ftz.f32 %f3296, %f3291, 0f34222000; sub.ftz.f32 %f3297, %f3295, %f3296; mul.ftz.f32 %f3298, %f3291, 0f2CB4611A; sub.ftz.f32 %f3299, %f3297, %f3298; add.ftz.f32 %f3300, %f3299, 0fBFC90FDB; add.ftz.f32 %f3301, %f3300, 0f3FC90FDB; mul.ftz.f32 %f3302, %f3301, %f3301; and.b32 %r274, %r273, 1; setp.eq.b32 %p282, %r274, 1; neg.ftz.f32 %f3303, %f3301; selp.f32 %f3304, %f3303, %f3301, %p282; fma.rn.ftz.f32 %f3305, %f3302, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f3306, %f3302, %f3305, 0f3C088777; fma.rn.ftz.f32 %f3307, %f3302, %f3306, 0fBE2AAAA6; mul.ftz.f32 %f3308, %f3304, %f3307; fma.rn.ftz.f32 %f3309, %f3302, %f3308, %f3304; abs.ftz.f32 %f3310, %f3309; setp.gt.ftz.f32 %p283, %f3310, 0f3F800000; selp.f32 %f773, 0f00000000, %f3309, %p283; mul.ftz.f32 %f3311, %f1, %f772; sub.ftz.f32 %f3312, %f3311, %f3256; mul.ftz.f32 %f3313, %f3312, 0f3EA2F983; mov.b32 %r275, %f3313; and.b32 %r276, %r275, -2147483648; or.b32 %r277, %r276, 1056964608; mov.b32 %f3314, %r277; add.ftz.f32 %f3315, %f3313, %f3314; cvt.rzi.ftz.s32.f32 %r278, %f3315; cvt.rn.f32.s32 %f3316, %r278; mul.ftz.f32 %f3317, %f3316, 0f40490000; sub.ftz.f32 %f3318, %f3312, %f3317; mul.ftz.f32 %f3319, %f3316, 0f3A7DA000; sub.ftz.f32 %f3320, %f3318, %f3319; mul.ftz.f32 %f3321, %f3316, 0f34222000; sub.ftz.f32 %f3322, %f3320, %f3321; mul.ftz.f32 %f3323, %f3316, 0f2CB4611A; sub.ftz.f32 %f3324, %f3322, %f3323; add.ftz.f32 %f3325, %f3324, 0fBFC90FDB; add.ftz.f32 %f3326, %f3325, 0f3FC90FDB; mul.ftz.f32 %f3327, %f3326, %f3326; and.b32 %r279, %r278, 1; setp.eq.b32 %p284, %r279, 1; neg.ftz.f32 %f3328, %f3326; selp.f32 %f3329, %f3328, %f3326, %p284; fma.rn.ftz.f32 %f3330, %f3327, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f3331, %f3327, %f3330, 0f3C088777; fma.rn.ftz.f32 %f3332, %f3327, %f3331, 0fBE2AAAA6; mul.ftz.f32 %f3333, %f3329, %f3332; fma.rn.ftz.f32 %f3334, %f3327, %f3333, %f3329; fma.rn.ftz.f32 %f3335, %f3327, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f3336, %f3327, %f3335, 0fBAB60B58; fma.rn.ftz.f32 %f3337, %f3327, %f3336, 0f3D2AAAAA; fma.rn.ftz.f32 %f3338, %f3327, %f3337, 0fBF000000; fma.rn.ftz.f32 %f3339, %f3327, %f3338, 0f3F800000; neg.ftz.f32 %f3340, %f3339; selp.f32 %f3341, %f3340, %f3339, %p284; abs.ftz.f32 %f3342, %f3334; setp.gt.ftz.f32 %p285, %f3342, 0f3F800000; selp.f32 %f774, 0f00000000, %f3334, %p285; abs.ftz.f32 %f3343, %f3341; setp.gt.ftz.f32 %p286, %f3343, 0f3F800000; selp.f32 %f775, 0f00000000, %f3341, %p286; sub.ftz.f32 %f776, %f775, %f771; fma.rn.ftz.f32 %f777, %f3287, %f773, %f774; mul.ftz.f32 %f3344, %f3916, %f3922; fma.rn.ftz.f32 %f3345, %f3915, %f3921, %f3344; fma.rn.ftz.f32 %f3346, %f3917, %f3923, %f3345; mul.ftz.f32 %f3347, %f3915, %f3346; mul.ftz.f32 %f3348, %f3916, %f3346; mul.ftz.f32 %f3349, %f3917, %f3346; sub.ftz.f32 %f3924, %f3921, %f3347; sub.ftz.f32 %f3925, %f3922, %f3348; sub.ftz.f32 %f3926, %f3923, %f3349; mul.ftz.f32 %f3350, %f3925, %f3925; fma.rn.ftz.f32 %f3351, %f3924, %f3924, %f3350; fma.rn.ftz.f32 %f3352, %f3926, %f3926, %f3351; sqrt.approx.ftz.f32 %f781, %f3352; setp.neu.ftz.f32 %p287, %f781, 0f00000000; @%p287 bra $L__BB19_237; bra.uni $L__BB19_238; $L__BB19_237: rcp.approx.ftz.f32 %f3353, %f781; mul.ftz.f32 %f3926, %f3926, %f3353; mul.ftz.f32 %f3925, %f3925, %f3353; mul.ftz.f32 %f3924, %f3924, %f3353; $L__BB19_238: mul.ftz.f32 %f3355, %f775, %f776; fma.rn.ftz.f32 %f3356, %f774, %f777, %f3355; mul.ftz.f32 %f791, %f773, %f3356; setp.eq.ftz.f32 %p288, %f791, 0f00000000; mov.f32 %f3927, %f3242; @%p288 bra $L__BB19_240; mul.ftz.f32 %f3357, %f777, %f775; mul.ftz.f32 %f3358, %f774, %f776; sub.ftz.f32 %f3359, %f3357, %f3358; mul.ftz.f32 %f3360, %f771, %f3359; sub.ftz.f32 %f3361, %f3360, %f777; div.approx.ftz.f32 %f3927, %f3361, %f791; $L__BB19_240: mul.ftz.f32 %f3362, %f3927, %f3927; sub.ftz.f32 %f3364, %f3242, %f3362; mov.f32 %f3934, 0f00000000; max.ftz.f32 %f3366, %f3364, %f3934; sqrt.approx.ftz.f32 %f3367, %f3366; mul.ftz.f32 %f3368, %f3924, %f3367; mul.ftz.f32 %f3369, %f3925, %f3367; mul.ftz.f32 %f3370, %f3926, %f3367; fma.rn.ftz.f32 %f3928, %f3915, %f3927, %f3368; fma.rn.ftz.f32 %f3929, %f3916, %f3927, %f3369; fma.rn.ftz.f32 %f3930, %f3917, %f3927, %f3370; mul.ftz.f32 %f3371, %f3929, %f3929; fma.rn.ftz.f32 %f3372, %f3928, %f3928, %f3371; fma.rn.ftz.f32 %f3373, %f3930, %f3930, %f3372; sqrt.approx.ftz.f32 %f797, %f3373; setp.neu.ftz.f32 %p289, %f797, 0f00000000; @%p289 bra $L__BB19_241; bra.uni $L__BB19_242; $L__BB19_241: rcp.approx.ftz.f32 %f3374, %f797; mul.ftz.f32 %f3930, %f3930, %f3374; mul.ftz.f32 %f3929, %f3929, %f3374; mul.ftz.f32 %f3928, %f3928, %f3374; $L__BB19_242: mul.ftz.f32 %f3375, %f3919, %f3929; fma.rn.ftz.f32 %f3376, %f3918, %f3928, %f3375; fma.rn.ftz.f32 %f807, %f3920, %f3930, %f3376; mul.ftz.f32 %f3377, %f3918, %f807; mul.ftz.f32 %f3378, %f3919, %f807; mul.ftz.f32 %f3379, %f3920, %f807; sub.ftz.f32 %f3931, %f3928, %f3377; sub.ftz.f32 %f3932, %f3929, %f3378; sub.ftz.f32 %f3933, %f3930, %f3379; mul.ftz.f32 %f3380, %f3932, %f3932; fma.rn.ftz.f32 %f3381, %f3931, %f3931, %f3380; fma.rn.ftz.f32 %f3382, %f3933, %f3933, %f3381; sqrt.approx.ftz.f32 %f811, %f3382; setp.neu.ftz.f32 %p290, %f811, 0f00000000; @%p290 bra $L__BB19_243; bra.uni $L__BB19_244; $L__BB19_243: rcp.approx.ftz.f32 %f3383, %f811; mul.ftz.f32 %f3933, %f3383, %f3933; mul.ftz.f32 %f3932, %f3383, %f3932; mul.ftz.f32 %f3931, %f3383, %f3931; $L__BB19_244: mov.f32 %f3384, 0f3F800000; sub.ftz.f32 %f3385, %f3384, %f807; mul.ftz.f32 %f3386, %f3385, %f932; sub.ftz.f32 %f3387, %f3384, %f3386; mul.ftz.f32 %f3388, %f3387, %f3387; sub.ftz.f32 %f3389, %f3384, %f3388; max.ftz.f32 %f3391, %f3389, %f3934; sqrt.approx.ftz.f32 %f3392, %f3391; mul.ftz.f32 %f3393, %f3392, %f3931; mul.ftz.f32 %f3394, %f3392, %f3932; mul.ftz.f32 %f3395, %f3392, %f3933; fma.rn.ftz.f32 %f821, %f3918, %f3387, %f3393; fma.rn.ftz.f32 %f822, %f3919, %f3387, %f3394; fma.rn.ftz.f32 %f823, %f3920, %f3387, %f3395; st.local.f32 [%rd1+24], %f821; st.local.f32 [%rd1+28], %f822; st.local.f32 [%rd1+32], %f823; add.ftz.f32 %f3396, %f582, %f684; sub.ftz.f32 %f824, %f683, %f581; mul.ftz.f32 %f3397, %f824, %f3396; add.ftz.f32 %f3398, %f581, %f683; sub.ftz.f32 %f825, %f684, %f582; mul.ftz.f32 %f3399, %f3398, %f825; sub.ftz.f32 %f3400, %f3397, %f3399; add.ftz.f32 %f3401, %f580, %f682; mul.ftz.f32 %f3402, %f3401, %f825; sub.ftz.f32 %f826, %f682, %f580; mul.ftz.f32 %f3403, %f826, %f3396; sub.ftz.f32 %f3404, %f3402, %f3403; mul.ftz.f32 %f3405, %f826, %f3398; mul.ftz.f32 %f3406, %f3401, %f824; sub.ftz.f32 %f3407, %f3405, %f3406; mul.ftz.f32 %f3408, %f3404, %f822; fma.rn.ftz.f32 %f3409, %f3400, %f821, %f3408; fma.rn.ftz.f32 %f827, %f3407, %f823, %f3409; add.ftz.f32 %f3410, %f582, %f685; sub.ftz.f32 %f828, %f581, %f686; mul.ftz.f32 %f3411, %f828, %f3410; add.ftz.f32 %f3412, %f581, %f686; sub.ftz.f32 %f829, %f582, %f685; mul.ftz.f32 %f3413, %f3412, %f829; sub.ftz.f32 %f3414, %f3411, %f3413; add.ftz.f32 %f3415, %f580, %f681; mul.ftz.f32 %f3416, %f3415, %f829; sub.ftz.f32 %f830, %f580, %f681; mul.ftz.f32 %f3417, %f830, %f3410; sub.ftz.f32 %f3418, %f3416, %f3417; mul.ftz.f32 %f3419, %f830, %f3412; mul.ftz.f32 %f3420, %f3415, %f828; sub.ftz.f32 %f3421, %f3419, %f3420; mul.ftz.f32 %f3422, %f3418, %f822; fma.rn.ftz.f32 %f3423, %f3414, %f821, %f3422; fma.rn.ftz.f32 %f831, %f3421, %f823, %f3423; add.ftz.f32 %f3424, %f685, %f684; sub.ftz.f32 %f3425, %f686, %f683; mul.ftz.f32 %f3426, %f3425, %f3424; add.ftz.f32 %f3427, %f686, %f683; sub.ftz.f32 %f3428, %f685, %f684; mul.ftz.f32 %f3429, %f3427, %f3428; sub.ftz.f32 %f3430, %f3426, %f3429; add.ftz.f32 %f3431, %f681, %f682; mul.ftz.f32 %f3432, %f3431, %f3428; sub.ftz.f32 %f3433, %f681, %f682; mul.ftz.f32 %f3434, %f3433, %f3424; sub.ftz.f32 %f3435, %f3432, %f3434; mul.ftz.f32 %f3436, %f3433, %f3427; mul.ftz.f32 %f3437, %f3431, %f3425; sub.ftz.f32 %f3438, %f3436, %f3437; mul.ftz.f32 %f3439, %f3435, %f822; fma.rn.ftz.f32 %f3440, %f3430, %f821, %f3439; fma.rn.ftz.f32 %f3441, %f3438, %f823, %f3440; add.ftz.f32 %f3442, %f827, %f831; add.ftz.f32 %f832, %f3441, %f3442; abs.ftz.f32 %f833, %f832; mul.ftz.f32 %f3443, %f833, 0f34000000; min.ftz.f32 %f3444, %f831, %f3441; min.ftz.f32 %f3445, %f827, %f3444; max.ftz.f32 %f3446, %f831, %f3441; max.ftz.f32 %f3447, %f827, %f3446; mul.ftz.f32 %f3448, %f833, 0fB4000000; setp.ltu.ftz.f32 %p291, %f3445, %f3448; setp.gtu.ftz.f32 %p292, %f3447, %f3443; and.pred %p293, %p291, %p292; @%p293 bra $L__BB19_262; mul.ftz.f32 %f3449, %f828, %f825; mul.ftz.f32 %f3450, %f829, %f824; sub.ftz.f32 %f3451, %f3449, %f3450; mul.ftz.f32 %f3452, %f830, %f825; mul.ftz.f32 %f3453, %f829, %f826; sub.ftz.f32 %f3454, %f3453, %f3452; mul.ftz.f32 %f3455, %f828, %f826; mul.ftz.f32 %f3456, %f830, %f824; sub.ftz.f32 %f3457, %f3456, %f3455; add.ftz.f32 %f834, %f3451, %f3451; add.ftz.f32 %f835, %f3454, %f3454; add.ftz.f32 %f836, %f3457, %f3457; mul.ftz.f32 %f3458, %f835, %f822; fma.rn.ftz.f32 %f3459, %f834, %f821, %f3458; fma.rn.ftz.f32 %f837, %f836, %f823, %f3459; setp.eq.ftz.f32 %p294, %f837, 0f00000000; @%p294 bra $L__BB19_262; mul.ftz.f32 %f3460, %f581, %f835; fma.rn.ftz.f32 %f3461, %f580, %f834, %f3460; fma.rn.ftz.f32 %f3462, %f582, %f836, %f3461; div.approx.ftz.f32 %f838, %f3462, %f837; setp.ltu.ftz.f32 %p295, %f838, 0f00000000; setp.gtu.ftz.f32 %p296, %f838, 0f7F7FFFFF; or.pred %p297, %p296, %p295; @%p297 bra $L__BB19_262; bra.uni $L__BB19_247; $L__BB19_262: mov.f32 %f3961, 0f00000000; mov.u32 %r290, 0; st.local.u32 [%rd1+48], %r290; bra.uni $L__BB19_263; $L__BB19_208: setp.lt.ftz.f32 %p260, %f1, %f932; @%p260 bra $L__BB19_210; bra.uni $L__BB19_209; $L__BB19_210: mul.ftz.f32 %f3877, %f1, 0f3F000000; sub.ftz.f32 %f3878, %f932, %f3877; bra.uni $L__BB19_211; $L__BB19_209: mul.ftz.f32 %f3878, %f932, 0f3F000000; sub.ftz.f32 %f3877, %f1, %f3878; $L__BB19_211: mov.f32 %f2904, 0f3F800000; sub.ftz.f32 %f2905, %f2904, %f3877; sub.ftz.f32 %f2906, %f2905, %f3878; mul.ftz.f32 %f2907, %f3865, %f3878; fma.rn.ftz.f32 %f2908, %f3862, %f3877, %f2907; mul.ftz.f32 %f2909, %f3866, %f3878; fma.rn.ftz.f32 %f2910, %f3863, %f3877, %f2909; mul.ftz.f32 %f2911, %f3867, %f3878; fma.rn.ftz.f32 %f2912, %f3864, %f3877, %f2911; fma.rn.ftz.f32 %f2913, %f3868, %f2906, %f2908; fma.rn.ftz.f32 %f2914, %f3869, %f2906, %f2910; fma.rn.ftz.f32 %f2915, %f3870, %f2906, %f2912; st.local.f32 [%rd1], %f2913; st.local.f32 [%rd1+4], %f2914; st.local.f32 [%rd1+8], %f2915; sub.ftz.f32 %f2916, %f2913, %f934; sub.ftz.f32 %f2917, %f2914, %f935; sub.ftz.f32 %f2918, %f2915, %f936; mul.ftz.f32 %f2919, %f2917, %f2917; fma.rn.ftz.f32 %f2920, %f2916, %f2916, %f2919; fma.rn.ftz.f32 %f2921, %f2918, %f2918, %f2920; sqrt.approx.ftz.f32 %f589, %f2921; st.local.f32 [%rd1+36], %f589; rsqrt.approx.ftz.f32 %f2922, %f2921; mul.ftz.f32 %f2923, %f2916, %f2922; mul.ftz.f32 %f2924, %f2917, %f2922; mul.ftz.f32 %f2925, %f2918, %f2922; st.local.f32 [%rd1+24], %f2923; st.local.f32 [%rd1+28], %f2924; st.local.f32 [%rd1+32], %f2925; mul.ftz.f32 %f2926, %f3876, %f2923; mul.ftz.f32 %f2927, %f3875, %f2924; neg.ftz.f32 %f2928, %f2927; sub.ftz.f32 %f2929, %f2928, %f2926; mul.ftz.f32 %f2930, %f2925, %f3874; sub.ftz.f32 %f2931, %f2929, %f2930; abs.ftz.f32 %f590, %f2931; setp.eq.ftz.f32 %p261, %f590, 0f00000000; mov.f32 %f3961, 0f00000000; @%p261 bra $L__BB19_213; ld.const.f32 %f2932, [kernel_params+1500]; mul.ftz.f32 %f2933, %f589, %f589; mul.ftz.f32 %f2934, %f2933, %f2932; div.approx.ftz.f32 %f3961, %f2934, %f590; $L__BB19_213: st.local.f32 [%rd1+48], %f3961; setp.eq.ftz.f32 %p262, %f3960, 0f00000000; setp.eq.s16 %p263, %rs15, 0; or.pred %p264, %p263, %p262; @%p264 bra $L__BB19_223; cvta.to.global.u64 %rd205, %rd271; shl.b64 %rd206, %rd4, 4; add.s64 %rd207, %rd205, %rd206; ld.const.u64 %rd208, [kernel_params+2256]; cvta.to.global.u64 %rd209, %rd208; ld.global.u32 %r261, [%rd207+12]; mul.wide.u32 %rd210, %r261, 12; add.s64 %rd211, %rd209, %rd210; ld.global.f32 %f3896, [%rd211]; ld.global.f32 %f3897, [%rd211+4]; ld.global.f32 %f3898, [%rd211+8]; add.s32 %r262, %r261, 1; mul.wide.u32 %rd212, %r262, 12; add.s64 %rd213, %rd209, %rd212; ld.global.f32 %f3899, [%rd213]; ld.global.f32 %f3900, [%rd213+4]; ld.global.f32 %f3901, [%rd213+8]; add.s32 %r263, %r261, 2; mul.wide.u32 %rd214, %r263, 12; add.s64 %rd215, %rd209, %rd214; ld.global.f32 %f3902, [%rd215]; ld.global.f32 %f3903, [%rd215+4]; ld.global.f32 %f3904, [%rd215+8]; @%p250 bra $L__BB19_222; and.b32 %r264, %r85, 2; setp.eq.s32 %p266, %r264, 0; ld.const.u64 %rd216, [kernel_params+2160]; cvta.to.global.u64 %rd217, %rd216; mul.lo.s64 %rd218, %rd27, 208; add.s64 %rd92, %rd217, %rd218; @%p266 bra $L__BB19_220; bra.uni $L__BB19_216; $L__BB19_220: ld.global.v4.f32 {%f3884, %f3885, %f3886, %f3887}, [%rd92]; ld.global.v4.f32 {%f3888, %f3889, %f3890, %f3891}, [%rd92+16]; ld.global.v4.f32 {%f3892, %f3893, %f3894, %f3895}, [%rd92+32]; bra.uni $L__BB19_221; $L__BB19_11: ld.global.f32 %f5, [%rd7+-16]; ld.global.f32 %f6, [%rd7+-12]; ld.global.f32 %f7, [%rd7+-8]; st.local.f32 [%rd1], %f5; st.local.f32 [%rd1+4], %f6; st.local.f32 [%rd1+8], %f7; setp.eq.s32 %p16, %r12, 4; @%p16 bra $L__BB19_48; bra.uni $L__BB19_12; $L__BB19_48: ld.global.f32 %f154, [%rd7+124]; ld.global.v4.f32 {%f1404, %f1405, %f1406, %f1407}, [%rd7+140]; sub.ftz.f32 %f1408, %f934, %f5; sub.ftz.f32 %f1409, %f935, %f6; mul.ftz.f32 %f1410, %f1409, %f1409; fma.rn.ftz.f32 %f1411, %f1408, %f1408, %f1410; sub.ftz.f32 %f1412, %f936, %f7; fma.rn.ftz.f32 %f1413, %f1412, %f1412, %f1411; sqrt.approx.ftz.f32 %f1414, %f1413; div.approx.ftz.f32 %f159, %f1408, %f1414; div.approx.ftz.f32 %f160, %f1409, %f1414; div.approx.ftz.f32 %f161, %f1412, %f1414; st.local.f32 [%rd1], %f5; st.local.f32 [%rd1+4], %f6; st.local.f32 [%rd1+8], %f7; setp.leu.ftz.f32 %p66, %f154, 0f00000000; @%p66 bra $L__BB19_50; setp.neu.ftz.f32 %p67, %f159, %f160; setp.neu.ftz.f32 %p68, %f159, %f161; or.pred %p69, %p67, %p68; neg.ftz.f32 %f1415, %f161; selp.f32 %f1416, %f1415, %f161, %p69; add.ftz.f32 %f1417, %f159, %f1416; neg.ftz.f32 %f1418, %f160; selp.f32 %f1419, %f160, %f1418, %p69; sub.ftz.f32 %f1420, %f1419, %f159; sub.ftz.f32 %f1421, %f161, %f160; mul.ftz.f32 %f1422, %f1417, %f1417; fma.rn.ftz.f32 %f1423, %f1421, %f1421, %f1422; fma.rn.ftz.f32 %f1424, %f1420, %f1420, %f1423; sqrt.approx.ftz.f32 %f1425, %f1424; div.approx.ftz.f32 %f1426, %f1421, %f1425; div.approx.ftz.f32 %f1427, %f1417, %f1425; div.approx.ftz.f32 %f1428, %f1420, %f1425; mul.ftz.f32 %f1429, %f160, %f1428; mul.ftz.f32 %f1430, %f161, %f1427; sub.ftz.f32 %f1431, %f1429, %f1430; mul.ftz.f32 %f1432, %f161, %f1426; mul.ftz.f32 %f1433, %f159, %f1428; sub.ftz.f32 %f1434, %f1432, %f1433; mul.ftz.f32 %f1435, %f159, %f1427; mul.ftz.f32 %f1436, %f160, %f1426; sub.ftz.f32 %f1437, %f1435, %f1436; mul.ftz.f32 %f1438, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f1439, %f1438; sqrt.approx.ftz.f32 %f1440, %f932; mul.ftz.f32 %f1441, %f1440, %f1439; sin.approx.ftz.f32 %f1442, %f1438; mul.ftz.f32 %f1443, %f1440, %f1442; mul.ftz.f32 %f1444, %f1431, %f1443; mul.ftz.f32 %f1445, %f1434, %f1443; mul.ftz.f32 %f1446, %f1437, %f1443; fma.rn.ftz.f32 %f1447, %f1426, %f1441, %f1444; fma.rn.ftz.f32 %f1448, %f1427, %f1441, %f1445; fma.rn.ftz.f32 %f1449, %f1428, %f1441, %f1446; fma.rn.ftz.f32 %f5, %f154, %f1447, %f5; fma.rn.ftz.f32 %f6, %f154, %f1448, %f6; fma.rn.ftz.f32 %f7, %f154, %f1449, %f7; st.local.f32 [%rd1], %f5; st.local.f32 [%rd1+4], %f6; st.local.f32 [%rd1+8], %f7; $L__BB19_50: ld.global.f32 %f168, [%rd7+128]; st.local.f32 [%rd1+48], %f168; sub.ftz.f32 %f1451, %f5, %f934; sub.ftz.f32 %f1452, %f6, %f935; mul.ftz.f32 %f1453, %f1452, %f1452; fma.rn.ftz.f32 %f1454, %f1451, %f1451, %f1453; sub.ftz.f32 %f1455, %f7, %f936; fma.rn.ftz.f32 %f1456, %f1455, %f1455, %f1454; sqrt.approx.ftz.f32 %f169, %f1456; st.local.f32 [%rd1+36], %f169; rsqrt.approx.ftz.f32 %f1457, %f1456; mul.ftz.f32 %f170, %f1451, %f1457; mul.ftz.f32 %f171, %f1452, %f1457; mul.ftz.f32 %f172, %f1455, %f1457; st.local.f32 [%rd1+24], %f170; st.local.f32 [%rd1+28], %f171; st.local.f32 [%rd1+32], %f172; neg.ftz.f32 %f173, %f170; neg.ftz.f32 %f174, %f171; neg.ftz.f32 %f175, %f172; st.local.f32 [%rd1+12], %f173; st.local.f32 [%rd1+16], %f174; st.local.f32 [%rd1+20], %f175; mul.ftz.f32 %f176, %f168, 0f3DA2F983; st.local.f32 [%rd1+52], %f176; ld.global.v2.f32 {%f1458, %f1459}, [%rd7+132]; mul.ftz.f32 %f1460, %f1404, %f170; mul.ftz.f32 %f1461, %f1405, %f171; neg.ftz.f32 %f1462, %f1461; sub.ftz.f32 %f1463, %f1462, %f1460; mul.ftz.f32 %f1464, %f1406, %f172; sub.ftz.f32 %f179, %f1463, %f1464; setp.le.ftz.f32 %p70, %f179, %f1458; mov.f32 %f3781, 0f00000000; @%p70 bra $L__BB19_53; sub.ftz.f32 %f180, %f179, %f1458; setp.geu.ftz.f32 %p71, %f180, %f1459; setp.eq.ftz.f32 %p72, %f1459, 0f00000000; or.pred %p73, %p72, %p71; mov.f32 %f3781, %f179; @%p73 bra $L__BB19_53; div.approx.ftz.f32 %f1465, %f180, %f1459; mul.ftz.f32 %f1466, %f1465, %f1465; mul.ftz.f32 %f1467, %f1466, 0fC0000000; mul.ftz.f32 %f1468, %f1465, %f1467; fma.rn.ftz.f32 %f1469, %f1466, 0f40400000, %f1468; mul.ftz.f32 %f3781, %f179, %f1469; $L__BB19_53: mul.ftz.f32 %f1470, %f3781, %f176; st.local.f32 [%rd1+52], %f1470; setp.eq.ftz.f32 %p75, %f1470, 0f00000000; @%p75 bra $L__BB19_264; mul.ftz.f32 %f1473, %f171, %f171; fma.rn.ftz.f32 %f1474, %f170, %f170, %f1473; fma.rn.ftz.f32 %f183, %f172, %f172, %f1474; setp.leu.ftz.f32 %p76, %f183, 0f00000000; mov.f32 %f3786, 0f00000000; mov.f32 %f3784, %f3786; mov.f32 %f3785, %f3786; @%p76 bra $L__BB19_63; setp.eq.ftz.f32 %p77, %f170, 0f80000000; mov.f32 %f3784, 0f00000000; setp.eq.ftz.f32 %p78, %f171, 0f80000000; and.pred %p79, %p77, %p78; @%p79 bra $L__BB19_62; abs.ftz.f32 %f184, %f174; setp.eq.ftz.f32 %p80, %f184, 0f00000000; abs.ftz.f32 %f185, %f173; setp.eq.ftz.f32 %p81, %f185, 0f00000000; and.pred %p82, %p80, %p81; mov.b32 %r17, %f174; mov.b32 %r131, %f173; and.b32 %r18, %r131, -2147483648; @%p82 bra $L__BB19_60; bra.uni $L__BB19_57; $L__BB19_60: shr.s32 %r136, %r17, 31; and.b32 %r137, %r136, 1078530011; or.b32 %r138, %r137, %r18; mov.b32 %f3782, %r138; bra.uni $L__BB19_61; $L__BB19_216: ld.const.u64 %rd219, [kernel_params+2176]; cvta.to.global.u64 %rd220, %rd219; ld.global.u32 %rd221, [%rd92+168]; ld.global.u32 %r265, [%rd92+152]; shl.b32 %r266, %r265, 1; cvt.rn.f32.s32 %f2935, %r266; mul.ftz.f32 %f2936, %f2935, 0f3F000000; cvt.rzi.ftz.s32.f32 %r267, %f2936; add.s32 %r268, %r266, -1; min.s32 %r269, %r267, %r268; cvt.rn.f32.s32 %f2937, %r269; sub.ftz.f32 %f602, %f2936, %f2937; cvt.s64.s32 %rd222, %r269; add.s64 %rd223, %rd222, %rd221; shl.b64 %rd224, %rd223, 6; add.s64 %rd93, %rd220, %rd224; ld.global.v4.f32 {%f2938, %f2939, %f2940, %f2941}, [%rd93]; ld.global.v4.f32 {%f2942, %f2943, %f2944, %f2945}, [%rd93+64]; mul.ftz.f32 %f2946, %f2939, %f2943; fma.rn.ftz.f32 %f2947, %f2938, %f2942, %f2946; mul.ftz.f32 %f2948, %f2941, %f2945; fma.rn.ftz.f32 %f2949, %f2940, %f2944, %f2948; add.ftz.f32 %f611, %f2947, %f2949; setp.gt.ftz.f32 %p267, %f611, 0f3F7FDF3B; sub.ftz.f32 %f612, %f2904, %f602; @%p267 bra $L__BB19_218; bra.uni $L__BB19_217; $L__BB19_218: mul.ftz.f32 %f3004, %f2938, %f612; fma.rn.ftz.f32 %f3005, %f2942, %f602, %f3004; mul.ftz.f32 %f3006, %f2939, %f612; fma.rn.ftz.f32 %f3007, %f2943, %f602, %f3006; mul.ftz.f32 %f3008, %f2940, %f612; fma.rn.ftz.f32 %f3009, %f2944, %f602, %f3008; mul.ftz.f32 %f3010, %f2941, %f612; fma.rn.ftz.f32 %f3011, %f2945, %f602, %f3010; mul.ftz.f32 %f3012, %f3007, %f3007; fma.rn.ftz.f32 %f3013, %f3005, %f3005, %f3012; mul.ftz.f32 %f3014, %f3011, %f3011; fma.rn.ftz.f32 %f3015, %f3009, %f3009, %f3014; add.ftz.f32 %f3016, %f3013, %f3015; rsqrt.approx.ftz.f32 %f3017, %f3016; mul.ftz.f32 %f3883, %f3011, %f3017; mul.ftz.f32 %f3882, %f3009, %f3017; mul.ftz.f32 %f3881, %f3007, %f3017; mul.ftz.f32 %f3880, %f3005, %f3017; bra.uni $L__BB19_219; $L__BB19_247: setp.lt.ftz.f32 %p298, %f833, 0f219392EF; @%p298 bra $L__BB19_249; rcp.approx.ftz.f32 %f3934, %f832; $L__BB19_249: mul.ftz.f32 %f3464, %f827, %f3934; min.ftz.f32 %f3466, %f3464, %f3384; st.local.f32 [%rd1+40], %f3466; mul.ftz.f32 %f3467, %f831, %f3934; min.ftz.f32 %f3468, %f3467, %f3384; st.local.f32 [%rd1+44], %f3468; st.local.f32 [%rd1+36], %f838; fma.rn.ftz.f32 %f3469, %f838, %f821, %f934; fma.rn.ftz.f32 %f3470, %f838, %f822, %f935; fma.rn.ftz.f32 %f3471, %f838, %f823, %f936; st.local.f32 [%rd1], %f3469; st.local.f32 [%rd1+4], %f3470; st.local.f32 [%rd1+8], %f3471; setp.eq.ftz.f32 %p299, %f772, 0f00000000; @%p299 bra $L__BB19_261; bra.uni $L__BB19_250; $L__BB19_261: mov.f32 %f3961, 0f00000000; mov.u32 %r289, 0; st.local.u32 [%rd1+48], %r289; bra.uni $L__BB19_263; $L__BB19_12: setp.eq.s32 %p17, %r12, 0; @%p17 bra $L__BB19_33; ld.global.v4.f32 {%f944, %f945, %f946, %f947}, [%rd7+124]; ld.global.v4.f32 {%f948, %f949, %f950, %f951}, [%rd7+140]; ld.global.v4.f32 {%f952, %f953, %f954, %f955}, [%rd7+156]; abs.ftz.f32 %f3836, %f947; sub.ftz.f32 %f956, %f5, %f934; sub.ftz.f32 %f957, %f6, %f935; mul.ftz.f32 %f958, %f953, %f957; fma.rn.ftz.f32 %f959, %f952, %f956, %f958; sub.ftz.f32 %f960, %f7, %f936; fma.rn.ftz.f32 %f961, %f954, %f960, %f959; setp.gt.ftz.f32 %p19, %f961, 0f00000000; @%p19 bra $L__BB19_264; setp.lt.ftz.f32 %p20, %f947, 0f00000000; @%p20 bra $L__BB19_26; bra.uni $L__BB19_15; $L__BB19_26: mul.ftz.f32 %f1221, %f944, 0f3F000000; mul.ftz.f32 %f1222, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f1223, %f1222; sqrt.approx.ftz.f32 %f1224, %f932; mul.ftz.f32 %f1225, %f1224, %f1223; sin.approx.ftz.f32 %f1226, %f1222; mul.ftz.f32 %f1227, %f1224, %f1226; mul.ftz.f32 %f1228, %f945, 0f3F000000; mul.ftz.f32 %f1229, %f946, 0f3F000000; mul.ftz.f32 %f1230, %f948, 0f3F000000; mul.ftz.f32 %f1231, %f1230, %f1227; mul.ftz.f32 %f1232, %f949, 0f3F000000; mul.ftz.f32 %f1233, %f1232, %f1227; mul.ftz.f32 %f1234, %f950, 0f3F000000; mul.ftz.f32 %f1235, %f1234, %f1227; fma.rn.ftz.f32 %f3766, %f1221, %f1225, %f1231; fma.rn.ftz.f32 %f3767, %f1228, %f1225, %f1233; fma.rn.ftz.f32 %f3768, %f1229, %f1225, %f1235; add.ftz.f32 %f3765, %f5, %f3766; add.ftz.f32 %f3764, %f6, %f3767; add.ftz.f32 %f3763, %f7, %f3768; st.local.f32 [%rd1], %f3765; st.local.f32 [%rd1+4], %f3764; st.local.f32 [%rd1+8], %f3763; st.local.f32 [%rd1+48], %f3836; ld.global.f32 %f3761, [%rd7+152]; bra.uni $L__BB19_27; $L__BB19_217: mov.f32 %f2951, 0fBF800000; max.ftz.f32 %f2952, %f611, %f2951; min.ftz.f32 %f2954, %f2952, %f2904; abs.ftz.f32 %f2955, %f2954; sub.ftz.f32 %f2956, %f2904, %f2955; mul.ftz.f32 %f2957, %f2956, 0f3F000000; sqrt.approx.ftz.f32 %f2958, %f2957; setp.gt.ftz.f32 %p268, %f2955, 0f3F11EB85; selp.f32 %f2959, %f2958, %f2955, %p268; mul.ftz.f32 %f2960, %f2959, %f2959; mov.f32 %f2961, 0f3C94D2E9; mov.f32 %f2962, 0f3D53F941; fma.rn.ftz.f32 %f2963, %f2962, %f2960, %f2961; mov.f32 %f2964, 0f3D3F841F; fma.rn.ftz.f32 %f2965, %f2963, %f2960, %f2964; mov.f32 %f2966, 0f3D994929; fma.rn.ftz.f32 %f2967, %f2965, %f2960, %f2966; mov.f32 %f2968, 0f3E2AAB94; fma.rn.ftz.f32 %f2969, %f2967, %f2960, %f2968; mul.ftz.f32 %f2970, %f2960, %f2969; fma.rn.ftz.f32 %f2971, %f2970, %f2959, %f2959; add.ftz.f32 %f2972, %f2971, %f2971; mov.f32 %f2973, 0f3FC90FDB; sub.ftz.f32 %f2974, %f2973, %f2971; selp.f32 %f2975, %f2972, %f2974, %p268; setp.lt.ftz.f32 %p269, %f2954, 0f00000000; mov.f32 %f2976, 0f40490FDB; sub.ftz.f32 %f2977, %f2976, %f2975; selp.f32 %f2978, %f2977, %f2975, %p269; mul.ftz.f32 %f2979, %f2938, %f611; sub.ftz.f32 %f2980, %f2942, %f2979; mul.ftz.f32 %f2981, %f2939, %f611; sub.ftz.f32 %f2982, %f2943, %f2981; mul.ftz.f32 %f2983, %f2940, %f611; sub.ftz.f32 %f2984, %f2944, %f2983; mul.ftz.f32 %f2985, %f2941, %f611; sub.ftz.f32 %f2986, %f2945, %f2985; mul.ftz.f32 %f2987, %f2982, %f2982; fma.rn.ftz.f32 %f2988, %f2980, %f2980, %f2987; mul.ftz.f32 %f2989, %f2986, %f2986; fma.rn.ftz.f32 %f2990, %f2984, %f2984, %f2989; add.ftz.f32 %f2991, %f2988, %f2990; rsqrt.approx.ftz.f32 %f2992, %f2991; mul.ftz.f32 %f2993, %f2980, %f2992; mul.ftz.f32 %f2994, %f2982, %f2992; mul.ftz.f32 %f2995, %f2984, %f2992; mul.ftz.f32 %f2996, %f2986, %f2992; mul.ftz.f32 %f2997, %f602, %f2978; cos.approx.ftz.f32 %f2998, %f2997; sin.approx.ftz.f32 %f2999, %f2997; mul.ftz.f32 %f3000, %f2993, %f2999; mul.ftz.f32 %f3001, %f2994, %f2999; mul.ftz.f32 %f3002, %f2995, %f2999; mul.ftz.f32 %f3003, %f2996, %f2999; fma.rn.ftz.f32 %f3883, %f2941, %f2998, %f3003; fma.rn.ftz.f32 %f3882, %f2940, %f2998, %f3002; fma.rn.ftz.f32 %f3881, %f2939, %f2998, %f3001; fma.rn.ftz.f32 %f3880, %f2938, %f2998, %f3000; $L__BB19_219: ld.global.v4.f32 {%f3018, %f3019, %f3020, %f3021}, [%rd93+16]; ld.global.v4.f32 {%f3026, %f3027, %f3028, %f3029}, [%rd93+80]; mul.ftz.f32 %f3034, %f602, %f3026; mul.ftz.f32 %f3035, %f602, %f3027; mul.ftz.f32 %f3036, %f602, %f3028; mul.ftz.f32 %f3037, %f602, %f3029; fma.rn.ftz.f32 %f3038, %f612, %f3021, %f3037; ld.global.v4.f32 {%f3039, %f3040, %f3041, %f3042}, [%rd93+32]; ld.global.v4.f32 {%f3047, %f3048, %f3049, %f3050}, [%rd93+96]; mul.ftz.f32 %f3055, %f602, %f3047; mul.ftz.f32 %f3056, %f602, %f3048; mul.ftz.f32 %f3057, %f602, %f3049; mul.ftz.f32 %f3058, %f602, %f3050; fma.rn.ftz.f32 %f3059, %f612, %f3039, %f3055; fma.rn.ftz.f32 %f3060, %f612, %f3040, %f3056; fma.rn.ftz.f32 %f3061, %f612, %f3041, %f3057; fma.rn.ftz.f32 %f3062, %f612, %f3042, %f3058; ld.global.v4.f32 {%f3063, %f3064, %f3065, %f3066}, [%rd93+48]; ld.global.v4.f32 {%f3071, %f3072, %f3073, %f3074}, [%rd93+112]; mul.ftz.f32 %f3079, %f602, %f3071; mul.ftz.f32 %f3080, %f602, %f3072; mul.ftz.f32 %f3081, %f602, %f3073; mul.ftz.f32 %f3082, %f602, %f3074; fma.rn.ftz.f32 %f3083, %f612, %f3063, %f3079; fma.rn.ftz.f32 %f3084, %f612, %f3064, %f3080; fma.rn.ftz.f32 %f3085, %f612, %f3065, %f3081; fma.rn.ftz.f32 %f3086, %f612, %f3066, %f3082; mul.ftz.f32 %f3087, %f3883, 0f3FB504F3; mul.ftz.f32 %f3088, %f3880, 0f3FB504F3; mul.ftz.f32 %f3089, %f3881, 0f3FB504F3; mul.ftz.f32 %f3090, %f3882, 0f3FB504F3; mul.ftz.f32 %f3091, %f3087, %f3088; mul.ftz.f32 %f3092, %f3087, %f3089; mul.ftz.f32 %f3093, %f3087, %f3090; mul.ftz.f32 %f3094, %f3088, %f3088; mul.ftz.f32 %f3095, %f3088, %f3089; mul.ftz.f32 %f3096, %f3088, %f3090; mul.ftz.f32 %f3097, %f3089, %f3089; mul.ftz.f32 %f3098, %f3089, %f3090; mul.ftz.f32 %f3099, %f3090, %f3090; sub.ftz.f32 %f3101, %f2904, %f3097; sub.ftz.f32 %f3102, %f3101, %f3099; sub.ftz.f32 %f3103, %f3095, %f3093; add.ftz.f32 %f3104, %f3092, %f3096; add.ftz.f32 %f3105, %f3093, %f3095; sub.ftz.f32 %f3106, %f2904, %f3094; sub.ftz.f32 %f3107, %f3106, %f3099; sub.ftz.f32 %f3108, %f3098, %f3091; sub.ftz.f32 %f3109, %f3096, %f3092; add.ftz.f32 %f3110, %f3091, %f3098; sub.ftz.f32 %f3111, %f3106, %f3097; mul.ftz.f32 %f3112, %f3103, %f3061; fma.rn.ftz.f32 %f3113, %f3102, %f3038, %f3112; mul.ftz.f32 %f3114, %f3103, %f3062; fma.rn.ftz.f32 %f3115, %f3102, %f3059, %f3114; mul.ftz.f32 %f3116, %f3103, %f3083; fma.rn.ftz.f32 %f3117, %f3102, %f3060, %f3116; mul.ftz.f32 %f3118, %f3107, %f3061; fma.rn.ftz.f32 %f3119, %f3105, %f3038, %f3118; mul.ftz.f32 %f3120, %f3107, %f3062; fma.rn.ftz.f32 %f3121, %f3105, %f3059, %f3120; mul.ftz.f32 %f3122, %f3107, %f3083; fma.rn.ftz.f32 %f3123, %f3105, %f3060, %f3122; mul.ftz.f32 %f3124, %f3110, %f3061; fma.rn.ftz.f32 %f3125, %f3109, %f3038, %f3124; mul.ftz.f32 %f3126, %f3110, %f3062; fma.rn.ftz.f32 %f3127, %f3109, %f3059, %f3126; mul.ftz.f32 %f3128, %f3110, %f3083; fma.rn.ftz.f32 %f3129, %f3109, %f3060, %f3128; fma.rn.ftz.f32 %f3890, %f3108, %f3086, %f3123; fma.rn.ftz.f32 %f3889, %f3108, %f3085, %f3121; fma.rn.ftz.f32 %f3888, %f3108, %f3084, %f3119; fma.rn.ftz.f32 %f3894, %f3111, %f3086, %f3129; fma.rn.ftz.f32 %f3893, %f3111, %f3085, %f3127; fma.rn.ftz.f32 %f3892, %f3111, %f3084, %f3125; fma.rn.ftz.f32 %f3886, %f3104, %f3086, %f3117; fma.rn.ftz.f32 %f3885, %f3104, %f3085, %f3115; fma.rn.ftz.f32 %f3884, %f3104, %f3084, %f3113; fma.rn.ftz.f32 %f3895, %f612, %f3020, %f3036; fma.rn.ftz.f32 %f3891, %f612, %f3019, %f3035; fma.rn.ftz.f32 %f3887, %f612, %f3018, %f3034; $L__BB19_221: mul.ftz.f32 %f3142, %f3897, %f3885; fma.rn.ftz.f32 %f3143, %f3896, %f3884, %f3142; fma.rn.ftz.f32 %f3144, %f3898, %f3886, %f3143; add.ftz.f32 %f661, %f3887, %f3144; mul.ftz.f32 %f3145, %f3897, %f3889; fma.rn.ftz.f32 %f3146, %f3896, %f3888, %f3145; fma.rn.ftz.f32 %f3147, %f3898, %f3890, %f3146; add.ftz.f32 %f662, %f3891, %f3147; mul.ftz.f32 %f3148, %f3897, %f3893; fma.rn.ftz.f32 %f3149, %f3896, %f3892, %f3148; fma.rn.ftz.f32 %f3150, %f3898, %f3894, %f3149; add.ftz.f32 %f3898, %f3895, %f3150; mul.ftz.f32 %f3151, %f3900, %f3885; fma.rn.ftz.f32 %f3152, %f3899, %f3884, %f3151; fma.rn.ftz.f32 %f3153, %f3901, %f3886, %f3152; add.ftz.f32 %f664, %f3887, %f3153; mul.ftz.f32 %f3154, %f3900, %f3889; fma.rn.ftz.f32 %f3155, %f3899, %f3888, %f3154; fma.rn.ftz.f32 %f3156, %f3901, %f3890, %f3155; add.ftz.f32 %f665, %f3891, %f3156; mul.ftz.f32 %f3157, %f3900, %f3893; fma.rn.ftz.f32 %f3158, %f3899, %f3892, %f3157; fma.rn.ftz.f32 %f3159, %f3901, %f3894, %f3158; add.ftz.f32 %f3901, %f3895, %f3159; mul.ftz.f32 %f3160, %f3903, %f3885; fma.rn.ftz.f32 %f3161, %f3902, %f3884, %f3160; fma.rn.ftz.f32 %f3162, %f3904, %f3886, %f3161; add.ftz.f32 %f667, %f3887, %f3162; mul.ftz.f32 %f3163, %f3903, %f3889; fma.rn.ftz.f32 %f3164, %f3902, %f3888, %f3163; fma.rn.ftz.f32 %f3165, %f3904, %f3890, %f3164; add.ftz.f32 %f668, %f3891, %f3165; mul.ftz.f32 %f3166, %f3903, %f3893; fma.rn.ftz.f32 %f3167, %f3902, %f3892, %f3166; fma.rn.ftz.f32 %f3168, %f3904, %f3894, %f3167; add.ftz.f32 %f3904, %f3895, %f3168; mov.f32 %f3896, %f661; mov.f32 %f3897, %f662; mov.f32 %f3899, %f664; mov.f32 %f3900, %f665; mov.f32 %f3902, %f667; mov.f32 %f3903, %f668; $L__BB19_222: sub.ftz.f32 %f3169, %f3898, %f3901; sub.ftz.f32 %f3170, %f3903, %f3900; mul.ftz.f32 %f3171, %f3169, %f3170; sub.ftz.f32 %f3172, %f3897, %f3900; sub.ftz.f32 %f3173, %f3904, %f3901; mul.ftz.f32 %f3174, %f3172, %f3173; sub.ftz.f32 %f3175, %f3171, %f3174; sub.ftz.f32 %f3176, %f3896, %f3899; mul.ftz.f32 %f3177, %f3176, %f3173; sub.ftz.f32 %f3178, %f3902, %f3899; mul.ftz.f32 %f3179, %f3169, %f3178; sub.ftz.f32 %f3180, %f3177, %f3179; mul.ftz.f32 %f3181, %f3172, %f3178; mul.ftz.f32 %f3182, %f3176, %f3170; sub.ftz.f32 %f3183, %f3181, %f3182; mul.ftz.f32 %f3184, %f3180, %f3180; fma.rn.ftz.f32 %f3185, %f3175, %f3175, %f3184; fma.rn.ftz.f32 %f3186, %f3183, %f3183, %f3185; sqrt.approx.ftz.f32 %f3187, %f3186; mul.ftz.f32 %f3188, %f3187, 0f3F000000; mul.ftz.f32 %f3189, %f3961, %f3188; div.approx.ftz.f32 %f3961, %f3189, %f3960; st.local.f32 [%rd1+48], %f3961; $L__BB19_223: st.local.f32 [%rd1+40], %f3877; st.local.f32 [%rd1+44], %f3878; $L__BB19_263: or.b32 %r291, %r105, %r84; st.local.u32 [%rd1+64], %r291; setp.gt.ftz.f32 %p310, %f3961, 0f00000000; $L__BB19_264: selp.u32 %r292, 1, 0, %p310; st.param.b32 [func_retval0+0], %r292; ret; $L__BB19_76: rcp.approx.ftz.f32 %f1622, %f204; mul.ftz.f32 %f205, %f3787, %f1622; mul.ftz.f32 %f206, %f1539, %f1622; mul.ftz.f32 %f207, %f1540, %f1622; add.ftz.f32 %f208, %f205, %f206; setp.gt.ftz.f32 %p106, %f205, %f1; @%p106 bra $L__BB19_93; bra.uni $L__BB19_77; $L__BB19_93: setp.eq.ftz.f32 %p124, %f205, 0f3F800000; @%p124 bra $L__BB19_95; div.approx.ftz.f32 %f1, %f1, %f205; $L__BB19_95: cvt.rn.f32.s32 %f232, %r305; mul.ftz.f32 %f1715, %f232, %f932; cvt.rzi.ftz.s32.f32 %r315, %f1715; cvt.rn.f32.s32 %f1716, %r315; sub.ftz.f32 %f233, %f1715, %f1716; ld.const.u32 %r53, [kernel_params+1048]; setp.lt.s32 %p125, %r53, 1; mov.f32 %f3830, 0f00000000; mov.u32 %r316, 0; mov.f32 %f3831, %f3830; mov.f32 %f3832, %f3830; @%p125 bra $L__BB19_113; mov.u32 %r316, 0; ld.const.u32 %r54, [kernel_params+1052]; $L__BB19_97: add.s32 %r184, %r54, %r316; mul.wide.s32 %rd121, %r184, 192; add.s64 %rd122, %rd6, %rd121; add.s64 %rd16, %rd122, 4; ld.global.v4.f32 {%f1717, %f1718, %f1719, %f1720}, [%rd122+176]; ld.global.f32 %f237, [%rd122+4]; sub.ftz.f32 %f1721, %f934, %f237; ld.global.f32 %f238, [%rd122+8]; sub.ftz.f32 %f1722, %f935, %f238; ld.global.f32 %f239, [%rd122+12]; sub.ftz.f32 %f1723, %f936, %f239; mul.ftz.f32 %f1724, %f1722, %f1718; fma.rn.ftz.f32 %f1725, %f1721, %f1717, %f1724; fma.rn.ftz.f32 %f1726, %f1723, %f1719, %f1725; setp.leu.ftz.f32 %p126, %f1726, 0f38D1B717; @%p126 bra $L__BB19_100; setp.eq.s32 %p127, %r315, 0; @%p127 bra $L__BB19_101; add.s32 %r315, %r315, -1; $L__BB19_100: add.s32 %r316, %r316, 1; setp.lt.s32 %p142, %r316, %r53; mov.f32 %f3831, %f3830; mov.f32 %f3832, %f3830; @%p142 bra $L__BB19_97; bra.uni $L__BB19_113; $L__BB19_250: setp.eq.s16 %p300, %rs15, 0; @%p300 bra $L__BB19_260; cvta.to.global.u64 %rd225, %rd271; shl.b64 %rd226, %rd4, 4; add.s64 %rd227, %rd225, %rd226; ld.const.u64 %rd228, [kernel_params+2256]; cvta.to.global.u64 %rd229, %rd228; ld.global.u32 %r280, [%rd227+12]; mul.wide.u32 %rd230, %r280, 12; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f3951, [%rd231]; ld.global.f32 %f3952, [%rd231+4]; ld.global.f32 %f3953, [%rd231+8]; add.s32 %r281, %r280, 1; mul.wide.u32 %rd232, %r281, 12; add.s64 %rd233, %rd229, %rd232; ld.global.f32 %f3954, [%rd233]; ld.global.f32 %f3955, [%rd233+4]; ld.global.f32 %f3956, [%rd233+8]; add.s32 %r282, %r280, 2; mul.wide.u32 %rd234, %r282, 12; add.s64 %rd235, %rd229, %rd234; ld.global.f32 %f3957, [%rd235]; ld.global.f32 %f3958, [%rd235+4]; ld.global.f32 %f3959, [%rd235+8]; @%p250 bra $L__BB19_259; and.b32 %r283, %r85, 2; setp.eq.s32 %p302, %r283, 0; ld.const.u64 %rd236, [kernel_params+2160]; cvta.to.global.u64 %rd237, %rd236; mul.lo.s64 %rd238, %rd27, 208; add.s64 %rd94, %rd237, %rd238; @%p302 bra $L__BB19_257; bra.uni $L__BB19_253; $L__BB19_257: ld.global.v4.f32 {%f3939, %f3940, %f3941, %f3942}, [%rd94]; ld.global.v4.f32 {%f3943, %f3944, %f3945, %f3946}, [%rd94+16]; ld.global.v4.f32 {%f3947, %f3948, %f3949, %f3950}, [%rd94+32]; bra.uni $L__BB19_258; $L__BB19_33: ld.global.f32 %f120, [%rd7+124]; sub.ftz.f32 %f1286, %f934, %f5; sub.ftz.f32 %f1287, %f935, %f6; mul.ftz.f32 %f1288, %f1287, %f1287; fma.rn.ftz.f32 %f1289, %f1286, %f1286, %f1288; sub.ftz.f32 %f1290, %f936, %f7; fma.rn.ftz.f32 %f1291, %f1290, %f1290, %f1289; sqrt.approx.ftz.f32 %f1292, %f1291; div.approx.ftz.f32 %f121, %f1286, %f1292; div.approx.ftz.f32 %f122, %f1287, %f1292; div.approx.ftz.f32 %f123, %f1290, %f1292; st.local.f32 [%rd1], %f5; st.local.f32 [%rd1+4], %f6; st.local.f32 [%rd1+8], %f7; setp.leu.ftz.f32 %p43, %f120, 0f00000000; @%p43 bra $L__BB19_35; setp.neu.ftz.f32 %p44, %f121, %f122; setp.neu.ftz.f32 %p45, %f121, %f123; or.pred %p46, %p44, %p45; neg.ftz.f32 %f1293, %f123; selp.f32 %f1294, %f1293, %f123, %p46; add.ftz.f32 %f1295, %f121, %f1294; neg.ftz.f32 %f1296, %f122; selp.f32 %f1297, %f122, %f1296, %p46; sub.ftz.f32 %f1298, %f1297, %f121; sub.ftz.f32 %f1299, %f123, %f122; mul.ftz.f32 %f1300, %f1295, %f1295; fma.rn.ftz.f32 %f1301, %f1299, %f1299, %f1300; fma.rn.ftz.f32 %f1302, %f1298, %f1298, %f1301; sqrt.approx.ftz.f32 %f1303, %f1302; div.approx.ftz.f32 %f1304, %f1299, %f1303; div.approx.ftz.f32 %f1305, %f1295, %f1303; div.approx.ftz.f32 %f1306, %f1298, %f1303; mul.ftz.f32 %f1307, %f122, %f1306; mul.ftz.f32 %f1308, %f123, %f1305; sub.ftz.f32 %f1309, %f1307, %f1308; mul.ftz.f32 %f1310, %f123, %f1304; mul.ftz.f32 %f1311, %f121, %f1306; sub.ftz.f32 %f1312, %f1310, %f1311; mul.ftz.f32 %f1313, %f121, %f1305; mul.ftz.f32 %f1314, %f122, %f1304; sub.ftz.f32 %f1315, %f1313, %f1314; mul.ftz.f32 %f1316, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f1317, %f1316; sqrt.approx.ftz.f32 %f1318, %f932; mul.ftz.f32 %f1319, %f1318, %f1317; sin.approx.ftz.f32 %f1320, %f1316; mul.ftz.f32 %f1321, %f1318, %f1320; mul.ftz.f32 %f1322, %f1309, %f1321; mul.ftz.f32 %f1323, %f1312, %f1321; mul.ftz.f32 %f1324, %f1315, %f1321; fma.rn.ftz.f32 %f1325, %f1304, %f1319, %f1322; fma.rn.ftz.f32 %f1326, %f1305, %f1319, %f1323; fma.rn.ftz.f32 %f1327, %f1306, %f1319, %f1324; fma.rn.ftz.f32 %f5, %f120, %f1325, %f5; fma.rn.ftz.f32 %f6, %f120, %f1326, %f6; fma.rn.ftz.f32 %f7, %f120, %f1327, %f7; st.local.f32 [%rd1], %f5; st.local.f32 [%rd1+4], %f6; st.local.f32 [%rd1+8], %f7; $L__BB19_35: ld.global.f32 %f130, [%rd7+128]; st.local.f32 [%rd1+48], %f130; sub.ftz.f32 %f1328, %f5, %f934; sub.ftz.f32 %f1329, %f6, %f935; mul.ftz.f32 %f1330, %f1329, %f1329; fma.rn.ftz.f32 %f1331, %f1328, %f1328, %f1330; sub.ftz.f32 %f1332, %f7, %f936; fma.rn.ftz.f32 %f1333, %f1332, %f1332, %f1331; sqrt.approx.ftz.f32 %f131, %f1333; st.local.f32 [%rd1+36], %f131; rsqrt.approx.ftz.f32 %f1334, %f1333; mul.ftz.f32 %f132, %f1328, %f1334; mul.ftz.f32 %f133, %f1329, %f1334; mul.ftz.f32 %f134, %f1332, %f1334; st.local.f32 [%rd1+24], %f132; st.local.f32 [%rd1+28], %f133; st.local.f32 [%rd1+32], %f134; neg.ftz.f32 %f135, %f132; neg.ftz.f32 %f136, %f133; neg.ftz.f32 %f137, %f134; st.local.f32 [%rd1+12], %f135; st.local.f32 [%rd1+16], %f136; st.local.f32 [%rd1+20], %f137; mul.ftz.f32 %f1335, %f130, 0f3DA2F983; st.local.f32 [%rd1+52], %f1335; setp.eq.ftz.f32 %p48, %f1335, 0f00000000; @%p48 bra $L__BB19_264; mul.ftz.f32 %f1338, %f133, %f133; fma.rn.ftz.f32 %f1339, %f132, %f132, %f1338; fma.rn.ftz.f32 %f138, %f134, %f134, %f1339; setp.leu.ftz.f32 %p49, %f138, 0f00000000; mov.f32 %f3777, 0f00000000; mov.f32 %f3775, %f3777; mov.f32 %f3776, %f3777; @%p49 bra $L__BB19_45; setp.eq.ftz.f32 %p50, %f132, 0f80000000; mov.f32 %f3775, 0f00000000; setp.eq.ftz.f32 %p51, %f133, 0f80000000; and.pred %p52, %p50, %p51; @%p52 bra $L__BB19_44; abs.ftz.f32 %f139, %f136; setp.eq.ftz.f32 %p53, %f139, 0f00000000; abs.ftz.f32 %f140, %f135; setp.eq.ftz.f32 %p54, %f140, 0f00000000; and.pred %p55, %p53, %p54; mov.b32 %r15, %f136; mov.b32 %r123, %f135; and.b32 %r16, %r123, -2147483648; @%p55 bra $L__BB19_42; bra.uni $L__BB19_39; $L__BB19_42: shr.s32 %r128, %r15, 31; and.b32 %r129, %r128, 1078530011; or.b32 %r130, %r129, %r16; mov.b32 %f3773, %r130; bra.uni $L__BB19_43; $L__BB19_77: setp.gt.ftz.f32 %p107, %f208, %f1; @%p107 bra $L__BB19_90; bra.uni $L__BB19_78; $L__BB19_90: setp.eq.ftz.f32 %p119, %f206, 0f3F800000; @%p119 bra $L__BB19_92; sub.ftz.f32 %f1657, %f1, %f205; div.approx.ftz.f32 %f1, %f1657, %f206; $L__BB19_92: ld.const.v4.f32 {%f1658, %f1659, %f1660, %f1661}, [kernel_params+1008]; mov.u32 %r322, 1; ld.const.f32 %f1665, [kernel_params+1020]; cos.approx.ftz.f32 %f1666, %f1665; mul.ftz.f32 %f1667, %f1, %f1666; sub.ftz.f32 %f1668, %f1666, %f1667; add.ftz.f32 %f1669, %f1, %f1668; mul.ftz.f32 %f1670, %f1669, %f1669; mov.f32 %f1671, 0f3F800000; sub.ftz.f32 %f1672, %f1671, %f1670; mov.f32 %f1673, 0f00000000; max.ftz.f32 %f1674, %f1672, %f1673; sqrt.approx.ftz.f32 %f1675, %f1674; mul.ftz.f32 %f1676, %f932, 0f40C90FDB; cos.approx.ftz.f32 %f1677, %f1676; mul.ftz.f32 %f1678, %f1675, %f1677; sin.approx.ftz.f32 %f1679, %f1676; mul.ftz.f32 %f1680, %f1675, %f1679; setp.neu.ftz.f32 %p120, %f1658, %f1659; setp.neu.ftz.f32 %p121, %f1658, %f1660; or.pred %p122, %p120, %p121; sub.ftz.f32 %f1681, %f1660, %f1659; neg.ftz.f32 %f1682, %f1659; neg.ftz.f32 %f1683, %f1660; selp.f32 %f1684, %f1683, %f1660, %p122; add.ftz.f32 %f1685, %f1658, %f1684; selp.f32 %f1686, %f1659, %f1682, %p122; sub.ftz.f32 %f1687, %f1686, %f1658; mul.ftz.f32 %f1688, %f1685, %f1685; fma.rn.ftz.f32 %f1689, %f1681, %f1681, %f1688; fma.rn.ftz.f32 %f1690, %f1687, %f1687, %f1689; sqrt.approx.ftz.f32 %f1691, %f1690; div.approx.ftz.f32 %f1692, %f1681, %f1691; div.approx.ftz.f32 %f1693, %f1685, %f1691; div.approx.ftz.f32 %f1694, %f1687, %f1691; mul.ftz.f32 %f1695, %f1659, %f1694; mul.ftz.f32 %f1696, %f1660, %f1693; sub.ftz.f32 %f1697, %f1695, %f1696; mul.ftz.f32 %f1698, %f1660, %f1692; mul.ftz.f32 %f1699, %f1658, %f1694; sub.ftz.f32 %f1700, %f1698, %f1699; mul.ftz.f32 %f1701, %f1658, %f1693; mul.ftz.f32 %f1702, %f1659, %f1692; sub.ftz.f32 %f1703, %f1701, %f1702; mul.ftz.f32 %f1704, %f1680, %f1697; mul.ftz.f32 %f1705, %f1680, %f1700; mul.ftz.f32 %f1706, %f1680, %f1703; fma.rn.ftz.f32 %f1707, %f1678, %f1692, %f1704; fma.rn.ftz.f32 %f1708, %f1678, %f1693, %f1705; fma.rn.ftz.f32 %f1709, %f1678, %f1694, %f1706; sub.ftz.f32 %f1710, %f1671, %f1666; mov.f32 %f1711, 0f3E22F983; div.approx.ftz.f32 %f3812, %f1711, %f1710; st.local.f32 [%rd1+48], %f3812; fma.rn.ftz.f32 %f3832, %f1660, %f1669, %f1709; fma.rn.ftz.f32 %f3831, %f1659, %f1669, %f1708; fma.rn.ftz.f32 %f3830, %f1658, %f1669, %f1707; mov.f32 %f3816, %f206; @%p119 bra $L__BB19_171; bra.uni $L__BB19_135; $L__BB19_78: setp.eq.ftz.f32 %p108, %f207, 0f3F800000; @%p108 bra $L__BB19_80; sub.ftz.f32 %f1623, %f1, %f208; div.approx.ftz.f32 %f1, %f1623, %f207; $L__BB19_80: ld.const.v2.u32 {%r158, %r159}, [kernel_params+1064]; mov.u32 %r309, 0; setp.lt.s32 %p109, %r159, 1; ld.const.u64 %rd12, [kernel_params+2376]; @%p109 bra $L__BB19_83; cvta.to.global.u64 %rd13, %rd12; mov.u32 %r309, 0; mov.u32 %r307, %r159; $L__BB19_82: shr.s32 %r161, %r307, 1; add.s32 %r162, %r161, %r309; mul.wide.s32 %rd107, %r162, 8; add.s64 %rd108, %rd13, %rd107; ld.global.f32 %f1624, [%rd108+4]; setp.lt.ftz.f32 %p110, %f1624, %f932; add.s32 %r163, %r162, 1; not.b32 %r164, %r161; add.s32 %r165, %r307, %r164; selp.b32 %r309, %r163, %r309, %p110; selp.b32 %r307, %r165, %r161, %p110; setp.gt.s32 %p111, %r307, 0; @%p111 bra $L__BB19_82; $L__BB19_83: add.s32 %r167, %r309, -1; mov.u32 %r312, -1; max.s32 %r168, %r167, 0; cvta.to.global.u64 %rd109, %rd12; mul.wide.s32 %rd110, %r168, 8; add.s64 %rd111, %rd109, %rd110; ld.global.v2.f32 {%f1625, %f1626}, [%rd111]; mul.wide.s32 %rd112, %r159, 8; add.s64 %rd113, %rd109, %rd112; ld.global.f32 %f212, [%rd113]; sub.ftz.f32 %f1628, %f932, %f1626; ld.global.f32 %f1629, [%rd111+12]; sub.ftz.f32 %f1630, %f1629, %f1626; div.approx.ftz.f32 %f1631, %f1628, %f1630; cvt.rn.f32.s32 %f1632, %r168; add.ftz.f32 %f1633, %f1631, %f1632; cvt.rn.f32.s32 %f1634, %r159; div.approx.ftz.f32 %f213, %f1633, %f1634; add.s32 %r169, %r158, 1; mul.lo.s32 %r45, %r168, %r169; ld.const.u64 %rd14, [kernel_params+2384]; setp.lt.s32 %p112, %r158, 1; @%p112 bra $L__BB19_87; cvta.to.global.u64 %rd15, %rd14; mov.u32 %r311, 0; mov.u32 %r310, %r158; $L__BB19_85: shr.s32 %r171, %r310, 1; add.s32 %r172, %r171, %r311; add.s32 %r173, %r172, %r45; mul.wide.s32 %rd114, %r173, 8; add.s64 %rd115, %rd15, %rd114; ld.global.f32 %f1635, [%rd115+4]; setp.lt.ftz.f32 %p113, %f1635, %f1; add.s32 %r174, %r172, 1; not.b32 %r175, %r171; add.s32 %r176, %r310, %r175; selp.b32 %r311, %r174, %r311, %p113; selp.b32 %r310, %r176, %r171, %p113; setp.gt.s32 %p114, %r310, 0; @%p114 bra $L__BB19_85; add.s32 %r312, %r311, -1; $L__BB19_87: max.s32 %r177, %r312, 0; add.s32 %r178, %r177, %r45; cvta.to.global.u64 %rd116, %rd14; mul.wide.s32 %rd117, %r178, 8; add.s64 %rd118, %rd116, %rd117; ld.global.v2.f32 {%f1637, %f1638}, [%rd118]; add.s32 %r179, %r45, %r158; mul.wide.s32 %rd119, %r179, 8; add.s64 %rd120, %rd116, %rd119; sub.ftz.f32 %f1639, %f1, %f1638; ld.global.f32 %f1640, [%rd118+12]; sub.ftz.f32 %f1641, %f1640, %f1638; div.approx.ftz.f32 %f1642, %f1639, %f1641; cvt.rn.f32.s32 %f1643, %r177; add.ftz.f32 %f1644, %f1642, %f1643; cvt.rn.f32.s32 %f1645, %r158; div.approx.ftz.f32 %f216, %f1644, %f1645; mul.ftz.f32 %f217, %f213, 0f40490FDB; sin.approx.ftz.f32 %f1646, %f217; mul.ftz.f32 %f1647, %f1646, 0f419DE9E7; ld.global.f32 %f1648, [%rd120]; mul.ftz.f32 %f1649, %f1648, %f1647; mul.ftz.f32 %f218, %f212, %f1649; setp.eq.ftz.f32 %p115, %f1646, 0f00000000; mov.f32 %f3812, 0f00000000; setp.eq.ftz.f32 %p116, %f218, 0f00000000; or.pred %p117, %p115, %p116; @%p117 bra $L__BB19_89; mul.ftz.f32 %f1650, %f1625, %f1637; div.approx.ftz.f32 %f3812, %f1650, %f218; $L__BB19_89: st.local.f32 [%rd1+48], %f3812; fma.rn.ftz.f32 %f1651, %f216, 0fC0C90FDB, 0f40490FDB; mov.f32 %f1652, 0f40490FDB; sub.ftz.f32 %f1653, %f1652, %f217; sin.approx.ftz.f32 %f1654, %f1653; cos.approx.ftz.f32 %f1655, %f1651; sin.approx.ftz.f32 %f1656, %f1651; mul.ftz.f32 %f3831, %f1654, %f1656; mul.ftz.f32 %f3830, %f1654, %f1655; cos.approx.ftz.f32 %f3832, %f1653; mov.u32 %r322, 2; mov.f32 %f3816, %f207; @%p108 bra $L__BB19_171; bra.uni $L__BB19_135; $L__BB19_15: ld.global.f32 %f3761, [%rd7+152]; setp.leu.ftz.f32 %p21, %f3761, 0f00000000; mov.f32 %f3746, %f7; mov.f32 %f3747, %f6; mov.f32 %f3748, %f5; mov.f32 %f3749, %f948; mov.f32 %f3750, %f949; mov.f32 %f3751, %f950; mov.f32 %f3752, %f944; mov.f32 %f3753, %f945; mov.f32 %f3754, %f946; @%p21 bra $L__BB19_19; sub.ftz.f32 %f962, %f934, %f5; sub.ftz.f32 %f963, %f935, %f6; mul.ftz.f32 %f964, %f953, %f963; fma.rn.ftz.f32 %f965, %f952, %f962, %f964; sub.ftz.f32 %f966, %f936, %f7; fma.rn.ftz.f32 %f967, %f954, %f966, %f965; mul.ftz.f32 %f968, %f952, %f967; mul.ftz.f32 %f969, %f953, %f967; mul.ftz.f32 %f970, %f954, %f967; sub.ftz.f32 %f971, %f934, %f968; sub.ftz.f32 %f972, %f935, %f969; sub.ftz.f32 %f973, %f936, %f970; sub.ftz.f32 %f974, %f971, %f934; sub.ftz.f32 %f975, %f972, %f935; sub.ftz.f32 %f976, %f973, %f936; mul.ftz.f32 %f977, %f975, %f975; fma.rn.ftz.f32 %f978, %f974, %f974, %f977; fma.rn.ftz.f32 %f979, %f976, %f976, %f978; sqrt.approx.ftz.f32 %f980, %f979; div.approx.ftz.f32 %f981, %f980, %f3761; mul.ftz.f32 %f982, %f945, %f945; fma.rn.ftz.f32 %f983, %f944, %f944, %f982; fma.rn.ftz.f32 %f984, %f946, %f946, %f983; sqrt.approx.ftz.f32 %f985, %f984; rsqrt.approx.ftz.f32 %f986, %f984; mul.ftz.f32 %f20, %f944, %f986; mul.ftz.f32 %f21, %f945, %f986; mul.ftz.f32 %f22, %f946, %f986; mul.ftz.f32 %f987, %f949, %f949; fma.rn.ftz.f32 %f988, %f948, %f948, %f987; fma.rn.ftz.f32 %f989, %f950, %f950, %f988; sqrt.approx.ftz.f32 %f990, %f989; rsqrt.approx.ftz.f32 %f991, %f989; mul.ftz.f32 %f23, %f948, %f991; mul.ftz.f32 %f24, %f949, %f991; mul.ftz.f32 %f25, %f950, %f991; sub.ftz.f32 %f992, %f971, %f5; sub.ftz.f32 %f993, %f972, %f6; sub.ftz.f32 %f994, %f973, %f7; mul.ftz.f32 %f995, %f21, %f993; fma.rn.ftz.f32 %f996, %f20, %f992, %f995; fma.rn.ftz.f32 %f997, %f22, %f994, %f996; mul.ftz.f32 %f998, %f24, %f993; fma.rn.ftz.f32 %f999, %f23, %f992, %f998; fma.rn.ftz.f32 %f1000, %f25, %f994, %f999; sub.ftz.f32 %f1001, %f997, %f981; mul.ftz.f32 %f1002, %f985, 0fBF000000; max.ftz.f32 %f26, %f1001, %f1002; add.ftz.f32 %f1003, %f981, %f997; mul.ftz.f32 %f1004, %f985, 0f3F000000; min.ftz.f32 %f27, %f1003, %f1004; sub.ftz.f32 %f1005, %f1000, %f981; mul.ftz.f32 %f1006, %f990, 0fBF000000; max.ftz.f32 %f28, %f1005, %f1006; add.ftz.f32 %f1007, %f981, %f1000; mul.ftz.f32 %f1008, %f990, 0f3F000000; min.ftz.f32 %f29, %f1007, %f1008; setp.ge.ftz.f32 %p23, %f26, %f27; setp.ge.ftz.f32 %p24, %f28, %f29; or.pred %p25, %p23, %p24; mov.pred %p309, -1; mov.f32 %f3746, %f7; mov.f32 %f3747, %f6; mov.f32 %f3748, %f5; mov.f32 %f3749, %f948; mov.f32 %f3750, %f949; mov.f32 %f3751, %f950; mov.f32 %f3752, %f944; mov.f32 %f3753, %f945; mov.f32 %f3754, %f946; @%p25 bra $L__BB19_18; add.ftz.f32 %f1009, %f26, %f27; mul.ftz.f32 %f1010, %f1009, 0f3F000000; add.ftz.f32 %f1011, %f28, %f29; mul.ftz.f32 %f1012, %f1011, 0f3F000000; fma.rn.ftz.f32 %f1013, %f20, %f1010, %f5; fma.rn.ftz.f32 %f1014, %f21, %f1010, %f6; fma.rn.ftz.f32 %f1015, %f22, %f1010, %f7; fma.rn.ftz.f32 %f3748, %f23, %f1012, %f1013; fma.rn.ftz.f32 %f3747, %f24, %f1012, %f1014; fma.rn.ftz.f32 %f3746, %f25, %f1012, %f1015; st.local.f32 [%rd1], %f3748; st.local.f32 [%rd1+4], %f3747; st.local.f32 [%rd1+8], %f3746; sub.ftz.f32 %f1016, %f27, %f26; mul.ftz.f32 %f3752, %f20, %f1016; mul.ftz.f32 %f3753, %f21, %f1016; mul.ftz.f32 %f3754, %f22, %f1016; sub.ftz.f32 %f1017, %f29, %f28; mul.ftz.f32 %f3749, %f23, %f1017; mul.ftz.f32 %f3750, %f24, %f1017; mul.ftz.f32 %f3751, %f25, %f1017; mov.pred %p309, 0; $L__BB19_18: @%p309 bra $L__BB19_264; $L__BB19_19: mul.ftz.f32 %f1018, %f3752, 0f3F000000; sub.ftz.f32 %f1019, %f3748, %f1018; mul.ftz.f32 %f1020, %f3753, 0f3F000000; sub.ftz.f32 %f1021, %f3747, %f1020; mul.ftz.f32 %f1022, %f3754, 0f3F000000; sub.ftz.f32 %f1023, %f3746, %f1022; mul.ftz.f32 %f1024, %f3749, 0f3F000000; sub.ftz.f32 %f1025, %f1019, %f1024; mul.ftz.f32 %f1026, %f3750, 0f3F000000; sub.ftz.f32 %f1027, %f1021, %f1026; mul.ftz.f32 %f1028, %f3751, 0f3F000000; sub.ftz.f32 %f1029, %f1023, %f1028; mul.ftz.f32 %f1030, %f3753, %f3753; fma.rn.ftz.f32 %f1031, %f3752, %f3752, %f1030; fma.rn.ftz.f32 %f1032, %f3754, %f3754, %f1031; sqrt.approx.ftz.f32 %f57, %f1032; rsqrt.approx.ftz.f32 %f1033, %f1032; mul.ftz.f32 %f58, %f3752, %f1033; mul.ftz.f32 %f59, %f3753, %f1033; mul.ftz.f32 %f60, %f3754, %f1033; mul.ftz.f32 %f1034, %f3750, %f3750; fma.rn.ftz.f32 %f1035, %f3749, %f3749, %f1034; fma.rn.ftz.f32 %f1036, %f3751, %f3751, %f1035; sqrt.approx.ftz.f32 %f61, %f1036; rsqrt.approx.ftz.f32 %f1037, %f1036; mul.ftz.f32 %f62, %f3749, %f1037; mul.ftz.f32 %f63, %f3750, %f1037; mul.ftz.f32 %f64, %f3751, %f1037; mul.ftz.f32 %f1038, %f59, %f64; mul.ftz.f32 %f1039, %f60, %f63; sub.ftz.f32 %f3755, %f1038, %f1039; mul.ftz.f32 %f1040, %f60, %f62; mul.ftz.f32 %f1041, %f58, %f64; sub.ftz.f32 %f3756, %f1040, %f1041; mul.ftz.f32 %f1042, %f58, %f63; mul.ftz.f32 %f1043, %f59, %f62; sub.ftz.f32 %f3757, %f1042, %f1043; sub.ftz.f32 %f68, %f1025, %f934; sub.ftz.f32 %f69, %f1027, %f935; sub.ftz.f32 %f70, %f1029, %f936; mul.ftz.f32 %f1044, %f69, %f3756; fma.rn.ftz.f32 %f1045, %f68, %f3755, %f1044; fma.rn.ftz.f32 %f3758, %f70, %f3757, %f1045; setp.leu.ftz.f32 %p28, %f3758, 0f00000000; @%p28 bra $L__BB19_21; neg.ftz.f32 %f3755, %f3755; neg.ftz.f32 %f3756, %f3756; neg.ftz.f32 %f3757, %f3757; neg.ftz.f32 %f3758, %f3758; $L__BB19_21: mul.ftz.f32 %f1046, %f69, %f59; fma.rn.ftz.f32 %f1047, %f68, %f58, %f1046; fma.rn.ftz.f32 %f1048, %f70, %f60, %f1047; mul.ftz.f32 %f1049, %f69, %f63; fma.rn.ftz.f32 %f1050, %f68, %f62, %f1049; fma.rn.ftz.f32 %f1051, %f70, %f64, %f1050; add.ftz.f32 %f1052, %f57, %f1048; add.ftz.f32 %f3759, %f61, %f1051; sub.ftz.f32 %f1053, %f1048, %f1052; sub.ftz.f32 %f1054, %f3759, %f1051; sub.ftz.f32 %f1055, %f1052, %f1048; sub.ftz.f32 %f1056, %f1051, %f3759; mul.ftz.f32 %f1057, %f1051, %f1053; mul.ftz.f32 %f1058, %f1052, %f1054; mul.ftz.f32 %f1059, %f3759, %f1055; mul.ftz.f32 %f1060, %f1048, %f1056; mul.ftz.f32 %f1061, %f3758, %f3758; mul.ftz.f32 %f1062, %f1053, %f1061; mul.ftz.f32 %f1063, %f1054, %f1061; mul.ftz.f32 %f1064, %f1055, %f1061; mul.ftz.f32 %f1065, %f1056, %f1061; mul.ftz.f32 %f1066, %f1053, %f1062; mul.ftz.f32 %f1067, %f1054, %f1063; mul.ftz.f32 %f1068, %f1055, %f1064; mul.ftz.f32 %f1069, %f1056, %f1065; fma.rn.ftz.f32 %f1070, %f1057, %f1057, %f1066; fma.rn.ftz.f32 %f1071, %f1058, %f1058, %f1067; fma.rn.ftz.f32 %f1072, %f1059, %f1059, %f1068; fma.rn.ftz.f32 %f1073, %f1060, %f1060, %f1069; sqrt.approx.ftz.f32 %f1074, %f1070; sqrt.approx.ftz.f32 %f1075, %f1071; sqrt.approx.ftz.f32 %f1076, %f1072; sqrt.approx.ftz.f32 %f1077, %f1073; div.approx.ftz.f32 %f1078, %f1057, %f1074; div.approx.ftz.f32 %f1079, %f1058, %f1075; div.approx.ftz.f32 %f1080, %f1059, %f1076; div.approx.ftz.f32 %f1081, %f1060, %f1077; mul.ftz.f32 %f1082, %f1078, %f1079; neg.ftz.f32 %f1083, %f1082; mov.f32 %f1084, 0fBF800000; max.ftz.f32 %f1085, %f1083, %f1084; mov.f32 %f1086, 0f3F800000; min.ftz.f32 %f1087, %f1085, %f1086; abs.ftz.f32 %f1088, %f1087; sub.ftz.f32 %f1089, %f1086, %f1088; mul.ftz.f32 %f1090, %f1089, 0f3F000000; sqrt.approx.ftz.f32 %f1091, %f1090; setp.gt.ftz.f32 %p29, %f1088, 0f3F11EB85; selp.f32 %f1092, %f1091, %f1088, %p29; mul.ftz.f32 %f1093, %f1092, %f1092; mov.f32 %f1094, 0f3C94D2E9; mov.f32 %f1095, 0f3D53F941; fma.rn.ftz.f32 %f1096, %f1095, %f1093, %f1094; mov.f32 %f1097, 0f3D3F841F; fma.rn.ftz.f32 %f1098, %f1096, %f1093, %f1097; mov.f32 %f1099, 0f3D994929; fma.rn.ftz.f32 %f1100, %f1098, %f1093, %f1099; mov.f32 %f1101, 0f3E2AAB94; fma.rn.ftz.f32 %f1102, %f1100, %f1093, %f1101; mul.ftz.f32 %f1103, %f1093, %f1102; fma.rn.ftz.f32 %f1104, %f1103, %f1092, %f1092; add.ftz.f32 %f1105, %f1104, %f1104; mov.f32 %f1106, 0f3FC90FDB; sub.ftz.f32 %f1107, %f1106, %f1104; selp.f32 %f1108, %f1105, %f1107, %p29; setp.lt.ftz.f32 %p30, %f1087, 0f00000000; mov.f32 %f1109, 0f40490FDB; sub.ftz.f32 %f1110, %f1109, %f1108; selp.f32 %f1111, %f1110, %f1108, %p30; mul.ftz.f32 %f1112, %f1079, %f1080; neg.ftz.f32 %f1113, %f1112; max.ftz.f32 %f1114, %f1113, %f1084; min.ftz.f32 %f1115, %f1114, %f1086; abs.ftz.f32 %f1116, %f1115; sub.ftz.f32 %f1117, %f1086, %f1116; mul.ftz.f32 %f1118, %f1117, 0f3F000000; sqrt.approx.ftz.f32 %f1119, %f1118; setp.gt.ftz.f32 %p31, %f1116, 0f3F11EB85; selp.f32 %f1120, %f1119, %f1116, %p31; mul.ftz.f32 %f1121, %f1120, %f1120; fma.rn.ftz.f32 %f1122, %f1095, %f1121, %f1094; fma.rn.ftz.f32 %f1123, %f1122, %f1121, %f1097; fma.rn.ftz.f32 %f1124, %f1123, %f1121, %f1099; fma.rn.ftz.f32 %f1125, %f1124, %f1121, %f1101; mul.ftz.f32 %f1126, %f1121, %f1125; fma.rn.ftz.f32 %f1127, %f1126, %f1120, %f1120; add.ftz.f32 %f1128, %f1127, %f1127; sub.ftz.f32 %f1129, %f1106, %f1127; selp.f32 %f1130, %f1128, %f1129, %p31; setp.lt.ftz.f32 %p32, %f1115, 0f00000000; sub.ftz.f32 %f1131, %f1109, %f1130; selp.f32 %f1132, %f1131, %f1130, %p32; mul.ftz.f32 %f1133, %f1080, %f1081; neg.ftz.f32 %f1134, %f1133; max.ftz.f32 %f1135, %f1134, %f1084; min.ftz.f32 %f1136, %f1135, %f1086; abs.ftz.f32 %f1137, %f1136; sub.ftz.f32 %f1138, %f1086, %f1137; mul.ftz.f32 %f1139, %f1138, 0f3F000000; sqrt.approx.ftz.f32 %f1140, %f1139; setp.gt.ftz.f32 %p33, %f1137, 0f3F11EB85; selp.f32 %f1141, %f1140, %f1137, %p33; mul.ftz.f32 %f1142, %f1141, %f1141; fma.rn.ftz.f32 %f1143, %f1095, %f1142, %f1094; fma.rn.ftz.f32 %f1144, %f1143, %f1142, %f1097; fma.rn.ftz.f32 %f1145, %f1144, %f1142, %f1099; fma.rn.ftz.f32 %f1146, %f1145, %f1142, %f1101; mul.ftz.f32 %f1147, %f1142, %f1146; fma.rn.ftz.f32 %f1148, %f1147, %f1141, %f1141; add.ftz.f32 %f1149, %f1148, %f1148; sub.ftz.f32 %f1150, %f1106, %f1148; selp.f32 %f1151, %f1149, %f1150, %p33; setp.lt.ftz.f32 %p34, %f1136, 0f00000000; sub.ftz.f32 %f1152, %f1109, %f1151; selp.f32 %f1153, %f1152, %f1151, %p34; mul.ftz.f32 %f1154, %f1078, %f1081; neg.ftz.f32 %f1155, %f1154; max.ftz.f32 %f1156, %f1155, %f1084; min.ftz.f32 %f1157, %f1156, %f1086; abs.ftz.f32 %f1158, %f1157; sub.ftz.f32 %f1159, %f1086, %f1158; mul.ftz.f32 %f1160, %f1159, 0f3F000000; sqrt.approx.ftz.f32 %f1161, %f1160; setp.gt.ftz.f32 %p35, %f1158, 0f3F11EB85; selp.f32 %f1162, %f1161, %f1158, %p35; mul.ftz.f32 %f1163, %f1162, %f1162; fma.rn.ftz.f32 %f1164, %f1095, %f1163, %f1094; fma.rn.ftz.f32 %f1165, %f1164, %f1163, %f1097; fma.rn.ftz.f32 %f1166, %f1165, %f1163, %f1099; fma.rn.ftz.f32 %f1167, %f1166, %f1163, %f1101; mul.ftz.f32 %f1168, %f1163, %f1167; fma.rn.ftz.f32 %f1169, %f1168, %f1162, %f1162; add.ftz.f32 %f1170, %f1169, %f1169; sub.ftz.f32 %f1171, %f1106, %f1169; selp.f32 %f1172, %f1170, %f1171, %p35; setp.lt.ftz.f32 %p36, %f1157, 0f00000000; sub.ftz.f32 %f1173, %f1109, %f1172; selp.f32 %f1174, %f1173, %f1172, %p36; mov.f32 %f1175, 0f40C90FDB; sub.ftz.f32 %f1176, %f1175, %f1153; sub.ftz.f32 %f1177, %f1176, %f1174; add.ftz.f32 %f1178, %f1111, %f1132; sub.ftz.f32 %f81, %f1178, %f1177; fma.rn.ftz.f32 %f1179, %f1, %f81, %f1177; cos.approx.ftz.f32 %f1180, %f1179; mul.ftz.f32 %f1181, %f1078, %f1180; sub.ftz.f32 %f1182, %f1181, %f1080; sin.approx.ftz.f32 %f1183, %f1179; div.approx.ftz.f32 %f1184, %f1182, %f1183; mul.ftz.f32 %f1185, %f1184, %f1184; fma.rn.ftz.f32 %f1186, %f1078, %f1078, %f1185; rsqrt.approx.ftz.f32 %f1187, %f1186; setp.gt.ftz.f32 %p37, %f1184, 0f00000000; selp.f32 %f1188, 0f3F800000, 0fBF800000, %p37; mul.ftz.f32 %f1189, %f1187, %f1188; max.ftz.f32 %f1190, %f1189, %f1084; min.ftz.f32 %f1191, %f1190, %f1086; mul.ftz.f32 %f1192, %f3758, %f1191; neg.ftz.f32 %f1193, %f1192; mul.ftz.f32 %f1194, %f1191, %f1191; sub.ftz.f32 %f1195, %f1086, %f1194; sqrt.approx.ftz.f32 %f1196, %f1195; mov.f32 %f1197, 0f33D6BF95; max.ftz.f32 %f1198, %f1196, %f1197; div.approx.ftz.f32 %f1199, %f1193, %f1198; max.ftz.f32 %f1200, %f1199, %f1048; min.ftz.f32 %f82, %f1200, %f1052; fma.rn.ftz.f32 %f1201, %f82, %f82, %f1061; sqrt.approx.ftz.f32 %f83, %f1201; mul.ftz.f32 %f1202, %f83, %f83; fma.rn.ftz.f32 %f1203, %f1051, %f1051, %f1202; sqrt.approx.ftz.f32 %f1204, %f1203; div.approx.ftz.f32 %f1205, %f1051, %f1204; fma.rn.ftz.f32 %f1206, %f3759, %f3759, %f1202; sqrt.approx.ftz.f32 %f1207, %f1206; div.approx.ftz.f32 %f1208, %f3759, %f1207; sub.ftz.f32 %f1209, %f1208, %f1205; fma.rn.ftz.f32 %f84, %f1209, %f932, %f1205; mul.ftz.f32 %f85, %f84, %f84; setp.geu.ftz.f32 %p38, %f85, 0f3F7FFFEF; @%p38 bra $L__BB19_23; mul.ftz.f32 %f1210, %f83, %f84; sub.ftz.f32 %f1212, %f1086, %f85; sqrt.approx.ftz.f32 %f1213, %f1212; div.approx.ftz.f32 %f3759, %f1210, %f1213; $L__BB19_23: ld.param.f32 %f3732, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+8]; ld.param.f32 %f3731, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+4]; ld.param.f32 %f3730, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3]; fma.rn.ftz.f32 %f1215, %f58, %f82, %f3730; fma.rn.ftz.f32 %f1216, %f59, %f82, %f3731; fma.rn.ftz.f32 %f1217, %f60, %f82, %f3732; fma.rn.ftz.f32 %f1218, %f62, %f3759, %f1215; fma.rn.ftz.f32 %f1219, %f63, %f3759, %f1216; fma.rn.ftz.f32 %f1220, %f64, %f3759, %f1217; fma.rn.ftz.f32 %f3765, %f3755, %f3758, %f1218; fma.rn.ftz.f32 %f3764, %f3756, %f3758, %f1219; fma.rn.ftz.f32 %f3763, %f3757, %f3758, %f1220; st.local.f32 [%rd1], %f3765; st.local.f32 [%rd1+4], %f3764; st.local.f32 [%rd1+8], %f3763; setp.eq.ftz.f32 %p39, %f81, 0f00000000; mov.f32 %f3836, 0f00000000; @%p39 bra $L__BB19_25; rcp.approx.ftz.f32 %f3836, %f81; $L__BB19_25: st.local.f32 [%rd1+48], %f3836; sub.ftz.f32 %f3766, %f3765, %f5; sub.ftz.f32 %f3767, %f3764, %f6; sub.ftz.f32 %f3768, %f3763, %f7; $L__BB19_27: abs.ftz.f32 %f3736, %f947; ld.param.f32 %f3735, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+8]; ld.param.f32 %f3734, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3+4]; ld.param.f32 %f3733, [_Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample_param_3]; mul.ftz.f32 %f1236, %f945, %f3767; fma.rn.ftz.f32 %f1237, %f944, %f3766, %f1236; fma.rn.ftz.f32 %f1238, %f946, %f3768, %f1237; mul.ftz.f32 %f1239, %f945, %f945; fma.rn.ftz.f32 %f1240, %f944, %f944, %f1239; fma.rn.ftz.f32 %f1241, %f946, %f946, %f1240; rcp.approx.ftz.f32 %f1242, %f1241; mul.ftz.f32 %f1243, %f1242, %f1238; mul.ftz.f32 %f1244, %f949, %f3767; fma.rn.ftz.f32 %f1245, %f948, %f3766, %f1244; fma.rn.ftz.f32 %f1246, %f950, %f3768, %f1245; mul.ftz.f32 %f1247, %f949, %f949; fma.rn.ftz.f32 %f1248, %f948, %f948, %f1247; fma.rn.ftz.f32 %f1249, %f950, %f950, %f1248; rcp.approx.ftz.f32 %f1250, %f1249; mul.ftz.f32 %f1251, %f1250, %f1246; add.ftz.f32 %f1252, %f1251, 0f3F000000; st.local.f32 [%rd1+40], %f1252; neg.ftz.f32 %f1253, %f1243; sub.ftz.f32 %f1254, %f1253, %f1251; st.local.f32 [%rd1+44], %f1254; st.local.f32 [%rd1+12], %f952; st.local.f32 [%rd1+16], %f953; st.local.f32 [%rd1+20], %f954; sub.ftz.f32 %f1255, %f3765, %f3733; sub.ftz.f32 %f1256, %f3764, %f3734; mul.ftz.f32 %f1257, %f1256, %f1256; fma.rn.ftz.f32 %f1258, %f1255, %f1255, %f1257; sub.ftz.f32 %f1259, %f3763, %f3735; fma.rn.ftz.f32 %f1260, %f1259, %f1259, %f1258; sqrt.approx.ftz.f32 %f111, %f1260; st.local.f32 [%rd1+36], %f111; rsqrt.approx.ftz.f32 %f1261, %f1260; mul.ftz.f32 %f112, %f1255, %f1261; mul.ftz.f32 %f113, %f1256, %f1261; mul.ftz.f32 %f114, %f1259, %f1261; st.local.f32 [%rd1+24], %f112; st.local.f32 [%rd1+28], %f113; st.local.f32 [%rd1+32], %f114; mul.ftz.f32 %f115, %f3736, 0f3E800000; st.local.f32 [%rd1+52], %f115; setp.leu.ftz.f32 %p40, %f3761, 0f00000000; @%p40 bra $L__BB19_29; mul.ftz.f32 %f1262, %f113, %f953; fma.rn.ftz.f32 %f1263, %f112, %f952, %f1262; fma.rn.ftz.f32 %f1264, %f114, %f954, %f1263; neg.ftz.f32 %f1265, %f1264; mul.ftz.f32 %f1266, %f1264, %f1264; mov.f32 %f1267, 0f3F800000; sub.ftz.f32 %f1268, %f1267, %f1266; mov.f32 %f1269, 0f00000000; max.ftz.f32 %f1270, %f1268, %f1269; sqrt.approx.ftz.f32 %f1271, %f1270; div.approx.ftz.f32 %f1272, %f1271, %f1265; mul.ftz.f32 %f1273, %f3761, %f1272; sub.ftz.f32 %f1274, %f1267, %f1273; ld.global.f32 %f1275, [%rd7+168]; mul.ftz.f32 %f1276, %f1275, %f1274; max.ftz.f32 %f1277, %f1276, %f1269; mul.ftz.f32 %f1278, %f1277, %f115; st.local.f32 [%rd1+52], %f1278; $L__BB19_29: setp.lt.ftz.f32 %p308, %f947, 0f00000000; @%p308 bra $L__BB19_30; bra.uni $L__BB19_175; $L__BB19_30: mul.ftz.f32 %f1280, %f952, %f112; mul.ftz.f32 %f1281, %f953, %f113; neg.ftz.f32 %f1282, %f1281; sub.ftz.f32 %f1283, %f1282, %f1280; mul.ftz.f32 %f1284, %f954, %f114; sub.ftz.f32 %f116, %f1283, %f1284; setp.le.ftz.f32 %p42, %f116, 0f00000000; mov.f32 %f3769, 0f00000000; @%p42 bra $L__BB19_32; mul.ftz.f32 %f1285, %f111, %f111; div.approx.ftz.f32 %f3769, %f1285, %f116; $L__BB19_32: mul.ftz.f32 %f3836, %f3769, %f3836; st.local.f32 [%rd1+48], %f3836; bra.uni $L__BB19_175; $L__BB19_101: ld.global.v4.f32 {%f1727, %f1728, %f1729, %f1730}, [%rd16+140]; ld.global.v4.f32 {%f1731, %f1732, %f1733, %f1734}, [%rd16+156]; setp.lt.ftz.f32 %p128, %f1730, 0f00000000; mul.ftz.f32 %f247, %f1727, 0f3F000000; mul.ftz.f32 %f248, %f1728, 0f3F000000; mul.ftz.f32 %f249, %f1729, 0f3F000000; mul.ftz.f32 %f250, %f1731, 0f3F000000; mul.ftz.f32 %f251, %f1732, 0f3F000000; mul.ftz.f32 %f252, %f1733, 0f3F000000; @%p128 bra $L__BB19_109; bra.uni $L__BB19_102; $L__BB19_109: mul.ftz.f32 %f1940, %f1, 0f40C90FDB; cos.approx.ftz.f32 %f1941, %f1940; sqrt.approx.ftz.f32 %f1942, %f233; mul.ftz.f32 %f1943, %f1942, %f1941; sin.approx.ftz.f32 %f1944, %f1940; mul.ftz.f32 %f1945, %f1942, %f1944; mul.ftz.f32 %f1946, %f250, %f1945; mul.ftz.f32 %f1947, %f251, %f1945; mul.ftz.f32 %f1948, %f252, %f1945; fma.rn.ftz.f32 %f1949, %f247, %f1943, %f1946; fma.rn.ftz.f32 %f1950, %f248, %f1943, %f1947; fma.rn.ftz.f32 %f1951, %f249, %f1943, %f1948; add.ftz.f32 %f1952, %f237, %f1949; add.ftz.f32 %f1953, %f238, %f1950; add.ftz.f32 %f1954, %f239, %f1951; sub.ftz.f32 %f1955, %f1952, %f934; sub.ftz.f32 %f1956, %f1953, %f935; sub.ftz.f32 %f1957, %f1954, %f936; mul.ftz.f32 %f1958, %f1956, %f1956; fma.rn.ftz.f32 %f1959, %f1955, %f1955, %f1958; fma.rn.ftz.f32 %f1960, %f1957, %f1957, %f1959; sqrt.approx.ftz.f32 %f292, %f1960; rsqrt.approx.ftz.f32 %f1961, %f1960; mul.ftz.f32 %f3830, %f1961, %f1955; mul.ftz.f32 %f3831, %f1961, %f1956; mul.ftz.f32 %f3832, %f1961, %f1957; mul.ftz.f32 %f1962, %f1717, %f3830; mul.ftz.f32 %f1963, %f1718, %f3831; neg.ftz.f32 %f1964, %f1963; sub.ftz.f32 %f1965, %f1964, %f1962; mul.ftz.f32 %f1966, %f1719, %f3832; sub.ftz.f32 %f296, %f1965, %f1966; setp.le.ftz.f32 %p141, %f296, 0f00000000; mov.f32 %f3798, 0f00000000; @%p141 bra $L__BB19_111; mul.ftz.f32 %f1967, %f292, %f292; div.approx.ftz.f32 %f3798, %f1967, %f296; $L__BB19_111: abs.ftz.f32 %f1968, %f1730; mul.ftz.f32 %f3799, %f1968, %f3798; st.local.f32 [%rd1+48], %f3799; bra.uni $L__BB19_112; $L__BB19_253: ld.const.u64 %rd239, [kernel_params+2176]; cvta.to.global.u64 %rd240, %rd239; ld.global.u32 %rd241, [%rd94+168]; ld.global.u32 %r284, [%rd94+152]; shl.b32 %r285, %r284, 1; cvt.rn.f32.s32 %f3472, %r285; mul.ftz.f32 %f3473, %f3472, 0f3F000000; cvt.rzi.ftz.s32.f32 %r286, %f3473; add.s32 %r287, %r285, -1; min.s32 %r288, %r286, %r287; cvt.rn.f32.s32 %f3474, %r288; sub.ftz.f32 %f850, %f3473, %f3474; cvt.s64.s32 %rd242, %r288; add.s64 %rd243, %rd242, %rd241; shl.b64 %rd244, %rd243, 6; add.s64 %rd95, %rd240, %rd244; ld.global.v4.f32 {%f3475, %f3476, %f3477, %f3478}, [%rd95]; ld.global.v4.f32 {%f3479, %f3480, %f3481, %f3482}, [%rd95+64]; mul.ftz.f32 %f3483, %f3476, %f3480; fma.rn.ftz.f32 %f3484, %f3475, %f3479, %f3483; mul.ftz.f32 %f3485, %f3478, %f3482; fma.rn.ftz.f32 %f3486, %f3477, %f3481, %f3485; add.ftz.f32 %f859, %f3484, %f3486; setp.gt.ftz.f32 %p303, %f859, 0f3F7FDF3B; sub.ftz.f32 %f860, %f3384, %f850; @%p303 bra $L__BB19_255; bra.uni $L__BB19_254; $L__BB19_255: mul.ftz.f32 %f3541, %f3475, %f860; fma.rn.ftz.f32 %f3542, %f3479, %f850, %f3541; mul.ftz.f32 %f3543, %f3476, %f860; fma.rn.ftz.f32 %f3544, %f3480, %f850, %f3543; mul.ftz.f32 %f3545, %f3477, %f860; fma.rn.ftz.f32 %f3546, %f3481, %f850, %f3545; mul.ftz.f32 %f3547, %f3478, %f860; fma.rn.ftz.f32 %f3548, %f3482, %f850, %f3547; mul.ftz.f32 %f3549, %f3544, %f3544; fma.rn.ftz.f32 %f3550, %f3542, %f3542, %f3549; mul.ftz.f32 %f3551, %f3548, %f3548; fma.rn.ftz.f32 %f3552, %f3546, %f3546, %f3551; add.ftz.f32 %f3553, %f3550, %f3552; rsqrt.approx.ftz.f32 %f3554, %f3553; mul.ftz.f32 %f3938, %f3548, %f3554; mul.ftz.f32 %f3937, %f3546, %f3554; mul.ftz.f32 %f3936, %f3544, %f3554; mul.ftz.f32 %f3935, %f3542, %f3554; bra.uni $L__BB19_256; $L__BB19_57: setp.eq.ftz.f32 %p83, %f184, 0f7F800000; setp.eq.ftz.f32 %p84, %f185, 0f7F800000; and.pred %p85, %p83, %p84; @%p85 bra $L__BB19_59; bra.uni $L__BB19_58; $L__BB19_59: setp.lt.s32 %p89, %r17, 0; selp.b32 %r134, 1075235812, 1061752795, %p89; or.b32 %r135, %r134, %r18; mov.b32 %f3782, %r135; bra.uni $L__BB19_61; $L__BB19_102: sub.ftz.f32 %f1735, %f237, %f247; sub.ftz.f32 %f1736, %f1735, %f250; sub.ftz.f32 %f1737, %f238, %f248; sub.ftz.f32 %f1738, %f1737, %f251; sub.ftz.f32 %f1739, %f239, %f249; sub.ftz.f32 %f1740, %f1739, %f252; mul.ftz.f32 %f1741, %f1728, %f1728; fma.rn.ftz.f32 %f1742, %f1727, %f1727, %f1741; fma.rn.ftz.f32 %f1743, %f1729, %f1729, %f1742; sqrt.approx.ftz.f32 %f253, %f1743; rsqrt.approx.ftz.f32 %f1744, %f1743; mul.ftz.f32 %f254, %f1727, %f1744; mul.ftz.f32 %f255, %f1728, %f1744; mul.ftz.f32 %f256, %f1729, %f1744; mul.ftz.f32 %f1745, %f1732, %f1732; fma.rn.ftz.f32 %f1746, %f1731, %f1731, %f1745; fma.rn.ftz.f32 %f1747, %f1733, %f1733, %f1746; sqrt.approx.ftz.f32 %f257, %f1747; rsqrt.approx.ftz.f32 %f1748, %f1747; mul.ftz.f32 %f258, %f1731, %f1748; mul.ftz.f32 %f259, %f1732, %f1748; mul.ftz.f32 %f260, %f1733, %f1748; mul.ftz.f32 %f1749, %f255, %f260; mul.ftz.f32 %f1750, %f256, %f259; sub.ftz.f32 %f3792, %f1749, %f1750; mul.ftz.f32 %f1751, %f256, %f258; mul.ftz.f32 %f1752, %f254, %f260; sub.ftz.f32 %f3793, %f1751, %f1752; mul.ftz.f32 %f1753, %f254, %f259; mul.ftz.f32 %f1754, %f255, %f258; sub.ftz.f32 %f3794, %f1753, %f1754; sub.ftz.f32 %f264, %f1736, %f934; sub.ftz.f32 %f265, %f1738, %f935; sub.ftz.f32 %f266, %f1740, %f936; mul.ftz.f32 %f1755, %f265, %f3793; fma.rn.ftz.f32 %f1756, %f264, %f3792, %f1755; fma.rn.ftz.f32 %f3795, %f266, %f3794, %f1756; setp.leu.ftz.f32 %p129, %f3795, 0f00000000; @%p129 bra $L__BB19_104; neg.ftz.f32 %f3792, %f3792; neg.ftz.f32 %f3793, %f3793; neg.ftz.f32 %f3794, %f3794; neg.ftz.f32 %f3795, %f3795; $L__BB19_104: mul.ftz.f32 %f1757, %f265, %f255; fma.rn.ftz.f32 %f1758, %f264, %f254, %f1757; fma.rn.ftz.f32 %f1759, %f266, %f256, %f1758; mul.ftz.f32 %f1760, %f265, %f259; fma.rn.ftz.f32 %f1761, %f264, %f258, %f1760; fma.rn.ftz.f32 %f1762, %f266, %f260, %f1761; add.ftz.f32 %f1763, %f253, %f1759; add.ftz.f32 %f3796, %f257, %f1762; sub.ftz.f32 %f1764, %f1759, %f1763; sub.ftz.f32 %f1765, %f3796, %f1762; sub.ftz.f32 %f1766, %f1763, %f1759; sub.ftz.f32 %f1767, %f1762, %f3796; mul.ftz.f32 %f1768, %f1762, %f1764; mul.ftz.f32 %f1769, %f1763, %f1765; mul.ftz.f32 %f1770, %f3796, %f1766; mul.ftz.f32 %f1771, %f1759, %f1767; mul.ftz.f32 %f1772, %f3795, %f3795; mul.ftz.f32 %f1773, %f1764, %f1772; mul.ftz.f32 %f1774, %f1765, %f1772; mul.ftz.f32 %f1775, %f1766, %f1772; mul.ftz.f32 %f1776, %f1767, %f1772; mul.ftz.f32 %f1777, %f1764, %f1773; mul.ftz.f32 %f1778, %f1765, %f1774; mul.ftz.f32 %f1779, %f1766, %f1775; mul.ftz.f32 %f1780, %f1767, %f1776; fma.rn.ftz.f32 %f1781, %f1768, %f1768, %f1777; fma.rn.ftz.f32 %f1782, %f1769, %f1769, %f1778; fma.rn.ftz.f32 %f1783, %f1770, %f1770, %f1779; fma.rn.ftz.f32 %f1784, %f1771, %f1771, %f1780; sqrt.approx.ftz.f32 %f1785, %f1781; sqrt.approx.ftz.f32 %f1786, %f1782; sqrt.approx.ftz.f32 %f1787, %f1783; sqrt.approx.ftz.f32 %f1788, %f1784; div.approx.ftz.f32 %f1789, %f1768, %f1785; div.approx.ftz.f32 %f1790, %f1769, %f1786; div.approx.ftz.f32 %f1791, %f1770, %f1787; div.approx.ftz.f32 %f1792, %f1771, %f1788; mul.ftz.f32 %f1793, %f1789, %f1790; neg.ftz.f32 %f1794, %f1793; mov.f32 %f1795, 0fBF800000; max.ftz.f32 %f1796, %f1794, %f1795; mov.f32 %f1797, 0f3F800000; min.ftz.f32 %f1798, %f1796, %f1797; abs.ftz.f32 %f1799, %f1798; sub.ftz.f32 %f1800, %f1797, %f1799; mul.ftz.f32 %f1801, %f1800, 0f3F000000; sqrt.approx.ftz.f32 %f1802, %f1801; setp.gt.ftz.f32 %p130, %f1799, 0f3F11EB85; selp.f32 %f1803, %f1802, %f1799, %p130; mul.ftz.f32 %f1804, %f1803, %f1803; mov.f32 %f1805, 0f3C94D2E9; mov.f32 %f1806, 0f3D53F941; fma.rn.ftz.f32 %f1807, %f1806, %f1804, %f1805; mov.f32 %f1808, 0f3D3F841F; fma.rn.ftz.f32 %f1809, %f1807, %f1804, %f1808; mov.f32 %f1810, 0f3D994929; fma.rn.ftz.f32 %f1811, %f1809, %f1804, %f1810; mov.f32 %f1812, 0f3E2AAB94; fma.rn.ftz.f32 %f1813, %f1811, %f1804, %f1812; mul.ftz.f32 %f1814, %f1804, %f1813; fma.rn.ftz.f32 %f1815, %f1814, %f1803, %f1803; add.ftz.f32 %f1816, %f1815, %f1815; mov.f32 %f1817, 0f3FC90FDB; sub.ftz.f32 %f1818, %f1817, %f1815; selp.f32 %f1819, %f1816, %f1818, %p130; setp.lt.ftz.f32 %p131, %f1798, 0f00000000; mov.f32 %f1820, 0f40490FDB; sub.ftz.f32 %f1821, %f1820, %f1819; selp.f32 %f1822, %f1821, %f1819, %p131; mul.ftz.f32 %f1823, %f1790, %f1791; neg.ftz.f32 %f1824, %f1823; max.ftz.f32 %f1825, %f1824, %f1795; min.ftz.f32 %f1826, %f1825, %f1797; abs.ftz.f32 %f1827, %f1826; sub.ftz.f32 %f1828, %f1797, %f1827; mul.ftz.f32 %f1829, %f1828, 0f3F000000; sqrt.approx.ftz.f32 %f1830, %f1829; setp.gt.ftz.f32 %p132, %f1827, 0f3F11EB85; selp.f32 %f1831, %f1830, %f1827, %p132; mul.ftz.f32 %f1832, %f1831, %f1831; fma.rn.ftz.f32 %f1833, %f1806, %f1832, %f1805; fma.rn.ftz.f32 %f1834, %f1833, %f1832, %f1808; fma.rn.ftz.f32 %f1835, %f1834, %f1832, %f1810; fma.rn.ftz.f32 %f1836, %f1835, %f1832, %f1812; mul.ftz.f32 %f1837, %f1832, %f1836; fma.rn.ftz.f32 %f1838, %f1837, %f1831, %f1831; add.ftz.f32 %f1839, %f1838, %f1838; sub.ftz.f32 %f1840, %f1817, %f1838; selp.f32 %f1841, %f1839, %f1840, %p132; setp.lt.ftz.f32 %p133, %f1826, 0f00000000; sub.ftz.f32 %f1842, %f1820, %f1841; selp.f32 %f1843, %f1842, %f1841, %p133; mul.ftz.f32 %f1844, %f1791, %f1792; neg.ftz.f32 %f1845, %f1844; max.ftz.f32 %f1846, %f1845, %f1795; min.ftz.f32 %f1847, %f1846, %f1797; abs.ftz.f32 %f1848, %f1847; sub.ftz.f32 %f1849, %f1797, %f1848; mul.ftz.f32 %f1850, %f1849, 0f3F000000; sqrt.approx.ftz.f32 %f1851, %f1850; setp.gt.ftz.f32 %p134, %f1848, 0f3F11EB85; selp.f32 %f1852, %f1851, %f1848, %p134; mul.ftz.f32 %f1853, %f1852, %f1852; fma.rn.ftz.f32 %f1854, %f1806, %f1853, %f1805; fma.rn.ftz.f32 %f1855, %f1854, %f1853, %f1808; fma.rn.ftz.f32 %f1856, %f1855, %f1853, %f1810; fma.rn.ftz.f32 %f1857, %f1856, %f1853, %f1812; mul.ftz.f32 %f1858, %f1853, %f1857; fma.rn.ftz.f32 %f1859, %f1858, %f1852, %f1852; add.ftz.f32 %f1860, %f1859, %f1859; sub.ftz.f32 %f1861, %f1817, %f1859; selp.f32 %f1862, %f1860, %f1861, %p134; setp.lt.ftz.f32 %p135, %f1847, 0f00000000; sub.ftz.f32 %f1863, %f1820, %f1862; selp.f32 %f1864, %f1863, %f1862, %p135; mul.ftz.f32 %f1865, %f1789, %f1792; neg.ftz.f32 %f1866, %f1865; max.ftz.f32 %f1867, %f1866, %f1795; min.ftz.f32 %f1868, %f1867, %f1797; abs.ftz.f32 %f1869, %f1868; sub.ftz.f32 %f1870, %f1797, %f1869; mul.ftz.f32 %f1871, %f1870, 0f3F000000; sqrt.approx.ftz.f32 %f1872, %f1871; setp.gt.ftz.f32 %p136, %f1869, 0f3F11EB85; selp.f32 %f1873, %f1872, %f1869, %p136; mul.ftz.f32 %f1874, %f1873, %f1873; fma.rn.ftz.f32 %f1875, %f1806, %f1874, %f1805; fma.rn.ftz.f32 %f1876, %f1875, %f1874, %f1808; fma.rn.ftz.f32 %f1877, %f1876, %f1874, %f1810; fma.rn.ftz.f32 %f1878, %f1877, %f1874, %f1812; mul.ftz.f32 %f1879, %f1874, %f1878; fma.rn.ftz.f32 %f1880, %f1879, %f1873, %f1873; add.ftz.f32 %f1881, %f1880, %f1880; sub.ftz.f32 %f1882, %f1817, %f1880; selp.f32 %f1883, %f1881, %f1882, %p136; setp.lt.ftz.f32 %p137, %f1868, 0f00000000; sub.ftz.f32 %f1884, %f1820, %f1883; selp.f32 %f1885, %f1884, %f1883, %p137; mov.f32 %f1886, 0f40C90FDB; sub.ftz.f32 %f1887, %f1886, %f1864; sub.ftz.f32 %f1888, %f1887, %f1885; add.ftz.f32 %f1889, %f1822, %f1843; sub.ftz.f32 %f277, %f1889, %f1888; fma.rn.ftz.f32 %f1890, %f1, %f277, %f1888; cos.approx.ftz.f32 %f1891, %f1890; mul.ftz.f32 %f1892, %f1789, %f1891; sub.ftz.f32 %f1893, %f1892, %f1791; sin.approx.ftz.f32 %f1894, %f1890; div.approx.ftz.f32 %f1895, %f1893, %f1894; mul.ftz.f32 %f1896, %f1895, %f1895; fma.rn.ftz.f32 %f1897, %f1789, %f1789, %f1896; rsqrt.approx.ftz.f32 %f1898, %f1897; setp.gt.ftz.f32 %p138, %f1895, 0f00000000; selp.f32 %f1899, 0f3F800000, 0fBF800000, %p138; mul.ftz.f32 %f1900, %f1898, %f1899; max.ftz.f32 %f1901, %f1900, %f1795; min.ftz.f32 %f1902, %f1901, %f1797; mul.ftz.f32 %f1903, %f3795, %f1902; neg.ftz.f32 %f1904, %f1903; mul.ftz.f32 %f1905, %f1902, %f1902; sub.ftz.f32 %f1906, %f1797, %f1905; sqrt.approx.ftz.f32 %f1907, %f1906; mov.f32 %f1908, 0f33D6BF95; max.ftz.f32 %f1909, %f1907, %f1908; div.approx.ftz.f32 %f1910, %f1904, %f1909; max.ftz.f32 %f1911, %f1910, %f1759; min.ftz.f32 %f278, %f1911, %f1763; fma.rn.ftz.f32 %f1912, %f278, %f278, %f1772; sqrt.approx.ftz.f32 %f279, %f1912; mul.ftz.f32 %f1913, %f279, %f279; fma.rn.ftz.f32 %f1914, %f1762, %f1762, %f1913; sqrt.approx.ftz.f32 %f1915, %f1914; div.approx.ftz.f32 %f1916, %f1762, %f1915; fma.rn.ftz.f32 %f1917, %f3796, %f3796, %f1913; sqrt.approx.ftz.f32 %f1918, %f1917; div.approx.ftz.f32 %f1919, %f3796, %f1918; sub.ftz.f32 %f1920, %f1919, %f1916; fma.rn.ftz.f32 %f280, %f233, %f1920, %f1916; mul.ftz.f32 %f281, %f280, %f280; setp.geu.ftz.f32 %p139, %f281, 0f3F7FFFEF; @%p139 bra $L__BB19_106; mul.ftz.f32 %f1921, %f279, %f280; sub.ftz.f32 %f1923, %f1797, %f281; sqrt.approx.ftz.f32 %f1924, %f1923; div.approx.ftz.f32 %f3796, %f1921, %f1924; $L__BB19_106: fma.rn.ftz.f32 %f1926, %f254, %f278, %f934; fma.rn.ftz.f32 %f1927, %f255, %f278, %f935; fma.rn.ftz.f32 %f1928, %f256, %f278, %f936; fma.rn.ftz.f32 %f1929, %f258, %f3796, %f1926; fma.rn.ftz.f32 %f1930, %f259, %f3796, %f1927; fma.rn.ftz.f32 %f1931, %f260, %f3796, %f1928; fma.rn.ftz.f32 %f284, %f3792, %f3795, %f1929; fma.rn.ftz.f32 %f285, %f3793, %f3795, %f1930; fma.rn.ftz.f32 %f286, %f3794, %f3795, %f1931; setp.eq.ftz.f32 %p140, %f277, 0f00000000; mov.f32 %f3799, 0f00000000; @%p140 bra $L__BB19_108; rcp.approx.ftz.f32 %f3799, %f277; $L__BB19_108: st.local.f32 [%rd1+48], %f3799; sub.ftz.f32 %f1932, %f284, %f934; sub.ftz.f32 %f1933, %f285, %f935; mul.ftz.f32 %f1934, %f1933, %f1933; fma.rn.ftz.f32 %f1935, %f1932, %f1932, %f1934; sub.ftz.f32 %f1936, %f286, %f936; fma.rn.ftz.f32 %f1937, %f1936, %f1936, %f1935; sqrt.approx.ftz.f32 %f1938, %f1937; div.approx.ftz.f32 %f3830, %f1932, %f1938; div.approx.ftz.f32 %f3831, %f1933, %f1938; div.approx.ftz.f32 %f3832, %f1936, %f1938; $L__BB19_112: cvt.rn.f32.s32 %f3729, %r305; div.approx.ftz.f32 %f1969, %f3799, %f3729; st.local.f32 [%rd1+48], %f1969; $L__BB19_113: setp.lt.s32 %p143, %r305, 2; @%p143 bra $L__BB19_133; setp.lt.s32 %p307, %r53, 1; mov.u32 %r320, 0; mov.f32 %f3811, 0f00000000; mov.f32 %f3809, %f3811; @%p307 bra $L__BB19_130; ld.const.u64 %rd248, [kernel_params+2368]; cvta.to.global.u64 %rd247, %rd248; ld.const.u32 %r188, [kernel_params+1052]; neg.s32 %r317, %r316; mov.u32 %r318, 0; mul.wide.s32 %rd123, %r188, 192; add.s64 %rd124, %rd247, %rd123; add.s64 %rd250, %rd124, 176; mov.f32 %f3809, 0f00000000; mov.u32 %r320, %r318; $L__BB19_116: setp.eq.s32 %p145, %r317, 0; @%p145 bra $L__BB19_129; add.s64 %rd19, %rd250, -172; ld.global.v4.f32 {%f1975, %f1976, %f1977, %f1978}, [%rd250]; ld.global.f32 %f317, [%rd250+-172]; sub.ftz.f32 %f1979, %f934, %f317; ld.global.f32 %f318, [%rd250+-168]; sub.ftz.f32 %f1980, %f935, %f318; ld.global.f32 %f319, [%rd250+-164]; sub.ftz.f32 %f1981, %f936, %f319; mul.ftz.f32 %f1982, %f1980, %f1976; fma.rn.ftz.f32 %f1983, %f1979, %f1975, %f1982; fma.rn.ftz.f32 %f1984, %f1981, %f1977, %f1983; setp.leu.ftz.f32 %p146, %f1984, 0f38D1B717; @%p146 bra $L__BB19_129; add.s32 %r320, %r320, 1; ld.global.v4.f32 {%f1985, %f1986, %f1987, %f1988}, [%rd19+140]; ld.global.v4.f32 {%f1989, %f1990, %f1991, %f1992}, [%rd19+156]; mul.ftz.f32 %f1993, %f935, %f1976; fma.rn.ftz.f32 %f1994, %f934, %f1975, %f1993; fma.rn.ftz.f32 %f1995, %f936, %f1977, %f1994; mul.ftz.f32 %f1996, %f318, %f1976; fma.rn.ftz.f32 %f1997, %f317, %f1975, %f1996; fma.rn.ftz.f32 %f1998, %f319, %f1977, %f1997; sub.ftz.f32 %f1999, %f1995, %f1998; neg.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f3831, %f1976; fma.rn.ftz.f32 %f2002, %f3830, %f1975, %f2001; fma.rn.ftz.f32 %f2003, %f3832, %f1977, %f2002; div.approx.ftz.f32 %f327, %f2000, %f2003; setp.leu.ftz.f32 %p147, %f327, 0f38D1B717; setp.geu.ftz.f32 %p148, %f327, 0f7F7FFFFF; or.pred %p149, %p148, %p147; @%p149 bra $L__BB19_129; fma.rn.ftz.f32 %f2004, %f3830, %f327, %f934; fma.rn.ftz.f32 %f2005, %f3831, %f327, %f935; fma.rn.ftz.f32 %f2006, %f3832, %f327, %f936; sub.ftz.f32 %f328, %f2004, %f317; sub.ftz.f32 %f329, %f2005, %f318; sub.ftz.f32 %f330, %f2006, %f319; mul.ftz.f32 %f2007, %f1986, %f329; fma.rn.ftz.f32 %f2008, %f1985, %f328, %f2007; fma.rn.ftz.f32 %f2009, %f1987, %f330, %f2008; mul.ftz.f32 %f2010, %f1986, %f1986; fma.rn.ftz.f32 %f2011, %f1985, %f1985, %f2010; fma.rn.ftz.f32 %f331, %f1987, %f1987, %f2011; div.approx.ftz.f32 %f332, %f2009, %f331; setp.lt.ftz.f32 %p150, %f332, 0fBF000000; setp.gt.ftz.f32 %p151, %f332, 0f3F000000; or.pred %p152, %p150, %p151; @%p152 bra $L__BB19_129; mul.ftz.f32 %f2012, %f1990, %f329; fma.rn.ftz.f32 %f2013, %f1989, %f328, %f2012; fma.rn.ftz.f32 %f2014, %f1991, %f330, %f2013; mul.ftz.f32 %f2015, %f1990, %f1990; fma.rn.ftz.f32 %f2016, %f1989, %f1989, %f2015; fma.rn.ftz.f32 %f333, %f1991, %f1991, %f2016; div.approx.ftz.f32 %f334, %f2014, %f333; setp.lt.ftz.f32 %p153, %f334, 0fBF000000; setp.gt.ftz.f32 %p154, %f334, 0f3F000000; or.pred %p155, %p153, %p154; @%p155 bra $L__BB19_129; setp.lt.ftz.f32 %p156, %f1988, 0f00000000; mul.ftz.f32 %f2017, %f334, %f334; fma.rn.ftz.f32 %f2018, %f332, %f332, %f2017; setp.gt.ftz.f32 %p157, %f2018, 0f3E800000; and.pred %p158, %p156, %p157; @%p158 bra $L__BB19_129; @%p156 bra $L__BB19_126; bra.uni $L__BB19_123; $L__BB19_126: sub.ftz.f32 %f2194, %f317, %f934; sub.ftz.f32 %f2195, %f318, %f935; mul.ftz.f32 %f2196, %f2195, %f2195; fma.rn.ftz.f32 %f2197, %f2194, %f2194, %f2196; sub.ftz.f32 %f2198, %f319, %f936; fma.rn.ftz.f32 %f2199, %f2198, %f2198, %f2197; sqrt.approx.ftz.f32 %f339, %f2199; rsqrt.approx.ftz.f32 %f2200, %f2199; mul.ftz.f32 %f2201, %f2194, %f2200; mul.ftz.f32 %f2202, %f2195, %f2200; mul.ftz.f32 %f2203, %f2198, %f2200; mul.ftz.f32 %f2204, %f1975, %f2201; mul.ftz.f32 %f2205, %f1976, %f2202; neg.ftz.f32 %f2206, %f2205; sub.ftz.f32 %f2207, %f2206, %f2204; mul.ftz.f32 %f2208, %f1977, %f2203; sub.ftz.f32 %f340, %f2207, %f2208; setp.le.ftz.f32 %p170, %f340, 0f00000000; mov.f32 %f3808, 0f00000000; @%p170 bra $L__BB19_128; mul.ftz.f32 %f2209, %f339, %f339; div.approx.ftz.f32 %f3808, %f2209, %f340; $L__BB19_128: abs.ftz.f32 %f2210, %f1988; fma.rn.ftz.f32 %f3809, %f2210, %f3808, %f3809; bra.uni $L__BB19_129; $L__BB19_123: mul.ftz.f32 %f2020, %f1985, 0f3F000000; sub.ftz.f32 %f2021, %f317, %f2020; mul.ftz.f32 %f2022, %f1986, 0f3F000000; sub.ftz.f32 %f2023, %f318, %f2022; mul.ftz.f32 %f2024, %f1987, 0f3F000000; sub.ftz.f32 %f2025, %f319, %f2024; mul.ftz.f32 %f2026, %f1989, 0f3F000000; sub.ftz.f32 %f2027, %f2021, %f2026; mul.ftz.f32 %f2028, %f1990, 0f3F000000; sub.ftz.f32 %f2029, %f2023, %f2028; mul.ftz.f32 %f2030, %f1991, 0f3F000000; sub.ftz.f32 %f2031, %f2025, %f2030; sqrt.approx.ftz.f32 %f2032, %f331; rsqrt.approx.ftz.f32 %f2033, %f331; mov.f32 %f2034, 0f3F800000; mul.ftz.f32 %f2035, %f1985, %f2033; mul.ftz.f32 %f2036, %f1986, %f2033; mul.ftz.f32 %f2037, %f1987, %f2033; sqrt.approx.ftz.f32 %f2038, %f333; rsqrt.approx.ftz.f32 %f2039, %f333; mul.ftz.f32 %f2040, %f1989, %f2039; mul.ftz.f32 %f2041, %f1990, %f2039; mul.ftz.f32 %f2042, %f1991, %f2039; mul.ftz.f32 %f2043, %f2036, %f2042; mul.ftz.f32 %f2044, %f2037, %f2041; sub.ftz.f32 %f2045, %f2043, %f2044; mul.ftz.f32 %f2046, %f2037, %f2040; mul.ftz.f32 %f2047, %f2035, %f2042; sub.ftz.f32 %f2048, %f2046, %f2047; mul.ftz.f32 %f2049, %f2035, %f2041; mul.ftz.f32 %f2050, %f2036, %f2040; sub.ftz.f32 %f2051, %f2049, %f2050; sub.ftz.f32 %f2052, %f2027, %f934; sub.ftz.f32 %f2053, %f2029, %f935; sub.ftz.f32 %f2054, %f2031, %f936; mul.ftz.f32 %f2055, %f2053, %f2048; fma.rn.ftz.f32 %f2056, %f2052, %f2045, %f2055; fma.rn.ftz.f32 %f2057, %f2054, %f2051, %f2056; setp.gt.ftz.f32 %p160, %f2057, 0f00000000; mov.f32 %f3807, 0f00000000; neg.ftz.f32 %f2058, %f2057; selp.f32 %f2059, %f2058, %f2057, %p160; mul.ftz.f32 %f2060, %f2053, %f2036; fma.rn.ftz.f32 %f2061, %f2052, %f2035, %f2060; fma.rn.ftz.f32 %f2062, %f2054, %f2037, %f2061; mul.ftz.f32 %f2063, %f2053, %f2041; fma.rn.ftz.f32 %f2064, %f2052, %f2040, %f2063; fma.rn.ftz.f32 %f2065, %f2054, %f2042, %f2064; add.ftz.f32 %f2066, %f2032, %f2062; add.ftz.f32 %f2067, %f2038, %f2065; sub.ftz.f32 %f2068, %f2062, %f2066; sub.ftz.f32 %f2069, %f2067, %f2065; sub.ftz.f32 %f2070, %f2066, %f2062; sub.ftz.f32 %f2071, %f2065, %f2067; mul.ftz.f32 %f2072, %f2065, %f2068; mul.ftz.f32 %f2073, %f2066, %f2069; mul.ftz.f32 %f2074, %f2067, %f2070; mul.ftz.f32 %f2075, %f2062, %f2071; mul.ftz.f32 %f2076, %f2059, %f2059; mul.ftz.f32 %f2077, %f2068, %f2076; mul.ftz.f32 %f2078, %f2069, %f2076; mul.ftz.f32 %f2079, %f2070, %f2076; mul.ftz.f32 %f2080, %f2071, %f2076; mul.ftz.f32 %f2081, %f2068, %f2077; mul.ftz.f32 %f2082, %f2069, %f2078; mul.ftz.f32 %f2083, %f2070, %f2079; mul.ftz.f32 %f2084, %f2071, %f2080; fma.rn.ftz.f32 %f2085, %f2072, %f2072, %f2081; fma.rn.ftz.f32 %f2086, %f2073, %f2073, %f2082; fma.rn.ftz.f32 %f2087, %f2074, %f2074, %f2083; fma.rn.ftz.f32 %f2088, %f2075, %f2075, %f2084; sqrt.approx.ftz.f32 %f2089, %f2085; sqrt.approx.ftz.f32 %f2090, %f2086; sqrt.approx.ftz.f32 %f2091, %f2087; sqrt.approx.ftz.f32 %f2092, %f2088; div.approx.ftz.f32 %f2093, %f2072, %f2089; div.approx.ftz.f32 %f2094, %f2073, %f2090; div.approx.ftz.f32 %f2095, %f2074, %f2091; div.approx.ftz.f32 %f2096, %f2075, %f2092; mul.ftz.f32 %f2097, %f2093, %f2094; neg.ftz.f32 %f2098, %f2097; mov.f32 %f2099, 0fBF800000; max.ftz.f32 %f2100, %f2098, %f2099; min.ftz.f32 %f2101, %f2100, %f2034; abs.ftz.f32 %f2102, %f2101; sub.ftz.f32 %f2103, %f2034, %f2102; mul.ftz.f32 %f2104, %f2103, 0f3F000000; sqrt.approx.ftz.f32 %f2105, %f2104; setp.gt.ftz.f32 %p161, %f2102, 0f3F11EB85; selp.f32 %f2106, %f2105, %f2102, %p161; mul.ftz.f32 %f2107, %f2106, %f2106; mov.f32 %f2108, 0f3C94D2E9; mov.f32 %f2109, 0f3D53F941; fma.rn.ftz.f32 %f2110, %f2109, %f2107, %f2108; mov.f32 %f2111, 0f3D3F841F; fma.rn.ftz.f32 %f2112, %f2110, %f2107, %f2111; mov.f32 %f2113, 0f3D994929; fma.rn.ftz.f32 %f2114, %f2112, %f2107, %f2113; mov.f32 %f2115, 0f3E2AAB94; fma.rn.ftz.f32 %f2116, %f2114, %f2107, %f2115; mul.ftz.f32 %f2117, %f2107, %f2116; fma.rn.ftz.f32 %f2118, %f2117, %f2106, %f2106; add.ftz.f32 %f2119, %f2118, %f2118; mov.f32 %f2120, 0f3FC90FDB; sub.ftz.f32 %f2121, %f2120, %f2118; selp.f32 %f2122, %f2119, %f2121, %p161; setp.lt.ftz.f32 %p162, %f2101, 0f00000000; mov.f32 %f2123, 0f40490FDB; sub.ftz.f32 %f2124, %f2123, %f2122; selp.f32 %f2125, %f2124, %f2122, %p162; mul.ftz.f32 %f2126, %f2094, %f2095; neg.ftz.f32 %f2127, %f2126; max.ftz.f32 %f2128, %f2127, %f2099; min.ftz.f32 %f2129, %f2128, %f2034; abs.ftz.f32 %f2130, %f2129; sub.ftz.f32 %f2131, %f2034, %f2130; mul.ftz.f32 %f2132, %f2131, 0f3F000000; sqrt.approx.ftz.f32 %f2133, %f2132; setp.gt.ftz.f32 %p163, %f2130, 0f3F11EB85; selp.f32 %f2134, %f2133, %f2130, %p163; mul.ftz.f32 %f2135, %f2134, %f2134; fma.rn.ftz.f32 %f2136, %f2109, %f2135, %f2108; fma.rn.ftz.f32 %f2137, %f2136, %f2135, %f2111; fma.rn.ftz.f32 %f2138, %f2137, %f2135, %f2113; fma.rn.ftz.f32 %f2139, %f2138, %f2135, %f2115; mul.ftz.f32 %f2140, %f2135, %f2139; fma.rn.ftz.f32 %f2141, %f2140, %f2134, %f2134; add.ftz.f32 %f2142, %f2141, %f2141; sub.ftz.f32 %f2143, %f2120, %f2141; selp.f32 %f2144, %f2142, %f2143, %p163; setp.lt.ftz.f32 %p164, %f2129, 0f00000000; sub.ftz.f32 %f2145, %f2123, %f2144; selp.f32 %f2146, %f2145, %f2144, %p164; mul.ftz.f32 %f2147, %f2095, %f2096; neg.ftz.f32 %f2148, %f2147; max.ftz.f32 %f2149, %f2148, %f2099; min.ftz.f32 %f2150, %f2149, %f2034; abs.ftz.f32 %f2151, %f2150; sub.ftz.f32 %f2152, %f2034, %f2151; mul.ftz.f32 %f2153, %f2152, 0f3F000000; sqrt.approx.ftz.f32 %f2154, %f2153; setp.gt.ftz.f32 %p165, %f2151, 0f3F11EB85; selp.f32 %f2155, %f2154, %f2151, %p165; mul.ftz.f32 %f2156, %f2155, %f2155; fma.rn.ftz.f32 %f2157, %f2109, %f2156, %f2108; fma.rn.ftz.f32 %f2158, %f2157, %f2156, %f2111; fma.rn.ftz.f32 %f2159, %f2158, %f2156, %f2113; fma.rn.ftz.f32 %f2160, %f2159, %f2156, %f2115; mul.ftz.f32 %f2161, %f2156, %f2160; fma.rn.ftz.f32 %f2162, %f2161, %f2155, %f2155; add.ftz.f32 %f2163, %f2162, %f2162; sub.ftz.f32 %f2164, %f2120, %f2162; selp.f32 %f2165, %f2163, %f2164, %p165; setp.lt.ftz.f32 %p166, %f2150, 0f00000000; sub.ftz.f32 %f2166, %f2123, %f2165; selp.f32 %f2167, %f2166, %f2165, %p166; mul.ftz.f32 %f2168, %f2093, %f2096; neg.ftz.f32 %f2169, %f2168; max.ftz.f32 %f2170, %f2169, %f2099; min.ftz.f32 %f2171, %f2170, %f2034; abs.ftz.f32 %f2172, %f2171; sub.ftz.f32 %f2173, %f2034, %f2172; mul.ftz.f32 %f2174, %f2173, 0f3F000000; sqrt.approx.ftz.f32 %f2175, %f2174; setp.gt.ftz.f32 %p167, %f2172, 0f3F11EB85; selp.f32 %f2176, %f2175, %f2172, %p167; mul.ftz.f32 %f2177, %f2176, %f2176; fma.rn.ftz.f32 %f2178, %f2109, %f2177, %f2108; fma.rn.ftz.f32 %f2179, %f2178, %f2177, %f2111; fma.rn.ftz.f32 %f2180, %f2179, %f2177, %f2113; fma.rn.ftz.f32 %f2181, %f2180, %f2177, %f2115; mul.ftz.f32 %f2182, %f2177, %f2181; fma.rn.ftz.f32 %f2183, %f2182, %f2176, %f2176; add.ftz.f32 %f2184, %f2183, %f2183; sub.ftz.f32 %f2185, %f2120, %f2183; selp.f32 %f2186, %f2184, %f2185, %p167; setp.lt.ftz.f32 %p168, %f2171, 0f00000000; sub.ftz.f32 %f2187, %f2123, %f2186; selp.f32 %f2188, %f2187, %f2186, %p168; mov.f32 %f2189, 0f40C90FDB; sub.ftz.f32 %f2190, %f2189, %f2167; sub.ftz.f32 %f2191, %f2190, %f2188; add.ftz.f32 %f2192, %f2125, %f2146; sub.ftz.f32 %f335, %f2192, %f2191; setp.eq.ftz.f32 %p169, %f335, 0f00000000; @%p169 bra $L__BB19_125; rcp.approx.ftz.f32 %f3807, %f335; $L__BB19_125: add.ftz.f32 %f3809, %f3809, %f3807; $L__BB19_129: add.s32 %r317, %r317, 1; add.s64 %rd250, %rd250, 192; add.s32 %r318, %r318, 1; setp.lt.s32 %p171, %r318, %r53; @%p171 bra $L__BB19_116; $L__BB19_130: setp.lt.s32 %p172, %r320, 0; @%p172 bra $L__BB19_132; add.s32 %r189, %r320, 1; cvt.rn.f32.s32 %f2212, %r189; div.approx.ftz.f32 %f3811, %f3809, %f2212; $L__BB19_132: ld.local.f32 %f2213, [%rd1+48]; add.ftz.f32 %f2214, %f3811, %f2213; st.local.f32 [%rd1+48], %f2214; $L__BB19_133: setp.eq.ftz.f32 %p306, %f205, 0f3F800000; @%p306 bra $L__BB19_171; ld.local.f32 %f3812, [%rd1+48]; mov.u32 %r322, 0; mov.f32 %f3816, %f205; $L__BB19_135: mul.ftz.f32 %f3825, %f3816, %f3812; st.local.f32 [%rd1+48], %f3825; setp.eq.ftz.f32 %p174, %f205, 0f00000000; setp.eq.s32 %p175, %r322, 0; or.pred %p176, %p174, %p175; @%p176 bra $L__BB19_154; ld.const.u32 %r71, [kernel_params+1048]; setp.lt.s32 %p177, %r71, 1; mov.u32 %r325, 0; mov.f32 %f3822, 0f00000000; mov.f32 %f3820, %f3822; @%p177 bra $L__BB19_151; ld.const.u64 %rd246, [kernel_params+2368]; cvta.to.global.u64 %rd245, %rd246; ld.const.u32 %r194, [kernel_params+1052]; mul.wide.s32 %rd125, %r194, 192; add.s64 %rd126, %rd245, %rd125; add.s64 %rd251, %rd126, 176; mov.f32 %f3820, 0f00000000; mov.u32 %r323, 0; mov.u32 %r325, %r323; $L__BB19_138: ld.global.v4.f32 {%f2217, %f2218, %f2219, %f2220}, [%rd251]; ld.global.f32 %f362, [%rd251+-172]; sub.ftz.f32 %f2221, %f934, %f362; ld.global.f32 %f363, [%rd251+-168]; sub.ftz.f32 %f2222, %f935, %f363; ld.global.f32 %f364, [%rd251+-164]; sub.ftz.f32 %f2223, %f936, %f364; mul.ftz.f32 %f2224, %f2222, %f2218; fma.rn.ftz.f32 %f2225, %f2221, %f2217, %f2224; fma.rn.ftz.f32 %f2226, %f2223, %f2219, %f2225; setp.leu.ftz.f32 %p178, %f2226, 0f38D1B717; @%p178 bra $L__BB19_150; add.s32 %r325, %r325, 1; ld.global.v4.f32 {%f2227, %f2228, %f2229, %f2230}, [%rd251+-32]; ld.global.v4.f32 {%f2231, %f2232, %f2233, %f2234}, [%rd251+-16]; mul.ftz.f32 %f2235, %f935, %f2218; fma.rn.ftz.f32 %f2236, %f934, %f2217, %f2235; fma.rn.ftz.f32 %f2237, %f936, %f2219, %f2236; mul.ftz.f32 %f2238, %f363, %f2218; fma.rn.ftz.f32 %f2239, %f362, %f2217, %f2238; fma.rn.ftz.f32 %f2240, %f364, %f2219, %f2239; sub.ftz.f32 %f2241, %f2237, %f2240; neg.ftz.f32 %f2242, %f2241; mul.ftz.f32 %f2243, %f3831, %f2218; fma.rn.ftz.f32 %f2244, %f3830, %f2217, %f2243; fma.rn.ftz.f32 %f2245, %f3832, %f2219, %f2244; div.approx.ftz.f32 %f372, %f2242, %f2245; setp.leu.ftz.f32 %p179, %f372, 0f38D1B717; setp.geu.ftz.f32 %p180, %f372, 0f7F7FFFFF; or.pred %p181, %p180, %p179; @%p181 bra $L__BB19_150; fma.rn.ftz.f32 %f2246, %f3830, %f372, %f934; fma.rn.ftz.f32 %f2247, %f3831, %f372, %f935; fma.rn.ftz.f32 %f2248, %f3832, %f372, %f936; sub.ftz.f32 %f373, %f2246, %f362; sub.ftz.f32 %f374, %f2247, %f363; sub.ftz.f32 %f375, %f2248, %f364; mul.ftz.f32 %f2249, %f2228, %f374; fma.rn.ftz.f32 %f2250, %f2227, %f373, %f2249; fma.rn.ftz.f32 %f2251, %f2229, %f375, %f2250; mul.ftz.f32 %f2252, %f2228, %f2228; fma.rn.ftz.f32 %f2253, %f2227, %f2227, %f2252; fma.rn.ftz.f32 %f376, %f2229, %f2229, %f2253; div.approx.ftz.f32 %f377, %f2251, %f376; setp.lt.ftz.f32 %p182, %f377, 0fBF000000; setp.gt.ftz.f32 %p183, %f377, 0f3F000000; or.pred %p184, %p182, %p183; @%p184 bra $L__BB19_150; mul.ftz.f32 %f2254, %f2232, %f374; fma.rn.ftz.f32 %f2255, %f2231, %f373, %f2254; fma.rn.ftz.f32 %f2256, %f2233, %f375, %f2255; mul.ftz.f32 %f2257, %f2232, %f2232; fma.rn.ftz.f32 %f2258, %f2231, %f2231, %f2257; fma.rn.ftz.f32 %f378, %f2233, %f2233, %f2258; div.approx.ftz.f32 %f379, %f2256, %f378; setp.lt.ftz.f32 %p185, %f379, 0fBF000000; setp.gt.ftz.f32 %p186, %f379, 0f3F000000; or.pred %p187, %p185, %p186; @%p187 bra $L__BB19_150; setp.lt.ftz.f32 %p188, %f2230, 0f00000000; mul.ftz.f32 %f2259, %f379, %f379; fma.rn.ftz.f32 %f2260, %f377, %f377, %f2259; setp.gt.ftz.f32 %p189, %f2260, 0f3E800000; and.pred %p190, %p188, %p189; @%p190 bra $L__BB19_150; @%p188 bra $L__BB19_147; bra.uni $L__BB19_144; $L__BB19_147: sub.ftz.f32 %f2436, %f362, %f934; sub.ftz.f32 %f2437, %f363, %f935; mul.ftz.f32 %f2438, %f2437, %f2437; fma.rn.ftz.f32 %f2439, %f2436, %f2436, %f2438; sub.ftz.f32 %f2440, %f364, %f936; fma.rn.ftz.f32 %f2441, %f2440, %f2440, %f2439; sqrt.approx.ftz.f32 %f384, %f2441; rsqrt.approx.ftz.f32 %f2442, %f2441; mul.ftz.f32 %f2443, %f2436, %f2442; mul.ftz.f32 %f2444, %f2437, %f2442; mul.ftz.f32 %f2445, %f2440, %f2442; mul.ftz.f32 %f2446, %f2217, %f2443; mul.ftz.f32 %f2447, %f2218, %f2444; neg.ftz.f32 %f2448, %f2447; sub.ftz.f32 %f2449, %f2448, %f2446; mul.ftz.f32 %f2450, %f2219, %f2445; sub.ftz.f32 %f385, %f2449, %f2450; setp.le.ftz.f32 %p202, %f385, 0f00000000; mov.f32 %f3819, 0f00000000; @%p202 bra $L__BB19_149; mul.ftz.f32 %f2451, %f384, %f384; div.approx.ftz.f32 %f3819, %f2451, %f385; $L__BB19_149: abs.ftz.f32 %f2452, %f2230; fma.rn.ftz.f32 %f3820, %f2452, %f3819, %f3820; bra.uni $L__BB19_150; $L__BB19_144: mul.ftz.f32 %f2262, %f2227, 0f3F000000; sub.ftz.f32 %f2263, %f362, %f2262; mul.ftz.f32 %f2264, %f2228, 0f3F000000; sub.ftz.f32 %f2265, %f363, %f2264; mul.ftz.f32 %f2266, %f2229, 0f3F000000; sub.ftz.f32 %f2267, %f364, %f2266; mul.ftz.f32 %f2268, %f2231, 0f3F000000; sub.ftz.f32 %f2269, %f2263, %f2268; mul.ftz.f32 %f2270, %f2232, 0f3F000000; sub.ftz.f32 %f2271, %f2265, %f2270; mul.ftz.f32 %f2272, %f2233, 0f3F000000; sub.ftz.f32 %f2273, %f2267, %f2272; sqrt.approx.ftz.f32 %f2274, %f376; rsqrt.approx.ftz.f32 %f2275, %f376; mov.f32 %f2276, 0f3F800000; mul.ftz.f32 %f2277, %f2227, %f2275; mul.ftz.f32 %f2278, %f2228, %f2275; mul.ftz.f32 %f2279, %f2229, %f2275; sqrt.approx.ftz.f32 %f2280, %f378; rsqrt.approx.ftz.f32 %f2281, %f378; mul.ftz.f32 %f2282, %f2231, %f2281; mul.ftz.f32 %f2283, %f2232, %f2281; mul.ftz.f32 %f2284, %f2233, %f2281; mul.ftz.f32 %f2285, %f2278, %f2284; mul.ftz.f32 %f2286, %f2279, %f2283; sub.ftz.f32 %f2287, %f2285, %f2286; mul.ftz.f32 %f2288, %f2279, %f2282; mul.ftz.f32 %f2289, %f2277, %f2284; sub.ftz.f32 %f2290, %f2288, %f2289; mul.ftz.f32 %f2291, %f2277, %f2283; mul.ftz.f32 %f2292, %f2278, %f2282; sub.ftz.f32 %f2293, %f2291, %f2292; sub.ftz.f32 %f2294, %f2269, %f934; sub.ftz.f32 %f2295, %f2271, %f935; sub.ftz.f32 %f2296, %f2273, %f936; mul.ftz.f32 %f2297, %f2295, %f2290; fma.rn.ftz.f32 %f2298, %f2294, %f2287, %f2297; fma.rn.ftz.f32 %f2299, %f2296, %f2293, %f2298; setp.gt.ftz.f32 %p192, %f2299, 0f00000000; mov.f32 %f3818, 0f00000000; neg.ftz.f32 %f2300, %f2299; selp.f32 %f2301, %f2300, %f2299, %p192; mul.ftz.f32 %f2302, %f2295, %f2278; fma.rn.ftz.f32 %f2303, %f2294, %f2277, %f2302; fma.rn.ftz.f32 %f2304, %f2296, %f2279, %f2303; mul.ftz.f32 %f2305, %f2295, %f2283; fma.rn.ftz.f32 %f2306, %f2294, %f2282, %f2305; fma.rn.ftz.f32 %f2307, %f2296, %f2284, %f2306; add.ftz.f32 %f2308, %f2274, %f2304; add.ftz.f32 %f2309, %f2280, %f2307; sub.ftz.f32 %f2310, %f2304, %f2308; sub.ftz.f32 %f2311, %f2309, %f2307; sub.ftz.f32 %f2312, %f2308, %f2304; sub.ftz.f32 %f2313, %f2307, %f2309; mul.ftz.f32 %f2314, %f2307, %f2310; mul.ftz.f32 %f2315, %f2308, %f2311; mul.ftz.f32 %f2316, %f2309, %f2312; mul.ftz.f32 %f2317, %f2304, %f2313; mul.ftz.f32 %f2318, %f2301, %f2301; mul.ftz.f32 %f2319, %f2310, %f2318; mul.ftz.f32 %f2320, %f2311, %f2318; mul.ftz.f32 %f2321, %f2312, %f2318; mul.ftz.f32 %f2322, %f2313, %f2318; mul.ftz.f32 %f2323, %f2310, %f2319; mul.ftz.f32 %f2324, %f2311, %f2320; mul.ftz.f32 %f2325, %f2312, %f2321; mul.ftz.f32 %f2326, %f2313, %f2322; fma.rn.ftz.f32 %f2327, %f2314, %f2314, %f2323; fma.rn.ftz.f32 %f2328, %f2315, %f2315, %f2324; fma.rn.ftz.f32 %f2329, %f2316, %f2316, %f2325; fma.rn.ftz.f32 %f2330, %f2317, %f2317, %f2326; sqrt.approx.ftz.f32 %f2331, %f2327; sqrt.approx.ftz.f32 %f2332, %f2328; sqrt.approx.ftz.f32 %f2333, %f2329; sqrt.approx.ftz.f32 %f2334, %f2330; div.approx.ftz.f32 %f2335, %f2314, %f2331; div.approx.ftz.f32 %f2336, %f2315, %f2332; div.approx.ftz.f32 %f2337, %f2316, %f2333; div.approx.ftz.f32 %f2338, %f2317, %f2334; mul.ftz.f32 %f2339, %f2335, %f2336; neg.ftz.f32 %f2340, %f2339; mov.f32 %f2341, 0fBF800000; max.ftz.f32 %f2342, %f2340, %f2341; min.ftz.f32 %f2343, %f2342, %f2276; abs.ftz.f32 %f2344, %f2343; sub.ftz.f32 %f2345, %f2276, %f2344; mul.ftz.f32 %f2346, %f2345, 0f3F000000; sqrt.approx.ftz.f32 %f2347, %f2346; setp.gt.ftz.f32 %p193, %f2344, 0f3F11EB85; selp.f32 %f2348, %f2347, %f2344, %p193; mul.ftz.f32 %f2349, %f2348, %f2348; mov.f32 %f2350, 0f3C94D2E9; mov.f32 %f2351, 0f3D53F941; fma.rn.ftz.f32 %f2352, %f2351, %f2349, %f2350; mov.f32 %f2353, 0f3D3F841F; fma.rn.ftz.f32 %f2354, %f2352, %f2349, %f2353; mov.f32 %f2355, 0f3D994929; fma.rn.ftz.f32 %f2356, %f2354, %f2349, %f2355; mov.f32 %f2357, 0f3E2AAB94; fma.rn.ftz.f32 %f2358, %f2356, %f2349, %f2357; mul.ftz.f32 %f2359, %f2349, %f2358; fma.rn.ftz.f32 %f2360, %f2359, %f2348, %f2348; add.ftz.f32 %f2361, %f2360, %f2360; mov.f32 %f2362, 0f3FC90FDB; sub.ftz.f32 %f2363, %f2362, %f2360; selp.f32 %f2364, %f2361, %f2363, %p193; setp.lt.ftz.f32 %p194, %f2343, 0f00000000; mov.f32 %f2365, 0f40490FDB; sub.ftz.f32 %f2366, %f2365, %f2364; selp.f32 %f2367, %f2366, %f2364, %p194; mul.ftz.f32 %f2368, %f2336, %f2337; neg.ftz.f32 %f2369, %f2368; max.ftz.f32 %f2370, %f2369, %f2341; min.ftz.f32 %f2371, %f2370, %f2276; abs.ftz.f32 %f2372, %f2371; sub.ftz.f32 %f2373, %f2276, %f2372; mul.ftz.f32 %f2374, %f2373, 0f3F000000; sqrt.approx.ftz.f32 %f2375, %f2374; setp.gt.ftz.f32 %p195, %f2372, 0f3F11EB85; selp.f32 %f2376, %f2375, %f2372, %p195; mul.ftz.f32 %f2377, %f2376, %f2376; fma.rn.ftz.f32 %f2378, %f2351, %f2377, %f2350; fma.rn.ftz.f32 %f2379, %f2378, %f2377, %f2353; fma.rn.ftz.f32 %f2380, %f2379, %f2377, %f2355; fma.rn.ftz.f32 %f2381, %f2380, %f2377, %f2357; mul.ftz.f32 %f2382, %f2377, %f2381; fma.rn.ftz.f32 %f2383, %f2382, %f2376, %f2376; add.ftz.f32 %f2384, %f2383, %f2383; sub.ftz.f32 %f2385, %f2362, %f2383; selp.f32 %f2386, %f2384, %f2385, %p195; setp.lt.ftz.f32 %p196, %f2371, 0f00000000; sub.ftz.f32 %f2387, %f2365, %f2386; selp.f32 %f2388, %f2387, %f2386, %p196; mul.ftz.f32 %f2389, %f2337, %f2338; neg.ftz.f32 %f2390, %f2389; max.ftz.f32 %f2391, %f2390, %f2341; min.ftz.f32 %f2392, %f2391, %f2276; abs.ftz.f32 %f2393, %f2392; sub.ftz.f32 %f2394, %f2276, %f2393; mul.ftz.f32 %f2395, %f2394, 0f3F000000; sqrt.approx.ftz.f32 %f2396, %f2395; setp.gt.ftz.f32 %p197, %f2393, 0f3F11EB85; selp.f32 %f2397, %f2396, %f2393, %p197; mul.ftz.f32 %f2398, %f2397, %f2397; fma.rn.ftz.f32 %f2399, %f2351, %f2398, %f2350; fma.rn.ftz.f32 %f2400, %f2399, %f2398, %f2353; fma.rn.ftz.f32 %f2401, %f2400, %f2398, %f2355; fma.rn.ftz.f32 %f2402, %f2401, %f2398, %f2357; mul.ftz.f32 %f2403, %f2398, %f2402; fma.rn.ftz.f32 %f2404, %f2403, %f2397, %f2397; add.ftz.f32 %f2405, %f2404, %f2404; sub.ftz.f32 %f2406, %f2362, %f2404; selp.f32 %f2407, %f2405, %f2406, %p197; setp.lt.ftz.f32 %p198, %f2392, 0f00000000; sub.ftz.f32 %f2408, %f2365, %f2407; selp.f32 %f2409, %f2408, %f2407, %p198; mul.ftz.f32 %f2410, %f2335, %f2338; neg.ftz.f32 %f2411, %f2410; max.ftz.f32 %f2412, %f2411, %f2341; min.ftz.f32 %f2413, %f2412, %f2276; abs.ftz.f32 %f2414, %f2413; sub.ftz.f32 %f2415, %f2276, %f2414; mul.ftz.f32 %f2416, %f2415, 0f3F000000; sqrt.approx.ftz.f32 %f2417, %f2416; setp.gt.ftz.f32 %p199, %f2414, 0f3F11EB85; selp.f32 %f2418, %f2417, %f2414, %p199; mul.ftz.f32 %f2419, %f2418, %f2418; fma.rn.ftz.f32 %f2420, %f2351, %f2419, %f2350; fma.rn.ftz.f32 %f2421, %f2420, %f2419, %f2353; fma.rn.ftz.f32 %f2422, %f2421, %f2419, %f2355; fma.rn.ftz.f32 %f2423, %f2422, %f2419, %f2357; mul.ftz.f32 %f2424, %f2419, %f2423; fma.rn.ftz.f32 %f2425, %f2424, %f2418, %f2418; add.ftz.f32 %f2426, %f2425, %f2425; sub.ftz.f32 %f2427, %f2362, %f2425; selp.f32 %f2428, %f2426, %f2427, %p199; setp.lt.ftz.f32 %p200, %f2413, 0f00000000; sub.ftz.f32 %f2429, %f2365, %f2428; selp.f32 %f2430, %f2429, %f2428, %p200; mov.f32 %f2431, 0f40C90FDB; sub.ftz.f32 %f2432, %f2431, %f2409; sub.ftz.f32 %f2433, %f2432, %f2430; add.ftz.f32 %f2434, %f2367, %f2388; sub.ftz.f32 %f380, %f2434, %f2433; setp.eq.ftz.f32 %p201, %f380, 0f00000000; @%p201 bra $L__BB19_146; rcp.approx.ftz.f32 %f3818, %f380; $L__BB19_146: add.ftz.f32 %f3820, %f3820, %f3818; $L__BB19_150: add.s64 %rd251, %rd251, 192; add.s32 %r323, %r323, 1; setp.lt.s32 %p203, %r323, %r71; @%p203 bra $L__BB19_138; $L__BB19_151: setp.lt.s32 %p204, %r325, 1; @%p204 bra $L__BB19_153; cvt.rn.f32.s32 %f2454, %r325; div.approx.ftz.f32 %f3822, %f3820, %f2454; $L__BB19_153: fma.rn.ftz.f32 %f3825, %f205, %f3822, %f3825; st.local.f32 [%rd1+48], %f3825; $L__BB19_154: setp.eq.ftz.f32 %p205, %f206, 0f00000000; setp.eq.s32 %p206, %r322, 1; or.pred %p207, %p205, %p206; @%p207 bra $L__BB19_158; ld.const.v4.f32 {%f2456, %f2457, %f2458, %f2459}, [kernel_params+1008]; ld.const.f32 %f2463, [kernel_params+1020]; cos.approx.ftz.f32 %f395, %f2463; mul.ftz.f32 %f2464, %f3831, %f2457; fma.rn.ftz.f32 %f2465, %f3830, %f2456, %f2464; fma.rn.ftz.f32 %f2466, %f3832, %f2458, %f2465; setp.leu.ftz.f32 %p208, %f2466, %f395; mov.f32 %f3824, 0f00000000; @%p208 bra $L__BB19_157; mov.f32 %f2467, 0f3F800000; sub.ftz.f32 %f2468, %f2467, %f395; mov.f32 %f2469, 0f3E22F983; div.approx.ftz.f32 %f3824, %f2469, %f2468; $L__BB19_157: fma.rn.ftz.f32 %f3825, %f206, %f3824, %f3825; st.local.f32 [%rd1+48], %f3825; $L__BB19_158: setp.eq.ftz.f32 %p209, %f207, 0f00000000; setp.eq.s32 %p210, %r322, 2; or.pred %p211, %p209, %p210; @%p211 bra $L__BB19_171; setp.eq.ftz.f32 %p212, %f3830, 0f00000000; setp.eq.ftz.f32 %p213, %f3831, 0f00000000; and.pred %p214, %p212, %p213; setp.eq.ftz.f32 %p215, %f3832, 0f00000000; and.pred %p216, %p215, %p214; mov.f32 %f3829, 0f00000000; mov.f32 %f3827, %f3829; mov.f32 %f3828, %f3829; @%p216 bra $L__BB19_166; abs.ftz.f32 %f403, %f3830; setp.eq.ftz.f32 %p217, %f403, 0f00000000; abs.ftz.f32 %f404, %f3831; setp.eq.ftz.f32 %p218, %f404, 0f00000000; and.pred %p219, %p217, %p218; mov.b32 %r78, %f3830; mov.b32 %r195, %f3831; and.b32 %r79, %r195, -2147483648; @%p219 bra $L__BB19_164; bra.uni $L__BB19_161; $L__BB19_164: shr.s32 %r200, %r78, 31; and.b32 %r201, %r200, 1078530011; or.b32 %r202, %r201, %r79; mov.b32 %f3826, %r202; bra.uni $L__BB19_165; $L__BB19_254: mov.f32 %f3488, 0fBF800000; max.ftz.f32 %f3489, %f859, %f3488; min.ftz.f32 %f3491, %f3489, %f3384; abs.ftz.f32 %f3492, %f3491; sub.ftz.f32 %f3493, %f3384, %f3492; mul.ftz.f32 %f3494, %f3493, 0f3F000000; sqrt.approx.ftz.f32 %f3495, %f3494; setp.gt.ftz.f32 %p304, %f3492, 0f3F11EB85; selp.f32 %f3496, %f3495, %f3492, %p304; mul.ftz.f32 %f3497, %f3496, %f3496; mov.f32 %f3498, 0f3C94D2E9; mov.f32 %f3499, 0f3D53F941; fma.rn.ftz.f32 %f3500, %f3499, %f3497, %f3498; mov.f32 %f3501, 0f3D3F841F; fma.rn.ftz.f32 %f3502, %f3500, %f3497, %f3501; mov.f32 %f3503, 0f3D994929; fma.rn.ftz.f32 %f3504, %f3502, %f3497, %f3503; mov.f32 %f3505, 0f3E2AAB94; fma.rn.ftz.f32 %f3506, %f3504, %f3497, %f3505; mul.ftz.f32 %f3507, %f3497, %f3506; fma.rn.ftz.f32 %f3508, %f3507, %f3496, %f3496; add.ftz.f32 %f3509, %f3508, %f3508; mov.f32 %f3510, 0f3FC90FDB; sub.ftz.f32 %f3511, %f3510, %f3508; selp.f32 %f3512, %f3509, %f3511, %p304; setp.lt.ftz.f32 %p305, %f3491, 0f00000000; sub.ftz.f32 %f3514, %f3254, %f3512; selp.f32 %f3515, %f3514, %f3512, %p305; mul.ftz.f32 %f3516, %f3475, %f859; sub.ftz.f32 %f3517, %f3479, %f3516; mul.ftz.f32 %f3518, %f3476, %f859; sub.ftz.f32 %f3519, %f3480, %f3518; mul.ftz.f32 %f3520, %f3477, %f859; sub.ftz.f32 %f3521, %f3481, %f3520; mul.ftz.f32 %f3522, %f3478, %f859; sub.ftz.f32 %f3523, %f3482, %f3522; mul.ftz.f32 %f3524, %f3519, %f3519; fma.rn.ftz.f32 %f3525, %f3517, %f3517, %f3524; mul.ftz.f32 %f3526, %f3523, %f3523; fma.rn.ftz.f32 %f3527, %f3521, %f3521, %f3526; add.ftz.f32 %f3528, %f3525, %f3527; rsqrt.approx.ftz.f32 %f3529, %f3528; mul.ftz.f32 %f3530, %f3517, %f3529; mul.ftz.f32 %f3531, %f3519, %f3529; mul.ftz.f32 %f3532, %f3521, %f3529; mul.ftz.f32 %f3533, %f3523, %f3529; mul.ftz.f32 %f3534, %f850, %f3515; cos.approx.ftz.f32 %f3535, %f3534; sin.approx.ftz.f32 %f3536, %f3534; mul.ftz.f32 %f3537, %f3530, %f3536; mul.ftz.f32 %f3538, %f3531, %f3536; mul.ftz.f32 %f3539, %f3532, %f3536; mul.ftz.f32 %f3540, %f3533, %f3536; fma.rn.ftz.f32 %f3938, %f3478, %f3535, %f3540; fma.rn.ftz.f32 %f3937, %f3477, %f3535, %f3539; fma.rn.ftz.f32 %f3936, %f3476, %f3535, %f3538; fma.rn.ftz.f32 %f3935, %f3475, %f3535, %f3537; $L__BB19_256: ld.global.v4.f32 {%f3555, %f3556, %f3557, %f3558}, [%rd95+16]; ld.global.v4.f32 {%f3563, %f3564, %f3565, %f3566}, [%rd95+80]; mul.ftz.f32 %f3571, %f850, %f3563; mul.ftz.f32 %f3572, %f850, %f3564; mul.ftz.f32 %f3573, %f850, %f3565; mul.ftz.f32 %f3574, %f850, %f3566; fma.rn.ftz.f32 %f3575, %f860, %f3558, %f3574; ld.global.v4.f32 {%f3576, %f3577, %f3578, %f3579}, [%rd95+32]; ld.global.v4.f32 {%f3584, %f3585, %f3586, %f3587}, [%rd95+96]; mul.ftz.f32 %f3592, %f850, %f3584; mul.ftz.f32 %f3593, %f850, %f3585; mul.ftz.f32 %f3594, %f850, %f3586; mul.ftz.f32 %f3595, %f850, %f3587; fma.rn.ftz.f32 %f3596, %f860, %f3576, %f3592; fma.rn.ftz.f32 %f3597, %f860, %f3577, %f3593; fma.rn.ftz.f32 %f3598, %f860, %f3578, %f3594; fma.rn.ftz.f32 %f3599, %f860, %f3579, %f3595; ld.global.v4.f32 {%f3600, %f3601, %f3602, %f3603}, [%rd95+48]; ld.global.v4.f32 {%f3608, %f3609, %f3610, %f3611}, [%rd95+112]; mul.ftz.f32 %f3616, %f850, %f3608; mul.ftz.f32 %f3617, %f850, %f3609; mul.ftz.f32 %f3618, %f850, %f3610; mul.ftz.f32 %f3619, %f850, %f3611; fma.rn.ftz.f32 %f3620, %f860, %f3600, %f3616; fma.rn.ftz.f32 %f3621, %f860, %f3601, %f3617; fma.rn.ftz.f32 %f3622, %f860, %f3602, %f3618; fma.rn.ftz.f32 %f3623, %f860, %f3603, %f3619; mul.ftz.f32 %f3624, %f3938, 0f3FB504F3; mul.ftz.f32 %f3625, %f3935, 0f3FB504F3; mul.ftz.f32 %f3626, %f3936, 0f3FB504F3; mul.ftz.f32 %f3627, %f3937, 0f3FB504F3; mul.ftz.f32 %f3628, %f3624, %f3625; mul.ftz.f32 %f3629, %f3624, %f3626; mul.ftz.f32 %f3630, %f3624, %f3627; mul.ftz.f32 %f3631, %f3625, %f3625; mul.ftz.f32 %f3632, %f3625, %f3626; mul.ftz.f32 %f3633, %f3625, %f3627; mul.ftz.f32 %f3634, %f3626, %f3626; mul.ftz.f32 %f3635, %f3626, %f3627; mul.ftz.f32 %f3636, %f3627, %f3627; sub.ftz.f32 %f3638, %f3384, %f3634; sub.ftz.f32 %f3639, %f3638, %f3636; sub.ftz.f32 %f3640, %f3632, %f3630; add.ftz.f32 %f3641, %f3629, %f3633; add.ftz.f32 %f3642, %f3630, %f3632; sub.ftz.f32 %f3643, %f3384, %f3631; sub.ftz.f32 %f3644, %f3643, %f3636; sub.ftz.f32 %f3645, %f3635, %f3628; sub.ftz.f32 %f3646, %f3633, %f3629; add.ftz.f32 %f3647, %f3628, %f3635; sub.ftz.f32 %f3648, %f3643, %f3634; mul.ftz.f32 %f3649, %f3640, %f3598; fma.rn.ftz.f32 %f3650, %f3639, %f3575, %f3649; mul.ftz.f32 %f3651, %f3640, %f3599; fma.rn.ftz.f32 %f3652, %f3639, %f3596, %f3651; mul.ftz.f32 %f3653, %f3640, %f3620; fma.rn.ftz.f32 %f3654, %f3639, %f3597, %f3653; mul.ftz.f32 %f3655, %f3644, %f3598; fma.rn.ftz.f32 %f3656, %f3642, %f3575, %f3655; mul.ftz.f32 %f3657, %f3644, %f3599; fma.rn.ftz.f32 %f3658, %f3642, %f3596, %f3657; mul.ftz.f32 %f3659, %f3644, %f3620; fma.rn.ftz.f32 %f3660, %f3642, %f3597, %f3659; mul.ftz.f32 %f3661, %f3647, %f3598; fma.rn.ftz.f32 %f3662, %f3646, %f3575, %f3661; mul.ftz.f32 %f3663, %f3647, %f3599; fma.rn.ftz.f32 %f3664, %f3646, %f3596, %f3663; mul.ftz.f32 %f3665, %f3647, %f3620; fma.rn.ftz.f32 %f3666, %f3646, %f3597, %f3665; fma.rn.ftz.f32 %f3945, %f3645, %f3623, %f3660; fma.rn.ftz.f32 %f3944, %f3645, %f3622, %f3658; fma.rn.ftz.f32 %f3943, %f3645, %f3621, %f3656; fma.rn.ftz.f32 %f3949, %f3648, %f3623, %f3666; fma.rn.ftz.f32 %f3948, %f3648, %f3622, %f3664; fma.rn.ftz.f32 %f3947, %f3648, %f3621, %f3662; fma.rn.ftz.f32 %f3941, %f3641, %f3623, %f3654; fma.rn.ftz.f32 %f3940, %f3641, %f3622, %f3652; fma.rn.ftz.f32 %f3939, %f3641, %f3621, %f3650; fma.rn.ftz.f32 %f3950, %f860, %f3557, %f3573; fma.rn.ftz.f32 %f3946, %f860, %f3556, %f3572; fma.rn.ftz.f32 %f3942, %f860, %f3555, %f3571; $L__BB19_258: mul.ftz.f32 %f3679, %f3952, %f3940; fma.rn.ftz.f32 %f3680, %f3951, %f3939, %f3679; fma.rn.ftz.f32 %f3681, %f3953, %f3941, %f3680; add.ftz.f32 %f909, %f3942, %f3681; mul.ftz.f32 %f3682, %f3952, %f3944; fma.rn.ftz.f32 %f3683, %f3951, %f3943, %f3682; fma.rn.ftz.f32 %f3684, %f3953, %f3945, %f3683; add.ftz.f32 %f910, %f3946, %f3684; mul.ftz.f32 %f3685, %f3952, %f3948; fma.rn.ftz.f32 %f3686, %f3951, %f3947, %f3685; fma.rn.ftz.f32 %f3687, %f3953, %f3949, %f3686; add.ftz.f32 %f3953, %f3950, %f3687; mul.ftz.f32 %f3688, %f3955, %f3940; fma.rn.ftz.f32 %f3689, %f3954, %f3939, %f3688; fma.rn.ftz.f32 %f3690, %f3956, %f3941, %f3689; add.ftz.f32 %f912, %f3942, %f3690; mul.ftz.f32 %f3691, %f3955, %f3944; fma.rn.ftz.f32 %f3692, %f3954, %f3943, %f3691; fma.rn.ftz.f32 %f3693, %f3956, %f3945, %f3692; add.ftz.f32 %f913, %f3946, %f3693; mul.ftz.f32 %f3694, %f3955, %f3948; fma.rn.ftz.f32 %f3695, %f3954, %f3947, %f3694; fma.rn.ftz.f32 %f3696, %f3956, %f3949, %f3695; add.ftz.f32 %f3956, %f3950, %f3696; mul.ftz.f32 %f3697, %f3958, %f3940; fma.rn.ftz.f32 %f3698, %f3957, %f3939, %f3697; fma.rn.ftz.f32 %f3699, %f3959, %f3941, %f3698; add.ftz.f32 %f915, %f3942, %f3699; mul.ftz.f32 %f3700, %f3958, %f3944; fma.rn.ftz.f32 %f3701, %f3957, %f3943, %f3700; fma.rn.ftz.f32 %f3702, %f3959, %f3945, %f3701; add.ftz.f32 %f916, %f3946, %f3702; mul.ftz.f32 %f3703, %f3958, %f3948; fma.rn.ftz.f32 %f3704, %f3957, %f3947, %f3703; fma.rn.ftz.f32 %f3705, %f3959, %f3949, %f3704; add.ftz.f32 %f3959, %f3950, %f3705; mov.f32 %f3951, %f909; mov.f32 %f3952, %f910; mov.f32 %f3954, %f912; mov.f32 %f3955, %f913; mov.f32 %f3957, %f915; mov.f32 %f3958, %f916; $L__BB19_259: sub.ftz.f32 %f3706, %f3953, %f3956; sub.ftz.f32 %f3707, %f3958, %f3955; mul.ftz.f32 %f3708, %f3706, %f3707; sub.ftz.f32 %f3709, %f3952, %f3955; sub.ftz.f32 %f3710, %f3959, %f3956; mul.ftz.f32 %f3711, %f3709, %f3710; sub.ftz.f32 %f3712, %f3708, %f3711; sub.ftz.f32 %f3713, %f3951, %f3954; mul.ftz.f32 %f3714, %f3713, %f3710; sub.ftz.f32 %f3715, %f3957, %f3954; mul.ftz.f32 %f3716, %f3706, %f3715; sub.ftz.f32 %f3717, %f3714, %f3716; mul.ftz.f32 %f3718, %f3709, %f3715; mul.ftz.f32 %f3719, %f3713, %f3707; sub.ftz.f32 %f3720, %f3718, %f3719; mul.ftz.f32 %f3721, %f3717, %f3717; fma.rn.ftz.f32 %f3722, %f3712, %f3712, %f3721; fma.rn.ftz.f32 %f3723, %f3720, %f3720, %f3722; sqrt.approx.ftz.f32 %f3724, %f3723; mul.ftz.f32 %f3960, %f3724, 0f3F000000; $L__BB19_260: ld.const.f32 %f3725, [kernel_params+1500]; mul.ftz.f32 %f3726, %f3960, %f3725; div.approx.ftz.f32 %f3961, %f3726, %f772; st.local.f32 [%rd1+48], %f3961; bra.uni $L__BB19_263; $L__BB19_161: setp.eq.ftz.f32 %p220, %f403, 0f7F800000; setp.eq.ftz.f32 %p221, %f404, 0f7F800000; and.pred %p222, %p220, %p221; @%p222 bra $L__BB19_163; bra.uni $L__BB19_162; $L__BB19_163: setp.lt.s32 %p226, %r78, 0; selp.b32 %r198, 1075235812, 1061752795, %p226; or.b32 %r199, %r198, %r79; mov.b32 %f3826, %r199; bra.uni $L__BB19_165; $L__BB19_58: setp.lt.s32 %p86, %r17, 0; min.ftz.f32 %f1476, %f185, %f184; max.ftz.f32 %f1477, %f185, %f184; div.full.ftz.f32 %f1478, %f1476, %f1477; mul.rn.ftz.f32 %f1479, %f1478, %f1478; mov.f32 %f1480, 0fC0B59883; mov.f32 %f1481, 0fBF52C7EA; fma.rn.ftz.f32 %f1482, %f1479, %f1481, %f1480; mov.f32 %f1483, 0fC0D21907; fma.rn.ftz.f32 %f1484, %f1482, %f1479, %f1483; mul.ftz.f32 %f1485, %f1479, %f1484; mul.ftz.f32 %f1486, %f1478, %f1485; add.ftz.f32 %f1487, %f1479, 0f41355DC0; mov.f32 %f1488, 0f41E6BD60; fma.rn.ftz.f32 %f1489, %f1487, %f1479, %f1488; mov.f32 %f1490, 0f419D92C8; fma.rn.ftz.f32 %f1491, %f1489, %f1479, %f1490; rcp.approx.ftz.f32 %f1492, %f1491; fma.rn.ftz.f32 %f1493, %f1486, %f1492, %f1478; mov.f32 %f1494, 0f3FC90FDB; sub.ftz.f32 %f1495, %f1494, %f1493; setp.gt.ftz.f32 %p87, %f185, %f184; selp.f32 %f1496, %f1495, %f1493, %p87; mov.f32 %f1497, 0f40490FDB; sub.ftz.f32 %f1498, %f1497, %f1496; selp.f32 %f1499, %f1498, %f1496, %p86; mov.b32 %r132, %f1499; or.b32 %r133, %r18, %r132; mov.b32 %f1500, %r133; add.ftz.f32 %f1501, %f184, %f185; setp.le.ftz.f32 %p88, %f1501, 0f7F800000; selp.f32 %f3782, %f1500, %f1501, %p88; $L__BB19_61: fma.rn.ftz.f32 %f3784, %f3782, 0fBE22F983, 0f3F000000; $L__BB19_62: sqrt.approx.ftz.f32 %f1502, %f183; div.approx.ftz.f32 %f1503, %f175, %f1502; mov.f32 %f1504, 0fBF800000; max.ftz.f32 %f1505, %f1503, %f1504; mov.f32 %f1506, 0f3F800000; min.ftz.f32 %f1507, %f1505, %f1506; abs.ftz.f32 %f1508, %f1507; sub.ftz.f32 %f1509, %f1506, %f1508; mul.ftz.f32 %f1510, %f1509, 0f3F000000; sqrt.approx.ftz.f32 %f1511, %f1510; setp.gt.ftz.f32 %p90, %f1508, 0f3F11EB85; selp.f32 %f1512, %f1511, %f1508, %p90; mul.ftz.f32 %f1513, %f1512, %f1512; mov.f32 %f1514, 0f3C94D2E9; mov.f32 %f1515, 0f3D53F941; fma.rn.ftz.f32 %f1516, %f1515, %f1513, %f1514; mov.f32 %f1517, 0f3D3F841F; fma.rn.ftz.f32 %f1518, %f1516, %f1513, %f1517; mov.f32 %f1519, 0f3D994929; fma.rn.ftz.f32 %f1520, %f1518, %f1513, %f1519; mov.f32 %f1521, 0f3E2AAB94; fma.rn.ftz.f32 %f1522, %f1520, %f1513, %f1521; mul.ftz.f32 %f1523, %f1513, %f1522; fma.rn.ftz.f32 %f1524, %f1523, %f1512, %f1512; add.ftz.f32 %f1525, %f1524, %f1524; mov.f32 %f1526, 0f3FC90FDB; sub.ftz.f32 %f1527, %f1526, %f1524; selp.f32 %f1528, %f1525, %f1527, %p90; setp.lt.ftz.f32 %p91, %f1507, 0f00000000; mov.f32 %f1529, 0f40490FDB; sub.ftz.f32 %f1530, %f1529, %f1528; selp.f32 %f1531, %f1530, %f1528, %p91; fma.rn.ftz.f32 %f3785, %f1531, 0fBEA2F983, 0f3F800000; $L__BB19_63: st.local.f32 [%rd1+40], %f3784; st.local.f32 [%rd1+44], %f3785; mul.ftz.f32 %f1533, %f160, %f171; neg.ftz.f32 %f1534, %f1533; mul.ftz.f32 %f1535, %f159, %f170; sub.ftz.f32 %f1536, %f1534, %f1535; mul.ftz.f32 %f1537, %f161, %f172; sub.ftz.f32 %f195, %f1536, %f1537; setp.le.ftz.f32 %p92, %f195, 0f00000000; @%p92 bra $L__BB19_65; mul.ftz.f32 %f1538, %f169, %f169; div.approx.ftz.f32 %f3786, %f1538, %f195; $L__BB19_65: mul.ftz.f32 %f3836, %f3786, %f168; st.local.f32 [%rd1+48], %f3836; bra.uni $L__BB19_175; $L__BB19_39: setp.eq.ftz.f32 %p56, %f139, 0f7F800000; setp.eq.ftz.f32 %p57, %f140, 0f7F800000; and.pred %p58, %p56, %p57; @%p58 bra $L__BB19_41; bra.uni $L__BB19_40; $L__BB19_41: setp.lt.s32 %p62, %r15, 0; selp.b32 %r126, 1075235812, 1061752795, %p62; or.b32 %r127, %r126, %r16; mov.b32 %f3773, %r127; bra.uni $L__BB19_43; $L__BB19_162: setp.lt.s32 %p223, %r78, 0; min.ftz.f32 %f2472, %f404, %f403; max.ftz.f32 %f2473, %f404, %f403; div.full.ftz.f32 %f2474, %f2472, %f2473; mul.rn.ftz.f32 %f2475, %f2474, %f2474; mov.f32 %f2476, 0fC0B59883; mov.f32 %f2477, 0fBF52C7EA; fma.rn.ftz.f32 %f2478, %f2475, %f2477, %f2476; mov.f32 %f2479, 0fC0D21907; fma.rn.ftz.f32 %f2480, %f2478, %f2475, %f2479; mul.ftz.f32 %f2481, %f2475, %f2480; mul.ftz.f32 %f2482, %f2474, %f2481; add.ftz.f32 %f2483, %f2475, 0f41355DC0; mov.f32 %f2484, 0f41E6BD60; fma.rn.ftz.f32 %f2485, %f2483, %f2475, %f2484; mov.f32 %f2486, 0f419D92C8; fma.rn.ftz.f32 %f2487, %f2485, %f2475, %f2486; rcp.approx.ftz.f32 %f2488, %f2487; fma.rn.ftz.f32 %f2489, %f2482, %f2488, %f2474; mov.f32 %f2490, 0f3FC90FDB; sub.ftz.f32 %f2491, %f2490, %f2489; setp.gt.ftz.f32 %p224, %f404, %f403; selp.f32 %f2492, %f2491, %f2489, %p224; mov.f32 %f2493, 0f40490FDB; sub.ftz.f32 %f2494, %f2493, %f2492; selp.f32 %f2495, %f2494, %f2492, %p223; mov.b32 %r196, %f2495; or.b32 %r197, %r79, %r196; mov.b32 %f2496, %r197; add.ftz.f32 %f2497, %f403, %f404; setp.le.ftz.f32 %p225, %f2497, 0f7F800000; selp.f32 %f3826, %f2496, %f2497, %p225; $L__BB19_165: add.ftz.f32 %f2498, %f3826, 0fC0490FDB; mov.f32 %f2499, 0fC0490FDB; mov.f32 %f2500, 0f40490FDB; mov.f32 %f2501, 0fC0C90FDB; mul.ftz.f32 %f2502, %f3831, %f3831; fma.rn.ftz.f32 %f2503, %f3830, %f3830, %f2502; fma.rn.ftz.f32 %f2504, %f3832, %f3832, %f2503; sqrt.approx.ftz.f32 %f2505, %f2504; div.approx.ftz.f32 %f2506, %f3832, %f2505; abs.ftz.f32 %f2507, %f2506; mov.f32 %f2508, 0f3F800000; sub.ftz.f32 %f2509, %f2508, %f2507; mul.ftz.f32 %f2510, %f2509, 0f3F000000; sqrt.approx.ftz.f32 %f2511, %f2510; setp.gt.ftz.f32 %p227, %f2507, 0f3F11EB85; selp.f32 %f2512, %f2511, %f2507, %p227; mul.ftz.f32 %f2513, %f2512, %f2512; mov.f32 %f2514, 0f3C94D2E9; mov.f32 %f2515, 0f3D53F941; fma.rn.ftz.f32 %f2516, %f2515, %f2513, %f2514; mov.f32 %f2517, 0f3D3F841F; fma.rn.ftz.f32 %f2518, %f2516, %f2513, %f2517; mov.f32 %f2519, 0f3D994929; fma.rn.ftz.f32 %f2520, %f2518, %f2513, %f2519; mov.f32 %f2521, 0f3E2AAB94; fma.rn.ftz.f32 %f2522, %f2520, %f2513, %f2521; mul.ftz.f32 %f2523, %f2513, %f2522; fma.rn.ftz.f32 %f2524, %f2523, %f2512, %f2512; add.ftz.f32 %f2525, %f2524, %f2524; mov.f32 %f2526, 0f3FC90FDB; sub.ftz.f32 %f2527, %f2526, %f2524; selp.f32 %f2528, %f2525, %f2527, %p227; setp.lt.ftz.f32 %p228, %f2506, 0f00000000; sub.ftz.f32 %f2529, %f2500, %f2528; selp.f32 %f2530, %f2529, %f2528, %p228; add.ftz.f32 %f2531, %f2530, 0fC0490FDB; div.approx.ftz.f32 %f3828, %f2531, %f2499; div.approx.ftz.f32 %f3827, %f2498, %f2501; $L__BB19_166: mul.ftz.f32 %f2533, %f3828, 0f40490FDB; sin.approx.ftz.f32 %f414, %f2533; setp.eq.ftz.f32 %p229, %f414, 0f00000000; @%p229 bra $L__BB19_169; ld.const.v2.u32 {%r203, %r204}, [kernel_params+1064]; add.s32 %r207, %r203, 1; cvt.rn.f32.s32 %f2535, %r203; mul.ftz.f32 %f2536, %f3827, %f2535; cvt.rzi.ftz.s32.f32 %r208, %f2536; max.s32 %r209, %r208, 0; add.s32 %r210, %r203, -1; min.s32 %r80, %r209, %r210; cvt.rn.f32.s32 %f2537, %r204; mul.ftz.f32 %f2538, %f3828, %f2537; cvt.rzi.ftz.s32.f32 %r211, %f2538; max.s32 %r212, %r211, 0; add.s32 %r213, %r204, -1; min.s32 %r81, %r212, %r213; ld.const.u64 %rd127, [kernel_params+2384]; cvta.to.global.u64 %rd25, %rd127; mul.lo.s32 %r82, %r81, %r207; add.s32 %r214, %r82, %r203; mul.wide.s32 %rd128, %r214, 8; add.s64 %rd129, %rd25, %rd128; ld.const.u64 %rd130, [kernel_params+2376]; cvta.to.global.u64 %rd26, %rd130; mul.wide.s32 %rd131, %r204, 8; add.s64 %rd132, %rd26, %rd131; mul.ftz.f32 %f2539, %f414, 0f419DE9E7; ld.global.f32 %f2540, [%rd129]; mul.ftz.f32 %f2541, %f2539, %f2540; ld.global.f32 %f2542, [%rd132]; mul.ftz.f32 %f415, %f2541, %f2542; setp.eq.ftz.f32 %p230, %f415, 0f00000000; @%p230 bra $L__BB19_169; add.s32 %r215, %r82, %r80; mul.wide.s32 %rd133, %r215, 8; add.s64 %rd134, %rd25, %rd133; mul.wide.s32 %rd135, %r81, 8; add.s64 %rd136, %rd26, %rd135; ld.global.f32 %f2543, [%rd136]; ld.global.f32 %f2544, [%rd134]; mul.ftz.f32 %f2545, %f2544, %f2543; div.approx.ftz.f32 %f3829, %f2545, %f415; $L__BB19_169: fma.rn.ftz.f32 %f2546, %f207, %f3829, %f3825; st.local.f32 [%rd1+48], %f2546; bra.uni $L__BB19_171; $L__BB19_40: setp.lt.s32 %p59, %r15, 0; min.ftz.f32 %f1341, %f140, %f139; max.ftz.f32 %f1342, %f140, %f139; div.full.ftz.f32 %f1343, %f1341, %f1342; mul.rn.ftz.f32 %f1344, %f1343, %f1343; mov.f32 %f1345, 0fC0B59883; mov.f32 %f1346, 0fBF52C7EA; fma.rn.ftz.f32 %f1347, %f1344, %f1346, %f1345; mov.f32 %f1348, 0fC0D21907; fma.rn.ftz.f32 %f1349, %f1347, %f1344, %f1348; mul.ftz.f32 %f1350, %f1344, %f1349; mul.ftz.f32 %f1351, %f1343, %f1350; add.ftz.f32 %f1352, %f1344, 0f41355DC0; mov.f32 %f1353, 0f41E6BD60; fma.rn.ftz.f32 %f1354, %f1352, %f1344, %f1353; mov.f32 %f1355, 0f419D92C8; fma.rn.ftz.f32 %f1356, %f1354, %f1344, %f1355; rcp.approx.ftz.f32 %f1357, %f1356; fma.rn.ftz.f32 %f1358, %f1351, %f1357, %f1343; mov.f32 %f1359, 0f3FC90FDB; sub.ftz.f32 %f1360, %f1359, %f1358; setp.gt.ftz.f32 %p60, %f140, %f139; selp.f32 %f1361, %f1360, %f1358, %p60; mov.f32 %f1362, 0f40490FDB; sub.ftz.f32 %f1363, %f1362, %f1361; selp.f32 %f1364, %f1363, %f1361, %p59; mov.b32 %r124, %f1364; or.b32 %r125, %r16, %r124; mov.b32 %f1365, %r125; add.ftz.f32 %f1366, %f139, %f140; setp.le.ftz.f32 %p61, %f1366, 0f7F800000; selp.f32 %f3773, %f1365, %f1366, %p61; $L__BB19_43: fma.rn.ftz.f32 %f3775, %f3773, 0fBE22F983, 0f3F000000; $L__BB19_44: sqrt.approx.ftz.f32 %f1367, %f138; div.approx.ftz.f32 %f1368, %f137, %f1367; mov.f32 %f1369, 0fBF800000; max.ftz.f32 %f1370, %f1368, %f1369; mov.f32 %f1371, 0f3F800000; min.ftz.f32 %f1372, %f1370, %f1371; abs.ftz.f32 %f1373, %f1372; sub.ftz.f32 %f1374, %f1371, %f1373; mul.ftz.f32 %f1375, %f1374, 0f3F000000; sqrt.approx.ftz.f32 %f1376, %f1375; setp.gt.ftz.f32 %p63, %f1373, 0f3F11EB85; selp.f32 %f1377, %f1376, %f1373, %p63; mul.ftz.f32 %f1378, %f1377, %f1377; mov.f32 %f1379, 0f3C94D2E9; mov.f32 %f1380, 0f3D53F941; fma.rn.ftz.f32 %f1381, %f1380, %f1378, %f1379; mov.f32 %f1382, 0f3D3F841F; fma.rn.ftz.f32 %f1383, %f1381, %f1378, %f1382; mov.f32 %f1384, 0f3D994929; fma.rn.ftz.f32 %f1385, %f1383, %f1378, %f1384; mov.f32 %f1386, 0f3E2AAB94; fma.rn.ftz.f32 %f1387, %f1385, %f1378, %f1386; mul.ftz.f32 %f1388, %f1378, %f1387; fma.rn.ftz.f32 %f1389, %f1388, %f1377, %f1377; add.ftz.f32 %f1390, %f1389, %f1389; mov.f32 %f1391, 0f3FC90FDB; sub.ftz.f32 %f1392, %f1391, %f1389; selp.f32 %f1393, %f1390, %f1392, %p63; setp.lt.ftz.f32 %p64, %f1372, 0f00000000; mov.f32 %f1394, 0f40490FDB; sub.ftz.f32 %f1395, %f1394, %f1393; selp.f32 %f1396, %f1395, %f1393, %p64; fma.rn.ftz.f32 %f3776, %f1396, 0fBEA2F983, 0f3F800000; $L__BB19_45: st.local.f32 [%rd1+40], %f3775; st.local.f32 [%rd1+44], %f3776; mul.ftz.f32 %f1398, %f122, %f133; neg.ftz.f32 %f1399, %f1398; mul.ftz.f32 %f1400, %f121, %f132; sub.ftz.f32 %f1401, %f1399, %f1400; mul.ftz.f32 %f1402, %f123, %f134; sub.ftz.f32 %f150, %f1401, %f1402; setp.le.ftz.f32 %p65, %f150, 0f00000000; @%p65 bra $L__BB19_47; mul.ftz.f32 %f1403, %f131, %f131; div.approx.ftz.f32 %f3777, %f1403, %f150; $L__BB19_47: mul.ftz.f32 %f3836, %f3777, %f130; st.local.f32 [%rd1+48], %f3836; bra.uni $L__BB19_175; } .func (.param .b32 func_retval0) _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj( .param .b64 _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0, .param .b64 _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1, .param .b32 _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2 ) { .local .align 16 .b8 __local_depot20[768]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<92>; .reg .b16 %rs<27>; .reg .f32 %f<508>; .reg .b32 %r<181>; .reg .b64 %rd<111>; mov.u64 %SPL, __local_depot20; ld.param.u64 %rd32, [_Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0]; ld.param.u64 %rd33, [_Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1]; ld.param.u32 %r69, [_Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2]; cvta.to.local.u64 %rd3, %rd33; add.u64 %rd1, %SPL, 0; mov.u32 %r73, 1985229328; st.local.u32 [%rd1], %r73; ld.const.u32 %r160, [kernel_params+1104]; cvta.to.local.u64 %rd2, %rd32; ld.local.f32 %f493, [%rd2]; ld.local.f32 %f494, [%rd2+4]; ld.local.f32 %f495, [%rd2+8]; ld.local.f32 %f149, [%rd2+12]; abs.ftz.f32 %f150, %f149; setp.gt.ftz.f32 %p5, %f150, 0f17800000; mov.b32 %r74, %f149; and.b32 %r75, %r74, -2147483648; or.b32 %r76, %r75, 394264576; mov.b32 %f151, %r76; selp.f32 %f488, %f149, %f151, %p5; ld.local.f32 %f152, [%rd2+16]; abs.ftz.f32 %f153, %f152; setp.gt.ftz.f32 %p6, %f153, 0f17800000; mov.b32 %r77, %f152; and.b32 %r78, %r77, -2147483648; or.b32 %r79, %r78, 394264576; mov.b32 %f154, %r79; selp.f32 %f487, %f152, %f154, %p6; ld.local.f32 %f155, [%rd2+20]; abs.ftz.f32 %f156, %f155; setp.gt.ftz.f32 %p7, %f156, 0f17800000; mov.b32 %r80, %f155; and.b32 %r81, %r80, -2147483648; or.b32 %r82, %r81, 394264576; mov.b32 %f157, %r82; selp.f32 %f492, %f155, %f157, %p7; rcp.approx.ftz.f32 %f491, %f488; rcp.approx.ftz.f32 %f490, %f487; rcp.approx.ftz.f32 %f489, %f492; ld.local.f32 %f10, [%rd2+24]; ld.local.f32 %f478, [%rd2+28]; st.local.f32 [%rd3], %f478; mov.u32 %r159, 0; st.local.u32 [%rd3+4], %r159; st.local.u32 [%rd3+8], %r159; mov.u32 %r173, -1; st.local.u32 [%rd3+12], %r173; st.local.u32 [%rd3+16], %r173; ld.const.v2.u64 {%rd35, %rd36}, [kernel_params+2096]; cvta.to.global.u64 %rd4, %rd35; cvta.to.global.u64 %rd5, %rd36; ld.const.v2.u64 {%rd39, %rd40}, [kernel_params+2128]; cvta.to.global.u64 %rd6, %rd40; ld.const.u64 %rd43, [kernel_params+2160]; cvta.to.global.u64 %rd7, %rd43; ld.const.v2.u64 {%rd44, %rd45}, [kernel_params+2144]; cvta.to.global.u64 %rd8, %rd44; cvta.to.global.u64 %rd9, %rd39; ld.const.u64 %rd48, [kernel_params+2232]; cvta.to.global.u64 %rd10, %rd48; ld.const.u64 %rd49, [kernel_params+2256]; cvta.to.global.u64 %rd11, %rd49; ld.const.u32 %r2, [kernel_params+1128]; ld.const.v2.u64 {%rd50, %rd51}, [kernel_params+2112]; cvta.to.global.u64 %rd12, %rd50; cvta.to.global.u64 %rd13, %rd45; and.b32 %r3, %r69, 256; cvta.to.global.u64 %rd14, %rd51; ld.const.u64 %rd54, [kernel_params+2296]; cvta.to.global.u64 %rd15, %rd54; ld.const.u64 %rd55, [kernel_params+2312]; cvta.to.global.u64 %rd16, %rd55; ld.const.u64 %rd56, [kernel_params+2344]; cvta.to.global.u64 %rd17, %rd56; setp.eq.s32 %p30, %r2, 0; mov.u16 %rs26, 1; mov.pred %p63, -1; mov.pred %p75, 0; mov.u32 %r177, %r173; bra.uni $L__BB20_1; $L__BB20_9: mul.wide.s32 %rd60, %r160, 16; xor.b64 %rd61, %rd60, -16; add.s64 %rd62, %rd5, %rd61; ld.global.v4.u32 {%r164, %r95, %r96, %r97}, [%rd62]; setp.gt.s32 %p25, %r164, -1; mul.wide.s32 %rd63, %r23, 4; add.s64 %rd19, %rd1, %rd63; @%p25 bra $L__BB20_11; bra.uni $L__BB20_10; $L__BB20_11: ld.local.u32 %r160, [%rd19]; add.s32 %r159, %r23, -1; setp.le.s32 %p29, %r95, %r164; @%p29 bra $L__BB20_55; and.b32 %r108, %r97, 64; setp.eq.s32 %p31, %r108, 0; or.pred %p1, %p31, %p30; mul.ftz.f32 %f282, %f488, %f488; fma.rn.ftz.f32 %f283, %f487, %f487, %f282; fma.rn.ftz.f32 %f284, %f492, %f492, %f283; rcp.approx.ftz.f32 %f53, %f284; $L__BB20_13: cvt.s64.s32 %rd20, %r164; setp.ne.s32 %p32, %r177, -1; mov.u32 %r165, %r177; @%p32 bra $L__BB20_15; shl.b64 %rd71, %rd20, 2; add.s64 %rd72, %rd6, %rd71; ld.global.u32 %r165, [%rd72]; $L__BB20_15: cvt.s64.s32 %rd105, %r164; shl.b64 %rd73, %rd105, 2; add.s64 %rd74, %rd9, %rd73; ld.global.u32 %r40, [%rd74]; cvt.s64.s32 %rd21, %r40; ld.local.u32 %r109, [%rd2+36]; setp.ne.s32 %p33, %r109, %r40; @%p33 bra $L__BB20_17; ld.local.u32 %r110, [%rd2+40]; setp.eq.s32 %p34, %r110, %r165; mov.u16 %rs23, %rs26; @%p34 bra $L__BB20_19; $L__BB20_17: ld.local.u32 %r111, [%rd2+44]; setp.ne.s32 %p35, %r111, %r40; mov.u16 %rs23, 0; @%p35 bra $L__BB20_19; ld.local.u32 %r112, [%rd2+48]; setp.eq.s32 %p36, %r112, %r165; selp.u16 %rs23, 1, 0, %p36; $L__BB20_19: setp.ne.s16 %p37, %rs23, 0; @%p37 bra $L__BB20_54; bra.uni $L__BB20_20; $L__BB20_54: add.s32 %r159, %r23, -1; cvt.u32.u64 %r133, %rd20; add.s32 %r164, %r133, 1; setp.lt.s32 %p81, %r164, %r95; @%p81 bra $L__BB20_13; bra.uni $L__BB20_55; $L__BB20_20: cvt.u16.u32 %rs18, %r97; and.b16 %rs17, %rs18, 127; setp.eq.s16 %p38, %rs17, 1; @%p38 bra $L__BB20_44; cvt.u16.u32 %rs20, %r97; and.b16 %rs19, %rs20, 127; setp.eq.s16 %p39, %rs19, 72; @%p39 bra $L__BB20_23; cvt.u16.u32 %rs22, %r97; and.b16 %rs21, %rs22, 127; setp.ne.s16 %p40, %rs21, 8; @%p40 bra $L__BB20_54; $L__BB20_23: @%p1 bra $L__BB20_25; shl.b64 %rd75, %rd20, 3; add.s64 %rd76, %rd13, %rd75; ld.global.v2.f32 {%f285, %f286}, [%rd76]; ld.local.f32 %f289, [%rd2+32]; setp.lt.ftz.f32 %p41, %f289, %f285; setp.gt.ftz.f32 %p42, %f289, %f286; or.pred %p43, %p41, %p42; @%p43 bra $L__BB20_54; $L__BB20_25: cvt.s64.s32 %rd107, %r164; shl.b64 %rd106, %rd105, 2; add.s64 %rd78, %rd12, %rd106; ld.global.u32 %r41, [%rd78]; and.b32 %r113, %r41, 64; setp.eq.s32 %p44, %r113, 0; @%p44 bra $L__BB20_39; bra.uni $L__BB20_26; $L__BB20_39: shl.b64 %rd89, %rd21, 4; add.s64 %rd90, %rd15, %rd89; ld.global.v4.f32 {%f467, %f468, %f469, %f470}, [%rd90]; bra.uni $L__BB20_40; $L__BB20_44: shl.b64 %rd91, %rd21, 4; add.s64 %rd92, %rd10, %rd91; ld.global.u32 %r127, [%rd92+12]; mul.wide.u32 %rd93, %r127, 12; add.s64 %rd94, %rd11, %rd93; add.s32 %r128, %r127, 1; mul.wide.u32 %rd95, %r128, 12; add.s64 %rd96, %rd11, %rd95; add.s32 %r129, %r127, 2; mul.wide.u32 %rd97, %r129, 12; add.s64 %rd98, %rd11, %rd97; ld.global.f32 %f333, [%rd94]; sub.ftz.f32 %f83, %f333, %f493; ld.global.f32 %f334, [%rd94+4]; sub.ftz.f32 %f84, %f334, %f494; ld.global.f32 %f335, [%rd94+8]; sub.ftz.f32 %f85, %f335, %f495; ld.global.f32 %f336, [%rd96]; sub.ftz.f32 %f337, %f336, %f493; ld.global.f32 %f338, [%rd96+4]; sub.ftz.f32 %f339, %f338, %f494; ld.global.f32 %f340, [%rd96+8]; sub.ftz.f32 %f341, %f340, %f495; ld.global.f32 %f342, [%rd98]; sub.ftz.f32 %f343, %f342, %f493; ld.global.f32 %f344, [%rd98+4]; sub.ftz.f32 %f345, %f344, %f494; ld.global.f32 %f346, [%rd98+8]; sub.ftz.f32 %f347, %f346, %f495; sub.ftz.f32 %f86, %f343, %f83; sub.ftz.f32 %f87, %f345, %f84; sub.ftz.f32 %f88, %f347, %f85; sub.ftz.f32 %f89, %f83, %f337; sub.ftz.f32 %f90, %f84, %f339; sub.ftz.f32 %f91, %f85, %f341; sub.ftz.f32 %f348, %f337, %f343; sub.ftz.f32 %f349, %f339, %f345; sub.ftz.f32 %f350, %f341, %f347; add.ftz.f32 %f351, %f83, %f343; add.ftz.f32 %f352, %f84, %f345; add.ftz.f32 %f353, %f85, %f347; mul.ftz.f32 %f354, %f87, %f353; mul.ftz.f32 %f355, %f352, %f88; sub.ftz.f32 %f356, %f354, %f355; mul.ftz.f32 %f357, %f351, %f88; mul.ftz.f32 %f358, %f86, %f353; sub.ftz.f32 %f359, %f357, %f358; mul.ftz.f32 %f360, %f86, %f352; mul.ftz.f32 %f361, %f351, %f87; sub.ftz.f32 %f362, %f360, %f361; mul.ftz.f32 %f363, %f487, %f359; fma.rn.ftz.f32 %f364, %f488, %f356, %f363; fma.rn.ftz.f32 %f92, %f492, %f362, %f364; add.ftz.f32 %f365, %f83, %f337; add.ftz.f32 %f366, %f84, %f339; add.ftz.f32 %f367, %f85, %f341; mul.ftz.f32 %f368, %f90, %f367; mul.ftz.f32 %f369, %f366, %f91; sub.ftz.f32 %f370, %f368, %f369; mul.ftz.f32 %f371, %f365, %f91; mul.ftz.f32 %f372, %f89, %f367; sub.ftz.f32 %f373, %f371, %f372; mul.ftz.f32 %f374, %f89, %f366; mul.ftz.f32 %f375, %f365, %f90; sub.ftz.f32 %f376, %f374, %f375; mul.ftz.f32 %f377, %f487, %f373; fma.rn.ftz.f32 %f378, %f488, %f370, %f377; fma.rn.ftz.f32 %f93, %f492, %f376, %f378; add.ftz.f32 %f379, %f337, %f343; add.ftz.f32 %f380, %f339, %f345; add.ftz.f32 %f381, %f341, %f347; mul.ftz.f32 %f382, %f349, %f381; mul.ftz.f32 %f383, %f380, %f350; sub.ftz.f32 %f384, %f382, %f383; mul.ftz.f32 %f385, %f379, %f350; mul.ftz.f32 %f386, %f348, %f381; sub.ftz.f32 %f387, %f385, %f386; mul.ftz.f32 %f388, %f348, %f380; mul.ftz.f32 %f389, %f379, %f349; sub.ftz.f32 %f390, %f388, %f389; mul.ftz.f32 %f391, %f487, %f387; fma.rn.ftz.f32 %f392, %f488, %f384, %f391; fma.rn.ftz.f32 %f393, %f492, %f390, %f392; add.ftz.f32 %f394, %f93, %f92; add.ftz.f32 %f94, %f393, %f394; abs.ftz.f32 %f95, %f94; mul.ftz.f32 %f395, %f95, 0f34000000; min.ftz.f32 %f396, %f93, %f393; min.ftz.f32 %f397, %f92, %f396; max.ftz.f32 %f398, %f93, %f393; max.ftz.f32 %f399, %f92, %f398; mul.ftz.f32 %f400, %f95, 0fB4000000; setp.ltu.ftz.f32 %p64, %f397, %f400; setp.gtu.ftz.f32 %p65, %f399, %f395; and.pred %p66, %p64, %p65; mov.pred %p90, %p63; @%p66 bra $L__BB20_50; mul.ftz.f32 %f401, %f90, %f88; mul.ftz.f32 %f402, %f91, %f87; sub.ftz.f32 %f403, %f401, %f402; mul.ftz.f32 %f404, %f89, %f88; mul.ftz.f32 %f405, %f91, %f86; sub.ftz.f32 %f406, %f405, %f404; mul.ftz.f32 %f407, %f90, %f86; mul.ftz.f32 %f408, %f89, %f87; sub.ftz.f32 %f409, %f408, %f407; add.ftz.f32 %f96, %f403, %f403; add.ftz.f32 %f97, %f406, %f406; add.ftz.f32 %f98, %f409, %f409; mul.ftz.f32 %f410, %f487, %f97; fma.rn.ftz.f32 %f411, %f488, %f96, %f410; fma.rn.ftz.f32 %f99, %f492, %f98, %f411; setp.eq.ftz.f32 %p68, %f99, 0f00000000; mov.pred %p90, %p63; @%p68 bra $L__BB20_50; mul.ftz.f32 %f412, %f84, %f97; fma.rn.ftz.f32 %f413, %f83, %f96, %f412; fma.rn.ftz.f32 %f414, %f85, %f98, %f413; div.approx.ftz.f32 %f100, %f414, %f99; setp.ltu.ftz.f32 %p70, %f100, %f10; setp.gtu.ftz.f32 %p71, %f100, %f478; or.pred %p72, %p71, %p70; mov.pred %p90, %p63; @%p72 bra $L__BB20_50; abs.ftz.f32 %f428, %f94; setp.lt.ftz.f32 %p73, %f428, 0f219392EF; mov.f32 %f473, 0f00000000; @%p73 bra $L__BB20_49; rcp.approx.ftz.f32 %f473, %f94; $L__BB20_49: mul.ftz.f32 %f416, %f92, %f473; mov.f32 %f417, 0f3F800000; min.ftz.f32 %f482, %f416, %f417; mul.ftz.f32 %f418, %f93, %f473; min.ftz.f32 %f481, %f418, %f417; mov.pred %p90, 0; mov.f32 %f483, %f100; $L__BB20_50: mov.pred %p91, %p75; @%p90 bra $L__BB20_53; cvt.s64.s32 %rd104, %r164; shl.b64 %rd103, %rd104, 2; add.s64 %rd100, %rd14, %rd103; ld.global.u32 %r130, [%rd100]; and.b32 %r131, %r130, %r69; setp.eq.s32 %p77, %r131, 0; mov.pred %p91, %p75; @%p77 bra $L__BB20_53; st.local.u32 [%rd3+16], %r165; st.local.u32 [%rd3+12], %r40; mov.u32 %r132, 1; st.local.u32 [%rd3+20], %r132; st.local.f32 [%rd3+4], %f482; st.local.f32 [%rd3+8], %f481; st.local.f32 [%rd3], %f483; mov.pred %p91, -1; mov.u32 %r173, %r40; mov.f32 %f478, %f483; $L__BB20_53: setp.ne.s32 %p79, %r3, 0; and.pred %p80, %p79, %p91; @%p80 bra $L__BB20_60; bra.uni $L__BB20_54; $L__BB20_26: mul.wide.s32 %rd79, %r165, 208; add.s64 %rd80, %rd7, %rd79; ld.global.u32 %r42, [%rd80+148]; ld.global.u32 %r43, [%rd80+152]; shl.b32 %r114, %r43, 1; cvt.rn.f32.s32 %f290, %r114; ld.local.f32 %f291, [%rd2+32]; mul.ftz.f32 %f60, %f291, %f290; cvt.rzi.ftz.s32.f32 %r115, %f60; add.s32 %r116, %r114, -1; min.s32 %r44, %r115, %r116; ld.global.u32 %r167, [%rd80+164]; mul.wide.u32 %rd81, %r167, 16; add.s64 %rd82, %rd16, %rd81; ld.global.u16 %rs24, [%rd82+12]; ld.global.u32 %r166, [%rd82+8]; ld.global.u64 %rd108, [%rd82]; setp.eq.s64 %p45, %rd108, 11; @%p45 bra $L__BB20_31; $L__BB20_27: setp.eq.s64 %p46, %rd108, 0; @%p46 bra $L__BB20_29; add.s32 %r167, %r167, 2; bra.uni $L__BB20_30; $L__BB20_29: setp.eq.s16 %p47, %rs24, 0; mov.u32 %r170, -1; mov.u32 %r167, %r166; @%p47 bra $L__BB20_32; $L__BB20_30: mul.wide.u32 %rd83, %r167, 16; add.s64 %rd84, %rd16, %rd83; ld.global.u16 %rs24, [%rd84+12]; ld.global.u32 %r166, [%rd84+8]; ld.global.u64 %rd108, [%rd84]; setp.ne.s64 %p48, %rd108, 11; @%p48 bra $L__BB20_27; $L__BB20_31: setp.eq.s16 %p49, %rs24, 0; selp.b32 %r170, -1, %r166, %p49; $L__BB20_32: setp.eq.s32 %p50, %r44, %r43; @%p50 bra $L__BB20_34; bra.uni $L__BB20_33; $L__BB20_34: shl.b64 %rd86, %rd21, 4; add.s64 %rd109, %rd15, %rd86; bra.uni $L__BB20_35; $L__BB20_33: setp.gt.s32 %p51, %r44, %r43; selp.b32 %r118, -1, 0, %p51; add.s32 %r119, %r44, %r118; mad.lo.s32 %r120, %r119, %r42, %r40; add.s32 %r121, %r120, %r170; mul.wide.s32 %rd85, %r121, 16; add.s64 %rd109, %rd17, %rd85; $L__BB20_35: ld.global.v4.f32 {%f292, %f293, %f294, %f295}, [%rd109]; add.s32 %r55, %r44, 1; setp.eq.s32 %p52, %r55, %r43; @%p52 bra $L__BB20_37; bra.uni $L__BB20_36; $L__BB20_37: shl.b64 %rd88, %rd21, 4; add.s64 %rd110, %rd15, %rd88; bra.uni $L__BB20_38; $L__BB20_36: setp.ge.s32 %p53, %r44, %r43; selp.b32 %r122, -1, 0, %p53; add.s32 %r123, %r55, %r122; mad.lo.s32 %r124, %r123, %r42, %r40; add.s32 %r125, %r124, %r170; mul.wide.s32 %rd87, %r125, 16; add.s64 %rd110, %rd17, %rd87; $L__BB20_38: shl.b32 %r148, %r43, 1; cvt.rn.f32.s32 %f430, %r148; mul.ftz.f32 %f429, %f291, %f430; cvt.rn.f32.s32 %f296, %r44; sub.ftz.f32 %f297, %f429, %f296; ld.global.v4.f32 {%f298, %f299, %f300, %f301}, [%rd110]; mov.f32 %f306, 0f3F800000; sub.ftz.f32 %f307, %f306, %f297; mul.ftz.f32 %f308, %f297, %f298; mul.ftz.f32 %f309, %f297, %f299; mul.ftz.f32 %f310, %f297, %f300; mul.ftz.f32 %f311, %f297, %f301; fma.rn.ftz.f32 %f470, %f307, %f295, %f311; fma.rn.ftz.f32 %f469, %f307, %f294, %f310; fma.rn.ftz.f32 %f468, %f307, %f293, %f309; fma.rn.ftz.f32 %f467, %f307, %f292, %f308; $L__BB20_40: sub.ftz.f32 %f316, %f467, %f493; sub.ftz.f32 %f317, %f468, %f494; sub.ftz.f32 %f318, %f469, %f495; mul.ftz.f32 %f319, %f487, %f317; fma.rn.ftz.f32 %f320, %f488, %f316, %f319; fma.rn.ftz.f32 %f321, %f492, %f318, %f320; mul.ftz.f32 %f77, %f53, %f321; mul.ftz.f32 %f322, %f488, %f77; mul.ftz.f32 %f323, %f487, %f77; mul.ftz.f32 %f324, %f492, %f77; sub.ftz.f32 %f325, %f316, %f322; sub.ftz.f32 %f326, %f317, %f323; sub.ftz.f32 %f327, %f318, %f324; mul.ftz.f32 %f328, %f326, %f326; fma.rn.ftz.f32 %f329, %f325, %f325, %f328; fma.rn.ftz.f32 %f78, %f327, %f327, %f329; mul.ftz.f32 %f79, %f470, %f470; setp.gtu.ftz.f32 %p55, %f78, %f79; mov.pred %p89, 0; @%p55 bra $L__BB20_43; sub.ftz.f32 %f330, %f79, %f78; mul.ftz.f32 %f331, %f53, %f330; sqrt.approx.ftz.f32 %f332, %f331; sub.ftz.f32 %f80, %f77, %f332; setp.ltu.ftz.f32 %p57, %f80, %f10; setp.gtu.ftz.f32 %p58, %f80, %f478; or.pred %p59, %p58, %p57; @%p59 bra $L__BB20_43; st.local.f32 [%rd3], %f80; st.local.u32 [%rd3+12], %r40; st.local.u32 [%rd3+16], %r165; st.local.u32 [%rd3+20], %r41; mov.u32 %r126, 0; st.local.u32 [%rd3+4], %r126; st.local.u32 [%rd3+8], %r126; mov.pred %p89, -1; mov.u32 %r173, %r40; mov.f32 %f478, %f80; $L__BB20_43: setp.ne.s32 %p61, %r3, 0; and.pred %p62, %p61, %p89; @%p62 bra $L__BB20_60; bra.uni $L__BB20_54; $L__BB20_10: mul.wide.s32 %rd64, %r164, 4; xor.b64 %rd65, %rd64, -4; add.s64 %rd66, %rd6, %rd65; ld.global.u32 %r177, [%rd66]; mul.wide.s32 %rd67, %r177, 208; add.s64 %rd68, %rd7, %rd67; ld.global.v4.f32 {%f228, %f229, %f230, %f231}, [%rd68+48]; ld.global.v4.f32 {%f236, %f237, %f238, %f239}, [%rd68+64]; ld.global.v4.f32 {%f244, %f245, %f246, %f247}, [%rd68+80]; ld.local.f32 %f252, [%rd2]; ld.local.f32 %f253, [%rd2+4]; mul.ftz.f32 %f254, %f229, %f253; fma.rn.ftz.f32 %f255, %f252, %f228, %f254; ld.local.f32 %f256, [%rd2+8]; fma.rn.ftz.f32 %f257, %f230, %f256, %f255; add.ftz.f32 %f493, %f231, %f257; mul.ftz.f32 %f258, %f237, %f253; fma.rn.ftz.f32 %f259, %f252, %f236, %f258; fma.rn.ftz.f32 %f260, %f238, %f256, %f259; add.ftz.f32 %f494, %f239, %f260; mul.ftz.f32 %f261, %f245, %f253; fma.rn.ftz.f32 %f262, %f252, %f244, %f261; fma.rn.ftz.f32 %f263, %f246, %f256, %f262; add.ftz.f32 %f495, %f247, %f263; ld.local.f32 %f264, [%rd2+12]; ld.local.f32 %f265, [%rd2+16]; mul.ftz.f32 %f266, %f229, %f265; fma.rn.ftz.f32 %f267, %f228, %f264, %f266; ld.local.f32 %f268, [%rd2+20]; fma.rn.ftz.f32 %f269, %f230, %f268, %f267; mul.ftz.f32 %f270, %f237, %f265; fma.rn.ftz.f32 %f271, %f236, %f264, %f270; fma.rn.ftz.f32 %f272, %f238, %f268, %f271; mul.ftz.f32 %f273, %f245, %f265; fma.rn.ftz.f32 %f274, %f244, %f264, %f273; fma.rn.ftz.f32 %f275, %f246, %f268, %f274; abs.ftz.f32 %f276, %f269; setp.gt.ftz.f32 %p26, %f276, 0f17800000; mov.b32 %r98, %f269; and.b32 %r99, %r98, -2147483648; or.b32 %r100, %r99, 394264576; mov.b32 %f277, %r100; selp.f32 %f488, %f269, %f277, %p26; abs.ftz.f32 %f278, %f272; setp.gt.ftz.f32 %p27, %f278, 0f17800000; mov.b32 %r101, %f272; and.b32 %r102, %r101, -2147483648; or.b32 %r103, %r102, 394264576; mov.b32 %f279, %r103; selp.f32 %f487, %f272, %f279, %p27; abs.ftz.f32 %f280, %f275; setp.gt.ftz.f32 %p28, %f280, 0f17800000; mov.b32 %r104, %f275; and.b32 %r105, %r104, -2147483648; or.b32 %r106, %r105, 394264576; mov.b32 %f281, %r106; selp.f32 %f492, %f275, %f281, %p28; rcp.approx.ftz.f32 %f491, %f488; rcp.approx.ftz.f32 %f490, %f487; rcp.approx.ftz.f32 %f489, %f492; add.s32 %r159, %r23, 1; mov.u32 %r107, 1985229328; st.local.u32 [%rd19+4], %r107; mul.wide.s32 %rd69, %r177, 4; add.s64 %rd70, %rd8, %rd69; ld.global.u32 %r160, [%rd70]; bra.uni $L__BB20_55; $L__BB20_1: setp.eq.s32 %p8, %r160, 1985229328; setp.lt.s32 %p9, %r160, 0; or.pred %p10, %p9, %p8; @%p10 bra $L__BB20_8; $L__BB20_2: mul.wide.s32 %rd57, %r160, 16; add.s64 %rd58, %rd4, %rd57; ld.global.v4.f32 {%f158, %f159, %f160, %f161}, [%rd58+16]; ld.global.v4.f32 {%f166, %f167, %f168, %f169}, [%rd58+32]; ld.global.v4.f32 {%f174, %f175, %f176, %f177}, [%rd58+48]; sub.ftz.f32 %f182, %f158, %f493; mul.ftz.f32 %f183, %f491, %f182; sub.ftz.f32 %f184, %f160, %f493; mul.ftz.f32 %f185, %f491, %f184; sub.ftz.f32 %f186, %f166, %f494; mul.ftz.f32 %f187, %f490, %f186; sub.ftz.f32 %f188, %f168, %f494; mul.ftz.f32 %f189, %f490, %f188; sub.ftz.f32 %f190, %f174, %f495; mul.ftz.f32 %f191, %f489, %f190; sub.ftz.f32 %f192, %f176, %f495; mul.ftz.f32 %f193, %f489, %f192; min.ftz.f32 %f194, %f183, %f185; min.ftz.f32 %f195, %f187, %f189; min.ftz.f32 %f196, %f191, %f193; max.ftz.f32 %f197, %f10, %f194; max.ftz.f32 %f198, %f195, %f196; max.ftz.f32 %f42, %f197, %f198; max.ftz.f32 %f199, %f183, %f185; max.ftz.f32 %f200, %f187, %f189; max.ftz.f32 %f201, %f191, %f193; min.ftz.f32 %f202, %f478, %f199; min.ftz.f32 %f203, %f200, %f201; min.ftz.f32 %f204, %f202, %f203; sub.ftz.f32 %f205, %f159, %f493; mul.ftz.f32 %f206, %f491, %f205; sub.ftz.f32 %f207, %f161, %f493; mul.ftz.f32 %f208, %f491, %f207; sub.ftz.f32 %f209, %f167, %f494; mul.ftz.f32 %f210, %f490, %f209; sub.ftz.f32 %f211, %f169, %f494; mul.ftz.f32 %f212, %f490, %f211; sub.ftz.f32 %f213, %f175, %f495; mul.ftz.f32 %f214, %f489, %f213; sub.ftz.f32 %f215, %f177, %f495; mul.ftz.f32 %f216, %f489, %f215; min.ftz.f32 %f217, %f206, %f208; min.ftz.f32 %f218, %f210, %f212; min.ftz.f32 %f219, %f214, %f216; max.ftz.f32 %f220, %f10, %f217; max.ftz.f32 %f221, %f218, %f219; max.ftz.f32 %f43, %f220, %f221; max.ftz.f32 %f222, %f206, %f208; max.ftz.f32 %f223, %f210, %f212; max.ftz.f32 %f224, %f214, %f216; min.ftz.f32 %f225, %f478, %f222; min.ftz.f32 %f226, %f223, %f224; min.ftz.f32 %f227, %f225, %f226; setp.ge.ftz.f32 %p11, %f204, %f42; ld.global.v4.u32 {%r83, %r84, %r85, %r160}, [%rd58]; and.b32 %r88, %r83, %r69; setp.ne.s32 %p12, %r88, 0; and.pred %p13, %p12, %p11; selp.u32 %r89, 1, 0, %p13; setp.ltu.ftz.f32 %p14, %f227, %f43; and.b32 %r91, %r84, %r69; setp.eq.s32 %p15, %r91, 0; or.pred %p16, %p15, %p14; selp.b32 %r92, 0, 2, %p16; or.b32 %r14, %r92, %r89; setp.eq.s32 %p17, %r14, 3; mul.wide.s32 %rd59, %r159, 4; add.s64 %rd18, %rd1, %rd59; @%p17 bra $L__BB20_6; bra.uni $L__BB20_3; $L__BB20_6: setp.lt.ftz.f32 %p20, %f43, %f42; selp.b32 %r19, %r160, %r85, %p20; selp.b32 %r93, %r85, %r160, %p20; add.s32 %r159, %r159, 1; st.local.u32 [%rd18+4], %r93; mov.u32 %r160, %r19; bra.uni $L__BB20_7; $L__BB20_3: cvt.u16.u32 %rs10, %r14; setp.eq.s16 %p18, %rs10, 2; @%p18 bra $L__BB20_7; setp.ne.s16 %p19, %rs10, 0; mov.u32 %r160, %r85; @%p19 bra $L__BB20_7; ld.local.u32 %r160, [%rd18]; add.s32 %r159, %r159, -1; $L__BB20_7: setp.ne.s32 %p21, %r160, 1985229328; setp.gt.s32 %p22, %r160, -1; and.pred %p23, %p22, %p21; @%p23 bra $L__BB20_2; $L__BB20_8: mov.u32 %r23, %r159; setp.gt.s32 %p24, %r160, -1; @%p24 bra $L__BB20_55; bra.uni $L__BB20_9; $L__BB20_55: setp.ne.s32 %p82, %r160, 1985229328; @%p82 bra $L__BB20_1; mov.u32 %r160, 1985229328; setp.lt.s32 %p83, %r159, 0; @%p83 bra $L__BB20_58; ld.local.f32 %f493, [%rd2]; ld.local.f32 %f494, [%rd2+4]; ld.local.f32 %f495, [%rd2+8]; ld.local.f32 %f419, [%rd2+12]; abs.ftz.f32 %f420, %f419; setp.gt.ftz.f32 %p84, %f420, 0f17800000; mov.b32 %r136, %f419; and.b32 %r137, %r136, -2147483648; or.b32 %r138, %r137, 394264576; mov.b32 %f421, %r138; selp.f32 %f488, %f419, %f421, %p84; ld.local.f32 %f422, [%rd2+16]; abs.ftz.f32 %f423, %f422; setp.gt.ftz.f32 %p85, %f423, 0f17800000; mov.b32 %r139, %f422; and.b32 %r140, %r139, -2147483648; or.b32 %r141, %r140, 394264576; mov.b32 %f424, %r141; selp.f32 %f487, %f422, %f424, %p85; ld.local.f32 %f425, [%rd2+20]; abs.ftz.f32 %f426, %f425; setp.gt.ftz.f32 %p86, %f426, 0f17800000; mov.b32 %r142, %f425; and.b32 %r143, %r142, -2147483648; or.b32 %r144, %r143, 394264576; mov.b32 %f427, %r144; selp.f32 %f492, %f425, %f427, %p86; rcp.approx.ftz.f32 %f491, %f488; rcp.approx.ftz.f32 %f490, %f487; rcp.approx.ftz.f32 %f489, %f492; mul.wide.s32 %rd101, %r159, 4; add.s64 %rd102, %rd1, %rd101; ld.local.u32 %r160, [%rd102]; add.s32 %r159, %r159, -1; mov.u32 %r177, -1; $L__BB20_58: setp.ne.s32 %p87, %r160, 1985229328; @%p87 bra $L__BB20_1; setp.ne.s32 %p88, %r173, -1; selp.u16 %rs26, 1, 0, %p88; $L__BB20_60: cvt.u32.u16 %r145, %rs26; st.param.b32 [func_retval0+0], %r145; ret; } .func (.param .b32 func_retval0) _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj( .param .b64 _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0, .param .b64 _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1, .param .b32 _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2 ) { .local .align 16 .b8 __local_depot21[816]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<212>; .reg .b16 %rs<71>; .reg .f32 %f<2508>; .reg .b32 %r<284>; .reg .b64 %rd<148>; mov.u64 %SPL, __local_depot21; ld.param.u64 %rd46, [_Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0]; ld.param.u64 %rd47, [_Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1]; ld.param.u32 %r115, [_Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2]; cvta.to.local.u64 %rd4, %rd47; add.u64 %rd1, %SPL, 0; add.u64 %rd2, %SPL, 48; mov.u32 %r119, 1985229328; st.local.u32 [%rd2], %r119; ld.const.u32 %r247, [kernel_params+1104]; cvta.to.local.u64 %rd3, %rd46; ld.local.f32 %f2490, [%rd3]; ld.local.f32 %f2489, [%rd3+4]; ld.local.f32 %f2488, [%rd3+8]; ld.local.f32 %f593, [%rd3+12]; abs.ftz.f32 %f594, %f593; setp.gt.ftz.f32 %p10, %f594, 0f17800000; mov.b32 %r120, %f593; and.b32 %r121, %r120, -2147483648; or.b32 %r122, %r121, 394264576; mov.b32 %f595, %r122; selp.f32 %f2493, %f593, %f595, %p10; ld.local.f32 %f596, [%rd3+16]; abs.ftz.f32 %f597, %f596; setp.gt.ftz.f32 %p11, %f597, 0f17800000; mov.b32 %r123, %f596; and.b32 %r124, %r123, -2147483648; or.b32 %r125, %r124, 394264576; mov.b32 %f598, %r125; selp.f32 %f2492, %f596, %f598, %p11; ld.local.f32 %f599, [%rd3+20]; abs.ftz.f32 %f600, %f599; setp.gt.ftz.f32 %p12, %f600, 0f17800000; mov.b32 %r126, %f599; and.b32 %r127, %r126, -2147483648; or.b32 %r128, %r127, 394264576; mov.b32 %f601, %r128; selp.f32 %f2491, %f599, %f601, %p12; rcp.approx.ftz.f32 %f2496, %f2493; mov.f32 %f602, 0f3F800000; rcp.approx.ftz.f32 %f2495, %f2492; rcp.approx.ftz.f32 %f2494, %f2491; ld.local.f32 %f10, [%rd3+24]; ld.local.f32 %f2466, [%rd3+28]; st.local.f32 [%rd4], %f2466; mov.u32 %r246, 0; st.local.u32 [%rd4+4], %r246; st.local.u32 [%rd4+8], %r246; mov.u32 %r276, -1; st.local.u32 [%rd4+12], %r276; st.local.u32 [%rd4+16], %r276; ld.const.v2.u64 {%rd50, %rd51}, [kernel_params+2096]; cvta.to.global.u64 %rd7, %rd50; ld.const.v2.u64 {%rd52, %rd53}, [kernel_params+2128]; cvta.to.global.u64 %rd8, %rd53; ld.const.u64 %rd56, [kernel_params+2160]; cvta.to.global.u64 %rd9, %rd56; ld.const.v2.u64 {%rd57, %rd58}, [kernel_params+2144]; cvta.to.global.u64 %rd12, %rd52; ld.const.u64 %rd59, [kernel_params+2232]; cvta.to.global.u64 %rd13, %rd59; ld.const.v2.u64 {%rd60, %rd61}, [kernel_params+2256]; ld.const.v2.u32 {%r129, %r130}, [kernel_params+1128]; ld.const.v2.u64 {%rd62, %rd63}, [kernel_params+2112]; ld.const.u64 %rd64, [kernel_params+2272]; cvta.to.global.u64 %rd18, %rd64; ld.const.u64 %rd65, [kernel_params+2312]; cvta.to.global.u64 %rd19, %rd65; ld.const.u64 %rd66, [kernel_params+2296]; cvta.to.global.u64 %rd20, %rd66; cvt.rn.f32.s32 %f603, %r130; rcp.approx.ftz.f32 %f12, %f603; sub.ftz.f32 %f604, %f602, %f12; mul.ftz.f32 %f605, %f12, %f604; mul.ftz.f32 %f606, %f604, %f605; neg.ftz.f32 %f13, %f606; mul.ftz.f32 %f607, %f12, %f12; fma.rn.ftz.f32 %f608, %f12, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f14, %f607, %f608, 0f40000000; mul.ftz.f32 %f609, %f604, %f604; fma.rn.ftz.f32 %f610, %f604, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f15, %f609, %f610, 0f40000000; mul.ftz.f32 %f611, %f12, %f605; neg.ftz.f32 %f16, %f611; ld.const.u64 %rd67, [kernel_params+2344]; cvta.to.global.u64 %rd21, %rd67; cvta.to.global.u64 %rd71, %rd51; setp.eq.s32 %p41, %r129, 0; mov.u16 %rs70, 1; cvta.to.global.u64 %rd88, %rd58; cvta.to.global.u64 %rd90, %rd62; cvta.to.global.u64 %rd132, %rd60; mov.pred %p178, -1; mov.pred %p190, 0; cvta.to.global.u64 %rd139, %rd63; cvta.to.global.u64 %rd104, %rd61; cvta.to.global.u64 %rd81, %rd57; mov.u32 %r280, %r276; bra.uni $L__BB21_1; $L__BB21_13: mul.wide.s32 %rd72, %r247, 16; xor.b64 %rd73, %rd72, -16; add.s64 %rd74, %rd71, %rd73; ld.global.v4.u32 {%r251, %r142, %r143, %r144}, [%rd74]; setp.gt.s32 %p36, %r251, -1; mul.wide.s32 %rd75, %r32, 4; add.s64 %rd24, %rd2, %rd75; @%p36 bra $L__BB21_15; bra.uni $L__BB21_14; $L__BB21_15: ld.local.u32 %r247, [%rd24]; add.s32 %r246, %r32, -1; setp.le.s32 %p40, %r142, %r251; @%p40 bra $L__BB21_145; cvt.u16.u32 %rs26, %r144; and.b16 %rs1, %rs26, 127; and.b32 %r155, %r144, 64; setp.eq.s32 %p42, %r155, 0; or.pred %p1, %p42, %p41; mul.ftz.f32 %f879, %f2493, %f2493; fma.rn.ftz.f32 %f880, %f2492, %f2492, %f879; fma.rn.ftz.f32 %f75, %f2491, %f2491, %f880; rcp.approx.ftz.f32 %f76, %f75; $L__BB21_17: cvt.s64.s32 %rd25, %r251; setp.ne.s32 %p43, %r280, -1; mov.u32 %r48, %r280; @%p43 bra $L__BB21_19; shl.b64 %rd84, %rd25, 2; add.s64 %rd85, %rd8, %rd84; ld.global.u32 %r48, [%rd85]; $L__BB21_19: shl.b64 %rd86, %rd25, 2; add.s64 %rd87, %rd12, %rd86; ld.global.u32 %r49, [%rd87]; cvt.s64.s32 %rd26, %r49; ld.local.u32 %r156, [%rd3+36]; setp.ne.s32 %p44, %r156, %r49; @%p44 bra $L__BB21_21; ld.local.u32 %r157, [%rd3+40]; setp.eq.s32 %p45, %r157, %r48; mov.u16 %rs55, %rs70; @%p45 bra $L__BB21_23; $L__BB21_21: ld.local.u32 %r158, [%rd3+44]; setp.ne.s32 %p46, %r158, %r49; mov.u16 %rs55, 0; @%p46 bra $L__BB21_23; ld.local.u32 %r159, [%rd3+48]; setp.eq.s32 %p47, %r159, %r48; selp.u16 %rs55, 1, 0, %p47; $L__BB21_23: setp.ne.s16 %p48, %rs55, 0; @%p48 bra $L__BB21_144; bra.uni $L__BB21_24; $L__BB21_144: add.s32 %r246, %r32, -1; cvt.u32.u64 %r220, %rd25; add.s32 %r251, %r220, 1; setp.lt.s32 %p196, %r251, %r142; @%p196 bra $L__BB21_17; bra.uni $L__BB21_145; $L__BB21_24: shl.b64 %rd89, %rd25, 3; add.s64 %rd28, %rd88, %rd89; add.s64 %rd29, %rd90, %rd86; mul.wide.s32 %rd92, %r48, 208; add.s64 %rd93, %rd9, %rd92; add.s64 %rd30, %rd93, 148; setp.gt.s16 %p49, %rs1, 7; @%p49 bra $L__BB21_28; bra.uni $L__BB21_25; $L__BB21_28: setp.gt.s16 %p50, %rs1, 67; @%p50 bra $L__BB21_37; setp.eq.s16 %p53, %rs1, 8; @%p53 bra $L__BB21_39; setp.eq.s16 %p54, %rs1, 66; @%p54 bra $L__BB21_31; bra.uni $L__BB21_144; $L__BB21_25: setp.eq.s16 %p55, %rs1, 1; @%p55 bra $L__BB21_134; setp.eq.s16 %p56, %rs1, 2; @%p56 bra $L__BB21_31; setp.eq.s16 %p57, %rs1, 4; @%p57 bra $L__BB21_31; bra.uni $L__BB21_144; $L__BB21_37: setp.eq.s16 %p51, %rs1, 68; @%p51 bra $L__BB21_31; setp.ne.s16 %p52, %rs1, 72; @%p52 bra $L__BB21_144; $L__BB21_39: @%p1 bra $L__BB21_41; ld.global.v2.f32 {%f881, %f882}, [%rd28]; ld.local.f32 %f885, [%rd3+32]; setp.lt.ftz.f32 %p58, %f885, %f881; setp.gt.ftz.f32 %p59, %f885, %f882; or.pred %p60, %p58, %p59; @%p60 bra $L__BB21_144; $L__BB21_41: ld.global.u32 %r50, [%rd29]; and.b32 %r160, %r50, 64; setp.eq.s32 %p61, %r160, 0; @%p61 bra $L__BB21_55; bra.uni $L__BB21_42; $L__BB21_55: shl.b64 %rd102, %rd26, 4; add.s64 %rd103, %rd20, %rd102; ld.global.v4.f32 {%f2323, %f2324, %f2325, %f2326}, [%rd103]; bra.uni $L__BB21_56; $L__BB21_31: @%p1 bra $L__BB21_33; ld.global.v2.f32 {%f929, %f930}, [%rd28]; ld.local.f32 %f933, [%rd3+32]; setp.lt.ftz.f32 %p80, %f933, %f929; setp.gt.ftz.f32 %p81, %f933, %f930; or.pred %p82, %p80, %p81; @%p82 bra $L__BB21_144; $L__BB21_33: ld.global.u32 %r66, [%rd29]; and.b32 %r176, %r66, 64; setp.eq.s32 %p83, %r176, 0; shl.b64 %rd105, %rd26, 4; add.s64 %rd106, %rd104, %rd105; shr.s32 %r177, %r66, 7; ld.global.u32 %r178, [%rd106+4]; add.s32 %r67, %r178, %r177; add.s32 %r179, %r67, -1; max.s32 %r68, %r179, %r178; add.s32 %r180, %r67, 2; ld.global.u32 %r181, [%rd106+8]; add.s32 %r182, %r178, %r181; add.s32 %r183, %r182, -1; min.s32 %r69, %r180, %r183; mul.wide.s32 %rd107, %r68, 16; add.s64 %rd40, %rd18, %rd107; mul.wide.s32 %rd108, %r67, 16; add.s64 %rd41, %rd18, %rd108; mul.wide.s32 %rd109, %r69, 16; add.s64 %rd42, %rd18, %rd109; @%p83 bra $L__BB21_70; bra.uni $L__BB21_34; $L__BB21_70: ld.global.v4.f32 {%f2362, %f2363, %f2364, %f245}, [%rd40]; ld.global.v4.f32 {%f2366, %f2367, %f2368, %f249}, [%rd41]; ld.global.v4.f32 {%f2370, %f2371, %f2372, %f253}, [%rd41+16]; ld.global.v4.f32 {%f2374, %f2375, %f2376, %f257}, [%rd42]; bra.uni $L__BB21_71; $L__BB21_34: ld.global.u32 %r70, [%rd30]; ld.global.u32 %r71, [%rd30+4]; shl.b32 %r184, %r71, 1; cvt.rn.f32.s32 %f934, %r184; ld.local.f32 %f935, [%rd3+32]; mul.ftz.f32 %f113, %f935, %f934; cvt.rzi.ftz.s32.f32 %r185, %f113; add.s32 %r186, %r184, -1; min.s32 %r72, %r185, %r186; ld.global.u32 %r260, [%rd30+16]; mul.wide.u32 %rd110, %r260, 16; add.s64 %rd111, %rd19, %rd110; ld.global.u16 %rs58, [%rd111+12]; ld.global.u32 %r259, [%rd111+8]; ld.global.u64 %rd147, [%rd111]; setp.eq.s64 %p84, %rd147, 11; @%p84 bra $L__BB21_62; $L__BB21_35: setp.eq.s64 %p85, %rd147, 0; @%p85 bra $L__BB21_60; add.s32 %r260, %r260, 2; bra.uni $L__BB21_61; $L__BB21_60: setp.eq.s16 %p86, %rs58, 0; mov.u32 %r263, -1; mov.u32 %r260, %r259; @%p86 bra $L__BB21_63; $L__BB21_61: mul.wide.u32 %rd112, %r260, 16; add.s64 %rd113, %rd19, %rd112; ld.global.u16 %rs58, [%rd113+12]; ld.global.u32 %r259, [%rd113+8]; ld.global.u64 %rd147, [%rd113]; setp.ne.s64 %p87, %rd147, 11; @%p87 bra $L__BB21_35; $L__BB21_62: setp.eq.s16 %p88, %rs58, 0; selp.b32 %r263, -1, %r259, %p88; $L__BB21_63: setp.eq.s32 %p89, %r72, %r71; @%p89 bra $L__BB21_65; bra.uni $L__BB21_64; $L__BB21_65: ld.global.v4.f32 {%f2330, %f2331, %f2332, %f2333}, [%rd40]; ld.global.v4.f32 {%f2334, %f2335, %f2336, %f2337}, [%rd41]; ld.global.v4.f32 {%f2338, %f2339, %f2340, %f2341}, [%rd41+16]; ld.global.v4.f32 {%f2342, %f2343, %f2344, %f2345}, [%rd42]; bra.uni $L__BB21_66; $L__BB21_134: shl.b64 %rd130, %rd26, 4; add.s64 %rd131, %rd13, %rd130; ld.global.u32 %r212, [%rd131+12]; mul.wide.u32 %rd133, %r212, 12; add.s64 %rd134, %rd132, %rd133; add.s32 %r213, %r212, 1; mul.wide.u32 %rd135, %r213, 12; add.s64 %rd136, %rd132, %rd135; add.s32 %r214, %r212, 2; mul.wide.u32 %rd137, %r214, 12; add.s64 %rd138, %rd132, %rd137; ld.global.f32 %f2149, [%rd134]; sub.ftz.f32 %f515, %f2149, %f2490; ld.global.f32 %f2150, [%rd134+4]; sub.ftz.f32 %f516, %f2150, %f2489; ld.global.f32 %f2151, [%rd134+8]; sub.ftz.f32 %f517, %f2151, %f2488; ld.global.f32 %f2152, [%rd136]; sub.ftz.f32 %f2153, %f2152, %f2490; ld.global.f32 %f2154, [%rd136+4]; sub.ftz.f32 %f2155, %f2154, %f2489; ld.global.f32 %f2156, [%rd136+8]; sub.ftz.f32 %f2157, %f2156, %f2488; ld.global.f32 %f2158, [%rd138]; sub.ftz.f32 %f2159, %f2158, %f2490; ld.global.f32 %f2160, [%rd138+4]; sub.ftz.f32 %f2161, %f2160, %f2489; ld.global.f32 %f2162, [%rd138+8]; sub.ftz.f32 %f2163, %f2162, %f2488; sub.ftz.f32 %f518, %f2159, %f515; sub.ftz.f32 %f519, %f2161, %f516; sub.ftz.f32 %f520, %f2163, %f517; sub.ftz.f32 %f521, %f515, %f2153; sub.ftz.f32 %f522, %f516, %f2155; sub.ftz.f32 %f523, %f517, %f2157; sub.ftz.f32 %f2164, %f2153, %f2159; sub.ftz.f32 %f2165, %f2155, %f2161; sub.ftz.f32 %f2166, %f2157, %f2163; add.ftz.f32 %f2167, %f515, %f2159; add.ftz.f32 %f2168, %f516, %f2161; add.ftz.f32 %f2169, %f517, %f2163; mul.ftz.f32 %f2170, %f519, %f2169; mul.ftz.f32 %f2171, %f2168, %f520; sub.ftz.f32 %f2172, %f2170, %f2171; mul.ftz.f32 %f2173, %f2167, %f520; mul.ftz.f32 %f2174, %f518, %f2169; sub.ftz.f32 %f2175, %f2173, %f2174; mul.ftz.f32 %f2176, %f518, %f2168; mul.ftz.f32 %f2177, %f2167, %f519; sub.ftz.f32 %f2178, %f2176, %f2177; mul.ftz.f32 %f2179, %f2492, %f2175; fma.rn.ftz.f32 %f2180, %f2493, %f2172, %f2179; fma.rn.ftz.f32 %f524, %f2491, %f2178, %f2180; add.ftz.f32 %f2181, %f515, %f2153; add.ftz.f32 %f2182, %f516, %f2155; add.ftz.f32 %f2183, %f517, %f2157; mul.ftz.f32 %f2184, %f522, %f2183; mul.ftz.f32 %f2185, %f2182, %f523; sub.ftz.f32 %f2186, %f2184, %f2185; mul.ftz.f32 %f2187, %f2181, %f523; mul.ftz.f32 %f2188, %f521, %f2183; sub.ftz.f32 %f2189, %f2187, %f2188; mul.ftz.f32 %f2190, %f521, %f2182; mul.ftz.f32 %f2191, %f2181, %f522; sub.ftz.f32 %f2192, %f2190, %f2191; mul.ftz.f32 %f2193, %f2492, %f2189; fma.rn.ftz.f32 %f2194, %f2493, %f2186, %f2193; fma.rn.ftz.f32 %f525, %f2491, %f2192, %f2194; add.ftz.f32 %f2195, %f2153, %f2159; add.ftz.f32 %f2196, %f2155, %f2161; add.ftz.f32 %f2197, %f2157, %f2163; mul.ftz.f32 %f2198, %f2165, %f2197; mul.ftz.f32 %f2199, %f2196, %f2166; sub.ftz.f32 %f2200, %f2198, %f2199; mul.ftz.f32 %f2201, %f2195, %f2166; mul.ftz.f32 %f2202, %f2164, %f2197; sub.ftz.f32 %f2203, %f2201, %f2202; mul.ftz.f32 %f2204, %f2164, %f2196; mul.ftz.f32 %f2205, %f2195, %f2165; sub.ftz.f32 %f2206, %f2204, %f2205; mul.ftz.f32 %f2207, %f2492, %f2203; fma.rn.ftz.f32 %f2208, %f2493, %f2200, %f2207; fma.rn.ftz.f32 %f2209, %f2491, %f2206, %f2208; add.ftz.f32 %f2210, %f525, %f524; add.ftz.f32 %f526, %f2209, %f2210; abs.ftz.f32 %f527, %f526; mul.ftz.f32 %f2211, %f527, 0f34000000; min.ftz.f32 %f2212, %f525, %f2209; min.ftz.f32 %f2213, %f524, %f2212; max.ftz.f32 %f2214, %f525, %f2209; max.ftz.f32 %f2215, %f524, %f2214; mul.ftz.f32 %f2216, %f527, 0fB4000000; setp.ltu.ftz.f32 %p179, %f2213, %f2216; setp.gtu.ftz.f32 %p180, %f2215, %f2211; and.pred %p181, %p179, %p180; mov.pred %p210, %p178; @%p181 bra $L__BB21_140; mul.ftz.f32 %f2217, %f522, %f520; mul.ftz.f32 %f2218, %f523, %f519; sub.ftz.f32 %f2219, %f2217, %f2218; mul.ftz.f32 %f2220, %f521, %f520; mul.ftz.f32 %f2221, %f523, %f518; sub.ftz.f32 %f2222, %f2221, %f2220; mul.ftz.f32 %f2223, %f522, %f518; mul.ftz.f32 %f2224, %f521, %f519; sub.ftz.f32 %f2225, %f2224, %f2223; add.ftz.f32 %f528, %f2219, %f2219; add.ftz.f32 %f529, %f2222, %f2222; add.ftz.f32 %f530, %f2225, %f2225; mul.ftz.f32 %f2226, %f2492, %f529; fma.rn.ftz.f32 %f2227, %f2493, %f528, %f2226; fma.rn.ftz.f32 %f531, %f2491, %f530, %f2227; setp.eq.ftz.f32 %p183, %f531, 0f00000000; mov.pred %p210, %p178; @%p183 bra $L__BB21_140; mul.ftz.f32 %f2228, %f516, %f529; fma.rn.ftz.f32 %f2229, %f515, %f528, %f2228; fma.rn.ftz.f32 %f2230, %f517, %f530, %f2229; div.approx.ftz.f32 %f532, %f2230, %f531; setp.ltu.ftz.f32 %p185, %f532, %f10; setp.gtu.ftz.f32 %p186, %f532, %f2466; or.pred %p187, %p186, %p185; mov.pred %p210, %p178; @%p187 bra $L__BB21_140; setp.lt.ftz.f32 %p188, %f527, 0f219392EF; mov.f32 %f2461, 0f00000000; @%p188 bra $L__BB21_139; rcp.approx.ftz.f32 %f2461, %f526; $L__BB21_139: mul.ftz.f32 %f2232, %f524, %f2461; mov.f32 %f2233, 0f3F800000; min.ftz.f32 %f2471, %f2232, %f2233; mul.ftz.f32 %f2234, %f525, %f2461; min.ftz.f32 %f2470, %f2234, %f2233; mov.pred %p210, 0; mov.f32 %f2472, %f532; $L__BB21_140: mov.pred %p211, %p190; @%p210 bra $L__BB21_143; add.s64 %rd141, %rd139, %rd86; ld.global.u32 %r215, [%rd141]; and.b32 %r216, %r215, %r115; setp.eq.s32 %p192, %r216, 0; mov.pred %p211, %p190; @%p192 bra $L__BB21_143; st.local.u32 [%rd4+16], %r48; st.local.u32 [%rd4+12], %r49; mov.u32 %r217, 1; st.local.u32 [%rd4+20], %r217; st.local.f32 [%rd4+4], %f2471; st.local.f32 [%rd4+8], %f2470; st.local.f32 [%rd4], %f2472; mov.pred %p211, -1; mov.u32 %r276, %r49; mov.f32 %f2466, %f2472; $L__BB21_143: shr.u32 %r218, %r115, 8; and.b32 %r219, %r218, 1; setp.eq.b32 %p194, %r219, 1; and.pred %p195, %p194, %p211; @%p195 bra $L__BB21_150; bra.uni $L__BB21_144; $L__BB21_42: ld.global.u32 %r51, [%rd30]; ld.global.u32 %r52, [%rd30+4]; shl.b32 %r161, %r52, 1; cvt.rn.f32.s32 %f886, %r161; ld.local.f32 %f887, [%rd3+32]; mul.ftz.f32 %f89, %f887, %f886; cvt.rzi.ftz.s32.f32 %r162, %f89; add.s32 %r163, %r161, -1; min.s32 %r53, %r162, %r163; ld.global.u32 %r254, [%rd30+16]; mul.wide.u32 %rd94, %r254, 16; add.s64 %rd95, %rd19, %rd94; ld.global.u16 %rs56, [%rd95+12]; ld.global.u32 %r253, [%rd95+8]; ld.global.u64 %rd144, [%rd95]; setp.eq.s64 %p62, %rd144, 11; @%p62 bra $L__BB21_47; $L__BB21_43: setp.eq.s64 %p63, %rd144, 0; @%p63 bra $L__BB21_45; add.s32 %r254, %r254, 2; bra.uni $L__BB21_46; $L__BB21_45: setp.eq.s16 %p64, %rs56, 0; mov.u32 %r257, -1; mov.u32 %r254, %r253; @%p64 bra $L__BB21_48; $L__BB21_46: mul.wide.u32 %rd96, %r254, 16; add.s64 %rd97, %rd19, %rd96; ld.global.u16 %rs56, [%rd97+12]; ld.global.u32 %r253, [%rd97+8]; ld.global.u64 %rd144, [%rd97]; setp.ne.s64 %p65, %rd144, 11; @%p65 bra $L__BB21_43; $L__BB21_47: setp.eq.s16 %p66, %rs56, 0; selp.b32 %r257, -1, %r253, %p66; $L__BB21_48: setp.eq.s32 %p67, %r53, %r52; @%p67 bra $L__BB21_50; bra.uni $L__BB21_49; $L__BB21_50: shl.b64 %rd99, %rd26, 4; add.s64 %rd145, %rd20, %rd99; bra.uni $L__BB21_51; $L__BB21_64: setp.gt.s32 %p90, %r72, %r71; selp.b32 %r188, -1, 0, %p90; add.s32 %r189, %r72, %r188; mad.lo.s32 %r190, %r189, %r70, %r263; add.s32 %r191, %r190, %r68; mul.wide.s32 %rd114, %r191, 16; add.s64 %rd115, %rd21, %rd114; ld.global.v4.f32 {%f2330, %f2331, %f2332, %f2333}, [%rd115]; add.s32 %r192, %r190, %r67; mul.wide.s32 %rd116, %r192, 16; add.s64 %rd117, %rd21, %rd116; ld.global.v4.f32 {%f2334, %f2335, %f2336, %f2337}, [%rd117]; ld.global.v4.f32 {%f2338, %f2339, %f2340, %f2341}, [%rd117+16]; add.s32 %r193, %r190, %r69; mul.wide.s32 %rd118, %r193, 16; add.s64 %rd119, %rd21, %rd118; ld.global.v4.f32 {%f2342, %f2343, %f2344, %f2345}, [%rd119]; $L__BB21_66: add.s32 %r83, %r72, 1; setp.eq.s32 %p91, %r83, %r71; @%p91 bra $L__BB21_68; bra.uni $L__BB21_67; $L__BB21_68: ld.global.v4.f32 {%f2346, %f2347, %f2348, %f2349}, [%rd40]; ld.global.v4.f32 {%f2350, %f2351, %f2352, %f2353}, [%rd41]; ld.global.v4.f32 {%f2354, %f2355, %f2356, %f2357}, [%rd41+16]; ld.global.v4.f32 {%f2358, %f2359, %f2360, %f2361}, [%rd42]; bra.uni $L__BB21_69; $L__BB21_67: setp.ge.s32 %p92, %r72, %r71; selp.b32 %r194, -1, 0, %p92; add.s32 %r195, %r83, %r194; mad.lo.s32 %r196, %r195, %r70, %r263; add.s32 %r197, %r196, %r68; mul.wide.s32 %rd120, %r197, 16; add.s64 %rd121, %rd21, %rd120; ld.global.v4.f32 {%f2346, %f2347, %f2348, %f2349}, [%rd121]; add.s32 %r198, %r196, %r67; mul.wide.s32 %rd122, %r198, 16; add.s64 %rd123, %rd21, %rd122; ld.global.v4.f32 {%f2350, %f2351, %f2352, %f2353}, [%rd123]; ld.global.v4.f32 {%f2354, %f2355, %f2356, %f2357}, [%rd123+16]; add.s32 %r199, %r196, %r69; mul.wide.s32 %rd124, %r199, 16; add.s64 %rd125, %rd21, %rd124; ld.global.v4.f32 {%f2358, %f2359, %f2360, %f2361}, [%rd125]; $L__BB21_69: cvt.rn.f32.s32 %f1000, %r72; sub.ftz.f32 %f1001, %f113, %f1000; sub.ftz.f32 %f1003, %f602, %f1001; mul.ftz.f32 %f1004, %f1001, %f2346; fma.rn.ftz.f32 %f2362, %f1003, %f2330, %f1004; mul.ftz.f32 %f1005, %f1001, %f2347; fma.rn.ftz.f32 %f2363, %f1003, %f2331, %f1005; mul.ftz.f32 %f1006, %f1001, %f2348; fma.rn.ftz.f32 %f2364, %f1003, %f2332, %f1006; mul.ftz.f32 %f1007, %f1001, %f2349; fma.rn.ftz.f32 %f245, %f1003, %f2333, %f1007; mul.ftz.f32 %f1008, %f1001, %f2350; fma.rn.ftz.f32 %f2366, %f1003, %f2334, %f1008; mul.ftz.f32 %f1009, %f1001, %f2351; fma.rn.ftz.f32 %f2367, %f1003, %f2335, %f1009; mul.ftz.f32 %f1010, %f1001, %f2352; fma.rn.ftz.f32 %f2368, %f1003, %f2336, %f1010; mul.ftz.f32 %f1011, %f1001, %f2353; fma.rn.ftz.f32 %f249, %f1003, %f2337, %f1011; mul.ftz.f32 %f1012, %f1001, %f2354; fma.rn.ftz.f32 %f2370, %f1003, %f2338, %f1012; mul.ftz.f32 %f1013, %f1001, %f2355; fma.rn.ftz.f32 %f2371, %f1003, %f2339, %f1013; mul.ftz.f32 %f1014, %f1001, %f2356; fma.rn.ftz.f32 %f2372, %f1003, %f2340, %f1014; mul.ftz.f32 %f1015, %f1001, %f2357; fma.rn.ftz.f32 %f253, %f1003, %f2341, %f1015; mul.ftz.f32 %f1016, %f1001, %f2358; fma.rn.ftz.f32 %f2374, %f1003, %f2342, %f1016; mul.ftz.f32 %f1017, %f1001, %f2359; fma.rn.ftz.f32 %f2375, %f1003, %f2343, %f1017; mul.ftz.f32 %f1018, %f1001, %f2360; fma.rn.ftz.f32 %f2376, %f1003, %f2344, %f1018; mul.ftz.f32 %f1019, %f1001, %f2361; fma.rn.ftz.f32 %f257, %f1003, %f2345, %f1019; $L__BB21_71: and.b32 %r200, %r66, 4; setp.eq.s32 %p93, %r200, 0; @%p93 bra $L__BB21_88; bra.uni $L__BB21_72; $L__BB21_88: mov.u16 %rs68, 0; add.ftz.f32 %f1383, %f2362, %f2366; add.ftz.f32 %f1384, %f1383, %f2370; add.ftz.f32 %f1385, %f2363, %f2367; add.ftz.f32 %f1386, %f1385, %f2371; add.ftz.f32 %f1387, %f2364, %f2368; add.ftz.f32 %f1388, %f1387, %f2372; add.ftz.f32 %f1389, %f1384, %f2374; add.ftz.f32 %f1390, %f1386, %f2375; add.ftz.f32 %f1391, %f1388, %f2376; mul.ftz.f32 %f1392, %f1389, 0f3E800000; mul.ftz.f32 %f1393, %f1390, 0f3E800000; mul.ftz.f32 %f1394, %f1391, 0f3E800000; sub.ftz.f32 %f1395, %f1392, %f2490; sub.ftz.f32 %f1396, %f1393, %f2489; sub.ftz.f32 %f1397, %f1394, %f2488; mul.ftz.f32 %f1398, %f2492, %f1396; fma.rn.ftz.f32 %f1399, %f2493, %f1395, %f1398; fma.rn.ftz.f32 %f1400, %f2491, %f1397, %f1399; div.approx.ftz.f32 %f348, %f1400, %f75; fma.rn.ftz.f32 %f1401, %f2493, %f348, %f2490; fma.rn.ftz.f32 %f1402, %f2492, %f348, %f2489; fma.rn.ftz.f32 %f1403, %f2491, %f348, %f2488; sub.ftz.f32 %f349, %f2362, %f1401; sub.ftz.f32 %f350, %f2363, %f1402; sub.ftz.f32 %f351, %f2364, %f1403; sub.ftz.f32 %f352, %f2366, %f1401; sub.ftz.f32 %f353, %f2367, %f1402; sub.ftz.f32 %f354, %f2368, %f1403; sub.ftz.f32 %f355, %f2370, %f1401; sub.ftz.f32 %f356, %f2371, %f1402; sub.ftz.f32 %f357, %f2372, %f1403; sub.ftz.f32 %f358, %f2374, %f1401; sub.ftz.f32 %f359, %f2375, %f1402; sub.ftz.f32 %f360, %f2376, %f1403; mov.u32 %r271, 0; mov.f32 %f2446, 0f00000000; mov.f32 %f2447, %f602; mov.u32 %r267, %r271; bra.uni $L__BB21_89; $L__BB21_133: add.s32 %r271, %r271, -1; mul.wide.s32 %rd128, %r271, 12; add.s64 %rd129, %rd1, %rd128; ld.local.f32 %f2446, [%rd129]; ld.local.f32 %f2447, [%rd129+4]; ld.local.u32 %r267, [%rd129+8]; $L__BB21_89: setp.gt.s32 %p117, %r267, 1; @%p117 bra $L__BB21_129; $L__BB21_90: mov.u32 %r272, %r267; cvt.rn.f32.s32 %f1406, %r272; mul.ftz.f32 %f376, %f1406, 0f3F000000; sub.ftz.f32 %f377, %f2447, %f2446; mul.ftz.f32 %f1407, %f377, 0f3EAAAAAB; mul.ftz.f32 %f1408, %f1407, 0f3F000000; fma.rn.ftz.f32 %f378, %f377, %f376, %f2446; add.ftz.f32 %f1409, %f376, 0f3F000000; fma.rn.ftz.f32 %f379, %f377, %f1409, %f2446; mov.f32 %f1410, 0f3F800000; sub.ftz.f32 %f1411, %f1410, %f378; mul.ftz.f32 %f1412, %f378, %f1411; mul.ftz.f32 %f1413, %f1411, %f1412; mul.ftz.f32 %f1414, %f378, %f378; mul.ftz.f32 %f1415, %f378, 0f40400000; add.ftz.f32 %f1416, %f1415, 0fC0A00000; fma.rn.ftz.f32 %f1417, %f1414, %f1416, 0f40000000; mul.ftz.f32 %f1418, %f1411, %f1411; mul.ftz.f32 %f1419, %f1411, 0f40400000; add.ftz.f32 %f1420, %f1419, 0fC0A00000; fma.rn.ftz.f32 %f1421, %f1418, %f1420, 0f40000000; mul.ftz.f32 %f1422, %f378, %f1412; mul.ftz.f32 %f1423, %f349, %f1413; mul.ftz.f32 %f1424, %f350, %f1413; mul.ftz.f32 %f1425, %f351, %f1413; mul.ftz.f32 %f1426, %f245, %f1413; mul.ftz.f32 %f1427, %f352, %f1417; mul.ftz.f32 %f1428, %f353, %f1417; mul.ftz.f32 %f1429, %f354, %f1417; mul.ftz.f32 %f1430, %f249, %f1417; sub.ftz.f32 %f1431, %f1427, %f1423; sub.ftz.f32 %f1432, %f1428, %f1424; sub.ftz.f32 %f1433, %f1429, %f1425; sub.ftz.f32 %f1434, %f1430, %f1426; fma.rn.ftz.f32 %f1435, %f355, %f1421, %f1431; fma.rn.ftz.f32 %f1436, %f356, %f1421, %f1432; fma.rn.ftz.f32 %f1437, %f357, %f1421, %f1433; fma.rn.ftz.f32 %f1438, %f253, %f1421, %f1434; mul.ftz.f32 %f1439, %f358, %f1422; mul.ftz.f32 %f1440, %f359, %f1422; mul.ftz.f32 %f1441, %f360, %f1422; mul.ftz.f32 %f1442, %f257, %f1422; sub.ftz.f32 %f1443, %f1435, %f1439; sub.ftz.f32 %f1444, %f1436, %f1440; sub.ftz.f32 %f1445, %f1437, %f1441; sub.ftz.f32 %f1446, %f1438, %f1442; mul.ftz.f32 %f1447, %f1443, 0f3F000000; mul.ftz.f32 %f1448, %f1444, 0f3F000000; mul.ftz.f32 %f1449, %f1445, 0f3F000000; mul.ftz.f32 %f1450, %f1446, 0f3F000000; add.ftz.f32 %f1451, %f1411, %f1411; mul.ftz.f32 %f1452, %f378, %f1451; sub.ftz.f32 %f1453, %f1452, %f1418; add.ftz.f32 %f1454, %f378, %f378; mul.ftz.f32 %f1455, %f1454, %f1416; fma.rn.ftz.f32 %f1456, %f378, %f1415, %f1455; add.ftz.f32 %f1457, %f1415, 0f40000000; mul.ftz.f32 %f1458, %f1451, %f1457; mul.ftz.f32 %f1459, %f1411, %f1419; sub.ftz.f32 %f1460, %f1458, %f1459; mul.ftz.f32 %f1461, %f1411, 0fC0000000; fma.rn.ftz.f32 %f1462, %f378, %f1461, %f1414; mul.ftz.f32 %f1463, %f352, %f1456; mul.ftz.f32 %f1464, %f353, %f1456; mul.ftz.f32 %f1465, %f354, %f1456; mul.ftz.f32 %f1466, %f249, %f1456; fma.rn.ftz.f32 %f1467, %f349, %f1453, %f1463; fma.rn.ftz.f32 %f1468, %f350, %f1453, %f1464; fma.rn.ftz.f32 %f1469, %f351, %f1453, %f1465; fma.rn.ftz.f32 %f1470, %f245, %f1453, %f1466; fma.rn.ftz.f32 %f1471, %f355, %f1460, %f1467; fma.rn.ftz.f32 %f1472, %f356, %f1460, %f1468; fma.rn.ftz.f32 %f1473, %f357, %f1460, %f1469; fma.rn.ftz.f32 %f1474, %f253, %f1460, %f1470; fma.rn.ftz.f32 %f1475, %f358, %f1462, %f1471; fma.rn.ftz.f32 %f1476, %f359, %f1462, %f1472; fma.rn.ftz.f32 %f1477, %f360, %f1462, %f1473; fma.rn.ftz.f32 %f1478, %f257, %f1462, %f1474; mul.ftz.f32 %f1479, %f1475, 0f3F000000; mul.ftz.f32 %f1480, %f1476, 0f3F000000; mul.ftz.f32 %f1481, %f1477, 0f3F000000; mul.ftz.f32 %f1482, %f1478, 0f3F000000; mul.ftz.f32 %f380, %f1408, %f1479; mul.ftz.f32 %f381, %f1408, %f1480; mul.ftz.f32 %f382, %f1408, %f1481; sub.ftz.f32 %f1483, %f1410, %f379; mul.ftz.f32 %f1484, %f379, %f1483; mul.ftz.f32 %f1485, %f1483, %f1484; mul.ftz.f32 %f1486, %f379, %f379; mul.ftz.f32 %f1487, %f379, 0f40400000; add.ftz.f32 %f1488, %f1487, 0fC0A00000; fma.rn.ftz.f32 %f1489, %f1486, %f1488, 0f40000000; mul.ftz.f32 %f1490, %f1483, %f1483; mul.ftz.f32 %f1491, %f1483, 0f40400000; add.ftz.f32 %f1492, %f1491, 0fC0A00000; fma.rn.ftz.f32 %f1493, %f1490, %f1492, 0f40000000; mul.ftz.f32 %f1494, %f379, %f1484; mul.ftz.f32 %f1495, %f349, %f1485; mul.ftz.f32 %f1496, %f350, %f1485; mul.ftz.f32 %f1497, %f351, %f1485; mul.ftz.f32 %f1498, %f245, %f1485; mul.ftz.f32 %f1499, %f352, %f1489; mul.ftz.f32 %f1500, %f353, %f1489; mul.ftz.f32 %f1501, %f354, %f1489; mul.ftz.f32 %f1502, %f249, %f1489; sub.ftz.f32 %f1503, %f1499, %f1495; sub.ftz.f32 %f1504, %f1500, %f1496; sub.ftz.f32 %f1505, %f1501, %f1497; sub.ftz.f32 %f1506, %f1502, %f1498; fma.rn.ftz.f32 %f1507, %f355, %f1493, %f1503; fma.rn.ftz.f32 %f1508, %f356, %f1493, %f1504; fma.rn.ftz.f32 %f1509, %f357, %f1493, %f1505; fma.rn.ftz.f32 %f1510, %f253, %f1493, %f1506; mul.ftz.f32 %f1511, %f358, %f1494; mul.ftz.f32 %f1512, %f359, %f1494; mul.ftz.f32 %f1513, %f360, %f1494; mul.ftz.f32 %f1514, %f257, %f1494; sub.ftz.f32 %f1515, %f1507, %f1511; sub.ftz.f32 %f1516, %f1508, %f1512; sub.ftz.f32 %f1517, %f1509, %f1513; sub.ftz.f32 %f1518, %f1510, %f1514; mul.ftz.f32 %f383, %f1515, 0f3F000000; mul.ftz.f32 %f384, %f1516, 0f3F000000; mul.ftz.f32 %f385, %f1517, 0f3F000000; mul.ftz.f32 %f1519, %f1518, 0f3F000000; add.ftz.f32 %f1520, %f1483, %f1483; mul.ftz.f32 %f1521, %f379, %f1520; sub.ftz.f32 %f1522, %f1521, %f1490; add.ftz.f32 %f1523, %f379, %f379; mul.ftz.f32 %f1524, %f1523, %f1488; fma.rn.ftz.f32 %f1525, %f379, %f1487, %f1524; add.ftz.f32 %f1526, %f1487, 0f40000000; mul.ftz.f32 %f1527, %f1520, %f1526; mul.ftz.f32 %f1528, %f1483, %f1491; sub.ftz.f32 %f1529, %f1527, %f1528; mul.ftz.f32 %f1530, %f1483, 0fC0000000; fma.rn.ftz.f32 %f1531, %f379, %f1530, %f1486; mul.ftz.f32 %f1532, %f352, %f1525; mul.ftz.f32 %f1533, %f353, %f1525; mul.ftz.f32 %f1534, %f354, %f1525; mul.ftz.f32 %f1535, %f249, %f1525; fma.rn.ftz.f32 %f1536, %f349, %f1522, %f1532; fma.rn.ftz.f32 %f1537, %f350, %f1522, %f1533; fma.rn.ftz.f32 %f1538, %f351, %f1522, %f1534; fma.rn.ftz.f32 %f1539, %f245, %f1522, %f1535; fma.rn.ftz.f32 %f1540, %f355, %f1529, %f1536; fma.rn.ftz.f32 %f1541, %f356, %f1529, %f1537; fma.rn.ftz.f32 %f1542, %f357, %f1529, %f1538; fma.rn.ftz.f32 %f1543, %f253, %f1529, %f1539; fma.rn.ftz.f32 %f1544, %f358, %f1531, %f1540; fma.rn.ftz.f32 %f1545, %f359, %f1531, %f1541; fma.rn.ftz.f32 %f1546, %f360, %f1531, %f1542; fma.rn.ftz.f32 %f1547, %f257, %f1531, %f1543; mul.ftz.f32 %f1548, %f1544, 0f3F000000; mul.ftz.f32 %f1549, %f1545, 0f3F000000; mul.ftz.f32 %f1550, %f1546, 0f3F000000; mul.ftz.f32 %f1551, %f1547, 0f3F000000; mul.ftz.f32 %f386, %f1408, %f1548; mul.ftz.f32 %f387, %f1408, %f1549; mul.ftz.f32 %f388, %f1408, %f1550; mul.ftz.f32 %f1552, %f1408, %f1551; fma.rn.ftz.f32 %f1553, %f1408, %f1482, %f1450; sub.ftz.f32 %f1554, %f1519, %f1552; sub.ftz.f32 %f1555, %f383, %f1447; sub.ftz.f32 %f1556, %f384, %f1448; sub.ftz.f32 %f1557, %f385, %f1449; mul.ftz.f32 %f1558, %f381, %f1557; mul.ftz.f32 %f1559, %f382, %f1556; sub.ftz.f32 %f1560, %f1558, %f1559; mul.ftz.f32 %f1561, %f382, %f1555; mul.ftz.f32 %f1562, %f380, %f1557; sub.ftz.f32 %f1563, %f1561, %f1562; mul.ftz.f32 %f1564, %f380, %f1556; mul.ftz.f32 %f1565, %f381, %f1555; sub.ftz.f32 %f1566, %f1564, %f1565; mul.ftz.f32 %f1567, %f1563, %f1563; fma.rn.ftz.f32 %f1568, %f1560, %f1560, %f1567; fma.rn.ftz.f32 %f1569, %f1566, %f1566, %f1568; mul.ftz.f32 %f1570, %f1556, %f1556; fma.rn.ftz.f32 %f1571, %f1555, %f1555, %f1570; fma.rn.ftz.f32 %f1572, %f1557, %f1557, %f1571; div.approx.ftz.f32 %f1573, %f1569, %f1572; mul.ftz.f32 %f1574, %f387, %f1557; mul.ftz.f32 %f1575, %f388, %f1556; sub.ftz.f32 %f1576, %f1574, %f1575; mul.ftz.f32 %f1577, %f388, %f1555; mul.ftz.f32 %f1578, %f386, %f1557; sub.ftz.f32 %f1579, %f1577, %f1578; mul.ftz.f32 %f1580, %f386, %f1556; mul.ftz.f32 %f1581, %f387, %f1555; sub.ftz.f32 %f1582, %f1580, %f1581; mul.ftz.f32 %f1583, %f1579, %f1579; fma.rn.ftz.f32 %f1584, %f1576, %f1576, %f1583; fma.rn.ftz.f32 %f1585, %f1582, %f1582, %f1584; div.approx.ftz.f32 %f1586, %f1585, %f1572; max.ftz.f32 %f1587, %f1573, %f1586; sqrt.approx.ftz.f32 %f1588, %f1587; max.ftz.f32 %f1589, %f1450, %f1553; max.ftz.f32 %f1590, %f1554, %f1519; max.ftz.f32 %f1591, %f1589, %f1590; add.ftz.f32 %f1592, %f1588, %f1591; min.ftz.f32 %f1593, %f1450, %f1553; min.ftz.f32 %f1594, %f1554, %f1519; min.ftz.f32 %f1595, %f1593, %f1594; sub.ftz.f32 %f389, %f1595, %f1588; mul.ftz.f32 %f1596, %f1592, 0f3F800002; rsqrt.approx.ftz.f32 %f390, %f1572; mul.ftz.f32 %f1597, %f1555, %f390; mul.ftz.f32 %f1598, %f1556, %f390; mul.ftz.f32 %f1599, %f1557, %f390; mul.ftz.f32 %f391, %f1443, 0fBF000000; mul.ftz.f32 %f392, %f1444, 0fBF000000; mul.ftz.f32 %f393, %f1445, 0fBF000000; mul.ftz.f32 %f1600, %f2493, %f391; mul.ftz.f32 %f1601, %f2492, %f1448; sub.ftz.f32 %f1602, %f1600, %f1601; mul.ftz.f32 %f1603, %f2491, %f1449; sub.ftz.f32 %f1604, %f1602, %f1603; mul.ftz.f32 %f1605, %f1448, %f1448; fma.rn.ftz.f32 %f1606, %f1447, %f1447, %f1605; fma.rn.ftz.f32 %f1607, %f1449, %f1449, %f1606; mul.ftz.f32 %f1608, %f2492, %f1598; fma.rn.ftz.f32 %f1609, %f2493, %f1597, %f1608; fma.rn.ftz.f32 %f394, %f2491, %f1599, %f1609; mul.ftz.f32 %f1610, %f1597, %f391; mul.ftz.f32 %f1611, %f1598, %f1448; sub.ftz.f32 %f1612, %f1610, %f1611; mul.ftz.f32 %f1613, %f1599, %f1449; sub.ftz.f32 %f395, %f1612, %f1613; mul.ftz.f32 %f396, %f394, %f394; sub.ftz.f32 %f397, %f75, %f396; mul.ftz.f32 %f1614, %f394, %f395; sub.ftz.f32 %f1615, %f1604, %f1614; add.ftz.f32 %f398, %f1615, %f1615; mul.ftz.f32 %f1616, %f395, %f395; sub.ftz.f32 %f399, %f1607, %f1616; mul.ftz.f32 %f1617, %f1596, %f1596; sub.ftz.f32 %f400, %f399, %f1617; mul.ftz.f32 %f401, %f398, %f398; mul.ftz.f32 %f402, %f397, 0f40800000; mul.ftz.f32 %f1618, %f402, %f400; sub.ftz.f32 %f403, %f401, %f1618; setp.ltu.ftz.f32 %p119, %f403, 0f00000000; mov.pred %p208, -1; mov.f32 %f2420, 0fFF7FFFFF; mov.f32 %f2419, 0f7F7FFFFF; @%p119 bra $L__BB21_94; abs.ftz.f32 %f1619, %f75; abs.ftz.f32 %f1620, %f396; max.ftz.f32 %f1621, %f1619, %f1620; mul.ftz.f32 %f1622, %f1621, 0f36000000; abs.ftz.f32 %f1623, %f397; setp.lt.ftz.f32 %p120, %f1623, %f1622; @%p120 bra $L__BB21_93; bra.uni $L__BB21_92; $L__BB21_93: setp.gtu.ftz.f32 %p208, %f400, 0f00000000; mov.f32 %f2420, 0f7F7FFFFF; mov.f32 %f2419, 0fFF7FFFFF; bra.uni $L__BB21_94; $L__BB21_92: sqrt.approx.ftz.f32 %f1624, %f403; add.ftz.f32 %f1625, %f397, %f397; rcp.approx.ftz.f32 %f1626, %f1625; neg.ftz.f32 %f1627, %f398; sub.ftz.f32 %f1628, %f1627, %f1624; mul.ftz.f32 %f2419, %f1626, %f1628; sub.ftz.f32 %f1629, %f1624, %f398; mul.ftz.f32 %f2420, %f1626, %f1629; fma.rn.ftz.f32 %f1630, %f394, %f2419, %f395; mul.ftz.f32 %f2477, %f390, %f1630; fma.rn.ftz.f32 %f1631, %f394, %f2420, %f395; mul.ftz.f32 %f2476, %f390, %f1631; mov.pred %p208, 0; $L__BB21_94: @%p208 bra $L__BB21_128; sub.ftz.f32 %f2244, %f10, %f348; sub.ftz.f32 %f1634, %f2466, %f348; min.ftz.f32 %f1635, %f1634, %f2420; mul.ftz.f32 %f1636, %f381, %f392; fma.rn.ftz.f32 %f1637, %f380, %f391, %f1636; fma.rn.ftz.f32 %f1638, %f382, %f393, %f1637; mul.ftz.f32 %f1639, %f2492, %f381; fma.rn.ftz.f32 %f1640, %f2493, %f380, %f1639; fma.rn.ftz.f32 %f1641, %f2491, %f382, %f1640; abs.ftz.f32 %f1642, %f1641; setp.lt.ftz.f32 %p122, %f1642, 0f219392EF; neg.ftz.f32 %f1643, %f1638; div.approx.ftz.f32 %f1644, %f1643, %f1641; setp.lt.ftz.f32 %p123, %f1641, 0f00000000; or.pred %p124, %p123, %p122; selp.f32 %f1645, 0fFF7FFFFF, %f1644, %p124; setp.gt.ftz.f32 %p125, %f1641, 0f00000000; or.pred %p126, %p125, %p122; selp.f32 %f1646, 0f7F7FFFFF, %f1644, %p126; max.ftz.f32 %f1647, %f2244, %f2419; max.ftz.f32 %f1648, %f1647, %f1645; min.ftz.f32 %f1649, %f1635, %f1646; mul.ftz.f32 %f1650, %f384, %f387; fma.rn.ftz.f32 %f1651, %f383, %f386, %f1650; fma.rn.ftz.f32 %f1652, %f385, %f388, %f1651; mul.ftz.f32 %f1653, %f2492, %f387; neg.ftz.f32 %f1654, %f1653; mul.ftz.f32 %f1655, %f2493, %f386; sub.ftz.f32 %f1656, %f1654, %f1655; mul.ftz.f32 %f1657, %f2491, %f388; sub.ftz.f32 %f1658, %f1656, %f1657; abs.ftz.f32 %f1659, %f1658; setp.lt.ftz.f32 %p127, %f1659, 0f219392EF; neg.ftz.f32 %f1660, %f1652; div.approx.ftz.f32 %f1661, %f1660, %f1658; setp.lt.ftz.f32 %p128, %f1658, 0f00000000; or.pred %p129, %p128, %p127; selp.f32 %f1662, 0fFF7FFFFF, %f1661, %p129; setp.gt.ftz.f32 %p130, %f1658, 0f00000000; or.pred %p131, %p130, %p127; selp.f32 %f1663, 0f7F7FFFFF, %f1661, %p131; max.ftz.f32 %f2423, %f1648, %f1662; min.ftz.f32 %f2432, %f1649, %f1663; setp.gtu.ftz.f32 %p132, %f2423, %f2432; @%p132 bra $L__BB21_128; mul.ftz.f32 %f2249, %f398, %f398; sub.ftz.f32 %f2248, %f2447, %f2446; cvt.rn.f32.s32 %f2247, %r272; mul.ftz.f32 %f2246, %f2247, 0f3F000000; mov.f32 %f2245, 0f3F800000; mul.ftz.f32 %f1666, %f389, 0f3F7FFFFC; mov.f32 %f1667, 0f00000000; max.ftz.f32 %f1668, %f1667, %f1666; max.ftz.f32 %f1669, %f2477, %f1667; min.ftz.f32 %f1671, %f1669, %f2245; max.ftz.f32 %f1672, %f2476, %f1667; min.ftz.f32 %f1673, %f1672, %f2245; add.ftz.f32 %f1674, %f2246, %f1671; mul.ftz.f32 %f1675, %f1674, 0f3EAAAAAB; fma.rn.ftz.f32 %f2477, %f2248, %f1675, %f2446; add.ftz.f32 %f1676, %f2246, %f1673; mul.ftz.f32 %f1677, %f1676, 0f3EAAAAAB; fma.rn.ftz.f32 %f2476, %f2248, %f1677, %f2446; mul.ftz.f32 %f1678, %f1668, %f1668; sub.ftz.f32 %f1679, %f399, %f1678; mul.ftz.f32 %f1680, %f402, %f1679; sub.ftz.f32 %f416, %f2249, %f1680; setp.ltu.ftz.f32 %p133, %f416, 0f00000000; mov.f32 %f2422, 0fFF7FFFFF; mov.f32 %f2421, 0f7F7FFFFF; @%p133 bra $L__BB21_99; abs.ftz.f32 %f1683, %f75; abs.ftz.f32 %f1684, %f396; max.ftz.f32 %f1685, %f1683, %f1684; mul.ftz.f32 %f1686, %f1685, 0f36000000; abs.ftz.f32 %f1687, %f397; setp.lt.ftz.f32 %p134, %f1687, %f1686; mov.f32 %f2422, 0f7F7FFFFF; mov.f32 %f2421, 0fFF7FFFFF; @%p134 bra $L__BB21_99; sqrt.approx.ftz.f32 %f1688, %f416; add.ftz.f32 %f1689, %f397, %f397; rcp.approx.ftz.f32 %f1690, %f1689; neg.ftz.f32 %f1691, %f398; sub.ftz.f32 %f1692, %f1691, %f1688; mul.ftz.f32 %f2421, %f1690, %f1692; sub.ftz.f32 %f1693, %f1688, %f398; mul.ftz.f32 %f2422, %f1690, %f1693; $L__BB21_99: min.ftz.f32 %f421, %f2432, %f2421; setp.gtu.ftz.f32 %p135, %f2423, %f421; max.ftz.f32 %f422, %f2423, %f2422; setp.gtu.ftz.f32 %p136, %f422, %f2432; and.pred %p137, %p135, %p136; @%p137 bra $L__BB21_128; mov.u16 %rs64, 0; @%p135 bra $L__BB21_113; setp.lt.s32 %p139, %r271, 4; mov.u16 %rs64, 1; @%p139 bra $L__BB21_113; sqrt.approx.ftz.f32 %f1694, %f75; min.ftz.f32 %f1695, %f355, %f358; min.ftz.f32 %f1696, %f349, %f352; min.ftz.f32 %f1697, %f1696, %f1695; min.ftz.f32 %f1698, %f356, %f359; min.ftz.f32 %f1699, %f350, %f353; min.ftz.f32 %f1700, %f1699, %f1698; min.ftz.f32 %f1701, %f357, %f360; min.ftz.f32 %f1702, %f351, %f354; min.ftz.f32 %f1703, %f1702, %f1701; min.ftz.f32 %f1704, %f253, %f257; min.ftz.f32 %f1705, %f245, %f249; min.ftz.f32 %f1706, %f1705, %f1704; max.ftz.f32 %f1707, %f355, %f358; max.ftz.f32 %f1708, %f1696, %f1707; max.ftz.f32 %f1709, %f356, %f359; max.ftz.f32 %f1710, %f1699, %f1709; max.ftz.f32 %f1711, %f357, %f360; max.ftz.f32 %f1712, %f1702, %f1711; max.ftz.f32 %f1713, %f253, %f257; max.ftz.f32 %f1714, %f1705, %f1713; abs.ftz.f32 %f1715, %f1697; abs.ftz.f32 %f1716, %f1700; abs.ftz.f32 %f1717, %f1703; abs.ftz.f32 %f1718, %f1706; abs.ftz.f32 %f1719, %f1708; abs.ftz.f32 %f1720, %f1710; abs.ftz.f32 %f1721, %f1712; abs.ftz.f32 %f1722, %f1714; max.ftz.f32 %f1723, %f1715, %f1719; max.ftz.f32 %f1724, %f1716, %f1720; max.ftz.f32 %f1725, %f1717, %f1721; max.ftz.f32 %f1726, %f1718, %f1722; max.ftz.f32 %f1727, %f1725, %f1726; max.ftz.f32 %f1728, %f1724, %f1727; max.ftz.f32 %f1729, %f1723, %f1728; mul.ftz.f32 %f423, %f1729, 0f36000000; mov.u32 %r269, 0; mov.f32 %f2424, %f2477; $L__BB21_103: mul.ftz.f32 %f2261, %f1694, 0f36000000; mov.f32 %f2255, 0f3F800000; sub.ftz.f32 %f1732, %f2255, %f2424; mul.ftz.f32 %f1733, %f2424, %f1732; mul.ftz.f32 %f1734, %f1732, %f1733; mul.ftz.f32 %f1735, %f2424, 0f40400000; add.ftz.f32 %f1736, %f1735, 0fC0A00000; mul.ftz.f32 %f1737, %f2424, %f2424; fma.rn.ftz.f32 %f1738, %f1737, %f1736, 0f40000000; mov.f32 %f1739, 0f40000000; mul.ftz.f32 %f1740, %f1732, %f1732; mul.ftz.f32 %f1741, %f1732, 0f40400000; add.ftz.f32 %f1742, %f1741, 0fC0A00000; fma.rn.ftz.f32 %f1743, %f1740, %f1742, 0f40000000; mul.ftz.f32 %f1744, %f2424, %f1733; mul.ftz.f32 %f1745, %f349, %f1734; mul.ftz.f32 %f1746, %f350, %f1734; mul.ftz.f32 %f1747, %f351, %f1734; mul.ftz.f32 %f1748, %f245, %f1734; mul.ftz.f32 %f1749, %f352, %f1738; mul.ftz.f32 %f1750, %f353, %f1738; mul.ftz.f32 %f1751, %f354, %f1738; mul.ftz.f32 %f1752, %f249, %f1738; sub.ftz.f32 %f1753, %f1749, %f1745; sub.ftz.f32 %f1754, %f1750, %f1746; sub.ftz.f32 %f1755, %f1751, %f1747; sub.ftz.f32 %f1756, %f1752, %f1748; fma.rn.ftz.f32 %f1757, %f355, %f1743, %f1753; fma.rn.ftz.f32 %f1758, %f356, %f1743, %f1754; fma.rn.ftz.f32 %f1759, %f357, %f1743, %f1755; fma.rn.ftz.f32 %f1760, %f253, %f1743, %f1756; mul.ftz.f32 %f1761, %f358, %f1744; mul.ftz.f32 %f1762, %f359, %f1744; mul.ftz.f32 %f1763, %f360, %f1744; mul.ftz.f32 %f1764, %f257, %f1744; sub.ftz.f32 %f1765, %f1757, %f1761; sub.ftz.f32 %f1766, %f1758, %f1762; sub.ftz.f32 %f1767, %f1759, %f1763; sub.ftz.f32 %f1768, %f1760, %f1764; mul.ftz.f32 %f1769, %f1765, 0f3F000000; mul.ftz.f32 %f1770, %f1766, 0f3F000000; mul.ftz.f32 %f1771, %f1767, 0f3F000000; mul.ftz.f32 %f428, %f1768, 0f3F000000; add.ftz.f32 %f1772, %f1732, %f1732; mul.ftz.f32 %f1773, %f2424, %f1772; sub.ftz.f32 %f1774, %f1773, %f1740; add.ftz.f32 %f1775, %f2424, %f2424; mul.ftz.f32 %f1776, %f1775, %f1736; fma.rn.ftz.f32 %f1777, %f2424, %f1735, %f1776; add.ftz.f32 %f1778, %f1735, 0f40000000; mul.ftz.f32 %f1779, %f1772, %f1778; mul.ftz.f32 %f1780, %f1732, %f1741; sub.ftz.f32 %f1781, %f1779, %f1780; mul.ftz.f32 %f1782, %f1732, 0fC0000000; fma.rn.ftz.f32 %f1783, %f2424, %f1782, %f1737; mul.ftz.f32 %f1784, %f352, %f1777; mul.ftz.f32 %f1785, %f353, %f1777; mul.ftz.f32 %f1786, %f354, %f1777; mul.ftz.f32 %f1787, %f249, %f1777; fma.rn.ftz.f32 %f1788, %f349, %f1774, %f1784; fma.rn.ftz.f32 %f1789, %f350, %f1774, %f1785; fma.rn.ftz.f32 %f1790, %f351, %f1774, %f1786; fma.rn.ftz.f32 %f1791, %f245, %f1774, %f1787; fma.rn.ftz.f32 %f1792, %f355, %f1781, %f1788; fma.rn.ftz.f32 %f1793, %f356, %f1781, %f1789; fma.rn.ftz.f32 %f1794, %f357, %f1781, %f1790; fma.rn.ftz.f32 %f1795, %f253, %f1781, %f1791; fma.rn.ftz.f32 %f1796, %f358, %f1783, %f1792; fma.rn.ftz.f32 %f1797, %f359, %f1783, %f1793; fma.rn.ftz.f32 %f1798, %f360, %f1783, %f1794; fma.rn.ftz.f32 %f1799, %f257, %f1783, %f1795; mul.ftz.f32 %f429, %f1796, 0f3F000000; mul.ftz.f32 %f430, %f1797, 0f3F000000; mul.ftz.f32 %f431, %f1798, 0f3F000000; mul.ftz.f32 %f432, %f1799, 0f3F000000; sub.ftz.f32 %f1800, %f1739, %f1735; mul.ftz.f32 %f1801, %f2424, 0f41100000; add.ftz.f32 %f1802, %f1801, 0fC0A00000; mov.f32 %f1803, 0f40800000; sub.ftz.f32 %f1804, %f1803, %f1801; add.ftz.f32 %f1805, %f1735, 0fBF800000; mul.ftz.f32 %f1806, %f352, %f1802; mul.ftz.f32 %f1807, %f353, %f1802; mul.ftz.f32 %f1808, %f354, %f1802; fma.rn.ftz.f32 %f1809, %f349, %f1800, %f1806; fma.rn.ftz.f32 %f1810, %f350, %f1800, %f1807; fma.rn.ftz.f32 %f1811, %f351, %f1800, %f1808; fma.rn.ftz.f32 %f1812, %f355, %f1804, %f1809; fma.rn.ftz.f32 %f1813, %f356, %f1804, %f1810; fma.rn.ftz.f32 %f1814, %f357, %f1804, %f1811; fma.rn.ftz.f32 %f1815, %f358, %f1805, %f1812; fma.rn.ftz.f32 %f1816, %f359, %f1805, %f1813; fma.rn.ftz.f32 %f1817, %f360, %f1805, %f1814; mul.ftz.f32 %f1818, %f2493, %f2423; sub.ftz.f32 %f433, %f1818, %f1769; mul.ftz.f32 %f1819, %f2492, %f2423; sub.ftz.f32 %f434, %f1819, %f1770; mul.ftz.f32 %f1820, %f2491, %f2423; sub.ftz.f32 %f435, %f1820, %f1771; mul.ftz.f32 %f1821, %f434, %f434; fma.rn.ftz.f32 %f1822, %f433, %f433, %f1821; fma.rn.ftz.f32 %f1823, %f435, %f435, %f1822; sqrt.approx.ftz.f32 %f436, %f1823; mul.ftz.f32 %f1824, %f2261, %f2423; max.ftz.f32 %f437, %f1824, %f423; mul.ftz.f32 %f1825, %f1796, 0fBF000000; mul.ftz.f32 %f1826, %f430, %f430; fma.rn.ftz.f32 %f1827, %f429, %f429, %f1826; fma.rn.ftz.f32 %f1828, %f431, %f431, %f1827; sqrt.approx.ftz.f32 %f1829, %f1828; div.approx.ftz.f32 %f1830, %f429, %f1829; div.approx.ftz.f32 %f1831, %f430, %f1829; div.approx.ftz.f32 %f1832, %f431, %f1829; mul.ftz.f32 %f1833, %f1816, %f430; fma.rn.ftz.f32 %f1834, %f1815, %f429, %f1833; fma.rn.ftz.f32 %f1835, %f1817, %f431, %f1834; mul.ftz.f32 %f1836, %f1815, %f1828; mul.ftz.f32 %f1837, %f1816, %f1828; mul.ftz.f32 %f1838, %f1817, %f1828; mul.ftz.f32 %f1839, %f429, %f1835; mul.ftz.f32 %f1840, %f430, %f1835; mul.ftz.f32 %f1841, %f431, %f1835; sub.ftz.f32 %f1842, %f1836, %f1839; sub.ftz.f32 %f1843, %f1837, %f1840; sub.ftz.f32 %f1844, %f1838, %f1841; mul.ftz.f32 %f1845, %f1828, %f1829; div.approx.ftz.f32 %f1846, %f1842, %f1845; div.approx.ftz.f32 %f1847, %f1843, %f1845; div.approx.ftz.f32 %f1848, %f1844, %f1845; div.approx.ftz.f32 %f1849, %f423, %f1829; mul.ftz.f32 %f1850, %f434, %f1831; fma.rn.ftz.f32 %f1851, %f433, %f1830, %f1850; fma.rn.ftz.f32 %f438, %f435, %f1832, %f1851; fma.rn.ftz.f32 %f1852, %f423, %f436, %f437; add.ftz.f32 %f1853, %f436, 0f3F800000; fma.rn.ftz.f32 %f439, %f1853, %f1849, %f1852; mul.ftz.f32 %f1854, %f1830, %f1825; mul.ftz.f32 %f1855, %f1831, %f430; sub.ftz.f32 %f1856, %f1854, %f1855; mul.ftz.f32 %f1857, %f1832, %f431; sub.ftz.f32 %f1858, %f1856, %f1857; mul.ftz.f32 %f1859, %f434, %f1847; fma.rn.ftz.f32 %f1860, %f433, %f1846, %f1859; fma.rn.ftz.f32 %f1861, %f435, %f1848, %f1860; add.ftz.f32 %f440, %f1858, %f1861; mul.ftz.f32 %f1862, %f2492, %f1831; fma.rn.ftz.f32 %f1863, %f2493, %f1830, %f1862; fma.rn.ftz.f32 %f441, %f2491, %f1832, %f1863; mul.ftz.f32 %f1864, %f438, %f438; sub.ftz.f32 %f1865, %f1823, %f1864; mul.ftz.f32 %f1866, %f433, %f1825; mul.ftz.f32 %f1867, %f434, %f430; sub.ftz.f32 %f1868, %f1866, %f1867; mul.ftz.f32 %f1869, %f435, %f431; sub.ftz.f32 %f442, %f1868, %f1869; mul.ftz.f32 %f1870, %f2492, %f434; fma.rn.ftz.f32 %f1871, %f2493, %f433, %f1870; fma.rn.ftz.f32 %f443, %f2491, %f435, %f1871; setp.leu.ftz.f32 %p140, %f1865, 0f00000000; mov.f32 %f2425, 0f00000000; sqrt.approx.ftz.f32 %f444, %f1865; @%p140 bra $L__BB21_105; rcp.approx.ftz.f32 %f2425, %f444; $L__BB21_105: max.ftz.f32 %f2259, %f253, %f257; min.ftz.f32 %f2258, %f245, %f249; max.ftz.f32 %f2257, %f2258, %f2259; mul.ftz.f32 %f2256, %f2257, 0f36000000; add.ftz.f32 %f1872, %f437, %f439; add.ftz.f32 %f447, %f2256, %f1872; mul.ftz.f32 %f1873, %f438, %f440; sub.ftz.f32 %f1874, %f442, %f1873; mul.ftz.f32 %f1875, %f1874, %f2425; sub.ftz.f32 %f1876, %f1875, %f432; mul.ftz.f32 %f1877, %f438, %f441; sub.ftz.f32 %f1878, %f443, %f1877; mul.ftz.f32 %f1879, %f1878, %f2425; mul.ftz.f32 %f1880, %f440, %f1879; mul.ftz.f32 %f1881, %f441, %f1876; sub.ftz.f32 %f1882, %f1880, %f1881; rcp.approx.ftz.f32 %f1883, %f1882; mul.ftz.f32 %f1884, %f438, %f1879; sub.ftz.f32 %f448, %f444, %f428; mul.ftz.f32 %f1885, %f441, %f448; sub.ftz.f32 %f1886, %f1884, %f1885; mul.ftz.f32 %f1887, %f1886, %f1883; sub.ftz.f32 %f2424, %f2424, %f1887; mul.ftz.f32 %f1888, %f438, %f1876; mul.ftz.f32 %f1889, %f440, %f448; sub.ftz.f32 %f1890, %f1889, %f1888; mul.ftz.f32 %f1891, %f1890, %f1883; sub.ftz.f32 %f2423, %f2423, %f1891; abs.ftz.f32 %f1892, %f438; setp.geu.ftz.f32 %p141, %f1892, %f439; @%p141 bra $L__BB21_107; abs.ftz.f32 %f1893, %f448; setp.lt.ftz.f32 %p142, %f1893, %f447; @%p142 bra $L__BB21_108; $L__BB21_107: add.s32 %r269, %r269, 1; setp.lt.u32 %p143, %r269, 5; mov.u16 %rs62, 0; @%p143 bra $L__BB21_103; bra.uni $L__BB21_112; $L__BB21_108: add.ftz.f32 %f451, %f348, %f2423; setp.ltu.ftz.f32 %p144, %f451, %f10; setp.gtu.ftz.f32 %p145, %f451, %f2466; or.pred %p146, %p144, %p145; mov.u16 %rs62, 0; @%p146 bra $L__BB21_112; setp.ltu.ftz.f32 %p147, %f2424, 0f00000000; setp.gtu.ftz.f32 %p148, %f2424, 0f3F800000; or.pred %p149, %p148, %p147; @%p149 bra $L__BB21_112; div.approx.ftz.f32 %f1894, %f433, %f436; div.approx.ftz.f32 %f1895, %f434, %f436; div.approx.ftz.f32 %f1896, %f435, %f436; fma.rn.ftz.f32 %f1897, %f432, %f1894, %f429; fma.rn.ftz.f32 %f1898, %f432, %f1895, %f430; fma.rn.ftz.f32 %f1899, %f432, %f1896, %f431; mul.ftz.f32 %f1900, %f430, %f1896; mul.ftz.f32 %f1901, %f431, %f1895; sub.ftz.f32 %f1902, %f1900, %f1901; mul.ftz.f32 %f1903, %f431, %f1894; mul.ftz.f32 %f1904, %f429, %f1896; sub.ftz.f32 %f1905, %f1903, %f1904; mul.ftz.f32 %f1906, %f429, %f1895; mul.ftz.f32 %f1907, %f430, %f1894; sub.ftz.f32 %f1908, %f1906, %f1907; mul.ftz.f32 %f1909, %f1905, %f1899; mul.ftz.f32 %f1910, %f1908, %f1898; sub.ftz.f32 %f1911, %f1909, %f1910; mul.ftz.f32 %f1912, %f1897, %f1908; mul.ftz.f32 %f1913, %f1902, %f1899; sub.ftz.f32 %f1914, %f1912, %f1913; mul.ftz.f32 %f1915, %f1898, %f1902; mul.ftz.f32 %f1916, %f1897, %f1905; sub.ftz.f32 %f1917, %f1915, %f1916; mul.ftz.f32 %f1918, %f2492, %f1914; fma.rn.ftz.f32 %f1919, %f2493, %f1911, %f1918; fma.rn.ftz.f32 %f1920, %f2491, %f1917, %f1919; setp.gt.ftz.f32 %p150, %f1920, 0f00000000; @%p150 bra $L__BB21_112; st.local.f32 [%rd4], %f451; st.local.f32 [%rd4+4], %f2424; mov.u32 %r205, 0; st.local.u32 [%rd4+8], %r205; mov.u16 %rs62, 1; mov.f32 %f2466, %f451; $L__BB21_112: mov.u16 %rs64, 0; or.b16 %rs40, %rs62, %rs68; and.b16 %rs41, %rs40, 255; setp.ne.s16 %p151, %rs41, 0; selp.u16 %rs68, 1, 0, %p151; $L__BB21_113: setp.gtu.ftz.f32 %p204, %f422, %f2432; add.ftz.f32 %f1921, %f348, %f422; setp.ltu.ftz.f32 %p152, %f1921, %f10; setp.gtu.ftz.f32 %p153, %f1921, %f2466; or.pred %p154, %p152, %p153; or.pred %p156, %p204, %p154; mov.u16 %rs67, %rs64; @%p156 bra $L__BB21_126; setp.lt.s32 %p157, %r271, 4; mov.u16 %rs67, 1; @%p157 bra $L__BB21_126; sqrt.approx.ftz.f32 %f1922, %f75; min.ftz.f32 %f1923, %f355, %f358; min.ftz.f32 %f1924, %f349, %f352; min.ftz.f32 %f1925, %f1924, %f1923; min.ftz.f32 %f1926, %f356, %f359; min.ftz.f32 %f1927, %f350, %f353; min.ftz.f32 %f1928, %f1927, %f1926; min.ftz.f32 %f1929, %f357, %f360; min.ftz.f32 %f1930, %f351, %f354; min.ftz.f32 %f1931, %f1930, %f1929; min.ftz.f32 %f1932, %f253, %f257; min.ftz.f32 %f1933, %f245, %f249; min.ftz.f32 %f1934, %f1933, %f1932; max.ftz.f32 %f1935, %f355, %f358; max.ftz.f32 %f1936, %f1924, %f1935; max.ftz.f32 %f1937, %f356, %f359; max.ftz.f32 %f1938, %f1927, %f1937; max.ftz.f32 %f1939, %f357, %f360; max.ftz.f32 %f1940, %f1930, %f1939; max.ftz.f32 %f1941, %f253, %f257; max.ftz.f32 %f1942, %f1933, %f1941; abs.ftz.f32 %f1943, %f1925; abs.ftz.f32 %f1944, %f1928; abs.ftz.f32 %f1945, %f1931; abs.ftz.f32 %f1946, %f1934; abs.ftz.f32 %f1947, %f1936; abs.ftz.f32 %f1948, %f1938; abs.ftz.f32 %f1949, %f1940; abs.ftz.f32 %f1950, %f1942; max.ftz.f32 %f1951, %f1943, %f1947; max.ftz.f32 %f1952, %f1944, %f1948; max.ftz.f32 %f1953, %f1945, %f1949; max.ftz.f32 %f1954, %f1946, %f1950; max.ftz.f32 %f1955, %f1953, %f1954; max.ftz.f32 %f1956, %f1952, %f1955; max.ftz.f32 %f1957, %f1951, %f1956; mul.ftz.f32 %f458, %f1957, 0f36000000; mov.u32 %r270, 0; mov.f32 %f2433, %f2476; $L__BB21_116: mul.ftz.f32 %f2266, %f1922, 0f36000000; mov.f32 %f2434, 0f00000000; mov.f32 %f2254, 0f3F800000; sub.ftz.f32 %f1960, %f2254, %f2433; mul.ftz.f32 %f1961, %f2433, %f1960; mul.ftz.f32 %f1962, %f1960, %f1961; mul.ftz.f32 %f1963, %f2433, 0f40400000; add.ftz.f32 %f1964, %f1963, 0fC0A00000; mul.ftz.f32 %f1965, %f2433, %f2433; fma.rn.ftz.f32 %f1966, %f1965, %f1964, 0f40000000; mov.f32 %f1967, 0f40000000; mul.ftz.f32 %f1968, %f1960, %f1960; mul.ftz.f32 %f1969, %f1960, 0f40400000; add.ftz.f32 %f1970, %f1969, 0fC0A00000; fma.rn.ftz.f32 %f1971, %f1968, %f1970, 0f40000000; mul.ftz.f32 %f1972, %f2433, %f1961; mul.ftz.f32 %f1973, %f349, %f1962; mul.ftz.f32 %f1974, %f350, %f1962; mul.ftz.f32 %f1975, %f351, %f1962; mul.ftz.f32 %f1976, %f245, %f1962; mul.ftz.f32 %f1977, %f352, %f1966; mul.ftz.f32 %f1978, %f353, %f1966; mul.ftz.f32 %f1979, %f354, %f1966; mul.ftz.f32 %f1980, %f249, %f1966; sub.ftz.f32 %f1981, %f1977, %f1973; sub.ftz.f32 %f1982, %f1978, %f1974; sub.ftz.f32 %f1983, %f1979, %f1975; sub.ftz.f32 %f1984, %f1980, %f1976; fma.rn.ftz.f32 %f1985, %f355, %f1971, %f1981; fma.rn.ftz.f32 %f1986, %f356, %f1971, %f1982; fma.rn.ftz.f32 %f1987, %f357, %f1971, %f1983; fma.rn.ftz.f32 %f1988, %f253, %f1971, %f1984; mul.ftz.f32 %f1989, %f358, %f1972; mul.ftz.f32 %f1990, %f359, %f1972; mul.ftz.f32 %f1991, %f360, %f1972; mul.ftz.f32 %f1992, %f257, %f1972; sub.ftz.f32 %f1993, %f1985, %f1989; sub.ftz.f32 %f1994, %f1986, %f1990; sub.ftz.f32 %f1995, %f1987, %f1991; sub.ftz.f32 %f1996, %f1988, %f1992; mul.ftz.f32 %f1997, %f1993, 0f3F000000; mul.ftz.f32 %f1998, %f1994, 0f3F000000; mul.ftz.f32 %f1999, %f1995, 0f3F000000; mul.ftz.f32 %f463, %f1996, 0f3F000000; add.ftz.f32 %f2000, %f1960, %f1960; mul.ftz.f32 %f2001, %f2433, %f2000; sub.ftz.f32 %f2002, %f2001, %f1968; add.ftz.f32 %f2003, %f2433, %f2433; mul.ftz.f32 %f2004, %f2003, %f1964; fma.rn.ftz.f32 %f2005, %f2433, %f1963, %f2004; add.ftz.f32 %f2006, %f1963, 0f40000000; mul.ftz.f32 %f2007, %f2000, %f2006; mul.ftz.f32 %f2008, %f1960, %f1969; sub.ftz.f32 %f2009, %f2007, %f2008; mul.ftz.f32 %f2010, %f1960, 0fC0000000; fma.rn.ftz.f32 %f2011, %f2433, %f2010, %f1965; mul.ftz.f32 %f2012, %f352, %f2005; mul.ftz.f32 %f2013, %f353, %f2005; mul.ftz.f32 %f2014, %f354, %f2005; mul.ftz.f32 %f2015, %f249, %f2005; fma.rn.ftz.f32 %f2016, %f349, %f2002, %f2012; fma.rn.ftz.f32 %f2017, %f350, %f2002, %f2013; fma.rn.ftz.f32 %f2018, %f351, %f2002, %f2014; fma.rn.ftz.f32 %f2019, %f245, %f2002, %f2015; fma.rn.ftz.f32 %f2020, %f355, %f2009, %f2016; fma.rn.ftz.f32 %f2021, %f356, %f2009, %f2017; fma.rn.ftz.f32 %f2022, %f357, %f2009, %f2018; fma.rn.ftz.f32 %f2023, %f253, %f2009, %f2019; fma.rn.ftz.f32 %f2024, %f358, %f2011, %f2020; fma.rn.ftz.f32 %f2025, %f359, %f2011, %f2021; fma.rn.ftz.f32 %f2026, %f360, %f2011, %f2022; fma.rn.ftz.f32 %f2027, %f257, %f2011, %f2023; mul.ftz.f32 %f464, %f2024, 0f3F000000; mul.ftz.f32 %f465, %f2025, 0f3F000000; mul.ftz.f32 %f466, %f2026, 0f3F000000; mul.ftz.f32 %f467, %f2027, 0f3F000000; sub.ftz.f32 %f2028, %f1967, %f1963; mul.ftz.f32 %f2029, %f2433, 0f41100000; add.ftz.f32 %f2030, %f2029, 0fC0A00000; mov.f32 %f2031, 0f40800000; sub.ftz.f32 %f2032, %f2031, %f2029; add.ftz.f32 %f2033, %f1963, 0fBF800000; mul.ftz.f32 %f2034, %f352, %f2030; mul.ftz.f32 %f2035, %f353, %f2030; mul.ftz.f32 %f2036, %f354, %f2030; fma.rn.ftz.f32 %f2037, %f349, %f2028, %f2034; fma.rn.ftz.f32 %f2038, %f350, %f2028, %f2035; fma.rn.ftz.f32 %f2039, %f351, %f2028, %f2036; fma.rn.ftz.f32 %f2040, %f355, %f2032, %f2037; fma.rn.ftz.f32 %f2041, %f356, %f2032, %f2038; fma.rn.ftz.f32 %f2042, %f357, %f2032, %f2039; fma.rn.ftz.f32 %f2043, %f358, %f2033, %f2040; fma.rn.ftz.f32 %f2044, %f359, %f2033, %f2041; fma.rn.ftz.f32 %f2045, %f360, %f2033, %f2042; mul.ftz.f32 %f2046, %f2493, %f2432; sub.ftz.f32 %f468, %f2046, %f1997; mul.ftz.f32 %f2047, %f2492, %f2432; sub.ftz.f32 %f469, %f2047, %f1998; mul.ftz.f32 %f2048, %f2491, %f2432; sub.ftz.f32 %f470, %f2048, %f1999; mul.ftz.f32 %f2049, %f469, %f469; fma.rn.ftz.f32 %f2050, %f468, %f468, %f2049; fma.rn.ftz.f32 %f2051, %f470, %f470, %f2050; sqrt.approx.ftz.f32 %f471, %f2051; mul.ftz.f32 %f2052, %f2266, %f2432; max.ftz.f32 %f472, %f2052, %f458; mul.ftz.f32 %f2053, %f2024, 0fBF000000; mul.ftz.f32 %f2054, %f465, %f465; fma.rn.ftz.f32 %f2055, %f464, %f464, %f2054; fma.rn.ftz.f32 %f2056, %f466, %f466, %f2055; sqrt.approx.ftz.f32 %f2057, %f2056; div.approx.ftz.f32 %f2058, %f464, %f2057; div.approx.ftz.f32 %f2059, %f465, %f2057; div.approx.ftz.f32 %f2060, %f466, %f2057; mul.ftz.f32 %f2061, %f2044, %f465; fma.rn.ftz.f32 %f2062, %f2043, %f464, %f2061; fma.rn.ftz.f32 %f2063, %f2045, %f466, %f2062; mul.ftz.f32 %f2064, %f2043, %f2056; mul.ftz.f32 %f2065, %f2044, %f2056; mul.ftz.f32 %f2066, %f2045, %f2056; mul.ftz.f32 %f2067, %f464, %f2063; mul.ftz.f32 %f2068, %f465, %f2063; mul.ftz.f32 %f2069, %f466, %f2063; sub.ftz.f32 %f2070, %f2064, %f2067; sub.ftz.f32 %f2071, %f2065, %f2068; sub.ftz.f32 %f2072, %f2066, %f2069; mul.ftz.f32 %f2073, %f2056, %f2057; div.approx.ftz.f32 %f2074, %f2070, %f2073; div.approx.ftz.f32 %f2075, %f2071, %f2073; div.approx.ftz.f32 %f2076, %f2072, %f2073; div.approx.ftz.f32 %f2077, %f458, %f2057; mul.ftz.f32 %f2078, %f469, %f2059; fma.rn.ftz.f32 %f2079, %f468, %f2058, %f2078; fma.rn.ftz.f32 %f473, %f470, %f2060, %f2079; fma.rn.ftz.f32 %f2080, %f458, %f471, %f472; add.ftz.f32 %f2081, %f471, 0f3F800000; fma.rn.ftz.f32 %f474, %f2081, %f2077, %f2080; mul.ftz.f32 %f2082, %f2058, %f2053; mul.ftz.f32 %f2083, %f2059, %f465; sub.ftz.f32 %f2084, %f2082, %f2083; mul.ftz.f32 %f2085, %f2060, %f466; sub.ftz.f32 %f2086, %f2084, %f2085; mul.ftz.f32 %f2087, %f469, %f2075; fma.rn.ftz.f32 %f2088, %f468, %f2074, %f2087; fma.rn.ftz.f32 %f2089, %f470, %f2076, %f2088; add.ftz.f32 %f475, %f2086, %f2089; mul.ftz.f32 %f2090, %f2492, %f2059; fma.rn.ftz.f32 %f2091, %f2493, %f2058, %f2090; fma.rn.ftz.f32 %f476, %f2491, %f2060, %f2091; mul.ftz.f32 %f2092, %f473, %f473; sub.ftz.f32 %f2093, %f2051, %f2092; mul.ftz.f32 %f2094, %f468, %f2053; mul.ftz.f32 %f2095, %f469, %f465; sub.ftz.f32 %f2096, %f2094, %f2095; mul.ftz.f32 %f2097, %f470, %f466; sub.ftz.f32 %f477, %f2096, %f2097; mul.ftz.f32 %f2098, %f2492, %f469; fma.rn.ftz.f32 %f2099, %f2493, %f468, %f2098; fma.rn.ftz.f32 %f478, %f2491, %f470, %f2099; setp.leu.ftz.f32 %p158, %f2093, 0f00000000; sqrt.approx.ftz.f32 %f479, %f2093; @%p158 bra $L__BB21_118; rcp.approx.ftz.f32 %f2434, %f479; $L__BB21_118: max.ftz.f32 %f2265, %f253, %f257; min.ftz.f32 %f2264, %f245, %f249; max.ftz.f32 %f2263, %f2264, %f2265; mul.ftz.f32 %f2262, %f2263, 0f36000000; add.ftz.f32 %f2100, %f472, %f474; add.ftz.f32 %f482, %f2262, %f2100; mul.ftz.f32 %f2101, %f473, %f475; sub.ftz.f32 %f2102, %f477, %f2101; mul.ftz.f32 %f2103, %f2102, %f2434; sub.ftz.f32 %f2104, %f2103, %f467; mul.ftz.f32 %f2105, %f473, %f476; sub.ftz.f32 %f2106, %f478, %f2105; mul.ftz.f32 %f2107, %f2106, %f2434; mul.ftz.f32 %f2108, %f475, %f2107; mul.ftz.f32 %f2109, %f476, %f2104; sub.ftz.f32 %f2110, %f2108, %f2109; rcp.approx.ftz.f32 %f2111, %f2110; mul.ftz.f32 %f2112, %f473, %f2107; sub.ftz.f32 %f483, %f479, %f463; mul.ftz.f32 %f2113, %f476, %f483; sub.ftz.f32 %f2114, %f2112, %f2113; mul.ftz.f32 %f2115, %f2114, %f2111; sub.ftz.f32 %f2433, %f2433, %f2115; mul.ftz.f32 %f2116, %f473, %f2104; mul.ftz.f32 %f2117, %f475, %f483; sub.ftz.f32 %f2118, %f2117, %f2116; mul.ftz.f32 %f2119, %f2118, %f2111; sub.ftz.f32 %f2432, %f2432, %f2119; abs.ftz.f32 %f2120, %f473; setp.geu.ftz.f32 %p159, %f2120, %f474; @%p159 bra $L__BB21_120; abs.ftz.f32 %f2121, %f483; setp.lt.ftz.f32 %p160, %f2121, %f482; @%p160 bra $L__BB21_121; $L__BB21_120: add.s32 %r270, %r270, 1; setp.lt.u32 %p161, %r270, 5; mov.u16 %rs65, 0; @%p161 bra $L__BB21_116; bra.uni $L__BB21_125; $L__BB21_121: add.ftz.f32 %f486, %f348, %f2432; setp.ltu.ftz.f32 %p162, %f486, %f10; setp.gtu.ftz.f32 %p163, %f486, %f2466; or.pred %p164, %p162, %p163; mov.u16 %rs65, 0; @%p164 bra $L__BB21_125; setp.ltu.ftz.f32 %p165, %f2433, 0f00000000; setp.gtu.ftz.f32 %p166, %f2433, 0f3F800000; or.pred %p167, %p166, %p165; @%p167 bra $L__BB21_125; div.approx.ftz.f32 %f2122, %f468, %f471; div.approx.ftz.f32 %f2123, %f469, %f471; div.approx.ftz.f32 %f2124, %f470, %f471; fma.rn.ftz.f32 %f2125, %f467, %f2122, %f464; fma.rn.ftz.f32 %f2126, %f467, %f2123, %f465; fma.rn.ftz.f32 %f2127, %f467, %f2124, %f466; mul.ftz.f32 %f2128, %f465, %f2124; mul.ftz.f32 %f2129, %f466, %f2123; sub.ftz.f32 %f2130, %f2128, %f2129; mul.ftz.f32 %f2131, %f466, %f2122; mul.ftz.f32 %f2132, %f464, %f2124; sub.ftz.f32 %f2133, %f2131, %f2132; mul.ftz.f32 %f2134, %f464, %f2123; mul.ftz.f32 %f2135, %f465, %f2122; sub.ftz.f32 %f2136, %f2134, %f2135; mul.ftz.f32 %f2137, %f2133, %f2127; mul.ftz.f32 %f2138, %f2136, %f2126; sub.ftz.f32 %f2139, %f2137, %f2138; mul.ftz.f32 %f2140, %f2125, %f2136; mul.ftz.f32 %f2141, %f2130, %f2127; sub.ftz.f32 %f2142, %f2140, %f2141; mul.ftz.f32 %f2143, %f2126, %f2130; mul.ftz.f32 %f2144, %f2125, %f2133; sub.ftz.f32 %f2145, %f2143, %f2144; mul.ftz.f32 %f2146, %f2492, %f2142; fma.rn.ftz.f32 %f2147, %f2493, %f2139, %f2146; fma.rn.ftz.f32 %f2148, %f2491, %f2145, %f2147; setp.gt.ftz.f32 %p168, %f2148, 0f00000000; @%p168 bra $L__BB21_125; st.local.f32 [%rd4], %f486; st.local.f32 [%rd4+4], %f2433; mov.u32 %r207, 0; st.local.u32 [%rd4+8], %r207; mov.u16 %rs65, 1; mov.f32 %f2466, %f486; $L__BB21_125: or.b16 %rs48, %rs65, %rs68; and.b16 %rs49, %rs48, 255; setp.ne.s16 %p169, %rs49, 0; selp.u16 %rs68, 1, 0, %p169; mov.u16 %rs67, %rs64; $L__BB21_126: setp.eq.s16 %p170, %rs67, 0; @%p170 bra $L__BB21_128; cvt.rn.f32.s32 %f2253, %r272; mul.ftz.f32 %f2252, %f2253, 0f3F000000; sub.ftz.f32 %f2251, %f2447, %f2446; fma.rn.ftz.f32 %f2250, %f2251, %f2252, %f2446; mul.wide.s32 %rd126, %r271, 12; add.s64 %rd127, %rd1, %rd126; st.local.f32 [%rd127], %f2446; st.local.f32 [%rd127+4], %f2447; add.s32 %r209, %r272, 1; st.local.u32 [%rd127+8], %r209; add.s32 %r271, %r271, 1; mov.u32 %r272, -1; mov.f32 %f2446, %f2250; mov.f32 %f2447, %f379; $L__BB21_128: add.s32 %r267, %r272, 1; setp.lt.s32 %p171, %r272, 1; @%p171 bra $L__BB21_90; $L__BB21_129: setp.gt.s32 %p172, %r271, 0; @%p172 bra $L__BB21_133; and.b16 %rs50, %rs68, 255; setp.eq.s16 %p174, %rs50, 0; mov.pred %p209, 0; @%p174 bra $L__BB21_132; st.local.u32 [%rd4+12], %r49; st.local.u32 [%rd4+16], %r48; st.local.u32 [%rd4+20], %r66; mov.pred %p209, -1; mov.u32 %r276, %r49; bra.uni $L__BB21_132; $L__BB21_72: rsqrt.approx.ftz.f32 %f258, %f75; mul.ftz.f32 %f1037, %f2493, %f258; mul.ftz.f32 %f1038, %f2492, %f258; mul.ftz.f32 %f1039, %f2491, %f258; neg.ftz.f32 %f1040, %f1038; neg.ftz.f32 %f1041, %f1039; fma.rn.ftz.f32 %f1042, %f1039, %f1039, 0f00000000; fma.rn.ftz.f32 %f1043, %f1038, %f1038, %f1042; fma.rn.ftz.f32 %f1044, %f1037, %f1037, %f1042; setp.gt.ftz.f32 %p94, %f1043, %f1044; selp.f32 %f1045, 0f00000000, %f1041, %p94; selp.f32 %f1046, %f1039, 0f00000000, %p94; selp.f32 %f1047, %f1040, %f1037, %p94; mul.ftz.f32 %f1048, %f1046, %f1046; fma.rn.ftz.f32 %f1049, %f1045, %f1045, %f1048; fma.rn.ftz.f32 %f1050, %f1047, %f1047, %f1049; sqrt.approx.ftz.f32 %f1051, %f1050; div.approx.ftz.f32 %f1052, %f1045, %f1051; div.approx.ftz.f32 %f1053, %f1046, %f1051; div.approx.ftz.f32 %f1054, %f1047, %f1051; mul.ftz.f32 %f1055, %f1038, %f1054; mul.ftz.f32 %f1056, %f1039, %f1053; sub.ftz.f32 %f1057, %f1055, %f1056; mul.ftz.f32 %f1058, %f1039, %f1052; mul.ftz.f32 %f1059, %f1037, %f1054; sub.ftz.f32 %f1060, %f1058, %f1059; mul.ftz.f32 %f1061, %f1037, %f1053; mul.ftz.f32 %f1062, %f1038, %f1052; sub.ftz.f32 %f1063, %f1061, %f1062; mul.ftz.f32 %f1064, %f1060, %f1060; fma.rn.ftz.f32 %f1065, %f1057, %f1057, %f1064; fma.rn.ftz.f32 %f1066, %f1063, %f1063, %f1065; sqrt.approx.ftz.f32 %f1067, %f1066; div.approx.ftz.f32 %f1068, %f1057, %f1067; div.approx.ftz.f32 %f1069, %f1060, %f1067; div.approx.ftz.f32 %f1070, %f1063, %f1067; mul.ftz.f32 %f1071, %f258, %f1037; mul.ftz.f32 %f1072, %f258, %f1038; mul.ftz.f32 %f1073, %f258, %f1039; sub.ftz.f32 %f1074, %f2362, %f2490; sub.ftz.f32 %f1075, %f2363, %f2489; mul.ftz.f32 %f1076, %f1075, %f1053; fma.rn.ftz.f32 %f1077, %f1074, %f1052, %f1076; sub.ftz.f32 %f1078, %f2364, %f2488; fma.rn.ftz.f32 %f259, %f1078, %f1054, %f1077; mul.ftz.f32 %f1079, %f1075, %f1069; fma.rn.ftz.f32 %f1080, %f1074, %f1068, %f1079; fma.rn.ftz.f32 %f260, %f1078, %f1070, %f1080; mul.ftz.f32 %f1081, %f1075, %f1072; fma.rn.ftz.f32 %f1082, %f1074, %f1071, %f1081; fma.rn.ftz.f32 %f261, %f1078, %f1073, %f1082; sub.ftz.f32 %f1083, %f2366, %f2490; sub.ftz.f32 %f1084, %f2367, %f2489; mul.ftz.f32 %f1085, %f1084, %f1053; fma.rn.ftz.f32 %f1086, %f1083, %f1052, %f1085; sub.ftz.f32 %f1087, %f2368, %f2488; fma.rn.ftz.f32 %f262, %f1087, %f1054, %f1086; mul.ftz.f32 %f1088, %f1084, %f1069; fma.rn.ftz.f32 %f1089, %f1083, %f1068, %f1088; fma.rn.ftz.f32 %f263, %f1087, %f1070, %f1089; mul.ftz.f32 %f1090, %f1084, %f1072; fma.rn.ftz.f32 %f1091, %f1083, %f1071, %f1090; fma.rn.ftz.f32 %f264, %f1087, %f1073, %f1091; sub.ftz.f32 %f1092, %f2370, %f2490; sub.ftz.f32 %f1093, %f2371, %f2489; mul.ftz.f32 %f1094, %f1093, %f1053; fma.rn.ftz.f32 %f1095, %f1092, %f1052, %f1094; sub.ftz.f32 %f1096, %f2372, %f2488; fma.rn.ftz.f32 %f265, %f1096, %f1054, %f1095; mul.ftz.f32 %f1097, %f1093, %f1069; fma.rn.ftz.f32 %f1098, %f1092, %f1068, %f1097; fma.rn.ftz.f32 %f266, %f1096, %f1070, %f1098; mul.ftz.f32 %f1099, %f1093, %f1072; fma.rn.ftz.f32 %f1100, %f1092, %f1071, %f1099; fma.rn.ftz.f32 %f267, %f1096, %f1073, %f1100; sub.ftz.f32 %f1101, %f2374, %f2490; sub.ftz.f32 %f1102, %f2375, %f2489; mul.ftz.f32 %f1103, %f1102, %f1053; fma.rn.ftz.f32 %f1104, %f1101, %f1052, %f1103; sub.ftz.f32 %f1105, %f2376, %f2488; fma.rn.ftz.f32 %f268, %f1105, %f1054, %f1104; mul.ftz.f32 %f1106, %f1102, %f1069; fma.rn.ftz.f32 %f1107, %f1101, %f1068, %f1106; fma.rn.ftz.f32 %f269, %f1105, %f1070, %f1107; mul.ftz.f32 %f1108, %f1102, %f1072; fma.rn.ftz.f32 %f1109, %f1101, %f1071, %f1108; fma.rn.ftz.f32 %f270, %f1105, %f1073, %f1109; abs.ftz.f32 %f1110, %f259; abs.ftz.f32 %f1111, %f260; abs.ftz.f32 %f1112, %f261; abs.ftz.f32 %f1113, %f262; abs.ftz.f32 %f1114, %f263; abs.ftz.f32 %f1115, %f264; max.ftz.f32 %f1116, %f1110, %f1113; max.ftz.f32 %f1117, %f1111, %f1114; max.ftz.f32 %f1118, %f1112, %f1115; abs.ftz.f32 %f1119, %f249; abs.ftz.f32 %f1120, %f245; max.ftz.f32 %f1121, %f1120, %f1119; abs.ftz.f32 %f1122, %f265; abs.ftz.f32 %f1123, %f266; abs.ftz.f32 %f1124, %f267; abs.ftz.f32 %f1125, %f268; abs.ftz.f32 %f1126, %f269; abs.ftz.f32 %f1127, %f270; max.ftz.f32 %f1128, %f1122, %f1125; max.ftz.f32 %f1129, %f1123, %f1126; max.ftz.f32 %f1130, %f1124, %f1127; abs.ftz.f32 %f1131, %f257; abs.ftz.f32 %f1132, %f253; max.ftz.f32 %f1133, %f1132, %f1131; max.ftz.f32 %f1134, %f1116, %f1128; max.ftz.f32 %f1135, %f1117, %f1129; max.ftz.f32 %f1136, %f1118, %f1130; max.ftz.f32 %f1137, %f1121, %f1133; max.ftz.f32 %f1138, %f1134, %f1135; max.ftz.f32 %f1139, %f1136, %f1137; max.ftz.f32 %f1140, %f1138, %f1139; mul.ftz.f32 %f271, %f1140, 0f35000000; sub.ftz.f32 %f1141, %f602, 0f00000000; mul.ftz.f32 %f1142, %f1141, 0f00000000; mul.ftz.f32 %f1143, %f1142, %f1141; mul.ftz.f32 %f1144, %f259, %f1143; mul.ftz.f32 %f1145, %f260, %f1143; mul.ftz.f32 %f1146, %f261, %f1143; add.ftz.f32 %f1147, %f262, %f262; add.ftz.f32 %f1148, %f263, %f263; add.ftz.f32 %f1149, %f264, %f264; sub.ftz.f32 %f1150, %f1147, %f1144; sub.ftz.f32 %f1151, %f1148, %f1145; sub.ftz.f32 %f1152, %f1149, %f1146; add.ftz.f32 %f1153, %f249, %f249; mul.ftz.f32 %f1154, %f245, %f1143; sub.ftz.f32 %f1155, %f1153, %f1154; mul.ftz.f32 %f1156, %f1141, %f1141; mul.ftz.f32 %f1157, %f1141, 0f40400000; add.ftz.f32 %f1158, %f1157, 0fC0A00000; fma.rn.ftz.f32 %f1159, %f1156, %f1158, 0f40000000; fma.rn.ftz.f32 %f1160, %f265, %f1159, %f1150; fma.rn.ftz.f32 %f1161, %f266, %f1159, %f1151; fma.rn.ftz.f32 %f1162, %f267, %f1159, %f1152; fma.rn.ftz.f32 %f1163, %f253, %f1159, %f1155; mul.ftz.f32 %f1164, %f1142, 0f00000000; mul.ftz.f32 %f1165, %f268, %f1164; mul.ftz.f32 %f1166, %f269, %f1164; mul.ftz.f32 %f1167, %f270, %f1164; sub.ftz.f32 %f1168, %f1160, %f1165; sub.ftz.f32 %f1169, %f1161, %f1166; sub.ftz.f32 %f272, %f1162, %f1167; mul.ftz.f32 %f1170, %f257, %f1164; sub.ftz.f32 %f1171, %f1163, %f1170; mul.ftz.f32 %f2383, %f1168, 0f3F000000; mul.ftz.f32 %f2382, %f1169, 0f3F000000; mul.ftz.f32 %f2380, %f1171, 0f3F000000; add.ftz.f32 %f1172, %f1141, %f1141; mul.ftz.f32 %f1173, %f1172, 0f00000000; sub.ftz.f32 %f1174, %f1173, %f1156; mul.ftz.f32 %f1175, %f262, 0f00000000; mul.ftz.f32 %f1176, %f263, 0f00000000; mul.ftz.f32 %f1177, %f264, 0f00000000; fma.rn.ftz.f32 %f1178, %f259, %f1174, %f1175; fma.rn.ftz.f32 %f1179, %f260, %f1174, %f1176; fma.rn.ftz.f32 %f1180, %f261, %f1174, %f1177; mul.ftz.f32 %f1181, %f1157, %f1141; add.ftz.f32 %f1182, %f1172, %f1172; sub.ftz.f32 %f1183, %f1182, %f1181; fma.rn.ftz.f32 %f1184, %f265, %f1183, %f1178; fma.rn.ftz.f32 %f1185, %f266, %f1183, %f1179; fma.rn.ftz.f32 %f1186, %f267, %f1183, %f1180; mul.ftz.f32 %f1187, %f1141, 0fC0000000; fma.rn.ftz.f32 %f1188, %f1187, 0f00000000, 0f00000000; fma.rn.ftz.f32 %f1189, %f268, %f1188, %f1184; fma.rn.ftz.f32 %f1190, %f269, %f1188, %f1185; fma.rn.ftz.f32 %f1191, %f270, %f1188, %f1186; mul.ftz.f32 %f2378, %f1189, 0f3F000000; mul.ftz.f32 %f2379, %f1190, 0f3F000000; mul.ftz.f32 %f1192, %f1191, 0f3F000000; abs.ftz.f32 %f1193, %f2378; abs.ftz.f32 %f1194, %f2379; abs.ftz.f32 %f1195, %f1192; max.ftz.f32 %f1196, %f1193, %f1194; max.ftz.f32 %f1197, %f1196, %f1195; setp.geu.ftz.f32 %p95, %f1197, %f271; @%p95 bra $L__BB21_74; mul.ftz.f32 %f1198, %f14, %f262; fma.rn.ftz.f32 %f1199, %f259, %f13, %f1198; mul.ftz.f32 %f1200, %f14, %f263; fma.rn.ftz.f32 %f1201, %f260, %f13, %f1200; fma.rn.ftz.f32 %f1202, %f265, %f15, %f1199; fma.rn.ftz.f32 %f1203, %f15, %f266, %f1201; fma.rn.ftz.f32 %f1204, %f268, %f16, %f1202; fma.rn.ftz.f32 %f1205, %f269, %f16, %f1203; mul.ftz.f32 %f1206, %f1204, 0f3F000000; mul.ftz.f32 %f1207, %f1205, 0f3F000000; sub.ftz.f32 %f2378, %f1206, %f2383; sub.ftz.f32 %f2379, %f1207, %f2382; $L__BB21_74: mul.ftz.f32 %f1208, %f2379, %f2379; fma.rn.ftz.f32 %f282, %f2378, %f2378, %f1208; setp.lt.s32 %p97, %r130, 1; mov.pred %p207, -1; @%p97 bra $L__BB21_86; add.ftz.f32 %f1209, %f282, 0f00000000; mov.f32 %f1210, 0f00000000; sqrt.approx.ftz.f32 %f1211, %f1209; mul.ftz.f32 %f2381, %f272, 0f3F000000; div.approx.ftz.f32 %f1212, %f1210, %f1211; mul.ftz.f32 %f2384, %f2380, %f1212; neg.ftz.f32 %f1213, %f2378; div.approx.ftz.f32 %f1214, %f1213, %f1211; mul.ftz.f32 %f2385, %f2380, %f1214; div.approx.ftz.f32 %f1215, %f2379, %f1211; mul.ftz.f32 %f2386, %f2380, %f1215; mov.u32 %r264, 0; $L__BB21_76: mov.f32 %f293, %f2386; mov.f32 %f292, %f2385; mov.f32 %f291, %f2384; mov.f32 %f290, %f2383; mov.f32 %f289, %f2382; mov.f32 %f288, %f2381; mov.f32 %f287, %f2380; cvt.rn.f32.s32 %f1216, %r264; mul.ftz.f32 %f297, %f12, %f1216; add.ftz.f32 %f1217, %f12, %f297; sub.ftz.f32 %f1219, %f602, %f1217; mul.ftz.f32 %f1220, %f1217, %f1219; mul.ftz.f32 %f1221, %f1219, %f1220; mul.ftz.f32 %f1222, %f1217, %f1217; mul.ftz.f32 %f1223, %f1217, 0f40400000; add.ftz.f32 %f1224, %f1223, 0fC0A00000; fma.rn.ftz.f32 %f1225, %f1222, %f1224, 0f40000000; mul.ftz.f32 %f1226, %f1219, %f1219; mul.ftz.f32 %f1227, %f1219, 0f40400000; add.ftz.f32 %f1228, %f1227, 0fC0A00000; fma.rn.ftz.f32 %f1229, %f1226, %f1228, 0f40000000; mul.ftz.f32 %f1230, %f1217, %f1220; mul.ftz.f32 %f1231, %f259, %f1221; mul.ftz.f32 %f1232, %f260, %f1221; mul.ftz.f32 %f1233, %f261, %f1221; mul.ftz.f32 %f1234, %f245, %f1221; mul.ftz.f32 %f1235, %f262, %f1225; mul.ftz.f32 %f1236, %f263, %f1225; mul.ftz.f32 %f1237, %f264, %f1225; mul.ftz.f32 %f1238, %f249, %f1225; sub.ftz.f32 %f1239, %f1235, %f1231; sub.ftz.f32 %f1240, %f1236, %f1232; sub.ftz.f32 %f1241, %f1237, %f1233; sub.ftz.f32 %f1242, %f1238, %f1234; fma.rn.ftz.f32 %f1243, %f265, %f1229, %f1239; fma.rn.ftz.f32 %f1244, %f266, %f1229, %f1240; fma.rn.ftz.f32 %f1245, %f267, %f1229, %f1241; fma.rn.ftz.f32 %f1246, %f253, %f1229, %f1242; mul.ftz.f32 %f1247, %f268, %f1230; mul.ftz.f32 %f1248, %f269, %f1230; mul.ftz.f32 %f1249, %f270, %f1230; mul.ftz.f32 %f1250, %f257, %f1230; sub.ftz.f32 %f1251, %f1243, %f1247; sub.ftz.f32 %f1252, %f1244, %f1248; sub.ftz.f32 %f1253, %f1245, %f1249; sub.ftz.f32 %f1254, %f1246, %f1250; mul.ftz.f32 %f2383, %f1251, 0f3F000000; mul.ftz.f32 %f2382, %f1252, 0f3F000000; mul.ftz.f32 %f2381, %f1253, 0f3F000000; mul.ftz.f32 %f2380, %f1254, 0f3F000000; sub.ftz.f32 %f302, %f2383, %f290; sub.ftz.f32 %f303, %f2382, %f289; mul.ftz.f32 %f1255, %f303, %f303; fma.rn.ftz.f32 %f304, %f302, %f302, %f1255; add.ftz.f32 %f1256, %f1219, %f1219; mul.ftz.f32 %f1257, %f1217, %f1256; sub.ftz.f32 %f1258, %f1257, %f1226; add.ftz.f32 %f1259, %f1217, %f1217; mul.ftz.f32 %f1260, %f1259, %f1224; fma.rn.ftz.f32 %f1261, %f1217, %f1223, %f1260; add.ftz.f32 %f1262, %f1223, 0f40000000; mul.ftz.f32 %f1263, %f1256, %f1262; mul.ftz.f32 %f1264, %f1219, %f1227; sub.ftz.f32 %f1265, %f1263, %f1264; mul.ftz.f32 %f1266, %f1219, 0fC0000000; fma.rn.ftz.f32 %f1267, %f1217, %f1266, %f1222; mul.ftz.f32 %f1268, %f262, %f1261; mul.ftz.f32 %f1269, %f263, %f1261; mul.ftz.f32 %f1270, %f264, %f1261; fma.rn.ftz.f32 %f1271, %f259, %f1258, %f1268; fma.rn.ftz.f32 %f1272, %f260, %f1258, %f1269; fma.rn.ftz.f32 %f1273, %f261, %f1258, %f1270; fma.rn.ftz.f32 %f1274, %f265, %f1265, %f1271; fma.rn.ftz.f32 %f1275, %f266, %f1265, %f1272; fma.rn.ftz.f32 %f1276, %f267, %f1265, %f1273; fma.rn.ftz.f32 %f1277, %f268, %f1267, %f1274; fma.rn.ftz.f32 %f1278, %f269, %f1267, %f1275; fma.rn.ftz.f32 %f1279, %f270, %f1267, %f1276; mul.ftz.f32 %f2390, %f1277, 0f3F000000; mul.ftz.f32 %f2391, %f1278, 0f3F000000; mul.ftz.f32 %f307, %f1279, 0f3F000000; abs.ftz.f32 %f1280, %f2390; abs.ftz.f32 %f1281, %f2391; abs.ftz.f32 %f1282, %f307; max.ftz.f32 %f1283, %f1280, %f1281; max.ftz.f32 %f1284, %f1283, %f1282; setp.lt.ftz.f32 %p98, %f1284, %f271; @%p98 bra $L__BB21_77; bra.uni $L__BB21_78; $L__BB21_77: mov.f32 %f2390, %f302; mov.f32 %f2391, %f303; $L__BB21_78: max.ftz.f32 %f1285, %f287, %f2380; mul.ftz.f32 %f1286, %f290, %f303; mul.ftz.f32 %f1287, %f289, %f302; sub.ftz.f32 %f1288, %f1287, %f1286; neg.ftz.f32 %f1289, %f2390; mul.ftz.f32 %f1290, %f2390, %f2390; fma.rn.ftz.f32 %f1291, %f2391, %f2391, %f1290; add.ftz.f32 %f1292, %f1291, 0f00000000; sqrt.approx.ftz.f32 %f1294, %f1292; div.approx.ftz.f32 %f1295, %f2391, %f1294; div.approx.ftz.f32 %f1296, %f1289, %f1294; div.approx.ftz.f32 %f1297, %f1210, %f1294; mul.ftz.f32 %f2386, %f2380, %f1295; mul.ftz.f32 %f2385, %f2380, %f1296; mul.ftz.f32 %f2384, %f2380, %f1297; mul.ftz.f32 %f1298, %f1285, %f1285; mul.ftz.f32 %f1299, %f304, %f1298; mul.ftz.f32 %f1300, %f1288, %f1288; setp.gtu.ftz.f32 %p99, %f1300, %f1299; @%p99 bra $L__BB21_85; add.ftz.f32 %f1301, %f293, %f290; sub.ftz.f32 %f1302, %f290, %f293; add.ftz.f32 %f1303, %f2383, %f2386; sub.ftz.f32 %f1304, %f1303, %f1302; sub.ftz.f32 %f1305, %f289, %f292; add.ftz.f32 %f1306, %f2382, %f2385; sub.ftz.f32 %f1307, %f1306, %f1305; sub.ftz.f32 %f1308, %f288, %f291; add.ftz.f32 %f1309, %f2381, %f2384; sub.ftz.f32 %f1310, %f1309, %f1308; mul.ftz.f32 %f1311, %f1305, %f1310; mul.ftz.f32 %f1312, %f1308, %f1307; sub.ftz.f32 %f1313, %f1311, %f1312; mul.ftz.f32 %f1314, %f1308, %f1304; mul.ftz.f32 %f1315, %f1302, %f1310; sub.ftz.f32 %f1316, %f1314, %f1315; mul.ftz.f32 %f1317, %f1302, %f1307; mul.ftz.f32 %f1318, %f1305, %f1304; sub.ftz.f32 %f1319, %f1317, %f1318; mul.ftz.f32 %f1320, %f1316, 0f00000000; fma.rn.ftz.f32 %f1321, %f1313, 0f00000000, %f1320; add.ftz.f32 %f320, %f1319, %f1321; setp.gtu.ftz.f32 %p101, %f320, 0f00000000; sub.ftz.f32 %f1322, %f2383, %f2386; selp.f32 %f321, %f1322, %f1301, %p101; sub.ftz.f32 %f1323, %f2382, %f2385; add.ftz.f32 %f1324, %f292, %f289; selp.f32 %f322, %f1323, %f1324, %p101; sub.ftz.f32 %f1325, %f2381, %f2384; add.ftz.f32 %f1326, %f291, %f288; selp.f32 %f323, %f1325, %f1326, %p101; selp.f32 %f1327, %f1302, %f1303, %p101; selp.f32 %f1328, %f1305, %f1306, %p101; selp.f32 %f1329, %f1308, %f1309, %p101; selp.f32 %f1330, %f1303, %f1302, %p101; selp.f32 %f1331, %f1306, %f1305, %p101; selp.f32 %f1332, %f1309, %f1308, %p101; sub.ftz.f32 %f324, %f1330, %f321; sub.ftz.f32 %f325, %f1331, %f322; sub.ftz.f32 %f326, %f1332, %f323; sub.ftz.f32 %f327, %f321, %f1327; sub.ftz.f32 %f328, %f322, %f1328; sub.ftz.f32 %f329, %f323, %f1329; mul.ftz.f32 %f1333, %f322, %f326; mul.ftz.f32 %f1334, %f323, %f325; sub.ftz.f32 %f1335, %f1333, %f1334; mul.ftz.f32 %f1336, %f323, %f324; mul.ftz.f32 %f1337, %f321, %f326; sub.ftz.f32 %f1338, %f1336, %f1337; mul.ftz.f32 %f1339, %f321, %f325; mul.ftz.f32 %f1340, %f322, %f324; sub.ftz.f32 %f1341, %f1339, %f1340; mul.ftz.f32 %f1342, %f1338, 0f00000000; fma.rn.ftz.f32 %f1343, %f1335, 0f00000000, %f1342; add.ftz.f32 %f330, %f1341, %f1343; mul.ftz.f32 %f1344, %f1328, %f329; mul.ftz.f32 %f1345, %f1329, %f328; sub.ftz.f32 %f1346, %f1344, %f1345; mul.ftz.f32 %f1347, %f1329, %f327; mul.ftz.f32 %f1348, %f1327, %f329; sub.ftz.f32 %f1349, %f1347, %f1348; mul.ftz.f32 %f1350, %f1327, %f328; mul.ftz.f32 %f1351, %f1328, %f327; sub.ftz.f32 %f1352, %f1350, %f1351; mul.ftz.f32 %f1353, %f1349, 0f00000000; fma.rn.ftz.f32 %f1354, %f1346, 0f00000000, %f1353; add.ftz.f32 %f331, %f1352, %f1354; max.ftz.f32 %f1355, %f330, %f331; setp.gtu.ftz.f32 %p102, %f1355, 0f00000000; mov.pred %p206, -1; @%p102 bra $L__BB21_82; mul.ftz.f32 %f1356, %f328, %f326; mul.ftz.f32 %f1357, %f329, %f325; sub.ftz.f32 %f1358, %f1356, %f1357; mul.ftz.f32 %f1359, %f327, %f326; mul.ftz.f32 %f1360, %f329, %f324; sub.ftz.f32 %f1361, %f1360, %f1359; mul.ftz.f32 %f1362, %f328, %f324; mul.ftz.f32 %f1363, %f327, %f325; sub.ftz.f32 %f1364, %f1363, %f1362; mul.ftz.f32 %f1365, %f1361, 0f00000000; fma.rn.ftz.f32 %f1366, %f1358, 0f00000000, %f1365; add.ftz.f32 %f1367, %f1364, %f1366; rcp.approx.ftz.f32 %f332, %f1367; mul.ftz.f32 %f1368, %f322, %f1361; fma.rn.ftz.f32 %f1369, %f321, %f1358, %f1368; fma.rn.ftz.f32 %f1370, %f323, %f1364, %f1369; mul.ftz.f32 %f333, %f1370, %f332; setp.ltu.ftz.f32 %p104, %f333, %f10; setp.gtu.ftz.f32 %p105, %f333, %f2466; or.pred %p106, %p105, %p104; setp.eq.ftz.f32 %p107, %f1367, 0f00000000; or.pred %p108, %p107, %p106; @%p108 bra $L__BB21_82; mul.ftz.f32 %f1371, %f330, %f332; sub.ftz.f32 %f1373, %f602, %f1371; setp.le.ftz.f32 %p110, %f320, 0f00000000; selp.f32 %f2475, %f1371, %f1373, %p110; mul.ftz.f32 %f1374, %f331, %f332; sub.ftz.f32 %f1375, %f602, %f1374; selp.f32 %f2474, %f1374, %f1375, %p110; mov.pred %p206, 0; mov.f32 %f2473, %f333; $L__BB21_82: @%p206 bra $L__BB21_85; sub.ftz.f32 %f1376, %f2380, %f287; fma.rn.ftz.f32 %f1377, %f1376, %f2475, %f287; add.ftz.f32 %f1378, %f1377, %f1377; mul.ftz.f32 %f1379, %f258, %f1378; setp.leu.ftz.f32 %p111, %f2473, %f1379; @%p111 bra $L__BB21_85; bra.uni $L__BB21_84; $L__BB21_85: add.s32 %r264, %r264, 1; setp.lt.s32 %p114, %r264, %r130; @%p114 bra $L__BB21_76; bra.uni $L__BB21_86; $L__BB21_49: setp.gt.s32 %p68, %r53, %r52; selp.b32 %r165, -1, 0, %p68; add.s32 %r166, %r53, %r165; mad.lo.s32 %r167, %r166, %r51, %r49; add.s32 %r168, %r167, %r257; mul.wide.s32 %rd98, %r168, 16; add.s64 %rd145, %rd21, %rd98; $L__BB21_51: ld.global.v4.f32 {%f888, %f889, %f890, %f891}, [%rd145]; add.s32 %r64, %r53, 1; setp.eq.s32 %p69, %r64, %r52; @%p69 bra $L__BB21_53; bra.uni $L__BB21_52; $L__BB21_53: shl.b64 %rd101, %rd26, 4; add.s64 %rd146, %rd20, %rd101; bra.uni $L__BB21_54; $L__BB21_52: setp.ge.s32 %p70, %r53, %r52; selp.b32 %r169, -1, 0, %p70; add.s32 %r170, %r64, %r169; mad.lo.s32 %r171, %r170, %r51, %r49; add.s32 %r172, %r171, %r257; mul.wide.s32 %rd100, %r172, 16; add.s64 %rd146, %rd21, %rd100; $L__BB21_54: cvt.rn.f32.s32 %f892, %r53; sub.ftz.f32 %f893, %f89, %f892; ld.global.v4.f32 {%f894, %f895, %f896, %f897}, [%rd146]; sub.ftz.f32 %f903, %f602, %f893; mul.ftz.f32 %f904, %f893, %f894; mul.ftz.f32 %f905, %f893, %f895; mul.ftz.f32 %f906, %f893, %f896; mul.ftz.f32 %f907, %f893, %f897; fma.rn.ftz.f32 %f2326, %f903, %f891, %f907; fma.rn.ftz.f32 %f2325, %f903, %f890, %f906; fma.rn.ftz.f32 %f2324, %f903, %f889, %f905; fma.rn.ftz.f32 %f2323, %f903, %f888, %f904; $L__BB21_56: sub.ftz.f32 %f912, %f2323, %f2490; sub.ftz.f32 %f913, %f2324, %f2489; sub.ftz.f32 %f914, %f2325, %f2488; mul.ftz.f32 %f915, %f2492, %f913; fma.rn.ftz.f32 %f916, %f2493, %f912, %f915; fma.rn.ftz.f32 %f917, %f2491, %f914, %f916; mul.ftz.f32 %f106, %f76, %f917; mul.ftz.f32 %f918, %f2493, %f106; mul.ftz.f32 %f919, %f2492, %f106; mul.ftz.f32 %f920, %f2491, %f106; sub.ftz.f32 %f921, %f912, %f918; sub.ftz.f32 %f922, %f913, %f919; sub.ftz.f32 %f923, %f914, %f920; mul.ftz.f32 %f924, %f922, %f922; fma.rn.ftz.f32 %f925, %f921, %f921, %f924; fma.rn.ftz.f32 %f107, %f923, %f923, %f925; mul.ftz.f32 %f108, %f2326, %f2326; setp.gtu.ftz.f32 %p72, %f107, %f108; mov.pred %p205, 0; @%p72 bra $L__BB21_59; sub.ftz.f32 %f926, %f108, %f107; mul.ftz.f32 %f927, %f76, %f926; sqrt.approx.ftz.f32 %f928, %f927; sub.ftz.f32 %f109, %f106, %f928; setp.ltu.ftz.f32 %p74, %f109, %f10; setp.gtu.ftz.f32 %p75, %f109, %f2466; or.pred %p76, %p75, %p74; @%p76 bra $L__BB21_59; st.local.f32 [%rd4], %f109; st.local.u32 [%rd4+12], %r49; st.local.u32 [%rd4+16], %r48; st.local.u32 [%rd4+20], %r50; mov.u32 %r173, 0; st.local.u32 [%rd4+4], %r173; st.local.u32 [%rd4+8], %r173; mov.pred %p205, -1; mov.u32 %r276, %r49; mov.f32 %f2466, %f109; $L__BB21_59: shr.u32 %r174, %r115, 8; and.b32 %r175, %r174, 1; setp.eq.b32 %p78, %r175, 1; and.pred %p79, %p78, %p205; @%p79 bra $L__BB21_150; bra.uni $L__BB21_144; $L__BB21_84: fma.rn.ftz.f32 %f2474, %f2474, 0f40000000, 0fBF800000; st.local.f32 [%rd4], %f2473; fma.rn.ftz.f32 %f1380, %f12, %f2475, %f297; st.local.f32 [%rd4+4], %f1380; st.local.f32 [%rd4+8], %f2474; mov.pred %p207, 0; mov.f32 %f2466, %f2473; $L__BB21_86: mov.pred %p209, 0; @%p207 bra $L__BB21_132; st.local.u32 [%rd4+12], %r49; st.local.u32 [%rd4+16], %r48; st.local.u32 [%rd4+20], %r66; mov.pred %p209, -1; mov.u32 %r276, %r49; $L__BB21_132: shr.u32 %r210, %r115, 8; and.b32 %r211, %r210, 1; setp.eq.b32 %p176, %r211, 1; and.pred %p177, %p176, %p209; @%p177 bra $L__BB21_150; bra.uni $L__BB21_144; $L__BB21_14: mul.wide.s32 %rd76, %r251, 4; xor.b64 %rd77, %rd76, -4; add.s64 %rd78, %rd8, %rd77; ld.global.u32 %r280, [%rd78]; mul.wide.s32 %rd79, %r280, 208; add.s64 %rd80, %rd9, %rd79; ld.global.v4.f32 {%f825, %f826, %f827, %f828}, [%rd80+48]; ld.global.v4.f32 {%f833, %f834, %f835, %f836}, [%rd80+64]; ld.global.v4.f32 {%f841, %f842, %f843, %f844}, [%rd80+80]; ld.local.f32 %f849, [%rd3]; ld.local.f32 %f850, [%rd3+4]; mul.ftz.f32 %f851, %f826, %f850; fma.rn.ftz.f32 %f852, %f849, %f825, %f851; ld.local.f32 %f853, [%rd3+8]; fma.rn.ftz.f32 %f854, %f827, %f853, %f852; add.ftz.f32 %f2490, %f828, %f854; mul.ftz.f32 %f855, %f834, %f850; fma.rn.ftz.f32 %f856, %f849, %f833, %f855; fma.rn.ftz.f32 %f857, %f835, %f853, %f856; add.ftz.f32 %f2489, %f836, %f857; mul.ftz.f32 %f858, %f842, %f850; fma.rn.ftz.f32 %f859, %f849, %f841, %f858; fma.rn.ftz.f32 %f860, %f843, %f853, %f859; add.ftz.f32 %f2488, %f844, %f860; ld.local.f32 %f861, [%rd3+12]; ld.local.f32 %f862, [%rd3+16]; mul.ftz.f32 %f863, %f826, %f862; fma.rn.ftz.f32 %f864, %f825, %f861, %f863; ld.local.f32 %f865, [%rd3+20]; fma.rn.ftz.f32 %f866, %f827, %f865, %f864; mul.ftz.f32 %f867, %f834, %f862; fma.rn.ftz.f32 %f868, %f833, %f861, %f867; fma.rn.ftz.f32 %f869, %f835, %f865, %f868; mul.ftz.f32 %f870, %f842, %f862; fma.rn.ftz.f32 %f871, %f841, %f861, %f870; fma.rn.ftz.f32 %f872, %f843, %f865, %f871; abs.ftz.f32 %f873, %f866; setp.gt.ftz.f32 %p37, %f873, 0f17800000; mov.b32 %r145, %f866; and.b32 %r146, %r145, -2147483648; or.b32 %r147, %r146, 394264576; mov.b32 %f874, %r147; selp.f32 %f2493, %f866, %f874, %p37; abs.ftz.f32 %f875, %f869; setp.gt.ftz.f32 %p38, %f875, 0f17800000; mov.b32 %r148, %f869; and.b32 %r149, %r148, -2147483648; or.b32 %r150, %r149, 394264576; mov.b32 %f876, %r150; selp.f32 %f2492, %f869, %f876, %p38; abs.ftz.f32 %f877, %f872; setp.gt.ftz.f32 %p39, %f877, 0f17800000; mov.b32 %r151, %f872; and.b32 %r152, %r151, -2147483648; or.b32 %r153, %r152, 394264576; mov.b32 %f878, %r153; selp.f32 %f2491, %f872, %f878, %p39; rcp.approx.ftz.f32 %f2496, %f2493; rcp.approx.ftz.f32 %f2495, %f2492; rcp.approx.ftz.f32 %f2494, %f2491; add.s32 %r246, %r32, 1; mov.u32 %r154, 1985229328; st.local.u32 [%rd24+4], %r154; mul.wide.s32 %rd82, %r280, 4; add.s64 %rd83, %rd81, %rd82; ld.global.u32 %r247, [%rd83]; bra.uni $L__BB21_145; $L__BB21_1: setp.eq.s32 %p13, %r247, 1985229328; setp.lt.s32 %p14, %r247, 0; or.pred %p15, %p14, %p13; @%p15 bra $L__BB21_12; $L__BB21_2: mul.wide.s32 %rd68, %r247, 16; add.s64 %rd22, %rd7, %rd68; ld.global.v4.f32 {%f612, %f613, %f614, %f615}, [%rd22]; mov.b32 %r247, %f615; mov.b32 %r17, %f614; mov.b32 %r16, %f613; mov.b32 %r15, %f612; mov.b64 %rd69, {%r15, %r16}; cvt.u32.u64 %r131, %rd69; and.b32 %r132, %r131, 1024; setp.eq.s32 %p16, %r132, 0; and.b32 %r19, %r15, %r115; @%p16 bra $L__BB21_5; bra.uni $L__BB21_3; $L__BB21_5: setp.ne.s32 %p22, %r19, 0; ld.global.v4.f32 {%f755, %f756, %f757, %f758}, [%rd22+16]; ld.global.v4.f32 {%f763, %f764, %f765, %f766}, [%rd22+32]; ld.global.v4.f32 {%f771, %f772, %f773, %f774}, [%rd22+48]; sub.ftz.f32 %f779, %f755, %f2490; mul.ftz.f32 %f780, %f2496, %f779; sub.ftz.f32 %f781, %f757, %f2490; mul.ftz.f32 %f782, %f2496, %f781; sub.ftz.f32 %f783, %f763, %f2489; mul.ftz.f32 %f784, %f2495, %f783; sub.ftz.f32 %f785, %f765, %f2489; mul.ftz.f32 %f786, %f2495, %f785; sub.ftz.f32 %f787, %f771, %f2488; mul.ftz.f32 %f788, %f2494, %f787; sub.ftz.f32 %f789, %f773, %f2488; mul.ftz.f32 %f790, %f2494, %f789; min.ftz.f32 %f791, %f780, %f782; min.ftz.f32 %f792, %f784, %f786; min.ftz.f32 %f793, %f788, %f790; max.ftz.f32 %f794, %f10, %f791; max.ftz.f32 %f795, %f792, %f793; max.ftz.f32 %f2310, %f794, %f795; max.ftz.f32 %f796, %f780, %f782; max.ftz.f32 %f797, %f784, %f786; max.ftz.f32 %f798, %f788, %f790; min.ftz.f32 %f799, %f2466, %f796; min.ftz.f32 %f800, %f797, %f798; min.ftz.f32 %f801, %f799, %f800; sub.ftz.f32 %f802, %f756, %f2490; mul.ftz.f32 %f803, %f2496, %f802; sub.ftz.f32 %f804, %f758, %f2490; mul.ftz.f32 %f805, %f2496, %f804; sub.ftz.f32 %f806, %f764, %f2489; mul.ftz.f32 %f807, %f2495, %f806; sub.ftz.f32 %f808, %f766, %f2489; mul.ftz.f32 %f809, %f2495, %f808; sub.ftz.f32 %f810, %f772, %f2488; mul.ftz.f32 %f811, %f2494, %f810; sub.ftz.f32 %f812, %f774, %f2488; mul.ftz.f32 %f813, %f2494, %f812; min.ftz.f32 %f814, %f803, %f805; min.ftz.f32 %f815, %f807, %f809; min.ftz.f32 %f816, %f811, %f813; max.ftz.f32 %f817, %f10, %f814; max.ftz.f32 %f818, %f815, %f816; max.ftz.f32 %f2309, %f817, %f818; max.ftz.f32 %f819, %f803, %f805; max.ftz.f32 %f820, %f807, %f809; max.ftz.f32 %f821, %f811, %f813; min.ftz.f32 %f822, %f2466, %f819; min.ftz.f32 %f823, %f820, %f821; min.ftz.f32 %f824, %f822, %f823; setp.ge.ftz.f32 %p23, %f801, %f2310; and.pred %p24, %p22, %p23; selp.u32 %r136, 1, 0, %p24; setp.ltu.ftz.f32 %p25, %f824, %f2309; and.b32 %r138, %r16, %r115; setp.eq.s32 %p26, %r138, 0; or.pred %p27, %p26, %p25; selp.b32 %r139, 0, 2, %p27; or.b32 %r245, %r139, %r136; bra.uni $L__BB21_6; $L__BB21_3: setp.ne.s32 %p17, %r19, 0; ld.global.v4.f32 {%f619, %f620, %f621, %f622}, [%rd22+16]; ld.global.v4.f32 {%f627, %f628, %f629, %f630}, [%rd22+32]; ld.global.v4.f32 {%f635, %f636, %f637, %f638}, [%rd22+48]; mul.ftz.f32 %f643, %f2492, %f620; fma.rn.ftz.f32 %f644, %f2493, %f619, %f643; fma.rn.ftz.f32 %f645, %f2491, %f621, %f644; mul.ftz.f32 %f646, %f2492, %f628; fma.rn.ftz.f32 %f647, %f2493, %f627, %f646; fma.rn.ftz.f32 %f648, %f2491, %f629, %f647; mul.ftz.f32 %f649, %f2492, %f636; fma.rn.ftz.f32 %f650, %f2493, %f635, %f649; fma.rn.ftz.f32 %f651, %f2491, %f637, %f650; mul.ftz.f32 %f652, %f2489, %f620; fma.rn.ftz.f32 %f653, %f2490, %f619, %f652; fma.rn.ftz.f32 %f654, %f2488, %f621, %f653; add.ftz.f32 %f655, %f622, %f654; mul.ftz.f32 %f656, %f2489, %f628; fma.rn.ftz.f32 %f657, %f2490, %f627, %f656; fma.rn.ftz.f32 %f658, %f2488, %f629, %f657; add.ftz.f32 %f659, %f630, %f658; mul.ftz.f32 %f660, %f2489, %f636; fma.rn.ftz.f32 %f661, %f2490, %f635, %f660; fma.rn.ftz.f32 %f662, %f2488, %f637, %f661; add.ftz.f32 %f663, %f638, %f662; rcp.approx.ftz.f32 %f664, %f645; rcp.approx.ftz.f32 %f665, %f648; rcp.approx.ftz.f32 %f666, %f651; mul.ftz.f32 %f667, %f664, %f655; neg.ftz.f32 %f668, %f667; mul.ftz.f32 %f669, %f665, %f659; neg.ftz.f32 %f670, %f669; mul.ftz.f32 %f671, %f666, %f663; neg.ftz.f32 %f672, %f671; sub.ftz.f32 %f673, %f664, %f667; sub.ftz.f32 %f674, %f665, %f669; sub.ftz.f32 %f675, %f666, %f671; min.ftz.f32 %f676, %f668, %f673; min.ftz.f32 %f677, %f670, %f674; min.ftz.f32 %f678, %f672, %f675; max.ftz.f32 %f679, %f668, %f673; max.ftz.f32 %f680, %f670, %f674; max.ftz.f32 %f681, %f672, %f675; max.ftz.f32 %f682, %f10, %f676; max.ftz.f32 %f683, %f677, %f678; max.ftz.f32 %f2310, %f682, %f683; min.ftz.f32 %f684, %f2466, %f679; min.ftz.f32 %f685, %f680, %f681; min.ftz.f32 %f686, %f684, %f685; setp.le.ftz.f32 %p18, %f2310, %f686; and.pred %p19, %p17, %p18; selp.u32 %r245, 1, 0, %p19; ld.global.v4.f32 {%f687, %f688, %f689, %f690}, [%rd22+64]; ld.global.v4.f32 {%f695, %f696, %f697, %f698}, [%rd22+80]; ld.global.v4.f32 {%f703, %f704, %f705, %f706}, [%rd22+96]; mul.ftz.f32 %f711, %f2492, %f688; fma.rn.ftz.f32 %f712, %f2493, %f687, %f711; fma.rn.ftz.f32 %f713, %f2491, %f689, %f712; mul.ftz.f32 %f714, %f2492, %f696; fma.rn.ftz.f32 %f715, %f2493, %f695, %f714; fma.rn.ftz.f32 %f716, %f2491, %f697, %f715; mul.ftz.f32 %f717, %f2492, %f704; fma.rn.ftz.f32 %f718, %f2493, %f703, %f717; fma.rn.ftz.f32 %f719, %f2491, %f705, %f718; mul.ftz.f32 %f720, %f2489, %f688; fma.rn.ftz.f32 %f721, %f2490, %f687, %f720; fma.rn.ftz.f32 %f722, %f2488, %f689, %f721; add.ftz.f32 %f723, %f690, %f722; mul.ftz.f32 %f724, %f2489, %f696; fma.rn.ftz.f32 %f725, %f2490, %f695, %f724; fma.rn.ftz.f32 %f726, %f2488, %f697, %f725; add.ftz.f32 %f727, %f698, %f726; mul.ftz.f32 %f728, %f2489, %f704; fma.rn.ftz.f32 %f729, %f2490, %f703, %f728; fma.rn.ftz.f32 %f730, %f2488, %f705, %f729; add.ftz.f32 %f731, %f706, %f730; rcp.approx.ftz.f32 %f732, %f713; rcp.approx.ftz.f32 %f733, %f716; rcp.approx.ftz.f32 %f734, %f719; mul.ftz.f32 %f735, %f732, %f723; neg.ftz.f32 %f736, %f735; mul.ftz.f32 %f737, %f733, %f727; neg.ftz.f32 %f738, %f737; mul.ftz.f32 %f739, %f734, %f731; neg.ftz.f32 %f740, %f739; sub.ftz.f32 %f741, %f732, %f735; sub.ftz.f32 %f742, %f733, %f737; sub.ftz.f32 %f743, %f734, %f739; min.ftz.f32 %f744, %f736, %f741; min.ftz.f32 %f745, %f738, %f742; min.ftz.f32 %f746, %f740, %f743; max.ftz.f32 %f747, %f736, %f741; max.ftz.f32 %f748, %f738, %f742; max.ftz.f32 %f749, %f740, %f743; max.ftz.f32 %f750, %f10, %f744; max.ftz.f32 %f751, %f745, %f746; max.ftz.f32 %f2309, %f750, %f751; min.ftz.f32 %f752, %f2466, %f747; min.ftz.f32 %f753, %f748, %f749; min.ftz.f32 %f754, %f752, %f753; setp.gtu.ftz.f32 %p20, %f2309, %f754; @%p20 bra $L__BB21_6; and.b32 %r134, %r16, %r115; setp.eq.s32 %p21, %r134, 0; or.b32 %r135, %r245, 2; selp.b32 %r245, %r245, %r135, %p21; $L__BB21_6: mul.wide.s32 %rd70, %r246, 4; add.s64 %rd23, %rd2, %rd70; setp.eq.s32 %p28, %r245, 3; @%p28 bra $L__BB21_10; bra.uni $L__BB21_7; $L__BB21_10: setp.lt.ftz.f32 %p31, %f2309, %f2310; selp.b32 %r28, %r247, %r17, %p31; selp.b32 %r140, %r17, %r247, %p31; add.s32 %r246, %r246, 1; st.local.u32 [%rd23+4], %r140; mov.u32 %r247, %r28; bra.uni $L__BB21_11; $L__BB21_7: setp.eq.s32 %p29, %r245, 2; @%p29 bra $L__BB21_11; setp.ne.s32 %p30, %r245, 0; mov.u32 %r247, %r17; @%p30 bra $L__BB21_11; ld.local.u32 %r247, [%rd23]; add.s32 %r246, %r246, -1; $L__BB21_11: setp.ne.s32 %p32, %r247, 1985229328; setp.gt.s32 %p33, %r247, -1; and.pred %p34, %p33, %p32; @%p34 bra $L__BB21_2; $L__BB21_12: mov.u32 %r32, %r246; setp.gt.s32 %p35, %r247, -1; @%p35 bra $L__BB21_145; bra.uni $L__BB21_13; $L__BB21_145: setp.ne.s32 %p197, %r247, 1985229328; @%p197 bra $L__BB21_1; mov.u32 %r247, 1985229328; setp.lt.s32 %p198, %r246, 0; @%p198 bra $L__BB21_148; ld.local.f32 %f2490, [%rd3]; ld.local.f32 %f2489, [%rd3+4]; ld.local.f32 %f2488, [%rd3+8]; ld.local.f32 %f2235, [%rd3+12]; abs.ftz.f32 %f2236, %f2235; setp.gt.ftz.f32 %p199, %f2236, 0f17800000; mov.b32 %r223, %f2235; and.b32 %r224, %r223, -2147483648; or.b32 %r225, %r224, 394264576; mov.b32 %f2237, %r225; selp.f32 %f2493, %f2235, %f2237, %p199; ld.local.f32 %f2238, [%rd3+16]; abs.ftz.f32 %f2239, %f2238; setp.gt.ftz.f32 %p200, %f2239, 0f17800000; mov.b32 %r226, %f2238; and.b32 %r227, %r226, -2147483648; or.b32 %r228, %r227, 394264576; mov.b32 %f2240, %r228; selp.f32 %f2492, %f2238, %f2240, %p200; ld.local.f32 %f2241, [%rd3+20]; abs.ftz.f32 %f2242, %f2241; setp.gt.ftz.f32 %p201, %f2242, 0f17800000; mov.b32 %r229, %f2241; and.b32 %r230, %r229, -2147483648; or.b32 %r231, %r230, 394264576; mov.b32 %f2243, %r231; selp.f32 %f2491, %f2241, %f2243, %p201; rcp.approx.ftz.f32 %f2496, %f2493; rcp.approx.ftz.f32 %f2495, %f2492; rcp.approx.ftz.f32 %f2494, %f2491; mul.wide.s32 %rd142, %r246, 4; add.s64 %rd143, %rd2, %rd142; ld.local.u32 %r247, [%rd143]; add.s32 %r246, %r246, -1; mov.u32 %r280, -1; $L__BB21_148: setp.ne.s32 %p202, %r247, 1985229328; @%p202 bra $L__BB21_1; setp.ne.s32 %p203, %r276, -1; selp.u16 %rs70, 1, 0, %p203; $L__BB21_150: cvt.u32.u16 %r232, %rs70; st.param.b32 [func_retval0+0], %r232; ret; } .func (.param .b32 func_retval0) _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj( .param .b64 _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0, .param .b64 _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1, .param .b32 _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2 ) { .local .align 16 .b8 __local_depot22[768]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<136>; .reg .b16 %rs<40>; .reg .f32 %f<1042>; .reg .b32 %r<242>; .reg .b64 %rd<221>; mov.u64 %SPL, __local_depot22; ld.param.u64 %rd100, [_Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0]; ld.param.u64 %rd101, [_Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1]; ld.param.u32 %r90, [_Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2]; cvta.to.local.u64 %rd3, %rd101; add.u64 %rd1, %SPL, 0; mov.u32 %r94, 1985229328; st.local.u32 [%rd1], %r94; ld.const.u32 %r215, [kernel_params+1104]; cvta.to.local.u64 %rd2, %rd100; ld.local.f32 %f1026, [%rd2]; ld.local.f32 %f1025, [%rd2+4]; ld.local.f32 %f1024, [%rd2+8]; ld.local.f32 %f299, [%rd2+12]; abs.ftz.f32 %f300, %f299; setp.gt.ftz.f32 %p7, %f300, 0f17800000; mov.b32 %r95, %f299; and.b32 %r96, %r95, -2147483648; or.b32 %r97, %r96, 394264576; mov.b32 %f301, %r97; selp.f32 %f1029, %f299, %f301, %p7; ld.local.f32 %f302, [%rd2+16]; abs.ftz.f32 %f303, %f302; setp.gt.ftz.f32 %p8, %f303, 0f17800000; mov.b32 %r98, %f302; and.b32 %r99, %r98, -2147483648; or.b32 %r100, %r99, 394264576; mov.b32 %f304, %r100; selp.f32 %f1028, %f302, %f304, %p8; ld.local.f32 %f305, [%rd2+20]; abs.ftz.f32 %f306, %f305; setp.gt.ftz.f32 %p9, %f306, 0f17800000; mov.b32 %r101, %f305; and.b32 %r102, %r101, -2147483648; or.b32 %r103, %r102, 394264576; mov.b32 %f307, %r103; selp.f32 %f1027, %f305, %f307, %p9; rcp.approx.ftz.f32 %f1032, %f1029; rcp.approx.ftz.f32 %f1031, %f1028; rcp.approx.ftz.f32 %f1030, %f1027; ld.local.f32 %f10, [%rd2+24]; ld.local.f32 %f1004, [%rd2+28]; st.local.f32 [%rd3], %f1004; mov.u32 %r214, 0; st.local.u32 [%rd3+4], %r214; st.local.u32 [%rd3+8], %r214; mov.u32 %r234, -1; st.local.u32 [%rd3+12], %r234; st.local.u32 [%rd3+16], %r234; ld.const.v2.u64 {%rd103, %rd104}, [kernel_params+2096]; cvta.to.global.u64 %rd6, %rd103; ld.const.v2.u64 {%rd105, %rd106}, [kernel_params+2128]; cvta.to.global.u64 %rd7, %rd106; ld.const.v2.u64 {%rd109, %rd110}, [kernel_params+2176]; ld.const.u64 %rd111, [kernel_params+2160]; cvta.to.global.u64 %rd10, %rd111; cvta.to.global.u64 %rd11, %rd105; ld.const.u64 %rd112, [kernel_params+2232]; cvta.to.global.u64 %rd12, %rd112; ld.const.u64 %rd113, [kernel_params+2256]; cvta.to.global.u64 %rd13, %rd113; ld.const.u64 %rd114, [kernel_params+2312]; cvta.to.global.u64 %rd14, %rd114; ld.const.v2.u64 {%rd115, %rd116}, [kernel_params+2112]; cvta.to.global.u64 %rd15, %rd115; ld.const.v2.u64 {%rd119, %rd120}, [kernel_params+2144]; cvta.to.global.u64 %rd18, %rd120; and.b32 %r3, %r90, 256; cvta.to.global.u64 %rd19, %rd116; ld.const.u64 %rd121, [kernel_params+2296]; cvta.to.global.u64 %rd20, %rd121; ld.const.v2.u64 {%rd122, %rd123}, [kernel_params+2336]; cvta.to.global.u64 %rd21, %rd122; cvta.to.global.u64 %rd22, %rd123; cvta.to.global.u64 %rd129, %rd104; mov.u16 %rs39, 1; mov.pred %p100, -1; mov.pred %p112, 0; cvta.to.global.u64 %rd137, %rd110; cvta.to.global.u64 %rd147, %rd119; cvta.to.global.u64 %rd145, %rd109; mov.u32 %r238, %r234; bra.uni $L__BB22_1; $L__BB22_9: mul.wide.s32 %rd130, %r215, 16; xor.b64 %rd131, %rd130, -16; add.s64 %rd132, %rd129, %rd131; ld.global.v4.u32 {%r219, %r116, %r117, %r118}, [%rd132]; setp.gt.s32 %p27, %r219, -1; mul.wide.s32 %rd133, %r23, 4; add.s64 %rd24, %rd1, %rd133; @%p27 bra $L__BB22_19; bra.uni $L__BB22_10; $L__BB22_19: ld.local.u32 %r215, [%rd24]; add.s32 %r214, %r23, -1; setp.le.s32 %p37, %r116, %r219; @%p37 bra $L__BB22_87; mul.ftz.f32 %f654, %f1029, %f1029; fma.rn.ftz.f32 %f655, %f1028, %f1028, %f654; fma.rn.ftz.f32 %f656, %f1027, %f1027, %f655; rcp.approx.ftz.f32 %f153, %f656; $L__BB22_21: cvt.s64.s32 %rd28, %r219; setp.ne.s32 %p40, %r238, -1; mov.u32 %r39, %r238; @%p40 bra $L__BB22_23; shl.b64 %rd150, %rd28, 2; add.s64 %rd151, %rd7, %rd150; ld.global.u32 %r39, [%rd151]; $L__BB22_23: shl.b64 %rd152, %rd28, 2; add.s64 %rd153, %rd11, %rd152; ld.global.u32 %r40, [%rd153]; cvt.s64.s32 %rd29, %r40; ld.local.u32 %r135, [%rd2+36]; setp.ne.s32 %p41, %r135, %r40; @%p41 bra $L__BB22_25; ld.local.u32 %r136, [%rd2+40]; setp.eq.s32 %p42, %r136, %r39; mov.u16 %rs34, %rs39; @%p42 bra $L__BB22_27; $L__BB22_25: ld.local.u32 %r137, [%rd2+44]; setp.ne.s32 %p43, %r137, %r40; mov.u16 %rs34, 0; @%p43 bra $L__BB22_27; ld.local.u32 %r138, [%rd2+48]; setp.eq.s32 %p44, %r138, %r39; selp.u16 %rs34, 1, 0, %p44; $L__BB22_27: setp.ne.s16 %p45, %rs34, 0; @%p45 bra $L__BB22_86; bra.uni $L__BB22_28; $L__BB22_86: add.s32 %r214, %r23, -1; cvt.u32.u64 %r186, %rd28; add.s32 %r219, %r186, 1; setp.lt.s32 %p118, %r219, %r116; @%p118 bra $L__BB22_21; bra.uni $L__BB22_87; $L__BB22_28: cvt.u16.u32 %rs25, %r118; and.b16 %rs24, %rs25, 127; mul.wide.s32 %rd154, %r39, 208; add.s64 %rd155, %rd10, %rd154; add.s64 %rd31, %rd155, 148; shl.b64 %rd156, %rd29, 4; add.s64 %rd32, %rd12, %rd156; add.s64 %rd33, %rd19, %rd152; setp.gt.s16 %p46, %rs24, 64; @%p46 bra $L__BB22_37; cvt.u16.u32 %rs31, %r118; and.b16 %rs30, %rs31, 127; setp.eq.s16 %p49, %rs30, 1; @%p49 bra $L__BB22_76; cvt.u16.u32 %rs33, %r118; and.b16 %rs32, %rs33, 127; setp.eq.s16 %p50, %rs32, 8; @%p50 bra $L__BB22_31; bra.uni $L__BB22_86; $L__BB22_37: cvt.u16.u32 %rs27, %r118; and.b16 %rs26, %rs27, 127; setp.eq.s16 %p47, %rs26, 72; @%p47 bra $L__BB22_31; cvt.u16.u32 %rs29, %r118; and.b16 %rs28, %rs29, 127; setp.ne.s16 %p48, %rs28, 65; @%p48 bra $L__BB22_86; ld.global.u32 %r41, [%rd31+8]; ld.global.u32 %r42, [%rd31+4]; shl.b32 %r139, %r42, 1; cvt.rn.f32.s32 %f657, %r139; ld.local.f32 %f658, [%rd2+32]; mul.ftz.f32 %f164, %f658, %f657; cvt.rzi.ftz.s32.f32 %r140, %f164; add.s32 %r141, %r139, -1; min.s32 %r43, %r140, %r141; ld.global.u32 %r222, [%rd31+16]; mul.wide.u32 %rd158, %r222, 16; add.s64 %rd159, %rd14, %rd158; ld.global.u16 %rs35, [%rd159+12]; ld.global.u32 %r221, [%rd159+8]; ld.global.u64 %rd199, [%rd159]; setp.eq.s64 %p51, %rd199, 11; @%p51 bra $L__BB22_44; $L__BB22_40: setp.eq.s64 %p52, %rd199, 0; @%p52 bra $L__BB22_42; add.s32 %r222, %r222, 2; bra.uni $L__BB22_43; $L__BB22_42: setp.eq.s16 %p53, %rs35, 0; mov.u32 %r225, -1; mov.u32 %r222, %r221; @%p53 bra $L__BB22_45; $L__BB22_43: mul.wide.u32 %rd160, %r222, 16; add.s64 %rd161, %rd14, %rd160; ld.global.u16 %rs35, [%rd161+12]; ld.global.u32 %r221, [%rd161+8]; ld.global.u64 %rd199, [%rd161]; setp.ne.s64 %p54, %rd199, 11; @%p54 bra $L__BB22_40; $L__BB22_44: setp.eq.s16 %p55, %rs35, 0; selp.b32 %r225, -1, %r221, %p55; $L__BB22_45: ld.global.v4.u32 {%r143, %r144, %r145, %r146}, [%rd32]; setp.eq.s32 %p56, %r43, %r42; @%p56 bra $L__BB22_47; bra.uni $L__BB22_46; $L__BB22_47: mul.wide.u32 %rd165, %r146, 12; add.s64 %rd200, %rd13, %rd165; add.s32 %r153, %r146, 1; mul.wide.u32 %rd166, %r153, 12; add.s64 %rd203, %rd13, %rd166; add.s32 %r154, %r146, 2; mul.wide.u32 %rd167, %r154, 12; add.s64 %rd206, %rd13, %rd167; add.s64 %rd201, %rd200, 4; add.s64 %rd202, %rd200, 8; add.s64 %rd204, %rd203, 4; add.s64 %rd205, %rd203, 8; add.s64 %rd207, %rd206, 4; add.s64 %rd208, %rd206, 8; bra.uni $L__BB22_48; $L__BB22_31: and.b32 %r202, %r118, 64; ld.const.u32 %r201, [kernel_params+1128]; setp.eq.s32 %p130, %r201, 0; setp.eq.s32 %p129, %r202, 0; or.pred %p128, %p129, %p130; @%p128 bra $L__BB22_33; shl.b64 %rd174, %rd28, 3; add.s64 %rd175, %rd18, %rd174; ld.global.v2.f32 {%f767, %f768}, [%rd175]; ld.local.f32 %f771, [%rd2+32]; setp.lt.ftz.f32 %p78, %f771, %f767; setp.gt.ftz.f32 %p79, %f771, %f768; or.pred %p80, %p78, %p79; @%p80 bra $L__BB22_86; $L__BB22_33: add.s64 %rd177, %rd15, %rd152; ld.global.u32 %r61, [%rd177]; and.b32 %r166, %r61, 64; setp.eq.s32 %p81, %r166, 0; @%p81 bra $L__BB22_71; bra.uni $L__BB22_34; $L__BB22_71: add.s64 %rd187, %rd20, %rd156; ld.global.v4.f32 {%f993, %f994, %f995, %f996}, [%rd187]; bra.uni $L__BB22_72; $L__BB22_76: ld.global.u32 %r180, [%rd32+12]; mul.wide.u32 %rd188, %r180, 12; add.s64 %rd189, %rd13, %rd188; add.s32 %r181, %r180, 1; mul.wide.u32 %rd190, %r181, 12; add.s64 %rd191, %rd13, %rd190; add.s32 %r182, %r180, 2; mul.wide.u32 %rd192, %r182, 12; add.s64 %rd193, %rd13, %rd192; ld.global.f32 %f815, [%rd189]; sub.ftz.f32 %f225, %f815, %f1026; ld.global.f32 %f816, [%rd189+4]; sub.ftz.f32 %f226, %f816, %f1025; ld.global.f32 %f817, [%rd189+8]; sub.ftz.f32 %f227, %f817, %f1024; ld.global.f32 %f818, [%rd191]; sub.ftz.f32 %f819, %f818, %f1026; ld.global.f32 %f820, [%rd191+4]; sub.ftz.f32 %f821, %f820, %f1025; ld.global.f32 %f822, [%rd191+8]; sub.ftz.f32 %f823, %f822, %f1024; ld.global.f32 %f824, [%rd193]; sub.ftz.f32 %f825, %f824, %f1026; ld.global.f32 %f826, [%rd193+4]; sub.ftz.f32 %f827, %f826, %f1025; ld.global.f32 %f828, [%rd193+8]; sub.ftz.f32 %f829, %f828, %f1024; sub.ftz.f32 %f228, %f825, %f225; sub.ftz.f32 %f229, %f827, %f226; sub.ftz.f32 %f230, %f829, %f227; sub.ftz.f32 %f231, %f225, %f819; sub.ftz.f32 %f232, %f226, %f821; sub.ftz.f32 %f233, %f227, %f823; sub.ftz.f32 %f830, %f819, %f825; sub.ftz.f32 %f831, %f821, %f827; sub.ftz.f32 %f832, %f823, %f829; add.ftz.f32 %f833, %f225, %f825; add.ftz.f32 %f834, %f226, %f827; add.ftz.f32 %f835, %f227, %f829; mul.ftz.f32 %f836, %f229, %f835; mul.ftz.f32 %f837, %f834, %f230; sub.ftz.f32 %f838, %f836, %f837; mul.ftz.f32 %f839, %f833, %f230; mul.ftz.f32 %f840, %f228, %f835; sub.ftz.f32 %f841, %f839, %f840; mul.ftz.f32 %f842, %f228, %f834; mul.ftz.f32 %f843, %f833, %f229; sub.ftz.f32 %f844, %f842, %f843; mul.ftz.f32 %f845, %f1028, %f841; fma.rn.ftz.f32 %f846, %f1029, %f838, %f845; fma.rn.ftz.f32 %f234, %f1027, %f844, %f846; add.ftz.f32 %f847, %f225, %f819; add.ftz.f32 %f848, %f226, %f821; add.ftz.f32 %f849, %f227, %f823; mul.ftz.f32 %f850, %f232, %f849; mul.ftz.f32 %f851, %f848, %f233; sub.ftz.f32 %f852, %f850, %f851; mul.ftz.f32 %f853, %f847, %f233; mul.ftz.f32 %f854, %f231, %f849; sub.ftz.f32 %f855, %f853, %f854; mul.ftz.f32 %f856, %f231, %f848; mul.ftz.f32 %f857, %f847, %f232; sub.ftz.f32 %f858, %f856, %f857; mul.ftz.f32 %f859, %f1028, %f855; fma.rn.ftz.f32 %f860, %f1029, %f852, %f859; fma.rn.ftz.f32 %f235, %f1027, %f858, %f860; add.ftz.f32 %f861, %f819, %f825; add.ftz.f32 %f862, %f821, %f827; add.ftz.f32 %f863, %f823, %f829; mul.ftz.f32 %f864, %f831, %f863; mul.ftz.f32 %f865, %f862, %f832; sub.ftz.f32 %f866, %f864, %f865; mul.ftz.f32 %f867, %f861, %f832; mul.ftz.f32 %f868, %f830, %f863; sub.ftz.f32 %f869, %f867, %f868; mul.ftz.f32 %f870, %f830, %f862; mul.ftz.f32 %f871, %f861, %f831; sub.ftz.f32 %f872, %f870, %f871; mul.ftz.f32 %f873, %f1028, %f869; fma.rn.ftz.f32 %f874, %f1029, %f866, %f873; fma.rn.ftz.f32 %f875, %f1027, %f872, %f874; add.ftz.f32 %f876, %f235, %f234; add.ftz.f32 %f236, %f875, %f876; abs.ftz.f32 %f237, %f236; mul.ftz.f32 %f877, %f237, 0f34000000; min.ftz.f32 %f878, %f235, %f875; min.ftz.f32 %f879, %f234, %f878; max.ftz.f32 %f880, %f235, %f875; max.ftz.f32 %f881, %f234, %f880; mul.ftz.f32 %f882, %f237, 0fB4000000; setp.ltu.ftz.f32 %p101, %f879, %f882; setp.gtu.ftz.f32 %p102, %f881, %f877; and.pred %p103, %p101, %p102; mov.pred %p134, %p100; @%p103 bra $L__BB22_82; mul.ftz.f32 %f883, %f232, %f230; mul.ftz.f32 %f884, %f233, %f229; sub.ftz.f32 %f885, %f883, %f884; mul.ftz.f32 %f886, %f231, %f230; mul.ftz.f32 %f887, %f233, %f228; sub.ftz.f32 %f888, %f887, %f886; mul.ftz.f32 %f889, %f232, %f228; mul.ftz.f32 %f890, %f231, %f229; sub.ftz.f32 %f891, %f890, %f889; add.ftz.f32 %f238, %f885, %f885; add.ftz.f32 %f239, %f888, %f888; add.ftz.f32 %f240, %f891, %f891; mul.ftz.f32 %f892, %f1028, %f239; fma.rn.ftz.f32 %f893, %f1029, %f238, %f892; fma.rn.ftz.f32 %f241, %f1027, %f240, %f893; setp.eq.ftz.f32 %p105, %f241, 0f00000000; mov.pred %p134, %p100; @%p105 bra $L__BB22_82; mul.ftz.f32 %f894, %f226, %f239; fma.rn.ftz.f32 %f895, %f225, %f238, %f894; fma.rn.ftz.f32 %f896, %f227, %f240, %f895; div.approx.ftz.f32 %f242, %f896, %f241; setp.ltu.ftz.f32 %p107, %f242, %f10; setp.gtu.ftz.f32 %p108, %f242, %f1004; or.pred %p109, %p108, %p107; mov.pred %p134, %p100; @%p109 bra $L__BB22_82; abs.ftz.f32 %f912, %f236; setp.lt.ftz.f32 %p110, %f912, 0f219392EF; mov.f32 %f999, 0f00000000; @%p110 bra $L__BB22_81; rcp.approx.ftz.f32 %f999, %f236; $L__BB22_81: mul.ftz.f32 %f898, %f234, %f999; mov.f32 %f899, 0f3F800000; min.ftz.f32 %f1012, %f898, %f899; mul.ftz.f32 %f900, %f235, %f999; min.ftz.f32 %f1013, %f900, %f899; mov.pred %p134, 0; mov.f32 %f1011, %f242; $L__BB22_82: mov.pred %p135, %p112; @%p134 bra $L__BB22_85; ld.global.u32 %r183, [%rd33]; and.b32 %r184, %r183, %r90; setp.eq.s32 %p114, %r184, 0; mov.pred %p135, %p112; @%p114 bra $L__BB22_85; st.local.u32 [%rd3+16], %r39; st.local.u32 [%rd3+12], %r40; mov.u32 %r185, 1; st.local.u32 [%rd3+20], %r185; st.local.f32 [%rd3+4], %f1012; st.local.f32 [%rd3+8], %f1013; st.local.f32 [%rd3], %f1011; mov.pred %p135, -1; mov.u32 %r234, %r40; mov.f32 %f1004, %f1011; $L__BB22_85: setp.ne.s32 %p116, %r3, 0; and.pred %p117, %p116, %p135; @%p117 bra $L__BB22_92; bra.uni $L__BB22_86; $L__BB22_34: ld.global.u32 %r62, [%rd31]; ld.global.u32 %r63, [%rd31+4]; shl.b32 %r167, %r63, 1; cvt.rn.f32.s32 %f772, %r167; ld.local.f32 %f773, [%rd2+32]; mul.ftz.f32 %f202, %f773, %f772; cvt.rzi.ftz.s32.f32 %r168, %f202; add.s32 %r169, %r167, -1; min.s32 %r64, %r168, %r169; ld.global.u32 %r228, [%rd31+16]; mul.wide.u32 %rd178, %r228, 16; add.s64 %rd179, %rd14, %rd178; ld.global.u16 %rs37, [%rd179+12]; ld.global.u32 %r227, [%rd179+8]; ld.global.u64 %rd218, [%rd179]; setp.eq.s64 %p82, %rd218, 11; @%p82 bra $L__BB22_63; $L__BB22_35: setp.eq.s64 %p83, %rd218, 0; @%p83 bra $L__BB22_61; add.s32 %r228, %r228, 2; bra.uni $L__BB22_62; $L__BB22_61: setp.eq.s16 %p84, %rs37, 0; mov.u32 %r231, -1; mov.u32 %r228, %r227; @%p84 bra $L__BB22_64; $L__BB22_62: mul.wide.u32 %rd180, %r228, 16; add.s64 %rd181, %rd14, %rd180; ld.global.u16 %rs37, [%rd181+12]; ld.global.u32 %r227, [%rd181+8]; ld.global.u64 %rd218, [%rd181]; setp.ne.s64 %p85, %rd218, 11; @%p85 bra $L__BB22_35; $L__BB22_63: setp.eq.s16 %p86, %rs37, 0; selp.b32 %r231, -1, %r227, %p86; $L__BB22_64: setp.eq.s32 %p87, %r64, %r63; @%p87 bra $L__BB22_66; bra.uni $L__BB22_65; $L__BB22_66: add.s64 %rd219, %rd20, %rd156; bra.uni $L__BB22_67; $L__BB22_46: setp.gt.s32 %p57, %r43, %r42; selp.b32 %r147, -1, 0, %p57; add.s32 %r148, %r43, %r147; mad.lo.s32 %r149, %r148, %r41, %r225; add.s32 %r150, %r143, %r149; mul.wide.u32 %rd162, %r150, 12; add.s64 %rd200, %rd21, %rd162; add.s32 %r151, %r144, %r149; mul.wide.u32 %rd163, %r151, 12; add.s64 %rd203, %rd21, %rd163; add.s32 %r152, %r145, %r149; mul.wide.u32 %rd164, %r152, 12; add.s64 %rd206, %rd21, %rd164; add.s64 %rd201, %rd200, 4; add.s64 %rd202, %rd200, 8; add.s64 %rd204, %rd203, 4; add.s64 %rd205, %rd203, 8; add.s64 %rd207, %rd206, 4; add.s64 %rd208, %rd206, 8; $L__BB22_48: ld.global.f32 %f165, [%rd208]; ld.global.f32 %f166, [%rd207]; ld.global.f32 %f167, [%rd206]; ld.global.f32 %f168, [%rd205]; ld.global.f32 %f169, [%rd204]; ld.global.f32 %f170, [%rd203]; ld.global.f32 %f171, [%rd202]; ld.global.f32 %f172, [%rd201]; ld.global.f32 %f173, [%rd200]; add.s32 %r59, %r43, 1; setp.eq.s32 %p58, %r59, %r42; @%p58 bra $L__BB22_50; bra.uni $L__BB22_49; $L__BB22_50: mul.wide.u32 %rd171, %r146, 12; add.s64 %rd209, %rd13, %rd171; add.s32 %r161, %r146, 1; mul.wide.u32 %rd172, %r161, 12; add.s64 %rd212, %rd13, %rd172; add.s32 %r162, %r146, 2; mul.wide.u32 %rd173, %r162, 12; add.s64 %rd215, %rd13, %rd173; add.s64 %rd210, %rd209, 4; add.s64 %rd211, %rd209, 8; add.s64 %rd213, %rd212, 4; add.s64 %rd214, %rd212, 8; add.s64 %rd216, %rd215, 4; add.s64 %rd217, %rd215, 8; bra.uni $L__BB22_51; $L__BB22_49: setp.ge.s32 %p59, %r43, %r42; selp.b32 %r155, -1, 0, %p59; add.s32 %r156, %r59, %r155; mad.lo.s32 %r157, %r156, %r41, %r225; add.s32 %r158, %r143, %r157; mul.wide.u32 %rd168, %r158, 12; add.s64 %rd209, %rd21, %rd168; add.s32 %r159, %r144, %r157; mul.wide.u32 %rd169, %r159, 12; add.s64 %rd212, %rd21, %rd169; add.s32 %r160, %r145, %r157; mul.wide.u32 %rd170, %r160, 12; add.s64 %rd215, %rd21, %rd170; add.s64 %rd210, %rd209, 4; add.s64 %rd211, %rd209, 8; add.s64 %rd213, %rd212, 4; add.s64 %rd214, %rd212, 8; add.s64 %rd216, %rd215, 4; add.s64 %rd217, %rd215, 8; $L__BB22_51: shl.b32 %r203, %r42, 1; cvt.rn.f32.s32 %f914, %r203; mul.ftz.f32 %f913, %f658, %f914; cvt.rn.f32.s32 %f659, %r43; sub.ftz.f32 %f660, %f913, %f659; mov.f32 %f661, 0f3F800000; sub.ftz.f32 %f662, %f661, %f660; ld.global.f32 %f663, [%rd209]; mul.ftz.f32 %f664, %f660, %f663; ld.global.f32 %f665, [%rd210]; mul.ftz.f32 %f666, %f660, %f665; ld.global.f32 %f667, [%rd211]; mul.ftz.f32 %f668, %f660, %f667; fma.rn.ftz.f32 %f669, %f662, %f173, %f664; fma.rn.ftz.f32 %f670, %f662, %f172, %f666; fma.rn.ftz.f32 %f671, %f662, %f171, %f668; ld.global.f32 %f672, [%rd212]; mul.ftz.f32 %f673, %f660, %f672; ld.global.f32 %f674, [%rd213]; mul.ftz.f32 %f675, %f660, %f674; ld.global.f32 %f676, [%rd214]; mul.ftz.f32 %f677, %f660, %f676; fma.rn.ftz.f32 %f678, %f662, %f170, %f673; fma.rn.ftz.f32 %f679, %f662, %f169, %f675; fma.rn.ftz.f32 %f680, %f662, %f168, %f677; ld.global.f32 %f681, [%rd215]; mul.ftz.f32 %f682, %f660, %f681; ld.global.f32 %f683, [%rd216]; mul.ftz.f32 %f684, %f660, %f683; ld.global.f32 %f685, [%rd217]; mul.ftz.f32 %f686, %f660, %f685; fma.rn.ftz.f32 %f687, %f662, %f167, %f682; fma.rn.ftz.f32 %f688, %f662, %f166, %f684; fma.rn.ftz.f32 %f689, %f662, %f165, %f686; sub.ftz.f32 %f174, %f669, %f1026; sub.ftz.f32 %f175, %f670, %f1025; sub.ftz.f32 %f176, %f671, %f1024; sub.ftz.f32 %f690, %f678, %f1026; sub.ftz.f32 %f691, %f679, %f1025; sub.ftz.f32 %f692, %f680, %f1024; sub.ftz.f32 %f693, %f687, %f1026; sub.ftz.f32 %f694, %f688, %f1025; sub.ftz.f32 %f695, %f689, %f1024; sub.ftz.f32 %f177, %f693, %f174; sub.ftz.f32 %f178, %f694, %f175; sub.ftz.f32 %f179, %f695, %f176; sub.ftz.f32 %f180, %f174, %f690; sub.ftz.f32 %f181, %f175, %f691; sub.ftz.f32 %f182, %f176, %f692; sub.ftz.f32 %f696, %f690, %f693; sub.ftz.f32 %f697, %f691, %f694; sub.ftz.f32 %f698, %f692, %f695; add.ftz.f32 %f699, %f693, %f174; add.ftz.f32 %f700, %f694, %f175; add.ftz.f32 %f701, %f695, %f176; mul.ftz.f32 %f702, %f701, %f178; mul.ftz.f32 %f703, %f179, %f700; sub.ftz.f32 %f704, %f702, %f703; mul.ftz.f32 %f705, %f179, %f699; mul.ftz.f32 %f706, %f701, %f177; sub.ftz.f32 %f707, %f705, %f706; mul.ftz.f32 %f708, %f700, %f177; mul.ftz.f32 %f709, %f178, %f699; sub.ftz.f32 %f710, %f708, %f709; mul.ftz.f32 %f711, %f1028, %f707; fma.rn.ftz.f32 %f712, %f1029, %f704, %f711; fma.rn.ftz.f32 %f183, %f1027, %f710, %f712; add.ftz.f32 %f713, %f690, %f174; add.ftz.f32 %f714, %f691, %f175; add.ftz.f32 %f715, %f692, %f176; mul.ftz.f32 %f716, %f715, %f181; mul.ftz.f32 %f717, %f182, %f714; sub.ftz.f32 %f718, %f716, %f717; mul.ftz.f32 %f719, %f182, %f713; mul.ftz.f32 %f720, %f715, %f180; sub.ftz.f32 %f721, %f719, %f720; mul.ftz.f32 %f722, %f714, %f180; mul.ftz.f32 %f723, %f181, %f713; sub.ftz.f32 %f724, %f722, %f723; mul.ftz.f32 %f725, %f1028, %f721; fma.rn.ftz.f32 %f726, %f1029, %f718, %f725; fma.rn.ftz.f32 %f184, %f1027, %f724, %f726; add.ftz.f32 %f727, %f693, %f690; add.ftz.f32 %f728, %f694, %f691; add.ftz.f32 %f729, %f695, %f692; mul.ftz.f32 %f730, %f729, %f697; mul.ftz.f32 %f731, %f698, %f728; sub.ftz.f32 %f732, %f730, %f731; mul.ftz.f32 %f733, %f698, %f727; mul.ftz.f32 %f734, %f729, %f696; sub.ftz.f32 %f735, %f733, %f734; mul.ftz.f32 %f736, %f728, %f696; mul.ftz.f32 %f737, %f697, %f727; sub.ftz.f32 %f738, %f736, %f737; mul.ftz.f32 %f739, %f1028, %f735; fma.rn.ftz.f32 %f740, %f1029, %f732, %f739; fma.rn.ftz.f32 %f741, %f1027, %f738, %f740; add.ftz.f32 %f742, %f183, %f184; add.ftz.f32 %f185, %f741, %f742; abs.ftz.f32 %f186, %f185; mul.ftz.f32 %f743, %f186, 0f34000000; min.ftz.f32 %f744, %f184, %f741; min.ftz.f32 %f745, %f183, %f744; max.ftz.f32 %f746, %f184, %f741; max.ftz.f32 %f747, %f183, %f746; mul.ftz.f32 %f748, %f186, 0fB4000000; setp.ltu.ftz.f32 %p61, %f745, %f748; setp.gtu.ftz.f32 %p62, %f747, %f743; and.pred %p63, %p61, %p62; mov.pred %p131, -1; @%p63 bra $L__BB22_57; mov.pred %p131, -1; mul.ftz.f32 %f749, %f179, %f181; mul.ftz.f32 %f750, %f182, %f178; sub.ftz.f32 %f751, %f749, %f750; mul.ftz.f32 %f752, %f179, %f180; mul.ftz.f32 %f753, %f182, %f177; sub.ftz.f32 %f754, %f753, %f752; mul.ftz.f32 %f755, %f181, %f177; mul.ftz.f32 %f756, %f178, %f180; sub.ftz.f32 %f757, %f756, %f755; add.ftz.f32 %f187, %f751, %f751; add.ftz.f32 %f188, %f754, %f754; add.ftz.f32 %f189, %f757, %f757; mul.ftz.f32 %f758, %f1028, %f188; fma.rn.ftz.f32 %f759, %f1029, %f187, %f758; fma.rn.ftz.f32 %f190, %f1027, %f189, %f759; setp.eq.ftz.f32 %p65, %f190, 0f00000000; @%p65 bra $L__BB22_57; mov.pred %p131, -1; mul.ftz.f32 %f760, %f175, %f188; fma.rn.ftz.f32 %f761, %f174, %f187, %f760; fma.rn.ftz.f32 %f762, %f176, %f189, %f761; div.approx.ftz.f32 %f191, %f762, %f190; setp.ltu.ftz.f32 %p67, %f191, %f10; setp.gtu.ftz.f32 %p68, %f191, %f1004; or.pred %p69, %p68, %p67; @%p69 bra $L__BB22_57; abs.ftz.f32 %f910, %f185; setp.lt.ftz.f32 %p70, %f910, 0f219392EF; mov.f32 %f986, 0f00000000; @%p70 bra $L__BB22_56; rcp.approx.ftz.f32 %f986, %f185; $L__BB22_56: mov.f32 %f911, 0f3F800000; mul.ftz.f32 %f764, %f183, %f986; min.ftz.f32 %f1009, %f764, %f911; mul.ftz.f32 %f766, %f184, %f986; min.ftz.f32 %f1008, %f766, %f911; mov.pred %p131, 0; mov.f32 %f1010, %f191; $L__BB22_57: mov.pred %p132, 0; @%p131 bra $L__BB22_60; ld.global.u32 %r163, [%rd33]; and.b32 %r164, %r163, %r90; setp.eq.s32 %p74, %r164, 0; @%p74 bra $L__BB22_60; st.local.f32 [%rd3], %f1010; st.local.f32 [%rd3+4], %f1009; st.local.f32 [%rd3+8], %f1008; st.local.u32 [%rd3+12], %r40; st.local.u32 [%rd3+16], %r39; mov.u32 %r165, 65; st.local.u32 [%rd3+20], %r165; mov.pred %p132, -1; mov.u32 %r234, %r40; mov.f32 %f1004, %f1010; $L__BB22_60: setp.ne.s32 %p76, %r3, 0; and.pred %p77, %p76, %p132; @%p77 bra $L__BB22_92; bra.uni $L__BB22_86; $L__BB22_65: setp.gt.s32 %p88, %r64, %r63; selp.b32 %r171, -1, 0, %p88; add.s32 %r172, %r64, %r171; mad.lo.s32 %r173, %r172, %r62, %r40; add.s32 %r174, %r173, %r231; mul.wide.s32 %rd182, %r174, 16; add.s64 %rd219, %rd22, %rd182; $L__BB22_67: ld.global.v4.f32 {%f774, %f775, %f776, %f777}, [%rd219]; add.s32 %r75, %r64, 1; setp.eq.s32 %p89, %r75, %r63; @%p89 bra $L__BB22_69; bra.uni $L__BB22_68; $L__BB22_69: add.s64 %rd220, %rd20, %rd156; bra.uni $L__BB22_70; $L__BB22_68: setp.ge.s32 %p90, %r64, %r63; selp.b32 %r175, -1, 0, %p90; add.s32 %r176, %r75, %r175; mad.lo.s32 %r177, %r176, %r62, %r40; add.s32 %r178, %r177, %r231; mul.wide.s32 %rd184, %r178, 16; add.s64 %rd220, %rd22, %rd184; $L__BB22_70: cvt.rn.f32.s32 %f778, %r64; sub.ftz.f32 %f779, %f202, %f778; ld.global.v4.f32 {%f780, %f781, %f782, %f783}, [%rd220]; mov.f32 %f788, 0f3F800000; sub.ftz.f32 %f789, %f788, %f779; mul.ftz.f32 %f790, %f779, %f780; mul.ftz.f32 %f791, %f779, %f781; mul.ftz.f32 %f792, %f779, %f782; mul.ftz.f32 %f793, %f779, %f783; fma.rn.ftz.f32 %f996, %f789, %f777, %f793; fma.rn.ftz.f32 %f995, %f789, %f776, %f792; fma.rn.ftz.f32 %f994, %f789, %f775, %f791; fma.rn.ftz.f32 %f993, %f789, %f774, %f790; $L__BB22_72: sub.ftz.f32 %f798, %f993, %f1026; sub.ftz.f32 %f799, %f994, %f1025; sub.ftz.f32 %f800, %f995, %f1024; mul.ftz.f32 %f801, %f1028, %f799; fma.rn.ftz.f32 %f802, %f1029, %f798, %f801; fma.rn.ftz.f32 %f803, %f1027, %f800, %f802; mul.ftz.f32 %f219, %f153, %f803; mul.ftz.f32 %f804, %f1029, %f219; mul.ftz.f32 %f805, %f1028, %f219; mul.ftz.f32 %f806, %f1027, %f219; sub.ftz.f32 %f807, %f798, %f804; sub.ftz.f32 %f808, %f799, %f805; sub.ftz.f32 %f809, %f800, %f806; mul.ftz.f32 %f810, %f808, %f808; fma.rn.ftz.f32 %f811, %f807, %f807, %f810; fma.rn.ftz.f32 %f220, %f809, %f809, %f811; mul.ftz.f32 %f221, %f996, %f996; setp.gtu.ftz.f32 %p92, %f220, %f221; mov.pred %p133, 0; @%p92 bra $L__BB22_75; sub.ftz.f32 %f812, %f221, %f220; mul.ftz.f32 %f813, %f153, %f812; sqrt.approx.ftz.f32 %f814, %f813; sub.ftz.f32 %f222, %f219, %f814; setp.ltu.ftz.f32 %p94, %f222, %f10; setp.gtu.ftz.f32 %p95, %f222, %f1004; or.pred %p96, %p95, %p94; @%p96 bra $L__BB22_75; st.local.f32 [%rd3], %f222; st.local.u32 [%rd3+12], %r40; st.local.u32 [%rd3+16], %r39; st.local.u32 [%rd3+20], %r61; mov.u32 %r179, 0; st.local.u32 [%rd3+4], %r179; st.local.u32 [%rd3+8], %r179; mov.pred %p133, -1; mov.u32 %r234, %r40; mov.f32 %f1004, %f222; $L__BB22_75: setp.ne.s32 %p98, %r3, 0; and.pred %p99, %p98, %p133; @%p99 bra $L__BB22_92; bra.uni $L__BB22_86; $L__BB22_10: mul.wide.s32 %rd134, %r219, 4; xor.b64 %rd135, %rd134, -4; add.s64 %rd136, %rd7, %rd135; ld.global.u32 %r238, [%rd136]; mul.wide.s32 %rd138, %r238, 4; add.s64 %rd139, %rd137, %rd138; ld.global.u8 %rs15, [%rd139]; and.b16 %rs16, %rs15, 2; setp.eq.s16 %p28, %rs16, 0; mul.wide.s32 %rd140, %r238, 208; add.s64 %rd141, %rd10, %rd140; add.s64 %rd26, %rd141, 48; @%p28 bra $L__BB22_17; bra.uni $L__BB22_11; $L__BB22_17: ld.global.v4.f32 {%f964, %f965, %f966, %f967}, [%rd26]; ld.global.v4.f32 {%f968, %f969, %f970, %f971}, [%rd26+16]; ld.global.v4.f32 {%f972, %f973, %f974, %f975}, [%rd26+32]; bra.uni $L__BB22_18; $L__BB22_11: ld.global.u32 %rd142, [%rd26+120]; ld.global.u32 %r119, [%rd26+104]; shl.b32 %r120, %r119, 1; cvt.rn.f32.s32 %f378, %r120; ld.local.f32 %f379, [%rd2+32]; mul.ftz.f32 %f380, %f379, %f378; cvt.rzi.ftz.s32.f32 %r121, %f380; add.s32 %r122, %r120, -1; min.s32 %r123, %r121, %r122; cvt.rn.f32.s32 %f381, %r123; sub.ftz.f32 %f52, %f380, %f381; cvt.s64.s32 %rd143, %r123; add.s64 %rd144, %rd143, %rd142; shl.b64 %rd146, %rd144, 6; add.s64 %rd27, %rd145, %rd146; ld.global.v4.f32 {%f382, %f383, %f384, %f385}, [%rd27]; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd27+64]; mul.ftz.f32 %f390, %f383, %f387; fma.rn.ftz.f32 %f391, %f382, %f386, %f390; mul.ftz.f32 %f392, %f385, %f389; fma.rn.ftz.f32 %f393, %f384, %f388, %f392; add.ftz.f32 %f61, %f391, %f393; setp.gt.ftz.f32 %p29, %f61, 0f3F7FDF3B; mov.f32 %f394, 0f3F800000; sub.ftz.f32 %f62, %f394, %f52; @%p29 bra $L__BB22_13; bra.uni $L__BB22_12; $L__BB22_13: mul.ftz.f32 %f448, %f382, %f62; fma.rn.ftz.f32 %f449, %f386, %f52, %f448; mul.ftz.f32 %f450, %f383, %f62; fma.rn.ftz.f32 %f451, %f387, %f52, %f450; mul.ftz.f32 %f452, %f384, %f62; fma.rn.ftz.f32 %f453, %f388, %f52, %f452; mul.ftz.f32 %f454, %f385, %f62; fma.rn.ftz.f32 %f455, %f389, %f52, %f454; mul.ftz.f32 %f456, %f451, %f451; fma.rn.ftz.f32 %f457, %f449, %f449, %f456; mul.ftz.f32 %f458, %f455, %f455; fma.rn.ftz.f32 %f459, %f453, %f453, %f458; add.ftz.f32 %f460, %f457, %f459; rsqrt.approx.ftz.f32 %f461, %f460; mul.ftz.f32 %f956, %f455, %f461; mul.ftz.f32 %f955, %f453, %f461; mul.ftz.f32 %f954, %f451, %f461; mul.ftz.f32 %f953, %f449, %f461; bra.uni $L__BB22_14; $L__BB22_12: mov.f32 %f395, 0fBF800000; max.ftz.f32 %f396, %f61, %f395; min.ftz.f32 %f398, %f396, %f394; abs.ftz.f32 %f399, %f398; sub.ftz.f32 %f400, %f394, %f399; mul.ftz.f32 %f401, %f400, 0f3F000000; sqrt.approx.ftz.f32 %f402, %f401; setp.gt.ftz.f32 %p30, %f399, 0f3F11EB85; selp.f32 %f403, %f402, %f399, %p30; mul.ftz.f32 %f404, %f403, %f403; mov.f32 %f405, 0f3C94D2E9; mov.f32 %f406, 0f3D53F941; fma.rn.ftz.f32 %f407, %f406, %f404, %f405; mov.f32 %f408, 0f3D3F841F; fma.rn.ftz.f32 %f409, %f407, %f404, %f408; mov.f32 %f410, 0f3D994929; fma.rn.ftz.f32 %f411, %f409, %f404, %f410; mov.f32 %f412, 0f3E2AAB94; fma.rn.ftz.f32 %f413, %f411, %f404, %f412; mul.ftz.f32 %f414, %f404, %f413; fma.rn.ftz.f32 %f415, %f414, %f403, %f403; add.ftz.f32 %f416, %f415, %f415; mov.f32 %f417, 0f3FC90FDB; sub.ftz.f32 %f418, %f417, %f415; selp.f32 %f419, %f416, %f418, %p30; setp.lt.ftz.f32 %p31, %f398, 0f00000000; mov.f32 %f420, 0f40490FDB; sub.ftz.f32 %f421, %f420, %f419; selp.f32 %f422, %f421, %f419, %p31; mul.ftz.f32 %f423, %f382, %f61; sub.ftz.f32 %f424, %f386, %f423; mul.ftz.f32 %f425, %f383, %f61; sub.ftz.f32 %f426, %f387, %f425; mul.ftz.f32 %f427, %f384, %f61; sub.ftz.f32 %f428, %f388, %f427; mul.ftz.f32 %f429, %f385, %f61; sub.ftz.f32 %f430, %f389, %f429; mul.ftz.f32 %f431, %f426, %f426; fma.rn.ftz.f32 %f432, %f424, %f424, %f431; mul.ftz.f32 %f433, %f430, %f430; fma.rn.ftz.f32 %f434, %f428, %f428, %f433; add.ftz.f32 %f435, %f432, %f434; rsqrt.approx.ftz.f32 %f436, %f435; mul.ftz.f32 %f437, %f424, %f436; mul.ftz.f32 %f438, %f426, %f436; mul.ftz.f32 %f439, %f428, %f436; mul.ftz.f32 %f440, %f430, %f436; mul.ftz.f32 %f441, %f52, %f422; cos.approx.ftz.f32 %f442, %f441; sin.approx.ftz.f32 %f443, %f441; mul.ftz.f32 %f444, %f437, %f443; mul.ftz.f32 %f445, %f438, %f443; mul.ftz.f32 %f446, %f439, %f443; mul.ftz.f32 %f447, %f440, %f443; fma.rn.ftz.f32 %f956, %f385, %f442, %f447; fma.rn.ftz.f32 %f955, %f384, %f442, %f446; fma.rn.ftz.f32 %f954, %f383, %f442, %f445; fma.rn.ftz.f32 %f953, %f382, %f442, %f444; $L__BB22_14: ld.global.v4.f32 {%f462, %f463, %f464, %f465}, [%rd27+16]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd27+80]; mul.ftz.f32 %f478, %f52, %f470; mul.ftz.f32 %f479, %f52, %f471; mul.ftz.f32 %f480, %f52, %f472; mul.ftz.f32 %f481, %f52, %f473; fma.rn.ftz.f32 %f75, %f62, %f462, %f478; fma.rn.ftz.f32 %f76, %f62, %f463, %f479; fma.rn.ftz.f32 %f77, %f62, %f464, %f480; fma.rn.ftz.f32 %f482, %f62, %f465, %f481; ld.global.v4.f32 {%f483, %f484, %f485, %f486}, [%rd27+32]; ld.global.v4.f32 {%f491, %f492, %f493, %f494}, [%rd27+96]; mul.ftz.f32 %f499, %f52, %f491; mul.ftz.f32 %f500, %f52, %f492; mul.ftz.f32 %f501, %f52, %f493; mul.ftz.f32 %f502, %f52, %f494; fma.rn.ftz.f32 %f503, %f62, %f483, %f499; fma.rn.ftz.f32 %f504, %f62, %f484, %f500; fma.rn.ftz.f32 %f505, %f62, %f485, %f501; fma.rn.ftz.f32 %f506, %f62, %f486, %f502; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd27+48]; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd27+112]; mul.ftz.f32 %f523, %f52, %f515; mul.ftz.f32 %f524, %f52, %f516; mul.ftz.f32 %f525, %f52, %f517; mul.ftz.f32 %f526, %f52, %f518; fma.rn.ftz.f32 %f527, %f62, %f507, %f523; fma.rn.ftz.f32 %f528, %f62, %f508, %f524; fma.rn.ftz.f32 %f529, %f62, %f509, %f525; fma.rn.ftz.f32 %f530, %f62, %f510, %f526; mul.ftz.f32 %f531, %f956, 0f3FB504F3; mul.ftz.f32 %f532, %f953, 0f3FB504F3; mul.ftz.f32 %f533, %f954, 0f3FB504F3; mul.ftz.f32 %f534, %f955, 0f3FB504F3; mul.ftz.f32 %f535, %f531, %f532; mul.ftz.f32 %f536, %f531, %f533; mul.ftz.f32 %f537, %f531, %f534; mul.ftz.f32 %f538, %f532, %f532; mul.ftz.f32 %f539, %f532, %f533; mul.ftz.f32 %f540, %f532, %f534; mul.ftz.f32 %f541, %f533, %f533; mul.ftz.f32 %f542, %f533, %f534; mul.ftz.f32 %f543, %f534, %f534; sub.ftz.f32 %f545, %f394, %f541; sub.ftz.f32 %f546, %f545, %f543; sub.ftz.f32 %f547, %f539, %f537; add.ftz.f32 %f548, %f536, %f540; add.ftz.f32 %f549, %f537, %f539; sub.ftz.f32 %f550, %f394, %f538; sub.ftz.f32 %f551, %f550, %f543; sub.ftz.f32 %f552, %f542, %f535; sub.ftz.f32 %f553, %f540, %f536; add.ftz.f32 %f554, %f535, %f542; sub.ftz.f32 %f555, %f550, %f541; mul.ftz.f32 %f556, %f547, %f505; fma.rn.ftz.f32 %f557, %f546, %f482, %f556; fma.rn.ftz.f32 %f962, %f548, %f528, %f557; mul.ftz.f32 %f558, %f547, %f506; fma.rn.ftz.f32 %f559, %f546, %f503, %f558; fma.rn.ftz.f32 %f79, %f548, %f529, %f559; mul.ftz.f32 %f560, %f547, %f527; fma.rn.ftz.f32 %f561, %f546, %f504, %f560; fma.rn.ftz.f32 %f80, %f548, %f530, %f561; mul.ftz.f32 %f562, %f551, %f505; fma.rn.ftz.f32 %f563, %f549, %f482, %f562; fma.rn.ftz.f32 %f81, %f552, %f528, %f563; mul.ftz.f32 %f564, %f551, %f506; fma.rn.ftz.f32 %f565, %f549, %f503, %f564; fma.rn.ftz.f32 %f961, %f552, %f529, %f565; mul.ftz.f32 %f566, %f551, %f527; fma.rn.ftz.f32 %f567, %f549, %f504, %f566; fma.rn.ftz.f32 %f83, %f552, %f530, %f567; mul.ftz.f32 %f568, %f554, %f505; fma.rn.ftz.f32 %f569, %f553, %f482, %f568; fma.rn.ftz.f32 %f84, %f555, %f528, %f569; mul.ftz.f32 %f570, %f554, %f506; fma.rn.ftz.f32 %f571, %f553, %f503, %f570; fma.rn.ftz.f32 %f85, %f555, %f529, %f571; mul.ftz.f32 %f572, %f554, %f527; fma.rn.ftz.f32 %f573, %f553, %f504, %f572; fma.rn.ftz.f32 %f960, %f555, %f530, %f573; mul.ftz.f32 %f574, %f961, %f960; mul.ftz.f32 %f87, %f85, %f83; sub.ftz.f32 %f959, %f574, %f87; mul.ftz.f32 %f89, %f85, %f80; mul.ftz.f32 %f575, %f79, %f960; sub.ftz.f32 %f958, %f89, %f575; mul.ftz.f32 %f91, %f79, %f83; mul.ftz.f32 %f576, %f961, %f80; sub.ftz.f32 %f957, %f91, %f576; mul.ftz.f32 %f577, %f81, %f958; fma.rn.ftz.f32 %f578, %f962, %f959, %f577; fma.rn.ftz.f32 %f963, %f84, %f957, %f578; setp.neu.ftz.f32 %p32, %f963, 0f00000000; @%p32 bra $L__BB22_16; add.ftz.f32 %f962, %f962, 0f322BCC77; add.ftz.f32 %f960, %f960, 0f322BCC77; add.ftz.f32 %f961, %f961, 0f322BCC77; mul.ftz.f32 %f579, %f961, %f960; sub.ftz.f32 %f959, %f579, %f87; mul.ftz.f32 %f580, %f79, %f960; sub.ftz.f32 %f958, %f89, %f580; mul.ftz.f32 %f581, %f961, %f80; sub.ftz.f32 %f957, %f91, %f581; mul.ftz.f32 %f582, %f81, %f958; fma.rn.ftz.f32 %f583, %f962, %f959, %f582; fma.rn.ftz.f32 %f584, %f84, %f957, %f583; setp.eq.ftz.f32 %p33, %f584, 0f00000000; selp.f32 %f963, 0f7F7FFFFF, %f584, %p33; $L__BB22_16: mul.ftz.f32 %f585, %f81, %f960; mul.ftz.f32 %f586, %f84, %f83; sub.ftz.f32 %f587, %f586, %f585; mul.ftz.f32 %f588, %f84, %f80; mul.ftz.f32 %f589, %f960, %f962; sub.ftz.f32 %f590, %f589, %f588; mul.ftz.f32 %f591, %f83, %f962; mul.ftz.f32 %f592, %f81, %f80; sub.ftz.f32 %f593, %f592, %f591; div.approx.ftz.f32 %f968, %f587, %f963; div.approx.ftz.f32 %f969, %f590, %f963; div.approx.ftz.f32 %f970, %f593, %f963; mul.ftz.f32 %f594, %f84, %f961; mul.ftz.f32 %f595, %f81, %f85; sub.ftz.f32 %f596, %f595, %f594; mul.ftz.f32 %f597, %f85, %f962; mul.ftz.f32 %f598, %f84, %f79; sub.ftz.f32 %f599, %f598, %f597; mul.ftz.f32 %f600, %f81, %f79; mul.ftz.f32 %f601, %f961, %f962; sub.ftz.f32 %f602, %f601, %f600; div.approx.ftz.f32 %f972, %f596, %f963; div.approx.ftz.f32 %f973, %f599, %f963; div.approx.ftz.f32 %f974, %f602, %f963; div.approx.ftz.f32 %f964, %f959, %f963; div.approx.ftz.f32 %f965, %f958, %f963; mul.ftz.f32 %f603, %f76, %f965; fma.rn.ftz.f32 %f604, %f75, %f964, %f603; div.approx.ftz.f32 %f966, %f957, %f963; fma.rn.ftz.f32 %f605, %f77, %f966, %f604; neg.ftz.f32 %f967, %f605; mul.ftz.f32 %f606, %f76, %f969; fma.rn.ftz.f32 %f607, %f75, %f968, %f606; fma.rn.ftz.f32 %f608, %f77, %f970, %f607; neg.ftz.f32 %f971, %f608; mul.ftz.f32 %f609, %f76, %f973; fma.rn.ftz.f32 %f610, %f75, %f972, %f609; fma.rn.ftz.f32 %f611, %f77, %f974, %f610; neg.ftz.f32 %f975, %f611; $L__BB22_18: cvt.s64.s32 %rd198, %r238; mul.wide.s32 %rd197, %r23, 4; add.s64 %rd196, %rd1, %rd197; ld.local.f32 %f624, [%rd2]; ld.local.f32 %f625, [%rd2+4]; mul.ftz.f32 %f626, %f965, %f625; fma.rn.ftz.f32 %f627, %f964, %f624, %f626; ld.local.f32 %f628, [%rd2+8]; fma.rn.ftz.f32 %f629, %f966, %f628, %f627; add.ftz.f32 %f1026, %f967, %f629; mul.ftz.f32 %f630, %f969, %f625; fma.rn.ftz.f32 %f631, %f968, %f624, %f630; fma.rn.ftz.f32 %f632, %f970, %f628, %f631; add.ftz.f32 %f1025, %f971, %f632; mul.ftz.f32 %f633, %f973, %f625; fma.rn.ftz.f32 %f634, %f972, %f624, %f633; fma.rn.ftz.f32 %f635, %f974, %f628, %f634; add.ftz.f32 %f1024, %f975, %f635; ld.local.f32 %f636, [%rd2+12]; ld.local.f32 %f637, [%rd2+16]; mul.ftz.f32 %f638, %f965, %f637; fma.rn.ftz.f32 %f639, %f964, %f636, %f638; ld.local.f32 %f640, [%rd2+20]; fma.rn.ftz.f32 %f641, %f966, %f640, %f639; mul.ftz.f32 %f642, %f969, %f637; fma.rn.ftz.f32 %f643, %f968, %f636, %f642; fma.rn.ftz.f32 %f644, %f970, %f640, %f643; mul.ftz.f32 %f645, %f973, %f637; fma.rn.ftz.f32 %f646, %f972, %f636, %f645; fma.rn.ftz.f32 %f647, %f974, %f640, %f646; abs.ftz.f32 %f648, %f641; setp.gt.ftz.f32 %p34, %f648, 0f17800000; mov.b32 %r124, %f641; and.b32 %r125, %r124, -2147483648; or.b32 %r126, %r125, 394264576; mov.b32 %f649, %r126; selp.f32 %f1029, %f641, %f649, %p34; abs.ftz.f32 %f650, %f644; setp.gt.ftz.f32 %p35, %f650, 0f17800000; mov.b32 %r127, %f644; and.b32 %r128, %r127, -2147483648; or.b32 %r129, %r128, 394264576; mov.b32 %f651, %r129; selp.f32 %f1028, %f644, %f651, %p35; abs.ftz.f32 %f652, %f647; setp.gt.ftz.f32 %p36, %f652, 0f17800000; mov.b32 %r130, %f647; and.b32 %r131, %r130, -2147483648; or.b32 %r132, %r131, 394264576; mov.b32 %f653, %r132; selp.f32 %f1027, %f647, %f653, %p36; rcp.approx.ftz.f32 %f1032, %f1029; rcp.approx.ftz.f32 %f1031, %f1028; rcp.approx.ftz.f32 %f1030, %f1027; add.s32 %r214, %r23, 1; mov.u32 %r133, 1985229328; st.local.u32 [%rd196+4], %r133; shl.b64 %rd148, %rd198, 2; add.s64 %rd149, %rd147, %rd148; ld.global.u32 %r215, [%rd149]; bra.uni $L__BB22_87; $L__BB22_1: setp.eq.s32 %p10, %r215, 1985229328; setp.lt.s32 %p11, %r215, 0; or.pred %p12, %p11, %p10; @%p12 bra $L__BB22_8; $L__BB22_2: mul.wide.s32 %rd126, %r215, 16; add.s64 %rd127, %rd6, %rd126; ld.global.v4.f32 {%f308, %f309, %f310, %f311}, [%rd127+16]; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd127+32]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd127+48]; sub.ftz.f32 %f332, %f308, %f1026; mul.ftz.f32 %f333, %f1032, %f332; sub.ftz.f32 %f334, %f310, %f1026; mul.ftz.f32 %f335, %f1032, %f334; sub.ftz.f32 %f336, %f316, %f1025; mul.ftz.f32 %f337, %f1031, %f336; sub.ftz.f32 %f338, %f318, %f1025; mul.ftz.f32 %f339, %f1031, %f338; sub.ftz.f32 %f340, %f324, %f1024; mul.ftz.f32 %f341, %f1030, %f340; sub.ftz.f32 %f342, %f326, %f1024; mul.ftz.f32 %f343, %f1030, %f342; min.ftz.f32 %f344, %f333, %f335; min.ftz.f32 %f345, %f337, %f339; min.ftz.f32 %f346, %f341, %f343; max.ftz.f32 %f347, %f10, %f344; max.ftz.f32 %f348, %f345, %f346; max.ftz.f32 %f50, %f347, %f348; max.ftz.f32 %f349, %f333, %f335; max.ftz.f32 %f350, %f337, %f339; max.ftz.f32 %f351, %f341, %f343; min.ftz.f32 %f352, %f1004, %f349; min.ftz.f32 %f353, %f350, %f351; min.ftz.f32 %f354, %f352, %f353; sub.ftz.f32 %f355, %f309, %f1026; mul.ftz.f32 %f356, %f1032, %f355; sub.ftz.f32 %f357, %f311, %f1026; mul.ftz.f32 %f358, %f1032, %f357; sub.ftz.f32 %f359, %f317, %f1025; mul.ftz.f32 %f360, %f1031, %f359; sub.ftz.f32 %f361, %f319, %f1025; mul.ftz.f32 %f362, %f1031, %f361; sub.ftz.f32 %f363, %f325, %f1024; mul.ftz.f32 %f364, %f1030, %f363; sub.ftz.f32 %f365, %f327, %f1024; mul.ftz.f32 %f366, %f1030, %f365; min.ftz.f32 %f367, %f356, %f358; min.ftz.f32 %f368, %f360, %f362; min.ftz.f32 %f369, %f364, %f366; max.ftz.f32 %f370, %f10, %f367; max.ftz.f32 %f371, %f368, %f369; max.ftz.f32 %f51, %f370, %f371; max.ftz.f32 %f372, %f356, %f358; max.ftz.f32 %f373, %f360, %f362; max.ftz.f32 %f374, %f364, %f366; min.ftz.f32 %f375, %f1004, %f372; min.ftz.f32 %f376, %f373, %f374; min.ftz.f32 %f377, %f375, %f376; setp.ge.ftz.f32 %p13, %f354, %f50; ld.global.v4.u32 {%r104, %r105, %r106, %r215}, [%rd127]; and.b32 %r109, %r104, %r90; setp.ne.s32 %p14, %r109, 0; and.pred %p15, %p14, %p13; selp.u32 %r110, 1, 0, %p15; setp.ltu.ftz.f32 %p16, %f377, %f51; and.b32 %r112, %r105, %r90; setp.eq.s32 %p17, %r112, 0; or.pred %p18, %p17, %p16; selp.b32 %r113, 0, 2, %p18; or.b32 %r14, %r113, %r110; setp.eq.s32 %p19, %r14, 3; mul.wide.s32 %rd128, %r214, 4; add.s64 %rd23, %rd1, %rd128; @%p19 bra $L__BB22_6; bra.uni $L__BB22_3; $L__BB22_6: setp.lt.ftz.f32 %p22, %f51, %f50; selp.b32 %r19, %r215, %r106, %p22; selp.b32 %r114, %r106, %r215, %p22; add.s32 %r214, %r214, 1; st.local.u32 [%rd23+4], %r114; mov.u32 %r215, %r19; bra.uni $L__BB22_7; $L__BB22_3: cvt.u16.u32 %rs14, %r14; setp.eq.s16 %p20, %rs14, 2; @%p20 bra $L__BB22_7; setp.ne.s16 %p21, %rs14, 0; mov.u32 %r215, %r106; @%p21 bra $L__BB22_7; ld.local.u32 %r215, [%rd23]; add.s32 %r214, %r214, -1; $L__BB22_7: setp.ne.s32 %p23, %r215, 1985229328; setp.gt.s32 %p24, %r215, -1; and.pred %p25, %p24, %p23; @%p25 bra $L__BB22_2; $L__BB22_8: mov.u32 %r23, %r214; setp.gt.s32 %p26, %r215, -1; @%p26 bra $L__BB22_87; bra.uni $L__BB22_9; $L__BB22_87: setp.ne.s32 %p119, %r215, 1985229328; @%p119 bra $L__BB22_1; mov.u32 %r215, 1985229328; setp.lt.s32 %p120, %r214, 0; @%p120 bra $L__BB22_90; ld.local.f32 %f1026, [%rd2]; ld.local.f32 %f1025, [%rd2+4]; ld.local.f32 %f1024, [%rd2+8]; ld.local.f32 %f901, [%rd2+12]; abs.ftz.f32 %f902, %f901; setp.gt.ftz.f32 %p121, %f902, 0f17800000; mov.b32 %r189, %f901; and.b32 %r190, %r189, -2147483648; or.b32 %r191, %r190, 394264576; mov.b32 %f903, %r191; selp.f32 %f1029, %f901, %f903, %p121; ld.local.f32 %f904, [%rd2+16]; abs.ftz.f32 %f905, %f904; setp.gt.ftz.f32 %p122, %f905, 0f17800000; mov.b32 %r192, %f904; and.b32 %r193, %r192, -2147483648; or.b32 %r194, %r193, 394264576; mov.b32 %f906, %r194; selp.f32 %f1028, %f904, %f906, %p122; ld.local.f32 %f907, [%rd2+20]; abs.ftz.f32 %f908, %f907; setp.gt.ftz.f32 %p123, %f908, 0f17800000; mov.b32 %r195, %f907; and.b32 %r196, %r195, -2147483648; or.b32 %r197, %r196, 394264576; mov.b32 %f909, %r197; selp.f32 %f1027, %f907, %f909, %p123; rcp.approx.ftz.f32 %f1032, %f1029; rcp.approx.ftz.f32 %f1031, %f1028; rcp.approx.ftz.f32 %f1030, %f1027; mul.wide.s32 %rd194, %r214, 4; add.s64 %rd195, %rd1, %rd194; ld.local.u32 %r215, [%rd195]; add.s32 %r214, %r214, -1; mov.u32 %r238, -1; $L__BB22_90: setp.ne.s32 %p124, %r215, 1985229328; @%p124 bra $L__BB22_1; setp.ne.s32 %p125, %r234, -1; selp.u16 %rs39, 1, 0, %p125; $L__BB22_92: cvt.u32.u16 %r198, %rs39; st.param.b32 [func_retval0+0], %r198; ret; } .func (.param .b32 func_retval0) _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj( .param .b64 _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0, .param .b64 _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1, .param .b32 _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2 ) { .local .align 16 .b8 __local_depot23[816]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<251>; .reg .b16 %rs<80>; .reg .f32 %f<3042>; .reg .b32 %r<336>; .reg .b64 %rd<270>; mov.u64 %SPL, __local_depot23; ld.param.u64 %rd108, [_Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_0]; ld.param.u64 %rd109, [_Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_1]; ld.param.u32 %r135, [_Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj_param_2]; cvta.to.local.u64 %rd4, %rd109; add.u64 %rd1, %SPL, 0; add.u64 %rd2, %SPL, 48; mov.u32 %r139, 1985229328; st.local.u32 [%rd2], %r139; ld.const.u32 %r293, [kernel_params+1104]; cvta.to.local.u64 %rd3, %rd108; ld.local.f32 %f3026, [%rd3]; ld.local.f32 %f3025, [%rd3+4]; ld.local.f32 %f3024, [%rd3+8]; ld.local.f32 %f744, [%rd3+12]; abs.ftz.f32 %f745, %f744; setp.gt.ftz.f32 %p12, %f745, 0f17800000; mov.b32 %r140, %f744; and.b32 %r141, %r140, -2147483648; or.b32 %r142, %r141, 394264576; mov.b32 %f746, %r142; selp.f32 %f3029, %f744, %f746, %p12; ld.local.f32 %f747, [%rd3+16]; abs.ftz.f32 %f748, %f747; setp.gt.ftz.f32 %p13, %f748, 0f17800000; mov.b32 %r143, %f747; and.b32 %r144, %r143, -2147483648; or.b32 %r145, %r144, 394264576; mov.b32 %f749, %r145; selp.f32 %f3028, %f747, %f749, %p13; ld.local.f32 %f750, [%rd3+20]; abs.ftz.f32 %f751, %f750; setp.gt.ftz.f32 %p14, %f751, 0f17800000; mov.b32 %r146, %f750; and.b32 %r147, %r146, -2147483648; or.b32 %r148, %r147, 394264576; mov.b32 %f752, %r148; selp.f32 %f3027, %f750, %f752, %p14; rcp.approx.ftz.f32 %f3032, %f3029; mov.f32 %f753, 0f3F800000; rcp.approx.ftz.f32 %f3031, %f3028; rcp.approx.ftz.f32 %f3030, %f3027; ld.local.f32 %f10, [%rd3+24]; ld.local.f32 %f2992, [%rd3+28]; st.local.f32 [%rd4], %f2992; mov.u32 %r292, 0; st.local.u32 [%rd4+4], %r292; st.local.u32 [%rd4+8], %r292; mov.u32 %r328, -1; st.local.u32 [%rd4+12], %r328; st.local.u32 [%rd4+16], %r328; ld.const.v2.u64 {%rd112, %rd113}, [kernel_params+2096]; cvta.to.global.u64 %rd7, %rd112; ld.const.v2.u64 {%rd114, %rd115}, [kernel_params+2128]; cvta.to.global.u64 %rd8, %rd115; ld.const.v2.u64 {%rd118, %rd119}, [kernel_params+2176]; ld.const.u64 %rd120, [kernel_params+2160]; cvta.to.global.u64 %rd11, %rd120; cvta.to.global.u64 %rd12, %rd114; ld.const.u64 %rd121, [kernel_params+2232]; cvta.to.global.u64 %rd13, %rd121; ld.const.v2.u64 {%rd122, %rd123}, [kernel_params+2256]; cvta.to.global.u64 %rd16, %rd122; ld.const.v2.u32 {%r149, %r150}, [kernel_params+1128]; ld.const.u64 %rd124, [kernel_params+2312]; cvta.to.global.u64 %rd17, %rd124; ld.const.v2.u64 {%rd125, %rd126}, [kernel_params+2112]; cvta.to.global.u64 %rd18, %rd125; ld.const.v2.u64 {%rd129, %rd130}, [kernel_params+2144]; cvta.to.global.u64 %rd21, %rd130; and.b32 %r5, %r135, 256; cvta.to.global.u64 %rd22, %rd126; ld.const.u64 %rd131, [kernel_params+2272]; cvta.to.global.u64 %rd23, %rd131; ld.const.u64 %rd132, [kernel_params+2296]; cvta.to.global.u64 %rd24, %rd132; ld.const.v2.u64 {%rd133, %rd134}, [kernel_params+2336]; cvta.to.global.u64 %rd25, %rd133; cvt.rn.f32.s32 %f754, %r150; rcp.approx.ftz.f32 %f12, %f754; sub.ftz.f32 %f755, %f753, %f12; mul.ftz.f32 %f756, %f12, %f755; mul.ftz.f32 %f757, %f755, %f756; neg.ftz.f32 %f13, %f757; mul.ftz.f32 %f758, %f12, %f12; fma.rn.ftz.f32 %f759, %f12, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f14, %f758, %f759, 0f40000000; mul.ftz.f32 %f760, %f755, %f755; fma.rn.ftz.f32 %f761, %f755, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f15, %f760, %f761, 0f40000000; mul.ftz.f32 %f762, %f12, %f756; neg.ftz.f32 %f16, %f762; cvta.to.global.u64 %rd26, %rd134; cvta.to.global.u64 %rd140, %rd113; setp.eq.s32 %p49, %r149, 0; mov.u16 %rs79, 1; mov.pred %p215, -1; mov.pred %p227, 0; cvta.to.global.u64 %rd207, %rd123; cvta.to.global.u64 %rd148, %rd119; cvta.to.global.u64 %rd158, %rd129; cvta.to.global.u64 %rd156, %rd118; mov.u32 %r332, %r328; bra.uni $L__BB23_1; $L__BB23_13: mul.wide.s32 %rd141, %r293, 16; xor.b64 %rd142, %rd141, -16; add.s64 %rd143, %rd140, %rd142; ld.global.v4.u32 {%r297, %r162, %r163, %r164}, [%rd143]; setp.gt.s32 %p38, %r297, -1; mul.wide.s32 %rd144, %r33, 4; add.s64 %rd29, %rd2, %rd144; @%p38 bra $L__BB23_23; bra.uni $L__BB23_14; $L__BB23_23: ld.local.u32 %r293, [%rd29]; add.s32 %r292, %r33, -1; setp.le.s32 %p48, %r162, %r297; @%p48 bra $L__BB23_177; cvt.u16.u32 %rs32, %r164; and.b16 %rs1, %rs32, 127; and.b32 %r180, %r164, 64; setp.eq.s32 %p50, %r180, 0; or.pred %p1, %p50, %p49; mul.ftz.f32 %f1252, %f3029, %f3029; fma.rn.ftz.f32 %f1253, %f3028, %f3028, %f1252; fma.rn.ftz.f32 %f175, %f3027, %f3027, %f1253; rcp.approx.ftz.f32 %f176, %f175; $L__BB23_25: cvt.s64.s32 %rd33, %r297; setp.ne.s32 %p51, %r332, -1; mov.u32 %r49, %r332; @%p51 bra $L__BB23_27; shl.b64 %rd161, %rd33, 2; add.s64 %rd162, %rd8, %rd161; ld.global.u32 %r49, [%rd162]; $L__BB23_27: shl.b64 %rd163, %rd33, 2; add.s64 %rd164, %rd12, %rd163; ld.global.u32 %r50, [%rd164]; cvt.s64.s32 %rd34, %r50; ld.local.u32 %r181, [%rd3+36]; setp.ne.s32 %p52, %r181, %r50; @%p52 bra $L__BB23_29; ld.local.u32 %r182, [%rd3+40]; setp.eq.s32 %p53, %r182, %r49; mov.u16 %rs62, %rs79; @%p53 bra $L__BB23_31; $L__BB23_29: ld.local.u32 %r183, [%rd3+44]; setp.ne.s32 %p54, %r183, %r50; mov.u16 %rs62, 0; @%p54 bra $L__BB23_31; ld.local.u32 %r184, [%rd3+48]; setp.eq.s32 %p55, %r184, %r49; selp.u16 %rs62, 1, 0, %p55; $L__BB23_31: setp.ne.s16 %p56, %rs62, 0; @%p56 bra $L__BB23_176; bra.uni $L__BB23_32; $L__BB23_176: add.s32 %r292, %r33, -1; cvt.u32.u64 %r266, %rd33; add.s32 %r297, %r266, 1; setp.lt.s32 %p233, %r297, %r162; @%p233 bra $L__BB23_25; bra.uni $L__BB23_177; $L__BB23_32: setp.gt.s16 %p57, %rs1, 64; @%p57 bra $L__BB23_38; setp.gt.s16 %p63, %rs1, 3; @%p63 bra $L__BB23_36; setp.eq.s16 %p66, %rs1, 1; @%p66 bra $L__BB23_166; setp.eq.s16 %p67, %rs1, 2; @%p67 bra $L__BB23_41; bra.uni $L__BB23_176; $L__BB23_38: setp.gt.s16 %p58, %rs1, 67; @%p58 bra $L__BB23_47; setp.eq.s16 %p61, %rs1, 65; @%p61 bra $L__BB23_55; setp.eq.s16 %p62, %rs1, 66; @%p62 bra $L__BB23_41; bra.uni $L__BB23_176; $L__BB23_36: setp.eq.s16 %p64, %rs1, 4; @%p64 bra $L__BB23_41; setp.eq.s16 %p65, %rs1, 8; @%p65 bra $L__BB23_49; bra.uni $L__BB23_176; $L__BB23_47: setp.eq.s16 %p59, %rs1, 68; @%p59 bra $L__BB23_41; setp.eq.s16 %p60, %rs1, 72; @%p60 bra $L__BB23_49; bra.uni $L__BB23_176; $L__BB23_49: @%p1 bra $L__BB23_51; shl.b64 %rd187, %rd33, 3; add.s64 %rd188, %rd21, %rd187; ld.global.v2.f32 {%f1364, %f1365}, [%rd188]; ld.local.f32 %f1368, [%rd3+32]; setp.lt.ftz.f32 %p95, %f1368, %f1364; setp.gt.ftz.f32 %p96, %f1368, %f1365; or.pred %p97, %p95, %p96; @%p97 bra $L__BB23_176; $L__BB23_51: add.s64 %rd190, %rd18, %rd163; ld.global.u32 %r70, [%rd190]; and.b32 %r212, %r70, 64; setp.eq.s32 %p98, %r212, 0; @%p98 bra $L__BB23_87; bra.uni $L__BB23_52; $L__BB23_87: shl.b64 %rd201, %rd34, 4; add.s64 %rd202, %rd24, %rd201; ld.global.v4.f32 {%f2849, %f2850, %f2851, %f2852}, [%rd202]; bra.uni $L__BB23_88; $L__BB23_41: @%p1 bra $L__BB23_43; shl.b64 %rd203, %rd33, 3; add.s64 %rd204, %rd21, %rd203; ld.global.v2.f32 {%f1412, %f1413}, [%rd204]; ld.local.f32 %f1416, [%rd3+32]; setp.lt.ftz.f32 %p117, %f1416, %f1412; setp.gt.ftz.f32 %p118, %f1416, %f1413; or.pred %p119, %p117, %p118; @%p119 bra $L__BB23_176; $L__BB23_43: add.s64 %rd206, %rd18, %rd163; ld.global.u32 %r86, [%rd206]; and.b32 %r226, %r86, 64; setp.eq.s32 %p120, %r226, 0; shl.b64 %rd208, %rd34, 4; add.s64 %rd209, %rd207, %rd208; shr.s32 %r227, %r86, 7; ld.global.u32 %r228, [%rd209+4]; add.s32 %r87, %r228, %r227; add.s32 %r229, %r87, -1; max.s32 %r88, %r229, %r228; add.s32 %r230, %r87, 2; ld.global.u32 %r231, [%rd209+8]; add.s32 %r232, %r228, %r231; add.s32 %r233, %r232, -1; min.s32 %r89, %r230, %r233; mul.wide.s32 %rd210, %r88, 16; add.s64 %rd102, %rd23, %rd210; mul.wide.s32 %rd211, %r87, 16; add.s64 %rd103, %rd23, %rd211; mul.wide.s32 %rd212, %r89, 16; add.s64 %rd104, %rd23, %rd212; @%p120 bra $L__BB23_102; bra.uni $L__BB23_44; $L__BB23_102: ld.global.v4.f32 {%f2888, %f2889, %f2890, %f388}, [%rd102]; ld.global.v4.f32 {%f2892, %f2893, %f2894, %f392}, [%rd103]; ld.global.v4.f32 {%f2896, %f2897, %f2898, %f396}, [%rd103+16]; ld.global.v4.f32 {%f2900, %f2901, %f2902, %f400}, [%rd104]; bra.uni $L__BB23_103; $L__BB23_44: mul.wide.s32 %rd213, %r49, 208; add.s64 %rd214, %rd11, %rd213; ld.global.u32 %r90, [%rd214+148]; ld.global.u32 %r91, [%rd214+152]; shl.b32 %r234, %r91, 1; cvt.rn.f32.s32 %f1417, %r234; ld.local.f32 %f1418, [%rd3+32]; mul.ftz.f32 %f256, %f1418, %f1417; cvt.rzi.ftz.s32.f32 %r235, %f256; add.s32 %r236, %r234, -1; min.s32 %r92, %r235, %r236; ld.global.u32 %r312, [%rd214+164]; mul.wide.u32 %rd215, %r312, 16; add.s64 %rd216, %rd17, %rd215; ld.global.u16 %rs67, [%rd216+12]; ld.global.u32 %r311, [%rd216+8]; ld.global.u64 %rd269, [%rd216]; setp.eq.s64 %p121, %rd269, 11; @%p121 bra $L__BB23_94; $L__BB23_45: setp.eq.s64 %p122, %rd269, 0; @%p122 bra $L__BB23_92; add.s32 %r312, %r312, 2; bra.uni $L__BB23_93; $L__BB23_92: setp.eq.s16 %p123, %rs67, 0; mov.u32 %r315, -1; mov.u32 %r312, %r311; @%p123 bra $L__BB23_95; $L__BB23_93: mul.wide.u32 %rd217, %r312, 16; add.s64 %rd218, %rd17, %rd217; ld.global.u16 %rs67, [%rd218+12]; ld.global.u32 %r311, [%rd218+8]; ld.global.u64 %rd269, [%rd218]; setp.ne.s64 %p124, %rd269, 11; @%p124 bra $L__BB23_45; $L__BB23_94: setp.eq.s16 %p125, %rs67, 0; selp.b32 %r315, -1, %r311, %p125; $L__BB23_95: setp.eq.s32 %p126, %r92, %r91; @%p126 bra $L__BB23_97; bra.uni $L__BB23_96; $L__BB23_97: ld.global.v4.f32 {%f2856, %f2857, %f2858, %f2859}, [%rd102]; ld.global.v4.f32 {%f2860, %f2861, %f2862, %f2863}, [%rd103]; ld.global.v4.f32 {%f2864, %f2865, %f2866, %f2867}, [%rd103+16]; ld.global.v4.f32 {%f2868, %f2869, %f2870, %f2871}, [%rd104]; bra.uni $L__BB23_98; $L__BB23_166: shl.b64 %rd235, %rd34, 4; add.s64 %rd236, %rd13, %rd235; ld.global.u32 %r260, [%rd236+12]; mul.wide.u32 %rd237, %r260, 12; add.s64 %rd238, %rd16, %rd237; add.s32 %r261, %r260, 1; mul.wide.u32 %rd239, %r261, 12; add.s64 %rd240, %rd16, %rd239; add.s32 %r262, %r260, 2; mul.wide.u32 %rd241, %r262, 12; add.s64 %rd242, %rd16, %rd241; ld.global.f32 %f2632, [%rd238]; sub.ftz.f32 %f658, %f2632, %f3026; ld.global.f32 %f2633, [%rd238+4]; sub.ftz.f32 %f659, %f2633, %f3025; ld.global.f32 %f2634, [%rd238+8]; sub.ftz.f32 %f660, %f2634, %f3024; ld.global.f32 %f2635, [%rd240]; sub.ftz.f32 %f2636, %f2635, %f3026; ld.global.f32 %f2637, [%rd240+4]; sub.ftz.f32 %f2638, %f2637, %f3025; ld.global.f32 %f2639, [%rd240+8]; sub.ftz.f32 %f2640, %f2639, %f3024; ld.global.f32 %f2641, [%rd242]; sub.ftz.f32 %f2642, %f2641, %f3026; ld.global.f32 %f2643, [%rd242+4]; sub.ftz.f32 %f2644, %f2643, %f3025; ld.global.f32 %f2645, [%rd242+8]; sub.ftz.f32 %f2646, %f2645, %f3024; sub.ftz.f32 %f661, %f2642, %f658; sub.ftz.f32 %f662, %f2644, %f659; sub.ftz.f32 %f663, %f2646, %f660; sub.ftz.f32 %f664, %f658, %f2636; sub.ftz.f32 %f665, %f659, %f2638; sub.ftz.f32 %f666, %f660, %f2640; sub.ftz.f32 %f2647, %f2636, %f2642; sub.ftz.f32 %f2648, %f2638, %f2644; sub.ftz.f32 %f2649, %f2640, %f2646; add.ftz.f32 %f2650, %f658, %f2642; add.ftz.f32 %f2651, %f659, %f2644; add.ftz.f32 %f2652, %f660, %f2646; mul.ftz.f32 %f2653, %f662, %f2652; mul.ftz.f32 %f2654, %f2651, %f663; sub.ftz.f32 %f2655, %f2653, %f2654; mul.ftz.f32 %f2656, %f2650, %f663; mul.ftz.f32 %f2657, %f661, %f2652; sub.ftz.f32 %f2658, %f2656, %f2657; mul.ftz.f32 %f2659, %f661, %f2651; mul.ftz.f32 %f2660, %f2650, %f662; sub.ftz.f32 %f2661, %f2659, %f2660; mul.ftz.f32 %f2662, %f3028, %f2658; fma.rn.ftz.f32 %f2663, %f3029, %f2655, %f2662; fma.rn.ftz.f32 %f667, %f3027, %f2661, %f2663; add.ftz.f32 %f2664, %f658, %f2636; add.ftz.f32 %f2665, %f659, %f2638; add.ftz.f32 %f2666, %f660, %f2640; mul.ftz.f32 %f2667, %f665, %f2666; mul.ftz.f32 %f2668, %f2665, %f666; sub.ftz.f32 %f2669, %f2667, %f2668; mul.ftz.f32 %f2670, %f2664, %f666; mul.ftz.f32 %f2671, %f664, %f2666; sub.ftz.f32 %f2672, %f2670, %f2671; mul.ftz.f32 %f2673, %f664, %f2665; mul.ftz.f32 %f2674, %f2664, %f665; sub.ftz.f32 %f2675, %f2673, %f2674; mul.ftz.f32 %f2676, %f3028, %f2672; fma.rn.ftz.f32 %f2677, %f3029, %f2669, %f2676; fma.rn.ftz.f32 %f668, %f3027, %f2675, %f2677; add.ftz.f32 %f2678, %f2636, %f2642; add.ftz.f32 %f2679, %f2638, %f2644; add.ftz.f32 %f2680, %f2640, %f2646; mul.ftz.f32 %f2681, %f2648, %f2680; mul.ftz.f32 %f2682, %f2679, %f2649; sub.ftz.f32 %f2683, %f2681, %f2682; mul.ftz.f32 %f2684, %f2678, %f2649; mul.ftz.f32 %f2685, %f2647, %f2680; sub.ftz.f32 %f2686, %f2684, %f2685; mul.ftz.f32 %f2687, %f2647, %f2679; mul.ftz.f32 %f2688, %f2678, %f2648; sub.ftz.f32 %f2689, %f2687, %f2688; mul.ftz.f32 %f2690, %f3028, %f2686; fma.rn.ftz.f32 %f2691, %f3029, %f2683, %f2690; fma.rn.ftz.f32 %f2692, %f3027, %f2689, %f2691; add.ftz.f32 %f2693, %f668, %f667; add.ftz.f32 %f669, %f2692, %f2693; abs.ftz.f32 %f670, %f669; mul.ftz.f32 %f2694, %f670, 0f34000000; min.ftz.f32 %f2695, %f668, %f2692; min.ftz.f32 %f2696, %f667, %f2695; max.ftz.f32 %f2697, %f668, %f2692; max.ftz.f32 %f2698, %f667, %f2697; mul.ftz.f32 %f2699, %f670, 0fB4000000; setp.ltu.ftz.f32 %p216, %f2696, %f2699; setp.gtu.ftz.f32 %p217, %f2698, %f2694; and.pred %p218, %p216, %p217; mov.pred %p249, %p215; @%p218 bra $L__BB23_172; mul.ftz.f32 %f2700, %f665, %f663; mul.ftz.f32 %f2701, %f666, %f662; sub.ftz.f32 %f2702, %f2700, %f2701; mul.ftz.f32 %f2703, %f664, %f663; mul.ftz.f32 %f2704, %f666, %f661; sub.ftz.f32 %f2705, %f2704, %f2703; mul.ftz.f32 %f2706, %f665, %f661; mul.ftz.f32 %f2707, %f664, %f662; sub.ftz.f32 %f2708, %f2707, %f2706; add.ftz.f32 %f671, %f2702, %f2702; add.ftz.f32 %f672, %f2705, %f2705; add.ftz.f32 %f673, %f2708, %f2708; mul.ftz.f32 %f2709, %f3028, %f672; fma.rn.ftz.f32 %f2710, %f3029, %f671, %f2709; fma.rn.ftz.f32 %f674, %f3027, %f673, %f2710; setp.eq.ftz.f32 %p220, %f674, 0f00000000; mov.pred %p249, %p215; @%p220 bra $L__BB23_172; mul.ftz.f32 %f2711, %f659, %f672; fma.rn.ftz.f32 %f2712, %f658, %f671, %f2711; fma.rn.ftz.f32 %f2713, %f660, %f673, %f2712; div.approx.ftz.f32 %f675, %f2713, %f674; setp.ltu.ftz.f32 %p222, %f675, %f10; setp.gtu.ftz.f32 %p223, %f675, %f2992; or.pred %p224, %p223, %p222; mov.pred %p249, %p215; @%p224 bra $L__BB23_172; setp.lt.ftz.f32 %p225, %f670, 0f219392EF; mov.f32 %f2987, 0f00000000; @%p225 bra $L__BB23_171; rcp.approx.ftz.f32 %f2987, %f669; $L__BB23_171: mul.ftz.f32 %f2715, %f667, %f2987; mov.f32 %f2716, 0f3F800000; min.ftz.f32 %f2998, %f2715, %f2716; mul.ftz.f32 %f2717, %f668, %f2987; min.ftz.f32 %f2997, %f2717, %f2716; mov.pred %p249, 0; mov.f32 %f2999, %f675; $L__BB23_172: mov.pred %p250, %p227; @%p249 bra $L__BB23_175; add.s64 %rd244, %rd22, %rd163; ld.global.u32 %r263, [%rd244]; and.b32 %r264, %r263, %r135; setp.eq.s32 %p229, %r264, 0; mov.pred %p250, %p227; @%p229 bra $L__BB23_175; st.local.u32 [%rd4+16], %r49; st.local.u32 [%rd4+12], %r50; mov.u32 %r265, 1; st.local.u32 [%rd4+20], %r265; st.local.f32 [%rd4+4], %f2998; st.local.f32 [%rd4+8], %f2997; st.local.f32 [%rd4], %f2999; mov.pred %p250, -1; mov.u32 %r328, %r50; mov.f32 %f2992, %f2999; $L__BB23_175: setp.ne.s32 %p231, %r5, 0; and.pred %p232, %p231, %p250; @%p232 bra $L__BB23_182; bra.uni $L__BB23_176; $L__BB23_55: mul.wide.s32 %rd165, %r49, 208; add.s64 %rd166, %rd11, %rd165; ld.global.u32 %r51, [%rd166+156]; ld.global.u32 %r52, [%rd166+152]; shl.b32 %r185, %r52, 1; cvt.rn.f32.s32 %f1254, %r185; ld.local.f32 %f1255, [%rd3+32]; mul.ftz.f32 %f193, %f1255, %f1254; cvt.rzi.ftz.s32.f32 %r186, %f193; add.s32 %r187, %r185, -1; min.s32 %r53, %r186, %r187; ld.global.u32 %r300, [%rd166+164]; mul.wide.u32 %rd167, %r300, 16; add.s64 %rd168, %rd17, %rd167; ld.global.u16 %rs63, [%rd168+12]; ld.global.u32 %r299, [%rd168+8]; ld.global.u64 %rd247, [%rd168]; setp.eq.s64 %p68, %rd247, 11; @%p68 bra $L__BB23_60; $L__BB23_56: setp.eq.s64 %p69, %rd247, 0; @%p69 bra $L__BB23_58; add.s32 %r300, %r300, 2; bra.uni $L__BB23_59; $L__BB23_58: setp.eq.s16 %p70, %rs63, 0; mov.u32 %r303, -1; mov.u32 %r300, %r299; @%p70 bra $L__BB23_61; $L__BB23_59: mul.wide.u32 %rd169, %r300, 16; add.s64 %rd170, %rd17, %rd169; ld.global.u16 %rs63, [%rd170+12]; ld.global.u32 %r299, [%rd170+8]; ld.global.u64 %rd247, [%rd170]; setp.ne.s64 %p71, %rd247, 11; @%p71 bra $L__BB23_56; $L__BB23_60: setp.eq.s16 %p72, %rs63, 0; selp.b32 %r303, -1, %r299, %p72; $L__BB23_61: shl.b64 %rd171, %rd34, 4; add.s64 %rd172, %rd13, %rd171; ld.global.v4.u32 {%r189, %r190, %r191, %r192}, [%rd172]; setp.eq.s32 %p73, %r53, %r52; @%p73 bra $L__BB23_63; bra.uni $L__BB23_62; $L__BB23_63: mul.wide.u32 %rd176, %r192, 12; add.s64 %rd248, %rd16, %rd176; add.s32 %r199, %r192, 1; mul.wide.u32 %rd177, %r199, 12; add.s64 %rd251, %rd16, %rd177; add.s32 %r200, %r192, 2; mul.wide.u32 %rd178, %r200, 12; add.s64 %rd254, %rd16, %rd178; add.s64 %rd249, %rd248, 4; add.s64 %rd250, %rd248, 8; add.s64 %rd252, %rd251, 4; add.s64 %rd253, %rd251, 8; add.s64 %rd255, %rd254, 4; add.s64 %rd256, %rd254, 8; bra.uni $L__BB23_64; $L__BB23_52: mul.wide.s32 %rd191, %r49, 208; add.s64 %rd192, %rd11, %rd191; ld.global.u32 %r71, [%rd192+148]; ld.global.u32 %r72, [%rd192+152]; shl.b32 %r213, %r72, 1; cvt.rn.f32.s32 %f1369, %r213; ld.local.f32 %f1370, [%rd3+32]; mul.ftz.f32 %f232, %f1370, %f1369; cvt.rzi.ftz.s32.f32 %r214, %f232; add.s32 %r215, %r213, -1; min.s32 %r73, %r214, %r215; ld.global.u32 %r306, [%rd192+164]; mul.wide.u32 %rd193, %r306, 16; add.s64 %rd194, %rd17, %rd193; ld.global.u16 %rs65, [%rd194+12]; ld.global.u32 %r305, [%rd194+8]; ld.global.u64 %rd266, [%rd194]; setp.eq.s64 %p99, %rd266, 11; @%p99 bra $L__BB23_79; $L__BB23_53: setp.eq.s64 %p100, %rd266, 0; @%p100 bra $L__BB23_77; add.s32 %r306, %r306, 2; bra.uni $L__BB23_78; $L__BB23_77: setp.eq.s16 %p101, %rs65, 0; mov.u32 %r309, -1; mov.u32 %r306, %r305; @%p101 bra $L__BB23_80; $L__BB23_78: mul.wide.u32 %rd195, %r306, 16; add.s64 %rd196, %rd17, %rd195; ld.global.u16 %rs65, [%rd196+12]; ld.global.u32 %r305, [%rd196+8]; ld.global.u64 %rd266, [%rd196]; setp.ne.s64 %p102, %rd266, 11; @%p102 bra $L__BB23_53; $L__BB23_79: setp.eq.s16 %p103, %rs65, 0; selp.b32 %r309, -1, %r305, %p103; $L__BB23_80: setp.eq.s32 %p104, %r73, %r72; @%p104 bra $L__BB23_82; bra.uni $L__BB23_81; $L__BB23_82: shl.b64 %rd198, %rd34, 4; add.s64 %rd267, %rd24, %rd198; bra.uni $L__BB23_83; $L__BB23_96: setp.gt.s32 %p127, %r92, %r91; selp.b32 %r238, -1, 0, %p127; add.s32 %r239, %r92, %r238; mad.lo.s32 %r240, %r239, %r90, %r315; add.s32 %r241, %r240, %r88; mul.wide.s32 %rd219, %r241, 16; add.s64 %rd220, %rd26, %rd219; ld.global.v4.f32 {%f2856, %f2857, %f2858, %f2859}, [%rd220]; add.s32 %r242, %r240, %r87; mul.wide.s32 %rd221, %r242, 16; add.s64 %rd222, %rd26, %rd221; ld.global.v4.f32 {%f2860, %f2861, %f2862, %f2863}, [%rd222]; ld.global.v4.f32 {%f2864, %f2865, %f2866, %f2867}, [%rd222+16]; add.s32 %r243, %r240, %r89; mul.wide.s32 %rd223, %r243, 16; add.s64 %rd224, %rd26, %rd223; ld.global.v4.f32 {%f2868, %f2869, %f2870, %f2871}, [%rd224]; $L__BB23_98: add.s32 %r103, %r92, 1; setp.eq.s32 %p128, %r103, %r91; @%p128 bra $L__BB23_100; bra.uni $L__BB23_99; $L__BB23_100: ld.global.v4.f32 {%f2872, %f2873, %f2874, %f2875}, [%rd102]; ld.global.v4.f32 {%f2876, %f2877, %f2878, %f2879}, [%rd103]; ld.global.v4.f32 {%f2880, %f2881, %f2882, %f2883}, [%rd103+16]; ld.global.v4.f32 {%f2884, %f2885, %f2886, %f2887}, [%rd104]; bra.uni $L__BB23_101; $L__BB23_99: setp.ge.s32 %p129, %r92, %r91; selp.b32 %r244, -1, 0, %p129; add.s32 %r245, %r103, %r244; mad.lo.s32 %r246, %r245, %r90, %r315; add.s32 %r247, %r246, %r88; mul.wide.s32 %rd225, %r247, 16; add.s64 %rd226, %rd26, %rd225; ld.global.v4.f32 {%f2872, %f2873, %f2874, %f2875}, [%rd226]; add.s32 %r248, %r246, %r87; mul.wide.s32 %rd227, %r248, 16; add.s64 %rd228, %rd26, %rd227; ld.global.v4.f32 {%f2876, %f2877, %f2878, %f2879}, [%rd228]; ld.global.v4.f32 {%f2880, %f2881, %f2882, %f2883}, [%rd228+16]; add.s32 %r249, %r246, %r89; mul.wide.s32 %rd229, %r249, 16; add.s64 %rd230, %rd26, %rd229; ld.global.v4.f32 {%f2884, %f2885, %f2886, %f2887}, [%rd230]; $L__BB23_101: cvt.rn.f32.s32 %f1483, %r92; sub.ftz.f32 %f1484, %f256, %f1483; sub.ftz.f32 %f1486, %f753, %f1484; mul.ftz.f32 %f1487, %f1484, %f2872; fma.rn.ftz.f32 %f2888, %f1486, %f2856, %f1487; mul.ftz.f32 %f1488, %f1484, %f2873; fma.rn.ftz.f32 %f2889, %f1486, %f2857, %f1488; mul.ftz.f32 %f1489, %f1484, %f2874; fma.rn.ftz.f32 %f2890, %f1486, %f2858, %f1489; mul.ftz.f32 %f1490, %f1484, %f2875; fma.rn.ftz.f32 %f388, %f1486, %f2859, %f1490; mul.ftz.f32 %f1491, %f1484, %f2876; fma.rn.ftz.f32 %f2892, %f1486, %f2860, %f1491; mul.ftz.f32 %f1492, %f1484, %f2877; fma.rn.ftz.f32 %f2893, %f1486, %f2861, %f1492; mul.ftz.f32 %f1493, %f1484, %f2878; fma.rn.ftz.f32 %f2894, %f1486, %f2862, %f1493; mul.ftz.f32 %f1494, %f1484, %f2879; fma.rn.ftz.f32 %f392, %f1486, %f2863, %f1494; mul.ftz.f32 %f1495, %f1484, %f2880; fma.rn.ftz.f32 %f2896, %f1486, %f2864, %f1495; mul.ftz.f32 %f1496, %f1484, %f2881; fma.rn.ftz.f32 %f2897, %f1486, %f2865, %f1496; mul.ftz.f32 %f1497, %f1484, %f2882; fma.rn.ftz.f32 %f2898, %f1486, %f2866, %f1497; mul.ftz.f32 %f1498, %f1484, %f2883; fma.rn.ftz.f32 %f396, %f1486, %f2867, %f1498; mul.ftz.f32 %f1499, %f1484, %f2884; fma.rn.ftz.f32 %f2900, %f1486, %f2868, %f1499; mul.ftz.f32 %f1500, %f1484, %f2885; fma.rn.ftz.f32 %f2901, %f1486, %f2869, %f1500; mul.ftz.f32 %f1501, %f1484, %f2886; fma.rn.ftz.f32 %f2902, %f1486, %f2870, %f1501; mul.ftz.f32 %f1502, %f1484, %f2887; fma.rn.ftz.f32 %f400, %f1486, %f2871, %f1502; $L__BB23_103: and.b32 %r250, %r86, 4; setp.eq.s32 %p130, %r250, 0; @%p130 bra $L__BB23_120; bra.uni $L__BB23_104; $L__BB23_120: mov.u16 %rs77, 0; add.ftz.f32 %f1866, %f2888, %f2892; add.ftz.f32 %f1867, %f1866, %f2896; add.ftz.f32 %f1868, %f2889, %f2893; add.ftz.f32 %f1869, %f1868, %f2897; add.ftz.f32 %f1870, %f2890, %f2894; add.ftz.f32 %f1871, %f1870, %f2898; add.ftz.f32 %f1872, %f1867, %f2900; add.ftz.f32 %f1873, %f1869, %f2901; add.ftz.f32 %f1874, %f1871, %f2902; mul.ftz.f32 %f1875, %f1872, 0f3E800000; mul.ftz.f32 %f1876, %f1873, 0f3E800000; mul.ftz.f32 %f1877, %f1874, 0f3E800000; sub.ftz.f32 %f1878, %f1875, %f3026; sub.ftz.f32 %f1879, %f1876, %f3025; sub.ftz.f32 %f1880, %f1877, %f3024; mul.ftz.f32 %f1881, %f3028, %f1879; fma.rn.ftz.f32 %f1882, %f3029, %f1878, %f1881; fma.rn.ftz.f32 %f1883, %f3027, %f1880, %f1882; div.approx.ftz.f32 %f491, %f1883, %f175; fma.rn.ftz.f32 %f1884, %f3029, %f491, %f3026; fma.rn.ftz.f32 %f1885, %f3028, %f491, %f3025; fma.rn.ftz.f32 %f1886, %f3027, %f491, %f3024; sub.ftz.f32 %f492, %f2888, %f1884; sub.ftz.f32 %f493, %f2889, %f1885; sub.ftz.f32 %f494, %f2890, %f1886; sub.ftz.f32 %f495, %f2892, %f1884; sub.ftz.f32 %f496, %f2893, %f1885; sub.ftz.f32 %f497, %f2894, %f1886; sub.ftz.f32 %f498, %f2896, %f1884; sub.ftz.f32 %f499, %f2897, %f1885; sub.ftz.f32 %f500, %f2898, %f1886; sub.ftz.f32 %f501, %f2900, %f1884; sub.ftz.f32 %f502, %f2901, %f1885; sub.ftz.f32 %f503, %f2902, %f1886; mov.u32 %r323, 0; mov.f32 %f2972, 0f00000000; mov.f32 %f2973, %f753; mov.u32 %r319, %r323; bra.uni $L__BB23_121; $L__BB23_165: add.s32 %r323, %r323, -1; mul.wide.s32 %rd233, %r323, 12; add.s64 %rd234, %rd1, %rd233; ld.local.f32 %f2972, [%rd234]; ld.local.f32 %f2973, [%rd234+4]; ld.local.u32 %r319, [%rd234+8]; $L__BB23_121: setp.gt.s32 %p154, %r319, 1; @%p154 bra $L__BB23_161; $L__BB23_122: mov.u32 %r324, %r319; cvt.rn.f32.s32 %f1889, %r324; mul.ftz.f32 %f519, %f1889, 0f3F000000; sub.ftz.f32 %f520, %f2973, %f2972; mul.ftz.f32 %f1890, %f520, 0f3EAAAAAB; mul.ftz.f32 %f1891, %f1890, 0f3F000000; fma.rn.ftz.f32 %f521, %f520, %f519, %f2972; add.ftz.f32 %f1892, %f519, 0f3F000000; fma.rn.ftz.f32 %f522, %f520, %f1892, %f2972; mov.f32 %f1893, 0f3F800000; sub.ftz.f32 %f1894, %f1893, %f521; mul.ftz.f32 %f1895, %f521, %f1894; mul.ftz.f32 %f1896, %f1894, %f1895; mul.ftz.f32 %f1897, %f521, %f521; mul.ftz.f32 %f1898, %f521, 0f40400000; add.ftz.f32 %f1899, %f1898, 0fC0A00000; fma.rn.ftz.f32 %f1900, %f1897, %f1899, 0f40000000; mul.ftz.f32 %f1901, %f1894, %f1894; mul.ftz.f32 %f1902, %f1894, 0f40400000; add.ftz.f32 %f1903, %f1902, 0fC0A00000; fma.rn.ftz.f32 %f1904, %f1901, %f1903, 0f40000000; mul.ftz.f32 %f1905, %f521, %f1895; mul.ftz.f32 %f1906, %f492, %f1896; mul.ftz.f32 %f1907, %f493, %f1896; mul.ftz.f32 %f1908, %f494, %f1896; mul.ftz.f32 %f1909, %f388, %f1896; mul.ftz.f32 %f1910, %f495, %f1900; mul.ftz.f32 %f1911, %f496, %f1900; mul.ftz.f32 %f1912, %f497, %f1900; mul.ftz.f32 %f1913, %f392, %f1900; sub.ftz.f32 %f1914, %f1910, %f1906; sub.ftz.f32 %f1915, %f1911, %f1907; sub.ftz.f32 %f1916, %f1912, %f1908; sub.ftz.f32 %f1917, %f1913, %f1909; fma.rn.ftz.f32 %f1918, %f498, %f1904, %f1914; fma.rn.ftz.f32 %f1919, %f499, %f1904, %f1915; fma.rn.ftz.f32 %f1920, %f500, %f1904, %f1916; fma.rn.ftz.f32 %f1921, %f396, %f1904, %f1917; mul.ftz.f32 %f1922, %f501, %f1905; mul.ftz.f32 %f1923, %f502, %f1905; mul.ftz.f32 %f1924, %f503, %f1905; mul.ftz.f32 %f1925, %f400, %f1905; sub.ftz.f32 %f1926, %f1918, %f1922; sub.ftz.f32 %f1927, %f1919, %f1923; sub.ftz.f32 %f1928, %f1920, %f1924; sub.ftz.f32 %f1929, %f1921, %f1925; mul.ftz.f32 %f1930, %f1926, 0f3F000000; mul.ftz.f32 %f1931, %f1927, 0f3F000000; mul.ftz.f32 %f1932, %f1928, 0f3F000000; mul.ftz.f32 %f1933, %f1929, 0f3F000000; add.ftz.f32 %f1934, %f1894, %f1894; mul.ftz.f32 %f1935, %f521, %f1934; sub.ftz.f32 %f1936, %f1935, %f1901; add.ftz.f32 %f1937, %f521, %f521; mul.ftz.f32 %f1938, %f1937, %f1899; fma.rn.ftz.f32 %f1939, %f521, %f1898, %f1938; add.ftz.f32 %f1940, %f1898, 0f40000000; mul.ftz.f32 %f1941, %f1934, %f1940; mul.ftz.f32 %f1942, %f1894, %f1902; sub.ftz.f32 %f1943, %f1941, %f1942; mul.ftz.f32 %f1944, %f1894, 0fC0000000; fma.rn.ftz.f32 %f1945, %f521, %f1944, %f1897; mul.ftz.f32 %f1946, %f495, %f1939; mul.ftz.f32 %f1947, %f496, %f1939; mul.ftz.f32 %f1948, %f497, %f1939; mul.ftz.f32 %f1949, %f392, %f1939; fma.rn.ftz.f32 %f1950, %f492, %f1936, %f1946; fma.rn.ftz.f32 %f1951, %f493, %f1936, %f1947; fma.rn.ftz.f32 %f1952, %f494, %f1936, %f1948; fma.rn.ftz.f32 %f1953, %f388, %f1936, %f1949; fma.rn.ftz.f32 %f1954, %f498, %f1943, %f1950; fma.rn.ftz.f32 %f1955, %f499, %f1943, %f1951; fma.rn.ftz.f32 %f1956, %f500, %f1943, %f1952; fma.rn.ftz.f32 %f1957, %f396, %f1943, %f1953; fma.rn.ftz.f32 %f1958, %f501, %f1945, %f1954; fma.rn.ftz.f32 %f1959, %f502, %f1945, %f1955; fma.rn.ftz.f32 %f1960, %f503, %f1945, %f1956; fma.rn.ftz.f32 %f1961, %f400, %f1945, %f1957; mul.ftz.f32 %f1962, %f1958, 0f3F000000; mul.ftz.f32 %f1963, %f1959, 0f3F000000; mul.ftz.f32 %f1964, %f1960, 0f3F000000; mul.ftz.f32 %f1965, %f1961, 0f3F000000; mul.ftz.f32 %f523, %f1891, %f1962; mul.ftz.f32 %f524, %f1891, %f1963; mul.ftz.f32 %f525, %f1891, %f1964; sub.ftz.f32 %f1966, %f1893, %f522; mul.ftz.f32 %f1967, %f522, %f1966; mul.ftz.f32 %f1968, %f1966, %f1967; mul.ftz.f32 %f1969, %f522, %f522; mul.ftz.f32 %f1970, %f522, 0f40400000; add.ftz.f32 %f1971, %f1970, 0fC0A00000; fma.rn.ftz.f32 %f1972, %f1969, %f1971, 0f40000000; mul.ftz.f32 %f1973, %f1966, %f1966; mul.ftz.f32 %f1974, %f1966, 0f40400000; add.ftz.f32 %f1975, %f1974, 0fC0A00000; fma.rn.ftz.f32 %f1976, %f1973, %f1975, 0f40000000; mul.ftz.f32 %f1977, %f522, %f1967; mul.ftz.f32 %f1978, %f492, %f1968; mul.ftz.f32 %f1979, %f493, %f1968; mul.ftz.f32 %f1980, %f494, %f1968; mul.ftz.f32 %f1981, %f388, %f1968; mul.ftz.f32 %f1982, %f495, %f1972; mul.ftz.f32 %f1983, %f496, %f1972; mul.ftz.f32 %f1984, %f497, %f1972; mul.ftz.f32 %f1985, %f392, %f1972; sub.ftz.f32 %f1986, %f1982, %f1978; sub.ftz.f32 %f1987, %f1983, %f1979; sub.ftz.f32 %f1988, %f1984, %f1980; sub.ftz.f32 %f1989, %f1985, %f1981; fma.rn.ftz.f32 %f1990, %f498, %f1976, %f1986; fma.rn.ftz.f32 %f1991, %f499, %f1976, %f1987; fma.rn.ftz.f32 %f1992, %f500, %f1976, %f1988; fma.rn.ftz.f32 %f1993, %f396, %f1976, %f1989; mul.ftz.f32 %f1994, %f501, %f1977; mul.ftz.f32 %f1995, %f502, %f1977; mul.ftz.f32 %f1996, %f503, %f1977; mul.ftz.f32 %f1997, %f400, %f1977; sub.ftz.f32 %f1998, %f1990, %f1994; sub.ftz.f32 %f1999, %f1991, %f1995; sub.ftz.f32 %f2000, %f1992, %f1996; sub.ftz.f32 %f2001, %f1993, %f1997; mul.ftz.f32 %f526, %f1998, 0f3F000000; mul.ftz.f32 %f527, %f1999, 0f3F000000; mul.ftz.f32 %f528, %f2000, 0f3F000000; mul.ftz.f32 %f2002, %f2001, 0f3F000000; add.ftz.f32 %f2003, %f1966, %f1966; mul.ftz.f32 %f2004, %f522, %f2003; sub.ftz.f32 %f2005, %f2004, %f1973; add.ftz.f32 %f2006, %f522, %f522; mul.ftz.f32 %f2007, %f2006, %f1971; fma.rn.ftz.f32 %f2008, %f522, %f1970, %f2007; add.ftz.f32 %f2009, %f1970, 0f40000000; mul.ftz.f32 %f2010, %f2003, %f2009; mul.ftz.f32 %f2011, %f1966, %f1974; sub.ftz.f32 %f2012, %f2010, %f2011; mul.ftz.f32 %f2013, %f1966, 0fC0000000; fma.rn.ftz.f32 %f2014, %f522, %f2013, %f1969; mul.ftz.f32 %f2015, %f495, %f2008; mul.ftz.f32 %f2016, %f496, %f2008; mul.ftz.f32 %f2017, %f497, %f2008; mul.ftz.f32 %f2018, %f392, %f2008; fma.rn.ftz.f32 %f2019, %f492, %f2005, %f2015; fma.rn.ftz.f32 %f2020, %f493, %f2005, %f2016; fma.rn.ftz.f32 %f2021, %f494, %f2005, %f2017; fma.rn.ftz.f32 %f2022, %f388, %f2005, %f2018; fma.rn.ftz.f32 %f2023, %f498, %f2012, %f2019; fma.rn.ftz.f32 %f2024, %f499, %f2012, %f2020; fma.rn.ftz.f32 %f2025, %f500, %f2012, %f2021; fma.rn.ftz.f32 %f2026, %f396, %f2012, %f2022; fma.rn.ftz.f32 %f2027, %f501, %f2014, %f2023; fma.rn.ftz.f32 %f2028, %f502, %f2014, %f2024; fma.rn.ftz.f32 %f2029, %f503, %f2014, %f2025; fma.rn.ftz.f32 %f2030, %f400, %f2014, %f2026; mul.ftz.f32 %f2031, %f2027, 0f3F000000; mul.ftz.f32 %f2032, %f2028, 0f3F000000; mul.ftz.f32 %f2033, %f2029, 0f3F000000; mul.ftz.f32 %f2034, %f2030, 0f3F000000; mul.ftz.f32 %f529, %f1891, %f2031; mul.ftz.f32 %f530, %f1891, %f2032; mul.ftz.f32 %f531, %f1891, %f2033; mul.ftz.f32 %f2035, %f1891, %f2034; fma.rn.ftz.f32 %f2036, %f1891, %f1965, %f1933; sub.ftz.f32 %f2037, %f2002, %f2035; sub.ftz.f32 %f2038, %f526, %f1930; sub.ftz.f32 %f2039, %f527, %f1931; sub.ftz.f32 %f2040, %f528, %f1932; mul.ftz.f32 %f2041, %f524, %f2040; mul.ftz.f32 %f2042, %f525, %f2039; sub.ftz.f32 %f2043, %f2041, %f2042; mul.ftz.f32 %f2044, %f525, %f2038; mul.ftz.f32 %f2045, %f523, %f2040; sub.ftz.f32 %f2046, %f2044, %f2045; mul.ftz.f32 %f2047, %f523, %f2039; mul.ftz.f32 %f2048, %f524, %f2038; sub.ftz.f32 %f2049, %f2047, %f2048; mul.ftz.f32 %f2050, %f2046, %f2046; fma.rn.ftz.f32 %f2051, %f2043, %f2043, %f2050; fma.rn.ftz.f32 %f2052, %f2049, %f2049, %f2051; mul.ftz.f32 %f2053, %f2039, %f2039; fma.rn.ftz.f32 %f2054, %f2038, %f2038, %f2053; fma.rn.ftz.f32 %f2055, %f2040, %f2040, %f2054; div.approx.ftz.f32 %f2056, %f2052, %f2055; mul.ftz.f32 %f2057, %f530, %f2040; mul.ftz.f32 %f2058, %f531, %f2039; sub.ftz.f32 %f2059, %f2057, %f2058; mul.ftz.f32 %f2060, %f531, %f2038; mul.ftz.f32 %f2061, %f529, %f2040; sub.ftz.f32 %f2062, %f2060, %f2061; mul.ftz.f32 %f2063, %f529, %f2039; mul.ftz.f32 %f2064, %f530, %f2038; sub.ftz.f32 %f2065, %f2063, %f2064; mul.ftz.f32 %f2066, %f2062, %f2062; fma.rn.ftz.f32 %f2067, %f2059, %f2059, %f2066; fma.rn.ftz.f32 %f2068, %f2065, %f2065, %f2067; div.approx.ftz.f32 %f2069, %f2068, %f2055; max.ftz.f32 %f2070, %f2056, %f2069; sqrt.approx.ftz.f32 %f2071, %f2070; max.ftz.f32 %f2072, %f1933, %f2036; max.ftz.f32 %f2073, %f2037, %f2002; max.ftz.f32 %f2074, %f2072, %f2073; add.ftz.f32 %f2075, %f2071, %f2074; min.ftz.f32 %f2076, %f1933, %f2036; min.ftz.f32 %f2077, %f2037, %f2002; min.ftz.f32 %f2078, %f2076, %f2077; sub.ftz.f32 %f532, %f2078, %f2071; mul.ftz.f32 %f2079, %f2075, 0f3F800002; rsqrt.approx.ftz.f32 %f533, %f2055; mul.ftz.f32 %f2080, %f2038, %f533; mul.ftz.f32 %f2081, %f2039, %f533; mul.ftz.f32 %f2082, %f2040, %f533; mul.ftz.f32 %f534, %f1926, 0fBF000000; mul.ftz.f32 %f535, %f1927, 0fBF000000; mul.ftz.f32 %f536, %f1928, 0fBF000000; mul.ftz.f32 %f2083, %f3029, %f534; mul.ftz.f32 %f2084, %f3028, %f1931; sub.ftz.f32 %f2085, %f2083, %f2084; mul.ftz.f32 %f2086, %f3027, %f1932; sub.ftz.f32 %f2087, %f2085, %f2086; mul.ftz.f32 %f2088, %f1931, %f1931; fma.rn.ftz.f32 %f2089, %f1930, %f1930, %f2088; fma.rn.ftz.f32 %f2090, %f1932, %f1932, %f2089; mul.ftz.f32 %f2091, %f3028, %f2081; fma.rn.ftz.f32 %f2092, %f3029, %f2080, %f2091; fma.rn.ftz.f32 %f537, %f3027, %f2082, %f2092; mul.ftz.f32 %f2093, %f2080, %f534; mul.ftz.f32 %f2094, %f2081, %f1931; sub.ftz.f32 %f2095, %f2093, %f2094; mul.ftz.f32 %f2096, %f2082, %f1932; sub.ftz.f32 %f538, %f2095, %f2096; mul.ftz.f32 %f539, %f537, %f537; sub.ftz.f32 %f540, %f175, %f539; mul.ftz.f32 %f2097, %f537, %f538; sub.ftz.f32 %f2098, %f2087, %f2097; add.ftz.f32 %f541, %f2098, %f2098; mul.ftz.f32 %f2099, %f538, %f538; sub.ftz.f32 %f542, %f2090, %f2099; mul.ftz.f32 %f2100, %f2079, %f2079; sub.ftz.f32 %f543, %f542, %f2100; mul.ftz.f32 %f544, %f541, %f541; mul.ftz.f32 %f545, %f540, 0f40800000; mul.ftz.f32 %f2101, %f545, %f543; sub.ftz.f32 %f546, %f544, %f2101; setp.ltu.ftz.f32 %p156, %f546, 0f00000000; mov.pred %p247, -1; mov.f32 %f2946, 0fFF7FFFFF; mov.f32 %f2945, 0f7F7FFFFF; @%p156 bra $L__BB23_126; abs.ftz.f32 %f2102, %f175; abs.ftz.f32 %f2103, %f539; max.ftz.f32 %f2104, %f2102, %f2103; mul.ftz.f32 %f2105, %f2104, 0f36000000; abs.ftz.f32 %f2106, %f540; setp.lt.ftz.f32 %p157, %f2106, %f2105; @%p157 bra $L__BB23_125; bra.uni $L__BB23_124; $L__BB23_125: setp.gtu.ftz.f32 %p247, %f543, 0f00000000; mov.f32 %f2946, 0f7F7FFFFF; mov.f32 %f2945, 0fFF7FFFFF; bra.uni $L__BB23_126; $L__BB23_124: sqrt.approx.ftz.f32 %f2107, %f546; add.ftz.f32 %f2108, %f540, %f540; rcp.approx.ftz.f32 %f2109, %f2108; neg.ftz.f32 %f2110, %f541; sub.ftz.f32 %f2111, %f2110, %f2107; mul.ftz.f32 %f2945, %f2109, %f2111; sub.ftz.f32 %f2112, %f2107, %f541; mul.ftz.f32 %f2946, %f2109, %f2112; fma.rn.ftz.f32 %f2113, %f537, %f2945, %f538; mul.ftz.f32 %f3007, %f533, %f2113; fma.rn.ftz.f32 %f2114, %f537, %f2946, %f538; mul.ftz.f32 %f3006, %f533, %f2114; mov.pred %p247, 0; $L__BB23_126: @%p247 bra $L__BB23_160; sub.ftz.f32 %f2727, %f10, %f491; sub.ftz.f32 %f2117, %f2992, %f491; min.ftz.f32 %f2118, %f2117, %f2946; mul.ftz.f32 %f2119, %f524, %f535; fma.rn.ftz.f32 %f2120, %f523, %f534, %f2119; fma.rn.ftz.f32 %f2121, %f525, %f536, %f2120; mul.ftz.f32 %f2122, %f3028, %f524; fma.rn.ftz.f32 %f2123, %f3029, %f523, %f2122; fma.rn.ftz.f32 %f2124, %f3027, %f525, %f2123; abs.ftz.f32 %f2125, %f2124; setp.lt.ftz.f32 %p159, %f2125, 0f219392EF; neg.ftz.f32 %f2126, %f2121; div.approx.ftz.f32 %f2127, %f2126, %f2124; setp.lt.ftz.f32 %p160, %f2124, 0f00000000; or.pred %p161, %p160, %p159; selp.f32 %f2128, 0fFF7FFFFF, %f2127, %p161; setp.gt.ftz.f32 %p162, %f2124, 0f00000000; or.pred %p163, %p162, %p159; selp.f32 %f2129, 0f7F7FFFFF, %f2127, %p163; max.ftz.f32 %f2130, %f2727, %f2945; max.ftz.f32 %f2131, %f2130, %f2128; min.ftz.f32 %f2132, %f2118, %f2129; mul.ftz.f32 %f2133, %f527, %f530; fma.rn.ftz.f32 %f2134, %f526, %f529, %f2133; fma.rn.ftz.f32 %f2135, %f528, %f531, %f2134; mul.ftz.f32 %f2136, %f3028, %f530; neg.ftz.f32 %f2137, %f2136; mul.ftz.f32 %f2138, %f3029, %f529; sub.ftz.f32 %f2139, %f2137, %f2138; mul.ftz.f32 %f2140, %f3027, %f531; sub.ftz.f32 %f2141, %f2139, %f2140; abs.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p164, %f2142, 0f219392EF; neg.ftz.f32 %f2143, %f2135; div.approx.ftz.f32 %f2144, %f2143, %f2141; setp.lt.ftz.f32 %p165, %f2141, 0f00000000; or.pred %p166, %p165, %p164; selp.f32 %f2145, 0fFF7FFFFF, %f2144, %p166; setp.gt.ftz.f32 %p167, %f2141, 0f00000000; or.pred %p168, %p167, %p164; selp.f32 %f2146, 0f7F7FFFFF, %f2144, %p168; max.ftz.f32 %f2949, %f2131, %f2145; min.ftz.f32 %f2958, %f2132, %f2146; setp.gtu.ftz.f32 %p169, %f2949, %f2958; @%p169 bra $L__BB23_160; mul.ftz.f32 %f2732, %f541, %f541; sub.ftz.f32 %f2731, %f2973, %f2972; cvt.rn.f32.s32 %f2730, %r324; mul.ftz.f32 %f2729, %f2730, 0f3F000000; mov.f32 %f2728, 0f3F800000; mul.ftz.f32 %f2149, %f532, 0f3F7FFFFC; mov.f32 %f2150, 0f00000000; max.ftz.f32 %f2151, %f2150, %f2149; max.ftz.f32 %f2152, %f3007, %f2150; min.ftz.f32 %f2154, %f2152, %f2728; max.ftz.f32 %f2155, %f3006, %f2150; min.ftz.f32 %f2156, %f2155, %f2728; add.ftz.f32 %f2157, %f2729, %f2154; mul.ftz.f32 %f2158, %f2157, 0f3EAAAAAB; fma.rn.ftz.f32 %f3007, %f2731, %f2158, %f2972; add.ftz.f32 %f2159, %f2729, %f2156; mul.ftz.f32 %f2160, %f2159, 0f3EAAAAAB; fma.rn.ftz.f32 %f3006, %f2731, %f2160, %f2972; mul.ftz.f32 %f2161, %f2151, %f2151; sub.ftz.f32 %f2162, %f542, %f2161; mul.ftz.f32 %f2163, %f545, %f2162; sub.ftz.f32 %f559, %f2732, %f2163; setp.ltu.ftz.f32 %p170, %f559, 0f00000000; mov.f32 %f2948, 0fFF7FFFFF; mov.f32 %f2947, 0f7F7FFFFF; @%p170 bra $L__BB23_131; abs.ftz.f32 %f2166, %f175; abs.ftz.f32 %f2167, %f539; max.ftz.f32 %f2168, %f2166, %f2167; mul.ftz.f32 %f2169, %f2168, 0f36000000; abs.ftz.f32 %f2170, %f540; setp.lt.ftz.f32 %p171, %f2170, %f2169; mov.f32 %f2948, 0f7F7FFFFF; mov.f32 %f2947, 0fFF7FFFFF; @%p171 bra $L__BB23_131; sqrt.approx.ftz.f32 %f2171, %f559; add.ftz.f32 %f2172, %f540, %f540; rcp.approx.ftz.f32 %f2173, %f2172; neg.ftz.f32 %f2174, %f541; sub.ftz.f32 %f2175, %f2174, %f2171; mul.ftz.f32 %f2947, %f2173, %f2175; sub.ftz.f32 %f2176, %f2171, %f541; mul.ftz.f32 %f2948, %f2173, %f2176; $L__BB23_131: min.ftz.f32 %f564, %f2958, %f2947; setp.gtu.ftz.f32 %p172, %f2949, %f564; max.ftz.f32 %f565, %f2949, %f2948; setp.gtu.ftz.f32 %p173, %f565, %f2958; and.pred %p174, %p172, %p173; @%p174 bra $L__BB23_160; mov.u16 %rs73, 0; @%p172 bra $L__BB23_145; setp.lt.s32 %p176, %r323, 4; mov.u16 %rs73, 1; @%p176 bra $L__BB23_145; sqrt.approx.ftz.f32 %f2177, %f175; min.ftz.f32 %f2178, %f498, %f501; min.ftz.f32 %f2179, %f492, %f495; min.ftz.f32 %f2180, %f2179, %f2178; min.ftz.f32 %f2181, %f499, %f502; min.ftz.f32 %f2182, %f493, %f496; min.ftz.f32 %f2183, %f2182, %f2181; min.ftz.f32 %f2184, %f500, %f503; min.ftz.f32 %f2185, %f494, %f497; min.ftz.f32 %f2186, %f2185, %f2184; min.ftz.f32 %f2187, %f396, %f400; min.ftz.f32 %f2188, %f388, %f392; min.ftz.f32 %f2189, %f2188, %f2187; max.ftz.f32 %f2190, %f498, %f501; max.ftz.f32 %f2191, %f2179, %f2190; max.ftz.f32 %f2192, %f499, %f502; max.ftz.f32 %f2193, %f2182, %f2192; max.ftz.f32 %f2194, %f500, %f503; max.ftz.f32 %f2195, %f2185, %f2194; max.ftz.f32 %f2196, %f396, %f400; max.ftz.f32 %f2197, %f2188, %f2196; abs.ftz.f32 %f2198, %f2180; abs.ftz.f32 %f2199, %f2183; abs.ftz.f32 %f2200, %f2186; abs.ftz.f32 %f2201, %f2189; abs.ftz.f32 %f2202, %f2191; abs.ftz.f32 %f2203, %f2193; abs.ftz.f32 %f2204, %f2195; abs.ftz.f32 %f2205, %f2197; max.ftz.f32 %f2206, %f2198, %f2202; max.ftz.f32 %f2207, %f2199, %f2203; max.ftz.f32 %f2208, %f2200, %f2204; max.ftz.f32 %f2209, %f2201, %f2205; max.ftz.f32 %f2210, %f2208, %f2209; max.ftz.f32 %f2211, %f2207, %f2210; max.ftz.f32 %f2212, %f2206, %f2211; mul.ftz.f32 %f566, %f2212, 0f36000000; mov.u32 %r321, 0; mov.f32 %f2950, %f3007; $L__BB23_135: mul.ftz.f32 %f2744, %f2177, 0f36000000; mov.f32 %f2738, 0f3F800000; sub.ftz.f32 %f2215, %f2738, %f2950; mul.ftz.f32 %f2216, %f2950, %f2215; mul.ftz.f32 %f2217, %f2215, %f2216; mul.ftz.f32 %f2218, %f2950, 0f40400000; add.ftz.f32 %f2219, %f2218, 0fC0A00000; mul.ftz.f32 %f2220, %f2950, %f2950; fma.rn.ftz.f32 %f2221, %f2220, %f2219, 0f40000000; mov.f32 %f2222, 0f40000000; mul.ftz.f32 %f2223, %f2215, %f2215; mul.ftz.f32 %f2224, %f2215, 0f40400000; add.ftz.f32 %f2225, %f2224, 0fC0A00000; fma.rn.ftz.f32 %f2226, %f2223, %f2225, 0f40000000; mul.ftz.f32 %f2227, %f2950, %f2216; mul.ftz.f32 %f2228, %f492, %f2217; mul.ftz.f32 %f2229, %f493, %f2217; mul.ftz.f32 %f2230, %f494, %f2217; mul.ftz.f32 %f2231, %f388, %f2217; mul.ftz.f32 %f2232, %f495, %f2221; mul.ftz.f32 %f2233, %f496, %f2221; mul.ftz.f32 %f2234, %f497, %f2221; mul.ftz.f32 %f2235, %f392, %f2221; sub.ftz.f32 %f2236, %f2232, %f2228; sub.ftz.f32 %f2237, %f2233, %f2229; sub.ftz.f32 %f2238, %f2234, %f2230; sub.ftz.f32 %f2239, %f2235, %f2231; fma.rn.ftz.f32 %f2240, %f498, %f2226, %f2236; fma.rn.ftz.f32 %f2241, %f499, %f2226, %f2237; fma.rn.ftz.f32 %f2242, %f500, %f2226, %f2238; fma.rn.ftz.f32 %f2243, %f396, %f2226, %f2239; mul.ftz.f32 %f2244, %f501, %f2227; mul.ftz.f32 %f2245, %f502, %f2227; mul.ftz.f32 %f2246, %f503, %f2227; mul.ftz.f32 %f2247, %f400, %f2227; sub.ftz.f32 %f2248, %f2240, %f2244; sub.ftz.f32 %f2249, %f2241, %f2245; sub.ftz.f32 %f2250, %f2242, %f2246; sub.ftz.f32 %f2251, %f2243, %f2247; mul.ftz.f32 %f2252, %f2248, 0f3F000000; mul.ftz.f32 %f2253, %f2249, 0f3F000000; mul.ftz.f32 %f2254, %f2250, 0f3F000000; mul.ftz.f32 %f571, %f2251, 0f3F000000; add.ftz.f32 %f2255, %f2215, %f2215; mul.ftz.f32 %f2256, %f2950, %f2255; sub.ftz.f32 %f2257, %f2256, %f2223; add.ftz.f32 %f2258, %f2950, %f2950; mul.ftz.f32 %f2259, %f2258, %f2219; fma.rn.ftz.f32 %f2260, %f2950, %f2218, %f2259; add.ftz.f32 %f2261, %f2218, 0f40000000; mul.ftz.f32 %f2262, %f2255, %f2261; mul.ftz.f32 %f2263, %f2215, %f2224; sub.ftz.f32 %f2264, %f2262, %f2263; mul.ftz.f32 %f2265, %f2215, 0fC0000000; fma.rn.ftz.f32 %f2266, %f2950, %f2265, %f2220; mul.ftz.f32 %f2267, %f495, %f2260; mul.ftz.f32 %f2268, %f496, %f2260; mul.ftz.f32 %f2269, %f497, %f2260; mul.ftz.f32 %f2270, %f392, %f2260; fma.rn.ftz.f32 %f2271, %f492, %f2257, %f2267; fma.rn.ftz.f32 %f2272, %f493, %f2257, %f2268; fma.rn.ftz.f32 %f2273, %f494, %f2257, %f2269; fma.rn.ftz.f32 %f2274, %f388, %f2257, %f2270; fma.rn.ftz.f32 %f2275, %f498, %f2264, %f2271; fma.rn.ftz.f32 %f2276, %f499, %f2264, %f2272; fma.rn.ftz.f32 %f2277, %f500, %f2264, %f2273; fma.rn.ftz.f32 %f2278, %f396, %f2264, %f2274; fma.rn.ftz.f32 %f2279, %f501, %f2266, %f2275; fma.rn.ftz.f32 %f2280, %f502, %f2266, %f2276; fma.rn.ftz.f32 %f2281, %f503, %f2266, %f2277; fma.rn.ftz.f32 %f2282, %f400, %f2266, %f2278; mul.ftz.f32 %f572, %f2279, 0f3F000000; mul.ftz.f32 %f573, %f2280, 0f3F000000; mul.ftz.f32 %f574, %f2281, 0f3F000000; mul.ftz.f32 %f575, %f2282, 0f3F000000; sub.ftz.f32 %f2283, %f2222, %f2218; mul.ftz.f32 %f2284, %f2950, 0f41100000; add.ftz.f32 %f2285, %f2284, 0fC0A00000; mov.f32 %f2286, 0f40800000; sub.ftz.f32 %f2287, %f2286, %f2284; add.ftz.f32 %f2288, %f2218, 0fBF800000; mul.ftz.f32 %f2289, %f495, %f2285; mul.ftz.f32 %f2290, %f496, %f2285; mul.ftz.f32 %f2291, %f497, %f2285; fma.rn.ftz.f32 %f2292, %f492, %f2283, %f2289; fma.rn.ftz.f32 %f2293, %f493, %f2283, %f2290; fma.rn.ftz.f32 %f2294, %f494, %f2283, %f2291; fma.rn.ftz.f32 %f2295, %f498, %f2287, %f2292; fma.rn.ftz.f32 %f2296, %f499, %f2287, %f2293; fma.rn.ftz.f32 %f2297, %f500, %f2287, %f2294; fma.rn.ftz.f32 %f2298, %f501, %f2288, %f2295; fma.rn.ftz.f32 %f2299, %f502, %f2288, %f2296; fma.rn.ftz.f32 %f2300, %f503, %f2288, %f2297; mul.ftz.f32 %f2301, %f3029, %f2949; sub.ftz.f32 %f576, %f2301, %f2252; mul.ftz.f32 %f2302, %f3028, %f2949; sub.ftz.f32 %f577, %f2302, %f2253; mul.ftz.f32 %f2303, %f3027, %f2949; sub.ftz.f32 %f578, %f2303, %f2254; mul.ftz.f32 %f2304, %f577, %f577; fma.rn.ftz.f32 %f2305, %f576, %f576, %f2304; fma.rn.ftz.f32 %f2306, %f578, %f578, %f2305; sqrt.approx.ftz.f32 %f579, %f2306; mul.ftz.f32 %f2307, %f2744, %f2949; max.ftz.f32 %f580, %f2307, %f566; mul.ftz.f32 %f2308, %f2279, 0fBF000000; mul.ftz.f32 %f2309, %f573, %f573; fma.rn.ftz.f32 %f2310, %f572, %f572, %f2309; fma.rn.ftz.f32 %f2311, %f574, %f574, %f2310; sqrt.approx.ftz.f32 %f2312, %f2311; div.approx.ftz.f32 %f2313, %f572, %f2312; div.approx.ftz.f32 %f2314, %f573, %f2312; div.approx.ftz.f32 %f2315, %f574, %f2312; mul.ftz.f32 %f2316, %f2299, %f573; fma.rn.ftz.f32 %f2317, %f2298, %f572, %f2316; fma.rn.ftz.f32 %f2318, %f2300, %f574, %f2317; mul.ftz.f32 %f2319, %f2298, %f2311; mul.ftz.f32 %f2320, %f2299, %f2311; mul.ftz.f32 %f2321, %f2300, %f2311; mul.ftz.f32 %f2322, %f572, %f2318; mul.ftz.f32 %f2323, %f573, %f2318; mul.ftz.f32 %f2324, %f574, %f2318; sub.ftz.f32 %f2325, %f2319, %f2322; sub.ftz.f32 %f2326, %f2320, %f2323; sub.ftz.f32 %f2327, %f2321, %f2324; mul.ftz.f32 %f2328, %f2311, %f2312; div.approx.ftz.f32 %f2329, %f2325, %f2328; div.approx.ftz.f32 %f2330, %f2326, %f2328; div.approx.ftz.f32 %f2331, %f2327, %f2328; div.approx.ftz.f32 %f2332, %f566, %f2312; mul.ftz.f32 %f2333, %f577, %f2314; fma.rn.ftz.f32 %f2334, %f576, %f2313, %f2333; fma.rn.ftz.f32 %f581, %f578, %f2315, %f2334; fma.rn.ftz.f32 %f2335, %f566, %f579, %f580; add.ftz.f32 %f2336, %f579, 0f3F800000; fma.rn.ftz.f32 %f582, %f2336, %f2332, %f2335; mul.ftz.f32 %f2337, %f2313, %f2308; mul.ftz.f32 %f2338, %f2314, %f573; sub.ftz.f32 %f2339, %f2337, %f2338; mul.ftz.f32 %f2340, %f2315, %f574; sub.ftz.f32 %f2341, %f2339, %f2340; mul.ftz.f32 %f2342, %f577, %f2330; fma.rn.ftz.f32 %f2343, %f576, %f2329, %f2342; fma.rn.ftz.f32 %f2344, %f578, %f2331, %f2343; add.ftz.f32 %f583, %f2341, %f2344; mul.ftz.f32 %f2345, %f3028, %f2314; fma.rn.ftz.f32 %f2346, %f3029, %f2313, %f2345; fma.rn.ftz.f32 %f584, %f3027, %f2315, %f2346; mul.ftz.f32 %f2347, %f581, %f581; sub.ftz.f32 %f2348, %f2306, %f2347; mul.ftz.f32 %f2349, %f576, %f2308; mul.ftz.f32 %f2350, %f577, %f573; sub.ftz.f32 %f2351, %f2349, %f2350; mul.ftz.f32 %f2352, %f578, %f574; sub.ftz.f32 %f585, %f2351, %f2352; mul.ftz.f32 %f2353, %f3028, %f577; fma.rn.ftz.f32 %f2354, %f3029, %f576, %f2353; fma.rn.ftz.f32 %f586, %f3027, %f578, %f2354; setp.leu.ftz.f32 %p177, %f2348, 0f00000000; mov.f32 %f2951, 0f00000000; sqrt.approx.ftz.f32 %f587, %f2348; @%p177 bra $L__BB23_137; rcp.approx.ftz.f32 %f2951, %f587; $L__BB23_137: max.ftz.f32 %f2742, %f396, %f400; min.ftz.f32 %f2741, %f388, %f392; max.ftz.f32 %f2740, %f2741, %f2742; mul.ftz.f32 %f2739, %f2740, 0f36000000; add.ftz.f32 %f2355, %f580, %f582; add.ftz.f32 %f590, %f2739, %f2355; mul.ftz.f32 %f2356, %f581, %f583; sub.ftz.f32 %f2357, %f585, %f2356; mul.ftz.f32 %f2358, %f2357, %f2951; sub.ftz.f32 %f2359, %f2358, %f575; mul.ftz.f32 %f2360, %f581, %f584; sub.ftz.f32 %f2361, %f586, %f2360; mul.ftz.f32 %f2362, %f2361, %f2951; mul.ftz.f32 %f2363, %f583, %f2362; mul.ftz.f32 %f2364, %f584, %f2359; sub.ftz.f32 %f2365, %f2363, %f2364; rcp.approx.ftz.f32 %f2366, %f2365; mul.ftz.f32 %f2367, %f581, %f2362; sub.ftz.f32 %f591, %f587, %f571; mul.ftz.f32 %f2368, %f584, %f591; sub.ftz.f32 %f2369, %f2367, %f2368; mul.ftz.f32 %f2370, %f2369, %f2366; sub.ftz.f32 %f2950, %f2950, %f2370; mul.ftz.f32 %f2371, %f581, %f2359; mul.ftz.f32 %f2372, %f583, %f591; sub.ftz.f32 %f2373, %f2372, %f2371; mul.ftz.f32 %f2374, %f2373, %f2366; sub.ftz.f32 %f2949, %f2949, %f2374; abs.ftz.f32 %f2375, %f581; setp.geu.ftz.f32 %p178, %f2375, %f582; @%p178 bra $L__BB23_139; abs.ftz.f32 %f2376, %f591; setp.lt.ftz.f32 %p179, %f2376, %f590; @%p179 bra $L__BB23_140; $L__BB23_139: add.s32 %r321, %r321, 1; setp.lt.u32 %p180, %r321, 5; mov.u16 %rs71, 0; @%p180 bra $L__BB23_135; bra.uni $L__BB23_144; $L__BB23_140: add.ftz.f32 %f594, %f491, %f2949; setp.ltu.ftz.f32 %p181, %f594, %f10; setp.gtu.ftz.f32 %p182, %f594, %f2992; or.pred %p183, %p181, %p182; mov.u16 %rs71, 0; @%p183 bra $L__BB23_144; setp.ltu.ftz.f32 %p184, %f2950, 0f00000000; setp.gtu.ftz.f32 %p185, %f2950, 0f3F800000; or.pred %p186, %p185, %p184; @%p186 bra $L__BB23_144; div.approx.ftz.f32 %f2377, %f576, %f579; div.approx.ftz.f32 %f2378, %f577, %f579; div.approx.ftz.f32 %f2379, %f578, %f579; fma.rn.ftz.f32 %f2380, %f575, %f2377, %f572; fma.rn.ftz.f32 %f2381, %f575, %f2378, %f573; fma.rn.ftz.f32 %f2382, %f575, %f2379, %f574; mul.ftz.f32 %f2383, %f573, %f2379; mul.ftz.f32 %f2384, %f574, %f2378; sub.ftz.f32 %f2385, %f2383, %f2384; mul.ftz.f32 %f2386, %f574, %f2377; mul.ftz.f32 %f2387, %f572, %f2379; sub.ftz.f32 %f2388, %f2386, %f2387; mul.ftz.f32 %f2389, %f572, %f2378; mul.ftz.f32 %f2390, %f573, %f2377; sub.ftz.f32 %f2391, %f2389, %f2390; mul.ftz.f32 %f2392, %f2388, %f2382; mul.ftz.f32 %f2393, %f2391, %f2381; sub.ftz.f32 %f2394, %f2392, %f2393; mul.ftz.f32 %f2395, %f2380, %f2391; mul.ftz.f32 %f2396, %f2385, %f2382; sub.ftz.f32 %f2397, %f2395, %f2396; mul.ftz.f32 %f2398, %f2381, %f2385; mul.ftz.f32 %f2399, %f2380, %f2388; sub.ftz.f32 %f2400, %f2398, %f2399; mul.ftz.f32 %f2401, %f3028, %f2397; fma.rn.ftz.f32 %f2402, %f3029, %f2394, %f2401; fma.rn.ftz.f32 %f2403, %f3027, %f2400, %f2402; setp.gt.ftz.f32 %p187, %f2403, 0f00000000; @%p187 bra $L__BB23_144; st.local.f32 [%rd4], %f594; st.local.f32 [%rd4+4], %f2950; mov.u32 %r255, 0; st.local.u32 [%rd4+8], %r255; mov.u16 %rs71, 1; mov.f32 %f2992, %f594; $L__BB23_144: mov.u16 %rs73, 0; or.b16 %rs47, %rs71, %rs77; and.b16 %rs48, %rs47, 255; setp.ne.s16 %p188, %rs48, 0; selp.u16 %rs77, 1, 0, %p188; $L__BB23_145: setp.gtu.ftz.f32 %p241, %f565, %f2958; add.ftz.f32 %f2404, %f491, %f565; setp.ltu.ftz.f32 %p189, %f2404, %f10; setp.gtu.ftz.f32 %p190, %f2404, %f2992; or.pred %p191, %p189, %p190; or.pred %p193, %p241, %p191; mov.u16 %rs76, %rs73; @%p193 bra $L__BB23_158; setp.lt.s32 %p194, %r323, 4; mov.u16 %rs76, 1; @%p194 bra $L__BB23_158; sqrt.approx.ftz.f32 %f2405, %f175; min.ftz.f32 %f2406, %f498, %f501; min.ftz.f32 %f2407, %f492, %f495; min.ftz.f32 %f2408, %f2407, %f2406; min.ftz.f32 %f2409, %f499, %f502; min.ftz.f32 %f2410, %f493, %f496; min.ftz.f32 %f2411, %f2410, %f2409; min.ftz.f32 %f2412, %f500, %f503; min.ftz.f32 %f2413, %f494, %f497; min.ftz.f32 %f2414, %f2413, %f2412; min.ftz.f32 %f2415, %f396, %f400; min.ftz.f32 %f2416, %f388, %f392; min.ftz.f32 %f2417, %f2416, %f2415; max.ftz.f32 %f2418, %f498, %f501; max.ftz.f32 %f2419, %f2407, %f2418; max.ftz.f32 %f2420, %f499, %f502; max.ftz.f32 %f2421, %f2410, %f2420; max.ftz.f32 %f2422, %f500, %f503; max.ftz.f32 %f2423, %f2413, %f2422; max.ftz.f32 %f2424, %f396, %f400; max.ftz.f32 %f2425, %f2416, %f2424; abs.ftz.f32 %f2426, %f2408; abs.ftz.f32 %f2427, %f2411; abs.ftz.f32 %f2428, %f2414; abs.ftz.f32 %f2429, %f2417; abs.ftz.f32 %f2430, %f2419; abs.ftz.f32 %f2431, %f2421; abs.ftz.f32 %f2432, %f2423; abs.ftz.f32 %f2433, %f2425; max.ftz.f32 %f2434, %f2426, %f2430; max.ftz.f32 %f2435, %f2427, %f2431; max.ftz.f32 %f2436, %f2428, %f2432; max.ftz.f32 %f2437, %f2429, %f2433; max.ftz.f32 %f2438, %f2436, %f2437; max.ftz.f32 %f2439, %f2435, %f2438; max.ftz.f32 %f2440, %f2434, %f2439; mul.ftz.f32 %f601, %f2440, 0f36000000; mov.u32 %r322, 0; mov.f32 %f2959, %f3006; $L__BB23_148: mul.ftz.f32 %f2749, %f2405, 0f36000000; mov.f32 %f2960, 0f00000000; mov.f32 %f2737, 0f3F800000; sub.ftz.f32 %f2443, %f2737, %f2959; mul.ftz.f32 %f2444, %f2959, %f2443; mul.ftz.f32 %f2445, %f2443, %f2444; mul.ftz.f32 %f2446, %f2959, 0f40400000; add.ftz.f32 %f2447, %f2446, 0fC0A00000; mul.ftz.f32 %f2448, %f2959, %f2959; fma.rn.ftz.f32 %f2449, %f2448, %f2447, 0f40000000; mov.f32 %f2450, 0f40000000; mul.ftz.f32 %f2451, %f2443, %f2443; mul.ftz.f32 %f2452, %f2443, 0f40400000; add.ftz.f32 %f2453, %f2452, 0fC0A00000; fma.rn.ftz.f32 %f2454, %f2451, %f2453, 0f40000000; mul.ftz.f32 %f2455, %f2959, %f2444; mul.ftz.f32 %f2456, %f492, %f2445; mul.ftz.f32 %f2457, %f493, %f2445; mul.ftz.f32 %f2458, %f494, %f2445; mul.ftz.f32 %f2459, %f388, %f2445; mul.ftz.f32 %f2460, %f495, %f2449; mul.ftz.f32 %f2461, %f496, %f2449; mul.ftz.f32 %f2462, %f497, %f2449; mul.ftz.f32 %f2463, %f392, %f2449; sub.ftz.f32 %f2464, %f2460, %f2456; sub.ftz.f32 %f2465, %f2461, %f2457; sub.ftz.f32 %f2466, %f2462, %f2458; sub.ftz.f32 %f2467, %f2463, %f2459; fma.rn.ftz.f32 %f2468, %f498, %f2454, %f2464; fma.rn.ftz.f32 %f2469, %f499, %f2454, %f2465; fma.rn.ftz.f32 %f2470, %f500, %f2454, %f2466; fma.rn.ftz.f32 %f2471, %f396, %f2454, %f2467; mul.ftz.f32 %f2472, %f501, %f2455; mul.ftz.f32 %f2473, %f502, %f2455; mul.ftz.f32 %f2474, %f503, %f2455; mul.ftz.f32 %f2475, %f400, %f2455; sub.ftz.f32 %f2476, %f2468, %f2472; sub.ftz.f32 %f2477, %f2469, %f2473; sub.ftz.f32 %f2478, %f2470, %f2474; sub.ftz.f32 %f2479, %f2471, %f2475; mul.ftz.f32 %f2480, %f2476, 0f3F000000; mul.ftz.f32 %f2481, %f2477, 0f3F000000; mul.ftz.f32 %f2482, %f2478, 0f3F000000; mul.ftz.f32 %f606, %f2479, 0f3F000000; add.ftz.f32 %f2483, %f2443, %f2443; mul.ftz.f32 %f2484, %f2959, %f2483; sub.ftz.f32 %f2485, %f2484, %f2451; add.ftz.f32 %f2486, %f2959, %f2959; mul.ftz.f32 %f2487, %f2486, %f2447; fma.rn.ftz.f32 %f2488, %f2959, %f2446, %f2487; add.ftz.f32 %f2489, %f2446, 0f40000000; mul.ftz.f32 %f2490, %f2483, %f2489; mul.ftz.f32 %f2491, %f2443, %f2452; sub.ftz.f32 %f2492, %f2490, %f2491; mul.ftz.f32 %f2493, %f2443, 0fC0000000; fma.rn.ftz.f32 %f2494, %f2959, %f2493, %f2448; mul.ftz.f32 %f2495, %f495, %f2488; mul.ftz.f32 %f2496, %f496, %f2488; mul.ftz.f32 %f2497, %f497, %f2488; mul.ftz.f32 %f2498, %f392, %f2488; fma.rn.ftz.f32 %f2499, %f492, %f2485, %f2495; fma.rn.ftz.f32 %f2500, %f493, %f2485, %f2496; fma.rn.ftz.f32 %f2501, %f494, %f2485, %f2497; fma.rn.ftz.f32 %f2502, %f388, %f2485, %f2498; fma.rn.ftz.f32 %f2503, %f498, %f2492, %f2499; fma.rn.ftz.f32 %f2504, %f499, %f2492, %f2500; fma.rn.ftz.f32 %f2505, %f500, %f2492, %f2501; fma.rn.ftz.f32 %f2506, %f396, %f2492, %f2502; fma.rn.ftz.f32 %f2507, %f501, %f2494, %f2503; fma.rn.ftz.f32 %f2508, %f502, %f2494, %f2504; fma.rn.ftz.f32 %f2509, %f503, %f2494, %f2505; fma.rn.ftz.f32 %f2510, %f400, %f2494, %f2506; mul.ftz.f32 %f607, %f2507, 0f3F000000; mul.ftz.f32 %f608, %f2508, 0f3F000000; mul.ftz.f32 %f609, %f2509, 0f3F000000; mul.ftz.f32 %f610, %f2510, 0f3F000000; sub.ftz.f32 %f2511, %f2450, %f2446; mul.ftz.f32 %f2512, %f2959, 0f41100000; add.ftz.f32 %f2513, %f2512, 0fC0A00000; mov.f32 %f2514, 0f40800000; sub.ftz.f32 %f2515, %f2514, %f2512; add.ftz.f32 %f2516, %f2446, 0fBF800000; mul.ftz.f32 %f2517, %f495, %f2513; mul.ftz.f32 %f2518, %f496, %f2513; mul.ftz.f32 %f2519, %f497, %f2513; fma.rn.ftz.f32 %f2520, %f492, %f2511, %f2517; fma.rn.ftz.f32 %f2521, %f493, %f2511, %f2518; fma.rn.ftz.f32 %f2522, %f494, %f2511, %f2519; fma.rn.ftz.f32 %f2523, %f498, %f2515, %f2520; fma.rn.ftz.f32 %f2524, %f499, %f2515, %f2521; fma.rn.ftz.f32 %f2525, %f500, %f2515, %f2522; fma.rn.ftz.f32 %f2526, %f501, %f2516, %f2523; fma.rn.ftz.f32 %f2527, %f502, %f2516, %f2524; fma.rn.ftz.f32 %f2528, %f503, %f2516, %f2525; mul.ftz.f32 %f2529, %f3029, %f2958; sub.ftz.f32 %f611, %f2529, %f2480; mul.ftz.f32 %f2530, %f3028, %f2958; sub.ftz.f32 %f612, %f2530, %f2481; mul.ftz.f32 %f2531, %f3027, %f2958; sub.ftz.f32 %f613, %f2531, %f2482; mul.ftz.f32 %f2532, %f612, %f612; fma.rn.ftz.f32 %f2533, %f611, %f611, %f2532; fma.rn.ftz.f32 %f2534, %f613, %f613, %f2533; sqrt.approx.ftz.f32 %f614, %f2534; mul.ftz.f32 %f2535, %f2749, %f2958; max.ftz.f32 %f615, %f2535, %f601; mul.ftz.f32 %f2536, %f2507, 0fBF000000; mul.ftz.f32 %f2537, %f608, %f608; fma.rn.ftz.f32 %f2538, %f607, %f607, %f2537; fma.rn.ftz.f32 %f2539, %f609, %f609, %f2538; sqrt.approx.ftz.f32 %f2540, %f2539; div.approx.ftz.f32 %f2541, %f607, %f2540; div.approx.ftz.f32 %f2542, %f608, %f2540; div.approx.ftz.f32 %f2543, %f609, %f2540; mul.ftz.f32 %f2544, %f2527, %f608; fma.rn.ftz.f32 %f2545, %f2526, %f607, %f2544; fma.rn.ftz.f32 %f2546, %f2528, %f609, %f2545; mul.ftz.f32 %f2547, %f2526, %f2539; mul.ftz.f32 %f2548, %f2527, %f2539; mul.ftz.f32 %f2549, %f2528, %f2539; mul.ftz.f32 %f2550, %f607, %f2546; mul.ftz.f32 %f2551, %f608, %f2546; mul.ftz.f32 %f2552, %f609, %f2546; sub.ftz.f32 %f2553, %f2547, %f2550; sub.ftz.f32 %f2554, %f2548, %f2551; sub.ftz.f32 %f2555, %f2549, %f2552; mul.ftz.f32 %f2556, %f2539, %f2540; div.approx.ftz.f32 %f2557, %f2553, %f2556; div.approx.ftz.f32 %f2558, %f2554, %f2556; div.approx.ftz.f32 %f2559, %f2555, %f2556; div.approx.ftz.f32 %f2560, %f601, %f2540; mul.ftz.f32 %f2561, %f612, %f2542; fma.rn.ftz.f32 %f2562, %f611, %f2541, %f2561; fma.rn.ftz.f32 %f616, %f613, %f2543, %f2562; fma.rn.ftz.f32 %f2563, %f601, %f614, %f615; add.ftz.f32 %f2564, %f614, 0f3F800000; fma.rn.ftz.f32 %f617, %f2564, %f2560, %f2563; mul.ftz.f32 %f2565, %f2541, %f2536; mul.ftz.f32 %f2566, %f2542, %f608; sub.ftz.f32 %f2567, %f2565, %f2566; mul.ftz.f32 %f2568, %f2543, %f609; sub.ftz.f32 %f2569, %f2567, %f2568; mul.ftz.f32 %f2570, %f612, %f2558; fma.rn.ftz.f32 %f2571, %f611, %f2557, %f2570; fma.rn.ftz.f32 %f2572, %f613, %f2559, %f2571; add.ftz.f32 %f618, %f2569, %f2572; mul.ftz.f32 %f2573, %f3028, %f2542; fma.rn.ftz.f32 %f2574, %f3029, %f2541, %f2573; fma.rn.ftz.f32 %f619, %f3027, %f2543, %f2574; mul.ftz.f32 %f2575, %f616, %f616; sub.ftz.f32 %f2576, %f2534, %f2575; mul.ftz.f32 %f2577, %f611, %f2536; mul.ftz.f32 %f2578, %f612, %f608; sub.ftz.f32 %f2579, %f2577, %f2578; mul.ftz.f32 %f2580, %f613, %f609; sub.ftz.f32 %f620, %f2579, %f2580; mul.ftz.f32 %f2581, %f3028, %f612; fma.rn.ftz.f32 %f2582, %f3029, %f611, %f2581; fma.rn.ftz.f32 %f621, %f3027, %f613, %f2582; setp.leu.ftz.f32 %p195, %f2576, 0f00000000; sqrt.approx.ftz.f32 %f622, %f2576; @%p195 bra $L__BB23_150; rcp.approx.ftz.f32 %f2960, %f622; $L__BB23_150: max.ftz.f32 %f2748, %f396, %f400; min.ftz.f32 %f2747, %f388, %f392; max.ftz.f32 %f2746, %f2747, %f2748; mul.ftz.f32 %f2745, %f2746, 0f36000000; add.ftz.f32 %f2583, %f615, %f617; add.ftz.f32 %f625, %f2745, %f2583; mul.ftz.f32 %f2584, %f616, %f618; sub.ftz.f32 %f2585, %f620, %f2584; mul.ftz.f32 %f2586, %f2585, %f2960; sub.ftz.f32 %f2587, %f2586, %f610; mul.ftz.f32 %f2588, %f616, %f619; sub.ftz.f32 %f2589, %f621, %f2588; mul.ftz.f32 %f2590, %f2589, %f2960; mul.ftz.f32 %f2591, %f618, %f2590; mul.ftz.f32 %f2592, %f619, %f2587; sub.ftz.f32 %f2593, %f2591, %f2592; rcp.approx.ftz.f32 %f2594, %f2593; mul.ftz.f32 %f2595, %f616, %f2590; sub.ftz.f32 %f626, %f622, %f606; mul.ftz.f32 %f2596, %f619, %f626; sub.ftz.f32 %f2597, %f2595, %f2596; mul.ftz.f32 %f2598, %f2597, %f2594; sub.ftz.f32 %f2959, %f2959, %f2598; mul.ftz.f32 %f2599, %f616, %f2587; mul.ftz.f32 %f2600, %f618, %f626; sub.ftz.f32 %f2601, %f2600, %f2599; mul.ftz.f32 %f2602, %f2601, %f2594; sub.ftz.f32 %f2958, %f2958, %f2602; abs.ftz.f32 %f2603, %f616; setp.geu.ftz.f32 %p196, %f2603, %f617; @%p196 bra $L__BB23_152; abs.ftz.f32 %f2604, %f626; setp.lt.ftz.f32 %p197, %f2604, %f625; @%p197 bra $L__BB23_153; $L__BB23_152: add.s32 %r322, %r322, 1; setp.lt.u32 %p198, %r322, 5; mov.u16 %rs74, 0; @%p198 bra $L__BB23_148; bra.uni $L__BB23_157; $L__BB23_153: add.ftz.f32 %f629, %f491, %f2958; setp.ltu.ftz.f32 %p199, %f629, %f10; setp.gtu.ftz.f32 %p200, %f629, %f2992; or.pred %p201, %p199, %p200; mov.u16 %rs74, 0; @%p201 bra $L__BB23_157; setp.ltu.ftz.f32 %p202, %f2959, 0f00000000; setp.gtu.ftz.f32 %p203, %f2959, 0f3F800000; or.pred %p204, %p203, %p202; @%p204 bra $L__BB23_157; div.approx.ftz.f32 %f2605, %f611, %f614; div.approx.ftz.f32 %f2606, %f612, %f614; div.approx.ftz.f32 %f2607, %f613, %f614; fma.rn.ftz.f32 %f2608, %f610, %f2605, %f607; fma.rn.ftz.f32 %f2609, %f610, %f2606, %f608; fma.rn.ftz.f32 %f2610, %f610, %f2607, %f609; mul.ftz.f32 %f2611, %f608, %f2607; mul.ftz.f32 %f2612, %f609, %f2606; sub.ftz.f32 %f2613, %f2611, %f2612; mul.ftz.f32 %f2614, %f609, %f2605; mul.ftz.f32 %f2615, %f607, %f2607; sub.ftz.f32 %f2616, %f2614, %f2615; mul.ftz.f32 %f2617, %f607, %f2606; mul.ftz.f32 %f2618, %f608, %f2605; sub.ftz.f32 %f2619, %f2617, %f2618; mul.ftz.f32 %f2620, %f2616, %f2610; mul.ftz.f32 %f2621, %f2619, %f2609; sub.ftz.f32 %f2622, %f2620, %f2621; mul.ftz.f32 %f2623, %f2608, %f2619; mul.ftz.f32 %f2624, %f2613, %f2610; sub.ftz.f32 %f2625, %f2623, %f2624; mul.ftz.f32 %f2626, %f2609, %f2613; mul.ftz.f32 %f2627, %f2608, %f2616; sub.ftz.f32 %f2628, %f2626, %f2627; mul.ftz.f32 %f2629, %f3028, %f2625; fma.rn.ftz.f32 %f2630, %f3029, %f2622, %f2629; fma.rn.ftz.f32 %f2631, %f3027, %f2628, %f2630; setp.gt.ftz.f32 %p205, %f2631, 0f00000000; @%p205 bra $L__BB23_157; st.local.f32 [%rd4], %f629; st.local.f32 [%rd4+4], %f2959; mov.u32 %r257, 0; st.local.u32 [%rd4+8], %r257; mov.u16 %rs74, 1; mov.f32 %f2992, %f629; $L__BB23_157: or.b16 %rs55, %rs74, %rs77; and.b16 %rs56, %rs55, 255; setp.ne.s16 %p206, %rs56, 0; selp.u16 %rs77, 1, 0, %p206; mov.u16 %rs76, %rs73; $L__BB23_158: setp.eq.s16 %p207, %rs76, 0; @%p207 bra $L__BB23_160; cvt.rn.f32.s32 %f2736, %r324; mul.ftz.f32 %f2735, %f2736, 0f3F000000; sub.ftz.f32 %f2734, %f2973, %f2972; fma.rn.ftz.f32 %f2733, %f2734, %f2735, %f2972; mul.wide.s32 %rd231, %r323, 12; add.s64 %rd232, %rd1, %rd231; st.local.f32 [%rd232], %f2972; st.local.f32 [%rd232+4], %f2973; add.s32 %r259, %r324, 1; st.local.u32 [%rd232+8], %r259; add.s32 %r323, %r323, 1; mov.u32 %r324, -1; mov.f32 %f2972, %f2733; mov.f32 %f2973, %f522; $L__BB23_160: add.s32 %r319, %r324, 1; setp.lt.s32 %p208, %r324, 1; @%p208 bra $L__BB23_122; $L__BB23_161: setp.gt.s32 %p209, %r323, 0; @%p209 bra $L__BB23_165; and.b16 %rs57, %rs77, 255; setp.eq.s16 %p211, %rs57, 0; mov.pred %p248, 0; @%p211 bra $L__BB23_164; st.local.u32 [%rd4+12], %r50; st.local.u32 [%rd4+16], %r49; st.local.u32 [%rd4+20], %r86; mov.pred %p248, -1; mov.u32 %r328, %r50; bra.uni $L__BB23_164; $L__BB23_104: rsqrt.approx.ftz.f32 %f401, %f175; mul.ftz.f32 %f1520, %f3029, %f401; mul.ftz.f32 %f1521, %f3028, %f401; mul.ftz.f32 %f1522, %f3027, %f401; neg.ftz.f32 %f1523, %f1521; neg.ftz.f32 %f1524, %f1522; fma.rn.ftz.f32 %f1525, %f1522, %f1522, 0f00000000; fma.rn.ftz.f32 %f1526, %f1521, %f1521, %f1525; fma.rn.ftz.f32 %f1527, %f1520, %f1520, %f1525; setp.gt.ftz.f32 %p131, %f1526, %f1527; selp.f32 %f1528, 0f00000000, %f1524, %p131; selp.f32 %f1529, %f1522, 0f00000000, %p131; selp.f32 %f1530, %f1523, %f1520, %p131; mul.ftz.f32 %f1531, %f1529, %f1529; fma.rn.ftz.f32 %f1532, %f1528, %f1528, %f1531; fma.rn.ftz.f32 %f1533, %f1530, %f1530, %f1532; sqrt.approx.ftz.f32 %f1534, %f1533; div.approx.ftz.f32 %f1535, %f1528, %f1534; div.approx.ftz.f32 %f1536, %f1529, %f1534; div.approx.ftz.f32 %f1537, %f1530, %f1534; mul.ftz.f32 %f1538, %f1521, %f1537; mul.ftz.f32 %f1539, %f1522, %f1536; sub.ftz.f32 %f1540, %f1538, %f1539; mul.ftz.f32 %f1541, %f1522, %f1535; mul.ftz.f32 %f1542, %f1520, %f1537; sub.ftz.f32 %f1543, %f1541, %f1542; mul.ftz.f32 %f1544, %f1520, %f1536; mul.ftz.f32 %f1545, %f1521, %f1535; sub.ftz.f32 %f1546, %f1544, %f1545; mul.ftz.f32 %f1547, %f1543, %f1543; fma.rn.ftz.f32 %f1548, %f1540, %f1540, %f1547; fma.rn.ftz.f32 %f1549, %f1546, %f1546, %f1548; sqrt.approx.ftz.f32 %f1550, %f1549; div.approx.ftz.f32 %f1551, %f1540, %f1550; div.approx.ftz.f32 %f1552, %f1543, %f1550; div.approx.ftz.f32 %f1553, %f1546, %f1550; mul.ftz.f32 %f1554, %f401, %f1520; mul.ftz.f32 %f1555, %f401, %f1521; mul.ftz.f32 %f1556, %f401, %f1522; sub.ftz.f32 %f1557, %f2888, %f3026; sub.ftz.f32 %f1558, %f2889, %f3025; mul.ftz.f32 %f1559, %f1558, %f1536; fma.rn.ftz.f32 %f1560, %f1557, %f1535, %f1559; sub.ftz.f32 %f1561, %f2890, %f3024; fma.rn.ftz.f32 %f402, %f1561, %f1537, %f1560; mul.ftz.f32 %f1562, %f1558, %f1552; fma.rn.ftz.f32 %f1563, %f1557, %f1551, %f1562; fma.rn.ftz.f32 %f403, %f1561, %f1553, %f1563; mul.ftz.f32 %f1564, %f1558, %f1555; fma.rn.ftz.f32 %f1565, %f1557, %f1554, %f1564; fma.rn.ftz.f32 %f404, %f1561, %f1556, %f1565; sub.ftz.f32 %f1566, %f2892, %f3026; sub.ftz.f32 %f1567, %f2893, %f3025; mul.ftz.f32 %f1568, %f1567, %f1536; fma.rn.ftz.f32 %f1569, %f1566, %f1535, %f1568; sub.ftz.f32 %f1570, %f2894, %f3024; fma.rn.ftz.f32 %f405, %f1570, %f1537, %f1569; mul.ftz.f32 %f1571, %f1567, %f1552; fma.rn.ftz.f32 %f1572, %f1566, %f1551, %f1571; fma.rn.ftz.f32 %f406, %f1570, %f1553, %f1572; mul.ftz.f32 %f1573, %f1567, %f1555; fma.rn.ftz.f32 %f1574, %f1566, %f1554, %f1573; fma.rn.ftz.f32 %f407, %f1570, %f1556, %f1574; sub.ftz.f32 %f1575, %f2896, %f3026; sub.ftz.f32 %f1576, %f2897, %f3025; mul.ftz.f32 %f1577, %f1576, %f1536; fma.rn.ftz.f32 %f1578, %f1575, %f1535, %f1577; sub.ftz.f32 %f1579, %f2898, %f3024; fma.rn.ftz.f32 %f408, %f1579, %f1537, %f1578; mul.ftz.f32 %f1580, %f1576, %f1552; fma.rn.ftz.f32 %f1581, %f1575, %f1551, %f1580; fma.rn.ftz.f32 %f409, %f1579, %f1553, %f1581; mul.ftz.f32 %f1582, %f1576, %f1555; fma.rn.ftz.f32 %f1583, %f1575, %f1554, %f1582; fma.rn.ftz.f32 %f410, %f1579, %f1556, %f1583; sub.ftz.f32 %f1584, %f2900, %f3026; sub.ftz.f32 %f1585, %f2901, %f3025; mul.ftz.f32 %f1586, %f1585, %f1536; fma.rn.ftz.f32 %f1587, %f1584, %f1535, %f1586; sub.ftz.f32 %f1588, %f2902, %f3024; fma.rn.ftz.f32 %f411, %f1588, %f1537, %f1587; mul.ftz.f32 %f1589, %f1585, %f1552; fma.rn.ftz.f32 %f1590, %f1584, %f1551, %f1589; fma.rn.ftz.f32 %f412, %f1588, %f1553, %f1590; mul.ftz.f32 %f1591, %f1585, %f1555; fma.rn.ftz.f32 %f1592, %f1584, %f1554, %f1591; fma.rn.ftz.f32 %f413, %f1588, %f1556, %f1592; abs.ftz.f32 %f1593, %f402; abs.ftz.f32 %f1594, %f403; abs.ftz.f32 %f1595, %f404; abs.ftz.f32 %f1596, %f405; abs.ftz.f32 %f1597, %f406; abs.ftz.f32 %f1598, %f407; max.ftz.f32 %f1599, %f1593, %f1596; max.ftz.f32 %f1600, %f1594, %f1597; max.ftz.f32 %f1601, %f1595, %f1598; abs.ftz.f32 %f1602, %f392; abs.ftz.f32 %f1603, %f388; max.ftz.f32 %f1604, %f1603, %f1602; abs.ftz.f32 %f1605, %f408; abs.ftz.f32 %f1606, %f409; abs.ftz.f32 %f1607, %f410; abs.ftz.f32 %f1608, %f411; abs.ftz.f32 %f1609, %f412; abs.ftz.f32 %f1610, %f413; max.ftz.f32 %f1611, %f1605, %f1608; max.ftz.f32 %f1612, %f1606, %f1609; max.ftz.f32 %f1613, %f1607, %f1610; abs.ftz.f32 %f1614, %f400; abs.ftz.f32 %f1615, %f396; max.ftz.f32 %f1616, %f1615, %f1614; max.ftz.f32 %f1617, %f1599, %f1611; max.ftz.f32 %f1618, %f1600, %f1612; max.ftz.f32 %f1619, %f1601, %f1613; max.ftz.f32 %f1620, %f1604, %f1616; max.ftz.f32 %f1621, %f1617, %f1618; max.ftz.f32 %f1622, %f1619, %f1620; max.ftz.f32 %f1623, %f1621, %f1622; mul.ftz.f32 %f414, %f1623, 0f35000000; sub.ftz.f32 %f1624, %f753, 0f00000000; mul.ftz.f32 %f1625, %f1624, 0f00000000; mul.ftz.f32 %f1626, %f1625, %f1624; mul.ftz.f32 %f1627, %f402, %f1626; mul.ftz.f32 %f1628, %f403, %f1626; mul.ftz.f32 %f1629, %f404, %f1626; add.ftz.f32 %f1630, %f405, %f405; add.ftz.f32 %f1631, %f406, %f406; add.ftz.f32 %f1632, %f407, %f407; sub.ftz.f32 %f1633, %f1630, %f1627; sub.ftz.f32 %f1634, %f1631, %f1628; sub.ftz.f32 %f1635, %f1632, %f1629; add.ftz.f32 %f1636, %f392, %f392; mul.ftz.f32 %f1637, %f388, %f1626; sub.ftz.f32 %f1638, %f1636, %f1637; mul.ftz.f32 %f1639, %f1624, %f1624; mul.ftz.f32 %f1640, %f1624, 0f40400000; add.ftz.f32 %f1641, %f1640, 0fC0A00000; fma.rn.ftz.f32 %f1642, %f1639, %f1641, 0f40000000; fma.rn.ftz.f32 %f1643, %f408, %f1642, %f1633; fma.rn.ftz.f32 %f1644, %f409, %f1642, %f1634; fma.rn.ftz.f32 %f1645, %f410, %f1642, %f1635; fma.rn.ftz.f32 %f1646, %f396, %f1642, %f1638; mul.ftz.f32 %f1647, %f1625, 0f00000000; mul.ftz.f32 %f1648, %f411, %f1647; mul.ftz.f32 %f1649, %f412, %f1647; mul.ftz.f32 %f1650, %f413, %f1647; sub.ftz.f32 %f1651, %f1643, %f1648; sub.ftz.f32 %f1652, %f1644, %f1649; sub.ftz.f32 %f415, %f1645, %f1650; mul.ftz.f32 %f1653, %f400, %f1647; sub.ftz.f32 %f1654, %f1646, %f1653; mul.ftz.f32 %f2909, %f1651, 0f3F000000; mul.ftz.f32 %f2908, %f1652, 0f3F000000; mul.ftz.f32 %f2906, %f1654, 0f3F000000; add.ftz.f32 %f1655, %f1624, %f1624; mul.ftz.f32 %f1656, %f1655, 0f00000000; sub.ftz.f32 %f1657, %f1656, %f1639; mul.ftz.f32 %f1658, %f405, 0f00000000; mul.ftz.f32 %f1659, %f406, 0f00000000; mul.ftz.f32 %f1660, %f407, 0f00000000; fma.rn.ftz.f32 %f1661, %f402, %f1657, %f1658; fma.rn.ftz.f32 %f1662, %f403, %f1657, %f1659; fma.rn.ftz.f32 %f1663, %f404, %f1657, %f1660; mul.ftz.f32 %f1664, %f1640, %f1624; add.ftz.f32 %f1665, %f1655, %f1655; sub.ftz.f32 %f1666, %f1665, %f1664; fma.rn.ftz.f32 %f1667, %f408, %f1666, %f1661; fma.rn.ftz.f32 %f1668, %f409, %f1666, %f1662; fma.rn.ftz.f32 %f1669, %f410, %f1666, %f1663; mul.ftz.f32 %f1670, %f1624, 0fC0000000; fma.rn.ftz.f32 %f1671, %f1670, 0f00000000, 0f00000000; fma.rn.ftz.f32 %f1672, %f411, %f1671, %f1667; fma.rn.ftz.f32 %f1673, %f412, %f1671, %f1668; fma.rn.ftz.f32 %f1674, %f413, %f1671, %f1669; mul.ftz.f32 %f2904, %f1672, 0f3F000000; mul.ftz.f32 %f2905, %f1673, 0f3F000000; mul.ftz.f32 %f1675, %f1674, 0f3F000000; abs.ftz.f32 %f1676, %f2904; abs.ftz.f32 %f1677, %f2905; abs.ftz.f32 %f1678, %f1675; max.ftz.f32 %f1679, %f1676, %f1677; max.ftz.f32 %f1680, %f1679, %f1678; setp.geu.ftz.f32 %p132, %f1680, %f414; @%p132 bra $L__BB23_106; mul.ftz.f32 %f1681, %f14, %f405; fma.rn.ftz.f32 %f1682, %f402, %f13, %f1681; mul.ftz.f32 %f1683, %f14, %f406; fma.rn.ftz.f32 %f1684, %f403, %f13, %f1683; fma.rn.ftz.f32 %f1685, %f408, %f15, %f1682; fma.rn.ftz.f32 %f1686, %f15, %f409, %f1684; fma.rn.ftz.f32 %f1687, %f411, %f16, %f1685; fma.rn.ftz.f32 %f1688, %f412, %f16, %f1686; mul.ftz.f32 %f1689, %f1687, 0f3F000000; mul.ftz.f32 %f1690, %f1688, 0f3F000000; sub.ftz.f32 %f2904, %f1689, %f2909; sub.ftz.f32 %f2905, %f1690, %f2908; $L__BB23_106: mul.ftz.f32 %f1691, %f2905, %f2905; fma.rn.ftz.f32 %f425, %f2904, %f2904, %f1691; setp.lt.s32 %p134, %r150, 1; mov.pred %p246, -1; @%p134 bra $L__BB23_118; add.ftz.f32 %f1692, %f425, 0f00000000; mov.f32 %f1693, 0f00000000; sqrt.approx.ftz.f32 %f1694, %f1692; mul.ftz.f32 %f2907, %f415, 0f3F000000; div.approx.ftz.f32 %f1695, %f1693, %f1694; mul.ftz.f32 %f2910, %f2906, %f1695; neg.ftz.f32 %f1696, %f2904; div.approx.ftz.f32 %f1697, %f1696, %f1694; mul.ftz.f32 %f2911, %f2906, %f1697; div.approx.ftz.f32 %f1698, %f2905, %f1694; mul.ftz.f32 %f2912, %f2906, %f1698; mov.u32 %r316, 0; $L__BB23_108: mov.f32 %f436, %f2912; mov.f32 %f435, %f2911; mov.f32 %f434, %f2910; mov.f32 %f433, %f2909; mov.f32 %f432, %f2908; mov.f32 %f431, %f2907; mov.f32 %f430, %f2906; cvt.rn.f32.s32 %f1699, %r316; mul.ftz.f32 %f440, %f12, %f1699; add.ftz.f32 %f1700, %f12, %f440; sub.ftz.f32 %f1702, %f753, %f1700; mul.ftz.f32 %f1703, %f1700, %f1702; mul.ftz.f32 %f1704, %f1702, %f1703; mul.ftz.f32 %f1705, %f1700, %f1700; mul.ftz.f32 %f1706, %f1700, 0f40400000; add.ftz.f32 %f1707, %f1706, 0fC0A00000; fma.rn.ftz.f32 %f1708, %f1705, %f1707, 0f40000000; mul.ftz.f32 %f1709, %f1702, %f1702; mul.ftz.f32 %f1710, %f1702, 0f40400000; add.ftz.f32 %f1711, %f1710, 0fC0A00000; fma.rn.ftz.f32 %f1712, %f1709, %f1711, 0f40000000; mul.ftz.f32 %f1713, %f1700, %f1703; mul.ftz.f32 %f1714, %f402, %f1704; mul.ftz.f32 %f1715, %f403, %f1704; mul.ftz.f32 %f1716, %f404, %f1704; mul.ftz.f32 %f1717, %f388, %f1704; mul.ftz.f32 %f1718, %f405, %f1708; mul.ftz.f32 %f1719, %f406, %f1708; mul.ftz.f32 %f1720, %f407, %f1708; mul.ftz.f32 %f1721, %f392, %f1708; sub.ftz.f32 %f1722, %f1718, %f1714; sub.ftz.f32 %f1723, %f1719, %f1715; sub.ftz.f32 %f1724, %f1720, %f1716; sub.ftz.f32 %f1725, %f1721, %f1717; fma.rn.ftz.f32 %f1726, %f408, %f1712, %f1722; fma.rn.ftz.f32 %f1727, %f409, %f1712, %f1723; fma.rn.ftz.f32 %f1728, %f410, %f1712, %f1724; fma.rn.ftz.f32 %f1729, %f396, %f1712, %f1725; mul.ftz.f32 %f1730, %f411, %f1713; mul.ftz.f32 %f1731, %f412, %f1713; mul.ftz.f32 %f1732, %f413, %f1713; mul.ftz.f32 %f1733, %f400, %f1713; sub.ftz.f32 %f1734, %f1726, %f1730; sub.ftz.f32 %f1735, %f1727, %f1731; sub.ftz.f32 %f1736, %f1728, %f1732; sub.ftz.f32 %f1737, %f1729, %f1733; mul.ftz.f32 %f2909, %f1734, 0f3F000000; mul.ftz.f32 %f2908, %f1735, 0f3F000000; mul.ftz.f32 %f2907, %f1736, 0f3F000000; mul.ftz.f32 %f2906, %f1737, 0f3F000000; sub.ftz.f32 %f445, %f2909, %f433; sub.ftz.f32 %f446, %f2908, %f432; mul.ftz.f32 %f1738, %f446, %f446; fma.rn.ftz.f32 %f447, %f445, %f445, %f1738; add.ftz.f32 %f1739, %f1702, %f1702; mul.ftz.f32 %f1740, %f1700, %f1739; sub.ftz.f32 %f1741, %f1740, %f1709; add.ftz.f32 %f1742, %f1700, %f1700; mul.ftz.f32 %f1743, %f1742, %f1707; fma.rn.ftz.f32 %f1744, %f1700, %f1706, %f1743; add.ftz.f32 %f1745, %f1706, 0f40000000; mul.ftz.f32 %f1746, %f1739, %f1745; mul.ftz.f32 %f1747, %f1702, %f1710; sub.ftz.f32 %f1748, %f1746, %f1747; mul.ftz.f32 %f1749, %f1702, 0fC0000000; fma.rn.ftz.f32 %f1750, %f1700, %f1749, %f1705; mul.ftz.f32 %f1751, %f405, %f1744; mul.ftz.f32 %f1752, %f406, %f1744; mul.ftz.f32 %f1753, %f407, %f1744; fma.rn.ftz.f32 %f1754, %f402, %f1741, %f1751; fma.rn.ftz.f32 %f1755, %f403, %f1741, %f1752; fma.rn.ftz.f32 %f1756, %f404, %f1741, %f1753; fma.rn.ftz.f32 %f1757, %f408, %f1748, %f1754; fma.rn.ftz.f32 %f1758, %f409, %f1748, %f1755; fma.rn.ftz.f32 %f1759, %f410, %f1748, %f1756; fma.rn.ftz.f32 %f1760, %f411, %f1750, %f1757; fma.rn.ftz.f32 %f1761, %f412, %f1750, %f1758; fma.rn.ftz.f32 %f1762, %f413, %f1750, %f1759; mul.ftz.f32 %f2916, %f1760, 0f3F000000; mul.ftz.f32 %f2917, %f1761, 0f3F000000; mul.ftz.f32 %f450, %f1762, 0f3F000000; abs.ftz.f32 %f1763, %f2916; abs.ftz.f32 %f1764, %f2917; abs.ftz.f32 %f1765, %f450; max.ftz.f32 %f1766, %f1763, %f1764; max.ftz.f32 %f1767, %f1766, %f1765; setp.lt.ftz.f32 %p135, %f1767, %f414; @%p135 bra $L__BB23_109; bra.uni $L__BB23_110; $L__BB23_109: mov.f32 %f2916, %f445; mov.f32 %f2917, %f446; $L__BB23_110: max.ftz.f32 %f1768, %f430, %f2906; mul.ftz.f32 %f1769, %f433, %f446; mul.ftz.f32 %f1770, %f432, %f445; sub.ftz.f32 %f1771, %f1770, %f1769; neg.ftz.f32 %f1772, %f2916; mul.ftz.f32 %f1773, %f2916, %f2916; fma.rn.ftz.f32 %f1774, %f2917, %f2917, %f1773; add.ftz.f32 %f1775, %f1774, 0f00000000; sqrt.approx.ftz.f32 %f1777, %f1775; div.approx.ftz.f32 %f1778, %f2917, %f1777; div.approx.ftz.f32 %f1779, %f1772, %f1777; div.approx.ftz.f32 %f1780, %f1693, %f1777; mul.ftz.f32 %f2912, %f2906, %f1778; mul.ftz.f32 %f2911, %f2906, %f1779; mul.ftz.f32 %f2910, %f2906, %f1780; mul.ftz.f32 %f1781, %f1768, %f1768; mul.ftz.f32 %f1782, %f447, %f1781; mul.ftz.f32 %f1783, %f1771, %f1771; setp.gtu.ftz.f32 %p136, %f1783, %f1782; @%p136 bra $L__BB23_117; add.ftz.f32 %f1784, %f436, %f433; sub.ftz.f32 %f1785, %f433, %f436; add.ftz.f32 %f1786, %f2909, %f2912; sub.ftz.f32 %f1787, %f1786, %f1785; sub.ftz.f32 %f1788, %f432, %f435; add.ftz.f32 %f1789, %f2908, %f2911; sub.ftz.f32 %f1790, %f1789, %f1788; sub.ftz.f32 %f1791, %f431, %f434; add.ftz.f32 %f1792, %f2907, %f2910; sub.ftz.f32 %f1793, %f1792, %f1791; mul.ftz.f32 %f1794, %f1788, %f1793; mul.ftz.f32 %f1795, %f1791, %f1790; sub.ftz.f32 %f1796, %f1794, %f1795; mul.ftz.f32 %f1797, %f1791, %f1787; mul.ftz.f32 %f1798, %f1785, %f1793; sub.ftz.f32 %f1799, %f1797, %f1798; mul.ftz.f32 %f1800, %f1785, %f1790; mul.ftz.f32 %f1801, %f1788, %f1787; sub.ftz.f32 %f1802, %f1800, %f1801; mul.ftz.f32 %f1803, %f1799, 0f00000000; fma.rn.ftz.f32 %f1804, %f1796, 0f00000000, %f1803; add.ftz.f32 %f463, %f1802, %f1804; setp.gtu.ftz.f32 %p138, %f463, 0f00000000; sub.ftz.f32 %f1805, %f2909, %f2912; selp.f32 %f464, %f1805, %f1784, %p138; sub.ftz.f32 %f1806, %f2908, %f2911; add.ftz.f32 %f1807, %f435, %f432; selp.f32 %f465, %f1806, %f1807, %p138; sub.ftz.f32 %f1808, %f2907, %f2910; add.ftz.f32 %f1809, %f434, %f431; selp.f32 %f466, %f1808, %f1809, %p138; selp.f32 %f1810, %f1785, %f1786, %p138; selp.f32 %f1811, %f1788, %f1789, %p138; selp.f32 %f1812, %f1791, %f1792, %p138; selp.f32 %f1813, %f1786, %f1785, %p138; selp.f32 %f1814, %f1789, %f1788, %p138; selp.f32 %f1815, %f1792, %f1791, %p138; sub.ftz.f32 %f467, %f1813, %f464; sub.ftz.f32 %f468, %f1814, %f465; sub.ftz.f32 %f469, %f1815, %f466; sub.ftz.f32 %f470, %f464, %f1810; sub.ftz.f32 %f471, %f465, %f1811; sub.ftz.f32 %f472, %f466, %f1812; mul.ftz.f32 %f1816, %f465, %f469; mul.ftz.f32 %f1817, %f466, %f468; sub.ftz.f32 %f1818, %f1816, %f1817; mul.ftz.f32 %f1819, %f466, %f467; mul.ftz.f32 %f1820, %f464, %f469; sub.ftz.f32 %f1821, %f1819, %f1820; mul.ftz.f32 %f1822, %f464, %f468; mul.ftz.f32 %f1823, %f465, %f467; sub.ftz.f32 %f1824, %f1822, %f1823; mul.ftz.f32 %f1825, %f1821, 0f00000000; fma.rn.ftz.f32 %f1826, %f1818, 0f00000000, %f1825; add.ftz.f32 %f473, %f1824, %f1826; mul.ftz.f32 %f1827, %f1811, %f472; mul.ftz.f32 %f1828, %f1812, %f471; sub.ftz.f32 %f1829, %f1827, %f1828; mul.ftz.f32 %f1830, %f1812, %f470; mul.ftz.f32 %f1831, %f1810, %f472; sub.ftz.f32 %f1832, %f1830, %f1831; mul.ftz.f32 %f1833, %f1810, %f471; mul.ftz.f32 %f1834, %f1811, %f470; sub.ftz.f32 %f1835, %f1833, %f1834; mul.ftz.f32 %f1836, %f1832, 0f00000000; fma.rn.ftz.f32 %f1837, %f1829, 0f00000000, %f1836; add.ftz.f32 %f474, %f1835, %f1837; max.ftz.f32 %f1838, %f473, %f474; setp.gtu.ftz.f32 %p139, %f1838, 0f00000000; mov.pred %p245, -1; @%p139 bra $L__BB23_114; mul.ftz.f32 %f1839, %f471, %f469; mul.ftz.f32 %f1840, %f472, %f468; sub.ftz.f32 %f1841, %f1839, %f1840; mul.ftz.f32 %f1842, %f470, %f469; mul.ftz.f32 %f1843, %f472, %f467; sub.ftz.f32 %f1844, %f1843, %f1842; mul.ftz.f32 %f1845, %f471, %f467; mul.ftz.f32 %f1846, %f470, %f468; sub.ftz.f32 %f1847, %f1846, %f1845; mul.ftz.f32 %f1848, %f1844, 0f00000000; fma.rn.ftz.f32 %f1849, %f1841, 0f00000000, %f1848; add.ftz.f32 %f1850, %f1847, %f1849; rcp.approx.ftz.f32 %f475, %f1850; mul.ftz.f32 %f1851, %f465, %f1844; fma.rn.ftz.f32 %f1852, %f464, %f1841, %f1851; fma.rn.ftz.f32 %f1853, %f466, %f1847, %f1852; mul.ftz.f32 %f476, %f1853, %f475; setp.ltu.ftz.f32 %p141, %f476, %f10; setp.gtu.ftz.f32 %p142, %f476, %f2992; or.pred %p143, %p142, %p141; setp.eq.ftz.f32 %p144, %f1850, 0f00000000; or.pred %p145, %p144, %p143; @%p145 bra $L__BB23_114; mul.ftz.f32 %f1854, %f473, %f475; sub.ftz.f32 %f1856, %f753, %f1854; setp.le.ftz.f32 %p147, %f463, 0f00000000; selp.f32 %f3005, %f1854, %f1856, %p147; mul.ftz.f32 %f1857, %f474, %f475; sub.ftz.f32 %f1858, %f753, %f1857; selp.f32 %f3004, %f1857, %f1858, %p147; mov.pred %p245, 0; mov.f32 %f3003, %f476; $L__BB23_114: @%p245 bra $L__BB23_117; sub.ftz.f32 %f1859, %f2906, %f430; fma.rn.ftz.f32 %f1860, %f1859, %f3005, %f430; add.ftz.f32 %f1861, %f1860, %f1860; mul.ftz.f32 %f1862, %f401, %f1861; setp.leu.ftz.f32 %p148, %f3003, %f1862; @%p148 bra $L__BB23_117; bra.uni $L__BB23_116; $L__BB23_117: add.s32 %r316, %r316, 1; setp.lt.s32 %p151, %r316, %r150; @%p151 bra $L__BB23_108; bra.uni $L__BB23_118; $L__BB23_62: setp.gt.s32 %p74, %r53, %r52; selp.b32 %r193, -1, 0, %p74; add.s32 %r194, %r53, %r193; mad.lo.s32 %r195, %r194, %r51, %r303; add.s32 %r196, %r189, %r195; mul.wide.u32 %rd173, %r196, 12; add.s64 %rd248, %rd25, %rd173; add.s32 %r197, %r190, %r195; mul.wide.u32 %rd174, %r197, 12; add.s64 %rd251, %rd25, %rd174; add.s32 %r198, %r191, %r195; mul.wide.u32 %rd175, %r198, 12; add.s64 %rd254, %rd25, %rd175; add.s64 %rd249, %rd248, 4; add.s64 %rd250, %rd248, 8; add.s64 %rd252, %rd251, 4; add.s64 %rd253, %rd251, 8; add.s64 %rd255, %rd254, 4; add.s64 %rd256, %rd254, 8; $L__BB23_64: ld.global.f32 %f194, [%rd256]; ld.global.f32 %f195, [%rd255]; ld.global.f32 %f196, [%rd254]; ld.global.f32 %f197, [%rd253]; ld.global.f32 %f198, [%rd252]; ld.global.f32 %f199, [%rd251]; ld.global.f32 %f200, [%rd250]; ld.global.f32 %f201, [%rd249]; ld.global.f32 %f202, [%rd248]; add.s32 %r68, %r53, 1; setp.eq.s32 %p75, %r68, %r52; @%p75 bra $L__BB23_66; bra.uni $L__BB23_65; $L__BB23_66: mul.wide.u32 %rd182, %r192, 12; add.s64 %rd257, %rd16, %rd182; add.s32 %r207, %r192, 1; mul.wide.u32 %rd183, %r207, 12; add.s64 %rd260, %rd16, %rd183; add.s32 %r208, %r192, 2; mul.wide.u32 %rd184, %r208, 12; add.s64 %rd263, %rd16, %rd184; add.s64 %rd258, %rd257, 4; add.s64 %rd259, %rd257, 8; add.s64 %rd261, %rd260, 4; add.s64 %rd262, %rd260, 8; add.s64 %rd264, %rd263, 4; add.s64 %rd265, %rd263, 8; bra.uni $L__BB23_67; $L__BB23_65: setp.ge.s32 %p76, %r53, %r52; selp.b32 %r201, -1, 0, %p76; add.s32 %r202, %r68, %r201; mad.lo.s32 %r203, %r202, %r51, %r303; add.s32 %r204, %r189, %r203; mul.wide.u32 %rd179, %r204, 12; add.s64 %rd257, %rd25, %rd179; add.s32 %r205, %r190, %r203; mul.wide.u32 %rd180, %r205, 12; add.s64 %rd260, %rd25, %rd180; add.s32 %r206, %r191, %r203; mul.wide.u32 %rd181, %r206, 12; add.s64 %rd263, %rd25, %rd181; add.s64 %rd258, %rd257, 4; add.s64 %rd259, %rd257, 8; add.s64 %rd261, %rd260, 4; add.s64 %rd262, %rd260, 8; add.s64 %rd264, %rd263, 4; add.s64 %rd265, %rd263, 8; $L__BB23_67: cvt.rn.f32.s32 %f1256, %r53; sub.ftz.f32 %f1257, %f193, %f1256; sub.ftz.f32 %f1259, %f753, %f1257; ld.global.f32 %f1260, [%rd257]; mul.ftz.f32 %f1261, %f1257, %f1260; ld.global.f32 %f1262, [%rd258]; mul.ftz.f32 %f1263, %f1257, %f1262; ld.global.f32 %f1264, [%rd259]; mul.ftz.f32 %f1265, %f1257, %f1264; fma.rn.ftz.f32 %f1266, %f1259, %f202, %f1261; fma.rn.ftz.f32 %f1267, %f1259, %f201, %f1263; fma.rn.ftz.f32 %f1268, %f1259, %f200, %f1265; ld.global.f32 %f1269, [%rd260]; mul.ftz.f32 %f1270, %f1257, %f1269; ld.global.f32 %f1271, [%rd261]; mul.ftz.f32 %f1272, %f1257, %f1271; ld.global.f32 %f1273, [%rd262]; mul.ftz.f32 %f1274, %f1257, %f1273; fma.rn.ftz.f32 %f1275, %f1259, %f199, %f1270; fma.rn.ftz.f32 %f1276, %f1259, %f198, %f1272; fma.rn.ftz.f32 %f1277, %f1259, %f197, %f1274; ld.global.f32 %f1278, [%rd263]; mul.ftz.f32 %f1279, %f1257, %f1278; ld.global.f32 %f1280, [%rd264]; mul.ftz.f32 %f1281, %f1257, %f1280; ld.global.f32 %f1282, [%rd265]; mul.ftz.f32 %f1283, %f1257, %f1282; fma.rn.ftz.f32 %f1284, %f1259, %f196, %f1279; fma.rn.ftz.f32 %f1285, %f1259, %f195, %f1281; fma.rn.ftz.f32 %f1286, %f1259, %f194, %f1283; sub.ftz.f32 %f203, %f1266, %f3026; sub.ftz.f32 %f204, %f1267, %f3025; sub.ftz.f32 %f205, %f1268, %f3024; sub.ftz.f32 %f1287, %f1275, %f3026; sub.ftz.f32 %f1288, %f1276, %f3025; sub.ftz.f32 %f1289, %f1277, %f3024; sub.ftz.f32 %f1290, %f1284, %f3026; sub.ftz.f32 %f1291, %f1285, %f3025; sub.ftz.f32 %f1292, %f1286, %f3024; sub.ftz.f32 %f206, %f1290, %f203; sub.ftz.f32 %f207, %f1291, %f204; sub.ftz.f32 %f208, %f1292, %f205; sub.ftz.f32 %f209, %f203, %f1287; sub.ftz.f32 %f210, %f204, %f1288; sub.ftz.f32 %f211, %f205, %f1289; sub.ftz.f32 %f1293, %f1287, %f1290; sub.ftz.f32 %f1294, %f1288, %f1291; sub.ftz.f32 %f1295, %f1289, %f1292; add.ftz.f32 %f1296, %f1290, %f203; add.ftz.f32 %f1297, %f1291, %f204; add.ftz.f32 %f1298, %f1292, %f205; mul.ftz.f32 %f1299, %f1298, %f207; mul.ftz.f32 %f1300, %f208, %f1297; sub.ftz.f32 %f1301, %f1299, %f1300; mul.ftz.f32 %f1302, %f208, %f1296; mul.ftz.f32 %f1303, %f1298, %f206; sub.ftz.f32 %f1304, %f1302, %f1303; mul.ftz.f32 %f1305, %f1297, %f206; mul.ftz.f32 %f1306, %f207, %f1296; sub.ftz.f32 %f1307, %f1305, %f1306; mul.ftz.f32 %f1308, %f3028, %f1304; fma.rn.ftz.f32 %f1309, %f3029, %f1301, %f1308; fma.rn.ftz.f32 %f212, %f3027, %f1307, %f1309; add.ftz.f32 %f1310, %f1287, %f203; add.ftz.f32 %f1311, %f1288, %f204; add.ftz.f32 %f1312, %f1289, %f205; mul.ftz.f32 %f1313, %f1312, %f210; mul.ftz.f32 %f1314, %f211, %f1311; sub.ftz.f32 %f1315, %f1313, %f1314; mul.ftz.f32 %f1316, %f211, %f1310; mul.ftz.f32 %f1317, %f1312, %f209; sub.ftz.f32 %f1318, %f1316, %f1317; mul.ftz.f32 %f1319, %f1311, %f209; mul.ftz.f32 %f1320, %f210, %f1310; sub.ftz.f32 %f1321, %f1319, %f1320; mul.ftz.f32 %f1322, %f3028, %f1318; fma.rn.ftz.f32 %f1323, %f3029, %f1315, %f1322; fma.rn.ftz.f32 %f213, %f3027, %f1321, %f1323; add.ftz.f32 %f1324, %f1290, %f1287; add.ftz.f32 %f1325, %f1291, %f1288; add.ftz.f32 %f1326, %f1292, %f1289; mul.ftz.f32 %f1327, %f1326, %f1294; mul.ftz.f32 %f1328, %f1295, %f1325; sub.ftz.f32 %f1329, %f1327, %f1328; mul.ftz.f32 %f1330, %f1295, %f1324; mul.ftz.f32 %f1331, %f1326, %f1293; sub.ftz.f32 %f1332, %f1330, %f1331; mul.ftz.f32 %f1333, %f1325, %f1293; mul.ftz.f32 %f1334, %f1294, %f1324; sub.ftz.f32 %f1335, %f1333, %f1334; mul.ftz.f32 %f1336, %f3028, %f1332; fma.rn.ftz.f32 %f1337, %f3029, %f1329, %f1336; fma.rn.ftz.f32 %f1338, %f3027, %f1335, %f1337; add.ftz.f32 %f1339, %f212, %f213; add.ftz.f32 %f214, %f1338, %f1339; abs.ftz.f32 %f215, %f214; mul.ftz.f32 %f1340, %f215, 0f34000000; min.ftz.f32 %f1341, %f213, %f1338; min.ftz.f32 %f1342, %f212, %f1341; max.ftz.f32 %f1343, %f213, %f1338; max.ftz.f32 %f1344, %f212, %f1343; mul.ftz.f32 %f1345, %f215, 0fB4000000; setp.ltu.ftz.f32 %p78, %f1342, %f1345; setp.gtu.ftz.f32 %p79, %f1344, %f1340; and.pred %p80, %p78, %p79; mov.pred %p242, -1; @%p80 bra $L__BB23_73; mul.ftz.f32 %f1346, %f208, %f210; mul.ftz.f32 %f1347, %f211, %f207; sub.ftz.f32 %f1348, %f1346, %f1347; mul.ftz.f32 %f1349, %f208, %f209; mul.ftz.f32 %f1350, %f211, %f206; sub.ftz.f32 %f1351, %f1350, %f1349; mul.ftz.f32 %f1352, %f210, %f206; mul.ftz.f32 %f1353, %f207, %f209; sub.ftz.f32 %f1354, %f1353, %f1352; add.ftz.f32 %f216, %f1348, %f1348; add.ftz.f32 %f217, %f1351, %f1351; add.ftz.f32 %f218, %f1354, %f1354; mul.ftz.f32 %f1355, %f3028, %f217; fma.rn.ftz.f32 %f1356, %f3029, %f216, %f1355; fma.rn.ftz.f32 %f219, %f3027, %f218, %f1356; setp.eq.ftz.f32 %p82, %f219, 0f00000000; @%p82 bra $L__BB23_73; mul.ftz.f32 %f1357, %f204, %f217; fma.rn.ftz.f32 %f1358, %f203, %f216, %f1357; fma.rn.ftz.f32 %f1359, %f205, %f218, %f1358; div.approx.ftz.f32 %f220, %f1359, %f219; setp.ltu.ftz.f32 %p84, %f220, %f10; setp.gtu.ftz.f32 %p85, %f220, %f2992; or.pred %p86, %p85, %p84; @%p86 bra $L__BB23_73; setp.lt.ftz.f32 %p87, %f215, 0f219392EF; mov.f32 %f2841, 0f00000000; @%p87 bra $L__BB23_72; rcp.approx.ftz.f32 %f2841, %f214; $L__BB23_72: mul.ftz.f32 %f1361, %f212, %f2841; min.ftz.f32 %f3001, %f1361, %f753; mul.ftz.f32 %f1363, %f213, %f2841; min.ftz.f32 %f3000, %f1363, %f753; mov.pred %p242, 0; mov.f32 %f3002, %f220; $L__BB23_73: mov.pred %p243, 0; @%p242 bra $L__BB23_76; add.s64 %rd186, %rd22, %rd163; ld.global.u32 %r209, [%rd186]; and.b32 %r210, %r209, %r135; setp.eq.s32 %p91, %r210, 0; @%p91 bra $L__BB23_76; st.local.f32 [%rd4], %f3002; st.local.f32 [%rd4+4], %f3001; st.local.f32 [%rd4+8], %f3000; st.local.u32 [%rd4+12], %r50; st.local.u32 [%rd4+16], %r49; mov.u32 %r211, 65; st.local.u32 [%rd4+20], %r211; mov.pred %p243, -1; mov.u32 %r328, %r50; mov.f32 %f2992, %f3002; $L__BB23_76: setp.ne.s32 %p93, %r5, 0; and.pred %p94, %p93, %p243; @%p94 bra $L__BB23_182; bra.uni $L__BB23_176; $L__BB23_81: setp.gt.s32 %p105, %r73, %r72; selp.b32 %r217, -1, 0, %p105; add.s32 %r218, %r73, %r217; mad.lo.s32 %r219, %r218, %r71, %r50; add.s32 %r220, %r219, %r309; mul.wide.s32 %rd197, %r220, 16; add.s64 %rd267, %rd26, %rd197; $L__BB23_83: ld.global.v4.f32 {%f1371, %f1372, %f1373, %f1374}, [%rd267]; add.s32 %r84, %r73, 1; setp.eq.s32 %p106, %r84, %r72; @%p106 bra $L__BB23_85; bra.uni $L__BB23_84; $L__BB23_85: shl.b64 %rd200, %rd34, 4; add.s64 %rd268, %rd24, %rd200; bra.uni $L__BB23_86; $L__BB23_84: setp.ge.s32 %p107, %r73, %r72; selp.b32 %r221, -1, 0, %p107; add.s32 %r222, %r84, %r221; mad.lo.s32 %r223, %r222, %r71, %r50; add.s32 %r224, %r223, %r309; mul.wide.s32 %rd199, %r224, 16; add.s64 %rd268, %rd26, %rd199; $L__BB23_86: cvt.rn.f32.s32 %f1375, %r73; sub.ftz.f32 %f1376, %f232, %f1375; ld.global.v4.f32 {%f1377, %f1378, %f1379, %f1380}, [%rd268]; sub.ftz.f32 %f1386, %f753, %f1376; mul.ftz.f32 %f1387, %f1376, %f1377; mul.ftz.f32 %f1388, %f1376, %f1378; mul.ftz.f32 %f1389, %f1376, %f1379; mul.ftz.f32 %f1390, %f1376, %f1380; fma.rn.ftz.f32 %f2852, %f1386, %f1374, %f1390; fma.rn.ftz.f32 %f2851, %f1386, %f1373, %f1389; fma.rn.ftz.f32 %f2850, %f1386, %f1372, %f1388; fma.rn.ftz.f32 %f2849, %f1386, %f1371, %f1387; $L__BB23_88: sub.ftz.f32 %f1395, %f2849, %f3026; sub.ftz.f32 %f1396, %f2850, %f3025; sub.ftz.f32 %f1397, %f2851, %f3024; mul.ftz.f32 %f1398, %f3028, %f1396; fma.rn.ftz.f32 %f1399, %f3029, %f1395, %f1398; fma.rn.ftz.f32 %f1400, %f3027, %f1397, %f1399; mul.ftz.f32 %f249, %f176, %f1400; mul.ftz.f32 %f1401, %f3029, %f249; mul.ftz.f32 %f1402, %f3028, %f249; mul.ftz.f32 %f1403, %f3027, %f249; sub.ftz.f32 %f1404, %f1395, %f1401; sub.ftz.f32 %f1405, %f1396, %f1402; sub.ftz.f32 %f1406, %f1397, %f1403; mul.ftz.f32 %f1407, %f1405, %f1405; fma.rn.ftz.f32 %f1408, %f1404, %f1404, %f1407; fma.rn.ftz.f32 %f250, %f1406, %f1406, %f1408; mul.ftz.f32 %f251, %f2852, %f2852; setp.gtu.ftz.f32 %p109, %f250, %f251; mov.pred %p244, 0; @%p109 bra $L__BB23_91; sub.ftz.f32 %f1409, %f251, %f250; mul.ftz.f32 %f1410, %f176, %f1409; sqrt.approx.ftz.f32 %f1411, %f1410; sub.ftz.f32 %f252, %f249, %f1411; setp.ltu.ftz.f32 %p111, %f252, %f10; setp.gtu.ftz.f32 %p112, %f252, %f2992; or.pred %p113, %p112, %p111; @%p113 bra $L__BB23_91; st.local.f32 [%rd4], %f252; st.local.u32 [%rd4+12], %r50; st.local.u32 [%rd4+16], %r49; st.local.u32 [%rd4+20], %r70; mov.u32 %r225, 0; st.local.u32 [%rd4+4], %r225; st.local.u32 [%rd4+8], %r225; mov.pred %p244, -1; mov.u32 %r328, %r50; mov.f32 %f2992, %f252; $L__BB23_91: setp.ne.s32 %p115, %r5, 0; and.pred %p116, %p115, %p244; @%p116 bra $L__BB23_182; bra.uni $L__BB23_176; $L__BB23_116: fma.rn.ftz.f32 %f3004, %f3004, 0f40000000, 0fBF800000; st.local.f32 [%rd4], %f3003; fma.rn.ftz.f32 %f1863, %f12, %f3005, %f440; st.local.f32 [%rd4+4], %f1863; st.local.f32 [%rd4+8], %f3004; mov.pred %p246, 0; mov.f32 %f2992, %f3003; $L__BB23_118: mov.pred %p248, 0; @%p246 bra $L__BB23_164; st.local.u32 [%rd4+12], %r50; st.local.u32 [%rd4+16], %r49; st.local.u32 [%rd4+20], %r86; mov.pred %p248, -1; mov.u32 %r328, %r50; $L__BB23_164: setp.ne.s32 %p213, %r5, 0; and.pred %p214, %p213, %p248; @%p214 bra $L__BB23_182; bra.uni $L__BB23_176; $L__BB23_14: mul.wide.s32 %rd145, %r297, 4; xor.b64 %rd146, %rd145, -4; add.s64 %rd147, %rd8, %rd146; ld.global.u32 %r332, [%rd147]; cvt.s64.s32 %rd30, %r332; mul.wide.s32 %rd149, %r332, 4; add.s64 %rd150, %rd148, %rd149; ld.global.u8 %rs30, [%rd150]; and.b16 %rs31, %rs30, 2; setp.eq.s16 %p39, %rs31, 0; mul.wide.s32 %rd151, %r332, 208; add.s64 %rd152, %rd11, %rd151; add.s64 %rd31, %rd152, 48; @%p39 bra $L__BB23_21; bra.uni $L__BB23_15; $L__BB23_21: ld.global.v4.f32 {%f2813, %f2814, %f2815, %f2816}, [%rd31]; ld.global.v4.f32 {%f2817, %f2818, %f2819, %f2820}, [%rd31+16]; ld.global.v4.f32 {%f2821, %f2822, %f2823, %f2824}, [%rd31+32]; bra.uni $L__BB23_22; $L__BB23_15: ld.global.u32 %rd153, [%rd31+120]; ld.global.u32 %r165, [%rd31+104]; shl.b32 %r166, %r165, 1; cvt.rn.f32.s32 %f976, %r166; ld.local.f32 %f977, [%rd3+32]; mul.ftz.f32 %f978, %f977, %f976; cvt.rzi.ftz.s32.f32 %r167, %f978; add.s32 %r168, %r166, -1; min.s32 %r169, %r167, %r168; cvt.rn.f32.s32 %f979, %r169; sub.ftz.f32 %f74, %f978, %f979; cvt.s64.s32 %rd154, %r169; add.s64 %rd155, %rd154, %rd153; shl.b64 %rd157, %rd155, 6; add.s64 %rd32, %rd156, %rd157; ld.global.v4.f32 {%f980, %f981, %f982, %f983}, [%rd32]; ld.global.v4.f32 {%f984, %f985, %f986, %f987}, [%rd32+64]; mul.ftz.f32 %f988, %f981, %f985; fma.rn.ftz.f32 %f989, %f980, %f984, %f988; mul.ftz.f32 %f990, %f983, %f987; fma.rn.ftz.f32 %f991, %f982, %f986, %f990; add.ftz.f32 %f83, %f989, %f991; setp.gt.ftz.f32 %p40, %f83, 0f3F7FDF3B; sub.ftz.f32 %f84, %f753, %f74; @%p40 bra $L__BB23_17; bra.uni $L__BB23_16; $L__BB23_17: mul.ftz.f32 %f1046, %f980, %f84; fma.rn.ftz.f32 %f1047, %f984, %f74, %f1046; mul.ftz.f32 %f1048, %f981, %f84; fma.rn.ftz.f32 %f1049, %f985, %f74, %f1048; mul.ftz.f32 %f1050, %f982, %f84; fma.rn.ftz.f32 %f1051, %f986, %f74, %f1050; mul.ftz.f32 %f1052, %f983, %f84; fma.rn.ftz.f32 %f1053, %f987, %f74, %f1052; mul.ftz.f32 %f1054, %f1049, %f1049; fma.rn.ftz.f32 %f1055, %f1047, %f1047, %f1054; mul.ftz.f32 %f1056, %f1053, %f1053; fma.rn.ftz.f32 %f1057, %f1051, %f1051, %f1056; add.ftz.f32 %f1058, %f1055, %f1057; rsqrt.approx.ftz.f32 %f1059, %f1058; mul.ftz.f32 %f2805, %f1053, %f1059; mul.ftz.f32 %f2804, %f1051, %f1059; mul.ftz.f32 %f2803, %f1049, %f1059; mul.ftz.f32 %f2802, %f1047, %f1059; bra.uni $L__BB23_18; $L__BB23_16: mov.f32 %f993, 0fBF800000; max.ftz.f32 %f994, %f83, %f993; min.ftz.f32 %f996, %f994, %f753; abs.ftz.f32 %f997, %f996; sub.ftz.f32 %f998, %f753, %f997; mul.ftz.f32 %f999, %f998, 0f3F000000; sqrt.approx.ftz.f32 %f1000, %f999; setp.gt.ftz.f32 %p41, %f997, 0f3F11EB85; selp.f32 %f1001, %f1000, %f997, %p41; mul.ftz.f32 %f1002, %f1001, %f1001; mov.f32 %f1003, 0f3C94D2E9; mov.f32 %f1004, 0f3D53F941; fma.rn.ftz.f32 %f1005, %f1004, %f1002, %f1003; mov.f32 %f1006, 0f3D3F841F; fma.rn.ftz.f32 %f1007, %f1005, %f1002, %f1006; mov.f32 %f1008, 0f3D994929; fma.rn.ftz.f32 %f1009, %f1007, %f1002, %f1008; mov.f32 %f1010, 0f3E2AAB94; fma.rn.ftz.f32 %f1011, %f1009, %f1002, %f1010; mul.ftz.f32 %f1012, %f1002, %f1011; fma.rn.ftz.f32 %f1013, %f1012, %f1001, %f1001; add.ftz.f32 %f1014, %f1013, %f1013; mov.f32 %f1015, 0f3FC90FDB; sub.ftz.f32 %f1016, %f1015, %f1013; selp.f32 %f1017, %f1014, %f1016, %p41; setp.lt.ftz.f32 %p42, %f996, 0f00000000; mov.f32 %f1018, 0f40490FDB; sub.ftz.f32 %f1019, %f1018, %f1017; selp.f32 %f1020, %f1019, %f1017, %p42; mul.ftz.f32 %f1021, %f980, %f83; sub.ftz.f32 %f1022, %f984, %f1021; mul.ftz.f32 %f1023, %f981, %f83; sub.ftz.f32 %f1024, %f985, %f1023; mul.ftz.f32 %f1025, %f982, %f83; sub.ftz.f32 %f1026, %f986, %f1025; mul.ftz.f32 %f1027, %f983, %f83; sub.ftz.f32 %f1028, %f987, %f1027; mul.ftz.f32 %f1029, %f1024, %f1024; fma.rn.ftz.f32 %f1030, %f1022, %f1022, %f1029; mul.ftz.f32 %f1031, %f1028, %f1028; fma.rn.ftz.f32 %f1032, %f1026, %f1026, %f1031; add.ftz.f32 %f1033, %f1030, %f1032; rsqrt.approx.ftz.f32 %f1034, %f1033; mul.ftz.f32 %f1035, %f1022, %f1034; mul.ftz.f32 %f1036, %f1024, %f1034; mul.ftz.f32 %f1037, %f1026, %f1034; mul.ftz.f32 %f1038, %f1028, %f1034; mul.ftz.f32 %f1039, %f74, %f1020; cos.approx.ftz.f32 %f1040, %f1039; sin.approx.ftz.f32 %f1041, %f1039; mul.ftz.f32 %f1042, %f1035, %f1041; mul.ftz.f32 %f1043, %f1036, %f1041; mul.ftz.f32 %f1044, %f1037, %f1041; mul.ftz.f32 %f1045, %f1038, %f1041; fma.rn.ftz.f32 %f2805, %f983, %f1040, %f1045; fma.rn.ftz.f32 %f2804, %f982, %f1040, %f1044; fma.rn.ftz.f32 %f2803, %f981, %f1040, %f1043; fma.rn.ftz.f32 %f2802, %f980, %f1040, %f1042; $L__BB23_18: ld.global.v4.f32 {%f1060, %f1061, %f1062, %f1063}, [%rd32+16]; ld.global.v4.f32 {%f1068, %f1069, %f1070, %f1071}, [%rd32+80]; mul.ftz.f32 %f1076, %f74, %f1068; mul.ftz.f32 %f1077, %f74, %f1069; mul.ftz.f32 %f1078, %f74, %f1070; mul.ftz.f32 %f1079, %f74, %f1071; fma.rn.ftz.f32 %f97, %f84, %f1060, %f1076; fma.rn.ftz.f32 %f98, %f84, %f1061, %f1077; fma.rn.ftz.f32 %f99, %f84, %f1062, %f1078; fma.rn.ftz.f32 %f1080, %f84, %f1063, %f1079; ld.global.v4.f32 {%f1081, %f1082, %f1083, %f1084}, [%rd32+32]; ld.global.v4.f32 {%f1089, %f1090, %f1091, %f1092}, [%rd32+96]; mul.ftz.f32 %f1097, %f74, %f1089; mul.ftz.f32 %f1098, %f74, %f1090; mul.ftz.f32 %f1099, %f74, %f1091; mul.ftz.f32 %f1100, %f74, %f1092; fma.rn.ftz.f32 %f1101, %f84, %f1081, %f1097; fma.rn.ftz.f32 %f1102, %f84, %f1082, %f1098; fma.rn.ftz.f32 %f1103, %f84, %f1083, %f1099; fma.rn.ftz.f32 %f1104, %f84, %f1084, %f1100; ld.global.v4.f32 {%f1105, %f1106, %f1107, %f1108}, [%rd32+48]; ld.global.v4.f32 {%f1113, %f1114, %f1115, %f1116}, [%rd32+112]; mul.ftz.f32 %f1121, %f74, %f1113; mul.ftz.f32 %f1122, %f74, %f1114; mul.ftz.f32 %f1123, %f74, %f1115; mul.ftz.f32 %f1124, %f74, %f1116; fma.rn.ftz.f32 %f1125, %f84, %f1105, %f1121; fma.rn.ftz.f32 %f1126, %f84, %f1106, %f1122; fma.rn.ftz.f32 %f1127, %f84, %f1107, %f1123; fma.rn.ftz.f32 %f1128, %f84, %f1108, %f1124; mul.ftz.f32 %f1129, %f2805, 0f3FB504F3; mul.ftz.f32 %f1130, %f2802, 0f3FB504F3; mul.ftz.f32 %f1131, %f2803, 0f3FB504F3; mul.ftz.f32 %f1132, %f2804, 0f3FB504F3; mul.ftz.f32 %f1133, %f1129, %f1130; mul.ftz.f32 %f1134, %f1129, %f1131; mul.ftz.f32 %f1135, %f1129, %f1132; mul.ftz.f32 %f1136, %f1130, %f1130; mul.ftz.f32 %f1137, %f1130, %f1131; mul.ftz.f32 %f1138, %f1130, %f1132; mul.ftz.f32 %f1139, %f1131, %f1131; mul.ftz.f32 %f1140, %f1131, %f1132; mul.ftz.f32 %f1141, %f1132, %f1132; sub.ftz.f32 %f1143, %f753, %f1139; sub.ftz.f32 %f1144, %f1143, %f1141; sub.ftz.f32 %f1145, %f1137, %f1135; add.ftz.f32 %f1146, %f1134, %f1138; add.ftz.f32 %f1147, %f1135, %f1137; sub.ftz.f32 %f1148, %f753, %f1136; sub.ftz.f32 %f1149, %f1148, %f1141; sub.ftz.f32 %f1150, %f1140, %f1133; sub.ftz.f32 %f1151, %f1138, %f1134; add.ftz.f32 %f1152, %f1133, %f1140; sub.ftz.f32 %f1153, %f1148, %f1139; mul.ftz.f32 %f1154, %f1145, %f1103; fma.rn.ftz.f32 %f1155, %f1144, %f1080, %f1154; fma.rn.ftz.f32 %f2811, %f1146, %f1126, %f1155; mul.ftz.f32 %f1156, %f1145, %f1104; fma.rn.ftz.f32 %f1157, %f1144, %f1101, %f1156; fma.rn.ftz.f32 %f101, %f1146, %f1127, %f1157; mul.ftz.f32 %f1158, %f1145, %f1125; fma.rn.ftz.f32 %f1159, %f1144, %f1102, %f1158; fma.rn.ftz.f32 %f102, %f1146, %f1128, %f1159; mul.ftz.f32 %f1160, %f1149, %f1103; fma.rn.ftz.f32 %f1161, %f1147, %f1080, %f1160; fma.rn.ftz.f32 %f103, %f1150, %f1126, %f1161; mul.ftz.f32 %f1162, %f1149, %f1104; fma.rn.ftz.f32 %f1163, %f1147, %f1101, %f1162; fma.rn.ftz.f32 %f2810, %f1150, %f1127, %f1163; mul.ftz.f32 %f1164, %f1149, %f1125; fma.rn.ftz.f32 %f1165, %f1147, %f1102, %f1164; fma.rn.ftz.f32 %f105, %f1150, %f1128, %f1165; mul.ftz.f32 %f1166, %f1152, %f1103; fma.rn.ftz.f32 %f1167, %f1151, %f1080, %f1166; fma.rn.ftz.f32 %f106, %f1153, %f1126, %f1167; mul.ftz.f32 %f1168, %f1152, %f1104; fma.rn.ftz.f32 %f1169, %f1151, %f1101, %f1168; fma.rn.ftz.f32 %f107, %f1153, %f1127, %f1169; mul.ftz.f32 %f1170, %f1152, %f1125; fma.rn.ftz.f32 %f1171, %f1151, %f1102, %f1170; fma.rn.ftz.f32 %f2809, %f1153, %f1128, %f1171; mul.ftz.f32 %f1172, %f2810, %f2809; mul.ftz.f32 %f109, %f107, %f105; sub.ftz.f32 %f2808, %f1172, %f109; mul.ftz.f32 %f111, %f107, %f102; mul.ftz.f32 %f1173, %f101, %f2809; sub.ftz.f32 %f2807, %f111, %f1173; mul.ftz.f32 %f113, %f101, %f105; mul.ftz.f32 %f1174, %f2810, %f102; sub.ftz.f32 %f2806, %f113, %f1174; mul.ftz.f32 %f1175, %f103, %f2807; fma.rn.ftz.f32 %f1176, %f2811, %f2808, %f1175; fma.rn.ftz.f32 %f2812, %f106, %f2806, %f1176; setp.neu.ftz.f32 %p43, %f2812, 0f00000000; @%p43 bra $L__BB23_20; add.ftz.f32 %f2811, %f2811, 0f322BCC77; add.ftz.f32 %f2809, %f2809, 0f322BCC77; add.ftz.f32 %f2810, %f2810, 0f322BCC77; mul.ftz.f32 %f1177, %f2810, %f2809; sub.ftz.f32 %f2808, %f1177, %f109; mul.ftz.f32 %f1178, %f101, %f2809; sub.ftz.f32 %f2807, %f111, %f1178; mul.ftz.f32 %f1179, %f2810, %f102; sub.ftz.f32 %f2806, %f113, %f1179; mul.ftz.f32 %f1180, %f103, %f2807; fma.rn.ftz.f32 %f1181, %f2811, %f2808, %f1180; fma.rn.ftz.f32 %f1182, %f106, %f2806, %f1181; setp.eq.ftz.f32 %p44, %f1182, 0f00000000; selp.f32 %f2812, 0f7F7FFFFF, %f1182, %p44; $L__BB23_20: mul.ftz.f32 %f1183, %f103, %f2809; mul.ftz.f32 %f1184, %f106, %f105; sub.ftz.f32 %f1185, %f1184, %f1183; mul.ftz.f32 %f1186, %f106, %f102; mul.ftz.f32 %f1187, %f2809, %f2811; sub.ftz.f32 %f1188, %f1187, %f1186; mul.ftz.f32 %f1189, %f105, %f2811; mul.ftz.f32 %f1190, %f103, %f102; sub.ftz.f32 %f1191, %f1190, %f1189; div.approx.ftz.f32 %f2817, %f1185, %f2812; div.approx.ftz.f32 %f2818, %f1188, %f2812; div.approx.ftz.f32 %f2819, %f1191, %f2812; mul.ftz.f32 %f1192, %f106, %f2810; mul.ftz.f32 %f1193, %f103, %f107; sub.ftz.f32 %f1194, %f1193, %f1192; mul.ftz.f32 %f1195, %f107, %f2811; mul.ftz.f32 %f1196, %f106, %f101; sub.ftz.f32 %f1197, %f1196, %f1195; mul.ftz.f32 %f1198, %f103, %f101; mul.ftz.f32 %f1199, %f2810, %f2811; sub.ftz.f32 %f1200, %f1199, %f1198; div.approx.ftz.f32 %f2821, %f1194, %f2812; div.approx.ftz.f32 %f2822, %f1197, %f2812; div.approx.ftz.f32 %f2823, %f1200, %f2812; div.approx.ftz.f32 %f2813, %f2808, %f2812; div.approx.ftz.f32 %f2814, %f2807, %f2812; mul.ftz.f32 %f1201, %f98, %f2814; fma.rn.ftz.f32 %f1202, %f97, %f2813, %f1201; div.approx.ftz.f32 %f2815, %f2806, %f2812; fma.rn.ftz.f32 %f1203, %f99, %f2815, %f1202; neg.ftz.f32 %f2816, %f1203; mul.ftz.f32 %f1204, %f98, %f2818; fma.rn.ftz.f32 %f1205, %f97, %f2817, %f1204; fma.rn.ftz.f32 %f1206, %f99, %f2819, %f1205; neg.ftz.f32 %f2820, %f1206; mul.ftz.f32 %f1207, %f98, %f2822; fma.rn.ftz.f32 %f1208, %f97, %f2821, %f1207; fma.rn.ftz.f32 %f1209, %f99, %f2823, %f1208; neg.ftz.f32 %f2824, %f1209; $L__BB23_22: ld.local.f32 %f1222, [%rd3]; ld.local.f32 %f1223, [%rd3+4]; mul.ftz.f32 %f1224, %f2814, %f1223; fma.rn.ftz.f32 %f1225, %f2813, %f1222, %f1224; ld.local.f32 %f1226, [%rd3+8]; fma.rn.ftz.f32 %f1227, %f2815, %f1226, %f1225; add.ftz.f32 %f3026, %f2816, %f1227; mul.ftz.f32 %f1228, %f2818, %f1223; fma.rn.ftz.f32 %f1229, %f2817, %f1222, %f1228; fma.rn.ftz.f32 %f1230, %f2819, %f1226, %f1229; add.ftz.f32 %f3025, %f2820, %f1230; mul.ftz.f32 %f1231, %f2822, %f1223; fma.rn.ftz.f32 %f1232, %f2821, %f1222, %f1231; fma.rn.ftz.f32 %f1233, %f2823, %f1226, %f1232; add.ftz.f32 %f3024, %f2824, %f1233; ld.local.f32 %f1234, [%rd3+12]; ld.local.f32 %f1235, [%rd3+16]; mul.ftz.f32 %f1236, %f2814, %f1235; fma.rn.ftz.f32 %f1237, %f2813, %f1234, %f1236; ld.local.f32 %f1238, [%rd3+20]; fma.rn.ftz.f32 %f1239, %f2815, %f1238, %f1237; mul.ftz.f32 %f1240, %f2818, %f1235; fma.rn.ftz.f32 %f1241, %f2817, %f1234, %f1240; fma.rn.ftz.f32 %f1242, %f2819, %f1238, %f1241; mul.ftz.f32 %f1243, %f2822, %f1235; fma.rn.ftz.f32 %f1244, %f2821, %f1234, %f1243; fma.rn.ftz.f32 %f1245, %f2823, %f1238, %f1244; abs.ftz.f32 %f1246, %f1239; setp.gt.ftz.f32 %p45, %f1246, 0f17800000; mov.b32 %r170, %f1239; and.b32 %r171, %r170, -2147483648; or.b32 %r172, %r171, 394264576; mov.b32 %f1247, %r172; selp.f32 %f3029, %f1239, %f1247, %p45; abs.ftz.f32 %f1248, %f1242; setp.gt.ftz.f32 %p46, %f1248, 0f17800000; mov.b32 %r173, %f1242; and.b32 %r174, %r173, -2147483648; or.b32 %r175, %r174, 394264576; mov.b32 %f1249, %r175; selp.f32 %f3028, %f1242, %f1249, %p46; abs.ftz.f32 %f1250, %f1245; setp.gt.ftz.f32 %p47, %f1250, 0f17800000; mov.b32 %r176, %f1245; and.b32 %r177, %r176, -2147483648; or.b32 %r178, %r177, 394264576; mov.b32 %f1251, %r178; selp.f32 %f3027, %f1245, %f1251, %p47; rcp.approx.ftz.f32 %f3032, %f3029; rcp.approx.ftz.f32 %f3031, %f3028; rcp.approx.ftz.f32 %f3030, %f3027; add.s32 %r292, %r33, 1; mov.u32 %r179, 1985229328; st.local.u32 [%rd29+4], %r179; shl.b64 %rd159, %rd30, 2; add.s64 %rd160, %rd158, %rd159; ld.global.u32 %r293, [%rd160]; bra.uni $L__BB23_177; $L__BB23_1: setp.eq.s32 %p15, %r293, 1985229328; setp.lt.s32 %p16, %r293, 0; or.pred %p17, %p16, %p15; @%p17 bra $L__BB23_12; $L__BB23_2: mul.wide.s32 %rd137, %r293, 16; add.s64 %rd27, %rd7, %rd137; ld.global.v4.f32 {%f763, %f764, %f765, %f766}, [%rd27]; mov.b32 %r293, %f766; mov.b32 %r18, %f765; mov.b32 %r17, %f764; mov.b32 %r16, %f763; mov.b64 %rd138, {%r16, %r17}; cvt.u32.u64 %r151, %rd138; and.b32 %r152, %r151, 1024; setp.eq.s32 %p18, %r152, 0; and.b32 %r20, %r16, %r135; @%p18 bra $L__BB23_5; bra.uni $L__BB23_3; $L__BB23_5: setp.ne.s32 %p24, %r20, 0; ld.global.v4.f32 {%f906, %f907, %f908, %f909}, [%rd27+16]; ld.global.v4.f32 {%f914, %f915, %f916, %f917}, [%rd27+32]; ld.global.v4.f32 {%f922, %f923, %f924, %f925}, [%rd27+48]; sub.ftz.f32 %f930, %f906, %f3026; mul.ftz.f32 %f931, %f3032, %f930; sub.ftz.f32 %f932, %f908, %f3026; mul.ftz.f32 %f933, %f3032, %f932; sub.ftz.f32 %f934, %f914, %f3025; mul.ftz.f32 %f935, %f3031, %f934; sub.ftz.f32 %f936, %f916, %f3025; mul.ftz.f32 %f937, %f3031, %f936; sub.ftz.f32 %f938, %f922, %f3024; mul.ftz.f32 %f939, %f3030, %f938; sub.ftz.f32 %f940, %f924, %f3024; mul.ftz.f32 %f941, %f3030, %f940; min.ftz.f32 %f942, %f931, %f933; min.ftz.f32 %f943, %f935, %f937; min.ftz.f32 %f944, %f939, %f941; max.ftz.f32 %f945, %f10, %f942; max.ftz.f32 %f946, %f943, %f944; max.ftz.f32 %f2801, %f945, %f946; max.ftz.f32 %f947, %f931, %f933; max.ftz.f32 %f948, %f935, %f937; max.ftz.f32 %f949, %f939, %f941; min.ftz.f32 %f950, %f2992, %f947; min.ftz.f32 %f951, %f948, %f949; min.ftz.f32 %f952, %f950, %f951; sub.ftz.f32 %f953, %f907, %f3026; mul.ftz.f32 %f954, %f3032, %f953; sub.ftz.f32 %f955, %f909, %f3026; mul.ftz.f32 %f956, %f3032, %f955; sub.ftz.f32 %f957, %f915, %f3025; mul.ftz.f32 %f958, %f3031, %f957; sub.ftz.f32 %f959, %f917, %f3025; mul.ftz.f32 %f960, %f3031, %f959; sub.ftz.f32 %f961, %f923, %f3024; mul.ftz.f32 %f962, %f3030, %f961; sub.ftz.f32 %f963, %f925, %f3024; mul.ftz.f32 %f964, %f3030, %f963; min.ftz.f32 %f965, %f954, %f956; min.ftz.f32 %f966, %f958, %f960; min.ftz.f32 %f967, %f962, %f964; max.ftz.f32 %f968, %f10, %f965; max.ftz.f32 %f969, %f966, %f967; max.ftz.f32 %f2800, %f968, %f969; max.ftz.f32 %f970, %f954, %f956; max.ftz.f32 %f971, %f958, %f960; max.ftz.f32 %f972, %f962, %f964; min.ftz.f32 %f973, %f2992, %f970; min.ftz.f32 %f974, %f971, %f972; min.ftz.f32 %f975, %f973, %f974; setp.ge.ftz.f32 %p25, %f952, %f2801; and.pred %p26, %p24, %p25; selp.u32 %r156, 1, 0, %p26; setp.ltu.ftz.f32 %p27, %f975, %f2800; and.b32 %r158, %r17, %r135; setp.eq.s32 %p28, %r158, 0; or.pred %p29, %p28, %p27; selp.b32 %r159, 0, 2, %p29; or.b32 %r291, %r159, %r156; bra.uni $L__BB23_6; $L__BB23_3: setp.ne.s32 %p19, %r20, 0; ld.global.v4.f32 {%f770, %f771, %f772, %f773}, [%rd27+16]; ld.global.v4.f32 {%f778, %f779, %f780, %f781}, [%rd27+32]; ld.global.v4.f32 {%f786, %f787, %f788, %f789}, [%rd27+48]; mul.ftz.f32 %f794, %f3028, %f771; fma.rn.ftz.f32 %f795, %f3029, %f770, %f794; fma.rn.ftz.f32 %f796, %f3027, %f772, %f795; mul.ftz.f32 %f797, %f3028, %f779; fma.rn.ftz.f32 %f798, %f3029, %f778, %f797; fma.rn.ftz.f32 %f799, %f3027, %f780, %f798; mul.ftz.f32 %f800, %f3028, %f787; fma.rn.ftz.f32 %f801, %f3029, %f786, %f800; fma.rn.ftz.f32 %f802, %f3027, %f788, %f801; mul.ftz.f32 %f803, %f3025, %f771; fma.rn.ftz.f32 %f804, %f3026, %f770, %f803; fma.rn.ftz.f32 %f805, %f3024, %f772, %f804; add.ftz.f32 %f806, %f773, %f805; mul.ftz.f32 %f807, %f3025, %f779; fma.rn.ftz.f32 %f808, %f3026, %f778, %f807; fma.rn.ftz.f32 %f809, %f3024, %f780, %f808; add.ftz.f32 %f810, %f781, %f809; mul.ftz.f32 %f811, %f3025, %f787; fma.rn.ftz.f32 %f812, %f3026, %f786, %f811; fma.rn.ftz.f32 %f813, %f3024, %f788, %f812; add.ftz.f32 %f814, %f789, %f813; rcp.approx.ftz.f32 %f815, %f796; rcp.approx.ftz.f32 %f816, %f799; rcp.approx.ftz.f32 %f817, %f802; mul.ftz.f32 %f818, %f815, %f806; neg.ftz.f32 %f819, %f818; mul.ftz.f32 %f820, %f816, %f810; neg.ftz.f32 %f821, %f820; mul.ftz.f32 %f822, %f817, %f814; neg.ftz.f32 %f823, %f822; sub.ftz.f32 %f824, %f815, %f818; sub.ftz.f32 %f825, %f816, %f820; sub.ftz.f32 %f826, %f817, %f822; min.ftz.f32 %f827, %f819, %f824; min.ftz.f32 %f828, %f821, %f825; min.ftz.f32 %f829, %f823, %f826; max.ftz.f32 %f830, %f819, %f824; max.ftz.f32 %f831, %f821, %f825; max.ftz.f32 %f832, %f823, %f826; max.ftz.f32 %f833, %f10, %f827; max.ftz.f32 %f834, %f828, %f829; max.ftz.f32 %f2801, %f833, %f834; min.ftz.f32 %f835, %f2992, %f830; min.ftz.f32 %f836, %f831, %f832; min.ftz.f32 %f837, %f835, %f836; setp.le.ftz.f32 %p20, %f2801, %f837; and.pred %p21, %p19, %p20; selp.u32 %r291, 1, 0, %p21; ld.global.v4.f32 {%f838, %f839, %f840, %f841}, [%rd27+64]; ld.global.v4.f32 {%f846, %f847, %f848, %f849}, [%rd27+80]; ld.global.v4.f32 {%f854, %f855, %f856, %f857}, [%rd27+96]; mul.ftz.f32 %f862, %f3028, %f839; fma.rn.ftz.f32 %f863, %f3029, %f838, %f862; fma.rn.ftz.f32 %f864, %f3027, %f840, %f863; mul.ftz.f32 %f865, %f3028, %f847; fma.rn.ftz.f32 %f866, %f3029, %f846, %f865; fma.rn.ftz.f32 %f867, %f3027, %f848, %f866; mul.ftz.f32 %f868, %f3028, %f855; fma.rn.ftz.f32 %f869, %f3029, %f854, %f868; fma.rn.ftz.f32 %f870, %f3027, %f856, %f869; mul.ftz.f32 %f871, %f3025, %f839; fma.rn.ftz.f32 %f872, %f3026, %f838, %f871; fma.rn.ftz.f32 %f873, %f3024, %f840, %f872; add.ftz.f32 %f874, %f841, %f873; mul.ftz.f32 %f875, %f3025, %f847; fma.rn.ftz.f32 %f876, %f3026, %f846, %f875; fma.rn.ftz.f32 %f877, %f3024, %f848, %f876; add.ftz.f32 %f878, %f849, %f877; mul.ftz.f32 %f879, %f3025, %f855; fma.rn.ftz.f32 %f880, %f3026, %f854, %f879; fma.rn.ftz.f32 %f881, %f3024, %f856, %f880; add.ftz.f32 %f882, %f857, %f881; rcp.approx.ftz.f32 %f883, %f864; rcp.approx.ftz.f32 %f884, %f867; rcp.approx.ftz.f32 %f885, %f870; mul.ftz.f32 %f886, %f883, %f874; neg.ftz.f32 %f887, %f886; mul.ftz.f32 %f888, %f884, %f878; neg.ftz.f32 %f889, %f888; mul.ftz.f32 %f890, %f885, %f882; neg.ftz.f32 %f891, %f890; sub.ftz.f32 %f892, %f883, %f886; sub.ftz.f32 %f893, %f884, %f888; sub.ftz.f32 %f894, %f885, %f890; min.ftz.f32 %f895, %f887, %f892; min.ftz.f32 %f896, %f889, %f893; min.ftz.f32 %f897, %f891, %f894; max.ftz.f32 %f898, %f887, %f892; max.ftz.f32 %f899, %f889, %f893; max.ftz.f32 %f900, %f891, %f894; max.ftz.f32 %f901, %f10, %f895; max.ftz.f32 %f902, %f896, %f897; max.ftz.f32 %f2800, %f901, %f902; min.ftz.f32 %f903, %f2992, %f898; min.ftz.f32 %f904, %f899, %f900; min.ftz.f32 %f905, %f903, %f904; setp.gtu.ftz.f32 %p22, %f2800, %f905; @%p22 bra $L__BB23_6; and.b32 %r154, %r17, %r135; setp.eq.s32 %p23, %r154, 0; or.b32 %r155, %r291, 2; selp.b32 %r291, %r291, %r155, %p23; $L__BB23_6: mul.wide.s32 %rd139, %r292, 4; add.s64 %rd28, %rd2, %rd139; setp.eq.s32 %p30, %r291, 3; @%p30 bra $L__BB23_10; bra.uni $L__BB23_7; $L__BB23_10: setp.lt.ftz.f32 %p33, %f2800, %f2801; selp.b32 %r29, %r293, %r18, %p33; selp.b32 %r160, %r18, %r293, %p33; add.s32 %r292, %r292, 1; st.local.u32 [%rd28+4], %r160; mov.u32 %r293, %r29; bra.uni $L__BB23_11; $L__BB23_7: setp.eq.s32 %p31, %r291, 2; @%p31 bra $L__BB23_11; setp.ne.s32 %p32, %r291, 0; mov.u32 %r293, %r18; @%p32 bra $L__BB23_11; ld.local.u32 %r293, [%rd28]; add.s32 %r292, %r292, -1; $L__BB23_11: setp.ne.s32 %p34, %r293, 1985229328; setp.gt.s32 %p35, %r293, -1; and.pred %p36, %p35, %p34; @%p36 bra $L__BB23_2; $L__BB23_12: mov.u32 %r33, %r292; setp.gt.s32 %p37, %r293, -1; @%p37 bra $L__BB23_177; bra.uni $L__BB23_13; $L__BB23_177: setp.ne.s32 %p234, %r293, 1985229328; @%p234 bra $L__BB23_1; mov.u32 %r293, 1985229328; setp.lt.s32 %p235, %r292, 0; @%p235 bra $L__BB23_180; ld.local.f32 %f3026, [%rd3]; ld.local.f32 %f3025, [%rd3+4]; ld.local.f32 %f3024, [%rd3+8]; ld.local.f32 %f2718, [%rd3+12]; abs.ftz.f32 %f2719, %f2718; setp.gt.ftz.f32 %p236, %f2719, 0f17800000; mov.b32 %r269, %f2718; and.b32 %r270, %r269, -2147483648; or.b32 %r271, %r270, 394264576; mov.b32 %f2720, %r271; selp.f32 %f3029, %f2718, %f2720, %p236; ld.local.f32 %f2721, [%rd3+16]; abs.ftz.f32 %f2722, %f2721; setp.gt.ftz.f32 %p237, %f2722, 0f17800000; mov.b32 %r272, %f2721; and.b32 %r273, %r272, -2147483648; or.b32 %r274, %r273, 394264576; mov.b32 %f2723, %r274; selp.f32 %f3028, %f2721, %f2723, %p237; ld.local.f32 %f2724, [%rd3+20]; abs.ftz.f32 %f2725, %f2724; setp.gt.ftz.f32 %p238, %f2725, 0f17800000; mov.b32 %r275, %f2724; and.b32 %r276, %r275, -2147483648; or.b32 %r277, %r276, 394264576; mov.b32 %f2726, %r277; selp.f32 %f3027, %f2724, %f2726, %p238; rcp.approx.ftz.f32 %f3032, %f3029; rcp.approx.ftz.f32 %f3031, %f3028; rcp.approx.ftz.f32 %f3030, %f3027; mul.wide.s32 %rd245, %r292, 4; add.s64 %rd246, %rd2, %rd245; ld.local.u32 %r293, [%rd246]; add.s32 %r292, %r292, -1; mov.u32 %r332, -1; $L__BB23_180: setp.ne.s32 %p239, %r293, 1985229328; @%p239 bra $L__BB23_1; setp.ne.s32 %p240, %r328, -1; selp.u16 %rs79, 1, 0, %p240; $L__BB23_182: cvt.u32.u16 %r278, %rs79; st.param.b32 [func_retval0+0], %r278; ret; } .func (.param .b32 func_retval0) _Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj( .param .b32 _Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_0, .param .b64 _Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_1, .param .b32 _Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_2 ) { .local .align 16 .b8 __local_depot24[816]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<808>; .reg .b16 %rs<246>; .reg .f32 %f<7322>; .reg .b32 %r<1539>; .reg .b64 %rd<1232>; mov.u64 %SPL, __local_depot24; ld.param.u32 %r670, [_Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_0]; ld.param.u64 %rd313, [_Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_1]; ld.param.u32 %r668, [_Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj_param_2]; cvta.to.local.u64 %rd4, %rd313; add.u64 %rd1, %SPL, 0; ld.const.u64 %rd316, [kernel_params+3384]; cvta.to.global.u64 %rd317, %rd316; cvt.s64.s32 %rd3, %r670; mul.wide.s32 %rd318, %r670, 2; add.s64 %rd319, %rd317, %rd318; ld.global.u16 %r671, [%rd319]; not.b32 %r672, %r671; ld.const.u32 %r673, [kernel_params+1556]; add.s32 %r674, %r673, %r672; mov.u32 %r1536, 0; max.s32 %r1, %r674, 0; ld.local.f32 %f6460, [%rd4]; setp.nan.ftz.f32 %p49, %f6460, %f6460; mov.u16 %rs242, 0; mov.f32 %f7321, 0f3F800000; @%p49 bra $L__BB24_594; mov.u16 %rs242, 0; mov.f32 %f7321, 0f3F800000; mov.u32 %r1536, 0; mov.b32 %r676, %f6460; or.b32 %r677, %r676, -2147483648; setp.ne.s32 %p50, %r677, -2147483648; add.ftz.f32 %f1917, %f6460, %f6460; setp.eq.ftz.f32 %p51, %f1917, %f6460; and.pred %p52, %p50, %p51; shl.b32 %r678, %r676, 1; setp.gt.u32 %p53, %r678, -16777216; or.pred %p54, %p53, %p52; @%p54 bra $L__BB24_594; mov.u16 %rs242, 0; mov.f32 %f7321, 0f3F800000; mov.u32 %r1536, 0; ld.local.f32 %f6457, [%rd4+12]; setp.nan.ftz.f32 %p55, %f6457, %f6457; @%p55 bra $L__BB24_594; mov.u16 %rs242, 0; mov.f32 %f7321, 0f3F800000; mov.u32 %r1536, 0; mov.b32 %r681, %f6457; or.b32 %r682, %r681, -2147483648; setp.ne.s32 %p56, %r682, -2147483648; add.ftz.f32 %f1920, %f6457, %f6457; setp.eq.ftz.f32 %p57, %f1920, %f6457; and.pred %p58, %p56, %p57; shl.b32 %r683, %r681, 1; setp.gt.u32 %p59, %r683, -16777216; or.pred %p60, %p59, %p58; @%p60 bra $L__BB24_594; mov.u16 %rs242, 0; mov.f32 %f7321, 0f3F800000; mov.u32 %r1536, 0; ld.local.f32 %f6456, [%rd4+16]; mul.ftz.f32 %f1922, %f6456, %f6456; fma.rn.ftz.f32 %f1923, %f6457, %f6457, %f1922; ld.local.f32 %f6455, [%rd4+20]; fma.rn.ftz.f32 %f1924, %f6455, %f6455, %f1923; setp.eq.ftz.f32 %p61, %f1924, 0f00000000; @%p61 bra $L__BB24_594; bra.uni $L__BB24_5; $L__BB24_594: setp.eq.ftz.f32 %p776, %f7321, 0f3F800000; @%p776 bra $L__BB24_596; ld.const.u64 %rd1113, [kernel_params+3432]; cvta.to.global.u64 %rd1114, %rd1113; mul.lo.s64 %rd1115, %rd3, 12; add.s64 %rd1116, %rd1114, %rd1115; ld.global.f32 %f6439, [%rd1116]; mul.ftz.f32 %f6440, %f7321, %f6439; ld.global.f32 %f6441, [%rd1116+4]; mul.ftz.f32 %f6442, %f7321, %f6441; ld.global.f32 %f6443, [%rd1116+8]; mul.ftz.f32 %f6444, %f7321, %f6443; st.global.f32 [%rd1116], %f6440; st.global.f32 [%rd1116+4], %f6442; st.global.f32 [%rd1116+8], %f6444; $L__BB24_596: setp.gt.u32 %p777, %r1536, %r1; selp.b16 %rs90, 1, %rs242, %p777; setp.eq.s16 %p778, %rs90, 0; @%p778 bra $L__BB24_598; ld.const.u64 %rd1117, [kernel_params+3464]; cvta.to.global.u64 %rd1118, %rd1117; shl.b64 %rd1119, %rd3, 1; add.s64 %rd1120, %rd1118, %rd1119; mov.u16 %rs181, 0; st.global.u16 [%rd1120], %rs181; bra.uni $L__BB24_607; $L__BB24_598: min.u32 %r1209, %r1, 4; min.u32 %r1537, %r1536, %r1209; setp.eq.s32 %p779, %r1537, 0; @%p779 bra $L__BB24_606; ld.const.u64 %rd306, [kernel_params+3552]; mov.u16 %rs182, 0; mov.u64 %rd1121, kernel_params; shl.b64 %rd1123, %rd3, 2; mov.u16 %rs184, 1; $L__BB24_600: add.s32 %r1537, %r1537, -1; setp.eq.s32 %p780, %r1537, 0; mov.u16 %rs244, %rs182; @%p780 bra $L__BB24_605; mov.u32 %r1538, 0; mov.u64 %rd1230, %rd1121; mov.u64 %rd1231, %rd306; mov.u16 %rs244, %rs182; $L__BB24_602: cvta.to.global.u64 %rd1122, %rd1231; add.s64 %rd1124, %rd1122, %rd1123; add.s32 %r1538, %r1538, 1; ld.const.u64 %rd1231, [%rd1230+3600]; cvta.to.global.u64 %rd1125, %rd1231; add.s64 %rd311, %rd1125, %rd1123; ld.global.f32 %f1914, [%rd311]; ld.global.f32 %f6445, [%rd1124]; setp.leu.ftz.f32 %p781, %f6445, %f1914; @%p781 bra $L__BB24_604; ld.const.v2.u64 {%rd1126, %rd1127}, [%rd1230+3568]; cvta.to.global.u64 %rd1130, %rd1127; add.s64 %rd1132, %rd1130, %rd1123; ld.global.u32 %r1211, [%rd1132]; ld.const.v2.u64 {%rd1133, %rd1134}, [%rd1230+3584]; cvta.to.global.u64 %rd1137, %rd1133; add.s64 %rd1138, %rd1137, %rd1123; ld.global.u32 %r1212, [%rd1138]; cvta.to.global.u64 %rd1139, %rd1134; add.s64 %rd1140, %rd1139, %rd1123; ld.global.u32 %r1213, [%rd1140]; ld.const.v2.u64 {%rd1141, %rd1142}, [%rd1230+3552]; cvta.to.global.u64 %rd1145, %rd1142; add.s64 %rd1146, %rd1145, %rd1123; ld.global.f32 %f6446, [%rd1146]; cvta.to.global.u64 %rd1147, %rd1126; add.s64 %rd1148, %rd1147, %rd1123; ld.global.f32 %f6447, [%rd1148]; cvta.to.global.u64 %rd1149, %rd1141; add.s64 %rd1150, %rd1149, %rd1123; ld.global.f32 %f6448, [%rd1150]; ld.const.v2.u64 {%rd1151, %rd1152}, [%rd1230+3616]; cvta.to.global.u64 %rd1155, %rd1152; add.s64 %rd1156, %rd1155, %rd1123; ld.global.u32 %r1214, [%rd1156]; ld.const.v2.u64 {%rd1157, %rd1158}, [%rd1230+3632]; cvta.to.global.u64 %rd1161, %rd1157; add.s64 %rd1162, %rd1161, %rd1123; ld.global.u32 %r1215, [%rd1162]; cvta.to.global.u64 %rd1163, %rd1158; add.s64 %rd1164, %rd1163, %rd1123; ld.global.u32 %r1216, [%rd1164]; ld.const.u64 %rd1165, [%rd1230+3608]; cvta.to.global.u64 %rd1166, %rd1165; add.s64 %rd1167, %rd1166, %rd1123; ld.global.f32 %f6449, [%rd1167]; cvta.to.global.u64 %rd1168, %rd1151; add.s64 %rd1169, %rd1168, %rd1123; ld.global.f32 %f6450, [%rd1169]; st.global.f32 [%rd1150], %f1914; st.global.f32 [%rd1146], %f6449; st.global.f32 [%rd1148], %f6450; st.global.u32 [%rd1138], %r1215; st.global.u32 [%rd1132], %r1214; st.global.u32 [%rd1140], %r1216; st.global.f32 [%rd311], %f6448; st.global.f32 [%rd1167], %f6446; st.global.f32 [%rd1169], %f6447; st.global.u32 [%rd1162], %r1212; st.global.u32 [%rd1156], %r1211; st.global.u32 [%rd1164], %r1213; mov.u16 %rs244, %rs184; $L__BB24_604: add.s64 %rd1230, %rd1230, 48; setp.lt.u32 %p782, %r1538, %r1537; @%p782 bra $L__BB24_602; $L__BB24_605: and.b16 %rs185, %rs244, 255; setp.ne.s16 %p783, %rs185, 0; @%p783 bra $L__BB24_600; $L__BB24_606: ld.const.u64 %rd1170, [kernel_params+3464]; cvta.to.global.u64 %rd1171, %rd1170; shl.b64 %rd1172, %rd3, 1; add.s64 %rd1173, %rd1171, %rd1172; st.global.u16 [%rd1173], %r1536; $L__BB24_607: cvt.u32.u16 %r1217, %rs90; st.param.b32 [func_retval0+0], %r1217; ret; $L__BB24_5: ld.const.u32 %r685, [kernel_params+1108]; setp.eq.s32 %p62, %r685, 0; ld.const.u32 %r2, [kernel_params+1112]; abs.ftz.f32 %f1925, %f6457; setp.gt.ftz.f32 %p63, %f1925, 0f17800000; and.b32 %r687, %r681, -2147483648; or.b32 %r688, %r687, 394264576; mov.b32 %f1926, %r688; selp.f32 %f5, %f6457, %f1926, %p63; abs.ftz.f32 %f1927, %f6456; setp.gt.ftz.f32 %p64, %f1927, 0f17800000; mov.b32 %r689, %f6456; and.b32 %r690, %r689, -2147483648; or.b32 %r691, %r690, 394264576; mov.b32 %f1928, %r691; selp.f32 %f6, %f6456, %f1928, %p64; abs.ftz.f32 %f1929, %f6455; setp.gt.ftz.f32 %p65, %f1929, 0f17800000; mov.b32 %r692, %f6455; and.b32 %r693, %r692, -2147483648; or.b32 %r694, %r693, 394264576; mov.b32 %f1930, %r694; selp.f32 %f7, %f6455, %f1930, %p65; rcp.approx.ftz.f32 %f8, %f5; rcp.approx.ftz.f32 %f9, %f6; rcp.approx.ftz.f32 %f10, %f7; min.u32 %r3, %r1, 4; add.u64 %rd6, %SPL, 48; @%p62 bra $L__BB24_330; setp.ne.s32 %p66, %r2, 0; @%p66 bra $L__BB24_125; bra.uni $L__BB24_7; $L__BB24_125: mov.u32 %r827, 1985229328; st.local.u32 [%rd6], %r827; ld.const.u32 %r1314, [kernel_params+1104]; ld.local.f32 %f6616, [%rd4+4]; ld.local.f32 %f6615, [%rd4+8]; ld.local.f32 %f340, [%rd4+24]; ld.local.f32 %f341, [%rd4+28]; ld.const.v2.u64 {%rd517, %rd518}, [kernel_params+2096]; mov.u32 %r1536, 0; cvta.to.global.u64 %rd112, %rd517; cvta.to.global.u64 %rd113, %rd518; ld.const.v2.u64 {%rd521, %rd522}, [kernel_params+2128]; cvta.to.global.u64 %rd114, %rd522; ld.const.v2.u64 {%rd525, %rd526}, [kernel_params+2176]; cvta.to.global.u64 %rd115, %rd526; ld.const.u64 %rd529, [kernel_params+2160]; cvta.to.global.u64 %rd116, %rd529; cvta.to.global.u64 %rd117, %rd521; ld.const.u64 %rd530, [kernel_params+2232]; cvta.to.global.u64 %rd118, %rd530; ld.const.v2.u64 {%rd531, %rd532}, [kernel_params+2256]; cvta.to.global.u64 %rd119, %rd531; ld.const.v2.u32 {%r828, %r829}, [kernel_params+1128]; ld.const.v2.u64 {%rd535, %rd536}, [kernel_params+2304]; cvta.to.global.u64 %rd120, %rd536; ld.const.v2.u64 {%rd539, %rd540}, [kernel_params+2112]; cvta.to.global.u64 %rd121, %rd539; cvta.to.global.u64 %rd122, %rd532; ld.const.v2.u64 {%rd543, %rd544}, [kernel_params+2144]; cvta.to.global.u64 %rd123, %rd544; ld.const.u64 %rd547, [kernel_params+2408]; cvta.to.global.u64 %rd124, %rd547; ld.const.u64 %rd548, [kernel_params+2216]; cvta.to.global.u64 %rd125, %rd548; cvta.to.global.u64 %rd126, %rd540; ld.const.u64 %rd549, [kernel_params+2272]; cvta.to.global.u64 %rd127, %rd549; ld.const.u64 %rd550, [kernel_params+2296]; cvta.to.global.u64 %rd128, %rd550; ld.const.v2.u64 {%rd551, %rd552}, [kernel_params+2336]; cvta.to.global.u64 %rd129, %rd551; cvta.to.global.u64 %rd130, %rd535; cvt.rn.f32.s32 %f2540, %r829; rcp.approx.ftz.f32 %f342, %f2540; mov.f32 %f7321, 0f3F800000; sub.ftz.f32 %f2541, %f7321, %f342; mul.ftz.f32 %f2542, %f342, %f2541; mul.ftz.f32 %f2543, %f2541, %f2542; neg.ftz.f32 %f343, %f2543; mul.ftz.f32 %f2544, %f342, %f342; fma.rn.ftz.f32 %f2545, %f342, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f344, %f2544, %f2545, 0f40000000; mul.ftz.f32 %f2546, %f2541, %f2541; fma.rn.ftz.f32 %f2547, %f2541, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f345, %f2546, %f2547, 0f40000000; mul.ftz.f32 %f2548, %f342, %f2542; neg.ftz.f32 %f346, %f2548; cvta.to.global.u64 %rd131, %rd552; ld.const.u64 %rd555, [kernel_params+2320]; cvta.to.global.u64 %rd132, %rd555; ld.const.u64 %rd556, [kernel_params+3552]; cvta.to.global.u64 %rd557, %rd556; shl.b64 %rd558, %rd3, 2; add.s64 %rd133, %rd557, %rd558; cvta.to.global.u64 %rd134, %rd543; cvta.to.global.u64 %rd135, %rd525; mov.u32 %r1377, -1; setp.eq.s32 %p241, %r828, 0; mov.u16 %rs114, 1; mov.pred %p249, 0; mov.pred %p395, -1; mov.u16 %rs117, 0; mov.f32 %f6903, %f6615; mov.f32 %f6904, %f6616; mov.f32 %f6905, %f6460; mov.f32 %f6906, %f7; mov.f32 %f6907, %f6; mov.f32 %f6908, %f5; mov.f32 %f6909, %f10; mov.f32 %f6910, %f9; mov.f32 %f6911, %f8; mov.f32 %f6916, %f341; mov.u32 %r1376, %r1536; mov.u32 %r1315, %r1536; bra.uni $L__BB24_126; $L__BB24_138: mul.wide.s32 %rd562, %r1314, 16; xor.b64 %rd563, %rd562, -16; add.s64 %rd564, %rd113, %rd563; ld.global.v4.u32 {%r1318, %r841, %r842, %r843}, [%rd564]; setp.gt.s32 %p230, %r1318, -1; mul.wide.s32 %rd567, %r1315, 4; add.s64 %rd138, %rd6, %rd567; @%p230 bra $L__BB24_148; bra.uni $L__BB24_139; $L__BB24_148: ld.local.u32 %r1314, [%rd138]; add.s32 %r1315, %r1315, -1; setp.le.s32 %p240, %r841, %r1318; @%p240 bra $L__BB24_326; cvt.u16.u32 %rs113, %r843; and.b16 %rs17, %rs113, 127; and.b32 %r859, %r843, 64; setp.eq.s32 %p242, %r859, 0; or.pred %p11, %p242, %p241; mul.ftz.f32 %f3032, %f6908, %f6908; fma.rn.ftz.f32 %f3033, %f6907, %f6907, %f3032; fma.rn.ftz.f32 %f511, %f6906, %f6906, %f3033; rcp.approx.ftz.f32 %f512, %f511; $L__BB24_150: cvt.s64.s32 %rd142, %r1318; setp.ne.s32 %p243, %r1377, -1; mov.u32 %r212, %r1377; @%p243 bra $L__BB24_152; shl.b64 %rd581, %rd142, 2; add.s64 %rd582, %rd114, %rd581; ld.global.u32 %r212, [%rd582]; $L__BB24_152: shl.b64 %rd583, %rd142, 2; add.s64 %rd584, %rd117, %rd583; ld.global.u32 %r860, [%rd584]; cvt.s64.s32 %rd143, %r860; ld.local.u32 %r861, [%rd4+36]; setp.ne.s32 %p244, %r861, %r860; @%p244 bra $L__BB24_154; ld.local.u32 %r862, [%rd4+40]; setp.eq.s32 %p245, %r862, %r212; mov.u16 %rs198, %rs114; @%p245 bra $L__BB24_156; $L__BB24_154: cvt.u32.u64 %r863, %rd143; ld.local.u32 %r864, [%rd4+44]; setp.ne.s32 %p246, %r864, %r863; mov.u16 %rs198, 0; @%p246 bra $L__BB24_156; ld.local.u32 %r865, [%rd4+48]; setp.eq.s32 %p247, %r865, %r212; selp.u16 %rs198, 1, 0, %p247; $L__BB24_156: setp.ne.s16 %p248, %rs198, 0; @%p248 bra $L__BB24_325; setp.gt.s16 %p250, %rs17, 64; @%p250 bra $L__BB24_163; setp.gt.s16 %p256, %rs17, 3; @%p256 bra $L__BB24_161; setp.eq.s16 %p259, %rs17, 1; @%p259 bra $L__BB24_285; setp.eq.s16 %p260, %rs17, 2; mov.pred %p795, %p249; @%p260 bra $L__BB24_166; bra.uni $L__BB24_293; $L__BB24_163: setp.gt.s16 %p251, %rs17, 67; @%p251 bra $L__BB24_172; setp.eq.s16 %p254, %rs17, 65; @%p254 bra $L__BB24_180; setp.eq.s16 %p255, %rs17, 66; mov.pred %p795, %p249; @%p255 bra $L__BB24_166; bra.uni $L__BB24_293; $L__BB24_161: setp.eq.s16 %p257, %rs17, 4; @%p257 bra $L__BB24_166; setp.eq.s16 %p258, %rs17, 8; mov.pred %p795, %p249; @%p258 bra $L__BB24_174; bra.uni $L__BB24_293; $L__BB24_172: setp.eq.s16 %p252, %rs17, 68; @%p252 bra $L__BB24_166; setp.eq.s16 %p253, %rs17, 72; mov.pred %p795, %p249; @%p253 bra $L__BB24_174; bra.uni $L__BB24_293; $L__BB24_174: @%p11 bra $L__BB24_176; shl.b64 %rd607, %rd142, 3; add.s64 %rd608, %rd123, %rd607; ld.global.v2.f32 {%f3144, %f3145}, [%rd608]; ld.local.f32 %f3148, [%rd4+32]; setp.lt.ftz.f32 %p285, %f3148, %f3144; setp.gt.ftz.f32 %p286, %f3148, %f3145; or.pred %p287, %p285, %p286; mov.pred %p795, %p249; @%p287 bra $L__BB24_293; $L__BB24_176: add.s64 %rd610, %rd121, %rd583; ld.global.u32 %r234, [%rd610]; and.b32 %r893, %r234, 64; setp.eq.s32 %p288, %r893, 0; @%p288 bra $L__BB24_210; bra.uni $L__BB24_177; $L__BB24_210: shl.b64 %rd621, %rd143, 4; add.s64 %rd622, %rd128, %rd621; ld.global.v4.f32 {%f6713, %f6714, %f6715, %f6716}, [%rd622]; bra.uni $L__BB24_211; $L__BB24_166: @%p11 bra $L__BB24_168; shl.b64 %rd623, %rd142, 3; add.s64 %rd624, %rd123, %rd623; ld.global.v2.f32 {%f3193, %f3194}, [%rd624]; ld.local.f32 %f3197, [%rd4+32]; setp.lt.ftz.f32 %p303, %f3197, %f3193; setp.gt.ftz.f32 %p304, %f3197, %f3194; or.pred %p305, %p303, %p304; mov.pred %p795, %p249; @%p305 bra $L__BB24_293; $L__BB24_168: add.s64 %rd626, %rd121, %rd583; ld.global.u32 %r252, [%rd626]; and.b32 %r909, %r252, 64; setp.eq.s32 %p306, %r909, 0; shl.b64 %rd627, %rd143, 4; add.s64 %rd628, %rd122, %rd627; shr.s32 %r910, %r252, 7; ld.global.u32 %r911, [%rd628+4]; add.s32 %r253, %r911, %r910; add.s32 %r912, %r253, -1; max.s32 %r254, %r912, %r911; add.s32 %r913, %r253, 2; ld.global.u32 %r914, [%rd628+8]; add.s32 %r915, %r911, %r914; add.s32 %r916, %r915, -1; min.s32 %r255, %r913, %r916; mul.wide.s32 %rd629, %r254, 16; add.s64 %rd210, %rd127, %rd629; mul.wide.s32 %rd630, %r253, 16; add.s64 %rd211, %rd127, %rd630; mul.wide.s32 %rd631, %r255, 16; add.s64 %rd212, %rd127, %rd631; @%p306 bra $L__BB24_224; bra.uni $L__BB24_169; $L__BB24_224: ld.global.v4.f32 {%f6765, %f6764, %f6763, %f732}, [%rd210]; ld.global.v4.f32 {%f6761, %f6760, %f6759, %f728}, [%rd211]; ld.global.v4.f32 {%f6757, %f6756, %f6755, %f724}, [%rd211+16]; ld.global.v4.f32 {%f6753, %f6752, %f6751, %f720}, [%rd212]; bra.uni $L__BB24_225; $L__BB24_169: mul.wide.s32 %rd632, %r212, 208; add.s64 %rd633, %rd116, %rd632; ld.global.u32 %r256, [%rd633+148]; ld.global.u32 %r257, [%rd633+152]; shl.b32 %r917, %r257, 1; cvt.rn.f32.s32 %f3198, %r917; ld.local.f32 %f3199, [%rd4+32]; mul.ftz.f32 %f591, %f3199, %f3198; cvt.rzi.ftz.s32.f32 %r918, %f591; add.s32 %r919, %r917, -1; min.s32 %r258, %r918, %r919; ld.global.u32 %r1336, [%rd633+164]; mul.wide.u32 %rd634, %r1336, 16; add.s64 %rd635, %rd120, %rd634; ld.global.u16 %rs203, [%rd635+12]; ld.global.u32 %r1335, [%rd635+8]; ld.global.u64 %rd1219, [%rd635]; setp.eq.s64 %p307, %rd1219, 11; @%p307 bra $L__BB24_216; $L__BB24_170: setp.eq.s64 %p308, %rd1219, 0; @%p308 bra $L__BB24_214; add.s32 %r1336, %r1336, 2; bra.uni $L__BB24_215; $L__BB24_214: setp.eq.s16 %p309, %rs203, 0; mov.u32 %r1339, -1; mov.u32 %r1336, %r1335; @%p309 bra $L__BB24_217; $L__BB24_215: mul.wide.u32 %rd636, %r1336, 16; add.s64 %rd637, %rd120, %rd636; ld.global.u16 %rs203, [%rd637+12]; ld.global.u32 %r1335, [%rd637+8]; ld.global.u64 %rd1219, [%rd637]; setp.ne.s64 %p310, %rd1219, 11; @%p310 bra $L__BB24_170; $L__BB24_216: setp.eq.s16 %p311, %rs203, 0; selp.b32 %r1339, -1, %r1335, %p311; $L__BB24_217: setp.eq.s32 %p312, %r258, %r257; @%p312 bra $L__BB24_219; bra.uni $L__BB24_218; $L__BB24_219: ld.global.v4.f32 {%f6733, %f6732, %f6731, %f6730}, [%rd210]; ld.global.v4.f32 {%f6729, %f6728, %f6727, %f6726}, [%rd211]; ld.global.v4.f32 {%f6725, %f6724, %f6723, %f6722}, [%rd211+16]; ld.global.v4.f32 {%f6721, %f6720, %f6719, %f6718}, [%rd212]; bra.uni $L__BB24_220; $L__BB24_285: shl.b64 %rd654, %rd143, 4; add.s64 %rd655, %rd118, %rd654; ld.global.u32 %r943, [%rd655+12]; mul.wide.u32 %rd656, %r943, 12; add.s64 %rd657, %rd119, %rd656; add.s32 %r944, %r943, 1; mul.wide.u32 %rd658, %r944, 12; add.s64 %rd659, %rd119, %rd658; add.s32 %r945, %r943, 2; mul.wide.u32 %rd660, %r945, 12; add.s64 %rd661, %rd119, %rd660; ld.global.f32 %f4414, [%rd657]; sub.ftz.f32 %f995, %f4414, %f6905; ld.global.f32 %f4415, [%rd657+4]; sub.ftz.f32 %f996, %f4415, %f6904; ld.global.f32 %f4416, [%rd657+8]; sub.ftz.f32 %f997, %f4416, %f6903; ld.global.f32 %f4417, [%rd659]; sub.ftz.f32 %f4418, %f4417, %f6905; ld.global.f32 %f4419, [%rd659+4]; sub.ftz.f32 %f4420, %f4419, %f6904; ld.global.f32 %f4421, [%rd659+8]; sub.ftz.f32 %f4422, %f4421, %f6903; ld.global.f32 %f4423, [%rd661]; sub.ftz.f32 %f4424, %f4423, %f6905; ld.global.f32 %f4425, [%rd661+4]; sub.ftz.f32 %f4426, %f4425, %f6904; ld.global.f32 %f4427, [%rd661+8]; sub.ftz.f32 %f4428, %f4427, %f6903; sub.ftz.f32 %f998, %f4424, %f995; sub.ftz.f32 %f999, %f4426, %f996; sub.ftz.f32 %f1000, %f4428, %f997; sub.ftz.f32 %f1001, %f995, %f4418; sub.ftz.f32 %f1002, %f996, %f4420; sub.ftz.f32 %f1003, %f997, %f4422; sub.ftz.f32 %f4429, %f4418, %f4424; sub.ftz.f32 %f4430, %f4420, %f4426; sub.ftz.f32 %f4431, %f4422, %f4428; add.ftz.f32 %f4432, %f995, %f4424; add.ftz.f32 %f4433, %f996, %f4426; add.ftz.f32 %f4434, %f997, %f4428; mul.ftz.f32 %f4435, %f999, %f4434; mul.ftz.f32 %f4436, %f4433, %f1000; sub.ftz.f32 %f4437, %f4435, %f4436; mul.ftz.f32 %f4438, %f4432, %f1000; mul.ftz.f32 %f4439, %f998, %f4434; sub.ftz.f32 %f4440, %f4438, %f4439; mul.ftz.f32 %f4441, %f998, %f4433; mul.ftz.f32 %f4442, %f4432, %f999; sub.ftz.f32 %f4443, %f4441, %f4442; mul.ftz.f32 %f4444, %f6907, %f4440; fma.rn.ftz.f32 %f4445, %f6908, %f4437, %f4444; fma.rn.ftz.f32 %f1004, %f6906, %f4443, %f4445; add.ftz.f32 %f4446, %f995, %f4418; add.ftz.f32 %f4447, %f996, %f4420; add.ftz.f32 %f4448, %f997, %f4422; mul.ftz.f32 %f4449, %f1002, %f4448; mul.ftz.f32 %f4450, %f4447, %f1003; sub.ftz.f32 %f4451, %f4449, %f4450; mul.ftz.f32 %f4452, %f4446, %f1003; mul.ftz.f32 %f4453, %f1001, %f4448; sub.ftz.f32 %f4454, %f4452, %f4453; mul.ftz.f32 %f4455, %f1001, %f4447; mul.ftz.f32 %f4456, %f4446, %f1002; sub.ftz.f32 %f4457, %f4455, %f4456; mul.ftz.f32 %f4458, %f6907, %f4454; fma.rn.ftz.f32 %f4459, %f6908, %f4451, %f4458; fma.rn.ftz.f32 %f1005, %f6906, %f4457, %f4459; add.ftz.f32 %f4460, %f4418, %f4424; add.ftz.f32 %f4461, %f4420, %f4426; add.ftz.f32 %f4462, %f4422, %f4428; mul.ftz.f32 %f4463, %f4430, %f4462; mul.ftz.f32 %f4464, %f4461, %f4431; sub.ftz.f32 %f4465, %f4463, %f4464; mul.ftz.f32 %f4466, %f4460, %f4431; mul.ftz.f32 %f4467, %f4429, %f4462; sub.ftz.f32 %f4468, %f4466, %f4467; mul.ftz.f32 %f4469, %f4429, %f4461; mul.ftz.f32 %f4470, %f4460, %f4430; sub.ftz.f32 %f4471, %f4469, %f4470; mul.ftz.f32 %f4472, %f6907, %f4468; fma.rn.ftz.f32 %f4473, %f6908, %f4465, %f4472; fma.rn.ftz.f32 %f4474, %f6906, %f4471, %f4473; add.ftz.f32 %f4475, %f1005, %f1004; add.ftz.f32 %f1006, %f4474, %f4475; abs.ftz.f32 %f1007, %f1006; mul.ftz.f32 %f4476, %f1007, 0f34000000; min.ftz.f32 %f4477, %f1005, %f4474; min.ftz.f32 %f4478, %f1004, %f4477; max.ftz.f32 %f4479, %f1005, %f4474; max.ftz.f32 %f4480, %f1004, %f4479; mul.ftz.f32 %f4481, %f1007, 0fB4000000; setp.ltu.ftz.f32 %p396, %f4478, %f4481; setp.gtu.ftz.f32 %p397, %f4480, %f4476; and.pred %p398, %p396, %p397; mov.pred %p794, %p395; @%p398 bra $L__BB24_291; mul.ftz.f32 %f4482, %f1002, %f1000; mul.ftz.f32 %f4483, %f1003, %f999; sub.ftz.f32 %f4484, %f4482, %f4483; mul.ftz.f32 %f4485, %f1001, %f1000; mul.ftz.f32 %f4486, %f1003, %f998; sub.ftz.f32 %f4487, %f4486, %f4485; mul.ftz.f32 %f4488, %f1002, %f998; mul.ftz.f32 %f4489, %f1001, %f999; sub.ftz.f32 %f4490, %f4489, %f4488; add.ftz.f32 %f1008, %f4484, %f4484; add.ftz.f32 %f1009, %f4487, %f4487; add.ftz.f32 %f1010, %f4490, %f4490; mul.ftz.f32 %f4491, %f6907, %f1009; fma.rn.ftz.f32 %f4492, %f6908, %f1008, %f4491; fma.rn.ftz.f32 %f1011, %f6906, %f1010, %f4492; setp.eq.ftz.f32 %p400, %f1011, 0f00000000; mov.pred %p794, %p395; @%p400 bra $L__BB24_291; mul.ftz.f32 %f4493, %f996, %f1009; fma.rn.ftz.f32 %f4494, %f995, %f1008, %f4493; fma.rn.ftz.f32 %f4495, %f997, %f1010, %f4494; div.approx.ftz.f32 %f1012, %f4495, %f1011; setp.ltu.ftz.f32 %p402, %f1012, %f340; setp.gtu.ftz.f32 %p403, %f1012, %f341; or.pred %p404, %p403, %p402; mov.pred %p794, %p395; @%p404 bra $L__BB24_291; setp.lt.ftz.f32 %p405, %f1007, 0f219392EF; mov.f32 %f6850, 0f00000000; @%p405 bra $L__BB24_290; rcp.approx.ftz.f32 %f6850, %f1006; $L__BB24_290: mul.ftz.f32 %f4497, %f1004, %f6850; mov.f32 %f4498, 0f3F800000; min.ftz.f32 %f6863, %f4497, %f4498; mul.ftz.f32 %f4499, %f1005, %f6850; min.ftz.f32 %f6864, %f4499, %f4498; mov.pred %p794, 0; mov.f32 %f6862, %f1012; $L__BB24_291: mov.pred %p795, %p249; @%p794 bra $L__BB24_293; cvt.u32.u64 %r946, %rd143; add.s64 %rd663, %rd126, %rd583; ld.global.u32 %r947, [%rd663]; and.b32 %r948, %r947, %r668; setp.eq.s32 %p408, %r948, 0; selp.b32 %r1350, %r1350, 1, %p408; selp.b32 %r1351, %r1351, %r212, %p408; selp.b32 %r1352, %r1352, %r946, %p408; selp.f32 %f6865, %f6865, %f6864, %p408; selp.f32 %f6866, %f6866, %f6863, %p408; selp.f32 %f6867, %f6867, %f6862, %p408; setp.ne.s32 %p795, %r948, 0; bra.uni $L__BB24_293; $L__BB24_180: mul.wide.s32 %rd585, %r212, 208; add.s64 %rd586, %rd116, %rd585; ld.global.u32 %r213, [%rd586+156]; ld.global.u32 %r214, [%rd586+152]; shl.b32 %r866, %r214, 1; cvt.rn.f32.s32 %f3034, %r866; ld.local.f32 %f3035, [%rd4+32]; mul.ftz.f32 %f529, %f3035, %f3034; cvt.rzi.ftz.s32.f32 %r867, %f529; add.s32 %r868, %r866, -1; min.s32 %r215, %r867, %r868; ld.global.u32 %r1326, [%rd586+164]; mul.wide.u32 %rd587, %r1326, 16; add.s64 %rd588, %rd120, %rd587; ld.global.u32 %r1325, [%rd588+8]; ld.global.u16 %rs199, [%rd588+12]; ld.global.u64 %rd1197, [%rd588]; setp.eq.s64 %p261, %rd1197, 11; @%p261 bra $L__BB24_185; $L__BB24_181: setp.eq.s64 %p262, %rd1197, 0; @%p262 bra $L__BB24_183; add.s32 %r1326, %r1326, 2; bra.uni $L__BB24_184; $L__BB24_183: setp.eq.s16 %p263, %rs199, 0; mov.u32 %r1329, -1; mov.u32 %r1326, %r1325; @%p263 bra $L__BB24_186; $L__BB24_184: mul.wide.u32 %rd589, %r1326, 16; add.s64 %rd590, %rd120, %rd589; ld.global.u32 %r1325, [%rd590+8]; ld.global.u16 %rs199, [%rd590+12]; ld.global.u64 %rd1197, [%rd590]; setp.ne.s64 %p264, %rd1197, 11; @%p264 bra $L__BB24_181; $L__BB24_185: setp.eq.s16 %p265, %rs199, 0; selp.b32 %r1329, -1, %r1325, %p265; $L__BB24_186: shl.b64 %rd591, %rd143, 4; add.s64 %rd592, %rd118, %rd591; ld.global.v4.u32 {%r870, %r871, %r872, %r873}, [%rd592]; setp.eq.s32 %p266, %r215, %r214; @%p266 bra $L__BB24_188; bra.uni $L__BB24_187; $L__BB24_188: mul.wide.u32 %rd596, %r873, 12; add.s64 %rd1206, %rd119, %rd596; add.s32 %r880, %r873, 1; mul.wide.u32 %rd597, %r880, 12; add.s64 %rd1203, %rd119, %rd597; add.s32 %r881, %r873, 2; mul.wide.u32 %rd598, %r881, 12; add.s64 %rd1200, %rd119, %rd598; add.s64 %rd1205, %rd1206, 4; add.s64 %rd1204, %rd1206, 8; add.s64 %rd1202, %rd1203, 4; add.s64 %rd1201, %rd1203, 8; add.s64 %rd1199, %rd1200, 4; add.s64 %rd1198, %rd1200, 8; bra.uni $L__BB24_189; $L__BB24_177: mul.wide.s32 %rd611, %r212, 208; add.s64 %rd612, %rd116, %rd611; ld.global.u32 %r235, [%rd612+148]; ld.global.u32 %r236, [%rd612+152]; shl.b32 %r894, %r236, 1; cvt.rn.f32.s32 %f3149, %r894; ld.local.f32 %f3150, [%rd4+32]; mul.ftz.f32 %f567, %f3150, %f3149; cvt.rzi.ftz.s32.f32 %r895, %f567; add.s32 %r896, %r894, -1; min.s32 %r237, %r895, %r896; ld.global.u32 %r1331, [%rd612+164]; mul.wide.u32 %rd613, %r1331, 16; add.s64 %rd614, %rd120, %rd613; ld.global.u16 %rs201, [%rd614+12]; ld.global.u32 %r1330, [%rd614+8]; ld.global.u64 %rd1216, [%rd614]; setp.eq.s64 %p289, %rd1216, 11; @%p289 bra $L__BB24_202; $L__BB24_178: setp.eq.s64 %p290, %rd1216, 0; @%p290 bra $L__BB24_200; add.s32 %r1331, %r1331, 2; bra.uni $L__BB24_201; $L__BB24_200: setp.eq.s16 %p291, %rs201, 0; mov.u32 %r1334, -1; mov.u32 %r1331, %r1330; @%p291 bra $L__BB24_203; $L__BB24_201: mul.wide.u32 %rd615, %r1331, 16; add.s64 %rd616, %rd120, %rd615; ld.global.u16 %rs201, [%rd616+12]; ld.global.u32 %r1330, [%rd616+8]; ld.global.u64 %rd1216, [%rd616]; setp.ne.s64 %p292, %rd1216, 11; @%p292 bra $L__BB24_178; $L__BB24_202: setp.eq.s16 %p293, %rs201, 0; selp.b32 %r1334, -1, %r1330, %p293; $L__BB24_203: setp.eq.s32 %p294, %r237, %r236; @%p294 bra $L__BB24_205; bra.uni $L__BB24_204; $L__BB24_205: shl.b64 %rd618, %rd143, 4; add.s64 %rd1217, %rd128, %rd618; bra.uni $L__BB24_206; $L__BB24_218: setp.gt.s32 %p313, %r258, %r257; selp.b32 %r921, -1, 0, %p313; add.s32 %r922, %r258, %r921; mad.lo.s32 %r923, %r922, %r256, %r1339; add.s32 %r924, %r923, %r254; mul.wide.s32 %rd638, %r924, 16; add.s64 %rd639, %rd131, %rd638; ld.global.v4.f32 {%f6733, %f6732, %f6731, %f6730}, [%rd639]; add.s32 %r925, %r923, %r253; mul.wide.s32 %rd640, %r925, 16; add.s64 %rd641, %rd131, %rd640; ld.global.v4.f32 {%f6729, %f6728, %f6727, %f6726}, [%rd641]; ld.global.v4.f32 {%f6725, %f6724, %f6723, %f6722}, [%rd641+16]; add.s32 %r926, %r923, %r255; mul.wide.s32 %rd642, %r926, 16; add.s64 %rd643, %rd131, %rd642; ld.global.v4.f32 {%f6721, %f6720, %f6719, %f6718}, [%rd643]; $L__BB24_220: add.s32 %r269, %r258, 1; setp.eq.s32 %p314, %r269, %r257; @%p314 bra $L__BB24_222; bra.uni $L__BB24_221; $L__BB24_222: ld.global.v4.f32 {%f6749, %f6748, %f6747, %f6746}, [%rd210]; ld.global.v4.f32 {%f6745, %f6744, %f6743, %f6742}, [%rd211]; ld.global.v4.f32 {%f6741, %f6740, %f6739, %f6738}, [%rd211+16]; ld.global.v4.f32 {%f6737, %f6736, %f6735, %f6734}, [%rd212]; bra.uni $L__BB24_223; $L__BB24_221: setp.ge.s32 %p315, %r258, %r257; selp.b32 %r927, -1, 0, %p315; add.s32 %r928, %r269, %r927; mad.lo.s32 %r929, %r928, %r256, %r1339; add.s32 %r930, %r929, %r254; mul.wide.s32 %rd644, %r930, 16; add.s64 %rd645, %rd131, %rd644; ld.global.v4.f32 {%f6749, %f6748, %f6747, %f6746}, [%rd645]; add.s32 %r931, %r929, %r253; mul.wide.s32 %rd646, %r931, 16; add.s64 %rd647, %rd131, %rd646; ld.global.v4.f32 {%f6745, %f6744, %f6743, %f6742}, [%rd647]; ld.global.v4.f32 {%f6741, %f6740, %f6739, %f6738}, [%rd647+16]; add.s32 %r932, %r929, %r255; mul.wide.s32 %rd648, %r932, 16; add.s64 %rd649, %rd131, %rd648; ld.global.v4.f32 {%f6737, %f6736, %f6735, %f6734}, [%rd649]; $L__BB24_223: cvt.rn.f32.s32 %f3264, %r258; sub.ftz.f32 %f3265, %f591, %f3264; mov.f32 %f3266, 0f3F800000; sub.ftz.f32 %f3267, %f3266, %f3265; mul.ftz.f32 %f3268, %f3265, %f6749; fma.rn.ftz.f32 %f6765, %f3267, %f6733, %f3268; mul.ftz.f32 %f3269, %f3265, %f6748; fma.rn.ftz.f32 %f6764, %f3267, %f6732, %f3269; mul.ftz.f32 %f3270, %f3265, %f6747; fma.rn.ftz.f32 %f6763, %f3267, %f6731, %f3270; mul.ftz.f32 %f3271, %f3265, %f6746; fma.rn.ftz.f32 %f732, %f3267, %f6730, %f3271; mul.ftz.f32 %f3272, %f3265, %f6745; fma.rn.ftz.f32 %f6761, %f3267, %f6729, %f3272; mul.ftz.f32 %f3273, %f3265, %f6744; fma.rn.ftz.f32 %f6760, %f3267, %f6728, %f3273; mul.ftz.f32 %f3274, %f3265, %f6743; fma.rn.ftz.f32 %f6759, %f3267, %f6727, %f3274; mul.ftz.f32 %f3275, %f3265, %f6742; fma.rn.ftz.f32 %f728, %f3267, %f6726, %f3275; mul.ftz.f32 %f3276, %f3265, %f6741; fma.rn.ftz.f32 %f6757, %f3267, %f6725, %f3276; mul.ftz.f32 %f3277, %f3265, %f6740; fma.rn.ftz.f32 %f6756, %f3267, %f6724, %f3277; mul.ftz.f32 %f3278, %f3265, %f6739; fma.rn.ftz.f32 %f6755, %f3267, %f6723, %f3278; mul.ftz.f32 %f3279, %f3265, %f6738; fma.rn.ftz.f32 %f724, %f3267, %f6722, %f3279; mul.ftz.f32 %f3280, %f3265, %f6737; fma.rn.ftz.f32 %f6753, %f3267, %f6721, %f3280; mul.ftz.f32 %f3281, %f3265, %f6736; fma.rn.ftz.f32 %f6752, %f3267, %f6720, %f3281; mul.ftz.f32 %f3282, %f3265, %f6735; fma.rn.ftz.f32 %f6751, %f3267, %f6719, %f3282; mul.ftz.f32 %f3283, %f3265, %f6734; fma.rn.ftz.f32 %f720, %f3267, %f6718, %f3283; $L__BB24_225: and.b32 %r933, %r252, 4; setp.eq.s32 %p316, %r933, 0; @%p316 bra $L__BB24_241; bra.uni $L__BB24_226; $L__BB24_241: add.ftz.f32 %f3646, %f6761, %f6765; add.ftz.f32 %f3647, %f6757, %f3646; add.ftz.f32 %f3648, %f6760, %f6764; add.ftz.f32 %f3649, %f6756, %f3648; add.ftz.f32 %f3650, %f6759, %f6763; add.ftz.f32 %f3651, %f6755, %f3650; add.ftz.f32 %f3652, %f6753, %f3647; add.ftz.f32 %f3653, %f6752, %f3649; add.ftz.f32 %f3654, %f6751, %f3651; mul.ftz.f32 %f3655, %f3652, 0f3E800000; mul.ftz.f32 %f3656, %f3653, 0f3E800000; mul.ftz.f32 %f3657, %f3654, 0f3E800000; sub.ftz.f32 %f3658, %f3655, %f6905; sub.ftz.f32 %f3659, %f3656, %f6904; sub.ftz.f32 %f3660, %f3657, %f6903; mul.ftz.f32 %f3661, %f6908, %f3658; fma.rn.ftz.f32 %f3662, %f6907, %f3659, %f3661; fma.rn.ftz.f32 %f3663, %f6906, %f3660, %f3662; div.approx.ftz.f32 %f828, %f3663, %f511; fma.rn.ftz.f32 %f3664, %f6908, %f828, %f6905; fma.rn.ftz.f32 %f3665, %f6907, %f828, %f6904; fma.rn.ftz.f32 %f3666, %f6906, %f828, %f6903; sub.ftz.f32 %f829, %f6765, %f3664; sub.ftz.f32 %f830, %f6764, %f3665; sub.ftz.f32 %f831, %f6763, %f3666; sub.ftz.f32 %f832, %f6761, %f3664; sub.ftz.f32 %f833, %f6760, %f3665; sub.ftz.f32 %f834, %f6759, %f3666; sub.ftz.f32 %f835, %f6757, %f3664; sub.ftz.f32 %f836, %f6756, %f3665; sub.ftz.f32 %f837, %f6755, %f3666; sub.ftz.f32 %f838, %f6753, %f3664; sub.ftz.f32 %f839, %f6752, %f3665; sub.ftz.f32 %f840, %f6751, %f3666; sub.ftz.f32 %f841, %f340, %f828; mov.u32 %r1347, 0; mov.f32 %f6843, 0f3F800000; mov.f32 %f6842, 0f00000000; mov.f32 %f6838, %f341; mov.u16 %rs213, %rs117; mov.u32 %r1343, %r1347; bra.uni $L__BB24_242; $L__BB24_284: add.s32 %r1347, %r1347, -1; mul.wide.s32 %rd652, %r1347, 12; add.s64 %rd653, %rd1, %rd652; ld.local.f32 %f6842, [%rd653]; ld.local.f32 %f6843, [%rd653+4]; ld.local.u32 %r1343, [%rd653+8]; $L__BB24_242: setp.gt.s32 %p338, %r1343, 1; @%p338 bra $L__BB24_282; $L__BB24_243: mov.u32 %r1348, %r1343; cvt.rn.f32.s32 %f3669, %r1348; mul.ftz.f32 %f858, %f3669, 0f3F000000; sub.ftz.f32 %f859, %f6843, %f6842; mul.ftz.f32 %f3670, %f859, 0f3EAAAAAB; mul.ftz.f32 %f3671, %f3670, 0f3F000000; fma.rn.ftz.f32 %f860, %f859, %f858, %f6842; add.ftz.f32 %f3672, %f858, 0f3F000000; fma.rn.ftz.f32 %f861, %f859, %f3672, %f6842; mov.f32 %f3673, 0f3F800000; sub.ftz.f32 %f3674, %f3673, %f860; mul.ftz.f32 %f3675, %f860, %f3674; mul.ftz.f32 %f3676, %f3674, %f3675; mul.ftz.f32 %f3677, %f860, %f860; mul.ftz.f32 %f3678, %f860, 0f40400000; add.ftz.f32 %f3679, %f3678, 0fC0A00000; fma.rn.ftz.f32 %f3680, %f3677, %f3679, 0f40000000; mul.ftz.f32 %f3681, %f3674, %f3674; mul.ftz.f32 %f3682, %f3674, 0f40400000; add.ftz.f32 %f3683, %f3682, 0fC0A00000; fma.rn.ftz.f32 %f3684, %f3681, %f3683, 0f40000000; mul.ftz.f32 %f3685, %f860, %f3675; mul.ftz.f32 %f3686, %f829, %f3676; mul.ftz.f32 %f3687, %f830, %f3676; mul.ftz.f32 %f3688, %f831, %f3676; mul.ftz.f32 %f3689, %f732, %f3676; mul.ftz.f32 %f3690, %f832, %f3680; mul.ftz.f32 %f3691, %f833, %f3680; mul.ftz.f32 %f3692, %f834, %f3680; mul.ftz.f32 %f3693, %f728, %f3680; sub.ftz.f32 %f3694, %f3690, %f3686; sub.ftz.f32 %f3695, %f3691, %f3687; sub.ftz.f32 %f3696, %f3692, %f3688; sub.ftz.f32 %f3697, %f3693, %f3689; fma.rn.ftz.f32 %f3698, %f835, %f3684, %f3694; fma.rn.ftz.f32 %f3699, %f836, %f3684, %f3695; fma.rn.ftz.f32 %f3700, %f837, %f3684, %f3696; fma.rn.ftz.f32 %f3701, %f724, %f3684, %f3697; mul.ftz.f32 %f3702, %f838, %f3685; mul.ftz.f32 %f3703, %f839, %f3685; mul.ftz.f32 %f3704, %f840, %f3685; mul.ftz.f32 %f3705, %f720, %f3685; sub.ftz.f32 %f3706, %f3698, %f3702; sub.ftz.f32 %f3707, %f3699, %f3703; sub.ftz.f32 %f3708, %f3700, %f3704; sub.ftz.f32 %f3709, %f3701, %f3705; mul.ftz.f32 %f3710, %f3706, 0f3F000000; mul.ftz.f32 %f3711, %f3707, 0f3F000000; mul.ftz.f32 %f3712, %f3708, 0f3F000000; mul.ftz.f32 %f3713, %f3709, 0f3F000000; add.ftz.f32 %f3714, %f3674, %f3674; mul.ftz.f32 %f3715, %f860, %f3714; sub.ftz.f32 %f3716, %f3715, %f3681; add.ftz.f32 %f3717, %f860, %f860; mul.ftz.f32 %f3718, %f3717, %f3679; fma.rn.ftz.f32 %f3719, %f860, %f3678, %f3718; add.ftz.f32 %f3720, %f3678, 0f40000000; mul.ftz.f32 %f3721, %f3714, %f3720; mul.ftz.f32 %f3722, %f3674, %f3682; sub.ftz.f32 %f3723, %f3721, %f3722; mul.ftz.f32 %f3724, %f3674, 0fC0000000; fma.rn.ftz.f32 %f3725, %f860, %f3724, %f3677; mul.ftz.f32 %f3726, %f832, %f3719; mul.ftz.f32 %f3727, %f833, %f3719; mul.ftz.f32 %f3728, %f834, %f3719; mul.ftz.f32 %f3729, %f728, %f3719; fma.rn.ftz.f32 %f3730, %f829, %f3716, %f3726; fma.rn.ftz.f32 %f3731, %f830, %f3716, %f3727; fma.rn.ftz.f32 %f3732, %f831, %f3716, %f3728; fma.rn.ftz.f32 %f3733, %f732, %f3716, %f3729; fma.rn.ftz.f32 %f3734, %f835, %f3723, %f3730; fma.rn.ftz.f32 %f3735, %f836, %f3723, %f3731; fma.rn.ftz.f32 %f3736, %f837, %f3723, %f3732; fma.rn.ftz.f32 %f3737, %f724, %f3723, %f3733; fma.rn.ftz.f32 %f3738, %f838, %f3725, %f3734; fma.rn.ftz.f32 %f3739, %f839, %f3725, %f3735; fma.rn.ftz.f32 %f3740, %f840, %f3725, %f3736; fma.rn.ftz.f32 %f3741, %f720, %f3725, %f3737; mul.ftz.f32 %f3742, %f3738, 0f3F000000; mul.ftz.f32 %f3743, %f3739, 0f3F000000; mul.ftz.f32 %f3744, %f3740, 0f3F000000; mul.ftz.f32 %f3745, %f3741, 0f3F000000; mul.ftz.f32 %f862, %f3671, %f3742; mul.ftz.f32 %f863, %f3671, %f3743; mul.ftz.f32 %f864, %f3671, %f3744; sub.ftz.f32 %f3746, %f3673, %f861; mul.ftz.f32 %f3747, %f861, %f3746; mul.ftz.f32 %f3748, %f3746, %f3747; mul.ftz.f32 %f3749, %f861, %f861; mul.ftz.f32 %f3750, %f861, 0f40400000; add.ftz.f32 %f3751, %f3750, 0fC0A00000; fma.rn.ftz.f32 %f3752, %f3749, %f3751, 0f40000000; mul.ftz.f32 %f3753, %f3746, %f3746; mul.ftz.f32 %f3754, %f3746, 0f40400000; add.ftz.f32 %f3755, %f3754, 0fC0A00000; fma.rn.ftz.f32 %f3756, %f3753, %f3755, 0f40000000; mul.ftz.f32 %f3757, %f861, %f3747; mul.ftz.f32 %f3758, %f829, %f3748; mul.ftz.f32 %f3759, %f830, %f3748; mul.ftz.f32 %f3760, %f831, %f3748; mul.ftz.f32 %f3761, %f732, %f3748; mul.ftz.f32 %f3762, %f832, %f3752; mul.ftz.f32 %f3763, %f833, %f3752; mul.ftz.f32 %f3764, %f834, %f3752; mul.ftz.f32 %f3765, %f728, %f3752; sub.ftz.f32 %f3766, %f3762, %f3758; sub.ftz.f32 %f3767, %f3763, %f3759; sub.ftz.f32 %f3768, %f3764, %f3760; sub.ftz.f32 %f3769, %f3765, %f3761; fma.rn.ftz.f32 %f3770, %f835, %f3756, %f3766; fma.rn.ftz.f32 %f3771, %f836, %f3756, %f3767; fma.rn.ftz.f32 %f3772, %f837, %f3756, %f3768; fma.rn.ftz.f32 %f3773, %f724, %f3756, %f3769; mul.ftz.f32 %f3774, %f838, %f3757; mul.ftz.f32 %f3775, %f839, %f3757; mul.ftz.f32 %f3776, %f840, %f3757; mul.ftz.f32 %f3777, %f720, %f3757; sub.ftz.f32 %f3778, %f3770, %f3774; sub.ftz.f32 %f3779, %f3771, %f3775; sub.ftz.f32 %f3780, %f3772, %f3776; sub.ftz.f32 %f3781, %f3773, %f3777; mul.ftz.f32 %f865, %f3778, 0f3F000000; mul.ftz.f32 %f866, %f3779, 0f3F000000; mul.ftz.f32 %f867, %f3780, 0f3F000000; mul.ftz.f32 %f3782, %f3781, 0f3F000000; add.ftz.f32 %f3783, %f3746, %f3746; mul.ftz.f32 %f3784, %f861, %f3783; sub.ftz.f32 %f3785, %f3784, %f3753; add.ftz.f32 %f3786, %f861, %f861; mul.ftz.f32 %f3787, %f3786, %f3751; fma.rn.ftz.f32 %f3788, %f861, %f3750, %f3787; add.ftz.f32 %f3789, %f3750, 0f40000000; mul.ftz.f32 %f3790, %f3783, %f3789; mul.ftz.f32 %f3791, %f3746, %f3754; sub.ftz.f32 %f3792, %f3790, %f3791; mul.ftz.f32 %f3793, %f3746, 0fC0000000; fma.rn.ftz.f32 %f3794, %f861, %f3793, %f3749; mul.ftz.f32 %f3795, %f832, %f3788; mul.ftz.f32 %f3796, %f833, %f3788; mul.ftz.f32 %f3797, %f834, %f3788; mul.ftz.f32 %f3798, %f728, %f3788; fma.rn.ftz.f32 %f3799, %f829, %f3785, %f3795; fma.rn.ftz.f32 %f3800, %f830, %f3785, %f3796; fma.rn.ftz.f32 %f3801, %f831, %f3785, %f3797; fma.rn.ftz.f32 %f3802, %f732, %f3785, %f3798; fma.rn.ftz.f32 %f3803, %f835, %f3792, %f3799; fma.rn.ftz.f32 %f3804, %f836, %f3792, %f3800; fma.rn.ftz.f32 %f3805, %f837, %f3792, %f3801; fma.rn.ftz.f32 %f3806, %f724, %f3792, %f3802; fma.rn.ftz.f32 %f3807, %f838, %f3794, %f3803; fma.rn.ftz.f32 %f3808, %f839, %f3794, %f3804; fma.rn.ftz.f32 %f3809, %f840, %f3794, %f3805; fma.rn.ftz.f32 %f3810, %f720, %f3794, %f3806; mul.ftz.f32 %f3811, %f3807, 0f3F000000; mul.ftz.f32 %f3812, %f3808, 0f3F000000; mul.ftz.f32 %f3813, %f3809, 0f3F000000; mul.ftz.f32 %f3814, %f3810, 0f3F000000; mul.ftz.f32 %f868, %f3671, %f3811; mul.ftz.f32 %f869, %f3671, %f3812; mul.ftz.f32 %f870, %f3671, %f3813; mul.ftz.f32 %f3815, %f3671, %f3814; fma.rn.ftz.f32 %f3816, %f3671, %f3745, %f3713; sub.ftz.f32 %f3817, %f3782, %f3815; sub.ftz.f32 %f3818, %f865, %f3710; sub.ftz.f32 %f3819, %f866, %f3711; sub.ftz.f32 %f3820, %f867, %f3712; mul.ftz.f32 %f3821, %f863, %f3820; mul.ftz.f32 %f3822, %f864, %f3819; sub.ftz.f32 %f3823, %f3821, %f3822; mul.ftz.f32 %f3824, %f864, %f3818; mul.ftz.f32 %f3825, %f862, %f3820; sub.ftz.f32 %f3826, %f3824, %f3825; mul.ftz.f32 %f3827, %f862, %f3819; mul.ftz.f32 %f3828, %f863, %f3818; sub.ftz.f32 %f3829, %f3827, %f3828; mul.ftz.f32 %f3830, %f3826, %f3826; fma.rn.ftz.f32 %f3831, %f3823, %f3823, %f3830; fma.rn.ftz.f32 %f3832, %f3829, %f3829, %f3831; mul.ftz.f32 %f3833, %f3819, %f3819; fma.rn.ftz.f32 %f3834, %f3818, %f3818, %f3833; fma.rn.ftz.f32 %f3835, %f3820, %f3820, %f3834; div.approx.ftz.f32 %f3836, %f3832, %f3835; mul.ftz.f32 %f3837, %f869, %f3820; mul.ftz.f32 %f3838, %f870, %f3819; sub.ftz.f32 %f3839, %f3837, %f3838; mul.ftz.f32 %f3840, %f870, %f3818; mul.ftz.f32 %f3841, %f868, %f3820; sub.ftz.f32 %f3842, %f3840, %f3841; mul.ftz.f32 %f3843, %f868, %f3819; mul.ftz.f32 %f3844, %f869, %f3818; sub.ftz.f32 %f3845, %f3843, %f3844; mul.ftz.f32 %f3846, %f3842, %f3842; fma.rn.ftz.f32 %f3847, %f3839, %f3839, %f3846; fma.rn.ftz.f32 %f3848, %f3845, %f3845, %f3847; div.approx.ftz.f32 %f3849, %f3848, %f3835; max.ftz.f32 %f3850, %f3836, %f3849; sqrt.approx.ftz.f32 %f3851, %f3850; max.ftz.f32 %f3852, %f3713, %f3816; max.ftz.f32 %f3853, %f3817, %f3782; max.ftz.f32 %f3854, %f3852, %f3853; add.ftz.f32 %f3855, %f3851, %f3854; min.ftz.f32 %f3856, %f3713, %f3816; min.ftz.f32 %f3857, %f3817, %f3782; min.ftz.f32 %f3858, %f3856, %f3857; sub.ftz.f32 %f871, %f3858, %f3851; mul.ftz.f32 %f3859, %f3855, 0f3F800002; rsqrt.approx.ftz.f32 %f872, %f3835; mul.ftz.f32 %f3860, %f3818, %f872; mul.ftz.f32 %f3861, %f3819, %f872; mul.ftz.f32 %f3862, %f3820, %f872; mul.ftz.f32 %f873, %f3706, 0fBF000000; mul.ftz.f32 %f874, %f3707, 0fBF000000; mul.ftz.f32 %f875, %f3708, 0fBF000000; mul.ftz.f32 %f3863, %f6908, %f873; mul.ftz.f32 %f3864, %f6907, %f3711; sub.ftz.f32 %f3865, %f3863, %f3864; mul.ftz.f32 %f3866, %f6906, %f3712; sub.ftz.f32 %f3867, %f3865, %f3866; mul.ftz.f32 %f3868, %f3711, %f3711; fma.rn.ftz.f32 %f3869, %f3710, %f3710, %f3868; fma.rn.ftz.f32 %f3870, %f3712, %f3712, %f3869; mul.ftz.f32 %f3871, %f6907, %f3861; fma.rn.ftz.f32 %f3872, %f6908, %f3860, %f3871; fma.rn.ftz.f32 %f876, %f6906, %f3862, %f3872; mul.ftz.f32 %f3873, %f3860, %f873; mul.ftz.f32 %f3874, %f3861, %f3711; sub.ftz.f32 %f3875, %f3873, %f3874; mul.ftz.f32 %f3876, %f3862, %f3712; sub.ftz.f32 %f877, %f3875, %f3876; mul.ftz.f32 %f878, %f876, %f876; sub.ftz.f32 %f879, %f511, %f878; mul.ftz.f32 %f3877, %f876, %f877; sub.ftz.f32 %f3878, %f3867, %f3877; add.ftz.f32 %f880, %f3878, %f3878; mul.ftz.f32 %f3879, %f877, %f877; sub.ftz.f32 %f881, %f3870, %f3879; mul.ftz.f32 %f3880, %f3859, %f3859; sub.ftz.f32 %f882, %f881, %f3880; mul.ftz.f32 %f883, %f880, %f880; mul.ftz.f32 %f884, %f879, 0f40800000; mul.ftz.f32 %f3881, %f884, %f882; sub.ftz.f32 %f885, %f883, %f3881; setp.ltu.ftz.f32 %p340, %f885, 0f00000000; mov.pred %p793, -1; mov.f32 %f6811, 0fFF7FFFFF; mov.f32 %f6810, 0f7F7FFFFF; @%p340 bra $L__BB24_247; abs.ftz.f32 %f3882, %f511; abs.ftz.f32 %f3883, %f878; max.ftz.f32 %f3884, %f3882, %f3883; mul.ftz.f32 %f3885, %f3884, 0f36000000; abs.ftz.f32 %f3886, %f879; setp.lt.ftz.f32 %p341, %f3886, %f3885; @%p341 bra $L__BB24_246; bra.uni $L__BB24_245; $L__BB24_246: setp.gtu.ftz.f32 %p793, %f882, 0f00000000; mov.f32 %f6811, 0f7F7FFFFF; mov.f32 %f6810, 0fFF7FFFFF; bra.uni $L__BB24_247; $L__BB24_245: sqrt.approx.ftz.f32 %f3887, %f885; add.ftz.f32 %f3888, %f879, %f879; rcp.approx.ftz.f32 %f3889, %f3888; neg.ftz.f32 %f3890, %f880; sub.ftz.f32 %f3891, %f3890, %f3887; mul.ftz.f32 %f6810, %f3889, %f3891; sub.ftz.f32 %f3892, %f3887, %f880; mul.ftz.f32 %f6811, %f3889, %f3892; fma.rn.ftz.f32 %f3893, %f876, %f6810, %f877; mul.ftz.f32 %f6837, %f872, %f3893; fma.rn.ftz.f32 %f3894, %f876, %f6811, %f877; mul.ftz.f32 %f6836, %f872, %f3894; mov.pred %p793, 0; $L__BB24_247: @%p793 bra $L__BB24_281; sub.ftz.f32 %f3897, %f6838, %f828; min.ftz.f32 %f3898, %f3897, %f6811; mul.ftz.f32 %f3899, %f863, %f874; fma.rn.ftz.f32 %f3900, %f862, %f873, %f3899; fma.rn.ftz.f32 %f3901, %f864, %f875, %f3900; mul.ftz.f32 %f3902, %f6907, %f863; fma.rn.ftz.f32 %f3903, %f6908, %f862, %f3902; fma.rn.ftz.f32 %f3904, %f6906, %f864, %f3903; abs.ftz.f32 %f3905, %f3904; setp.lt.ftz.f32 %p343, %f3905, 0f219392EF; neg.ftz.f32 %f3906, %f3901; div.approx.ftz.f32 %f3907, %f3906, %f3904; setp.lt.ftz.f32 %p344, %f3904, 0f00000000; or.pred %p345, %p344, %p343; selp.f32 %f3908, 0fFF7FFFFF, %f3907, %p345; setp.gt.ftz.f32 %p346, %f3904, 0f00000000; or.pred %p347, %p346, %p343; selp.f32 %f3909, 0f7F7FFFFF, %f3907, %p347; max.ftz.f32 %f3910, %f841, %f6810; max.ftz.f32 %f3911, %f3910, %f3908; min.ftz.f32 %f3912, %f3898, %f3909; mul.ftz.f32 %f3913, %f866, %f869; fma.rn.ftz.f32 %f3914, %f865, %f868, %f3913; fma.rn.ftz.f32 %f3915, %f867, %f870, %f3914; mul.ftz.f32 %f3916, %f6907, %f869; neg.ftz.f32 %f3917, %f3916; mul.ftz.f32 %f3918, %f6908, %f868; sub.ftz.f32 %f3919, %f3917, %f3918; mul.ftz.f32 %f3920, %f6906, %f870; sub.ftz.f32 %f3921, %f3919, %f3920; abs.ftz.f32 %f3922, %f3921; setp.lt.ftz.f32 %p348, %f3922, 0f219392EF; neg.ftz.f32 %f3923, %f3915; div.approx.ftz.f32 %f3924, %f3923, %f3921; setp.lt.ftz.f32 %p349, %f3921, 0f00000000; or.pred %p350, %p349, %p348; selp.f32 %f3925, 0fFF7FFFFF, %f3924, %p350; setp.gt.ftz.f32 %p351, %f3921, 0f00000000; or.pred %p352, %p351, %p348; selp.f32 %f3926, 0f7F7FFFFF, %f3924, %p352; max.ftz.f32 %f6814, %f3911, %f3925; min.ftz.f32 %f6825, %f3912, %f3926; setp.gtu.ftz.f32 %p353, %f6814, %f6825; @%p353 bra $L__BB24_281; mul.ftz.f32 %f3929, %f871, 0f3F7FFFFC; mov.f32 %f3930, 0f00000000; max.ftz.f32 %f3931, %f3930, %f3929; max.ftz.f32 %f3932, %f6837, %f3930; min.ftz.f32 %f3934, %f3932, %f3673; max.ftz.f32 %f3935, %f6836, %f3930; min.ftz.f32 %f3936, %f3935, %f3673; add.ftz.f32 %f3937, %f858, %f3934; mul.ftz.f32 %f3938, %f3937, 0f3EAAAAAB; fma.rn.ftz.f32 %f6837, %f859, %f3938, %f6842; add.ftz.f32 %f3939, %f858, %f3936; mul.ftz.f32 %f3940, %f3939, 0f3EAAAAAB; fma.rn.ftz.f32 %f6836, %f859, %f3940, %f6842; mul.ftz.f32 %f3941, %f3931, %f3931; sub.ftz.f32 %f3942, %f881, %f3941; mul.ftz.f32 %f3943, %f884, %f3942; sub.ftz.f32 %f898, %f883, %f3943; setp.ltu.ftz.f32 %p354, %f898, 0f00000000; mov.f32 %f6813, 0fFF7FFFFF; mov.f32 %f6812, 0f7F7FFFFF; @%p354 bra $L__BB24_252; abs.ftz.f32 %f3946, %f511; abs.ftz.f32 %f3947, %f878; max.ftz.f32 %f3948, %f3946, %f3947; mul.ftz.f32 %f3949, %f3948, 0f36000000; abs.ftz.f32 %f3950, %f879; setp.lt.ftz.f32 %p355, %f3950, %f3949; mov.f32 %f6813, 0f7F7FFFFF; mov.f32 %f6812, 0fFF7FFFFF; @%p355 bra $L__BB24_252; sqrt.approx.ftz.f32 %f3951, %f898; add.ftz.f32 %f3952, %f879, %f879; rcp.approx.ftz.f32 %f3953, %f3952; neg.ftz.f32 %f3954, %f880; sub.ftz.f32 %f3955, %f3954, %f3951; mul.ftz.f32 %f6812, %f3953, %f3955; sub.ftz.f32 %f3956, %f3951, %f880; mul.ftz.f32 %f6813, %f3953, %f3956; $L__BB24_252: min.ftz.f32 %f903, %f6825, %f6812; setp.gtu.ftz.f32 %p356, %f6814, %f903; max.ftz.f32 %f904, %f6814, %f6813; setp.gtu.ftz.f32 %p357, %f904, %f6825; and.pred %p358, %p356, %p357; @%p358 bra $L__BB24_281; mov.u16 %rs118, 0; mov.u16 %rs209, %rs118; @%p356 bra $L__BB24_266; setp.lt.s32 %p360, %r1347, 4; mov.u16 %rs209, 1; @%p360 bra $L__BB24_266; sqrt.approx.ftz.f32 %f3957, %f511; min.ftz.f32 %f3958, %f835, %f838; min.ftz.f32 %f3959, %f829, %f832; min.ftz.f32 %f3960, %f3959, %f3958; min.ftz.f32 %f3961, %f836, %f839; min.ftz.f32 %f3962, %f830, %f833; min.ftz.f32 %f3963, %f3962, %f3961; min.ftz.f32 %f3964, %f837, %f840; min.ftz.f32 %f3965, %f831, %f834; min.ftz.f32 %f3966, %f3965, %f3964; min.ftz.f32 %f3967, %f724, %f720; min.ftz.f32 %f3968, %f732, %f728; min.ftz.f32 %f3969, %f3968, %f3967; max.ftz.f32 %f3970, %f835, %f838; max.ftz.f32 %f3971, %f3959, %f3970; max.ftz.f32 %f3972, %f836, %f839; max.ftz.f32 %f3973, %f3962, %f3972; max.ftz.f32 %f3974, %f837, %f840; max.ftz.f32 %f3975, %f3965, %f3974; max.ftz.f32 %f3976, %f724, %f720; max.ftz.f32 %f3977, %f3968, %f3976; abs.ftz.f32 %f3978, %f3960; abs.ftz.f32 %f3979, %f3963; abs.ftz.f32 %f3980, %f3966; abs.ftz.f32 %f3981, %f3969; abs.ftz.f32 %f3982, %f3971; abs.ftz.f32 %f3983, %f3973; abs.ftz.f32 %f3984, %f3975; abs.ftz.f32 %f3985, %f3977; max.ftz.f32 %f3986, %f3978, %f3982; max.ftz.f32 %f3987, %f3979, %f3983; max.ftz.f32 %f3988, %f3980, %f3984; max.ftz.f32 %f3989, %f3981, %f3985; max.ftz.f32 %f3990, %f3988, %f3989; max.ftz.f32 %f3991, %f3987, %f3990; max.ftz.f32 %f3992, %f3986, %f3991; mul.ftz.f32 %f905, %f3992, 0f36000000; mul.ftz.f32 %f906, %f3957, 0f36000000; mul.ftz.f32 %f907, %f3977, 0f36000000; mov.u32 %r1345, 0; mov.f32 %f6815, %f6837; $L__BB24_256: sub.ftz.f32 %f3995, %f3673, %f6815; mul.ftz.f32 %f3996, %f6815, %f3995; mul.ftz.f32 %f3997, %f3995, %f3996; mul.ftz.f32 %f3998, %f6815, 0f40400000; add.ftz.f32 %f3999, %f3998, 0fC0A00000; mul.ftz.f32 %f4000, %f6815, %f6815; fma.rn.ftz.f32 %f4001, %f4000, %f3999, 0f40000000; mov.f32 %f4002, 0f40000000; mul.ftz.f32 %f4003, %f3995, %f3995; mul.ftz.f32 %f4004, %f3995, 0f40400000; add.ftz.f32 %f4005, %f4004, 0fC0A00000; fma.rn.ftz.f32 %f4006, %f4003, %f4005, 0f40000000; mul.ftz.f32 %f4007, %f6815, %f3996; mul.ftz.f32 %f4008, %f829, %f3997; mul.ftz.f32 %f4009, %f830, %f3997; mul.ftz.f32 %f4010, %f831, %f3997; mul.ftz.f32 %f4011, %f732, %f3997; mul.ftz.f32 %f4012, %f832, %f4001; mul.ftz.f32 %f4013, %f833, %f4001; mul.ftz.f32 %f4014, %f834, %f4001; mul.ftz.f32 %f4015, %f728, %f4001; sub.ftz.f32 %f4016, %f4012, %f4008; sub.ftz.f32 %f4017, %f4013, %f4009; sub.ftz.f32 %f4018, %f4014, %f4010; sub.ftz.f32 %f4019, %f4015, %f4011; fma.rn.ftz.f32 %f4020, %f835, %f4006, %f4016; fma.rn.ftz.f32 %f4021, %f836, %f4006, %f4017; fma.rn.ftz.f32 %f4022, %f837, %f4006, %f4018; fma.rn.ftz.f32 %f4023, %f724, %f4006, %f4019; mul.ftz.f32 %f4024, %f838, %f4007; mul.ftz.f32 %f4025, %f839, %f4007; mul.ftz.f32 %f4026, %f840, %f4007; mul.ftz.f32 %f4027, %f720, %f4007; sub.ftz.f32 %f4028, %f4020, %f4024; sub.ftz.f32 %f4029, %f4021, %f4025; sub.ftz.f32 %f4030, %f4022, %f4026; sub.ftz.f32 %f4031, %f4023, %f4027; mul.ftz.f32 %f4032, %f4028, 0f3F000000; mul.ftz.f32 %f4033, %f4029, 0f3F000000; mul.ftz.f32 %f4034, %f4030, 0f3F000000; mul.ftz.f32 %f910, %f4031, 0f3F000000; add.ftz.f32 %f4035, %f3995, %f3995; mul.ftz.f32 %f4036, %f6815, %f4035; sub.ftz.f32 %f4037, %f4036, %f4003; add.ftz.f32 %f4038, %f6815, %f6815; mul.ftz.f32 %f4039, %f4038, %f3999; fma.rn.ftz.f32 %f4040, %f6815, %f3998, %f4039; add.ftz.f32 %f4041, %f3998, 0f40000000; mul.ftz.f32 %f4042, %f4035, %f4041; mul.ftz.f32 %f4043, %f3995, %f4004; sub.ftz.f32 %f4044, %f4042, %f4043; mul.ftz.f32 %f4045, %f3995, 0fC0000000; fma.rn.ftz.f32 %f4046, %f6815, %f4045, %f4000; mul.ftz.f32 %f4047, %f832, %f4040; mul.ftz.f32 %f4048, %f833, %f4040; mul.ftz.f32 %f4049, %f834, %f4040; mul.ftz.f32 %f4050, %f728, %f4040; fma.rn.ftz.f32 %f4051, %f829, %f4037, %f4047; fma.rn.ftz.f32 %f4052, %f830, %f4037, %f4048; fma.rn.ftz.f32 %f4053, %f831, %f4037, %f4049; fma.rn.ftz.f32 %f4054, %f732, %f4037, %f4050; fma.rn.ftz.f32 %f4055, %f835, %f4044, %f4051; fma.rn.ftz.f32 %f4056, %f836, %f4044, %f4052; fma.rn.ftz.f32 %f4057, %f837, %f4044, %f4053; fma.rn.ftz.f32 %f4058, %f724, %f4044, %f4054; fma.rn.ftz.f32 %f4059, %f838, %f4046, %f4055; fma.rn.ftz.f32 %f4060, %f839, %f4046, %f4056; fma.rn.ftz.f32 %f4061, %f840, %f4046, %f4057; fma.rn.ftz.f32 %f4062, %f720, %f4046, %f4058; mul.ftz.f32 %f911, %f4059, 0f3F000000; mul.ftz.f32 %f912, %f4060, 0f3F000000; mul.ftz.f32 %f913, %f4061, 0f3F000000; mul.ftz.f32 %f914, %f4062, 0f3F000000; sub.ftz.f32 %f4063, %f4002, %f3998; mul.ftz.f32 %f4064, %f6815, 0f41100000; add.ftz.f32 %f4065, %f4064, 0fC0A00000; mov.f32 %f4066, 0f40800000; sub.ftz.f32 %f4067, %f4066, %f4064; add.ftz.f32 %f4068, %f3998, 0fBF800000; mul.ftz.f32 %f4069, %f832, %f4065; mul.ftz.f32 %f4070, %f833, %f4065; mul.ftz.f32 %f4071, %f834, %f4065; fma.rn.ftz.f32 %f4072, %f829, %f4063, %f4069; fma.rn.ftz.f32 %f4073, %f830, %f4063, %f4070; fma.rn.ftz.f32 %f4074, %f831, %f4063, %f4071; fma.rn.ftz.f32 %f4075, %f835, %f4067, %f4072; fma.rn.ftz.f32 %f4076, %f836, %f4067, %f4073; fma.rn.ftz.f32 %f4077, %f837, %f4067, %f4074; fma.rn.ftz.f32 %f4078, %f838, %f4068, %f4075; fma.rn.ftz.f32 %f4079, %f839, %f4068, %f4076; fma.rn.ftz.f32 %f4080, %f840, %f4068, %f4077; mul.ftz.f32 %f4081, %f6908, %f6814; sub.ftz.f32 %f915, %f4081, %f4032; mul.ftz.f32 %f4082, %f6907, %f6814; sub.ftz.f32 %f916, %f4082, %f4033; mul.ftz.f32 %f4083, %f6906, %f6814; sub.ftz.f32 %f917, %f4083, %f4034; mul.ftz.f32 %f4084, %f916, %f916; fma.rn.ftz.f32 %f4085, %f915, %f915, %f4084; fma.rn.ftz.f32 %f4086, %f917, %f917, %f4085; sqrt.approx.ftz.f32 %f918, %f4086; mul.ftz.f32 %f4087, %f906, %f6814; max.ftz.f32 %f919, %f4087, %f905; mul.ftz.f32 %f4088, %f4059, 0fBF000000; mul.ftz.f32 %f4089, %f912, %f912; fma.rn.ftz.f32 %f4090, %f911, %f911, %f4089; fma.rn.ftz.f32 %f4091, %f913, %f913, %f4090; sqrt.approx.ftz.f32 %f4092, %f4091; div.approx.ftz.f32 %f4093, %f911, %f4092; div.approx.ftz.f32 %f4094, %f912, %f4092; div.approx.ftz.f32 %f4095, %f913, %f4092; mul.ftz.f32 %f4096, %f4079, %f912; fma.rn.ftz.f32 %f4097, %f4078, %f911, %f4096; fma.rn.ftz.f32 %f4098, %f4080, %f913, %f4097; mul.ftz.f32 %f4099, %f4078, %f4091; mul.ftz.f32 %f4100, %f4079, %f4091; mul.ftz.f32 %f4101, %f4080, %f4091; mul.ftz.f32 %f4102, %f911, %f4098; mul.ftz.f32 %f4103, %f912, %f4098; mul.ftz.f32 %f4104, %f913, %f4098; sub.ftz.f32 %f4105, %f4099, %f4102; sub.ftz.f32 %f4106, %f4100, %f4103; sub.ftz.f32 %f4107, %f4101, %f4104; mul.ftz.f32 %f4108, %f4091, %f4092; div.approx.ftz.f32 %f4109, %f4105, %f4108; div.approx.ftz.f32 %f4110, %f4106, %f4108; div.approx.ftz.f32 %f4111, %f4107, %f4108; div.approx.ftz.f32 %f4112, %f905, %f4092; mul.ftz.f32 %f4113, %f916, %f4094; fma.rn.ftz.f32 %f4114, %f915, %f4093, %f4113; fma.rn.ftz.f32 %f920, %f917, %f4095, %f4114; fma.rn.ftz.f32 %f4115, %f905, %f918, %f919; add.ftz.f32 %f4116, %f918, 0f3F800000; fma.rn.ftz.f32 %f921, %f4116, %f4112, %f4115; mul.ftz.f32 %f4117, %f4093, %f4088; mul.ftz.f32 %f4118, %f4094, %f912; sub.ftz.f32 %f4119, %f4117, %f4118; mul.ftz.f32 %f4120, %f4095, %f913; sub.ftz.f32 %f4121, %f4119, %f4120; mul.ftz.f32 %f4122, %f916, %f4110; fma.rn.ftz.f32 %f4123, %f915, %f4109, %f4122; fma.rn.ftz.f32 %f4124, %f917, %f4111, %f4123; add.ftz.f32 %f922, %f4121, %f4124; mul.ftz.f32 %f4125, %f6907, %f4094; fma.rn.ftz.f32 %f4126, %f6908, %f4093, %f4125; fma.rn.ftz.f32 %f923, %f6906, %f4095, %f4126; mul.ftz.f32 %f4127, %f920, %f920; sub.ftz.f32 %f4128, %f4086, %f4127; mul.ftz.f32 %f4129, %f915, %f4088; mul.ftz.f32 %f4130, %f916, %f912; sub.ftz.f32 %f4131, %f4129, %f4130; mul.ftz.f32 %f4132, %f917, %f913; sub.ftz.f32 %f924, %f4131, %f4132; mul.ftz.f32 %f4133, %f6907, %f916; fma.rn.ftz.f32 %f4134, %f6908, %f915, %f4133; fma.rn.ftz.f32 %f925, %f6906, %f917, %f4134; setp.leu.ftz.f32 %p361, %f4128, 0f00000000; mov.f32 %f6816, 0f00000000; sqrt.approx.ftz.f32 %f926, %f4128; @%p361 bra $L__BB24_258; rcp.approx.ftz.f32 %f6816, %f926; $L__BB24_258: add.ftz.f32 %f4135, %f919, %f921; add.ftz.f32 %f929, %f907, %f4135; mul.ftz.f32 %f4136, %f920, %f922; sub.ftz.f32 %f4137, %f924, %f4136; mul.ftz.f32 %f4138, %f4137, %f6816; sub.ftz.f32 %f4139, %f4138, %f914; mul.ftz.f32 %f4140, %f920, %f923; sub.ftz.f32 %f4141, %f925, %f4140; mul.ftz.f32 %f4142, %f4141, %f6816; mul.ftz.f32 %f4143, %f922, %f4142; mul.ftz.f32 %f4144, %f923, %f4139; sub.ftz.f32 %f4145, %f4143, %f4144; rcp.approx.ftz.f32 %f4146, %f4145; mul.ftz.f32 %f4147, %f920, %f4142; sub.ftz.f32 %f930, %f926, %f910; mul.ftz.f32 %f4148, %f923, %f930; sub.ftz.f32 %f4149, %f4147, %f4148; mul.ftz.f32 %f4150, %f4149, %f4146; sub.ftz.f32 %f6815, %f6815, %f4150; mul.ftz.f32 %f4151, %f920, %f4139; mul.ftz.f32 %f4152, %f922, %f930; sub.ftz.f32 %f4153, %f4152, %f4151; mul.ftz.f32 %f4154, %f4153, %f4146; sub.ftz.f32 %f6814, %f6814, %f4154; abs.ftz.f32 %f4155, %f920; setp.geu.ftz.f32 %p362, %f4155, %f921; @%p362 bra $L__BB24_260; abs.ftz.f32 %f4156, %f930; setp.lt.ftz.f32 %p363, %f4156, %f929; @%p363 bra $L__BB24_261; $L__BB24_260: add.s32 %r1345, %r1345, 1; setp.lt.u32 %p364, %r1345, 5; mov.u16 %rs207, 0; @%p364 bra $L__BB24_256; bra.uni $L__BB24_265; $L__BB24_261: add.ftz.f32 %f933, %f828, %f6814; setp.ltu.ftz.f32 %p365, %f933, %f340; setp.gtu.ftz.f32 %p366, %f933, %f6838; or.pred %p367, %p365, %p366; mov.u16 %rs207, 0; @%p367 bra $L__BB24_265; setp.ltu.ftz.f32 %p368, %f6815, 0f00000000; setp.gtu.ftz.f32 %p369, %f6815, 0f3F800000; or.pred %p370, %p369, %p368; @%p370 bra $L__BB24_265; div.approx.ftz.f32 %f4157, %f915, %f918; div.approx.ftz.f32 %f4158, %f916, %f918; div.approx.ftz.f32 %f4159, %f917, %f918; fma.rn.ftz.f32 %f4160, %f914, %f4157, %f911; fma.rn.ftz.f32 %f4161, %f914, %f4158, %f912; fma.rn.ftz.f32 %f4162, %f914, %f4159, %f913; mul.ftz.f32 %f4163, %f912, %f4159; mul.ftz.f32 %f4164, %f913, %f4158; sub.ftz.f32 %f4165, %f4163, %f4164; mul.ftz.f32 %f4166, %f913, %f4157; mul.ftz.f32 %f4167, %f911, %f4159; sub.ftz.f32 %f4168, %f4166, %f4167; mul.ftz.f32 %f4169, %f911, %f4158; mul.ftz.f32 %f4170, %f912, %f4157; sub.ftz.f32 %f4171, %f4169, %f4170; mul.ftz.f32 %f4172, %f4168, %f4162; mul.ftz.f32 %f4173, %f4171, %f4161; sub.ftz.f32 %f4174, %f4172, %f4173; mul.ftz.f32 %f4175, %f4160, %f4171; mul.ftz.f32 %f4176, %f4165, %f4162; sub.ftz.f32 %f4177, %f4175, %f4176; mul.ftz.f32 %f4178, %f4161, %f4165; mul.ftz.f32 %f4179, %f4160, %f4168; sub.ftz.f32 %f4180, %f4178, %f4179; mul.ftz.f32 %f4181, %f6907, %f4177; fma.rn.ftz.f32 %f4182, %f6908, %f4174, %f4181; fma.rn.ftz.f32 %f4183, %f6906, %f4180, %f4182; setp.gt.ftz.f32 %p371, %f4183, 0f00000000; @%p371 bra $L__BB24_265; mov.u16 %rs207, 1; mov.f32 %f6865, 0f00000000; mov.f32 %f6838, %f933; mov.f32 %f6866, %f6815; mov.f32 %f6867, %f933; $L__BB24_265: or.b16 %rs126, %rs207, %rs213; and.b16 %rs127, %rs126, 255; setp.ne.s16 %p372, %rs127, 0; selp.u16 %rs213, 1, 0, %p372; mov.u16 %rs209, %rs118; $L__BB24_266: add.ftz.f32 %f4185, %f828, %f904; setp.ltu.ftz.f32 %p373, %f4185, %f340; setp.gtu.ftz.f32 %p374, %f4185, %f6838; or.pred %p375, %p373, %p374; or.pred %p377, %p357, %p375; mov.u16 %rs212, %rs209; @%p377 bra $L__BB24_279; setp.lt.s32 %p378, %r1347, 4; mov.u16 %rs212, 1; @%p378 bra $L__BB24_279; sqrt.approx.ftz.f32 %f4186, %f511; min.ftz.f32 %f4187, %f835, %f838; min.ftz.f32 %f4188, %f829, %f832; min.ftz.f32 %f4189, %f4188, %f4187; min.ftz.f32 %f4190, %f836, %f839; min.ftz.f32 %f4191, %f830, %f833; min.ftz.f32 %f4192, %f4191, %f4190; min.ftz.f32 %f4193, %f837, %f840; min.ftz.f32 %f4194, %f831, %f834; min.ftz.f32 %f4195, %f4194, %f4193; min.ftz.f32 %f4196, %f724, %f720; min.ftz.f32 %f4197, %f732, %f728; min.ftz.f32 %f4198, %f4197, %f4196; max.ftz.f32 %f4199, %f835, %f838; max.ftz.f32 %f4200, %f4188, %f4199; max.ftz.f32 %f4201, %f836, %f839; max.ftz.f32 %f4202, %f4191, %f4201; max.ftz.f32 %f4203, %f837, %f840; max.ftz.f32 %f4204, %f4194, %f4203; max.ftz.f32 %f4205, %f724, %f720; max.ftz.f32 %f4206, %f4197, %f4205; abs.ftz.f32 %f4207, %f4189; abs.ftz.f32 %f4208, %f4192; abs.ftz.f32 %f4209, %f4195; abs.ftz.f32 %f4210, %f4198; abs.ftz.f32 %f4211, %f4200; abs.ftz.f32 %f4212, %f4202; abs.ftz.f32 %f4213, %f4204; abs.ftz.f32 %f4214, %f4206; max.ftz.f32 %f4215, %f4207, %f4211; max.ftz.f32 %f4216, %f4208, %f4212; max.ftz.f32 %f4217, %f4209, %f4213; max.ftz.f32 %f4218, %f4210, %f4214; max.ftz.f32 %f4219, %f4217, %f4218; max.ftz.f32 %f4220, %f4216, %f4219; max.ftz.f32 %f4221, %f4215, %f4220; mul.ftz.f32 %f942, %f4221, 0f36000000; mul.ftz.f32 %f943, %f4186, 0f36000000; mul.ftz.f32 %f944, %f4206, 0f36000000; mov.u32 %r1346, 0; mov.f32 %f6826, %f6836; $L__BB24_269: sub.ftz.f32 %f4224, %f3673, %f6826; mul.ftz.f32 %f4225, %f6826, %f4224; mul.ftz.f32 %f4226, %f4224, %f4225; mul.ftz.f32 %f4227, %f6826, 0f40400000; add.ftz.f32 %f4228, %f4227, 0fC0A00000; mul.ftz.f32 %f4229, %f6826, %f6826; fma.rn.ftz.f32 %f4230, %f4229, %f4228, 0f40000000; mov.f32 %f4231, 0f40000000; mul.ftz.f32 %f4232, %f4224, %f4224; mul.ftz.f32 %f4233, %f4224, 0f40400000; add.ftz.f32 %f4234, %f4233, 0fC0A00000; fma.rn.ftz.f32 %f4235, %f4232, %f4234, 0f40000000; mul.ftz.f32 %f4236, %f6826, %f4225; mul.ftz.f32 %f4237, %f829, %f4226; mul.ftz.f32 %f4238, %f830, %f4226; mul.ftz.f32 %f4239, %f831, %f4226; mul.ftz.f32 %f4240, %f732, %f4226; mul.ftz.f32 %f4241, %f832, %f4230; mul.ftz.f32 %f4242, %f833, %f4230; mul.ftz.f32 %f4243, %f834, %f4230; mul.ftz.f32 %f4244, %f728, %f4230; sub.ftz.f32 %f4245, %f4241, %f4237; sub.ftz.f32 %f4246, %f4242, %f4238; sub.ftz.f32 %f4247, %f4243, %f4239; sub.ftz.f32 %f4248, %f4244, %f4240; fma.rn.ftz.f32 %f4249, %f835, %f4235, %f4245; fma.rn.ftz.f32 %f4250, %f836, %f4235, %f4246; fma.rn.ftz.f32 %f4251, %f837, %f4235, %f4247; fma.rn.ftz.f32 %f4252, %f724, %f4235, %f4248; mul.ftz.f32 %f4253, %f838, %f4236; mul.ftz.f32 %f4254, %f839, %f4236; mul.ftz.f32 %f4255, %f840, %f4236; mul.ftz.f32 %f4256, %f720, %f4236; sub.ftz.f32 %f4257, %f4249, %f4253; sub.ftz.f32 %f4258, %f4250, %f4254; sub.ftz.f32 %f4259, %f4251, %f4255; sub.ftz.f32 %f4260, %f4252, %f4256; mul.ftz.f32 %f4261, %f4257, 0f3F000000; mul.ftz.f32 %f4262, %f4258, 0f3F000000; mul.ftz.f32 %f4263, %f4259, 0f3F000000; mul.ftz.f32 %f947, %f4260, 0f3F000000; add.ftz.f32 %f4264, %f4224, %f4224; mul.ftz.f32 %f4265, %f6826, %f4264; sub.ftz.f32 %f4266, %f4265, %f4232; add.ftz.f32 %f4267, %f6826, %f6826; mul.ftz.f32 %f4268, %f4267, %f4228; fma.rn.ftz.f32 %f4269, %f6826, %f4227, %f4268; add.ftz.f32 %f4270, %f4227, 0f40000000; mul.ftz.f32 %f4271, %f4264, %f4270; mul.ftz.f32 %f4272, %f4224, %f4233; sub.ftz.f32 %f4273, %f4271, %f4272; mul.ftz.f32 %f4274, %f4224, 0fC0000000; fma.rn.ftz.f32 %f4275, %f6826, %f4274, %f4229; mul.ftz.f32 %f4276, %f832, %f4269; mul.ftz.f32 %f4277, %f833, %f4269; mul.ftz.f32 %f4278, %f834, %f4269; mul.ftz.f32 %f4279, %f728, %f4269; fma.rn.ftz.f32 %f4280, %f829, %f4266, %f4276; fma.rn.ftz.f32 %f4281, %f830, %f4266, %f4277; fma.rn.ftz.f32 %f4282, %f831, %f4266, %f4278; fma.rn.ftz.f32 %f4283, %f732, %f4266, %f4279; fma.rn.ftz.f32 %f4284, %f835, %f4273, %f4280; fma.rn.ftz.f32 %f4285, %f836, %f4273, %f4281; fma.rn.ftz.f32 %f4286, %f837, %f4273, %f4282; fma.rn.ftz.f32 %f4287, %f724, %f4273, %f4283; fma.rn.ftz.f32 %f4288, %f838, %f4275, %f4284; fma.rn.ftz.f32 %f4289, %f839, %f4275, %f4285; fma.rn.ftz.f32 %f4290, %f840, %f4275, %f4286; fma.rn.ftz.f32 %f4291, %f720, %f4275, %f4287; mul.ftz.f32 %f948, %f4288, 0f3F000000; mul.ftz.f32 %f949, %f4289, 0f3F000000; mul.ftz.f32 %f950, %f4290, 0f3F000000; mul.ftz.f32 %f951, %f4291, 0f3F000000; sub.ftz.f32 %f4292, %f4231, %f4227; mul.ftz.f32 %f4293, %f6826, 0f41100000; add.ftz.f32 %f4294, %f4293, 0fC0A00000; mov.f32 %f4295, 0f40800000; sub.ftz.f32 %f4296, %f4295, %f4293; add.ftz.f32 %f4297, %f4227, 0fBF800000; mul.ftz.f32 %f4298, %f832, %f4294; mul.ftz.f32 %f4299, %f833, %f4294; mul.ftz.f32 %f4300, %f834, %f4294; fma.rn.ftz.f32 %f4301, %f829, %f4292, %f4298; fma.rn.ftz.f32 %f4302, %f830, %f4292, %f4299; fma.rn.ftz.f32 %f4303, %f831, %f4292, %f4300; fma.rn.ftz.f32 %f4304, %f835, %f4296, %f4301; fma.rn.ftz.f32 %f4305, %f836, %f4296, %f4302; fma.rn.ftz.f32 %f4306, %f837, %f4296, %f4303; fma.rn.ftz.f32 %f4307, %f838, %f4297, %f4304; fma.rn.ftz.f32 %f4308, %f839, %f4297, %f4305; fma.rn.ftz.f32 %f4309, %f840, %f4297, %f4306; mul.ftz.f32 %f4310, %f6908, %f6825; sub.ftz.f32 %f952, %f4310, %f4261; mul.ftz.f32 %f4311, %f6907, %f6825; sub.ftz.f32 %f953, %f4311, %f4262; mul.ftz.f32 %f4312, %f6906, %f6825; sub.ftz.f32 %f954, %f4312, %f4263; mul.ftz.f32 %f4313, %f953, %f953; fma.rn.ftz.f32 %f4314, %f952, %f952, %f4313; fma.rn.ftz.f32 %f4315, %f954, %f954, %f4314; sqrt.approx.ftz.f32 %f955, %f4315; mul.ftz.f32 %f4316, %f943, %f6825; max.ftz.f32 %f956, %f4316, %f942; mul.ftz.f32 %f4317, %f4288, 0fBF000000; mul.ftz.f32 %f4318, %f949, %f949; fma.rn.ftz.f32 %f4319, %f948, %f948, %f4318; fma.rn.ftz.f32 %f4320, %f950, %f950, %f4319; sqrt.approx.ftz.f32 %f4321, %f4320; div.approx.ftz.f32 %f4322, %f948, %f4321; div.approx.ftz.f32 %f4323, %f949, %f4321; div.approx.ftz.f32 %f4324, %f950, %f4321; mul.ftz.f32 %f4325, %f4308, %f949; fma.rn.ftz.f32 %f4326, %f4307, %f948, %f4325; fma.rn.ftz.f32 %f4327, %f4309, %f950, %f4326; mul.ftz.f32 %f4328, %f4307, %f4320; mul.ftz.f32 %f4329, %f4308, %f4320; mul.ftz.f32 %f4330, %f4309, %f4320; mul.ftz.f32 %f4331, %f948, %f4327; mul.ftz.f32 %f4332, %f949, %f4327; mul.ftz.f32 %f4333, %f950, %f4327; sub.ftz.f32 %f4334, %f4328, %f4331; sub.ftz.f32 %f4335, %f4329, %f4332; sub.ftz.f32 %f4336, %f4330, %f4333; mul.ftz.f32 %f4337, %f4320, %f4321; div.approx.ftz.f32 %f4338, %f4334, %f4337; div.approx.ftz.f32 %f4339, %f4335, %f4337; div.approx.ftz.f32 %f4340, %f4336, %f4337; div.approx.ftz.f32 %f4341, %f942, %f4321; mul.ftz.f32 %f4342, %f953, %f4323; fma.rn.ftz.f32 %f4343, %f952, %f4322, %f4342; fma.rn.ftz.f32 %f957, %f954, %f4324, %f4343; fma.rn.ftz.f32 %f4344, %f942, %f955, %f956; add.ftz.f32 %f4345, %f955, 0f3F800000; fma.rn.ftz.f32 %f958, %f4345, %f4341, %f4344; mul.ftz.f32 %f4346, %f4322, %f4317; mul.ftz.f32 %f4347, %f4323, %f949; sub.ftz.f32 %f4348, %f4346, %f4347; mul.ftz.f32 %f4349, %f4324, %f950; sub.ftz.f32 %f4350, %f4348, %f4349; mul.ftz.f32 %f4351, %f953, %f4339; fma.rn.ftz.f32 %f4352, %f952, %f4338, %f4351; fma.rn.ftz.f32 %f4353, %f954, %f4340, %f4352; add.ftz.f32 %f959, %f4350, %f4353; mul.ftz.f32 %f4354, %f6907, %f4323; fma.rn.ftz.f32 %f4355, %f6908, %f4322, %f4354; fma.rn.ftz.f32 %f960, %f6906, %f4324, %f4355; mul.ftz.f32 %f4356, %f957, %f957; sub.ftz.f32 %f4357, %f4315, %f4356; mul.ftz.f32 %f4358, %f952, %f4317; mul.ftz.f32 %f4359, %f953, %f949; sub.ftz.f32 %f4360, %f4358, %f4359; mul.ftz.f32 %f4361, %f954, %f950; sub.ftz.f32 %f961, %f4360, %f4361; mul.ftz.f32 %f4362, %f6907, %f953; fma.rn.ftz.f32 %f4363, %f6908, %f952, %f4362; fma.rn.ftz.f32 %f962, %f6906, %f954, %f4363; setp.leu.ftz.f32 %p379, %f4357, 0f00000000; sqrt.approx.ftz.f32 %f963, %f4357; mov.f32 %f6827, %f3930; @%p379 bra $L__BB24_271; rcp.approx.ftz.f32 %f6827, %f963; $L__BB24_271: add.ftz.f32 %f4364, %f956, %f958; add.ftz.f32 %f966, %f944, %f4364; mul.ftz.f32 %f4365, %f957, %f959; sub.ftz.f32 %f4366, %f961, %f4365; mul.ftz.f32 %f4367, %f4366, %f6827; sub.ftz.f32 %f4368, %f4367, %f951; mul.ftz.f32 %f4369, %f957, %f960; sub.ftz.f32 %f4370, %f962, %f4369; mul.ftz.f32 %f4371, %f4370, %f6827; mul.ftz.f32 %f4372, %f959, %f4371; mul.ftz.f32 %f4373, %f960, %f4368; sub.ftz.f32 %f4374, %f4372, %f4373; rcp.approx.ftz.f32 %f4375, %f4374; mul.ftz.f32 %f4376, %f957, %f4371; sub.ftz.f32 %f967, %f963, %f947; mul.ftz.f32 %f4377, %f960, %f967; sub.ftz.f32 %f4378, %f4376, %f4377; mul.ftz.f32 %f4379, %f4378, %f4375; sub.ftz.f32 %f6826, %f6826, %f4379; mul.ftz.f32 %f4380, %f957, %f4368; mul.ftz.f32 %f4381, %f959, %f967; sub.ftz.f32 %f4382, %f4381, %f4380; mul.ftz.f32 %f4383, %f4382, %f4375; sub.ftz.f32 %f6825, %f6825, %f4383; abs.ftz.f32 %f4384, %f957; setp.geu.ftz.f32 %p380, %f4384, %f958; @%p380 bra $L__BB24_273; abs.ftz.f32 %f4385, %f967; setp.lt.ftz.f32 %p381, %f4385, %f966; @%p381 bra $L__BB24_274; $L__BB24_273: add.s32 %r1346, %r1346, 1; setp.lt.u32 %p382, %r1346, 5; mov.u16 %rs210, 0; @%p382 bra $L__BB24_269; bra.uni $L__BB24_278; $L__BB24_274: add.ftz.f32 %f970, %f828, %f6825; setp.ltu.ftz.f32 %p383, %f970, %f340; setp.gtu.ftz.f32 %p384, %f970, %f6838; or.pred %p385, %p383, %p384; mov.u16 %rs210, 0; @%p385 bra $L__BB24_278; setp.ltu.ftz.f32 %p386, %f6826, 0f00000000; setp.gtu.ftz.f32 %p387, %f6826, 0f3F800000; or.pred %p388, %p387, %p386; @%p388 bra $L__BB24_278; div.approx.ftz.f32 %f4386, %f952, %f955; div.approx.ftz.f32 %f4387, %f953, %f955; div.approx.ftz.f32 %f4388, %f954, %f955; fma.rn.ftz.f32 %f4389, %f951, %f4386, %f948; fma.rn.ftz.f32 %f4390, %f951, %f4387, %f949; fma.rn.ftz.f32 %f4391, %f951, %f4388, %f950; mul.ftz.f32 %f4392, %f949, %f4388; mul.ftz.f32 %f4393, %f950, %f4387; sub.ftz.f32 %f4394, %f4392, %f4393; mul.ftz.f32 %f4395, %f950, %f4386; mul.ftz.f32 %f4396, %f948, %f4388; sub.ftz.f32 %f4397, %f4395, %f4396; mul.ftz.f32 %f4398, %f948, %f4387; mul.ftz.f32 %f4399, %f949, %f4386; sub.ftz.f32 %f4400, %f4398, %f4399; mul.ftz.f32 %f4401, %f4397, %f4391; mul.ftz.f32 %f4402, %f4400, %f4390; sub.ftz.f32 %f4403, %f4401, %f4402; mul.ftz.f32 %f4404, %f4389, %f4400; mul.ftz.f32 %f4405, %f4394, %f4391; sub.ftz.f32 %f4406, %f4404, %f4405; mul.ftz.f32 %f4407, %f4390, %f4394; mul.ftz.f32 %f4408, %f4389, %f4397; sub.ftz.f32 %f4409, %f4407, %f4408; mul.ftz.f32 %f4410, %f6907, %f4406; fma.rn.ftz.f32 %f4411, %f6908, %f4403, %f4410; fma.rn.ftz.f32 %f4412, %f6906, %f4409, %f4411; setp.gt.ftz.f32 %p389, %f4412, 0f00000000; @%p389 bra $L__BB24_278; mov.u16 %rs210, 1; mov.f32 %f6865, 0f00000000; mov.f32 %f6838, %f970; mov.f32 %f6866, %f6826; mov.f32 %f6867, %f970; $L__BB24_278: or.b16 %rs134, %rs210, %rs213; and.b16 %rs135, %rs134, 255; setp.ne.s16 %p390, %rs135, 0; selp.u16 %rs213, 1, 0, %p390; mov.u16 %rs212, %rs209; $L__BB24_279: setp.eq.s16 %p391, %rs212, 0; @%p391 bra $L__BB24_281; mul.wide.s32 %rd650, %r1347, 12; add.s64 %rd651, %rd1, %rd650; st.local.f32 [%rd651], %f6842; st.local.f32 [%rd651+4], %f6843; add.s32 %r941, %r1348, 1; st.local.u32 [%rd651+8], %r941; add.s32 %r1347, %r1347, 1; mov.u32 %r1348, -1; mov.f32 %f6842, %f860; mov.f32 %f6843, %f861; $L__BB24_281: add.s32 %r1343, %r1348, 1; setp.lt.s32 %p392, %r1348, 1; @%p392 bra $L__BB24_243; $L__BB24_282: setp.gt.s32 %p393, %r1347, 0; @%p393 bra $L__BB24_284; cvt.u32.u64 %r942, %rd143; and.b16 %rs136, %rs213, 255; setp.eq.s16 %p394, %rs136, 0; selp.b32 %r1350, %r1350, %r252, %p394; selp.b32 %r1351, %r1351, %r212, %p394; selp.b32 %r1352, %r1352, %r942, %p394; setp.ne.s16 %p795, %rs136, 0; bra.uni $L__BB24_293; $L__BB24_226: rsqrt.approx.ftz.f32 %f736, %f511; mov.f32 %f3300, 0f3F800000; mul.ftz.f32 %f3301, %f6908, %f736; mul.ftz.f32 %f3302, %f6907, %f736; mul.ftz.f32 %f3303, %f6906, %f736; neg.ftz.f32 %f3304, %f3302; neg.ftz.f32 %f3305, %f3303; fma.rn.ftz.f32 %f3306, %f3303, %f3303, 0f00000000; fma.rn.ftz.f32 %f3307, %f3302, %f3302, %f3306; fma.rn.ftz.f32 %f3308, %f3301, %f3301, %f3306; setp.gt.ftz.f32 %p317, %f3307, %f3308; selp.f32 %f3309, 0f00000000, %f3305, %p317; selp.f32 %f3310, %f3303, 0f00000000, %p317; selp.f32 %f3311, %f3304, %f3301, %p317; mul.ftz.f32 %f3312, %f3310, %f3310; fma.rn.ftz.f32 %f3313, %f3309, %f3309, %f3312; fma.rn.ftz.f32 %f3314, %f3311, %f3311, %f3313; sqrt.approx.ftz.f32 %f3315, %f3314; div.approx.ftz.f32 %f3316, %f3309, %f3315; div.approx.ftz.f32 %f3317, %f3310, %f3315; div.approx.ftz.f32 %f3318, %f3311, %f3315; mul.ftz.f32 %f3319, %f3302, %f3318; mul.ftz.f32 %f3320, %f3303, %f3317; sub.ftz.f32 %f3321, %f3319, %f3320; mul.ftz.f32 %f3322, %f3303, %f3316; mul.ftz.f32 %f3323, %f3301, %f3318; sub.ftz.f32 %f3324, %f3322, %f3323; mul.ftz.f32 %f3325, %f3301, %f3317; mul.ftz.f32 %f3326, %f3302, %f3316; sub.ftz.f32 %f3327, %f3325, %f3326; mul.ftz.f32 %f3328, %f3324, %f3324; fma.rn.ftz.f32 %f3329, %f3321, %f3321, %f3328; fma.rn.ftz.f32 %f3330, %f3327, %f3327, %f3329; sqrt.approx.ftz.f32 %f3331, %f3330; div.approx.ftz.f32 %f3332, %f3321, %f3331; div.approx.ftz.f32 %f3333, %f3324, %f3331; div.approx.ftz.f32 %f3334, %f3327, %f3331; mul.ftz.f32 %f3335, %f736, %f3301; mul.ftz.f32 %f3336, %f736, %f3302; mul.ftz.f32 %f3337, %f736, %f3303; sub.ftz.f32 %f3338, %f6765, %f6905; sub.ftz.f32 %f3339, %f6764, %f6904; mul.ftz.f32 %f3340, %f3339, %f3317; fma.rn.ftz.f32 %f3341, %f3338, %f3316, %f3340; sub.ftz.f32 %f3342, %f6763, %f6903; fma.rn.ftz.f32 %f737, %f3342, %f3318, %f3341; mul.ftz.f32 %f3343, %f3339, %f3333; fma.rn.ftz.f32 %f3344, %f3338, %f3332, %f3343; fma.rn.ftz.f32 %f738, %f3342, %f3334, %f3344; mul.ftz.f32 %f3345, %f3339, %f3336; fma.rn.ftz.f32 %f3346, %f3338, %f3335, %f3345; fma.rn.ftz.f32 %f739, %f3342, %f3337, %f3346; sub.ftz.f32 %f3347, %f6761, %f6905; sub.ftz.f32 %f3348, %f6760, %f6904; mul.ftz.f32 %f3349, %f3348, %f3317; fma.rn.ftz.f32 %f3350, %f3347, %f3316, %f3349; sub.ftz.f32 %f3351, %f6759, %f6903; fma.rn.ftz.f32 %f740, %f3351, %f3318, %f3350; mul.ftz.f32 %f3352, %f3348, %f3333; fma.rn.ftz.f32 %f3353, %f3347, %f3332, %f3352; fma.rn.ftz.f32 %f741, %f3351, %f3334, %f3353; mul.ftz.f32 %f3354, %f3348, %f3336; fma.rn.ftz.f32 %f3355, %f3347, %f3335, %f3354; fma.rn.ftz.f32 %f742, %f3351, %f3337, %f3355; sub.ftz.f32 %f3356, %f6757, %f6905; sub.ftz.f32 %f3357, %f6756, %f6904; mul.ftz.f32 %f3358, %f3357, %f3317; fma.rn.ftz.f32 %f3359, %f3356, %f3316, %f3358; sub.ftz.f32 %f3360, %f6755, %f6903; fma.rn.ftz.f32 %f743, %f3360, %f3318, %f3359; mul.ftz.f32 %f3361, %f3357, %f3333; fma.rn.ftz.f32 %f3362, %f3356, %f3332, %f3361; fma.rn.ftz.f32 %f744, %f3360, %f3334, %f3362; mul.ftz.f32 %f3363, %f3357, %f3336; fma.rn.ftz.f32 %f3364, %f3356, %f3335, %f3363; fma.rn.ftz.f32 %f745, %f3360, %f3337, %f3364; sub.ftz.f32 %f3365, %f6753, %f6905; sub.ftz.f32 %f3366, %f6752, %f6904; mul.ftz.f32 %f3367, %f3366, %f3317; fma.rn.ftz.f32 %f3368, %f3365, %f3316, %f3367; sub.ftz.f32 %f3369, %f6751, %f6903; fma.rn.ftz.f32 %f746, %f3369, %f3318, %f3368; mul.ftz.f32 %f3370, %f3366, %f3333; fma.rn.ftz.f32 %f3371, %f3365, %f3332, %f3370; fma.rn.ftz.f32 %f747, %f3369, %f3334, %f3371; mul.ftz.f32 %f3372, %f3366, %f3336; fma.rn.ftz.f32 %f3373, %f3365, %f3335, %f3372; fma.rn.ftz.f32 %f748, %f3369, %f3337, %f3373; abs.ftz.f32 %f3374, %f737; abs.ftz.f32 %f3375, %f738; abs.ftz.f32 %f3376, %f739; abs.ftz.f32 %f3377, %f740; abs.ftz.f32 %f3378, %f741; abs.ftz.f32 %f3379, %f742; max.ftz.f32 %f3380, %f3374, %f3377; max.ftz.f32 %f3381, %f3375, %f3378; max.ftz.f32 %f3382, %f3376, %f3379; abs.ftz.f32 %f3383, %f728; abs.ftz.f32 %f3384, %f732; max.ftz.f32 %f3385, %f3384, %f3383; abs.ftz.f32 %f3386, %f743; abs.ftz.f32 %f3387, %f744; abs.ftz.f32 %f3388, %f745; abs.ftz.f32 %f3389, %f746; abs.ftz.f32 %f3390, %f747; abs.ftz.f32 %f3391, %f748; max.ftz.f32 %f3392, %f3386, %f3389; max.ftz.f32 %f3393, %f3387, %f3390; max.ftz.f32 %f3394, %f3388, %f3391; abs.ftz.f32 %f3395, %f720; abs.ftz.f32 %f3396, %f724; max.ftz.f32 %f3397, %f3396, %f3395; max.ftz.f32 %f3398, %f3380, %f3392; max.ftz.f32 %f3399, %f3381, %f3393; max.ftz.f32 %f3400, %f3382, %f3394; max.ftz.f32 %f3401, %f3385, %f3397; max.ftz.f32 %f3402, %f3398, %f3399; max.ftz.f32 %f3403, %f3400, %f3401; max.ftz.f32 %f3404, %f3402, %f3403; mul.ftz.f32 %f749, %f3404, 0f35000000; sub.ftz.f32 %f3405, %f3300, 0f00000000; mul.ftz.f32 %f3406, %f3405, 0f00000000; mul.ftz.f32 %f3407, %f3406, %f3405; mul.ftz.f32 %f3408, %f737, %f3407; mul.ftz.f32 %f3409, %f738, %f3407; mul.ftz.f32 %f3410, %f739, %f3407; add.ftz.f32 %f3411, %f740, %f740; add.ftz.f32 %f3412, %f741, %f741; add.ftz.f32 %f3413, %f742, %f742; sub.ftz.f32 %f3414, %f3411, %f3408; sub.ftz.f32 %f3415, %f3412, %f3409; sub.ftz.f32 %f3416, %f3413, %f3410; add.ftz.f32 %f3417, %f728, %f728; mul.ftz.f32 %f3418, %f732, %f3407; sub.ftz.f32 %f3419, %f3417, %f3418; mul.ftz.f32 %f3420, %f3405, %f3405; mul.ftz.f32 %f3421, %f3405, 0f40400000; add.ftz.f32 %f3422, %f3421, 0fC0A00000; fma.rn.ftz.f32 %f3423, %f3420, %f3422, 0f40000000; fma.rn.ftz.f32 %f3424, %f743, %f3423, %f3414; fma.rn.ftz.f32 %f3425, %f744, %f3423, %f3415; fma.rn.ftz.f32 %f3426, %f745, %f3423, %f3416; fma.rn.ftz.f32 %f3427, %f724, %f3423, %f3419; mul.ftz.f32 %f3428, %f3406, 0f00000000; mul.ftz.f32 %f3429, %f746, %f3428; mul.ftz.f32 %f3430, %f747, %f3428; mul.ftz.f32 %f3431, %f748, %f3428; sub.ftz.f32 %f3432, %f3424, %f3429; sub.ftz.f32 %f3433, %f3425, %f3430; sub.ftz.f32 %f750, %f3426, %f3431; mul.ftz.f32 %f3434, %f720, %f3428; sub.ftz.f32 %f3435, %f3427, %f3434; mul.ftz.f32 %f6771, %f3432, 0f3F000000; mul.ftz.f32 %f6770, %f3433, 0f3F000000; mul.ftz.f32 %f6768, %f3435, 0f3F000000; add.ftz.f32 %f3436, %f3405, %f3405; mul.ftz.f32 %f3437, %f3436, 0f00000000; sub.ftz.f32 %f3438, %f3437, %f3420; mul.ftz.f32 %f3439, %f740, 0f00000000; mul.ftz.f32 %f3440, %f741, 0f00000000; mul.ftz.f32 %f3441, %f742, 0f00000000; fma.rn.ftz.f32 %f3442, %f737, %f3438, %f3439; fma.rn.ftz.f32 %f3443, %f738, %f3438, %f3440; fma.rn.ftz.f32 %f3444, %f739, %f3438, %f3441; mul.ftz.f32 %f3445, %f3421, %f3405; add.ftz.f32 %f3446, %f3436, %f3436; sub.ftz.f32 %f3447, %f3446, %f3445; fma.rn.ftz.f32 %f3448, %f743, %f3447, %f3442; fma.rn.ftz.f32 %f3449, %f744, %f3447, %f3443; fma.rn.ftz.f32 %f3450, %f745, %f3447, %f3444; mul.ftz.f32 %f3451, %f3405, 0fC0000000; fma.rn.ftz.f32 %f3452, %f3451, 0f00000000, 0f00000000; fma.rn.ftz.f32 %f3453, %f746, %f3452, %f3448; fma.rn.ftz.f32 %f3454, %f747, %f3452, %f3449; fma.rn.ftz.f32 %f3455, %f748, %f3452, %f3450; mul.ftz.f32 %f6766, %f3453, 0f3F000000; mul.ftz.f32 %f6767, %f3454, 0f3F000000; mul.ftz.f32 %f3456, %f3455, 0f3F000000; abs.ftz.f32 %f3457, %f6766; abs.ftz.f32 %f3458, %f6767; abs.ftz.f32 %f3459, %f3456; max.ftz.f32 %f3460, %f3457, %f3458; max.ftz.f32 %f3461, %f3460, %f3459; setp.geu.ftz.f32 %p318, %f3461, %f749; @%p318 bra $L__BB24_228; mul.ftz.f32 %f3462, %f344, %f740; fma.rn.ftz.f32 %f3463, %f737, %f343, %f3462; mul.ftz.f32 %f3464, %f344, %f741; fma.rn.ftz.f32 %f3465, %f738, %f343, %f3464; fma.rn.ftz.f32 %f3466, %f743, %f345, %f3463; fma.rn.ftz.f32 %f3467, %f345, %f744, %f3465; fma.rn.ftz.f32 %f3468, %f746, %f346, %f3466; fma.rn.ftz.f32 %f3469, %f747, %f346, %f3467; mul.ftz.f32 %f3470, %f3468, 0f3F000000; mul.ftz.f32 %f3471, %f3469, 0f3F000000; sub.ftz.f32 %f6766, %f3470, %f6771; sub.ftz.f32 %f6767, %f3471, %f6770; $L__BB24_228: mul.ftz.f32 %f3472, %f6767, %f6767; fma.rn.ftz.f32 %f760, %f6766, %f6766, %f3472; setp.lt.s32 %p320, %r829, 1; mov.pred %p792, -1; @%p320 bra $L__BB24_240; add.ftz.f32 %f3473, %f760, 0f00000000; mov.f32 %f3474, 0f00000000; sqrt.approx.ftz.f32 %f3475, %f3473; mul.ftz.f32 %f6769, %f750, 0f3F000000; div.approx.ftz.f32 %f3476, %f3474, %f3475; mul.ftz.f32 %f6772, %f6768, %f3476; neg.ftz.f32 %f3477, %f6766; div.approx.ftz.f32 %f3478, %f3477, %f3475; mul.ftz.f32 %f6773, %f6768, %f3478; div.approx.ftz.f32 %f3479, %f6767, %f3475; mul.ftz.f32 %f6774, %f6768, %f3479; mov.u32 %r1340, 0; $L__BB24_230: mov.f32 %f771, %f6774; mov.f32 %f770, %f6773; mov.f32 %f769, %f6772; mov.f32 %f768, %f6771; mov.f32 %f767, %f6770; mov.f32 %f766, %f6769; mov.f32 %f765, %f6768; cvt.rn.f32.s32 %f3480, %r1340; mul.ftz.f32 %f775, %f342, %f3480; add.ftz.f32 %f3481, %f342, %f775; sub.ftz.f32 %f3483, %f3300, %f3481; mul.ftz.f32 %f3484, %f3481, %f3483; mul.ftz.f32 %f3485, %f3483, %f3484; mul.ftz.f32 %f3486, %f3481, %f3481; mul.ftz.f32 %f3487, %f3481, 0f40400000; add.ftz.f32 %f3488, %f3487, 0fC0A00000; fma.rn.ftz.f32 %f3489, %f3486, %f3488, 0f40000000; mul.ftz.f32 %f3490, %f3483, %f3483; mul.ftz.f32 %f3491, %f3483, 0f40400000; add.ftz.f32 %f3492, %f3491, 0fC0A00000; fma.rn.ftz.f32 %f3493, %f3490, %f3492, 0f40000000; mul.ftz.f32 %f3494, %f3481, %f3484; mul.ftz.f32 %f3495, %f737, %f3485; mul.ftz.f32 %f3496, %f738, %f3485; mul.ftz.f32 %f3497, %f739, %f3485; mul.ftz.f32 %f3498, %f732, %f3485; mul.ftz.f32 %f3499, %f740, %f3489; mul.ftz.f32 %f3500, %f741, %f3489; mul.ftz.f32 %f3501, %f742, %f3489; mul.ftz.f32 %f3502, %f728, %f3489; sub.ftz.f32 %f3503, %f3499, %f3495; sub.ftz.f32 %f3504, %f3500, %f3496; sub.ftz.f32 %f3505, %f3501, %f3497; sub.ftz.f32 %f3506, %f3502, %f3498; fma.rn.ftz.f32 %f3507, %f743, %f3493, %f3503; fma.rn.ftz.f32 %f3508, %f744, %f3493, %f3504; fma.rn.ftz.f32 %f3509, %f745, %f3493, %f3505; fma.rn.ftz.f32 %f3510, %f724, %f3493, %f3506; mul.ftz.f32 %f3511, %f746, %f3494; mul.ftz.f32 %f3512, %f747, %f3494; mul.ftz.f32 %f3513, %f748, %f3494; mul.ftz.f32 %f3514, %f720, %f3494; sub.ftz.f32 %f3515, %f3507, %f3511; sub.ftz.f32 %f3516, %f3508, %f3512; sub.ftz.f32 %f3517, %f3509, %f3513; sub.ftz.f32 %f3518, %f3510, %f3514; mul.ftz.f32 %f6771, %f3515, 0f3F000000; mul.ftz.f32 %f6770, %f3516, 0f3F000000; mul.ftz.f32 %f6769, %f3517, 0f3F000000; mul.ftz.f32 %f6768, %f3518, 0f3F000000; sub.ftz.f32 %f780, %f6771, %f768; sub.ftz.f32 %f781, %f6770, %f767; mul.ftz.f32 %f3519, %f781, %f781; fma.rn.ftz.f32 %f782, %f780, %f780, %f3519; add.ftz.f32 %f3520, %f3483, %f3483; mul.ftz.f32 %f3521, %f3481, %f3520; sub.ftz.f32 %f3522, %f3521, %f3490; add.ftz.f32 %f3523, %f3481, %f3481; mul.ftz.f32 %f3524, %f3523, %f3488; fma.rn.ftz.f32 %f3525, %f3481, %f3487, %f3524; add.ftz.f32 %f3526, %f3487, 0f40000000; mul.ftz.f32 %f3527, %f3520, %f3526; mul.ftz.f32 %f3528, %f3483, %f3491; sub.ftz.f32 %f3529, %f3527, %f3528; mul.ftz.f32 %f3530, %f3483, 0fC0000000; fma.rn.ftz.f32 %f3531, %f3481, %f3530, %f3486; mul.ftz.f32 %f3532, %f740, %f3525; mul.ftz.f32 %f3533, %f741, %f3525; mul.ftz.f32 %f3534, %f742, %f3525; fma.rn.ftz.f32 %f3535, %f737, %f3522, %f3532; fma.rn.ftz.f32 %f3536, %f738, %f3522, %f3533; fma.rn.ftz.f32 %f3537, %f739, %f3522, %f3534; fma.rn.ftz.f32 %f3538, %f743, %f3529, %f3535; fma.rn.ftz.f32 %f3539, %f744, %f3529, %f3536; fma.rn.ftz.f32 %f3540, %f745, %f3529, %f3537; fma.rn.ftz.f32 %f3541, %f746, %f3531, %f3538; fma.rn.ftz.f32 %f3542, %f747, %f3531, %f3539; fma.rn.ftz.f32 %f3543, %f748, %f3531, %f3540; mul.ftz.f32 %f6778, %f3541, 0f3F000000; mul.ftz.f32 %f6779, %f3542, 0f3F000000; mul.ftz.f32 %f785, %f3543, 0f3F000000; abs.ftz.f32 %f3544, %f6778; abs.ftz.f32 %f3545, %f6779; abs.ftz.f32 %f3546, %f785; max.ftz.f32 %f3547, %f3544, %f3545; max.ftz.f32 %f3548, %f3547, %f3546; setp.lt.ftz.f32 %p321, %f3548, %f749; @%p321 bra $L__BB24_231; bra.uni $L__BB24_232; $L__BB24_231: mov.f32 %f6778, %f780; mov.f32 %f6779, %f781; $L__BB24_232: max.ftz.f32 %f3549, %f765, %f6768; mul.ftz.f32 %f3550, %f768, %f781; mul.ftz.f32 %f3551, %f767, %f780; sub.ftz.f32 %f3552, %f3551, %f3550; neg.ftz.f32 %f3553, %f6778; mul.ftz.f32 %f3554, %f6778, %f6778; fma.rn.ftz.f32 %f3555, %f6779, %f6779, %f3554; add.ftz.f32 %f3556, %f3555, 0f00000000; sqrt.approx.ftz.f32 %f3558, %f3556; div.approx.ftz.f32 %f3559, %f6779, %f3558; div.approx.ftz.f32 %f3560, %f3553, %f3558; div.approx.ftz.f32 %f3561, %f3474, %f3558; mul.ftz.f32 %f6774, %f6768, %f3559; mul.ftz.f32 %f6773, %f6768, %f3560; mul.ftz.f32 %f6772, %f6768, %f3561; mul.ftz.f32 %f3562, %f3549, %f3549; mul.ftz.f32 %f3563, %f782, %f3562; mul.ftz.f32 %f3564, %f3552, %f3552; setp.gtu.ftz.f32 %p322, %f3564, %f3563; @%p322 bra $L__BB24_239; add.ftz.f32 %f3565, %f771, %f768; sub.ftz.f32 %f3566, %f768, %f771; add.ftz.f32 %f3567, %f6771, %f6774; sub.ftz.f32 %f3568, %f3567, %f3566; sub.ftz.f32 %f3569, %f767, %f770; add.ftz.f32 %f3570, %f6770, %f6773; sub.ftz.f32 %f3571, %f3570, %f3569; sub.ftz.f32 %f3572, %f766, %f769; add.ftz.f32 %f3573, %f6769, %f6772; sub.ftz.f32 %f3574, %f3573, %f3572; mul.ftz.f32 %f3575, %f3569, %f3574; mul.ftz.f32 %f3576, %f3572, %f3571; sub.ftz.f32 %f3577, %f3575, %f3576; mul.ftz.f32 %f3578, %f3572, %f3568; mul.ftz.f32 %f3579, %f3566, %f3574; sub.ftz.f32 %f3580, %f3578, %f3579; mul.ftz.f32 %f3581, %f3566, %f3571; mul.ftz.f32 %f3582, %f3569, %f3568; sub.ftz.f32 %f3583, %f3581, %f3582; mul.ftz.f32 %f3584, %f3580, 0f00000000; fma.rn.ftz.f32 %f3585, %f3577, 0f00000000, %f3584; add.ftz.f32 %f798, %f3583, %f3585; setp.gtu.ftz.f32 %p324, %f798, 0f00000000; sub.ftz.f32 %f3586, %f6771, %f6774; selp.f32 %f799, %f3586, %f3565, %p324; sub.ftz.f32 %f3587, %f6770, %f6773; add.ftz.f32 %f3588, %f770, %f767; selp.f32 %f800, %f3587, %f3588, %p324; sub.ftz.f32 %f3589, %f6769, %f6772; add.ftz.f32 %f3590, %f769, %f766; selp.f32 %f801, %f3589, %f3590, %p324; selp.f32 %f3591, %f3566, %f3567, %p324; selp.f32 %f3592, %f3569, %f3570, %p324; selp.f32 %f3593, %f3572, %f3573, %p324; selp.f32 %f3594, %f3567, %f3566, %p324; selp.f32 %f3595, %f3570, %f3569, %p324; selp.f32 %f3596, %f3573, %f3572, %p324; sub.ftz.f32 %f802, %f3594, %f799; sub.ftz.f32 %f803, %f3595, %f800; sub.ftz.f32 %f804, %f3596, %f801; sub.ftz.f32 %f805, %f799, %f3591; sub.ftz.f32 %f806, %f800, %f3592; sub.ftz.f32 %f807, %f801, %f3593; mul.ftz.f32 %f3597, %f800, %f804; mul.ftz.f32 %f3598, %f801, %f803; sub.ftz.f32 %f3599, %f3597, %f3598; mul.ftz.f32 %f3600, %f801, %f802; mul.ftz.f32 %f3601, %f799, %f804; sub.ftz.f32 %f3602, %f3600, %f3601; mul.ftz.f32 %f3603, %f799, %f803; mul.ftz.f32 %f3604, %f800, %f802; sub.ftz.f32 %f3605, %f3603, %f3604; mul.ftz.f32 %f3606, %f3602, 0f00000000; fma.rn.ftz.f32 %f3607, %f3599, 0f00000000, %f3606; add.ftz.f32 %f808, %f3605, %f3607; mul.ftz.f32 %f3608, %f3592, %f807; mul.ftz.f32 %f3609, %f3593, %f806; sub.ftz.f32 %f3610, %f3608, %f3609; mul.ftz.f32 %f3611, %f3593, %f805; mul.ftz.f32 %f3612, %f3591, %f807; sub.ftz.f32 %f3613, %f3611, %f3612; mul.ftz.f32 %f3614, %f3591, %f806; mul.ftz.f32 %f3615, %f3592, %f805; sub.ftz.f32 %f3616, %f3614, %f3615; mul.ftz.f32 %f3617, %f3613, 0f00000000; fma.rn.ftz.f32 %f3618, %f3610, 0f00000000, %f3617; add.ftz.f32 %f809, %f3616, %f3618; max.ftz.f32 %f3619, %f808, %f809; setp.gtu.ftz.f32 %p325, %f3619, 0f00000000; mov.pred %p791, -1; @%p325 bra $L__BB24_236; mul.ftz.f32 %f3620, %f806, %f804; mul.ftz.f32 %f3621, %f807, %f803; sub.ftz.f32 %f3622, %f3620, %f3621; mul.ftz.f32 %f3623, %f805, %f804; mul.ftz.f32 %f3624, %f807, %f802; sub.ftz.f32 %f3625, %f3624, %f3623; mul.ftz.f32 %f3626, %f806, %f802; mul.ftz.f32 %f3627, %f805, %f803; sub.ftz.f32 %f3628, %f3627, %f3626; mul.ftz.f32 %f3629, %f3625, 0f00000000; fma.rn.ftz.f32 %f3630, %f3622, 0f00000000, %f3629; add.ftz.f32 %f3631, %f3628, %f3630; rcp.approx.ftz.f32 %f810, %f3631; mul.ftz.f32 %f3632, %f800, %f3625; fma.rn.ftz.f32 %f3633, %f799, %f3622, %f3632; fma.rn.ftz.f32 %f3634, %f801, %f3628, %f3633; mul.ftz.f32 %f811, %f3634, %f810; setp.ltu.ftz.f32 %p327, %f811, %f340; setp.gtu.ftz.f32 %p328, %f811, %f341; or.pred %p329, %p328, %p327; setp.eq.ftz.f32 %p330, %f3631, 0f00000000; or.pred %p331, %p330, %p329; @%p331 bra $L__BB24_236; mul.ftz.f32 %f3635, %f808, %f810; sub.ftz.f32 %f3637, %f3300, %f3635; setp.le.ftz.f32 %p333, %f798, 0f00000000; selp.f32 %f6856, %f3635, %f3637, %p333; mul.ftz.f32 %f3638, %f809, %f810; sub.ftz.f32 %f3639, %f3300, %f3638; selp.f32 %f6855, %f3638, %f3639, %p333; mov.pred %p791, 0; mov.f32 %f6854, %f811; $L__BB24_236: @%p791 bra $L__BB24_239; sub.ftz.f32 %f3640, %f6768, %f765; fma.rn.ftz.f32 %f3641, %f3640, %f6856, %f765; add.ftz.f32 %f3642, %f3641, %f3641; mul.ftz.f32 %f3643, %f736, %f3642; setp.leu.ftz.f32 %p334, %f6854, %f3643; @%p334 bra $L__BB24_239; bra.uni $L__BB24_238; $L__BB24_239: add.s32 %r1340, %r1340, 1; setp.lt.s32 %p337, %r1340, %r829; @%p337 bra $L__BB24_230; bra.uni $L__BB24_240; $L__BB24_187: setp.gt.s32 %p267, %r215, %r214; selp.b32 %r874, -1, 0, %p267; add.s32 %r875, %r215, %r874; mad.lo.s32 %r876, %r875, %r213, %r1329; add.s32 %r877, %r870, %r876; mul.wide.u32 %rd593, %r877, 12; add.s64 %rd1206, %rd129, %rd593; add.s32 %r878, %r871, %r876; mul.wide.u32 %rd594, %r878, 12; add.s64 %rd1203, %rd129, %rd594; add.s32 %r879, %r872, %r876; mul.wide.u32 %rd595, %r879, 12; add.s64 %rd1200, %rd129, %rd595; add.s64 %rd1205, %rd1206, 4; add.s64 %rd1204, %rd1206, 8; add.s64 %rd1202, %rd1203, 4; add.s64 %rd1201, %rd1203, 8; add.s64 %rd1199, %rd1200, 4; add.s64 %rd1198, %rd1200, 8; $L__BB24_189: ld.global.f32 %f530, [%rd1206]; ld.global.f32 %f531, [%rd1205]; ld.global.f32 %f532, [%rd1204]; ld.global.f32 %f533, [%rd1203]; ld.global.f32 %f534, [%rd1202]; ld.global.f32 %f535, [%rd1201]; ld.global.f32 %f536, [%rd1200]; ld.global.f32 %f537, [%rd1199]; ld.global.f32 %f538, [%rd1198]; add.s32 %r230, %r215, 1; setp.eq.s32 %p268, %r230, %r214; @%p268 bra $L__BB24_191; bra.uni $L__BB24_190; $L__BB24_191: mul.wide.u32 %rd602, %r873, 12; add.s64 %rd1215, %rd119, %rd602; add.s32 %r888, %r873, 1; mul.wide.u32 %rd603, %r888, 12; add.s64 %rd1212, %rd119, %rd603; add.s32 %r889, %r873, 2; mul.wide.u32 %rd604, %r889, 12; add.s64 %rd1209, %rd119, %rd604; add.s64 %rd1214, %rd1215, 4; add.s64 %rd1213, %rd1215, 8; add.s64 %rd1211, %rd1212, 4; add.s64 %rd1210, %rd1212, 8; add.s64 %rd1208, %rd1209, 4; add.s64 %rd1207, %rd1209, 8; bra.uni $L__BB24_192; $L__BB24_190: setp.ge.s32 %p269, %r215, %r214; selp.b32 %r882, -1, 0, %p269; add.s32 %r883, %r230, %r882; mad.lo.s32 %r884, %r883, %r213, %r1329; add.s32 %r885, %r870, %r884; mul.wide.u32 %rd599, %r885, 12; add.s64 %rd1215, %rd129, %rd599; add.s32 %r886, %r871, %r884; mul.wide.u32 %rd600, %r886, 12; add.s64 %rd1212, %rd129, %rd600; add.s32 %r887, %r872, %r884; mul.wide.u32 %rd601, %r887, 12; add.s64 %rd1209, %rd129, %rd601; add.s64 %rd1214, %rd1215, 4; add.s64 %rd1213, %rd1215, 8; add.s64 %rd1211, %rd1212, 4; add.s64 %rd1210, %rd1212, 8; add.s64 %rd1208, %rd1209, 4; add.s64 %rd1207, %rd1209, 8; $L__BB24_192: cvt.rn.f32.s32 %f3036, %r215; sub.ftz.f32 %f3037, %f529, %f3036; mov.f32 %f3038, 0f3F800000; sub.ftz.f32 %f3039, %f3038, %f3037; ld.global.f32 %f3040, [%rd1215]; mul.ftz.f32 %f3041, %f3037, %f3040; ld.global.f32 %f3042, [%rd1214]; mul.ftz.f32 %f3043, %f3037, %f3042; ld.global.f32 %f3044, [%rd1213]; mul.ftz.f32 %f3045, %f3037, %f3044; fma.rn.ftz.f32 %f3046, %f3039, %f530, %f3041; fma.rn.ftz.f32 %f3047, %f3039, %f531, %f3043; fma.rn.ftz.f32 %f3048, %f3039, %f532, %f3045; ld.global.f32 %f3049, [%rd1212]; mul.ftz.f32 %f3050, %f3037, %f3049; ld.global.f32 %f3051, [%rd1211]; mul.ftz.f32 %f3052, %f3037, %f3051; ld.global.f32 %f3053, [%rd1210]; mul.ftz.f32 %f3054, %f3037, %f3053; fma.rn.ftz.f32 %f3055, %f3039, %f533, %f3050; fma.rn.ftz.f32 %f3056, %f3039, %f534, %f3052; fma.rn.ftz.f32 %f3057, %f3039, %f535, %f3054; ld.global.f32 %f3058, [%rd1209]; mul.ftz.f32 %f3059, %f3037, %f3058; ld.global.f32 %f3060, [%rd1208]; mul.ftz.f32 %f3061, %f3037, %f3060; ld.global.f32 %f3062, [%rd1207]; mul.ftz.f32 %f3063, %f3037, %f3062; fma.rn.ftz.f32 %f3064, %f3039, %f536, %f3059; fma.rn.ftz.f32 %f3065, %f3039, %f537, %f3061; fma.rn.ftz.f32 %f3066, %f3039, %f538, %f3063; sub.ftz.f32 %f539, %f3046, %f6905; sub.ftz.f32 %f540, %f3047, %f6904; sub.ftz.f32 %f541, %f3048, %f6903; sub.ftz.f32 %f3067, %f3055, %f6905; sub.ftz.f32 %f3068, %f3056, %f6904; sub.ftz.f32 %f3069, %f3057, %f6903; sub.ftz.f32 %f3070, %f3064, %f6905; sub.ftz.f32 %f3071, %f3065, %f6904; sub.ftz.f32 %f3072, %f3066, %f6903; sub.ftz.f32 %f542, %f3070, %f539; sub.ftz.f32 %f543, %f3071, %f540; sub.ftz.f32 %f544, %f3072, %f541; sub.ftz.f32 %f545, %f539, %f3067; sub.ftz.f32 %f546, %f540, %f3068; sub.ftz.f32 %f547, %f541, %f3069; sub.ftz.f32 %f3073, %f3067, %f3070; sub.ftz.f32 %f3074, %f3068, %f3071; sub.ftz.f32 %f3075, %f3069, %f3072; add.ftz.f32 %f3076, %f539, %f3070; add.ftz.f32 %f3077, %f540, %f3071; add.ftz.f32 %f3078, %f541, %f3072; mul.ftz.f32 %f3079, %f543, %f3078; mul.ftz.f32 %f3080, %f3077, %f544; sub.ftz.f32 %f3081, %f3079, %f3080; mul.ftz.f32 %f3082, %f3076, %f544; mul.ftz.f32 %f3083, %f542, %f3078; sub.ftz.f32 %f3084, %f3082, %f3083; mul.ftz.f32 %f3085, %f542, %f3077; mul.ftz.f32 %f3086, %f3076, %f543; sub.ftz.f32 %f3087, %f3085, %f3086; mul.ftz.f32 %f3088, %f6907, %f3084; fma.rn.ftz.f32 %f3089, %f6908, %f3081, %f3088; fma.rn.ftz.f32 %f548, %f6906, %f3087, %f3089; add.ftz.f32 %f3090, %f539, %f3067; add.ftz.f32 %f3091, %f540, %f3068; add.ftz.f32 %f3092, %f541, %f3069; mul.ftz.f32 %f3093, %f546, %f3092; mul.ftz.f32 %f3094, %f3091, %f547; sub.ftz.f32 %f3095, %f3093, %f3094; mul.ftz.f32 %f3096, %f3090, %f547; mul.ftz.f32 %f3097, %f545, %f3092; sub.ftz.f32 %f3098, %f3096, %f3097; mul.ftz.f32 %f3099, %f545, %f3091; mul.ftz.f32 %f3100, %f3090, %f546; sub.ftz.f32 %f3101, %f3099, %f3100; mul.ftz.f32 %f3102, %f6907, %f3098; fma.rn.ftz.f32 %f3103, %f6908, %f3095, %f3102; fma.rn.ftz.f32 %f549, %f6906, %f3101, %f3103; add.ftz.f32 %f3104, %f3067, %f3070; add.ftz.f32 %f3105, %f3068, %f3071; add.ftz.f32 %f3106, %f3069, %f3072; mul.ftz.f32 %f3107, %f3074, %f3106; mul.ftz.f32 %f3108, %f3105, %f3075; sub.ftz.f32 %f3109, %f3107, %f3108; mul.ftz.f32 %f3110, %f3104, %f3075; mul.ftz.f32 %f3111, %f3073, %f3106; sub.ftz.f32 %f3112, %f3110, %f3111; mul.ftz.f32 %f3113, %f3073, %f3105; mul.ftz.f32 %f3114, %f3104, %f3074; sub.ftz.f32 %f3115, %f3113, %f3114; mul.ftz.f32 %f3116, %f6907, %f3112; fma.rn.ftz.f32 %f3117, %f6908, %f3109, %f3116; fma.rn.ftz.f32 %f3118, %f6906, %f3115, %f3117; add.ftz.f32 %f3119, %f549, %f548; add.ftz.f32 %f550, %f3118, %f3119; abs.ftz.f32 %f551, %f550; mul.ftz.f32 %f3120, %f551, 0f34000000; min.ftz.f32 %f3121, %f549, %f3118; min.ftz.f32 %f3122, %f548, %f3121; max.ftz.f32 %f3123, %f549, %f3118; max.ftz.f32 %f3124, %f548, %f3123; mul.ftz.f32 %f3125, %f551, 0fB4000000; setp.ltu.ftz.f32 %p271, %f3122, %f3125; setp.gtu.ftz.f32 %p272, %f3124, %f3120; and.pred %p273, %p271, %p272; mov.pred %p789, -1; @%p273 bra $L__BB24_198; mul.ftz.f32 %f3126, %f546, %f544; mul.ftz.f32 %f3127, %f547, %f543; sub.ftz.f32 %f3128, %f3126, %f3127; mul.ftz.f32 %f3129, %f545, %f544; mul.ftz.f32 %f3130, %f547, %f542; sub.ftz.f32 %f3131, %f3130, %f3129; mul.ftz.f32 %f3132, %f546, %f542; mul.ftz.f32 %f3133, %f545, %f543; sub.ftz.f32 %f3134, %f3133, %f3132; add.ftz.f32 %f552, %f3128, %f3128; add.ftz.f32 %f553, %f3131, %f3131; add.ftz.f32 %f554, %f3134, %f3134; mul.ftz.f32 %f3135, %f6907, %f553; fma.rn.ftz.f32 %f3136, %f6908, %f552, %f3135; fma.rn.ftz.f32 %f555, %f6906, %f554, %f3136; setp.eq.ftz.f32 %p275, %f555, 0f00000000; @%p275 bra $L__BB24_198; mul.ftz.f32 %f3137, %f540, %f553; fma.rn.ftz.f32 %f3138, %f539, %f552, %f3137; fma.rn.ftz.f32 %f3139, %f541, %f554, %f3138; div.approx.ftz.f32 %f556, %f3139, %f555; setp.ltu.ftz.f32 %p277, %f556, %f340; setp.gtu.ftz.f32 %p278, %f556, %f341; or.pred %p279, %p278, %p277; @%p279 bra $L__BB24_198; setp.lt.ftz.f32 %p280, %f551, 0f219392EF; mov.f32 %f6709, 0f00000000; @%p280 bra $L__BB24_197; rcp.approx.ftz.f32 %f6709, %f550; $L__BB24_197: mul.ftz.f32 %f3141, %f548, %f6709; min.ftz.f32 %f6860, %f3141, %f3038; mul.ftz.f32 %f3143, %f549, %f6709; min.ftz.f32 %f6861, %f3143, %f3038; mov.pred %p789, 0; mov.f32 %f6859, %f556; $L__BB24_198: mov.pred %p795, %p249; @%p789 bra $L__BB24_293; cvt.u32.u64 %r890, %rd143; add.s64 %rd606, %rd126, %rd583; ld.global.u32 %r891, [%rd606]; and.b32 %r892, %r891, %r668; setp.eq.s32 %p283, %r892, 0; selp.b32 %r1350, %r1350, 65, %p283; selp.b32 %r1351, %r1351, %r212, %p283; selp.b32 %r1352, %r1352, %r890, %p283; selp.f32 %f6865, %f6865, %f6861, %p283; selp.f32 %f6866, %f6866, %f6860, %p283; selp.f32 %f6867, %f6867, %f6859, %p283; setp.ne.s32 %p795, %r892, 0; bra.uni $L__BB24_293; $L__BB24_204: cvt.u32.u64 %r898, %rd143; setp.gt.s32 %p295, %r237, %r236; selp.b32 %r899, -1, 0, %p295; add.s32 %r900, %r237, %r899; mad.lo.s32 %r901, %r900, %r235, %r898; add.s32 %r902, %r901, %r1334; mul.wide.s32 %rd617, %r902, 16; add.s64 %rd1217, %rd131, %rd617; $L__BB24_206: ld.global.v4.f32 {%f3151, %f3152, %f3153, %f3154}, [%rd1217]; add.s32 %r248, %r237, 1; setp.eq.s32 %p296, %r248, %r236; @%p296 bra $L__BB24_208; bra.uni $L__BB24_207; $L__BB24_208: shl.b64 %rd620, %rd143, 4; add.s64 %rd1218, %rd128, %rd620; bra.uni $L__BB24_209; $L__BB24_207: cvt.u32.u64 %r903, %rd143; setp.ge.s32 %p297, %r237, %r236; selp.b32 %r904, -1, 0, %p297; add.s32 %r905, %r248, %r904; mad.lo.s32 %r906, %r905, %r235, %r903; add.s32 %r907, %r906, %r1334; mul.wide.s32 %rd619, %r907, 16; add.s64 %rd1218, %rd131, %rd619; $L__BB24_209: cvt.rn.f32.s32 %f3155, %r237; sub.ftz.f32 %f3156, %f567, %f3155; ld.global.v4.f32 {%f3157, %f3158, %f3159, %f3160}, [%rd1218]; mov.f32 %f3165, 0f3F800000; sub.ftz.f32 %f3166, %f3165, %f3156; mul.ftz.f32 %f3167, %f3156, %f3157; mul.ftz.f32 %f3168, %f3156, %f3158; mul.ftz.f32 %f3169, %f3156, %f3159; mul.ftz.f32 %f3170, %f3156, %f3160; fma.rn.ftz.f32 %f6716, %f3166, %f3154, %f3170; fma.rn.ftz.f32 %f6715, %f3166, %f3153, %f3169; fma.rn.ftz.f32 %f6714, %f3166, %f3152, %f3168; fma.rn.ftz.f32 %f6713, %f3166, %f3151, %f3167; $L__BB24_211: sub.ftz.f32 %f3175, %f6713, %f6905; sub.ftz.f32 %f3176, %f6714, %f6904; sub.ftz.f32 %f3177, %f6715, %f6903; mul.ftz.f32 %f3178, %f6907, %f3176; fma.rn.ftz.f32 %f3179, %f6908, %f3175, %f3178; fma.rn.ftz.f32 %f3180, %f6906, %f3177, %f3179; mul.ftz.f32 %f584, %f512, %f3180; mul.ftz.f32 %f3181, %f6908, %f584; mul.ftz.f32 %f3182, %f6907, %f584; mul.ftz.f32 %f3183, %f6906, %f584; sub.ftz.f32 %f3184, %f3175, %f3181; sub.ftz.f32 %f3185, %f3176, %f3182; sub.ftz.f32 %f3186, %f3177, %f3183; mul.ftz.f32 %f3187, %f3185, %f3185; fma.rn.ftz.f32 %f3188, %f3184, %f3184, %f3187; fma.rn.ftz.f32 %f585, %f3186, %f3186, %f3188; mul.ftz.f32 %f586, %f6716, %f6716; setp.gtu.ftz.f32 %p299, %f585, %f586; mov.pred %p790, -1; @%p299 bra $L__BB24_213; sub.ftz.f32 %f3189, %f586, %f585; mul.ftz.f32 %f3190, %f512, %f3189; sqrt.approx.ftz.f32 %f3191, %f3190; sub.ftz.f32 %f3192, %f584, %f3191; setp.ltu.ftz.f32 %p300, %f3192, %f340; setp.gtu.ftz.f32 %p301, %f3192, %f341; or.pred %p790, %p301, %p300; selp.f32 %f6867, %f6867, %f3192, %p790; $L__BB24_213: selp.b32 %r1350, %r1350, %r234, %p790; selp.b32 %r1351, %r1351, %r212, %p790; cvt.u32.u64 %r908, %rd143; selp.b32 %r1352, %r1352, %r908, %p790; selp.f32 %f6865, %f6865, 0f00000000, %p790; selp.f32 %f6866, %f6866, 0f00000000, %p790; not.pred %p795, %p790; bra.uni $L__BB24_293; $L__BB24_238: fma.rn.ftz.f32 %f6855, %f6855, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f6866, %f342, %f6856, %f775; mov.pred %p792, 0; mov.f32 %f6865, %f6855; mov.f32 %f6867, %f6854; $L__BB24_240: selp.b32 %r1350, %r1350, %r252, %p792; selp.b32 %r1351, %r1351, %r212, %p792; cvt.u32.u64 %r935, %rd143; selp.b32 %r1352, %r1352, %r935, %p792; not.pred %p795, %p792; $L__BB24_293: not.pred %p409, %p795; @%p409 bra $L__BB24_325; and.b32 %r949, %r1350, 1; setp.eq.b32 %p410, %r949, 1; mov.pred %p411, 0; xor.pred %p412, %p410, %p411; not.pred %p413, %p412; mul.wide.s32 %rd664, %r1352, 16; add.s64 %rd216, %rd122, %rd664; @%p413 bra $L__BB24_296; bra.uni $L__BB24_295; $L__BB24_296: and.b32 %r950, %r1350, 8; setp.eq.s32 %p414, %r950, 0; @%p414 bra $L__BB24_298; bra.uni $L__BB24_297; $L__BB24_298: and.b32 %r952, %r1350, 6; setp.eq.s32 %p415, %r952, 0; mov.u32 %r1353, 0; @%p415 bra $L__BB24_300; ld.global.u32 %r1353, [%rd216]; bra.uni $L__BB24_300; $L__BB24_295: mul.wide.s32 %rd665, %r1352, 4; add.s64 %rd666, %rd125, %rd665; ld.global.u32 %r1353, [%rd666]; bra.uni $L__BB24_300; $L__BB24_297: mul.wide.s32 %rd667, %r1352, 4; add.s64 %rd668, %rd130, %rd667; ld.global.u32 %r1353, [%rd668]; $L__BB24_300: shl.b32 %r953, %r1353, 5; cvt.u64.u32 %rd669, %r953; and.b64 %rd670, %rd669, 134217696; add.s64 %rd671, %rd124, %rd670; ld.global.u8 %rs138, [%rd671+18]; and.b16 %rs139, %rs138, 2; setp.eq.s16 %p416, %rs139, 0; setp.ge.u32 %p417, %r1376, %r1; or.pred %p418, %p417, %p416; mov.u16 %rs242, %rs114; @%p418 bra $L__BB24_594; and.b32 %r954, %r1350, 6; setp.eq.s32 %p420, %r954, 0; mov.pred %p796, -1; @%p420 bra $L__BB24_311; mul.wide.s32 %rd672, %r1351, 208; add.s64 %rd673, %rd116, %rd672; ld.global.u32 %r1355, [%rd673+164]; mul.wide.u32 %rd674, %r1355, 16; add.s64 %rd675, %rd120, %rd674; ld.global.u16 %rs215, [%rd675+12]; ld.global.u32 %r1354, [%rd675+8]; ld.global.u64 %rd1220, [%rd675]; setp.eq.s64 %p421, %rd1220, 31; @%p421 bra $L__BB24_307; $L__BB24_303: setp.eq.s64 %p422, %rd1220, 0; @%p422 bra $L__BB24_305; add.s32 %r1355, %r1355, 2; bra.uni $L__BB24_306; $L__BB24_305: setp.eq.s16 %p423, %rs215, 0; mov.u32 %r1358, -1; mov.u32 %r1355, %r1354; @%p423 bra $L__BB24_308; $L__BB24_306: mul.wide.u32 %rd676, %r1355, 16; add.s64 %rd677, %rd120, %rd676; ld.global.u16 %rs215, [%rd677+12]; ld.global.u32 %r1354, [%rd677+8]; ld.global.u64 %rd1220, [%rd677]; setp.ne.s64 %p424, %rd1220, 31; @%p424 bra $L__BB24_303; $L__BB24_307: setp.eq.s16 %p425, %rs215, 0; selp.b32 %r1358, -1, %r1354, %p425; $L__BB24_308: setp.eq.s32 %p426, %r1358, -1; mov.f32 %f6868, 0f00000000; @%p426 bra $L__BB24_310; shr.s32 %r956, %r1350, 7; ld.global.u32 %r957, [%rd216+4]; add.s32 %r958, %r957, %r956; add.s32 %r959, %r958, %r1358; mul.wide.s32 %rd678, %r959, 4; add.s64 %rd679, %rd132, %rd678; mov.f32 %f4501, 0f3F800000; sub.ftz.f32 %f4502, %f4501, %f6866; ld.global.f32 %f4503, [%rd679]; ld.global.f32 %f4504, [%rd679+4]; mul.ftz.f32 %f4505, %f6866, %f4504; fma.rn.ftz.f32 %f6868, %f4502, %f4503, %f4505; $L__BB24_310: mul.ftz.f32 %f7321, %f7321, %f6868; setp.lt.ftz.f32 %p428, %f7321, 0f3A83126F; mov.pred %p796, %p411; mov.u16 %rs242, %rs114; @%p428 bra $L__BB24_594; $L__BB24_311: add.s32 %r1376, %r1376, 1; not.pred %p429, %p796; @%p429 bra $L__BB24_325; setp.ge.u32 %p430, %r1536, %r3; setp.geu.ftz.f32 %p431, %f6867, %f6916; and.pred %p432, %p430, %p431; @%p432 bra $L__BB24_324; min.u32 %r1366, %r1536, %r3; add.s32 %r960, %r1366, 1; setp.lt.u32 %p433, %r960, %r3; @%p433 bra $L__BB24_323; ld.global.f32 %f6873, [%rd133]; setp.lt.u32 %p434, %r1366, 2; mov.u32 %r1365, 0; @%p434 bra $L__BB24_322; min.u32 %r965, %r1, %r1536; not.b32 %r966, %r965; max.u32 %r967, %r966, -5; mov.u32 %r968, -2; sub.s32 %r315, %r968, %r967; mov.u32 %r969, -3; sub.s32 %r970, %r969, %r967; and.b32 %r316, %r315, 3; setp.lt.u32 %p435, %r970, 3; mov.u32 %r1365, 0; mov.u32 %r1363, 1; @%p435 bra $L__BB24_318; sub.s32 %r1361, %r315, %r316; mov.u32 %r1365, 0; mov.u32 %r1363, 1; $L__BB24_317: mul.wide.u32 %rd680, %r1363, 48; mov.u64 %rd681, kernel_params; add.s64 %rd682, %rd681, 3552; add.s64 %rd683, %rd682, %rd680; ld.const.u64 %rd684, [%rd683]; cvta.to.global.u64 %rd685, %rd684; add.s64 %rd687, %rd685, %rd558; ld.global.f32 %f4507, [%rd687]; setp.gt.ftz.f32 %p436, %f4507, %f6873; selp.f32 %f4508, %f4507, %f6873, %p436; selp.b32 %r973, %r1363, %r1365, %p436; ld.const.u64 %rd688, [%rd683+48]; cvta.to.global.u64 %rd689, %rd688; add.s64 %rd690, %rd689, %rd558; ld.global.f32 %f4509, [%rd690]; setp.gt.ftz.f32 %p437, %f4509, %f4508; selp.f32 %f4510, %f4509, %f4508, %p437; add.s32 %r974, %r1363, 1; selp.b32 %r975, %r974, %r973, %p437; ld.const.u64 %rd691, [%rd683+96]; cvta.to.global.u64 %rd692, %rd691; add.s64 %rd693, %rd692, %rd558; ld.global.f32 %f4511, [%rd693]; setp.gt.ftz.f32 %p438, %f4511, %f4510; selp.f32 %f4512, %f4511, %f4510, %p438; add.s32 %r976, %r1363, 2; selp.b32 %r977, %r976, %r975, %p438; add.s32 %r978, %r1363, 3; mul.wide.u32 %rd694, %r978, 48; add.s64 %rd695, %rd682, %rd694; ld.const.u64 %rd696, [%rd695]; cvta.to.global.u64 %rd697, %rd696; add.s64 %rd698, %rd697, %rd558; ld.global.f32 %f4513, [%rd698]; setp.gt.ftz.f32 %p439, %f4513, %f4512; selp.f32 %f6873, %f4513, %f4512, %p439; selp.b32 %r1365, %r978, %r977, %p439; add.s32 %r1363, %r1363, 4; add.s32 %r1361, %r1361, -4; setp.ne.s32 %p440, %r1361, 0; @%p440 bra $L__BB24_317; $L__BB24_318: setp.eq.s32 %p441, %r316, 0; @%p441 bra $L__BB24_322; mul.wide.u32 %rd699, %r1363, 48; mov.u64 %rd700, kernel_params; add.s64 %rd701, %rd700, %rd699; ld.const.u64 %rd702, [%rd701+3552]; cvta.to.global.u64 %rd703, %rd702; add.s64 %rd705, %rd703, %rd558; ld.global.f32 %f4514, [%rd705]; setp.gt.ftz.f32 %p442, %f4514, %f6873; selp.f32 %f6873, %f4514, %f6873, %p442; selp.b32 %r1365, %r1363, %r1365, %p442; setp.eq.s32 %p443, %r316, 1; @%p443 bra $L__BB24_322; add.s32 %r979, %r1363, 1; mul.wide.u32 %rd706, %r979, 48; add.s64 %rd708, %rd700, %rd706; ld.const.u64 %rd709, [%rd708+3552]; cvta.to.global.u64 %rd710, %rd709; add.s64 %rd712, %rd710, %rd558; ld.global.f32 %f4515, [%rd712]; setp.gt.ftz.f32 %p444, %f4515, %f6873; selp.f32 %f6873, %f4515, %f6873, %p444; selp.b32 %r1365, %r979, %r1365, %p444; setp.eq.s32 %p445, %r316, 2; @%p445 bra $L__BB24_322; add.s32 %r980, %r1363, 2; mul.wide.u32 %rd713, %r980, 48; add.s64 %rd715, %rd700, %rd713; ld.const.u64 %rd716, [%rd715+3552]; cvta.to.global.u64 %rd717, %rd716; add.s64 %rd719, %rd717, %rd558; ld.global.f32 %f4516, [%rd719]; setp.gt.ftz.f32 %p446, %f4516, %f6873; selp.f32 %f6873, %f4516, %f6873, %p446; selp.b32 %r1365, %r980, %r1365, %p446; $L__BB24_322: setp.lt.u32 %p447, %r1536, %r3; selp.b32 %r1366, %r1366, %r1365, %p447; max.ftz.f32 %f6916, %f6867, %f6873; $L__BB24_323: mul.wide.s32 %rd720, %r1366, 48; mov.u64 %rd721, kernel_params; add.s64 %rd722, %rd721, %rd720; ld.const.v2.u64 {%rd723, %rd724}, [%rd722+3552]; cvta.to.global.u64 %rd727, %rd723; add.s64 %rd729, %rd727, %rd558; st.global.f32 [%rd729], %f6867; cvta.to.global.u64 %rd730, %rd724; add.s64 %rd731, %rd730, %rd558; st.global.f32 [%rd731], %f6866; ld.const.v2.u64 {%rd732, %rd733}, [%rd722+3568]; cvta.to.global.u64 %rd736, %rd732; add.s64 %rd737, %rd736, %rd558; st.global.f32 [%rd737], %f6865; ld.const.v2.u64 {%rd738, %rd739}, [%rd722+3584]; cvta.to.global.u64 %rd742, %rd738; add.s64 %rd743, %rd742, %rd558; st.global.u32 [%rd743], %r1351; cvta.to.global.u64 %rd744, %rd733; add.s64 %rd745, %rd744, %rd558; st.global.u32 [%rd745], %r1352; cvta.to.global.u64 %rd746, %rd739; add.s64 %rd747, %rd746, %rd558; st.global.u32 [%rd747], %r1350; $L__BB24_324: add.s32 %r1536, %r1536, 1; $L__BB24_325: cvt.u32.u64 %r981, %rd142; add.s32 %r1318, %r981, 1; setp.lt.s32 %p448, %r1318, %r841; @%p448 bra $L__BB24_150; bra.uni $L__BB24_326; $L__BB24_139: mul.wide.s32 %rd568, %r1318, 4; xor.b64 %rd569, %rd568, -4; add.s64 %rd570, %rd114, %rd569; ld.global.u32 %r1377, [%rd570]; cvt.s64.s32 %rd139, %r1377; mul.wide.s32 %rd571, %r1377, 4; add.s64 %rd572, %rd115, %rd571; ld.global.u8 %rs111, [%rd572]; and.b16 %rs112, %rs111, 2; setp.eq.s16 %p231, %rs112, 0; mul.wide.s32 %rd573, %r1377, 208; add.s64 %rd574, %rd116, %rd573; add.s64 %rd140, %rd574, 48; @%p231 bra $L__BB24_146; bra.uni $L__BB24_140; $L__BB24_146: ld.global.v4.f32 {%f6681, %f6682, %f6683, %f6684}, [%rd140]; ld.global.v4.f32 {%f6685, %f6686, %f6687, %f6688}, [%rd140+16]; ld.global.v4.f32 {%f6689, %f6690, %f6691, %f6692}, [%rd140+32]; bra.uni $L__BB24_147; $L__BB24_140: ld.global.u32 %rd575, [%rd140+120]; ld.global.u32 %r844, [%rd140+104]; shl.b32 %r845, %r844, 1; cvt.rn.f32.s32 %f2762, %r845; ld.local.f32 %f2763, [%rd4+32]; mul.ftz.f32 %f2764, %f2763, %f2762; cvt.rzi.ftz.s32.f32 %r846, %f2764; add.s32 %r847, %r845, -1; min.s32 %r848, %r846, %r847; cvt.rn.f32.s32 %f2765, %r848; sub.ftz.f32 %f410, %f2764, %f2765; cvt.s64.s32 %rd576, %r848; add.s64 %rd577, %rd576, %rd575; shl.b64 %rd578, %rd577, 6; add.s64 %rd141, %rd135, %rd578; ld.global.v4.f32 {%f2766, %f2767, %f2768, %f2769}, [%rd141]; ld.global.v4.f32 {%f2770, %f2771, %f2772, %f2773}, [%rd141+64]; mul.ftz.f32 %f2774, %f2767, %f2771; fma.rn.ftz.f32 %f2775, %f2766, %f2770, %f2774; mul.ftz.f32 %f2776, %f2769, %f2773; fma.rn.ftz.f32 %f2777, %f2768, %f2772, %f2776; add.ftz.f32 %f419, %f2775, %f2777; setp.gt.ftz.f32 %p232, %f419, 0f3F7FDF3B; mov.f32 %f2778, 0f3F800000; sub.ftz.f32 %f420, %f2778, %f410; @%p232 bra $L__BB24_142; bra.uni $L__BB24_141; $L__BB24_142: mul.ftz.f32 %f2832, %f2766, %f420; fma.rn.ftz.f32 %f2833, %f2770, %f410, %f2832; mul.ftz.f32 %f2834, %f2767, %f420; fma.rn.ftz.f32 %f2835, %f2771, %f410, %f2834; mul.ftz.f32 %f2836, %f2768, %f420; fma.rn.ftz.f32 %f2837, %f2772, %f410, %f2836; mul.ftz.f32 %f2838, %f2769, %f420; fma.rn.ftz.f32 %f2839, %f2773, %f410, %f2838; mul.ftz.f32 %f2840, %f2835, %f2835; fma.rn.ftz.f32 %f2841, %f2833, %f2833, %f2840; mul.ftz.f32 %f2842, %f2839, %f2839; fma.rn.ftz.f32 %f2843, %f2837, %f2837, %f2842; add.ftz.f32 %f2844, %f2841, %f2843; rsqrt.approx.ftz.f32 %f2845, %f2844; mul.ftz.f32 %f6673, %f2839, %f2845; mul.ftz.f32 %f6672, %f2837, %f2845; mul.ftz.f32 %f6671, %f2835, %f2845; mul.ftz.f32 %f6670, %f2833, %f2845; bra.uni $L__BB24_143; $L__BB24_141: mov.f32 %f2779, 0fBF800000; max.ftz.f32 %f2780, %f419, %f2779; min.ftz.f32 %f2782, %f2780, %f2778; abs.ftz.f32 %f2783, %f2782; sub.ftz.f32 %f2784, %f2778, %f2783; mul.ftz.f32 %f2785, %f2784, 0f3F000000; sqrt.approx.ftz.f32 %f2786, %f2785; setp.gt.ftz.f32 %p233, %f2783, 0f3F11EB85; selp.f32 %f2787, %f2786, %f2783, %p233; mul.ftz.f32 %f2788, %f2787, %f2787; mov.f32 %f2789, 0f3C94D2E9; mov.f32 %f2790, 0f3D53F941; fma.rn.ftz.f32 %f2791, %f2790, %f2788, %f2789; mov.f32 %f2792, 0f3D3F841F; fma.rn.ftz.f32 %f2793, %f2791, %f2788, %f2792; mov.f32 %f2794, 0f3D994929; fma.rn.ftz.f32 %f2795, %f2793, %f2788, %f2794; mov.f32 %f2796, 0f3E2AAB94; fma.rn.ftz.f32 %f2797, %f2795, %f2788, %f2796; mul.ftz.f32 %f2798, %f2788, %f2797; fma.rn.ftz.f32 %f2799, %f2798, %f2787, %f2787; add.ftz.f32 %f2800, %f2799, %f2799; mov.f32 %f2801, 0f3FC90FDB; sub.ftz.f32 %f2802, %f2801, %f2799; selp.f32 %f2803, %f2800, %f2802, %p233; setp.lt.ftz.f32 %p234, %f2782, 0f00000000; mov.f32 %f2804, 0f40490FDB; sub.ftz.f32 %f2805, %f2804, %f2803; selp.f32 %f2806, %f2805, %f2803, %p234; mul.ftz.f32 %f2807, %f2766, %f419; sub.ftz.f32 %f2808, %f2770, %f2807; mul.ftz.f32 %f2809, %f2767, %f419; sub.ftz.f32 %f2810, %f2771, %f2809; mul.ftz.f32 %f2811, %f2768, %f419; sub.ftz.f32 %f2812, %f2772, %f2811; mul.ftz.f32 %f2813, %f2769, %f419; sub.ftz.f32 %f2814, %f2773, %f2813; mul.ftz.f32 %f2815, %f2810, %f2810; fma.rn.ftz.f32 %f2816, %f2808, %f2808, %f2815; mul.ftz.f32 %f2817, %f2814, %f2814; fma.rn.ftz.f32 %f2818, %f2812, %f2812, %f2817; add.ftz.f32 %f2819, %f2816, %f2818; rsqrt.approx.ftz.f32 %f2820, %f2819; mul.ftz.f32 %f2821, %f2808, %f2820; mul.ftz.f32 %f2822, %f2810, %f2820; mul.ftz.f32 %f2823, %f2812, %f2820; mul.ftz.f32 %f2824, %f2814, %f2820; mul.ftz.f32 %f2825, %f410, %f2806; cos.approx.ftz.f32 %f2826, %f2825; sin.approx.ftz.f32 %f2827, %f2825; mul.ftz.f32 %f2828, %f2821, %f2827; mul.ftz.f32 %f2829, %f2822, %f2827; mul.ftz.f32 %f2830, %f2823, %f2827; mul.ftz.f32 %f2831, %f2824, %f2827; fma.rn.ftz.f32 %f6673, %f2769, %f2826, %f2831; fma.rn.ftz.f32 %f6672, %f2768, %f2826, %f2830; fma.rn.ftz.f32 %f6671, %f2767, %f2826, %f2829; fma.rn.ftz.f32 %f6670, %f2766, %f2826, %f2828; $L__BB24_143: ld.global.v4.f32 {%f2846, %f2847, %f2848, %f2849}, [%rd141+16]; ld.global.v4.f32 {%f2854, %f2855, %f2856, %f2857}, [%rd141+80]; mul.ftz.f32 %f2862, %f410, %f2854; mul.ftz.f32 %f2863, %f410, %f2855; mul.ftz.f32 %f2864, %f410, %f2856; mul.ftz.f32 %f2865, %f410, %f2857; fma.rn.ftz.f32 %f433, %f420, %f2846, %f2862; fma.rn.ftz.f32 %f434, %f420, %f2847, %f2863; fma.rn.ftz.f32 %f435, %f420, %f2848, %f2864; fma.rn.ftz.f32 %f2866, %f420, %f2849, %f2865; ld.global.v4.f32 {%f2867, %f2868, %f2869, %f2870}, [%rd141+32]; ld.global.v4.f32 {%f2875, %f2876, %f2877, %f2878}, [%rd141+96]; mul.ftz.f32 %f2883, %f410, %f2875; mul.ftz.f32 %f2884, %f410, %f2876; mul.ftz.f32 %f2885, %f410, %f2877; mul.ftz.f32 %f2886, %f410, %f2878; fma.rn.ftz.f32 %f2887, %f420, %f2867, %f2883; fma.rn.ftz.f32 %f2888, %f420, %f2868, %f2884; fma.rn.ftz.f32 %f2889, %f420, %f2869, %f2885; fma.rn.ftz.f32 %f2890, %f420, %f2870, %f2886; ld.global.v4.f32 {%f2891, %f2892, %f2893, %f2894}, [%rd141+48]; ld.global.v4.f32 {%f2899, %f2900, %f2901, %f2902}, [%rd141+112]; mul.ftz.f32 %f2907, %f410, %f2899; mul.ftz.f32 %f2908, %f410, %f2900; mul.ftz.f32 %f2909, %f410, %f2901; mul.ftz.f32 %f2910, %f410, %f2902; fma.rn.ftz.f32 %f2911, %f420, %f2891, %f2907; fma.rn.ftz.f32 %f2912, %f420, %f2892, %f2908; fma.rn.ftz.f32 %f2913, %f420, %f2893, %f2909; fma.rn.ftz.f32 %f2914, %f420, %f2894, %f2910; mul.ftz.f32 %f2915, %f6673, 0f3FB504F3; mul.ftz.f32 %f2916, %f6670, 0f3FB504F3; mul.ftz.f32 %f2917, %f6671, 0f3FB504F3; mul.ftz.f32 %f2918, %f6672, 0f3FB504F3; mul.ftz.f32 %f2919, %f2915, %f2916; mul.ftz.f32 %f2920, %f2915, %f2917; mul.ftz.f32 %f2921, %f2915, %f2918; mul.ftz.f32 %f2922, %f2916, %f2916; mul.ftz.f32 %f2923, %f2916, %f2917; mul.ftz.f32 %f2924, %f2916, %f2918; mul.ftz.f32 %f2925, %f2917, %f2917; mul.ftz.f32 %f2926, %f2917, %f2918; mul.ftz.f32 %f2927, %f2918, %f2918; sub.ftz.f32 %f2929, %f2778, %f2925; sub.ftz.f32 %f2930, %f2929, %f2927; sub.ftz.f32 %f2931, %f2923, %f2921; add.ftz.f32 %f2932, %f2920, %f2924; add.ftz.f32 %f2933, %f2921, %f2923; sub.ftz.f32 %f2934, %f2778, %f2922; sub.ftz.f32 %f2935, %f2934, %f2927; sub.ftz.f32 %f2936, %f2926, %f2919; sub.ftz.f32 %f2937, %f2924, %f2920; add.ftz.f32 %f2938, %f2919, %f2926; sub.ftz.f32 %f2939, %f2934, %f2925; mul.ftz.f32 %f2940, %f2931, %f2889; fma.rn.ftz.f32 %f2941, %f2930, %f2866, %f2940; fma.rn.ftz.f32 %f6679, %f2932, %f2912, %f2941; mul.ftz.f32 %f2942, %f2931, %f2890; fma.rn.ftz.f32 %f2943, %f2930, %f2887, %f2942; fma.rn.ftz.f32 %f437, %f2932, %f2913, %f2943; mul.ftz.f32 %f2944, %f2931, %f2911; fma.rn.ftz.f32 %f2945, %f2930, %f2888, %f2944; fma.rn.ftz.f32 %f438, %f2932, %f2914, %f2945; mul.ftz.f32 %f2946, %f2935, %f2889; fma.rn.ftz.f32 %f2947, %f2933, %f2866, %f2946; fma.rn.ftz.f32 %f439, %f2936, %f2912, %f2947; mul.ftz.f32 %f2948, %f2935, %f2890; fma.rn.ftz.f32 %f2949, %f2933, %f2887, %f2948; fma.rn.ftz.f32 %f6678, %f2936, %f2913, %f2949; mul.ftz.f32 %f2950, %f2935, %f2911; fma.rn.ftz.f32 %f2951, %f2933, %f2888, %f2950; fma.rn.ftz.f32 %f441, %f2936, %f2914, %f2951; mul.ftz.f32 %f2952, %f2938, %f2889; fma.rn.ftz.f32 %f2953, %f2937, %f2866, %f2952; fma.rn.ftz.f32 %f442, %f2939, %f2912, %f2953; mul.ftz.f32 %f2954, %f2938, %f2890; fma.rn.ftz.f32 %f2955, %f2937, %f2887, %f2954; fma.rn.ftz.f32 %f443, %f2939, %f2913, %f2955; mul.ftz.f32 %f2956, %f2938, %f2911; fma.rn.ftz.f32 %f2957, %f2937, %f2888, %f2956; fma.rn.ftz.f32 %f6677, %f2939, %f2914, %f2957; mul.ftz.f32 %f2958, %f6678, %f6677; mul.ftz.f32 %f445, %f443, %f441; sub.ftz.f32 %f6676, %f2958, %f445; mul.ftz.f32 %f447, %f443, %f438; mul.ftz.f32 %f2959, %f437, %f6677; sub.ftz.f32 %f6675, %f447, %f2959; mul.ftz.f32 %f449, %f437, %f441; mul.ftz.f32 %f2960, %f6678, %f438; sub.ftz.f32 %f6674, %f449, %f2960; mul.ftz.f32 %f2961, %f439, %f6675; fma.rn.ftz.f32 %f2962, %f6679, %f6676, %f2961; fma.rn.ftz.f32 %f6680, %f442, %f6674, %f2962; setp.neu.ftz.f32 %p235, %f6680, 0f00000000; @%p235 bra $L__BB24_145; add.ftz.f32 %f6679, %f6679, 0f322BCC77; add.ftz.f32 %f6677, %f6677, 0f322BCC77; add.ftz.f32 %f6678, %f6678, 0f322BCC77; mul.ftz.f32 %f2963, %f6678, %f6677; sub.ftz.f32 %f6676, %f2963, %f445; mul.ftz.f32 %f2964, %f437, %f6677; sub.ftz.f32 %f6675, %f447, %f2964; mul.ftz.f32 %f2965, %f6678, %f438; sub.ftz.f32 %f6674, %f449, %f2965; mul.ftz.f32 %f2966, %f439, %f6675; fma.rn.ftz.f32 %f2967, %f6679, %f6676, %f2966; fma.rn.ftz.f32 %f2968, %f442, %f6674, %f2967; setp.eq.ftz.f32 %p236, %f2968, 0f00000000; selp.f32 %f6680, 0f7F7FFFFF, %f2968, %p236; $L__BB24_145: mul.ftz.f32 %f2969, %f439, %f6677; mul.ftz.f32 %f2970, %f442, %f441; sub.ftz.f32 %f2971, %f2970, %f2969; mul.ftz.f32 %f2972, %f442, %f438; mul.ftz.f32 %f2973, %f6677, %f6679; sub.ftz.f32 %f2974, %f2973, %f2972; mul.ftz.f32 %f2975, %f441, %f6679; mul.ftz.f32 %f2976, %f439, %f438; sub.ftz.f32 %f2977, %f2976, %f2975; div.approx.ftz.f32 %f6685, %f2971, %f6680; div.approx.ftz.f32 %f6686, %f2974, %f6680; div.approx.ftz.f32 %f6687, %f2977, %f6680; mul.ftz.f32 %f2978, %f442, %f6678; mul.ftz.f32 %f2979, %f439, %f443; sub.ftz.f32 %f2980, %f2979, %f2978; mul.ftz.f32 %f2981, %f443, %f6679; mul.ftz.f32 %f2982, %f442, %f437; sub.ftz.f32 %f2983, %f2982, %f2981; mul.ftz.f32 %f2984, %f439, %f437; mul.ftz.f32 %f2985, %f6678, %f6679; sub.ftz.f32 %f2986, %f2985, %f2984; div.approx.ftz.f32 %f6689, %f2980, %f6680; div.approx.ftz.f32 %f6690, %f2983, %f6680; div.approx.ftz.f32 %f6691, %f2986, %f6680; div.approx.ftz.f32 %f6681, %f6676, %f6680; div.approx.ftz.f32 %f6682, %f6675, %f6680; mul.ftz.f32 %f2987, %f434, %f6682; fma.rn.ftz.f32 %f2988, %f433, %f6681, %f2987; div.approx.ftz.f32 %f6683, %f6674, %f6680; fma.rn.ftz.f32 %f2989, %f435, %f6683, %f2988; neg.ftz.f32 %f6684, %f2989; mul.ftz.f32 %f2990, %f434, %f6686; fma.rn.ftz.f32 %f2991, %f433, %f6685, %f2990; fma.rn.ftz.f32 %f2992, %f435, %f6687, %f2991; neg.ftz.f32 %f6688, %f2992; mul.ftz.f32 %f2993, %f434, %f6690; fma.rn.ftz.f32 %f2994, %f433, %f6689, %f2993; fma.rn.ftz.f32 %f2995, %f435, %f6691, %f2994; neg.ftz.f32 %f6692, %f2995; $L__BB24_147: mul.ftz.f32 %f3008, %f6682, %f6616; fma.rn.ftz.f32 %f3009, %f6681, %f6460, %f3008; fma.rn.ftz.f32 %f3010, %f6683, %f6615, %f3009; add.ftz.f32 %f6905, %f6684, %f3010; mul.ftz.f32 %f3011, %f6686, %f6616; fma.rn.ftz.f32 %f3012, %f6685, %f6460, %f3011; fma.rn.ftz.f32 %f3013, %f6687, %f6615, %f3012; add.ftz.f32 %f6904, %f6688, %f3013; mul.ftz.f32 %f3014, %f6690, %f6616; fma.rn.ftz.f32 %f3015, %f6689, %f6460, %f3014; fma.rn.ftz.f32 %f3016, %f6691, %f6615, %f3015; add.ftz.f32 %f6903, %f6692, %f3016; mul.ftz.f32 %f3017, %f6682, %f6456; fma.rn.ftz.f32 %f3018, %f6681, %f6457, %f3017; fma.rn.ftz.f32 %f3019, %f6683, %f6455, %f3018; mul.ftz.f32 %f3020, %f6686, %f6456; fma.rn.ftz.f32 %f3021, %f6685, %f6457, %f3020; fma.rn.ftz.f32 %f3022, %f6687, %f6455, %f3021; mul.ftz.f32 %f3023, %f6690, %f6456; fma.rn.ftz.f32 %f3024, %f6689, %f6457, %f3023; fma.rn.ftz.f32 %f3025, %f6691, %f6455, %f3024; abs.ftz.f32 %f3026, %f3019; setp.gt.ftz.f32 %p237, %f3026, 0f17800000; mov.b32 %r849, %f3019; and.b32 %r850, %r849, -2147483648; or.b32 %r851, %r850, 394264576; mov.b32 %f3027, %r851; selp.f32 %f6908, %f3019, %f3027, %p237; abs.ftz.f32 %f3028, %f3022; setp.gt.ftz.f32 %p238, %f3028, 0f17800000; mov.b32 %r852, %f3022; and.b32 %r853, %r852, -2147483648; or.b32 %r854, %r853, 394264576; mov.b32 %f3029, %r854; selp.f32 %f6907, %f3022, %f3029, %p238; abs.ftz.f32 %f3030, %f3025; setp.gt.ftz.f32 %p239, %f3030, 0f17800000; mov.b32 %r855, %f3025; and.b32 %r856, %r855, -2147483648; or.b32 %r857, %r856, 394264576; mov.b32 %f3031, %r857; selp.f32 %f6906, %f3025, %f3031, %p239; rcp.approx.ftz.f32 %f6911, %f6908; rcp.approx.ftz.f32 %f6910, %f6907; rcp.approx.ftz.f32 %f6909, %f6906; add.s32 %r1315, %r1315, 1; mov.u32 %r858, 1985229328; st.local.u32 [%rd138+4], %r858; shl.b64 %rd579, %rd139, 2; add.s64 %rd580, %rd134, %rd579; ld.global.u32 %r1314, [%rd580]; bra.uni $L__BB24_326; $L__BB24_126: setp.eq.s32 %p207, %r1314, 1985229328; setp.lt.s32 %p208, %r1314, 0; or.pred %p209, %p208, %p207; @%p209 bra $L__BB24_137; $L__BB24_127: mul.wide.s32 %rd559, %r1314, 16; add.s64 %rd136, %rd112, %rd559; ld.global.v4.f32 {%f2549, %f2550, %f2551, %f2552}, [%rd136]; mov.b32 %r1314, %f2552; mov.b32 %r177, %f2551; mov.b32 %r176, %f2550; mov.b32 %r175, %f2549; mov.b64 %rd560, {%r175, %r176}; cvt.u32.u64 %r830, %rd560; and.b32 %r831, %r830, 1024; setp.eq.s32 %p210, %r831, 0; and.b32 %r179, %r175, %r668; @%p210 bra $L__BB24_130; bra.uni $L__BB24_128; $L__BB24_130: setp.ne.s32 %p216, %r179, 0; ld.global.v4.f32 {%f2692, %f2693, %f2694, %f2695}, [%rd136+16]; ld.global.v4.f32 {%f2700, %f2701, %f2702, %f2703}, [%rd136+32]; ld.global.v4.f32 {%f2708, %f2709, %f2710, %f2711}, [%rd136+48]; sub.ftz.f32 %f2716, %f2692, %f6905; mul.ftz.f32 %f2717, %f6911, %f2716; sub.ftz.f32 %f2718, %f2694, %f6905; mul.ftz.f32 %f2719, %f6911, %f2718; sub.ftz.f32 %f2720, %f2700, %f6904; mul.ftz.f32 %f2721, %f6910, %f2720; sub.ftz.f32 %f2722, %f2702, %f6904; mul.ftz.f32 %f2723, %f6910, %f2722; sub.ftz.f32 %f2724, %f2708, %f6903; mul.ftz.f32 %f2725, %f6909, %f2724; sub.ftz.f32 %f2726, %f2710, %f6903; mul.ftz.f32 %f2727, %f6909, %f2726; min.ftz.f32 %f2728, %f2717, %f2719; min.ftz.f32 %f2729, %f2721, %f2723; min.ftz.f32 %f2730, %f2725, %f2727; max.ftz.f32 %f2731, %f340, %f2728; max.ftz.f32 %f2732, %f2729, %f2730; max.ftz.f32 %f6669, %f2731, %f2732; max.ftz.f32 %f2733, %f2717, %f2719; max.ftz.f32 %f2734, %f2721, %f2723; max.ftz.f32 %f2735, %f2725, %f2727; min.ftz.f32 %f2736, %f341, %f2733; min.ftz.f32 %f2737, %f2734, %f2735; min.ftz.f32 %f2738, %f2736, %f2737; sub.ftz.f32 %f2739, %f2693, %f6905; mul.ftz.f32 %f2740, %f6911, %f2739; sub.ftz.f32 %f2741, %f2695, %f6905; mul.ftz.f32 %f2742, %f6911, %f2741; sub.ftz.f32 %f2743, %f2701, %f6904; mul.ftz.f32 %f2744, %f6910, %f2743; sub.ftz.f32 %f2745, %f2703, %f6904; mul.ftz.f32 %f2746, %f6910, %f2745; sub.ftz.f32 %f2747, %f2709, %f6903; mul.ftz.f32 %f2748, %f6909, %f2747; sub.ftz.f32 %f2749, %f2711, %f6903; mul.ftz.f32 %f2750, %f6909, %f2749; min.ftz.f32 %f2751, %f2740, %f2742; min.ftz.f32 %f2752, %f2744, %f2746; min.ftz.f32 %f2753, %f2748, %f2750; max.ftz.f32 %f2754, %f340, %f2751; max.ftz.f32 %f2755, %f2752, %f2753; max.ftz.f32 %f6668, %f2754, %f2755; max.ftz.f32 %f2756, %f2740, %f2742; max.ftz.f32 %f2757, %f2744, %f2746; max.ftz.f32 %f2758, %f2748, %f2750; min.ftz.f32 %f2759, %f341, %f2756; min.ftz.f32 %f2760, %f2757, %f2758; min.ftz.f32 %f2761, %f2759, %f2760; setp.ge.ftz.f32 %p217, %f2738, %f6669; and.pred %p218, %p216, %p217; selp.u32 %r835, 1, 0, %p218; setp.ltu.ftz.f32 %p219, %f2761, %f6668; and.b32 %r837, %r176, %r668; setp.eq.s32 %p220, %r837, 0; or.pred %p221, %p220, %p219; selp.b32 %r838, 0, 2, %p221; or.b32 %r1313, %r838, %r835; bra.uni $L__BB24_131; $L__BB24_128: setp.ne.s32 %p211, %r179, 0; ld.global.v4.f32 {%f2556, %f2557, %f2558, %f2559}, [%rd136+16]; ld.global.v4.f32 {%f2564, %f2565, %f2566, %f2567}, [%rd136+32]; ld.global.v4.f32 {%f2572, %f2573, %f2574, %f2575}, [%rd136+48]; mul.ftz.f32 %f2580, %f6907, %f2557; fma.rn.ftz.f32 %f2581, %f6908, %f2556, %f2580; fma.rn.ftz.f32 %f2582, %f6906, %f2558, %f2581; mul.ftz.f32 %f2583, %f6907, %f2565; fma.rn.ftz.f32 %f2584, %f6908, %f2564, %f2583; fma.rn.ftz.f32 %f2585, %f6906, %f2566, %f2584; mul.ftz.f32 %f2586, %f6907, %f2573; fma.rn.ftz.f32 %f2587, %f6908, %f2572, %f2586; fma.rn.ftz.f32 %f2588, %f6906, %f2574, %f2587; mul.ftz.f32 %f2589, %f6904, %f2557; fma.rn.ftz.f32 %f2590, %f6905, %f2556, %f2589; fma.rn.ftz.f32 %f2591, %f6903, %f2558, %f2590; add.ftz.f32 %f2592, %f2559, %f2591; mul.ftz.f32 %f2593, %f6904, %f2565; fma.rn.ftz.f32 %f2594, %f6905, %f2564, %f2593; fma.rn.ftz.f32 %f2595, %f6903, %f2566, %f2594; add.ftz.f32 %f2596, %f2567, %f2595; mul.ftz.f32 %f2597, %f6904, %f2573; fma.rn.ftz.f32 %f2598, %f6905, %f2572, %f2597; fma.rn.ftz.f32 %f2599, %f6903, %f2574, %f2598; add.ftz.f32 %f2600, %f2575, %f2599; rcp.approx.ftz.f32 %f2601, %f2582; rcp.approx.ftz.f32 %f2602, %f2585; rcp.approx.ftz.f32 %f2603, %f2588; mul.ftz.f32 %f2604, %f2601, %f2592; neg.ftz.f32 %f2605, %f2604; mul.ftz.f32 %f2606, %f2602, %f2596; neg.ftz.f32 %f2607, %f2606; mul.ftz.f32 %f2608, %f2603, %f2600; neg.ftz.f32 %f2609, %f2608; sub.ftz.f32 %f2610, %f2601, %f2604; sub.ftz.f32 %f2611, %f2602, %f2606; sub.ftz.f32 %f2612, %f2603, %f2608; min.ftz.f32 %f2613, %f2605, %f2610; min.ftz.f32 %f2614, %f2607, %f2611; min.ftz.f32 %f2615, %f2609, %f2612; max.ftz.f32 %f2616, %f2605, %f2610; max.ftz.f32 %f2617, %f2607, %f2611; max.ftz.f32 %f2618, %f2609, %f2612; max.ftz.f32 %f2619, %f340, %f2613; max.ftz.f32 %f2620, %f2614, %f2615; max.ftz.f32 %f6669, %f2619, %f2620; min.ftz.f32 %f2621, %f341, %f2616; min.ftz.f32 %f2622, %f2617, %f2618; min.ftz.f32 %f2623, %f2621, %f2622; setp.le.ftz.f32 %p212, %f6669, %f2623; and.pred %p213, %p211, %p212; selp.u32 %r1313, 1, 0, %p213; ld.global.v4.f32 {%f2624, %f2625, %f2626, %f2627}, [%rd136+64]; ld.global.v4.f32 {%f2632, %f2633, %f2634, %f2635}, [%rd136+80]; ld.global.v4.f32 {%f2640, %f2641, %f2642, %f2643}, [%rd136+96]; mul.ftz.f32 %f2648, %f6907, %f2625; fma.rn.ftz.f32 %f2649, %f6908, %f2624, %f2648; fma.rn.ftz.f32 %f2650, %f6906, %f2626, %f2649; mul.ftz.f32 %f2651, %f6907, %f2633; fma.rn.ftz.f32 %f2652, %f6908, %f2632, %f2651; fma.rn.ftz.f32 %f2653, %f6906, %f2634, %f2652; mul.ftz.f32 %f2654, %f6907, %f2641; fma.rn.ftz.f32 %f2655, %f6908, %f2640, %f2654; fma.rn.ftz.f32 %f2656, %f6906, %f2642, %f2655; mul.ftz.f32 %f2657, %f6904, %f2625; fma.rn.ftz.f32 %f2658, %f6905, %f2624, %f2657; fma.rn.ftz.f32 %f2659, %f6903, %f2626, %f2658; add.ftz.f32 %f2660, %f2627, %f2659; mul.ftz.f32 %f2661, %f6904, %f2633; fma.rn.ftz.f32 %f2662, %f6905, %f2632, %f2661; fma.rn.ftz.f32 %f2663, %f6903, %f2634, %f2662; add.ftz.f32 %f2664, %f2635, %f2663; mul.ftz.f32 %f2665, %f6904, %f2641; fma.rn.ftz.f32 %f2666, %f6905, %f2640, %f2665; fma.rn.ftz.f32 %f2667, %f6903, %f2642, %f2666; add.ftz.f32 %f2668, %f2643, %f2667; rcp.approx.ftz.f32 %f2669, %f2650; rcp.approx.ftz.f32 %f2670, %f2653; rcp.approx.ftz.f32 %f2671, %f2656; mul.ftz.f32 %f2672, %f2669, %f2660; neg.ftz.f32 %f2673, %f2672; mul.ftz.f32 %f2674, %f2670, %f2664; neg.ftz.f32 %f2675, %f2674; mul.ftz.f32 %f2676, %f2671, %f2668; neg.ftz.f32 %f2677, %f2676; sub.ftz.f32 %f2678, %f2669, %f2672; sub.ftz.f32 %f2679, %f2670, %f2674; sub.ftz.f32 %f2680, %f2671, %f2676; min.ftz.f32 %f2681, %f2673, %f2678; min.ftz.f32 %f2682, %f2675, %f2679; min.ftz.f32 %f2683, %f2677, %f2680; max.ftz.f32 %f2684, %f2673, %f2678; max.ftz.f32 %f2685, %f2675, %f2679; max.ftz.f32 %f2686, %f2677, %f2680; max.ftz.f32 %f2687, %f340, %f2681; max.ftz.f32 %f2688, %f2682, %f2683; max.ftz.f32 %f6668, %f2687, %f2688; min.ftz.f32 %f2689, %f341, %f2684; min.ftz.f32 %f2690, %f2685, %f2686; min.ftz.f32 %f2691, %f2689, %f2690; setp.gtu.ftz.f32 %p214, %f6668, %f2691; @%p214 bra $L__BB24_131; and.b32 %r833, %r176, %r668; setp.eq.s32 %p215, %r833, 0; or.b32 %r834, %r1313, 2; selp.b32 %r1313, %r1313, %r834, %p215; $L__BB24_131: mul.wide.s32 %rd561, %r1315, 4; add.s64 %rd137, %rd6, %rd561; setp.eq.s32 %p222, %r1313, 3; @%p222 bra $L__BB24_135; bra.uni $L__BB24_132; $L__BB24_135: setp.lt.ftz.f32 %p225, %f6668, %f6669; selp.b32 %r839, %r177, %r1314, %p225; selp.b32 %r1314, %r1314, %r177, %p225; add.s32 %r1315, %r1315, 1; st.local.u32 [%rd137+4], %r839; bra.uni $L__BB24_136; $L__BB24_132: setp.eq.s32 %p223, %r1313, 2; @%p223 bra $L__BB24_136; setp.ne.s32 %p224, %r1313, 0; mov.u32 %r1314, %r177; @%p224 bra $L__BB24_136; ld.local.u32 %r1314, [%rd137]; add.s32 %r1315, %r1315, -1; $L__BB24_136: setp.ne.s32 %p226, %r1314, 1985229328; setp.gt.s32 %p227, %r1314, -1; and.pred %p228, %p227, %p226; @%p228 bra $L__BB24_127; $L__BB24_137: setp.gt.s32 %p229, %r1314, -1; @%p229 bra $L__BB24_326; bra.uni $L__BB24_138; $L__BB24_326: setp.ne.s32 %p449, %r1314, 1985229328; @%p449 bra $L__BB24_126; setp.lt.s32 %p450, %r1315, 0; mov.u32 %r1314, %r827; @%p450 bra $L__BB24_329; mul.wide.s32 %rd750, %r1315, 4; add.s64 %rd751, %rd6, %rd750; ld.local.u32 %r1314, [%rd751]; add.s32 %r1315, %r1315, -1; mov.u32 %r1377, -1; mov.f32 %f6903, %f6615; mov.f32 %f6904, %f6616; mov.f32 %f6905, %f6460; mov.f32 %f6906, %f7; mov.f32 %f6907, %f6; mov.f32 %f6908, %f5; mov.f32 %f6909, %f10; mov.f32 %f6910, %f9; mov.f32 %f6911, %f8; $L__BB24_329: setp.ne.s32 %p451, %r1314, 1985229328; mov.u16 %rs242, 0; @%p451 bra $L__BB24_126; bra.uni $L__BB24_594; $L__BB24_330: setp.ne.s32 %p452, %r2, 0; @%p452 bra $L__BB24_419; bra.uni $L__BB24_331; $L__BB24_419: mov.u32 %r1084, 1985229328; st.local.u32 [%rd6], %r1084; ld.const.u32 %r1471, [kernel_params+1104]; ld.local.f32 %f7049, [%rd4+4]; ld.local.f32 %f7048, [%rd4+8]; ld.local.f32 %f1291, [%rd4+24]; ld.local.f32 %f1292, [%rd4+28]; ld.const.v2.u64 {%rd921, %rd922}, [kernel_params+2096]; mov.u32 %r1536, 0; cvta.to.global.u64 %rd258, %rd921; cvta.to.global.u64 %rd259, %rd922; ld.const.v2.u64 {%rd925, %rd926}, [kernel_params+2128]; cvta.to.global.u64 %rd260, %rd926; ld.const.u64 %rd929, [kernel_params+2160]; cvta.to.global.u64 %rd261, %rd929; ld.const.v2.u64 {%rd930, %rd931}, [kernel_params+2144]; cvta.to.global.u64 %rd262, %rd930; cvta.to.global.u64 %rd263, %rd925; ld.const.u64 %rd934, [kernel_params+2232]; cvta.to.global.u64 %rd264, %rd934; ld.const.v2.u64 {%rd935, %rd936}, [kernel_params+2256]; cvta.to.global.u64 %rd265, %rd935; ld.const.v2.u32 {%r1085, %r1086}, [kernel_params+1128]; ld.const.v2.u64 {%rd939, %rd940}, [kernel_params+2112]; cvta.to.global.u64 %rd266, %rd939; cvta.to.global.u64 %rd267, %rd936; cvta.to.global.u64 %rd268, %rd931; ld.const.u64 %rd943, [kernel_params+2408]; cvta.to.global.u64 %rd269, %rd943; ld.const.u64 %rd944, [kernel_params+2216]; cvta.to.global.u64 %rd270, %rd944; cvta.to.global.u64 %rd271, %rd940; ld.const.u64 %rd945, [kernel_params+2272]; cvta.to.global.u64 %rd272, %rd945; ld.const.v2.u64 {%rd946, %rd947}, [kernel_params+2304]; cvta.to.global.u64 %rd273, %rd947; ld.const.u64 %rd950, [kernel_params+2296]; cvta.to.global.u64 %rd274, %rd950; cvta.to.global.u64 %rd275, %rd946; cvt.rn.f32.s32 %f4794, %r1086; rcp.approx.ftz.f32 %f1293, %f4794; mov.f32 %f7321, 0f3F800000; sub.ftz.f32 %f4795, %f7321, %f1293; mul.ftz.f32 %f4796, %f1293, %f4795; mul.ftz.f32 %f4797, %f4795, %f4796; neg.ftz.f32 %f1294, %f4797; mul.ftz.f32 %f4798, %f1293, %f1293; fma.rn.ftz.f32 %f4799, %f1293, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f1295, %f4798, %f4799, 0f40000000; mul.ftz.f32 %f4800, %f4795, %f4795; fma.rn.ftz.f32 %f4801, %f4795, 0f40400000, 0fC0A00000; fma.rn.ftz.f32 %f1296, %f4800, %f4801, 0f40000000; mul.ftz.f32 %f4802, %f1293, %f4796; neg.ftz.f32 %f1297, %f4802; ld.const.u64 %rd951, [kernel_params+2344]; cvta.to.global.u64 %rd276, %rd951; ld.const.u64 %rd952, [kernel_params+2320]; cvta.to.global.u64 %rd277, %rd952; ld.const.u64 %rd953, [kernel_params+3552]; cvta.to.global.u64 %rd954, %rd953; shl.b64 %rd955, %rd3, 2; add.s64 %rd278, %rd954, %rd955; mov.u32 %r1529, -1; setp.eq.s32 %p590, %r1085, 0; mov.u16 %rs153, 1; mov.pred %p598, 0; mov.pred %p719, -1; mov.u16 %rs156, 0; mov.f32 %f7286, %f8; mov.f32 %f7287, %f9; mov.f32 %f7288, %f10; mov.f32 %f7289, %f5; mov.f32 %f7290, %f6; mov.f32 %f7291, %f7; mov.f32 %f7292, %f6460; mov.f32 %f7293, %f7049; mov.f32 %f7294, %f7048; mov.f32 %f7304, %f1292; mov.u32 %r1528, %r1536; mov.u32 %r1472, %r1536; bra.uni $L__BB24_420; $L__BB24_432: mul.wide.s32 %rd959, %r1471, 16; xor.b64 %rd960, %rd959, -16; add.s64 %rd961, %rd259, %rd960; ld.global.v4.u32 {%r1475, %r1098, %r1099, %r1100}, [%rd961]; setp.gt.s32 %p585, %r1475, -1; mul.wide.s32 %rd964, %r1472, 4; add.s64 %rd281, %rd6, %rd964; @%p585 bra $L__BB24_434; bra.uni $L__BB24_433; $L__BB24_434: ld.local.u32 %r1471, [%rd281]; add.s32 %r1472, %r1472, -1; setp.le.s32 %p589, %r1098, %r1475; @%p589 bra $L__BB24_590; cvt.u16.u32 %rs152, %r1100; and.b16 %rs61, %rs152, 127; and.b32 %r1111, %r1100, 64; setp.eq.s32 %p591, %r1111, 0; or.pred %p35, %p591, %p590; mul.ftz.f32 %f5064, %f7290, %f7290; fma.rn.ftz.f32 %f5065, %f7289, %f7289, %f5064; fma.rn.ftz.f32 %f1364, %f7291, %f7291, %f5065; rcp.approx.ftz.f32 %f1365, %f1364; $L__BB24_436: cvt.s64.s32 %rd282, %r1475; setp.ne.s32 %p592, %r1529, -1; mov.u32 %r541, %r1529; @%p592 bra $L__BB24_438; shl.b64 %rd972, %rd282, 2; add.s64 %rd973, %rd260, %rd972; ld.global.u32 %r541, [%rd973]; $L__BB24_438: shl.b64 %rd974, %rd282, 2; add.s64 %rd975, %rd263, %rd974; ld.global.u32 %r1112, [%rd975]; cvt.s64.s32 %rd283, %r1112; ld.local.u32 %r1113, [%rd4+36]; setp.ne.s32 %p593, %r1113, %r1112; @%p593 bra $L__BB24_440; ld.local.u32 %r1114, [%rd4+40]; setp.eq.s32 %p594, %r1114, %r541; mov.u16 %rs224, %rs153; @%p594 bra $L__BB24_442; $L__BB24_440: cvt.u32.u64 %r1115, %rd283; ld.local.u32 %r1116, [%rd4+44]; setp.ne.s32 %p595, %r1116, %r1115; mov.u16 %rs224, 0; @%p595 bra $L__BB24_442; ld.local.u32 %r1117, [%rd4+48]; setp.eq.s32 %p596, %r1117, %r541; selp.u16 %rs224, 1, 0, %p596; $L__BB24_442: setp.ne.s16 %p597, %rs224, 0; @%p597 bra $L__BB24_589; shl.b64 %rd976, %rd282, 3; add.s64 %rd284, %rd268, %rd976; add.s64 %rd285, %rd266, %rd974; mul.wide.s32 %rd978, %r541, 208; add.s64 %rd979, %rd261, %rd978; add.s64 %rd286, %rd979, 148; setp.gt.s16 %p599, %rs61, 7; @%p599 bra $L__BB24_447; bra.uni $L__BB24_444; $L__BB24_447: setp.gt.s16 %p600, %rs61, 67; @%p600 bra $L__BB24_456; setp.eq.s16 %p603, %rs61, 8; @%p603 bra $L__BB24_458; setp.eq.s16 %p604, %rs61, 66; mov.pred %p806, %p598; @%p604 bra $L__BB24_450; bra.uni $L__BB24_557; $L__BB24_444: setp.eq.s16 %p605, %rs61, 1; @%p605 bra $L__BB24_549; setp.eq.s16 %p606, %rs61, 2; @%p606 bra $L__BB24_450; setp.eq.s16 %p607, %rs61, 4; mov.pred %p806, %p598; @%p607 bra $L__BB24_450; bra.uni $L__BB24_557; $L__BB24_456: setp.eq.s16 %p601, %rs61, 68; @%p601 bra $L__BB24_450; setp.ne.s16 %p602, %rs61, 72; mov.pred %p806, %p598; @%p602 bra $L__BB24_557; $L__BB24_458: @%p35 bra $L__BB24_460; ld.global.v2.f32 {%f5066, %f5067}, [%rd284]; ld.local.f32 %f5070, [%rd4+32]; setp.lt.ftz.f32 %p609, %f5070, %f5066; setp.gt.ftz.f32 %p610, %f5070, %f5067; or.pred %p611, %p609, %p610; mov.pred %p806, %p598; @%p611 bra $L__BB24_557; $L__BB24_460: ld.global.u32 %r542, [%rd285]; and.b32 %r1118, %r542, 64; setp.eq.s32 %p612, %r1118, 0; @%p612 bra $L__BB24_474; bra.uni $L__BB24_461; $L__BB24_474: shl.b64 %rd988, %rd283, 4; add.s64 %rd989, %rd274, %rd988; ld.global.v4.f32 {%f7110, %f7111, %f7112, %f7113}, [%rd989]; bra.uni $L__BB24_475; $L__BB24_450: @%p35 bra $L__BB24_452; ld.global.v2.f32 {%f5115, %f5116}, [%rd284]; ld.local.f32 %f5119, [%rd4+32]; setp.lt.ftz.f32 %p627, %f5119, %f5115; setp.gt.ftz.f32 %p628, %f5119, %f5116; or.pred %p629, %p627, %p628; mov.pred %p806, %p598; @%p629 bra $L__BB24_557; $L__BB24_452: ld.global.u32 %r560, [%rd285]; and.b32 %r1134, %r560, 64; setp.eq.s32 %p630, %r1134, 0; shl.b64 %rd990, %rd283, 4; add.s64 %rd991, %rd267, %rd990; shr.s32 %r1135, %r560, 7; ld.global.u32 %r1136, [%rd991+4]; add.s32 %r561, %r1136, %r1135; add.s32 %r1137, %r561, -1; max.s32 %r562, %r1137, %r1136; add.s32 %r1138, %r561, 2; ld.global.u32 %r1139, [%rd991+8]; add.s32 %r1140, %r1136, %r1139; add.s32 %r1141, %r1140, -1; min.s32 %r563, %r1138, %r1141; mul.wide.s32 %rd992, %r562, 16; add.s64 %rd296, %rd272, %rd992; mul.wide.s32 %rd993, %r561, 16; add.s64 %rd297, %rd272, %rd993; mul.wide.s32 %rd994, %r563, 16; add.s64 %rd298, %rd272, %rd994; @%p630 bra $L__BB24_488; bra.uni $L__BB24_453; $L__BB24_488: ld.global.v4.f32 {%f7147, %f7148, %f7149, %f1535}, [%rd296]; ld.global.v4.f32 {%f7151, %f7152, %f7153, %f1539}, [%rd297]; ld.global.v4.f32 {%f7155, %f7156, %f7157, %f1543}, [%rd297+16]; ld.global.v4.f32 {%f7159, %f7160, %f7161, %f1547}, [%rd298]; bra.uni $L__BB24_489; $L__BB24_453: ld.global.u32 %r564, [%rd286]; ld.global.u32 %r565, [%rd286+4]; shl.b32 %r1142, %r565, 1; cvt.rn.f32.s32 %f5120, %r1142; ld.local.f32 %f5121, [%rd4+32]; mul.ftz.f32 %f1403, %f5121, %f5120; cvt.rzi.ftz.s32.f32 %r1143, %f1403; add.s32 %r1144, %r1142, -1; min.s32 %r566, %r1143, %r1144; ld.global.u32 %r1488, [%rd286+16]; mul.wide.u32 %rd995, %r1488, 16; add.s64 %rd996, %rd273, %rd995; ld.global.u16 %rs227, [%rd996+12]; ld.global.u32 %r1487, [%rd996+8]; ld.global.u64 %rd1228, [%rd996]; setp.eq.s64 %p631, %rd1228, 11; @%p631 bra $L__BB24_480; $L__BB24_454: setp.eq.s64 %p632, %rd1228, 0; @%p632 bra $L__BB24_478; add.s32 %r1488, %r1488, 2; bra.uni $L__BB24_479; $L__BB24_478: setp.eq.s16 %p633, %rs227, 0; mov.u32 %r1491, -1; mov.u32 %r1488, %r1487; @%p633 bra $L__BB24_481; $L__BB24_479: mul.wide.u32 %rd997, %r1488, 16; add.s64 %rd998, %rd273, %rd997; ld.global.u16 %rs227, [%rd998+12]; ld.global.u32 %r1487, [%rd998+8]; ld.global.u64 %rd1228, [%rd998]; setp.ne.s64 %p634, %rd1228, 11; @%p634 bra $L__BB24_454; $L__BB24_480: setp.eq.s16 %p635, %rs227, 0; selp.b32 %r1491, -1, %r1487, %p635; $L__BB24_481: setp.eq.s32 %p636, %r566, %r565; @%p636 bra $L__BB24_483; bra.uni $L__BB24_482; $L__BB24_483: ld.global.v4.f32 {%f7115, %f7116, %f7117, %f7118}, [%rd296]; ld.global.v4.f32 {%f7119, %f7120, %f7121, %f7122}, [%rd297]; ld.global.v4.f32 {%f7123, %f7124, %f7125, %f7126}, [%rd297+16]; ld.global.v4.f32 {%f7127, %f7128, %f7129, %f7130}, [%rd298]; bra.uni $L__BB24_484; $L__BB24_549: shl.b64 %rd1015, %rd283, 4; add.s64 %rd1016, %rd264, %rd1015; ld.global.u32 %r1168, [%rd1016+12]; mul.wide.u32 %rd1017, %r1168, 12; add.s64 %rd1018, %rd265, %rd1017; add.s32 %r1169, %r1168, 1; mul.wide.u32 %rd1019, %r1169, 12; add.s64 %rd1020, %rd265, %rd1019; add.s32 %r1170, %r1168, 2; mul.wide.u32 %rd1021, %r1170, 12; add.s64 %rd1022, %rd265, %rd1021; ld.global.f32 %f6336, [%rd1018]; sub.ftz.f32 %f1807, %f6336, %f7292; ld.global.f32 %f6337, [%rd1018+4]; sub.ftz.f32 %f1808, %f6337, %f7293; ld.global.f32 %f6338, [%rd1018+8]; sub.ftz.f32 %f1809, %f6338, %f7294; ld.global.f32 %f6339, [%rd1020]; sub.ftz.f32 %f6340, %f6339, %f7292; ld.global.f32 %f6341, [%rd1020+4]; sub.ftz.f32 %f6342, %f6341, %f7293; ld.global.f32 %f6343, [%rd1020+8]; sub.ftz.f32 %f6344, %f6343, %f7294; ld.global.f32 %f6345, [%rd1022]; sub.ftz.f32 %f6346, %f6345, %f7292; ld.global.f32 %f6347, [%rd1022+4]; sub.ftz.f32 %f6348, %f6347, %f7293; ld.global.f32 %f6349, [%rd1022+8]; sub.ftz.f32 %f6350, %f6349, %f7294; sub.ftz.f32 %f1810, %f6346, %f1807; sub.ftz.f32 %f1811, %f6348, %f1808; sub.ftz.f32 %f1812, %f6350, %f1809; sub.ftz.f32 %f1813, %f1807, %f6340; sub.ftz.f32 %f1814, %f1808, %f6342; sub.ftz.f32 %f1815, %f1809, %f6344; sub.ftz.f32 %f6351, %f6340, %f6346; sub.ftz.f32 %f6352, %f6342, %f6348; sub.ftz.f32 %f6353, %f6344, %f6350; add.ftz.f32 %f6354, %f1807, %f6346; add.ftz.f32 %f6355, %f1808, %f6348; add.ftz.f32 %f6356, %f1809, %f6350; mul.ftz.f32 %f6357, %f1811, %f6356; mul.ftz.f32 %f6358, %f6355, %f1812; sub.ftz.f32 %f6359, %f6357, %f6358; mul.ftz.f32 %f6360, %f6354, %f1812; mul.ftz.f32 %f6361, %f1810, %f6356; sub.ftz.f32 %f6362, %f6360, %f6361; mul.ftz.f32 %f6363, %f1810, %f6355; mul.ftz.f32 %f6364, %f6354, %f1811; sub.ftz.f32 %f6365, %f6363, %f6364; mul.ftz.f32 %f6366, %f7290, %f6362; fma.rn.ftz.f32 %f6367, %f7289, %f6359, %f6366; fma.rn.ftz.f32 %f1816, %f7291, %f6365, %f6367; add.ftz.f32 %f6368, %f1807, %f6340; add.ftz.f32 %f6369, %f1808, %f6342; add.ftz.f32 %f6370, %f1809, %f6344; mul.ftz.f32 %f6371, %f1814, %f6370; mul.ftz.f32 %f6372, %f6369, %f1815; sub.ftz.f32 %f6373, %f6371, %f6372; mul.ftz.f32 %f6374, %f6368, %f1815; mul.ftz.f32 %f6375, %f1813, %f6370; sub.ftz.f32 %f6376, %f6374, %f6375; mul.ftz.f32 %f6377, %f1813, %f6369; mul.ftz.f32 %f6378, %f6368, %f1814; sub.ftz.f32 %f6379, %f6377, %f6378; mul.ftz.f32 %f6380, %f7290, %f6376; fma.rn.ftz.f32 %f6381, %f7289, %f6373, %f6380; fma.rn.ftz.f32 %f1817, %f7291, %f6379, %f6381; add.ftz.f32 %f6382, %f6340, %f6346; add.ftz.f32 %f6383, %f6342, %f6348; add.ftz.f32 %f6384, %f6344, %f6350; mul.ftz.f32 %f6385, %f6352, %f6384; mul.ftz.f32 %f6386, %f6383, %f6353; sub.ftz.f32 %f6387, %f6385, %f6386; mul.ftz.f32 %f6388, %f6382, %f6353; mul.ftz.f32 %f6389, %f6351, %f6384; sub.ftz.f32 %f6390, %f6388, %f6389; mul.ftz.f32 %f6391, %f6351, %f6383; mul.ftz.f32 %f6392, %f6382, %f6352; sub.ftz.f32 %f6393, %f6391, %f6392; mul.ftz.f32 %f6394, %f7290, %f6390; fma.rn.ftz.f32 %f6395, %f7289, %f6387, %f6394; fma.rn.ftz.f32 %f6396, %f7291, %f6393, %f6395; add.ftz.f32 %f6397, %f1817, %f1816; add.ftz.f32 %f1818, %f6396, %f6397; abs.ftz.f32 %f1819, %f1818; mul.ftz.f32 %f6398, %f1819, 0f34000000; min.ftz.f32 %f6399, %f1817, %f6396; min.ftz.f32 %f6400, %f1816, %f6399; max.ftz.f32 %f6401, %f1817, %f6396; max.ftz.f32 %f6402, %f1816, %f6401; mul.ftz.f32 %f6403, %f1819, 0fB4000000; setp.ltu.ftz.f32 %p720, %f6400, %f6403; setp.gtu.ftz.f32 %p721, %f6402, %f6398; and.pred %p722, %p720, %p721; mov.pred %p805, %p719; @%p722 bra $L__BB24_555; mul.ftz.f32 %f6404, %f1814, %f1812; mul.ftz.f32 %f6405, %f1815, %f1811; sub.ftz.f32 %f6406, %f6404, %f6405; mul.ftz.f32 %f6407, %f1813, %f1812; mul.ftz.f32 %f6408, %f1815, %f1810; sub.ftz.f32 %f6409, %f6408, %f6407; mul.ftz.f32 %f6410, %f1814, %f1810; mul.ftz.f32 %f6411, %f1813, %f1811; sub.ftz.f32 %f6412, %f6411, %f6410; add.ftz.f32 %f1820, %f6406, %f6406; add.ftz.f32 %f1821, %f6409, %f6409; add.ftz.f32 %f1822, %f6412, %f6412; mul.ftz.f32 %f6413, %f7290, %f1821; fma.rn.ftz.f32 %f6414, %f7289, %f1820, %f6413; fma.rn.ftz.f32 %f1823, %f7291, %f1822, %f6414; setp.eq.ftz.f32 %p724, %f1823, 0f00000000; mov.pred %p805, %p719; @%p724 bra $L__BB24_555; mul.ftz.f32 %f6415, %f1808, %f1821; fma.rn.ftz.f32 %f6416, %f1807, %f1820, %f6415; fma.rn.ftz.f32 %f6417, %f1809, %f1822, %f6416; div.approx.ftz.f32 %f1824, %f6417, %f1823; setp.ltu.ftz.f32 %p726, %f1824, %f1291; setp.gtu.ftz.f32 %p727, %f1824, %f1292; or.pred %p728, %p727, %p726; mov.pred %p805, %p719; @%p728 bra $L__BB24_555; setp.lt.ftz.f32 %p729, %f1819, 0f219392EF; mov.f32 %f7247, 0f00000000; @%p729 bra $L__BB24_554; rcp.approx.ftz.f32 %f7247, %f1818; $L__BB24_554: mul.ftz.f32 %f6419, %f1816, %f7247; mov.f32 %f6420, 0f3F800000; min.ftz.f32 %f7255, %f6419, %f6420; mul.ftz.f32 %f6421, %f1817, %f7247; min.ftz.f32 %f7254, %f6421, %f6420; mov.pred %p805, 0; mov.f32 %f7256, %f1824; $L__BB24_555: mov.pred %p806, %p598; @%p805 bra $L__BB24_557; cvt.u32.u64 %r1171, %rd283; add.s64 %rd1024, %rd271, %rd974; ld.global.u32 %r1172, [%rd1024]; and.b32 %r1173, %r1172, %r668; setp.eq.s32 %p732, %r1173, 0; selp.f32 %f7251, %f7251, %f7256, %p732; selp.f32 %f7252, %f7252, %f7255, %p732; selp.f32 %f7253, %f7253, %f7254, %p732; selp.b32 %r1502, %r1502, %r1171, %p732; selp.b32 %r1503, %r1503, %r541, %p732; selp.b32 %r1504, %r1504, 1, %p732; setp.ne.s32 %p806, %r1173, 0; bra.uni $L__BB24_557; $L__BB24_461: ld.global.u32 %r543, [%rd286]; ld.global.u32 %r544, [%rd286+4]; shl.b32 %r1119, %r544, 1; cvt.rn.f32.s32 %f5071, %r1119; ld.local.f32 %f5072, [%rd4+32]; mul.ftz.f32 %f1379, %f5072, %f5071; cvt.rzi.ftz.s32.f32 %r1120, %f1379; add.s32 %r1121, %r1119, -1; min.s32 %r545, %r1120, %r1121; ld.global.u32 %r1483, [%rd286+16]; mul.wide.u32 %rd980, %r1483, 16; add.s64 %rd981, %rd273, %rd980; ld.global.u16 %rs225, [%rd981+12]; ld.global.u32 %r1482, [%rd981+8]; ld.global.u64 %rd1225, [%rd981]; setp.eq.s64 %p613, %rd1225, 11; @%p613 bra $L__BB24_466; $L__BB24_462: setp.eq.s64 %p614, %rd1225, 0; @%p614 bra $L__BB24_464; add.s32 %r1483, %r1483, 2; bra.uni $L__BB24_465; $L__BB24_464: setp.eq.s16 %p615, %rs225, 0; mov.u32 %r1486, -1; mov.u32 %r1483, %r1482; @%p615 bra $L__BB24_467; $L__BB24_465: mul.wide.u32 %rd982, %r1483, 16; add.s64 %rd983, %rd273, %rd982; ld.global.u16 %rs225, [%rd983+12]; ld.global.u32 %r1482, [%rd983+8]; ld.global.u64 %rd1225, [%rd983]; setp.ne.s64 %p616, %rd1225, 11; @%p616 bra $L__BB24_462; $L__BB24_466: setp.eq.s16 %p617, %rs225, 0; selp.b32 %r1486, -1, %r1482, %p617; $L__BB24_467: setp.eq.s32 %p618, %r545, %r544; @%p618 bra $L__BB24_469; bra.uni $L__BB24_468; $L__BB24_469: shl.b64 %rd985, %rd283, 4; add.s64 %rd1226, %rd274, %rd985; bra.uni $L__BB24_470; $L__BB24_482: setp.gt.s32 %p637, %r566, %r565; selp.b32 %r1146, -1, 0, %p637; add.s32 %r1147, %r566, %r1146; mad.lo.s32 %r1148, %r1147, %r564, %r1491; add.s32 %r1149, %r1148, %r562; mul.wide.s32 %rd999, %r1149, 16; add.s64 %rd1000, %rd276, %rd999; ld.global.v4.f32 {%f7115, %f7116, %f7117, %f7118}, [%rd1000]; add.s32 %r1150, %r1148, %r561; mul.wide.s32 %rd1001, %r1150, 16; add.s64 %rd1002, %rd276, %rd1001; ld.global.v4.f32 {%f7119, %f7120, %f7121, %f7122}, [%rd1002]; ld.global.v4.f32 {%f7123, %f7124, %f7125, %f7126}, [%rd1002+16]; add.s32 %r1151, %r1148, %r563; mul.wide.s32 %rd1003, %r1151, 16; add.s64 %rd1004, %rd276, %rd1003; ld.global.v4.f32 {%f7127, %f7128, %f7129, %f7130}, [%rd1004]; $L__BB24_484: add.s32 %r577, %r566, 1; setp.eq.s32 %p638, %r577, %r565; @%p638 bra $L__BB24_486; bra.uni $L__BB24_485; $L__BB24_486: ld.global.v4.f32 {%f7131, %f7132, %f7133, %f7134}, [%rd296]; ld.global.v4.f32 {%f7135, %f7136, %f7137, %f7138}, [%rd297]; ld.global.v4.f32 {%f7139, %f7140, %f7141, %f7142}, [%rd297+16]; ld.global.v4.f32 {%f7143, %f7144, %f7145, %f7146}, [%rd298]; bra.uni $L__BB24_487; $L__BB24_485: setp.ge.s32 %p639, %r566, %r565; selp.b32 %r1152, -1, 0, %p639; add.s32 %r1153, %r577, %r1152; mad.lo.s32 %r1154, %r1153, %r564, %r1491; add.s32 %r1155, %r1154, %r562; mul.wide.s32 %rd1005, %r1155, 16; add.s64 %rd1006, %rd276, %rd1005; ld.global.v4.f32 {%f7131, %f7132, %f7133, %f7134}, [%rd1006]; add.s32 %r1156, %r1154, %r561; mul.wide.s32 %rd1007, %r1156, 16; add.s64 %rd1008, %rd276, %rd1007; ld.global.v4.f32 {%f7135, %f7136, %f7137, %f7138}, [%rd1008]; ld.global.v4.f32 {%f7139, %f7140, %f7141, %f7142}, [%rd1008+16]; add.s32 %r1157, %r1154, %r563; mul.wide.s32 %rd1009, %r1157, 16; add.s64 %rd1010, %rd276, %rd1009; ld.global.v4.f32 {%f7143, %f7144, %f7145, %f7146}, [%rd1010]; $L__BB24_487: cvt.rn.f32.s32 %f5186, %r566; sub.ftz.f32 %f5187, %f1403, %f5186; mov.f32 %f5188, 0f3F800000; sub.ftz.f32 %f5189, %f5188, %f5187; mul.ftz.f32 %f5190, %f5187, %f7131; fma.rn.ftz.f32 %f7147, %f5189, %f7115, %f5190; mul.ftz.f32 %f5191, %f5187, %f7132; fma.rn.ftz.f32 %f7148, %f5189, %f7116, %f5191; mul.ftz.f32 %f5192, %f5187, %f7133; fma.rn.ftz.f32 %f7149, %f5189, %f7117, %f5192; mul.ftz.f32 %f5193, %f5187, %f7134; fma.rn.ftz.f32 %f1535, %f5189, %f7118, %f5193; mul.ftz.f32 %f5194, %f5187, %f7135; fma.rn.ftz.f32 %f7151, %f5189, %f7119, %f5194; mul.ftz.f32 %f5195, %f5187, %f7136; fma.rn.ftz.f32 %f7152, %f5189, %f7120, %f5195; mul.ftz.f32 %f5196, %f5187, %f7137; fma.rn.ftz.f32 %f7153, %f5189, %f7121, %f5196; mul.ftz.f32 %f5197, %f5187, %f7138; fma.rn.ftz.f32 %f1539, %f5189, %f7122, %f5197; mul.ftz.f32 %f5198, %f5187, %f7139; fma.rn.ftz.f32 %f7155, %f5189, %f7123, %f5198; mul.ftz.f32 %f5199, %f5187, %f7140; fma.rn.ftz.f32 %f7156, %f5189, %f7124, %f5199; mul.ftz.f32 %f5200, %f5187, %f7141; fma.rn.ftz.f32 %f7157, %f5189, %f7125, %f5200; mul.ftz.f32 %f5201, %f5187, %f7142; fma.rn.ftz.f32 %f1543, %f5189, %f7126, %f5201; mul.ftz.f32 %f5202, %f5187, %f7143; fma.rn.ftz.f32 %f7159, %f5189, %f7127, %f5202; mul.ftz.f32 %f5203, %f5187, %f7144; fma.rn.ftz.f32 %f7160, %f5189, %f7128, %f5203; mul.ftz.f32 %f5204, %f5187, %f7145; fma.rn.ftz.f32 %f7161, %f5189, %f7129, %f5204; mul.ftz.f32 %f5205, %f5187, %f7146; fma.rn.ftz.f32 %f1547, %f5189, %f7130, %f5205; $L__BB24_489: and.b32 %r1158, %r560, 4; setp.eq.s32 %p640, %r1158, 0; @%p640 bra $L__BB24_505; bra.uni $L__BB24_490; $L__BB24_505: add.ftz.f32 %f5568, %f7147, %f7151; add.ftz.f32 %f5569, %f5568, %f7155; add.ftz.f32 %f5570, %f7148, %f7152; add.ftz.f32 %f5571, %f5570, %f7156; add.ftz.f32 %f5572, %f7149, %f7153; add.ftz.f32 %f5573, %f5572, %f7157; add.ftz.f32 %f5574, %f5569, %f7159; add.ftz.f32 %f5575, %f5571, %f7160; add.ftz.f32 %f5576, %f5573, %f7161; mul.ftz.f32 %f5577, %f5574, 0f3E800000; mul.ftz.f32 %f5578, %f5575, 0f3E800000; mul.ftz.f32 %f5579, %f5576, 0f3E800000; sub.ftz.f32 %f5580, %f5577, %f7292; sub.ftz.f32 %f5581, %f5578, %f7293; sub.ftz.f32 %f5582, %f5579, %f7294; mul.ftz.f32 %f5583, %f7290, %f5581; fma.rn.ftz.f32 %f5584, %f7289, %f5580, %f5583; fma.rn.ftz.f32 %f5585, %f7291, %f5582, %f5584; div.approx.ftz.f32 %f1640, %f5585, %f1364; fma.rn.ftz.f32 %f5586, %f7289, %f1640, %f7292; fma.rn.ftz.f32 %f5587, %f7290, %f1640, %f7293; fma.rn.ftz.f32 %f5588, %f7291, %f1640, %f7294; sub.ftz.f32 %f1641, %f7147, %f5586; sub.ftz.f32 %f1642, %f7148, %f5587; sub.ftz.f32 %f1643, %f7149, %f5588; sub.ftz.f32 %f1644, %f7151, %f5586; sub.ftz.f32 %f1645, %f7152, %f5587; sub.ftz.f32 %f1646, %f7153, %f5588; sub.ftz.f32 %f1647, %f7155, %f5586; sub.ftz.f32 %f1648, %f7156, %f5587; sub.ftz.f32 %f1649, %f7157, %f5588; sub.ftz.f32 %f1650, %f7159, %f5586; sub.ftz.f32 %f1651, %f7160, %f5587; sub.ftz.f32 %f1652, %f7161, %f5588; sub.ftz.f32 %f1653, %f1291, %f1640; mov.u32 %r1499, 0; mov.f32 %f7240, 0f3F800000; mov.f32 %f7239, 0f00000000; mov.f32 %f7238, %f1292; mov.u16 %rs237, %rs156; mov.u32 %r1495, %r1499; bra.uni $L__BB24_506; $L__BB24_548: add.s32 %r1499, %r1499, -1; mul.wide.s32 %rd1013, %r1499, 12; add.s64 %rd1014, %rd1, %rd1013; ld.local.f32 %f7239, [%rd1014]; ld.local.f32 %f7240, [%rd1014+4]; ld.local.u32 %r1495, [%rd1014+8]; $L__BB24_506: setp.gt.s32 %p662, %r1495, 1; @%p662 bra $L__BB24_546; $L__BB24_507: mov.u32 %r1500, %r1495; cvt.rn.f32.s32 %f5591, %r1500; mul.ftz.f32 %f1670, %f5591, 0f3F000000; sub.ftz.f32 %f1671, %f7240, %f7239; mul.ftz.f32 %f5592, %f1671, 0f3EAAAAAB; mul.ftz.f32 %f5593, %f5592, 0f3F000000; fma.rn.ftz.f32 %f1672, %f1671, %f1670, %f7239; add.ftz.f32 %f5594, %f1670, 0f3F000000; fma.rn.ftz.f32 %f1673, %f1671, %f5594, %f7239; mov.f32 %f5595, 0f3F800000; sub.ftz.f32 %f5596, %f5595, %f1672; mul.ftz.f32 %f5597, %f1672, %f5596; mul.ftz.f32 %f5598, %f5596, %f5597; mul.ftz.f32 %f5599, %f1672, %f1672; mul.ftz.f32 %f5600, %f1672, 0f40400000; add.ftz.f32 %f5601, %f5600, 0fC0A00000; fma.rn.ftz.f32 %f5602, %f5599, %f5601, 0f40000000; mul.ftz.f32 %f5603, %f5596, %f5596; mul.ftz.f32 %f5604, %f5596, 0f40400000; add.ftz.f32 %f5605, %f5604, 0fC0A00000; fma.rn.ftz.f32 %f5606, %f5603, %f5605, 0f40000000; mul.ftz.f32 %f5607, %f1672, %f5597; mul.ftz.f32 %f5608, %f1641, %f5598; mul.ftz.f32 %f5609, %f1642, %f5598; mul.ftz.f32 %f5610, %f1643, %f5598; mul.ftz.f32 %f5611, %f1535, %f5598; mul.ftz.f32 %f5612, %f1644, %f5602; mul.ftz.f32 %f5613, %f1645, %f5602; mul.ftz.f32 %f5614, %f1646, %f5602; mul.ftz.f32 %f5615, %f1539, %f5602; sub.ftz.f32 %f5616, %f5612, %f5608; sub.ftz.f32 %f5617, %f5613, %f5609; sub.ftz.f32 %f5618, %f5614, %f5610; sub.ftz.f32 %f5619, %f5615, %f5611; fma.rn.ftz.f32 %f5620, %f1647, %f5606, %f5616; fma.rn.ftz.f32 %f5621, %f1648, %f5606, %f5617; fma.rn.ftz.f32 %f5622, %f1649, %f5606, %f5618; fma.rn.ftz.f32 %f5623, %f1543, %f5606, %f5619; mul.ftz.f32 %f5624, %f1650, %f5607; mul.ftz.f32 %f5625, %f1651, %f5607; mul.ftz.f32 %f5626, %f1652, %f5607; mul.ftz.f32 %f5627, %f1547, %f5607; sub.ftz.f32 %f5628, %f5620, %f5624; sub.ftz.f32 %f5629, %f5621, %f5625; sub.ftz.f32 %f5630, %f5622, %f5626; sub.ftz.f32 %f5631, %f5623, %f5627; mul.ftz.f32 %f5632, %f5628, 0f3F000000; mul.ftz.f32 %f5633, %f5629, 0f3F000000; mul.ftz.f32 %f5634, %f5630, 0f3F000000; mul.ftz.f32 %f5635, %f5631, 0f3F000000; add.ftz.f32 %f5636, %f5596, %f5596; mul.ftz.f32 %f5637, %f1672, %f5636; sub.ftz.f32 %f5638, %f5637, %f5603; add.ftz.f32 %f5639, %f1672, %f1672; mul.ftz.f32 %f5640, %f5639, %f5601; fma.rn.ftz.f32 %f5641, %f1672, %f5600, %f5640; add.ftz.f32 %f5642, %f5600, 0f40000000; mul.ftz.f32 %f5643, %f5636, %f5642; mul.ftz.f32 %f5644, %f5596, %f5604; sub.ftz.f32 %f5645, %f5643, %f5644; mul.ftz.f32 %f5646, %f5596, 0fC0000000; fma.rn.ftz.f32 %f5647, %f1672, %f5646, %f5599; mul.ftz.f32 %f5648, %f1644, %f5641; mul.ftz.f32 %f5649, %f1645, %f5641; mul.ftz.f32 %f5650, %f1646, %f5641; mul.ftz.f32 %f5651, %f1539, %f5641; fma.rn.ftz.f32 %f5652, %f1641, %f5638, %f5648; fma.rn.ftz.f32 %f5653, %f1642, %f5638, %f5649; fma.rn.ftz.f32 %f5654, %f1643, %f5638, %f5650; fma.rn.ftz.f32 %f5655, %f1535, %f5638, %f5651; fma.rn.ftz.f32 %f5656, %f1647, %f5645, %f5652; fma.rn.ftz.f32 %f5657, %f1648, %f5645, %f5653; fma.rn.ftz.f32 %f5658, %f1649, %f5645, %f5654; fma.rn.ftz.f32 %f5659, %f1543, %f5645, %f5655; fma.rn.ftz.f32 %f5660, %f1650, %f5647, %f5656; fma.rn.ftz.f32 %f5661, %f1651, %f5647, %f5657; fma.rn.ftz.f32 %f5662, %f1652, %f5647, %f5658; fma.rn.ftz.f32 %f5663, %f1547, %f5647, %f5659; mul.ftz.f32 %f5664, %f5660, 0f3F000000; mul.ftz.f32 %f5665, %f5661, 0f3F000000; mul.ftz.f32 %f5666, %f5662, 0f3F000000; mul.ftz.f32 %f5667, %f5663, 0f3F000000; mul.ftz.f32 %f1674, %f5593, %f5664; mul.ftz.f32 %f1675, %f5593, %f5665; mul.ftz.f32 %f1676, %f5593, %f5666; sub.ftz.f32 %f5668, %f5595, %f1673; mul.ftz.f32 %f5669, %f1673, %f5668; mul.ftz.f32 %f5670, %f5668, %f5669; mul.ftz.f32 %f5671, %f1673, %f1673; mul.ftz.f32 %f5672, %f1673, 0f40400000; add.ftz.f32 %f5673, %f5672, 0fC0A00000; fma.rn.ftz.f32 %f5674, %f5671, %f5673, 0f40000000; mul.ftz.f32 %f5675, %f5668, %f5668; mul.ftz.f32 %f5676, %f5668, 0f40400000; add.ftz.f32 %f5677, %f5676, 0fC0A00000; fma.rn.ftz.f32 %f5678, %f5675, %f5677, 0f40000000; mul.ftz.f32 %f5679, %f1673, %f5669; mul.ftz.f32 %f5680, %f1641, %f5670; mul.ftz.f32 %f5681, %f1642, %f5670; mul.ftz.f32 %f5682, %f1643, %f5670; mul.ftz.f32 %f5683, %f1535, %f5670; mul.ftz.f32 %f5684, %f1644, %f5674; mul.ftz.f32 %f5685, %f1645, %f5674; mul.ftz.f32 %f5686, %f1646, %f5674; mul.ftz.f32 %f5687, %f1539, %f5674; sub.ftz.f32 %f5688, %f5684, %f5680; sub.ftz.f32 %f5689, %f5685, %f5681; sub.ftz.f32 %f5690, %f5686, %f5682; sub.ftz.f32 %f5691, %f5687, %f5683; fma.rn.ftz.f32 %f5692, %f1647, %f5678, %f5688; fma.rn.ftz.f32 %f5693, %f1648, %f5678, %f5689; fma.rn.ftz.f32 %f5694, %f1649, %f5678, %f5690; fma.rn.ftz.f32 %f5695, %f1543, %f5678, %f5691; mul.ftz.f32 %f5696, %f1650, %f5679; mul.ftz.f32 %f5697, %f1651, %f5679; mul.ftz.f32 %f5698, %f1652, %f5679; mul.ftz.f32 %f5699, %f1547, %f5679; sub.ftz.f32 %f5700, %f5692, %f5696; sub.ftz.f32 %f5701, %f5693, %f5697; sub.ftz.f32 %f5702, %f5694, %f5698; sub.ftz.f32 %f5703, %f5695, %f5699; mul.ftz.f32 %f1677, %f5700, 0f3F000000; mul.ftz.f32 %f1678, %f5701, 0f3F000000; mul.ftz.f32 %f1679, %f5702, 0f3F000000; mul.ftz.f32 %f5704, %f5703, 0f3F000000; add.ftz.f32 %f5705, %f5668, %f5668; mul.ftz.f32 %f5706, %f1673, %f5705; sub.ftz.f32 %f5707, %f5706, %f5675; add.ftz.f32 %f5708, %f1673, %f1673; mul.ftz.f32 %f5709, %f5708, %f5673; fma.rn.ftz.f32 %f5710, %f1673, %f5672, %f5709; add.ftz.f32 %f5711, %f5672, 0f40000000; mul.ftz.f32 %f5712, %f5705, %f5711; mul.ftz.f32 %f5713, %f5668, %f5676; sub.ftz.f32 %f5714, %f5712, %f5713; mul.ftz.f32 %f5715, %f5668, 0fC0000000; fma.rn.ftz.f32 %f5716, %f1673, %f5715, %f5671; mul.ftz.f32 %f5717, %f1644, %f5710; mul.ftz.f32 %f5718, %f1645, %f5710; mul.ftz.f32 %f5719, %f1646, %f5710; mul.ftz.f32 %f5720, %f1539, %f5710; fma.rn.ftz.f32 %f5721, %f1641, %f5707, %f5717; fma.rn.ftz.f32 %f5722, %f1642, %f5707, %f5718; fma.rn.ftz.f32 %f5723, %f1643, %f5707, %f5719; fma.rn.ftz.f32 %f5724, %f1535, %f5707, %f5720; fma.rn.ftz.f32 %f5725, %f1647, %f5714, %f5721; fma.rn.ftz.f32 %f5726, %f1648, %f5714, %f5722; fma.rn.ftz.f32 %f5727, %f1649, %f5714, %f5723; fma.rn.ftz.f32 %f5728, %f1543, %f5714, %f5724; fma.rn.ftz.f32 %f5729, %f1650, %f5716, %f5725; fma.rn.ftz.f32 %f5730, %f1651, %f5716, %f5726; fma.rn.ftz.f32 %f5731, %f1652, %f5716, %f5727; fma.rn.ftz.f32 %f5732, %f1547, %f5716, %f5728; mul.ftz.f32 %f5733, %f5729, 0f3F000000; mul.ftz.f32 %f5734, %f5730, 0f3F000000; mul.ftz.f32 %f5735, %f5731, 0f3F000000; mul.ftz.f32 %f5736, %f5732, 0f3F000000; mul.ftz.f32 %f1680, %f5593, %f5733; mul.ftz.f32 %f1681, %f5593, %f5734; mul.ftz.f32 %f1682, %f5593, %f5735; mul.ftz.f32 %f5737, %f5593, %f5736; fma.rn.ftz.f32 %f5738, %f5593, %f5667, %f5635; sub.ftz.f32 %f5739, %f5704, %f5737; sub.ftz.f32 %f5740, %f1677, %f5632; sub.ftz.f32 %f5741, %f1678, %f5633; sub.ftz.f32 %f5742, %f1679, %f5634; mul.ftz.f32 %f5743, %f1675, %f5742; mul.ftz.f32 %f5744, %f1676, %f5741; sub.ftz.f32 %f5745, %f5743, %f5744; mul.ftz.f32 %f5746, %f1676, %f5740; mul.ftz.f32 %f5747, %f1674, %f5742; sub.ftz.f32 %f5748, %f5746, %f5747; mul.ftz.f32 %f5749, %f1674, %f5741; mul.ftz.f32 %f5750, %f1675, %f5740; sub.ftz.f32 %f5751, %f5749, %f5750; mul.ftz.f32 %f5752, %f5748, %f5748; fma.rn.ftz.f32 %f5753, %f5745, %f5745, %f5752; fma.rn.ftz.f32 %f5754, %f5751, %f5751, %f5753; mul.ftz.f32 %f5755, %f5741, %f5741; fma.rn.ftz.f32 %f5756, %f5740, %f5740, %f5755; fma.rn.ftz.f32 %f5757, %f5742, %f5742, %f5756; div.approx.ftz.f32 %f5758, %f5754, %f5757; mul.ftz.f32 %f5759, %f1681, %f5742; mul.ftz.f32 %f5760, %f1682, %f5741; sub.ftz.f32 %f5761, %f5759, %f5760; mul.ftz.f32 %f5762, %f1682, %f5740; mul.ftz.f32 %f5763, %f1680, %f5742; sub.ftz.f32 %f5764, %f5762, %f5763; mul.ftz.f32 %f5765, %f1680, %f5741; mul.ftz.f32 %f5766, %f1681, %f5740; sub.ftz.f32 %f5767, %f5765, %f5766; mul.ftz.f32 %f5768, %f5764, %f5764; fma.rn.ftz.f32 %f5769, %f5761, %f5761, %f5768; fma.rn.ftz.f32 %f5770, %f5767, %f5767, %f5769; div.approx.ftz.f32 %f5771, %f5770, %f5757; max.ftz.f32 %f5772, %f5758, %f5771; sqrt.approx.ftz.f32 %f5773, %f5772; max.ftz.f32 %f5774, %f5635, %f5738; max.ftz.f32 %f5775, %f5739, %f5704; max.ftz.f32 %f5776, %f5774, %f5775; add.ftz.f32 %f5777, %f5773, %f5776; min.ftz.f32 %f5778, %f5635, %f5738; min.ftz.f32 %f5779, %f5739, %f5704; min.ftz.f32 %f5780, %f5778, %f5779; sub.ftz.f32 %f1683, %f5780, %f5773; mul.ftz.f32 %f5781, %f5777, 0f3F800002; rsqrt.approx.ftz.f32 %f1684, %f5757; mul.ftz.f32 %f5782, %f5740, %f1684; mul.ftz.f32 %f5783, %f5741, %f1684; mul.ftz.f32 %f5784, %f5742, %f1684; mul.ftz.f32 %f1685, %f5628, 0fBF000000; mul.ftz.f32 %f1686, %f5629, 0fBF000000; mul.ftz.f32 %f1687, %f5630, 0fBF000000; mul.ftz.f32 %f5785, %f7289, %f1685; mul.ftz.f32 %f5786, %f7290, %f5633; sub.ftz.f32 %f5787, %f5785, %f5786; mul.ftz.f32 %f5788, %f7291, %f5634; sub.ftz.f32 %f5789, %f5787, %f5788; mul.ftz.f32 %f5790, %f5633, %f5633; fma.rn.ftz.f32 %f5791, %f5632, %f5632, %f5790; fma.rn.ftz.f32 %f5792, %f5634, %f5634, %f5791; mul.ftz.f32 %f5793, %f7290, %f5783; fma.rn.ftz.f32 %f5794, %f7289, %f5782, %f5793; fma.rn.ftz.f32 %f1688, %f7291, %f5784, %f5794; mul.ftz.f32 %f5795, %f5782, %f1685; mul.ftz.f32 %f5796, %f5783, %f5633; sub.ftz.f32 %f5797, %f5795, %f5796; mul.ftz.f32 %f5798, %f5784, %f5634; sub.ftz.f32 %f1689, %f5797, %f5798; mul.ftz.f32 %f1690, %f1688, %f1688; sub.ftz.f32 %f1691, %f1364, %f1690; mul.ftz.f32 %f5799, %f1688, %f1689; sub.ftz.f32 %f5800, %f5789, %f5799; add.ftz.f32 %f1692, %f5800, %f5800; mul.ftz.f32 %f5801, %f1689, %f1689; sub.ftz.f32 %f1693, %f5792, %f5801; mul.ftz.f32 %f5802, %f5781, %f5781; sub.ftz.f32 %f1694, %f1693, %f5802; mul.ftz.f32 %f1695, %f1692, %f1692; mul.ftz.f32 %f1696, %f1691, 0f40800000; mul.ftz.f32 %f5803, %f1696, %f1694; sub.ftz.f32 %f1697, %f1695, %f5803; setp.ltu.ftz.f32 %p664, %f1697, 0f00000000; mov.pred %p804, -1; mov.f32 %f7208, 0fFF7FFFFF; mov.f32 %f7207, 0f7F7FFFFF; @%p664 bra $L__BB24_511; abs.ftz.f32 %f5804, %f1364; abs.ftz.f32 %f5805, %f1690; max.ftz.f32 %f5806, %f5804, %f5805; mul.ftz.f32 %f5807, %f5806, 0f36000000; abs.ftz.f32 %f5808, %f1691; setp.lt.ftz.f32 %p665, %f5808, %f5807; @%p665 bra $L__BB24_510; bra.uni $L__BB24_509; $L__BB24_510: setp.gtu.ftz.f32 %p804, %f1694, 0f00000000; mov.f32 %f7208, 0f7F7FFFFF; mov.f32 %f7207, 0fFF7FFFFF; bra.uni $L__BB24_511; $L__BB24_509: sqrt.approx.ftz.f32 %f5809, %f1697; add.ftz.f32 %f5810, %f1691, %f1691; rcp.approx.ftz.f32 %f5811, %f5810; neg.ftz.f32 %f5812, %f1692; sub.ftz.f32 %f5813, %f5812, %f5809; mul.ftz.f32 %f7207, %f5811, %f5813; sub.ftz.f32 %f5814, %f5809, %f1692; mul.ftz.f32 %f7208, %f5811, %f5814; fma.rn.ftz.f32 %f5815, %f1688, %f7207, %f1689; mul.ftz.f32 %f7237, %f1684, %f5815; fma.rn.ftz.f32 %f5816, %f1688, %f7208, %f1689; mul.ftz.f32 %f7236, %f1684, %f5816; mov.pred %p804, 0; $L__BB24_511: @%p804 bra $L__BB24_545; sub.ftz.f32 %f5819, %f7238, %f1640; min.ftz.f32 %f5820, %f5819, %f7208; mul.ftz.f32 %f5821, %f1675, %f1686; fma.rn.ftz.f32 %f5822, %f1674, %f1685, %f5821; fma.rn.ftz.f32 %f5823, %f1676, %f1687, %f5822; mul.ftz.f32 %f5824, %f7290, %f1675; fma.rn.ftz.f32 %f5825, %f7289, %f1674, %f5824; fma.rn.ftz.f32 %f5826, %f7291, %f1676, %f5825; abs.ftz.f32 %f5827, %f5826; setp.lt.ftz.f32 %p667, %f5827, 0f219392EF; neg.ftz.f32 %f5828, %f5823; div.approx.ftz.f32 %f5829, %f5828, %f5826; setp.lt.ftz.f32 %p668, %f5826, 0f00000000; or.pred %p669, %p668, %p667; selp.f32 %f5830, 0fFF7FFFFF, %f5829, %p669; setp.gt.ftz.f32 %p670, %f5826, 0f00000000; or.pred %p671, %p670, %p667; selp.f32 %f5831, 0f7F7FFFFF, %f5829, %p671; max.ftz.f32 %f5832, %f1653, %f7207; max.ftz.f32 %f5833, %f5832, %f5830; min.ftz.f32 %f5834, %f5820, %f5831; mul.ftz.f32 %f5835, %f1678, %f1681; fma.rn.ftz.f32 %f5836, %f1677, %f1680, %f5835; fma.rn.ftz.f32 %f5837, %f1679, %f1682, %f5836; mul.ftz.f32 %f5838, %f7290, %f1681; neg.ftz.f32 %f5839, %f5838; mul.ftz.f32 %f5840, %f7289, %f1680; sub.ftz.f32 %f5841, %f5839, %f5840; mul.ftz.f32 %f5842, %f7291, %f1682; sub.ftz.f32 %f5843, %f5841, %f5842; abs.ftz.f32 %f5844, %f5843; setp.lt.ftz.f32 %p672, %f5844, 0f219392EF; neg.ftz.f32 %f5845, %f5837; div.approx.ftz.f32 %f5846, %f5845, %f5843; setp.lt.ftz.f32 %p673, %f5843, 0f00000000; or.pred %p674, %p673, %p672; selp.f32 %f5847, 0fFF7FFFFF, %f5846, %p674; setp.gt.ftz.f32 %p675, %f5843, 0f00000000; or.pred %p676, %p675, %p672; selp.f32 %f5848, 0f7F7FFFFF, %f5846, %p676; max.ftz.f32 %f7211, %f5833, %f5847; min.ftz.f32 %f7222, %f5834, %f5848; setp.gtu.ftz.f32 %p677, %f7211, %f7222; @%p677 bra $L__BB24_545; mul.ftz.f32 %f5851, %f1683, 0f3F7FFFFC; mov.f32 %f5852, 0f00000000; max.ftz.f32 %f5853, %f5852, %f5851; max.ftz.f32 %f5854, %f7237, %f5852; min.ftz.f32 %f5856, %f5854, %f5595; max.ftz.f32 %f5857, %f7236, %f5852; min.ftz.f32 %f5858, %f5857, %f5595; add.ftz.f32 %f5859, %f1670, %f5856; mul.ftz.f32 %f5860, %f5859, 0f3EAAAAAB; fma.rn.ftz.f32 %f7237, %f1671, %f5860, %f7239; add.ftz.f32 %f5861, %f1670, %f5858; mul.ftz.f32 %f5862, %f5861, 0f3EAAAAAB; fma.rn.ftz.f32 %f7236, %f1671, %f5862, %f7239; mul.ftz.f32 %f5863, %f5853, %f5853; sub.ftz.f32 %f5864, %f1693, %f5863; mul.ftz.f32 %f5865, %f1696, %f5864; sub.ftz.f32 %f1710, %f1695, %f5865; setp.ltu.ftz.f32 %p678, %f1710, 0f00000000; mov.f32 %f7210, 0fFF7FFFFF; mov.f32 %f7209, 0f7F7FFFFF; @%p678 bra $L__BB24_516; abs.ftz.f32 %f5868, %f1364; abs.ftz.f32 %f5869, %f1690; max.ftz.f32 %f5870, %f5868, %f5869; mul.ftz.f32 %f5871, %f5870, 0f36000000; abs.ftz.f32 %f5872, %f1691; setp.lt.ftz.f32 %p679, %f5872, %f5871; mov.f32 %f7210, 0f7F7FFFFF; mov.f32 %f7209, 0fFF7FFFFF; @%p679 bra $L__BB24_516; sqrt.approx.ftz.f32 %f5873, %f1710; add.ftz.f32 %f5874, %f1691, %f1691; rcp.approx.ftz.f32 %f5875, %f5874; neg.ftz.f32 %f5876, %f1692; sub.ftz.f32 %f5877, %f5876, %f5873; mul.ftz.f32 %f7209, %f5875, %f5877; sub.ftz.f32 %f5878, %f5873, %f1692; mul.ftz.f32 %f7210, %f5875, %f5878; $L__BB24_516: min.ftz.f32 %f1715, %f7222, %f7209; setp.gtu.ftz.f32 %p680, %f7211, %f1715; max.ftz.f32 %f1716, %f7211, %f7210; setp.gtu.ftz.f32 %p681, %f1716, %f7222; and.pred %p682, %p680, %p681; @%p682 bra $L__BB24_545; mov.u16 %rs157, 0; mov.u16 %rs233, %rs157; @%p680 bra $L__BB24_530; setp.lt.s32 %p684, %r1499, 4; mov.u16 %rs233, 1; @%p684 bra $L__BB24_530; sqrt.approx.ftz.f32 %f5879, %f1364; min.ftz.f32 %f5880, %f1647, %f1650; min.ftz.f32 %f5881, %f1641, %f1644; min.ftz.f32 %f5882, %f5881, %f5880; min.ftz.f32 %f5883, %f1648, %f1651; min.ftz.f32 %f5884, %f1642, %f1645; min.ftz.f32 %f5885, %f5884, %f5883; min.ftz.f32 %f5886, %f1649, %f1652; min.ftz.f32 %f5887, %f1643, %f1646; min.ftz.f32 %f5888, %f5887, %f5886; min.ftz.f32 %f5889, %f1543, %f1547; min.ftz.f32 %f5890, %f1535, %f1539; min.ftz.f32 %f5891, %f5890, %f5889; max.ftz.f32 %f5892, %f1647, %f1650; max.ftz.f32 %f5893, %f5881, %f5892; max.ftz.f32 %f5894, %f1648, %f1651; max.ftz.f32 %f5895, %f5884, %f5894; max.ftz.f32 %f5896, %f1649, %f1652; max.ftz.f32 %f5897, %f5887, %f5896; max.ftz.f32 %f5898, %f1543, %f1547; max.ftz.f32 %f5899, %f5890, %f5898; abs.ftz.f32 %f5900, %f5882; abs.ftz.f32 %f5901, %f5885; abs.ftz.f32 %f5902, %f5888; abs.ftz.f32 %f5903, %f5891; abs.ftz.f32 %f5904, %f5893; abs.ftz.f32 %f5905, %f5895; abs.ftz.f32 %f5906, %f5897; abs.ftz.f32 %f5907, %f5899; max.ftz.f32 %f5908, %f5900, %f5904; max.ftz.f32 %f5909, %f5901, %f5905; max.ftz.f32 %f5910, %f5902, %f5906; max.ftz.f32 %f5911, %f5903, %f5907; max.ftz.f32 %f5912, %f5910, %f5911; max.ftz.f32 %f5913, %f5909, %f5912; max.ftz.f32 %f5914, %f5908, %f5913; mul.ftz.f32 %f1717, %f5914, 0f36000000; mul.ftz.f32 %f1718, %f5879, 0f36000000; mul.ftz.f32 %f1719, %f5899, 0f36000000; mov.u32 %r1497, 0; mov.f32 %f7212, %f7237; $L__BB24_520: sub.ftz.f32 %f5917, %f5595, %f7212; mul.ftz.f32 %f5918, %f7212, %f5917; mul.ftz.f32 %f5919, %f5917, %f5918; mul.ftz.f32 %f5920, %f7212, 0f40400000; add.ftz.f32 %f5921, %f5920, 0fC0A00000; mul.ftz.f32 %f5922, %f7212, %f7212; fma.rn.ftz.f32 %f5923, %f5922, %f5921, 0f40000000; mov.f32 %f5924, 0f40000000; mul.ftz.f32 %f5925, %f5917, %f5917; mul.ftz.f32 %f5926, %f5917, 0f40400000; add.ftz.f32 %f5927, %f5926, 0fC0A00000; fma.rn.ftz.f32 %f5928, %f5925, %f5927, 0f40000000; mul.ftz.f32 %f5929, %f7212, %f5918; mul.ftz.f32 %f5930, %f1641, %f5919; mul.ftz.f32 %f5931, %f1642, %f5919; mul.ftz.f32 %f5932, %f1643, %f5919; mul.ftz.f32 %f5933, %f1535, %f5919; mul.ftz.f32 %f5934, %f1644, %f5923; mul.ftz.f32 %f5935, %f1645, %f5923; mul.ftz.f32 %f5936, %f1646, %f5923; mul.ftz.f32 %f5937, %f1539, %f5923; sub.ftz.f32 %f5938, %f5934, %f5930; sub.ftz.f32 %f5939, %f5935, %f5931; sub.ftz.f32 %f5940, %f5936, %f5932; sub.ftz.f32 %f5941, %f5937, %f5933; fma.rn.ftz.f32 %f5942, %f1647, %f5928, %f5938; fma.rn.ftz.f32 %f5943, %f1648, %f5928, %f5939; fma.rn.ftz.f32 %f5944, %f1649, %f5928, %f5940; fma.rn.ftz.f32 %f5945, %f1543, %f5928, %f5941; mul.ftz.f32 %f5946, %f1650, %f5929; mul.ftz.f32 %f5947, %f1651, %f5929; mul.ftz.f32 %f5948, %f1652, %f5929; mul.ftz.f32 %f5949, %f1547, %f5929; sub.ftz.f32 %f5950, %f5942, %f5946; sub.ftz.f32 %f5951, %f5943, %f5947; sub.ftz.f32 %f5952, %f5944, %f5948; sub.ftz.f32 %f5953, %f5945, %f5949; mul.ftz.f32 %f5954, %f5950, 0f3F000000; mul.ftz.f32 %f5955, %f5951, 0f3F000000; mul.ftz.f32 %f5956, %f5952, 0f3F000000; mul.ftz.f32 %f1722, %f5953, 0f3F000000; add.ftz.f32 %f5957, %f5917, %f5917; mul.ftz.f32 %f5958, %f7212, %f5957; sub.ftz.f32 %f5959, %f5958, %f5925; add.ftz.f32 %f5960, %f7212, %f7212; mul.ftz.f32 %f5961, %f5960, %f5921; fma.rn.ftz.f32 %f5962, %f7212, %f5920, %f5961; add.ftz.f32 %f5963, %f5920, 0f40000000; mul.ftz.f32 %f5964, %f5957, %f5963; mul.ftz.f32 %f5965, %f5917, %f5926; sub.ftz.f32 %f5966, %f5964, %f5965; mul.ftz.f32 %f5967, %f5917, 0fC0000000; fma.rn.ftz.f32 %f5968, %f7212, %f5967, %f5922; mul.ftz.f32 %f5969, %f1644, %f5962; mul.ftz.f32 %f5970, %f1645, %f5962; mul.ftz.f32 %f5971, %f1646, %f5962; mul.ftz.f32 %f5972, %f1539, %f5962; fma.rn.ftz.f32 %f5973, %f1641, %f5959, %f5969; fma.rn.ftz.f32 %f5974, %f1642, %f5959, %f5970; fma.rn.ftz.f32 %f5975, %f1643, %f5959, %f5971; fma.rn.ftz.f32 %f5976, %f1535, %f5959, %f5972; fma.rn.ftz.f32 %f5977, %f1647, %f5966, %f5973; fma.rn.ftz.f32 %f5978, %f1648, %f5966, %f5974; fma.rn.ftz.f32 %f5979, %f1649, %f5966, %f5975; fma.rn.ftz.f32 %f5980, %f1543, %f5966, %f5976; fma.rn.ftz.f32 %f5981, %f1650, %f5968, %f5977; fma.rn.ftz.f32 %f5982, %f1651, %f5968, %f5978; fma.rn.ftz.f32 %f5983, %f1652, %f5968, %f5979; fma.rn.ftz.f32 %f5984, %f1547, %f5968, %f5980; mul.ftz.f32 %f1723, %f5981, 0f3F000000; mul.ftz.f32 %f1724, %f5982, 0f3F000000; mul.ftz.f32 %f1725, %f5983, 0f3F000000; mul.ftz.f32 %f1726, %f5984, 0f3F000000; sub.ftz.f32 %f5985, %f5924, %f5920; mul.ftz.f32 %f5986, %f7212, 0f41100000; add.ftz.f32 %f5987, %f5986, 0fC0A00000; mov.f32 %f5988, 0f40800000; sub.ftz.f32 %f5989, %f5988, %f5986; add.ftz.f32 %f5990, %f5920, 0fBF800000; mul.ftz.f32 %f5991, %f1644, %f5987; mul.ftz.f32 %f5992, %f1645, %f5987; mul.ftz.f32 %f5993, %f1646, %f5987; fma.rn.ftz.f32 %f5994, %f1641, %f5985, %f5991; fma.rn.ftz.f32 %f5995, %f1642, %f5985, %f5992; fma.rn.ftz.f32 %f5996, %f1643, %f5985, %f5993; fma.rn.ftz.f32 %f5997, %f1647, %f5989, %f5994; fma.rn.ftz.f32 %f5998, %f1648, %f5989, %f5995; fma.rn.ftz.f32 %f5999, %f1649, %f5989, %f5996; fma.rn.ftz.f32 %f6000, %f1650, %f5990, %f5997; fma.rn.ftz.f32 %f6001, %f1651, %f5990, %f5998; fma.rn.ftz.f32 %f6002, %f1652, %f5990, %f5999; mul.ftz.f32 %f6003, %f7289, %f7211; sub.ftz.f32 %f1727, %f6003, %f5954; mul.ftz.f32 %f6004, %f7290, %f7211; sub.ftz.f32 %f1728, %f6004, %f5955; mul.ftz.f32 %f6005, %f7291, %f7211; sub.ftz.f32 %f1729, %f6005, %f5956; mul.ftz.f32 %f6006, %f1728, %f1728; fma.rn.ftz.f32 %f6007, %f1727, %f1727, %f6006; fma.rn.ftz.f32 %f6008, %f1729, %f1729, %f6007; sqrt.approx.ftz.f32 %f1730, %f6008; mul.ftz.f32 %f6009, %f1718, %f7211; max.ftz.f32 %f1731, %f6009, %f1717; mul.ftz.f32 %f6010, %f5981, 0fBF000000; mul.ftz.f32 %f6011, %f1724, %f1724; fma.rn.ftz.f32 %f6012, %f1723, %f1723, %f6011; fma.rn.ftz.f32 %f6013, %f1725, %f1725, %f6012; sqrt.approx.ftz.f32 %f6014, %f6013; div.approx.ftz.f32 %f6015, %f1723, %f6014; div.approx.ftz.f32 %f6016, %f1724, %f6014; div.approx.ftz.f32 %f6017, %f1725, %f6014; mul.ftz.f32 %f6018, %f6001, %f1724; fma.rn.ftz.f32 %f6019, %f6000, %f1723, %f6018; fma.rn.ftz.f32 %f6020, %f6002, %f1725, %f6019; mul.ftz.f32 %f6021, %f6000, %f6013; mul.ftz.f32 %f6022, %f6001, %f6013; mul.ftz.f32 %f6023, %f6002, %f6013; mul.ftz.f32 %f6024, %f1723, %f6020; mul.ftz.f32 %f6025, %f1724, %f6020; mul.ftz.f32 %f6026, %f1725, %f6020; sub.ftz.f32 %f6027, %f6021, %f6024; sub.ftz.f32 %f6028, %f6022, %f6025; sub.ftz.f32 %f6029, %f6023, %f6026; mul.ftz.f32 %f6030, %f6013, %f6014; div.approx.ftz.f32 %f6031, %f6027, %f6030; div.approx.ftz.f32 %f6032, %f6028, %f6030; div.approx.ftz.f32 %f6033, %f6029, %f6030; div.approx.ftz.f32 %f6034, %f1717, %f6014; mul.ftz.f32 %f6035, %f1728, %f6016; fma.rn.ftz.f32 %f6036, %f1727, %f6015, %f6035; fma.rn.ftz.f32 %f1732, %f1729, %f6017, %f6036; fma.rn.ftz.f32 %f6037, %f1717, %f1730, %f1731; add.ftz.f32 %f6038, %f1730, 0f3F800000; fma.rn.ftz.f32 %f1733, %f6038, %f6034, %f6037; mul.ftz.f32 %f6039, %f6015, %f6010; mul.ftz.f32 %f6040, %f6016, %f1724; sub.ftz.f32 %f6041, %f6039, %f6040; mul.ftz.f32 %f6042, %f6017, %f1725; sub.ftz.f32 %f6043, %f6041, %f6042; mul.ftz.f32 %f6044, %f1728, %f6032; fma.rn.ftz.f32 %f6045, %f1727, %f6031, %f6044; fma.rn.ftz.f32 %f6046, %f1729, %f6033, %f6045; add.ftz.f32 %f1734, %f6043, %f6046; mul.ftz.f32 %f6047, %f7290, %f6016; fma.rn.ftz.f32 %f6048, %f7289, %f6015, %f6047; fma.rn.ftz.f32 %f1735, %f7291, %f6017, %f6048; mul.ftz.f32 %f6049, %f1732, %f1732; sub.ftz.f32 %f6050, %f6008, %f6049; mul.ftz.f32 %f6051, %f1727, %f6010; mul.ftz.f32 %f6052, %f1728, %f1724; sub.ftz.f32 %f6053, %f6051, %f6052; mul.ftz.f32 %f6054, %f1729, %f1725; sub.ftz.f32 %f1736, %f6053, %f6054; mul.ftz.f32 %f6055, %f7290, %f1728; fma.rn.ftz.f32 %f6056, %f7289, %f1727, %f6055; fma.rn.ftz.f32 %f1737, %f7291, %f1729, %f6056; setp.leu.ftz.f32 %p685, %f6050, 0f00000000; mov.f32 %f7213, 0f00000000; sqrt.approx.ftz.f32 %f1738, %f6050; @%p685 bra $L__BB24_522; rcp.approx.ftz.f32 %f7213, %f1738; $L__BB24_522: add.ftz.f32 %f6057, %f1731, %f1733; add.ftz.f32 %f1741, %f1719, %f6057; mul.ftz.f32 %f6058, %f1732, %f1734; sub.ftz.f32 %f6059, %f1736, %f6058; mul.ftz.f32 %f6060, %f6059, %f7213; sub.ftz.f32 %f6061, %f6060, %f1726; mul.ftz.f32 %f6062, %f1732, %f1735; sub.ftz.f32 %f6063, %f1737, %f6062; mul.ftz.f32 %f6064, %f6063, %f7213; mul.ftz.f32 %f6065, %f1734, %f6064; mul.ftz.f32 %f6066, %f1735, %f6061; sub.ftz.f32 %f6067, %f6065, %f6066; rcp.approx.ftz.f32 %f6068, %f6067; mul.ftz.f32 %f6069, %f1732, %f6064; sub.ftz.f32 %f1742, %f1738, %f1722; mul.ftz.f32 %f6070, %f1735, %f1742; sub.ftz.f32 %f6071, %f6069, %f6070; mul.ftz.f32 %f6072, %f6071, %f6068; sub.ftz.f32 %f7212, %f7212, %f6072; mul.ftz.f32 %f6073, %f1732, %f6061; mul.ftz.f32 %f6074, %f1734, %f1742; sub.ftz.f32 %f6075, %f6074, %f6073; mul.ftz.f32 %f6076, %f6075, %f6068; sub.ftz.f32 %f7211, %f7211, %f6076; abs.ftz.f32 %f6077, %f1732; setp.geu.ftz.f32 %p686, %f6077, %f1733; @%p686 bra $L__BB24_524; abs.ftz.f32 %f6078, %f1742; setp.lt.ftz.f32 %p687, %f6078, %f1741; @%p687 bra $L__BB24_525; $L__BB24_524: add.s32 %r1497, %r1497, 1; setp.lt.u32 %p688, %r1497, 5; mov.u16 %rs231, 0; @%p688 bra $L__BB24_520; bra.uni $L__BB24_529; $L__BB24_525: add.ftz.f32 %f1745, %f1640, %f7211; setp.ltu.ftz.f32 %p689, %f1745, %f1291; setp.gtu.ftz.f32 %p690, %f1745, %f7238; or.pred %p691, %p689, %p690; mov.u16 %rs231, 0; @%p691 bra $L__BB24_529; setp.ltu.ftz.f32 %p692, %f7212, 0f00000000; setp.gtu.ftz.f32 %p693, %f7212, 0f3F800000; or.pred %p694, %p693, %p692; @%p694 bra $L__BB24_529; div.approx.ftz.f32 %f6079, %f1727, %f1730; div.approx.ftz.f32 %f6080, %f1728, %f1730; div.approx.ftz.f32 %f6081, %f1729, %f1730; fma.rn.ftz.f32 %f6082, %f1726, %f6079, %f1723; fma.rn.ftz.f32 %f6083, %f1726, %f6080, %f1724; fma.rn.ftz.f32 %f6084, %f1726, %f6081, %f1725; mul.ftz.f32 %f6085, %f1724, %f6081; mul.ftz.f32 %f6086, %f1725, %f6080; sub.ftz.f32 %f6087, %f6085, %f6086; mul.ftz.f32 %f6088, %f1725, %f6079; mul.ftz.f32 %f6089, %f1723, %f6081; sub.ftz.f32 %f6090, %f6088, %f6089; mul.ftz.f32 %f6091, %f1723, %f6080; mul.ftz.f32 %f6092, %f1724, %f6079; sub.ftz.f32 %f6093, %f6091, %f6092; mul.ftz.f32 %f6094, %f6090, %f6084; mul.ftz.f32 %f6095, %f6093, %f6083; sub.ftz.f32 %f6096, %f6094, %f6095; mul.ftz.f32 %f6097, %f6082, %f6093; mul.ftz.f32 %f6098, %f6087, %f6084; sub.ftz.f32 %f6099, %f6097, %f6098; mul.ftz.f32 %f6100, %f6083, %f6087; mul.ftz.f32 %f6101, %f6082, %f6090; sub.ftz.f32 %f6102, %f6100, %f6101; mul.ftz.f32 %f6103, %f7290, %f6099; fma.rn.ftz.f32 %f6104, %f7289, %f6096, %f6103; fma.rn.ftz.f32 %f6105, %f7291, %f6102, %f6104; setp.gt.ftz.f32 %p695, %f6105, 0f00000000; @%p695 bra $L__BB24_529; mov.u16 %rs231, 1; mov.f32 %f7253, 0f00000000; mov.f32 %f7251, %f1745; mov.f32 %f7252, %f7212; mov.f32 %f7238, %f1745; $L__BB24_529: or.b16 %rs165, %rs231, %rs237; and.b16 %rs166, %rs165, 255; setp.ne.s16 %p696, %rs166, 0; selp.u16 %rs237, 1, 0, %p696; mov.u16 %rs233, %rs157; $L__BB24_530: add.ftz.f32 %f6107, %f1640, %f1716; setp.ltu.ftz.f32 %p697, %f6107, %f1291; setp.gtu.ftz.f32 %p698, %f6107, %f7238; or.pred %p699, %p697, %p698; or.pred %p701, %p681, %p699; mov.u16 %rs236, %rs233; @%p701 bra $L__BB24_543; setp.lt.s32 %p702, %r1499, 4; mov.u16 %rs236, 1; @%p702 bra $L__BB24_543; sqrt.approx.ftz.f32 %f6108, %f1364; min.ftz.f32 %f6109, %f1647, %f1650; min.ftz.f32 %f6110, %f1641, %f1644; min.ftz.f32 %f6111, %f6110, %f6109; min.ftz.f32 %f6112, %f1648, %f1651; min.ftz.f32 %f6113, %f1642, %f1645; min.ftz.f32 %f6114, %f6113, %f6112; min.ftz.f32 %f6115, %f1649, %f1652; min.ftz.f32 %f6116, %f1643, %f1646; min.ftz.f32 %f6117, %f6116, %f6115; min.ftz.f32 %f6118, %f1543, %f1547; min.ftz.f32 %f6119, %f1535, %f1539; min.ftz.f32 %f6120, %f6119, %f6118; max.ftz.f32 %f6121, %f1647, %f1650; max.ftz.f32 %f6122, %f6110, %f6121; max.ftz.f32 %f6123, %f1648, %f1651; max.ftz.f32 %f6124, %f6113, %f6123; max.ftz.f32 %f6125, %f1649, %f1652; max.ftz.f32 %f6126, %f6116, %f6125; max.ftz.f32 %f6127, %f1543, %f1547; max.ftz.f32 %f6128, %f6119, %f6127; abs.ftz.f32 %f6129, %f6111; abs.ftz.f32 %f6130, %f6114; abs.ftz.f32 %f6131, %f6117; abs.ftz.f32 %f6132, %f6120; abs.ftz.f32 %f6133, %f6122; abs.ftz.f32 %f6134, %f6124; abs.ftz.f32 %f6135, %f6126; abs.ftz.f32 %f6136, %f6128; max.ftz.f32 %f6137, %f6129, %f6133; max.ftz.f32 %f6138, %f6130, %f6134; max.ftz.f32 %f6139, %f6131, %f6135; max.ftz.f32 %f6140, %f6132, %f6136; max.ftz.f32 %f6141, %f6139, %f6140; max.ftz.f32 %f6142, %f6138, %f6141; max.ftz.f32 %f6143, %f6137, %f6142; mul.ftz.f32 %f1754, %f6143, 0f36000000; mul.ftz.f32 %f1755, %f6108, 0f36000000; mul.ftz.f32 %f1756, %f6128, 0f36000000; mov.u32 %r1498, 0; mov.f32 %f7223, %f7236; $L__BB24_533: sub.ftz.f32 %f6146, %f5595, %f7223; mul.ftz.f32 %f6147, %f7223, %f6146; mul.ftz.f32 %f6148, %f6146, %f6147; mul.ftz.f32 %f6149, %f7223, 0f40400000; add.ftz.f32 %f6150, %f6149, 0fC0A00000; mul.ftz.f32 %f6151, %f7223, %f7223; fma.rn.ftz.f32 %f6152, %f6151, %f6150, 0f40000000; mov.f32 %f6153, 0f40000000; mul.ftz.f32 %f6154, %f6146, %f6146; mul.ftz.f32 %f6155, %f6146, 0f40400000; add.ftz.f32 %f6156, %f6155, 0fC0A00000; fma.rn.ftz.f32 %f6157, %f6154, %f6156, 0f40000000; mul.ftz.f32 %f6158, %f7223, %f6147; mul.ftz.f32 %f6159, %f1641, %f6148; mul.ftz.f32 %f6160, %f1642, %f6148; mul.ftz.f32 %f6161, %f1643, %f6148; mul.ftz.f32 %f6162, %f1535, %f6148; mul.ftz.f32 %f6163, %f1644, %f6152; mul.ftz.f32 %f6164, %f1645, %f6152; mul.ftz.f32 %f6165, %f1646, %f6152; mul.ftz.f32 %f6166, %f1539, %f6152; sub.ftz.f32 %f6167, %f6163, %f6159; sub.ftz.f32 %f6168, %f6164, %f6160; sub.ftz.f32 %f6169, %f6165, %f6161; sub.ftz.f32 %f6170, %f6166, %f6162; fma.rn.ftz.f32 %f6171, %f1647, %f6157, %f6167; fma.rn.ftz.f32 %f6172, %f1648, %f6157, %f6168; fma.rn.ftz.f32 %f6173, %f1649, %f6157, %f6169; fma.rn.ftz.f32 %f6174, %f1543, %f6157, %f6170; mul.ftz.f32 %f6175, %f1650, %f6158; mul.ftz.f32 %f6176, %f1651, %f6158; mul.ftz.f32 %f6177, %f1652, %f6158; mul.ftz.f32 %f6178, %f1547, %f6158; sub.ftz.f32 %f6179, %f6171, %f6175; sub.ftz.f32 %f6180, %f6172, %f6176; sub.ftz.f32 %f6181, %f6173, %f6177; sub.ftz.f32 %f6182, %f6174, %f6178; mul.ftz.f32 %f6183, %f6179, 0f3F000000; mul.ftz.f32 %f6184, %f6180, 0f3F000000; mul.ftz.f32 %f6185, %f6181, 0f3F000000; mul.ftz.f32 %f1759, %f6182, 0f3F000000; add.ftz.f32 %f6186, %f6146, %f6146; mul.ftz.f32 %f6187, %f7223, %f6186; sub.ftz.f32 %f6188, %f6187, %f6154; add.ftz.f32 %f6189, %f7223, %f7223; mul.ftz.f32 %f6190, %f6189, %f6150; fma.rn.ftz.f32 %f6191, %f7223, %f6149, %f6190; add.ftz.f32 %f6192, %f6149, 0f40000000; mul.ftz.f32 %f6193, %f6186, %f6192; mul.ftz.f32 %f6194, %f6146, %f6155; sub.ftz.f32 %f6195, %f6193, %f6194; mul.ftz.f32 %f6196, %f6146, 0fC0000000; fma.rn.ftz.f32 %f6197, %f7223, %f6196, %f6151; mul.ftz.f32 %f6198, %f1644, %f6191; mul.ftz.f32 %f6199, %f1645, %f6191; mul.ftz.f32 %f6200, %f1646, %f6191; mul.ftz.f32 %f6201, %f1539, %f6191; fma.rn.ftz.f32 %f6202, %f1641, %f6188, %f6198; fma.rn.ftz.f32 %f6203, %f1642, %f6188, %f6199; fma.rn.ftz.f32 %f6204, %f1643, %f6188, %f6200; fma.rn.ftz.f32 %f6205, %f1535, %f6188, %f6201; fma.rn.ftz.f32 %f6206, %f1647, %f6195, %f6202; fma.rn.ftz.f32 %f6207, %f1648, %f6195, %f6203; fma.rn.ftz.f32 %f6208, %f1649, %f6195, %f6204; fma.rn.ftz.f32 %f6209, %f1543, %f6195, %f6205; fma.rn.ftz.f32 %f6210, %f1650, %f6197, %f6206; fma.rn.ftz.f32 %f6211, %f1651, %f6197, %f6207; fma.rn.ftz.f32 %f6212, %f1652, %f6197, %f6208; fma.rn.ftz.f32 %f6213, %f1547, %f6197, %f6209; mul.ftz.f32 %f1760, %f6210, 0f3F000000; mul.ftz.f32 %f1761, %f6211, 0f3F000000; mul.ftz.f32 %f1762, %f6212, 0f3F000000; mul.ftz.f32 %f1763, %f6213, 0f3F000000; sub.ftz.f32 %f6214, %f6153, %f6149; mul.ftz.f32 %f6215, %f7223, 0f41100000; add.ftz.f32 %f6216, %f6215, 0fC0A00000; mov.f32 %f6217, 0f40800000; sub.ftz.f32 %f6218, %f6217, %f6215; add.ftz.f32 %f6219, %f6149, 0fBF800000; mul.ftz.f32 %f6220, %f1644, %f6216; mul.ftz.f32 %f6221, %f1645, %f6216; mul.ftz.f32 %f6222, %f1646, %f6216; fma.rn.ftz.f32 %f6223, %f1641, %f6214, %f6220; fma.rn.ftz.f32 %f6224, %f1642, %f6214, %f6221; fma.rn.ftz.f32 %f6225, %f1643, %f6214, %f6222; fma.rn.ftz.f32 %f6226, %f1647, %f6218, %f6223; fma.rn.ftz.f32 %f6227, %f1648, %f6218, %f6224; fma.rn.ftz.f32 %f6228, %f1649, %f6218, %f6225; fma.rn.ftz.f32 %f6229, %f1650, %f6219, %f6226; fma.rn.ftz.f32 %f6230, %f1651, %f6219, %f6227; fma.rn.ftz.f32 %f6231, %f1652, %f6219, %f6228; mul.ftz.f32 %f6232, %f7289, %f7222; sub.ftz.f32 %f1764, %f6232, %f6183; mul.ftz.f32 %f6233, %f7290, %f7222; sub.ftz.f32 %f1765, %f6233, %f6184; mul.ftz.f32 %f6234, %f7291, %f7222; sub.ftz.f32 %f1766, %f6234, %f6185; mul.ftz.f32 %f6235, %f1765, %f1765; fma.rn.ftz.f32 %f6236, %f1764, %f1764, %f6235; fma.rn.ftz.f32 %f6237, %f1766, %f1766, %f6236; sqrt.approx.ftz.f32 %f1767, %f6237; mul.ftz.f32 %f6238, %f1755, %f7222; max.ftz.f32 %f1768, %f6238, %f1754; mul.ftz.f32 %f6239, %f6210, 0fBF000000; mul.ftz.f32 %f6240, %f1761, %f1761; fma.rn.ftz.f32 %f6241, %f1760, %f1760, %f6240; fma.rn.ftz.f32 %f6242, %f1762, %f1762, %f6241; sqrt.approx.ftz.f32 %f6243, %f6242; div.approx.ftz.f32 %f6244, %f1760, %f6243; div.approx.ftz.f32 %f6245, %f1761, %f6243; div.approx.ftz.f32 %f6246, %f1762, %f6243; mul.ftz.f32 %f6247, %f6230, %f1761; fma.rn.ftz.f32 %f6248, %f6229, %f1760, %f6247; fma.rn.ftz.f32 %f6249, %f6231, %f1762, %f6248; mul.ftz.f32 %f6250, %f6229, %f6242; mul.ftz.f32 %f6251, %f6230, %f6242; mul.ftz.f32 %f6252, %f6231, %f6242; mul.ftz.f32 %f6253, %f1760, %f6249; mul.ftz.f32 %f6254, %f1761, %f6249; mul.ftz.f32 %f6255, %f1762, %f6249; sub.ftz.f32 %f6256, %f6250, %f6253; sub.ftz.f32 %f6257, %f6251, %f6254; sub.ftz.f32 %f6258, %f6252, %f6255; mul.ftz.f32 %f6259, %f6242, %f6243; div.approx.ftz.f32 %f6260, %f6256, %f6259; div.approx.ftz.f32 %f6261, %f6257, %f6259; div.approx.ftz.f32 %f6262, %f6258, %f6259; div.approx.ftz.f32 %f6263, %f1754, %f6243; mul.ftz.f32 %f6264, %f1765, %f6245; fma.rn.ftz.f32 %f6265, %f1764, %f6244, %f6264; fma.rn.ftz.f32 %f1769, %f1766, %f6246, %f6265; fma.rn.ftz.f32 %f6266, %f1754, %f1767, %f1768; add.ftz.f32 %f6267, %f1767, 0f3F800000; fma.rn.ftz.f32 %f1770, %f6267, %f6263, %f6266; mul.ftz.f32 %f6268, %f6244, %f6239; mul.ftz.f32 %f6269, %f6245, %f1761; sub.ftz.f32 %f6270, %f6268, %f6269; mul.ftz.f32 %f6271, %f6246, %f1762; sub.ftz.f32 %f6272, %f6270, %f6271; mul.ftz.f32 %f6273, %f1765, %f6261; fma.rn.ftz.f32 %f6274, %f1764, %f6260, %f6273; fma.rn.ftz.f32 %f6275, %f1766, %f6262, %f6274; add.ftz.f32 %f1771, %f6272, %f6275; mul.ftz.f32 %f6276, %f7290, %f6245; fma.rn.ftz.f32 %f6277, %f7289, %f6244, %f6276; fma.rn.ftz.f32 %f1772, %f7291, %f6246, %f6277; mul.ftz.f32 %f6278, %f1769, %f1769; sub.ftz.f32 %f6279, %f6237, %f6278; mul.ftz.f32 %f6280, %f1764, %f6239; mul.ftz.f32 %f6281, %f1765, %f1761; sub.ftz.f32 %f6282, %f6280, %f6281; mul.ftz.f32 %f6283, %f1766, %f1762; sub.ftz.f32 %f1773, %f6282, %f6283; mul.ftz.f32 %f6284, %f7290, %f1765; fma.rn.ftz.f32 %f6285, %f7289, %f1764, %f6284; fma.rn.ftz.f32 %f1774, %f7291, %f1766, %f6285; setp.leu.ftz.f32 %p703, %f6279, 0f00000000; sqrt.approx.ftz.f32 %f1775, %f6279; mov.f32 %f7224, %f5852; @%p703 bra $L__BB24_535; rcp.approx.ftz.f32 %f7224, %f1775; $L__BB24_535: add.ftz.f32 %f6286, %f1768, %f1770; add.ftz.f32 %f1778, %f1756, %f6286; mul.ftz.f32 %f6287, %f1769, %f1771; sub.ftz.f32 %f6288, %f1773, %f6287; mul.ftz.f32 %f6289, %f6288, %f7224; sub.ftz.f32 %f6290, %f6289, %f1763; mul.ftz.f32 %f6291, %f1769, %f1772; sub.ftz.f32 %f6292, %f1774, %f6291; mul.ftz.f32 %f6293, %f6292, %f7224; mul.ftz.f32 %f6294, %f1771, %f6293; mul.ftz.f32 %f6295, %f1772, %f6290; sub.ftz.f32 %f6296, %f6294, %f6295; rcp.approx.ftz.f32 %f6297, %f6296; mul.ftz.f32 %f6298, %f1769, %f6293; sub.ftz.f32 %f1779, %f1775, %f1759; mul.ftz.f32 %f6299, %f1772, %f1779; sub.ftz.f32 %f6300, %f6298, %f6299; mul.ftz.f32 %f6301, %f6300, %f6297; sub.ftz.f32 %f7223, %f7223, %f6301; mul.ftz.f32 %f6302, %f1769, %f6290; mul.ftz.f32 %f6303, %f1771, %f1779; sub.ftz.f32 %f6304, %f6303, %f6302; mul.ftz.f32 %f6305, %f6304, %f6297; sub.ftz.f32 %f7222, %f7222, %f6305; abs.ftz.f32 %f6306, %f1769; setp.geu.ftz.f32 %p704, %f6306, %f1770; @%p704 bra $L__BB24_537; abs.ftz.f32 %f6307, %f1779; setp.lt.ftz.f32 %p705, %f6307, %f1778; @%p705 bra $L__BB24_538; $L__BB24_537: add.s32 %r1498, %r1498, 1; setp.lt.u32 %p706, %r1498, 5; mov.u16 %rs234, 0; @%p706 bra $L__BB24_533; bra.uni $L__BB24_542; $L__BB24_538: add.ftz.f32 %f1782, %f1640, %f7222; setp.ltu.ftz.f32 %p707, %f1782, %f1291; setp.gtu.ftz.f32 %p708, %f1782, %f7238; or.pred %p709, %p707, %p708; mov.u16 %rs234, 0; @%p709 bra $L__BB24_542; setp.ltu.ftz.f32 %p710, %f7223, 0f00000000; setp.gtu.ftz.f32 %p711, %f7223, 0f3F800000; or.pred %p712, %p711, %p710; @%p712 bra $L__BB24_542; div.approx.ftz.f32 %f6308, %f1764, %f1767; div.approx.ftz.f32 %f6309, %f1765, %f1767; div.approx.ftz.f32 %f6310, %f1766, %f1767; fma.rn.ftz.f32 %f6311, %f1763, %f6308, %f1760; fma.rn.ftz.f32 %f6312, %f1763, %f6309, %f1761; fma.rn.ftz.f32 %f6313, %f1763, %f6310, %f1762; mul.ftz.f32 %f6314, %f1761, %f6310; mul.ftz.f32 %f6315, %f1762, %f6309; sub.ftz.f32 %f6316, %f6314, %f6315; mul.ftz.f32 %f6317, %f1762, %f6308; mul.ftz.f32 %f6318, %f1760, %f6310; sub.ftz.f32 %f6319, %f6317, %f6318; mul.ftz.f32 %f6320, %f1760, %f6309; mul.ftz.f32 %f6321, %f1761, %f6308; sub.ftz.f32 %f6322, %f6320, %f6321; mul.ftz.f32 %f6323, %f6319, %f6313; mul.ftz.f32 %f6324, %f6322, %f6312; sub.ftz.f32 %f6325, %f6323, %f6324; mul.ftz.f32 %f6326, %f6311, %f6322; mul.ftz.f32 %f6327, %f6316, %f6313; sub.ftz.f32 %f6328, %f6326, %f6327; mul.ftz.f32 %f6329, %f6312, %f6316; mul.ftz.f32 %f6330, %f6311, %f6319; sub.ftz.f32 %f6331, %f6329, %f6330; mul.ftz.f32 %f6332, %f7290, %f6328; fma.rn.ftz.f32 %f6333, %f7289, %f6325, %f6332; fma.rn.ftz.f32 %f6334, %f7291, %f6331, %f6333; setp.gt.ftz.f32 %p713, %f6334, 0f00000000; @%p713 bra $L__BB24_542; mov.u16 %rs234, 1; mov.f32 %f7253, 0f00000000; mov.f32 %f7251, %f1782; mov.f32 %f7252, %f7223; mov.f32 %f7238, %f1782; $L__BB24_542: or.b16 %rs173, %rs234, %rs237; and.b16 %rs174, %rs173, 255; setp.ne.s16 %p714, %rs174, 0; selp.u16 %rs237, 1, 0, %p714; mov.u16 %rs236, %rs233; $L__BB24_543: setp.eq.s16 %p715, %rs236, 0; @%p715 bra $L__BB24_545; mul.wide.s32 %rd1011, %r1499, 12; add.s64 %rd1012, %rd1, %rd1011; st.local.f32 [%rd1012], %f7239; st.local.f32 [%rd1012+4], %f7240; add.s32 %r1166, %r1500, 1; st.local.u32 [%rd1012+8], %r1166; add.s32 %r1499, %r1499, 1; mov.u32 %r1500, -1; mov.f32 %f7239, %f1672; mov.f32 %f7240, %f1673; $L__BB24_545: add.s32 %r1495, %r1500, 1; setp.lt.s32 %p716, %r1500, 1; @%p716 bra $L__BB24_507; $L__BB24_546: setp.gt.s32 %p717, %r1499, 0; @%p717 bra $L__BB24_548; cvt.u32.u64 %r1167, %rd283; and.b16 %rs175, %rs237, 255; setp.eq.s16 %p718, %rs175, 0; selp.b32 %r1502, %r1502, %r1167, %p718; selp.b32 %r1503, %r1503, %r541, %p718; selp.b32 %r1504, %r1504, %r560, %p718; setp.ne.s16 %p806, %rs175, 0; bra.uni $L__BB24_557; $L__BB24_490: rsqrt.approx.ftz.f32 %f1548, %f1364; mov.f32 %f5222, 0f3F800000; mul.ftz.f32 %f5223, %f7289, %f1548; mul.ftz.f32 %f5224, %f7290, %f1548; mul.ftz.f32 %f5225, %f7291, %f1548; neg.ftz.f32 %f5226, %f5224; neg.ftz.f32 %f5227, %f5225; fma.rn.ftz.f32 %f5228, %f5225, %f5225, 0f00000000; fma.rn.ftz.f32 %f5229, %f5224, %f5224, %f5228; fma.rn.ftz.f32 %f5230, %f5223, %f5223, %f5228; setp.gt.ftz.f32 %p641, %f5229, %f5230; selp.f32 %f5231, 0f00000000, %f5227, %p641; selp.f32 %f5232, %f5225, 0f00000000, %p641; selp.f32 %f5233, %f5226, %f5223, %p641; mul.ftz.f32 %f5234, %f5232, %f5232; fma.rn.ftz.f32 %f5235, %f5231, %f5231, %f5234; fma.rn.ftz.f32 %f5236, %f5233, %f5233, %f5235; sqrt.approx.ftz.f32 %f5237, %f5236; div.approx.ftz.f32 %f5238, %f5231, %f5237; div.approx.ftz.f32 %f5239, %f5232, %f5237; div.approx.ftz.f32 %f5240, %f5233, %f5237; mul.ftz.f32 %f5241, %f5224, %f5240; mul.ftz.f32 %f5242, %f5225, %f5239; sub.ftz.f32 %f5243, %f5241, %f5242; mul.ftz.f32 %f5244, %f5225, %f5238; mul.ftz.f32 %f5245, %f5223, %f5240; sub.ftz.f32 %f5246, %f5244, %f5245; mul.ftz.f32 %f5247, %f5223, %f5239; mul.ftz.f32 %f5248, %f5224, %f5238; sub.ftz.f32 %f5249, %f5247, %f5248; mul.ftz.f32 %f5250, %f5246, %f5246; fma.rn.ftz.f32 %f5251, %f5243, %f5243, %f5250; fma.rn.ftz.f32 %f5252, %f5249, %f5249, %f5251; sqrt.approx.ftz.f32 %f5253, %f5252; div.approx.ftz.f32 %f5254, %f5243, %f5253; div.approx.ftz.f32 %f5255, %f5246, %f5253; div.approx.ftz.f32 %f5256, %f5249, %f5253; mul.ftz.f32 %f5257, %f1548, %f5223; mul.ftz.f32 %f5258, %f1548, %f5224; mul.ftz.f32 %f5259, %f1548, %f5225; sub.ftz.f32 %f5260, %f7147, %f7292; sub.ftz.f32 %f5261, %f7148, %f7293; mul.ftz.f32 %f5262, %f5261, %f5239; fma.rn.ftz.f32 %f5263, %f5260, %f5238, %f5262; sub.ftz.f32 %f5264, %f7149, %f7294; fma.rn.ftz.f32 %f1549, %f5264, %f5240, %f5263; mul.ftz.f32 %f5265, %f5261, %f5255; fma.rn.ftz.f32 %f5266, %f5260, %f5254, %f5265; fma.rn.ftz.f32 %f1550, %f5264, %f5256, %f5266; mul.ftz.f32 %f5267, %f5261, %f5258; fma.rn.ftz.f32 %f5268, %f5260, %f5257, %f5267; fma.rn.ftz.f32 %f1551, %f5264, %f5259, %f5268; sub.ftz.f32 %f5269, %f7151, %f7292; sub.ftz.f32 %f5270, %f7152, %f7293; mul.ftz.f32 %f5271, %f5270, %f5239; fma.rn.ftz.f32 %f5272, %f5269, %f5238, %f5271; sub.ftz.f32 %f5273, %f7153, %f7294; fma.rn.ftz.f32 %f1552, %f5273, %f5240, %f5272; mul.ftz.f32 %f5274, %f5270, %f5255; fma.rn.ftz.f32 %f5275, %f5269, %f5254, %f5274; fma.rn.ftz.f32 %f1553, %f5273, %f5256, %f5275; mul.ftz.f32 %f5276, %f5270, %f5258; fma.rn.ftz.f32 %f5277, %f5269, %f5257, %f5276; fma.rn.ftz.f32 %f1554, %f5273, %f5259, %f5277; sub.ftz.f32 %f5278, %f7155, %f7292; sub.ftz.f32 %f5279, %f7156, %f7293; mul.ftz.f32 %f5280, %f5279, %f5239; fma.rn.ftz.f32 %f5281, %f5278, %f5238, %f5280; sub.ftz.f32 %f5282, %f7157, %f7294; fma.rn.ftz.f32 %f1555, %f5282, %f5240, %f5281; mul.ftz.f32 %f5283, %f5279, %f5255; fma.rn.ftz.f32 %f5284, %f5278, %f5254, %f5283; fma.rn.ftz.f32 %f1556, %f5282, %f5256, %f5284; mul.ftz.f32 %f5285, %f5279, %f5258; fma.rn.ftz.f32 %f5286, %f5278, %f5257, %f5285; fma.rn.ftz.f32 %f1557, %f5282, %f5259, %f5286; sub.ftz.f32 %f5287, %f7159, %f7292; sub.ftz.f32 %f5288, %f7160, %f7293; mul.ftz.f32 %f5289, %f5288, %f5239; fma.rn.ftz.f32 %f5290, %f5287, %f5238, %f5289; sub.ftz.f32 %f5291, %f7161, %f7294; fma.rn.ftz.f32 %f1558, %f5291, %f5240, %f5290; mul.ftz.f32 %f5292, %f5288, %f5255; fma.rn.ftz.f32 %f5293, %f5287, %f5254, %f5292; fma.rn.ftz.f32 %f1559, %f5291, %f5256, %f5293; mul.ftz.f32 %f5294, %f5288, %f5258; fma.rn.ftz.f32 %f5295, %f5287, %f5257, %f5294; fma.rn.ftz.f32 %f1560, %f5291, %f5259, %f5295; abs.ftz.f32 %f5296, %f1549; abs.ftz.f32 %f5297, %f1550; abs.ftz.f32 %f5298, %f1551; abs.ftz.f32 %f5299, %f1552; abs.ftz.f32 %f5300, %f1553; abs.ftz.f32 %f5301, %f1554; max.ftz.f32 %f5302, %f5296, %f5299; max.ftz.f32 %f5303, %f5297, %f5300; max.ftz.f32 %f5304, %f5298, %f5301; abs.ftz.f32 %f5305, %f1539; abs.ftz.f32 %f5306, %f1535; max.ftz.f32 %f5307, %f5306, %f5305; abs.ftz.f32 %f5308, %f1555; abs.ftz.f32 %f5309, %f1556; abs.ftz.f32 %f5310, %f1557; abs.ftz.f32 %f5311, %f1558; abs.ftz.f32 %f5312, %f1559; abs.ftz.f32 %f5313, %f1560; max.ftz.f32 %f5314, %f5308, %f5311; max.ftz.f32 %f5315, %f5309, %f5312; max.ftz.f32 %f5316, %f5310, %f5313; abs.ftz.f32 %f5317, %f1547; abs.ftz.f32 %f5318, %f1543; max.ftz.f32 %f5319, %f5318, %f5317; max.ftz.f32 %f5320, %f5302, %f5314; max.ftz.f32 %f5321, %f5303, %f5315; max.ftz.f32 %f5322, %f5304, %f5316; max.ftz.f32 %f5323, %f5307, %f5319; max.ftz.f32 %f5324, %f5320, %f5321; max.ftz.f32 %f5325, %f5322, %f5323; max.ftz.f32 %f5326, %f5324, %f5325; mul.ftz.f32 %f1561, %f5326, 0f35000000; sub.ftz.f32 %f5327, %f5222, 0f00000000; mul.ftz.f32 %f5328, %f5327, 0f00000000; mul.ftz.f32 %f5329, %f5328, %f5327; mul.ftz.f32 %f5330, %f1549, %f5329; mul.ftz.f32 %f5331, %f1550, %f5329; mul.ftz.f32 %f5332, %f1551, %f5329; add.ftz.f32 %f5333, %f1552, %f1552; add.ftz.f32 %f5334, %f1553, %f1553; add.ftz.f32 %f5335, %f1554, %f1554; sub.ftz.f32 %f5336, %f5333, %f5330; sub.ftz.f32 %f5337, %f5334, %f5331; sub.ftz.f32 %f5338, %f5335, %f5332; add.ftz.f32 %f5339, %f1539, %f1539; mul.ftz.f32 %f5340, %f1535, %f5329; sub.ftz.f32 %f5341, %f5339, %f5340; mul.ftz.f32 %f5342, %f5327, %f5327; mul.ftz.f32 %f5343, %f5327, 0f40400000; add.ftz.f32 %f5344, %f5343, 0fC0A00000; fma.rn.ftz.f32 %f5345, %f5342, %f5344, 0f40000000; fma.rn.ftz.f32 %f5346, %f1555, %f5345, %f5336; fma.rn.ftz.f32 %f5347, %f1556, %f5345, %f5337; fma.rn.ftz.f32 %f5348, %f1557, %f5345, %f5338; fma.rn.ftz.f32 %f5349, %f1543, %f5345, %f5341; mul.ftz.f32 %f5350, %f5328, 0f00000000; mul.ftz.f32 %f5351, %f1558, %f5350; mul.ftz.f32 %f5352, %f1559, %f5350; mul.ftz.f32 %f5353, %f1560, %f5350; sub.ftz.f32 %f5354, %f5346, %f5351; sub.ftz.f32 %f5355, %f5347, %f5352; sub.ftz.f32 %f1562, %f5348, %f5353; mul.ftz.f32 %f5356, %f1547, %f5350; sub.ftz.f32 %f5357, %f5349, %f5356; mul.ftz.f32 %f7168, %f5354, 0f3F000000; mul.ftz.f32 %f7167, %f5355, 0f3F000000; mul.ftz.f32 %f7165, %f5357, 0f3F000000; add.ftz.f32 %f5358, %f5327, %f5327; mul.ftz.f32 %f5359, %f5358, 0f00000000; sub.ftz.f32 %f5360, %f5359, %f5342; mul.ftz.f32 %f5361, %f1552, 0f00000000; mul.ftz.f32 %f5362, %f1553, 0f00000000; mul.ftz.f32 %f5363, %f1554, 0f00000000; fma.rn.ftz.f32 %f5364, %f1549, %f5360, %f5361; fma.rn.ftz.f32 %f5365, %f1550, %f5360, %f5362; fma.rn.ftz.f32 %f5366, %f1551, %f5360, %f5363; mul.ftz.f32 %f5367, %f5343, %f5327; add.ftz.f32 %f5368, %f5358, %f5358; sub.ftz.f32 %f5369, %f5368, %f5367; fma.rn.ftz.f32 %f5370, %f1555, %f5369, %f5364; fma.rn.ftz.f32 %f5371, %f1556, %f5369, %f5365; fma.rn.ftz.f32 %f5372, %f1557, %f5369, %f5366; mul.ftz.f32 %f5373, %f5327, 0fC0000000; fma.rn.ftz.f32 %f5374, %f5373, 0f00000000, 0f00000000; fma.rn.ftz.f32 %f5375, %f1558, %f5374, %f5370; fma.rn.ftz.f32 %f5376, %f1559, %f5374, %f5371; fma.rn.ftz.f32 %f5377, %f1560, %f5374, %f5372; mul.ftz.f32 %f7163, %f5375, 0f3F000000; mul.ftz.f32 %f7164, %f5376, 0f3F000000; mul.ftz.f32 %f5378, %f5377, 0f3F000000; abs.ftz.f32 %f5379, %f7163; abs.ftz.f32 %f5380, %f7164; abs.ftz.f32 %f5381, %f5378; max.ftz.f32 %f5382, %f5379, %f5380; max.ftz.f32 %f5383, %f5382, %f5381; setp.geu.ftz.f32 %p642, %f5383, %f1561; @%p642 bra $L__BB24_492; mul.ftz.f32 %f5384, %f1295, %f1552; fma.rn.ftz.f32 %f5385, %f1549, %f1294, %f5384; mul.ftz.f32 %f5386, %f1295, %f1553; fma.rn.ftz.f32 %f5387, %f1550, %f1294, %f5386; fma.rn.ftz.f32 %f5388, %f1555, %f1296, %f5385; fma.rn.ftz.f32 %f5389, %f1296, %f1556, %f5387; fma.rn.ftz.f32 %f5390, %f1558, %f1297, %f5388; fma.rn.ftz.f32 %f5391, %f1559, %f1297, %f5389; mul.ftz.f32 %f5392, %f5390, 0f3F000000; mul.ftz.f32 %f5393, %f5391, 0f3F000000; sub.ftz.f32 %f7163, %f5392, %f7168; sub.ftz.f32 %f7164, %f5393, %f7167; $L__BB24_492: mul.ftz.f32 %f5394, %f7164, %f7164; fma.rn.ftz.f32 %f1572, %f7163, %f7163, %f5394; setp.lt.s32 %p644, %r1086, 1; mov.pred %p803, -1; @%p644 bra $L__BB24_504; add.ftz.f32 %f5395, %f1572, 0f00000000; mov.f32 %f5396, 0f00000000; sqrt.approx.ftz.f32 %f5397, %f5395; mul.ftz.f32 %f7166, %f1562, 0f3F000000; div.approx.ftz.f32 %f5398, %f5396, %f5397; mul.ftz.f32 %f7169, %f7165, %f5398; neg.ftz.f32 %f5399, %f7163; div.approx.ftz.f32 %f5400, %f5399, %f5397; mul.ftz.f32 %f7170, %f7165, %f5400; div.approx.ftz.f32 %f5401, %f7164, %f5397; mul.ftz.f32 %f7171, %f7165, %f5401; mov.u32 %r1492, 0; $L__BB24_494: mov.f32 %f1583, %f7171; mov.f32 %f1582, %f7170; mov.f32 %f1581, %f7169; mov.f32 %f1580, %f7168; mov.f32 %f1579, %f7167; mov.f32 %f1578, %f7166; mov.f32 %f1577, %f7165; cvt.rn.f32.s32 %f5402, %r1492; mul.ftz.f32 %f1587, %f1293, %f5402; add.ftz.f32 %f5403, %f1293, %f1587; sub.ftz.f32 %f5405, %f5222, %f5403; mul.ftz.f32 %f5406, %f5403, %f5405; mul.ftz.f32 %f5407, %f5405, %f5406; mul.ftz.f32 %f5408, %f5403, %f5403; mul.ftz.f32 %f5409, %f5403, 0f40400000; add.ftz.f32 %f5410, %f5409, 0fC0A00000; fma.rn.ftz.f32 %f5411, %f5408, %f5410, 0f40000000; mul.ftz.f32 %f5412, %f5405, %f5405; mul.ftz.f32 %f5413, %f5405, 0f40400000; add.ftz.f32 %f5414, %f5413, 0fC0A00000; fma.rn.ftz.f32 %f5415, %f5412, %f5414, 0f40000000; mul.ftz.f32 %f5416, %f5403, %f5406; mul.ftz.f32 %f5417, %f1549, %f5407; mul.ftz.f32 %f5418, %f1550, %f5407; mul.ftz.f32 %f5419, %f1551, %f5407; mul.ftz.f32 %f5420, %f1535, %f5407; mul.ftz.f32 %f5421, %f1552, %f5411; mul.ftz.f32 %f5422, %f1553, %f5411; mul.ftz.f32 %f5423, %f1554, %f5411; mul.ftz.f32 %f5424, %f1539, %f5411; sub.ftz.f32 %f5425, %f5421, %f5417; sub.ftz.f32 %f5426, %f5422, %f5418; sub.ftz.f32 %f5427, %f5423, %f5419; sub.ftz.f32 %f5428, %f5424, %f5420; fma.rn.ftz.f32 %f5429, %f1555, %f5415, %f5425; fma.rn.ftz.f32 %f5430, %f1556, %f5415, %f5426; fma.rn.ftz.f32 %f5431, %f1557, %f5415, %f5427; fma.rn.ftz.f32 %f5432, %f1543, %f5415, %f5428; mul.ftz.f32 %f5433, %f1558, %f5416; mul.ftz.f32 %f5434, %f1559, %f5416; mul.ftz.f32 %f5435, %f1560, %f5416; mul.ftz.f32 %f5436, %f1547, %f5416; sub.ftz.f32 %f5437, %f5429, %f5433; sub.ftz.f32 %f5438, %f5430, %f5434; sub.ftz.f32 %f5439, %f5431, %f5435; sub.ftz.f32 %f5440, %f5432, %f5436; mul.ftz.f32 %f7168, %f5437, 0f3F000000; mul.ftz.f32 %f7167, %f5438, 0f3F000000; mul.ftz.f32 %f7166, %f5439, 0f3F000000; mul.ftz.f32 %f7165, %f5440, 0f3F000000; sub.ftz.f32 %f1592, %f7168, %f1580; sub.ftz.f32 %f1593, %f7167, %f1579; mul.ftz.f32 %f5441, %f1593, %f1593; fma.rn.ftz.f32 %f1594, %f1592, %f1592, %f5441; add.ftz.f32 %f5442, %f5405, %f5405; mul.ftz.f32 %f5443, %f5403, %f5442; sub.ftz.f32 %f5444, %f5443, %f5412; add.ftz.f32 %f5445, %f5403, %f5403; mul.ftz.f32 %f5446, %f5445, %f5410; fma.rn.ftz.f32 %f5447, %f5403, %f5409, %f5446; add.ftz.f32 %f5448, %f5409, 0f40000000; mul.ftz.f32 %f5449, %f5442, %f5448; mul.ftz.f32 %f5450, %f5405, %f5413; sub.ftz.f32 %f5451, %f5449, %f5450; mul.ftz.f32 %f5452, %f5405, 0fC0000000; fma.rn.ftz.f32 %f5453, %f5403, %f5452, %f5408; mul.ftz.f32 %f5454, %f1552, %f5447; mul.ftz.f32 %f5455, %f1553, %f5447; mul.ftz.f32 %f5456, %f1554, %f5447; fma.rn.ftz.f32 %f5457, %f1549, %f5444, %f5454; fma.rn.ftz.f32 %f5458, %f1550, %f5444, %f5455; fma.rn.ftz.f32 %f5459, %f1551, %f5444, %f5456; fma.rn.ftz.f32 %f5460, %f1555, %f5451, %f5457; fma.rn.ftz.f32 %f5461, %f1556, %f5451, %f5458; fma.rn.ftz.f32 %f5462, %f1557, %f5451, %f5459; fma.rn.ftz.f32 %f5463, %f1558, %f5453, %f5460; fma.rn.ftz.f32 %f5464, %f1559, %f5453, %f5461; fma.rn.ftz.f32 %f5465, %f1560, %f5453, %f5462; mul.ftz.f32 %f7175, %f5463, 0f3F000000; mul.ftz.f32 %f7176, %f5464, 0f3F000000; mul.ftz.f32 %f1597, %f5465, 0f3F000000; abs.ftz.f32 %f5466, %f7175; abs.ftz.f32 %f5467, %f7176; abs.ftz.f32 %f5468, %f1597; max.ftz.f32 %f5469, %f5466, %f5467; max.ftz.f32 %f5470, %f5469, %f5468; setp.lt.ftz.f32 %p645, %f5470, %f1561; @%p645 bra $L__BB24_495; bra.uni $L__BB24_496; $L__BB24_495: mov.f32 %f7175, %f1592; mov.f32 %f7176, %f1593; $L__BB24_496: max.ftz.f32 %f5471, %f1577, %f7165; mul.ftz.f32 %f5472, %f1580, %f1593; mul.ftz.f32 %f5473, %f1579, %f1592; sub.ftz.f32 %f5474, %f5473, %f5472; neg.ftz.f32 %f5475, %f7175; mul.ftz.f32 %f5476, %f7175, %f7175; fma.rn.ftz.f32 %f5477, %f7176, %f7176, %f5476; add.ftz.f32 %f5478, %f5477, 0f00000000; sqrt.approx.ftz.f32 %f5480, %f5478; div.approx.ftz.f32 %f5481, %f7176, %f5480; div.approx.ftz.f32 %f5482, %f5475, %f5480; div.approx.ftz.f32 %f5483, %f5396, %f5480; mul.ftz.f32 %f7171, %f7165, %f5481; mul.ftz.f32 %f7170, %f7165, %f5482; mul.ftz.f32 %f7169, %f7165, %f5483; mul.ftz.f32 %f5484, %f5471, %f5471; mul.ftz.f32 %f5485, %f1594, %f5484; mul.ftz.f32 %f5486, %f5474, %f5474; setp.gtu.ftz.f32 %p646, %f5486, %f5485; @%p646 bra $L__BB24_503; add.ftz.f32 %f5487, %f1583, %f1580; sub.ftz.f32 %f5488, %f1580, %f1583; add.ftz.f32 %f5489, %f7168, %f7171; sub.ftz.f32 %f5490, %f5489, %f5488; sub.ftz.f32 %f5491, %f1579, %f1582; add.ftz.f32 %f5492, %f7167, %f7170; sub.ftz.f32 %f5493, %f5492, %f5491; sub.ftz.f32 %f5494, %f1578, %f1581; add.ftz.f32 %f5495, %f7166, %f7169; sub.ftz.f32 %f5496, %f5495, %f5494; mul.ftz.f32 %f5497, %f5491, %f5496; mul.ftz.f32 %f5498, %f5494, %f5493; sub.ftz.f32 %f5499, %f5497, %f5498; mul.ftz.f32 %f5500, %f5494, %f5490; mul.ftz.f32 %f5501, %f5488, %f5496; sub.ftz.f32 %f5502, %f5500, %f5501; mul.ftz.f32 %f5503, %f5488, %f5493; mul.ftz.f32 %f5504, %f5491, %f5490; sub.ftz.f32 %f5505, %f5503, %f5504; mul.ftz.f32 %f5506, %f5502, 0f00000000; fma.rn.ftz.f32 %f5507, %f5499, 0f00000000, %f5506; add.ftz.f32 %f1610, %f5505, %f5507; setp.gtu.ftz.f32 %p648, %f1610, 0f00000000; sub.ftz.f32 %f5508, %f7168, %f7171; selp.f32 %f1611, %f5508, %f5487, %p648; sub.ftz.f32 %f5509, %f7167, %f7170; add.ftz.f32 %f5510, %f1582, %f1579; selp.f32 %f1612, %f5509, %f5510, %p648; sub.ftz.f32 %f5511, %f7166, %f7169; add.ftz.f32 %f5512, %f1581, %f1578; selp.f32 %f1613, %f5511, %f5512, %p648; selp.f32 %f5513, %f5488, %f5489, %p648; selp.f32 %f5514, %f5491, %f5492, %p648; selp.f32 %f5515, %f5494, %f5495, %p648; selp.f32 %f5516, %f5489, %f5488, %p648; selp.f32 %f5517, %f5492, %f5491, %p648; selp.f32 %f5518, %f5495, %f5494, %p648; sub.ftz.f32 %f1614, %f5516, %f1611; sub.ftz.f32 %f1615, %f5517, %f1612; sub.ftz.f32 %f1616, %f5518, %f1613; sub.ftz.f32 %f1617, %f1611, %f5513; sub.ftz.f32 %f1618, %f1612, %f5514; sub.ftz.f32 %f1619, %f1613, %f5515; mul.ftz.f32 %f5519, %f1612, %f1616; mul.ftz.f32 %f5520, %f1613, %f1615; sub.ftz.f32 %f5521, %f5519, %f5520; mul.ftz.f32 %f5522, %f1613, %f1614; mul.ftz.f32 %f5523, %f1611, %f1616; sub.ftz.f32 %f5524, %f5522, %f5523; mul.ftz.f32 %f5525, %f1611, %f1615; mul.ftz.f32 %f5526, %f1612, %f1614; sub.ftz.f32 %f5527, %f5525, %f5526; mul.ftz.f32 %f5528, %f5524, 0f00000000; fma.rn.ftz.f32 %f5529, %f5521, 0f00000000, %f5528; add.ftz.f32 %f1620, %f5527, %f5529; mul.ftz.f32 %f5530, %f5514, %f1619; mul.ftz.f32 %f5531, %f5515, %f1618; sub.ftz.f32 %f5532, %f5530, %f5531; mul.ftz.f32 %f5533, %f5515, %f1617; mul.ftz.f32 %f5534, %f5513, %f1619; sub.ftz.f32 %f5535, %f5533, %f5534; mul.ftz.f32 %f5536, %f5513, %f1618; mul.ftz.f32 %f5537, %f5514, %f1617; sub.ftz.f32 %f5538, %f5536, %f5537; mul.ftz.f32 %f5539, %f5535, 0f00000000; fma.rn.ftz.f32 %f5540, %f5532, 0f00000000, %f5539; add.ftz.f32 %f1621, %f5538, %f5540; max.ftz.f32 %f5541, %f1620, %f1621; setp.gtu.ftz.f32 %p649, %f5541, 0f00000000; mov.pred %p802, -1; @%p649 bra $L__BB24_500; mul.ftz.f32 %f5542, %f1618, %f1616; mul.ftz.f32 %f5543, %f1619, %f1615; sub.ftz.f32 %f5544, %f5542, %f5543; mul.ftz.f32 %f5545, %f1617, %f1616; mul.ftz.f32 %f5546, %f1619, %f1614; sub.ftz.f32 %f5547, %f5546, %f5545; mul.ftz.f32 %f5548, %f1618, %f1614; mul.ftz.f32 %f5549, %f1617, %f1615; sub.ftz.f32 %f5550, %f5549, %f5548; mul.ftz.f32 %f5551, %f5547, 0f00000000; fma.rn.ftz.f32 %f5552, %f5544, 0f00000000, %f5551; add.ftz.f32 %f5553, %f5550, %f5552; rcp.approx.ftz.f32 %f1622, %f5553; mul.ftz.f32 %f5554, %f1612, %f5547; fma.rn.ftz.f32 %f5555, %f1611, %f5544, %f5554; fma.rn.ftz.f32 %f5556, %f1613, %f5550, %f5555; mul.ftz.f32 %f1623, %f5556, %f1622; setp.ltu.ftz.f32 %p651, %f1623, %f1291; setp.gtu.ftz.f32 %p652, %f1623, %f1292; or.pred %p653, %p652, %p651; setp.eq.ftz.f32 %p654, %f5553, 0f00000000; or.pred %p655, %p654, %p653; @%p655 bra $L__BB24_500; mul.ftz.f32 %f5557, %f1620, %f1622; sub.ftz.f32 %f5559, %f5222, %f5557; setp.le.ftz.f32 %p657, %f1610, 0f00000000; selp.f32 %f7259, %f5557, %f5559, %p657; mul.ftz.f32 %f5560, %f1621, %f1622; sub.ftz.f32 %f5561, %f5222, %f5560; selp.f32 %f7258, %f5560, %f5561, %p657; mov.pred %p802, 0; mov.f32 %f7257, %f1623; $L__BB24_500: @%p802 bra $L__BB24_503; sub.ftz.f32 %f5562, %f7165, %f1577; fma.rn.ftz.f32 %f5563, %f5562, %f7259, %f1577; add.ftz.f32 %f5564, %f5563, %f5563; mul.ftz.f32 %f5565, %f1548, %f5564; setp.leu.ftz.f32 %p658, %f7257, %f5565; @%p658 bra $L__BB24_503; bra.uni $L__BB24_502; $L__BB24_503: add.s32 %r1492, %r1492, 1; setp.lt.s32 %p661, %r1492, %r1086; @%p661 bra $L__BB24_494; bra.uni $L__BB24_504; $L__BB24_468: cvt.u32.u64 %r1123, %rd283; setp.gt.s32 %p619, %r545, %r544; selp.b32 %r1124, -1, 0, %p619; add.s32 %r1125, %r545, %r1124; mad.lo.s32 %r1126, %r1125, %r543, %r1123; add.s32 %r1127, %r1126, %r1486; mul.wide.s32 %rd984, %r1127, 16; add.s64 %rd1226, %rd276, %rd984; $L__BB24_470: ld.global.v4.f32 {%f5073, %f5074, %f5075, %f5076}, [%rd1226]; add.s32 %r556, %r545, 1; setp.eq.s32 %p620, %r556, %r544; @%p620 bra $L__BB24_472; bra.uni $L__BB24_471; $L__BB24_472: shl.b64 %rd987, %rd283, 4; add.s64 %rd1227, %rd274, %rd987; bra.uni $L__BB24_473; $L__BB24_471: cvt.u32.u64 %r1128, %rd283; setp.ge.s32 %p621, %r545, %r544; selp.b32 %r1129, -1, 0, %p621; add.s32 %r1130, %r556, %r1129; mad.lo.s32 %r1131, %r1130, %r543, %r1128; add.s32 %r1132, %r1131, %r1486; mul.wide.s32 %rd986, %r1132, 16; add.s64 %rd1227, %rd276, %rd986; $L__BB24_473: cvt.rn.f32.s32 %f5077, %r545; sub.ftz.f32 %f5078, %f1379, %f5077; ld.global.v4.f32 {%f5079, %f5080, %f5081, %f5082}, [%rd1227]; mov.f32 %f5087, 0f3F800000; sub.ftz.f32 %f5088, %f5087, %f5078; mul.ftz.f32 %f5089, %f5078, %f5079; mul.ftz.f32 %f5090, %f5078, %f5080; mul.ftz.f32 %f5091, %f5078, %f5081; mul.ftz.f32 %f5092, %f5078, %f5082; fma.rn.ftz.f32 %f7113, %f5088, %f5076, %f5092; fma.rn.ftz.f32 %f7112, %f5088, %f5075, %f5091; fma.rn.ftz.f32 %f7111, %f5088, %f5074, %f5090; fma.rn.ftz.f32 %f7110, %f5088, %f5073, %f5089; $L__BB24_475: sub.ftz.f32 %f5097, %f7110, %f7292; sub.ftz.f32 %f5098, %f7111, %f7293; sub.ftz.f32 %f5099, %f7112, %f7294; mul.ftz.f32 %f5100, %f7290, %f5098; fma.rn.ftz.f32 %f5101, %f7289, %f5097, %f5100; fma.rn.ftz.f32 %f5102, %f7291, %f5099, %f5101; mul.ftz.f32 %f1396, %f1365, %f5102; mul.ftz.f32 %f5103, %f7289, %f1396; mul.ftz.f32 %f5104, %f7290, %f1396; mul.ftz.f32 %f5105, %f7291, %f1396; sub.ftz.f32 %f5106, %f5097, %f5103; sub.ftz.f32 %f5107, %f5098, %f5104; sub.ftz.f32 %f5108, %f5099, %f5105; mul.ftz.f32 %f5109, %f5107, %f5107; fma.rn.ftz.f32 %f5110, %f5106, %f5106, %f5109; fma.rn.ftz.f32 %f1397, %f5108, %f5108, %f5110; mul.ftz.f32 %f1398, %f7113, %f7113; setp.gtu.ftz.f32 %p623, %f1397, %f1398; mov.pred %p801, -1; @%p623 bra $L__BB24_477; sub.ftz.f32 %f5111, %f1398, %f1397; mul.ftz.f32 %f5112, %f1365, %f5111; sqrt.approx.ftz.f32 %f5113, %f5112; sub.ftz.f32 %f5114, %f1396, %f5113; setp.ltu.ftz.f32 %p624, %f5114, %f1291; setp.gtu.ftz.f32 %p625, %f5114, %f1292; or.pred %p801, %p625, %p624; selp.f32 %f7251, %f7251, %f5114, %p801; $L__BB24_477: selp.f32 %f7252, %f7252, 0f00000000, %p801; selp.f32 %f7253, %f7253, 0f00000000, %p801; cvt.u32.u64 %r1133, %rd283; selp.b32 %r1502, %r1502, %r1133, %p801; selp.b32 %r1503, %r1503, %r541, %p801; selp.b32 %r1504, %r1504, %r542, %p801; not.pred %p806, %p801; bra.uni $L__BB24_557; $L__BB24_502: fma.rn.ftz.f32 %f7253, %f7258, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f7252, %f1293, %f7259, %f1587; mov.pred %p803, 0; mov.f32 %f7251, %f7257; mov.f32 %f7258, %f7253; $L__BB24_504: cvt.u32.u64 %r1160, %rd283; selp.b32 %r1502, %r1502, %r1160, %p803; selp.b32 %r1503, %r1503, %r541, %p803; selp.b32 %r1504, %r1504, %r560, %p803; not.pred %p806, %p803; $L__BB24_557: not.pred %p733, %p806; @%p733 bra $L__BB24_589; and.b32 %r1174, %r1504, 1; setp.eq.b32 %p734, %r1174, 1; mov.pred %p735, 0; xor.pred %p736, %p734, %p735; not.pred %p737, %p736; mul.wide.s32 %rd1025, %r1502, 16; add.s64 %rd302, %rd267, %rd1025; @%p737 bra $L__BB24_560; bra.uni $L__BB24_559; $L__BB24_560: and.b32 %r1175, %r1504, 8; setp.eq.s32 %p738, %r1175, 0; @%p738 bra $L__BB24_562; bra.uni $L__BB24_561; $L__BB24_562: and.b32 %r1177, %r1504, 6; setp.eq.s32 %p739, %r1177, 0; mov.u32 %r1505, 0; @%p739 bra $L__BB24_564; ld.global.u32 %r1505, [%rd302]; bra.uni $L__BB24_564; $L__BB24_559: mul.wide.s32 %rd1026, %r1502, 4; add.s64 %rd1027, %rd270, %rd1026; ld.global.u32 %r1505, [%rd1027]; bra.uni $L__BB24_564; $L__BB24_561: mul.wide.s32 %rd1028, %r1502, 4; add.s64 %rd1029, %rd275, %rd1028; ld.global.u32 %r1505, [%rd1029]; $L__BB24_564: shl.b32 %r1178, %r1505, 5; cvt.u64.u32 %rd1030, %r1178; and.b64 %rd1031, %rd1030, 134217696; add.s64 %rd1032, %rd269, %rd1031; ld.global.u8 %rs177, [%rd1032+18]; and.b16 %rs178, %rs177, 2; setp.eq.s16 %p740, %rs178, 0; setp.ge.u32 %p741, %r1528, %r1; or.pred %p742, %p741, %p740; mov.u16 %rs242, %rs153; @%p742 bra $L__BB24_594; and.b32 %r1179, %r1504, 6; setp.eq.s32 %p744, %r1179, 0; mov.pred %p807, -1; @%p744 bra $L__BB24_575; mul.wide.s32 %rd1033, %r1503, 208; add.s64 %rd1034, %rd261, %rd1033; ld.global.u32 %r1507, [%rd1034+164]; mul.wide.u32 %rd1035, %r1507, 16; add.s64 %rd1036, %rd273, %rd1035; ld.global.u16 %rs239, [%rd1036+12]; ld.global.u32 %r1506, [%rd1036+8]; ld.global.u64 %rd1229, [%rd1036]; setp.eq.s64 %p745, %rd1229, 31; @%p745 bra $L__BB24_571; $L__BB24_567: setp.eq.s64 %p746, %rd1229, 0; @%p746 bra $L__BB24_569; add.s32 %r1507, %r1507, 2; bra.uni $L__BB24_570; $L__BB24_569: setp.eq.s16 %p747, %rs239, 0; mov.u32 %r1510, -1; mov.u32 %r1507, %r1506; @%p747 bra $L__BB24_572; $L__BB24_570: mul.wide.u32 %rd1037, %r1507, 16; add.s64 %rd1038, %rd273, %rd1037; ld.global.u16 %rs239, [%rd1038+12]; ld.global.u32 %r1506, [%rd1038+8]; ld.global.u64 %rd1229, [%rd1038]; setp.ne.s64 %p748, %rd1229, 31; @%p748 bra $L__BB24_567; $L__BB24_571: setp.eq.s16 %p749, %rs239, 0; selp.b32 %r1510, -1, %r1506, %p749; $L__BB24_572: setp.eq.s32 %p750, %r1510, -1; mov.f32 %f7262, 0f00000000; @%p750 bra $L__BB24_574; shr.s32 %r1181, %r1504, 7; ld.global.u32 %r1182, [%rd302+4]; add.s32 %r1183, %r1182, %r1181; add.s32 %r1184, %r1183, %r1510; mul.wide.s32 %rd1039, %r1184, 4; add.s64 %rd1040, %rd277, %rd1039; mov.f32 %f6423, 0f3F800000; sub.ftz.f32 %f6424, %f6423, %f7252; ld.global.f32 %f6425, [%rd1040]; ld.global.f32 %f6426, [%rd1040+4]; mul.ftz.f32 %f6427, %f7252, %f6426; fma.rn.ftz.f32 %f7262, %f6424, %f6425, %f6427; $L__BB24_574: mul.ftz.f32 %f7321, %f7321, %f7262; setp.lt.ftz.f32 %p752, %f7321, 0f3A83126F; mov.pred %p807, %p735; mov.u16 %rs242, %rs153; @%p752 bra $L__BB24_594; $L__BB24_575: add.s32 %r1528, %r1528, 1; not.pred %p753, %p807; @%p753 bra $L__BB24_589; setp.ge.u32 %p754, %r1536, %r3; setp.geu.ftz.f32 %p755, %f7251, %f7304; and.pred %p756, %p754, %p755; @%p756 bra $L__BB24_588; min.u32 %r1518, %r1536, %r3; add.s32 %r1185, %r1518, 1; setp.lt.u32 %p757, %r1185, %r3; @%p757 bra $L__BB24_587; ld.global.f32 %f7267, [%rd278]; setp.lt.u32 %p758, %r1518, 2; mov.u32 %r1517, 0; @%p758 bra $L__BB24_586; min.u32 %r1190, %r1, %r1536; not.b32 %r1191, %r1190; max.u32 %r1192, %r1191, -5; mov.u32 %r1193, -2; sub.s32 %r623, %r1193, %r1192; mov.u32 %r1194, -3; sub.s32 %r1195, %r1194, %r1192; and.b32 %r624, %r623, 3; setp.lt.u32 %p759, %r1195, 3; mov.u32 %r1517, 0; mov.u32 %r1515, 1; @%p759 bra $L__BB24_582; sub.s32 %r1513, %r623, %r624; mov.u32 %r1517, 0; mov.u32 %r1515, 1; $L__BB24_581: mul.wide.u32 %rd1041, %r1515, 48; mov.u64 %rd1042, kernel_params; add.s64 %rd1043, %rd1042, 3552; add.s64 %rd1044, %rd1043, %rd1041; ld.const.u64 %rd1045, [%rd1044]; cvta.to.global.u64 %rd1046, %rd1045; add.s64 %rd1048, %rd1046, %rd955; ld.global.f32 %f6429, [%rd1048]; setp.gt.ftz.f32 %p760, %f6429, %f7267; selp.f32 %f6430, %f6429, %f7267, %p760; selp.b32 %r1198, %r1515, %r1517, %p760; ld.const.u64 %rd1049, [%rd1044+48]; cvta.to.global.u64 %rd1050, %rd1049; add.s64 %rd1051, %rd1050, %rd955; ld.global.f32 %f6431, [%rd1051]; setp.gt.ftz.f32 %p761, %f6431, %f6430; selp.f32 %f6432, %f6431, %f6430, %p761; add.s32 %r1199, %r1515, 1; selp.b32 %r1200, %r1199, %r1198, %p761; ld.const.u64 %rd1052, [%rd1044+96]; cvta.to.global.u64 %rd1053, %rd1052; add.s64 %rd1054, %rd1053, %rd955; ld.global.f32 %f6433, [%rd1054]; setp.gt.ftz.f32 %p762, %f6433, %f6432; selp.f32 %f6434, %f6433, %f6432, %p762; add.s32 %r1201, %r1515, 2; selp.b32 %r1202, %r1201, %r1200, %p762; add.s32 %r1203, %r1515, 3; mul.wide.u32 %rd1055, %r1203, 48; add.s64 %rd1056, %rd1043, %rd1055; ld.const.u64 %rd1057, [%rd1056]; cvta.to.global.u64 %rd1058, %rd1057; add.s64 %rd1059, %rd1058, %rd955; ld.global.f32 %f6435, [%rd1059]; setp.gt.ftz.f32 %p763, %f6435, %f6434; selp.f32 %f7267, %f6435, %f6434, %p763; selp.b32 %r1517, %r1203, %r1202, %p763; add.s32 %r1515, %r1515, 4; add.s32 %r1513, %r1513, -4; setp.ne.s32 %p764, %r1513, 0; @%p764 bra $L__BB24_581; $L__BB24_582: setp.eq.s32 %p765, %r624, 0; @%p765 bra $L__BB24_586; mul.wide.u32 %rd1060, %r1515, 48; mov.u64 %rd1061, kernel_params; add.s64 %rd1062, %rd1061, %rd1060; ld.const.u64 %rd1063, [%rd1062+3552]; cvta.to.global.u64 %rd1064, %rd1063; add.s64 %rd1066, %rd1064, %rd955; ld.global.f32 %f6436, [%rd1066]; setp.gt.ftz.f32 %p766, %f6436, %f7267; selp.f32 %f7267, %f6436, %f7267, %p766; selp.b32 %r1517, %r1515, %r1517, %p766; setp.eq.s32 %p767, %r624, 1; @%p767 bra $L__BB24_586; add.s32 %r1204, %r1515, 1; mul.wide.u32 %rd1067, %r1204, 48; add.s64 %rd1069, %rd1061, %rd1067; ld.const.u64 %rd1070, [%rd1069+3552]; cvta.to.global.u64 %rd1071, %rd1070; add.s64 %rd1073, %rd1071, %rd955; ld.global.f32 %f6437, [%rd1073]; setp.gt.ftz.f32 %p768, %f6437, %f7267; selp.f32 %f7267, %f6437, %f7267, %p768; selp.b32 %r1517, %r1204, %r1517, %p768; setp.eq.s32 %p769, %r624, 2; @%p769 bra $L__BB24_586; add.s32 %r1205, %r1515, 2; mul.wide.u32 %rd1074, %r1205, 48; add.s64 %rd1076, %rd1061, %rd1074; ld.const.u64 %rd1077, [%rd1076+3552]; cvta.to.global.u64 %rd1078, %rd1077; add.s64 %rd1080, %rd1078, %rd955; ld.global.f32 %f6438, [%rd1080]; setp.gt.ftz.f32 %p770, %f6438, %f7267; selp.f32 %f7267, %f6438, %f7267, %p770; selp.b32 %r1517, %r1205, %r1517, %p770; $L__BB24_586: setp.lt.u32 %p771, %r1536, %r3; selp.b32 %r1518, %r1518, %r1517, %p771; max.ftz.f32 %f7304, %f7251, %f7267; $L__BB24_587: mul.wide.s32 %rd1081, %r1518, 48; mov.u64 %rd1082, kernel_params; add.s64 %rd1083, %rd1082, %rd1081; ld.const.v2.u64 {%rd1084, %rd1085}, [%rd1083+3552]; cvta.to.global.u64 %rd1088, %rd1084; add.s64 %rd1090, %rd1088, %rd955; st.global.f32 [%rd1090], %f7251; cvta.to.global.u64 %rd1091, %rd1085; add.s64 %rd1092, %rd1091, %rd955; st.global.f32 [%rd1092], %f7252; ld.const.v2.u64 {%rd1093, %rd1094}, [%rd1083+3568]; cvta.to.global.u64 %rd1097, %rd1093; add.s64 %rd1098, %rd1097, %rd955; st.global.f32 [%rd1098], %f7253; ld.const.v2.u64 {%rd1099, %rd1100}, [%rd1083+3584]; cvta.to.global.u64 %rd1103, %rd1099; add.s64 %rd1104, %rd1103, %rd955; st.global.u32 [%rd1104], %r1503; cvta.to.global.u64 %rd1105, %rd1094; add.s64 %rd1106, %rd1105, %rd955; st.global.u32 [%rd1106], %r1502; cvta.to.global.u64 %rd1107, %rd1100; add.s64 %rd1108, %rd1107, %rd955; st.global.u32 [%rd1108], %r1504; $L__BB24_588: add.s32 %r1536, %r1536, 1; $L__BB24_589: cvt.u32.u64 %r1206, %rd282; add.s32 %r1475, %r1206, 1; setp.lt.s32 %p772, %r1475, %r1098; @%p772 bra $L__BB24_436; bra.uni $L__BB24_590; $L__BB24_433: mul.wide.s32 %rd965, %r1475, 4; xor.b64 %rd966, %rd965, -4; add.s64 %rd967, %rd260, %rd966; ld.global.u32 %r1529, [%rd967]; mul.wide.s32 %rd968, %r1529, 208; add.s64 %rd969, %rd261, %rd968; ld.global.v4.f32 {%f5016, %f5017, %f5018, %f5019}, [%rd969+48]; ld.global.v4.f32 {%f5024, %f5025, %f5026, %f5027}, [%rd969+64]; ld.global.v4.f32 {%f5032, %f5033, %f5034, %f5035}, [%rd969+80]; mul.ftz.f32 %f5040, %f5017, %f7049; fma.rn.ftz.f32 %f5041, %f6460, %f5016, %f5040; fma.rn.ftz.f32 %f5042, %f5018, %f7048, %f5041; add.ftz.f32 %f7292, %f5019, %f5042; mul.ftz.f32 %f5043, %f5025, %f7049; fma.rn.ftz.f32 %f5044, %f6460, %f5024, %f5043; fma.rn.ftz.f32 %f5045, %f5026, %f7048, %f5044; add.ftz.f32 %f7293, %f5027, %f5045; mul.ftz.f32 %f5046, %f5033, %f7049; fma.rn.ftz.f32 %f5047, %f6460, %f5032, %f5046; fma.rn.ftz.f32 %f5048, %f5034, %f7048, %f5047; add.ftz.f32 %f7294, %f5035, %f5048; mul.ftz.f32 %f5049, %f5017, %f6456; fma.rn.ftz.f32 %f5050, %f5016, %f6457, %f5049; fma.rn.ftz.f32 %f5051, %f5018, %f6455, %f5050; mul.ftz.f32 %f5052, %f5025, %f6456; fma.rn.ftz.f32 %f5053, %f5024, %f6457, %f5052; fma.rn.ftz.f32 %f5054, %f5026, %f6455, %f5053; mul.ftz.f32 %f5055, %f5033, %f6456; fma.rn.ftz.f32 %f5056, %f5032, %f6457, %f5055; fma.rn.ftz.f32 %f5057, %f5034, %f6455, %f5056; abs.ftz.f32 %f5058, %f5051; setp.gt.ftz.f32 %p586, %f5058, 0f17800000; mov.b32 %r1101, %f5051; and.b32 %r1102, %r1101, -2147483648; or.b32 %r1103, %r1102, 394264576; mov.b32 %f5059, %r1103; selp.f32 %f7289, %f5051, %f5059, %p586; abs.ftz.f32 %f5060, %f5054; setp.gt.ftz.f32 %p587, %f5060, 0f17800000; mov.b32 %r1104, %f5054; and.b32 %r1105, %r1104, -2147483648; or.b32 %r1106, %r1105, 394264576; mov.b32 %f5061, %r1106; selp.f32 %f7290, %f5054, %f5061, %p587; abs.ftz.f32 %f5062, %f5057; setp.gt.ftz.f32 %p588, %f5062, 0f17800000; mov.b32 %r1107, %f5057; and.b32 %r1108, %r1107, -2147483648; or.b32 %r1109, %r1108, 394264576; mov.b32 %f5063, %r1109; selp.f32 %f7291, %f5057, %f5063, %p588; rcp.approx.ftz.f32 %f7286, %f7289; rcp.approx.ftz.f32 %f7287, %f7290; rcp.approx.ftz.f32 %f7288, %f7291; add.s32 %r1472, %r1472, 1; mov.u32 %r1110, 1985229328; st.local.u32 [%rd281+4], %r1110; mul.wide.s32 %rd970, %r1529, 4; add.s64 %rd971, %rd262, %rd970; ld.global.u32 %r1471, [%rd971]; bra.uni $L__BB24_590; $L__BB24_420: setp.eq.s32 %p562, %r1471, 1985229328; setp.lt.s32 %p563, %r1471, 0; or.pred %p564, %p563, %p562; @%p564 bra $L__BB24_431; $L__BB24_421: mul.wide.s32 %rd956, %r1471, 16; add.s64 %rd279, %rd258, %rd956; ld.global.v4.f32 {%f4803, %f4804, %f4805, %f4806}, [%rd279]; mov.b32 %r1471, %f4806; mov.b32 %r506, %f4805; mov.b32 %r505, %f4804; mov.b32 %r504, %f4803; mov.b64 %rd957, {%r504, %r505}; cvt.u32.u64 %r1087, %rd957; and.b32 %r1088, %r1087, 1024; setp.eq.s32 %p565, %r1088, 0; and.b32 %r508, %r504, %r668; @%p565 bra $L__BB24_424; bra.uni $L__BB24_422; $L__BB24_424: setp.ne.s32 %p571, %r508, 0; ld.global.v4.f32 {%f4946, %f4947, %f4948, %f4949}, [%rd279+16]; ld.global.v4.f32 {%f4954, %f4955, %f4956, %f4957}, [%rd279+32]; ld.global.v4.f32 {%f4962, %f4963, %f4964, %f4965}, [%rd279+48]; sub.ftz.f32 %f4970, %f4946, %f7292; mul.ftz.f32 %f4971, %f7286, %f4970; sub.ftz.f32 %f4972, %f4948, %f7292; mul.ftz.f32 %f4973, %f7286, %f4972; sub.ftz.f32 %f4974, %f4954, %f7293; mul.ftz.f32 %f4975, %f7287, %f4974; sub.ftz.f32 %f4976, %f4956, %f7293; mul.ftz.f32 %f4977, %f7287, %f4976; sub.ftz.f32 %f4978, %f4962, %f7294; mul.ftz.f32 %f4979, %f7288, %f4978; sub.ftz.f32 %f4980, %f4964, %f7294; mul.ftz.f32 %f4981, %f7288, %f4980; min.ftz.f32 %f4982, %f4971, %f4973; min.ftz.f32 %f4983, %f4975, %f4977; min.ftz.f32 %f4984, %f4979, %f4981; max.ftz.f32 %f4985, %f1291, %f4982; max.ftz.f32 %f4986, %f4983, %f4984; max.ftz.f32 %f7095, %f4985, %f4986; max.ftz.f32 %f4987, %f4971, %f4973; max.ftz.f32 %f4988, %f4975, %f4977; max.ftz.f32 %f4989, %f4979, %f4981; min.ftz.f32 %f4990, %f1292, %f4987; min.ftz.f32 %f4991, %f4988, %f4989; min.ftz.f32 %f4992, %f4990, %f4991; sub.ftz.f32 %f4993, %f4947, %f7292; mul.ftz.f32 %f4994, %f7286, %f4993; sub.ftz.f32 %f4995, %f4949, %f7292; mul.ftz.f32 %f4996, %f7286, %f4995; sub.ftz.f32 %f4997, %f4955, %f7293; mul.ftz.f32 %f4998, %f7287, %f4997; sub.ftz.f32 %f4999, %f4957, %f7293; mul.ftz.f32 %f5000, %f7287, %f4999; sub.ftz.f32 %f5001, %f4963, %f7294; mul.ftz.f32 %f5002, %f7288, %f5001; sub.ftz.f32 %f5003, %f4965, %f7294; mul.ftz.f32 %f5004, %f7288, %f5003; min.ftz.f32 %f5005, %f4994, %f4996; min.ftz.f32 %f5006, %f4998, %f5000; min.ftz.f32 %f5007, %f5002, %f5004; max.ftz.f32 %f5008, %f1291, %f5005; max.ftz.f32 %f5009, %f5006, %f5007; max.ftz.f32 %f7096, %f5008, %f5009; max.ftz.f32 %f5010, %f4994, %f4996; max.ftz.f32 %f5011, %f4998, %f5000; max.ftz.f32 %f5012, %f5002, %f5004; min.ftz.f32 %f5013, %f1292, %f5010; min.ftz.f32 %f5014, %f5011, %f5012; min.ftz.f32 %f5015, %f5013, %f5014; setp.ge.ftz.f32 %p572, %f4992, %f7095; and.pred %p573, %p571, %p572; selp.u32 %r1092, 1, 0, %p573; setp.ltu.ftz.f32 %p574, %f5015, %f7096; and.b32 %r1094, %r505, %r668; setp.eq.s32 %p575, %r1094, 0; or.pred %p576, %p575, %p574; selp.b32 %r1095, 0, 2, %p576; or.b32 %r1470, %r1095, %r1092; bra.uni $L__BB24_425; $L__BB24_422: setp.ne.s32 %p566, %r508, 0; ld.global.v4.f32 {%f4810, %f4811, %f4812, %f4813}, [%rd279+16]; ld.global.v4.f32 {%f4818, %f4819, %f4820, %f4821}, [%rd279+32]; ld.global.v4.f32 {%f4826, %f4827, %f4828, %f4829}, [%rd279+48]; mul.ftz.f32 %f4834, %f7290, %f4811; fma.rn.ftz.f32 %f4835, %f7289, %f4810, %f4834; fma.rn.ftz.f32 %f4836, %f7291, %f4812, %f4835; mul.ftz.f32 %f4837, %f7290, %f4819; fma.rn.ftz.f32 %f4838, %f7289, %f4818, %f4837; fma.rn.ftz.f32 %f4839, %f7291, %f4820, %f4838; mul.ftz.f32 %f4840, %f7290, %f4827; fma.rn.ftz.f32 %f4841, %f7289, %f4826, %f4840; fma.rn.ftz.f32 %f4842, %f7291, %f4828, %f4841; mul.ftz.f32 %f4843, %f7293, %f4811; fma.rn.ftz.f32 %f4844, %f7292, %f4810, %f4843; fma.rn.ftz.f32 %f4845, %f7294, %f4812, %f4844; add.ftz.f32 %f4846, %f4813, %f4845; mul.ftz.f32 %f4847, %f7293, %f4819; fma.rn.ftz.f32 %f4848, %f7292, %f4818, %f4847; fma.rn.ftz.f32 %f4849, %f7294, %f4820, %f4848; add.ftz.f32 %f4850, %f4821, %f4849; mul.ftz.f32 %f4851, %f7293, %f4827; fma.rn.ftz.f32 %f4852, %f7292, %f4826, %f4851; fma.rn.ftz.f32 %f4853, %f7294, %f4828, %f4852; add.ftz.f32 %f4854, %f4829, %f4853; rcp.approx.ftz.f32 %f4855, %f4836; rcp.approx.ftz.f32 %f4856, %f4839; rcp.approx.ftz.f32 %f4857, %f4842; mul.ftz.f32 %f4858, %f4855, %f4846; neg.ftz.f32 %f4859, %f4858; mul.ftz.f32 %f4860, %f4856, %f4850; neg.ftz.f32 %f4861, %f4860; mul.ftz.f32 %f4862, %f4857, %f4854; neg.ftz.f32 %f4863, %f4862; sub.ftz.f32 %f4864, %f4855, %f4858; sub.ftz.f32 %f4865, %f4856, %f4860; sub.ftz.f32 %f4866, %f4857, %f4862; min.ftz.f32 %f4867, %f4859, %f4864; min.ftz.f32 %f4868, %f4861, %f4865; min.ftz.f32 %f4869, %f4863, %f4866; max.ftz.f32 %f4870, %f4859, %f4864; max.ftz.f32 %f4871, %f4861, %f4865; max.ftz.f32 %f4872, %f4863, %f4866; max.ftz.f32 %f4873, %f1291, %f4867; max.ftz.f32 %f4874, %f4868, %f4869; max.ftz.f32 %f7095, %f4873, %f4874; min.ftz.f32 %f4875, %f1292, %f4870; min.ftz.f32 %f4876, %f4871, %f4872; min.ftz.f32 %f4877, %f4875, %f4876; setp.le.ftz.f32 %p567, %f7095, %f4877; and.pred %p568, %p566, %p567; selp.u32 %r1470, 1, 0, %p568; ld.global.v4.f32 {%f4878, %f4879, %f4880, %f4881}, [%rd279+64]; ld.global.v4.f32 {%f4886, %f4887, %f4888, %f4889}, [%rd279+80]; ld.global.v4.f32 {%f4894, %f4895, %f4896, %f4897}, [%rd279+96]; mul.ftz.f32 %f4902, %f7290, %f4879; fma.rn.ftz.f32 %f4903, %f7289, %f4878, %f4902; fma.rn.ftz.f32 %f4904, %f7291, %f4880, %f4903; mul.ftz.f32 %f4905, %f7290, %f4887; fma.rn.ftz.f32 %f4906, %f7289, %f4886, %f4905; fma.rn.ftz.f32 %f4907, %f7291, %f4888, %f4906; mul.ftz.f32 %f4908, %f7290, %f4895; fma.rn.ftz.f32 %f4909, %f7289, %f4894, %f4908; fma.rn.ftz.f32 %f4910, %f7291, %f4896, %f4909; mul.ftz.f32 %f4911, %f7293, %f4879; fma.rn.ftz.f32 %f4912, %f7292, %f4878, %f4911; fma.rn.ftz.f32 %f4913, %f7294, %f4880, %f4912; add.ftz.f32 %f4914, %f4881, %f4913; mul.ftz.f32 %f4915, %f7293, %f4887; fma.rn.ftz.f32 %f4916, %f7292, %f4886, %f4915; fma.rn.ftz.f32 %f4917, %f7294, %f4888, %f4916; add.ftz.f32 %f4918, %f4889, %f4917; mul.ftz.f32 %f4919, %f7293, %f4895; fma.rn.ftz.f32 %f4920, %f7292, %f4894, %f4919; fma.rn.ftz.f32 %f4921, %f7294, %f4896, %f4920; add.ftz.f32 %f4922, %f4897, %f4921; rcp.approx.ftz.f32 %f4923, %f4904; rcp.approx.ftz.f32 %f4924, %f4907; rcp.approx.ftz.f32 %f4925, %f4910; mul.ftz.f32 %f4926, %f4923, %f4914; neg.ftz.f32 %f4927, %f4926; mul.ftz.f32 %f4928, %f4924, %f4918; neg.ftz.f32 %f4929, %f4928; mul.ftz.f32 %f4930, %f4925, %f4922; neg.ftz.f32 %f4931, %f4930; sub.ftz.f32 %f4932, %f4923, %f4926; sub.ftz.f32 %f4933, %f4924, %f4928; sub.ftz.f32 %f4934, %f4925, %f4930; min.ftz.f32 %f4935, %f4927, %f4932; min.ftz.f32 %f4936, %f4929, %f4933; min.ftz.f32 %f4937, %f4931, %f4934; max.ftz.f32 %f4938, %f4927, %f4932; max.ftz.f32 %f4939, %f4929, %f4933; max.ftz.f32 %f4940, %f4931, %f4934; max.ftz.f32 %f4941, %f1291, %f4935; max.ftz.f32 %f4942, %f4936, %f4937; max.ftz.f32 %f7096, %f4941, %f4942; min.ftz.f32 %f4943, %f1292, %f4938; min.ftz.f32 %f4944, %f4939, %f4940; min.ftz.f32 %f4945, %f4943, %f4944; setp.gtu.ftz.f32 %p569, %f7096, %f4945; @%p569 bra $L__BB24_425; and.b32 %r1090, %r505, %r668; setp.eq.s32 %p570, %r1090, 0; or.b32 %r1091, %r1470, 2; selp.b32 %r1470, %r1470, %r1091, %p570; $L__BB24_425: mul.wide.s32 %rd958, %r1472, 4; add.s64 %rd280, %rd6, %rd958; setp.eq.s32 %p577, %r1470, 3; @%p577 bra $L__BB24_429; bra.uni $L__BB24_426; $L__BB24_429: setp.lt.ftz.f32 %p580, %f7096, %f7095; selp.b32 %r1096, %r506, %r1471, %p580; selp.b32 %r1471, %r1471, %r506, %p580; add.s32 %r1472, %r1472, 1; st.local.u32 [%rd280+4], %r1096; bra.uni $L__BB24_430; $L__BB24_426: setp.eq.s32 %p578, %r1470, 2; @%p578 bra $L__BB24_430; setp.ne.s32 %p579, %r1470, 0; mov.u32 %r1471, %r506; @%p579 bra $L__BB24_430; ld.local.u32 %r1471, [%rd280]; add.s32 %r1472, %r1472, -1; $L__BB24_430: setp.ne.s32 %p581, %r1471, 1985229328; setp.gt.s32 %p582, %r1471, -1; and.pred %p583, %p582, %p581; @%p583 bra $L__BB24_421; $L__BB24_431: setp.gt.s32 %p584, %r1471, -1; @%p584 bra $L__BB24_590; bra.uni $L__BB24_432; $L__BB24_590: setp.ne.s32 %p773, %r1471, 1985229328; @%p773 bra $L__BB24_420; setp.lt.s32 %p774, %r1472, 0; mov.u32 %r1471, %r1084; @%p774 bra $L__BB24_593; mul.wide.s32 %rd1111, %r1472, 4; add.s64 %rd1112, %rd6, %rd1111; ld.local.u32 %r1471, [%rd1112]; add.s32 %r1472, %r1472, -1; mov.u32 %r1529, -1; mov.f32 %f7286, %f8; mov.f32 %f7287, %f9; mov.f32 %f7288, %f10; mov.f32 %f7289, %f5; mov.f32 %f7290, %f6; mov.f32 %f7291, %f7; mov.f32 %f7292, %f6460; mov.f32 %f7293, %f7049; mov.f32 %f7294, %f7048; $L__BB24_593: setp.ne.s32 %p775, %r1471, 1985229328; mov.u16 %rs242, 0; @%p775 bra $L__BB24_420; bra.uni $L__BB24_594; $L__BB24_7: mov.u32 %r700, 1985229328; st.local.u32 [%rd6], %r700; ld.const.u32 %r1240, [kernel_params+1104]; ld.local.f32 %f6459, [%rd4+4]; ld.local.f32 %f6458, [%rd4+8]; ld.local.f32 %f13, [%rd4+24]; ld.local.f32 %f14, [%rd4+28]; ld.const.v2.u64 {%rd322, %rd323}, [kernel_params+2096]; mov.u32 %r1536, 0; cvta.to.global.u64 %rd9, %rd322; cvta.to.global.u64 %rd10, %rd323; ld.const.v2.u64 {%rd326, %rd327}, [kernel_params+2128]; cvta.to.global.u64 %rd11, %rd327; ld.const.v2.u64 {%rd330, %rd331}, [kernel_params+2176]; cvta.to.global.u64 %rd12, %rd331; ld.const.u64 %rd334, [kernel_params+2160]; cvta.to.global.u64 %rd13, %rd334; cvta.to.global.u64 %rd14, %rd326; ld.const.u64 %rd335, [kernel_params+2232]; cvta.to.global.u64 %rd15, %rd335; ld.const.v2.u64 {%rd336, %rd337}, [kernel_params+2256]; cvta.to.global.u64 %rd16, %rd336; ld.const.u32 %r5, [kernel_params+1128]; ld.const.v2.u64 {%rd340, %rd341}, [kernel_params+2304]; cvta.to.global.u64 %rd17, %rd341; ld.const.v2.u64 {%rd344, %rd345}, [kernel_params+2112]; cvta.to.global.u64 %rd18, %rd344; ld.const.v2.u64 {%rd348, %rd349}, [kernel_params+2144]; cvta.to.global.u64 %rd19, %rd349; ld.const.u64 %rd352, [kernel_params+2408]; cvta.to.global.u64 %rd20, %rd352; ld.const.u64 %rd353, [kernel_params+2216]; cvta.to.global.u64 %rd21, %rd353; cvta.to.global.u64 %rd22, %rd345; ld.const.u64 %rd354, [kernel_params+2296]; cvta.to.global.u64 %rd23, %rd354; ld.const.v2.u64 {%rd355, %rd356}, [kernel_params+2336]; cvta.to.global.u64 %rd24, %rd355; cvta.to.global.u64 %rd25, %rd340; cvta.to.global.u64 %rd26, %rd337; cvta.to.global.u64 %rd27, %rd356; ld.const.u64 %rd359, [kernel_params+2320]; cvta.to.global.u64 %rd28, %rd359; ld.const.u64 %rd360, [kernel_params+3552]; cvta.to.global.u64 %rd361, %rd360; shl.b64 %rd362, %rd3, 2; add.s64 %rd29, %rd361, %rd362; cvta.to.global.u64 %rd30, %rd348; cvta.to.global.u64 %rd31, %rd330; mov.u32 %r1288, -1; mov.f32 %f7321, 0f3F800000; setp.eq.s32 %p95, %r5, 0; mov.u16 %rs103, 1; mov.pred %p103, 0; mov.pred %p150, -1; mov.f32 %f6579, %f8; mov.f32 %f6580, %f9; mov.f32 %f6581, %f10; mov.f32 %f6582, %f5; mov.f32 %f6583, %f6; mov.f32 %f6584, %f7; mov.f32 %f6585, %f6460; mov.f32 %f6586, %f6459; mov.f32 %f6587, %f6458; mov.f32 %f6595, %f14; mov.u32 %r1287, %r1536; mov.u32 %r1241, %r1536; bra.uni $L__BB24_8; $L__BB24_16: mul.wide.s32 %rd366, %r1240, 16; xor.b64 %rd367, %rd366, -16; add.s64 %rd368, %rd10, %rd367; ld.global.v4.u32 {%r1244, %r713, %r714, %r715}, [%rd368]; setp.gt.s32 %p84, %r1244, -1; mul.wide.s32 %rd371, %r1241, 4; add.s64 %rd33, %rd6, %rd371; @%p84 bra $L__BB24_26; bra.uni $L__BB24_17; $L__BB24_26: ld.local.u32 %r1240, [%rd33]; add.s32 %r1241, %r1241, -1; setp.le.s32 %p94, %r713, %r1244; @%p94 bra $L__BB24_121; cvt.u16.u32 %rs102, %r715; and.b16 %rs1, %rs102, 127; and.b32 %r731, %r715, 64; setp.eq.s32 %p96, %r731, 0; or.pred %p1, %p96, %p95; mul.ftz.f32 %f2273, %f6583, %f6583; fma.rn.ftz.f32 %f2274, %f6582, %f6582, %f2273; fma.rn.ftz.f32 %f2275, %f6584, %f6584, %f2274; rcp.approx.ftz.f32 %f164, %f2275; $L__BB24_28: cvt.s64.s32 %rd37, %r1244; setp.ne.s32 %p97, %r1288, -1; mov.u32 %r53, %r1288; @%p97 bra $L__BB24_30; shl.b64 %rd385, %rd37, 2; add.s64 %rd386, %rd11, %rd385; ld.global.u32 %r53, [%rd386]; $L__BB24_30: shl.b64 %rd387, %rd37, 2; add.s64 %rd388, %rd14, %rd387; ld.global.u32 %r732, [%rd388]; cvt.s64.s32 %rd38, %r732; ld.local.u32 %r733, [%rd4+36]; setp.ne.s32 %p98, %r733, %r732; @%p98 bra $L__BB24_32; ld.local.u32 %r734, [%rd4+40]; setp.eq.s32 %p99, %r734, %r53; mov.u16 %rs190, %rs103; @%p99 bra $L__BB24_34; $L__BB24_32: cvt.u32.u64 %r735, %rd38; ld.local.u32 %r736, [%rd4+44]; setp.ne.s32 %p100, %r736, %r735; mov.u16 %rs190, 0; @%p100 bra $L__BB24_34; ld.local.u32 %r737, [%rd4+48]; setp.eq.s32 %p101, %r737, %r53; selp.u16 %rs190, 1, 0, %p101; $L__BB24_34: setp.ne.s16 %p102, %rs190, 0; @%p102 bra $L__BB24_120; mul.wide.s32 %rd389, %r53, 208; add.s64 %rd390, %rd13, %rd389; add.s64 %rd39, %rd390, 148; shl.b64 %rd391, %rd38, 4; add.s64 %rd40, %rd15, %rd391; add.s64 %rd41, %rd22, %rd387; setp.gt.s16 %p104, %rs1, 64; @%p104 bra $L__BB24_44; setp.eq.s16 %p107, %rs1, 1; @%p107 bra $L__BB24_80; setp.eq.s16 %p108, %rs1, 8; mov.pred %p787, %p103; @%p108 bra $L__BB24_38; bra.uni $L__BB24_88; $L__BB24_44: setp.eq.s16 %p105, %rs1, 72; @%p105 bra $L__BB24_38; setp.ne.s16 %p106, %rs1, 65; mov.pred %p787, %p103; @%p106 bra $L__BB24_88; ld.global.u32 %r54, [%rd39+8]; ld.global.u32 %r55, [%rd39+4]; shl.b32 %r738, %r55, 1; cvt.rn.f32.s32 %f2276, %r738; ld.local.f32 %f2277, [%rd4+32]; mul.ftz.f32 %f176, %f2277, %f2276; cvt.rzi.ftz.s32.f32 %r739, %f176; add.s32 %r740, %r738, -1; min.s32 %r56, %r739, %r740; ld.global.u32 %r1252, [%rd39+16]; mul.wide.u32 %rd393, %r1252, 16; add.s64 %rd394, %rd17, %rd393; ld.global.u16 %rs191, [%rd394+12]; ld.global.u32 %r1251, [%rd394+8]; ld.global.u64 %rd1174, [%rd394]; setp.eq.s64 %p109, %rd1174, 11; @%p109 bra $L__BB24_51; $L__BB24_47: setp.eq.s64 %p110, %rd1174, 0; @%p110 bra $L__BB24_49; add.s32 %r1252, %r1252, 2; bra.uni $L__BB24_50; $L__BB24_49: setp.eq.s16 %p111, %rs191, 0; mov.u32 %r1255, -1; mov.u32 %r1252, %r1251; @%p111 bra $L__BB24_52; $L__BB24_50: mul.wide.u32 %rd395, %r1252, 16; add.s64 %rd396, %rd17, %rd395; ld.global.u16 %rs191, [%rd396+12]; ld.global.u32 %r1251, [%rd396+8]; ld.global.u64 %rd1174, [%rd396]; setp.ne.s64 %p112, %rd1174, 11; @%p112 bra $L__BB24_47; $L__BB24_51: setp.eq.s16 %p113, %rs191, 0; selp.b32 %r1255, -1, %r1251, %p113; $L__BB24_52: ld.global.v4.u32 {%r742, %r743, %r744, %r745}, [%rd40]; setp.eq.s32 %p114, %r56, %r55; @%p114 bra $L__BB24_54; bra.uni $L__BB24_53; $L__BB24_54: mul.wide.u32 %rd400, %r745, 12; add.s64 %rd1175, %rd16, %rd400; add.s32 %r752, %r745, 1; mul.wide.u32 %rd401, %r752, 12; add.s64 %rd1178, %rd16, %rd401; add.s32 %r753, %r745, 2; mul.wide.u32 %rd402, %r753, 12; add.s64 %rd1181, %rd16, %rd402; add.s64 %rd1176, %rd1175, 4; add.s64 %rd1177, %rd1175, 8; add.s64 %rd1179, %rd1178, 4; add.s64 %rd1180, %rd1178, 8; add.s64 %rd1182, %rd1181, 4; add.s64 %rd1183, %rd1181, 8; bra.uni $L__BB24_55; $L__BB24_38: @%p1 bra $L__BB24_40; shl.b64 %rd409, %rd37, 3; add.s64 %rd410, %rd19, %rd409; ld.global.v2.f32 {%f2386, %f2387}, [%rd410]; ld.local.f32 %f2390, [%rd4+32]; setp.lt.ftz.f32 %p133, %f2390, %f2386; setp.gt.ftz.f32 %p134, %f2390, %f2387; or.pred %p135, %p133, %p134; mov.pred %p787, %p103; @%p135 bra $L__BB24_88; $L__BB24_40: add.s64 %rd412, %rd18, %rd387; ld.global.u32 %r75, [%rd412]; and.b32 %r765, %r75, 64; setp.eq.s32 %p136, %r765, 0; @%p136 bra $L__BB24_76; bra.uni $L__BB24_41; $L__BB24_76: add.s64 %rd422, %rd23, %rd391; ld.global.v4.f32 {%f6539, %f6540, %f6541, %f6542}, [%rd422]; bra.uni $L__BB24_77; $L__BB24_80: ld.global.u32 %r781, [%rd40+12]; mul.wide.u32 %rd423, %r781, 12; add.s64 %rd424, %rd16, %rd423; add.s32 %r782, %r781, 1; mul.wide.u32 %rd425, %r782, 12; add.s64 %rd426, %rd16, %rd425; add.s32 %r783, %r781, 2; mul.wide.u32 %rd427, %r783, 12; add.s64 %rd428, %rd16, %rd427; ld.global.f32 %f2435, [%rd424]; sub.ftz.f32 %f238, %f2435, %f6585; ld.global.f32 %f2436, [%rd424+4]; sub.ftz.f32 %f239, %f2436, %f6586; ld.global.f32 %f2437, [%rd424+8]; sub.ftz.f32 %f240, %f2437, %f6587; ld.global.f32 %f2438, [%rd426]; sub.ftz.f32 %f2439, %f2438, %f6585; ld.global.f32 %f2440, [%rd426+4]; sub.ftz.f32 %f2441, %f2440, %f6586; ld.global.f32 %f2442, [%rd426+8]; sub.ftz.f32 %f2443, %f2442, %f6587; ld.global.f32 %f2444, [%rd428]; sub.ftz.f32 %f2445, %f2444, %f6585; ld.global.f32 %f2446, [%rd428+4]; sub.ftz.f32 %f2447, %f2446, %f6586; ld.global.f32 %f2448, [%rd428+8]; sub.ftz.f32 %f2449, %f2448, %f6587; sub.ftz.f32 %f241, %f2445, %f238; sub.ftz.f32 %f242, %f2447, %f239; sub.ftz.f32 %f243, %f2449, %f240; sub.ftz.f32 %f244, %f238, %f2439; sub.ftz.f32 %f245, %f239, %f2441; sub.ftz.f32 %f246, %f240, %f2443; sub.ftz.f32 %f2450, %f2439, %f2445; sub.ftz.f32 %f2451, %f2441, %f2447; sub.ftz.f32 %f2452, %f2443, %f2449; add.ftz.f32 %f2453, %f238, %f2445; add.ftz.f32 %f2454, %f239, %f2447; add.ftz.f32 %f2455, %f240, %f2449; mul.ftz.f32 %f2456, %f242, %f2455; mul.ftz.f32 %f2457, %f2454, %f243; sub.ftz.f32 %f2458, %f2456, %f2457; mul.ftz.f32 %f2459, %f2453, %f243; mul.ftz.f32 %f2460, %f241, %f2455; sub.ftz.f32 %f2461, %f2459, %f2460; mul.ftz.f32 %f2462, %f241, %f2454; mul.ftz.f32 %f2463, %f2453, %f242; sub.ftz.f32 %f2464, %f2462, %f2463; mul.ftz.f32 %f2465, %f6583, %f2461; fma.rn.ftz.f32 %f2466, %f6582, %f2458, %f2465; fma.rn.ftz.f32 %f247, %f6584, %f2464, %f2466; add.ftz.f32 %f2467, %f238, %f2439; add.ftz.f32 %f2468, %f239, %f2441; add.ftz.f32 %f2469, %f240, %f2443; mul.ftz.f32 %f2470, %f245, %f2469; mul.ftz.f32 %f2471, %f2468, %f246; sub.ftz.f32 %f2472, %f2470, %f2471; mul.ftz.f32 %f2473, %f2467, %f246; mul.ftz.f32 %f2474, %f244, %f2469; sub.ftz.f32 %f2475, %f2473, %f2474; mul.ftz.f32 %f2476, %f244, %f2468; mul.ftz.f32 %f2477, %f2467, %f245; sub.ftz.f32 %f2478, %f2476, %f2477; mul.ftz.f32 %f2479, %f6583, %f2475; fma.rn.ftz.f32 %f2480, %f6582, %f2472, %f2479; fma.rn.ftz.f32 %f248, %f6584, %f2478, %f2480; add.ftz.f32 %f2481, %f2439, %f2445; add.ftz.f32 %f2482, %f2441, %f2447; add.ftz.f32 %f2483, %f2443, %f2449; mul.ftz.f32 %f2484, %f2451, %f2483; mul.ftz.f32 %f2485, %f2482, %f2452; sub.ftz.f32 %f2486, %f2484, %f2485; mul.ftz.f32 %f2487, %f2481, %f2452; mul.ftz.f32 %f2488, %f2450, %f2483; sub.ftz.f32 %f2489, %f2487, %f2488; mul.ftz.f32 %f2490, %f2450, %f2482; mul.ftz.f32 %f2491, %f2481, %f2451; sub.ftz.f32 %f2492, %f2490, %f2491; mul.ftz.f32 %f2493, %f6583, %f2489; fma.rn.ftz.f32 %f2494, %f6582, %f2486, %f2493; fma.rn.ftz.f32 %f2495, %f6584, %f2492, %f2494; add.ftz.f32 %f2496, %f248, %f247; add.ftz.f32 %f249, %f2495, %f2496; abs.ftz.f32 %f250, %f249; mul.ftz.f32 %f2497, %f250, 0f34000000; min.ftz.f32 %f2498, %f248, %f2495; min.ftz.f32 %f2499, %f247, %f2498; max.ftz.f32 %f2500, %f248, %f2495; max.ftz.f32 %f2501, %f247, %f2500; mul.ftz.f32 %f2502, %f250, 0fB4000000; setp.ltu.ftz.f32 %p151, %f2499, %f2502; setp.gtu.ftz.f32 %p152, %f2501, %f2497; and.pred %p153, %p151, %p152; mov.pred %p786, %p150; @%p153 bra $L__BB24_86; mul.ftz.f32 %f2503, %f245, %f243; mul.ftz.f32 %f2504, %f246, %f242; sub.ftz.f32 %f2505, %f2503, %f2504; mul.ftz.f32 %f2506, %f244, %f243; mul.ftz.f32 %f2507, %f246, %f241; sub.ftz.f32 %f2508, %f2507, %f2506; mul.ftz.f32 %f2509, %f245, %f241; mul.ftz.f32 %f2510, %f244, %f242; sub.ftz.f32 %f2511, %f2510, %f2509; add.ftz.f32 %f251, %f2505, %f2505; add.ftz.f32 %f252, %f2508, %f2508; add.ftz.f32 %f253, %f2511, %f2511; mul.ftz.f32 %f2512, %f6583, %f252; fma.rn.ftz.f32 %f2513, %f6582, %f251, %f2512; fma.rn.ftz.f32 %f254, %f6584, %f253, %f2513; setp.eq.ftz.f32 %p155, %f254, 0f00000000; mov.pred %p786, %p150; @%p155 bra $L__BB24_86; mul.ftz.f32 %f2514, %f239, %f252; fma.rn.ftz.f32 %f2515, %f238, %f251, %f2514; fma.rn.ftz.f32 %f2516, %f240, %f253, %f2515; div.approx.ftz.f32 %f255, %f2516, %f254; setp.ltu.ftz.f32 %p157, %f255, %f13; setp.gtu.ftz.f32 %p158, %f255, %f14; or.pred %p159, %p158, %p157; mov.pred %p786, %p150; @%p159 bra $L__BB24_86; setp.lt.ftz.f32 %p160, %f250, 0f219392EF; mov.f32 %f6544, 0f00000000; @%p160 bra $L__BB24_85; rcp.approx.ftz.f32 %f6544, %f249; $L__BB24_85: mul.ftz.f32 %f2518, %f247, %f6544; mov.f32 %f2519, 0f3F800000; min.ftz.f32 %f6552, %f2518, %f2519; mul.ftz.f32 %f2520, %f248, %f6544; min.ftz.f32 %f6551, %f2520, %f2519; mov.pred %p786, 0; mov.f32 %f6553, %f255; $L__BB24_86: mov.pred %p787, %p103; @%p786 bra $L__BB24_88; cvt.u32.u64 %r784, %rd38; ld.global.u32 %r785, [%rd41]; and.b32 %r786, %r785, %r668; setp.eq.s32 %p163, %r786, 0; selp.f32 %f6548, %f6548, %f6553, %p163; selp.f32 %f6549, %f6549, %f6552, %p163; selp.f32 %f6550, %f6550, %f6551, %p163; selp.b32 %r1261, %r1261, %r784, %p163; selp.b32 %r1262, %r1262, %r53, %p163; selp.b32 %r1263, %r1263, 1, %p163; setp.ne.s32 %p787, %r786, 0; bra.uni $L__BB24_88; $L__BB24_41: ld.global.u32 %r76, [%rd39]; ld.global.u32 %r77, [%rd39+4]; shl.b32 %r766, %r77, 1; cvt.rn.f32.s32 %f2391, %r766; ld.local.f32 %f2392, [%rd4+32]; mul.ftz.f32 %f214, %f2392, %f2391; cvt.rzi.ftz.s32.f32 %r767, %f214; add.s32 %r768, %r766, -1; min.s32 %r78, %r767, %r768; ld.global.u32 %r1257, [%rd39+16]; mul.wide.u32 %rd413, %r1257, 16; add.s64 %rd414, %rd17, %rd413; ld.global.u16 %rs193, [%rd414+12]; ld.global.u32 %r1256, [%rd414+8]; ld.global.u64 %rd1193, [%rd414]; setp.eq.s64 %p137, %rd1193, 11; @%p137 bra $L__BB24_68; $L__BB24_42: setp.eq.s64 %p138, %rd1193, 0; @%p138 bra $L__BB24_66; add.s32 %r1257, %r1257, 2; bra.uni $L__BB24_67; $L__BB24_66: setp.eq.s16 %p139, %rs193, 0; mov.u32 %r1260, -1; mov.u32 %r1257, %r1256; @%p139 bra $L__BB24_69; $L__BB24_67: mul.wide.u32 %rd415, %r1257, 16; add.s64 %rd416, %rd17, %rd415; ld.global.u16 %rs193, [%rd416+12]; ld.global.u32 %r1256, [%rd416+8]; ld.global.u64 %rd1193, [%rd416]; setp.ne.s64 %p140, %rd1193, 11; @%p140 bra $L__BB24_42; $L__BB24_68: setp.eq.s16 %p141, %rs193, 0; selp.b32 %r1260, -1, %r1256, %p141; $L__BB24_69: setp.eq.s32 %p142, %r78, %r77; @%p142 bra $L__BB24_71; bra.uni $L__BB24_70; $L__BB24_71: add.s64 %rd1194, %rd23, %rd391; bra.uni $L__BB24_72; $L__BB24_53: setp.gt.s32 %p115, %r56, %r55; selp.b32 %r746, -1, 0, %p115; add.s32 %r747, %r56, %r746; mad.lo.s32 %r748, %r747, %r54, %r1255; add.s32 %r749, %r742, %r748; mul.wide.u32 %rd397, %r749, 12; add.s64 %rd1175, %rd24, %rd397; add.s32 %r750, %r743, %r748; mul.wide.u32 %rd398, %r750, 12; add.s64 %rd1178, %rd24, %rd398; add.s32 %r751, %r744, %r748; mul.wide.u32 %rd399, %r751, 12; add.s64 %rd1181, %rd24, %rd399; add.s64 %rd1176, %rd1175, 4; add.s64 %rd1177, %rd1175, 8; add.s64 %rd1179, %rd1178, 4; add.s64 %rd1180, %rd1178, 8; add.s64 %rd1182, %rd1181, 4; add.s64 %rd1183, %rd1181, 8; $L__BB24_55: ld.global.f32 %f177, [%rd1183]; ld.global.f32 %f178, [%rd1182]; ld.global.f32 %f179, [%rd1181]; ld.global.f32 %f180, [%rd1180]; ld.global.f32 %f181, [%rd1179]; ld.global.f32 %f182, [%rd1178]; ld.global.f32 %f183, [%rd1177]; ld.global.f32 %f184, [%rd1176]; ld.global.f32 %f185, [%rd1175]; add.s32 %r71, %r56, 1; setp.eq.s32 %p116, %r71, %r55; @%p116 bra $L__BB24_57; bra.uni $L__BB24_56; $L__BB24_57: mul.wide.u32 %rd406, %r745, 12; add.s64 %rd1184, %rd16, %rd406; add.s32 %r760, %r745, 1; mul.wide.u32 %rd407, %r760, 12; add.s64 %rd1187, %rd16, %rd407; add.s32 %r761, %r745, 2; mul.wide.u32 %rd408, %r761, 12; add.s64 %rd1190, %rd16, %rd408; add.s64 %rd1185, %rd1184, 4; add.s64 %rd1186, %rd1184, 8; add.s64 %rd1188, %rd1187, 4; add.s64 %rd1189, %rd1187, 8; add.s64 %rd1191, %rd1190, 4; add.s64 %rd1192, %rd1190, 8; bra.uni $L__BB24_58; $L__BB24_56: setp.ge.s32 %p117, %r56, %r55; selp.b32 %r754, -1, 0, %p117; add.s32 %r755, %r71, %r754; mad.lo.s32 %r756, %r755, %r54, %r1255; add.s32 %r757, %r742, %r756; mul.wide.u32 %rd403, %r757, 12; add.s64 %rd1184, %rd24, %rd403; add.s32 %r758, %r743, %r756; mul.wide.u32 %rd404, %r758, 12; add.s64 %rd1187, %rd24, %rd404; add.s32 %r759, %r744, %r756; mul.wide.u32 %rd405, %r759, 12; add.s64 %rd1190, %rd24, %rd405; add.s64 %rd1185, %rd1184, 4; add.s64 %rd1186, %rd1184, 8; add.s64 %rd1188, %rd1187, 4; add.s64 %rd1189, %rd1187, 8; add.s64 %rd1191, %rd1190, 4; add.s64 %rd1192, %rd1190, 8; $L__BB24_58: cvt.rn.f32.s32 %f2278, %r56; sub.ftz.f32 %f2279, %f176, %f2278; mov.f32 %f2280, 0f3F800000; sub.ftz.f32 %f2281, %f2280, %f2279; ld.global.f32 %f2282, [%rd1184]; mul.ftz.f32 %f2283, %f2279, %f2282; ld.global.f32 %f2284, [%rd1185]; mul.ftz.f32 %f2285, %f2279, %f2284; ld.global.f32 %f2286, [%rd1186]; mul.ftz.f32 %f2287, %f2279, %f2286; fma.rn.ftz.f32 %f2288, %f2281, %f185, %f2283; fma.rn.ftz.f32 %f2289, %f2281, %f184, %f2285; fma.rn.ftz.f32 %f2290, %f2281, %f183, %f2287; ld.global.f32 %f2291, [%rd1187]; mul.ftz.f32 %f2292, %f2279, %f2291; ld.global.f32 %f2293, [%rd1188]; mul.ftz.f32 %f2294, %f2279, %f2293; ld.global.f32 %f2295, [%rd1189]; mul.ftz.f32 %f2296, %f2279, %f2295; fma.rn.ftz.f32 %f2297, %f2281, %f182, %f2292; fma.rn.ftz.f32 %f2298, %f2281, %f181, %f2294; fma.rn.ftz.f32 %f2299, %f2281, %f180, %f2296; ld.global.f32 %f2300, [%rd1190]; mul.ftz.f32 %f2301, %f2279, %f2300; ld.global.f32 %f2302, [%rd1191]; mul.ftz.f32 %f2303, %f2279, %f2302; ld.global.f32 %f2304, [%rd1192]; mul.ftz.f32 %f2305, %f2279, %f2304; fma.rn.ftz.f32 %f2306, %f2281, %f179, %f2301; fma.rn.ftz.f32 %f2307, %f2281, %f178, %f2303; fma.rn.ftz.f32 %f2308, %f2281, %f177, %f2305; sub.ftz.f32 %f186, %f2288, %f6585; sub.ftz.f32 %f187, %f2289, %f6586; sub.ftz.f32 %f188, %f2290, %f6587; sub.ftz.f32 %f2309, %f2297, %f6585; sub.ftz.f32 %f2310, %f2298, %f6586; sub.ftz.f32 %f2311, %f2299, %f6587; sub.ftz.f32 %f2312, %f2306, %f6585; sub.ftz.f32 %f2313, %f2307, %f6586; sub.ftz.f32 %f2314, %f2308, %f6587; sub.ftz.f32 %f189, %f2312, %f186; sub.ftz.f32 %f190, %f2313, %f187; sub.ftz.f32 %f191, %f2314, %f188; sub.ftz.f32 %f192, %f186, %f2309; sub.ftz.f32 %f193, %f187, %f2310; sub.ftz.f32 %f194, %f188, %f2311; sub.ftz.f32 %f2315, %f2309, %f2312; sub.ftz.f32 %f2316, %f2310, %f2313; sub.ftz.f32 %f2317, %f2311, %f2314; add.ftz.f32 %f2318, %f2312, %f186; add.ftz.f32 %f2319, %f2313, %f187; add.ftz.f32 %f2320, %f2314, %f188; mul.ftz.f32 %f2321, %f2320, %f190; mul.ftz.f32 %f2322, %f191, %f2319; sub.ftz.f32 %f2323, %f2321, %f2322; mul.ftz.f32 %f2324, %f191, %f2318; mul.ftz.f32 %f2325, %f2320, %f189; sub.ftz.f32 %f2326, %f2324, %f2325; mul.ftz.f32 %f2327, %f2319, %f189; mul.ftz.f32 %f2328, %f190, %f2318; sub.ftz.f32 %f2329, %f2327, %f2328; mul.ftz.f32 %f2330, %f6583, %f2326; fma.rn.ftz.f32 %f2331, %f6582, %f2323, %f2330; fma.rn.ftz.f32 %f195, %f6584, %f2329, %f2331; add.ftz.f32 %f2332, %f2309, %f186; add.ftz.f32 %f2333, %f2310, %f187; add.ftz.f32 %f2334, %f2311, %f188; mul.ftz.f32 %f2335, %f2334, %f193; mul.ftz.f32 %f2336, %f194, %f2333; sub.ftz.f32 %f2337, %f2335, %f2336; mul.ftz.f32 %f2338, %f194, %f2332; mul.ftz.f32 %f2339, %f2334, %f192; sub.ftz.f32 %f2340, %f2338, %f2339; mul.ftz.f32 %f2341, %f2333, %f192; mul.ftz.f32 %f2342, %f193, %f2332; sub.ftz.f32 %f2343, %f2341, %f2342; mul.ftz.f32 %f2344, %f6583, %f2340; fma.rn.ftz.f32 %f2345, %f6582, %f2337, %f2344; fma.rn.ftz.f32 %f196, %f6584, %f2343, %f2345; add.ftz.f32 %f2346, %f2312, %f2309; add.ftz.f32 %f2347, %f2313, %f2310; add.ftz.f32 %f2348, %f2314, %f2311; mul.ftz.f32 %f2349, %f2348, %f2316; mul.ftz.f32 %f2350, %f2317, %f2347; sub.ftz.f32 %f2351, %f2349, %f2350; mul.ftz.f32 %f2352, %f2317, %f2346; mul.ftz.f32 %f2353, %f2348, %f2315; sub.ftz.f32 %f2354, %f2352, %f2353; mul.ftz.f32 %f2355, %f2347, %f2315; mul.ftz.f32 %f2356, %f2316, %f2346; sub.ftz.f32 %f2357, %f2355, %f2356; mul.ftz.f32 %f2358, %f6583, %f2354; fma.rn.ftz.f32 %f2359, %f6582, %f2351, %f2358; fma.rn.ftz.f32 %f2360, %f6584, %f2357, %f2359; add.ftz.f32 %f2361, %f195, %f196; add.ftz.f32 %f197, %f2360, %f2361; abs.ftz.f32 %f198, %f197; mul.ftz.f32 %f2362, %f198, 0f34000000; min.ftz.f32 %f2363, %f196, %f2360; min.ftz.f32 %f2364, %f195, %f2363; max.ftz.f32 %f2365, %f196, %f2360; max.ftz.f32 %f2366, %f195, %f2365; mul.ftz.f32 %f2367, %f198, 0fB4000000; setp.ltu.ftz.f32 %p119, %f2364, %f2367; setp.gtu.ftz.f32 %p120, %f2366, %f2362; and.pred %p121, %p119, %p120; mov.pred %p784, -1; @%p121 bra $L__BB24_64; mul.ftz.f32 %f2368, %f191, %f193; mul.ftz.f32 %f2369, %f194, %f190; sub.ftz.f32 %f2370, %f2368, %f2369; mul.ftz.f32 %f2371, %f191, %f192; mul.ftz.f32 %f2372, %f194, %f189; sub.ftz.f32 %f2373, %f2372, %f2371; mul.ftz.f32 %f2374, %f193, %f189; mul.ftz.f32 %f2375, %f190, %f192; sub.ftz.f32 %f2376, %f2375, %f2374; add.ftz.f32 %f199, %f2370, %f2370; add.ftz.f32 %f200, %f2373, %f2373; add.ftz.f32 %f201, %f2376, %f2376; mul.ftz.f32 %f2377, %f6583, %f200; fma.rn.ftz.f32 %f2378, %f6582, %f199, %f2377; fma.rn.ftz.f32 %f202, %f6584, %f201, %f2378; setp.eq.ftz.f32 %p123, %f202, 0f00000000; @%p123 bra $L__BB24_64; mul.ftz.f32 %f2379, %f187, %f200; fma.rn.ftz.f32 %f2380, %f186, %f199, %f2379; fma.rn.ftz.f32 %f2381, %f188, %f201, %f2380; div.approx.ftz.f32 %f203, %f2381, %f202; setp.ltu.ftz.f32 %p125, %f203, %f13; setp.gtu.ftz.f32 %p126, %f203, %f14; or.pred %p127, %p126, %p125; @%p127 bra $L__BB24_64; setp.lt.ftz.f32 %p128, %f198, 0f219392EF; mov.f32 %f6535, 0f00000000; @%p128 bra $L__BB24_63; rcp.approx.ftz.f32 %f6535, %f197; $L__BB24_63: mul.ftz.f32 %f2383, %f195, %f6535; min.ftz.f32 %f6555, %f2383, %f2280; mul.ftz.f32 %f2385, %f196, %f6535; min.ftz.f32 %f6554, %f2385, %f2280; mov.pred %p784, 0; mov.f32 %f6556, %f203; $L__BB24_64: mov.pred %p787, %p103; @%p784 bra $L__BB24_88; cvt.u32.u64 %r762, %rd38; ld.global.u32 %r763, [%rd41]; and.b32 %r764, %r763, %r668; setp.eq.s32 %p131, %r764, 0; selp.f32 %f6548, %f6548, %f6556, %p131; selp.f32 %f6549, %f6549, %f6555, %p131; selp.f32 %f6550, %f6550, %f6554, %p131; selp.b32 %r1261, %r1261, %r762, %p131; selp.b32 %r1262, %r1262, %r53, %p131; selp.b32 %r1263, %r1263, 65, %p131; setp.ne.s32 %p787, %r764, 0; bra.uni $L__BB24_88; $L__BB24_70: cvt.u32.u64 %r770, %rd38; setp.gt.s32 %p143, %r78, %r77; selp.b32 %r771, -1, 0, %p143; add.s32 %r772, %r78, %r771; mad.lo.s32 %r773, %r772, %r76, %r770; add.s32 %r774, %r773, %r1260; mul.wide.s32 %rd417, %r774, 16; add.s64 %rd1194, %rd27, %rd417; $L__BB24_72: ld.global.v4.f32 {%f2393, %f2394, %f2395, %f2396}, [%rd1194]; add.s32 %r89, %r78, 1; setp.eq.s32 %p144, %r89, %r77; @%p144 bra $L__BB24_74; bra.uni $L__BB24_73; $L__BB24_74: add.s64 %rd1195, %rd23, %rd391; bra.uni $L__BB24_75; $L__BB24_73: cvt.u32.u64 %r775, %rd38; setp.ge.s32 %p145, %r78, %r77; selp.b32 %r776, -1, 0, %p145; add.s32 %r777, %r89, %r776; mad.lo.s32 %r778, %r777, %r76, %r775; add.s32 %r779, %r778, %r1260; mul.wide.s32 %rd419, %r779, 16; add.s64 %rd1195, %rd27, %rd419; $L__BB24_75: cvt.rn.f32.s32 %f2397, %r78; sub.ftz.f32 %f2398, %f214, %f2397; ld.global.v4.f32 {%f2399, %f2400, %f2401, %f2402}, [%rd1195]; mov.f32 %f2407, 0f3F800000; sub.ftz.f32 %f2408, %f2407, %f2398; mul.ftz.f32 %f2409, %f2398, %f2399; mul.ftz.f32 %f2410, %f2398, %f2400; mul.ftz.f32 %f2411, %f2398, %f2401; mul.ftz.f32 %f2412, %f2398, %f2402; fma.rn.ftz.f32 %f6542, %f2408, %f2396, %f2412; fma.rn.ftz.f32 %f6541, %f2408, %f2395, %f2411; fma.rn.ftz.f32 %f6540, %f2408, %f2394, %f2410; fma.rn.ftz.f32 %f6539, %f2408, %f2393, %f2409; $L__BB24_77: sub.ftz.f32 %f2417, %f6539, %f6585; sub.ftz.f32 %f2418, %f6540, %f6586; sub.ftz.f32 %f2419, %f6541, %f6587; mul.ftz.f32 %f2420, %f6583, %f2418; fma.rn.ftz.f32 %f2421, %f6582, %f2417, %f2420; fma.rn.ftz.f32 %f2422, %f6584, %f2419, %f2421; mul.ftz.f32 %f231, %f164, %f2422; mul.ftz.f32 %f2423, %f6582, %f231; mul.ftz.f32 %f2424, %f6583, %f231; mul.ftz.f32 %f2425, %f6584, %f231; sub.ftz.f32 %f2426, %f2417, %f2423; sub.ftz.f32 %f2427, %f2418, %f2424; sub.ftz.f32 %f2428, %f2419, %f2425; mul.ftz.f32 %f2429, %f2427, %f2427; fma.rn.ftz.f32 %f2430, %f2426, %f2426, %f2429; fma.rn.ftz.f32 %f232, %f2428, %f2428, %f2430; mul.ftz.f32 %f233, %f6542, %f6542; setp.gtu.ftz.f32 %p147, %f232, %f233; mov.pred %p785, -1; @%p147 bra $L__BB24_79; sub.ftz.f32 %f2431, %f233, %f232; mul.ftz.f32 %f2432, %f164, %f2431; sqrt.approx.ftz.f32 %f2433, %f2432; sub.ftz.f32 %f2434, %f231, %f2433; setp.ltu.ftz.f32 %p148, %f2434, %f13; setp.gtu.ftz.f32 %p149, %f2434, %f14; or.pred %p785, %p149, %p148; selp.f32 %f6548, %f6548, %f2434, %p785; $L__BB24_79: selp.f32 %f6549, %f6549, 0f00000000, %p785; selp.f32 %f6550, %f6550, 0f00000000, %p785; cvt.u32.u64 %r780, %rd38; selp.b32 %r1261, %r1261, %r780, %p785; selp.b32 %r1262, %r1262, %r53, %p785; selp.b32 %r1263, %r1263, %r75, %p785; not.pred %p787, %p785; $L__BB24_88: not.pred %p164, %p787; @%p164 bra $L__BB24_120; and.b32 %r787, %r1263, 1; setp.eq.b32 %p165, %r787, 1; mov.pred %p166, 0; xor.pred %p167, %p165, %p166; not.pred %p168, %p167; mul.wide.s32 %rd429, %r1261, 16; add.s64 %rd108, %rd26, %rd429; @%p168 bra $L__BB24_91; bra.uni $L__BB24_90; $L__BB24_91: and.b32 %r788, %r1263, 8; setp.eq.s32 %p169, %r788, 0; @%p169 bra $L__BB24_93; bra.uni $L__BB24_92; $L__BB24_93: and.b32 %r790, %r1263, 6; setp.eq.s32 %p170, %r790, 0; mov.u32 %r1264, 0; @%p170 bra $L__BB24_95; ld.global.u32 %r1264, [%rd108]; bra.uni $L__BB24_95; $L__BB24_90: mul.wide.s32 %rd430, %r1261, 4; add.s64 %rd431, %rd21, %rd430; ld.global.u32 %r1264, [%rd431]; bra.uni $L__BB24_95; $L__BB24_92: mul.wide.s32 %rd432, %r1261, 4; add.s64 %rd433, %rd25, %rd432; ld.global.u32 %r1264, [%rd433]; $L__BB24_95: shl.b32 %r791, %r1264, 5; cvt.u64.u32 %rd434, %r791; and.b64 %rd435, %rd434, 134217696; add.s64 %rd436, %rd20, %rd435; ld.global.u8 %rs107, [%rd436+18]; and.b16 %rs108, %rs107, 2; setp.eq.s16 %p171, %rs108, 0; setp.ge.u32 %p172, %r1287, %r1; or.pred %p173, %p172, %p171; mov.u16 %rs242, %rs103; @%p173 bra $L__BB24_594; and.b32 %r792, %r1263, 6; setp.eq.s32 %p175, %r792, 0; mov.pred %p788, -1; @%p175 bra $L__BB24_106; mul.wide.s32 %rd437, %r1262, 208; add.s64 %rd438, %rd13, %rd437; ld.global.u32 %r1266, [%rd438+164]; mul.wide.u32 %rd439, %r1266, 16; add.s64 %rd440, %rd17, %rd439; ld.global.u16 %rs195, [%rd440+12]; ld.global.u32 %r1265, [%rd440+8]; ld.global.u64 %rd1196, [%rd440]; setp.eq.s64 %p176, %rd1196, 31; @%p176 bra $L__BB24_102; $L__BB24_98: setp.eq.s64 %p177, %rd1196, 0; @%p177 bra $L__BB24_100; add.s32 %r1266, %r1266, 2; bra.uni $L__BB24_101; $L__BB24_100: setp.eq.s16 %p178, %rs195, 0; mov.u32 %r1269, -1; mov.u32 %r1266, %r1265; @%p178 bra $L__BB24_103; $L__BB24_101: mul.wide.u32 %rd441, %r1266, 16; add.s64 %rd442, %rd17, %rd441; ld.global.u16 %rs195, [%rd442+12]; ld.global.u32 %r1265, [%rd442+8]; ld.global.u64 %rd1196, [%rd442]; setp.ne.s64 %p179, %rd1196, 31; @%p179 bra $L__BB24_98; $L__BB24_102: setp.eq.s16 %p180, %rs195, 0; selp.b32 %r1269, -1, %r1265, %p180; $L__BB24_103: setp.eq.s32 %p181, %r1269, -1; mov.f32 %f6557, 0f00000000; @%p181 bra $L__BB24_105; shr.s32 %r794, %r1263, 7; ld.global.u32 %r795, [%rd108+4]; add.s32 %r796, %r795, %r794; add.s32 %r797, %r796, %r1269; mul.wide.s32 %rd443, %r797, 4; add.s64 %rd444, %rd28, %rd443; mov.f32 %f2522, 0f3F800000; sub.ftz.f32 %f2523, %f2522, %f6549; ld.global.f32 %f2524, [%rd444]; ld.global.f32 %f2525, [%rd444+4]; mul.ftz.f32 %f2526, %f6549, %f2525; fma.rn.ftz.f32 %f6557, %f2523, %f2524, %f2526; $L__BB24_105: mul.ftz.f32 %f7321, %f7321, %f6557; setp.lt.ftz.f32 %p183, %f7321, 0f3A83126F; mov.pred %p788, %p166; mov.u16 %rs242, %rs103; @%p183 bra $L__BB24_594; $L__BB24_106: add.s32 %r1287, %r1287, 1; not.pred %p184, %p788; @%p184 bra $L__BB24_120; setp.ge.u32 %p185, %r1536, %r3; setp.geu.ftz.f32 %p186, %f6548, %f6595; and.pred %p187, %p185, %p186; @%p187 bra $L__BB24_119; min.u32 %r1277, %r1536, %r3; add.s32 %r798, %r1277, 1; setp.lt.u32 %p188, %r798, %r3; @%p188 bra $L__BB24_118; ld.global.f32 %f6562, [%rd29]; setp.lt.u32 %p189, %r1277, 2; mov.u32 %r1276, 0; @%p189 bra $L__BB24_117; min.u32 %r803, %r1, %r1536; not.b32 %r804, %r803; max.u32 %r805, %r804, -5; mov.u32 %r806, -2; sub.s32 %r115, %r806, %r805; mov.u32 %r807, -3; sub.s32 %r808, %r807, %r805; and.b32 %r116, %r115, 3; setp.lt.u32 %p190, %r808, 3; mov.u32 %r1276, 0; mov.u32 %r1274, 1; @%p190 bra $L__BB24_113; sub.s32 %r1272, %r115, %r116; mov.u32 %r1276, 0; mov.u32 %r1274, 1; $L__BB24_112: mul.wide.u32 %rd445, %r1274, 48; mov.u64 %rd446, kernel_params; add.s64 %rd447, %rd446, 3552; add.s64 %rd448, %rd447, %rd445; ld.const.u64 %rd449, [%rd448]; cvta.to.global.u64 %rd450, %rd449; add.s64 %rd452, %rd450, %rd362; ld.global.f32 %f2528, [%rd452]; setp.gt.ftz.f32 %p191, %f2528, %f6562; selp.f32 %f2529, %f2528, %f6562, %p191; selp.b32 %r811, %r1274, %r1276, %p191; ld.const.u64 %rd453, [%rd448+48]; cvta.to.global.u64 %rd454, %rd453; add.s64 %rd455, %rd454, %rd362; ld.global.f32 %f2530, [%rd455]; setp.gt.ftz.f32 %p192, %f2530, %f2529; selp.f32 %f2531, %f2530, %f2529, %p192; add.s32 %r812, %r1274, 1; selp.b32 %r813, %r812, %r811, %p192; ld.const.u64 %rd456, [%rd448+96]; cvta.to.global.u64 %rd457, %rd456; add.s64 %rd458, %rd457, %rd362; ld.global.f32 %f2532, [%rd458]; setp.gt.ftz.f32 %p193, %f2532, %f2531; selp.f32 %f2533, %f2532, %f2531, %p193; add.s32 %r814, %r1274, 2; selp.b32 %r815, %r814, %r813, %p193; add.s32 %r816, %r1274, 3; mul.wide.u32 %rd459, %r816, 48; add.s64 %rd460, %rd447, %rd459; ld.const.u64 %rd461, [%rd460]; cvta.to.global.u64 %rd462, %rd461; add.s64 %rd463, %rd462, %rd362; ld.global.f32 %f2534, [%rd463]; setp.gt.ftz.f32 %p194, %f2534, %f2533; selp.f32 %f6562, %f2534, %f2533, %p194; selp.b32 %r1276, %r816, %r815, %p194; add.s32 %r1274, %r1274, 4; add.s32 %r1272, %r1272, -4; setp.ne.s32 %p195, %r1272, 0; @%p195 bra $L__BB24_112; $L__BB24_113: setp.eq.s32 %p196, %r116, 0; @%p196 bra $L__BB24_117; mul.wide.u32 %rd464, %r1274, 48; mov.u64 %rd465, kernel_params; add.s64 %rd466, %rd465, %rd464; ld.const.u64 %rd467, [%rd466+3552]; cvta.to.global.u64 %rd468, %rd467; add.s64 %rd470, %rd468, %rd362; ld.global.f32 %f2535, [%rd470]; setp.gt.ftz.f32 %p197, %f2535, %f6562; selp.f32 %f6562, %f2535, %f6562, %p197; selp.b32 %r1276, %r1274, %r1276, %p197; setp.eq.s32 %p198, %r116, 1; @%p198 bra $L__BB24_117; add.s32 %r817, %r1274, 1; mul.wide.u32 %rd471, %r817, 48; add.s64 %rd473, %rd465, %rd471; ld.const.u64 %rd474, [%rd473+3552]; cvta.to.global.u64 %rd475, %rd474; add.s64 %rd477, %rd475, %rd362; ld.global.f32 %f2536, [%rd477]; setp.gt.ftz.f32 %p199, %f2536, %f6562; selp.f32 %f6562, %f2536, %f6562, %p199; selp.b32 %r1276, %r817, %r1276, %p199; setp.eq.s32 %p200, %r116, 2; @%p200 bra $L__BB24_117; add.s32 %r818, %r1274, 2; mul.wide.u32 %rd478, %r818, 48; add.s64 %rd480, %rd465, %rd478; ld.const.u64 %rd481, [%rd480+3552]; cvta.to.global.u64 %rd482, %rd481; add.s64 %rd484, %rd482, %rd362; ld.global.f32 %f2537, [%rd484]; setp.gt.ftz.f32 %p201, %f2537, %f6562; selp.f32 %f6562, %f2537, %f6562, %p201; selp.b32 %r1276, %r818, %r1276, %p201; $L__BB24_117: setp.lt.u32 %p202, %r1536, %r3; selp.b32 %r1277, %r1277, %r1276, %p202; max.ftz.f32 %f6595, %f6548, %f6562; $L__BB24_118: mul.wide.s32 %rd485, %r1277, 48; mov.u64 %rd486, kernel_params; add.s64 %rd487, %rd486, %rd485; ld.const.v2.u64 {%rd488, %rd489}, [%rd487+3552]; cvta.to.global.u64 %rd492, %rd488; add.s64 %rd494, %rd492, %rd362; st.global.f32 [%rd494], %f6548; cvta.to.global.u64 %rd495, %rd489; add.s64 %rd496, %rd495, %rd362; st.global.f32 [%rd496], %f6549; ld.const.v2.u64 {%rd497, %rd498}, [%rd487+3568]; cvta.to.global.u64 %rd501, %rd497; add.s64 %rd502, %rd501, %rd362; st.global.f32 [%rd502], %f6550; ld.const.v2.u64 {%rd503, %rd504}, [%rd487+3584]; cvta.to.global.u64 %rd507, %rd503; add.s64 %rd508, %rd507, %rd362; st.global.u32 [%rd508], %r1262; cvta.to.global.u64 %rd509, %rd498; add.s64 %rd510, %rd509, %rd362; st.global.u32 [%rd510], %r1261; cvta.to.global.u64 %rd511, %rd504; add.s64 %rd512, %rd511, %rd362; st.global.u32 [%rd512], %r1263; $L__BB24_119: add.s32 %r1536, %r1536, 1; $L__BB24_120: cvt.u32.u64 %r819, %rd37; add.s32 %r1244, %r819, 1; setp.lt.s32 %p203, %r1244, %r713; @%p203 bra $L__BB24_28; bra.uni $L__BB24_121; $L__BB24_17: mul.wide.s32 %rd372, %r1244, 4; xor.b64 %rd373, %rd372, -4; add.s64 %rd374, %rd11, %rd373; ld.global.u32 %r1288, [%rd374]; cvt.s64.s32 %rd34, %r1288; mul.wide.s32 %rd375, %r1288, 4; add.s64 %rd376, %rd12, %rd375; ld.global.u8 %rs100, [%rd376]; and.b16 %rs101, %rs100, 2; setp.eq.s16 %p85, %rs101, 0; mul.wide.s32 %rd377, %r1288, 208; add.s64 %rd378, %rd13, %rd377; add.s64 %rd35, %rd378, 48; @%p85 bra $L__BB24_24; bra.uni $L__BB24_18; $L__BB24_24: ld.global.v4.f32 {%f6512, %f6513, %f6514, %f6515}, [%rd35]; ld.global.v4.f32 {%f6516, %f6517, %f6518, %f6519}, [%rd35+16]; ld.global.v4.f32 {%f6520, %f6521, %f6522, %f6523}, [%rd35+32]; bra.uni $L__BB24_25; $L__BB24_18: ld.global.u32 %rd379, [%rd35+120]; ld.global.u32 %r716, [%rd35+104]; shl.b32 %r717, %r716, 1; cvt.rn.f32.s32 %f2003, %r717; ld.local.f32 %f2004, [%rd4+32]; mul.ftz.f32 %f2005, %f2004, %f2003; cvt.rzi.ftz.s32.f32 %r718, %f2005; add.s32 %r719, %r717, -1; min.s32 %r720, %r718, %r719; cvt.rn.f32.s32 %f2006, %r720; sub.ftz.f32 %f63, %f2005, %f2006; cvt.s64.s32 %rd380, %r720; add.s64 %rd381, %rd380, %rd379; shl.b64 %rd382, %rd381, 6; add.s64 %rd36, %rd31, %rd382; ld.global.v4.f32 {%f2007, %f2008, %f2009, %f2010}, [%rd36]; ld.global.v4.f32 {%f2011, %f2012, %f2013, %f2014}, [%rd36+64]; mul.ftz.f32 %f2015, %f2008, %f2012; fma.rn.ftz.f32 %f2016, %f2007, %f2011, %f2015; mul.ftz.f32 %f2017, %f2010, %f2014; fma.rn.ftz.f32 %f2018, %f2009, %f2013, %f2017; add.ftz.f32 %f72, %f2016, %f2018; setp.gt.ftz.f32 %p86, %f72, 0f3F7FDF3B; mov.f32 %f2019, 0f3F800000; sub.ftz.f32 %f73, %f2019, %f63; @%p86 bra $L__BB24_20; bra.uni $L__BB24_19; $L__BB24_20: mul.ftz.f32 %f2073, %f2007, %f73; fma.rn.ftz.f32 %f2074, %f2011, %f63, %f2073; mul.ftz.f32 %f2075, %f2008, %f73; fma.rn.ftz.f32 %f2076, %f2012, %f63, %f2075; mul.ftz.f32 %f2077, %f2009, %f73; fma.rn.ftz.f32 %f2078, %f2013, %f63, %f2077; mul.ftz.f32 %f2079, %f2010, %f73; fma.rn.ftz.f32 %f2080, %f2014, %f63, %f2079; mul.ftz.f32 %f2081, %f2076, %f2076; fma.rn.ftz.f32 %f2082, %f2074, %f2074, %f2081; mul.ftz.f32 %f2083, %f2080, %f2080; fma.rn.ftz.f32 %f2084, %f2078, %f2078, %f2083; add.ftz.f32 %f2085, %f2082, %f2084; rsqrt.approx.ftz.f32 %f2086, %f2085; mul.ftz.f32 %f6504, %f2080, %f2086; mul.ftz.f32 %f6503, %f2078, %f2086; mul.ftz.f32 %f6502, %f2076, %f2086; mul.ftz.f32 %f6501, %f2074, %f2086; bra.uni $L__BB24_21; $L__BB24_19: mov.f32 %f2020, 0fBF800000; max.ftz.f32 %f2021, %f72, %f2020; min.ftz.f32 %f2023, %f2021, %f2019; abs.ftz.f32 %f2024, %f2023; sub.ftz.f32 %f2025, %f2019, %f2024; mul.ftz.f32 %f2026, %f2025, 0f3F000000; sqrt.approx.ftz.f32 %f2027, %f2026; setp.gt.ftz.f32 %p87, %f2024, 0f3F11EB85; selp.f32 %f2028, %f2027, %f2024, %p87; mul.ftz.f32 %f2029, %f2028, %f2028; mov.f32 %f2030, 0f3C94D2E9; mov.f32 %f2031, 0f3D53F941; fma.rn.ftz.f32 %f2032, %f2031, %f2029, %f2030; mov.f32 %f2033, 0f3D3F841F; fma.rn.ftz.f32 %f2034, %f2032, %f2029, %f2033; mov.f32 %f2035, 0f3D994929; fma.rn.ftz.f32 %f2036, %f2034, %f2029, %f2035; mov.f32 %f2037, 0f3E2AAB94; fma.rn.ftz.f32 %f2038, %f2036, %f2029, %f2037; mul.ftz.f32 %f2039, %f2029, %f2038; fma.rn.ftz.f32 %f2040, %f2039, %f2028, %f2028; add.ftz.f32 %f2041, %f2040, %f2040; mov.f32 %f2042, 0f3FC90FDB; sub.ftz.f32 %f2043, %f2042, %f2040; selp.f32 %f2044, %f2041, %f2043, %p87; setp.lt.ftz.f32 %p88, %f2023, 0f00000000; mov.f32 %f2045, 0f40490FDB; sub.ftz.f32 %f2046, %f2045, %f2044; selp.f32 %f2047, %f2046, %f2044, %p88; mul.ftz.f32 %f2048, %f2007, %f72; sub.ftz.f32 %f2049, %f2011, %f2048; mul.ftz.f32 %f2050, %f2008, %f72; sub.ftz.f32 %f2051, %f2012, %f2050; mul.ftz.f32 %f2052, %f2009, %f72; sub.ftz.f32 %f2053, %f2013, %f2052; mul.ftz.f32 %f2054, %f2010, %f72; sub.ftz.f32 %f2055, %f2014, %f2054; mul.ftz.f32 %f2056, %f2051, %f2051; fma.rn.ftz.f32 %f2057, %f2049, %f2049, %f2056; mul.ftz.f32 %f2058, %f2055, %f2055; fma.rn.ftz.f32 %f2059, %f2053, %f2053, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; rsqrt.approx.ftz.f32 %f2061, %f2060; mul.ftz.f32 %f2062, %f2049, %f2061; mul.ftz.f32 %f2063, %f2051, %f2061; mul.ftz.f32 %f2064, %f2053, %f2061; mul.ftz.f32 %f2065, %f2055, %f2061; mul.ftz.f32 %f2066, %f63, %f2047; cos.approx.ftz.f32 %f2067, %f2066; sin.approx.ftz.f32 %f2068, %f2066; mul.ftz.f32 %f2069, %f2062, %f2068; mul.ftz.f32 %f2070, %f2063, %f2068; mul.ftz.f32 %f2071, %f2064, %f2068; mul.ftz.f32 %f2072, %f2065, %f2068; fma.rn.ftz.f32 %f6504, %f2010, %f2067, %f2072; fma.rn.ftz.f32 %f6503, %f2009, %f2067, %f2071; fma.rn.ftz.f32 %f6502, %f2008, %f2067, %f2070; fma.rn.ftz.f32 %f6501, %f2007, %f2067, %f2069; $L__BB24_21: ld.global.v4.f32 {%f2087, %f2088, %f2089, %f2090}, [%rd36+16]; ld.global.v4.f32 {%f2095, %f2096, %f2097, %f2098}, [%rd36+80]; mul.ftz.f32 %f2103, %f63, %f2095; mul.ftz.f32 %f2104, %f63, %f2096; mul.ftz.f32 %f2105, %f63, %f2097; mul.ftz.f32 %f2106, %f63, %f2098; fma.rn.ftz.f32 %f86, %f73, %f2087, %f2103; fma.rn.ftz.f32 %f87, %f73, %f2088, %f2104; fma.rn.ftz.f32 %f88, %f73, %f2089, %f2105; fma.rn.ftz.f32 %f2107, %f73, %f2090, %f2106; ld.global.v4.f32 {%f2108, %f2109, %f2110, %f2111}, [%rd36+32]; ld.global.v4.f32 {%f2116, %f2117, %f2118, %f2119}, [%rd36+96]; mul.ftz.f32 %f2124, %f63, %f2116; mul.ftz.f32 %f2125, %f63, %f2117; mul.ftz.f32 %f2126, %f63, %f2118; mul.ftz.f32 %f2127, %f63, %f2119; fma.rn.ftz.f32 %f2128, %f73, %f2108, %f2124; fma.rn.ftz.f32 %f2129, %f73, %f2109, %f2125; fma.rn.ftz.f32 %f2130, %f73, %f2110, %f2126; fma.rn.ftz.f32 %f2131, %f73, %f2111, %f2127; ld.global.v4.f32 {%f2132, %f2133, %f2134, %f2135}, [%rd36+48]; ld.global.v4.f32 {%f2140, %f2141, %f2142, %f2143}, [%rd36+112]; mul.ftz.f32 %f2148, %f63, %f2140; mul.ftz.f32 %f2149, %f63, %f2141; mul.ftz.f32 %f2150, %f63, %f2142; mul.ftz.f32 %f2151, %f63, %f2143; fma.rn.ftz.f32 %f2152, %f73, %f2132, %f2148; fma.rn.ftz.f32 %f2153, %f73, %f2133, %f2149; fma.rn.ftz.f32 %f2154, %f73, %f2134, %f2150; fma.rn.ftz.f32 %f2155, %f73, %f2135, %f2151; mul.ftz.f32 %f2156, %f6504, 0f3FB504F3; mul.ftz.f32 %f2157, %f6501, 0f3FB504F3; mul.ftz.f32 %f2158, %f6502, 0f3FB504F3; mul.ftz.f32 %f2159, %f6503, 0f3FB504F3; mul.ftz.f32 %f2160, %f2156, %f2157; mul.ftz.f32 %f2161, %f2156, %f2158; mul.ftz.f32 %f2162, %f2156, %f2159; mul.ftz.f32 %f2163, %f2157, %f2157; mul.ftz.f32 %f2164, %f2157, %f2158; mul.ftz.f32 %f2165, %f2157, %f2159; mul.ftz.f32 %f2166, %f2158, %f2158; mul.ftz.f32 %f2167, %f2158, %f2159; mul.ftz.f32 %f2168, %f2159, %f2159; sub.ftz.f32 %f2170, %f2019, %f2166; sub.ftz.f32 %f2171, %f2170, %f2168; sub.ftz.f32 %f2172, %f2164, %f2162; add.ftz.f32 %f2173, %f2161, %f2165; add.ftz.f32 %f2174, %f2162, %f2164; sub.ftz.f32 %f2175, %f2019, %f2163; sub.ftz.f32 %f2176, %f2175, %f2168; sub.ftz.f32 %f2177, %f2167, %f2160; sub.ftz.f32 %f2178, %f2165, %f2161; add.ftz.f32 %f2179, %f2160, %f2167; sub.ftz.f32 %f2180, %f2175, %f2166; mul.ftz.f32 %f2181, %f2172, %f2130; fma.rn.ftz.f32 %f2182, %f2171, %f2107, %f2181; fma.rn.ftz.f32 %f6510, %f2173, %f2153, %f2182; mul.ftz.f32 %f2183, %f2172, %f2131; fma.rn.ftz.f32 %f2184, %f2171, %f2128, %f2183; fma.rn.ftz.f32 %f90, %f2173, %f2154, %f2184; mul.ftz.f32 %f2185, %f2172, %f2152; fma.rn.ftz.f32 %f2186, %f2171, %f2129, %f2185; fma.rn.ftz.f32 %f91, %f2173, %f2155, %f2186; mul.ftz.f32 %f2187, %f2176, %f2130; fma.rn.ftz.f32 %f2188, %f2174, %f2107, %f2187; fma.rn.ftz.f32 %f92, %f2177, %f2153, %f2188; mul.ftz.f32 %f2189, %f2176, %f2131; fma.rn.ftz.f32 %f2190, %f2174, %f2128, %f2189; fma.rn.ftz.f32 %f6509, %f2177, %f2154, %f2190; mul.ftz.f32 %f2191, %f2176, %f2152; fma.rn.ftz.f32 %f2192, %f2174, %f2129, %f2191; fma.rn.ftz.f32 %f94, %f2177, %f2155, %f2192; mul.ftz.f32 %f2193, %f2179, %f2130; fma.rn.ftz.f32 %f2194, %f2178, %f2107, %f2193; fma.rn.ftz.f32 %f95, %f2180, %f2153, %f2194; mul.ftz.f32 %f2195, %f2179, %f2131; fma.rn.ftz.f32 %f2196, %f2178, %f2128, %f2195; fma.rn.ftz.f32 %f96, %f2180, %f2154, %f2196; mul.ftz.f32 %f2197, %f2179, %f2152; fma.rn.ftz.f32 %f2198, %f2178, %f2129, %f2197; fma.rn.ftz.f32 %f6508, %f2180, %f2155, %f2198; mul.ftz.f32 %f2199, %f6509, %f6508; mul.ftz.f32 %f98, %f96, %f94; sub.ftz.f32 %f6507, %f2199, %f98; mul.ftz.f32 %f100, %f96, %f91; mul.ftz.f32 %f2200, %f90, %f6508; sub.ftz.f32 %f6506, %f100, %f2200; mul.ftz.f32 %f102, %f90, %f94; mul.ftz.f32 %f2201, %f6509, %f91; sub.ftz.f32 %f6505, %f102, %f2201; mul.ftz.f32 %f2202, %f92, %f6506; fma.rn.ftz.f32 %f2203, %f6510, %f6507, %f2202; fma.rn.ftz.f32 %f6511, %f95, %f6505, %f2203; setp.neu.ftz.f32 %p89, %f6511, 0f00000000; @%p89 bra $L__BB24_23; add.ftz.f32 %f6510, %f6510, 0f322BCC77; add.ftz.f32 %f6508, %f6508, 0f322BCC77; add.ftz.f32 %f6509, %f6509, 0f322BCC77; mul.ftz.f32 %f2204, %f6509, %f6508; sub.ftz.f32 %f6507, %f2204, %f98; mul.ftz.f32 %f2205, %f90, %f6508; sub.ftz.f32 %f6506, %f100, %f2205; mul.ftz.f32 %f2206, %f6509, %f91; sub.ftz.f32 %f6505, %f102, %f2206; mul.ftz.f32 %f2207, %f92, %f6506; fma.rn.ftz.f32 %f2208, %f6510, %f6507, %f2207; fma.rn.ftz.f32 %f2209, %f95, %f6505, %f2208; setp.eq.ftz.f32 %p90, %f2209, 0f00000000; selp.f32 %f6511, 0f7F7FFFFF, %f2209, %p90; $L__BB24_23: mul.ftz.f32 %f2210, %f92, %f6508; mul.ftz.f32 %f2211, %f95, %f94; sub.ftz.f32 %f2212, %f2211, %f2210; mul.ftz.f32 %f2213, %f95, %f91; mul.ftz.f32 %f2214, %f6508, %f6510; sub.ftz.f32 %f2215, %f2214, %f2213; mul.ftz.f32 %f2216, %f94, %f6510; mul.ftz.f32 %f2217, %f92, %f91; sub.ftz.f32 %f2218, %f2217, %f2216; div.approx.ftz.f32 %f6516, %f2212, %f6511; div.approx.ftz.f32 %f6517, %f2215, %f6511; div.approx.ftz.f32 %f6518, %f2218, %f6511; mul.ftz.f32 %f2219, %f95, %f6509; mul.ftz.f32 %f2220, %f92, %f96; sub.ftz.f32 %f2221, %f2220, %f2219; mul.ftz.f32 %f2222, %f96, %f6510; mul.ftz.f32 %f2223, %f95, %f90; sub.ftz.f32 %f2224, %f2223, %f2222; mul.ftz.f32 %f2225, %f92, %f90; mul.ftz.f32 %f2226, %f6509, %f6510; sub.ftz.f32 %f2227, %f2226, %f2225; div.approx.ftz.f32 %f6520, %f2221, %f6511; div.approx.ftz.f32 %f6521, %f2224, %f6511; div.approx.ftz.f32 %f6522, %f2227, %f6511; div.approx.ftz.f32 %f6512, %f6507, %f6511; div.approx.ftz.f32 %f6513, %f6506, %f6511; mul.ftz.f32 %f2228, %f87, %f6513; fma.rn.ftz.f32 %f2229, %f86, %f6512, %f2228; div.approx.ftz.f32 %f6514, %f6505, %f6511; fma.rn.ftz.f32 %f2230, %f88, %f6514, %f2229; neg.ftz.f32 %f6515, %f2230; mul.ftz.f32 %f2231, %f87, %f6517; fma.rn.ftz.f32 %f2232, %f86, %f6516, %f2231; fma.rn.ftz.f32 %f2233, %f88, %f6518, %f2232; neg.ftz.f32 %f6519, %f2233; mul.ftz.f32 %f2234, %f87, %f6521; fma.rn.ftz.f32 %f2235, %f86, %f6520, %f2234; fma.rn.ftz.f32 %f2236, %f88, %f6522, %f2235; neg.ftz.f32 %f6523, %f2236; $L__BB24_25: mul.ftz.f32 %f2249, %f6513, %f6459; fma.rn.ftz.f32 %f2250, %f6512, %f6460, %f2249; fma.rn.ftz.f32 %f2251, %f6514, %f6458, %f2250; add.ftz.f32 %f6585, %f6515, %f2251; mul.ftz.f32 %f2252, %f6517, %f6459; fma.rn.ftz.f32 %f2253, %f6516, %f6460, %f2252; fma.rn.ftz.f32 %f2254, %f6518, %f6458, %f2253; add.ftz.f32 %f6586, %f6519, %f2254; mul.ftz.f32 %f2255, %f6521, %f6459; fma.rn.ftz.f32 %f2256, %f6520, %f6460, %f2255; fma.rn.ftz.f32 %f2257, %f6522, %f6458, %f2256; add.ftz.f32 %f6587, %f6523, %f2257; mul.ftz.f32 %f2258, %f6513, %f6456; fma.rn.ftz.f32 %f2259, %f6512, %f6457, %f2258; fma.rn.ftz.f32 %f2260, %f6514, %f6455, %f2259; mul.ftz.f32 %f2261, %f6517, %f6456; fma.rn.ftz.f32 %f2262, %f6516, %f6457, %f2261; fma.rn.ftz.f32 %f2263, %f6518, %f6455, %f2262; mul.ftz.f32 %f2264, %f6521, %f6456; fma.rn.ftz.f32 %f2265, %f6520, %f6457, %f2264; fma.rn.ftz.f32 %f2266, %f6522, %f6455, %f2265; abs.ftz.f32 %f2267, %f2260; setp.gt.ftz.f32 %p91, %f2267, 0f17800000; mov.b32 %r721, %f2260; and.b32 %r722, %r721, -2147483648; or.b32 %r723, %r722, 394264576; mov.b32 %f2268, %r723; selp.f32 %f6582, %f2260, %f2268, %p91; abs.ftz.f32 %f2269, %f2263; setp.gt.ftz.f32 %p92, %f2269, 0f17800000; mov.b32 %r724, %f2263; and.b32 %r725, %r724, -2147483648; or.b32 %r726, %r725, 394264576; mov.b32 %f2270, %r726; selp.f32 %f6583, %f2263, %f2270, %p92; abs.ftz.f32 %f2271, %f2266; setp.gt.ftz.f32 %p93, %f2271, 0f17800000; mov.b32 %r727, %f2266; and.b32 %r728, %r727, -2147483648; or.b32 %r729, %r728, 394264576; mov.b32 %f2272, %r729; selp.f32 %f6584, %f2266, %f2272, %p93; rcp.approx.ftz.f32 %f6579, %f6582; rcp.approx.ftz.f32 %f6580, %f6583; rcp.approx.ftz.f32 %f6581, %f6584; add.s32 %r1241, %r1241, 1; mov.u32 %r730, 1985229328; st.local.u32 [%rd33+4], %r730; shl.b64 %rd383, %rd34, 2; add.s64 %rd384, %rd30, %rd383; ld.global.u32 %r1240, [%rd384]; bra.uni $L__BB24_121; $L__BB24_8: setp.eq.s32 %p67, %r1240, 1985229328; setp.lt.s32 %p68, %r1240, 0; or.pred %p69, %p68, %p67; @%p69 bra $L__BB24_15; $L__BB24_9: mul.wide.s32 %rd363, %r1240, 16; add.s64 %rd364, %rd9, %rd363; ld.global.v4.f32 {%f1933, %f1934, %f1935, %f1936}, [%rd364+16]; ld.global.v4.f32 {%f1941, %f1942, %f1943, %f1944}, [%rd364+32]; ld.global.v4.f32 {%f1949, %f1950, %f1951, %f1952}, [%rd364+48]; sub.ftz.f32 %f1957, %f1933, %f6585; mul.ftz.f32 %f1958, %f6579, %f1957; sub.ftz.f32 %f1959, %f1935, %f6585; mul.ftz.f32 %f1960, %f6579, %f1959; sub.ftz.f32 %f1961, %f1941, %f6586; mul.ftz.f32 %f1962, %f6580, %f1961; sub.ftz.f32 %f1963, %f1943, %f6586; mul.ftz.f32 %f1964, %f6580, %f1963; sub.ftz.f32 %f1965, %f1949, %f6587; mul.ftz.f32 %f1966, %f6581, %f1965; sub.ftz.f32 %f1967, %f1951, %f6587; mul.ftz.f32 %f1968, %f6581, %f1967; min.ftz.f32 %f1969, %f1958, %f1960; min.ftz.f32 %f1970, %f1962, %f1964; min.ftz.f32 %f1971, %f1966, %f1968; max.ftz.f32 %f1972, %f13, %f1969; max.ftz.f32 %f1973, %f1970, %f1971; max.ftz.f32 %f61, %f1972, %f1973; max.ftz.f32 %f1974, %f1958, %f1960; max.ftz.f32 %f1975, %f1962, %f1964; max.ftz.f32 %f1976, %f1966, %f1968; min.ftz.f32 %f1977, %f14, %f1974; min.ftz.f32 %f1978, %f1975, %f1976; min.ftz.f32 %f1979, %f1977, %f1978; sub.ftz.f32 %f1980, %f1934, %f6585; mul.ftz.f32 %f1981, %f6579, %f1980; sub.ftz.f32 %f1982, %f1936, %f6585; mul.ftz.f32 %f1983, %f6579, %f1982; sub.ftz.f32 %f1984, %f1942, %f6586; mul.ftz.f32 %f1985, %f6580, %f1984; sub.ftz.f32 %f1986, %f1944, %f6586; mul.ftz.f32 %f1987, %f6580, %f1986; sub.ftz.f32 %f1988, %f1950, %f6587; mul.ftz.f32 %f1989, %f6581, %f1988; sub.ftz.f32 %f1990, %f1952, %f6587; mul.ftz.f32 %f1991, %f6581, %f1990; min.ftz.f32 %f1992, %f1981, %f1983; min.ftz.f32 %f1993, %f1985, %f1987; min.ftz.f32 %f1994, %f1989, %f1991; max.ftz.f32 %f1995, %f13, %f1992; max.ftz.f32 %f1996, %f1993, %f1994; max.ftz.f32 %f62, %f1995, %f1996; max.ftz.f32 %f1997, %f1981, %f1983; max.ftz.f32 %f1998, %f1985, %f1987; max.ftz.f32 %f1999, %f1989, %f1991; min.ftz.f32 %f2000, %f14, %f1997; min.ftz.f32 %f2001, %f1998, %f1999; min.ftz.f32 %f2002, %f2000, %f2001; setp.ge.ftz.f32 %p70, %f1979, %f61; ld.global.v4.u32 {%r701, %r702, %r703, %r1240}, [%rd364]; and.b32 %r706, %r701, %r668; setp.ne.s32 %p71, %r706, 0; and.pred %p72, %p71, %p70; selp.u32 %r707, 1, 0, %p72; setp.ltu.ftz.f32 %p73, %f2002, %f62; and.b32 %r709, %r702, %r668; setp.eq.s32 %p74, %r709, 0; or.pred %p75, %p74, %p73; selp.b32 %r710, 0, 2, %p75; or.b32 %r24, %r710, %r707; setp.eq.s32 %p76, %r24, 3; mul.wide.s32 %rd365, %r1241, 4; add.s64 %rd32, %rd6, %rd365; @%p76 bra $L__BB24_13; bra.uni $L__BB24_10; $L__BB24_13: setp.lt.ftz.f32 %p79, %f62, %f61; selp.b32 %r711, %r703, %r1240, %p79; selp.b32 %r1240, %r1240, %r703, %p79; add.s32 %r1241, %r1241, 1; st.local.u32 [%rd32+4], %r711; bra.uni $L__BB24_14; $L__BB24_10: cvt.u16.u32 %rs99, %r24; setp.eq.s16 %p77, %rs99, 2; @%p77 bra $L__BB24_14; setp.ne.s16 %p78, %rs99, 0; mov.u32 %r1240, %r703; @%p78 bra $L__BB24_14; ld.local.u32 %r1240, [%rd32]; add.s32 %r1241, %r1241, -1; $L__BB24_14: setp.ne.s32 %p80, %r1240, 1985229328; setp.gt.s32 %p81, %r1240, -1; and.pred %p82, %p81, %p80; @%p82 bra $L__BB24_9; $L__BB24_15: setp.gt.s32 %p83, %r1240, -1; @%p83 bra $L__BB24_121; bra.uni $L__BB24_16; $L__BB24_121: setp.ne.s32 %p204, %r1240, 1985229328; @%p204 bra $L__BB24_8; setp.lt.s32 %p205, %r1241, 0; mov.u32 %r1240, %r700; @%p205 bra $L__BB24_124; mul.wide.s32 %rd515, %r1241, 4; add.s64 %rd516, %rd6, %rd515; ld.local.u32 %r1240, [%rd516]; add.s32 %r1241, %r1241, -1; mov.u32 %r1288, -1; mov.f32 %f6579, %f8; mov.f32 %f6580, %f9; mov.f32 %f6581, %f10; mov.f32 %f6582, %f5; mov.f32 %f6583, %f6; mov.f32 %f6584, %f7; mov.f32 %f6585, %f6460; mov.f32 %f6586, %f6459; mov.f32 %f6587, %f6458; $L__BB24_124: setp.ne.s32 %p206, %r1240, 1985229328; mov.u16 %rs242, 0; @%p206 bra $L__BB24_8; bra.uni $L__BB24_594; $L__BB24_331: mov.u32 %r989, 1985229328; st.local.u32 [%rd6], %r989; ld.const.u32 %r1402, [kernel_params+1104]; ld.local.f32 %f6937, [%rd4+4]; ld.local.f32 %f6936, [%rd4+8]; ld.local.f32 %f1112, [%rd4+24]; ld.local.f32 %f1113, [%rd4+28]; ld.const.v2.u64 {%rd752, %rd753}, [kernel_params+2096]; mov.u32 %r1536, 0; cvta.to.global.u64 %rd220, %rd752; cvta.to.global.u64 %rd221, %rd753; ld.const.v2.u64 {%rd756, %rd757}, [kernel_params+2128]; cvta.to.global.u64 %rd222, %rd757; ld.const.u64 %rd760, [kernel_params+2160]; cvta.to.global.u64 %rd223, %rd760; ld.const.v2.u64 {%rd761, %rd762}, [kernel_params+2144]; cvta.to.global.u64 %rd224, %rd761; cvta.to.global.u64 %rd225, %rd756; ld.const.u64 %rd765, [kernel_params+2232]; cvta.to.global.u64 %rd226, %rd765; ld.const.v2.u64 {%rd766, %rd767}, [kernel_params+2256]; cvta.to.global.u64 %rd227, %rd766; ld.const.u32 %r355, [kernel_params+1128]; ld.const.v2.u64 {%rd770, %rd771}, [kernel_params+2112]; cvta.to.global.u64 %rd228, %rd770; cvta.to.global.u64 %rd229, %rd762; ld.const.u64 %rd774, [kernel_params+2408]; cvta.to.global.u64 %rd230, %rd774; ld.const.u64 %rd775, [kernel_params+2216]; cvta.to.global.u64 %rd231, %rd775; cvta.to.global.u64 %rd232, %rd771; ld.const.u64 %rd776, [kernel_params+2296]; cvta.to.global.u64 %rd233, %rd776; ld.const.v2.u64 {%rd777, %rd778}, [kernel_params+2304]; cvta.to.global.u64 %rd234, %rd778; cvta.to.global.u64 %rd235, %rd777; cvta.to.global.u64 %rd236, %rd767; ld.const.u64 %rd781, [kernel_params+2344]; cvta.to.global.u64 %rd237, %rd781; ld.const.u64 %rd782, [kernel_params+2320]; cvta.to.global.u64 %rd238, %rd782; ld.const.u64 %rd783, [kernel_params+3552]; cvta.to.global.u64 %rd784, %rd783; shl.b64 %rd785, %rd3, 2; add.s64 %rd239, %rd784, %rd785; mov.u32 %r1445, -1; mov.f32 %f7321, 0f3F800000; setp.eq.s32 %p475, %r355, 0; mov.u16 %rs144, 1; mov.pred %p483, 0; mov.pred %p505, -1; mov.f32 %f7015, %f8; mov.f32 %f7016, %f9; mov.f32 %f7017, %f10; mov.f32 %f7018, %f5; mov.f32 %f7019, %f6; mov.f32 %f7020, %f7; mov.f32 %f7021, %f6460; mov.f32 %f7022, %f6937; mov.f32 %f7023, %f6936; mov.f32 %f7028, %f1113; mov.u32 %r1444, %r1536; mov.u32 %r1403, %r1536; bra.uni $L__BB24_332; $L__BB24_340: mul.wide.s32 %rd789, %r1402, 16; xor.b64 %rd790, %rd789, -16; add.s64 %rd791, %rd221, %rd790; ld.global.v4.u32 {%r1406, %r1002, %r1003, %r1004}, [%rd791]; setp.gt.s32 %p470, %r1406, -1; mul.wide.s32 %rd794, %r1403, 4; add.s64 %rd241, %rd6, %rd794; @%p470 bra $L__BB24_342; bra.uni $L__BB24_341; $L__BB24_342: ld.local.u32 %r1402, [%rd241]; add.s32 %r1403, %r1403, -1; setp.le.s32 %p474, %r1002, %r1406; @%p474 bra $L__BB24_415; cvt.u16.u32 %rs143, %r1004; and.b16 %rs49, %rs143, 127; and.b32 %r1015, %r1004, 64; setp.eq.s32 %p476, %r1015, 0; or.pred %p27, %p476, %p475; mul.ftz.f32 %f4637, %f7019, %f7019; fma.rn.ftz.f32 %f4638, %f7018, %f7018, %f4637; fma.rn.ftz.f32 %f4639, %f7020, %f7020, %f4638; rcp.approx.ftz.f32 %f1165, %f4639; $L__BB24_344: cvt.s64.s32 %rd242, %r1406; setp.ne.s32 %p477, %r1445, -1; mov.u32 %r1412, %r1445; @%p477 bra $L__BB24_346; shl.b64 %rd802, %rd242, 2; add.s64 %rd803, %rd222, %rd802; ld.global.u32 %r1412, [%rd803]; $L__BB24_346: shl.b64 %rd804, %rd242, 2; add.s64 %rd805, %rd225, %rd804; ld.global.u32 %r1016, [%rd805]; cvt.s64.s32 %rd243, %r1016; ld.local.u32 %r1017, [%rd4+36]; setp.ne.s32 %p478, %r1017, %r1016; @%p478 bra $L__BB24_348; ld.local.u32 %r1018, [%rd4+40]; setp.eq.s32 %p479, %r1018, %r1412; mov.u16 %rs218, %rs144; @%p479 bra $L__BB24_350; $L__BB24_348: cvt.u32.u64 %r1019, %rd243; ld.local.u32 %r1020, [%rd4+44]; setp.ne.s32 %p480, %r1020, %r1019; mov.u16 %rs218, 0; @%p480 bra $L__BB24_350; ld.local.u32 %r1021, [%rd4+48]; setp.eq.s32 %p481, %r1021, %r1412; selp.u16 %rs218, 1, 0, %p481; $L__BB24_350: setp.ne.s16 %p482, %rs218, 0; @%p482 bra $L__BB24_414; setp.eq.s16 %p484, %rs49, 1; @%p484 bra $L__BB24_374; setp.eq.s16 %p485, %rs49, 72; @%p485 bra $L__BB24_354; setp.ne.s16 %p486, %rs49, 8; mov.pred %p799, %p483; @%p486 bra $L__BB24_382; $L__BB24_354: @%p27 bra $L__BB24_356; shl.b64 %rd806, %rd242, 3; add.s64 %rd807, %rd229, %rd806; ld.global.v2.f32 {%f4640, %f4641}, [%rd807]; ld.local.f32 %f4644, [%rd4+32]; setp.lt.ftz.f32 %p488, %f4644, %f4640; setp.gt.ftz.f32 %p489, %f4644, %f4641; or.pred %p490, %p488, %p489; mov.pred %p799, %p483; @%p490 bra $L__BB24_382; $L__BB24_356: add.s64 %rd809, %rd228, %rd804; ld.global.u32 %r404, [%rd809]; and.b32 %r1022, %r404, 64; setp.eq.s32 %p491, %r1022, 0; @%p491 bra $L__BB24_370; bra.uni $L__BB24_357; $L__BB24_370: shl.b64 %rd820, %rd243, 4; add.s64 %rd821, %rd233, %rd820; ld.global.v4.f32 {%f6981, %f6982, %f6983, %f6984}, [%rd821]; bra.uni $L__BB24_371; $L__BB24_374: shl.b64 %rd822, %rd243, 4; add.s64 %rd823, %rd226, %rd822; ld.global.u32 %r1038, [%rd823+12]; mul.wide.u32 %rd824, %r1038, 12; add.s64 %rd825, %rd227, %rd824; add.s32 %r1039, %r1038, 1; mul.wide.u32 %rd826, %r1039, 12; add.s64 %rd827, %rd227, %rd826; add.s32 %r1040, %r1038, 2; mul.wide.u32 %rd828, %r1040, 12; add.s64 %rd829, %rd227, %rd828; ld.global.f32 %f4689, [%rd825]; sub.ftz.f32 %f1198, %f4689, %f7021; ld.global.f32 %f4690, [%rd825+4]; sub.ftz.f32 %f1199, %f4690, %f7022; ld.global.f32 %f4691, [%rd825+8]; sub.ftz.f32 %f1200, %f4691, %f7023; ld.global.f32 %f4692, [%rd827]; sub.ftz.f32 %f4693, %f4692, %f7021; ld.global.f32 %f4694, [%rd827+4]; sub.ftz.f32 %f4695, %f4694, %f7022; ld.global.f32 %f4696, [%rd827+8]; sub.ftz.f32 %f4697, %f4696, %f7023; ld.global.f32 %f4698, [%rd829]; sub.ftz.f32 %f4699, %f4698, %f7021; ld.global.f32 %f4700, [%rd829+4]; sub.ftz.f32 %f4701, %f4700, %f7022; ld.global.f32 %f4702, [%rd829+8]; sub.ftz.f32 %f4703, %f4702, %f7023; sub.ftz.f32 %f1201, %f4699, %f1198; sub.ftz.f32 %f1202, %f4701, %f1199; sub.ftz.f32 %f1203, %f4703, %f1200; sub.ftz.f32 %f1204, %f1198, %f4693; sub.ftz.f32 %f1205, %f1199, %f4695; sub.ftz.f32 %f1206, %f1200, %f4697; sub.ftz.f32 %f4704, %f4693, %f4699; sub.ftz.f32 %f4705, %f4695, %f4701; sub.ftz.f32 %f4706, %f4697, %f4703; add.ftz.f32 %f4707, %f1198, %f4699; add.ftz.f32 %f4708, %f1199, %f4701; add.ftz.f32 %f4709, %f1200, %f4703; mul.ftz.f32 %f4710, %f1202, %f4709; mul.ftz.f32 %f4711, %f4708, %f1203; sub.ftz.f32 %f4712, %f4710, %f4711; mul.ftz.f32 %f4713, %f4707, %f1203; mul.ftz.f32 %f4714, %f1201, %f4709; sub.ftz.f32 %f4715, %f4713, %f4714; mul.ftz.f32 %f4716, %f1201, %f4708; mul.ftz.f32 %f4717, %f4707, %f1202; sub.ftz.f32 %f4718, %f4716, %f4717; mul.ftz.f32 %f4719, %f7019, %f4715; fma.rn.ftz.f32 %f4720, %f7018, %f4712, %f4719; fma.rn.ftz.f32 %f1207, %f7020, %f4718, %f4720; add.ftz.f32 %f4721, %f1198, %f4693; add.ftz.f32 %f4722, %f1199, %f4695; add.ftz.f32 %f4723, %f1200, %f4697; mul.ftz.f32 %f4724, %f1205, %f4723; mul.ftz.f32 %f4725, %f4722, %f1206; sub.ftz.f32 %f4726, %f4724, %f4725; mul.ftz.f32 %f4727, %f4721, %f1206; mul.ftz.f32 %f4728, %f1204, %f4723; sub.ftz.f32 %f4729, %f4727, %f4728; mul.ftz.f32 %f4730, %f1204, %f4722; mul.ftz.f32 %f4731, %f4721, %f1205; sub.ftz.f32 %f4732, %f4730, %f4731; mul.ftz.f32 %f4733, %f7019, %f4729; fma.rn.ftz.f32 %f4734, %f7018, %f4726, %f4733; fma.rn.ftz.f32 %f1208, %f7020, %f4732, %f4734; add.ftz.f32 %f4735, %f4693, %f4699; add.ftz.f32 %f4736, %f4695, %f4701; add.ftz.f32 %f4737, %f4697, %f4703; mul.ftz.f32 %f4738, %f4705, %f4737; mul.ftz.f32 %f4739, %f4736, %f4706; sub.ftz.f32 %f4740, %f4738, %f4739; mul.ftz.f32 %f4741, %f4735, %f4706; mul.ftz.f32 %f4742, %f4704, %f4737; sub.ftz.f32 %f4743, %f4741, %f4742; mul.ftz.f32 %f4744, %f4704, %f4736; mul.ftz.f32 %f4745, %f4735, %f4705; sub.ftz.f32 %f4746, %f4744, %f4745; mul.ftz.f32 %f4747, %f7019, %f4743; fma.rn.ftz.f32 %f4748, %f7018, %f4740, %f4747; fma.rn.ftz.f32 %f4749, %f7020, %f4746, %f4748; add.ftz.f32 %f4750, %f1208, %f1207; add.ftz.f32 %f1209, %f4749, %f4750; abs.ftz.f32 %f1210, %f1209; mul.ftz.f32 %f4751, %f1210, 0f34000000; min.ftz.f32 %f4752, %f1208, %f4749; min.ftz.f32 %f4753, %f1207, %f4752; max.ftz.f32 %f4754, %f1208, %f4749; max.ftz.f32 %f4755, %f1207, %f4754; mul.ftz.f32 %f4756, %f1210, 0fB4000000; setp.ltu.ftz.f32 %p506, %f4753, %f4756; setp.gtu.ftz.f32 %p507, %f4755, %f4751; and.pred %p508, %p506, %p507; mov.pred %p798, %p505; @%p508 bra $L__BB24_380; mul.ftz.f32 %f4757, %f1205, %f1203; mul.ftz.f32 %f4758, %f1206, %f1202; sub.ftz.f32 %f4759, %f4757, %f4758; mul.ftz.f32 %f4760, %f1204, %f1203; mul.ftz.f32 %f4761, %f1206, %f1201; sub.ftz.f32 %f4762, %f4761, %f4760; mul.ftz.f32 %f4763, %f1205, %f1201; mul.ftz.f32 %f4764, %f1204, %f1202; sub.ftz.f32 %f4765, %f4764, %f4763; add.ftz.f32 %f1211, %f4759, %f4759; add.ftz.f32 %f1212, %f4762, %f4762; add.ftz.f32 %f1213, %f4765, %f4765; mul.ftz.f32 %f4766, %f7019, %f1212; fma.rn.ftz.f32 %f4767, %f7018, %f1211, %f4766; fma.rn.ftz.f32 %f1214, %f7020, %f1213, %f4767; setp.eq.ftz.f32 %p510, %f1214, 0f00000000; mov.pred %p798, %p505; @%p510 bra $L__BB24_380; mul.ftz.f32 %f4768, %f1199, %f1212; fma.rn.ftz.f32 %f4769, %f1198, %f1211, %f4768; fma.rn.ftz.f32 %f4770, %f1200, %f1213, %f4769; div.approx.ftz.f32 %f1215, %f4770, %f1214; setp.ltu.ftz.f32 %p512, %f1215, %f1112; setp.gtu.ftz.f32 %p513, %f1215, %f1113; or.pred %p514, %p513, %p512; mov.pred %p798, %p505; @%p514 bra $L__BB24_380; setp.lt.ftz.f32 %p515, %f1210, 0f219392EF; mov.f32 %f6986, 0f00000000; @%p515 bra $L__BB24_379; rcp.approx.ftz.f32 %f6986, %f1209; $L__BB24_379: mul.ftz.f32 %f4772, %f1207, %f6986; mov.f32 %f4773, 0f3F800000; min.ftz.f32 %f6994, %f4772, %f4773; mul.ftz.f32 %f4774, %f1208, %f6986; min.ftz.f32 %f6993, %f4774, %f4773; mov.pred %p798, 0; mov.f32 %f6995, %f1215; $L__BB24_380: mov.pred %p799, %p483; @%p798 bra $L__BB24_382; cvt.u32.u64 %r1041, %rd243; add.s64 %rd831, %rd232, %rd804; ld.global.u32 %r1042, [%rd831]; and.b32 %r1043, %r1042, %r668; setp.eq.s32 %p518, %r1043, 0; selp.f32 %f6990, %f6990, %f6995, %p518; selp.f32 %f6991, %f6991, %f6994, %p518; selp.f32 %f6992, %f6992, %f6993, %p518; selp.b32 %r1418, %r1418, %r1041, %p518; selp.b32 %r1419, %r1419, %r1412, %p518; selp.b32 %r1420, %r1420, 1, %p518; setp.ne.s32 %p799, %r1043, 0; bra.uni $L__BB24_382; $L__BB24_357: mul.wide.s32 %rd810, %r1412, 208; add.s64 %rd811, %rd223, %rd810; ld.global.u32 %r405, [%rd811+148]; ld.global.u32 %r406, [%rd811+152]; shl.b32 %r1023, %r406, 1; cvt.rn.f32.s32 %f4645, %r1023; ld.local.f32 %f4646, [%rd4+32]; mul.ftz.f32 %f1174, %f4646, %f4645; cvt.rzi.ftz.s32.f32 %r1024, %f1174; add.s32 %r1025, %r1023, -1; min.s32 %r407, %r1024, %r1025; ld.global.u32 %r1414, [%rd811+164]; mul.wide.u32 %rd812, %r1414, 16; add.s64 %rd813, %rd234, %rd812; ld.global.u16 %rs219, [%rd813+12]; ld.global.u32 %r1413, [%rd813+8]; ld.global.u64 %rd1221, [%rd813]; setp.eq.s64 %p492, %rd1221, 11; @%p492 bra $L__BB24_362; $L__BB24_358: setp.eq.s64 %p493, %rd1221, 0; @%p493 bra $L__BB24_360; add.s32 %r1414, %r1414, 2; bra.uni $L__BB24_361; $L__BB24_360: setp.eq.s16 %p494, %rs219, 0; mov.u32 %r1417, -1; mov.u32 %r1414, %r1413; @%p494 bra $L__BB24_363; $L__BB24_361: mul.wide.u32 %rd814, %r1414, 16; add.s64 %rd815, %rd234, %rd814; ld.global.u16 %rs219, [%rd815+12]; ld.global.u32 %r1413, [%rd815+8]; ld.global.u64 %rd1221, [%rd815]; setp.ne.s64 %p495, %rd1221, 11; @%p495 bra $L__BB24_358; $L__BB24_362: setp.eq.s16 %p496, %rs219, 0; selp.b32 %r1417, -1, %r1413, %p496; $L__BB24_363: setp.eq.s32 %p497, %r407, %r406; @%p497 bra $L__BB24_365; bra.uni $L__BB24_364; $L__BB24_365: shl.b64 %rd817, %rd243, 4; add.s64 %rd1222, %rd233, %rd817; bra.uni $L__BB24_366; $L__BB24_364: cvt.u32.u64 %r1027, %rd243; setp.gt.s32 %p498, %r407, %r406; selp.b32 %r1028, -1, 0, %p498; add.s32 %r1029, %r407, %r1028; mad.lo.s32 %r1030, %r1029, %r405, %r1027; add.s32 %r1031, %r1030, %r1417; mul.wide.s32 %rd816, %r1031, 16; add.s64 %rd1222, %rd237, %rd816; $L__BB24_366: ld.global.v4.f32 {%f4647, %f4648, %f4649, %f4650}, [%rd1222]; add.s32 %r418, %r407, 1; setp.eq.s32 %p499, %r418, %r406; @%p499 bra $L__BB24_368; bra.uni $L__BB24_367; $L__BB24_368: shl.b64 %rd819, %rd243, 4; add.s64 %rd1223, %rd233, %rd819; bra.uni $L__BB24_369; $L__BB24_367: cvt.u32.u64 %r1032, %rd243; setp.ge.s32 %p500, %r407, %r406; selp.b32 %r1033, -1, 0, %p500; add.s32 %r1034, %r418, %r1033; mad.lo.s32 %r1035, %r1034, %r405, %r1032; add.s32 %r1036, %r1035, %r1417; mul.wide.s32 %rd818, %r1036, 16; add.s64 %rd1223, %rd237, %rd818; $L__BB24_369: cvt.rn.f32.s32 %f4651, %r407; sub.ftz.f32 %f4652, %f1174, %f4651; ld.global.v4.f32 {%f4653, %f4654, %f4655, %f4656}, [%rd1223]; mov.f32 %f4661, 0f3F800000; sub.ftz.f32 %f4662, %f4661, %f4652; mul.ftz.f32 %f4663, %f4652, %f4653; mul.ftz.f32 %f4664, %f4652, %f4654; mul.ftz.f32 %f4665, %f4652, %f4655; mul.ftz.f32 %f4666, %f4652, %f4656; fma.rn.ftz.f32 %f6984, %f4662, %f4650, %f4666; fma.rn.ftz.f32 %f6983, %f4662, %f4649, %f4665; fma.rn.ftz.f32 %f6982, %f4662, %f4648, %f4664; fma.rn.ftz.f32 %f6981, %f4662, %f4647, %f4663; $L__BB24_371: sub.ftz.f32 %f4671, %f6981, %f7021; sub.ftz.f32 %f4672, %f6982, %f7022; sub.ftz.f32 %f4673, %f6983, %f7023; mul.ftz.f32 %f4674, %f7019, %f4672; fma.rn.ftz.f32 %f4675, %f7018, %f4671, %f4674; fma.rn.ftz.f32 %f4676, %f7020, %f4673, %f4675; mul.ftz.f32 %f1191, %f1165, %f4676; mul.ftz.f32 %f4677, %f7018, %f1191; mul.ftz.f32 %f4678, %f7019, %f1191; mul.ftz.f32 %f4679, %f7020, %f1191; sub.ftz.f32 %f4680, %f4671, %f4677; sub.ftz.f32 %f4681, %f4672, %f4678; sub.ftz.f32 %f4682, %f4673, %f4679; mul.ftz.f32 %f4683, %f4681, %f4681; fma.rn.ftz.f32 %f4684, %f4680, %f4680, %f4683; fma.rn.ftz.f32 %f1192, %f4682, %f4682, %f4684; mul.ftz.f32 %f1193, %f6984, %f6984; setp.gtu.ftz.f32 %p502, %f1192, %f1193; mov.pred %p797, -1; @%p502 bra $L__BB24_373; sub.ftz.f32 %f4685, %f1193, %f1192; mul.ftz.f32 %f4686, %f1165, %f4685; sqrt.approx.ftz.f32 %f4687, %f4686; sub.ftz.f32 %f4688, %f1191, %f4687; setp.ltu.ftz.f32 %p503, %f4688, %f1112; setp.gtu.ftz.f32 %p504, %f4688, %f1113; or.pred %p797, %p504, %p503; selp.f32 %f6990, %f6990, %f4688, %p797; $L__BB24_373: selp.f32 %f6991, %f6991, 0f00000000, %p797; selp.f32 %f6992, %f6992, 0f00000000, %p797; cvt.u32.u64 %r1037, %rd243; selp.b32 %r1418, %r1418, %r1037, %p797; selp.b32 %r1419, %r1419, %r1412, %p797; selp.b32 %r1420, %r1420, %r404, %p797; not.pred %p799, %p797; $L__BB24_382: not.pred %p519, %p799; @%p519 bra $L__BB24_414; and.b32 %r1044, %r1420, 1; setp.eq.b32 %p520, %r1044, 1; mov.pred %p521, 0; xor.pred %p522, %p520, %p521; not.pred %p523, %p522; mul.wide.s32 %rd832, %r1418, 16; add.s64 %rd253, %rd236, %rd832; @%p523 bra $L__BB24_385; bra.uni $L__BB24_384; $L__BB24_385: and.b32 %r1045, %r1420, 8; setp.eq.s32 %p524, %r1045, 0; @%p524 bra $L__BB24_387; bra.uni $L__BB24_386; $L__BB24_387: and.b32 %r1047, %r1420, 6; setp.eq.s32 %p525, %r1047, 0; mov.u32 %r1421, 0; @%p525 bra $L__BB24_389; ld.global.u32 %r1421, [%rd253]; bra.uni $L__BB24_389; $L__BB24_384: mul.wide.s32 %rd833, %r1418, 4; add.s64 %rd834, %rd231, %rd833; ld.global.u32 %r1421, [%rd834]; bra.uni $L__BB24_389; $L__BB24_386: mul.wide.s32 %rd835, %r1418, 4; add.s64 %rd836, %rd235, %rd835; ld.global.u32 %r1421, [%rd836]; $L__BB24_389: shl.b32 %r1048, %r1421, 5; cvt.u64.u32 %rd837, %r1048; and.b64 %rd838, %rd837, 134217696; add.s64 %rd839, %rd230, %rd838; ld.global.u8 %rs148, [%rd839+18]; and.b16 %rs149, %rs148, 2; setp.eq.s16 %p526, %rs149, 0; setp.ge.u32 %p527, %r1444, %r1; or.pred %p528, %p527, %p526; mov.u16 %rs242, %rs144; @%p528 bra $L__BB24_594; and.b32 %r1049, %r1420, 6; setp.eq.s32 %p530, %r1049, 0; mov.pred %p800, -1; @%p530 bra $L__BB24_400; mul.wide.s32 %rd840, %r1419, 208; add.s64 %rd841, %rd223, %rd840; ld.global.u32 %r1423, [%rd841+164]; mul.wide.u32 %rd842, %r1423, 16; add.s64 %rd843, %rd234, %rd842; ld.global.u16 %rs221, [%rd843+12]; ld.global.u32 %r1422, [%rd843+8]; ld.global.u64 %rd1224, [%rd843]; setp.eq.s64 %p531, %rd1224, 31; @%p531 bra $L__BB24_396; $L__BB24_392: setp.eq.s64 %p532, %rd1224, 0; @%p532 bra $L__BB24_394; add.s32 %r1423, %r1423, 2; bra.uni $L__BB24_395; $L__BB24_394: setp.eq.s16 %p533, %rs221, 0; mov.u32 %r1426, -1; mov.u32 %r1423, %r1422; @%p533 bra $L__BB24_397; $L__BB24_395: mul.wide.u32 %rd844, %r1423, 16; add.s64 %rd845, %rd234, %rd844; ld.global.u16 %rs221, [%rd845+12]; ld.global.u32 %r1422, [%rd845+8]; ld.global.u64 %rd1224, [%rd845]; setp.ne.s64 %p534, %rd1224, 31; @%p534 bra $L__BB24_392; $L__BB24_396: setp.eq.s16 %p535, %rs221, 0; selp.b32 %r1426, -1, %r1422, %p535; $L__BB24_397: setp.eq.s32 %p536, %r1426, -1; mov.f32 %f6996, 0f00000000; @%p536 bra $L__BB24_399; shr.s32 %r1051, %r1420, 7; ld.global.u32 %r1052, [%rd253+4]; add.s32 %r1053, %r1052, %r1051; add.s32 %r1054, %r1053, %r1426; mul.wide.s32 %rd846, %r1054, 4; add.s64 %rd847, %rd238, %rd846; mov.f32 %f4776, 0f3F800000; sub.ftz.f32 %f4777, %f4776, %f6991; ld.global.f32 %f4778, [%rd847]; ld.global.f32 %f4779, [%rd847+4]; mul.ftz.f32 %f4780, %f6991, %f4779; fma.rn.ftz.f32 %f6996, %f4777, %f4778, %f4780; $L__BB24_399: mul.ftz.f32 %f7321, %f7321, %f6996; setp.lt.ftz.f32 %p538, %f7321, 0f3A83126F; mov.pred %p800, %p521; mov.u16 %rs242, %rs144; @%p538 bra $L__BB24_594; $L__BB24_400: add.s32 %r1444, %r1444, 1; not.pred %p539, %p800; @%p539 bra $L__BB24_414; setp.ge.u32 %p540, %r1536, %r3; setp.geu.ftz.f32 %p541, %f6990, %f7028; and.pred %p542, %p540, %p541; @%p542 bra $L__BB24_413; min.u32 %r1434, %r1536, %r3; add.s32 %r1055, %r1434, 1; setp.lt.u32 %p543, %r1055, %r3; @%p543 bra $L__BB24_412; ld.global.f32 %f7001, [%rd239]; setp.lt.u32 %p544, %r1434, 2; mov.u32 %r1433, 0; @%p544 bra $L__BB24_411; min.u32 %r1060, %r1, %r1536; not.b32 %r1061, %r1060; max.u32 %r1062, %r1061, -5; mov.u32 %r1063, -2; sub.s32 %r444, %r1063, %r1062; mov.u32 %r1064, -3; sub.s32 %r1065, %r1064, %r1062; and.b32 %r445, %r444, 3; setp.lt.u32 %p545, %r1065, 3; mov.u32 %r1433, 0; mov.u32 %r1431, 1; @%p545 bra $L__BB24_407; sub.s32 %r1429, %r444, %r445; mov.u32 %r1433, 0; mov.u32 %r1431, 1; $L__BB24_406: mul.wide.u32 %rd848, %r1431, 48; mov.u64 %rd849, kernel_params; add.s64 %rd850, %rd849, 3552; add.s64 %rd851, %rd850, %rd848; ld.const.u64 %rd852, [%rd851]; cvta.to.global.u64 %rd853, %rd852; add.s64 %rd855, %rd853, %rd785; ld.global.f32 %f4782, [%rd855]; setp.gt.ftz.f32 %p546, %f4782, %f7001; selp.f32 %f4783, %f4782, %f7001, %p546; selp.b32 %r1068, %r1431, %r1433, %p546; ld.const.u64 %rd856, [%rd851+48]; cvta.to.global.u64 %rd857, %rd856; add.s64 %rd858, %rd857, %rd785; ld.global.f32 %f4784, [%rd858]; setp.gt.ftz.f32 %p547, %f4784, %f4783; selp.f32 %f4785, %f4784, %f4783, %p547; add.s32 %r1069, %r1431, 1; selp.b32 %r1070, %r1069, %r1068, %p547; ld.const.u64 %rd859, [%rd851+96]; cvta.to.global.u64 %rd860, %rd859; add.s64 %rd861, %rd860, %rd785; ld.global.f32 %f4786, [%rd861]; setp.gt.ftz.f32 %p548, %f4786, %f4785; selp.f32 %f4787, %f4786, %f4785, %p548; add.s32 %r1071, %r1431, 2; selp.b32 %r1072, %r1071, %r1070, %p548; add.s32 %r1073, %r1431, 3; mul.wide.u32 %rd862, %r1073, 48; add.s64 %rd863, %rd850, %rd862; ld.const.u64 %rd864, [%rd863]; cvta.to.global.u64 %rd865, %rd864; add.s64 %rd866, %rd865, %rd785; ld.global.f32 %f4788, [%rd866]; setp.gt.ftz.f32 %p549, %f4788, %f4787; selp.f32 %f7001, %f4788, %f4787, %p549; selp.b32 %r1433, %r1073, %r1072, %p549; add.s32 %r1431, %r1431, 4; add.s32 %r1429, %r1429, -4; setp.ne.s32 %p550, %r1429, 0; @%p550 bra $L__BB24_406; $L__BB24_407: setp.eq.s32 %p551, %r445, 0; @%p551 bra $L__BB24_411; mul.wide.u32 %rd867, %r1431, 48; mov.u64 %rd868, kernel_params; add.s64 %rd869, %rd868, %rd867; ld.const.u64 %rd870, [%rd869+3552]; cvta.to.global.u64 %rd871, %rd870; add.s64 %rd873, %rd871, %rd785; ld.global.f32 %f4789, [%rd873]; setp.gt.ftz.f32 %p552, %f4789, %f7001; selp.f32 %f7001, %f4789, %f7001, %p552; selp.b32 %r1433, %r1431, %r1433, %p552; setp.eq.s32 %p553, %r445, 1; @%p553 bra $L__BB24_411; add.s32 %r1074, %r1431, 1; mul.wide.u32 %rd874, %r1074, 48; add.s64 %rd876, %rd868, %rd874; ld.const.u64 %rd877, [%rd876+3552]; cvta.to.global.u64 %rd878, %rd877; add.s64 %rd880, %rd878, %rd785; ld.global.f32 %f4790, [%rd880]; setp.gt.ftz.f32 %p554, %f4790, %f7001; selp.f32 %f7001, %f4790, %f7001, %p554; selp.b32 %r1433, %r1074, %r1433, %p554; setp.eq.s32 %p555, %r445, 2; @%p555 bra $L__BB24_411; add.s32 %r1075, %r1431, 2; mul.wide.u32 %rd881, %r1075, 48; add.s64 %rd883, %rd868, %rd881; ld.const.u64 %rd884, [%rd883+3552]; cvta.to.global.u64 %rd885, %rd884; add.s64 %rd887, %rd885, %rd785; ld.global.f32 %f4791, [%rd887]; setp.gt.ftz.f32 %p556, %f4791, %f7001; selp.f32 %f7001, %f4791, %f7001, %p556; selp.b32 %r1433, %r1075, %r1433, %p556; $L__BB24_411: setp.lt.u32 %p557, %r1536, %r3; selp.b32 %r1434, %r1434, %r1433, %p557; max.ftz.f32 %f7028, %f6990, %f7001; $L__BB24_412: mul.wide.s32 %rd888, %r1434, 48; mov.u64 %rd889, kernel_params; add.s64 %rd890, %rd889, %rd888; ld.const.v2.u64 {%rd891, %rd892}, [%rd890+3552]; cvta.to.global.u64 %rd895, %rd891; add.s64 %rd897, %rd895, %rd785; st.global.f32 [%rd897], %f6990; cvta.to.global.u64 %rd898, %rd892; add.s64 %rd899, %rd898, %rd785; st.global.f32 [%rd899], %f6991; ld.const.v2.u64 {%rd900, %rd901}, [%rd890+3568]; cvta.to.global.u64 %rd904, %rd900; add.s64 %rd905, %rd904, %rd785; st.global.f32 [%rd905], %f6992; ld.const.v2.u64 {%rd906, %rd907}, [%rd890+3584]; cvta.to.global.u64 %rd910, %rd906; add.s64 %rd911, %rd910, %rd785; st.global.u32 [%rd911], %r1419; cvta.to.global.u64 %rd912, %rd901; add.s64 %rd913, %rd912, %rd785; st.global.u32 [%rd913], %r1418; cvta.to.global.u64 %rd914, %rd907; add.s64 %rd915, %rd914, %rd785; st.global.u32 [%rd915], %r1420; $L__BB24_413: add.s32 %r1536, %r1536, 1; $L__BB24_414: cvt.u32.u64 %r1076, %rd242; add.s32 %r1406, %r1076, 1; setp.lt.s32 %p558, %r1406, %r1002; @%p558 bra $L__BB24_344; bra.uni $L__BB24_415; $L__BB24_341: mul.wide.s32 %rd795, %r1406, 4; xor.b64 %rd796, %rd795, -4; add.s64 %rd797, %rd222, %rd796; ld.global.u32 %r1445, [%rd797]; mul.wide.s32 %rd798, %r1445, 208; add.s64 %rd799, %rd223, %rd798; ld.global.v4.f32 {%f4589, %f4590, %f4591, %f4592}, [%rd799+48]; ld.global.v4.f32 {%f4597, %f4598, %f4599, %f4600}, [%rd799+64]; ld.global.v4.f32 {%f4605, %f4606, %f4607, %f4608}, [%rd799+80]; mul.ftz.f32 %f4613, %f4590, %f6937; fma.rn.ftz.f32 %f4614, %f6460, %f4589, %f4613; fma.rn.ftz.f32 %f4615, %f4591, %f6936, %f4614; add.ftz.f32 %f7021, %f4592, %f4615; mul.ftz.f32 %f4616, %f4598, %f6937; fma.rn.ftz.f32 %f4617, %f6460, %f4597, %f4616; fma.rn.ftz.f32 %f4618, %f4599, %f6936, %f4617; add.ftz.f32 %f7022, %f4600, %f4618; mul.ftz.f32 %f4619, %f4606, %f6937; fma.rn.ftz.f32 %f4620, %f6460, %f4605, %f4619; fma.rn.ftz.f32 %f4621, %f4607, %f6936, %f4620; add.ftz.f32 %f7023, %f4608, %f4621; mul.ftz.f32 %f4622, %f4590, %f6456; fma.rn.ftz.f32 %f4623, %f4589, %f6457, %f4622; fma.rn.ftz.f32 %f4624, %f4591, %f6455, %f4623; mul.ftz.f32 %f4625, %f4598, %f6456; fma.rn.ftz.f32 %f4626, %f4597, %f6457, %f4625; fma.rn.ftz.f32 %f4627, %f4599, %f6455, %f4626; mul.ftz.f32 %f4628, %f4606, %f6456; fma.rn.ftz.f32 %f4629, %f4605, %f6457, %f4628; fma.rn.ftz.f32 %f4630, %f4607, %f6455, %f4629; abs.ftz.f32 %f4631, %f4624; setp.gt.ftz.f32 %p471, %f4631, 0f17800000; mov.b32 %r1005, %f4624; and.b32 %r1006, %r1005, -2147483648; or.b32 %r1007, %r1006, 394264576; mov.b32 %f4632, %r1007; selp.f32 %f7018, %f4624, %f4632, %p471; abs.ftz.f32 %f4633, %f4627; setp.gt.ftz.f32 %p472, %f4633, 0f17800000; mov.b32 %r1008, %f4627; and.b32 %r1009, %r1008, -2147483648; or.b32 %r1010, %r1009, 394264576; mov.b32 %f4634, %r1010; selp.f32 %f7019, %f4627, %f4634, %p472; abs.ftz.f32 %f4635, %f4630; setp.gt.ftz.f32 %p473, %f4635, 0f17800000; mov.b32 %r1011, %f4630; and.b32 %r1012, %r1011, -2147483648; or.b32 %r1013, %r1012, 394264576; mov.b32 %f4636, %r1013; selp.f32 %f7020, %f4630, %f4636, %p473; rcp.approx.ftz.f32 %f7015, %f7018; rcp.approx.ftz.f32 %f7016, %f7019; rcp.approx.ftz.f32 %f7017, %f7020; add.s32 %r1403, %r1403, 1; mov.u32 %r1014, 1985229328; st.local.u32 [%rd241+4], %r1014; mul.wide.s32 %rd800, %r1445, 4; add.s64 %rd801, %rd224, %rd800; ld.global.u32 %r1402, [%rd801]; bra.uni $L__BB24_415; $L__BB24_332: setp.eq.s32 %p453, %r1402, 1985229328; setp.lt.s32 %p454, %r1402, 0; or.pred %p455, %p454, %p453; @%p455 bra $L__BB24_339; $L__BB24_333: mul.wide.s32 %rd786, %r1402, 16; add.s64 %rd787, %rd220, %rd786; ld.global.v4.f32 {%f4519, %f4520, %f4521, %f4522}, [%rd787+16]; ld.global.v4.f32 {%f4527, %f4528, %f4529, %f4530}, [%rd787+32]; ld.global.v4.f32 {%f4535, %f4536, %f4537, %f4538}, [%rd787+48]; sub.ftz.f32 %f4543, %f4519, %f7021; mul.ftz.f32 %f4544, %f7015, %f4543; sub.ftz.f32 %f4545, %f4521, %f7021; mul.ftz.f32 %f4546, %f7015, %f4545; sub.ftz.f32 %f4547, %f4527, %f7022; mul.ftz.f32 %f4548, %f7016, %f4547; sub.ftz.f32 %f4549, %f4529, %f7022; mul.ftz.f32 %f4550, %f7016, %f4549; sub.ftz.f32 %f4551, %f4535, %f7023; mul.ftz.f32 %f4552, %f7017, %f4551; sub.ftz.f32 %f4553, %f4537, %f7023; mul.ftz.f32 %f4554, %f7017, %f4553; min.ftz.f32 %f4555, %f4544, %f4546; min.ftz.f32 %f4556, %f4548, %f4550; min.ftz.f32 %f4557, %f4552, %f4554; max.ftz.f32 %f4558, %f1112, %f4555; max.ftz.f32 %f4559, %f4556, %f4557; max.ftz.f32 %f1154, %f4558, %f4559; max.ftz.f32 %f4560, %f4544, %f4546; max.ftz.f32 %f4561, %f4548, %f4550; max.ftz.f32 %f4562, %f4552, %f4554; min.ftz.f32 %f4563, %f1113, %f4560; min.ftz.f32 %f4564, %f4561, %f4562; min.ftz.f32 %f4565, %f4563, %f4564; sub.ftz.f32 %f4566, %f4520, %f7021; mul.ftz.f32 %f4567, %f7015, %f4566; sub.ftz.f32 %f4568, %f4522, %f7021; mul.ftz.f32 %f4569, %f7015, %f4568; sub.ftz.f32 %f4570, %f4528, %f7022; mul.ftz.f32 %f4571, %f7016, %f4570; sub.ftz.f32 %f4572, %f4530, %f7022; mul.ftz.f32 %f4573, %f7016, %f4572; sub.ftz.f32 %f4574, %f4536, %f7023; mul.ftz.f32 %f4575, %f7017, %f4574; sub.ftz.f32 %f4576, %f4538, %f7023; mul.ftz.f32 %f4577, %f7017, %f4576; min.ftz.f32 %f4578, %f4567, %f4569; min.ftz.f32 %f4579, %f4571, %f4573; min.ftz.f32 %f4580, %f4575, %f4577; max.ftz.f32 %f4581, %f1112, %f4578; max.ftz.f32 %f4582, %f4579, %f4580; max.ftz.f32 %f1155, %f4581, %f4582; max.ftz.f32 %f4583, %f4567, %f4569; max.ftz.f32 %f4584, %f4571, %f4573; max.ftz.f32 %f4585, %f4575, %f4577; min.ftz.f32 %f4586, %f1113, %f4583; min.ftz.f32 %f4587, %f4584, %f4585; min.ftz.f32 %f4588, %f4586, %f4587; setp.ge.ftz.f32 %p456, %f4565, %f1154; ld.global.v4.u32 {%r990, %r991, %r992, %r1402}, [%rd787]; and.b32 %r995, %r990, %r668; setp.ne.s32 %p457, %r995, 0; and.pred %p458, %p457, %p456; selp.u32 %r996, 1, 0, %p458; setp.ltu.ftz.f32 %p459, %f4588, %f1155; and.b32 %r998, %r991, %r668; setp.eq.s32 %p460, %r998, 0; or.pred %p461, %p460, %p459; selp.b32 %r999, 0, 2, %p461; or.b32 %r374, %r999, %r996; setp.eq.s32 %p462, %r374, 3; mul.wide.s32 %rd788, %r1403, 4; add.s64 %rd240, %rd6, %rd788; @%p462 bra $L__BB24_337; bra.uni $L__BB24_334; $L__BB24_337: setp.lt.ftz.f32 %p465, %f1155, %f1154; selp.b32 %r1000, %r992, %r1402, %p465; selp.b32 %r1402, %r1402, %r992, %p465; add.s32 %r1403, %r1403, 1; st.local.u32 [%rd240+4], %r1000; bra.uni $L__BB24_338; $L__BB24_334: cvt.u16.u32 %rs142, %r374; setp.eq.s16 %p463, %rs142, 2; @%p463 bra $L__BB24_338; setp.ne.s16 %p464, %rs142, 0; mov.u32 %r1402, %r992; @%p464 bra $L__BB24_338; ld.local.u32 %r1402, [%rd240]; add.s32 %r1403, %r1403, -1; $L__BB24_338: setp.ne.s32 %p466, %r1402, 1985229328; setp.gt.s32 %p467, %r1402, -1; and.pred %p468, %p467, %p466; @%p468 bra $L__BB24_333; $L__BB24_339: setp.gt.s32 %p469, %r1402, -1; @%p469 bra $L__BB24_415; bra.uni $L__BB24_340; $L__BB24_415: setp.ne.s32 %p559, %r1402, 1985229328; @%p559 bra $L__BB24_332; setp.lt.s32 %p560, %r1403, 0; mov.u32 %r1402, %r989; @%p560 bra $L__BB24_418; mul.wide.s32 %rd918, %r1403, 4; add.s64 %rd919, %rd6, %rd918; ld.local.u32 %r1402, [%rd919]; add.s32 %r1403, %r1403, -1; mov.u32 %r1445, -1; mov.f32 %f7015, %f8; mov.f32 %f7016, %f9; mov.f32 %f7017, %f10; mov.f32 %f7018, %f5; mov.f32 %f7019, %f6; mov.f32 %f7020, %f7; mov.f32 %f7021, %f6460; mov.f32 %f7022, %f6937; mov.f32 %f7023, %f6936; $L__BB24_418: setp.ne.s32 %p561, %r1402, 1985229328; mov.u16 %rs242, 0; @%p561 bra $L__BB24_332; bra.uni $L__BB24_594; } .func _Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2_( .param .b32 _Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_0, .param .align 4 .b8 _Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_1[12], .param .align 4 .b8 _Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_2[12] ) { .local .align 16 .b8 __local_depot25[1072]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<252>; .reg .b16 %rs<26>; .reg .f32 %f<2844>; .reg .b32 %r<557>; .reg .b64 %rd<710>; mov.u64 %SPL, __local_depot25; cvta.local.u64 %SP, %SPL; ld.param.u32 %r239, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_0]; ld.param.f32 %f3, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_1+8]; ld.param.f32 %f724, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_2+8]; ld.param.f32 %f2, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_1+4]; ld.param.f32 %f725, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_2+4]; ld.param.f32 %f1, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_1]; ld.param.f32 %f726, [_Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2__param_2]; add.u64 %rd1, %SPL, 0; sub.ftz.f32 %f727, %f726, %f1; sub.ftz.f32 %f728, %f725, %f2; sub.ftz.f32 %f729, %f724, %f3; mul.ftz.f32 %f730, %f728, %f728; fma.rn.ftz.f32 %f731, %f727, %f727, %f730; fma.rn.ftz.f32 %f732, %f729, %f729, %f731; sqrt.approx.ftz.f32 %f4, %f732; rsqrt.approx.ftz.f32 %f733, %f732; mul.ftz.f32 %f5, %f727, %f733; mul.ftz.f32 %f6, %f728, %f733; mul.ftz.f32 %f7, %f729, %f733; cvt.s64.s32 %rd2, %r239; ld.const.u64 %rd89, [kernel_params+2544]; cvta.to.global.u64 %rd90, %rd89; mul.wide.s32 %rd91, %r239, 4; add.s64 %rd92, %rd90, %rd91; ld.global.u8 %rs9, [%rd92+3]; and.b16 %rs10, %rs9, 64; setp.eq.s16 %p8, %rs10, 0; selp.b32 %r1, 2047, 134152192, %p8; ld.const.u32 %r2, [kernel_params+12]; shl.b32 %r3, %r2, 1; setp.eq.s32 %p9, %r3, 0; @%p9 bra $L__BB25_211; ld.const.u64 %rd93, [kernel_params+2720]; ld.const.u64 %rd94, [kernel_params+2712]; cvta.to.global.u64 %rd95, %rd94; shl.b64 %rd96, %rd2, 2; add.s64 %rd97, %rd95, %rd96; ld.global.u32 %r538, [%rd97]; cvta.to.global.u64 %rd98, %rd93; add.s64 %rd99, %rd98, %rd96; ld.global.u32 %r537, [%rd99]; mov.b32 %r242, %f1; or.b32 %r243, %r242, -2147483648; setp.eq.s32 %p10, %r243, -2147483648; add.ftz.f32 %f736, %f1, %f1; setp.neu.ftz.f32 %p11, %f736, %f1; or.pred %p12, %p11, %p10; shl.b32 %r244, %r242, 1; setp.lt.u32 %p13, %r244, -16777215; ld.const.v2.u32 {%r245, %r246}, [kernel_params+1104]; mov.b32 %r247, %f5; or.b32 %r248, %r247, -2147483648; setp.eq.s32 %p14, %r248, -2147483648; add.ftz.f32 %f737, %f5, %f5; setp.neu.ftz.f32 %p15, %f737, %f5; or.pred %p16, %p15, %p14; ld.const.u64 %rd100, [kernel_params+2184]; cvta.to.global.u64 %rd3, %rd100; ld.const.v2.u64 {%rd101, %rd102}, [kernel_params+2096]; mov.u32 %r448, 0; cvta.to.global.u64 %rd6, %rd101; ld.const.v2.u64 {%rd103, %rd104}, [kernel_params+2128]; cvta.to.global.u64 %rd7, %rd104; cvta.to.global.u64 %rd8, %rd103; ld.const.v2.u64 {%rd107, %rd108}, [kernel_params+2224]; cvta.to.global.u64 %rd11, %rd108; ld.const.v2.u64 {%rd109, %rd110}, [kernel_params+2256]; cvta.to.global.u64 %rd14, %rd109; ld.const.u64 %rd111, [kernel_params+2120]; cvta.to.global.u64 %rd15, %rd111; ld.const.u64 %rd112, [kernel_params+2160]; cvta.to.global.u64 %rd16, %rd112; ld.const.u64 %rd113, [kernel_params+2144]; cvta.to.global.u64 %rd17, %rd113; ld.const.v2.u64 {%rd114, %rd115}, [kernel_params+2304]; cvta.to.global.u64 %rd20, %rd115; shl.b32 %r249, %r247, 1; setp.lt.u32 %p17, %r249, -16777215; mul.ftz.f32 %f738, %f6, %f6; fma.rn.ftz.f32 %f739, %f5, %f5, %f738; fma.rn.ftz.f32 %f740, %f7, %f7, %f739; setp.neu.ftz.f32 %p18, %f740, 0f00000000; mov.f32 %f2517, 0f00000000; ld.const.u64 %rd116, [kernel_params+2408]; cvta.to.global.u64 %rd21, %rd116; ld.const.u64 %rd117, [kernel_params+2296]; cvta.to.global.u64 %rd22, %rd117; ld.const.u64 %rd118, [kernel_params+2272]; cvta.to.global.u64 %rd23, %rd118; ld.const.u64 %rd119, [kernel_params+2216]; cvta.to.global.u64 %rd24, %rd119; ld.const.u64 %rd120, [kernel_params+2344]; cvta.to.global.u64 %rd25, %rd120; ld.const.u64 %rd121, [kernel_params+2776]; cvta.to.global.u64 %rd122, %rd121; add.s64 %rd26, %rd122, %rd96; and.pred %p19, %p12, %p13; setp.num.ftz.f32 %p20, %f1, %f1; and.pred %p21, %p20, %p19; and.pred %p22, %p16, %p17; setp.num.ftz.f32 %p23, %f5, %f5; and.pred %p24, %p23, %p22; and.pred %p25, %p18, %p24; and.pred %p1, %p25, %p21; not.pred %p27, %p1; setp.eq.s32 %p28, %r246, 0; mov.pred %p148, -1; cvta.to.global.u64 %rd433, %rd107; cvta.to.global.u64 %rd358, %rd114; bra.uni $L__BB25_2; $L__BB25_197: and.b32 %r413, %r550, 1; setp.eq.b32 %p225, %r413, 1; mov.pred %p226, 0; xor.pred %p227, %p225, %p226; not.pred %p228, %p227; ld.global.u32 %r551, [%rd26]; @%p228 bra $L__BB25_204; bra.uni $L__BB25_198; $L__BB25_204: mov.u64 %rd639, kernel_params; add.s64 %rd709, %rd639, 2776; add.s64 %rd708, %rd639, 2768; setp.eq.s32 %p235, %r551, -1; mov.u32 %r556, 0; @%p235 bra $L__BB25_208; add.s64 %rd642, %rd207, 76; cvta.to.local.u64 %rd643, %rd642; ld.local.u32 %r232, [%rd643]; add.s64 %rd708, %rd639, 2768; mov.u32 %r556, 0; $L__BB25_206: ld.const.u64 %rd645, [%rd708]; cvta.to.global.u64 %rd646, %rd645; add.s64 %rd648, %rd646, %rd96; ld.global.u32 %r420, [%rd648]; setp.eq.s32 %p236, %r420, %r232; setp.eq.s32 %p237, %r551, %r219; and.pred %p238, %p236, %p237; @%p238 bra $L__BB25_210; add.s32 %r556, %r556, 1; mul.wide.s32 %rd649, %r556, 16; add.s64 %rd651, %rd639, %rd649; add.s64 %rd708, %rd651, 2768; add.s64 %rd709, %rd651, 2776; ld.const.u64 %rd652, [%rd651+2776]; cvta.to.global.u64 %rd653, %rd652; add.s64 %rd655, %rd653, %rd91; ld.global.u32 %r551, [%rd655]; setp.ne.s32 %p239, %r551, -1; @%p239 bra $L__BB25_206; $L__BB25_208: ld.const.u32 %r433, [kernel_params+12]; add.s32 %r432, %r433, -1; setp.ge.u32 %p240, %r556, %r432; @%p240 bra $L__BB25_210; add.s64 %rd657, %rd207, 76; cvta.to.local.u64 %rd658, %rd657; ld.local.u32 %r421, [%rd658]; ld.const.u64 %rd659, [%rd708]; cvta.to.global.u64 %rd660, %rd659; add.s64 %rd662, %rd660, %rd91; st.global.u32 [%rd662], %r421; ld.const.u64 %rd663, [%rd709]; cvta.to.global.u64 %rd664, %rd663; add.s64 %rd665, %rd664, %rd91; st.global.u32 [%rd665], %r219; add.s32 %r422, %r556, 1; mul.wide.s32 %rd666, %r422, 16; add.s64 %rd668, %rd639, %rd666; ld.const.v2.u64 {%rd669, %rd670}, [%rd668+2768]; cvta.to.global.u64 %rd673, %rd669; add.s64 %rd674, %rd673, %rd91; mov.u32 %r423, -1; st.global.u32 [%rd674], %r423; cvta.to.global.u64 %rd675, %rd670; add.s64 %rd676, %rd675, %rd91; st.global.u32 [%rd676], %r423; bra.uni $L__BB25_210; $L__BB25_198: setp.eq.s32 %p229, %r551, -1; @%p229 bra $L__BB25_210; mov.u32 %r552, 0; add.s64 %rd604, %rd207, 76; cvta.to.local.u64 %rd605, %rd604; ld.local.u32 %r224, [%rd605]; mov.u64 %rd606, kernel_params; add.s64 %rd704, %rd606, 2768; $L__BB25_200: ld.const.u64 %rd607, [%rd704]; cvta.to.global.u64 %rd608, %rd607; add.s64 %rd610, %rd608, %rd96; ld.global.u32 %r415, [%rd610]; setp.eq.s32 %p230, %r415, %r224; setp.eq.s32 %p231, %r551, %r219; and.pred %p232, %p230, %p231; add.s32 %r227, %r552, 1; mul.wide.s32 %rd611, %r227, 16; add.s64 %rd613, %rd606, %rd611; add.s64 %rd74, %rd613, 2776; @%p232 bra $L__BB25_202; add.s64 %rd704, %rd613, 2768; ld.const.u64 %rd617, [%rd74]; cvta.to.global.u64 %rd618, %rd617; add.s64 %rd620, %rd618, %rd91; ld.global.u32 %r551, [%rd620]; setp.eq.s32 %p233, %r551, -1; mov.u32 %r552, %r227; @%p233 bra $L__BB25_210; bra.uni $L__BB25_200; $L__BB25_202: ld.const.v2.u64 {%rd706, %rd705}, [%rd74+-24]; $L__BB25_203: add.s32 %r552, %r552, 1; mul.wide.s32 %rd623, %r552, 16; add.s64 %rd625, %rd606, %rd623; ld.const.v2.u64 {%rd626, %rd627}, [%rd625+2768]; cvta.to.global.u64 %rd628, %rd626; add.s64 %rd630, %rd628, %rd96; ld.global.u32 %r416, [%rd630]; cvta.to.global.u64 %rd631, %rd627; add.s64 %rd632, %rd631, %rd96; ld.global.u32 %r417, [%rd632]; cvta.to.global.u64 %rd633, %rd706; add.s64 %rd634, %rd633, %rd96; st.global.u32 [%rd634], %r416; cvta.to.global.u64 %rd635, %rd705; add.s64 %rd636, %rd635, %rd96; st.global.u32 [%rd636], %r417; setp.eq.s32 %p234, %r417, -1; mov.u64 %rd705, %rd627; mov.u64 %rd706, %rd626; @%p234 bra $L__BB25_210; bra.uni $L__BB25_203; $L__BB25_2: mov.pred %p251, 0; @%p27 bra $L__BB25_101; abs.ftz.f32 %f741, %f5; setp.gt.ftz.f32 %p29, %f741, 0f17800000; and.b32 %r251, %r247, -2147483648; or.b32 %r252, %r251, 394264576; mov.b32 %f742, %r252; selp.f32 %f21, %f5, %f742, %p29; abs.ftz.f32 %f743, %f6; setp.gt.ftz.f32 %p30, %f743, 0f17800000; mov.b32 %r253, %f6; and.b32 %r254, %r253, -2147483648; or.b32 %r255, %r254, 394264576; mov.b32 %f744, %r255; selp.f32 %f22, %f6, %f744, %p30; abs.ftz.f32 %f745, %f7; setp.gt.ftz.f32 %p31, %f745, 0f17800000; mov.b32 %r256, %f7; and.b32 %r257, %r256, -2147483648; or.b32 %r258, %r257, 394264576; mov.b32 %f746, %r258; selp.f32 %f23, %f7, %f746, %p31; rcp.approx.ftz.f32 %f24, %f21; rcp.approx.ftz.f32 %f25, %f22; rcp.approx.ftz.f32 %f26, %f23; @%p28 bra $L__BB25_65; mov.u32 %r263, 1985229328; st.local.u32 [%rd1], %r263; mov.u32 %r492, -1; mov.u32 %r494, 0; mov.f32 %f2627, 0f00000000; mov.f32 %f2618, %f3; mov.f32 %f2619, %f2; mov.f32 %f2620, %f1; mov.f32 %f2621, %f23; mov.f32 %f2622, %f22; mov.f32 %f2623, %f21; mov.f32 %f2624, %f26; mov.f32 %f2625, %f25; mov.f32 %f2626, %f24; mov.u32 %r493, %r492; mov.f32 %f2628, %f2627; mov.f32 %f2629, %f4; mov.u32 %r470, %r245; mov.u32 %r496, %r492; bra.uni $L__BB25_5; $L__BB25_17: cvta.to.global.u64 %rd126, %rd102; mul.wide.s32 %rd127, %r470, 16; xor.b64 %rd128, %rd127, -16; add.s64 %rd129, %rd126, %rd128; ld.global.v4.u32 {%r473, %r275, %r276, %r277}, [%rd129]; setp.gt.s32 %p55, %r473, -1; mul.wide.s32 %rd130, %r47, 4; add.s64 %rd30, %rd1, %rd130; @%p55 bra $L__BB25_26; bra.uni $L__BB25_18; $L__BB25_26: ld.local.u32 %r470, [%rd30]; cvt.u16.u32 %rs12, %r277; and.b16 %rs11, %rs12, 127; add.s32 %r494, %r47, -1; setp.eq.s16 %p63, %rs11, 1; @%p63 bra $L__BB25_46; setp.ne.s16 %p64, %rs11, 65; @%p64 bra $L__BB25_60; add.s32 %r494, %r47, -1; setp.le.s32 %p65, %r275, %r473; @%p65 bra $L__BB25_60; mov.f32 %f1033, 0f7FC00000; add.ftz.f32 %f1034, %f1033, 0f7FC00000; sub.ftz.f32 %f1035, %f1033, 0f7FC00000; mul.ftz.f32 %f1036, %f1035, %f1034; sub.ftz.f32 %f1037, %f1036, %f1036; mul.ftz.f32 %f1038, %f2623, %f1037; fma.rn.ftz.f32 %f1039, %f2622, %f1037, %f1038; fma.rn.ftz.f32 %f113, %f2621, %f1037, %f1039; add.ftz.f32 %f1040, %f113, %f113; add.ftz.f32 %f114, %f113, %f1040; mul.ftz.f32 %f1041, %f1035, %f1035; sub.ftz.f32 %f1042, %f1041, %f1041; add.ftz.f32 %f1043, %f1042, %f1042; mul.ftz.f32 %f1044, %f2623, %f1043; fma.rn.ftz.f32 %f1045, %f2622, %f1043, %f1044; fma.rn.ftz.f32 %f115, %f2621, %f1043, %f1045; rcp.approx.ftz.f32 %f116, %f114; $L__BB25_30: cvt.s64.s32 %rd32, %r473; setp.ne.s32 %p66, %r496, -1; mov.u32 %r477, %r496; @%p66 bra $L__BB25_32; shl.b64 %rd140, %rd32, 2; add.s64 %rd141, %rd7, %rd140; ld.global.u32 %r477, [%rd141]; $L__BB25_32: shl.b64 %rd142, %rd32, 2; add.s64 %rd143, %rd8, %rd142; ld.global.u32 %r69, [%rd143]; setp.eq.s32 %p67, %r538, %r69; setp.eq.s32 %p68, %r537, %r477; and.pred %p69, %p68, %p67; @%p69 bra $L__BB25_45; cvt.s64.s32 %rd33, %r477; mul.wide.s32 %rd144, %r477, 4; add.s64 %rd145, %rd3, %rd144; ld.global.u8 %rs13, [%rd145]; and.b16 %rs14, %rs13, 16; setp.eq.s16 %p70, %rs14, 0; @%p70 bra $L__BB25_45; mul.lo.s64 %rd146, %rd33, 208; add.s64 %rd147, %rd16, %rd146; ld.global.u32 %r478, [%rd147+164]; mul.wide.u32 %rd148, %r478, 16; add.s64 %rd149, %rd20, %rd148; add.s64 %rd698, %rd149, 8; add.s64 %rd699, %rd149, 12; ld.global.u64 %rd697, [%rd149]; setp.eq.s64 %p71, %rd697, 11; @%p71 bra $L__BB25_39; $L__BB25_35: setp.eq.s64 %p72, %rd697, 0; @%p72 bra $L__BB25_37; add.s32 %r478, %r478, 2; bra.uni $L__BB25_38; $L__BB25_37: ld.global.u16 %rs15, [%rd699]; ld.global.u32 %r478, [%rd698]; setp.eq.s16 %p73, %rs15, 0; @%p73 bra $L__BB25_39; $L__BB25_38: mul.wide.u32 %rd150, %r478, 16; add.s64 %rd151, %rd20, %rd150; add.s64 %rd698, %rd151, 8; ld.global.u64 %rd697, [%rd151]; setp.ne.s64 %p74, %rd697, 11; add.s64 %rd699, %rd151, 12; @%p74 bra $L__BB25_35; $L__BB25_39: abs.ftz.f32 %f123, %f114; mul.ftz.f32 %f1046, %f123, 0f34000000; mul.ftz.f32 %f1047, %f123, 0fB4000000; setp.ltu.ftz.f32 %p76, %f113, %f1047; setp.gtu.ftz.f32 %p77, %f113, %f1046; and.pred %p78, %p76, %p77; mov.pred %p248, -1; @%p78 bra $L__BB25_43; setp.eq.ftz.f32 %p80, %f115, 0f00000000; @%p80 bra $L__BB25_43; mul.ftz.f32 %f1053, %f1043, 0f7FC00000; fma.rn.ftz.f32 %f1054, %f1043, 0f7FC00000, %f1053; fma.rn.ftz.f32 %f1055, %f1043, 0f7FC00000, %f1054; div.approx.ftz.f32 %f124, %f1055, %f115; setp.ltu.ftz.f32 %p82, %f124, %f2517; setp.gtu.ftz.f32 %p83, %f124, %f2629; or.pred %p84, %p83, %p82; @%p84 bra $L__BB25_43; setp.lt.ftz.f32 %p86, %f123, 0f219392EF; selp.f32 %f1056, 0f00000000, %f116, %p86; mul.ftz.f32 %f1057, %f113, %f1056; mov.f32 %f1058, 0f3F800000; min.ftz.f32 %f2591, %f1057, %f1058; mov.pred %p248, 0; mov.f32 %f2592, %f124; $L__BB25_43: @%p248 bra $L__BB25_45; add.s64 %rd153, %rd15, %rd142; ld.global.u32 %r291, [%rd153]; and.b32 %r292, %r291, %r1; setp.eq.s32 %p87, %r292, 0; selp.b32 %r491, %r491, 65, %p87; selp.b32 %r492, %r492, %r477, %p87; selp.b32 %r493, %r493, %r69, %p87; selp.f32 %f2627, %f2627, %f2591, %p87; selp.f32 %f2628, %f2628, %f2591, %p87; selp.f32 %f2629, %f2629, %f2592, %p87; $L__BB25_45: add.s32 %r494, %r47, -1; cvt.u32.u64 %r293, %rd32; add.s32 %r473, %r293, 1; setp.lt.s32 %p88, %r473, %r275; @%p88 bra $L__BB25_30; bra.uni $L__BB25_60; $L__BB25_18: mul.wide.s32 %rd131, %r473, 4; xor.b64 %rd132, %rd131, -4; add.s64 %rd133, %rd7, %rd132; ld.global.u32 %r496, [%rd133]; cvt.s64.s32 %rd31, %r496; mul.wide.s32 %rd134, %r496, 4; add.s64 %rd135, %rd3, %rd134; ld.global.u32 %r55, [%rd135]; and.b32 %r278, %r55, 16; setp.eq.s32 %p56, %r278, 0; @%p56 bra $L__BB25_25; bra.uni $L__BB25_19; $L__BB25_25: ld.local.u32 %r470, [%rd30]; add.s32 %r494, %r47, -1; mov.u32 %r496, -1; bra.uni $L__BB25_60; $L__BB25_46: setp.le.s32 %p89, %r275, %r473; @%p89 bra $L__BB25_60; $L__BB25_47: cvt.s64.s32 %rd43, %r473; setp.ne.s32 %p90, %r496, -1; mov.u32 %r487, %r496; @%p90 bra $L__BB25_49; shl.b64 %rd154, %rd43, 2; add.s64 %rd155, %rd7, %rd154; ld.global.u32 %r487, [%rd155]; $L__BB25_49: cvt.s64.s32 %rd679, %r473; shl.b64 %rd156, %rd679, 2; add.s64 %rd157, %rd8, %rd156; ld.global.u32 %r294, [%rd157]; cvt.s64.s32 %rd44, %r294; setp.eq.s32 %p91, %r538, %r294; setp.eq.s32 %p92, %r537, %r487; and.pred %p93, %p92, %p91; @%p93 bra $L__BB25_59; mul.wide.s32 %rd158, %r487, 4; add.s64 %rd159, %rd3, %rd158; ld.global.u8 %rs16, [%rd159]; and.b16 %rs17, %rs16, 16; setp.eq.s16 %p94, %rs17, 0; @%p94 bra $L__BB25_59; shl.b64 %rd160, %rd44, 4; add.s64 %rd161, %rd11, %rd160; ld.global.u32 %r295, [%rd161+12]; mul.wide.u32 %rd162, %r295, 12; add.s64 %rd163, %rd14, %rd162; add.s32 %r296, %r295, 1; mul.wide.u32 %rd164, %r296, 12; add.s64 %rd165, %rd14, %rd164; add.s32 %r297, %r295, 2; mul.wide.u32 %rd166, %r297, 12; add.s64 %rd167, %rd14, %rd166; ld.global.f32 %f1059, [%rd163]; sub.ftz.f32 %f144, %f1059, %f2620; ld.global.f32 %f1060, [%rd163+4]; sub.ftz.f32 %f145, %f1060, %f2619; ld.global.f32 %f1061, [%rd163+8]; sub.ftz.f32 %f146, %f1061, %f2618; ld.global.f32 %f1062, [%rd165]; sub.ftz.f32 %f1063, %f1062, %f2620; ld.global.f32 %f1064, [%rd165+4]; sub.ftz.f32 %f1065, %f1064, %f2619; ld.global.f32 %f1066, [%rd165+8]; sub.ftz.f32 %f1067, %f1066, %f2618; ld.global.f32 %f1068, [%rd167]; sub.ftz.f32 %f1069, %f1068, %f2620; ld.global.f32 %f1070, [%rd167+4]; sub.ftz.f32 %f1071, %f1070, %f2619; ld.global.f32 %f1072, [%rd167+8]; sub.ftz.f32 %f1073, %f1072, %f2618; sub.ftz.f32 %f147, %f1069, %f144; sub.ftz.f32 %f148, %f1071, %f145; sub.ftz.f32 %f149, %f1073, %f146; sub.ftz.f32 %f150, %f144, %f1063; sub.ftz.f32 %f151, %f145, %f1065; sub.ftz.f32 %f152, %f146, %f1067; sub.ftz.f32 %f1074, %f1063, %f1069; sub.ftz.f32 %f1075, %f1065, %f1071; sub.ftz.f32 %f1076, %f1067, %f1073; add.ftz.f32 %f1077, %f144, %f1069; add.ftz.f32 %f1078, %f145, %f1071; add.ftz.f32 %f1079, %f146, %f1073; mul.ftz.f32 %f1080, %f148, %f1079; mul.ftz.f32 %f1081, %f1078, %f149; sub.ftz.f32 %f1082, %f1080, %f1081; mul.ftz.f32 %f1083, %f1077, %f149; mul.ftz.f32 %f1084, %f147, %f1079; sub.ftz.f32 %f1085, %f1083, %f1084; mul.ftz.f32 %f1086, %f147, %f1078; mul.ftz.f32 %f1087, %f1077, %f148; sub.ftz.f32 %f1088, %f1086, %f1087; mul.ftz.f32 %f1089, %f2622, %f1085; fma.rn.ftz.f32 %f1090, %f2623, %f1082, %f1089; fma.rn.ftz.f32 %f153, %f2621, %f1088, %f1090; add.ftz.f32 %f1091, %f144, %f1063; add.ftz.f32 %f1092, %f145, %f1065; add.ftz.f32 %f1093, %f146, %f1067; mul.ftz.f32 %f1094, %f151, %f1093; mul.ftz.f32 %f1095, %f1092, %f152; sub.ftz.f32 %f1096, %f1094, %f1095; mul.ftz.f32 %f1097, %f1091, %f152; mul.ftz.f32 %f1098, %f150, %f1093; sub.ftz.f32 %f1099, %f1097, %f1098; mul.ftz.f32 %f1100, %f150, %f1092; mul.ftz.f32 %f1101, %f1091, %f151; sub.ftz.f32 %f1102, %f1100, %f1101; mul.ftz.f32 %f1103, %f2622, %f1099; fma.rn.ftz.f32 %f1104, %f2623, %f1096, %f1103; fma.rn.ftz.f32 %f154, %f2621, %f1102, %f1104; add.ftz.f32 %f1105, %f1063, %f1069; add.ftz.f32 %f1106, %f1065, %f1071; add.ftz.f32 %f1107, %f1067, %f1073; mul.ftz.f32 %f1108, %f1075, %f1107; mul.ftz.f32 %f1109, %f1106, %f1076; sub.ftz.f32 %f1110, %f1108, %f1109; mul.ftz.f32 %f1111, %f1105, %f1076; mul.ftz.f32 %f1112, %f1074, %f1107; sub.ftz.f32 %f1113, %f1111, %f1112; mul.ftz.f32 %f1114, %f1074, %f1106; mul.ftz.f32 %f1115, %f1105, %f1075; sub.ftz.f32 %f1116, %f1114, %f1115; mul.ftz.f32 %f1117, %f2622, %f1113; fma.rn.ftz.f32 %f1118, %f2623, %f1110, %f1117; fma.rn.ftz.f32 %f1119, %f2621, %f1116, %f1118; add.ftz.f32 %f1120, %f154, %f153; add.ftz.f32 %f155, %f1119, %f1120; abs.ftz.f32 %f156, %f155; mul.ftz.f32 %f1121, %f156, 0f34000000; min.ftz.f32 %f1122, %f154, %f1119; min.ftz.f32 %f1123, %f153, %f1122; max.ftz.f32 %f1124, %f154, %f1119; max.ftz.f32 %f1125, %f153, %f1124; mul.ftz.f32 %f1126, %f156, 0fB4000000; setp.ltu.ftz.f32 %p96, %f1123, %f1126; setp.gtu.ftz.f32 %p97, %f1125, %f1121; and.pred %p98, %p96, %p97; mov.pred %p249, -1; @%p98 bra $L__BB25_57; mov.pred %p249, -1; mul.ftz.f32 %f1127, %f151, %f149; mul.ftz.f32 %f1128, %f152, %f148; sub.ftz.f32 %f1129, %f1127, %f1128; mul.ftz.f32 %f1130, %f150, %f149; mul.ftz.f32 %f1131, %f152, %f147; sub.ftz.f32 %f1132, %f1131, %f1130; mul.ftz.f32 %f1133, %f151, %f147; mul.ftz.f32 %f1134, %f150, %f148; sub.ftz.f32 %f1135, %f1134, %f1133; add.ftz.f32 %f157, %f1129, %f1129; add.ftz.f32 %f158, %f1132, %f1132; add.ftz.f32 %f159, %f1135, %f1135; mul.ftz.f32 %f1136, %f2622, %f158; fma.rn.ftz.f32 %f1137, %f2623, %f157, %f1136; fma.rn.ftz.f32 %f160, %f2621, %f159, %f1137; setp.eq.ftz.f32 %p100, %f160, 0f00000000; @%p100 bra $L__BB25_57; mov.pred %p249, -1; mul.ftz.f32 %f1138, %f145, %f158; fma.rn.ftz.f32 %f1139, %f144, %f157, %f1138; fma.rn.ftz.f32 %f1140, %f146, %f159, %f1139; div.approx.ftz.f32 %f161, %f1140, %f160; setp.ltu.ftz.f32 %p102, %f161, %f2517; setp.gtu.ftz.f32 %p103, %f161, %f2629; or.pred %p104, %p103, %p102; @%p104 bra $L__BB25_57; abs.ftz.f32 %f2508, %f155; setp.lt.ftz.f32 %p105, %f2508, 0f219392EF; mov.f32 %f2602, 0f00000000; @%p105 bra $L__BB25_56; rcp.approx.ftz.f32 %f2602, %f155; $L__BB25_56: mul.ftz.f32 %f1142, %f153, %f2602; mov.f32 %f1143, 0f3F800000; min.ftz.f32 %f2607, %f1142, %f1143; mul.ftz.f32 %f1144, %f154, %f2602; min.ftz.f32 %f2606, %f1144, %f1143; mov.pred %p249, 0; mov.f32 %f2608, %f161; $L__BB25_57: @%p249 bra $L__BB25_59; cvt.s64.s32 %rd678, %r473; shl.b64 %rd677, %rd678, 2; cvt.u32.u64 %r298, %rd44; add.s64 %rd169, %rd15, %rd677; ld.global.u32 %r299, [%rd169]; and.b32 %r300, %r299, %r1; setp.eq.s32 %p107, %r300, 0; selp.b32 %r491, %r491, 1, %p107; selp.b32 %r492, %r492, %r487, %p107; selp.b32 %r493, %r493, %r298, %p107; selp.f32 %f2627, %f2627, %f2606, %p107; selp.f32 %f2628, %f2628, %f2607, %p107; selp.f32 %f2629, %f2629, %f2608, %p107; $L__BB25_59: add.s32 %r494, %r47, -1; cvt.u32.u64 %r301, %rd43; add.s32 %r473, %r301, 1; setp.lt.s32 %p108, %r473, %r275; @%p108 bra $L__BB25_47; bra.uni $L__BB25_60; $L__BB25_19: and.b32 %r279, %r55, 2; setp.eq.s32 %p57, %r279, 0; @%p57 bra $L__BB25_23; bra.uni $L__BB25_20; $L__BB25_23: mul.lo.s64 %rd136, %rd31, 208; add.s64 %rd137, %rd16, %rd136; ld.global.v4.f32 {%f2569, %f2570, %f2571, %f2572}, [%rd137+48]; ld.global.v4.f32 {%f2573, %f2574, %f2575, %f2576}, [%rd137+64]; ld.global.v4.f32 {%f2577, %f2578, %f2579, %f2580}, [%rd137+80]; bra.uni $L__BB25_24; $L__BB25_20: mov.f32 %f967, 0f7FC00000; add.ftz.f32 %f968, %f967, 0f7FC00000; add.ftz.f32 %f965, %f968, 0f7FC00000; mul.ftz.f32 %f969, %f965, %f965; sub.ftz.f32 %f2565, %f969, %f969; mul.ftz.f32 %f970, %f965, %f2565; fma.rn.ftz.f32 %f971, %f965, %f2565, %f970; fma.rn.ftz.f32 %f2568, %f965, %f2565, %f971; setp.neu.ftz.f32 %p58, %f2568, 0f00000000; mov.f32 %f2566, %f2565; mov.f32 %f2567, %f965; @%p58 bra $L__BB25_22; add.ftz.f32 %f979, %f968, 0f7FC00000; mul.ftz.f32 %f980, %f979, %f979; add.ftz.f32 %f2567, %f979, 0f322BCC77; mul.ftz.f32 %f981, %f2567, %f2567; sub.ftz.f32 %f2566, %f981, %f980; mul.ftz.f32 %f982, %f2567, %f979; sub.ftz.f32 %f2565, %f980, %f982; mul.ftz.f32 %f983, %f979, %f2565; fma.rn.ftz.f32 %f984, %f2567, %f2566, %f983; fma.rn.ftz.f32 %f985, %f979, %f2565, %f984; setp.eq.ftz.f32 %p59, %f985, 0f00000000; selp.f32 %f2568, 0f7F7FFFFF, %f985, %p59; $L__BB25_22: div.approx.ftz.f32 %f2569, %f2566, %f2568; div.approx.ftz.f32 %f2570, %f2565, %f2568; div.approx.ftz.f32 %f2571, %f2565, %f2568; mul.ftz.f32 %f992, %f2567, %f965; sub.ftz.f32 %f994, %f969, %f992; mul.ftz.f32 %f995, %f2567, %f2567; sub.ftz.f32 %f996, %f995, %f969; div.approx.ftz.f32 %f2573, %f994, %f2568; div.approx.ftz.f32 %f2574, %f996, %f2568; neg.ftz.f32 %f2572, %f965; mov.f32 %f2575, %f2573; mov.f32 %f2576, %f2572; mov.f32 %f2577, %f2573; mov.f32 %f2578, %f2573; mov.f32 %f2579, %f2574; mov.f32 %f2580, %f2572; $L__BB25_24: mul.ftz.f32 %f1009, %f2, %f2570; fma.rn.ftz.f32 %f1010, %f1, %f2569, %f1009; fma.rn.ftz.f32 %f1011, %f3, %f2571, %f1010; add.ftz.f32 %f2620, %f2572, %f1011; mul.ftz.f32 %f1012, %f2, %f2574; fma.rn.ftz.f32 %f1013, %f1, %f2573, %f1012; fma.rn.ftz.f32 %f1014, %f3, %f2575, %f1013; add.ftz.f32 %f2619, %f2576, %f1014; mul.ftz.f32 %f1015, %f2, %f2578; fma.rn.ftz.f32 %f1016, %f1, %f2577, %f1015; fma.rn.ftz.f32 %f1017, %f3, %f2579, %f1016; add.ftz.f32 %f2618, %f2580, %f1017; mul.ftz.f32 %f1018, %f6, %f2570; fma.rn.ftz.f32 %f1019, %f5, %f2569, %f1018; fma.rn.ftz.f32 %f1020, %f7, %f2571, %f1019; mul.ftz.f32 %f1021, %f6, %f2574; fma.rn.ftz.f32 %f1022, %f5, %f2573, %f1021; fma.rn.ftz.f32 %f1023, %f7, %f2575, %f1022; mul.ftz.f32 %f1024, %f6, %f2578; fma.rn.ftz.f32 %f1025, %f5, %f2577, %f1024; fma.rn.ftz.f32 %f1026, %f7, %f2579, %f1025; abs.ftz.f32 %f1027, %f1020; setp.gt.ftz.f32 %p60, %f1027, 0f17800000; mov.b32 %r280, %f1020; and.b32 %r281, %r280, -2147483648; or.b32 %r282, %r281, 394264576; mov.b32 %f1028, %r282; selp.f32 %f2623, %f1020, %f1028, %p60; abs.ftz.f32 %f1029, %f1023; setp.gt.ftz.f32 %p61, %f1029, 0f17800000; mov.b32 %r283, %f1023; and.b32 %r284, %r283, -2147483648; or.b32 %r285, %r284, 394264576; mov.b32 %f1030, %r285; selp.f32 %f2622, %f1023, %f1030, %p61; abs.ftz.f32 %f1031, %f1026; setp.gt.ftz.f32 %p62, %f1031, 0f17800000; mov.b32 %r286, %f1026; and.b32 %r287, %r286, -2147483648; or.b32 %r288, %r287, 394264576; mov.b32 %f1032, %r288; selp.f32 %f2621, %f1026, %f1032, %p62; rcp.approx.ftz.f32 %f2626, %f2623; rcp.approx.ftz.f32 %f2625, %f2622; rcp.approx.ftz.f32 %f2624, %f2621; add.s32 %r494, %r47, 1; mov.u32 %r289, 1985229328; st.local.u32 [%rd30+4], %r289; shl.b64 %rd138, %rd31, 2; add.s64 %rd139, %rd17, %rd138; ld.global.u32 %r470, [%rd139]; bra.uni $L__BB25_60; $L__BB25_5: setp.eq.s32 %p32, %r470, 1985229328; setp.lt.s32 %p33, %r470, 0; or.pred %p34, %p33, %p32; @%p34 bra $L__BB25_16; $L__BB25_6: mul.wide.s32 %rd123, %r470, 16; add.s64 %rd28, %rd6, %rd123; ld.global.v4.f32 {%f749, %f750, %f751, %f752}, [%rd28]; mov.b32 %r470, %f752; mov.b32 %r32, %f751; mov.b32 %r31, %f750; mov.b32 %r30, %f749; mov.b64 %rd124, {%r30, %r31}; cvt.u32.u64 %r264, %rd124; and.b32 %r265, %r264, 1024; setp.eq.s32 %p35, %r265, 0; and.b32 %r34, %r1, %r30; @%p35 bra $L__BB25_9; bra.uni $L__BB25_7; $L__BB25_9: setp.ne.s32 %p41, %r34, 0; ld.global.v4.f32 {%f892, %f893, %f894, %f895}, [%rd28+16]; ld.global.v4.f32 {%f900, %f901, %f902, %f903}, [%rd28+32]; ld.global.v4.f32 {%f908, %f909, %f910, %f911}, [%rd28+48]; sub.ftz.f32 %f916, %f892, %f2620; mul.ftz.f32 %f917, %f2626, %f916; sub.ftz.f32 %f918, %f894, %f2620; mul.ftz.f32 %f919, %f2626, %f918; sub.ftz.f32 %f920, %f900, %f2619; mul.ftz.f32 %f921, %f2625, %f920; sub.ftz.f32 %f922, %f902, %f2619; mul.ftz.f32 %f923, %f2625, %f922; sub.ftz.f32 %f924, %f908, %f2618; mul.ftz.f32 %f925, %f2624, %f924; sub.ftz.f32 %f926, %f910, %f2618; mul.ftz.f32 %f927, %f2624, %f926; min.ftz.f32 %f928, %f917, %f919; min.ftz.f32 %f929, %f921, %f923; min.ftz.f32 %f930, %f925, %f927; max.ftz.f32 %f931, %f2517, %f928; max.ftz.f32 %f932, %f929, %f930; max.ftz.f32 %f2564, %f931, %f932; max.ftz.f32 %f933, %f917, %f919; max.ftz.f32 %f934, %f921, %f923; max.ftz.f32 %f935, %f925, %f927; min.ftz.f32 %f936, %f2629, %f933; min.ftz.f32 %f937, %f934, %f935; min.ftz.f32 %f938, %f936, %f937; sub.ftz.f32 %f939, %f893, %f2620; mul.ftz.f32 %f940, %f2626, %f939; sub.ftz.f32 %f941, %f895, %f2620; mul.ftz.f32 %f942, %f2626, %f941; sub.ftz.f32 %f943, %f901, %f2619; mul.ftz.f32 %f944, %f2625, %f943; sub.ftz.f32 %f945, %f903, %f2619; mul.ftz.f32 %f946, %f2625, %f945; sub.ftz.f32 %f947, %f909, %f2618; mul.ftz.f32 %f948, %f2624, %f947; sub.ftz.f32 %f949, %f911, %f2618; mul.ftz.f32 %f950, %f2624, %f949; min.ftz.f32 %f951, %f940, %f942; min.ftz.f32 %f952, %f944, %f946; min.ftz.f32 %f953, %f948, %f950; max.ftz.f32 %f954, %f2517, %f951; max.ftz.f32 %f955, %f952, %f953; max.ftz.f32 %f2563, %f954, %f955; max.ftz.f32 %f956, %f940, %f942; max.ftz.f32 %f957, %f944, %f946; max.ftz.f32 %f958, %f948, %f950; min.ftz.f32 %f959, %f2629, %f956; min.ftz.f32 %f960, %f957, %f958; min.ftz.f32 %f961, %f959, %f960; setp.ge.ftz.f32 %p42, %f938, %f2564; and.pred %p43, %p41, %p42; selp.u32 %r269, 1, 0, %p43; setp.ltu.ftz.f32 %p44, %f961, %f2563; and.b32 %r271, %r1, %r31; setp.eq.s32 %p45, %r271, 0; or.pred %p46, %p45, %p44; selp.b32 %r272, 0, 2, %p46; or.b32 %r468, %r272, %r269; bra.uni $L__BB25_10; $L__BB25_7: setp.ne.s32 %p36, %r34, 0; ld.global.v4.f32 {%f756, %f757, %f758, %f759}, [%rd28+16]; ld.global.v4.f32 {%f764, %f765, %f766, %f767}, [%rd28+32]; ld.global.v4.f32 {%f772, %f773, %f774, %f775}, [%rd28+48]; mul.ftz.f32 %f780, %f2622, %f757; fma.rn.ftz.f32 %f781, %f2623, %f756, %f780; fma.rn.ftz.f32 %f782, %f2621, %f758, %f781; mul.ftz.f32 %f783, %f2622, %f765; fma.rn.ftz.f32 %f784, %f2623, %f764, %f783; fma.rn.ftz.f32 %f785, %f2621, %f766, %f784; mul.ftz.f32 %f786, %f2622, %f773; fma.rn.ftz.f32 %f787, %f2623, %f772, %f786; fma.rn.ftz.f32 %f788, %f2621, %f774, %f787; mul.ftz.f32 %f789, %f2619, %f757; fma.rn.ftz.f32 %f790, %f2620, %f756, %f789; fma.rn.ftz.f32 %f791, %f2618, %f758, %f790; add.ftz.f32 %f792, %f759, %f791; mul.ftz.f32 %f793, %f2619, %f765; fma.rn.ftz.f32 %f794, %f2620, %f764, %f793; fma.rn.ftz.f32 %f795, %f2618, %f766, %f794; add.ftz.f32 %f796, %f767, %f795; mul.ftz.f32 %f797, %f2619, %f773; fma.rn.ftz.f32 %f798, %f2620, %f772, %f797; fma.rn.ftz.f32 %f799, %f2618, %f774, %f798; add.ftz.f32 %f800, %f775, %f799; rcp.approx.ftz.f32 %f801, %f782; rcp.approx.ftz.f32 %f802, %f785; rcp.approx.ftz.f32 %f803, %f788; mul.ftz.f32 %f804, %f801, %f792; neg.ftz.f32 %f805, %f804; mul.ftz.f32 %f806, %f802, %f796; neg.ftz.f32 %f807, %f806; mul.ftz.f32 %f808, %f803, %f800; neg.ftz.f32 %f809, %f808; sub.ftz.f32 %f810, %f801, %f804; sub.ftz.f32 %f811, %f802, %f806; sub.ftz.f32 %f812, %f803, %f808; min.ftz.f32 %f813, %f805, %f810; min.ftz.f32 %f814, %f807, %f811; min.ftz.f32 %f815, %f809, %f812; max.ftz.f32 %f816, %f805, %f810; max.ftz.f32 %f817, %f807, %f811; max.ftz.f32 %f818, %f809, %f812; max.ftz.f32 %f819, %f2517, %f813; max.ftz.f32 %f820, %f814, %f815; max.ftz.f32 %f2564, %f819, %f820; min.ftz.f32 %f821, %f2629, %f816; min.ftz.f32 %f822, %f817, %f818; min.ftz.f32 %f823, %f821, %f822; setp.le.ftz.f32 %p37, %f2564, %f823; and.pred %p38, %p36, %p37; selp.u32 %r468, 1, 0, %p38; ld.global.v4.f32 {%f824, %f825, %f826, %f827}, [%rd28+64]; ld.global.v4.f32 {%f832, %f833, %f834, %f835}, [%rd28+80]; ld.global.v4.f32 {%f840, %f841, %f842, %f843}, [%rd28+96]; mul.ftz.f32 %f848, %f2622, %f825; fma.rn.ftz.f32 %f849, %f2623, %f824, %f848; fma.rn.ftz.f32 %f850, %f2621, %f826, %f849; mul.ftz.f32 %f851, %f2622, %f833; fma.rn.ftz.f32 %f852, %f2623, %f832, %f851; fma.rn.ftz.f32 %f853, %f2621, %f834, %f852; mul.ftz.f32 %f854, %f2622, %f841; fma.rn.ftz.f32 %f855, %f2623, %f840, %f854; fma.rn.ftz.f32 %f856, %f2621, %f842, %f855; mul.ftz.f32 %f857, %f2619, %f825; fma.rn.ftz.f32 %f858, %f2620, %f824, %f857; fma.rn.ftz.f32 %f859, %f2618, %f826, %f858; add.ftz.f32 %f860, %f827, %f859; mul.ftz.f32 %f861, %f2619, %f833; fma.rn.ftz.f32 %f862, %f2620, %f832, %f861; fma.rn.ftz.f32 %f863, %f2618, %f834, %f862; add.ftz.f32 %f864, %f835, %f863; mul.ftz.f32 %f865, %f2619, %f841; fma.rn.ftz.f32 %f866, %f2620, %f840, %f865; fma.rn.ftz.f32 %f867, %f2618, %f842, %f866; add.ftz.f32 %f868, %f843, %f867; rcp.approx.ftz.f32 %f869, %f850; rcp.approx.ftz.f32 %f870, %f853; rcp.approx.ftz.f32 %f871, %f856; mul.ftz.f32 %f872, %f869, %f860; neg.ftz.f32 %f873, %f872; mul.ftz.f32 %f874, %f870, %f864; neg.ftz.f32 %f875, %f874; mul.ftz.f32 %f876, %f871, %f868; neg.ftz.f32 %f877, %f876; sub.ftz.f32 %f878, %f869, %f872; sub.ftz.f32 %f879, %f870, %f874; sub.ftz.f32 %f880, %f871, %f876; min.ftz.f32 %f881, %f873, %f878; min.ftz.f32 %f882, %f875, %f879; min.ftz.f32 %f883, %f877, %f880; max.ftz.f32 %f884, %f873, %f878; max.ftz.f32 %f885, %f875, %f879; max.ftz.f32 %f886, %f877, %f880; max.ftz.f32 %f887, %f2517, %f881; max.ftz.f32 %f888, %f882, %f883; max.ftz.f32 %f2563, %f887, %f888; min.ftz.f32 %f889, %f2629, %f884; min.ftz.f32 %f890, %f885, %f886; min.ftz.f32 %f891, %f889, %f890; setp.gtu.ftz.f32 %p39, %f2563, %f891; @%p39 bra $L__BB25_10; and.b32 %r267, %r1, %r31; setp.eq.s32 %p40, %r267, 0; or.b32 %r268, %r468, 2; selp.b32 %r468, %r468, %r268, %p40; $L__BB25_10: mul.wide.s32 %rd125, %r494, 4; add.s64 %rd29, %rd1, %rd125; setp.eq.s32 %p47, %r468, 3; @%p47 bra $L__BB25_14; bra.uni $L__BB25_11; $L__BB25_14: setp.lt.ftz.f32 %p50, %f2563, %f2564; selp.b32 %r43, %r470, %r32, %p50; selp.b32 %r273, %r32, %r470, %p50; add.s32 %r494, %r494, 1; st.local.u32 [%rd29+4], %r273; mov.u32 %r470, %r43; bra.uni $L__BB25_15; $L__BB25_11: setp.eq.s32 %p48, %r468, 2; @%p48 bra $L__BB25_15; setp.ne.s32 %p49, %r468, 0; mov.u32 %r470, %r32; @%p49 bra $L__BB25_15; ld.local.u32 %r470, [%rd29]; add.s32 %r494, %r494, -1; $L__BB25_15: setp.ne.s32 %p51, %r470, 1985229328; setp.gt.s32 %p52, %r470, -1; and.pred %p53, %p52, %p51; @%p53 bra $L__BB25_6; $L__BB25_16: mov.u32 %r47, %r494; setp.gt.s32 %p54, %r470, -1; @%p54 bra $L__BB25_60; bra.uni $L__BB25_17; $L__BB25_60: setp.ne.s32 %p109, %r470, 1985229328; @%p109 bra $L__BB25_5; mov.u32 %r470, 1985229328; setp.lt.s32 %p110, %r494, 0; @%p110 bra $L__BB25_63; mul.wide.s32 %rd170, %r494, 4; add.s64 %rd171, %rd1, %rd170; ld.local.u32 %r470, [%rd171]; add.s32 %r494, %r494, -1; mov.u32 %r496, -1; mov.f32 %f2618, %f3; mov.f32 %f2619, %f2; mov.f32 %f2620, %f1; mov.f32 %f2621, %f23; mov.f32 %f2622, %f22; mov.f32 %f2623, %f21; mov.f32 %f2624, %f26; mov.f32 %f2625, %f25; mov.f32 %f2626, %f24; $L__BB25_63: setp.ne.s32 %p111, %r470, 1985229328; @%p111 bra $L__BB25_5; setp.ne.s32 %p251, %r493, -1; mov.u32 %r537, %r492; mov.u32 %r538, %r493; bra.uni $L__BB25_101; $L__BB25_65: mov.u32 %r308, 1985229328; st.local.u32 [%rd1], %r308; mov.u32 %r528, -1; mov.u32 %r515, 0; mov.f32 %f2627, 0f00000000; mov.f32 %f2687, %f24; mov.f32 %f2688, %f25; mov.f32 %f2689, %f26; mov.f32 %f2690, %f21; mov.f32 %f2691, %f22; mov.f32 %f2692, %f23; mov.f32 %f2693, %f1; mov.f32 %f2694, %f2; mov.f32 %f2695, %f3; mov.u32 %r529, %r528; mov.f32 %f2628, %f2627; mov.f32 %f2629, %f4; mov.u32 %r516, %r245; mov.u32 %r532, %r528; bra.uni $L__BB25_66; $L__BB25_78: cvta.to.global.u64 %rd175, %rd102; mul.wide.s32 %rd176, %r516, 16; xor.b64 %rd177, %rd176, -16; add.s64 %rd178, %rd175, %rd177; ld.global.v4.u32 {%r519, %r320, %r321, %r322}, [%rd178]; setp.gt.s32 %p135, %r519, -1; mul.wide.s32 %rd179, %r138, 4; add.s64 %rd47, %rd1, %rd179; @%p135 bra $L__BB25_82; bra.uni $L__BB25_79; $L__BB25_82: ld.local.u32 %r516, [%rd47]; add.s32 %r515, %r138, -1; and.b32 %r334, %r322, 127; setp.ne.s32 %p140, %r334, 1; setp.le.s32 %p141, %r320, %r519; or.pred %p142, %p140, %p141; @%p142 bra $L__BB25_96; $L__BB25_83: cvt.s64.s32 %rd49, %r519; setp.ne.s32 %p143, %r532, -1; mov.u32 %r523, %r532; @%p143 bra $L__BB25_85; shl.b64 %rd189, %rd49, 2; add.s64 %rd190, %rd7, %rd189; ld.global.u32 %r523, [%rd190]; $L__BB25_85: cvt.s64.s32 %rd682, %r519; shl.b64 %rd191, %rd682, 2; add.s64 %rd192, %rd8, %rd191; ld.global.u32 %r335, [%rd192]; cvt.s64.s32 %rd50, %r335; setp.eq.s32 %p144, %r538, %r335; setp.eq.s32 %p145, %r537, %r523; and.pred %p146, %p145, %p144; @%p146 bra $L__BB25_95; mul.wide.s32 %rd193, %r523, 4; add.s64 %rd194, %rd3, %rd193; ld.global.u8 %rs20, [%rd194]; and.b16 %rs21, %rs20, 16; setp.eq.s16 %p147, %rs21, 0; @%p147 bra $L__BB25_95; shl.b64 %rd195, %rd50, 4; add.s64 %rd196, %rd11, %rd195; ld.global.u32 %r336, [%rd196+12]; mul.wide.u32 %rd197, %r336, 12; add.s64 %rd198, %rd14, %rd197; add.s32 %r337, %r336, 1; mul.wide.u32 %rd199, %r337, 12; add.s64 %rd200, %rd14, %rd199; add.s32 %r338, %r336, 2; mul.wide.u32 %rd201, %r338, 12; add.s64 %rd202, %rd14, %rd201; ld.global.f32 %f1408, [%rd198]; sub.ftz.f32 %f257, %f1408, %f2693; ld.global.f32 %f1409, [%rd198+4]; sub.ftz.f32 %f258, %f1409, %f2694; ld.global.f32 %f1410, [%rd198+8]; sub.ftz.f32 %f259, %f1410, %f2695; ld.global.f32 %f1411, [%rd200]; sub.ftz.f32 %f1412, %f1411, %f2693; ld.global.f32 %f1413, [%rd200+4]; sub.ftz.f32 %f1414, %f1413, %f2694; ld.global.f32 %f1415, [%rd200+8]; sub.ftz.f32 %f1416, %f1415, %f2695; ld.global.f32 %f1417, [%rd202]; sub.ftz.f32 %f1418, %f1417, %f2693; ld.global.f32 %f1419, [%rd202+4]; sub.ftz.f32 %f1420, %f1419, %f2694; ld.global.f32 %f1421, [%rd202+8]; sub.ftz.f32 %f1422, %f1421, %f2695; sub.ftz.f32 %f260, %f1418, %f257; sub.ftz.f32 %f261, %f1420, %f258; sub.ftz.f32 %f262, %f1422, %f259; sub.ftz.f32 %f263, %f257, %f1412; sub.ftz.f32 %f264, %f258, %f1414; sub.ftz.f32 %f265, %f259, %f1416; sub.ftz.f32 %f1423, %f1412, %f1418; sub.ftz.f32 %f1424, %f1414, %f1420; sub.ftz.f32 %f1425, %f1416, %f1422; add.ftz.f32 %f1426, %f257, %f1418; add.ftz.f32 %f1427, %f258, %f1420; add.ftz.f32 %f1428, %f259, %f1422; mul.ftz.f32 %f1429, %f261, %f1428; mul.ftz.f32 %f1430, %f1427, %f262; sub.ftz.f32 %f1431, %f1429, %f1430; mul.ftz.f32 %f1432, %f1426, %f262; mul.ftz.f32 %f1433, %f260, %f1428; sub.ftz.f32 %f1434, %f1432, %f1433; mul.ftz.f32 %f1435, %f260, %f1427; mul.ftz.f32 %f1436, %f1426, %f261; sub.ftz.f32 %f1437, %f1435, %f1436; mul.ftz.f32 %f1438, %f2691, %f1434; fma.rn.ftz.f32 %f1439, %f2690, %f1431, %f1438; fma.rn.ftz.f32 %f266, %f2692, %f1437, %f1439; add.ftz.f32 %f1440, %f257, %f1412; add.ftz.f32 %f1441, %f258, %f1414; add.ftz.f32 %f1442, %f259, %f1416; mul.ftz.f32 %f1443, %f264, %f1442; mul.ftz.f32 %f1444, %f1441, %f265; sub.ftz.f32 %f1445, %f1443, %f1444; mul.ftz.f32 %f1446, %f1440, %f265; mul.ftz.f32 %f1447, %f263, %f1442; sub.ftz.f32 %f1448, %f1446, %f1447; mul.ftz.f32 %f1449, %f263, %f1441; mul.ftz.f32 %f1450, %f1440, %f264; sub.ftz.f32 %f1451, %f1449, %f1450; mul.ftz.f32 %f1452, %f2691, %f1448; fma.rn.ftz.f32 %f1453, %f2690, %f1445, %f1452; fma.rn.ftz.f32 %f267, %f2692, %f1451, %f1453; add.ftz.f32 %f1454, %f1412, %f1418; add.ftz.f32 %f1455, %f1414, %f1420; add.ftz.f32 %f1456, %f1416, %f1422; mul.ftz.f32 %f1457, %f1424, %f1456; mul.ftz.f32 %f1458, %f1455, %f1425; sub.ftz.f32 %f1459, %f1457, %f1458; mul.ftz.f32 %f1460, %f1454, %f1425; mul.ftz.f32 %f1461, %f1423, %f1456; sub.ftz.f32 %f1462, %f1460, %f1461; mul.ftz.f32 %f1463, %f1423, %f1455; mul.ftz.f32 %f1464, %f1454, %f1424; sub.ftz.f32 %f1465, %f1463, %f1464; mul.ftz.f32 %f1466, %f2691, %f1462; fma.rn.ftz.f32 %f1467, %f2690, %f1459, %f1466; fma.rn.ftz.f32 %f1468, %f2692, %f1465, %f1467; add.ftz.f32 %f1469, %f267, %f266; add.ftz.f32 %f268, %f1468, %f1469; abs.ftz.f32 %f269, %f268; mul.ftz.f32 %f1470, %f269, 0f34000000; min.ftz.f32 %f1471, %f267, %f1468; min.ftz.f32 %f1472, %f266, %f1471; max.ftz.f32 %f1473, %f267, %f1468; max.ftz.f32 %f1474, %f266, %f1473; mul.ftz.f32 %f1475, %f269, 0fB4000000; setp.ltu.ftz.f32 %p149, %f1472, %f1475; setp.gtu.ftz.f32 %p150, %f1474, %f1470; and.pred %p151, %p149, %p150; mov.pred %p250, %p148; @%p151 bra $L__BB25_93; mul.ftz.f32 %f1476, %f264, %f262; mul.ftz.f32 %f1477, %f265, %f261; sub.ftz.f32 %f1478, %f1476, %f1477; mul.ftz.f32 %f1479, %f263, %f262; mul.ftz.f32 %f1480, %f265, %f260; sub.ftz.f32 %f1481, %f1480, %f1479; mul.ftz.f32 %f1482, %f264, %f260; mul.ftz.f32 %f1483, %f263, %f261; sub.ftz.f32 %f1484, %f1483, %f1482; add.ftz.f32 %f270, %f1478, %f1478; add.ftz.f32 %f271, %f1481, %f1481; add.ftz.f32 %f272, %f1484, %f1484; mul.ftz.f32 %f1485, %f2691, %f271; fma.rn.ftz.f32 %f1486, %f2690, %f270, %f1485; fma.rn.ftz.f32 %f273, %f2692, %f272, %f1486; setp.eq.ftz.f32 %p153, %f273, 0f00000000; mov.pred %p250, %p148; @%p153 bra $L__BB25_93; mul.ftz.f32 %f1487, %f258, %f271; fma.rn.ftz.f32 %f1488, %f257, %f270, %f1487; fma.rn.ftz.f32 %f1489, %f259, %f272, %f1488; div.approx.ftz.f32 %f274, %f1489, %f273; setp.ltu.ftz.f32 %p155, %f274, %f2517; setp.gtu.ftz.f32 %p156, %f274, %f2629; or.pred %p157, %p156, %p155; mov.pred %p250, %p148; @%p157 bra $L__BB25_93; abs.ftz.f32 %f2509, %f268; setp.lt.ftz.f32 %p158, %f2509, 0f219392EF; mov.f32 %f2677, 0f00000000; @%p158 bra $L__BB25_92; rcp.approx.ftz.f32 %f2677, %f268; $L__BB25_92: mul.ftz.f32 %f1491, %f266, %f2677; mov.f32 %f1492, 0f3F800000; min.ftz.f32 %f2682, %f1491, %f1492; mul.ftz.f32 %f1493, %f267, %f2677; min.ftz.f32 %f2681, %f1493, %f1492; mov.pred %p250, 0; mov.f32 %f2683, %f274; $L__BB25_93: @%p250 bra $L__BB25_95; cvt.s64.s32 %rd681, %r519; shl.b64 %rd680, %rd681, 2; cvt.u32.u64 %r339, %rd50; add.s64 %rd204, %rd15, %rd680; ld.global.u32 %r340, [%rd204]; and.b32 %r341, %r340, %r1; setp.eq.s32 %p160, %r341, 0; selp.b32 %r491, %r491, 1, %p160; selp.b32 %r528, %r528, %r523, %p160; selp.b32 %r529, %r529, %r339, %p160; selp.f32 %f2627, %f2627, %f2681, %p160; selp.f32 %f2628, %f2628, %f2682, %p160; selp.f32 %f2629, %f2629, %f2683, %p160; $L__BB25_95: add.s32 %r515, %r138, -1; cvt.u32.u64 %r342, %rd49; add.s32 %r519, %r342, 1; setp.lt.s32 %p161, %r519, %r320; @%p161 bra $L__BB25_83; bra.uni $L__BB25_96; $L__BB25_79: mul.wide.s32 %rd180, %r519, 4; xor.b64 %rd181, %rd180, -4; add.s64 %rd182, %rd7, %rd181; ld.global.u32 %r532, [%rd182]; cvt.s64.s32 %rd48, %r532; mul.wide.s32 %rd183, %r532, 4; add.s64 %rd184, %rd3, %rd183; ld.global.u8 %rs18, [%rd184]; and.b16 %rs19, %rs18, 16; setp.eq.s16 %p136, %rs19, 0; @%p136 bra $L__BB25_81; bra.uni $L__BB25_80; $L__BB25_81: ld.local.u32 %r516, [%rd47]; add.s32 %r515, %r138, -1; mov.u32 %r532, -1; bra.uni $L__BB25_96; $L__BB25_80: mul.lo.s64 %rd185, %rd48, 208; add.s64 %rd186, %rd16, %rd185; ld.global.v4.f32 {%f1360, %f1361, %f1362, %f1363}, [%rd186+48]; ld.global.v4.f32 {%f1368, %f1369, %f1370, %f1371}, [%rd186+64]; ld.global.v4.f32 {%f1376, %f1377, %f1378, %f1379}, [%rd186+80]; mul.ftz.f32 %f1384, %f2, %f1361; fma.rn.ftz.f32 %f1385, %f1, %f1360, %f1384; fma.rn.ftz.f32 %f1386, %f3, %f1362, %f1385; add.ftz.f32 %f2693, %f1363, %f1386; mul.ftz.f32 %f1387, %f2, %f1369; fma.rn.ftz.f32 %f1388, %f1, %f1368, %f1387; fma.rn.ftz.f32 %f1389, %f3, %f1370, %f1388; add.ftz.f32 %f2694, %f1371, %f1389; mul.ftz.f32 %f1390, %f2, %f1377; fma.rn.ftz.f32 %f1391, %f1, %f1376, %f1390; fma.rn.ftz.f32 %f1392, %f3, %f1378, %f1391; add.ftz.f32 %f2695, %f1379, %f1392; mul.ftz.f32 %f1393, %f6, %f1361; fma.rn.ftz.f32 %f1394, %f5, %f1360, %f1393; fma.rn.ftz.f32 %f1395, %f7, %f1362, %f1394; mul.ftz.f32 %f1396, %f6, %f1369; fma.rn.ftz.f32 %f1397, %f5, %f1368, %f1396; fma.rn.ftz.f32 %f1398, %f7, %f1370, %f1397; mul.ftz.f32 %f1399, %f6, %f1377; fma.rn.ftz.f32 %f1400, %f5, %f1376, %f1399; fma.rn.ftz.f32 %f1401, %f7, %f1378, %f1400; abs.ftz.f32 %f1402, %f1395; setp.gt.ftz.f32 %p137, %f1402, 0f17800000; mov.b32 %r323, %f1395; and.b32 %r324, %r323, -2147483648; or.b32 %r325, %r324, 394264576; mov.b32 %f1403, %r325; selp.f32 %f2690, %f1395, %f1403, %p137; abs.ftz.f32 %f1404, %f1398; setp.gt.ftz.f32 %p138, %f1404, 0f17800000; mov.b32 %r326, %f1398; and.b32 %r327, %r326, -2147483648; or.b32 %r328, %r327, 394264576; mov.b32 %f1405, %r328; selp.f32 %f2691, %f1398, %f1405, %p138; abs.ftz.f32 %f1406, %f1401; setp.gt.ftz.f32 %p139, %f1406, 0f17800000; mov.b32 %r329, %f1401; and.b32 %r330, %r329, -2147483648; or.b32 %r331, %r330, 394264576; mov.b32 %f1407, %r331; selp.f32 %f2692, %f1401, %f1407, %p139; rcp.approx.ftz.f32 %f2687, %f2690; rcp.approx.ftz.f32 %f2688, %f2691; rcp.approx.ftz.f32 %f2689, %f2692; add.s32 %r515, %r138, 1; mov.u32 %r332, 1985229328; st.local.u32 [%rd47+4], %r332; shl.b64 %rd187, %rd48, 2; add.s64 %rd188, %rd17, %rd187; ld.global.u32 %r516, [%rd188]; bra.uni $L__BB25_96; $L__BB25_66: setp.eq.s32 %p112, %r516, 1985229328; setp.lt.s32 %p113, %r516, 0; or.pred %p114, %p113, %p112; @%p114 bra $L__BB25_77; $L__BB25_67: mul.wide.s32 %rd172, %r516, 16; add.s64 %rd45, %rd6, %rd172; ld.global.v4.f32 {%f1147, %f1148, %f1149, %f1150}, [%rd45]; mov.b32 %r516, %f1150; mov.b32 %r123, %f1149; mov.b32 %r122, %f1148; mov.b32 %r121, %f1147; mov.b64 %rd173, {%r121, %r122}; cvt.u32.u64 %r309, %rd173; and.b32 %r310, %r309, 1024; setp.eq.s32 %p115, %r310, 0; and.b32 %r125, %r1, %r121; @%p115 bra $L__BB25_70; bra.uni $L__BB25_68; $L__BB25_70: setp.ne.s32 %p121, %r125, 0; ld.global.v4.f32 {%f1290, %f1291, %f1292, %f1293}, [%rd45+16]; ld.global.v4.f32 {%f1298, %f1299, %f1300, %f1301}, [%rd45+32]; ld.global.v4.f32 {%f1306, %f1307, %f1308, %f1309}, [%rd45+48]; sub.ftz.f32 %f1314, %f1290, %f2693; mul.ftz.f32 %f1315, %f2687, %f1314; sub.ftz.f32 %f1316, %f1292, %f2693; mul.ftz.f32 %f1317, %f2687, %f1316; sub.ftz.f32 %f1318, %f1298, %f2694; mul.ftz.f32 %f1319, %f2688, %f1318; sub.ftz.f32 %f1320, %f1300, %f2694; mul.ftz.f32 %f1321, %f2688, %f1320; sub.ftz.f32 %f1322, %f1306, %f2695; mul.ftz.f32 %f1323, %f2689, %f1322; sub.ftz.f32 %f1324, %f1308, %f2695; mul.ftz.f32 %f1325, %f2689, %f1324; min.ftz.f32 %f1326, %f1315, %f1317; min.ftz.f32 %f1327, %f1319, %f1321; min.ftz.f32 %f1328, %f1323, %f1325; max.ftz.f32 %f1329, %f2517, %f1326; max.ftz.f32 %f1330, %f1327, %f1328; max.ftz.f32 %f2670, %f1329, %f1330; max.ftz.f32 %f1331, %f1315, %f1317; max.ftz.f32 %f1332, %f1319, %f1321; max.ftz.f32 %f1333, %f1323, %f1325; min.ftz.f32 %f1334, %f2629, %f1331; min.ftz.f32 %f1335, %f1332, %f1333; min.ftz.f32 %f1336, %f1334, %f1335; sub.ftz.f32 %f1337, %f1291, %f2693; mul.ftz.f32 %f1338, %f2687, %f1337; sub.ftz.f32 %f1339, %f1293, %f2693; mul.ftz.f32 %f1340, %f2687, %f1339; sub.ftz.f32 %f1341, %f1299, %f2694; mul.ftz.f32 %f1342, %f2688, %f1341; sub.ftz.f32 %f1343, %f1301, %f2694; mul.ftz.f32 %f1344, %f2688, %f1343; sub.ftz.f32 %f1345, %f1307, %f2695; mul.ftz.f32 %f1346, %f2689, %f1345; sub.ftz.f32 %f1347, %f1309, %f2695; mul.ftz.f32 %f1348, %f2689, %f1347; min.ftz.f32 %f1349, %f1338, %f1340; min.ftz.f32 %f1350, %f1342, %f1344; min.ftz.f32 %f1351, %f1346, %f1348; max.ftz.f32 %f1352, %f2517, %f1349; max.ftz.f32 %f1353, %f1350, %f1351; max.ftz.f32 %f2669, %f1352, %f1353; max.ftz.f32 %f1354, %f1338, %f1340; max.ftz.f32 %f1355, %f1342, %f1344; max.ftz.f32 %f1356, %f1346, %f1348; min.ftz.f32 %f1357, %f2629, %f1354; min.ftz.f32 %f1358, %f1355, %f1356; min.ftz.f32 %f1359, %f1357, %f1358; setp.ge.ftz.f32 %p122, %f1336, %f2670; and.pred %p123, %p121, %p122; selp.u32 %r314, 1, 0, %p123; setp.ltu.ftz.f32 %p124, %f1359, %f2669; and.b32 %r316, %r1, %r122; setp.eq.s32 %p125, %r316, 0; or.pred %p126, %p125, %p124; selp.b32 %r317, 0, 2, %p126; or.b32 %r514, %r317, %r314; bra.uni $L__BB25_71; $L__BB25_68: setp.ne.s32 %p116, %r125, 0; ld.global.v4.f32 {%f1154, %f1155, %f1156, %f1157}, [%rd45+16]; ld.global.v4.f32 {%f1162, %f1163, %f1164, %f1165}, [%rd45+32]; ld.global.v4.f32 {%f1170, %f1171, %f1172, %f1173}, [%rd45+48]; mul.ftz.f32 %f1178, %f2691, %f1155; fma.rn.ftz.f32 %f1179, %f2690, %f1154, %f1178; fma.rn.ftz.f32 %f1180, %f2692, %f1156, %f1179; mul.ftz.f32 %f1181, %f2691, %f1163; fma.rn.ftz.f32 %f1182, %f2690, %f1162, %f1181; fma.rn.ftz.f32 %f1183, %f2692, %f1164, %f1182; mul.ftz.f32 %f1184, %f2691, %f1171; fma.rn.ftz.f32 %f1185, %f2690, %f1170, %f1184; fma.rn.ftz.f32 %f1186, %f2692, %f1172, %f1185; mul.ftz.f32 %f1187, %f2694, %f1155; fma.rn.ftz.f32 %f1188, %f2693, %f1154, %f1187; fma.rn.ftz.f32 %f1189, %f2695, %f1156, %f1188; add.ftz.f32 %f1190, %f1157, %f1189; mul.ftz.f32 %f1191, %f2694, %f1163; fma.rn.ftz.f32 %f1192, %f2693, %f1162, %f1191; fma.rn.ftz.f32 %f1193, %f2695, %f1164, %f1192; add.ftz.f32 %f1194, %f1165, %f1193; mul.ftz.f32 %f1195, %f2694, %f1171; fma.rn.ftz.f32 %f1196, %f2693, %f1170, %f1195; fma.rn.ftz.f32 %f1197, %f2695, %f1172, %f1196; add.ftz.f32 %f1198, %f1173, %f1197; rcp.approx.ftz.f32 %f1199, %f1180; rcp.approx.ftz.f32 %f1200, %f1183; rcp.approx.ftz.f32 %f1201, %f1186; mul.ftz.f32 %f1202, %f1199, %f1190; neg.ftz.f32 %f1203, %f1202; mul.ftz.f32 %f1204, %f1200, %f1194; neg.ftz.f32 %f1205, %f1204; mul.ftz.f32 %f1206, %f1201, %f1198; neg.ftz.f32 %f1207, %f1206; sub.ftz.f32 %f1208, %f1199, %f1202; sub.ftz.f32 %f1209, %f1200, %f1204; sub.ftz.f32 %f1210, %f1201, %f1206; min.ftz.f32 %f1211, %f1203, %f1208; min.ftz.f32 %f1212, %f1205, %f1209; min.ftz.f32 %f1213, %f1207, %f1210; max.ftz.f32 %f1214, %f1203, %f1208; max.ftz.f32 %f1215, %f1205, %f1209; max.ftz.f32 %f1216, %f1207, %f1210; max.ftz.f32 %f1217, %f2517, %f1211; max.ftz.f32 %f1218, %f1212, %f1213; max.ftz.f32 %f2670, %f1217, %f1218; min.ftz.f32 %f1219, %f2629, %f1214; min.ftz.f32 %f1220, %f1215, %f1216; min.ftz.f32 %f1221, %f1219, %f1220; setp.le.ftz.f32 %p117, %f2670, %f1221; and.pred %p118, %p116, %p117; selp.u32 %r514, 1, 0, %p118; ld.global.v4.f32 {%f1222, %f1223, %f1224, %f1225}, [%rd45+64]; ld.global.v4.f32 {%f1230, %f1231, %f1232, %f1233}, [%rd45+80]; ld.global.v4.f32 {%f1238, %f1239, %f1240, %f1241}, [%rd45+96]; mul.ftz.f32 %f1246, %f2691, %f1223; fma.rn.ftz.f32 %f1247, %f2690, %f1222, %f1246; fma.rn.ftz.f32 %f1248, %f2692, %f1224, %f1247; mul.ftz.f32 %f1249, %f2691, %f1231; fma.rn.ftz.f32 %f1250, %f2690, %f1230, %f1249; fma.rn.ftz.f32 %f1251, %f2692, %f1232, %f1250; mul.ftz.f32 %f1252, %f2691, %f1239; fma.rn.ftz.f32 %f1253, %f2690, %f1238, %f1252; fma.rn.ftz.f32 %f1254, %f2692, %f1240, %f1253; mul.ftz.f32 %f1255, %f2694, %f1223; fma.rn.ftz.f32 %f1256, %f2693, %f1222, %f1255; fma.rn.ftz.f32 %f1257, %f2695, %f1224, %f1256; add.ftz.f32 %f1258, %f1225, %f1257; mul.ftz.f32 %f1259, %f2694, %f1231; fma.rn.ftz.f32 %f1260, %f2693, %f1230, %f1259; fma.rn.ftz.f32 %f1261, %f2695, %f1232, %f1260; add.ftz.f32 %f1262, %f1233, %f1261; mul.ftz.f32 %f1263, %f2694, %f1239; fma.rn.ftz.f32 %f1264, %f2693, %f1238, %f1263; fma.rn.ftz.f32 %f1265, %f2695, %f1240, %f1264; add.ftz.f32 %f1266, %f1241, %f1265; rcp.approx.ftz.f32 %f1267, %f1248; rcp.approx.ftz.f32 %f1268, %f1251; rcp.approx.ftz.f32 %f1269, %f1254; mul.ftz.f32 %f1270, %f1267, %f1258; neg.ftz.f32 %f1271, %f1270; mul.ftz.f32 %f1272, %f1268, %f1262; neg.ftz.f32 %f1273, %f1272; mul.ftz.f32 %f1274, %f1269, %f1266; neg.ftz.f32 %f1275, %f1274; sub.ftz.f32 %f1276, %f1267, %f1270; sub.ftz.f32 %f1277, %f1268, %f1272; sub.ftz.f32 %f1278, %f1269, %f1274; min.ftz.f32 %f1279, %f1271, %f1276; min.ftz.f32 %f1280, %f1273, %f1277; min.ftz.f32 %f1281, %f1275, %f1278; max.ftz.f32 %f1282, %f1271, %f1276; max.ftz.f32 %f1283, %f1273, %f1277; max.ftz.f32 %f1284, %f1275, %f1278; max.ftz.f32 %f1285, %f2517, %f1279; max.ftz.f32 %f1286, %f1280, %f1281; max.ftz.f32 %f2669, %f1285, %f1286; min.ftz.f32 %f1287, %f2629, %f1282; min.ftz.f32 %f1288, %f1283, %f1284; min.ftz.f32 %f1289, %f1287, %f1288; setp.gtu.ftz.f32 %p119, %f2669, %f1289; @%p119 bra $L__BB25_71; and.b32 %r312, %r1, %r122; setp.eq.s32 %p120, %r312, 0; or.b32 %r313, %r514, 2; selp.b32 %r514, %r514, %r313, %p120; $L__BB25_71: mul.wide.s32 %rd174, %r515, 4; add.s64 %rd46, %rd1, %rd174; setp.eq.s32 %p127, %r514, 3; @%p127 bra $L__BB25_75; bra.uni $L__BB25_72; $L__BB25_75: setp.lt.ftz.f32 %p130, %f2669, %f2670; selp.b32 %r134, %r516, %r123, %p130; selp.b32 %r318, %r123, %r516, %p130; add.s32 %r515, %r515, 1; st.local.u32 [%rd46+4], %r318; mov.u32 %r516, %r134; bra.uni $L__BB25_76; $L__BB25_72: setp.eq.s32 %p128, %r514, 2; @%p128 bra $L__BB25_76; setp.ne.s32 %p129, %r514, 0; mov.u32 %r516, %r123; @%p129 bra $L__BB25_76; ld.local.u32 %r516, [%rd46]; add.s32 %r515, %r515, -1; $L__BB25_76: setp.ne.s32 %p131, %r516, 1985229328; setp.gt.s32 %p132, %r516, -1; and.pred %p133, %p132, %p131; @%p133 bra $L__BB25_67; $L__BB25_77: mov.u32 %r138, %r515; setp.gt.s32 %p134, %r516, -1; @%p134 bra $L__BB25_96; bra.uni $L__BB25_78; $L__BB25_96: setp.ne.s32 %p162, %r516, 1985229328; @%p162 bra $L__BB25_66; mov.u32 %r516, 1985229328; setp.lt.s32 %p163, %r515, 0; @%p163 bra $L__BB25_99; mul.wide.s32 %rd205, %r515, 4; add.s64 %rd206, %rd1, %rd205; ld.local.u32 %r516, [%rd206]; add.s32 %r515, %r515, -1; mov.u32 %r532, -1; mov.f32 %f2687, %f24; mov.f32 %f2688, %f25; mov.f32 %f2689, %f26; mov.f32 %f2690, %f21; mov.f32 %f2691, %f22; mov.f32 %f2692, %f23; mov.f32 %f2693, %f1; mov.f32 %f2694, %f2; mov.f32 %f2695, %f3; $L__BB25_99: setp.ne.s32 %p164, %r516, 1985229328; @%p164 bra $L__BB25_66; setp.ne.s32 %p251, %r529, -1; mov.u32 %r537, %r528; mov.u32 %r538, %r529; $L__BB25_101: not.pred %p165, %p251; @%p165 bra $L__BB25_211; add.u64 %rd207, %SP, 768; add.s64 %rd208, %rd207, 88; cvta.to.local.u64 %rd209, %rd208; st.local.f32 [%rd209], %f2629; mov.u32 %r345, 0; mov.b32 %r346, %f2627; mov.b32 %r347, %f2628; add.s64 %rd210, %rd207, 64; cvta.to.local.u64 %rd211, %rd210; st.local.v4.u32 [%rd211], {%r491, %r347, %r346, %r537}; cvt.s64.s32 %rd51, %r537; mul.wide.s32 %rd212, %r537, 4; add.s64 %rd213, %rd3, %rd212; ld.global.u32 %r549, [%rd213]; add.s64 %rd214, %rd207, 56; cvta.to.local.u64 %rd215, %rd214; st.local.u32 [%rd215], %r549; add.s64 %rd216, %rd207, 60; cvta.to.local.u64 %rd217, %rd216; st.local.u32 [%rd217], %r538; add.s64 %rd218, %rd207, 80; cvta.to.local.u64 %rd219, %rd218; mov.u32 %r348, -1; st.local.u32 [%rd219], %r348; add.s64 %rd220, %rd207, 52; cvta.to.local.u64 %rd221, %rd220; st.local.u32 [%rd221], %r345; and.b32 %r181, %r549, 2; setp.eq.s32 %p166, %r181, 0; @%p166 bra $L__BB25_106; add.s64 %rd223, %rd207, 144; cvta.to.local.u64 %rd224, %rd223; mov.f32 %f1499, 0f7FC00000; add.ftz.f32 %f1500, %f1499, 0f7FC00000; add.ftz.f32 %f1497, %f1500, 0f7FC00000; st.local.v4.f32 [%rd224], {%f1497, %f1497, %f1497, %f1499}; add.s64 %rd225, %rd207, 160; cvta.to.local.u64 %rd226, %rd225; st.local.v4.f32 [%rd226], {%f1497, %f1497, %f1497, %f1499}; add.s64 %rd227, %rd207, 176; cvta.to.local.u64 %rd228, %rd227; st.local.v4.f32 [%rd228], {%f1497, %f1497, %f1497, %f1499}; mul.ftz.f32 %f1501, %f1497, %f1497; sub.ftz.f32 %f2723, %f1501, %f1501; mul.ftz.f32 %f1502, %f1497, %f2723; fma.rn.ftz.f32 %f1503, %f1497, %f2723, %f1502; fma.rn.ftz.f32 %f2726, %f1497, %f2723, %f1503; setp.neu.ftz.f32 %p167, %f2726, 0f00000000; mov.f32 %f2724, %f2723; mov.f32 %f2725, %f1497; @%p167 bra $L__BB25_105; add.ftz.f32 %f1511, %f1500, 0f7FC00000; mul.ftz.f32 %f1512, %f1511, %f1511; add.ftz.f32 %f2725, %f1511, 0f322BCC77; mul.ftz.f32 %f1513, %f2725, %f2725; sub.ftz.f32 %f2724, %f1513, %f1512; mul.ftz.f32 %f1514, %f2725, %f1511; sub.ftz.f32 %f2723, %f1512, %f1514; mul.ftz.f32 %f1515, %f1511, %f2723; fma.rn.ftz.f32 %f1516, %f2725, %f2724, %f1515; fma.rn.ftz.f32 %f1517, %f1511, %f2723, %f1516; setp.eq.ftz.f32 %p168, %f1517, 0f00000000; selp.f32 %f2726, 0f7F7FFFFF, %f1517, %p168; $L__BB25_105: mul.ftz.f32 %f1521, %f2725, %f1497; sub.ftz.f32 %f1523, %f1501, %f1521; mul.ftz.f32 %f1524, %f2725, %f2725; sub.ftz.f32 %f1525, %f1524, %f1501; div.approx.ftz.f32 %f1526, %f2724, %f2726; neg.ftz.f32 %f1527, %f1497; div.approx.ftz.f32 %f1528, %f2723, %f2726; div.approx.ftz.f32 %f1529, %f2723, %f2726; add.s64 %rd230, %rd207, 192; cvta.to.local.u64 %rd231, %rd230; st.local.v4.f32 [%rd231], {%f1526, %f1528, %f1529, %f1527}; div.approx.ftz.f32 %f1530, %f1525, %f2726; add.s64 %rd232, %rd207, 208; cvta.to.local.u64 %rd233, %rd232; div.approx.ftz.f32 %f1531, %f1523, %f2726; st.local.v4.f32 [%rd233], {%f1531, %f1530, %f1531, %f1527}; add.s64 %rd234, %rd207, 224; cvta.to.local.u64 %rd235, %rd234; st.local.v4.f32 [%rd235], {%f1531, %f1531, %f1530, %f1527}; $L__BB25_106: neg.ftz.f32 %f1532, %f5; add.s64 %rd237, %rd207, 36; cvta.to.local.u64 %rd238, %rd237; st.local.f32 [%rd238], %f1532; add.s64 %rd239, %rd207, 40; cvta.to.local.u64 %rd240, %rd239; neg.ftz.f32 %f1533, %f6; st.local.f32 [%rd240], %f1533; add.s64 %rd241, %rd207, 44; cvta.to.local.u64 %rd242, %rd241; neg.ftz.f32 %f1534, %f7; st.local.f32 [%rd242], %f1534; and.b32 %r349, %r491, 6; setp.eq.s32 %p169, %r349, 0; mul.lo.s64 %rd243, %rd51, 208; add.s64 %rd52, %rd16, %rd243; @%p169 bra $L__BB25_144; and.b32 %r182, %r549, 4; setp.ne.s32 %p170, %r182, 0; mov.f32 %f2742, %f2629; mov.f32 %f2743, %f5; mov.f32 %f2744, %f6; mov.f32 %f2745, %f7; mov.f32 %f2746, %f1; mov.f32 %f2747, %f2; mov.f32 %f2748, %f3; @%p170 bra $L__BB25_113; @%p166 bra $L__BB25_110; bra.uni $L__BB25_109; $L__BB25_110: ld.global.v4.f32 {%f2727, %f2728, %f2729, %f2730}, [%rd52+48]; ld.global.v4.f32 {%f2731, %f2732, %f2733, %f2734}, [%rd52+64]; ld.global.v4.f32 {%f2735, %f2736, %f2737, %f2738}, [%rd52+80]; bra.uni $L__BB25_111; $L__BB25_144: and.b32 %r376, %r491, 8; setp.eq.s32 %p194, %r376, 0; @%p194 bra $L__BB25_165; bra.uni $L__BB25_145; $L__BB25_165: setp.eq.s32 %p207, %r491, 1; @%p207 bra $L__BB25_167; bra.uni $L__BB25_166; $L__BB25_167: cvt.s64.s32 %rd70, %r538; mul.wide.s32 %rd407, %r538, 16; add.s64 %rd408, %rd11, %rd407; add.s64 %rd71, %rd408, 12; ld.global.u32 %r393, [%rd408+12]; mul.wide.u32 %rd409, %r393, 12; add.s64 %rd410, %rd14, %rd409; add.s32 %r394, %r393, 1; mul.wide.u32 %rd411, %r394, 12; add.s64 %rd412, %rd14, %rd411; add.s32 %r395, %r393, 2; mul.wide.u32 %rd413, %r395, 12; add.s64 %rd414, %rd14, %rd413; and.b32 %r396, %r549, 8; setp.eq.s32 %p208, %r396, 0; ld.global.f32 %f2092, [%rd414]; ld.global.f32 %f605, [%rd410]; sub.ftz.f32 %f606, %f2092, %f605; ld.global.f32 %f2093, [%rd414+4]; ld.global.f32 %f607, [%rd410+4]; sub.ftz.f32 %f608, %f2093, %f607; ld.global.f32 %f2094, [%rd414+8]; ld.global.f32 %f609, [%rd410+8]; sub.ftz.f32 %f610, %f2094, %f609; ld.global.f32 %f2095, [%rd412]; sub.ftz.f32 %f611, %f2095, %f605; ld.global.f32 %f2096, [%rd412+4]; sub.ftz.f32 %f612, %f2096, %f607; ld.global.f32 %f2097, [%rd412+8]; sub.ftz.f32 %f613, %f2097, %f609; mul.ftz.f32 %f614, %f613, %f608; mul.ftz.f32 %f615, %f612, %f610; mul.ftz.f32 %f616, %f611, %f610; mul.ftz.f32 %f617, %f613, %f606; mul.ftz.f32 %f618, %f612, %f606; mul.ftz.f32 %f619, %f611, %f608; @%p208 bra $L__BB25_169; bra.uni $L__BB25_168; $L__BB25_169: sub.ftz.f32 %f2105, %f615, %f614; sub.ftz.f32 %f2106, %f617, %f616; mul.ftz.f32 %f2107, %f2106, %f2106; fma.rn.ftz.f32 %f2108, %f2105, %f2105, %f2107; sub.ftz.f32 %f2109, %f619, %f618; fma.rn.ftz.f32 %f2110, %f2109, %f2109, %f2108; sqrt.approx.ftz.f32 %f2111, %f2110; div.approx.ftz.f32 %f2813, %f2109, %f2111; div.approx.ftz.f32 %f2812, %f2106, %f2111; div.approx.ftz.f32 %f2811, %f2105, %f2111; bra.uni $L__BB25_170; $L__BB25_145: cvt.s64.s32 %rd60, %r538; mul.wide.s32 %rd359, %r538, 4; add.s64 %rd360, %rd358, %rd359; ld.global.u32 %r377, [%rd360]; add.s64 %rd362, %rd207, 48; cvta.to.local.u64 %rd363, %rd362; st.local.u32 [%rd363], %r377; fma.rn.ftz.f32 %f570, %f5, %f2629, %f1; fma.rn.ftz.f32 %f571, %f6, %f2629, %f2; fma.rn.ftz.f32 %f572, %f7, %f2629, %f3; cvta.to.local.u64 %rd364, %rd207; st.local.v2.f32 [%rd364], {%f570, %f571}; or.b64 %rd365, %rd207, 8; cvta.to.local.u64 %rd366, %rd365; st.local.f32 [%rd366], %f572; and.b32 %r378, %r491, 64; setp.eq.s32 %p195, %r378, 0; @%p195 bra $L__BB25_159; bra.uni $L__BB25_146; $L__BB25_159: shl.b64 %rd378, %rd60, 4; add.s64 %rd379, %rd22, %rd378; ld.global.v4.f32 {%f2808, %f2809, %f2810, %f2017}, [%rd379]; bra.uni $L__BB25_160; $L__BB25_109: add.s64 %rd245, %rd207, 192; cvta.to.local.u64 %rd246, %rd245; ld.local.v4.f32 {%f2727, %f2728, %f2729, %f2730}, [%rd246]; add.s64 %rd247, %rd207, 208; cvta.to.local.u64 %rd248, %rd247; ld.local.v4.f32 {%f2731, %f2732, %f2733, %f2734}, [%rd248]; add.s64 %rd249, %rd207, 224; cvta.to.local.u64 %rd250, %rd249; ld.local.v4.f32 {%f2735, %f2736, %f2737, %f2738}, [%rd250]; $L__BB25_111: mul.ftz.f32 %f1559, %f2, %f2728; fma.rn.ftz.f32 %f1560, %f1, %f2727, %f1559; fma.rn.ftz.f32 %f1561, %f3, %f2729, %f1560; add.ftz.f32 %f2746, %f2730, %f1561; mul.ftz.f32 %f1562, %f2, %f2732; fma.rn.ftz.f32 %f1563, %f1, %f2731, %f1562; fma.rn.ftz.f32 %f1564, %f3, %f2733, %f1563; add.ftz.f32 %f2747, %f2734, %f1564; mul.ftz.f32 %f1565, %f2, %f2736; fma.rn.ftz.f32 %f1566, %f1, %f2735, %f1565; fma.rn.ftz.f32 %f1567, %f3, %f2737, %f1566; add.ftz.f32 %f2748, %f2738, %f1567; mul.ftz.f32 %f1568, %f5, %f2629; mul.ftz.f32 %f1569, %f6, %f2629; mul.ftz.f32 %f1570, %f1569, %f2728; fma.rn.ftz.f32 %f1571, %f1568, %f2727, %f1570; mul.ftz.f32 %f1572, %f7, %f2629; fma.rn.ftz.f32 %f2743, %f1572, %f2729, %f1571; mul.ftz.f32 %f1573, %f1569, %f2732; fma.rn.ftz.f32 %f1574, %f1568, %f2731, %f1573; fma.rn.ftz.f32 %f2744, %f1572, %f2733, %f1574; mul.ftz.f32 %f1575, %f1569, %f2736; fma.rn.ftz.f32 %f1576, %f1568, %f2735, %f1575; fma.rn.ftz.f32 %f2745, %f1572, %f2737, %f1576; mul.ftz.f32 %f1577, %f2744, %f2744; fma.rn.ftz.f32 %f1578, %f2743, %f2743, %f1577; fma.rn.ftz.f32 %f1579, %f2745, %f2745, %f1578; sqrt.approx.ftz.f32 %f2742, %f1579; setp.neu.ftz.f32 %p172, %f2742, 0f00000000; @%p172 bra $L__BB25_112; bra.uni $L__BB25_113; $L__BB25_112: div.approx.ftz.f32 %f2745, %f2745, %f2742; div.approx.ftz.f32 %f2744, %f2744, %f2742; div.approx.ftz.f32 %f2743, %f2743, %f2742; $L__BB25_113: cvta.to.global.u64 %rd694, %rd110; mul.wide.s32 %rd252, %r538, 16; add.s64 %rd253, %rd694, %rd252; shr.s32 %r350, %r491, 7; ld.global.u32 %r351, [%rd253+4]; add.s32 %r183, %r350, %r351; add.s32 %r352, %r183, -1; max.s32 %r184, %r352, %r351; add.s32 %r353, %r183, 2; ld.global.u32 %r354, [%rd253+8]; add.s32 %r355, %r351, %r354; add.s32 %r356, %r355, -1; min.s32 %r185, %r353, %r356; and.b32 %r357, %r491, 64; setp.eq.s32 %p173, %r357, 0; mul.wide.s32 %rd254, %r184, 16; add.s64 %rd54, %rd23, %rd254; mul.wide.s32 %rd255, %r183, 16; add.s64 %rd55, %rd23, %rd255; mul.wide.s32 %rd256, %r185, 16; add.s64 %rd56, %rd23, %rd256; @%p173 bra $L__BB25_127; bra.uni $L__BB25_114; $L__BB25_127: ld.global.v4.f32 {%f2773, %f2774, %f2775, %f1665}, [%rd54]; ld.global.v4.f32 {%f2776, %f2777, %f2778, %f1669}, [%rd55]; ld.global.v4.f32 {%f2779, %f2780, %f2781, %f1673}, [%rd55+16]; ld.global.v4.f32 {%f2782, %f2783, %f2784, %f1677}, [%rd56]; bra.uni $L__BB25_128; $L__BB25_114: add.s64 %rd258, %rd207, 84; cvta.to.local.u64 %rd259, %rd258; ld.global.u32 %r186, [%rd52+148]; ld.global.u32 %r187, [%rd52+152]; shl.b32 %r358, %r187, 1; cvt.rn.f32.s32 %f1580, %r358; ld.local.f32 %f1581, [%rd259]; mul.ftz.f32 %f394, %f1581, %f1580; cvt.rzi.ftz.s32.f32 %r359, %f394; add.s32 %r360, %r358, -1; min.s32 %r188, %r359, %r360; ld.global.u32 %r540, [%rd52+164]; mul.wide.u32 %rd260, %r540, 16; add.s64 %rd261, %rd20, %rd260; ld.global.u16 %rs22, [%rd261+12]; ld.global.u32 %r539, [%rd261+8]; ld.global.u64 %rd700, [%rd261]; setp.eq.s64 %p174, %rd700, 11; @%p174 bra $L__BB25_119; $L__BB25_115: setp.eq.s64 %p175, %rd700, 0; @%p175 bra $L__BB25_117; add.s32 %r540, %r540, 2; bra.uni $L__BB25_118; $L__BB25_117: mov.u32 %r543, -1; setp.eq.s16 %p176, %rs22, 0; mov.u32 %r540, %r539; @%p176 bra $L__BB25_120; $L__BB25_118: mul.wide.u32 %rd262, %r540, 16; add.s64 %rd263, %rd20, %rd262; ld.global.u16 %rs22, [%rd263+12]; ld.global.u32 %r539, [%rd263+8]; ld.global.u64 %rd700, [%rd263]; setp.ne.s64 %p177, %rd700, 11; @%p177 bra $L__BB25_115; $L__BB25_119: setp.eq.s16 %p178, %rs22, 0; selp.b32 %r543, -1, %r539, %p178; $L__BB25_120: setp.eq.s32 %p179, %r188, %r187; @%p179 bra $L__BB25_122; bra.uni $L__BB25_121; $L__BB25_122: ld.global.v4.f32 {%f2749, %f2750, %f2751, %f1601}, [%rd54]; ld.global.v4.f32 {%f2752, %f2753, %f2754, %f1605}, [%rd55]; ld.global.v4.f32 {%f2755, %f2756, %f2757, %f1609}, [%rd55+16]; ld.global.v4.f32 {%f2758, %f2759, %f2760, %f1613}, [%rd56]; bra.uni $L__BB25_123; $L__BB25_121: shr.s32 %r440, %r491, 7; add.s32 %r439, %r440, %r351; setp.gt.s32 %p180, %r188, %r187; selp.b32 %r362, -1, 0, %p180; add.s32 %r363, %r188, %r362; mad.lo.s32 %r364, %r363, %r186, %r543; add.s32 %r365, %r364, %r184; mul.wide.s32 %rd264, %r365, 16; add.s64 %rd265, %rd25, %rd264; ld.global.v4.f32 {%f2749, %f2750, %f2751, %f1585}, [%rd265]; add.s32 %r366, %r364, %r439; mul.wide.s32 %rd266, %r366, 16; add.s64 %rd267, %rd25, %rd266; ld.global.v4.f32 {%f2752, %f2753, %f2754, %f1589}, [%rd267]; ld.global.v4.f32 {%f2755, %f2756, %f2757, %f1593}, [%rd267+16]; add.s32 %r367, %r364, %r185; mul.wide.s32 %rd268, %r367, 16; add.s64 %rd269, %rd25, %rd268; ld.global.v4.f32 {%f2758, %f2759, %f2760, %f1597}, [%rd269]; $L__BB25_123: add.s32 %r199, %r188, 1; setp.eq.s32 %p181, %r199, %r187; @%p181 bra $L__BB25_125; bra.uni $L__BB25_124; $L__BB25_125: ld.global.v4.f32 {%f2761, %f2762, %f2763, %f1633}, [%rd54]; ld.global.v4.f32 {%f2764, %f2765, %f2766, %f1637}, [%rd55]; ld.global.v4.f32 {%f2767, %f2768, %f2769, %f1641}, [%rd55+16]; ld.global.v4.f32 {%f2770, %f2771, %f2772, %f1645}, [%rd56]; bra.uni $L__BB25_126; $L__BB25_124: shr.s32 %r442, %r491, 7; add.s32 %r441, %r442, %r351; setp.ge.s32 %p182, %r188, %r187; selp.b32 %r368, -1, 0, %p182; add.s32 %r369, %r199, %r368; mad.lo.s32 %r370, %r369, %r186, %r543; add.s32 %r371, %r370, %r184; mul.wide.s32 %rd270, %r371, 16; add.s64 %rd271, %rd25, %rd270; ld.global.v4.f32 {%f2761, %f2762, %f2763, %f1617}, [%rd271]; add.s32 %r372, %r370, %r441; mul.wide.s32 %rd272, %r372, 16; add.s64 %rd273, %rd25, %rd272; ld.global.v4.f32 {%f2764, %f2765, %f2766, %f1621}, [%rd273]; ld.global.v4.f32 {%f2767, %f2768, %f2769, %f1625}, [%rd273+16]; add.s32 %r373, %r370, %r185; mul.wide.s32 %rd274, %r373, 16; add.s64 %rd275, %rd25, %rd274; ld.global.v4.f32 {%f2770, %f2771, %f2772, %f1629}, [%rd275]; $L__BB25_126: shl.b32 %r443, %r187, 1; cvt.rn.f32.s32 %f2511, %r443; mul.ftz.f32 %f2510, %f1581, %f2511; cvt.rn.f32.s32 %f1646, %r188; sub.ftz.f32 %f1647, %f2510, %f1646; mov.f32 %f1648, 0f3F800000; sub.ftz.f32 %f1649, %f1648, %f1647; mul.ftz.f32 %f1650, %f1647, %f2761; fma.rn.ftz.f32 %f2773, %f1649, %f2749, %f1650; mul.ftz.f32 %f1651, %f1647, %f2762; fma.rn.ftz.f32 %f2774, %f1649, %f2750, %f1651; mul.ftz.f32 %f1652, %f1647, %f2763; fma.rn.ftz.f32 %f2775, %f1649, %f2751, %f1652; mul.ftz.f32 %f1653, %f1647, %f2764; fma.rn.ftz.f32 %f2776, %f1649, %f2752, %f1653; mul.ftz.f32 %f1654, %f1647, %f2765; fma.rn.ftz.f32 %f2777, %f1649, %f2753, %f1654; mul.ftz.f32 %f1655, %f1647, %f2766; fma.rn.ftz.f32 %f2778, %f1649, %f2754, %f1655; mul.ftz.f32 %f1656, %f1647, %f2767; fma.rn.ftz.f32 %f2779, %f1649, %f2755, %f1656; mul.ftz.f32 %f1657, %f1647, %f2768; fma.rn.ftz.f32 %f2780, %f1649, %f2756, %f1657; mul.ftz.f32 %f1658, %f1647, %f2769; fma.rn.ftz.f32 %f2781, %f1649, %f2757, %f1658; mul.ftz.f32 %f1659, %f1647, %f2770; fma.rn.ftz.f32 %f2782, %f1649, %f2758, %f1659; mul.ftz.f32 %f1660, %f1647, %f2771; fma.rn.ftz.f32 %f2783, %f1649, %f2759, %f1660; mul.ftz.f32 %f1661, %f1647, %f2772; fma.rn.ftz.f32 %f2784, %f1649, %f2760, %f1661; $L__BB25_128: fma.rn.ftz.f32 %f2802, %f2742, %f2743, %f2746; fma.rn.ftz.f32 %f2803, %f2742, %f2744, %f2747; fma.rn.ftz.f32 %f2804, %f2742, %f2745, %f2748; mov.f32 %f1678, 0f3F800000; sub.ftz.f32 %f506, %f1678, %f2628; mul.ftz.f32 %f507, %f506, %f506; add.ftz.f32 %f1679, %f506, %f506; mul.ftz.f32 %f1680, %f2628, %f1679; sub.ftz.f32 %f1681, %f1680, %f507; mul.ftz.f32 %f1682, %f2628, 0f40400000; add.ftz.f32 %f508, %f1682, 0fC0A00000; add.ftz.f32 %f1683, %f2628, %f2628; mul.ftz.f32 %f1684, %f1683, %f508; fma.rn.ftz.f32 %f1685, %f2628, %f1682, %f1684; add.ftz.f32 %f1686, %f1682, 0f40000000; mul.ftz.f32 %f1687, %f1679, %f1686; mul.ftz.f32 %f509, %f506, 0f40400000; mul.ftz.f32 %f1688, %f506, %f509; sub.ftz.f32 %f1689, %f1687, %f1688; mul.ftz.f32 %f1690, %f506, 0fC0000000; mul.ftz.f32 %f510, %f2628, %f2628; fma.rn.ftz.f32 %f1691, %f2628, %f1690, %f510; mul.ftz.f32 %f1692, %f2776, %f1685; mul.ftz.f32 %f1693, %f2777, %f1685; mul.ftz.f32 %f1694, %f2778, %f1685; fma.rn.ftz.f32 %f1695, %f2773, %f1681, %f1692; fma.rn.ftz.f32 %f1696, %f2774, %f1681, %f1693; fma.rn.ftz.f32 %f1697, %f2775, %f1681, %f1694; fma.rn.ftz.f32 %f1698, %f2779, %f1689, %f1695; fma.rn.ftz.f32 %f1699, %f2780, %f1689, %f1696; fma.rn.ftz.f32 %f1700, %f2781, %f1689, %f1697; fma.rn.ftz.f32 %f1701, %f2782, %f1691, %f1698; fma.rn.ftz.f32 %f1702, %f2783, %f1691, %f1699; fma.rn.ftz.f32 %f1703, %f2784, %f1691, %f1700; mul.ftz.f32 %f2797, %f1701, 0f3F000000; mul.ftz.f32 %f2799, %f1702, 0f3F000000; mul.ftz.f32 %f2798, %f1703, 0f3F000000; and.b32 %r200, %r491, 4; setp.eq.s32 %p183, %r200, 0; @%p183 bra $L__BB25_130; bra.uni $L__BB25_129; $L__BB25_130: mul.ftz.f32 %f1762, %f2628, %f506; mul.ftz.f32 %f1763, %f506, %f1762; fma.rn.ftz.f32 %f1764, %f510, %f508, 0f40000000; add.ftz.f32 %f1765, %f509, 0fC0A00000; fma.rn.ftz.f32 %f1766, %f507, %f1765, 0f40000000; mul.ftz.f32 %f1767, %f2628, %f1762; mul.ftz.f32 %f1768, %f2773, %f1763; mul.ftz.f32 %f1769, %f2774, %f1763; mul.ftz.f32 %f1770, %f2775, %f1763; mul.ftz.f32 %f1771, %f2776, %f1764; mul.ftz.f32 %f1772, %f2777, %f1764; mul.ftz.f32 %f1773, %f2778, %f1764; sub.ftz.f32 %f1774, %f1771, %f1768; sub.ftz.f32 %f1775, %f1772, %f1769; sub.ftz.f32 %f1776, %f1773, %f1770; fma.rn.ftz.f32 %f1777, %f2779, %f1766, %f1774; fma.rn.ftz.f32 %f1778, %f2780, %f1766, %f1775; fma.rn.ftz.f32 %f1779, %f2781, %f1766, %f1776; mul.ftz.f32 %f1780, %f2782, %f1767; mul.ftz.f32 %f1781, %f2783, %f1767; mul.ftz.f32 %f1782, %f2784, %f1767; sub.ftz.f32 %f1783, %f1777, %f1780; sub.ftz.f32 %f1784, %f1778, %f1781; sub.ftz.f32 %f1785, %f1779, %f1782; mul.ftz.f32 %f517, %f1783, 0f3F000000; mul.ftz.f32 %f518, %f1784, 0f3F000000; mul.ftz.f32 %f519, %f1785, 0f3F000000; setp.eq.ftz.f32 %p184, %f2802, %f517; setp.eq.ftz.f32 %p185, %f2803, %f518; and.pred %p186, %p184, %p185; setp.eq.ftz.f32 %p187, %f2804, %f519; and.pred %p188, %p187, %p186; mov.f32 %f2801, %f5; mov.f32 %f2800, %f6; mov.f32 %f2788, %f7; @%p188 bra $L__BB25_132; sub.ftz.f32 %f1786, %f2802, %f517; sub.ftz.f32 %f1787, %f2803, %f518; mul.ftz.f32 %f1788, %f1787, %f1787; fma.rn.ftz.f32 %f1789, %f1786, %f1786, %f1788; sub.ftz.f32 %f1790, %f2804, %f519; fma.rn.ftz.f32 %f1791, %f1790, %f1790, %f1789; sqrt.approx.ftz.f32 %f1792, %f1791; div.approx.ftz.f32 %f2788, %f1790, %f1792; div.approx.ftz.f32 %f2800, %f1787, %f1792; div.approx.ftz.f32 %f2801, %f1786, %f1792; $L__BB25_132: mov.u32 %r446, 0; or.b64 %rd284, %rd207, 12; cvta.to.local.u64 %rd285, %rd284; st.local.f32 [%rd285], %f2801; add.s64 %rd286, %rd207, 16; cvta.to.local.u64 %rd287, %rd286; st.local.f32 [%rd287], %f2800; add.s64 %rd288, %rd207, 20; cvta.to.local.u64 %rd289, %rd288; st.local.f32 [%rd289], %f2788; add.s64 %rd290, %rd207, 72; cvta.to.local.u64 %rd291, %rd290; st.local.u32 [%rd291], %r446; bra.uni $L__BB25_133; $L__BB25_129: mul.ftz.f32 %f1704, %f2799, %f2799; fma.rn.ftz.f32 %f1705, %f2797, %f2797, %f1704; fma.rn.ftz.f32 %f1706, %f2798, %f2798, %f1705; sqrt.approx.ftz.f32 %f1707, %f1706; div.approx.ftz.f32 %f1708, %f2797, %f1707; div.approx.ftz.f32 %f1709, %f2799, %f1707; div.approx.ftz.f32 %f1710, %f2798, %f1707; mul.ftz.f32 %f1711, %f1709, %f2745; mul.ftz.f32 %f1712, %f1710, %f2744; sub.ftz.f32 %f1713, %f1712, %f1711; mul.ftz.f32 %f1714, %f1710, %f2743; mul.ftz.f32 %f1715, %f1708, %f2745; sub.ftz.f32 %f1716, %f1715, %f1714; mul.ftz.f32 %f1717, %f1708, %f2744; mul.ftz.f32 %f1718, %f1709, %f2743; sub.ftz.f32 %f1719, %f1718, %f1717; mul.ftz.f32 %f1720, %f1716, %f1716; fma.rn.ftz.f32 %f1721, %f1713, %f1713, %f1720; fma.rn.ftz.f32 %f1722, %f1719, %f1719, %f1721; sqrt.approx.ftz.f32 %f1723, %f1722; div.approx.ftz.f32 %f1724, %f1713, %f1723; div.approx.ftz.f32 %f1725, %f1716, %f1723; div.approx.ftz.f32 %f1726, %f1719, %f1723; mul.ftz.f32 %f1727, %f2627, %f2627; sub.ftz.f32 %f1729, %f1678, %f1727; mov.f32 %f1730, 0f00000000; max.ftz.f32 %f1731, %f1729, %f1730; sqrt.approx.ftz.f32 %f1732, %f1731; mul.ftz.f32 %f1733, %f1724, %f2627; mul.ftz.f32 %f1734, %f1725, %f2627; mul.ftz.f32 %f1735, %f1726, %f2627; mul.ftz.f32 %f1736, %f1709, %f1726; mul.ftz.f32 %f1737, %f1710, %f1725; sub.ftz.f32 %f1738, %f1736, %f1737; mul.ftz.f32 %f1739, %f1710, %f1724; mul.ftz.f32 %f1740, %f1708, %f1726; sub.ftz.f32 %f1741, %f1739, %f1740; mul.ftz.f32 %f1742, %f1708, %f1725; mul.ftz.f32 %f1743, %f1709, %f1724; sub.ftz.f32 %f1744, %f1742, %f1743; mul.ftz.f32 %f1745, %f1741, %f1741; fma.rn.ftz.f32 %f1746, %f1738, %f1738, %f1745; fma.rn.ftz.f32 %f1747, %f1744, %f1744, %f1746; sqrt.approx.ftz.f32 %f1748, %f1747; div.approx.ftz.f32 %f1749, %f1738, %f1748; div.approx.ftz.f32 %f1750, %f1741, %f1748; div.approx.ftz.f32 %f1751, %f1744, %f1748; mul.ftz.f32 %f1752, %f1732, %f1749; mul.ftz.f32 %f1753, %f1732, %f1750; mul.ftz.f32 %f1754, %f1732, %f1751; sub.ftz.f32 %f1755, %f1733, %f1752; sub.ftz.f32 %f1756, %f1734, %f1753; sub.ftz.f32 %f1757, %f1735, %f1754; mul.ftz.f32 %f1758, %f1756, %f1756; fma.rn.ftz.f32 %f1759, %f1755, %f1755, %f1758; fma.rn.ftz.f32 %f1760, %f1757, %f1757, %f1759; sqrt.approx.ftz.f32 %f1761, %f1760; div.approx.ftz.f32 %f2801, %f1755, %f1761; div.approx.ftz.f32 %f2800, %f1756, %f1761; div.approx.ftz.f32 %f2788, %f1757, %f1761; or.b64 %rd277, %rd207, 12; cvta.to.local.u64 %rd278, %rd277; st.local.f32 [%rd278], %f2801; add.s64 %rd279, %rd207, 16; cvta.to.local.u64 %rd280, %rd279; st.local.f32 [%rd280], %f2800; add.s64 %rd281, %rd207, 20; cvta.to.local.u64 %rd282, %rd281; st.local.f32 [%rd282], %f2788; $L__BB25_133: and.b32 %r444, %r549, 4; setp.ne.s32 %p246, %r444, 0; add.s64 %rd293, %rd207, 116; cvta.to.local.u64 %rd294, %rd293; st.local.f32 [%rd294], %f2797; add.s64 %rd295, %rd207, 120; cvta.to.local.u64 %rd296, %rd295; st.local.f32 [%rd296], %f2799; add.s64 %rd297, %rd207, 124; cvta.to.local.u64 %rd298, %rd297; st.local.f32 [%rd298], %f2798; @%p246 bra $L__BB25_143; @%p166 bra $L__BB25_136; bra.uni $L__BB25_135; $L__BB25_136: ld.global.v4.f32 {%f1826, %f1827, %f1828, %f1829}, [%rd52]; ld.global.v4.f32 {%f1834, %f1835, %f1836, %f1837}, [%rd52+16]; ld.global.v4.f32 {%f1842, %f1843, %f1844, %f1845}, [%rd52+32]; mul.ftz.f32 %f1850, %f2803, %f1827; fma.rn.ftz.f32 %f1851, %f2802, %f1826, %f1850; fma.rn.ftz.f32 %f1852, %f2804, %f1828, %f1851; add.ftz.f32 %f2791, %f1829, %f1852; mul.ftz.f32 %f1853, %f2803, %f1835; fma.rn.ftz.f32 %f1854, %f2802, %f1834, %f1853; fma.rn.ftz.f32 %f1855, %f2804, %f1836, %f1854; add.ftz.f32 %f2792, %f1837, %f1855; mul.ftz.f32 %f1856, %f2803, %f1843; fma.rn.ftz.f32 %f1857, %f2802, %f1842, %f1856; fma.rn.ftz.f32 %f1858, %f2804, %f1844, %f1857; add.ftz.f32 %f2804, %f1845, %f1858; bra.uni $L__BB25_137; $L__BB25_135: add.s64 %rd300, %rd207, 144; cvta.to.local.u64 %rd301, %rd300; ld.local.v4.f32 {%f1793, %f1794, %f1795, %f1796}, [%rd301]; mul.ftz.f32 %f1801, %f2803, %f1794; fma.rn.ftz.f32 %f1802, %f2802, %f1793, %f1801; fma.rn.ftz.f32 %f1803, %f2804, %f1795, %f1802; add.ftz.f32 %f2791, %f1796, %f1803; add.s64 %rd302, %rd207, 160; cvta.to.local.u64 %rd303, %rd302; ld.local.v4.f32 {%f1804, %f1805, %f1806, %f1807}, [%rd303]; mul.ftz.f32 %f1812, %f2803, %f1805; fma.rn.ftz.f32 %f1813, %f2802, %f1804, %f1812; fma.rn.ftz.f32 %f1814, %f2804, %f1806, %f1813; add.ftz.f32 %f2792, %f1807, %f1814; add.s64 %rd304, %rd207, 176; cvta.to.local.u64 %rd305, %rd304; ld.local.v4.f32 {%f1815, %f1816, %f1817, %f1818}, [%rd305]; mul.ftz.f32 %f1823, %f2803, %f1816; fma.rn.ftz.f32 %f1824, %f2802, %f1815, %f1823; fma.rn.ftz.f32 %f1825, %f2804, %f1817, %f1824; add.ftz.f32 %f2804, %f1818, %f1825; $L__BB25_137: @%p166 bra $L__BB25_139; bra.uni $L__BB25_138; $L__BB25_139: ld.global.v4.f32 {%f1893, %f1894, %f1895, %f1896}, [%rd52+48]; ld.global.v4.f32 {%f1900, %f1901, %f1902, %f1903}, [%rd52+64]; ld.global.v4.f32 {%f1907, %f1908, %f1909, %f1910}, [%rd52+80]; mul.ftz.f32 %f1914, %f1900, %f2800; fma.rn.ftz.f32 %f1915, %f1893, %f2801, %f1914; fma.rn.ftz.f32 %f1916, %f1907, %f2788, %f1915; mul.ftz.f32 %f1917, %f1901, %f2800; fma.rn.ftz.f32 %f1918, %f1894, %f2801, %f1917; fma.rn.ftz.f32 %f1919, %f1908, %f2788, %f1918; mul.ftz.f32 %f1920, %f1902, %f2800; fma.rn.ftz.f32 %f1921, %f1895, %f2801, %f1920; fma.rn.ftz.f32 %f1922, %f1909, %f2788, %f1921; mul.ftz.f32 %f1923, %f1919, %f1919; fma.rn.ftz.f32 %f1924, %f1916, %f1916, %f1923; fma.rn.ftz.f32 %f1925, %f1922, %f1922, %f1924; sqrt.approx.ftz.f32 %f1926, %f1925; div.approx.ftz.f32 %f2801, %f1916, %f1926; div.approx.ftz.f32 %f2800, %f1919, %f1926; div.approx.ftz.f32 %f2794, %f1922, %f1926; bra.uni $L__BB25_140; $L__BB25_138: add.s64 %rd307, %rd207, 192; cvta.to.local.u64 %rd308, %rd307; ld.local.v4.f32 {%f1859, %f1860, %f1861, %f1862}, [%rd308]; add.s64 %rd309, %rd207, 208; cvta.to.local.u64 %rd310, %rd309; ld.local.v4.f32 {%f1866, %f1867, %f1868, %f1869}, [%rd310]; add.s64 %rd311, %rd207, 224; cvta.to.local.u64 %rd312, %rd311; ld.local.v4.f32 {%f1873, %f1874, %f1875, %f1876}, [%rd312]; mul.ftz.f32 %f1880, %f2800, %f1866; fma.rn.ftz.f32 %f1881, %f2801, %f1859, %f1880; fma.rn.ftz.f32 %f1882, %f2788, %f1873, %f1881; mul.ftz.f32 %f1883, %f2800, %f1867; fma.rn.ftz.f32 %f1884, %f2801, %f1860, %f1883; fma.rn.ftz.f32 %f1885, %f2788, %f1874, %f1884; mul.ftz.f32 %f1886, %f2800, %f1868; fma.rn.ftz.f32 %f1887, %f2801, %f1861, %f1886; fma.rn.ftz.f32 %f1888, %f2788, %f1875, %f1887; mul.ftz.f32 %f1889, %f1885, %f1885; fma.rn.ftz.f32 %f1890, %f1882, %f1882, %f1889; fma.rn.ftz.f32 %f1891, %f1888, %f1888, %f1890; sqrt.approx.ftz.f32 %f1892, %f1891; div.approx.ftz.f32 %f2801, %f1882, %f1892; div.approx.ftz.f32 %f2800, %f1885, %f1892; div.approx.ftz.f32 %f2794, %f1888, %f1892; $L__BB25_140: or.b64 %rd314, %rd207, 12; cvta.to.local.u64 %rd315, %rd314; st.local.f32 [%rd315], %f2801; add.s64 %rd316, %rd207, 16; cvta.to.local.u64 %rd317, %rd316; st.local.f32 [%rd317], %f2800; add.s64 %rd318, %rd207, 20; cvta.to.local.u64 %rd319, %rd318; st.local.f32 [%rd319], %f2794; @%p166 bra $L__BB25_142; bra.uni $L__BB25_141; $L__BB25_142: ld.global.v4.f32 {%f1954, %f1955, %f1956, %f1957}, [%rd52]; ld.global.v4.f32 {%f1961, %f1962, %f1963, %f1964}, [%rd52+16]; ld.global.v4.f32 {%f1968, %f1969, %f1970, %f1971}, [%rd52+32]; mul.ftz.f32 %f1975, %f1955, %f2799; fma.rn.ftz.f32 %f1976, %f2797, %f1954, %f1975; fma.rn.ftz.f32 %f556, %f1956, %f2798, %f1976; mul.ftz.f32 %f1977, %f1962, %f2799; fma.rn.ftz.f32 %f1978, %f2797, %f1961, %f1977; fma.rn.ftz.f32 %f557, %f1963, %f2798, %f1978; mul.ftz.f32 %f1979, %f1969, %f2799; fma.rn.ftz.f32 %f1980, %f2797, %f1968, %f1979; fma.rn.ftz.f32 %f2798, %f1970, %f2798, %f1980; st.local.f32 [%rd294], %f556; st.local.f32 [%rd296], %f557; st.local.f32 [%rd298], %f2798; mov.f32 %f2797, %f556; mov.f32 %f2799, %f557; mov.f32 %f2802, %f2791; mov.f32 %f2803, %f2792; bra.uni $L__BB25_143; $L__BB25_141: add.s64 %rd321, %rd207, 144; cvta.to.local.u64 %rd322, %rd321; ld.local.v4.f32 {%f1927, %f1928, %f1929, %f1930}, [%rd322]; mul.ftz.f32 %f1934, %f2799, %f1928; fma.rn.ftz.f32 %f1935, %f2797, %f1927, %f1934; fma.rn.ftz.f32 %f553, %f2798, %f1929, %f1935; add.s64 %rd323, %rd207, 160; cvta.to.local.u64 %rd324, %rd323; ld.local.v4.f32 {%f1936, %f1937, %f1938, %f1939}, [%rd324]; mul.ftz.f32 %f1943, %f2799, %f1937; fma.rn.ftz.f32 %f1944, %f2797, %f1936, %f1943; fma.rn.ftz.f32 %f554, %f2798, %f1938, %f1944; add.s64 %rd325, %rd207, 176; cvta.to.local.u64 %rd326, %rd325; ld.local.v4.f32 {%f1945, %f1946, %f1947, %f1948}, [%rd326]; mul.ftz.f32 %f1952, %f2799, %f1946; fma.rn.ftz.f32 %f1953, %f2797, %f1945, %f1952; fma.rn.ftz.f32 %f2798, %f2798, %f1947, %f1953; st.local.f32 [%rd294], %f553; st.local.f32 [%rd296], %f554; st.local.f32 [%rd298], %f2798; mov.f32 %f2797, %f553; mov.f32 %f2799, %f554; mov.f32 %f2802, %f2791; mov.f32 %f2803, %f2792; $L__BB25_143: and.b32 %r447, %r491, 4; setp.eq.s32 %p247, %r447, 0; mul.wide.s32 %rd687, %r538, 16; cvta.to.global.u64 %rd686, %rd110; add.s64 %rd685, %rd686, %rd687; add.s64 %rd684, %rd685, 4; add.s64 %rd683, %rd207, 44; neg.ftz.f32 %f2513, %f6; neg.ftz.f32 %f2512, %f5; add.s64 %rd341, %rd207, 20; cvta.to.local.u64 %rd343, %rd207; st.local.v2.f32 [%rd343], {%f2802, %f2803}; or.b64 %rd344, %rd207, 8; cvta.to.local.u64 %rd345, %rd344; st.local.f32 [%rd345], %f2804; selp.f32 %f2843, %f2801, %f2512, %p247; selp.f32 %f2842, %f2800, %f2513, %p247; selp.b64 %rd346, %rd341, %rd683, %p247; cvta.to.local.u64 %rd347, %rd346; ld.local.f32 %f2841, [%rd347]; add.s64 %rd348, %rd207, 24; cvta.to.local.u64 %rd349, %rd348; st.local.v2.f32 [%rd349], {%f2843, %f2842}; add.s64 %rd350, %rd207, 32; cvta.to.local.u64 %rd351, %rd350; st.local.f32 [%rd351], %f2841; mul.ftz.f32 %f1983, %f2799, %f2841; mul.ftz.f32 %f1984, %f2798, %f2842; mul.ftz.f32 %f1985, %f2798, %f2843; mul.ftz.f32 %f1986, %f2797, %f2841; mul.ftz.f32 %f1987, %f2797, %f2842; mul.ftz.f32 %f1988, %f2799, %f2843; sub.ftz.f32 %f1989, %f1987, %f1988; sub.ftz.f32 %f1990, %f1985, %f1986; sub.ftz.f32 %f1991, %f1983, %f1984; add.s64 %rd352, %rd207, 128; cvta.to.local.u64 %rd353, %rd352; st.local.v2.f32 [%rd353], {%f1991, %f1990}; add.s64 %rd354, %rd207, 136; cvta.to.local.u64 %rd355, %rd354; st.local.f32 [%rd355], %f1989; ld.global.u32 %r375, [%rd684+-4]; add.s64 %rd356, %rd207, 48; cvta.to.local.u64 %rd357, %rd356; st.local.u32 [%rd357], %r375; bra.uni $L__BB25_194; $L__BB25_166: { // callseq 25, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd207; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 25 ld.local.u32 %r549, [%rd215]; bra.uni $L__BB25_180; $L__BB25_146: add.s64 %rd368, %rd207, 84; cvta.to.local.u64 %rd369, %rd368; ld.global.u32 %r201, [%rd52+148]; ld.global.u32 %r202, [%rd52+152]; shl.b32 %r379, %r202, 1; cvt.rn.f32.s32 %f1992, %r379; ld.local.f32 %f1993, [%rd369]; mul.ftz.f32 %f573, %f1993, %f1992; cvt.rzi.ftz.s32.f32 %r380, %f573; add.s32 %r381, %r379, -1; min.s32 %r203, %r380, %r381; ld.global.u32 %r545, [%rd52+164]; mul.wide.u32 %rd370, %r545, 16; add.s64 %rd371, %rd20, %rd370; ld.global.u16 %rs24, [%rd371+12]; ld.global.u32 %r544, [%rd371+8]; ld.global.u64 %rd701, [%rd371]; setp.eq.s64 %p196, %rd701, 11; @%p196 bra $L__BB25_151; $L__BB25_147: setp.eq.s64 %p197, %rd701, 0; @%p197 bra $L__BB25_149; add.s32 %r545, %r545, 2; bra.uni $L__BB25_150; $L__BB25_149: setp.eq.s16 %p198, %rs24, 0; mov.u32 %r548, -1; mov.u32 %r545, %r544; @%p198 bra $L__BB25_152; $L__BB25_150: mul.wide.u32 %rd372, %r545, 16; add.s64 %rd373, %rd20, %rd372; ld.global.u16 %rs24, [%rd373+12]; ld.global.u32 %r544, [%rd373+8]; ld.global.u64 %rd701, [%rd373]; setp.ne.s64 %p199, %rd701, 11; @%p199 bra $L__BB25_147; $L__BB25_151: setp.eq.s16 %p200, %rs24, 0; selp.b32 %r548, -1, %r544, %p200; $L__BB25_152: setp.eq.s32 %p201, %r203, %r202; @%p201 bra $L__BB25_154; bra.uni $L__BB25_153; $L__BB25_154: shl.b64 %rd375, %rd60, 4; add.s64 %rd702, %rd22, %rd375; bra.uni $L__BB25_155; $L__BB25_168: sub.ftz.f32 %f2098, %f614, %f615; sub.ftz.f32 %f2099, %f616, %f617; mul.ftz.f32 %f2100, %f2099, %f2099; fma.rn.ftz.f32 %f2101, %f2098, %f2098, %f2100; sub.ftz.f32 %f2102, %f618, %f619; fma.rn.ftz.f32 %f2103, %f2102, %f2102, %f2101; sqrt.approx.ftz.f32 %f2104, %f2103; div.approx.ftz.f32 %f2813, %f2102, %f2104; div.approx.ftz.f32 %f2812, %f2099, %f2104; div.approx.ftz.f32 %f2811, %f2098, %f2104; $L__BB25_170: shl.b64 %rd415, %rd70, 2; add.s64 %rd416, %rd24, %rd415; ld.global.u32 %r216, [%rd416]; add.s64 %rd418, %rd207, 48; cvta.to.local.u64 %rd419, %rd418; st.local.u32 [%rd419], %r216; fma.rn.ftz.f32 %f2112, %f2628, %f611, %f605; fma.rn.ftz.f32 %f2113, %f2628, %f612, %f607; fma.rn.ftz.f32 %f2114, %f2628, %f613, %f609; fma.rn.ftz.f32 %f2826, %f2627, %f606, %f2112; fma.rn.ftz.f32 %f2827, %f2627, %f608, %f2113; fma.rn.ftz.f32 %f2828, %f2627, %f610, %f2114; and.b32 %r397, %r549, 4; setp.ne.s32 %p209, %r397, 0; @%p209 bra $L__BB25_175; @%p166 bra $L__BB25_173; bra.uni $L__BB25_172; $L__BB25_173: ld.global.v4.f32 {%f2814, %f2815, %f2816, %f2817}, [%rd52]; ld.global.v4.f32 {%f2818, %f2819, %f2820, %f2821}, [%rd52+16]; ld.global.v4.f32 {%f2822, %f2823, %f2824, %f2825}, [%rd52+32]; bra.uni $L__BB25_174; $L__BB25_153: setp.gt.s32 %p202, %r203, %r202; selp.b32 %r383, -1, 0, %p202; add.s32 %r384, %r203, %r383; mad.lo.s32 %r385, %r384, %r201, %r538; add.s32 %r386, %r385, %r548; mul.wide.s32 %rd374, %r386, 16; add.s64 %rd702, %rd25, %rd374; $L__BB25_155: ld.global.v4.f32 {%f1994, %f1995, %f1996, %f1997}, [%rd702]; add.s32 %r214, %r203, 1; setp.eq.s32 %p203, %r214, %r202; @%p203 bra $L__BB25_157; bra.uni $L__BB25_156; $L__BB25_157: shl.b64 %rd377, %rd60, 4; add.s64 %rd703, %rd22, %rd377; bra.uni $L__BB25_158; $L__BB25_156: setp.ge.s32 %p204, %r203, %r202; selp.b32 %r387, -1, 0, %p204; add.s32 %r388, %r214, %r387; mad.lo.s32 %r389, %r388, %r201, %r538; add.s32 %r390, %r389, %r548; mul.wide.s32 %rd376, %r390, 16; add.s64 %rd703, %rd25, %rd376; $L__BB25_158: cvt.rn.f32.s32 %f1998, %r203; sub.ftz.f32 %f1999, %f573, %f1998; ld.global.v4.f32 {%f2000, %f2001, %f2002, %f2003}, [%rd703]; mov.f32 %f2008, 0f3F800000; sub.ftz.f32 %f2009, %f2008, %f1999; mul.ftz.f32 %f2010, %f1999, %f2000; mul.ftz.f32 %f2011, %f1999, %f2001; mul.ftz.f32 %f2012, %f1999, %f2002; fma.rn.ftz.f32 %f2810, %f2009, %f1996, %f2012; fma.rn.ftz.f32 %f2809, %f2009, %f1995, %f2011; fma.rn.ftz.f32 %f2808, %f2009, %f1994, %f2010; $L__BB25_160: and.b32 %r391, %r549, 4; setp.ne.s32 %p205, %r391, 0; @%p205 bra $L__BB25_164; @%p166 bra $L__BB25_163; bra.uni $L__BB25_162; $L__BB25_163: ld.global.v4.f32 {%f2051, %f2052, %f2053, %f2054}, [%rd52]; ld.global.v4.f32 {%f2059, %f2060, %f2061, %f2062}, [%rd52+16]; ld.global.v4.f32 {%f2067, %f2068, %f2069, %f2070}, [%rd52+32]; mul.ftz.f32 %f2075, %f2809, %f2052; fma.rn.ftz.f32 %f2076, %f2808, %f2051, %f2075; fma.rn.ftz.f32 %f2077, %f2810, %f2053, %f2076; add.ftz.f32 %f596, %f2054, %f2077; mul.ftz.f32 %f2078, %f2809, %f2060; fma.rn.ftz.f32 %f2079, %f2808, %f2059, %f2078; fma.rn.ftz.f32 %f2080, %f2810, %f2061, %f2079; add.ftz.f32 %f597, %f2062, %f2080; mul.ftz.f32 %f2081, %f2809, %f2068; fma.rn.ftz.f32 %f2082, %f2808, %f2067, %f2081; fma.rn.ftz.f32 %f2083, %f2810, %f2069, %f2082; add.ftz.f32 %f2810, %f2070, %f2083; mov.f32 %f2808, %f596; mov.f32 %f2809, %f597; bra.uni $L__BB25_164; $L__BB25_162: add.s64 %rd381, %rd207, 144; cvta.to.local.u64 %rd382, %rd381; ld.local.v4.f32 {%f2018, %f2019, %f2020, %f2021}, [%rd382]; mul.ftz.f32 %f2026, %f2809, %f2019; fma.rn.ftz.f32 %f2027, %f2808, %f2018, %f2026; fma.rn.ftz.f32 %f2028, %f2810, %f2020, %f2027; add.ftz.f32 %f593, %f2021, %f2028; add.s64 %rd383, %rd207, 160; cvta.to.local.u64 %rd384, %rd383; ld.local.v4.f32 {%f2029, %f2030, %f2031, %f2032}, [%rd384]; mul.ftz.f32 %f2037, %f2809, %f2030; fma.rn.ftz.f32 %f2038, %f2808, %f2029, %f2037; fma.rn.ftz.f32 %f2039, %f2810, %f2031, %f2038; add.ftz.f32 %f594, %f2032, %f2039; add.s64 %rd385, %rd207, 176; cvta.to.local.u64 %rd386, %rd385; ld.local.v4.f32 {%f2040, %f2041, %f2042, %f2043}, [%rd386]; mul.ftz.f32 %f2048, %f2809, %f2041; fma.rn.ftz.f32 %f2049, %f2808, %f2040, %f2048; fma.rn.ftz.f32 %f2050, %f2810, %f2042, %f2049; add.ftz.f32 %f2810, %f2043, %f2050; mov.f32 %f2808, %f593; mov.f32 %f2809, %f594; $L__BB25_164: sub.ftz.f32 %f2084, %f570, %f2808; sub.ftz.f32 %f2085, %f571, %f2809; mul.ftz.f32 %f2086, %f2085, %f2085; fma.rn.ftz.f32 %f2087, %f2084, %f2084, %f2086; sub.ftz.f32 %f2088, %f572, %f2810; fma.rn.ftz.f32 %f2089, %f2088, %f2088, %f2087; sqrt.approx.ftz.f32 %f2090, %f2089; div.approx.ftz.f32 %f2843, %f2084, %f2090; div.approx.ftz.f32 %f2842, %f2085, %f2090; div.approx.ftz.f32 %f2841, %f2088, %f2090; add.s64 %rd388, %rd207, 32; cvta.to.local.u64 %rd389, %rd388; st.local.f32 [%rd389], %f2841; or.b64 %rd390, %rd207, 12; cvta.to.local.u64 %rd391, %rd390; st.local.f32 [%rd391], %f2843; add.s64 %rd392, %rd207, 16; cvta.to.local.u64 %rd393, %rd392; st.local.v4.f32 [%rd393], {%f2842, %f2841, %f2843, %f2842}; add.s64 %rd394, %rd207, 116; cvta.to.local.u64 %rd395, %rd394; mov.u32 %r392, 0; st.local.u32 [%rd395], %r392; add.s64 %rd396, %rd207, 120; cvta.to.local.u64 %rd397, %rd396; st.local.u32 [%rd397], %r392; add.s64 %rd398, %rd207, 124; cvta.to.local.u64 %rd399, %rd398; st.local.u32 [%rd399], %r392; add.s64 %rd400, %rd207, 128; cvta.to.local.u64 %rd401, %rd400; mov.f32 %f2091, 0f00000000; st.local.v2.f32 [%rd401], {%f2091, %f2091}; add.s64 %rd402, %rd207, 136; cvta.to.local.u64 %rd403, %rd402; st.local.u32 [%rd403], %r392; bra.uni $L__BB25_194; $L__BB25_172: add.s64 %rd421, %rd207, 144; cvta.to.local.u64 %rd422, %rd421; ld.local.v4.f32 {%f2814, %f2815, %f2816, %f2817}, [%rd422]; add.s64 %rd423, %rd207, 160; cvta.to.local.u64 %rd424, %rd423; ld.local.v4.f32 {%f2818, %f2819, %f2820, %f2821}, [%rd424]; add.s64 %rd425, %rd207, 176; cvta.to.local.u64 %rd426, %rd425; ld.local.v4.f32 {%f2822, %f2823, %f2824, %f2825}, [%rd426]; $L__BB25_174: mul.ftz.f32 %f2139, %f2827, %f2815; fma.rn.ftz.f32 %f2140, %f2826, %f2814, %f2139; fma.rn.ftz.f32 %f2141, %f2828, %f2816, %f2140; mul.ftz.f32 %f2142, %f2827, %f2819; fma.rn.ftz.f32 %f2143, %f2826, %f2818, %f2142; fma.rn.ftz.f32 %f2144, %f2828, %f2820, %f2143; mul.ftz.f32 %f2145, %f2827, %f2823; fma.rn.ftz.f32 %f2146, %f2826, %f2822, %f2145; fma.rn.ftz.f32 %f2147, %f2828, %f2824, %f2146; add.ftz.f32 %f2826, %f2817, %f2141; add.ftz.f32 %f2827, %f2821, %f2144; add.ftz.f32 %f2828, %f2825, %f2147; $L__BB25_175: add.s64 %rd428, %rd207, 32; cvta.to.local.u64 %rd429, %rd428; st.local.f32 [%rd429], %f2813; cvta.to.local.u64 %rd430, %rd207; st.local.v4.f32 [%rd430], {%f2826, %f2827, %f2828, %f2811}; add.s64 %rd431, %rd207, 16; cvta.to.local.u64 %rd432, %rd431; st.local.v4.f32 [%rd432], {%f2812, %f2813, %f2811, %f2812}; setp.gt.s32 %p211, %r216, -1; @%p211 bra $L__BB25_179; ld.global.v4.u32 {%r398, %r399, %r400, %r401}, [%rd71+-12]; mul.wide.u32 %rd434, %r398, 12; add.s64 %rd435, %rd433, %rd434; mul.wide.u32 %rd436, %r399, 12; add.s64 %rd437, %rd433, %rd436; mul.wide.u32 %rd438, %r400, 12; add.s64 %rd439, %rd433, %rd438; mov.f32 %f2148, 0f3F800000; sub.ftz.f32 %f2149, %f2148, %f2628; sub.ftz.f32 %f2150, %f2149, %f2627; ld.global.f32 %f2151, [%rd435]; ld.global.f32 %f2152, [%rd435+4]; ld.global.f32 %f2153, [%rd435+8]; ld.global.f32 %f2154, [%rd437]; mul.ftz.f32 %f2155, %f2628, %f2154; ld.global.f32 %f2156, [%rd437+4]; mul.ftz.f32 %f2157, %f2628, %f2156; ld.global.f32 %f2158, [%rd437+8]; mul.ftz.f32 %f2159, %f2628, %f2158; fma.rn.ftz.f32 %f2160, %f2150, %f2151, %f2155; fma.rn.ftz.f32 %f2161, %f2150, %f2152, %f2157; fma.rn.ftz.f32 %f2162, %f2150, %f2153, %f2159; ld.global.f32 %f2163, [%rd439]; ld.global.f32 %f2164, [%rd439+4]; ld.global.f32 %f2165, [%rd439+8]; fma.rn.ftz.f32 %f2829, %f2627, %f2163, %f2160; fma.rn.ftz.f32 %f2830, %f2627, %f2164, %f2161; fma.rn.ftz.f32 %f2831, %f2627, %f2165, %f2162; mul.ftz.f32 %f2166, %f2830, %f2830; fma.rn.ftz.f32 %f2167, %f2829, %f2829, %f2166; fma.rn.ftz.f32 %f2168, %f2831, %f2831, %f2167; sqrt.approx.ftz.f32 %f680, %f2168; setp.neu.ftz.f32 %p212, %f680, 0f00000000; @%p212 bra $L__BB25_177; bra.uni $L__BB25_178; $L__BB25_177: rcp.approx.ftz.f32 %f2169, %f680; mul.ftz.f32 %f2831, %f2831, %f2169; mul.ftz.f32 %f2830, %f2830, %f2169; mul.ftz.f32 %f2829, %f2829, %f2169; $L__BB25_178: setp.eq.ftz.f32 %p213, %f2829, 0f00000000; setp.eq.ftz.f32 %p214, %f2830, 0f00000000; and.pred %p215, %p213, %p214; setp.eq.ftz.f32 %p216, %f2831, 0f00000000; and.pred %p217, %p216, %p215; selp.f32 %f2170, %f2813, %f2831, %p217; selp.f32 %f2171, %f2812, %f2830, %p217; selp.f32 %f2172, %f2811, %f2829, %p217; or.b64 %rd441, %rd207, 12; cvta.to.local.u64 %rd442, %rd441; st.local.f32 [%rd442], %f2172; st.local.f32 [%rd432], %f2171; add.s64 %rd445, %rd207, 20; cvta.to.local.u64 %rd446, %rd445; st.local.f32 [%rd446], %f2170; $L__BB25_179: add.s64 %rd448, %rd207, 116; cvta.to.local.u64 %rd449, %rd448; st.local.f32 [%rd449], %f611; add.s64 %rd450, %rd207, 120; cvta.to.local.u64 %rd451, %rd450; st.local.f32 [%rd451], %f612; add.s64 %rd452, %rd207, 124; cvta.to.local.u64 %rd453, %rd452; st.local.f32 [%rd453], %f613; add.s64 %rd454, %rd207, 128; cvta.to.local.u64 %rd455, %rd454; st.local.v2.f32 [%rd455], {%f606, %f608}; add.s64 %rd456, %rd207, 136; cvta.to.local.u64 %rd457, %rd456; st.local.f32 [%rd457], %f610; $L__BB25_180: and.b32 %r405, %r549, 4; setp.eq.s32 %p218, %r405, 0; @%p218 bra $L__BB25_182; bra.uni $L__BB25_181; $L__BB25_182: and.b32 %r218, %r549, 2; setp.eq.s32 %p219, %r218, 0; @%p219 bra $L__BB25_184; bra.uni $L__BB25_183; $L__BB25_184: add.s64 %rd475, %rd207, 16; add.s64 %rd476, %rd207, 76; cvta.to.local.u64 %rd477, %rd476; ld.local.u32 %r406, [%rd477]; mul.wide.s32 %rd478, %r406, 208; add.s64 %rd479, %rd16, %rd478; ld.global.v4.f32 {%f2214, %f2215, %f2216, %f2217}, [%rd479+48]; ld.global.v4.f32 {%f2221, %f2222, %f2223, %f2224}, [%rd479+64]; ld.global.v4.f32 {%f2228, %f2229, %f2230, %f2231}, [%rd479+80]; or.b64 %rd480, %rd207, 12; cvta.to.local.u64 %rd481, %rd480; cvta.to.local.u64 %rd482, %rd475; ld.local.v2.f32 {%f2235, %f2236}, [%rd482]; ld.local.f32 %f2239, [%rd481]; mul.ftz.f32 %f2240, %f2221, %f2235; fma.rn.ftz.f32 %f2241, %f2214, %f2239, %f2240; fma.rn.ftz.f32 %f2242, %f2228, %f2236, %f2241; mul.ftz.f32 %f2243, %f2222, %f2235; fma.rn.ftz.f32 %f2244, %f2215, %f2239, %f2243; fma.rn.ftz.f32 %f2245, %f2229, %f2236, %f2244; mul.ftz.f32 %f2246, %f2223, %f2235; fma.rn.ftz.f32 %f2247, %f2216, %f2239, %f2246; fma.rn.ftz.f32 %f2248, %f2230, %f2236, %f2247; mul.ftz.f32 %f2249, %f2245, %f2245; fma.rn.ftz.f32 %f2250, %f2242, %f2242, %f2249; fma.rn.ftz.f32 %f2251, %f2248, %f2248, %f2250; sqrt.approx.ftz.f32 %f2252, %f2251; div.approx.ftz.f32 %f2834, %f2242, %f2252; div.approx.ftz.f32 %f2833, %f2245, %f2252; div.approx.ftz.f32 %f2832, %f2248, %f2252; bra.uni $L__BB25_185; $L__BB25_181: add.s64 %rd459, %rd207, 24; cvta.to.local.u64 %rd460, %rd459; ld.local.v2.f32 {%f2843, %f2842}, [%rd460]; add.s64 %rd461, %rd207, 32; cvta.to.local.u64 %rd462, %rd461; ld.local.f32 %f2841, [%rd462]; bra.uni $L__BB25_194; $L__BB25_183: add.s64 %rd464, %rd207, 16; or.b64 %rd465, %rd207, 12; cvta.to.local.u64 %rd466, %rd465; cvta.to.local.u64 %rd467, %rd464; ld.local.v2.f32 {%f2175, %f2176}, [%rd467]; add.s64 %rd468, %rd207, 192; cvta.to.local.u64 %rd469, %rd468; ld.local.v4.f32 {%f2179, %f2180, %f2181, %f2182}, [%rd469]; add.s64 %rd470, %rd207, 208; cvta.to.local.u64 %rd471, %rd470; ld.local.v4.f32 {%f2186, %f2187, %f2188, %f2189}, [%rd471]; add.s64 %rd472, %rd207, 224; cvta.to.local.u64 %rd473, %rd472; ld.local.v4.f32 {%f2193, %f2194, %f2195, %f2196}, [%rd473]; ld.local.f32 %f2200, [%rd466]; mul.ftz.f32 %f2201, %f2175, %f2186; fma.rn.ftz.f32 %f2202, %f2200, %f2179, %f2201; fma.rn.ftz.f32 %f2203, %f2176, %f2193, %f2202; mul.ftz.f32 %f2204, %f2175, %f2187; fma.rn.ftz.f32 %f2205, %f2200, %f2180, %f2204; fma.rn.ftz.f32 %f2206, %f2176, %f2194, %f2205; mul.ftz.f32 %f2207, %f2175, %f2188; fma.rn.ftz.f32 %f2208, %f2200, %f2181, %f2207; fma.rn.ftz.f32 %f2209, %f2176, %f2195, %f2208; mul.ftz.f32 %f2210, %f2206, %f2206; fma.rn.ftz.f32 %f2211, %f2203, %f2203, %f2210; fma.rn.ftz.f32 %f2212, %f2209, %f2209, %f2211; sqrt.approx.ftz.f32 %f2213, %f2212; div.approx.ftz.f32 %f2834, %f2203, %f2213; div.approx.ftz.f32 %f2833, %f2206, %f2213; div.approx.ftz.f32 %f2832, %f2209, %f2213; $L__BB25_185: or.b64 %rd484, %rd207, 12; cvta.to.local.u64 %rd485, %rd484; st.local.f32 [%rd485], %f2834; add.s64 %rd486, %rd207, 16; cvta.to.local.u64 %rd487, %rd486; st.local.f32 [%rd487], %f2833; add.s64 %rd488, %rd207, 20; cvta.to.local.u64 %rd489, %rd488; st.local.f32 [%rd489], %f2832; @%p219 bra $L__BB25_187; bra.uni $L__BB25_186; $L__BB25_187: add.s64 %rd502, %rd207, 24; add.s64 %rd503, %rd207, 76; cvta.to.local.u64 %rd504, %rd503; ld.local.u32 %r407, [%rd504]; mul.wide.s32 %rd505, %r407, 208; add.s64 %rd506, %rd16, %rd505; ld.global.v4.f32 {%f2292, %f2293, %f2294, %f2295}, [%rd506+48]; ld.global.v4.f32 {%f2299, %f2300, %f2301, %f2302}, [%rd506+64]; ld.global.v4.f32 {%f2306, %f2307, %f2308, %f2309}, [%rd506+80]; cvta.to.local.u64 %rd507, %rd502; ld.local.v2.f32 {%f2313, %f2314}, [%rd507]; add.s64 %rd508, %rd207, 32; cvta.to.local.u64 %rd509, %rd508; mul.ftz.f32 %f2317, %f2299, %f2314; fma.rn.ftz.f32 %f2318, %f2292, %f2313, %f2317; ld.local.f32 %f2319, [%rd509]; fma.rn.ftz.f32 %f2320, %f2306, %f2319, %f2318; mul.ftz.f32 %f2321, %f2300, %f2314; fma.rn.ftz.f32 %f2322, %f2293, %f2313, %f2321; fma.rn.ftz.f32 %f2323, %f2307, %f2319, %f2322; mul.ftz.f32 %f2324, %f2301, %f2314; fma.rn.ftz.f32 %f2325, %f2294, %f2313, %f2324; fma.rn.ftz.f32 %f2326, %f2308, %f2319, %f2325; mul.ftz.f32 %f2327, %f2323, %f2323; fma.rn.ftz.f32 %f2328, %f2320, %f2320, %f2327; fma.rn.ftz.f32 %f2329, %f2326, %f2326, %f2328; sqrt.approx.ftz.f32 %f2330, %f2329; div.approx.ftz.f32 %f2843, %f2320, %f2330; div.approx.ftz.f32 %f2842, %f2323, %f2330; div.approx.ftz.f32 %f2841, %f2326, %f2330; bra.uni $L__BB25_188; $L__BB25_186: add.s64 %rd491, %rd207, 24; cvta.to.local.u64 %rd492, %rd491; ld.local.v2.f32 {%f2253, %f2254}, [%rd492]; add.s64 %rd493, %rd207, 32; cvta.to.local.u64 %rd494, %rd493; add.s64 %rd495, %rd207, 192; cvta.to.local.u64 %rd496, %rd495; ld.local.v4.f32 {%f2257, %f2258, %f2259, %f2260}, [%rd496]; add.s64 %rd497, %rd207, 208; cvta.to.local.u64 %rd498, %rd497; ld.local.v4.f32 {%f2264, %f2265, %f2266, %f2267}, [%rd498]; add.s64 %rd499, %rd207, 224; cvta.to.local.u64 %rd500, %rd499; ld.local.v4.f32 {%f2271, %f2272, %f2273, %f2274}, [%rd500]; mul.ftz.f32 %f2278, %f2254, %f2264; fma.rn.ftz.f32 %f2279, %f2253, %f2257, %f2278; ld.local.f32 %f2280, [%rd494]; fma.rn.ftz.f32 %f2281, %f2280, %f2271, %f2279; mul.ftz.f32 %f2282, %f2254, %f2265; fma.rn.ftz.f32 %f2283, %f2253, %f2258, %f2282; fma.rn.ftz.f32 %f2284, %f2280, %f2272, %f2283; mul.ftz.f32 %f2285, %f2254, %f2266; fma.rn.ftz.f32 %f2286, %f2253, %f2259, %f2285; fma.rn.ftz.f32 %f2287, %f2280, %f2273, %f2286; mul.ftz.f32 %f2288, %f2284, %f2284; fma.rn.ftz.f32 %f2289, %f2281, %f2281, %f2288; fma.rn.ftz.f32 %f2290, %f2287, %f2287, %f2289; sqrt.approx.ftz.f32 %f2291, %f2290; div.approx.ftz.f32 %f2843, %f2281, %f2291; div.approx.ftz.f32 %f2842, %f2284, %f2291; div.approx.ftz.f32 %f2841, %f2287, %f2291; $L__BB25_188: add.s64 %rd511, %rd207, 24; cvta.to.local.u64 %rd512, %rd511; st.local.v2.f32 [%rd512], {%f2843, %f2842}; add.s64 %rd513, %rd207, 32; cvta.to.local.u64 %rd514, %rd513; st.local.f32 [%rd514], %f2841; @%p219 bra $L__BB25_190; bra.uni $L__BB25_189; $L__BB25_190: add.s64 %rd527, %rd207, 120; add.s64 %rd528, %rd207, 76; cvta.to.local.u64 %rd529, %rd528; ld.local.u32 %r408, [%rd529]; mul.wide.s32 %rd530, %r408, 208; add.s64 %rd531, %rd16, %rd530; ld.global.v4.f32 {%f2363, %f2364, %f2365, %f2366}, [%rd531]; ld.global.v4.f32 {%f2370, %f2371, %f2372, %f2373}, [%rd531+16]; ld.global.v4.f32 {%f2377, %f2378, %f2379, %f2380}, [%rd531+32]; add.s64 %rd532, %rd207, 116; cvta.to.local.u64 %rd533, %rd532; cvta.to.local.u64 %rd534, %rd527; ld.local.v2.f32 {%f2384, %f2385}, [%rd534]; ld.local.f32 %f2388, [%rd533]; mul.ftz.f32 %f2389, %f2364, %f2384; fma.rn.ftz.f32 %f2390, %f2388, %f2363, %f2389; fma.rn.ftz.f32 %f2840, %f2365, %f2385, %f2390; mul.ftz.f32 %f2391, %f2371, %f2384; fma.rn.ftz.f32 %f2392, %f2388, %f2370, %f2391; fma.rn.ftz.f32 %f2839, %f2372, %f2385, %f2392; mul.ftz.f32 %f2393, %f2378, %f2384; fma.rn.ftz.f32 %f2394, %f2388, %f2377, %f2393; fma.rn.ftz.f32 %f2838, %f2379, %f2385, %f2394; bra.uni $L__BB25_191; $L__BB25_189: add.s64 %rd516, %rd207, 120; add.s64 %rd517, %rd207, 116; cvta.to.local.u64 %rd518, %rd517; cvta.to.local.u64 %rd519, %rd516; ld.local.v2.f32 {%f2331, %f2332}, [%rd519]; add.s64 %rd520, %rd207, 144; cvta.to.local.u64 %rd521, %rd520; ld.local.v4.f32 {%f2335, %f2336, %f2337, %f2338}, [%rd521]; ld.local.f32 %f2342, [%rd518]; mul.ftz.f32 %f2343, %f2331, %f2336; fma.rn.ftz.f32 %f2344, %f2342, %f2335, %f2343; fma.rn.ftz.f32 %f2840, %f2332, %f2337, %f2344; add.s64 %rd522, %rd207, 160; cvta.to.local.u64 %rd523, %rd522; ld.local.v4.f32 {%f2345, %f2346, %f2347, %f2348}, [%rd523]; mul.ftz.f32 %f2352, %f2331, %f2346; fma.rn.ftz.f32 %f2353, %f2342, %f2345, %f2352; fma.rn.ftz.f32 %f2839, %f2332, %f2347, %f2353; add.s64 %rd524, %rd207, 176; cvta.to.local.u64 %rd525, %rd524; ld.local.v4.f32 {%f2354, %f2355, %f2356, %f2357}, [%rd525]; mul.ftz.f32 %f2361, %f2331, %f2355; fma.rn.ftz.f32 %f2362, %f2342, %f2354, %f2361; fma.rn.ftz.f32 %f2838, %f2332, %f2356, %f2362; $L__BB25_191: add.s64 %rd536, %rd207, 116; cvta.to.local.u64 %rd537, %rd536; st.local.f32 [%rd537], %f2840; add.s64 %rd538, %rd207, 120; cvta.to.local.u64 %rd539, %rd538; st.local.f32 [%rd539], %f2839; add.s64 %rd540, %rd207, 124; cvta.to.local.u64 %rd541, %rd540; st.local.f32 [%rd541], %f2838; @%p219 bra $L__BB25_193; bra.uni $L__BB25_192; $L__BB25_193: add.s64 %rd554, %rd207, 128; add.s64 %rd555, %rd207, 76; cvta.to.local.u64 %rd556, %rd555; ld.local.u32 %r409, [%rd556]; mul.wide.s32 %rd557, %r409, 208; add.s64 %rd558, %rd16, %rd557; ld.global.v4.f32 {%f2432, %f2433, %f2434, %f2435}, [%rd558]; ld.global.v4.f32 {%f2439, %f2440, %f2441, %f2442}, [%rd558+16]; ld.global.v4.f32 {%f2446, %f2447, %f2448, %f2449}, [%rd558+32]; cvta.to.local.u64 %rd559, %rd554; ld.local.v4.f32 {%f2453, %f2454, %f2455, %f2456}, [%rd559]; mul.ftz.f32 %f2460, %f2433, %f2454; fma.rn.ftz.f32 %f2461, %f2453, %f2432, %f2460; mul.ftz.f32 %f2462, %f2440, %f2454; fma.rn.ftz.f32 %f2463, %f2453, %f2439, %f2462; mul.ftz.f32 %f2464, %f2447, %f2454; fma.rn.ftz.f32 %f2465, %f2453, %f2446, %f2464; fma.rn.ftz.f32 %f2466, %f2448, %f2455, %f2465; fma.rn.ftz.f32 %f2467, %f2441, %f2455, %f2463; fma.rn.ftz.f32 %f2468, %f2434, %f2455, %f2461; st.local.v2.f32 [%rd559], {%f2468, %f2467}; add.s64 %rd560, %rd207, 136; cvta.to.local.u64 %rd561, %rd560; st.local.f32 [%rd561], %f2466; bra.uni $L__BB25_194; $L__BB25_192: add.s64 %rd543, %rd207, 128; cvta.to.local.u64 %rd544, %rd543; ld.local.v4.f32 {%f2395, %f2396, %f2397, %f2398}, [%rd544]; add.s64 %rd545, %rd207, 144; cvta.to.local.u64 %rd546, %rd545; ld.local.v4.f32 {%f2402, %f2403, %f2404, %f2405}, [%rd546]; mul.ftz.f32 %f2409, %f2396, %f2403; fma.rn.ftz.f32 %f2410, %f2395, %f2402, %f2409; add.s64 %rd547, %rd207, 160; cvta.to.local.u64 %rd548, %rd547; ld.local.v4.f32 {%f2411, %f2412, %f2413, %f2414}, [%rd548]; mul.ftz.f32 %f2418, %f2396, %f2412; fma.rn.ftz.f32 %f2419, %f2395, %f2411, %f2418; add.s64 %rd549, %rd207, 176; cvta.to.local.u64 %rd550, %rd549; ld.local.v4.f32 {%f2420, %f2421, %f2422, %f2423}, [%rd550]; mul.ftz.f32 %f2427, %f2396, %f2421; fma.rn.ftz.f32 %f2428, %f2395, %f2420, %f2427; fma.rn.ftz.f32 %f2429, %f2397, %f2422, %f2428; fma.rn.ftz.f32 %f2430, %f2397, %f2413, %f2419; fma.rn.ftz.f32 %f2431, %f2397, %f2404, %f2410; st.local.v2.f32 [%rd544], {%f2431, %f2430}; add.s64 %rd551, %rd207, 136; cvta.to.local.u64 %rd552, %rd551; st.local.f32 [%rd552], %f2429; $L__BB25_194: add.s64 %rd693, %rd207, 40; cvta.to.local.u64 %rd692, %rd693; add.s64 %rd691, %rd207, 36; cvta.to.local.u64 %rd690, %rd691; add.s64 %rd689, %rd207, 52; cvta.to.local.u64 %rd688, %rd689; add.s64 %rd563, %rd207, 48; cvta.to.local.u64 %rd564, %rd563; ld.local.u32 %r219, [%rd564]; shl.b32 %r410, %r219, 5; cvt.u64.u32 %rd565, %r410; and.b64 %rd566, %rd565, 134217696; add.s64 %rd567, %rd21, %rd566; ld.global.u32 %r550, [%rd567+16]; st.local.u32 [%rd688], %r550; ld.local.f32 %f2469, [%rd690]; ld.local.v2.f32 {%f2470, %f2471}, [%rd692]; mul.ftz.f32 %f2474, %f2842, %f2470; fma.rn.ftz.f32 %f2475, %f2843, %f2469, %f2474; fma.rn.ftz.f32 %f2476, %f2841, %f2471, %f2475; setp.geu.ftz.f32 %p223, %f2476, 0f00000000; @%p223 bra $L__BB25_196; add.s64 %rd696, %rd207, 52; cvta.to.local.u64 %rd695, %rd696; or.b32 %r550, %r550, 1; st.local.u32 [%rd695], %r550; add.s64 %rd577, %rd207, 32; cvta.to.local.u64 %rd578, %rd577; neg.ftz.f32 %f2477, %f2841; st.local.f32 [%rd578], %f2477; or.b64 %rd579, %rd207, 12; cvta.to.local.u64 %rd580, %rd579; ld.local.f32 %f2478, [%rd580]; neg.ftz.f32 %f2479, %f2478; add.s64 %rd581, %rd207, 16; cvta.to.local.u64 %rd582, %rd581; ld.local.v2.f32 {%f2480, %f2481}, [%rd582]; st.local.f32 [%rd580], %f2479; neg.ftz.f32 %f2484, %f2842; neg.ftz.f32 %f2485, %f2843; neg.ftz.f32 %f2486, %f2481; neg.ftz.f32 %f2487, %f2480; st.local.v4.f32 [%rd582], {%f2487, %f2486, %f2485, %f2484}; add.s64 %rd583, %rd207, 116; cvta.to.local.u64 %rd584, %rd583; ld.local.f32 %f2488, [%rd584]; neg.ftz.f32 %f2489, %f2488; add.s64 %rd585, %rd207, 120; cvta.to.local.u64 %rd586, %rd585; ld.local.v2.f32 {%f2490, %f2491}, [%rd586]; neg.ftz.f32 %f2494, %f2490; neg.ftz.f32 %f2495, %f2491; st.local.f32 [%rd584], %f2489; st.local.f32 [%rd586], %f2494; add.s64 %rd587, %rd207, 124; cvta.to.local.u64 %rd588, %rd587; st.local.f32 [%rd588], %f2495; add.s64 %rd589, %rd207, 128; cvta.to.local.u64 %rd590, %rd589; ld.local.v4.f32 {%f2496, %f2497, %f2498, %f2499}, [%rd590]; neg.ftz.f32 %f2503, %f2498; neg.ftz.f32 %f2504, %f2497; neg.ftz.f32 %f2505, %f2496; st.local.v2.f32 [%rd590], {%f2505, %f2504}; add.s64 %rd591, %rd207, 136; cvta.to.local.u64 %rd592, %rd591; st.local.f32 [%rd592], %f2503; $L__BB25_196: add.s64 %rd594, %rd207, 92; cvta.to.local.u64 %rd595, %rd594; mov.u32 %r411, 2143289344; st.local.u32 [%rd595], %r411; add.s64 %rd596, %rd207, 100; cvta.to.local.u64 %rd597, %rd596; st.local.u32 [%rd597], %r411; add.s64 %rd598, %rd207, 104; cvta.to.local.u64 %rd599, %rd598; mov.f32 %f2506, 0f7FC00000; st.local.v2.f32 [%rd599], {%f2506, %f2506}; add.s64 %rd600, %rd207, 112; cvta.to.local.u64 %rd601, %rd600; st.local.u32 [%rd601], %r411; and.b32 %r412, %r550, 262144; setp.eq.s32 %p224, %r412, 0; @%p224 bra $L__BB25_210; bra.uni $L__BB25_197; $L__BB25_210: ld.const.u32 %r435, [kernel_params+12]; shl.b32 %r434, %r435, 1; setp.eq.ftz.f32 %p241, %f2629, 0f00000000; mov.b32 %r424, %f2629; add.s32 %r425, %r424, 1; mov.b32 %f2507, %r425; selp.f32 %f2517, 0f00000001, %f2507, %p241; add.s32 %r448, %r448, 1; setp.lt.u32 %p242, %r448, %r434; @%p242 bra $L__BB25_2; $L__BB25_211: ret; } .func (.param .b32 func_retval0) _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection( .param .b32 _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_0, .param .align 4 .b8 _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1[12], .param .b64 _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_2, .param .b64 _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_3 ) { .local .align 16 .b8 __local_depot26[816]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<291>; .reg .b16 %rs<56>; .reg .f32 %f<2454>; .reg .b32 %r<612>; .reg .b64 %rd<354>; mov.u64 %SPL, __local_depot26; ld.param.u32 %r167, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_0]; ld.param.u64 %rd117, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_2]; ld.param.u64 %rd118, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_3]; ld.param.u32 %r1, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1]; ld.param.u32 %r2, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1+4]; ld.param.u32 %r4, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1+8]; cvta.to.local.u64 %rd1, %rd118; add.u64 %rd2, %SPL, 0; ld.const.u32 %r3, [kernel_params+1596]; setp.eq.s32 %p11, %r3, 0; brev.b32 %r168, %r4; mul.lo.s32 %r169, %r168, 1025551850; xor.b32 %r5, %r169, %r168; @%p11 bra $L__BB26_3; ld.const.v2.u32 {%r170, %r171}, [kernel_params+1600]; mov.b32 %f1, %r171; setp.geu.ftz.f32 %p12, %f1, 0f3F800000; setp.lt.ftz.f32 %p13, %f1, 0f3F800000; ld.const.u32 %r174, [kernel_params+1576]; selp.b32 %r175, %r174, %r1, %p13; and.b32 %r176, %r2, 255; xor.b32 %r177, %r175, %r176; mul.lo.s32 %r178, %r177, -512718531; shr.u32 %r179, %r175, 16; xor.b32 %r180, %r178, %r179; bfe.u32 %r181, %r180, 4, 4; shr.u32 %r182, %r175, 8; xor.b32 %r183, %r180, %r182; xor.b32 %r184, %r183, %r181; mul.lo.s32 %r185, %r184, 153742143; shr.u32 %r186, %r175, 23; xor.b32 %r187, %r185, %r186; bfe.u32 %r188, %r187, 1, 7; xor.b32 %r189, %r188, %r187; shr.u32 %r190, %r175, 27; or.b32 %r191, %r190, 1; mul.lo.s32 %r192, %r191, %r189; mul.lo.s32 %r193, %r192, -1074439621; bfe.u32 %r194, %r193, 2, 6; xor.b32 %r195, %r194, %r193; mul.lo.s32 %r196, %r195, -1638916925; bfe.u32 %r197, %r196, 2, 6; xor.b32 %r198, %r197, %r196; mul.lo.s32 %r199, %r198, -933190689; and.b32 %r200, %r199, 255; bfe.u32 %r201, %r199, 5, 3; xor.b32 %r202, %r201, %r200; add.s32 %r203, %r170, -1; xor.b32 %r6, %r2, 61; xor.b32 %r204, %r175, %r6; mul.lo.s32 %r205, %r204, 9; shr.u32 %r206, %r205, 4; xor.b32 %r207, %r206, %r205; mul.lo.s32 %r208, %r207, 668265261; add.s32 %r209, %r208, %r5; shr.u32 %r210, %r208, 16; or.b32 %r211, %r210, 1; mul.lo.s32 %r212, %r211, %r209; mul.lo.s32 %r213, %r212, 89287766; xor.b32 %r214, %r213, %r212; mul.lo.s32 %r215, %r214, 1403136100; xor.b32 %r216, %r215, %r214; brev.b32 %r217, %r216; neg.s32 %r218, %r170; and.b32 %r219, %r4, %r218; and.b32 %r220, %r217, %r203; or.b32 %r221, %r220, %r219; mad.lo.s32 %r222, %r202, %r170, %r221; shl.b32 %r223, %r170, 8; rem.u32 %r224, %r222, %r223; ld.const.u64 %rd121, [kernel_params+2424]; cvta.to.global.u64 %rd122, %rd121; shl.b32 %r225, %r224, 1; mul.wide.u32 %rd123, %r225, 4; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f2303, [%rd124]; ld.global.f32 %f13, [%rd124+4]; @%p12 bra $L__BB26_7; xor.b32 %r226, %r6, %r1; mul.lo.s32 %r227, %r226, 9; shr.u32 %r228, %r227, 4; xor.b32 %r229, %r228, %r227; mul.lo.s32 %r230, %r229, 668265261; cvt.rn.f32.u32 %f552, %r230; mul.ftz.f32 %f553, %f552, 0f2F7FFFFE; xor.b32 %r231, %r1, %r2; xor.b32 %r232, %r231, -905047700; mul.lo.s32 %r233, %r232, 9; shr.u32 %r234, %r233, 4; xor.b32 %r235, %r234, %r233; mul.lo.s32 %r236, %r235, 668265261; cvt.rn.f32.u32 %f554, %r236; mul.ftz.f32 %f555, %f554, 0f2F7FFFFE; fma.rn.ftz.f32 %f556, %f553, %f1, %f2303; fma.rn.ftz.f32 %f557, %f555, %f1, %f13; cvt.rmi.ftz.f32.f32 %f558, %f556; sub.ftz.f32 %f2303, %f556, %f558; cvt.rmi.ftz.f32.f32 %f559, %f557; sub.ftz.f32 %f13, %f557, %f559; bra.uni $L__BB26_7; $L__BB26_3: shr.u32 %r238, %r2, 16; xor.b32 %r239, %r238, %r2; mul.lo.s32 %r240, %r239, 569420461; shr.u32 %r241, %r240, 15; xor.b32 %r242, %r241, %r240; mul.lo.s32 %r243, %r242, -749064809; shr.u32 %r244, %r243, 15; xor.b32 %r245, %r1, %r243; xor.b32 %r246, %r245, %r244; xor.b32 %r7, %r246, -419546133; xor.b32 %r247, %r246, 508809841; add.s32 %r248, %r247, %r5; shr.u32 %r249, %r247, 16; or.b32 %r250, %r249, 1; mul.lo.s32 %r251, %r250, %r248; mul.lo.s32 %r252, %r251, 89287766; xor.b32 %r253, %r252, %r251; mul.lo.s32 %r254, %r253, 1403136100; xor.b32 %r8, %r254, %r253; setp.eq.s32 %p14, %r8, 0; mov.u32 %r557, 0; @%p14 bra $L__BB26_6; mov.u32 %r554, 0; mov.u64 %rd126, sobol_burley_table; mov.u32 %r555, %r8; mov.u32 %r557, %r554; $L__BB26_5: clz.b32 %r257, %r555; add.s32 %r258, %r257, %r554; mul.wide.u32 %rd125, %r258, 4; add.s64 %rd127, %rd126, %rd125; ld.const.u32 %r259, [%rd127+128]; xor.b32 %r557, %r259, %r557; add.s32 %r554, %r258, 1; shl.b32 %r260, %r555, %r257; shl.b32 %r555, %r260, 1; setp.ne.s32 %p15, %r555, 0; @%p15 bra $L__BB26_5; $L__BB26_6: xor.b32 %r261, %r7, -525684874; brev.b32 %r262, %r8; shr.u32 %r263, %r261, 16; or.b32 %r264, %r263, 1; mul.lo.s32 %r265, %r262, 1025551850; xor.b32 %r266, %r265, %r262; add.s32 %r267, %r266, %r261; mul.lo.s32 %r268, %r267, %r264; mul.lo.s32 %r269, %r268, 89287766; xor.b32 %r270, %r269, %r268; xor.b32 %r271, %r7, -1802089138; mul.lo.s32 %r272, %r557, 1025551850; xor.b32 %r273, %r272, %r557; add.s32 %r274, %r273, %r271; shr.u32 %r275, %r271, 16; or.b32 %r276, %r275, 1; mul.lo.s32 %r277, %r274, %r276; mul.lo.s32 %r278, %r277, 89287766; xor.b32 %r279, %r278, %r277; mul.lo.s32 %r280, %r279, 1403136100; xor.b32 %r281, %r280, %r279; brev.b32 %r282, %r281; cvt.rn.f32.u32 %f560, %r282; mul.lo.s32 %r283, %r270, 1403136100; xor.b32 %r284, %r283, %r270; brev.b32 %r285, %r284; cvt.rn.f32.u32 %f561, %r285; mul.ftz.f32 %f13, %f560, 0f2F7FFFFE; mul.ftz.f32 %f2303, %f561, 0f2F7FFFFE; $L__BB26_7: ld.const.u64 %rd128, [kernel_params+2632]; cvta.to.global.u64 %rd129, %rd128; mul.wide.s32 %rd130, %r167, 12; add.s64 %rd131, %rd129, %rd130; ld.global.f32 %f14, [%rd131]; ld.global.f32 %f15, [%rd131+4]; ld.global.f32 %f16, [%rd131+8]; ld.const.u64 %rd132, [kernel_params+2672]; cvta.to.global.u64 %rd133, %rd132; mul.wide.s32 %rd134, %r167, 4; add.s64 %rd135, %rd133, %rd134; ld.global.f32 %f17, [%rd135]; ld.const.u64 %rd136, [kernel_params+2664]; cvta.to.global.u64 %rd137, %rd136; add.s64 %rd138, %rd137, %rd134; ld.global.f32 %f18, [%rd138]; ld.const.u64 %rd139, [kernel_params+2760]; cvta.to.global.u64 %rd140, %rd139; add.s64 %rd141, %rd140, %rd130; ld.const.u64 %rd142, [kernel_params+2720]; cvta.to.global.u64 %rd143, %rd142; add.s64 %rd144, %rd143, %rd134; ld.global.u32 %r16, [%rd144]; cvt.s64.s32 %rd5, %r16; ld.const.u64 %rd145, [kernel_params+2544]; cvta.to.global.u64 %rd146, %rd145; add.s64 %rd147, %rd146, %rd134; ld.global.u32 %r17, [%rd147]; ld.const.u64 %rd148, [kernel_params+2744]; cvta.to.global.u64 %rd149, %rd148; add.s64 %rd150, %rd149, %rd130; ld.global.f32 %f19, [%rd150]; ld.global.f32 %f20, [%rd150+4]; ld.global.f32 %f21, [%rd150+8]; ld.global.f32 %f2313, [%rd141+4]; ld.global.f32 %f2314, [%rd141]; setp.neu.ftz.f32 %p16, %f2314, %f2313; ld.global.f32 %f2312, [%rd141+8]; setp.neu.ftz.f32 %p17, %f2314, %f2312; or.pred %p18, %p16, %p17; sub.ftz.f32 %f562, %f2312, %f2313; neg.ftz.f32 %f563, %f2313; selp.f32 %f564, %f2313, %f563, %p18; neg.ftz.f32 %f565, %f2312; selp.f32 %f566, %f565, %f2312, %p18; add.ftz.f32 %f567, %f2314, %f566; sub.ftz.f32 %f568, %f564, %f2314; mul.ftz.f32 %f569, %f567, %f567; fma.rn.ftz.f32 %f570, %f562, %f562, %f569; fma.rn.ftz.f32 %f571, %f568, %f568, %f570; sqrt.approx.ftz.f32 %f572, %f571; div.approx.ftz.f32 %f2311, %f562, %f572; div.approx.ftz.f32 %f2310, %f567, %f572; div.approx.ftz.f32 %f2309, %f568, %f572; mul.ftz.f32 %f573, %f2313, %f2309; mul.ftz.f32 %f574, %f2312, %f2310; sub.ftz.f32 %f28, %f573, %f574; mul.ftz.f32 %f575, %f2312, %f2311; mul.ftz.f32 %f576, %f2314, %f2309; sub.ftz.f32 %f29, %f575, %f576; mul.ftz.f32 %f577, %f2314, %f2310; mul.ftz.f32 %f578, %f2313, %f2311; sub.ftz.f32 %f30, %f577, %f578; setp.lt.ftz.f32 %p19, %f13, 0f3F000000; @%p19 bra $L__BB26_11; bra.uni $L__BB26_8; $L__BB26_11: add.ftz.f32 %f2305, %f13, %f13; mov.f32 %f2316, 0f3E800000; mov.f32 %f2315, 0f3F000000; mov.f32 %f2306, %f2312; mov.f32 %f2307, %f2313; mov.f32 %f2308, %f2314; mov.f32 %f2312, %f30; mov.f32 %f2313, %f29; mov.f32 %f2314, %f28; mov.f32 %f2317, %f2316; bra.uni $L__BB26_12; $L__BB26_8: setp.lt.ftz.f32 %p20, %f13, 0f3F400000; @%p20 bra $L__BB26_10; bra.uni $L__BB26_9; $L__BB26_10: add.ftz.f32 %f586, %f13, 0fBF000000; mov.f32 %f2316, 0f3F000000; mul.ftz.f32 %f2305, %f586, 0f40800000; mov.f32 %f2315, 0f3E800000; mov.f32 %f2306, %f2309; mov.f32 %f2307, %f2310; mov.f32 %f2308, %f2311; mov.f32 %f2309, %f2312; mov.f32 %f2310, %f2313; mov.f32 %f2311, %f2314; mov.f32 %f2312, %f30; mov.f32 %f2313, %f29; mov.f32 %f2314, %f28; mov.f32 %f2317, %f2315; bra.uni $L__BB26_12; $L__BB26_9: add.ftz.f32 %f582, %f13, 0fBF400000; mul.ftz.f32 %f2305, %f582, 0f40800000; mov.f32 %f2317, 0f3F000000; mov.f32 %f2315, 0f3E800000; mov.f32 %f2306, %f30; mov.f32 %f2307, %f29; mov.f32 %f2308, %f28; mov.f32 %f2316, %f2315; $L__BB26_12: setp.leu.ftz.f32 %p21, %f19, 0f00000000; setp.gt.ftz.f32 %p22, %f19, 0f00000000; mov.f32 %f2318, 0f00000000; selp.f32 %f591, 0f3F800000, 0f00000000, %p22; add.ftz.f32 %f592, %f591, 0f3F800000; setp.gt.ftz.f32 %p23, %f20, 0f00000000; selp.f32 %f593, %f592, %f591, %p23; add.ftz.f32 %f594, %f593, 0f3F800000; setp.gt.ftz.f32 %p24, %f21, 0f00000000; selp.f32 %f48, %f594, %f593, %p24; mul.ftz.f32 %f2323, %f2303, %f48; @%p21 bra $L__BB26_14; setp.lt.ftz.f32 %p25, %f2323, 0f3F800000; mov.f32 %f2318, 0f3F800000; mov.f32 %f2320, 0f00000000; mov.f32 %f2321, %f19; @%p25 bra $L__BB26_17; $L__BB26_14: setp.leu.ftz.f32 %p26, %f20, 0f00000000; mov.f32 %f2319, %f2318; @%p26 bra $L__BB26_16; add.ftz.f32 %f2319, %f2318, 0f3F800000; setp.lt.ftz.f32 %p27, %f2323, %f2319; mov.f32 %f2320, %f2318; mov.f32 %f2321, %f20; @%p27 bra $L__BB26_17; $L__BB26_16: add.ftz.f32 %f598, %f2319, 0f3F800000; setp.geu.ftz.f32 %p28, %f2323, %f598; setp.leu.ftz.f32 %p29, %f21, 0f00000000; mov.f32 %f2322, 0f00000000; or.pred %p30, %p29, %p28; mov.f32 %f2320, %f2319; mov.f32 %f2321, %f21; @%p30 bra $L__BB26_18; $L__BB26_17: sub.ftz.f32 %f2323, %f2323, %f2320; mov.f32 %f2322, %f2321; $L__BB26_18: mul.ftz.f32 %f58, %f2323, 0f3F7F12B2; setp.gtu.ftz.f32 %p31, %f58, 0f3F666666; mov.f32 %f2325, 0f41700000; @%p31 bra $L__BB26_20; mul.ftz.f32 %f600, %f58, %f58; mul.ftz.f32 %f601, %f600, 0f4019999A; mul.ftz.f32 %f602, %f601, 0f3FB8AA3B; ex2.approx.ftz.f32 %f603, %f602; add.ftz.f32 %f2325, %f603, 0fBF800000; $L__BB26_20: neg.ftz.f32 %f604, %f2325; mov.f32 %f605, 0f40400000; div.approx.ftz.f32 %f606, %f604, %f605; mul.ftz.f32 %f607, %f606, 0f3FB8AA3B; ex2.approx.ftz.f32 %f608, %f607; mul.ftz.f32 %f609, %f608, %f608; mul.ftz.f32 %f610, %f608, %f609; mul.ftz.f32 %f611, %f610, 0f3E800000; mov.f32 %f612, 0f3F800000; sub.ftz.f32 %f613, %f612, %f611; fma.rn.ftz.f32 %f614, %f608, 0fBF400000, %f613; sub.ftz.f32 %f61, %f614, %f58; fma.rn.ftz.f32 %f62, %f608, 0f3E800000, %f611; abs.ftz.f32 %f615, %f61; setp.lt.ftz.f32 %p32, %f615, 0f358637BD; setp.eq.ftz.f32 %p33, %f62, 0f00000000; or.pred %p34, %p32, %p33; @%p34 bra $L__BB26_31; div.approx.ftz.f32 %f616, %f61, %f62; sub.ftz.f32 %f617, %f2325, %f616; setp.lt.ftz.f32 %p35, %f617, 0f00000000; selp.f32 %f2325, 0f00000000, %f617, %p35; neg.ftz.f32 %f618, %f2325; div.approx.ftz.f32 %f620, %f618, %f605; mul.ftz.f32 %f621, %f620, 0f3FB8AA3B; ex2.approx.ftz.f32 %f622, %f621; mul.ftz.f32 %f623, %f622, %f622; mul.ftz.f32 %f624, %f622, %f623; mul.ftz.f32 %f625, %f624, 0f3E800000; sub.ftz.f32 %f627, %f612, %f625; fma.rn.ftz.f32 %f628, %f622, 0fBF400000, %f627; sub.ftz.f32 %f64, %f628, %f58; fma.rn.ftz.f32 %f65, %f622, 0f3E800000, %f625; abs.ftz.f32 %f629, %f64; setp.lt.ftz.f32 %p36, %f629, 0f358637BD; setp.eq.ftz.f32 %p37, %f65, 0f00000000; or.pred %p38, %p36, %p37; @%p38 bra $L__BB26_31; div.approx.ftz.f32 %f630, %f64, %f65; sub.ftz.f32 %f631, %f2325, %f630; setp.lt.ftz.f32 %p39, %f631, 0f00000000; selp.f32 %f2325, 0f00000000, %f631, %p39; neg.ftz.f32 %f632, %f2325; div.approx.ftz.f32 %f634, %f632, %f605; mul.ftz.f32 %f635, %f634, 0f3FB8AA3B; ex2.approx.ftz.f32 %f636, %f635; mul.ftz.f32 %f637, %f636, %f636; mul.ftz.f32 %f638, %f636, %f637; mul.ftz.f32 %f639, %f638, 0f3E800000; sub.ftz.f32 %f641, %f612, %f639; fma.rn.ftz.f32 %f642, %f636, 0fBF400000, %f641; sub.ftz.f32 %f67, %f642, %f58; fma.rn.ftz.f32 %f68, %f636, 0f3E800000, %f639; abs.ftz.f32 %f643, %f67; setp.lt.ftz.f32 %p40, %f643, 0f358637BD; setp.eq.ftz.f32 %p41, %f68, 0f00000000; or.pred %p42, %p40, %p41; @%p42 bra $L__BB26_31; div.approx.ftz.f32 %f644, %f67, %f68; sub.ftz.f32 %f645, %f2325, %f644; setp.lt.ftz.f32 %p43, %f645, 0f00000000; selp.f32 %f2325, 0f00000000, %f645, %p43; neg.ftz.f32 %f646, %f2325; div.approx.ftz.f32 %f648, %f646, %f605; mul.ftz.f32 %f649, %f648, 0f3FB8AA3B; ex2.approx.ftz.f32 %f650, %f649; mul.ftz.f32 %f651, %f650, %f650; mul.ftz.f32 %f652, %f650, %f651; mul.ftz.f32 %f653, %f652, 0f3E800000; sub.ftz.f32 %f655, %f612, %f653; fma.rn.ftz.f32 %f656, %f650, 0fBF400000, %f655; sub.ftz.f32 %f70, %f656, %f58; fma.rn.ftz.f32 %f71, %f650, 0f3E800000, %f653; abs.ftz.f32 %f657, %f70; setp.lt.ftz.f32 %p44, %f657, 0f358637BD; setp.eq.ftz.f32 %p45, %f71, 0f00000000; or.pred %p46, %p44, %p45; @%p46 bra $L__BB26_31; div.approx.ftz.f32 %f658, %f70, %f71; sub.ftz.f32 %f659, %f2325, %f658; setp.lt.ftz.f32 %p47, %f659, 0f00000000; selp.f32 %f2325, 0f00000000, %f659, %p47; neg.ftz.f32 %f660, %f2325; div.approx.ftz.f32 %f662, %f660, %f605; mul.ftz.f32 %f663, %f662, 0f3FB8AA3B; ex2.approx.ftz.f32 %f664, %f663; mul.ftz.f32 %f665, %f664, %f664; mul.ftz.f32 %f666, %f664, %f665; mul.ftz.f32 %f667, %f666, 0f3E800000; sub.ftz.f32 %f669, %f612, %f667; fma.rn.ftz.f32 %f670, %f664, 0fBF400000, %f669; sub.ftz.f32 %f73, %f670, %f58; fma.rn.ftz.f32 %f74, %f664, 0f3E800000, %f667; abs.ftz.f32 %f671, %f73; setp.lt.ftz.f32 %p48, %f671, 0f358637BD; setp.eq.ftz.f32 %p49, %f74, 0f00000000; or.pred %p50, %p48, %p49; @%p50 bra $L__BB26_31; div.approx.ftz.f32 %f672, %f73, %f74; sub.ftz.f32 %f673, %f2325, %f672; setp.lt.ftz.f32 %p51, %f673, 0f00000000; selp.f32 %f2325, 0f00000000, %f673, %p51; neg.ftz.f32 %f674, %f2325; div.approx.ftz.f32 %f676, %f674, %f605; mul.ftz.f32 %f677, %f676, 0f3FB8AA3B; ex2.approx.ftz.f32 %f678, %f677; mul.ftz.f32 %f679, %f678, %f678; mul.ftz.f32 %f680, %f678, %f679; mul.ftz.f32 %f681, %f680, 0f3E800000; sub.ftz.f32 %f683, %f612, %f681; fma.rn.ftz.f32 %f684, %f678, 0fBF400000, %f683; sub.ftz.f32 %f76, %f684, %f58; fma.rn.ftz.f32 %f77, %f678, 0f3E800000, %f681; abs.ftz.f32 %f685, %f76; setp.lt.ftz.f32 %p52, %f685, 0f358637BD; setp.eq.ftz.f32 %p53, %f77, 0f00000000; or.pred %p54, %p52, %p53; @%p54 bra $L__BB26_31; div.approx.ftz.f32 %f686, %f76, %f77; sub.ftz.f32 %f687, %f2325, %f686; setp.lt.ftz.f32 %p55, %f687, 0f00000000; selp.f32 %f2325, 0f00000000, %f687, %p55; neg.ftz.f32 %f688, %f2325; div.approx.ftz.f32 %f690, %f688, %f605; mul.ftz.f32 %f691, %f690, 0f3FB8AA3B; ex2.approx.ftz.f32 %f692, %f691; mul.ftz.f32 %f693, %f692, %f692; mul.ftz.f32 %f694, %f692, %f693; mul.ftz.f32 %f695, %f694, 0f3E800000; sub.ftz.f32 %f697, %f612, %f695; fma.rn.ftz.f32 %f698, %f692, 0fBF400000, %f697; sub.ftz.f32 %f79, %f698, %f58; fma.rn.ftz.f32 %f80, %f692, 0f3E800000, %f695; abs.ftz.f32 %f699, %f79; setp.lt.ftz.f32 %p56, %f699, 0f358637BD; setp.eq.ftz.f32 %p57, %f80, 0f00000000; or.pred %p58, %p56, %p57; @%p58 bra $L__BB26_31; div.approx.ftz.f32 %f700, %f79, %f80; sub.ftz.f32 %f701, %f2325, %f700; setp.lt.ftz.f32 %p59, %f701, 0f00000000; selp.f32 %f2325, 0f00000000, %f701, %p59; neg.ftz.f32 %f702, %f2325; div.approx.ftz.f32 %f704, %f702, %f605; mul.ftz.f32 %f705, %f704, 0f3FB8AA3B; ex2.approx.ftz.f32 %f706, %f705; mul.ftz.f32 %f707, %f706, %f706; mul.ftz.f32 %f708, %f706, %f707; mul.ftz.f32 %f709, %f708, 0f3E800000; sub.ftz.f32 %f711, %f612, %f709; fma.rn.ftz.f32 %f712, %f706, 0fBF400000, %f711; sub.ftz.f32 %f82, %f712, %f58; fma.rn.ftz.f32 %f83, %f706, 0f3E800000, %f709; abs.ftz.f32 %f713, %f82; setp.lt.ftz.f32 %p60, %f713, 0f358637BD; setp.eq.ftz.f32 %p61, %f83, 0f00000000; or.pred %p62, %p60, %p61; @%p62 bra $L__BB26_31; div.approx.ftz.f32 %f714, %f82, %f83; sub.ftz.f32 %f715, %f2325, %f714; setp.lt.ftz.f32 %p63, %f715, 0f00000000; selp.f32 %f2325, 0f00000000, %f715, %p63; neg.ftz.f32 %f716, %f2325; div.approx.ftz.f32 %f718, %f716, %f605; mul.ftz.f32 %f719, %f718, 0f3FB8AA3B; ex2.approx.ftz.f32 %f720, %f719; mul.ftz.f32 %f721, %f720, %f720; mul.ftz.f32 %f722, %f720, %f721; mul.ftz.f32 %f723, %f722, 0f3E800000; sub.ftz.f32 %f725, %f612, %f723; fma.rn.ftz.f32 %f726, %f720, 0fBF400000, %f725; sub.ftz.f32 %f85, %f726, %f58; fma.rn.ftz.f32 %f86, %f720, 0f3E800000, %f723; abs.ftz.f32 %f727, %f85; setp.lt.ftz.f32 %p64, %f727, 0f358637BD; setp.eq.ftz.f32 %p65, %f86, 0f00000000; or.pred %p66, %p64, %p65; @%p66 bra $L__BB26_31; div.approx.ftz.f32 %f728, %f85, %f86; sub.ftz.f32 %f729, %f2325, %f728; setp.lt.ftz.f32 %p67, %f729, 0f00000000; selp.f32 %f2325, 0f00000000, %f729, %p67; neg.ftz.f32 %f730, %f2325; div.approx.ftz.f32 %f732, %f730, %f605; mul.ftz.f32 %f733, %f732, 0f3FB8AA3B; ex2.approx.ftz.f32 %f734, %f733; mul.ftz.f32 %f735, %f734, %f734; mul.ftz.f32 %f736, %f734, %f735; mul.ftz.f32 %f737, %f736, 0f3E800000; sub.ftz.f32 %f739, %f612, %f737; fma.rn.ftz.f32 %f740, %f734, 0fBF400000, %f739; sub.ftz.f32 %f88, %f740, %f58; fma.rn.ftz.f32 %f89, %f734, 0f3E800000, %f737; abs.ftz.f32 %f741, %f88; setp.lt.ftz.f32 %p68, %f741, 0f358637BD; setp.eq.ftz.f32 %p69, %f89, 0f00000000; or.pred %p70, %p68, %p69; @%p70 bra $L__BB26_31; div.approx.ftz.f32 %f742, %f88, %f89; sub.ftz.f32 %f743, %f2325, %f742; setp.lt.ftz.f32 %p71, %f743, 0f00000000; selp.f32 %f2325, 0f00000000, %f743, %p71; $L__BB26_31: mul.ftz.f32 %f744, %f2305, 0f40C90FDB; mul.ftz.f32 %f745, %f2322, 0f41800000; mul.ftz.f32 %f92, %f2322, %f2325; mul.ftz.f32 %f746, %f92, %f92; mul.ftz.f32 %f747, %f745, %f745; sub.ftz.f32 %f748, %f747, %f746; mov.f32 %f749, 0f00000000; max.ftz.f32 %f750, %f748, %f749; mov.u32 %r586, 0; sqrt.approx.ftz.f32 %f751, %f750; cos.approx.ftz.f32 %f752, %f744; mul.ftz.f32 %f753, %f92, %f752; sin.approx.ftz.f32 %f754, %f744; mul.ftz.f32 %f755, %f92, %f754; mul.ftz.f32 %f756, %f2314, %f755; mul.ftz.f32 %f757, %f2313, %f755; mul.ftz.f32 %f758, %f2312, %f755; fma.rn.ftz.f32 %f759, %f2311, %f753, %f756; fma.rn.ftz.f32 %f760, %f2310, %f753, %f757; fma.rn.ftz.f32 %f761, %f2309, %f753, %f758; fma.rn.ftz.f32 %f762, %f2308, %f751, %f14; fma.rn.ftz.f32 %f763, %f2307, %f751, %f15; fma.rn.ftz.f32 %f764, %f2306, %f751, %f16; add.ftz.f32 %f2392, %f762, %f759; add.ftz.f32 %f2393, %f763, %f760; add.ftz.f32 %f2394, %f764, %f761; cvta.to.local.u64 %rd6, %rd117; st.local.f32 [%rd6], %f2392; st.local.f32 [%rd6+4], %f2393; st.local.f32 [%rd6+8], %f2394; neg.ftz.f32 %f96, %f2308; st.local.f32 [%rd6+12], %f96; neg.ftz.f32 %f97, %f2307; st.local.f32 [%rd6+16], %f97; neg.ftz.f32 %f98, %f2306; st.local.f32 [%rd6+20], %f98; st.local.u32 [%rd6+24], %r586; add.ftz.f32 %f99, %f751, %f751; st.local.f32 [%rd6+28], %f99; st.local.f32 [%rd6+52], %f17; st.local.u32 [%rd6+56], %r586; st.local.f32 [%rd6+32], %f18; mov.u32 %r287, -1; st.local.u32 [%rd6+40], %r287; st.local.u32 [%rd6+36], %r287; st.local.u32 [%rd6+48], %r287; st.local.u32 [%rd6+44], %r287; add.s32 %r288, %r1, %r2; mad.lo.s32 %r289, %r4, 1757159915, %r288; mad.lo.s32 %r585, %r289, 1103515245, 12345; mov.b32 %r19, %f2392; setp.nan.ftz.f32 %p72, %f2392, %f2392; mov.u16 %rs46, 0; mov.u16 %rs44, %rs46; @%p72 bra $L__BB26_34; or.b32 %r290, %r19, -2147483648; setp.ne.s32 %p73, %r290, -2147483648; add.ftz.f32 %f765, %f2392, %f2392; setp.eq.ftz.f32 %p74, %f765, %f2392; and.pred %p75, %p74, %p73; mov.u16 %rs44, %rs46; @%p75 bra $L__BB26_34; shl.b32 %r291, %r19, 1; setp.lt.u32 %p76, %r291, -16777215; selp.u16 %rs44, 1, 0, %p76; $L__BB26_34: setp.eq.s16 %p77, %rs44, 0; @%p77 bra $L__BB26_40; mov.b32 %r20, %f96; setp.nan.ftz.f32 %p78, %f2308, %f2308; mov.u16 %rs46, 0; mov.u16 %rs45, %rs46; @%p78 bra $L__BB26_38; or.b32 %r292, %r20, -2147483648; setp.ne.s32 %p79, %r292, -2147483648; mul.ftz.f32 %f766, %f2308, 0fC0000000; setp.eq.ftz.f32 %p80, %f766, %f96; and.pred %p81, %p80, %p79; mov.u16 %rs45, %rs46; @%p81 bra $L__BB26_38; shl.b32 %r293, %r20, 1; setp.lt.u32 %p82, %r293, -16777215; selp.u16 %rs45, 1, 0, %p82; $L__BB26_38: setp.eq.s16 %p83, %rs45, 0; @%p83 bra $L__BB26_40; mul.ftz.f32 %f767, %f2307, %f2307; fma.rn.ftz.f32 %f768, %f2308, %f2308, %f767; fma.rn.ftz.f32 %f769, %f2306, %f2306, %f768; setp.neu.ftz.f32 %p84, %f769, 0f00000000; selp.u16 %rs46, 1, 0, %p84; $L__BB26_40: setp.eq.s16 %p85, %rs46, 0; @%p85 bra $L__BB26_155; ld.const.u32 %r294, [kernel_params+1108]; setp.eq.s32 %p86, %r294, 0; abs.ftz.f32 %f770, %f96; setp.gt.ftz.f32 %p87, %f770, 0f17800000; mov.b32 %r295, %f96; and.b32 %r296, %r295, -2147483648; or.b32 %r297, %r296, 394264576; mov.b32 %f771, %r297; selp.f32 %f2389, %f96, %f771, %p87; abs.ftz.f32 %f772, %f97; setp.gt.ftz.f32 %p88, %f772, 0f17800000; mov.b32 %r298, %f97; and.b32 %r299, %r298, -2147483648; or.b32 %r300, %r299, 394264576; mov.b32 %f773, %r300; selp.f32 %f2390, %f97, %f773, %p88; abs.ftz.f32 %f774, %f98; setp.gt.ftz.f32 %p89, %f774, 0f17800000; mov.b32 %r301, %f98; and.b32 %r302, %r301, -2147483648; or.b32 %r303, %r302, 394264576; mov.b32 %f775, %r303; selp.f32 %f2391, %f98, %f775, %p89; @%p86 bra $L__BB26_119; mov.u32 %r304, 1985229328; st.local.u32 [%rd2], %r304; ld.const.u64 %rd151, [kernel_params+2144]; cvta.to.global.u64 %rd152, %rd151; shl.b64 %rd153, %rd5, 2; add.s64 %rd154, %rd152, %rd153; ld.global.u32 %r565, [%rd154]; setp.eq.s64 %p90, %rd118, 0; @%p90 bra $L__BB26_44; st.local.u32 [%rd1], %r586; $L__BB26_44: ld.const.u64 %rd155, [kernel_params+2184]; cvta.to.global.u64 %rd156, %rd155; add.s64 %rd158, %rd156, %rd153; ld.global.u32 %r22, [%rd158]; and.b32 %r306, %r22, 4; setp.eq.s32 %p91, %r306, 0; @%p91 bra $L__BB26_46; bra.uni $L__BB26_45; $L__BB26_46: and.b32 %r307, %r22, 2; setp.eq.s32 %p93, %r307, 0; ld.const.u64 %rd329, [kernel_params+2160]; cvta.to.global.u64 %rd159, %rd329; mul.lo.s64 %rd160, %rd5, 208; add.s64 %rd161, %rd159, %rd160; add.s64 %rd11, %rd161, 48; @%p93 bra $L__BB26_53; bra.uni $L__BB26_47; $L__BB26_53: ld.global.v4.f32 {%f2337, %f2338, %f2339, %f2340}, [%rd11]; ld.global.v4.f32 {%f2341, %f2342, %f2343, %f2344}, [%rd11+16]; ld.global.v4.f32 {%f2345, %f2346, %f2347, %f2348}, [%rd11+32]; bra.uni $L__BB26_54; $L__BB26_155: setp.eq.s64 %p232, %rd118, 0; @%p232 bra $L__BB26_157; mov.u32 %r409, 0; st.local.u32 [%rd1], %r409; bra.uni $L__BB26_157; $L__BB26_119: mov.u32 %r373, 1985229328; st.local.u32 [%rd2], %r373; ld.const.u64 %rd237, [kernel_params+2144]; cvta.to.global.u64 %rd238, %rd237; shl.b64 %rd239, %rd5, 2; add.s64 %rd240, %rd238, %rd239; ld.global.u32 %r595, [%rd240]; setp.eq.s64 %p179, %rd118, 0; @%p179 bra $L__BB26_121; mov.u32 %r374, 0; st.local.u32 [%rd1], %r374; $L__BB26_121: ld.const.u64 %rd241, [kernel_params+2184]; cvta.to.global.u64 %rd242, %rd241; add.s64 %rd244, %rd242, %rd239; ld.global.u8 %rs26, [%rd244]; and.b16 %rs27, %rs26, 4; setp.ne.s16 %p181, %rs27, 0; mov.pred %p180, -1; mov.pred %p289, %p180; @%p181 bra $L__BB26_123; ld.const.u64 %rd245, [kernel_params+2160]; cvta.to.global.u64 %rd246, %rd245; mul.lo.s64 %rd247, %rd5, 208; add.s64 %rd248, %rd246, %rd247; ld.global.v4.f32 {%f1492, %f1493, %f1494, %f1495}, [%rd248+48]; ld.global.v4.f32 {%f1500, %f1501, %f1502, %f1503}, [%rd248+64]; ld.global.v4.f32 {%f1508, %f1509, %f1510, %f1511}, [%rd248+80]; ld.local.f32 %f1516, [%rd6]; ld.local.f32 %f1517, [%rd6+4]; mul.ftz.f32 %f1518, %f1493, %f1517; fma.rn.ftz.f32 %f1519, %f1516, %f1492, %f1518; ld.local.f32 %f1520, [%rd6+8]; fma.rn.ftz.f32 %f1521, %f1494, %f1520, %f1519; add.ftz.f32 %f2392, %f1495, %f1521; mul.ftz.f32 %f1522, %f1501, %f1517; fma.rn.ftz.f32 %f1523, %f1516, %f1500, %f1522; fma.rn.ftz.f32 %f1524, %f1502, %f1520, %f1523; add.ftz.f32 %f2393, %f1503, %f1524; mul.ftz.f32 %f1525, %f1509, %f1517; fma.rn.ftz.f32 %f1526, %f1516, %f1508, %f1525; fma.rn.ftz.f32 %f1527, %f1510, %f1520, %f1526; add.ftz.f32 %f2394, %f1511, %f1527; ld.local.f32 %f1528, [%rd6+12]; ld.local.f32 %f1529, [%rd6+16]; mul.ftz.f32 %f1530, %f1493, %f1529; fma.rn.ftz.f32 %f1531, %f1492, %f1528, %f1530; ld.local.f32 %f1532, [%rd6+20]; fma.rn.ftz.f32 %f1533, %f1494, %f1532, %f1531; mul.ftz.f32 %f1534, %f1501, %f1529; fma.rn.ftz.f32 %f1535, %f1500, %f1528, %f1534; fma.rn.ftz.f32 %f1536, %f1502, %f1532, %f1535; mul.ftz.f32 %f1537, %f1509, %f1529; fma.rn.ftz.f32 %f1538, %f1508, %f1528, %f1537; fma.rn.ftz.f32 %f1539, %f1510, %f1532, %f1538; abs.ftz.f32 %f1540, %f1533; setp.gt.ftz.f32 %p182, %f1540, 0f17800000; mov.b32 %r375, %f1533; and.b32 %r376, %r375, -2147483648; or.b32 %r377, %r376, 394264576; mov.b32 %f1541, %r377; selp.f32 %f2389, %f1533, %f1541, %p182; abs.ftz.f32 %f1542, %f1536; setp.gt.ftz.f32 %p183, %f1542, 0f17800000; mov.b32 %r378, %f1536; and.b32 %r379, %r378, -2147483648; or.b32 %r380, %r379, 394264576; mov.b32 %f1543, %r380; selp.f32 %f2390, %f1536, %f1543, %p183; abs.ftz.f32 %f1544, %f1539; setp.gt.ftz.f32 %p184, %f1544, 0f17800000; mov.b32 %r381, %f1539; and.b32 %r382, %r381, -2147483648; or.b32 %r383, %r382, 394264576; mov.b32 %f1545, %r383; selp.f32 %f2391, %f1539, %f1545, %p184; setp.eq.s32 %p289, %r16, -1; $L__BB26_123: rcp.approx.ftz.f32 %f331, %f2390; rcp.approx.ftz.f32 %f332, %f2389; rcp.approx.ftz.f32 %f333, %f2391; ld.const.v2.u64 {%rd249, %rd250}, [kernel_params+2096]; mov.u32 %r594, 0; cvta.to.global.u64 %rd85, %rd249; cvta.to.global.u64 %rd86, %rd250; ld.const.v2.u64 {%rd253, %rd254}, [kernel_params+2128]; cvta.to.global.u64 %rd87, %rd254; cvta.to.global.u64 %rd88, %rd253; ld.const.u64 %rd257, [kernel_params+2232]; cvta.to.global.u64 %rd89, %rd257; ld.const.u64 %rd258, [kernel_params+2256]; cvta.to.global.u64 %rd90, %rd258; add.s64 %rd91, %rd1, -44; not.pred %p211, %p289; mov.pred %p225, 0; $L__BB26_124: setp.eq.s32 %p185, %r595, 1985229328; setp.lt.s32 %p186, %r595, 0; or.pred %p187, %p186, %p185; @%p187 bra $L__BB26_135; $L__BB26_125: mul.wide.s32 %rd259, %r595, 16; add.s64 %rd92, %rd85, %rd259; ld.global.v4.f32 {%f1547, %f1548, %f1549, %f1550}, [%rd92]; mov.b32 %r595, %f1550; mov.b32 %r108, %f1549; mov.b32 %r107, %f1548; mov.b32 %r106, %f1547; mov.b64 %rd260, {%r106, %r107}; cvt.u32.u64 %r385, %rd260; and.b32 %r386, %r385, 1024; setp.eq.s32 %p188, %r386, 0; and.b32 %r110, %r106, 2047; @%p188 bra $L__BB26_128; bra.uni $L__BB26_126; $L__BB26_128: setp.ne.s32 %p194, %r110, 0; ld.global.v4.f32 {%f1691, %f1692, %f1693, %f1694}, [%rd92+16]; ld.global.v4.f32 {%f1699, %f1700, %f1701, %f1702}, [%rd92+32]; ld.global.v4.f32 {%f1707, %f1708, %f1709, %f1710}, [%rd92+48]; sub.ftz.f32 %f1715, %f1691, %f2392; mul.ftz.f32 %f1716, %f332, %f1715; sub.ftz.f32 %f1717, %f1693, %f2392; mul.ftz.f32 %f1718, %f332, %f1717; sub.ftz.f32 %f1719, %f1699, %f2393; mul.ftz.f32 %f1720, %f331, %f1719; sub.ftz.f32 %f1721, %f1701, %f2393; mul.ftz.f32 %f1722, %f331, %f1721; sub.ftz.f32 %f1723, %f1707, %f2394; mul.ftz.f32 %f1724, %f333, %f1723; sub.ftz.f32 %f1725, %f1709, %f2394; mul.ftz.f32 %f1726, %f333, %f1725; min.ftz.f32 %f1727, %f1716, %f1718; min.ftz.f32 %f1728, %f1720, %f1722; min.ftz.f32 %f1729, %f1724, %f1726; mov.f32 %f1730, 0f00000000; max.ftz.f32 %f1731, %f1730, %f1727; max.ftz.f32 %f1732, %f1728, %f1729; max.ftz.f32 %f2398, %f1731, %f1732; max.ftz.f32 %f1733, %f1716, %f1718; max.ftz.f32 %f1734, %f1720, %f1722; max.ftz.f32 %f1735, %f1724, %f1726; min.ftz.f32 %f1736, %f99, %f1733; min.ftz.f32 %f1737, %f1734, %f1735; min.ftz.f32 %f1738, %f1736, %f1737; sub.ftz.f32 %f1739, %f1692, %f2392; mul.ftz.f32 %f1740, %f332, %f1739; sub.ftz.f32 %f1741, %f1694, %f2392; mul.ftz.f32 %f1742, %f332, %f1741; sub.ftz.f32 %f1743, %f1700, %f2393; mul.ftz.f32 %f1744, %f331, %f1743; sub.ftz.f32 %f1745, %f1702, %f2393; mul.ftz.f32 %f1746, %f331, %f1745; sub.ftz.f32 %f1747, %f1708, %f2394; mul.ftz.f32 %f1748, %f333, %f1747; sub.ftz.f32 %f1749, %f1710, %f2394; mul.ftz.f32 %f1750, %f333, %f1749; min.ftz.f32 %f1751, %f1740, %f1742; min.ftz.f32 %f1752, %f1744, %f1746; min.ftz.f32 %f1753, %f1748, %f1750; max.ftz.f32 %f1754, %f1730, %f1751; max.ftz.f32 %f1755, %f1752, %f1753; max.ftz.f32 %f2399, %f1754, %f1755; max.ftz.f32 %f1756, %f1740, %f1742; max.ftz.f32 %f1757, %f1744, %f1746; max.ftz.f32 %f1758, %f1748, %f1750; min.ftz.f32 %f1759, %f99, %f1756; min.ftz.f32 %f1760, %f1757, %f1758; min.ftz.f32 %f1761, %f1759, %f1760; setp.ge.ftz.f32 %p195, %f1738, %f2398; and.pred %p196, %p194, %p195; selp.u32 %r390, 1, 0, %p196; setp.ltu.ftz.f32 %p197, %f1761, %f2399; and.b32 %r392, %r107, 2047; setp.eq.s32 %p198, %r392, 0; or.pred %p199, %p198, %p197; selp.b32 %r393, 0, 2, %p199; or.b32 %r593, %r393, %r390; bra.uni $L__BB26_129; $L__BB26_126: setp.ne.s32 %p189, %r110, 0; ld.global.v4.f32 {%f1554, %f1555, %f1556, %f1557}, [%rd92+16]; ld.global.v4.f32 {%f1562, %f1563, %f1564, %f1565}, [%rd92+32]; ld.global.v4.f32 {%f1570, %f1571, %f1572, %f1573}, [%rd92+48]; mul.ftz.f32 %f1578, %f2390, %f1555; fma.rn.ftz.f32 %f1579, %f2389, %f1554, %f1578; fma.rn.ftz.f32 %f1580, %f2391, %f1556, %f1579; mul.ftz.f32 %f1581, %f2390, %f1563; fma.rn.ftz.f32 %f1582, %f2389, %f1562, %f1581; fma.rn.ftz.f32 %f1583, %f2391, %f1564, %f1582; mul.ftz.f32 %f1584, %f2390, %f1571; fma.rn.ftz.f32 %f1585, %f2389, %f1570, %f1584; fma.rn.ftz.f32 %f1586, %f2391, %f1572, %f1585; mul.ftz.f32 %f1587, %f2393, %f1555; fma.rn.ftz.f32 %f1588, %f2392, %f1554, %f1587; fma.rn.ftz.f32 %f1589, %f2394, %f1556, %f1588; add.ftz.f32 %f1590, %f1557, %f1589; mul.ftz.f32 %f1591, %f2393, %f1563; fma.rn.ftz.f32 %f1592, %f2392, %f1562, %f1591; fma.rn.ftz.f32 %f1593, %f2394, %f1564, %f1592; add.ftz.f32 %f1594, %f1565, %f1593; mul.ftz.f32 %f1595, %f2393, %f1571; fma.rn.ftz.f32 %f1596, %f2392, %f1570, %f1595; fma.rn.ftz.f32 %f1597, %f2394, %f1572, %f1596; add.ftz.f32 %f1598, %f1573, %f1597; rcp.approx.ftz.f32 %f1599, %f1580; rcp.approx.ftz.f32 %f1600, %f1583; rcp.approx.ftz.f32 %f1601, %f1586; mul.ftz.f32 %f1602, %f1599, %f1590; neg.ftz.f32 %f1603, %f1602; mul.ftz.f32 %f1604, %f1600, %f1594; neg.ftz.f32 %f1605, %f1604; mul.ftz.f32 %f1606, %f1601, %f1598; neg.ftz.f32 %f1607, %f1606; sub.ftz.f32 %f1608, %f1599, %f1602; sub.ftz.f32 %f1609, %f1600, %f1604; sub.ftz.f32 %f1610, %f1601, %f1606; min.ftz.f32 %f1611, %f1603, %f1608; min.ftz.f32 %f1612, %f1605, %f1609; min.ftz.f32 %f1613, %f1607, %f1610; max.ftz.f32 %f1614, %f1603, %f1608; max.ftz.f32 %f1615, %f1605, %f1609; max.ftz.f32 %f1616, %f1607, %f1610; mov.f32 %f1617, 0f00000000; max.ftz.f32 %f1618, %f1617, %f1611; max.ftz.f32 %f1619, %f1612, %f1613; max.ftz.f32 %f2398, %f1618, %f1619; min.ftz.f32 %f1620, %f99, %f1614; min.ftz.f32 %f1621, %f1615, %f1616; min.ftz.f32 %f1622, %f1620, %f1621; setp.le.ftz.f32 %p190, %f2398, %f1622; and.pred %p191, %p189, %p190; selp.u32 %r593, 1, 0, %p191; ld.global.v4.f32 {%f1623, %f1624, %f1625, %f1626}, [%rd92+64]; ld.global.v4.f32 {%f1631, %f1632, %f1633, %f1634}, [%rd92+80]; ld.global.v4.f32 {%f1639, %f1640, %f1641, %f1642}, [%rd92+96]; mul.ftz.f32 %f1647, %f2390, %f1624; fma.rn.ftz.f32 %f1648, %f2389, %f1623, %f1647; fma.rn.ftz.f32 %f1649, %f2391, %f1625, %f1648; mul.ftz.f32 %f1650, %f2390, %f1632; fma.rn.ftz.f32 %f1651, %f2389, %f1631, %f1650; fma.rn.ftz.f32 %f1652, %f2391, %f1633, %f1651; mul.ftz.f32 %f1653, %f2390, %f1640; fma.rn.ftz.f32 %f1654, %f2389, %f1639, %f1653; fma.rn.ftz.f32 %f1655, %f2391, %f1641, %f1654; mul.ftz.f32 %f1656, %f2393, %f1624; fma.rn.ftz.f32 %f1657, %f2392, %f1623, %f1656; fma.rn.ftz.f32 %f1658, %f2394, %f1625, %f1657; add.ftz.f32 %f1659, %f1626, %f1658; mul.ftz.f32 %f1660, %f2393, %f1632; fma.rn.ftz.f32 %f1661, %f2392, %f1631, %f1660; fma.rn.ftz.f32 %f1662, %f2394, %f1633, %f1661; add.ftz.f32 %f1663, %f1634, %f1662; mul.ftz.f32 %f1664, %f2393, %f1640; fma.rn.ftz.f32 %f1665, %f2392, %f1639, %f1664; fma.rn.ftz.f32 %f1666, %f2394, %f1641, %f1665; add.ftz.f32 %f1667, %f1642, %f1666; rcp.approx.ftz.f32 %f1668, %f1649; rcp.approx.ftz.f32 %f1669, %f1652; rcp.approx.ftz.f32 %f1670, %f1655; mul.ftz.f32 %f1671, %f1668, %f1659; neg.ftz.f32 %f1672, %f1671; mul.ftz.f32 %f1673, %f1669, %f1663; neg.ftz.f32 %f1674, %f1673; mul.ftz.f32 %f1675, %f1670, %f1667; neg.ftz.f32 %f1676, %f1675; sub.ftz.f32 %f1677, %f1668, %f1671; sub.ftz.f32 %f1678, %f1669, %f1673; sub.ftz.f32 %f1679, %f1670, %f1675; min.ftz.f32 %f1680, %f1672, %f1677; min.ftz.f32 %f1681, %f1674, %f1678; min.ftz.f32 %f1682, %f1676, %f1679; max.ftz.f32 %f1683, %f1672, %f1677; max.ftz.f32 %f1684, %f1674, %f1678; max.ftz.f32 %f1685, %f1676, %f1679; max.ftz.f32 %f1686, %f1617, %f1680; max.ftz.f32 %f1687, %f1681, %f1682; max.ftz.f32 %f2399, %f1686, %f1687; min.ftz.f32 %f1688, %f99, %f1683; min.ftz.f32 %f1689, %f1684, %f1685; min.ftz.f32 %f1690, %f1688, %f1689; setp.gtu.ftz.f32 %p192, %f2399, %f1690; @%p192 bra $L__BB26_129; and.b32 %r388, %r107, 2047; setp.eq.s32 %p193, %r388, 0; or.b32 %r389, %r593, 2; selp.b32 %r593, %r593, %r389, %p193; $L__BB26_129: mul.wide.s32 %rd261, %r594, 4; add.s64 %rd93, %rd2, %rd261; setp.eq.s32 %p200, %r593, 3; @%p200 bra $L__BB26_133; bra.uni $L__BB26_130; $L__BB26_133: setp.lt.ftz.f32 %p203, %f2399, %f2398; selp.b32 %r119, %r595, %r108, %p203; selp.b32 %r394, %r108, %r595, %p203; add.s32 %r594, %r594, 1; st.local.u32 [%rd93+4], %r394; mov.u32 %r595, %r119; bra.uni $L__BB26_134; $L__BB26_130: setp.eq.s32 %p201, %r593, 2; @%p201 bra $L__BB26_134; setp.ne.s32 %p202, %r593, 0; mov.u32 %r595, %r108; @%p202 bra $L__BB26_134; ld.local.u32 %r595, [%rd93]; add.s32 %r594, %r594, -1; $L__BB26_134: setp.ne.s32 %p204, %r595, 1985229328; setp.gt.s32 %p205, %r595, -1; and.pred %p206, %p205, %p204; @%p206 bra $L__BB26_125; $L__BB26_135: mov.u32 %r123, %r594; setp.gt.s32 %p207, %r595, -1; @%p207 bra $L__BB26_154; mul.wide.s32 %rd262, %r595, 16; xor.b64 %rd263, %rd262, -16; add.s64 %rd264, %rd86, %rd263; ld.global.v4.u32 {%r598, %r396, %r397, %r398}, [%rd264]; mul.wide.s32 %rd265, %r123, 4; add.s64 %rd266, %rd2, %rd265; ld.local.u32 %r595, [%rd266]; add.s32 %r594, %r123, -1; and.b32 %r400, %r398, 127; setp.ne.s32 %p208, %r400, 1; setp.ge.s32 %p209, %r598, %r396; or.pred %p210, %p208, %p209; @%p210 bra $L__BB26_154; $L__BB26_137: cvt.s64.s32 %rd94, %r598; @%p211 bra $L__BB26_139; shl.b64 %rd267, %rd94, 2; add.s64 %rd268, %rd87, %rd267; ld.global.u32 %r401, [%rd268]; setp.ne.s32 %p212, %r401, %r16; @%p212 bra $L__BB26_153; $L__BB26_139: shl.b64 %rd269, %rd94, 2; add.s64 %rd270, %rd88, %rd269; ld.global.u32 %r131, [%rd270]; ld.local.u32 %r402, [%rd6+36]; setp.eq.s32 %p213, %r402, %r131; @%p213 bra $L__BB26_153; mul.wide.s32 %rd271, %r131, 16; add.s64 %rd272, %rd89, %rd271; ld.global.u32 %r403, [%rd272+12]; mul.wide.u32 %rd273, %r403, 12; add.s64 %rd274, %rd90, %rd273; add.s32 %r404, %r403, 1; mul.wide.u32 %rd275, %r404, 12; add.s64 %rd276, %rd90, %rd275; add.s32 %r405, %r403, 2; mul.wide.u32 %rd277, %r405, 12; add.s64 %rd278, %rd90, %rd277; ld.global.f32 %f347, [%rd274]; sub.ftz.f32 %f348, %f347, %f2392; ld.global.f32 %f349, [%rd274+4]; sub.ftz.f32 %f350, %f349, %f2393; ld.global.f32 %f351, [%rd274+8]; sub.ftz.f32 %f352, %f351, %f2394; ld.global.f32 %f353, [%rd276]; sub.ftz.f32 %f1762, %f353, %f2392; ld.global.f32 %f354, [%rd276+4]; sub.ftz.f32 %f1763, %f354, %f2393; ld.global.f32 %f355, [%rd276+8]; sub.ftz.f32 %f1764, %f355, %f2394; ld.global.f32 %f356, [%rd278]; sub.ftz.f32 %f1765, %f356, %f2392; ld.global.f32 %f357, [%rd278+4]; sub.ftz.f32 %f1766, %f357, %f2393; ld.global.f32 %f358, [%rd278+8]; sub.ftz.f32 %f1767, %f358, %f2394; sub.ftz.f32 %f359, %f1765, %f348; sub.ftz.f32 %f360, %f1766, %f350; sub.ftz.f32 %f361, %f1767, %f352; sub.ftz.f32 %f362, %f348, %f1762; sub.ftz.f32 %f363, %f350, %f1763; sub.ftz.f32 %f364, %f352, %f1764; sub.ftz.f32 %f1768, %f1762, %f1765; sub.ftz.f32 %f1769, %f1763, %f1766; sub.ftz.f32 %f1770, %f1764, %f1767; add.ftz.f32 %f1771, %f348, %f1765; add.ftz.f32 %f1772, %f350, %f1766; add.ftz.f32 %f1773, %f352, %f1767; mul.ftz.f32 %f1774, %f360, %f1773; mul.ftz.f32 %f1775, %f1772, %f361; sub.ftz.f32 %f1776, %f1774, %f1775; mul.ftz.f32 %f1777, %f1771, %f361; mul.ftz.f32 %f1778, %f359, %f1773; sub.ftz.f32 %f1779, %f1777, %f1778; mul.ftz.f32 %f1780, %f359, %f1772; mul.ftz.f32 %f1781, %f1771, %f360; sub.ftz.f32 %f1782, %f1780, %f1781; mul.ftz.f32 %f1783, %f2390, %f1779; fma.rn.ftz.f32 %f1784, %f2389, %f1776, %f1783; fma.rn.ftz.f32 %f365, %f2391, %f1782, %f1784; add.ftz.f32 %f1785, %f348, %f1762; add.ftz.f32 %f1786, %f350, %f1763; add.ftz.f32 %f1787, %f352, %f1764; mul.ftz.f32 %f1788, %f363, %f1787; mul.ftz.f32 %f1789, %f1786, %f364; sub.ftz.f32 %f1790, %f1788, %f1789; mul.ftz.f32 %f1791, %f1785, %f364; mul.ftz.f32 %f1792, %f362, %f1787; sub.ftz.f32 %f1793, %f1791, %f1792; mul.ftz.f32 %f1794, %f362, %f1786; mul.ftz.f32 %f1795, %f1785, %f363; sub.ftz.f32 %f1796, %f1794, %f1795; mul.ftz.f32 %f1797, %f2390, %f1793; fma.rn.ftz.f32 %f1798, %f2389, %f1790, %f1797; fma.rn.ftz.f32 %f366, %f2391, %f1796, %f1798; add.ftz.f32 %f1799, %f1762, %f1765; add.ftz.f32 %f1800, %f1763, %f1766; add.ftz.f32 %f1801, %f1764, %f1767; mul.ftz.f32 %f1802, %f1769, %f1801; mul.ftz.f32 %f1803, %f1800, %f1770; sub.ftz.f32 %f1804, %f1802, %f1803; mul.ftz.f32 %f1805, %f1799, %f1770; mul.ftz.f32 %f1806, %f1768, %f1801; sub.ftz.f32 %f1807, %f1805, %f1806; mul.ftz.f32 %f1808, %f1768, %f1800; mul.ftz.f32 %f1809, %f1799, %f1769; sub.ftz.f32 %f1810, %f1808, %f1809; mul.ftz.f32 %f1811, %f2390, %f1807; fma.rn.ftz.f32 %f1812, %f2389, %f1804, %f1811; fma.rn.ftz.f32 %f1813, %f2391, %f1810, %f1812; add.ftz.f32 %f1814, %f366, %f365; add.ftz.f32 %f367, %f1813, %f1814; abs.ftz.f32 %f368, %f367; mul.ftz.f32 %f1815, %f368, 0f34000000; min.ftz.f32 %f1816, %f366, %f1813; min.ftz.f32 %f1817, %f365, %f1816; max.ftz.f32 %f1818, %f366, %f1813; max.ftz.f32 %f1819, %f365, %f1818; mul.ftz.f32 %f1820, %f368, 0fB4000000; setp.ltu.ftz.f32 %p215, %f1817, %f1820; setp.gtu.ftz.f32 %p216, %f1819, %f1815; and.pred %p217, %p215, %p216; mov.pred %p290, %p180; @%p217 bra $L__BB26_146; mul.ftz.f32 %f1821, %f363, %f361; mul.ftz.f32 %f1822, %f364, %f360; sub.ftz.f32 %f1823, %f1821, %f1822; mul.ftz.f32 %f1824, %f362, %f361; mul.ftz.f32 %f1825, %f364, %f359; sub.ftz.f32 %f1826, %f1825, %f1824; mul.ftz.f32 %f1827, %f363, %f359; mul.ftz.f32 %f1828, %f362, %f360; sub.ftz.f32 %f1829, %f1828, %f1827; add.ftz.f32 %f369, %f1823, %f1823; add.ftz.f32 %f370, %f1826, %f1826; add.ftz.f32 %f371, %f1829, %f1829; mul.ftz.f32 %f1830, %f2390, %f370; fma.rn.ftz.f32 %f1831, %f2389, %f369, %f1830; fma.rn.ftz.f32 %f372, %f2391, %f371, %f1831; setp.eq.ftz.f32 %p219, %f372, 0f00000000; mov.pred %p290, %p180; @%p219 bra $L__BB26_146; sub.ftz.f32 %f2300, %f351, %f2394; sub.ftz.f32 %f2299, %f347, %f2392; sub.ftz.f32 %f2298, %f349, %f2393; mul.ftz.f32 %f1832, %f2298, %f370; fma.rn.ftz.f32 %f1833, %f2299, %f369, %f1832; fma.rn.ftz.f32 %f1834, %f2300, %f371, %f1833; div.approx.ftz.f32 %f373, %f1834, %f372; setp.ltu.ftz.f32 %p221, %f373, 0f00000000; setp.gtu.ftz.f32 %p222, %f373, %f99; or.pred %p223, %p222, %p221; mov.pred %p290, %p180; @%p223 bra $L__BB26_146; setp.lt.ftz.f32 %p224, %f368, 0f219392EF; mov.f32 %f2403, 0f00000000; @%p224 bra $L__BB26_145; rcp.approx.ftz.f32 %f2403, %f367; $L__BB26_145: mul.ftz.f32 %f1836, %f365, %f2403; min.ftz.f32 %f2408, %f1836, %f612; mul.ftz.f32 %f1838, %f366, %f2403; min.ftz.f32 %f2407, %f1838, %f612; mov.f32 %f2409, %f373; mov.pred %p290, %p225; $L__BB26_146: @%p290 bra $L__BB26_153; ld.local.u32 %r602, [%rd1]; min.s32 %r406, %r602, 4; add.s32 %r600, %r406, 1; mul.wide.s32 %rd279, %r600, 24; add.s64 %rd349, %rd91, %rd279; $L__BB26_148: add.s32 %r600, %r600, -1; setp.gt.s32 %p226, %r600, 0; @%p226 bra $L__BB26_152; bra.uni $L__BB26_149; $L__BB26_152: add.s64 %rd97, %rd349, -24; ld.local.f32 %f1861, [%rd349]; setp.neu.ftz.f32 %p229, %f1861, %f2409; mov.u64 %rd349, %rd97; @%p229 bra $L__BB26_148; bra.uni $L__BB26_153; $L__BB26_149: setp.lt.s32 %p227, %r602, 4; add.s32 %r136, %r602, 1; st.local.u32 [%rd1], %r136; @%p227 bra $L__BB26_151; mad.lo.s32 %r585, %r585, 1103515245, 12345; rem.u32 %r602, %r585, %r136; setp.gt.s32 %p228, %r602, 3; @%p228 bra $L__BB26_153; $L__BB26_151: mul.wide.s32 %rd280, %r602, 24; add.s64 %rd281, %rd1, %rd280; st.local.u32 [%rd281+16], %r131; st.local.u32 [%rd281+20], %r16; mov.u32 %r407, 1; st.local.u32 [%rd281+24], %r407; st.local.f32 [%rd281+8], %f2408; st.local.f32 [%rd281+12], %f2407; st.local.f32 [%rd281+4], %f2409; sub.ftz.f32 %f1839, %f358, %f351; sub.ftz.f32 %f1840, %f354, %f349; mul.ftz.f32 %f1841, %f1840, %f1839; sub.ftz.f32 %f1842, %f357, %f349; sub.ftz.f32 %f1843, %f355, %f351; mul.ftz.f32 %f1844, %f1843, %f1842; sub.ftz.f32 %f1845, %f1841, %f1844; sub.ftz.f32 %f1846, %f356, %f347; mul.ftz.f32 %f1847, %f1843, %f1846; sub.ftz.f32 %f1848, %f353, %f347; mul.ftz.f32 %f1849, %f1848, %f1839; sub.ftz.f32 %f1850, %f1847, %f1849; mul.ftz.f32 %f1851, %f1848, %f1842; mul.ftz.f32 %f1852, %f1840, %f1846; sub.ftz.f32 %f1853, %f1851, %f1852; mul.ftz.f32 %f1854, %f1850, %f1850; fma.rn.ftz.f32 %f1855, %f1845, %f1845, %f1854; fma.rn.ftz.f32 %f1856, %f1853, %f1853, %f1855; sqrt.approx.ftz.f32 %f1857, %f1856; div.approx.ftz.f32 %f1858, %f1845, %f1857; div.approx.ftz.f32 %f1859, %f1850, %f1857; div.approx.ftz.f32 %f1860, %f1853, %f1857; mul.wide.s32 %rd282, %r602, 12; add.s64 %rd283, %rd1, %rd282; st.local.f32 [%rd283+100], %f1858; st.local.f32 [%rd283+104], %f1859; st.local.f32 [%rd283+108], %f1860; $L__BB26_153: add.s32 %r594, %r123, -1; cvt.u32.u64 %r408, %rd94; add.s32 %r598, %r408, 1; setp.lt.s32 %p230, %r598, %r396; @%p230 bra $L__BB26_137; $L__BB26_154: setp.ne.s32 %p231, %r595, 1985229328; @%p231 bra $L__BB26_124; bra.uni $L__BB26_157; $L__BB26_45: mov.pred %p286, -1; ld.const.u64 %rd329, [kernel_params+2160]; bra.uni $L__BB26_55; $L__BB26_47: ld.const.u64 %rd162, [kernel_params+2176]; cvta.to.global.u64 %rd163, %rd162; ld.global.u32 %rd164, [%rd11+120]; ld.global.u32 %r308, [%rd11+104]; shl.b32 %r309, %r308, 1; cvt.rn.f32.s32 %f776, %r309; ld.local.f32 %f777, [%rd6+32]; mul.ftz.f32 %f778, %f777, %f776; cvt.rzi.ftz.s32.f32 %r310, %f778; add.s32 %r311, %r309, -1; min.s32 %r312, %r310, %r311; cvt.rn.f32.s32 %f779, %r312; sub.ftz.f32 %f103, %f778, %f779; cvt.s64.s32 %rd165, %r312; add.s64 %rd166, %rd165, %rd164; shl.b64 %rd167, %rd166, 6; add.s64 %rd12, %rd163, %rd167; ld.global.v4.f32 {%f780, %f781, %f782, %f783}, [%rd12]; ld.global.v4.f32 {%f784, %f785, %f786, %f787}, [%rd12+64]; mul.ftz.f32 %f788, %f781, %f785; fma.rn.ftz.f32 %f789, %f780, %f784, %f788; mul.ftz.f32 %f790, %f783, %f787; fma.rn.ftz.f32 %f791, %f782, %f786, %f790; add.ftz.f32 %f112, %f789, %f791; setp.gt.ftz.f32 %p94, %f112, 0f3F7FDF3B; sub.ftz.f32 %f113, %f612, %f103; @%p94 bra $L__BB26_49; bra.uni $L__BB26_48; $L__BB26_49: mul.ftz.f32 %f846, %f780, %f113; fma.rn.ftz.f32 %f847, %f784, %f103, %f846; mul.ftz.f32 %f848, %f781, %f113; fma.rn.ftz.f32 %f849, %f785, %f103, %f848; mul.ftz.f32 %f850, %f782, %f113; fma.rn.ftz.f32 %f851, %f786, %f103, %f850; mul.ftz.f32 %f852, %f783, %f113; fma.rn.ftz.f32 %f853, %f787, %f103, %f852; mul.ftz.f32 %f854, %f849, %f849; fma.rn.ftz.f32 %f855, %f847, %f847, %f854; mul.ftz.f32 %f856, %f853, %f853; fma.rn.ftz.f32 %f857, %f851, %f851, %f856; add.ftz.f32 %f858, %f855, %f857; rsqrt.approx.ftz.f32 %f859, %f858; mul.ftz.f32 %f2329, %f853, %f859; mul.ftz.f32 %f2328, %f851, %f859; mul.ftz.f32 %f2327, %f849, %f859; mul.ftz.f32 %f2326, %f847, %f859; bra.uni $L__BB26_50; $L__BB26_48: mov.f32 %f793, 0fBF800000; max.ftz.f32 %f794, %f112, %f793; min.ftz.f32 %f796, %f794, %f612; abs.ftz.f32 %f797, %f796; sub.ftz.f32 %f798, %f612, %f797; mul.ftz.f32 %f799, %f798, 0f3F000000; sqrt.approx.ftz.f32 %f800, %f799; setp.gt.ftz.f32 %p95, %f797, 0f3F11EB85; selp.f32 %f801, %f800, %f797, %p95; mul.ftz.f32 %f802, %f801, %f801; mov.f32 %f803, 0f3C94D2E9; mov.f32 %f804, 0f3D53F941; fma.rn.ftz.f32 %f805, %f804, %f802, %f803; mov.f32 %f806, 0f3D3F841F; fma.rn.ftz.f32 %f807, %f805, %f802, %f806; mov.f32 %f808, 0f3D994929; fma.rn.ftz.f32 %f809, %f807, %f802, %f808; mov.f32 %f810, 0f3E2AAB94; fma.rn.ftz.f32 %f811, %f809, %f802, %f810; mul.ftz.f32 %f812, %f802, %f811; fma.rn.ftz.f32 %f813, %f812, %f801, %f801; add.ftz.f32 %f814, %f813, %f813; mov.f32 %f815, 0f3FC90FDB; sub.ftz.f32 %f816, %f815, %f813; selp.f32 %f817, %f814, %f816, %p95; setp.lt.ftz.f32 %p96, %f796, 0f00000000; mov.f32 %f818, 0f40490FDB; sub.ftz.f32 %f819, %f818, %f817; selp.f32 %f820, %f819, %f817, %p96; mul.ftz.f32 %f821, %f780, %f112; sub.ftz.f32 %f822, %f784, %f821; mul.ftz.f32 %f823, %f781, %f112; sub.ftz.f32 %f824, %f785, %f823; mul.ftz.f32 %f825, %f782, %f112; sub.ftz.f32 %f826, %f786, %f825; mul.ftz.f32 %f827, %f783, %f112; sub.ftz.f32 %f828, %f787, %f827; mul.ftz.f32 %f829, %f824, %f824; fma.rn.ftz.f32 %f830, %f822, %f822, %f829; mul.ftz.f32 %f831, %f828, %f828; fma.rn.ftz.f32 %f832, %f826, %f826, %f831; add.ftz.f32 %f833, %f830, %f832; rsqrt.approx.ftz.f32 %f834, %f833; mul.ftz.f32 %f835, %f822, %f834; mul.ftz.f32 %f836, %f824, %f834; mul.ftz.f32 %f837, %f826, %f834; mul.ftz.f32 %f838, %f828, %f834; mul.ftz.f32 %f839, %f103, %f820; cos.approx.ftz.f32 %f840, %f839; sin.approx.ftz.f32 %f841, %f839; mul.ftz.f32 %f842, %f835, %f841; mul.ftz.f32 %f843, %f836, %f841; mul.ftz.f32 %f844, %f837, %f841; mul.ftz.f32 %f845, %f838, %f841; fma.rn.ftz.f32 %f2329, %f783, %f840, %f845; fma.rn.ftz.f32 %f2328, %f782, %f840, %f844; fma.rn.ftz.f32 %f2327, %f781, %f840, %f843; fma.rn.ftz.f32 %f2326, %f780, %f840, %f842; $L__BB26_50: ld.global.v4.f32 {%f860, %f861, %f862, %f863}, [%rd12+16]; ld.global.v4.f32 {%f868, %f869, %f870, %f871}, [%rd12+80]; mul.ftz.f32 %f876, %f103, %f868; mul.ftz.f32 %f877, %f103, %f869; mul.ftz.f32 %f878, %f103, %f870; mul.ftz.f32 %f879, %f103, %f871; fma.rn.ftz.f32 %f126, %f113, %f860, %f876; fma.rn.ftz.f32 %f127, %f113, %f861, %f877; fma.rn.ftz.f32 %f128, %f113, %f862, %f878; fma.rn.ftz.f32 %f880, %f113, %f863, %f879; ld.global.v4.f32 {%f881, %f882, %f883, %f884}, [%rd12+32]; ld.global.v4.f32 {%f889, %f890, %f891, %f892}, [%rd12+96]; mul.ftz.f32 %f897, %f103, %f889; mul.ftz.f32 %f898, %f103, %f890; mul.ftz.f32 %f899, %f103, %f891; mul.ftz.f32 %f900, %f103, %f892; fma.rn.ftz.f32 %f901, %f113, %f881, %f897; fma.rn.ftz.f32 %f902, %f113, %f882, %f898; fma.rn.ftz.f32 %f903, %f113, %f883, %f899; fma.rn.ftz.f32 %f904, %f113, %f884, %f900; ld.global.v4.f32 {%f905, %f906, %f907, %f908}, [%rd12+48]; ld.global.v4.f32 {%f913, %f914, %f915, %f916}, [%rd12+112]; mul.ftz.f32 %f921, %f103, %f913; mul.ftz.f32 %f922, %f103, %f914; mul.ftz.f32 %f923, %f103, %f915; mul.ftz.f32 %f924, %f103, %f916; fma.rn.ftz.f32 %f925, %f113, %f905, %f921; fma.rn.ftz.f32 %f926, %f113, %f906, %f922; fma.rn.ftz.f32 %f927, %f113, %f907, %f923; fma.rn.ftz.f32 %f928, %f113, %f908, %f924; mul.ftz.f32 %f929, %f2329, 0f3FB504F3; mul.ftz.f32 %f930, %f2326, 0f3FB504F3; mul.ftz.f32 %f931, %f2327, 0f3FB504F3; mul.ftz.f32 %f932, %f2328, 0f3FB504F3; mul.ftz.f32 %f933, %f929, %f930; mul.ftz.f32 %f934, %f929, %f931; mul.ftz.f32 %f935, %f929, %f932; mul.ftz.f32 %f936, %f930, %f930; mul.ftz.f32 %f937, %f930, %f931; mul.ftz.f32 %f938, %f930, %f932; mul.ftz.f32 %f939, %f931, %f931; mul.ftz.f32 %f940, %f931, %f932; mul.ftz.f32 %f941, %f932, %f932; sub.ftz.f32 %f943, %f612, %f939; sub.ftz.f32 %f944, %f943, %f941; sub.ftz.f32 %f945, %f937, %f935; add.ftz.f32 %f946, %f934, %f938; add.ftz.f32 %f947, %f935, %f937; sub.ftz.f32 %f948, %f612, %f936; sub.ftz.f32 %f949, %f948, %f941; sub.ftz.f32 %f950, %f940, %f933; sub.ftz.f32 %f951, %f938, %f934; add.ftz.f32 %f952, %f933, %f940; sub.ftz.f32 %f953, %f948, %f939; mul.ftz.f32 %f954, %f945, %f903; fma.rn.ftz.f32 %f955, %f944, %f880, %f954; fma.rn.ftz.f32 %f2335, %f946, %f926, %f955; mul.ftz.f32 %f956, %f945, %f904; fma.rn.ftz.f32 %f957, %f944, %f901, %f956; fma.rn.ftz.f32 %f130, %f946, %f927, %f957; mul.ftz.f32 %f958, %f945, %f925; fma.rn.ftz.f32 %f959, %f944, %f902, %f958; fma.rn.ftz.f32 %f131, %f946, %f928, %f959; mul.ftz.f32 %f960, %f949, %f903; fma.rn.ftz.f32 %f961, %f947, %f880, %f960; fma.rn.ftz.f32 %f132, %f950, %f926, %f961; mul.ftz.f32 %f962, %f949, %f904; fma.rn.ftz.f32 %f963, %f947, %f901, %f962; fma.rn.ftz.f32 %f2334, %f950, %f927, %f963; mul.ftz.f32 %f964, %f949, %f925; fma.rn.ftz.f32 %f965, %f947, %f902, %f964; fma.rn.ftz.f32 %f134, %f950, %f928, %f965; mul.ftz.f32 %f966, %f952, %f903; fma.rn.ftz.f32 %f967, %f951, %f880, %f966; fma.rn.ftz.f32 %f135, %f953, %f926, %f967; mul.ftz.f32 %f968, %f952, %f904; fma.rn.ftz.f32 %f969, %f951, %f901, %f968; fma.rn.ftz.f32 %f136, %f953, %f927, %f969; mul.ftz.f32 %f970, %f952, %f925; fma.rn.ftz.f32 %f971, %f951, %f902, %f970; fma.rn.ftz.f32 %f2333, %f953, %f928, %f971; mul.ftz.f32 %f972, %f2334, %f2333; mul.ftz.f32 %f138, %f136, %f134; sub.ftz.f32 %f2332, %f972, %f138; mul.ftz.f32 %f140, %f136, %f131; mul.ftz.f32 %f973, %f130, %f2333; sub.ftz.f32 %f2331, %f140, %f973; mul.ftz.f32 %f142, %f130, %f134; mul.ftz.f32 %f974, %f2334, %f131; sub.ftz.f32 %f2330, %f142, %f974; mul.ftz.f32 %f975, %f132, %f2331; fma.rn.ftz.f32 %f976, %f2335, %f2332, %f975; fma.rn.ftz.f32 %f2336, %f135, %f2330, %f976; setp.neu.ftz.f32 %p97, %f2336, 0f00000000; @%p97 bra $L__BB26_52; add.ftz.f32 %f2335, %f2335, 0f322BCC77; add.ftz.f32 %f2333, %f2333, 0f322BCC77; add.ftz.f32 %f2334, %f2334, 0f322BCC77; mul.ftz.f32 %f977, %f2334, %f2333; sub.ftz.f32 %f2332, %f977, %f138; mul.ftz.f32 %f978, %f130, %f2333; sub.ftz.f32 %f2331, %f140, %f978; mul.ftz.f32 %f979, %f2334, %f131; sub.ftz.f32 %f2330, %f142, %f979; mul.ftz.f32 %f980, %f132, %f2331; fma.rn.ftz.f32 %f981, %f2335, %f2332, %f980; fma.rn.ftz.f32 %f982, %f135, %f2330, %f981; setp.eq.ftz.f32 %p98, %f982, 0f00000000; selp.f32 %f2336, 0f7F7FFFFF, %f982, %p98; $L__BB26_52: mul.ftz.f32 %f983, %f132, %f2333; mul.ftz.f32 %f984, %f135, %f134; sub.ftz.f32 %f985, %f984, %f983; mul.ftz.f32 %f986, %f135, %f131; mul.ftz.f32 %f987, %f2333, %f2335; sub.ftz.f32 %f988, %f987, %f986; mul.ftz.f32 %f989, %f134, %f2335; mul.ftz.f32 %f990, %f132, %f131; sub.ftz.f32 %f991, %f990, %f989; div.approx.ftz.f32 %f2341, %f985, %f2336; div.approx.ftz.f32 %f2342, %f988, %f2336; div.approx.ftz.f32 %f2343, %f991, %f2336; mul.ftz.f32 %f992, %f135, %f2334; mul.ftz.f32 %f993, %f132, %f136; sub.ftz.f32 %f994, %f993, %f992; mul.ftz.f32 %f995, %f136, %f2335; mul.ftz.f32 %f996, %f135, %f130; sub.ftz.f32 %f997, %f996, %f995; mul.ftz.f32 %f998, %f132, %f130; mul.ftz.f32 %f999, %f2334, %f2335; sub.ftz.f32 %f1000, %f999, %f998; div.approx.ftz.f32 %f2345, %f994, %f2336; div.approx.ftz.f32 %f2346, %f997, %f2336; div.approx.ftz.f32 %f2347, %f1000, %f2336; div.approx.ftz.f32 %f2337, %f2332, %f2336; div.approx.ftz.f32 %f2338, %f2331, %f2336; mul.ftz.f32 %f1001, %f127, %f2338; fma.rn.ftz.f32 %f1002, %f126, %f2337, %f1001; div.approx.ftz.f32 %f2339, %f2330, %f2336; fma.rn.ftz.f32 %f1003, %f128, %f2339, %f1002; neg.ftz.f32 %f2340, %f1003; mul.ftz.f32 %f1004, %f127, %f2342; fma.rn.ftz.f32 %f1005, %f126, %f2341, %f1004; fma.rn.ftz.f32 %f1006, %f128, %f2343, %f1005; neg.ftz.f32 %f2344, %f1006; mul.ftz.f32 %f1007, %f127, %f2346; fma.rn.ftz.f32 %f1008, %f126, %f2345, %f1007; fma.rn.ftz.f32 %f1009, %f128, %f2347, %f1008; neg.ftz.f32 %f2348, %f1009; $L__BB26_54: ld.local.f32 %f1022, [%rd6]; ld.local.f32 %f1023, [%rd6+4]; mul.ftz.f32 %f1024, %f2338, %f1023; fma.rn.ftz.f32 %f1025, %f2337, %f1022, %f1024; ld.local.f32 %f1026, [%rd6+8]; fma.rn.ftz.f32 %f1027, %f2339, %f1026, %f1025; add.ftz.f32 %f2392, %f2340, %f1027; mul.ftz.f32 %f1028, %f2342, %f1023; fma.rn.ftz.f32 %f1029, %f2341, %f1022, %f1028; fma.rn.ftz.f32 %f1030, %f2343, %f1026, %f1029; add.ftz.f32 %f2393, %f2344, %f1030; mul.ftz.f32 %f1031, %f2346, %f1023; fma.rn.ftz.f32 %f1032, %f2345, %f1022, %f1031; fma.rn.ftz.f32 %f1033, %f2347, %f1026, %f1032; add.ftz.f32 %f2394, %f2348, %f1033; ld.local.f32 %f1034, [%rd6+12]; ld.local.f32 %f1035, [%rd6+16]; mul.ftz.f32 %f1036, %f2338, %f1035; fma.rn.ftz.f32 %f1037, %f2337, %f1034, %f1036; ld.local.f32 %f1038, [%rd6+20]; fma.rn.ftz.f32 %f1039, %f2339, %f1038, %f1037; mul.ftz.f32 %f1040, %f2342, %f1035; fma.rn.ftz.f32 %f1041, %f2341, %f1034, %f1040; fma.rn.ftz.f32 %f1042, %f2343, %f1038, %f1041; mul.ftz.f32 %f1043, %f2346, %f1035; fma.rn.ftz.f32 %f1044, %f2345, %f1034, %f1043; fma.rn.ftz.f32 %f1045, %f2347, %f1038, %f1044; abs.ftz.f32 %f1046, %f1039; setp.gt.ftz.f32 %p99, %f1046, 0f17800000; mov.b32 %r313, %f1039; and.b32 %r314, %r313, -2147483648; or.b32 %r315, %r314, 394264576; mov.b32 %f1047, %r315; selp.f32 %f2389, %f1039, %f1047, %p99; abs.ftz.f32 %f1048, %f1042; setp.gt.ftz.f32 %p100, %f1048, 0f17800000; mov.b32 %r316, %f1042; and.b32 %r317, %r316, -2147483648; or.b32 %r318, %r317, 394264576; mov.b32 %f1049, %r318; selp.f32 %f2390, %f1042, %f1049, %p100; abs.ftz.f32 %f1050, %f1045; setp.gt.ftz.f32 %p101, %f1050, 0f17800000; mov.b32 %r319, %f1045; and.b32 %r320, %r319, -2147483648; or.b32 %r321, %r320, 394264576; mov.b32 %f1051, %r321; selp.f32 %f2391, %f1045, %f1051, %p101; setp.eq.s32 %p286, %r16, -1; $L__BB26_55: rcp.approx.ftz.f32 %f207, %f2390; rcp.approx.ftz.f32 %f208, %f2389; rcp.approx.ftz.f32 %f209, %f2391; ld.const.v2.u64 {%rd168, %rd169}, [kernel_params+2096]; cvta.to.global.u64 %rd14, %rd168; cvta.to.global.u64 %rd15, %rd169; ld.const.v2.u64 {%rd172, %rd173}, [kernel_params+2128]; cvta.to.global.u64 %rd16, %rd173; cvta.to.global.u64 %rd17, %rd172; ld.const.u64 %rd176, [kernel_params+2232]; cvta.to.global.u64 %rd18, %rd176; ld.const.u64 %rd177, [kernel_params+2256]; cvta.to.global.u64 %rd19, %rd177; cvta.to.global.u64 %rd178, %rd329; mul.lo.s64 %rd179, %rd5, 208; add.s64 %rd180, %rd178, %rd179; add.s64 %rd20, %rd180, 152; ld.const.u64 %rd181, [kernel_params+2312]; cvta.to.global.u64 %rd21, %rd181; ld.const.u64 %rd182, [kernel_params+2336]; cvta.to.global.u64 %rd22, %rd182; not.pred %p158, %p286; mov.pred %p161, -1; bra.uni $L__BB26_56; $L__BB26_68: mul.wide.s32 %rd186, %r565, 16; xor.b64 %rd187, %rd186, -16; add.s64 %rd188, %rd15, %rd187; ld.global.v4.u32 {%r568, %r334, %r335, %r336}, [%rd188]; mul.wide.s32 %rd189, %r45, 4; add.s64 %rd190, %rd2, %rd189; ld.local.u32 %r565, [%rd190]; add.s32 %r586, %r45, -1; cvt.u16.u32 %rs25, %r336; and.b16 %rs24, %rs25, 127; setp.eq.s16 %p125, %rs24, 1; @%p125 bra $L__BB26_100; setp.ne.s16 %p126, %rs24, 65; @%p126 bra $L__BB26_118; setp.ge.s32 %p127, %r568, %r334; @%p127 bra $L__BB26_118; $L__BB26_71: cvt.s64.s32 %rd25, %r568; @%p158 bra $L__BB26_73; shl.b64 %rd191, %rd25, 2; add.s64 %rd192, %rd16, %rd191; ld.global.u32 %r338, [%rd192]; setp.ne.s32 %p129, %r338, %r16; @%p129 bra $L__BB26_99; $L__BB26_73: shl.b64 %rd193, %rd25, 2; add.s64 %rd194, %rd17, %rd193; ld.global.u32 %r53, [%rd194]; ld.local.u32 %r339, [%rd6+36]; setp.eq.s32 %p130, %r339, %r53; @%p130 bra $L__BB26_99; ld.global.u32 %r54, [%rd20+4]; ld.global.u32 %r55, [%rd20]; shl.b32 %r340, %r55, 1; cvt.rn.f32.s32 %f1268, %r340; ld.local.f32 %f1269, [%rd6+32]; mul.ftz.f32 %f226, %f1269, %f1268; cvt.rzi.ftz.s32.f32 %r341, %f226; add.s32 %r342, %r340, -1; min.s32 %r56, %r341, %r342; ld.global.u32 %r571, [%rd20+12]; mul.wide.u32 %rd195, %r571, 16; add.s64 %rd196, %rd21, %rd195; ld.global.u16 %rs47, [%rd196+12]; ld.global.u32 %r570, [%rd196+8]; ld.global.u64 %rd330, [%rd196]; setp.eq.s64 %p131, %rd330, 11; @%p131 bra $L__BB26_79; $L__BB26_75: setp.eq.s64 %p132, %rd330, 0; @%p132 bra $L__BB26_77; add.s32 %r571, %r571, 2; bra.uni $L__BB26_78; $L__BB26_77: setp.eq.s16 %p133, %rs47, 0; mov.u32 %r571, %r570; mov.u32 %r574, %r287; @%p133 bra $L__BB26_80; $L__BB26_78: mul.wide.u32 %rd197, %r571, 16; add.s64 %rd198, %rd21, %rd197; ld.global.u16 %rs47, [%rd198+12]; ld.global.u32 %r570, [%rd198+8]; ld.global.u64 %rd330, [%rd198]; setp.ne.s64 %p134, %rd330, 11; @%p134 bra $L__BB26_75; $L__BB26_79: setp.eq.s16 %p135, %rs47, 0; selp.b32 %r574, -1, %r570, %p135; $L__BB26_80: mul.wide.s32 %rd199, %r53, 16; add.s64 %rd200, %rd18, %rd199; ld.global.v4.u32 {%r344, %r345, %r346, %r347}, [%rd200]; setp.eq.s32 %p136, %r56, %r55; @%p136 bra $L__BB26_82; bra.uni $L__BB26_81; $L__BB26_82: mul.wide.u32 %rd204, %r347, 12; add.s64 %rd331, %rd19, %rd204; add.s32 %r354, %r347, 1; mul.wide.u32 %rd205, %r354, 12; add.s64 %rd334, %rd19, %rd205; add.s32 %r355, %r347, 2; mul.wide.u32 %rd206, %r355, 12; add.s64 %rd337, %rd19, %rd206; add.s64 %rd332, %rd331, 4; add.s64 %rd333, %rd331, 8; add.s64 %rd335, %rd334, 4; add.s64 %rd336, %rd334, 8; add.s64 %rd338, %rd337, 4; add.s64 %rd339, %rd337, 8; bra.uni $L__BB26_83; $L__BB26_81: setp.gt.s32 %p137, %r56, %r55; selp.b32 %r348, -1, 0, %p137; add.s32 %r349, %r56, %r348; mad.lo.s32 %r350, %r349, %r54, %r574; add.s32 %r351, %r344, %r350; mul.wide.u32 %rd201, %r351, 12; add.s64 %rd331, %rd22, %rd201; add.s32 %r352, %r345, %r350; mul.wide.u32 %rd202, %r352, 12; add.s64 %rd334, %rd22, %rd202; add.s32 %r353, %r346, %r350; mul.wide.u32 %rd203, %r353, 12; add.s64 %rd337, %rd22, %rd203; add.s64 %rd332, %rd331, 4; add.s64 %rd333, %rd331, 8; add.s64 %rd335, %rd334, 4; add.s64 %rd336, %rd334, 8; add.s64 %rd338, %rd337, 4; add.s64 %rd339, %rd337, 8; $L__BB26_83: ld.global.f32 %f227, [%rd339]; ld.global.f32 %f228, [%rd338]; ld.global.f32 %f229, [%rd337]; ld.global.f32 %f230, [%rd336]; ld.global.f32 %f231, [%rd335]; ld.global.f32 %f232, [%rd334]; ld.global.f32 %f233, [%rd333]; ld.global.f32 %f234, [%rd332]; ld.global.f32 %f235, [%rd331]; add.s32 %r71, %r56, 1; setp.eq.s32 %p138, %r71, %r55; @%p138 bra $L__BB26_85; bra.uni $L__BB26_84; $L__BB26_85: mul.wide.u32 %rd210, %r347, 12; add.s64 %rd340, %rd19, %rd210; add.s32 %r362, %r347, 1; mul.wide.u32 %rd211, %r362, 12; add.s64 %rd343, %rd19, %rd211; add.s32 %r363, %r347, 2; mul.wide.u32 %rd212, %r363, 12; add.s64 %rd346, %rd19, %rd212; add.s64 %rd341, %rd340, 4; add.s64 %rd342, %rd340, 8; add.s64 %rd344, %rd343, 4; add.s64 %rd345, %rd343, 8; add.s64 %rd347, %rd346, 4; add.s64 %rd348, %rd346, 8; bra.uni $L__BB26_86; $L__BB26_84: setp.ge.s32 %p139, %r56, %r55; selp.b32 %r356, -1, 0, %p139; add.s32 %r357, %r71, %r356; mad.lo.s32 %r358, %r357, %r54, %r574; add.s32 %r359, %r344, %r358; mul.wide.u32 %rd207, %r359, 12; add.s64 %rd340, %rd22, %rd207; add.s32 %r360, %r345, %r358; mul.wide.u32 %rd208, %r360, 12; add.s64 %rd343, %rd22, %rd208; add.s32 %r361, %r346, %r358; mul.wide.u32 %rd209, %r361, 12; add.s64 %rd346, %rd22, %rd209; add.s64 %rd341, %rd340, 4; add.s64 %rd342, %rd340, 8; add.s64 %rd344, %rd343, 4; add.s64 %rd345, %rd343, 8; add.s64 %rd347, %rd346, 4; add.s64 %rd348, %rd346, 8; $L__BB26_86: cvt.rn.f32.s32 %f1270, %r56; sub.ftz.f32 %f1271, %f226, %f1270; sub.ftz.f32 %f1273, %f612, %f1271; ld.global.f32 %f1274, [%rd340]; mul.ftz.f32 %f1275, %f1271, %f1274; ld.global.f32 %f1276, [%rd341]; mul.ftz.f32 %f1277, %f1271, %f1276; ld.global.f32 %f1278, [%rd342]; mul.ftz.f32 %f1279, %f1271, %f1278; fma.rn.ftz.f32 %f236, %f1273, %f235, %f1275; fma.rn.ftz.f32 %f237, %f1273, %f234, %f1277; fma.rn.ftz.f32 %f238, %f1273, %f233, %f1279; ld.global.f32 %f1280, [%rd343]; mul.ftz.f32 %f1281, %f1271, %f1280; ld.global.f32 %f1282, [%rd344]; mul.ftz.f32 %f1283, %f1271, %f1282; ld.global.f32 %f1284, [%rd345]; mul.ftz.f32 %f1285, %f1271, %f1284; fma.rn.ftz.f32 %f239, %f1273, %f232, %f1281; fma.rn.ftz.f32 %f240, %f1273, %f231, %f1283; fma.rn.ftz.f32 %f241, %f1273, %f230, %f1285; ld.global.f32 %f1286, [%rd346]; mul.ftz.f32 %f1287, %f1271, %f1286; ld.global.f32 %f1288, [%rd347]; mul.ftz.f32 %f1289, %f1271, %f1288; ld.global.f32 %f1290, [%rd348]; mul.ftz.f32 %f1291, %f1271, %f1290; fma.rn.ftz.f32 %f242, %f1273, %f229, %f1287; fma.rn.ftz.f32 %f243, %f1273, %f228, %f1289; fma.rn.ftz.f32 %f244, %f1273, %f227, %f1291; sub.ftz.f32 %f245, %f236, %f2392; sub.ftz.f32 %f246, %f237, %f2393; sub.ftz.f32 %f247, %f238, %f2394; sub.ftz.f32 %f1292, %f239, %f2392; sub.ftz.f32 %f1293, %f240, %f2393; sub.ftz.f32 %f1294, %f241, %f2394; sub.ftz.f32 %f1295, %f242, %f2392; sub.ftz.f32 %f1296, %f243, %f2393; sub.ftz.f32 %f1297, %f244, %f2394; sub.ftz.f32 %f248, %f1295, %f245; sub.ftz.f32 %f249, %f1296, %f246; sub.ftz.f32 %f250, %f1297, %f247; sub.ftz.f32 %f251, %f245, %f1292; sub.ftz.f32 %f252, %f246, %f1293; sub.ftz.f32 %f253, %f247, %f1294; sub.ftz.f32 %f1298, %f1292, %f1295; sub.ftz.f32 %f1299, %f1293, %f1296; sub.ftz.f32 %f1300, %f1294, %f1297; add.ftz.f32 %f1301, %f1295, %f245; add.ftz.f32 %f1302, %f1296, %f246; add.ftz.f32 %f1303, %f1297, %f247; mul.ftz.f32 %f1304, %f1303, %f249; mul.ftz.f32 %f1305, %f250, %f1302; sub.ftz.f32 %f1306, %f1304, %f1305; mul.ftz.f32 %f1307, %f250, %f1301; mul.ftz.f32 %f1308, %f1303, %f248; sub.ftz.f32 %f1309, %f1307, %f1308; mul.ftz.f32 %f1310, %f1302, %f248; mul.ftz.f32 %f1311, %f249, %f1301; sub.ftz.f32 %f1312, %f1310, %f1311; mul.ftz.f32 %f1313, %f2390, %f1309; fma.rn.ftz.f32 %f1314, %f2389, %f1306, %f1313; fma.rn.ftz.f32 %f254, %f2391, %f1312, %f1314; add.ftz.f32 %f1315, %f1292, %f245; add.ftz.f32 %f1316, %f1293, %f246; add.ftz.f32 %f1317, %f1294, %f247; mul.ftz.f32 %f1318, %f1317, %f252; mul.ftz.f32 %f1319, %f253, %f1316; sub.ftz.f32 %f1320, %f1318, %f1319; mul.ftz.f32 %f1321, %f253, %f1315; mul.ftz.f32 %f1322, %f1317, %f251; sub.ftz.f32 %f1323, %f1321, %f1322; mul.ftz.f32 %f1324, %f1316, %f251; mul.ftz.f32 %f1325, %f252, %f1315; sub.ftz.f32 %f1326, %f1324, %f1325; mul.ftz.f32 %f1327, %f2390, %f1323; fma.rn.ftz.f32 %f1328, %f2389, %f1320, %f1327; fma.rn.ftz.f32 %f255, %f2391, %f1326, %f1328; add.ftz.f32 %f1329, %f1295, %f1292; add.ftz.f32 %f1330, %f1296, %f1293; add.ftz.f32 %f1331, %f1297, %f1294; mul.ftz.f32 %f1332, %f1331, %f1299; mul.ftz.f32 %f1333, %f1300, %f1330; sub.ftz.f32 %f1334, %f1332, %f1333; mul.ftz.f32 %f1335, %f1300, %f1329; mul.ftz.f32 %f1336, %f1331, %f1298; sub.ftz.f32 %f1337, %f1335, %f1336; mul.ftz.f32 %f1338, %f1330, %f1298; mul.ftz.f32 %f1339, %f1299, %f1329; sub.ftz.f32 %f1340, %f1338, %f1339; mul.ftz.f32 %f1341, %f2390, %f1337; fma.rn.ftz.f32 %f1342, %f2389, %f1334, %f1341; fma.rn.ftz.f32 %f1343, %f2391, %f1340, %f1342; add.ftz.f32 %f1344, %f254, %f255; add.ftz.f32 %f256, %f1343, %f1344; abs.ftz.f32 %f257, %f256; mul.ftz.f32 %f1345, %f257, 0f34000000; min.ftz.f32 %f1346, %f255, %f1343; min.ftz.f32 %f1347, %f254, %f1346; max.ftz.f32 %f1348, %f255, %f1343; max.ftz.f32 %f1349, %f254, %f1348; mul.ftz.f32 %f1350, %f257, 0fB4000000; setp.ltu.ftz.f32 %p141, %f1347, %f1350; setp.gtu.ftz.f32 %p142, %f1349, %f1345; and.pred %p143, %p141, %p142; mov.pred %p287, -1; @%p143 bra $L__BB26_92; mov.pred %p287, -1; mul.ftz.f32 %f1351, %f250, %f252; mul.ftz.f32 %f1352, %f253, %f249; sub.ftz.f32 %f1353, %f1351, %f1352; mul.ftz.f32 %f1354, %f250, %f251; mul.ftz.f32 %f1355, %f253, %f248; sub.ftz.f32 %f1356, %f1355, %f1354; mul.ftz.f32 %f1357, %f252, %f248; mul.ftz.f32 %f1358, %f249, %f251; sub.ftz.f32 %f1359, %f1358, %f1357; add.ftz.f32 %f258, %f1353, %f1353; add.ftz.f32 %f259, %f1356, %f1356; add.ftz.f32 %f260, %f1359, %f1359; mul.ftz.f32 %f1360, %f2390, %f259; fma.rn.ftz.f32 %f1361, %f2389, %f258, %f1360; fma.rn.ftz.f32 %f261, %f2391, %f260, %f1361; setp.eq.ftz.f32 %p145, %f261, 0f00000000; @%p145 bra $L__BB26_92; mov.pred %p287, -1; mul.ftz.f32 %f1362, %f246, %f259; fma.rn.ftz.f32 %f1363, %f245, %f258, %f1362; fma.rn.ftz.f32 %f1364, %f247, %f260, %f1363; div.approx.ftz.f32 %f262, %f1364, %f261; setp.ltu.ftz.f32 %p147, %f262, 0f00000000; setp.gtu.ftz.f32 %p148, %f262, %f99; or.pred %p149, %p148, %p147; @%p149 bra $L__BB26_92; abs.ftz.f32 %f2297, %f256; setp.lt.ftz.f32 %p150, %f2297, 0f219392EF; mov.f32 %f2366, %f749; @%p150 bra $L__BB26_91; rcp.approx.ftz.f32 %f2366, %f256; $L__BB26_91: mul.ftz.f32 %f1366, %f254, %f2366; min.ftz.f32 %f2387, %f1366, %f612; mul.ftz.f32 %f1368, %f255, %f2366; min.ftz.f32 %f2386, %f1368, %f612; mov.pred %p287, 0; mov.f32 %f2388, %f262; $L__BB26_92: @%p287 bra $L__BB26_99; ld.local.u32 %r577, [%rd1]; min.s32 %r575, %r577, 4; $L__BB26_94: add.s32 %r75, %r575, -1; setp.gt.s32 %p152, %r575, 0; @%p152 bra $L__BB26_98; bra.uni $L__BB26_95; $L__BB26_98: mul.wide.s32 %rd217, %r75, 24; add.s64 %rd218, %rd1, %rd217; ld.local.f32 %f1391, [%rd218+4]; setp.neu.ftz.f32 %p155, %f1391, %f2388; mov.u32 %r575, %r75; @%p155 bra $L__BB26_94; bra.uni $L__BB26_99; $L__BB26_95: setp.lt.s32 %p153, %r577, 4; add.s32 %r76, %r577, 1; st.local.u32 [%rd1], %r76; @%p153 bra $L__BB26_97; mad.lo.s32 %r585, %r585, 1103515245, 12345; rem.u32 %r577, %r585, %r76; setp.gt.s32 %p154, %r577, 3; @%p154 bra $L__BB26_99; $L__BB26_97: mul.wide.s32 %rd213, %r577, 24; add.s64 %rd214, %rd1, %rd213; st.local.f32 [%rd214+4], %f2388; st.local.f32 [%rd214+8], %f2387; st.local.f32 [%rd214+12], %f2386; st.local.u32 [%rd214+16], %r53; st.local.u32 [%rd214+20], %r16; mov.u32 %r364, 65; st.local.u32 [%rd214+24], %r364; sub.ftz.f32 %f1369, %f244, %f238; sub.ftz.f32 %f1370, %f240, %f237; mul.ftz.f32 %f1371, %f1369, %f1370; sub.ftz.f32 %f1372, %f243, %f237; sub.ftz.f32 %f1373, %f241, %f238; mul.ftz.f32 %f1374, %f1373, %f1372; sub.ftz.f32 %f1375, %f1371, %f1374; sub.ftz.f32 %f1376, %f242, %f236; mul.ftz.f32 %f1377, %f1373, %f1376; sub.ftz.f32 %f1378, %f239, %f236; mul.ftz.f32 %f1379, %f1369, %f1378; sub.ftz.f32 %f1380, %f1377, %f1379; mul.ftz.f32 %f1381, %f1372, %f1378; mul.ftz.f32 %f1382, %f1370, %f1376; sub.ftz.f32 %f1383, %f1381, %f1382; mul.ftz.f32 %f1384, %f1380, %f1380; fma.rn.ftz.f32 %f1385, %f1375, %f1375, %f1384; fma.rn.ftz.f32 %f1386, %f1383, %f1383, %f1385; sqrt.approx.ftz.f32 %f1387, %f1386; div.approx.ftz.f32 %f1388, %f1375, %f1387; div.approx.ftz.f32 %f1389, %f1380, %f1387; div.approx.ftz.f32 %f1390, %f1383, %f1387; mul.wide.s32 %rd215, %r577, 12; add.s64 %rd216, %rd1, %rd215; st.local.f32 [%rd216+100], %f1388; st.local.f32 [%rd216+104], %f1389; st.local.f32 [%rd216+108], %f1390; $L__BB26_99: add.s32 %r586, %r45, -1; cvt.u32.u64 %r365, %rd25; add.s32 %r568, %r365, 1; setp.lt.s32 %p156, %r568, %r334; @%p156 bra $L__BB26_71; bra.uni $L__BB26_118; $L__BB26_100: setp.ge.s32 %p157, %r568, %r334; @%p157 bra $L__BB26_118; $L__BB26_101: cvt.s64.s32 %rd84, %r568; @%p158 bra $L__BB26_103; shl.b64 %rd219, %rd84, 2; add.s64 %rd220, %rd16, %rd219; ld.global.u32 %r366, [%rd220]; setp.ne.s32 %p159, %r366, %r16; @%p159 bra $L__BB26_117; $L__BB26_103: shl.b64 %rd221, %rd84, 2; add.s64 %rd222, %rd17, %rd221; ld.global.u32 %r85, [%rd222]; ld.local.u32 %r367, [%rd6+36]; setp.eq.s32 %p160, %r367, %r85; @%p160 bra $L__BB26_117; mul.wide.s32 %rd223, %r85, 16; add.s64 %rd224, %rd18, %rd223; ld.global.u32 %r368, [%rd224+12]; mul.wide.u32 %rd225, %r368, 12; add.s64 %rd226, %rd19, %rd225; add.s32 %r369, %r368, 1; mul.wide.u32 %rd227, %r369, 12; add.s64 %rd228, %rd19, %rd227; add.s32 %r370, %r368, 2; mul.wide.u32 %rd229, %r370, 12; add.s64 %rd230, %rd19, %rd229; ld.global.f32 %f276, [%rd226]; sub.ftz.f32 %f277, %f276, %f2392; ld.global.f32 %f278, [%rd226+4]; sub.ftz.f32 %f279, %f278, %f2393; ld.global.f32 %f280, [%rd226+8]; sub.ftz.f32 %f281, %f280, %f2394; ld.global.f32 %f282, [%rd228]; sub.ftz.f32 %f1392, %f282, %f2392; ld.global.f32 %f283, [%rd228+4]; sub.ftz.f32 %f1393, %f283, %f2393; ld.global.f32 %f284, [%rd228+8]; sub.ftz.f32 %f1394, %f284, %f2394; ld.global.f32 %f285, [%rd230]; sub.ftz.f32 %f1395, %f285, %f2392; ld.global.f32 %f286, [%rd230+4]; sub.ftz.f32 %f1396, %f286, %f2393; ld.global.f32 %f287, [%rd230+8]; sub.ftz.f32 %f1397, %f287, %f2394; sub.ftz.f32 %f288, %f1395, %f277; sub.ftz.f32 %f289, %f1396, %f279; sub.ftz.f32 %f290, %f1397, %f281; sub.ftz.f32 %f291, %f277, %f1392; sub.ftz.f32 %f292, %f279, %f1393; sub.ftz.f32 %f293, %f281, %f1394; sub.ftz.f32 %f1398, %f1392, %f1395; sub.ftz.f32 %f1399, %f1393, %f1396; sub.ftz.f32 %f1400, %f1394, %f1397; add.ftz.f32 %f1401, %f277, %f1395; add.ftz.f32 %f1402, %f279, %f1396; add.ftz.f32 %f1403, %f281, %f1397; mul.ftz.f32 %f1404, %f289, %f1403; mul.ftz.f32 %f1405, %f1402, %f290; sub.ftz.f32 %f1406, %f1404, %f1405; mul.ftz.f32 %f1407, %f1401, %f290; mul.ftz.f32 %f1408, %f288, %f1403; sub.ftz.f32 %f1409, %f1407, %f1408; mul.ftz.f32 %f1410, %f288, %f1402; mul.ftz.f32 %f1411, %f1401, %f289; sub.ftz.f32 %f1412, %f1410, %f1411; mul.ftz.f32 %f1413, %f2390, %f1409; fma.rn.ftz.f32 %f1414, %f2389, %f1406, %f1413; fma.rn.ftz.f32 %f294, %f2391, %f1412, %f1414; add.ftz.f32 %f1415, %f277, %f1392; add.ftz.f32 %f1416, %f279, %f1393; add.ftz.f32 %f1417, %f281, %f1394; mul.ftz.f32 %f1418, %f292, %f1417; mul.ftz.f32 %f1419, %f1416, %f293; sub.ftz.f32 %f1420, %f1418, %f1419; mul.ftz.f32 %f1421, %f1415, %f293; mul.ftz.f32 %f1422, %f291, %f1417; sub.ftz.f32 %f1423, %f1421, %f1422; mul.ftz.f32 %f1424, %f291, %f1416; mul.ftz.f32 %f1425, %f1415, %f292; sub.ftz.f32 %f1426, %f1424, %f1425; mul.ftz.f32 %f1427, %f2390, %f1423; fma.rn.ftz.f32 %f1428, %f2389, %f1420, %f1427; fma.rn.ftz.f32 %f295, %f2391, %f1426, %f1428; add.ftz.f32 %f1429, %f1392, %f1395; add.ftz.f32 %f1430, %f1393, %f1396; add.ftz.f32 %f1431, %f1394, %f1397; mul.ftz.f32 %f1432, %f1399, %f1431; mul.ftz.f32 %f1433, %f1430, %f1400; sub.ftz.f32 %f1434, %f1432, %f1433; mul.ftz.f32 %f1435, %f1429, %f1400; mul.ftz.f32 %f1436, %f1398, %f1431; sub.ftz.f32 %f1437, %f1435, %f1436; mul.ftz.f32 %f1438, %f1398, %f1430; mul.ftz.f32 %f1439, %f1429, %f1399; sub.ftz.f32 %f1440, %f1438, %f1439; mul.ftz.f32 %f1441, %f2390, %f1437; fma.rn.ftz.f32 %f1442, %f2389, %f1434, %f1441; fma.rn.ftz.f32 %f1443, %f2391, %f1440, %f1442; add.ftz.f32 %f1444, %f295, %f294; add.ftz.f32 %f296, %f1443, %f1444; abs.ftz.f32 %f297, %f296; mul.ftz.f32 %f1445, %f297, 0f34000000; min.ftz.f32 %f1446, %f295, %f1443; min.ftz.f32 %f1447, %f294, %f1446; max.ftz.f32 %f1448, %f295, %f1443; max.ftz.f32 %f1449, %f294, %f1448; mul.ftz.f32 %f1450, %f297, 0fB4000000; setp.ltu.ftz.f32 %p162, %f1447, %f1450; setp.gtu.ftz.f32 %p163, %f1449, %f1445; and.pred %p164, %p162, %p163; mov.pred %p288, %p161; @%p164 bra $L__BB26_110; mul.ftz.f32 %f1451, %f292, %f290; mul.ftz.f32 %f1452, %f293, %f289; sub.ftz.f32 %f1453, %f1451, %f1452; mul.ftz.f32 %f1454, %f291, %f290; mul.ftz.f32 %f1455, %f293, %f288; sub.ftz.f32 %f1456, %f1455, %f1454; mul.ftz.f32 %f1457, %f292, %f288; mul.ftz.f32 %f1458, %f291, %f289; sub.ftz.f32 %f1459, %f1458, %f1457; add.ftz.f32 %f298, %f1453, %f1453; add.ftz.f32 %f299, %f1456, %f1456; add.ftz.f32 %f300, %f1459, %f1459; mul.ftz.f32 %f1460, %f2390, %f299; fma.rn.ftz.f32 %f1461, %f2389, %f298, %f1460; fma.rn.ftz.f32 %f301, %f2391, %f300, %f1461; setp.eq.ftz.f32 %p166, %f301, 0f00000000; mov.pred %p288, %p161; @%p166 bra $L__BB26_110; sub.ftz.f32 %f2296, %f280, %f2394; sub.ftz.f32 %f2295, %f276, %f2392; sub.ftz.f32 %f2294, %f278, %f2393; mul.ftz.f32 %f1462, %f2294, %f299; fma.rn.ftz.f32 %f1463, %f2295, %f298, %f1462; fma.rn.ftz.f32 %f1464, %f2296, %f300, %f1463; div.approx.ftz.f32 %f302, %f1464, %f301; setp.ltu.ftz.f32 %p168, %f302, 0f00000000; setp.gtu.ftz.f32 %p169, %f302, %f99; or.pred %p170, %p169, %p168; mov.pred %p288, %p161; @%p170 bra $L__BB26_110; setp.lt.ftz.f32 %p171, %f297, 0f219392EF; mov.f32 %f2376, 0f00000000; @%p171 bra $L__BB26_109; rcp.approx.ftz.f32 %f2376, %f296; $L__BB26_109: mul.ftz.f32 %f1466, %f294, %f2376; min.ftz.f32 %f2381, %f1466, %f612; mul.ftz.f32 %f1468, %f295, %f2376; min.ftz.f32 %f2380, %f1468, %f612; mov.pred %p288, 0; mov.f32 %f2382, %f302; $L__BB26_110: @%p288 bra $L__BB26_117; ld.local.u32 %r583, [%rd1]; min.s32 %r581, %r583, 4; $L__BB26_112: add.s32 %r89, %r581, -1; setp.gt.s32 %p173, %r581, 0; @%p173 bra $L__BB26_116; bra.uni $L__BB26_113; $L__BB26_116: mul.wide.s32 %rd235, %r89, 24; add.s64 %rd236, %rd1, %rd235; ld.local.f32 %f1491, [%rd236+4]; setp.neu.ftz.f32 %p176, %f1491, %f2382; mov.u32 %r581, %r89; @%p176 bra $L__BB26_112; bra.uni $L__BB26_117; $L__BB26_113: setp.lt.s32 %p174, %r583, 4; add.s32 %r90, %r583, 1; st.local.u32 [%rd1], %r90; @%p174 bra $L__BB26_115; mad.lo.s32 %r585, %r585, 1103515245, 12345; rem.u32 %r583, %r585, %r90; setp.gt.s32 %p175, %r583, 3; @%p175 bra $L__BB26_117; $L__BB26_115: mul.wide.s32 %rd231, %r583, 24; add.s64 %rd232, %rd1, %rd231; st.local.u32 [%rd232+16], %r85; st.local.u32 [%rd232+20], %r16; mov.u32 %r371, 1; st.local.u32 [%rd232+24], %r371; st.local.f32 [%rd232+8], %f2381; st.local.f32 [%rd232+12], %f2380; st.local.f32 [%rd232+4], %f2382; sub.ftz.f32 %f1469, %f287, %f280; sub.ftz.f32 %f1470, %f283, %f278; mul.ftz.f32 %f1471, %f1470, %f1469; sub.ftz.f32 %f1472, %f286, %f278; sub.ftz.f32 %f1473, %f284, %f280; mul.ftz.f32 %f1474, %f1473, %f1472; sub.ftz.f32 %f1475, %f1471, %f1474; sub.ftz.f32 %f1476, %f285, %f276; mul.ftz.f32 %f1477, %f1473, %f1476; sub.ftz.f32 %f1478, %f282, %f276; mul.ftz.f32 %f1479, %f1478, %f1469; sub.ftz.f32 %f1480, %f1477, %f1479; mul.ftz.f32 %f1481, %f1478, %f1472; mul.ftz.f32 %f1482, %f1470, %f1476; sub.ftz.f32 %f1483, %f1481, %f1482; mul.ftz.f32 %f1484, %f1480, %f1480; fma.rn.ftz.f32 %f1485, %f1475, %f1475, %f1484; fma.rn.ftz.f32 %f1486, %f1483, %f1483, %f1485; sqrt.approx.ftz.f32 %f1487, %f1486; div.approx.ftz.f32 %f1488, %f1475, %f1487; div.approx.ftz.f32 %f1489, %f1480, %f1487; div.approx.ftz.f32 %f1490, %f1483, %f1487; mul.wide.s32 %rd233, %r583, 12; add.s64 %rd234, %rd1, %rd233; st.local.f32 [%rd234+100], %f1488; st.local.f32 [%rd234+104], %f1489; st.local.f32 [%rd234+108], %f1490; $L__BB26_117: add.s32 %r586, %r45, -1; cvt.u32.u64 %r372, %rd84; add.s32 %r568, %r372, 1; setp.lt.s32 %p177, %r568, %r334; @%p177 bra $L__BB26_101; bra.uni $L__BB26_118; $L__BB26_56: setp.eq.s32 %p102, %r565, 1985229328; setp.lt.s32 %p103, %r565, 0; or.pred %p104, %p103, %p102; @%p104 bra $L__BB26_67; $L__BB26_57: mul.wide.s32 %rd183, %r565, 16; add.s64 %rd23, %rd14, %rd183; ld.global.v4.f32 {%f1053, %f1054, %f1055, %f1056}, [%rd23]; mov.b32 %r565, %f1056; mov.b32 %r30, %f1055; mov.b32 %r29, %f1054; mov.b32 %r28, %f1053; mov.b64 %rd184, {%r28, %r29}; cvt.u32.u64 %r323, %rd184; and.b32 %r324, %r323, 1024; setp.eq.s32 %p105, %r324, 0; and.b32 %r32, %r28, 2047; @%p105 bra $L__BB26_60; bra.uni $L__BB26_58; $L__BB26_60: setp.ne.s32 %p111, %r32, 0; ld.global.v4.f32 {%f1197, %f1198, %f1199, %f1200}, [%rd23+16]; ld.global.v4.f32 {%f1205, %f1206, %f1207, %f1208}, [%rd23+32]; ld.global.v4.f32 {%f1213, %f1214, %f1215, %f1216}, [%rd23+48]; sub.ftz.f32 %f1221, %f1197, %f2392; mul.ftz.f32 %f1222, %f208, %f1221; sub.ftz.f32 %f1223, %f1199, %f2392; mul.ftz.f32 %f1224, %f208, %f1223; sub.ftz.f32 %f1225, %f1205, %f2393; mul.ftz.f32 %f1226, %f207, %f1225; sub.ftz.f32 %f1227, %f1207, %f2393; mul.ftz.f32 %f1228, %f207, %f1227; sub.ftz.f32 %f1229, %f1213, %f2394; mul.ftz.f32 %f1230, %f209, %f1229; sub.ftz.f32 %f1231, %f1215, %f2394; mul.ftz.f32 %f1232, %f209, %f1231; min.ftz.f32 %f1233, %f1222, %f1224; min.ftz.f32 %f1234, %f1226, %f1228; min.ftz.f32 %f1235, %f1230, %f1232; mov.f32 %f1236, 0f00000000; max.ftz.f32 %f1237, %f1236, %f1233; max.ftz.f32 %f1238, %f1234, %f1235; max.ftz.f32 %f2361, %f1237, %f1238; max.ftz.f32 %f1239, %f1222, %f1224; max.ftz.f32 %f1240, %f1226, %f1228; max.ftz.f32 %f1241, %f1230, %f1232; min.ftz.f32 %f1242, %f99, %f1239; min.ftz.f32 %f1243, %f1240, %f1241; min.ftz.f32 %f1244, %f1242, %f1243; sub.ftz.f32 %f1245, %f1198, %f2392; mul.ftz.f32 %f1246, %f208, %f1245; sub.ftz.f32 %f1247, %f1200, %f2392; mul.ftz.f32 %f1248, %f208, %f1247; sub.ftz.f32 %f1249, %f1206, %f2393; mul.ftz.f32 %f1250, %f207, %f1249; sub.ftz.f32 %f1251, %f1208, %f2393; mul.ftz.f32 %f1252, %f207, %f1251; sub.ftz.f32 %f1253, %f1214, %f2394; mul.ftz.f32 %f1254, %f209, %f1253; sub.ftz.f32 %f1255, %f1216, %f2394; mul.ftz.f32 %f1256, %f209, %f1255; min.ftz.f32 %f1257, %f1246, %f1248; min.ftz.f32 %f1258, %f1250, %f1252; min.ftz.f32 %f1259, %f1254, %f1256; max.ftz.f32 %f1260, %f1236, %f1257; max.ftz.f32 %f1261, %f1258, %f1259; max.ftz.f32 %f2362, %f1260, %f1261; max.ftz.f32 %f1262, %f1246, %f1248; max.ftz.f32 %f1263, %f1250, %f1252; max.ftz.f32 %f1264, %f1254, %f1256; min.ftz.f32 %f1265, %f99, %f1262; min.ftz.f32 %f1266, %f1263, %f1264; min.ftz.f32 %f1267, %f1265, %f1266; setp.ge.ftz.f32 %p112, %f1244, %f2361; and.pred %p113, %p111, %p112; selp.u32 %r328, 1, 0, %p113; setp.ltu.ftz.f32 %p114, %f1267, %f2362; and.b32 %r330, %r29, 2047; setp.eq.s32 %p115, %r330, 0; or.pred %p116, %p115, %p114; selp.b32 %r331, 0, 2, %p116; or.b32 %r563, %r331, %r328; bra.uni $L__BB26_61; $L__BB26_58: setp.ne.s32 %p106, %r32, 0; ld.global.v4.f32 {%f1060, %f1061, %f1062, %f1063}, [%rd23+16]; ld.global.v4.f32 {%f1068, %f1069, %f1070, %f1071}, [%rd23+32]; ld.global.v4.f32 {%f1076, %f1077, %f1078, %f1079}, [%rd23+48]; mul.ftz.f32 %f1084, %f2390, %f1061; fma.rn.ftz.f32 %f1085, %f2389, %f1060, %f1084; fma.rn.ftz.f32 %f1086, %f2391, %f1062, %f1085; mul.ftz.f32 %f1087, %f2390, %f1069; fma.rn.ftz.f32 %f1088, %f2389, %f1068, %f1087; fma.rn.ftz.f32 %f1089, %f2391, %f1070, %f1088; mul.ftz.f32 %f1090, %f2390, %f1077; fma.rn.ftz.f32 %f1091, %f2389, %f1076, %f1090; fma.rn.ftz.f32 %f1092, %f2391, %f1078, %f1091; mul.ftz.f32 %f1093, %f2393, %f1061; fma.rn.ftz.f32 %f1094, %f2392, %f1060, %f1093; fma.rn.ftz.f32 %f1095, %f2394, %f1062, %f1094; add.ftz.f32 %f1096, %f1063, %f1095; mul.ftz.f32 %f1097, %f2393, %f1069; fma.rn.ftz.f32 %f1098, %f2392, %f1068, %f1097; fma.rn.ftz.f32 %f1099, %f2394, %f1070, %f1098; add.ftz.f32 %f1100, %f1071, %f1099; mul.ftz.f32 %f1101, %f2393, %f1077; fma.rn.ftz.f32 %f1102, %f2392, %f1076, %f1101; fma.rn.ftz.f32 %f1103, %f2394, %f1078, %f1102; add.ftz.f32 %f1104, %f1079, %f1103; rcp.approx.ftz.f32 %f1105, %f1086; rcp.approx.ftz.f32 %f1106, %f1089; rcp.approx.ftz.f32 %f1107, %f1092; mul.ftz.f32 %f1108, %f1105, %f1096; neg.ftz.f32 %f1109, %f1108; mul.ftz.f32 %f1110, %f1106, %f1100; neg.ftz.f32 %f1111, %f1110; mul.ftz.f32 %f1112, %f1107, %f1104; neg.ftz.f32 %f1113, %f1112; sub.ftz.f32 %f1114, %f1105, %f1108; sub.ftz.f32 %f1115, %f1106, %f1110; sub.ftz.f32 %f1116, %f1107, %f1112; min.ftz.f32 %f1117, %f1109, %f1114; min.ftz.f32 %f1118, %f1111, %f1115; min.ftz.f32 %f1119, %f1113, %f1116; max.ftz.f32 %f1120, %f1109, %f1114; max.ftz.f32 %f1121, %f1111, %f1115; max.ftz.f32 %f1122, %f1113, %f1116; mov.f32 %f1123, 0f00000000; max.ftz.f32 %f1124, %f1123, %f1117; max.ftz.f32 %f1125, %f1118, %f1119; max.ftz.f32 %f2361, %f1124, %f1125; min.ftz.f32 %f1126, %f99, %f1120; min.ftz.f32 %f1127, %f1121, %f1122; min.ftz.f32 %f1128, %f1126, %f1127; setp.le.ftz.f32 %p107, %f2361, %f1128; and.pred %p108, %p106, %p107; selp.u32 %r563, 1, 0, %p108; ld.global.v4.f32 {%f1129, %f1130, %f1131, %f1132}, [%rd23+64]; ld.global.v4.f32 {%f1137, %f1138, %f1139, %f1140}, [%rd23+80]; ld.global.v4.f32 {%f1145, %f1146, %f1147, %f1148}, [%rd23+96]; mul.ftz.f32 %f1153, %f2390, %f1130; fma.rn.ftz.f32 %f1154, %f2389, %f1129, %f1153; fma.rn.ftz.f32 %f1155, %f2391, %f1131, %f1154; mul.ftz.f32 %f1156, %f2390, %f1138; fma.rn.ftz.f32 %f1157, %f2389, %f1137, %f1156; fma.rn.ftz.f32 %f1158, %f2391, %f1139, %f1157; mul.ftz.f32 %f1159, %f2390, %f1146; fma.rn.ftz.f32 %f1160, %f2389, %f1145, %f1159; fma.rn.ftz.f32 %f1161, %f2391, %f1147, %f1160; mul.ftz.f32 %f1162, %f2393, %f1130; fma.rn.ftz.f32 %f1163, %f2392, %f1129, %f1162; fma.rn.ftz.f32 %f1164, %f2394, %f1131, %f1163; add.ftz.f32 %f1165, %f1132, %f1164; mul.ftz.f32 %f1166, %f2393, %f1138; fma.rn.ftz.f32 %f1167, %f2392, %f1137, %f1166; fma.rn.ftz.f32 %f1168, %f2394, %f1139, %f1167; add.ftz.f32 %f1169, %f1140, %f1168; mul.ftz.f32 %f1170, %f2393, %f1146; fma.rn.ftz.f32 %f1171, %f2392, %f1145, %f1170; fma.rn.ftz.f32 %f1172, %f2394, %f1147, %f1171; add.ftz.f32 %f1173, %f1148, %f1172; rcp.approx.ftz.f32 %f1174, %f1155; rcp.approx.ftz.f32 %f1175, %f1158; rcp.approx.ftz.f32 %f1176, %f1161; mul.ftz.f32 %f1177, %f1174, %f1165; neg.ftz.f32 %f1178, %f1177; mul.ftz.f32 %f1179, %f1175, %f1169; neg.ftz.f32 %f1180, %f1179; mul.ftz.f32 %f1181, %f1176, %f1173; neg.ftz.f32 %f1182, %f1181; sub.ftz.f32 %f1183, %f1174, %f1177; sub.ftz.f32 %f1184, %f1175, %f1179; sub.ftz.f32 %f1185, %f1176, %f1181; min.ftz.f32 %f1186, %f1178, %f1183; min.ftz.f32 %f1187, %f1180, %f1184; min.ftz.f32 %f1188, %f1182, %f1185; max.ftz.f32 %f1189, %f1178, %f1183; max.ftz.f32 %f1190, %f1180, %f1184; max.ftz.f32 %f1191, %f1182, %f1185; max.ftz.f32 %f1192, %f1123, %f1186; max.ftz.f32 %f1193, %f1187, %f1188; max.ftz.f32 %f2362, %f1192, %f1193; min.ftz.f32 %f1194, %f99, %f1189; min.ftz.f32 %f1195, %f1190, %f1191; min.ftz.f32 %f1196, %f1194, %f1195; setp.gtu.ftz.f32 %p109, %f2362, %f1196; @%p109 bra $L__BB26_61; and.b32 %r326, %r29, 2047; setp.eq.s32 %p110, %r326, 0; or.b32 %r327, %r563, 2; selp.b32 %r563, %r563, %r327, %p110; $L__BB26_61: mul.wide.s32 %rd185, %r586, 4; add.s64 %rd24, %rd2, %rd185; setp.eq.s32 %p117, %r563, 3; @%p117 bra $L__BB26_65; bra.uni $L__BB26_62; $L__BB26_65: setp.lt.ftz.f32 %p120, %f2362, %f2361; selp.b32 %r41, %r565, %r30, %p120; selp.b32 %r332, %r30, %r565, %p120; add.s32 %r586, %r586, 1; st.local.u32 [%rd24+4], %r332; mov.u32 %r565, %r41; bra.uni $L__BB26_66; $L__BB26_62: setp.eq.s32 %p118, %r563, 2; @%p118 bra $L__BB26_66; setp.ne.s32 %p119, %r563, 0; mov.u32 %r565, %r30; @%p119 bra $L__BB26_66; ld.local.u32 %r565, [%rd24]; add.s32 %r586, %r586, -1; $L__BB26_66: setp.ne.s32 %p121, %r565, 1985229328; setp.gt.s32 %p122, %r565, -1; and.pred %p123, %p122, %p121; @%p123 bra $L__BB26_57; $L__BB26_67: mov.u32 %r45, %r586; setp.gt.s32 %p124, %r565, -1; @%p124 bra $L__BB26_118; bra.uni $L__BB26_68; $L__BB26_118: setp.ne.s32 %p178, %r565, 1985229328; @%p178 bra $L__BB26_56; $L__BB26_157: ld.local.u32 %r146, [%rd1]; min.s32 %r147, %r146, 4; setp.eq.s32 %p233, %r147, 0; mov.u16 %rs55, 0; @%p233 bra $L__BB26_214; not.b32 %r411, %r146; max.s32 %r412, %r411, -5; mov.u32 %r413, -2; sub.s32 %r148, %r413, %r412; mov.u32 %r414, -3; sub.s32 %r149, %r414, %r412; mov.u32 %r607, 0; mov.u16 %rs29, 0; mov.u16 %rs35, 1; $L__BB26_159: sub.s32 %r151, %r148, %r607; setp.eq.s32 %p234, %r151, 0; mov.u16 %rs50, %rs29; @%p234 bra $L__BB26_170; and.b32 %r152, %r151, 1; setp.eq.s32 %p235, %r149, %r607; mov.u32 %r610, 0; mov.u16 %rs50, %rs29; @%p235 bra $L__BB26_167; sub.s32 %r609, %r151, %r152; mov.u32 %r610, 0; mov.u16 %rs50, %rs29; $L__BB26_162: cvt.s64.s32 %rd98, %r610; mul.wide.s32 %rd284, %r610, 24; add.s64 %rd285, %rd1, %rd284; add.s64 %rd99, %rd285, 4; ld.local.f32 %f2413, [%rd285+28]; ld.local.f32 %f388, [%rd285+4]; setp.leu.ftz.f32 %p236, %f388, %f2413; mul.wide.s32 %rd286, %r610, 12; add.s64 %rd287, %rd1, %rd286; add.s64 %rd100, %rd287, 100; @%p236 bra $L__BB26_164; ld.local.f32 %f1862, [%rd99+4]; ld.local.f32 %f1863, [%rd99+8]; ld.local.u32 %r417, [%rd99+12]; ld.local.u32 %r418, [%rd99+16]; ld.local.u32 %r419, [%rd99+20]; ld.local.f32 %f1864, [%rd100]; ld.local.f32 %f1865, [%rd100+4]; ld.local.f32 %f1866, [%rd100+8]; ld.local.f32 %f1867, [%rd99+28]; ld.local.f32 %f1868, [%rd99+32]; ld.local.u32 %r420, [%rd99+36]; ld.local.u32 %r421, [%rd99+40]; ld.local.u32 %r422, [%rd99+44]; st.local.f32 [%rd99], %f2413; st.local.f32 [%rd99+4], %f1867; st.local.f32 [%rd99+8], %f1868; st.local.u32 [%rd99+12], %r420; st.local.u32 [%rd99+16], %r421; st.local.u32 [%rd99+20], %r422; ld.local.f32 %f1869, [%rd100+12]; ld.local.f32 %f1870, [%rd100+16]; ld.local.f32 %f1871, [%rd100+20]; st.local.f32 [%rd100], %f1869; st.local.f32 [%rd100+4], %f1870; st.local.f32 [%rd100+8], %f1871; st.local.f32 [%rd99+24], %f388; st.local.f32 [%rd99+28], %f1862; st.local.f32 [%rd99+32], %f1863; st.local.u32 [%rd99+36], %r417; st.local.u32 [%rd99+40], %r418; st.local.u32 [%rd99+44], %r419; st.local.f32 [%rd100+12], %f1864; st.local.f32 [%rd100+16], %f1865; st.local.f32 [%rd100+20], %f1866; ld.local.f32 %f2413, [%rd99+24]; mov.u16 %rs50, 1; $L__BB26_164: cvt.u32.u64 %r423, %rd98; add.s32 %r610, %r423, 2; ld.local.f32 %f391, [%rd99+48]; setp.leu.ftz.f32 %p237, %f2413, %f391; @%p237 bra $L__BB26_166; ld.local.f32 %f1872, [%rd99+28]; ld.local.f32 %f1873, [%rd99+32]; ld.local.u32 %r424, [%rd99+36]; ld.local.u32 %r425, [%rd99+40]; ld.local.u32 %r426, [%rd99+44]; ld.local.f32 %f1874, [%rd100+12]; ld.local.f32 %f1875, [%rd100+16]; ld.local.f32 %f1876, [%rd100+20]; ld.local.f32 %f1877, [%rd99+52]; ld.local.f32 %f1878, [%rd99+56]; ld.local.u32 %r427, [%rd99+60]; ld.local.u32 %r428, [%rd99+64]; ld.local.u32 %r429, [%rd99+68]; st.local.f32 [%rd99+24], %f391; st.local.f32 [%rd99+28], %f1877; st.local.f32 [%rd99+32], %f1878; st.local.u32 [%rd99+36], %r427; st.local.u32 [%rd99+40], %r428; st.local.u32 [%rd99+44], %r429; ld.local.f32 %f1879, [%rd100+24]; ld.local.f32 %f1880, [%rd100+28]; ld.local.f32 %f1881, [%rd100+32]; st.local.f32 [%rd100+12], %f1879; st.local.f32 [%rd100+16], %f1880; st.local.f32 [%rd100+20], %f1881; st.local.f32 [%rd99+48], %f2413; st.local.f32 [%rd99+52], %f1872; st.local.f32 [%rd99+56], %f1873; st.local.u32 [%rd99+60], %r424; st.local.u32 [%rd99+64], %r425; st.local.u32 [%rd99+68], %r426; st.local.f32 [%rd100+24], %f1874; st.local.f32 [%rd100+28], %f1875; st.local.f32 [%rd100+32], %f1876; mov.u16 %rs50, 1; $L__BB26_166: add.s32 %r609, %r609, -2; setp.ne.s32 %p238, %r609, 0; @%p238 bra $L__BB26_162; $L__BB26_167: setp.eq.s32 %p239, %r152, 0; @%p239 bra $L__BB26_170; cvt.s64.s32 %rd101, %r610; mul.wide.s32 %rd288, %r610, 24; add.s64 %rd289, %rd1, %rd288; add.s64 %rd102, %rd289, 4; ld.local.f32 %f392, [%rd289+28]; ld.local.f32 %f393, [%rd289+4]; setp.leu.ftz.f32 %p240, %f393, %f392; @%p240 bra $L__BB26_170; ld.local.f32 %f1882, [%rd102+4]; ld.local.f32 %f1883, [%rd102+8]; ld.local.u32 %r430, [%rd102+12]; ld.local.u32 %r431, [%rd102+16]; ld.local.u32 %r432, [%rd102+20]; mul.lo.s64 %rd290, %rd101, 12; add.s64 %rd291, %rd1, %rd290; ld.local.f32 %f1884, [%rd291+100]; ld.local.f32 %f1885, [%rd291+104]; ld.local.f32 %f1886, [%rd291+108]; ld.local.f32 %f1887, [%rd102+28]; ld.local.f32 %f1888, [%rd102+32]; ld.local.u32 %r433, [%rd102+36]; ld.local.u32 %r434, [%rd102+40]; ld.local.u32 %r435, [%rd102+44]; st.local.f32 [%rd102], %f392; st.local.f32 [%rd102+4], %f1887; st.local.f32 [%rd102+8], %f1888; st.local.u32 [%rd102+12], %r433; st.local.u32 [%rd102+16], %r434; st.local.u32 [%rd102+20], %r435; ld.local.f32 %f1889, [%rd291+112]; ld.local.f32 %f1890, [%rd291+116]; ld.local.f32 %f1891, [%rd291+120]; st.local.f32 [%rd291+100], %f1889; st.local.f32 [%rd291+104], %f1890; st.local.f32 [%rd291+108], %f1891; st.local.f32 [%rd102+24], %f393; st.local.f32 [%rd102+28], %f1882; st.local.f32 [%rd102+32], %f1883; st.local.u32 [%rd102+36], %r430; st.local.u32 [%rd102+40], %r431; st.local.u32 [%rd102+44], %r432; st.local.f32 [%rd291+112], %f1884; st.local.f32 [%rd291+116], %f1885; st.local.f32 [%rd291+120], %f1886; mov.u16 %rs50, %rs35; $L__BB26_170: add.s32 %r607, %r607, 1; and.b16 %rs36, %rs50, 255; setp.ne.s16 %p241, %rs36, 0; @%p241 bra $L__BB26_159; setp.lt.s32 %p242, %r146, 1; mov.f32 %f2448, 0f00000000; @%p242 bra $L__BB26_200; add.u64 %rd351, %SPL, 768; ld.const.v2.u64 {%rd292, %rd293}, [kernel_params+2176]; mov.u32 %r611, 0; cvta.to.global.u64 %rd103, %rd293; and.b32 %r160, %r17, 8388608; ld.const.u64 %rd296, [kernel_params+2160]; cvta.to.global.u64 %rd104, %rd296; cvta.to.global.u64 %rd105, %rd292; mul.ftz.f32 %f394, %f19, 0f40400000; mul.ftz.f32 %f395, %f19, 0f40800000; mul.ftz.f32 %f396, %f20, 0f40400000; mul.ftz.f32 %f397, %f20, 0f40800000; mul.ftz.f32 %f398, %f21, 0f40400000; mul.ftz.f32 %f399, %f21, 0f40800000; neg.ftz.f32 %f400, %f92; mul.ftz.f32 %f401, %f19, 0f41800000; setp.gtu.ftz.f32 %p243, %f401, %f92; setp.neu.ftz.f32 %p244, %f92, 0f00000000; mov.f32 %f2448, 0f00000000; and.pred %p8, %p244, %p243; mul.ftz.f32 %f402, %f20, 0f41800000; setp.gtu.ftz.f32 %p245, %f402, %f92; and.pred %p9, %p244, %p245; mul.ftz.f32 %f403, %f21, 0f41800000; setp.gtu.ftz.f32 %p246, %f403, %f92; and.pred %p10, %p244, %p246; not.pred %p266, %p8; not.pred %p267, %p9; not.pred %p268, %p10; mov.u64 %rd350, %rd1; mov.u64 %rd352, %rd1; $L__BB26_173: ld.local.u32 %r437, [%rd352+20]; cvt.s64.s32 %rd109, %r437; mul.wide.s32 %rd297, %r437, 4; add.s64 %rd298, %rd103, %rd297; ld.local.f32 %f1894, [%rd350+100]; neg.ftz.f32 %f1895, %f1894; ld.local.f32 %f1896, [%rd350+104]; neg.ftz.f32 %f1897, %f1896; ld.local.f32 %f1898, [%rd350+108]; neg.ftz.f32 %f1899, %f1898; setp.eq.s32 %p247, %r160, 0; selp.f32 %f1900, %f1898, %f1899, %p247; selp.f32 %f1901, %f1896, %f1897, %p247; selp.f32 %f1902, %f1894, %f1895, %p247; ld.global.u32 %r162, [%rd298]; and.b32 %r438, %r162, 8; setp.eq.s32 %p248, %r438, 0; neg.ftz.f32 %f1903, %f1902; neg.ftz.f32 %f1904, %f1901; neg.ftz.f32 %f1905, %f1900; selp.f32 %f2435, %f1900, %f1905, %p248; selp.f32 %f2436, %f1901, %f1904, %p248; selp.f32 %f2437, %f1902, %f1903, %p248; and.b32 %r439, %r162, 4; setp.ne.s32 %p249, %r439, 0; @%p249 bra $L__BB26_183; and.b32 %r440, %r162, 2; setp.eq.s32 %p250, %r440, 0; mul.lo.s64 %rd299, %rd109, 208; add.s64 %rd300, %rd104, %rd299; add.s64 %rd110, %rd300, 48; @%p250 bra $L__BB26_181; bra.uni $L__BB26_175; $L__BB26_181: ld.global.v4.f32 {%f2434, %f2433, %f2432, %f2116}, [%rd110]; ld.global.v4.f32 {%f2431, %f2430, %f2429, %f2120}, [%rd110+16]; ld.global.v4.f32 {%f2428, %f2427, %f2426, %f2124}, [%rd110+32]; bra.uni $L__BB26_182; $L__BB26_175: ld.global.u32 %rd301, [%rd110+120]; ld.global.u32 %r441, [%rd110+104]; shl.b32 %r442, %r441, 1; cvt.rn.f32.s32 %f1906, %r442; mul.ftz.f32 %f1907, %f18, %f1906; cvt.rzi.ftz.s32.f32 %r443, %f1907; add.s32 %r444, %r442, -1; min.s32 %r445, %r443, %r444; cvt.rn.f32.s32 %f1908, %r445; sub.ftz.f32 %f408, %f1907, %f1908; cvt.s64.s32 %rd302, %r445; add.s64 %rd303, %rd302, %rd301; shl.b64 %rd304, %rd303, 6; add.s64 %rd111, %rd105, %rd304; ld.global.v4.f32 {%f1909, %f1910, %f1911, %f1912}, [%rd111]; ld.global.v4.f32 {%f1913, %f1914, %f1915, %f1916}, [%rd111+64]; mul.ftz.f32 %f1917, %f1910, %f1914; fma.rn.ftz.f32 %f1918, %f1909, %f1913, %f1917; mul.ftz.f32 %f1919, %f1912, %f1916; fma.rn.ftz.f32 %f1920, %f1911, %f1915, %f1919; add.ftz.f32 %f417, %f1918, %f1920; setp.gt.ftz.f32 %p251, %f417, 0f3F7FDF3B; sub.ftz.f32 %f418, %f612, %f408; @%p251 bra $L__BB26_177; bra.uni $L__BB26_176; $L__BB26_177: mul.ftz.f32 %f1975, %f1909, %f418; fma.rn.ftz.f32 %f1976, %f1913, %f408, %f1975; mul.ftz.f32 %f1977, %f1910, %f418; fma.rn.ftz.f32 %f1978, %f1914, %f408, %f1977; mul.ftz.f32 %f1979, %f1911, %f418; fma.rn.ftz.f32 %f1980, %f1915, %f408, %f1979; mul.ftz.f32 %f1981, %f1912, %f418; fma.rn.ftz.f32 %f1982, %f1916, %f408, %f1981; mul.ftz.f32 %f1983, %f1978, %f1978; fma.rn.ftz.f32 %f1984, %f1976, %f1976, %f1983; mul.ftz.f32 %f1985, %f1982, %f1982; fma.rn.ftz.f32 %f1986, %f1980, %f1980, %f1985; add.ftz.f32 %f1987, %f1984, %f1986; rsqrt.approx.ftz.f32 %f1988, %f1987; mul.ftz.f32 %f2418, %f1982, %f1988; mul.ftz.f32 %f2417, %f1980, %f1988; mul.ftz.f32 %f2416, %f1978, %f1988; mul.ftz.f32 %f2415, %f1976, %f1988; bra.uni $L__BB26_178; $L__BB26_176: mov.f32 %f1922, 0fBF800000; max.ftz.f32 %f1923, %f417, %f1922; min.ftz.f32 %f1925, %f1923, %f612; abs.ftz.f32 %f1926, %f1925; sub.ftz.f32 %f1927, %f612, %f1926; mul.ftz.f32 %f1928, %f1927, 0f3F000000; sqrt.approx.ftz.f32 %f1929, %f1928; setp.gt.ftz.f32 %p252, %f1926, 0f3F11EB85; selp.f32 %f1930, %f1929, %f1926, %p252; mul.ftz.f32 %f1931, %f1930, %f1930; mov.f32 %f1932, 0f3C94D2E9; mov.f32 %f1933, 0f3D53F941; fma.rn.ftz.f32 %f1934, %f1933, %f1931, %f1932; mov.f32 %f1935, 0f3D3F841F; fma.rn.ftz.f32 %f1936, %f1934, %f1931, %f1935; mov.f32 %f1937, 0f3D994929; fma.rn.ftz.f32 %f1938, %f1936, %f1931, %f1937; mov.f32 %f1939, 0f3E2AAB94; fma.rn.ftz.f32 %f1940, %f1938, %f1931, %f1939; mul.ftz.f32 %f1941, %f1931, %f1940; fma.rn.ftz.f32 %f1942, %f1941, %f1930, %f1930; add.ftz.f32 %f1943, %f1942, %f1942; mov.f32 %f1944, 0f3FC90FDB; sub.ftz.f32 %f1945, %f1944, %f1942; selp.f32 %f1946, %f1943, %f1945, %p252; setp.lt.ftz.f32 %p253, %f1925, 0f00000000; mov.f32 %f1947, 0f40490FDB; sub.ftz.f32 %f1948, %f1947, %f1946; selp.f32 %f1949, %f1948, %f1946, %p253; mul.ftz.f32 %f1950, %f1909, %f417; sub.ftz.f32 %f1951, %f1913, %f1950; mul.ftz.f32 %f1952, %f1910, %f417; sub.ftz.f32 %f1953, %f1914, %f1952; mul.ftz.f32 %f1954, %f1911, %f417; sub.ftz.f32 %f1955, %f1915, %f1954; mul.ftz.f32 %f1956, %f1912, %f417; sub.ftz.f32 %f1957, %f1916, %f1956; mul.ftz.f32 %f1958, %f1953, %f1953; fma.rn.ftz.f32 %f1959, %f1951, %f1951, %f1958; mul.ftz.f32 %f1960, %f1957, %f1957; fma.rn.ftz.f32 %f1961, %f1955, %f1955, %f1960; add.ftz.f32 %f1962, %f1959, %f1961; rsqrt.approx.ftz.f32 %f1963, %f1962; mul.ftz.f32 %f1964, %f1951, %f1963; mul.ftz.f32 %f1965, %f1953, %f1963; mul.ftz.f32 %f1966, %f1955, %f1963; mul.ftz.f32 %f1967, %f1957, %f1963; mul.ftz.f32 %f1968, %f408, %f1949; cos.approx.ftz.f32 %f1969, %f1968; sin.approx.ftz.f32 %f1970, %f1968; mul.ftz.f32 %f1971, %f1964, %f1970; mul.ftz.f32 %f1972, %f1965, %f1970; mul.ftz.f32 %f1973, %f1966, %f1970; mul.ftz.f32 %f1974, %f1967, %f1970; fma.rn.ftz.f32 %f2418, %f1912, %f1969, %f1974; fma.rn.ftz.f32 %f2417, %f1911, %f1969, %f1973; fma.rn.ftz.f32 %f2416, %f1910, %f1969, %f1972; fma.rn.ftz.f32 %f2415, %f1909, %f1969, %f1971; $L__BB26_178: ld.global.f32 %f1989, [%rd111+28]; ld.global.f32 %f1990, [%rd111+92]; mul.ftz.f32 %f1991, %f408, %f1990; fma.rn.ftz.f32 %f1992, %f418, %f1989, %f1991; ld.global.v4.f32 {%f1993, %f1994, %f1995, %f1996}, [%rd111+32]; ld.global.v4.f32 {%f2001, %f2002, %f2003, %f2004}, [%rd111+96]; mul.ftz.f32 %f2009, %f408, %f2001; mul.ftz.f32 %f2010, %f408, %f2002; mul.ftz.f32 %f2011, %f408, %f2003; mul.ftz.f32 %f2012, %f408, %f2004; fma.rn.ftz.f32 %f2013, %f418, %f1993, %f2009; fma.rn.ftz.f32 %f2014, %f418, %f1994, %f2010; fma.rn.ftz.f32 %f2015, %f418, %f1995, %f2011; fma.rn.ftz.f32 %f2016, %f418, %f1996, %f2012; ld.global.v4.f32 {%f2017, %f2018, %f2019, %f2020}, [%rd111+48]; ld.global.v4.f32 {%f2025, %f2026, %f2027, %f2028}, [%rd111+112]; mul.ftz.f32 %f2033, %f408, %f2025; mul.ftz.f32 %f2034, %f408, %f2026; mul.ftz.f32 %f2035, %f408, %f2027; mul.ftz.f32 %f2036, %f408, %f2028; fma.rn.ftz.f32 %f2037, %f418, %f2017, %f2033; fma.rn.ftz.f32 %f2038, %f418, %f2018, %f2034; fma.rn.ftz.f32 %f2039, %f418, %f2019, %f2035; fma.rn.ftz.f32 %f2040, %f418, %f2020, %f2036; mul.ftz.f32 %f2041, %f2418, 0f3FB504F3; mul.ftz.f32 %f2042, %f2415, 0f3FB504F3; mul.ftz.f32 %f2043, %f2416, 0f3FB504F3; mul.ftz.f32 %f2044, %f2417, 0f3FB504F3; mul.ftz.f32 %f2045, %f2041, %f2042; mul.ftz.f32 %f2046, %f2041, %f2043; mul.ftz.f32 %f2047, %f2041, %f2044; mul.ftz.f32 %f2048, %f2042, %f2042; mul.ftz.f32 %f2049, %f2042, %f2043; mul.ftz.f32 %f2050, %f2042, %f2044; mul.ftz.f32 %f2051, %f2043, %f2043; mul.ftz.f32 %f2052, %f2043, %f2044; mul.ftz.f32 %f2053, %f2044, %f2044; sub.ftz.f32 %f2055, %f612, %f2051; sub.ftz.f32 %f2056, %f2055, %f2053; sub.ftz.f32 %f2057, %f2049, %f2047; add.ftz.f32 %f2058, %f2046, %f2050; add.ftz.f32 %f2059, %f2047, %f2049; sub.ftz.f32 %f2060, %f612, %f2048; sub.ftz.f32 %f2061, %f2060, %f2053; sub.ftz.f32 %f2062, %f2052, %f2045; sub.ftz.f32 %f2063, %f2050, %f2046; add.ftz.f32 %f2064, %f2045, %f2052; sub.ftz.f32 %f2065, %f2060, %f2051; mul.ftz.f32 %f2066, %f2057, %f2015; fma.rn.ftz.f32 %f2067, %f1992, %f2056, %f2066; fma.rn.ftz.f32 %f2424, %f2058, %f2038, %f2067; mul.ftz.f32 %f2068, %f2057, %f2016; fma.rn.ftz.f32 %f2069, %f2056, %f2013, %f2068; fma.rn.ftz.f32 %f432, %f2058, %f2039, %f2069; mul.ftz.f32 %f2070, %f2057, %f2037; fma.rn.ftz.f32 %f2071, %f2056, %f2014, %f2070; fma.rn.ftz.f32 %f433, %f2058, %f2040, %f2071; mul.ftz.f32 %f2072, %f2061, %f2015; fma.rn.ftz.f32 %f2073, %f2059, %f1992, %f2072; fma.rn.ftz.f32 %f434, %f2062, %f2038, %f2073; mul.ftz.f32 %f2074, %f2061, %f2016; fma.rn.ftz.f32 %f2075, %f2059, %f2013, %f2074; fma.rn.ftz.f32 %f2423, %f2062, %f2039, %f2075; mul.ftz.f32 %f2076, %f2061, %f2037; fma.rn.ftz.f32 %f2077, %f2059, %f2014, %f2076; fma.rn.ftz.f32 %f436, %f2062, %f2040, %f2077; mul.ftz.f32 %f2078, %f2064, %f2015; fma.rn.ftz.f32 %f2079, %f2063, %f1992, %f2078; fma.rn.ftz.f32 %f437, %f2065, %f2038, %f2079; mul.ftz.f32 %f2080, %f2064, %f2016; fma.rn.ftz.f32 %f2081, %f2063, %f2013, %f2080; fma.rn.ftz.f32 %f438, %f2065, %f2039, %f2081; mul.ftz.f32 %f2082, %f2064, %f2037; fma.rn.ftz.f32 %f2083, %f2063, %f2014, %f2082; fma.rn.ftz.f32 %f2422, %f2065, %f2040, %f2083; mul.ftz.f32 %f2084, %f2423, %f2422; mul.ftz.f32 %f440, %f438, %f436; sub.ftz.f32 %f2421, %f2084, %f440; mul.ftz.f32 %f442, %f438, %f433; mul.ftz.f32 %f2085, %f432, %f2422; sub.ftz.f32 %f2420, %f442, %f2085; mul.ftz.f32 %f444, %f432, %f436; mul.ftz.f32 %f2086, %f2423, %f433; sub.ftz.f32 %f2419, %f444, %f2086; mul.ftz.f32 %f2087, %f434, %f2420; fma.rn.ftz.f32 %f2088, %f2424, %f2421, %f2087; fma.rn.ftz.f32 %f2425, %f437, %f2419, %f2088; setp.neu.ftz.f32 %p254, %f2425, 0f00000000; @%p254 bra $L__BB26_180; add.ftz.f32 %f2424, %f2424, 0f322BCC77; add.ftz.f32 %f2422, %f2422, 0f322BCC77; add.ftz.f32 %f2423, %f2423, 0f322BCC77; mul.ftz.f32 %f2089, %f2423, %f2422; sub.ftz.f32 %f2421, %f2089, %f440; mul.ftz.f32 %f2090, %f432, %f2422; sub.ftz.f32 %f2420, %f442, %f2090; mul.ftz.f32 %f2091, %f2423, %f433; sub.ftz.f32 %f2419, %f444, %f2091; mul.ftz.f32 %f2092, %f434, %f2420; fma.rn.ftz.f32 %f2093, %f2424, %f2421, %f2092; fma.rn.ftz.f32 %f2094, %f437, %f2419, %f2093; setp.eq.ftz.f32 %p255, %f2094, 0f00000000; selp.f32 %f2425, 0f7F7FFFFF, %f2094, %p255; $L__BB26_180: div.approx.ftz.f32 %f2434, %f2421, %f2425; div.approx.ftz.f32 %f2433, %f2420, %f2425; div.approx.ftz.f32 %f2432, %f2419, %f2425; mul.ftz.f32 %f2095, %f434, %f2422; mul.ftz.f32 %f2096, %f437, %f436; sub.ftz.f32 %f2097, %f2096, %f2095; mul.ftz.f32 %f2098, %f437, %f433; mul.ftz.f32 %f2099, %f2422, %f2424; sub.ftz.f32 %f2100, %f2099, %f2098; mul.ftz.f32 %f2101, %f436, %f2424; mul.ftz.f32 %f2102, %f434, %f433; sub.ftz.f32 %f2103, %f2102, %f2101; div.approx.ftz.f32 %f2431, %f2097, %f2425; div.approx.ftz.f32 %f2430, %f2100, %f2425; div.approx.ftz.f32 %f2429, %f2103, %f2425; mul.ftz.f32 %f2104, %f437, %f2423; mul.ftz.f32 %f2105, %f434, %f438; sub.ftz.f32 %f2106, %f2105, %f2104; mul.ftz.f32 %f2107, %f438, %f2424; mul.ftz.f32 %f2108, %f437, %f432; sub.ftz.f32 %f2109, %f2108, %f2107; mul.ftz.f32 %f2110, %f434, %f432; mul.ftz.f32 %f2111, %f2423, %f2424; sub.ftz.f32 %f2112, %f2111, %f2110; div.approx.ftz.f32 %f2428, %f2106, %f2425; div.approx.ftz.f32 %f2427, %f2109, %f2425; div.approx.ftz.f32 %f2426, %f2112, %f2425; $L__BB26_182: mul.ftz.f32 %f2125, %f2437, %f2434; fma.rn.ftz.f32 %f2126, %f2436, %f2431, %f2125; fma.rn.ftz.f32 %f2127, %f2435, %f2428, %f2126; mul.ftz.f32 %f2128, %f2437, %f2433; fma.rn.ftz.f32 %f2129, %f2436, %f2430, %f2128; fma.rn.ftz.f32 %f2130, %f2435, %f2427, %f2129; mul.ftz.f32 %f2131, %f2437, %f2432; fma.rn.ftz.f32 %f2132, %f2436, %f2429, %f2131; fma.rn.ftz.f32 %f2133, %f2435, %f2426, %f2132; mul.ftz.f32 %f2134, %f2127, %f2127; fma.rn.ftz.f32 %f2135, %f2130, %f2130, %f2134; fma.rn.ftz.f32 %f2136, %f2133, %f2133, %f2135; sqrt.approx.ftz.f32 %f2137, %f2136; div.approx.ftz.f32 %f2437, %f2127, %f2137; div.approx.ftz.f32 %f2436, %f2130, %f2137; div.approx.ftz.f32 %f2435, %f2133, %f2137; $L__BB26_183: ld.local.f32 %f2138, [%rd6+12]; ld.local.f32 %f2139, [%rd352+4]; ld.local.f32 %f2140, [%rd6+16]; ld.local.f32 %f2141, [%rd6+20]; ld.local.f32 %f2142, [%rd6]; fma.rn.ftz.f32 %f494, %f2139, %f2138, %f2142; ld.local.f32 %f2143, [%rd6+4]; fma.rn.ftz.f32 %f495, %f2139, %f2140, %f2143; ld.local.f32 %f2144, [%rd6+8]; fma.rn.ftz.f32 %f496, %f2139, %f2141, %f2144; mul.ftz.f32 %f2145, %f2308, %f2437; fma.rn.ftz.f32 %f2146, %f2307, %f2436, %f2145; fma.rn.ftz.f32 %f2147, %f2306, %f2435, %f2146; abs.ftz.f32 %f2148, %f2147; mul.ftz.f32 %f2149, %f2315, %f2148; mul.ftz.f32 %f2150, %f2311, %f2437; fma.rn.ftz.f32 %f2151, %f2310, %f2436, %f2150; fma.rn.ftz.f32 %f2152, %f2309, %f2435, %f2151; abs.ftz.f32 %f2153, %f2152; mul.ftz.f32 %f2154, %f2316, %f2153; mul.ftz.f32 %f2155, %f2314, %f2437; fma.rn.ftz.f32 %f2156, %f2313, %f2436, %f2155; fma.rn.ftz.f32 %f2157, %f2312, %f2435, %f2156; abs.ftz.f32 %f2158, %f2157; mul.ftz.f32 %f2159, %f2317, %f2158; mul.ftz.f32 %f2160, %f2154, %f2154; fma.rn.ftz.f32 %f2161, %f2149, %f2149, %f2160; fma.rn.ftz.f32 %f2162, %f2159, %f2159, %f2161; div.approx.ftz.f32 %f2438, %f2149, %f2162; ld.local.u32 %r163, [%rd1]; setp.lt.s32 %p256, %r163, 5; @%p256 bra $L__BB26_185; cvt.rn.f32.s32 %f2163, %r163; mov.f32 %f2164, 0f40800000; div.approx.ftz.f32 %f2165, %f2163, %f2164; mul.ftz.f32 %f2438, %f2438, %f2165; $L__BB26_185: sub.ftz.f32 %f2167, %f494, %f14; sub.ftz.f32 %f2168, %f495, %f15; mul.ftz.f32 %f2169, %f2168, %f2168; fma.rn.ftz.f32 %f2170, %f2167, %f2167, %f2169; sub.ftz.f32 %f2171, %f496, %f16; fma.rn.ftz.f32 %f2172, %f2171, %f2171, %f2170; sqrt.approx.ftz.f32 %f500, %f2172; setp.eq.ftz.f32 %p257, %f500, 0f00000000; mov.f32 %f2445, 0f00000000; neg.ftz.f32 %f501, %f500; setp.le.ftz.f32 %p258, %f401, %f500; or.pred %p259, %p257, %p258; mov.f32 %f2439, %f2445; @%p259 bra $L__BB26_187; div.approx.ftz.f32 %f2173, %f501, %f394; mul.ftz.f32 %f2174, %f2173, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2175, %f2174; mul.ftz.f32 %f2176, %f2175, %f2175; fma.rn.ftz.f32 %f2177, %f2175, %f2176, %f2175; div.approx.ftz.f32 %f2178, %f2177, %f395; mul.ftz.f32 %f2439, %f2178, 0f3F807715; $L__BB26_187: setp.le.ftz.f32 %p260, %f402, %f500; or.pred %p262, %p257, %p260; mov.f32 %f2440, %f2445; @%p262 bra $L__BB26_189; div.approx.ftz.f32 %f2180, %f501, %f396; mul.ftz.f32 %f2181, %f2180, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2182, %f2181; mul.ftz.f32 %f2183, %f2182, %f2182; fma.rn.ftz.f32 %f2184, %f2182, %f2183, %f2182; div.approx.ftz.f32 %f2185, %f2184, %f397; mul.ftz.f32 %f2440, %f2185, 0f3F807715; $L__BB26_189: setp.le.ftz.f32 %p263, %f403, %f500; or.pred %p265, %p257, %p263; mov.f32 %f2441, %f2445; @%p265 bra $L__BB26_191; div.approx.ftz.f32 %f2187, %f501, %f398; mul.ftz.f32 %f2188, %f2187, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2189, %f2188; mul.ftz.f32 %f2190, %f2189, %f2189; fma.rn.ftz.f32 %f2191, %f2189, %f2190, %f2189; div.approx.ftz.f32 %f2192, %f2191, %f399; mul.ftz.f32 %f2441, %f2192, 0f3F807715; $L__BB26_191: mov.f32 %f2442, %f2445; @%p266 bra $L__BB26_193; div.approx.ftz.f32 %f2194, %f400, %f394; mul.ftz.f32 %f2195, %f2194, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2196, %f2195; mul.ftz.f32 %f2197, %f2196, %f2196; fma.rn.ftz.f32 %f2198, %f2196, %f2197, %f2196; div.approx.ftz.f32 %f2199, %f2198, %f395; mul.ftz.f32 %f2442, %f2199, 0f3F807715; $L__BB26_193: mov.f32 %f2443, %f2445; @%p267 bra $L__BB26_195; div.approx.ftz.f32 %f2201, %f400, %f396; mul.ftz.f32 %f2202, %f2201, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2203, %f2202; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2203, %f2204, %f2203; div.approx.ftz.f32 %f2206, %f2205, %f397; mul.ftz.f32 %f2443, %f2206, 0f3F807715; $L__BB26_195: mov.f32 %f2444, %f2445; @%p268 bra $L__BB26_197; div.approx.ftz.f32 %f2208, %f400, %f398; mul.ftz.f32 %f2209, %f2208, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2210, %f2209; mul.ftz.f32 %f2211, %f2210, %f2210; fma.rn.ftz.f32 %f2212, %f2210, %f2211, %f2210; div.approx.ftz.f32 %f2213, %f2212, %f399; mul.ftz.f32 %f2444, %f2213, 0f3F807715; $L__BB26_197: add.ftz.f32 %f2217, %f2442, %f2443; add.ftz.f32 %f2218, %f2217, %f2444; div.approx.ftz.f32 %f514, %f2218, %f48; setp.leu.ftz.f32 %p269, %f514, 0f00000000; mov.f32 %f2446, %f2445; mov.f32 %f2447, %f2445; @%p269 bra $L__BB26_199; div.approx.ftz.f32 %f2447, %f2441, %f514; div.approx.ftz.f32 %f2446, %f2440, %f514; div.approx.ftz.f32 %f2445, %f2439, %f514; $L__BB26_199: mul.ftz.f32 %f2219, %f2438, %f2445; mul.ftz.f32 %f2220, %f2438, %f2446; mul.ftz.f32 %f2221, %f2438, %f2447; st.local.f32 [%rd350+100], %f2437; st.local.f32 [%rd350+104], %f2436; st.local.f32 [%rd350+108], %f2435; st.local.f32 [%rd351], %f2219; st.local.f32 [%rd351+4], %f2220; st.local.f32 [%rd351+8], %f2221; abs.ftz.f32 %f2222, %f2219; abs.ftz.f32 %f2223, %f2220; abs.ftz.f32 %f2224, %f2221; add.ftz.f32 %f2225, %f2222, %f2223; add.ftz.f32 %f2226, %f2224, %f2225; fma.rn.ftz.f32 %f2448, %f2226, 0f3EAAAAAB, %f2448; add.s64 %rd352, %rd352, 24; add.s64 %rd351, %rd351, 12; add.s64 %rd350, %rd350, 12; add.s32 %r611, %r611, 1; setp.lt.s32 %p270, %r611, %r147; @%p270 bra $L__BB26_173; $L__BB26_200: setp.eq.ftz.f32 %p271, %f2448, 0f00000000; mov.u16 %rs55, 0; @%p271 bra $L__BB26_214; bra.uni $L__BB26_201; $L__BB26_214: cvt.u32.u16 %r543, %rs55; st.param.b32 [func_retval0+0], %r543; ret; $L__BB26_201: ld.const.u32 %r545, [kernel_params+1596]; setp.eq.s32 %p283, %r545, 0; ld.param.u32 %r544, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1+4]; add.s32 %r165, %r544, 1; @%p283 bra $L__BB26_204; ld.param.u32 %r547, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1+8]; ld.param.u32 %r546, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1]; ld.const.v2.u32 {%r446, %r447}, [kernel_params+1600]; mov.b32 %f523, %r447; setp.geu.ftz.f32 %p273, %f523, 0f3F800000; setp.lt.ftz.f32 %p274, %f523, 0f3F800000; ld.const.u32 %r450, [kernel_params+1576]; selp.b32 %r451, %r450, %r546, %p274; and.b32 %r452, %r165, 255; xor.b32 %r453, %r451, %r452; mul.lo.s32 %r454, %r453, -512718531; shr.u32 %r455, %r451, 16; xor.b32 %r456, %r454, %r455; bfe.u32 %r457, %r456, 4, 4; shr.u32 %r458, %r451, 8; xor.b32 %r459, %r456, %r458; xor.b32 %r460, %r459, %r457; mul.lo.s32 %r461, %r460, 153742143; shr.u32 %r462, %r451, 23; xor.b32 %r463, %r461, %r462; bfe.u32 %r464, %r463, 1, 7; xor.b32 %r465, %r464, %r463; shr.u32 %r466, %r451, 27; or.b32 %r467, %r466, 1; mul.lo.s32 %r468, %r467, %r465; mul.lo.s32 %r469, %r468, -1074439621; bfe.u32 %r470, %r469, 2, 6; xor.b32 %r471, %r470, %r469; mul.lo.s32 %r472, %r471, -1638916925; bfe.u32 %r473, %r472, 2, 6; xor.b32 %r474, %r473, %r472; mul.lo.s32 %r475, %r474, -933190689; and.b32 %r476, %r475, 255; bfe.u32 %r477, %r475, 5, 3; xor.b32 %r478, %r477, %r476; add.s32 %r479, %r446, -1; xor.b32 %r166, %r165, 61; xor.b32 %r480, %r451, %r166; mul.lo.s32 %r481, %r480, 9; shr.u32 %r482, %r481, 4; xor.b32 %r483, %r482, %r481; mul.lo.s32 %r484, %r483, 668265261; add.s32 %r485, %r484, %r5; shr.u32 %r486, %r484, 16; or.b32 %r487, %r486, 1; mul.lo.s32 %r488, %r487, %r485; mul.lo.s32 %r489, %r488, 89287766; xor.b32 %r490, %r489, %r488; mul.lo.s32 %r491, %r490, 1403136100; xor.b32 %r492, %r491, %r490; brev.b32 %r493, %r492; neg.s32 %r494, %r446; and.b32 %r495, %r547, %r494; and.b32 %r496, %r493, %r479; or.b32 %r497, %r496, %r495; mad.lo.s32 %r498, %r478, %r446, %r497; shl.b32 %r499, %r446, 8; rem.u32 %r500, %r498, %r499; ld.const.u64 %rd305, [kernel_params+2424]; cvta.to.global.u64 %rd306, %rd305; shl.b32 %r501, %r500, 1; mul.wide.u32 %rd307, %r501, 4; add.s64 %rd308, %rd306, %rd307; ld.global.f32 %f2449, [%rd308]; @%p273 bra $L__BB26_205; ld.param.u32 %r549, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1]; xor.b32 %r502, %r166, %r549; mul.lo.s32 %r503, %r502, 9; shr.u32 %r504, %r503, 4; xor.b32 %r505, %r504, %r503; mul.lo.s32 %r506, %r505, 668265261; cvt.rn.f32.u32 %f2227, %r506; mul.ftz.f32 %f2228, %f2227, 0f2F7FFFFE; fma.rn.ftz.f32 %f2229, %f2228, %f523, %f2449; cvt.rmi.ftz.f32.f32 %f2230, %f2229; sub.ftz.f32 %f2449, %f2229, %f2230; bra.uni $L__BB26_205; $L__BB26_204: ld.param.u32 %r550, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_1]; shr.u32 %r507, %r165, 16; xor.b32 %r508, %r507, %r165; mul.lo.s32 %r509, %r508, 569420461; shr.u32 %r510, %r509, 15; xor.b32 %r511, %r510, %r509; mul.lo.s32 %r512, %r511, -749064809; shr.u32 %r513, %r512, 15; xor.b32 %r514, %r550, %r512; xor.b32 %r515, %r514, %r513; xor.b32 %r516, %r515, 1493655573; add.s32 %r517, %r516, %r5; shr.u32 %r518, %r516, 16; or.b32 %r519, %r518, 1; mul.lo.s32 %r520, %r519, %r517; mul.lo.s32 %r521, %r520, 89287766; xor.b32 %r522, %r521, %r520; mul.lo.s32 %r523, %r522, 1403136100; xor.b32 %r524, %r523, %r522; xor.b32 %r525, %r515, -2052961194; brev.b32 %r526, %r524; mul.lo.s32 %r527, %r526, 1025551850; xor.b32 %r528, %r527, %r526; add.s32 %r529, %r528, %r525; shr.u32 %r530, %r525, 16; or.b32 %r531, %r530, 1; mul.lo.s32 %r532, %r529, %r531; mul.lo.s32 %r533, %r532, 89287766; xor.b32 %r534, %r533, %r532; mul.lo.s32 %r535, %r534, 1403136100; xor.b32 %r536, %r535, %r534; brev.b32 %r537, %r536; cvt.rn.f32.u32 %f2231, %r537; mul.ftz.f32 %f2449, %f2231, 0f2F7FFFFE; $L__BB26_205: mul.ftz.f32 %f528, %f2448, %f2449; @%p242 bra $L__BB26_214; add.u64 %rd311, %SPL, 768; ld.local.v4.f32 {%f2450, %f2451, %f2452, %f2235}, [%rd311]; mov.u64 %rd353, 0; abs.ftz.f32 %f2236, %f2450; abs.ftz.f32 %f2237, %f2451; abs.ftz.f32 %f2238, %f2452; add.ftz.f32 %f2239, %f2236, %f2237; add.ftz.f32 %f2240, %f2238, %f2239; mul.ftz.f32 %f2453, %f2240, 0f3EAAAAAB; add.ftz.f32 %f533, %f2453, 0f00000000; setp.lt.ftz.f32 %p276, %f528, %f533; @%p276 bra $L__BB26_213; bra.uni $L__BB26_207; $L__BB26_213: ld.param.u32 %r548, [_Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection_param_0]; cvt.s64.s32 %rd326, %r548; mul.ftz.f32 %f2263, %f2448, %f2450; div.approx.ftz.f32 %f2264, %f2263, %f2453; mul.ftz.f32 %f2265, %f2448, %f2451; div.approx.ftz.f32 %f2266, %f2265, %f2453; mul.ftz.f32 %f2267, %f2448, %f2452; div.approx.ftz.f32 %f2268, %f2267, %f2453; ld.const.u64 %rd315, [kernel_params+2584]; cvta.to.global.u64 %rd316, %rd315; mul.lo.s64 %rd317, %rd326, 12; add.s64 %rd318, %rd316, %rd317; ld.global.f32 %f2269, [%rd318]; mul.ftz.f32 %f2270, %f2264, %f2269; ld.global.f32 %f2271, [%rd318+4]; mul.ftz.f32 %f2272, %f2266, %f2271; ld.global.f32 %f2273, [%rd318+8]; mul.ftz.f32 %f2274, %f2268, %f2273; st.global.f32 [%rd318], %f2270; st.global.f32 [%rd318+4], %f2272; st.global.f32 [%rd318+8], %f2274; mul.lo.s64 %rd319, %rd353, 24; add.s64 %rd320, %rd1, %rd319; ld.local.f32 %f2275, [%rd320+4]; ld.local.f32 %f2276, [%rd320+8]; ld.local.f32 %f2277, [%rd320+12]; ld.local.u32 %r538, [%rd320+16]; ld.local.u32 %r539, [%rd320+20]; ld.local.u32 %r540, [%rd320+24]; st.local.f32 [%rd1+4], %f2275; st.local.f32 [%rd1+8], %f2276; st.local.f32 [%rd1+12], %f2277; st.local.u32 [%rd1+16], %r538; st.local.u32 [%rd1+20], %r539; st.local.u32 [%rd1+24], %r540; mul.lo.s64 %rd321, %rd353, 12; add.s64 %rd322, %rd1, %rd321; ld.local.f32 %f2278, [%rd322+100]; ld.local.f32 %f2279, [%rd322+104]; ld.local.f32 %f2280, [%rd322+108]; st.local.f32 [%rd1+100], %f2278; st.local.f32 [%rd1+104], %f2279; st.local.f32 [%rd1+108], %f2280; ld.local.f32 %f2281, [%rd6+12]; ld.local.f32 %f2282, [%rd320+4]; ld.local.f32 %f2283, [%rd6+16]; ld.local.f32 %f2284, [%rd6+20]; ld.local.f32 %f2285, [%rd6]; fma.rn.ftz.f32 %f2286, %f2282, %f2281, %f2285; ld.local.f32 %f2287, [%rd6+4]; fma.rn.ftz.f32 %f2288, %f2282, %f2283, %f2287; ld.local.f32 %f2289, [%rd6+8]; fma.rn.ftz.f32 %f2290, %f2282, %f2284, %f2289; st.local.f32 [%rd6], %f2286; st.local.f32 [%rd6+4], %f2288; st.local.f32 [%rd6+8], %f2290; ld.local.f32 %f2291, [%rd322+100]; ld.local.f32 %f2292, [%rd322+104]; ld.local.f32 %f2293, [%rd322+108]; st.local.f32 [%rd6+12], %f2291; st.local.f32 [%rd6+16], %f2292; st.local.f32 [%rd6+20], %f2293; mov.u32 %r541, 0; st.local.u32 [%rd6+24], %r541; mov.u32 %r542, 1065353216; st.local.u32 [%rd6+28], %r542; mov.u16 %rs55, 1; bra.uni $L__BB26_214; $L__BB26_207: setp.lt.s32 %p277, %r147, 2; @%p277 bra $L__BB26_214; add.u64 %rd324, %SPL, 768; add.s64 %rd115, %rd324, 12; ld.local.v2.f32 {%f2451, %f2452}, [%rd324+16]; mov.u64 %rd353, 1; ld.local.f32 %f2450, [%rd324+12]; abs.ftz.f32 %f2243, %f2450; abs.ftz.f32 %f2244, %f2451; abs.ftz.f32 %f2245, %f2452; add.ftz.f32 %f2246, %f2243, %f2244; add.ftz.f32 %f2247, %f2245, %f2246; mul.ftz.f32 %f2453, %f2247, 0f3EAAAAAB; add.ftz.f32 %f538, %f533, %f2453; setp.lt.ftz.f32 %p278, %f528, %f538; @%p278 bra $L__BB26_213; setp.lt.s32 %p279, %r147, 3; @%p279 bra $L__BB26_214; ld.local.v2.f32 {%f2450, %f2451}, [%rd115+12]; abs.ftz.f32 %f2250, %f2450; abs.ftz.f32 %f2251, %f2451; ld.local.f32 %f2452, [%rd115+20]; abs.ftz.f32 %f2252, %f2452; add.ftz.f32 %f2253, %f2250, %f2251; add.ftz.f32 %f2254, %f2252, %f2253; mul.ftz.f32 %f2453, %f2254, 0f3EAAAAAB; add.ftz.f32 %f543, %f538, %f2453; setp.lt.ftz.f32 %p280, %f528, %f543; mov.u64 %rd353, 2; @%p280 bra $L__BB26_213; setp.lt.s32 %p281, %r147, 4; @%p281 bra $L__BB26_214; ld.local.v2.f32 {%f2451, %f2452}, [%rd115+28]; ld.local.f32 %f2450, [%rd115+24]; abs.ftz.f32 %f2257, %f2450; abs.ftz.f32 %f2258, %f2451; abs.ftz.f32 %f2259, %f2452; add.ftz.f32 %f2260, %f2257, %f2258; add.ftz.f32 %f2261, %f2259, %f2260; mul.ftz.f32 %f2453, %f2261, 0f3EAAAAAB; add.ftz.f32 %f2262, %f543, %f2453; setp.geu.ftz.f32 %p282, %f528, %f2262; mov.u64 %rd353, 3; @%p282 bra $L__BB26_214; bra.uni $L__BB26_213; } .func _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .local .align 16 .b8 __local_depot27[48]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<677>; .reg .b32 %r<249>; .reg .b64 %rd<302>; mov.u64 %SPL, __local_depot27; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd65, [_Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd66, [_Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; add.u64 %rd67, %SP, 0; add.u64 %rd1, %SPL, 0; cvta.to.local.u64 %rd68, %rd65; add.s64 %rd2, %rd68, 76; ld.local.u32 %r3, [%rd68+76]; cvt.s64.s32 %rd3, %r3; setp.eq.s32 %p1, %r3, -1; @%p1 bra $L__BB27_16; cvt.u64.u32 %rd4, %r62; ld.local.s32 %rd5, [%rd2+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r4, [%rd72+164]; ld.local.u32 %r31, [%rd2+-12]; and.b32 %r66, %r31, 1; setp.eq.b32 %p2, %r66, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r231, 0; @%p5 bra $L__BB27_5; cvt.u32.u64 %r68, %rd5; setp.eq.s32 %p6, %r68, -1; mov.u32 %r230, -1; @%p6 bra $L__BB27_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r230, [%rd76]; $L__BB27_4: setp.ne.s32 %p7, %r230, -1; selp.u32 %r231, 1, 0, %p7; $L__BB27_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r233, %r231, %r4; mul.wide.u32 %rd78, %r233, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd288, %rd79, 14; ld.global.u32 %r232, [%rd79+8]; ld.global.u16 %rs37, [%rd79+12]; ld.global.u64 %rd287, [%rd79]; setp.eq.s64 %p8, %rd287, %rd4; add.s64 %rd289, %rd79, 15; @%p8 bra $L__BB27_11; $L__BB27_6: setp.eq.s64 %p9, %rd287, 0; @%p9 bra $L__BB27_8; add.s32 %r233, %r233, 2; bra.uni $L__BB27_9; $L__BB27_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r239, -1; mov.u32 %r236, 0; mov.u32 %r233, %r232; mov.u32 %r237, %r236; mov.u32 %r238, %r236; @%p10 bra $L__BB27_15; $L__BB27_9: cvt.u64.u32 %rd11, %r233; mul.wide.u32 %rd80, %r233, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r232, [%rd81+8]; ld.global.u16 %rs37, [%rd81+12]; ld.global.u64 %rd287, [%rd81]; setp.ne.s64 %p11, %rd287, %rd4; @%p11 bra $L__BB27_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd288, %rd83, 14; add.s64 %rd289, %rd83, 15; $L__BB27_11: ld.global.u8 %rs5, [%rd289]; ld.global.u8 %rs6, [%rd288]; cvt.u32.u64 %r73, %rd5; setp.ne.s32 %p12, %r73, -1; @%p12 bra $L__BB27_14; mov.u32 %r239, -1; mov.u32 %r236, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB27_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r237, %r236; mov.u32 %r238, %r236; @%p14 bra $L__BB27_15; $L__BB27_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r239, -1, %r232, %p15; cvt.u32.u16 %r236, %rs37; cvt.u32.u16 %r78, %rs5; and.b32 %r238, %r78, 255; cvt.u32.u16 %r79, %rs6; and.b32 %r237, %r79, 255; $L__BB27_15: setp.eq.s32 %p16, %r239, -1; selp.b32 %r32, 0, %r239, %p16; selp.b32 %r33, 0, %r238, %p16; selp.b32 %r243, %r64, %r237, %p16; selp.b32 %r35, 0, %r236, %p16; bra.uni $L__BB27_17; $L__BB27_16: ld.local.u32 %r31, [%rd2+-12]; mov.u32 %r32, 0; mov.u32 %r33, %r32; mov.u32 %r243, %r64; mov.u32 %r35, %r32; $L__BB27_17: cvta.to.local.u64 %rd84, %rd66; mul.wide.u32 %rd85, %r63, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r31, 16; @%p17 bra $L__BB27_159; bra.uni $L__BB27_18; $L__BB27_159: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB27_163; bra.uni $L__BB27_160; $L__BB27_163: mov.u32 %r229, 0; st.local.u32 [%rd17], %r229; bra.uni $L__BB27_164; $L__BB27_18: setp.eq.s32 %p18, %r62, 7; setp.eq.s32 %p19, %r35, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB27_150; bra.uni $L__BB27_19; $L__BB27_150: cvt.u32.u64 %r225, %rd3; setp.eq.s32 %p98, %r225, -1; ld.local.f32 %f548, [%rd2+-48]; ld.local.f32 %f549, [%rd2+-52]; setp.neu.ftz.f32 %p99, %f549, %f548; ld.local.f32 %f550, [%rd2+-44]; setp.neu.ftz.f32 %p100, %f549, %f550; or.pred %p101, %p99, %p100; sub.ftz.f32 %f551, %f550, %f548; neg.ftz.f32 %f552, %f548; neg.ftz.f32 %f553, %f550; selp.f32 %f554, %f553, %f550, %p101; add.ftz.f32 %f555, %f549, %f554; selp.f32 %f556, %f548, %f552, %p101; sub.ftz.f32 %f557, %f556, %f549; mul.ftz.f32 %f558, %f555, %f555; fma.rn.ftz.f32 %f559, %f551, %f551, %f558; fma.rn.ftz.f32 %f560, %f557, %f557, %f559; sqrt.approx.ftz.f32 %f561, %f560; div.approx.ftz.f32 %f562, %f551, %f561; div.approx.ftz.f32 %f563, %f555, %f561; div.approx.ftz.f32 %f564, %f557, %f561; ld.local.f32 %f565, [%rd2+16]; ld.local.f32 %f566, [%rd2+-76]; fma.rn.ftz.f32 %f676, %f565, %f562, %f566; ld.local.f32 %f567, [%rd2+-72]; fma.rn.ftz.f32 %f675, %f565, %f563, %f567; ld.local.f32 %f568, [%rd2+-68]; fma.rn.ftz.f32 %f674, %f565, %f564, %f568; @%p98 bra $L__BB27_154; ld.local.u8 %rs35, [%rd2+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p102, %rs36, 0; @%p102 bra $L__BB27_153; bra.uni $L__BB27_152; $L__BB27_153: ld.const.u64 %rd283, [kernel_params+2160]; cvta.to.global.u64 %rd284, %rd283; mul.wide.s32 %rd285, %r3, 208; add.s64 %rd286, %rd284, %rd285; ld.global.v4.f32 {%f602, %f603, %f604, %f605}, [%rd286+48]; ld.global.v4.f32 {%f610, %f611, %f612, %f613}, [%rd286+64]; ld.global.v4.f32 {%f618, %f619, %f620, %f621}, [%rd286+80]; mul.ftz.f32 %f626, %f675, %f603; fma.rn.ftz.f32 %f627, %f676, %f602, %f626; fma.rn.ftz.f32 %f628, %f674, %f604, %f627; add.ftz.f32 %f176, %f605, %f628; mul.ftz.f32 %f629, %f675, %f611; fma.rn.ftz.f32 %f630, %f676, %f610, %f629; fma.rn.ftz.f32 %f631, %f674, %f612, %f630; add.ftz.f32 %f177, %f613, %f631; mul.ftz.f32 %f632, %f675, %f619; fma.rn.ftz.f32 %f633, %f676, %f618, %f632; fma.rn.ftz.f32 %f634, %f674, %f620, %f633; add.ftz.f32 %f674, %f621, %f634; mov.f32 %f675, %f177; mov.f32 %f676, %f176; bra.uni $L__BB27_154; $L__BB27_160: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB27_162; mov.u32 %r227, 1065353216; st.local.u32 [%rd17], %r227; bra.uni $L__BB27_164; $L__BB27_19: setp.eq.s32 %p21, %r243, 0; and.b32 %r40, %r31, 1; @%p21 bra $L__BB27_125; setp.eq.s32 %p22, %r243, 1; @%p22 bra $L__BB27_100; bra.uni $L__BB27_21; $L__BB27_100: setp.eq.s32 %p66, %r40, 0; mov.u32 %r50, 1; @%p66 bra $L__BB27_111; bra.uni $L__BB27_101; $L__BB27_111: and.b32 %r174, %r31, 6; setp.eq.s32 %p73, %r174, 0; @%p73 bra $L__BB27_118; and.b32 %r175, %r35, 768; setp.eq.s32 %p74, %r175, 0; @%p74 bra $L__BB27_114; ld.const.u64 %rd218, [kernel_params+2264]; cvta.to.global.u64 %rd219, %rd218; ld.local.u32 %r176, [%rd2+-16]; mul.wide.s32 %rd220, %r176, 16; add.s64 %rd221, %rd219, %rd220; shr.s32 %r177, %r31, 7; ld.global.u32 %r178, [%rd221+4]; add.s32 %r179, %r177, %r178; ld.const.u64 %rd222, [kernel_params+2328]; cvta.to.global.u64 %rd223, %rd222; add.s32 %r180, %r179, %r32; mul.wide.s32 %rd224, %r180, 8; add.s64 %rd225, %rd223, %rd224; ld.global.v2.f32 {%f481, %f482}, [%rd225]; ld.global.v2.f32 {%f485, %f486}, [%rd225+8]; sub.ftz.f32 %f489, %f485, %f481; sub.ftz.f32 %f490, %f486, %f482; ld.local.f32 %f491, [%rd2+24]; mul.ftz.f32 %f492, %f491, %f490; mul.ftz.f32 %f493, %f491, %f489; add.u64 %rd227, %SPL, 8; st.local.v2.f32 [%rd227], {%f493, %f492}; ld.local.f32 %f494, [%rd2+-8]; mov.f32 %f495, 0f3F800000; sub.ftz.f32 %f496, %f495, %f494; mul.ftz.f32 %f497, %f481, %f496; mul.ftz.f32 %f498, %f482, %f496; fma.rn.ftz.f32 %f672, %f486, %f494, %f498; fma.rn.ftz.f32 %f671, %f485, %f494, %f497; bra.uni $L__BB27_120; $L__BB27_162: mov.u32 %r228, 0; st.local.u32 [%rd17], %r228; st.local.u32 [%rd17+4], %r228; st.local.u32 [%rd17+8], %r228; bra.uni $L__BB27_164; $L__BB27_125: setp.eq.s32 %p82, %r40, 0; @%p82 bra $L__BB27_136; bra.uni $L__BB27_126; $L__BB27_136: and.b32 %r210, %r31, 6; setp.eq.s32 %p89, %r210, 0; @%p89 bra $L__BB27_143; and.b32 %r211, %r35, 768; setp.eq.s32 %p90, %r211, 0; @%p90 bra $L__BB27_139; ld.const.u64 %rd267, [kernel_params+2264]; cvta.to.global.u64 %rd268, %rd267; ld.local.u32 %r212, [%rd2+-16]; mul.wide.s32 %rd269, %r212, 16; add.s64 %rd270, %rd268, %rd269; shr.s32 %r213, %r31, 7; ld.global.u32 %r214, [%rd270+4]; add.s32 %r215, %r213, %r214; ld.const.u64 %rd271, [kernel_params+2320]; cvta.to.global.u64 %rd272, %rd271; add.s32 %r216, %r215, %r32; mul.wide.s32 %rd273, %r216, 4; add.s64 %rd274, %rd272, %rd273; ld.global.f32 %f533, [%rd274+4]; ld.global.f32 %f534, [%rd274]; sub.ftz.f32 %f535, %f533, %f534; ld.local.f32 %f536, [%rd2+24]; mul.ftz.f32 %f537, %f536, %f535; st.local.f32 [%rd1], %f537; ld.local.f32 %f538, [%rd2+-8]; mov.f32 %f539, 0f3F800000; sub.ftz.f32 %f540, %f539, %f538; mul.ftz.f32 %f541, %f534, %f540; fma.rn.ftz.f32 %f673, %f533, %f538, %f541; bra.uni $L__BB27_145; $L__BB27_152: ld.local.v4.f32 {%f569, %f570, %f571, %f572}, [%rd2+116]; mul.ftz.f32 %f577, %f675, %f570; fma.rn.ftz.f32 %f578, %f676, %f569, %f577; fma.rn.ftz.f32 %f579, %f674, %f571, %f578; add.ftz.f32 %f173, %f572, %f579; ld.local.v4.f32 {%f580, %f581, %f582, %f583}, [%rd2+132]; mul.ftz.f32 %f588, %f675, %f581; fma.rn.ftz.f32 %f589, %f676, %f580, %f588; fma.rn.ftz.f32 %f590, %f674, %f582, %f589; add.ftz.f32 %f174, %f583, %f590; ld.local.v4.f32 {%f591, %f592, %f593, %f594}, [%rd2+148]; mul.ftz.f32 %f599, %f675, %f592; fma.rn.ftz.f32 %f600, %f676, %f591, %f599; fma.rn.ftz.f32 %f601, %f674, %f593, %f600; add.ftz.f32 %f674, %f594, %f601; mov.f32 %f675, %f174; mov.f32 %f676, %f173; $L__BB27_154: setp.eq.s32 %p103, %r64, 1; @%p103 bra $L__BB27_158; bra.uni $L__BB27_155; $L__BB27_158: add.ftz.f32 %f635, %f675, %f676; add.ftz.f32 %f636, %f674, %f635; mul.ftz.f32 %f637, %f636, 0f3EAAAAAB; st.local.f32 [%rd17], %f637; bra.uni $L__BB27_164; $L__BB27_155: setp.eq.s32 %p104, %r64, 0; @%p104 bra $L__BB27_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB27_164; $L__BB27_21: add.s32 %r86, %r243, -3; setp.lt.u32 %p23, %r86, 2; @%p23 bra $L__BB27_47; bra.uni $L__BB27_22; $L__BB27_47: setp.eq.s32 %p40, %r40, 0; @%p40 bra $L__BB27_86; bra.uni $L__BB27_48; $L__BB27_86: and.b32 %r143, %r31, 6; setp.eq.s32 %p57, %r143, 0; @%p57 bra $L__BB27_93; and.b32 %r144, %r35, 768; setp.eq.s32 %p58, %r144, 0; @%p58 bra $L__BB27_89; ld.const.u64 %rd160, [kernel_params+2264]; cvta.to.global.u64 %rd161, %rd160; ld.local.u32 %r145, [%rd2+-16]; mul.wide.s32 %rd162, %r145, 16; add.s64 %rd163, %rd161, %rd162; shr.s32 %r146, %r31, 7; ld.global.u32 %r147, [%rd163+4]; add.s32 %r148, %r146, %r147; ld.const.u64 %rd164, [kernel_params+2344]; cvta.to.global.u64 %rd165, %rd164; add.s32 %r149, %r148, %r32; mul.wide.s32 %rd166, %r149, 16; add.s64 %rd167, %rd165, %rd166; ld.global.v4.f32 {%f372, %f373, %f374, %f375}, [%rd167]; ld.global.v4.f32 {%f380, %f381, %f382, %f383}, [%rd167+16]; sub.ftz.f32 %f388, %f380, %f372; sub.ftz.f32 %f389, %f381, %f373; sub.ftz.f32 %f390, %f382, %f374; sub.ftz.f32 %f391, %f383, %f375; ld.local.f32 %f392, [%rd2+24]; mul.ftz.f32 %f393, %f392, %f391; mul.ftz.f32 %f394, %f392, %f390; mul.ftz.f32 %f395, %f392, %f389; mul.ftz.f32 %f396, %f392, %f388; add.u64 %rd169, %SPL, 16; st.local.v4.f32 [%rd169], {%f396, %f395, %f394, %f393}; ld.local.f32 %f397, [%rd2+-8]; mov.f32 %f398, 0f3F800000; sub.ftz.f32 %f399, %f398, %f397; mul.ftz.f32 %f400, %f372, %f399; mul.ftz.f32 %f401, %f373, %f399; mul.ftz.f32 %f402, %f374, %f399; mul.ftz.f32 %f403, %f375, %f399; fma.rn.ftz.f32 %f664, %f383, %f397, %f403; fma.rn.ftz.f32 %f663, %f382, %f397, %f402; fma.rn.ftz.f32 %f662, %f381, %f397, %f401; fma.rn.ftz.f32 %f661, %f380, %f397, %f400; bra.uni $L__BB27_95; $L__BB27_126: ld.local.s32 %rd55, [%rd2+-16]; setp.eq.s64 %p83, %rd55, -1; @%p83 bra $L__BB27_129; ld.const.u64 %rd244, [kernel_params+2240]; cvta.to.global.u64 %rd245, %rd244; shl.b64 %rd246, %rd55, 2; add.s64 %rd247, %rd245, %rd246; ld.global.u32 %r188, [%rd247]; setp.eq.s32 %p84, %r188, -1; @%p84 bra $L__BB27_129; mov.u32 %r189, 0; mov.u64 %rd248, 0; { // callseq 29, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r189; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd67; .param .b64 param3; st.param.b64 [param3+0], %rd248; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f673, [retval0+0]; } // callseq 29 bra.uni $L__BB27_145; $L__BB27_157: st.local.f32 [%rd17], %f676; st.local.f32 [%rd17+4], %f675; st.local.f32 [%rd17+8], %f674; bra.uni $L__BB27_164; $L__BB27_101: ld.local.s32 %rd53, [%rd2+-16]; setp.eq.s64 %p67, %rd53, -1; @%p67 bra $L__BB27_104; ld.const.u64 %rd188, [kernel_params+2240]; cvta.to.global.u64 %rd189, %rd188; shl.b64 %rd190, %rd53, 2; add.s64 %rd191, %rd189, %rd190; ld.global.u32 %r155, [%rd191]; setp.eq.s32 %p68, %r155, -1; @%p68 bra $L__BB27_104; mov.u64 %rd192, 0; add.u64 %rd193, %SP, 8; { // callseq 28, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r50; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd193; .param .b64 param3; st.param.b64 [param3+0], %rd192; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f671, [retval0+0]; ld.param.f32 %f672, [retval0+4]; } // callseq 28 bra.uni $L__BB27_120; $L__BB27_22: setp.eq.s32 %p24, %r40, 0; add.u64 %rd86, %SP, 32; add.u64 %rd18, %SPL, 32; @%p24 bra $L__BB27_33; bra.uni $L__BB27_23; $L__BB27_33: and.b32 %r109, %r31, 6; setp.eq.s32 %p31, %r109, 0; @%p31 bra $L__BB27_40; and.b32 %r110, %r35, 768; setp.eq.s32 %p32, %r110, 0; @%p32 bra $L__BB27_36; ld.const.u64 %rd111, [kernel_params+2264]; cvta.to.global.u64 %rd112, %rd111; ld.local.u32 %r111, [%rd2+-16]; mul.wide.s32 %rd113, %r111, 16; add.s64 %rd114, %rd112, %rd113; shr.s32 %r112, %r31, 7; ld.global.u32 %r113, [%rd114+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd115, [kernel_params+2336]; cvta.to.global.u64 %rd116, %rd115; add.s32 %r115, %r114, %r32; mul.wide.s32 %rd117, %r115, 12; add.s64 %rd118, %rd116, %rd117; ld.global.f32 %f220, [%rd118+12]; ld.global.f32 %f221, [%rd118]; sub.ftz.f32 %f222, %f220, %f221; ld.global.f32 %f223, [%rd118+16]; ld.global.f32 %f224, [%rd118+4]; sub.ftz.f32 %f225, %f223, %f224; ld.global.f32 %f226, [%rd118+20]; ld.global.f32 %f227, [%rd118+8]; sub.ftz.f32 %f228, %f226, %f227; ld.local.f32 %f229, [%rd2+24]; mul.ftz.f32 %f230, %f222, %f229; mul.ftz.f32 %f231, %f225, %f229; mul.ftz.f32 %f232, %f228, %f229; st.local.f32 [%rd18], %f230; st.local.f32 [%rd18+4], %f231; st.local.f32 [%rd18+8], %f232; ld.local.f32 %f233, [%rd2+-8]; mov.f32 %f234, 0f3F800000; sub.ftz.f32 %f235, %f234, %f233; mul.ftz.f32 %f236, %f221, %f235; mul.ftz.f32 %f237, %f224, %f235; mul.ftz.f32 %f238, %f227, %f235; fma.rn.ftz.f32 %f640, %f226, %f233, %f238; fma.rn.ftz.f32 %f639, %f223, %f233, %f237; fma.rn.ftz.f32 %f638, %f220, %f233, %f236; bra.uni $L__BB27_42; $L__BB27_129: and.b32 %r190, %r35, 56; setp.eq.s32 %p85, %r190, 0; @%p85 bra $L__BB27_134; and.b32 %r191, %r35, 24; setp.eq.s32 %p86, %r191, 0; @%p86 bra $L__BB27_132; ld.const.u64 %rd250, [kernel_params+2232]; cvta.to.global.u64 %rd251, %rd250; shl.b64 %rd252, %rd55, 4; add.s64 %rd253, %rd251, %rd252; ld.global.v4.u32 {%r192, %r193, %r194, %r195}, [%rd253]; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r199, %r192, %r32; mul.wide.u32 %rd256, %r199, 4; add.s64 %rd299, %rd255, %rd256; add.s32 %r200, %r193, %r32; mul.wide.u32 %rd257, %r200, 4; add.s64 %rd300, %rd255, %rd257; add.s32 %r201, %r194, %r32; mul.wide.u32 %rd258, %r201, 4; add.s64 %rd301, %rd255, %rd258; bra.uni $L__BB27_133; $L__BB27_104: and.b32 %r156, %r35, 56; setp.eq.s32 %p69, %r156, 0; @%p69 bra $L__BB27_109; and.b32 %r157, %r35, 24; setp.eq.s32 %p70, %r157, 0; @%p70 bra $L__BB27_107; ld.const.u64 %rd194, [kernel_params+2232]; cvta.to.global.u64 %rd195, %rd194; shl.b64 %rd196, %rd53, 4; add.s64 %rd197, %rd195, %rd196; ld.global.v4.u32 {%r158, %r159, %r160, %r161}, [%rd197]; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r165, %r158, %r32; mul.wide.u32 %rd200, %r165, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f669, %f670}, [%rd201]; add.s32 %r166, %r159, %r32; mul.wide.u32 %rd202, %r166, 8; add.s64 %rd203, %rd199, %rd202; ld.global.v2.f32 {%f667, %f668}, [%rd203]; add.s32 %r167, %r160, %r32; mul.wide.u32 %rd204, %r167, 8; add.s64 %rd205, %rd199, %rd204; ld.global.v2.f32 {%f665, %f666}, [%rd205]; bra.uni $L__BB27_108; $L__BB27_48: ld.local.s32 %rd49, [%rd2+-16]; setp.eq.s64 %p41, %rd49, -1; @%p41 bra $L__BB27_51; ld.const.u64 %rd127, [kernel_params+2240]; cvta.to.global.u64 %rd128, %rd127; shl.b64 %rd129, %rd49, 2; add.s64 %rd130, %rd128, %rd129; ld.global.u32 %r124, [%rd130]; setp.eq.s32 %p42, %r124, -1; @%p42 bra $L__BB27_51; mov.u64 %rd131, 0; add.u64 %rd132, %SP, 16; { // callseq 27, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r243; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd132; .param .b64 param3; st.param.b64 [param3+0], %rd131; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f661, [retval0+0]; ld.param.f32 %f662, [retval0+4]; ld.param.f32 %f663, [retval0+8]; ld.param.f32 %f664, [retval0+12]; } // callseq 27 bra.uni $L__BB27_95; $L__BB27_23: ld.local.s32 %rd20, [%rd2+-16]; setp.eq.s64 %p25, %rd20, -1; @%p25 bra $L__BB27_26; ld.const.u64 %rd87, [kernel_params+2240]; cvta.to.global.u64 %rd88, %rd87; shl.b64 %rd89, %rd20, 2; add.s64 %rd90, %rd88, %rd89; ld.global.u32 %r87, [%rd90]; setp.eq.s32 %p26, %r87, -1; @%p26 bra $L__BB27_26; mov.u64 %rd91, 0; { // callseq 26, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r243; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd86; .param .b64 param3; st.param.b64 [param3+0], %rd91; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f638, [retval0+0]; ld.param.f32 %f639, [retval0+4]; ld.param.f32 %f640, [retval0+8]; } // callseq 26 bra.uni $L__BB27_42; $L__BB27_143: and.b32 %r220, %r31, 8; setp.eq.s32 %p93, %r220, 0; mov.u32 %r221, 0; mov.f32 %f673, 0f00000000; st.local.u32 [%rd1], %r221; setp.ne.s32 %p94, %r35, 8; or.pred %p95, %p94, %p93; @%p95 bra $L__BB27_145; ld.const.u64 %rd279, [kernel_params+2320]; cvta.to.global.u64 %rd280, %rd279; ld.local.u32 %r222, [%rd2+-16]; add.s32 %r223, %r222, %r32; mul.wide.s32 %rd281, %r223, 4; add.s64 %rd282, %rd280, %rd281; ld.global.f32 %f673, [%rd282]; bra.uni $L__BB27_145; $L__BB27_118: and.b32 %r183, %r31, 8; setp.eq.s32 %p77, %r183, 0; add.u64 %rd235, %SPL, 8; mov.f32 %f671, 0f00000000; st.local.v2.f32 [%rd235], {%f671, %f671}; setp.ne.s32 %p78, %r35, 8; or.pred %p79, %p78, %p77; mov.f32 %f672, %f671; @%p79 bra $L__BB27_120; ld.const.u64 %rd236, [kernel_params+2328]; cvta.to.global.u64 %rd237, %rd236; ld.local.u32 %r184, [%rd2+-16]; add.s32 %r185, %r184, %r32; mul.wide.s32 %rd238, %r185, 8; add.s64 %rd239, %rd237, %rd238; ld.global.v2.f32 {%f671, %f672}, [%rd239]; bra.uni $L__BB27_120; $L__BB27_51: and.b32 %r125, %r35, 120; setp.eq.s32 %p43, %r125, 0; @%p43 bra $L__BB27_84; and.b32 %r126, %r35, 24; setp.eq.s32 %p44, %r126, 0; @%p44 bra $L__BB27_54; ld.const.u64 %rd133, [kernel_params+2232]; cvta.to.global.u64 %rd134, %rd133; shl.b64 %rd135, %rd49, 4; add.s64 %rd136, %rd134, %rd135; ld.global.v4.u32 {%r127, %r128, %r129, %r130}, [%rd136]; ld.const.u64 %rd137, [kernel_params+2344]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r134, %r127, %r32; mul.wide.u32 %rd139, %r134, 16; add.s64 %rd140, %rd138, %rd139; ld.global.v4.f32 {%f657, %f658, %f659, %f660}, [%rd140]; add.s32 %r135, %r128, %r32; mul.wide.u32 %rd141, %r135, 16; add.s64 %rd142, %rd138, %rd141; ld.global.v4.f32 {%f653, %f654, %f655, %f656}, [%rd142]; add.s32 %r136, %r129, %r32; mul.wide.u32 %rd143, %r136, 16; add.s64 %rd144, %rd138, %rd143; ld.global.v4.f32 {%f649, %f650, %f651, %f652}, [%rd144]; bra.uni $L__BB27_83; $L__BB27_26: and.b32 %r88, %r35, 56; setp.eq.s32 %p27, %r88, 0; @%p27 bra $L__BB27_31; and.b32 %r89, %r35, 24; setp.eq.s32 %p28, %r89, 0; @%p28 bra $L__BB27_29; ld.const.u64 %rd93, [kernel_params+2232]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd20, 4; add.s64 %rd96, %rd94, %rd95; ld.global.v4.u32 {%r90, %r91, %r92, %r93}, [%rd96]; ld.const.u64 %rd97, [kernel_params+2336]; cvta.to.global.u64 %rd98, %rd97; add.s32 %r97, %r90, %r32; mul.wide.u32 %rd99, %r97, 12; add.s64 %rd296, %rd98, %rd99; add.s64 %rd297, %rd296, 4; add.s64 %rd298, %rd296, 8; add.s32 %r98, %r91, %r32; mul.wide.u32 %rd100, %r98, 12; add.s64 %rd293, %rd98, %rd100; add.s64 %rd294, %rd293, 4; add.s64 %rd295, %rd293, 8; add.s32 %r99, %r92, %r32; mul.wide.u32 %rd101, %r99, 12; add.s64 %rd290, %rd98, %rd101; add.s64 %rd291, %rd290, 4; add.s64 %rd292, %rd290, 8; bra.uni $L__BB27_30; $L__BB27_139: mov.f32 %f673, 0f00000000; mov.u32 %r217, 0; st.local.u32 [%rd1], %r217; and.b32 %r218, %r35, 131; setp.eq.s32 %p91, %r218, 0; @%p91 bra $L__BB27_145; setp.ne.s32 %p92, %r35, 128; @%p92 bra $L__BB27_142; ld.local.u32 %r219, [%rd2+-16]; add.s32 %r32, %r219, %r32; $L__BB27_142: ld.const.u64 %rd275, [kernel_params+2320]; cvta.to.global.u64 %rd276, %rd275; mul.wide.s32 %rd277, %r32, 4; add.s64 %rd278, %rd276, %rd277; ld.global.f32 %f673, [%rd278]; bra.uni $L__BB27_145; $L__BB27_134: mov.f32 %f673, 0f00000000; mov.u32 %r205, 0; st.local.u32 [%rd1], %r205; and.b32 %r206, %r35, 7; setp.eq.s32 %p87, %r206, 0; @%p87 bra $L__BB27_145; cvt.u32.u64 %r207, %rd55; setp.eq.s32 %p88, %r35, 4; selp.b32 %r208, %r207, 0, %p88; add.s32 %r209, %r32, %r208; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; mul.wide.s32 %rd265, %r209, 4; add.s64 %rd266, %rd264, %rd265; ld.global.f32 %f673, [%rd266]; bra.uni $L__BB27_145; $L__BB27_114: add.u64 %rd229, %SPL, 8; mov.f32 %f671, 0f00000000; st.local.v2.f32 [%rd229], {%f671, %f671}; and.b32 %r181, %r35, 131; setp.eq.s32 %p75, %r181, 0; mov.f32 %f672, %f671; @%p75 bra $L__BB27_120; setp.ne.s32 %p76, %r35, 128; @%p76 bra $L__BB27_117; ld.local.u32 %r182, [%rd2+-16]; add.s32 %r32, %r182, %r32; $L__BB27_117: ld.const.u64 %rd230, [kernel_params+2328]; cvta.to.global.u64 %rd231, %rd230; mul.wide.s32 %rd232, %r32, 8; add.s64 %rd233, %rd231, %rd232; ld.global.v2.f32 {%f671, %f672}, [%rd233]; bra.uni $L__BB27_120; $L__BB27_109: add.u64 %rd213, %SPL, 8; mov.f32 %f671, 0f00000000; st.local.v2.f32 [%rd213], {%f671, %f671}; and.b32 %r170, %r35, 7; setp.eq.s32 %p71, %r170, 0; mov.f32 %f672, %f671; @%p71 bra $L__BB27_120; cvt.u32.u64 %r171, %rd53; setp.eq.s32 %p72, %r35, 4; selp.b32 %r172, %r171, 0, %p72; add.s32 %r173, %r32, %r172; ld.const.u64 %rd214, [kernel_params+2328]; cvta.to.global.u64 %rd215, %rd214; mul.wide.s32 %rd216, %r173, 8; add.s64 %rd217, %rd215, %rd216; ld.global.v2.f32 {%f671, %f672}, [%rd217]; bra.uni $L__BB27_120; $L__BB27_93: and.b32 %r152, %r31, 8; setp.eq.s32 %p61, %r152, 0; add.u64 %rd177, %SPL, 16; mov.f32 %f661, 0f00000000; st.local.v4.f32 [%rd177], {%f661, %f661, %f661, %f661}; setp.ne.s32 %p62, %r35, 8; or.pred %p63, %p62, %p61; mov.f32 %f662, %f661; mov.f32 %f663, %f661; mov.f32 %f664, %f661; @%p63 bra $L__BB27_95; ld.const.u64 %rd178, [kernel_params+2344]; cvta.to.global.u64 %rd179, %rd178; ld.local.u32 %r153, [%rd2+-16]; add.s32 %r154, %r153, %r32; mul.wide.s32 %rd180, %r154, 16; add.s64 %rd181, %rd179, %rd180; ld.global.v4.f32 {%f661, %f662, %f663, %f664}, [%rd181]; bra.uni $L__BB27_95; $L__BB27_40: and.b32 %r119, %r31, 8; setp.eq.s32 %p35, %r119, 0; mov.u32 %r120, 0; st.local.u32 [%rd18], %r120; st.local.u32 [%rd18+4], %r120; st.local.u32 [%rd18+8], %r120; setp.ne.s32 %p36, %r35, 8; mov.f32 %f638, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f639, %f638; mov.f32 %f640, %f638; @%p37 bra $L__BB27_42; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r121, [%rd2+-16]; add.s32 %r122, %r121, %r32; mul.wide.s32 %rd125, %r122, 12; add.s64 %rd126, %rd124, %rd125; ld.global.f32 %f640, [%rd126+8]; ld.global.f32 %f639, [%rd126+4]; ld.global.f32 %f638, [%rd126]; bra.uni $L__BB27_42; $L__BB27_132: cvt.u32.u64 %r202, %rd55; mad.lo.s32 %r203, %r202, 3, %r32; ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r203, 4; add.s64 %rd299, %rd260, %rd261; add.s64 %rd300, %rd299, 4; add.s32 %r204, %r203, 2; mul.wide.s32 %rd262, %r204, 4; add.s64 %rd301, %rd260, %rd262; $L__BB27_133: ld.local.f32 %f515, [%rd2+24]; ld.global.f32 %f516, [%rd300]; ld.local.f32 %f517, [%rd2+32]; ld.global.f32 %f518, [%rd301]; mul.ftz.f32 %f519, %f518, %f517; fma.rn.ftz.f32 %f520, %f516, %f515, %f519; add.ftz.f32 %f521, %f515, %f517; ld.global.f32 %f522, [%rd299]; mul.ftz.f32 %f523, %f522, %f521; sub.ftz.f32 %f524, %f520, %f523; st.local.f32 [%rd1], %f524; ld.local.f32 %f525, [%rd2+-8]; ld.local.f32 %f526, [%rd2+-4]; mul.ftz.f32 %f527, %f518, %f526; fma.rn.ftz.f32 %f528, %f516, %f525, %f527; mov.f32 %f529, 0f3F800000; sub.ftz.f32 %f530, %f529, %f525; sub.ftz.f32 %f531, %f530, %f526; fma.rn.ftz.f32 %f673, %f522, %f531, %f528; $L__BB27_145: setp.eq.s32 %p96, %r64, 1; @%p96 bra $L__BB27_149; bra.uni $L__BB27_146; $L__BB27_149: ld.local.f32 %f546, [%rd1]; add.ftz.f32 %f547, %f673, %f546; st.local.f32 [%rd17], %f547; bra.uni $L__BB27_164; $L__BB27_146: setp.eq.s32 %p97, %r64, 0; @%p97 bra $L__BB27_148; mov.u32 %r224, 1065353216; st.local.u32 [%rd17], %r224; bra.uni $L__BB27_164; $L__BB27_148: ld.local.f32 %f544, [%rd1]; add.ftz.f32 %f545, %f673, %f544; st.local.f32 [%rd17], %f545; st.local.f32 [%rd17+4], %f545; st.local.f32 [%rd17+8], %f545; bra.uni $L__BB27_164; $L__BB27_107: cvt.u32.u64 %r168, %rd53; mad.lo.s32 %r169, %r168, 3, %r32; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; mul.wide.s32 %rd208, %r169, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f669, %f670}, [%rd209]; ld.global.v2.f32 {%f667, %f668}, [%rd209+8]; ld.global.v2.f32 {%f665, %f666}, [%rd209+16]; $L__BB27_108: ld.local.f32 %f457, [%rd2+24]; ld.local.f32 %f458, [%rd2+32]; mul.ftz.f32 %f459, %f665, %f458; mul.ftz.f32 %f460, %f666, %f458; fma.rn.ftz.f32 %f461, %f667, %f457, %f459; fma.rn.ftz.f32 %f462, %f668, %f457, %f460; add.ftz.f32 %f463, %f458, %f457; mul.ftz.f32 %f464, %f669, %f463; mul.ftz.f32 %f465, %f670, %f463; sub.ftz.f32 %f466, %f462, %f465; sub.ftz.f32 %f467, %f461, %f464; add.u64 %rd211, %SPL, 8; st.local.v2.f32 [%rd211], {%f467, %f466}; ld.local.f32 %f468, [%rd2+-8]; ld.local.f32 %f469, [%rd2+-4]; mul.ftz.f32 %f470, %f665, %f469; mul.ftz.f32 %f471, %f666, %f469; fma.rn.ftz.f32 %f472, %f667, %f468, %f470; fma.rn.ftz.f32 %f473, %f668, %f468, %f471; mov.f32 %f474, 0f3F800000; sub.ftz.f32 %f475, %f474, %f468; sub.ftz.f32 %f476, %f475, %f469; fma.rn.ftz.f32 %f672, %f670, %f476, %f473; fma.rn.ftz.f32 %f671, %f669, %f476, %f472; $L__BB27_120: setp.eq.s32 %p80, %r64, 1; @%p80 bra $L__BB27_124; bra.uni $L__BB27_121; $L__BB27_124: add.u64 %rd243, %SPL, 8; ld.local.f32 %f513, [%rd243]; add.ftz.f32 %f514, %f671, %f513; st.local.f32 [%rd17], %f514; bra.uni $L__BB27_164; $L__BB27_121: setp.eq.s32 %p81, %r64, 0; @%p81 bra $L__BB27_123; mov.u32 %r186, 1065353216; st.local.u32 [%rd17], %r186; bra.uni $L__BB27_164; $L__BB27_123: add.u64 %rd241, %SPL, 8; ld.local.v2.f32 {%f507, %f508}, [%rd241]; add.ftz.f32 %f511, %f671, %f507; add.ftz.f32 %f512, %f672, %f508; st.local.f32 [%rd17], %f511; st.local.f32 [%rd17+4], %f512; mov.u32 %r187, 0; st.local.u32 [%rd17+8], %r187; bra.uni $L__BB27_164; $L__BB27_89: add.u64 %rd171, %SPL, 16; mov.f32 %f661, 0f00000000; st.local.v4.f32 [%rd171], {%f661, %f661, %f661, %f661}; and.b32 %r150, %r35, 131; setp.eq.s32 %p59, %r150, 0; mov.f32 %f662, %f661; mov.f32 %f663, %f661; mov.f32 %f664, %f661; @%p59 bra $L__BB27_95; setp.ne.s32 %p60, %r35, 128; @%p60 bra $L__BB27_92; ld.local.u32 %r151, [%rd2+-16]; add.s32 %r32, %r151, %r32; $L__BB27_92: ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; mul.wide.s32 %rd174, %r32, 16; add.s64 %rd175, %rd173, %rd174; ld.global.v4.f32 {%f661, %f662, %f663, %f664}, [%rd175]; bra.uni $L__BB27_95; $L__BB27_36: mov.u32 %r116, 0; st.local.u32 [%rd18], %r116; st.local.u32 [%rd18+4], %r116; st.local.u32 [%rd18+8], %r116; and.b32 %r117, %r35, 131; setp.eq.s32 %p33, %r117, 0; mov.f32 %f638, 0f00000000; mov.f32 %f639, %f638; mov.f32 %f640, %f638; @%p33 bra $L__BB27_42; setp.ne.s32 %p34, %r35, 128; @%p34 bra $L__BB27_39; ld.local.u32 %r118, [%rd2+-16]; add.s32 %r32, %r118, %r32; $L__BB27_39: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r32, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f640, [%rd122+8]; ld.global.f32 %f639, [%rd122+4]; ld.global.f32 %f638, [%rd122]; bra.uni $L__BB27_42; $L__BB27_84: add.u64 %rd155, %SPL, 16; mov.f32 %f661, 0f00000000; st.local.v4.f32 [%rd155], {%f661, %f661, %f661, %f661}; and.b32 %r139, %r35, 7; setp.eq.s32 %p55, %r139, 0; mov.f32 %f662, %f661; mov.f32 %f663, %f661; mov.f32 %f664, %f661; @%p55 bra $L__BB27_95; cvt.u32.u64 %r140, %rd49; setp.eq.s32 %p56, %r35, 4; selp.b32 %r141, %r140, 0, %p56; add.s32 %r142, %r32, %r141; ld.const.u64 %rd156, [kernel_params+2344]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r142, 16; add.s64 %rd159, %rd157, %rd158; ld.global.v4.f32 {%f661, %f662, %f663, %f664}, [%rd159]; bra.uni $L__BB27_95; $L__BB27_31: mov.u32 %r104, 0; st.local.u32 [%rd18], %r104; st.local.u32 [%rd18+4], %r104; st.local.u32 [%rd18+8], %r104; and.b32 %r105, %r35, 7; setp.eq.s32 %p29, %r105, 0; mov.f32 %f638, 0f00000000; mov.f32 %f639, %f638; mov.f32 %f640, %f638; @%p29 bra $L__BB27_42; cvt.u32.u64 %r106, %rd20; setp.eq.s32 %p30, %r35, 4; selp.b32 %r107, %r106, 0, %p30; add.s32 %r108, %r32, %r107; ld.const.u64 %rd107, [kernel_params+2336]; cvta.to.global.u64 %rd108, %rd107; mul.wide.s32 %rd109, %r108, 12; add.s64 %rd110, %rd108, %rd109; ld.global.f32 %f640, [%rd110+8]; ld.global.f32 %f639, [%rd110+4]; ld.global.f32 %f638, [%rd110]; bra.uni $L__BB27_42; $L__BB27_54: cvt.u32.u64 %r137, %rd49; mad.lo.s32 %r138, %r137, 3, %r32; cvt.s64.s32 %rd50, %r138; setp.eq.s32 %p45, %r35, 32; @%p45 bra $L__BB27_82; bra.uni $L__BB27_55; $L__BB27_82: ld.const.u64 %rd148, [kernel_params+2344]; cvta.to.global.u64 %rd149, %rd148; shl.b64 %rd150, %rd50, 4; add.s64 %rd151, %rd149, %rd150; ld.global.v4.f32 {%f657, %f658, %f659, %f660}, [%rd151]; ld.global.v4.f32 {%f653, %f654, %f655, %f656}, [%rd151+16]; ld.global.v4.f32 {%f649, %f650, %f651, %f652}, [%rd151+32]; bra.uni $L__BB27_83; $L__BB27_29: cvt.u32.u64 %r100, %rd20; mad.lo.s32 %r101, %r100, 3, %r32; ld.const.u64 %rd102, [kernel_params+2336]; cvta.to.global.u64 %rd103, %rd102; mul.wide.s32 %rd104, %r101, 12; add.s64 %rd296, %rd103, %rd104; add.s64 %rd297, %rd296, 4; add.s64 %rd298, %rd296, 8; add.s32 %r102, %r101, 1; mul.wide.s32 %rd105, %r102, 12; add.s64 %rd293, %rd103, %rd105; add.s64 %rd294, %rd293, 4; add.s64 %rd295, %rd293, 8; add.s32 %r103, %r101, 2; mul.wide.s32 %rd106, %r103, 12; add.s64 %rd290, %rd103, %rd106; add.s64 %rd291, %rd290, 4; add.s64 %rd292, %rd290, 8; $L__BB27_30: ld.local.f32 %f182, [%rd2+24]; ld.global.f32 %f183, [%rd293]; ld.global.f32 %f184, [%rd294]; ld.global.f32 %f185, [%rd295]; ld.local.f32 %f186, [%rd2+32]; ld.global.f32 %f187, [%rd290]; mul.ftz.f32 %f188, %f187, %f186; ld.global.f32 %f189, [%rd291]; mul.ftz.f32 %f190, %f189, %f186; ld.global.f32 %f191, [%rd292]; mul.ftz.f32 %f192, %f191, %f186; fma.rn.ftz.f32 %f193, %f183, %f182, %f188; fma.rn.ftz.f32 %f194, %f184, %f182, %f190; fma.rn.ftz.f32 %f195, %f185, %f182, %f192; add.ftz.f32 %f196, %f182, %f186; ld.global.f32 %f197, [%rd296]; mul.ftz.f32 %f198, %f197, %f196; ld.global.f32 %f199, [%rd297]; mul.ftz.f32 %f200, %f199, %f196; ld.global.f32 %f201, [%rd298]; mul.ftz.f32 %f202, %f201, %f196; sub.ftz.f32 %f203, %f193, %f198; sub.ftz.f32 %f204, %f194, %f200; sub.ftz.f32 %f205, %f195, %f202; st.local.f32 [%rd18], %f203; st.local.f32 [%rd18+4], %f204; st.local.f32 [%rd18+8], %f205; ld.local.f32 %f206, [%rd2+-8]; ld.local.f32 %f207, [%rd2+-4]; mul.ftz.f32 %f208, %f187, %f207; mul.ftz.f32 %f209, %f189, %f207; mul.ftz.f32 %f210, %f191, %f207; fma.rn.ftz.f32 %f211, %f183, %f206, %f208; fma.rn.ftz.f32 %f212, %f184, %f206, %f209; fma.rn.ftz.f32 %f213, %f185, %f206, %f210; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f206; sub.ftz.f32 %f216, %f215, %f207; fma.rn.ftz.f32 %f640, %f201, %f216, %f213; fma.rn.ftz.f32 %f639, %f199, %f216, %f212; fma.rn.ftz.f32 %f638, %f197, %f216, %f211; $L__BB27_42: setp.eq.s32 %p38, %r64, 1; @%p38 bra $L__BB27_46; bra.uni $L__BB27_43; $L__BB27_46: ld.local.f32 %f251, [%rd18]; add.ftz.f32 %f252, %f638, %f251; ld.local.f32 %f253, [%rd18+4]; add.ftz.f32 %f254, %f639, %f253; ld.local.f32 %f255, [%rd18+8]; add.ftz.f32 %f256, %f640, %f255; add.ftz.f32 %f257, %f252, %f254; add.ftz.f32 %f258, %f256, %f257; mul.ftz.f32 %f259, %f258, 0f3EAAAAAB; st.local.f32 [%rd17], %f259; bra.uni $L__BB27_164; $L__BB27_43: setp.eq.s32 %p39, %r64, 0; @%p39 bra $L__BB27_45; mov.u32 %r123, 1065353216; st.local.u32 [%rd17], %r123; bra.uni $L__BB27_164; $L__BB27_45: ld.local.f32 %f245, [%rd18]; add.ftz.f32 %f246, %f638, %f245; ld.local.f32 %f247, [%rd18+4]; add.ftz.f32 %f248, %f639, %f247; ld.local.f32 %f249, [%rd18+8]; add.ftz.f32 %f250, %f640, %f249; st.local.f32 [%rd17], %f246; st.local.f32 [%rd17+4], %f248; st.local.f32 [%rd17+8], %f250; bra.uni $L__BB27_164; $L__BB27_55: ld.const.u64 %rd145, [kernel_params+2352]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd50, 2; add.s64 %rd51, %rd146, %rd147; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd51]; cvt.rn.f32.u16 %f272, %rs20; mul.ftz.f32 %f41, %f272, 0f3B808081; cvt.rn.f32.u16 %f273, %rs21; mul.ftz.f32 %f42, %f273, 0f3B808081; cvt.rn.f32.u16 %f274, %rs22; mul.ftz.f32 %f43, %f274, 0f3B808081; setp.lt.ftz.f32 %p46, %f41, 0f3D25AEE6; @%p46 bra $L__BB27_57; bra.uni $L__BB27_56; $L__BB27_57: mul.ftz.f32 %f657, %f41, 0f3D9E8391; bra.uni $L__BB27_58; $L__BB27_56: add.ftz.f32 %f275, %f41, 0f3D6147AE; mul.ftz.f32 %f276, %f275, 0f3F72A76F; lg2.approx.ftz.f32 %f277, %f276; mul.ftz.f32 %f278, %f277, 0f4019999A; ex2.approx.ftz.f32 %f657, %f278; $L__BB27_58: setp.lt.ftz.f32 %p47, %f42, 0f3D25AEE6; @%p47 bra $L__BB27_60; bra.uni $L__BB27_59; $L__BB27_60: mul.ftz.f32 %f658, %f42, 0f3D9E8391; bra.uni $L__BB27_61; $L__BB27_59: add.ftz.f32 %f279, %f42, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f658, %f282; $L__BB27_61: setp.lt.ftz.f32 %p48, %f43, 0f3D25AEE6; @%p48 bra $L__BB27_63; bra.uni $L__BB27_62; $L__BB27_63: mul.ftz.f32 %f659, %f43, 0f3D9E8391; bra.uni $L__BB27_64; $L__BB27_62: add.ftz.f32 %f283, %f43, 0f3D6147AE; mul.ftz.f32 %f284, %f283, 0f3F72A76F; lg2.approx.ftz.f32 %f285, %f284; mul.ftz.f32 %f286, %f285, 0f4019999A; ex2.approx.ftz.f32 %f659, %f286; $L__BB27_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd51+4]; cvt.rn.f32.u16 %f287, %rs24; mul.ftz.f32 %f53, %f287, 0f3B808081; cvt.rn.f32.u16 %f288, %rs25; mul.ftz.f32 %f54, %f288, 0f3B808081; cvt.rn.f32.u16 %f289, %rs26; mul.ftz.f32 %f55, %f289, 0f3B808081; setp.lt.ftz.f32 %p49, %f53, 0f3D25AEE6; @%p49 bra $L__BB27_66; bra.uni $L__BB27_65; $L__BB27_66: mul.ftz.f32 %f653, %f53, 0f3D9E8391; bra.uni $L__BB27_67; $L__BB27_65: add.ftz.f32 %f290, %f53, 0f3D6147AE; mul.ftz.f32 %f291, %f290, 0f3F72A76F; lg2.approx.ftz.f32 %f292, %f291; mul.ftz.f32 %f293, %f292, 0f4019999A; ex2.approx.ftz.f32 %f653, %f293; $L__BB27_67: setp.lt.ftz.f32 %p50, %f54, 0f3D25AEE6; @%p50 bra $L__BB27_69; bra.uni $L__BB27_68; $L__BB27_69: mul.ftz.f32 %f654, %f54, 0f3D9E8391; bra.uni $L__BB27_70; $L__BB27_68: add.ftz.f32 %f294, %f54, 0f3D6147AE; mul.ftz.f32 %f295, %f294, 0f3F72A76F; lg2.approx.ftz.f32 %f296, %f295; mul.ftz.f32 %f297, %f296, 0f4019999A; ex2.approx.ftz.f32 %f654, %f297; $L__BB27_70: setp.lt.ftz.f32 %p51, %f55, 0f3D25AEE6; @%p51 bra $L__BB27_72; bra.uni $L__BB27_71; $L__BB27_72: mul.ftz.f32 %f655, %f55, 0f3D9E8391; bra.uni $L__BB27_73; $L__BB27_71: add.ftz.f32 %f298, %f55, 0f3D6147AE; mul.ftz.f32 %f299, %f298, 0f3F72A76F; lg2.approx.ftz.f32 %f300, %f299; mul.ftz.f32 %f301, %f300, 0f4019999A; ex2.approx.ftz.f32 %f655, %f301; $L__BB27_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd51+8]; cvt.rn.f32.u16 %f302, %rs28; mul.ftz.f32 %f65, %f302, 0f3B808081; cvt.rn.f32.u16 %f303, %rs29; mul.ftz.f32 %f66, %f303, 0f3B808081; cvt.rn.f32.u16 %f304, %rs30; mul.ftz.f32 %f67, %f304, 0f3B808081; setp.lt.ftz.f32 %p52, %f65, 0f3D25AEE6; @%p52 bra $L__BB27_75; bra.uni $L__BB27_74; $L__BB27_75: mul.ftz.f32 %f649, %f65, 0f3D9E8391; bra.uni $L__BB27_76; $L__BB27_74: add.ftz.f32 %f305, %f65, 0f3D6147AE; mul.ftz.f32 %f306, %f305, 0f3F72A76F; lg2.approx.ftz.f32 %f307, %f306; mul.ftz.f32 %f308, %f307, 0f4019999A; ex2.approx.ftz.f32 %f649, %f308; $L__BB27_76: setp.lt.ftz.f32 %p53, %f66, 0f3D25AEE6; @%p53 bra $L__BB27_78; bra.uni $L__BB27_77; $L__BB27_78: mul.ftz.f32 %f650, %f66, 0f3D9E8391; bra.uni $L__BB27_79; $L__BB27_77: add.ftz.f32 %f309, %f66, 0f3D6147AE; mul.ftz.f32 %f310, %f309, 0f3F72A76F; lg2.approx.ftz.f32 %f311, %f310; mul.ftz.f32 %f312, %f311, 0f4019999A; ex2.approx.ftz.f32 %f650, %f312; $L__BB27_79: cvt.rn.f32.u16 %f313, %rs23; mul.ftz.f32 %f660, %f313, 0f3B808081; cvt.rn.f32.u16 %f314, %rs27; mul.ftz.f32 %f656, %f314, 0f3B808081; cvt.rn.f32.u16 %f315, %rs31; mul.ftz.f32 %f652, %f315, 0f3B808081; setp.lt.ftz.f32 %p54, %f67, 0f3D25AEE6; @%p54 bra $L__BB27_81; bra.uni $L__BB27_80; $L__BB27_81: mul.ftz.f32 %f651, %f67, 0f3D9E8391; bra.uni $L__BB27_83; $L__BB27_80: add.ftz.f32 %f316, %f67, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f651, %f319; $L__BB27_83: ld.local.f32 %f332, [%rd2+24]; ld.local.f32 %f333, [%rd2+32]; mul.ftz.f32 %f334, %f649, %f333; mul.ftz.f32 %f335, %f650, %f333; mul.ftz.f32 %f336, %f651, %f333; mul.ftz.f32 %f337, %f652, %f333; fma.rn.ftz.f32 %f338, %f653, %f332, %f334; fma.rn.ftz.f32 %f339, %f654, %f332, %f335; fma.rn.ftz.f32 %f340, %f655, %f332, %f336; fma.rn.ftz.f32 %f341, %f656, %f332, %f337; add.ftz.f32 %f342, %f332, %f333; mul.ftz.f32 %f343, %f657, %f342; mul.ftz.f32 %f344, %f658, %f342; mul.ftz.f32 %f345, %f659, %f342; mul.ftz.f32 %f346, %f660, %f342; sub.ftz.f32 %f347, %f341, %f346; sub.ftz.f32 %f348, %f340, %f345; sub.ftz.f32 %f349, %f339, %f344; sub.ftz.f32 %f350, %f338, %f343; add.u64 %rd153, %SPL, 16; st.local.v4.f32 [%rd153], {%f350, %f349, %f348, %f347}; ld.local.f32 %f351, [%rd2+-8]; ld.local.f32 %f352, [%rd2+-4]; mul.ftz.f32 %f353, %f649, %f352; mul.ftz.f32 %f354, %f650, %f352; mul.ftz.f32 %f355, %f651, %f352; mul.ftz.f32 %f356, %f652, %f352; fma.rn.ftz.f32 %f357, %f653, %f351, %f353; fma.rn.ftz.f32 %f358, %f654, %f351, %f354; fma.rn.ftz.f32 %f359, %f655, %f351, %f355; fma.rn.ftz.f32 %f360, %f656, %f351, %f356; mov.f32 %f361, 0f3F800000; sub.ftz.f32 %f362, %f361, %f351; sub.ftz.f32 %f363, %f362, %f352; fma.rn.ftz.f32 %f664, %f660, %f363, %f360; fma.rn.ftz.f32 %f663, %f659, %f363, %f359; fma.rn.ftz.f32 %f662, %f658, %f363, %f358; fma.rn.ftz.f32 %f661, %f657, %f363, %f357; $L__BB27_95: setp.eq.s32 %p64, %r64, 1; @%p64 bra $L__BB27_99; bra.uni $L__BB27_96; $L__BB27_99: add.u64 %rd187, %SPL, 16; ld.local.v4.f32 {%f432, %f433, %f434, %f435}, [%rd187]; add.ftz.f32 %f439, %f661, %f432; add.ftz.f32 %f440, %f662, %f433; add.ftz.f32 %f441, %f663, %f434; add.ftz.f32 %f442, %f439, %f440; add.ftz.f32 %f443, %f441, %f442; mul.ftz.f32 %f444, %f443, 0f3EAAAAAB; st.local.f32 [%rd17], %f444; bra.uni $L__BB27_164; $L__BB27_96: setp.eq.s32 %p65, %r64, 0; @%p65 bra $L__BB27_98; add.u64 %rd183, %SPL, 16; ld.local.f32 %f420, [%rd183+12]; add.ftz.f32 %f421, %f664, %f420; st.local.f32 [%rd17], %f421; bra.uni $L__BB27_164; $L__BB27_98: add.u64 %rd185, %SPL, 16; ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd185]; add.ftz.f32 %f429, %f661, %f422; add.ftz.f32 %f430, %f662, %f423; add.ftz.f32 %f431, %f663, %f424; st.local.f32 [%rd17], %f429; st.local.f32 [%rd17+4], %f430; st.local.f32 [%rd17+8], %f431; $L__BB27_164: ret; } .func _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .local .align 16 .b8 __local_depot28[48]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<664>; .reg .b32 %r<251>; .reg .b64 %rd<302>; mov.u64 %SPL, __local_depot28; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd65, [_Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd66, [_Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; add.u64 %rd67, %SP, 0; add.u64 %rd1, %SPL, 0; cvta.to.local.u64 %rd68, %rd65; add.s64 %rd2, %rd68, 76; ld.local.u32 %r3, [%rd68+76]; cvt.s64.s32 %rd3, %r3; setp.eq.s32 %p1, %r3, -1; @%p1 bra $L__BB28_16; cvt.u64.u32 %rd4, %r62; ld.local.s32 %rd5, [%rd2+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r4, [%rd72+164]; ld.local.u32 %r31, [%rd2+-12]; and.b32 %r66, %r31, 1; setp.eq.b32 %p2, %r66, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r233, 0; @%p5 bra $L__BB28_5; cvt.u32.u64 %r68, %rd5; setp.eq.s32 %p6, %r68, -1; mov.u32 %r232, -1; @%p6 bra $L__BB28_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r232, [%rd76]; $L__BB28_4: setp.ne.s32 %p7, %r232, -1; selp.u32 %r233, 1, 0, %p7; $L__BB28_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r235, %r233, %r4; mul.wide.u32 %rd78, %r235, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd288, %rd79, 14; ld.global.u32 %r234, [%rd79+8]; ld.global.u16 %rs37, [%rd79+12]; ld.global.u64 %rd287, [%rd79]; setp.eq.s64 %p8, %rd287, %rd4; add.s64 %rd289, %rd79, 15; @%p8 bra $L__BB28_11; $L__BB28_6: setp.eq.s64 %p9, %rd287, 0; @%p9 bra $L__BB28_8; add.s32 %r235, %r235, 2; bra.uni $L__BB28_9; $L__BB28_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r241, -1; mov.u32 %r238, 0; mov.u32 %r235, %r234; mov.u32 %r239, %r238; mov.u32 %r240, %r238; @%p10 bra $L__BB28_15; $L__BB28_9: cvt.u64.u32 %rd11, %r235; mul.wide.u32 %rd80, %r235, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r234, [%rd81+8]; ld.global.u16 %rs37, [%rd81+12]; ld.global.u64 %rd287, [%rd81]; setp.ne.s64 %p11, %rd287, %rd4; @%p11 bra $L__BB28_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd288, %rd83, 14; add.s64 %rd289, %rd83, 15; $L__BB28_11: ld.global.u8 %rs5, [%rd289]; ld.global.u8 %rs6, [%rd288]; cvt.u32.u64 %r73, %rd5; setp.ne.s32 %p12, %r73, -1; @%p12 bra $L__BB28_14; mov.u32 %r241, -1; mov.u32 %r238, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB28_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r239, %r238; mov.u32 %r240, %r238; @%p14 bra $L__BB28_15; $L__BB28_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r241, -1, %r234, %p15; cvt.u32.u16 %r238, %rs37; cvt.u32.u16 %r78, %rs5; and.b32 %r240, %r78, 255; cvt.u32.u16 %r79, %rs6; and.b32 %r239, %r79, 255; $L__BB28_15: setp.eq.s32 %p16, %r241, -1; selp.b32 %r32, 0, %r241, %p16; selp.b32 %r33, 0, %r240, %p16; selp.b32 %r245, %r64, %r239, %p16; selp.b32 %r35, 0, %r238, %p16; bra.uni $L__BB28_17; $L__BB28_16: ld.local.u32 %r31, [%rd2+-12]; mov.u32 %r32, 0; mov.u32 %r33, %r32; mov.u32 %r245, %r64; mov.u32 %r35, %r32; $L__BB28_17: cvta.to.local.u64 %rd84, %rd66; mul.wide.u32 %rd85, %r63, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r31, 16; @%p17 bra $L__BB28_159; bra.uni $L__BB28_18; $L__BB28_159: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB28_163; bra.uni $L__BB28_160; $L__BB28_163: mov.u32 %r231, 0; st.local.u32 [%rd17], %r231; bra.uni $L__BB28_164; $L__BB28_18: setp.eq.s32 %p18, %r62, 7; setp.eq.s32 %p19, %r35, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB28_150; bra.uni $L__BB28_19; $L__BB28_150: cvt.u32.u64 %r227, %rd3; setp.eq.s32 %p98, %r227, -1; ld.local.f32 %f526, [%rd2+-48]; ld.local.f32 %f527, [%rd2+-52]; setp.neu.ftz.f32 %p99, %f527, %f526; ld.local.f32 %f528, [%rd2+-44]; setp.neu.ftz.f32 %p100, %f527, %f528; or.pred %p101, %p99, %p100; sub.ftz.f32 %f529, %f528, %f526; neg.ftz.f32 %f530, %f526; neg.ftz.f32 %f531, %f528; selp.f32 %f532, %f531, %f528, %p101; add.ftz.f32 %f533, %f527, %f532; selp.f32 %f534, %f526, %f530, %p101; sub.ftz.f32 %f535, %f534, %f527; mul.ftz.f32 %f536, %f533, %f533; fma.rn.ftz.f32 %f537, %f529, %f529, %f536; fma.rn.ftz.f32 %f538, %f535, %f535, %f537; sqrt.approx.ftz.f32 %f539, %f538; div.approx.ftz.f32 %f540, %f529, %f539; div.approx.ftz.f32 %f541, %f533, %f539; div.approx.ftz.f32 %f542, %f535, %f539; mul.ftz.f32 %f543, %f526, %f542; mul.ftz.f32 %f544, %f528, %f541; sub.ftz.f32 %f545, %f543, %f544; mul.ftz.f32 %f546, %f528, %f540; mul.ftz.f32 %f547, %f527, %f542; sub.ftz.f32 %f548, %f546, %f547; mul.ftz.f32 %f549, %f527, %f541; mul.ftz.f32 %f550, %f526, %f540; sub.ftz.f32 %f551, %f549, %f550; ld.local.f32 %f552, [%rd2+16]; ld.local.f32 %f553, [%rd2+-76]; fma.rn.ftz.f32 %f663, %f552, %f545, %f553; ld.local.f32 %f554, [%rd2+-72]; fma.rn.ftz.f32 %f662, %f552, %f548, %f554; ld.local.f32 %f555, [%rd2+-68]; fma.rn.ftz.f32 %f661, %f552, %f551, %f555; @%p98 bra $L__BB28_154; ld.local.u8 %rs35, [%rd2+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p102, %rs36, 0; @%p102 bra $L__BB28_153; bra.uni $L__BB28_152; $L__BB28_153: ld.const.u64 %rd283, [kernel_params+2160]; cvta.to.global.u64 %rd284, %rd283; mul.wide.s32 %rd285, %r3, 208; add.s64 %rd286, %rd284, %rd285; ld.global.v4.f32 {%f589, %f590, %f591, %f592}, [%rd286+48]; ld.global.v4.f32 {%f597, %f598, %f599, %f600}, [%rd286+64]; ld.global.v4.f32 {%f605, %f606, %f607, %f608}, [%rd286+80]; mul.ftz.f32 %f613, %f662, %f590; fma.rn.ftz.f32 %f614, %f663, %f589, %f613; fma.rn.ftz.f32 %f615, %f661, %f591, %f614; add.ftz.f32 %f176, %f592, %f615; mul.ftz.f32 %f616, %f662, %f598; fma.rn.ftz.f32 %f617, %f663, %f597, %f616; fma.rn.ftz.f32 %f618, %f661, %f599, %f617; add.ftz.f32 %f177, %f600, %f618; mul.ftz.f32 %f619, %f662, %f606; fma.rn.ftz.f32 %f620, %f663, %f605, %f619; fma.rn.ftz.f32 %f621, %f661, %f607, %f620; add.ftz.f32 %f661, %f608, %f621; mov.f32 %f662, %f177; mov.f32 %f663, %f176; bra.uni $L__BB28_154; $L__BB28_160: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB28_162; mov.u32 %r229, 1065353216; st.local.u32 [%rd17], %r229; bra.uni $L__BB28_164; $L__BB28_19: setp.eq.s32 %p21, %r245, 0; and.b32 %r40, %r31, 1; @%p21 bra $L__BB28_125; setp.eq.s32 %p22, %r245, 1; @%p22 bra $L__BB28_100; bra.uni $L__BB28_21; $L__BB28_100: setp.eq.s32 %p66, %r40, 0; mov.u32 %r50, 1; @%p66 bra $L__BB28_111; bra.uni $L__BB28_101; $L__BB28_111: and.b32 %r175, %r31, 6; setp.eq.s32 %p73, %r175, 0; @%p73 bra $L__BB28_118; and.b32 %r176, %r35, 768; setp.eq.s32 %p74, %r176, 0; @%p74 bra $L__BB28_114; ld.const.u64 %rd218, [kernel_params+2264]; cvta.to.global.u64 %rd219, %rd218; ld.local.u32 %r177, [%rd2+-16]; mul.wide.s32 %rd220, %r177, 16; add.s64 %rd221, %rd219, %rd220; shr.s32 %r178, %r31, 7; ld.global.u32 %r179, [%rd221+4]; add.s32 %r180, %r178, %r179; ld.const.u64 %rd222, [kernel_params+2328]; cvta.to.global.u64 %rd223, %rd222; add.s32 %r181, %r180, %r32; mul.wide.s32 %rd224, %r181, 8; add.s64 %rd225, %rd223, %rd224; ld.global.v2.f32 {%f466, %f467}, [%rd225]; ld.global.v2.f32 {%f470, %f471}, [%rd225+8]; add.u64 %rd227, %SPL, 8; mov.f32 %f474, 0f00000000; st.local.v2.f32 [%rd227], {%f474, %f474}; ld.local.f32 %f475, [%rd2+-8]; mov.f32 %f476, 0f3F800000; sub.ftz.f32 %f477, %f476, %f475; mul.ftz.f32 %f478, %f466, %f477; mul.ftz.f32 %f479, %f467, %f477; fma.rn.ftz.f32 %f659, %f471, %f475, %f479; fma.rn.ftz.f32 %f658, %f470, %f475, %f478; bra.uni $L__BB28_120; $L__BB28_162: mov.u32 %r230, 0; st.local.u32 [%rd17], %r230; st.local.u32 [%rd17+4], %r230; st.local.u32 [%rd17+8], %r230; bra.uni $L__BB28_164; $L__BB28_125: setp.eq.s32 %p82, %r40, 0; @%p82 bra $L__BB28_136; bra.uni $L__BB28_126; $L__BB28_136: and.b32 %r211, %r31, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB28_143; and.b32 %r212, %r35, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB28_139; ld.const.u64 %rd267, [kernel_params+2264]; cvta.to.global.u64 %rd268, %rd267; ld.local.u32 %r213, [%rd2+-16]; mul.wide.s32 %rd269, %r213, 16; add.s64 %rd270, %rd268, %rd269; shr.s32 %r214, %r31, 7; ld.global.u32 %r215, [%rd270+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd271, [kernel_params+2320]; cvta.to.global.u64 %rd272, %rd271; add.s32 %r217, %r216, %r32; mul.wide.s32 %rd273, %r217, 4; add.s64 %rd274, %rd272, %rd273; ld.global.f32 %f514, [%rd274]; ld.global.f32 %f515, [%rd274+4]; mov.u32 %r218, 0; st.local.u32 [%rd1], %r218; ld.local.f32 %f516, [%rd2+-8]; mov.f32 %f517, 0f3F800000; sub.ftz.f32 %f518, %f517, %f516; mul.ftz.f32 %f519, %f514, %f518; fma.rn.ftz.f32 %f660, %f515, %f516, %f519; bra.uni $L__BB28_145; $L__BB28_152: ld.local.v4.f32 {%f556, %f557, %f558, %f559}, [%rd2+116]; mul.ftz.f32 %f564, %f662, %f557; fma.rn.ftz.f32 %f565, %f663, %f556, %f564; fma.rn.ftz.f32 %f566, %f661, %f558, %f565; add.ftz.f32 %f173, %f559, %f566; ld.local.v4.f32 {%f567, %f568, %f569, %f570}, [%rd2+132]; mul.ftz.f32 %f575, %f662, %f568; fma.rn.ftz.f32 %f576, %f663, %f567, %f575; fma.rn.ftz.f32 %f577, %f661, %f569, %f576; add.ftz.f32 %f174, %f570, %f577; ld.local.v4.f32 {%f578, %f579, %f580, %f581}, [%rd2+148]; mul.ftz.f32 %f586, %f662, %f579; fma.rn.ftz.f32 %f587, %f663, %f578, %f586; fma.rn.ftz.f32 %f588, %f661, %f580, %f587; add.ftz.f32 %f661, %f581, %f588; mov.f32 %f662, %f174; mov.f32 %f663, %f173; $L__BB28_154: setp.eq.s32 %p103, %r64, 1; @%p103 bra $L__BB28_158; bra.uni $L__BB28_155; $L__BB28_158: add.ftz.f32 %f622, %f662, %f663; add.ftz.f32 %f623, %f661, %f622; mul.ftz.f32 %f624, %f623, 0f3EAAAAAB; st.local.f32 [%rd17], %f624; bra.uni $L__BB28_164; $L__BB28_155: setp.eq.s32 %p104, %r64, 0; @%p104 bra $L__BB28_157; mov.u32 %r228, 1065353216; st.local.u32 [%rd17], %r228; bra.uni $L__BB28_164; $L__BB28_21: add.s32 %r86, %r245, -3; setp.lt.u32 %p23, %r86, 2; @%p23 bra $L__BB28_47; bra.uni $L__BB28_22; $L__BB28_47: setp.eq.s32 %p40, %r40, 0; @%p40 bra $L__BB28_86; bra.uni $L__BB28_48; $L__BB28_86: and.b32 %r144, %r31, 6; setp.eq.s32 %p57, %r144, 0; @%p57 bra $L__BB28_93; and.b32 %r145, %r35, 768; setp.eq.s32 %p58, %r145, 0; @%p58 bra $L__BB28_89; ld.const.u64 %rd160, [kernel_params+2264]; cvta.to.global.u64 %rd161, %rd160; ld.local.u32 %r146, [%rd2+-16]; mul.wide.s32 %rd162, %r146, 16; add.s64 %rd163, %rd161, %rd162; shr.s32 %r147, %r31, 7; ld.global.u32 %r148, [%rd163+4]; add.s32 %r149, %r147, %r148; ld.const.u64 %rd164, [kernel_params+2344]; cvta.to.global.u64 %rd165, %rd164; add.s32 %r150, %r149, %r32; mul.wide.s32 %rd166, %r150, 16; add.s64 %rd167, %rd165, %rd166; ld.global.v4.f32 {%f365, %f366, %f367, %f368}, [%rd167]; ld.global.v4.f32 {%f373, %f374, %f375, %f376}, [%rd167+16]; add.u64 %rd169, %SPL, 16; mov.f32 %f381, 0f00000000; st.local.v4.f32 [%rd169], {%f381, %f381, %f381, %f381}; ld.local.f32 %f382, [%rd2+-8]; mov.f32 %f383, 0f3F800000; sub.ftz.f32 %f384, %f383, %f382; mul.ftz.f32 %f385, %f365, %f384; mul.ftz.f32 %f386, %f366, %f384; mul.ftz.f32 %f387, %f367, %f384; mul.ftz.f32 %f388, %f368, %f384; fma.rn.ftz.f32 %f651, %f376, %f382, %f388; fma.rn.ftz.f32 %f650, %f375, %f382, %f387; fma.rn.ftz.f32 %f649, %f374, %f382, %f386; fma.rn.ftz.f32 %f648, %f373, %f382, %f385; bra.uni $L__BB28_95; $L__BB28_126: ld.local.s32 %rd55, [%rd2+-16]; setp.eq.s64 %p83, %rd55, -1; @%p83 bra $L__BB28_129; ld.const.u64 %rd244, [kernel_params+2240]; cvta.to.global.u64 %rd245, %rd244; shl.b64 %rd246, %rd55, 2; add.s64 %rd247, %rd245, %rd246; ld.global.u32 %r189, [%rd247]; setp.eq.s32 %p84, %r189, -1; @%p84 bra $L__BB28_129; mov.u32 %r190, 0; mov.u64 %rd248, 0; { // callseq 33, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r190; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd248; .param .b64 param3; st.param.b64 [param3+0], %rd67; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f660, [retval0+0]; } // callseq 33 bra.uni $L__BB28_145; $L__BB28_157: st.local.f32 [%rd17], %f663; st.local.f32 [%rd17+4], %f662; st.local.f32 [%rd17+8], %f661; bra.uni $L__BB28_164; $L__BB28_101: ld.local.s32 %rd53, [%rd2+-16]; setp.eq.s64 %p67, %rd53, -1; @%p67 bra $L__BB28_104; ld.const.u64 %rd188, [kernel_params+2240]; cvta.to.global.u64 %rd189, %rd188; shl.b64 %rd190, %rd53, 2; add.s64 %rd191, %rd189, %rd190; ld.global.u32 %r156, [%rd191]; setp.eq.s32 %p68, %r156, -1; @%p68 bra $L__BB28_104; mov.u64 %rd192, 0; add.u64 %rd193, %SP, 8; { // callseq 32, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r50; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd192; .param .b64 param3; st.param.b64 [param3+0], %rd193; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f658, [retval0+0]; ld.param.f32 %f659, [retval0+4]; } // callseq 32 bra.uni $L__BB28_120; $L__BB28_22: setp.eq.s32 %p24, %r40, 0; add.u64 %rd86, %SP, 32; add.u64 %rd18, %SPL, 32; @%p24 bra $L__BB28_33; bra.uni $L__BB28_23; $L__BB28_33: and.b32 %r109, %r31, 6; setp.eq.s32 %p31, %r109, 0; @%p31 bra $L__BB28_40; and.b32 %r110, %r35, 768; setp.eq.s32 %p32, %r110, 0; @%p32 bra $L__BB28_36; ld.const.u64 %rd111, [kernel_params+2264]; cvta.to.global.u64 %rd112, %rd111; ld.local.u32 %r111, [%rd2+-16]; mul.wide.s32 %rd113, %r111, 16; add.s64 %rd114, %rd112, %rd113; shr.s32 %r112, %r31, 7; ld.global.u32 %r113, [%rd114+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd115, [kernel_params+2336]; cvta.to.global.u64 %rd116, %rd115; add.s32 %r115, %r114, %r32; mul.wide.s32 %rd117, %r115, 12; add.s64 %rd118, %rd116, %rd117; ld.global.f32 %f220, [%rd118]; ld.global.f32 %f221, [%rd118+4]; ld.global.f32 %f222, [%rd118+8]; ld.global.f32 %f223, [%rd118+12]; ld.global.f32 %f224, [%rd118+16]; ld.global.f32 %f225, [%rd118+20]; mov.u32 %r116, 0; st.local.u32 [%rd18], %r116; st.local.u32 [%rd18+4], %r116; st.local.u32 [%rd18+8], %r116; ld.local.f32 %f226, [%rd2+-8]; mov.f32 %f227, 0f3F800000; sub.ftz.f32 %f228, %f227, %f226; mul.ftz.f32 %f229, %f220, %f228; mul.ftz.f32 %f230, %f221, %f228; mul.ftz.f32 %f231, %f222, %f228; fma.rn.ftz.f32 %f627, %f225, %f226, %f231; fma.rn.ftz.f32 %f626, %f224, %f226, %f230; fma.rn.ftz.f32 %f625, %f223, %f226, %f229; bra.uni $L__BB28_42; $L__BB28_129: and.b32 %r191, %r35, 56; setp.eq.s32 %p85, %r191, 0; @%p85 bra $L__BB28_134; and.b32 %r192, %r35, 24; setp.eq.s32 %p86, %r192, 0; @%p86 bra $L__BB28_132; ld.const.u64 %rd250, [kernel_params+2232]; cvta.to.global.u64 %rd251, %rd250; shl.b64 %rd252, %rd55, 4; add.s64 %rd253, %rd251, %rd252; ld.global.v4.u32 {%r193, %r194, %r195, %r196}, [%rd253]; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r200, %r193, %r32; mul.wide.u32 %rd256, %r200, 4; add.s64 %rd299, %rd255, %rd256; add.s32 %r201, %r194, %r32; mul.wide.u32 %rd257, %r201, 4; add.s64 %rd300, %rd255, %rd257; add.s32 %r202, %r195, %r32; mul.wide.u32 %rd258, %r202, 4; add.s64 %rd301, %rd255, %rd258; bra.uni $L__BB28_133; $L__BB28_104: and.b32 %r157, %r35, 56; setp.eq.s32 %p69, %r157, 0; @%p69 bra $L__BB28_109; and.b32 %r158, %r35, 24; setp.eq.s32 %p70, %r158, 0; @%p70 bra $L__BB28_107; ld.const.u64 %rd194, [kernel_params+2232]; cvta.to.global.u64 %rd195, %rd194; shl.b64 %rd196, %rd53, 4; add.s64 %rd197, %rd195, %rd196; ld.global.v4.u32 {%r159, %r160, %r161, %r162}, [%rd197]; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r166, %r159, %r32; mul.wide.u32 %rd200, %r166, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f656, %f657}, [%rd201]; add.s32 %r167, %r160, %r32; mul.wide.u32 %rd202, %r167, 8; add.s64 %rd203, %rd199, %rd202; ld.global.v2.f32 {%f654, %f655}, [%rd203]; add.s32 %r168, %r161, %r32; mul.wide.u32 %rd204, %r168, 8; add.s64 %rd205, %rd199, %rd204; ld.global.v2.f32 {%f652, %f653}, [%rd205]; bra.uni $L__BB28_108; $L__BB28_48: ld.local.s32 %rd49, [%rd2+-16]; setp.eq.s64 %p41, %rd49, -1; @%p41 bra $L__BB28_51; ld.const.u64 %rd127, [kernel_params+2240]; cvta.to.global.u64 %rd128, %rd127; shl.b64 %rd129, %rd49, 2; add.s64 %rd130, %rd128, %rd129; ld.global.u32 %r125, [%rd130]; setp.eq.s32 %p42, %r125, -1; @%p42 bra $L__BB28_51; mov.u64 %rd131, 0; add.u64 %rd132, %SP, 16; { // callseq 31, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r245; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd131; .param .b64 param3; st.param.b64 [param3+0], %rd132; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f648, [retval0+0]; ld.param.f32 %f649, [retval0+4]; ld.param.f32 %f650, [retval0+8]; ld.param.f32 %f651, [retval0+12]; } // callseq 31 bra.uni $L__BB28_95; $L__BB28_23: ld.local.s32 %rd20, [%rd2+-16]; setp.eq.s64 %p25, %rd20, -1; @%p25 bra $L__BB28_26; ld.const.u64 %rd87, [kernel_params+2240]; cvta.to.global.u64 %rd88, %rd87; shl.b64 %rd89, %rd20, 2; add.s64 %rd90, %rd88, %rd89; ld.global.u32 %r87, [%rd90]; setp.eq.s32 %p26, %r87, -1; @%p26 bra $L__BB28_26; mov.u64 %rd91, 0; { // callseq 30, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r35; st.param.b32 [param1+4], %r245; st.param.b32 [param1+8], %r33; st.param.b32 [param1+12], %r32; .param .b64 param2; st.param.b64 [param2+0], %rd91; .param .b64 param3; st.param.b64 [param3+0], %rd86; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f625, [retval0+0]; ld.param.f32 %f626, [retval0+4]; ld.param.f32 %f627, [retval0+8]; } // callseq 30 bra.uni $L__BB28_42; $L__BB28_143: and.b32 %r222, %r31, 8; setp.eq.s32 %p93, %r222, 0; mov.u32 %r223, 0; mov.f32 %f660, 0f00000000; st.local.u32 [%rd1], %r223; setp.ne.s32 %p94, %r35, 8; or.pred %p95, %p94, %p93; @%p95 bra $L__BB28_145; ld.const.u64 %rd279, [kernel_params+2320]; cvta.to.global.u64 %rd280, %rd279; ld.local.u32 %r224, [%rd2+-16]; add.s32 %r225, %r224, %r32; mul.wide.s32 %rd281, %r225, 4; add.s64 %rd282, %rd280, %rd281; ld.global.f32 %f660, [%rd282]; bra.uni $L__BB28_145; $L__BB28_118: and.b32 %r184, %r31, 8; setp.eq.s32 %p77, %r184, 0; add.u64 %rd235, %SPL, 8; mov.f32 %f658, 0f00000000; st.local.v2.f32 [%rd235], {%f658, %f658}; setp.ne.s32 %p78, %r35, 8; or.pred %p79, %p78, %p77; mov.f32 %f659, %f658; @%p79 bra $L__BB28_120; ld.const.u64 %rd236, [kernel_params+2328]; cvta.to.global.u64 %rd237, %rd236; ld.local.u32 %r185, [%rd2+-16]; add.s32 %r186, %r185, %r32; mul.wide.s32 %rd238, %r186, 8; add.s64 %rd239, %rd237, %rd238; ld.global.v2.f32 {%f658, %f659}, [%rd239]; bra.uni $L__BB28_120; $L__BB28_51: and.b32 %r126, %r35, 120; setp.eq.s32 %p43, %r126, 0; @%p43 bra $L__BB28_84; and.b32 %r127, %r35, 24; setp.eq.s32 %p44, %r127, 0; @%p44 bra $L__BB28_54; ld.const.u64 %rd133, [kernel_params+2232]; cvta.to.global.u64 %rd134, %rd133; shl.b64 %rd135, %rd49, 4; add.s64 %rd136, %rd134, %rd135; ld.global.v4.u32 {%r128, %r129, %r130, %r131}, [%rd136]; ld.const.u64 %rd137, [kernel_params+2344]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r135, %r128, %r32; mul.wide.u32 %rd139, %r135, 16; add.s64 %rd140, %rd138, %rd139; ld.global.v4.f32 {%f644, %f645, %f646, %f647}, [%rd140]; add.s32 %r136, %r129, %r32; mul.wide.u32 %rd141, %r136, 16; add.s64 %rd142, %rd138, %rd141; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd142]; add.s32 %r137, %r130, %r32; mul.wide.u32 %rd143, %r137, 16; add.s64 %rd144, %rd138, %rd143; ld.global.v4.f32 {%f636, %f637, %f638, %f639}, [%rd144]; bra.uni $L__BB28_83; $L__BB28_26: and.b32 %r88, %r35, 56; setp.eq.s32 %p27, %r88, 0; @%p27 bra $L__BB28_31; and.b32 %r89, %r35, 24; setp.eq.s32 %p28, %r89, 0; @%p28 bra $L__BB28_29; ld.const.u64 %rd93, [kernel_params+2232]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd20, 4; add.s64 %rd96, %rd94, %rd95; ld.global.v4.u32 {%r90, %r91, %r92, %r93}, [%rd96]; ld.const.u64 %rd97, [kernel_params+2336]; cvta.to.global.u64 %rd98, %rd97; add.s32 %r97, %r90, %r32; mul.wide.u32 %rd99, %r97, 12; add.s64 %rd296, %rd98, %rd99; add.s64 %rd297, %rd296, 4; add.s64 %rd298, %rd296, 8; add.s32 %r98, %r91, %r32; mul.wide.u32 %rd100, %r98, 12; add.s64 %rd293, %rd98, %rd100; add.s64 %rd294, %rd293, 4; add.s64 %rd295, %rd293, 8; add.s32 %r99, %r92, %r32; mul.wide.u32 %rd101, %r99, 12; add.s64 %rd290, %rd98, %rd101; add.s64 %rd291, %rd290, 4; add.s64 %rd292, %rd290, 8; bra.uni $L__BB28_30; $L__BB28_139: mov.f32 %f660, 0f00000000; mov.u32 %r219, 0; st.local.u32 [%rd1], %r219; and.b32 %r220, %r35, 131; setp.eq.s32 %p91, %r220, 0; @%p91 bra $L__BB28_145; setp.ne.s32 %p92, %r35, 128; @%p92 bra $L__BB28_142; ld.local.u32 %r221, [%rd2+-16]; add.s32 %r32, %r221, %r32; $L__BB28_142: ld.const.u64 %rd275, [kernel_params+2320]; cvta.to.global.u64 %rd276, %rd275; mul.wide.s32 %rd277, %r32, 4; add.s64 %rd278, %rd276, %rd277; ld.global.f32 %f660, [%rd278]; bra.uni $L__BB28_145; $L__BB28_134: mov.f32 %f660, 0f00000000; mov.u32 %r206, 0; st.local.u32 [%rd1], %r206; and.b32 %r207, %r35, 7; setp.eq.s32 %p87, %r207, 0; @%p87 bra $L__BB28_145; cvt.u32.u64 %r208, %rd55; setp.eq.s32 %p88, %r35, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r32, %r209; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; mul.wide.s32 %rd265, %r210, 4; add.s64 %rd266, %rd264, %rd265; ld.global.f32 %f660, [%rd266]; bra.uni $L__BB28_145; $L__BB28_114: add.u64 %rd229, %SPL, 8; mov.f32 %f658, 0f00000000; st.local.v2.f32 [%rd229], {%f658, %f658}; and.b32 %r182, %r35, 131; setp.eq.s32 %p75, %r182, 0; mov.f32 %f659, %f658; @%p75 bra $L__BB28_120; setp.ne.s32 %p76, %r35, 128; @%p76 bra $L__BB28_117; ld.local.u32 %r183, [%rd2+-16]; add.s32 %r32, %r183, %r32; $L__BB28_117: ld.const.u64 %rd230, [kernel_params+2328]; cvta.to.global.u64 %rd231, %rd230; mul.wide.s32 %rd232, %r32, 8; add.s64 %rd233, %rd231, %rd232; ld.global.v2.f32 {%f658, %f659}, [%rd233]; bra.uni $L__BB28_120; $L__BB28_109: add.u64 %rd213, %SPL, 8; mov.f32 %f658, 0f00000000; st.local.v2.f32 [%rd213], {%f658, %f658}; and.b32 %r171, %r35, 7; setp.eq.s32 %p71, %r171, 0; mov.f32 %f659, %f658; @%p71 bra $L__BB28_120; cvt.u32.u64 %r172, %rd53; setp.eq.s32 %p72, %r35, 4; selp.b32 %r173, %r172, 0, %p72; add.s32 %r174, %r32, %r173; ld.const.u64 %rd214, [kernel_params+2328]; cvta.to.global.u64 %rd215, %rd214; mul.wide.s32 %rd216, %r174, 8; add.s64 %rd217, %rd215, %rd216; ld.global.v2.f32 {%f658, %f659}, [%rd217]; bra.uni $L__BB28_120; $L__BB28_93: and.b32 %r153, %r31, 8; setp.eq.s32 %p61, %r153, 0; add.u64 %rd177, %SPL, 16; mov.f32 %f648, 0f00000000; st.local.v4.f32 [%rd177], {%f648, %f648, %f648, %f648}; setp.ne.s32 %p62, %r35, 8; or.pred %p63, %p62, %p61; mov.f32 %f649, %f648; mov.f32 %f650, %f648; mov.f32 %f651, %f648; @%p63 bra $L__BB28_95; ld.const.u64 %rd178, [kernel_params+2344]; cvta.to.global.u64 %rd179, %rd178; ld.local.u32 %r154, [%rd2+-16]; add.s32 %r155, %r154, %r32; mul.wide.s32 %rd180, %r155, 16; add.s64 %rd181, %rd179, %rd180; ld.global.v4.f32 {%f648, %f649, %f650, %f651}, [%rd181]; bra.uni $L__BB28_95; $L__BB28_40: and.b32 %r120, %r31, 8; setp.eq.s32 %p35, %r120, 0; mov.u32 %r121, 0; st.local.u32 [%rd18], %r121; st.local.u32 [%rd18+4], %r121; st.local.u32 [%rd18+8], %r121; setp.ne.s32 %p36, %r35, 8; mov.f32 %f625, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f626, %f625; mov.f32 %f627, %f625; @%p37 bra $L__BB28_42; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r122, [%rd2+-16]; add.s32 %r123, %r122, %r32; mul.wide.s32 %rd125, %r123, 12; add.s64 %rd126, %rd124, %rd125; ld.global.f32 %f627, [%rd126+8]; ld.global.f32 %f626, [%rd126+4]; ld.global.f32 %f625, [%rd126]; bra.uni $L__BB28_42; $L__BB28_132: cvt.u32.u64 %r203, %rd55; mad.lo.s32 %r204, %r203, 3, %r32; ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r204, 4; add.s64 %rd299, %rd260, %rd261; add.s64 %rd300, %rd299, 4; add.s32 %r205, %r204, 2; mul.wide.s32 %rd262, %r205, 4; add.s64 %rd301, %rd260, %rd262; $L__BB28_133: ld.local.f32 %f496, [%rd2+28]; ld.global.f32 %f497, [%rd300]; ld.local.f32 %f498, [%rd2+36]; ld.global.f32 %f499, [%rd301]; mul.ftz.f32 %f500, %f499, %f498; fma.rn.ftz.f32 %f501, %f497, %f496, %f500; add.ftz.f32 %f502, %f496, %f498; ld.global.f32 %f503, [%rd299]; mul.ftz.f32 %f504, %f503, %f502; sub.ftz.f32 %f505, %f501, %f504; st.local.f32 [%rd1], %f505; ld.local.f32 %f506, [%rd2+-8]; ld.local.f32 %f507, [%rd2+-4]; mul.ftz.f32 %f508, %f499, %f507; fma.rn.ftz.f32 %f509, %f497, %f506, %f508; mov.f32 %f510, 0f3F800000; sub.ftz.f32 %f511, %f510, %f506; sub.ftz.f32 %f512, %f511, %f507; fma.rn.ftz.f32 %f660, %f503, %f512, %f509; $L__BB28_145: setp.eq.s32 %p96, %r64, 1; @%p96 bra $L__BB28_149; bra.uni $L__BB28_146; $L__BB28_149: ld.local.f32 %f524, [%rd1]; add.ftz.f32 %f525, %f660, %f524; st.local.f32 [%rd17], %f525; bra.uni $L__BB28_164; $L__BB28_146: setp.eq.s32 %p97, %r64, 0; @%p97 bra $L__BB28_148; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB28_164; $L__BB28_148: ld.local.f32 %f522, [%rd1]; add.ftz.f32 %f523, %f660, %f522; st.local.f32 [%rd17], %f523; st.local.f32 [%rd17+4], %f523; st.local.f32 [%rd17+8], %f523; bra.uni $L__BB28_164; $L__BB28_107: cvt.u32.u64 %r169, %rd53; mad.lo.s32 %r170, %r169, 3, %r32; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; mul.wide.s32 %rd208, %r170, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f656, %f657}, [%rd209]; ld.global.v2.f32 {%f654, %f655}, [%rd209+8]; ld.global.v2.f32 {%f652, %f653}, [%rd209+16]; $L__BB28_108: ld.local.f32 %f442, [%rd2+28]; ld.local.f32 %f443, [%rd2+36]; mul.ftz.f32 %f444, %f652, %f443; mul.ftz.f32 %f445, %f653, %f443; fma.rn.ftz.f32 %f446, %f654, %f442, %f444; fma.rn.ftz.f32 %f447, %f655, %f442, %f445; add.ftz.f32 %f448, %f443, %f442; mul.ftz.f32 %f449, %f656, %f448; mul.ftz.f32 %f450, %f657, %f448; sub.ftz.f32 %f451, %f447, %f450; sub.ftz.f32 %f452, %f446, %f449; add.u64 %rd211, %SPL, 8; st.local.v2.f32 [%rd211], {%f452, %f451}; ld.local.f32 %f453, [%rd2+-8]; ld.local.f32 %f454, [%rd2+-4]; mul.ftz.f32 %f455, %f652, %f454; mul.ftz.f32 %f456, %f653, %f454; fma.rn.ftz.f32 %f457, %f654, %f453, %f455; fma.rn.ftz.f32 %f458, %f655, %f453, %f456; mov.f32 %f459, 0f3F800000; sub.ftz.f32 %f460, %f459, %f453; sub.ftz.f32 %f461, %f460, %f454; fma.rn.ftz.f32 %f659, %f657, %f461, %f458; fma.rn.ftz.f32 %f658, %f656, %f461, %f457; $L__BB28_120: setp.eq.s32 %p80, %r64, 1; @%p80 bra $L__BB28_124; bra.uni $L__BB28_121; $L__BB28_124: add.u64 %rd243, %SPL, 8; ld.local.f32 %f494, [%rd243]; add.ftz.f32 %f495, %f658, %f494; st.local.f32 [%rd17], %f495; bra.uni $L__BB28_164; $L__BB28_121: setp.eq.s32 %p81, %r64, 0; @%p81 bra $L__BB28_123; mov.u32 %r187, 1065353216; st.local.u32 [%rd17], %r187; bra.uni $L__BB28_164; $L__BB28_123: add.u64 %rd241, %SPL, 8; ld.local.v2.f32 {%f488, %f489}, [%rd241]; add.ftz.f32 %f492, %f658, %f488; add.ftz.f32 %f493, %f659, %f489; st.local.f32 [%rd17], %f492; st.local.f32 [%rd17+4], %f493; mov.u32 %r188, 0; st.local.u32 [%rd17+8], %r188; bra.uni $L__BB28_164; $L__BB28_89: add.u64 %rd171, %SPL, 16; mov.f32 %f648, 0f00000000; st.local.v4.f32 [%rd171], {%f648, %f648, %f648, %f648}; and.b32 %r151, %r35, 131; setp.eq.s32 %p59, %r151, 0; mov.f32 %f649, %f648; mov.f32 %f650, %f648; mov.f32 %f651, %f648; @%p59 bra $L__BB28_95; setp.ne.s32 %p60, %r35, 128; @%p60 bra $L__BB28_92; ld.local.u32 %r152, [%rd2+-16]; add.s32 %r32, %r152, %r32; $L__BB28_92: ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; mul.wide.s32 %rd174, %r32, 16; add.s64 %rd175, %rd173, %rd174; ld.global.v4.f32 {%f648, %f649, %f650, %f651}, [%rd175]; bra.uni $L__BB28_95; $L__BB28_36: mov.u32 %r117, 0; st.local.u32 [%rd18], %r117; st.local.u32 [%rd18+4], %r117; st.local.u32 [%rd18+8], %r117; and.b32 %r118, %r35, 131; setp.eq.s32 %p33, %r118, 0; mov.f32 %f625, 0f00000000; mov.f32 %f626, %f625; mov.f32 %f627, %f625; @%p33 bra $L__BB28_42; setp.ne.s32 %p34, %r35, 128; @%p34 bra $L__BB28_39; ld.local.u32 %r119, [%rd2+-16]; add.s32 %r32, %r119, %r32; $L__BB28_39: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r32, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f627, [%rd122+8]; ld.global.f32 %f626, [%rd122+4]; ld.global.f32 %f625, [%rd122]; bra.uni $L__BB28_42; $L__BB28_84: add.u64 %rd155, %SPL, 16; mov.f32 %f648, 0f00000000; st.local.v4.f32 [%rd155], {%f648, %f648, %f648, %f648}; and.b32 %r140, %r35, 7; setp.eq.s32 %p55, %r140, 0; mov.f32 %f649, %f648; mov.f32 %f650, %f648; mov.f32 %f651, %f648; @%p55 bra $L__BB28_95; cvt.u32.u64 %r141, %rd49; setp.eq.s32 %p56, %r35, 4; selp.b32 %r142, %r141, 0, %p56; add.s32 %r143, %r32, %r142; ld.const.u64 %rd156, [kernel_params+2344]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r143, 16; add.s64 %rd159, %rd157, %rd158; ld.global.v4.f32 {%f648, %f649, %f650, %f651}, [%rd159]; bra.uni $L__BB28_95; $L__BB28_31: mov.u32 %r104, 0; st.local.u32 [%rd18], %r104; st.local.u32 [%rd18+4], %r104; st.local.u32 [%rd18+8], %r104; and.b32 %r105, %r35, 7; setp.eq.s32 %p29, %r105, 0; mov.f32 %f625, 0f00000000; mov.f32 %f626, %f625; mov.f32 %f627, %f625; @%p29 bra $L__BB28_42; cvt.u32.u64 %r106, %rd20; setp.eq.s32 %p30, %r35, 4; selp.b32 %r107, %r106, 0, %p30; add.s32 %r108, %r32, %r107; ld.const.u64 %rd107, [kernel_params+2336]; cvta.to.global.u64 %rd108, %rd107; mul.wide.s32 %rd109, %r108, 12; add.s64 %rd110, %rd108, %rd109; ld.global.f32 %f627, [%rd110+8]; ld.global.f32 %f626, [%rd110+4]; ld.global.f32 %f625, [%rd110]; bra.uni $L__BB28_42; $L__BB28_54: cvt.u32.u64 %r138, %rd49; mad.lo.s32 %r139, %r138, 3, %r32; cvt.s64.s32 %rd50, %r139; setp.eq.s32 %p45, %r35, 32; @%p45 bra $L__BB28_82; bra.uni $L__BB28_55; $L__BB28_82: ld.const.u64 %rd148, [kernel_params+2344]; cvta.to.global.u64 %rd149, %rd148; shl.b64 %rd150, %rd50, 4; add.s64 %rd151, %rd149, %rd150; ld.global.v4.f32 {%f644, %f645, %f646, %f647}, [%rd151]; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd151+16]; ld.global.v4.f32 {%f636, %f637, %f638, %f639}, [%rd151+32]; bra.uni $L__BB28_83; $L__BB28_29: cvt.u32.u64 %r100, %rd20; mad.lo.s32 %r101, %r100, 3, %r32; ld.const.u64 %rd102, [kernel_params+2336]; cvta.to.global.u64 %rd103, %rd102; mul.wide.s32 %rd104, %r101, 12; add.s64 %rd296, %rd103, %rd104; add.s64 %rd297, %rd296, 4; add.s64 %rd298, %rd296, 8; add.s32 %r102, %r101, 1; mul.wide.s32 %rd105, %r102, 12; add.s64 %rd293, %rd103, %rd105; add.s64 %rd294, %rd293, 4; add.s64 %rd295, %rd293, 8; add.s32 %r103, %r101, 2; mul.wide.s32 %rd106, %r103, 12; add.s64 %rd290, %rd103, %rd106; add.s64 %rd291, %rd290, 4; add.s64 %rd292, %rd290, 8; $L__BB28_30: ld.local.f32 %f182, [%rd2+28]; ld.global.f32 %f183, [%rd293]; ld.global.f32 %f184, [%rd294]; ld.global.f32 %f185, [%rd295]; ld.local.f32 %f186, [%rd2+36]; ld.global.f32 %f187, [%rd290]; mul.ftz.f32 %f188, %f187, %f186; ld.global.f32 %f189, [%rd291]; mul.ftz.f32 %f190, %f189, %f186; ld.global.f32 %f191, [%rd292]; mul.ftz.f32 %f192, %f191, %f186; fma.rn.ftz.f32 %f193, %f183, %f182, %f188; fma.rn.ftz.f32 %f194, %f184, %f182, %f190; fma.rn.ftz.f32 %f195, %f185, %f182, %f192; add.ftz.f32 %f196, %f182, %f186; ld.global.f32 %f197, [%rd296]; mul.ftz.f32 %f198, %f197, %f196; ld.global.f32 %f199, [%rd297]; mul.ftz.f32 %f200, %f199, %f196; ld.global.f32 %f201, [%rd298]; mul.ftz.f32 %f202, %f201, %f196; sub.ftz.f32 %f203, %f193, %f198; sub.ftz.f32 %f204, %f194, %f200; sub.ftz.f32 %f205, %f195, %f202; st.local.f32 [%rd18], %f203; st.local.f32 [%rd18+4], %f204; st.local.f32 [%rd18+8], %f205; ld.local.f32 %f206, [%rd2+-8]; ld.local.f32 %f207, [%rd2+-4]; mul.ftz.f32 %f208, %f187, %f207; mul.ftz.f32 %f209, %f189, %f207; mul.ftz.f32 %f210, %f191, %f207; fma.rn.ftz.f32 %f211, %f183, %f206, %f208; fma.rn.ftz.f32 %f212, %f184, %f206, %f209; fma.rn.ftz.f32 %f213, %f185, %f206, %f210; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f206; sub.ftz.f32 %f216, %f215, %f207; fma.rn.ftz.f32 %f627, %f201, %f216, %f213; fma.rn.ftz.f32 %f626, %f199, %f216, %f212; fma.rn.ftz.f32 %f625, %f197, %f216, %f211; $L__BB28_42: setp.eq.s32 %p38, %r64, 1; @%p38 bra $L__BB28_46; bra.uni $L__BB28_43; $L__BB28_46: ld.local.f32 %f244, [%rd18]; add.ftz.f32 %f245, %f625, %f244; ld.local.f32 %f246, [%rd18+4]; add.ftz.f32 %f247, %f626, %f246; ld.local.f32 %f248, [%rd18+8]; add.ftz.f32 %f249, %f627, %f248; add.ftz.f32 %f250, %f245, %f247; add.ftz.f32 %f251, %f249, %f250; mul.ftz.f32 %f252, %f251, 0f3EAAAAAB; st.local.f32 [%rd17], %f252; bra.uni $L__BB28_164; $L__BB28_43: setp.eq.s32 %p39, %r64, 0; @%p39 bra $L__BB28_45; mov.u32 %r124, 1065353216; st.local.u32 [%rd17], %r124; bra.uni $L__BB28_164; $L__BB28_45: ld.local.f32 %f238, [%rd18]; add.ftz.f32 %f239, %f625, %f238; ld.local.f32 %f240, [%rd18+4]; add.ftz.f32 %f241, %f626, %f240; ld.local.f32 %f242, [%rd18+8]; add.ftz.f32 %f243, %f627, %f242; st.local.f32 [%rd17], %f239; st.local.f32 [%rd17+4], %f241; st.local.f32 [%rd17+8], %f243; bra.uni $L__BB28_164; $L__BB28_55: ld.const.u64 %rd145, [kernel_params+2352]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd50, 2; add.s64 %rd51, %rd146, %rd147; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd51]; cvt.rn.f32.u16 %f265, %rs20; mul.ftz.f32 %f41, %f265, 0f3B808081; cvt.rn.f32.u16 %f266, %rs21; mul.ftz.f32 %f42, %f266, 0f3B808081; cvt.rn.f32.u16 %f267, %rs22; mul.ftz.f32 %f43, %f267, 0f3B808081; setp.lt.ftz.f32 %p46, %f41, 0f3D25AEE6; @%p46 bra $L__BB28_57; bra.uni $L__BB28_56; $L__BB28_57: mul.ftz.f32 %f644, %f41, 0f3D9E8391; bra.uni $L__BB28_58; $L__BB28_56: add.ftz.f32 %f268, %f41, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f644, %f271; $L__BB28_58: setp.lt.ftz.f32 %p47, %f42, 0f3D25AEE6; @%p47 bra $L__BB28_60; bra.uni $L__BB28_59; $L__BB28_60: mul.ftz.f32 %f645, %f42, 0f3D9E8391; bra.uni $L__BB28_61; $L__BB28_59: add.ftz.f32 %f272, %f42, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f645, %f275; $L__BB28_61: setp.lt.ftz.f32 %p48, %f43, 0f3D25AEE6; @%p48 bra $L__BB28_63; bra.uni $L__BB28_62; $L__BB28_63: mul.ftz.f32 %f646, %f43, 0f3D9E8391; bra.uni $L__BB28_64; $L__BB28_62: add.ftz.f32 %f276, %f43, 0f3D6147AE; mul.ftz.f32 %f277, %f276, 0f3F72A76F; lg2.approx.ftz.f32 %f278, %f277; mul.ftz.f32 %f279, %f278, 0f4019999A; ex2.approx.ftz.f32 %f646, %f279; $L__BB28_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd51+4]; cvt.rn.f32.u16 %f280, %rs24; mul.ftz.f32 %f53, %f280, 0f3B808081; cvt.rn.f32.u16 %f281, %rs25; mul.ftz.f32 %f54, %f281, 0f3B808081; cvt.rn.f32.u16 %f282, %rs26; mul.ftz.f32 %f55, %f282, 0f3B808081; setp.lt.ftz.f32 %p49, %f53, 0f3D25AEE6; @%p49 bra $L__BB28_66; bra.uni $L__BB28_65; $L__BB28_66: mul.ftz.f32 %f640, %f53, 0f3D9E8391; bra.uni $L__BB28_67; $L__BB28_65: add.ftz.f32 %f283, %f53, 0f3D6147AE; mul.ftz.f32 %f284, %f283, 0f3F72A76F; lg2.approx.ftz.f32 %f285, %f284; mul.ftz.f32 %f286, %f285, 0f4019999A; ex2.approx.ftz.f32 %f640, %f286; $L__BB28_67: setp.lt.ftz.f32 %p50, %f54, 0f3D25AEE6; @%p50 bra $L__BB28_69; bra.uni $L__BB28_68; $L__BB28_69: mul.ftz.f32 %f641, %f54, 0f3D9E8391; bra.uni $L__BB28_70; $L__BB28_68: add.ftz.f32 %f287, %f54, 0f3D6147AE; mul.ftz.f32 %f288, %f287, 0f3F72A76F; lg2.approx.ftz.f32 %f289, %f288; mul.ftz.f32 %f290, %f289, 0f4019999A; ex2.approx.ftz.f32 %f641, %f290; $L__BB28_70: setp.lt.ftz.f32 %p51, %f55, 0f3D25AEE6; @%p51 bra $L__BB28_72; bra.uni $L__BB28_71; $L__BB28_72: mul.ftz.f32 %f642, %f55, 0f3D9E8391; bra.uni $L__BB28_73; $L__BB28_71: add.ftz.f32 %f291, %f55, 0f3D6147AE; mul.ftz.f32 %f292, %f291, 0f3F72A76F; lg2.approx.ftz.f32 %f293, %f292; mul.ftz.f32 %f294, %f293, 0f4019999A; ex2.approx.ftz.f32 %f642, %f294; $L__BB28_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd51+8]; cvt.rn.f32.u16 %f295, %rs28; mul.ftz.f32 %f65, %f295, 0f3B808081; cvt.rn.f32.u16 %f296, %rs29; mul.ftz.f32 %f66, %f296, 0f3B808081; cvt.rn.f32.u16 %f297, %rs30; mul.ftz.f32 %f67, %f297, 0f3B808081; setp.lt.ftz.f32 %p52, %f65, 0f3D25AEE6; @%p52 bra $L__BB28_75; bra.uni $L__BB28_74; $L__BB28_75: mul.ftz.f32 %f636, %f65, 0f3D9E8391; bra.uni $L__BB28_76; $L__BB28_74: add.ftz.f32 %f298, %f65, 0f3D6147AE; mul.ftz.f32 %f299, %f298, 0f3F72A76F; lg2.approx.ftz.f32 %f300, %f299; mul.ftz.f32 %f301, %f300, 0f4019999A; ex2.approx.ftz.f32 %f636, %f301; $L__BB28_76: setp.lt.ftz.f32 %p53, %f66, 0f3D25AEE6; @%p53 bra $L__BB28_78; bra.uni $L__BB28_77; $L__BB28_78: mul.ftz.f32 %f637, %f66, 0f3D9E8391; bra.uni $L__BB28_79; $L__BB28_77: add.ftz.f32 %f302, %f66, 0f3D6147AE; mul.ftz.f32 %f303, %f302, 0f3F72A76F; lg2.approx.ftz.f32 %f304, %f303; mul.ftz.f32 %f305, %f304, 0f4019999A; ex2.approx.ftz.f32 %f637, %f305; $L__BB28_79: cvt.rn.f32.u16 %f306, %rs23; mul.ftz.f32 %f647, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs27; mul.ftz.f32 %f643, %f307, 0f3B808081; cvt.rn.f32.u16 %f308, %rs31; mul.ftz.f32 %f639, %f308, 0f3B808081; setp.lt.ftz.f32 %p54, %f67, 0f3D25AEE6; @%p54 bra $L__BB28_81; bra.uni $L__BB28_80; $L__BB28_81: mul.ftz.f32 %f638, %f67, 0f3D9E8391; bra.uni $L__BB28_83; $L__BB28_80: add.ftz.f32 %f309, %f67, 0f3D6147AE; mul.ftz.f32 %f310, %f309, 0f3F72A76F; lg2.approx.ftz.f32 %f311, %f310; mul.ftz.f32 %f312, %f311, 0f4019999A; ex2.approx.ftz.f32 %f638, %f312; $L__BB28_83: ld.local.f32 %f325, [%rd2+28]; ld.local.f32 %f326, [%rd2+36]; mul.ftz.f32 %f327, %f636, %f326; mul.ftz.f32 %f328, %f637, %f326; mul.ftz.f32 %f329, %f638, %f326; mul.ftz.f32 %f330, %f639, %f326; fma.rn.ftz.f32 %f331, %f640, %f325, %f327; fma.rn.ftz.f32 %f332, %f641, %f325, %f328; fma.rn.ftz.f32 %f333, %f642, %f325, %f329; fma.rn.ftz.f32 %f334, %f643, %f325, %f330; add.ftz.f32 %f335, %f325, %f326; mul.ftz.f32 %f336, %f644, %f335; mul.ftz.f32 %f337, %f645, %f335; mul.ftz.f32 %f338, %f646, %f335; mul.ftz.f32 %f339, %f647, %f335; sub.ftz.f32 %f340, %f334, %f339; sub.ftz.f32 %f341, %f333, %f338; sub.ftz.f32 %f342, %f332, %f337; sub.ftz.f32 %f343, %f331, %f336; add.u64 %rd153, %SPL, 16; st.local.v4.f32 [%rd153], {%f343, %f342, %f341, %f340}; ld.local.f32 %f344, [%rd2+-8]; ld.local.f32 %f345, [%rd2+-4]; mul.ftz.f32 %f346, %f636, %f345; mul.ftz.f32 %f347, %f637, %f345; mul.ftz.f32 %f348, %f638, %f345; mul.ftz.f32 %f349, %f639, %f345; fma.rn.ftz.f32 %f350, %f640, %f344, %f346; fma.rn.ftz.f32 %f351, %f641, %f344, %f347; fma.rn.ftz.f32 %f352, %f642, %f344, %f348; fma.rn.ftz.f32 %f353, %f643, %f344, %f349; mov.f32 %f354, 0f3F800000; sub.ftz.f32 %f355, %f354, %f344; sub.ftz.f32 %f356, %f355, %f345; fma.rn.ftz.f32 %f651, %f647, %f356, %f353; fma.rn.ftz.f32 %f650, %f646, %f356, %f352; fma.rn.ftz.f32 %f649, %f645, %f356, %f351; fma.rn.ftz.f32 %f648, %f644, %f356, %f350; $L__BB28_95: setp.eq.s32 %p64, %r64, 1; @%p64 bra $L__BB28_99; bra.uni $L__BB28_96; $L__BB28_99: add.u64 %rd187, %SPL, 16; ld.local.v4.f32 {%f417, %f418, %f419, %f420}, [%rd187]; add.ftz.f32 %f424, %f648, %f417; add.ftz.f32 %f425, %f649, %f418; add.ftz.f32 %f426, %f650, %f419; add.ftz.f32 %f427, %f424, %f425; add.ftz.f32 %f428, %f426, %f427; mul.ftz.f32 %f429, %f428, 0f3EAAAAAB; st.local.f32 [%rd17], %f429; bra.uni $L__BB28_164; $L__BB28_96: setp.eq.s32 %p65, %r64, 0; @%p65 bra $L__BB28_98; add.u64 %rd183, %SPL, 16; ld.local.f32 %f405, [%rd183+12]; add.ftz.f32 %f406, %f651, %f405; st.local.f32 [%rd17], %f406; bra.uni $L__BB28_164; $L__BB28_98: add.u64 %rd185, %SPL, 16; ld.local.v4.f32 {%f407, %f408, %f409, %f410}, [%rd185]; add.ftz.f32 %f414, %f648, %f407; add.ftz.f32 %f415, %f649, %f408; add.ftz.f32 %f416, %f650, %f409; st.local.f32 [%rd17], %f414; st.local.f32 [%rd17+4], %f415; st.local.f32 [%rd17+8], %f416; $L__BB28_164: ret; } .func _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b32 _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2 ) { .reg .pred %p<10>; .reg .f32 %f<80>; .reg .b32 %r<4>; .reg .b64 %rd<23>; ld.param.u64 %rd4, [_Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u32 %r2, [_Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r1, [_Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; cvta.to.local.u64 %rd1, %rd4; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f1, [%rd6]; setp.ge.ftz.f32 %p1, %f1, 0f463B8000; mov.f32 %f76, 0f3FC87CC5; mov.f32 %f75, 0f3F7E9814; mov.f32 %f74, 0f3F53881A; @%p1 bra $L__BB29_7; setp.lt.ftz.f32 %p2, %f1, 0f44480000; mov.f32 %f76, 0fBDA87489; mov.f32 %f75, 0fBE501210; mov.f32 %f74, 0f40AD39B5; @%p2 bra $L__BB29_7; setp.ge.ftz.f32 %p3, %f1, 0f45C6E800; mov.u64 %rd22, 6; @%p3 bra $L__BB29_6; setp.ge.ftz.f32 %p4, %f1, 0f454F3000; mov.u64 %rd22, 5; @%p4 bra $L__BB29_6; setp.ge.ftz.f32 %p5, %f1, 0f44EDC000; mov.u64 %rd22, 4; @%p5 bra $L__BB29_6; setp.ge.ftz.f32 %p6, %f1, 0f44B52000; setp.ge.ftz.f32 %p7, %f1, 0f44714000; selp.u64 %rd10, 1, 0, %p7; setp.ge.ftz.f32 %p8, %f1, 0f4491E000; selp.b64 %rd11, 2, %rd10, %p8; selp.b64 %rd22, 3, %rd11, %p6; $L__BB29_6: mul.lo.s64 %rd12, %rd22, 12; mov.u64 %rd13, blackbody_table_r; add.s64 %rd14, %rd13, %rd12; ld.const.f32 %f26, [%rd14]; rcp.approx.ftz.f32 %f27, %f1; ld.const.f32 %f28, [%rd14+4]; mul.ftz.f32 %f29, %f1, %f28; fma.rn.ftz.f32 %f30, %f27, %f26, %f29; ld.const.f32 %f31, [%rd14+8]; mov.u64 %rd15, blackbody_table_g; add.s64 %rd16, %rd15, %rd12; ld.const.f32 %f32, [%rd16]; ld.const.f32 %f33, [%rd16+4]; mul.ftz.f32 %f34, %f1, %f33; fma.rn.ftz.f32 %f35, %f27, %f32, %f34; ld.const.f32 %f36, [%rd16+8]; shl.b64 %rd17, %rd22, 4; mov.u64 %rd18, blackbody_table_b; add.s64 %rd19, %rd18, %rd17; ld.const.f32 %f37, [%rd19]; ld.const.f32 %f38, [%rd19+4]; fma.rn.ftz.f32 %f39, %f1, %f37, %f38; ld.const.f32 %f40, [%rd19+8]; fma.rn.ftz.f32 %f41, %f1, %f39, %f40; ld.const.f32 %f42, [%rd19+12]; fma.rn.ftz.f32 %f76, %f1, %f41, %f42; add.ftz.f32 %f75, %f36, %f35; add.ftz.f32 %f74, %f31, %f30; $L__BB29_7: ld.const.u32 %r3, [kernel_params+1248]; setp.eq.s32 %p9, %r3, 0; @%p9 bra $L__BB29_8; bra.uni $L__BB29_9; $L__BB29_8: ld.const.v4.f32 {%f43, %f44, %f45, %f46}, [kernel_params+1200]; mul.ftz.f32 %f50, %f75, %f44; fma.rn.ftz.f32 %f51, %f74, %f43, %f50; ld.const.v4.f32 {%f52, %f53, %f54, %f55}, [kernel_params+1216]; mul.ftz.f32 %f59, %f75, %f53; fma.rn.ftz.f32 %f60, %f74, %f52, %f59; ld.const.v4.f32 {%f61, %f62, %f63, %f64}, [kernel_params+1232]; mul.ftz.f32 %f68, %f75, %f62; fma.rn.ftz.f32 %f69, %f74, %f61, %f68; fma.rn.ftz.f32 %f16, %f76, %f63, %f69; fma.rn.ftz.f32 %f75, %f76, %f54, %f60; fma.rn.ftz.f32 %f74, %f76, %f45, %f51; mov.f32 %f76, %f16; $L__BB29_9: mov.f32 %f70, 0f00000000; max.ftz.f32 %f71, %f74, %f70; max.ftz.f32 %f72, %f75, %f70; max.ftz.f32 %f73, %f76, %f70; mul.wide.u32 %rd20, %r1, 4; add.s64 %rd21, %rd1, %rd20; st.local.f32 [%rd21], %f71; st.local.f32 [%rd21+4], %f72; st.local.f32 [%rd21+8], %f73; ret; } .func (.param .b32 func_retval0) _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<18>; .reg .f32 %f<104>; .reg .b32 %r<50>; .reg .b64 %rd<34>; ld.param.u64 %rd2, [_Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r20, [_Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r21, [_Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; ld.param.u32 %r22, [_Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r23, [_Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd2; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r20, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r24, %r25, %r26, %r27}, [%rd6]; ld.global.v4.u32 {%r28, %r29, %r30, %r31}, [%rd6+16]; ld.global.f32 %f94, [%rd6+32]; shr.u32 %r32, %r23, 24; and.b32 %r9, %r22, 255; bfe.u32 %r10, %r22, 8, 8; bfe.u32 %r11, %r22, 16, 8; shr.u32 %r12, %r22, 24; and.b32 %r13, %r21, 255; bfe.u32 %r14, %r21, 8, 8; bfe.u32 %r15, %r21, 16, 8; shr.u32 %r16, %r21, 24; and.b32 %r17, %r24, 255; bfe.u32 %r18, %r24, 8, 8; shl.b32 %r33, %r23, 2; cvt.u64.u32 %rd7, %r33; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f2, [%rd9]; ld.local.f32 %f3, [%rd9+4]; shr.u32 %r34, %r23, 6; cvt.u64.u32 %rd10, %r34; and.b64 %rd11, %rd10, 1020; add.s64 %rd12, %rd1, %rd11; ld.local.f32 %f101, [%rd12]; ld.local.f32 %f102, [%rd12+4]; ld.local.f32 %f103, [%rd12+8]; shr.u32 %r35, %r23, 14; cvt.u64.u32 %rd13, %r35; and.b64 %rd14, %rd13, 1020; add.s64 %rd15, %rd1, %rd14; ld.local.f32 %f7, [%rd15]; ld.local.f32 %f8, [%rd15+4]; ld.local.f32 %f9, [%rd15+8]; mul.wide.u32 %rd16, %r32, 4; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f10, [%rd17]; ld.local.f32 %f11, [%rd17+4]; ld.local.f32 %f12, [%rd17+8]; setp.eq.s32 %p1, %r9, 255; @%p1 bra $L__BB30_2; bra.uni $L__BB30_1; $L__BB30_2: mov.b32 %f92, %r25; bra.uni $L__BB30_3; $L__BB30_1: mul.wide.u32 %rd18, %r9, 4; add.s64 %rd19, %rd1, %rd18; ld.local.f32 %f92, [%rd19]; $L__BB30_3: setp.eq.s32 %p2, %r10, 255; @%p2 bra $L__BB30_5; bra.uni $L__BB30_4; $L__BB30_5: mov.b32 %f93, %r26; bra.uni $L__BB30_6; $L__BB30_4: mul.wide.u32 %rd20, %r10, 4; add.s64 %rd21, %rd1, %rd20; ld.local.f32 %f93, [%rd21]; $L__BB30_6: setp.eq.s32 %p3, %r16, 255; @%p3 bra $L__BB30_8; mul.wide.u32 %rd22, %r16, 4; add.s64 %rd23, %rd1, %rd22; ld.local.f32 %f94, [%rd23]; $L__BB30_8: setp.eq.s32 %p4, %r11, 255; @%p4 bra $L__BB30_10; bra.uni $L__BB30_9; $L__BB30_10: mov.b32 %f95, %r27; bra.uni $L__BB30_11; $L__BB30_9: mul.wide.u32 %rd24, %r11, 4; add.s64 %rd25, %rd1, %rd24; ld.local.f32 %f95, [%rd25]; $L__BB30_11: setp.eq.s32 %p5, %r12, 255; @%p5 bra $L__BB30_13; bra.uni $L__BB30_12; $L__BB30_13: mov.b32 %f98, %r28; bra.uni $L__BB30_14; $L__BB30_12: mul.wide.u32 %rd26, %r12, 4; add.s64 %rd27, %rd1, %rd26; ld.local.f32 %f98, [%rd27]; $L__BB30_14: setp.eq.s32 %p6, %r13, 255; @%p6 bra $L__BB30_16; bra.uni $L__BB30_15; $L__BB30_16: mov.b32 %f97, %r29; bra.uni $L__BB30_17; $L__BB30_15: mul.wide.u32 %rd28, %r13, 4; add.s64 %rd29, %rd1, %rd28; ld.local.f32 %f97, [%rd29]; $L__BB30_17: mul.ftz.f32 %f30, %f2, %f92; mul.ftz.f32 %f31, %f3, %f92; div.approx.ftz.f32 %f48, %f31, %f97; cvt.rmi.ftz.f32.f32 %f49, %f48; cvt.rzi.ftz.s32.f32 %r19, %f49; setp.eq.s32 %p7, %r18, 0; setp.eq.s32 %p8, %r17, 0; mov.f32 %f100, 0f00000000; or.pred %p9, %p8, %p7; mov.f32 %f99, %f100; @%p9 bra $L__BB30_19; mov.b32 %f50, %r31; mov.b32 %f51, %r30; rem.s32 %r36, %r19, %r18; setp.eq.s32 %p10, %r36, 0; selp.f32 %f52, %f50, 0f3F800000, %p10; mul.ftz.f32 %f98, %f98, %f52; rem.s32 %r37, %r19, %r17; setp.eq.s32 %p11, %r37, 0; mul.ftz.f32 %f53, %f98, %f51; selp.f32 %f99, %f53, 0f00000000, %p11; $L__BB30_19: add.ftz.f32 %f55, %f30, %f99; div.approx.ftz.f32 %f56, %f55, %f98; cvt.rmi.ftz.f32.f32 %f57, %f56; cvt.rzi.ftz.s32.f32 %r38, %f57; cvt.rn.f32.s32 %f58, %r38; mul.ftz.f32 %f59, %f98, %f58; sub.ftz.f32 %f60, %f55, %f59; cvt.rn.f32.s32 %f61, %r19; mul.ftz.f32 %f62, %f97, %f61; sub.ftz.f32 %f63, %f31, %f62; and.b32 %r39, %r38, 65535; prmt.b32 %r40, %r19, %r39, 4180; add.s32 %r41, %r40, 1013; and.b32 %r42, %r41, 2147483647; bfe.u32 %r43, %r41, 13, 18; xor.b32 %r44, %r43, %r42; mul.lo.s32 %r45, %r44, 60493; mad.lo.s32 %r46, %r45, %r44, 19990303; mad.lo.s32 %r47, %r46, %r44, 1376312589; and.b32 %r48, %r47, 2147483647; cvt.rn.f32.u32 %f64, %r48; mov.f32 %f65, 0f4E800000; div.approx.ftz.f32 %f66, %f64, %f65; fma.rn.ftz.f32 %f36, %f66, 0f3F000000, %f95; min.ftz.f32 %f67, %f60, %f63; sub.ftz.f32 %f68, %f98, %f60; sub.ftz.f32 %f69, %f97, %f63; min.ftz.f32 %f70, %f68, %f69; min.ftz.f32 %f37, %f67, %f70; setp.ge.ftz.f32 %p12, %f37, %f93; @%p12 bra $L__BB30_23; setp.eq.ftz.f32 %p13, %f94, 0f00000000; mov.f32 %f100, 0f3F800000; @%p13 bra $L__BB30_23; div.approx.ftz.f32 %f73, %f37, %f93; mov.f32 %f100, 0f3F800000; sub.ftz.f32 %f38, %f100, %f73; setp.geu.ftz.f32 %p14, %f38, %f94; @%p14 bra $L__BB30_23; div.approx.ftz.f32 %f74, %f38, %f94; mul.ftz.f32 %f75, %f74, %f74; mul.ftz.f32 %f76, %f75, 0fC0000000; mul.ftz.f32 %f77, %f74, %f76; fma.rn.ftz.f32 %f100, %f75, 0f40400000, %f77; $L__BB30_23: setp.eq.ftz.f32 %p15, %f100, 0f3F800000; @%p15 bra $L__BB30_25; cvt.ftz.sat.f32.f32 %f78, %f36; mov.f32 %f79, 0f3F800000; sub.ftz.f32 %f80, %f79, %f78; mul.ftz.f32 %f81, %f101, %f80; mul.ftz.f32 %f82, %f102, %f80; mul.ftz.f32 %f83, %f103, %f80; fma.rn.ftz.f32 %f101, %f7, %f78, %f81; fma.rn.ftz.f32 %f102, %f8, %f78, %f82; fma.rn.ftz.f32 %f103, %f9, %f78, %f83; $L__BB30_25: setp.eq.s32 %p16, %r14, 255; @%p16 bra $L__BB30_27; mov.f32 %f84, 0f3F800000; sub.ftz.f32 %f85, %f84, %f100; mul.ftz.f32 %f86, %f85, %f101; mul.ftz.f32 %f87, %f85, %f102; mul.ftz.f32 %f88, %f85, %f103; fma.rn.ftz.f32 %f89, %f10, %f100, %f86; fma.rn.ftz.f32 %f90, %f11, %f100, %f87; fma.rn.ftz.f32 %f91, %f12, %f100, %f88; mul.wide.u32 %rd30, %r14, 4; add.s64 %rd31, %rd1, %rd30; st.local.f32 [%rd31], %f89; st.local.f32 [%rd31+4], %f90; st.local.f32 [%rd31+8], %f91; $L__BB30_27: setp.eq.s32 %p17, %r15, 255; @%p17 bra $L__BB30_29; mul.wide.u32 %rd32, %r15, 4; add.s64 %rd33, %rd1, %rd32; st.local.f32 [%rd33], %f100; $L__BB30_29: add.s32 %r49, %r20, 3; st.param.b32 [func_retval0+0], %r49; ret; } .func _Z19svm_node_brightnessP10ShaderDataPfjjj( .param .b64 _Z19svm_node_brightnessP10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_brightnessP10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_brightnessP10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_brightnessP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<2>; .reg .f32 %f<15>; .reg .b32 %r<6>; .reg .b64 %rd<13>; ld.param.u64 %rd2, [_Z19svm_node_brightnessP10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z19svm_node_brightnessP10ShaderDataPfjjj_param_1]; ld.param.u32 %r1, [_Z19svm_node_brightnessP10ShaderDataPfjjj_param_2]; ld.param.u32 %r3, [_Z19svm_node_brightnessP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; mul.wide.u32 %rd3, %r2, 4; add.s64 %rd4, %rd1, %rd3; shl.b32 %r4, %r3, 2; cvt.u64.u32 %rd5, %r4; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; shr.u32 %r5, %r3, 6; cvt.u64.u32 %rd8, %r5; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f4, [%rd10]; add.ftz.f32 %f5, %f4, 0f3F800000; ld.local.f32 %f6, [%rd7]; fma.rn.ftz.f32 %f7, %f4, 0fBF000000, %f6; ld.local.f32 %f8, [%rd4]; fma.rn.ftz.f32 %f1, %f8, %f5, %f7; ld.local.f32 %f9, [%rd4+4]; fma.rn.ftz.f32 %f2, %f9, %f5, %f7; ld.local.f32 %f10, [%rd4+8]; fma.rn.ftz.f32 %f3, %f10, %f5, %f7; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB31_2; mov.f32 %f11, 0f00000000; max.ftz.f32 %f12, %f1, %f11; mul.wide.u32 %rd11, %r1, 4; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f12; max.ftz.f32 %f13, %f2, %f11; st.local.f32 [%rd12+4], %f13; max.ftz.f32 %f14, %f3, %f11; st.local.f32 [%rd12+8], %f14; $L__BB31_2: ret; } .func _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj( .param .b64 _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_0, .param .b64 _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_1, .param .b32 _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_2 ) { .local .align 4 .b8 __local_depot32[24]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<37>; .reg .b16 %rs<10>; .reg .f32 %f<346>; .reg .b32 %r<103>; .reg .b64 %rd<134>; mov.u64 %SPL, __local_depot32; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd45, [_Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_0]; ld.param.u64 %rd46, [_Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_1]; ld.param.u32 %r31, [_Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_2]; cvta.to.local.u64 %rd47, %rd46; cvta.to.local.u64 %rd48, %rd45; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd48]; mul.wide.u32 %rd49, %r31, 4; add.s64 %rd50, %rd47, %rd49; st.local.f32 [%rd50], %f52; st.local.f32 [%rd50+4], %f53; st.local.f32 [%rd50+8], %f54; add.s64 %rd1, %rd48, 92; ld.local.f32 %f59, [%rd48+92]; add.s32 %r32, %r31, 3; mul.wide.u32 %rd51, %r32, 4; add.s64 %rd52, %rd47, %rd51; st.local.f32 [%rd52], %f59; ld.local.u32 %r1, [%rd48+76]; ld.local.s32 %rd5, [%rd48+60]; setp.eq.s32 %p1, %r1, -1; mov.u32 %r22, -1; mov.u32 %r19, 0; mov.u32 %r100, %r19; mov.u32 %r101, %r19; @%p1 bra $L__BB32_15; ld.const.u64 %rd53, [kernel_params+2160]; cvta.to.global.u64 %rd54, %rd53; mul.wide.s32 %rd55, %r1, 208; add.s64 %rd56, %rd54, %rd55; ld.global.u32 %r2, [%rd56+164]; ld.local.u32 %r34, [%rd1+-28]; and.b32 %r35, %r34, 1; setp.eq.b32 %p2, %r35, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r94, 0; @%p5 bra $L__BB32_5; cvt.u32.u64 %r37, %rd5; setp.eq.s32 %p6, %r37, -1; mov.u32 %r93, -1; @%p6 bra $L__BB32_4; ld.const.u64 %rd57, [kernel_params+2240]; cvta.to.global.u64 %rd58, %rd57; shl.b64 %rd59, %rd5, 2; add.s64 %rd60, %rd58, %rd59; ld.global.u32 %r93, [%rd60]; $L__BB32_4: setp.ne.s32 %p7, %r93, -1; selp.u32 %r94, 1, 0, %p7; $L__BB32_5: ld.const.u64 %rd61, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd61; add.s32 %r96, %r94, %r2; mul.wide.u32 %rd62, %r96, 16; add.s64 %rd63, %rd6, %rd62; add.s64 %rd124, %rd63, 14; add.s64 %rd123, %rd63, 15; ld.global.u16 %rs8, [%rd63+12]; ld.global.u32 %r95, [%rd63+8]; ld.global.u64 %rd122, [%rd63]; setp.eq.s64 %p8, %rd122, 10; @%p8 bra $L__BB32_11; $L__BB32_6: setp.eq.s64 %p9, %rd122, 0; @%p9 bra $L__BB32_8; add.s32 %r96, %r96, 2; bra.uni $L__BB32_9; $L__BB32_8: setp.eq.s16 %p10, %rs8, 0; mov.u32 %r19, 0; mov.u32 %r96, %r95; mov.u32 %r100, %r19; mov.u32 %r101, %r19; @%p10 bra $L__BB32_15; $L__BB32_9: cvt.u64.u32 %rd11, %r96; mul.wide.u32 %rd64, %r96, 16; add.s64 %rd65, %rd6, %rd64; ld.global.u16 %rs8, [%rd65+12]; ld.global.u32 %r95, [%rd65+8]; ld.global.u64 %rd122, [%rd65]; setp.ne.s64 %p11, %rd122, 10; @%p11 bra $L__BB32_6; shl.b64 %rd66, %rd11, 4; add.s64 %rd67, %rd6, %rd66; add.s64 %rd124, %rd67, 14; add.s64 %rd123, %rd67, 15; $L__BB32_11: ld.global.u8 %rs5, [%rd124]; ld.global.u8 %rs6, [%rd123]; cvt.u32.u64 %r42, %rd5; setp.ne.s32 %p12, %r42, -1; @%p12 bra $L__BB32_14; mov.u32 %r19, 0; add.s16 %rs7, %rs8, -1; setp.lt.u16 %p13, %rs7, 2; @%p13 bra $L__BB32_14; setp.ne.s16 %p14, %rs8, 1024; mov.u32 %r100, %r19; mov.u32 %r101, %r19; @%p14 bra $L__BB32_15; $L__BB32_14: setp.eq.s16 %p15, %rs8, 0; selp.b32 %r22, -1, %r95, %p15; cvt.u32.u16 %r19, %rs8; cvt.u32.u16 %r47, %rs6; and.b32 %r101, %r47, 255; cvt.u32.u16 %r48, %rs5; and.b32 %r100, %r48, 255; $L__BB32_15: setp.eq.s32 %p16, %r22, -1; @%p16 bra $L__BB32_44; ld.local.u32 %r25, [%rd1+-28]; and.b32 %r49, %r25, 1; setp.eq.b32 %p17, %r49, 1; mov.pred %p18, 0; xor.pred %p19, %p17, %p18; not.pred %p20, %p19; add.u64 %rd68, %SP, 0; add.u64 %rd17, %SPL, 0; @%p20 bra $L__BB32_27; bra.uni $L__BB32_17; $L__BB32_27: and.b32 %r73, %r25, 6; setp.eq.s32 %p27, %r73, 0; @%p27 bra $L__BB32_32; and.b32 %r74, %r19, 768; setp.eq.s32 %p28, %r74, 0; @%p28 bra $L__BB32_30; ld.const.u64 %rd93, [kernel_params+2264]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 4; add.s64 %rd96, %rd94, %rd95; shr.s32 %r75, %r25, 7; ld.global.u32 %r76, [%rd96+4]; add.s32 %r77, %r75, %r76; ld.const.u64 %rd97, [kernel_params+2336]; cvta.to.global.u64 %rd98, %rd97; add.s32 %r78, %r77, %r22; mul.wide.s32 %rd99, %r78, 12; add.s64 %rd100, %rd98, %rd99; ld.global.f32 %f113, [%rd100+12]; ld.global.f32 %f114, [%rd100]; sub.ftz.f32 %f115, %f113, %f114; ld.global.f32 %f116, [%rd100+16]; ld.global.f32 %f117, [%rd100+4]; sub.ftz.f32 %f118, %f116, %f117; ld.global.f32 %f119, [%rd100+20]; ld.global.f32 %f120, [%rd100+8]; sub.ftz.f32 %f121, %f119, %f120; ld.local.f32 %f122, [%rd1+8]; mul.ftz.f32 %f123, %f115, %f122; mul.ftz.f32 %f124, %f118, %f122; mul.ftz.f32 %f125, %f121, %f122; st.local.f32 [%rd17], %f123; st.local.f32 [%rd17+4], %f124; st.local.f32 [%rd17+8], %f125; mov.u32 %r79, 0; st.local.u32 [%rd17+12], %r79; st.local.u32 [%rd17+16], %r79; st.local.u32 [%rd17+20], %r79; ld.local.f32 %f126, [%rd1+-24]; mov.f32 %f127, 0f3F800000; sub.ftz.f32 %f128, %f127, %f126; mul.ftz.f32 %f129, %f114, %f128; mul.ftz.f32 %f130, %f117, %f128; mul.ftz.f32 %f131, %f120, %f128; fma.rn.ftz.f32 %f336, %f119, %f126, %f131; fma.rn.ftz.f32 %f335, %f116, %f126, %f130; fma.rn.ftz.f32 %f334, %f113, %f126, %f129; bra.uni $L__BB32_34; $L__BB32_17: cvt.u32.u64 %r50, %rd5; setp.eq.s32 %p21, %r50, -1; @%p21 bra $L__BB32_20; ld.const.u64 %rd69, [kernel_params+2240]; cvta.to.global.u64 %rd70, %rd69; shl.b64 %rd71, %rd5, 2; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r51, [%rd72]; setp.eq.s32 %p22, %r51, -1; @%p22 bra $L__BB32_20; add.s64 %rd74, %rd68, 12; { // callseq 34, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd45; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r19; st.param.b32 [param1+4], %r100; st.param.b32 [param1+8], %r101; st.param.b32 [param1+12], %r22; .param .b64 param2; st.param.b64 [param2+0], %rd68; .param .b64 param3; st.param.b64 [param3+0], %rd74; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f334, [retval0+0]; ld.param.f32 %f335, [retval0+4]; ld.param.f32 %f336, [retval0+8]; } // callseq 34 bra.uni $L__BB32_34; $L__BB32_20: and.b32 %r52, %r19, 56; setp.eq.s32 %p23, %r52, 0; @%p23 bra $L__BB32_25; and.b32 %r53, %r19, 24; setp.eq.s32 %p24, %r53, 0; @%p24 bra $L__BB32_23; ld.const.u64 %rd75, [kernel_params+2232]; cvta.to.global.u64 %rd76, %rd75; shl.b64 %rd77, %rd5, 4; add.s64 %rd78, %rd76, %rd77; ld.global.v4.u32 {%r54, %r55, %r56, %r57}, [%rd78]; ld.const.u64 %rd79, [kernel_params+2336]; cvta.to.global.u64 %rd80, %rd79; add.s32 %r61, %r54, %r22; mul.wide.u32 %rd81, %r61, 12; add.s64 %rd131, %rd80, %rd81; add.s64 %rd132, %rd131, 4; add.s64 %rd133, %rd131, 8; add.s32 %r62, %r55, %r22; mul.wide.u32 %rd82, %r62, 12; add.s64 %rd128, %rd80, %rd82; add.s64 %rd129, %rd128, 4; add.s64 %rd130, %rd128, 8; add.s32 %r63, %r56, %r22; mul.wide.u32 %rd83, %r63, 12; add.s64 %rd125, %rd80, %rd83; add.s64 %rd126, %rd125, 4; add.s64 %rd127, %rd125, 8; bra.uni $L__BB32_24; $L__BB32_32: and.b32 %r85, %r25, 8; setp.eq.s32 %p31, %r85, 0; mov.u32 %r86, 0; st.local.u32 [%rd17], %r86; st.local.u32 [%rd17+4], %r86; st.local.u32 [%rd17+8], %r86; st.local.u32 [%rd17+12], %r86; st.local.u32 [%rd17+16], %r86; st.local.u32 [%rd17+20], %r86; setp.ne.s32 %p32, %r19, 8; mov.f32 %f334, 0f00000000; or.pred %p33, %p32, %p31; mov.f32 %f335, %f334; mov.f32 %f336, %f334; @%p33 bra $L__BB32_34; cvt.u32.u64 %r87, %rd5; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; add.s32 %r88, %r87, %r22; mul.wide.s32 %rd107, %r88, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f336, [%rd108+8]; ld.global.f32 %f335, [%rd108+4]; ld.global.f32 %f334, [%rd108]; bra.uni $L__BB32_34; $L__BB32_30: mov.u32 %r80, 0; st.local.u32 [%rd17], %r80; st.local.u32 [%rd17+4], %r80; st.local.u32 [%rd17+8], %r80; st.local.u32 [%rd17+12], %r80; st.local.u32 [%rd17+16], %r80; st.local.u32 [%rd17+20], %r80; and.b32 %r81, %r19, 131; setp.eq.s32 %p29, %r81, 0; mov.f32 %f334, 0f00000000; mov.f32 %f335, %f334; mov.f32 %f336, %f334; @%p29 bra $L__BB32_34; cvt.u32.u64 %r82, %rd5; setp.eq.s32 %p30, %r19, 128; selp.b32 %r83, %r82, 0, %p30; add.s32 %r84, %r22, %r83; ld.const.u64 %rd101, [kernel_params+2336]; cvta.to.global.u64 %rd102, %rd101; mul.wide.s32 %rd103, %r84, 12; add.s64 %rd104, %rd102, %rd103; ld.global.f32 %f336, [%rd104+8]; ld.global.f32 %f335, [%rd104+4]; ld.global.f32 %f334, [%rd104]; bra.uni $L__BB32_34; $L__BB32_25: mov.u32 %r68, 0; st.local.u32 [%rd17], %r68; st.local.u32 [%rd17+4], %r68; st.local.u32 [%rd17+8], %r68; st.local.u32 [%rd17+12], %r68; st.local.u32 [%rd17+16], %r68; st.local.u32 [%rd17+20], %r68; and.b32 %r69, %r19, 7; setp.eq.s32 %p25, %r69, 0; mov.f32 %f334, 0f00000000; mov.f32 %f335, %f334; mov.f32 %f336, %f334; @%p25 bra $L__BB32_34; setp.eq.s32 %p26, %r19, 4; selp.b32 %r71, %r50, 0, %p26; add.s32 %r72, %r22, %r71; ld.const.u64 %rd89, [kernel_params+2336]; cvta.to.global.u64 %rd90, %rd89; mul.wide.s32 %rd91, %r72, 12; add.s64 %rd92, %rd90, %rd91; ld.global.f32 %f336, [%rd92+8]; ld.global.f32 %f335, [%rd92+4]; ld.global.f32 %f334, [%rd92]; bra.uni $L__BB32_34; $L__BB32_23: mad.lo.s32 %r65, %r50, 3, %r22; ld.const.u64 %rd84, [kernel_params+2336]; cvta.to.global.u64 %rd85, %rd84; mul.wide.s32 %rd86, %r65, 12; add.s64 %rd131, %rd85, %rd86; add.s64 %rd132, %rd131, 4; add.s64 %rd133, %rd131, 8; add.s32 %r66, %r65, 1; mul.wide.s32 %rd87, %r66, 12; add.s64 %rd128, %rd85, %rd87; add.s64 %rd129, %rd128, 4; add.s64 %rd130, %rd128, 8; add.s32 %r67, %r65, 2; mul.wide.s32 %rd88, %r67, 12; add.s64 %rd125, %rd85, %rd88; add.s64 %rd126, %rd125, 4; add.s64 %rd127, %rd125, 8; $L__BB32_24: ld.local.f32 %f60, [%rd1+8]; ld.global.f32 %f61, [%rd128]; ld.global.f32 %f62, [%rd129]; ld.global.f32 %f63, [%rd130]; ld.local.f32 %f64, [%rd1+16]; ld.global.f32 %f65, [%rd125]; mul.ftz.f32 %f66, %f65, %f64; ld.global.f32 %f67, [%rd126]; mul.ftz.f32 %f68, %f67, %f64; ld.global.f32 %f69, [%rd127]; mul.ftz.f32 %f70, %f69, %f64; fma.rn.ftz.f32 %f71, %f61, %f60, %f66; fma.rn.ftz.f32 %f72, %f62, %f60, %f68; fma.rn.ftz.f32 %f73, %f63, %f60, %f70; add.ftz.f32 %f74, %f60, %f64; ld.global.f32 %f75, [%rd131]; mul.ftz.f32 %f76, %f75, %f74; ld.global.f32 %f77, [%rd132]; mul.ftz.f32 %f78, %f77, %f74; ld.global.f32 %f79, [%rd133]; mul.ftz.f32 %f80, %f79, %f74; sub.ftz.f32 %f81, %f71, %f76; sub.ftz.f32 %f82, %f72, %f78; sub.ftz.f32 %f83, %f73, %f80; st.local.f32 [%rd17], %f81; st.local.f32 [%rd17+4], %f82; st.local.f32 [%rd17+8], %f83; ld.local.f32 %f84, [%rd1+12]; ld.local.f32 %f85, [%rd1+20]; mul.ftz.f32 %f86, %f65, %f85; mul.ftz.f32 %f87, %f67, %f85; mul.ftz.f32 %f88, %f69, %f85; fma.rn.ftz.f32 %f89, %f61, %f84, %f86; fma.rn.ftz.f32 %f90, %f62, %f84, %f87; fma.rn.ftz.f32 %f91, %f63, %f84, %f88; add.ftz.f32 %f92, %f84, %f85; mul.ftz.f32 %f93, %f75, %f92; mul.ftz.f32 %f94, %f77, %f92; mul.ftz.f32 %f95, %f79, %f92; sub.ftz.f32 %f96, %f89, %f93; sub.ftz.f32 %f97, %f90, %f94; sub.ftz.f32 %f98, %f91, %f95; st.local.f32 [%rd17+12], %f96; st.local.f32 [%rd17+16], %f97; st.local.f32 [%rd17+20], %f98; ld.local.f32 %f99, [%rd1+-24]; ld.local.f32 %f100, [%rd1+-20]; mul.ftz.f32 %f101, %f65, %f100; mul.ftz.f32 %f102, %f67, %f100; mul.ftz.f32 %f103, %f69, %f100; fma.rn.ftz.f32 %f104, %f61, %f99, %f101; fma.rn.ftz.f32 %f105, %f62, %f99, %f102; fma.rn.ftz.f32 %f106, %f63, %f99, %f103; mov.f32 %f107, 0f3F800000; sub.ftz.f32 %f108, %f107, %f99; sub.ftz.f32 %f109, %f108, %f100; fma.rn.ftz.f32 %f336, %f79, %f109, %f106; fma.rn.ftz.f32 %f335, %f77, %f109, %f105; fma.rn.ftz.f32 %f334, %f75, %f109, %f104; $L__BB32_34: ld.local.u32 %r89, [%rd1+-36]; and.b32 %r26, %r89, 2; setp.eq.s32 %p34, %r26, 0; @%p34 bra $L__BB32_36; bra.uni $L__BB32_35; $L__BB32_36: ld.const.u64 %rd109, [kernel_params+2160]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r90, [%rd1+-16]; mul.wide.s32 %rd111, %r90, 208; add.s64 %rd112, %rd110, %rd111; ld.global.v4.f32 {%f171, %f172, %f173, %f174}, [%rd112]; ld.global.v4.f32 {%f179, %f180, %f181, %f182}, [%rd112+16]; ld.global.v4.f32 {%f187, %f188, %f189, %f190}, [%rd112+32]; mul.ftz.f32 %f195, %f335, %f172; fma.rn.ftz.f32 %f196, %f334, %f171, %f195; fma.rn.ftz.f32 %f197, %f336, %f173, %f196; add.ftz.f32 %f339, %f174, %f197; mul.ftz.f32 %f198, %f335, %f180; fma.rn.ftz.f32 %f199, %f334, %f179, %f198; fma.rn.ftz.f32 %f200, %f336, %f181, %f199; add.ftz.f32 %f338, %f182, %f200; mul.ftz.f32 %f201, %f335, %f188; fma.rn.ftz.f32 %f202, %f334, %f187, %f201; fma.rn.ftz.f32 %f203, %f336, %f189, %f202; add.ftz.f32 %f337, %f190, %f203; bra.uni $L__BB32_37; $L__BB32_35: ld.local.v4.f32 {%f138, %f139, %f140, %f141}, [%rd1+52]; mul.ftz.f32 %f146, %f335, %f139; fma.rn.ftz.f32 %f147, %f334, %f138, %f146; fma.rn.ftz.f32 %f148, %f336, %f140, %f147; add.ftz.f32 %f339, %f141, %f148; ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd1+68]; mul.ftz.f32 %f157, %f335, %f150; fma.rn.ftz.f32 %f158, %f334, %f149, %f157; fma.rn.ftz.f32 %f159, %f336, %f151, %f158; add.ftz.f32 %f338, %f152, %f159; ld.local.v4.f32 {%f160, %f161, %f162, %f163}, [%rd1+84]; mul.ftz.f32 %f168, %f335, %f161; fma.rn.ftz.f32 %f169, %f334, %f160, %f168; fma.rn.ftz.f32 %f170, %f336, %f162, %f169; add.ftz.f32 %f337, %f163, %f170; $L__BB32_37: @%p34 bra $L__BB32_39; bra.uni $L__BB32_38; $L__BB32_39: ld.const.u64 %rd113, [kernel_params+2160]; cvta.to.global.u64 %rd114, %rd113; ld.local.u32 %r91, [%rd1+-16]; mul.wide.s32 %rd115, %r91, 208; add.s64 %rd116, %rd114, %rd115; ld.global.v4.f32 {%f234, %f235, %f236, %f237}, [%rd116]; ld.global.v4.f32 {%f241, %f242, %f243, %f244}, [%rd116+16]; ld.global.v4.f32 {%f248, %f249, %f250, %f251}, [%rd116+32]; ld.local.f32 %f255, [%rd17]; ld.local.f32 %f256, [%rd17+4]; mul.ftz.f32 %f257, %f235, %f256; fma.rn.ftz.f32 %f258, %f255, %f234, %f257; ld.local.f32 %f259, [%rd17+8]; fma.rn.ftz.f32 %f340, %f236, %f259, %f258; mul.ftz.f32 %f260, %f242, %f256; fma.rn.ftz.f32 %f261, %f255, %f241, %f260; fma.rn.ftz.f32 %f341, %f243, %f259, %f261; mul.ftz.f32 %f262, %f249, %f256; fma.rn.ftz.f32 %f263, %f255, %f248, %f262; fma.rn.ftz.f32 %f342, %f250, %f259, %f263; bra.uni $L__BB32_40; $L__BB32_38: ld.local.f32 %f204, [%rd17]; ld.local.v4.f32 {%f205, %f206, %f207, %f208}, [%rd1+52]; ld.local.f32 %f212, [%rd17+4]; mul.ftz.f32 %f213, %f212, %f206; fma.rn.ftz.f32 %f214, %f204, %f205, %f213; ld.local.f32 %f215, [%rd17+8]; fma.rn.ftz.f32 %f340, %f215, %f207, %f214; ld.local.v4.f32 {%f216, %f217, %f218, %f219}, [%rd1+68]; mul.ftz.f32 %f223, %f212, %f217; fma.rn.ftz.f32 %f224, %f204, %f216, %f223; fma.rn.ftz.f32 %f341, %f215, %f218, %f224; ld.local.v4.f32 {%f225, %f226, %f227, %f228}, [%rd1+84]; mul.ftz.f32 %f232, %f212, %f226; fma.rn.ftz.f32 %f233, %f204, %f225, %f232; fma.rn.ftz.f32 %f342, %f215, %f227, %f233; $L__BB32_40: st.local.f32 [%rd17], %f340; st.local.f32 [%rd17+4], %f341; st.local.f32 [%rd17+8], %f342; @%p34 bra $L__BB32_42; bra.uni $L__BB32_41; $L__BB32_42: ld.const.u64 %rd117, [kernel_params+2160]; cvta.to.global.u64 %rd118, %rd117; ld.local.u32 %r92, [%rd1+-16]; mul.wide.s32 %rd119, %r92, 208; add.s64 %rd120, %rd118, %rd119; ld.global.v4.f32 {%f294, %f295, %f296, %f297}, [%rd120]; ld.global.v4.f32 {%f301, %f302, %f303, %f304}, [%rd120+16]; ld.global.v4.f32 {%f308, %f309, %f310, %f311}, [%rd120+32]; ld.local.f32 %f315, [%rd17+12]; ld.local.f32 %f316, [%rd17+16]; mul.ftz.f32 %f317, %f295, %f316; fma.rn.ftz.f32 %f318, %f315, %f294, %f317; ld.local.f32 %f319, [%rd17+20]; fma.rn.ftz.f32 %f343, %f296, %f319, %f318; mul.ftz.f32 %f320, %f302, %f316; fma.rn.ftz.f32 %f321, %f315, %f301, %f320; fma.rn.ftz.f32 %f344, %f303, %f319, %f321; mul.ftz.f32 %f322, %f309, %f316; fma.rn.ftz.f32 %f323, %f315, %f308, %f322; fma.rn.ftz.f32 %f345, %f310, %f319, %f323; bra.uni $L__BB32_43; $L__BB32_41: ld.local.v4.f32 {%f264, %f265, %f266, %f267}, [%rd1+52]; ld.local.f32 %f271, [%rd17+12]; ld.local.f32 %f272, [%rd17+16]; mul.ftz.f32 %f273, %f272, %f265; fma.rn.ftz.f32 %f274, %f271, %f264, %f273; ld.local.f32 %f275, [%rd17+20]; fma.rn.ftz.f32 %f343, %f275, %f266, %f274; ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd1+68]; mul.ftz.f32 %f283, %f272, %f277; fma.rn.ftz.f32 %f284, %f271, %f276, %f283; fma.rn.ftz.f32 %f344, %f275, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd1+84]; mul.ftz.f32 %f292, %f272, %f286; fma.rn.ftz.f32 %f293, %f271, %f285, %f292; fma.rn.ftz.f32 %f345, %f275, %f287, %f293; $L__BB32_43: st.local.f32 [%rd17+12], %f343; st.local.f32 [%rd17+16], %f344; st.local.f32 [%rd17+20], %f345; st.local.v2.f32 [%rd48], {%f339, %f338}; st.local.f32 [%rd1+-84], %f337; mul.ftz.f32 %f324, %f341, %f341; fma.rn.ftz.f32 %f325, %f340, %f340, %f324; fma.rn.ftz.f32 %f326, %f342, %f342, %f325; sqrt.approx.ftz.f32 %f327, %f326; mul.ftz.f32 %f328, %f344, %f344; fma.rn.ftz.f32 %f329, %f343, %f343, %f328; fma.rn.ftz.f32 %f330, %f345, %f345, %f329; sqrt.approx.ftz.f32 %f331, %f330; add.ftz.f32 %f332, %f327, %f331; mul.ftz.f32 %f333, %f332, 0f3F000000; st.local.f32 [%rd1], %f333; $L__BB32_44: ret; } .func _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj( .param .b64 _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_0, .param .b64 _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_1, .param .b32 _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_2 ) { .reg .f32 %f<5>; .reg .b32 %r<3>; .reg .b64 %rd<9>; ld.param.u64 %rd1, [_Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_0]; ld.param.u64 %rd2, [_Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_1]; ld.param.u32 %r1, [_Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj_param_2]; cvta.to.local.u64 %rd3, %rd2; mul.wide.u32 %rd4, %r1, 4; add.s64 %rd5, %rd3, %rd4; ld.local.f32 %f1, [%rd5+8]; ld.local.f32 %f2, [%rd5+4]; ld.local.f32 %f3, [%rd5]; cvta.to.local.u64 %rd6, %rd1; st.local.v2.f32 [%rd6], {%f3, %f2}; st.local.f32 [%rd6+8], %f1; add.s32 %r2, %r1, 3; mul.wide.u32 %rd7, %r2, 4; add.s64 %rd8, %rd3, %rd7; ld.local.f32 %f4, [%rd8]; st.local.f32 [%rd6+92], %f4; ret; } .func _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .align 4 .b8 _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0[12], .param .b64 _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3, .param .b32 _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4 ) { .reg .pred %p<4>; .reg .f32 %f<47>; .reg .b32 %r<4>; .reg .b64 %rd<9>; ld.param.u64 %rd2, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r1, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4]; ld.param.f32 %f5, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0+8]; ld.param.f32 %f6, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.f32 %f7, [_Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0+4]; cvta.to.local.u64 %rd1, %rd2; ld.const.v4.f32 {%f8, %f9, %f10, %f11}, [kernel_params+688]; ld.const.v4.f32 {%f16, %f17, %f18, %f19}, [kernel_params+704]; ld.const.v4.f32 {%f24, %f25, %f26, %f27}, [kernel_params+720]; mul.ftz.f32 %f32, %f7, %f9; fma.rn.ftz.f32 %f33, %f6, %f8, %f32; fma.rn.ftz.f32 %f34, %f5, %f10, %f33; add.ftz.f32 %f1, %f11, %f34; mul.ftz.f32 %f35, %f7, %f17; fma.rn.ftz.f32 %f36, %f6, %f16, %f35; fma.rn.ftz.f32 %f37, %f5, %f18, %f36; add.ftz.f32 %f2, %f19, %f37; mul.ftz.f32 %f38, %f7, %f25; fma.rn.ftz.f32 %f39, %f6, %f24, %f38; fma.rn.ftz.f32 %f40, %f5, %f26, %f39; add.ftz.f32 %f3, %f27, %f40; mul.ftz.f32 %f41, %f2, %f2; fma.rn.ftz.f32 %f42, %f1, %f1, %f41; fma.rn.ftz.f32 %f43, %f3, %f3, %f42; sqrt.approx.ftz.f32 %f4, %f43; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB34_2; div.approx.ftz.f32 %f44, %f1, %f4; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; st.local.f32 [%rd4], %f44; div.approx.ftz.f32 %f45, %f2, %f4; st.local.f32 [%rd4+4], %f45; div.approx.ftz.f32 %f46, %f3, %f4; st.local.f32 [%rd4+8], %f46; $L__BB34_2: setp.eq.s32 %p2, %r2, 255; @%p2 bra $L__BB34_4; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd1, %rd5; st.local.f32 [%rd6], %f3; $L__BB34_4: setp.eq.s32 %p3, %r3, 255; @%p3 bra $L__BB34_6; mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; st.local.f32 [%rd8], %f4; $L__BB34_6: ret; } .func _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<7>; .reg .f32 %f<27>; .reg .b32 %r<30>; .reg .b64 %rd<18>; ld.param.u64 %rd2, [_Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r9, [_Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; ld.param.u32 %r8, [_Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r7, [_Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r7, 24; and.b32 %r2, %r8, 255; bfe.u32 %r3, %r8, 8, 8; shl.b32 %r10, %r7, 2; cvt.u64.u32 %rd3, %r10; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f1, [%rd5]; ld.local.f32 %f2, [%rd5+4]; ld.local.f32 %f3, [%rd5+8]; shr.u32 %r11, %r7, 6; cvt.u64.u32 %rd6, %r11; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f6, [%rd8+8]; ld.local.f32 %f5, [%rd8+4]; ld.local.f32 %f4, [%rd8]; shr.u32 %r12, %r7, 14; cvt.u64.u32 %rd9, %r12; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f9, [%rd11+8]; ld.local.f32 %f8, [%rd11+4]; ld.local.f32 %f7, [%rd11]; setp.eq.s32 %p2, %r1, 255; @%p2 bra $L__BB35_2; bra.uni $L__BB35_1; $L__BB35_2: mov.b32 %f26, %r9; bra.uni $L__BB35_3; $L__BB35_1: mul.wide.u32 %rd12, %r1, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f26, [%rd13]; $L__BB35_3: fma.rn.ftz.f32 %f13, %f1, %f26, 0f358637BD; mul.ftz.f32 %f14, %f13, 0f3F7FFFEF; fma.rn.ftz.f32 %f15, %f2, %f26, 0f358637BD; mul.ftz.f32 %f16, %f15, 0f3F7FFFEF; fma.rn.ftz.f32 %f17, %f3, %f26, 0f358637BD; mul.ftz.f32 %f18, %f17, 0f3F7FFFEF; cvt.rmi.ftz.f32.f32 %f19, %f14; cvt.rzi.ftz.s32.f32 %r13, %f19; cvt.rmi.ftz.f32.f32 %f20, %f16; cvt.rzi.ftz.s32.f32 %r14, %f20; cvt.rmi.ftz.f32.f32 %f21, %f18; cvt.rzi.ftz.s32.f32 %r15, %f21; abs.s32 %r16, %r13; shr.u32 %r17, %r16, 31; add.s32 %r18, %r16, %r17; and.b32 %r19, %r18, -2; sub.s32 %r20, %r16, %r19; abs.s32 %r21, %r14; shr.u32 %r22, %r21, 31; add.s32 %r23, %r21, %r22; and.b32 %r24, %r23, -2; sub.s32 %r25, %r21, %r24; setp.eq.s32 %p3, %r20, %r25; selp.u32 %r4, 1, 0, %p3; abs.s32 %r26, %r15; shr.u32 %r27, %r26, 31; add.s32 %r28, %r26, %r27; and.b32 %r29, %r28, -2; sub.s32 %r5, %r26, %r29; setp.eq.s32 %p4, %r2, 255; @%p4 bra $L__BB35_5; setp.eq.s32 %p5, %r5, %r4; selp.f32 %f22, %f6, %f9, %p5; selp.f32 %f23, %f5, %f8, %p5; selp.f32 %f24, %f4, %f7, %p5; mul.wide.u32 %rd14, %r2, 4; add.s64 %rd15, %rd1, %rd14; st.local.f32 [%rd15], %f24; st.local.f32 [%rd15+4], %f23; st.local.f32 [%rd15+8], %f22; $L__BB35_5: setp.eq.s32 %p6, %r3, 255; setp.eq.s32 %p1, %r5, %r4; @%p6 bra $L__BB35_7; mul.wide.u32 %rd16, %r3, 4; add.s64 %rd17, %rd1, %rd16; selp.f32 %f25, 0f3F800000, 0f00000000, %p1; st.local.f32 [%rd17], %f25; $L__BB35_7: ret; } .func (.param .b32 func_retval0) _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<6>; .reg .f32 %f<14>; .reg .b32 %r<17>; .reg .b64 %rd<15>; ld.param.u64 %rd2, [_Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r10, [_Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r7, [_Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r8, [_Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r9, [_Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; bfe.u32 %r1, %r7, 8, 8; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r9, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r11, %r12, %r13, %r14}, [%rd6]; mul.wide.u32 %rd7, %r10, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f1, [%rd8]; and.b32 %r6, %r7, 255; setp.eq.s32 %p1, %r6, 255; @%p1 bra $L__BB36_2; bra.uni $L__BB36_1; $L__BB36_2: mov.b32 %f12, %r11; bra.uni $L__BB36_3; $L__BB36_1: mul.wide.u32 %rd9, %r6, 4; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f12, [%rd10]; $L__BB36_3: setp.eq.s32 %p2, %r1, 255; @%p2 bra $L__BB36_5; bra.uni $L__BB36_4; $L__BB36_5: mov.b32 %f13, %r12; bra.uni $L__BB36_6; $L__BB36_4: mul.wide.u32 %rd11, %r1, 4; add.s64 %rd12, %rd1, %rd11; ld.local.f32 %f13, [%rd12]; $L__BB36_6: setp.gt.ftz.f32 %p3, %f12, %f13; and.b32 %r15, %r7, 16711680; setp.eq.s32 %p4, %r15, 65536; and.pred %p5, %p4, %p3; mul.wide.u32 %rd13, %r8, 4; add.s64 %rd14, %rd1, %rd13; selp.f32 %f8, %f12, %f13, %p5; selp.f32 %f9, %f13, %f12, %p5; max.ftz.f32 %f10, %f1, %f9; min.ftz.f32 %f11, %f10, %f8; st.local.f32 [%rd14], %f11; add.s32 %r16, %r9, 1; st.param.b32 [func_retval0+0], %r16; ret; } .func _Z25svm_node_closure_emissionP10ShaderDataPf5uint4( .param .b64 _Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_0, .param .b64 _Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_1, .param .align 16 .b8 _Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<4>; .reg .f32 %f<20>; .reg .b32 %r<5>; .reg .b64 %rd<8>; ld.param.u64 %rd3, [_Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd2, [_Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r1, [_Z25svm_node_closure_emissionP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd4, %rd3; add.s64 %rd1, %rd4, 268; ld.local.f32 %f19, [%rd4+268]; ld.local.f32 %f18, [%rd4+272]; ld.local.f32 %f17, [%rd4+276]; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB37_3; cvta.to.local.u64 %rd5, %rd2; mul.wide.u32 %rd6, %r1, 4; add.s64 %rd7, %rd5, %rd6; ld.local.f32 %f4, [%rd7]; setp.eq.ftz.f32 %p2, %f4, 0f00000000; @%p2 bra $L__BB37_6; mul.ftz.f32 %f19, %f19, %f4; mul.ftz.f32 %f18, %f18, %f4; mul.ftz.f32 %f17, %f17, %f4; $L__BB37_3: ld.local.u32 %r2, [%rd1+-216]; and.b32 %r3, %r2, 2; setp.eq.s32 %p3, %r3, 0; @%p3 bra $L__BB37_5; bra.uni $L__BB37_4; $L__BB37_5: or.b32 %r4, %r2, 2; st.local.u32 [%rd1+-216], %r4; st.local.v2.f32 [%rd1+12], {%f19, %f18}; st.local.f32 [%rd1+20], %f17; bra.uni $L__BB37_6; $L__BB37_4: ld.local.f32 %f11, [%rd1+12]; add.ftz.f32 %f12, %f19, %f11; ld.local.f32 %f13, [%rd1+16]; add.ftz.f32 %f14, %f18, %f13; ld.local.f32 %f15, [%rd1+20]; add.ftz.f32 %f16, %f17, %f15; st.local.f32 [%rd1+12], %f12; st.local.f32 [%rd1+16], %f14; st.local.f32 [%rd1+20], %f16; $L__BB37_6: ret; } .func _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4( .param .b64 _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_0, .param .b64 _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_1, .param .align 16 .b8 _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<4>; .reg .f32 %f<20>; .reg .b32 %r<5>; .reg .b64 %rd<8>; ld.param.u64 %rd3, [_Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd2, [_Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r1, [_Z27svm_node_closure_backgroundP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd4, %rd3; add.s64 %rd1, %rd4, 268; ld.local.f32 %f19, [%rd4+268]; ld.local.f32 %f18, [%rd4+272]; ld.local.f32 %f17, [%rd4+276]; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB38_3; cvta.to.local.u64 %rd5, %rd2; mul.wide.u32 %rd6, %r1, 4; add.s64 %rd7, %rd5, %rd6; ld.local.f32 %f4, [%rd7]; setp.eq.ftz.f32 %p2, %f4, 0f00000000; @%p2 bra $L__BB38_6; mul.ftz.f32 %f19, %f19, %f4; mul.ftz.f32 %f18, %f18, %f4; mul.ftz.f32 %f17, %f17, %f4; $L__BB38_3: ld.local.u32 %r2, [%rd1+-216]; and.b32 %r3, %r2, 2; setp.eq.s32 %p3, %r3, 0; @%p3 bra $L__BB38_5; bra.uni $L__BB38_4; $L__BB38_5: or.b32 %r4, %r2, 2; st.local.u32 [%rd1+-216], %r4; st.local.v2.f32 [%rd1+12], {%f19, %f18}; st.local.f32 [%rd1+20], %f17; bra.uni $L__BB38_6; $L__BB38_4: ld.local.f32 %f11, [%rd1+12]; add.ftz.f32 %f12, %f19, %f11; ld.local.f32 %f13, [%rd1+16]; add.ftz.f32 %f14, %f18, %f13; ld.local.f32 %f15, [%rd1+20]; add.ftz.f32 %f16, %f17, %f15; st.local.f32 [%rd1+12], %f12; st.local.f32 [%rd1+16], %f14; st.local.f32 [%rd1+20], %f16; $L__BB38_6: ret; } .func _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4( .param .b64 _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_0, .param .b64 _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_1, .param .align 16 .b8 _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<5>; .reg .f32 %f<8>; .reg .b32 %r<24>; .reg .b64 %rd<12>; ld.param.u64 %rd4, [_Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd3, [_Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r1, [_Z24svm_node_closure_holdoutP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd1, %rd4; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd1, 264; @%p1 bra $L__BB39_4; bra.uni $L__BB39_1; $L__BB39_4: ld.local.u32 %r5, [%rd2]; setp.eq.s32 %p4, %r5, 0; @%p4 bra $L__BB39_6; ld.local.u32 %r15, [%rd2+-4]; mul.wide.s32 %rd10, %r15, 80; add.s64 %rd11, %rd1, %rd10; ld.local.u32 %r16, [%rd2+12]; ld.local.u32 %r17, [%rd2+8]; ld.local.u32 %r18, [%rd2+4]; mov.u32 %r19, 35; st.local.v4.u32 [%rd11+304], {%r18, %r17, %r16, %r19}; add.s32 %r20, %r15, 1; st.local.u32 [%rd2+-4], %r20; add.s32 %r21, %r5, -1; st.local.u32 [%rd2], %r21; bra.uni $L__BB39_6; $L__BB39_1: cvta.to.local.u64 %rd5, %rd3; mul.wide.u32 %rd6, %r1, 4; add.s64 %rd7, %rd5, %rd6; ld.local.f32 %f1, [%rd7]; setp.eq.ftz.f32 %p2, %f1, 0f00000000; @%p2 bra $L__BB39_7; ld.local.v2.u32 {%r6, %r7}, [%rd2]; setp.eq.s32 %p3, %r6, 0; @%p3 bra $L__BB39_6; ld.local.f32 %f2, [%rd2+8]; ld.local.f32 %f3, [%rd2+12]; ld.local.u32 %r8, [%rd2+-4]; mul.wide.s32 %rd8, %r8, 80; add.s64 %rd9, %rd1, %rd8; mov.b32 %f4, %r7; mul.ftz.f32 %f5, %f1, %f4; mul.ftz.f32 %f6, %f1, %f2; mul.ftz.f32 %f7, %f1, %f3; mov.b32 %r9, %f7; mov.b32 %r10, %f6; mov.b32 %r11, %f5; mov.u32 %r12, 35; st.local.v4.u32 [%rd9+304], {%r11, %r10, %r9, %r12}; add.s32 %r13, %r8, 1; st.local.u32 [%rd2+-4], %r13; add.s32 %r14, %r6, -1; st.local.u32 [%rd2], %r14; $L__BB39_6: ld.local.u32 %r22, [%rd2+-212]; or.b32 %r23, %r22, 32; st.local.u32 [%rd2+-212], %r23; $L__BB39_7: ret; } .func _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .f32 %f<8>; .reg .b32 %r<3>; .reg .b64 %rd<9>; ld.param.u64 %rd1, [_Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd2, [_Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r1, [_Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; ld.param.u32 %r2, [_Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; cvta.to.local.u64 %rd3, %rd1; cvta.to.local.u64 %rd4, %rd2; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd4, %rd5; ld.local.f32 %f1, [%rd6]; mul.wide.u32 %rd7, %r1, 4; add.s64 %rd8, %rd4, %rd7; ld.local.f32 %f2, [%rd8]; ld.local.f32 %f3, [%rd8+4]; ld.local.f32 %f4, [%rd8+8]; mul.ftz.f32 %f5, %f1, %f2; mul.ftz.f32 %f6, %f1, %f3; mul.ftz.f32 %f7, %f1, %f4; st.local.f32 [%rd3+268], %f5; st.local.f32 [%rd3+272], %f6; st.local.f32 [%rd3+276], %f7; ret; } .func _Z20svm_node_mix_closureP10ShaderDataPf5uint4( .param .b64 _Z20svm_node_mix_closureP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z20svm_node_mix_closureP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<4>; .reg .f32 %f<11>; .reg .b32 %r<6>; .reg .b64 %rd<12>; ld.param.u64 %rd2, [_Z20svm_node_mix_closureP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r4, [_Z20svm_node_mix_closureP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; bfe.u32 %r1, %r4, 8, 8; bfe.u32 %r2, %r4, 16, 8; shr.u32 %r3, %r4, 24; shl.b32 %r5, %r4, 2; cvt.u64.u32 %rd3, %r5; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f5, [%rd5]; cvt.ftz.sat.f32.f32 %f1, %f5; setp.eq.s32 %p1, %r1, 255; mov.f32 %f10, 0f3F800000; @%p1 bra $L__BB41_2; mul.wide.u32 %rd6, %r1, 4; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f10, [%rd7]; $L__BB41_2: setp.eq.s32 %p2, %r2, 255; @%p2 bra $L__BB41_4; mov.f32 %f6, 0f3F800000; sub.ftz.f32 %f7, %f6, %f1; mul.ftz.f32 %f8, %f7, %f10; mul.wide.u32 %rd8, %r2, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f8; $L__BB41_4: setp.eq.s32 %p3, %r3, 255; @%p3 bra $L__BB41_6; mul.ftz.f32 %f9, %f1, %f10; mul.wide.u32 %rd10, %r3, 4; add.s64 %rd11, %rd1, %rd10; st.local.f32 [%rd11], %f9; $L__BB41_6: ret; } .func _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<12>; .reg .f32 %f<43>; .reg .b32 %r<9>; .reg .b64 %rd<7>; ld.param.u64 %rd3, [_Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r1, [_Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r2, [_Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r3, [_Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd4, %rd3; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd1, %rd4, %rd5; mul.wide.u32 %rd6, %r3, 4; add.s64 %rd2, %rd4, %rd6; setp.gt.s32 %p1, %r1, 3; @%p1 bra $L__BB42_8; setp.gt.s32 %p7, %r1, 1; @%p7 bra $L__BB42_5; setp.eq.s32 %p10, %r1, 0; @%p10 bra $L__BB42_18; setp.eq.s32 %p11, %r1, 1; @%p11 bra $L__BB42_4; bra.uni $L__BB42_19; $L__BB42_4: ld.local.f32 %f41, [%rd1]; cvt.rzi.ftz.s32.f32 %r8, %f41; st.local.u32 [%rd2], %r8; bra.uni $L__BB42_19; $L__BB42_8: setp.gt.s32 %p2, %r1, 5; @%p2 bra $L__BB42_12; setp.eq.s32 %p5, %r1, 4; @%p5 bra $L__BB42_16; setp.eq.s32 %p6, %r1, 5; @%p6 bra $L__BB42_11; bra.uni $L__BB42_19; $L__BB42_11: ld.local.f32 %f3, [%rd1]; ld.local.f32 %f4, [%rd1+4]; add.ftz.f32 %f5, %f3, %f4; ld.local.f32 %f6, [%rd1+8]; add.ftz.f32 %f7, %f6, %f5; mul.ftz.f32 %f8, %f7, 0f3EAAAAAB; cvt.rzi.ftz.s32.f32 %r6, %f8; st.local.u32 [%rd2], %r6; bra.uni $L__BB42_19; $L__BB42_5: setp.eq.s32 %p8, %r1, 2; @%p8 bra $L__BB42_17; setp.eq.s32 %p9, %r1, 3; @%p9 bra $L__BB42_7; bra.uni $L__BB42_19; $L__BB42_7: ld.local.f32 %f15, [%rd1]; ld.const.v4.f32 {%f16, %f17, %f18, %f19}, [kernel_params+1184]; ld.local.f32 %f23, [%rd1+4]; mul.ftz.f32 %f24, %f23, %f17; fma.rn.ftz.f32 %f25, %f15, %f16, %f24; ld.local.f32 %f26, [%rd1+8]; fma.rn.ftz.f32 %f27, %f26, %f18, %f25; cvt.rzi.ftz.s32.f32 %r7, %f27; st.local.u32 [%rd2], %r7; bra.uni $L__BB42_19; $L__BB42_12: setp.eq.s32 %p3, %r1, 6; @%p3 bra $L__BB42_15; setp.ne.s32 %p4, %r1, 7; @%p4 bra $L__BB42_19; ld.local.u32 %r4, [%rd1]; cvt.rn.f32.s32 %f1, %r4; st.local.f32 [%rd2], %f1; st.local.f32 [%rd2+4], %f1; st.local.f32 [%rd2+8], %f1; bra.uni $L__BB42_19; $L__BB42_18: ld.local.f32 %f42, [%rd1]; st.local.f32 [%rd2], %f42; st.local.f32 [%rd2+4], %f42; st.local.f32 [%rd2+8], %f42; bra.uni $L__BB42_19; $L__BB42_16: ld.local.f32 %f9, [%rd1]; ld.local.f32 %f10, [%rd1+4]; add.ftz.f32 %f11, %f9, %f10; ld.local.f32 %f12, [%rd1+8]; add.ftz.f32 %f13, %f12, %f11; mul.ftz.f32 %f14, %f13, 0f3EAAAAAB; st.local.f32 [%rd2], %f14; bra.uni $L__BB42_19; $L__BB42_17: ld.local.f32 %f28, [%rd1]; ld.const.v4.f32 {%f29, %f30, %f31, %f32}, [kernel_params+1184]; ld.local.f32 %f36, [%rd1+4]; mul.ftz.f32 %f37, %f36, %f30; fma.rn.ftz.f32 %f38, %f28, %f29, %f37; ld.local.f32 %f39, [%rd1+8]; fma.rn.ftz.f32 %f40, %f39, %f31, %f38; st.local.f32 [%rd2], %f40; bra.uni $L__BB42_19; $L__BB42_15: ld.local.u32 %r5, [%rd1]; cvt.rn.f32.s32 %f2, %r5; st.local.f32 [%rd2], %f2; $L__BB42_19: ret; } .func _Z16svm_node_fresnelP10ShaderDataPfjjj( .param .b64 _Z16svm_node_fresnelP10ShaderDataPfjjj_param_0, .param .b64 _Z16svm_node_fresnelP10ShaderDataPfjjj_param_1, .param .b32 _Z16svm_node_fresnelP10ShaderDataPfjjj_param_2, .param .b32 _Z16svm_node_fresnelP10ShaderDataPfjjj_param_3, .param .b32 _Z16svm_node_fresnelP10ShaderDataPfjjj_param_4 ) { .reg .pred %p<8>; .reg .f32 %f<45>; .reg .b32 %r<8>; .reg .b64 %rd<13>; ld.param.u64 %rd3, [_Z16svm_node_fresnelP10ShaderDataPfjjj_param_0]; ld.param.u64 %rd4, [_Z16svm_node_fresnelP10ShaderDataPfjjj_param_1]; ld.param.u32 %r2, [_Z16svm_node_fresnelP10ShaderDataPfjjj_param_2]; ld.param.u32 %r3, [_Z16svm_node_fresnelP10ShaderDataPfjjj_param_3]; ld.param.u32 %r4, [_Z16svm_node_fresnelP10ShaderDataPfjjj_param_4]; cvta.to.local.u64 %rd1, %rd4; and.b32 %r1, %r4, 255; setp.eq.s32 %p1, %r2, 255; @%p1 bra $L__BB43_2; bra.uni $L__BB43_1; $L__BB43_2: mov.b32 %f39, %r3; bra.uni $L__BB43_3; $L__BB43_1: mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f39, [%rd6]; $L__BB43_3: cvta.to.local.u64 %rd7, %rd3; add.s64 %rd2, %rd7, 12; setp.eq.s32 %p2, %r1, 255; @%p2 bra $L__BB43_5; bra.uni $L__BB43_4; $L__BB43_5: ld.local.f32 %f42, [%rd2]; ld.local.f32 %f41, [%rd2+4]; ld.local.f32 %f40, [%rd2+8]; bra.uni $L__BB43_6; $L__BB43_4: mul.wide.u32 %rd8, %r1, 4; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f42, [%rd9]; ld.local.f32 %f41, [%rd9+4]; ld.local.f32 %f40, [%rd9+8]; $L__BB43_6: mov.f32 %f20, 0f3727C5AC; max.ftz.f32 %f43, %f39, %f20; ld.local.u32 %r5, [%rd2+40]; and.b32 %r6, %r5, 1; setp.eq.b32 %p3, %r6, 1; mov.pred %p4, 0; xor.pred %p5, %p3, %p4; not.pred %p6, %p5; @%p6 bra $L__BB43_8; rcp.approx.ftz.f32 %f43, %f43; $L__BB43_8: ld.local.f32 %f22, [%rd2+24]; ld.local.f32 %f23, [%rd2+28]; mul.ftz.f32 %f24, %f41, %f23; fma.rn.ftz.f32 %f25, %f42, %f22, %f24; ld.local.f32 %f26, [%rd2+32]; fma.rn.ftz.f32 %f27, %f40, %f26, %f25; abs.ftz.f32 %f16, %f27; fma.rn.ftz.f32 %f28, %f43, %f43, 0fBF800000; mov.f32 %f44, 0f3F800000; fma.rn.ftz.f32 %f17, %f16, %f16, %f28; setp.leu.ftz.f32 %p7, %f17, 0f00000000; @%p7 bra $L__BB43_10; sqrt.approx.ftz.f32 %f29, %f17; sub.ftz.f32 %f30, %f29, %f16; add.ftz.f32 %f31, %f16, %f29; div.approx.ftz.f32 %f32, %f30, %f31; fma.rn.ftz.f32 %f33, %f16, %f31, 0fBF800000; fma.rn.ftz.f32 %f34, %f16, %f30, 0f3F800000; div.approx.ftz.f32 %f35, %f33, %f34; mul.ftz.f32 %f36, %f32, 0f3F000000; mul.ftz.f32 %f37, %f32, %f36; fma.rn.ftz.f32 %f38, %f35, %f35, 0f3F800000; mul.ftz.f32 %f44, %f37, %f38; $L__BB43_10: shr.u32 %r7, %r4, 6; cvt.u64.u32 %rd10, %r7; and.b64 %rd11, %rd10, 1020; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f44; ret; } .func _Z21svm_node_layer_weightP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_1, .param .align 16 .b8 _Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<10>; .reg .f32 %f<72>; .reg .b32 %r<12>; .reg .b64 %rd<13>; ld.param.u64 %rd3, [_Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd4, [_Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r7, [_Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r6, [_Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r5, [_Z21svm_node_layer_weightP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd1, %rd4; bfe.u32 %r2, %r7, 8, 8; setp.eq.s32 %p1, %r5, 255; @%p1 bra $L__BB44_2; bra.uni $L__BB44_1; $L__BB44_2: mov.b32 %f64, %r6; bra.uni $L__BB44_3; $L__BB44_1: mul.wide.u32 %rd5, %r5, 4; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f64, [%rd6]; $L__BB44_3: cvta.to.local.u64 %rd7, %rd3; add.s64 %rd2, %rd7, 12; setp.eq.s32 %p2, %r2, 255; @%p2 bra $L__BB44_5; bra.uni $L__BB44_4; $L__BB44_5: ld.local.f32 %f67, [%rd2]; ld.local.f32 %f66, [%rd2+4]; ld.local.f32 %f65, [%rd2+8]; bra.uni $L__BB44_6; $L__BB44_4: mul.wide.u32 %rd8, %r2, 4; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f67, [%rd9]; ld.local.f32 %f66, [%rd9+4]; ld.local.f32 %f65, [%rd9+8]; $L__BB44_6: and.b32 %r8, %r7, 255; setp.eq.s32 %p3, %r8, 0; @%p3 bra $L__BB44_13; ld.local.f32 %f28, [%rd2+24]; ld.local.f32 %f29, [%rd2+28]; mul.ftz.f32 %f30, %f66, %f29; fma.rn.ftz.f32 %f31, %f67, %f28, %f30; ld.local.f32 %f32, [%rd2+32]; fma.rn.ftz.f32 %f33, %f65, %f32, %f31; abs.ftz.f32 %f69, %f33; setp.eq.ftz.f32 %p4, %f64, 0f3F000000; @%p4 bra $L__BB44_12; mov.f32 %f34, 0f00000000; max.ftz.f32 %f35, %f64, %f34; mov.f32 %f36, 0f3F7FFF58; min.ftz.f32 %f14, %f35, %f36; setp.lt.ftz.f32 %p5, %f14, 0f3F000000; @%p5 bra $L__BB44_10; bra.uni $L__BB44_9; $L__BB44_10: add.ftz.f32 %f68, %f14, %f14; bra.uni $L__BB44_11; $L__BB44_13: mov.f32 %f71, 0f3F800000; sub.ftz.f32 %f44, %f71, %f64; mov.f32 %f45, 0f3727C5AC; max.ftz.f32 %f70, %f44, %f45; ld.local.u32 %r9, [%rd2+40]; and.b32 %r10, %r9, 1; setp.eq.b32 %p6, %r10, 1; mov.pred %p7, 0; xor.pred %p8, %p6, %p7; @%p8 bra $L__BB44_15; rcp.approx.ftz.f32 %f70, %f70; $L__BB44_15: ld.local.f32 %f47, [%rd2+24]; ld.local.f32 %f48, [%rd2+28]; mul.ftz.f32 %f49, %f66, %f48; fma.rn.ftz.f32 %f50, %f67, %f47, %f49; ld.local.f32 %f51, [%rd2+32]; fma.rn.ftz.f32 %f52, %f65, %f51, %f50; abs.ftz.f32 %f24, %f52; fma.rn.ftz.f32 %f53, %f70, %f70, 0fBF800000; fma.rn.ftz.f32 %f25, %f24, %f24, %f53; setp.leu.ftz.f32 %p9, %f25, 0f00000000; @%p9 bra $L__BB44_17; sqrt.approx.ftz.f32 %f54, %f25; sub.ftz.f32 %f55, %f54, %f24; add.ftz.f32 %f56, %f24, %f54; div.approx.ftz.f32 %f57, %f55, %f56; fma.rn.ftz.f32 %f58, %f24, %f56, 0fBF800000; fma.rn.ftz.f32 %f59, %f24, %f55, 0f3F800000; div.approx.ftz.f32 %f60, %f58, %f59; mul.ftz.f32 %f61, %f57, 0f3F000000; mul.ftz.f32 %f62, %f57, %f61; fma.rn.ftz.f32 %f63, %f60, %f60, 0f3F800000; mul.ftz.f32 %f71, %f62, %f63; bra.uni $L__BB44_17; $L__BB44_9: mov.f32 %f37, 0f3F800000; sub.ftz.f32 %f38, %f37, %f14; mov.f32 %f39, 0f3F000000; div.approx.ftz.f32 %f68, %f39, %f38; $L__BB44_11: lg2.approx.ftz.f32 %f40, %f69; mul.ftz.f32 %f41, %f68, %f40; ex2.approx.ftz.f32 %f69, %f41; $L__BB44_12: mov.f32 %f42, 0f3F800000; sub.ftz.f32 %f71, %f42, %f69; $L__BB44_17: shr.u32 %r11, %r7, 14; cvt.u64.u32 %rd10, %r11; and.b64 %rd11, %rd10, 1020; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f71; ret; } .func _Z14svm_node_gammaP10ShaderDataPfjjj( .param .b64 _Z14svm_node_gammaP10ShaderDataPfjjj_param_0, .param .b32 _Z14svm_node_gammaP10ShaderDataPfjjj_param_1, .param .b32 _Z14svm_node_gammaP10ShaderDataPfjjj_param_2, .param .b32 _Z14svm_node_gammaP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<6>; .reg .f32 %f<32>; .reg .b32 %r<4>; .reg .b64 %rd<9>; ld.param.u64 %rd2, [_Z14svm_node_gammaP10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z14svm_node_gammaP10ShaderDataPfjjj_param_1]; ld.param.u32 %r3, [_Z14svm_node_gammaP10ShaderDataPfjjj_param_2]; ld.param.u32 %r1, [_Z14svm_node_gammaP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd1, %rd3; ld.local.f32 %f26, [%rd4]; ld.local.f32 %f27, [%rd4+4]; ld.local.f32 %f28, [%rd4+8]; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f4, [%rd6]; setp.eq.ftz.f32 %p1, %f4, 0f00000000; mov.f32 %f29, 0f3F800000; mov.f32 %f30, %f29; mov.f32 %f31, %f29; @%p1 bra $L__BB45_8; setp.leu.ftz.f32 %p2, %f26, 0f00000000; @%p2 bra $L__BB45_3; lg2.approx.ftz.f32 %f20, %f26; mul.ftz.f32 %f21, %f4, %f20; ex2.approx.ftz.f32 %f26, %f21; $L__BB45_3: setp.leu.ftz.f32 %p3, %f27, 0f00000000; @%p3 bra $L__BB45_5; lg2.approx.ftz.f32 %f22, %f27; mul.ftz.f32 %f23, %f4, %f22; ex2.approx.ftz.f32 %f27, %f23; $L__BB45_5: setp.leu.ftz.f32 %p4, %f28, 0f00000000; @%p4 bra $L__BB45_7; lg2.approx.ftz.f32 %f24, %f28; mul.ftz.f32 %f25, %f4, %f24; ex2.approx.ftz.f32 %f28, %f25; $L__BB45_7: mov.f32 %f29, %f26; mov.f32 %f30, %f27; mov.f32 %f31, %f28; $L__BB45_8: setp.eq.s32 %p5, %r1, 255; @%p5 bra $L__BB45_10; mul.wide.u32 %rd7, %r1, 4; add.s64 %rd8, %rd1, %rd7; st.local.f32 [%rd8], %f29; st.local.f32 [%rd8+4], %f30; st.local.f32 [%rd8+8], %f31; $L__BB45_10: ret; } .func _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<43>; .reg .b16 %rs<12>; .reg .f32 %f<229>; .reg .b32 %r<95>; .reg .b64 %rd<118>; ld.param.u64 %rd44, [_Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd45, [_Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r26, [_Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r27, [_Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd46, %rd44; add.s64 %rd1, %rd46, 68; mov.f32 %f228, 0f00000000; setp.gt.s32 %p1, %r26, 2; @%p1 bra $L__BB46_6; setp.eq.s32 %p5, %r26, 0; @%p5 bra $L__BB46_52; setp.eq.s32 %p6, %r26, 1; @%p6 bra $L__BB46_51; setp.eq.s32 %p7, %r26, 2; mov.f32 %f226, %f228; mov.f32 %f227, %f228; @%p7 bra $L__BB46_4; bra.uni $L__BB46_53; $L__BB46_4: ld.local.u32 %r1, [%rd1+-4]; and.b32 %r28, %r1, 14; setp.eq.s32 %p8, %r28, 0; @%p8 bra $L__BB46_12; ld.local.f32 %f68, [%rd1+48]; ld.local.f32 %f69, [%rd1+52]; mul.ftz.f32 %f70, %f69, %f69; fma.rn.ftz.f32 %f71, %f68, %f68, %f70; ld.local.f32 %f72, [%rd1+56]; fma.rn.ftz.f32 %f73, %f72, %f72, %f71; sqrt.approx.ftz.f32 %f74, %f73; div.approx.ftz.f32 %f228, %f72, %f74; div.approx.ftz.f32 %f227, %f69, %f74; div.approx.ftz.f32 %f226, %f68, %f74; bra.uni $L__BB46_53; $L__BB46_6: setp.eq.s32 %p2, %r26, 3; @%p2 bra $L__BB46_11; setp.eq.s32 %p3, %r26, 4; @%p3 bra $L__BB46_10; setp.ne.s32 %p4, %r26, 5; mov.f32 %f226, %f228; mov.f32 %f227, %f228; @%p4 bra $L__BB46_53; mov.f32 %f63, 0f3F800000; ld.local.f32 %f227, [%rd1]; sub.ftz.f32 %f64, %f63, %f227; ld.local.f32 %f65, [%rd1+4]; sub.ftz.f32 %f226, %f64, %f65; bra.uni $L__BB46_53; $L__BB46_52: ld.local.v4.f32 {%f226, %f227, %f228, %f220}, [%rd46]; bra.uni $L__BB46_53; $L__BB46_51: ld.local.f32 %f228, [%rd1+-48]; ld.local.f32 %f227, [%rd1+-52]; ld.local.f32 %f226, [%rd1+-56]; bra.uni $L__BB46_53; $L__BB46_11: ld.local.f32 %f228, [%rd1+-24]; ld.local.f32 %f227, [%rd1+-28]; ld.local.f32 %f226, [%rd1+-32]; bra.uni $L__BB46_53; $L__BB46_10: ld.local.v2.f32 {%f226, %f227}, [%rd1+-44]; ld.local.f32 %f228, [%rd1+-36]; $L__BB46_53: cvta.to.local.u64 %rd106, %rd45; mul.wide.u32 %rd107, %r27, 4; add.s64 %rd108, %rd106, %rd107; st.local.f32 [%rd108], %f226; st.local.f32 [%rd108+4], %f227; st.local.f32 [%rd108+8], %f228; ret; $L__BB46_12: ld.local.u32 %r2, [%rd1+8]; ld.local.s32 %rd5, [%rd1+-8]; setp.eq.s32 %p9, %r2, -1; mov.u32 %r23, -1; mov.u32 %r20, 0; mov.u32 %r92, %r20; mov.u32 %r93, %r20; @%p9 bra $L__BB46_27; ld.const.u64 %rd47, [kernel_params+2160]; cvta.to.global.u64 %rd48, %rd47; mul.wide.s32 %rd49, %r2, 208; add.s64 %rd50, %rd48, %rd49; ld.global.u32 %r3, [%rd50+164]; and.b32 %r34, %r1, 1; setp.eq.b32 %p10, %r34, 1; mov.pred %p11, 0; xor.pred %p12, %p10, %p11; not.pred %p13, %p12; mov.u32 %r86, 0; @%p13 bra $L__BB46_17; cvt.u32.u64 %r36, %rd5; setp.eq.s32 %p14, %r36, -1; mov.u32 %r85, -1; @%p14 bra $L__BB46_16; ld.const.u64 %rd51, [kernel_params+2240]; cvta.to.global.u64 %rd52, %rd51; shl.b64 %rd53, %rd5, 2; add.s64 %rd54, %rd52, %rd53; ld.global.u32 %r85, [%rd54]; $L__BB46_16: setp.ne.s32 %p15, %r85, -1; selp.u32 %r86, 1, 0, %p15; $L__BB46_17: ld.const.u64 %rd55, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd55; add.s32 %r88, %r86, %r3; mul.wide.u32 %rd56, %r88, 16; add.s64 %rd57, %rd6, %rd56; add.s64 %rd111, %rd57, 14; add.s64 %rd110, %rd57, 15; ld.global.u16 %rs10, [%rd57+12]; ld.global.u32 %r87, [%rd57+8]; ld.global.u64 %rd109, [%rd57]; setp.eq.s64 %p16, %rd109, 7; @%p16 bra $L__BB46_23; $L__BB46_18: setp.eq.s64 %p17, %rd109, 0; @%p17 bra $L__BB46_20; add.s32 %r88, %r88, 2; bra.uni $L__BB46_21; $L__BB46_20: setp.eq.s16 %p18, %rs10, 0; mov.u32 %r20, 0; mov.u32 %r88, %r87; mov.u32 %r92, %r20; mov.u32 %r93, %r20; @%p18 bra $L__BB46_27; $L__BB46_21: cvt.u64.u32 %rd11, %r88; mul.wide.u32 %rd58, %r88, 16; add.s64 %rd59, %rd6, %rd58; ld.global.u16 %rs10, [%rd59+12]; ld.global.u32 %r87, [%rd59+8]; ld.global.u64 %rd109, [%rd59]; setp.ne.s64 %p19, %rd109, 7; @%p19 bra $L__BB46_18; shl.b64 %rd60, %rd11, 4; add.s64 %rd61, %rd6, %rd60; add.s64 %rd111, %rd61, 14; add.s64 %rd110, %rd61, 15; $L__BB46_23: ld.global.u8 %rs5, [%rd111]; ld.global.u8 %rs6, [%rd110]; cvt.u32.u64 %r41, %rd5; setp.ne.s32 %p20, %r41, -1; @%p20 bra $L__BB46_26; mov.u32 %r20, 0; add.s16 %rs7, %rs10, -1; setp.lt.u16 %p21, %rs7, 2; @%p21 bra $L__BB46_26; setp.ne.s16 %p22, %rs10, 1024; mov.u32 %r92, %r20; mov.u32 %r93, %r20; @%p22 bra $L__BB46_27; $L__BB46_26: setp.eq.s16 %p23, %rs10, 0; selp.b32 %r23, -1, %r87, %p23; cvt.u32.u16 %r20, %rs10; cvt.u32.u16 %r46, %rs6; and.b32 %r93, %r46, 255; cvt.u32.u16 %r47, %rs5; and.b32 %r92, %r47, 255; $L__BB46_27: setp.eq.s32 %p24, %r23, -1; @%p24 bra $L__BB46_50; and.b32 %r48, %r1, 1; setp.eq.b32 %p25, %r48, 1; mov.pred %p26, 0; xor.pred %p27, %p25, %p26; not.pred %p28, %p27; @%p28 bra $L__BB46_39; bra.uni $L__BB46_29; $L__BB46_39: and.b32 %r71, %r1, 6; setp.eq.s32 %p35, %r71, 0; @%p35 bra $L__BB46_44; and.b32 %r72, %r20, 768; setp.eq.s32 %p36, %r72, 0; @%p36 bra $L__BB46_42; ld.const.u64 %rd85, [kernel_params+2264]; cvta.to.global.u64 %rd86, %rd85; shl.b64 %rd87, %rd5, 4; add.s64 %rd88, %rd86, %rd87; shr.s32 %r73, %r1, 7; ld.global.u32 %r74, [%rd88+4]; add.s32 %r75, %r73, %r74; ld.const.u64 %rd89, [kernel_params+2336]; cvta.to.global.u64 %rd90, %rd89; add.s32 %r76, %r75, %r23; mul.wide.s32 %rd91, %r76, 12; add.s64 %rd92, %rd90, %rd91; ld.local.f32 %f98, [%rd1]; mov.f32 %f99, 0f3F800000; sub.ftz.f32 %f100, %f99, %f98; ld.global.f32 %f101, [%rd92]; mul.ftz.f32 %f102, %f101, %f100; ld.global.f32 %f103, [%rd92+4]; mul.ftz.f32 %f104, %f103, %f100; ld.global.f32 %f107, [%rd92+12]; ld.global.f32 %f108, [%rd92+16]; fma.rn.ftz.f32 %f222, %f108, %f98, %f104; fma.rn.ftz.f32 %f221, %f107, %f98, %f102; bra.uni $L__BB46_46; $L__BB46_50: ld.local.f32 %f210, [%rd1+48]; ld.local.f32 %f211, [%rd1+52]; mul.ftz.f32 %f212, %f211, %f211; fma.rn.ftz.f32 %f213, %f210, %f210, %f212; ld.local.f32 %f214, [%rd1+56]; fma.rn.ftz.f32 %f215, %f214, %f214, %f213; sqrt.approx.ftz.f32 %f216, %f215; div.approx.ftz.f32 %f228, %f214, %f216; div.approx.ftz.f32 %f227, %f211, %f216; div.approx.ftz.f32 %f226, %f210, %f216; bra.uni $L__BB46_53; $L__BB46_29: cvt.u32.u64 %r49, %rd5; setp.eq.s32 %p29, %r49, -1; @%p29 bra $L__BB46_32; ld.const.u64 %rd62, [kernel_params+2240]; cvta.to.global.u64 %rd63, %rd62; shl.b64 %rd64, %rd5, 2; add.s64 %rd65, %rd63, %rd64; ld.global.u32 %r50, [%rd65]; setp.eq.s32 %p30, %r50, -1; @%p30 bra $L__BB46_32; mov.u64 %rd66, 0; { // callseq 35, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd44; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r20; st.param.b32 [param1+4], %r92; st.param.b32 [param1+8], %r93; st.param.b32 [param1+12], %r23; .param .b64 param2; st.param.b64 [param2+0], %rd66; .param .b64 param3; st.param.b64 [param3+0], %rd66; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f221, [retval0+0]; ld.param.f32 %f222, [retval0+4]; ld.param.f32 %f15, [retval0+8]; } // callseq 35 bra.uni $L__BB46_46; $L__BB46_32: and.b32 %r51, %r20, 56; setp.eq.s32 %p31, %r51, 0; @%p31 bra $L__BB46_37; and.b32 %r52, %r20, 24; setp.eq.s32 %p32, %r52, 0; @%p32 bra $L__BB46_35; ld.const.u64 %rd67, [kernel_params+2232]; cvta.to.global.u64 %rd68, %rd67; shl.b64 %rd69, %rd5, 4; add.s64 %rd70, %rd68, %rd69; ld.global.v4.u32 {%r53, %r54, %r55, %r56}, [%rd70]; ld.const.u64 %rd71, [kernel_params+2336]; cvta.to.global.u64 %rd72, %rd71; add.s32 %r60, %r53, %r23; mul.wide.u32 %rd73, %r60, 12; add.s64 %rd116, %rd72, %rd73; add.s64 %rd117, %rd116, 4; add.s32 %r61, %r54, %r23; mul.wide.u32 %rd74, %r61, 12; add.s64 %rd114, %rd72, %rd74; add.s64 %rd115, %rd114, 4; add.s32 %r62, %r55, %r23; mul.wide.u32 %rd75, %r62, 12; add.s64 %rd112, %rd72, %rd75; add.s64 %rd113, %rd112, 4; bra.uni $L__BB46_36; $L__BB46_44: and.b32 %r81, %r1, 8; setp.eq.s32 %p39, %r81, 0; setp.ne.s32 %p40, %r20, 8; mov.f32 %f221, 0f00000000; or.pred %p41, %p40, %p39; mov.f32 %f222, %f221; @%p41 bra $L__BB46_46; cvt.u32.u64 %r82, %rd5; ld.const.u64 %rd97, [kernel_params+2336]; cvta.to.global.u64 %rd98, %rd97; add.s32 %r83, %r82, %r23; mul.wide.s32 %rd99, %r83, 12; add.s64 %rd100, %rd98, %rd99; ld.global.f32 %f222, [%rd100+4]; ld.global.f32 %f221, [%rd100]; bra.uni $L__BB46_46; $L__BB46_42: and.b32 %r77, %r20, 131; setp.eq.s32 %p37, %r77, 0; mov.f32 %f221, 0f00000000; mov.f32 %f222, %f221; @%p37 bra $L__BB46_46; cvt.u32.u64 %r78, %rd5; setp.eq.s32 %p38, %r20, 128; selp.b32 %r79, %r78, 0, %p38; add.s32 %r80, %r23, %r79; ld.const.u64 %rd93, [kernel_params+2336]; cvta.to.global.u64 %rd94, %rd93; mul.wide.s32 %rd95, %r80, 12; add.s64 %rd96, %rd94, %rd95; ld.global.f32 %f222, [%rd96+4]; ld.global.f32 %f221, [%rd96]; bra.uni $L__BB46_46; $L__BB46_37: and.b32 %r67, %r20, 7; setp.eq.s32 %p33, %r67, 0; mov.f32 %f221, 0f00000000; mov.f32 %f222, %f221; @%p33 bra $L__BB46_46; setp.eq.s32 %p34, %r20, 4; selp.b32 %r69, %r49, 0, %p34; add.s32 %r70, %r23, %r69; ld.const.u64 %rd81, [kernel_params+2336]; cvta.to.global.u64 %rd82, %rd81; mul.wide.s32 %rd83, %r70, 12; add.s64 %rd84, %rd82, %rd83; ld.global.f32 %f222, [%rd84+4]; ld.global.f32 %f221, [%rd84]; bra.uni $L__BB46_46; $L__BB46_35: mad.lo.s32 %r64, %r49, 3, %r23; ld.const.u64 %rd76, [kernel_params+2336]; cvta.to.global.u64 %rd77, %rd76; mul.wide.s32 %rd78, %r64, 12; add.s64 %rd116, %rd77, %rd78; add.s64 %rd117, %rd116, 4; add.s32 %r65, %r64, 1; mul.wide.s32 %rd79, %r65, 12; add.s64 %rd114, %rd77, %rd79; add.s64 %rd115, %rd114, 4; add.s32 %r66, %r64, 2; mul.wide.s32 %rd80, %r66, 12; add.s64 %rd112, %rd77, %rd80; add.s64 %rd113, %rd112, 4; $L__BB46_36: ld.local.f32 %f75, [%rd1]; ld.global.f32 %f76, [%rd114]; ld.global.f32 %f77, [%rd115]; ld.local.f32 %f79, [%rd1+4]; ld.global.f32 %f80, [%rd112]; mul.ftz.f32 %f81, %f80, %f79; ld.global.f32 %f82, [%rd113]; mul.ftz.f32 %f83, %f82, %f79; fma.rn.ftz.f32 %f86, %f76, %f75, %f81; fma.rn.ftz.f32 %f87, %f77, %f75, %f83; mov.f32 %f89, 0f3F800000; sub.ftz.f32 %f90, %f89, %f75; sub.ftz.f32 %f91, %f90, %f79; ld.global.f32 %f92, [%rd116]; ld.global.f32 %f93, [%rd117]; fma.rn.ftz.f32 %f222, %f93, %f91, %f87; fma.rn.ftz.f32 %f221, %f92, %f91, %f86; $L__BB46_46: add.ftz.f32 %f116, %f222, 0fBF000000; neg.ftz.f32 %f34, %f116; add.ftz.f32 %f35, %f221, 0fBF000000; ld.local.u8 %rs8, [%rd1+-12]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p42, %rs9, 0; @%p42 bra $L__BB46_48; bra.uni $L__BB46_47; $L__BB46_48: ld.const.u64 %rd101, [kernel_params+2160]; cvta.to.global.u64 %rd102, %rd101; ld.local.u32 %r84, [%rd1+8]; mul.wide.s32 %rd103, %r84, 208; add.s64 %rd104, %rd102, %rd103; ld.global.v4.f32 {%f151, %f152, %f153, %f154}, [%rd104+48]; ld.global.v4.f32 {%f158, %f159, %f160, %f161}, [%rd104+64]; ld.global.v4.f32 {%f165, %f166, %f167, %f168}, [%rd104+80]; mul.ftz.f32 %f172, %f35, %f158; fma.rn.ftz.f32 %f173, %f151, %f34, %f172; fma.rn.ftz.f32 %f174, %f165, 0f00000000, %f173; mul.ftz.f32 %f175, %f35, %f159; fma.rn.ftz.f32 %f176, %f152, %f34, %f175; fma.rn.ftz.f32 %f177, %f166, 0f00000000, %f176; mul.ftz.f32 %f178, %f35, %f160; fma.rn.ftz.f32 %f179, %f153, %f34, %f178; fma.rn.ftz.f32 %f180, %f167, 0f00000000, %f179; mul.ftz.f32 %f181, %f177, %f177; fma.rn.ftz.f32 %f182, %f174, %f174, %f181; fma.rn.ftz.f32 %f183, %f180, %f180, %f182; sqrt.approx.ftz.f32 %f184, %f183; div.approx.ftz.f32 %f225, %f174, %f184; div.approx.ftz.f32 %f224, %f177, %f184; div.approx.ftz.f32 %f223, %f180, %f184; bra.uni $L__BB46_49; $L__BB46_47: ld.local.v4.f32 {%f117, %f118, %f119, %f120}, [%rd1+124]; ld.local.v4.f32 {%f124, %f125, %f126, %f127}, [%rd1+140]; ld.local.v4.f32 {%f131, %f132, %f133, %f134}, [%rd1+156]; mul.ftz.f32 %f138, %f35, %f124; fma.rn.ftz.f32 %f139, %f117, %f34, %f138; fma.rn.ftz.f32 %f140, %f131, 0f00000000, %f139; mul.ftz.f32 %f141, %f35, %f125; fma.rn.ftz.f32 %f142, %f118, %f34, %f141; fma.rn.ftz.f32 %f143, %f132, 0f00000000, %f142; mul.ftz.f32 %f144, %f35, %f126; fma.rn.ftz.f32 %f145, %f119, %f34, %f144; fma.rn.ftz.f32 %f146, %f133, 0f00000000, %f145; mul.ftz.f32 %f147, %f143, %f143; fma.rn.ftz.f32 %f148, %f140, %f140, %f147; fma.rn.ftz.f32 %f149, %f146, %f146, %f148; sqrt.approx.ftz.f32 %f150, %f149; div.approx.ftz.f32 %f225, %f140, %f150; div.approx.ftz.f32 %f224, %f143, %f150; div.approx.ftz.f32 %f223, %f146, %f150; $L__BB46_49: ld.local.f32 %f185, [%rd1+-48]; mul.ftz.f32 %f186, %f224, %f185; ld.local.f32 %f187, [%rd1+-52]; mul.ftz.f32 %f188, %f223, %f187; sub.ftz.f32 %f189, %f186, %f188; ld.local.f32 %f190, [%rd1+-56]; mul.ftz.f32 %f191, %f223, %f190; mul.ftz.f32 %f192, %f225, %f185; sub.ftz.f32 %f193, %f191, %f192; mul.ftz.f32 %f194, %f225, %f187; mul.ftz.f32 %f195, %f224, %f190; sub.ftz.f32 %f196, %f194, %f195; mul.ftz.f32 %f197, %f193, %f193; fma.rn.ftz.f32 %f198, %f189, %f189, %f197; fma.rn.ftz.f32 %f199, %f196, %f196, %f198; sqrt.approx.ftz.f32 %f200, %f199; div.approx.ftz.f32 %f201, %f189, %f200; div.approx.ftz.f32 %f202, %f193, %f200; div.approx.ftz.f32 %f203, %f196, %f200; mul.ftz.f32 %f204, %f203, %f187; mul.ftz.f32 %f205, %f202, %f185; mul.ftz.f32 %f206, %f201, %f185; mul.ftz.f32 %f207, %f203, %f190; mul.ftz.f32 %f208, %f202, %f190; mul.ftz.f32 %f209, %f201, %f187; sub.ftz.f32 %f228, %f208, %f209; sub.ftz.f32 %f227, %f206, %f207; sub.ftz.f32 %f226, %f204, %f205; bra.uni $L__BB46_53; } .func _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<6>; .reg .f32 %f<42>; .reg .b32 %r<3>; .reg .b64 %rd<7>; ld.param.u64 %rd2, [_Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd3, [_Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r1, [_Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r2, [_Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd1, %rd2; setp.eq.s32 %p1, %r1, 0; @%p1 bra $L__BB47_3; setp.ne.s32 %p2, %r1, 5; @%p2 bra $L__BB47_5; ld.local.f32 %f10, [%rd1+68]; mov.f32 %f11, 0f3F800000; sub.ftz.f32 %f12, %f11, %f10; ld.local.f32 %f13, [%rd1+100]; sub.ftz.f32 %f14, %f12, %f13; ld.local.f32 %f15, [%rd1+72]; sub.ftz.f32 %f16, %f14, %f15; ld.local.f32 %f17, [%rd1+108]; add.ftz.f32 %f40, %f10, %f13; sub.ftz.f32 %f39, %f16, %f17; mov.f32 %f41, 0f00000000; bra.uni $L__BB47_4; $L__BB47_3: ld.local.f32 %f18, [%rd1+28]; ld.local.f32 %f19, [%rd1+24]; setp.neu.ftz.f32 %p3, %f19, %f18; ld.local.f32 %f20, [%rd1+32]; setp.neu.ftz.f32 %p4, %f19, %f20; or.pred %p5, %p3, %p4; sub.ftz.f32 %f21, %f20, %f18; neg.ftz.f32 %f22, %f18; neg.ftz.f32 %f23, %f20; selp.f32 %f24, %f23, %f20, %p5; add.ftz.f32 %f25, %f19, %f24; selp.f32 %f26, %f18, %f22, %p5; sub.ftz.f32 %f27, %f26, %f19; mul.ftz.f32 %f28, %f25, %f25; fma.rn.ftz.f32 %f29, %f21, %f21, %f28; fma.rn.ftz.f32 %f30, %f27, %f27, %f29; sqrt.approx.ftz.f32 %f31, %f30; div.approx.ftz.f32 %f32, %f21, %f31; div.approx.ftz.f32 %f33, %f25, %f31; div.approx.ftz.f32 %f34, %f27, %f31; ld.local.f32 %f35, [%rd1+92]; ld.local.f32 %f36, [%rd1]; ld.local.f32 %f37, [%rd1+4]; ld.local.f32 %f38, [%rd1+8]; fma.rn.ftz.f32 %f41, %f35, %f34, %f38; fma.rn.ftz.f32 %f40, %f35, %f33, %f37; fma.rn.ftz.f32 %f39, %f35, %f32, %f36; $L__BB47_4: cvta.to.local.u64 %rd4, %rd3; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd4, %rd5; st.local.f32 [%rd6], %f39; st.local.f32 [%rd6+4], %f40; st.local.f32 [%rd6+8], %f41; bra.uni $L__BB47_6; $L__BB47_5: { // callseq 36, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2; .param .b64 param1; st.param.b64 [param1+0], %rd3; .param .b32 param2; st.param.b32 [param2+0], %r1; .param .b32 param3; st.param.b32 [param3+0], %r2; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 36 $L__BB47_6: ret; } .func _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<6>; .reg .f32 %f<51>; .reg .b32 %r<3>; .reg .b64 %rd<7>; ld.param.u64 %rd2, [_Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd3, [_Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r1, [_Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r2, [_Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd1, %rd2; setp.eq.s32 %p1, %r1, 0; @%p1 bra $L__BB48_3; setp.ne.s32 %p2, %r1, 5; @%p2 bra $L__BB48_5; ld.local.f32 %f10, [%rd1+68]; mov.f32 %f11, 0f3F800000; sub.ftz.f32 %f12, %f11, %f10; ld.local.f32 %f13, [%rd1+104]; sub.ftz.f32 %f14, %f12, %f13; ld.local.f32 %f15, [%rd1+72]; sub.ftz.f32 %f16, %f14, %f15; ld.local.f32 %f17, [%rd1+112]; add.ftz.f32 %f49, %f10, %f13; sub.ftz.f32 %f48, %f16, %f17; mov.f32 %f50, 0f00000000; bra.uni $L__BB48_4; $L__BB48_3: ld.local.f32 %f18, [%rd1+28]; ld.local.f32 %f19, [%rd1+24]; setp.neu.ftz.f32 %p3, %f19, %f18; ld.local.f32 %f20, [%rd1+32]; setp.neu.ftz.f32 %p4, %f19, %f20; or.pred %p5, %p3, %p4; sub.ftz.f32 %f21, %f20, %f18; neg.ftz.f32 %f22, %f18; neg.ftz.f32 %f23, %f20; selp.f32 %f24, %f23, %f20, %p5; add.ftz.f32 %f25, %f19, %f24; selp.f32 %f26, %f18, %f22, %p5; sub.ftz.f32 %f27, %f26, %f19; mul.ftz.f32 %f28, %f25, %f25; fma.rn.ftz.f32 %f29, %f21, %f21, %f28; fma.rn.ftz.f32 %f30, %f27, %f27, %f29; sqrt.approx.ftz.f32 %f31, %f30; div.approx.ftz.f32 %f32, %f21, %f31; div.approx.ftz.f32 %f33, %f25, %f31; div.approx.ftz.f32 %f34, %f27, %f31; mul.ftz.f32 %f35, %f18, %f34; mul.ftz.f32 %f36, %f20, %f33; sub.ftz.f32 %f37, %f35, %f36; mul.ftz.f32 %f38, %f20, %f32; mul.ftz.f32 %f39, %f19, %f34; sub.ftz.f32 %f40, %f38, %f39; mul.ftz.f32 %f41, %f19, %f33; mul.ftz.f32 %f42, %f18, %f32; sub.ftz.f32 %f43, %f41, %f42; ld.local.f32 %f44, [%rd1+92]; ld.local.f32 %f45, [%rd1]; ld.local.f32 %f46, [%rd1+4]; ld.local.f32 %f47, [%rd1+8]; fma.rn.ftz.f32 %f50, %f44, %f43, %f47; fma.rn.ftz.f32 %f49, %f44, %f40, %f46; fma.rn.ftz.f32 %f48, %f44, %f37, %f45; $L__BB48_4: cvta.to.local.u64 %rd4, %rd3; mul.wide.u32 %rd5, %r2, 4; add.s64 %rd6, %rd4, %rd5; st.local.f32 [%rd6], %f48; st.local.f32 [%rd6+4], %f49; st.local.f32 [%rd6+8], %f50; bra.uni $L__BB48_6; $L__BB48_5: { // callseq 37, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2; .param .b64 param1; st.param.b64 [param1+0], %rd3; .param .b32 param2; st.param.b32 [param2+0], %r1; .param .b32 param3; st.param.b32 [param3+0], %r2; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 37 $L__BB48_6: ret; } .func _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<15>; .reg .b16 %rs<3>; .reg .f32 %f<51>; .reg .b32 %r<10>; .reg .b64 %rd<38>; ld.param.u64 %rd3, [_Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd4, [_Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r7, [_Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r8, [_Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd5, %rd4; cvta.to.local.u64 %rd6, %rd3; add.s64 %rd1, %rd6, 80; mul.wide.u32 %rd7, %r8, 4; add.s64 %rd2, %rd5, %rd7; mov.f32 %f50, 0f00000000; setp.gt.s32 %p1, %r7, 2; @%p1 bra $L__BB49_6; setp.eq.s32 %p5, %r7, 0; @%p5 bra $L__BB49_18; setp.eq.s32 %p6, %r7, 1; @%p6 bra $L__BB49_15; setp.eq.s32 %p7, %r7, 2; @%p7 bra $L__BB49_4; bra.uni $L__BB49_24; $L__BB49_4: ld.local.u32 %r4, [%rd1+-4]; setp.eq.s32 %p11, %r4, -1; @%p11 bra $L__BB49_24; ld.const.u64 %rd26, [kernel_params+2160]; cvta.to.global.u64 %rd27, %rd26; mul.wide.s32 %rd28, %r4, 208; add.s64 %rd29, %rd27, %rd28; ld.global.f32 %f50, [%rd29+120]; bra.uni $L__BB49_24; $L__BB49_6: setp.eq.s32 %p2, %r7, 3; @%p2 bra $L__BB49_13; setp.eq.s32 %p3, %r7, 4; @%p3 bra $L__BB49_23; setp.ne.s32 %p4, %r7, 5; @%p4 bra $L__BB49_24; ld.local.u32 %r1, [%rd1]; setp.eq.s32 %p8, %r1, -1; @%p8 bra $L__BB49_11; ld.const.u64 %rd8, [kernel_params+2368]; cvta.to.global.u64 %rd9, %rd8; mul.wide.s32 %rd10, %r1, 192; add.s64 %rd11, %rd9, %rd10; ld.global.f32 %f50, [%rd11+24]; bra.uni $L__BB49_24; $L__BB49_18: ld.local.u32 %r6, [%rd1+-4]; setp.eq.s32 %p13, %r6, -1; mov.f32 %f47, 0f00000000; mov.f32 %f48, %f47; mov.f32 %f49, %f47; @%p13 bra $L__BB49_22; ld.local.u8 %rs1, [%rd1+-24]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p14, %rs2, 0; @%p14 bra $L__BB49_21; bra.uni $L__BB49_20; $L__BB49_21: ld.const.u64 %rd34, [kernel_params+2160]; cvta.to.global.u64 %rd35, %rd34; mul.wide.s32 %rd36, %r6, 208; add.s64 %rd37, %rd35, %rd36; ld.global.v4.f32 {%f32, %f33, %f34, %f47}, [%rd37]; ld.global.v4.f32 {%f36, %f37, %f38, %f48}, [%rd37+16]; ld.global.v4.f32 {%f40, %f41, %f42, %f49}, [%rd37+32]; bra.uni $L__BB49_22; $L__BB49_15: ld.local.u32 %r5, [%rd1+-4]; setp.eq.s32 %p12, %r5, -1; mov.f32 %f44, 0f00000000; mov.f32 %f45, %f44; mov.f32 %f46, %f44; @%p12 bra $L__BB49_17; ld.const.u64 %rd30, [kernel_params+2160]; cvta.to.global.u64 %rd31, %rd30; mul.wide.s32 %rd32, %r5, 208; add.s64 %rd33, %rd31, %rd32; ld.global.f32 %f46, [%rd33+116]; ld.global.f32 %f45, [%rd33+112]; ld.global.f32 %f44, [%rd33+108]; $L__BB49_17: st.local.f32 [%rd2], %f44; st.local.f32 [%rd2+4], %f45; st.local.f32 [%rd2+8], %f46; bra.uni $L__BB49_25; $L__BB49_13: ld.local.u32 %r3, [%rd1+-4]; setp.eq.s32 %p10, %r3, -1; @%p10 bra $L__BB49_24; ld.const.u64 %rd22, [kernel_params+2160]; cvta.to.global.u64 %rd23, %rd22; mul.wide.s32 %rd24, %r3, 208; add.s64 %rd25, %rd23, %rd24; ld.global.f32 %f50, [%rd25+100]; bra.uni $L__BB49_24; $L__BB49_23: ld.const.u64 %rd16, [kernel_params+2408]; cvta.to.global.u64 %rd17, %rd16; ld.local.u32 %rd18, [%rd1+-32]; shl.b64 %rd19, %rd18, 5; and.b64 %rd20, %rd19, 134217696; add.s64 %rd21, %rd17, %rd20; ld.global.u32 %r9, [%rd21+20]; cvt.rn.f32.s32 %f50, %r9; bra.uni $L__BB49_24; $L__BB49_11: ld.local.u32 %r2, [%rd1+-4]; setp.eq.s32 %p9, %r2, -1; @%p9 bra $L__BB49_24; ld.const.u64 %rd12, [kernel_params+2160]; cvta.to.global.u64 %rd13, %rd12; mul.wide.s32 %rd14, %r2, 208; add.s64 %rd15, %rd13, %rd14; ld.global.f32 %f50, [%rd15+104]; $L__BB49_24: st.local.f32 [%rd2], %f50; $L__BB49_25: ret; $L__BB49_20: ld.local.f32 %f49, [%rd1+108]; ld.local.f32 %f48, [%rd1+92]; ld.local.f32 %f47, [%rd1+76]; $L__BB49_22: st.local.f32 [%rd2], %f47; st.local.f32 [%rd2+4], %f48; st.local.f32 [%rd2+8], %f49; bra.uni $L__BB49_25; } .func _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<20>; .reg .f32 %f<28>; .reg .b32 %r<31>; .reg .b64 %rd<104>; ld.param.u64 %rd19, [_Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd20, [_Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r9, [_Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r10, [_Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd21, %rd20; cvta.to.local.u64 %rd22, %rd19; add.s64 %rd1, %rd22, 76; mul.wide.u32 %rd23, %r10, 4; add.s64 %rd2, %rd21, %rd23; setp.gt.s32 %p1, %r9, 3; @%p1 bra $L__BB50_12; setp.gt.s32 %p7, %r9, 1; @%p7 bra $L__BB50_7; setp.eq.s32 %p10, %r9, 0; @%p10 bra $L__BB50_32; setp.eq.s32 %p11, %r9, 1; @%p11 bra $L__BB50_4; bra.uni $L__BB50_35; $L__BB50_4: ld.local.u32 %r7, [%rd1]; mov.u64 %rd102, 0; setp.eq.s32 %p18, %r7, -1; @%p18 bra $L__BB50_6; ld.const.u64 %rd79, [kernel_params+2160]; cvta.to.global.u64 %rd80, %rd79; mul.wide.s32 %rd81, %r7, 208; add.s64 %rd82, %rd80, %rd81; ld.global.s32 %rd102, [%rd82+124]; $L__BB50_6: ld.const.u64 %rd83, [kernel_params+2392]; cvta.to.global.u64 %rd84, %rd83; mul.lo.s64 %rd85, %rd102, 80; add.s64 %rd86, %rd84, %rd85; ld.global.u32 %r11, [%rd86]; add.s32 %r12, %r11, -559038716; xor.b32 %r13, %r12, -1874933675; add.s32 %r14, %r13, 163402623; xor.b32 %r15, %r14, -559038716; shf.l.wrap.b32 %r16, %r14, %r14, 25; sub.s32 %r17, %r15, %r16; xor.b32 %r18, %r17, -1874933675; shf.l.wrap.b32 %r19, %r17, %r17, 16; sub.s32 %r20, %r18, %r19; xor.b32 %r21, %r20, %r14; shf.l.wrap.b32 %r22, %r20, %r20, 4; sub.s32 %r23, %r21, %r22; xor.b32 %r24, %r23, %r17; shf.l.wrap.b32 %r25, %r23, %r23, 14; sub.s32 %r26, %r24, %r25; xor.b32 %r27, %r26, %r20; shf.l.wrap.b32 %r28, %r26, %r26, 24; sub.s32 %r29, %r27, %r28; cvt.rn.f32.u32 %f25, %r29; mul.ftz.f32 %f26, %f25, 0f2F800000; st.local.f32 [%rd2], %f26; bra.uni $L__BB50_35; $L__BB50_12: setp.gt.s32 %p2, %r9, 6; @%p2 bra $L__BB50_18; setp.eq.s32 %p5, %r9, 4; @%p5 bra $L__BB50_26; setp.eq.s32 %p6, %r9, 6; @%p6 bra $L__BB50_15; bra.uni $L__BB50_35; $L__BB50_15: ld.local.u32 %r3, [%rd1]; mov.u64 %rd98, 0; setp.eq.s32 %p14, %r3, -1; @%p14 bra $L__BB50_17; ld.const.u64 %rd43, [kernel_params+2160]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r3, 208; add.s64 %rd46, %rd44, %rd45; ld.global.s32 %rd98, [%rd46+124]; $L__BB50_17: ld.const.u64 %rd47, [kernel_params+2392]; cvta.to.global.u64 %rd48, %rd47; mul.lo.s64 %rd49, %rd98, 80; add.s64 %rd50, %rd48, %rd49; ld.global.f32 %f15, [%rd50+12]; st.local.f32 [%rd2], %f15; bra.uni $L__BB50_35; $L__BB50_7: setp.eq.s32 %p8, %r9, 2; @%p8 bra $L__BB50_29; setp.eq.s32 %p9, %r9, 3; @%p9 bra $L__BB50_9; bra.uni $L__BB50_35; $L__BB50_9: ld.local.u32 %r5, [%rd1]; mov.u64 %rd100, 0; setp.eq.s32 %p16, %r5, -1; @%p16 bra $L__BB50_11; ld.const.u64 %rd61, [kernel_params+2160]; cvta.to.global.u64 %rd62, %rd61; mul.wide.s32 %rd63, %r5, 208; add.s64 %rd64, %rd62, %rd63; ld.global.s32 %rd100, [%rd64+124]; $L__BB50_11: ld.const.u64 %rd65, [kernel_params+2392]; cvta.to.global.u64 %rd66, %rd65; mul.lo.s64 %rd67, %rd100, 80; add.s64 %rd68, %rd66, %rd67; ld.global.f32 %f23, [%rd68+8]; st.local.f32 [%rd2], %f23; bra.uni $L__BB50_35; $L__BB50_18: setp.eq.s32 %p3, %r9, 7; @%p3 bra $L__BB50_23; setp.ne.s32 %p4, %r9, 8; @%p4 bra $L__BB50_35; ld.local.u32 %r1, [%rd1]; mov.u64 %rd96, 0; setp.eq.s32 %p12, %r1, -1; @%p12 bra $L__BB50_22; ld.const.u64 %rd25, [kernel_params+2160]; cvta.to.global.u64 %rd26, %rd25; mul.wide.s32 %rd27, %r1, 208; add.s64 %rd28, %rd26, %rd27; ld.global.s32 %rd96, [%rd28+124]; $L__BB50_22: ld.const.u64 %rd29, [kernel_params+2392]; cvta.to.global.u64 %rd30, %rd29; mul.lo.s64 %rd31, %rd96, 80; add.s64 %rd32, %rd30, %rd31; ld.global.v4.f32 {%f1, %f2, %f3, %f4}, [%rd32+64]; st.local.f32 [%rd2], %f1; st.local.f32 [%rd2+4], %f2; st.local.f32 [%rd2+8], %f3; bra.uni $L__BB50_35; $L__BB50_32: ld.local.u32 %r8, [%rd1]; mov.u64 %rd103, 0; setp.eq.s32 %p19, %r8, -1; @%p19 bra $L__BB50_34; ld.const.u64 %rd88, [kernel_params+2160]; cvta.to.global.u64 %rd89, %rd88; mul.wide.s32 %rd90, %r8, 208; add.s64 %rd91, %rd89, %rd90; ld.global.s32 %rd103, [%rd91+124]; $L__BB50_34: ld.const.u64 %rd92, [kernel_params+2392]; cvta.to.global.u64 %rd93, %rd92; mul.lo.s64 %rd94, %rd103, 80; add.s64 %rd95, %rd93, %rd94; ld.global.u32 %r30, [%rd95]; cvt.rn.f32.u32 %f27, %r30; st.local.f32 [%rd2], %f27; bra.uni $L__BB50_35; $L__BB50_26: ld.local.u32 %r4, [%rd1]; mov.u64 %rd99, 0; setp.eq.s32 %p15, %r4, -1; @%p15 bra $L__BB50_28; ld.const.u64 %rd52, [kernel_params+2160]; cvta.to.global.u64 %rd53, %rd52; mul.wide.s32 %rd54, %r4, 208; add.s64 %rd55, %rd53, %rd54; ld.global.s32 %rd99, [%rd55+124]; $L__BB50_28: ld.const.u64 %rd56, [kernel_params+2392]; cvta.to.global.u64 %rd57, %rd56; mul.lo.s64 %rd58, %rd99, 80; add.s64 %rd59, %rd57, %rd58; ld.global.v4.f32 {%f16, %f17, %f18, %f19}, [%rd59+32]; st.local.f32 [%rd2], %f16; st.local.f32 [%rd2+4], %f17; st.local.f32 [%rd2+8], %f18; bra.uni $L__BB50_35; $L__BB50_29: ld.local.u32 %r6, [%rd1]; mov.u64 %rd101, 0; setp.eq.s32 %p17, %r6, -1; @%p17 bra $L__BB50_31; ld.const.u64 %rd70, [kernel_params+2160]; cvta.to.global.u64 %rd71, %rd70; mul.wide.s32 %rd72, %r6, 208; add.s64 %rd73, %rd71, %rd72; ld.global.s32 %rd101, [%rd73+124]; $L__BB50_31: ld.const.u64 %rd74, [kernel_params+2392]; cvta.to.global.u64 %rd75, %rd74; mul.lo.s64 %rd76, %rd101, 80; add.s64 %rd77, %rd75, %rd76; ld.global.f32 %f24, [%rd77+4]; st.local.f32 [%rd2], %f24; bra.uni $L__BB50_35; $L__BB50_23: ld.local.u32 %r2, [%rd1]; mov.u64 %rd97, 0; setp.eq.s32 %p13, %r2, -1; @%p13 bra $L__BB50_25; ld.const.u64 %rd34, [kernel_params+2160]; cvta.to.global.u64 %rd35, %rd34; mul.wide.s32 %rd36, %r2, 208; add.s64 %rd37, %rd35, %rd36; ld.global.s32 %rd97, [%rd37+124]; $L__BB50_25: ld.const.u64 %rd38, [kernel_params+2392]; cvta.to.global.u64 %rd39, %rd38; mul.lo.s64 %rd40, %rd97, 80; add.s64 %rd41, %rd39, %rd40; ld.global.v4.f32 {%f8, %f9, %f10, %f11}, [%rd41+48]; st.local.f32 [%rd2], %f8; st.local.f32 [%rd2+4], %f9; st.local.f32 [%rd2+8], %f10; $L__BB50_35: ret; } .func _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<17>; .reg .b16 %rs<9>; .reg .f32 %f<74>; .reg .b32 %r<47>; .reg .b64 %rd<66>; ld.param.u64 %rd19, [_Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd20, [_Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r18, [_Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r19, [_Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd21, %rd20; cvta.to.local.u64 %rd22, %rd19; add.s64 %rd1, %rd22, 64; mul.wide.u32 %rd23, %r19, 4; add.s64 %rd2, %rd21, %rd23; setp.eq.s32 %p1, %r18, 0; @%p1 bra $L__BB51_24; setp.eq.s32 %p2, %r18, 4; @%p2 bra $L__BB51_21; setp.ne.s32 %p3, %r18, 3; @%p3 bra $L__BB51_25; ld.local.u32 %r1, [%rd1]; and.b32 %r20, %r1, 6; setp.eq.s32 %p4, %r20, 0; mov.f32 %f70, 0f00000000; @%p4 bra $L__BB51_20; ld.const.u64 %rd24, [kernel_params+2264]; cvta.to.global.u64 %rd25, %rd24; ld.local.u32 %r21, [%rd1+-4]; mul.wide.s32 %rd26, %r21, 16; add.s64 %rd27, %rd25, %rd26; shr.s32 %r22, %r1, 7; ld.global.u32 %r23, [%rd27+4]; add.s32 %r2, %r22, %r23; add.s32 %r3, %r2, 1; and.b32 %r24, %r1, 64; setp.eq.s32 %p5, %r24, 0; @%p5 bra $L__BB51_18; bra.uni $L__BB51_5; $L__BB51_18: ld.const.u64 %rd57, [kernel_params+2272]; cvta.to.global.u64 %rd58, %rd57; mul.wide.s32 %rd59, %r2, 16; add.s64 %rd60, %rd58, %rd59; ld.global.f32 %f68, [%rd60+12]; ld.global.f32 %f69, [%rd60+28]; bra.uni $L__BB51_19; $L__BB51_24: ld.local.u32 %r40, [%rd1]; and.b32 %r41, %r40, 6; setp.ne.s32 %p16, %r41, 0; selp.f32 %f67, 0f3F800000, 0f00000000, %p16; st.local.f32 [%rd2], %f67; bra.uni $L__BB51_25; $L__BB51_21: ld.local.u8 %rs5, [%rd1]; and.b16 %rs6, %rs5, 6; setp.eq.s16 %p15, %rs6, 0; mov.f32 %f71, 0f00000000; mov.f32 %f72, %f71; mov.f32 %f73, %f71; @%p15 bra $L__BB51_23; ld.local.f32 %f35, [%rd1+-28]; neg.ftz.f32 %f36, %f35; ld.local.f32 %f37, [%rd1+-24]; neg.ftz.f32 %f38, %f37; ld.local.f32 %f39, [%rd1+-20]; neg.ftz.f32 %f40, %f39; ld.local.f32 %f41, [%rd1+52]; mul.ftz.f32 %f42, %f35, %f41; ld.local.f32 %f43, [%rd1+56]; mul.ftz.f32 %f44, %f37, %f43; neg.ftz.f32 %f45, %f44; sub.ftz.f32 %f46, %f45, %f42; ld.local.f32 %f47, [%rd1+60]; mul.ftz.f32 %f48, %f39, %f47; sub.ftz.f32 %f49, %f46, %f48; mul.ftz.f32 %f50, %f43, %f43; fma.rn.ftz.f32 %f51, %f41, %f41, %f50; fma.rn.ftz.f32 %f52, %f47, %f47, %f51; div.approx.ftz.f32 %f53, %f49, %f52; mul.ftz.f32 %f54, %f53, %f41; mul.ftz.f32 %f55, %f53, %f43; mul.ftz.f32 %f56, %f53, %f47; sub.ftz.f32 %f57, %f36, %f54; sub.ftz.f32 %f58, %f38, %f55; sub.ftz.f32 %f59, %f40, %f56; neg.ftz.f32 %f60, %f57; neg.ftz.f32 %f61, %f58; neg.ftz.f32 %f62, %f59; mul.ftz.f32 %f63, %f58, %f58; fma.rn.ftz.f32 %f64, %f57, %f57, %f63; fma.rn.ftz.f32 %f65, %f59, %f59, %f64; sqrt.approx.ftz.f32 %f66, %f65; div.approx.ftz.f32 %f73, %f60, %f66; div.approx.ftz.f32 %f72, %f61, %f66; div.approx.ftz.f32 %f71, %f62, %f66; $L__BB51_23: st.local.f32 [%rd2], %f73; st.local.f32 [%rd2+4], %f72; st.local.f32 [%rd2+8], %f71; bra.uni $L__BB51_25; $L__BB51_5: ld.const.u64 %rd28, [kernel_params+2160]; cvta.to.global.u64 %rd29, %rd28; ld.local.u32 %r25, [%rd1+12]; mul.wide.s32 %rd30, %r25, 208; add.s64 %rd31, %rd29, %rd30; ld.global.u32 %r4, [%rd31+148]; ld.global.u32 %r5, [%rd31+152]; shl.b32 %r26, %r5, 1; cvt.rn.f32.s32 %f19, %r26; ld.local.f32 %f20, [%rd1+20]; mul.ftz.f32 %f1, %f20, %f19; cvt.rzi.ftz.s32.f32 %r27, %f1; add.s32 %r28, %r26, -1; min.s32 %r6, %r27, %r28; ld.const.u64 %rd32, [kernel_params+2312]; cvta.to.global.u64 %rd3, %rd32; ld.global.u32 %r43, [%rd31+164]; mul.wide.u32 %rd33, %r43, 16; add.s64 %rd34, %rd3, %rd33; ld.global.u16 %rs7, [%rd34+12]; ld.global.u32 %r42, [%rd34+8]; ld.global.u64 %rd61, [%rd34]; setp.eq.s64 %p6, %rd61, 11; @%p6 bra $L__BB51_10; $L__BB51_6: setp.eq.s64 %p7, %rd61, 0; @%p7 bra $L__BB51_8; add.s32 %r43, %r43, 2; bra.uni $L__BB51_9; $L__BB51_8: setp.eq.s16 %p8, %rs7, 0; mov.u32 %r46, -1; mov.u32 %r43, %r42; @%p8 bra $L__BB51_11; $L__BB51_9: mul.wide.u32 %rd35, %r43, 16; add.s64 %rd36, %rd3, %rd35; ld.global.u16 %rs7, [%rd36+12]; ld.global.u32 %r42, [%rd36+8]; ld.global.u64 %rd61, [%rd36]; setp.ne.s64 %p9, %rd61, 11; @%p9 bra $L__BB51_6; $L__BB51_10: setp.eq.s16 %p10, %rs7, 0; selp.b32 %r46, -1, %r42, %p10; $L__BB51_11: setp.eq.s32 %p11, %r6, %r5; @%p11 bra $L__BB51_13; bra.uni $L__BB51_12; $L__BB51_13: ld.const.u64 %rd43, [kernel_params+2272]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r2, 16; add.s64 %rd46, %rd44, %rd45; add.s64 %rd62, %rd46, 12; add.s64 %rd63, %rd46, 28; bra.uni $L__BB51_14; $L__BB51_12: setp.gt.s32 %p12, %r6, %r5; selp.b32 %r30, -1, 0, %p12; add.s32 %r31, %r6, %r30; mad.lo.s32 %r32, %r31, %r4, %r46; ld.const.u64 %rd37, [kernel_params+2344]; cvta.to.global.u64 %rd38, %rd37; add.s32 %r33, %r32, %r2; mul.wide.s32 %rd39, %r33, 16; add.s64 %rd40, %rd38, %rd39; add.s64 %rd62, %rd40, 12; add.s32 %r34, %r32, %r3; mul.wide.s32 %rd41, %r34, 16; add.s64 %rd42, %rd38, %rd41; add.s64 %rd63, %rd42, 12; $L__BB51_14: ld.global.f32 %f2, [%rd63]; ld.global.f32 %f3, [%rd62]; add.s32 %r17, %r6, 1; setp.eq.s32 %p13, %r17, %r5; @%p13 bra $L__BB51_16; bra.uni $L__BB51_15; $L__BB51_16: ld.const.u64 %rd53, [kernel_params+2272]; cvta.to.global.u64 %rd54, %rd53; mul.wide.s32 %rd55, %r2, 16; add.s64 %rd56, %rd54, %rd55; add.s64 %rd64, %rd56, 12; add.s64 %rd65, %rd56, 28; bra.uni $L__BB51_17; $L__BB51_15: setp.ge.s32 %p14, %r6, %r5; selp.b32 %r35, -1, 0, %p14; add.s32 %r36, %r17, %r35; mad.lo.s32 %r37, %r36, %r4, %r46; ld.const.u64 %rd47, [kernel_params+2344]; cvta.to.global.u64 %rd48, %rd47; add.s32 %r38, %r37, %r2; mul.wide.s32 %rd49, %r38, 16; add.s64 %rd50, %rd48, %rd49; add.s64 %rd64, %rd50, 12; add.s32 %r39, %r37, %r3; mul.wide.s32 %rd51, %r39, 16; add.s64 %rd52, %rd48, %rd51; add.s64 %rd65, %rd52, 12; $L__BB51_17: cvt.rn.f32.s32 %f21, %r6; sub.ftz.f32 %f22, %f1, %f21; mov.f32 %f23, 0f3F800000; sub.ftz.f32 %f24, %f23, %f22; ld.global.f32 %f25, [%rd64]; mul.ftz.f32 %f26, %f22, %f25; fma.rn.ftz.f32 %f68, %f24, %f3, %f26; ld.global.f32 %f27, [%rd65]; mul.ftz.f32 %f28, %f22, %f27; fma.rn.ftz.f32 %f69, %f24, %f2, %f28; $L__BB51_19: ld.local.f32 %f29, [%rd1+4]; sub.ftz.f32 %f30, %f69, %f68; fma.rn.ftz.f32 %f31, %f30, %f29, %f68; add.ftz.f32 %f70, %f31, %f31; $L__BB51_20: st.local.f32 [%rd2], %f70; $L__BB51_25: ret; } .func _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b64 _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2, .param .b32 _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3 ) { .reg .pred %p<19>; .reg .b16 %rs<9>; .reg .f32 %f<207>; .reg .b32 %r<52>; .reg .b64 %rd<60>; ld.param.u64 %rd15, [_Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u64 %rd16, [_Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r18, [_Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; ld.param.u32 %r19, [_Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj_param_3]; cvta.to.local.u64 %rd17, %rd16; cvta.to.local.u64 %rd18, %rd15; add.s64 %rd1, %rd18, 64; mul.wide.u32 %rd19, %r19, 4; add.s64 %rd2, %rd17, %rd19; setp.eq.s32 %p1, %r18, 0; @%p1 bra $L__BB52_9; setp.ne.s32 %p2, %r18, 1; @%p2 bra $L__BB52_30; ld.local.u8 %rs5, [%rd1]; and.b16 %rs6, %rs5, 8; setp.eq.s16 %p3, %rs6, 0; mov.f32 %f197, 0f00000000; @%p3 bra $L__BB52_8; ld.const.u64 %rd20, [kernel_params+2296]; cvta.to.global.u64 %rd21, %rd20; ld.local.v2.u32 {%r20, %r21}, [%rd1+-8]; mul.wide.s32 %rd22, %r21, 16; add.s64 %rd23, %rd21, %rd22; ld.global.f32 %f197, [%rd23+12]; and.b32 %r23, %r20, 4; setp.ne.s32 %p4, %r23, 0; @%p4 bra $L__BB52_8; mul.ftz.f32 %f2, %f197, 0f3F13CD3A; and.b32 %r24, %r20, 2; setp.eq.s32 %p5, %r24, 0; @%p5 bra $L__BB52_6; bra.uni $L__BB52_5; $L__BB52_6: ld.const.u64 %rd24, [kernel_params+2160]; cvta.to.global.u64 %rd25, %rd24; ld.local.u32 %r25, [%rd1+12]; mul.wide.s32 %rd26, %r25, 208; add.s64 %rd27, %rd25, %rd26; ld.global.v4.f32 {%f71, %f72, %f73, %f74}, [%rd27]; ld.global.v4.f32 {%f78, %f79, %f80, %f81}, [%rd27+16]; ld.global.v4.f32 {%f85, %f86, %f87, %f88}, [%rd27+32]; mul.ftz.f32 %f92, %f2, %f72; fma.rn.ftz.f32 %f93, %f2, %f71, %f92; fma.rn.ftz.f32 %f194, %f2, %f73, %f93; mul.ftz.f32 %f94, %f2, %f79; fma.rn.ftz.f32 %f95, %f2, %f78, %f94; fma.rn.ftz.f32 %f195, %f2, %f80, %f95; mul.ftz.f32 %f96, %f2, %f86; fma.rn.ftz.f32 %f97, %f2, %f85, %f96; fma.rn.ftz.f32 %f196, %f2, %f87, %f97; bra.uni $L__BB52_7; $L__BB52_9: ld.local.u32 %r2, [%rd1]; and.b32 %r26, %r2, 8; setp.eq.s32 %p6, %r26, 0; mov.f32 %f203, 0f00000000; mov.f32 %f202, %f203; mov.f32 %f201, %f203; @%p6 bra $L__BB52_29; and.b32 %r27, %r2, 64; setp.eq.s32 %p7, %r27, 0; @%p7 bra $L__BB52_24; bra.uni $L__BB52_11; $L__BB52_24: ld.const.u64 %rd49, [kernel_params+2296]; cvta.to.global.u64 %rd50, %rd49; ld.local.u32 %r43, [%rd1+-4]; mul.wide.s32 %rd51, %r43, 16; add.s64 %rd52, %rd50, %rd51; ld.global.v4.f32 {%f203, %f202, %f201, %f127}, [%rd52]; bra.uni $L__BB52_25; $L__BB52_11: ld.local.s32 %rd4, [%rd1+-4]; ld.const.u64 %rd28, [kernel_params+2160]; cvta.to.global.u64 %rd29, %rd28; ld.local.u32 %r28, [%rd1+12]; mul.wide.s32 %rd30, %r28, 208; add.s64 %rd31, %rd29, %rd30; ld.global.u32 %r3, [%rd31+148]; ld.global.u32 %r4, [%rd31+152]; shl.b32 %r29, %r4, 1; cvt.rn.f32.s32 %f104, %r29; ld.local.f32 %f105, [%rd1+20]; mul.ftz.f32 %f14, %f105, %f104; cvt.rzi.ftz.s32.f32 %r30, %f14; add.s32 %r31, %r29, -1; min.s32 %r5, %r30, %r31; ld.const.u64 %rd32, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd32; ld.global.u32 %r48, [%rd31+164]; mul.wide.u32 %rd33, %r48, 16; add.s64 %rd34, %rd5, %rd33; ld.global.u16 %rs7, [%rd34+12]; ld.global.u32 %r47, [%rd34+8]; ld.global.u64 %rd57, [%rd34]; setp.eq.s64 %p8, %rd57, 11; @%p8 bra $L__BB52_16; $L__BB52_12: setp.eq.s64 %p9, %rd57, 0; @%p9 bra $L__BB52_14; add.s32 %r48, %r48, 2; bra.uni $L__BB52_15; $L__BB52_14: setp.eq.s16 %p10, %rs7, 0; mov.u32 %r51, -1; mov.u32 %r48, %r47; @%p10 bra $L__BB52_17; $L__BB52_15: mul.wide.u32 %rd35, %r48, 16; add.s64 %rd36, %rd5, %rd35; ld.global.u16 %rs7, [%rd36+12]; ld.global.u32 %r47, [%rd36+8]; ld.global.u64 %rd57, [%rd36]; setp.ne.s64 %p11, %rd57, 11; @%p11 bra $L__BB52_12; $L__BB52_16: setp.eq.s16 %p12, %rs7, 0; selp.b32 %r51, -1, %r47, %p12; $L__BB52_17: setp.eq.s32 %p13, %r5, %r4; @%p13 bra $L__BB52_19; bra.uni $L__BB52_18; $L__BB52_19: ld.const.u64 %rd40, [kernel_params+2296]; cvta.to.global.u64 %rd41, %rd40; shl.b64 %rd42, %rd4, 4; add.s64 %rd58, %rd41, %rd42; bra.uni $L__BB52_20; $L__BB52_5: ld.local.v4.f32 {%f44, %f45, %f46, %f47}, [%rd1+80]; mul.ftz.f32 %f51, %f2, %f45; fma.rn.ftz.f32 %f52, %f2, %f44, %f51; fma.rn.ftz.f32 %f194, %f2, %f46, %f52; ld.local.v4.f32 {%f53, %f54, %f55, %f56}, [%rd1+96]; mul.ftz.f32 %f60, %f2, %f54; fma.rn.ftz.f32 %f61, %f2, %f53, %f60; fma.rn.ftz.f32 %f195, %f2, %f55, %f61; ld.local.v4.f32 {%f62, %f63, %f64, %f65}, [%rd1+112]; mul.ftz.f32 %f69, %f2, %f63; fma.rn.ftz.f32 %f70, %f2, %f62, %f69; fma.rn.ftz.f32 %f196, %f2, %f64, %f70; $L__BB52_7: mul.ftz.f32 %f98, %f195, %f195; fma.rn.ftz.f32 %f99, %f194, %f194, %f98; fma.rn.ftz.f32 %f100, %f196, %f196, %f99; sqrt.approx.ftz.f32 %f197, %f100; $L__BB52_8: st.local.f32 [%rd2], %f197; bra.uni $L__BB52_30; $L__BB52_18: cvt.u32.u64 %r33, %rd4; setp.gt.s32 %p14, %r5, %r4; selp.b32 %r34, -1, 0, %p14; add.s32 %r35, %r5, %r34; ld.const.u64 %rd37, [kernel_params+2344]; cvta.to.global.u64 %rd38, %rd37; mad.lo.s32 %r36, %r35, %r3, %r33; add.s32 %r37, %r36, %r51; mul.wide.s32 %rd39, %r37, 16; add.s64 %rd58, %rd38, %rd39; $L__BB52_20: ld.global.v4.f32 {%f106, %f107, %f108, %f109}, [%rd58]; add.s32 %r16, %r5, 1; setp.eq.s32 %p15, %r16, %r4; @%p15 bra $L__BB52_22; bra.uni $L__BB52_21; $L__BB52_22: ld.const.u64 %rd46, [kernel_params+2296]; cvta.to.global.u64 %rd47, %rd46; shl.b64 %rd48, %rd4, 4; add.s64 %rd59, %rd47, %rd48; bra.uni $L__BB52_23; $L__BB52_21: cvt.u32.u64 %r38, %rd4; setp.ge.s32 %p16, %r5, %r4; selp.b32 %r39, -1, 0, %p16; add.s32 %r40, %r16, %r39; ld.const.u64 %rd43, [kernel_params+2344]; cvta.to.global.u64 %rd44, %rd43; mad.lo.s32 %r41, %r40, %r3, %r38; add.s32 %r42, %r41, %r51; mul.wide.s32 %rd45, %r42, 16; add.s64 %rd59, %rd44, %rd45; $L__BB52_23: cvt.rn.f32.s32 %f110, %r5; sub.ftz.f32 %f111, %f14, %f110; ld.global.v4.f32 {%f112, %f113, %f114, %f115}, [%rd59]; mov.f32 %f119, 0f3F800000; sub.ftz.f32 %f120, %f119, %f111; mul.ftz.f32 %f121, %f111, %f112; mul.ftz.f32 %f122, %f111, %f113; mul.ftz.f32 %f123, %f111, %f114; fma.rn.ftz.f32 %f203, %f120, %f106, %f121; fma.rn.ftz.f32 %f202, %f120, %f107, %f122; fma.rn.ftz.f32 %f201, %f120, %f108, %f123; $L__BB52_25: ld.local.u32 %r17, [%rd1+-8]; and.b32 %r44, %r17, 4; setp.ne.s32 %p17, %r44, 0; @%p17 bra $L__BB52_29; and.b32 %r45, %r17, 2; setp.eq.s32 %p18, %r45, 0; @%p18 bra $L__BB52_28; bra.uni $L__BB52_27; $L__BB52_28: ld.const.u64 %rd53, [kernel_params+2160]; cvta.to.global.u64 %rd54, %rd53; ld.local.u32 %r46, [%rd1+12]; mul.wide.s32 %rd55, %r46, 208; add.s64 %rd56, %rd54, %rd55; ld.global.v4.f32 {%f161, %f162, %f163, %f164}, [%rd56]; ld.global.v4.f32 {%f169, %f170, %f171, %f172}, [%rd56+16]; ld.global.v4.f32 {%f177, %f178, %f179, %f180}, [%rd56+32]; mul.ftz.f32 %f185, %f202, %f162; fma.rn.ftz.f32 %f186, %f203, %f161, %f185; fma.rn.ftz.f32 %f187, %f201, %f163, %f186; add.ftz.f32 %f31, %f164, %f187; mul.ftz.f32 %f188, %f202, %f170; fma.rn.ftz.f32 %f189, %f203, %f169, %f188; fma.rn.ftz.f32 %f190, %f201, %f171, %f189; add.ftz.f32 %f32, %f172, %f190; mul.ftz.f32 %f191, %f202, %f178; fma.rn.ftz.f32 %f192, %f203, %f177, %f191; fma.rn.ftz.f32 %f193, %f201, %f179, %f192; add.ftz.f32 %f201, %f180, %f193; mov.f32 %f202, %f32; mov.f32 %f203, %f31; bra.uni $L__BB52_29; $L__BB52_27: ld.local.v4.f32 {%f128, %f129, %f130, %f131}, [%rd1+80]; mul.ftz.f32 %f136, %f202, %f129; fma.rn.ftz.f32 %f137, %f203, %f128, %f136; fma.rn.ftz.f32 %f138, %f201, %f130, %f137; add.ftz.f32 %f28, %f131, %f138; ld.local.v4.f32 {%f139, %f140, %f141, %f142}, [%rd1+96]; mul.ftz.f32 %f147, %f202, %f140; fma.rn.ftz.f32 %f148, %f203, %f139, %f147; fma.rn.ftz.f32 %f149, %f201, %f141, %f148; add.ftz.f32 %f29, %f142, %f149; ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd1+112]; mul.ftz.f32 %f158, %f202, %f151; fma.rn.ftz.f32 %f159, %f203, %f150, %f158; fma.rn.ftz.f32 %f160, %f201, %f152, %f159; add.ftz.f32 %f201, %f153, %f160; mov.f32 %f202, %f29; mov.f32 %f203, %f28; $L__BB52_29: st.local.f32 [%rd2], %f203; st.local.f32 [%rd2+4], %f202; st.local.f32 [%rd2+8], %f201; $L__BB52_30: ret; } .func _Z21svm_node_tex_gradientP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_tex_gradientP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_tex_gradientP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<20>; .reg .f32 %f<66>; .reg .b32 %r<16>; .reg .b64 %rd<11>; ld.param.u64 %rd3, [_Z21svm_node_tex_gradientP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r6, [_Z21svm_node_tex_gradientP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd3; and.b32 %r1, %r6, 255; bfe.u32 %r2, %r6, 16, 8; shr.u32 %r3, %r6, 24; shr.u32 %r7, %r6, 6; cvt.u64.u32 %rd4, %r7; and.b64 %rd5, %rd4, 1020; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f65, [%rd6]; ld.local.f32 %f2, [%rd6+4]; add.s64 %rd2, %rd6, 8; setp.eq.s32 %p1, %r1, 0; @%p1 bra $L__BB53_17; setp.eq.s32 %p2, %r1, 1; @%p2 bra $L__BB53_16; bra.uni $L__BB53_2; $L__BB53_16: mov.f32 %f62, 0f00000000; max.ftz.f32 %f63, %f65, %f62; mul.ftz.f32 %f65, %f63, %f63; bra.uni $L__BB53_17; $L__BB53_2: setp.eq.s32 %p3, %r1, 2; @%p3 bra $L__BB53_15; bra.uni $L__BB53_3; $L__BB53_15: mov.f32 %f55, 0f00000000; max.ftz.f32 %f56, %f65, %f55; mov.f32 %f57, 0f3F800000; min.ftz.f32 %f58, %f56, %f57; mul.ftz.f32 %f59, %f58, %f58; mul.ftz.f32 %f60, %f59, 0fC0000000; mul.ftz.f32 %f61, %f58, %f60; fma.rn.ftz.f32 %f65, %f59, 0f40400000, %f61; bra.uni $L__BB53_17; $L__BB53_3: setp.eq.s32 %p4, %r1, 3; @%p4 bra $L__BB53_14; bra.uni $L__BB53_4; $L__BB53_14: add.ftz.f32 %f54, %f65, %f2; mul.ftz.f32 %f65, %f54, 0f3F000000; bra.uni $L__BB53_17; $L__BB53_4: setp.eq.s32 %p5, %r1, 4; @%p5 bra $L__BB53_8; bra.uni $L__BB53_5; $L__BB53_8: abs.ftz.f32 %f6, %f65; setp.eq.ftz.f32 %p8, %f6, 0f00000000; abs.ftz.f32 %f7, %f2; setp.eq.ftz.f32 %p9, %f7, 0f00000000; and.pred %p10, %p8, %p9; mov.b32 %r4, %f65; mov.b32 %r8, %f2; and.b32 %r5, %r8, -2147483648; @%p10 bra $L__BB53_12; bra.uni $L__BB53_9; $L__BB53_12: shr.s32 %r13, %r4, 31; and.b32 %r14, %r13, 1078530011; or.b32 %r15, %r14, %r5; mov.b32 %f64, %r15; bra.uni $L__BB53_13; $L__BB53_5: ld.local.f32 %f18, [%rd2]; mul.ftz.f32 %f19, %f2, %f2; fma.rn.ftz.f32 %f20, %f65, %f65, %f19; fma.rn.ftz.f32 %f21, %f18, %f18, %f20; sqrt.approx.ftz.f32 %f22, %f21; mov.f32 %f23, 0f3F7FFFEF; sub.ftz.f32 %f24, %f23, %f22; mov.f32 %f25, 0f00000000; max.ftz.f32 %f3, %f24, %f25; setp.eq.s32 %p6, %r1, 5; @%p6 bra $L__BB53_7; bra.uni $L__BB53_6; $L__BB53_7: mul.ftz.f32 %f65, %f3, %f3; bra.uni $L__BB53_17; $L__BB53_9: setp.eq.ftz.f32 %p11, %f6, 0f7F800000; setp.eq.ftz.f32 %p12, %f7, 0f7F800000; and.pred %p13, %p11, %p12; @%p13 bra $L__BB53_11; bra.uni $L__BB53_10; $L__BB53_11: setp.lt.s32 %p17, %r4, 0; selp.b32 %r11, 1075235812, 1061752795, %p17; or.b32 %r12, %r11, %r5; mov.b32 %f64, %r12; bra.uni $L__BB53_13; $L__BB53_6: setp.eq.s32 %p7, %r1, 6; selp.f32 %f65, %f3, 0f00000000, %p7; bra.uni $L__BB53_17; $L__BB53_10: setp.lt.s32 %p14, %r4, 0; min.ftz.f32 %f26, %f7, %f6; max.ftz.f32 %f27, %f7, %f6; div.full.ftz.f32 %f28, %f26, %f27; mul.rn.ftz.f32 %f29, %f28, %f28; mov.f32 %f30, 0fC0B59883; mov.f32 %f31, 0fBF52C7EA; fma.rn.ftz.f32 %f32, %f29, %f31, %f30; mov.f32 %f33, 0fC0D21907; fma.rn.ftz.f32 %f34, %f32, %f29, %f33; mul.ftz.f32 %f35, %f29, %f34; mul.ftz.f32 %f36, %f28, %f35; add.ftz.f32 %f37, %f29, 0f41355DC0; mov.f32 %f38, 0f41E6BD60; fma.rn.ftz.f32 %f39, %f37, %f29, %f38; mov.f32 %f40, 0f419D92C8; fma.rn.ftz.f32 %f41, %f39, %f29, %f40; rcp.approx.ftz.f32 %f42, %f41; fma.rn.ftz.f32 %f43, %f36, %f42, %f28; mov.f32 %f44, 0f3FC90FDB; sub.ftz.f32 %f45, %f44, %f43; setp.gt.ftz.f32 %p15, %f7, %f6; selp.f32 %f46, %f45, %f43, %p15; mov.f32 %f47, 0f40490FDB; sub.ftz.f32 %f48, %f47, %f46; selp.f32 %f49, %f48, %f46, %p14; mov.b32 %r9, %f49; or.b32 %r10, %r5, %r9; mov.b32 %f50, %r10; add.ftz.f32 %f51, %f6, %f7; setp.le.ftz.f32 %p16, %f51, 0f7F800000; selp.f32 %f64, %f50, %f51, %p16; $L__BB53_13: mov.f32 %f52, 0f40C90FDB; div.approx.ftz.f32 %f53, %f64, %f52; add.ftz.f32 %f65, %f53, 0f3F000000; $L__BB53_17: cvt.ftz.sat.f32.f32 %f17, %f65; setp.eq.s32 %p18, %r2, 255; @%p18 bra $L__BB53_19; mul.wide.u32 %rd7, %r2, 4; add.s64 %rd8, %rd1, %rd7; st.local.f32 [%rd8], %f17; $L__BB53_19: setp.eq.s32 %p19, %r3, 255; @%p19 bra $L__BB53_21; mul.wide.u32 %rd9, %r3, 4; add.s64 %rd10, %rd1, %rd9; st.local.f32 [%rd10], %f17; st.local.f32 [%rd10+4], %f17; st.local.f32 [%rd10+8], %f17; $L__BB53_21: ret; } .func _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<27>; .reg .f32 %f<121>; .reg .b32 %r<40>; .reg .b64 %rd<21>; ld.param.u64 %rd1, [_Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r16, [_Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r17, [_Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd2, %rd1; bfe.u32 %r1, %r17, 16, 8; shr.u32 %r18, %r17, 6; cvt.u64.u32 %rd3, %r18; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd2, %rd4; ld.local.f32 %f1, [%rd5]; shl.b32 %r19, %r17, 2; cvt.u64.u32 %rd6, %r19; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd2, %rd7; shl.b32 %r20, %r16, 2; cvt.u64.u32 %rd9, %r20; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd2, %rd10; ld.local.f32 %f2, [%rd11]; shr.u32 %r21, %r16, 6; cvt.u64.u32 %rd12, %r21; and.b64 %rd13, %rd12, 1020; add.s64 %rd14, %rd2, %rd13; ld.local.f32 %f3, [%rd14]; shr.u32 %r22, %r16, 14; cvt.u64.u32 %rd15, %r22; and.b64 %rd16, %rd15, 1020; add.s64 %rd17, %rd2, %rd16; ld.local.f32 %f4, [%rd17]; ld.local.f32 %f5, [%rd8+8]; ld.local.f32 %f6, [%rd8+4]; max.ftz.f32 %f66, %f6, %f5; ld.local.f32 %f7, [%rd8]; max.ftz.f32 %f8, %f7, %f66; min.ftz.f32 %f67, %f6, %f5; min.ftz.f32 %f68, %f7, %f67; sub.ftz.f32 %f9, %f8, %f68; setp.eq.ftz.f32 %p1, %f8, 0f00000000; mov.f32 %f116, 0f00000000; mov.f32 %f114, %f116; @%p1 bra $L__BB54_2; div.approx.ftz.f32 %f114, %f9, %f8; $L__BB54_2: setp.eq.ftz.f32 %p2, %f114, 0f00000000; @%p2 bra $L__BB54_9; sub.ftz.f32 %f70, %f8, %f7; div.approx.ftz.f32 %f12, %f70, %f9; sub.ftz.f32 %f71, %f8, %f6; div.approx.ftz.f32 %f13, %f71, %f9; sub.ftz.f32 %f72, %f8, %f5; div.approx.ftz.f32 %f14, %f72, %f9; setp.eq.ftz.f32 %p3, %f7, %f8; @%p3 bra $L__BB54_7; bra.uni $L__BB54_4; $L__BB54_7: sub.ftz.f32 %f115, %f14, %f13; bra.uni $L__BB54_8; $L__BB54_4: setp.eq.ftz.f32 %p4, %f6, %f8; @%p4 bra $L__BB54_6; bra.uni $L__BB54_5; $L__BB54_6: add.ftz.f32 %f74, %f12, 0f40000000; sub.ftz.f32 %f115, %f74, %f14; bra.uni $L__BB54_8; $L__BB54_5: add.ftz.f32 %f73, %f13, 0f40800000; sub.ftz.f32 %f115, %f73, %f12; $L__BB54_8: mov.f32 %f75, 0f40C00000; div.approx.ftz.f32 %f76, %f115, %f75; setp.lt.ftz.f32 %p5, %f76, 0f00000000; add.ftz.f32 %f77, %f76, 0f3F800000; selp.f32 %f116, %f77, %f76, %p5; $L__BB54_9: add.ftz.f32 %f78, %f2, %f116; add.ftz.f32 %f21, %f78, 0f3F000000; abs.ftz.f32 %f22, %f21; mov.f32 %f79, 0f3F800000; abs.ftz.f32 %f23, %f79; mul.ftz.f32 %f24, %f23, 0f4B000000; setp.lt.ftz.f32 %p6, %f22, %f23; mov.b32 %r39, %f22; @%p6 bra $L__BB54_21; setp.le.ftz.f32 %p7, %f22, %f24; setp.leu.ftz.f32 %p8, %f23, 0f00000000; or.pred %p9, %p8, %p7; setp.gt.u32 %p10, %r39, 2139095039; or.pred %p11, %p10, %p9; @%p11 bra $L__BB54_15; bra.uni $L__BB54_11; $L__BB54_15: div.approx.ftz.f32 %f84, %f22, %f23; cvt.rzi.ftz.f32.f32 %f117, %f84; neg.ftz.f32 %f26, %f23; fma.rn.f32 %f27, %f26, %f117, %f22; mov.b32 %r31, %f27; mov.b32 %r32, %f23; setp.lt.u32 %p14, %r31, %r32; @%p14 bra $L__BB54_20; setp.gt.u32 %p15, %r31, -2147483648; @%p15 bra $L__BB54_19; bra.uni $L__BB54_17; $L__BB54_19: add.ftz.f32 %f90, %f117, 0fBF800000; add.ftz.f32 %f91, %f90, 0fBF800000; setp.lt.ftz.f32 %p18, %f27, %f26; selp.f32 %f117, %f91, %f90, %p18; bra.uni $L__BB54_20; $L__BB54_11: shr.u32 %r23, %r39, 23; mov.b32 %r3, %f24; shr.u32 %r24, %r3, 23; and.b32 %r25, %r39, 8388607; or.b32 %r38, %r25, 8388608; add.s32 %r26, %r23, 23; sub.s32 %r36, %r26, %r24; setp.eq.s32 %p12, %r36, 0; @%p12 bra $L__BB54_14; and.b32 %r27, %r3, 8388607; or.b32 %r6, %r27, 8388608; $L__BB54_13: min.s32 %r28, %r36, 8; shl.b32 %r29, %r38, %r28; rem.u32 %r38, %r29, %r6; sub.s32 %r36, %r36, %r28; setp.ne.s32 %p13, %r36, 0; @%p13 bra $L__BB54_13; $L__BB54_14: and.b32 %r30, %r3, -8388608; mov.b32 %f80, %r30; cvt.rn.f32.u32 %f81, %r38; mul.ftz.f32 %f82, %f81, 0f28800000; mul.ftz.f32 %f83, %f82, %f80; mov.b32 %r39, %f83; bra.uni $L__BB54_21; $L__BB54_17: add.ftz.f32 %f117, %f117, 0f3F800000; add.ftz.f32 %f85, %f23, %f23; setp.ltu.ftz.f32 %p16, %f27, %f85; @%p16 bra $L__BB54_20; add.ftz.f32 %f86, %f117, 0f3F800000; mov.f32 %f87, 0fC0400000; fma.rn.ftz.f32 %f88, %f87, %f23, %f27; setp.ge.ftz.f32 %p17, %f88, 0f00000000; add.ftz.f32 %f89, %f86, 0f3F800000; selp.f32 %f117, %f89, %f86, %p17; $L__BB54_20: fma.rn.ftz.f32 %f92, %f26, %f117, %f22; mov.b32 %r39, %f92; $L__BB54_21: mov.b32 %r34, %f21; and.b32 %r35, %r34, -2147483648; or.b32 %r15, %r39, %r35; mul.ftz.f32 %f93, %f3, %f114; cvt.ftz.sat.f32.f32 %f32, %f93; mul.ftz.f32 %f33, %f4, %f8; setp.neu.ftz.f32 %p19, %f32, 0f00000000; mov.f32 %f118, %f33; mov.f32 %f119, %f33; mov.f32 %f120, %f33; @%p19 bra $L__BB54_22; bra.uni $L__BB54_29; $L__BB54_22: mov.b32 %f94, %r15; setp.eq.ftz.f32 %p20, %f94, 0f3F800000; mul.ftz.f32 %f96, %f94, 0f40C00000; selp.f32 %f97, 0f00000000, %f96, %p20; cvt.rmi.ftz.f32.f32 %f37, %f97; sub.ftz.f32 %f98, %f97, %f37; sub.ftz.f32 %f99, %f79, %f32; mul.ftz.f32 %f38, %f33, %f99; mul.ftz.f32 %f100, %f32, %f98; sub.ftz.f32 %f101, %f79, %f100; mul.ftz.f32 %f39, %f33, %f101; sub.ftz.f32 %f102, %f79, %f98; mul.ftz.f32 %f103, %f32, %f102; sub.ftz.f32 %f104, %f79, %f103; mul.ftz.f32 %f40, %f33, %f104; setp.eq.ftz.f32 %p21, %f37, 0f00000000; @%p21 bra $L__BB54_28; bra.uni $L__BB54_23; $L__BB54_28: mov.f32 %f118, %f33; mov.f32 %f119, %f40; mov.f32 %f120, %f38; bra.uni $L__BB54_29; $L__BB54_23: setp.eq.ftz.f32 %p22, %f37, 0f3F800000; mov.f32 %f118, %f39; mov.f32 %f119, %f33; mov.f32 %f120, %f38; @%p22 bra $L__BB54_29; setp.eq.ftz.f32 %p23, %f37, 0f40000000; mov.f32 %f118, %f38; mov.f32 %f119, %f33; mov.f32 %f120, %f40; @%p23 bra $L__BB54_29; setp.eq.ftz.f32 %p24, %f37, 0f40400000; mov.f32 %f118, %f38; mov.f32 %f119, %f39; mov.f32 %f120, %f33; @%p24 bra $L__BB54_29; setp.eq.ftz.f32 %p25, %f37, 0f40800000; mov.f32 %f118, %f33; mov.f32 %f119, %f38; mov.f32 %f120, %f39; @%p25 bra $L__BB54_27; bra.uni $L__BB54_29; $L__BB54_27: mov.f32 %f118, %f40; mov.f32 %f119, %f38; mov.f32 %f120, %f33; $L__BB54_29: mul.ftz.f32 %f105, %f1, %f118; sub.ftz.f32 %f107, %f79, %f1; fma.rn.ftz.f32 %f62, %f107, %f7, %f105; mul.ftz.f32 %f108, %f1, %f119; fma.rn.ftz.f32 %f63, %f107, %f6, %f108; mul.ftz.f32 %f109, %f1, %f120; fma.rn.ftz.f32 %f64, %f107, %f5, %f109; setp.eq.s32 %p26, %r1, 255; @%p26 bra $L__BB54_31; mov.f32 %f110, 0f00000000; max.ftz.f32 %f111, %f64, %f110; mul.wide.u32 %rd19, %r1, 4; add.s64 %rd20, %rd2, %rd19; max.ftz.f32 %f112, %f62, %f110; st.local.f32 [%rd20], %f112; max.ftz.f32 %f113, %f63, %f110; st.local.f32 [%rd20+4], %f113; st.local.f32 [%rd20+8], %f111; $L__BB54_31: ret; } .func _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<22>; .reg .f32 %f<177>; .reg .b32 %r<63>; .reg .b64 %rd<46>; ld.param.u64 %rd5, [_Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r19, [_Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; ld.param.u32 %r18, [_Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r17, [_Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd5; and.b32 %r1, %r17, 255; bfe.u32 %r2, %r17, 16, 8; shr.u32 %r20, %r17, 6; cvt.u64.u32 %rd6, %r20; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f1, [%rd8]; ld.local.f32 %f2, [%rd8+4]; ld.local.f32 %f3, [%rd8+8]; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB55_2; bra.uni $L__BB55_1; $L__BB55_2: mov.b32 %f174, %r19; bra.uni $L__BB55_3; $L__BB55_1: mul.wide.u32 %rd9, %r1, 4; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f174, [%rd10]; $L__BB55_3: mul.ftz.f32 %f20, %f2, %f2; fma.rn.ftz.f32 %f21, %f1, %f1, %f20; fma.rn.ftz.f32 %f22, %f3, %f3, %f21; sqrt.approx.ftz.f32 %f23, %f22; div.approx.ftz.f32 %f24, %f1, %f23; div.approx.ftz.f32 %f25, %f2, %f23; div.approx.ftz.f32 %f26, %f3, %f23; neg.ftz.f32 %f27, %f26; mov.f32 %f28, 0fBF800000; max.ftz.f32 %f29, %f27, %f28; mov.f32 %f30, 0f3F800000; min.ftz.f32 %f31, %f29, %f30; abs.ftz.f32 %f32, %f31; sub.ftz.f32 %f33, %f30, %f32; mul.ftz.f32 %f34, %f33, 0f3F000000; sqrt.approx.ftz.f32 %f35, %f34; setp.gt.ftz.f32 %p2, %f32, 0f3F11EB85; selp.f32 %f36, %f35, %f32, %p2; mul.ftz.f32 %f37, %f36, %f36; mov.f32 %f38, 0f3C94D2E9; mov.f32 %f39, 0f3D53F941; fma.rn.ftz.f32 %f40, %f39, %f37, %f38; mov.f32 %f41, 0f3D3F841F; fma.rn.ftz.f32 %f42, %f40, %f37, %f41; mov.f32 %f43, 0f3D994929; fma.rn.ftz.f32 %f44, %f42, %f37, %f43; mov.f32 %f45, 0f3E2AAB94; fma.rn.ftz.f32 %f46, %f44, %f37, %f45; mul.ftz.f32 %f47, %f37, %f46; fma.rn.ftz.f32 %f48, %f47, %f36, %f36; add.ftz.f32 %f49, %f48, %f48; mov.f32 %f50, 0f3FC90FDB; sub.ftz.f32 %f51, %f50, %f48; selp.f32 %f52, %f49, %f51, %p2; setp.lt.ftz.f32 %p3, %f31, 0f00000000; mov.f32 %f53, 0f40490FDB; sub.ftz.f32 %f54, %f53, %f52; selp.f32 %f7, %f54, %f52, %p3; abs.ftz.f32 %f8, %f25; abs.ftz.f32 %f9, %f24; setp.eq.ftz.f32 %p4, %f8, 0f00000000; setp.eq.ftz.f32 %p5, %f9, 0f00000000; and.pred %p6, %p4, %p5; mov.b32 %r3, %f25; mov.b32 %r21, %f24; and.b32 %r4, %r21, -2147483648; @%p6 bra $L__BB55_7; bra.uni $L__BB55_4; $L__BB55_7: shr.s32 %r26, %r3, 31; and.b32 %r27, %r26, 1078530011; or.b32 %r28, %r27, %r4; mov.b32 %f175, %r28; bra.uni $L__BB55_8; $L__BB55_4: setp.eq.ftz.f32 %p7, %f8, 0f7F800000; setp.eq.ftz.f32 %p8, %f9, 0f7F800000; and.pred %p9, %p7, %p8; @%p9 bra $L__BB55_6; bra.uni $L__BB55_5; $L__BB55_6: setp.lt.s32 %p13, %r3, 0; selp.b32 %r24, 1075235812, 1061752795, %p13; or.b32 %r25, %r24, %r4; mov.b32 %f175, %r25; bra.uni $L__BB55_8; $L__BB55_5: setp.lt.s32 %p10, %r3, 0; min.ftz.f32 %f55, %f9, %f8; max.ftz.f32 %f56, %f9, %f8; div.full.ftz.f32 %f57, %f55, %f56; mul.rn.ftz.f32 %f58, %f57, %f57; mov.f32 %f59, 0fC0B59883; mov.f32 %f60, 0fBF52C7EA; fma.rn.ftz.f32 %f61, %f58, %f60, %f59; mov.f32 %f62, 0fC0D21907; fma.rn.ftz.f32 %f63, %f61, %f58, %f62; mul.ftz.f32 %f64, %f58, %f63; mul.ftz.f32 %f65, %f57, %f64; add.ftz.f32 %f66, %f58, 0f41355DC0; mov.f32 %f67, 0f41E6BD60; fma.rn.ftz.f32 %f68, %f66, %f58, %f67; mov.f32 %f69, 0f419D92C8; fma.rn.ftz.f32 %f70, %f68, %f58, %f69; rcp.approx.ftz.f32 %f71, %f70; fma.rn.ftz.f32 %f72, %f65, %f71, %f57; sub.ftz.f32 %f74, %f50, %f72; setp.gt.ftz.f32 %p11, %f9, %f8; selp.f32 %f75, %f74, %f72, %p11; sub.ftz.f32 %f77, %f53, %f75; selp.f32 %f78, %f77, %f75, %p10; mov.b32 %r22, %f78; or.b32 %r23, %r4, %r22; mov.b32 %f79, %r23; add.ftz.f32 %f80, %f8, %f9; setp.le.ftz.f32 %p12, %f80, 0f7F800000; selp.f32 %f175, %f79, %f80, %p12; $L__BB55_8: add.ftz.f32 %f14, %f175, 0f40490FDB; ld.const.u64 %rd11, [kernel_params+2440]; cvta.to.global.u64 %rd2, %rd11; mul.wide.s32 %rd12, %r18, 4; add.s64 %rd13, %rd2, %rd12; ld.global.u32 %r5, [%rd13]; setp.eq.s32 %p14, %r5, -1; mov.f32 %f176, 0f42C80000; @%p14 bra $L__BB55_15; mul.wide.s32 %rd14, %r5, 4; add.s64 %rd15, %rd2, %rd14; ld.global.u32 %r7, [%rd15]; add.s32 %r29, %r7, %r5; add.s32 %r8, %r29, 2; ld.global.u32 %r9, [%rd15+4]; add.s32 %r10, %r9, -1; add.s32 %r30, %r10, %r8; mul.wide.s32 %rd16, %r30, 4; add.s64 %rd17, %rd2, %rd16; ld.global.f32 %f83, [%rd17]; setp.le.ftz.f32 %p15, %f83, %f7; mov.f32 %f176, 0f00000000; @%p15 bra $L__BB55_15; mov.u32 %r61, 0; $L__BB55_11: mov.u32 %r12, %r61; add.s32 %r61, %r12, 1; add.s32 %r32, %r12, %r5; mul.wide.s32 %rd18, %r32, 4; add.s64 %rd3, %rd2, %rd18; ld.global.f32 %f15, [%rd3+12]; setp.lt.ftz.f32 %p16, %f15, %f14; @%p16 bra $L__BB55_11; mov.u32 %r62, 0; $L__BB55_13: mov.u32 %r14, %r62; add.s32 %r62, %r14, 1; add.s32 %r34, %r14, %r29; mul.wide.s32 %rd19, %r34, 4; add.s64 %rd4, %rd2, %rd19; ld.global.f32 %f16, [%rd4+12]; setp.lt.ftz.f32 %p17, %f16, %f7; @%p17 bra $L__BB55_13; ld.global.f32 %f84, [%rd3+8]; sub.ftz.f32 %f85, %f14, %f84; sub.ftz.f32 %f86, %f15, %f84; div.approx.ftz.f32 %f87, %f85, %f86; ld.global.f32 %f88, [%rd4+8]; sub.ftz.f32 %f89, %f7, %f88; sub.ftz.f32 %f90, %f16, %f88; div.approx.ftz.f32 %f91, %f89, %f90; add.s32 %r35, %r12, -1; add.s32 %r36, %r7, -2; setp.eq.s32 %p18, %r12, 0; selp.b32 %r37, %r36, %r35, %p18; add.s32 %r38, %r8, %r9; mad.lo.s32 %r39, %r37, %r9, %r38; add.s32 %r40, %r14, -1; setp.eq.s32 %p19, %r14, 0; selp.b32 %r41, 1, %r40, %p19; add.s32 %r42, %r41, %r39; mul.wide.s32 %rd20, %r42, 4; add.s64 %rd21, %rd2, %rd20; add.s32 %r43, %r14, %r39; mul.wide.s32 %rd22, %r43, 4; add.s64 %rd23, %rd2, %rd22; add.s32 %r44, %r14, 2; min.s32 %r45, %r44, %r10; add.s32 %r46, %r45, %r39; mul.wide.s32 %rd24, %r46, 4; add.s64 %rd25, %rd2, %rd24; ld.global.f32 %f92, [%rd23+4]; ld.global.f32 %f93, [%rd23]; sub.ftz.f32 %f94, %f93, %f92; ld.global.f32 %f95, [%rd25]; fma.rn.ftz.f32 %f96, %f94, 0f40400000, %f95; ld.global.f32 %f97, [%rd21]; sub.ftz.f32 %f98, %f96, %f97; add.ftz.f32 %f99, %f97, %f97; mul.ftz.f32 %f100, %f93, 0f40A00000; sub.ftz.f32 %f101, %f99, %f100; fma.rn.ftz.f32 %f102, %f92, 0f40800000, %f101; sub.ftz.f32 %f103, %f102, %f95; fma.rn.ftz.f32 %f104, %f91, %f98, %f103; sub.ftz.f32 %f105, %f92, %f97; fma.rn.ftz.f32 %f106, %f91, %f104, %f105; mul.ftz.f32 %f107, %f106, 0f3F000000; fma.rn.ftz.f32 %f108, %f91, %f107, %f93; mad.lo.s32 %r47, %r12, %r9, %r38; add.s32 %r48, %r41, %r47; mul.wide.s32 %rd26, %r48, 4; add.s64 %rd27, %rd2, %rd26; add.s32 %r49, %r14, %r47; mul.wide.s32 %rd28, %r49, 4; add.s64 %rd29, %rd2, %rd28; add.s32 %r50, %r45, %r47; mul.wide.s32 %rd30, %r50, 4; add.s64 %rd31, %rd2, %rd30; ld.global.f32 %f109, [%rd29+4]; ld.global.f32 %f110, [%rd29]; sub.ftz.f32 %f111, %f110, %f109; ld.global.f32 %f112, [%rd31]; fma.rn.ftz.f32 %f113, %f111, 0f40400000, %f112; ld.global.f32 %f114, [%rd27]; sub.ftz.f32 %f115, %f113, %f114; add.ftz.f32 %f116, %f114, %f114; mul.ftz.f32 %f117, %f110, 0f40A00000; sub.ftz.f32 %f118, %f116, %f117; fma.rn.ftz.f32 %f119, %f109, 0f40800000, %f118; sub.ftz.f32 %f120, %f119, %f112; fma.rn.ftz.f32 %f121, %f91, %f115, %f120; sub.ftz.f32 %f122, %f109, %f114; fma.rn.ftz.f32 %f123, %f91, %f121, %f122; mul.ftz.f32 %f124, %f123, 0f3F000000; fma.rn.ftz.f32 %f125, %f91, %f124, %f110; add.s32 %r51, %r47, %r9; add.s32 %r52, %r41, %r51; mul.wide.s32 %rd32, %r52, 4; add.s64 %rd33, %rd2, %rd32; add.s32 %r53, %r14, %r51; mul.wide.s32 %rd34, %r53, 4; add.s64 %rd35, %rd2, %rd34; add.s32 %r54, %r45, %r51; mul.wide.s32 %rd36, %r54, 4; add.s64 %rd37, %rd2, %rd36; ld.global.f32 %f126, [%rd35+4]; ld.global.f32 %f127, [%rd35]; sub.ftz.f32 %f128, %f127, %f126; ld.global.f32 %f129, [%rd37]; fma.rn.ftz.f32 %f130, %f128, 0f40400000, %f129; ld.global.f32 %f131, [%rd33]; sub.ftz.f32 %f132, %f130, %f131; add.ftz.f32 %f133, %f131, %f131; mul.ftz.f32 %f134, %f127, 0f40A00000; sub.ftz.f32 %f135, %f133, %f134; fma.rn.ftz.f32 %f136, %f126, 0f40800000, %f135; sub.ftz.f32 %f137, %f136, %f129; fma.rn.ftz.f32 %f138, %f91, %f132, %f137; sub.ftz.f32 %f139, %f126, %f131; fma.rn.ftz.f32 %f140, %f91, %f138, %f139; mul.ftz.f32 %f141, %f140, 0f3F000000; fma.rn.ftz.f32 %f142, %f91, %f141, %f127; add.s32 %r55, %r12, 2; setp.eq.s32 %p20, %r55, %r7; selp.b32 %r56, 1, %r55, %p20; mad.lo.s32 %r57, %r56, %r9, %r38; add.s32 %r58, %r41, %r57; mul.wide.s32 %rd38, %r58, 4; add.s64 %rd39, %rd2, %rd38; add.s32 %r59, %r14, %r57; mul.wide.s32 %rd40, %r59, 4; add.s64 %rd41, %rd2, %rd40; add.s32 %r60, %r45, %r57; mul.wide.s32 %rd42, %r60, 4; add.s64 %rd43, %rd2, %rd42; ld.global.f32 %f143, [%rd41+4]; ld.global.f32 %f144, [%rd41]; sub.ftz.f32 %f145, %f144, %f143; ld.global.f32 %f146, [%rd43]; fma.rn.ftz.f32 %f147, %f145, 0f40400000, %f146; ld.global.f32 %f148, [%rd39]; sub.ftz.f32 %f149, %f147, %f148; add.ftz.f32 %f150, %f148, %f148; mul.ftz.f32 %f151, %f144, 0f40A00000; sub.ftz.f32 %f152, %f150, %f151; fma.rn.ftz.f32 %f153, %f143, 0f40800000, %f152; sub.ftz.f32 %f154, %f153, %f146; fma.rn.ftz.f32 %f155, %f91, %f149, %f154; sub.ftz.f32 %f156, %f143, %f148; fma.rn.ftz.f32 %f157, %f91, %f155, %f156; mul.ftz.f32 %f158, %f157, 0f3F000000; fma.rn.ftz.f32 %f159, %f91, %f158, %f144; sub.ftz.f32 %f160, %f125, %f142; fma.rn.ftz.f32 %f161, %f160, 0f40400000, %f159; sub.ftz.f32 %f162, %f161, %f108; add.ftz.f32 %f163, %f108, %f108; mul.ftz.f32 %f164, %f125, 0f40A00000; sub.ftz.f32 %f165, %f163, %f164; fma.rn.ftz.f32 %f166, %f142, 0f40800000, %f165; sub.ftz.f32 %f167, %f166, %f159; fma.rn.ftz.f32 %f168, %f87, %f162, %f167; sub.ftz.f32 %f169, %f142, %f108; fma.rn.ftz.f32 %f170, %f87, %f168, %f169; mul.ftz.f32 %f171, %f170, 0f3F000000; fma.rn.ftz.f32 %f172, %f87, %f171, %f125; mov.f32 %f173, 0f00000000; max.ftz.f32 %f176, %f172, %f173; $L__BB55_15: mul.ftz.f32 %f19, %f174, %f176; setp.eq.s32 %p21, %r2, 255; @%p21 bra $L__BB55_17; mul.wide.u32 %rd44, %r2, 4; add.s64 %rd45, %rd1, %rd44; st.local.f32 [%rd45], %f19; $L__BB55_17: ret; } .func (.param .b32 func_retval0) _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<57>; .reg .f32 %f<235>; .reg .b32 %r<80>; .reg .b64 %rd<23>; ld.param.u64 %rd4, [_Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r76, [_Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r32, [_Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; ld.param.u32 %r31, [_Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r30, [_Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd4; bfe.u32 %r2, %r31, 8, 8; bfe.u32 %r3, %r31, 16, 8; shl.b32 %r34, %r31, 2; cvt.u64.u32 %rd5, %r34; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f217, [%rd7]; ld.local.f32 %f216, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; setp.eq.s32 %p1, %r32, 2; @%p1 bra $L__BB56_9; bra.uni $L__BB56_1; $L__BB56_9: add.ftz.f32 %f124, %f217, 0fBF000000; add.ftz.f32 %f15, %f124, %f124; add.ftz.f32 %f125, %f216, 0fBF000000; add.ftz.f32 %f16, %f125, %f125; add.ftz.f32 %f126, %f3, 0fBF000000; add.ftz.f32 %f17, %f126, %f126; mul.ftz.f32 %f127, %f16, %f16; fma.rn.ftz.f32 %f128, %f15, %f15, %f127; fma.rn.ftz.f32 %f18, %f17, %f17, %f128; setp.leu.ftz.f32 %p14, %f18, 0f00000000; mov.f32 %f216, 0f00000000; mov.f32 %f217, %f216; @%p14 bra $L__BB56_18; setp.eq.ftz.f32 %p15, %f15, 0f00000000; mov.f32 %f217, 0f00000000; setp.eq.ftz.f32 %p16, %f16, 0f00000000; and.pred %p17, %p15, %p16; @%p17 bra $L__BB56_17; abs.ftz.f32 %f19, %f16; setp.eq.ftz.f32 %p18, %f19, 0f00000000; abs.ftz.f32 %f20, %f15; setp.eq.ftz.f32 %p19, %f20, 0f00000000; and.pred %p20, %p18, %p19; mov.b32 %r7, %f16; mov.b32 %r43, %f15; and.b32 %r8, %r43, -2147483648; @%p20 bra $L__BB56_15; bra.uni $L__BB56_12; $L__BB56_15: shr.s32 %r48, %r7, 31; and.b32 %r49, %r48, 1078530011; or.b32 %r50, %r49, %r8; mov.b32 %f212, %r50; bra.uni $L__BB56_16; $L__BB56_1: setp.ne.s32 %p2, %r32, 3; @%p2 bra $L__BB56_18; add.ftz.f32 %f84, %f217, 0fBF000000; add.ftz.f32 %f4, %f84, %f84; add.ftz.f32 %f85, %f216, 0fBF000000; add.ftz.f32 %f5, %f85, %f85; mul.ftz.f32 %f86, %f5, %f5; fma.rn.ftz.f32 %f87, %f4, %f4, %f86; sqrt.approx.ftz.f32 %f6, %f87; setp.leu.ftz.f32 %p3, %f6, 0f00000000; mov.f32 %f216, 0f00000000; mov.f32 %f217, %f216; @%p3 bra $L__BB56_18; div.approx.ftz.f32 %f88, %f4, %f6; div.approx.ftz.f32 %f89, %f5, %f6; abs.ftz.f32 %f7, %f89; abs.ftz.f32 %f8, %f88; setp.eq.ftz.f32 %p4, %f7, 0f00000000; setp.eq.ftz.f32 %p5, %f8, 0f00000000; and.pred %p6, %p4, %p5; mov.b32 %r5, %f89; mov.b32 %r35, %f88; and.b32 %r6, %r35, -2147483648; @%p6 bra $L__BB56_7; bra.uni $L__BB56_4; $L__BB56_7: shr.s32 %r40, %r5, 31; and.b32 %r41, %r40, 1078530011; or.b32 %r42, %r41, %r6; mov.b32 %f211, %r42; bra.uni $L__BB56_8; $L__BB56_12: setp.eq.ftz.f32 %p21, %f19, 0f7F800000; setp.eq.ftz.f32 %p22, %f20, 0f7F800000; and.pred %p23, %p21, %p22; @%p23 bra $L__BB56_14; bra.uni $L__BB56_13; $L__BB56_14: setp.lt.s32 %p27, %r7, 0; selp.b32 %r46, 1075235812, 1061752795, %p27; or.b32 %r47, %r46, %r8; mov.b32 %f212, %r47; bra.uni $L__BB56_16; $L__BB56_4: setp.eq.ftz.f32 %p7, %f7, 0f7F800000; setp.eq.ftz.f32 %p8, %f8, 0f7F800000; and.pred %p9, %p7, %p8; @%p9 bra $L__BB56_6; bra.uni $L__BB56_5; $L__BB56_6: setp.lt.s32 %p13, %r5, 0; selp.b32 %r38, 1075235812, 1061752795, %p13; or.b32 %r39, %r38, %r6; mov.b32 %f211, %r39; bra.uni $L__BB56_8; $L__BB56_13: setp.lt.s32 %p24, %r7, 0; min.ftz.f32 %f130, %f20, %f19; max.ftz.f32 %f131, %f20, %f19; div.full.ftz.f32 %f132, %f130, %f131; mul.rn.ftz.f32 %f133, %f132, %f132; mov.f32 %f134, 0fC0B59883; mov.f32 %f135, 0fBF52C7EA; fma.rn.ftz.f32 %f136, %f133, %f135, %f134; mov.f32 %f137, 0fC0D21907; fma.rn.ftz.f32 %f138, %f136, %f133, %f137; mul.ftz.f32 %f139, %f133, %f138; mul.ftz.f32 %f140, %f132, %f139; add.ftz.f32 %f141, %f133, 0f41355DC0; mov.f32 %f142, 0f41E6BD60; fma.rn.ftz.f32 %f143, %f141, %f133, %f142; mov.f32 %f144, 0f419D92C8; fma.rn.ftz.f32 %f145, %f143, %f133, %f144; rcp.approx.ftz.f32 %f146, %f145; fma.rn.ftz.f32 %f147, %f140, %f146, %f132; mov.f32 %f148, 0f3FC90FDB; sub.ftz.f32 %f149, %f148, %f147; setp.gt.ftz.f32 %p25, %f20, %f19; selp.f32 %f150, %f149, %f147, %p25; mov.f32 %f151, 0f40490FDB; sub.ftz.f32 %f152, %f151, %f150; selp.f32 %f153, %f152, %f150, %p24; mov.b32 %r44, %f153; or.b32 %r45, %r8, %r44; mov.b32 %f154, %r45; add.ftz.f32 %f155, %f19, %f20; setp.le.ftz.f32 %p26, %f155, 0f7F800000; selp.f32 %f212, %f154, %f155, %p26; $L__BB56_16: fma.rn.ftz.f32 %f217, %f212, 0fBE22F983, 0f3F000000; $L__BB56_17: sqrt.approx.ftz.f32 %f156, %f18; div.approx.ftz.f32 %f157, %f17, %f156; mov.f32 %f158, 0fBF800000; max.ftz.f32 %f159, %f157, %f158; mov.f32 %f160, 0f3F800000; min.ftz.f32 %f161, %f159, %f160; abs.ftz.f32 %f162, %f161; sub.ftz.f32 %f163, %f160, %f162; mul.ftz.f32 %f164, %f163, 0f3F000000; sqrt.approx.ftz.f32 %f165, %f164; setp.gt.ftz.f32 %p28, %f162, 0f3F11EB85; selp.f32 %f166, %f165, %f162, %p28; mul.ftz.f32 %f167, %f166, %f166; mov.f32 %f168, 0f3C94D2E9; mov.f32 %f169, 0f3D53F941; fma.rn.ftz.f32 %f170, %f169, %f167, %f168; mov.f32 %f171, 0f3D3F841F; fma.rn.ftz.f32 %f172, %f170, %f167, %f171; mov.f32 %f173, 0f3D994929; fma.rn.ftz.f32 %f174, %f172, %f167, %f173; mov.f32 %f175, 0f3E2AAB94; fma.rn.ftz.f32 %f176, %f174, %f167, %f175; mul.ftz.f32 %f177, %f167, %f176; fma.rn.ftz.f32 %f178, %f177, %f166, %f166; add.ftz.f32 %f179, %f178, %f178; mov.f32 %f180, 0f3FC90FDB; sub.ftz.f32 %f181, %f180, %f178; selp.f32 %f182, %f179, %f181, %p28; setp.lt.ftz.f32 %p29, %f161, 0f00000000; mov.f32 %f183, 0f40490FDB; sub.ftz.f32 %f184, %f183, %f182; selp.f32 %f185, %f184, %f182, %p29; fma.rn.ftz.f32 %f216, %f185, 0fBEA2F983, 0f3F800000; bra.uni $L__BB56_18; $L__BB56_5: setp.lt.s32 %p10, %r5, 0; min.ftz.f32 %f90, %f8, %f7; max.ftz.f32 %f91, %f8, %f7; div.full.ftz.f32 %f92, %f90, %f91; mul.rn.ftz.f32 %f93, %f92, %f92; mov.f32 %f94, 0fC0B59883; mov.f32 %f95, 0fBF52C7EA; fma.rn.ftz.f32 %f96, %f93, %f95, %f94; mov.f32 %f97, 0fC0D21907; fma.rn.ftz.f32 %f98, %f96, %f93, %f97; mul.ftz.f32 %f99, %f93, %f98; mul.ftz.f32 %f100, %f92, %f99; add.ftz.f32 %f101, %f93, 0f41355DC0; mov.f32 %f102, 0f41E6BD60; fma.rn.ftz.f32 %f103, %f101, %f93, %f102; mov.f32 %f104, 0f419D92C8; fma.rn.ftz.f32 %f105, %f103, %f93, %f104; rcp.approx.ftz.f32 %f106, %f105; fma.rn.ftz.f32 %f107, %f100, %f106, %f92; mov.f32 %f108, 0f3FC90FDB; sub.ftz.f32 %f109, %f108, %f107; setp.gt.ftz.f32 %p11, %f8, %f7; selp.f32 %f110, %f109, %f107, %p11; mov.f32 %f111, 0f40490FDB; sub.ftz.f32 %f112, %f111, %f110; selp.f32 %f113, %f112, %f110, %p10; mov.b32 %r36, %f113; or.b32 %r37, %r6, %r36; mov.b32 %f114, %r37; add.ftz.f32 %f115, %f7, %f8; setp.le.ftz.f32 %p12, %f115, 0f7F800000; selp.f32 %f211, %f114, %f115, %p12; $L__BB56_8: mov.f32 %f116, 0f40490FDB; div.approx.ftz.f32 %f117, %f211, %f116; mov.f32 %f118, 0f3F800000; sub.ftz.f32 %f119, %f118, %f117; mul.ftz.f32 %f217, %f119, 0f3F000000; add.ftz.f32 %f120, %f3, 0fBF000000; fma.rn.ftz.f32 %f121, %f120, 0f40000000, 0f3F800000; mul.ftz.f32 %f216, %f121, 0f3F000000; $L__BB56_18: setp.gt.s32 %p30, %r30, 0; @%p30 bra $L__BB56_20; bra.uni $L__BB56_19; $L__BB56_20: add.s32 %r78, %r30, %r76; cvt.rzi.ftz.s32.f32 %r12, %f216; cvt.rzi.ftz.s32.f32 %r13, %f217; or.b32 %r52, %r12, %r13; setp.lt.s32 %p31, %r52, 0; mov.u32 %r79, -1; setp.gt.s32 %p32, %r13, 9; or.pred %p33, %p32, %p31; @%p33 bra $L__BB56_28; mad.lo.s32 %r54, %r12, 10, %r13; add.s32 %r14, %r54, 1001; ld.const.u64 %rd8, [kernel_params+2400]; cvta.to.global.u64 %rd2, %rd8; mov.u32 %r75, 0; $L__BB56_22: mul.wide.s32 %rd9, %r76, 16; add.s64 %rd10, %rd2, %rd9; ld.global.v4.u32 {%r55, %r56, %r57, %r77}, [%rd10]; setp.eq.s32 %p34, %r55, %r14; @%p34 bra $L__BB56_25; setp.eq.s32 %p35, %r57, %r14; add.s32 %r75, %r75, 1; @%p35 bra $L__BB56_26; add.s32 %r76, %r76, 1; setp.lt.s32 %p36, %r75, %r30; mov.u32 %r77, -1; @%p36 bra $L__BB56_22; bra.uni $L__BB56_26; $L__BB56_19: neg.s32 %r79, %r30; mov.u32 %r78, %r76; bra.uni $L__BB56_28; $L__BB56_25: mov.u32 %r77, %r56; $L__BB56_26: setp.eq.s32 %p37, %r77, -1; @%p37 bra $L__BB56_28; cvt.rn.f32.s32 %f186, %r13; sub.ftz.f32 %f217, %f217, %f186; cvt.rn.f32.s32 %f187, %r12; sub.ftz.f32 %f216, %f216, %f187; mov.u32 %r79, %r77; $L__BB56_28: setp.eq.s32 %p38, %r79, -1; mov.f32 %f219, 0f3F800000; mov.f32 %f220, 0f00000000; mov.f32 %f221, %f219; mov.f32 %f222, %f219; @%p38 bra $L__BB56_50; ld.const.u64 %rd11, [kernel_params+2432]; cvta.to.global.u64 %rd12, %rd11; mul.wide.s32 %rd13, %r79, 96; add.s64 %rd14, %rd12, %rd13; add.s64 %rd3, %rd14, 8; ld.global.v2.u32 {%r62, %r63}, [%rd14+8]; setp.gt.u32 %p39, %r62, 6; @%p39 bra $L__BB56_34; mov.u32 %r64, 1; shl.b32 %r65, %r64, %r62; and.b32 %r66, %r65, 71; setp.ne.s32 %p40, %r66, 0; @%p40 bra $L__BB56_31; bra.uni $L__BB56_34; $L__BB56_31: and.b32 %r67, %r63, -2; setp.eq.s32 %p41, %r67, 2; @%p41 bra $L__BB56_33; bra.uni $L__BB56_32; $L__BB56_33: ld.global.u64 %rd16, [%rd3+-8]; ld.global.u32 %r68, [%rd3+12]; ld.global.u32 %r69, [%rd3+16]; { // callseq 38, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.b32 [param1+0], %r68; .param .b32 param2; st.param.b32 [param2+0], %r69; .param .b32 param3; st.param.f32 [param3+0], %f217; .param .b32 param4; st.param.f32 [param4+0], %f216; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f219, [retval0+0]; ld.param.f32 %f220, [retval0+4]; ld.param.f32 %f221, [retval0+8]; ld.param.f32 %f222, [retval0+12]; } // callseq 38 bra.uni $L__BB56_38; $L__BB56_32: ld.global.u64 %rd15, [%rd3+-8]; tex.2d.v4.f32.f32 {%f219, %f220, %f221, %f222}, [%rd15, {%f217, %f216}]; bra.uni $L__BB56_38; $L__BB56_34: and.b32 %r70, %r63, -2; setp.eq.s32 %p42, %r70, 2; @%p42 bra $L__BB56_36; bra.uni $L__BB56_35; $L__BB56_36: ld.global.u64 %rd18, [%rd3+-8]; ld.global.u32 %r71, [%rd3+12]; ld.global.u32 %r72, [%rd3+16]; { // callseq 39, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd18; .param .b32 param1; st.param.b32 [param1+0], %r71; .param .b32 param2; st.param.b32 [param2+0], %r72; .param .b32 param3; st.param.f32 [param3+0], %f217; .param .b32 param4; st.param.f32 [param4+0], %f216; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f219, [retval0+0]; } // callseq 39 bra.uni $L__BB56_37; $L__BB56_35: ld.global.u64 %rd17, [%rd3+-8]; tex.2d.v4.f32.f32 {%f219, %f192, %f193, %f194}, [%rd17, {%f217, %f216}]; $L__BB56_37: mov.f32 %f222, 0f3F800000; mov.f32 %f220, %f219; mov.f32 %f221, %f219; $L__BB56_38: and.b32 %r73, %r31, 33554432; setp.eq.s32 %p43, %r73, 0; setp.eq.ftz.f32 %p44, %f222, 0f3F800000; or.pred %p45, %p43, %p44; setp.eq.ftz.f32 %p46, %f222, 0f00000000; or.pred %p47, %p46, %p45; @%p47 bra $L__BB56_40; rcp.approx.ftz.f32 %f195, %f222; mul.ftz.f32 %f219, %f219, %f195; mul.ftz.f32 %f220, %f220, %f195; mul.ftz.f32 %f221, %f221, %f195; $L__BB56_40: and.b32 %r74, %r31, 16777216; setp.eq.s32 %p48, %r74, 0; @%p48 bra $L__BB56_50; setp.lt.ftz.f32 %p49, %f219, 0f3D25AEE6; @%p49 bra $L__BB56_43; bra.uni $L__BB56_42; $L__BB56_43: setp.lt.ftz.f32 %p50, %f219, 0f00000000; mul.ftz.f32 %f200, %f219, 0f3D9E8391; selp.f32 %f219, 0f00000000, %f200, %p50; bra.uni $L__BB56_44; $L__BB56_42: add.ftz.f32 %f196, %f219, 0f3D6147AE; mul.ftz.f32 %f197, %f196, 0f3F72A76F; lg2.approx.ftz.f32 %f198, %f197; mul.ftz.f32 %f199, %f198, 0f4019999A; ex2.approx.ftz.f32 %f219, %f199; $L__BB56_44: setp.lt.ftz.f32 %p51, %f220, 0f3D25AEE6; @%p51 bra $L__BB56_46; bra.uni $L__BB56_45; $L__BB56_46: setp.lt.ftz.f32 %p52, %f220, 0f00000000; mul.ftz.f32 %f205, %f220, 0f3D9E8391; selp.f32 %f220, 0f00000000, %f205, %p52; bra.uni $L__BB56_47; $L__BB56_45: add.ftz.f32 %f201, %f220, 0f3D6147AE; mul.ftz.f32 %f202, %f201, 0f3F72A76F; lg2.approx.ftz.f32 %f203, %f202; mul.ftz.f32 %f204, %f203, 0f4019999A; ex2.approx.ftz.f32 %f220, %f204; $L__BB56_47: setp.lt.ftz.f32 %p53, %f221, 0f3D25AEE6; @%p53 bra $L__BB56_49; bra.uni $L__BB56_48; $L__BB56_49: setp.lt.ftz.f32 %p54, %f221, 0f00000000; mul.ftz.f32 %f210, %f221, 0f3D9E8391; selp.f32 %f221, 0f00000000, %f210, %p54; bra.uni $L__BB56_50; $L__BB56_48: add.ftz.f32 %f206, %f221, 0f3D6147AE; mul.ftz.f32 %f207, %f206, 0f3F72A76F; lg2.approx.ftz.f32 %f208, %f207; mul.ftz.f32 %f209, %f208, 0f4019999A; ex2.approx.ftz.f32 %f221, %f209; $L__BB56_50: setp.eq.s32 %p55, %r2, 255; @%p55 bra $L__BB56_52; mul.wide.u32 %rd19, %r2, 4; add.s64 %rd20, %rd1, %rd19; st.local.f32 [%rd20], %f219; st.local.f32 [%rd20+4], %f220; st.local.f32 [%rd20+8], %f221; $L__BB56_52: setp.eq.s32 %p56, %r3, 255; @%p56 bra $L__BB56_54; mul.wide.u32 %rd21, %r3, 4; add.s64 %rd22, %rd1, %rd21; st.local.f32 [%rd22], %f222; $L__BB56_54: st.param.b32 [func_retval0+0], %r78; ret; } .func _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<75>; .reg .b16 %rs<3>; .reg .f32 %f<514>; .reg .b32 %r<60>; .reg .b64 %rd<49>; ld.param.u64 %rd7, [_Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd8, [_Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r11, [_Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r10, [_Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r9, [_Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd1, %rd8; cvta.to.local.u64 %rd9, %rd7; add.s64 %rd2, %rd9, 12; ld.local.f32 %f445, [%rd9+12]; ld.local.f32 %f446, [%rd9+16]; ld.local.f32 %f447, [%rd9+20]; ld.local.u8 %rs1, [%rd9+56]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p1, %rs2, 0; ld.local.s32 %rd3, [%rd9+76]; @%p1 bra $L__BB57_4; bra.uni $L__BB57_1; $L__BB57_4: cvt.u32.u64 %r14, %rd3; setp.ne.s32 %p4, %r14, -1; @%p4 bra $L__BB57_7; bra.uni $L__BB57_5; $L__BB57_7: ld.const.u64 %rd14, [kernel_params+2160]; cvta.to.global.u64 %rd15, %rd14; mul.lo.s64 %rd16, %rd3, 208; add.s64 %rd17, %rd15, %rd16; ld.global.v4.f32 {%f280, %f281, %f282, %f283}, [%rd17]; ld.global.v4.f32 {%f287, %f288, %f289, %f290}, [%rd17+16]; ld.global.v4.f32 {%f294, %f295, %f296, %f297}, [%rd17+32]; mul.ftz.f32 %f301, %f446, %f287; fma.rn.ftz.f32 %f302, %f445, %f280, %f301; fma.rn.ftz.f32 %f303, %f447, %f294, %f302; mul.ftz.f32 %f304, %f446, %f288; fma.rn.ftz.f32 %f305, %f445, %f281, %f304; fma.rn.ftz.f32 %f306, %f447, %f295, %f305; mul.ftz.f32 %f307, %f446, %f289; fma.rn.ftz.f32 %f308, %f445, %f282, %f307; fma.rn.ftz.f32 %f309, %f447, %f296, %f308; mul.ftz.f32 %f310, %f306, %f306; fma.rn.ftz.f32 %f311, %f303, %f303, %f310; fma.rn.ftz.f32 %f312, %f309, %f309, %f311; sqrt.approx.ftz.f32 %f313, %f312; div.approx.ftz.f32 %f445, %f303, %f313; div.approx.ftz.f32 %f446, %f306, %f313; div.approx.ftz.f32 %f447, %f309, %f313; bra.uni $L__BB57_8; $L__BB57_1: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p2, %r12, -1; @%p2 bra $L__BB57_3; ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p3, %r13, 32; @%p3 bra $L__BB57_8; $L__BB57_3: ld.local.v4.f32 {%f212, %f213, %f214, %f215}, [%rd2+132]; ld.local.v4.f32 {%f219, %f220, %f221, %f222}, [%rd2+148]; ld.local.v4.f32 {%f226, %f227, %f228, %f229}, [%rd2+164]; mul.ftz.f32 %f233, %f446, %f219; fma.rn.ftz.f32 %f234, %f445, %f212, %f233; fma.rn.ftz.f32 %f235, %f447, %f226, %f234; mul.ftz.f32 %f236, %f446, %f220; fma.rn.ftz.f32 %f237, %f445, %f213, %f236; fma.rn.ftz.f32 %f238, %f447, %f227, %f237; mul.ftz.f32 %f239, %f446, %f221; fma.rn.ftz.f32 %f240, %f445, %f214, %f239; fma.rn.ftz.f32 %f241, %f447, %f228, %f240; mul.ftz.f32 %f242, %f238, %f238; fma.rn.ftz.f32 %f243, %f235, %f235, %f242; fma.rn.ftz.f32 %f244, %f241, %f241, %f243; sqrt.approx.ftz.f32 %f245, %f244; div.approx.ftz.f32 %f445, %f235, %f245; div.approx.ftz.f32 %f446, %f238, %f245; div.approx.ftz.f32 %f447, %f241, %f245; $L__BB57_8: abs.ftz.f32 %f314, %f446; abs.ftz.f32 %f315, %f445; add.ftz.f32 %f316, %f315, %f314; abs.ftz.f32 %f317, %f447; add.ftz.f32 %f318, %f316, %f317; rcp.approx.ftz.f32 %f319, %f318; mul.ftz.f32 %f16, %f315, %f319; mul.ftz.f32 %f17, %f314, %f319; mul.ftz.f32 %f18, %f317, %f319; mov.b32 %f19, %r11; add.ftz.f32 %f320, %f19, 0f3F800000; mul.ftz.f32 %f20, %f320, 0f3F000000; add.ftz.f32 %f21, %f16, %f17; mul.ftz.f32 %f22, %f20, %f21; setp.leu.ftz.f32 %p6, %f16, %f22; @%p6 bra $L__BB57_10; add.ftz.f32 %f324, %f16, %f18; mul.ftz.f32 %f325, %f20, %f324; setp.gt.ftz.f32 %p7, %f16, %f325; mov.f32 %f449, 0f00000000; mov.f32 %f448, 0f3F800000; mov.f32 %f450, %f449; @%p7 bra $L__BB57_22; $L__BB57_10: setp.leu.ftz.f32 %p8, %f17, %f22; @%p8 bra $L__BB57_12; add.ftz.f32 %f329, %f17, %f18; mul.ftz.f32 %f330, %f20, %f329; setp.gt.ftz.f32 %p9, %f17, %f330; mov.f32 %f448, 0f00000000; mov.f32 %f449, 0f3F800000; mov.f32 %f450, %f448; @%p9 bra $L__BB57_22; $L__BB57_12: add.ftz.f32 %f23, %f16, %f18; mul.ftz.f32 %f331, %f20, %f23; setp.leu.ftz.f32 %p10, %f18, %f331; @%p10 bra $L__BB57_14; add.ftz.f32 %f335, %f17, %f18; mul.ftz.f32 %f336, %f20, %f335; setp.gt.ftz.f32 %p11, %f18, %f336; mov.f32 %f450, 0f3F800000; mov.f32 %f448, 0f00000000; mov.f32 %f449, %f448; @%p11 bra $L__BB57_22; $L__BB57_14: setp.leu.ftz.f32 %p12, %f19, 0f00000000; mov.f32 %f450, 0f00000000; mov.f32 %f448, 0f3F800000; mov.f32 %f449, %f450; @%p12 bra $L__BB57_22; mov.f32 %f340, 0f3F800000; sub.ftz.f32 %f24, %f340, %f20; mul.ftz.f32 %f341, %f24, %f21; setp.lt.ftz.f32 %p13, %f18, %f341; @%p13 bra $L__BB57_21; bra.uni $L__BB57_16; $L__BB57_21: div.approx.ftz.f32 %f364, %f16, %f21; sub.ftz.f32 %f366, %f340, %f19; fma.rn.ftz.f32 %f367, %f366, 0fBF000000, %f364; div.approx.ftz.f32 %f368, %f367, %f19; cvt.ftz.sat.f32.f32 %f448, %f368; sub.ftz.f32 %f449, %f340, %f448; bra.uni $L__BB57_22; $L__BB57_5: ld.local.u32 %r15, [%rd2+52]; setp.ne.s32 %p5, %r15, 32; @%p5 bra $L__BB57_8; ld.const.u64 %rd10, [kernel_params+2368]; cvta.to.global.u64 %rd11, %rd10; ld.local.u32 %r16, [%rd2+68]; mul.wide.s32 %rd12, %r16, 192; add.s64 %rd13, %rd11, %rd12; ld.global.v4.f32 {%f246, %f247, %f248, %f249}, [%rd13+48]; ld.global.v4.f32 {%f253, %f254, %f255, %f256}, [%rd13+64]; ld.global.v4.f32 {%f260, %f261, %f262, %f263}, [%rd13+80]; mul.ftz.f32 %f267, %f446, %f253; fma.rn.ftz.f32 %f268, %f445, %f246, %f267; fma.rn.ftz.f32 %f269, %f447, %f260, %f268; mul.ftz.f32 %f270, %f446, %f254; fma.rn.ftz.f32 %f271, %f445, %f247, %f270; fma.rn.ftz.f32 %f272, %f447, %f261, %f271; mul.ftz.f32 %f273, %f446, %f255; fma.rn.ftz.f32 %f274, %f445, %f248, %f273; fma.rn.ftz.f32 %f275, %f447, %f262, %f274; mul.ftz.f32 %f276, %f272, %f272; fma.rn.ftz.f32 %f277, %f269, %f269, %f276; fma.rn.ftz.f32 %f278, %f275, %f275, %f277; sqrt.approx.ftz.f32 %f279, %f278; div.approx.ftz.f32 %f445, %f269, %f279; div.approx.ftz.f32 %f446, %f272, %f279; div.approx.ftz.f32 %f447, %f275, %f279; bra.uni $L__BB57_8; $L__BB57_16: add.ftz.f32 %f25, %f17, %f18; mul.ftz.f32 %f342, %f24, %f25; setp.lt.ftz.f32 %p14, %f16, %f342; @%p14 bra $L__BB57_20; bra.uni $L__BB57_17; $L__BB57_20: div.approx.ftz.f32 %f358, %f17, %f25; sub.ftz.f32 %f360, %f340, %f19; fma.rn.ftz.f32 %f361, %f360, 0fBF000000, %f358; div.approx.ftz.f32 %f362, %f361, %f19; cvt.ftz.sat.f32.f32 %f449, %f362; sub.ftz.f32 %f450, %f340, %f449; mov.f32 %f448, 0f00000000; bra.uni $L__BB57_22; $L__BB57_17: mul.ftz.f32 %f343, %f24, %f23; setp.lt.ftz.f32 %p15, %f17, %f343; @%p15 bra $L__BB57_19; bra.uni $L__BB57_18; $L__BB57_19: div.approx.ftz.f32 %f352, %f16, %f23; sub.ftz.f32 %f354, %f340, %f19; fma.rn.ftz.f32 %f355, %f354, 0fBF000000, %f352; div.approx.ftz.f32 %f356, %f355, %f19; cvt.ftz.sat.f32.f32 %f448, %f356; sub.ftz.f32 %f30, %f340, %f448; mov.f32 %f449, %f450; mov.f32 %f450, %f30; bra.uni $L__BB57_22; $L__BB57_18: mov.f32 %f344, 0f40000000; sub.ftz.f32 %f345, %f344, %f20; add.ftz.f32 %f346, %f20, 0fBF800000; fma.rn.ftz.f32 %f347, %f345, %f16, %f346; fma.rn.ftz.f32 %f348, %f20, 0f40000000, 0fBF800000; div.approx.ftz.f32 %f448, %f347, %f348; fma.rn.ftz.f32 %f349, %f345, %f17, %f346; div.approx.ftz.f32 %f449, %f349, %f348; fma.rn.ftz.f32 %f350, %f345, %f18, %f346; div.approx.ftz.f32 %f450, %f350, %f348; $L__BB57_22: bfe.u32 %r3, %r10, 8, 8; bfe.u32 %r4, %r10, 16, 8; shl.b32 %r17, %r10, 2; cvt.u64.u32 %rd18, %r17; and.b64 %rd19, %rd18, 1020; add.s64 %rd20, %rd1, %rd19; ld.local.f32 %f38, [%rd20]; ld.local.f32 %f39, [%rd20+4]; ld.local.f32 %f40, [%rd20+8]; setp.leu.ftz.f32 %p16, %f448, 0f00000000; mov.f32 %f489, 0f00000000; mov.f32 %f490, %f489; mov.f32 %f491, %f489; mov.f32 %f492, %f489; @%p16 bra $L__BB57_46; setp.lt.ftz.f32 %p17, %f445, 0f00000000; mov.f32 %f452, 0f3F800000; sub.ftz.f32 %f377, %f452, %f39; selp.f32 %f41, %f377, %f39, %p17; setp.eq.s32 %p18, %r9, -1; mov.f32 %f453, 0f00000000; mov.f32 %f454, %f452; mov.f32 %f455, %f452; @%p18 bra $L__BB57_45; ld.const.u64 %rd21, [kernel_params+2432]; cvta.to.global.u64 %rd22, %rd21; mul.wide.s32 %rd23, %r9, 96; add.s64 %rd24, %rd22, %rd23; add.s64 %rd4, %rd24, 8; ld.global.v2.u32 {%r19, %r20}, [%rd24+8]; setp.gt.u32 %p19, %r19, 6; @%p19 bra $L__BB57_29; mov.u32 %r21, 1; shl.b32 %r22, %r21, %r19; and.b32 %r23, %r22, 71; setp.ne.s32 %p20, %r23, 0; @%p20 bra $L__BB57_26; bra.uni $L__BB57_29; $L__BB57_26: and.b32 %r24, %r20, -2; setp.eq.s32 %p21, %r24, 2; @%p21 bra $L__BB57_28; bra.uni $L__BB57_27; $L__BB57_28: ld.global.u64 %rd26, [%rd4+-8]; ld.global.u32 %r25, [%rd4+12]; ld.global.u32 %r26, [%rd4+16]; { // callseq 40, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd26; .param .b32 param1; st.param.b32 [param1+0], %r25; .param .b32 param2; st.param.b32 [param2+0], %r26; .param .b32 param3; st.param.f32 [param3+0], %f41; .param .b32 param4; st.param.f32 [param4+0], %f40; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f452, [retval0+0]; ld.param.f32 %f453, [retval0+4]; ld.param.f32 %f454, [retval0+8]; ld.param.f32 %f455, [retval0+12]; } // callseq 40 bra.uni $L__BB57_33; $L__BB57_27: ld.global.u64 %rd25, [%rd4+-8]; tex.2d.v4.f32.f32 {%f452, %f453, %f454, %f455}, [%rd25, {%f41, %f40}]; bra.uni $L__BB57_33; $L__BB57_29: and.b32 %r27, %r20, -2; setp.eq.s32 %p22, %r27, 2; @%p22 bra $L__BB57_31; bra.uni $L__BB57_30; $L__BB57_31: ld.global.u64 %rd28, [%rd4+-8]; ld.global.u32 %r28, [%rd4+12]; ld.global.u32 %r29, [%rd4+16]; { // callseq 41, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd28; .param .b32 param1; st.param.b32 [param1+0], %r28; .param .b32 param2; st.param.b32 [param2+0], %r29; .param .b32 param3; st.param.f32 [param3+0], %f41; .param .b32 param4; st.param.f32 [param4+0], %f40; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f452, [retval0+0]; } // callseq 41 bra.uni $L__BB57_32; $L__BB57_30: ld.global.u64 %rd27, [%rd4+-8]; tex.2d.v4.f32.f32 {%f452, %f378, %f379, %f380}, [%rd27, {%f41, %f40}]; $L__BB57_32: mov.f32 %f455, 0f3F800000; mov.f32 %f453, %f452; mov.f32 %f454, %f452; $L__BB57_33: and.b32 %r30, %r10, 33554432; setp.eq.s32 %p23, %r30, 0; setp.eq.ftz.f32 %p24, %f455, 0f3F800000; or.pred %p25, %p23, %p24; setp.eq.ftz.f32 %p26, %f455, 0f00000000; or.pred %p27, %p26, %p25; @%p27 bra $L__BB57_35; rcp.approx.ftz.f32 %f381, %f455; mul.ftz.f32 %f452, %f452, %f381; mul.ftz.f32 %f453, %f453, %f381; mul.ftz.f32 %f454, %f454, %f381; $L__BB57_35: and.b32 %r31, %r10, 16777216; setp.eq.s32 %p28, %r31, 0; @%p28 bra $L__BB57_45; setp.lt.ftz.f32 %p29, %f452, 0f3D25AEE6; @%p29 bra $L__BB57_38; bra.uni $L__BB57_37; $L__BB57_38: setp.lt.ftz.f32 %p30, %f452, 0f00000000; mul.ftz.f32 %f386, %f452, 0f3D9E8391; selp.f32 %f452, 0f00000000, %f386, %p30; bra.uni $L__BB57_39; $L__BB57_37: add.ftz.f32 %f382, %f452, 0f3D6147AE; mul.ftz.f32 %f383, %f382, 0f3F72A76F; lg2.approx.ftz.f32 %f384, %f383; mul.ftz.f32 %f385, %f384, 0f4019999A; ex2.approx.ftz.f32 %f452, %f385; $L__BB57_39: setp.lt.ftz.f32 %p31, %f453, 0f3D25AEE6; @%p31 bra $L__BB57_41; bra.uni $L__BB57_40; $L__BB57_41: setp.lt.ftz.f32 %p32, %f453, 0f00000000; mul.ftz.f32 %f391, %f453, 0f3D9E8391; selp.f32 %f453, 0f00000000, %f391, %p32; bra.uni $L__BB57_42; $L__BB57_40: add.ftz.f32 %f387, %f453, 0f3D6147AE; mul.ftz.f32 %f388, %f387, 0f3F72A76F; lg2.approx.ftz.f32 %f389, %f388; mul.ftz.f32 %f390, %f389, 0f4019999A; ex2.approx.ftz.f32 %f453, %f390; $L__BB57_42: setp.lt.ftz.f32 %p33, %f454, 0f3D25AEE6; @%p33 bra $L__BB57_44; bra.uni $L__BB57_43; $L__BB57_44: setp.lt.ftz.f32 %p34, %f454, 0f00000000; mul.ftz.f32 %f396, %f454, 0f3D9E8391; selp.f32 %f454, 0f00000000, %f396, %p34; bra.uni $L__BB57_45; $L__BB57_43: add.ftz.f32 %f392, %f454, 0f3D6147AE; mul.ftz.f32 %f393, %f392, 0f3F72A76F; lg2.approx.ftz.f32 %f394, %f393; mul.ftz.f32 %f395, %f394, 0f4019999A; ex2.approx.ftz.f32 %f454, %f395; $L__BB57_45: fma.rn.ftz.f32 %f489, %f448, %f452, 0f00000000; fma.rn.ftz.f32 %f490, %f448, %f453, 0f00000000; fma.rn.ftz.f32 %f491, %f448, %f454, 0f00000000; fma.rn.ftz.f32 %f492, %f448, %f455, 0f00000000; $L__BB57_46: setp.leu.ftz.f32 %p35, %f449, 0f00000000; @%p35 bra $L__BB57_70; setp.gt.ftz.f32 %p36, %f446, 0f00000000; mov.f32 %f473, 0f3F800000; sub.ftz.f32 %f401, %f473, %f38; selp.f32 %f98, %f401, %f38, %p36; setp.eq.s32 %p37, %r9, -1; mov.f32 %f474, 0f00000000; mov.f32 %f475, %f473; mov.f32 %f476, %f473; @%p37 bra $L__BB57_69; ld.const.u64 %rd29, [kernel_params+2432]; cvta.to.global.u64 %rd30, %rd29; mul.wide.s32 %rd31, %r9, 96; add.s64 %rd32, %rd30, %rd31; add.s64 %rd5, %rd32, 8; ld.global.v2.u32 {%r33, %r34}, [%rd32+8]; setp.gt.u32 %p38, %r33, 6; @%p38 bra $L__BB57_53; mov.u32 %r35, 1; shl.b32 %r36, %r35, %r33; and.b32 %r37, %r36, 71; setp.ne.s32 %p39, %r37, 0; @%p39 bra $L__BB57_50; bra.uni $L__BB57_53; $L__BB57_50: and.b32 %r38, %r34, -2; setp.eq.s32 %p40, %r38, 2; @%p40 bra $L__BB57_52; bra.uni $L__BB57_51; $L__BB57_52: ld.global.u64 %rd34, [%rd5+-8]; ld.global.u32 %r39, [%rd5+12]; ld.global.u32 %r40, [%rd5+16]; { // callseq 42, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd34; .param .b32 param1; st.param.b32 [param1+0], %r39; .param .b32 param2; st.param.b32 [param2+0], %r40; .param .b32 param3; st.param.f32 [param3+0], %f98; .param .b32 param4; st.param.f32 [param4+0], %f40; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f473, [retval0+0]; ld.param.f32 %f474, [retval0+4]; ld.param.f32 %f475, [retval0+8]; ld.param.f32 %f476, [retval0+12]; } // callseq 42 bra.uni $L__BB57_57; $L__BB57_51: ld.global.u64 %rd33, [%rd5+-8]; tex.2d.v4.f32.f32 {%f473, %f474, %f475, %f476}, [%rd33, {%f98, %f40}]; bra.uni $L__BB57_57; $L__BB57_53: and.b32 %r41, %r34, -2; setp.eq.s32 %p41, %r41, 2; @%p41 bra $L__BB57_55; bra.uni $L__BB57_54; $L__BB57_55: ld.global.u64 %rd36, [%rd5+-8]; ld.global.u32 %r42, [%rd5+12]; ld.global.u32 %r43, [%rd5+16]; { // callseq 43, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.b32 [param1+0], %r42; .param .b32 param2; st.param.b32 [param2+0], %r43; .param .b32 param3; st.param.f32 [param3+0], %f98; .param .b32 param4; st.param.f32 [param4+0], %f40; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f473, [retval0+0]; } // callseq 43 bra.uni $L__BB57_56; $L__BB57_54: ld.global.u64 %rd35, [%rd5+-8]; tex.2d.v4.f32.f32 {%f473, %f402, %f403, %f404}, [%rd35, {%f98, %f40}]; $L__BB57_56: mov.f32 %f476, 0f3F800000; mov.f32 %f474, %f473; mov.f32 %f475, %f473; $L__BB57_57: and.b32 %r44, %r10, 33554432; setp.eq.s32 %p42, %r44, 0; setp.eq.ftz.f32 %p43, %f476, 0f3F800000; or.pred %p44, %p42, %p43; setp.eq.ftz.f32 %p45, %f476, 0f00000000; or.pred %p46, %p45, %p44; @%p46 bra $L__BB57_59; rcp.approx.ftz.f32 %f405, %f476; mul.ftz.f32 %f473, %f473, %f405; mul.ftz.f32 %f474, %f474, %f405; mul.ftz.f32 %f475, %f475, %f405; $L__BB57_59: and.b32 %r45, %r10, 16777216; setp.eq.s32 %p47, %r45, 0; @%p47 bra $L__BB57_69; setp.lt.ftz.f32 %p48, %f473, 0f3D25AEE6; @%p48 bra $L__BB57_62; bra.uni $L__BB57_61; $L__BB57_62: setp.lt.ftz.f32 %p49, %f473, 0f00000000; mul.ftz.f32 %f410, %f473, 0f3D9E8391; selp.f32 %f473, 0f00000000, %f410, %p49; bra.uni $L__BB57_63; $L__BB57_61: add.ftz.f32 %f406, %f473, 0f3D6147AE; mul.ftz.f32 %f407, %f406, 0f3F72A76F; lg2.approx.ftz.f32 %f408, %f407; mul.ftz.f32 %f409, %f408, 0f4019999A; ex2.approx.ftz.f32 %f473, %f409; $L__BB57_63: setp.lt.ftz.f32 %p50, %f474, 0f3D25AEE6; @%p50 bra $L__BB57_65; bra.uni $L__BB57_64; $L__BB57_65: setp.lt.ftz.f32 %p51, %f474, 0f00000000; mul.ftz.f32 %f415, %f474, 0f3D9E8391; selp.f32 %f474, 0f00000000, %f415, %p51; bra.uni $L__BB57_66; $L__BB57_64: add.ftz.f32 %f411, %f474, 0f3D6147AE; mul.ftz.f32 %f412, %f411, 0f3F72A76F; lg2.approx.ftz.f32 %f413, %f412; mul.ftz.f32 %f414, %f413, 0f4019999A; ex2.approx.ftz.f32 %f474, %f414; $L__BB57_66: setp.lt.ftz.f32 %p52, %f475, 0f3D25AEE6; @%p52 bra $L__BB57_68; bra.uni $L__BB57_67; $L__BB57_68: setp.lt.ftz.f32 %p53, %f475, 0f00000000; mul.ftz.f32 %f420, %f475, 0f3D9E8391; selp.f32 %f475, 0f00000000, %f420, %p53; bra.uni $L__BB57_69; $L__BB57_67: add.ftz.f32 %f416, %f475, 0f3D6147AE; mul.ftz.f32 %f417, %f416, 0f3F72A76F; lg2.approx.ftz.f32 %f418, %f417; mul.ftz.f32 %f419, %f418, 0f4019999A; ex2.approx.ftz.f32 %f475, %f419; $L__BB57_69: fma.rn.ftz.f32 %f489, %f449, %f473, %f489; fma.rn.ftz.f32 %f490, %f449, %f474, %f490; fma.rn.ftz.f32 %f491, %f449, %f475, %f491; fma.rn.ftz.f32 %f492, %f449, %f476, %f492; $L__BB57_70: setp.leu.ftz.f32 %p54, %f450, 0f00000000; @%p54 bra $L__BB57_94; setp.gt.ftz.f32 %p55, %f447, 0f00000000; mov.f32 %f494, 0f3F800000; sub.ftz.f32 %f425, %f494, %f39; selp.f32 %f155, %f425, %f39, %p55; setp.eq.s32 %p56, %r9, -1; mov.f32 %f495, 0f00000000; mov.f32 %f496, %f494; mov.f32 %f497, %f494; @%p56 bra $L__BB57_93; ld.const.u64 %rd37, [kernel_params+2432]; cvta.to.global.u64 %rd38, %rd37; mul.wide.s32 %rd39, %r9, 96; add.s64 %rd40, %rd38, %rd39; add.s64 %rd6, %rd40, 8; ld.global.v2.u32 {%r47, %r48}, [%rd40+8]; setp.gt.u32 %p57, %r47, 6; @%p57 bra $L__BB57_77; mov.u32 %r49, 1; shl.b32 %r50, %r49, %r47; and.b32 %r51, %r50, 71; setp.ne.s32 %p58, %r51, 0; @%p58 bra $L__BB57_74; bra.uni $L__BB57_77; $L__BB57_74: and.b32 %r52, %r48, -2; setp.eq.s32 %p59, %r52, 2; @%p59 bra $L__BB57_76; bra.uni $L__BB57_75; $L__BB57_76: ld.global.u64 %rd42, [%rd6+-8]; ld.global.u32 %r53, [%rd6+12]; ld.global.u32 %r54, [%rd6+16]; { // callseq 44, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd42; .param .b32 param1; st.param.b32 [param1+0], %r53; .param .b32 param2; st.param.b32 [param2+0], %r54; .param .b32 param3; st.param.f32 [param3+0], %f155; .param .b32 param4; st.param.f32 [param4+0], %f38; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f494, [retval0+0]; ld.param.f32 %f495, [retval0+4]; ld.param.f32 %f496, [retval0+8]; ld.param.f32 %f497, [retval0+12]; } // callseq 44 bra.uni $L__BB57_81; $L__BB57_75: ld.global.u64 %rd41, [%rd6+-8]; tex.2d.v4.f32.f32 {%f494, %f495, %f496, %f497}, [%rd41, {%f155, %f38}]; bra.uni $L__BB57_81; $L__BB57_77: and.b32 %r55, %r48, -2; setp.eq.s32 %p60, %r55, 2; @%p60 bra $L__BB57_79; bra.uni $L__BB57_78; $L__BB57_79: ld.global.u64 %rd44, [%rd6+-8]; ld.global.u32 %r56, [%rd6+12]; ld.global.u32 %r57, [%rd6+16]; { // callseq 45, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd44; .param .b32 param1; st.param.b32 [param1+0], %r56; .param .b32 param2; st.param.b32 [param2+0], %r57; .param .b32 param3; st.param.f32 [param3+0], %f155; .param .b32 param4; st.param.f32 [param4+0], %f38; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f494, [retval0+0]; } // callseq 45 bra.uni $L__BB57_80; $L__BB57_78: ld.global.u64 %rd43, [%rd6+-8]; tex.2d.v4.f32.f32 {%f494, %f426, %f427, %f428}, [%rd43, {%f155, %f38}]; $L__BB57_80: mov.f32 %f497, 0f3F800000; mov.f32 %f495, %f494; mov.f32 %f496, %f494; $L__BB57_81: and.b32 %r58, %r10, 33554432; setp.eq.s32 %p61, %r58, 0; setp.eq.ftz.f32 %p62, %f497, 0f3F800000; or.pred %p63, %p61, %p62; setp.eq.ftz.f32 %p64, %f497, 0f00000000; or.pred %p65, %p64, %p63; @%p65 bra $L__BB57_83; rcp.approx.ftz.f32 %f429, %f497; mul.ftz.f32 %f494, %f494, %f429; mul.ftz.f32 %f495, %f495, %f429; mul.ftz.f32 %f496, %f496, %f429; $L__BB57_83: and.b32 %r59, %r10, 16777216; setp.eq.s32 %p66, %r59, 0; @%p66 bra $L__BB57_93; setp.lt.ftz.f32 %p67, %f494, 0f3D25AEE6; @%p67 bra $L__BB57_86; bra.uni $L__BB57_85; $L__BB57_86: setp.lt.ftz.f32 %p68, %f494, 0f00000000; mul.ftz.f32 %f434, %f494, 0f3D9E8391; selp.f32 %f494, 0f00000000, %f434, %p68; bra.uni $L__BB57_87; $L__BB57_85: add.ftz.f32 %f430, %f494, 0f3D6147AE; mul.ftz.f32 %f431, %f430, 0f3F72A76F; lg2.approx.ftz.f32 %f432, %f431; mul.ftz.f32 %f433, %f432, 0f4019999A; ex2.approx.ftz.f32 %f494, %f433; $L__BB57_87: setp.lt.ftz.f32 %p69, %f495, 0f3D25AEE6; @%p69 bra $L__BB57_89; bra.uni $L__BB57_88; $L__BB57_89: setp.lt.ftz.f32 %p70, %f495, 0f00000000; mul.ftz.f32 %f439, %f495, 0f3D9E8391; selp.f32 %f495, 0f00000000, %f439, %p70; bra.uni $L__BB57_90; $L__BB57_88: add.ftz.f32 %f435, %f495, 0f3D6147AE; mul.ftz.f32 %f436, %f435, 0f3F72A76F; lg2.approx.ftz.f32 %f437, %f436; mul.ftz.f32 %f438, %f437, 0f4019999A; ex2.approx.ftz.f32 %f495, %f438; $L__BB57_90: setp.lt.ftz.f32 %p71, %f496, 0f3D25AEE6; @%p71 bra $L__BB57_92; bra.uni $L__BB57_91; $L__BB57_92: setp.lt.ftz.f32 %p72, %f496, 0f00000000; mul.ftz.f32 %f444, %f496, 0f3D9E8391; selp.f32 %f496, 0f00000000, %f444, %p72; bra.uni $L__BB57_93; $L__BB57_91: add.ftz.f32 %f440, %f496, 0f3D6147AE; mul.ftz.f32 %f441, %f440, 0f3F72A76F; lg2.approx.ftz.f32 %f442, %f441; mul.ftz.f32 %f443, %f442, 0f4019999A; ex2.approx.ftz.f32 %f496, %f443; $L__BB57_93: fma.rn.ftz.f32 %f489, %f450, %f494, %f489; fma.rn.ftz.f32 %f490, %f450, %f495, %f490; fma.rn.ftz.f32 %f491, %f450, %f496, %f491; fma.rn.ftz.f32 %f492, %f450, %f497, %f492; $L__BB57_94: setp.eq.s32 %p73, %r3, 255; @%p73 bra $L__BB57_96; mul.wide.u32 %rd45, %r3, 4; add.s64 %rd46, %rd1, %rd45; st.local.f32 [%rd46], %f489; st.local.f32 [%rd46+4], %f490; st.local.f32 [%rd46+8], %f491; $L__BB57_96: setp.eq.s32 %p74, %r4, 255; @%p74 bra $L__BB57_98; mul.wide.u32 %rd47, %r4, 4; add.s64 %rd48, %rd1, %rd47; st.local.f32 [%rd48], %f492; $L__BB57_98: ret; } .func _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<40>; .reg .f32 %f<210>; .reg .b32 %r<35>; .reg .b64 %rd<19>; ld.param.u64 %rd3, [_Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r11, [_Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; ld.param.u32 %r10, [_Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r9, [_Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd3; bfe.u32 %r2, %r10, 8, 8; bfe.u32 %r3, %r10, 16, 8; shl.b32 %r12, %r10, 2; cvt.u64.u32 %rd4, %r12; and.b64 %rd5, %rd4, 1020; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f183, [%rd6]; ld.local.f32 %f184, [%rd6+4]; mul.ftz.f32 %f86, %f184, %f184; fma.rn.ftz.f32 %f87, %f183, %f183, %f86; ld.local.f32 %f185, [%rd6+8]; fma.rn.ftz.f32 %f88, %f185, %f185, %f87; sqrt.approx.ftz.f32 %f4, %f88; setp.neu.ftz.f32 %p1, %f4, 0f00000000; @%p1 bra $L__BB58_1; bra.uni $L__BB58_2; $L__BB58_1: rcp.approx.ftz.f32 %f89, %f4; mul.ftz.f32 %f185, %f185, %f89; mul.ftz.f32 %f184, %f184, %f89; mul.ftz.f32 %f183, %f183, %f89; $L__BB58_2: setp.eq.s32 %p2, %r11, 0; @%p2 bra $L__BB58_6; add.ftz.f32 %f90, %f184, 0fBF800000; mul.ftz.f32 %f91, %f90, 0fBF000000; mov.f32 %f92, 0f00000000; max.ftz.f32 %f93, %f91, %f92; sqrt.approx.ftz.f32 %f94, %f93; add.ftz.f32 %f17, %f94, %f94; setp.leu.ftz.f32 %p3, %f17, 0f00000000; @%p3 bra $L__BB58_5; rcp.approx.ftz.f32 %f95, %f17; mul.ftz.f32 %f183, %f183, %f95; mul.ftz.f32 %f185, %f185, %f95; $L__BB58_5: add.ftz.f32 %f96, %f183, 0f3F800000; mul.ftz.f32 %f37, %f96, 0f3F000000; add.ftz.f32 %f97, %f185, 0f3F800000; mul.ftz.f32 %f36, %f97, 0f3F000000; bra.uni $L__BB58_13; $L__BB58_6: setp.eq.ftz.f32 %p4, %f183, 0f00000000; setp.eq.ftz.f32 %p5, %f184, 0f00000000; and.pred %p6, %p4, %p5; setp.eq.ftz.f32 %p7, %f185, 0f00000000; and.pred %p8, %p7, %p6; mov.f32 %f37, 0f00000000; mov.f32 %f36, %f37; @%p8 bra $L__BB58_13; abs.ftz.f32 %f24, %f183; setp.eq.ftz.f32 %p9, %f24, 0f00000000; abs.ftz.f32 %f25, %f184; setp.eq.ftz.f32 %p10, %f25, 0f00000000; and.pred %p11, %p9, %p10; mov.b32 %r4, %f183; mov.b32 %r13, %f184; and.b32 %r5, %r13, -2147483648; @%p11 bra $L__BB58_11; bra.uni $L__BB58_8; $L__BB58_11: shr.s32 %r18, %r4, 31; and.b32 %r19, %r18, 1078530011; or.b32 %r20, %r19, %r5; mov.b32 %f188, %r20; bra.uni $L__BB58_12; $L__BB58_8: setp.eq.ftz.f32 %p12, %f24, 0f7F800000; setp.eq.ftz.f32 %p13, %f25, 0f7F800000; and.pred %p14, %p12, %p13; @%p14 bra $L__BB58_10; bra.uni $L__BB58_9; $L__BB58_10: setp.lt.s32 %p18, %r4, 0; selp.b32 %r16, 1075235812, 1061752795, %p18; or.b32 %r17, %r16, %r5; mov.b32 %f188, %r17; bra.uni $L__BB58_12; $L__BB58_9: setp.lt.s32 %p15, %r4, 0; min.ftz.f32 %f100, %f25, %f24; max.ftz.f32 %f101, %f25, %f24; div.full.ftz.f32 %f102, %f100, %f101; mul.rn.ftz.f32 %f103, %f102, %f102; mov.f32 %f104, 0fC0B59883; mov.f32 %f105, 0fBF52C7EA; fma.rn.ftz.f32 %f106, %f103, %f105, %f104; mov.f32 %f107, 0fC0D21907; fma.rn.ftz.f32 %f108, %f106, %f103, %f107; mul.ftz.f32 %f109, %f103, %f108; mul.ftz.f32 %f110, %f102, %f109; add.ftz.f32 %f111, %f103, 0f41355DC0; mov.f32 %f112, 0f41E6BD60; fma.rn.ftz.f32 %f113, %f111, %f103, %f112; mov.f32 %f114, 0f419D92C8; fma.rn.ftz.f32 %f115, %f113, %f103, %f114; rcp.approx.ftz.f32 %f116, %f115; fma.rn.ftz.f32 %f117, %f110, %f116, %f102; mov.f32 %f118, 0f3FC90FDB; sub.ftz.f32 %f119, %f118, %f117; setp.gt.ftz.f32 %p16, %f25, %f24; selp.f32 %f120, %f119, %f117, %p16; mov.f32 %f121, 0f40490FDB; sub.ftz.f32 %f122, %f121, %f120; selp.f32 %f123, %f122, %f120, %p15; mov.b32 %r14, %f123; or.b32 %r15, %r5, %r14; mov.b32 %f124, %r15; add.ftz.f32 %f125, %f24, %f25; setp.le.ftz.f32 %p17, %f125, 0f7F800000; selp.f32 %f188, %f124, %f125, %p17; $L__BB58_12: add.ftz.f32 %f126, %f188, 0fC0490FDB; mov.f32 %f127, 0fC0490FDB; mov.f32 %f128, 0f40490FDB; mov.f32 %f129, 0fC0C90FDB; mul.ftz.f32 %f130, %f184, %f184; fma.rn.ftz.f32 %f131, %f183, %f183, %f130; fma.rn.ftz.f32 %f132, %f185, %f185, %f131; sqrt.approx.ftz.f32 %f133, %f132; div.approx.ftz.f32 %f134, %f185, %f133; abs.ftz.f32 %f135, %f134; mov.f32 %f136, 0f3F800000; sub.ftz.f32 %f137, %f136, %f135; mul.ftz.f32 %f138, %f137, 0f3F000000; sqrt.approx.ftz.f32 %f139, %f138; setp.gt.ftz.f32 %p19, %f135, 0f3F11EB85; selp.f32 %f140, %f139, %f135, %p19; mul.ftz.f32 %f141, %f140, %f140; mov.f32 %f142, 0f3C94D2E9; mov.f32 %f143, 0f3D53F941; fma.rn.ftz.f32 %f144, %f143, %f141, %f142; mov.f32 %f145, 0f3D3F841F; fma.rn.ftz.f32 %f146, %f144, %f141, %f145; mov.f32 %f147, 0f3D994929; fma.rn.ftz.f32 %f148, %f146, %f141, %f147; mov.f32 %f149, 0f3E2AAB94; fma.rn.ftz.f32 %f150, %f148, %f141, %f149; mul.ftz.f32 %f151, %f141, %f150; fma.rn.ftz.f32 %f152, %f151, %f140, %f140; add.ftz.f32 %f153, %f152, %f152; mov.f32 %f154, 0f3FC90FDB; sub.ftz.f32 %f155, %f154, %f152; selp.f32 %f156, %f153, %f155, %p19; setp.lt.ftz.f32 %p20, %f134, 0f00000000; sub.ftz.f32 %f157, %f128, %f156; selp.f32 %f158, %f157, %f156, %p20; add.ftz.f32 %f159, %f158, 0fC0490FDB; div.approx.ftz.f32 %f36, %f159, %f127; div.approx.ftz.f32 %f37, %f126, %f129; $L__BB58_13: setp.eq.s32 %p21, %r9, -1; mov.f32 %f194, 0f3F800000; mov.f32 %f195, 0f00000000; mov.f32 %f196, %f194; mov.f32 %f197, %f194; @%p21 bra $L__BB58_35; ld.const.u64 %rd7, [kernel_params+2432]; cvta.to.global.u64 %rd8, %rd7; mul.wide.s32 %rd9, %r9, 96; add.s64 %rd10, %rd8, %rd9; add.s64 %rd2, %rd10, 8; ld.global.v2.u32 {%r22, %r23}, [%rd10+8]; setp.gt.u32 %p22, %r22, 6; @%p22 bra $L__BB58_19; mov.u32 %r24, 1; shl.b32 %r25, %r24, %r22; and.b32 %r26, %r25, 71; setp.ne.s32 %p23, %r26, 0; @%p23 bra $L__BB58_16; bra.uni $L__BB58_19; $L__BB58_16: and.b32 %r27, %r23, -2; setp.eq.s32 %p24, %r27, 2; @%p24 bra $L__BB58_18; bra.uni $L__BB58_17; $L__BB58_18: ld.global.u64 %rd12, [%rd2+-8]; ld.global.u32 %r28, [%rd2+12]; ld.global.u32 %r29, [%rd2+16]; { // callseq 46, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd12; .param .b32 param1; st.param.b32 [param1+0], %r28; .param .b32 param2; st.param.b32 [param2+0], %r29; .param .b32 param3; st.param.f32 [param3+0], %f37; .param .b32 param4; st.param.f32 [param4+0], %f36; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f194, [retval0+0]; ld.param.f32 %f195, [retval0+4]; ld.param.f32 %f196, [retval0+8]; ld.param.f32 %f197, [retval0+12]; } // callseq 46 bra.uni $L__BB58_23; $L__BB58_17: ld.global.u64 %rd11, [%rd2+-8]; tex.2d.v4.f32.f32 {%f194, %f195, %f196, %f197}, [%rd11, {%f37, %f36}]; bra.uni $L__BB58_23; $L__BB58_19: and.b32 %r30, %r23, -2; setp.eq.s32 %p25, %r30, 2; @%p25 bra $L__BB58_21; bra.uni $L__BB58_20; $L__BB58_21: ld.global.u64 %rd14, [%rd2+-8]; ld.global.u32 %r31, [%rd2+12]; ld.global.u32 %r32, [%rd2+16]; { // callseq 47, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd14; .param .b32 param1; st.param.b32 [param1+0], %r31; .param .b32 param2; st.param.b32 [param2+0], %r32; .param .b32 param3; st.param.f32 [param3+0], %f37; .param .b32 param4; st.param.f32 [param4+0], %f36; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f194, [retval0+0]; } // callseq 47 bra.uni $L__BB58_22; $L__BB58_20: ld.global.u64 %rd13, [%rd2+-8]; tex.2d.v4.f32.f32 {%f194, %f164, %f165, %f166}, [%rd13, {%f37, %f36}]; $L__BB58_22: mov.f32 %f197, 0f3F800000; mov.f32 %f195, %f194; mov.f32 %f196, %f194; $L__BB58_23: and.b32 %r33, %r10, 33554432; setp.eq.s32 %p26, %r33, 0; setp.eq.ftz.f32 %p27, %f197, 0f3F800000; or.pred %p28, %p26, %p27; setp.eq.ftz.f32 %p29, %f197, 0f00000000; or.pred %p30, %p29, %p28; @%p30 bra $L__BB58_25; rcp.approx.ftz.f32 %f167, %f197; mul.ftz.f32 %f194, %f194, %f167; mul.ftz.f32 %f195, %f195, %f167; mul.ftz.f32 %f196, %f196, %f167; $L__BB58_25: and.b32 %r34, %r10, 16777216; setp.eq.s32 %p31, %r34, 0; @%p31 bra $L__BB58_35; setp.lt.ftz.f32 %p32, %f194, 0f3D25AEE6; @%p32 bra $L__BB58_28; bra.uni $L__BB58_27; $L__BB58_28: setp.lt.ftz.f32 %p33, %f194, 0f00000000; mul.ftz.f32 %f172, %f194, 0f3D9E8391; selp.f32 %f194, 0f00000000, %f172, %p33; bra.uni $L__BB58_29; $L__BB58_27: add.ftz.f32 %f168, %f194, 0f3D6147AE; mul.ftz.f32 %f169, %f168, 0f3F72A76F; lg2.approx.ftz.f32 %f170, %f169; mul.ftz.f32 %f171, %f170, 0f4019999A; ex2.approx.ftz.f32 %f194, %f171; $L__BB58_29: setp.lt.ftz.f32 %p34, %f195, 0f3D25AEE6; @%p34 bra $L__BB58_31; bra.uni $L__BB58_30; $L__BB58_31: setp.lt.ftz.f32 %p35, %f195, 0f00000000; mul.ftz.f32 %f177, %f195, 0f3D9E8391; selp.f32 %f195, 0f00000000, %f177, %p35; bra.uni $L__BB58_32; $L__BB58_30: add.ftz.f32 %f173, %f195, 0f3D6147AE; mul.ftz.f32 %f174, %f173, 0f3F72A76F; lg2.approx.ftz.f32 %f175, %f174; mul.ftz.f32 %f176, %f175, 0f4019999A; ex2.approx.ftz.f32 %f195, %f176; $L__BB58_32: setp.lt.ftz.f32 %p36, %f196, 0f3D25AEE6; @%p36 bra $L__BB58_34; bra.uni $L__BB58_33; $L__BB58_34: setp.lt.ftz.f32 %p37, %f196, 0f00000000; mul.ftz.f32 %f182, %f196, 0f3D9E8391; selp.f32 %f196, 0f00000000, %f182, %p37; bra.uni $L__BB58_35; $L__BB58_33: add.ftz.f32 %f178, %f196, 0f3D6147AE; mul.ftz.f32 %f179, %f178, 0f3F72A76F; lg2.approx.ftz.f32 %f180, %f179; mul.ftz.f32 %f181, %f180, 0f4019999A; ex2.approx.ftz.f32 %f196, %f181; $L__BB58_35: setp.eq.s32 %p38, %r2, 255; @%p38 bra $L__BB58_37; mul.wide.u32 %rd15, %r2, 4; add.s64 %rd16, %rd1, %rd15; st.local.f32 [%rd16], %f194; st.local.f32 [%rd16+4], %f195; st.local.f32 [%rd16+8], %f196; $L__BB58_37: setp.eq.s32 %p39, %r3, 255; @%p39 bra $L__BB58_39; mul.wide.u32 %rd17, %r3, 4; add.s64 %rd18, %rd1, %rd17; st.local.f32 [%rd18], %f197; $L__BB58_39: ret; } .func _Z15svm_node_invertP10ShaderDataPfjjj( .param .b64 _Z15svm_node_invertP10ShaderDataPfjjj_param_0, .param .b32 _Z15svm_node_invertP10ShaderDataPfjjj_param_1, .param .b32 _Z15svm_node_invertP10ShaderDataPfjjj_param_2, .param .b32 _Z15svm_node_invertP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<2>; .reg .f32 %f<16>; .reg .b32 %r<4>; .reg .b64 %rd<9>; ld.param.u64 %rd2, [_Z15svm_node_invertP10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z15svm_node_invertP10ShaderDataPfjjj_param_1]; ld.param.u32 %r3, [_Z15svm_node_invertP10ShaderDataPfjjj_param_2]; ld.param.u32 %r1, [_Z15svm_node_invertP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; mul.wide.u32 %rd3, %r2, 4; add.s64 %rd4, %rd1, %rd3; mul.wide.u32 %rd5, %r3, 4; add.s64 %rd6, %rd1, %rd5; ld.local.f32 %f4, [%rd6]; mov.f32 %f5, 0f3F800000; sub.ftz.f32 %f6, %f5, %f4; ld.local.f32 %f7, [%rd4]; mul.ftz.f32 %f8, %f7, %f6; sub.ftz.f32 %f9, %f5, %f7; fma.rn.ftz.f32 %f1, %f9, %f4, %f8; ld.local.f32 %f10, [%rd6+4]; sub.ftz.f32 %f11, %f5, %f10; mul.ftz.f32 %f12, %f7, %f11; fma.rn.ftz.f32 %f2, %f9, %f10, %f12; ld.local.f32 %f13, [%rd6+8]; sub.ftz.f32 %f14, %f5, %f13; mul.ftz.f32 %f15, %f7, %f14; fma.rn.ftz.f32 %f3, %f9, %f13, %f15; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB59_2; mul.wide.u32 %rd7, %r1, 4; add.s64 %rd8, %rd1, %rd7; st.local.f32 [%rd8], %f1; st.local.f32 [%rd8+4], %f2; st.local.f32 [%rd8+8], %f3; $L__BB59_2: ret; } .func _Z22svm_node_light_falloffP10ShaderDataPf5uint4( .param .b64 _Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_0, .param .b64 _Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_1, .param .align 16 .b8 _Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<5>; .reg .f32 %f<18>; .reg .b32 %r<6>; .reg .b64 %rd<15>; ld.param.u64 %rd3, [_Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd4, [_Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r2, [_Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_2+4]; ld.param.u32 %r1, [_Z22svm_node_light_falloffP10ShaderDataPf5uint4_param_2+8]; cvta.to.local.u64 %rd5, %rd3; cvta.to.local.u64 %rd1, %rd4; shl.b32 %r3, %r1, 2; cvt.u64.u32 %rd6, %r3; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f17, [%rd8]; add.s64 %rd2, %rd5, 88; setp.eq.s32 %p1, %r2, 2; @%p1 bra $L__BB60_3; setp.ne.s32 %p2, %r2, 1; @%p2 bra $L__BB60_4; ld.local.f32 %f9, [%rd2]; mul.ftz.f32 %f17, %f17, %f9; bra.uni $L__BB60_4; $L__BB60_3: ld.local.f32 %f10, [%rd2]; mul.ftz.f32 %f11, %f10, %f10; mul.ftz.f32 %f17, %f17, %f11; $L__BB60_4: shr.u32 %r4, %r1, 6; cvt.u64.u32 %rd9, %r4; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f5, [%rd11]; setp.leu.ftz.f32 %p3, %f5, 0f00000000; @%p3 bra $L__BB60_7; ld.local.f32 %f12, [%rd2]; mul.ftz.f32 %f6, %f12, %f12; abs.ftz.f32 %f13, %f6; setp.geu.ftz.f32 %p4, %f13, 0f7F800000; @%p4 bra $L__BB60_7; add.ftz.f32 %f14, %f5, %f6; div.approx.ftz.f32 %f15, %f6, %f14; mul.ftz.f32 %f17, %f17, %f15; $L__BB60_7: shr.u32 %r5, %r1, 14; cvt.u64.u32 %rd12, %r5; and.b64 %rd13, %rd12, 1020; add.s64 %rd14, %rd1, %rd13; st.local.f32 [%rd14], %f17; ret; } .func (.param .b32 func_retval0) _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<55>; .reg .f32 %f<173>; .reg .b32 %r<111>; .reg .b64 %rd<18>; ld.param.u64 %rd2, [_Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r54, [_Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r55, [_Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r56, [_Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd2; and.b32 %r1, %r56, 255; bfe.u32 %r2, %r56, 8, 8; bfe.u32 %r3, %r56, 16, 8; bfe.u32 %r4, %r55, 8, 8; bfe.u32 %r5, %r55, 16, 8; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r54, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r57, %r58, %r59, %r60}, [%rd6]; shl.b32 %r61, %r55, 2; cvt.u64.u32 %rd7, %r61; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r4, 255; @%p1 bra $L__BB61_2; bra.uni $L__BB61_1; $L__BB61_2: mov.b32 %f162, %r57; bra.uni $L__BB61_3; $L__BB61_1: mul.wide.u32 %rd10, %r4, 4; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f162, [%rd11]; $L__BB61_3: setp.eq.s32 %p2, %r5, 255; @%p2 bra $L__BB61_5; bra.uni $L__BB61_4; $L__BB61_5: mov.b32 %f163, %r58; bra.uni $L__BB61_6; $L__BB61_4: mul.wide.u32 %rd12, %r5, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f163, [%rd13]; $L__BB61_6: mov.f32 %f69, 0f40C90FDB; abs.ftz.f32 %f10, %f69; mul.ftz.f32 %f11, %f10, 0f4B000000; abs.ftz.f32 %f12, %f1; setp.lt.ftz.f32 %p3, %f12, %f10; mov.b32 %r102, %f12; @%p3 bra $L__BB61_18; setp.le.ftz.f32 %p4, %f12, %f11; setp.leu.ftz.f32 %p5, %f10, 0f00000000; or.pred %p6, %p5, %p4; setp.gt.u32 %p7, %r102, 2139095039; or.pred %p8, %p7, %p6; @%p8 bra $L__BB61_12; bra.uni $L__BB61_8; $L__BB61_12: div.approx.ftz.f32 %f74, %f12, %f10; cvt.rzi.ftz.f32.f32 %f164, %f74; neg.ftz.f32 %f14, %f10; fma.rn.f32 %f15, %f14, %f164, %f12; mov.b32 %r21, %f15; mov.b32 %r70, %f10; setp.lt.u32 %p11, %r21, %r70; @%p11 bra $L__BB61_17; setp.gt.u32 %p12, %r21, -2147483648; @%p12 bra $L__BB61_16; bra.uni $L__BB61_14; $L__BB61_16: add.ftz.f32 %f80, %f164, 0fBF800000; add.ftz.f32 %f81, %f80, 0fBF800000; setp.lt.ftz.f32 %p15, %f15, %f14; selp.f32 %f164, %f81, %f80, %p15; bra.uni $L__BB61_17; $L__BB61_8: shr.u32 %r62, %r102, 23; mov.b32 %r11, %f11; shr.u32 %r63, %r11, 23; and.b32 %r64, %r102, 8388607; or.b32 %r101, %r64, 8388608; add.s32 %r65, %r62, 23; sub.s32 %r99, %r65, %r63; setp.eq.s32 %p9, %r99, 0; @%p9 bra $L__BB61_11; and.b32 %r66, %r11, 8388607; or.b32 %r14, %r66, 8388608; $L__BB61_10: min.s32 %r67, %r99, 8; shl.b32 %r68, %r101, %r67; rem.u32 %r101, %r68, %r14; sub.s32 %r99, %r99, %r67; setp.ne.s32 %p10, %r99, 0; @%p10 bra $L__BB61_10; $L__BB61_11: and.b32 %r69, %r11, -8388608; mov.b32 %f70, %r69; cvt.rn.f32.u32 %f71, %r101; mul.ftz.f32 %f72, %f71, 0f28800000; mul.ftz.f32 %f73, %f72, %f70; mov.b32 %r102, %f73; bra.uni $L__BB61_18; $L__BB61_14: add.ftz.f32 %f164, %f164, 0f3F800000; add.ftz.f32 %f75, %f10, %f10; setp.ltu.ftz.f32 %p13, %f15, %f75; @%p13 bra $L__BB61_17; add.ftz.f32 %f76, %f164, 0f3F800000; mov.f32 %f77, 0fC0400000; fma.rn.ftz.f32 %f78, %f77, %f10, %f15; setp.ge.ftz.f32 %p14, %f78, 0f00000000; add.ftz.f32 %f79, %f76, 0f3F800000; selp.f32 %f164, %f79, %f76, %p14; $L__BB61_17: fma.rn.ftz.f32 %f82, %f14, %f164, %f12; mov.b32 %r102, %f82; $L__BB61_18: mov.b32 %r71, %f1; and.b32 %r72, %r71, -2147483648; or.b32 %r24, %r102, %r72; abs.ftz.f32 %f20, %f2; setp.lt.ftz.f32 %p16, %f20, %f10; mov.b32 %r106, %f20; @%p16 bra $L__BB61_30; setp.le.ftz.f32 %p17, %f20, %f11; setp.leu.ftz.f32 %p18, %f10, 0f00000000; or.pred %p19, %p18, %p17; setp.gt.u32 %p20, %r106, 2139095039; or.pred %p21, %p20, %p19; @%p21 bra $L__BB61_24; bra.uni $L__BB61_20; $L__BB61_24: div.approx.ftz.f32 %f87, %f20, %f10; cvt.rzi.ftz.f32.f32 %f165, %f87; neg.ftz.f32 %f22, %f10; fma.rn.f32 %f23, %f22, %f165, %f20; mov.b32 %r36, %f23; mov.b32 %r82, %f10; setp.lt.u32 %p24, %r36, %r82; @%p24 bra $L__BB61_29; setp.gt.u32 %p25, %r36, -2147483648; @%p25 bra $L__BB61_28; bra.uni $L__BB61_26; $L__BB61_28: add.ftz.f32 %f93, %f165, 0fBF800000; add.ftz.f32 %f94, %f93, 0fBF800000; setp.lt.ftz.f32 %p28, %f23, %f22; selp.f32 %f165, %f94, %f93, %p28; bra.uni $L__BB61_29; $L__BB61_20: shr.u32 %r73, %r106, 23; mov.u32 %r74, 23; mov.b32 %r26, %f11; shr.u32 %r75, %r26, 23; and.b32 %r76, %r106, 8388607; or.b32 %r105, %r76, 8388608; sub.s32 %r77, %r74, %r75; add.s32 %r103, %r77, %r73; setp.eq.s32 %p22, %r103, 0; @%p22 bra $L__BB61_23; and.b32 %r78, %r26, 8388607; or.b32 %r29, %r78, 8388608; $L__BB61_22: min.s32 %r79, %r103, 8; shl.b32 %r80, %r105, %r79; rem.u32 %r105, %r80, %r29; sub.s32 %r103, %r103, %r79; setp.ne.s32 %p23, %r103, 0; @%p23 bra $L__BB61_22; $L__BB61_23: and.b32 %r81, %r26, -8388608; mov.b32 %f83, %r81; cvt.rn.f32.u32 %f84, %r105; mul.ftz.f32 %f85, %f84, 0f28800000; mul.ftz.f32 %f86, %f85, %f83; mov.b32 %r106, %f86; bra.uni $L__BB61_30; $L__BB61_26: add.ftz.f32 %f165, %f165, 0f3F800000; add.ftz.f32 %f88, %f10, %f10; setp.ltu.ftz.f32 %p26, %f23, %f88; @%p26 bra $L__BB61_29; add.ftz.f32 %f89, %f165, 0f3F800000; mov.f32 %f90, 0fC0400000; fma.rn.ftz.f32 %f91, %f90, %f10, %f23; setp.ge.ftz.f32 %p27, %f91, 0f00000000; add.ftz.f32 %f92, %f89, 0f3F800000; selp.f32 %f165, %f92, %f89, %p27; $L__BB61_29: fma.rn.ftz.f32 %f95, %f22, %f165, %f20; mov.b32 %r106, %f95; $L__BB61_30: mov.b32 %f28, %r24; mov.b32 %r83, %f2; and.b32 %r84, %r83, -2147483648; or.b32 %r39, %r106, %r84; abs.ftz.f32 %f29, %f3; setp.lt.ftz.f32 %p29, %f29, %f10; mov.b32 %r110, %f29; @%p29 bra $L__BB61_42; setp.le.ftz.f32 %p30, %f29, %f11; setp.leu.ftz.f32 %p31, %f10, 0f00000000; or.pred %p32, %p31, %p30; setp.gt.u32 %p33, %r110, 2139095039; or.pred %p34, %p33, %p32; @%p34 bra $L__BB61_36; bra.uni $L__BB61_32; $L__BB61_36: div.approx.ftz.f32 %f100, %f29, %f10; cvt.rzi.ftz.f32.f32 %f166, %f100; neg.ftz.f32 %f31, %f10; fma.rn.f32 %f32, %f31, %f166, %f29; mov.b32 %r51, %f32; mov.b32 %r94, %f10; setp.lt.u32 %p37, %r51, %r94; @%p37 bra $L__BB61_41; setp.gt.u32 %p38, %r51, -2147483648; @%p38 bra $L__BB61_40; bra.uni $L__BB61_38; $L__BB61_40: add.ftz.f32 %f106, %f166, 0fBF800000; add.ftz.f32 %f107, %f106, 0fBF800000; setp.lt.ftz.f32 %p41, %f32, %f31; selp.f32 %f166, %f107, %f106, %p41; bra.uni $L__BB61_41; $L__BB61_32: shr.u32 %r85, %r110, 23; mov.u32 %r86, 23; mov.b32 %r41, %f11; shr.u32 %r87, %r41, 23; and.b32 %r88, %r110, 8388607; or.b32 %r109, %r88, 8388608; sub.s32 %r89, %r86, %r87; add.s32 %r107, %r89, %r85; setp.eq.s32 %p35, %r107, 0; @%p35 bra $L__BB61_35; and.b32 %r90, %r41, 8388607; or.b32 %r44, %r90, 8388608; $L__BB61_34: min.s32 %r91, %r107, 8; shl.b32 %r92, %r109, %r91; rem.u32 %r109, %r92, %r44; sub.s32 %r107, %r107, %r91; setp.ne.s32 %p36, %r107, 0; @%p36 bra $L__BB61_34; $L__BB61_35: and.b32 %r93, %r41, -8388608; mov.b32 %f96, %r93; cvt.rn.f32.u32 %f97, %r109; mul.ftz.f32 %f98, %f97, 0f28800000; mul.ftz.f32 %f99, %f98, %f96; mov.b32 %r110, %f99; bra.uni $L__BB61_42; $L__BB61_38: add.ftz.f32 %f166, %f166, 0f3F800000; add.ftz.f32 %f101, %f10, %f10; setp.ltu.ftz.f32 %p39, %f32, %f101; @%p39 bra $L__BB61_41; add.ftz.f32 %f102, %f166, 0f3F800000; mov.f32 %f103, 0fC0400000; fma.rn.ftz.f32 %f104, %f103, %f10, %f32; setp.ge.ftz.f32 %p40, %f104, 0f00000000; add.ftz.f32 %f105, %f102, 0f3F800000; selp.f32 %f166, %f105, %f102, %p40; $L__BB61_41: fma.rn.ftz.f32 %f108, %f31, %f166, %f29; mov.b32 %r110, %f108; $L__BB61_42: mov.b32 %f109, %r39; mov.b32 %r95, %f3; and.b32 %r96, %r95, -2147483648; or.b32 %r97, %r110, %r96; mov.b32 %f110, %r97; add.ftz.f32 %f111, %f28, %f109; add.ftz.f32 %f112, %f111, %f110; mul.ftz.f32 %f113, %f112, 0f40A00000; mul.ftz.f32 %f114, %f162, %f113; sin.approx.ftz.f32 %f167, %f114; sub.ftz.f32 %f115, %f109, %f28; sub.ftz.f32 %f116, %f115, %f110; mul.ftz.f32 %f117, %f116, 0f40A00000; mul.ftz.f32 %f118, %f162, %f117; cos.approx.ftz.f32 %f168, %f118; neg.ftz.f32 %f119, %f28; sub.ftz.f32 %f120, %f119, %f109; add.ftz.f32 %f121, %f120, %f110; mul.ftz.f32 %f122, %f121, 0f40A00000; mul.ftz.f32 %f123, %f162, %f122; cos.approx.ftz.f32 %f124, %f123; neg.ftz.f32 %f169, %f124; setp.eq.s32 %p42, %r1, 0; @%p42 bra $L__BB61_53; mul.ftz.f32 %f167, %f163, %f167; mul.ftz.f32 %f125, %f163, %f168; sub.ftz.f32 %f126, %f167, %f125; mul.ftz.f32 %f169, %f163, %f169; add.ftz.f32 %f127, %f126, %f169; cos.approx.ftz.f32 %f128, %f127; mul.ftz.f32 %f42, %f163, %f128; neg.ftz.f32 %f168, %f42; setp.eq.s32 %p43, %r1, 1; @%p43 bra $L__BB61_53; add.ftz.f32 %f129, %f167, %f42; sub.ftz.f32 %f130, %f129, %f169; cos.approx.ftz.f32 %f131, %f130; mul.ftz.f32 %f167, %f163, %f131; setp.lt.u32 %p44, %r1, 3; @%p44 bra $L__BB61_53; sub.ftz.f32 %f132, %f42, %f167; sub.ftz.f32 %f133, %f132, %f169; sin.approx.ftz.f32 %f134, %f133; mul.ftz.f32 %f169, %f163, %f134; setp.eq.s32 %p45, %r1, 3; @%p45 bra $L__BB61_53; sub.ftz.f32 %f135, %f168, %f167; sub.ftz.f32 %f136, %f135, %f169; cos.approx.ftz.f32 %f137, %f136; mul.ftz.f32 %f46, %f163, %f137; neg.ftz.f32 %f167, %f46; setp.lt.u32 %p46, %r1, 5; @%p46 bra $L__BB61_53; sub.ftz.f32 %f138, %f46, %f42; add.ftz.f32 %f139, %f169, %f138; sin.approx.ftz.f32 %f140, %f139; mul.ftz.f32 %f48, %f163, %f140; neg.ftz.f32 %f168, %f48; setp.eq.s32 %p47, %r1, 5; @%p47 bra $L__BB61_53; sub.ftz.f32 %f141, %f46, %f48; add.ftz.f32 %f142, %f169, %f141; cos.approx.ftz.f32 %f143, %f142; mul.ftz.f32 %f50, %f163, %f143; neg.ftz.f32 %f168, %f50; setp.lt.u32 %p48, %r1, 7; @%p48 bra $L__BB61_53; sub.ftz.f32 %f144, %f168, %f46; add.ftz.f32 %f145, %f169, %f144; cos.approx.ftz.f32 %f146, %f145; mul.ftz.f32 %f167, %f163, %f146; setp.eq.s32 %p49, %r1, 7; @%p49 bra $L__BB61_53; sub.ftz.f32 %f147, %f167, %f50; sub.ftz.f32 %f148, %f147, %f169; sin.approx.ftz.f32 %f149, %f148; mul.ftz.f32 %f169, %f163, %f149; setp.lt.u32 %p50, %r1, 9; @%p50 bra $L__BB61_53; sub.ftz.f32 %f150, %f50, %f167; add.ftz.f32 %f151, %f150, %f169; cos.approx.ftz.f32 %f152, %f151; mul.ftz.f32 %f54, %f163, %f152; neg.ftz.f32 %f167, %f54; setp.eq.s32 %p51, %r1, 9; @%p51 bra $L__BB61_53; sub.ftz.f32 %f153, %f50, %f54; add.ftz.f32 %f154, %f169, %f153; sin.approx.ftz.f32 %f155, %f154; mul.ftz.f32 %f156, %f163, %f155; neg.ftz.f32 %f168, %f156; $L__BB61_53: setp.eq.ftz.f32 %p52, %f163, 0f00000000; @%p52 bra $L__BB61_55; add.ftz.f32 %f157, %f163, %f163; div.approx.ftz.f32 %f167, %f167, %f157; div.approx.ftz.f32 %f168, %f168, %f157; div.approx.ftz.f32 %f169, %f169, %f157; $L__BB61_55: mov.f32 %f158, 0f3F000000; sub.ftz.f32 %f66, %f158, %f167; sub.ftz.f32 %f67, %f158, %f168; sub.ftz.f32 %f68, %f158, %f169; setp.eq.s32 %p53, %r3, 255; @%p53 bra $L__BB61_57; add.ftz.f32 %f159, %f66, %f67; add.ftz.f32 %f160, %f68, %f159; mul.ftz.f32 %f161, %f160, 0f3EAAAAAB; mul.wide.u32 %rd14, %r3, 4; add.s64 %rd15, %rd1, %rd14; st.local.f32 [%rd15], %f161; $L__BB61_57: setp.eq.s32 %p54, %r2, 255; @%p54 bra $L__BB61_59; mul.wide.u32 %rd16, %r2, 4; add.s64 %rd17, %rd1, %rd16; st.local.f32 [%rd17], %f66; st.local.f32 [%rd17+4], %f67; st.local.f32 [%rd17+8], %f68; $L__BB61_59: add.s32 %r98, %r54, 1; st.param.b32 [func_retval0+0], %r98; ret; } .func (.param .b32 func_retval0) _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<19>; .reg .b16 %rs<3>; .reg .f32 %f<95>; .reg .b32 %r<23>; .reg .b64 %rd<24>; ld.param.u64 %rd2, [_Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r10, [_Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r7, [_Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r8, [_Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r9, [_Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; bfe.u32 %r1, %r8, 8, 8; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r9, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r11, %r12, %r13, %r14}, [%rd6]; ld.global.f32 %f89, [%rd6+16]; mul.wide.u32 %rd7, %r10, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f2, [%rd8]; and.b32 %r6, %r7, 255; setp.eq.s32 %p1, %r6, 255; @%p1 bra $L__BB62_2; bra.uni $L__BB62_1; $L__BB62_2: mov.b32 %f5, %r11; bra.uni $L__BB62_3; $L__BB62_1: mul.wide.u32 %rd9, %r6, 4; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f5, [%rd10]; $L__BB62_3: bfe.u32 %r15, %r7, 8, 8; setp.eq.s32 %p2, %r15, 255; @%p2 bra $L__BB62_5; bra.uni $L__BB62_4; $L__BB62_5: mov.b32 %f8, %r12; bra.uni $L__BB62_6; $L__BB62_4: shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd11, %r16; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f8, [%rd13]; $L__BB62_6: bfe.u32 %r17, %r7, 16, 8; setp.eq.s32 %p3, %r17, 255; @%p3 bra $L__BB62_8; bra.uni $L__BB62_7; $L__BB62_8: mov.b32 %f87, %r13; bra.uni $L__BB62_9; $L__BB62_7: shr.u32 %r18, %r7, 14; cvt.u64.u32 %rd14, %r18; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f87, [%rd16]; $L__BB62_9: shr.u32 %r19, %r7, 24; setp.eq.s32 %p4, %r19, 255; @%p4 bra $L__BB62_11; bra.uni $L__BB62_10; $L__BB62_11: mov.b32 %f88, %r14; bra.uni $L__BB62_12; $L__BB62_10: mul.wide.u32 %rd17, %r19, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f88, [%rd18]; $L__BB62_12: setp.eq.s32 %p5, %r1, 255; @%p5 bra $L__BB62_14; mul.wide.u32 %rd19, %r1, 4; add.s64 %rd20, %rd1, %rd19; ld.local.f32 %f89, [%rd20]; $L__BB62_14: setp.eq.ftz.f32 %p6, %f8, %f5; mov.f32 %f94, 0f00000000; @%p6 bra $L__BB62_37; cvt.u16.u32 %rs2, %r8; and.b16 %rs1, %rs2, 255; setp.eq.s16 %p7, %rs1, 1; @%p7 bra $L__BB62_26; setp.eq.s16 %p8, %rs1, 3; @%p8 bra $L__BB62_28; setp.ne.s16 %p9, %rs1, 2; @%p9 bra $L__BB62_35; setp.gt.ftz.f32 %p10, %f5, %f8; @%p10 bra $L__BB62_22; bra.uni $L__BB62_19; $L__BB62_22: setp.lt.ftz.f32 %p13, %f2, %f8; mov.f32 %f90, 0f00000000; @%p13 bra $L__BB62_25; setp.ge.ftz.f32 %p14, %f2, %f5; mov.f32 %f90, 0f3F800000; @%p14 bra $L__BB62_25; sub.ftz.f32 %f49, %f2, %f8; sub.ftz.f32 %f50, %f5, %f8; div.approx.ftz.f32 %f51, %f49, %f50; add.ftz.f32 %f52, %f51, %f51; mov.f32 %f53, 0f40400000; sub.ftz.f32 %f54, %f53, %f52; mul.ftz.f32 %f55, %f51, %f51; mul.ftz.f32 %f90, %f55, %f54; $L__BB62_25: mov.f32 %f56, 0f3F800000; sub.ftz.f32 %f93, %f56, %f90; bra.uni $L__BB62_36; $L__BB62_26: sub.ftz.f32 %f58, %f2, %f5; sub.ftz.f32 %f59, %f8, %f5; div.approx.ftz.f32 %f21, %f58, %f59; setp.leu.ftz.f32 %p15, %f89, 0f00000000; mov.f32 %f93, 0f00000000; @%p15 bra $L__BB62_36; add.ftz.f32 %f60, %f89, 0f3F800000; mul.ftz.f32 %f61, %f60, %f21; cvt.rmi.ftz.f32.f32 %f62, %f61; div.approx.ftz.f32 %f93, %f62, %f89; bra.uni $L__BB62_36; $L__BB62_28: setp.gt.ftz.f32 %p16, %f5, %f8; @%p16 bra $L__BB62_32; bra.uni $L__BB62_29; $L__BB62_32: sub.ftz.f32 %f28, %f2, %f8; sub.ftz.f32 %f29, %f5, %f8; setp.eq.ftz.f32 %p18, %f29, 0f00000000; mov.f32 %f72, 0f00000000; mov.f32 %f92, %f72; @%p18 bra $L__BB62_34; div.approx.ftz.f32 %f92, %f28, %f29; $L__BB62_34: max.ftz.f32 %f74, %f92, %f72; mov.f32 %f75, 0f3F800000; min.ftz.f32 %f76, %f74, %f75; mul.ftz.f32 %f77, %f76, %f76; mul.ftz.f32 %f78, %f76, %f77; fma.rn.ftz.f32 %f79, %f76, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f80, %f76, %f79, 0f41200000; mul.ftz.f32 %f81, %f78, %f80; sub.ftz.f32 %f93, %f75, %f81; bra.uni $L__BB62_36; $L__BB62_35: sub.ftz.f32 %f82, %f2, %f5; sub.ftz.f32 %f83, %f8, %f5; div.approx.ftz.f32 %f93, %f82, %f83; bra.uni $L__BB62_36; $L__BB62_29: sub.ftz.f32 %f23, %f2, %f5; sub.ftz.f32 %f24, %f8, %f5; setp.eq.ftz.f32 %p17, %f24, 0f00000000; mov.f32 %f63, 0f00000000; mov.f32 %f91, %f63; @%p17 bra $L__BB62_31; div.approx.ftz.f32 %f91, %f23, %f24; $L__BB62_31: max.ftz.f32 %f65, %f91, %f63; mov.f32 %f66, 0f3F800000; min.ftz.f32 %f67, %f65, %f66; mul.ftz.f32 %f68, %f67, %f67; mul.ftz.f32 %f69, %f67, %f68; fma.rn.ftz.f32 %f70, %f67, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f71, %f67, %f70, 0f41200000; mul.ftz.f32 %f93, %f69, %f71; bra.uni $L__BB62_36; $L__BB62_19: setp.lt.ftz.f32 %p11, %f2, %f5; mov.f32 %f93, 0f00000000; @%p11 bra $L__BB62_36; setp.ge.ftz.f32 %p12, %f2, %f8; mov.f32 %f93, 0f3F800000; @%p12 bra $L__BB62_36; sub.ftz.f32 %f40, %f2, %f5; sub.ftz.f32 %f41, %f8, %f5; div.approx.ftz.f32 %f42, %f40, %f41; add.ftz.f32 %f43, %f42, %f42; mov.f32 %f44, 0f40400000; sub.ftz.f32 %f45, %f44, %f43; mul.ftz.f32 %f46, %f42, %f42; mul.ftz.f32 %f93, %f46, %f45; $L__BB62_36: sub.ftz.f32 %f84, %f88, %f87; fma.rn.ftz.f32 %f94, %f84, %f93, %f87; $L__BB62_37: shr.u32 %r21, %r8, 14; cvt.u64.u32 %rd21, %r21; and.b64 %rd22, %rd21, 1020; add.s64 %rd23, %rd1, %rd22; st.local.f32 [%rd23], %f94; add.s32 %r22, %r9, 2; st.param.b32 [func_retval0+0], %r22; ret; } .func (.param .b32 func_retval0) _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<16>; .reg .b16 %rs<3>; .reg .f32 %f<123>; .reg .b32 %r<15>; .reg .b64 %rd<21>; ld.param.u64 %rd3, [_Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r4, [_Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r5, [_Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r2, [_Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r3, [_Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd3; mul.wide.u32 %rd4, %r4, 4; add.s64 %rd2, %rd1, %rd4; shl.b32 %r6, %r5, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; shr.u32 %r7, %r5, 6; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; shr.u32 %r8, %r5, 14; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f1, [%rd13]; ld.local.f32 %f2, [%rd13+4]; ld.local.f32 %f3, [%rd13+8]; shr.u32 %r9, %r5, 24; mul.wide.u32 %rd14, %r9, 4; add.s64 %rd15, %rd1, %rd14; ld.local.f32 %f4, [%rd15]; ld.local.f32 %f5, [%rd15+4]; ld.local.f32 %f6, [%rd15+8]; shl.b32 %r10, %r2, 2; cvt.u64.u32 %rd16, %r10; and.b64 %rd17, %rd16, 1020; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f7, [%rd18]; ld.local.f32 %f8, [%rd18+4]; ld.local.f32 %f9, [%rd18+8]; ld.local.f32 %f10, [%rd7]; ld.local.f32 %f11, [%rd7+4]; ld.local.f32 %f12, [%rd2+4]; ld.local.f32 %f13, [%rd7+8]; ld.local.f32 %f14, [%rd2+8]; ld.local.f32 %f61, [%rd10]; sub.ftz.f32 %f15, %f61, %f10; ld.local.f32 %f62, [%rd10+4]; sub.ftz.f32 %f16, %f62, %f11; ld.local.f32 %f63, [%rd10+8]; sub.ftz.f32 %f17, %f63, %f13; setp.eq.ftz.f32 %p1, %f15, 0f00000000; mov.f32 %f60, 0f00000000; mov.f32 %f110, %f60; @%p1 bra $L__BB63_2; ld.local.f32 %f64, [%rd2]; sub.ftz.f32 %f65, %f64, %f10; div.approx.ftz.f32 %f110, %f65, %f15; $L__BB63_2: setp.eq.ftz.f32 %p2, %f16, 0f00000000; sub.ftz.f32 %f20, %f14, %f13; sub.ftz.f32 %f21, %f12, %f11; mov.f32 %f111, %f60; @%p2 bra $L__BB63_4; div.approx.ftz.f32 %f111, %f21, %f16; $L__BB63_4: setp.eq.ftz.f32 %p3, %f17, 0f00000000; mov.f32 %f112, %f60; @%p3 bra $L__BB63_6; div.approx.ftz.f32 %f112, %f20, %f17; $L__BB63_6: shr.u32 %r1, %r2, 16; max.ftz.f32 %f69, %f110, %f60; mov.f32 %f70, 0f3F800000; min.ftz.f32 %f26, %f69, %f70; max.ftz.f32 %f71, %f111, %f60; min.ftz.f32 %f27, %f71, %f70; max.ftz.f32 %f72, %f112, %f60; min.ftz.f32 %f28, %f72, %f70; mul.ftz.f32 %f29, %f26, %f26; mul.ftz.f32 %f30, %f27, %f27; mul.ftz.f32 %f31, %f28, %f28; cvt.u16.u32 %rs2, %r1; and.b16 %rs1, %rs2, 255; setp.eq.s16 %p4, %rs1, 1; @%p4 bra $L__BB63_10; setp.eq.s16 %p5, %rs1, 3; @%p5 bra $L__BB63_16; setp.ne.s16 %p6, %rs1, 2; mov.f32 %f115, %f110; mov.f32 %f116, %f111; mov.f32 %f117, %f112; @%p6 bra $L__BB63_17; add.ftz.f32 %f73, %f26, %f26; mov.f32 %f74, 0f40400000; sub.ftz.f32 %f75, %f74, %f73; add.ftz.f32 %f76, %f27, %f27; sub.ftz.f32 %f77, %f74, %f76; add.ftz.f32 %f78, %f28, %f28; sub.ftz.f32 %f79, %f74, %f78; mul.ftz.f32 %f115, %f29, %f75; mul.ftz.f32 %f116, %f30, %f77; mul.ftz.f32 %f117, %f31, %f79; bra.uni $L__BB63_17; $L__BB63_16: mul.ftz.f32 %f92, %f26, %f29; fma.rn.ftz.f32 %f93, %f26, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f94, %f27, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f95, %f28, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f96, %f26, %f93, 0f41200000; fma.rn.ftz.f32 %f97, %f27, %f94, 0f41200000; fma.rn.ftz.f32 %f98, %f28, %f95, 0f41200000; mul.ftz.f32 %f115, %f92, %f96; mul.ftz.f32 %f99, %f27, %f30; mul.ftz.f32 %f116, %f99, %f97; mul.ftz.f32 %f100, %f28, %f31; mul.ftz.f32 %f117, %f100, %f98; bra.uni $L__BB63_17; $L__BB63_10: setp.leu.ftz.f32 %p7, %f7, 0f00000000; mov.f32 %f117, 0f00000000; mov.f32 %f115, %f117; @%p7 bra $L__BB63_12; add.ftz.f32 %f81, %f7, 0f3F800000; mul.ftz.f32 %f82, %f81, %f110; cvt.rmi.ftz.f32.f32 %f83, %f82; div.approx.ftz.f32 %f115, %f83, %f7; $L__BB63_12: setp.leu.ftz.f32 %p8, %f8, 0f00000000; mov.f32 %f116, %f117; @%p8 bra $L__BB63_14; add.ftz.f32 %f85, %f8, 0f3F800000; mul.ftz.f32 %f86, %f85, %f111; cvt.rmi.ftz.f32.f32 %f87, %f86; div.approx.ftz.f32 %f116, %f87, %f8; $L__BB63_14: setp.leu.ftz.f32 %p9, %f9, 0f00000000; @%p9 bra $L__BB63_17; add.ftz.f32 %f89, %f9, 0f3F800000; mul.ftz.f32 %f90, %f89, %f112; cvt.rmi.ftz.f32.f32 %f91, %f90; div.approx.ftz.f32 %f117, %f91, %f9; $L__BB63_17: sub.ftz.f32 %f101, %f4, %f1; sub.ftz.f32 %f102, %f5, %f2; sub.ftz.f32 %f103, %f6, %f3; fma.rn.ftz.f32 %f120, %f101, %f115, %f1; fma.rn.ftz.f32 %f121, %f102, %f116, %f2; fma.rn.ftz.f32 %f122, %f103, %f117, %f3; and.b32 %r11, %r2, 65280; setp.eq.s32 %p10, %r11, 0; and.b32 %r12, %r1, 254; or.b32 %r13, %r12, 1; setp.eq.s32 %p11, %r13, 3; or.pred %p12, %p10, %p11; @%p12 bra $L__BB63_27; setp.gt.ftz.f32 %p13, %f1, %f4; @%p13 bra $L__BB63_20; bra.uni $L__BB63_19; $L__BB63_20: max.ftz.f32 %f105, %f120, %f4; min.ftz.f32 %f120, %f105, %f1; bra.uni $L__BB63_21; $L__BB63_19: max.ftz.f32 %f104, %f120, %f1; min.ftz.f32 %f120, %f104, %f4; $L__BB63_21: setp.gt.ftz.f32 %p14, %f2, %f5; @%p14 bra $L__BB63_23; bra.uni $L__BB63_22; $L__BB63_23: max.ftz.f32 %f107, %f121, %f5; min.ftz.f32 %f121, %f107, %f2; bra.uni $L__BB63_24; $L__BB63_22: max.ftz.f32 %f106, %f121, %f2; min.ftz.f32 %f121, %f106, %f5; $L__BB63_24: setp.gt.ftz.f32 %p15, %f3, %f6; @%p15 bra $L__BB63_26; bra.uni $L__BB63_25; $L__BB63_26: max.ftz.f32 %f109, %f122, %f6; min.ftz.f32 %f122, %f109, %f3; bra.uni $L__BB63_27; $L__BB63_25: max.ftz.f32 %f108, %f122, %f3; min.ftz.f32 %f122, %f108, %f6; $L__BB63_27: shr.u32 %r14, %r2, 24; mul.wide.u32 %rd19, %r14, 4; add.s64 %rd20, %rd1, %rd19; st.local.f32 [%rd20], %f120; st.local.f32 [%rd20+4], %f121; st.local.f32 [%rd20+8], %f122; st.param.b32 [func_retval0+0], %r3; ret; } .func _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<13>; .reg .f32 %f<131>; .reg .b32 %r<8>; .reg .b64 %rd<16>; ld.param.u64 %rd2, [_Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r1, [_Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r3, [_Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; shl.b32 %r4, %r3, 2; cvt.u64.u32 %rd3, %r4; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f1, [%rd5]; ld.local.f32 %f2, [%rd5+4]; ld.local.f32 %f3, [%rd5+8]; shr.u32 %r5, %r3, 6; cvt.u64.u32 %rd6, %r5; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f4, [%rd8]; ld.local.f32 %f5, [%rd8+4]; ld.local.f32 %f6, [%rd8+8]; shr.u32 %r6, %r3, 14; cvt.u64.u32 %rd9, %r6; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; shr.u32 %r7, %r3, 24; mul.wide.u32 %rd12, %r7, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f7, [%rd13]; ld.local.f32 %f8, [%rd13+4]; ld.local.f32 %f9, [%rd13+8]; ld.local.f32 %f59, [%rd11]; cos.approx.ftz.f32 %f60, %f59; ld.local.f32 %f61, [%rd11+4]; cos.approx.ftz.f32 %f62, %f61; ld.local.f32 %f63, [%rd11+8]; cos.approx.ftz.f32 %f64, %f63; sin.approx.ftz.f32 %f65, %f59; sin.approx.ftz.f32 %f66, %f61; sin.approx.ftz.f32 %f67, %f63; mul.ftz.f32 %f10, %f62, %f64; mul.ftz.f32 %f11, %f62, %f67; neg.ftz.f32 %f12, %f66; mul.ftz.f32 %f68, %f65, %f66; mul.ftz.f32 %f69, %f64, %f68; mul.ftz.f32 %f70, %f60, %f67; sub.ftz.f32 %f13, %f69, %f70; mul.ftz.f32 %f71, %f68, %f67; fma.rn.ftz.f32 %f14, %f60, %f64, %f71; mul.ftz.f32 %f15, %f62, %f65; mul.ftz.f32 %f72, %f60, %f66; mul.ftz.f32 %f73, %f65, %f67; fma.rn.ftz.f32 %f16, %f64, %f72, %f73; mul.ftz.f32 %f74, %f72, %f67; mul.ftz.f32 %f75, %f64, %f65; sub.ftz.f32 %f17, %f74, %f75; mul.ftz.f32 %f18, %f60, %f62; mov.f32 %f128, 0f00000000; setp.gt.s32 %p1, %r1, 1; @%p1 bra $L__BB64_9; setp.eq.s32 %p4, %r1, 0; @%p4 bra $L__BB64_20; setp.eq.s32 %p5, %r1, 1; mov.f32 %f129, %f128; mov.f32 %f130, %f128; @%p5 bra $L__BB64_3; bra.uni $L__BB64_21; $L__BB64_3: sub.ftz.f32 %f99, %f1, %f4; sub.ftz.f32 %f100, %f2, %f5; mul.ftz.f32 %f101, %f100, %f11; fma.rn.ftz.f32 %f102, %f99, %f10, %f101; sub.ftz.f32 %f103, %f3, %f6; fma.rn.ftz.f32 %f38, %f103, %f12, %f102; mul.ftz.f32 %f104, %f100, %f14; fma.rn.ftz.f32 %f105, %f99, %f13, %f104; fma.rn.ftz.f32 %f39, %f103, %f15, %f105; mul.ftz.f32 %f106, %f100, %f17; fma.rn.ftz.f32 %f107, %f99, %f16, %f106; fma.rn.ftz.f32 %f40, %f103, %f18, %f107; setp.eq.ftz.f32 %p10, %f7, 0f00000000; mov.f32 %f130, 0f00000000; mov.f32 %f128, %f130; @%p10 bra $L__BB64_5; div.approx.ftz.f32 %f128, %f38, %f7; $L__BB64_5: setp.eq.ftz.f32 %p11, %f8, 0f00000000; mov.f32 %f129, %f130; @%p11 bra $L__BB64_7; div.approx.ftz.f32 %f129, %f39, %f8; $L__BB64_7: setp.eq.ftz.f32 %p12, %f9, 0f00000000; @%p12 bra $L__BB64_21; div.approx.ftz.f32 %f130, %f40, %f9; bra.uni $L__BB64_21; $L__BB64_9: setp.eq.s32 %p2, %r1, 2; @%p2 bra $L__BB64_19; setp.ne.s32 %p3, %r1, 3; mov.f32 %f129, %f128; mov.f32 %f130, %f128; @%p3 bra $L__BB64_21; setp.eq.ftz.f32 %p6, %f7, 0f00000000; mov.f32 %f124, 0f00000000; mov.f32 %f122, %f124; @%p6 bra $L__BB64_13; div.approx.ftz.f32 %f122, %f1, %f7; $L__BB64_13: setp.eq.ftz.f32 %p7, %f8, 0f00000000; mov.f32 %f123, %f124; @%p7 bra $L__BB64_15; div.approx.ftz.f32 %f123, %f2, %f8; $L__BB64_15: setp.eq.ftz.f32 %p8, %f9, 0f00000000; @%p8 bra $L__BB64_17; div.approx.ftz.f32 %f124, %f3, %f9; $L__BB64_17: mul.ftz.f32 %f79, %f13, %f123; fma.rn.ftz.f32 %f80, %f10, %f122, %f79; fma.rn.ftz.f32 %f128, %f16, %f124, %f80; mul.ftz.f32 %f81, %f14, %f123; fma.rn.ftz.f32 %f82, %f11, %f122, %f81; fma.rn.ftz.f32 %f129, %f17, %f124, %f82; mul.ftz.f32 %f83, %f15, %f123; fma.rn.ftz.f32 %f84, %f122, %f12, %f83; fma.rn.ftz.f32 %f130, %f18, %f124, %f84; mul.ftz.f32 %f85, %f129, %f129; fma.rn.ftz.f32 %f86, %f128, %f128, %f85; fma.rn.ftz.f32 %f87, %f130, %f130, %f86; sqrt.approx.ftz.f32 %f28, %f87; setp.neu.ftz.f32 %p9, %f28, 0f00000000; @%p9 bra $L__BB64_18; bra.uni $L__BB64_21; $L__BB64_18: rcp.approx.ftz.f32 %f88, %f28; mul.ftz.f32 %f130, %f130, %f88; mul.ftz.f32 %f129, %f129, %f88; mul.ftz.f32 %f128, %f128, %f88; bra.uni $L__BB64_21; $L__BB64_20: mul.ftz.f32 %f110, %f1, %f7; mul.ftz.f32 %f111, %f2, %f8; mul.ftz.f32 %f112, %f111, %f13; fma.rn.ftz.f32 %f113, %f110, %f10, %f112; mul.ftz.f32 %f114, %f3, %f9; fma.rn.ftz.f32 %f115, %f114, %f16, %f113; mul.ftz.f32 %f116, %f111, %f14; fma.rn.ftz.f32 %f117, %f110, %f11, %f116; fma.rn.ftz.f32 %f118, %f114, %f17, %f117; mul.ftz.f32 %f119, %f111, %f15; fma.rn.ftz.f32 %f120, %f110, %f12, %f119; fma.rn.ftz.f32 %f121, %f114, %f18, %f120; add.ftz.f32 %f130, %f6, %f121; add.ftz.f32 %f129, %f5, %f118; add.ftz.f32 %f128, %f4, %f115; bra.uni $L__BB64_21; $L__BB64_19: mul.ftz.f32 %f89, %f1, %f7; mul.ftz.f32 %f90, %f2, %f8; mul.ftz.f32 %f91, %f90, %f13; fma.rn.ftz.f32 %f92, %f89, %f10, %f91; mul.ftz.f32 %f93, %f3, %f9; mul.ftz.f32 %f94, %f90, %f14; fma.rn.ftz.f32 %f95, %f89, %f11, %f94; mul.ftz.f32 %f96, %f90, %f15; fma.rn.ftz.f32 %f97, %f89, %f12, %f96; fma.rn.ftz.f32 %f130, %f93, %f18, %f97; fma.rn.ftz.f32 %f129, %f93, %f17, %f95; fma.rn.ftz.f32 %f128, %f93, %f16, %f92; $L__BB64_21: mul.wide.u32 %rd14, %r2, 4; add.s64 %rd15, %rd1, %rd14; st.local.f32 [%rd15], %f128; st.local.f32 [%rd15+4], %f129; st.local.f32 [%rd15+8], %f130; ret; } .func (.param .b32 func_retval0) _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji( .param .b64 _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_0, .param .b32 _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_1, .param .b32 _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_2, .param .b32 _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_3 ) { .reg .f32 %f<40>; .reg .b32 %r<5>; .reg .b64 %rd<11>; ld.param.u64 %rd1, [_Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_0]; ld.param.u32 %r1, [_Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_1]; ld.param.u32 %r2, [_Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_2]; ld.param.u32 %r3, [_Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji_param_3]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; ld.local.f32 %f1, [%rd4]; ld.local.f32 %f2, [%rd4+4]; ld.local.f32 %f3, [%rd4+8]; ld.const.u64 %rd5, [kernel_params+2400]; cvta.to.global.u64 %rd6, %rd5; mul.wide.s32 %rd7, %r3, 16; add.s64 %rd8, %rd6, %rd7; ld.global.v4.f32 {%f4, %f5, %f6, %f7}, [%rd8]; ld.global.v4.f32 {%f12, %f13, %f14, %f15}, [%rd8+16]; ld.global.v4.f32 {%f20, %f21, %f22, %f23}, [%rd8+32]; add.s32 %r4, %r3, 3; mul.ftz.f32 %f28, %f2, %f5; fma.rn.ftz.f32 %f29, %f1, %f4, %f28; fma.rn.ftz.f32 %f30, %f3, %f6, %f29; add.ftz.f32 %f31, %f30, %f7; mul.ftz.f32 %f32, %f2, %f13; fma.rn.ftz.f32 %f33, %f1, %f12, %f32; fma.rn.ftz.f32 %f34, %f3, %f14, %f33; add.ftz.f32 %f35, %f34, %f15; mul.ftz.f32 %f36, %f2, %f21; fma.rn.ftz.f32 %f37, %f1, %f20, %f36; fma.rn.ftz.f32 %f38, %f3, %f22, %f37; add.ftz.f32 %f39, %f38, %f23; mul.wide.u32 %rd9, %r2, 4; add.s64 %rd10, %rd2, %rd9; st.local.f32 [%rd10], %f31; st.local.f32 [%rd10+4], %f35; st.local.f32 [%rd10+8], %f39; st.param.b32 [func_retval0+0], %r4; ret; } .func (.param .b32 func_retval0) _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji( .param .b64 _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_0, .param .b32 _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_1, .param .b32 _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_2, .param .b32 _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_3 ) { .reg .f32 %f<24>; .reg .b32 %r<5>; .reg .b64 %rd<11>; ld.param.u64 %rd1, [_Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_0]; ld.param.u32 %r1, [_Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_1]; ld.param.u32 %r2, [_Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_2]; ld.param.u32 %r3, [_Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji_param_3]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; ld.local.f32 %f1, [%rd4]; ld.local.f32 %f2, [%rd4+4]; ld.local.f32 %f3, [%rd4+8]; ld.const.u64 %rd5, [kernel_params+2400]; cvta.to.global.u64 %rd6, %rd5; mul.wide.s32 %rd7, %r3, 16; add.s64 %rd8, %rd6, %rd7; ld.global.v4.f32 {%f4, %f5, %f6, %f7}, [%rd8]; ld.global.v4.f32 {%f11, %f12, %f13, %f14}, [%rd8+16]; add.s32 %r4, %r3, 2; max.ftz.f32 %f18, %f4, %f1; max.ftz.f32 %f19, %f5, %f2; max.ftz.f32 %f20, %f6, %f3; min.ftz.f32 %f21, %f18, %f11; min.ftz.f32 %f22, %f19, %f12; min.ftz.f32 %f23, %f20, %f13; mul.wide.u32 %rd9, %r2, 4; add.s64 %rd10, %rd2, %rd9; st.local.f32 [%rd10], %f21; st.local.f32 [%rd10+4], %f22; st.local.f32 [%rd10+8], %f23; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<131>; .reg .f32 %f<355>; .reg .b32 %r<105>; .reg .b64 %rd<15>; ld.param.u64 %rd1, [_Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r30, [_Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r32, [_Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r31, [_Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd2, %rd1; shl.b32 %r33, %r32, 2; cvt.u64.u32 %rd3, %r33; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd2, %rd4; ld.local.f32 %f1, [%rd5]; shr.u32 %r34, %r32, 6; cvt.u64.u32 %rd6, %r34; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd2, %rd7; ld.local.f32 %f2, [%rd8]; shr.u32 %r35, %r32, 14; cvt.u64.u32 %rd9, %r35; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd2, %rd10; ld.local.f32 %f3, [%rd11]; mov.f32 %f87, 0f00000000; setp.gt.s32 %p1, %r30, 19; @%p1 bra $L__BB67_42; setp.gt.s32 %p29, %r30, 9; @%p29 bra $L__BB67_21; setp.gt.s32 %p43, %r30, 4; @%p43 bra $L__BB67_11; setp.gt.s32 %p50, %r30, 1; @%p50 bra $L__BB67_7; bra.uni $L__BB67_4; $L__BB67_7: setp.eq.s32 %p51, %r30, 2; @%p51 bra $L__BB67_149; setp.eq.s32 %p52, %r30, 3; @%p52 bra $L__BB67_147; setp.eq.s32 %p53, %r30, 4; mov.f32 %f354, %f87; @%p53 bra $L__BB67_10; bra.uni $L__BB67_151; $L__BB67_10: sin.approx.ftz.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_42: setp.gt.s32 %p2, %r30, 29; @%p2 bra $L__BB67_62; setp.gt.s32 %p16, %r30, 24; @%p16 bra $L__BB67_53; setp.gt.s32 %p23, %r30, 21; @%p23 bra $L__BB67_48; bra.uni $L__BB67_45; $L__BB67_48: setp.eq.s32 %p24, %r30, 22; @%p24 bra $L__BB67_102; setp.eq.s32 %p25, %r30, 23; @%p25 bra $L__BB67_101; setp.eq.s32 %p26, %r30, 24; mov.f32 %f354, %f87; @%p26 bra $L__BB67_51; bra.uni $L__BB67_151; $L__BB67_51: setp.leu.ftz.f32 %p73, %f1, 0f00000000; mov.f32 %f354, %f87; @%p73 bra $L__BB67_151; rsqrt.approx.ftz.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_21: setp.gt.s32 %p30, %r30, 14; @%p30 bra $L__BB67_32; setp.gt.s32 %p37, %r30, 11; @%p37 bra $L__BB67_28; bra.uni $L__BB67_23; $L__BB67_28: setp.eq.s32 %p38, %r30, 12; @%p38 bra $L__BB67_123; setp.eq.s32 %p39, %r30, 13; @%p39 bra $L__BB67_122; setp.eq.s32 %p40, %r30, 14; mov.f32 %f354, %f87; @%p40 bra $L__BB67_31; bra.uni $L__BB67_151; $L__BB67_31: add.ftz.f32 %f242, %f1, 0f3F000000; cvt.rmi.ftz.f32.f32 %f354, %f242; bra.uni $L__BB67_151; $L__BB67_62: setp.gt.s32 %p3, %r30, 34; @%p3 bra $L__BB67_73; setp.gt.s32 %p10, %r30, 31; @%p10 bra $L__BB67_68; bra.uni $L__BB67_64; $L__BB67_68: setp.eq.s32 %p11, %r30, 32; @%p11 bra $L__BB67_92; setp.eq.s32 %p12, %r30, 33; @%p12 bra $L__BB67_89; setp.eq.s32 %p13, %r30, 34; mov.f32 %f354, %f87; @%p13 bra $L__BB67_71; bra.uni $L__BB67_151; $L__BB67_71: sub.ftz.f32 %f13, %f2, %f3; setp.eq.ftz.f32 %p61, %f13, 0f00000000; mov.f32 %f354, %f3; @%p61 bra $L__BB67_151; sub.ftz.f32 %f122, %f1, %f3; div.approx.ftz.f32 %f123, %f122, %f13; cvt.rmi.ftz.f32.f32 %f124, %f123; mul.ftz.f32 %f125, %f13, %f124; sub.ftz.f32 %f354, %f1, %f125; bra.uni $L__BB67_151; $L__BB67_11: setp.gt.s32 %p44, %r30, 6; @%p44 bra $L__BB67_15; bra.uni $L__BB67_12; $L__BB67_15: setp.eq.s32 %p45, %r30, 7; @%p45 bra $L__BB67_145; setp.eq.s32 %p46, %r30, 8; @%p46 bra $L__BB67_144; setp.eq.s32 %p47, %r30, 9; mov.f32 %f354, %f87; @%p47 bra $L__BB67_18; bra.uni $L__BB67_151; $L__BB67_18: abs.ftz.f32 %f73, %f1; setp.leu.ftz.f32 %p123, %f73, 0f3F800000; mov.f32 %f353, %f73; @%p123 bra $L__BB67_20; rcp.approx.ftz.f32 %f353, %f73; $L__BB67_20: mul.rn.ftz.f32 %f274, %f353, %f353; mov.f32 %f275, 0fC0B59883; mov.f32 %f276, 0fBF52C7EA; fma.rn.ftz.f32 %f277, %f274, %f276, %f275; mov.f32 %f278, 0fC0D21907; fma.rn.ftz.f32 %f279, %f277, %f274, %f278; mul.ftz.f32 %f280, %f274, %f279; mul.ftz.f32 %f281, %f353, %f280; add.ftz.f32 %f282, %f274, 0f41355DC0; mov.f32 %f283, 0f41E6BD60; fma.rn.ftz.f32 %f284, %f282, %f274, %f283; mov.f32 %f285, 0f419D92C8; fma.rn.ftz.f32 %f286, %f284, %f274, %f285; rcp.approx.ftz.f32 %f287, %f286; fma.rn.ftz.f32 %f288, %f281, %f287, %f353; mov.f32 %f289, 0f3FC90FDB; sub.ftz.f32 %f290, %f289, %f288; setp.gt.ftz.f32 %p124, %f73, 0f3F800000; selp.f32 %f291, %f290, %f288, %p124; mov.b32 %r89, %f291; mov.b32 %r90, %f1; and.b32 %r91, %r90, -2147483648; or.b32 %r92, %r91, %r89; mov.b32 %f292, %r92; setp.le.ftz.f32 %p125, %f73, 0f7F800000; selp.f32 %f354, %f292, %f291, %p125; bra.uni $L__BB67_151; $L__BB67_53: setp.gt.s32 %p17, %r30, 26; @%p17 bra $L__BB67_57; bra.uni $L__BB67_54; $L__BB67_57: setp.eq.s32 %p18, %r30, 27; @%p18 bra $L__BB67_99; setp.eq.s32 %p19, %r30, 28; @%p19 bra $L__BB67_98; setp.eq.s32 %p20, %r30, 29; mov.f32 %f354, %f87; @%p20 bra $L__BB67_60; bra.uni $L__BB67_151; $L__BB67_60: abs.ftz.f32 %f24, %f1; setp.ltu.ftz.f32 %p68, %f24, 0f3F800000; @%p68 bra $L__BB67_97; bra.uni $L__BB67_61; $L__BB67_97: mul.ftz.f32 %f187, %f1, %f1; mov.f32 %f188, 0f394FFF49; mov.f32 %f189, 0f363D0ADA; fma.rn.ftz.f32 %f190, %f189, %f187, %f188; mov.f32 %f191, 0f3C08889A; fma.rn.ftz.f32 %f192, %f190, %f187, %f191; mov.f32 %f193, 0f3E2AAAAB; fma.rn.ftz.f32 %f194, %f192, %f187, %f193; mul.ftz.f32 %f195, %f187, %f194; fma.rn.ftz.f32 %f354, %f195, %f1, %f1; bra.uni $L__BB67_151; $L__BB67_32: setp.gt.s32 %p31, %r30, 16; @%p31 bra $L__BB67_36; bra.uni $L__BB67_33; $L__BB67_36: setp.eq.s32 %p32, %r30, 17; @%p32 bra $L__BB67_107; setp.eq.s32 %p33, %r30, 18; @%p33 bra $L__BB67_106; setp.eq.s32 %p34, %r30, 19; mov.f32 %f354, %f87; @%p34 bra $L__BB67_39; bra.uni $L__BB67_151; $L__BB67_39: abs.ftz.f32 %f36, %f2; setp.eq.ftz.f32 %p74, %f36, 0f00000000; abs.ftz.f32 %f37, %f1; setp.eq.ftz.f32 %p75, %f37, 0f00000000; and.pred %p76, %p74, %p75; mov.b32 %r1, %f2; mov.b32 %r54, %f1; and.b32 %r2, %r54, -2147483648; @%p76 bra $L__BB67_105; bra.uni $L__BB67_40; $L__BB67_105: shr.s32 %r59, %r1, 31; and.b32 %r60, %r59, 1078530011; or.b32 %r61, %r60, %r2; mov.b32 %f354, %r61; bra.uni $L__BB67_151; $L__BB67_73: setp.gt.s32 %p4, %r30, 36; @%p4 bra $L__BB67_77; bra.uni $L__BB67_74; $L__BB67_77: setp.eq.s32 %p5, %r30, 37; @%p5 bra $L__BB67_85; setp.eq.s32 %p6, %r30, 38; @%p6 bra $L__BB67_83; setp.ne.s32 %p7, %r30, 39; mov.f32 %f354, %f87; @%p7 bra $L__BB67_151; neg.ftz.f32 %f88, %f1; neg.ftz.f32 %f89, %f2; min.ftz.f32 %f349, %f88, %f89; setp.eq.ftz.f32 %p56, %f3, 0f00000000; @%p56 bra $L__BB67_82; sub.ftz.f32 %f90, %f2, %f1; abs.ftz.f32 %f91, %f90; sub.ftz.f32 %f92, %f3, %f91; mov.f32 %f93, 0f00000000; max.ftz.f32 %f94, %f92, %f93; div.approx.ftz.f32 %f95, %f94, %f3; mul.ftz.f32 %f96, %f95, %f95; mul.ftz.f32 %f97, %f95, %f96; mul.ftz.f32 %f98, %f3, %f97; fma.rn.ftz.f32 %f349, %f98, 0fBE2AAAAB, %f349; $L__BB67_82: neg.ftz.f32 %f354, %f349; bra.uni $L__BB67_151; $L__BB67_4: setp.eq.s32 %p54, %r30, 0; @%p54 bra $L__BB67_150; setp.eq.s32 %p55, %r30, 1; mov.f32 %f354, %f87; @%p55 bra $L__BB67_6; bra.uni $L__BB67_151; $L__BB67_6: sub.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_45: setp.eq.s32 %p27, %r30, 20; @%p27 bra $L__BB67_103; setp.eq.s32 %p28, %r30, 21; mov.f32 %f354, %f87; @%p28 bra $L__BB67_47; bra.uni $L__BB67_151; $L__BB67_47: cvt.rpi.ftz.f32.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_23: setp.eq.s32 %p41, %r30, 10; @%p41 bra $L__BB67_124; setp.eq.s32 %p42, %r30, 11; mov.f32 %f354, %f87; @%p42 bra $L__BB67_25; bra.uni $L__BB67_151; $L__BB67_25: setp.le.ftz.f32 %p100, %f1, 0f00000000; setp.le.ftz.f32 %p101, %f2, 0f00000000; or.pred %p102, %p100, %p101; mov.f32 %f354, %f87; @%p102 bra $L__BB67_151; lg2.approx.ftz.f32 %f245, %f2; mul.ftz.f32 %f58, %f245, 0f3F317218; setp.eq.ftz.f32 %p103, %f58, 0f00000000; mov.f32 %f354, %f87; @%p103 bra $L__BB67_151; lg2.approx.ftz.f32 %f246, %f1; mul.ftz.f32 %f247, %f246, 0f3F317218; div.approx.ftz.f32 %f354, %f247, %f58; bra.uni $L__BB67_151; $L__BB67_64: setp.eq.s32 %p14, %r30, 30; @%p14 bra $L__BB67_96; setp.eq.s32 %p15, %r30, 31; mov.f32 %f354, %f87; @%p15 bra $L__BB67_66; bra.uni $L__BB67_151; $L__BB67_66: abs.ftz.f32 %f20, %f1; setp.ltu.ftz.f32 %p64, %f20, 0f3F19999A; @%p64 bra $L__BB67_95; bra.uni $L__BB67_67; $L__BB67_95: mul.ftz.f32 %f136, %f1, %f1; mov.f32 %f137, 0fBD563CAE; mov.f32 %f138, 0f3C80F082; fma.rn.ftz.f32 %f139, %f138, %f136, %f137; mov.f32 %f140, 0f3E085941; fma.rn.ftz.f32 %f141, %f139, %f136, %f140; mov.f32 %f142, 0fBEAAA9ED; fma.rn.ftz.f32 %f143, %f141, %f136, %f142; mov.f32 %f144, 0f00000000; fma.rn.ftz.f32 %f145, %f143, %f136, %f144; fma.rn.ftz.f32 %f354, %f145, %f1, %f1; bra.uni $L__BB67_151; $L__BB67_12: setp.eq.s32 %p48, %r30, 5; @%p48 bra $L__BB67_146; setp.eq.s32 %p49, %r30, 6; mov.f32 %f354, %f87; @%p49 bra $L__BB67_14; bra.uni $L__BB67_151; $L__BB67_14: sin.approx.ftz.f32 %f346, %f1; cos.approx.ftz.f32 %f347, %f1; div.approx.ftz.f32 %f354, %f346, %f347; bra.uni $L__BB67_151; $L__BB67_54: setp.eq.s32 %p21, %r30, 25; @%p21 bra $L__BB67_100; setp.eq.s32 %p22, %r30, 26; mov.f32 %f354, %f87; @%p22 bra $L__BB67_56; bra.uni $L__BB67_151; $L__BB67_56: mul.ftz.f32 %f196, %f1, 0f3FB8AA3B; ex2.approx.ftz.f32 %f354, %f196; bra.uni $L__BB67_151; $L__BB67_33: setp.eq.s32 %p35, %r30, 15; @%p35 bra $L__BB67_121; setp.eq.s32 %p36, %r30, 16; mov.f32 %f354, %f87; @%p36 bra $L__BB67_35; bra.uni $L__BB67_151; $L__BB67_35: setp.gt.ftz.f32 %p98, %f1, %f2; selp.f32 %f354, 0f3F800000, 0f00000000, %p98; bra.uni $L__BB67_151; $L__BB67_74: setp.eq.s32 %p8, %r30, 35; @%p8 bra $L__BB67_87; setp.eq.s32 %p9, %r30, 36; mov.f32 %f354, %f87; @%p9 bra $L__BB67_76; bra.uni $L__BB67_151; $L__BB67_76: fma.rn.ftz.f32 %f354, %f1, %f2, %f3; bra.uni $L__BB67_151; $L__BB67_149: mul.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_147: setp.eq.ftz.f32 %p130, %f2, 0f00000000; mov.f32 %f354, %f87; @%p130 bra $L__BB67_151; div.approx.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_102: cvt.rmi.ftz.f32.f32 %f201, %f1; sub.ftz.f32 %f354, %f1, %f201; bra.uni $L__BB67_151; $L__BB67_101: mov.f32 %f199, 0f00000000; max.ftz.f32 %f200, %f1, %f199; sqrt.approx.ftz.f32 %f354, %f200; bra.uni $L__BB67_151; $L__BB67_123: min.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_122: max.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_92: setp.ltu.ftz.f32 %p63, %f1, 0f00000000; @%p63 bra $L__BB67_94; bra.uni $L__BB67_93; $L__BB67_94: cvt.rpi.ftz.f32.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_89: setp.eq.ftz.f32 %p62, %f2, 0f00000000; mov.f32 %f350, 0f00000000; @%p62 bra $L__BB67_91; div.approx.ftz.f32 %f350, %f1, %f2; $L__BB67_91: cvt.rmi.ftz.f32.f32 %f127, %f350; mul.ftz.f32 %f354, %f2, %f127; bra.uni $L__BB67_151; $L__BB67_145: mov.f32 %f320, 0fBF800000; max.ftz.f32 %f321, %f1, %f320; mov.f32 %f322, 0f3F800000; min.ftz.f32 %f323, %f321, %f322; abs.ftz.f32 %f324, %f323; sub.ftz.f32 %f325, %f322, %f324; mul.ftz.f32 %f326, %f325, 0f3F000000; sqrt.approx.ftz.f32 %f327, %f326; setp.gt.ftz.f32 %p128, %f324, 0f3F11EB85; selp.f32 %f328, %f327, %f324, %p128; mul.ftz.f32 %f329, %f328, %f328; mov.f32 %f330, 0f3C94D2E9; mov.f32 %f331, 0f3D53F941; fma.rn.ftz.f32 %f332, %f331, %f329, %f330; mov.f32 %f333, 0f3D3F841F; fma.rn.ftz.f32 %f334, %f332, %f329, %f333; mov.f32 %f335, 0f3D994929; fma.rn.ftz.f32 %f336, %f334, %f329, %f335; mov.f32 %f337, 0f3E2AAB94; fma.rn.ftz.f32 %f338, %f336, %f329, %f337; mul.ftz.f32 %f339, %f329, %f338; fma.rn.ftz.f32 %f340, %f339, %f328, %f328; mov.f32 %f341, 0f3FC90FDB; mov.f32 %f342, 0fC0000000; fma.rn.ftz.f32 %f343, %f342, %f340, %f341; selp.f32 %f344, %f343, %f340, %p128; setp.le.ftz.f32 %p129, %f344, 0f7F800000; mov.b32 %r93, %f344; mov.b32 %r94, %f323; and.b32 %r95, %r94, -2147483648; or.b32 %r96, %r95, %r93; mov.b32 %f345, %r96; selp.f32 %f354, %f345, %f344, %p129; bra.uni $L__BB67_151; $L__BB67_144: mov.f32 %f293, 0fBF800000; max.ftz.f32 %f294, %f1, %f293; mov.f32 %f295, 0f3F800000; min.ftz.f32 %f296, %f294, %f295; abs.ftz.f32 %f297, %f296; sub.ftz.f32 %f298, %f295, %f297; mul.ftz.f32 %f299, %f298, 0f3F000000; sqrt.approx.ftz.f32 %f300, %f299; setp.gt.ftz.f32 %p126, %f297, 0f3F11EB85; selp.f32 %f301, %f300, %f297, %p126; mul.ftz.f32 %f302, %f301, %f301; mov.f32 %f303, 0f3C94D2E9; mov.f32 %f304, 0f3D53F941; fma.rn.ftz.f32 %f305, %f304, %f302, %f303; mov.f32 %f306, 0f3D3F841F; fma.rn.ftz.f32 %f307, %f305, %f302, %f306; mov.f32 %f308, 0f3D994929; fma.rn.ftz.f32 %f309, %f307, %f302, %f308; mov.f32 %f310, 0f3E2AAB94; fma.rn.ftz.f32 %f311, %f309, %f302, %f310; mul.ftz.f32 %f312, %f302, %f311; fma.rn.ftz.f32 %f313, %f312, %f301, %f301; add.ftz.f32 %f314, %f313, %f313; mov.f32 %f315, 0f3FC90FDB; sub.ftz.f32 %f316, %f315, %f313; selp.f32 %f317, %f314, %f316, %p126; setp.lt.ftz.f32 %p127, %f296, 0f00000000; mov.f32 %f318, 0f40490FDB; sub.ftz.f32 %f319, %f318, %f317; selp.f32 %f354, %f319, %f317, %p127; bra.uni $L__BB67_151; $L__BB67_99: mul.ftz.f32 %f354, %f1, 0f3C8EFA35; bra.uni $L__BB67_151; $L__BB67_98: mul.ftz.f32 %f354, %f1, 0f42652EE0; bra.uni $L__BB67_151; $L__BB67_107: setp.eq.ftz.f32 %p84, %f2, 0f00000000; mov.f32 %f354, %f87; @%p84 bra $L__BB67_151; abs.ftz.f32 %f42, %f1; abs.ftz.f32 %f43, %f2; mul.ftz.f32 %f44, %f43, 0f4B000000; setp.lt.ftz.f32 %p85, %f42, %f43; mov.b32 %r100, %f42; @%p85 bra $L__BB67_120; setp.le.ftz.f32 %p86, %f42, %f44; setp.leu.ftz.f32 %p87, %f43, 0f00000000; or.pred %p88, %p87, %p86; setp.gt.u32 %p89, %r100, 2139095039; or.pred %p90, %p89, %p88; @%p90 bra $L__BB67_114; bra.uni $L__BB67_110; $L__BB67_114: div.approx.ftz.f32 %f233, %f42, %f43; cvt.rzi.ftz.f32.f32 %f351, %f233; neg.ftz.f32 %f46, %f43; fma.rn.f32 %f47, %f46, %f351, %f42; mov.b32 %r14, %f47; mov.b32 %r70, %f43; setp.lt.u32 %p93, %r14, %r70; @%p93 bra $L__BB67_119; setp.gt.u32 %p94, %r14, -2147483648; @%p94 bra $L__BB67_118; bra.uni $L__BB67_116; $L__BB67_118: add.ftz.f32 %f239, %f351, 0fBF800000; add.ftz.f32 %f240, %f239, 0fBF800000; setp.lt.ftz.f32 %p97, %f47, %f46; selp.f32 %f351, %f240, %f239, %p97; bra.uni $L__BB67_119; $L__BB67_106: abs.ftz.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_85: setp.eq.ftz.f32 %p58, %f2, 0f00000000; mov.f32 %f354, %f87; @%p58 bra $L__BB67_151; sub.ftz.f32 %f109, %f1, %f2; add.ftz.f32 %f110, %f2, %f2; div.approx.ftz.f32 %f111, %f109, %f110; cvt.rmi.ftz.f32.f32 %f112, %f111; sub.ftz.f32 %f113, %f111, %f112; mul.ftz.f32 %f114, %f2, %f113; fma.rn.ftz.f32 %f115, %f2, %f113, %f114; sub.ftz.f32 %f116, %f115, %f2; abs.ftz.f32 %f354, %f116; bra.uni $L__BB67_151; $L__BB67_83: setp.eq.ftz.f32 %p57, %f3, 0f00000000; min.ftz.f32 %f354, %f1, %f2; @%p57 bra $L__BB67_151; sub.ftz.f32 %f99, %f1, %f2; abs.ftz.f32 %f100, %f99; sub.ftz.f32 %f101, %f3, %f100; mov.f32 %f102, 0f00000000; max.ftz.f32 %f103, %f101, %f102; div.approx.ftz.f32 %f104, %f103, %f3; mul.ftz.f32 %f105, %f104, %f104; mul.ftz.f32 %f106, %f104, %f105; mul.ftz.f32 %f107, %f3, %f106; fma.rn.ftz.f32 %f354, %f107, 0fBE2AAAAB, %f354; bra.uni $L__BB67_151; $L__BB67_150: add.ftz.f32 %f354, %f1, %f2; bra.uni $L__BB67_151; $L__BB67_103: cvt.rmi.ftz.f32.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_124: setp.geu.ftz.f32 %p104, %f1, 0f00000000; @%p104 bra $L__BB67_126; cvt.rzi.ftz.s32.f32 %r74, %f2; cvt.rn.f32.s32 %f249, %r74; setp.neu.ftz.f32 %p105, %f2, %f249; mov.f32 %f354, %f87; @%p105 bra $L__BB67_151; $L__BB67_126: setp.eq.ftz.f32 %p106, %f2, 0f00000000; mov.f32 %f354, 0f3F800000; @%p106 bra $L__BB67_151; setp.lt.ftz.f32 %p107, %f1, 0f00000000; @%p107 bra $L__BB67_130; bra.uni $L__BB67_128; $L__BB67_130: neg.ftz.f32 %f254, %f2; abs.ftz.f32 %f61, %f254; mov.f32 %f255, 0f40000000; abs.ftz.f32 %f62, %f255; mul.ftz.f32 %f63, %f62, 0f4B000000; setp.lt.ftz.f32 %p109, %f61, %f62; mov.b32 %r104, %f61; @%p109 bra $L__BB67_142; setp.le.ftz.f32 %p110, %f61, %f63; setp.leu.ftz.f32 %p111, %f62, 0f00000000; or.pred %p112, %p111, %p110; setp.gt.u32 %p113, %r104, 2139095039; or.pred %p114, %p113, %p112; @%p114 bra $L__BB67_136; bra.uni $L__BB67_132; $L__BB67_136: div.approx.ftz.f32 %f260, %f61, %f62; cvt.rzi.ftz.f32.f32 %f352, %f260; neg.ftz.f32 %f65, %f62; fma.rn.f32 %f66, %f65, %f352, %f61; mov.b32 %r83, %f66; mov.b32 %r84, %f62; setp.lt.u32 %p117, %r83, %r84; @%p117 bra $L__BB67_141; setp.gt.u32 %p118, %r83, -2147483648; @%p118 bra $L__BB67_140; bra.uni $L__BB67_138; $L__BB67_140: add.ftz.f32 %f266, %f352, 0fBF800000; add.ftz.f32 %f267, %f266, 0fBF800000; setp.lt.ftz.f32 %p121, %f66, %f65; selp.f32 %f352, %f267, %f266, %p121; bra.uni $L__BB67_141; $L__BB67_96: abs.ftz.f32 %f146, %f1; mov.f32 %f147, 0f3FB8AA3B; mul.rn.ftz.f32 %f148, %f146, %f147; cvt.rzi.ftz.f32.f32 %f149, %f148; abs.ftz.f32 %f150, %f149; setp.gt.ftz.f32 %p66, %f150, 0f42FC0000; mov.b32 %r40, %f149; and.b32 %r41, %r40, -2147483648; or.b32 %r42, %r41, 1123811328; mov.b32 %f151, %r42; selp.f32 %f152, %f151, %f149, %p66; mov.f32 %f153, 0fBF317218; fma.rn.ftz.f32 %f154, %f152, %f153, %f146; mov.f32 %f155, 0f3102E308; fma.rn.ftz.f32 %f156, %f152, %f155, %f154; mul.ftz.f32 %f157, %f156, 0f3FB8AA3B; add.ftz.f32 %f158, %f152, 0f4B40007D; mov.b32 %r43, %f158; shl.b32 %r44, %r43, 23; mov.b32 %f159, %r44; ex2.approx.ftz.f32 %f160, %f157; mul.ftz.f32 %f161, %f160, %f159; mov.f32 %f162, 0f3E000000; div.approx.ftz.f32 %f163, %f162, %f161; mov.f32 %f164, 0f40000000; fma.rn.ftz.f32 %f165, %f164, %f161, %f163; setp.ge.ftz.f32 %p67, %f146, 0f42B40000; selp.f32 %f354, 0f7F800000, %f165, %p67; bra.uni $L__BB67_151; $L__BB67_146: cos.approx.ftz.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_100: setp.eq.ftz.f32 %p71, %f1, 0f00000000; setp.lt.ftz.f32 %p72, %f1, 0f00000000; selp.f32 %f197, 0fBF800000, 0f3F800000, %p72; selp.f32 %f354, 0f00000000, %f197, %p71; bra.uni $L__BB67_151; $L__BB67_121: setp.lt.ftz.f32 %p99, %f1, %f2; selp.f32 %f354, 0f3F800000, 0f00000000, %p99; bra.uni $L__BB67_151; $L__BB67_87: setp.eq.ftz.f32 %p59, %f1, %f2; mov.f32 %f354, 0f3F800000; @%p59 bra $L__BB67_151; sub.ftz.f32 %f118, %f1, %f2; abs.ftz.f32 %f119, %f118; mov.f32 %f120, 0f34000000; max.ftz.f32 %f121, %f3, %f120; setp.le.ftz.f32 %p60, %f119, %f121; selp.f32 %f354, 0f3F800000, 0f00000000, %p60; bra.uni $L__BB67_151; $L__BB67_61: mov.f32 %f166, 0f3FB8AA3B; mul.rn.ftz.f32 %f167, %f24, %f166; cvt.rzi.ftz.f32.f32 %f168, %f167; abs.ftz.f32 %f169, %f168; setp.gt.ftz.f32 %p69, %f169, 0f42FC0000; mov.b32 %r45, %f168; and.b32 %r46, %r45, -2147483648; or.b32 %r47, %r46, 1123811328; mov.b32 %f170, %r47; selp.f32 %f171, %f170, %f168, %p69; mov.f32 %f172, 0fBF317218; fma.rn.ftz.f32 %f173, %f171, %f172, %f24; mov.f32 %f174, 0f3102E308; fma.rn.ftz.f32 %f175, %f171, %f174, %f173; mul.ftz.f32 %f176, %f175, 0f3FB8AA3B; add.ftz.f32 %f177, %f171, 0f4B40007D; mov.b32 %r48, %f177; shl.b32 %r49, %r48, 23; mov.b32 %f178, %r49; ex2.approx.ftz.f32 %f179, %f176; mul.ftz.f32 %f180, %f179, %f178; mov.f32 %f181, 0f3E000000; div.approx.ftz.f32 %f182, %f181, %f180; neg.ftz.f32 %f183, %f182; mov.f32 %f184, 0f40000000; fma.rn.ftz.f32 %f185, %f184, %f180, %f183; setp.ge.ftz.f32 %p70, %f24, 0f42B40000; selp.f32 %f186, 0f7F800000, %f185, %p70; mov.b32 %r50, %f186; mov.b32 %r51, %f1; and.b32 %r52, %r51, -2147483648; or.b32 %r53, %r52, %r50; mov.b32 %f354, %r53; bra.uni $L__BB67_151; $L__BB67_40: setp.eq.ftz.f32 %p77, %f36, 0f7F800000; setp.eq.ftz.f32 %p78, %f37, 0f7F800000; and.pred %p79, %p77, %p78; @%p79 bra $L__BB67_104; bra.uni $L__BB67_41; $L__BB67_104: setp.lt.s32 %p83, %r1, 0; selp.b32 %r57, 1075235812, 1061752795, %p83; or.b32 %r58, %r57, %r2; mov.b32 %f354, %r58; bra.uni $L__BB67_151; $L__BB67_67: mul.ftz.f32 %f128, %f20, 0f4038AA3B; ex2.approx.ftz.f32 %f129, %f128; add.ftz.f32 %f130, %f129, 0f3F800000; mov.f32 %f131, 0f3F800000; rcp.approx.ftz.f32 %f132, %f130; mov.f32 %f133, 0fC0000000; fma.rn.ftz.f32 %f134, %f132, %f133, %f131; setp.ge.ftz.f32 %p65, %f20, 0f41102CB4; selp.f32 %f135, 0f3F800000, %f134, %p65; mov.b32 %r36, %f135; mov.b32 %r37, %f1; and.b32 %r38, %r37, -2147483648; or.b32 %r39, %r38, %r36; mov.b32 %f354, %r39; bra.uni $L__BB67_151; $L__BB67_93: cvt.rmi.ftz.f32.f32 %f354, %f1; bra.uni $L__BB67_151; $L__BB67_41: setp.lt.s32 %p80, %r1, 0; min.ftz.f32 %f202, %f37, %f36; max.ftz.f32 %f203, %f37, %f36; div.full.ftz.f32 %f204, %f202, %f203; mul.rn.ftz.f32 %f205, %f204, %f204; mov.f32 %f206, 0fC0B59883; mov.f32 %f207, 0fBF52C7EA; fma.rn.ftz.f32 %f208, %f205, %f207, %f206; mov.f32 %f209, 0fC0D21907; fma.rn.ftz.f32 %f210, %f208, %f205, %f209; mul.ftz.f32 %f211, %f205, %f210; mul.ftz.f32 %f212, %f204, %f211; add.ftz.f32 %f213, %f205, 0f41355DC0; mov.f32 %f214, 0f41E6BD60; fma.rn.ftz.f32 %f215, %f213, %f205, %f214; mov.f32 %f216, 0f419D92C8; fma.rn.ftz.f32 %f217, %f215, %f205, %f216; rcp.approx.ftz.f32 %f218, %f217; fma.rn.ftz.f32 %f219, %f212, %f218, %f204; mov.f32 %f220, 0f3FC90FDB; sub.ftz.f32 %f221, %f220, %f219; setp.gt.ftz.f32 %p81, %f37, %f36; selp.f32 %f222, %f221, %f219, %p81; mov.f32 %f223, 0f40490FDB; sub.ftz.f32 %f224, %f223, %f222; selp.f32 %f225, %f224, %f222, %p80; mov.b32 %r55, %f225; or.b32 %r56, %r2, %r55; mov.b32 %f226, %r56; add.ftz.f32 %f227, %f36, %f37; setp.le.ftz.f32 %p82, %f227, 0f7F800000; selp.f32 %f354, %f226, %f227, %p82; bra.uni $L__BB67_151; $L__BB67_110: shr.u32 %r62, %r100, 23; mov.b32 %r4, %f44; shr.u32 %r63, %r4, 23; and.b32 %r64, %r100, 8388607; or.b32 %r99, %r64, 8388608; add.s32 %r65, %r62, 23; sub.s32 %r97, %r65, %r63; setp.eq.s32 %p91, %r97, 0; @%p91 bra $L__BB67_113; and.b32 %r66, %r4, 8388607; or.b32 %r7, %r66, 8388608; $L__BB67_112: min.s32 %r67, %r97, 8; shl.b32 %r68, %r99, %r67; rem.u32 %r99, %r68, %r7; sub.s32 %r97, %r97, %r67; setp.ne.s32 %p92, %r97, 0; @%p92 bra $L__BB67_112; $L__BB67_113: and.b32 %r69, %r4, -8388608; mov.b32 %f229, %r69; cvt.rn.f32.u32 %f230, %r99; mul.ftz.f32 %f231, %f230, 0f28800000; mul.ftz.f32 %f232, %f231, %f229; mov.b32 %r100, %f232; bra.uni $L__BB67_120; $L__BB67_128: setp.eq.ftz.f32 %p108, %f1, 0f00000000; mov.f32 %f354, %f87; @%p108 bra $L__BB67_151; lg2.approx.ftz.f32 %f252, %f1; mul.ftz.f32 %f253, %f2, %f252; ex2.approx.ftz.f32 %f354, %f253; bra.uni $L__BB67_151; $L__BB67_116: add.ftz.f32 %f351, %f351, 0f3F800000; add.ftz.f32 %f234, %f43, %f43; setp.ltu.ftz.f32 %p95, %f47, %f234; @%p95 bra $L__BB67_119; add.ftz.f32 %f235, %f351, 0f3F800000; mov.f32 %f236, 0fC0400000; fma.rn.ftz.f32 %f237, %f236, %f43, %f47; setp.ge.ftz.f32 %p96, %f237, 0f00000000; add.ftz.f32 %f238, %f235, 0f3F800000; selp.f32 %f351, %f238, %f235, %p96; $L__BB67_119: fma.rn.ftz.f32 %f241, %f46, %f351, %f42; mov.b32 %r100, %f241; $L__BB67_120: mov.b32 %r71, %f1; and.b32 %r72, %r71, -2147483648; or.b32 %r73, %r100, %r72; mov.b32 %f354, %r73; $L__BB67_151: mul.wide.u32 %rd13, %r31, 4; add.s64 %rd14, %rd2, %rd13; st.local.f32 [%rd14], %f354; ret; $L__BB67_132: shr.u32 %r75, %r104, 23; mov.b32 %r18, %f63; shr.u32 %r76, %r18, 23; and.b32 %r77, %r104, 8388607; or.b32 %r103, %r77, 8388608; add.s32 %r78, %r75, 23; sub.s32 %r101, %r78, %r76; setp.eq.s32 %p115, %r101, 0; @%p115 bra $L__BB67_135; and.b32 %r79, %r18, 8388607; or.b32 %r21, %r79, 8388608; $L__BB67_134: min.s32 %r80, %r101, 8; shl.b32 %r81, %r103, %r80; rem.u32 %r103, %r81, %r21; sub.s32 %r101, %r101, %r80; setp.ne.s32 %p116, %r101, 0; @%p116 bra $L__BB67_134; $L__BB67_135: and.b32 %r82, %r18, -8388608; mov.b32 %f256, %r82; cvt.rn.f32.u32 %f257, %r103; mul.ftz.f32 %f258, %f257, 0f28800000; mul.ftz.f32 %f259, %f258, %f256; mov.b32 %r104, %f259; bra.uni $L__BB67_142; $L__BB67_138: add.ftz.f32 %f352, %f352, 0f3F800000; add.ftz.f32 %f261, %f62, %f62; setp.ltu.ftz.f32 %p119, %f66, %f261; @%p119 bra $L__BB67_141; add.ftz.f32 %f262, %f352, 0f3F800000; mov.f32 %f263, 0fC0400000; fma.rn.ftz.f32 %f264, %f263, %f62, %f66; setp.ge.ftz.f32 %p120, %f264, 0f00000000; add.ftz.f32 %f265, %f262, 0f3F800000; selp.f32 %f352, %f265, %f262, %p120; $L__BB67_141: fma.rn.ftz.f32 %f268, %f65, %f352, %f61; mov.b32 %r104, %f268; $L__BB67_142: mov.b32 %r86, %f254; and.b32 %r87, %r86, -2147483648; or.b32 %r88, %r104, %r87; mov.b32 %f270, %r88; setp.eq.ftz.f32 %p122, %f270, 0f00000000; neg.ftz.f32 %f271, %f1; lg2.approx.ftz.f32 %f272, %f271; mul.ftz.f32 %f273, %f2, %f272; ex2.approx.ftz.f32 %f354, %f273; @%p122 bra $L__BB67_151; neg.ftz.f32 %f354, %f354; bra.uni $L__BB67_151; } .func (.param .b32 func_retval0) _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<97>; .reg .f32 %f<392>; .reg .b32 %r<108>; .reg .b64 %rd<22>; ld.param.u64 %rd2, [_Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r47, [_Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r50, [_Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r51, [_Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r95, [_Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; and.b32 %r1, %r51, 255; bfe.u32 %r2, %r51, 8, 8; shl.b32 %r52, %r50, 2; cvt.u64.u32 %rd3, %r52; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f1, [%rd5]; ld.local.f32 %f2, [%rd5+4]; ld.local.f32 %f3, [%rd5+8]; shr.u32 %r53, %r50, 6; cvt.u64.u32 %rd6, %r53; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f4, [%rd8]; ld.local.f32 %f5, [%rd8+4]; ld.local.f32 %f6, [%rd8+8]; shr.u32 %r54, %r50, 14; cvt.u64.u32 %rd9, %r54; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f7, [%rd11]; mov.f32 %f11, 0f00000000; setp.gt.u32 %p1, %r47, 26; mov.f32 %f389, %f11; mov.f32 %f390, %f11; @%p1 bra $L__BB68_3; mov.u32 %r55, 1; shl.b32 %r56, %r55, %r47; and.b32 %r57, %r56, 101711872; setp.ne.s32 %p2, %r57, 0; mov.f32 %f389, %f11; mov.f32 %f390, %f11; @%p2 bra $L__BB68_2; bra.uni $L__BB68_3; $L__BB68_2: ld.const.u64 %rd12, [kernel_params+2400]; cvta.to.global.u64 %rd13, %rd12; mul.wide.s32 %rd14, %r95, 16; add.s64 %rd15, %rd13, %rd14; add.s32 %r95, %r95, 1; ld.global.u32 %r58, [%rd15]; mul.wide.u32 %rd16, %r58, 4; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f390, [%rd17+8]; ld.local.f32 %f389, [%rd17+4]; ld.local.f32 %f11, [%rd17]; $L__BB68_3: mul.ftz.f32 %f391, %f1, %f4; mul.ftz.f32 %f18, %f2, %f5; mul.ftz.f32 %f19, %f3, %f6; setp.gt.s32 %p3, %r47, 12; @%p3 bra $L__BB68_33; bra.uni $L__BB68_4; $L__BB68_33: setp.gt.s32 %p4, %r47, 19; @%p4 bra $L__BB68_52; setp.gt.s32 %p14, %r47, 15; @%p14 bra $L__BB68_45; bra.uni $L__BB68_35; $L__BB68_45: setp.gt.s32 %p15, %r47, 17; @%p15 bra $L__BB68_49; setp.eq.s32 %p18, %r47, 16; @%p18 bra $L__BB68_78; setp.eq.s32 %p19, %r47, 17; @%p19 bra $L__BB68_48; bra.uni $L__BB68_60; $L__BB68_48: abs.ftz.f32 %f391, %f1; abs.ftz.f32 %f389, %f2; abs.ftz.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_4: setp.gt.s32 %p23, %r47, 5; @%p23 bra $L__BB68_15; bra.uni $L__BB68_5; $L__BB68_15: setp.gt.s32 %p24, %r47, 8; @%p24 bra $L__BB68_20; bra.uni $L__BB68_16; $L__BB68_20: setp.gt.s32 %p25, %r47, 10; @%p25 bra $L__BB68_24; setp.eq.s32 %p28, %r47, 9; @%p28 bra $L__BB68_119; setp.eq.s32 %p29, %r47, 10; @%p29 bra $L__BB68_23; bra.uni $L__BB68_60; $L__BB68_23: mul.ftz.f32 %f391, %f1, %f7; mul.ftz.f32 %f389, %f2, %f7; mul.ftz.f32 %f390, %f3, %f7; bra.uni $L__BB68_131; $L__BB68_52: setp.gt.s32 %p5, %r47, 22; @%p5 bra $L__BB68_57; bra.uni $L__BB68_53; $L__BB68_57: setp.gt.s32 %p6, %r47, 24; @%p6 bra $L__BB68_61; setp.eq.s32 %p9, %r47, 23; @%p9 bra $L__BB68_69; setp.eq.s32 %p10, %r47, 24; @%p10 bra $L__BB68_67; bra.uni $L__BB68_60; $L__BB68_67: mul.ftz.f32 %f190, %f5, %f5; fma.rn.ftz.f32 %f191, %f4, %f4, %f190; fma.rn.ftz.f32 %f192, %f6, %f6, %f191; sqrt.approx.ftz.f32 %f193, %f192; div.approx.ftz.f32 %f35, %f4, %f193; div.approx.ftz.f32 %f36, %f5, %f193; div.approx.ftz.f32 %f37, %f6, %f193; mul.ftz.f32 %f194, %f2, %f36; fma.rn.ftz.f32 %f195, %f1, %f35, %f194; fma.rn.ftz.f32 %f38, %f3, %f37, %f195; mul.ftz.f32 %f196, %f38, %f38; mov.f32 %f197, 0f3F800000; sub.ftz.f32 %f198, %f197, %f196; mul.ftz.f32 %f199, %f7, %f7; mul.ftz.f32 %f200, %f199, %f198; sub.ftz.f32 %f39, %f197, %f200; setp.lt.ftz.f32 %p41, %f39, 0f00000000; mov.f32 %f391, 0f00000000; mov.f32 %f389, %f391; mov.f32 %f390, %f391; @%p41 bra $L__BB68_131; mul.ftz.f32 %f201, %f1, %f7; sqrt.approx.ftz.f32 %f202, %f39; fma.rn.ftz.f32 %f203, %f7, %f38, %f202; mul.ftz.f32 %f204, %f35, %f203; mul.ftz.f32 %f205, %f36, %f203; mul.ftz.f32 %f206, %f37, %f203; mul.ftz.f32 %f207, %f2, %f7; mul.ftz.f32 %f208, %f3, %f7; sub.ftz.f32 %f390, %f208, %f206; sub.ftz.f32 %f389, %f207, %f205; sub.ftz.f32 %f391, %f201, %f204; bra.uni $L__BB68_131; $L__BB68_5: setp.gt.s32 %p33, %r47, 2; @%p33 bra $L__BB68_9; setp.eq.s32 %p37, %r47, 0; @%p37 bra $L__BB68_130; setp.eq.s32 %p38, %r47, 1; @%p38 bra $L__BB68_129; setp.eq.s32 %p39, %r47, 2; mov.f32 %f389, %f18; mov.f32 %f390, %f19; @%p39 bra $L__BB68_131; bra.uni $L__BB68_60; $L__BB68_35: setp.eq.s32 %p20, %r47, 13; @%p20 bra $L__BB68_115; setp.eq.s32 %p21, %r47, 14; @%p21 bra $L__BB68_114; setp.eq.s32 %p22, %r47, 15; @%p22 bra $L__BB68_38; bra.uni $L__BB68_60; $L__BB68_38: setp.eq.ftz.f32 %p45, %f4, 0f00000000; mov.f32 %f390, 0f00000000; mov.f32 %f391, %f390; @%p45 bra $L__BB68_86; abs.ftz.f32 %f78, %f1; abs.ftz.f32 %f79, %f4; mul.ftz.f32 %f80, %f79, 0f4B000000; setp.lt.ftz.f32 %p46, %f78, %f79; mov.b32 %r99, %f78; @%p46 bra $L__BB68_85; setp.le.ftz.f32 %p47, %f78, %f80; setp.leu.ftz.f32 %p48, %f79, 0f00000000; or.pred %p49, %p48, %p47; setp.gt.u32 %p50, %r99, 2139095039; or.pred %p51, %p50, %p49; @%p51 bra $L__BB68_79; bra.uni $L__BB68_41; $L__BB68_79: div.approx.ftz.f32 %f245, %f78, %f79; cvt.rzi.ftz.f32.f32 %f372, %f245; neg.ftz.f32 %f82, %f79; fma.rn.f32 %f83, %f82, %f372, %f78; mov.b32 %r16, %f83; mov.b32 %r67, %f79; setp.lt.u32 %p54, %r16, %r67; @%p54 bra $L__BB68_84; setp.gt.u32 %p55, %r16, -2147483648; @%p55 bra $L__BB68_83; bra.uni $L__BB68_81; $L__BB68_83: add.ftz.f32 %f251, %f372, 0fBF800000; add.ftz.f32 %f252, %f251, 0fBF800000; setp.lt.ftz.f32 %p58, %f83, %f82; selp.f32 %f372, %f252, %f251, %p58; bra.uni $L__BB68_84; $L__BB68_16: setp.eq.s32 %p30, %r47, 6; @%p30 bra $L__BB68_121; setp.eq.s32 %p31, %r47, 7; @%p31 bra $L__BB68_120; setp.eq.s32 %p32, %r47, 8; @%p32 bra $L__BB68_19; bra.uni $L__BB68_60; $L__BB68_19: sub.ftz.f32 %f304, %f1, %f4; sub.ftz.f32 %f305, %f2, %f5; mul.ftz.f32 %f306, %f305, %f305; fma.rn.ftz.f32 %f307, %f304, %f304, %f306; sub.ftz.f32 %f308, %f3, %f6; fma.rn.ftz.f32 %f309, %f308, %f308, %f307; sqrt.approx.ftz.f32 %f391, %f309; bra.uni $L__BB68_131; $L__BB68_53: setp.eq.s32 %p11, %r47, 20; @%p11 bra $L__BB68_71; setp.eq.s32 %p12, %r47, 21; @%p12 bra $L__BB68_70; setp.eq.s32 %p13, %r47, 22; @%p13 bra $L__BB68_56; bra.uni $L__BB68_60; $L__BB68_56: cos.approx.ftz.f32 %f391, %f1; cos.approx.ftz.f32 %f389, %f2; cos.approx.ftz.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_9: setp.eq.s32 %p34, %r47, 3; @%p34 bra $L__BB68_123; setp.eq.s32 %p35, %r47, 4; @%p35 bra $L__BB68_122; setp.eq.s32 %p36, %r47, 5; @%p36 bra $L__BB68_12; bra.uni $L__BB68_60; $L__BB68_12: mul.ftz.f32 %f332, %f5, %f5; fma.rn.ftz.f32 %f333, %f4, %f4, %f332; fma.rn.ftz.f32 %f150, %f6, %f6, %f333; setp.eq.ftz.f32 %p91, %f150, 0f00000000; mov.f32 %f383, 0f00000000; mov.f32 %f389, %f383; mov.f32 %f390, %f383; @%p91 bra $L__BB68_14; add.ftz.f32 %f334, %f391, %f18; add.ftz.f32 %f335, %f334, %f19; div.approx.ftz.f32 %f336, %f335, %f150; mul.ftz.f32 %f390, %f6, %f336; mul.ftz.f32 %f389, %f5, %f336; mul.ftz.f32 %f383, %f4, %f336; $L__BB68_14: mov.f32 %f391, %f383; bra.uni $L__BB68_131; $L__BB68_49: setp.eq.s32 %p16, %r47, 18; @%p16 bra $L__BB68_77; setp.eq.s32 %p17, %r47, 19; @%p17 bra $L__BB68_51; bra.uni $L__BB68_60; $L__BB68_51: max.ftz.f32 %f391, %f1, %f4; max.ftz.f32 %f389, %f2, %f5; max.ftz.f32 %f390, %f3, %f6; bra.uni $L__BB68_131; $L__BB68_24: setp.eq.s32 %p26, %r47, 11; @%p26 bra $L__BB68_116; setp.eq.s32 %p27, %r47, 12; @%p27 bra $L__BB68_26; bra.uni $L__BB68_60; $L__BB68_26: setp.eq.ftz.f32 %p87, %f4, 0f00000000; mov.f32 %f379, 0f00000000; mov.f32 %f377, %f379; @%p87 bra $L__BB68_28; div.approx.ftz.f32 %f377, %f1, %f4; $L__BB68_28: setp.eq.ftz.f32 %p88, %f5, 0f00000000; mov.f32 %f378, %f379; @%p88 bra $L__BB68_30; div.approx.ftz.f32 %f378, %f2, %f5; $L__BB68_30: setp.eq.ftz.f32 %p89, %f6, 0f00000000; @%p89 bra $L__BB68_32; div.approx.ftz.f32 %f379, %f3, %f6; $L__BB68_32: cvt.rmi.ftz.f32.f32 %f290, %f377; mul.ftz.f32 %f391, %f4, %f290; cvt.rmi.ftz.f32.f32 %f291, %f378; mul.ftz.f32 %f389, %f5, %f291; cvt.rmi.ftz.f32.f32 %f292, %f379; mul.ftz.f32 %f390, %f6, %f292; bra.uni $L__BB68_131; $L__BB68_61: setp.eq.s32 %p7, %r47, 25; @%p7 bra $L__BB68_64; setp.ne.s32 %p8, %r47, 26; @%p8 bra $L__BB68_60; add.ftz.f32 %f391, %f391, %f11; add.ftz.f32 %f389, %f18, %f389; add.ftz.f32 %f390, %f19, %f390; bra.uni $L__BB68_131; $L__BB68_78: cvt.rmi.ftz.f32.f32 %f237, %f1; sub.ftz.f32 %f391, %f1, %f237; cvt.rmi.ftz.f32.f32 %f238, %f2; sub.ftz.f32 %f389, %f2, %f238; cvt.rmi.ftz.f32.f32 %f239, %f3; sub.ftz.f32 %f390, %f3, %f239; bra.uni $L__BB68_131; $L__BB68_119: mul.ftz.f32 %f300, %f2, %f2; fma.rn.ftz.f32 %f301, %f1, %f1, %f300; fma.rn.ftz.f32 %f302, %f3, %f3, %f301; sqrt.approx.ftz.f32 %f391, %f302; bra.uni $L__BB68_131; $L__BB68_69: sin.approx.ftz.f32 %f211, %f1; cos.approx.ftz.f32 %f212, %f1; div.approx.ftz.f32 %f391, %f211, %f212; cos.approx.ftz.f32 %f213, %f2; sin.approx.ftz.f32 %f214, %f2; div.approx.ftz.f32 %f389, %f214, %f213; cos.approx.ftz.f32 %f215, %f3; sin.approx.ftz.f32 %f216, %f3; div.approx.ftz.f32 %f390, %f216, %f215; bra.uni $L__BB68_131; $L__BB68_130: add.ftz.f32 %f391, %f1, %f4; add.ftz.f32 %f389, %f2, %f5; add.ftz.f32 %f390, %f3, %f6; bra.uni $L__BB68_131; $L__BB68_129: sub.ftz.f32 %f391, %f1, %f4; sub.ftz.f32 %f389, %f2, %f5; sub.ftz.f32 %f390, %f3, %f6; bra.uni $L__BB68_131; $L__BB68_115: cvt.rmi.ftz.f32.f32 %f391, %f1; cvt.rmi.ftz.f32.f32 %f389, %f2; cvt.rmi.ftz.f32.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_114: cvt.rpi.ftz.f32.f32 %f391, %f1; cvt.rpi.ftz.f32.f32 %f389, %f2; cvt.rpi.ftz.f32.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_121: mul.ftz.f32 %f313, %f5, %f5; fma.rn.ftz.f32 %f314, %f4, %f4, %f313; fma.rn.ftz.f32 %f315, %f6, %f6, %f314; sqrt.approx.ftz.f32 %f316, %f315; div.approx.ftz.f32 %f317, %f4, %f316; div.approx.ftz.f32 %f318, %f5, %f316; div.approx.ftz.f32 %f319, %f6, %f316; add.ftz.f32 %f320, %f317, %f317; add.ftz.f32 %f321, %f318, %f318; add.ftz.f32 %f322, %f319, %f319; mul.ftz.f32 %f323, %f2, %f318; fma.rn.ftz.f32 %f324, %f1, %f317, %f323; fma.rn.ftz.f32 %f325, %f3, %f319, %f324; mul.ftz.f32 %f326, %f320, %f325; mul.ftz.f32 %f327, %f321, %f325; mul.ftz.f32 %f328, %f322, %f325; sub.ftz.f32 %f391, %f1, %f326; sub.ftz.f32 %f389, %f2, %f327; sub.ftz.f32 %f390, %f3, %f328; bra.uni $L__BB68_131; $L__BB68_120: add.ftz.f32 %f311, %f391, %f18; add.ftz.f32 %f391, %f311, %f19; bra.uni $L__BB68_131; $L__BB68_71: sub.ftz.f32 %f58, %f4, %f11; setp.eq.ftz.f32 %p42, %f58, 0f00000000; mov.f32 %f391, %f11; @%p42 bra $L__BB68_73; sub.ftz.f32 %f219, %f1, %f11; div.approx.ftz.f32 %f220, %f219, %f58; cvt.rmi.ftz.f32.f32 %f221, %f220; mul.ftz.f32 %f222, %f58, %f221; sub.ftz.f32 %f391, %f1, %f222; $L__BB68_73: sub.ftz.f32 %f61, %f5, %f389; setp.eq.ftz.f32 %p43, %f61, 0f00000000; @%p43 bra $L__BB68_75; sub.ftz.f32 %f223, %f2, %f389; div.approx.ftz.f32 %f224, %f223, %f61; cvt.rmi.ftz.f32.f32 %f225, %f224; mul.ftz.f32 %f226, %f61, %f225; sub.ftz.f32 %f389, %f2, %f226; $L__BB68_75: sub.ftz.f32 %f64, %f6, %f390; setp.eq.ftz.f32 %p44, %f64, 0f00000000; @%p44 bra $L__BB68_131; sub.ftz.f32 %f229, %f3, %f390; div.approx.ftz.f32 %f230, %f229, %f64; cvt.rmi.ftz.f32.f32 %f231, %f230; mul.ftz.f32 %f232, %f64, %f231; sub.ftz.f32 %f390, %f3, %f232; bra.uni $L__BB68_131; $L__BB68_70: sin.approx.ftz.f32 %f391, %f1; sin.approx.ftz.f32 %f389, %f2; sin.approx.ftz.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_123: setp.eq.ftz.f32 %p92, %f4, 0f00000000; mov.f32 %f390, 0f00000000; mov.f32 %f391, %f390; @%p92 bra $L__BB68_125; div.approx.ftz.f32 %f391, %f1, %f4; $L__BB68_125: setp.eq.ftz.f32 %p93, %f5, 0f00000000; mov.f32 %f389, %f390; @%p93 bra $L__BB68_127; div.approx.ftz.f32 %f389, %f2, %f5; $L__BB68_127: setp.eq.ftz.f32 %p94, %f6, 0f00000000; @%p94 bra $L__BB68_131; div.approx.ftz.f32 %f390, %f3, %f6; bra.uni $L__BB68_131; $L__BB68_122: mul.ftz.f32 %f339, %f2, %f6; mul.ftz.f32 %f340, %f3, %f5; sub.ftz.f32 %f391, %f339, %f340; mul.ftz.f32 %f341, %f1, %f6; mul.ftz.f32 %f342, %f3, %f4; sub.ftz.f32 %f389, %f342, %f341; mul.ftz.f32 %f343, %f2, %f4; mul.ftz.f32 %f344, %f1, %f5; sub.ftz.f32 %f390, %f344, %f343; bra.uni $L__BB68_131; $L__BB68_77: min.ftz.f32 %f391, %f1, %f4; min.ftz.f32 %f389, %f2, %f5; min.ftz.f32 %f390, %f3, %f6; bra.uni $L__BB68_131; $L__BB68_116: mul.ftz.f32 %f293, %f2, %f2; fma.rn.ftz.f32 %f294, %f1, %f1, %f293; fma.rn.ftz.f32 %f295, %f3, %f3, %f294; sqrt.approx.ftz.f32 %f128, %f295; setp.neu.ftz.f32 %p90, %f128, 0f00000000; @%p90 bra $L__BB68_117; bra.uni $L__BB68_118; $L__BB68_117: rcp.approx.ftz.f32 %f296, %f128; mul.ftz.f32 %f3, %f3, %f296; mul.ftz.f32 %f2, %f2, %f296; mul.ftz.f32 %f1, %f1, %f296; $L__BB68_118: mov.f32 %f391, %f1; mov.f32 %f389, %f2; mov.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_64: mul.ftz.f32 %f183, %f5, %f389; fma.rn.ftz.f32 %f184, %f4, %f11, %f183; fma.rn.ftz.f32 %f185, %f6, %f390, %f184; setp.lt.ftz.f32 %p40, %f185, 0f00000000; @%p40 bra $L__BB68_66; bra.uni $L__BB68_65; $L__BB68_66: mov.f32 %f391, %f1; mov.f32 %f389, %f2; mov.f32 %f390, %f3; bra.uni $L__BB68_131; $L__BB68_60: mov.f32 %f391, 0f00000000; mov.f32 %f389, %f391; mov.f32 %f390, %f391; bra.uni $L__BB68_131; $L__BB68_65: neg.ftz.f32 %f390, %f3; neg.ftz.f32 %f389, %f2; neg.ftz.f32 %f391, %f1; bra.uni $L__BB68_131; $L__BB68_41: shr.u32 %r59, %r99, 23; mov.b32 %r6, %f80; shr.u32 %r60, %r6, 23; and.b32 %r61, %r99, 8388607; or.b32 %r98, %r61, 8388608; add.s32 %r62, %r59, 23; sub.s32 %r96, %r62, %r60; setp.eq.s32 %p52, %r96, 0; @%p52 bra $L__BB68_44; and.b32 %r63, %r6, 8388607; or.b32 %r9, %r63, 8388608; $L__BB68_43: min.s32 %r64, %r96, 8; shl.b32 %r65, %r98, %r64; rem.u32 %r98, %r65, %r9; sub.s32 %r96, %r96, %r64; setp.ne.s32 %p53, %r96, 0; @%p53 bra $L__BB68_43; $L__BB68_44: and.b32 %r66, %r6, -8388608; mov.b32 %f241, %r66; cvt.rn.f32.u32 %f242, %r98; mul.ftz.f32 %f243, %f242, 0f28800000; mul.ftz.f32 %f244, %f243, %f241; mov.b32 %r99, %f244; bra.uni $L__BB68_85; $L__BB68_81: add.ftz.f32 %f372, %f372, 0f3F800000; add.ftz.f32 %f246, %f79, %f79; setp.ltu.ftz.f32 %p56, %f83, %f246; @%p56 bra $L__BB68_84; add.ftz.f32 %f247, %f372, 0f3F800000; mov.f32 %f248, 0fC0400000; fma.rn.ftz.f32 %f249, %f248, %f79, %f83; setp.ge.ftz.f32 %p57, %f249, 0f00000000; add.ftz.f32 %f250, %f247, 0f3F800000; selp.f32 %f372, %f250, %f247, %p57; $L__BB68_84: fma.rn.ftz.f32 %f253, %f82, %f372, %f78; mov.b32 %r99, %f253; $L__BB68_85: mov.b32 %r68, %f1; and.b32 %r69, %r68, -2147483648; or.b32 %r70, %r99, %r69; mov.b32 %f391, %r70; $L__BB68_86: setp.eq.ftz.f32 %p59, %f5, 0f00000000; mov.f32 %f389, %f390; @%p59 bra $L__BB68_100; abs.ftz.f32 %f90, %f2; abs.ftz.f32 %f91, %f5; mul.ftz.f32 %f92, %f91, 0f4B000000; setp.lt.ftz.f32 %p60, %f90, %f91; mov.b32 %r103, %f90; @%p60 bra $L__BB68_99; setp.le.ftz.f32 %p61, %f90, %f92; setp.leu.ftz.f32 %p62, %f91, 0f00000000; or.pred %p63, %p62, %p61; setp.gt.u32 %p64, %r103, 2139095039; or.pred %p65, %p64, %p63; @%p65 bra $L__BB68_93; bra.uni $L__BB68_89; $L__BB68_93: div.approx.ftz.f32 %f259, %f90, %f91; cvt.rzi.ftz.f32.f32 %f374, %f259; neg.ftz.f32 %f94, %f91; fma.rn.f32 %f95, %f94, %f374, %f90; mov.b32 %r30, %f95; mov.b32 %r79, %f91; setp.lt.u32 %p68, %r30, %r79; @%p68 bra $L__BB68_98; setp.gt.u32 %p69, %r30, -2147483648; @%p69 bra $L__BB68_97; bra.uni $L__BB68_95; $L__BB68_97: add.ftz.f32 %f265, %f374, 0fBF800000; add.ftz.f32 %f266, %f265, 0fBF800000; setp.lt.ftz.f32 %p72, %f95, %f94; selp.f32 %f374, %f266, %f265, %p72; bra.uni $L__BB68_98; $L__BB68_89: shr.u32 %r71, %r103, 23; mov.b32 %r20, %f92; shr.u32 %r72, %r20, 23; and.b32 %r73, %r103, 8388607; or.b32 %r102, %r73, 8388608; add.s32 %r74, %r71, 23; sub.s32 %r100, %r74, %r72; setp.eq.s32 %p66, %r100, 0; @%p66 bra $L__BB68_92; and.b32 %r75, %r20, 8388607; or.b32 %r23, %r75, 8388608; $L__BB68_91: min.s32 %r76, %r100, 8; shl.b32 %r77, %r102, %r76; rem.u32 %r102, %r77, %r23; sub.s32 %r100, %r100, %r76; setp.ne.s32 %p67, %r100, 0; @%p67 bra $L__BB68_91; $L__BB68_92: and.b32 %r78, %r20, -8388608; mov.b32 %f255, %r78; cvt.rn.f32.u32 %f256, %r102; mul.ftz.f32 %f257, %f256, 0f28800000; mul.ftz.f32 %f258, %f257, %f255; mov.b32 %r103, %f258; bra.uni $L__BB68_99; $L__BB68_95: add.ftz.f32 %f374, %f374, 0f3F800000; add.ftz.f32 %f260, %f91, %f91; setp.ltu.ftz.f32 %p70, %f95, %f260; @%p70 bra $L__BB68_98; add.ftz.f32 %f261, %f374, 0f3F800000; mov.f32 %f262, 0fC0400000; fma.rn.ftz.f32 %f263, %f262, %f91, %f95; setp.ge.ftz.f32 %p71, %f263, 0f00000000; add.ftz.f32 %f264, %f261, 0f3F800000; selp.f32 %f374, %f264, %f261, %p71; $L__BB68_98: fma.rn.ftz.f32 %f267, %f94, %f374, %f90; mov.b32 %r103, %f267; $L__BB68_99: mov.b32 %r80, %f2; and.b32 %r81, %r80, -2147483648; or.b32 %r82, %r103, %r81; mov.b32 %f389, %r82; $L__BB68_100: setp.eq.ftz.f32 %p73, %f6, 0f00000000; @%p73 bra $L__BB68_131; abs.ftz.f32 %f102, %f3; abs.ftz.f32 %f103, %f6; mul.ftz.f32 %f104, %f103, 0f4B000000; setp.lt.ftz.f32 %p74, %f102, %f103; mov.b32 %r107, %f102; @%p74 bra $L__BB68_113; setp.le.ftz.f32 %p75, %f102, %f104; setp.leu.ftz.f32 %p76, %f103, 0f00000000; or.pred %p77, %p76, %p75; setp.gt.u32 %p78, %r107, 2139095039; or.pred %p79, %p78, %p77; @%p79 bra $L__BB68_107; bra.uni $L__BB68_103; $L__BB68_107: div.approx.ftz.f32 %f274, %f102, %f103; cvt.rzi.ftz.f32.f32 %f376, %f274; neg.ftz.f32 %f106, %f103; fma.rn.f32 %f107, %f106, %f376, %f102; mov.b32 %r44, %f107; mov.b32 %r91, %f103; setp.lt.u32 %p82, %r44, %r91; @%p82 bra $L__BB68_112; setp.gt.u32 %p83, %r44, -2147483648; @%p83 bra $L__BB68_111; bra.uni $L__BB68_109; $L__BB68_111: add.ftz.f32 %f280, %f376, 0fBF800000; add.ftz.f32 %f281, %f280, 0fBF800000; setp.lt.ftz.f32 %p86, %f107, %f106; selp.f32 %f376, %f281, %f280, %p86; bra.uni $L__BB68_112; $L__BB68_103: shr.u32 %r83, %r107, 23; mov.b32 %r34, %f104; shr.u32 %r84, %r34, 23; and.b32 %r85, %r107, 8388607; or.b32 %r106, %r85, 8388608; add.s32 %r86, %r83, 23; sub.s32 %r104, %r86, %r84; setp.eq.s32 %p80, %r104, 0; @%p80 bra $L__BB68_106; and.b32 %r87, %r34, 8388607; or.b32 %r37, %r87, 8388608; $L__BB68_105: min.s32 %r88, %r104, 8; shl.b32 %r89, %r106, %r88; rem.u32 %r106, %r89, %r37; sub.s32 %r104, %r104, %r88; setp.ne.s32 %p81, %r104, 0; @%p81 bra $L__BB68_105; $L__BB68_106: and.b32 %r90, %r34, -8388608; mov.b32 %f270, %r90; cvt.rn.f32.u32 %f271, %r106; mul.ftz.f32 %f272, %f271, 0f28800000; mul.ftz.f32 %f273, %f272, %f270; mov.b32 %r107, %f273; bra.uni $L__BB68_113; $L__BB68_109: add.ftz.f32 %f376, %f376, 0f3F800000; add.ftz.f32 %f275, %f103, %f103; setp.ltu.ftz.f32 %p84, %f107, %f275; @%p84 bra $L__BB68_112; add.ftz.f32 %f276, %f376, 0f3F800000; mov.f32 %f277, 0fC0400000; fma.rn.ftz.f32 %f278, %f277, %f103, %f107; setp.ge.ftz.f32 %p85, %f278, 0f00000000; add.ftz.f32 %f279, %f276, 0f3F800000; selp.f32 %f376, %f279, %f276, %p85; $L__BB68_112: fma.rn.ftz.f32 %f282, %f106, %f376, %f102; mov.b32 %r107, %f282; $L__BB68_113: mov.b32 %r92, %f3; and.b32 %r93, %r92, -2147483648; or.b32 %r94, %r107, %r93; mov.b32 %f390, %r94; $L__BB68_131: setp.eq.s32 %p95, %r1, 255; @%p95 bra $L__BB68_133; mul.wide.u32 %rd18, %r1, 4; add.s64 %rd19, %rd1, %rd18; st.local.f32 [%rd19], %f391; $L__BB68_133: setp.eq.s32 %p96, %r2, 255; @%p96 bra $L__BB68_135; mul.wide.u32 %rd20, %r2, 4; add.s64 %rd21, %rd1, %rd20; st.local.f32 [%rd21], %f391; st.local.f32 [%rd21+4], %f389; st.local.f32 [%rd21+8], %f390; $L__BB68_135: st.param.b32 [func_retval0+0], %r95; ret; } .func (.param .b32 func_retval0) _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<107>; .reg .f32 %f<577>; .reg .b32 %r<15>; .reg .b64 %rd<16>; ld.param.u64 %rd1, [_Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r7, [_Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r8, [_Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r9, [_Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r5, [_Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r5, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r10, %r11, %r12, %r13}, [%rd6]; mul.wide.u32 %rd7, %r7, 4; add.s64 %rd8, %rd2, %rd7; mul.wide.u32 %rd9, %r8, 4; add.s64 %rd10, %rd2, %rd9; ld.local.f32 %f1, [%rd10]; ld.local.f32 %f2, [%rd10+4]; ld.local.f32 %f3, [%rd10+8]; mul.wide.u32 %rd11, %r9, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f7, [%rd12]; ld.local.f32 %f8, [%rd12+4]; ld.local.f32 %f9, [%rd12+8]; ld.local.f32 %f283, [%rd8]; cvt.ftz.sat.f32.f32 %f10, %f283; mov.f32 %f574, 0f00000000; setp.gt.s32 %p3, %r11, 8; @%p3 bra $L__BB69_23; bra.uni $L__BB69_1; $L__BB69_23: setp.gt.s32 %p4, %r11, 13; @%p4 bra $L__BB69_45; setp.gt.s32 %p11, %r11, 10; @%p11 bra $L__BB69_36; bra.uni $L__BB69_25; $L__BB69_36: setp.eq.s32 %p12, %r11, 11; @%p12 bra $L__BB69_124; setp.eq.s32 %p13, %r11, 12; @%p13 bra $L__BB69_103; setp.eq.s32 %p14, %r11, 13; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p14 bra $L__BB69_39; bra.uni $L__BB69_147; $L__BB69_39: max.ftz.f32 %f369, %f2, %f3; max.ftz.f32 %f107, %f1, %f369; min.ftz.f32 %f370, %f2, %f3; min.ftz.f32 %f371, %f1, %f370; sub.ftz.f32 %f108, %f107, %f371; setp.eq.ftz.f32 %p53, %f107, 0f00000000; mov.f32 %f552, 0f00000000; mov.f32 %f550, %f552; @%p53 bra $L__BB69_41; div.approx.ftz.f32 %f550, %f108, %f107; $L__BB69_41: setp.eq.ftz.f32 %p54, %f550, 0f00000000; @%p54 bra $L__BB69_92; sub.ftz.f32 %f373, %f107, %f1; div.approx.ftz.f32 %f111, %f373, %f108; sub.ftz.f32 %f374, %f107, %f2; div.approx.ftz.f32 %f112, %f374, %f108; sub.ftz.f32 %f375, %f107, %f3; div.approx.ftz.f32 %f113, %f375, %f108; setp.eq.ftz.f32 %p55, %f1, %f107; @%p55 bra $L__BB69_90; bra.uni $L__BB69_43; $L__BB69_90: sub.ftz.f32 %f551, %f113, %f112; bra.uni $L__BB69_91; $L__BB69_1: setp.gt.s32 %p17, %r11, 3; @%p17 bra $L__BB69_9; bra.uni $L__BB69_2; $L__BB69_9: setp.gt.s32 %p18, %r11, 5; @%p18 bra $L__BB69_19; bra.uni $L__BB69_10; $L__BB69_19: setp.eq.s32 %p19, %r11, 6; @%p19 bra $L__BB69_143; setp.eq.s32 %p20, %r11, 7; @%p20 bra $L__BB69_142; setp.eq.s32 %p21, %r11, 8; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p21 bra $L__BB69_22; bra.uni $L__BB69_147; $L__BB69_22: max.ftz.f32 %f473, %f1, %f7; sub.ftz.f32 %f474, %f473, %f1; max.ftz.f32 %f475, %f2, %f8; sub.ftz.f32 %f476, %f475, %f2; max.ftz.f32 %f477, %f3, %f9; sub.ftz.f32 %f478, %f477, %f3; fma.rn.ftz.f32 %f576, %f10, %f478, %f3; fma.rn.ftz.f32 %f575, %f10, %f476, %f2; fma.rn.ftz.f32 %f574, %f10, %f474, %f1; bra.uni $L__BB69_147; $L__BB69_45: setp.gt.s32 %p5, %r11, 15; @%p5 bra $L__BB69_54; bra.uni $L__BB69_46; $L__BB69_54: setp.eq.s32 %p6, %r11, 16; @%p6 bra $L__BB69_59; setp.eq.s32 %p7, %r11, 17; @%p7 bra $L__BB69_58; setp.ne.s32 %p8, %r11, 18; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p8 bra $L__BB69_147; cvt.ftz.sat.f32.f32 %f576, %f3; cvt.ftz.sat.f32.f32 %f575, %f2; cvt.ftz.sat.f32.f32 %f574, %f1; bra.uni $L__BB69_147; $L__BB69_2: setp.gt.s32 %p24, %r11, 1; @%p24 bra $L__BB69_6; setp.eq.s32 %p27, %r11, 0; @%p27 bra $L__BB69_146; setp.eq.s32 %p28, %r11, 1; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p28 bra $L__BB69_5; bra.uni $L__BB69_147; $L__BB69_5: add.ftz.f32 %f527, %f1, %f7; sub.ftz.f32 %f528, %f527, %f1; add.ftz.f32 %f529, %f2, %f8; sub.ftz.f32 %f530, %f529, %f2; add.ftz.f32 %f531, %f3, %f9; sub.ftz.f32 %f532, %f531, %f3; fma.rn.ftz.f32 %f576, %f10, %f532, %f3; fma.rn.ftz.f32 %f575, %f530, %f10, %f2; fma.rn.ftz.f32 %f574, %f528, %f10, %f1; bra.uni $L__BB69_147; $L__BB69_25: setp.eq.s32 %p15, %r11, 9; @%p15 bra $L__BB69_133; setp.eq.s32 %p16, %r11, 10; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p16 bra $L__BB69_27; bra.uni $L__BB69_147; $L__BB69_27: setp.eq.ftz.f32 %p91, %f1, 0f00000000; mov.f32 %f574, %f1; @%p91 bra $L__BB69_30; mul.ftz.f32 %f446, %f7, %f10; mov.f32 %f574, 0f3F800000; sub.ftz.f32 %f217, %f574, %f446; setp.le.ftz.f32 %p92, %f217, 0f00000000; @%p92 bra $L__BB69_30; div.approx.ftz.f32 %f447, %f1, %f217; setp.gt.ftz.f32 %p93, %f447, 0f3F800000; selp.f32 %f574, 0f3F800000, %f447, %p93; $L__BB69_30: setp.eq.ftz.f32 %p94, %f2, 0f00000000; mov.f32 %f575, %f2; @%p94 bra $L__BB69_33; mul.ftz.f32 %f449, %f8, %f10; mov.f32 %f575, 0f3F800000; sub.ftz.f32 %f220, %f575, %f449; setp.le.ftz.f32 %p95, %f220, 0f00000000; @%p95 bra $L__BB69_33; div.approx.ftz.f32 %f450, %f2, %f220; setp.gt.ftz.f32 %p96, %f450, 0f3F800000; selp.f32 %f575, 0f3F800000, %f450, %p96; $L__BB69_33: setp.eq.ftz.f32 %p97, %f3, 0f00000000; mov.f32 %f576, %f3; @%p97 bra $L__BB69_147; mul.ftz.f32 %f452, %f9, %f10; mov.f32 %f576, 0f3F800000; sub.ftz.f32 %f223, %f576, %f452; setp.le.ftz.f32 %p98, %f223, 0f00000000; @%p98 bra $L__BB69_147; div.approx.ftz.f32 %f453, %f3, %f223; setp.gt.ftz.f32 %p99, %f453, 0f3F800000; selp.f32 %f576, 0f3F800000, %f453, %p99; bra.uni $L__BB69_147; $L__BB69_10: setp.eq.s32 %p22, %r11, 4; @%p22 bra $L__BB69_144; setp.eq.s32 %p23, %r11, 5; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p23 bra $L__BB69_12; bra.uni $L__BB69_147; $L__BB69_12: mov.f32 %f494, 0f3F800000; sub.ftz.f32 %f252, %f494, %f10; setp.eq.ftz.f32 %p103, %f7, 0f00000000; @%p103 bra $L__BB69_14; mul.ftz.f32 %f495, %f1, %f10; div.approx.ftz.f32 %f496, %f495, %f7; fma.rn.ftz.f32 %f1, %f1, %f252, %f496; $L__BB69_14: setp.eq.ftz.f32 %p104, %f8, 0f00000000; @%p104 bra $L__BB69_16; mul.ftz.f32 %f497, %f2, %f10; div.approx.ftz.f32 %f498, %f497, %f8; fma.rn.ftz.f32 %f2, %f2, %f252, %f498; $L__BB69_16: setp.eq.ftz.f32 %p105, %f9, 0f00000000; @%p105 bra $L__BB69_18; mul.ftz.f32 %f499, %f3, %f10; div.approx.ftz.f32 %f500, %f499, %f9; fma.rn.ftz.f32 %f3, %f3, %f252, %f500; $L__BB69_18: mov.f32 %f574, %f1; mov.f32 %f575, %f2; mov.f32 %f576, %f3; bra.uni $L__BB69_147; $L__BB69_46: setp.eq.s32 %p9, %r11, 14; @%p9 bra $L__BB69_72; setp.eq.s32 %p10, %r11, 15; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p10 bra $L__BB69_48; bra.uni $L__BB69_147; $L__BB69_48: max.ftz.f32 %f314, %f8, %f9; max.ftz.f32 %f20, %f7, %f314; min.ftz.f32 %f315, %f8, %f9; min.ftz.f32 %f316, %f7, %f315; sub.ftz.f32 %f21, %f20, %f316; setp.eq.ftz.f32 %p29, %f20, 0f00000000; mov.f32 %f540, 0f00000000; mov.f32 %f538, %f540; @%p29 bra $L__BB69_50; div.approx.ftz.f32 %f538, %f21, %f20; $L__BB69_50: setp.eq.ftz.f32 %p30, %f538, 0f00000000; @%p30 bra $L__BB69_63; sub.ftz.f32 %f318, %f20, %f7; div.approx.ftz.f32 %f24, %f318, %f21; sub.ftz.f32 %f319, %f20, %f8; div.approx.ftz.f32 %f25, %f319, %f21; sub.ftz.f32 %f320, %f20, %f9; div.approx.ftz.f32 %f26, %f320, %f21; setp.eq.ftz.f32 %p31, %f7, %f20; @%p31 bra $L__BB69_61; bra.uni $L__BB69_52; $L__BB69_61: sub.ftz.f32 %f539, %f26, %f25; bra.uni $L__BB69_62; $L__BB69_6: setp.eq.s32 %p25, %r11, 2; @%p25 bra $L__BB69_145; setp.eq.s32 %p26, %r11, 3; mov.f32 %f575, %f574; mov.f32 %f576, %f574; @%p26 bra $L__BB69_8; bra.uni $L__BB69_147; $L__BB69_8: sub.ftz.f32 %f515, %f1, %f7; sub.ftz.f32 %f516, %f515, %f1; sub.ftz.f32 %f517, %f2, %f8; sub.ftz.f32 %f518, %f517, %f2; sub.ftz.f32 %f519, %f3, %f9; sub.ftz.f32 %f520, %f519, %f3; fma.rn.ftz.f32 %f576, %f10, %f520, %f3; fma.rn.ftz.f32 %f575, %f518, %f10, %f2; fma.rn.ftz.f32 %f574, %f516, %f10, %f1; bra.uni $L__BB69_147; $L__BB69_124: mov.f32 %f430, 0f3F800000; sub.ftz.f32 %f201, %f430, %f10; fma.rn.ftz.f32 %f202, %f7, %f10, %f201; setp.le.ftz.f32 %p82, %f202, 0f00000000; mov.f32 %f576, 0f00000000; mov.f32 %f574, %f576; @%p82 bra $L__BB69_127; sub.ftz.f32 %f433, %f430, %f1; div.approx.ftz.f32 %f434, %f433, %f202; sub.ftz.f32 %f203, %f430, %f434; setp.lt.ftz.f32 %p83, %f203, 0f00000000; mov.f32 %f574, %f576; @%p83 bra $L__BB69_127; setp.gt.ftz.f32 %p84, %f203, 0f3F800000; selp.f32 %f574, 0f3F800000, %f203, %p84; $L__BB69_127: fma.rn.ftz.f32 %f206, %f8, %f10, %f201; setp.le.ftz.f32 %p85, %f206, 0f00000000; mov.f32 %f575, %f576; @%p85 bra $L__BB69_130; sub.ftz.f32 %f438, %f430, %f2; div.approx.ftz.f32 %f439, %f438, %f206; sub.ftz.f32 %f207, %f430, %f439; setp.lt.ftz.f32 %p86, %f207, 0f00000000; mov.f32 %f575, %f576; @%p86 bra $L__BB69_130; setp.gt.ftz.f32 %p87, %f207, 0f3F800000; selp.f32 %f575, 0f3F800000, %f207, %p87; $L__BB69_130: fma.rn.ftz.f32 %f210, %f9, %f10, %f201; setp.le.ftz.f32 %p88, %f210, 0f00000000; @%p88 bra $L__BB69_147; sub.ftz.f32 %f443, %f430, %f3; div.approx.ftz.f32 %f444, %f443, %f210; sub.ftz.f32 %f211, %f430, %f444; setp.lt.ftz.f32 %p89, %f211, 0f00000000; @%p89 bra $L__BB69_147; setp.gt.ftz.f32 %p90, %f211, 0f3F800000; selp.f32 %f576, 0f3F800000, %f211, %p90; bra.uni $L__BB69_147; $L__BB69_103: max.ftz.f32 %f398, %f8, %f9; max.ftz.f32 %f156, %f7, %f398; min.ftz.f32 %f399, %f8, %f9; min.ftz.f32 %f400, %f7, %f399; sub.ftz.f32 %f157, %f156, %f400; setp.eq.ftz.f32 %p68, %f156, 0f00000000; mov.pred %p106, 0; @%p68 bra $L__BB69_105; div.approx.ftz.f32 %f401, %f157, %f156; setp.neu.ftz.f32 %p106, %f401, 0f00000000; $L__BB69_105: mov.f32 %f558, 0f00000000; not.pred %p69, %p106; @%p69 bra $L__BB69_112; sub.ftz.f32 %f403, %f156, %f7; div.approx.ftz.f32 %f158, %f403, %f157; sub.ftz.f32 %f404, %f156, %f8; div.approx.ftz.f32 %f159, %f404, %f157; sub.ftz.f32 %f405, %f156, %f9; div.approx.ftz.f32 %f160, %f405, %f157; setp.eq.ftz.f32 %p70, %f7, %f156; @%p70 bra $L__BB69_110; bra.uni $L__BB69_107; $L__BB69_110: sub.ftz.f32 %f557, %f160, %f159; bra.uni $L__BB69_111; $L__BB69_143: sub.ftz.f32 %f485, %f1, %f7; abs.ftz.f32 %f486, %f485; sub.ftz.f32 %f487, %f2, %f8; abs.ftz.f32 %f488, %f487; sub.ftz.f32 %f489, %f3, %f9; abs.ftz.f32 %f490, %f489; sub.ftz.f32 %f491, %f486, %f1; sub.ftz.f32 %f492, %f488, %f2; sub.ftz.f32 %f493, %f490, %f3; fma.rn.ftz.f32 %f576, %f10, %f493, %f3; fma.rn.ftz.f32 %f575, %f10, %f492, %f2; fma.rn.ftz.f32 %f574, %f10, %f491, %f1; bra.uni $L__BB69_147; $L__BB69_142: min.ftz.f32 %f479, %f1, %f7; sub.ftz.f32 %f480, %f479, %f1; min.ftz.f32 %f481, %f2, %f8; sub.ftz.f32 %f482, %f481, %f2; min.ftz.f32 %f483, %f3, %f9; sub.ftz.f32 %f484, %f483, %f3; fma.rn.ftz.f32 %f576, %f10, %f484, %f3; fma.rn.ftz.f32 %f575, %f10, %f482, %f2; fma.rn.ftz.f32 %f574, %f10, %f480, %f1; bra.uni $L__BB69_147; $L__BB69_59: mov.f32 %f287, 0f3F800000; sub.ftz.f32 %f288, %f287, %f10; sub.ftz.f32 %f289, %f287, %f1; sub.ftz.f32 %f290, %f287, %f7; mul.ftz.f32 %f291, %f289, %f290; sub.ftz.f32 %f292, %f287, %f2; sub.ftz.f32 %f293, %f287, %f8; mul.ftz.f32 %f294, %f292, %f293; sub.ftz.f32 %f295, %f287, %f3; sub.ftz.f32 %f296, %f287, %f9; mul.ftz.f32 %f297, %f295, %f296; sub.ftz.f32 %f298, %f287, %f291; sub.ftz.f32 %f299, %f287, %f294; sub.ftz.f32 %f300, %f287, %f297; mul.ftz.f32 %f301, %f289, %f7; mul.ftz.f32 %f302, %f292, %f8; mul.ftz.f32 %f303, %f295, %f9; mul.ftz.f32 %f304, %f1, %f298; mul.ftz.f32 %f305, %f2, %f299; mul.ftz.f32 %f306, %f3, %f300; fma.rn.ftz.f32 %f307, %f1, %f301, %f304; fma.rn.ftz.f32 %f308, %f2, %f302, %f305; fma.rn.ftz.f32 %f309, %f3, %f303, %f306; mul.ftz.f32 %f310, %f10, %f307; mul.ftz.f32 %f311, %f10, %f308; mul.ftz.f32 %f312, %f10, %f309; fma.rn.ftz.f32 %f576, %f3, %f288, %f312; fma.rn.ftz.f32 %f575, %f2, %f288, %f311; fma.rn.ftz.f32 %f574, %f1, %f288, %f310; bra.uni $L__BB69_147; $L__BB69_58: fma.rn.ftz.f32 %f284, %f7, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f285, %f8, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f286, %f9, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f576, %f10, %f286, %f3; fma.rn.ftz.f32 %f575, %f285, %f10, %f2; fma.rn.ftz.f32 %f574, %f284, %f10, %f1; bra.uni $L__BB69_147; $L__BB69_146: sub.ftz.f32 %f533, %f7, %f1; sub.ftz.f32 %f534, %f8, %f2; sub.ftz.f32 %f535, %f9, %f3; fma.rn.ftz.f32 %f576, %f535, %f10, %f3; fma.rn.ftz.f32 %f575, %f534, %f10, %f2; fma.rn.ftz.f32 %f574, %f533, %f10, %f1; bra.uni $L__BB69_147; $L__BB69_133: mov.f32 %f454, 0f3F800000; sub.ftz.f32 %f229, %f454, %f10; add.ftz.f32 %f230, %f10, %f10; setp.lt.ftz.f32 %p100, %f1, 0f3F000000; @%p100 bra $L__BB69_135; bra.uni $L__BB69_134; $L__BB69_135: fma.rn.ftz.f32 %f460, %f7, %f230, %f229; mul.ftz.f32 %f574, %f1, %f460; bra.uni $L__BB69_136; $L__BB69_144: mov.f32 %f501, 0f3F800000; sub.ftz.f32 %f502, %f501, %f10; sub.ftz.f32 %f503, %f501, %f7; sub.ftz.f32 %f504, %f501, %f8; sub.ftz.f32 %f505, %f501, %f9; fma.rn.ftz.f32 %f506, %f503, %f10, %f502; fma.rn.ftz.f32 %f507, %f504, %f10, %f502; fma.rn.ftz.f32 %f508, %f505, %f10, %f502; sub.ftz.f32 %f509, %f501, %f1; mul.ftz.f32 %f510, %f509, %f506; sub.ftz.f32 %f511, %f501, %f2; mul.ftz.f32 %f512, %f511, %f507; sub.ftz.f32 %f513, %f501, %f3; mul.ftz.f32 %f514, %f513, %f508; sub.ftz.f32 %f576, %f501, %f514; sub.ftz.f32 %f575, %f501, %f512; sub.ftz.f32 %f574, %f501, %f510; bra.uni $L__BB69_147; $L__BB69_72: max.ftz.f32 %f341, %f2, %f3; max.ftz.f32 %f62, %f1, %f341; min.ftz.f32 %f342, %f2, %f3; min.ftz.f32 %f343, %f1, %f342; sub.ftz.f32 %f63, %f62, %f343; setp.eq.ftz.f32 %p41, %f62, 0f00000000; mov.f32 %f546, 0f00000000; mov.f32 %f544, %f546; @%p41 bra $L__BB69_74; div.approx.ftz.f32 %f544, %f63, %f62; $L__BB69_74: setp.eq.ftz.f32 %p42, %f544, 0f00000000; @%p42 bra $L__BB69_81; sub.ftz.f32 %f345, %f62, %f1; div.approx.ftz.f32 %f66, %f345, %f63; sub.ftz.f32 %f346, %f62, %f2; div.approx.ftz.f32 %f67, %f346, %f63; sub.ftz.f32 %f347, %f62, %f3; div.approx.ftz.f32 %f68, %f347, %f63; setp.eq.ftz.f32 %p43, %f1, %f62; @%p43 bra $L__BB69_79; bra.uni $L__BB69_76; $L__BB69_79: sub.ftz.f32 %f545, %f68, %f67; bra.uni $L__BB69_80; $L__BB69_145: mul.ftz.f32 %f521, %f1, %f7; sub.ftz.f32 %f522, %f521, %f1; mul.ftz.f32 %f523, %f2, %f8; sub.ftz.f32 %f524, %f523, %f2; mul.ftz.f32 %f525, %f3, %f9; sub.ftz.f32 %f526, %f525, %f3; fma.rn.ftz.f32 %f576, %f10, %f526, %f3; fma.rn.ftz.f32 %f575, %f524, %f10, %f2; fma.rn.ftz.f32 %f574, %f522, %f10, %f1; bra.uni $L__BB69_147; $L__BB69_134: sub.ftz.f32 %f456, %f454, %f7; fma.rn.ftz.f32 %f457, %f456, %f230, %f229; sub.ftz.f32 %f458, %f454, %f1; mul.ftz.f32 %f459, %f458, %f457; sub.ftz.f32 %f574, %f454, %f459; $L__BB69_136: setp.lt.ftz.f32 %p101, %f2, 0f3F000000; @%p101 bra $L__BB69_138; bra.uni $L__BB69_137; $L__BB69_138: fma.rn.ftz.f32 %f466, %f8, %f230, %f229; mul.ftz.f32 %f575, %f2, %f466; bra.uni $L__BB69_139; $L__BB69_137: sub.ftz.f32 %f462, %f454, %f8; fma.rn.ftz.f32 %f463, %f462, %f230, %f229; sub.ftz.f32 %f464, %f454, %f2; mul.ftz.f32 %f465, %f464, %f463; sub.ftz.f32 %f575, %f454, %f465; $L__BB69_139: setp.lt.ftz.f32 %p102, %f3, 0f3F000000; @%p102 bra $L__BB69_141; bra.uni $L__BB69_140; $L__BB69_141: fma.rn.ftz.f32 %f472, %f9, %f230, %f229; mul.ftz.f32 %f576, %f3, %f472; bra.uni $L__BB69_147; $L__BB69_140: sub.ftz.f32 %f468, %f454, %f9; fma.rn.ftz.f32 %f469, %f468, %f230, %f229; sub.ftz.f32 %f470, %f454, %f3; mul.ftz.f32 %f471, %f470, %f469; sub.ftz.f32 %f576, %f454, %f471; bra.uni $L__BB69_147; $L__BB69_43: setp.eq.ftz.f32 %p56, %f2, %f107; @%p56 bra $L__BB69_89; bra.uni $L__BB69_44; $L__BB69_89: add.ftz.f32 %f377, %f111, 0f40000000; sub.ftz.f32 %f551, %f377, %f113; bra.uni $L__BB69_91; $L__BB69_52: setp.eq.ftz.f32 %p32, %f8, %f20; @%p32 bra $L__BB69_60; bra.uni $L__BB69_53; $L__BB69_60: add.ftz.f32 %f322, %f24, 0f40000000; sub.ftz.f32 %f539, %f322, %f26; bra.uni $L__BB69_62; $L__BB69_76: setp.eq.ftz.f32 %p44, %f2, %f62; @%p44 bra $L__BB69_78; bra.uni $L__BB69_77; $L__BB69_78: add.ftz.f32 %f349, %f66, 0f40000000; sub.ftz.f32 %f545, %f349, %f68; bra.uni $L__BB69_80; $L__BB69_107: setp.eq.ftz.f32 %p71, %f8, %f156; @%p71 bra $L__BB69_109; bra.uni $L__BB69_108; $L__BB69_109: add.ftz.f32 %f407, %f158, 0f40000000; sub.ftz.f32 %f557, %f407, %f160; bra.uni $L__BB69_111; $L__BB69_44: add.ftz.f32 %f376, %f112, 0f40800000; sub.ftz.f32 %f551, %f376, %f111; $L__BB69_91: mov.f32 %f378, 0f40C00000; div.approx.ftz.f32 %f379, %f551, %f378; setp.lt.ftz.f32 %p57, %f379, 0f00000000; add.ftz.f32 %f380, %f379, 0f3F800000; selp.f32 %f552, %f380, %f379, %p57; $L__BB69_92: mov.f32 %f574, %f1; mov.f32 %f575, %f2; mov.f32 %f576, %f3; @%p54 bra $L__BB69_147; max.ftz.f32 %f382, %f8, %f9; max.ftz.f32 %f120, %f7, %f382; min.ftz.f32 %f383, %f8, %f9; min.ftz.f32 %f121, %f7, %f383; setp.eq.ftz.f32 %p59, %f120, 0f00000000; mov.f32 %f553, 0f00000000; @%p59 bra $L__BB69_95; sub.ftz.f32 %f384, %f120, %f121; div.approx.ftz.f32 %f553, %f384, %f120; $L__BB69_95: mul.ftz.f32 %f385, %f10, %f553; mov.f32 %f386, 0f3F800000; sub.ftz.f32 %f387, %f386, %f10; fma.rn.ftz.f32 %f124, %f387, %f550, %f385; setp.neu.ftz.f32 %p60, %f124, 0f00000000; mov.f32 %f574, %f107; mov.f32 %f575, %f107; mov.f32 %f576, %f107; @%p60 bra $L__BB69_96; bra.uni $L__BB69_147; $L__BB69_96: setp.eq.ftz.f32 %p61, %f552, 0f3F800000; mul.ftz.f32 %f389, %f552, 0f40C00000; selp.f32 %f390, 0f00000000, %f389, %p61; cvt.rmi.ftz.f32.f32 %f128, %f390; sub.ftz.f32 %f391, %f390, %f128; sub.ftz.f32 %f392, %f386, %f124; mul.ftz.f32 %f129, %f107, %f392; mul.ftz.f32 %f393, %f124, %f391; sub.ftz.f32 %f394, %f386, %f393; mul.ftz.f32 %f130, %f107, %f394; sub.ftz.f32 %f395, %f386, %f391; mul.ftz.f32 %f396, %f124, %f395; sub.ftz.f32 %f397, %f386, %f396; mul.ftz.f32 %f131, %f107, %f397; setp.eq.ftz.f32 %p62, %f128, 0f00000000; @%p62 bra $L__BB69_102; bra.uni $L__BB69_97; $L__BB69_102: mov.f32 %f574, %f107; mov.f32 %f575, %f131; mov.f32 %f576, %f129; bra.uni $L__BB69_147; $L__BB69_97: setp.eq.ftz.f32 %p63, %f128, 0f3F800000; mov.f32 %f574, %f130; mov.f32 %f575, %f107; mov.f32 %f576, %f129; @%p63 bra $L__BB69_147; setp.eq.ftz.f32 %p64, %f128, 0f40000000; mov.f32 %f574, %f129; mov.f32 %f575, %f107; mov.f32 %f576, %f131; @%p64 bra $L__BB69_147; setp.eq.ftz.f32 %p65, %f128, 0f40400000; mov.f32 %f574, %f129; mov.f32 %f575, %f130; mov.f32 %f576, %f107; @%p65 bra $L__BB69_147; setp.eq.ftz.f32 %p66, %f128, 0f40800000; mov.f32 %f574, %f107; mov.f32 %f575, %f129; mov.f32 %f576, %f130; @%p66 bra $L__BB69_101; bra.uni $L__BB69_147; $L__BB69_101: mov.f32 %f574, %f131; mov.f32 %f575, %f129; mov.f32 %f576, %f107; bra.uni $L__BB69_147; $L__BB69_53: add.ftz.f32 %f321, %f25, 0f40800000; sub.ftz.f32 %f539, %f321, %f24; $L__BB69_62: mov.f32 %f323, 0f40C00000; div.approx.ftz.f32 %f324, %f539, %f323; setp.lt.ftz.f32 %p33, %f324, 0f00000000; add.ftz.f32 %f325, %f324, 0f3F800000; selp.f32 %f540, %f325, %f324, %p33; $L__BB69_63: mov.f32 %f574, %f1; mov.f32 %f575, %f2; mov.f32 %f576, %f3; @%p30 bra $L__BB69_147; max.ftz.f32 %f326, %f2, %f3; max.ftz.f32 %f33, %f1, %f326; mul.ftz.f32 %f327, %f540, 0f40C00000; setp.eq.ftz.f32 %p35, %f540, 0f3F800000; mov.f32 %f328, 0f3F800000; selp.f32 %f329, 0f00000000, %f327, %p35; cvt.rmi.ftz.f32.f32 %f34, %f329; sub.ftz.f32 %f330, %f329, %f34; sub.ftz.f32 %f331, %f328, %f538; mul.ftz.f32 %f35, %f331, %f33; mul.ftz.f32 %f332, %f538, %f330; sub.ftz.f32 %f333, %f328, %f332; mul.ftz.f32 %f36, %f33, %f333; sub.ftz.f32 %f334, %f328, %f330; mul.ftz.f32 %f335, %f538, %f334; sub.ftz.f32 %f336, %f328, %f335; mul.ftz.f32 %f37, %f33, %f336; setp.eq.ftz.f32 %p36, %f34, 0f00000000; @%p36 bra $L__BB69_70; bra.uni $L__BB69_65; $L__BB69_70: mov.f32 %f541, %f33; mov.f32 %f542, %f37; mov.f32 %f543, %f35; bra.uni $L__BB69_71; $L__BB69_65: setp.eq.ftz.f32 %p37, %f34, 0f3F800000; mov.f32 %f541, %f36; mov.f32 %f542, %f33; mov.f32 %f543, %f35; @%p37 bra $L__BB69_71; setp.eq.ftz.f32 %p38, %f34, 0f40000000; mov.f32 %f541, %f35; mov.f32 %f542, %f33; mov.f32 %f543, %f37; @%p38 bra $L__BB69_71; setp.eq.ftz.f32 %p39, %f34, 0f40400000; mov.f32 %f541, %f35; mov.f32 %f542, %f36; mov.f32 %f543, %f33; @%p39 bra $L__BB69_71; setp.eq.ftz.f32 %p40, %f34, 0f40800000; mov.f32 %f541, %f33; mov.f32 %f542, %f35; mov.f32 %f543, %f36; @%p40 bra $L__BB69_69; bra.uni $L__BB69_71; $L__BB69_69: mov.f32 %f541, %f37; mov.f32 %f542, %f35; mov.f32 %f543, %f33; $L__BB69_71: sub.ftz.f32 %f337, %f541, %f1; sub.ftz.f32 %f338, %f542, %f2; sub.ftz.f32 %f339, %f543, %f3; fma.rn.ftz.f32 %f574, %f10, %f337, %f1; fma.rn.ftz.f32 %f575, %f10, %f338, %f2; fma.rn.ftz.f32 %f576, %f10, %f339, %f3; bra.uni $L__BB69_147; $L__BB69_77: add.ftz.f32 %f348, %f67, 0f40800000; sub.ftz.f32 %f545, %f348, %f66; $L__BB69_80: mov.f32 %f350, 0f40C00000; div.approx.ftz.f32 %f351, %f545, %f350; setp.lt.ftz.f32 %p45, %f351, 0f00000000; add.ftz.f32 %f352, %f351, 0f3F800000; selp.f32 %f546, %f352, %f351, %p45; $L__BB69_81: max.ftz.f32 %f353, %f8, %f9; max.ftz.f32 %f354, %f7, %f353; mul.ftz.f32 %f355, %f10, %f354; mov.f32 %f356, 0f3F800000; sub.ftz.f32 %f357, %f356, %f10; fma.rn.ftz.f32 %f75, %f357, %f62, %f355; setp.neu.ftz.f32 %p46, %f544, 0f00000000; mov.f32 %f574, %f75; mov.f32 %f575, %f75; mov.f32 %f576, %f75; @%p46 bra $L__BB69_82; bra.uni $L__BB69_147; $L__BB69_82: setp.eq.ftz.f32 %p47, %f546, 0f3F800000; mul.ftz.f32 %f359, %f546, 0f40C00000; selp.f32 %f360, 0f00000000, %f359, %p47; cvt.rmi.ftz.f32.f32 %f79, %f360; sub.ftz.f32 %f361, %f360, %f79; sub.ftz.f32 %f362, %f356, %f544; mul.ftz.f32 %f80, %f362, %f75; mul.ftz.f32 %f363, %f544, %f361; sub.ftz.f32 %f364, %f356, %f363; mul.ftz.f32 %f81, %f75, %f364; sub.ftz.f32 %f365, %f356, %f361; mul.ftz.f32 %f366, %f544, %f365; sub.ftz.f32 %f367, %f356, %f366; mul.ftz.f32 %f82, %f75, %f367; setp.eq.ftz.f32 %p48, %f79, 0f00000000; @%p48 bra $L__BB69_88; bra.uni $L__BB69_83; $L__BB69_88: mov.f32 %f574, %f75; mov.f32 %f575, %f82; mov.f32 %f576, %f80; bra.uni $L__BB69_147; $L__BB69_83: setp.eq.ftz.f32 %p49, %f79, 0f3F800000; mov.f32 %f574, %f81; mov.f32 %f575, %f75; mov.f32 %f576, %f80; @%p49 bra $L__BB69_147; setp.eq.ftz.f32 %p50, %f79, 0f40000000; mov.f32 %f574, %f80; mov.f32 %f575, %f75; mov.f32 %f576, %f82; @%p50 bra $L__BB69_147; setp.eq.ftz.f32 %p51, %f79, 0f40400000; mov.f32 %f574, %f80; mov.f32 %f575, %f81; mov.f32 %f576, %f75; @%p51 bra $L__BB69_147; setp.eq.ftz.f32 %p52, %f79, 0f40800000; mov.f32 %f574, %f75; mov.f32 %f575, %f80; mov.f32 %f576, %f81; @%p52 bra $L__BB69_87; bra.uni $L__BB69_147; $L__BB69_87: mov.f32 %f574, %f82; mov.f32 %f575, %f80; mov.f32 %f576, %f75; bra.uni $L__BB69_147; $L__BB69_108: add.ftz.f32 %f406, %f159, 0f40800000; sub.ftz.f32 %f557, %f406, %f158; $L__BB69_111: mov.f32 %f408, 0f40C00000; div.approx.ftz.f32 %f409, %f557, %f408; setp.lt.ftz.f32 %p72, %f409, 0f00000000; add.ftz.f32 %f410, %f409, 0f3F800000; selp.f32 %f558, %f410, %f409, %p72; $L__BB69_112: mov.f32 %f574, %f1; mov.f32 %f575, %f2; mov.f32 %f576, %f3; @%p69 bra $L__BB69_147; max.ftz.f32 %f412, %f2, %f3; max.ftz.f32 %f167, %f1, %f412; setp.eq.ftz.f32 %p74, %f167, 0f00000000; mov.f32 %f559, 0f00000000; @%p74 bra $L__BB69_115; min.ftz.f32 %f413, %f2, %f3; min.ftz.f32 %f414, %f1, %f413; sub.ftz.f32 %f415, %f167, %f414; div.approx.ftz.f32 %f559, %f415, %f167; $L__BB69_115: setp.neu.ftz.f32 %p75, %f559, 0f00000000; mov.f32 %f560, %f167; mov.f32 %f561, %f167; mov.f32 %f562, %f167; @%p75 bra $L__BB69_116; bra.uni $L__BB69_123; $L__BB69_116: setp.eq.ftz.f32 %p76, %f558, 0f3F800000; mov.f32 %f416, 0f3F800000; mul.ftz.f32 %f417, %f558, 0f40C00000; selp.f32 %f418, 0f00000000, %f417, %p76; cvt.rmi.ftz.f32.f32 %f173, %f418; sub.ftz.f32 %f419, %f418, %f173; sub.ftz.f32 %f420, %f416, %f559; mul.ftz.f32 %f174, %f167, %f420; mul.ftz.f32 %f421, %f559, %f419; sub.ftz.f32 %f422, %f416, %f421; mul.ftz.f32 %f175, %f167, %f422; sub.ftz.f32 %f423, %f416, %f419; mul.ftz.f32 %f424, %f559, %f423; sub.ftz.f32 %f425, %f416, %f424; mul.ftz.f32 %f176, %f167, %f425; setp.eq.ftz.f32 %p77, %f173, 0f00000000; @%p77 bra $L__BB69_122; bra.uni $L__BB69_117; $L__BB69_122: mov.f32 %f560, %f167; mov.f32 %f561, %f176; mov.f32 %f562, %f174; bra.uni $L__BB69_123; $L__BB69_117: setp.eq.ftz.f32 %p78, %f173, 0f3F800000; mov.f32 %f560, %f175; mov.f32 %f561, %f167; mov.f32 %f562, %f174; @%p78 bra $L__BB69_123; setp.eq.ftz.f32 %p79, %f173, 0f40000000; mov.f32 %f560, %f174; mov.f32 %f561, %f167; mov.f32 %f562, %f176; @%p79 bra $L__BB69_123; setp.eq.ftz.f32 %p80, %f173, 0f40400000; mov.f32 %f560, %f174; mov.f32 %f561, %f175; mov.f32 %f562, %f167; @%p80 bra $L__BB69_123; setp.eq.ftz.f32 %p81, %f173, 0f40800000; mov.f32 %f560, %f167; mov.f32 %f561, %f174; mov.f32 %f562, %f175; @%p81 bra $L__BB69_121; bra.uni $L__BB69_123; $L__BB69_121: mov.f32 %f560, %f176; mov.f32 %f561, %f174; mov.f32 %f562, %f167; $L__BB69_123: sub.ftz.f32 %f426, %f560, %f1; sub.ftz.f32 %f427, %f561, %f2; sub.ftz.f32 %f428, %f562, %f3; fma.rn.ftz.f32 %f574, %f10, %f426, %f1; fma.rn.ftz.f32 %f575, %f10, %f427, %f2; fma.rn.ftz.f32 %f576, %f10, %f428, %f3; $L__BB69_147: mul.wide.u32 %rd14, %r12, 4; add.s64 %rd15, %rd2, %rd14; st.local.f32 [%rd15], %f574; st.local.f32 [%rd15+4], %f575; st.local.f32 [%rd15+8], %f576; add.s32 %r14, %r5, 1; st.param.b32 [func_retval0+0], %r14; ret; } .func _Z18svm_node_mix_colorP10ShaderDataPfjjj( .param .b64 _Z18svm_node_mix_colorP10ShaderDataPfjjj_param_0, .param .b32 _Z18svm_node_mix_colorP10ShaderDataPfjjj_param_1, .param .b32 _Z18svm_node_mix_colorP10ShaderDataPfjjj_param_2, .param .b32 _Z18svm_node_mix_colorP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<109>; .reg .b16 %rs<3>; .reg .f32 %f<588>; .reg .b32 %r<9>; .reg .b64 %rd<15>; ld.param.u64 %rd2, [_Z18svm_node_mix_colorP10ShaderDataPfjjj_param_0]; ld.param.u32 %r1, [_Z18svm_node_mix_colorP10ShaderDataPfjjj_param_1]; ld.param.u32 %r2, [_Z18svm_node_mix_colorP10ShaderDataPfjjj_param_2]; ld.param.u32 %r3, [_Z18svm_node_mix_colorP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; shl.b32 %r4, %r2, 2; cvt.u64.u32 %rd3, %r4; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f3, [%rd5]; and.b32 %r5, %r1, 255; setp.eq.s32 %p3, %r5, 0; @%p3 bra $L__BB70_2; cvt.ftz.sat.f32.f32 %f3, %f3; $L__BB70_2: shr.u32 %r6, %r2, 6; cvt.u64.u32 %rd6, %r6; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f4, [%rd8]; ld.local.f32 %f5, [%rd8+4]; ld.local.f32 %f6, [%rd8+8]; shr.u32 %r7, %r2, 14; cvt.u64.u32 %rd9, %r7; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f10, [%rd11]; ld.local.f32 %f11, [%rd11+4]; ld.local.f32 %f12, [%rd11+8]; cvt.u16.u32 %rs2, %r1; shr.u16 %rs1, %rs2, 8; mov.f32 %f582, 0f00000000; setp.gt.s16 %p4, %rs1, 8; @%p4 bra $L__BB70_25; bra.uni $L__BB70_3; $L__BB70_25: setp.gt.s16 %p5, %rs1, 13; @%p5 bra $L__BB70_47; setp.gt.s16 %p12, %rs1, 10; @%p12 bra $L__BB70_38; bra.uni $L__BB70_27; $L__BB70_38: setp.eq.s16 %p13, %rs1, 11; @%p13 bra $L__BB70_126; setp.eq.s16 %p14, %rs1, 12; @%p14 bra $L__BB70_105; setp.eq.s16 %p15, %rs1, 13; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p15 bra $L__BB70_41; bra.uni $L__BB70_149; $L__BB70_41: max.ftz.f32 %f376, %f5, %f6; max.ftz.f32 %f109, %f4, %f376; min.ftz.f32 %f377, %f5, %f6; min.ftz.f32 %f378, %f4, %f377; sub.ftz.f32 %f110, %f109, %f378; setp.eq.ftz.f32 %p54, %f109, 0f00000000; mov.f32 %f560, 0f00000000; mov.f32 %f558, %f560; @%p54 bra $L__BB70_43; div.approx.ftz.f32 %f558, %f110, %f109; $L__BB70_43: setp.eq.ftz.f32 %p55, %f558, 0f00000000; @%p55 bra $L__BB70_94; sub.ftz.f32 %f380, %f109, %f4; div.approx.ftz.f32 %f113, %f380, %f110; sub.ftz.f32 %f381, %f109, %f5; div.approx.ftz.f32 %f114, %f381, %f110; sub.ftz.f32 %f382, %f109, %f6; div.approx.ftz.f32 %f115, %f382, %f110; setp.eq.ftz.f32 %p56, %f4, %f109; @%p56 bra $L__BB70_92; bra.uni $L__BB70_45; $L__BB70_92: sub.ftz.f32 %f559, %f115, %f114; bra.uni $L__BB70_93; $L__BB70_3: setp.gt.s16 %p18, %rs1, 3; @%p18 bra $L__BB70_11; bra.uni $L__BB70_4; $L__BB70_11: setp.gt.s16 %p19, %rs1, 5; @%p19 bra $L__BB70_21; bra.uni $L__BB70_12; $L__BB70_21: setp.eq.s16 %p20, %rs1, 6; @%p20 bra $L__BB70_145; setp.eq.s16 %p21, %rs1, 7; @%p21 bra $L__BB70_144; setp.eq.s16 %p22, %rs1, 8; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p22 bra $L__BB70_24; bra.uni $L__BB70_149; $L__BB70_24: max.ftz.f32 %f480, %f4, %f10; sub.ftz.f32 %f481, %f480, %f4; max.ftz.f32 %f482, %f5, %f11; sub.ftz.f32 %f483, %f482, %f5; max.ftz.f32 %f484, %f6, %f12; sub.ftz.f32 %f485, %f484, %f6; fma.rn.ftz.f32 %f584, %f3, %f485, %f6; fma.rn.ftz.f32 %f583, %f3, %f483, %f5; fma.rn.ftz.f32 %f582, %f3, %f481, %f4; bra.uni $L__BB70_149; $L__BB70_47: setp.gt.s16 %p6, %rs1, 15; @%p6 bra $L__BB70_56; bra.uni $L__BB70_48; $L__BB70_56: setp.eq.s16 %p7, %rs1, 16; @%p7 bra $L__BB70_61; setp.eq.s16 %p8, %rs1, 17; @%p8 bra $L__BB70_60; setp.ne.s16 %p9, %rs1, 18; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p9 bra $L__BB70_149; cvt.ftz.sat.f32.f32 %f584, %f6; cvt.ftz.sat.f32.f32 %f583, %f5; cvt.ftz.sat.f32.f32 %f582, %f4; bra.uni $L__BB70_149; $L__BB70_4: setp.gt.s16 %p25, %rs1, 1; @%p25 bra $L__BB70_8; setp.eq.s16 %p28, %rs1, 0; @%p28 bra $L__BB70_148; setp.eq.s16 %p29, %rs1, 1; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p29 bra $L__BB70_7; bra.uni $L__BB70_149; $L__BB70_7: add.ftz.f32 %f534, %f4, %f10; sub.ftz.f32 %f535, %f534, %f4; add.ftz.f32 %f536, %f5, %f11; sub.ftz.f32 %f537, %f536, %f5; add.ftz.f32 %f538, %f6, %f12; sub.ftz.f32 %f539, %f538, %f6; fma.rn.ftz.f32 %f584, %f3, %f539, %f6; fma.rn.ftz.f32 %f583, %f3, %f537, %f5; fma.rn.ftz.f32 %f582, %f3, %f535, %f4; bra.uni $L__BB70_149; $L__BB70_27: setp.eq.s16 %p16, %rs1, 9; @%p16 bra $L__BB70_135; setp.eq.s16 %p17, %rs1, 10; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p17 bra $L__BB70_29; bra.uni $L__BB70_149; $L__BB70_29: setp.eq.ftz.f32 %p92, %f4, 0f00000000; mov.f32 %f582, %f4; @%p92 bra $L__BB70_32; mul.ftz.f32 %f453, %f3, %f10; mov.f32 %f582, 0f3F800000; sub.ftz.f32 %f219, %f582, %f453; setp.le.ftz.f32 %p93, %f219, 0f00000000; @%p93 bra $L__BB70_32; div.approx.ftz.f32 %f454, %f4, %f219; setp.gt.ftz.f32 %p94, %f454, 0f3F800000; selp.f32 %f582, 0f3F800000, %f454, %p94; $L__BB70_32: setp.eq.ftz.f32 %p95, %f5, 0f00000000; mov.f32 %f583, %f5; @%p95 bra $L__BB70_35; mul.ftz.f32 %f456, %f3, %f11; mov.f32 %f583, 0f3F800000; sub.ftz.f32 %f222, %f583, %f456; setp.le.ftz.f32 %p96, %f222, 0f00000000; @%p96 bra $L__BB70_35; div.approx.ftz.f32 %f457, %f5, %f222; setp.gt.ftz.f32 %p97, %f457, 0f3F800000; selp.f32 %f583, 0f3F800000, %f457, %p97; $L__BB70_35: setp.eq.ftz.f32 %p98, %f6, 0f00000000; mov.f32 %f584, %f6; @%p98 bra $L__BB70_149; mul.ftz.f32 %f459, %f3, %f12; mov.f32 %f584, 0f3F800000; sub.ftz.f32 %f225, %f584, %f459; setp.le.ftz.f32 %p99, %f225, 0f00000000; @%p99 bra $L__BB70_149; div.approx.ftz.f32 %f460, %f6, %f225; setp.gt.ftz.f32 %p100, %f460, 0f3F800000; selp.f32 %f584, 0f3F800000, %f460, %p100; bra.uni $L__BB70_149; $L__BB70_12: setp.eq.s16 %p23, %rs1, 4; @%p23 bra $L__BB70_146; setp.eq.s16 %p24, %rs1, 5; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p24 bra $L__BB70_14; bra.uni $L__BB70_149; $L__BB70_14: mov.f32 %f501, 0f3F800000; sub.ftz.f32 %f254, %f501, %f3; setp.eq.ftz.f32 %p104, %f10, 0f00000000; @%p104 bra $L__BB70_16; mul.ftz.f32 %f502, %f3, %f4; div.approx.ftz.f32 %f503, %f502, %f10; fma.rn.ftz.f32 %f4, %f254, %f4, %f503; $L__BB70_16: setp.eq.ftz.f32 %p105, %f11, 0f00000000; @%p105 bra $L__BB70_18; mul.ftz.f32 %f504, %f3, %f5; div.approx.ftz.f32 %f505, %f504, %f11; fma.rn.ftz.f32 %f5, %f254, %f5, %f505; $L__BB70_18: setp.eq.ftz.f32 %p106, %f12, 0f00000000; @%p106 bra $L__BB70_20; mul.ftz.f32 %f506, %f3, %f6; div.approx.ftz.f32 %f507, %f506, %f12; fma.rn.ftz.f32 %f6, %f254, %f6, %f507; $L__BB70_20: mov.f32 %f582, %f4; mov.f32 %f583, %f5; mov.f32 %f584, %f6; bra.uni $L__BB70_149; $L__BB70_48: setp.eq.s16 %p10, %rs1, 14; @%p10 bra $L__BB70_74; setp.eq.s16 %p11, %rs1, 15; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p11 bra $L__BB70_50; bra.uni $L__BB70_149; $L__BB70_50: max.ftz.f32 %f321, %f11, %f12; max.ftz.f32 %f22, %f10, %f321; min.ftz.f32 %f322, %f11, %f12; min.ftz.f32 %f323, %f10, %f322; sub.ftz.f32 %f23, %f22, %f323; setp.eq.ftz.f32 %p30, %f22, 0f00000000; mov.f32 %f548, 0f00000000; mov.f32 %f546, %f548; @%p30 bra $L__BB70_52; div.approx.ftz.f32 %f546, %f23, %f22; $L__BB70_52: setp.eq.ftz.f32 %p31, %f546, 0f00000000; @%p31 bra $L__BB70_65; sub.ftz.f32 %f325, %f22, %f10; div.approx.ftz.f32 %f26, %f325, %f23; sub.ftz.f32 %f326, %f22, %f11; div.approx.ftz.f32 %f27, %f326, %f23; sub.ftz.f32 %f327, %f22, %f12; div.approx.ftz.f32 %f28, %f327, %f23; setp.eq.ftz.f32 %p32, %f10, %f22; @%p32 bra $L__BB70_63; bra.uni $L__BB70_54; $L__BB70_63: sub.ftz.f32 %f547, %f28, %f27; bra.uni $L__BB70_64; $L__BB70_8: setp.eq.s16 %p26, %rs1, 2; @%p26 bra $L__BB70_147; setp.eq.s16 %p27, %rs1, 3; mov.f32 %f583, %f582; mov.f32 %f584, %f582; @%p27 bra $L__BB70_10; bra.uni $L__BB70_149; $L__BB70_10: sub.ftz.f32 %f522, %f4, %f10; sub.ftz.f32 %f523, %f522, %f4; sub.ftz.f32 %f524, %f5, %f11; sub.ftz.f32 %f525, %f524, %f5; sub.ftz.f32 %f526, %f6, %f12; sub.ftz.f32 %f527, %f526, %f6; fma.rn.ftz.f32 %f584, %f3, %f527, %f6; fma.rn.ftz.f32 %f583, %f3, %f525, %f5; fma.rn.ftz.f32 %f582, %f3, %f523, %f4; bra.uni $L__BB70_149; $L__BB70_126: mov.f32 %f437, 0f3F800000; sub.ftz.f32 %f203, %f437, %f3; fma.rn.ftz.f32 %f204, %f3, %f10, %f203; setp.le.ftz.f32 %p83, %f204, 0f00000000; mov.f32 %f584, 0f00000000; mov.f32 %f582, %f584; @%p83 bra $L__BB70_129; sub.ftz.f32 %f440, %f437, %f4; div.approx.ftz.f32 %f441, %f440, %f204; sub.ftz.f32 %f205, %f437, %f441; setp.lt.ftz.f32 %p84, %f205, 0f00000000; mov.f32 %f582, %f584; @%p84 bra $L__BB70_129; setp.gt.ftz.f32 %p85, %f205, 0f3F800000; selp.f32 %f582, 0f3F800000, %f205, %p85; $L__BB70_129: fma.rn.ftz.f32 %f208, %f3, %f11, %f203; setp.le.ftz.f32 %p86, %f208, 0f00000000; mov.f32 %f583, %f584; @%p86 bra $L__BB70_132; sub.ftz.f32 %f445, %f437, %f5; div.approx.ftz.f32 %f446, %f445, %f208; sub.ftz.f32 %f209, %f437, %f446; setp.lt.ftz.f32 %p87, %f209, 0f00000000; mov.f32 %f583, %f584; @%p87 bra $L__BB70_132; setp.gt.ftz.f32 %p88, %f209, 0f3F800000; selp.f32 %f583, 0f3F800000, %f209, %p88; $L__BB70_132: fma.rn.ftz.f32 %f212, %f3, %f12, %f203; setp.le.ftz.f32 %p89, %f212, 0f00000000; @%p89 bra $L__BB70_149; sub.ftz.f32 %f450, %f437, %f6; div.approx.ftz.f32 %f451, %f450, %f212; sub.ftz.f32 %f213, %f437, %f451; setp.lt.ftz.f32 %p90, %f213, 0f00000000; @%p90 bra $L__BB70_149; setp.gt.ftz.f32 %p91, %f213, 0f3F800000; selp.f32 %f584, 0f3F800000, %f213, %p91; bra.uni $L__BB70_149; $L__BB70_105: max.ftz.f32 %f405, %f11, %f12; max.ftz.f32 %f158, %f10, %f405; min.ftz.f32 %f406, %f11, %f12; min.ftz.f32 %f407, %f10, %f406; sub.ftz.f32 %f159, %f158, %f407; setp.eq.ftz.f32 %p69, %f158, 0f00000000; mov.pred %p108, 0; @%p69 bra $L__BB70_107; div.approx.ftz.f32 %f408, %f159, %f158; setp.neu.ftz.f32 %p108, %f408, 0f00000000; $L__BB70_107: mov.f32 %f566, 0f00000000; not.pred %p70, %p108; @%p70 bra $L__BB70_114; sub.ftz.f32 %f410, %f158, %f10; div.approx.ftz.f32 %f160, %f410, %f159; sub.ftz.f32 %f411, %f158, %f11; div.approx.ftz.f32 %f161, %f411, %f159; sub.ftz.f32 %f412, %f158, %f12; div.approx.ftz.f32 %f162, %f412, %f159; setp.eq.ftz.f32 %p71, %f10, %f158; @%p71 bra $L__BB70_112; bra.uni $L__BB70_109; $L__BB70_112: sub.ftz.f32 %f565, %f162, %f161; bra.uni $L__BB70_113; $L__BB70_145: sub.ftz.f32 %f492, %f4, %f10; abs.ftz.f32 %f493, %f492; sub.ftz.f32 %f494, %f5, %f11; abs.ftz.f32 %f495, %f494; sub.ftz.f32 %f496, %f6, %f12; abs.ftz.f32 %f497, %f496; sub.ftz.f32 %f498, %f493, %f4; sub.ftz.f32 %f499, %f495, %f5; sub.ftz.f32 %f500, %f497, %f6; fma.rn.ftz.f32 %f584, %f3, %f500, %f6; fma.rn.ftz.f32 %f583, %f3, %f499, %f5; fma.rn.ftz.f32 %f582, %f3, %f498, %f4; bra.uni $L__BB70_149; $L__BB70_144: min.ftz.f32 %f486, %f4, %f10; sub.ftz.f32 %f487, %f486, %f4; min.ftz.f32 %f488, %f5, %f11; sub.ftz.f32 %f489, %f488, %f5; min.ftz.f32 %f490, %f6, %f12; sub.ftz.f32 %f491, %f490, %f6; fma.rn.ftz.f32 %f584, %f3, %f491, %f6; fma.rn.ftz.f32 %f583, %f3, %f489, %f5; fma.rn.ftz.f32 %f582, %f3, %f487, %f4; bra.uni $L__BB70_149; $L__BB70_61: mov.f32 %f294, 0f3F800000; sub.ftz.f32 %f295, %f294, %f3; sub.ftz.f32 %f296, %f294, %f4; sub.ftz.f32 %f297, %f294, %f10; mul.ftz.f32 %f298, %f296, %f297; sub.ftz.f32 %f299, %f294, %f5; sub.ftz.f32 %f300, %f294, %f11; mul.ftz.f32 %f301, %f299, %f300; sub.ftz.f32 %f302, %f294, %f6; sub.ftz.f32 %f303, %f294, %f12; mul.ftz.f32 %f304, %f302, %f303; sub.ftz.f32 %f305, %f294, %f298; sub.ftz.f32 %f306, %f294, %f301; sub.ftz.f32 %f307, %f294, %f304; mul.ftz.f32 %f308, %f296, %f10; mul.ftz.f32 %f309, %f299, %f11; mul.ftz.f32 %f310, %f302, %f12; mul.ftz.f32 %f311, %f4, %f305; mul.ftz.f32 %f312, %f5, %f306; mul.ftz.f32 %f313, %f6, %f307; fma.rn.ftz.f32 %f314, %f4, %f308, %f311; fma.rn.ftz.f32 %f315, %f5, %f309, %f312; fma.rn.ftz.f32 %f316, %f6, %f310, %f313; mul.ftz.f32 %f317, %f3, %f314; mul.ftz.f32 %f318, %f3, %f315; mul.ftz.f32 %f319, %f3, %f316; fma.rn.ftz.f32 %f584, %f295, %f6, %f319; fma.rn.ftz.f32 %f583, %f295, %f5, %f318; fma.rn.ftz.f32 %f582, %f4, %f295, %f317; bra.uni $L__BB70_149; $L__BB70_60: fma.rn.ftz.f32 %f291, %f10, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f292, %f11, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f293, %f12, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f584, %f3, %f293, %f6; fma.rn.ftz.f32 %f583, %f3, %f292, %f5; fma.rn.ftz.f32 %f582, %f3, %f291, %f4; bra.uni $L__BB70_149; $L__BB70_148: sub.ftz.f32 %f540, %f10, %f4; sub.ftz.f32 %f541, %f11, %f5; sub.ftz.f32 %f542, %f12, %f6; fma.rn.ftz.f32 %f584, %f3, %f542, %f6; fma.rn.ftz.f32 %f583, %f3, %f541, %f5; fma.rn.ftz.f32 %f582, %f3, %f540, %f4; bra.uni $L__BB70_149; $L__BB70_135: mov.f32 %f461, 0f3F800000; sub.ftz.f32 %f231, %f461, %f3; add.ftz.f32 %f232, %f3, %f3; setp.lt.ftz.f32 %p101, %f4, 0f3F000000; @%p101 bra $L__BB70_137; bra.uni $L__BB70_136; $L__BB70_137: fma.rn.ftz.f32 %f467, %f232, %f10, %f231; mul.ftz.f32 %f582, %f4, %f467; bra.uni $L__BB70_138; $L__BB70_146: mov.f32 %f508, 0f3F800000; sub.ftz.f32 %f509, %f508, %f3; sub.ftz.f32 %f510, %f508, %f10; sub.ftz.f32 %f511, %f508, %f11; sub.ftz.f32 %f512, %f508, %f12; fma.rn.ftz.f32 %f513, %f3, %f510, %f509; fma.rn.ftz.f32 %f514, %f3, %f511, %f509; fma.rn.ftz.f32 %f515, %f3, %f512, %f509; sub.ftz.f32 %f516, %f508, %f4; mul.ftz.f32 %f517, %f516, %f513; sub.ftz.f32 %f518, %f508, %f5; mul.ftz.f32 %f519, %f518, %f514; sub.ftz.f32 %f520, %f508, %f6; mul.ftz.f32 %f521, %f520, %f515; sub.ftz.f32 %f584, %f508, %f521; sub.ftz.f32 %f583, %f508, %f519; sub.ftz.f32 %f582, %f508, %f517; bra.uni $L__BB70_149; $L__BB70_74: max.ftz.f32 %f348, %f5, %f6; max.ftz.f32 %f64, %f4, %f348; min.ftz.f32 %f349, %f5, %f6; min.ftz.f32 %f350, %f4, %f349; sub.ftz.f32 %f65, %f64, %f350; setp.eq.ftz.f32 %p42, %f64, 0f00000000; mov.f32 %f554, 0f00000000; mov.f32 %f552, %f554; @%p42 bra $L__BB70_76; div.approx.ftz.f32 %f552, %f65, %f64; $L__BB70_76: setp.eq.ftz.f32 %p43, %f552, 0f00000000; @%p43 bra $L__BB70_83; sub.ftz.f32 %f352, %f64, %f4; div.approx.ftz.f32 %f68, %f352, %f65; sub.ftz.f32 %f353, %f64, %f5; div.approx.ftz.f32 %f69, %f353, %f65; sub.ftz.f32 %f354, %f64, %f6; div.approx.ftz.f32 %f70, %f354, %f65; setp.eq.ftz.f32 %p44, %f4, %f64; @%p44 bra $L__BB70_81; bra.uni $L__BB70_78; $L__BB70_81: sub.ftz.f32 %f553, %f70, %f69; bra.uni $L__BB70_82; $L__BB70_147: mul.ftz.f32 %f528, %f4, %f10; sub.ftz.f32 %f529, %f528, %f4; mul.ftz.f32 %f530, %f5, %f11; sub.ftz.f32 %f531, %f530, %f5; mul.ftz.f32 %f532, %f6, %f12; sub.ftz.f32 %f533, %f532, %f6; fma.rn.ftz.f32 %f584, %f3, %f533, %f6; fma.rn.ftz.f32 %f583, %f3, %f531, %f5; fma.rn.ftz.f32 %f582, %f3, %f529, %f4; bra.uni $L__BB70_149; $L__BB70_136: sub.ftz.f32 %f463, %f461, %f10; fma.rn.ftz.f32 %f464, %f232, %f463, %f231; sub.ftz.f32 %f465, %f461, %f4; mul.ftz.f32 %f466, %f465, %f464; sub.ftz.f32 %f582, %f461, %f466; $L__BB70_138: setp.lt.ftz.f32 %p102, %f5, 0f3F000000; @%p102 bra $L__BB70_140; bra.uni $L__BB70_139; $L__BB70_140: fma.rn.ftz.f32 %f473, %f232, %f11, %f231; mul.ftz.f32 %f583, %f5, %f473; bra.uni $L__BB70_141; $L__BB70_139: sub.ftz.f32 %f469, %f461, %f11; fma.rn.ftz.f32 %f470, %f232, %f469, %f231; sub.ftz.f32 %f471, %f461, %f5; mul.ftz.f32 %f472, %f471, %f470; sub.ftz.f32 %f583, %f461, %f472; $L__BB70_141: setp.lt.ftz.f32 %p103, %f6, 0f3F000000; @%p103 bra $L__BB70_143; bra.uni $L__BB70_142; $L__BB70_143: fma.rn.ftz.f32 %f479, %f232, %f12, %f231; mul.ftz.f32 %f584, %f6, %f479; bra.uni $L__BB70_149; $L__BB70_142: sub.ftz.f32 %f475, %f461, %f12; fma.rn.ftz.f32 %f476, %f232, %f475, %f231; sub.ftz.f32 %f477, %f461, %f6; mul.ftz.f32 %f478, %f477, %f476; sub.ftz.f32 %f584, %f461, %f478; bra.uni $L__BB70_149; $L__BB70_45: setp.eq.ftz.f32 %p57, %f5, %f109; @%p57 bra $L__BB70_91; bra.uni $L__BB70_46; $L__BB70_91: add.ftz.f32 %f384, %f113, 0f40000000; sub.ftz.f32 %f559, %f384, %f115; bra.uni $L__BB70_93; $L__BB70_54: setp.eq.ftz.f32 %p33, %f11, %f22; @%p33 bra $L__BB70_62; bra.uni $L__BB70_55; $L__BB70_62: add.ftz.f32 %f329, %f26, 0f40000000; sub.ftz.f32 %f547, %f329, %f28; bra.uni $L__BB70_64; $L__BB70_78: setp.eq.ftz.f32 %p45, %f5, %f64; @%p45 bra $L__BB70_80; bra.uni $L__BB70_79; $L__BB70_80: add.ftz.f32 %f356, %f68, 0f40000000; sub.ftz.f32 %f553, %f356, %f70; bra.uni $L__BB70_82; $L__BB70_109: setp.eq.ftz.f32 %p72, %f11, %f158; @%p72 bra $L__BB70_111; bra.uni $L__BB70_110; $L__BB70_111: add.ftz.f32 %f414, %f160, 0f40000000; sub.ftz.f32 %f565, %f414, %f162; bra.uni $L__BB70_113; $L__BB70_46: add.ftz.f32 %f383, %f114, 0f40800000; sub.ftz.f32 %f559, %f383, %f113; $L__BB70_93: mov.f32 %f385, 0f40C00000; div.approx.ftz.f32 %f386, %f559, %f385; setp.lt.ftz.f32 %p58, %f386, 0f00000000; add.ftz.f32 %f387, %f386, 0f3F800000; selp.f32 %f560, %f387, %f386, %p58; $L__BB70_94: mov.f32 %f582, %f4; mov.f32 %f583, %f5; mov.f32 %f584, %f6; @%p55 bra $L__BB70_149; max.ftz.f32 %f389, %f11, %f12; max.ftz.f32 %f122, %f10, %f389; min.ftz.f32 %f390, %f11, %f12; min.ftz.f32 %f123, %f10, %f390; setp.eq.ftz.f32 %p60, %f122, 0f00000000; mov.f32 %f561, 0f00000000; @%p60 bra $L__BB70_97; sub.ftz.f32 %f391, %f122, %f123; div.approx.ftz.f32 %f561, %f391, %f122; $L__BB70_97: mul.ftz.f32 %f392, %f3, %f561; mov.f32 %f393, 0f3F800000; sub.ftz.f32 %f394, %f393, %f3; fma.rn.ftz.f32 %f126, %f394, %f558, %f392; setp.neu.ftz.f32 %p61, %f126, 0f00000000; mov.f32 %f582, %f109; mov.f32 %f583, %f109; mov.f32 %f584, %f109; @%p61 bra $L__BB70_98; bra.uni $L__BB70_149; $L__BB70_98: setp.eq.ftz.f32 %p62, %f560, 0f3F800000; mul.ftz.f32 %f396, %f560, 0f40C00000; selp.f32 %f397, 0f00000000, %f396, %p62; cvt.rmi.ftz.f32.f32 %f130, %f397; sub.ftz.f32 %f398, %f397, %f130; sub.ftz.f32 %f399, %f393, %f126; mul.ftz.f32 %f131, %f109, %f399; mul.ftz.f32 %f400, %f126, %f398; sub.ftz.f32 %f401, %f393, %f400; mul.ftz.f32 %f132, %f109, %f401; sub.ftz.f32 %f402, %f393, %f398; mul.ftz.f32 %f403, %f126, %f402; sub.ftz.f32 %f404, %f393, %f403; mul.ftz.f32 %f133, %f109, %f404; setp.eq.ftz.f32 %p63, %f130, 0f00000000; @%p63 bra $L__BB70_104; bra.uni $L__BB70_99; $L__BB70_104: mov.f32 %f582, %f109; mov.f32 %f583, %f133; mov.f32 %f584, %f131; bra.uni $L__BB70_149; $L__BB70_99: setp.eq.ftz.f32 %p64, %f130, 0f3F800000; mov.f32 %f582, %f132; mov.f32 %f583, %f109; mov.f32 %f584, %f131; @%p64 bra $L__BB70_149; setp.eq.ftz.f32 %p65, %f130, 0f40000000; mov.f32 %f582, %f131; mov.f32 %f583, %f109; mov.f32 %f584, %f133; @%p65 bra $L__BB70_149; setp.eq.ftz.f32 %p66, %f130, 0f40400000; mov.f32 %f582, %f131; mov.f32 %f583, %f132; mov.f32 %f584, %f109; @%p66 bra $L__BB70_149; setp.eq.ftz.f32 %p67, %f130, 0f40800000; mov.f32 %f582, %f109; mov.f32 %f583, %f131; mov.f32 %f584, %f132; @%p67 bra $L__BB70_103; bra.uni $L__BB70_149; $L__BB70_103: mov.f32 %f582, %f133; mov.f32 %f583, %f131; mov.f32 %f584, %f109; bra.uni $L__BB70_149; $L__BB70_55: add.ftz.f32 %f328, %f27, 0f40800000; sub.ftz.f32 %f547, %f328, %f26; $L__BB70_64: mov.f32 %f330, 0f40C00000; div.approx.ftz.f32 %f331, %f547, %f330; setp.lt.ftz.f32 %p34, %f331, 0f00000000; add.ftz.f32 %f332, %f331, 0f3F800000; selp.f32 %f548, %f332, %f331, %p34; $L__BB70_65: mov.f32 %f582, %f4; mov.f32 %f583, %f5; mov.f32 %f584, %f6; @%p31 bra $L__BB70_149; max.ftz.f32 %f333, %f5, %f6; max.ftz.f32 %f35, %f4, %f333; mul.ftz.f32 %f334, %f548, 0f40C00000; setp.eq.ftz.f32 %p36, %f548, 0f3F800000; mov.f32 %f335, 0f3F800000; selp.f32 %f336, 0f00000000, %f334, %p36; cvt.rmi.ftz.f32.f32 %f36, %f336; sub.ftz.f32 %f337, %f336, %f36; sub.ftz.f32 %f338, %f335, %f546; mul.ftz.f32 %f37, %f338, %f35; mul.ftz.f32 %f339, %f546, %f337; sub.ftz.f32 %f340, %f335, %f339; mul.ftz.f32 %f38, %f35, %f340; sub.ftz.f32 %f341, %f335, %f337; mul.ftz.f32 %f342, %f546, %f341; sub.ftz.f32 %f343, %f335, %f342; mul.ftz.f32 %f39, %f35, %f343; setp.eq.ftz.f32 %p37, %f36, 0f00000000; @%p37 bra $L__BB70_72; bra.uni $L__BB70_67; $L__BB70_72: mov.f32 %f549, %f35; mov.f32 %f550, %f39; mov.f32 %f551, %f37; bra.uni $L__BB70_73; $L__BB70_67: setp.eq.ftz.f32 %p38, %f36, 0f3F800000; mov.f32 %f549, %f38; mov.f32 %f550, %f35; mov.f32 %f551, %f37; @%p38 bra $L__BB70_73; setp.eq.ftz.f32 %p39, %f36, 0f40000000; mov.f32 %f549, %f37; mov.f32 %f550, %f35; mov.f32 %f551, %f39; @%p39 bra $L__BB70_73; setp.eq.ftz.f32 %p40, %f36, 0f40400000; mov.f32 %f549, %f37; mov.f32 %f550, %f38; mov.f32 %f551, %f35; @%p40 bra $L__BB70_73; setp.eq.ftz.f32 %p41, %f36, 0f40800000; mov.f32 %f549, %f35; mov.f32 %f550, %f37; mov.f32 %f551, %f38; @%p41 bra $L__BB70_71; bra.uni $L__BB70_73; $L__BB70_71: mov.f32 %f549, %f39; mov.f32 %f550, %f37; mov.f32 %f551, %f35; $L__BB70_73: sub.ftz.f32 %f344, %f549, %f4; sub.ftz.f32 %f345, %f550, %f5; sub.ftz.f32 %f346, %f551, %f6; fma.rn.ftz.f32 %f582, %f3, %f344, %f4; fma.rn.ftz.f32 %f583, %f3, %f345, %f5; fma.rn.ftz.f32 %f584, %f3, %f346, %f6; bra.uni $L__BB70_149; $L__BB70_79: add.ftz.f32 %f355, %f69, 0f40800000; sub.ftz.f32 %f553, %f355, %f68; $L__BB70_82: mov.f32 %f357, 0f40C00000; div.approx.ftz.f32 %f358, %f553, %f357; setp.lt.ftz.f32 %p46, %f358, 0f00000000; add.ftz.f32 %f359, %f358, 0f3F800000; selp.f32 %f554, %f359, %f358, %p46; $L__BB70_83: max.ftz.f32 %f360, %f11, %f12; max.ftz.f32 %f361, %f10, %f360; mul.ftz.f32 %f362, %f3, %f361; mov.f32 %f363, 0f3F800000; sub.ftz.f32 %f364, %f363, %f3; fma.rn.ftz.f32 %f77, %f364, %f64, %f362; setp.neu.ftz.f32 %p47, %f552, 0f00000000; mov.f32 %f582, %f77; mov.f32 %f583, %f77; mov.f32 %f584, %f77; @%p47 bra $L__BB70_84; bra.uni $L__BB70_149; $L__BB70_84: setp.eq.ftz.f32 %p48, %f554, 0f3F800000; mul.ftz.f32 %f366, %f554, 0f40C00000; selp.f32 %f367, 0f00000000, %f366, %p48; cvt.rmi.ftz.f32.f32 %f81, %f367; sub.ftz.f32 %f368, %f367, %f81; sub.ftz.f32 %f369, %f363, %f552; mul.ftz.f32 %f82, %f369, %f77; mul.ftz.f32 %f370, %f552, %f368; sub.ftz.f32 %f371, %f363, %f370; mul.ftz.f32 %f83, %f77, %f371; sub.ftz.f32 %f372, %f363, %f368; mul.ftz.f32 %f373, %f552, %f372; sub.ftz.f32 %f374, %f363, %f373; mul.ftz.f32 %f84, %f77, %f374; setp.eq.ftz.f32 %p49, %f81, 0f00000000; @%p49 bra $L__BB70_90; bra.uni $L__BB70_85; $L__BB70_90: mov.f32 %f582, %f77; mov.f32 %f583, %f84; mov.f32 %f584, %f82; bra.uni $L__BB70_149; $L__BB70_85: setp.eq.ftz.f32 %p50, %f81, 0f3F800000; mov.f32 %f582, %f83; mov.f32 %f583, %f77; mov.f32 %f584, %f82; @%p50 bra $L__BB70_149; setp.eq.ftz.f32 %p51, %f81, 0f40000000; mov.f32 %f582, %f82; mov.f32 %f583, %f77; mov.f32 %f584, %f84; @%p51 bra $L__BB70_149; setp.eq.ftz.f32 %p52, %f81, 0f40400000; mov.f32 %f582, %f82; mov.f32 %f583, %f83; mov.f32 %f584, %f77; @%p52 bra $L__BB70_149; setp.eq.ftz.f32 %p53, %f81, 0f40800000; mov.f32 %f582, %f77; mov.f32 %f583, %f82; mov.f32 %f584, %f83; @%p53 bra $L__BB70_89; bra.uni $L__BB70_149; $L__BB70_89: mov.f32 %f582, %f84; mov.f32 %f583, %f82; mov.f32 %f584, %f77; bra.uni $L__BB70_149; $L__BB70_110: add.ftz.f32 %f413, %f161, 0f40800000; sub.ftz.f32 %f565, %f413, %f160; $L__BB70_113: mov.f32 %f415, 0f40C00000; div.approx.ftz.f32 %f416, %f565, %f415; setp.lt.ftz.f32 %p73, %f416, 0f00000000; add.ftz.f32 %f417, %f416, 0f3F800000; selp.f32 %f566, %f417, %f416, %p73; $L__BB70_114: mov.f32 %f582, %f4; mov.f32 %f583, %f5; mov.f32 %f584, %f6; @%p70 bra $L__BB70_149; max.ftz.f32 %f419, %f5, %f6; max.ftz.f32 %f169, %f4, %f419; setp.eq.ftz.f32 %p75, %f169, 0f00000000; mov.f32 %f567, 0f00000000; @%p75 bra $L__BB70_117; min.ftz.f32 %f420, %f5, %f6; min.ftz.f32 %f421, %f4, %f420; sub.ftz.f32 %f422, %f169, %f421; div.approx.ftz.f32 %f567, %f422, %f169; $L__BB70_117: setp.neu.ftz.f32 %p76, %f567, 0f00000000; mov.f32 %f568, %f169; mov.f32 %f569, %f169; mov.f32 %f570, %f169; @%p76 bra $L__BB70_118; bra.uni $L__BB70_125; $L__BB70_118: setp.eq.ftz.f32 %p77, %f566, 0f3F800000; mov.f32 %f423, 0f3F800000; mul.ftz.f32 %f424, %f566, 0f40C00000; selp.f32 %f425, 0f00000000, %f424, %p77; cvt.rmi.ftz.f32.f32 %f175, %f425; sub.ftz.f32 %f426, %f425, %f175; sub.ftz.f32 %f427, %f423, %f567; mul.ftz.f32 %f176, %f169, %f427; mul.ftz.f32 %f428, %f567, %f426; sub.ftz.f32 %f429, %f423, %f428; mul.ftz.f32 %f177, %f169, %f429; sub.ftz.f32 %f430, %f423, %f426; mul.ftz.f32 %f431, %f567, %f430; sub.ftz.f32 %f432, %f423, %f431; mul.ftz.f32 %f178, %f169, %f432; setp.eq.ftz.f32 %p78, %f175, 0f00000000; @%p78 bra $L__BB70_124; bra.uni $L__BB70_119; $L__BB70_124: mov.f32 %f568, %f169; mov.f32 %f569, %f178; mov.f32 %f570, %f176; bra.uni $L__BB70_125; $L__BB70_119: setp.eq.ftz.f32 %p79, %f175, 0f3F800000; mov.f32 %f568, %f177; mov.f32 %f569, %f169; mov.f32 %f570, %f176; @%p79 bra $L__BB70_125; setp.eq.ftz.f32 %p80, %f175, 0f40000000; mov.f32 %f568, %f176; mov.f32 %f569, %f169; mov.f32 %f570, %f178; @%p80 bra $L__BB70_125; setp.eq.ftz.f32 %p81, %f175, 0f40400000; mov.f32 %f568, %f176; mov.f32 %f569, %f177; mov.f32 %f570, %f169; @%p81 bra $L__BB70_125; setp.eq.ftz.f32 %p82, %f175, 0f40800000; mov.f32 %f568, %f169; mov.f32 %f569, %f176; mov.f32 %f570, %f177; @%p82 bra $L__BB70_123; bra.uni $L__BB70_125; $L__BB70_123: mov.f32 %f568, %f178; mov.f32 %f569, %f176; mov.f32 %f570, %f169; $L__BB70_125: sub.ftz.f32 %f433, %f568, %f4; sub.ftz.f32 %f434, %f569, %f5; sub.ftz.f32 %f435, %f570, %f6; fma.rn.ftz.f32 %f582, %f3, %f433, %f4; fma.rn.ftz.f32 %f583, %f3, %f434, %f5; fma.rn.ftz.f32 %f584, %f3, %f435, %f6; $L__BB70_149: and.b32 %r8, %r1, 16711680; setp.eq.s32 %p107, %r8, 0; @%p107 bra $L__BB70_151; cvt.ftz.sat.f32.f32 %f582, %f582; cvt.ftz.sat.f32.f32 %f583, %f583; cvt.ftz.sat.f32.f32 %f584, %f584; $L__BB70_151: mul.wide.u32 %rd13, %r3, 4; add.s64 %rd14, %rd1, %rd13; st.local.f32 [%rd14], %f582; st.local.f32 [%rd14+4], %f583; st.local.f32 [%rd14+8], %f584; ret; } .func _Z18svm_node_mix_floatP10ShaderDataPfjjj( .param .b64 _Z18svm_node_mix_floatP10ShaderDataPfjjj_param_0, .param .b32 _Z18svm_node_mix_floatP10ShaderDataPfjjj_param_1, .param .b32 _Z18svm_node_mix_floatP10ShaderDataPfjjj_param_2, .param .b32 _Z18svm_node_mix_floatP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<2>; .reg .f32 %f<11>; .reg .b32 %r<7>; .reg .b64 %rd<14>; ld.param.u64 %rd2, [_Z18svm_node_mix_floatP10ShaderDataPfjjj_param_0]; ld.param.u32 %r3, [_Z18svm_node_mix_floatP10ShaderDataPfjjj_param_1]; ld.param.u32 %r1, [_Z18svm_node_mix_floatP10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z18svm_node_mix_floatP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; shl.b32 %r4, %r1, 2; cvt.u64.u32 %rd3, %r4; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f10, [%rd5]; setp.eq.s32 %p1, %r3, 0; @%p1 bra $L__BB71_2; cvt.ftz.sat.f32.f32 %f10, %f10; $L__BB71_2: shr.u32 %r5, %r1, 6; cvt.u64.u32 %rd6, %r5; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; shr.u32 %r6, %r1, 14; cvt.u64.u32 %rd9, %r6; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; mov.f32 %f4, 0f3F800000; sub.ftz.f32 %f5, %f4, %f10; ld.local.f32 %f6, [%rd8]; ld.local.f32 %f7, [%rd11]; mul.ftz.f32 %f8, %f10, %f7; fma.rn.ftz.f32 %f9, %f6, %f5, %f8; mul.wide.u32 %rd12, %r2, 4; add.s64 %rd13, %rd1, %rd12; st.local.f32 [%rd13], %f9; ret; } .func _Z19svm_node_mix_vectorP10ShaderDataPfjj( .param .b64 _Z19svm_node_mix_vectorP10ShaderDataPfjj_param_0, .param .b32 _Z19svm_node_mix_vectorP10ShaderDataPfjj_param_1, .param .b32 _Z19svm_node_mix_vectorP10ShaderDataPfjj_param_2 ) { .reg .pred %p<2>; .reg .f32 %f<19>; .reg .b32 %r<7>; .reg .b64 %rd<13>; ld.param.u64 %rd2, [_Z19svm_node_mix_vectorP10ShaderDataPfjj_param_0]; ld.param.u32 %r1, [_Z19svm_node_mix_vectorP10ShaderDataPfjj_param_1]; ld.param.u32 %r2, [_Z19svm_node_mix_vectorP10ShaderDataPfjj_param_2]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r3, %r1, 6; cvt.u64.u32 %rd3, %r3; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f18, [%rd5]; and.b32 %r4, %r1, 255; setp.eq.s32 %p1, %r4, 0; @%p1 bra $L__BB72_2; cvt.ftz.sat.f32.f32 %f18, %f18; $L__BB72_2: shr.u32 %r5, %r1, 14; cvt.u64.u32 %rd6, %r5; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; shr.u32 %r6, %r1, 24; mul.wide.u32 %rd9, %r6, 4; add.s64 %rd10, %rd1, %rd9; mov.f32 %f4, 0f3F800000; sub.ftz.f32 %f5, %f4, %f18; ld.local.f32 %f6, [%rd8]; mul.ftz.f32 %f7, %f6, %f5; ld.local.f32 %f8, [%rd8+4]; mul.ftz.f32 %f9, %f8, %f5; ld.local.f32 %f10, [%rd8+8]; mul.ftz.f32 %f11, %f10, %f5; ld.local.f32 %f12, [%rd10]; ld.local.f32 %f13, [%rd10+4]; ld.local.f32 %f14, [%rd10+8]; fma.rn.ftz.f32 %f15, %f18, %f12, %f7; fma.rn.ftz.f32 %f16, %f18, %f13, %f9; fma.rn.ftz.f32 %f17, %f18, %f14, %f11; mul.wide.u32 %rd11, %r2, 4; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f15; st.local.f32 [%rd12+4], %f16; st.local.f32 [%rd12+8], %f17; ret; } .func _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj( .param .b64 _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_0, .param .b32 _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_1, .param .b32 _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_2 ) { .reg .pred %p<2>; .reg .f32 %f<29>; .reg .b32 %r<7>; .reg .b64 %rd<13>; ld.param.u64 %rd2, [_Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_0]; ld.param.u32 %r1, [_Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_1]; ld.param.u32 %r2, [_Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj_param_2]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r3, %r1, 6; cvt.u64.u32 %rd3, %r3; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f26, [%rd5]; ld.local.f32 %f27, [%rd5+4]; ld.local.f32 %f28, [%rd5+8]; and.b32 %r4, %r1, 255; setp.eq.s32 %p1, %r4, 0; @%p1 bra $L__BB73_2; cvt.ftz.sat.f32.f32 %f26, %f26; cvt.ftz.sat.f32.f32 %f27, %f27; cvt.ftz.sat.f32.f32 %f28, %f28; $L__BB73_2: shr.u32 %r5, %r1, 14; cvt.u64.u32 %rd6, %r5; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; shr.u32 %r6, %r1, 24; mul.wide.u32 %rd9, %r6, 4; add.s64 %rd10, %rd1, %rd9; mov.f32 %f10, 0f3F800000; sub.ftz.f32 %f11, %f10, %f26; ld.local.f32 %f12, [%rd8]; sub.ftz.f32 %f13, %f10, %f27; ld.local.f32 %f14, [%rd8+4]; sub.ftz.f32 %f15, %f10, %f28; ld.local.f32 %f16, [%rd8+8]; ld.local.f32 %f17, [%rd10]; mul.ftz.f32 %f18, %f26, %f17; ld.local.f32 %f19, [%rd10+4]; mul.ftz.f32 %f20, %f27, %f19; ld.local.f32 %f21, [%rd10+8]; mul.ftz.f32 %f22, %f28, %f21; fma.rn.ftz.f32 %f23, %f11, %f12, %f18; fma.rn.ftz.f32 %f24, %f13, %f14, %f20; fma.rn.ftz.f32 %f25, %f15, %f16, %f22; mul.wide.u32 %rd11, %r2, 4; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f23; st.local.f32 [%rd12+4], %f24; st.local.f32 [%rd12+8], %f25; ret; } .func (.param .b32 func_retval0) _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<1995>; .reg .b16 %rs<275>; .reg .f32 %f<5828>; .reg .b32 %r<10038>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r484, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r487, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r485, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r486, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; and.b32 %r1, %r487, 255; bfe.u32 %r2, %r487, 8, 8; bfe.u32 %r3, %r487, 16, 8; shr.u32 %r4, %r487, 24; and.b32 %r5, %r485, 255; bfe.u32 %r6, %r485, 8, 8; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r486, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r488, %r489, %r490, %r491}, [%rd6]; ld.global.v4.u32 {%r492, %r493, %r494, %r495}, [%rd6+16]; shr.u32 %r496, %r484, 14; cvt.u64.u32 %rd7, %r496; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; shr.u32 %r15, %r484, 24; setp.eq.s32 %p5, %r15, 255; @%p5 bra $L__BB74_2; bra.uni $L__BB74_1; $L__BB74_2: mov.b32 %f6, %r488; bra.uni $L__BB74_3; $L__BB74_1: mul.wide.u32 %rd10, %r15, 4; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f6, [%rd11]; $L__BB74_3: setp.eq.s32 %p6, %r1, 255; @%p6 bra $L__BB74_5; bra.uni $L__BB74_4; $L__BB74_5: mov.b32 %f9, %r489; bra.uni $L__BB74_6; $L__BB74_4: mul.wide.u32 %rd12, %r1, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f9, [%rd13]; $L__BB74_6: setp.eq.s32 %p7, %r2, 255; @%p7 bra $L__BB74_8; bra.uni $L__BB74_7; $L__BB74_8: mov.b32 %f5516, %r490; bra.uni $L__BB74_9; $L__BB74_7: mul.wide.u32 %rd14, %r2, 4; add.s64 %rd15, %rd1, %rd14; ld.local.f32 %f5516, [%rd15]; $L__BB74_9: setp.eq.s32 %p8, %r3, 255; @%p8 bra $L__BB74_11; bra.uni $L__BB74_10; $L__BB74_11: mov.b32 %f5517, %r491; bra.uni $L__BB74_12; $L__BB74_10: mul.wide.u32 %rd16, %r3, 4; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f5517, [%rd17]; $L__BB74_12: setp.eq.s32 %p9, %r4, 255; @%p9 bra $L__BB74_14; bra.uni $L__BB74_13; $L__BB74_14: mov.b32 %f5518, %r492; bra.uni $L__BB74_15; $L__BB74_13: mul.wide.u32 %rd18, %r4, 4; add.s64 %rd19, %rd1, %rd18; ld.local.f32 %f5518, [%rd19]; $L__BB74_15: setp.eq.s32 %p10, %r5, 255; @%p10 bra $L__BB74_17; bra.uni $L__BB74_16; $L__BB74_17: mov.b32 %f21, %r493; bra.uni $L__BB74_18; $L__BB74_16: mul.wide.u32 %rd20, %r5, 4; add.s64 %rd21, %rd1, %rd20; ld.local.f32 %f21, [%rd21]; $L__BB74_18: setp.eq.s32 %p11, %r6, 255; @%p11 bra $L__BB74_20; bra.uni $L__BB74_19; $L__BB74_20: mov.b32 %f24, %r494; bra.uni $L__BB74_21; $L__BB74_19: mul.wide.u32 %rd22, %r6, 4; add.s64 %rd23, %rd1, %rd22; ld.local.f32 %f24, [%rd23]; $L__BB74_21: mov.f32 %f816, 0f3727C5AC; max.ftz.f32 %f25, %f5517, %f816; mov.f32 %f5827, 0f00000000; max.ftz.f32 %f817, %f5516, %f5827; mov.f32 %f818, 0f41700000; min.ftz.f32 %f26, %f817, %f818; max.ftz.f32 %f27, %f5518, %f816; cvt.u16.u32 %rs2, %r484; shr.u16 %rs1, %rs2, 8; setp.gt.s16 %p12, %rs1, 2; @%p12 bra $L__BB74_38; setp.eq.s16 %p15, %rs1, 1; @%p15 bra $L__BB74_388; setp.eq.s16 %p16, %rs1, 2; @%p16 bra $L__BB74_24; bra.uni $L__BB74_450; $L__BB74_24: mul.ftz.f32 %f5771, %f1, %f9; mul.ftz.f32 %f5772, %f2, %f9; and.b16 %rs271, %rs2, 255; setp.gt.s16 %p1675, %rs271, 1; @%p1675 bra $L__BB74_340; bra.uni $L__BB74_25; $L__BB74_340: setp.eq.s16 %p1676, %rs271, 2; @%p1676 bra $L__BB74_366; setp.eq.s16 %p1677, %rs271, 3; @%p1677 bra $L__BB74_357; setp.ne.s16 %p1678, %rs271, 4; @%p1678 bra $L__BB74_450; lg2.approx.ftz.f32 %f4521, %f27; mul.ftz.f32 %f4522, %f25, %f4521; neg.ftz.f32 %f4523, %f4522; ex2.approx.ftz.f32 %f604, %f4523; cvt.rmi.ftz.f32.f32 %f4524, %f5771; cvt.rzi.ftz.s32.f32 %r8453, %f4524; sub.ftz.f32 %f4525, %f5771, %f4524; cvt.rmi.ftz.f32.f32 %f4526, %f5772; cvt.rzi.ftz.s32.f32 %r8454, %f4526; sub.ftz.f32 %f4527, %f5772, %f4526; mul.ftz.f32 %f4528, %f4525, %f4525; mul.ftz.f32 %f4529, %f4525, %f4528; fma.rn.ftz.f32 %f4530, %f4525, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4531, %f4525, %f4530, 0f41200000; mul.ftz.f32 %f4532, %f4529, %f4531; mul.ftz.f32 %f4533, %f4527, %f4527; mul.ftz.f32 %f4534, %f4527, %f4533; fma.rn.ftz.f32 %f4535, %f4527, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4536, %f4527, %f4535, 0f41200000; mul.ftz.f32 %f4537, %f4534, %f4536; add.s32 %r8455, %r8454, -559038716; add.s32 %r8456, %r8453, -559038716; xor.b32 %r8457, %r8455, -559038716; shf.l.wrap.b32 %r8458, %r8455, %r8455, 14; sub.s32 %r8459, %r8457, %r8458; xor.b32 %r8460, %r8459, %r8456; shf.l.wrap.b32 %r8461, %r8459, %r8459, 11; sub.s32 %r8462, %r8460, %r8461; xor.b32 %r8463, %r8462, %r8455; shf.l.wrap.b32 %r8464, %r8462, %r8462, 25; sub.s32 %r8465, %r8463, %r8464; xor.b32 %r8466, %r8465, %r8459; shf.l.wrap.b32 %r8467, %r8465, %r8465, 16; sub.s32 %r8468, %r8466, %r8467; xor.b32 %r8469, %r8468, %r8462; shf.l.wrap.b32 %r8470, %r8468, %r8468, 4; sub.s32 %r8471, %r8469, %r8470; xor.b32 %r8472, %r8471, %r8465; shf.l.wrap.b32 %r8473, %r8471, %r8471, 14; sub.s32 %r8474, %r8472, %r8473; xor.b32 %r8475, %r8474, %r8468; shf.l.wrap.b32 %r8476, %r8474, %r8474, 24; sub.s32 %r8477, %r8475, %r8476; and.b32 %r8478, %r8477, 4; setp.eq.s32 %p1681, %r8478, 0; selp.f32 %f4538, %f4525, %f4527, %p1681; selp.f32 %f4539, %f4527, %f4525, %p1681; add.ftz.f32 %f4540, %f4539, %f4539; and.b32 %r8479, %r8477, 1; setp.eq.b32 %p1682, %r8479, 1; neg.ftz.f32 %f4541, %f4538; selp.f32 %f4542, %f4541, %f4538, %p1682; and.b32 %r8480, %r8477, 2; setp.eq.s32 %p1683, %r8480, 0; neg.ftz.f32 %f4543, %f4540; selp.f32 %f4544, %f4540, %f4543, %p1683; add.ftz.f32 %f4545, %f4542, %f4544; add.s32 %r8481, %r8453, -559038715; xor.b32 %r8482, %r8459, %r8481; sub.s32 %r8483, %r8482, %r8461; xor.b32 %r8484, %r8483, %r8455; shf.l.wrap.b32 %r8485, %r8483, %r8483, 25; sub.s32 %r8486, %r8484, %r8485; xor.b32 %r8487, %r8486, %r8459; shf.l.wrap.b32 %r8488, %r8486, %r8486, 16; sub.s32 %r8489, %r8487, %r8488; xor.b32 %r8490, %r8489, %r8483; shf.l.wrap.b32 %r8491, %r8489, %r8489, 4; sub.s32 %r8492, %r8490, %r8491; xor.b32 %r8493, %r8492, %r8486; shf.l.wrap.b32 %r8494, %r8492, %r8492, 14; sub.s32 %r8495, %r8493, %r8494; xor.b32 %r8496, %r8495, %r8489; shf.l.wrap.b32 %r8497, %r8495, %r8495, 24; sub.s32 %r8498, %r8496, %r8497; add.ftz.f32 %f4546, %f4525, 0fBF800000; mov.f32 %f4547, 0f3F800000; and.b32 %r8499, %r8498, 4; setp.eq.s32 %p1684, %r8499, 0; selp.f32 %f4548, %f4546, %f4527, %p1684; selp.f32 %f4549, %f4527, %f4546, %p1684; add.ftz.f32 %f4550, %f4549, %f4549; and.b32 %r8500, %r8498, 1; setp.eq.b32 %p1685, %r8500, 1; neg.ftz.f32 %f4551, %f4548; selp.f32 %f4552, %f4551, %f4548, %p1685; and.b32 %r8501, %r8498, 2; setp.eq.s32 %p1686, %r8501, 0; neg.ftz.f32 %f4553, %f4550; selp.f32 %f4554, %f4550, %f4553, %p1686; add.ftz.f32 %f4555, %f4552, %f4554; add.s32 %r8502, %r8454, -559038715; xor.b32 %r8503, %r8502, -559038716; shf.l.wrap.b32 %r8504, %r8502, %r8502, 14; sub.s32 %r8505, %r8503, %r8504; xor.b32 %r8506, %r8505, %r8456; shf.l.wrap.b32 %r8507, %r8505, %r8505, 11; sub.s32 %r8508, %r8506, %r8507; xor.b32 %r8509, %r8508, %r8502; shf.l.wrap.b32 %r8510, %r8508, %r8508, 25; sub.s32 %r8511, %r8509, %r8510; xor.b32 %r8512, %r8511, %r8505; shf.l.wrap.b32 %r8513, %r8511, %r8511, 16; sub.s32 %r8514, %r8512, %r8513; xor.b32 %r8515, %r8514, %r8508; shf.l.wrap.b32 %r8516, %r8514, %r8514, 4; sub.s32 %r8517, %r8515, %r8516; xor.b32 %r8518, %r8517, %r8511; shf.l.wrap.b32 %r8519, %r8517, %r8517, 14; sub.s32 %r8520, %r8518, %r8519; xor.b32 %r8521, %r8520, %r8514; shf.l.wrap.b32 %r8522, %r8520, %r8520, 24; sub.s32 %r8523, %r8521, %r8522; add.ftz.f32 %f4556, %f4527, 0fBF800000; and.b32 %r8524, %r8523, 4; setp.eq.s32 %p1687, %r8524, 0; selp.f32 %f4557, %f4525, %f4556, %p1687; selp.f32 %f4558, %f4556, %f4525, %p1687; add.ftz.f32 %f4559, %f4558, %f4558; and.b32 %r8525, %r8523, 1; setp.eq.b32 %p1688, %r8525, 1; neg.ftz.f32 %f4560, %f4557; selp.f32 %f4561, %f4560, %f4557, %p1688; and.b32 %r8526, %r8523, 2; setp.eq.s32 %p1689, %r8526, 0; neg.ftz.f32 %f4562, %f4559; selp.f32 %f4563, %f4559, %f4562, %p1689; add.ftz.f32 %f4564, %f4561, %f4563; xor.b32 %r8527, %r8505, %r8481; sub.s32 %r8528, %r8527, %r8507; xor.b32 %r8529, %r8528, %r8502; shf.l.wrap.b32 %r8530, %r8528, %r8528, 25; sub.s32 %r8531, %r8529, %r8530; xor.b32 %r8532, %r8531, %r8505; shf.l.wrap.b32 %r8533, %r8531, %r8531, 16; sub.s32 %r8534, %r8532, %r8533; xor.b32 %r8535, %r8534, %r8528; shf.l.wrap.b32 %r8536, %r8534, %r8534, 4; sub.s32 %r8537, %r8535, %r8536; xor.b32 %r8538, %r8537, %r8531; shf.l.wrap.b32 %r8539, %r8537, %r8537, 14; sub.s32 %r8540, %r8538, %r8539; xor.b32 %r8541, %r8540, %r8534; shf.l.wrap.b32 %r8542, %r8540, %r8540, 24; sub.s32 %r8543, %r8541, %r8542; and.b32 %r8544, %r8543, 4; setp.eq.s32 %p1690, %r8544, 0; selp.f32 %f4565, %f4546, %f4556, %p1690; selp.f32 %f4566, %f4556, %f4546, %p1690; add.ftz.f32 %f4567, %f4566, %f4566; and.b32 %r8545, %r8543, 1; setp.eq.b32 %p1691, %r8545, 1; neg.ftz.f32 %f4568, %f4565; selp.f32 %f4569, %f4568, %f4565, %p1691; and.b32 %r8546, %r8543, 2; setp.eq.s32 %p1692, %r8546, 0; neg.ftz.f32 %f4570, %f4567; selp.f32 %f4571, %f4567, %f4570, %p1692; add.ftz.f32 %f4572, %f4569, %f4571; sub.ftz.f32 %f4573, %f4547, %f4532; sub.ftz.f32 %f4574, %f4547, %f4537; mul.ftz.f32 %f4575, %f4573, %f4545; fma.rn.ftz.f32 %f4576, %f4532, %f4555, %f4575; mul.ftz.f32 %f4577, %f4573, %f4564; fma.rn.ftz.f32 %f4578, %f4532, %f4572, %f4577; mul.ftz.f32 %f4579, %f4537, %f4578; fma.rn.ftz.f32 %f605, %f4574, %f4576, %f4579; mov.b32 %r432, %f605; setp.nan.ftz.f32 %p1693, %f605, %f605; mov.f32 %f5736, 0f00000000; @%p1693 bra $L__BB74_346; or.b32 %r8547, %r432, -2147483648; setp.ne.s32 %p1694, %r8547, -2147483648; add.ftz.f32 %f4581, %f605, %f605; setp.eq.ftz.f32 %p1695, %f4581, %f605; and.pred %p1696, %p1695, %p1694; @%p1696 bra $L__BB74_346; shl.b32 %r8548, %r432, 1; setp.lt.u32 %p1697, %r8548, -16777215; mul.ftz.f32 %f4582, %f605, 0f3F295E9E; selp.f32 %f5736, %f4582, 0f00000000, %p1697; $L__BB74_346: add.ftz.f32 %f5827, %f21, %f5736; mul.ftz.f32 %f5742, %f5771, %f27; mul.ftz.f32 %f5743, %f5772, %f27; cvt.rzi.ftz.s32.f32 %r433, %f26; setp.lt.s32 %p1698, %r433, 2; mov.f32 %f5744, %f604; @%p1698 bra $L__BB74_352; mov.u32 %r10028, 1; mov.f32 %f5744, %f604; $L__BB74_348: cvt.rmi.ftz.f32.f32 %f4584, %f5742; cvt.rzi.ftz.s32.f32 %r8550, %f4584; sub.ftz.f32 %f4585, %f5742, %f4584; cvt.rmi.ftz.f32.f32 %f4586, %f5743; cvt.rzi.ftz.s32.f32 %r8551, %f4586; sub.ftz.f32 %f4587, %f5743, %f4586; mul.ftz.f32 %f4588, %f4585, %f4585; mul.ftz.f32 %f4589, %f4585, %f4588; fma.rn.ftz.f32 %f4590, %f4585, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4591, %f4585, %f4590, 0f41200000; mul.ftz.f32 %f4592, %f4589, %f4591; mul.ftz.f32 %f4593, %f4587, %f4587; mul.ftz.f32 %f4594, %f4587, %f4593; fma.rn.ftz.f32 %f4595, %f4587, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4596, %f4587, %f4595, 0f41200000; mul.ftz.f32 %f4597, %f4594, %f4596; add.s32 %r8552, %r8551, -559038716; add.s32 %r8553, %r8550, -559038716; xor.b32 %r8554, %r8552, -559038716; shf.l.wrap.b32 %r8555, %r8552, %r8552, 14; sub.s32 %r8556, %r8554, %r8555; xor.b32 %r8557, %r8556, %r8553; shf.l.wrap.b32 %r8558, %r8556, %r8556, 11; sub.s32 %r8559, %r8557, %r8558; xor.b32 %r8560, %r8559, %r8552; shf.l.wrap.b32 %r8561, %r8559, %r8559, 25; sub.s32 %r8562, %r8560, %r8561; xor.b32 %r8563, %r8562, %r8556; shf.l.wrap.b32 %r8564, %r8562, %r8562, 16; sub.s32 %r8565, %r8563, %r8564; xor.b32 %r8566, %r8565, %r8559; shf.l.wrap.b32 %r8567, %r8565, %r8565, 4; sub.s32 %r8568, %r8566, %r8567; xor.b32 %r8569, %r8568, %r8562; shf.l.wrap.b32 %r8570, %r8568, %r8568, 14; sub.s32 %r8571, %r8569, %r8570; xor.b32 %r8572, %r8571, %r8565; shf.l.wrap.b32 %r8573, %r8571, %r8571, 24; sub.s32 %r8574, %r8572, %r8573; and.b32 %r8575, %r8574, 4; setp.eq.s32 %p1699, %r8575, 0; selp.f32 %f4598, %f4585, %f4587, %p1699; selp.f32 %f4599, %f4587, %f4585, %p1699; add.ftz.f32 %f4600, %f4599, %f4599; and.b32 %r8576, %r8574, 1; setp.eq.b32 %p1700, %r8576, 1; neg.ftz.f32 %f4601, %f4598; selp.f32 %f4602, %f4601, %f4598, %p1700; and.b32 %r8577, %r8574, 2; setp.eq.s32 %p1701, %r8577, 0; neg.ftz.f32 %f4603, %f4600; selp.f32 %f4604, %f4600, %f4603, %p1701; add.ftz.f32 %f4605, %f4602, %f4604; add.s32 %r8578, %r8550, -559038715; xor.b32 %r8579, %r8556, %r8578; sub.s32 %r8580, %r8579, %r8558; xor.b32 %r8581, %r8580, %r8552; shf.l.wrap.b32 %r8582, %r8580, %r8580, 25; sub.s32 %r8583, %r8581, %r8582; xor.b32 %r8584, %r8583, %r8556; shf.l.wrap.b32 %r8585, %r8583, %r8583, 16; sub.s32 %r8586, %r8584, %r8585; xor.b32 %r8587, %r8586, %r8580; shf.l.wrap.b32 %r8588, %r8586, %r8586, 4; sub.s32 %r8589, %r8587, %r8588; xor.b32 %r8590, %r8589, %r8583; shf.l.wrap.b32 %r8591, %r8589, %r8589, 14; sub.s32 %r8592, %r8590, %r8591; xor.b32 %r8593, %r8592, %r8586; shf.l.wrap.b32 %r8594, %r8592, %r8592, 24; sub.s32 %r8595, %r8593, %r8594; add.ftz.f32 %f4606, %f4585, 0fBF800000; and.b32 %r8596, %r8595, 4; setp.eq.s32 %p1702, %r8596, 0; selp.f32 %f4608, %f4606, %f4587, %p1702; selp.f32 %f4609, %f4587, %f4606, %p1702; add.ftz.f32 %f4610, %f4609, %f4609; and.b32 %r8597, %r8595, 1; setp.eq.b32 %p1703, %r8597, 1; neg.ftz.f32 %f4611, %f4608; selp.f32 %f4612, %f4611, %f4608, %p1703; and.b32 %r8598, %r8595, 2; setp.eq.s32 %p1704, %r8598, 0; neg.ftz.f32 %f4613, %f4610; selp.f32 %f4614, %f4610, %f4613, %p1704; add.ftz.f32 %f4615, %f4612, %f4614; add.s32 %r8599, %r8551, -559038715; xor.b32 %r8600, %r8599, -559038716; shf.l.wrap.b32 %r8601, %r8599, %r8599, 14; sub.s32 %r8602, %r8600, %r8601; xor.b32 %r8603, %r8602, %r8553; shf.l.wrap.b32 %r8604, %r8602, %r8602, 11; sub.s32 %r8605, %r8603, %r8604; xor.b32 %r8606, %r8605, %r8599; shf.l.wrap.b32 %r8607, %r8605, %r8605, 25; sub.s32 %r8608, %r8606, %r8607; xor.b32 %r8609, %r8608, %r8602; shf.l.wrap.b32 %r8610, %r8608, %r8608, 16; sub.s32 %r8611, %r8609, %r8610; xor.b32 %r8612, %r8611, %r8605; shf.l.wrap.b32 %r8613, %r8611, %r8611, 4; sub.s32 %r8614, %r8612, %r8613; xor.b32 %r8615, %r8614, %r8608; shf.l.wrap.b32 %r8616, %r8614, %r8614, 14; sub.s32 %r8617, %r8615, %r8616; xor.b32 %r8618, %r8617, %r8611; shf.l.wrap.b32 %r8619, %r8617, %r8617, 24; sub.s32 %r8620, %r8618, %r8619; add.ftz.f32 %f4616, %f4587, 0fBF800000; and.b32 %r8621, %r8620, 4; setp.eq.s32 %p1705, %r8621, 0; selp.f32 %f4617, %f4585, %f4616, %p1705; selp.f32 %f4618, %f4616, %f4585, %p1705; add.ftz.f32 %f4619, %f4618, %f4618; and.b32 %r8622, %r8620, 1; setp.eq.b32 %p1706, %r8622, 1; neg.ftz.f32 %f4620, %f4617; selp.f32 %f4621, %f4620, %f4617, %p1706; and.b32 %r8623, %r8620, 2; setp.eq.s32 %p1707, %r8623, 0; neg.ftz.f32 %f4622, %f4619; selp.f32 %f4623, %f4619, %f4622, %p1707; add.ftz.f32 %f4624, %f4621, %f4623; xor.b32 %r8624, %r8602, %r8578; sub.s32 %r8625, %r8624, %r8604; xor.b32 %r8626, %r8625, %r8599; shf.l.wrap.b32 %r8627, %r8625, %r8625, 25; sub.s32 %r8628, %r8626, %r8627; xor.b32 %r8629, %r8628, %r8602; shf.l.wrap.b32 %r8630, %r8628, %r8628, 16; sub.s32 %r8631, %r8629, %r8630; xor.b32 %r8632, %r8631, %r8625; shf.l.wrap.b32 %r8633, %r8631, %r8631, 4; sub.s32 %r8634, %r8632, %r8633; xor.b32 %r8635, %r8634, %r8628; shf.l.wrap.b32 %r8636, %r8634, %r8634, 14; sub.s32 %r8637, %r8635, %r8636; xor.b32 %r8638, %r8637, %r8631; shf.l.wrap.b32 %r8639, %r8637, %r8637, 24; sub.s32 %r8640, %r8638, %r8639; and.b32 %r8641, %r8640, 4; setp.eq.s32 %p1708, %r8641, 0; selp.f32 %f4625, %f4606, %f4616, %p1708; selp.f32 %f4626, %f4616, %f4606, %p1708; add.ftz.f32 %f4627, %f4626, %f4626; and.b32 %r8642, %r8640, 1; setp.eq.b32 %p1709, %r8642, 1; neg.ftz.f32 %f4628, %f4625; selp.f32 %f4629, %f4628, %f4625, %p1709; and.b32 %r8643, %r8640, 2; setp.eq.s32 %p1710, %r8643, 0; neg.ftz.f32 %f4630, %f4627; selp.f32 %f4631, %f4627, %f4630, %p1710; add.ftz.f32 %f4632, %f4629, %f4631; sub.ftz.f32 %f4633, %f4547, %f4592; sub.ftz.f32 %f4634, %f4547, %f4597; mul.ftz.f32 %f4635, %f4633, %f4605; fma.rn.ftz.f32 %f4636, %f4592, %f4615, %f4635; mul.ftz.f32 %f4637, %f4633, %f4624; fma.rn.ftz.f32 %f4638, %f4592, %f4632, %f4637; mul.ftz.f32 %f4639, %f4597, %f4638; fma.rn.ftz.f32 %f615, %f4634, %f4636, %f4639; mov.b32 %r435, %f615; setp.nan.ftz.f32 %p1711, %f615, %f615; mov.f32 %f5741, 0f00000000; @%p1711 bra $L__BB74_351; or.b32 %r8644, %r435, -2147483648; setp.ne.s32 %p1712, %r8644, -2147483648; add.ftz.f32 %f4641, %f615, %f615; setp.eq.ftz.f32 %p1713, %f4641, %f615; and.pred %p1714, %p1713, %p1712; @%p1714 bra $L__BB74_351; shl.b32 %r8645, %r435, 1; setp.lt.u32 %p1715, %r8645, -16777215; selp.f32 %f5741, %f615, 0f00000000, %p1715; $L__BB74_351: fma.rn.ftz.f32 %f4642, %f5741, 0f3F295E9E, %f21; mul.ftz.f32 %f4643, %f5744, %f4642; fma.rn.ftz.f32 %f5827, %f5827, %f4643, %f5827; mul.ftz.f32 %f5744, %f604, %f5744; mul.ftz.f32 %f5742, %f27, %f5742; mul.ftz.f32 %f5743, %f27, %f5743; add.s32 %r10028, %r10028, 1; setp.lt.s32 %p1716, %r10028, %r433; @%p1716 bra $L__BB74_348; $L__BB74_352: cvt.rmi.ftz.f32.f32 %f4644, %f26; sub.ftz.f32 %f626, %f26, %f4644; setp.eq.ftz.f32 %p1717, %f626, 0f00000000; @%p1717 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f4646, %f5742; cvt.rzi.ftz.s32.f32 %r8646, %f4646; sub.ftz.f32 %f4647, %f5742, %f4646; cvt.rmi.ftz.f32.f32 %f4648, %f5743; cvt.rzi.ftz.s32.f32 %r8647, %f4648; sub.ftz.f32 %f4649, %f5743, %f4648; mul.ftz.f32 %f4650, %f4647, %f4647; mul.ftz.f32 %f4651, %f4647, %f4650; fma.rn.ftz.f32 %f4652, %f4647, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4653, %f4647, %f4652, 0f41200000; mul.ftz.f32 %f4654, %f4651, %f4653; mul.ftz.f32 %f4655, %f4649, %f4649; mul.ftz.f32 %f4656, %f4649, %f4655; fma.rn.ftz.f32 %f4657, %f4649, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4658, %f4649, %f4657, 0f41200000; mul.ftz.f32 %f4659, %f4656, %f4658; add.s32 %r8648, %r8647, -559038716; add.s32 %r8649, %r8646, -559038716; xor.b32 %r8650, %r8648, -559038716; shf.l.wrap.b32 %r8651, %r8648, %r8648, 14; sub.s32 %r8652, %r8650, %r8651; xor.b32 %r8653, %r8652, %r8649; shf.l.wrap.b32 %r8654, %r8652, %r8652, 11; sub.s32 %r8655, %r8653, %r8654; xor.b32 %r8656, %r8655, %r8648; shf.l.wrap.b32 %r8657, %r8655, %r8655, 25; sub.s32 %r8658, %r8656, %r8657; xor.b32 %r8659, %r8658, %r8652; shf.l.wrap.b32 %r8660, %r8658, %r8658, 16; sub.s32 %r8661, %r8659, %r8660; xor.b32 %r8662, %r8661, %r8655; shf.l.wrap.b32 %r8663, %r8661, %r8661, 4; sub.s32 %r8664, %r8662, %r8663; xor.b32 %r8665, %r8664, %r8658; shf.l.wrap.b32 %r8666, %r8664, %r8664, 14; sub.s32 %r8667, %r8665, %r8666; xor.b32 %r8668, %r8667, %r8661; shf.l.wrap.b32 %r8669, %r8667, %r8667, 24; sub.s32 %r8670, %r8668, %r8669; and.b32 %r8671, %r8670, 4; setp.eq.s32 %p1718, %r8671, 0; selp.f32 %f4660, %f4647, %f4649, %p1718; selp.f32 %f4661, %f4649, %f4647, %p1718; add.ftz.f32 %f4662, %f4661, %f4661; and.b32 %r8672, %r8670, 1; setp.eq.b32 %p1719, %r8672, 1; neg.ftz.f32 %f4663, %f4660; selp.f32 %f4664, %f4663, %f4660, %p1719; and.b32 %r8673, %r8670, 2; setp.eq.s32 %p1720, %r8673, 0; neg.ftz.f32 %f4665, %f4662; selp.f32 %f4666, %f4662, %f4665, %p1720; add.ftz.f32 %f4667, %f4664, %f4666; add.s32 %r8674, %r8646, -559038715; xor.b32 %r8675, %r8652, %r8674; sub.s32 %r8676, %r8675, %r8654; xor.b32 %r8677, %r8676, %r8648; shf.l.wrap.b32 %r8678, %r8676, %r8676, 25; sub.s32 %r8679, %r8677, %r8678; xor.b32 %r8680, %r8679, %r8652; shf.l.wrap.b32 %r8681, %r8679, %r8679, 16; sub.s32 %r8682, %r8680, %r8681; xor.b32 %r8683, %r8682, %r8676; shf.l.wrap.b32 %r8684, %r8682, %r8682, 4; sub.s32 %r8685, %r8683, %r8684; xor.b32 %r8686, %r8685, %r8679; shf.l.wrap.b32 %r8687, %r8685, %r8685, 14; sub.s32 %r8688, %r8686, %r8687; xor.b32 %r8689, %r8688, %r8682; shf.l.wrap.b32 %r8690, %r8688, %r8688, 24; sub.s32 %r8691, %r8689, %r8690; add.ftz.f32 %f4668, %f4647, 0fBF800000; and.b32 %r8692, %r8691, 4; setp.eq.s32 %p1721, %r8692, 0; selp.f32 %f4670, %f4668, %f4649, %p1721; selp.f32 %f4671, %f4649, %f4668, %p1721; add.ftz.f32 %f4672, %f4671, %f4671; and.b32 %r8693, %r8691, 1; setp.eq.b32 %p1722, %r8693, 1; neg.ftz.f32 %f4673, %f4670; selp.f32 %f4674, %f4673, %f4670, %p1722; and.b32 %r8694, %r8691, 2; setp.eq.s32 %p1723, %r8694, 0; neg.ftz.f32 %f4675, %f4672; selp.f32 %f4676, %f4672, %f4675, %p1723; add.ftz.f32 %f4677, %f4674, %f4676; add.s32 %r8695, %r8647, -559038715; xor.b32 %r8696, %r8695, -559038716; shf.l.wrap.b32 %r8697, %r8695, %r8695, 14; sub.s32 %r8698, %r8696, %r8697; xor.b32 %r8699, %r8698, %r8649; shf.l.wrap.b32 %r8700, %r8698, %r8698, 11; sub.s32 %r8701, %r8699, %r8700; xor.b32 %r8702, %r8701, %r8695; shf.l.wrap.b32 %r8703, %r8701, %r8701, 25; sub.s32 %r8704, %r8702, %r8703; xor.b32 %r8705, %r8704, %r8698; shf.l.wrap.b32 %r8706, %r8704, %r8704, 16; sub.s32 %r8707, %r8705, %r8706; xor.b32 %r8708, %r8707, %r8701; shf.l.wrap.b32 %r8709, %r8707, %r8707, 4; sub.s32 %r8710, %r8708, %r8709; xor.b32 %r8711, %r8710, %r8704; shf.l.wrap.b32 %r8712, %r8710, %r8710, 14; sub.s32 %r8713, %r8711, %r8712; xor.b32 %r8714, %r8713, %r8707; shf.l.wrap.b32 %r8715, %r8713, %r8713, 24; sub.s32 %r8716, %r8714, %r8715; add.ftz.f32 %f4678, %f4649, 0fBF800000; and.b32 %r8717, %r8716, 4; setp.eq.s32 %p1724, %r8717, 0; selp.f32 %f4679, %f4647, %f4678, %p1724; selp.f32 %f4680, %f4678, %f4647, %p1724; add.ftz.f32 %f4681, %f4680, %f4680; and.b32 %r8718, %r8716, 1; setp.eq.b32 %p1725, %r8718, 1; neg.ftz.f32 %f4682, %f4679; selp.f32 %f4683, %f4682, %f4679, %p1725; and.b32 %r8719, %r8716, 2; setp.eq.s32 %p1726, %r8719, 0; neg.ftz.f32 %f4684, %f4681; selp.f32 %f4685, %f4681, %f4684, %p1726; add.ftz.f32 %f4686, %f4683, %f4685; xor.b32 %r8720, %r8698, %r8674; sub.s32 %r8721, %r8720, %r8700; xor.b32 %r8722, %r8721, %r8695; shf.l.wrap.b32 %r8723, %r8721, %r8721, 25; sub.s32 %r8724, %r8722, %r8723; xor.b32 %r8725, %r8724, %r8698; shf.l.wrap.b32 %r8726, %r8724, %r8724, 16; sub.s32 %r8727, %r8725, %r8726; xor.b32 %r8728, %r8727, %r8721; shf.l.wrap.b32 %r8729, %r8727, %r8727, 4; sub.s32 %r8730, %r8728, %r8729; xor.b32 %r8731, %r8730, %r8724; shf.l.wrap.b32 %r8732, %r8730, %r8730, 14; sub.s32 %r8733, %r8731, %r8732; xor.b32 %r8734, %r8733, %r8727; shf.l.wrap.b32 %r8735, %r8733, %r8733, 24; sub.s32 %r8736, %r8734, %r8735; and.b32 %r8737, %r8736, 4; setp.eq.s32 %p1727, %r8737, 0; selp.f32 %f4687, %f4668, %f4678, %p1727; selp.f32 %f4688, %f4678, %f4668, %p1727; add.ftz.f32 %f4689, %f4688, %f4688; and.b32 %r8738, %r8736, 1; setp.eq.b32 %p1728, %r8738, 1; neg.ftz.f32 %f4690, %f4687; selp.f32 %f4691, %f4690, %f4687, %p1728; and.b32 %r8739, %r8736, 2; setp.eq.s32 %p1729, %r8739, 0; neg.ftz.f32 %f4692, %f4689; selp.f32 %f4693, %f4689, %f4692, %p1729; add.ftz.f32 %f4694, %f4691, %f4693; sub.ftz.f32 %f4695, %f4547, %f4654; sub.ftz.f32 %f4696, %f4547, %f4659; mul.ftz.f32 %f4697, %f4695, %f4667; fma.rn.ftz.f32 %f4698, %f4654, %f4677, %f4697; mul.ftz.f32 %f4699, %f4695, %f4686; fma.rn.ftz.f32 %f4700, %f4654, %f4694, %f4699; mul.ftz.f32 %f4701, %f4659, %f4700; fma.rn.ftz.f32 %f627, %f4696, %f4698, %f4701; mov.b32 %r437, %f627; setp.nan.ftz.f32 %p1730, %f627, %f627; mov.f32 %f5746, 0f00000000; @%p1730 bra $L__BB74_356; or.b32 %r8740, %r437, -2147483648; setp.ne.s32 %p1731, %r8740, -2147483648; add.ftz.f32 %f4703, %f627, %f627; setp.eq.ftz.f32 %p1732, %f4703, %f627; and.pred %p1733, %p1732, %p1731; @%p1733 bra $L__BB74_356; shl.b32 %r8741, %r437, 1; setp.lt.u32 %p1734, %r8741, -16777215; mul.ftz.f32 %f4704, %f627, 0f3F295E9E; selp.f32 %f5746, %f4704, 0f00000000, %p1734; $L__BB74_356: add.ftz.f32 %f4705, %f21, %f5746; mul.ftz.f32 %f4706, %f5744, %f4705; mul.ftz.f32 %f4707, %f5827, %f4706; fma.rn.ftz.f32 %f5827, %f626, %f4707, %f5827; bra.uni $L__BB74_450; $L__BB74_38: setp.eq.s16 %p13, %rs1, 3; @%p13 bra $L__BB74_102; setp.ne.s16 %p14, %rs1, 4; @%p14 bra $L__BB74_450; mul.ftz.f32 %f5568, %f1, %f9; mul.ftz.f32 %f5569, %f2, %f9; mul.ftz.f32 %f5570, %f3, %f9; mul.ftz.f32 %f5571, %f6, %f9; and.b16 %rs3, %rs2, 255; setp.gt.s16 %p17, %rs3, 1; @%p17 bra $L__BB74_54; bra.uni $L__BB74_41; $L__BB74_54: setp.eq.s16 %p18, %rs3, 2; @%p18 bra $L__BB74_80; setp.eq.s16 %p19, %rs3, 3; @%p19 bra $L__BB74_71; setp.ne.s16 %p20, %rs3, 4; @%p20 bra $L__BB74_450; lg2.approx.ftz.f32 %f821, %f27; mul.ftz.f32 %f822, %f25, %f821; neg.ftz.f32 %f823, %f822; ex2.approx.ftz.f32 %f32, %f823; cvt.rmi.ftz.f32.f32 %f824, %f5568; cvt.rzi.ftz.s32.f32 %r497, %f824; sub.ftz.f32 %f825, %f5568, %f824; cvt.rmi.ftz.f32.f32 %f826, %f5569; cvt.rzi.ftz.s32.f32 %r498, %f826; sub.ftz.f32 %f827, %f5569, %f826; cvt.rmi.ftz.f32.f32 %f828, %f5570; cvt.rzi.ftz.s32.f32 %r499, %f828; sub.ftz.f32 %f829, %f5570, %f828; cvt.rmi.ftz.f32.f32 %f830, %f5571; cvt.rzi.ftz.s32.f32 %r500, %f830; sub.ftz.f32 %f831, %f5571, %f830; mul.ftz.f32 %f832, %f825, %f825; mul.ftz.f32 %f833, %f825, %f832; fma.rn.ftz.f32 %f834, %f825, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f835, %f825, %f834, 0f41200000; mul.ftz.f32 %f836, %f833, %f835; mul.ftz.f32 %f837, %f827, %f827; mul.ftz.f32 %f838, %f827, %f837; fma.rn.ftz.f32 %f839, %f827, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f840, %f827, %f839, 0f41200000; mul.ftz.f32 %f841, %f838, %f840; mul.ftz.f32 %f842, %f829, %f829; mul.ftz.f32 %f843, %f829, %f842; fma.rn.ftz.f32 %f844, %f829, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f845, %f829, %f844, 0f41200000; mul.ftz.f32 %f846, %f843, %f845; mul.ftz.f32 %f847, %f831, %f831; mul.ftz.f32 %f848, %f831, %f847; fma.rn.ftz.f32 %f849, %f831, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f850, %f831, %f849, 0f41200000; mul.ftz.f32 %f851, %f848, %f850; add.s32 %r501, %r498, -559038708; add.s32 %r502, %r499, -559038708; sub.s32 %r503, %r497, %r499; shf.l.wrap.b32 %r504, %r502, %r502, 4; xor.b32 %r505, %r504, %r503; add.s32 %r506, %r502, %r501; sub.s32 %r507, %r501, %r505; shf.l.wrap.b32 %r508, %r505, %r505, 6; xor.b32 %r509, %r508, %r507; add.s32 %r510, %r505, %r506; sub.s32 %r511, %r506, %r509; shf.l.wrap.b32 %r512, %r509, %r509, 8; xor.b32 %r513, %r512, %r511; add.s32 %r514, %r509, %r510; sub.s32 %r515, %r510, %r513; shf.l.wrap.b32 %r516, %r513, %r513, 16; xor.b32 %r517, %r516, %r515; add.s32 %r518, %r513, %r514; sub.s32 %r519, %r514, %r517; shf.l.wrap.b32 %r520, %r517, %r517, 19; xor.b32 %r521, %r520, %r519; add.s32 %r522, %r517, %r518; sub.s32 %r523, %r518, %r521; shf.l.wrap.b32 %r524, %r521, %r521, 4; add.s32 %r525, %r521, %r522; add.s32 %r526, %r522, %r500; xor.b32 %r527, %r525, %r523; xor.b32 %r528, %r527, %r524; shf.l.wrap.b32 %r529, %r525, %r525, 14; sub.s32 %r530, %r528, %r529; xor.b32 %r531, %r530, %r526; shf.l.wrap.b32 %r532, %r530, %r530, 11; sub.s32 %r533, %r531, %r532; xor.b32 %r534, %r533, %r525; shf.l.wrap.b32 %r535, %r533, %r533, 25; sub.s32 %r536, %r534, %r535; xor.b32 %r537, %r536, %r530; shf.l.wrap.b32 %r538, %r536, %r536, 16; sub.s32 %r539, %r537, %r538; xor.b32 %r540, %r539, %r533; shf.l.wrap.b32 %r541, %r539, %r539, 4; sub.s32 %r542, %r540, %r541; xor.b32 %r543, %r542, %r536; shf.l.wrap.b32 %r544, %r542, %r542, 14; sub.s32 %r545, %r543, %r544; xor.b32 %r546, %r545, %r539; shf.l.wrap.b32 %r547, %r545, %r545, 24; sub.s32 %r548, %r546, %r547; and.b32 %r549, %r548, 31; setp.lt.u32 %p23, %r549, 24; selp.f32 %f852, %f825, %f827, %p23; setp.lt.u32 %p24, %r549, 16; selp.f32 %f853, %f827, %f829, %p24; setp.lt.u32 %p25, %r549, 8; selp.f32 %f854, %f829, %f831, %p25; and.b32 %r550, %r548, 1; setp.eq.b32 %p26, %r550, 1; neg.ftz.f32 %f855, %f852; selp.f32 %f856, %f855, %f852, %p26; and.b32 %r551, %r548, 2; setp.eq.s32 %p27, %r551, 0; neg.ftz.f32 %f857, %f853; selp.f32 %f858, %f853, %f857, %p27; add.ftz.f32 %f859, %f856, %f858; and.b32 %r552, %r548, 4; setp.eq.s32 %p28, %r552, 0; neg.ftz.f32 %f860, %f854; selp.f32 %f861, %f854, %f860, %p28; add.ftz.f32 %f862, %f861, %f859; add.s32 %r553, %r497, 1; sub.s32 %r554, %r553, %r499; xor.b32 %r555, %r504, %r554; sub.s32 %r556, %r501, %r555; shf.l.wrap.b32 %r557, %r555, %r555, 6; xor.b32 %r558, %r557, %r556; add.s32 %r559, %r555, %r506; sub.s32 %r560, %r506, %r558; shf.l.wrap.b32 %r561, %r558, %r558, 8; xor.b32 %r562, %r561, %r560; add.s32 %r563, %r558, %r559; sub.s32 %r564, %r559, %r562; shf.l.wrap.b32 %r565, %r562, %r562, 16; xor.b32 %r566, %r565, %r564; add.s32 %r567, %r562, %r563; sub.s32 %r568, %r563, %r566; shf.l.wrap.b32 %r569, %r566, %r566, 19; xor.b32 %r570, %r569, %r568; add.s32 %r571, %r566, %r567; sub.s32 %r572, %r567, %r570; shf.l.wrap.b32 %r573, %r570, %r570, 4; add.s32 %r574, %r570, %r571; add.s32 %r575, %r571, %r500; xor.b32 %r576, %r574, %r572; xor.b32 %r577, %r576, %r573; shf.l.wrap.b32 %r578, %r574, %r574, 14; sub.s32 %r579, %r577, %r578; xor.b32 %r580, %r579, %r575; shf.l.wrap.b32 %r581, %r579, %r579, 11; sub.s32 %r582, %r580, %r581; xor.b32 %r583, %r582, %r574; shf.l.wrap.b32 %r584, %r582, %r582, 25; sub.s32 %r585, %r583, %r584; xor.b32 %r586, %r585, %r579; shf.l.wrap.b32 %r587, %r585, %r585, 16; sub.s32 %r588, %r586, %r587; xor.b32 %r589, %r588, %r582; shf.l.wrap.b32 %r590, %r588, %r588, 4; sub.s32 %r591, %r589, %r590; xor.b32 %r592, %r591, %r585; shf.l.wrap.b32 %r593, %r591, %r591, 14; sub.s32 %r594, %r592, %r593; xor.b32 %r595, %r594, %r588; shf.l.wrap.b32 %r596, %r594, %r594, 24; sub.s32 %r597, %r595, %r596; add.ftz.f32 %f863, %f825, 0fBF800000; mov.f32 %f864, 0f3F800000; and.b32 %r598, %r597, 31; setp.lt.u32 %p29, %r598, 24; selp.f32 %f865, %f863, %f827, %p29; setp.lt.u32 %p30, %r598, 16; selp.f32 %f866, %f827, %f829, %p30; setp.lt.u32 %p31, %r598, 8; selp.f32 %f867, %f829, %f831, %p31; and.b32 %r599, %r597, 1; setp.eq.b32 %p32, %r599, 1; neg.ftz.f32 %f868, %f865; selp.f32 %f869, %f868, %f865, %p32; and.b32 %r600, %r597, 2; setp.eq.s32 %p33, %r600, 0; neg.ftz.f32 %f870, %f866; selp.f32 %f871, %f866, %f870, %p33; add.ftz.f32 %f872, %f869, %f871; and.b32 %r601, %r597, 4; setp.eq.s32 %p34, %r601, 0; neg.ftz.f32 %f873, %f867; selp.f32 %f874, %f867, %f873, %p34; add.ftz.f32 %f875, %f874, %f872; add.s32 %r602, %r498, -559038707; add.s32 %r603, %r502, %r602; sub.s32 %r604, %r602, %r505; xor.b32 %r605, %r508, %r604; add.s32 %r606, %r505, %r603; sub.s32 %r607, %r603, %r605; shf.l.wrap.b32 %r608, %r605, %r605, 8; xor.b32 %r609, %r608, %r607; add.s32 %r610, %r605, %r606; sub.s32 %r611, %r606, %r609; shf.l.wrap.b32 %r612, %r609, %r609, 16; xor.b32 %r613, %r612, %r611; add.s32 %r614, %r609, %r610; sub.s32 %r615, %r610, %r613; shf.l.wrap.b32 %r616, %r613, %r613, 19; xor.b32 %r617, %r616, %r615; add.s32 %r618, %r613, %r614; sub.s32 %r619, %r614, %r617; shf.l.wrap.b32 %r620, %r617, %r617, 4; add.s32 %r621, %r617, %r618; add.s32 %r622, %r618, %r500; xor.b32 %r623, %r621, %r619; xor.b32 %r624, %r623, %r620; shf.l.wrap.b32 %r625, %r621, %r621, 14; sub.s32 %r626, %r624, %r625; xor.b32 %r627, %r626, %r622; shf.l.wrap.b32 %r628, %r626, %r626, 11; sub.s32 %r629, %r627, %r628; xor.b32 %r630, %r629, %r621; shf.l.wrap.b32 %r631, %r629, %r629, 25; sub.s32 %r632, %r630, %r631; xor.b32 %r633, %r632, %r626; shf.l.wrap.b32 %r634, %r632, %r632, 16; sub.s32 %r635, %r633, %r634; xor.b32 %r636, %r635, %r629; shf.l.wrap.b32 %r637, %r635, %r635, 4; sub.s32 %r638, %r636, %r637; xor.b32 %r639, %r638, %r632; shf.l.wrap.b32 %r640, %r638, %r638, 14; sub.s32 %r641, %r639, %r640; xor.b32 %r642, %r641, %r635; shf.l.wrap.b32 %r643, %r641, %r641, 24; sub.s32 %r644, %r642, %r643; add.ftz.f32 %f876, %f827, 0fBF800000; and.b32 %r645, %r644, 31; setp.lt.u32 %p35, %r645, 24; selp.f32 %f877, %f825, %f876, %p35; setp.lt.u32 %p36, %r645, 16; selp.f32 %f878, %f876, %f829, %p36; setp.lt.u32 %p37, %r645, 8; selp.f32 %f879, %f829, %f831, %p37; and.b32 %r646, %r644, 1; setp.eq.b32 %p38, %r646, 1; neg.ftz.f32 %f880, %f877; selp.f32 %f881, %f880, %f877, %p38; and.b32 %r647, %r644, 2; setp.eq.s32 %p39, %r647, 0; neg.ftz.f32 %f882, %f878; selp.f32 %f883, %f878, %f882, %p39; add.ftz.f32 %f884, %f881, %f883; and.b32 %r648, %r644, 4; setp.eq.s32 %p40, %r648, 0; neg.ftz.f32 %f885, %f879; selp.f32 %f886, %f879, %f885, %p40; add.ftz.f32 %f887, %f886, %f884; sub.s32 %r649, %r602, %r555; xor.b32 %r650, %r557, %r649; add.s32 %r651, %r555, %r603; sub.s32 %r652, %r603, %r650; shf.l.wrap.b32 %r653, %r650, %r650, 8; xor.b32 %r654, %r653, %r652; add.s32 %r655, %r650, %r651; sub.s32 %r656, %r651, %r654; shf.l.wrap.b32 %r657, %r654, %r654, 16; xor.b32 %r658, %r657, %r656; add.s32 %r659, %r654, %r655; sub.s32 %r660, %r655, %r658; shf.l.wrap.b32 %r661, %r658, %r658, 19; xor.b32 %r662, %r661, %r660; add.s32 %r663, %r658, %r659; sub.s32 %r664, %r659, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 4; add.s32 %r666, %r662, %r663; add.s32 %r667, %r663, %r500; xor.b32 %r668, %r666, %r664; xor.b32 %r669, %r668, %r665; shf.l.wrap.b32 %r670, %r666, %r666, 14; sub.s32 %r671, %r669, %r670; xor.b32 %r672, %r671, %r667; shf.l.wrap.b32 %r673, %r671, %r671, 11; sub.s32 %r674, %r672, %r673; xor.b32 %r675, %r674, %r666; shf.l.wrap.b32 %r676, %r674, %r674, 25; sub.s32 %r677, %r675, %r676; xor.b32 %r678, %r677, %r671; shf.l.wrap.b32 %r679, %r677, %r677, 16; sub.s32 %r680, %r678, %r679; xor.b32 %r681, %r680, %r674; shf.l.wrap.b32 %r682, %r680, %r680, 4; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r677; shf.l.wrap.b32 %r685, %r683, %r683, 14; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r680; shf.l.wrap.b32 %r688, %r686, %r686, 24; sub.s32 %r689, %r687, %r688; and.b32 %r690, %r689, 31; setp.lt.u32 %p41, %r690, 24; selp.f32 %f888, %f863, %f876, %p41; setp.lt.u32 %p42, %r690, 16; selp.f32 %f889, %f876, %f829, %p42; setp.lt.u32 %p43, %r690, 8; selp.f32 %f890, %f829, %f831, %p43; and.b32 %r691, %r689, 1; setp.eq.b32 %p44, %r691, 1; neg.ftz.f32 %f891, %f888; selp.f32 %f892, %f891, %f888, %p44; and.b32 %r692, %r689, 2; setp.eq.s32 %p45, %r692, 0; neg.ftz.f32 %f893, %f889; selp.f32 %f894, %f889, %f893, %p45; add.ftz.f32 %f895, %f892, %f894; and.b32 %r693, %r689, 4; setp.eq.s32 %p46, %r693, 0; neg.ftz.f32 %f896, %f890; selp.f32 %f897, %f890, %f896, %p46; add.ftz.f32 %f898, %f897, %f895; not.b32 %r694, %r499; add.s32 %r695, %r499, -559038707; add.s32 %r696, %r497, %r694; shf.l.wrap.b32 %r697, %r695, %r695, 4; xor.b32 %r698, %r697, %r696; add.s32 %r699, %r695, %r501; sub.s32 %r700, %r501, %r698; shf.l.wrap.b32 %r701, %r698, %r698, 6; xor.b32 %r702, %r701, %r700; add.s32 %r703, %r698, %r699; sub.s32 %r704, %r699, %r702; shf.l.wrap.b32 %r705, %r702, %r702, 8; xor.b32 %r706, %r705, %r704; add.s32 %r707, %r702, %r703; sub.s32 %r708, %r703, %r706; shf.l.wrap.b32 %r709, %r706, %r706, 16; xor.b32 %r710, %r709, %r708; add.s32 %r711, %r706, %r707; sub.s32 %r712, %r707, %r710; shf.l.wrap.b32 %r713, %r710, %r710, 19; xor.b32 %r714, %r713, %r712; add.s32 %r715, %r710, %r711; sub.s32 %r716, %r711, %r714; shf.l.wrap.b32 %r717, %r714, %r714, 4; add.s32 %r718, %r714, %r715; add.s32 %r719, %r715, %r500; xor.b32 %r720, %r718, %r716; xor.b32 %r721, %r720, %r717; shf.l.wrap.b32 %r722, %r718, %r718, 14; sub.s32 %r723, %r721, %r722; xor.b32 %r724, %r723, %r719; shf.l.wrap.b32 %r725, %r723, %r723, 11; sub.s32 %r726, %r724, %r725; xor.b32 %r727, %r726, %r718; shf.l.wrap.b32 %r728, %r726, %r726, 25; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r723; shf.l.wrap.b32 %r731, %r729, %r729, 16; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r726; shf.l.wrap.b32 %r734, %r732, %r732, 4; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 14; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 24; sub.s32 %r741, %r739, %r740; add.ftz.f32 %f899, %f829, 0fBF800000; and.b32 %r742, %r741, 31; setp.lt.u32 %p47, %r742, 24; selp.f32 %f900, %f825, %f827, %p47; setp.lt.u32 %p48, %r742, 16; selp.f32 %f901, %f827, %f899, %p48; setp.lt.u32 %p49, %r742, 8; selp.f32 %f902, %f899, %f831, %p49; and.b32 %r743, %r741, 1; setp.eq.b32 %p50, %r743, 1; neg.ftz.f32 %f903, %f900; selp.f32 %f904, %f903, %f900, %p50; and.b32 %r744, %r741, 2; setp.eq.s32 %p51, %r744, 0; neg.ftz.f32 %f905, %f901; selp.f32 %f906, %f901, %f905, %p51; add.ftz.f32 %f907, %f904, %f906; and.b32 %r745, %r741, 4; setp.eq.s32 %p52, %r745, 0; neg.ftz.f32 %f908, %f902; selp.f32 %f909, %f902, %f908, %p52; add.ftz.f32 %f910, %f909, %f907; xor.b32 %r746, %r697, %r503; sub.s32 %r747, %r501, %r746; shf.l.wrap.b32 %r748, %r746, %r746, 6; xor.b32 %r749, %r748, %r747; add.s32 %r750, %r746, %r699; sub.s32 %r751, %r699, %r749; shf.l.wrap.b32 %r752, %r749, %r749, 8; xor.b32 %r753, %r752, %r751; add.s32 %r754, %r749, %r750; sub.s32 %r755, %r750, %r753; shf.l.wrap.b32 %r756, %r753, %r753, 16; xor.b32 %r757, %r756, %r755; add.s32 %r758, %r753, %r754; sub.s32 %r759, %r754, %r757; shf.l.wrap.b32 %r760, %r757, %r757, 19; xor.b32 %r761, %r760, %r759; add.s32 %r762, %r757, %r758; sub.s32 %r763, %r758, %r761; shf.l.wrap.b32 %r764, %r761, %r761, 4; add.s32 %r765, %r761, %r762; add.s32 %r766, %r762, %r500; xor.b32 %r767, %r765, %r763; xor.b32 %r768, %r767, %r764; shf.l.wrap.b32 %r769, %r765, %r765, 14; sub.s32 %r770, %r768, %r769; xor.b32 %r771, %r770, %r766; shf.l.wrap.b32 %r772, %r770, %r770, 11; sub.s32 %r773, %r771, %r772; xor.b32 %r774, %r773, %r765; shf.l.wrap.b32 %r775, %r773, %r773, 25; sub.s32 %r776, %r774, %r775; xor.b32 %r777, %r776, %r770; shf.l.wrap.b32 %r778, %r776, %r776, 16; sub.s32 %r779, %r777, %r778; xor.b32 %r780, %r779, %r773; shf.l.wrap.b32 %r781, %r779, %r779, 4; sub.s32 %r782, %r780, %r781; xor.b32 %r783, %r782, %r776; shf.l.wrap.b32 %r784, %r782, %r782, 14; sub.s32 %r785, %r783, %r784; xor.b32 %r786, %r785, %r779; shf.l.wrap.b32 %r787, %r785, %r785, 24; sub.s32 %r788, %r786, %r787; and.b32 %r789, %r788, 31; setp.lt.u32 %p53, %r789, 24; selp.f32 %f911, %f863, %f827, %p53; setp.lt.u32 %p54, %r789, 16; selp.f32 %f912, %f827, %f899, %p54; setp.lt.u32 %p55, %r789, 8; selp.f32 %f913, %f899, %f831, %p55; and.b32 %r790, %r788, 1; setp.eq.b32 %p56, %r790, 1; neg.ftz.f32 %f914, %f911; selp.f32 %f915, %f914, %f911, %p56; and.b32 %r791, %r788, 2; setp.eq.s32 %p57, %r791, 0; neg.ftz.f32 %f916, %f912; selp.f32 %f917, %f912, %f916, %p57; add.ftz.f32 %f918, %f915, %f917; and.b32 %r792, %r788, 4; setp.eq.s32 %p58, %r792, 0; neg.ftz.f32 %f919, %f913; selp.f32 %f920, %f913, %f919, %p58; add.ftz.f32 %f921, %f920, %f918; add.s32 %r793, %r695, %r602; sub.s32 %r794, %r602, %r698; xor.b32 %r795, %r701, %r794; add.s32 %r796, %r698, %r793; sub.s32 %r797, %r793, %r795; shf.l.wrap.b32 %r798, %r795, %r795, 8; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r795, %r796; sub.s32 %r801, %r796, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 16; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 19; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 4; add.s32 %r811, %r807, %r808; add.s32 %r812, %r808, %r500; xor.b32 %r813, %r811, %r809; xor.b32 %r814, %r813, %r810; shf.l.wrap.b32 %r815, %r811, %r811, 14; sub.s32 %r816, %r814, %r815; xor.b32 %r817, %r816, %r812; shf.l.wrap.b32 %r818, %r816, %r816, 11; sub.s32 %r819, %r817, %r818; xor.b32 %r820, %r819, %r811; shf.l.wrap.b32 %r821, %r819, %r819, 25; sub.s32 %r822, %r820, %r821; xor.b32 %r823, %r822, %r816; shf.l.wrap.b32 %r824, %r822, %r822, 16; sub.s32 %r825, %r823, %r824; xor.b32 %r826, %r825, %r819; shf.l.wrap.b32 %r827, %r825, %r825, 4; sub.s32 %r828, %r826, %r827; xor.b32 %r829, %r828, %r822; shf.l.wrap.b32 %r830, %r828, %r828, 14; sub.s32 %r831, %r829, %r830; xor.b32 %r832, %r831, %r825; shf.l.wrap.b32 %r833, %r831, %r831, 24; sub.s32 %r834, %r832, %r833; and.b32 %r835, %r834, 31; setp.lt.u32 %p59, %r835, 24; selp.f32 %f922, %f825, %f876, %p59; setp.lt.u32 %p60, %r835, 16; selp.f32 %f923, %f876, %f899, %p60; setp.lt.u32 %p61, %r835, 8; selp.f32 %f924, %f899, %f831, %p61; and.b32 %r836, %r834, 1; setp.eq.b32 %p62, %r836, 1; neg.ftz.f32 %f925, %f922; selp.f32 %f926, %f925, %f922, %p62; and.b32 %r837, %r834, 2; setp.eq.s32 %p63, %r837, 0; neg.ftz.f32 %f927, %f923; selp.f32 %f928, %f923, %f927, %p63; add.ftz.f32 %f929, %f926, %f928; and.b32 %r838, %r834, 4; setp.eq.s32 %p64, %r838, 0; neg.ftz.f32 %f930, %f924; selp.f32 %f931, %f924, %f930, %p64; add.ftz.f32 %f932, %f931, %f929; sub.s32 %r839, %r602, %r746; xor.b32 %r840, %r748, %r839; add.s32 %r841, %r746, %r793; sub.s32 %r842, %r793, %r840; shf.l.wrap.b32 %r843, %r840, %r840, 8; xor.b32 %r844, %r843, %r842; add.s32 %r845, %r840, %r841; sub.s32 %r846, %r841, %r844; shf.l.wrap.b32 %r847, %r844, %r844, 16; xor.b32 %r848, %r847, %r846; add.s32 %r849, %r844, %r845; sub.s32 %r850, %r845, %r848; shf.l.wrap.b32 %r851, %r848, %r848, 19; xor.b32 %r852, %r851, %r850; add.s32 %r853, %r848, %r849; sub.s32 %r854, %r849, %r852; shf.l.wrap.b32 %r855, %r852, %r852, 4; add.s32 %r856, %r852, %r853; add.s32 %r857, %r853, %r500; xor.b32 %r858, %r856, %r854; xor.b32 %r859, %r858, %r855; shf.l.wrap.b32 %r860, %r856, %r856, 14; sub.s32 %r861, %r859, %r860; xor.b32 %r862, %r861, %r857; shf.l.wrap.b32 %r863, %r861, %r861, 11; sub.s32 %r864, %r862, %r863; xor.b32 %r865, %r864, %r856; shf.l.wrap.b32 %r866, %r864, %r864, 25; sub.s32 %r867, %r865, %r866; xor.b32 %r868, %r867, %r861; shf.l.wrap.b32 %r869, %r867, %r867, 16; sub.s32 %r870, %r868, %r869; xor.b32 %r871, %r870, %r864; shf.l.wrap.b32 %r872, %r870, %r870, 4; sub.s32 %r873, %r871, %r872; xor.b32 %r874, %r873, %r867; shf.l.wrap.b32 %r875, %r873, %r873, 14; sub.s32 %r876, %r874, %r875; xor.b32 %r877, %r876, %r870; shf.l.wrap.b32 %r878, %r876, %r876, 24; sub.s32 %r879, %r877, %r878; and.b32 %r880, %r879, 31; setp.lt.u32 %p65, %r880, 24; selp.f32 %f933, %f863, %f876, %p65; setp.lt.u32 %p66, %r880, 16; selp.f32 %f934, %f876, %f899, %p66; setp.lt.u32 %p67, %r880, 8; selp.f32 %f935, %f899, %f831, %p67; and.b32 %r881, %r879, 1; setp.eq.b32 %p68, %r881, 1; neg.ftz.f32 %f936, %f933; selp.f32 %f937, %f936, %f933, %p68; and.b32 %r882, %r879, 2; setp.eq.s32 %p69, %r882, 0; neg.ftz.f32 %f938, %f934; selp.f32 %f939, %f934, %f938, %p69; add.ftz.f32 %f940, %f937, %f939; and.b32 %r883, %r879, 4; setp.eq.s32 %p70, %r883, 0; neg.ftz.f32 %f941, %f935; selp.f32 %f942, %f935, %f941, %p70; add.ftz.f32 %f943, %f942, %f940; add.s32 %r884, %r500, 1; add.s32 %r885, %r522, %r884; xor.b32 %r886, %r530, %r885; sub.s32 %r887, %r886, %r532; xor.b32 %r888, %r887, %r525; shf.l.wrap.b32 %r889, %r887, %r887, 25; sub.s32 %r890, %r888, %r889; xor.b32 %r891, %r890, %r530; shf.l.wrap.b32 %r892, %r890, %r890, 16; sub.s32 %r893, %r891, %r892; xor.b32 %r894, %r893, %r887; shf.l.wrap.b32 %r895, %r893, %r893, 4; sub.s32 %r896, %r894, %r895; xor.b32 %r897, %r896, %r890; shf.l.wrap.b32 %r898, %r896, %r896, 14; sub.s32 %r899, %r897, %r898; xor.b32 %r900, %r899, %r893; shf.l.wrap.b32 %r901, %r899, %r899, 24; sub.s32 %r902, %r900, %r901; add.ftz.f32 %f944, %f831, 0fBF800000; and.b32 %r903, %r902, 31; setp.lt.u32 %p71, %r903, 24; selp.f32 %f945, %f825, %f827, %p71; setp.lt.u32 %p72, %r903, 16; selp.f32 %f946, %f827, %f829, %p72; setp.lt.u32 %p73, %r903, 8; selp.f32 %f947, %f829, %f944, %p73; and.b32 %r904, %r902, 1; setp.eq.b32 %p74, %r904, 1; neg.ftz.f32 %f948, %f945; selp.f32 %f949, %f948, %f945, %p74; and.b32 %r905, %r902, 2; setp.eq.s32 %p75, %r905, 0; neg.ftz.f32 %f950, %f946; selp.f32 %f951, %f946, %f950, %p75; add.ftz.f32 %f952, %f949, %f951; and.b32 %r906, %r902, 4; setp.eq.s32 %p76, %r906, 0; neg.ftz.f32 %f953, %f947; selp.f32 %f954, %f947, %f953, %p76; add.ftz.f32 %f955, %f954, %f952; add.s32 %r907, %r571, %r884; xor.b32 %r908, %r579, %r907; sub.s32 %r909, %r908, %r581; xor.b32 %r910, %r909, %r574; shf.l.wrap.b32 %r911, %r909, %r909, 25; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r579; shf.l.wrap.b32 %r914, %r912, %r912, 16; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 4; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 14; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 24; sub.s32 %r924, %r922, %r923; and.b32 %r925, %r924, 31; setp.lt.u32 %p77, %r925, 24; selp.f32 %f956, %f863, %f827, %p77; setp.lt.u32 %p78, %r925, 16; selp.f32 %f957, %f827, %f829, %p78; setp.lt.u32 %p79, %r925, 8; selp.f32 %f958, %f829, %f944, %p79; and.b32 %r926, %r924, 1; setp.eq.b32 %p80, %r926, 1; neg.ftz.f32 %f959, %f956; selp.f32 %f960, %f959, %f956, %p80; and.b32 %r927, %r924, 2; setp.eq.s32 %p81, %r927, 0; neg.ftz.f32 %f961, %f957; selp.f32 %f962, %f957, %f961, %p81; add.ftz.f32 %f963, %f960, %f962; and.b32 %r928, %r924, 4; setp.eq.s32 %p82, %r928, 0; neg.ftz.f32 %f964, %f958; selp.f32 %f965, %f958, %f964, %p82; add.ftz.f32 %f966, %f965, %f963; add.s32 %r929, %r618, %r884; xor.b32 %r930, %r626, %r929; sub.s32 %r931, %r930, %r628; xor.b32 %r932, %r931, %r621; shf.l.wrap.b32 %r933, %r931, %r931, 25; sub.s32 %r934, %r932, %r933; xor.b32 %r935, %r934, %r626; shf.l.wrap.b32 %r936, %r934, %r934, 16; sub.s32 %r937, %r935, %r936; xor.b32 %r938, %r937, %r931; shf.l.wrap.b32 %r939, %r937, %r937, 4; sub.s32 %r940, %r938, %r939; xor.b32 %r941, %r940, %r934; shf.l.wrap.b32 %r942, %r940, %r940, 14; sub.s32 %r943, %r941, %r942; xor.b32 %r944, %r943, %r937; shf.l.wrap.b32 %r945, %r943, %r943, 24; sub.s32 %r946, %r944, %r945; and.b32 %r947, %r946, 31; setp.lt.u32 %p83, %r947, 24; selp.f32 %f967, %f825, %f876, %p83; setp.lt.u32 %p84, %r947, 16; selp.f32 %f968, %f876, %f829, %p84; setp.lt.u32 %p85, %r947, 8; selp.f32 %f969, %f829, %f944, %p85; and.b32 %r948, %r946, 1; setp.eq.b32 %p86, %r948, 1; neg.ftz.f32 %f970, %f967; selp.f32 %f971, %f970, %f967, %p86; and.b32 %r949, %r946, 2; setp.eq.s32 %p87, %r949, 0; neg.ftz.f32 %f972, %f968; selp.f32 %f973, %f968, %f972, %p87; add.ftz.f32 %f974, %f971, %f973; and.b32 %r950, %r946, 4; setp.eq.s32 %p88, %r950, 0; neg.ftz.f32 %f975, %f969; selp.f32 %f976, %f969, %f975, %p88; add.ftz.f32 %f977, %f976, %f974; add.s32 %r951, %r663, %r884; xor.b32 %r952, %r671, %r951; sub.s32 %r953, %r952, %r673; xor.b32 %r954, %r953, %r666; shf.l.wrap.b32 %r955, %r953, %r953, 25; sub.s32 %r956, %r954, %r955; xor.b32 %r957, %r956, %r671; shf.l.wrap.b32 %r958, %r956, %r956, 16; sub.s32 %r959, %r957, %r958; xor.b32 %r960, %r959, %r953; shf.l.wrap.b32 %r961, %r959, %r959, 4; sub.s32 %r962, %r960, %r961; xor.b32 %r963, %r962, %r956; shf.l.wrap.b32 %r964, %r962, %r962, 14; sub.s32 %r965, %r963, %r964; xor.b32 %r966, %r965, %r959; shf.l.wrap.b32 %r967, %r965, %r965, 24; sub.s32 %r968, %r966, %r967; and.b32 %r969, %r968, 31; setp.lt.u32 %p89, %r969, 24; selp.f32 %f978, %f863, %f876, %p89; setp.lt.u32 %p90, %r969, 16; selp.f32 %f979, %f876, %f829, %p90; setp.lt.u32 %p91, %r969, 8; selp.f32 %f980, %f829, %f944, %p91; and.b32 %r970, %r968, 1; setp.eq.b32 %p92, %r970, 1; neg.ftz.f32 %f981, %f978; selp.f32 %f982, %f981, %f978, %p92; and.b32 %r971, %r968, 2; setp.eq.s32 %p93, %r971, 0; neg.ftz.f32 %f983, %f979; selp.f32 %f984, %f979, %f983, %p93; add.ftz.f32 %f985, %f982, %f984; and.b32 %r972, %r968, 4; setp.eq.s32 %p94, %r972, 0; neg.ftz.f32 %f986, %f980; selp.f32 %f987, %f980, %f986, %p94; add.ftz.f32 %f988, %f987, %f985; add.s32 %r973, %r715, %r884; xor.b32 %r974, %r723, %r973; sub.s32 %r975, %r974, %r725; xor.b32 %r976, %r975, %r718; shf.l.wrap.b32 %r977, %r975, %r975, 25; sub.s32 %r978, %r976, %r977; xor.b32 %r979, %r978, %r723; shf.l.wrap.b32 %r980, %r978, %r978, 16; sub.s32 %r981, %r979, %r980; xor.b32 %r982, %r981, %r975; shf.l.wrap.b32 %r983, %r981, %r981, 4; sub.s32 %r984, %r982, %r983; xor.b32 %r985, %r984, %r978; shf.l.wrap.b32 %r986, %r984, %r984, 14; sub.s32 %r987, %r985, %r986; xor.b32 %r988, %r987, %r981; shf.l.wrap.b32 %r989, %r987, %r987, 24; sub.s32 %r990, %r988, %r989; and.b32 %r991, %r990, 31; setp.lt.u32 %p95, %r991, 24; selp.f32 %f989, %f825, %f827, %p95; setp.lt.u32 %p96, %r991, 16; selp.f32 %f990, %f827, %f899, %p96; setp.lt.u32 %p97, %r991, 8; selp.f32 %f991, %f899, %f944, %p97; and.b32 %r992, %r990, 1; setp.eq.b32 %p98, %r992, 1; neg.ftz.f32 %f992, %f989; selp.f32 %f993, %f992, %f989, %p98; and.b32 %r993, %r990, 2; setp.eq.s32 %p99, %r993, 0; neg.ftz.f32 %f994, %f990; selp.f32 %f995, %f990, %f994, %p99; add.ftz.f32 %f996, %f993, %f995; and.b32 %r994, %r990, 4; setp.eq.s32 %p100, %r994, 0; neg.ftz.f32 %f997, %f991; selp.f32 %f998, %f991, %f997, %p100; add.ftz.f32 %f999, %f998, %f996; add.s32 %r995, %r762, %r884; xor.b32 %r996, %r770, %r995; sub.s32 %r997, %r996, %r772; xor.b32 %r998, %r997, %r765; shf.l.wrap.b32 %r999, %r997, %r997, 25; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r770; shf.l.wrap.b32 %r1002, %r1000, %r1000, 16; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r997; shf.l.wrap.b32 %r1005, %r1003, %r1003, 4; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 14; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 24; sub.s32 %r1012, %r1010, %r1011; and.b32 %r1013, %r1012, 31; setp.lt.u32 %p101, %r1013, 24; selp.f32 %f1000, %f863, %f827, %p101; setp.lt.u32 %p102, %r1013, 16; selp.f32 %f1001, %f827, %f899, %p102; setp.lt.u32 %p103, %r1013, 8; selp.f32 %f1002, %f899, %f944, %p103; and.b32 %r1014, %r1012, 1; setp.eq.b32 %p104, %r1014, 1; neg.ftz.f32 %f1003, %f1000; selp.f32 %f1004, %f1003, %f1000, %p104; and.b32 %r1015, %r1012, 2; setp.eq.s32 %p105, %r1015, 0; neg.ftz.f32 %f1005, %f1001; selp.f32 %f1006, %f1001, %f1005, %p105; add.ftz.f32 %f1007, %f1004, %f1006; and.b32 %r1016, %r1012, 4; setp.eq.s32 %p106, %r1016, 0; neg.ftz.f32 %f1008, %f1002; selp.f32 %f1009, %f1002, %f1008, %p106; add.ftz.f32 %f1010, %f1009, %f1007; add.s32 %r1017, %r808, %r884; xor.b32 %r1018, %r816, %r1017; sub.s32 %r1019, %r1018, %r818; xor.b32 %r1020, %r1019, %r811; shf.l.wrap.b32 %r1021, %r1019, %r1019, 25; sub.s32 %r1022, %r1020, %r1021; xor.b32 %r1023, %r1022, %r816; shf.l.wrap.b32 %r1024, %r1022, %r1022, 16; sub.s32 %r1025, %r1023, %r1024; xor.b32 %r1026, %r1025, %r1019; shf.l.wrap.b32 %r1027, %r1025, %r1025, 4; sub.s32 %r1028, %r1026, %r1027; xor.b32 %r1029, %r1028, %r1022; shf.l.wrap.b32 %r1030, %r1028, %r1028, 14; sub.s32 %r1031, %r1029, %r1030; xor.b32 %r1032, %r1031, %r1025; shf.l.wrap.b32 %r1033, %r1031, %r1031, 24; sub.s32 %r1034, %r1032, %r1033; and.b32 %r1035, %r1034, 31; setp.lt.u32 %p107, %r1035, 24; selp.f32 %f1011, %f825, %f876, %p107; setp.lt.u32 %p108, %r1035, 16; selp.f32 %f1012, %f876, %f899, %p108; setp.lt.u32 %p109, %r1035, 8; selp.f32 %f1013, %f899, %f944, %p109; and.b32 %r1036, %r1034, 1; setp.eq.b32 %p110, %r1036, 1; neg.ftz.f32 %f1014, %f1011; selp.f32 %f1015, %f1014, %f1011, %p110; and.b32 %r1037, %r1034, 2; setp.eq.s32 %p111, %r1037, 0; neg.ftz.f32 %f1016, %f1012; selp.f32 %f1017, %f1012, %f1016, %p111; add.ftz.f32 %f1018, %f1015, %f1017; and.b32 %r1038, %r1034, 4; setp.eq.s32 %p112, %r1038, 0; neg.ftz.f32 %f1019, %f1013; selp.f32 %f1020, %f1013, %f1019, %p112; add.ftz.f32 %f1021, %f1020, %f1018; add.s32 %r1039, %r853, %r884; xor.b32 %r1040, %r861, %r1039; sub.s32 %r1041, %r1040, %r863; xor.b32 %r1042, %r1041, %r856; shf.l.wrap.b32 %r1043, %r1041, %r1041, 25; sub.s32 %r1044, %r1042, %r1043; xor.b32 %r1045, %r1044, %r861; shf.l.wrap.b32 %r1046, %r1044, %r1044, 16; sub.s32 %r1047, %r1045, %r1046; xor.b32 %r1048, %r1047, %r1041; shf.l.wrap.b32 %r1049, %r1047, %r1047, 4; sub.s32 %r1050, %r1048, %r1049; xor.b32 %r1051, %r1050, %r1044; shf.l.wrap.b32 %r1052, %r1050, %r1050, 14; sub.s32 %r1053, %r1051, %r1052; xor.b32 %r1054, %r1053, %r1047; shf.l.wrap.b32 %r1055, %r1053, %r1053, 24; sub.s32 %r1056, %r1054, %r1055; and.b32 %r1057, %r1056, 31; setp.lt.u32 %p113, %r1057, 24; selp.f32 %f1022, %f863, %f876, %p113; setp.lt.u32 %p114, %r1057, 16; selp.f32 %f1023, %f876, %f899, %p114; setp.lt.u32 %p115, %r1057, 8; selp.f32 %f1024, %f899, %f944, %p115; and.b32 %r1058, %r1056, 1; setp.eq.b32 %p116, %r1058, 1; neg.ftz.f32 %f1025, %f1022; selp.f32 %f1026, %f1025, %f1022, %p116; and.b32 %r1059, %r1056, 2; setp.eq.s32 %p117, %r1059, 0; neg.ftz.f32 %f1027, %f1023; selp.f32 %f1028, %f1023, %f1027, %p117; add.ftz.f32 %f1029, %f1026, %f1028; and.b32 %r1060, %r1056, 4; setp.eq.s32 %p118, %r1060, 0; neg.ftz.f32 %f1030, %f1024; selp.f32 %f1031, %f1024, %f1030, %p118; add.ftz.f32 %f1032, %f1031, %f1029; sub.ftz.f32 %f1033, %f864, %f836; sub.ftz.f32 %f1034, %f864, %f841; sub.ftz.f32 %f1035, %f864, %f846; mul.ftz.f32 %f1036, %f1033, %f862; fma.rn.ftz.f32 %f1037, %f836, %f875, %f1036; mul.ftz.f32 %f1038, %f1033, %f887; fma.rn.ftz.f32 %f1039, %f836, %f898, %f1038; mul.ftz.f32 %f1040, %f841, %f1039; fma.rn.ftz.f32 %f1041, %f1034, %f1037, %f1040; mul.ftz.f32 %f1042, %f1033, %f910; fma.rn.ftz.f32 %f1043, %f836, %f921, %f1042; mul.ftz.f32 %f1044, %f1033, %f932; fma.rn.ftz.f32 %f1045, %f836, %f943, %f1044; mul.ftz.f32 %f1046, %f841, %f1045; fma.rn.ftz.f32 %f1047, %f1034, %f1043, %f1046; mul.ftz.f32 %f1048, %f846, %f1047; fma.rn.ftz.f32 %f1049, %f1035, %f1041, %f1048; mul.ftz.f32 %f1050, %f1033, %f955; fma.rn.ftz.f32 %f1051, %f836, %f966, %f1050; mul.ftz.f32 %f1052, %f1033, %f977; fma.rn.ftz.f32 %f1053, %f836, %f988, %f1052; mul.ftz.f32 %f1054, %f841, %f1053; fma.rn.ftz.f32 %f1055, %f1034, %f1051, %f1054; mul.ftz.f32 %f1056, %f1033, %f999; fma.rn.ftz.f32 %f1057, %f836, %f1010, %f1056; mul.ftz.f32 %f1058, %f1033, %f1021; fma.rn.ftz.f32 %f1059, %f836, %f1032, %f1058; mul.ftz.f32 %f1060, %f841, %f1059; fma.rn.ftz.f32 %f1061, %f1034, %f1057, %f1060; mul.ftz.f32 %f1062, %f846, %f1061; fma.rn.ftz.f32 %f1063, %f1035, %f1055, %f1062; sub.ftz.f32 %f1064, %f1063, %f1049; fma.rn.ftz.f32 %f33, %f851, %f1064, %f1049; mov.b32 %r16, %f33; setp.nan.ftz.f32 %p119, %f33, %f33; mov.f32 %f5521, 0f00000000; @%p119 bra $L__BB74_60; or.b32 %r1061, %r16, -2147483648; setp.ne.s32 %p120, %r1061, -2147483648; add.ftz.f32 %f1066, %f33, %f33; setp.eq.ftz.f32 %p121, %f1066, %f33; and.pred %p122, %p121, %p120; @%p122 bra $L__BB74_60; shl.b32 %r1062, %r16, 1; setp.lt.u32 %p123, %r1062, -16777215; mul.ftz.f32 %f1067, %f33, 0f3F559B3D; selp.f32 %f5521, %f1067, 0f00000000, %p123; $L__BB74_60: add.ftz.f32 %f5827, %f21, %f5521; mul.ftz.f32 %f5532, %f5568, %f27; mul.ftz.f32 %f5531, %f5569, %f27; mul.ftz.f32 %f5530, %f5570, %f27; mul.ftz.f32 %f5529, %f5571, %f27; cvt.rzi.ftz.s32.f32 %r17, %f26; setp.lt.s32 %p124, %r17, 2; mov.f32 %f5533, %f32; @%p124 bra $L__BB74_66; mov.u32 %r10018, 1; mov.f32 %f5533, %f32; $L__BB74_62: cvt.rmi.ftz.f32.f32 %f1069, %f5532; cvt.rzi.ftz.s32.f32 %r1064, %f1069; sub.ftz.f32 %f1070, %f5532, %f1069; cvt.rmi.ftz.f32.f32 %f1071, %f5531; cvt.rzi.ftz.s32.f32 %r1065, %f1071; sub.ftz.f32 %f1072, %f5531, %f1071; cvt.rmi.ftz.f32.f32 %f1073, %f5530; cvt.rzi.ftz.s32.f32 %r1066, %f1073; sub.ftz.f32 %f1074, %f5530, %f1073; cvt.rmi.ftz.f32.f32 %f1075, %f5529; cvt.rzi.ftz.s32.f32 %r1067, %f1075; sub.ftz.f32 %f1076, %f5529, %f1075; mul.ftz.f32 %f1077, %f1070, %f1070; mul.ftz.f32 %f1078, %f1070, %f1077; fma.rn.ftz.f32 %f1079, %f1070, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1080, %f1070, %f1079, 0f41200000; mul.ftz.f32 %f1081, %f1078, %f1080; mul.ftz.f32 %f1082, %f1072, %f1072; mul.ftz.f32 %f1083, %f1072, %f1082; fma.rn.ftz.f32 %f1084, %f1072, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1085, %f1072, %f1084, 0f41200000; mul.ftz.f32 %f1086, %f1083, %f1085; mul.ftz.f32 %f1087, %f1074, %f1074; mul.ftz.f32 %f1088, %f1074, %f1087; fma.rn.ftz.f32 %f1089, %f1074, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1090, %f1074, %f1089, 0f41200000; mul.ftz.f32 %f1091, %f1088, %f1090; mul.ftz.f32 %f1092, %f1076, %f1076; mul.ftz.f32 %f1093, %f1076, %f1092; fma.rn.ftz.f32 %f1094, %f1076, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1095, %f1076, %f1094, 0f41200000; mul.ftz.f32 %f1096, %f1093, %f1095; add.s32 %r1068, %r1065, -559038708; add.s32 %r1069, %r1066, -559038708; sub.s32 %r1070, %r1064, %r1066; shf.l.wrap.b32 %r1071, %r1069, %r1069, 4; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1069, %r1068; sub.s32 %r1074, %r1068, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 6; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 8; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 16; xor.b32 %r1084, %r1083, %r1082; add.s32 %r1085, %r1080, %r1081; sub.s32 %r1086, %r1081, %r1084; shf.l.wrap.b32 %r1087, %r1084, %r1084, 19; xor.b32 %r1088, %r1087, %r1086; add.s32 %r1089, %r1084, %r1085; sub.s32 %r1090, %r1085, %r1088; shf.l.wrap.b32 %r1091, %r1088, %r1088, 4; add.s32 %r1092, %r1088, %r1089; add.s32 %r1093, %r1089, %r1067; xor.b32 %r1094, %r1092, %r1090; xor.b32 %r1095, %r1094, %r1091; shf.l.wrap.b32 %r1096, %r1092, %r1092, 14; sub.s32 %r1097, %r1095, %r1096; xor.b32 %r1098, %r1097, %r1093; shf.l.wrap.b32 %r1099, %r1097, %r1097, 11; sub.s32 %r1100, %r1098, %r1099; xor.b32 %r1101, %r1100, %r1092; shf.l.wrap.b32 %r1102, %r1100, %r1100, 25; sub.s32 %r1103, %r1101, %r1102; xor.b32 %r1104, %r1103, %r1097; shf.l.wrap.b32 %r1105, %r1103, %r1103, 16; sub.s32 %r1106, %r1104, %r1105; xor.b32 %r1107, %r1106, %r1100; shf.l.wrap.b32 %r1108, %r1106, %r1106, 4; sub.s32 %r1109, %r1107, %r1108; xor.b32 %r1110, %r1109, %r1103; shf.l.wrap.b32 %r1111, %r1109, %r1109, 14; sub.s32 %r1112, %r1110, %r1111; xor.b32 %r1113, %r1112, %r1106; shf.l.wrap.b32 %r1114, %r1112, %r1112, 24; sub.s32 %r1115, %r1113, %r1114; and.b32 %r1116, %r1115, 31; setp.lt.u32 %p125, %r1116, 24; selp.f32 %f1097, %f1070, %f1072, %p125; setp.lt.u32 %p126, %r1116, 16; selp.f32 %f1098, %f1072, %f1074, %p126; setp.lt.u32 %p127, %r1116, 8; selp.f32 %f1099, %f1074, %f1076, %p127; and.b32 %r1117, %r1115, 1; setp.eq.b32 %p128, %r1117, 1; neg.ftz.f32 %f1100, %f1097; selp.f32 %f1101, %f1100, %f1097, %p128; and.b32 %r1118, %r1115, 2; setp.eq.s32 %p129, %r1118, 0; neg.ftz.f32 %f1102, %f1098; selp.f32 %f1103, %f1098, %f1102, %p129; add.ftz.f32 %f1104, %f1101, %f1103; and.b32 %r1119, %r1115, 4; setp.eq.s32 %p130, %r1119, 0; neg.ftz.f32 %f1105, %f1099; selp.f32 %f1106, %f1099, %f1105, %p130; add.ftz.f32 %f1107, %f1106, %f1104; add.s32 %r1120, %r1064, 1; sub.s32 %r1121, %r1120, %r1066; xor.b32 %r1122, %r1071, %r1121; sub.s32 %r1123, %r1068, %r1122; shf.l.wrap.b32 %r1124, %r1122, %r1122, 6; xor.b32 %r1125, %r1124, %r1123; add.s32 %r1126, %r1122, %r1073; sub.s32 %r1127, %r1073, %r1125; shf.l.wrap.b32 %r1128, %r1125, %r1125, 8; xor.b32 %r1129, %r1128, %r1127; add.s32 %r1130, %r1125, %r1126; sub.s32 %r1131, %r1126, %r1129; shf.l.wrap.b32 %r1132, %r1129, %r1129, 16; xor.b32 %r1133, %r1132, %r1131; add.s32 %r1134, %r1129, %r1130; sub.s32 %r1135, %r1130, %r1133; shf.l.wrap.b32 %r1136, %r1133, %r1133, 19; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1133, %r1134; sub.s32 %r1139, %r1134, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 4; add.s32 %r1141, %r1137, %r1138; add.s32 %r1142, %r1138, %r1067; xor.b32 %r1143, %r1141, %r1139; xor.b32 %r1144, %r1143, %r1140; shf.l.wrap.b32 %r1145, %r1141, %r1141, 14; sub.s32 %r1146, %r1144, %r1145; xor.b32 %r1147, %r1146, %r1142; shf.l.wrap.b32 %r1148, %r1146, %r1146, 11; sub.s32 %r1149, %r1147, %r1148; xor.b32 %r1150, %r1149, %r1141; shf.l.wrap.b32 %r1151, %r1149, %r1149, 25; sub.s32 %r1152, %r1150, %r1151; xor.b32 %r1153, %r1152, %r1146; shf.l.wrap.b32 %r1154, %r1152, %r1152, 16; sub.s32 %r1155, %r1153, %r1154; xor.b32 %r1156, %r1155, %r1149; shf.l.wrap.b32 %r1157, %r1155, %r1155, 4; sub.s32 %r1158, %r1156, %r1157; xor.b32 %r1159, %r1158, %r1152; shf.l.wrap.b32 %r1160, %r1158, %r1158, 14; sub.s32 %r1161, %r1159, %r1160; xor.b32 %r1162, %r1161, %r1155; shf.l.wrap.b32 %r1163, %r1161, %r1161, 24; sub.s32 %r1164, %r1162, %r1163; add.ftz.f32 %f1108, %f1070, 0fBF800000; and.b32 %r1165, %r1164, 31; setp.lt.u32 %p131, %r1165, 24; selp.f32 %f1110, %f1108, %f1072, %p131; setp.lt.u32 %p132, %r1165, 16; selp.f32 %f1111, %f1072, %f1074, %p132; setp.lt.u32 %p133, %r1165, 8; selp.f32 %f1112, %f1074, %f1076, %p133; and.b32 %r1166, %r1164, 1; setp.eq.b32 %p134, %r1166, 1; neg.ftz.f32 %f1113, %f1110; selp.f32 %f1114, %f1113, %f1110, %p134; and.b32 %r1167, %r1164, 2; setp.eq.s32 %p135, %r1167, 0; neg.ftz.f32 %f1115, %f1111; selp.f32 %f1116, %f1111, %f1115, %p135; add.ftz.f32 %f1117, %f1114, %f1116; and.b32 %r1168, %r1164, 4; setp.eq.s32 %p136, %r1168, 0; neg.ftz.f32 %f1118, %f1112; selp.f32 %f1119, %f1112, %f1118, %p136; add.ftz.f32 %f1120, %f1119, %f1117; add.s32 %r1169, %r1065, -559038707; add.s32 %r1170, %r1069, %r1169; sub.s32 %r1171, %r1169, %r1072; xor.b32 %r1172, %r1075, %r1171; add.s32 %r1173, %r1072, %r1170; sub.s32 %r1174, %r1170, %r1172; shf.l.wrap.b32 %r1175, %r1172, %r1172, 8; xor.b32 %r1176, %r1175, %r1174; add.s32 %r1177, %r1172, %r1173; sub.s32 %r1178, %r1173, %r1176; shf.l.wrap.b32 %r1179, %r1176, %r1176, 16; xor.b32 %r1180, %r1179, %r1178; add.s32 %r1181, %r1176, %r1177; sub.s32 %r1182, %r1177, %r1180; shf.l.wrap.b32 %r1183, %r1180, %r1180, 19; xor.b32 %r1184, %r1183, %r1182; add.s32 %r1185, %r1180, %r1181; sub.s32 %r1186, %r1181, %r1184; shf.l.wrap.b32 %r1187, %r1184, %r1184, 4; add.s32 %r1188, %r1184, %r1185; add.s32 %r1189, %r1185, %r1067; xor.b32 %r1190, %r1188, %r1186; xor.b32 %r1191, %r1190, %r1187; shf.l.wrap.b32 %r1192, %r1188, %r1188, 14; sub.s32 %r1193, %r1191, %r1192; xor.b32 %r1194, %r1193, %r1189; shf.l.wrap.b32 %r1195, %r1193, %r1193, 11; sub.s32 %r1196, %r1194, %r1195; xor.b32 %r1197, %r1196, %r1188; shf.l.wrap.b32 %r1198, %r1196, %r1196, 25; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1193; shf.l.wrap.b32 %r1201, %r1199, %r1199, 16; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1196; shf.l.wrap.b32 %r1204, %r1202, %r1202, 4; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 14; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 24; sub.s32 %r1211, %r1209, %r1210; add.ftz.f32 %f1121, %f1072, 0fBF800000; and.b32 %r1212, %r1211, 31; setp.lt.u32 %p137, %r1212, 24; selp.f32 %f1122, %f1070, %f1121, %p137; setp.lt.u32 %p138, %r1212, 16; selp.f32 %f1123, %f1121, %f1074, %p138; setp.lt.u32 %p139, %r1212, 8; selp.f32 %f1124, %f1074, %f1076, %p139; and.b32 %r1213, %r1211, 1; setp.eq.b32 %p140, %r1213, 1; neg.ftz.f32 %f1125, %f1122; selp.f32 %f1126, %f1125, %f1122, %p140; and.b32 %r1214, %r1211, 2; setp.eq.s32 %p141, %r1214, 0; neg.ftz.f32 %f1127, %f1123; selp.f32 %f1128, %f1123, %f1127, %p141; add.ftz.f32 %f1129, %f1126, %f1128; and.b32 %r1215, %r1211, 4; setp.eq.s32 %p142, %r1215, 0; neg.ftz.f32 %f1130, %f1124; selp.f32 %f1131, %f1124, %f1130, %p142; add.ftz.f32 %f1132, %f1131, %f1129; sub.s32 %r1216, %r1169, %r1122; xor.b32 %r1217, %r1124, %r1216; add.s32 %r1218, %r1122, %r1170; sub.s32 %r1219, %r1170, %r1217; shf.l.wrap.b32 %r1220, %r1217, %r1217, 8; xor.b32 %r1221, %r1220, %r1219; add.s32 %r1222, %r1217, %r1218; sub.s32 %r1223, %r1218, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 16; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 19; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 4; add.s32 %r1233, %r1229, %r1230; add.s32 %r1234, %r1230, %r1067; xor.b32 %r1235, %r1233, %r1231; xor.b32 %r1236, %r1235, %r1232; shf.l.wrap.b32 %r1237, %r1233, %r1233, 14; sub.s32 %r1238, %r1236, %r1237; xor.b32 %r1239, %r1238, %r1234; shf.l.wrap.b32 %r1240, %r1238, %r1238, 11; sub.s32 %r1241, %r1239, %r1240; xor.b32 %r1242, %r1241, %r1233; shf.l.wrap.b32 %r1243, %r1241, %r1241, 25; sub.s32 %r1244, %r1242, %r1243; xor.b32 %r1245, %r1244, %r1238; shf.l.wrap.b32 %r1246, %r1244, %r1244, 16; sub.s32 %r1247, %r1245, %r1246; xor.b32 %r1248, %r1247, %r1241; shf.l.wrap.b32 %r1249, %r1247, %r1247, 4; sub.s32 %r1250, %r1248, %r1249; xor.b32 %r1251, %r1250, %r1244; shf.l.wrap.b32 %r1252, %r1250, %r1250, 14; sub.s32 %r1253, %r1251, %r1252; xor.b32 %r1254, %r1253, %r1247; shf.l.wrap.b32 %r1255, %r1253, %r1253, 24; sub.s32 %r1256, %r1254, %r1255; and.b32 %r1257, %r1256, 31; setp.lt.u32 %p143, %r1257, 24; selp.f32 %f1133, %f1108, %f1121, %p143; setp.lt.u32 %p144, %r1257, 16; selp.f32 %f1134, %f1121, %f1074, %p144; setp.lt.u32 %p145, %r1257, 8; selp.f32 %f1135, %f1074, %f1076, %p145; and.b32 %r1258, %r1256, 1; setp.eq.b32 %p146, %r1258, 1; neg.ftz.f32 %f1136, %f1133; selp.f32 %f1137, %f1136, %f1133, %p146; and.b32 %r1259, %r1256, 2; setp.eq.s32 %p147, %r1259, 0; neg.ftz.f32 %f1138, %f1134; selp.f32 %f1139, %f1134, %f1138, %p147; add.ftz.f32 %f1140, %f1137, %f1139; and.b32 %r1260, %r1256, 4; setp.eq.s32 %p148, %r1260, 0; neg.ftz.f32 %f1141, %f1135; selp.f32 %f1142, %f1135, %f1141, %p148; add.ftz.f32 %f1143, %f1142, %f1140; not.b32 %r1261, %r1066; add.s32 %r1262, %r1066, -559038707; add.s32 %r1263, %r1064, %r1261; shf.l.wrap.b32 %r1264, %r1262, %r1262, 4; xor.b32 %r1265, %r1264, %r1263; add.s32 %r1266, %r1262, %r1068; sub.s32 %r1267, %r1068, %r1265; shf.l.wrap.b32 %r1268, %r1265, %r1265, 6; xor.b32 %r1269, %r1268, %r1267; add.s32 %r1270, %r1265, %r1266; sub.s32 %r1271, %r1266, %r1269; shf.l.wrap.b32 %r1272, %r1269, %r1269, 8; xor.b32 %r1273, %r1272, %r1271; add.s32 %r1274, %r1269, %r1270; sub.s32 %r1275, %r1270, %r1273; shf.l.wrap.b32 %r1276, %r1273, %r1273, 16; xor.b32 %r1277, %r1276, %r1275; add.s32 %r1278, %r1273, %r1274; sub.s32 %r1279, %r1274, %r1277; shf.l.wrap.b32 %r1280, %r1277, %r1277, 19; xor.b32 %r1281, %r1280, %r1279; add.s32 %r1282, %r1277, %r1278; sub.s32 %r1283, %r1278, %r1281; shf.l.wrap.b32 %r1284, %r1281, %r1281, 4; add.s32 %r1285, %r1281, %r1282; add.s32 %r1286, %r1282, %r1067; xor.b32 %r1287, %r1285, %r1283; xor.b32 %r1288, %r1287, %r1284; shf.l.wrap.b32 %r1289, %r1285, %r1285, 14; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1286; shf.l.wrap.b32 %r1292, %r1290, %r1290, 11; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1285; shf.l.wrap.b32 %r1295, %r1293, %r1293, 25; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 16; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 4; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 14; sub.s32 %r1305, %r1303, %r1304; xor.b32 %r1306, %r1305, %r1299; shf.l.wrap.b32 %r1307, %r1305, %r1305, 24; sub.s32 %r1308, %r1306, %r1307; add.ftz.f32 %f1144, %f1074, 0fBF800000; and.b32 %r1309, %r1308, 31; setp.lt.u32 %p149, %r1309, 24; selp.f32 %f1145, %f1070, %f1072, %p149; setp.lt.u32 %p150, %r1309, 16; selp.f32 %f1146, %f1072, %f1144, %p150; setp.lt.u32 %p151, %r1309, 8; selp.f32 %f1147, %f1144, %f1076, %p151; and.b32 %r1310, %r1308, 1; setp.eq.b32 %p152, %r1310, 1; neg.ftz.f32 %f1148, %f1145; selp.f32 %f1149, %f1148, %f1145, %p152; and.b32 %r1311, %r1308, 2; setp.eq.s32 %p153, %r1311, 0; neg.ftz.f32 %f1150, %f1146; selp.f32 %f1151, %f1146, %f1150, %p153; add.ftz.f32 %f1152, %f1149, %f1151; and.b32 %r1312, %r1308, 4; setp.eq.s32 %p154, %r1312, 0; neg.ftz.f32 %f1153, %f1147; selp.f32 %f1154, %f1147, %f1153, %p154; add.ftz.f32 %f1155, %f1154, %f1152; xor.b32 %r1313, %r1264, %r1070; sub.s32 %r1314, %r1068, %r1313; shf.l.wrap.b32 %r1315, %r1313, %r1313, 6; xor.b32 %r1316, %r1315, %r1314; add.s32 %r1317, %r1313, %r1266; sub.s32 %r1318, %r1266, %r1316; shf.l.wrap.b32 %r1319, %r1316, %r1316, 8; xor.b32 %r1320, %r1319, %r1318; add.s32 %r1321, %r1316, %r1317; sub.s32 %r1322, %r1317, %r1320; shf.l.wrap.b32 %r1323, %r1320, %r1320, 16; xor.b32 %r1324, %r1323, %r1322; add.s32 %r1325, %r1320, %r1321; sub.s32 %r1326, %r1321, %r1324; shf.l.wrap.b32 %r1327, %r1324, %r1324, 19; xor.b32 %r1328, %r1327, %r1326; add.s32 %r1329, %r1324, %r1325; sub.s32 %r1330, %r1325, %r1328; shf.l.wrap.b32 %r1331, %r1328, %r1328, 4; add.s32 %r1332, %r1328, %r1329; add.s32 %r1333, %r1329, %r1067; xor.b32 %r1334, %r1332, %r1330; xor.b32 %r1335, %r1334, %r1331; shf.l.wrap.b32 %r1336, %r1332, %r1332, 14; sub.s32 %r1337, %r1335, %r1336; xor.b32 %r1338, %r1337, %r1333; shf.l.wrap.b32 %r1339, %r1337, %r1337, 11; sub.s32 %r1340, %r1338, %r1339; xor.b32 %r1341, %r1340, %r1332; shf.l.wrap.b32 %r1342, %r1340, %r1340, 25; sub.s32 %r1343, %r1341, %r1342; xor.b32 %r1344, %r1343, %r1337; shf.l.wrap.b32 %r1345, %r1343, %r1343, 16; sub.s32 %r1346, %r1344, %r1345; xor.b32 %r1347, %r1346, %r1340; shf.l.wrap.b32 %r1348, %r1346, %r1346, 4; sub.s32 %r1349, %r1347, %r1348; xor.b32 %r1350, %r1349, %r1343; shf.l.wrap.b32 %r1351, %r1349, %r1349, 14; sub.s32 %r1352, %r1350, %r1351; xor.b32 %r1353, %r1352, %r1346; shf.l.wrap.b32 %r1354, %r1352, %r1352, 24; sub.s32 %r1355, %r1353, %r1354; and.b32 %r1356, %r1355, 31; setp.lt.u32 %p155, %r1356, 24; selp.f32 %f1156, %f1108, %f1072, %p155; setp.lt.u32 %p156, %r1356, 16; selp.f32 %f1157, %f1072, %f1144, %p156; setp.lt.u32 %p157, %r1356, 8; selp.f32 %f1158, %f1144, %f1076, %p157; and.b32 %r1357, %r1355, 1; setp.eq.b32 %p158, %r1357, 1; neg.ftz.f32 %f1159, %f1156; selp.f32 %f1160, %f1159, %f1156, %p158; and.b32 %r1358, %r1355, 2; setp.eq.s32 %p159, %r1358, 0; neg.ftz.f32 %f1161, %f1157; selp.f32 %f1162, %f1157, %f1161, %p159; add.ftz.f32 %f1163, %f1160, %f1162; and.b32 %r1359, %r1355, 4; setp.eq.s32 %p160, %r1359, 0; neg.ftz.f32 %f1164, %f1158; selp.f32 %f1165, %f1158, %f1164, %p160; add.ftz.f32 %f1166, %f1165, %f1163; add.s32 %r1360, %r1262, %r1169; sub.s32 %r1361, %r1169, %r1265; xor.b32 %r1362, %r1268, %r1361; add.s32 %r1363, %r1265, %r1360; sub.s32 %r1364, %r1360, %r1362; shf.l.wrap.b32 %r1365, %r1362, %r1362, 8; xor.b32 %r1366, %r1365, %r1364; add.s32 %r1367, %r1362, %r1363; sub.s32 %r1368, %r1363, %r1366; shf.l.wrap.b32 %r1369, %r1366, %r1366, 16; xor.b32 %r1370, %r1369, %r1368; add.s32 %r1371, %r1366, %r1367; sub.s32 %r1372, %r1367, %r1370; shf.l.wrap.b32 %r1373, %r1370, %r1370, 19; xor.b32 %r1374, %r1373, %r1372; add.s32 %r1375, %r1370, %r1371; sub.s32 %r1376, %r1371, %r1374; shf.l.wrap.b32 %r1377, %r1374, %r1374, 4; add.s32 %r1378, %r1374, %r1375; add.s32 %r1379, %r1375, %r1067; xor.b32 %r1380, %r1378, %r1376; xor.b32 %r1381, %r1380, %r1377; shf.l.wrap.b32 %r1382, %r1378, %r1378, 14; sub.s32 %r1383, %r1381, %r1382; xor.b32 %r1384, %r1383, %r1379; shf.l.wrap.b32 %r1385, %r1383, %r1383, 11; sub.s32 %r1386, %r1384, %r1385; xor.b32 %r1387, %r1386, %r1378; shf.l.wrap.b32 %r1388, %r1386, %r1386, 25; sub.s32 %r1389, %r1387, %r1388; xor.b32 %r1390, %r1389, %r1383; shf.l.wrap.b32 %r1391, %r1389, %r1389, 16; sub.s32 %r1392, %r1390, %r1391; xor.b32 %r1393, %r1392, %r1386; shf.l.wrap.b32 %r1394, %r1392, %r1392, 4; sub.s32 %r1395, %r1393, %r1394; xor.b32 %r1396, %r1395, %r1389; shf.l.wrap.b32 %r1397, %r1395, %r1395, 14; sub.s32 %r1398, %r1396, %r1397; xor.b32 %r1399, %r1398, %r1392; shf.l.wrap.b32 %r1400, %r1398, %r1398, 24; sub.s32 %r1401, %r1399, %r1400; and.b32 %r1402, %r1401, 31; setp.lt.u32 %p161, %r1402, 24; selp.f32 %f1167, %f1070, %f1121, %p161; setp.lt.u32 %p162, %r1402, 16; selp.f32 %f1168, %f1121, %f1144, %p162; setp.lt.u32 %p163, %r1402, 8; selp.f32 %f1169, %f1144, %f1076, %p163; and.b32 %r1403, %r1401, 1; setp.eq.b32 %p164, %r1403, 1; neg.ftz.f32 %f1170, %f1167; selp.f32 %f1171, %f1170, %f1167, %p164; and.b32 %r1404, %r1401, 2; setp.eq.s32 %p165, %r1404, 0; neg.ftz.f32 %f1172, %f1168; selp.f32 %f1173, %f1168, %f1172, %p165; add.ftz.f32 %f1174, %f1171, %f1173; and.b32 %r1405, %r1401, 4; setp.eq.s32 %p166, %r1405, 0; neg.ftz.f32 %f1175, %f1169; selp.f32 %f1176, %f1169, %f1175, %p166; add.ftz.f32 %f1177, %f1176, %f1174; sub.s32 %r1406, %r1169, %r1313; xor.b32 %r1407, %r1315, %r1406; add.s32 %r1408, %r1313, %r1360; sub.s32 %r1409, %r1360, %r1407; shf.l.wrap.b32 %r1410, %r1407, %r1407, 8; xor.b32 %r1411, %r1410, %r1409; add.s32 %r1412, %r1407, %r1408; sub.s32 %r1413, %r1408, %r1411; shf.l.wrap.b32 %r1414, %r1411, %r1411, 16; xor.b32 %r1415, %r1414, %r1413; add.s32 %r1416, %r1411, %r1412; sub.s32 %r1417, %r1412, %r1415; shf.l.wrap.b32 %r1418, %r1415, %r1415, 19; xor.b32 %r1419, %r1418, %r1417; add.s32 %r1420, %r1415, %r1416; sub.s32 %r1421, %r1416, %r1419; shf.l.wrap.b32 %r1422, %r1419, %r1419, 4; add.s32 %r1423, %r1419, %r1420; add.s32 %r1424, %r1420, %r1067; xor.b32 %r1425, %r1423, %r1421; xor.b32 %r1426, %r1425, %r1422; shf.l.wrap.b32 %r1427, %r1423, %r1423, 14; sub.s32 %r1428, %r1426, %r1427; xor.b32 %r1429, %r1428, %r1424; shf.l.wrap.b32 %r1430, %r1428, %r1428, 11; sub.s32 %r1431, %r1429, %r1430; xor.b32 %r1432, %r1431, %r1423; shf.l.wrap.b32 %r1433, %r1431, %r1431, 25; sub.s32 %r1434, %r1432, %r1433; xor.b32 %r1435, %r1434, %r1428; shf.l.wrap.b32 %r1436, %r1434, %r1434, 16; sub.s32 %r1437, %r1435, %r1436; xor.b32 %r1438, %r1437, %r1431; shf.l.wrap.b32 %r1439, %r1437, %r1437, 4; sub.s32 %r1440, %r1438, %r1439; xor.b32 %r1441, %r1440, %r1434; shf.l.wrap.b32 %r1442, %r1440, %r1440, 14; sub.s32 %r1443, %r1441, %r1442; xor.b32 %r1444, %r1443, %r1437; shf.l.wrap.b32 %r1445, %r1443, %r1443, 24; sub.s32 %r1446, %r1444, %r1445; and.b32 %r1447, %r1446, 31; setp.lt.u32 %p167, %r1447, 24; selp.f32 %f1178, %f1108, %f1121, %p167; setp.lt.u32 %p168, %r1447, 16; selp.f32 %f1179, %f1121, %f1144, %p168; setp.lt.u32 %p169, %r1447, 8; selp.f32 %f1180, %f1144, %f1076, %p169; and.b32 %r1448, %r1446, 1; setp.eq.b32 %p170, %r1448, 1; neg.ftz.f32 %f1181, %f1178; selp.f32 %f1182, %f1181, %f1178, %p170; and.b32 %r1449, %r1446, 2; setp.eq.s32 %p171, %r1449, 0; neg.ftz.f32 %f1183, %f1179; selp.f32 %f1184, %f1179, %f1183, %p171; add.ftz.f32 %f1185, %f1182, %f1184; and.b32 %r1450, %r1446, 4; setp.eq.s32 %p172, %r1450, 0; neg.ftz.f32 %f1186, %f1180; selp.f32 %f1187, %f1180, %f1186, %p172; add.ftz.f32 %f1188, %f1187, %f1185; add.s32 %r1451, %r1067, 1; add.s32 %r1452, %r1089, %r1451; xor.b32 %r1453, %r1097, %r1452; sub.s32 %r1454, %r1453, %r1099; xor.b32 %r1455, %r1454, %r1092; shf.l.wrap.b32 %r1456, %r1454, %r1454, 25; sub.s32 %r1457, %r1455, %r1456; xor.b32 %r1458, %r1457, %r1097; shf.l.wrap.b32 %r1459, %r1457, %r1457, 16; sub.s32 %r1460, %r1458, %r1459; xor.b32 %r1461, %r1460, %r1454; shf.l.wrap.b32 %r1462, %r1460, %r1460, 4; sub.s32 %r1463, %r1461, %r1462; xor.b32 %r1464, %r1463, %r1457; shf.l.wrap.b32 %r1465, %r1463, %r1463, 14; sub.s32 %r1466, %r1464, %r1465; xor.b32 %r1467, %r1466, %r1460; shf.l.wrap.b32 %r1468, %r1466, %r1466, 24; sub.s32 %r1469, %r1467, %r1468; add.ftz.f32 %f1189, %f1076, 0fBF800000; and.b32 %r1470, %r1469, 31; setp.lt.u32 %p173, %r1470, 24; selp.f32 %f1190, %f1070, %f1072, %p173; setp.lt.u32 %p174, %r1470, 16; selp.f32 %f1191, %f1072, %f1074, %p174; setp.lt.u32 %p175, %r1470, 8; selp.f32 %f1192, %f1074, %f1189, %p175; and.b32 %r1471, %r1469, 1; setp.eq.b32 %p176, %r1471, 1; neg.ftz.f32 %f1193, %f1190; selp.f32 %f1194, %f1193, %f1190, %p176; and.b32 %r1472, %r1469, 2; setp.eq.s32 %p177, %r1472, 0; neg.ftz.f32 %f1195, %f1191; selp.f32 %f1196, %f1191, %f1195, %p177; add.ftz.f32 %f1197, %f1194, %f1196; and.b32 %r1473, %r1469, 4; setp.eq.s32 %p178, %r1473, 0; neg.ftz.f32 %f1198, %f1192; selp.f32 %f1199, %f1192, %f1198, %p178; add.ftz.f32 %f1200, %f1199, %f1197; add.s32 %r1474, %r1138, %r1451; xor.b32 %r1475, %r1146, %r1474; sub.s32 %r1476, %r1475, %r1148; xor.b32 %r1477, %r1476, %r1141; shf.l.wrap.b32 %r1478, %r1476, %r1476, 25; sub.s32 %r1479, %r1477, %r1478; xor.b32 %r1480, %r1479, %r1146; shf.l.wrap.b32 %r1481, %r1479, %r1479, 16; sub.s32 %r1482, %r1480, %r1481; xor.b32 %r1483, %r1482, %r1476; shf.l.wrap.b32 %r1484, %r1482, %r1482, 4; sub.s32 %r1485, %r1483, %r1484; xor.b32 %r1486, %r1485, %r1479; shf.l.wrap.b32 %r1487, %r1485, %r1485, 14; sub.s32 %r1488, %r1486, %r1487; xor.b32 %r1489, %r1488, %r1482; shf.l.wrap.b32 %r1490, %r1488, %r1488, 24; sub.s32 %r1491, %r1489, %r1490; and.b32 %r1492, %r1491, 31; setp.lt.u32 %p179, %r1492, 24; selp.f32 %f1201, %f1108, %f1072, %p179; setp.lt.u32 %p180, %r1492, 16; selp.f32 %f1202, %f1072, %f1074, %p180; setp.lt.u32 %p181, %r1492, 8; selp.f32 %f1203, %f1074, %f1189, %p181; and.b32 %r1493, %r1491, 1; setp.eq.b32 %p182, %r1493, 1; neg.ftz.f32 %f1204, %f1201; selp.f32 %f1205, %f1204, %f1201, %p182; and.b32 %r1494, %r1491, 2; setp.eq.s32 %p183, %r1494, 0; neg.ftz.f32 %f1206, %f1202; selp.f32 %f1207, %f1202, %f1206, %p183; add.ftz.f32 %f1208, %f1205, %f1207; and.b32 %r1495, %r1491, 4; setp.eq.s32 %p184, %r1495, 0; neg.ftz.f32 %f1209, %f1203; selp.f32 %f1210, %f1203, %f1209, %p184; add.ftz.f32 %f1211, %f1210, %f1208; add.s32 %r1496, %r1185, %r1451; xor.b32 %r1497, %r1193, %r1496; sub.s32 %r1498, %r1497, %r1195; xor.b32 %r1499, %r1498, %r1188; shf.l.wrap.b32 %r1500, %r1498, %r1498, 25; sub.s32 %r1501, %r1499, %r1500; xor.b32 %r1502, %r1501, %r1193; shf.l.wrap.b32 %r1503, %r1501, %r1501, 16; sub.s32 %r1504, %r1502, %r1503; xor.b32 %r1505, %r1504, %r1498; shf.l.wrap.b32 %r1506, %r1504, %r1504, 4; sub.s32 %r1507, %r1505, %r1506; xor.b32 %r1508, %r1507, %r1501; shf.l.wrap.b32 %r1509, %r1507, %r1507, 14; sub.s32 %r1510, %r1508, %r1509; xor.b32 %r1511, %r1510, %r1504; shf.l.wrap.b32 %r1512, %r1510, %r1510, 24; sub.s32 %r1513, %r1511, %r1512; and.b32 %r1514, %r1513, 31; setp.lt.u32 %p185, %r1514, 24; selp.f32 %f1212, %f1070, %f1121, %p185; setp.lt.u32 %p186, %r1514, 16; selp.f32 %f1213, %f1121, %f1074, %p186; setp.lt.u32 %p187, %r1514, 8; selp.f32 %f1214, %f1074, %f1189, %p187; and.b32 %r1515, %r1513, 1; setp.eq.b32 %p188, %r1515, 1; neg.ftz.f32 %f1215, %f1212; selp.f32 %f1216, %f1215, %f1212, %p188; and.b32 %r1516, %r1513, 2; setp.eq.s32 %p189, %r1516, 0; neg.ftz.f32 %f1217, %f1213; selp.f32 %f1218, %f1213, %f1217, %p189; add.ftz.f32 %f1219, %f1216, %f1218; and.b32 %r1517, %r1513, 4; setp.eq.s32 %p190, %r1517, 0; neg.ftz.f32 %f1220, %f1214; selp.f32 %f1221, %f1214, %f1220, %p190; add.ftz.f32 %f1222, %f1221, %f1219; add.s32 %r1518, %r1230, %r1451; xor.b32 %r1519, %r1238, %r1518; sub.s32 %r1520, %r1519, %r1240; xor.b32 %r1521, %r1520, %r1233; shf.l.wrap.b32 %r1522, %r1520, %r1520, 25; sub.s32 %r1523, %r1521, %r1522; xor.b32 %r1524, %r1523, %r1238; shf.l.wrap.b32 %r1525, %r1523, %r1523, 16; sub.s32 %r1526, %r1524, %r1525; xor.b32 %r1527, %r1526, %r1520; shf.l.wrap.b32 %r1528, %r1526, %r1526, 4; sub.s32 %r1529, %r1527, %r1528; xor.b32 %r1530, %r1529, %r1523; shf.l.wrap.b32 %r1531, %r1529, %r1529, 14; sub.s32 %r1532, %r1530, %r1531; xor.b32 %r1533, %r1532, %r1526; shf.l.wrap.b32 %r1534, %r1532, %r1532, 24; sub.s32 %r1535, %r1533, %r1534; and.b32 %r1536, %r1535, 31; setp.lt.u32 %p191, %r1536, 24; selp.f32 %f1223, %f1108, %f1121, %p191; setp.lt.u32 %p192, %r1536, 16; selp.f32 %f1224, %f1121, %f1074, %p192; setp.lt.u32 %p193, %r1536, 8; selp.f32 %f1225, %f1074, %f1189, %p193; and.b32 %r1537, %r1535, 1; setp.eq.b32 %p194, %r1537, 1; neg.ftz.f32 %f1226, %f1223; selp.f32 %f1227, %f1226, %f1223, %p194; and.b32 %r1538, %r1535, 2; setp.eq.s32 %p195, %r1538, 0; neg.ftz.f32 %f1228, %f1224; selp.f32 %f1229, %f1224, %f1228, %p195; add.ftz.f32 %f1230, %f1227, %f1229; and.b32 %r1539, %r1535, 4; setp.eq.s32 %p196, %r1539, 0; neg.ftz.f32 %f1231, %f1225; selp.f32 %f1232, %f1225, %f1231, %p196; add.ftz.f32 %f1233, %f1232, %f1230; add.s32 %r1540, %r1282, %r1451; xor.b32 %r1541, %r1290, %r1540; sub.s32 %r1542, %r1541, %r1292; xor.b32 %r1543, %r1542, %r1285; shf.l.wrap.b32 %r1544, %r1542, %r1542, 25; sub.s32 %r1545, %r1543, %r1544; xor.b32 %r1546, %r1545, %r1290; shf.l.wrap.b32 %r1547, %r1545, %r1545, 16; sub.s32 %r1548, %r1546, %r1547; xor.b32 %r1549, %r1548, %r1542; shf.l.wrap.b32 %r1550, %r1548, %r1548, 4; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r1545; shf.l.wrap.b32 %r1553, %r1551, %r1551, 14; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 24; sub.s32 %r1557, %r1555, %r1556; and.b32 %r1558, %r1557, 31; setp.lt.u32 %p197, %r1558, 24; selp.f32 %f1234, %f1070, %f1072, %p197; setp.lt.u32 %p198, %r1558, 16; selp.f32 %f1235, %f1072, %f1144, %p198; setp.lt.u32 %p199, %r1558, 8; selp.f32 %f1236, %f1144, %f1189, %p199; and.b32 %r1559, %r1557, 1; setp.eq.b32 %p200, %r1559, 1; neg.ftz.f32 %f1237, %f1234; selp.f32 %f1238, %f1237, %f1234, %p200; and.b32 %r1560, %r1557, 2; setp.eq.s32 %p201, %r1560, 0; neg.ftz.f32 %f1239, %f1235; selp.f32 %f1240, %f1235, %f1239, %p201; add.ftz.f32 %f1241, %f1238, %f1240; and.b32 %r1561, %r1557, 4; setp.eq.s32 %p202, %r1561, 0; neg.ftz.f32 %f1242, %f1236; selp.f32 %f1243, %f1236, %f1242, %p202; add.ftz.f32 %f1244, %f1243, %f1241; add.s32 %r1562, %r1329, %r1451; xor.b32 %r1563, %r1337, %r1562; sub.s32 %r1564, %r1563, %r1339; xor.b32 %r1565, %r1564, %r1332; shf.l.wrap.b32 %r1566, %r1564, %r1564, 25; sub.s32 %r1567, %r1565, %r1566; xor.b32 %r1568, %r1567, %r1337; shf.l.wrap.b32 %r1569, %r1567, %r1567, 16; sub.s32 %r1570, %r1568, %r1569; xor.b32 %r1571, %r1570, %r1564; shf.l.wrap.b32 %r1572, %r1570, %r1570, 4; sub.s32 %r1573, %r1571, %r1572; xor.b32 %r1574, %r1573, %r1567; shf.l.wrap.b32 %r1575, %r1573, %r1573, 14; sub.s32 %r1576, %r1574, %r1575; xor.b32 %r1577, %r1576, %r1570; shf.l.wrap.b32 %r1578, %r1576, %r1576, 24; sub.s32 %r1579, %r1577, %r1578; and.b32 %r1580, %r1579, 31; setp.lt.u32 %p203, %r1580, 24; selp.f32 %f1245, %f1108, %f1072, %p203; setp.lt.u32 %p204, %r1580, 16; selp.f32 %f1246, %f1072, %f1144, %p204; setp.lt.u32 %p205, %r1580, 8; selp.f32 %f1247, %f1144, %f1189, %p205; and.b32 %r1581, %r1579, 1; setp.eq.b32 %p206, %r1581, 1; neg.ftz.f32 %f1248, %f1245; selp.f32 %f1249, %f1248, %f1245, %p206; and.b32 %r1582, %r1579, 2; setp.eq.s32 %p207, %r1582, 0; neg.ftz.f32 %f1250, %f1246; selp.f32 %f1251, %f1246, %f1250, %p207; add.ftz.f32 %f1252, %f1249, %f1251; and.b32 %r1583, %r1579, 4; setp.eq.s32 %p208, %r1583, 0; neg.ftz.f32 %f1253, %f1247; selp.f32 %f1254, %f1247, %f1253, %p208; add.ftz.f32 %f1255, %f1254, %f1252; add.s32 %r1584, %r1375, %r1451; xor.b32 %r1585, %r1383, %r1584; sub.s32 %r1586, %r1585, %r1385; xor.b32 %r1587, %r1586, %r1378; shf.l.wrap.b32 %r1588, %r1586, %r1586, 25; sub.s32 %r1589, %r1587, %r1588; xor.b32 %r1590, %r1589, %r1383; shf.l.wrap.b32 %r1591, %r1589, %r1589, 16; sub.s32 %r1592, %r1590, %r1591; xor.b32 %r1593, %r1592, %r1586; shf.l.wrap.b32 %r1594, %r1592, %r1592, 4; sub.s32 %r1595, %r1593, %r1594; xor.b32 %r1596, %r1595, %r1589; shf.l.wrap.b32 %r1597, %r1595, %r1595, 14; sub.s32 %r1598, %r1596, %r1597; xor.b32 %r1599, %r1598, %r1592; shf.l.wrap.b32 %r1600, %r1598, %r1598, 24; sub.s32 %r1601, %r1599, %r1600; and.b32 %r1602, %r1601, 31; setp.lt.u32 %p209, %r1602, 24; selp.f32 %f1256, %f1070, %f1121, %p209; setp.lt.u32 %p210, %r1602, 16; selp.f32 %f1257, %f1121, %f1144, %p210; setp.lt.u32 %p211, %r1602, 8; selp.f32 %f1258, %f1144, %f1189, %p211; and.b32 %r1603, %r1601, 1; setp.eq.b32 %p212, %r1603, 1; neg.ftz.f32 %f1259, %f1256; selp.f32 %f1260, %f1259, %f1256, %p212; and.b32 %r1604, %r1601, 2; setp.eq.s32 %p213, %r1604, 0; neg.ftz.f32 %f1261, %f1257; selp.f32 %f1262, %f1257, %f1261, %p213; add.ftz.f32 %f1263, %f1260, %f1262; and.b32 %r1605, %r1601, 4; setp.eq.s32 %p214, %r1605, 0; neg.ftz.f32 %f1264, %f1258; selp.f32 %f1265, %f1258, %f1264, %p214; add.ftz.f32 %f1266, %f1265, %f1263; add.s32 %r1606, %r1420, %r1451; xor.b32 %r1607, %r1428, %r1606; sub.s32 %r1608, %r1607, %r1430; xor.b32 %r1609, %r1608, %r1423; shf.l.wrap.b32 %r1610, %r1608, %r1608, 25; sub.s32 %r1611, %r1609, %r1610; xor.b32 %r1612, %r1611, %r1428; shf.l.wrap.b32 %r1613, %r1611, %r1611, 16; sub.s32 %r1614, %r1612, %r1613; xor.b32 %r1615, %r1614, %r1608; shf.l.wrap.b32 %r1616, %r1614, %r1614, 4; sub.s32 %r1617, %r1615, %r1616; xor.b32 %r1618, %r1617, %r1611; shf.l.wrap.b32 %r1619, %r1617, %r1617, 14; sub.s32 %r1620, %r1618, %r1619; xor.b32 %r1621, %r1620, %r1614; shf.l.wrap.b32 %r1622, %r1620, %r1620, 24; sub.s32 %r1623, %r1621, %r1622; and.b32 %r1624, %r1623, 31; setp.lt.u32 %p215, %r1624, 24; selp.f32 %f1267, %f1108, %f1121, %p215; setp.lt.u32 %p216, %r1624, 16; selp.f32 %f1268, %f1121, %f1144, %p216; setp.lt.u32 %p217, %r1624, 8; selp.f32 %f1269, %f1144, %f1189, %p217; and.b32 %r1625, %r1623, 1; setp.eq.b32 %p218, %r1625, 1; neg.ftz.f32 %f1270, %f1267; selp.f32 %f1271, %f1270, %f1267, %p218; and.b32 %r1626, %r1623, 2; setp.eq.s32 %p219, %r1626, 0; neg.ftz.f32 %f1272, %f1268; selp.f32 %f1273, %f1268, %f1272, %p219; add.ftz.f32 %f1274, %f1271, %f1273; and.b32 %r1627, %r1623, 4; setp.eq.s32 %p220, %r1627, 0; neg.ftz.f32 %f1275, %f1269; selp.f32 %f1276, %f1269, %f1275, %p220; add.ftz.f32 %f1277, %f1276, %f1274; sub.ftz.f32 %f1278, %f864, %f1081; sub.ftz.f32 %f1279, %f864, %f1086; sub.ftz.f32 %f1280, %f864, %f1091; mul.ftz.f32 %f1281, %f1278, %f1107; fma.rn.ftz.f32 %f1282, %f1081, %f1120, %f1281; mul.ftz.f32 %f1283, %f1278, %f1132; fma.rn.ftz.f32 %f1284, %f1081, %f1143, %f1283; mul.ftz.f32 %f1285, %f1086, %f1284; fma.rn.ftz.f32 %f1286, %f1279, %f1282, %f1285; mul.ftz.f32 %f1287, %f1278, %f1155; fma.rn.ftz.f32 %f1288, %f1081, %f1166, %f1287; mul.ftz.f32 %f1289, %f1278, %f1177; fma.rn.ftz.f32 %f1290, %f1081, %f1188, %f1289; mul.ftz.f32 %f1291, %f1086, %f1290; fma.rn.ftz.f32 %f1292, %f1279, %f1288, %f1291; mul.ftz.f32 %f1293, %f1091, %f1292; fma.rn.ftz.f32 %f1294, %f1280, %f1286, %f1293; mul.ftz.f32 %f1295, %f1278, %f1200; fma.rn.ftz.f32 %f1296, %f1081, %f1211, %f1295; mul.ftz.f32 %f1297, %f1278, %f1222; fma.rn.ftz.f32 %f1298, %f1081, %f1233, %f1297; mul.ftz.f32 %f1299, %f1086, %f1298; fma.rn.ftz.f32 %f1300, %f1279, %f1296, %f1299; mul.ftz.f32 %f1301, %f1278, %f1244; fma.rn.ftz.f32 %f1302, %f1081, %f1255, %f1301; mul.ftz.f32 %f1303, %f1278, %f1266; fma.rn.ftz.f32 %f1304, %f1081, %f1277, %f1303; mul.ftz.f32 %f1305, %f1086, %f1304; fma.rn.ftz.f32 %f1306, %f1279, %f1302, %f1305; mul.ftz.f32 %f1307, %f1091, %f1306; fma.rn.ftz.f32 %f1308, %f1280, %f1300, %f1307; sub.ftz.f32 %f1309, %f1308, %f1294; fma.rn.ftz.f32 %f47, %f1096, %f1309, %f1294; mov.b32 %r19, %f47; setp.nan.ftz.f32 %p221, %f47, %f47; mov.f32 %f5528, 0f00000000; @%p221 bra $L__BB74_65; or.b32 %r1628, %r19, -2147483648; setp.ne.s32 %p222, %r1628, -2147483648; add.ftz.f32 %f1311, %f47, %f47; setp.eq.ftz.f32 %p223, %f1311, %f47; and.pred %p224, %p223, %p222; @%p224 bra $L__BB74_65; shl.b32 %r1629, %r19, 1; setp.lt.u32 %p225, %r1629, -16777215; selp.f32 %f5528, %f47, 0f00000000, %p225; $L__BB74_65: fma.rn.ftz.f32 %f1312, %f5528, 0f3F559B3D, %f21; mul.ftz.f32 %f1313, %f5533, %f1312; fma.rn.ftz.f32 %f5827, %f5827, %f1313, %f5827; mul.ftz.f32 %f5533, %f32, %f5533; mul.ftz.f32 %f5532, %f27, %f5532; mul.ftz.f32 %f5531, %f27, %f5531; mul.ftz.f32 %f5530, %f27, %f5530; mul.ftz.f32 %f5529, %f27, %f5529; add.s32 %r10018, %r10018, 1; setp.lt.s32 %p226, %r10018, %r17; @%p226 bra $L__BB74_62; $L__BB74_66: cvt.rmi.ftz.f32.f32 %f1314, %f26; sub.ftz.f32 %f62, %f26, %f1314; setp.eq.ftz.f32 %p227, %f62, 0f00000000; @%p227 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f1316, %f5532; cvt.rzi.ftz.s32.f32 %r1630, %f1316; sub.ftz.f32 %f1317, %f5532, %f1316; cvt.rmi.ftz.f32.f32 %f1318, %f5531; cvt.rzi.ftz.s32.f32 %r1631, %f1318; sub.ftz.f32 %f1319, %f5531, %f1318; cvt.rmi.ftz.f32.f32 %f1320, %f5530; cvt.rzi.ftz.s32.f32 %r1632, %f1320; sub.ftz.f32 %f1321, %f5530, %f1320; cvt.rmi.ftz.f32.f32 %f1322, %f5529; cvt.rzi.ftz.s32.f32 %r1633, %f1322; sub.ftz.f32 %f1323, %f5529, %f1322; mul.ftz.f32 %f1324, %f1317, %f1317; mul.ftz.f32 %f1325, %f1317, %f1324; fma.rn.ftz.f32 %f1326, %f1317, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1327, %f1317, %f1326, 0f41200000; mul.ftz.f32 %f1328, %f1325, %f1327; mul.ftz.f32 %f1329, %f1319, %f1319; mul.ftz.f32 %f1330, %f1319, %f1329; fma.rn.ftz.f32 %f1331, %f1319, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1332, %f1319, %f1331, 0f41200000; mul.ftz.f32 %f1333, %f1330, %f1332; mul.ftz.f32 %f1334, %f1321, %f1321; mul.ftz.f32 %f1335, %f1321, %f1334; fma.rn.ftz.f32 %f1336, %f1321, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1337, %f1321, %f1336, 0f41200000; mul.ftz.f32 %f1338, %f1335, %f1337; mul.ftz.f32 %f1339, %f1323, %f1323; mul.ftz.f32 %f1340, %f1323, %f1339; fma.rn.ftz.f32 %f1341, %f1323, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1342, %f1323, %f1341, 0f41200000; mul.ftz.f32 %f1343, %f1340, %f1342; add.s32 %r1634, %r1631, -559038708; add.s32 %r1635, %r1632, -559038708; sub.s32 %r1636, %r1630, %r1632; shf.l.wrap.b32 %r1637, %r1635, %r1635, 4; xor.b32 %r1638, %r1637, %r1636; add.s32 %r1639, %r1635, %r1634; sub.s32 %r1640, %r1634, %r1638; shf.l.wrap.b32 %r1641, %r1638, %r1638, 6; xor.b32 %r1642, %r1641, %r1640; add.s32 %r1643, %r1638, %r1639; sub.s32 %r1644, %r1639, %r1642; shf.l.wrap.b32 %r1645, %r1642, %r1642, 8; xor.b32 %r1646, %r1645, %r1644; add.s32 %r1647, %r1642, %r1643; sub.s32 %r1648, %r1643, %r1646; shf.l.wrap.b32 %r1649, %r1646, %r1646, 16; xor.b32 %r1650, %r1649, %r1648; add.s32 %r1651, %r1646, %r1647; sub.s32 %r1652, %r1647, %r1650; shf.l.wrap.b32 %r1653, %r1650, %r1650, 19; xor.b32 %r1654, %r1653, %r1652; add.s32 %r1655, %r1650, %r1651; sub.s32 %r1656, %r1651, %r1654; shf.l.wrap.b32 %r1657, %r1654, %r1654, 4; add.s32 %r1658, %r1654, %r1655; add.s32 %r1659, %r1655, %r1633; xor.b32 %r1660, %r1658, %r1656; xor.b32 %r1661, %r1660, %r1657; shf.l.wrap.b32 %r1662, %r1658, %r1658, 14; sub.s32 %r1663, %r1661, %r1662; xor.b32 %r1664, %r1663, %r1659; shf.l.wrap.b32 %r1665, %r1663, %r1663, 11; sub.s32 %r1666, %r1664, %r1665; xor.b32 %r1667, %r1666, %r1658; shf.l.wrap.b32 %r1668, %r1666, %r1666, 25; sub.s32 %r1669, %r1667, %r1668; xor.b32 %r1670, %r1669, %r1663; shf.l.wrap.b32 %r1671, %r1669, %r1669, 16; sub.s32 %r1672, %r1670, %r1671; xor.b32 %r1673, %r1672, %r1666; shf.l.wrap.b32 %r1674, %r1672, %r1672, 4; sub.s32 %r1675, %r1673, %r1674; xor.b32 %r1676, %r1675, %r1669; shf.l.wrap.b32 %r1677, %r1675, %r1675, 14; sub.s32 %r1678, %r1676, %r1677; xor.b32 %r1679, %r1678, %r1672; shf.l.wrap.b32 %r1680, %r1678, %r1678, 24; sub.s32 %r1681, %r1679, %r1680; and.b32 %r1682, %r1681, 31; setp.lt.u32 %p228, %r1682, 24; selp.f32 %f1344, %f1317, %f1319, %p228; setp.lt.u32 %p229, %r1682, 16; selp.f32 %f1345, %f1319, %f1321, %p229; setp.lt.u32 %p230, %r1682, 8; selp.f32 %f1346, %f1321, %f1323, %p230; and.b32 %r1683, %r1681, 1; setp.eq.b32 %p231, %r1683, 1; neg.ftz.f32 %f1347, %f1344; selp.f32 %f1348, %f1347, %f1344, %p231; and.b32 %r1684, %r1681, 2; setp.eq.s32 %p232, %r1684, 0; neg.ftz.f32 %f1349, %f1345; selp.f32 %f1350, %f1345, %f1349, %p232; add.ftz.f32 %f1351, %f1348, %f1350; and.b32 %r1685, %r1681, 4; setp.eq.s32 %p233, %r1685, 0; neg.ftz.f32 %f1352, %f1346; selp.f32 %f1353, %f1346, %f1352, %p233; add.ftz.f32 %f1354, %f1353, %f1351; add.s32 %r1686, %r1630, 1; sub.s32 %r1687, %r1686, %r1632; xor.b32 %r1688, %r1637, %r1687; sub.s32 %r1689, %r1634, %r1688; shf.l.wrap.b32 %r1690, %r1688, %r1688, 6; xor.b32 %r1691, %r1690, %r1689; add.s32 %r1692, %r1688, %r1639; sub.s32 %r1693, %r1639, %r1691; shf.l.wrap.b32 %r1694, %r1691, %r1691, 8; xor.b32 %r1695, %r1694, %r1693; add.s32 %r1696, %r1691, %r1692; sub.s32 %r1697, %r1692, %r1695; shf.l.wrap.b32 %r1698, %r1695, %r1695, 16; xor.b32 %r1699, %r1698, %r1697; add.s32 %r1700, %r1695, %r1696; sub.s32 %r1701, %r1696, %r1699; shf.l.wrap.b32 %r1702, %r1699, %r1699, 19; xor.b32 %r1703, %r1702, %r1701; add.s32 %r1704, %r1699, %r1700; sub.s32 %r1705, %r1700, %r1703; shf.l.wrap.b32 %r1706, %r1703, %r1703, 4; add.s32 %r1707, %r1703, %r1704; add.s32 %r1708, %r1704, %r1633; xor.b32 %r1709, %r1707, %r1705; xor.b32 %r1710, %r1709, %r1706; shf.l.wrap.b32 %r1711, %r1707, %r1707, 14; sub.s32 %r1712, %r1710, %r1711; xor.b32 %r1713, %r1712, %r1708; shf.l.wrap.b32 %r1714, %r1712, %r1712, 11; sub.s32 %r1715, %r1713, %r1714; xor.b32 %r1716, %r1715, %r1707; shf.l.wrap.b32 %r1717, %r1715, %r1715, 25; sub.s32 %r1718, %r1716, %r1717; xor.b32 %r1719, %r1718, %r1712; shf.l.wrap.b32 %r1720, %r1718, %r1718, 16; sub.s32 %r1721, %r1719, %r1720; xor.b32 %r1722, %r1721, %r1715; shf.l.wrap.b32 %r1723, %r1721, %r1721, 4; sub.s32 %r1724, %r1722, %r1723; xor.b32 %r1725, %r1724, %r1718; shf.l.wrap.b32 %r1726, %r1724, %r1724, 14; sub.s32 %r1727, %r1725, %r1726; xor.b32 %r1728, %r1727, %r1721; shf.l.wrap.b32 %r1729, %r1727, %r1727, 24; sub.s32 %r1730, %r1728, %r1729; add.ftz.f32 %f1355, %f1317, 0fBF800000; and.b32 %r1731, %r1730, 31; setp.lt.u32 %p234, %r1731, 24; selp.f32 %f1357, %f1355, %f1319, %p234; setp.lt.u32 %p235, %r1731, 16; selp.f32 %f1358, %f1319, %f1321, %p235; setp.lt.u32 %p236, %r1731, 8; selp.f32 %f1359, %f1321, %f1323, %p236; and.b32 %r1732, %r1730, 1; setp.eq.b32 %p237, %r1732, 1; neg.ftz.f32 %f1360, %f1357; selp.f32 %f1361, %f1360, %f1357, %p237; and.b32 %r1733, %r1730, 2; setp.eq.s32 %p238, %r1733, 0; neg.ftz.f32 %f1362, %f1358; selp.f32 %f1363, %f1358, %f1362, %p238; add.ftz.f32 %f1364, %f1361, %f1363; and.b32 %r1734, %r1730, 4; setp.eq.s32 %p239, %r1734, 0; neg.ftz.f32 %f1365, %f1359; selp.f32 %f1366, %f1359, %f1365, %p239; add.ftz.f32 %f1367, %f1366, %f1364; add.s32 %r1735, %r1631, -559038707; add.s32 %r1736, %r1635, %r1735; sub.s32 %r1737, %r1735, %r1638; xor.b32 %r1738, %r1641, %r1737; add.s32 %r1739, %r1638, %r1736; sub.s32 %r1740, %r1736, %r1738; shf.l.wrap.b32 %r1741, %r1738, %r1738, 8; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1738, %r1739; sub.s32 %r1744, %r1739, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 16; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 19; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 4; add.s32 %r1754, %r1750, %r1751; add.s32 %r1755, %r1751, %r1633; xor.b32 %r1756, %r1754, %r1752; xor.b32 %r1757, %r1756, %r1753; shf.l.wrap.b32 %r1758, %r1754, %r1754, 14; sub.s32 %r1759, %r1757, %r1758; xor.b32 %r1760, %r1759, %r1755; shf.l.wrap.b32 %r1761, %r1759, %r1759, 11; sub.s32 %r1762, %r1760, %r1761; xor.b32 %r1763, %r1762, %r1754; shf.l.wrap.b32 %r1764, %r1762, %r1762, 25; sub.s32 %r1765, %r1763, %r1764; xor.b32 %r1766, %r1765, %r1759; shf.l.wrap.b32 %r1767, %r1765, %r1765, 16; sub.s32 %r1768, %r1766, %r1767; xor.b32 %r1769, %r1768, %r1762; shf.l.wrap.b32 %r1770, %r1768, %r1768, 4; sub.s32 %r1771, %r1769, %r1770; xor.b32 %r1772, %r1771, %r1765; shf.l.wrap.b32 %r1773, %r1771, %r1771, 14; sub.s32 %r1774, %r1772, %r1773; xor.b32 %r1775, %r1774, %r1768; shf.l.wrap.b32 %r1776, %r1774, %r1774, 24; sub.s32 %r1777, %r1775, %r1776; add.ftz.f32 %f1368, %f1319, 0fBF800000; and.b32 %r1778, %r1777, 31; setp.lt.u32 %p240, %r1778, 24; selp.f32 %f1369, %f1317, %f1368, %p240; setp.lt.u32 %p241, %r1778, 16; selp.f32 %f1370, %f1368, %f1321, %p241; setp.lt.u32 %p242, %r1778, 8; selp.f32 %f1371, %f1321, %f1323, %p242; and.b32 %r1779, %r1777, 1; setp.eq.b32 %p243, %r1779, 1; neg.ftz.f32 %f1372, %f1369; selp.f32 %f1373, %f1372, %f1369, %p243; and.b32 %r1780, %r1777, 2; setp.eq.s32 %p244, %r1780, 0; neg.ftz.f32 %f1374, %f1370; selp.f32 %f1375, %f1370, %f1374, %p244; add.ftz.f32 %f1376, %f1373, %f1375; and.b32 %r1781, %r1777, 4; setp.eq.s32 %p245, %r1781, 0; neg.ftz.f32 %f1377, %f1371; selp.f32 %f1378, %f1371, %f1377, %p245; add.ftz.f32 %f1379, %f1378, %f1376; sub.s32 %r1782, %r1735, %r1688; xor.b32 %r1783, %r1690, %r1782; add.s32 %r1784, %r1688, %r1736; sub.s32 %r1785, %r1736, %r1783; shf.l.wrap.b32 %r1786, %r1783, %r1783, 8; xor.b32 %r1787, %r1786, %r1785; add.s32 %r1788, %r1783, %r1784; sub.s32 %r1789, %r1784, %r1787; shf.l.wrap.b32 %r1790, %r1787, %r1787, 16; xor.b32 %r1791, %r1790, %r1789; add.s32 %r1792, %r1787, %r1788; sub.s32 %r1793, %r1788, %r1791; shf.l.wrap.b32 %r1794, %r1791, %r1791, 19; xor.b32 %r1795, %r1794, %r1793; add.s32 %r1796, %r1791, %r1792; sub.s32 %r1797, %r1792, %r1795; shf.l.wrap.b32 %r1798, %r1795, %r1795, 4; add.s32 %r1799, %r1795, %r1796; add.s32 %r1800, %r1796, %r1633; xor.b32 %r1801, %r1799, %r1797; xor.b32 %r1802, %r1801, %r1798; shf.l.wrap.b32 %r1803, %r1799, %r1799, 14; sub.s32 %r1804, %r1802, %r1803; xor.b32 %r1805, %r1804, %r1800; shf.l.wrap.b32 %r1806, %r1804, %r1804, 11; sub.s32 %r1807, %r1805, %r1806; xor.b32 %r1808, %r1807, %r1799; shf.l.wrap.b32 %r1809, %r1807, %r1807, 25; sub.s32 %r1810, %r1808, %r1809; xor.b32 %r1811, %r1810, %r1804; shf.l.wrap.b32 %r1812, %r1810, %r1810, 16; sub.s32 %r1813, %r1811, %r1812; xor.b32 %r1814, %r1813, %r1807; shf.l.wrap.b32 %r1815, %r1813, %r1813, 4; sub.s32 %r1816, %r1814, %r1815; xor.b32 %r1817, %r1816, %r1810; shf.l.wrap.b32 %r1818, %r1816, %r1816, 14; sub.s32 %r1819, %r1817, %r1818; xor.b32 %r1820, %r1819, %r1813; shf.l.wrap.b32 %r1821, %r1819, %r1819, 24; sub.s32 %r1822, %r1820, %r1821; and.b32 %r1823, %r1822, 31; setp.lt.u32 %p246, %r1823, 24; selp.f32 %f1380, %f1355, %f1368, %p246; setp.lt.u32 %p247, %r1823, 16; selp.f32 %f1381, %f1368, %f1321, %p247; setp.lt.u32 %p248, %r1823, 8; selp.f32 %f1382, %f1321, %f1323, %p248; and.b32 %r1824, %r1822, 1; setp.eq.b32 %p249, %r1824, 1; neg.ftz.f32 %f1383, %f1380; selp.f32 %f1384, %f1383, %f1380, %p249; and.b32 %r1825, %r1822, 2; setp.eq.s32 %p250, %r1825, 0; neg.ftz.f32 %f1385, %f1381; selp.f32 %f1386, %f1381, %f1385, %p250; add.ftz.f32 %f1387, %f1384, %f1386; and.b32 %r1826, %r1822, 4; setp.eq.s32 %p251, %r1826, 0; neg.ftz.f32 %f1388, %f1382; selp.f32 %f1389, %f1382, %f1388, %p251; add.ftz.f32 %f1390, %f1389, %f1387; not.b32 %r1827, %r1632; add.s32 %r1828, %r1632, -559038707; add.s32 %r1829, %r1630, %r1827; shf.l.wrap.b32 %r1830, %r1828, %r1828, 4; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1828, %r1634; sub.s32 %r1833, %r1634, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 6; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 8; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 16; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 19; xor.b32 %r1847, %r1846, %r1845; add.s32 %r1848, %r1843, %r1844; sub.s32 %r1849, %r1844, %r1847; shf.l.wrap.b32 %r1850, %r1847, %r1847, 4; add.s32 %r1851, %r1847, %r1848; add.s32 %r1852, %r1848, %r1633; xor.b32 %r1853, %r1851, %r1849; xor.b32 %r1854, %r1853, %r1850; shf.l.wrap.b32 %r1855, %r1851, %r1851, 14; sub.s32 %r1856, %r1854, %r1855; xor.b32 %r1857, %r1856, %r1852; shf.l.wrap.b32 %r1858, %r1856, %r1856, 11; sub.s32 %r1859, %r1857, %r1858; xor.b32 %r1860, %r1859, %r1851; shf.l.wrap.b32 %r1861, %r1859, %r1859, 25; sub.s32 %r1862, %r1860, %r1861; xor.b32 %r1863, %r1862, %r1856; shf.l.wrap.b32 %r1864, %r1862, %r1862, 16; sub.s32 %r1865, %r1863, %r1864; xor.b32 %r1866, %r1865, %r1859; shf.l.wrap.b32 %r1867, %r1865, %r1865, 4; sub.s32 %r1868, %r1866, %r1867; xor.b32 %r1869, %r1868, %r1862; shf.l.wrap.b32 %r1870, %r1868, %r1868, 14; sub.s32 %r1871, %r1869, %r1870; xor.b32 %r1872, %r1871, %r1865; shf.l.wrap.b32 %r1873, %r1871, %r1871, 24; sub.s32 %r1874, %r1872, %r1873; add.ftz.f32 %f1391, %f1321, 0fBF800000; and.b32 %r1875, %r1874, 31; setp.lt.u32 %p252, %r1875, 24; selp.f32 %f1392, %f1317, %f1319, %p252; setp.lt.u32 %p253, %r1875, 16; selp.f32 %f1393, %f1319, %f1391, %p253; setp.lt.u32 %p254, %r1875, 8; selp.f32 %f1394, %f1391, %f1323, %p254; and.b32 %r1876, %r1874, 1; setp.eq.b32 %p255, %r1876, 1; neg.ftz.f32 %f1395, %f1392; selp.f32 %f1396, %f1395, %f1392, %p255; and.b32 %r1877, %r1874, 2; setp.eq.s32 %p256, %r1877, 0; neg.ftz.f32 %f1397, %f1393; selp.f32 %f1398, %f1393, %f1397, %p256; add.ftz.f32 %f1399, %f1396, %f1398; and.b32 %r1878, %r1874, 4; setp.eq.s32 %p257, %r1878, 0; neg.ftz.f32 %f1400, %f1394; selp.f32 %f1401, %f1394, %f1400, %p257; add.ftz.f32 %f1402, %f1401, %f1399; xor.b32 %r1879, %r1830, %r1636; sub.s32 %r1880, %r1634, %r1879; shf.l.wrap.b32 %r1881, %r1879, %r1879, 6; xor.b32 %r1882, %r1881, %r1880; add.s32 %r1883, %r1879, %r1832; sub.s32 %r1884, %r1832, %r1882; shf.l.wrap.b32 %r1885, %r1882, %r1882, 8; xor.b32 %r1886, %r1885, %r1884; add.s32 %r1887, %r1882, %r1883; sub.s32 %r1888, %r1883, %r1886; shf.l.wrap.b32 %r1889, %r1886, %r1886, 16; xor.b32 %r1890, %r1889, %r1888; add.s32 %r1891, %r1886, %r1887; sub.s32 %r1892, %r1887, %r1890; shf.l.wrap.b32 %r1893, %r1890, %r1890, 19; xor.b32 %r1894, %r1893, %r1892; add.s32 %r1895, %r1890, %r1891; sub.s32 %r1896, %r1891, %r1894; shf.l.wrap.b32 %r1897, %r1894, %r1894, 4; add.s32 %r1898, %r1894, %r1895; add.s32 %r1899, %r1895, %r1633; xor.b32 %r1900, %r1898, %r1896; xor.b32 %r1901, %r1900, %r1897; shf.l.wrap.b32 %r1902, %r1898, %r1898, 14; sub.s32 %r1903, %r1901, %r1902; xor.b32 %r1904, %r1903, %r1899; shf.l.wrap.b32 %r1905, %r1903, %r1903, 11; sub.s32 %r1906, %r1904, %r1905; xor.b32 %r1907, %r1906, %r1898; shf.l.wrap.b32 %r1908, %r1906, %r1906, 25; sub.s32 %r1909, %r1907, %r1908; xor.b32 %r1910, %r1909, %r1903; shf.l.wrap.b32 %r1911, %r1909, %r1909, 16; sub.s32 %r1912, %r1910, %r1911; xor.b32 %r1913, %r1912, %r1906; shf.l.wrap.b32 %r1914, %r1912, %r1912, 4; sub.s32 %r1915, %r1913, %r1914; xor.b32 %r1916, %r1915, %r1909; shf.l.wrap.b32 %r1917, %r1915, %r1915, 14; sub.s32 %r1918, %r1916, %r1917; xor.b32 %r1919, %r1918, %r1912; shf.l.wrap.b32 %r1920, %r1918, %r1918, 24; sub.s32 %r1921, %r1919, %r1920; and.b32 %r1922, %r1921, 31; setp.lt.u32 %p258, %r1922, 24; selp.f32 %f1403, %f1355, %f1319, %p258; setp.lt.u32 %p259, %r1922, 16; selp.f32 %f1404, %f1319, %f1391, %p259; setp.lt.u32 %p260, %r1922, 8; selp.f32 %f1405, %f1391, %f1323, %p260; and.b32 %r1923, %r1921, 1; setp.eq.b32 %p261, %r1923, 1; neg.ftz.f32 %f1406, %f1403; selp.f32 %f1407, %f1406, %f1403, %p261; and.b32 %r1924, %r1921, 2; setp.eq.s32 %p262, %r1924, 0; neg.ftz.f32 %f1408, %f1404; selp.f32 %f1409, %f1404, %f1408, %p262; add.ftz.f32 %f1410, %f1407, %f1409; and.b32 %r1925, %r1921, 4; setp.eq.s32 %p263, %r1925, 0; neg.ftz.f32 %f1411, %f1405; selp.f32 %f1412, %f1405, %f1411, %p263; add.ftz.f32 %f1413, %f1412, %f1410; add.s32 %r1926, %r1828, %r1735; sub.s32 %r1927, %r1735, %r1831; xor.b32 %r1928, %r1834, %r1927; add.s32 %r1929, %r1831, %r1926; sub.s32 %r1930, %r1926, %r1928; shf.l.wrap.b32 %r1931, %r1928, %r1928, 8; xor.b32 %r1932, %r1931, %r1930; add.s32 %r1933, %r1928, %r1929; sub.s32 %r1934, %r1929, %r1932; shf.l.wrap.b32 %r1935, %r1932, %r1932, 16; xor.b32 %r1936, %r1935, %r1934; add.s32 %r1937, %r1932, %r1933; sub.s32 %r1938, %r1933, %r1936; shf.l.wrap.b32 %r1939, %r1936, %r1936, 19; xor.b32 %r1940, %r1939, %r1938; add.s32 %r1941, %r1936, %r1937; sub.s32 %r1942, %r1937, %r1940; shf.l.wrap.b32 %r1943, %r1940, %r1940, 4; add.s32 %r1944, %r1940, %r1941; add.s32 %r1945, %r1941, %r1633; xor.b32 %r1946, %r1944, %r1942; xor.b32 %r1947, %r1946, %r1943; shf.l.wrap.b32 %r1948, %r1944, %r1944, 14; sub.s32 %r1949, %r1947, %r1948; xor.b32 %r1950, %r1949, %r1945; shf.l.wrap.b32 %r1951, %r1949, %r1949, 11; sub.s32 %r1952, %r1950, %r1951; xor.b32 %r1953, %r1952, %r1944; shf.l.wrap.b32 %r1954, %r1952, %r1952, 25; sub.s32 %r1955, %r1953, %r1954; xor.b32 %r1956, %r1955, %r1949; shf.l.wrap.b32 %r1957, %r1955, %r1955, 16; sub.s32 %r1958, %r1956, %r1957; xor.b32 %r1959, %r1958, %r1952; shf.l.wrap.b32 %r1960, %r1958, %r1958, 4; sub.s32 %r1961, %r1959, %r1960; xor.b32 %r1962, %r1961, %r1955; shf.l.wrap.b32 %r1963, %r1961, %r1961, 14; sub.s32 %r1964, %r1962, %r1963; xor.b32 %r1965, %r1964, %r1958; shf.l.wrap.b32 %r1966, %r1964, %r1964, 24; sub.s32 %r1967, %r1965, %r1966; and.b32 %r1968, %r1967, 31; setp.lt.u32 %p264, %r1968, 24; selp.f32 %f1414, %f1317, %f1368, %p264; setp.lt.u32 %p265, %r1968, 16; selp.f32 %f1415, %f1368, %f1391, %p265; setp.lt.u32 %p266, %r1968, 8; selp.f32 %f1416, %f1391, %f1323, %p266; and.b32 %r1969, %r1967, 1; setp.eq.b32 %p267, %r1969, 1; neg.ftz.f32 %f1417, %f1414; selp.f32 %f1418, %f1417, %f1414, %p267; and.b32 %r1970, %r1967, 2; setp.eq.s32 %p268, %r1970, 0; neg.ftz.f32 %f1419, %f1415; selp.f32 %f1420, %f1415, %f1419, %p268; add.ftz.f32 %f1421, %f1418, %f1420; and.b32 %r1971, %r1967, 4; setp.eq.s32 %p269, %r1971, 0; neg.ftz.f32 %f1422, %f1416; selp.f32 %f1423, %f1416, %f1422, %p269; add.ftz.f32 %f1424, %f1423, %f1421; sub.s32 %r1972, %r1735, %r1879; xor.b32 %r1973, %r1881, %r1972; add.s32 %r1974, %r1879, %r1926; sub.s32 %r1975, %r1926, %r1973; shf.l.wrap.b32 %r1976, %r1973, %r1973, 8; xor.b32 %r1977, %r1976, %r1975; add.s32 %r1978, %r1973, %r1974; sub.s32 %r1979, %r1974, %r1977; shf.l.wrap.b32 %r1980, %r1977, %r1977, 16; xor.b32 %r1981, %r1980, %r1979; add.s32 %r1982, %r1977, %r1978; sub.s32 %r1983, %r1978, %r1981; shf.l.wrap.b32 %r1984, %r1981, %r1981, 19; xor.b32 %r1985, %r1984, %r1983; add.s32 %r1986, %r1981, %r1982; sub.s32 %r1987, %r1982, %r1985; shf.l.wrap.b32 %r1988, %r1985, %r1985, 4; add.s32 %r1989, %r1985, %r1986; add.s32 %r1990, %r1986, %r1633; xor.b32 %r1991, %r1989, %r1987; xor.b32 %r1992, %r1991, %r1988; shf.l.wrap.b32 %r1993, %r1989, %r1989, 14; sub.s32 %r1994, %r1992, %r1993; xor.b32 %r1995, %r1994, %r1990; shf.l.wrap.b32 %r1996, %r1994, %r1994, 11; sub.s32 %r1997, %r1995, %r1996; xor.b32 %r1998, %r1997, %r1989; shf.l.wrap.b32 %r1999, %r1997, %r1997, 25; sub.s32 %r2000, %r1998, %r1999; xor.b32 %r2001, %r2000, %r1994; shf.l.wrap.b32 %r2002, %r2000, %r2000, 16; sub.s32 %r2003, %r2001, %r2002; xor.b32 %r2004, %r2003, %r1997; shf.l.wrap.b32 %r2005, %r2003, %r2003, 4; sub.s32 %r2006, %r2004, %r2005; xor.b32 %r2007, %r2006, %r2000; shf.l.wrap.b32 %r2008, %r2006, %r2006, 14; sub.s32 %r2009, %r2007, %r2008; xor.b32 %r2010, %r2009, %r2003; shf.l.wrap.b32 %r2011, %r2009, %r2009, 24; sub.s32 %r2012, %r2010, %r2011; and.b32 %r2013, %r2012, 31; setp.lt.u32 %p270, %r2013, 24; selp.f32 %f1425, %f1355, %f1368, %p270; setp.lt.u32 %p271, %r2013, 16; selp.f32 %f1426, %f1368, %f1391, %p271; setp.lt.u32 %p272, %r2013, 8; selp.f32 %f1427, %f1391, %f1323, %p272; and.b32 %r2014, %r2012, 1; setp.eq.b32 %p273, %r2014, 1; neg.ftz.f32 %f1428, %f1425; selp.f32 %f1429, %f1428, %f1425, %p273; and.b32 %r2015, %r2012, 2; setp.eq.s32 %p274, %r2015, 0; neg.ftz.f32 %f1430, %f1426; selp.f32 %f1431, %f1426, %f1430, %p274; add.ftz.f32 %f1432, %f1429, %f1431; and.b32 %r2016, %r2012, 4; setp.eq.s32 %p275, %r2016, 0; neg.ftz.f32 %f1433, %f1427; selp.f32 %f1434, %f1427, %f1433, %p275; add.ftz.f32 %f1435, %f1434, %f1432; add.s32 %r2017, %r1633, 1; add.s32 %r2018, %r1655, %r2017; xor.b32 %r2019, %r1663, %r2018; sub.s32 %r2020, %r2019, %r1665; xor.b32 %r2021, %r2020, %r1658; shf.l.wrap.b32 %r2022, %r2020, %r2020, 25; sub.s32 %r2023, %r2021, %r2022; xor.b32 %r2024, %r2023, %r1663; shf.l.wrap.b32 %r2025, %r2023, %r2023, 16; sub.s32 %r2026, %r2024, %r2025; xor.b32 %r2027, %r2026, %r2020; shf.l.wrap.b32 %r2028, %r2026, %r2026, 4; sub.s32 %r2029, %r2027, %r2028; xor.b32 %r2030, %r2029, %r2023; shf.l.wrap.b32 %r2031, %r2029, %r2029, 14; sub.s32 %r2032, %r2030, %r2031; xor.b32 %r2033, %r2032, %r2026; shf.l.wrap.b32 %r2034, %r2032, %r2032, 24; sub.s32 %r2035, %r2033, %r2034; add.ftz.f32 %f1436, %f1323, 0fBF800000; and.b32 %r2036, %r2035, 31; setp.lt.u32 %p276, %r2036, 24; selp.f32 %f1437, %f1317, %f1319, %p276; setp.lt.u32 %p277, %r2036, 16; selp.f32 %f1438, %f1319, %f1321, %p277; setp.lt.u32 %p278, %r2036, 8; selp.f32 %f1439, %f1321, %f1436, %p278; and.b32 %r2037, %r2035, 1; setp.eq.b32 %p279, %r2037, 1; neg.ftz.f32 %f1440, %f1437; selp.f32 %f1441, %f1440, %f1437, %p279; and.b32 %r2038, %r2035, 2; setp.eq.s32 %p280, %r2038, 0; neg.ftz.f32 %f1442, %f1438; selp.f32 %f1443, %f1438, %f1442, %p280; add.ftz.f32 %f1444, %f1441, %f1443; and.b32 %r2039, %r2035, 4; setp.eq.s32 %p281, %r2039, 0; neg.ftz.f32 %f1445, %f1439; selp.f32 %f1446, %f1439, %f1445, %p281; add.ftz.f32 %f1447, %f1446, %f1444; add.s32 %r2040, %r1704, %r2017; xor.b32 %r2041, %r1712, %r2040; sub.s32 %r2042, %r2041, %r1714; xor.b32 %r2043, %r2042, %r1707; shf.l.wrap.b32 %r2044, %r2042, %r2042, 25; sub.s32 %r2045, %r2043, %r2044; xor.b32 %r2046, %r2045, %r1712; shf.l.wrap.b32 %r2047, %r2045, %r2045, 16; sub.s32 %r2048, %r2046, %r2047; xor.b32 %r2049, %r2048, %r2042; shf.l.wrap.b32 %r2050, %r2048, %r2048, 4; sub.s32 %r2051, %r2049, %r2050; xor.b32 %r2052, %r2051, %r2045; shf.l.wrap.b32 %r2053, %r2051, %r2051, 14; sub.s32 %r2054, %r2052, %r2053; xor.b32 %r2055, %r2054, %r2048; shf.l.wrap.b32 %r2056, %r2054, %r2054, 24; sub.s32 %r2057, %r2055, %r2056; and.b32 %r2058, %r2057, 31; setp.lt.u32 %p282, %r2058, 24; selp.f32 %f1448, %f1355, %f1319, %p282; setp.lt.u32 %p283, %r2058, 16; selp.f32 %f1449, %f1319, %f1321, %p283; setp.lt.u32 %p284, %r2058, 8; selp.f32 %f1450, %f1321, %f1436, %p284; and.b32 %r2059, %r2057, 1; setp.eq.b32 %p285, %r2059, 1; neg.ftz.f32 %f1451, %f1448; selp.f32 %f1452, %f1451, %f1448, %p285; and.b32 %r2060, %r2057, 2; setp.eq.s32 %p286, %r2060, 0; neg.ftz.f32 %f1453, %f1449; selp.f32 %f1454, %f1449, %f1453, %p286; add.ftz.f32 %f1455, %f1452, %f1454; and.b32 %r2061, %r2057, 4; setp.eq.s32 %p287, %r2061, 0; neg.ftz.f32 %f1456, %f1450; selp.f32 %f1457, %f1450, %f1456, %p287; add.ftz.f32 %f1458, %f1457, %f1455; add.s32 %r2062, %r1751, %r2017; xor.b32 %r2063, %r1759, %r2062; sub.s32 %r2064, %r2063, %r1761; xor.b32 %r2065, %r2064, %r1754; shf.l.wrap.b32 %r2066, %r2064, %r2064, 25; sub.s32 %r2067, %r2065, %r2066; xor.b32 %r2068, %r2067, %r1759; shf.l.wrap.b32 %r2069, %r2067, %r2067, 16; sub.s32 %r2070, %r2068, %r2069; xor.b32 %r2071, %r2070, %r2064; shf.l.wrap.b32 %r2072, %r2070, %r2070, 4; sub.s32 %r2073, %r2071, %r2072; xor.b32 %r2074, %r2073, %r2067; shf.l.wrap.b32 %r2075, %r2073, %r2073, 14; sub.s32 %r2076, %r2074, %r2075; xor.b32 %r2077, %r2076, %r2070; shf.l.wrap.b32 %r2078, %r2076, %r2076, 24; sub.s32 %r2079, %r2077, %r2078; and.b32 %r2080, %r2079, 31; setp.lt.u32 %p288, %r2080, 24; selp.f32 %f1459, %f1317, %f1368, %p288; setp.lt.u32 %p289, %r2080, 16; selp.f32 %f1460, %f1368, %f1321, %p289; setp.lt.u32 %p290, %r2080, 8; selp.f32 %f1461, %f1321, %f1436, %p290; and.b32 %r2081, %r2079, 1; setp.eq.b32 %p291, %r2081, 1; neg.ftz.f32 %f1462, %f1459; selp.f32 %f1463, %f1462, %f1459, %p291; and.b32 %r2082, %r2079, 2; setp.eq.s32 %p292, %r2082, 0; neg.ftz.f32 %f1464, %f1460; selp.f32 %f1465, %f1460, %f1464, %p292; add.ftz.f32 %f1466, %f1463, %f1465; and.b32 %r2083, %r2079, 4; setp.eq.s32 %p293, %r2083, 0; neg.ftz.f32 %f1467, %f1461; selp.f32 %f1468, %f1461, %f1467, %p293; add.ftz.f32 %f1469, %f1468, %f1466; add.s32 %r2084, %r1796, %r2017; xor.b32 %r2085, %r1804, %r2084; sub.s32 %r2086, %r2085, %r1806; xor.b32 %r2087, %r2086, %r1799; shf.l.wrap.b32 %r2088, %r2086, %r2086, 25; sub.s32 %r2089, %r2087, %r2088; xor.b32 %r2090, %r2089, %r1804; shf.l.wrap.b32 %r2091, %r2089, %r2089, 16; sub.s32 %r2092, %r2090, %r2091; xor.b32 %r2093, %r2092, %r2086; shf.l.wrap.b32 %r2094, %r2092, %r2092, 4; sub.s32 %r2095, %r2093, %r2094; xor.b32 %r2096, %r2095, %r2089; shf.l.wrap.b32 %r2097, %r2095, %r2095, 14; sub.s32 %r2098, %r2096, %r2097; xor.b32 %r2099, %r2098, %r2092; shf.l.wrap.b32 %r2100, %r2098, %r2098, 24; sub.s32 %r2101, %r2099, %r2100; and.b32 %r2102, %r2101, 31; setp.lt.u32 %p294, %r2102, 24; selp.f32 %f1470, %f1355, %f1368, %p294; setp.lt.u32 %p295, %r2102, 16; selp.f32 %f1471, %f1368, %f1321, %p295; setp.lt.u32 %p296, %r2102, 8; selp.f32 %f1472, %f1321, %f1436, %p296; and.b32 %r2103, %r2101, 1; setp.eq.b32 %p297, %r2103, 1; neg.ftz.f32 %f1473, %f1470; selp.f32 %f1474, %f1473, %f1470, %p297; and.b32 %r2104, %r2101, 2; setp.eq.s32 %p298, %r2104, 0; neg.ftz.f32 %f1475, %f1471; selp.f32 %f1476, %f1471, %f1475, %p298; add.ftz.f32 %f1477, %f1474, %f1476; and.b32 %r2105, %r2101, 4; setp.eq.s32 %p299, %r2105, 0; neg.ftz.f32 %f1478, %f1472; selp.f32 %f1479, %f1472, %f1478, %p299; add.ftz.f32 %f1480, %f1479, %f1477; add.s32 %r2106, %r1848, %r2017; xor.b32 %r2107, %r1856, %r2106; sub.s32 %r2108, %r2107, %r1858; xor.b32 %r2109, %r2108, %r1851; shf.l.wrap.b32 %r2110, %r2108, %r2108, 25; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r1856; shf.l.wrap.b32 %r2113, %r2111, %r2111, 16; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 4; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 14; sub.s32 %r2120, %r2118, %r2119; xor.b32 %r2121, %r2120, %r2114; shf.l.wrap.b32 %r2122, %r2120, %r2120, 24; sub.s32 %r2123, %r2121, %r2122; and.b32 %r2124, %r2123, 31; setp.lt.u32 %p300, %r2124, 24; selp.f32 %f1481, %f1317, %f1319, %p300; setp.lt.u32 %p301, %r2124, 16; selp.f32 %f1482, %f1319, %f1391, %p301; setp.lt.u32 %p302, %r2124, 8; selp.f32 %f1483, %f1391, %f1436, %p302; and.b32 %r2125, %r2123, 1; setp.eq.b32 %p303, %r2125, 1; neg.ftz.f32 %f1484, %f1481; selp.f32 %f1485, %f1484, %f1481, %p303; and.b32 %r2126, %r2123, 2; setp.eq.s32 %p304, %r2126, 0; neg.ftz.f32 %f1486, %f1482; selp.f32 %f1487, %f1482, %f1486, %p304; add.ftz.f32 %f1488, %f1485, %f1487; and.b32 %r2127, %r2123, 4; setp.eq.s32 %p305, %r2127, 0; neg.ftz.f32 %f1489, %f1483; selp.f32 %f1490, %f1483, %f1489, %p305; add.ftz.f32 %f1491, %f1490, %f1488; add.s32 %r2128, %r1895, %r2017; xor.b32 %r2129, %r1903, %r2128; sub.s32 %r2130, %r2129, %r1905; xor.b32 %r2131, %r2130, %r1898; shf.l.wrap.b32 %r2132, %r2130, %r2130, 25; sub.s32 %r2133, %r2131, %r2132; xor.b32 %r2134, %r2133, %r1903; shf.l.wrap.b32 %r2135, %r2133, %r2133, 16; sub.s32 %r2136, %r2134, %r2135; xor.b32 %r2137, %r2136, %r2130; shf.l.wrap.b32 %r2138, %r2136, %r2136, 4; sub.s32 %r2139, %r2137, %r2138; xor.b32 %r2140, %r2139, %r2133; shf.l.wrap.b32 %r2141, %r2139, %r2139, 14; sub.s32 %r2142, %r2140, %r2141; xor.b32 %r2143, %r2142, %r2136; shf.l.wrap.b32 %r2144, %r2142, %r2142, 24; sub.s32 %r2145, %r2143, %r2144; and.b32 %r2146, %r2145, 31; setp.lt.u32 %p306, %r2146, 24; selp.f32 %f1492, %f1355, %f1319, %p306; setp.lt.u32 %p307, %r2146, 16; selp.f32 %f1493, %f1319, %f1391, %p307; setp.lt.u32 %p308, %r2146, 8; selp.f32 %f1494, %f1391, %f1436, %p308; and.b32 %r2147, %r2145, 1; setp.eq.b32 %p309, %r2147, 1; neg.ftz.f32 %f1495, %f1492; selp.f32 %f1496, %f1495, %f1492, %p309; and.b32 %r2148, %r2145, 2; setp.eq.s32 %p310, %r2148, 0; neg.ftz.f32 %f1497, %f1493; selp.f32 %f1498, %f1493, %f1497, %p310; add.ftz.f32 %f1499, %f1496, %f1498; and.b32 %r2149, %r2145, 4; setp.eq.s32 %p311, %r2149, 0; neg.ftz.f32 %f1500, %f1494; selp.f32 %f1501, %f1494, %f1500, %p311; add.ftz.f32 %f1502, %f1501, %f1499; add.s32 %r2150, %r1941, %r2017; xor.b32 %r2151, %r1949, %r2150; sub.s32 %r2152, %r2151, %r1951; xor.b32 %r2153, %r2152, %r1944; shf.l.wrap.b32 %r2154, %r2152, %r2152, 25; sub.s32 %r2155, %r2153, %r2154; xor.b32 %r2156, %r2155, %r1949; shf.l.wrap.b32 %r2157, %r2155, %r2155, 16; sub.s32 %r2158, %r2156, %r2157; xor.b32 %r2159, %r2158, %r2152; shf.l.wrap.b32 %r2160, %r2158, %r2158, 4; sub.s32 %r2161, %r2159, %r2160; xor.b32 %r2162, %r2161, %r2155; shf.l.wrap.b32 %r2163, %r2161, %r2161, 14; sub.s32 %r2164, %r2162, %r2163; xor.b32 %r2165, %r2164, %r2158; shf.l.wrap.b32 %r2166, %r2164, %r2164, 24; sub.s32 %r2167, %r2165, %r2166; and.b32 %r2168, %r2167, 31; setp.lt.u32 %p312, %r2168, 24; selp.f32 %f1503, %f1317, %f1368, %p312; setp.lt.u32 %p313, %r2168, 16; selp.f32 %f1504, %f1368, %f1391, %p313; setp.lt.u32 %p314, %r2168, 8; selp.f32 %f1505, %f1391, %f1436, %p314; and.b32 %r2169, %r2167, 1; setp.eq.b32 %p315, %r2169, 1; neg.ftz.f32 %f1506, %f1503; selp.f32 %f1507, %f1506, %f1503, %p315; and.b32 %r2170, %r2167, 2; setp.eq.s32 %p316, %r2170, 0; neg.ftz.f32 %f1508, %f1504; selp.f32 %f1509, %f1504, %f1508, %p316; add.ftz.f32 %f1510, %f1507, %f1509; and.b32 %r2171, %r2167, 4; setp.eq.s32 %p317, %r2171, 0; neg.ftz.f32 %f1511, %f1505; selp.f32 %f1512, %f1505, %f1511, %p317; add.ftz.f32 %f1513, %f1512, %f1510; add.s32 %r2172, %r1986, %r2017; xor.b32 %r2173, %r1994, %r2172; sub.s32 %r2174, %r2173, %r1996; xor.b32 %r2175, %r2174, %r1989; shf.l.wrap.b32 %r2176, %r2174, %r2174, 25; sub.s32 %r2177, %r2175, %r2176; xor.b32 %r2178, %r2177, %r1994; shf.l.wrap.b32 %r2179, %r2177, %r2177, 16; sub.s32 %r2180, %r2178, %r2179; xor.b32 %r2181, %r2180, %r2174; shf.l.wrap.b32 %r2182, %r2180, %r2180, 4; sub.s32 %r2183, %r2181, %r2182; xor.b32 %r2184, %r2183, %r2177; shf.l.wrap.b32 %r2185, %r2183, %r2183, 14; sub.s32 %r2186, %r2184, %r2185; xor.b32 %r2187, %r2186, %r2180; shf.l.wrap.b32 %r2188, %r2186, %r2186, 24; sub.s32 %r2189, %r2187, %r2188; and.b32 %r2190, %r2189, 31; setp.lt.u32 %p318, %r2190, 24; selp.f32 %f1514, %f1355, %f1368, %p318; setp.lt.u32 %p319, %r2190, 16; selp.f32 %f1515, %f1368, %f1391, %p319; setp.lt.u32 %p320, %r2190, 8; selp.f32 %f1516, %f1391, %f1436, %p320; and.b32 %r2191, %r2189, 1; setp.eq.b32 %p321, %r2191, 1; neg.ftz.f32 %f1517, %f1514; selp.f32 %f1518, %f1517, %f1514, %p321; and.b32 %r2192, %r2189, 2; setp.eq.s32 %p322, %r2192, 0; neg.ftz.f32 %f1519, %f1515; selp.f32 %f1520, %f1515, %f1519, %p322; add.ftz.f32 %f1521, %f1518, %f1520; and.b32 %r2193, %r2189, 4; setp.eq.s32 %p323, %r2193, 0; neg.ftz.f32 %f1522, %f1516; selp.f32 %f1523, %f1516, %f1522, %p323; add.ftz.f32 %f1524, %f1523, %f1521; sub.ftz.f32 %f1525, %f864, %f1328; sub.ftz.f32 %f1526, %f864, %f1333; sub.ftz.f32 %f1527, %f864, %f1338; mul.ftz.f32 %f1528, %f1525, %f1354; fma.rn.ftz.f32 %f1529, %f1328, %f1367, %f1528; mul.ftz.f32 %f1530, %f1525, %f1379; fma.rn.ftz.f32 %f1531, %f1328, %f1390, %f1530; mul.ftz.f32 %f1532, %f1333, %f1531; fma.rn.ftz.f32 %f1533, %f1526, %f1529, %f1532; mul.ftz.f32 %f1534, %f1525, %f1402; fma.rn.ftz.f32 %f1535, %f1328, %f1413, %f1534; mul.ftz.f32 %f1536, %f1525, %f1424; fma.rn.ftz.f32 %f1537, %f1328, %f1435, %f1536; mul.ftz.f32 %f1538, %f1333, %f1537; fma.rn.ftz.f32 %f1539, %f1526, %f1535, %f1538; mul.ftz.f32 %f1540, %f1338, %f1539; fma.rn.ftz.f32 %f1541, %f1527, %f1533, %f1540; mul.ftz.f32 %f1542, %f1525, %f1447; fma.rn.ftz.f32 %f1543, %f1328, %f1458, %f1542; mul.ftz.f32 %f1544, %f1525, %f1469; fma.rn.ftz.f32 %f1545, %f1328, %f1480, %f1544; mul.ftz.f32 %f1546, %f1333, %f1545; fma.rn.ftz.f32 %f1547, %f1526, %f1543, %f1546; mul.ftz.f32 %f1548, %f1525, %f1491; fma.rn.ftz.f32 %f1549, %f1328, %f1502, %f1548; mul.ftz.f32 %f1550, %f1525, %f1513; fma.rn.ftz.f32 %f1551, %f1328, %f1524, %f1550; mul.ftz.f32 %f1552, %f1333, %f1551; fma.rn.ftz.f32 %f1553, %f1526, %f1549, %f1552; mul.ftz.f32 %f1554, %f1338, %f1553; fma.rn.ftz.f32 %f1555, %f1527, %f1547, %f1554; sub.ftz.f32 %f1556, %f1555, %f1541; fma.rn.ftz.f32 %f63, %f1343, %f1556, %f1541; mov.b32 %r21, %f63; setp.nan.ftz.f32 %p324, %f63, %f63; mov.f32 %f5535, 0f00000000; @%p324 bra $L__BB74_70; or.b32 %r2194, %r21, -2147483648; setp.ne.s32 %p325, %r2194, -2147483648; add.ftz.f32 %f1558, %f63, %f63; setp.eq.ftz.f32 %p326, %f1558, %f63; and.pred %p327, %p326, %p325; @%p327 bra $L__BB74_70; shl.b32 %r2195, %r21, 1; setp.lt.u32 %p328, %r2195, -16777215; mul.ftz.f32 %f1559, %f63, 0f3F559B3D; selp.f32 %f5535, %f1559, 0f00000000, %p328; $L__BB74_70: add.ftz.f32 %f1560, %f21, %f5535; mul.ftz.f32 %f1561, %f5533, %f1560; mul.ftz.f32 %f1562, %f5827, %f1561; fma.rn.ftz.f32 %f5827, %f62, %f1562, %f5827; bra.uni $L__BB74_450; $L__BB74_388: mul.ftz.f32 %f5815, %f6, %f9; and.b16 %rs273, %rs2, 255; setp.gt.s16 %p1888, %rs273, 1; @%p1888 bra $L__BB74_402; bra.uni $L__BB74_389; $L__BB74_402: setp.eq.s16 %p1889, %rs273, 2; @%p1889 bra $L__BB74_428; setp.eq.s16 %p1890, %rs273, 3; @%p1890 bra $L__BB74_419; setp.ne.s16 %p1891, %rs273, 4; @%p1891 bra $L__BB74_450; lg2.approx.ftz.f32 %f5233, %f27; mul.ftz.f32 %f5234, %f25, %f5233; neg.ftz.f32 %f5235, %f5234; ex2.approx.ftz.f32 %f717, %f5235; cvt.rmi.ftz.f32.f32 %f5236, %f5815; cvt.rzi.ftz.s32.f32 %r9514, %f5236; sub.ftz.f32 %f5237, %f5815, %f5236; mul.ftz.f32 %f5238, %f5237, %f5237; mul.ftz.f32 %f5239, %f5237, %f5238; fma.rn.ftz.f32 %f5240, %f5237, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5241, %f5237, %f5240, 0f41200000; mul.ftz.f32 %f5242, %f5239, %f5241; add.s32 %r9515, %r9514, -559038720; xor.b32 %r9516, %r9515, -1874868139; add.s32 %r9517, %r9516, 29184895; xor.b32 %r9518, %r9517, -559038720; shf.l.wrap.b32 %r9519, %r9517, %r9517, 25; sub.s32 %r9520, %r9518, %r9519; xor.b32 %r9521, %r9520, -1874868139; shf.l.wrap.b32 %r9522, %r9520, %r9520, 16; sub.s32 %r9523, %r9521, %r9522; xor.b32 %r9524, %r9523, %r9517; shf.l.wrap.b32 %r9525, %r9523, %r9523, 4; sub.s32 %r9526, %r9524, %r9525; xor.b32 %r9527, %r9526, %r9520; shf.l.wrap.b32 %r9528, %r9526, %r9526, 14; sub.s32 %r9529, %r9527, %r9528; xor.b32 %r9530, %r9529, %r9523; shf.l.wrap.b32 %r9531, %r9529, %r9529, 24; sub.s32 %r9532, %r9530, %r9531; and.b32 %r9533, %r9532, 7; add.s32 %r9534, %r9533, 1; cvt.rn.f32.s32 %f5243, %r9534; and.b32 %r9535, %r9532, 8; setp.eq.s32 %p1894, %r9535, 0; neg.ftz.f32 %f5244, %f5243; selp.f32 %f5245, %f5243, %f5244, %p1894; mul.ftz.f32 %f5246, %f5237, %f5245; add.s32 %r9536, %r9514, -559038719; xor.b32 %r9537, %r9536, -1874868139; add.s32 %r9538, %r9537, 29184895; xor.b32 %r9539, %r9538, -559038720; shf.l.wrap.b32 %r9540, %r9538, %r9538, 25; sub.s32 %r9541, %r9539, %r9540; xor.b32 %r9542, %r9541, -1874868139; shf.l.wrap.b32 %r9543, %r9541, %r9541, 16; sub.s32 %r9544, %r9542, %r9543; xor.b32 %r9545, %r9544, %r9538; shf.l.wrap.b32 %r9546, %r9544, %r9544, 4; sub.s32 %r9547, %r9545, %r9546; xor.b32 %r9548, %r9547, %r9541; shf.l.wrap.b32 %r9549, %r9547, %r9547, 14; sub.s32 %r9550, %r9548, %r9549; xor.b32 %r9551, %r9550, %r9544; shf.l.wrap.b32 %r9552, %r9550, %r9550, 24; sub.s32 %r9553, %r9551, %r9552; add.ftz.f32 %f5247, %f5237, 0fBF800000; and.b32 %r9554, %r9553, 7; add.s32 %r9555, %r9554, 1; cvt.rn.f32.s32 %f5248, %r9555; and.b32 %r9556, %r9553, 8; setp.eq.s32 %p1895, %r9556, 0; neg.ftz.f32 %f5249, %f5248; selp.f32 %f5250, %f5248, %f5249, %p1895; mul.ftz.f32 %f5251, %f5247, %f5250; sub.ftz.f32 %f5252, %f5251, %f5246; fma.rn.ftz.f32 %f718, %f5242, %f5252, %f5246; mov.b32 %r458, %f718; setp.nan.ftz.f32 %p1896, %f718, %f718; mov.f32 %f5786, 0f00000000; @%p1896 bra $L__BB74_408; or.b32 %r9557, %r458, -2147483648; setp.ne.s32 %p1897, %r9557, -2147483648; add.ftz.f32 %f5254, %f718, %f718; setp.eq.ftz.f32 %p1898, %f5254, %f718; and.pred %p1899, %p1898, %p1897; @%p1899 bra $L__BB74_408; shl.b32 %r9558, %r458, 1; setp.lt.u32 %p1900, %r9558, -16777215; mul.ftz.f32 %f5255, %f718, 0f3E800000; selp.f32 %f5786, %f5255, 0f00000000, %p1900; $L__BB74_408: add.ftz.f32 %f5827, %f21, %f5786; mul.ftz.f32 %f5793, %f27, %f5815; cvt.rzi.ftz.s32.f32 %r459, %f26; setp.lt.s32 %p1901, %r459, 2; mov.f32 %f5791, %f717; @%p1901 bra $L__BB74_414; mov.u32 %r10033, 1; mov.f32 %f5791, %f717; $L__BB74_410: cvt.rmi.ftz.f32.f32 %f5257, %f5793; cvt.rzi.ftz.s32.f32 %r9560, %f5257; sub.ftz.f32 %f5258, %f5793, %f5257; mul.ftz.f32 %f5259, %f5258, %f5258; mul.ftz.f32 %f5260, %f5258, %f5259; fma.rn.ftz.f32 %f5261, %f5258, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5262, %f5258, %f5261, 0f41200000; mul.ftz.f32 %f5263, %f5260, %f5262; add.s32 %r9561, %r9560, -559038720; xor.b32 %r9562, %r9561, -1874868139; add.s32 %r9563, %r9562, 29184895; xor.b32 %r9564, %r9563, -559038720; shf.l.wrap.b32 %r9565, %r9563, %r9563, 25; sub.s32 %r9566, %r9564, %r9565; xor.b32 %r9567, %r9566, -1874868139; shf.l.wrap.b32 %r9568, %r9566, %r9566, 16; sub.s32 %r9569, %r9567, %r9568; xor.b32 %r9570, %r9569, %r9563; shf.l.wrap.b32 %r9571, %r9569, %r9569, 4; sub.s32 %r9572, %r9570, %r9571; xor.b32 %r9573, %r9572, %r9566; shf.l.wrap.b32 %r9574, %r9572, %r9572, 14; sub.s32 %r9575, %r9573, %r9574; xor.b32 %r9576, %r9575, %r9569; shf.l.wrap.b32 %r9577, %r9575, %r9575, 24; sub.s32 %r9578, %r9576, %r9577; and.b32 %r9579, %r9578, 7; add.s32 %r9580, %r9579, 1; cvt.rn.f32.s32 %f5264, %r9580; and.b32 %r9581, %r9578, 8; setp.eq.s32 %p1902, %r9581, 0; neg.ftz.f32 %f5265, %f5264; selp.f32 %f5266, %f5264, %f5265, %p1902; mul.ftz.f32 %f5267, %f5258, %f5266; add.s32 %r9582, %r9560, -559038719; xor.b32 %r9583, %r9582, -1874868139; add.s32 %r9584, %r9583, 29184895; xor.b32 %r9585, %r9584, -559038720; shf.l.wrap.b32 %r9586, %r9584, %r9584, 25; sub.s32 %r9587, %r9585, %r9586; xor.b32 %r9588, %r9587, -1874868139; shf.l.wrap.b32 %r9589, %r9587, %r9587, 16; sub.s32 %r9590, %r9588, %r9589; xor.b32 %r9591, %r9590, %r9584; shf.l.wrap.b32 %r9592, %r9590, %r9590, 4; sub.s32 %r9593, %r9591, %r9592; xor.b32 %r9594, %r9593, %r9587; shf.l.wrap.b32 %r9595, %r9593, %r9593, 14; sub.s32 %r9596, %r9594, %r9595; xor.b32 %r9597, %r9596, %r9590; shf.l.wrap.b32 %r9598, %r9596, %r9596, 24; sub.s32 %r9599, %r9597, %r9598; add.ftz.f32 %f5268, %f5258, 0fBF800000; and.b32 %r9600, %r9599, 7; add.s32 %r9601, %r9600, 1; cvt.rn.f32.s32 %f5269, %r9601; and.b32 %r9602, %r9599, 8; setp.eq.s32 %p1903, %r9602, 0; neg.ftz.f32 %f5270, %f5269; selp.f32 %f5271, %f5269, %f5270, %p1903; mul.ftz.f32 %f5272, %f5268, %f5271; sub.ftz.f32 %f5273, %f5272, %f5267; fma.rn.ftz.f32 %f726, %f5263, %f5273, %f5267; mov.b32 %r461, %f726; setp.nan.ftz.f32 %p1904, %f726, %f726; mov.f32 %f5790, 0f00000000; @%p1904 bra $L__BB74_413; or.b32 %r9603, %r461, -2147483648; setp.ne.s32 %p1905, %r9603, -2147483648; add.ftz.f32 %f5275, %f726, %f726; setp.eq.ftz.f32 %p1906, %f5275, %f726; and.pred %p1907, %p1906, %p1905; @%p1907 bra $L__BB74_413; shl.b32 %r9604, %r461, 1; setp.lt.u32 %p1908, %r9604, -16777215; selp.f32 %f5790, %f726, 0f00000000, %p1908; $L__BB74_413: fma.rn.ftz.f32 %f5276, %f5790, 0f3E800000, %f21; mul.ftz.f32 %f5277, %f5791, %f5276; fma.rn.ftz.f32 %f5827, %f5827, %f5277, %f5827; mul.ftz.f32 %f5791, %f717, %f5791; mul.ftz.f32 %f5793, %f27, %f5793; add.s32 %r10033, %r10033, 1; setp.lt.s32 %p1909, %r10033, %r459; @%p1909 bra $L__BB74_410; $L__BB74_414: cvt.rmi.ftz.f32.f32 %f5278, %f26; sub.ftz.f32 %f735, %f26, %f5278; setp.eq.ftz.f32 %p1910, %f735, 0f00000000; @%p1910 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f5280, %f5793; cvt.rzi.ftz.s32.f32 %r9605, %f5280; sub.ftz.f32 %f5281, %f5793, %f5280; mul.ftz.f32 %f5282, %f5281, %f5281; mul.ftz.f32 %f5283, %f5281, %f5282; fma.rn.ftz.f32 %f5284, %f5281, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5285, %f5281, %f5284, 0f41200000; mul.ftz.f32 %f5286, %f5283, %f5285; add.s32 %r9606, %r9605, -559038720; xor.b32 %r9607, %r9606, -1874868139; add.s32 %r9608, %r9607, 29184895; xor.b32 %r9609, %r9608, -559038720; shf.l.wrap.b32 %r9610, %r9608, %r9608, 25; sub.s32 %r9611, %r9609, %r9610; xor.b32 %r9612, %r9611, -1874868139; shf.l.wrap.b32 %r9613, %r9611, %r9611, 16; sub.s32 %r9614, %r9612, %r9613; xor.b32 %r9615, %r9614, %r9608; shf.l.wrap.b32 %r9616, %r9614, %r9614, 4; sub.s32 %r9617, %r9615, %r9616; xor.b32 %r9618, %r9617, %r9611; shf.l.wrap.b32 %r9619, %r9617, %r9617, 14; sub.s32 %r9620, %r9618, %r9619; xor.b32 %r9621, %r9620, %r9614; shf.l.wrap.b32 %r9622, %r9620, %r9620, 24; sub.s32 %r9623, %r9621, %r9622; and.b32 %r9624, %r9623, 7; add.s32 %r9625, %r9624, 1; cvt.rn.f32.s32 %f5287, %r9625; and.b32 %r9626, %r9623, 8; setp.eq.s32 %p1911, %r9626, 0; neg.ftz.f32 %f5288, %f5287; selp.f32 %f5289, %f5287, %f5288, %p1911; mul.ftz.f32 %f5290, %f5281, %f5289; add.s32 %r9627, %r9605, -559038719; xor.b32 %r9628, %r9627, -1874868139; add.s32 %r9629, %r9628, 29184895; xor.b32 %r9630, %r9629, -559038720; shf.l.wrap.b32 %r9631, %r9629, %r9629, 25; sub.s32 %r9632, %r9630, %r9631; xor.b32 %r9633, %r9632, -1874868139; shf.l.wrap.b32 %r9634, %r9632, %r9632, 16; sub.s32 %r9635, %r9633, %r9634; xor.b32 %r9636, %r9635, %r9629; shf.l.wrap.b32 %r9637, %r9635, %r9635, 4; sub.s32 %r9638, %r9636, %r9637; xor.b32 %r9639, %r9638, %r9632; shf.l.wrap.b32 %r9640, %r9638, %r9638, 14; sub.s32 %r9641, %r9639, %r9640; xor.b32 %r9642, %r9641, %r9635; shf.l.wrap.b32 %r9643, %r9641, %r9641, 24; sub.s32 %r9644, %r9642, %r9643; add.ftz.f32 %f5291, %f5281, 0fBF800000; and.b32 %r9645, %r9644, 7; add.s32 %r9646, %r9645, 1; cvt.rn.f32.s32 %f5292, %r9646; and.b32 %r9647, %r9644, 8; setp.eq.s32 %p1912, %r9647, 0; neg.ftz.f32 %f5293, %f5292; selp.f32 %f5294, %f5292, %f5293, %p1912; mul.ftz.f32 %f5295, %f5291, %f5294; sub.ftz.f32 %f5296, %f5295, %f5290; fma.rn.ftz.f32 %f736, %f5286, %f5296, %f5290; mov.b32 %r463, %f736; setp.nan.ftz.f32 %p1913, %f736, %f736; mov.f32 %f5794, 0f00000000; @%p1913 bra $L__BB74_418; or.b32 %r9648, %r463, -2147483648; setp.ne.s32 %p1914, %r9648, -2147483648; add.ftz.f32 %f5298, %f736, %f736; setp.eq.ftz.f32 %p1915, %f5298, %f736; and.pred %p1916, %p1915, %p1914; @%p1916 bra $L__BB74_418; shl.b32 %r9649, %r463, 1; setp.lt.u32 %p1917, %r9649, -16777215; mul.ftz.f32 %f5299, %f736, 0f3E800000; selp.f32 %f5794, %f5299, 0f00000000, %p1917; $L__BB74_418: add.ftz.f32 %f5300, %f21, %f5794; mul.ftz.f32 %f5301, %f5791, %f5300; mul.ftz.f32 %f5302, %f5827, %f5301; fma.rn.ftz.f32 %f5827, %f735, %f5302, %f5827; bra.uni $L__BB74_450; $L__BB74_102: mul.ftz.f32 %f5694, %f1, %f9; mul.ftz.f32 %f5695, %f2, %f9; mul.ftz.f32 %f5696, %f3, %f9; and.b16 %rs5, %rs2, 255; setp.gt.s16 %p1154, %rs5, 1; @%p1154 bra $L__BB74_148; bra.uni $L__BB74_103; $L__BB74_148: setp.eq.s16 %p1155, %rs5, 2; @%p1155 bra $L__BB74_254; setp.eq.s16 %p1156, %rs5, 3; @%p1156 bra $L__BB74_213; setp.ne.s16 %p1157, %rs5, 4; @%p1157 bra $L__BB74_450; lg2.approx.ftz.f32 %f3567, %f27; mul.ftz.f32 %f3568, %f25, %f3567; neg.ftz.f32 %f3569, %f3568; ex2.approx.ftz.f32 %f177, %f3569; cvt.rmi.ftz.f32.f32 %f3570, %f5694; cvt.rzi.ftz.s32.f32 %r42, %f3570; sub.ftz.f32 %f178, %f5694, %f3570; cvt.rmi.ftz.f32.f32 %f3571, %f5695; cvt.rzi.ftz.s32.f32 %r43, %f3571; sub.ftz.f32 %f179, %f5695, %f3571; cvt.rmi.ftz.f32.f32 %f3572, %f5696; cvt.rzi.ftz.s32.f32 %r44, %f3572; sub.ftz.f32 %f180, %f5696, %f3572; fma.rn.ftz.f32 %f3573, %f178, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f181, %f178, %f3573, 0f41200000; fma.rn.ftz.f32 %f3574, %f179, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f182, %f179, %f3574, 0f41200000; fma.rn.ftz.f32 %f3575, %f180, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f183, %f180, %f3575, 0f41200000; add.s32 %r6728, %r43, -559038712; add.s32 %r6729, %r44, -559038712; xor.b32 %r6730, %r6729, %r6728; shf.l.wrap.b32 %r45, %r6728, %r6728, 14; sub.s32 %r46, %r6730, %r45; add.s32 %r6731, %r42, -559038712; xor.b32 %r6732, %r46, %r6731; shf.l.wrap.b32 %r47, %r46, %r46, 11; sub.s32 %r6733, %r6732, %r47; xor.b32 %r6734, %r6733, %r6728; shf.l.wrap.b32 %r6735, %r6733, %r6733, 25; sub.s32 %r6736, %r6734, %r6735; xor.b32 %r6737, %r6736, %r46; shf.l.wrap.b32 %r6738, %r6736, %r6736, 16; sub.s32 %r6739, %r6737, %r6738; xor.b32 %r6740, %r6739, %r6733; shf.l.wrap.b32 %r6741, %r6739, %r6739, 4; sub.s32 %r6742, %r6740, %r6741; xor.b32 %r6743, %r6742, %r6736; shf.l.wrap.b32 %r6744, %r6742, %r6742, 14; sub.s32 %r6745, %r6743, %r6744; xor.b32 %r6746, %r6745, %r6739; shf.l.wrap.b32 %r6747, %r6745, %r6745, 24; sub.s32 %r48, %r6746, %r6747; and.b32 %r49, %r48, 15; cvt.u16.u32 %rs7, %r48; or.b16 %rs8, %rs7, 2; and.b16 %rs9, %rs8, 15; setp.eq.s16 %p1160, %rs9, 14; mov.f32 %f5589, %f178; @%p1160 bra $L__BB74_153; mov.f32 %f5589, %f180; $L__BB74_153: mul.ftz.f32 %f3576, %f178, %f178; mul.ftz.f32 %f3577, %f178, %f3576; mul.ftz.f32 %f185, %f3577, %f181; mul.ftz.f32 %f3578, %f179, %f179; mul.ftz.f32 %f3579, %f179, %f3578; mul.ftz.f32 %f186, %f3579, %f182; mul.ftz.f32 %f3580, %f180, %f180; mul.ftz.f32 %f3581, %f180, %f3580; mul.ftz.f32 %f187, %f3581, %f183; setp.lt.u32 %p1161, %r49, 8; selp.f32 %f3582, %f178, %f179, %p1161; setp.lt.u32 %p1162, %r49, 4; selp.f32 %f3583, %f179, %f5589, %p1162; and.b32 %r6748, %r48, 1; setp.eq.b32 %p1163, %r6748, 1; neg.ftz.f32 %f3584, %f3582; selp.f32 %f3585, %f3584, %f3582, %p1163; and.b32 %r6749, %r48, 2; setp.eq.s32 %p1164, %r6749, 0; neg.ftz.f32 %f3586, %f3583; selp.f32 %f3587, %f3583, %f3586, %p1164; add.ftz.f32 %f188, %f3585, %f3587; add.s32 %r50, %r42, -559038711; xor.b32 %r6750, %r46, %r50; sub.s32 %r6751, %r6750, %r47; xor.b32 %r6753, %r6751, %r6728; shf.l.wrap.b32 %r6754, %r6751, %r6751, 25; sub.s32 %r6755, %r6753, %r6754; xor.b32 %r6756, %r6755, %r46; shf.l.wrap.b32 %r6757, %r6755, %r6755, 16; sub.s32 %r6758, %r6756, %r6757; xor.b32 %r6759, %r6758, %r6751; shf.l.wrap.b32 %r6760, %r6758, %r6758, 4; sub.s32 %r6761, %r6759, %r6760; xor.b32 %r6762, %r6761, %r6755; shf.l.wrap.b32 %r6763, %r6761, %r6761, 14; sub.s32 %r6764, %r6762, %r6763; xor.b32 %r6765, %r6764, %r6758; shf.l.wrap.b32 %r6766, %r6764, %r6764, 24; sub.s32 %r51, %r6765, %r6766; and.b32 %r52, %r51, 15; add.ftz.f32 %f189, %f178, 0fBF800000; cvt.u16.u32 %rs10, %r51; or.b16 %rs11, %rs10, 2; and.b16 %rs12, %rs11, 15; setp.eq.s16 %p1165, %rs12, 14; mov.f32 %f5590, %f189; @%p1165 bra $L__BB74_155; mov.f32 %f5590, %f180; $L__BB74_155: setp.lt.u32 %p1166, %r52, 8; selp.f32 %f3588, %f189, %f179, %p1166; setp.lt.u32 %p1167, %r52, 4; selp.f32 %f3589, %f179, %f5590, %p1167; and.b32 %r6767, %r51, 1; setp.eq.b32 %p1168, %r6767, 1; neg.ftz.f32 %f3590, %f3588; selp.f32 %f3591, %f3590, %f3588, %p1168; and.b32 %r6768, %r51, 2; setp.eq.s32 %p1169, %r6768, 0; neg.ftz.f32 %f3592, %f3589; selp.f32 %f3593, %f3589, %f3592, %p1169; add.ftz.f32 %f191, %f3591, %f3593; add.s32 %r53, %r43, -559038711; xor.b32 %r6770, %r6729, %r53; shf.l.wrap.b32 %r54, %r53, %r53, 14; sub.s32 %r55, %r6770, %r54; xor.b32 %r6772, %r55, %r6731; shf.l.wrap.b32 %r56, %r55, %r55, 11; sub.s32 %r6773, %r6772, %r56; xor.b32 %r6774, %r6773, %r53; shf.l.wrap.b32 %r6775, %r6773, %r6773, 25; sub.s32 %r6776, %r6774, %r6775; xor.b32 %r6777, %r6776, %r55; shf.l.wrap.b32 %r6778, %r6776, %r6776, 16; sub.s32 %r6779, %r6777, %r6778; xor.b32 %r6780, %r6779, %r6773; shf.l.wrap.b32 %r6781, %r6779, %r6779, 4; sub.s32 %r6782, %r6780, %r6781; xor.b32 %r6783, %r6782, %r6776; shf.l.wrap.b32 %r6784, %r6782, %r6782, 14; sub.s32 %r6785, %r6783, %r6784; xor.b32 %r6786, %r6785, %r6779; shf.l.wrap.b32 %r6787, %r6785, %r6785, 24; sub.s32 %r57, %r6786, %r6787; and.b32 %r58, %r57, 15; add.ftz.f32 %f192, %f179, 0fBF800000; cvt.u16.u32 %rs13, %r57; or.b16 %rs14, %rs13, 2; and.b16 %rs15, %rs14, 15; setp.eq.s16 %p1170, %rs15, 14; mov.f32 %f5591, %f178; @%p1170 bra $L__BB74_157; mov.f32 %f5591, %f180; $L__BB74_157: setp.lt.u32 %p1171, %r58, 8; selp.f32 %f3594, %f178, %f192, %p1171; setp.lt.u32 %p1172, %r58, 4; selp.f32 %f3595, %f192, %f5591, %p1172; and.b32 %r6788, %r57, 1; setp.eq.b32 %p1173, %r6788, 1; neg.ftz.f32 %f3596, %f3594; selp.f32 %f3597, %f3596, %f3594, %p1173; and.b32 %r6789, %r57, 2; setp.eq.s32 %p1174, %r6789, 0; neg.ftz.f32 %f3598, %f3595; selp.f32 %f3599, %f3595, %f3598, %p1174; add.ftz.f32 %f194, %f3597, %f3599; xor.b32 %r6790, %r55, %r50; sub.s32 %r6791, %r6790, %r56; xor.b32 %r6792, %r6791, %r53; shf.l.wrap.b32 %r6793, %r6791, %r6791, 25; sub.s32 %r6794, %r6792, %r6793; xor.b32 %r6795, %r6794, %r55; shf.l.wrap.b32 %r6796, %r6794, %r6794, 16; sub.s32 %r6797, %r6795, %r6796; xor.b32 %r6798, %r6797, %r6791; shf.l.wrap.b32 %r6799, %r6797, %r6797, 4; sub.s32 %r6800, %r6798, %r6799; xor.b32 %r6801, %r6800, %r6794; shf.l.wrap.b32 %r6802, %r6800, %r6800, 14; sub.s32 %r6803, %r6801, %r6802; xor.b32 %r6804, %r6803, %r6797; shf.l.wrap.b32 %r6805, %r6803, %r6803, 24; sub.s32 %r59, %r6804, %r6805; and.b32 %r60, %r59, 15; cvt.u16.u32 %rs16, %r59; or.b16 %rs17, %rs16, 2; and.b16 %rs18, %rs17, 15; setp.eq.s16 %p1175, %rs18, 14; mov.f32 %f5592, %f189; @%p1175 bra $L__BB74_159; mov.f32 %f5592, %f180; $L__BB74_159: setp.lt.u32 %p1176, %r60, 8; selp.f32 %f3600, %f189, %f192, %p1176; setp.lt.u32 %p1177, %r60, 4; selp.f32 %f3601, %f192, %f5592, %p1177; and.b32 %r6806, %r59, 1; setp.eq.b32 %p1178, %r6806, 1; neg.ftz.f32 %f3602, %f3600; selp.f32 %f3603, %f3602, %f3600, %p1178; and.b32 %r6807, %r59, 2; setp.eq.s32 %p1179, %r6807, 0; neg.ftz.f32 %f3604, %f3601; selp.f32 %f3605, %f3601, %f3604, %p1179; add.ftz.f32 %f196, %f3603, %f3605; add.s32 %r61, %r44, -559038711; xor.b32 %r6809, %r61, %r6728; sub.s32 %r62, %r6809, %r45; xor.b32 %r6811, %r62, %r6731; shf.l.wrap.b32 %r63, %r62, %r62, 11; sub.s32 %r6812, %r6811, %r63; xor.b32 %r6813, %r6812, %r6728; shf.l.wrap.b32 %r6814, %r6812, %r6812, 25; sub.s32 %r6815, %r6813, %r6814; xor.b32 %r6816, %r6815, %r62; shf.l.wrap.b32 %r6817, %r6815, %r6815, 16; sub.s32 %r6818, %r6816, %r6817; xor.b32 %r6819, %r6818, %r6812; shf.l.wrap.b32 %r6820, %r6818, %r6818, 4; sub.s32 %r6821, %r6819, %r6820; xor.b32 %r6822, %r6821, %r6815; shf.l.wrap.b32 %r6823, %r6821, %r6821, 14; sub.s32 %r6824, %r6822, %r6823; xor.b32 %r6825, %r6824, %r6818; shf.l.wrap.b32 %r6826, %r6824, %r6824, 24; sub.s32 %r64, %r6825, %r6826; add.ftz.f32 %f197, %f180, 0fBF800000; and.b32 %r65, %r64, 15; cvt.u16.u32 %rs19, %r64; or.b16 %rs20, %rs19, 2; and.b16 %rs21, %rs20, 15; setp.eq.s16 %p1180, %rs21, 14; mov.f32 %f5593, %f178; @%p1180 bra $L__BB74_161; mov.f32 %f5593, %f197; $L__BB74_161: setp.lt.u32 %p1181, %r65, 8; selp.f32 %f3606, %f178, %f179, %p1181; setp.lt.u32 %p1182, %r65, 4; selp.f32 %f3607, %f179, %f5593, %p1182; and.b32 %r6827, %r64, 1; setp.eq.b32 %p1183, %r6827, 1; neg.ftz.f32 %f3608, %f3606; selp.f32 %f3609, %f3608, %f3606, %p1183; and.b32 %r6828, %r64, 2; setp.eq.s32 %p1184, %r6828, 0; neg.ftz.f32 %f3610, %f3607; selp.f32 %f3611, %f3607, %f3610, %p1184; add.ftz.f32 %f199, %f3609, %f3611; xor.b32 %r6829, %r62, %r50; sub.s32 %r6830, %r6829, %r63; xor.b32 %r6832, %r6830, %r6728; shf.l.wrap.b32 %r6833, %r6830, %r6830, 25; sub.s32 %r6834, %r6832, %r6833; xor.b32 %r6835, %r6834, %r62; shf.l.wrap.b32 %r6836, %r6834, %r6834, 16; sub.s32 %r6837, %r6835, %r6836; xor.b32 %r6838, %r6837, %r6830; shf.l.wrap.b32 %r6839, %r6837, %r6837, 4; sub.s32 %r6840, %r6838, %r6839; xor.b32 %r6841, %r6840, %r6834; shf.l.wrap.b32 %r6842, %r6840, %r6840, 14; sub.s32 %r6843, %r6841, %r6842; xor.b32 %r6844, %r6843, %r6837; shf.l.wrap.b32 %r6845, %r6843, %r6843, 24; sub.s32 %r66, %r6844, %r6845; and.b32 %r67, %r66, 15; cvt.u16.u32 %rs22, %r66; or.b16 %rs23, %rs22, 2; and.b16 %rs24, %rs23, 15; setp.eq.s16 %p1185, %rs24, 14; mov.f32 %f5594, %f189; @%p1185 bra $L__BB74_163; mov.f32 %f5594, %f197; $L__BB74_163: setp.lt.u32 %p1186, %r67, 8; selp.f32 %f3612, %f189, %f179, %p1186; setp.lt.u32 %p1187, %r67, 4; selp.f32 %f3613, %f179, %f5594, %p1187; and.b32 %r6846, %r66, 1; setp.eq.b32 %p1188, %r6846, 1; neg.ftz.f32 %f3614, %f3612; selp.f32 %f3615, %f3614, %f3612, %p1188; and.b32 %r6847, %r66, 2; setp.eq.s32 %p1189, %r6847, 0; neg.ftz.f32 %f3616, %f3613; selp.f32 %f3617, %f3613, %f3616, %p1189; add.ftz.f32 %f201, %f3615, %f3617; xor.b32 %r6848, %r61, %r53; sub.s32 %r68, %r6848, %r54; xor.b32 %r6850, %r68, %r6731; shf.l.wrap.b32 %r69, %r68, %r68, 11; sub.s32 %r6851, %r6850, %r69; xor.b32 %r6852, %r6851, %r53; shf.l.wrap.b32 %r6853, %r6851, %r6851, 25; sub.s32 %r6854, %r6852, %r6853; xor.b32 %r6855, %r6854, %r68; shf.l.wrap.b32 %r6856, %r6854, %r6854, 16; sub.s32 %r6857, %r6855, %r6856; xor.b32 %r6858, %r6857, %r6851; shf.l.wrap.b32 %r6859, %r6857, %r6857, 4; sub.s32 %r6860, %r6858, %r6859; xor.b32 %r6861, %r6860, %r6854; shf.l.wrap.b32 %r6862, %r6860, %r6860, 14; sub.s32 %r6863, %r6861, %r6862; xor.b32 %r6864, %r6863, %r6857; shf.l.wrap.b32 %r6865, %r6863, %r6863, 24; sub.s32 %r70, %r6864, %r6865; and.b32 %r71, %r70, 15; cvt.u16.u32 %rs25, %r70; or.b16 %rs26, %rs25, 2; and.b16 %rs27, %rs26, 15; setp.eq.s16 %p1190, %rs27, 14; mov.f32 %f5595, %f178; @%p1190 bra $L__BB74_165; mov.f32 %f5595, %f197; $L__BB74_165: setp.lt.u32 %p1191, %r71, 8; selp.f32 %f3618, %f178, %f192, %p1191; setp.lt.u32 %p1192, %r71, 4; selp.f32 %f3619, %f192, %f5595, %p1192; and.b32 %r6866, %r70, 1; setp.eq.b32 %p1193, %r6866, 1; neg.ftz.f32 %f3620, %f3618; selp.f32 %f3621, %f3620, %f3618, %p1193; and.b32 %r6867, %r70, 2; setp.eq.s32 %p1194, %r6867, 0; neg.ftz.f32 %f3622, %f3619; selp.f32 %f3623, %f3619, %f3622, %p1194; add.ftz.f32 %f203, %f3621, %f3623; xor.b32 %r6868, %r68, %r50; sub.s32 %r6869, %r6868, %r69; xor.b32 %r6870, %r6869, %r53; shf.l.wrap.b32 %r6871, %r6869, %r6869, 25; sub.s32 %r6872, %r6870, %r6871; xor.b32 %r6873, %r6872, %r68; shf.l.wrap.b32 %r6874, %r6872, %r6872, 16; sub.s32 %r6875, %r6873, %r6874; xor.b32 %r6876, %r6875, %r6869; shf.l.wrap.b32 %r6877, %r6875, %r6875, 4; sub.s32 %r6878, %r6876, %r6877; xor.b32 %r6879, %r6878, %r6872; shf.l.wrap.b32 %r6880, %r6878, %r6878, 14; sub.s32 %r6881, %r6879, %r6880; xor.b32 %r6882, %r6881, %r6875; shf.l.wrap.b32 %r6883, %r6881, %r6881, 24; sub.s32 %r72, %r6882, %r6883; and.b32 %r73, %r72, 15; cvt.u16.u32 %rs28, %r72; or.b16 %rs29, %rs28, 2; and.b16 %rs30, %rs29, 15; setp.eq.s16 %p1195, %rs30, 14; mov.f32 %f5596, %f189; @%p1195 bra $L__BB74_167; mov.f32 %f5596, %f197; $L__BB74_167: setp.lt.u32 %p1196, %r73, 8; selp.f32 %f3625, %f189, %f192, %p1196; setp.lt.u32 %p1197, %r73, 4; selp.f32 %f3626, %f192, %f5596, %p1197; and.b32 %r6884, %r72, 1; setp.eq.b32 %p1198, %r6884, 1; neg.ftz.f32 %f3627, %f3625; selp.f32 %f3628, %f3627, %f3625, %p1198; and.b32 %r6885, %r72, 2; setp.eq.s32 %p1199, %r6885, 0; neg.ftz.f32 %f3629, %f3626; selp.f32 %f3630, %f3626, %f3629, %p1199; add.ftz.f32 %f3631, %f3628, %f3630; mov.f32 %f3632, 0f3F800000; sub.ftz.f32 %f3633, %f3632, %f185; mul.ftz.f32 %f3634, %f185, %f191; fma.rn.ftz.f32 %f3635, %f3633, %f188, %f3634; sub.ftz.f32 %f3636, %f3632, %f186; mul.ftz.f32 %f3637, %f185, %f196; fma.rn.ftz.f32 %f3638, %f3633, %f194, %f3637; mul.ftz.f32 %f3639, %f186, %f3638; fma.rn.ftz.f32 %f3640, %f3636, %f3635, %f3639; sub.ftz.f32 %f3641, %f3632, %f187; mul.ftz.f32 %f3642, %f185, %f201; fma.rn.ftz.f32 %f3643, %f3633, %f199, %f3642; mul.ftz.f32 %f3644, %f185, %f3631; fma.rn.ftz.f32 %f3645, %f3633, %f203, %f3644; mul.ftz.f32 %f3646, %f186, %f3645; fma.rn.ftz.f32 %f3647, %f3636, %f3643, %f3646; mul.ftz.f32 %f3648, %f187, %f3647; fma.rn.ftz.f32 %f205, %f3641, %f3640, %f3648; mov.b32 %r74, %f205; setp.nan.ftz.f32 %p1200, %f205, %f205; mov.f32 %f5597, 0f00000000; @%p1200 bra $L__BB74_170; or.b32 %r6886, %r74, -2147483648; setp.ne.s32 %p1201, %r6886, -2147483648; add.ftz.f32 %f3650, %f205, %f205; setp.eq.ftz.f32 %p1202, %f3650, %f205; and.pred %p1203, %p1202, %p1201; @%p1203 bra $L__BB74_170; shl.b32 %r6887, %r74, 1; setp.lt.u32 %p1204, %r6887, -16777215; mul.ftz.f32 %f3651, %f205, 0f3F7B645A; selp.f32 %f5597, %f3651, 0f00000000, %p1204; $L__BB74_170: add.ftz.f32 %f5827, %f21, %f5597; mul.ftz.f32 %f5612, %f5694, %f27; mul.ftz.f32 %f5613, %f5695, %f27; mul.ftz.f32 %f5614, %f5696, %f27; cvt.rzi.ftz.s32.f32 %r75, %f26; setp.lt.s32 %p1205, %r75, 2; mov.f32 %f5615, %f177; @%p1205 bra $L__BB74_192; mov.u32 %r10023, 1; mov.f32 %f5615, %f177; $L__BB74_172: cvt.rmi.ftz.f32.f32 %f3652, %f5612; cvt.rzi.ftz.s32.f32 %r77, %f3652; sub.ftz.f32 %f217, %f5612, %f3652; cvt.rmi.ftz.f32.f32 %f3653, %f5613; cvt.rzi.ftz.s32.f32 %r78, %f3653; sub.ftz.f32 %f218, %f5613, %f3653; cvt.rmi.ftz.f32.f32 %f3654, %f5614; cvt.rzi.ftz.s32.f32 %r79, %f3654; sub.ftz.f32 %f219, %f5614, %f3654; fma.rn.ftz.f32 %f3655, %f217, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f220, %f217, %f3655, 0f41200000; fma.rn.ftz.f32 %f3656, %f218, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f221, %f218, %f3656, 0f41200000; fma.rn.ftz.f32 %f3657, %f219, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f222, %f219, %f3657, 0f41200000; add.s32 %r6889, %r78, -559038712; add.s32 %r6890, %r79, -559038712; xor.b32 %r6891, %r6890, %r6889; shf.l.wrap.b32 %r80, %r6889, %r6889, 14; sub.s32 %r81, %r6891, %r80; add.s32 %r6892, %r77, -559038712; xor.b32 %r6893, %r81, %r6892; shf.l.wrap.b32 %r82, %r81, %r81, 11; sub.s32 %r6894, %r6893, %r82; xor.b32 %r6895, %r6894, %r6889; shf.l.wrap.b32 %r6896, %r6894, %r6894, 25; sub.s32 %r6897, %r6895, %r6896; xor.b32 %r6898, %r6897, %r81; shf.l.wrap.b32 %r6899, %r6897, %r6897, 16; sub.s32 %r6900, %r6898, %r6899; xor.b32 %r6901, %r6900, %r6894; shf.l.wrap.b32 %r6902, %r6900, %r6900, 4; sub.s32 %r6903, %r6901, %r6902; xor.b32 %r6904, %r6903, %r6897; shf.l.wrap.b32 %r6905, %r6903, %r6903, 14; sub.s32 %r6906, %r6904, %r6905; xor.b32 %r6907, %r6906, %r6900; shf.l.wrap.b32 %r6908, %r6906, %r6906, 24; sub.s32 %r83, %r6907, %r6908; and.b32 %r84, %r83, 15; cvt.u16.u32 %rs31, %r83; or.b16 %rs32, %rs31, 2; and.b16 %rs33, %rs32, 15; setp.eq.s16 %p1206, %rs33, 14; mov.f32 %f5603, %f217; @%p1206 bra $L__BB74_174; mov.f32 %f5603, %f219; $L__BB74_174: mul.ftz.f32 %f3658, %f217, %f217; mul.ftz.f32 %f3659, %f217, %f3658; mul.ftz.f32 %f224, %f3659, %f220; mul.ftz.f32 %f3660, %f218, %f218; mul.ftz.f32 %f3661, %f218, %f3660; mul.ftz.f32 %f225, %f3661, %f221; mul.ftz.f32 %f3662, %f219, %f219; mul.ftz.f32 %f3663, %f219, %f3662; mul.ftz.f32 %f226, %f3663, %f222; setp.lt.u32 %p1207, %r84, 8; selp.f32 %f3664, %f217, %f218, %p1207; setp.lt.u32 %p1208, %r84, 4; selp.f32 %f3665, %f218, %f5603, %p1208; and.b32 %r6909, %r83, 1; setp.eq.b32 %p1209, %r6909, 1; neg.ftz.f32 %f3666, %f3664; selp.f32 %f3667, %f3666, %f3664, %p1209; and.b32 %r6910, %r83, 2; setp.eq.s32 %p1210, %r6910, 0; neg.ftz.f32 %f3668, %f3665; selp.f32 %f3669, %f3665, %f3668, %p1210; add.ftz.f32 %f227, %f3667, %f3669; add.s32 %r85, %r77, -559038711; xor.b32 %r6911, %r81, %r85; sub.s32 %r6912, %r6911, %r82; xor.b32 %r6914, %r6912, %r6889; shf.l.wrap.b32 %r6915, %r6912, %r6912, 25; sub.s32 %r6916, %r6914, %r6915; xor.b32 %r6917, %r6916, %r81; shf.l.wrap.b32 %r6918, %r6916, %r6916, 16; sub.s32 %r6919, %r6917, %r6918; xor.b32 %r6920, %r6919, %r6912; shf.l.wrap.b32 %r6921, %r6919, %r6919, 4; sub.s32 %r6922, %r6920, %r6921; xor.b32 %r6923, %r6922, %r6916; shf.l.wrap.b32 %r6924, %r6922, %r6922, 14; sub.s32 %r6925, %r6923, %r6924; xor.b32 %r6926, %r6925, %r6919; shf.l.wrap.b32 %r6927, %r6925, %r6925, 24; sub.s32 %r86, %r6926, %r6927; and.b32 %r87, %r86, 15; add.ftz.f32 %f228, %f217, 0fBF800000; cvt.u16.u32 %rs34, %r86; or.b16 %rs35, %rs34, 2; and.b16 %rs36, %rs35, 15; setp.eq.s16 %p1211, %rs36, 14; mov.f32 %f5604, %f228; @%p1211 bra $L__BB74_176; mov.f32 %f5604, %f219; $L__BB74_176: setp.lt.u32 %p1212, %r87, 8; selp.f32 %f3670, %f228, %f218, %p1212; setp.lt.u32 %p1213, %r87, 4; selp.f32 %f3671, %f218, %f5604, %p1213; and.b32 %r6928, %r86, 1; setp.eq.b32 %p1214, %r6928, 1; neg.ftz.f32 %f3672, %f3670; selp.f32 %f3673, %f3672, %f3670, %p1214; and.b32 %r6929, %r86, 2; setp.eq.s32 %p1215, %r6929, 0; neg.ftz.f32 %f3674, %f3671; selp.f32 %f3675, %f3671, %f3674, %p1215; add.ftz.f32 %f230, %f3673, %f3675; add.s32 %r88, %r78, -559038711; xor.b32 %r6931, %r6890, %r88; shf.l.wrap.b32 %r89, %r88, %r88, 14; sub.s32 %r90, %r6931, %r89; xor.b32 %r6933, %r90, %r6892; shf.l.wrap.b32 %r91, %r90, %r90, 11; sub.s32 %r6934, %r6933, %r91; xor.b32 %r6935, %r6934, %r88; shf.l.wrap.b32 %r6936, %r6934, %r6934, 25; sub.s32 %r6937, %r6935, %r6936; xor.b32 %r6938, %r6937, %r90; shf.l.wrap.b32 %r6939, %r6937, %r6937, 16; sub.s32 %r6940, %r6938, %r6939; xor.b32 %r6941, %r6940, %r6934; shf.l.wrap.b32 %r6942, %r6940, %r6940, 4; sub.s32 %r6943, %r6941, %r6942; xor.b32 %r6944, %r6943, %r6937; shf.l.wrap.b32 %r6945, %r6943, %r6943, 14; sub.s32 %r6946, %r6944, %r6945; xor.b32 %r6947, %r6946, %r6940; shf.l.wrap.b32 %r6948, %r6946, %r6946, 24; sub.s32 %r92, %r6947, %r6948; and.b32 %r93, %r92, 15; add.ftz.f32 %f231, %f218, 0fBF800000; cvt.u16.u32 %rs37, %r92; or.b16 %rs38, %rs37, 2; and.b16 %rs39, %rs38, 15; setp.eq.s16 %p1216, %rs39, 14; mov.f32 %f5605, %f217; @%p1216 bra $L__BB74_178; mov.f32 %f5605, %f219; $L__BB74_178: setp.lt.u32 %p1217, %r93, 8; selp.f32 %f3676, %f217, %f231, %p1217; setp.lt.u32 %p1218, %r93, 4; selp.f32 %f3677, %f231, %f5605, %p1218; and.b32 %r6949, %r92, 1; setp.eq.b32 %p1219, %r6949, 1; neg.ftz.f32 %f3678, %f3676; selp.f32 %f3679, %f3678, %f3676, %p1219; and.b32 %r6950, %r92, 2; setp.eq.s32 %p1220, %r6950, 0; neg.ftz.f32 %f3680, %f3677; selp.f32 %f3681, %f3677, %f3680, %p1220; add.ftz.f32 %f233, %f3679, %f3681; xor.b32 %r6951, %r90, %r85; sub.s32 %r6952, %r6951, %r91; xor.b32 %r6953, %r6952, %r88; shf.l.wrap.b32 %r6954, %r6952, %r6952, 25; sub.s32 %r6955, %r6953, %r6954; xor.b32 %r6956, %r6955, %r90; shf.l.wrap.b32 %r6957, %r6955, %r6955, 16; sub.s32 %r6958, %r6956, %r6957; xor.b32 %r6959, %r6958, %r6952; shf.l.wrap.b32 %r6960, %r6958, %r6958, 4; sub.s32 %r6961, %r6959, %r6960; xor.b32 %r6962, %r6961, %r6955; shf.l.wrap.b32 %r6963, %r6961, %r6961, 14; sub.s32 %r6964, %r6962, %r6963; xor.b32 %r6965, %r6964, %r6958; shf.l.wrap.b32 %r6966, %r6964, %r6964, 24; sub.s32 %r94, %r6965, %r6966; and.b32 %r95, %r94, 15; cvt.u16.u32 %rs40, %r94; or.b16 %rs41, %rs40, 2; and.b16 %rs42, %rs41, 15; setp.eq.s16 %p1221, %rs42, 14; mov.f32 %f5606, %f228; @%p1221 bra $L__BB74_180; mov.f32 %f5606, %f219; $L__BB74_180: setp.lt.u32 %p1222, %r95, 8; selp.f32 %f3682, %f228, %f231, %p1222; setp.lt.u32 %p1223, %r95, 4; selp.f32 %f3683, %f231, %f5606, %p1223; and.b32 %r6967, %r94, 1; setp.eq.b32 %p1224, %r6967, 1; neg.ftz.f32 %f3684, %f3682; selp.f32 %f3685, %f3684, %f3682, %p1224; and.b32 %r6968, %r94, 2; setp.eq.s32 %p1225, %r6968, 0; neg.ftz.f32 %f3686, %f3683; selp.f32 %f3687, %f3683, %f3686, %p1225; add.ftz.f32 %f235, %f3685, %f3687; add.s32 %r96, %r79, -559038711; xor.b32 %r6970, %r96, %r6889; sub.s32 %r97, %r6970, %r80; xor.b32 %r6972, %r97, %r6892; shf.l.wrap.b32 %r98, %r97, %r97, 11; sub.s32 %r6973, %r6972, %r98; xor.b32 %r6974, %r6973, %r6889; shf.l.wrap.b32 %r6975, %r6973, %r6973, 25; sub.s32 %r6976, %r6974, %r6975; xor.b32 %r6977, %r6976, %r97; shf.l.wrap.b32 %r6978, %r6976, %r6976, 16; sub.s32 %r6979, %r6977, %r6978; xor.b32 %r6980, %r6979, %r6973; shf.l.wrap.b32 %r6981, %r6979, %r6979, 4; sub.s32 %r6982, %r6980, %r6981; xor.b32 %r6983, %r6982, %r6976; shf.l.wrap.b32 %r6984, %r6982, %r6982, 14; sub.s32 %r6985, %r6983, %r6984; xor.b32 %r6986, %r6985, %r6979; shf.l.wrap.b32 %r6987, %r6985, %r6985, 24; sub.s32 %r99, %r6986, %r6987; add.ftz.f32 %f236, %f219, 0fBF800000; and.b32 %r100, %r99, 15; cvt.u16.u32 %rs43, %r99; or.b16 %rs44, %rs43, 2; and.b16 %rs45, %rs44, 15; setp.eq.s16 %p1226, %rs45, 14; mov.f32 %f5607, %f217; @%p1226 bra $L__BB74_182; mov.f32 %f5607, %f236; $L__BB74_182: setp.lt.u32 %p1227, %r100, 8; selp.f32 %f3688, %f217, %f218, %p1227; setp.lt.u32 %p1228, %r100, 4; selp.f32 %f3689, %f218, %f5607, %p1228; and.b32 %r6988, %r99, 1; setp.eq.b32 %p1229, %r6988, 1; neg.ftz.f32 %f3690, %f3688; selp.f32 %f3691, %f3690, %f3688, %p1229; and.b32 %r6989, %r99, 2; setp.eq.s32 %p1230, %r6989, 0; neg.ftz.f32 %f3692, %f3689; selp.f32 %f3693, %f3689, %f3692, %p1230; add.ftz.f32 %f238, %f3691, %f3693; xor.b32 %r6990, %r97, %r85; sub.s32 %r6991, %r6990, %r98; xor.b32 %r6993, %r6991, %r6889; shf.l.wrap.b32 %r6994, %r6991, %r6991, 25; sub.s32 %r6995, %r6993, %r6994; xor.b32 %r6996, %r6995, %r97; shf.l.wrap.b32 %r6997, %r6995, %r6995, 16; sub.s32 %r6998, %r6996, %r6997; xor.b32 %r6999, %r6998, %r6991; shf.l.wrap.b32 %r7000, %r6998, %r6998, 4; sub.s32 %r7001, %r6999, %r7000; xor.b32 %r7002, %r7001, %r6995; shf.l.wrap.b32 %r7003, %r7001, %r7001, 14; sub.s32 %r7004, %r7002, %r7003; xor.b32 %r7005, %r7004, %r6998; shf.l.wrap.b32 %r7006, %r7004, %r7004, 24; sub.s32 %r101, %r7005, %r7006; and.b32 %r102, %r101, 15; cvt.u16.u32 %rs46, %r101; or.b16 %rs47, %rs46, 2; and.b16 %rs48, %rs47, 15; setp.eq.s16 %p1231, %rs48, 14; mov.f32 %f5608, %f228; @%p1231 bra $L__BB74_184; mov.f32 %f5608, %f236; $L__BB74_184: setp.lt.u32 %p1232, %r102, 8; selp.f32 %f3694, %f228, %f218, %p1232; setp.lt.u32 %p1233, %r102, 4; selp.f32 %f3695, %f218, %f5608, %p1233; and.b32 %r7007, %r101, 1; setp.eq.b32 %p1234, %r7007, 1; neg.ftz.f32 %f3696, %f3694; selp.f32 %f3697, %f3696, %f3694, %p1234; and.b32 %r7008, %r101, 2; setp.eq.s32 %p1235, %r7008, 0; neg.ftz.f32 %f3698, %f3695; selp.f32 %f3699, %f3695, %f3698, %p1235; add.ftz.f32 %f240, %f3697, %f3699; xor.b32 %r7009, %r96, %r88; sub.s32 %r103, %r7009, %r89; xor.b32 %r7011, %r103, %r6892; shf.l.wrap.b32 %r104, %r103, %r103, 11; sub.s32 %r7012, %r7011, %r104; xor.b32 %r7013, %r7012, %r88; shf.l.wrap.b32 %r7014, %r7012, %r7012, 25; sub.s32 %r7015, %r7013, %r7014; xor.b32 %r7016, %r7015, %r103; shf.l.wrap.b32 %r7017, %r7015, %r7015, 16; sub.s32 %r7018, %r7016, %r7017; xor.b32 %r7019, %r7018, %r7012; shf.l.wrap.b32 %r7020, %r7018, %r7018, 4; sub.s32 %r7021, %r7019, %r7020; xor.b32 %r7022, %r7021, %r7015; shf.l.wrap.b32 %r7023, %r7021, %r7021, 14; sub.s32 %r7024, %r7022, %r7023; xor.b32 %r7025, %r7024, %r7018; shf.l.wrap.b32 %r7026, %r7024, %r7024, 24; sub.s32 %r105, %r7025, %r7026; and.b32 %r106, %r105, 15; cvt.u16.u32 %rs49, %r105; or.b16 %rs50, %rs49, 2; and.b16 %rs51, %rs50, 15; setp.eq.s16 %p1236, %rs51, 14; mov.f32 %f5609, %f217; @%p1236 bra $L__BB74_186; mov.f32 %f5609, %f236; $L__BB74_186: setp.lt.u32 %p1237, %r106, 8; selp.f32 %f3700, %f217, %f231, %p1237; setp.lt.u32 %p1238, %r106, 4; selp.f32 %f3701, %f231, %f5609, %p1238; and.b32 %r7027, %r105, 1; setp.eq.b32 %p1239, %r7027, 1; neg.ftz.f32 %f3702, %f3700; selp.f32 %f3703, %f3702, %f3700, %p1239; and.b32 %r7028, %r105, 2; setp.eq.s32 %p1240, %r7028, 0; neg.ftz.f32 %f3704, %f3701; selp.f32 %f3705, %f3701, %f3704, %p1240; add.ftz.f32 %f242, %f3703, %f3705; xor.b32 %r7029, %r103, %r85; sub.s32 %r7030, %r7029, %r104; xor.b32 %r7031, %r7030, %r88; shf.l.wrap.b32 %r7032, %r7030, %r7030, 25; sub.s32 %r7033, %r7031, %r7032; xor.b32 %r7034, %r7033, %r103; shf.l.wrap.b32 %r7035, %r7033, %r7033, 16; sub.s32 %r7036, %r7034, %r7035; xor.b32 %r7037, %r7036, %r7030; shf.l.wrap.b32 %r7038, %r7036, %r7036, 4; sub.s32 %r7039, %r7037, %r7038; xor.b32 %r7040, %r7039, %r7033; shf.l.wrap.b32 %r7041, %r7039, %r7039, 14; sub.s32 %r7042, %r7040, %r7041; xor.b32 %r7043, %r7042, %r7036; shf.l.wrap.b32 %r7044, %r7042, %r7042, 24; sub.s32 %r107, %r7043, %r7044; and.b32 %r108, %r107, 15; cvt.u16.u32 %rs52, %r107; or.b16 %rs53, %rs52, 2; and.b16 %rs54, %rs53, 15; setp.eq.s16 %p1241, %rs54, 14; mov.f32 %f5610, %f228; @%p1241 bra $L__BB74_188; mov.f32 %f5610, %f236; $L__BB74_188: setp.lt.u32 %p1242, %r108, 8; selp.f32 %f3707, %f228, %f231, %p1242; setp.lt.u32 %p1243, %r108, 4; selp.f32 %f3708, %f231, %f5610, %p1243; and.b32 %r7045, %r107, 1; setp.eq.b32 %p1244, %r7045, 1; neg.ftz.f32 %f3709, %f3707; selp.f32 %f3710, %f3709, %f3707, %p1244; and.b32 %r7046, %r107, 2; setp.eq.s32 %p1245, %r7046, 0; neg.ftz.f32 %f3711, %f3708; selp.f32 %f3712, %f3708, %f3711, %p1245; add.ftz.f32 %f3713, %f3710, %f3712; sub.ftz.f32 %f3715, %f3632, %f224; mul.ftz.f32 %f3716, %f224, %f230; fma.rn.ftz.f32 %f3717, %f3715, %f227, %f3716; sub.ftz.f32 %f3718, %f3632, %f225; mul.ftz.f32 %f3719, %f224, %f235; fma.rn.ftz.f32 %f3720, %f3715, %f233, %f3719; mul.ftz.f32 %f3721, %f225, %f3720; fma.rn.ftz.f32 %f3722, %f3718, %f3717, %f3721; sub.ftz.f32 %f3723, %f3632, %f226; mul.ftz.f32 %f3724, %f224, %f240; fma.rn.ftz.f32 %f3725, %f3715, %f238, %f3724; mul.ftz.f32 %f3726, %f224, %f3713; fma.rn.ftz.f32 %f3727, %f3715, %f242, %f3726; mul.ftz.f32 %f3728, %f225, %f3727; fma.rn.ftz.f32 %f3729, %f3718, %f3725, %f3728; mul.ftz.f32 %f3730, %f226, %f3729; fma.rn.ftz.f32 %f244, %f3723, %f3722, %f3730; mov.b32 %r109, %f244; setp.nan.ftz.f32 %p1246, %f244, %f244; mov.f32 %f5611, 0f00000000; @%p1246 bra $L__BB74_191; or.b32 %r7047, %r109, -2147483648; setp.ne.s32 %p1247, %r7047, -2147483648; add.ftz.f32 %f3732, %f244, %f244; setp.eq.ftz.f32 %p1248, %f3732, %f244; and.pred %p1249, %p1248, %p1247; @%p1249 bra $L__BB74_191; shl.b32 %r7048, %r109, 1; setp.lt.u32 %p1250, %r7048, -16777215; selp.f32 %f5611, %f244, 0f00000000, %p1250; $L__BB74_191: fma.rn.ftz.f32 %f3733, %f5611, 0f3F7B645A, %f21; mul.ftz.f32 %f3734, %f5615, %f3733; fma.rn.ftz.f32 %f5827, %f5827, %f3734, %f5827; mul.ftz.f32 %f5615, %f177, %f5615; mul.ftz.f32 %f5612, %f27, %f5612; mul.ftz.f32 %f5613, %f27, %f5613; mul.ftz.f32 %f5614, %f27, %f5614; add.s32 %r10023, %r10023, 1; setp.lt.s32 %p1251, %r10023, %r75; @%p1251 bra $L__BB74_172; $L__BB74_192: cvt.rmi.ftz.f32.f32 %f3735, %f26; sub.ftz.f32 %f257, %f26, %f3735; setp.eq.ftz.f32 %p1252, %f257, 0f00000000; @%p1252 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f3736, %f5612; cvt.rzi.ftz.s32.f32 %r111, %f3736; sub.ftz.f32 %f258, %f5612, %f3736; cvt.rmi.ftz.f32.f32 %f3737, %f5613; cvt.rzi.ftz.s32.f32 %r112, %f3737; sub.ftz.f32 %f259, %f5613, %f3737; cvt.rmi.ftz.f32.f32 %f3738, %f5614; cvt.rzi.ftz.s32.f32 %r113, %f3738; sub.ftz.f32 %f260, %f5614, %f3738; fma.rn.ftz.f32 %f3739, %f258, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f261, %f258, %f3739, 0f41200000; fma.rn.ftz.f32 %f3740, %f259, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f262, %f259, %f3740, 0f41200000; fma.rn.ftz.f32 %f3741, %f260, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f263, %f260, %f3741, 0f41200000; add.s32 %r7049, %r112, -559038712; add.s32 %r7050, %r113, -559038712; xor.b32 %r7051, %r7050, %r7049; shf.l.wrap.b32 %r114, %r7049, %r7049, 14; sub.s32 %r115, %r7051, %r114; add.s32 %r7052, %r111, -559038712; xor.b32 %r7053, %r115, %r7052; shf.l.wrap.b32 %r116, %r115, %r115, 11; sub.s32 %r7054, %r7053, %r116; xor.b32 %r7055, %r7054, %r7049; shf.l.wrap.b32 %r7056, %r7054, %r7054, 25; sub.s32 %r7057, %r7055, %r7056; xor.b32 %r7058, %r7057, %r115; shf.l.wrap.b32 %r7059, %r7057, %r7057, 16; sub.s32 %r7060, %r7058, %r7059; xor.b32 %r7061, %r7060, %r7054; shf.l.wrap.b32 %r7062, %r7060, %r7060, 4; sub.s32 %r7063, %r7061, %r7062; xor.b32 %r7064, %r7063, %r7057; shf.l.wrap.b32 %r7065, %r7063, %r7063, 14; sub.s32 %r7066, %r7064, %r7065; xor.b32 %r7067, %r7066, %r7060; shf.l.wrap.b32 %r7068, %r7066, %r7066, 24; sub.s32 %r117, %r7067, %r7068; and.b32 %r118, %r117, 15; cvt.u16.u32 %rs55, %r117; or.b16 %rs56, %rs55, 2; and.b16 %rs57, %rs56, 15; setp.eq.s16 %p1253, %rs57, 14; mov.f32 %f5617, %f258; @%p1253 bra $L__BB74_195; mov.f32 %f5617, %f260; $L__BB74_195: mul.ftz.f32 %f3742, %f258, %f258; mul.ftz.f32 %f3743, %f258, %f3742; mul.ftz.f32 %f265, %f3743, %f261; mul.ftz.f32 %f3744, %f259, %f259; mul.ftz.f32 %f3745, %f259, %f3744; mul.ftz.f32 %f266, %f3745, %f262; mul.ftz.f32 %f3746, %f260, %f260; mul.ftz.f32 %f3747, %f260, %f3746; mul.ftz.f32 %f267, %f3747, %f263; setp.lt.u32 %p1254, %r118, 8; selp.f32 %f3748, %f258, %f259, %p1254; setp.lt.u32 %p1255, %r118, 4; selp.f32 %f3749, %f259, %f5617, %p1255; and.b32 %r7069, %r117, 1; setp.eq.b32 %p1256, %r7069, 1; neg.ftz.f32 %f3750, %f3748; selp.f32 %f3751, %f3750, %f3748, %p1256; and.b32 %r7070, %r117, 2; setp.eq.s32 %p1257, %r7070, 0; neg.ftz.f32 %f3752, %f3749; selp.f32 %f3753, %f3749, %f3752, %p1257; add.ftz.f32 %f268, %f3751, %f3753; add.s32 %r119, %r111, -559038711; xor.b32 %r7071, %r115, %r119; sub.s32 %r7072, %r7071, %r116; xor.b32 %r7074, %r7072, %r7049; shf.l.wrap.b32 %r7075, %r7072, %r7072, 25; sub.s32 %r7076, %r7074, %r7075; xor.b32 %r7077, %r7076, %r115; shf.l.wrap.b32 %r7078, %r7076, %r7076, 16; sub.s32 %r7079, %r7077, %r7078; xor.b32 %r7080, %r7079, %r7072; shf.l.wrap.b32 %r7081, %r7079, %r7079, 4; sub.s32 %r7082, %r7080, %r7081; xor.b32 %r7083, %r7082, %r7076; shf.l.wrap.b32 %r7084, %r7082, %r7082, 14; sub.s32 %r7085, %r7083, %r7084; xor.b32 %r7086, %r7085, %r7079; shf.l.wrap.b32 %r7087, %r7085, %r7085, 24; sub.s32 %r120, %r7086, %r7087; and.b32 %r121, %r120, 15; add.ftz.f32 %f269, %f258, 0fBF800000; cvt.u16.u32 %rs58, %r120; or.b16 %rs59, %rs58, 2; and.b16 %rs60, %rs59, 15; setp.eq.s16 %p1258, %rs60, 14; mov.f32 %f5618, %f269; @%p1258 bra $L__BB74_197; mov.f32 %f5618, %f260; $L__BB74_197: setp.lt.u32 %p1259, %r121, 8; selp.f32 %f3754, %f269, %f259, %p1259; setp.lt.u32 %p1260, %r121, 4; selp.f32 %f3755, %f259, %f5618, %p1260; and.b32 %r7088, %r120, 1; setp.eq.b32 %p1261, %r7088, 1; neg.ftz.f32 %f3756, %f3754; selp.f32 %f3757, %f3756, %f3754, %p1261; and.b32 %r7089, %r120, 2; setp.eq.s32 %p1262, %r7089, 0; neg.ftz.f32 %f3758, %f3755; selp.f32 %f3759, %f3755, %f3758, %p1262; add.ftz.f32 %f271, %f3757, %f3759; add.s32 %r122, %r112, -559038711; xor.b32 %r7091, %r7050, %r122; shf.l.wrap.b32 %r123, %r122, %r122, 14; sub.s32 %r124, %r7091, %r123; xor.b32 %r7093, %r124, %r7052; shf.l.wrap.b32 %r125, %r124, %r124, 11; sub.s32 %r7094, %r7093, %r125; xor.b32 %r7095, %r7094, %r122; shf.l.wrap.b32 %r7096, %r7094, %r7094, 25; sub.s32 %r7097, %r7095, %r7096; xor.b32 %r7098, %r7097, %r124; shf.l.wrap.b32 %r7099, %r7097, %r7097, 16; sub.s32 %r7100, %r7098, %r7099; xor.b32 %r7101, %r7100, %r7094; shf.l.wrap.b32 %r7102, %r7100, %r7100, 4; sub.s32 %r7103, %r7101, %r7102; xor.b32 %r7104, %r7103, %r7097; shf.l.wrap.b32 %r7105, %r7103, %r7103, 14; sub.s32 %r7106, %r7104, %r7105; xor.b32 %r7107, %r7106, %r7100; shf.l.wrap.b32 %r7108, %r7106, %r7106, 24; sub.s32 %r126, %r7107, %r7108; and.b32 %r127, %r126, 15; add.ftz.f32 %f272, %f259, 0fBF800000; cvt.u16.u32 %rs61, %r126; or.b16 %rs62, %rs61, 2; and.b16 %rs63, %rs62, 15; setp.eq.s16 %p1263, %rs63, 14; mov.f32 %f5619, %f258; @%p1263 bra $L__BB74_199; mov.f32 %f5619, %f260; $L__BB74_199: setp.lt.u32 %p1264, %r127, 8; selp.f32 %f3760, %f258, %f272, %p1264; setp.lt.u32 %p1265, %r127, 4; selp.f32 %f3761, %f272, %f5619, %p1265; and.b32 %r7109, %r126, 1; setp.eq.b32 %p1266, %r7109, 1; neg.ftz.f32 %f3762, %f3760; selp.f32 %f3763, %f3762, %f3760, %p1266; and.b32 %r7110, %r126, 2; setp.eq.s32 %p1267, %r7110, 0; neg.ftz.f32 %f3764, %f3761; selp.f32 %f3765, %f3761, %f3764, %p1267; add.ftz.f32 %f274, %f3763, %f3765; xor.b32 %r7111, %r124, %r119; sub.s32 %r7112, %r7111, %r125; xor.b32 %r7113, %r7112, %r122; shf.l.wrap.b32 %r7114, %r7112, %r7112, 25; sub.s32 %r7115, %r7113, %r7114; xor.b32 %r7116, %r7115, %r124; shf.l.wrap.b32 %r7117, %r7115, %r7115, 16; sub.s32 %r7118, %r7116, %r7117; xor.b32 %r7119, %r7118, %r7112; shf.l.wrap.b32 %r7120, %r7118, %r7118, 4; sub.s32 %r7121, %r7119, %r7120; xor.b32 %r7122, %r7121, %r7115; shf.l.wrap.b32 %r7123, %r7121, %r7121, 14; sub.s32 %r7124, %r7122, %r7123; xor.b32 %r7125, %r7124, %r7118; shf.l.wrap.b32 %r7126, %r7124, %r7124, 24; sub.s32 %r128, %r7125, %r7126; and.b32 %r129, %r128, 15; cvt.u16.u32 %rs64, %r128; or.b16 %rs65, %rs64, 2; and.b16 %rs66, %rs65, 15; setp.eq.s16 %p1268, %rs66, 14; mov.f32 %f5620, %f269; @%p1268 bra $L__BB74_201; mov.f32 %f5620, %f260; $L__BB74_201: setp.lt.u32 %p1269, %r129, 8; selp.f32 %f3766, %f269, %f272, %p1269; setp.lt.u32 %p1270, %r129, 4; selp.f32 %f3767, %f272, %f5620, %p1270; and.b32 %r7127, %r128, 1; setp.eq.b32 %p1271, %r7127, 1; neg.ftz.f32 %f3768, %f3766; selp.f32 %f3769, %f3768, %f3766, %p1271; and.b32 %r7128, %r128, 2; setp.eq.s32 %p1272, %r7128, 0; neg.ftz.f32 %f3770, %f3767; selp.f32 %f3771, %f3767, %f3770, %p1272; add.ftz.f32 %f276, %f3769, %f3771; add.s32 %r130, %r113, -559038711; xor.b32 %r7130, %r130, %r7049; sub.s32 %r131, %r7130, %r114; xor.b32 %r7132, %r131, %r7052; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r7133, %r7132, %r132; xor.b32 %r7134, %r7133, %r7049; shf.l.wrap.b32 %r7135, %r7133, %r7133, 25; sub.s32 %r7136, %r7134, %r7135; xor.b32 %r7137, %r7136, %r131; shf.l.wrap.b32 %r7138, %r7136, %r7136, 16; sub.s32 %r7139, %r7137, %r7138; xor.b32 %r7140, %r7139, %r7133; shf.l.wrap.b32 %r7141, %r7139, %r7139, 4; sub.s32 %r7142, %r7140, %r7141; xor.b32 %r7143, %r7142, %r7136; shf.l.wrap.b32 %r7144, %r7142, %r7142, 14; sub.s32 %r7145, %r7143, %r7144; xor.b32 %r7146, %r7145, %r7139; shf.l.wrap.b32 %r7147, %r7145, %r7145, 24; sub.s32 %r133, %r7146, %r7147; add.ftz.f32 %f277, %f260, 0fBF800000; and.b32 %r134, %r133, 15; cvt.u16.u32 %rs67, %r133; or.b16 %rs68, %rs67, 2; and.b16 %rs69, %rs68, 15; setp.eq.s16 %p1273, %rs69, 14; mov.f32 %f5621, %f258; @%p1273 bra $L__BB74_203; mov.f32 %f5621, %f277; $L__BB74_203: setp.lt.u32 %p1274, %r134, 8; selp.f32 %f3772, %f258, %f259, %p1274; setp.lt.u32 %p1275, %r134, 4; selp.f32 %f3773, %f259, %f5621, %p1275; and.b32 %r7148, %r133, 1; setp.eq.b32 %p1276, %r7148, 1; neg.ftz.f32 %f3774, %f3772; selp.f32 %f3775, %f3774, %f3772, %p1276; and.b32 %r7149, %r133, 2; setp.eq.s32 %p1277, %r7149, 0; neg.ftz.f32 %f3776, %f3773; selp.f32 %f3777, %f3773, %f3776, %p1277; add.ftz.f32 %f279, %f3775, %f3777; xor.b32 %r7150, %r131, %r119; sub.s32 %r7151, %r7150, %r132; xor.b32 %r7153, %r7151, %r7049; shf.l.wrap.b32 %r7154, %r7151, %r7151, 25; sub.s32 %r7155, %r7153, %r7154; xor.b32 %r7156, %r7155, %r131; shf.l.wrap.b32 %r7157, %r7155, %r7155, 16; sub.s32 %r7158, %r7156, %r7157; xor.b32 %r7159, %r7158, %r7151; shf.l.wrap.b32 %r7160, %r7158, %r7158, 4; sub.s32 %r7161, %r7159, %r7160; xor.b32 %r7162, %r7161, %r7155; shf.l.wrap.b32 %r7163, %r7161, %r7161, 14; sub.s32 %r7164, %r7162, %r7163; xor.b32 %r7165, %r7164, %r7158; shf.l.wrap.b32 %r7166, %r7164, %r7164, 24; sub.s32 %r135, %r7165, %r7166; and.b32 %r136, %r135, 15; cvt.u16.u32 %rs70, %r135; or.b16 %rs71, %rs70, 2; and.b16 %rs72, %rs71, 15; setp.eq.s16 %p1278, %rs72, 14; mov.f32 %f5622, %f269; @%p1278 bra $L__BB74_205; mov.f32 %f5622, %f277; $L__BB74_205: setp.lt.u32 %p1279, %r136, 8; selp.f32 %f3778, %f269, %f259, %p1279; setp.lt.u32 %p1280, %r136, 4; selp.f32 %f3779, %f259, %f5622, %p1280; and.b32 %r7167, %r135, 1; setp.eq.b32 %p1281, %r7167, 1; neg.ftz.f32 %f3780, %f3778; selp.f32 %f3781, %f3780, %f3778, %p1281; and.b32 %r7168, %r135, 2; setp.eq.s32 %p1282, %r7168, 0; neg.ftz.f32 %f3782, %f3779; selp.f32 %f3783, %f3779, %f3782, %p1282; add.ftz.f32 %f281, %f3781, %f3783; xor.b32 %r7169, %r130, %r122; sub.s32 %r137, %r7169, %r123; xor.b32 %r7171, %r137, %r7052; shf.l.wrap.b32 %r138, %r137, %r137, 11; sub.s32 %r7172, %r7171, %r138; xor.b32 %r7173, %r7172, %r122; shf.l.wrap.b32 %r7174, %r7172, %r7172, 25; sub.s32 %r7175, %r7173, %r7174; xor.b32 %r7176, %r7175, %r137; shf.l.wrap.b32 %r7177, %r7175, %r7175, 16; sub.s32 %r7178, %r7176, %r7177; xor.b32 %r7179, %r7178, %r7172; shf.l.wrap.b32 %r7180, %r7178, %r7178, 4; sub.s32 %r7181, %r7179, %r7180; xor.b32 %r7182, %r7181, %r7175; shf.l.wrap.b32 %r7183, %r7181, %r7181, 14; sub.s32 %r7184, %r7182, %r7183; xor.b32 %r7185, %r7184, %r7178; shf.l.wrap.b32 %r7186, %r7184, %r7184, 24; sub.s32 %r139, %r7185, %r7186; and.b32 %r140, %r139, 15; cvt.u16.u32 %rs73, %r139; or.b16 %rs74, %rs73, 2; and.b16 %rs75, %rs74, 15; setp.eq.s16 %p1283, %rs75, 14; mov.f32 %f5623, %f258; @%p1283 bra $L__BB74_207; mov.f32 %f5623, %f277; $L__BB74_207: setp.lt.u32 %p1284, %r140, 8; selp.f32 %f3784, %f258, %f272, %p1284; setp.lt.u32 %p1285, %r140, 4; selp.f32 %f3785, %f272, %f5623, %p1285; and.b32 %r7187, %r139, 1; setp.eq.b32 %p1286, %r7187, 1; neg.ftz.f32 %f3786, %f3784; selp.f32 %f3787, %f3786, %f3784, %p1286; and.b32 %r7188, %r139, 2; setp.eq.s32 %p1287, %r7188, 0; neg.ftz.f32 %f3788, %f3785; selp.f32 %f3789, %f3785, %f3788, %p1287; add.ftz.f32 %f283, %f3787, %f3789; xor.b32 %r7189, %r137, %r119; sub.s32 %r7190, %r7189, %r138; xor.b32 %r7191, %r7190, %r122; shf.l.wrap.b32 %r7192, %r7190, %r7190, 25; sub.s32 %r7193, %r7191, %r7192; xor.b32 %r7194, %r7193, %r137; shf.l.wrap.b32 %r7195, %r7193, %r7193, 16; sub.s32 %r7196, %r7194, %r7195; xor.b32 %r7197, %r7196, %r7190; shf.l.wrap.b32 %r7198, %r7196, %r7196, 4; sub.s32 %r7199, %r7197, %r7198; xor.b32 %r7200, %r7199, %r7193; shf.l.wrap.b32 %r7201, %r7199, %r7199, 14; sub.s32 %r7202, %r7200, %r7201; xor.b32 %r7203, %r7202, %r7196; shf.l.wrap.b32 %r7204, %r7202, %r7202, 24; sub.s32 %r141, %r7203, %r7204; and.b32 %r142, %r141, 15; cvt.u16.u32 %rs76, %r141; or.b16 %rs77, %rs76, 2; and.b16 %rs78, %rs77, 15; setp.eq.s16 %p1288, %rs78, 14; mov.f32 %f5624, %f269; @%p1288 bra $L__BB74_209; mov.f32 %f5624, %f277; $L__BB74_209: setp.lt.u32 %p1289, %r142, 8; selp.f32 %f3791, %f269, %f272, %p1289; setp.lt.u32 %p1290, %r142, 4; selp.f32 %f3792, %f272, %f5624, %p1290; and.b32 %r7205, %r141, 1; setp.eq.b32 %p1291, %r7205, 1; neg.ftz.f32 %f3793, %f3791; selp.f32 %f3794, %f3793, %f3791, %p1291; and.b32 %r7206, %r141, 2; setp.eq.s32 %p1292, %r7206, 0; neg.ftz.f32 %f3795, %f3792; selp.f32 %f3796, %f3792, %f3795, %p1292; add.ftz.f32 %f3797, %f3794, %f3796; sub.ftz.f32 %f3799, %f3632, %f265; mul.ftz.f32 %f3800, %f265, %f271; fma.rn.ftz.f32 %f3801, %f3799, %f268, %f3800; sub.ftz.f32 %f3802, %f3632, %f266; mul.ftz.f32 %f3803, %f265, %f276; fma.rn.ftz.f32 %f3804, %f3799, %f274, %f3803; mul.ftz.f32 %f3805, %f266, %f3804; fma.rn.ftz.f32 %f3806, %f3802, %f3801, %f3805; sub.ftz.f32 %f3807, %f3632, %f267; mul.ftz.f32 %f3808, %f265, %f281; fma.rn.ftz.f32 %f3809, %f3799, %f279, %f3808; mul.ftz.f32 %f3810, %f265, %f3797; fma.rn.ftz.f32 %f3811, %f3799, %f283, %f3810; mul.ftz.f32 %f3812, %f266, %f3811; fma.rn.ftz.f32 %f3813, %f3802, %f3809, %f3812; mul.ftz.f32 %f3814, %f267, %f3813; fma.rn.ftz.f32 %f285, %f3807, %f3806, %f3814; mov.b32 %r143, %f285; setp.nan.ftz.f32 %p1293, %f285, %f285; mov.f32 %f5625, 0f00000000; @%p1293 bra $L__BB74_212; or.b32 %r7207, %r143, -2147483648; setp.ne.s32 %p1294, %r7207, -2147483648; add.ftz.f32 %f3816, %f285, %f285; setp.eq.ftz.f32 %p1295, %f3816, %f285; and.pred %p1296, %p1295, %p1294; @%p1296 bra $L__BB74_212; shl.b32 %r7208, %r143, 1; setp.lt.u32 %p1297, %r7208, -16777215; mul.ftz.f32 %f3817, %f285, 0f3F7B645A; selp.f32 %f5625, %f3817, 0f00000000, %p1297; $L__BB74_212: add.ftz.f32 %f3818, %f21, %f5625; mul.ftz.f32 %f3819, %f5615, %f3818; mul.ftz.f32 %f3820, %f5827, %f3819; fma.rn.ftz.f32 %f5827, %f257, %f3820, %f5827; bra.uni $L__BB74_450; $L__BB74_25: setp.eq.s16 %p1679, %rs271, 0; @%p1679 bra $L__BB74_377; setp.eq.s16 %p1680, %rs271, 1; @%p1680 bra $L__BB74_27; bra.uni $L__BB74_450; $L__BB74_27: cvt.rzi.ftz.s32.f32 %r448, %f26; setp.lt.s32 %p1814, %r448, 1; mov.f32 %f5774, 0f3F800000; mov.f32 %f5827, 0f00000000; @%p1814 bra $L__BB74_33; lg2.approx.ftz.f32 %f4976, %f27; mul.ftz.f32 %f4977, %f25, %f4976; neg.ftz.f32 %f4978, %f4977; ex2.approx.ftz.f32 %f673, %f4978; mov.f32 %f4975, 0f00000000; mov.f32 %f4974, 0f3F800000; mov.u32 %r10031, 0; mov.f32 %f5774, %f4974; mov.f32 %f5827, %f4975; $L__BB74_29: cvt.rmi.ftz.f32.f32 %f4980, %f5771; cvt.rzi.ftz.s32.f32 %r9129, %f4980; sub.ftz.f32 %f4981, %f5771, %f4980; cvt.rmi.ftz.f32.f32 %f4982, %f5772; cvt.rzi.ftz.s32.f32 %r9130, %f4982; sub.ftz.f32 %f4983, %f5772, %f4982; mul.ftz.f32 %f4984, %f4981, %f4981; mul.ftz.f32 %f4985, %f4981, %f4984; fma.rn.ftz.f32 %f4986, %f4981, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4987, %f4981, %f4986, 0f41200000; mul.ftz.f32 %f4988, %f4985, %f4987; mul.ftz.f32 %f4989, %f4983, %f4983; mul.ftz.f32 %f4990, %f4983, %f4989; fma.rn.ftz.f32 %f4991, %f4983, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4992, %f4983, %f4991, 0f41200000; mul.ftz.f32 %f4993, %f4990, %f4992; add.s32 %r9131, %r9130, -559038716; add.s32 %r9132, %r9129, -559038716; xor.b32 %r9133, %r9131, -559038716; shf.l.wrap.b32 %r9134, %r9131, %r9131, 14; sub.s32 %r9135, %r9133, %r9134; xor.b32 %r9136, %r9135, %r9132; shf.l.wrap.b32 %r9137, %r9135, %r9135, 11; sub.s32 %r9138, %r9136, %r9137; xor.b32 %r9139, %r9138, %r9131; shf.l.wrap.b32 %r9140, %r9138, %r9138, 25; sub.s32 %r9141, %r9139, %r9140; xor.b32 %r9142, %r9141, %r9135; shf.l.wrap.b32 %r9143, %r9141, %r9141, 16; sub.s32 %r9144, %r9142, %r9143; xor.b32 %r9145, %r9144, %r9138; shf.l.wrap.b32 %r9146, %r9144, %r9144, 4; sub.s32 %r9147, %r9145, %r9146; xor.b32 %r9148, %r9147, %r9141; shf.l.wrap.b32 %r9149, %r9147, %r9147, 14; sub.s32 %r9150, %r9148, %r9149; xor.b32 %r9151, %r9150, %r9144; shf.l.wrap.b32 %r9152, %r9150, %r9150, 24; sub.s32 %r9153, %r9151, %r9152; and.b32 %r9154, %r9153, 4; setp.eq.s32 %p1815, %r9154, 0; selp.f32 %f4994, %f4981, %f4983, %p1815; selp.f32 %f4995, %f4983, %f4981, %p1815; add.ftz.f32 %f4996, %f4995, %f4995; and.b32 %r9155, %r9153, 1; setp.eq.b32 %p1816, %r9155, 1; neg.ftz.f32 %f4997, %f4994; selp.f32 %f4998, %f4997, %f4994, %p1816; and.b32 %r9156, %r9153, 2; setp.eq.s32 %p1817, %r9156, 0; neg.ftz.f32 %f4999, %f4996; selp.f32 %f5000, %f4996, %f4999, %p1817; add.ftz.f32 %f5001, %f4998, %f5000; add.s32 %r9157, %r9129, -559038715; xor.b32 %r9158, %r9135, %r9157; sub.s32 %r9159, %r9158, %r9137; xor.b32 %r9160, %r9159, %r9131; shf.l.wrap.b32 %r9161, %r9159, %r9159, 25; sub.s32 %r9162, %r9160, %r9161; xor.b32 %r9163, %r9162, %r9135; shf.l.wrap.b32 %r9164, %r9162, %r9162, 16; sub.s32 %r9165, %r9163, %r9164; xor.b32 %r9166, %r9165, %r9159; shf.l.wrap.b32 %r9167, %r9165, %r9165, 4; sub.s32 %r9168, %r9166, %r9167; xor.b32 %r9169, %r9168, %r9162; shf.l.wrap.b32 %r9170, %r9168, %r9168, 14; sub.s32 %r9171, %r9169, %r9170; xor.b32 %r9172, %r9171, %r9165; shf.l.wrap.b32 %r9173, %r9171, %r9171, 24; sub.s32 %r9174, %r9172, %r9173; add.ftz.f32 %f5002, %f4981, 0fBF800000; and.b32 %r9175, %r9174, 4; setp.eq.s32 %p1818, %r9175, 0; selp.f32 %f5004, %f5002, %f4983, %p1818; selp.f32 %f5005, %f4983, %f5002, %p1818; add.ftz.f32 %f5006, %f5005, %f5005; and.b32 %r9176, %r9174, 1; setp.eq.b32 %p1819, %r9176, 1; neg.ftz.f32 %f5007, %f5004; selp.f32 %f5008, %f5007, %f5004, %p1819; and.b32 %r9177, %r9174, 2; setp.eq.s32 %p1820, %r9177, 0; neg.ftz.f32 %f5009, %f5006; selp.f32 %f5010, %f5006, %f5009, %p1820; add.ftz.f32 %f5011, %f5008, %f5010; add.s32 %r9178, %r9130, -559038715; xor.b32 %r9179, %r9178, -559038716; shf.l.wrap.b32 %r9180, %r9178, %r9178, 14; sub.s32 %r9181, %r9179, %r9180; xor.b32 %r9182, %r9181, %r9132; shf.l.wrap.b32 %r9183, %r9181, %r9181, 11; sub.s32 %r9184, %r9182, %r9183; xor.b32 %r9185, %r9184, %r9178; shf.l.wrap.b32 %r9186, %r9184, %r9184, 25; sub.s32 %r9187, %r9185, %r9186; xor.b32 %r9188, %r9187, %r9181; shf.l.wrap.b32 %r9189, %r9187, %r9187, 16; sub.s32 %r9190, %r9188, %r9189; xor.b32 %r9191, %r9190, %r9184; shf.l.wrap.b32 %r9192, %r9190, %r9190, 4; sub.s32 %r9193, %r9191, %r9192; xor.b32 %r9194, %r9193, %r9187; shf.l.wrap.b32 %r9195, %r9193, %r9193, 14; sub.s32 %r9196, %r9194, %r9195; xor.b32 %r9197, %r9196, %r9190; shf.l.wrap.b32 %r9198, %r9196, %r9196, 24; sub.s32 %r9199, %r9197, %r9198; add.ftz.f32 %f5012, %f4983, 0fBF800000; and.b32 %r9200, %r9199, 4; setp.eq.s32 %p1821, %r9200, 0; selp.f32 %f5013, %f4981, %f5012, %p1821; selp.f32 %f5014, %f5012, %f4981, %p1821; add.ftz.f32 %f5015, %f5014, %f5014; and.b32 %r9201, %r9199, 1; setp.eq.b32 %p1822, %r9201, 1; neg.ftz.f32 %f5016, %f5013; selp.f32 %f5017, %f5016, %f5013, %p1822; and.b32 %r9202, %r9199, 2; setp.eq.s32 %p1823, %r9202, 0; neg.ftz.f32 %f5018, %f5015; selp.f32 %f5019, %f5015, %f5018, %p1823; add.ftz.f32 %f5020, %f5017, %f5019; xor.b32 %r9203, %r9181, %r9157; sub.s32 %r9204, %r9203, %r9183; xor.b32 %r9205, %r9204, %r9178; shf.l.wrap.b32 %r9206, %r9204, %r9204, 25; sub.s32 %r9207, %r9205, %r9206; xor.b32 %r9208, %r9207, %r9181; shf.l.wrap.b32 %r9209, %r9207, %r9207, 16; sub.s32 %r9210, %r9208, %r9209; xor.b32 %r9211, %r9210, %r9204; shf.l.wrap.b32 %r9212, %r9210, %r9210, 4; sub.s32 %r9213, %r9211, %r9212; xor.b32 %r9214, %r9213, %r9207; shf.l.wrap.b32 %r9215, %r9213, %r9213, 14; sub.s32 %r9216, %r9214, %r9215; xor.b32 %r9217, %r9216, %r9210; shf.l.wrap.b32 %r9218, %r9216, %r9216, 24; sub.s32 %r9219, %r9217, %r9218; and.b32 %r9220, %r9219, 4; setp.eq.s32 %p1824, %r9220, 0; selp.f32 %f5021, %f5002, %f5012, %p1824; selp.f32 %f5022, %f5012, %f5002, %p1824; add.ftz.f32 %f5023, %f5022, %f5022; and.b32 %r9221, %r9219, 1; setp.eq.b32 %p1825, %r9221, 1; neg.ftz.f32 %f5024, %f5021; selp.f32 %f5025, %f5024, %f5021, %p1825; and.b32 %r9222, %r9219, 2; setp.eq.s32 %p1826, %r9222, 0; neg.ftz.f32 %f5026, %f5023; selp.f32 %f5027, %f5023, %f5026, %p1826; add.ftz.f32 %f5028, %f5025, %f5027; sub.ftz.f32 %f5029, %f4974, %f4988; sub.ftz.f32 %f5030, %f4974, %f4993; mul.ftz.f32 %f5031, %f5029, %f5001; fma.rn.ftz.f32 %f5032, %f4988, %f5011, %f5031; mul.ftz.f32 %f5033, %f5029, %f5020; fma.rn.ftz.f32 %f5034, %f4988, %f5028, %f5033; mul.ftz.f32 %f5035, %f4993, %f5034; fma.rn.ftz.f32 %f678, %f5030, %f5032, %f5035; mov.b32 %r450, %f678; setp.nan.ftz.f32 %p1827, %f678, %f678; mov.f32 %f5770, %f4975; @%p1827 bra $L__BB74_32; or.b32 %r9223, %r450, -2147483648; setp.ne.s32 %p1828, %r9223, -2147483648; add.ftz.f32 %f5037, %f678, %f678; setp.eq.ftz.f32 %p1829, %f5037, %f678; and.pred %p1830, %p1829, %p1828; mov.f32 %f5770, %f4975; @%p1830 bra $L__BB74_32; shl.b32 %r9224, %r450, 1; setp.lt.u32 %p1831, %r9224, -16777215; selp.f32 %f5770, %f678, 0f00000000, %p1831; $L__BB74_32: mul.ftz.f32 %f5038, %f5770, 0f3F295E9E; fma.rn.ftz.f32 %f5827, %f5774, %f5038, %f5827; mul.ftz.f32 %f5774, %f673, %f5774; mul.ftz.f32 %f5771, %f27, %f5771; mul.ftz.f32 %f5772, %f27, %f5772; add.s32 %r10031, %r10031, 1; setp.lt.s32 %p1832, %r10031, %r448; @%p1832 bra $L__BB74_29; $L__BB74_33: cvt.rmi.ftz.f32.f32 %f5039, %f26; sub.ftz.f32 %f689, %f26, %f5039; setp.eq.ftz.f32 %p1833, %f689, 0f00000000; @%p1833 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f5041, %f5771; cvt.rzi.ftz.s32.f32 %r9225, %f5041; sub.ftz.f32 %f5042, %f5771, %f5041; cvt.rmi.ftz.f32.f32 %f5043, %f5772; cvt.rzi.ftz.s32.f32 %r9226, %f5043; sub.ftz.f32 %f5044, %f5772, %f5043; mul.ftz.f32 %f5045, %f5042, %f5042; mul.ftz.f32 %f5046, %f5042, %f5045; fma.rn.ftz.f32 %f5047, %f5042, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5048, %f5042, %f5047, 0f41200000; mul.ftz.f32 %f5049, %f5046, %f5048; mul.ftz.f32 %f5050, %f5044, %f5044; mul.ftz.f32 %f5051, %f5044, %f5050; fma.rn.ftz.f32 %f5052, %f5044, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5053, %f5044, %f5052, 0f41200000; mul.ftz.f32 %f5054, %f5051, %f5053; add.s32 %r9227, %r9226, -559038716; add.s32 %r9228, %r9225, -559038716; xor.b32 %r9229, %r9227, -559038716; shf.l.wrap.b32 %r9230, %r9227, %r9227, 14; sub.s32 %r9231, %r9229, %r9230; xor.b32 %r9232, %r9231, %r9228; shf.l.wrap.b32 %r9233, %r9231, %r9231, 11; sub.s32 %r9234, %r9232, %r9233; xor.b32 %r9235, %r9234, %r9227; shf.l.wrap.b32 %r9236, %r9234, %r9234, 25; sub.s32 %r9237, %r9235, %r9236; xor.b32 %r9238, %r9237, %r9231; shf.l.wrap.b32 %r9239, %r9237, %r9237, 16; sub.s32 %r9240, %r9238, %r9239; xor.b32 %r9241, %r9240, %r9234; shf.l.wrap.b32 %r9242, %r9240, %r9240, 4; sub.s32 %r9243, %r9241, %r9242; xor.b32 %r9244, %r9243, %r9237; shf.l.wrap.b32 %r9245, %r9243, %r9243, 14; sub.s32 %r9246, %r9244, %r9245; xor.b32 %r9247, %r9246, %r9240; shf.l.wrap.b32 %r9248, %r9246, %r9246, 24; sub.s32 %r9249, %r9247, %r9248; and.b32 %r9250, %r9249, 4; setp.eq.s32 %p1834, %r9250, 0; selp.f32 %f5055, %f5042, %f5044, %p1834; selp.f32 %f5056, %f5044, %f5042, %p1834; add.ftz.f32 %f5057, %f5056, %f5056; and.b32 %r9251, %r9249, 1; setp.eq.b32 %p1835, %r9251, 1; neg.ftz.f32 %f5058, %f5055; selp.f32 %f5059, %f5058, %f5055, %p1835; and.b32 %r9252, %r9249, 2; setp.eq.s32 %p1836, %r9252, 0; neg.ftz.f32 %f5060, %f5057; selp.f32 %f5061, %f5057, %f5060, %p1836; add.ftz.f32 %f5062, %f5059, %f5061; add.s32 %r9253, %r9225, -559038715; xor.b32 %r9254, %r9231, %r9253; sub.s32 %r9255, %r9254, %r9233; xor.b32 %r9256, %r9255, %r9227; shf.l.wrap.b32 %r9257, %r9255, %r9255, 25; sub.s32 %r9258, %r9256, %r9257; xor.b32 %r9259, %r9258, %r9231; shf.l.wrap.b32 %r9260, %r9258, %r9258, 16; sub.s32 %r9261, %r9259, %r9260; xor.b32 %r9262, %r9261, %r9255; shf.l.wrap.b32 %r9263, %r9261, %r9261, 4; sub.s32 %r9264, %r9262, %r9263; xor.b32 %r9265, %r9264, %r9258; shf.l.wrap.b32 %r9266, %r9264, %r9264, 14; sub.s32 %r9267, %r9265, %r9266; xor.b32 %r9268, %r9267, %r9261; shf.l.wrap.b32 %r9269, %r9267, %r9267, 24; sub.s32 %r9270, %r9268, %r9269; add.ftz.f32 %f5063, %f5042, 0fBF800000; mov.f32 %f5064, 0f3F800000; and.b32 %r9271, %r9270, 4; setp.eq.s32 %p1837, %r9271, 0; selp.f32 %f5065, %f5063, %f5044, %p1837; selp.f32 %f5066, %f5044, %f5063, %p1837; add.ftz.f32 %f5067, %f5066, %f5066; and.b32 %r9272, %r9270, 1; setp.eq.b32 %p1838, %r9272, 1; neg.ftz.f32 %f5068, %f5065; selp.f32 %f5069, %f5068, %f5065, %p1838; and.b32 %r9273, %r9270, 2; setp.eq.s32 %p1839, %r9273, 0; neg.ftz.f32 %f5070, %f5067; selp.f32 %f5071, %f5067, %f5070, %p1839; add.ftz.f32 %f5072, %f5069, %f5071; add.s32 %r9274, %r9226, -559038715; xor.b32 %r9275, %r9274, -559038716; shf.l.wrap.b32 %r9276, %r9274, %r9274, 14; sub.s32 %r9277, %r9275, %r9276; xor.b32 %r9278, %r9277, %r9228; shf.l.wrap.b32 %r9279, %r9277, %r9277, 11; sub.s32 %r9280, %r9278, %r9279; xor.b32 %r9281, %r9280, %r9274; shf.l.wrap.b32 %r9282, %r9280, %r9280, 25; sub.s32 %r9283, %r9281, %r9282; xor.b32 %r9284, %r9283, %r9277; shf.l.wrap.b32 %r9285, %r9283, %r9283, 16; sub.s32 %r9286, %r9284, %r9285; xor.b32 %r9287, %r9286, %r9280; shf.l.wrap.b32 %r9288, %r9286, %r9286, 4; sub.s32 %r9289, %r9287, %r9288; xor.b32 %r9290, %r9289, %r9283; shf.l.wrap.b32 %r9291, %r9289, %r9289, 14; sub.s32 %r9292, %r9290, %r9291; xor.b32 %r9293, %r9292, %r9286; shf.l.wrap.b32 %r9294, %r9292, %r9292, 24; sub.s32 %r9295, %r9293, %r9294; add.ftz.f32 %f5073, %f5044, 0fBF800000; and.b32 %r9296, %r9295, 4; setp.eq.s32 %p1840, %r9296, 0; selp.f32 %f5074, %f5042, %f5073, %p1840; selp.f32 %f5075, %f5073, %f5042, %p1840; add.ftz.f32 %f5076, %f5075, %f5075; and.b32 %r9297, %r9295, 1; setp.eq.b32 %p1841, %r9297, 1; neg.ftz.f32 %f5077, %f5074; selp.f32 %f5078, %f5077, %f5074, %p1841; and.b32 %r9298, %r9295, 2; setp.eq.s32 %p1842, %r9298, 0; neg.ftz.f32 %f5079, %f5076; selp.f32 %f5080, %f5076, %f5079, %p1842; add.ftz.f32 %f5081, %f5078, %f5080; xor.b32 %r9299, %r9277, %r9253; sub.s32 %r9300, %r9299, %r9279; xor.b32 %r9301, %r9300, %r9274; shf.l.wrap.b32 %r9302, %r9300, %r9300, 25; sub.s32 %r9303, %r9301, %r9302; xor.b32 %r9304, %r9303, %r9277; shf.l.wrap.b32 %r9305, %r9303, %r9303, 16; sub.s32 %r9306, %r9304, %r9305; xor.b32 %r9307, %r9306, %r9300; shf.l.wrap.b32 %r9308, %r9306, %r9306, 4; sub.s32 %r9309, %r9307, %r9308; xor.b32 %r9310, %r9309, %r9303; shf.l.wrap.b32 %r9311, %r9309, %r9309, 14; sub.s32 %r9312, %r9310, %r9311; xor.b32 %r9313, %r9312, %r9306; shf.l.wrap.b32 %r9314, %r9312, %r9312, 24; sub.s32 %r9315, %r9313, %r9314; and.b32 %r9316, %r9315, 4; setp.eq.s32 %p1843, %r9316, 0; selp.f32 %f5082, %f5063, %f5073, %p1843; selp.f32 %f5083, %f5073, %f5063, %p1843; add.ftz.f32 %f5084, %f5083, %f5083; and.b32 %r9317, %r9315, 1; setp.eq.b32 %p1844, %r9317, 1; neg.ftz.f32 %f5085, %f5082; selp.f32 %f5086, %f5085, %f5082, %p1844; and.b32 %r9318, %r9315, 2; setp.eq.s32 %p1845, %r9318, 0; neg.ftz.f32 %f5087, %f5084; selp.f32 %f5088, %f5084, %f5087, %p1845; add.ftz.f32 %f5089, %f5086, %f5088; sub.ftz.f32 %f5090, %f5064, %f5049; sub.ftz.f32 %f5091, %f5064, %f5054; mul.ftz.f32 %f5092, %f5090, %f5062; fma.rn.ftz.f32 %f5093, %f5049, %f5072, %f5092; mul.ftz.f32 %f5094, %f5090, %f5081; fma.rn.ftz.f32 %f5095, %f5049, %f5089, %f5094; mul.ftz.f32 %f5096, %f5054, %f5095; fma.rn.ftz.f32 %f690, %f5091, %f5093, %f5096; mov.b32 %r452, %f690; setp.nan.ftz.f32 %p1846, %f690, %f690; mov.f32 %f5775, 0f00000000; @%p1846 bra $L__BB74_37; or.b32 %r9319, %r452, -2147483648; setp.ne.s32 %p1847, %r9319, -2147483648; add.ftz.f32 %f5098, %f690, %f690; setp.eq.ftz.f32 %p1848, %f5098, %f690; and.pred %p1849, %p1848, %p1847; @%p1849 bra $L__BB74_37; shl.b32 %r9320, %r452, 1; setp.lt.u32 %p1850, %r9320, -16777215; mul.ftz.f32 %f5099, %f690, 0f3F295E9E; selp.f32 %f5775, %f5099, 0f00000000, %p1850; $L__BB74_37: mul.ftz.f32 %f5100, %f689, %f5775; fma.rn.ftz.f32 %f5827, %f5774, %f5100, %f5827; bra.uni $L__BB74_450; $L__BB74_41: setp.eq.s16 %p21, %rs3, 0; @%p21 bra $L__BB74_91; setp.eq.s16 %p22, %rs3, 1; @%p22 bra $L__BB74_43; bra.uni $L__BB74_450; $L__BB74_43: cvt.rzi.ftz.s32.f32 %r32, %f26; setp.lt.s32 %p744, %r32, 1; mov.f32 %f5573, 0f3F800000; mov.f32 %f5827, 0f00000000; @%p744 bra $L__BB74_49; lg2.approx.ftz.f32 %f2571, %f27; mul.ftz.f32 %f2572, %f25, %f2571; neg.ftz.f32 %f2573, %f2572; ex2.approx.ftz.f32 %f119, %f2573; mov.f32 %f2570, 0f00000000; mov.f32 %f2569, 0f3F800000; mov.u32 %r10021, 0; mov.f32 %f5573, %f2569; mov.f32 %f5827, %f2570; $L__BB74_45: cvt.rmi.ftz.f32.f32 %f2575, %f5568; cvt.rzi.ftz.s32.f32 %r4463, %f2575; sub.ftz.f32 %f2576, %f5568, %f2575; cvt.rmi.ftz.f32.f32 %f2577, %f5569; cvt.rzi.ftz.s32.f32 %r4464, %f2577; sub.ftz.f32 %f2578, %f5569, %f2577; cvt.rmi.ftz.f32.f32 %f2579, %f5570; cvt.rzi.ftz.s32.f32 %r4465, %f2579; sub.ftz.f32 %f2580, %f5570, %f2579; cvt.rmi.ftz.f32.f32 %f2581, %f5571; cvt.rzi.ftz.s32.f32 %r4466, %f2581; sub.ftz.f32 %f2582, %f5571, %f2581; mul.ftz.f32 %f2583, %f2576, %f2576; mul.ftz.f32 %f2584, %f2576, %f2583; fma.rn.ftz.f32 %f2585, %f2576, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2586, %f2576, %f2585, 0f41200000; mul.ftz.f32 %f2587, %f2584, %f2586; mul.ftz.f32 %f2588, %f2578, %f2578; mul.ftz.f32 %f2589, %f2578, %f2588; fma.rn.ftz.f32 %f2590, %f2578, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2591, %f2578, %f2590, 0f41200000; mul.ftz.f32 %f2592, %f2589, %f2591; mul.ftz.f32 %f2593, %f2580, %f2580; mul.ftz.f32 %f2594, %f2580, %f2593; fma.rn.ftz.f32 %f2595, %f2580, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2596, %f2580, %f2595, 0f41200000; mul.ftz.f32 %f2597, %f2594, %f2596; mul.ftz.f32 %f2598, %f2582, %f2582; mul.ftz.f32 %f2599, %f2582, %f2598; fma.rn.ftz.f32 %f2600, %f2582, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2601, %f2582, %f2600, 0f41200000; mul.ftz.f32 %f2602, %f2599, %f2601; add.s32 %r4467, %r4464, -559038708; add.s32 %r4468, %r4465, -559038708; sub.s32 %r4469, %r4463, %r4465; shf.l.wrap.b32 %r4470, %r4468, %r4468, 4; xor.b32 %r4471, %r4470, %r4469; add.s32 %r4472, %r4468, %r4467; sub.s32 %r4473, %r4467, %r4471; shf.l.wrap.b32 %r4474, %r4471, %r4471, 6; xor.b32 %r4475, %r4474, %r4473; add.s32 %r4476, %r4471, %r4472; sub.s32 %r4477, %r4472, %r4475; shf.l.wrap.b32 %r4478, %r4475, %r4475, 8; xor.b32 %r4479, %r4478, %r4477; add.s32 %r4480, %r4475, %r4476; sub.s32 %r4481, %r4476, %r4479; shf.l.wrap.b32 %r4482, %r4479, %r4479, 16; xor.b32 %r4483, %r4482, %r4481; add.s32 %r4484, %r4479, %r4480; sub.s32 %r4485, %r4480, %r4483; shf.l.wrap.b32 %r4486, %r4483, %r4483, 19; xor.b32 %r4487, %r4486, %r4485; add.s32 %r4488, %r4483, %r4484; sub.s32 %r4489, %r4484, %r4487; shf.l.wrap.b32 %r4490, %r4487, %r4487, 4; add.s32 %r4491, %r4487, %r4488; add.s32 %r4492, %r4488, %r4466; xor.b32 %r4493, %r4491, %r4489; xor.b32 %r4494, %r4493, %r4490; shf.l.wrap.b32 %r4495, %r4491, %r4491, 14; sub.s32 %r4496, %r4494, %r4495; xor.b32 %r4497, %r4496, %r4492; shf.l.wrap.b32 %r4498, %r4496, %r4496, 11; sub.s32 %r4499, %r4497, %r4498; xor.b32 %r4500, %r4499, %r4491; shf.l.wrap.b32 %r4501, %r4499, %r4499, 25; sub.s32 %r4502, %r4500, %r4501; xor.b32 %r4503, %r4502, %r4496; shf.l.wrap.b32 %r4504, %r4502, %r4502, 16; sub.s32 %r4505, %r4503, %r4504; xor.b32 %r4506, %r4505, %r4499; shf.l.wrap.b32 %r4507, %r4505, %r4505, 4; sub.s32 %r4508, %r4506, %r4507; xor.b32 %r4509, %r4508, %r4502; shf.l.wrap.b32 %r4510, %r4508, %r4508, 14; sub.s32 %r4511, %r4509, %r4510; xor.b32 %r4512, %r4511, %r4505; shf.l.wrap.b32 %r4513, %r4511, %r4511, 24; sub.s32 %r4514, %r4512, %r4513; and.b32 %r4515, %r4514, 31; setp.lt.u32 %p745, %r4515, 24; selp.f32 %f2603, %f2576, %f2578, %p745; setp.lt.u32 %p746, %r4515, 16; selp.f32 %f2604, %f2578, %f2580, %p746; setp.lt.u32 %p747, %r4515, 8; selp.f32 %f2605, %f2580, %f2582, %p747; and.b32 %r4516, %r4514, 1; setp.eq.b32 %p748, %r4516, 1; neg.ftz.f32 %f2606, %f2603; selp.f32 %f2607, %f2606, %f2603, %p748; and.b32 %r4517, %r4514, 2; setp.eq.s32 %p749, %r4517, 0; neg.ftz.f32 %f2608, %f2604; selp.f32 %f2609, %f2604, %f2608, %p749; add.ftz.f32 %f2610, %f2607, %f2609; and.b32 %r4518, %r4514, 4; setp.eq.s32 %p750, %r4518, 0; neg.ftz.f32 %f2611, %f2605; selp.f32 %f2612, %f2605, %f2611, %p750; add.ftz.f32 %f2613, %f2612, %f2610; add.s32 %r4519, %r4463, 1; sub.s32 %r4520, %r4519, %r4465; xor.b32 %r4521, %r4470, %r4520; sub.s32 %r4522, %r4467, %r4521; shf.l.wrap.b32 %r4523, %r4521, %r4521, 6; xor.b32 %r4524, %r4523, %r4522; add.s32 %r4525, %r4521, %r4472; sub.s32 %r4526, %r4472, %r4524; shf.l.wrap.b32 %r4527, %r4524, %r4524, 8; xor.b32 %r4528, %r4527, %r4526; add.s32 %r4529, %r4524, %r4525; sub.s32 %r4530, %r4525, %r4528; shf.l.wrap.b32 %r4531, %r4528, %r4528, 16; xor.b32 %r4532, %r4531, %r4530; add.s32 %r4533, %r4528, %r4529; sub.s32 %r4534, %r4529, %r4532; shf.l.wrap.b32 %r4535, %r4532, %r4532, 19; xor.b32 %r4536, %r4535, %r4534; add.s32 %r4537, %r4532, %r4533; sub.s32 %r4538, %r4533, %r4536; shf.l.wrap.b32 %r4539, %r4536, %r4536, 4; add.s32 %r4540, %r4536, %r4537; add.s32 %r4541, %r4537, %r4466; xor.b32 %r4542, %r4540, %r4538; xor.b32 %r4543, %r4542, %r4539; shf.l.wrap.b32 %r4544, %r4540, %r4540, 14; sub.s32 %r4545, %r4543, %r4544; xor.b32 %r4546, %r4545, %r4541; shf.l.wrap.b32 %r4547, %r4545, %r4545, 11; sub.s32 %r4548, %r4546, %r4547; xor.b32 %r4549, %r4548, %r4540; shf.l.wrap.b32 %r4550, %r4548, %r4548, 25; sub.s32 %r4551, %r4549, %r4550; xor.b32 %r4552, %r4551, %r4545; shf.l.wrap.b32 %r4553, %r4551, %r4551, 16; sub.s32 %r4554, %r4552, %r4553; xor.b32 %r4555, %r4554, %r4548; shf.l.wrap.b32 %r4556, %r4554, %r4554, 4; sub.s32 %r4557, %r4555, %r4556; xor.b32 %r4558, %r4557, %r4551; shf.l.wrap.b32 %r4559, %r4557, %r4557, 14; sub.s32 %r4560, %r4558, %r4559; xor.b32 %r4561, %r4560, %r4554; shf.l.wrap.b32 %r4562, %r4560, %r4560, 24; sub.s32 %r4563, %r4561, %r4562; add.ftz.f32 %f2614, %f2576, 0fBF800000; and.b32 %r4564, %r4563, 31; setp.lt.u32 %p751, %r4564, 24; selp.f32 %f2616, %f2614, %f2578, %p751; setp.lt.u32 %p752, %r4564, 16; selp.f32 %f2617, %f2578, %f2580, %p752; setp.lt.u32 %p753, %r4564, 8; selp.f32 %f2618, %f2580, %f2582, %p753; and.b32 %r4565, %r4563, 1; setp.eq.b32 %p754, %r4565, 1; neg.ftz.f32 %f2619, %f2616; selp.f32 %f2620, %f2619, %f2616, %p754; and.b32 %r4566, %r4563, 2; setp.eq.s32 %p755, %r4566, 0; neg.ftz.f32 %f2621, %f2617; selp.f32 %f2622, %f2617, %f2621, %p755; add.ftz.f32 %f2623, %f2620, %f2622; and.b32 %r4567, %r4563, 4; setp.eq.s32 %p756, %r4567, 0; neg.ftz.f32 %f2624, %f2618; selp.f32 %f2625, %f2618, %f2624, %p756; add.ftz.f32 %f2626, %f2625, %f2623; add.s32 %r4568, %r4464, -559038707; add.s32 %r4569, %r4468, %r4568; sub.s32 %r4570, %r4568, %r4471; xor.b32 %r4571, %r4474, %r4570; add.s32 %r4572, %r4471, %r4569; sub.s32 %r4573, %r4569, %r4571; shf.l.wrap.b32 %r4574, %r4571, %r4571, 8; xor.b32 %r4575, %r4574, %r4573; add.s32 %r4576, %r4571, %r4572; sub.s32 %r4577, %r4572, %r4575; shf.l.wrap.b32 %r4578, %r4575, %r4575, 16; xor.b32 %r4579, %r4578, %r4577; add.s32 %r4580, %r4575, %r4576; sub.s32 %r4581, %r4576, %r4579; shf.l.wrap.b32 %r4582, %r4579, %r4579, 19; xor.b32 %r4583, %r4582, %r4581; add.s32 %r4584, %r4579, %r4580; sub.s32 %r4585, %r4580, %r4583; shf.l.wrap.b32 %r4586, %r4583, %r4583, 4; add.s32 %r4587, %r4583, %r4584; add.s32 %r4588, %r4584, %r4466; xor.b32 %r4589, %r4587, %r4585; xor.b32 %r4590, %r4589, %r4586; shf.l.wrap.b32 %r4591, %r4587, %r4587, 14; sub.s32 %r4592, %r4590, %r4591; xor.b32 %r4593, %r4592, %r4588; shf.l.wrap.b32 %r4594, %r4592, %r4592, 11; sub.s32 %r4595, %r4593, %r4594; xor.b32 %r4596, %r4595, %r4587; shf.l.wrap.b32 %r4597, %r4595, %r4595, 25; sub.s32 %r4598, %r4596, %r4597; xor.b32 %r4599, %r4598, %r4592; shf.l.wrap.b32 %r4600, %r4598, %r4598, 16; sub.s32 %r4601, %r4599, %r4600; xor.b32 %r4602, %r4601, %r4595; shf.l.wrap.b32 %r4603, %r4601, %r4601, 4; sub.s32 %r4604, %r4602, %r4603; xor.b32 %r4605, %r4604, %r4598; shf.l.wrap.b32 %r4606, %r4604, %r4604, 14; sub.s32 %r4607, %r4605, %r4606; xor.b32 %r4608, %r4607, %r4601; shf.l.wrap.b32 %r4609, %r4607, %r4607, 24; sub.s32 %r4610, %r4608, %r4609; add.ftz.f32 %f2627, %f2578, 0fBF800000; and.b32 %r4611, %r4610, 31; setp.lt.u32 %p757, %r4611, 24; selp.f32 %f2628, %f2576, %f2627, %p757; setp.lt.u32 %p758, %r4611, 16; selp.f32 %f2629, %f2627, %f2580, %p758; setp.lt.u32 %p759, %r4611, 8; selp.f32 %f2630, %f2580, %f2582, %p759; and.b32 %r4612, %r4610, 1; setp.eq.b32 %p760, %r4612, 1; neg.ftz.f32 %f2631, %f2628; selp.f32 %f2632, %f2631, %f2628, %p760; and.b32 %r4613, %r4610, 2; setp.eq.s32 %p761, %r4613, 0; neg.ftz.f32 %f2633, %f2629; selp.f32 %f2634, %f2629, %f2633, %p761; add.ftz.f32 %f2635, %f2632, %f2634; and.b32 %r4614, %r4610, 4; setp.eq.s32 %p762, %r4614, 0; neg.ftz.f32 %f2636, %f2630; selp.f32 %f2637, %f2630, %f2636, %p762; add.ftz.f32 %f2638, %f2637, %f2635; sub.s32 %r4615, %r4568, %r4521; xor.b32 %r4616, %r4523, %r4615; add.s32 %r4617, %r4521, %r4569; sub.s32 %r4618, %r4569, %r4616; shf.l.wrap.b32 %r4619, %r4616, %r4616, 8; xor.b32 %r4620, %r4619, %r4618; add.s32 %r4621, %r4616, %r4617; sub.s32 %r4622, %r4617, %r4620; shf.l.wrap.b32 %r4623, %r4620, %r4620, 16; xor.b32 %r4624, %r4623, %r4622; add.s32 %r4625, %r4620, %r4621; sub.s32 %r4626, %r4621, %r4624; shf.l.wrap.b32 %r4627, %r4624, %r4624, 19; xor.b32 %r4628, %r4627, %r4626; add.s32 %r4629, %r4624, %r4625; sub.s32 %r4630, %r4625, %r4628; shf.l.wrap.b32 %r4631, %r4628, %r4628, 4; add.s32 %r4632, %r4628, %r4629; add.s32 %r4633, %r4629, %r4466; xor.b32 %r4634, %r4632, %r4630; xor.b32 %r4635, %r4634, %r4631; shf.l.wrap.b32 %r4636, %r4632, %r4632, 14; sub.s32 %r4637, %r4635, %r4636; xor.b32 %r4638, %r4637, %r4633; shf.l.wrap.b32 %r4639, %r4637, %r4637, 11; sub.s32 %r4640, %r4638, %r4639; xor.b32 %r4641, %r4640, %r4632; shf.l.wrap.b32 %r4642, %r4640, %r4640, 25; sub.s32 %r4643, %r4641, %r4642; xor.b32 %r4644, %r4643, %r4637; shf.l.wrap.b32 %r4645, %r4643, %r4643, 16; sub.s32 %r4646, %r4644, %r4645; xor.b32 %r4647, %r4646, %r4640; shf.l.wrap.b32 %r4648, %r4646, %r4646, 4; sub.s32 %r4649, %r4647, %r4648; xor.b32 %r4650, %r4649, %r4643; shf.l.wrap.b32 %r4651, %r4649, %r4649, 14; sub.s32 %r4652, %r4650, %r4651; xor.b32 %r4653, %r4652, %r4646; shf.l.wrap.b32 %r4654, %r4652, %r4652, 24; sub.s32 %r4655, %r4653, %r4654; and.b32 %r4656, %r4655, 31; setp.lt.u32 %p763, %r4656, 24; selp.f32 %f2639, %f2614, %f2627, %p763; setp.lt.u32 %p764, %r4656, 16; selp.f32 %f2640, %f2627, %f2580, %p764; setp.lt.u32 %p765, %r4656, 8; selp.f32 %f2641, %f2580, %f2582, %p765; and.b32 %r4657, %r4655, 1; setp.eq.b32 %p766, %r4657, 1; neg.ftz.f32 %f2642, %f2639; selp.f32 %f2643, %f2642, %f2639, %p766; and.b32 %r4658, %r4655, 2; setp.eq.s32 %p767, %r4658, 0; neg.ftz.f32 %f2644, %f2640; selp.f32 %f2645, %f2640, %f2644, %p767; add.ftz.f32 %f2646, %f2643, %f2645; and.b32 %r4659, %r4655, 4; setp.eq.s32 %p768, %r4659, 0; neg.ftz.f32 %f2647, %f2641; selp.f32 %f2648, %f2641, %f2647, %p768; add.ftz.f32 %f2649, %f2648, %f2646; not.b32 %r4660, %r4465; add.s32 %r4661, %r4465, -559038707; add.s32 %r4662, %r4463, %r4660; shf.l.wrap.b32 %r4663, %r4661, %r4661, 4; xor.b32 %r4664, %r4663, %r4662; add.s32 %r4665, %r4661, %r4467; sub.s32 %r4666, %r4467, %r4664; shf.l.wrap.b32 %r4667, %r4664, %r4664, 6; xor.b32 %r4668, %r4667, %r4666; add.s32 %r4669, %r4664, %r4665; sub.s32 %r4670, %r4665, %r4668; shf.l.wrap.b32 %r4671, %r4668, %r4668, 8; xor.b32 %r4672, %r4671, %r4670; add.s32 %r4673, %r4668, %r4669; sub.s32 %r4674, %r4669, %r4672; shf.l.wrap.b32 %r4675, %r4672, %r4672, 16; xor.b32 %r4676, %r4675, %r4674; add.s32 %r4677, %r4672, %r4673; sub.s32 %r4678, %r4673, %r4676; shf.l.wrap.b32 %r4679, %r4676, %r4676, 19; xor.b32 %r4680, %r4679, %r4678; add.s32 %r4681, %r4676, %r4677; sub.s32 %r4682, %r4677, %r4680; shf.l.wrap.b32 %r4683, %r4680, %r4680, 4; add.s32 %r4684, %r4680, %r4681; add.s32 %r4685, %r4681, %r4466; xor.b32 %r4686, %r4684, %r4682; xor.b32 %r4687, %r4686, %r4683; shf.l.wrap.b32 %r4688, %r4684, %r4684, 14; sub.s32 %r4689, %r4687, %r4688; xor.b32 %r4690, %r4689, %r4685; shf.l.wrap.b32 %r4691, %r4689, %r4689, 11; sub.s32 %r4692, %r4690, %r4691; xor.b32 %r4693, %r4692, %r4684; shf.l.wrap.b32 %r4694, %r4692, %r4692, 25; sub.s32 %r4695, %r4693, %r4694; xor.b32 %r4696, %r4695, %r4689; shf.l.wrap.b32 %r4697, %r4695, %r4695, 16; sub.s32 %r4698, %r4696, %r4697; xor.b32 %r4699, %r4698, %r4692; shf.l.wrap.b32 %r4700, %r4698, %r4698, 4; sub.s32 %r4701, %r4699, %r4700; xor.b32 %r4702, %r4701, %r4695; shf.l.wrap.b32 %r4703, %r4701, %r4701, 14; sub.s32 %r4704, %r4702, %r4703; xor.b32 %r4705, %r4704, %r4698; shf.l.wrap.b32 %r4706, %r4704, %r4704, 24; sub.s32 %r4707, %r4705, %r4706; add.ftz.f32 %f2650, %f2580, 0fBF800000; and.b32 %r4708, %r4707, 31; setp.lt.u32 %p769, %r4708, 24; selp.f32 %f2651, %f2576, %f2578, %p769; setp.lt.u32 %p770, %r4708, 16; selp.f32 %f2652, %f2578, %f2650, %p770; setp.lt.u32 %p771, %r4708, 8; selp.f32 %f2653, %f2650, %f2582, %p771; and.b32 %r4709, %r4707, 1; setp.eq.b32 %p772, %r4709, 1; neg.ftz.f32 %f2654, %f2651; selp.f32 %f2655, %f2654, %f2651, %p772; and.b32 %r4710, %r4707, 2; setp.eq.s32 %p773, %r4710, 0; neg.ftz.f32 %f2656, %f2652; selp.f32 %f2657, %f2652, %f2656, %p773; add.ftz.f32 %f2658, %f2655, %f2657; and.b32 %r4711, %r4707, 4; setp.eq.s32 %p774, %r4711, 0; neg.ftz.f32 %f2659, %f2653; selp.f32 %f2660, %f2653, %f2659, %p774; add.ftz.f32 %f2661, %f2660, %f2658; xor.b32 %r4712, %r4663, %r4469; sub.s32 %r4713, %r4467, %r4712; shf.l.wrap.b32 %r4714, %r4712, %r4712, 6; xor.b32 %r4715, %r4714, %r4713; add.s32 %r4716, %r4712, %r4665; sub.s32 %r4717, %r4665, %r4715; shf.l.wrap.b32 %r4718, %r4715, %r4715, 8; xor.b32 %r4719, %r4718, %r4717; add.s32 %r4720, %r4715, %r4716; sub.s32 %r4721, %r4716, %r4719; shf.l.wrap.b32 %r4722, %r4719, %r4719, 16; xor.b32 %r4723, %r4722, %r4721; add.s32 %r4724, %r4719, %r4720; sub.s32 %r4725, %r4720, %r4723; shf.l.wrap.b32 %r4726, %r4723, %r4723, 19; xor.b32 %r4727, %r4726, %r4725; add.s32 %r4728, %r4723, %r4724; sub.s32 %r4729, %r4724, %r4727; shf.l.wrap.b32 %r4730, %r4727, %r4727, 4; add.s32 %r4731, %r4727, %r4728; add.s32 %r4732, %r4728, %r4466; xor.b32 %r4733, %r4731, %r4729; xor.b32 %r4734, %r4733, %r4730; shf.l.wrap.b32 %r4735, %r4731, %r4731, 14; sub.s32 %r4736, %r4734, %r4735; xor.b32 %r4737, %r4736, %r4732; shf.l.wrap.b32 %r4738, %r4736, %r4736, 11; sub.s32 %r4739, %r4737, %r4738; xor.b32 %r4740, %r4739, %r4731; shf.l.wrap.b32 %r4741, %r4739, %r4739, 25; sub.s32 %r4742, %r4740, %r4741; xor.b32 %r4743, %r4742, %r4736; shf.l.wrap.b32 %r4744, %r4742, %r4742, 16; sub.s32 %r4745, %r4743, %r4744; xor.b32 %r4746, %r4745, %r4739; shf.l.wrap.b32 %r4747, %r4745, %r4745, 4; sub.s32 %r4748, %r4746, %r4747; xor.b32 %r4749, %r4748, %r4742; shf.l.wrap.b32 %r4750, %r4748, %r4748, 14; sub.s32 %r4751, %r4749, %r4750; xor.b32 %r4752, %r4751, %r4745; shf.l.wrap.b32 %r4753, %r4751, %r4751, 24; sub.s32 %r4754, %r4752, %r4753; and.b32 %r4755, %r4754, 31; setp.lt.u32 %p775, %r4755, 24; selp.f32 %f2662, %f2614, %f2578, %p775; setp.lt.u32 %p776, %r4755, 16; selp.f32 %f2663, %f2578, %f2650, %p776; setp.lt.u32 %p777, %r4755, 8; selp.f32 %f2664, %f2650, %f2582, %p777; and.b32 %r4756, %r4754, 1; setp.eq.b32 %p778, %r4756, 1; neg.ftz.f32 %f2665, %f2662; selp.f32 %f2666, %f2665, %f2662, %p778; and.b32 %r4757, %r4754, 2; setp.eq.s32 %p779, %r4757, 0; neg.ftz.f32 %f2667, %f2663; selp.f32 %f2668, %f2663, %f2667, %p779; add.ftz.f32 %f2669, %f2666, %f2668; and.b32 %r4758, %r4754, 4; setp.eq.s32 %p780, %r4758, 0; neg.ftz.f32 %f2670, %f2664; selp.f32 %f2671, %f2664, %f2670, %p780; add.ftz.f32 %f2672, %f2671, %f2669; add.s32 %r4759, %r4661, %r4568; sub.s32 %r4760, %r4568, %r4664; xor.b32 %r4761, %r4667, %r4760; add.s32 %r4762, %r4664, %r4759; sub.s32 %r4763, %r4759, %r4761; shf.l.wrap.b32 %r4764, %r4761, %r4761, 8; xor.b32 %r4765, %r4764, %r4763; add.s32 %r4766, %r4761, %r4762; sub.s32 %r4767, %r4762, %r4765; shf.l.wrap.b32 %r4768, %r4765, %r4765, 16; xor.b32 %r4769, %r4768, %r4767; add.s32 %r4770, %r4765, %r4766; sub.s32 %r4771, %r4766, %r4769; shf.l.wrap.b32 %r4772, %r4769, %r4769, 19; xor.b32 %r4773, %r4772, %r4771; add.s32 %r4774, %r4769, %r4770; sub.s32 %r4775, %r4770, %r4773; shf.l.wrap.b32 %r4776, %r4773, %r4773, 4; add.s32 %r4777, %r4773, %r4774; add.s32 %r4778, %r4774, %r4466; xor.b32 %r4779, %r4777, %r4775; xor.b32 %r4780, %r4779, %r4776; shf.l.wrap.b32 %r4781, %r4777, %r4777, 14; sub.s32 %r4782, %r4780, %r4781; xor.b32 %r4783, %r4782, %r4778; shf.l.wrap.b32 %r4784, %r4782, %r4782, 11; sub.s32 %r4785, %r4783, %r4784; xor.b32 %r4786, %r4785, %r4777; shf.l.wrap.b32 %r4787, %r4785, %r4785, 25; sub.s32 %r4788, %r4786, %r4787; xor.b32 %r4789, %r4788, %r4782; shf.l.wrap.b32 %r4790, %r4788, %r4788, 16; sub.s32 %r4791, %r4789, %r4790; xor.b32 %r4792, %r4791, %r4785; shf.l.wrap.b32 %r4793, %r4791, %r4791, 4; sub.s32 %r4794, %r4792, %r4793; xor.b32 %r4795, %r4794, %r4788; shf.l.wrap.b32 %r4796, %r4794, %r4794, 14; sub.s32 %r4797, %r4795, %r4796; xor.b32 %r4798, %r4797, %r4791; shf.l.wrap.b32 %r4799, %r4797, %r4797, 24; sub.s32 %r4800, %r4798, %r4799; and.b32 %r4801, %r4800, 31; setp.lt.u32 %p781, %r4801, 24; selp.f32 %f2673, %f2576, %f2627, %p781; setp.lt.u32 %p782, %r4801, 16; selp.f32 %f2674, %f2627, %f2650, %p782; setp.lt.u32 %p783, %r4801, 8; selp.f32 %f2675, %f2650, %f2582, %p783; and.b32 %r4802, %r4800, 1; setp.eq.b32 %p784, %r4802, 1; neg.ftz.f32 %f2676, %f2673; selp.f32 %f2677, %f2676, %f2673, %p784; and.b32 %r4803, %r4800, 2; setp.eq.s32 %p785, %r4803, 0; neg.ftz.f32 %f2678, %f2674; selp.f32 %f2679, %f2674, %f2678, %p785; add.ftz.f32 %f2680, %f2677, %f2679; and.b32 %r4804, %r4800, 4; setp.eq.s32 %p786, %r4804, 0; neg.ftz.f32 %f2681, %f2675; selp.f32 %f2682, %f2675, %f2681, %p786; add.ftz.f32 %f2683, %f2682, %f2680; sub.s32 %r4805, %r4568, %r4712; xor.b32 %r4806, %r4714, %r4805; add.s32 %r4807, %r4712, %r4759; sub.s32 %r4808, %r4759, %r4806; shf.l.wrap.b32 %r4809, %r4806, %r4806, 8; xor.b32 %r4810, %r4809, %r4808; add.s32 %r4811, %r4806, %r4807; sub.s32 %r4812, %r4807, %r4810; shf.l.wrap.b32 %r4813, %r4810, %r4810, 16; xor.b32 %r4814, %r4813, %r4812; add.s32 %r4815, %r4810, %r4811; sub.s32 %r4816, %r4811, %r4814; shf.l.wrap.b32 %r4817, %r4814, %r4814, 19; xor.b32 %r4818, %r4817, %r4816; add.s32 %r4819, %r4814, %r4815; sub.s32 %r4820, %r4815, %r4818; shf.l.wrap.b32 %r4821, %r4818, %r4818, 4; add.s32 %r4822, %r4818, %r4819; add.s32 %r4823, %r4819, %r4466; xor.b32 %r4824, %r4822, %r4820; xor.b32 %r4825, %r4824, %r4821; shf.l.wrap.b32 %r4826, %r4822, %r4822, 14; sub.s32 %r4827, %r4825, %r4826; xor.b32 %r4828, %r4827, %r4823; shf.l.wrap.b32 %r4829, %r4827, %r4827, 11; sub.s32 %r4830, %r4828, %r4829; xor.b32 %r4831, %r4830, %r4822; shf.l.wrap.b32 %r4832, %r4830, %r4830, 25; sub.s32 %r4833, %r4831, %r4832; xor.b32 %r4834, %r4833, %r4827; shf.l.wrap.b32 %r4835, %r4833, %r4833, 16; sub.s32 %r4836, %r4834, %r4835; xor.b32 %r4837, %r4836, %r4830; shf.l.wrap.b32 %r4838, %r4836, %r4836, 4; sub.s32 %r4839, %r4837, %r4838; xor.b32 %r4840, %r4839, %r4833; shf.l.wrap.b32 %r4841, %r4839, %r4839, 14; sub.s32 %r4842, %r4840, %r4841; xor.b32 %r4843, %r4842, %r4836; shf.l.wrap.b32 %r4844, %r4842, %r4842, 24; sub.s32 %r4845, %r4843, %r4844; and.b32 %r4846, %r4845, 31; setp.lt.u32 %p787, %r4846, 24; selp.f32 %f2684, %f2614, %f2627, %p787; setp.lt.u32 %p788, %r4846, 16; selp.f32 %f2685, %f2627, %f2650, %p788; setp.lt.u32 %p789, %r4846, 8; selp.f32 %f2686, %f2650, %f2582, %p789; and.b32 %r4847, %r4845, 1; setp.eq.b32 %p790, %r4847, 1; neg.ftz.f32 %f2687, %f2684; selp.f32 %f2688, %f2687, %f2684, %p790; and.b32 %r4848, %r4845, 2; setp.eq.s32 %p791, %r4848, 0; neg.ftz.f32 %f2689, %f2685; selp.f32 %f2690, %f2685, %f2689, %p791; add.ftz.f32 %f2691, %f2688, %f2690; and.b32 %r4849, %r4845, 4; setp.eq.s32 %p792, %r4849, 0; neg.ftz.f32 %f2692, %f2686; selp.f32 %f2693, %f2686, %f2692, %p792; add.ftz.f32 %f2694, %f2693, %f2691; add.s32 %r4850, %r4466, 1; add.s32 %r4851, %r4488, %r4850; xor.b32 %r4852, %r4496, %r4851; sub.s32 %r4853, %r4852, %r4498; xor.b32 %r4854, %r4853, %r4491; shf.l.wrap.b32 %r4855, %r4853, %r4853, 25; sub.s32 %r4856, %r4854, %r4855; xor.b32 %r4857, %r4856, %r4496; shf.l.wrap.b32 %r4858, %r4856, %r4856, 16; sub.s32 %r4859, %r4857, %r4858; xor.b32 %r4860, %r4859, %r4853; shf.l.wrap.b32 %r4861, %r4859, %r4859, 4; sub.s32 %r4862, %r4860, %r4861; xor.b32 %r4863, %r4862, %r4856; shf.l.wrap.b32 %r4864, %r4862, %r4862, 14; sub.s32 %r4865, %r4863, %r4864; xor.b32 %r4866, %r4865, %r4859; shf.l.wrap.b32 %r4867, %r4865, %r4865, 24; sub.s32 %r4868, %r4866, %r4867; add.ftz.f32 %f2695, %f2582, 0fBF800000; and.b32 %r4869, %r4868, 31; setp.lt.u32 %p793, %r4869, 24; selp.f32 %f2696, %f2576, %f2578, %p793; setp.lt.u32 %p794, %r4869, 16; selp.f32 %f2697, %f2578, %f2580, %p794; setp.lt.u32 %p795, %r4869, 8; selp.f32 %f2698, %f2580, %f2695, %p795; and.b32 %r4870, %r4868, 1; setp.eq.b32 %p796, %r4870, 1; neg.ftz.f32 %f2699, %f2696; selp.f32 %f2700, %f2699, %f2696, %p796; and.b32 %r4871, %r4868, 2; setp.eq.s32 %p797, %r4871, 0; neg.ftz.f32 %f2701, %f2697; selp.f32 %f2702, %f2697, %f2701, %p797; add.ftz.f32 %f2703, %f2700, %f2702; and.b32 %r4872, %r4868, 4; setp.eq.s32 %p798, %r4872, 0; neg.ftz.f32 %f2704, %f2698; selp.f32 %f2705, %f2698, %f2704, %p798; add.ftz.f32 %f2706, %f2705, %f2703; add.s32 %r4873, %r4537, %r4850; xor.b32 %r4874, %r4545, %r4873; sub.s32 %r4875, %r4874, %r4547; xor.b32 %r4876, %r4875, %r4540; shf.l.wrap.b32 %r4877, %r4875, %r4875, 25; sub.s32 %r4878, %r4876, %r4877; xor.b32 %r4879, %r4878, %r4545; shf.l.wrap.b32 %r4880, %r4878, %r4878, 16; sub.s32 %r4881, %r4879, %r4880; xor.b32 %r4882, %r4881, %r4875; shf.l.wrap.b32 %r4883, %r4881, %r4881, 4; sub.s32 %r4884, %r4882, %r4883; xor.b32 %r4885, %r4884, %r4878; shf.l.wrap.b32 %r4886, %r4884, %r4884, 14; sub.s32 %r4887, %r4885, %r4886; xor.b32 %r4888, %r4887, %r4881; shf.l.wrap.b32 %r4889, %r4887, %r4887, 24; sub.s32 %r4890, %r4888, %r4889; and.b32 %r4891, %r4890, 31; setp.lt.u32 %p799, %r4891, 24; selp.f32 %f2707, %f2614, %f2578, %p799; setp.lt.u32 %p800, %r4891, 16; selp.f32 %f2708, %f2578, %f2580, %p800; setp.lt.u32 %p801, %r4891, 8; selp.f32 %f2709, %f2580, %f2695, %p801; and.b32 %r4892, %r4890, 1; setp.eq.b32 %p802, %r4892, 1; neg.ftz.f32 %f2710, %f2707; selp.f32 %f2711, %f2710, %f2707, %p802; and.b32 %r4893, %r4890, 2; setp.eq.s32 %p803, %r4893, 0; neg.ftz.f32 %f2712, %f2708; selp.f32 %f2713, %f2708, %f2712, %p803; add.ftz.f32 %f2714, %f2711, %f2713; and.b32 %r4894, %r4890, 4; setp.eq.s32 %p804, %r4894, 0; neg.ftz.f32 %f2715, %f2709; selp.f32 %f2716, %f2709, %f2715, %p804; add.ftz.f32 %f2717, %f2716, %f2714; add.s32 %r4895, %r4584, %r4850; xor.b32 %r4896, %r4592, %r4895; sub.s32 %r4897, %r4896, %r4594; xor.b32 %r4898, %r4897, %r4587; shf.l.wrap.b32 %r4899, %r4897, %r4897, 25; sub.s32 %r4900, %r4898, %r4899; xor.b32 %r4901, %r4900, %r4592; shf.l.wrap.b32 %r4902, %r4900, %r4900, 16; sub.s32 %r4903, %r4901, %r4902; xor.b32 %r4904, %r4903, %r4897; shf.l.wrap.b32 %r4905, %r4903, %r4903, 4; sub.s32 %r4906, %r4904, %r4905; xor.b32 %r4907, %r4906, %r4900; shf.l.wrap.b32 %r4908, %r4906, %r4906, 14; sub.s32 %r4909, %r4907, %r4908; xor.b32 %r4910, %r4909, %r4903; shf.l.wrap.b32 %r4911, %r4909, %r4909, 24; sub.s32 %r4912, %r4910, %r4911; and.b32 %r4913, %r4912, 31; setp.lt.u32 %p805, %r4913, 24; selp.f32 %f2718, %f2576, %f2627, %p805; setp.lt.u32 %p806, %r4913, 16; selp.f32 %f2719, %f2627, %f2580, %p806; setp.lt.u32 %p807, %r4913, 8; selp.f32 %f2720, %f2580, %f2695, %p807; and.b32 %r4914, %r4912, 1; setp.eq.b32 %p808, %r4914, 1; neg.ftz.f32 %f2721, %f2718; selp.f32 %f2722, %f2721, %f2718, %p808; and.b32 %r4915, %r4912, 2; setp.eq.s32 %p809, %r4915, 0; neg.ftz.f32 %f2723, %f2719; selp.f32 %f2724, %f2719, %f2723, %p809; add.ftz.f32 %f2725, %f2722, %f2724; and.b32 %r4916, %r4912, 4; setp.eq.s32 %p810, %r4916, 0; neg.ftz.f32 %f2726, %f2720; selp.f32 %f2727, %f2720, %f2726, %p810; add.ftz.f32 %f2728, %f2727, %f2725; add.s32 %r4917, %r4629, %r4850; xor.b32 %r4918, %r4637, %r4917; sub.s32 %r4919, %r4918, %r4639; xor.b32 %r4920, %r4919, %r4632; shf.l.wrap.b32 %r4921, %r4919, %r4919, 25; sub.s32 %r4922, %r4920, %r4921; xor.b32 %r4923, %r4922, %r4637; shf.l.wrap.b32 %r4924, %r4922, %r4922, 16; sub.s32 %r4925, %r4923, %r4924; xor.b32 %r4926, %r4925, %r4919; shf.l.wrap.b32 %r4927, %r4925, %r4925, 4; sub.s32 %r4928, %r4926, %r4927; xor.b32 %r4929, %r4928, %r4922; shf.l.wrap.b32 %r4930, %r4928, %r4928, 14; sub.s32 %r4931, %r4929, %r4930; xor.b32 %r4932, %r4931, %r4925; shf.l.wrap.b32 %r4933, %r4931, %r4931, 24; sub.s32 %r4934, %r4932, %r4933; and.b32 %r4935, %r4934, 31; setp.lt.u32 %p811, %r4935, 24; selp.f32 %f2729, %f2614, %f2627, %p811; setp.lt.u32 %p812, %r4935, 16; selp.f32 %f2730, %f2627, %f2580, %p812; setp.lt.u32 %p813, %r4935, 8; selp.f32 %f2731, %f2580, %f2695, %p813; and.b32 %r4936, %r4934, 1; setp.eq.b32 %p814, %r4936, 1; neg.ftz.f32 %f2732, %f2729; selp.f32 %f2733, %f2732, %f2729, %p814; and.b32 %r4937, %r4934, 2; setp.eq.s32 %p815, %r4937, 0; neg.ftz.f32 %f2734, %f2730; selp.f32 %f2735, %f2730, %f2734, %p815; add.ftz.f32 %f2736, %f2733, %f2735; and.b32 %r4938, %r4934, 4; setp.eq.s32 %p816, %r4938, 0; neg.ftz.f32 %f2737, %f2731; selp.f32 %f2738, %f2731, %f2737, %p816; add.ftz.f32 %f2739, %f2738, %f2736; add.s32 %r4939, %r4681, %r4850; xor.b32 %r4940, %r4689, %r4939; sub.s32 %r4941, %r4940, %r4691; xor.b32 %r4942, %r4941, %r4684; shf.l.wrap.b32 %r4943, %r4941, %r4941, 25; sub.s32 %r4944, %r4942, %r4943; xor.b32 %r4945, %r4944, %r4689; shf.l.wrap.b32 %r4946, %r4944, %r4944, 16; sub.s32 %r4947, %r4945, %r4946; xor.b32 %r4948, %r4947, %r4941; shf.l.wrap.b32 %r4949, %r4947, %r4947, 4; sub.s32 %r4950, %r4948, %r4949; xor.b32 %r4951, %r4950, %r4944; shf.l.wrap.b32 %r4952, %r4950, %r4950, 14; sub.s32 %r4953, %r4951, %r4952; xor.b32 %r4954, %r4953, %r4947; shf.l.wrap.b32 %r4955, %r4953, %r4953, 24; sub.s32 %r4956, %r4954, %r4955; and.b32 %r4957, %r4956, 31; setp.lt.u32 %p817, %r4957, 24; selp.f32 %f2740, %f2576, %f2578, %p817; setp.lt.u32 %p818, %r4957, 16; selp.f32 %f2741, %f2578, %f2650, %p818; setp.lt.u32 %p819, %r4957, 8; selp.f32 %f2742, %f2650, %f2695, %p819; and.b32 %r4958, %r4956, 1; setp.eq.b32 %p820, %r4958, 1; neg.ftz.f32 %f2743, %f2740; selp.f32 %f2744, %f2743, %f2740, %p820; and.b32 %r4959, %r4956, 2; setp.eq.s32 %p821, %r4959, 0; neg.ftz.f32 %f2745, %f2741; selp.f32 %f2746, %f2741, %f2745, %p821; add.ftz.f32 %f2747, %f2744, %f2746; and.b32 %r4960, %r4956, 4; setp.eq.s32 %p822, %r4960, 0; neg.ftz.f32 %f2748, %f2742; selp.f32 %f2749, %f2742, %f2748, %p822; add.ftz.f32 %f2750, %f2749, %f2747; add.s32 %r4961, %r4728, %r4850; xor.b32 %r4962, %r4736, %r4961; sub.s32 %r4963, %r4962, %r4738; xor.b32 %r4964, %r4963, %r4731; shf.l.wrap.b32 %r4965, %r4963, %r4963, 25; sub.s32 %r4966, %r4964, %r4965; xor.b32 %r4967, %r4966, %r4736; shf.l.wrap.b32 %r4968, %r4966, %r4966, 16; sub.s32 %r4969, %r4967, %r4968; xor.b32 %r4970, %r4969, %r4963; shf.l.wrap.b32 %r4971, %r4969, %r4969, 4; sub.s32 %r4972, %r4970, %r4971; xor.b32 %r4973, %r4972, %r4966; shf.l.wrap.b32 %r4974, %r4972, %r4972, 14; sub.s32 %r4975, %r4973, %r4974; xor.b32 %r4976, %r4975, %r4969; shf.l.wrap.b32 %r4977, %r4975, %r4975, 24; sub.s32 %r4978, %r4976, %r4977; and.b32 %r4979, %r4978, 31; setp.lt.u32 %p823, %r4979, 24; selp.f32 %f2751, %f2614, %f2578, %p823; setp.lt.u32 %p824, %r4979, 16; selp.f32 %f2752, %f2578, %f2650, %p824; setp.lt.u32 %p825, %r4979, 8; selp.f32 %f2753, %f2650, %f2695, %p825; and.b32 %r4980, %r4978, 1; setp.eq.b32 %p826, %r4980, 1; neg.ftz.f32 %f2754, %f2751; selp.f32 %f2755, %f2754, %f2751, %p826; and.b32 %r4981, %r4978, 2; setp.eq.s32 %p827, %r4981, 0; neg.ftz.f32 %f2756, %f2752; selp.f32 %f2757, %f2752, %f2756, %p827; add.ftz.f32 %f2758, %f2755, %f2757; and.b32 %r4982, %r4978, 4; setp.eq.s32 %p828, %r4982, 0; neg.ftz.f32 %f2759, %f2753; selp.f32 %f2760, %f2753, %f2759, %p828; add.ftz.f32 %f2761, %f2760, %f2758; add.s32 %r4983, %r4774, %r4850; xor.b32 %r4984, %r4782, %r4983; sub.s32 %r4985, %r4984, %r4784; xor.b32 %r4986, %r4985, %r4777; shf.l.wrap.b32 %r4987, %r4985, %r4985, 25; sub.s32 %r4988, %r4986, %r4987; xor.b32 %r4989, %r4988, %r4782; shf.l.wrap.b32 %r4990, %r4988, %r4988, 16; sub.s32 %r4991, %r4989, %r4990; xor.b32 %r4992, %r4991, %r4985; shf.l.wrap.b32 %r4993, %r4991, %r4991, 4; sub.s32 %r4994, %r4992, %r4993; xor.b32 %r4995, %r4994, %r4988; shf.l.wrap.b32 %r4996, %r4994, %r4994, 14; sub.s32 %r4997, %r4995, %r4996; xor.b32 %r4998, %r4997, %r4991; shf.l.wrap.b32 %r4999, %r4997, %r4997, 24; sub.s32 %r5000, %r4998, %r4999; and.b32 %r5001, %r5000, 31; setp.lt.u32 %p829, %r5001, 24; selp.f32 %f2762, %f2576, %f2627, %p829; setp.lt.u32 %p830, %r5001, 16; selp.f32 %f2763, %f2627, %f2650, %p830; setp.lt.u32 %p831, %r5001, 8; selp.f32 %f2764, %f2650, %f2695, %p831; and.b32 %r5002, %r5000, 1; setp.eq.b32 %p832, %r5002, 1; neg.ftz.f32 %f2765, %f2762; selp.f32 %f2766, %f2765, %f2762, %p832; and.b32 %r5003, %r5000, 2; setp.eq.s32 %p833, %r5003, 0; neg.ftz.f32 %f2767, %f2763; selp.f32 %f2768, %f2763, %f2767, %p833; add.ftz.f32 %f2769, %f2766, %f2768; and.b32 %r5004, %r5000, 4; setp.eq.s32 %p834, %r5004, 0; neg.ftz.f32 %f2770, %f2764; selp.f32 %f2771, %f2764, %f2770, %p834; add.ftz.f32 %f2772, %f2771, %f2769; add.s32 %r5005, %r4819, %r4850; xor.b32 %r5006, %r4827, %r5005; sub.s32 %r5007, %r5006, %r4829; xor.b32 %r5008, %r5007, %r4822; shf.l.wrap.b32 %r5009, %r5007, %r5007, 25; sub.s32 %r5010, %r5008, %r5009; xor.b32 %r5011, %r5010, %r4827; shf.l.wrap.b32 %r5012, %r5010, %r5010, 16; sub.s32 %r5013, %r5011, %r5012; xor.b32 %r5014, %r5013, %r5007; shf.l.wrap.b32 %r5015, %r5013, %r5013, 4; sub.s32 %r5016, %r5014, %r5015; xor.b32 %r5017, %r5016, %r5010; shf.l.wrap.b32 %r5018, %r5016, %r5016, 14; sub.s32 %r5019, %r5017, %r5018; xor.b32 %r5020, %r5019, %r5013; shf.l.wrap.b32 %r5021, %r5019, %r5019, 24; sub.s32 %r5022, %r5020, %r5021; and.b32 %r5023, %r5022, 31; setp.lt.u32 %p835, %r5023, 24; selp.f32 %f2773, %f2614, %f2627, %p835; setp.lt.u32 %p836, %r5023, 16; selp.f32 %f2774, %f2627, %f2650, %p836; setp.lt.u32 %p837, %r5023, 8; selp.f32 %f2775, %f2650, %f2695, %p837; and.b32 %r5024, %r5022, 1; setp.eq.b32 %p838, %r5024, 1; neg.ftz.f32 %f2776, %f2773; selp.f32 %f2777, %f2776, %f2773, %p838; and.b32 %r5025, %r5022, 2; setp.eq.s32 %p839, %r5025, 0; neg.ftz.f32 %f2778, %f2774; selp.f32 %f2779, %f2774, %f2778, %p839; add.ftz.f32 %f2780, %f2777, %f2779; and.b32 %r5026, %r5022, 4; setp.eq.s32 %p840, %r5026, 0; neg.ftz.f32 %f2781, %f2775; selp.f32 %f2782, %f2775, %f2781, %p840; add.ftz.f32 %f2783, %f2782, %f2780; sub.ftz.f32 %f2784, %f2569, %f2587; sub.ftz.f32 %f2785, %f2569, %f2592; sub.ftz.f32 %f2786, %f2569, %f2597; mul.ftz.f32 %f2787, %f2784, %f2613; fma.rn.ftz.f32 %f2788, %f2587, %f2626, %f2787; mul.ftz.f32 %f2789, %f2784, %f2638; fma.rn.ftz.f32 %f2790, %f2587, %f2649, %f2789; mul.ftz.f32 %f2791, %f2592, %f2790; fma.rn.ftz.f32 %f2792, %f2785, %f2788, %f2791; mul.ftz.f32 %f2793, %f2784, %f2661; fma.rn.ftz.f32 %f2794, %f2587, %f2672, %f2793; mul.ftz.f32 %f2795, %f2784, %f2683; fma.rn.ftz.f32 %f2796, %f2587, %f2694, %f2795; mul.ftz.f32 %f2797, %f2592, %f2796; fma.rn.ftz.f32 %f2798, %f2785, %f2794, %f2797; mul.ftz.f32 %f2799, %f2597, %f2798; fma.rn.ftz.f32 %f2800, %f2786, %f2792, %f2799; mul.ftz.f32 %f2801, %f2784, %f2706; fma.rn.ftz.f32 %f2802, %f2587, %f2717, %f2801; mul.ftz.f32 %f2803, %f2784, %f2728; fma.rn.ftz.f32 %f2804, %f2587, %f2739, %f2803; mul.ftz.f32 %f2805, %f2592, %f2804; fma.rn.ftz.f32 %f2806, %f2785, %f2802, %f2805; mul.ftz.f32 %f2807, %f2784, %f2750; fma.rn.ftz.f32 %f2808, %f2587, %f2761, %f2807; mul.ftz.f32 %f2809, %f2784, %f2772; fma.rn.ftz.f32 %f2810, %f2587, %f2783, %f2809; mul.ftz.f32 %f2811, %f2592, %f2810; fma.rn.ftz.f32 %f2812, %f2785, %f2808, %f2811; mul.ftz.f32 %f2813, %f2597, %f2812; fma.rn.ftz.f32 %f2814, %f2786, %f2806, %f2813; sub.ftz.f32 %f2815, %f2814, %f2800; fma.rn.ftz.f32 %f126, %f2602, %f2815, %f2800; mov.b32 %r34, %f126; setp.nan.ftz.f32 %p841, %f126, %f126; mov.f32 %f5567, %f2570; @%p841 bra $L__BB74_48; or.b32 %r5027, %r34, -2147483648; setp.ne.s32 %p842, %r5027, -2147483648; add.ftz.f32 %f2817, %f126, %f126; setp.eq.ftz.f32 %p843, %f2817, %f126; and.pred %p844, %p843, %p842; mov.f32 %f5567, %f2570; @%p844 bra $L__BB74_48; shl.b32 %r5028, %r34, 1; setp.lt.u32 %p845, %r5028, -16777215; selp.f32 %f5567, %f126, 0f00000000, %p845; $L__BB74_48: mul.ftz.f32 %f2818, %f5567, 0f3F559B3D; fma.rn.ftz.f32 %f5827, %f5573, %f2818, %f5827; mul.ftz.f32 %f5573, %f119, %f5573; mul.ftz.f32 %f5568, %f27, %f5568; mul.ftz.f32 %f5569, %f27, %f5569; mul.ftz.f32 %f5570, %f27, %f5570; mul.ftz.f32 %f5571, %f27, %f5571; add.s32 %r10021, %r10021, 1; setp.lt.s32 %p846, %r10021, %r32; @%p846 bra $L__BB74_45; $L__BB74_49: cvt.rmi.ftz.f32.f32 %f2819, %f26; sub.ftz.f32 %f141, %f26, %f2819; setp.eq.ftz.f32 %p847, %f141, 0f00000000; @%p847 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f2821, %f5568; cvt.rzi.ftz.s32.f32 %r5029, %f2821; sub.ftz.f32 %f2822, %f5568, %f2821; cvt.rmi.ftz.f32.f32 %f2823, %f5569; cvt.rzi.ftz.s32.f32 %r5030, %f2823; sub.ftz.f32 %f2824, %f5569, %f2823; cvt.rmi.ftz.f32.f32 %f2825, %f5570; cvt.rzi.ftz.s32.f32 %r5031, %f2825; sub.ftz.f32 %f2826, %f5570, %f2825; cvt.rmi.ftz.f32.f32 %f2827, %f5571; cvt.rzi.ftz.s32.f32 %r5032, %f2827; sub.ftz.f32 %f2828, %f5571, %f2827; mul.ftz.f32 %f2829, %f2822, %f2822; mul.ftz.f32 %f2830, %f2822, %f2829; fma.rn.ftz.f32 %f2831, %f2822, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2832, %f2822, %f2831, 0f41200000; mul.ftz.f32 %f2833, %f2830, %f2832; mul.ftz.f32 %f2834, %f2824, %f2824; mul.ftz.f32 %f2835, %f2824, %f2834; fma.rn.ftz.f32 %f2836, %f2824, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2837, %f2824, %f2836, 0f41200000; mul.ftz.f32 %f2838, %f2835, %f2837; mul.ftz.f32 %f2839, %f2826, %f2826; mul.ftz.f32 %f2840, %f2826, %f2839; fma.rn.ftz.f32 %f2841, %f2826, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2842, %f2826, %f2841, 0f41200000; mul.ftz.f32 %f2843, %f2840, %f2842; mul.ftz.f32 %f2844, %f2828, %f2828; mul.ftz.f32 %f2845, %f2828, %f2844; fma.rn.ftz.f32 %f2846, %f2828, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2847, %f2828, %f2846, 0f41200000; mul.ftz.f32 %f2848, %f2845, %f2847; add.s32 %r5033, %r5030, -559038708; add.s32 %r5034, %r5031, -559038708; sub.s32 %r5035, %r5029, %r5031; shf.l.wrap.b32 %r5036, %r5034, %r5034, 4; xor.b32 %r5037, %r5036, %r5035; add.s32 %r5038, %r5034, %r5033; sub.s32 %r5039, %r5033, %r5037; shf.l.wrap.b32 %r5040, %r5037, %r5037, 6; xor.b32 %r5041, %r5040, %r5039; add.s32 %r5042, %r5037, %r5038; sub.s32 %r5043, %r5038, %r5041; shf.l.wrap.b32 %r5044, %r5041, %r5041, 8; xor.b32 %r5045, %r5044, %r5043; add.s32 %r5046, %r5041, %r5042; sub.s32 %r5047, %r5042, %r5045; shf.l.wrap.b32 %r5048, %r5045, %r5045, 16; xor.b32 %r5049, %r5048, %r5047; add.s32 %r5050, %r5045, %r5046; sub.s32 %r5051, %r5046, %r5049; shf.l.wrap.b32 %r5052, %r5049, %r5049, 19; xor.b32 %r5053, %r5052, %r5051; add.s32 %r5054, %r5049, %r5050; sub.s32 %r5055, %r5050, %r5053; shf.l.wrap.b32 %r5056, %r5053, %r5053, 4; add.s32 %r5057, %r5053, %r5054; add.s32 %r5058, %r5054, %r5032; xor.b32 %r5059, %r5057, %r5055; xor.b32 %r5060, %r5059, %r5056; shf.l.wrap.b32 %r5061, %r5057, %r5057, 14; sub.s32 %r5062, %r5060, %r5061; xor.b32 %r5063, %r5062, %r5058; shf.l.wrap.b32 %r5064, %r5062, %r5062, 11; sub.s32 %r5065, %r5063, %r5064; xor.b32 %r5066, %r5065, %r5057; shf.l.wrap.b32 %r5067, %r5065, %r5065, 25; sub.s32 %r5068, %r5066, %r5067; xor.b32 %r5069, %r5068, %r5062; shf.l.wrap.b32 %r5070, %r5068, %r5068, 16; sub.s32 %r5071, %r5069, %r5070; xor.b32 %r5072, %r5071, %r5065; shf.l.wrap.b32 %r5073, %r5071, %r5071, 4; sub.s32 %r5074, %r5072, %r5073; xor.b32 %r5075, %r5074, %r5068; shf.l.wrap.b32 %r5076, %r5074, %r5074, 14; sub.s32 %r5077, %r5075, %r5076; xor.b32 %r5078, %r5077, %r5071; shf.l.wrap.b32 %r5079, %r5077, %r5077, 24; sub.s32 %r5080, %r5078, %r5079; and.b32 %r5081, %r5080, 31; setp.lt.u32 %p848, %r5081, 24; selp.f32 %f2849, %f2822, %f2824, %p848; setp.lt.u32 %p849, %r5081, 16; selp.f32 %f2850, %f2824, %f2826, %p849; setp.lt.u32 %p850, %r5081, 8; selp.f32 %f2851, %f2826, %f2828, %p850; and.b32 %r5082, %r5080, 1; setp.eq.b32 %p851, %r5082, 1; neg.ftz.f32 %f2852, %f2849; selp.f32 %f2853, %f2852, %f2849, %p851; and.b32 %r5083, %r5080, 2; setp.eq.s32 %p852, %r5083, 0; neg.ftz.f32 %f2854, %f2850; selp.f32 %f2855, %f2850, %f2854, %p852; add.ftz.f32 %f2856, %f2853, %f2855; and.b32 %r5084, %r5080, 4; setp.eq.s32 %p853, %r5084, 0; neg.ftz.f32 %f2857, %f2851; selp.f32 %f2858, %f2851, %f2857, %p853; add.ftz.f32 %f2859, %f2858, %f2856; add.s32 %r5085, %r5029, 1; sub.s32 %r5086, %r5085, %r5031; xor.b32 %r5087, %r5036, %r5086; sub.s32 %r5088, %r5033, %r5087; shf.l.wrap.b32 %r5089, %r5087, %r5087, 6; xor.b32 %r5090, %r5089, %r5088; add.s32 %r5091, %r5087, %r5038; sub.s32 %r5092, %r5038, %r5090; shf.l.wrap.b32 %r5093, %r5090, %r5090, 8; xor.b32 %r5094, %r5093, %r5092; add.s32 %r5095, %r5090, %r5091; sub.s32 %r5096, %r5091, %r5094; shf.l.wrap.b32 %r5097, %r5094, %r5094, 16; xor.b32 %r5098, %r5097, %r5096; add.s32 %r5099, %r5094, %r5095; sub.s32 %r5100, %r5095, %r5098; shf.l.wrap.b32 %r5101, %r5098, %r5098, 19; xor.b32 %r5102, %r5101, %r5100; add.s32 %r5103, %r5098, %r5099; sub.s32 %r5104, %r5099, %r5102; shf.l.wrap.b32 %r5105, %r5102, %r5102, 4; add.s32 %r5106, %r5102, %r5103; add.s32 %r5107, %r5103, %r5032; xor.b32 %r5108, %r5106, %r5104; xor.b32 %r5109, %r5108, %r5105; shf.l.wrap.b32 %r5110, %r5106, %r5106, 14; sub.s32 %r5111, %r5109, %r5110; xor.b32 %r5112, %r5111, %r5107; shf.l.wrap.b32 %r5113, %r5111, %r5111, 11; sub.s32 %r5114, %r5112, %r5113; xor.b32 %r5115, %r5114, %r5106; shf.l.wrap.b32 %r5116, %r5114, %r5114, 25; sub.s32 %r5117, %r5115, %r5116; xor.b32 %r5118, %r5117, %r5111; shf.l.wrap.b32 %r5119, %r5117, %r5117, 16; sub.s32 %r5120, %r5118, %r5119; xor.b32 %r5121, %r5120, %r5114; shf.l.wrap.b32 %r5122, %r5120, %r5120, 4; sub.s32 %r5123, %r5121, %r5122; xor.b32 %r5124, %r5123, %r5117; shf.l.wrap.b32 %r5125, %r5123, %r5123, 14; sub.s32 %r5126, %r5124, %r5125; xor.b32 %r5127, %r5126, %r5120; shf.l.wrap.b32 %r5128, %r5126, %r5126, 24; sub.s32 %r5129, %r5127, %r5128; add.ftz.f32 %f2860, %f2822, 0fBF800000; mov.f32 %f2861, 0f3F800000; and.b32 %r5130, %r5129, 31; setp.lt.u32 %p854, %r5130, 24; selp.f32 %f2862, %f2860, %f2824, %p854; setp.lt.u32 %p855, %r5130, 16; selp.f32 %f2863, %f2824, %f2826, %p855; setp.lt.u32 %p856, %r5130, 8; selp.f32 %f2864, %f2826, %f2828, %p856; and.b32 %r5131, %r5129, 1; setp.eq.b32 %p857, %r5131, 1; neg.ftz.f32 %f2865, %f2862; selp.f32 %f2866, %f2865, %f2862, %p857; and.b32 %r5132, %r5129, 2; setp.eq.s32 %p858, %r5132, 0; neg.ftz.f32 %f2867, %f2863; selp.f32 %f2868, %f2863, %f2867, %p858; add.ftz.f32 %f2869, %f2866, %f2868; and.b32 %r5133, %r5129, 4; setp.eq.s32 %p859, %r5133, 0; neg.ftz.f32 %f2870, %f2864; selp.f32 %f2871, %f2864, %f2870, %p859; add.ftz.f32 %f2872, %f2871, %f2869; add.s32 %r5134, %r5030, -559038707; add.s32 %r5135, %r5034, %r5134; sub.s32 %r5136, %r5134, %r5037; xor.b32 %r5137, %r5040, %r5136; add.s32 %r5138, %r5037, %r5135; sub.s32 %r5139, %r5135, %r5137; shf.l.wrap.b32 %r5140, %r5137, %r5137, 8; xor.b32 %r5141, %r5140, %r5139; add.s32 %r5142, %r5137, %r5138; sub.s32 %r5143, %r5138, %r5141; shf.l.wrap.b32 %r5144, %r5141, %r5141, 16; xor.b32 %r5145, %r5144, %r5143; add.s32 %r5146, %r5141, %r5142; sub.s32 %r5147, %r5142, %r5145; shf.l.wrap.b32 %r5148, %r5145, %r5145, 19; xor.b32 %r5149, %r5148, %r5147; add.s32 %r5150, %r5145, %r5146; sub.s32 %r5151, %r5146, %r5149; shf.l.wrap.b32 %r5152, %r5149, %r5149, 4; add.s32 %r5153, %r5149, %r5150; add.s32 %r5154, %r5150, %r5032; xor.b32 %r5155, %r5153, %r5151; xor.b32 %r5156, %r5155, %r5152; shf.l.wrap.b32 %r5157, %r5153, %r5153, 14; sub.s32 %r5158, %r5156, %r5157; xor.b32 %r5159, %r5158, %r5154; shf.l.wrap.b32 %r5160, %r5158, %r5158, 11; sub.s32 %r5161, %r5159, %r5160; xor.b32 %r5162, %r5161, %r5153; shf.l.wrap.b32 %r5163, %r5161, %r5161, 25; sub.s32 %r5164, %r5162, %r5163; xor.b32 %r5165, %r5164, %r5158; shf.l.wrap.b32 %r5166, %r5164, %r5164, 16; sub.s32 %r5167, %r5165, %r5166; xor.b32 %r5168, %r5167, %r5161; shf.l.wrap.b32 %r5169, %r5167, %r5167, 4; sub.s32 %r5170, %r5168, %r5169; xor.b32 %r5171, %r5170, %r5164; shf.l.wrap.b32 %r5172, %r5170, %r5170, 14; sub.s32 %r5173, %r5171, %r5172; xor.b32 %r5174, %r5173, %r5167; shf.l.wrap.b32 %r5175, %r5173, %r5173, 24; sub.s32 %r5176, %r5174, %r5175; add.ftz.f32 %f2873, %f2824, 0fBF800000; and.b32 %r5177, %r5176, 31; setp.lt.u32 %p860, %r5177, 24; selp.f32 %f2874, %f2822, %f2873, %p860; setp.lt.u32 %p861, %r5177, 16; selp.f32 %f2875, %f2873, %f2826, %p861; setp.lt.u32 %p862, %r5177, 8; selp.f32 %f2876, %f2826, %f2828, %p862; and.b32 %r5178, %r5176, 1; setp.eq.b32 %p863, %r5178, 1; neg.ftz.f32 %f2877, %f2874; selp.f32 %f2878, %f2877, %f2874, %p863; and.b32 %r5179, %r5176, 2; setp.eq.s32 %p864, %r5179, 0; neg.ftz.f32 %f2879, %f2875; selp.f32 %f2880, %f2875, %f2879, %p864; add.ftz.f32 %f2881, %f2878, %f2880; and.b32 %r5180, %r5176, 4; setp.eq.s32 %p865, %r5180, 0; neg.ftz.f32 %f2882, %f2876; selp.f32 %f2883, %f2876, %f2882, %p865; add.ftz.f32 %f2884, %f2883, %f2881; sub.s32 %r5181, %r5134, %r5087; xor.b32 %r5182, %r5089, %r5181; add.s32 %r5183, %r5087, %r5135; sub.s32 %r5184, %r5135, %r5182; shf.l.wrap.b32 %r5185, %r5182, %r5182, 8; xor.b32 %r5186, %r5185, %r5184; add.s32 %r5187, %r5182, %r5183; sub.s32 %r5188, %r5183, %r5186; shf.l.wrap.b32 %r5189, %r5186, %r5186, 16; xor.b32 %r5190, %r5189, %r5188; add.s32 %r5191, %r5186, %r5187; sub.s32 %r5192, %r5187, %r5190; shf.l.wrap.b32 %r5193, %r5190, %r5190, 19; xor.b32 %r5194, %r5193, %r5192; add.s32 %r5195, %r5190, %r5191; sub.s32 %r5196, %r5191, %r5194; shf.l.wrap.b32 %r5197, %r5194, %r5194, 4; add.s32 %r5198, %r5194, %r5195; add.s32 %r5199, %r5195, %r5032; xor.b32 %r5200, %r5198, %r5196; xor.b32 %r5201, %r5200, %r5197; shf.l.wrap.b32 %r5202, %r5198, %r5198, 14; sub.s32 %r5203, %r5201, %r5202; xor.b32 %r5204, %r5203, %r5199; shf.l.wrap.b32 %r5205, %r5203, %r5203, 11; sub.s32 %r5206, %r5204, %r5205; xor.b32 %r5207, %r5206, %r5198; shf.l.wrap.b32 %r5208, %r5206, %r5206, 25; sub.s32 %r5209, %r5207, %r5208; xor.b32 %r5210, %r5209, %r5203; shf.l.wrap.b32 %r5211, %r5209, %r5209, 16; sub.s32 %r5212, %r5210, %r5211; xor.b32 %r5213, %r5212, %r5206; shf.l.wrap.b32 %r5214, %r5212, %r5212, 4; sub.s32 %r5215, %r5213, %r5214; xor.b32 %r5216, %r5215, %r5209; shf.l.wrap.b32 %r5217, %r5215, %r5215, 14; sub.s32 %r5218, %r5216, %r5217; xor.b32 %r5219, %r5218, %r5212; shf.l.wrap.b32 %r5220, %r5218, %r5218, 24; sub.s32 %r5221, %r5219, %r5220; and.b32 %r5222, %r5221, 31; setp.lt.u32 %p866, %r5222, 24; selp.f32 %f2885, %f2860, %f2873, %p866; setp.lt.u32 %p867, %r5222, 16; selp.f32 %f2886, %f2873, %f2826, %p867; setp.lt.u32 %p868, %r5222, 8; selp.f32 %f2887, %f2826, %f2828, %p868; and.b32 %r5223, %r5221, 1; setp.eq.b32 %p869, %r5223, 1; neg.ftz.f32 %f2888, %f2885; selp.f32 %f2889, %f2888, %f2885, %p869; and.b32 %r5224, %r5221, 2; setp.eq.s32 %p870, %r5224, 0; neg.ftz.f32 %f2890, %f2886; selp.f32 %f2891, %f2886, %f2890, %p870; add.ftz.f32 %f2892, %f2889, %f2891; and.b32 %r5225, %r5221, 4; setp.eq.s32 %p871, %r5225, 0; neg.ftz.f32 %f2893, %f2887; selp.f32 %f2894, %f2887, %f2893, %p871; add.ftz.f32 %f2895, %f2894, %f2892; not.b32 %r5226, %r5031; add.s32 %r5227, %r5031, -559038707; add.s32 %r5228, %r5029, %r5226; shf.l.wrap.b32 %r5229, %r5227, %r5227, 4; xor.b32 %r5230, %r5229, %r5228; add.s32 %r5231, %r5227, %r5033; sub.s32 %r5232, %r5033, %r5230; shf.l.wrap.b32 %r5233, %r5230, %r5230, 6; xor.b32 %r5234, %r5233, %r5232; add.s32 %r5235, %r5230, %r5231; sub.s32 %r5236, %r5231, %r5234; shf.l.wrap.b32 %r5237, %r5234, %r5234, 8; xor.b32 %r5238, %r5237, %r5236; add.s32 %r5239, %r5234, %r5235; sub.s32 %r5240, %r5235, %r5238; shf.l.wrap.b32 %r5241, %r5238, %r5238, 16; xor.b32 %r5242, %r5241, %r5240; add.s32 %r5243, %r5238, %r5239; sub.s32 %r5244, %r5239, %r5242; shf.l.wrap.b32 %r5245, %r5242, %r5242, 19; xor.b32 %r5246, %r5245, %r5244; add.s32 %r5247, %r5242, %r5243; sub.s32 %r5248, %r5243, %r5246; shf.l.wrap.b32 %r5249, %r5246, %r5246, 4; add.s32 %r5250, %r5246, %r5247; add.s32 %r5251, %r5247, %r5032; xor.b32 %r5252, %r5250, %r5248; xor.b32 %r5253, %r5252, %r5249; shf.l.wrap.b32 %r5254, %r5250, %r5250, 14; sub.s32 %r5255, %r5253, %r5254; xor.b32 %r5256, %r5255, %r5251; shf.l.wrap.b32 %r5257, %r5255, %r5255, 11; sub.s32 %r5258, %r5256, %r5257; xor.b32 %r5259, %r5258, %r5250; shf.l.wrap.b32 %r5260, %r5258, %r5258, 25; sub.s32 %r5261, %r5259, %r5260; xor.b32 %r5262, %r5261, %r5255; shf.l.wrap.b32 %r5263, %r5261, %r5261, 16; sub.s32 %r5264, %r5262, %r5263; xor.b32 %r5265, %r5264, %r5258; shf.l.wrap.b32 %r5266, %r5264, %r5264, 4; sub.s32 %r5267, %r5265, %r5266; xor.b32 %r5268, %r5267, %r5261; shf.l.wrap.b32 %r5269, %r5267, %r5267, 14; sub.s32 %r5270, %r5268, %r5269; xor.b32 %r5271, %r5270, %r5264; shf.l.wrap.b32 %r5272, %r5270, %r5270, 24; sub.s32 %r5273, %r5271, %r5272; add.ftz.f32 %f2896, %f2826, 0fBF800000; and.b32 %r5274, %r5273, 31; setp.lt.u32 %p872, %r5274, 24; selp.f32 %f2897, %f2822, %f2824, %p872; setp.lt.u32 %p873, %r5274, 16; selp.f32 %f2898, %f2824, %f2896, %p873; setp.lt.u32 %p874, %r5274, 8; selp.f32 %f2899, %f2896, %f2828, %p874; and.b32 %r5275, %r5273, 1; setp.eq.b32 %p875, %r5275, 1; neg.ftz.f32 %f2900, %f2897; selp.f32 %f2901, %f2900, %f2897, %p875; and.b32 %r5276, %r5273, 2; setp.eq.s32 %p876, %r5276, 0; neg.ftz.f32 %f2902, %f2898; selp.f32 %f2903, %f2898, %f2902, %p876; add.ftz.f32 %f2904, %f2901, %f2903; and.b32 %r5277, %r5273, 4; setp.eq.s32 %p877, %r5277, 0; neg.ftz.f32 %f2905, %f2899; selp.f32 %f2906, %f2899, %f2905, %p877; add.ftz.f32 %f2907, %f2906, %f2904; xor.b32 %r5278, %r5229, %r5035; sub.s32 %r5279, %r5033, %r5278; shf.l.wrap.b32 %r5280, %r5278, %r5278, 6; xor.b32 %r5281, %r5280, %r5279; add.s32 %r5282, %r5278, %r5231; sub.s32 %r5283, %r5231, %r5281; shf.l.wrap.b32 %r5284, %r5281, %r5281, 8; xor.b32 %r5285, %r5284, %r5283; add.s32 %r5286, %r5281, %r5282; sub.s32 %r5287, %r5282, %r5285; shf.l.wrap.b32 %r5288, %r5285, %r5285, 16; xor.b32 %r5289, %r5288, %r5287; add.s32 %r5290, %r5285, %r5286; sub.s32 %r5291, %r5286, %r5289; shf.l.wrap.b32 %r5292, %r5289, %r5289, 19; xor.b32 %r5293, %r5292, %r5291; add.s32 %r5294, %r5289, %r5290; sub.s32 %r5295, %r5290, %r5293; shf.l.wrap.b32 %r5296, %r5293, %r5293, 4; add.s32 %r5297, %r5293, %r5294; add.s32 %r5298, %r5294, %r5032; xor.b32 %r5299, %r5297, %r5295; xor.b32 %r5300, %r5299, %r5296; shf.l.wrap.b32 %r5301, %r5297, %r5297, 14; sub.s32 %r5302, %r5300, %r5301; xor.b32 %r5303, %r5302, %r5298; shf.l.wrap.b32 %r5304, %r5302, %r5302, 11; sub.s32 %r5305, %r5303, %r5304; xor.b32 %r5306, %r5305, %r5297; shf.l.wrap.b32 %r5307, %r5305, %r5305, 25; sub.s32 %r5308, %r5306, %r5307; xor.b32 %r5309, %r5308, %r5302; shf.l.wrap.b32 %r5310, %r5308, %r5308, 16; sub.s32 %r5311, %r5309, %r5310; xor.b32 %r5312, %r5311, %r5305; shf.l.wrap.b32 %r5313, %r5311, %r5311, 4; sub.s32 %r5314, %r5312, %r5313; xor.b32 %r5315, %r5314, %r5308; shf.l.wrap.b32 %r5316, %r5314, %r5314, 14; sub.s32 %r5317, %r5315, %r5316; xor.b32 %r5318, %r5317, %r5311; shf.l.wrap.b32 %r5319, %r5317, %r5317, 24; sub.s32 %r5320, %r5318, %r5319; and.b32 %r5321, %r5320, 31; setp.lt.u32 %p878, %r5321, 24; selp.f32 %f2908, %f2860, %f2824, %p878; setp.lt.u32 %p879, %r5321, 16; selp.f32 %f2909, %f2824, %f2896, %p879; setp.lt.u32 %p880, %r5321, 8; selp.f32 %f2910, %f2896, %f2828, %p880; and.b32 %r5322, %r5320, 1; setp.eq.b32 %p881, %r5322, 1; neg.ftz.f32 %f2911, %f2908; selp.f32 %f2912, %f2911, %f2908, %p881; and.b32 %r5323, %r5320, 2; setp.eq.s32 %p882, %r5323, 0; neg.ftz.f32 %f2913, %f2909; selp.f32 %f2914, %f2909, %f2913, %p882; add.ftz.f32 %f2915, %f2912, %f2914; and.b32 %r5324, %r5320, 4; setp.eq.s32 %p883, %r5324, 0; neg.ftz.f32 %f2916, %f2910; selp.f32 %f2917, %f2910, %f2916, %p883; add.ftz.f32 %f2918, %f2917, %f2915; add.s32 %r5325, %r5227, %r5134; sub.s32 %r5326, %r5134, %r5230; xor.b32 %r5327, %r5233, %r5326; add.s32 %r5328, %r5230, %r5325; sub.s32 %r5329, %r5325, %r5327; shf.l.wrap.b32 %r5330, %r5327, %r5327, 8; xor.b32 %r5331, %r5330, %r5329; add.s32 %r5332, %r5327, %r5328; sub.s32 %r5333, %r5328, %r5331; shf.l.wrap.b32 %r5334, %r5331, %r5331, 16; xor.b32 %r5335, %r5334, %r5333; add.s32 %r5336, %r5331, %r5332; sub.s32 %r5337, %r5332, %r5335; shf.l.wrap.b32 %r5338, %r5335, %r5335, 19; xor.b32 %r5339, %r5338, %r5337; add.s32 %r5340, %r5335, %r5336; sub.s32 %r5341, %r5336, %r5339; shf.l.wrap.b32 %r5342, %r5339, %r5339, 4; add.s32 %r5343, %r5339, %r5340; add.s32 %r5344, %r5340, %r5032; xor.b32 %r5345, %r5343, %r5341; xor.b32 %r5346, %r5345, %r5342; shf.l.wrap.b32 %r5347, %r5343, %r5343, 14; sub.s32 %r5348, %r5346, %r5347; xor.b32 %r5349, %r5348, %r5344; shf.l.wrap.b32 %r5350, %r5348, %r5348, 11; sub.s32 %r5351, %r5349, %r5350; xor.b32 %r5352, %r5351, %r5343; shf.l.wrap.b32 %r5353, %r5351, %r5351, 25; sub.s32 %r5354, %r5352, %r5353; xor.b32 %r5355, %r5354, %r5348; shf.l.wrap.b32 %r5356, %r5354, %r5354, 16; sub.s32 %r5357, %r5355, %r5356; xor.b32 %r5358, %r5357, %r5351; shf.l.wrap.b32 %r5359, %r5357, %r5357, 4; sub.s32 %r5360, %r5358, %r5359; xor.b32 %r5361, %r5360, %r5354; shf.l.wrap.b32 %r5362, %r5360, %r5360, 14; sub.s32 %r5363, %r5361, %r5362; xor.b32 %r5364, %r5363, %r5357; shf.l.wrap.b32 %r5365, %r5363, %r5363, 24; sub.s32 %r5366, %r5364, %r5365; and.b32 %r5367, %r5366, 31; setp.lt.u32 %p884, %r5367, 24; selp.f32 %f2919, %f2822, %f2873, %p884; setp.lt.u32 %p885, %r5367, 16; selp.f32 %f2920, %f2873, %f2896, %p885; setp.lt.u32 %p886, %r5367, 8; selp.f32 %f2921, %f2896, %f2828, %p886; and.b32 %r5368, %r5366, 1; setp.eq.b32 %p887, %r5368, 1; neg.ftz.f32 %f2922, %f2919; selp.f32 %f2923, %f2922, %f2919, %p887; and.b32 %r5369, %r5366, 2; setp.eq.s32 %p888, %r5369, 0; neg.ftz.f32 %f2924, %f2920; selp.f32 %f2925, %f2920, %f2924, %p888; add.ftz.f32 %f2926, %f2923, %f2925; and.b32 %r5370, %r5366, 4; setp.eq.s32 %p889, %r5370, 0; neg.ftz.f32 %f2927, %f2921; selp.f32 %f2928, %f2921, %f2927, %p889; add.ftz.f32 %f2929, %f2928, %f2926; sub.s32 %r5371, %r5134, %r5278; xor.b32 %r5372, %r5280, %r5371; add.s32 %r5373, %r5278, %r5325; sub.s32 %r5374, %r5325, %r5372; shf.l.wrap.b32 %r5375, %r5372, %r5372, 8; xor.b32 %r5376, %r5375, %r5374; add.s32 %r5377, %r5372, %r5373; sub.s32 %r5378, %r5373, %r5376; shf.l.wrap.b32 %r5379, %r5376, %r5376, 16; xor.b32 %r5380, %r5379, %r5378; add.s32 %r5381, %r5376, %r5377; sub.s32 %r5382, %r5377, %r5380; shf.l.wrap.b32 %r5383, %r5380, %r5380, 19; xor.b32 %r5384, %r5383, %r5382; add.s32 %r5385, %r5380, %r5381; sub.s32 %r5386, %r5381, %r5384; shf.l.wrap.b32 %r5387, %r5384, %r5384, 4; add.s32 %r5388, %r5384, %r5385; add.s32 %r5389, %r5385, %r5032; xor.b32 %r5390, %r5388, %r5386; xor.b32 %r5391, %r5390, %r5387; shf.l.wrap.b32 %r5392, %r5388, %r5388, 14; sub.s32 %r5393, %r5391, %r5392; xor.b32 %r5394, %r5393, %r5389; shf.l.wrap.b32 %r5395, %r5393, %r5393, 11; sub.s32 %r5396, %r5394, %r5395; xor.b32 %r5397, %r5396, %r5388; shf.l.wrap.b32 %r5398, %r5396, %r5396, 25; sub.s32 %r5399, %r5397, %r5398; xor.b32 %r5400, %r5399, %r5393; shf.l.wrap.b32 %r5401, %r5399, %r5399, 16; sub.s32 %r5402, %r5400, %r5401; xor.b32 %r5403, %r5402, %r5396; shf.l.wrap.b32 %r5404, %r5402, %r5402, 4; sub.s32 %r5405, %r5403, %r5404; xor.b32 %r5406, %r5405, %r5399; shf.l.wrap.b32 %r5407, %r5405, %r5405, 14; sub.s32 %r5408, %r5406, %r5407; xor.b32 %r5409, %r5408, %r5402; shf.l.wrap.b32 %r5410, %r5408, %r5408, 24; sub.s32 %r5411, %r5409, %r5410; and.b32 %r5412, %r5411, 31; setp.lt.u32 %p890, %r5412, 24; selp.f32 %f2930, %f2860, %f2873, %p890; setp.lt.u32 %p891, %r5412, 16; selp.f32 %f2931, %f2873, %f2896, %p891; setp.lt.u32 %p892, %r5412, 8; selp.f32 %f2932, %f2896, %f2828, %p892; and.b32 %r5413, %r5411, 1; setp.eq.b32 %p893, %r5413, 1; neg.ftz.f32 %f2933, %f2930; selp.f32 %f2934, %f2933, %f2930, %p893; and.b32 %r5414, %r5411, 2; setp.eq.s32 %p894, %r5414, 0; neg.ftz.f32 %f2935, %f2931; selp.f32 %f2936, %f2931, %f2935, %p894; add.ftz.f32 %f2937, %f2934, %f2936; and.b32 %r5415, %r5411, 4; setp.eq.s32 %p895, %r5415, 0; neg.ftz.f32 %f2938, %f2932; selp.f32 %f2939, %f2932, %f2938, %p895; add.ftz.f32 %f2940, %f2939, %f2937; add.s32 %r5416, %r5032, 1; add.s32 %r5417, %r5054, %r5416; xor.b32 %r5418, %r5062, %r5417; sub.s32 %r5419, %r5418, %r5064; xor.b32 %r5420, %r5419, %r5057; shf.l.wrap.b32 %r5421, %r5419, %r5419, 25; sub.s32 %r5422, %r5420, %r5421; xor.b32 %r5423, %r5422, %r5062; shf.l.wrap.b32 %r5424, %r5422, %r5422, 16; sub.s32 %r5425, %r5423, %r5424; xor.b32 %r5426, %r5425, %r5419; shf.l.wrap.b32 %r5427, %r5425, %r5425, 4; sub.s32 %r5428, %r5426, %r5427; xor.b32 %r5429, %r5428, %r5422; shf.l.wrap.b32 %r5430, %r5428, %r5428, 14; sub.s32 %r5431, %r5429, %r5430; xor.b32 %r5432, %r5431, %r5425; shf.l.wrap.b32 %r5433, %r5431, %r5431, 24; sub.s32 %r5434, %r5432, %r5433; add.ftz.f32 %f2941, %f2828, 0fBF800000; and.b32 %r5435, %r5434, 31; setp.lt.u32 %p896, %r5435, 24; selp.f32 %f2942, %f2822, %f2824, %p896; setp.lt.u32 %p897, %r5435, 16; selp.f32 %f2943, %f2824, %f2826, %p897; setp.lt.u32 %p898, %r5435, 8; selp.f32 %f2944, %f2826, %f2941, %p898; and.b32 %r5436, %r5434, 1; setp.eq.b32 %p899, %r5436, 1; neg.ftz.f32 %f2945, %f2942; selp.f32 %f2946, %f2945, %f2942, %p899; and.b32 %r5437, %r5434, 2; setp.eq.s32 %p900, %r5437, 0; neg.ftz.f32 %f2947, %f2943; selp.f32 %f2948, %f2943, %f2947, %p900; add.ftz.f32 %f2949, %f2946, %f2948; and.b32 %r5438, %r5434, 4; setp.eq.s32 %p901, %r5438, 0; neg.ftz.f32 %f2950, %f2944; selp.f32 %f2951, %f2944, %f2950, %p901; add.ftz.f32 %f2952, %f2951, %f2949; add.s32 %r5439, %r5103, %r5416; xor.b32 %r5440, %r5111, %r5439; sub.s32 %r5441, %r5440, %r5113; xor.b32 %r5442, %r5441, %r5106; shf.l.wrap.b32 %r5443, %r5441, %r5441, 25; sub.s32 %r5444, %r5442, %r5443; xor.b32 %r5445, %r5444, %r5111; shf.l.wrap.b32 %r5446, %r5444, %r5444, 16; sub.s32 %r5447, %r5445, %r5446; xor.b32 %r5448, %r5447, %r5441; shf.l.wrap.b32 %r5449, %r5447, %r5447, 4; sub.s32 %r5450, %r5448, %r5449; xor.b32 %r5451, %r5450, %r5444; shf.l.wrap.b32 %r5452, %r5450, %r5450, 14; sub.s32 %r5453, %r5451, %r5452; xor.b32 %r5454, %r5453, %r5447; shf.l.wrap.b32 %r5455, %r5453, %r5453, 24; sub.s32 %r5456, %r5454, %r5455; and.b32 %r5457, %r5456, 31; setp.lt.u32 %p902, %r5457, 24; selp.f32 %f2953, %f2860, %f2824, %p902; setp.lt.u32 %p903, %r5457, 16; selp.f32 %f2954, %f2824, %f2826, %p903; setp.lt.u32 %p904, %r5457, 8; selp.f32 %f2955, %f2826, %f2941, %p904; and.b32 %r5458, %r5456, 1; setp.eq.b32 %p905, %r5458, 1; neg.ftz.f32 %f2956, %f2953; selp.f32 %f2957, %f2956, %f2953, %p905; and.b32 %r5459, %r5456, 2; setp.eq.s32 %p906, %r5459, 0; neg.ftz.f32 %f2958, %f2954; selp.f32 %f2959, %f2954, %f2958, %p906; add.ftz.f32 %f2960, %f2957, %f2959; and.b32 %r5460, %r5456, 4; setp.eq.s32 %p907, %r5460, 0; neg.ftz.f32 %f2961, %f2955; selp.f32 %f2962, %f2955, %f2961, %p907; add.ftz.f32 %f2963, %f2962, %f2960; add.s32 %r5461, %r5150, %r5416; xor.b32 %r5462, %r5158, %r5461; sub.s32 %r5463, %r5462, %r5160; xor.b32 %r5464, %r5463, %r5153; shf.l.wrap.b32 %r5465, %r5463, %r5463, 25; sub.s32 %r5466, %r5464, %r5465; xor.b32 %r5467, %r5466, %r5158; shf.l.wrap.b32 %r5468, %r5466, %r5466, 16; sub.s32 %r5469, %r5467, %r5468; xor.b32 %r5470, %r5469, %r5463; shf.l.wrap.b32 %r5471, %r5469, %r5469, 4; sub.s32 %r5472, %r5470, %r5471; xor.b32 %r5473, %r5472, %r5466; shf.l.wrap.b32 %r5474, %r5472, %r5472, 14; sub.s32 %r5475, %r5473, %r5474; xor.b32 %r5476, %r5475, %r5469; shf.l.wrap.b32 %r5477, %r5475, %r5475, 24; sub.s32 %r5478, %r5476, %r5477; and.b32 %r5479, %r5478, 31; setp.lt.u32 %p908, %r5479, 24; selp.f32 %f2964, %f2822, %f2873, %p908; setp.lt.u32 %p909, %r5479, 16; selp.f32 %f2965, %f2873, %f2826, %p909; setp.lt.u32 %p910, %r5479, 8; selp.f32 %f2966, %f2826, %f2941, %p910; and.b32 %r5480, %r5478, 1; setp.eq.b32 %p911, %r5480, 1; neg.ftz.f32 %f2967, %f2964; selp.f32 %f2968, %f2967, %f2964, %p911; and.b32 %r5481, %r5478, 2; setp.eq.s32 %p912, %r5481, 0; neg.ftz.f32 %f2969, %f2965; selp.f32 %f2970, %f2965, %f2969, %p912; add.ftz.f32 %f2971, %f2968, %f2970; and.b32 %r5482, %r5478, 4; setp.eq.s32 %p913, %r5482, 0; neg.ftz.f32 %f2972, %f2966; selp.f32 %f2973, %f2966, %f2972, %p913; add.ftz.f32 %f2974, %f2973, %f2971; add.s32 %r5483, %r5195, %r5416; xor.b32 %r5484, %r5203, %r5483; sub.s32 %r5485, %r5484, %r5205; xor.b32 %r5486, %r5485, %r5198; shf.l.wrap.b32 %r5487, %r5485, %r5485, 25; sub.s32 %r5488, %r5486, %r5487; xor.b32 %r5489, %r5488, %r5203; shf.l.wrap.b32 %r5490, %r5488, %r5488, 16; sub.s32 %r5491, %r5489, %r5490; xor.b32 %r5492, %r5491, %r5485; shf.l.wrap.b32 %r5493, %r5491, %r5491, 4; sub.s32 %r5494, %r5492, %r5493; xor.b32 %r5495, %r5494, %r5488; shf.l.wrap.b32 %r5496, %r5494, %r5494, 14; sub.s32 %r5497, %r5495, %r5496; xor.b32 %r5498, %r5497, %r5491; shf.l.wrap.b32 %r5499, %r5497, %r5497, 24; sub.s32 %r5500, %r5498, %r5499; and.b32 %r5501, %r5500, 31; setp.lt.u32 %p914, %r5501, 24; selp.f32 %f2975, %f2860, %f2873, %p914; setp.lt.u32 %p915, %r5501, 16; selp.f32 %f2976, %f2873, %f2826, %p915; setp.lt.u32 %p916, %r5501, 8; selp.f32 %f2977, %f2826, %f2941, %p916; and.b32 %r5502, %r5500, 1; setp.eq.b32 %p917, %r5502, 1; neg.ftz.f32 %f2978, %f2975; selp.f32 %f2979, %f2978, %f2975, %p917; and.b32 %r5503, %r5500, 2; setp.eq.s32 %p918, %r5503, 0; neg.ftz.f32 %f2980, %f2976; selp.f32 %f2981, %f2976, %f2980, %p918; add.ftz.f32 %f2982, %f2979, %f2981; and.b32 %r5504, %r5500, 4; setp.eq.s32 %p919, %r5504, 0; neg.ftz.f32 %f2983, %f2977; selp.f32 %f2984, %f2977, %f2983, %p919; add.ftz.f32 %f2985, %f2984, %f2982; add.s32 %r5505, %r5247, %r5416; xor.b32 %r5506, %r5255, %r5505; sub.s32 %r5507, %r5506, %r5257; xor.b32 %r5508, %r5507, %r5250; shf.l.wrap.b32 %r5509, %r5507, %r5507, 25; sub.s32 %r5510, %r5508, %r5509; xor.b32 %r5511, %r5510, %r5255; shf.l.wrap.b32 %r5512, %r5510, %r5510, 16; sub.s32 %r5513, %r5511, %r5512; xor.b32 %r5514, %r5513, %r5507; shf.l.wrap.b32 %r5515, %r5513, %r5513, 4; sub.s32 %r5516, %r5514, %r5515; xor.b32 %r5517, %r5516, %r5510; shf.l.wrap.b32 %r5518, %r5516, %r5516, 14; sub.s32 %r5519, %r5517, %r5518; xor.b32 %r5520, %r5519, %r5513; shf.l.wrap.b32 %r5521, %r5519, %r5519, 24; sub.s32 %r5522, %r5520, %r5521; and.b32 %r5523, %r5522, 31; setp.lt.u32 %p920, %r5523, 24; selp.f32 %f2986, %f2822, %f2824, %p920; setp.lt.u32 %p921, %r5523, 16; selp.f32 %f2987, %f2824, %f2896, %p921; setp.lt.u32 %p922, %r5523, 8; selp.f32 %f2988, %f2896, %f2941, %p922; and.b32 %r5524, %r5522, 1; setp.eq.b32 %p923, %r5524, 1; neg.ftz.f32 %f2989, %f2986; selp.f32 %f2990, %f2989, %f2986, %p923; and.b32 %r5525, %r5522, 2; setp.eq.s32 %p924, %r5525, 0; neg.ftz.f32 %f2991, %f2987; selp.f32 %f2992, %f2987, %f2991, %p924; add.ftz.f32 %f2993, %f2990, %f2992; and.b32 %r5526, %r5522, 4; setp.eq.s32 %p925, %r5526, 0; neg.ftz.f32 %f2994, %f2988; selp.f32 %f2995, %f2988, %f2994, %p925; add.ftz.f32 %f2996, %f2995, %f2993; add.s32 %r5527, %r5294, %r5416; xor.b32 %r5528, %r5302, %r5527; sub.s32 %r5529, %r5528, %r5304; xor.b32 %r5530, %r5529, %r5297; shf.l.wrap.b32 %r5531, %r5529, %r5529, 25; sub.s32 %r5532, %r5530, %r5531; xor.b32 %r5533, %r5532, %r5302; shf.l.wrap.b32 %r5534, %r5532, %r5532, 16; sub.s32 %r5535, %r5533, %r5534; xor.b32 %r5536, %r5535, %r5529; shf.l.wrap.b32 %r5537, %r5535, %r5535, 4; sub.s32 %r5538, %r5536, %r5537; xor.b32 %r5539, %r5538, %r5532; shf.l.wrap.b32 %r5540, %r5538, %r5538, 14; sub.s32 %r5541, %r5539, %r5540; xor.b32 %r5542, %r5541, %r5535; shf.l.wrap.b32 %r5543, %r5541, %r5541, 24; sub.s32 %r5544, %r5542, %r5543; and.b32 %r5545, %r5544, 31; setp.lt.u32 %p926, %r5545, 24; selp.f32 %f2997, %f2860, %f2824, %p926; setp.lt.u32 %p927, %r5545, 16; selp.f32 %f2998, %f2824, %f2896, %p927; setp.lt.u32 %p928, %r5545, 8; selp.f32 %f2999, %f2896, %f2941, %p928; and.b32 %r5546, %r5544, 1; setp.eq.b32 %p929, %r5546, 1; neg.ftz.f32 %f3000, %f2997; selp.f32 %f3001, %f3000, %f2997, %p929; and.b32 %r5547, %r5544, 2; setp.eq.s32 %p930, %r5547, 0; neg.ftz.f32 %f3002, %f2998; selp.f32 %f3003, %f2998, %f3002, %p930; add.ftz.f32 %f3004, %f3001, %f3003; and.b32 %r5548, %r5544, 4; setp.eq.s32 %p931, %r5548, 0; neg.ftz.f32 %f3005, %f2999; selp.f32 %f3006, %f2999, %f3005, %p931; add.ftz.f32 %f3007, %f3006, %f3004; add.s32 %r5549, %r5340, %r5416; xor.b32 %r5550, %r5348, %r5549; sub.s32 %r5551, %r5550, %r5350; xor.b32 %r5552, %r5551, %r5343; shf.l.wrap.b32 %r5553, %r5551, %r5551, 25; sub.s32 %r5554, %r5552, %r5553; xor.b32 %r5555, %r5554, %r5348; shf.l.wrap.b32 %r5556, %r5554, %r5554, 16; sub.s32 %r5557, %r5555, %r5556; xor.b32 %r5558, %r5557, %r5551; shf.l.wrap.b32 %r5559, %r5557, %r5557, 4; sub.s32 %r5560, %r5558, %r5559; xor.b32 %r5561, %r5560, %r5554; shf.l.wrap.b32 %r5562, %r5560, %r5560, 14; sub.s32 %r5563, %r5561, %r5562; xor.b32 %r5564, %r5563, %r5557; shf.l.wrap.b32 %r5565, %r5563, %r5563, 24; sub.s32 %r5566, %r5564, %r5565; and.b32 %r5567, %r5566, 31; setp.lt.u32 %p932, %r5567, 24; selp.f32 %f3008, %f2822, %f2873, %p932; setp.lt.u32 %p933, %r5567, 16; selp.f32 %f3009, %f2873, %f2896, %p933; setp.lt.u32 %p934, %r5567, 8; selp.f32 %f3010, %f2896, %f2941, %p934; and.b32 %r5568, %r5566, 1; setp.eq.b32 %p935, %r5568, 1; neg.ftz.f32 %f3011, %f3008; selp.f32 %f3012, %f3011, %f3008, %p935; and.b32 %r5569, %r5566, 2; setp.eq.s32 %p936, %r5569, 0; neg.ftz.f32 %f3013, %f3009; selp.f32 %f3014, %f3009, %f3013, %p936; add.ftz.f32 %f3015, %f3012, %f3014; and.b32 %r5570, %r5566, 4; setp.eq.s32 %p937, %r5570, 0; neg.ftz.f32 %f3016, %f3010; selp.f32 %f3017, %f3010, %f3016, %p937; add.ftz.f32 %f3018, %f3017, %f3015; add.s32 %r5571, %r5385, %r5416; xor.b32 %r5572, %r5393, %r5571; sub.s32 %r5573, %r5572, %r5395; xor.b32 %r5574, %r5573, %r5388; shf.l.wrap.b32 %r5575, %r5573, %r5573, 25; sub.s32 %r5576, %r5574, %r5575; xor.b32 %r5577, %r5576, %r5393; shf.l.wrap.b32 %r5578, %r5576, %r5576, 16; sub.s32 %r5579, %r5577, %r5578; xor.b32 %r5580, %r5579, %r5573; shf.l.wrap.b32 %r5581, %r5579, %r5579, 4; sub.s32 %r5582, %r5580, %r5581; xor.b32 %r5583, %r5582, %r5576; shf.l.wrap.b32 %r5584, %r5582, %r5582, 14; sub.s32 %r5585, %r5583, %r5584; xor.b32 %r5586, %r5585, %r5579; shf.l.wrap.b32 %r5587, %r5585, %r5585, 24; sub.s32 %r5588, %r5586, %r5587; and.b32 %r5589, %r5588, 31; setp.lt.u32 %p938, %r5589, 24; selp.f32 %f3019, %f2860, %f2873, %p938; setp.lt.u32 %p939, %r5589, 16; selp.f32 %f3020, %f2873, %f2896, %p939; setp.lt.u32 %p940, %r5589, 8; selp.f32 %f3021, %f2896, %f2941, %p940; and.b32 %r5590, %r5588, 1; setp.eq.b32 %p941, %r5590, 1; neg.ftz.f32 %f3022, %f3019; selp.f32 %f3023, %f3022, %f3019, %p941; and.b32 %r5591, %r5588, 2; setp.eq.s32 %p942, %r5591, 0; neg.ftz.f32 %f3024, %f3020; selp.f32 %f3025, %f3020, %f3024, %p942; add.ftz.f32 %f3026, %f3023, %f3025; and.b32 %r5592, %r5588, 4; setp.eq.s32 %p943, %r5592, 0; neg.ftz.f32 %f3027, %f3021; selp.f32 %f3028, %f3021, %f3027, %p943; add.ftz.f32 %f3029, %f3028, %f3026; sub.ftz.f32 %f3030, %f2861, %f2833; sub.ftz.f32 %f3031, %f2861, %f2838; sub.ftz.f32 %f3032, %f2861, %f2843; mul.ftz.f32 %f3033, %f3030, %f2859; fma.rn.ftz.f32 %f3034, %f2833, %f2872, %f3033; mul.ftz.f32 %f3035, %f3030, %f2884; fma.rn.ftz.f32 %f3036, %f2833, %f2895, %f3035; mul.ftz.f32 %f3037, %f2838, %f3036; fma.rn.ftz.f32 %f3038, %f3031, %f3034, %f3037; mul.ftz.f32 %f3039, %f3030, %f2907; fma.rn.ftz.f32 %f3040, %f2833, %f2918, %f3039; mul.ftz.f32 %f3041, %f3030, %f2929; fma.rn.ftz.f32 %f3042, %f2833, %f2940, %f3041; mul.ftz.f32 %f3043, %f2838, %f3042; fma.rn.ftz.f32 %f3044, %f3031, %f3040, %f3043; mul.ftz.f32 %f3045, %f2843, %f3044; fma.rn.ftz.f32 %f3046, %f3032, %f3038, %f3045; mul.ftz.f32 %f3047, %f3030, %f2952; fma.rn.ftz.f32 %f3048, %f2833, %f2963, %f3047; mul.ftz.f32 %f3049, %f3030, %f2974; fma.rn.ftz.f32 %f3050, %f2833, %f2985, %f3049; mul.ftz.f32 %f3051, %f2838, %f3050; fma.rn.ftz.f32 %f3052, %f3031, %f3048, %f3051; mul.ftz.f32 %f3053, %f3030, %f2996; fma.rn.ftz.f32 %f3054, %f2833, %f3007, %f3053; mul.ftz.f32 %f3055, %f3030, %f3018; fma.rn.ftz.f32 %f3056, %f2833, %f3029, %f3055; mul.ftz.f32 %f3057, %f2838, %f3056; fma.rn.ftz.f32 %f3058, %f3031, %f3054, %f3057; mul.ftz.f32 %f3059, %f2843, %f3058; fma.rn.ftz.f32 %f3060, %f3032, %f3052, %f3059; sub.ftz.f32 %f3061, %f3060, %f3046; fma.rn.ftz.f32 %f142, %f2848, %f3061, %f3046; mov.b32 %r36, %f142; setp.nan.ftz.f32 %p944, %f142, %f142; mov.f32 %f5574, 0f00000000; @%p944 bra $L__BB74_53; or.b32 %r5593, %r36, -2147483648; setp.ne.s32 %p945, %r5593, -2147483648; add.ftz.f32 %f3063, %f142, %f142; setp.eq.ftz.f32 %p946, %f3063, %f142; and.pred %p947, %p946, %p945; @%p947 bra $L__BB74_53; shl.b32 %r5594, %r36, 1; setp.lt.u32 %p948, %r5594, -16777215; mul.ftz.f32 %f3064, %f142, 0f3F559B3D; selp.f32 %f5574, %f3064, 0f00000000, %p948; $L__BB74_53: mul.ftz.f32 %f3065, %f141, %f5574; fma.rn.ftz.f32 %f5827, %f5573, %f3065, %f5827; bra.uni $L__BB74_450; $L__BB74_389: setp.eq.s16 %p1892, %rs273, 0; @%p1892 bra $L__BB74_439; setp.eq.s16 %p1893, %rs273, 1; @%p1893 bra $L__BB74_391; bra.uni $L__BB74_450; $L__BB74_391: cvt.rzi.ftz.s32.f32 %r474, %f26; setp.lt.s32 %p1957, %r474, 1; mov.f32 %f5817, 0f3F800000; mov.f32 %f5827, 0f00000000; @%p1957 bra $L__BB74_397; lg2.approx.ftz.f32 %f5415, %f27; mul.ftz.f32 %f5416, %f25, %f5415; neg.ftz.f32 %f5417, %f5416; ex2.approx.ftz.f32 %f777, %f5417; mov.f32 %f5414, 0f00000000; mov.f32 %f5817, 0f3F800000; mov.u32 %r10036, 0; mov.f32 %f5827, %f5414; $L__BB74_393: cvt.rmi.ftz.f32.f32 %f5419, %f5815; cvt.rzi.ftz.s32.f32 %r9833, %f5419; sub.ftz.f32 %f5420, %f5815, %f5419; mul.ftz.f32 %f5421, %f5420, %f5420; mul.ftz.f32 %f5422, %f5420, %f5421; fma.rn.ftz.f32 %f5423, %f5420, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5424, %f5420, %f5423, 0f41200000; mul.ftz.f32 %f5425, %f5422, %f5424; add.s32 %r9834, %r9833, -559038720; xor.b32 %r9835, %r9834, -1874868139; add.s32 %r9836, %r9835, 29184895; xor.b32 %r9837, %r9836, -559038720; shf.l.wrap.b32 %r9838, %r9836, %r9836, 25; sub.s32 %r9839, %r9837, %r9838; xor.b32 %r9840, %r9839, -1874868139; shf.l.wrap.b32 %r9841, %r9839, %r9839, 16; sub.s32 %r9842, %r9840, %r9841; xor.b32 %r9843, %r9842, %r9836; shf.l.wrap.b32 %r9844, %r9842, %r9842, 4; sub.s32 %r9845, %r9843, %r9844; xor.b32 %r9846, %r9845, %r9839; shf.l.wrap.b32 %r9847, %r9845, %r9845, 14; sub.s32 %r9848, %r9846, %r9847; xor.b32 %r9849, %r9848, %r9842; shf.l.wrap.b32 %r9850, %r9848, %r9848, 24; sub.s32 %r9851, %r9849, %r9850; and.b32 %r9852, %r9851, 7; add.s32 %r9853, %r9852, 1; cvt.rn.f32.s32 %f5426, %r9853; and.b32 %r9854, %r9851, 8; setp.eq.s32 %p1958, %r9854, 0; neg.ftz.f32 %f5427, %f5426; selp.f32 %f5428, %f5426, %f5427, %p1958; mul.ftz.f32 %f5429, %f5420, %f5428; add.s32 %r9855, %r9833, -559038719; xor.b32 %r9856, %r9855, -1874868139; add.s32 %r9857, %r9856, 29184895; xor.b32 %r9858, %r9857, -559038720; shf.l.wrap.b32 %r9859, %r9857, %r9857, 25; sub.s32 %r9860, %r9858, %r9859; xor.b32 %r9861, %r9860, -1874868139; shf.l.wrap.b32 %r9862, %r9860, %r9860, 16; sub.s32 %r9863, %r9861, %r9862; xor.b32 %r9864, %r9863, %r9857; shf.l.wrap.b32 %r9865, %r9863, %r9863, 4; sub.s32 %r9866, %r9864, %r9865; xor.b32 %r9867, %r9866, %r9860; shf.l.wrap.b32 %r9868, %r9866, %r9866, 14; sub.s32 %r9869, %r9867, %r9868; xor.b32 %r9870, %r9869, %r9863; shf.l.wrap.b32 %r9871, %r9869, %r9869, 24; sub.s32 %r9872, %r9870, %r9871; add.ftz.f32 %f5430, %f5420, 0fBF800000; and.b32 %r9873, %r9872, 7; add.s32 %r9874, %r9873, 1; cvt.rn.f32.s32 %f5431, %r9874; and.b32 %r9875, %r9872, 8; setp.eq.s32 %p1959, %r9875, 0; neg.ftz.f32 %f5432, %f5431; selp.f32 %f5433, %f5431, %f5432, %p1959; mul.ftz.f32 %f5434, %f5430, %f5433; sub.ftz.f32 %f5435, %f5434, %f5429; fma.rn.ftz.f32 %f781, %f5425, %f5435, %f5429; mov.b32 %r476, %f781; setp.nan.ftz.f32 %p1960, %f781, %f781; mov.f32 %f5814, %f5414; @%p1960 bra $L__BB74_396; or.b32 %r9876, %r476, -2147483648; setp.ne.s32 %p1961, %r9876, -2147483648; add.ftz.f32 %f5437, %f781, %f781; setp.eq.ftz.f32 %p1962, %f5437, %f781; and.pred %p1963, %p1962, %p1961; mov.f32 %f5814, %f5414; @%p1963 bra $L__BB74_396; shl.b32 %r9877, %r476, 1; setp.lt.u32 %p1964, %r9877, -16777215; selp.f32 %f5814, %f781, 0f00000000, %p1964; $L__BB74_396: mul.ftz.f32 %f5438, %f5814, 0f3E800000; fma.rn.ftz.f32 %f5827, %f5817, %f5438, %f5827; mul.ftz.f32 %f5817, %f777, %f5817; mul.ftz.f32 %f5815, %f27, %f5815; add.s32 %r10036, %r10036, 1; setp.lt.s32 %p1965, %r10036, %r474; @%p1965 bra $L__BB74_393; $L__BB74_397: cvt.rmi.ftz.f32.f32 %f5439, %f26; sub.ftz.f32 %f790, %f26, %f5439; setp.eq.ftz.f32 %p1966, %f790, 0f00000000; @%p1966 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f5441, %f5815; cvt.rzi.ftz.s32.f32 %r9878, %f5441; sub.ftz.f32 %f5442, %f5815, %f5441; mul.ftz.f32 %f5443, %f5442, %f5442; mul.ftz.f32 %f5444, %f5442, %f5443; fma.rn.ftz.f32 %f5445, %f5442, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5446, %f5442, %f5445, 0f41200000; mul.ftz.f32 %f5447, %f5444, %f5446; add.s32 %r9879, %r9878, -559038720; xor.b32 %r9880, %r9879, -1874868139; add.s32 %r9881, %r9880, 29184895; xor.b32 %r9882, %r9881, -559038720; shf.l.wrap.b32 %r9883, %r9881, %r9881, 25; sub.s32 %r9884, %r9882, %r9883; xor.b32 %r9885, %r9884, -1874868139; shf.l.wrap.b32 %r9886, %r9884, %r9884, 16; sub.s32 %r9887, %r9885, %r9886; xor.b32 %r9888, %r9887, %r9881; shf.l.wrap.b32 %r9889, %r9887, %r9887, 4; sub.s32 %r9890, %r9888, %r9889; xor.b32 %r9891, %r9890, %r9884; shf.l.wrap.b32 %r9892, %r9890, %r9890, 14; sub.s32 %r9893, %r9891, %r9892; xor.b32 %r9894, %r9893, %r9887; shf.l.wrap.b32 %r9895, %r9893, %r9893, 24; sub.s32 %r9896, %r9894, %r9895; and.b32 %r9897, %r9896, 7; add.s32 %r9898, %r9897, 1; cvt.rn.f32.s32 %f5448, %r9898; and.b32 %r9899, %r9896, 8; setp.eq.s32 %p1967, %r9899, 0; neg.ftz.f32 %f5449, %f5448; selp.f32 %f5450, %f5448, %f5449, %p1967; mul.ftz.f32 %f5451, %f5442, %f5450; add.s32 %r9900, %r9878, -559038719; xor.b32 %r9901, %r9900, -1874868139; add.s32 %r9902, %r9901, 29184895; xor.b32 %r9903, %r9902, -559038720; shf.l.wrap.b32 %r9904, %r9902, %r9902, 25; sub.s32 %r9905, %r9903, %r9904; xor.b32 %r9906, %r9905, -1874868139; shf.l.wrap.b32 %r9907, %r9905, %r9905, 16; sub.s32 %r9908, %r9906, %r9907; xor.b32 %r9909, %r9908, %r9902; shf.l.wrap.b32 %r9910, %r9908, %r9908, 4; sub.s32 %r9911, %r9909, %r9910; xor.b32 %r9912, %r9911, %r9905; shf.l.wrap.b32 %r9913, %r9911, %r9911, 14; sub.s32 %r9914, %r9912, %r9913; xor.b32 %r9915, %r9914, %r9908; shf.l.wrap.b32 %r9916, %r9914, %r9914, 24; sub.s32 %r9917, %r9915, %r9916; add.ftz.f32 %f5452, %f5442, 0fBF800000; and.b32 %r9918, %r9917, 7; add.s32 %r9919, %r9918, 1; cvt.rn.f32.s32 %f5453, %r9919; and.b32 %r9920, %r9917, 8; setp.eq.s32 %p1968, %r9920, 0; neg.ftz.f32 %f5454, %f5453; selp.f32 %f5455, %f5453, %f5454, %p1968; mul.ftz.f32 %f5456, %f5452, %f5455; sub.ftz.f32 %f5457, %f5456, %f5451; fma.rn.ftz.f32 %f791, %f5447, %f5457, %f5451; mov.b32 %r478, %f791; setp.nan.ftz.f32 %p1969, %f791, %f791; mov.f32 %f5818, 0f00000000; @%p1969 bra $L__BB74_401; or.b32 %r9921, %r478, -2147483648; setp.ne.s32 %p1970, %r9921, -2147483648; add.ftz.f32 %f5459, %f791, %f791; setp.eq.ftz.f32 %p1971, %f5459, %f791; and.pred %p1972, %p1971, %p1970; @%p1972 bra $L__BB74_401; shl.b32 %r9922, %r478, 1; setp.lt.u32 %p1973, %r9922, -16777215; mul.ftz.f32 %f5460, %f791, 0f3E800000; selp.f32 %f5818, %f5460, 0f00000000, %p1973; $L__BB74_401: mul.ftz.f32 %f5461, %f790, %f5818; fma.rn.ftz.f32 %f5827, %f5817, %f5461, %f5827; bra.uni $L__BB74_450; $L__BB74_103: setp.eq.s16 %p1158, %rs5, 0; @%p1158 bra $L__BB74_297; setp.eq.s16 %p1159, %rs5, 1; @%p1159 bra $L__BB74_105; bra.uni $L__BB74_450; $L__BB74_105: cvt.rzi.ftz.s32.f32 %r282, %f26; setp.lt.s32 %p1489, %r282, 1; mov.f32 %f5698, 0f3F800000; mov.f32 %f5827, 0f00000000; @%p1489 bra $L__BB74_127; lg2.approx.ftz.f32 %f4176, %f27; mul.ftz.f32 %f4177, %f25, %f4176; neg.ftz.f32 %f4178, %f4177; ex2.approx.ftz.f32 %f445, %f4178; mov.f32 %f5827, 0f00000000; mov.f32 %f5698, 0f3F800000; mov.u32 %r10026, 0; $L__BB74_107: cvt.rmi.ftz.f32.f32 %f4179, %f5694; cvt.rzi.ftz.s32.f32 %r284, %f4179; sub.ftz.f32 %f451, %f5694, %f4179; cvt.rmi.ftz.f32.f32 %f4180, %f5695; cvt.rzi.ftz.s32.f32 %r285, %f4180; sub.ftz.f32 %f452, %f5695, %f4180; cvt.rmi.ftz.f32.f32 %f4181, %f5696; cvt.rzi.ftz.s32.f32 %r286, %f4181; sub.ftz.f32 %f453, %f5696, %f4181; fma.rn.ftz.f32 %f4182, %f451, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f454, %f451, %f4182, 0f41200000; fma.rn.ftz.f32 %f4183, %f452, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f455, %f452, %f4183, 0f41200000; fma.rn.ftz.f32 %f4184, %f453, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f456, %f453, %f4184, 0f41200000; add.s32 %r287, %r286, -559038712; add.s32 %r288, %r285, -559038712; add.s32 %r289, %r284, -559038712; xor.b32 %r7852, %r287, %r288; shf.l.wrap.b32 %r290, %r288, %r288, 14; sub.s32 %r291, %r7852, %r290; xor.b32 %r7853, %r291, %r289; shf.l.wrap.b32 %r292, %r291, %r291, 11; sub.s32 %r7854, %r7853, %r292; xor.b32 %r7855, %r7854, %r288; shf.l.wrap.b32 %r7856, %r7854, %r7854, 25; sub.s32 %r7857, %r7855, %r7856; xor.b32 %r7858, %r7857, %r291; shf.l.wrap.b32 %r7859, %r7857, %r7857, 16; sub.s32 %r7860, %r7858, %r7859; xor.b32 %r7861, %r7860, %r7854; shf.l.wrap.b32 %r7862, %r7860, %r7860, 4; sub.s32 %r7863, %r7861, %r7862; xor.b32 %r7864, %r7863, %r7857; shf.l.wrap.b32 %r7865, %r7863, %r7863, 14; sub.s32 %r7866, %r7864, %r7865; xor.b32 %r7867, %r7866, %r7860; shf.l.wrap.b32 %r7868, %r7866, %r7866, 24; sub.s32 %r293, %r7867, %r7868; and.b32 %r294, %r293, 15; cvt.u16.u32 %rs175, %r293; or.b16 %rs176, %rs175, 2; and.b16 %rs177, %rs176, 15; setp.eq.s16 %p1490, %rs177, 14; mov.f32 %f5685, %f451; @%p1490 bra $L__BB74_109; mov.f32 %f5685, %f453; $L__BB74_109: mul.ftz.f32 %f4185, %f451, %f451; mul.ftz.f32 %f4186, %f451, %f4185; mul.ftz.f32 %f458, %f4186, %f454; mul.ftz.f32 %f4187, %f452, %f452; mul.ftz.f32 %f4188, %f452, %f4187; mul.ftz.f32 %f459, %f4188, %f455; mul.ftz.f32 %f4189, %f453, %f453; mul.ftz.f32 %f4190, %f453, %f4189; mul.ftz.f32 %f460, %f4190, %f456; setp.lt.u32 %p1491, %r294, 8; selp.f32 %f4191, %f451, %f452, %p1491; setp.lt.u32 %p1492, %r294, 4; selp.f32 %f4192, %f452, %f5685, %p1492; and.b32 %r7869, %r293, 1; setp.eq.b32 %p1493, %r7869, 1; neg.ftz.f32 %f4193, %f4191; selp.f32 %f4194, %f4193, %f4191, %p1493; and.b32 %r7870, %r293, 2; setp.eq.s32 %p1494, %r7870, 0; neg.ftz.f32 %f4195, %f4192; selp.f32 %f4196, %f4192, %f4195, %p1494; add.ftz.f32 %f461, %f4194, %f4196; add.s32 %r295, %r284, -559038711; xor.b32 %r7871, %r291, %r295; sub.s32 %r7872, %r7871, %r292; xor.b32 %r7873, %r7872, %r288; shf.l.wrap.b32 %r7874, %r7872, %r7872, 25; sub.s32 %r7875, %r7873, %r7874; xor.b32 %r7876, %r7875, %r291; shf.l.wrap.b32 %r7877, %r7875, %r7875, 16; sub.s32 %r7878, %r7876, %r7877; xor.b32 %r7879, %r7878, %r7872; shf.l.wrap.b32 %r7880, %r7878, %r7878, 4; sub.s32 %r7881, %r7879, %r7880; xor.b32 %r7882, %r7881, %r7875; shf.l.wrap.b32 %r7883, %r7881, %r7881, 14; sub.s32 %r7884, %r7882, %r7883; xor.b32 %r7885, %r7884, %r7878; shf.l.wrap.b32 %r7886, %r7884, %r7884, 24; sub.s32 %r296, %r7885, %r7886; and.b32 %r297, %r296, 15; add.ftz.f32 %f462, %f451, 0fBF800000; cvt.u16.u32 %rs178, %r296; or.b16 %rs179, %rs178, 2; and.b16 %rs180, %rs179, 15; setp.eq.s16 %p1495, %rs180, 14; mov.f32 %f5686, %f462; @%p1495 bra $L__BB74_111; mov.f32 %f5686, %f453; $L__BB74_111: setp.lt.u32 %p1496, %r297, 8; selp.f32 %f4197, %f462, %f452, %p1496; setp.lt.u32 %p1497, %r297, 4; selp.f32 %f4198, %f452, %f5686, %p1497; and.b32 %r7887, %r296, 1; setp.eq.b32 %p1498, %r7887, 1; neg.ftz.f32 %f4199, %f4197; selp.f32 %f4200, %f4199, %f4197, %p1498; and.b32 %r7888, %r296, 2; setp.eq.s32 %p1499, %r7888, 0; neg.ftz.f32 %f4201, %f4198; selp.f32 %f4202, %f4198, %f4201, %p1499; add.ftz.f32 %f464, %f4200, %f4202; add.s32 %r298, %r285, -559038711; xor.b32 %r7889, %r287, %r298; shf.l.wrap.b32 %r299, %r298, %r298, 14; sub.s32 %r300, %r7889, %r299; xor.b32 %r7890, %r300, %r289; shf.l.wrap.b32 %r301, %r300, %r300, 11; sub.s32 %r7891, %r7890, %r301; xor.b32 %r7892, %r7891, %r298; shf.l.wrap.b32 %r7893, %r7891, %r7891, 25; sub.s32 %r7894, %r7892, %r7893; xor.b32 %r7895, %r7894, %r300; shf.l.wrap.b32 %r7896, %r7894, %r7894, 16; sub.s32 %r7897, %r7895, %r7896; xor.b32 %r7898, %r7897, %r7891; shf.l.wrap.b32 %r7899, %r7897, %r7897, 4; sub.s32 %r7900, %r7898, %r7899; xor.b32 %r7901, %r7900, %r7894; shf.l.wrap.b32 %r7902, %r7900, %r7900, 14; sub.s32 %r7903, %r7901, %r7902; xor.b32 %r7904, %r7903, %r7897; shf.l.wrap.b32 %r7905, %r7903, %r7903, 24; sub.s32 %r302, %r7904, %r7905; and.b32 %r303, %r302, 15; add.ftz.f32 %f465, %f452, 0fBF800000; cvt.u16.u32 %rs181, %r302; or.b16 %rs182, %rs181, 2; and.b16 %rs183, %rs182, 15; setp.eq.s16 %p1500, %rs183, 14; mov.f32 %f5687, %f451; @%p1500 bra $L__BB74_113; mov.f32 %f5687, %f453; $L__BB74_113: setp.lt.u32 %p1501, %r303, 8; selp.f32 %f4203, %f451, %f465, %p1501; setp.lt.u32 %p1502, %r303, 4; selp.f32 %f4204, %f465, %f5687, %p1502; and.b32 %r7906, %r302, 1; setp.eq.b32 %p1503, %r7906, 1; neg.ftz.f32 %f4205, %f4203; selp.f32 %f4206, %f4205, %f4203, %p1503; and.b32 %r7907, %r302, 2; setp.eq.s32 %p1504, %r7907, 0; neg.ftz.f32 %f4207, %f4204; selp.f32 %f4208, %f4204, %f4207, %p1504; add.ftz.f32 %f467, %f4206, %f4208; xor.b32 %r7908, %r300, %r295; sub.s32 %r7909, %r7908, %r301; xor.b32 %r7910, %r7909, %r298; shf.l.wrap.b32 %r7911, %r7909, %r7909, 25; sub.s32 %r7912, %r7910, %r7911; xor.b32 %r7913, %r7912, %r300; shf.l.wrap.b32 %r7914, %r7912, %r7912, 16; sub.s32 %r7915, %r7913, %r7914; xor.b32 %r7916, %r7915, %r7909; shf.l.wrap.b32 %r7917, %r7915, %r7915, 4; sub.s32 %r7918, %r7916, %r7917; xor.b32 %r7919, %r7918, %r7912; shf.l.wrap.b32 %r7920, %r7918, %r7918, 14; sub.s32 %r7921, %r7919, %r7920; xor.b32 %r7922, %r7921, %r7915; shf.l.wrap.b32 %r7923, %r7921, %r7921, 24; sub.s32 %r304, %r7922, %r7923; and.b32 %r305, %r304, 15; cvt.u16.u32 %rs184, %r304; or.b16 %rs185, %rs184, 2; and.b16 %rs186, %rs185, 15; setp.eq.s16 %p1505, %rs186, 14; mov.f32 %f5688, %f462; @%p1505 bra $L__BB74_115; mov.f32 %f5688, %f453; $L__BB74_115: setp.lt.u32 %p1506, %r305, 8; selp.f32 %f4209, %f462, %f465, %p1506; setp.lt.u32 %p1507, %r305, 4; selp.f32 %f4210, %f465, %f5688, %p1507; and.b32 %r7924, %r304, 1; setp.eq.b32 %p1508, %r7924, 1; neg.ftz.f32 %f4211, %f4209; selp.f32 %f4212, %f4211, %f4209, %p1508; and.b32 %r7925, %r304, 2; setp.eq.s32 %p1509, %r7925, 0; neg.ftz.f32 %f4213, %f4210; selp.f32 %f4214, %f4210, %f4213, %p1509; add.ftz.f32 %f469, %f4212, %f4214; add.s32 %r306, %r286, -559038711; xor.b32 %r7926, %r306, %r288; sub.s32 %r307, %r7926, %r290; xor.b32 %r7927, %r307, %r289; shf.l.wrap.b32 %r308, %r307, %r307, 11; sub.s32 %r7928, %r7927, %r308; xor.b32 %r7929, %r7928, %r288; shf.l.wrap.b32 %r7930, %r7928, %r7928, 25; sub.s32 %r7931, %r7929, %r7930; xor.b32 %r7932, %r7931, %r307; shf.l.wrap.b32 %r7933, %r7931, %r7931, 16; sub.s32 %r7934, %r7932, %r7933; xor.b32 %r7935, %r7934, %r7928; shf.l.wrap.b32 %r7936, %r7934, %r7934, 4; sub.s32 %r7937, %r7935, %r7936; xor.b32 %r7938, %r7937, %r7931; shf.l.wrap.b32 %r7939, %r7937, %r7937, 14; sub.s32 %r7940, %r7938, %r7939; xor.b32 %r7941, %r7940, %r7934; shf.l.wrap.b32 %r7942, %r7940, %r7940, 24; sub.s32 %r309, %r7941, %r7942; add.ftz.f32 %f470, %f453, 0fBF800000; and.b32 %r310, %r309, 15; cvt.u16.u32 %rs187, %r309; or.b16 %rs188, %rs187, 2; and.b16 %rs189, %rs188, 15; setp.eq.s16 %p1510, %rs189, 14; mov.f32 %f5689, %f451; @%p1510 bra $L__BB74_117; mov.f32 %f5689, %f470; $L__BB74_117: setp.lt.u32 %p1511, %r310, 8; selp.f32 %f4215, %f451, %f452, %p1511; setp.lt.u32 %p1512, %r310, 4; selp.f32 %f4216, %f452, %f5689, %p1512; and.b32 %r7943, %r309, 1; setp.eq.b32 %p1513, %r7943, 1; neg.ftz.f32 %f4217, %f4215; selp.f32 %f4218, %f4217, %f4215, %p1513; and.b32 %r7944, %r309, 2; setp.eq.s32 %p1514, %r7944, 0; neg.ftz.f32 %f4219, %f4216; selp.f32 %f4220, %f4216, %f4219, %p1514; add.ftz.f32 %f472, %f4218, %f4220; xor.b32 %r7945, %r307, %r295; sub.s32 %r7946, %r7945, %r308; xor.b32 %r7947, %r7946, %r288; shf.l.wrap.b32 %r7948, %r7946, %r7946, 25; sub.s32 %r7949, %r7947, %r7948; xor.b32 %r7950, %r7949, %r307; shf.l.wrap.b32 %r7951, %r7949, %r7949, 16; sub.s32 %r7952, %r7950, %r7951; xor.b32 %r7953, %r7952, %r7946; shf.l.wrap.b32 %r7954, %r7952, %r7952, 4; sub.s32 %r7955, %r7953, %r7954; xor.b32 %r7956, %r7955, %r7949; shf.l.wrap.b32 %r7957, %r7955, %r7955, 14; sub.s32 %r7958, %r7956, %r7957; xor.b32 %r7959, %r7958, %r7952; shf.l.wrap.b32 %r7960, %r7958, %r7958, 24; sub.s32 %r311, %r7959, %r7960; and.b32 %r312, %r311, 15; cvt.u16.u32 %rs190, %r311; or.b16 %rs191, %rs190, 2; and.b16 %rs192, %rs191, 15; setp.eq.s16 %p1515, %rs192, 14; mov.f32 %f5690, %f462; @%p1515 bra $L__BB74_119; mov.f32 %f5690, %f470; $L__BB74_119: setp.lt.u32 %p1516, %r312, 8; selp.f32 %f4221, %f462, %f452, %p1516; setp.lt.u32 %p1517, %r312, 4; selp.f32 %f4222, %f452, %f5690, %p1517; and.b32 %r7961, %r311, 1; setp.eq.b32 %p1518, %r7961, 1; neg.ftz.f32 %f4223, %f4221; selp.f32 %f4224, %f4223, %f4221, %p1518; and.b32 %r7962, %r311, 2; setp.eq.s32 %p1519, %r7962, 0; neg.ftz.f32 %f4225, %f4222; selp.f32 %f4226, %f4222, %f4225, %p1519; add.ftz.f32 %f474, %f4224, %f4226; xor.b32 %r7963, %r306, %r298; sub.s32 %r313, %r7963, %r299; xor.b32 %r7964, %r313, %r289; shf.l.wrap.b32 %r314, %r313, %r313, 11; sub.s32 %r7965, %r7964, %r314; xor.b32 %r7966, %r7965, %r298; shf.l.wrap.b32 %r7967, %r7965, %r7965, 25; sub.s32 %r7968, %r7966, %r7967; xor.b32 %r7969, %r7968, %r313; shf.l.wrap.b32 %r7970, %r7968, %r7968, 16; sub.s32 %r7971, %r7969, %r7970; xor.b32 %r7972, %r7971, %r7965; shf.l.wrap.b32 %r7973, %r7971, %r7971, 4; sub.s32 %r7974, %r7972, %r7973; xor.b32 %r7975, %r7974, %r7968; shf.l.wrap.b32 %r7976, %r7974, %r7974, 14; sub.s32 %r7977, %r7975, %r7976; xor.b32 %r7978, %r7977, %r7971; shf.l.wrap.b32 %r7979, %r7977, %r7977, 24; sub.s32 %r315, %r7978, %r7979; and.b32 %r316, %r315, 15; cvt.u16.u32 %rs193, %r315; or.b16 %rs194, %rs193, 2; and.b16 %rs195, %rs194, 15; setp.eq.s16 %p1520, %rs195, 14; mov.f32 %f5691, %f451; @%p1520 bra $L__BB74_121; mov.f32 %f5691, %f470; $L__BB74_121: setp.lt.u32 %p1521, %r316, 8; selp.f32 %f4227, %f451, %f465, %p1521; setp.lt.u32 %p1522, %r316, 4; selp.f32 %f4228, %f465, %f5691, %p1522; and.b32 %r7980, %r315, 1; setp.eq.b32 %p1523, %r7980, 1; neg.ftz.f32 %f4229, %f4227; selp.f32 %f4230, %f4229, %f4227, %p1523; and.b32 %r7981, %r315, 2; setp.eq.s32 %p1524, %r7981, 0; neg.ftz.f32 %f4231, %f4228; selp.f32 %f4232, %f4228, %f4231, %p1524; add.ftz.f32 %f476, %f4230, %f4232; xor.b32 %r7982, %r313, %r295; sub.s32 %r7983, %r7982, %r314; xor.b32 %r7984, %r7983, %r298; shf.l.wrap.b32 %r7985, %r7983, %r7983, 25; sub.s32 %r7986, %r7984, %r7985; xor.b32 %r7987, %r7986, %r313; shf.l.wrap.b32 %r7988, %r7986, %r7986, 16; sub.s32 %r7989, %r7987, %r7988; xor.b32 %r7990, %r7989, %r7983; shf.l.wrap.b32 %r7991, %r7989, %r7989, 4; sub.s32 %r7992, %r7990, %r7991; xor.b32 %r7993, %r7992, %r7986; shf.l.wrap.b32 %r7994, %r7992, %r7992, 14; sub.s32 %r7995, %r7993, %r7994; xor.b32 %r7996, %r7995, %r7989; shf.l.wrap.b32 %r7997, %r7995, %r7995, 24; sub.s32 %r317, %r7996, %r7997; and.b32 %r318, %r317, 15; cvt.u16.u32 %rs196, %r317; or.b16 %rs197, %rs196, 2; and.b16 %rs198, %rs197, 15; setp.eq.s16 %p1525, %rs198, 14; mov.f32 %f5692, %f462; @%p1525 bra $L__BB74_123; mov.f32 %f5692, %f470; $L__BB74_123: setp.lt.u32 %p1526, %r318, 8; selp.f32 %f4234, %f462, %f465, %p1526; setp.lt.u32 %p1527, %r318, 4; selp.f32 %f4235, %f465, %f5692, %p1527; and.b32 %r7998, %r317, 1; setp.eq.b32 %p1528, %r7998, 1; neg.ftz.f32 %f4236, %f4234; selp.f32 %f4237, %f4236, %f4234, %p1528; and.b32 %r7999, %r317, 2; setp.eq.s32 %p1529, %r7999, 0; neg.ftz.f32 %f4238, %f4235; selp.f32 %f4239, %f4235, %f4238, %p1529; add.ftz.f32 %f4240, %f4237, %f4239; mov.f32 %f4241, 0f3F800000; sub.ftz.f32 %f4242, %f4241, %f458; mul.ftz.f32 %f4243, %f458, %f464; fma.rn.ftz.f32 %f4244, %f4242, %f461, %f4243; sub.ftz.f32 %f4245, %f4241, %f459; mul.ftz.f32 %f4246, %f458, %f469; fma.rn.ftz.f32 %f4247, %f4242, %f467, %f4246; mul.ftz.f32 %f4248, %f459, %f4247; fma.rn.ftz.f32 %f4249, %f4245, %f4244, %f4248; sub.ftz.f32 %f4250, %f4241, %f460; mul.ftz.f32 %f4251, %f458, %f474; fma.rn.ftz.f32 %f4252, %f4242, %f472, %f4251; mul.ftz.f32 %f4253, %f458, %f4240; fma.rn.ftz.f32 %f4254, %f4242, %f476, %f4253; mul.ftz.f32 %f4255, %f459, %f4254; fma.rn.ftz.f32 %f4256, %f4245, %f4252, %f4255; mul.ftz.f32 %f4257, %f460, %f4256; fma.rn.ftz.f32 %f478, %f4250, %f4249, %f4257; mov.b32 %r319, %f478; setp.nan.ftz.f32 %p1530, %f478, %f478; mov.f32 %f5693, 0f00000000; @%p1530 bra $L__BB74_126; or.b32 %r8000, %r319, -2147483648; setp.ne.s32 %p1531, %r8000, -2147483648; add.ftz.f32 %f4259, %f478, %f478; setp.eq.ftz.f32 %p1532, %f4259, %f478; and.pred %p1533, %p1532, %p1531; @%p1533 bra $L__BB74_126; shl.b32 %r8001, %r319, 1; setp.lt.u32 %p1534, %r8001, -16777215; selp.f32 %f5693, %f478, 0f00000000, %p1534; $L__BB74_126: mul.ftz.f32 %f4260, %f5693, 0f3F7B645A; fma.rn.ftz.f32 %f5827, %f5698, %f4260, %f5827; mul.ftz.f32 %f5698, %f445, %f5698; mul.ftz.f32 %f5694, %f27, %f5694; mul.ftz.f32 %f5695, %f27, %f5695; mul.ftz.f32 %f5696, %f27, %f5696; add.s32 %r10026, %r10026, 1; setp.lt.s32 %p1535, %r10026, %r282; @%p1535 bra $L__BB74_107; $L__BB74_127: cvt.rmi.ftz.f32.f32 %f4261, %f26; sub.ftz.f32 %f491, %f26, %f4261; setp.eq.ftz.f32 %p1536, %f491, 0f00000000; @%p1536 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f4262, %f5694; cvt.rzi.ftz.s32.f32 %r321, %f4262; sub.ftz.f32 %f492, %f5694, %f4262; cvt.rmi.ftz.f32.f32 %f4263, %f5695; cvt.rzi.ftz.s32.f32 %r322, %f4263; sub.ftz.f32 %f493, %f5695, %f4263; cvt.rmi.ftz.f32.f32 %f4264, %f5696; cvt.rzi.ftz.s32.f32 %r323, %f4264; sub.ftz.f32 %f494, %f5696, %f4264; fma.rn.ftz.f32 %f4265, %f492, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f495, %f492, %f4265, 0f41200000; fma.rn.ftz.f32 %f4266, %f493, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f496, %f493, %f4266, 0f41200000; fma.rn.ftz.f32 %f4267, %f494, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f497, %f494, %f4267, 0f41200000; add.s32 %r324, %r323, -559038712; add.s32 %r325, %r322, -559038712; add.s32 %r326, %r321, -559038712; xor.b32 %r8002, %r324, %r325; shf.l.wrap.b32 %r327, %r325, %r325, 14; sub.s32 %r328, %r8002, %r327; xor.b32 %r8003, %r328, %r326; shf.l.wrap.b32 %r329, %r328, %r328, 11; sub.s32 %r8004, %r8003, %r329; xor.b32 %r8005, %r8004, %r325; shf.l.wrap.b32 %r8006, %r8004, %r8004, 25; sub.s32 %r8007, %r8005, %r8006; xor.b32 %r8008, %r8007, %r328; shf.l.wrap.b32 %r8009, %r8007, %r8007, 16; sub.s32 %r8010, %r8008, %r8009; xor.b32 %r8011, %r8010, %r8004; shf.l.wrap.b32 %r8012, %r8010, %r8010, 4; sub.s32 %r8013, %r8011, %r8012; xor.b32 %r8014, %r8013, %r8007; shf.l.wrap.b32 %r8015, %r8013, %r8013, 14; sub.s32 %r8016, %r8014, %r8015; xor.b32 %r8017, %r8016, %r8010; shf.l.wrap.b32 %r8018, %r8016, %r8016, 24; sub.s32 %r330, %r8017, %r8018; and.b32 %r331, %r330, 15; cvt.u16.u32 %rs199, %r330; or.b16 %rs200, %rs199, 2; and.b16 %rs201, %rs200, 15; setp.eq.s16 %p1537, %rs201, 14; mov.f32 %f5699, %f492; @%p1537 bra $L__BB74_130; mov.f32 %f5699, %f494; $L__BB74_130: mul.ftz.f32 %f4268, %f492, %f492; mul.ftz.f32 %f4269, %f492, %f4268; mul.ftz.f32 %f499, %f4269, %f495; mul.ftz.f32 %f4270, %f493, %f493; mul.ftz.f32 %f4271, %f493, %f4270; mul.ftz.f32 %f500, %f4271, %f496; mul.ftz.f32 %f4272, %f494, %f494; mul.ftz.f32 %f4273, %f494, %f4272; mul.ftz.f32 %f501, %f4273, %f497; setp.lt.u32 %p1538, %r331, 8; selp.f32 %f4274, %f492, %f493, %p1538; setp.lt.u32 %p1539, %r331, 4; selp.f32 %f4275, %f493, %f5699, %p1539; and.b32 %r8019, %r330, 1; setp.eq.b32 %p1540, %r8019, 1; neg.ftz.f32 %f4276, %f4274; selp.f32 %f4277, %f4276, %f4274, %p1540; and.b32 %r8020, %r330, 2; setp.eq.s32 %p1541, %r8020, 0; neg.ftz.f32 %f4278, %f4275; selp.f32 %f4279, %f4275, %f4278, %p1541; add.ftz.f32 %f502, %f4277, %f4279; add.s32 %r332, %r321, -559038711; xor.b32 %r8021, %r328, %r332; sub.s32 %r8022, %r8021, %r329; xor.b32 %r8023, %r8022, %r325; shf.l.wrap.b32 %r8024, %r8022, %r8022, 25; sub.s32 %r8025, %r8023, %r8024; xor.b32 %r8026, %r8025, %r328; shf.l.wrap.b32 %r8027, %r8025, %r8025, 16; sub.s32 %r8028, %r8026, %r8027; xor.b32 %r8029, %r8028, %r8022; shf.l.wrap.b32 %r8030, %r8028, %r8028, 4; sub.s32 %r8031, %r8029, %r8030; xor.b32 %r8032, %r8031, %r8025; shf.l.wrap.b32 %r8033, %r8031, %r8031, 14; sub.s32 %r8034, %r8032, %r8033; xor.b32 %r8035, %r8034, %r8028; shf.l.wrap.b32 %r8036, %r8034, %r8034, 24; sub.s32 %r333, %r8035, %r8036; and.b32 %r334, %r333, 15; add.ftz.f32 %f503, %f492, 0fBF800000; cvt.u16.u32 %rs202, %r333; or.b16 %rs203, %rs202, 2; and.b16 %rs204, %rs203, 15; setp.eq.s16 %p1542, %rs204, 14; mov.f32 %f5700, %f503; @%p1542 bra $L__BB74_132; mov.f32 %f5700, %f494; $L__BB74_132: setp.lt.u32 %p1543, %r334, 8; selp.f32 %f4280, %f503, %f493, %p1543; setp.lt.u32 %p1544, %r334, 4; selp.f32 %f4281, %f493, %f5700, %p1544; and.b32 %r8037, %r333, 1; setp.eq.b32 %p1545, %r8037, 1; neg.ftz.f32 %f4282, %f4280; selp.f32 %f4283, %f4282, %f4280, %p1545; and.b32 %r8038, %r333, 2; setp.eq.s32 %p1546, %r8038, 0; neg.ftz.f32 %f4284, %f4281; selp.f32 %f4285, %f4281, %f4284, %p1546; add.ftz.f32 %f505, %f4283, %f4285; add.s32 %r335, %r322, -559038711; xor.b32 %r8039, %r324, %r335; shf.l.wrap.b32 %r336, %r335, %r335, 14; sub.s32 %r337, %r8039, %r336; xor.b32 %r8040, %r337, %r326; shf.l.wrap.b32 %r338, %r337, %r337, 11; sub.s32 %r8041, %r8040, %r338; xor.b32 %r8042, %r8041, %r335; shf.l.wrap.b32 %r8043, %r8041, %r8041, 25; sub.s32 %r8044, %r8042, %r8043; xor.b32 %r8045, %r8044, %r337; shf.l.wrap.b32 %r8046, %r8044, %r8044, 16; sub.s32 %r8047, %r8045, %r8046; xor.b32 %r8048, %r8047, %r8041; shf.l.wrap.b32 %r8049, %r8047, %r8047, 4; sub.s32 %r8050, %r8048, %r8049; xor.b32 %r8051, %r8050, %r8044; shf.l.wrap.b32 %r8052, %r8050, %r8050, 14; sub.s32 %r8053, %r8051, %r8052; xor.b32 %r8054, %r8053, %r8047; shf.l.wrap.b32 %r8055, %r8053, %r8053, 24; sub.s32 %r339, %r8054, %r8055; and.b32 %r340, %r339, 15; add.ftz.f32 %f506, %f493, 0fBF800000; cvt.u16.u32 %rs205, %r339; or.b16 %rs206, %rs205, 2; and.b16 %rs207, %rs206, 15; setp.eq.s16 %p1547, %rs207, 14; mov.f32 %f5701, %f492; @%p1547 bra $L__BB74_134; mov.f32 %f5701, %f494; $L__BB74_134: setp.lt.u32 %p1548, %r340, 8; selp.f32 %f4286, %f492, %f506, %p1548; setp.lt.u32 %p1549, %r340, 4; selp.f32 %f4287, %f506, %f5701, %p1549; and.b32 %r8056, %r339, 1; setp.eq.b32 %p1550, %r8056, 1; neg.ftz.f32 %f4288, %f4286; selp.f32 %f4289, %f4288, %f4286, %p1550; and.b32 %r8057, %r339, 2; setp.eq.s32 %p1551, %r8057, 0; neg.ftz.f32 %f4290, %f4287; selp.f32 %f4291, %f4287, %f4290, %p1551; add.ftz.f32 %f508, %f4289, %f4291; xor.b32 %r8058, %r337, %r332; sub.s32 %r8059, %r8058, %r338; xor.b32 %r8060, %r8059, %r335; shf.l.wrap.b32 %r8061, %r8059, %r8059, 25; sub.s32 %r8062, %r8060, %r8061; xor.b32 %r8063, %r8062, %r337; shf.l.wrap.b32 %r8064, %r8062, %r8062, 16; sub.s32 %r8065, %r8063, %r8064; xor.b32 %r8066, %r8065, %r8059; shf.l.wrap.b32 %r8067, %r8065, %r8065, 4; sub.s32 %r8068, %r8066, %r8067; xor.b32 %r8069, %r8068, %r8062; shf.l.wrap.b32 %r8070, %r8068, %r8068, 14; sub.s32 %r8071, %r8069, %r8070; xor.b32 %r8072, %r8071, %r8065; shf.l.wrap.b32 %r8073, %r8071, %r8071, 24; sub.s32 %r341, %r8072, %r8073; and.b32 %r342, %r341, 15; cvt.u16.u32 %rs208, %r341; or.b16 %rs209, %rs208, 2; and.b16 %rs210, %rs209, 15; setp.eq.s16 %p1552, %rs210, 14; mov.f32 %f5702, %f503; @%p1552 bra $L__BB74_136; mov.f32 %f5702, %f494; $L__BB74_136: setp.lt.u32 %p1553, %r342, 8; selp.f32 %f4292, %f503, %f506, %p1553; setp.lt.u32 %p1554, %r342, 4; selp.f32 %f4293, %f506, %f5702, %p1554; and.b32 %r8074, %r341, 1; setp.eq.b32 %p1555, %r8074, 1; neg.ftz.f32 %f4294, %f4292; selp.f32 %f4295, %f4294, %f4292, %p1555; and.b32 %r8075, %r341, 2; setp.eq.s32 %p1556, %r8075, 0; neg.ftz.f32 %f4296, %f4293; selp.f32 %f4297, %f4293, %f4296, %p1556; add.ftz.f32 %f510, %f4295, %f4297; add.s32 %r343, %r323, -559038711; xor.b32 %r8076, %r343, %r325; sub.s32 %r344, %r8076, %r327; xor.b32 %r8077, %r344, %r326; shf.l.wrap.b32 %r345, %r344, %r344, 11; sub.s32 %r8078, %r8077, %r345; xor.b32 %r8079, %r8078, %r325; shf.l.wrap.b32 %r8080, %r8078, %r8078, 25; sub.s32 %r8081, %r8079, %r8080; xor.b32 %r8082, %r8081, %r344; shf.l.wrap.b32 %r8083, %r8081, %r8081, 16; sub.s32 %r8084, %r8082, %r8083; xor.b32 %r8085, %r8084, %r8078; shf.l.wrap.b32 %r8086, %r8084, %r8084, 4; sub.s32 %r8087, %r8085, %r8086; xor.b32 %r8088, %r8087, %r8081; shf.l.wrap.b32 %r8089, %r8087, %r8087, 14; sub.s32 %r8090, %r8088, %r8089; xor.b32 %r8091, %r8090, %r8084; shf.l.wrap.b32 %r8092, %r8090, %r8090, 24; sub.s32 %r346, %r8091, %r8092; add.ftz.f32 %f511, %f494, 0fBF800000; and.b32 %r347, %r346, 15; cvt.u16.u32 %rs211, %r346; or.b16 %rs212, %rs211, 2; and.b16 %rs213, %rs212, 15; setp.eq.s16 %p1557, %rs213, 14; mov.f32 %f5703, %f492; @%p1557 bra $L__BB74_138; mov.f32 %f5703, %f511; $L__BB74_138: setp.lt.u32 %p1558, %r347, 8; selp.f32 %f4298, %f492, %f493, %p1558; setp.lt.u32 %p1559, %r347, 4; selp.f32 %f4299, %f493, %f5703, %p1559; and.b32 %r8093, %r346, 1; setp.eq.b32 %p1560, %r8093, 1; neg.ftz.f32 %f4300, %f4298; selp.f32 %f4301, %f4300, %f4298, %p1560; and.b32 %r8094, %r346, 2; setp.eq.s32 %p1561, %r8094, 0; neg.ftz.f32 %f4302, %f4299; selp.f32 %f4303, %f4299, %f4302, %p1561; add.ftz.f32 %f513, %f4301, %f4303; xor.b32 %r8095, %r344, %r332; sub.s32 %r8096, %r8095, %r345; xor.b32 %r8097, %r8096, %r325; shf.l.wrap.b32 %r8098, %r8096, %r8096, 25; sub.s32 %r8099, %r8097, %r8098; xor.b32 %r8100, %r8099, %r344; shf.l.wrap.b32 %r8101, %r8099, %r8099, 16; sub.s32 %r8102, %r8100, %r8101; xor.b32 %r8103, %r8102, %r8096; shf.l.wrap.b32 %r8104, %r8102, %r8102, 4; sub.s32 %r8105, %r8103, %r8104; xor.b32 %r8106, %r8105, %r8099; shf.l.wrap.b32 %r8107, %r8105, %r8105, 14; sub.s32 %r8108, %r8106, %r8107; xor.b32 %r8109, %r8108, %r8102; shf.l.wrap.b32 %r8110, %r8108, %r8108, 24; sub.s32 %r348, %r8109, %r8110; and.b32 %r349, %r348, 15; cvt.u16.u32 %rs214, %r348; or.b16 %rs215, %rs214, 2; and.b16 %rs216, %rs215, 15; setp.eq.s16 %p1562, %rs216, 14; mov.f32 %f5704, %f503; @%p1562 bra $L__BB74_140; mov.f32 %f5704, %f511; $L__BB74_140: setp.lt.u32 %p1563, %r349, 8; selp.f32 %f4304, %f503, %f493, %p1563; setp.lt.u32 %p1564, %r349, 4; selp.f32 %f4305, %f493, %f5704, %p1564; and.b32 %r8111, %r348, 1; setp.eq.b32 %p1565, %r8111, 1; neg.ftz.f32 %f4306, %f4304; selp.f32 %f4307, %f4306, %f4304, %p1565; and.b32 %r8112, %r348, 2; setp.eq.s32 %p1566, %r8112, 0; neg.ftz.f32 %f4308, %f4305; selp.f32 %f4309, %f4305, %f4308, %p1566; add.ftz.f32 %f515, %f4307, %f4309; xor.b32 %r8113, %r343, %r335; sub.s32 %r350, %r8113, %r336; xor.b32 %r8114, %r350, %r326; shf.l.wrap.b32 %r351, %r350, %r350, 11; sub.s32 %r8115, %r8114, %r351; xor.b32 %r8116, %r8115, %r335; shf.l.wrap.b32 %r8117, %r8115, %r8115, 25; sub.s32 %r8118, %r8116, %r8117; xor.b32 %r8119, %r8118, %r350; shf.l.wrap.b32 %r8120, %r8118, %r8118, 16; sub.s32 %r8121, %r8119, %r8120; xor.b32 %r8122, %r8121, %r8115; shf.l.wrap.b32 %r8123, %r8121, %r8121, 4; sub.s32 %r8124, %r8122, %r8123; xor.b32 %r8125, %r8124, %r8118; shf.l.wrap.b32 %r8126, %r8124, %r8124, 14; sub.s32 %r8127, %r8125, %r8126; xor.b32 %r8128, %r8127, %r8121; shf.l.wrap.b32 %r8129, %r8127, %r8127, 24; sub.s32 %r352, %r8128, %r8129; and.b32 %r353, %r352, 15; cvt.u16.u32 %rs217, %r352; or.b16 %rs218, %rs217, 2; and.b16 %rs219, %rs218, 15; setp.eq.s16 %p1567, %rs219, 14; mov.f32 %f5705, %f492; @%p1567 bra $L__BB74_142; mov.f32 %f5705, %f511; $L__BB74_142: setp.lt.u32 %p1568, %r353, 8; selp.f32 %f4310, %f492, %f506, %p1568; setp.lt.u32 %p1569, %r353, 4; selp.f32 %f4311, %f506, %f5705, %p1569; and.b32 %r8130, %r352, 1; setp.eq.b32 %p1570, %r8130, 1; neg.ftz.f32 %f4312, %f4310; selp.f32 %f4313, %f4312, %f4310, %p1570; and.b32 %r8131, %r352, 2; setp.eq.s32 %p1571, %r8131, 0; neg.ftz.f32 %f4314, %f4311; selp.f32 %f4315, %f4311, %f4314, %p1571; add.ftz.f32 %f517, %f4313, %f4315; xor.b32 %r8132, %r350, %r332; sub.s32 %r8133, %r8132, %r351; xor.b32 %r8134, %r8133, %r335; shf.l.wrap.b32 %r8135, %r8133, %r8133, 25; sub.s32 %r8136, %r8134, %r8135; xor.b32 %r8137, %r8136, %r350; shf.l.wrap.b32 %r8138, %r8136, %r8136, 16; sub.s32 %r8139, %r8137, %r8138; xor.b32 %r8140, %r8139, %r8133; shf.l.wrap.b32 %r8141, %r8139, %r8139, 4; sub.s32 %r8142, %r8140, %r8141; xor.b32 %r8143, %r8142, %r8136; shf.l.wrap.b32 %r8144, %r8142, %r8142, 14; sub.s32 %r8145, %r8143, %r8144; xor.b32 %r8146, %r8145, %r8139; shf.l.wrap.b32 %r8147, %r8145, %r8145, 24; sub.s32 %r354, %r8146, %r8147; and.b32 %r355, %r354, 15; cvt.u16.u32 %rs220, %r354; or.b16 %rs221, %rs220, 2; and.b16 %rs222, %rs221, 15; setp.eq.s16 %p1572, %rs222, 14; mov.f32 %f5706, %f503; @%p1572 bra $L__BB74_144; mov.f32 %f5706, %f511; $L__BB74_144: setp.lt.u32 %p1573, %r355, 8; selp.f32 %f4317, %f503, %f506, %p1573; setp.lt.u32 %p1574, %r355, 4; selp.f32 %f4318, %f506, %f5706, %p1574; and.b32 %r8148, %r354, 1; setp.eq.b32 %p1575, %r8148, 1; neg.ftz.f32 %f4319, %f4317; selp.f32 %f4320, %f4319, %f4317, %p1575; and.b32 %r8149, %r354, 2; setp.eq.s32 %p1576, %r8149, 0; neg.ftz.f32 %f4321, %f4318; selp.f32 %f4322, %f4318, %f4321, %p1576; add.ftz.f32 %f4323, %f4320, %f4322; mov.f32 %f4324, 0f3F800000; sub.ftz.f32 %f4325, %f4324, %f499; mul.ftz.f32 %f4326, %f499, %f505; fma.rn.ftz.f32 %f4327, %f4325, %f502, %f4326; sub.ftz.f32 %f4328, %f4324, %f500; mul.ftz.f32 %f4329, %f499, %f510; fma.rn.ftz.f32 %f4330, %f4325, %f508, %f4329; mul.ftz.f32 %f4331, %f500, %f4330; fma.rn.ftz.f32 %f4332, %f4328, %f4327, %f4331; sub.ftz.f32 %f4333, %f4324, %f501; mul.ftz.f32 %f4334, %f499, %f515; fma.rn.ftz.f32 %f4335, %f4325, %f513, %f4334; mul.ftz.f32 %f4336, %f499, %f4323; fma.rn.ftz.f32 %f4337, %f4325, %f517, %f4336; mul.ftz.f32 %f4338, %f500, %f4337; fma.rn.ftz.f32 %f4339, %f4328, %f4335, %f4338; mul.ftz.f32 %f4340, %f501, %f4339; fma.rn.ftz.f32 %f519, %f4333, %f4332, %f4340; mov.b32 %r356, %f519; setp.nan.ftz.f32 %p1577, %f519, %f519; mov.f32 %f5707, 0f00000000; @%p1577 bra $L__BB74_147; or.b32 %r8150, %r356, -2147483648; setp.ne.s32 %p1578, %r8150, -2147483648; add.ftz.f32 %f4342, %f519, %f519; setp.eq.ftz.f32 %p1579, %f4342, %f519; and.pred %p1580, %p1579, %p1578; @%p1580 bra $L__BB74_147; shl.b32 %r8151, %r356, 1; setp.lt.u32 %p1581, %r8151, -16777215; mul.ftz.f32 %f4343, %f519, 0f3F7B645A; selp.f32 %f5707, %f4343, 0f00000000, %p1581; $L__BB74_147: mul.ftz.f32 %f4344, %f491, %f5707; fma.rn.ftz.f32 %f5827, %f5698, %f4344, %f5827; bra.uni $L__BB74_450; $L__BB74_366: lg2.approx.ftz.f32 %f4841, %f27; mul.ftz.f32 %f4842, %f25, %f4841; neg.ftz.f32 %f4843, %f4842; ex2.approx.ftz.f32 %f649, %f4843; cvt.rzi.ftz.s32.f32 %r443, %f26; mov.f32 %f5764, 0f3F800000; mov.f32 %f5827, 0f00000000; mov.u32 %r10030, 0; mov.pred %p1771, -1; mov.f32 %f5762, %f5764; $L__BB74_367: setp.ge.s32 %p1772, %r10030, %r443; mov.pred %p1993, %p1771; @%p1772 bra $L__BB74_372; cvt.rmi.ftz.f32.f32 %f4845, %f5771; cvt.rzi.ftz.s32.f32 %r8936, %f4845; sub.ftz.f32 %f4846, %f5771, %f4845; cvt.rmi.ftz.f32.f32 %f4847, %f5772; cvt.rzi.ftz.s32.f32 %r8937, %f4847; sub.ftz.f32 %f4848, %f5772, %f4847; mul.ftz.f32 %f4849, %f4846, %f4846; mul.ftz.f32 %f4850, %f4846, %f4849; fma.rn.ftz.f32 %f4851, %f4846, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4852, %f4846, %f4851, 0f41200000; mul.ftz.f32 %f4853, %f4850, %f4852; mul.ftz.f32 %f4854, %f4848, %f4848; mul.ftz.f32 %f4855, %f4848, %f4854; fma.rn.ftz.f32 %f4856, %f4848, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4857, %f4848, %f4856, 0f41200000; mul.ftz.f32 %f4858, %f4855, %f4857; add.s32 %r8938, %r8937, -559038716; add.s32 %r8939, %r8936, -559038716; xor.b32 %r8940, %r8938, -559038716; shf.l.wrap.b32 %r8941, %r8938, %r8938, 14; sub.s32 %r8942, %r8940, %r8941; xor.b32 %r8943, %r8942, %r8939; shf.l.wrap.b32 %r8944, %r8942, %r8942, 11; sub.s32 %r8945, %r8943, %r8944; xor.b32 %r8946, %r8945, %r8938; shf.l.wrap.b32 %r8947, %r8945, %r8945, 25; sub.s32 %r8948, %r8946, %r8947; xor.b32 %r8949, %r8948, %r8942; shf.l.wrap.b32 %r8950, %r8948, %r8948, 16; sub.s32 %r8951, %r8949, %r8950; xor.b32 %r8952, %r8951, %r8945; shf.l.wrap.b32 %r8953, %r8951, %r8951, 4; sub.s32 %r8954, %r8952, %r8953; xor.b32 %r8955, %r8954, %r8948; shf.l.wrap.b32 %r8956, %r8954, %r8954, 14; sub.s32 %r8957, %r8955, %r8956; xor.b32 %r8958, %r8957, %r8951; shf.l.wrap.b32 %r8959, %r8957, %r8957, 24; sub.s32 %r8960, %r8958, %r8959; and.b32 %r8961, %r8960, 4; setp.eq.s32 %p1773, %r8961, 0; selp.f32 %f4859, %f4846, %f4848, %p1773; selp.f32 %f4860, %f4848, %f4846, %p1773; add.ftz.f32 %f4861, %f4860, %f4860; and.b32 %r8962, %r8960, 1; setp.eq.b32 %p1774, %r8962, 1; neg.ftz.f32 %f4862, %f4859; selp.f32 %f4863, %f4862, %f4859, %p1774; and.b32 %r8963, %r8960, 2; setp.eq.s32 %p1775, %r8963, 0; neg.ftz.f32 %f4864, %f4861; selp.f32 %f4865, %f4861, %f4864, %p1775; add.ftz.f32 %f4866, %f4863, %f4865; add.s32 %r8964, %r8936, -559038715; xor.b32 %r8965, %r8942, %r8964; sub.s32 %r8966, %r8965, %r8944; xor.b32 %r8967, %r8966, %r8938; shf.l.wrap.b32 %r8968, %r8966, %r8966, 25; sub.s32 %r8969, %r8967, %r8968; xor.b32 %r8970, %r8969, %r8942; shf.l.wrap.b32 %r8971, %r8969, %r8969, 16; sub.s32 %r8972, %r8970, %r8971; xor.b32 %r8973, %r8972, %r8966; shf.l.wrap.b32 %r8974, %r8972, %r8972, 4; sub.s32 %r8975, %r8973, %r8974; xor.b32 %r8976, %r8975, %r8969; shf.l.wrap.b32 %r8977, %r8975, %r8975, 14; sub.s32 %r8978, %r8976, %r8977; xor.b32 %r8979, %r8978, %r8972; shf.l.wrap.b32 %r8980, %r8978, %r8978, 24; sub.s32 %r8981, %r8979, %r8980; add.ftz.f32 %f4867, %f4846, 0fBF800000; mov.f32 %f4868, 0f3F800000; and.b32 %r8982, %r8981, 4; setp.eq.s32 %p1776, %r8982, 0; selp.f32 %f4869, %f4867, %f4848, %p1776; selp.f32 %f4870, %f4848, %f4867, %p1776; add.ftz.f32 %f4871, %f4870, %f4870; and.b32 %r8983, %r8981, 1; setp.eq.b32 %p1777, %r8983, 1; neg.ftz.f32 %f4872, %f4869; selp.f32 %f4873, %f4872, %f4869, %p1777; and.b32 %r8984, %r8981, 2; setp.eq.s32 %p1778, %r8984, 0; neg.ftz.f32 %f4874, %f4871; selp.f32 %f4875, %f4871, %f4874, %p1778; add.ftz.f32 %f4876, %f4873, %f4875; add.s32 %r8985, %r8937, -559038715; xor.b32 %r8986, %r8985, -559038716; shf.l.wrap.b32 %r8987, %r8985, %r8985, 14; sub.s32 %r8988, %r8986, %r8987; xor.b32 %r8989, %r8988, %r8939; shf.l.wrap.b32 %r8990, %r8988, %r8988, 11; sub.s32 %r8991, %r8989, %r8990; xor.b32 %r8992, %r8991, %r8985; shf.l.wrap.b32 %r8993, %r8991, %r8991, 25; sub.s32 %r8994, %r8992, %r8993; xor.b32 %r8995, %r8994, %r8988; shf.l.wrap.b32 %r8996, %r8994, %r8994, 16; sub.s32 %r8997, %r8995, %r8996; xor.b32 %r8998, %r8997, %r8991; shf.l.wrap.b32 %r8999, %r8997, %r8997, 4; sub.s32 %r9000, %r8998, %r8999; xor.b32 %r9001, %r9000, %r8994; shf.l.wrap.b32 %r9002, %r9000, %r9000, 14; sub.s32 %r9003, %r9001, %r9002; xor.b32 %r9004, %r9003, %r8997; shf.l.wrap.b32 %r9005, %r9003, %r9003, 24; sub.s32 %r9006, %r9004, %r9005; add.ftz.f32 %f4877, %f4848, 0fBF800000; and.b32 %r9007, %r9006, 4; setp.eq.s32 %p1779, %r9007, 0; selp.f32 %f4878, %f4846, %f4877, %p1779; selp.f32 %f4879, %f4877, %f4846, %p1779; add.ftz.f32 %f4880, %f4879, %f4879; and.b32 %r9008, %r9006, 1; setp.eq.b32 %p1780, %r9008, 1; neg.ftz.f32 %f4881, %f4878; selp.f32 %f4882, %f4881, %f4878, %p1780; and.b32 %r9009, %r9006, 2; setp.eq.s32 %p1781, %r9009, 0; neg.ftz.f32 %f4883, %f4880; selp.f32 %f4884, %f4880, %f4883, %p1781; add.ftz.f32 %f4885, %f4882, %f4884; xor.b32 %r9010, %r8988, %r8964; sub.s32 %r9011, %r9010, %r8990; xor.b32 %r9012, %r9011, %r8985; shf.l.wrap.b32 %r9013, %r9011, %r9011, 25; sub.s32 %r9014, %r9012, %r9013; xor.b32 %r9015, %r9014, %r8988; shf.l.wrap.b32 %r9016, %r9014, %r9014, 16; sub.s32 %r9017, %r9015, %r9016; xor.b32 %r9018, %r9017, %r9011; shf.l.wrap.b32 %r9019, %r9017, %r9017, 4; sub.s32 %r9020, %r9018, %r9019; xor.b32 %r9021, %r9020, %r9014; shf.l.wrap.b32 %r9022, %r9020, %r9020, 14; sub.s32 %r9023, %r9021, %r9022; xor.b32 %r9024, %r9023, %r9017; shf.l.wrap.b32 %r9025, %r9023, %r9023, 24; sub.s32 %r9026, %r9024, %r9025; and.b32 %r9027, %r9026, 4; setp.eq.s32 %p1782, %r9027, 0; selp.f32 %f4886, %f4867, %f4877, %p1782; selp.f32 %f4887, %f4877, %f4867, %p1782; add.ftz.f32 %f4888, %f4887, %f4887; and.b32 %r9028, %r9026, 1; setp.eq.b32 %p1783, %r9028, 1; neg.ftz.f32 %f4889, %f4886; selp.f32 %f4890, %f4889, %f4886, %p1783; and.b32 %r9029, %r9026, 2; setp.eq.s32 %p1784, %r9029, 0; neg.ftz.f32 %f4891, %f4888; selp.f32 %f4892, %f4888, %f4891, %p1784; add.ftz.f32 %f4893, %f4890, %f4892; sub.ftz.f32 %f4894, %f4868, %f4853; sub.ftz.f32 %f4895, %f4868, %f4858; mul.ftz.f32 %f4896, %f4894, %f4866; fma.rn.ftz.f32 %f4897, %f4853, %f4876, %f4896; mul.ftz.f32 %f4898, %f4894, %f4885; fma.rn.ftz.f32 %f4899, %f4853, %f4893, %f4898; mul.ftz.f32 %f4900, %f4858, %f4899; fma.rn.ftz.f32 %f655, %f4895, %f4897, %f4900; mov.b32 %r445, %f655; setp.nan.ftz.f32 %p1785, %f655, %f655; mov.f32 %f5759, 0f00000000; @%p1785 bra $L__BB74_371; or.b32 %r9030, %r445, -2147483648; setp.ne.s32 %p1786, %r9030, -2147483648; add.ftz.f32 %f4902, %f655, %f655; setp.eq.ftz.f32 %p1787, %f4902, %f655; and.pred %p1788, %p1787, %p1786; @%p1788 bra $L__BB74_371; shl.b32 %r9031, %r445, 1; setp.lt.u32 %p1789, %r9031, -16777215; selp.f32 %f5759, %f655, 0f00000000, %p1789; $L__BB74_371: setp.gt.ftz.f32 %p1791, %f5764, 0f3F800000; selp.f32 %f4903, 0f3F800000, %f5764, %p1791; fma.rn.ftz.f32 %f4904, %f5759, 0f3F295E9E, %f21; mul.ftz.f32 %f4905, %f5762, %f4904; mul.ftz.f32 %f5762, %f649, %f5762; fma.rn.ftz.f32 %f5827, %f4903, %f4905, %f5827; mul.ftz.f32 %f4906, %f24, %f4905; mul.ftz.f32 %f5764, %f4903, %f4906; mul.ftz.f32 %f5771, %f27, %f5771; mul.ftz.f32 %f5772, %f27, %f5772; add.s32 %r10030, %r10030, 1; setp.gt.ftz.f32 %p1792, %f5764, 0f3A83126F; mov.pred %p1993, 0; @%p1792 bra $L__BB74_367; $L__BB74_372: cvt.rmi.ftz.f32.f32 %f4907, %f26; sub.ftz.f32 %f668, %f26, %f4907; setp.eq.ftz.f32 %p1793, %f668, 0f00000000; not.pred %p1794, %p1993; or.pred %p1795, %p1794, %p1793; @%p1795 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f4909, %f5771; cvt.rzi.ftz.s32.f32 %r9032, %f4909; sub.ftz.f32 %f4910, %f5771, %f4909; cvt.rmi.ftz.f32.f32 %f4911, %f5772; cvt.rzi.ftz.s32.f32 %r9033, %f4911; sub.ftz.f32 %f4912, %f5772, %f4911; mul.ftz.f32 %f4913, %f4910, %f4910; mul.ftz.f32 %f4914, %f4910, %f4913; fma.rn.ftz.f32 %f4915, %f4910, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4916, %f4910, %f4915, 0f41200000; mul.ftz.f32 %f4917, %f4914, %f4916; mul.ftz.f32 %f4918, %f4912, %f4912; mul.ftz.f32 %f4919, %f4912, %f4918; fma.rn.ftz.f32 %f4920, %f4912, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4921, %f4912, %f4920, 0f41200000; mul.ftz.f32 %f4922, %f4919, %f4921; add.s32 %r9034, %r9033, -559038716; add.s32 %r9035, %r9032, -559038716; xor.b32 %r9036, %r9034, -559038716; shf.l.wrap.b32 %r9037, %r9034, %r9034, 14; sub.s32 %r9038, %r9036, %r9037; xor.b32 %r9039, %r9038, %r9035; shf.l.wrap.b32 %r9040, %r9038, %r9038, 11; sub.s32 %r9041, %r9039, %r9040; xor.b32 %r9042, %r9041, %r9034; shf.l.wrap.b32 %r9043, %r9041, %r9041, 25; sub.s32 %r9044, %r9042, %r9043; xor.b32 %r9045, %r9044, %r9038; shf.l.wrap.b32 %r9046, %r9044, %r9044, 16; sub.s32 %r9047, %r9045, %r9046; xor.b32 %r9048, %r9047, %r9041; shf.l.wrap.b32 %r9049, %r9047, %r9047, 4; sub.s32 %r9050, %r9048, %r9049; xor.b32 %r9051, %r9050, %r9044; shf.l.wrap.b32 %r9052, %r9050, %r9050, 14; sub.s32 %r9053, %r9051, %r9052; xor.b32 %r9054, %r9053, %r9047; shf.l.wrap.b32 %r9055, %r9053, %r9053, 24; sub.s32 %r9056, %r9054, %r9055; and.b32 %r9057, %r9056, 4; setp.eq.s32 %p1796, %r9057, 0; selp.f32 %f4923, %f4910, %f4912, %p1796; selp.f32 %f4924, %f4912, %f4910, %p1796; add.ftz.f32 %f4925, %f4924, %f4924; and.b32 %r9058, %r9056, 1; setp.eq.b32 %p1797, %r9058, 1; neg.ftz.f32 %f4926, %f4923; selp.f32 %f4927, %f4926, %f4923, %p1797; and.b32 %r9059, %r9056, 2; setp.eq.s32 %p1798, %r9059, 0; neg.ftz.f32 %f4928, %f4925; selp.f32 %f4929, %f4925, %f4928, %p1798; add.ftz.f32 %f4930, %f4927, %f4929; add.s32 %r9060, %r9032, -559038715; xor.b32 %r9061, %r9038, %r9060; sub.s32 %r9062, %r9061, %r9040; xor.b32 %r9063, %r9062, %r9034; shf.l.wrap.b32 %r9064, %r9062, %r9062, 25; sub.s32 %r9065, %r9063, %r9064; xor.b32 %r9066, %r9065, %r9038; shf.l.wrap.b32 %r9067, %r9065, %r9065, 16; sub.s32 %r9068, %r9066, %r9067; xor.b32 %r9069, %r9068, %r9062; shf.l.wrap.b32 %r9070, %r9068, %r9068, 4; sub.s32 %r9071, %r9069, %r9070; xor.b32 %r9072, %r9071, %r9065; shf.l.wrap.b32 %r9073, %r9071, %r9071, 14; sub.s32 %r9074, %r9072, %r9073; xor.b32 %r9075, %r9074, %r9068; shf.l.wrap.b32 %r9076, %r9074, %r9074, 24; sub.s32 %r9077, %r9075, %r9076; add.ftz.f32 %f4931, %f4910, 0fBF800000; mov.f32 %f4932, 0f3F800000; and.b32 %r9078, %r9077, 4; setp.eq.s32 %p1799, %r9078, 0; selp.f32 %f4933, %f4931, %f4912, %p1799; selp.f32 %f4934, %f4912, %f4931, %p1799; add.ftz.f32 %f4935, %f4934, %f4934; and.b32 %r9079, %r9077, 1; setp.eq.b32 %p1800, %r9079, 1; neg.ftz.f32 %f4936, %f4933; selp.f32 %f4937, %f4936, %f4933, %p1800; and.b32 %r9080, %r9077, 2; setp.eq.s32 %p1801, %r9080, 0; neg.ftz.f32 %f4938, %f4935; selp.f32 %f4939, %f4935, %f4938, %p1801; add.ftz.f32 %f4940, %f4937, %f4939; add.s32 %r9081, %r9033, -559038715; xor.b32 %r9082, %r9081, -559038716; shf.l.wrap.b32 %r9083, %r9081, %r9081, 14; sub.s32 %r9084, %r9082, %r9083; xor.b32 %r9085, %r9084, %r9035; shf.l.wrap.b32 %r9086, %r9084, %r9084, 11; sub.s32 %r9087, %r9085, %r9086; xor.b32 %r9088, %r9087, %r9081; shf.l.wrap.b32 %r9089, %r9087, %r9087, 25; sub.s32 %r9090, %r9088, %r9089; xor.b32 %r9091, %r9090, %r9084; shf.l.wrap.b32 %r9092, %r9090, %r9090, 16; sub.s32 %r9093, %r9091, %r9092; xor.b32 %r9094, %r9093, %r9087; shf.l.wrap.b32 %r9095, %r9093, %r9093, 4; sub.s32 %r9096, %r9094, %r9095; xor.b32 %r9097, %r9096, %r9090; shf.l.wrap.b32 %r9098, %r9096, %r9096, 14; sub.s32 %r9099, %r9097, %r9098; xor.b32 %r9100, %r9099, %r9093; shf.l.wrap.b32 %r9101, %r9099, %r9099, 24; sub.s32 %r9102, %r9100, %r9101; add.ftz.f32 %f4941, %f4912, 0fBF800000; and.b32 %r9103, %r9102, 4; setp.eq.s32 %p1802, %r9103, 0; selp.f32 %f4942, %f4910, %f4941, %p1802; selp.f32 %f4943, %f4941, %f4910, %p1802; add.ftz.f32 %f4944, %f4943, %f4943; and.b32 %r9104, %r9102, 1; setp.eq.b32 %p1803, %r9104, 1; neg.ftz.f32 %f4945, %f4942; selp.f32 %f4946, %f4945, %f4942, %p1803; and.b32 %r9105, %r9102, 2; setp.eq.s32 %p1804, %r9105, 0; neg.ftz.f32 %f4947, %f4944; selp.f32 %f4948, %f4944, %f4947, %p1804; add.ftz.f32 %f4949, %f4946, %f4948; xor.b32 %r9106, %r9084, %r9060; sub.s32 %r9107, %r9106, %r9086; xor.b32 %r9108, %r9107, %r9081; shf.l.wrap.b32 %r9109, %r9107, %r9107, 25; sub.s32 %r9110, %r9108, %r9109; xor.b32 %r9111, %r9110, %r9084; shf.l.wrap.b32 %r9112, %r9110, %r9110, 16; sub.s32 %r9113, %r9111, %r9112; xor.b32 %r9114, %r9113, %r9107; shf.l.wrap.b32 %r9115, %r9113, %r9113, 4; sub.s32 %r9116, %r9114, %r9115; xor.b32 %r9117, %r9116, %r9110; shf.l.wrap.b32 %r9118, %r9116, %r9116, 14; sub.s32 %r9119, %r9117, %r9118; xor.b32 %r9120, %r9119, %r9113; shf.l.wrap.b32 %r9121, %r9119, %r9119, 24; sub.s32 %r9122, %r9120, %r9121; and.b32 %r9123, %r9122, 4; setp.eq.s32 %p1805, %r9123, 0; selp.f32 %f4950, %f4931, %f4941, %p1805; selp.f32 %f4951, %f4941, %f4931, %p1805; add.ftz.f32 %f4952, %f4951, %f4951; and.b32 %r9124, %r9122, 1; setp.eq.b32 %p1806, %r9124, 1; neg.ftz.f32 %f4953, %f4950; selp.f32 %f4954, %f4953, %f4950, %p1806; and.b32 %r9125, %r9122, 2; setp.eq.s32 %p1807, %r9125, 0; neg.ftz.f32 %f4955, %f4952; selp.f32 %f4956, %f4952, %f4955, %p1807; add.ftz.f32 %f4957, %f4954, %f4956; sub.ftz.f32 %f4958, %f4932, %f4917; sub.ftz.f32 %f4959, %f4932, %f4922; mul.ftz.f32 %f4960, %f4958, %f4930; fma.rn.ftz.f32 %f4961, %f4917, %f4940, %f4960; mul.ftz.f32 %f4962, %f4958, %f4949; fma.rn.ftz.f32 %f4963, %f4917, %f4957, %f4962; mul.ftz.f32 %f4964, %f4922, %f4963; fma.rn.ftz.f32 %f669, %f4959, %f4961, %f4964; mov.b32 %r447, %f669; setp.nan.ftz.f32 %p1808, %f669, %f669; mov.f32 %f5765, 0f00000000; @%p1808 bra $L__BB74_376; or.b32 %r9126, %r447, -2147483648; setp.ne.s32 %p1809, %r9126, -2147483648; add.ftz.f32 %f4966, %f669, %f669; setp.eq.ftz.f32 %p1810, %f4966, %f669; and.pred %p1811, %p1810, %p1809; @%p1811 bra $L__BB74_376; shl.b32 %r9127, %r447, 1; setp.lt.u32 %p1812, %r9127, -16777215; mul.ftz.f32 %f4967, %f669, 0f3F295E9E; selp.f32 %f5765, %f4967, 0f00000000, %p1812; $L__BB74_376: add.ftz.f32 %f4968, %f21, %f5765; mul.ftz.f32 %f4969, %f5762, %f4968; setp.gt.ftz.f32 %p1813, %f5764, 0f3F800000; selp.f32 %f4970, 0f3F800000, %f5764, %p1813; mul.ftz.f32 %f4971, %f4970, %f668; fma.rn.ftz.f32 %f5827, %f4971, %f4969, %f5827; bra.uni $L__BB74_450; $L__BB74_357: lg2.approx.ftz.f32 %f4709, %f27; mul.ftz.f32 %f4710, %f25, %f4709; neg.ftz.f32 %f4711, %f4710; ex2.approx.ftz.f32 %f631, %f4711; cvt.rmi.ftz.f32.f32 %f4712, %f5771; cvt.rzi.ftz.s32.f32 %r8742, %f4712; sub.ftz.f32 %f4713, %f5771, %f4712; cvt.rmi.ftz.f32.f32 %f4714, %f5772; cvt.rzi.ftz.s32.f32 %r8743, %f4714; sub.ftz.f32 %f4715, %f5772, %f4714; mul.ftz.f32 %f4716, %f4713, %f4713; mul.ftz.f32 %f4717, %f4713, %f4716; fma.rn.ftz.f32 %f4718, %f4713, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4719, %f4713, %f4718, 0f41200000; mul.ftz.f32 %f4720, %f4717, %f4719; mul.ftz.f32 %f4721, %f4715, %f4715; mul.ftz.f32 %f4722, %f4715, %f4721; fma.rn.ftz.f32 %f4723, %f4715, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4724, %f4715, %f4723, 0f41200000; mul.ftz.f32 %f4725, %f4722, %f4724; add.s32 %r8744, %r8743, -559038716; add.s32 %r8745, %r8742, -559038716; xor.b32 %r8746, %r8744, -559038716; shf.l.wrap.b32 %r8747, %r8744, %r8744, 14; sub.s32 %r8748, %r8746, %r8747; xor.b32 %r8749, %r8748, %r8745; shf.l.wrap.b32 %r8750, %r8748, %r8748, 11; sub.s32 %r8751, %r8749, %r8750; xor.b32 %r8752, %r8751, %r8744; shf.l.wrap.b32 %r8753, %r8751, %r8751, 25; sub.s32 %r8754, %r8752, %r8753; xor.b32 %r8755, %r8754, %r8748; shf.l.wrap.b32 %r8756, %r8754, %r8754, 16; sub.s32 %r8757, %r8755, %r8756; xor.b32 %r8758, %r8757, %r8751; shf.l.wrap.b32 %r8759, %r8757, %r8757, 4; sub.s32 %r8760, %r8758, %r8759; xor.b32 %r8761, %r8760, %r8754; shf.l.wrap.b32 %r8762, %r8760, %r8760, 14; sub.s32 %r8763, %r8761, %r8762; xor.b32 %r8764, %r8763, %r8757; shf.l.wrap.b32 %r8765, %r8763, %r8763, 24; sub.s32 %r8766, %r8764, %r8765; and.b32 %r8767, %r8766, 4; setp.eq.s32 %p1735, %r8767, 0; selp.f32 %f4726, %f4713, %f4715, %p1735; selp.f32 %f4727, %f4715, %f4713, %p1735; add.ftz.f32 %f4728, %f4727, %f4727; and.b32 %r8768, %r8766, 1; setp.eq.b32 %p1736, %r8768, 1; neg.ftz.f32 %f4729, %f4726; selp.f32 %f4730, %f4729, %f4726, %p1736; and.b32 %r8769, %r8766, 2; setp.eq.s32 %p1737, %r8769, 0; neg.ftz.f32 %f4731, %f4728; selp.f32 %f4732, %f4728, %f4731, %p1737; add.ftz.f32 %f4733, %f4730, %f4732; add.s32 %r8770, %r8742, -559038715; xor.b32 %r8771, %r8748, %r8770; sub.s32 %r8772, %r8771, %r8750; xor.b32 %r8773, %r8772, %r8744; shf.l.wrap.b32 %r8774, %r8772, %r8772, 25; sub.s32 %r8775, %r8773, %r8774; xor.b32 %r8776, %r8775, %r8748; shf.l.wrap.b32 %r8777, %r8775, %r8775, 16; sub.s32 %r8778, %r8776, %r8777; xor.b32 %r8779, %r8778, %r8772; shf.l.wrap.b32 %r8780, %r8778, %r8778, 4; sub.s32 %r8781, %r8779, %r8780; xor.b32 %r8782, %r8781, %r8775; shf.l.wrap.b32 %r8783, %r8781, %r8781, 14; sub.s32 %r8784, %r8782, %r8783; xor.b32 %r8785, %r8784, %r8778; shf.l.wrap.b32 %r8786, %r8784, %r8784, 24; sub.s32 %r8787, %r8785, %r8786; add.ftz.f32 %f4734, %f4713, 0fBF800000; mov.f32 %f4735, 0f3F800000; and.b32 %r8788, %r8787, 4; setp.eq.s32 %p1738, %r8788, 0; selp.f32 %f4736, %f4734, %f4715, %p1738; selp.f32 %f4737, %f4715, %f4734, %p1738; add.ftz.f32 %f4738, %f4737, %f4737; and.b32 %r8789, %r8787, 1; setp.eq.b32 %p1739, %r8789, 1; neg.ftz.f32 %f4739, %f4736; selp.f32 %f4740, %f4739, %f4736, %p1739; and.b32 %r8790, %r8787, 2; setp.eq.s32 %p1740, %r8790, 0; neg.ftz.f32 %f4741, %f4738; selp.f32 %f4742, %f4738, %f4741, %p1740; add.ftz.f32 %f4743, %f4740, %f4742; add.s32 %r8791, %r8743, -559038715; xor.b32 %r8792, %r8791, -559038716; shf.l.wrap.b32 %r8793, %r8791, %r8791, 14; sub.s32 %r8794, %r8792, %r8793; xor.b32 %r8795, %r8794, %r8745; shf.l.wrap.b32 %r8796, %r8794, %r8794, 11; sub.s32 %r8797, %r8795, %r8796; xor.b32 %r8798, %r8797, %r8791; shf.l.wrap.b32 %r8799, %r8797, %r8797, 25; sub.s32 %r8800, %r8798, %r8799; xor.b32 %r8801, %r8800, %r8794; shf.l.wrap.b32 %r8802, %r8800, %r8800, 16; sub.s32 %r8803, %r8801, %r8802; xor.b32 %r8804, %r8803, %r8797; shf.l.wrap.b32 %r8805, %r8803, %r8803, 4; sub.s32 %r8806, %r8804, %r8805; xor.b32 %r8807, %r8806, %r8800; shf.l.wrap.b32 %r8808, %r8806, %r8806, 14; sub.s32 %r8809, %r8807, %r8808; xor.b32 %r8810, %r8809, %r8803; shf.l.wrap.b32 %r8811, %r8809, %r8809, 24; sub.s32 %r8812, %r8810, %r8811; add.ftz.f32 %f4744, %f4715, 0fBF800000; and.b32 %r8813, %r8812, 4; setp.eq.s32 %p1741, %r8813, 0; selp.f32 %f4745, %f4713, %f4744, %p1741; selp.f32 %f4746, %f4744, %f4713, %p1741; add.ftz.f32 %f4747, %f4746, %f4746; and.b32 %r8814, %r8812, 1; setp.eq.b32 %p1742, %r8814, 1; neg.ftz.f32 %f4748, %f4745; selp.f32 %f4749, %f4748, %f4745, %p1742; and.b32 %r8815, %r8812, 2; setp.eq.s32 %p1743, %r8815, 0; neg.ftz.f32 %f4750, %f4747; selp.f32 %f4751, %f4747, %f4750, %p1743; add.ftz.f32 %f4752, %f4749, %f4751; xor.b32 %r8816, %r8794, %r8770; sub.s32 %r8817, %r8816, %r8796; xor.b32 %r8818, %r8817, %r8791; shf.l.wrap.b32 %r8819, %r8817, %r8817, 25; sub.s32 %r8820, %r8818, %r8819; xor.b32 %r8821, %r8820, %r8794; shf.l.wrap.b32 %r8822, %r8820, %r8820, 16; sub.s32 %r8823, %r8821, %r8822; xor.b32 %r8824, %r8823, %r8817; shf.l.wrap.b32 %r8825, %r8823, %r8823, 4; sub.s32 %r8826, %r8824, %r8825; xor.b32 %r8827, %r8826, %r8820; shf.l.wrap.b32 %r8828, %r8826, %r8826, 14; sub.s32 %r8829, %r8827, %r8828; xor.b32 %r8830, %r8829, %r8823; shf.l.wrap.b32 %r8831, %r8829, %r8829, 24; sub.s32 %r8832, %r8830, %r8831; and.b32 %r8833, %r8832, 4; setp.eq.s32 %p1744, %r8833, 0; selp.f32 %f4753, %f4734, %f4744, %p1744; selp.f32 %f4754, %f4744, %f4734, %p1744; add.ftz.f32 %f4755, %f4754, %f4754; and.b32 %r8834, %r8832, 1; setp.eq.b32 %p1745, %r8834, 1; neg.ftz.f32 %f4756, %f4753; selp.f32 %f4757, %f4756, %f4753, %p1745; and.b32 %r8835, %r8832, 2; setp.eq.s32 %p1746, %r8835, 0; neg.ftz.f32 %f4758, %f4755; selp.f32 %f4759, %f4755, %f4758, %p1746; add.ftz.f32 %f4760, %f4757, %f4759; sub.ftz.f32 %f4761, %f4735, %f4720; sub.ftz.f32 %f4762, %f4735, %f4725; mul.ftz.f32 %f4763, %f4761, %f4733; fma.rn.ftz.f32 %f4764, %f4720, %f4743, %f4763; mul.ftz.f32 %f4765, %f4761, %f4752; fma.rn.ftz.f32 %f4766, %f4720, %f4760, %f4765; mul.ftz.f32 %f4767, %f4725, %f4766; fma.rn.ftz.f32 %f632, %f4762, %f4764, %f4767; mov.b32 %r438, %f632; setp.nan.ftz.f32 %p1747, %f632, %f632; mov.f32 %f5747, 0f00000000; @%p1747 bra $L__BB74_360; or.b32 %r8836, %r438, -2147483648; setp.ne.s32 %p1748, %r8836, -2147483648; add.ftz.f32 %f4769, %f632, %f632; setp.eq.ftz.f32 %p1749, %f4769, %f632; and.pred %p1750, %p1749, %p1748; @%p1750 bra $L__BB74_360; shl.b32 %r8837, %r438, 1; setp.lt.u32 %p1751, %r8837, -16777215; mul.ftz.f32 %f4770, %f632, 0f3F295E9E; selp.f32 %f5747, %f4770, 0f00000000, %p1751; $L__BB74_360: abs.ftz.f32 %f4771, %f5747; sub.ftz.f32 %f4772, %f21, %f4771; mul.ftz.f32 %f5827, %f4772, %f4772; cvt.rzi.ftz.s32.f32 %r439, %f26; setp.lt.s32 %p1752, %r439, 2; @%p1752 bra $L__BB74_450; mov.u32 %r10029, 1; mov.f32 %f5749, %f5827; mov.f32 %f5750, %f631; $L__BB74_362: mul.ftz.f32 %f5771, %f27, %f5771; cvt.rmi.ftz.f32.f32 %f4774, %f5771; cvt.rzi.ftz.s32.f32 %r8839, %f4774; sub.ftz.f32 %f4775, %f5771, %f4774; mul.ftz.f32 %f5772, %f27, %f5772; cvt.rmi.ftz.f32.f32 %f4776, %f5772; cvt.rzi.ftz.s32.f32 %r8840, %f4776; sub.ftz.f32 %f4777, %f5772, %f4776; mul.ftz.f32 %f4778, %f4775, %f4775; mul.ftz.f32 %f4779, %f4775, %f4778; fma.rn.ftz.f32 %f4780, %f4775, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4781, %f4775, %f4780, 0f41200000; mul.ftz.f32 %f4782, %f4779, %f4781; mul.ftz.f32 %f4783, %f4777, %f4777; mul.ftz.f32 %f4784, %f4777, %f4783; fma.rn.ftz.f32 %f4785, %f4777, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f4786, %f4777, %f4785, 0f41200000; mul.ftz.f32 %f4787, %f4784, %f4786; add.s32 %r8841, %r8840, -559038716; add.s32 %r8842, %r8839, -559038716; xor.b32 %r8843, %r8841, -559038716; shf.l.wrap.b32 %r8844, %r8841, %r8841, 14; sub.s32 %r8845, %r8843, %r8844; xor.b32 %r8846, %r8845, %r8842; shf.l.wrap.b32 %r8847, %r8845, %r8845, 11; sub.s32 %r8848, %r8846, %r8847; xor.b32 %r8849, %r8848, %r8841; shf.l.wrap.b32 %r8850, %r8848, %r8848, 25; sub.s32 %r8851, %r8849, %r8850; xor.b32 %r8852, %r8851, %r8845; shf.l.wrap.b32 %r8853, %r8851, %r8851, 16; sub.s32 %r8854, %r8852, %r8853; xor.b32 %r8855, %r8854, %r8848; shf.l.wrap.b32 %r8856, %r8854, %r8854, 4; sub.s32 %r8857, %r8855, %r8856; xor.b32 %r8858, %r8857, %r8851; shf.l.wrap.b32 %r8859, %r8857, %r8857, 14; sub.s32 %r8860, %r8858, %r8859; xor.b32 %r8861, %r8860, %r8854; shf.l.wrap.b32 %r8862, %r8860, %r8860, 24; sub.s32 %r8863, %r8861, %r8862; and.b32 %r8864, %r8863, 4; setp.eq.s32 %p1753, %r8864, 0; selp.f32 %f4788, %f4775, %f4777, %p1753; selp.f32 %f4789, %f4777, %f4775, %p1753; add.ftz.f32 %f4790, %f4789, %f4789; and.b32 %r8865, %r8863, 1; setp.eq.b32 %p1754, %r8865, 1; neg.ftz.f32 %f4791, %f4788; selp.f32 %f4792, %f4791, %f4788, %p1754; and.b32 %r8866, %r8863, 2; setp.eq.s32 %p1755, %r8866, 0; neg.ftz.f32 %f4793, %f4790; selp.f32 %f4794, %f4790, %f4793, %p1755; add.ftz.f32 %f4795, %f4792, %f4794; add.s32 %r8867, %r8839, -559038715; xor.b32 %r8868, %r8845, %r8867; sub.s32 %r8869, %r8868, %r8847; xor.b32 %r8870, %r8869, %r8841; shf.l.wrap.b32 %r8871, %r8869, %r8869, 25; sub.s32 %r8872, %r8870, %r8871; xor.b32 %r8873, %r8872, %r8845; shf.l.wrap.b32 %r8874, %r8872, %r8872, 16; sub.s32 %r8875, %r8873, %r8874; xor.b32 %r8876, %r8875, %r8869; shf.l.wrap.b32 %r8877, %r8875, %r8875, 4; sub.s32 %r8878, %r8876, %r8877; xor.b32 %r8879, %r8878, %r8872; shf.l.wrap.b32 %r8880, %r8878, %r8878, 14; sub.s32 %r8881, %r8879, %r8880; xor.b32 %r8882, %r8881, %r8875; shf.l.wrap.b32 %r8883, %r8881, %r8881, 24; sub.s32 %r8884, %r8882, %r8883; add.ftz.f32 %f4796, %f4775, 0fBF800000; and.b32 %r8885, %r8884, 4; setp.eq.s32 %p1756, %r8885, 0; selp.f32 %f4798, %f4796, %f4777, %p1756; selp.f32 %f4799, %f4777, %f4796, %p1756; add.ftz.f32 %f4800, %f4799, %f4799; and.b32 %r8886, %r8884, 1; setp.eq.b32 %p1757, %r8886, 1; neg.ftz.f32 %f4801, %f4798; selp.f32 %f4802, %f4801, %f4798, %p1757; and.b32 %r8887, %r8884, 2; setp.eq.s32 %p1758, %r8887, 0; neg.ftz.f32 %f4803, %f4800; selp.f32 %f4804, %f4800, %f4803, %p1758; add.ftz.f32 %f4805, %f4802, %f4804; add.s32 %r8888, %r8840, -559038715; xor.b32 %r8889, %r8888, -559038716; shf.l.wrap.b32 %r8890, %r8888, %r8888, 14; sub.s32 %r8891, %r8889, %r8890; xor.b32 %r8892, %r8891, %r8842; shf.l.wrap.b32 %r8893, %r8891, %r8891, 11; sub.s32 %r8894, %r8892, %r8893; xor.b32 %r8895, %r8894, %r8888; shf.l.wrap.b32 %r8896, %r8894, %r8894, 25; sub.s32 %r8897, %r8895, %r8896; xor.b32 %r8898, %r8897, %r8891; shf.l.wrap.b32 %r8899, %r8897, %r8897, 16; sub.s32 %r8900, %r8898, %r8899; xor.b32 %r8901, %r8900, %r8894; shf.l.wrap.b32 %r8902, %r8900, %r8900, 4; sub.s32 %r8903, %r8901, %r8902; xor.b32 %r8904, %r8903, %r8897; shf.l.wrap.b32 %r8905, %r8903, %r8903, 14; sub.s32 %r8906, %r8904, %r8905; xor.b32 %r8907, %r8906, %r8900; shf.l.wrap.b32 %r8908, %r8906, %r8906, 24; sub.s32 %r8909, %r8907, %r8908; add.ftz.f32 %f4806, %f4777, 0fBF800000; and.b32 %r8910, %r8909, 4; setp.eq.s32 %p1759, %r8910, 0; selp.f32 %f4807, %f4775, %f4806, %p1759; selp.f32 %f4808, %f4806, %f4775, %p1759; add.ftz.f32 %f4809, %f4808, %f4808; and.b32 %r8911, %r8909, 1; setp.eq.b32 %p1760, %r8911, 1; neg.ftz.f32 %f4810, %f4807; selp.f32 %f4811, %f4810, %f4807, %p1760; and.b32 %r8912, %r8909, 2; setp.eq.s32 %p1761, %r8912, 0; neg.ftz.f32 %f4812, %f4809; selp.f32 %f4813, %f4809, %f4812, %p1761; add.ftz.f32 %f4814, %f4811, %f4813; xor.b32 %r8913, %r8891, %r8867; sub.s32 %r8914, %r8913, %r8893; xor.b32 %r8915, %r8914, %r8888; shf.l.wrap.b32 %r8916, %r8914, %r8914, 25; sub.s32 %r8917, %r8915, %r8916; xor.b32 %r8918, %r8917, %r8891; shf.l.wrap.b32 %r8919, %r8917, %r8917, 16; sub.s32 %r8920, %r8918, %r8919; xor.b32 %r8921, %r8920, %r8914; shf.l.wrap.b32 %r8922, %r8920, %r8920, 4; sub.s32 %r8923, %r8921, %r8922; xor.b32 %r8924, %r8923, %r8917; shf.l.wrap.b32 %r8925, %r8923, %r8923, 14; sub.s32 %r8926, %r8924, %r8925; xor.b32 %r8927, %r8926, %r8920; shf.l.wrap.b32 %r8928, %r8926, %r8926, 24; sub.s32 %r8929, %r8927, %r8928; and.b32 %r8930, %r8929, 4; setp.eq.s32 %p1762, %r8930, 0; selp.f32 %f4815, %f4796, %f4806, %p1762; selp.f32 %f4816, %f4806, %f4796, %p1762; add.ftz.f32 %f4817, %f4816, %f4816; and.b32 %r8931, %r8929, 1; setp.eq.b32 %p1763, %r8931, 1; neg.ftz.f32 %f4818, %f4815; selp.f32 %f4819, %f4818, %f4815, %p1763; and.b32 %r8932, %r8929, 2; setp.eq.s32 %p1764, %r8932, 0; neg.ftz.f32 %f4820, %f4817; selp.f32 %f4821, %f4817, %f4820, %p1764; add.ftz.f32 %f4822, %f4819, %f4821; sub.ftz.f32 %f4823, %f4735, %f4782; sub.ftz.f32 %f4824, %f4735, %f4787; mul.ftz.f32 %f4825, %f4823, %f4795; fma.rn.ftz.f32 %f4826, %f4782, %f4805, %f4825; mul.ftz.f32 %f4827, %f4823, %f4814; fma.rn.ftz.f32 %f4828, %f4782, %f4822, %f4827; mul.ftz.f32 %f4829, %f4787, %f4828; fma.rn.ftz.f32 %f643, %f4824, %f4826, %f4829; mov.b32 %r441, %f643; setp.nan.ftz.f32 %p1765, %f643, %f643; mov.f32 %f5753, 0f00000000; @%p1765 bra $L__BB74_365; or.b32 %r8933, %r441, -2147483648; setp.ne.s32 %p1766, %r8933, -2147483648; add.ftz.f32 %f4831, %f643, %f643; setp.eq.ftz.f32 %p1767, %f4831, %f643; and.pred %p1768, %p1767, %p1766; @%p1768 bra $L__BB74_365; shl.b32 %r8934, %r441, 1; setp.lt.u32 %p1769, %r8934, -16777215; selp.f32 %f5753, %f643, 0f00000000, %p1769; $L__BB74_365: mul.ftz.f32 %f4832, %f5753, 0f3F295E9E; abs.ftz.f32 %f4833, %f4832; sub.ftz.f32 %f4834, %f21, %f4833; mul.ftz.f32 %f4835, %f4834, %f4834; mul.ftz.f32 %f4836, %f24, %f5749; cvt.ftz.sat.f32.f32 %f4837, %f4836; mul.ftz.f32 %f5749, %f4837, %f4835; fma.rn.ftz.f32 %f5827, %f5750, %f5749, %f5827; mul.ftz.f32 %f5750, %f631, %f5750; add.s32 %r10029, %r10029, 1; setp.lt.s32 %p1770, %r10029, %r439; @%p1770 bra $L__BB74_362; bra.uni $L__BB74_450; $L__BB74_80: lg2.approx.ftz.f32 %f2066, %f27; mul.ftz.f32 %f2067, %f25, %f2066; neg.ftz.f32 %f2068, %f2067; ex2.approx.ftz.f32 %f89, %f2068; cvt.rzi.ftz.s32.f32 %r27, %f26; mov.f32 %f5559, 0f3F800000; mov.f32 %f5827, 0f00000000; mov.u32 %r10020, 0; mov.pred %p533, -1; mov.pred %p636, 0; mov.f32 %f5557, %f5559; $L__BB74_81: setp.ge.s32 %p534, %r10020, %r27; mov.pred %p1991, %p533; @%p534 bra $L__BB74_86; cvt.rmi.ftz.f32.f32 %f2070, %f5568; cvt.rzi.ftz.s32.f32 %r3330, %f2070; sub.ftz.f32 %f2071, %f5568, %f2070; cvt.rmi.ftz.f32.f32 %f2072, %f5569; cvt.rzi.ftz.s32.f32 %r3331, %f2072; sub.ftz.f32 %f2073, %f5569, %f2072; cvt.rmi.ftz.f32.f32 %f2074, %f5570; cvt.rzi.ftz.s32.f32 %r3332, %f2074; sub.ftz.f32 %f2075, %f5570, %f2074; cvt.rmi.ftz.f32.f32 %f2076, %f5571; cvt.rzi.ftz.s32.f32 %r3333, %f2076; sub.ftz.f32 %f2077, %f5571, %f2076; mul.ftz.f32 %f2078, %f2071, %f2071; mul.ftz.f32 %f2079, %f2071, %f2078; fma.rn.ftz.f32 %f2080, %f2071, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2081, %f2071, %f2080, 0f41200000; mul.ftz.f32 %f2082, %f2079, %f2081; mul.ftz.f32 %f2083, %f2073, %f2073; mul.ftz.f32 %f2084, %f2073, %f2083; fma.rn.ftz.f32 %f2085, %f2073, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2086, %f2073, %f2085, 0f41200000; mul.ftz.f32 %f2087, %f2084, %f2086; mul.ftz.f32 %f2088, %f2075, %f2075; mul.ftz.f32 %f2089, %f2075, %f2088; fma.rn.ftz.f32 %f2090, %f2075, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2091, %f2075, %f2090, 0f41200000; mul.ftz.f32 %f2092, %f2089, %f2091; mul.ftz.f32 %f2093, %f2077, %f2077; mul.ftz.f32 %f2094, %f2077, %f2093; fma.rn.ftz.f32 %f2095, %f2077, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2096, %f2077, %f2095, 0f41200000; mul.ftz.f32 %f2097, %f2094, %f2096; add.s32 %r3334, %r3331, -559038708; add.s32 %r3335, %r3332, -559038708; sub.s32 %r3336, %r3330, %r3332; shf.l.wrap.b32 %r3337, %r3335, %r3335, 4; xor.b32 %r3338, %r3337, %r3336; add.s32 %r3339, %r3335, %r3334; sub.s32 %r3340, %r3334, %r3338; shf.l.wrap.b32 %r3341, %r3338, %r3338, 6; xor.b32 %r3342, %r3341, %r3340; add.s32 %r3343, %r3338, %r3339; sub.s32 %r3344, %r3339, %r3342; shf.l.wrap.b32 %r3345, %r3342, %r3342, 8; xor.b32 %r3346, %r3345, %r3344; add.s32 %r3347, %r3342, %r3343; sub.s32 %r3348, %r3343, %r3346; shf.l.wrap.b32 %r3349, %r3346, %r3346, 16; xor.b32 %r3350, %r3349, %r3348; add.s32 %r3351, %r3346, %r3347; sub.s32 %r3352, %r3347, %r3350; shf.l.wrap.b32 %r3353, %r3350, %r3350, 19; xor.b32 %r3354, %r3353, %r3352; add.s32 %r3355, %r3350, %r3351; sub.s32 %r3356, %r3351, %r3354; shf.l.wrap.b32 %r3357, %r3354, %r3354, 4; add.s32 %r3358, %r3354, %r3355; add.s32 %r3359, %r3355, %r3333; xor.b32 %r3360, %r3358, %r3356; xor.b32 %r3361, %r3360, %r3357; shf.l.wrap.b32 %r3362, %r3358, %r3358, 14; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3359; shf.l.wrap.b32 %r3365, %r3363, %r3363, 11; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3358; shf.l.wrap.b32 %r3368, %r3366, %r3366, 25; sub.s32 %r3369, %r3367, %r3368; xor.b32 %r3370, %r3369, %r3363; shf.l.wrap.b32 %r3371, %r3369, %r3369, 16; sub.s32 %r3372, %r3370, %r3371; xor.b32 %r3373, %r3372, %r3366; shf.l.wrap.b32 %r3374, %r3372, %r3372, 4; sub.s32 %r3375, %r3373, %r3374; xor.b32 %r3376, %r3375, %r3369; shf.l.wrap.b32 %r3377, %r3375, %r3375, 14; sub.s32 %r3378, %r3376, %r3377; xor.b32 %r3379, %r3378, %r3372; shf.l.wrap.b32 %r3380, %r3378, %r3378, 24; sub.s32 %r3381, %r3379, %r3380; and.b32 %r3382, %r3381, 31; setp.lt.u32 %p535, %r3382, 24; selp.f32 %f2098, %f2071, %f2073, %p535; setp.lt.u32 %p536, %r3382, 16; selp.f32 %f2099, %f2073, %f2075, %p536; setp.lt.u32 %p537, %r3382, 8; selp.f32 %f2100, %f2075, %f2077, %p537; and.b32 %r3383, %r3381, 1; setp.eq.b32 %p538, %r3383, 1; neg.ftz.f32 %f2101, %f2098; selp.f32 %f2102, %f2101, %f2098, %p538; and.b32 %r3384, %r3381, 2; setp.eq.s32 %p539, %r3384, 0; neg.ftz.f32 %f2103, %f2099; selp.f32 %f2104, %f2099, %f2103, %p539; add.ftz.f32 %f2105, %f2102, %f2104; and.b32 %r3385, %r3381, 4; setp.eq.s32 %p540, %r3385, 0; neg.ftz.f32 %f2106, %f2100; selp.f32 %f2107, %f2100, %f2106, %p540; add.ftz.f32 %f2108, %f2107, %f2105; add.s32 %r3386, %r3330, 1; sub.s32 %r3387, %r3386, %r3332; xor.b32 %r3388, %r3337, %r3387; sub.s32 %r3389, %r3334, %r3388; shf.l.wrap.b32 %r3390, %r3388, %r3388, 6; xor.b32 %r3391, %r3390, %r3389; add.s32 %r3392, %r3388, %r3339; sub.s32 %r3393, %r3339, %r3391; shf.l.wrap.b32 %r3394, %r3391, %r3391, 8; xor.b32 %r3395, %r3394, %r3393; add.s32 %r3396, %r3391, %r3392; sub.s32 %r3397, %r3392, %r3395; shf.l.wrap.b32 %r3398, %r3395, %r3395, 16; xor.b32 %r3399, %r3398, %r3397; add.s32 %r3400, %r3395, %r3396; sub.s32 %r3401, %r3396, %r3399; shf.l.wrap.b32 %r3402, %r3399, %r3399, 19; xor.b32 %r3403, %r3402, %r3401; add.s32 %r3404, %r3399, %r3400; sub.s32 %r3405, %r3400, %r3403; shf.l.wrap.b32 %r3406, %r3403, %r3403, 4; add.s32 %r3407, %r3403, %r3404; add.s32 %r3408, %r3404, %r3333; xor.b32 %r3409, %r3407, %r3405; xor.b32 %r3410, %r3409, %r3406; shf.l.wrap.b32 %r3411, %r3407, %r3407, 14; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3408; shf.l.wrap.b32 %r3414, %r3412, %r3412, 11; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3407; shf.l.wrap.b32 %r3417, %r3415, %r3415, 25; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 16; sub.s32 %r3421, %r3419, %r3420; xor.b32 %r3422, %r3421, %r3415; shf.l.wrap.b32 %r3423, %r3421, %r3421, 4; sub.s32 %r3424, %r3422, %r3423; xor.b32 %r3425, %r3424, %r3418; shf.l.wrap.b32 %r3426, %r3424, %r3424, 14; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r3421; shf.l.wrap.b32 %r3429, %r3427, %r3427, 24; sub.s32 %r3430, %r3428, %r3429; add.ftz.f32 %f2109, %f2071, 0fBF800000; mov.f32 %f2110, 0f3F800000; and.b32 %r3431, %r3430, 31; setp.lt.u32 %p541, %r3431, 24; selp.f32 %f2111, %f2109, %f2073, %p541; setp.lt.u32 %p542, %r3431, 16; selp.f32 %f2112, %f2073, %f2075, %p542; setp.lt.u32 %p543, %r3431, 8; selp.f32 %f2113, %f2075, %f2077, %p543; and.b32 %r3432, %r3430, 1; setp.eq.b32 %p544, %r3432, 1; neg.ftz.f32 %f2114, %f2111; selp.f32 %f2115, %f2114, %f2111, %p544; and.b32 %r3433, %r3430, 2; setp.eq.s32 %p545, %r3433, 0; neg.ftz.f32 %f2116, %f2112; selp.f32 %f2117, %f2112, %f2116, %p545; add.ftz.f32 %f2118, %f2115, %f2117; and.b32 %r3434, %r3430, 4; setp.eq.s32 %p546, %r3434, 0; neg.ftz.f32 %f2119, %f2113; selp.f32 %f2120, %f2113, %f2119, %p546; add.ftz.f32 %f2121, %f2120, %f2118; add.s32 %r3435, %r3331, -559038707; add.s32 %r3436, %r3335, %r3435; sub.s32 %r3437, %r3435, %r3338; xor.b32 %r3438, %r3341, %r3437; add.s32 %r3439, %r3338, %r3436; sub.s32 %r3440, %r3436, %r3438; shf.l.wrap.b32 %r3441, %r3438, %r3438, 8; xor.b32 %r3442, %r3441, %r3440; add.s32 %r3443, %r3438, %r3439; sub.s32 %r3444, %r3439, %r3442; shf.l.wrap.b32 %r3445, %r3442, %r3442, 16; xor.b32 %r3446, %r3445, %r3444; add.s32 %r3447, %r3442, %r3443; sub.s32 %r3448, %r3443, %r3446; shf.l.wrap.b32 %r3449, %r3446, %r3446, 19; xor.b32 %r3450, %r3449, %r3448; add.s32 %r3451, %r3446, %r3447; sub.s32 %r3452, %r3447, %r3450; shf.l.wrap.b32 %r3453, %r3450, %r3450, 4; add.s32 %r3454, %r3450, %r3451; add.s32 %r3455, %r3451, %r3333; xor.b32 %r3456, %r3454, %r3452; xor.b32 %r3457, %r3456, %r3453; shf.l.wrap.b32 %r3458, %r3454, %r3454, 14; sub.s32 %r3459, %r3457, %r3458; xor.b32 %r3460, %r3459, %r3455; shf.l.wrap.b32 %r3461, %r3459, %r3459, 11; sub.s32 %r3462, %r3460, %r3461; xor.b32 %r3463, %r3462, %r3454; shf.l.wrap.b32 %r3464, %r3462, %r3462, 25; sub.s32 %r3465, %r3463, %r3464; xor.b32 %r3466, %r3465, %r3459; shf.l.wrap.b32 %r3467, %r3465, %r3465, 16; sub.s32 %r3468, %r3466, %r3467; xor.b32 %r3469, %r3468, %r3462; shf.l.wrap.b32 %r3470, %r3468, %r3468, 4; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r3465; shf.l.wrap.b32 %r3473, %r3471, %r3471, 14; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 24; sub.s32 %r3477, %r3475, %r3476; add.ftz.f32 %f2122, %f2073, 0fBF800000; and.b32 %r3478, %r3477, 31; setp.lt.u32 %p547, %r3478, 24; selp.f32 %f2123, %f2071, %f2122, %p547; setp.lt.u32 %p548, %r3478, 16; selp.f32 %f2124, %f2122, %f2075, %p548; setp.lt.u32 %p549, %r3478, 8; selp.f32 %f2125, %f2075, %f2077, %p549; and.b32 %r3479, %r3477, 1; setp.eq.b32 %p550, %r3479, 1; neg.ftz.f32 %f2126, %f2123; selp.f32 %f2127, %f2126, %f2123, %p550; and.b32 %r3480, %r3477, 2; setp.eq.s32 %p551, %r3480, 0; neg.ftz.f32 %f2128, %f2124; selp.f32 %f2129, %f2124, %f2128, %p551; add.ftz.f32 %f2130, %f2127, %f2129; and.b32 %r3481, %r3477, 4; setp.eq.s32 %p552, %r3481, 0; neg.ftz.f32 %f2131, %f2125; selp.f32 %f2132, %f2125, %f2131, %p552; add.ftz.f32 %f2133, %f2132, %f2130; sub.s32 %r3482, %r3435, %r3388; xor.b32 %r3483, %r3390, %r3482; add.s32 %r3484, %r3388, %r3436; sub.s32 %r3485, %r3436, %r3483; shf.l.wrap.b32 %r3486, %r3483, %r3483, 8; xor.b32 %r3487, %r3486, %r3485; add.s32 %r3488, %r3483, %r3484; sub.s32 %r3489, %r3484, %r3487; shf.l.wrap.b32 %r3490, %r3487, %r3487, 16; xor.b32 %r3491, %r3490, %r3489; add.s32 %r3492, %r3487, %r3488; sub.s32 %r3493, %r3488, %r3491; shf.l.wrap.b32 %r3494, %r3491, %r3491, 19; xor.b32 %r3495, %r3494, %r3493; add.s32 %r3496, %r3491, %r3492; sub.s32 %r3497, %r3492, %r3495; shf.l.wrap.b32 %r3498, %r3495, %r3495, 4; add.s32 %r3499, %r3495, %r3496; add.s32 %r3500, %r3496, %r3333; xor.b32 %r3501, %r3499, %r3497; xor.b32 %r3502, %r3501, %r3498; shf.l.wrap.b32 %r3503, %r3499, %r3499, 14; sub.s32 %r3504, %r3502, %r3503; xor.b32 %r3505, %r3504, %r3500; shf.l.wrap.b32 %r3506, %r3504, %r3504, 11; sub.s32 %r3507, %r3505, %r3506; xor.b32 %r3508, %r3507, %r3499; shf.l.wrap.b32 %r3509, %r3507, %r3507, 25; sub.s32 %r3510, %r3508, %r3509; xor.b32 %r3511, %r3510, %r3504; shf.l.wrap.b32 %r3512, %r3510, %r3510, 16; sub.s32 %r3513, %r3511, %r3512; xor.b32 %r3514, %r3513, %r3507; shf.l.wrap.b32 %r3515, %r3513, %r3513, 4; sub.s32 %r3516, %r3514, %r3515; xor.b32 %r3517, %r3516, %r3510; shf.l.wrap.b32 %r3518, %r3516, %r3516, 14; sub.s32 %r3519, %r3517, %r3518; xor.b32 %r3520, %r3519, %r3513; shf.l.wrap.b32 %r3521, %r3519, %r3519, 24; sub.s32 %r3522, %r3520, %r3521; and.b32 %r3523, %r3522, 31; setp.lt.u32 %p553, %r3523, 24; selp.f32 %f2134, %f2109, %f2122, %p553; setp.lt.u32 %p554, %r3523, 16; selp.f32 %f2135, %f2122, %f2075, %p554; setp.lt.u32 %p555, %r3523, 8; selp.f32 %f2136, %f2075, %f2077, %p555; and.b32 %r3524, %r3522, 1; setp.eq.b32 %p556, %r3524, 1; neg.ftz.f32 %f2137, %f2134; selp.f32 %f2138, %f2137, %f2134, %p556; and.b32 %r3525, %r3522, 2; setp.eq.s32 %p557, %r3525, 0; neg.ftz.f32 %f2139, %f2135; selp.f32 %f2140, %f2135, %f2139, %p557; add.ftz.f32 %f2141, %f2138, %f2140; and.b32 %r3526, %r3522, 4; setp.eq.s32 %p558, %r3526, 0; neg.ftz.f32 %f2142, %f2136; selp.f32 %f2143, %f2136, %f2142, %p558; add.ftz.f32 %f2144, %f2143, %f2141; not.b32 %r3527, %r3332; add.s32 %r3528, %r3332, -559038707; add.s32 %r3529, %r3330, %r3527; shf.l.wrap.b32 %r3530, %r3528, %r3528, 4; xor.b32 %r3531, %r3530, %r3529; add.s32 %r3532, %r3528, %r3334; sub.s32 %r3533, %r3334, %r3531; shf.l.wrap.b32 %r3534, %r3531, %r3531, 6; xor.b32 %r3535, %r3534, %r3533; add.s32 %r3536, %r3531, %r3532; sub.s32 %r3537, %r3532, %r3535; shf.l.wrap.b32 %r3538, %r3535, %r3535, 8; xor.b32 %r3539, %r3538, %r3537; add.s32 %r3540, %r3535, %r3536; sub.s32 %r3541, %r3536, %r3539; shf.l.wrap.b32 %r3542, %r3539, %r3539, 16; xor.b32 %r3543, %r3542, %r3541; add.s32 %r3544, %r3539, %r3540; sub.s32 %r3545, %r3540, %r3543; shf.l.wrap.b32 %r3546, %r3543, %r3543, 19; xor.b32 %r3547, %r3546, %r3545; add.s32 %r3548, %r3543, %r3544; sub.s32 %r3549, %r3544, %r3547; shf.l.wrap.b32 %r3550, %r3547, %r3547, 4; add.s32 %r3551, %r3547, %r3548; add.s32 %r3552, %r3548, %r3333; xor.b32 %r3553, %r3551, %r3549; xor.b32 %r3554, %r3553, %r3550; shf.l.wrap.b32 %r3555, %r3551, %r3551, 14; sub.s32 %r3556, %r3554, %r3555; xor.b32 %r3557, %r3556, %r3552; shf.l.wrap.b32 %r3558, %r3556, %r3556, 11; sub.s32 %r3559, %r3557, %r3558; xor.b32 %r3560, %r3559, %r3551; shf.l.wrap.b32 %r3561, %r3559, %r3559, 25; sub.s32 %r3562, %r3560, %r3561; xor.b32 %r3563, %r3562, %r3556; shf.l.wrap.b32 %r3564, %r3562, %r3562, 16; sub.s32 %r3565, %r3563, %r3564; xor.b32 %r3566, %r3565, %r3559; shf.l.wrap.b32 %r3567, %r3565, %r3565, 4; sub.s32 %r3568, %r3566, %r3567; xor.b32 %r3569, %r3568, %r3562; shf.l.wrap.b32 %r3570, %r3568, %r3568, 14; sub.s32 %r3571, %r3569, %r3570; xor.b32 %r3572, %r3571, %r3565; shf.l.wrap.b32 %r3573, %r3571, %r3571, 24; sub.s32 %r3574, %r3572, %r3573; add.ftz.f32 %f2145, %f2075, 0fBF800000; and.b32 %r3575, %r3574, 31; setp.lt.u32 %p559, %r3575, 24; selp.f32 %f2146, %f2071, %f2073, %p559; setp.lt.u32 %p560, %r3575, 16; selp.f32 %f2147, %f2073, %f2145, %p560; setp.lt.u32 %p561, %r3575, 8; selp.f32 %f2148, %f2145, %f2077, %p561; and.b32 %r3576, %r3574, 1; setp.eq.b32 %p562, %r3576, 1; neg.ftz.f32 %f2149, %f2146; selp.f32 %f2150, %f2149, %f2146, %p562; and.b32 %r3577, %r3574, 2; setp.eq.s32 %p563, %r3577, 0; neg.ftz.f32 %f2151, %f2147; selp.f32 %f2152, %f2147, %f2151, %p563; add.ftz.f32 %f2153, %f2150, %f2152; and.b32 %r3578, %r3574, 4; setp.eq.s32 %p564, %r3578, 0; neg.ftz.f32 %f2154, %f2148; selp.f32 %f2155, %f2148, %f2154, %p564; add.ftz.f32 %f2156, %f2155, %f2153; xor.b32 %r3579, %r3530, %r3336; sub.s32 %r3580, %r3334, %r3579; shf.l.wrap.b32 %r3581, %r3579, %r3579, 6; xor.b32 %r3582, %r3581, %r3580; add.s32 %r3583, %r3579, %r3532; sub.s32 %r3584, %r3532, %r3582; shf.l.wrap.b32 %r3585, %r3582, %r3582, 8; xor.b32 %r3586, %r3585, %r3584; add.s32 %r3587, %r3582, %r3583; sub.s32 %r3588, %r3583, %r3586; shf.l.wrap.b32 %r3589, %r3586, %r3586, 16; xor.b32 %r3590, %r3589, %r3588; add.s32 %r3591, %r3586, %r3587; sub.s32 %r3592, %r3587, %r3590; shf.l.wrap.b32 %r3593, %r3590, %r3590, 19; xor.b32 %r3594, %r3593, %r3592; add.s32 %r3595, %r3590, %r3591; sub.s32 %r3596, %r3591, %r3594; shf.l.wrap.b32 %r3597, %r3594, %r3594, 4; add.s32 %r3598, %r3594, %r3595; add.s32 %r3599, %r3595, %r3333; xor.b32 %r3600, %r3598, %r3596; xor.b32 %r3601, %r3600, %r3597; shf.l.wrap.b32 %r3602, %r3598, %r3598, 14; sub.s32 %r3603, %r3601, %r3602; xor.b32 %r3604, %r3603, %r3599; shf.l.wrap.b32 %r3605, %r3603, %r3603, 11; sub.s32 %r3606, %r3604, %r3605; xor.b32 %r3607, %r3606, %r3598; shf.l.wrap.b32 %r3608, %r3606, %r3606, 25; sub.s32 %r3609, %r3607, %r3608; xor.b32 %r3610, %r3609, %r3603; shf.l.wrap.b32 %r3611, %r3609, %r3609, 16; sub.s32 %r3612, %r3610, %r3611; xor.b32 %r3613, %r3612, %r3606; shf.l.wrap.b32 %r3614, %r3612, %r3612, 4; sub.s32 %r3615, %r3613, %r3614; xor.b32 %r3616, %r3615, %r3609; shf.l.wrap.b32 %r3617, %r3615, %r3615, 14; sub.s32 %r3618, %r3616, %r3617; xor.b32 %r3619, %r3618, %r3612; shf.l.wrap.b32 %r3620, %r3618, %r3618, 24; sub.s32 %r3621, %r3619, %r3620; and.b32 %r3622, %r3621, 31; setp.lt.u32 %p565, %r3622, 24; selp.f32 %f2157, %f2109, %f2073, %p565; setp.lt.u32 %p566, %r3622, 16; selp.f32 %f2158, %f2073, %f2145, %p566; setp.lt.u32 %p567, %r3622, 8; selp.f32 %f2159, %f2145, %f2077, %p567; and.b32 %r3623, %r3621, 1; setp.eq.b32 %p568, %r3623, 1; neg.ftz.f32 %f2160, %f2157; selp.f32 %f2161, %f2160, %f2157, %p568; and.b32 %r3624, %r3621, 2; setp.eq.s32 %p569, %r3624, 0; neg.ftz.f32 %f2162, %f2158; selp.f32 %f2163, %f2158, %f2162, %p569; add.ftz.f32 %f2164, %f2161, %f2163; and.b32 %r3625, %r3621, 4; setp.eq.s32 %p570, %r3625, 0; neg.ftz.f32 %f2165, %f2159; selp.f32 %f2166, %f2159, %f2165, %p570; add.ftz.f32 %f2167, %f2166, %f2164; add.s32 %r3626, %r3528, %r3435; sub.s32 %r3627, %r3435, %r3531; xor.b32 %r3628, %r3534, %r3627; add.s32 %r3629, %r3531, %r3626; sub.s32 %r3630, %r3626, %r3628; shf.l.wrap.b32 %r3631, %r3628, %r3628, 8; xor.b32 %r3632, %r3631, %r3630; add.s32 %r3633, %r3628, %r3629; sub.s32 %r3634, %r3629, %r3632; shf.l.wrap.b32 %r3635, %r3632, %r3632, 16; xor.b32 %r3636, %r3635, %r3634; add.s32 %r3637, %r3632, %r3633; sub.s32 %r3638, %r3633, %r3636; shf.l.wrap.b32 %r3639, %r3636, %r3636, 19; xor.b32 %r3640, %r3639, %r3638; add.s32 %r3641, %r3636, %r3637; sub.s32 %r3642, %r3637, %r3640; shf.l.wrap.b32 %r3643, %r3640, %r3640, 4; add.s32 %r3644, %r3640, %r3641; add.s32 %r3645, %r3641, %r3333; xor.b32 %r3646, %r3644, %r3642; xor.b32 %r3647, %r3646, %r3643; shf.l.wrap.b32 %r3648, %r3644, %r3644, 14; sub.s32 %r3649, %r3647, %r3648; xor.b32 %r3650, %r3649, %r3645; shf.l.wrap.b32 %r3651, %r3649, %r3649, 11; sub.s32 %r3652, %r3650, %r3651; xor.b32 %r3653, %r3652, %r3644; shf.l.wrap.b32 %r3654, %r3652, %r3652, 25; sub.s32 %r3655, %r3653, %r3654; xor.b32 %r3656, %r3655, %r3649; shf.l.wrap.b32 %r3657, %r3655, %r3655, 16; sub.s32 %r3658, %r3656, %r3657; xor.b32 %r3659, %r3658, %r3652; shf.l.wrap.b32 %r3660, %r3658, %r3658, 4; sub.s32 %r3661, %r3659, %r3660; xor.b32 %r3662, %r3661, %r3655; shf.l.wrap.b32 %r3663, %r3661, %r3661, 14; sub.s32 %r3664, %r3662, %r3663; xor.b32 %r3665, %r3664, %r3658; shf.l.wrap.b32 %r3666, %r3664, %r3664, 24; sub.s32 %r3667, %r3665, %r3666; and.b32 %r3668, %r3667, 31; setp.lt.u32 %p571, %r3668, 24; selp.f32 %f2168, %f2071, %f2122, %p571; setp.lt.u32 %p572, %r3668, 16; selp.f32 %f2169, %f2122, %f2145, %p572; setp.lt.u32 %p573, %r3668, 8; selp.f32 %f2170, %f2145, %f2077, %p573; and.b32 %r3669, %r3667, 1; setp.eq.b32 %p574, %r3669, 1; neg.ftz.f32 %f2171, %f2168; selp.f32 %f2172, %f2171, %f2168, %p574; and.b32 %r3670, %r3667, 2; setp.eq.s32 %p575, %r3670, 0; neg.ftz.f32 %f2173, %f2169; selp.f32 %f2174, %f2169, %f2173, %p575; add.ftz.f32 %f2175, %f2172, %f2174; and.b32 %r3671, %r3667, 4; setp.eq.s32 %p576, %r3671, 0; neg.ftz.f32 %f2176, %f2170; selp.f32 %f2177, %f2170, %f2176, %p576; add.ftz.f32 %f2178, %f2177, %f2175; sub.s32 %r3672, %r3435, %r3579; xor.b32 %r3673, %r3581, %r3672; add.s32 %r3674, %r3579, %r3626; sub.s32 %r3675, %r3626, %r3673; shf.l.wrap.b32 %r3676, %r3673, %r3673, 8; xor.b32 %r3677, %r3676, %r3675; add.s32 %r3678, %r3673, %r3674; sub.s32 %r3679, %r3674, %r3677; shf.l.wrap.b32 %r3680, %r3677, %r3677, 16; xor.b32 %r3681, %r3680, %r3679; add.s32 %r3682, %r3677, %r3678; sub.s32 %r3683, %r3678, %r3681; shf.l.wrap.b32 %r3684, %r3681, %r3681, 19; xor.b32 %r3685, %r3684, %r3683; add.s32 %r3686, %r3681, %r3682; sub.s32 %r3687, %r3682, %r3685; shf.l.wrap.b32 %r3688, %r3685, %r3685, 4; add.s32 %r3689, %r3685, %r3686; add.s32 %r3690, %r3686, %r3333; xor.b32 %r3691, %r3689, %r3687; xor.b32 %r3692, %r3691, %r3688; shf.l.wrap.b32 %r3693, %r3689, %r3689, 14; sub.s32 %r3694, %r3692, %r3693; xor.b32 %r3695, %r3694, %r3690; shf.l.wrap.b32 %r3696, %r3694, %r3694, 11; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r3689; shf.l.wrap.b32 %r3699, %r3697, %r3697, 25; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 16; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 4; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 14; sub.s32 %r3709, %r3707, %r3708; xor.b32 %r3710, %r3709, %r3703; shf.l.wrap.b32 %r3711, %r3709, %r3709, 24; sub.s32 %r3712, %r3710, %r3711; and.b32 %r3713, %r3712, 31; setp.lt.u32 %p577, %r3713, 24; selp.f32 %f2179, %f2109, %f2122, %p577; setp.lt.u32 %p578, %r3713, 16; selp.f32 %f2180, %f2122, %f2145, %p578; setp.lt.u32 %p579, %r3713, 8; selp.f32 %f2181, %f2145, %f2077, %p579; and.b32 %r3714, %r3712, 1; setp.eq.b32 %p580, %r3714, 1; neg.ftz.f32 %f2182, %f2179; selp.f32 %f2183, %f2182, %f2179, %p580; and.b32 %r3715, %r3712, 2; setp.eq.s32 %p581, %r3715, 0; neg.ftz.f32 %f2184, %f2180; selp.f32 %f2185, %f2180, %f2184, %p581; add.ftz.f32 %f2186, %f2183, %f2185; and.b32 %r3716, %r3712, 4; setp.eq.s32 %p582, %r3716, 0; neg.ftz.f32 %f2187, %f2181; selp.f32 %f2188, %f2181, %f2187, %p582; add.ftz.f32 %f2189, %f2188, %f2186; add.s32 %r3717, %r3333, 1; add.s32 %r3718, %r3355, %r3717; xor.b32 %r3719, %r3363, %r3718; sub.s32 %r3720, %r3719, %r3365; xor.b32 %r3721, %r3720, %r3358; shf.l.wrap.b32 %r3722, %r3720, %r3720, 25; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3363; shf.l.wrap.b32 %r3725, %r3723, %r3723, 16; sub.s32 %r3726, %r3724, %r3725; xor.b32 %r3727, %r3726, %r3720; shf.l.wrap.b32 %r3728, %r3726, %r3726, 4; sub.s32 %r3729, %r3727, %r3728; xor.b32 %r3730, %r3729, %r3723; shf.l.wrap.b32 %r3731, %r3729, %r3729, 14; sub.s32 %r3732, %r3730, %r3731; xor.b32 %r3733, %r3732, %r3726; shf.l.wrap.b32 %r3734, %r3732, %r3732, 24; sub.s32 %r3735, %r3733, %r3734; add.ftz.f32 %f2190, %f2077, 0fBF800000; and.b32 %r3736, %r3735, 31; setp.lt.u32 %p583, %r3736, 24; selp.f32 %f2191, %f2071, %f2073, %p583; setp.lt.u32 %p584, %r3736, 16; selp.f32 %f2192, %f2073, %f2075, %p584; setp.lt.u32 %p585, %r3736, 8; selp.f32 %f2193, %f2075, %f2190, %p585; and.b32 %r3737, %r3735, 1; setp.eq.b32 %p586, %r3737, 1; neg.ftz.f32 %f2194, %f2191; selp.f32 %f2195, %f2194, %f2191, %p586; and.b32 %r3738, %r3735, 2; setp.eq.s32 %p587, %r3738, 0; neg.ftz.f32 %f2196, %f2192; selp.f32 %f2197, %f2192, %f2196, %p587; add.ftz.f32 %f2198, %f2195, %f2197; and.b32 %r3739, %r3735, 4; setp.eq.s32 %p588, %r3739, 0; neg.ftz.f32 %f2199, %f2193; selp.f32 %f2200, %f2193, %f2199, %p588; add.ftz.f32 %f2201, %f2200, %f2198; add.s32 %r3740, %r3404, %r3717; xor.b32 %r3741, %r3412, %r3740; sub.s32 %r3742, %r3741, %r3414; xor.b32 %r3743, %r3742, %r3407; shf.l.wrap.b32 %r3744, %r3742, %r3742, 25; sub.s32 %r3745, %r3743, %r3744; xor.b32 %r3746, %r3745, %r3412; shf.l.wrap.b32 %r3747, %r3745, %r3745, 16; sub.s32 %r3748, %r3746, %r3747; xor.b32 %r3749, %r3748, %r3742; shf.l.wrap.b32 %r3750, %r3748, %r3748, 4; sub.s32 %r3751, %r3749, %r3750; xor.b32 %r3752, %r3751, %r3745; shf.l.wrap.b32 %r3753, %r3751, %r3751, 14; sub.s32 %r3754, %r3752, %r3753; xor.b32 %r3755, %r3754, %r3748; shf.l.wrap.b32 %r3756, %r3754, %r3754, 24; sub.s32 %r3757, %r3755, %r3756; and.b32 %r3758, %r3757, 31; setp.lt.u32 %p589, %r3758, 24; selp.f32 %f2202, %f2109, %f2073, %p589; setp.lt.u32 %p590, %r3758, 16; selp.f32 %f2203, %f2073, %f2075, %p590; setp.lt.u32 %p591, %r3758, 8; selp.f32 %f2204, %f2075, %f2190, %p591; and.b32 %r3759, %r3757, 1; setp.eq.b32 %p592, %r3759, 1; neg.ftz.f32 %f2205, %f2202; selp.f32 %f2206, %f2205, %f2202, %p592; and.b32 %r3760, %r3757, 2; setp.eq.s32 %p593, %r3760, 0; neg.ftz.f32 %f2207, %f2203; selp.f32 %f2208, %f2203, %f2207, %p593; add.ftz.f32 %f2209, %f2206, %f2208; and.b32 %r3761, %r3757, 4; setp.eq.s32 %p594, %r3761, 0; neg.ftz.f32 %f2210, %f2204; selp.f32 %f2211, %f2204, %f2210, %p594; add.ftz.f32 %f2212, %f2211, %f2209; add.s32 %r3762, %r3451, %r3717; xor.b32 %r3763, %r3459, %r3762; sub.s32 %r3764, %r3763, %r3461; xor.b32 %r3765, %r3764, %r3454; shf.l.wrap.b32 %r3766, %r3764, %r3764, 25; sub.s32 %r3767, %r3765, %r3766; xor.b32 %r3768, %r3767, %r3459; shf.l.wrap.b32 %r3769, %r3767, %r3767, 16; sub.s32 %r3770, %r3768, %r3769; xor.b32 %r3771, %r3770, %r3764; shf.l.wrap.b32 %r3772, %r3770, %r3770, 4; sub.s32 %r3773, %r3771, %r3772; xor.b32 %r3774, %r3773, %r3767; shf.l.wrap.b32 %r3775, %r3773, %r3773, 14; sub.s32 %r3776, %r3774, %r3775; xor.b32 %r3777, %r3776, %r3770; shf.l.wrap.b32 %r3778, %r3776, %r3776, 24; sub.s32 %r3779, %r3777, %r3778; and.b32 %r3780, %r3779, 31; setp.lt.u32 %p595, %r3780, 24; selp.f32 %f2213, %f2071, %f2122, %p595; setp.lt.u32 %p596, %r3780, 16; selp.f32 %f2214, %f2122, %f2075, %p596; setp.lt.u32 %p597, %r3780, 8; selp.f32 %f2215, %f2075, %f2190, %p597; and.b32 %r3781, %r3779, 1; setp.eq.b32 %p598, %r3781, 1; neg.ftz.f32 %f2216, %f2213; selp.f32 %f2217, %f2216, %f2213, %p598; and.b32 %r3782, %r3779, 2; setp.eq.s32 %p599, %r3782, 0; neg.ftz.f32 %f2218, %f2214; selp.f32 %f2219, %f2214, %f2218, %p599; add.ftz.f32 %f2220, %f2217, %f2219; and.b32 %r3783, %r3779, 4; setp.eq.s32 %p600, %r3783, 0; neg.ftz.f32 %f2221, %f2215; selp.f32 %f2222, %f2215, %f2221, %p600; add.ftz.f32 %f2223, %f2222, %f2220; add.s32 %r3784, %r3496, %r3717; xor.b32 %r3785, %r3504, %r3784; sub.s32 %r3786, %r3785, %r3506; xor.b32 %r3787, %r3786, %r3499; shf.l.wrap.b32 %r3788, %r3786, %r3786, 25; sub.s32 %r3789, %r3787, %r3788; xor.b32 %r3790, %r3789, %r3504; shf.l.wrap.b32 %r3791, %r3789, %r3789, 16; sub.s32 %r3792, %r3790, %r3791; xor.b32 %r3793, %r3792, %r3786; shf.l.wrap.b32 %r3794, %r3792, %r3792, 4; sub.s32 %r3795, %r3793, %r3794; xor.b32 %r3796, %r3795, %r3789; shf.l.wrap.b32 %r3797, %r3795, %r3795, 14; sub.s32 %r3798, %r3796, %r3797; xor.b32 %r3799, %r3798, %r3792; shf.l.wrap.b32 %r3800, %r3798, %r3798, 24; sub.s32 %r3801, %r3799, %r3800; and.b32 %r3802, %r3801, 31; setp.lt.u32 %p601, %r3802, 24; selp.f32 %f2224, %f2109, %f2122, %p601; setp.lt.u32 %p602, %r3802, 16; selp.f32 %f2225, %f2122, %f2075, %p602; setp.lt.u32 %p603, %r3802, 8; selp.f32 %f2226, %f2075, %f2190, %p603; and.b32 %r3803, %r3801, 1; setp.eq.b32 %p604, %r3803, 1; neg.ftz.f32 %f2227, %f2224; selp.f32 %f2228, %f2227, %f2224, %p604; and.b32 %r3804, %r3801, 2; setp.eq.s32 %p605, %r3804, 0; neg.ftz.f32 %f2229, %f2225; selp.f32 %f2230, %f2225, %f2229, %p605; add.ftz.f32 %f2231, %f2228, %f2230; and.b32 %r3805, %r3801, 4; setp.eq.s32 %p606, %r3805, 0; neg.ftz.f32 %f2232, %f2226; selp.f32 %f2233, %f2226, %f2232, %p606; add.ftz.f32 %f2234, %f2233, %f2231; add.s32 %r3806, %r3548, %r3717; xor.b32 %r3807, %r3556, %r3806; sub.s32 %r3808, %r3807, %r3558; xor.b32 %r3809, %r3808, %r3551; shf.l.wrap.b32 %r3810, %r3808, %r3808, 25; sub.s32 %r3811, %r3809, %r3810; xor.b32 %r3812, %r3811, %r3556; shf.l.wrap.b32 %r3813, %r3811, %r3811, 16; sub.s32 %r3814, %r3812, %r3813; xor.b32 %r3815, %r3814, %r3808; shf.l.wrap.b32 %r3816, %r3814, %r3814, 4; sub.s32 %r3817, %r3815, %r3816; xor.b32 %r3818, %r3817, %r3811; shf.l.wrap.b32 %r3819, %r3817, %r3817, 14; sub.s32 %r3820, %r3818, %r3819; xor.b32 %r3821, %r3820, %r3814; shf.l.wrap.b32 %r3822, %r3820, %r3820, 24; sub.s32 %r3823, %r3821, %r3822; and.b32 %r3824, %r3823, 31; setp.lt.u32 %p607, %r3824, 24; selp.f32 %f2235, %f2071, %f2073, %p607; setp.lt.u32 %p608, %r3824, 16; selp.f32 %f2236, %f2073, %f2145, %p608; setp.lt.u32 %p609, %r3824, 8; selp.f32 %f2237, %f2145, %f2190, %p609; and.b32 %r3825, %r3823, 1; setp.eq.b32 %p610, %r3825, 1; neg.ftz.f32 %f2238, %f2235; selp.f32 %f2239, %f2238, %f2235, %p610; and.b32 %r3826, %r3823, 2; setp.eq.s32 %p611, %r3826, 0; neg.ftz.f32 %f2240, %f2236; selp.f32 %f2241, %f2236, %f2240, %p611; add.ftz.f32 %f2242, %f2239, %f2241; and.b32 %r3827, %r3823, 4; setp.eq.s32 %p612, %r3827, 0; neg.ftz.f32 %f2243, %f2237; selp.f32 %f2244, %f2237, %f2243, %p612; add.ftz.f32 %f2245, %f2244, %f2242; add.s32 %r3828, %r3595, %r3717; xor.b32 %r3829, %r3603, %r3828; sub.s32 %r3830, %r3829, %r3605; xor.b32 %r3831, %r3830, %r3598; shf.l.wrap.b32 %r3832, %r3830, %r3830, 25; sub.s32 %r3833, %r3831, %r3832; xor.b32 %r3834, %r3833, %r3603; shf.l.wrap.b32 %r3835, %r3833, %r3833, 16; sub.s32 %r3836, %r3834, %r3835; xor.b32 %r3837, %r3836, %r3830; shf.l.wrap.b32 %r3838, %r3836, %r3836, 4; sub.s32 %r3839, %r3837, %r3838; xor.b32 %r3840, %r3839, %r3833; shf.l.wrap.b32 %r3841, %r3839, %r3839, 14; sub.s32 %r3842, %r3840, %r3841; xor.b32 %r3843, %r3842, %r3836; shf.l.wrap.b32 %r3844, %r3842, %r3842, 24; sub.s32 %r3845, %r3843, %r3844; and.b32 %r3846, %r3845, 31; setp.lt.u32 %p613, %r3846, 24; selp.f32 %f2246, %f2109, %f2073, %p613; setp.lt.u32 %p614, %r3846, 16; selp.f32 %f2247, %f2073, %f2145, %p614; setp.lt.u32 %p615, %r3846, 8; selp.f32 %f2248, %f2145, %f2190, %p615; and.b32 %r3847, %r3845, 1; setp.eq.b32 %p616, %r3847, 1; neg.ftz.f32 %f2249, %f2246; selp.f32 %f2250, %f2249, %f2246, %p616; and.b32 %r3848, %r3845, 2; setp.eq.s32 %p617, %r3848, 0; neg.ftz.f32 %f2251, %f2247; selp.f32 %f2252, %f2247, %f2251, %p617; add.ftz.f32 %f2253, %f2250, %f2252; and.b32 %r3849, %r3845, 4; setp.eq.s32 %p618, %r3849, 0; neg.ftz.f32 %f2254, %f2248; selp.f32 %f2255, %f2248, %f2254, %p618; add.ftz.f32 %f2256, %f2255, %f2253; add.s32 %r3850, %r3641, %r3717; xor.b32 %r3851, %r3649, %r3850; sub.s32 %r3852, %r3851, %r3651; xor.b32 %r3853, %r3852, %r3644; shf.l.wrap.b32 %r3854, %r3852, %r3852, 25; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3649; shf.l.wrap.b32 %r3857, %r3855, %r3855, 16; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 4; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 14; sub.s32 %r3864, %r3862, %r3863; xor.b32 %r3865, %r3864, %r3858; shf.l.wrap.b32 %r3866, %r3864, %r3864, 24; sub.s32 %r3867, %r3865, %r3866; and.b32 %r3868, %r3867, 31; setp.lt.u32 %p619, %r3868, 24; selp.f32 %f2257, %f2071, %f2122, %p619; setp.lt.u32 %p620, %r3868, 16; selp.f32 %f2258, %f2122, %f2145, %p620; setp.lt.u32 %p621, %r3868, 8; selp.f32 %f2259, %f2145, %f2190, %p621; and.b32 %r3869, %r3867, 1; setp.eq.b32 %p622, %r3869, 1; neg.ftz.f32 %f2260, %f2257; selp.f32 %f2261, %f2260, %f2257, %p622; and.b32 %r3870, %r3867, 2; setp.eq.s32 %p623, %r3870, 0; neg.ftz.f32 %f2262, %f2258; selp.f32 %f2263, %f2258, %f2262, %p623; add.ftz.f32 %f2264, %f2261, %f2263; and.b32 %r3871, %r3867, 4; setp.eq.s32 %p624, %r3871, 0; neg.ftz.f32 %f2265, %f2259; selp.f32 %f2266, %f2259, %f2265, %p624; add.ftz.f32 %f2267, %f2266, %f2264; add.s32 %r3872, %r3686, %r3717; xor.b32 %r3873, %r3694, %r3872; sub.s32 %r3874, %r3873, %r3696; xor.b32 %r3875, %r3874, %r3689; shf.l.wrap.b32 %r3876, %r3874, %r3874, 25; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3694; shf.l.wrap.b32 %r3879, %r3877, %r3877, 16; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 4; sub.s32 %r3883, %r3881, %r3882; xor.b32 %r3884, %r3883, %r3877; shf.l.wrap.b32 %r3885, %r3883, %r3883, 14; sub.s32 %r3886, %r3884, %r3885; xor.b32 %r3887, %r3886, %r3880; shf.l.wrap.b32 %r3888, %r3886, %r3886, 24; sub.s32 %r3889, %r3887, %r3888; and.b32 %r3890, %r3889, 31; setp.lt.u32 %p625, %r3890, 24; selp.f32 %f2268, %f2109, %f2122, %p625; setp.lt.u32 %p626, %r3890, 16; selp.f32 %f2269, %f2122, %f2145, %p626; setp.lt.u32 %p627, %r3890, 8; selp.f32 %f2270, %f2145, %f2190, %p627; and.b32 %r3891, %r3889, 1; setp.eq.b32 %p628, %r3891, 1; neg.ftz.f32 %f2271, %f2268; selp.f32 %f2272, %f2271, %f2268, %p628; and.b32 %r3892, %r3889, 2; setp.eq.s32 %p629, %r3892, 0; neg.ftz.f32 %f2273, %f2269; selp.f32 %f2274, %f2269, %f2273, %p629; add.ftz.f32 %f2275, %f2272, %f2274; and.b32 %r3893, %r3889, 4; setp.eq.s32 %p630, %r3893, 0; neg.ftz.f32 %f2276, %f2270; selp.f32 %f2277, %f2270, %f2276, %p630; add.ftz.f32 %f2278, %f2277, %f2275; sub.ftz.f32 %f2279, %f2110, %f2082; sub.ftz.f32 %f2280, %f2110, %f2087; sub.ftz.f32 %f2281, %f2110, %f2092; mul.ftz.f32 %f2282, %f2279, %f2108; fma.rn.ftz.f32 %f2283, %f2082, %f2121, %f2282; mul.ftz.f32 %f2284, %f2279, %f2133; fma.rn.ftz.f32 %f2285, %f2082, %f2144, %f2284; mul.ftz.f32 %f2286, %f2087, %f2285; fma.rn.ftz.f32 %f2287, %f2280, %f2283, %f2286; mul.ftz.f32 %f2288, %f2279, %f2156; fma.rn.ftz.f32 %f2289, %f2082, %f2167, %f2288; mul.ftz.f32 %f2290, %f2279, %f2178; fma.rn.ftz.f32 %f2291, %f2082, %f2189, %f2290; mul.ftz.f32 %f2292, %f2087, %f2291; fma.rn.ftz.f32 %f2293, %f2280, %f2289, %f2292; mul.ftz.f32 %f2294, %f2092, %f2293; fma.rn.ftz.f32 %f2295, %f2281, %f2287, %f2294; mul.ftz.f32 %f2296, %f2279, %f2201; fma.rn.ftz.f32 %f2297, %f2082, %f2212, %f2296; mul.ftz.f32 %f2298, %f2279, %f2223; fma.rn.ftz.f32 %f2299, %f2082, %f2234, %f2298; mul.ftz.f32 %f2300, %f2087, %f2299; fma.rn.ftz.f32 %f2301, %f2280, %f2297, %f2300; mul.ftz.f32 %f2302, %f2279, %f2245; fma.rn.ftz.f32 %f2303, %f2082, %f2256, %f2302; mul.ftz.f32 %f2304, %f2279, %f2267; fma.rn.ftz.f32 %f2305, %f2082, %f2278, %f2304; mul.ftz.f32 %f2306, %f2087, %f2305; fma.rn.ftz.f32 %f2307, %f2280, %f2303, %f2306; mul.ftz.f32 %f2308, %f2092, %f2307; fma.rn.ftz.f32 %f2309, %f2281, %f2301, %f2308; sub.ftz.f32 %f2310, %f2309, %f2295; fma.rn.ftz.f32 %f97, %f2097, %f2310, %f2295; mov.b32 %r29, %f97; setp.nan.ftz.f32 %p631, %f97, %f97; mov.f32 %f5552, 0f00000000; @%p631 bra $L__BB74_85; or.b32 %r3894, %r29, -2147483648; setp.ne.s32 %p632, %r3894, -2147483648; add.ftz.f32 %f2312, %f97, %f97; setp.eq.ftz.f32 %p633, %f2312, %f97; and.pred %p634, %p633, %p632; @%p634 bra $L__BB74_85; shl.b32 %r3895, %r29, 1; setp.lt.u32 %p635, %r3895, -16777215; selp.f32 %f5552, %f97, 0f00000000, %p635; $L__BB74_85: setp.gt.ftz.f32 %p637, %f5559, 0f3F800000; selp.f32 %f2313, 0f3F800000, %f5559, %p637; fma.rn.ftz.f32 %f2314, %f5552, 0f3F559B3D, %f21; mul.ftz.f32 %f2315, %f5557, %f2314; mul.ftz.f32 %f5557, %f89, %f5557; fma.rn.ftz.f32 %f5827, %f2313, %f2315, %f5827; mul.ftz.f32 %f2316, %f24, %f2315; mul.ftz.f32 %f5559, %f2313, %f2316; mul.ftz.f32 %f5568, %f27, %f5568; mul.ftz.f32 %f5569, %f27, %f5569; mul.ftz.f32 %f5570, %f27, %f5570; mul.ftz.f32 %f5571, %f27, %f5571; add.s32 %r10020, %r10020, 1; setp.gt.ftz.f32 %p638, %f5559, 0f3A83126F; mov.pred %p1991, %p636; @%p638 bra $L__BB74_81; $L__BB74_86: cvt.rmi.ftz.f32.f32 %f2317, %f26; sub.ftz.f32 %f114, %f26, %f2317; setp.eq.ftz.f32 %p639, %f114, 0f00000000; not.pred %p640, %p1991; or.pred %p641, %p640, %p639; @%p641 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f2319, %f5568; cvt.rzi.ftz.s32.f32 %r3896, %f2319; sub.ftz.f32 %f2320, %f5568, %f2319; cvt.rmi.ftz.f32.f32 %f2321, %f5569; cvt.rzi.ftz.s32.f32 %r3897, %f2321; sub.ftz.f32 %f2322, %f5569, %f2321; cvt.rmi.ftz.f32.f32 %f2323, %f5570; cvt.rzi.ftz.s32.f32 %r3898, %f2323; sub.ftz.f32 %f2324, %f5570, %f2323; cvt.rmi.ftz.f32.f32 %f2325, %f5571; cvt.rzi.ftz.s32.f32 %r3899, %f2325; sub.ftz.f32 %f2326, %f5571, %f2325; mul.ftz.f32 %f2327, %f2320, %f2320; mul.ftz.f32 %f2328, %f2320, %f2327; fma.rn.ftz.f32 %f2329, %f2320, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2330, %f2320, %f2329, 0f41200000; mul.ftz.f32 %f2331, %f2328, %f2330; mul.ftz.f32 %f2332, %f2322, %f2322; mul.ftz.f32 %f2333, %f2322, %f2332; fma.rn.ftz.f32 %f2334, %f2322, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2335, %f2322, %f2334, 0f41200000; mul.ftz.f32 %f2336, %f2333, %f2335; mul.ftz.f32 %f2337, %f2324, %f2324; mul.ftz.f32 %f2338, %f2324, %f2337; fma.rn.ftz.f32 %f2339, %f2324, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2340, %f2324, %f2339, 0f41200000; mul.ftz.f32 %f2341, %f2338, %f2340; mul.ftz.f32 %f2342, %f2326, %f2326; mul.ftz.f32 %f2343, %f2326, %f2342; fma.rn.ftz.f32 %f2344, %f2326, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f2345, %f2326, %f2344, 0f41200000; mul.ftz.f32 %f2346, %f2343, %f2345; add.s32 %r3900, %r3897, -559038708; add.s32 %r3901, %r3898, -559038708; sub.s32 %r3902, %r3896, %r3898; shf.l.wrap.b32 %r3903, %r3901, %r3901, 4; xor.b32 %r3904, %r3903, %r3902; add.s32 %r3905, %r3901, %r3900; sub.s32 %r3906, %r3900, %r3904; shf.l.wrap.b32 %r3907, %r3904, %r3904, 6; xor.b32 %r3908, %r3907, %r3906; add.s32 %r3909, %r3904, %r3905; sub.s32 %r3910, %r3905, %r3908; shf.l.wrap.b32 %r3911, %r3908, %r3908, 8; xor.b32 %r3912, %r3911, %r3910; add.s32 %r3913, %r3908, %r3909; sub.s32 %r3914, %r3909, %r3912; shf.l.wrap.b32 %r3915, %r3912, %r3912, 16; xor.b32 %r3916, %r3915, %r3914; add.s32 %r3917, %r3912, %r3913; sub.s32 %r3918, %r3913, %r3916; shf.l.wrap.b32 %r3919, %r3916, %r3916, 19; xor.b32 %r3920, %r3919, %r3918; add.s32 %r3921, %r3916, %r3917; sub.s32 %r3922, %r3917, %r3920; shf.l.wrap.b32 %r3923, %r3920, %r3920, 4; add.s32 %r3924, %r3920, %r3921; add.s32 %r3925, %r3921, %r3899; xor.b32 %r3926, %r3924, %r3922; xor.b32 %r3927, %r3926, %r3923; shf.l.wrap.b32 %r3928, %r3924, %r3924, 14; sub.s32 %r3929, %r3927, %r3928; xor.b32 %r3930, %r3929, %r3925; shf.l.wrap.b32 %r3931, %r3929, %r3929, 11; sub.s32 %r3932, %r3930, %r3931; xor.b32 %r3933, %r3932, %r3924; shf.l.wrap.b32 %r3934, %r3932, %r3932, 25; sub.s32 %r3935, %r3933, %r3934; xor.b32 %r3936, %r3935, %r3929; shf.l.wrap.b32 %r3937, %r3935, %r3935, 16; sub.s32 %r3938, %r3936, %r3937; xor.b32 %r3939, %r3938, %r3932; shf.l.wrap.b32 %r3940, %r3938, %r3938, 4; sub.s32 %r3941, %r3939, %r3940; xor.b32 %r3942, %r3941, %r3935; shf.l.wrap.b32 %r3943, %r3941, %r3941, 14; sub.s32 %r3944, %r3942, %r3943; xor.b32 %r3945, %r3944, %r3938; shf.l.wrap.b32 %r3946, %r3944, %r3944, 24; sub.s32 %r3947, %r3945, %r3946; and.b32 %r3948, %r3947, 31; setp.lt.u32 %p642, %r3948, 24; selp.f32 %f2347, %f2320, %f2322, %p642; setp.lt.u32 %p643, %r3948, 16; selp.f32 %f2348, %f2322, %f2324, %p643; setp.lt.u32 %p644, %r3948, 8; selp.f32 %f2349, %f2324, %f2326, %p644; and.b32 %r3949, %r3947, 1; setp.eq.b32 %p645, %r3949, 1; neg.ftz.f32 %f2350, %f2347; selp.f32 %f2351, %f2350, %f2347, %p645; and.b32 %r3950, %r3947, 2; setp.eq.s32 %p646, %r3950, 0; neg.ftz.f32 %f2352, %f2348; selp.f32 %f2353, %f2348, %f2352, %p646; add.ftz.f32 %f2354, %f2351, %f2353; and.b32 %r3951, %r3947, 4; setp.eq.s32 %p647, %r3951, 0; neg.ftz.f32 %f2355, %f2349; selp.f32 %f2356, %f2349, %f2355, %p647; add.ftz.f32 %f2357, %f2356, %f2354; add.s32 %r3952, %r3896, 1; sub.s32 %r3953, %r3952, %r3898; xor.b32 %r3954, %r3903, %r3953; sub.s32 %r3955, %r3900, %r3954; shf.l.wrap.b32 %r3956, %r3954, %r3954, 6; xor.b32 %r3957, %r3956, %r3955; add.s32 %r3958, %r3954, %r3905; sub.s32 %r3959, %r3905, %r3957; shf.l.wrap.b32 %r3960, %r3957, %r3957, 8; xor.b32 %r3961, %r3960, %r3959; add.s32 %r3962, %r3957, %r3958; sub.s32 %r3963, %r3958, %r3961; shf.l.wrap.b32 %r3964, %r3961, %r3961, 16; xor.b32 %r3965, %r3964, %r3963; add.s32 %r3966, %r3961, %r3962; sub.s32 %r3967, %r3962, %r3965; shf.l.wrap.b32 %r3968, %r3965, %r3965, 19; xor.b32 %r3969, %r3968, %r3967; add.s32 %r3970, %r3965, %r3966; sub.s32 %r3971, %r3966, %r3969; shf.l.wrap.b32 %r3972, %r3969, %r3969, 4; add.s32 %r3973, %r3969, %r3970; add.s32 %r3974, %r3970, %r3899; xor.b32 %r3975, %r3973, %r3971; xor.b32 %r3976, %r3975, %r3972; shf.l.wrap.b32 %r3977, %r3973, %r3973, 14; sub.s32 %r3978, %r3976, %r3977; xor.b32 %r3979, %r3978, %r3974; shf.l.wrap.b32 %r3980, %r3978, %r3978, 11; sub.s32 %r3981, %r3979, %r3980; xor.b32 %r3982, %r3981, %r3973; shf.l.wrap.b32 %r3983, %r3981, %r3981, 25; sub.s32 %r3984, %r3982, %r3983; xor.b32 %r3985, %r3984, %r3978; shf.l.wrap.b32 %r3986, %r3984, %r3984, 16; sub.s32 %r3987, %r3985, %r3986; xor.b32 %r3988, %r3987, %r3981; shf.l.wrap.b32 %r3989, %r3987, %r3987, 4; sub.s32 %r3990, %r3988, %r3989; xor.b32 %r3991, %r3990, %r3984; shf.l.wrap.b32 %r3992, %r3990, %r3990, 14; sub.s32 %r3993, %r3991, %r3992; xor.b32 %r3994, %r3993, %r3987; shf.l.wrap.b32 %r3995, %r3993, %r3993, 24; sub.s32 %r3996, %r3994, %r3995; add.ftz.f32 %f2358, %f2320, 0fBF800000; mov.f32 %f2359, 0f3F800000; and.b32 %r3997, %r3996, 31; setp.lt.u32 %p648, %r3997, 24; selp.f32 %f2360, %f2358, %f2322, %p648; setp.lt.u32 %p649, %r3997, 16; selp.f32 %f2361, %f2322, %f2324, %p649; setp.lt.u32 %p650, %r3997, 8; selp.f32 %f2362, %f2324, %f2326, %p650; and.b32 %r3998, %r3996, 1; setp.eq.b32 %p651, %r3998, 1; neg.ftz.f32 %f2363, %f2360; selp.f32 %f2364, %f2363, %f2360, %p651; and.b32 %r3999, %r3996, 2; setp.eq.s32 %p652, %r3999, 0; neg.ftz.f32 %f2365, %f2361; selp.f32 %f2366, %f2361, %f2365, %p652; add.ftz.f32 %f2367, %f2364, %f2366; and.b32 %r4000, %r3996, 4; setp.eq.s32 %p653, %r4000, 0; neg.ftz.f32 %f2368, %f2362; selp.f32 %f2369, %f2362, %f2368, %p653; add.ftz.f32 %f2370, %f2369, %f2367; add.s32 %r4001, %r3897, -559038707; add.s32 %r4002, %r3901, %r4001; sub.s32 %r4003, %r4001, %r3904; xor.b32 %r4004, %r3907, %r4003; add.s32 %r4005, %r3904, %r4002; sub.s32 %r4006, %r4002, %r4004; shf.l.wrap.b32 %r4007, %r4004, %r4004, 8; xor.b32 %r4008, %r4007, %r4006; add.s32 %r4009, %r4004, %r4005; sub.s32 %r4010, %r4005, %r4008; shf.l.wrap.b32 %r4011, %r4008, %r4008, 16; xor.b32 %r4012, %r4011, %r4010; add.s32 %r4013, %r4008, %r4009; sub.s32 %r4014, %r4009, %r4012; shf.l.wrap.b32 %r4015, %r4012, %r4012, 19; xor.b32 %r4016, %r4015, %r4014; add.s32 %r4017, %r4012, %r4013; sub.s32 %r4018, %r4013, %r4016; shf.l.wrap.b32 %r4019, %r4016, %r4016, 4; add.s32 %r4020, %r4016, %r4017; add.s32 %r4021, %r4017, %r3899; xor.b32 %r4022, %r4020, %r4018; xor.b32 %r4023, %r4022, %r4019; shf.l.wrap.b32 %r4024, %r4020, %r4020, 14; sub.s32 %r4025, %r4023, %r4024; xor.b32 %r4026, %r4025, %r4021; shf.l.wrap.b32 %r4027, %r4025, %r4025, 11; sub.s32 %r4028, %r4026, %r4027; xor.b32 %r4029, %r4028, %r4020; shf.l.wrap.b32 %r4030, %r4028, %r4028, 25; sub.s32 %r4031, %r4029, %r4030; xor.b32 %r4032, %r4031, %r4025; shf.l.wrap.b32 %r4033, %r4031, %r4031, 16; sub.s32 %r4034, %r4032, %r4033; xor.b32 %r4035, %r4034, %r4028; shf.l.wrap.b32 %r4036, %r4034, %r4034, 4; sub.s32 %r4037, %r4035, %r4036; xor.b32 %r4038, %r4037, %r4031; shf.l.wrap.b32 %r4039, %r4037, %r4037, 14; sub.s32 %r4040, %r4038, %r4039; xor.b32 %r4041, %r4040, %r4034; shf.l.wrap.b32 %r4042, %r4040, %r4040, 24; sub.s32 %r4043, %r4041, %r4042; add.ftz.f32 %f2371, %f2322, 0fBF800000; and.b32 %r4044, %r4043, 31; setp.lt.u32 %p654, %r4044, 24; selp.f32 %f2372, %f2320, %f2371, %p654; setp.lt.u32 %p655, %r4044, 16; selp.f32 %f2373, %f2371, %f2324, %p655; setp.lt.u32 %p656, %r4044, 8; selp.f32 %f2374, %f2324, %f2326, %p656; and.b32 %r4045, %r4043, 1; setp.eq.b32 %p657, %r4045, 1; neg.ftz.f32 %f2375, %f2372; selp.f32 %f2376, %f2375, %f2372, %p657; and.b32 %r4046, %r4043, 2; setp.eq.s32 %p658, %r4046, 0; neg.ftz.f32 %f2377, %f2373; selp.f32 %f2378, %f2373, %f2377, %p658; add.ftz.f32 %f2379, %f2376, %f2378; and.b32 %r4047, %r4043, 4; setp.eq.s32 %p659, %r4047, 0; neg.ftz.f32 %f2380, %f2374; selp.f32 %f2381, %f2374, %f2380, %p659; add.ftz.f32 %f2382, %f2381, %f2379; sub.s32 %r4048, %r4001, %r3954; xor.b32 %r4049, %r3956, %r4048; add.s32 %r4050, %r3954, %r4002; sub.s32 %r4051, %r4002, %r4049; shf.l.wrap.b32 %r4052, %r4049, %r4049, 8; xor.b32 %r4053, %r4052, %r4051; add.s32 %r4054, %r4049, %r4050; sub.s32 %r4055, %r4050, %r4053; shf.l.wrap.b32 %r4056, %r4053, %r4053, 16; xor.b32 %r4057, %r4056, %r4055; add.s32 %r4058, %r4053, %r4054; sub.s32 %r4059, %r4054, %r4057; shf.l.wrap.b32 %r4060, %r4057, %r4057, 19; xor.b32 %r4061, %r4060, %r4059; add.s32 %r4062, %r4057, %r4058; sub.s32 %r4063, %r4058, %r4061; shf.l.wrap.b32 %r4064, %r4061, %r4061, 4; add.s32 %r4065, %r4061, %r4062; add.s32 %r4066, %r4062, %r3899; xor.b32 %r4067, %r4065, %r4063; xor.b32 %r4068, %r4067, %r4064; shf.l.wrap.b32 %r4069, %r4065, %r4065, 14; sub.s32 %r4070, %r4068, %r4069; xor.b32 %r4071, %r4070, %r4066; shf.l.wrap.b32 %r4072, %r4070, %r4070, 11; sub.s32 %r4073, %r4071, %r4072; xor.b32 %r4074, %r4073, %r4065; shf.l.wrap.b32 %r4075, %r4073, %r4073, 25; sub.s32 %r4076, %r4074, %r4075; xor.b32 %r4077, %r4076, %r4070; shf.l.wrap.b32 %r4078, %r4076, %r4076, 16; sub.s32 %r4079, %r4077, %r4078; xor.b32 %r4080, %r4079, %r4073; shf.l.wrap.b32 %r4081, %r4079, %r4079, 4; sub.s32 %r4082, %r4080, %r4081; xor.b32 %r4083, %r4082, %r4076; shf.l.wrap.b32 %r4084, %r4082, %r4082, 14; sub.s32 %r4085, %r4083, %r4084; xor.b32 %r4086, %r4085, %r4079; shf.l.wrap.b32 %r4087, %r4085, %r4085, 24; sub.s32 %r4088, %r4086, %r4087; and.b32 %r4089, %r4088, 31; setp.lt.u32 %p660, %r4089, 24; selp.f32 %f2383, %f2358, %f2371, %p660; setp.lt.u32 %p661, %r4089, 16; selp.f32 %f2384, %f2371, %f2324, %p661; setp.lt.u32 %p662, %r4089, 8; selp.f32 %f2385, %f2324, %f2326, %p662; and.b32 %r4090, %r4088, 1; setp.eq.b32 %p663, %r4090, 1; neg.ftz.f32 %f2386, %f2383; selp.f32 %f2387, %f2386, %f2383, %p663; and.b32 %r4091, %r4088, 2; setp.eq.s32 %p664, %r4091, 0; neg.ftz.f32 %f2388, %f2384; selp.f32 %f2389, %f2384, %f2388, %p664; add.ftz.f32 %f2390, %f2387, %f2389; and.b32 %r4092, %r4088, 4; setp.eq.s32 %p665, %r4092, 0; neg.ftz.f32 %f2391, %f2385; selp.f32 %f2392, %f2385, %f2391, %p665; add.ftz.f32 %f2393, %f2392, %f2390; not.b32 %r4093, %r3898; add.s32 %r4094, %r3898, -559038707; add.s32 %r4095, %r3896, %r4093; shf.l.wrap.b32 %r4096, %r4094, %r4094, 4; xor.b32 %r4097, %r4096, %r4095; add.s32 %r4098, %r4094, %r3900; sub.s32 %r4099, %r3900, %r4097; shf.l.wrap.b32 %r4100, %r4097, %r4097, 6; xor.b32 %r4101, %r4100, %r4099; add.s32 %r4102, %r4097, %r4098; sub.s32 %r4103, %r4098, %r4101; shf.l.wrap.b32 %r4104, %r4101, %r4101, 8; xor.b32 %r4105, %r4104, %r4103; add.s32 %r4106, %r4101, %r4102; sub.s32 %r4107, %r4102, %r4105; shf.l.wrap.b32 %r4108, %r4105, %r4105, 16; xor.b32 %r4109, %r4108, %r4107; add.s32 %r4110, %r4105, %r4106; sub.s32 %r4111, %r4106, %r4109; shf.l.wrap.b32 %r4112, %r4109, %r4109, 19; xor.b32 %r4113, %r4112, %r4111; add.s32 %r4114, %r4109, %r4110; sub.s32 %r4115, %r4110, %r4113; shf.l.wrap.b32 %r4116, %r4113, %r4113, 4; add.s32 %r4117, %r4113, %r4114; add.s32 %r4118, %r4114, %r3899; xor.b32 %r4119, %r4117, %r4115; xor.b32 %r4120, %r4119, %r4116; shf.l.wrap.b32 %r4121, %r4117, %r4117, 14; sub.s32 %r4122, %r4120, %r4121; xor.b32 %r4123, %r4122, %r4118; shf.l.wrap.b32 %r4124, %r4122, %r4122, 11; sub.s32 %r4125, %r4123, %r4124; xor.b32 %r4126, %r4125, %r4117; shf.l.wrap.b32 %r4127, %r4125, %r4125, 25; sub.s32 %r4128, %r4126, %r4127; xor.b32 %r4129, %r4128, %r4122; shf.l.wrap.b32 %r4130, %r4128, %r4128, 16; sub.s32 %r4131, %r4129, %r4130; xor.b32 %r4132, %r4131, %r4125; shf.l.wrap.b32 %r4133, %r4131, %r4131, 4; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, %r4128; shf.l.wrap.b32 %r4136, %r4134, %r4134, 14; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 24; sub.s32 %r4140, %r4138, %r4139; add.ftz.f32 %f2394, %f2324, 0fBF800000; and.b32 %r4141, %r4140, 31; setp.lt.u32 %p666, %r4141, 24; selp.f32 %f2395, %f2320, %f2322, %p666; setp.lt.u32 %p667, %r4141, 16; selp.f32 %f2396, %f2322, %f2394, %p667; setp.lt.u32 %p668, %r4141, 8; selp.f32 %f2397, %f2394, %f2326, %p668; and.b32 %r4142, %r4140, 1; setp.eq.b32 %p669, %r4142, 1; neg.ftz.f32 %f2398, %f2395; selp.f32 %f2399, %f2398, %f2395, %p669; and.b32 %r4143, %r4140, 2; setp.eq.s32 %p670, %r4143, 0; neg.ftz.f32 %f2400, %f2396; selp.f32 %f2401, %f2396, %f2400, %p670; add.ftz.f32 %f2402, %f2399, %f2401; and.b32 %r4144, %r4140, 4; setp.eq.s32 %p671, %r4144, 0; neg.ftz.f32 %f2403, %f2397; selp.f32 %f2404, %f2397, %f2403, %p671; add.ftz.f32 %f2405, %f2404, %f2402; xor.b32 %r4145, %r4096, %r3902; sub.s32 %r4146, %r3900, %r4145; shf.l.wrap.b32 %r4147, %r4145, %r4145, 6; xor.b32 %r4148, %r4147, %r4146; add.s32 %r4149, %r4145, %r4098; sub.s32 %r4150, %r4098, %r4148; shf.l.wrap.b32 %r4151, %r4148, %r4148, 8; xor.b32 %r4152, %r4151, %r4150; add.s32 %r4153, %r4148, %r4149; sub.s32 %r4154, %r4149, %r4152; shf.l.wrap.b32 %r4155, %r4152, %r4152, 16; xor.b32 %r4156, %r4155, %r4154; add.s32 %r4157, %r4152, %r4153; sub.s32 %r4158, %r4153, %r4156; shf.l.wrap.b32 %r4159, %r4156, %r4156, 19; xor.b32 %r4160, %r4159, %r4158; add.s32 %r4161, %r4156, %r4157; sub.s32 %r4162, %r4157, %r4160; shf.l.wrap.b32 %r4163, %r4160, %r4160, 4; add.s32 %r4164, %r4160, %r4161; add.s32 %r4165, %r4161, %r3899; xor.b32 %r4166, %r4164, %r4162; xor.b32 %r4167, %r4166, %r4163; shf.l.wrap.b32 %r4168, %r4164, %r4164, 14; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, %r4165; shf.l.wrap.b32 %r4171, %r4169, %r4169, 11; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4164; shf.l.wrap.b32 %r4174, %r4172, %r4172, 25; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 16; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 4; sub.s32 %r4181, %r4179, %r4180; xor.b32 %r4182, %r4181, %r4175; shf.l.wrap.b32 %r4183, %r4181, %r4181, 14; sub.s32 %r4184, %r4182, %r4183; xor.b32 %r4185, %r4184, %r4178; shf.l.wrap.b32 %r4186, %r4184, %r4184, 24; sub.s32 %r4187, %r4185, %r4186; and.b32 %r4188, %r4187, 31; setp.lt.u32 %p672, %r4188, 24; selp.f32 %f2406, %f2358, %f2322, %p672; setp.lt.u32 %p673, %r4188, 16; selp.f32 %f2407, %f2322, %f2394, %p673; setp.lt.u32 %p674, %r4188, 8; selp.f32 %f2408, %f2394, %f2326, %p674; and.b32 %r4189, %r4187, 1; setp.eq.b32 %p675, %r4189, 1; neg.ftz.f32 %f2409, %f2406; selp.f32 %f2410, %f2409, %f2406, %p675; and.b32 %r4190, %r4187, 2; setp.eq.s32 %p676, %r4190, 0; neg.ftz.f32 %f2411, %f2407; selp.f32 %f2412, %f2407, %f2411, %p676; add.ftz.f32 %f2413, %f2410, %f2412; and.b32 %r4191, %r4187, 4; setp.eq.s32 %p677, %r4191, 0; neg.ftz.f32 %f2414, %f2408; selp.f32 %f2415, %f2408, %f2414, %p677; add.ftz.f32 %f2416, %f2415, %f2413; add.s32 %r4192, %r4094, %r4001; sub.s32 %r4193, %r4001, %r4097; xor.b32 %r4194, %r4100, %r4193; add.s32 %r4195, %r4097, %r4192; sub.s32 %r4196, %r4192, %r4194; shf.l.wrap.b32 %r4197, %r4194, %r4194, 8; xor.b32 %r4198, %r4197, %r4196; add.s32 %r4199, %r4194, %r4195; sub.s32 %r4200, %r4195, %r4198; shf.l.wrap.b32 %r4201, %r4198, %r4198, 16; xor.b32 %r4202, %r4201, %r4200; add.s32 %r4203, %r4198, %r4199; sub.s32 %r4204, %r4199, %r4202; shf.l.wrap.b32 %r4205, %r4202, %r4202, 19; xor.b32 %r4206, %r4205, %r4204; add.s32 %r4207, %r4202, %r4203; sub.s32 %r4208, %r4203, %r4206; shf.l.wrap.b32 %r4209, %r4206, %r4206, 4; add.s32 %r4210, %r4206, %r4207; add.s32 %r4211, %r4207, %r3899; xor.b32 %r4212, %r4210, %r4208; xor.b32 %r4213, %r4212, %r4209; shf.l.wrap.b32 %r4214, %r4210, %r4210, 14; sub.s32 %r4215, %r4213, %r4214; xor.b32 %r4216, %r4215, %r4211; shf.l.wrap.b32 %r4217, %r4215, %r4215, 11; sub.s32 %r4218, %r4216, %r4217; xor.b32 %r4219, %r4218, %r4210; shf.l.wrap.b32 %r4220, %r4218, %r4218, 25; sub.s32 %r4221, %r4219, %r4220; xor.b32 %r4222, %r4221, %r4215; shf.l.wrap.b32 %r4223, %r4221, %r4221, 16; sub.s32 %r4224, %r4222, %r4223; xor.b32 %r4225, %r4224, %r4218; shf.l.wrap.b32 %r4226, %r4224, %r4224, 4; sub.s32 %r4227, %r4225, %r4226; xor.b32 %r4228, %r4227, %r4221; shf.l.wrap.b32 %r4229, %r4227, %r4227, 14; sub.s32 %r4230, %r4228, %r4229; xor.b32 %r4231, %r4230, %r4224; shf.l.wrap.b32 %r4232, %r4230, %r4230, 24; sub.s32 %r4233, %r4231, %r4232; and.b32 %r4234, %r4233, 31; setp.lt.u32 %p678, %r4234, 24; selp.f32 %f2417, %f2320, %f2371, %p678; setp.lt.u32 %p679, %r4234, 16; selp.f32 %f2418, %f2371, %f2394, %p679; setp.lt.u32 %p680, %r4234, 8; selp.f32 %f2419, %f2394, %f2326, %p680; and.b32 %r4235, %r4233, 1; setp.eq.b32 %p681, %r4235, 1; neg.ftz.f32 %f2420, %f2417; selp.f32 %f2421, %f2420, %f2417, %p681; and.b32 %r4236, %r4233, 2; setp.eq.s32 %p682, %r4236, 0; neg.ftz.f32 %f2422, %f2418; selp.f32 %f2423, %f2418, %f2422, %p682; add.ftz.f32 %f2424, %f2421, %f2423; and.b32 %r4237, %r4233, 4; setp.eq.s32 %p683, %r4237, 0; neg.ftz.f32 %f2425, %f2419; selp.f32 %f2426, %f2419, %f2425, %p683; add.ftz.f32 %f2427, %f2426, %f2424; sub.s32 %r4238, %r4001, %r4145; xor.b32 %r4239, %r4147, %r4238; add.s32 %r4240, %r4145, %r4192; sub.s32 %r4241, %r4192, %r4239; shf.l.wrap.b32 %r4242, %r4239, %r4239, 8; xor.b32 %r4243, %r4242, %r4241; add.s32 %r4244, %r4239, %r4240; sub.s32 %r4245, %r4240, %r4243; shf.l.wrap.b32 %r4246, %r4243, %r4243, 16; xor.b32 %r4247, %r4246, %r4245; add.s32 %r4248, %r4243, %r4244; sub.s32 %r4249, %r4244, %r4247; shf.l.wrap.b32 %r4250, %r4247, %r4247, 19; xor.b32 %r4251, %r4250, %r4249; add.s32 %r4252, %r4247, %r4248; sub.s32 %r4253, %r4248, %r4251; shf.l.wrap.b32 %r4254, %r4251, %r4251, 4; add.s32 %r4255, %r4251, %r4252; add.s32 %r4256, %r4252, %r3899; xor.b32 %r4257, %r4255, %r4253; xor.b32 %r4258, %r4257, %r4254; shf.l.wrap.b32 %r4259, %r4255, %r4255, 14; sub.s32 %r4260, %r4258, %r4259; xor.b32 %r4261, %r4260, %r4256; shf.l.wrap.b32 %r4262, %r4260, %r4260, 11; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, %r4255; shf.l.wrap.b32 %r4265, %r4263, %r4263, 25; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 16; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 4; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 14; sub.s32 %r4275, %r4273, %r4274; xor.b32 %r4276, %r4275, %r4269; shf.l.wrap.b32 %r4277, %r4275, %r4275, 24; sub.s32 %r4278, %r4276, %r4277; and.b32 %r4279, %r4278, 31; setp.lt.u32 %p684, %r4279, 24; selp.f32 %f2428, %f2358, %f2371, %p684; setp.lt.u32 %p685, %r4279, 16; selp.f32 %f2429, %f2371, %f2394, %p685; setp.lt.u32 %p686, %r4279, 8; selp.f32 %f2430, %f2394, %f2326, %p686; and.b32 %r4280, %r4278, 1; setp.eq.b32 %p687, %r4280, 1; neg.ftz.f32 %f2431, %f2428; selp.f32 %f2432, %f2431, %f2428, %p687; and.b32 %r4281, %r4278, 2; setp.eq.s32 %p688, %r4281, 0; neg.ftz.f32 %f2433, %f2429; selp.f32 %f2434, %f2429, %f2433, %p688; add.ftz.f32 %f2435, %f2432, %f2434; and.b32 %r4282, %r4278, 4; setp.eq.s32 %p689, %r4282, 0; neg.ftz.f32 %f2436, %f2430; selp.f32 %f2437, %f2430, %f2436, %p689; add.ftz.f32 %f2438, %f2437, %f2435; add.s32 %r4283, %r3899, 1; add.s32 %r4284, %r3921, %r4283; xor.b32 %r4285, %r3929, %r4284; sub.s32 %r4286, %r4285, %r3931; xor.b32 %r4287, %r4286, %r3924; shf.l.wrap.b32 %r4288, %r4286, %r4286, 25; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r3929; shf.l.wrap.b32 %r4291, %r4289, %r4289, 16; sub.s32 %r4292, %r4290, %r4291; xor.b32 %r4293, %r4292, %r4286; shf.l.wrap.b32 %r4294, %r4292, %r4292, 4; sub.s32 %r4295, %r4293, %r4294; xor.b32 %r4296, %r4295, %r4289; shf.l.wrap.b32 %r4297, %r4295, %r4295, 14; sub.s32 %r4298, %r4296, %r4297; xor.b32 %r4299, %r4298, %r4292; shf.l.wrap.b32 %r4300, %r4298, %r4298, 24; sub.s32 %r4301, %r4299, %r4300; add.ftz.f32 %f2439, %f2326, 0fBF800000; and.b32 %r4302, %r4301, 31; setp.lt.u32 %p690, %r4302, 24; selp.f32 %f2440, %f2320, %f2322, %p690; setp.lt.u32 %p691, %r4302, 16; selp.f32 %f2441, %f2322, %f2324, %p691; setp.lt.u32 %p692, %r4302, 8; selp.f32 %f2442, %f2324, %f2439, %p692; and.b32 %r4303, %r4301, 1; setp.eq.b32 %p693, %r4303, 1; neg.ftz.f32 %f2443, %f2440; selp.f32 %f2444, %f2443, %f2440, %p693; and.b32 %r4304, %r4301, 2; setp.eq.s32 %p694, %r4304, 0; neg.ftz.f32 %f2445, %f2441; selp.f32 %f2446, %f2441, %f2445, %p694; add.ftz.f32 %f2447, %f2444, %f2446; and.b32 %r4305, %r4301, 4; setp.eq.s32 %p695, %r4305, 0; neg.ftz.f32 %f2448, %f2442; selp.f32 %f2449, %f2442, %f2448, %p695; add.ftz.f32 %f2450, %f2449, %f2447; add.s32 %r4306, %r3970, %r4283; xor.b32 %r4307, %r3978, %r4306; sub.s32 %r4308, %r4307, %r3980; xor.b32 %r4309, %r4308, %r3973; shf.l.wrap.b32 %r4310, %r4308, %r4308, 25; sub.s32 %r4311, %r4309, %r4310; xor.b32 %r4312, %r4311, %r3978; shf.l.wrap.b32 %r4313, %r4311, %r4311, 16; sub.s32 %r4314, %r4312, %r4313; xor.b32 %r4315, %r4314, %r4308; shf.l.wrap.b32 %r4316, %r4314, %r4314, 4; sub.s32 %r4317, %r4315, %r4316; xor.b32 %r4318, %r4317, %r4311; shf.l.wrap.b32 %r4319, %r4317, %r4317, 14; sub.s32 %r4320, %r4318, %r4319; xor.b32 %r4321, %r4320, %r4314; shf.l.wrap.b32 %r4322, %r4320, %r4320, 24; sub.s32 %r4323, %r4321, %r4322; and.b32 %r4324, %r4323, 31; setp.lt.u32 %p696, %r4324, 24; selp.f32 %f2451, %f2358, %f2322, %p696; setp.lt.u32 %p697, %r4324, 16; selp.f32 %f2452, %f2322, %f2324, %p697; setp.lt.u32 %p698, %r4324, 8; selp.f32 %f2453, %f2324, %f2439, %p698; and.b32 %r4325, %r4323, 1; setp.eq.b32 %p699, %r4325, 1; neg.ftz.f32 %f2454, %f2451; selp.f32 %f2455, %f2454, %f2451, %p699; and.b32 %r4326, %r4323, 2; setp.eq.s32 %p700, %r4326, 0; neg.ftz.f32 %f2456, %f2452; selp.f32 %f2457, %f2452, %f2456, %p700; add.ftz.f32 %f2458, %f2455, %f2457; and.b32 %r4327, %r4323, 4; setp.eq.s32 %p701, %r4327, 0; neg.ftz.f32 %f2459, %f2453; selp.f32 %f2460, %f2453, %f2459, %p701; add.ftz.f32 %f2461, %f2460, %f2458; add.s32 %r4328, %r4017, %r4283; xor.b32 %r4329, %r4025, %r4328; sub.s32 %r4330, %r4329, %r4027; xor.b32 %r4331, %r4330, %r4020; shf.l.wrap.b32 %r4332, %r4330, %r4330, 25; sub.s32 %r4333, %r4331, %r4332; xor.b32 %r4334, %r4333, %r4025; shf.l.wrap.b32 %r4335, %r4333, %r4333, 16; sub.s32 %r4336, %r4334, %r4335; xor.b32 %r4337, %r4336, %r4330; shf.l.wrap.b32 %r4338, %r4336, %r4336, 4; sub.s32 %r4339, %r4337, %r4338; xor.b32 %r4340, %r4339, %r4333; shf.l.wrap.b32 %r4341, %r4339, %r4339, 14; sub.s32 %r4342, %r4340, %r4341; xor.b32 %r4343, %r4342, %r4336; shf.l.wrap.b32 %r4344, %r4342, %r4342, 24; sub.s32 %r4345, %r4343, %r4344; and.b32 %r4346, %r4345, 31; setp.lt.u32 %p702, %r4346, 24; selp.f32 %f2462, %f2320, %f2371, %p702; setp.lt.u32 %p703, %r4346, 16; selp.f32 %f2463, %f2371, %f2324, %p703; setp.lt.u32 %p704, %r4346, 8; selp.f32 %f2464, %f2324, %f2439, %p704; and.b32 %r4347, %r4345, 1; setp.eq.b32 %p705, %r4347, 1; neg.ftz.f32 %f2465, %f2462; selp.f32 %f2466, %f2465, %f2462, %p705; and.b32 %r4348, %r4345, 2; setp.eq.s32 %p706, %r4348, 0; neg.ftz.f32 %f2467, %f2463; selp.f32 %f2468, %f2463, %f2467, %p706; add.ftz.f32 %f2469, %f2466, %f2468; and.b32 %r4349, %r4345, 4; setp.eq.s32 %p707, %r4349, 0; neg.ftz.f32 %f2470, %f2464; selp.f32 %f2471, %f2464, %f2470, %p707; add.ftz.f32 %f2472, %f2471, %f2469; add.s32 %r4350, %r4062, %r4283; xor.b32 %r4351, %r4070, %r4350; sub.s32 %r4352, %r4351, %r4072; xor.b32 %r4353, %r4352, %r4065; shf.l.wrap.b32 %r4354, %r4352, %r4352, 25; sub.s32 %r4355, %r4353, %r4354; xor.b32 %r4356, %r4355, %r4070; shf.l.wrap.b32 %r4357, %r4355, %r4355, 16; sub.s32 %r4358, %r4356, %r4357; xor.b32 %r4359, %r4358, %r4352; shf.l.wrap.b32 %r4360, %r4358, %r4358, 4; sub.s32 %r4361, %r4359, %r4360; xor.b32 %r4362, %r4361, %r4355; shf.l.wrap.b32 %r4363, %r4361, %r4361, 14; sub.s32 %r4364, %r4362, %r4363; xor.b32 %r4365, %r4364, %r4358; shf.l.wrap.b32 %r4366, %r4364, %r4364, 24; sub.s32 %r4367, %r4365, %r4366; and.b32 %r4368, %r4367, 31; setp.lt.u32 %p708, %r4368, 24; selp.f32 %f2473, %f2358, %f2371, %p708; setp.lt.u32 %p709, %r4368, 16; selp.f32 %f2474, %f2371, %f2324, %p709; setp.lt.u32 %p710, %r4368, 8; selp.f32 %f2475, %f2324, %f2439, %p710; and.b32 %r4369, %r4367, 1; setp.eq.b32 %p711, %r4369, 1; neg.ftz.f32 %f2476, %f2473; selp.f32 %f2477, %f2476, %f2473, %p711; and.b32 %r4370, %r4367, 2; setp.eq.s32 %p712, %r4370, 0; neg.ftz.f32 %f2478, %f2474; selp.f32 %f2479, %f2474, %f2478, %p712; add.ftz.f32 %f2480, %f2477, %f2479; and.b32 %r4371, %r4367, 4; setp.eq.s32 %p713, %r4371, 0; neg.ftz.f32 %f2481, %f2475; selp.f32 %f2482, %f2475, %f2481, %p713; add.ftz.f32 %f2483, %f2482, %f2480; add.s32 %r4372, %r4114, %r4283; xor.b32 %r4373, %r4122, %r4372; sub.s32 %r4374, %r4373, %r4124; xor.b32 %r4375, %r4374, %r4117; shf.l.wrap.b32 %r4376, %r4374, %r4374, 25; sub.s32 %r4377, %r4375, %r4376; xor.b32 %r4378, %r4377, %r4122; shf.l.wrap.b32 %r4379, %r4377, %r4377, 16; sub.s32 %r4380, %r4378, %r4379; xor.b32 %r4381, %r4380, %r4374; shf.l.wrap.b32 %r4382, %r4380, %r4380, 4; sub.s32 %r4383, %r4381, %r4382; xor.b32 %r4384, %r4383, %r4377; shf.l.wrap.b32 %r4385, %r4383, %r4383, 14; sub.s32 %r4386, %r4384, %r4385; xor.b32 %r4387, %r4386, %r4380; shf.l.wrap.b32 %r4388, %r4386, %r4386, 24; sub.s32 %r4389, %r4387, %r4388; and.b32 %r4390, %r4389, 31; setp.lt.u32 %p714, %r4390, 24; selp.f32 %f2484, %f2320, %f2322, %p714; setp.lt.u32 %p715, %r4390, 16; selp.f32 %f2485, %f2322, %f2394, %p715; setp.lt.u32 %p716, %r4390, 8; selp.f32 %f2486, %f2394, %f2439, %p716; and.b32 %r4391, %r4389, 1; setp.eq.b32 %p717, %r4391, 1; neg.ftz.f32 %f2487, %f2484; selp.f32 %f2488, %f2487, %f2484, %p717; and.b32 %r4392, %r4389, 2; setp.eq.s32 %p718, %r4392, 0; neg.ftz.f32 %f2489, %f2485; selp.f32 %f2490, %f2485, %f2489, %p718; add.ftz.f32 %f2491, %f2488, %f2490; and.b32 %r4393, %r4389, 4; setp.eq.s32 %p719, %r4393, 0; neg.ftz.f32 %f2492, %f2486; selp.f32 %f2493, %f2486, %f2492, %p719; add.ftz.f32 %f2494, %f2493, %f2491; add.s32 %r4394, %r4161, %r4283; xor.b32 %r4395, %r4169, %r4394; sub.s32 %r4396, %r4395, %r4171; xor.b32 %r4397, %r4396, %r4164; shf.l.wrap.b32 %r4398, %r4396, %r4396, 25; sub.s32 %r4399, %r4397, %r4398; xor.b32 %r4400, %r4399, %r4169; shf.l.wrap.b32 %r4401, %r4399, %r4399, 16; sub.s32 %r4402, %r4400, %r4401; xor.b32 %r4403, %r4402, %r4396; shf.l.wrap.b32 %r4404, %r4402, %r4402, 4; sub.s32 %r4405, %r4403, %r4404; xor.b32 %r4406, %r4405, %r4399; shf.l.wrap.b32 %r4407, %r4405, %r4405, 14; sub.s32 %r4408, %r4406, %r4407; xor.b32 %r4409, %r4408, %r4402; shf.l.wrap.b32 %r4410, %r4408, %r4408, 24; sub.s32 %r4411, %r4409, %r4410; and.b32 %r4412, %r4411, 31; setp.lt.u32 %p720, %r4412, 24; selp.f32 %f2495, %f2358, %f2322, %p720; setp.lt.u32 %p721, %r4412, 16; selp.f32 %f2496, %f2322, %f2394, %p721; setp.lt.u32 %p722, %r4412, 8; selp.f32 %f2497, %f2394, %f2439, %p722; and.b32 %r4413, %r4411, 1; setp.eq.b32 %p723, %r4413, 1; neg.ftz.f32 %f2498, %f2495; selp.f32 %f2499, %f2498, %f2495, %p723; and.b32 %r4414, %r4411, 2; setp.eq.s32 %p724, %r4414, 0; neg.ftz.f32 %f2500, %f2496; selp.f32 %f2501, %f2496, %f2500, %p724; add.ftz.f32 %f2502, %f2499, %f2501; and.b32 %r4415, %r4411, 4; setp.eq.s32 %p725, %r4415, 0; neg.ftz.f32 %f2503, %f2497; selp.f32 %f2504, %f2497, %f2503, %p725; add.ftz.f32 %f2505, %f2504, %f2502; add.s32 %r4416, %r4207, %r4283; xor.b32 %r4417, %r4215, %r4416; sub.s32 %r4418, %r4417, %r4217; xor.b32 %r4419, %r4418, %r4210; shf.l.wrap.b32 %r4420, %r4418, %r4418, 25; sub.s32 %r4421, %r4419, %r4420; xor.b32 %r4422, %r4421, %r4215; shf.l.wrap.b32 %r4423, %r4421, %r4421, 16; sub.s32 %r4424, %r4422, %r4423; xor.b32 %r4425, %r4424, %r4418; shf.l.wrap.b32 %r4426, %r4424, %r4424, 4; sub.s32 %r4427, %r4425, %r4426; xor.b32 %r4428, %r4427, %r4421; shf.l.wrap.b32 %r4429, %r4427, %r4427, 14; sub.s32 %r4430, %r4428, %r4429; xor.b32 %r4431, %r4430, %r4424; shf.l.wrap.b32 %r4432, %r4430, %r4430, 24; sub.s32 %r4433, %r4431, %r4432; and.b32 %r4434, %r4433, 31; setp.lt.u32 %p726, %r4434, 24; selp.f32 %f2506, %f2320, %f2371, %p726; setp.lt.u32 %p727, %r4434, 16; selp.f32 %f2507, %f2371, %f2394, %p727; setp.lt.u32 %p728, %r4434, 8; selp.f32 %f2508, %f2394, %f2439, %p728; and.b32 %r4435, %r4433, 1; setp.eq.b32 %p729, %r4435, 1; neg.ftz.f32 %f2509, %f2506; selp.f32 %f2510, %f2509, %f2506, %p729; and.b32 %r4436, %r4433, 2; setp.eq.s32 %p730, %r4436, 0; neg.ftz.f32 %f2511, %f2507; selp.f32 %f2512, %f2507, %f2511, %p730; add.ftz.f32 %f2513, %f2510, %f2512; and.b32 %r4437, %r4433, 4; setp.eq.s32 %p731, %r4437, 0; neg.ftz.f32 %f2514, %f2508; selp.f32 %f2515, %f2508, %f2514, %p731; add.ftz.f32 %f2516, %f2515, %f2513; add.s32 %r4438, %r4252, %r4283; xor.b32 %r4439, %r4260, %r4438; sub.s32 %r4440, %r4439, %r4262; xor.b32 %r4441, %r4440, %r4255; shf.l.wrap.b32 %r4442, %r4440, %r4440, 25; sub.s32 %r4443, %r4441, %r4442; xor.b32 %r4444, %r4443, %r4260; shf.l.wrap.b32 %r4445, %r4443, %r4443, 16; sub.s32 %r4446, %r4444, %r4445; xor.b32 %r4447, %r4446, %r4440; shf.l.wrap.b32 %r4448, %r4446, %r4446, 4; sub.s32 %r4449, %r4447, %r4448; xor.b32 %r4450, %r4449, %r4443; shf.l.wrap.b32 %r4451, %r4449, %r4449, 14; sub.s32 %r4452, %r4450, %r4451; xor.b32 %r4453, %r4452, %r4446; shf.l.wrap.b32 %r4454, %r4452, %r4452, 24; sub.s32 %r4455, %r4453, %r4454; and.b32 %r4456, %r4455, 31; setp.lt.u32 %p732, %r4456, 24; selp.f32 %f2517, %f2358, %f2371, %p732; setp.lt.u32 %p733, %r4456, 16; selp.f32 %f2518, %f2371, %f2394, %p733; setp.lt.u32 %p734, %r4456, 8; selp.f32 %f2519, %f2394, %f2439, %p734; and.b32 %r4457, %r4455, 1; setp.eq.b32 %p735, %r4457, 1; neg.ftz.f32 %f2520, %f2517; selp.f32 %f2521, %f2520, %f2517, %p735; and.b32 %r4458, %r4455, 2; setp.eq.s32 %p736, %r4458, 0; neg.ftz.f32 %f2522, %f2518; selp.f32 %f2523, %f2518, %f2522, %p736; add.ftz.f32 %f2524, %f2521, %f2523; and.b32 %r4459, %r4455, 4; setp.eq.s32 %p737, %r4459, 0; neg.ftz.f32 %f2525, %f2519; selp.f32 %f2526, %f2519, %f2525, %p737; add.ftz.f32 %f2527, %f2526, %f2524; sub.ftz.f32 %f2528, %f2359, %f2331; sub.ftz.f32 %f2529, %f2359, %f2336; sub.ftz.f32 %f2530, %f2359, %f2341; mul.ftz.f32 %f2531, %f2528, %f2357; fma.rn.ftz.f32 %f2532, %f2331, %f2370, %f2531; mul.ftz.f32 %f2533, %f2528, %f2382; fma.rn.ftz.f32 %f2534, %f2331, %f2393, %f2533; mul.ftz.f32 %f2535, %f2336, %f2534; fma.rn.ftz.f32 %f2536, %f2529, %f2532, %f2535; mul.ftz.f32 %f2537, %f2528, %f2405; fma.rn.ftz.f32 %f2538, %f2331, %f2416, %f2537; mul.ftz.f32 %f2539, %f2528, %f2427; fma.rn.ftz.f32 %f2540, %f2331, %f2438, %f2539; mul.ftz.f32 %f2541, %f2336, %f2540; fma.rn.ftz.f32 %f2542, %f2529, %f2538, %f2541; mul.ftz.f32 %f2543, %f2341, %f2542; fma.rn.ftz.f32 %f2544, %f2530, %f2536, %f2543; mul.ftz.f32 %f2545, %f2528, %f2450; fma.rn.ftz.f32 %f2546, %f2331, %f2461, %f2545; mul.ftz.f32 %f2547, %f2528, %f2472; fma.rn.ftz.f32 %f2548, %f2331, %f2483, %f2547; mul.ftz.f32 %f2549, %f2336, %f2548; fma.rn.ftz.f32 %f2550, %f2529, %f2546, %f2549; mul.ftz.f32 %f2551, %f2528, %f2494; fma.rn.ftz.f32 %f2552, %f2331, %f2505, %f2551; mul.ftz.f32 %f2553, %f2528, %f2516; fma.rn.ftz.f32 %f2554, %f2331, %f2527, %f2553; mul.ftz.f32 %f2555, %f2336, %f2554; fma.rn.ftz.f32 %f2556, %f2529, %f2552, %f2555; mul.ftz.f32 %f2557, %f2341, %f2556; fma.rn.ftz.f32 %f2558, %f2530, %f2550, %f2557; sub.ftz.f32 %f2559, %f2558, %f2544; fma.rn.ftz.f32 %f115, %f2346, %f2559, %f2544; mov.b32 %r31, %f115; setp.nan.ftz.f32 %p738, %f115, %f115; mov.f32 %f5560, 0f00000000; @%p738 bra $L__BB74_90; or.b32 %r4460, %r31, -2147483648; setp.ne.s32 %p739, %r4460, -2147483648; add.ftz.f32 %f2561, %f115, %f115; setp.eq.ftz.f32 %p740, %f2561, %f115; and.pred %p741, %p740, %p739; @%p741 bra $L__BB74_90; shl.b32 %r4461, %r31, 1; setp.lt.u32 %p742, %r4461, -16777215; mul.ftz.f32 %f2562, %f115, 0f3F559B3D; selp.f32 %f5560, %f2562, 0f00000000, %p742; $L__BB74_90: add.ftz.f32 %f2563, %f21, %f5560; mul.ftz.f32 %f2564, %f5557, %f2563; setp.gt.ftz.f32 %p743, %f5559, 0f3F800000; selp.f32 %f2565, 0f3F800000, %f5559, %p743; mul.ftz.f32 %f2566, %f2565, %f114; fma.rn.ftz.f32 %f5827, %f2566, %f2564, %f5827; bra.uni $L__BB74_450; $L__BB74_71: lg2.approx.ftz.f32 %f1564, %f27; mul.ftz.f32 %f1565, %f25, %f1564; neg.ftz.f32 %f1566, %f1565; ex2.approx.ftz.f32 %f67, %f1566; cvt.rmi.ftz.f32.f32 %f1567, %f5568; cvt.rzi.ftz.s32.f32 %r2196, %f1567; sub.ftz.f32 %f1568, %f5568, %f1567; cvt.rmi.ftz.f32.f32 %f1569, %f5569; cvt.rzi.ftz.s32.f32 %r2197, %f1569; sub.ftz.f32 %f1570, %f5569, %f1569; cvt.rmi.ftz.f32.f32 %f1571, %f5570; cvt.rzi.ftz.s32.f32 %r2198, %f1571; sub.ftz.f32 %f1572, %f5570, %f1571; cvt.rmi.ftz.f32.f32 %f1573, %f5571; cvt.rzi.ftz.s32.f32 %r2199, %f1573; sub.ftz.f32 %f1574, %f5571, %f1573; mul.ftz.f32 %f1575, %f1568, %f1568; mul.ftz.f32 %f1576, %f1568, %f1575; fma.rn.ftz.f32 %f1577, %f1568, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1578, %f1568, %f1577, 0f41200000; mul.ftz.f32 %f1579, %f1576, %f1578; mul.ftz.f32 %f1580, %f1570, %f1570; mul.ftz.f32 %f1581, %f1570, %f1580; fma.rn.ftz.f32 %f1582, %f1570, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1583, %f1570, %f1582, 0f41200000; mul.ftz.f32 %f1584, %f1581, %f1583; mul.ftz.f32 %f1585, %f1572, %f1572; mul.ftz.f32 %f1586, %f1572, %f1585; fma.rn.ftz.f32 %f1587, %f1572, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1588, %f1572, %f1587, 0f41200000; mul.ftz.f32 %f1589, %f1586, %f1588; mul.ftz.f32 %f1590, %f1574, %f1574; mul.ftz.f32 %f1591, %f1574, %f1590; fma.rn.ftz.f32 %f1592, %f1574, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1593, %f1574, %f1592, 0f41200000; mul.ftz.f32 %f1594, %f1591, %f1593; add.s32 %r2200, %r2197, -559038708; add.s32 %r2201, %r2198, -559038708; sub.s32 %r2202, %r2196, %r2198; shf.l.wrap.b32 %r2203, %r2201, %r2201, 4; xor.b32 %r2204, %r2203, %r2202; add.s32 %r2205, %r2201, %r2200; sub.s32 %r2206, %r2200, %r2204; shf.l.wrap.b32 %r2207, %r2204, %r2204, 6; xor.b32 %r2208, %r2207, %r2206; add.s32 %r2209, %r2204, %r2205; sub.s32 %r2210, %r2205, %r2208; shf.l.wrap.b32 %r2211, %r2208, %r2208, 8; xor.b32 %r2212, %r2211, %r2210; add.s32 %r2213, %r2208, %r2209; sub.s32 %r2214, %r2209, %r2212; shf.l.wrap.b32 %r2215, %r2212, %r2212, 16; xor.b32 %r2216, %r2215, %r2214; add.s32 %r2217, %r2212, %r2213; sub.s32 %r2218, %r2213, %r2216; shf.l.wrap.b32 %r2219, %r2216, %r2216, 19; xor.b32 %r2220, %r2219, %r2218; add.s32 %r2221, %r2216, %r2217; sub.s32 %r2222, %r2217, %r2220; shf.l.wrap.b32 %r2223, %r2220, %r2220, 4; add.s32 %r2224, %r2220, %r2221; add.s32 %r2225, %r2221, %r2199; xor.b32 %r2226, %r2224, %r2222; xor.b32 %r2227, %r2226, %r2223; shf.l.wrap.b32 %r2228, %r2224, %r2224, 14; sub.s32 %r2229, %r2227, %r2228; xor.b32 %r2230, %r2229, %r2225; shf.l.wrap.b32 %r2231, %r2229, %r2229, 11; sub.s32 %r2232, %r2230, %r2231; xor.b32 %r2233, %r2232, %r2224; shf.l.wrap.b32 %r2234, %r2232, %r2232, 25; sub.s32 %r2235, %r2233, %r2234; xor.b32 %r2236, %r2235, %r2229; shf.l.wrap.b32 %r2237, %r2235, %r2235, 16; sub.s32 %r2238, %r2236, %r2237; xor.b32 %r2239, %r2238, %r2232; shf.l.wrap.b32 %r2240, %r2238, %r2238, 4; sub.s32 %r2241, %r2239, %r2240; xor.b32 %r2242, %r2241, %r2235; shf.l.wrap.b32 %r2243, %r2241, %r2241, 14; sub.s32 %r2244, %r2242, %r2243; xor.b32 %r2245, %r2244, %r2238; shf.l.wrap.b32 %r2246, %r2244, %r2244, 24; sub.s32 %r2247, %r2245, %r2246; and.b32 %r2248, %r2247, 31; setp.lt.u32 %p329, %r2248, 24; selp.f32 %f1595, %f1568, %f1570, %p329; setp.lt.u32 %p330, %r2248, 16; selp.f32 %f1596, %f1570, %f1572, %p330; setp.lt.u32 %p331, %r2248, 8; selp.f32 %f1597, %f1572, %f1574, %p331; and.b32 %r2249, %r2247, 1; setp.eq.b32 %p332, %r2249, 1; neg.ftz.f32 %f1598, %f1595; selp.f32 %f1599, %f1598, %f1595, %p332; and.b32 %r2250, %r2247, 2; setp.eq.s32 %p333, %r2250, 0; neg.ftz.f32 %f1600, %f1596; selp.f32 %f1601, %f1596, %f1600, %p333; add.ftz.f32 %f1602, %f1599, %f1601; and.b32 %r2251, %r2247, 4; setp.eq.s32 %p334, %r2251, 0; neg.ftz.f32 %f1603, %f1597; selp.f32 %f1604, %f1597, %f1603, %p334; add.ftz.f32 %f1605, %f1604, %f1602; add.s32 %r2252, %r2196, 1; sub.s32 %r2253, %r2252, %r2198; xor.b32 %r2254, %r2203, %r2253; sub.s32 %r2255, %r2200, %r2254; shf.l.wrap.b32 %r2256, %r2254, %r2254, 6; xor.b32 %r2257, %r2256, %r2255; add.s32 %r2258, %r2254, %r2205; sub.s32 %r2259, %r2205, %r2257; shf.l.wrap.b32 %r2260, %r2257, %r2257, 8; xor.b32 %r2261, %r2260, %r2259; add.s32 %r2262, %r2257, %r2258; sub.s32 %r2263, %r2258, %r2261; shf.l.wrap.b32 %r2264, %r2261, %r2261, 16; xor.b32 %r2265, %r2264, %r2263; add.s32 %r2266, %r2261, %r2262; sub.s32 %r2267, %r2262, %r2265; shf.l.wrap.b32 %r2268, %r2265, %r2265, 19; xor.b32 %r2269, %r2268, %r2267; add.s32 %r2270, %r2265, %r2266; sub.s32 %r2271, %r2266, %r2269; shf.l.wrap.b32 %r2272, %r2269, %r2269, 4; add.s32 %r2273, %r2269, %r2270; add.s32 %r2274, %r2270, %r2199; xor.b32 %r2275, %r2273, %r2271; xor.b32 %r2276, %r2275, %r2272; shf.l.wrap.b32 %r2277, %r2273, %r2273, 14; sub.s32 %r2278, %r2276, %r2277; xor.b32 %r2279, %r2278, %r2274; shf.l.wrap.b32 %r2280, %r2278, %r2278, 11; sub.s32 %r2281, %r2279, %r2280; xor.b32 %r2282, %r2281, %r2273; shf.l.wrap.b32 %r2283, %r2281, %r2281, 25; sub.s32 %r2284, %r2282, %r2283; xor.b32 %r2285, %r2284, %r2278; shf.l.wrap.b32 %r2286, %r2284, %r2284, 16; sub.s32 %r2287, %r2285, %r2286; xor.b32 %r2288, %r2287, %r2281; shf.l.wrap.b32 %r2289, %r2287, %r2287, 4; sub.s32 %r2290, %r2288, %r2289; xor.b32 %r2291, %r2290, %r2284; shf.l.wrap.b32 %r2292, %r2290, %r2290, 14; sub.s32 %r2293, %r2291, %r2292; xor.b32 %r2294, %r2293, %r2287; shf.l.wrap.b32 %r2295, %r2293, %r2293, 24; sub.s32 %r2296, %r2294, %r2295; add.ftz.f32 %f1606, %f1568, 0fBF800000; mov.f32 %f1607, 0f3F800000; and.b32 %r2297, %r2296, 31; setp.lt.u32 %p335, %r2297, 24; selp.f32 %f1608, %f1606, %f1570, %p335; setp.lt.u32 %p336, %r2297, 16; selp.f32 %f1609, %f1570, %f1572, %p336; setp.lt.u32 %p337, %r2297, 8; selp.f32 %f1610, %f1572, %f1574, %p337; and.b32 %r2298, %r2296, 1; setp.eq.b32 %p338, %r2298, 1; neg.ftz.f32 %f1611, %f1608; selp.f32 %f1612, %f1611, %f1608, %p338; and.b32 %r2299, %r2296, 2; setp.eq.s32 %p339, %r2299, 0; neg.ftz.f32 %f1613, %f1609; selp.f32 %f1614, %f1609, %f1613, %p339; add.ftz.f32 %f1615, %f1612, %f1614; and.b32 %r2300, %r2296, 4; setp.eq.s32 %p340, %r2300, 0; neg.ftz.f32 %f1616, %f1610; selp.f32 %f1617, %f1610, %f1616, %p340; add.ftz.f32 %f1618, %f1617, %f1615; add.s32 %r2301, %r2197, -559038707; add.s32 %r2302, %r2201, %r2301; sub.s32 %r2303, %r2301, %r2204; xor.b32 %r2304, %r2207, %r2303; add.s32 %r2305, %r2204, %r2302; sub.s32 %r2306, %r2302, %r2304; shf.l.wrap.b32 %r2307, %r2304, %r2304, 8; xor.b32 %r2308, %r2307, %r2306; add.s32 %r2309, %r2304, %r2305; sub.s32 %r2310, %r2305, %r2308; shf.l.wrap.b32 %r2311, %r2308, %r2308, 16; xor.b32 %r2312, %r2311, %r2310; add.s32 %r2313, %r2308, %r2309; sub.s32 %r2314, %r2309, %r2312; shf.l.wrap.b32 %r2315, %r2312, %r2312, 19; xor.b32 %r2316, %r2315, %r2314; add.s32 %r2317, %r2312, %r2313; sub.s32 %r2318, %r2313, %r2316; shf.l.wrap.b32 %r2319, %r2316, %r2316, 4; add.s32 %r2320, %r2316, %r2317; add.s32 %r2321, %r2317, %r2199; xor.b32 %r2322, %r2320, %r2318; xor.b32 %r2323, %r2322, %r2319; shf.l.wrap.b32 %r2324, %r2320, %r2320, 14; sub.s32 %r2325, %r2323, %r2324; xor.b32 %r2326, %r2325, %r2321; shf.l.wrap.b32 %r2327, %r2325, %r2325, 11; sub.s32 %r2328, %r2326, %r2327; xor.b32 %r2329, %r2328, %r2320; shf.l.wrap.b32 %r2330, %r2328, %r2328, 25; sub.s32 %r2331, %r2329, %r2330; xor.b32 %r2332, %r2331, %r2325; shf.l.wrap.b32 %r2333, %r2331, %r2331, 16; sub.s32 %r2334, %r2332, %r2333; xor.b32 %r2335, %r2334, %r2328; shf.l.wrap.b32 %r2336, %r2334, %r2334, 4; sub.s32 %r2337, %r2335, %r2336; xor.b32 %r2338, %r2337, %r2331; shf.l.wrap.b32 %r2339, %r2337, %r2337, 14; sub.s32 %r2340, %r2338, %r2339; xor.b32 %r2341, %r2340, %r2334; shf.l.wrap.b32 %r2342, %r2340, %r2340, 24; sub.s32 %r2343, %r2341, %r2342; add.ftz.f32 %f1619, %f1570, 0fBF800000; and.b32 %r2344, %r2343, 31; setp.lt.u32 %p341, %r2344, 24; selp.f32 %f1620, %f1568, %f1619, %p341; setp.lt.u32 %p342, %r2344, 16; selp.f32 %f1621, %f1619, %f1572, %p342; setp.lt.u32 %p343, %r2344, 8; selp.f32 %f1622, %f1572, %f1574, %p343; and.b32 %r2345, %r2343, 1; setp.eq.b32 %p344, %r2345, 1; neg.ftz.f32 %f1623, %f1620; selp.f32 %f1624, %f1623, %f1620, %p344; and.b32 %r2346, %r2343, 2; setp.eq.s32 %p345, %r2346, 0; neg.ftz.f32 %f1625, %f1621; selp.f32 %f1626, %f1621, %f1625, %p345; add.ftz.f32 %f1627, %f1624, %f1626; and.b32 %r2347, %r2343, 4; setp.eq.s32 %p346, %r2347, 0; neg.ftz.f32 %f1628, %f1622; selp.f32 %f1629, %f1622, %f1628, %p346; add.ftz.f32 %f1630, %f1629, %f1627; sub.s32 %r2348, %r2301, %r2254; xor.b32 %r2349, %r2256, %r2348; add.s32 %r2350, %r2254, %r2302; sub.s32 %r2351, %r2302, %r2349; shf.l.wrap.b32 %r2352, %r2349, %r2349, 8; xor.b32 %r2353, %r2352, %r2351; add.s32 %r2354, %r2349, %r2350; sub.s32 %r2355, %r2350, %r2353; shf.l.wrap.b32 %r2356, %r2353, %r2353, 16; xor.b32 %r2357, %r2356, %r2355; add.s32 %r2358, %r2353, %r2354; sub.s32 %r2359, %r2354, %r2357; shf.l.wrap.b32 %r2360, %r2357, %r2357, 19; xor.b32 %r2361, %r2360, %r2359; add.s32 %r2362, %r2357, %r2358; sub.s32 %r2363, %r2358, %r2361; shf.l.wrap.b32 %r2364, %r2361, %r2361, 4; add.s32 %r2365, %r2361, %r2362; add.s32 %r2366, %r2362, %r2199; xor.b32 %r2367, %r2365, %r2363; xor.b32 %r2368, %r2367, %r2364; shf.l.wrap.b32 %r2369, %r2365, %r2365, 14; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2366; shf.l.wrap.b32 %r2372, %r2370, %r2370, 11; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2365; shf.l.wrap.b32 %r2375, %r2373, %r2373, 25; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 16; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 4; sub.s32 %r2382, %r2380, %r2381; xor.b32 %r2383, %r2382, %r2376; shf.l.wrap.b32 %r2384, %r2382, %r2382, 14; sub.s32 %r2385, %r2383, %r2384; xor.b32 %r2386, %r2385, %r2379; shf.l.wrap.b32 %r2387, %r2385, %r2385, 24; sub.s32 %r2388, %r2386, %r2387; and.b32 %r2389, %r2388, 31; setp.lt.u32 %p347, %r2389, 24; selp.f32 %f1631, %f1606, %f1619, %p347; setp.lt.u32 %p348, %r2389, 16; selp.f32 %f1632, %f1619, %f1572, %p348; setp.lt.u32 %p349, %r2389, 8; selp.f32 %f1633, %f1572, %f1574, %p349; and.b32 %r2390, %r2388, 1; setp.eq.b32 %p350, %r2390, 1; neg.ftz.f32 %f1634, %f1631; selp.f32 %f1635, %f1634, %f1631, %p350; and.b32 %r2391, %r2388, 2; setp.eq.s32 %p351, %r2391, 0; neg.ftz.f32 %f1636, %f1632; selp.f32 %f1637, %f1632, %f1636, %p351; add.ftz.f32 %f1638, %f1635, %f1637; and.b32 %r2392, %r2388, 4; setp.eq.s32 %p352, %r2392, 0; neg.ftz.f32 %f1639, %f1633; selp.f32 %f1640, %f1633, %f1639, %p352; add.ftz.f32 %f1641, %f1640, %f1638; not.b32 %r2393, %r2198; add.s32 %r2394, %r2198, -559038707; add.s32 %r2395, %r2196, %r2393; shf.l.wrap.b32 %r2396, %r2394, %r2394, 4; xor.b32 %r2397, %r2396, %r2395; add.s32 %r2398, %r2394, %r2200; sub.s32 %r2399, %r2200, %r2397; shf.l.wrap.b32 %r2400, %r2397, %r2397, 6; xor.b32 %r2401, %r2400, %r2399; add.s32 %r2402, %r2397, %r2398; sub.s32 %r2403, %r2398, %r2401; shf.l.wrap.b32 %r2404, %r2401, %r2401, 8; xor.b32 %r2405, %r2404, %r2403; add.s32 %r2406, %r2401, %r2402; sub.s32 %r2407, %r2402, %r2405; shf.l.wrap.b32 %r2408, %r2405, %r2405, 16; xor.b32 %r2409, %r2408, %r2407; add.s32 %r2410, %r2405, %r2406; sub.s32 %r2411, %r2406, %r2409; shf.l.wrap.b32 %r2412, %r2409, %r2409, 19; xor.b32 %r2413, %r2412, %r2411; add.s32 %r2414, %r2409, %r2410; sub.s32 %r2415, %r2410, %r2413; shf.l.wrap.b32 %r2416, %r2413, %r2413, 4; add.s32 %r2417, %r2413, %r2414; add.s32 %r2418, %r2414, %r2199; xor.b32 %r2419, %r2417, %r2415; xor.b32 %r2420, %r2419, %r2416; shf.l.wrap.b32 %r2421, %r2417, %r2417, 14; sub.s32 %r2422, %r2420, %r2421; xor.b32 %r2423, %r2422, %r2418; shf.l.wrap.b32 %r2424, %r2422, %r2422, 11; sub.s32 %r2425, %r2423, %r2424; xor.b32 %r2426, %r2425, %r2417; shf.l.wrap.b32 %r2427, %r2425, %r2425, 25; sub.s32 %r2428, %r2426, %r2427; xor.b32 %r2429, %r2428, %r2422; shf.l.wrap.b32 %r2430, %r2428, %r2428, 16; sub.s32 %r2431, %r2429, %r2430; xor.b32 %r2432, %r2431, %r2425; shf.l.wrap.b32 %r2433, %r2431, %r2431, 4; sub.s32 %r2434, %r2432, %r2433; xor.b32 %r2435, %r2434, %r2428; shf.l.wrap.b32 %r2436, %r2434, %r2434, 14; sub.s32 %r2437, %r2435, %r2436; xor.b32 %r2438, %r2437, %r2431; shf.l.wrap.b32 %r2439, %r2437, %r2437, 24; sub.s32 %r2440, %r2438, %r2439; add.ftz.f32 %f1642, %f1572, 0fBF800000; and.b32 %r2441, %r2440, 31; setp.lt.u32 %p353, %r2441, 24; selp.f32 %f1643, %f1568, %f1570, %p353; setp.lt.u32 %p354, %r2441, 16; selp.f32 %f1644, %f1570, %f1642, %p354; setp.lt.u32 %p355, %r2441, 8; selp.f32 %f1645, %f1642, %f1574, %p355; and.b32 %r2442, %r2440, 1; setp.eq.b32 %p356, %r2442, 1; neg.ftz.f32 %f1646, %f1643; selp.f32 %f1647, %f1646, %f1643, %p356; and.b32 %r2443, %r2440, 2; setp.eq.s32 %p357, %r2443, 0; neg.ftz.f32 %f1648, %f1644; selp.f32 %f1649, %f1644, %f1648, %p357; add.ftz.f32 %f1650, %f1647, %f1649; and.b32 %r2444, %r2440, 4; setp.eq.s32 %p358, %r2444, 0; neg.ftz.f32 %f1651, %f1645; selp.f32 %f1652, %f1645, %f1651, %p358; add.ftz.f32 %f1653, %f1652, %f1650; xor.b32 %r2445, %r2396, %r2202; sub.s32 %r2446, %r2200, %r2445; shf.l.wrap.b32 %r2447, %r2445, %r2445, 6; xor.b32 %r2448, %r2447, %r2446; add.s32 %r2449, %r2445, %r2398; sub.s32 %r2450, %r2398, %r2448; shf.l.wrap.b32 %r2451, %r2448, %r2448, 8; xor.b32 %r2452, %r2451, %r2450; add.s32 %r2453, %r2448, %r2449; sub.s32 %r2454, %r2449, %r2452; shf.l.wrap.b32 %r2455, %r2452, %r2452, 16; xor.b32 %r2456, %r2455, %r2454; add.s32 %r2457, %r2452, %r2453; sub.s32 %r2458, %r2453, %r2456; shf.l.wrap.b32 %r2459, %r2456, %r2456, 19; xor.b32 %r2460, %r2459, %r2458; add.s32 %r2461, %r2456, %r2457; sub.s32 %r2462, %r2457, %r2460; shf.l.wrap.b32 %r2463, %r2460, %r2460, 4; add.s32 %r2464, %r2460, %r2461; add.s32 %r2465, %r2461, %r2199; xor.b32 %r2466, %r2464, %r2462; xor.b32 %r2467, %r2466, %r2463; shf.l.wrap.b32 %r2468, %r2464, %r2464, 14; sub.s32 %r2469, %r2467, %r2468; xor.b32 %r2470, %r2469, %r2465; shf.l.wrap.b32 %r2471, %r2469, %r2469, 11; sub.s32 %r2472, %r2470, %r2471; xor.b32 %r2473, %r2472, %r2464; shf.l.wrap.b32 %r2474, %r2472, %r2472, 25; sub.s32 %r2475, %r2473, %r2474; xor.b32 %r2476, %r2475, %r2469; shf.l.wrap.b32 %r2477, %r2475, %r2475, 16; sub.s32 %r2478, %r2476, %r2477; xor.b32 %r2479, %r2478, %r2472; shf.l.wrap.b32 %r2480, %r2478, %r2478, 4; sub.s32 %r2481, %r2479, %r2480; xor.b32 %r2482, %r2481, %r2475; shf.l.wrap.b32 %r2483, %r2481, %r2481, 14; sub.s32 %r2484, %r2482, %r2483; xor.b32 %r2485, %r2484, %r2478; shf.l.wrap.b32 %r2486, %r2484, %r2484, 24; sub.s32 %r2487, %r2485, %r2486; and.b32 %r2488, %r2487, 31; setp.lt.u32 %p359, %r2488, 24; selp.f32 %f1654, %f1606, %f1570, %p359; setp.lt.u32 %p360, %r2488, 16; selp.f32 %f1655, %f1570, %f1642, %p360; setp.lt.u32 %p361, %r2488, 8; selp.f32 %f1656, %f1642, %f1574, %p361; and.b32 %r2489, %r2487, 1; setp.eq.b32 %p362, %r2489, 1; neg.ftz.f32 %f1657, %f1654; selp.f32 %f1658, %f1657, %f1654, %p362; and.b32 %r2490, %r2487, 2; setp.eq.s32 %p363, %r2490, 0; neg.ftz.f32 %f1659, %f1655; selp.f32 %f1660, %f1655, %f1659, %p363; add.ftz.f32 %f1661, %f1658, %f1660; and.b32 %r2491, %r2487, 4; setp.eq.s32 %p364, %r2491, 0; neg.ftz.f32 %f1662, %f1656; selp.f32 %f1663, %f1656, %f1662, %p364; add.ftz.f32 %f1664, %f1663, %f1661; add.s32 %r2492, %r2394, %r2301; sub.s32 %r2493, %r2301, %r2397; xor.b32 %r2494, %r2400, %r2493; add.s32 %r2495, %r2397, %r2492; sub.s32 %r2496, %r2492, %r2494; shf.l.wrap.b32 %r2497, %r2494, %r2494, 8; xor.b32 %r2498, %r2497, %r2496; add.s32 %r2499, %r2494, %r2495; sub.s32 %r2500, %r2495, %r2498; shf.l.wrap.b32 %r2501, %r2498, %r2498, 16; xor.b32 %r2502, %r2501, %r2500; add.s32 %r2503, %r2498, %r2499; sub.s32 %r2504, %r2499, %r2502; shf.l.wrap.b32 %r2505, %r2502, %r2502, 19; xor.b32 %r2506, %r2505, %r2504; add.s32 %r2507, %r2502, %r2503; sub.s32 %r2508, %r2503, %r2506; shf.l.wrap.b32 %r2509, %r2506, %r2506, 4; add.s32 %r2510, %r2506, %r2507; add.s32 %r2511, %r2507, %r2199; xor.b32 %r2512, %r2510, %r2508; xor.b32 %r2513, %r2512, %r2509; shf.l.wrap.b32 %r2514, %r2510, %r2510, 14; sub.s32 %r2515, %r2513, %r2514; xor.b32 %r2516, %r2515, %r2511; shf.l.wrap.b32 %r2517, %r2515, %r2515, 11; sub.s32 %r2518, %r2516, %r2517; xor.b32 %r2519, %r2518, %r2510; shf.l.wrap.b32 %r2520, %r2518, %r2518, 25; sub.s32 %r2521, %r2519, %r2520; xor.b32 %r2522, %r2521, %r2515; shf.l.wrap.b32 %r2523, %r2521, %r2521, 16; sub.s32 %r2524, %r2522, %r2523; xor.b32 %r2525, %r2524, %r2518; shf.l.wrap.b32 %r2526, %r2524, %r2524, 4; sub.s32 %r2527, %r2525, %r2526; xor.b32 %r2528, %r2527, %r2521; shf.l.wrap.b32 %r2529, %r2527, %r2527, 14; sub.s32 %r2530, %r2528, %r2529; xor.b32 %r2531, %r2530, %r2524; shf.l.wrap.b32 %r2532, %r2530, %r2530, 24; sub.s32 %r2533, %r2531, %r2532; and.b32 %r2534, %r2533, 31; setp.lt.u32 %p365, %r2534, 24; selp.f32 %f1665, %f1568, %f1619, %p365; setp.lt.u32 %p366, %r2534, 16; selp.f32 %f1666, %f1619, %f1642, %p366; setp.lt.u32 %p367, %r2534, 8; selp.f32 %f1667, %f1642, %f1574, %p367; and.b32 %r2535, %r2533, 1; setp.eq.b32 %p368, %r2535, 1; neg.ftz.f32 %f1668, %f1665; selp.f32 %f1669, %f1668, %f1665, %p368; and.b32 %r2536, %r2533, 2; setp.eq.s32 %p369, %r2536, 0; neg.ftz.f32 %f1670, %f1666; selp.f32 %f1671, %f1666, %f1670, %p369; add.ftz.f32 %f1672, %f1669, %f1671; and.b32 %r2537, %r2533, 4; setp.eq.s32 %p370, %r2537, 0; neg.ftz.f32 %f1673, %f1667; selp.f32 %f1674, %f1667, %f1673, %p370; add.ftz.f32 %f1675, %f1674, %f1672; sub.s32 %r2538, %r2301, %r2445; xor.b32 %r2539, %r2447, %r2538; add.s32 %r2540, %r2445, %r2492; sub.s32 %r2541, %r2492, %r2539; shf.l.wrap.b32 %r2542, %r2539, %r2539, 8; xor.b32 %r2543, %r2542, %r2541; add.s32 %r2544, %r2539, %r2540; sub.s32 %r2545, %r2540, %r2543; shf.l.wrap.b32 %r2546, %r2543, %r2543, 16; xor.b32 %r2547, %r2546, %r2545; add.s32 %r2548, %r2543, %r2544; sub.s32 %r2549, %r2544, %r2547; shf.l.wrap.b32 %r2550, %r2547, %r2547, 19; xor.b32 %r2551, %r2550, %r2549; add.s32 %r2552, %r2547, %r2548; sub.s32 %r2553, %r2548, %r2551; shf.l.wrap.b32 %r2554, %r2551, %r2551, 4; add.s32 %r2555, %r2551, %r2552; add.s32 %r2556, %r2552, %r2199; xor.b32 %r2557, %r2555, %r2553; xor.b32 %r2558, %r2557, %r2554; shf.l.wrap.b32 %r2559, %r2555, %r2555, 14; sub.s32 %r2560, %r2558, %r2559; xor.b32 %r2561, %r2560, %r2556; shf.l.wrap.b32 %r2562, %r2560, %r2560, 11; sub.s32 %r2563, %r2561, %r2562; xor.b32 %r2564, %r2563, %r2555; shf.l.wrap.b32 %r2565, %r2563, %r2563, 25; sub.s32 %r2566, %r2564, %r2565; xor.b32 %r2567, %r2566, %r2560; shf.l.wrap.b32 %r2568, %r2566, %r2566, 16; sub.s32 %r2569, %r2567, %r2568; xor.b32 %r2570, %r2569, %r2563; shf.l.wrap.b32 %r2571, %r2569, %r2569, 4; sub.s32 %r2572, %r2570, %r2571; xor.b32 %r2573, %r2572, %r2566; shf.l.wrap.b32 %r2574, %r2572, %r2572, 14; sub.s32 %r2575, %r2573, %r2574; xor.b32 %r2576, %r2575, %r2569; shf.l.wrap.b32 %r2577, %r2575, %r2575, 24; sub.s32 %r2578, %r2576, %r2577; and.b32 %r2579, %r2578, 31; setp.lt.u32 %p371, %r2579, 24; selp.f32 %f1676, %f1606, %f1619, %p371; setp.lt.u32 %p372, %r2579, 16; selp.f32 %f1677, %f1619, %f1642, %p372; setp.lt.u32 %p373, %r2579, 8; selp.f32 %f1678, %f1642, %f1574, %p373; and.b32 %r2580, %r2578, 1; setp.eq.b32 %p374, %r2580, 1; neg.ftz.f32 %f1679, %f1676; selp.f32 %f1680, %f1679, %f1676, %p374; and.b32 %r2581, %r2578, 2; setp.eq.s32 %p375, %r2581, 0; neg.ftz.f32 %f1681, %f1677; selp.f32 %f1682, %f1677, %f1681, %p375; add.ftz.f32 %f1683, %f1680, %f1682; and.b32 %r2582, %r2578, 4; setp.eq.s32 %p376, %r2582, 0; neg.ftz.f32 %f1684, %f1678; selp.f32 %f1685, %f1678, %f1684, %p376; add.ftz.f32 %f1686, %f1685, %f1683; add.s32 %r2583, %r2199, 1; add.s32 %r2584, %r2221, %r2583; xor.b32 %r2585, %r2229, %r2584; sub.s32 %r2586, %r2585, %r2231; xor.b32 %r2587, %r2586, %r2224; shf.l.wrap.b32 %r2588, %r2586, %r2586, 25; sub.s32 %r2589, %r2587, %r2588; xor.b32 %r2590, %r2589, %r2229; shf.l.wrap.b32 %r2591, %r2589, %r2589, 16; sub.s32 %r2592, %r2590, %r2591; xor.b32 %r2593, %r2592, %r2586; shf.l.wrap.b32 %r2594, %r2592, %r2592, 4; sub.s32 %r2595, %r2593, %r2594; xor.b32 %r2596, %r2595, %r2589; shf.l.wrap.b32 %r2597, %r2595, %r2595, 14; sub.s32 %r2598, %r2596, %r2597; xor.b32 %r2599, %r2598, %r2592; shf.l.wrap.b32 %r2600, %r2598, %r2598, 24; sub.s32 %r2601, %r2599, %r2600; add.ftz.f32 %f1687, %f1574, 0fBF800000; and.b32 %r2602, %r2601, 31; setp.lt.u32 %p377, %r2602, 24; selp.f32 %f1688, %f1568, %f1570, %p377; setp.lt.u32 %p378, %r2602, 16; selp.f32 %f1689, %f1570, %f1572, %p378; setp.lt.u32 %p379, %r2602, 8; selp.f32 %f1690, %f1572, %f1687, %p379; and.b32 %r2603, %r2601, 1; setp.eq.b32 %p380, %r2603, 1; neg.ftz.f32 %f1691, %f1688; selp.f32 %f1692, %f1691, %f1688, %p380; and.b32 %r2604, %r2601, 2; setp.eq.s32 %p381, %r2604, 0; neg.ftz.f32 %f1693, %f1689; selp.f32 %f1694, %f1689, %f1693, %p381; add.ftz.f32 %f1695, %f1692, %f1694; and.b32 %r2605, %r2601, 4; setp.eq.s32 %p382, %r2605, 0; neg.ftz.f32 %f1696, %f1690; selp.f32 %f1697, %f1690, %f1696, %p382; add.ftz.f32 %f1698, %f1697, %f1695; add.s32 %r2606, %r2270, %r2583; xor.b32 %r2607, %r2278, %r2606; sub.s32 %r2608, %r2607, %r2280; xor.b32 %r2609, %r2608, %r2273; shf.l.wrap.b32 %r2610, %r2608, %r2608, 25; sub.s32 %r2611, %r2609, %r2610; xor.b32 %r2612, %r2611, %r2278; shf.l.wrap.b32 %r2613, %r2611, %r2611, 16; sub.s32 %r2614, %r2612, %r2613; xor.b32 %r2615, %r2614, %r2608; shf.l.wrap.b32 %r2616, %r2614, %r2614, 4; sub.s32 %r2617, %r2615, %r2616; xor.b32 %r2618, %r2617, %r2611; shf.l.wrap.b32 %r2619, %r2617, %r2617, 14; sub.s32 %r2620, %r2618, %r2619; xor.b32 %r2621, %r2620, %r2614; shf.l.wrap.b32 %r2622, %r2620, %r2620, 24; sub.s32 %r2623, %r2621, %r2622; and.b32 %r2624, %r2623, 31; setp.lt.u32 %p383, %r2624, 24; selp.f32 %f1699, %f1606, %f1570, %p383; setp.lt.u32 %p384, %r2624, 16; selp.f32 %f1700, %f1570, %f1572, %p384; setp.lt.u32 %p385, %r2624, 8; selp.f32 %f1701, %f1572, %f1687, %p385; and.b32 %r2625, %r2623, 1; setp.eq.b32 %p386, %r2625, 1; neg.ftz.f32 %f1702, %f1699; selp.f32 %f1703, %f1702, %f1699, %p386; and.b32 %r2626, %r2623, 2; setp.eq.s32 %p387, %r2626, 0; neg.ftz.f32 %f1704, %f1700; selp.f32 %f1705, %f1700, %f1704, %p387; add.ftz.f32 %f1706, %f1703, %f1705; and.b32 %r2627, %r2623, 4; setp.eq.s32 %p388, %r2627, 0; neg.ftz.f32 %f1707, %f1701; selp.f32 %f1708, %f1701, %f1707, %p388; add.ftz.f32 %f1709, %f1708, %f1706; add.s32 %r2628, %r2317, %r2583; xor.b32 %r2629, %r2325, %r2628; sub.s32 %r2630, %r2629, %r2327; xor.b32 %r2631, %r2630, %r2320; shf.l.wrap.b32 %r2632, %r2630, %r2630, 25; sub.s32 %r2633, %r2631, %r2632; xor.b32 %r2634, %r2633, %r2325; shf.l.wrap.b32 %r2635, %r2633, %r2633, 16; sub.s32 %r2636, %r2634, %r2635; xor.b32 %r2637, %r2636, %r2630; shf.l.wrap.b32 %r2638, %r2636, %r2636, 4; sub.s32 %r2639, %r2637, %r2638; xor.b32 %r2640, %r2639, %r2633; shf.l.wrap.b32 %r2641, %r2639, %r2639, 14; sub.s32 %r2642, %r2640, %r2641; xor.b32 %r2643, %r2642, %r2636; shf.l.wrap.b32 %r2644, %r2642, %r2642, 24; sub.s32 %r2645, %r2643, %r2644; and.b32 %r2646, %r2645, 31; setp.lt.u32 %p389, %r2646, 24; selp.f32 %f1710, %f1568, %f1619, %p389; setp.lt.u32 %p390, %r2646, 16; selp.f32 %f1711, %f1619, %f1572, %p390; setp.lt.u32 %p391, %r2646, 8; selp.f32 %f1712, %f1572, %f1687, %p391; and.b32 %r2647, %r2645, 1; setp.eq.b32 %p392, %r2647, 1; neg.ftz.f32 %f1713, %f1710; selp.f32 %f1714, %f1713, %f1710, %p392; and.b32 %r2648, %r2645, 2; setp.eq.s32 %p393, %r2648, 0; neg.ftz.f32 %f1715, %f1711; selp.f32 %f1716, %f1711, %f1715, %p393; add.ftz.f32 %f1717, %f1714, %f1716; and.b32 %r2649, %r2645, 4; setp.eq.s32 %p394, %r2649, 0; neg.ftz.f32 %f1718, %f1712; selp.f32 %f1719, %f1712, %f1718, %p394; add.ftz.f32 %f1720, %f1719, %f1717; add.s32 %r2650, %r2362, %r2583; xor.b32 %r2651, %r2370, %r2650; sub.s32 %r2652, %r2651, %r2372; xor.b32 %r2653, %r2652, %r2365; shf.l.wrap.b32 %r2654, %r2652, %r2652, 25; sub.s32 %r2655, %r2653, %r2654; xor.b32 %r2656, %r2655, %r2370; shf.l.wrap.b32 %r2657, %r2655, %r2655, 16; sub.s32 %r2658, %r2656, %r2657; xor.b32 %r2659, %r2658, %r2652; shf.l.wrap.b32 %r2660, %r2658, %r2658, 4; sub.s32 %r2661, %r2659, %r2660; xor.b32 %r2662, %r2661, %r2655; shf.l.wrap.b32 %r2663, %r2661, %r2661, 14; sub.s32 %r2664, %r2662, %r2663; xor.b32 %r2665, %r2664, %r2658; shf.l.wrap.b32 %r2666, %r2664, %r2664, 24; sub.s32 %r2667, %r2665, %r2666; and.b32 %r2668, %r2667, 31; setp.lt.u32 %p395, %r2668, 24; selp.f32 %f1721, %f1606, %f1619, %p395; setp.lt.u32 %p396, %r2668, 16; selp.f32 %f1722, %f1619, %f1572, %p396; setp.lt.u32 %p397, %r2668, 8; selp.f32 %f1723, %f1572, %f1687, %p397; and.b32 %r2669, %r2667, 1; setp.eq.b32 %p398, %r2669, 1; neg.ftz.f32 %f1724, %f1721; selp.f32 %f1725, %f1724, %f1721, %p398; and.b32 %r2670, %r2667, 2; setp.eq.s32 %p399, %r2670, 0; neg.ftz.f32 %f1726, %f1722; selp.f32 %f1727, %f1722, %f1726, %p399; add.ftz.f32 %f1728, %f1725, %f1727; and.b32 %r2671, %r2667, 4; setp.eq.s32 %p400, %r2671, 0; neg.ftz.f32 %f1729, %f1723; selp.f32 %f1730, %f1723, %f1729, %p400; add.ftz.f32 %f1731, %f1730, %f1728; add.s32 %r2672, %r2414, %r2583; xor.b32 %r2673, %r2422, %r2672; sub.s32 %r2674, %r2673, %r2424; xor.b32 %r2675, %r2674, %r2417; shf.l.wrap.b32 %r2676, %r2674, %r2674, 25; sub.s32 %r2677, %r2675, %r2676; xor.b32 %r2678, %r2677, %r2422; shf.l.wrap.b32 %r2679, %r2677, %r2677, 16; sub.s32 %r2680, %r2678, %r2679; xor.b32 %r2681, %r2680, %r2674; shf.l.wrap.b32 %r2682, %r2680, %r2680, 4; sub.s32 %r2683, %r2681, %r2682; xor.b32 %r2684, %r2683, %r2677; shf.l.wrap.b32 %r2685, %r2683, %r2683, 14; sub.s32 %r2686, %r2684, %r2685; xor.b32 %r2687, %r2686, %r2680; shf.l.wrap.b32 %r2688, %r2686, %r2686, 24; sub.s32 %r2689, %r2687, %r2688; and.b32 %r2690, %r2689, 31; setp.lt.u32 %p401, %r2690, 24; selp.f32 %f1732, %f1568, %f1570, %p401; setp.lt.u32 %p402, %r2690, 16; selp.f32 %f1733, %f1570, %f1642, %p402; setp.lt.u32 %p403, %r2690, 8; selp.f32 %f1734, %f1642, %f1687, %p403; and.b32 %r2691, %r2689, 1; setp.eq.b32 %p404, %r2691, 1; neg.ftz.f32 %f1735, %f1732; selp.f32 %f1736, %f1735, %f1732, %p404; and.b32 %r2692, %r2689, 2; setp.eq.s32 %p405, %r2692, 0; neg.ftz.f32 %f1737, %f1733; selp.f32 %f1738, %f1733, %f1737, %p405; add.ftz.f32 %f1739, %f1736, %f1738; and.b32 %r2693, %r2689, 4; setp.eq.s32 %p406, %r2693, 0; neg.ftz.f32 %f1740, %f1734; selp.f32 %f1741, %f1734, %f1740, %p406; add.ftz.f32 %f1742, %f1741, %f1739; add.s32 %r2694, %r2461, %r2583; xor.b32 %r2695, %r2469, %r2694; sub.s32 %r2696, %r2695, %r2471; xor.b32 %r2697, %r2696, %r2464; shf.l.wrap.b32 %r2698, %r2696, %r2696, 25; sub.s32 %r2699, %r2697, %r2698; xor.b32 %r2700, %r2699, %r2469; shf.l.wrap.b32 %r2701, %r2699, %r2699, 16; sub.s32 %r2702, %r2700, %r2701; xor.b32 %r2703, %r2702, %r2696; shf.l.wrap.b32 %r2704, %r2702, %r2702, 4; sub.s32 %r2705, %r2703, %r2704; xor.b32 %r2706, %r2705, %r2699; shf.l.wrap.b32 %r2707, %r2705, %r2705, 14; sub.s32 %r2708, %r2706, %r2707; xor.b32 %r2709, %r2708, %r2702; shf.l.wrap.b32 %r2710, %r2708, %r2708, 24; sub.s32 %r2711, %r2709, %r2710; and.b32 %r2712, %r2711, 31; setp.lt.u32 %p407, %r2712, 24; selp.f32 %f1743, %f1606, %f1570, %p407; setp.lt.u32 %p408, %r2712, 16; selp.f32 %f1744, %f1570, %f1642, %p408; setp.lt.u32 %p409, %r2712, 8; selp.f32 %f1745, %f1642, %f1687, %p409; and.b32 %r2713, %r2711, 1; setp.eq.b32 %p410, %r2713, 1; neg.ftz.f32 %f1746, %f1743; selp.f32 %f1747, %f1746, %f1743, %p410; and.b32 %r2714, %r2711, 2; setp.eq.s32 %p411, %r2714, 0; neg.ftz.f32 %f1748, %f1744; selp.f32 %f1749, %f1744, %f1748, %p411; add.ftz.f32 %f1750, %f1747, %f1749; and.b32 %r2715, %r2711, 4; setp.eq.s32 %p412, %r2715, 0; neg.ftz.f32 %f1751, %f1745; selp.f32 %f1752, %f1745, %f1751, %p412; add.ftz.f32 %f1753, %f1752, %f1750; add.s32 %r2716, %r2507, %r2583; xor.b32 %r2717, %r2515, %r2716; sub.s32 %r2718, %r2717, %r2517; xor.b32 %r2719, %r2718, %r2510; shf.l.wrap.b32 %r2720, %r2718, %r2718, 25; sub.s32 %r2721, %r2719, %r2720; xor.b32 %r2722, %r2721, %r2515; shf.l.wrap.b32 %r2723, %r2721, %r2721, 16; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2718; shf.l.wrap.b32 %r2726, %r2724, %r2724, 4; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 14; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 24; sub.s32 %r2733, %r2731, %r2732; and.b32 %r2734, %r2733, 31; setp.lt.u32 %p413, %r2734, 24; selp.f32 %f1754, %f1568, %f1619, %p413; setp.lt.u32 %p414, %r2734, 16; selp.f32 %f1755, %f1619, %f1642, %p414; setp.lt.u32 %p415, %r2734, 8; selp.f32 %f1756, %f1642, %f1687, %p415; and.b32 %r2735, %r2733, 1; setp.eq.b32 %p416, %r2735, 1; neg.ftz.f32 %f1757, %f1754; selp.f32 %f1758, %f1757, %f1754, %p416; and.b32 %r2736, %r2733, 2; setp.eq.s32 %p417, %r2736, 0; neg.ftz.f32 %f1759, %f1755; selp.f32 %f1760, %f1755, %f1759, %p417; add.ftz.f32 %f1761, %f1758, %f1760; and.b32 %r2737, %r2733, 4; setp.eq.s32 %p418, %r2737, 0; neg.ftz.f32 %f1762, %f1756; selp.f32 %f1763, %f1756, %f1762, %p418; add.ftz.f32 %f1764, %f1763, %f1761; add.s32 %r2738, %r2552, %r2583; xor.b32 %r2739, %r2560, %r2738; sub.s32 %r2740, %r2739, %r2562; xor.b32 %r2741, %r2740, %r2555; shf.l.wrap.b32 %r2742, %r2740, %r2740, 25; sub.s32 %r2743, %r2741, %r2742; xor.b32 %r2744, %r2743, %r2560; shf.l.wrap.b32 %r2745, %r2743, %r2743, 16; sub.s32 %r2746, %r2744, %r2745; xor.b32 %r2747, %r2746, %r2740; shf.l.wrap.b32 %r2748, %r2746, %r2746, 4; sub.s32 %r2749, %r2747, %r2748; xor.b32 %r2750, %r2749, %r2743; shf.l.wrap.b32 %r2751, %r2749, %r2749, 14; sub.s32 %r2752, %r2750, %r2751; xor.b32 %r2753, %r2752, %r2746; shf.l.wrap.b32 %r2754, %r2752, %r2752, 24; sub.s32 %r2755, %r2753, %r2754; and.b32 %r2756, %r2755, 31; setp.lt.u32 %p419, %r2756, 24; selp.f32 %f1765, %f1606, %f1619, %p419; setp.lt.u32 %p420, %r2756, 16; selp.f32 %f1766, %f1619, %f1642, %p420; setp.lt.u32 %p421, %r2756, 8; selp.f32 %f1767, %f1642, %f1687, %p421; and.b32 %r2757, %r2755, 1; setp.eq.b32 %p422, %r2757, 1; neg.ftz.f32 %f1768, %f1765; selp.f32 %f1769, %f1768, %f1765, %p422; and.b32 %r2758, %r2755, 2; setp.eq.s32 %p423, %r2758, 0; neg.ftz.f32 %f1770, %f1766; selp.f32 %f1771, %f1766, %f1770, %p423; add.ftz.f32 %f1772, %f1769, %f1771; and.b32 %r2759, %r2755, 4; setp.eq.s32 %p424, %r2759, 0; neg.ftz.f32 %f1773, %f1767; selp.f32 %f1774, %f1767, %f1773, %p424; add.ftz.f32 %f1775, %f1774, %f1772; sub.ftz.f32 %f1776, %f1607, %f1579; sub.ftz.f32 %f1777, %f1607, %f1584; sub.ftz.f32 %f1778, %f1607, %f1589; mul.ftz.f32 %f1779, %f1776, %f1605; fma.rn.ftz.f32 %f1780, %f1579, %f1618, %f1779; mul.ftz.f32 %f1781, %f1776, %f1630; fma.rn.ftz.f32 %f1782, %f1579, %f1641, %f1781; mul.ftz.f32 %f1783, %f1584, %f1782; fma.rn.ftz.f32 %f1784, %f1777, %f1780, %f1783; mul.ftz.f32 %f1785, %f1776, %f1653; fma.rn.ftz.f32 %f1786, %f1579, %f1664, %f1785; mul.ftz.f32 %f1787, %f1776, %f1675; fma.rn.ftz.f32 %f1788, %f1579, %f1686, %f1787; mul.ftz.f32 %f1789, %f1584, %f1788; fma.rn.ftz.f32 %f1790, %f1777, %f1786, %f1789; mul.ftz.f32 %f1791, %f1589, %f1790; fma.rn.ftz.f32 %f1792, %f1778, %f1784, %f1791; mul.ftz.f32 %f1793, %f1776, %f1698; fma.rn.ftz.f32 %f1794, %f1579, %f1709, %f1793; mul.ftz.f32 %f1795, %f1776, %f1720; fma.rn.ftz.f32 %f1796, %f1579, %f1731, %f1795; mul.ftz.f32 %f1797, %f1584, %f1796; fma.rn.ftz.f32 %f1798, %f1777, %f1794, %f1797; mul.ftz.f32 %f1799, %f1776, %f1742; fma.rn.ftz.f32 %f1800, %f1579, %f1753, %f1799; mul.ftz.f32 %f1801, %f1776, %f1764; fma.rn.ftz.f32 %f1802, %f1579, %f1775, %f1801; mul.ftz.f32 %f1803, %f1584, %f1802; fma.rn.ftz.f32 %f1804, %f1777, %f1800, %f1803; mul.ftz.f32 %f1805, %f1589, %f1804; fma.rn.ftz.f32 %f1806, %f1778, %f1798, %f1805; sub.ftz.f32 %f1807, %f1806, %f1792; fma.rn.ftz.f32 %f68, %f1594, %f1807, %f1792; mov.b32 %r22, %f68; setp.nan.ftz.f32 %p425, %f68, %f68; mov.f32 %f5536, 0f00000000; @%p425 bra $L__BB74_74; or.b32 %r2760, %r22, -2147483648; setp.ne.s32 %p426, %r2760, -2147483648; add.ftz.f32 %f1809, %f68, %f68; setp.eq.ftz.f32 %p427, %f1809, %f68; and.pred %p428, %p427, %p426; @%p428 bra $L__BB74_74; shl.b32 %r2761, %r22, 1; setp.lt.u32 %p429, %r2761, -16777215; mul.ftz.f32 %f1810, %f68, 0f3F559B3D; selp.f32 %f5536, %f1810, 0f00000000, %p429; $L__BB74_74: abs.ftz.f32 %f1811, %f5536; sub.ftz.f32 %f1812, %f21, %f1811; mul.ftz.f32 %f5827, %f1812, %f1812; cvt.rzi.ftz.s32.f32 %r23, %f26; setp.lt.s32 %p430, %r23, 2; @%p430 bra $L__BB74_450; mov.u32 %r10019, 1; mov.f32 %f5538, %f5827; mov.f32 %f5539, %f67; $L__BB74_76: mul.ftz.f32 %f5568, %f27, %f5568; cvt.rmi.ftz.f32.f32 %f1814, %f5568; cvt.rzi.ftz.s32.f32 %r2763, %f1814; sub.ftz.f32 %f1815, %f5568, %f1814; mul.ftz.f32 %f5569, %f27, %f5569; cvt.rmi.ftz.f32.f32 %f1816, %f5569; cvt.rzi.ftz.s32.f32 %r2764, %f1816; sub.ftz.f32 %f1817, %f5569, %f1816; mul.ftz.f32 %f5570, %f27, %f5570; cvt.rmi.ftz.f32.f32 %f1818, %f5570; cvt.rzi.ftz.s32.f32 %r2765, %f1818; sub.ftz.f32 %f1819, %f5570, %f1818; mul.ftz.f32 %f5571, %f27, %f5571; cvt.rmi.ftz.f32.f32 %f1820, %f5571; cvt.rzi.ftz.s32.f32 %r2766, %f1820; sub.ftz.f32 %f1821, %f5571, %f1820; mul.ftz.f32 %f1822, %f1815, %f1815; mul.ftz.f32 %f1823, %f1815, %f1822; fma.rn.ftz.f32 %f1824, %f1815, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1825, %f1815, %f1824, 0f41200000; mul.ftz.f32 %f1826, %f1823, %f1825; mul.ftz.f32 %f1827, %f1817, %f1817; mul.ftz.f32 %f1828, %f1817, %f1827; fma.rn.ftz.f32 %f1829, %f1817, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1830, %f1817, %f1829, 0f41200000; mul.ftz.f32 %f1831, %f1828, %f1830; mul.ftz.f32 %f1832, %f1819, %f1819; mul.ftz.f32 %f1833, %f1819, %f1832; fma.rn.ftz.f32 %f1834, %f1819, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1835, %f1819, %f1834, 0f41200000; mul.ftz.f32 %f1836, %f1833, %f1835; mul.ftz.f32 %f1837, %f1821, %f1821; mul.ftz.f32 %f1838, %f1821, %f1837; fma.rn.ftz.f32 %f1839, %f1821, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1840, %f1821, %f1839, 0f41200000; mul.ftz.f32 %f1841, %f1838, %f1840; add.s32 %r2767, %r2764, -559038708; add.s32 %r2768, %r2765, -559038708; sub.s32 %r2769, %r2763, %r2765; shf.l.wrap.b32 %r2770, %r2768, %r2768, 4; xor.b32 %r2771, %r2770, %r2769; add.s32 %r2772, %r2768, %r2767; sub.s32 %r2773, %r2767, %r2771; shf.l.wrap.b32 %r2774, %r2771, %r2771, 6; xor.b32 %r2775, %r2774, %r2773; add.s32 %r2776, %r2771, %r2772; sub.s32 %r2777, %r2772, %r2775; shf.l.wrap.b32 %r2778, %r2775, %r2775, 8; xor.b32 %r2779, %r2778, %r2777; add.s32 %r2780, %r2775, %r2776; sub.s32 %r2781, %r2776, %r2779; shf.l.wrap.b32 %r2782, %r2779, %r2779, 16; xor.b32 %r2783, %r2782, %r2781; add.s32 %r2784, %r2779, %r2780; sub.s32 %r2785, %r2780, %r2783; shf.l.wrap.b32 %r2786, %r2783, %r2783, 19; xor.b32 %r2787, %r2786, %r2785; add.s32 %r2788, %r2783, %r2784; sub.s32 %r2789, %r2784, %r2787; shf.l.wrap.b32 %r2790, %r2787, %r2787, 4; add.s32 %r2791, %r2787, %r2788; add.s32 %r2792, %r2788, %r2766; xor.b32 %r2793, %r2791, %r2789; xor.b32 %r2794, %r2793, %r2790; shf.l.wrap.b32 %r2795, %r2791, %r2791, 14; sub.s32 %r2796, %r2794, %r2795; xor.b32 %r2797, %r2796, %r2792; shf.l.wrap.b32 %r2798, %r2796, %r2796, 11; sub.s32 %r2799, %r2797, %r2798; xor.b32 %r2800, %r2799, %r2791; shf.l.wrap.b32 %r2801, %r2799, %r2799, 25; sub.s32 %r2802, %r2800, %r2801; xor.b32 %r2803, %r2802, %r2796; shf.l.wrap.b32 %r2804, %r2802, %r2802, 16; sub.s32 %r2805, %r2803, %r2804; xor.b32 %r2806, %r2805, %r2799; shf.l.wrap.b32 %r2807, %r2805, %r2805, 4; sub.s32 %r2808, %r2806, %r2807; xor.b32 %r2809, %r2808, %r2802; shf.l.wrap.b32 %r2810, %r2808, %r2808, 14; sub.s32 %r2811, %r2809, %r2810; xor.b32 %r2812, %r2811, %r2805; shf.l.wrap.b32 %r2813, %r2811, %r2811, 24; sub.s32 %r2814, %r2812, %r2813; and.b32 %r2815, %r2814, 31; setp.lt.u32 %p431, %r2815, 24; selp.f32 %f1842, %f1815, %f1817, %p431; setp.lt.u32 %p432, %r2815, 16; selp.f32 %f1843, %f1817, %f1819, %p432; setp.lt.u32 %p433, %r2815, 8; selp.f32 %f1844, %f1819, %f1821, %p433; and.b32 %r2816, %r2814, 1; setp.eq.b32 %p434, %r2816, 1; neg.ftz.f32 %f1845, %f1842; selp.f32 %f1846, %f1845, %f1842, %p434; and.b32 %r2817, %r2814, 2; setp.eq.s32 %p435, %r2817, 0; neg.ftz.f32 %f1847, %f1843; selp.f32 %f1848, %f1843, %f1847, %p435; add.ftz.f32 %f1849, %f1846, %f1848; and.b32 %r2818, %r2814, 4; setp.eq.s32 %p436, %r2818, 0; neg.ftz.f32 %f1850, %f1844; selp.f32 %f1851, %f1844, %f1850, %p436; add.ftz.f32 %f1852, %f1851, %f1849; add.s32 %r2819, %r2763, 1; sub.s32 %r2820, %r2819, %r2765; xor.b32 %r2821, %r2770, %r2820; sub.s32 %r2822, %r2767, %r2821; shf.l.wrap.b32 %r2823, %r2821, %r2821, 6; xor.b32 %r2824, %r2823, %r2822; add.s32 %r2825, %r2821, %r2772; sub.s32 %r2826, %r2772, %r2824; shf.l.wrap.b32 %r2827, %r2824, %r2824, 8; xor.b32 %r2828, %r2827, %r2826; add.s32 %r2829, %r2824, %r2825; sub.s32 %r2830, %r2825, %r2828; shf.l.wrap.b32 %r2831, %r2828, %r2828, 16; xor.b32 %r2832, %r2831, %r2830; add.s32 %r2833, %r2828, %r2829; sub.s32 %r2834, %r2829, %r2832; shf.l.wrap.b32 %r2835, %r2832, %r2832, 19; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r2832, %r2833; sub.s32 %r2838, %r2833, %r2836; shf.l.wrap.b32 %r2839, %r2836, %r2836, 4; add.s32 %r2840, %r2836, %r2837; add.s32 %r2841, %r2837, %r2766; xor.b32 %r2842, %r2840, %r2838; xor.b32 %r2843, %r2842, %r2839; shf.l.wrap.b32 %r2844, %r2840, %r2840, 14; sub.s32 %r2845, %r2843, %r2844; xor.b32 %r2846, %r2845, %r2841; shf.l.wrap.b32 %r2847, %r2845, %r2845, 11; sub.s32 %r2848, %r2846, %r2847; xor.b32 %r2849, %r2848, %r2840; shf.l.wrap.b32 %r2850, %r2848, %r2848, 25; sub.s32 %r2851, %r2849, %r2850; xor.b32 %r2852, %r2851, %r2845; shf.l.wrap.b32 %r2853, %r2851, %r2851, 16; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2848; shf.l.wrap.b32 %r2856, %r2854, %r2854, 4; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2851; shf.l.wrap.b32 %r2859, %r2857, %r2857, 14; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 24; sub.s32 %r2863, %r2861, %r2862; add.ftz.f32 %f1853, %f1815, 0fBF800000; and.b32 %r2864, %r2863, 31; setp.lt.u32 %p437, %r2864, 24; selp.f32 %f1855, %f1853, %f1817, %p437; setp.lt.u32 %p438, %r2864, 16; selp.f32 %f1856, %f1817, %f1819, %p438; setp.lt.u32 %p439, %r2864, 8; selp.f32 %f1857, %f1819, %f1821, %p439; and.b32 %r2865, %r2863, 1; setp.eq.b32 %p440, %r2865, 1; neg.ftz.f32 %f1858, %f1855; selp.f32 %f1859, %f1858, %f1855, %p440; and.b32 %r2866, %r2863, 2; setp.eq.s32 %p441, %r2866, 0; neg.ftz.f32 %f1860, %f1856; selp.f32 %f1861, %f1856, %f1860, %p441; add.ftz.f32 %f1862, %f1859, %f1861; and.b32 %r2867, %r2863, 4; setp.eq.s32 %p442, %r2867, 0; neg.ftz.f32 %f1863, %f1857; selp.f32 %f1864, %f1857, %f1863, %p442; add.ftz.f32 %f1865, %f1864, %f1862; add.s32 %r2868, %r2764, -559038707; add.s32 %r2869, %r2768, %r2868; sub.s32 %r2870, %r2868, %r2771; xor.b32 %r2871, %r2774, %r2870; add.s32 %r2872, %r2771, %r2869; sub.s32 %r2873, %r2869, %r2871; shf.l.wrap.b32 %r2874, %r2871, %r2871, 8; xor.b32 %r2875, %r2874, %r2873; add.s32 %r2876, %r2871, %r2872; sub.s32 %r2877, %r2872, %r2875; shf.l.wrap.b32 %r2878, %r2875, %r2875, 16; xor.b32 %r2879, %r2878, %r2877; add.s32 %r2880, %r2875, %r2876; sub.s32 %r2881, %r2876, %r2879; shf.l.wrap.b32 %r2882, %r2879, %r2879, 19; xor.b32 %r2883, %r2882, %r2881; add.s32 %r2884, %r2879, %r2880; sub.s32 %r2885, %r2880, %r2883; shf.l.wrap.b32 %r2886, %r2883, %r2883, 4; add.s32 %r2887, %r2883, %r2884; add.s32 %r2888, %r2884, %r2766; xor.b32 %r2889, %r2887, %r2885; xor.b32 %r2890, %r2889, %r2886; shf.l.wrap.b32 %r2891, %r2887, %r2887, 14; sub.s32 %r2892, %r2890, %r2891; xor.b32 %r2893, %r2892, %r2888; shf.l.wrap.b32 %r2894, %r2892, %r2892, 11; sub.s32 %r2895, %r2893, %r2894; xor.b32 %r2896, %r2895, %r2887; shf.l.wrap.b32 %r2897, %r2895, %r2895, 25; sub.s32 %r2898, %r2896, %r2897; xor.b32 %r2899, %r2898, %r2892; shf.l.wrap.b32 %r2900, %r2898, %r2898, 16; sub.s32 %r2901, %r2899, %r2900; xor.b32 %r2902, %r2901, %r2895; shf.l.wrap.b32 %r2903, %r2901, %r2901, 4; sub.s32 %r2904, %r2902, %r2903; xor.b32 %r2905, %r2904, %r2898; shf.l.wrap.b32 %r2906, %r2904, %r2904, 14; sub.s32 %r2907, %r2905, %r2906; xor.b32 %r2908, %r2907, %r2901; shf.l.wrap.b32 %r2909, %r2907, %r2907, 24; sub.s32 %r2910, %r2908, %r2909; add.ftz.f32 %f1866, %f1817, 0fBF800000; and.b32 %r2911, %r2910, 31; setp.lt.u32 %p443, %r2911, 24; selp.f32 %f1867, %f1815, %f1866, %p443; setp.lt.u32 %p444, %r2911, 16; selp.f32 %f1868, %f1866, %f1819, %p444; setp.lt.u32 %p445, %r2911, 8; selp.f32 %f1869, %f1819, %f1821, %p445; and.b32 %r2912, %r2910, 1; setp.eq.b32 %p446, %r2912, 1; neg.ftz.f32 %f1870, %f1867; selp.f32 %f1871, %f1870, %f1867, %p446; and.b32 %r2913, %r2910, 2; setp.eq.s32 %p447, %r2913, 0; neg.ftz.f32 %f1872, %f1868; selp.f32 %f1873, %f1868, %f1872, %p447; add.ftz.f32 %f1874, %f1871, %f1873; and.b32 %r2914, %r2910, 4; setp.eq.s32 %p448, %r2914, 0; neg.ftz.f32 %f1875, %f1869; selp.f32 %f1876, %f1869, %f1875, %p448; add.ftz.f32 %f1877, %f1876, %f1874; sub.s32 %r2915, %r2868, %r2821; xor.b32 %r2916, %r2823, %r2915; add.s32 %r2917, %r2821, %r2869; sub.s32 %r2918, %r2869, %r2916; shf.l.wrap.b32 %r2919, %r2916, %r2916, 8; xor.b32 %r2920, %r2919, %r2918; add.s32 %r2921, %r2916, %r2917; sub.s32 %r2922, %r2917, %r2920; shf.l.wrap.b32 %r2923, %r2920, %r2920, 16; xor.b32 %r2924, %r2923, %r2922; add.s32 %r2925, %r2920, %r2921; sub.s32 %r2926, %r2921, %r2924; shf.l.wrap.b32 %r2927, %r2924, %r2924, 19; xor.b32 %r2928, %r2927, %r2926; add.s32 %r2929, %r2924, %r2925; sub.s32 %r2930, %r2925, %r2928; shf.l.wrap.b32 %r2931, %r2928, %r2928, 4; add.s32 %r2932, %r2928, %r2929; add.s32 %r2933, %r2929, %r2766; xor.b32 %r2934, %r2932, %r2930; xor.b32 %r2935, %r2934, %r2931; shf.l.wrap.b32 %r2936, %r2932, %r2932, 14; sub.s32 %r2937, %r2935, %r2936; xor.b32 %r2938, %r2937, %r2933; shf.l.wrap.b32 %r2939, %r2937, %r2937, 11; sub.s32 %r2940, %r2938, %r2939; xor.b32 %r2941, %r2940, %r2932; shf.l.wrap.b32 %r2942, %r2940, %r2940, 25; sub.s32 %r2943, %r2941, %r2942; xor.b32 %r2944, %r2943, %r2937; shf.l.wrap.b32 %r2945, %r2943, %r2943, 16; sub.s32 %r2946, %r2944, %r2945; xor.b32 %r2947, %r2946, %r2940; shf.l.wrap.b32 %r2948, %r2946, %r2946, 4; sub.s32 %r2949, %r2947, %r2948; xor.b32 %r2950, %r2949, %r2943; shf.l.wrap.b32 %r2951, %r2949, %r2949, 14; sub.s32 %r2952, %r2950, %r2951; xor.b32 %r2953, %r2952, %r2946; shf.l.wrap.b32 %r2954, %r2952, %r2952, 24; sub.s32 %r2955, %r2953, %r2954; and.b32 %r2956, %r2955, 31; setp.lt.u32 %p449, %r2956, 24; selp.f32 %f1878, %f1853, %f1866, %p449; setp.lt.u32 %p450, %r2956, 16; selp.f32 %f1879, %f1866, %f1819, %p450; setp.lt.u32 %p451, %r2956, 8; selp.f32 %f1880, %f1819, %f1821, %p451; and.b32 %r2957, %r2955, 1; setp.eq.b32 %p452, %r2957, 1; neg.ftz.f32 %f1881, %f1878; selp.f32 %f1882, %f1881, %f1878, %p452; and.b32 %r2958, %r2955, 2; setp.eq.s32 %p453, %r2958, 0; neg.ftz.f32 %f1883, %f1879; selp.f32 %f1884, %f1879, %f1883, %p453; add.ftz.f32 %f1885, %f1882, %f1884; and.b32 %r2959, %r2955, 4; setp.eq.s32 %p454, %r2959, 0; neg.ftz.f32 %f1886, %f1880; selp.f32 %f1887, %f1880, %f1886, %p454; add.ftz.f32 %f1888, %f1887, %f1885; not.b32 %r2960, %r2765; add.s32 %r2961, %r2765, -559038707; add.s32 %r2962, %r2763, %r2960; shf.l.wrap.b32 %r2963, %r2961, %r2961, 4; xor.b32 %r2964, %r2963, %r2962; add.s32 %r2965, %r2961, %r2767; sub.s32 %r2966, %r2767, %r2964; shf.l.wrap.b32 %r2967, %r2964, %r2964, 6; xor.b32 %r2968, %r2967, %r2966; add.s32 %r2969, %r2964, %r2965; sub.s32 %r2970, %r2965, %r2968; shf.l.wrap.b32 %r2971, %r2968, %r2968, 8; xor.b32 %r2972, %r2971, %r2970; add.s32 %r2973, %r2968, %r2969; sub.s32 %r2974, %r2969, %r2972; shf.l.wrap.b32 %r2975, %r2972, %r2972, 16; xor.b32 %r2976, %r2975, %r2974; add.s32 %r2977, %r2972, %r2973; sub.s32 %r2978, %r2973, %r2976; shf.l.wrap.b32 %r2979, %r2976, %r2976, 19; xor.b32 %r2980, %r2979, %r2978; add.s32 %r2981, %r2976, %r2977; sub.s32 %r2982, %r2977, %r2980; shf.l.wrap.b32 %r2983, %r2980, %r2980, 4; add.s32 %r2984, %r2980, %r2981; add.s32 %r2985, %r2981, %r2766; xor.b32 %r2986, %r2984, %r2982; xor.b32 %r2987, %r2986, %r2983; shf.l.wrap.b32 %r2988, %r2984, %r2984, 14; sub.s32 %r2989, %r2987, %r2988; xor.b32 %r2990, %r2989, %r2985; shf.l.wrap.b32 %r2991, %r2989, %r2989, 11; sub.s32 %r2992, %r2990, %r2991; xor.b32 %r2993, %r2992, %r2984; shf.l.wrap.b32 %r2994, %r2992, %r2992, 25; sub.s32 %r2995, %r2993, %r2994; xor.b32 %r2996, %r2995, %r2989; shf.l.wrap.b32 %r2997, %r2995, %r2995, 16; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r2992; shf.l.wrap.b32 %r3000, %r2998, %r2998, 4; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 14; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 24; sub.s32 %r3007, %r3005, %r3006; add.ftz.f32 %f1889, %f1819, 0fBF800000; and.b32 %r3008, %r3007, 31; setp.lt.u32 %p455, %r3008, 24; selp.f32 %f1890, %f1815, %f1817, %p455; setp.lt.u32 %p456, %r3008, 16; selp.f32 %f1891, %f1817, %f1889, %p456; setp.lt.u32 %p457, %r3008, 8; selp.f32 %f1892, %f1889, %f1821, %p457; and.b32 %r3009, %r3007, 1; setp.eq.b32 %p458, %r3009, 1; neg.ftz.f32 %f1893, %f1890; selp.f32 %f1894, %f1893, %f1890, %p458; and.b32 %r3010, %r3007, 2; setp.eq.s32 %p459, %r3010, 0; neg.ftz.f32 %f1895, %f1891; selp.f32 %f1896, %f1891, %f1895, %p459; add.ftz.f32 %f1897, %f1894, %f1896; and.b32 %r3011, %r3007, 4; setp.eq.s32 %p460, %r3011, 0; neg.ftz.f32 %f1898, %f1892; selp.f32 %f1899, %f1892, %f1898, %p460; add.ftz.f32 %f1900, %f1899, %f1897; xor.b32 %r3012, %r2963, %r2769; sub.s32 %r3013, %r2767, %r3012; shf.l.wrap.b32 %r3014, %r3012, %r3012, 6; xor.b32 %r3015, %r3014, %r3013; add.s32 %r3016, %r3012, %r2965; sub.s32 %r3017, %r2965, %r3015; shf.l.wrap.b32 %r3018, %r3015, %r3015, 8; xor.b32 %r3019, %r3018, %r3017; add.s32 %r3020, %r3015, %r3016; sub.s32 %r3021, %r3016, %r3019; shf.l.wrap.b32 %r3022, %r3019, %r3019, 16; xor.b32 %r3023, %r3022, %r3021; add.s32 %r3024, %r3019, %r3020; sub.s32 %r3025, %r3020, %r3023; shf.l.wrap.b32 %r3026, %r3023, %r3023, 19; xor.b32 %r3027, %r3026, %r3025; add.s32 %r3028, %r3023, %r3024; sub.s32 %r3029, %r3024, %r3027; shf.l.wrap.b32 %r3030, %r3027, %r3027, 4; add.s32 %r3031, %r3027, %r3028; add.s32 %r3032, %r3028, %r2766; xor.b32 %r3033, %r3031, %r3029; xor.b32 %r3034, %r3033, %r3030; shf.l.wrap.b32 %r3035, %r3031, %r3031, 14; sub.s32 %r3036, %r3034, %r3035; xor.b32 %r3037, %r3036, %r3032; shf.l.wrap.b32 %r3038, %r3036, %r3036, 11; sub.s32 %r3039, %r3037, %r3038; xor.b32 %r3040, %r3039, %r3031; shf.l.wrap.b32 %r3041, %r3039, %r3039, 25; sub.s32 %r3042, %r3040, %r3041; xor.b32 %r3043, %r3042, %r3036; shf.l.wrap.b32 %r3044, %r3042, %r3042, 16; sub.s32 %r3045, %r3043, %r3044; xor.b32 %r3046, %r3045, %r3039; shf.l.wrap.b32 %r3047, %r3045, %r3045, 4; sub.s32 %r3048, %r3046, %r3047; xor.b32 %r3049, %r3048, %r3042; shf.l.wrap.b32 %r3050, %r3048, %r3048, 14; sub.s32 %r3051, %r3049, %r3050; xor.b32 %r3052, %r3051, %r3045; shf.l.wrap.b32 %r3053, %r3051, %r3051, 24; sub.s32 %r3054, %r3052, %r3053; and.b32 %r3055, %r3054, 31; setp.lt.u32 %p461, %r3055, 24; selp.f32 %f1901, %f1853, %f1817, %p461; setp.lt.u32 %p462, %r3055, 16; selp.f32 %f1902, %f1817, %f1889, %p462; setp.lt.u32 %p463, %r3055, 8; selp.f32 %f1903, %f1889, %f1821, %p463; and.b32 %r3056, %r3054, 1; setp.eq.b32 %p464, %r3056, 1; neg.ftz.f32 %f1904, %f1901; selp.f32 %f1905, %f1904, %f1901, %p464; and.b32 %r3057, %r3054, 2; setp.eq.s32 %p465, %r3057, 0; neg.ftz.f32 %f1906, %f1902; selp.f32 %f1907, %f1902, %f1906, %p465; add.ftz.f32 %f1908, %f1905, %f1907; and.b32 %r3058, %r3054, 4; setp.eq.s32 %p466, %r3058, 0; neg.ftz.f32 %f1909, %f1903; selp.f32 %f1910, %f1903, %f1909, %p466; add.ftz.f32 %f1911, %f1910, %f1908; add.s32 %r3059, %r2961, %r2868; sub.s32 %r3060, %r2868, %r2964; xor.b32 %r3061, %r2967, %r3060; add.s32 %r3062, %r2964, %r3059; sub.s32 %r3063, %r3059, %r3061; shf.l.wrap.b32 %r3064, %r3061, %r3061, 8; xor.b32 %r3065, %r3064, %r3063; add.s32 %r3066, %r3061, %r3062; sub.s32 %r3067, %r3062, %r3065; shf.l.wrap.b32 %r3068, %r3065, %r3065, 16; xor.b32 %r3069, %r3068, %r3067; add.s32 %r3070, %r3065, %r3066; sub.s32 %r3071, %r3066, %r3069; shf.l.wrap.b32 %r3072, %r3069, %r3069, 19; xor.b32 %r3073, %r3072, %r3071; add.s32 %r3074, %r3069, %r3070; sub.s32 %r3075, %r3070, %r3073; shf.l.wrap.b32 %r3076, %r3073, %r3073, 4; add.s32 %r3077, %r3073, %r3074; add.s32 %r3078, %r3074, %r2766; xor.b32 %r3079, %r3077, %r3075; xor.b32 %r3080, %r3079, %r3076; shf.l.wrap.b32 %r3081, %r3077, %r3077, 14; sub.s32 %r3082, %r3080, %r3081; xor.b32 %r3083, %r3082, %r3078; shf.l.wrap.b32 %r3084, %r3082, %r3082, 11; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r3077; shf.l.wrap.b32 %r3087, %r3085, %r3085, 25; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3082; shf.l.wrap.b32 %r3090, %r3088, %r3088, 16; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 4; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 14; sub.s32 %r3097, %r3095, %r3096; xor.b32 %r3098, %r3097, %r3091; shf.l.wrap.b32 %r3099, %r3097, %r3097, 24; sub.s32 %r3100, %r3098, %r3099; and.b32 %r3101, %r3100, 31; setp.lt.u32 %p467, %r3101, 24; selp.f32 %f1912, %f1815, %f1866, %p467; setp.lt.u32 %p468, %r3101, 16; selp.f32 %f1913, %f1866, %f1889, %p468; setp.lt.u32 %p469, %r3101, 8; selp.f32 %f1914, %f1889, %f1821, %p469; and.b32 %r3102, %r3100, 1; setp.eq.b32 %p470, %r3102, 1; neg.ftz.f32 %f1915, %f1912; selp.f32 %f1916, %f1915, %f1912, %p470; and.b32 %r3103, %r3100, 2; setp.eq.s32 %p471, %r3103, 0; neg.ftz.f32 %f1917, %f1913; selp.f32 %f1918, %f1913, %f1917, %p471; add.ftz.f32 %f1919, %f1916, %f1918; and.b32 %r3104, %r3100, 4; setp.eq.s32 %p472, %r3104, 0; neg.ftz.f32 %f1920, %f1914; selp.f32 %f1921, %f1914, %f1920, %p472; add.ftz.f32 %f1922, %f1921, %f1919; sub.s32 %r3105, %r2868, %r3012; xor.b32 %r3106, %r3014, %r3105; add.s32 %r3107, %r3012, %r3059; sub.s32 %r3108, %r3059, %r3106; shf.l.wrap.b32 %r3109, %r3106, %r3106, 8; xor.b32 %r3110, %r3109, %r3108; add.s32 %r3111, %r3106, %r3107; sub.s32 %r3112, %r3107, %r3110; shf.l.wrap.b32 %r3113, %r3110, %r3110, 16; xor.b32 %r3114, %r3113, %r3112; add.s32 %r3115, %r3110, %r3111; sub.s32 %r3116, %r3111, %r3114; shf.l.wrap.b32 %r3117, %r3114, %r3114, 19; xor.b32 %r3118, %r3117, %r3116; add.s32 %r3119, %r3114, %r3115; sub.s32 %r3120, %r3115, %r3118; shf.l.wrap.b32 %r3121, %r3118, %r3118, 4; add.s32 %r3122, %r3118, %r3119; add.s32 %r3123, %r3119, %r2766; xor.b32 %r3124, %r3122, %r3120; xor.b32 %r3125, %r3124, %r3121; shf.l.wrap.b32 %r3126, %r3122, %r3122, 14; sub.s32 %r3127, %r3125, %r3126; xor.b32 %r3128, %r3127, %r3123; shf.l.wrap.b32 %r3129, %r3127, %r3127, 11; sub.s32 %r3130, %r3128, %r3129; xor.b32 %r3131, %r3130, %r3122; shf.l.wrap.b32 %r3132, %r3130, %r3130, 25; sub.s32 %r3133, %r3131, %r3132; xor.b32 %r3134, %r3133, %r3127; shf.l.wrap.b32 %r3135, %r3133, %r3133, 16; sub.s32 %r3136, %r3134, %r3135; xor.b32 %r3137, %r3136, %r3130; shf.l.wrap.b32 %r3138, %r3136, %r3136, 4; sub.s32 %r3139, %r3137, %r3138; xor.b32 %r3140, %r3139, %r3133; shf.l.wrap.b32 %r3141, %r3139, %r3139, 14; sub.s32 %r3142, %r3140, %r3141; xor.b32 %r3143, %r3142, %r3136; shf.l.wrap.b32 %r3144, %r3142, %r3142, 24; sub.s32 %r3145, %r3143, %r3144; and.b32 %r3146, %r3145, 31; setp.lt.u32 %p473, %r3146, 24; selp.f32 %f1923, %f1853, %f1866, %p473; setp.lt.u32 %p474, %r3146, 16; selp.f32 %f1924, %f1866, %f1889, %p474; setp.lt.u32 %p475, %r3146, 8; selp.f32 %f1925, %f1889, %f1821, %p475; and.b32 %r3147, %r3145, 1; setp.eq.b32 %p476, %r3147, 1; neg.ftz.f32 %f1926, %f1923; selp.f32 %f1927, %f1926, %f1923, %p476; and.b32 %r3148, %r3145, 2; setp.eq.s32 %p477, %r3148, 0; neg.ftz.f32 %f1928, %f1924; selp.f32 %f1929, %f1924, %f1928, %p477; add.ftz.f32 %f1930, %f1927, %f1929; and.b32 %r3149, %r3145, 4; setp.eq.s32 %p478, %r3149, 0; neg.ftz.f32 %f1931, %f1925; selp.f32 %f1932, %f1925, %f1931, %p478; add.ftz.f32 %f1933, %f1932, %f1930; add.s32 %r3150, %r2766, 1; add.s32 %r3151, %r2788, %r3150; xor.b32 %r3152, %r2796, %r3151; sub.s32 %r3153, %r3152, %r2798; xor.b32 %r3154, %r3153, %r2791; shf.l.wrap.b32 %r3155, %r3153, %r3153, 25; sub.s32 %r3156, %r3154, %r3155; xor.b32 %r3157, %r3156, %r2796; shf.l.wrap.b32 %r3158, %r3156, %r3156, 16; sub.s32 %r3159, %r3157, %r3158; xor.b32 %r3160, %r3159, %r3153; shf.l.wrap.b32 %r3161, %r3159, %r3159, 4; sub.s32 %r3162, %r3160, %r3161; xor.b32 %r3163, %r3162, %r3156; shf.l.wrap.b32 %r3164, %r3162, %r3162, 14; sub.s32 %r3165, %r3163, %r3164; xor.b32 %r3166, %r3165, %r3159; shf.l.wrap.b32 %r3167, %r3165, %r3165, 24; sub.s32 %r3168, %r3166, %r3167; add.ftz.f32 %f1934, %f1821, 0fBF800000; and.b32 %r3169, %r3168, 31; setp.lt.u32 %p479, %r3169, 24; selp.f32 %f1935, %f1815, %f1817, %p479; setp.lt.u32 %p480, %r3169, 16; selp.f32 %f1936, %f1817, %f1819, %p480; setp.lt.u32 %p481, %r3169, 8; selp.f32 %f1937, %f1819, %f1934, %p481; and.b32 %r3170, %r3168, 1; setp.eq.b32 %p482, %r3170, 1; neg.ftz.f32 %f1938, %f1935; selp.f32 %f1939, %f1938, %f1935, %p482; and.b32 %r3171, %r3168, 2; setp.eq.s32 %p483, %r3171, 0; neg.ftz.f32 %f1940, %f1936; selp.f32 %f1941, %f1936, %f1940, %p483; add.ftz.f32 %f1942, %f1939, %f1941; and.b32 %r3172, %r3168, 4; setp.eq.s32 %p484, %r3172, 0; neg.ftz.f32 %f1943, %f1937; selp.f32 %f1944, %f1937, %f1943, %p484; add.ftz.f32 %f1945, %f1944, %f1942; add.s32 %r3173, %r2837, %r3150; xor.b32 %r3174, %r2845, %r3173; sub.s32 %r3175, %r3174, %r2847; xor.b32 %r3176, %r3175, %r2840; shf.l.wrap.b32 %r3177, %r3175, %r3175, 25; sub.s32 %r3178, %r3176, %r3177; xor.b32 %r3179, %r3178, %r2845; shf.l.wrap.b32 %r3180, %r3178, %r3178, 16; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r3175; shf.l.wrap.b32 %r3183, %r3181, %r3181, 4; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 14; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 24; sub.s32 %r3190, %r3188, %r3189; and.b32 %r3191, %r3190, 31; setp.lt.u32 %p485, %r3191, 24; selp.f32 %f1946, %f1853, %f1817, %p485; setp.lt.u32 %p486, %r3191, 16; selp.f32 %f1947, %f1817, %f1819, %p486; setp.lt.u32 %p487, %r3191, 8; selp.f32 %f1948, %f1819, %f1934, %p487; and.b32 %r3192, %r3190, 1; setp.eq.b32 %p488, %r3192, 1; neg.ftz.f32 %f1949, %f1946; selp.f32 %f1950, %f1949, %f1946, %p488; and.b32 %r3193, %r3190, 2; setp.eq.s32 %p489, %r3193, 0; neg.ftz.f32 %f1951, %f1947; selp.f32 %f1952, %f1947, %f1951, %p489; add.ftz.f32 %f1953, %f1950, %f1952; and.b32 %r3194, %r3190, 4; setp.eq.s32 %p490, %r3194, 0; neg.ftz.f32 %f1954, %f1948; selp.f32 %f1955, %f1948, %f1954, %p490; add.ftz.f32 %f1956, %f1955, %f1953; add.s32 %r3195, %r2884, %r3150; xor.b32 %r3196, %r2892, %r3195; sub.s32 %r3197, %r3196, %r2894; xor.b32 %r3198, %r3197, %r2887; shf.l.wrap.b32 %r3199, %r3197, %r3197, 25; sub.s32 %r3200, %r3198, %r3199; xor.b32 %r3201, %r3200, %r2892; shf.l.wrap.b32 %r3202, %r3200, %r3200, 16; sub.s32 %r3203, %r3201, %r3202; xor.b32 %r3204, %r3203, %r3197; shf.l.wrap.b32 %r3205, %r3203, %r3203, 4; sub.s32 %r3206, %r3204, %r3205; xor.b32 %r3207, %r3206, %r3200; shf.l.wrap.b32 %r3208, %r3206, %r3206, 14; sub.s32 %r3209, %r3207, %r3208; xor.b32 %r3210, %r3209, %r3203; shf.l.wrap.b32 %r3211, %r3209, %r3209, 24; sub.s32 %r3212, %r3210, %r3211; and.b32 %r3213, %r3212, 31; setp.lt.u32 %p491, %r3213, 24; selp.f32 %f1957, %f1815, %f1866, %p491; setp.lt.u32 %p492, %r3213, 16; selp.f32 %f1958, %f1866, %f1819, %p492; setp.lt.u32 %p493, %r3213, 8; selp.f32 %f1959, %f1819, %f1934, %p493; and.b32 %r3214, %r3212, 1; setp.eq.b32 %p494, %r3214, 1; neg.ftz.f32 %f1960, %f1957; selp.f32 %f1961, %f1960, %f1957, %p494; and.b32 %r3215, %r3212, 2; setp.eq.s32 %p495, %r3215, 0; neg.ftz.f32 %f1962, %f1958; selp.f32 %f1963, %f1958, %f1962, %p495; add.ftz.f32 %f1964, %f1961, %f1963; and.b32 %r3216, %r3212, 4; setp.eq.s32 %p496, %r3216, 0; neg.ftz.f32 %f1965, %f1959; selp.f32 %f1966, %f1959, %f1965, %p496; add.ftz.f32 %f1967, %f1966, %f1964; add.s32 %r3217, %r2929, %r3150; xor.b32 %r3218, %r2937, %r3217; sub.s32 %r3219, %r3218, %r2939; xor.b32 %r3220, %r3219, %r2932; shf.l.wrap.b32 %r3221, %r3219, %r3219, 25; sub.s32 %r3222, %r3220, %r3221; xor.b32 %r3223, %r3222, %r2937; shf.l.wrap.b32 %r3224, %r3222, %r3222, 16; sub.s32 %r3225, %r3223, %r3224; xor.b32 %r3226, %r3225, %r3219; shf.l.wrap.b32 %r3227, %r3225, %r3225, 4; sub.s32 %r3228, %r3226, %r3227; xor.b32 %r3229, %r3228, %r3222; shf.l.wrap.b32 %r3230, %r3228, %r3228, 14; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3225; shf.l.wrap.b32 %r3233, %r3231, %r3231, 24; sub.s32 %r3234, %r3232, %r3233; and.b32 %r3235, %r3234, 31; setp.lt.u32 %p497, %r3235, 24; selp.f32 %f1968, %f1853, %f1866, %p497; setp.lt.u32 %p498, %r3235, 16; selp.f32 %f1969, %f1866, %f1819, %p498; setp.lt.u32 %p499, %r3235, 8; selp.f32 %f1970, %f1819, %f1934, %p499; and.b32 %r3236, %r3234, 1; setp.eq.b32 %p500, %r3236, 1; neg.ftz.f32 %f1971, %f1968; selp.f32 %f1972, %f1971, %f1968, %p500; and.b32 %r3237, %r3234, 2; setp.eq.s32 %p501, %r3237, 0; neg.ftz.f32 %f1973, %f1969; selp.f32 %f1974, %f1969, %f1973, %p501; add.ftz.f32 %f1975, %f1972, %f1974; and.b32 %r3238, %r3234, 4; setp.eq.s32 %p502, %r3238, 0; neg.ftz.f32 %f1976, %f1970; selp.f32 %f1977, %f1970, %f1976, %p502; add.ftz.f32 %f1978, %f1977, %f1975; add.s32 %r3239, %r2981, %r3150; xor.b32 %r3240, %r2989, %r3239; sub.s32 %r3241, %r3240, %r2991; xor.b32 %r3242, %r3241, %r2984; shf.l.wrap.b32 %r3243, %r3241, %r3241, 25; sub.s32 %r3244, %r3242, %r3243; xor.b32 %r3245, %r3244, %r2989; shf.l.wrap.b32 %r3246, %r3244, %r3244, 16; sub.s32 %r3247, %r3245, %r3246; xor.b32 %r3248, %r3247, %r3241; shf.l.wrap.b32 %r3249, %r3247, %r3247, 4; sub.s32 %r3250, %r3248, %r3249; xor.b32 %r3251, %r3250, %r3244; shf.l.wrap.b32 %r3252, %r3250, %r3250, 14; sub.s32 %r3253, %r3251, %r3252; xor.b32 %r3254, %r3253, %r3247; shf.l.wrap.b32 %r3255, %r3253, %r3253, 24; sub.s32 %r3256, %r3254, %r3255; and.b32 %r3257, %r3256, 31; setp.lt.u32 %p503, %r3257, 24; selp.f32 %f1979, %f1815, %f1817, %p503; setp.lt.u32 %p504, %r3257, 16; selp.f32 %f1980, %f1817, %f1889, %p504; setp.lt.u32 %p505, %r3257, 8; selp.f32 %f1981, %f1889, %f1934, %p505; and.b32 %r3258, %r3256, 1; setp.eq.b32 %p506, %r3258, 1; neg.ftz.f32 %f1982, %f1979; selp.f32 %f1983, %f1982, %f1979, %p506; and.b32 %r3259, %r3256, 2; setp.eq.s32 %p507, %r3259, 0; neg.ftz.f32 %f1984, %f1980; selp.f32 %f1985, %f1980, %f1984, %p507; add.ftz.f32 %f1986, %f1983, %f1985; and.b32 %r3260, %r3256, 4; setp.eq.s32 %p508, %r3260, 0; neg.ftz.f32 %f1987, %f1981; selp.f32 %f1988, %f1981, %f1987, %p508; add.ftz.f32 %f1989, %f1988, %f1986; add.s32 %r3261, %r3028, %r3150; xor.b32 %r3262, %r3036, %r3261; sub.s32 %r3263, %r3262, %r3038; xor.b32 %r3264, %r3263, %r3031; shf.l.wrap.b32 %r3265, %r3263, %r3263, 25; sub.s32 %r3266, %r3264, %r3265; xor.b32 %r3267, %r3266, %r3036; shf.l.wrap.b32 %r3268, %r3266, %r3266, 16; sub.s32 %r3269, %r3267, %r3268; xor.b32 %r3270, %r3269, %r3263; shf.l.wrap.b32 %r3271, %r3269, %r3269, 4; sub.s32 %r3272, %r3270, %r3271; xor.b32 %r3273, %r3272, %r3266; shf.l.wrap.b32 %r3274, %r3272, %r3272, 14; sub.s32 %r3275, %r3273, %r3274; xor.b32 %r3276, %r3275, %r3269; shf.l.wrap.b32 %r3277, %r3275, %r3275, 24; sub.s32 %r3278, %r3276, %r3277; and.b32 %r3279, %r3278, 31; setp.lt.u32 %p509, %r3279, 24; selp.f32 %f1990, %f1853, %f1817, %p509; setp.lt.u32 %p510, %r3279, 16; selp.f32 %f1991, %f1817, %f1889, %p510; setp.lt.u32 %p511, %r3279, 8; selp.f32 %f1992, %f1889, %f1934, %p511; and.b32 %r3280, %r3278, 1; setp.eq.b32 %p512, %r3280, 1; neg.ftz.f32 %f1993, %f1990; selp.f32 %f1994, %f1993, %f1990, %p512; and.b32 %r3281, %r3278, 2; setp.eq.s32 %p513, %r3281, 0; neg.ftz.f32 %f1995, %f1991; selp.f32 %f1996, %f1991, %f1995, %p513; add.ftz.f32 %f1997, %f1994, %f1996; and.b32 %r3282, %r3278, 4; setp.eq.s32 %p514, %r3282, 0; neg.ftz.f32 %f1998, %f1992; selp.f32 %f1999, %f1992, %f1998, %p514; add.ftz.f32 %f2000, %f1999, %f1997; add.s32 %r3283, %r3074, %r3150; xor.b32 %r3284, %r3082, %r3283; sub.s32 %r3285, %r3284, %r3084; xor.b32 %r3286, %r3285, %r3077; shf.l.wrap.b32 %r3287, %r3285, %r3285, 25; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3082; shf.l.wrap.b32 %r3290, %r3288, %r3288, 16; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 4; sub.s32 %r3294, %r3292, %r3293; xor.b32 %r3295, %r3294, %r3288; shf.l.wrap.b32 %r3296, %r3294, %r3294, 14; sub.s32 %r3297, %r3295, %r3296; xor.b32 %r3298, %r3297, %r3291; shf.l.wrap.b32 %r3299, %r3297, %r3297, 24; sub.s32 %r3300, %r3298, %r3299; and.b32 %r3301, %r3300, 31; setp.lt.u32 %p515, %r3301, 24; selp.f32 %f2001, %f1815, %f1866, %p515; setp.lt.u32 %p516, %r3301, 16; selp.f32 %f2002, %f1866, %f1889, %p516; setp.lt.u32 %p517, %r3301, 8; selp.f32 %f2003, %f1889, %f1934, %p517; and.b32 %r3302, %r3300, 1; setp.eq.b32 %p518, %r3302, 1; neg.ftz.f32 %f2004, %f2001; selp.f32 %f2005, %f2004, %f2001, %p518; and.b32 %r3303, %r3300, 2; setp.eq.s32 %p519, %r3303, 0; neg.ftz.f32 %f2006, %f2002; selp.f32 %f2007, %f2002, %f2006, %p519; add.ftz.f32 %f2008, %f2005, %f2007; and.b32 %r3304, %r3300, 4; setp.eq.s32 %p520, %r3304, 0; neg.ftz.f32 %f2009, %f2003; selp.f32 %f2010, %f2003, %f2009, %p520; add.ftz.f32 %f2011, %f2010, %f2008; add.s32 %r3305, %r3119, %r3150; xor.b32 %r3306, %r3127, %r3305; sub.s32 %r3307, %r3306, %r3129; xor.b32 %r3308, %r3307, %r3122; shf.l.wrap.b32 %r3309, %r3307, %r3307, 25; sub.s32 %r3310, %r3308, %r3309; xor.b32 %r3311, %r3310, %r3127; shf.l.wrap.b32 %r3312, %r3310, %r3310, 16; sub.s32 %r3313, %r3311, %r3312; xor.b32 %r3314, %r3313, %r3307; shf.l.wrap.b32 %r3315, %r3313, %r3313, 4; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3310; shf.l.wrap.b32 %r3318, %r3316, %r3316, 14; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 24; sub.s32 %r3322, %r3320, %r3321; and.b32 %r3323, %r3322, 31; setp.lt.u32 %p521, %r3323, 24; selp.f32 %f2012, %f1853, %f1866, %p521; setp.lt.u32 %p522, %r3323, 16; selp.f32 %f2013, %f1866, %f1889, %p522; setp.lt.u32 %p523, %r3323, 8; selp.f32 %f2014, %f1889, %f1934, %p523; and.b32 %r3324, %r3322, 1; setp.eq.b32 %p524, %r3324, 1; neg.ftz.f32 %f2015, %f2012; selp.f32 %f2016, %f2015, %f2012, %p524; and.b32 %r3325, %r3322, 2; setp.eq.s32 %p525, %r3325, 0; neg.ftz.f32 %f2017, %f2013; selp.f32 %f2018, %f2013, %f2017, %p525; add.ftz.f32 %f2019, %f2016, %f2018; and.b32 %r3326, %r3322, 4; setp.eq.s32 %p526, %r3326, 0; neg.ftz.f32 %f2020, %f2014; selp.f32 %f2021, %f2014, %f2020, %p526; add.ftz.f32 %f2022, %f2021, %f2019; sub.ftz.f32 %f2023, %f1607, %f1826; sub.ftz.f32 %f2024, %f1607, %f1831; sub.ftz.f32 %f2025, %f1607, %f1836; mul.ftz.f32 %f2026, %f2023, %f1852; fma.rn.ftz.f32 %f2027, %f1826, %f1865, %f2026; mul.ftz.f32 %f2028, %f2023, %f1877; fma.rn.ftz.f32 %f2029, %f1826, %f1888, %f2028; mul.ftz.f32 %f2030, %f1831, %f2029; fma.rn.ftz.f32 %f2031, %f2024, %f2027, %f2030; mul.ftz.f32 %f2032, %f2023, %f1900; fma.rn.ftz.f32 %f2033, %f1826, %f1911, %f2032; mul.ftz.f32 %f2034, %f2023, %f1922; fma.rn.ftz.f32 %f2035, %f1826, %f1933, %f2034; mul.ftz.f32 %f2036, %f1831, %f2035; fma.rn.ftz.f32 %f2037, %f2024, %f2033, %f2036; mul.ftz.f32 %f2038, %f1836, %f2037; fma.rn.ftz.f32 %f2039, %f2025, %f2031, %f2038; mul.ftz.f32 %f2040, %f2023, %f1945; fma.rn.ftz.f32 %f2041, %f1826, %f1956, %f2040; mul.ftz.f32 %f2042, %f2023, %f1967; fma.rn.ftz.f32 %f2043, %f1826, %f1978, %f2042; mul.ftz.f32 %f2044, %f1831, %f2043; fma.rn.ftz.f32 %f2045, %f2024, %f2041, %f2044; mul.ftz.f32 %f2046, %f2023, %f1989; fma.rn.ftz.f32 %f2047, %f1826, %f2000, %f2046; mul.ftz.f32 %f2048, %f2023, %f2011; fma.rn.ftz.f32 %f2049, %f1826, %f2022, %f2048; mul.ftz.f32 %f2050, %f1831, %f2049; fma.rn.ftz.f32 %f2051, %f2024, %f2047, %f2050; mul.ftz.f32 %f2052, %f1836, %f2051; fma.rn.ftz.f32 %f2053, %f2025, %f2045, %f2052; sub.ftz.f32 %f2054, %f2053, %f2039; fma.rn.ftz.f32 %f83, %f1841, %f2054, %f2039; mov.b32 %r25, %f83; setp.nan.ftz.f32 %p527, %f83, %f83; mov.f32 %f5544, 0f00000000; @%p527 bra $L__BB74_79; or.b32 %r3327, %r25, -2147483648; setp.ne.s32 %p528, %r3327, -2147483648; add.ftz.f32 %f2056, %f83, %f83; setp.eq.ftz.f32 %p529, %f2056, %f83; and.pred %p530, %p529, %p528; @%p530 bra $L__BB74_79; shl.b32 %r3328, %r25, 1; setp.lt.u32 %p531, %r3328, -16777215; selp.f32 %f5544, %f83, 0f00000000, %p531; $L__BB74_79: mul.ftz.f32 %f2057, %f5544, 0f3F559B3D; abs.ftz.f32 %f2058, %f2057; sub.ftz.f32 %f2059, %f21, %f2058; mul.ftz.f32 %f2060, %f2059, %f2059; mul.ftz.f32 %f2061, %f24, %f5538; cvt.ftz.sat.f32.f32 %f2062, %f2061; mul.ftz.f32 %f5538, %f2062, %f2060; fma.rn.ftz.f32 %f5827, %f5539, %f5538, %f5827; mul.ftz.f32 %f5539, %f67, %f5539; add.s32 %r10019, %r10019, 1; setp.lt.s32 %p532, %r10019, %r23; @%p532 bra $L__BB74_76; bra.uni $L__BB74_450; $L__BB74_428: lg2.approx.ftz.f32 %f5358, %f27; mul.ftz.f32 %f5359, %f25, %f5358; neg.ftz.f32 %f5360, %f5359; ex2.approx.ftz.f32 %f756, %f5360; cvt.rzi.ftz.s32.f32 %r469, %f26; mov.f32 %f5809, 0f3F800000; mov.f32 %f5827, 0f00000000; mov.u32 %r10035, 0; mov.pred %p1934, -1; mov.pred %p1943, 0; mov.f32 %f5807, %f5809; $L__BB74_429: setp.ge.s32 %p1935, %r10035, %r469; mov.pred %p1994, %p1934; @%p1935 bra $L__BB74_434; cvt.rmi.ftz.f32.f32 %f5362, %f5815; cvt.rzi.ftz.s32.f32 %r9742, %f5362; sub.ftz.f32 %f5363, %f5815, %f5362; mul.ftz.f32 %f5364, %f5363, %f5363; mul.ftz.f32 %f5365, %f5363, %f5364; fma.rn.ftz.f32 %f5366, %f5363, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5367, %f5363, %f5366, 0f41200000; mul.ftz.f32 %f5368, %f5365, %f5367; add.s32 %r9743, %r9742, -559038720; xor.b32 %r9744, %r9743, -1874868139; add.s32 %r9745, %r9744, 29184895; xor.b32 %r9746, %r9745, -559038720; shf.l.wrap.b32 %r9747, %r9745, %r9745, 25; sub.s32 %r9748, %r9746, %r9747; xor.b32 %r9749, %r9748, -1874868139; shf.l.wrap.b32 %r9750, %r9748, %r9748, 16; sub.s32 %r9751, %r9749, %r9750; xor.b32 %r9752, %r9751, %r9745; shf.l.wrap.b32 %r9753, %r9751, %r9751, 4; sub.s32 %r9754, %r9752, %r9753; xor.b32 %r9755, %r9754, %r9748; shf.l.wrap.b32 %r9756, %r9754, %r9754, 14; sub.s32 %r9757, %r9755, %r9756; xor.b32 %r9758, %r9757, %r9751; shf.l.wrap.b32 %r9759, %r9757, %r9757, 24; sub.s32 %r9760, %r9758, %r9759; and.b32 %r9761, %r9760, 7; add.s32 %r9762, %r9761, 1; cvt.rn.f32.s32 %f5369, %r9762; and.b32 %r9763, %r9760, 8; setp.eq.s32 %p1936, %r9763, 0; neg.ftz.f32 %f5370, %f5369; selp.f32 %f5371, %f5369, %f5370, %p1936; mul.ftz.f32 %f5372, %f5363, %f5371; add.s32 %r9764, %r9742, -559038719; xor.b32 %r9765, %r9764, -1874868139; add.s32 %r9766, %r9765, 29184895; xor.b32 %r9767, %r9766, -559038720; shf.l.wrap.b32 %r9768, %r9766, %r9766, 25; sub.s32 %r9769, %r9767, %r9768; xor.b32 %r9770, %r9769, -1874868139; shf.l.wrap.b32 %r9771, %r9769, %r9769, 16; sub.s32 %r9772, %r9770, %r9771; xor.b32 %r9773, %r9772, %r9766; shf.l.wrap.b32 %r9774, %r9772, %r9772, 4; sub.s32 %r9775, %r9773, %r9774; xor.b32 %r9776, %r9775, %r9769; shf.l.wrap.b32 %r9777, %r9775, %r9775, 14; sub.s32 %r9778, %r9776, %r9777; xor.b32 %r9779, %r9778, %r9772; shf.l.wrap.b32 %r9780, %r9778, %r9778, 24; sub.s32 %r9781, %r9779, %r9780; add.ftz.f32 %f5373, %f5363, 0fBF800000; and.b32 %r9782, %r9781, 7; add.s32 %r9783, %r9782, 1; cvt.rn.f32.s32 %f5374, %r9783; and.b32 %r9784, %r9781, 8; setp.eq.s32 %p1937, %r9784, 0; neg.ftz.f32 %f5375, %f5374; selp.f32 %f5376, %f5374, %f5375, %p1937; mul.ftz.f32 %f5377, %f5373, %f5376; sub.ftz.f32 %f5378, %f5377, %f5372; fma.rn.ftz.f32 %f761, %f5368, %f5378, %f5372; mov.b32 %r471, %f761; setp.nan.ftz.f32 %p1938, %f761, %f761; mov.f32 %f5805, 0f00000000; @%p1938 bra $L__BB74_433; or.b32 %r9785, %r471, -2147483648; setp.ne.s32 %p1939, %r9785, -2147483648; add.ftz.f32 %f5380, %f761, %f761; setp.eq.ftz.f32 %p1940, %f5380, %f761; and.pred %p1941, %p1940, %p1939; @%p1941 bra $L__BB74_433; shl.b32 %r9786, %r471, 1; setp.lt.u32 %p1942, %r9786, -16777215; selp.f32 %f5805, %f761, 0f00000000, %p1942; $L__BB74_433: setp.gt.ftz.f32 %p1944, %f5809, 0f3F800000; selp.f32 %f5381, 0f3F800000, %f5809, %p1944; fma.rn.ftz.f32 %f5382, %f5805, 0f3E800000, %f21; mul.ftz.f32 %f5383, %f5807, %f5382; mul.ftz.f32 %f5807, %f756, %f5807; fma.rn.ftz.f32 %f5827, %f5381, %f5383, %f5827; mul.ftz.f32 %f5384, %f24, %f5383; mul.ftz.f32 %f5809, %f5381, %f5384; mul.ftz.f32 %f5815, %f27, %f5815; add.s32 %r10035, %r10035, 1; setp.gt.ftz.f32 %p1945, %f5809, 0f3A83126F; mov.pred %p1994, %p1943; @%p1945 bra $L__BB74_429; $L__BB74_434: cvt.rmi.ftz.f32.f32 %f5385, %f26; sub.ftz.f32 %f772, %f26, %f5385; setp.eq.ftz.f32 %p1946, %f772, 0f00000000; not.pred %p1947, %p1994; or.pred %p1948, %p1947, %p1946; @%p1948 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f5387, %f5815; cvt.rzi.ftz.s32.f32 %r9787, %f5387; sub.ftz.f32 %f5388, %f5815, %f5387; mul.ftz.f32 %f5389, %f5388, %f5388; mul.ftz.f32 %f5390, %f5388, %f5389; fma.rn.ftz.f32 %f5391, %f5388, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5392, %f5388, %f5391, 0f41200000; mul.ftz.f32 %f5393, %f5390, %f5392; add.s32 %r9788, %r9787, -559038720; xor.b32 %r9789, %r9788, -1874868139; add.s32 %r9790, %r9789, 29184895; xor.b32 %r9791, %r9790, -559038720; shf.l.wrap.b32 %r9792, %r9790, %r9790, 25; sub.s32 %r9793, %r9791, %r9792; xor.b32 %r9794, %r9793, -1874868139; shf.l.wrap.b32 %r9795, %r9793, %r9793, 16; sub.s32 %r9796, %r9794, %r9795; xor.b32 %r9797, %r9796, %r9790; shf.l.wrap.b32 %r9798, %r9796, %r9796, 4; sub.s32 %r9799, %r9797, %r9798; xor.b32 %r9800, %r9799, %r9793; shf.l.wrap.b32 %r9801, %r9799, %r9799, 14; sub.s32 %r9802, %r9800, %r9801; xor.b32 %r9803, %r9802, %r9796; shf.l.wrap.b32 %r9804, %r9802, %r9802, 24; sub.s32 %r9805, %r9803, %r9804; and.b32 %r9806, %r9805, 7; add.s32 %r9807, %r9806, 1; cvt.rn.f32.s32 %f5394, %r9807; and.b32 %r9808, %r9805, 8; setp.eq.s32 %p1949, %r9808, 0; neg.ftz.f32 %f5395, %f5394; selp.f32 %f5396, %f5394, %f5395, %p1949; mul.ftz.f32 %f5397, %f5388, %f5396; add.s32 %r9809, %r9787, -559038719; xor.b32 %r9810, %r9809, -1874868139; add.s32 %r9811, %r9810, 29184895; xor.b32 %r9812, %r9811, -559038720; shf.l.wrap.b32 %r9813, %r9811, %r9811, 25; sub.s32 %r9814, %r9812, %r9813; xor.b32 %r9815, %r9814, -1874868139; shf.l.wrap.b32 %r9816, %r9814, %r9814, 16; sub.s32 %r9817, %r9815, %r9816; xor.b32 %r9818, %r9817, %r9811; shf.l.wrap.b32 %r9819, %r9817, %r9817, 4; sub.s32 %r9820, %r9818, %r9819; xor.b32 %r9821, %r9820, %r9814; shf.l.wrap.b32 %r9822, %r9820, %r9820, 14; sub.s32 %r9823, %r9821, %r9822; xor.b32 %r9824, %r9823, %r9817; shf.l.wrap.b32 %r9825, %r9823, %r9823, 24; sub.s32 %r9826, %r9824, %r9825; add.ftz.f32 %f5398, %f5388, 0fBF800000; and.b32 %r9827, %r9826, 7; add.s32 %r9828, %r9827, 1; cvt.rn.f32.s32 %f5399, %r9828; and.b32 %r9829, %r9826, 8; setp.eq.s32 %p1950, %r9829, 0; neg.ftz.f32 %f5400, %f5399; selp.f32 %f5401, %f5399, %f5400, %p1950; mul.ftz.f32 %f5402, %f5398, %f5401; sub.ftz.f32 %f5403, %f5402, %f5397; fma.rn.ftz.f32 %f773, %f5393, %f5403, %f5397; mov.b32 %r473, %f773; setp.nan.ftz.f32 %p1951, %f773, %f773; mov.f32 %f5810, 0f00000000; @%p1951 bra $L__BB74_438; or.b32 %r9830, %r473, -2147483648; setp.ne.s32 %p1952, %r9830, -2147483648; add.ftz.f32 %f5405, %f773, %f773; setp.eq.ftz.f32 %p1953, %f5405, %f773; and.pred %p1954, %p1953, %p1952; @%p1954 bra $L__BB74_438; shl.b32 %r9831, %r473, 1; setp.lt.u32 %p1955, %r9831, -16777215; mul.ftz.f32 %f5406, %f773, 0f3E800000; selp.f32 %f5810, %f5406, 0f00000000, %p1955; $L__BB74_438: add.ftz.f32 %f5407, %f21, %f5810; mul.ftz.f32 %f5408, %f5807, %f5407; setp.gt.ftz.f32 %p1956, %f5809, 0f3F800000; selp.f32 %f5409, 0f3F800000, %f5809, %p1956; mul.ftz.f32 %f5410, %f5409, %f772; fma.rn.ftz.f32 %f5827, %f5410, %f5408, %f5827; bra.uni $L__BB74_450; $L__BB74_419: lg2.approx.ftz.f32 %f5304, %f27; mul.ftz.f32 %f5305, %f25, %f5304; neg.ftz.f32 %f5306, %f5305; ex2.approx.ftz.f32 %f740, %f5306; cvt.rmi.ftz.f32.f32 %f5307, %f5815; cvt.rzi.ftz.s32.f32 %r9650, %f5307; sub.ftz.f32 %f5308, %f5815, %f5307; mul.ftz.f32 %f5309, %f5308, %f5308; mul.ftz.f32 %f5310, %f5308, %f5309; fma.rn.ftz.f32 %f5311, %f5308, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5312, %f5308, %f5311, 0f41200000; mul.ftz.f32 %f5313, %f5310, %f5312; add.s32 %r9651, %r9650, -559038720; xor.b32 %r9652, %r9651, -1874868139; add.s32 %r9653, %r9652, 29184895; xor.b32 %r9654, %r9653, -559038720; shf.l.wrap.b32 %r9655, %r9653, %r9653, 25; sub.s32 %r9656, %r9654, %r9655; xor.b32 %r9657, %r9656, -1874868139; shf.l.wrap.b32 %r9658, %r9656, %r9656, 16; sub.s32 %r9659, %r9657, %r9658; xor.b32 %r9660, %r9659, %r9653; shf.l.wrap.b32 %r9661, %r9659, %r9659, 4; sub.s32 %r9662, %r9660, %r9661; xor.b32 %r9663, %r9662, %r9656; shf.l.wrap.b32 %r9664, %r9662, %r9662, 14; sub.s32 %r9665, %r9663, %r9664; xor.b32 %r9666, %r9665, %r9659; shf.l.wrap.b32 %r9667, %r9665, %r9665, 24; sub.s32 %r9668, %r9666, %r9667; and.b32 %r9669, %r9668, 7; add.s32 %r9670, %r9669, 1; cvt.rn.f32.s32 %f5314, %r9670; and.b32 %r9671, %r9668, 8; setp.eq.s32 %p1918, %r9671, 0; neg.ftz.f32 %f5315, %f5314; selp.f32 %f5316, %f5314, %f5315, %p1918; mul.ftz.f32 %f5317, %f5308, %f5316; add.s32 %r9672, %r9650, -559038719; xor.b32 %r9673, %r9672, -1874868139; add.s32 %r9674, %r9673, 29184895; xor.b32 %r9675, %r9674, -559038720; shf.l.wrap.b32 %r9676, %r9674, %r9674, 25; sub.s32 %r9677, %r9675, %r9676; xor.b32 %r9678, %r9677, -1874868139; shf.l.wrap.b32 %r9679, %r9677, %r9677, 16; sub.s32 %r9680, %r9678, %r9679; xor.b32 %r9681, %r9680, %r9674; shf.l.wrap.b32 %r9682, %r9680, %r9680, 4; sub.s32 %r9683, %r9681, %r9682; xor.b32 %r9684, %r9683, %r9677; shf.l.wrap.b32 %r9685, %r9683, %r9683, 14; sub.s32 %r9686, %r9684, %r9685; xor.b32 %r9687, %r9686, %r9680; shf.l.wrap.b32 %r9688, %r9686, %r9686, 24; sub.s32 %r9689, %r9687, %r9688; add.ftz.f32 %f5318, %f5308, 0fBF800000; and.b32 %r9690, %r9689, 7; add.s32 %r9691, %r9690, 1; cvt.rn.f32.s32 %f5319, %r9691; and.b32 %r9692, %r9689, 8; setp.eq.s32 %p1919, %r9692, 0; neg.ftz.f32 %f5320, %f5319; selp.f32 %f5321, %f5319, %f5320, %p1919; mul.ftz.f32 %f5322, %f5318, %f5321; sub.ftz.f32 %f5323, %f5322, %f5317; fma.rn.ftz.f32 %f741, %f5313, %f5323, %f5317; mov.b32 %r464, %f741; setp.nan.ftz.f32 %p1920, %f741, %f741; mov.f32 %f5795, 0f00000000; @%p1920 bra $L__BB74_422; or.b32 %r9693, %r464, -2147483648; setp.ne.s32 %p1921, %r9693, -2147483648; add.ftz.f32 %f5325, %f741, %f741; setp.eq.ftz.f32 %p1922, %f5325, %f741; and.pred %p1923, %p1922, %p1921; @%p1923 bra $L__BB74_422; shl.b32 %r9694, %r464, 1; setp.lt.u32 %p1924, %r9694, -16777215; mul.ftz.f32 %f5326, %f741, 0f3E800000; selp.f32 %f5795, %f5326, 0f00000000, %p1924; $L__BB74_422: abs.ftz.f32 %f5327, %f5795; sub.ftz.f32 %f5328, %f21, %f5327; mul.ftz.f32 %f5827, %f5328, %f5328; cvt.rzi.ftz.s32.f32 %r465, %f26; setp.lt.s32 %p1925, %r465, 2; @%p1925 bra $L__BB74_450; mov.u32 %r10034, 1; mov.f32 %f5797, %f5827; mov.f32 %f5798, %f740; $L__BB74_424: mul.ftz.f32 %f5815, %f27, %f5815; cvt.rmi.ftz.f32.f32 %f5330, %f5815; cvt.rzi.ftz.s32.f32 %r9696, %f5330; sub.ftz.f32 %f5331, %f5815, %f5330; mul.ftz.f32 %f5332, %f5331, %f5331; mul.ftz.f32 %f5333, %f5331, %f5332; fma.rn.ftz.f32 %f5334, %f5331, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5335, %f5331, %f5334, 0f41200000; mul.ftz.f32 %f5336, %f5333, %f5335; add.s32 %r9697, %r9696, -559038720; xor.b32 %r9698, %r9697, -1874868139; add.s32 %r9699, %r9698, 29184895; xor.b32 %r9700, %r9699, -559038720; shf.l.wrap.b32 %r9701, %r9699, %r9699, 25; sub.s32 %r9702, %r9700, %r9701; xor.b32 %r9703, %r9702, -1874868139; shf.l.wrap.b32 %r9704, %r9702, %r9702, 16; sub.s32 %r9705, %r9703, %r9704; xor.b32 %r9706, %r9705, %r9699; shf.l.wrap.b32 %r9707, %r9705, %r9705, 4; sub.s32 %r9708, %r9706, %r9707; xor.b32 %r9709, %r9708, %r9702; shf.l.wrap.b32 %r9710, %r9708, %r9708, 14; sub.s32 %r9711, %r9709, %r9710; xor.b32 %r9712, %r9711, %r9705; shf.l.wrap.b32 %r9713, %r9711, %r9711, 24; sub.s32 %r9714, %r9712, %r9713; and.b32 %r9715, %r9714, 7; add.s32 %r9716, %r9715, 1; cvt.rn.f32.s32 %f5337, %r9716; and.b32 %r9717, %r9714, 8; setp.eq.s32 %p1926, %r9717, 0; neg.ftz.f32 %f5338, %f5337; selp.f32 %f5339, %f5337, %f5338, %p1926; mul.ftz.f32 %f5340, %f5331, %f5339; add.s32 %r9718, %r9696, -559038719; xor.b32 %r9719, %r9718, -1874868139; add.s32 %r9720, %r9719, 29184895; xor.b32 %r9721, %r9720, -559038720; shf.l.wrap.b32 %r9722, %r9720, %r9720, 25; sub.s32 %r9723, %r9721, %r9722; xor.b32 %r9724, %r9723, -1874868139; shf.l.wrap.b32 %r9725, %r9723, %r9723, 16; sub.s32 %r9726, %r9724, %r9725; xor.b32 %r9727, %r9726, %r9720; shf.l.wrap.b32 %r9728, %r9726, %r9726, 4; sub.s32 %r9729, %r9727, %r9728; xor.b32 %r9730, %r9729, %r9723; shf.l.wrap.b32 %r9731, %r9729, %r9729, 14; sub.s32 %r9732, %r9730, %r9731; xor.b32 %r9733, %r9732, %r9726; shf.l.wrap.b32 %r9734, %r9732, %r9732, 24; sub.s32 %r9735, %r9733, %r9734; add.ftz.f32 %f5341, %f5331, 0fBF800000; and.b32 %r9736, %r9735, 7; add.s32 %r9737, %r9736, 1; cvt.rn.f32.s32 %f5342, %r9737; and.b32 %r9738, %r9735, 8; setp.eq.s32 %p1927, %r9738, 0; neg.ftz.f32 %f5343, %f5342; selp.f32 %f5344, %f5342, %f5343, %p1927; mul.ftz.f32 %f5345, %f5341, %f5344; sub.ftz.f32 %f5346, %f5345, %f5340; fma.rn.ftz.f32 %f750, %f5336, %f5346, %f5340; mov.b32 %r467, %f750; setp.nan.ftz.f32 %p1928, %f750, %f750; mov.f32 %f5800, 0f00000000; @%p1928 bra $L__BB74_427; or.b32 %r9739, %r467, -2147483648; setp.ne.s32 %p1929, %r9739, -2147483648; add.ftz.f32 %f5348, %f750, %f750; setp.eq.ftz.f32 %p1930, %f5348, %f750; and.pred %p1931, %p1930, %p1929; @%p1931 bra $L__BB74_427; shl.b32 %r9740, %r467, 1; setp.lt.u32 %p1932, %r9740, -16777215; selp.f32 %f5800, %f750, 0f00000000, %p1932; $L__BB74_427: mul.ftz.f32 %f5349, %f5800, 0f3E800000; abs.ftz.f32 %f5350, %f5349; sub.ftz.f32 %f5351, %f21, %f5350; mul.ftz.f32 %f5352, %f5351, %f5351; mul.ftz.f32 %f5353, %f24, %f5797; cvt.ftz.sat.f32.f32 %f5354, %f5353; mul.ftz.f32 %f5797, %f5354, %f5352; fma.rn.ftz.f32 %f5827, %f5798, %f5797, %f5827; mul.ftz.f32 %f5798, %f740, %f5798; add.s32 %r10034, %r10034, 1; setp.lt.s32 %p1933, %r10034, %r465; @%p1933 bra $L__BB74_424; bra.uni $L__BB74_450; $L__BB74_254: lg2.approx.ftz.f32 %f3998, %f27; mul.ftz.f32 %f3999, %f25, %f3998; neg.ftz.f32 %f4000, %f3999; ex2.approx.ftz.f32 %f363, %f4000; cvt.rzi.ftz.s32.f32 %r213, %f26; mov.f32 %f5670, 0f3F800000; mov.f32 %f5827, 0f00000000; mov.u32 %r10025, 0; mov.pred %p1390, -1; mov.pred %p1438, 0; mov.f32 %f5668, %f5670; $L__BB74_255: setp.ge.s32 %p1391, %r10025, %r213; mov.pred %p1992, %p1390; @%p1391 bra $L__BB74_276; cvt.rmi.ftz.f32.f32 %f4001, %f5694; cvt.rzi.ftz.s32.f32 %r215, %f4001; sub.ftz.f32 %f370, %f5694, %f4001; cvt.rmi.ftz.f32.f32 %f4002, %f5695; cvt.rzi.ftz.s32.f32 %r216, %f4002; sub.ftz.f32 %f371, %f5695, %f4002; cvt.rmi.ftz.f32.f32 %f4003, %f5696; cvt.rzi.ftz.s32.f32 %r217, %f4003; sub.ftz.f32 %f372, %f5696, %f4003; fma.rn.ftz.f32 %f4004, %f370, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f373, %f370, %f4004, 0f41200000; fma.rn.ftz.f32 %f4005, %f371, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f374, %f371, %f4005, 0f41200000; fma.rn.ftz.f32 %f4006, %f372, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f375, %f372, %f4006, 0f41200000; add.s32 %r7531, %r216, -559038712; add.s32 %r7532, %r217, -559038712; xor.b32 %r7533, %r7532, %r7531; shf.l.wrap.b32 %r218, %r7531, %r7531, 14; sub.s32 %r219, %r7533, %r218; add.s32 %r7534, %r215, -559038712; xor.b32 %r7535, %r219, %r7534; shf.l.wrap.b32 %r220, %r219, %r219, 11; sub.s32 %r7536, %r7535, %r220; xor.b32 %r7537, %r7536, %r7531; shf.l.wrap.b32 %r7538, %r7536, %r7536, 25; sub.s32 %r7539, %r7537, %r7538; xor.b32 %r7540, %r7539, %r219; shf.l.wrap.b32 %r7541, %r7539, %r7539, 16; sub.s32 %r7542, %r7540, %r7541; xor.b32 %r7543, %r7542, %r7536; shf.l.wrap.b32 %r7544, %r7542, %r7542, 4; sub.s32 %r7545, %r7543, %r7544; xor.b32 %r7546, %r7545, %r7539; shf.l.wrap.b32 %r7547, %r7545, %r7545, 14; sub.s32 %r7548, %r7546, %r7547; xor.b32 %r7549, %r7548, %r7542; shf.l.wrap.b32 %r7550, %r7548, %r7548, 24; sub.s32 %r221, %r7549, %r7550; and.b32 %r222, %r221, 15; cvt.u16.u32 %rs127, %r221; or.b16 %rs128, %rs127, 2; and.b16 %rs129, %rs128, 15; setp.eq.s16 %p1392, %rs129, 14; mov.f32 %f5656, %f370; @%p1392 bra $L__BB74_258; mov.f32 %f5656, %f372; $L__BB74_258: setp.gt.ftz.f32 %p1393, %f5670, 0f3F800000; selp.f32 %f377, 0f3F800000, %f5670, %p1393; mul.ftz.f32 %f4007, %f370, %f370; mul.ftz.f32 %f4008, %f370, %f4007; mul.ftz.f32 %f378, %f4008, %f373; mul.ftz.f32 %f4009, %f371, %f371; mul.ftz.f32 %f4010, %f371, %f4009; mul.ftz.f32 %f379, %f4010, %f374; mul.ftz.f32 %f4011, %f372, %f372; mul.ftz.f32 %f4012, %f372, %f4011; mul.ftz.f32 %f380, %f4012, %f375; setp.lt.u32 %p1394, %r222, 8; selp.f32 %f4013, %f370, %f371, %p1394; setp.lt.u32 %p1395, %r222, 4; selp.f32 %f4014, %f371, %f5656, %p1395; and.b32 %r7551, %r221, 1; setp.eq.b32 %p1396, %r7551, 1; neg.ftz.f32 %f4015, %f4013; selp.f32 %f4016, %f4015, %f4013, %p1396; and.b32 %r7552, %r221, 2; setp.eq.s32 %p1397, %r7552, 0; neg.ftz.f32 %f4017, %f4014; selp.f32 %f4018, %f4014, %f4017, %p1397; add.ftz.f32 %f381, %f4016, %f4018; add.s32 %r223, %r215, -559038711; xor.b32 %r7553, %r219, %r223; sub.s32 %r7554, %r7553, %r220; xor.b32 %r7556, %r7554, %r7531; shf.l.wrap.b32 %r7557, %r7554, %r7554, 25; sub.s32 %r7558, %r7556, %r7557; xor.b32 %r7559, %r7558, %r219; shf.l.wrap.b32 %r7560, %r7558, %r7558, 16; sub.s32 %r7561, %r7559, %r7560; xor.b32 %r7562, %r7561, %r7554; shf.l.wrap.b32 %r7563, %r7561, %r7561, 4; sub.s32 %r7564, %r7562, %r7563; xor.b32 %r7565, %r7564, %r7558; shf.l.wrap.b32 %r7566, %r7564, %r7564, 14; sub.s32 %r7567, %r7565, %r7566; xor.b32 %r7568, %r7567, %r7561; shf.l.wrap.b32 %r7569, %r7567, %r7567, 24; sub.s32 %r224, %r7568, %r7569; and.b32 %r225, %r224, 15; add.ftz.f32 %f382, %f370, 0fBF800000; cvt.u16.u32 %rs130, %r224; or.b16 %rs131, %rs130, 2; and.b16 %rs132, %rs131, 15; setp.eq.s16 %p1398, %rs132, 14; mov.f32 %f5657, %f382; @%p1398 bra $L__BB74_260; mov.f32 %f5657, %f372; $L__BB74_260: setp.lt.u32 %p1399, %r225, 8; selp.f32 %f4019, %f382, %f371, %p1399; setp.lt.u32 %p1400, %r225, 4; selp.f32 %f4020, %f371, %f5657, %p1400; and.b32 %r7570, %r224, 1; setp.eq.b32 %p1401, %r7570, 1; neg.ftz.f32 %f4021, %f4019; selp.f32 %f4022, %f4021, %f4019, %p1401; and.b32 %r7571, %r224, 2; setp.eq.s32 %p1402, %r7571, 0; neg.ftz.f32 %f4023, %f4020; selp.f32 %f4024, %f4020, %f4023, %p1402; add.ftz.f32 %f384, %f4022, %f4024; add.s32 %r226, %r216, -559038711; xor.b32 %r7573, %r7532, %r226; shf.l.wrap.b32 %r227, %r226, %r226, 14; sub.s32 %r228, %r7573, %r227; xor.b32 %r7575, %r228, %r7534; shf.l.wrap.b32 %r229, %r228, %r228, 11; sub.s32 %r7576, %r7575, %r229; xor.b32 %r7577, %r7576, %r226; shf.l.wrap.b32 %r7578, %r7576, %r7576, 25; sub.s32 %r7579, %r7577, %r7578; xor.b32 %r7580, %r7579, %r228; shf.l.wrap.b32 %r7581, %r7579, %r7579, 16; sub.s32 %r7582, %r7580, %r7581; xor.b32 %r7583, %r7582, %r7576; shf.l.wrap.b32 %r7584, %r7582, %r7582, 4; sub.s32 %r7585, %r7583, %r7584; xor.b32 %r7586, %r7585, %r7579; shf.l.wrap.b32 %r7587, %r7585, %r7585, 14; sub.s32 %r7588, %r7586, %r7587; xor.b32 %r7589, %r7588, %r7582; shf.l.wrap.b32 %r7590, %r7588, %r7588, 24; sub.s32 %r230, %r7589, %r7590; and.b32 %r231, %r230, 15; add.ftz.f32 %f385, %f371, 0fBF800000; cvt.u16.u32 %rs133, %r230; or.b16 %rs134, %rs133, 2; and.b16 %rs135, %rs134, 15; setp.eq.s16 %p1403, %rs135, 14; mov.f32 %f5658, %f370; @%p1403 bra $L__BB74_262; mov.f32 %f5658, %f372; $L__BB74_262: setp.lt.u32 %p1404, %r231, 8; selp.f32 %f4025, %f370, %f385, %p1404; setp.lt.u32 %p1405, %r231, 4; selp.f32 %f4026, %f385, %f5658, %p1405; and.b32 %r7591, %r230, 1; setp.eq.b32 %p1406, %r7591, 1; neg.ftz.f32 %f4027, %f4025; selp.f32 %f4028, %f4027, %f4025, %p1406; and.b32 %r7592, %r230, 2; setp.eq.s32 %p1407, %r7592, 0; neg.ftz.f32 %f4029, %f4026; selp.f32 %f4030, %f4026, %f4029, %p1407; add.ftz.f32 %f387, %f4028, %f4030; xor.b32 %r7593, %r228, %r223; sub.s32 %r7594, %r7593, %r229; xor.b32 %r7595, %r7594, %r226; shf.l.wrap.b32 %r7596, %r7594, %r7594, 25; sub.s32 %r7597, %r7595, %r7596; xor.b32 %r7598, %r7597, %r228; shf.l.wrap.b32 %r7599, %r7597, %r7597, 16; sub.s32 %r7600, %r7598, %r7599; xor.b32 %r7601, %r7600, %r7594; shf.l.wrap.b32 %r7602, %r7600, %r7600, 4; sub.s32 %r7603, %r7601, %r7602; xor.b32 %r7604, %r7603, %r7597; shf.l.wrap.b32 %r7605, %r7603, %r7603, 14; sub.s32 %r7606, %r7604, %r7605; xor.b32 %r7607, %r7606, %r7600; shf.l.wrap.b32 %r7608, %r7606, %r7606, 24; sub.s32 %r232, %r7607, %r7608; and.b32 %r233, %r232, 15; cvt.u16.u32 %rs136, %r232; or.b16 %rs137, %rs136, 2; and.b16 %rs138, %rs137, 15; setp.eq.s16 %p1408, %rs138, 14; mov.f32 %f5659, %f382; @%p1408 bra $L__BB74_264; mov.f32 %f5659, %f372; $L__BB74_264: setp.lt.u32 %p1409, %r233, 8; selp.f32 %f4031, %f382, %f385, %p1409; setp.lt.u32 %p1410, %r233, 4; selp.f32 %f4032, %f385, %f5659, %p1410; and.b32 %r7609, %r232, 1; setp.eq.b32 %p1411, %r7609, 1; neg.ftz.f32 %f4033, %f4031; selp.f32 %f4034, %f4033, %f4031, %p1411; and.b32 %r7610, %r232, 2; setp.eq.s32 %p1412, %r7610, 0; neg.ftz.f32 %f4035, %f4032; selp.f32 %f4036, %f4032, %f4035, %p1412; add.ftz.f32 %f389, %f4034, %f4036; add.s32 %r234, %r217, -559038711; xor.b32 %r7612, %r234, %r7531; sub.s32 %r235, %r7612, %r218; xor.b32 %r7614, %r235, %r7534; shf.l.wrap.b32 %r236, %r235, %r235, 11; sub.s32 %r7615, %r7614, %r236; xor.b32 %r7616, %r7615, %r7531; shf.l.wrap.b32 %r7617, %r7615, %r7615, 25; sub.s32 %r7618, %r7616, %r7617; xor.b32 %r7619, %r7618, %r235; shf.l.wrap.b32 %r7620, %r7618, %r7618, 16; sub.s32 %r7621, %r7619, %r7620; xor.b32 %r7622, %r7621, %r7615; shf.l.wrap.b32 %r7623, %r7621, %r7621, 4; sub.s32 %r7624, %r7622, %r7623; xor.b32 %r7625, %r7624, %r7618; shf.l.wrap.b32 %r7626, %r7624, %r7624, 14; sub.s32 %r7627, %r7625, %r7626; xor.b32 %r7628, %r7627, %r7621; shf.l.wrap.b32 %r7629, %r7627, %r7627, 24; sub.s32 %r237, %r7628, %r7629; add.ftz.f32 %f390, %f372, 0fBF800000; and.b32 %r238, %r237, 15; cvt.u16.u32 %rs139, %r237; or.b16 %rs140, %rs139, 2; and.b16 %rs141, %rs140, 15; setp.eq.s16 %p1413, %rs141, 14; mov.f32 %f5660, %f370; @%p1413 bra $L__BB74_266; mov.f32 %f5660, %f390; $L__BB74_266: setp.lt.u32 %p1414, %r238, 8; selp.f32 %f4037, %f370, %f371, %p1414; setp.lt.u32 %p1415, %r238, 4; selp.f32 %f4038, %f371, %f5660, %p1415; and.b32 %r7630, %r237, 1; setp.eq.b32 %p1416, %r7630, 1; neg.ftz.f32 %f4039, %f4037; selp.f32 %f4040, %f4039, %f4037, %p1416; and.b32 %r7631, %r237, 2; setp.eq.s32 %p1417, %r7631, 0; neg.ftz.f32 %f4041, %f4038; selp.f32 %f4042, %f4038, %f4041, %p1417; add.ftz.f32 %f392, %f4040, %f4042; xor.b32 %r7632, %r235, %r223; sub.s32 %r7633, %r7632, %r236; xor.b32 %r7635, %r7633, %r7531; shf.l.wrap.b32 %r7636, %r7633, %r7633, 25; sub.s32 %r7637, %r7635, %r7636; xor.b32 %r7638, %r7637, %r235; shf.l.wrap.b32 %r7639, %r7637, %r7637, 16; sub.s32 %r7640, %r7638, %r7639; xor.b32 %r7641, %r7640, %r7633; shf.l.wrap.b32 %r7642, %r7640, %r7640, 4; sub.s32 %r7643, %r7641, %r7642; xor.b32 %r7644, %r7643, %r7637; shf.l.wrap.b32 %r7645, %r7643, %r7643, 14; sub.s32 %r7646, %r7644, %r7645; xor.b32 %r7647, %r7646, %r7640; shf.l.wrap.b32 %r7648, %r7646, %r7646, 24; sub.s32 %r239, %r7647, %r7648; and.b32 %r240, %r239, 15; cvt.u16.u32 %rs142, %r239; or.b16 %rs143, %rs142, 2; and.b16 %rs144, %rs143, 15; setp.eq.s16 %p1418, %rs144, 14; mov.f32 %f5661, %f382; @%p1418 bra $L__BB74_268; mov.f32 %f5661, %f390; $L__BB74_268: setp.lt.u32 %p1419, %r240, 8; selp.f32 %f4043, %f382, %f371, %p1419; setp.lt.u32 %p1420, %r240, 4; selp.f32 %f4044, %f371, %f5661, %p1420; and.b32 %r7649, %r239, 1; setp.eq.b32 %p1421, %r7649, 1; neg.ftz.f32 %f4045, %f4043; selp.f32 %f4046, %f4045, %f4043, %p1421; and.b32 %r7650, %r239, 2; setp.eq.s32 %p1422, %r7650, 0; neg.ftz.f32 %f4047, %f4044; selp.f32 %f4048, %f4044, %f4047, %p1422; add.ftz.f32 %f394, %f4046, %f4048; xor.b32 %r7651, %r234, %r226; sub.s32 %r241, %r7651, %r227; xor.b32 %r7653, %r241, %r7534; shf.l.wrap.b32 %r242, %r241, %r241, 11; sub.s32 %r7654, %r7653, %r242; xor.b32 %r7655, %r7654, %r226; shf.l.wrap.b32 %r7656, %r7654, %r7654, 25; sub.s32 %r7657, %r7655, %r7656; xor.b32 %r7658, %r7657, %r241; shf.l.wrap.b32 %r7659, %r7657, %r7657, 16; sub.s32 %r7660, %r7658, %r7659; xor.b32 %r7661, %r7660, %r7654; shf.l.wrap.b32 %r7662, %r7660, %r7660, 4; sub.s32 %r7663, %r7661, %r7662; xor.b32 %r7664, %r7663, %r7657; shf.l.wrap.b32 %r7665, %r7663, %r7663, 14; sub.s32 %r7666, %r7664, %r7665; xor.b32 %r7667, %r7666, %r7660; shf.l.wrap.b32 %r7668, %r7666, %r7666, 24; sub.s32 %r243, %r7667, %r7668; and.b32 %r244, %r243, 15; cvt.u16.u32 %rs145, %r243; or.b16 %rs146, %rs145, 2; and.b16 %rs147, %rs146, 15; setp.eq.s16 %p1423, %rs147, 14; mov.f32 %f5662, %f370; @%p1423 bra $L__BB74_270; mov.f32 %f5662, %f390; $L__BB74_270: setp.lt.u32 %p1424, %r244, 8; selp.f32 %f4049, %f370, %f385, %p1424; setp.lt.u32 %p1425, %r244, 4; selp.f32 %f4050, %f385, %f5662, %p1425; and.b32 %r7669, %r243, 1; setp.eq.b32 %p1426, %r7669, 1; neg.ftz.f32 %f4051, %f4049; selp.f32 %f4052, %f4051, %f4049, %p1426; and.b32 %r7670, %r243, 2; setp.eq.s32 %p1427, %r7670, 0; neg.ftz.f32 %f4053, %f4050; selp.f32 %f4054, %f4050, %f4053, %p1427; add.ftz.f32 %f396, %f4052, %f4054; xor.b32 %r7671, %r241, %r223; sub.s32 %r7672, %r7671, %r242; xor.b32 %r7673, %r7672, %r226; shf.l.wrap.b32 %r7674, %r7672, %r7672, 25; sub.s32 %r7675, %r7673, %r7674; xor.b32 %r7676, %r7675, %r241; shf.l.wrap.b32 %r7677, %r7675, %r7675, 16; sub.s32 %r7678, %r7676, %r7677; xor.b32 %r7679, %r7678, %r7672; shf.l.wrap.b32 %r7680, %r7678, %r7678, 4; sub.s32 %r7681, %r7679, %r7680; xor.b32 %r7682, %r7681, %r7675; shf.l.wrap.b32 %r7683, %r7681, %r7681, 14; sub.s32 %r7684, %r7682, %r7683; xor.b32 %r7685, %r7684, %r7678; shf.l.wrap.b32 %r7686, %r7684, %r7684, 24; sub.s32 %r245, %r7685, %r7686; and.b32 %r246, %r245, 15; cvt.u16.u32 %rs148, %r245; or.b16 %rs149, %rs148, 2; and.b16 %rs150, %rs149, 15; setp.eq.s16 %p1428, %rs150, 14; mov.f32 %f5663, %f382; @%p1428 bra $L__BB74_272; mov.f32 %f5663, %f390; $L__BB74_272: setp.lt.u32 %p1429, %r246, 8; selp.f32 %f4056, %f382, %f385, %p1429; setp.lt.u32 %p1430, %r246, 4; selp.f32 %f4057, %f385, %f5663, %p1430; and.b32 %r7687, %r245, 1; setp.eq.b32 %p1431, %r7687, 1; neg.ftz.f32 %f4058, %f4056; selp.f32 %f4059, %f4058, %f4056, %p1431; and.b32 %r7688, %r245, 2; setp.eq.s32 %p1432, %r7688, 0; neg.ftz.f32 %f4060, %f4057; selp.f32 %f4061, %f4057, %f4060, %p1432; add.ftz.f32 %f4062, %f4059, %f4061; mov.f32 %f4063, 0f3F800000; sub.ftz.f32 %f4064, %f4063, %f378; mul.ftz.f32 %f4065, %f378, %f384; fma.rn.ftz.f32 %f4066, %f4064, %f381, %f4065; sub.ftz.f32 %f4067, %f4063, %f379; mul.ftz.f32 %f4068, %f378, %f389; fma.rn.ftz.f32 %f4069, %f4064, %f387, %f4068; mul.ftz.f32 %f4070, %f379, %f4069; fma.rn.ftz.f32 %f4071, %f4067, %f4066, %f4070; sub.ftz.f32 %f4072, %f4063, %f380; mul.ftz.f32 %f4073, %f378, %f394; fma.rn.ftz.f32 %f4074, %f4064, %f392, %f4073; mul.ftz.f32 %f4075, %f378, %f4062; fma.rn.ftz.f32 %f4076, %f4064, %f396, %f4075; mul.ftz.f32 %f4077, %f379, %f4076; fma.rn.ftz.f32 %f4078, %f4067, %f4074, %f4077; mul.ftz.f32 %f4079, %f380, %f4078; fma.rn.ftz.f32 %f398, %f4072, %f4071, %f4079; mov.b32 %r247, %f398; setp.nan.ftz.f32 %p1433, %f398, %f398; mov.f32 %f5664, 0f00000000; @%p1433 bra $L__BB74_275; or.b32 %r7689, %r247, -2147483648; setp.ne.s32 %p1434, %r7689, -2147483648; add.ftz.f32 %f4081, %f398, %f398; setp.eq.ftz.f32 %p1435, %f4081, %f398; and.pred %p1436, %p1435, %p1434; @%p1436 bra $L__BB74_275; shl.b32 %r7690, %r247, 1; setp.lt.u32 %p1437, %r7690, -16777215; selp.f32 %f5664, %f398, 0f00000000, %p1437; $L__BB74_275: fma.rn.ftz.f32 %f4082, %f5664, 0f3F7B645A, %f21; mul.ftz.f32 %f4083, %f5668, %f4082; mul.ftz.f32 %f5668, %f363, %f5668; fma.rn.ftz.f32 %f5827, %f377, %f4083, %f5827; mul.ftz.f32 %f4084, %f24, %f4083; mul.ftz.f32 %f5670, %f377, %f4084; mul.ftz.f32 %f5694, %f27, %f5694; mul.ftz.f32 %f5695, %f27, %f5695; mul.ftz.f32 %f5696, %f27, %f5696; add.s32 %r10025, %r10025, 1; setp.gt.ftz.f32 %p1439, %f5670, 0f3A83126F; mov.pred %p1992, %p1438; @%p1439 bra $L__BB74_255; $L__BB74_276: cvt.rmi.ftz.f32.f32 %f4085, %f26; sub.ftz.f32 %f413, %f26, %f4085; setp.eq.ftz.f32 %p1440, %f413, 0f00000000; not.pred %p1441, %p1992; or.pred %p1442, %p1441, %p1440; @%p1442 bra $L__BB74_450; cvt.rmi.ftz.f32.f32 %f4086, %f5694; cvt.rzi.ftz.s32.f32 %r249, %f4086; sub.ftz.f32 %f414, %f5694, %f4086; cvt.rmi.ftz.f32.f32 %f4087, %f5695; cvt.rzi.ftz.s32.f32 %r250, %f4087; sub.ftz.f32 %f415, %f5695, %f4087; cvt.rmi.ftz.f32.f32 %f4088, %f5696; cvt.rzi.ftz.s32.f32 %r251, %f4088; sub.ftz.f32 %f416, %f5696, %f4088; fma.rn.ftz.f32 %f4089, %f414, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f417, %f414, %f4089, 0f41200000; fma.rn.ftz.f32 %f4090, %f415, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f418, %f415, %f4090, 0f41200000; fma.rn.ftz.f32 %f4091, %f416, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f419, %f416, %f4091, 0f41200000; add.s32 %r7691, %r250, -559038712; add.s32 %r7692, %r251, -559038712; xor.b32 %r7693, %r7692, %r7691; shf.l.wrap.b32 %r252, %r7691, %r7691, 14; sub.s32 %r253, %r7693, %r252; add.s32 %r7694, %r249, -559038712; xor.b32 %r7695, %r253, %r7694; shf.l.wrap.b32 %r254, %r253, %r253, 11; sub.s32 %r7696, %r7695, %r254; xor.b32 %r7697, %r7696, %r7691; shf.l.wrap.b32 %r7698, %r7696, %r7696, 25; sub.s32 %r7699, %r7697, %r7698; xor.b32 %r7700, %r7699, %r253; shf.l.wrap.b32 %r7701, %r7699, %r7699, 16; sub.s32 %r7702, %r7700, %r7701; xor.b32 %r7703, %r7702, %r7696; shf.l.wrap.b32 %r7704, %r7702, %r7702, 4; sub.s32 %r7705, %r7703, %r7704; xor.b32 %r7706, %r7705, %r7699; shf.l.wrap.b32 %r7707, %r7705, %r7705, 14; sub.s32 %r7708, %r7706, %r7707; xor.b32 %r7709, %r7708, %r7702; shf.l.wrap.b32 %r7710, %r7708, %r7708, 24; sub.s32 %r255, %r7709, %r7710; and.b32 %r256, %r255, 15; cvt.u16.u32 %rs151, %r255; or.b16 %rs152, %rs151, 2; and.b16 %rs153, %rs152, 15; setp.eq.s16 %p1443, %rs153, 14; mov.f32 %f5671, %f414; @%p1443 bra $L__BB74_279; mov.f32 %f5671, %f416; $L__BB74_279: mul.ftz.f32 %f4092, %f414, %f414; mul.ftz.f32 %f4093, %f414, %f4092; mul.ftz.f32 %f421, %f4093, %f417; mul.ftz.f32 %f4094, %f415, %f415; mul.ftz.f32 %f4095, %f415, %f4094; mul.ftz.f32 %f422, %f4095, %f418; mul.ftz.f32 %f4096, %f416, %f416; mul.ftz.f32 %f4097, %f416, %f4096; mul.ftz.f32 %f423, %f4097, %f419; setp.lt.u32 %p1444, %r256, 8; selp.f32 %f4098, %f414, %f415, %p1444; setp.lt.u32 %p1445, %r256, 4; selp.f32 %f4099, %f415, %f5671, %p1445; and.b32 %r7711, %r255, 1; setp.eq.b32 %p1446, %r7711, 1; neg.ftz.f32 %f4100, %f4098; selp.f32 %f4101, %f4100, %f4098, %p1446; and.b32 %r7712, %r255, 2; setp.eq.s32 %p1447, %r7712, 0; neg.ftz.f32 %f4102, %f4099; selp.f32 %f4103, %f4099, %f4102, %p1447; add.ftz.f32 %f424, %f4101, %f4103; add.s32 %r257, %r249, -559038711; xor.b32 %r7713, %r253, %r257; sub.s32 %r7714, %r7713, %r254; xor.b32 %r7716, %r7714, %r7691; shf.l.wrap.b32 %r7717, %r7714, %r7714, 25; sub.s32 %r7718, %r7716, %r7717; xor.b32 %r7719, %r7718, %r253; shf.l.wrap.b32 %r7720, %r7718, %r7718, 16; sub.s32 %r7721, %r7719, %r7720; xor.b32 %r7722, %r7721, %r7714; shf.l.wrap.b32 %r7723, %r7721, %r7721, 4; sub.s32 %r7724, %r7722, %r7723; xor.b32 %r7725, %r7724, %r7718; shf.l.wrap.b32 %r7726, %r7724, %r7724, 14; sub.s32 %r7727, %r7725, %r7726; xor.b32 %r7728, %r7727, %r7721; shf.l.wrap.b32 %r7729, %r7727, %r7727, 24; sub.s32 %r258, %r7728, %r7729; and.b32 %r259, %r258, 15; add.ftz.f32 %f425, %f414, 0fBF800000; cvt.u16.u32 %rs154, %r258; or.b16 %rs155, %rs154, 2; and.b16 %rs156, %rs155, 15; setp.eq.s16 %p1448, %rs156, 14; mov.f32 %f5672, %f425; @%p1448 bra $L__BB74_281; mov.f32 %f5672, %f416; $L__BB74_281: setp.lt.u32 %p1449, %r259, 8; selp.f32 %f4104, %f425, %f415, %p1449; setp.lt.u32 %p1450, %r259, 4; selp.f32 %f4105, %f415, %f5672, %p1450; and.b32 %r7730, %r258, 1; setp.eq.b32 %p1451, %r7730, 1; neg.ftz.f32 %f4106, %f4104; selp.f32 %f4107, %f4106, %f4104, %p1451; and.b32 %r7731, %r258, 2; setp.eq.s32 %p1452, %r7731, 0; neg.ftz.f32 %f4108, %f4105; selp.f32 %f4109, %f4105, %f4108, %p1452; add.ftz.f32 %f427, %f4107, %f4109; add.s32 %r260, %r250, -559038711; xor.b32 %r7733, %r7692, %r260; shf.l.wrap.b32 %r261, %r260, %r260, 14; sub.s32 %r262, %r7733, %r261; xor.b32 %r7735, %r262, %r7694; shf.l.wrap.b32 %r263, %r262, %r262, 11; sub.s32 %r7736, %r7735, %r263; xor.b32 %r7737, %r7736, %r260; shf.l.wrap.b32 %r7738, %r7736, %r7736, 25; sub.s32 %r7739, %r7737, %r7738; xor.b32 %r7740, %r7739, %r262; shf.l.wrap.b32 %r7741, %r7739, %r7739, 16; sub.s32 %r7742, %r7740, %r7741; xor.b32 %r7743, %r7742, %r7736; shf.l.wrap.b32 %r7744, %r7742, %r7742, 4; sub.s32 %r7745, %r7743, %r7744; xor.b32 %r7746, %r7745, %r7739; shf.l.wrap.b32 %r7747, %r7745, %r7745, 14; sub.s32 %r7748, %r7746, %r7747; xor.b32 %r7749, %r7748, %r7742; shf.l.wrap.b32 %r7750, %r7748, %r7748, 24; sub.s32 %r264, %r7749, %r7750; and.b32 %r265, %r264, 15; add.ftz.f32 %f428, %f415, 0fBF800000; cvt.u16.u32 %rs157, %r264; or.b16 %rs158, %rs157, 2; and.b16 %rs159, %rs158, 15; setp.eq.s16 %p1453, %rs159, 14; mov.f32 %f5673, %f414; @%p1453 bra $L__BB74_283; mov.f32 %f5673, %f416; $L__BB74_283: setp.lt.u32 %p1454, %r265, 8; selp.f32 %f4110, %f414, %f428, %p1454; setp.lt.u32 %p1455, %r265, 4; selp.f32 %f4111, %f428, %f5673, %p1455; and.b32 %r7751, %r264, 1; setp.eq.b32 %p1456, %r7751, 1; neg.ftz.f32 %f4112, %f4110; selp.f32 %f4113, %f4112, %f4110, %p1456; and.b32 %r7752, %r264, 2; setp.eq.s32 %p1457, %r7752, 0; neg.ftz.f32 %f4114, %f4111; selp.f32 %f4115, %f4111, %f4114, %p1457; add.ftz.f32 %f430, %f4113, %f4115; xor.b32 %r7753, %r262, %r257; sub.s32 %r7754, %r7753, %r263; xor.b32 %r7755, %r7754, %r260; shf.l.wrap.b32 %r7756, %r7754, %r7754, 25; sub.s32 %r7757, %r7755, %r7756; xor.b32 %r7758, %r7757, %r262; shf.l.wrap.b32 %r7759, %r7757, %r7757, 16; sub.s32 %r7760, %r7758, %r7759; xor.b32 %r7761, %r7760, %r7754; shf.l.wrap.b32 %r7762, %r7760, %r7760, 4; sub.s32 %r7763, %r7761, %r7762; xor.b32 %r7764, %r7763, %r7757; shf.l.wrap.b32 %r7765, %r7763, %r7763, 14; sub.s32 %r7766, %r7764, %r7765; xor.b32 %r7767, %r7766, %r7760; shf.l.wrap.b32 %r7768, %r7766, %r7766, 24; sub.s32 %r266, %r7767, %r7768; and.b32 %r267, %r266, 15; cvt.u16.u32 %rs160, %r266; or.b16 %rs161, %rs160, 2; and.b16 %rs162, %rs161, 15; setp.eq.s16 %p1458, %rs162, 14; mov.f32 %f5674, %f425; @%p1458 bra $L__BB74_285; mov.f32 %f5674, %f416; $L__BB74_285: setp.lt.u32 %p1459, %r267, 8; selp.f32 %f4116, %f425, %f428, %p1459; setp.lt.u32 %p1460, %r267, 4; selp.f32 %f4117, %f428, %f5674, %p1460; and.b32 %r7769, %r266, 1; setp.eq.b32 %p1461, %r7769, 1; neg.ftz.f32 %f4118, %f4116; selp.f32 %f4119, %f4118, %f4116, %p1461; and.b32 %r7770, %r266, 2; setp.eq.s32 %p1462, %r7770, 0; neg.ftz.f32 %f4120, %f4117; selp.f32 %f4121, %f4117, %f4120, %p1462; add.ftz.f32 %f432, %f4119, %f4121; add.s32 %r268, %r251, -559038711; xor.b32 %r7772, %r268, %r7691; sub.s32 %r269, %r7772, %r252; xor.b32 %r7774, %r269, %r7694; shf.l.wrap.b32 %r270, %r269, %r269, 11; sub.s32 %r7775, %r7774, %r270; xor.b32 %r7776, %r7775, %r7691; shf.l.wrap.b32 %r7777, %r7775, %r7775, 25; sub.s32 %r7778, %r7776, %r7777; xor.b32 %r7779, %r7778, %r269; shf.l.wrap.b32 %r7780, %r7778, %r7778, 16; sub.s32 %r7781, %r7779, %r7780; xor.b32 %r7782, %r7781, %r7775; shf.l.wrap.b32 %r7783, %r7781, %r7781, 4; sub.s32 %r7784, %r7782, %r7783; xor.b32 %r7785, %r7784, %r7778; shf.l.wrap.b32 %r7786, %r7784, %r7784, 14; sub.s32 %r7787, %r7785, %r7786; xor.b32 %r7788, %r7787, %r7781; shf.l.wrap.b32 %r7789, %r7787, %r7787, 24; sub.s32 %r271, %r7788, %r7789; add.ftz.f32 %f433, %f416, 0fBF800000; and.b32 %r272, %r271, 15; cvt.u16.u32 %rs163, %r271; or.b16 %rs164, %rs163, 2; and.b16 %rs165, %rs164, 15; setp.eq.s16 %p1463, %rs165, 14; mov.f32 %f5675, %f414; @%p1463 bra $L__BB74_287; mov.f32 %f5675, %f433; $L__BB74_287: setp.lt.u32 %p1464, %r272, 8; selp.f32 %f4122, %f414, %f415, %p1464; setp.lt.u32 %p1465, %r272, 4; selp.f32 %f4123, %f415, %f5675, %p1465; and.b32 %r7790, %r271, 1; setp.eq.b32 %p1466, %r7790, 1; neg.ftz.f32 %f4124, %f4122; selp.f32 %f4125, %f4124, %f4122, %p1466; and.b32 %r7791, %r271, 2; setp.eq.s32 %p1467, %r7791, 0; neg.ftz.f32 %f4126, %f4123; selp.f32 %f4127, %f4123, %f4126, %p1467; add.ftz.f32 %f435, %f4125, %f4127; xor.b32 %r7792, %r269, %r257; sub.s32 %r7793, %r7792, %r270; xor.b32 %r7795, %r7793, %r7691; shf.l.wrap.b32 %r7796, %r7793, %r7793, 25; sub.s32 %r7797, %r7795, %r7796; xor.b32 %r7798, %r7797, %r269; shf.l.wrap.b32 %r7799, %r7797, %r7797, 16; sub.s32 %r7800, %r7798, %r7799; xor.b32 %r7801, %r7800, %r7793; shf.l.wrap.b32 %r7802, %r7800, %r7800, 4; sub.s32 %r7803, %r7801, %r7802; xor.b32 %r7804, %r7803, %r7797; shf.l.wrap.b32 %r7805, %r7803, %r7803, 14; sub.s32 %r7806, %r7804, %r7805; xor.b32 %r7807, %r7806, %r7800; shf.l.wrap.b32 %r7808, %r7806, %r7806, 24; sub.s32 %r273, %r7807, %r7808; and.b32 %r274, %r273, 15; cvt.u16.u32 %rs166, %r273; or.b16 %rs167, %rs166, 2; and.b16 %rs168, %rs167, 15; setp.eq.s16 %p1468, %rs168, 14; mov.f32 %f5676, %f425; @%p1468 bra $L__BB74_289; mov.f32 %f5676, %f433; $L__BB74_289: setp.lt.u32 %p1469, %r274, 8; selp.f32 %f4128, %f425, %f415, %p1469; setp.lt.u32 %p1470, %r274, 4; selp.f32 %f4129, %f415, %f5676, %p1470; and.b32 %r7809, %r273, 1; setp.eq.b32 %p1471, %r7809, 1; neg.ftz.f32 %f4130, %f4128; selp.f32 %f4131, %f4130, %f4128, %p1471; and.b32 %r7810, %r273, 2; setp.eq.s32 %p1472, %r7810, 0; neg.ftz.f32 %f4132, %f4129; selp.f32 %f4133, %f4129, %f4132, %p1472; add.ftz.f32 %f437, %f4131, %f4133; xor.b32 %r7811, %r268, %r260; sub.s32 %r275, %r7811, %r261; xor.b32 %r7813, %r275, %r7694; shf.l.wrap.b32 %r276, %r275, %r275, 11; sub.s32 %r7814, %r7813, %r276; xor.b32 %r7815, %r7814, %r260; shf.l.wrap.b32 %r7816, %r7814, %r7814, 25; sub.s32 %r7817, %r7815, %r7816; xor.b32 %r7818, %r7817, %r275; shf.l.wrap.b32 %r7819, %r7817, %r7817, 16; sub.s32 %r7820, %r7818, %r7819; xor.b32 %r7821, %r7820, %r7814; shf.l.wrap.b32 %r7822, %r7820, %r7820, 4; sub.s32 %r7823, %r7821, %r7822; xor.b32 %r7824, %r7823, %r7817; shf.l.wrap.b32 %r7825, %r7823, %r7823, 14; sub.s32 %r7826, %r7824, %r7825; xor.b32 %r7827, %r7826, %r7820; shf.l.wrap.b32 %r7828, %r7826, %r7826, 24; sub.s32 %r277, %r7827, %r7828; and.b32 %r278, %r277, 15; cvt.u16.u32 %rs169, %r277; or.b16 %rs170, %rs169, 2; and.b16 %rs171, %rs170, 15; setp.eq.s16 %p1473, %rs171, 14; mov.f32 %f5677, %f414; @%p1473 bra $L__BB74_291; mov.f32 %f5677, %f433; $L__BB74_291: setp.lt.u32 %p1474, %r278, 8; selp.f32 %f4134, %f414, %f428, %p1474; setp.lt.u32 %p1475, %r278, 4; selp.f32 %f4135, %f428, %f5677, %p1475; and.b32 %r7829, %r277, 1; setp.eq.b32 %p1476, %r7829, 1; neg.ftz.f32 %f4136, %f4134; selp.f32 %f4137, %f4136, %f4134, %p1476; and.b32 %r7830, %r277, 2; setp.eq.s32 %p1477, %r7830, 0; neg.ftz.f32 %f4138, %f4135; selp.f32 %f4139, %f4135, %f4138, %p1477; add.ftz.f32 %f439, %f4137, %f4139; xor.b32 %r7831, %r275, %r257; sub.s32 %r7832, %r7831, %r276; xor.b32 %r7833, %r7832, %r260; shf.l.wrap.b32 %r7834, %r7832, %r7832, 25; sub.s32 %r7835, %r7833, %r7834; xor.b32 %r7836, %r7835, %r275; shf.l.wrap.b32 %r7837, %r7835, %r7835, 16; sub.s32 %r7838, %r7836, %r7837; xor.b32 %r7839, %r7838, %r7832; shf.l.wrap.b32 %r7840, %r7838, %r7838, 4; sub.s32 %r7841, %r7839, %r7840; xor.b32 %r7842, %r7841, %r7835; shf.l.wrap.b32 %r7843, %r7841, %r7841, 14; sub.s32 %r7844, %r7842, %r7843; xor.b32 %r7845, %r7844, %r7838; shf.l.wrap.b32 %r7846, %r7844, %r7844, 24; sub.s32 %r279, %r7845, %r7846; and.b32 %r280, %r279, 15; cvt.u16.u32 %rs172, %r279; or.b16 %rs173, %rs172, 2; and.b16 %rs174, %rs173, 15; setp.eq.s16 %p1478, %rs174, 14; mov.f32 %f5678, %f425; @%p1478 bra $L__BB74_293; mov.f32 %f5678, %f433; $L__BB74_293: setp.lt.u32 %p1479, %r280, 8; selp.f32 %f4141, %f425, %f428, %p1479; setp.lt.u32 %p1480, %r280, 4; selp.f32 %f4142, %f428, %f5678, %p1480; and.b32 %r7847, %r279, 1; setp.eq.b32 %p1481, %r7847, 1; neg.ftz.f32 %f4143, %f4141; selp.f32 %f4144, %f4143, %f4141, %p1481; and.b32 %r7848, %r279, 2; setp.eq.s32 %p1482, %r7848, 0; neg.ftz.f32 %f4145, %f4142; selp.f32 %f4146, %f4142, %f4145, %p1482; add.ftz.f32 %f4147, %f4144, %f4146; mov.f32 %f4148, 0f3F800000; sub.ftz.f32 %f4149, %f4148, %f421; mul.ftz.f32 %f4150, %f421, %f427; fma.rn.ftz.f32 %f4151, %f4149, %f424, %f4150; sub.ftz.f32 %f4152, %f4148, %f422; mul.ftz.f32 %f4153, %f421, %f432; fma.rn.ftz.f32 %f4154, %f4149, %f430, %f4153; mul.ftz.f32 %f4155, %f422, %f4154; fma.rn.ftz.f32 %f4156, %f4152, %f4151, %f4155; sub.ftz.f32 %f4157, %f4148, %f423; mul.ftz.f32 %f4158, %f421, %f437; fma.rn.ftz.f32 %f4159, %f4149, %f435, %f4158; mul.ftz.f32 %f4160, %f421, %f4147; fma.rn.ftz.f32 %f4161, %f4149, %f439, %f4160; mul.ftz.f32 %f4162, %f422, %f4161; fma.rn.ftz.f32 %f4163, %f4152, %f4159, %f4162; mul.ftz.f32 %f4164, %f423, %f4163; fma.rn.ftz.f32 %f441, %f4157, %f4156, %f4164; mov.b32 %r281, %f441; setp.nan.ftz.f32 %p1483, %f441, %f441; mov.f32 %f5679, 0f00000000; @%p1483 bra $L__BB74_296; or.b32 %r7849, %r281, -2147483648; setp.ne.s32 %p1484, %r7849, -2147483648; add.ftz.f32 %f4166, %f441, %f441; setp.eq.ftz.f32 %p1485, %f4166, %f441; and.pred %p1486, %p1485, %p1484; @%p1486 bra $L__BB74_296; shl.b32 %r7850, %r281, 1; setp.lt.u32 %p1487, %r7850, -16777215; mul.ftz.f32 %f4167, %f441, 0f3F7B645A; selp.f32 %f5679, %f4167, 0f00000000, %p1487; $L__BB74_296: add.ftz.f32 %f4168, %f21, %f5679; mul.ftz.f32 %f4169, %f5668, %f4168; setp.gt.ftz.f32 %p1488, %f5670, 0f3F800000; selp.f32 %f4170, 0f3F800000, %f5670, %p1488; mul.ftz.f32 %f4171, %f4170, %f413; fma.rn.ftz.f32 %f5827, %f4171, %f4169, %f5827; bra.uni $L__BB74_450; $L__BB74_213: lg2.approx.ftz.f32 %f3821, %f27; mul.ftz.f32 %f3822, %f25, %f3821; neg.ftz.f32 %f3823, %f3822; ex2.approx.ftz.f32 %f289, %f3823; cvt.rmi.ftz.f32.f32 %f3824, %f5694; cvt.rzi.ftz.s32.f32 %r144, %f3824; sub.ftz.f32 %f290, %f5694, %f3824; cvt.rmi.ftz.f32.f32 %f3825, %f5695; cvt.rzi.ftz.s32.f32 %r145, %f3825; sub.ftz.f32 %f291, %f5695, %f3825; cvt.rmi.ftz.f32.f32 %f3826, %f5696; cvt.rzi.ftz.s32.f32 %r146, %f3826; sub.ftz.f32 %f292, %f5696, %f3826; fma.rn.ftz.f32 %f3827, %f290, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f293, %f290, %f3827, 0f41200000; fma.rn.ftz.f32 %f3828, %f291, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f294, %f291, %f3828, 0f41200000; fma.rn.ftz.f32 %f3829, %f292, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f295, %f292, %f3829, 0f41200000; add.s32 %r7209, %r145, -559038712; add.s32 %r7210, %r146, -559038712; xor.b32 %r7211, %r7210, %r7209; shf.l.wrap.b32 %r147, %r7209, %r7209, 14; sub.s32 %r148, %r7211, %r147; add.s32 %r7212, %r144, -559038712; xor.b32 %r7213, %r148, %r7212; shf.l.wrap.b32 %r149, %r148, %r148, 11; sub.s32 %r7214, %r7213, %r149; xor.b32 %r7215, %r7214, %r7209; shf.l.wrap.b32 %r7216, %r7214, %r7214, 25; sub.s32 %r7217, %r7215, %r7216; xor.b32 %r7218, %r7217, %r148; shf.l.wrap.b32 %r7219, %r7217, %r7217, 16; sub.s32 %r7220, %r7218, %r7219; xor.b32 %r7221, %r7220, %r7214; shf.l.wrap.b32 %r7222, %r7220, %r7220, 4; sub.s32 %r7223, %r7221, %r7222; xor.b32 %r7224, %r7223, %r7217; shf.l.wrap.b32 %r7225, %r7223, %r7223, 14; sub.s32 %r7226, %r7224, %r7225; xor.b32 %r7227, %r7226, %r7220; shf.l.wrap.b32 %r7228, %r7226, %r7226, 24; sub.s32 %r150, %r7227, %r7228; and.b32 %r151, %r150, 15; cvt.u16.u32 %rs79, %r150; or.b16 %rs80, %rs79, 2; and.b16 %rs81, %rs80, 15; setp.eq.s16 %p1298, %rs81, 14; mov.f32 %f5626, %f290; @%p1298 bra $L__BB74_215; mov.f32 %f5626, %f292; $L__BB74_215: mul.ftz.f32 %f3830, %f290, %f290; mul.ftz.f32 %f3831, %f290, %f3830; mul.ftz.f32 %f297, %f3831, %f293; mul.ftz.f32 %f3832, %f291, %f291; mul.ftz.f32 %f3833, %f291, %f3832; mul.ftz.f32 %f298, %f3833, %f294; mul.ftz.f32 %f3834, %f292, %f292; mul.ftz.f32 %f3835, %f292, %f3834; mul.ftz.f32 %f299, %f3835, %f295; setp.lt.u32 %p1299, %r151, 8; selp.f32 %f3836, %f290, %f291, %p1299; setp.lt.u32 %p1300, %r151, 4; selp.f32 %f3837, %f291, %f5626, %p1300; and.b32 %r7229, %r150, 1; setp.eq.b32 %p1301, %r7229, 1; neg.ftz.f32 %f3838, %f3836; selp.f32 %f3839, %f3838, %f3836, %p1301; and.b32 %r7230, %r150, 2; setp.eq.s32 %p1302, %r7230, 0; neg.ftz.f32 %f3840, %f3837; selp.f32 %f3841, %f3837, %f3840, %p1302; add.ftz.f32 %f300, %f3839, %f3841; add.s32 %r152, %r144, -559038711; xor.b32 %r7231, %r148, %r152; sub.s32 %r7232, %r7231, %r149; xor.b32 %r7234, %r7232, %r7209; shf.l.wrap.b32 %r7235, %r7232, %r7232, 25; sub.s32 %r7236, %r7234, %r7235; xor.b32 %r7237, %r7236, %r148; shf.l.wrap.b32 %r7238, %r7236, %r7236, 16; sub.s32 %r7239, %r7237, %r7238; xor.b32 %r7240, %r7239, %r7232; shf.l.wrap.b32 %r7241, %r7239, %r7239, 4; sub.s32 %r7242, %r7240, %r7241; xor.b32 %r7243, %r7242, %r7236; shf.l.wrap.b32 %r7244, %r7242, %r7242, 14; sub.s32 %r7245, %r7243, %r7244; xor.b32 %r7246, %r7245, %r7239; shf.l.wrap.b32 %r7247, %r7245, %r7245, 24; sub.s32 %r153, %r7246, %r7247; and.b32 %r154, %r153, 15; add.ftz.f32 %f301, %f290, 0fBF800000; cvt.u16.u32 %rs82, %r153; or.b16 %rs83, %rs82, 2; and.b16 %rs84, %rs83, 15; setp.eq.s16 %p1303, %rs84, 14; mov.f32 %f5627, %f301; @%p1303 bra $L__BB74_217; mov.f32 %f5627, %f292; $L__BB74_217: setp.lt.u32 %p1304, %r154, 8; selp.f32 %f3842, %f301, %f291, %p1304; setp.lt.u32 %p1305, %r154, 4; selp.f32 %f3843, %f291, %f5627, %p1305; and.b32 %r7248, %r153, 1; setp.eq.b32 %p1306, %r7248, 1; neg.ftz.f32 %f3844, %f3842; selp.f32 %f3845, %f3844, %f3842, %p1306; and.b32 %r7249, %r153, 2; setp.eq.s32 %p1307, %r7249, 0; neg.ftz.f32 %f3846, %f3843; selp.f32 %f3847, %f3843, %f3846, %p1307; add.ftz.f32 %f303, %f3845, %f3847; add.s32 %r155, %r145, -559038711; xor.b32 %r7251, %r7210, %r155; shf.l.wrap.b32 %r156, %r155, %r155, 14; sub.s32 %r157, %r7251, %r156; xor.b32 %r7253, %r157, %r7212; shf.l.wrap.b32 %r158, %r157, %r157, 11; sub.s32 %r7254, %r7253, %r158; xor.b32 %r7255, %r7254, %r155; shf.l.wrap.b32 %r7256, %r7254, %r7254, 25; sub.s32 %r7257, %r7255, %r7256; xor.b32 %r7258, %r7257, %r157; shf.l.wrap.b32 %r7259, %r7257, %r7257, 16; sub.s32 %r7260, %r7258, %r7259; xor.b32 %r7261, %r7260, %r7254; shf.l.wrap.b32 %r7262, %r7260, %r7260, 4; sub.s32 %r7263, %r7261, %r7262; xor.b32 %r7264, %r7263, %r7257; shf.l.wrap.b32 %r7265, %r7263, %r7263, 14; sub.s32 %r7266, %r7264, %r7265; xor.b32 %r7267, %r7266, %r7260; shf.l.wrap.b32 %r7268, %r7266, %r7266, 24; sub.s32 %r159, %r7267, %r7268; and.b32 %r160, %r159, 15; add.ftz.f32 %f304, %f291, 0fBF800000; cvt.u16.u32 %rs85, %r159; or.b16 %rs86, %rs85, 2; and.b16 %rs87, %rs86, 15; setp.eq.s16 %p1308, %rs87, 14; mov.f32 %f5628, %f290; @%p1308 bra $L__BB74_219; mov.f32 %f5628, %f292; $L__BB74_219: setp.lt.u32 %p1309, %r160, 8; selp.f32 %f3848, %f290, %f304, %p1309; setp.lt.u32 %p1310, %r160, 4; selp.f32 %f3849, %f304, %f5628, %p1310; and.b32 %r7269, %r159, 1; setp.eq.b32 %p1311, %r7269, 1; neg.ftz.f32 %f3850, %f3848; selp.f32 %f3851, %f3850, %f3848, %p1311; and.b32 %r7270, %r159, 2; setp.eq.s32 %p1312, %r7270, 0; neg.ftz.f32 %f3852, %f3849; selp.f32 %f3853, %f3849, %f3852, %p1312; add.ftz.f32 %f306, %f3851, %f3853; xor.b32 %r7271, %r157, %r152; sub.s32 %r7272, %r7271, %r158; xor.b32 %r7273, %r7272, %r155; shf.l.wrap.b32 %r7274, %r7272, %r7272, 25; sub.s32 %r7275, %r7273, %r7274; xor.b32 %r7276, %r7275, %r157; shf.l.wrap.b32 %r7277, %r7275, %r7275, 16; sub.s32 %r7278, %r7276, %r7277; xor.b32 %r7279, %r7278, %r7272; shf.l.wrap.b32 %r7280, %r7278, %r7278, 4; sub.s32 %r7281, %r7279, %r7280; xor.b32 %r7282, %r7281, %r7275; shf.l.wrap.b32 %r7283, %r7281, %r7281, 14; sub.s32 %r7284, %r7282, %r7283; xor.b32 %r7285, %r7284, %r7278; shf.l.wrap.b32 %r7286, %r7284, %r7284, 24; sub.s32 %r161, %r7285, %r7286; and.b32 %r162, %r161, 15; cvt.u16.u32 %rs88, %r161; or.b16 %rs89, %rs88, 2; and.b16 %rs90, %rs89, 15; setp.eq.s16 %p1313, %rs90, 14; mov.f32 %f5629, %f301; @%p1313 bra $L__BB74_221; mov.f32 %f5629, %f292; $L__BB74_221: setp.lt.u32 %p1314, %r162, 8; selp.f32 %f3854, %f301, %f304, %p1314; setp.lt.u32 %p1315, %r162, 4; selp.f32 %f3855, %f304, %f5629, %p1315; and.b32 %r7287, %r161, 1; setp.eq.b32 %p1316, %r7287, 1; neg.ftz.f32 %f3856, %f3854; selp.f32 %f3857, %f3856, %f3854, %p1316; and.b32 %r7288, %r161, 2; setp.eq.s32 %p1317, %r7288, 0; neg.ftz.f32 %f3858, %f3855; selp.f32 %f3859, %f3855, %f3858, %p1317; add.ftz.f32 %f308, %f3857, %f3859; add.s32 %r163, %r146, -559038711; xor.b32 %r7290, %r163, %r7209; sub.s32 %r164, %r7290, %r147; xor.b32 %r7292, %r164, %r7212; shf.l.wrap.b32 %r165, %r164, %r164, 11; sub.s32 %r7293, %r7292, %r165; xor.b32 %r7294, %r7293, %r7209; shf.l.wrap.b32 %r7295, %r7293, %r7293, 25; sub.s32 %r7296, %r7294, %r7295; xor.b32 %r7297, %r7296, %r164; shf.l.wrap.b32 %r7298, %r7296, %r7296, 16; sub.s32 %r7299, %r7297, %r7298; xor.b32 %r7300, %r7299, %r7293; shf.l.wrap.b32 %r7301, %r7299, %r7299, 4; sub.s32 %r7302, %r7300, %r7301; xor.b32 %r7303, %r7302, %r7296; shf.l.wrap.b32 %r7304, %r7302, %r7302, 14; sub.s32 %r7305, %r7303, %r7304; xor.b32 %r7306, %r7305, %r7299; shf.l.wrap.b32 %r7307, %r7305, %r7305, 24; sub.s32 %r166, %r7306, %r7307; add.ftz.f32 %f309, %f292, 0fBF800000; and.b32 %r167, %r166, 15; cvt.u16.u32 %rs91, %r166; or.b16 %rs92, %rs91, 2; and.b16 %rs93, %rs92, 15; setp.eq.s16 %p1318, %rs93, 14; mov.f32 %f5630, %f290; @%p1318 bra $L__BB74_223; mov.f32 %f5630, %f309; $L__BB74_223: setp.lt.u32 %p1319, %r167, 8; selp.f32 %f3860, %f290, %f291, %p1319; setp.lt.u32 %p1320, %r167, 4; selp.f32 %f3861, %f291, %f5630, %p1320; and.b32 %r7308, %r166, 1; setp.eq.b32 %p1321, %r7308, 1; neg.ftz.f32 %f3862, %f3860; selp.f32 %f3863, %f3862, %f3860, %p1321; and.b32 %r7309, %r166, 2; setp.eq.s32 %p1322, %r7309, 0; neg.ftz.f32 %f3864, %f3861; selp.f32 %f3865, %f3861, %f3864, %p1322; add.ftz.f32 %f311, %f3863, %f3865; xor.b32 %r7310, %r164, %r152; sub.s32 %r7311, %r7310, %r165; xor.b32 %r7313, %r7311, %r7209; shf.l.wrap.b32 %r7314, %r7311, %r7311, 25; sub.s32 %r7315, %r7313, %r7314; xor.b32 %r7316, %r7315, %r164; shf.l.wrap.b32 %r7317, %r7315, %r7315, 16; sub.s32 %r7318, %r7316, %r7317; xor.b32 %r7319, %r7318, %r7311; shf.l.wrap.b32 %r7320, %r7318, %r7318, 4; sub.s32 %r7321, %r7319, %r7320; xor.b32 %r7322, %r7321, %r7315; shf.l.wrap.b32 %r7323, %r7321, %r7321, 14; sub.s32 %r7324, %r7322, %r7323; xor.b32 %r7325, %r7324, %r7318; shf.l.wrap.b32 %r7326, %r7324, %r7324, 24; sub.s32 %r168, %r7325, %r7326; and.b32 %r169, %r168, 15; cvt.u16.u32 %rs94, %r168; or.b16 %rs95, %rs94, 2; and.b16 %rs96, %rs95, 15; setp.eq.s16 %p1323, %rs96, 14; mov.f32 %f5631, %f301; @%p1323 bra $L__BB74_225; mov.f32 %f5631, %f309; $L__BB74_225: setp.lt.u32 %p1324, %r169, 8; selp.f32 %f3866, %f301, %f291, %p1324; setp.lt.u32 %p1325, %r169, 4; selp.f32 %f3867, %f291, %f5631, %p1325; and.b32 %r7327, %r168, 1; setp.eq.b32 %p1326, %r7327, 1; neg.ftz.f32 %f3868, %f3866; selp.f32 %f3869, %f3868, %f3866, %p1326; and.b32 %r7328, %r168, 2; setp.eq.s32 %p1327, %r7328, 0; neg.ftz.f32 %f3870, %f3867; selp.f32 %f3871, %f3867, %f3870, %p1327; add.ftz.f32 %f313, %f3869, %f3871; xor.b32 %r7329, %r163, %r155; sub.s32 %r170, %r7329, %r156; xor.b32 %r7331, %r170, %r7212; shf.l.wrap.b32 %r171, %r170, %r170, 11; sub.s32 %r7332, %r7331, %r171; xor.b32 %r7333, %r7332, %r155; shf.l.wrap.b32 %r7334, %r7332, %r7332, 25; sub.s32 %r7335, %r7333, %r7334; xor.b32 %r7336, %r7335, %r170; shf.l.wrap.b32 %r7337, %r7335, %r7335, 16; sub.s32 %r7338, %r7336, %r7337; xor.b32 %r7339, %r7338, %r7332; shf.l.wrap.b32 %r7340, %r7338, %r7338, 4; sub.s32 %r7341, %r7339, %r7340; xor.b32 %r7342, %r7341, %r7335; shf.l.wrap.b32 %r7343, %r7341, %r7341, 14; sub.s32 %r7344, %r7342, %r7343; xor.b32 %r7345, %r7344, %r7338; shf.l.wrap.b32 %r7346, %r7344, %r7344, 24; sub.s32 %r172, %r7345, %r7346; and.b32 %r173, %r172, 15; cvt.u16.u32 %rs97, %r172; or.b16 %rs98, %rs97, 2; and.b16 %rs99, %rs98, 15; setp.eq.s16 %p1328, %rs99, 14; mov.f32 %f5632, %f290; @%p1328 bra $L__BB74_227; mov.f32 %f5632, %f309; $L__BB74_227: setp.lt.u32 %p1329, %r173, 8; selp.f32 %f3872, %f290, %f304, %p1329; setp.lt.u32 %p1330, %r173, 4; selp.f32 %f3873, %f304, %f5632, %p1330; and.b32 %r7347, %r172, 1; setp.eq.b32 %p1331, %r7347, 1; neg.ftz.f32 %f3874, %f3872; selp.f32 %f3875, %f3874, %f3872, %p1331; and.b32 %r7348, %r172, 2; setp.eq.s32 %p1332, %r7348, 0; neg.ftz.f32 %f3876, %f3873; selp.f32 %f3877, %f3873, %f3876, %p1332; add.ftz.f32 %f315, %f3875, %f3877; xor.b32 %r7349, %r170, %r152; sub.s32 %r7350, %r7349, %r171; xor.b32 %r7351, %r7350, %r155; shf.l.wrap.b32 %r7352, %r7350, %r7350, 25; sub.s32 %r7353, %r7351, %r7352; xor.b32 %r7354, %r7353, %r170; shf.l.wrap.b32 %r7355, %r7353, %r7353, 16; sub.s32 %r7356, %r7354, %r7355; xor.b32 %r7357, %r7356, %r7350; shf.l.wrap.b32 %r7358, %r7356, %r7356, 4; sub.s32 %r7359, %r7357, %r7358; xor.b32 %r7360, %r7359, %r7353; shf.l.wrap.b32 %r7361, %r7359, %r7359, 14; sub.s32 %r7362, %r7360, %r7361; xor.b32 %r7363, %r7362, %r7356; shf.l.wrap.b32 %r7364, %r7362, %r7362, 24; sub.s32 %r174, %r7363, %r7364; and.b32 %r175, %r174, 15; cvt.u16.u32 %rs100, %r174; or.b16 %rs101, %rs100, 2; and.b16 %rs102, %rs101, 15; setp.eq.s16 %p1333, %rs102, 14; mov.f32 %f5633, %f301; @%p1333 bra $L__BB74_229; mov.f32 %f5633, %f309; $L__BB74_229: setp.lt.u32 %p1334, %r175, 8; selp.f32 %f3879, %f301, %f304, %p1334; setp.lt.u32 %p1335, %r175, 4; selp.f32 %f3880, %f304, %f5633, %p1335; and.b32 %r7365, %r174, 1; setp.eq.b32 %p1336, %r7365, 1; neg.ftz.f32 %f3881, %f3879; selp.f32 %f3882, %f3881, %f3879, %p1336; and.b32 %r7366, %r174, 2; setp.eq.s32 %p1337, %r7366, 0; neg.ftz.f32 %f3883, %f3880; selp.f32 %f3884, %f3880, %f3883, %p1337; add.ftz.f32 %f3885, %f3882, %f3884; mov.f32 %f3886, 0f3F800000; sub.ftz.f32 %f3887, %f3886, %f297; mul.ftz.f32 %f3888, %f297, %f303; fma.rn.ftz.f32 %f3889, %f3887, %f300, %f3888; sub.ftz.f32 %f3890, %f3886, %f298; mul.ftz.f32 %f3891, %f297, %f308; fma.rn.ftz.f32 %f3892, %f3887, %f306, %f3891; mul.ftz.f32 %f3893, %f298, %f3892; fma.rn.ftz.f32 %f3894, %f3890, %f3889, %f3893; sub.ftz.f32 %f3895, %f3886, %f299; mul.ftz.f32 %f3896, %f297, %f313; fma.rn.ftz.f32 %f3897, %f3887, %f311, %f3896; mul.ftz.f32 %f3898, %f297, %f3885; fma.rn.ftz.f32 %f3899, %f3887, %f315, %f3898; mul.ftz.f32 %f3900, %f298, %f3899; fma.rn.ftz.f32 %f3901, %f3890, %f3897, %f3900; mul.ftz.f32 %f3902, %f299, %f3901; fma.rn.ftz.f32 %f317, %f3895, %f3894, %f3902; mov.b32 %r176, %f317; setp.nan.ftz.f32 %p1338, %f317, %f317; mov.f32 %f5634, 0f00000000; @%p1338 bra $L__BB74_232; or.b32 %r7367, %r176, -2147483648; setp.ne.s32 %p1339, %r7367, -2147483648; add.ftz.f32 %f3904, %f317, %f317; setp.eq.ftz.f32 %p1340, %f3904, %f317; and.pred %p1341, %p1340, %p1339; @%p1341 bra $L__BB74_232; shl.b32 %r7368, %r176, 1; setp.lt.u32 %p1342, %r7368, -16777215; mul.ftz.f32 %f3905, %f317, 0f3F7B645A; selp.f32 %f5634, %f3905, 0f00000000, %p1342; $L__BB74_232: abs.ftz.f32 %f3906, %f5634; sub.ftz.f32 %f3907, %f21, %f3906; mul.ftz.f32 %f5827, %f3907, %f3907; cvt.rzi.ftz.s32.f32 %r177, %f26; setp.lt.s32 %p1343, %r177, 2; @%p1343 bra $L__BB74_450; mov.u32 %r10024, 1; mov.f32 %f5636, %f5827; mov.f32 %f5637, %f289; $L__BB74_234: mul.ftz.f32 %f5694, %f27, %f5694; cvt.rmi.ftz.f32.f32 %f3908, %f5694; cvt.rzi.ftz.s32.f32 %r179, %f3908; sub.ftz.f32 %f328, %f5694, %f3908; mul.ftz.f32 %f5695, %f27, %f5695; cvt.rmi.ftz.f32.f32 %f3909, %f5695; cvt.rzi.ftz.s32.f32 %r180, %f3909; sub.ftz.f32 %f330, %f5695, %f3909; mul.ftz.f32 %f5696, %f27, %f5696; cvt.rmi.ftz.f32.f32 %f3910, %f5696; cvt.rzi.ftz.s32.f32 %r181, %f3910; sub.ftz.f32 %f332, %f5696, %f3910; fma.rn.ftz.f32 %f3911, %f328, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f333, %f328, %f3911, 0f41200000; fma.rn.ftz.f32 %f3912, %f330, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f334, %f330, %f3912, 0f41200000; fma.rn.ftz.f32 %f3913, %f332, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f335, %f332, %f3913, 0f41200000; add.s32 %r7370, %r180, -559038712; add.s32 %r7371, %r181, -559038712; xor.b32 %r7372, %r7371, %r7370; shf.l.wrap.b32 %r182, %r7370, %r7370, 14; sub.s32 %r183, %r7372, %r182; add.s32 %r7373, %r179, -559038712; xor.b32 %r7374, %r183, %r7373; shf.l.wrap.b32 %r184, %r183, %r183, 11; sub.s32 %r7375, %r7374, %r184; xor.b32 %r7376, %r7375, %r7370; shf.l.wrap.b32 %r7377, %r7375, %r7375, 25; sub.s32 %r7378, %r7376, %r7377; xor.b32 %r7379, %r7378, %r183; shf.l.wrap.b32 %r7380, %r7378, %r7378, 16; sub.s32 %r7381, %r7379, %r7380; xor.b32 %r7382, %r7381, %r7375; shf.l.wrap.b32 %r7383, %r7381, %r7381, 4; sub.s32 %r7384, %r7382, %r7383; xor.b32 %r7385, %r7384, %r7378; shf.l.wrap.b32 %r7386, %r7384, %r7384, 14; sub.s32 %r7387, %r7385, %r7386; xor.b32 %r7388, %r7387, %r7381; shf.l.wrap.b32 %r7389, %r7387, %r7387, 24; sub.s32 %r185, %r7388, %r7389; and.b32 %r186, %r185, 15; cvt.u16.u32 %rs103, %r185; or.b16 %rs104, %rs103, 2; and.b16 %rs105, %rs104, 15; setp.eq.s16 %p1344, %rs105, 14; mov.f32 %f5641, %f328; @%p1344 bra $L__BB74_236; mov.f32 %f5641, %f332; $L__BB74_236: mul.ftz.f32 %f3914, %f328, %f328; mul.ftz.f32 %f3915, %f328, %f3914; mul.ftz.f32 %f337, %f3915, %f333; mul.ftz.f32 %f3916, %f330, %f330; mul.ftz.f32 %f3917, %f330, %f3916; mul.ftz.f32 %f338, %f3917, %f334; mul.ftz.f32 %f3918, %f332, %f332; mul.ftz.f32 %f3919, %f332, %f3918; mul.ftz.f32 %f339, %f3919, %f335; setp.lt.u32 %p1345, %r186, 8; selp.f32 %f3920, %f328, %f330, %p1345; setp.lt.u32 %p1346, %r186, 4; selp.f32 %f3921, %f330, %f5641, %p1346; and.b32 %r7390, %r185, 1; setp.eq.b32 %p1347, %r7390, 1; neg.ftz.f32 %f3922, %f3920; selp.f32 %f3923, %f3922, %f3920, %p1347; and.b32 %r7391, %r185, 2; setp.eq.s32 %p1348, %r7391, 0; neg.ftz.f32 %f3924, %f3921; selp.f32 %f3925, %f3921, %f3924, %p1348; add.ftz.f32 %f340, %f3923, %f3925; add.s32 %r187, %r179, -559038711; xor.b32 %r7392, %r183, %r187; sub.s32 %r7393, %r7392, %r184; xor.b32 %r7395, %r7393, %r7370; shf.l.wrap.b32 %r7396, %r7393, %r7393, 25; sub.s32 %r7397, %r7395, %r7396; xor.b32 %r7398, %r7397, %r183; shf.l.wrap.b32 %r7399, %r7397, %r7397, 16; sub.s32 %r7400, %r7398, %r7399; xor.b32 %r7401, %r7400, %r7393; shf.l.wrap.b32 %r7402, %r7400, %r7400, 4; sub.s32 %r7403, %r7401, %r7402; xor.b32 %r7404, %r7403, %r7397; shf.l.wrap.b32 %r7405, %r7403, %r7403, 14; sub.s32 %r7406, %r7404, %r7405; xor.b32 %r7407, %r7406, %r7400; shf.l.wrap.b32 %r7408, %r7406, %r7406, 24; sub.s32 %r188, %r7407, %r7408; and.b32 %r189, %r188, 15; add.ftz.f32 %f341, %f328, 0fBF800000; cvt.u16.u32 %rs106, %r188; or.b16 %rs107, %rs106, 2; and.b16 %rs108, %rs107, 15; setp.eq.s16 %p1349, %rs108, 14; mov.f32 %f5642, %f341; @%p1349 bra $L__BB74_238; mov.f32 %f5642, %f332; $L__BB74_238: setp.lt.u32 %p1350, %r189, 8; selp.f32 %f3926, %f341, %f330, %p1350; setp.lt.u32 %p1351, %r189, 4; selp.f32 %f3927, %f330, %f5642, %p1351; and.b32 %r7409, %r188, 1; setp.eq.b32 %p1352, %r7409, 1; neg.ftz.f32 %f3928, %f3926; selp.f32 %f3929, %f3928, %f3926, %p1352; and.b32 %r7410, %r188, 2; setp.eq.s32 %p1353, %r7410, 0; neg.ftz.f32 %f3930, %f3927; selp.f32 %f3931, %f3927, %f3930, %p1353; add.ftz.f32 %f343, %f3929, %f3931; add.s32 %r190, %r180, -559038711; xor.b32 %r7412, %r7371, %r190; shf.l.wrap.b32 %r191, %r190, %r190, 14; sub.s32 %r192, %r7412, %r191; xor.b32 %r7414, %r192, %r7373; shf.l.wrap.b32 %r193, %r192, %r192, 11; sub.s32 %r7415, %r7414, %r193; xor.b32 %r7416, %r7415, %r190; shf.l.wrap.b32 %r7417, %r7415, %r7415, 25; sub.s32 %r7418, %r7416, %r7417; xor.b32 %r7419, %r7418, %r192; shf.l.wrap.b32 %r7420, %r7418, %r7418, 16; sub.s32 %r7421, %r7419, %r7420; xor.b32 %r7422, %r7421, %r7415; shf.l.wrap.b32 %r7423, %r7421, %r7421, 4; sub.s32 %r7424, %r7422, %r7423; xor.b32 %r7425, %r7424, %r7418; shf.l.wrap.b32 %r7426, %r7424, %r7424, 14; sub.s32 %r7427, %r7425, %r7426; xor.b32 %r7428, %r7427, %r7421; shf.l.wrap.b32 %r7429, %r7427, %r7427, 24; sub.s32 %r194, %r7428, %r7429; and.b32 %r195, %r194, 15; add.ftz.f32 %f344, %f330, 0fBF800000; cvt.u16.u32 %rs109, %r194; or.b16 %rs110, %rs109, 2; and.b16 %rs111, %rs110, 15; setp.eq.s16 %p1354, %rs111, 14; mov.f32 %f5643, %f328; @%p1354 bra $L__BB74_240; mov.f32 %f5643, %f332; $L__BB74_240: setp.lt.u32 %p1355, %r195, 8; selp.f32 %f3932, %f328, %f344, %p1355; setp.lt.u32 %p1356, %r195, 4; selp.f32 %f3933, %f344, %f5643, %p1356; and.b32 %r7430, %r194, 1; setp.eq.b32 %p1357, %r7430, 1; neg.ftz.f32 %f3934, %f3932; selp.f32 %f3935, %f3934, %f3932, %p1357; and.b32 %r7431, %r194, 2; setp.eq.s32 %p1358, %r7431, 0; neg.ftz.f32 %f3936, %f3933; selp.f32 %f3937, %f3933, %f3936, %p1358; add.ftz.f32 %f346, %f3935, %f3937; xor.b32 %r7432, %r192, %r187; sub.s32 %r7433, %r7432, %r193; xor.b32 %r7434, %r7433, %r190; shf.l.wrap.b32 %r7435, %r7433, %r7433, 25; sub.s32 %r7436, %r7434, %r7435; xor.b32 %r7437, %r7436, %r192; shf.l.wrap.b32 %r7438, %r7436, %r7436, 16; sub.s32 %r7439, %r7437, %r7438; xor.b32 %r7440, %r7439, %r7433; shf.l.wrap.b32 %r7441, %r7439, %r7439, 4; sub.s32 %r7442, %r7440, %r7441; xor.b32 %r7443, %r7442, %r7436; shf.l.wrap.b32 %r7444, %r7442, %r7442, 14; sub.s32 %r7445, %r7443, %r7444; xor.b32 %r7446, %r7445, %r7439; shf.l.wrap.b32 %r7447, %r7445, %r7445, 24; sub.s32 %r196, %r7446, %r7447; and.b32 %r197, %r196, 15; cvt.u16.u32 %rs112, %r196; or.b16 %rs113, %rs112, 2; and.b16 %rs114, %rs113, 15; setp.eq.s16 %p1359, %rs114, 14; mov.f32 %f5644, %f341; @%p1359 bra $L__BB74_242; mov.f32 %f5644, %f332; $L__BB74_242: setp.lt.u32 %p1360, %r197, 8; selp.f32 %f3938, %f341, %f344, %p1360; setp.lt.u32 %p1361, %r197, 4; selp.f32 %f3939, %f344, %f5644, %p1361; and.b32 %r7448, %r196, 1; setp.eq.b32 %p1362, %r7448, 1; neg.ftz.f32 %f3940, %f3938; selp.f32 %f3941, %f3940, %f3938, %p1362; and.b32 %r7449, %r196, 2; setp.eq.s32 %p1363, %r7449, 0; neg.ftz.f32 %f3942, %f3939; selp.f32 %f3943, %f3939, %f3942, %p1363; add.ftz.f32 %f348, %f3941, %f3943; add.s32 %r198, %r181, -559038711; xor.b32 %r7451, %r198, %r7370; sub.s32 %r199, %r7451, %r182; xor.b32 %r7453, %r199, %r7373; shf.l.wrap.b32 %r200, %r199, %r199, 11; sub.s32 %r7454, %r7453, %r200; xor.b32 %r7455, %r7454, %r7370; shf.l.wrap.b32 %r7456, %r7454, %r7454, 25; sub.s32 %r7457, %r7455, %r7456; xor.b32 %r7458, %r7457, %r199; shf.l.wrap.b32 %r7459, %r7457, %r7457, 16; sub.s32 %r7460, %r7458, %r7459; xor.b32 %r7461, %r7460, %r7454; shf.l.wrap.b32 %r7462, %r7460, %r7460, 4; sub.s32 %r7463, %r7461, %r7462; xor.b32 %r7464, %r7463, %r7457; shf.l.wrap.b32 %r7465, %r7463, %r7463, 14; sub.s32 %r7466, %r7464, %r7465; xor.b32 %r7467, %r7466, %r7460; shf.l.wrap.b32 %r7468, %r7466, %r7466, 24; sub.s32 %r201, %r7467, %r7468; add.ftz.f32 %f349, %f332, 0fBF800000; and.b32 %r202, %r201, 15; cvt.u16.u32 %rs115, %r201; or.b16 %rs116, %rs115, 2; and.b16 %rs117, %rs116, 15; setp.eq.s16 %p1364, %rs117, 14; mov.f32 %f5645, %f328; @%p1364 bra $L__BB74_244; mov.f32 %f5645, %f349; $L__BB74_244: setp.lt.u32 %p1365, %r202, 8; selp.f32 %f3944, %f328, %f330, %p1365; setp.lt.u32 %p1366, %r202, 4; selp.f32 %f3945, %f330, %f5645, %p1366; and.b32 %r7469, %r201, 1; setp.eq.b32 %p1367, %r7469, 1; neg.ftz.f32 %f3946, %f3944; selp.f32 %f3947, %f3946, %f3944, %p1367; and.b32 %r7470, %r201, 2; setp.eq.s32 %p1368, %r7470, 0; neg.ftz.f32 %f3948, %f3945; selp.f32 %f3949, %f3945, %f3948, %p1368; add.ftz.f32 %f351, %f3947, %f3949; xor.b32 %r7471, %r199, %r187; sub.s32 %r7472, %r7471, %r200; xor.b32 %r7474, %r7472, %r7370; shf.l.wrap.b32 %r7475, %r7472, %r7472, 25; sub.s32 %r7476, %r7474, %r7475; xor.b32 %r7477, %r7476, %r199; shf.l.wrap.b32 %r7478, %r7476, %r7476, 16; sub.s32 %r7479, %r7477, %r7478; xor.b32 %r7480, %r7479, %r7472; shf.l.wrap.b32 %r7481, %r7479, %r7479, 4; sub.s32 %r7482, %r7480, %r7481; xor.b32 %r7483, %r7482, %r7476; shf.l.wrap.b32 %r7484, %r7482, %r7482, 14; sub.s32 %r7485, %r7483, %r7484; xor.b32 %r7486, %r7485, %r7479; shf.l.wrap.b32 %r7487, %r7485, %r7485, 24; sub.s32 %r203, %r7486, %r7487; and.b32 %r204, %r203, 15; cvt.u16.u32 %rs118, %r203; or.b16 %rs119, %rs118, 2; and.b16 %rs120, %rs119, 15; setp.eq.s16 %p1369, %rs120, 14; mov.f32 %f5646, %f341; @%p1369 bra $L__BB74_246; mov.f32 %f5646, %f349; $L__BB74_246: setp.lt.u32 %p1370, %r204, 8; selp.f32 %f3950, %f341, %f330, %p1370; setp.lt.u32 %p1371, %r204, 4; selp.f32 %f3951, %f330, %f5646, %p1371; and.b32 %r7488, %r203, 1; setp.eq.b32 %p1372, %r7488, 1; neg.ftz.f32 %f3952, %f3950; selp.f32 %f3953, %f3952, %f3950, %p1372; and.b32 %r7489, %r203, 2; setp.eq.s32 %p1373, %r7489, 0; neg.ftz.f32 %f3954, %f3951; selp.f32 %f3955, %f3951, %f3954, %p1373; add.ftz.f32 %f353, %f3953, %f3955; xor.b32 %r7490, %r198, %r190; sub.s32 %r205, %r7490, %r191; xor.b32 %r7492, %r205, %r7373; shf.l.wrap.b32 %r206, %r205, %r205, 11; sub.s32 %r7493, %r7492, %r206; xor.b32 %r7494, %r7493, %r190; shf.l.wrap.b32 %r7495, %r7493, %r7493, 25; sub.s32 %r7496, %r7494, %r7495; xor.b32 %r7497, %r7496, %r205; shf.l.wrap.b32 %r7498, %r7496, %r7496, 16; sub.s32 %r7499, %r7497, %r7498; xor.b32 %r7500, %r7499, %r7493; shf.l.wrap.b32 %r7501, %r7499, %r7499, 4; sub.s32 %r7502, %r7500, %r7501; xor.b32 %r7503, %r7502, %r7496; shf.l.wrap.b32 %r7504, %r7502, %r7502, 14; sub.s32 %r7505, %r7503, %r7504; xor.b32 %r7506, %r7505, %r7499; shf.l.wrap.b32 %r7507, %r7505, %r7505, 24; sub.s32 %r207, %r7506, %r7507; and.b32 %r208, %r207, 15; cvt.u16.u32 %rs121, %r207; or.b16 %rs122, %rs121, 2; and.b16 %rs123, %rs122, 15; setp.eq.s16 %p1374, %rs123, 14; mov.f32 %f5647, %f328; @%p1374 bra $L__BB74_248; mov.f32 %f5647, %f349; $L__BB74_248: setp.lt.u32 %p1375, %r208, 8; selp.f32 %f3956, %f328, %f344, %p1375; setp.lt.u32 %p1376, %r208, 4; selp.f32 %f3957, %f344, %f5647, %p1376; and.b32 %r7508, %r207, 1; setp.eq.b32 %p1377, %r7508, 1; neg.ftz.f32 %f3958, %f3956; selp.f32 %f3959, %f3958, %f3956, %p1377; and.b32 %r7509, %r207, 2; setp.eq.s32 %p1378, %r7509, 0; neg.ftz.f32 %f3960, %f3957; selp.f32 %f3961, %f3957, %f3960, %p1378; add.ftz.f32 %f355, %f3959, %f3961; xor.b32 %r7510, %r205, %r187; sub.s32 %r7511, %r7510, %r206; xor.b32 %r7512, %r7511, %r190; shf.l.wrap.b32 %r7513, %r7511, %r7511, 25; sub.s32 %r7514, %r7512, %r7513; xor.b32 %r7515, %r7514, %r205; shf.l.wrap.b32 %r7516, %r7514, %r7514, 16; sub.s32 %r7517, %r7515, %r7516; xor.b32 %r7518, %r7517, %r7511; shf.l.wrap.b32 %r7519, %r7517, %r7517, 4; sub.s32 %r7520, %r7518, %r7519; xor.b32 %r7521, %r7520, %r7514; shf.l.wrap.b32 %r7522, %r7520, %r7520, 14; sub.s32 %r7523, %r7521, %r7522; xor.b32 %r7524, %r7523, %r7517; shf.l.wrap.b32 %r7525, %r7523, %r7523, 24; sub.s32 %r209, %r7524, %r7525; and.b32 %r210, %r209, 15; cvt.u16.u32 %rs124, %r209; or.b16 %rs125, %rs124, 2; and.b16 %rs126, %rs125, 15; setp.eq.s16 %p1379, %rs126, 14; mov.f32 %f5648, %f341; @%p1379 bra $L__BB74_250; mov.f32 %f5648, %f349; $L__BB74_250: setp.lt.u32 %p1380, %r210, 8; selp.f32 %f3963, %f341, %f344, %p1380; setp.lt.u32 %p1381, %r210, 4; selp.f32 %f3964, %f344, %f5648, %p1381; and.b32 %r7526, %r209, 1; setp.eq.b32 %p1382, %r7526, 1; neg.ftz.f32 %f3965, %f3963; selp.f32 %f3966, %f3965, %f3963, %p1382; and.b32 %r7527, %r209, 2; setp.eq.s32 %p1383, %r7527, 0; neg.ftz.f32 %f3967, %f3964; selp.f32 %f3968, %f3964, %f3967, %p1383; add.ftz.f32 %f3969, %f3966, %f3968; sub.ftz.f32 %f3971, %f3886, %f337; mul.ftz.f32 %f3972, %f337, %f343; fma.rn.ftz.f32 %f3973, %f3971, %f340, %f3972; sub.ftz.f32 %f3974, %f3886, %f338; mul.ftz.f32 %f3975, %f337, %f348; fma.rn.ftz.f32 %f3976, %f3971, %f346, %f3975; mul.ftz.f32 %f3977, %f338, %f3976; fma.rn.ftz.f32 %f3978, %f3974, %f3973, %f3977; sub.ftz.f32 %f3979, %f3886, %f339; mul.ftz.f32 %f3980, %f337, %f353; fma.rn.ftz.f32 %f3981, %f3971, %f351, %f3980; mul.ftz.f32 %f3982, %f337, %f3969; fma.rn.ftz.f32 %f3983, %f3971, %f355, %f3982; mul.ftz.f32 %f3984, %f338, %f3983; fma.rn.ftz.f32 %f3985, %f3974, %f3981, %f3984; mul.ftz.f32 %f3986, %f339, %f3985; fma.rn.ftz.f32 %f357, %f3979, %f3978, %f3986; mov.b32 %r211, %f357; setp.nan.ftz.f32 %p1384, %f357, %f357; mov.f32 %f5649, 0f00000000; @%p1384 bra $L__BB74_253; or.b32 %r7528, %r211, -2147483648; setp.ne.s32 %p1385, %r7528, -2147483648; add.ftz.f32 %f3988, %f357, %f357; setp.eq.ftz.f32 %p1386, %f3988, %f357; and.pred %p1387, %p1386, %p1385; @%p1387 bra $L__BB74_253; shl.b32 %r7529, %r211, 1; setp.lt.u32 %p1388, %r7529, -16777215; selp.f32 %f5649, %f357, 0f00000000, %p1388; $L__BB74_253: mul.ftz.f32 %f3989, %f5649, 0f3F7B645A; abs.ftz.f32 %f3990, %f3989; sub.ftz.f32 %f3991, %f21, %f3990; mul.ftz.f32 %f3992, %f3991, %f3991; mul.ftz.f32 %f3993, %f24, %f5636; cvt.ftz.sat.f32.f32 %f3994, %f3993; mul.ftz.f32 %f5636, %f3994, %f3992; fma.rn.ftz.f32 %f5827, %f5637, %f5636, %f5827; mul.ftz.f32 %f5637, %f289, %f5637; add.s32 %r10024, %r10024, 1; setp.lt.s32 %p1389, %r10024, %r177; @%p1389 bra $L__BB74_234; bra.uni $L__BB74_450; $L__BB74_377: cvt.rzi.ftz.s32.f32 %r453, %f26; setp.lt.s32 %p1851, %r453, 1; mov.f32 %f5827, 0f3F800000; mov.f32 %f5784, %f5827; @%p1851 bra $L__BB74_383; lg2.approx.ftz.f32 %f5105, %f27; mul.ftz.f32 %f5106, %f25, %f5105; neg.ftz.f32 %f5107, %f5106; ex2.approx.ftz.f32 %f694, %f5107; mov.f32 %f5104, 0f3F800000; mov.u32 %r10032, 0; mov.f32 %f5784, %f5104; mov.f32 %f5827, %f5104; $L__BB74_379: cvt.rmi.ftz.f32.f32 %f5109, %f5771; cvt.rzi.ftz.s32.f32 %r9322, %f5109; sub.ftz.f32 %f5110, %f5771, %f5109; cvt.rmi.ftz.f32.f32 %f5111, %f5772; cvt.rzi.ftz.s32.f32 %r9323, %f5111; sub.ftz.f32 %f5112, %f5772, %f5111; mul.ftz.f32 %f5113, %f5110, %f5110; mul.ftz.f32 %f5114, %f5110, %f5113; fma.rn.ftz.f32 %f5115, %f5110, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5116, %f5110, %f5115, 0f41200000; mul.ftz.f32 %f5117, %f5114, %f5116; mul.ftz.f32 %f5118, %f5112, %f5112; mul.ftz.f32 %f5119, %f5112, %f5118; fma.rn.ftz.f32 %f5120, %f5112, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5121, %f5112, %f5120, 0f41200000; mul.ftz.f32 %f5122, %f5119, %f5121; add.s32 %r9324, %r9323, -559038716; add.s32 %r9325, %r9322, -559038716; xor.b32 %r9326, %r9324, -559038716; shf.l.wrap.b32 %r9327, %r9324, %r9324, 14; sub.s32 %r9328, %r9326, %r9327; xor.b32 %r9329, %r9328, %r9325; shf.l.wrap.b32 %r9330, %r9328, %r9328, 11; sub.s32 %r9331, %r9329, %r9330; xor.b32 %r9332, %r9331, %r9324; shf.l.wrap.b32 %r9333, %r9331, %r9331, 25; sub.s32 %r9334, %r9332, %r9333; xor.b32 %r9335, %r9334, %r9328; shf.l.wrap.b32 %r9336, %r9334, %r9334, 16; sub.s32 %r9337, %r9335, %r9336; xor.b32 %r9338, %r9337, %r9331; shf.l.wrap.b32 %r9339, %r9337, %r9337, 4; sub.s32 %r9340, %r9338, %r9339; xor.b32 %r9341, %r9340, %r9334; shf.l.wrap.b32 %r9342, %r9340, %r9340, 14; sub.s32 %r9343, %r9341, %r9342; xor.b32 %r9344, %r9343, %r9337; shf.l.wrap.b32 %r9345, %r9343, %r9343, 24; sub.s32 %r9346, %r9344, %r9345; and.b32 %r9347, %r9346, 4; setp.eq.s32 %p1852, %r9347, 0; selp.f32 %f5123, %f5110, %f5112, %p1852; selp.f32 %f5124, %f5112, %f5110, %p1852; add.ftz.f32 %f5125, %f5124, %f5124; and.b32 %r9348, %r9346, 1; setp.eq.b32 %p1853, %r9348, 1; neg.ftz.f32 %f5126, %f5123; selp.f32 %f5127, %f5126, %f5123, %p1853; and.b32 %r9349, %r9346, 2; setp.eq.s32 %p1854, %r9349, 0; neg.ftz.f32 %f5128, %f5125; selp.f32 %f5129, %f5125, %f5128, %p1854; add.ftz.f32 %f5130, %f5127, %f5129; add.s32 %r9350, %r9322, -559038715; xor.b32 %r9351, %r9328, %r9350; sub.s32 %r9352, %r9351, %r9330; xor.b32 %r9353, %r9352, %r9324; shf.l.wrap.b32 %r9354, %r9352, %r9352, 25; sub.s32 %r9355, %r9353, %r9354; xor.b32 %r9356, %r9355, %r9328; shf.l.wrap.b32 %r9357, %r9355, %r9355, 16; sub.s32 %r9358, %r9356, %r9357; xor.b32 %r9359, %r9358, %r9352; shf.l.wrap.b32 %r9360, %r9358, %r9358, 4; sub.s32 %r9361, %r9359, %r9360; xor.b32 %r9362, %r9361, %r9355; shf.l.wrap.b32 %r9363, %r9361, %r9361, 14; sub.s32 %r9364, %r9362, %r9363; xor.b32 %r9365, %r9364, %r9358; shf.l.wrap.b32 %r9366, %r9364, %r9364, 24; sub.s32 %r9367, %r9365, %r9366; add.ftz.f32 %f5131, %f5110, 0fBF800000; and.b32 %r9368, %r9367, 4; setp.eq.s32 %p1855, %r9368, 0; selp.f32 %f5133, %f5131, %f5112, %p1855; selp.f32 %f5134, %f5112, %f5131, %p1855; add.ftz.f32 %f5135, %f5134, %f5134; and.b32 %r9369, %r9367, 1; setp.eq.b32 %p1856, %r9369, 1; neg.ftz.f32 %f5136, %f5133; selp.f32 %f5137, %f5136, %f5133, %p1856; and.b32 %r9370, %r9367, 2; setp.eq.s32 %p1857, %r9370, 0; neg.ftz.f32 %f5138, %f5135; selp.f32 %f5139, %f5135, %f5138, %p1857; add.ftz.f32 %f5140, %f5137, %f5139; add.s32 %r9371, %r9323, -559038715; xor.b32 %r9372, %r9371, -559038716; shf.l.wrap.b32 %r9373, %r9371, %r9371, 14; sub.s32 %r9374, %r9372, %r9373; xor.b32 %r9375, %r9374, %r9325; shf.l.wrap.b32 %r9376, %r9374, %r9374, 11; sub.s32 %r9377, %r9375, %r9376; xor.b32 %r9378, %r9377, %r9371; shf.l.wrap.b32 %r9379, %r9377, %r9377, 25; sub.s32 %r9380, %r9378, %r9379; xor.b32 %r9381, %r9380, %r9374; shf.l.wrap.b32 %r9382, %r9380, %r9380, 16; sub.s32 %r9383, %r9381, %r9382; xor.b32 %r9384, %r9383, %r9377; shf.l.wrap.b32 %r9385, %r9383, %r9383, 4; sub.s32 %r9386, %r9384, %r9385; xor.b32 %r9387, %r9386, %r9380; shf.l.wrap.b32 %r9388, %r9386, %r9386, 14; sub.s32 %r9389, %r9387, %r9388; xor.b32 %r9390, %r9389, %r9383; shf.l.wrap.b32 %r9391, %r9389, %r9389, 24; sub.s32 %r9392, %r9390, %r9391; add.ftz.f32 %f5141, %f5112, 0fBF800000; and.b32 %r9393, %r9392, 4; setp.eq.s32 %p1858, %r9393, 0; selp.f32 %f5142, %f5110, %f5141, %p1858; selp.f32 %f5143, %f5141, %f5110, %p1858; add.ftz.f32 %f5144, %f5143, %f5143; and.b32 %r9394, %r9392, 1; setp.eq.b32 %p1859, %r9394, 1; neg.ftz.f32 %f5145, %f5142; selp.f32 %f5146, %f5145, %f5142, %p1859; and.b32 %r9395, %r9392, 2; setp.eq.s32 %p1860, %r9395, 0; neg.ftz.f32 %f5147, %f5144; selp.f32 %f5148, %f5144, %f5147, %p1860; add.ftz.f32 %f5149, %f5146, %f5148; xor.b32 %r9396, %r9374, %r9350; sub.s32 %r9397, %r9396, %r9376; xor.b32 %r9398, %r9397, %r9371; shf.l.wrap.b32 %r9399, %r9397, %r9397, 25; sub.s32 %r9400, %r9398, %r9399; xor.b32 %r9401, %r9400, %r9374; shf.l.wrap.b32 %r9402, %r9400, %r9400, 16; sub.s32 %r9403, %r9401, %r9402; xor.b32 %r9404, %r9403, %r9397; shf.l.wrap.b32 %r9405, %r9403, %r9403, 4; sub.s32 %r9406, %r9404, %r9405; xor.b32 %r9407, %r9406, %r9400; shf.l.wrap.b32 %r9408, %r9406, %r9406, 14; sub.s32 %r9409, %r9407, %r9408; xor.b32 %r9410, %r9409, %r9403; shf.l.wrap.b32 %r9411, %r9409, %r9409, 24; sub.s32 %r9412, %r9410, %r9411; and.b32 %r9413, %r9412, 4; setp.eq.s32 %p1861, %r9413, 0; selp.f32 %f5150, %f5131, %f5141, %p1861; selp.f32 %f5151, %f5141, %f5131, %p1861; add.ftz.f32 %f5152, %f5151, %f5151; and.b32 %r9414, %r9412, 1; setp.eq.b32 %p1862, %r9414, 1; neg.ftz.f32 %f5153, %f5150; selp.f32 %f5154, %f5153, %f5150, %p1862; and.b32 %r9415, %r9412, 2; setp.eq.s32 %p1863, %r9415, 0; neg.ftz.f32 %f5155, %f5152; selp.f32 %f5156, %f5152, %f5155, %p1863; add.ftz.f32 %f5157, %f5154, %f5156; sub.ftz.f32 %f5158, %f5104, %f5117; sub.ftz.f32 %f5159, %f5104, %f5122; mul.ftz.f32 %f5160, %f5158, %f5130; fma.rn.ftz.f32 %f5161, %f5117, %f5140, %f5160; mul.ftz.f32 %f5162, %f5158, %f5149; fma.rn.ftz.f32 %f5163, %f5117, %f5157, %f5162; mul.ftz.f32 %f5164, %f5122, %f5163; fma.rn.ftz.f32 %f699, %f5159, %f5161, %f5164; mov.b32 %r455, %f699; setp.nan.ftz.f32 %p1864, %f699, %f699; mov.f32 %f5780, 0f00000000; @%p1864 bra $L__BB74_382; or.b32 %r9416, %r455, -2147483648; setp.ne.s32 %p1865, %r9416, -2147483648; add.ftz.f32 %f5166, %f699, %f699; setp.eq.ftz.f32 %p1866, %f5166, %f699; and.pred %p1867, %p1866, %p1865; @%p1867 bra $L__BB74_382; shl.b32 %r9417, %r455, 1; setp.lt.u32 %p1868, %r9417, -16777215; selp.f32 %f5780, %f699, 0f00000000, %p1868; $L__BB74_382: mul.ftz.f32 %f5167, %f5780, 0f3F295E9E; fma.rn.ftz.f32 %f5168, %f5784, %f5167, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f5168; mul.ftz.f32 %f5784, %f694, %f5784; mul.ftz.f32 %f5771, %f27, %f5771; mul.ftz.f32 %f5772, %f27, %f5772; add.s32 %r10032, %r10032, 1; setp.lt.s32 %p1869, %r10032, %r453; @%p1869 bra $L__BB74_379; $L__BB74_383: cvt.rmi.ftz.f32.f32 %f5169, %f26; sub.ftz.f32 %f710, %f26, %f5169; setp.eq.ftz.f32 %p1870, %f710, 0f00000000; @%p1870 bra $L__BB74_450; mul.ftz.f32 %f711, %f5784, %f710; cvt.rmi.ftz.f32.f32 %f5171, %f5771; cvt.rzi.ftz.s32.f32 %r9418, %f5171; sub.ftz.f32 %f5172, %f5771, %f5171; cvt.rmi.ftz.f32.f32 %f5173, %f5772; cvt.rzi.ftz.s32.f32 %r9419, %f5173; sub.ftz.f32 %f5174, %f5772, %f5173; mul.ftz.f32 %f5175, %f5172, %f5172; mul.ftz.f32 %f5176, %f5172, %f5175; fma.rn.ftz.f32 %f5177, %f5172, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5178, %f5172, %f5177, 0f41200000; mul.ftz.f32 %f5179, %f5176, %f5178; mul.ftz.f32 %f5180, %f5174, %f5174; mul.ftz.f32 %f5181, %f5174, %f5180; fma.rn.ftz.f32 %f5182, %f5174, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5183, %f5174, %f5182, 0f41200000; mul.ftz.f32 %f5184, %f5181, %f5183; add.s32 %r9420, %r9419, -559038716; add.s32 %r9421, %r9418, -559038716; xor.b32 %r9422, %r9420, -559038716; shf.l.wrap.b32 %r9423, %r9420, %r9420, 14; sub.s32 %r9424, %r9422, %r9423; xor.b32 %r9425, %r9424, %r9421; shf.l.wrap.b32 %r9426, %r9424, %r9424, 11; sub.s32 %r9427, %r9425, %r9426; xor.b32 %r9428, %r9427, %r9420; shf.l.wrap.b32 %r9429, %r9427, %r9427, 25; sub.s32 %r9430, %r9428, %r9429; xor.b32 %r9431, %r9430, %r9424; shf.l.wrap.b32 %r9432, %r9430, %r9430, 16; sub.s32 %r9433, %r9431, %r9432; xor.b32 %r9434, %r9433, %r9427; shf.l.wrap.b32 %r9435, %r9433, %r9433, 4; sub.s32 %r9436, %r9434, %r9435; xor.b32 %r9437, %r9436, %r9430; shf.l.wrap.b32 %r9438, %r9436, %r9436, 14; sub.s32 %r9439, %r9437, %r9438; xor.b32 %r9440, %r9439, %r9433; shf.l.wrap.b32 %r9441, %r9439, %r9439, 24; sub.s32 %r9442, %r9440, %r9441; and.b32 %r9443, %r9442, 4; setp.eq.s32 %p1871, %r9443, 0; selp.f32 %f5185, %f5172, %f5174, %p1871; selp.f32 %f5186, %f5174, %f5172, %p1871; add.ftz.f32 %f5187, %f5186, %f5186; and.b32 %r9444, %r9442, 1; setp.eq.b32 %p1872, %r9444, 1; neg.ftz.f32 %f5188, %f5185; selp.f32 %f5189, %f5188, %f5185, %p1872; and.b32 %r9445, %r9442, 2; setp.eq.s32 %p1873, %r9445, 0; neg.ftz.f32 %f5190, %f5187; selp.f32 %f5191, %f5187, %f5190, %p1873; add.ftz.f32 %f5192, %f5189, %f5191; add.s32 %r9446, %r9418, -559038715; xor.b32 %r9447, %r9424, %r9446; sub.s32 %r9448, %r9447, %r9426; xor.b32 %r9449, %r9448, %r9420; shf.l.wrap.b32 %r9450, %r9448, %r9448, 25; sub.s32 %r9451, %r9449, %r9450; xor.b32 %r9452, %r9451, %r9424; shf.l.wrap.b32 %r9453, %r9451, %r9451, 16; sub.s32 %r9454, %r9452, %r9453; xor.b32 %r9455, %r9454, %r9448; shf.l.wrap.b32 %r9456, %r9454, %r9454, 4; sub.s32 %r9457, %r9455, %r9456; xor.b32 %r9458, %r9457, %r9451; shf.l.wrap.b32 %r9459, %r9457, %r9457, 14; sub.s32 %r9460, %r9458, %r9459; xor.b32 %r9461, %r9460, %r9454; shf.l.wrap.b32 %r9462, %r9460, %r9460, 24; sub.s32 %r9463, %r9461, %r9462; add.ftz.f32 %f5193, %f5172, 0fBF800000; mov.f32 %f5194, 0f3F800000; and.b32 %r9464, %r9463, 4; setp.eq.s32 %p1874, %r9464, 0; selp.f32 %f5195, %f5193, %f5174, %p1874; selp.f32 %f5196, %f5174, %f5193, %p1874; add.ftz.f32 %f5197, %f5196, %f5196; and.b32 %r9465, %r9463, 1; setp.eq.b32 %p1875, %r9465, 1; neg.ftz.f32 %f5198, %f5195; selp.f32 %f5199, %f5198, %f5195, %p1875; and.b32 %r9466, %r9463, 2; setp.eq.s32 %p1876, %r9466, 0; neg.ftz.f32 %f5200, %f5197; selp.f32 %f5201, %f5197, %f5200, %p1876; add.ftz.f32 %f5202, %f5199, %f5201; add.s32 %r9467, %r9419, -559038715; xor.b32 %r9468, %r9467, -559038716; shf.l.wrap.b32 %r9469, %r9467, %r9467, 14; sub.s32 %r9470, %r9468, %r9469; xor.b32 %r9471, %r9470, %r9421; shf.l.wrap.b32 %r9472, %r9470, %r9470, 11; sub.s32 %r9473, %r9471, %r9472; xor.b32 %r9474, %r9473, %r9467; shf.l.wrap.b32 %r9475, %r9473, %r9473, 25; sub.s32 %r9476, %r9474, %r9475; xor.b32 %r9477, %r9476, %r9470; shf.l.wrap.b32 %r9478, %r9476, %r9476, 16; sub.s32 %r9479, %r9477, %r9478; xor.b32 %r9480, %r9479, %r9473; shf.l.wrap.b32 %r9481, %r9479, %r9479, 4; sub.s32 %r9482, %r9480, %r9481; xor.b32 %r9483, %r9482, %r9476; shf.l.wrap.b32 %r9484, %r9482, %r9482, 14; sub.s32 %r9485, %r9483, %r9484; xor.b32 %r9486, %r9485, %r9479; shf.l.wrap.b32 %r9487, %r9485, %r9485, 24; sub.s32 %r9488, %r9486, %r9487; add.ftz.f32 %f5203, %f5174, 0fBF800000; and.b32 %r9489, %r9488, 4; setp.eq.s32 %p1877, %r9489, 0; selp.f32 %f5204, %f5172, %f5203, %p1877; selp.f32 %f5205, %f5203, %f5172, %p1877; add.ftz.f32 %f5206, %f5205, %f5205; and.b32 %r9490, %r9488, 1; setp.eq.b32 %p1878, %r9490, 1; neg.ftz.f32 %f5207, %f5204; selp.f32 %f5208, %f5207, %f5204, %p1878; and.b32 %r9491, %r9488, 2; setp.eq.s32 %p1879, %r9491, 0; neg.ftz.f32 %f5209, %f5206; selp.f32 %f5210, %f5206, %f5209, %p1879; add.ftz.f32 %f5211, %f5208, %f5210; xor.b32 %r9492, %r9470, %r9446; sub.s32 %r9493, %r9492, %r9472; xor.b32 %r9494, %r9493, %r9467; shf.l.wrap.b32 %r9495, %r9493, %r9493, 25; sub.s32 %r9496, %r9494, %r9495; xor.b32 %r9497, %r9496, %r9470; shf.l.wrap.b32 %r9498, %r9496, %r9496, 16; sub.s32 %r9499, %r9497, %r9498; xor.b32 %r9500, %r9499, %r9493; shf.l.wrap.b32 %r9501, %r9499, %r9499, 4; sub.s32 %r9502, %r9500, %r9501; xor.b32 %r9503, %r9502, %r9496; shf.l.wrap.b32 %r9504, %r9502, %r9502, 14; sub.s32 %r9505, %r9503, %r9504; xor.b32 %r9506, %r9505, %r9499; shf.l.wrap.b32 %r9507, %r9505, %r9505, 24; sub.s32 %r9508, %r9506, %r9507; and.b32 %r9509, %r9508, 4; setp.eq.s32 %p1880, %r9509, 0; selp.f32 %f5212, %f5193, %f5203, %p1880; selp.f32 %f5213, %f5203, %f5193, %p1880; add.ftz.f32 %f5214, %f5213, %f5213; and.b32 %r9510, %r9508, 1; setp.eq.b32 %p1881, %r9510, 1; neg.ftz.f32 %f5215, %f5212; selp.f32 %f5216, %f5215, %f5212, %p1881; and.b32 %r9511, %r9508, 2; setp.eq.s32 %p1882, %r9511, 0; neg.ftz.f32 %f5217, %f5214; selp.f32 %f5218, %f5214, %f5217, %p1882; add.ftz.f32 %f5219, %f5216, %f5218; sub.ftz.f32 %f5220, %f5194, %f5179; sub.ftz.f32 %f5221, %f5194, %f5184; mul.ftz.f32 %f5222, %f5220, %f5192; fma.rn.ftz.f32 %f5223, %f5179, %f5202, %f5222; mul.ftz.f32 %f5224, %f5220, %f5211; fma.rn.ftz.f32 %f5225, %f5179, %f5219, %f5224; mul.ftz.f32 %f5226, %f5184, %f5225; fma.rn.ftz.f32 %f712, %f5221, %f5223, %f5226; mov.b32 %r457, %f712; setp.nan.ftz.f32 %p1883, %f712, %f712; mov.f32 %f5785, 0f00000000; @%p1883 bra $L__BB74_387; or.b32 %r9512, %r457, -2147483648; setp.ne.s32 %p1884, %r9512, -2147483648; add.ftz.f32 %f5228, %f712, %f712; setp.eq.ftz.f32 %p1885, %f5228, %f712; and.pred %p1886, %p1885, %p1884; @%p1886 bra $L__BB74_387; shl.b32 %r9513, %r457, 1; setp.lt.u32 %p1887, %r9513, -16777215; mul.ftz.f32 %f5229, %f712, 0f3F295E9E; selp.f32 %f5785, %f5229, 0f00000000, %p1887; $L__BB74_387: fma.rn.ftz.f32 %f5230, %f711, %f5785, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f5230; bra.uni $L__BB74_450; $L__BB74_91: cvt.rzi.ftz.s32.f32 %r37, %f26; setp.lt.s32 %p949, %r37, 1; mov.f32 %f5827, 0f3F800000; mov.f32 %f5587, %f5827; @%p949 bra $L__BB74_97; lg2.approx.ftz.f32 %f3070, %f27; mul.ftz.f32 %f3071, %f25, %f3070; neg.ftz.f32 %f3072, %f3071; ex2.approx.ftz.f32 %f146, %f3072; mov.f32 %f3069, 0f3F800000; mov.u32 %r10022, 0; mov.f32 %f5587, %f3069; mov.f32 %f5827, %f3069; $L__BB74_93: cvt.rmi.ftz.f32.f32 %f3074, %f5568; cvt.rzi.ftz.s32.f32 %r5596, %f3074; sub.ftz.f32 %f3075, %f5568, %f3074; cvt.rmi.ftz.f32.f32 %f3076, %f5569; cvt.rzi.ftz.s32.f32 %r5597, %f3076; sub.ftz.f32 %f3077, %f5569, %f3076; cvt.rmi.ftz.f32.f32 %f3078, %f5570; cvt.rzi.ftz.s32.f32 %r5598, %f3078; sub.ftz.f32 %f3079, %f5570, %f3078; cvt.rmi.ftz.f32.f32 %f3080, %f5571; cvt.rzi.ftz.s32.f32 %r5599, %f3080; sub.ftz.f32 %f3081, %f5571, %f3080; mul.ftz.f32 %f3082, %f3075, %f3075; mul.ftz.f32 %f3083, %f3075, %f3082; fma.rn.ftz.f32 %f3084, %f3075, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3085, %f3075, %f3084, 0f41200000; mul.ftz.f32 %f3086, %f3083, %f3085; mul.ftz.f32 %f3087, %f3077, %f3077; mul.ftz.f32 %f3088, %f3077, %f3087; fma.rn.ftz.f32 %f3089, %f3077, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3090, %f3077, %f3089, 0f41200000; mul.ftz.f32 %f3091, %f3088, %f3090; mul.ftz.f32 %f3092, %f3079, %f3079; mul.ftz.f32 %f3093, %f3079, %f3092; fma.rn.ftz.f32 %f3094, %f3079, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3095, %f3079, %f3094, 0f41200000; mul.ftz.f32 %f3096, %f3093, %f3095; mul.ftz.f32 %f3097, %f3081, %f3081; mul.ftz.f32 %f3098, %f3081, %f3097; fma.rn.ftz.f32 %f3099, %f3081, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3100, %f3081, %f3099, 0f41200000; mul.ftz.f32 %f3101, %f3098, %f3100; add.s32 %r5600, %r5597, -559038708; add.s32 %r5601, %r5598, -559038708; sub.s32 %r5602, %r5596, %r5598; shf.l.wrap.b32 %r5603, %r5601, %r5601, 4; xor.b32 %r5604, %r5603, %r5602; add.s32 %r5605, %r5601, %r5600; sub.s32 %r5606, %r5600, %r5604; shf.l.wrap.b32 %r5607, %r5604, %r5604, 6; xor.b32 %r5608, %r5607, %r5606; add.s32 %r5609, %r5604, %r5605; sub.s32 %r5610, %r5605, %r5608; shf.l.wrap.b32 %r5611, %r5608, %r5608, 8; xor.b32 %r5612, %r5611, %r5610; add.s32 %r5613, %r5608, %r5609; sub.s32 %r5614, %r5609, %r5612; shf.l.wrap.b32 %r5615, %r5612, %r5612, 16; xor.b32 %r5616, %r5615, %r5614; add.s32 %r5617, %r5612, %r5613; sub.s32 %r5618, %r5613, %r5616; shf.l.wrap.b32 %r5619, %r5616, %r5616, 19; xor.b32 %r5620, %r5619, %r5618; add.s32 %r5621, %r5616, %r5617; sub.s32 %r5622, %r5617, %r5620; shf.l.wrap.b32 %r5623, %r5620, %r5620, 4; add.s32 %r5624, %r5620, %r5621; add.s32 %r5625, %r5621, %r5599; xor.b32 %r5626, %r5624, %r5622; xor.b32 %r5627, %r5626, %r5623; shf.l.wrap.b32 %r5628, %r5624, %r5624, 14; sub.s32 %r5629, %r5627, %r5628; xor.b32 %r5630, %r5629, %r5625; shf.l.wrap.b32 %r5631, %r5629, %r5629, 11; sub.s32 %r5632, %r5630, %r5631; xor.b32 %r5633, %r5632, %r5624; shf.l.wrap.b32 %r5634, %r5632, %r5632, 25; sub.s32 %r5635, %r5633, %r5634; xor.b32 %r5636, %r5635, %r5629; shf.l.wrap.b32 %r5637, %r5635, %r5635, 16; sub.s32 %r5638, %r5636, %r5637; xor.b32 %r5639, %r5638, %r5632; shf.l.wrap.b32 %r5640, %r5638, %r5638, 4; sub.s32 %r5641, %r5639, %r5640; xor.b32 %r5642, %r5641, %r5635; shf.l.wrap.b32 %r5643, %r5641, %r5641, 14; sub.s32 %r5644, %r5642, %r5643; xor.b32 %r5645, %r5644, %r5638; shf.l.wrap.b32 %r5646, %r5644, %r5644, 24; sub.s32 %r5647, %r5645, %r5646; and.b32 %r5648, %r5647, 31; setp.lt.u32 %p950, %r5648, 24; selp.f32 %f3102, %f3075, %f3077, %p950; setp.lt.u32 %p951, %r5648, 16; selp.f32 %f3103, %f3077, %f3079, %p951; setp.lt.u32 %p952, %r5648, 8; selp.f32 %f3104, %f3079, %f3081, %p952; and.b32 %r5649, %r5647, 1; setp.eq.b32 %p953, %r5649, 1; neg.ftz.f32 %f3105, %f3102; selp.f32 %f3106, %f3105, %f3102, %p953; and.b32 %r5650, %r5647, 2; setp.eq.s32 %p954, %r5650, 0; neg.ftz.f32 %f3107, %f3103; selp.f32 %f3108, %f3103, %f3107, %p954; add.ftz.f32 %f3109, %f3106, %f3108; and.b32 %r5651, %r5647, 4; setp.eq.s32 %p955, %r5651, 0; neg.ftz.f32 %f3110, %f3104; selp.f32 %f3111, %f3104, %f3110, %p955; add.ftz.f32 %f3112, %f3111, %f3109; add.s32 %r5652, %r5596, 1; sub.s32 %r5653, %r5652, %r5598; xor.b32 %r5654, %r5603, %r5653; sub.s32 %r5655, %r5600, %r5654; shf.l.wrap.b32 %r5656, %r5654, %r5654, 6; xor.b32 %r5657, %r5656, %r5655; add.s32 %r5658, %r5654, %r5605; sub.s32 %r5659, %r5605, %r5657; shf.l.wrap.b32 %r5660, %r5657, %r5657, 8; xor.b32 %r5661, %r5660, %r5659; add.s32 %r5662, %r5657, %r5658; sub.s32 %r5663, %r5658, %r5661; shf.l.wrap.b32 %r5664, %r5661, %r5661, 16; xor.b32 %r5665, %r5664, %r5663; add.s32 %r5666, %r5661, %r5662; sub.s32 %r5667, %r5662, %r5665; shf.l.wrap.b32 %r5668, %r5665, %r5665, 19; xor.b32 %r5669, %r5668, %r5667; add.s32 %r5670, %r5665, %r5666; sub.s32 %r5671, %r5666, %r5669; shf.l.wrap.b32 %r5672, %r5669, %r5669, 4; add.s32 %r5673, %r5669, %r5670; add.s32 %r5674, %r5670, %r5599; xor.b32 %r5675, %r5673, %r5671; xor.b32 %r5676, %r5675, %r5672; shf.l.wrap.b32 %r5677, %r5673, %r5673, 14; sub.s32 %r5678, %r5676, %r5677; xor.b32 %r5679, %r5678, %r5674; shf.l.wrap.b32 %r5680, %r5678, %r5678, 11; sub.s32 %r5681, %r5679, %r5680; xor.b32 %r5682, %r5681, %r5673; shf.l.wrap.b32 %r5683, %r5681, %r5681, 25; sub.s32 %r5684, %r5682, %r5683; xor.b32 %r5685, %r5684, %r5678; shf.l.wrap.b32 %r5686, %r5684, %r5684, 16; sub.s32 %r5687, %r5685, %r5686; xor.b32 %r5688, %r5687, %r5681; shf.l.wrap.b32 %r5689, %r5687, %r5687, 4; sub.s32 %r5690, %r5688, %r5689; xor.b32 %r5691, %r5690, %r5684; shf.l.wrap.b32 %r5692, %r5690, %r5690, 14; sub.s32 %r5693, %r5691, %r5692; xor.b32 %r5694, %r5693, %r5687; shf.l.wrap.b32 %r5695, %r5693, %r5693, 24; sub.s32 %r5696, %r5694, %r5695; add.ftz.f32 %f3113, %f3075, 0fBF800000; and.b32 %r5697, %r5696, 31; setp.lt.u32 %p956, %r5697, 24; selp.f32 %f3115, %f3113, %f3077, %p956; setp.lt.u32 %p957, %r5697, 16; selp.f32 %f3116, %f3077, %f3079, %p957; setp.lt.u32 %p958, %r5697, 8; selp.f32 %f3117, %f3079, %f3081, %p958; and.b32 %r5698, %r5696, 1; setp.eq.b32 %p959, %r5698, 1; neg.ftz.f32 %f3118, %f3115; selp.f32 %f3119, %f3118, %f3115, %p959; and.b32 %r5699, %r5696, 2; setp.eq.s32 %p960, %r5699, 0; neg.ftz.f32 %f3120, %f3116; selp.f32 %f3121, %f3116, %f3120, %p960; add.ftz.f32 %f3122, %f3119, %f3121; and.b32 %r5700, %r5696, 4; setp.eq.s32 %p961, %r5700, 0; neg.ftz.f32 %f3123, %f3117; selp.f32 %f3124, %f3117, %f3123, %p961; add.ftz.f32 %f3125, %f3124, %f3122; add.s32 %r5701, %r5597, -559038707; add.s32 %r5702, %r5601, %r5701; sub.s32 %r5703, %r5701, %r5604; xor.b32 %r5704, %r5607, %r5703; add.s32 %r5705, %r5604, %r5702; sub.s32 %r5706, %r5702, %r5704; shf.l.wrap.b32 %r5707, %r5704, %r5704, 8; xor.b32 %r5708, %r5707, %r5706; add.s32 %r5709, %r5704, %r5705; sub.s32 %r5710, %r5705, %r5708; shf.l.wrap.b32 %r5711, %r5708, %r5708, 16; xor.b32 %r5712, %r5711, %r5710; add.s32 %r5713, %r5708, %r5709; sub.s32 %r5714, %r5709, %r5712; shf.l.wrap.b32 %r5715, %r5712, %r5712, 19; xor.b32 %r5716, %r5715, %r5714; add.s32 %r5717, %r5712, %r5713; sub.s32 %r5718, %r5713, %r5716; shf.l.wrap.b32 %r5719, %r5716, %r5716, 4; add.s32 %r5720, %r5716, %r5717; add.s32 %r5721, %r5717, %r5599; xor.b32 %r5722, %r5720, %r5718; xor.b32 %r5723, %r5722, %r5719; shf.l.wrap.b32 %r5724, %r5720, %r5720, 14; sub.s32 %r5725, %r5723, %r5724; xor.b32 %r5726, %r5725, %r5721; shf.l.wrap.b32 %r5727, %r5725, %r5725, 11; sub.s32 %r5728, %r5726, %r5727; xor.b32 %r5729, %r5728, %r5720; shf.l.wrap.b32 %r5730, %r5728, %r5728, 25; sub.s32 %r5731, %r5729, %r5730; xor.b32 %r5732, %r5731, %r5725; shf.l.wrap.b32 %r5733, %r5731, %r5731, 16; sub.s32 %r5734, %r5732, %r5733; xor.b32 %r5735, %r5734, %r5728; shf.l.wrap.b32 %r5736, %r5734, %r5734, 4; sub.s32 %r5737, %r5735, %r5736; xor.b32 %r5738, %r5737, %r5731; shf.l.wrap.b32 %r5739, %r5737, %r5737, 14; sub.s32 %r5740, %r5738, %r5739; xor.b32 %r5741, %r5740, %r5734; shf.l.wrap.b32 %r5742, %r5740, %r5740, 24; sub.s32 %r5743, %r5741, %r5742; add.ftz.f32 %f3126, %f3077, 0fBF800000; and.b32 %r5744, %r5743, 31; setp.lt.u32 %p962, %r5744, 24; selp.f32 %f3127, %f3075, %f3126, %p962; setp.lt.u32 %p963, %r5744, 16; selp.f32 %f3128, %f3126, %f3079, %p963; setp.lt.u32 %p964, %r5744, 8; selp.f32 %f3129, %f3079, %f3081, %p964; and.b32 %r5745, %r5743, 1; setp.eq.b32 %p965, %r5745, 1; neg.ftz.f32 %f3130, %f3127; selp.f32 %f3131, %f3130, %f3127, %p965; and.b32 %r5746, %r5743, 2; setp.eq.s32 %p966, %r5746, 0; neg.ftz.f32 %f3132, %f3128; selp.f32 %f3133, %f3128, %f3132, %p966; add.ftz.f32 %f3134, %f3131, %f3133; and.b32 %r5747, %r5743, 4; setp.eq.s32 %p967, %r5747, 0; neg.ftz.f32 %f3135, %f3129; selp.f32 %f3136, %f3129, %f3135, %p967; add.ftz.f32 %f3137, %f3136, %f3134; sub.s32 %r5748, %r5701, %r5654; xor.b32 %r5749, %r5656, %r5748; add.s32 %r5750, %r5654, %r5702; sub.s32 %r5751, %r5702, %r5749; shf.l.wrap.b32 %r5752, %r5749, %r5749, 8; xor.b32 %r5753, %r5752, %r5751; add.s32 %r5754, %r5749, %r5750; sub.s32 %r5755, %r5750, %r5753; shf.l.wrap.b32 %r5756, %r5753, %r5753, 16; xor.b32 %r5757, %r5756, %r5755; add.s32 %r5758, %r5753, %r5754; sub.s32 %r5759, %r5754, %r5757; shf.l.wrap.b32 %r5760, %r5757, %r5757, 19; xor.b32 %r5761, %r5760, %r5759; add.s32 %r5762, %r5757, %r5758; sub.s32 %r5763, %r5758, %r5761; shf.l.wrap.b32 %r5764, %r5761, %r5761, 4; add.s32 %r5765, %r5761, %r5762; add.s32 %r5766, %r5762, %r5599; xor.b32 %r5767, %r5765, %r5763; xor.b32 %r5768, %r5767, %r5764; shf.l.wrap.b32 %r5769, %r5765, %r5765, 14; sub.s32 %r5770, %r5768, %r5769; xor.b32 %r5771, %r5770, %r5766; shf.l.wrap.b32 %r5772, %r5770, %r5770, 11; sub.s32 %r5773, %r5771, %r5772; xor.b32 %r5774, %r5773, %r5765; shf.l.wrap.b32 %r5775, %r5773, %r5773, 25; sub.s32 %r5776, %r5774, %r5775; xor.b32 %r5777, %r5776, %r5770; shf.l.wrap.b32 %r5778, %r5776, %r5776, 16; sub.s32 %r5779, %r5777, %r5778; xor.b32 %r5780, %r5779, %r5773; shf.l.wrap.b32 %r5781, %r5779, %r5779, 4; sub.s32 %r5782, %r5780, %r5781; xor.b32 %r5783, %r5782, %r5776; shf.l.wrap.b32 %r5784, %r5782, %r5782, 14; sub.s32 %r5785, %r5783, %r5784; xor.b32 %r5786, %r5785, %r5779; shf.l.wrap.b32 %r5787, %r5785, %r5785, 24; sub.s32 %r5788, %r5786, %r5787; and.b32 %r5789, %r5788, 31; setp.lt.u32 %p968, %r5789, 24; selp.f32 %f3138, %f3113, %f3126, %p968; setp.lt.u32 %p969, %r5789, 16; selp.f32 %f3139, %f3126, %f3079, %p969; setp.lt.u32 %p970, %r5789, 8; selp.f32 %f3140, %f3079, %f3081, %p970; and.b32 %r5790, %r5788, 1; setp.eq.b32 %p971, %r5790, 1; neg.ftz.f32 %f3141, %f3138; selp.f32 %f3142, %f3141, %f3138, %p971; and.b32 %r5791, %r5788, 2; setp.eq.s32 %p972, %r5791, 0; neg.ftz.f32 %f3143, %f3139; selp.f32 %f3144, %f3139, %f3143, %p972; add.ftz.f32 %f3145, %f3142, %f3144; and.b32 %r5792, %r5788, 4; setp.eq.s32 %p973, %r5792, 0; neg.ftz.f32 %f3146, %f3140; selp.f32 %f3147, %f3140, %f3146, %p973; add.ftz.f32 %f3148, %f3147, %f3145; not.b32 %r5793, %r5598; add.s32 %r5794, %r5598, -559038707; add.s32 %r5795, %r5596, %r5793; shf.l.wrap.b32 %r5796, %r5794, %r5794, 4; xor.b32 %r5797, %r5796, %r5795; add.s32 %r5798, %r5794, %r5600; sub.s32 %r5799, %r5600, %r5797; shf.l.wrap.b32 %r5800, %r5797, %r5797, 6; xor.b32 %r5801, %r5800, %r5799; add.s32 %r5802, %r5797, %r5798; sub.s32 %r5803, %r5798, %r5801; shf.l.wrap.b32 %r5804, %r5801, %r5801, 8; xor.b32 %r5805, %r5804, %r5803; add.s32 %r5806, %r5801, %r5802; sub.s32 %r5807, %r5802, %r5805; shf.l.wrap.b32 %r5808, %r5805, %r5805, 16; xor.b32 %r5809, %r5808, %r5807; add.s32 %r5810, %r5805, %r5806; sub.s32 %r5811, %r5806, %r5809; shf.l.wrap.b32 %r5812, %r5809, %r5809, 19; xor.b32 %r5813, %r5812, %r5811; add.s32 %r5814, %r5809, %r5810; sub.s32 %r5815, %r5810, %r5813; shf.l.wrap.b32 %r5816, %r5813, %r5813, 4; add.s32 %r5817, %r5813, %r5814; add.s32 %r5818, %r5814, %r5599; xor.b32 %r5819, %r5817, %r5815; xor.b32 %r5820, %r5819, %r5816; shf.l.wrap.b32 %r5821, %r5817, %r5817, 14; sub.s32 %r5822, %r5820, %r5821; xor.b32 %r5823, %r5822, %r5818; shf.l.wrap.b32 %r5824, %r5822, %r5822, 11; sub.s32 %r5825, %r5823, %r5824; xor.b32 %r5826, %r5825, %r5817; shf.l.wrap.b32 %r5827, %r5825, %r5825, 25; sub.s32 %r5828, %r5826, %r5827; xor.b32 %r5829, %r5828, %r5822; shf.l.wrap.b32 %r5830, %r5828, %r5828, 16; sub.s32 %r5831, %r5829, %r5830; xor.b32 %r5832, %r5831, %r5825; shf.l.wrap.b32 %r5833, %r5831, %r5831, 4; sub.s32 %r5834, %r5832, %r5833; xor.b32 %r5835, %r5834, %r5828; shf.l.wrap.b32 %r5836, %r5834, %r5834, 14; sub.s32 %r5837, %r5835, %r5836; xor.b32 %r5838, %r5837, %r5831; shf.l.wrap.b32 %r5839, %r5837, %r5837, 24; sub.s32 %r5840, %r5838, %r5839; add.ftz.f32 %f3149, %f3079, 0fBF800000; and.b32 %r5841, %r5840, 31; setp.lt.u32 %p974, %r5841, 24; selp.f32 %f3150, %f3075, %f3077, %p974; setp.lt.u32 %p975, %r5841, 16; selp.f32 %f3151, %f3077, %f3149, %p975; setp.lt.u32 %p976, %r5841, 8; selp.f32 %f3152, %f3149, %f3081, %p976; and.b32 %r5842, %r5840, 1; setp.eq.b32 %p977, %r5842, 1; neg.ftz.f32 %f3153, %f3150; selp.f32 %f3154, %f3153, %f3150, %p977; and.b32 %r5843, %r5840, 2; setp.eq.s32 %p978, %r5843, 0; neg.ftz.f32 %f3155, %f3151; selp.f32 %f3156, %f3151, %f3155, %p978; add.ftz.f32 %f3157, %f3154, %f3156; and.b32 %r5844, %r5840, 4; setp.eq.s32 %p979, %r5844, 0; neg.ftz.f32 %f3158, %f3152; selp.f32 %f3159, %f3152, %f3158, %p979; add.ftz.f32 %f3160, %f3159, %f3157; xor.b32 %r5845, %r5796, %r5602; sub.s32 %r5846, %r5600, %r5845; shf.l.wrap.b32 %r5847, %r5845, %r5845, 6; xor.b32 %r5848, %r5847, %r5846; add.s32 %r5849, %r5845, %r5798; sub.s32 %r5850, %r5798, %r5848; shf.l.wrap.b32 %r5851, %r5848, %r5848, 8; xor.b32 %r5852, %r5851, %r5850; add.s32 %r5853, %r5848, %r5849; sub.s32 %r5854, %r5849, %r5852; shf.l.wrap.b32 %r5855, %r5852, %r5852, 16; xor.b32 %r5856, %r5855, %r5854; add.s32 %r5857, %r5852, %r5853; sub.s32 %r5858, %r5853, %r5856; shf.l.wrap.b32 %r5859, %r5856, %r5856, 19; xor.b32 %r5860, %r5859, %r5858; add.s32 %r5861, %r5856, %r5857; sub.s32 %r5862, %r5857, %r5860; shf.l.wrap.b32 %r5863, %r5860, %r5860, 4; add.s32 %r5864, %r5860, %r5861; add.s32 %r5865, %r5861, %r5599; xor.b32 %r5866, %r5864, %r5862; xor.b32 %r5867, %r5866, %r5863; shf.l.wrap.b32 %r5868, %r5864, %r5864, 14; sub.s32 %r5869, %r5867, %r5868; xor.b32 %r5870, %r5869, %r5865; shf.l.wrap.b32 %r5871, %r5869, %r5869, 11; sub.s32 %r5872, %r5870, %r5871; xor.b32 %r5873, %r5872, %r5864; shf.l.wrap.b32 %r5874, %r5872, %r5872, 25; sub.s32 %r5875, %r5873, %r5874; xor.b32 %r5876, %r5875, %r5869; shf.l.wrap.b32 %r5877, %r5875, %r5875, 16; sub.s32 %r5878, %r5876, %r5877; xor.b32 %r5879, %r5878, %r5872; shf.l.wrap.b32 %r5880, %r5878, %r5878, 4; sub.s32 %r5881, %r5879, %r5880; xor.b32 %r5882, %r5881, %r5875; shf.l.wrap.b32 %r5883, %r5881, %r5881, 14; sub.s32 %r5884, %r5882, %r5883; xor.b32 %r5885, %r5884, %r5878; shf.l.wrap.b32 %r5886, %r5884, %r5884, 24; sub.s32 %r5887, %r5885, %r5886; and.b32 %r5888, %r5887, 31; setp.lt.u32 %p980, %r5888, 24; selp.f32 %f3161, %f3113, %f3077, %p980; setp.lt.u32 %p981, %r5888, 16; selp.f32 %f3162, %f3077, %f3149, %p981; setp.lt.u32 %p982, %r5888, 8; selp.f32 %f3163, %f3149, %f3081, %p982; and.b32 %r5889, %r5887, 1; setp.eq.b32 %p983, %r5889, 1; neg.ftz.f32 %f3164, %f3161; selp.f32 %f3165, %f3164, %f3161, %p983; and.b32 %r5890, %r5887, 2; setp.eq.s32 %p984, %r5890, 0; neg.ftz.f32 %f3166, %f3162; selp.f32 %f3167, %f3162, %f3166, %p984; add.ftz.f32 %f3168, %f3165, %f3167; and.b32 %r5891, %r5887, 4; setp.eq.s32 %p985, %r5891, 0; neg.ftz.f32 %f3169, %f3163; selp.f32 %f3170, %f3163, %f3169, %p985; add.ftz.f32 %f3171, %f3170, %f3168; add.s32 %r5892, %r5794, %r5701; sub.s32 %r5893, %r5701, %r5797; xor.b32 %r5894, %r5800, %r5893; add.s32 %r5895, %r5797, %r5892; sub.s32 %r5896, %r5892, %r5894; shf.l.wrap.b32 %r5897, %r5894, %r5894, 8; xor.b32 %r5898, %r5897, %r5896; add.s32 %r5899, %r5894, %r5895; sub.s32 %r5900, %r5895, %r5898; shf.l.wrap.b32 %r5901, %r5898, %r5898, 16; xor.b32 %r5902, %r5901, %r5900; add.s32 %r5903, %r5898, %r5899; sub.s32 %r5904, %r5899, %r5902; shf.l.wrap.b32 %r5905, %r5902, %r5902, 19; xor.b32 %r5906, %r5905, %r5904; add.s32 %r5907, %r5902, %r5903; sub.s32 %r5908, %r5903, %r5906; shf.l.wrap.b32 %r5909, %r5906, %r5906, 4; add.s32 %r5910, %r5906, %r5907; add.s32 %r5911, %r5907, %r5599; xor.b32 %r5912, %r5910, %r5908; xor.b32 %r5913, %r5912, %r5909; shf.l.wrap.b32 %r5914, %r5910, %r5910, 14; sub.s32 %r5915, %r5913, %r5914; xor.b32 %r5916, %r5915, %r5911; shf.l.wrap.b32 %r5917, %r5915, %r5915, 11; sub.s32 %r5918, %r5916, %r5917; xor.b32 %r5919, %r5918, %r5910; shf.l.wrap.b32 %r5920, %r5918, %r5918, 25; sub.s32 %r5921, %r5919, %r5920; xor.b32 %r5922, %r5921, %r5915; shf.l.wrap.b32 %r5923, %r5921, %r5921, 16; sub.s32 %r5924, %r5922, %r5923; xor.b32 %r5925, %r5924, %r5918; shf.l.wrap.b32 %r5926, %r5924, %r5924, 4; sub.s32 %r5927, %r5925, %r5926; xor.b32 %r5928, %r5927, %r5921; shf.l.wrap.b32 %r5929, %r5927, %r5927, 14; sub.s32 %r5930, %r5928, %r5929; xor.b32 %r5931, %r5930, %r5924; shf.l.wrap.b32 %r5932, %r5930, %r5930, 24; sub.s32 %r5933, %r5931, %r5932; and.b32 %r5934, %r5933, 31; setp.lt.u32 %p986, %r5934, 24; selp.f32 %f3172, %f3075, %f3126, %p986; setp.lt.u32 %p987, %r5934, 16; selp.f32 %f3173, %f3126, %f3149, %p987; setp.lt.u32 %p988, %r5934, 8; selp.f32 %f3174, %f3149, %f3081, %p988; and.b32 %r5935, %r5933, 1; setp.eq.b32 %p989, %r5935, 1; neg.ftz.f32 %f3175, %f3172; selp.f32 %f3176, %f3175, %f3172, %p989; and.b32 %r5936, %r5933, 2; setp.eq.s32 %p990, %r5936, 0; neg.ftz.f32 %f3177, %f3173; selp.f32 %f3178, %f3173, %f3177, %p990; add.ftz.f32 %f3179, %f3176, %f3178; and.b32 %r5937, %r5933, 4; setp.eq.s32 %p991, %r5937, 0; neg.ftz.f32 %f3180, %f3174; selp.f32 %f3181, %f3174, %f3180, %p991; add.ftz.f32 %f3182, %f3181, %f3179; sub.s32 %r5938, %r5701, %r5845; xor.b32 %r5939, %r5847, %r5938; add.s32 %r5940, %r5845, %r5892; sub.s32 %r5941, %r5892, %r5939; shf.l.wrap.b32 %r5942, %r5939, %r5939, 8; xor.b32 %r5943, %r5942, %r5941; add.s32 %r5944, %r5939, %r5940; sub.s32 %r5945, %r5940, %r5943; shf.l.wrap.b32 %r5946, %r5943, %r5943, 16; xor.b32 %r5947, %r5946, %r5945; add.s32 %r5948, %r5943, %r5944; sub.s32 %r5949, %r5944, %r5947; shf.l.wrap.b32 %r5950, %r5947, %r5947, 19; xor.b32 %r5951, %r5950, %r5949; add.s32 %r5952, %r5947, %r5948; sub.s32 %r5953, %r5948, %r5951; shf.l.wrap.b32 %r5954, %r5951, %r5951, 4; add.s32 %r5955, %r5951, %r5952; add.s32 %r5956, %r5952, %r5599; xor.b32 %r5957, %r5955, %r5953; xor.b32 %r5958, %r5957, %r5954; shf.l.wrap.b32 %r5959, %r5955, %r5955, 14; sub.s32 %r5960, %r5958, %r5959; xor.b32 %r5961, %r5960, %r5956; shf.l.wrap.b32 %r5962, %r5960, %r5960, 11; sub.s32 %r5963, %r5961, %r5962; xor.b32 %r5964, %r5963, %r5955; shf.l.wrap.b32 %r5965, %r5963, %r5963, 25; sub.s32 %r5966, %r5964, %r5965; xor.b32 %r5967, %r5966, %r5960; shf.l.wrap.b32 %r5968, %r5966, %r5966, 16; sub.s32 %r5969, %r5967, %r5968; xor.b32 %r5970, %r5969, %r5963; shf.l.wrap.b32 %r5971, %r5969, %r5969, 4; sub.s32 %r5972, %r5970, %r5971; xor.b32 %r5973, %r5972, %r5966; shf.l.wrap.b32 %r5974, %r5972, %r5972, 14; sub.s32 %r5975, %r5973, %r5974; xor.b32 %r5976, %r5975, %r5969; shf.l.wrap.b32 %r5977, %r5975, %r5975, 24; sub.s32 %r5978, %r5976, %r5977; and.b32 %r5979, %r5978, 31; setp.lt.u32 %p992, %r5979, 24; selp.f32 %f3183, %f3113, %f3126, %p992; setp.lt.u32 %p993, %r5979, 16; selp.f32 %f3184, %f3126, %f3149, %p993; setp.lt.u32 %p994, %r5979, 8; selp.f32 %f3185, %f3149, %f3081, %p994; and.b32 %r5980, %r5978, 1; setp.eq.b32 %p995, %r5980, 1; neg.ftz.f32 %f3186, %f3183; selp.f32 %f3187, %f3186, %f3183, %p995; and.b32 %r5981, %r5978, 2; setp.eq.s32 %p996, %r5981, 0; neg.ftz.f32 %f3188, %f3184; selp.f32 %f3189, %f3184, %f3188, %p996; add.ftz.f32 %f3190, %f3187, %f3189; and.b32 %r5982, %r5978, 4; setp.eq.s32 %p997, %r5982, 0; neg.ftz.f32 %f3191, %f3185; selp.f32 %f3192, %f3185, %f3191, %p997; add.ftz.f32 %f3193, %f3192, %f3190; add.s32 %r5983, %r5599, 1; add.s32 %r5984, %r5621, %r5983; xor.b32 %r5985, %r5629, %r5984; sub.s32 %r5986, %r5985, %r5631; xor.b32 %r5987, %r5986, %r5624; shf.l.wrap.b32 %r5988, %r5986, %r5986, 25; sub.s32 %r5989, %r5987, %r5988; xor.b32 %r5990, %r5989, %r5629; shf.l.wrap.b32 %r5991, %r5989, %r5989, 16; sub.s32 %r5992, %r5990, %r5991; xor.b32 %r5993, %r5992, %r5986; shf.l.wrap.b32 %r5994, %r5992, %r5992, 4; sub.s32 %r5995, %r5993, %r5994; xor.b32 %r5996, %r5995, %r5989; shf.l.wrap.b32 %r5997, %r5995, %r5995, 14; sub.s32 %r5998, %r5996, %r5997; xor.b32 %r5999, %r5998, %r5992; shf.l.wrap.b32 %r6000, %r5998, %r5998, 24; sub.s32 %r6001, %r5999, %r6000; add.ftz.f32 %f3194, %f3081, 0fBF800000; and.b32 %r6002, %r6001, 31; setp.lt.u32 %p998, %r6002, 24; selp.f32 %f3195, %f3075, %f3077, %p998; setp.lt.u32 %p999, %r6002, 16; selp.f32 %f3196, %f3077, %f3079, %p999; setp.lt.u32 %p1000, %r6002, 8; selp.f32 %f3197, %f3079, %f3194, %p1000; and.b32 %r6003, %r6001, 1; setp.eq.b32 %p1001, %r6003, 1; neg.ftz.f32 %f3198, %f3195; selp.f32 %f3199, %f3198, %f3195, %p1001; and.b32 %r6004, %r6001, 2; setp.eq.s32 %p1002, %r6004, 0; neg.ftz.f32 %f3200, %f3196; selp.f32 %f3201, %f3196, %f3200, %p1002; add.ftz.f32 %f3202, %f3199, %f3201; and.b32 %r6005, %r6001, 4; setp.eq.s32 %p1003, %r6005, 0; neg.ftz.f32 %f3203, %f3197; selp.f32 %f3204, %f3197, %f3203, %p1003; add.ftz.f32 %f3205, %f3204, %f3202; add.s32 %r6006, %r5670, %r5983; xor.b32 %r6007, %r5678, %r6006; sub.s32 %r6008, %r6007, %r5680; xor.b32 %r6009, %r6008, %r5673; shf.l.wrap.b32 %r6010, %r6008, %r6008, 25; sub.s32 %r6011, %r6009, %r6010; xor.b32 %r6012, %r6011, %r5678; shf.l.wrap.b32 %r6013, %r6011, %r6011, 16; sub.s32 %r6014, %r6012, %r6013; xor.b32 %r6015, %r6014, %r6008; shf.l.wrap.b32 %r6016, %r6014, %r6014, 4; sub.s32 %r6017, %r6015, %r6016; xor.b32 %r6018, %r6017, %r6011; shf.l.wrap.b32 %r6019, %r6017, %r6017, 14; sub.s32 %r6020, %r6018, %r6019; xor.b32 %r6021, %r6020, %r6014; shf.l.wrap.b32 %r6022, %r6020, %r6020, 24; sub.s32 %r6023, %r6021, %r6022; and.b32 %r6024, %r6023, 31; setp.lt.u32 %p1004, %r6024, 24; selp.f32 %f3206, %f3113, %f3077, %p1004; setp.lt.u32 %p1005, %r6024, 16; selp.f32 %f3207, %f3077, %f3079, %p1005; setp.lt.u32 %p1006, %r6024, 8; selp.f32 %f3208, %f3079, %f3194, %p1006; and.b32 %r6025, %r6023, 1; setp.eq.b32 %p1007, %r6025, 1; neg.ftz.f32 %f3209, %f3206; selp.f32 %f3210, %f3209, %f3206, %p1007; and.b32 %r6026, %r6023, 2; setp.eq.s32 %p1008, %r6026, 0; neg.ftz.f32 %f3211, %f3207; selp.f32 %f3212, %f3207, %f3211, %p1008; add.ftz.f32 %f3213, %f3210, %f3212; and.b32 %r6027, %r6023, 4; setp.eq.s32 %p1009, %r6027, 0; neg.ftz.f32 %f3214, %f3208; selp.f32 %f3215, %f3208, %f3214, %p1009; add.ftz.f32 %f3216, %f3215, %f3213; add.s32 %r6028, %r5717, %r5983; xor.b32 %r6029, %r5725, %r6028; sub.s32 %r6030, %r6029, %r5727; xor.b32 %r6031, %r6030, %r5720; shf.l.wrap.b32 %r6032, %r6030, %r6030, 25; sub.s32 %r6033, %r6031, %r6032; xor.b32 %r6034, %r6033, %r5725; shf.l.wrap.b32 %r6035, %r6033, %r6033, 16; sub.s32 %r6036, %r6034, %r6035; xor.b32 %r6037, %r6036, %r6030; shf.l.wrap.b32 %r6038, %r6036, %r6036, 4; sub.s32 %r6039, %r6037, %r6038; xor.b32 %r6040, %r6039, %r6033; shf.l.wrap.b32 %r6041, %r6039, %r6039, 14; sub.s32 %r6042, %r6040, %r6041; xor.b32 %r6043, %r6042, %r6036; shf.l.wrap.b32 %r6044, %r6042, %r6042, 24; sub.s32 %r6045, %r6043, %r6044; and.b32 %r6046, %r6045, 31; setp.lt.u32 %p1010, %r6046, 24; selp.f32 %f3217, %f3075, %f3126, %p1010; setp.lt.u32 %p1011, %r6046, 16; selp.f32 %f3218, %f3126, %f3079, %p1011; setp.lt.u32 %p1012, %r6046, 8; selp.f32 %f3219, %f3079, %f3194, %p1012; and.b32 %r6047, %r6045, 1; setp.eq.b32 %p1013, %r6047, 1; neg.ftz.f32 %f3220, %f3217; selp.f32 %f3221, %f3220, %f3217, %p1013; and.b32 %r6048, %r6045, 2; setp.eq.s32 %p1014, %r6048, 0; neg.ftz.f32 %f3222, %f3218; selp.f32 %f3223, %f3218, %f3222, %p1014; add.ftz.f32 %f3224, %f3221, %f3223; and.b32 %r6049, %r6045, 4; setp.eq.s32 %p1015, %r6049, 0; neg.ftz.f32 %f3225, %f3219; selp.f32 %f3226, %f3219, %f3225, %p1015; add.ftz.f32 %f3227, %f3226, %f3224; add.s32 %r6050, %r5762, %r5983; xor.b32 %r6051, %r5770, %r6050; sub.s32 %r6052, %r6051, %r5772; xor.b32 %r6053, %r6052, %r5765; shf.l.wrap.b32 %r6054, %r6052, %r6052, 25; sub.s32 %r6055, %r6053, %r6054; xor.b32 %r6056, %r6055, %r5770; shf.l.wrap.b32 %r6057, %r6055, %r6055, 16; sub.s32 %r6058, %r6056, %r6057; xor.b32 %r6059, %r6058, %r6052; shf.l.wrap.b32 %r6060, %r6058, %r6058, 4; sub.s32 %r6061, %r6059, %r6060; xor.b32 %r6062, %r6061, %r6055; shf.l.wrap.b32 %r6063, %r6061, %r6061, 14; sub.s32 %r6064, %r6062, %r6063; xor.b32 %r6065, %r6064, %r6058; shf.l.wrap.b32 %r6066, %r6064, %r6064, 24; sub.s32 %r6067, %r6065, %r6066; and.b32 %r6068, %r6067, 31; setp.lt.u32 %p1016, %r6068, 24; selp.f32 %f3228, %f3113, %f3126, %p1016; setp.lt.u32 %p1017, %r6068, 16; selp.f32 %f3229, %f3126, %f3079, %p1017; setp.lt.u32 %p1018, %r6068, 8; selp.f32 %f3230, %f3079, %f3194, %p1018; and.b32 %r6069, %r6067, 1; setp.eq.b32 %p1019, %r6069, 1; neg.ftz.f32 %f3231, %f3228; selp.f32 %f3232, %f3231, %f3228, %p1019; and.b32 %r6070, %r6067, 2; setp.eq.s32 %p1020, %r6070, 0; neg.ftz.f32 %f3233, %f3229; selp.f32 %f3234, %f3229, %f3233, %p1020; add.ftz.f32 %f3235, %f3232, %f3234; and.b32 %r6071, %r6067, 4; setp.eq.s32 %p1021, %r6071, 0; neg.ftz.f32 %f3236, %f3230; selp.f32 %f3237, %f3230, %f3236, %p1021; add.ftz.f32 %f3238, %f3237, %f3235; add.s32 %r6072, %r5814, %r5983; xor.b32 %r6073, %r5822, %r6072; sub.s32 %r6074, %r6073, %r5824; xor.b32 %r6075, %r6074, %r5817; shf.l.wrap.b32 %r6076, %r6074, %r6074, 25; sub.s32 %r6077, %r6075, %r6076; xor.b32 %r6078, %r6077, %r5822; shf.l.wrap.b32 %r6079, %r6077, %r6077, 16; sub.s32 %r6080, %r6078, %r6079; xor.b32 %r6081, %r6080, %r6074; shf.l.wrap.b32 %r6082, %r6080, %r6080, 4; sub.s32 %r6083, %r6081, %r6082; xor.b32 %r6084, %r6083, %r6077; shf.l.wrap.b32 %r6085, %r6083, %r6083, 14; sub.s32 %r6086, %r6084, %r6085; xor.b32 %r6087, %r6086, %r6080; shf.l.wrap.b32 %r6088, %r6086, %r6086, 24; sub.s32 %r6089, %r6087, %r6088; and.b32 %r6090, %r6089, 31; setp.lt.u32 %p1022, %r6090, 24; selp.f32 %f3239, %f3075, %f3077, %p1022; setp.lt.u32 %p1023, %r6090, 16; selp.f32 %f3240, %f3077, %f3149, %p1023; setp.lt.u32 %p1024, %r6090, 8; selp.f32 %f3241, %f3149, %f3194, %p1024; and.b32 %r6091, %r6089, 1; setp.eq.b32 %p1025, %r6091, 1; neg.ftz.f32 %f3242, %f3239; selp.f32 %f3243, %f3242, %f3239, %p1025; and.b32 %r6092, %r6089, 2; setp.eq.s32 %p1026, %r6092, 0; neg.ftz.f32 %f3244, %f3240; selp.f32 %f3245, %f3240, %f3244, %p1026; add.ftz.f32 %f3246, %f3243, %f3245; and.b32 %r6093, %r6089, 4; setp.eq.s32 %p1027, %r6093, 0; neg.ftz.f32 %f3247, %f3241; selp.f32 %f3248, %f3241, %f3247, %p1027; add.ftz.f32 %f3249, %f3248, %f3246; add.s32 %r6094, %r5861, %r5983; xor.b32 %r6095, %r5869, %r6094; sub.s32 %r6096, %r6095, %r5871; xor.b32 %r6097, %r6096, %r5864; shf.l.wrap.b32 %r6098, %r6096, %r6096, 25; sub.s32 %r6099, %r6097, %r6098; xor.b32 %r6100, %r6099, %r5869; shf.l.wrap.b32 %r6101, %r6099, %r6099, 16; sub.s32 %r6102, %r6100, %r6101; xor.b32 %r6103, %r6102, %r6096; shf.l.wrap.b32 %r6104, %r6102, %r6102, 4; sub.s32 %r6105, %r6103, %r6104; xor.b32 %r6106, %r6105, %r6099; shf.l.wrap.b32 %r6107, %r6105, %r6105, 14; sub.s32 %r6108, %r6106, %r6107; xor.b32 %r6109, %r6108, %r6102; shf.l.wrap.b32 %r6110, %r6108, %r6108, 24; sub.s32 %r6111, %r6109, %r6110; and.b32 %r6112, %r6111, 31; setp.lt.u32 %p1028, %r6112, 24; selp.f32 %f3250, %f3113, %f3077, %p1028; setp.lt.u32 %p1029, %r6112, 16; selp.f32 %f3251, %f3077, %f3149, %p1029; setp.lt.u32 %p1030, %r6112, 8; selp.f32 %f3252, %f3149, %f3194, %p1030; and.b32 %r6113, %r6111, 1; setp.eq.b32 %p1031, %r6113, 1; neg.ftz.f32 %f3253, %f3250; selp.f32 %f3254, %f3253, %f3250, %p1031; and.b32 %r6114, %r6111, 2; setp.eq.s32 %p1032, %r6114, 0; neg.ftz.f32 %f3255, %f3251; selp.f32 %f3256, %f3251, %f3255, %p1032; add.ftz.f32 %f3257, %f3254, %f3256; and.b32 %r6115, %r6111, 4; setp.eq.s32 %p1033, %r6115, 0; neg.ftz.f32 %f3258, %f3252; selp.f32 %f3259, %f3252, %f3258, %p1033; add.ftz.f32 %f3260, %f3259, %f3257; add.s32 %r6116, %r5907, %r5983; xor.b32 %r6117, %r5915, %r6116; sub.s32 %r6118, %r6117, %r5917; xor.b32 %r6119, %r6118, %r5910; shf.l.wrap.b32 %r6120, %r6118, %r6118, 25; sub.s32 %r6121, %r6119, %r6120; xor.b32 %r6122, %r6121, %r5915; shf.l.wrap.b32 %r6123, %r6121, %r6121, 16; sub.s32 %r6124, %r6122, %r6123; xor.b32 %r6125, %r6124, %r6118; shf.l.wrap.b32 %r6126, %r6124, %r6124, 4; sub.s32 %r6127, %r6125, %r6126; xor.b32 %r6128, %r6127, %r6121; shf.l.wrap.b32 %r6129, %r6127, %r6127, 14; sub.s32 %r6130, %r6128, %r6129; xor.b32 %r6131, %r6130, %r6124; shf.l.wrap.b32 %r6132, %r6130, %r6130, 24; sub.s32 %r6133, %r6131, %r6132; and.b32 %r6134, %r6133, 31; setp.lt.u32 %p1034, %r6134, 24; selp.f32 %f3261, %f3075, %f3126, %p1034; setp.lt.u32 %p1035, %r6134, 16; selp.f32 %f3262, %f3126, %f3149, %p1035; setp.lt.u32 %p1036, %r6134, 8; selp.f32 %f3263, %f3149, %f3194, %p1036; and.b32 %r6135, %r6133, 1; setp.eq.b32 %p1037, %r6135, 1; neg.ftz.f32 %f3264, %f3261; selp.f32 %f3265, %f3264, %f3261, %p1037; and.b32 %r6136, %r6133, 2; setp.eq.s32 %p1038, %r6136, 0; neg.ftz.f32 %f3266, %f3262; selp.f32 %f3267, %f3262, %f3266, %p1038; add.ftz.f32 %f3268, %f3265, %f3267; and.b32 %r6137, %r6133, 4; setp.eq.s32 %p1039, %r6137, 0; neg.ftz.f32 %f3269, %f3263; selp.f32 %f3270, %f3263, %f3269, %p1039; add.ftz.f32 %f3271, %f3270, %f3268; add.s32 %r6138, %r5952, %r5983; xor.b32 %r6139, %r5960, %r6138; sub.s32 %r6140, %r6139, %r5962; xor.b32 %r6141, %r6140, %r5955; shf.l.wrap.b32 %r6142, %r6140, %r6140, 25; sub.s32 %r6143, %r6141, %r6142; xor.b32 %r6144, %r6143, %r5960; shf.l.wrap.b32 %r6145, %r6143, %r6143, 16; sub.s32 %r6146, %r6144, %r6145; xor.b32 %r6147, %r6146, %r6140; shf.l.wrap.b32 %r6148, %r6146, %r6146, 4; sub.s32 %r6149, %r6147, %r6148; xor.b32 %r6150, %r6149, %r6143; shf.l.wrap.b32 %r6151, %r6149, %r6149, 14; sub.s32 %r6152, %r6150, %r6151; xor.b32 %r6153, %r6152, %r6146; shf.l.wrap.b32 %r6154, %r6152, %r6152, 24; sub.s32 %r6155, %r6153, %r6154; and.b32 %r6156, %r6155, 31; setp.lt.u32 %p1040, %r6156, 24; selp.f32 %f3272, %f3113, %f3126, %p1040; setp.lt.u32 %p1041, %r6156, 16; selp.f32 %f3273, %f3126, %f3149, %p1041; setp.lt.u32 %p1042, %r6156, 8; selp.f32 %f3274, %f3149, %f3194, %p1042; and.b32 %r6157, %r6155, 1; setp.eq.b32 %p1043, %r6157, 1; neg.ftz.f32 %f3275, %f3272; selp.f32 %f3276, %f3275, %f3272, %p1043; and.b32 %r6158, %r6155, 2; setp.eq.s32 %p1044, %r6158, 0; neg.ftz.f32 %f3277, %f3273; selp.f32 %f3278, %f3273, %f3277, %p1044; add.ftz.f32 %f3279, %f3276, %f3278; and.b32 %r6159, %r6155, 4; setp.eq.s32 %p1045, %r6159, 0; neg.ftz.f32 %f3280, %f3274; selp.f32 %f3281, %f3274, %f3280, %p1045; add.ftz.f32 %f3282, %f3281, %f3279; sub.ftz.f32 %f3283, %f3069, %f3086; sub.ftz.f32 %f3284, %f3069, %f3091; sub.ftz.f32 %f3285, %f3069, %f3096; mul.ftz.f32 %f3286, %f3283, %f3112; fma.rn.ftz.f32 %f3287, %f3086, %f3125, %f3286; mul.ftz.f32 %f3288, %f3283, %f3137; fma.rn.ftz.f32 %f3289, %f3086, %f3148, %f3288; mul.ftz.f32 %f3290, %f3091, %f3289; fma.rn.ftz.f32 %f3291, %f3284, %f3287, %f3290; mul.ftz.f32 %f3292, %f3283, %f3160; fma.rn.ftz.f32 %f3293, %f3086, %f3171, %f3292; mul.ftz.f32 %f3294, %f3283, %f3182; fma.rn.ftz.f32 %f3295, %f3086, %f3193, %f3294; mul.ftz.f32 %f3296, %f3091, %f3295; fma.rn.ftz.f32 %f3297, %f3284, %f3293, %f3296; mul.ftz.f32 %f3298, %f3096, %f3297; fma.rn.ftz.f32 %f3299, %f3285, %f3291, %f3298; mul.ftz.f32 %f3300, %f3283, %f3205; fma.rn.ftz.f32 %f3301, %f3086, %f3216, %f3300; mul.ftz.f32 %f3302, %f3283, %f3227; fma.rn.ftz.f32 %f3303, %f3086, %f3238, %f3302; mul.ftz.f32 %f3304, %f3091, %f3303; fma.rn.ftz.f32 %f3305, %f3284, %f3301, %f3304; mul.ftz.f32 %f3306, %f3283, %f3249; fma.rn.ftz.f32 %f3307, %f3086, %f3260, %f3306; mul.ftz.f32 %f3308, %f3283, %f3271; fma.rn.ftz.f32 %f3309, %f3086, %f3282, %f3308; mul.ftz.f32 %f3310, %f3091, %f3309; fma.rn.ftz.f32 %f3311, %f3284, %f3307, %f3310; mul.ftz.f32 %f3312, %f3096, %f3311; fma.rn.ftz.f32 %f3313, %f3285, %f3305, %f3312; sub.ftz.f32 %f3314, %f3313, %f3299; fma.rn.ftz.f32 %f153, %f3101, %f3314, %f3299; mov.b32 %r39, %f153; setp.nan.ftz.f32 %p1046, %f153, %f153; mov.f32 %f5581, 0f00000000; @%p1046 bra $L__BB74_96; or.b32 %r6160, %r39, -2147483648; setp.ne.s32 %p1047, %r6160, -2147483648; add.ftz.f32 %f3316, %f153, %f153; setp.eq.ftz.f32 %p1048, %f3316, %f153; and.pred %p1049, %p1048, %p1047; @%p1049 bra $L__BB74_96; shl.b32 %r6161, %r39, 1; setp.lt.u32 %p1050, %r6161, -16777215; selp.f32 %f5581, %f153, 0f00000000, %p1050; $L__BB74_96: mul.ftz.f32 %f3317, %f5581, 0f3F559B3D; fma.rn.ftz.f32 %f3318, %f5587, %f3317, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f3318; mul.ftz.f32 %f5587, %f146, %f5587; mul.ftz.f32 %f5568, %f27, %f5568; mul.ftz.f32 %f5569, %f27, %f5569; mul.ftz.f32 %f5570, %f27, %f5570; mul.ftz.f32 %f5571, %f27, %f5571; add.s32 %r10022, %r10022, 1; setp.lt.s32 %p1051, %r10022, %r37; @%p1051 bra $L__BB74_93; $L__BB74_97: cvt.rmi.ftz.f32.f32 %f3319, %f26; sub.ftz.f32 %f168, %f26, %f3319; setp.eq.ftz.f32 %p1052, %f168, 0f00000000; @%p1052 bra $L__BB74_450; mul.ftz.f32 %f169, %f5587, %f168; cvt.rmi.ftz.f32.f32 %f3321, %f5568; cvt.rzi.ftz.s32.f32 %r6162, %f3321; sub.ftz.f32 %f3322, %f5568, %f3321; cvt.rmi.ftz.f32.f32 %f3323, %f5569; cvt.rzi.ftz.s32.f32 %r6163, %f3323; sub.ftz.f32 %f3324, %f5569, %f3323; cvt.rmi.ftz.f32.f32 %f3325, %f5570; cvt.rzi.ftz.s32.f32 %r6164, %f3325; sub.ftz.f32 %f3326, %f5570, %f3325; cvt.rmi.ftz.f32.f32 %f3327, %f5571; cvt.rzi.ftz.s32.f32 %r6165, %f3327; sub.ftz.f32 %f3328, %f5571, %f3327; mul.ftz.f32 %f3329, %f3322, %f3322; mul.ftz.f32 %f3330, %f3322, %f3329; fma.rn.ftz.f32 %f3331, %f3322, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3332, %f3322, %f3331, 0f41200000; mul.ftz.f32 %f3333, %f3330, %f3332; mul.ftz.f32 %f3334, %f3324, %f3324; mul.ftz.f32 %f3335, %f3324, %f3334; fma.rn.ftz.f32 %f3336, %f3324, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3337, %f3324, %f3336, 0f41200000; mul.ftz.f32 %f3338, %f3335, %f3337; mul.ftz.f32 %f3339, %f3326, %f3326; mul.ftz.f32 %f3340, %f3326, %f3339; fma.rn.ftz.f32 %f3341, %f3326, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3342, %f3326, %f3341, 0f41200000; mul.ftz.f32 %f3343, %f3340, %f3342; mul.ftz.f32 %f3344, %f3328, %f3328; mul.ftz.f32 %f3345, %f3328, %f3344; fma.rn.ftz.f32 %f3346, %f3328, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f3347, %f3328, %f3346, 0f41200000; mul.ftz.f32 %f3348, %f3345, %f3347; add.s32 %r6166, %r6163, -559038708; add.s32 %r6167, %r6164, -559038708; sub.s32 %r6168, %r6162, %r6164; shf.l.wrap.b32 %r6169, %r6167, %r6167, 4; xor.b32 %r6170, %r6169, %r6168; add.s32 %r6171, %r6167, %r6166; sub.s32 %r6172, %r6166, %r6170; shf.l.wrap.b32 %r6173, %r6170, %r6170, 6; xor.b32 %r6174, %r6173, %r6172; add.s32 %r6175, %r6170, %r6171; sub.s32 %r6176, %r6171, %r6174; shf.l.wrap.b32 %r6177, %r6174, %r6174, 8; xor.b32 %r6178, %r6177, %r6176; add.s32 %r6179, %r6174, %r6175; sub.s32 %r6180, %r6175, %r6178; shf.l.wrap.b32 %r6181, %r6178, %r6178, 16; xor.b32 %r6182, %r6181, %r6180; add.s32 %r6183, %r6178, %r6179; sub.s32 %r6184, %r6179, %r6182; shf.l.wrap.b32 %r6185, %r6182, %r6182, 19; xor.b32 %r6186, %r6185, %r6184; add.s32 %r6187, %r6182, %r6183; sub.s32 %r6188, %r6183, %r6186; shf.l.wrap.b32 %r6189, %r6186, %r6186, 4; add.s32 %r6190, %r6186, %r6187; add.s32 %r6191, %r6187, %r6165; xor.b32 %r6192, %r6190, %r6188; xor.b32 %r6193, %r6192, %r6189; shf.l.wrap.b32 %r6194, %r6190, %r6190, 14; sub.s32 %r6195, %r6193, %r6194; xor.b32 %r6196, %r6195, %r6191; shf.l.wrap.b32 %r6197, %r6195, %r6195, 11; sub.s32 %r6198, %r6196, %r6197; xor.b32 %r6199, %r6198, %r6190; shf.l.wrap.b32 %r6200, %r6198, %r6198, 25; sub.s32 %r6201, %r6199, %r6200; xor.b32 %r6202, %r6201, %r6195; shf.l.wrap.b32 %r6203, %r6201, %r6201, 16; sub.s32 %r6204, %r6202, %r6203; xor.b32 %r6205, %r6204, %r6198; shf.l.wrap.b32 %r6206, %r6204, %r6204, 4; sub.s32 %r6207, %r6205, %r6206; xor.b32 %r6208, %r6207, %r6201; shf.l.wrap.b32 %r6209, %r6207, %r6207, 14; sub.s32 %r6210, %r6208, %r6209; xor.b32 %r6211, %r6210, %r6204; shf.l.wrap.b32 %r6212, %r6210, %r6210, 24; sub.s32 %r6213, %r6211, %r6212; and.b32 %r6214, %r6213, 31; setp.lt.u32 %p1053, %r6214, 24; selp.f32 %f3349, %f3322, %f3324, %p1053; setp.lt.u32 %p1054, %r6214, 16; selp.f32 %f3350, %f3324, %f3326, %p1054; setp.lt.u32 %p1055, %r6214, 8; selp.f32 %f3351, %f3326, %f3328, %p1055; and.b32 %r6215, %r6213, 1; setp.eq.b32 %p1056, %r6215, 1; neg.ftz.f32 %f3352, %f3349; selp.f32 %f3353, %f3352, %f3349, %p1056; and.b32 %r6216, %r6213, 2; setp.eq.s32 %p1057, %r6216, 0; neg.ftz.f32 %f3354, %f3350; selp.f32 %f3355, %f3350, %f3354, %p1057; add.ftz.f32 %f3356, %f3353, %f3355; and.b32 %r6217, %r6213, 4; setp.eq.s32 %p1058, %r6217, 0; neg.ftz.f32 %f3357, %f3351; selp.f32 %f3358, %f3351, %f3357, %p1058; add.ftz.f32 %f3359, %f3358, %f3356; add.s32 %r6218, %r6162, 1; sub.s32 %r6219, %r6218, %r6164; xor.b32 %r6220, %r6169, %r6219; sub.s32 %r6221, %r6166, %r6220; shf.l.wrap.b32 %r6222, %r6220, %r6220, 6; xor.b32 %r6223, %r6222, %r6221; add.s32 %r6224, %r6220, %r6171; sub.s32 %r6225, %r6171, %r6223; shf.l.wrap.b32 %r6226, %r6223, %r6223, 8; xor.b32 %r6227, %r6226, %r6225; add.s32 %r6228, %r6223, %r6224; sub.s32 %r6229, %r6224, %r6227; shf.l.wrap.b32 %r6230, %r6227, %r6227, 16; xor.b32 %r6231, %r6230, %r6229; add.s32 %r6232, %r6227, %r6228; sub.s32 %r6233, %r6228, %r6231; shf.l.wrap.b32 %r6234, %r6231, %r6231, 19; xor.b32 %r6235, %r6234, %r6233; add.s32 %r6236, %r6231, %r6232; sub.s32 %r6237, %r6232, %r6235; shf.l.wrap.b32 %r6238, %r6235, %r6235, 4; add.s32 %r6239, %r6235, %r6236; add.s32 %r6240, %r6236, %r6165; xor.b32 %r6241, %r6239, %r6237; xor.b32 %r6242, %r6241, %r6238; shf.l.wrap.b32 %r6243, %r6239, %r6239, 14; sub.s32 %r6244, %r6242, %r6243; xor.b32 %r6245, %r6244, %r6240; shf.l.wrap.b32 %r6246, %r6244, %r6244, 11; sub.s32 %r6247, %r6245, %r6246; xor.b32 %r6248, %r6247, %r6239; shf.l.wrap.b32 %r6249, %r6247, %r6247, 25; sub.s32 %r6250, %r6248, %r6249; xor.b32 %r6251, %r6250, %r6244; shf.l.wrap.b32 %r6252, %r6250, %r6250, 16; sub.s32 %r6253, %r6251, %r6252; xor.b32 %r6254, %r6253, %r6247; shf.l.wrap.b32 %r6255, %r6253, %r6253, 4; sub.s32 %r6256, %r6254, %r6255; xor.b32 %r6257, %r6256, %r6250; shf.l.wrap.b32 %r6258, %r6256, %r6256, 14; sub.s32 %r6259, %r6257, %r6258; xor.b32 %r6260, %r6259, %r6253; shf.l.wrap.b32 %r6261, %r6259, %r6259, 24; sub.s32 %r6262, %r6260, %r6261; add.ftz.f32 %f3360, %f3322, 0fBF800000; mov.f32 %f3361, 0f3F800000; and.b32 %r6263, %r6262, 31; setp.lt.u32 %p1059, %r6263, 24; selp.f32 %f3362, %f3360, %f3324, %p1059; setp.lt.u32 %p1060, %r6263, 16; selp.f32 %f3363, %f3324, %f3326, %p1060; setp.lt.u32 %p1061, %r6263, 8; selp.f32 %f3364, %f3326, %f3328, %p1061; and.b32 %r6264, %r6262, 1; setp.eq.b32 %p1062, %r6264, 1; neg.ftz.f32 %f3365, %f3362; selp.f32 %f3366, %f3365, %f3362, %p1062; and.b32 %r6265, %r6262, 2; setp.eq.s32 %p1063, %r6265, 0; neg.ftz.f32 %f3367, %f3363; selp.f32 %f3368, %f3363, %f3367, %p1063; add.ftz.f32 %f3369, %f3366, %f3368; and.b32 %r6266, %r6262, 4; setp.eq.s32 %p1064, %r6266, 0; neg.ftz.f32 %f3370, %f3364; selp.f32 %f3371, %f3364, %f3370, %p1064; add.ftz.f32 %f3372, %f3371, %f3369; add.s32 %r6267, %r6163, -559038707; add.s32 %r6268, %r6167, %r6267; sub.s32 %r6269, %r6267, %r6170; xor.b32 %r6270, %r6173, %r6269; add.s32 %r6271, %r6170, %r6268; sub.s32 %r6272, %r6268, %r6270; shf.l.wrap.b32 %r6273, %r6270, %r6270, 8; xor.b32 %r6274, %r6273, %r6272; add.s32 %r6275, %r6270, %r6271; sub.s32 %r6276, %r6271, %r6274; shf.l.wrap.b32 %r6277, %r6274, %r6274, 16; xor.b32 %r6278, %r6277, %r6276; add.s32 %r6279, %r6274, %r6275; sub.s32 %r6280, %r6275, %r6278; shf.l.wrap.b32 %r6281, %r6278, %r6278, 19; xor.b32 %r6282, %r6281, %r6280; add.s32 %r6283, %r6278, %r6279; sub.s32 %r6284, %r6279, %r6282; shf.l.wrap.b32 %r6285, %r6282, %r6282, 4; add.s32 %r6286, %r6282, %r6283; add.s32 %r6287, %r6283, %r6165; xor.b32 %r6288, %r6286, %r6284; xor.b32 %r6289, %r6288, %r6285; shf.l.wrap.b32 %r6290, %r6286, %r6286, 14; sub.s32 %r6291, %r6289, %r6290; xor.b32 %r6292, %r6291, %r6287; shf.l.wrap.b32 %r6293, %r6291, %r6291, 11; sub.s32 %r6294, %r6292, %r6293; xor.b32 %r6295, %r6294, %r6286; shf.l.wrap.b32 %r6296, %r6294, %r6294, 25; sub.s32 %r6297, %r6295, %r6296; xor.b32 %r6298, %r6297, %r6291; shf.l.wrap.b32 %r6299, %r6297, %r6297, 16; sub.s32 %r6300, %r6298, %r6299; xor.b32 %r6301, %r6300, %r6294; shf.l.wrap.b32 %r6302, %r6300, %r6300, 4; sub.s32 %r6303, %r6301, %r6302; xor.b32 %r6304, %r6303, %r6297; shf.l.wrap.b32 %r6305, %r6303, %r6303, 14; sub.s32 %r6306, %r6304, %r6305; xor.b32 %r6307, %r6306, %r6300; shf.l.wrap.b32 %r6308, %r6306, %r6306, 24; sub.s32 %r6309, %r6307, %r6308; add.ftz.f32 %f3373, %f3324, 0fBF800000; and.b32 %r6310, %r6309, 31; setp.lt.u32 %p1065, %r6310, 24; selp.f32 %f3374, %f3322, %f3373, %p1065; setp.lt.u32 %p1066, %r6310, 16; selp.f32 %f3375, %f3373, %f3326, %p1066; setp.lt.u32 %p1067, %r6310, 8; selp.f32 %f3376, %f3326, %f3328, %p1067; and.b32 %r6311, %r6309, 1; setp.eq.b32 %p1068, %r6311, 1; neg.ftz.f32 %f3377, %f3374; selp.f32 %f3378, %f3377, %f3374, %p1068; and.b32 %r6312, %r6309, 2; setp.eq.s32 %p1069, %r6312, 0; neg.ftz.f32 %f3379, %f3375; selp.f32 %f3380, %f3375, %f3379, %p1069; add.ftz.f32 %f3381, %f3378, %f3380; and.b32 %r6313, %r6309, 4; setp.eq.s32 %p1070, %r6313, 0; neg.ftz.f32 %f3382, %f3376; selp.f32 %f3383, %f3376, %f3382, %p1070; add.ftz.f32 %f3384, %f3383, %f3381; sub.s32 %r6314, %r6267, %r6220; xor.b32 %r6315, %r6222, %r6314; add.s32 %r6316, %r6220, %r6268; sub.s32 %r6317, %r6268, %r6315; shf.l.wrap.b32 %r6318, %r6315, %r6315, 8; xor.b32 %r6319, %r6318, %r6317; add.s32 %r6320, %r6315, %r6316; sub.s32 %r6321, %r6316, %r6319; shf.l.wrap.b32 %r6322, %r6319, %r6319, 16; xor.b32 %r6323, %r6322, %r6321; add.s32 %r6324, %r6319, %r6320; sub.s32 %r6325, %r6320, %r6323; shf.l.wrap.b32 %r6326, %r6323, %r6323, 19; xor.b32 %r6327, %r6326, %r6325; add.s32 %r6328, %r6323, %r6324; sub.s32 %r6329, %r6324, %r6327; shf.l.wrap.b32 %r6330, %r6327, %r6327, 4; add.s32 %r6331, %r6327, %r6328; add.s32 %r6332, %r6328, %r6165; xor.b32 %r6333, %r6331, %r6329; xor.b32 %r6334, %r6333, %r6330; shf.l.wrap.b32 %r6335, %r6331, %r6331, 14; sub.s32 %r6336, %r6334, %r6335; xor.b32 %r6337, %r6336, %r6332; shf.l.wrap.b32 %r6338, %r6336, %r6336, 11; sub.s32 %r6339, %r6337, %r6338; xor.b32 %r6340, %r6339, %r6331; shf.l.wrap.b32 %r6341, %r6339, %r6339, 25; sub.s32 %r6342, %r6340, %r6341; xor.b32 %r6343, %r6342, %r6336; shf.l.wrap.b32 %r6344, %r6342, %r6342, 16; sub.s32 %r6345, %r6343, %r6344; xor.b32 %r6346, %r6345, %r6339; shf.l.wrap.b32 %r6347, %r6345, %r6345, 4; sub.s32 %r6348, %r6346, %r6347; xor.b32 %r6349, %r6348, %r6342; shf.l.wrap.b32 %r6350, %r6348, %r6348, 14; sub.s32 %r6351, %r6349, %r6350; xor.b32 %r6352, %r6351, %r6345; shf.l.wrap.b32 %r6353, %r6351, %r6351, 24; sub.s32 %r6354, %r6352, %r6353; and.b32 %r6355, %r6354, 31; setp.lt.u32 %p1071, %r6355, 24; selp.f32 %f3385, %f3360, %f3373, %p1071; setp.lt.u32 %p1072, %r6355, 16; selp.f32 %f3386, %f3373, %f3326, %p1072; setp.lt.u32 %p1073, %r6355, 8; selp.f32 %f3387, %f3326, %f3328, %p1073; and.b32 %r6356, %r6354, 1; setp.eq.b32 %p1074, %r6356, 1; neg.ftz.f32 %f3388, %f3385; selp.f32 %f3389, %f3388, %f3385, %p1074; and.b32 %r6357, %r6354, 2; setp.eq.s32 %p1075, %r6357, 0; neg.ftz.f32 %f3390, %f3386; selp.f32 %f3391, %f3386, %f3390, %p1075; add.ftz.f32 %f3392, %f3389, %f3391; and.b32 %r6358, %r6354, 4; setp.eq.s32 %p1076, %r6358, 0; neg.ftz.f32 %f3393, %f3387; selp.f32 %f3394, %f3387, %f3393, %p1076; add.ftz.f32 %f3395, %f3394, %f3392; not.b32 %r6359, %r6164; add.s32 %r6360, %r6164, -559038707; add.s32 %r6361, %r6162, %r6359; shf.l.wrap.b32 %r6362, %r6360, %r6360, 4; xor.b32 %r6363, %r6362, %r6361; add.s32 %r6364, %r6360, %r6166; sub.s32 %r6365, %r6166, %r6363; shf.l.wrap.b32 %r6366, %r6363, %r6363, 6; xor.b32 %r6367, %r6366, %r6365; add.s32 %r6368, %r6363, %r6364; sub.s32 %r6369, %r6364, %r6367; shf.l.wrap.b32 %r6370, %r6367, %r6367, 8; xor.b32 %r6371, %r6370, %r6369; add.s32 %r6372, %r6367, %r6368; sub.s32 %r6373, %r6368, %r6371; shf.l.wrap.b32 %r6374, %r6371, %r6371, 16; xor.b32 %r6375, %r6374, %r6373; add.s32 %r6376, %r6371, %r6372; sub.s32 %r6377, %r6372, %r6375; shf.l.wrap.b32 %r6378, %r6375, %r6375, 19; xor.b32 %r6379, %r6378, %r6377; add.s32 %r6380, %r6375, %r6376; sub.s32 %r6381, %r6376, %r6379; shf.l.wrap.b32 %r6382, %r6379, %r6379, 4; add.s32 %r6383, %r6379, %r6380; add.s32 %r6384, %r6380, %r6165; xor.b32 %r6385, %r6383, %r6381; xor.b32 %r6386, %r6385, %r6382; shf.l.wrap.b32 %r6387, %r6383, %r6383, 14; sub.s32 %r6388, %r6386, %r6387; xor.b32 %r6389, %r6388, %r6384; shf.l.wrap.b32 %r6390, %r6388, %r6388, 11; sub.s32 %r6391, %r6389, %r6390; xor.b32 %r6392, %r6391, %r6383; shf.l.wrap.b32 %r6393, %r6391, %r6391, 25; sub.s32 %r6394, %r6392, %r6393; xor.b32 %r6395, %r6394, %r6388; shf.l.wrap.b32 %r6396, %r6394, %r6394, 16; sub.s32 %r6397, %r6395, %r6396; xor.b32 %r6398, %r6397, %r6391; shf.l.wrap.b32 %r6399, %r6397, %r6397, 4; sub.s32 %r6400, %r6398, %r6399; xor.b32 %r6401, %r6400, %r6394; shf.l.wrap.b32 %r6402, %r6400, %r6400, 14; sub.s32 %r6403, %r6401, %r6402; xor.b32 %r6404, %r6403, %r6397; shf.l.wrap.b32 %r6405, %r6403, %r6403, 24; sub.s32 %r6406, %r6404, %r6405; add.ftz.f32 %f3396, %f3326, 0fBF800000; and.b32 %r6407, %r6406, 31; setp.lt.u32 %p1077, %r6407, 24; selp.f32 %f3397, %f3322, %f3324, %p1077; setp.lt.u32 %p1078, %r6407, 16; selp.f32 %f3398, %f3324, %f3396, %p1078; setp.lt.u32 %p1079, %r6407, 8; selp.f32 %f3399, %f3396, %f3328, %p1079; and.b32 %r6408, %r6406, 1; setp.eq.b32 %p1080, %r6408, 1; neg.ftz.f32 %f3400, %f3397; selp.f32 %f3401, %f3400, %f3397, %p1080; and.b32 %r6409, %r6406, 2; setp.eq.s32 %p1081, %r6409, 0; neg.ftz.f32 %f3402, %f3398; selp.f32 %f3403, %f3398, %f3402, %p1081; add.ftz.f32 %f3404, %f3401, %f3403; and.b32 %r6410, %r6406, 4; setp.eq.s32 %p1082, %r6410, 0; neg.ftz.f32 %f3405, %f3399; selp.f32 %f3406, %f3399, %f3405, %p1082; add.ftz.f32 %f3407, %f3406, %f3404; xor.b32 %r6411, %r6362, %r6168; sub.s32 %r6412, %r6166, %r6411; shf.l.wrap.b32 %r6413, %r6411, %r6411, 6; xor.b32 %r6414, %r6413, %r6412; add.s32 %r6415, %r6411, %r6364; sub.s32 %r6416, %r6364, %r6414; shf.l.wrap.b32 %r6417, %r6414, %r6414, 8; xor.b32 %r6418, %r6417, %r6416; add.s32 %r6419, %r6414, %r6415; sub.s32 %r6420, %r6415, %r6418; shf.l.wrap.b32 %r6421, %r6418, %r6418, 16; xor.b32 %r6422, %r6421, %r6420; add.s32 %r6423, %r6418, %r6419; sub.s32 %r6424, %r6419, %r6422; shf.l.wrap.b32 %r6425, %r6422, %r6422, 19; xor.b32 %r6426, %r6425, %r6424; add.s32 %r6427, %r6422, %r6423; sub.s32 %r6428, %r6423, %r6426; shf.l.wrap.b32 %r6429, %r6426, %r6426, 4; add.s32 %r6430, %r6426, %r6427; add.s32 %r6431, %r6427, %r6165; xor.b32 %r6432, %r6430, %r6428; xor.b32 %r6433, %r6432, %r6429; shf.l.wrap.b32 %r6434, %r6430, %r6430, 14; sub.s32 %r6435, %r6433, %r6434; xor.b32 %r6436, %r6435, %r6431; shf.l.wrap.b32 %r6437, %r6435, %r6435, 11; sub.s32 %r6438, %r6436, %r6437; xor.b32 %r6439, %r6438, %r6430; shf.l.wrap.b32 %r6440, %r6438, %r6438, 25; sub.s32 %r6441, %r6439, %r6440; xor.b32 %r6442, %r6441, %r6435; shf.l.wrap.b32 %r6443, %r6441, %r6441, 16; sub.s32 %r6444, %r6442, %r6443; xor.b32 %r6445, %r6444, %r6438; shf.l.wrap.b32 %r6446, %r6444, %r6444, 4; sub.s32 %r6447, %r6445, %r6446; xor.b32 %r6448, %r6447, %r6441; shf.l.wrap.b32 %r6449, %r6447, %r6447, 14; sub.s32 %r6450, %r6448, %r6449; xor.b32 %r6451, %r6450, %r6444; shf.l.wrap.b32 %r6452, %r6450, %r6450, 24; sub.s32 %r6453, %r6451, %r6452; and.b32 %r6454, %r6453, 31; setp.lt.u32 %p1083, %r6454, 24; selp.f32 %f3408, %f3360, %f3324, %p1083; setp.lt.u32 %p1084, %r6454, 16; selp.f32 %f3409, %f3324, %f3396, %p1084; setp.lt.u32 %p1085, %r6454, 8; selp.f32 %f3410, %f3396, %f3328, %p1085; and.b32 %r6455, %r6453, 1; setp.eq.b32 %p1086, %r6455, 1; neg.ftz.f32 %f3411, %f3408; selp.f32 %f3412, %f3411, %f3408, %p1086; and.b32 %r6456, %r6453, 2; setp.eq.s32 %p1087, %r6456, 0; neg.ftz.f32 %f3413, %f3409; selp.f32 %f3414, %f3409, %f3413, %p1087; add.ftz.f32 %f3415, %f3412, %f3414; and.b32 %r6457, %r6453, 4; setp.eq.s32 %p1088, %r6457, 0; neg.ftz.f32 %f3416, %f3410; selp.f32 %f3417, %f3410, %f3416, %p1088; add.ftz.f32 %f3418, %f3417, %f3415; add.s32 %r6458, %r6360, %r6267; sub.s32 %r6459, %r6267, %r6363; xor.b32 %r6460, %r6366, %r6459; add.s32 %r6461, %r6363, %r6458; sub.s32 %r6462, %r6458, %r6460; shf.l.wrap.b32 %r6463, %r6460, %r6460, 8; xor.b32 %r6464, %r6463, %r6462; add.s32 %r6465, %r6460, %r6461; sub.s32 %r6466, %r6461, %r6464; shf.l.wrap.b32 %r6467, %r6464, %r6464, 16; xor.b32 %r6468, %r6467, %r6466; add.s32 %r6469, %r6464, %r6465; sub.s32 %r6470, %r6465, %r6468; shf.l.wrap.b32 %r6471, %r6468, %r6468, 19; xor.b32 %r6472, %r6471, %r6470; add.s32 %r6473, %r6468, %r6469; sub.s32 %r6474, %r6469, %r6472; shf.l.wrap.b32 %r6475, %r6472, %r6472, 4; add.s32 %r6476, %r6472, %r6473; add.s32 %r6477, %r6473, %r6165; xor.b32 %r6478, %r6476, %r6474; xor.b32 %r6479, %r6478, %r6475; shf.l.wrap.b32 %r6480, %r6476, %r6476, 14; sub.s32 %r6481, %r6479, %r6480; xor.b32 %r6482, %r6481, %r6477; shf.l.wrap.b32 %r6483, %r6481, %r6481, 11; sub.s32 %r6484, %r6482, %r6483; xor.b32 %r6485, %r6484, %r6476; shf.l.wrap.b32 %r6486, %r6484, %r6484, 25; sub.s32 %r6487, %r6485, %r6486; xor.b32 %r6488, %r6487, %r6481; shf.l.wrap.b32 %r6489, %r6487, %r6487, 16; sub.s32 %r6490, %r6488, %r6489; xor.b32 %r6491, %r6490, %r6484; shf.l.wrap.b32 %r6492, %r6490, %r6490, 4; sub.s32 %r6493, %r6491, %r6492; xor.b32 %r6494, %r6493, %r6487; shf.l.wrap.b32 %r6495, %r6493, %r6493, 14; sub.s32 %r6496, %r6494, %r6495; xor.b32 %r6497, %r6496, %r6490; shf.l.wrap.b32 %r6498, %r6496, %r6496, 24; sub.s32 %r6499, %r6497, %r6498; and.b32 %r6500, %r6499, 31; setp.lt.u32 %p1089, %r6500, 24; selp.f32 %f3419, %f3322, %f3373, %p1089; setp.lt.u32 %p1090, %r6500, 16; selp.f32 %f3420, %f3373, %f3396, %p1090; setp.lt.u32 %p1091, %r6500, 8; selp.f32 %f3421, %f3396, %f3328, %p1091; and.b32 %r6501, %r6499, 1; setp.eq.b32 %p1092, %r6501, 1; neg.ftz.f32 %f3422, %f3419; selp.f32 %f3423, %f3422, %f3419, %p1092; and.b32 %r6502, %r6499, 2; setp.eq.s32 %p1093, %r6502, 0; neg.ftz.f32 %f3424, %f3420; selp.f32 %f3425, %f3420, %f3424, %p1093; add.ftz.f32 %f3426, %f3423, %f3425; and.b32 %r6503, %r6499, 4; setp.eq.s32 %p1094, %r6503, 0; neg.ftz.f32 %f3427, %f3421; selp.f32 %f3428, %f3421, %f3427, %p1094; add.ftz.f32 %f3429, %f3428, %f3426; sub.s32 %r6504, %r6267, %r6411; xor.b32 %r6505, %r6413, %r6504; add.s32 %r6506, %r6411, %r6458; sub.s32 %r6507, %r6458, %r6505; shf.l.wrap.b32 %r6508, %r6505, %r6505, 8; xor.b32 %r6509, %r6508, %r6507; add.s32 %r6510, %r6505, %r6506; sub.s32 %r6511, %r6506, %r6509; shf.l.wrap.b32 %r6512, %r6509, %r6509, 16; xor.b32 %r6513, %r6512, %r6511; add.s32 %r6514, %r6509, %r6510; sub.s32 %r6515, %r6510, %r6513; shf.l.wrap.b32 %r6516, %r6513, %r6513, 19; xor.b32 %r6517, %r6516, %r6515; add.s32 %r6518, %r6513, %r6514; sub.s32 %r6519, %r6514, %r6517; shf.l.wrap.b32 %r6520, %r6517, %r6517, 4; add.s32 %r6521, %r6517, %r6518; add.s32 %r6522, %r6518, %r6165; xor.b32 %r6523, %r6521, %r6519; xor.b32 %r6524, %r6523, %r6520; shf.l.wrap.b32 %r6525, %r6521, %r6521, 14; sub.s32 %r6526, %r6524, %r6525; xor.b32 %r6527, %r6526, %r6522; shf.l.wrap.b32 %r6528, %r6526, %r6526, 11; sub.s32 %r6529, %r6527, %r6528; xor.b32 %r6530, %r6529, %r6521; shf.l.wrap.b32 %r6531, %r6529, %r6529, 25; sub.s32 %r6532, %r6530, %r6531; xor.b32 %r6533, %r6532, %r6526; shf.l.wrap.b32 %r6534, %r6532, %r6532, 16; sub.s32 %r6535, %r6533, %r6534; xor.b32 %r6536, %r6535, %r6529; shf.l.wrap.b32 %r6537, %r6535, %r6535, 4; sub.s32 %r6538, %r6536, %r6537; xor.b32 %r6539, %r6538, %r6532; shf.l.wrap.b32 %r6540, %r6538, %r6538, 14; sub.s32 %r6541, %r6539, %r6540; xor.b32 %r6542, %r6541, %r6535; shf.l.wrap.b32 %r6543, %r6541, %r6541, 24; sub.s32 %r6544, %r6542, %r6543; and.b32 %r6545, %r6544, 31; setp.lt.u32 %p1095, %r6545, 24; selp.f32 %f3430, %f3360, %f3373, %p1095; setp.lt.u32 %p1096, %r6545, 16; selp.f32 %f3431, %f3373, %f3396, %p1096; setp.lt.u32 %p1097, %r6545, 8; selp.f32 %f3432, %f3396, %f3328, %p1097; and.b32 %r6546, %r6544, 1; setp.eq.b32 %p1098, %r6546, 1; neg.ftz.f32 %f3433, %f3430; selp.f32 %f3434, %f3433, %f3430, %p1098; and.b32 %r6547, %r6544, 2; setp.eq.s32 %p1099, %r6547, 0; neg.ftz.f32 %f3435, %f3431; selp.f32 %f3436, %f3431, %f3435, %p1099; add.ftz.f32 %f3437, %f3434, %f3436; and.b32 %r6548, %r6544, 4; setp.eq.s32 %p1100, %r6548, 0; neg.ftz.f32 %f3438, %f3432; selp.f32 %f3439, %f3432, %f3438, %p1100; add.ftz.f32 %f3440, %f3439, %f3437; add.s32 %r6549, %r6165, 1; add.s32 %r6550, %r6187, %r6549; xor.b32 %r6551, %r6195, %r6550; sub.s32 %r6552, %r6551, %r6197; xor.b32 %r6553, %r6552, %r6190; shf.l.wrap.b32 %r6554, %r6552, %r6552, 25; sub.s32 %r6555, %r6553, %r6554; xor.b32 %r6556, %r6555, %r6195; shf.l.wrap.b32 %r6557, %r6555, %r6555, 16; sub.s32 %r6558, %r6556, %r6557; xor.b32 %r6559, %r6558, %r6552; shf.l.wrap.b32 %r6560, %r6558, %r6558, 4; sub.s32 %r6561, %r6559, %r6560; xor.b32 %r6562, %r6561, %r6555; shf.l.wrap.b32 %r6563, %r6561, %r6561, 14; sub.s32 %r6564, %r6562, %r6563; xor.b32 %r6565, %r6564, %r6558; shf.l.wrap.b32 %r6566, %r6564, %r6564, 24; sub.s32 %r6567, %r6565, %r6566; add.ftz.f32 %f3441, %f3328, 0fBF800000; and.b32 %r6568, %r6567, 31; setp.lt.u32 %p1101, %r6568, 24; selp.f32 %f3442, %f3322, %f3324, %p1101; setp.lt.u32 %p1102, %r6568, 16; selp.f32 %f3443, %f3324, %f3326, %p1102; setp.lt.u32 %p1103, %r6568, 8; selp.f32 %f3444, %f3326, %f3441, %p1103; and.b32 %r6569, %r6567, 1; setp.eq.b32 %p1104, %r6569, 1; neg.ftz.f32 %f3445, %f3442; selp.f32 %f3446, %f3445, %f3442, %p1104; and.b32 %r6570, %r6567, 2; setp.eq.s32 %p1105, %r6570, 0; neg.ftz.f32 %f3447, %f3443; selp.f32 %f3448, %f3443, %f3447, %p1105; add.ftz.f32 %f3449, %f3446, %f3448; and.b32 %r6571, %r6567, 4; setp.eq.s32 %p1106, %r6571, 0; neg.ftz.f32 %f3450, %f3444; selp.f32 %f3451, %f3444, %f3450, %p1106; add.ftz.f32 %f3452, %f3451, %f3449; add.s32 %r6572, %r6236, %r6549; xor.b32 %r6573, %r6244, %r6572; sub.s32 %r6574, %r6573, %r6246; xor.b32 %r6575, %r6574, %r6239; shf.l.wrap.b32 %r6576, %r6574, %r6574, 25; sub.s32 %r6577, %r6575, %r6576; xor.b32 %r6578, %r6577, %r6244; shf.l.wrap.b32 %r6579, %r6577, %r6577, 16; sub.s32 %r6580, %r6578, %r6579; xor.b32 %r6581, %r6580, %r6574; shf.l.wrap.b32 %r6582, %r6580, %r6580, 4; sub.s32 %r6583, %r6581, %r6582; xor.b32 %r6584, %r6583, %r6577; shf.l.wrap.b32 %r6585, %r6583, %r6583, 14; sub.s32 %r6586, %r6584, %r6585; xor.b32 %r6587, %r6586, %r6580; shf.l.wrap.b32 %r6588, %r6586, %r6586, 24; sub.s32 %r6589, %r6587, %r6588; and.b32 %r6590, %r6589, 31; setp.lt.u32 %p1107, %r6590, 24; selp.f32 %f3453, %f3360, %f3324, %p1107; setp.lt.u32 %p1108, %r6590, 16; selp.f32 %f3454, %f3324, %f3326, %p1108; setp.lt.u32 %p1109, %r6590, 8; selp.f32 %f3455, %f3326, %f3441, %p1109; and.b32 %r6591, %r6589, 1; setp.eq.b32 %p1110, %r6591, 1; neg.ftz.f32 %f3456, %f3453; selp.f32 %f3457, %f3456, %f3453, %p1110; and.b32 %r6592, %r6589, 2; setp.eq.s32 %p1111, %r6592, 0; neg.ftz.f32 %f3458, %f3454; selp.f32 %f3459, %f3454, %f3458, %p1111; add.ftz.f32 %f3460, %f3457, %f3459; and.b32 %r6593, %r6589, 4; setp.eq.s32 %p1112, %r6593, 0; neg.ftz.f32 %f3461, %f3455; selp.f32 %f3462, %f3455, %f3461, %p1112; add.ftz.f32 %f3463, %f3462, %f3460; add.s32 %r6594, %r6283, %r6549; xor.b32 %r6595, %r6291, %r6594; sub.s32 %r6596, %r6595, %r6293; xor.b32 %r6597, %r6596, %r6286; shf.l.wrap.b32 %r6598, %r6596, %r6596, 25; sub.s32 %r6599, %r6597, %r6598; xor.b32 %r6600, %r6599, %r6291; shf.l.wrap.b32 %r6601, %r6599, %r6599, 16; sub.s32 %r6602, %r6600, %r6601; xor.b32 %r6603, %r6602, %r6596; shf.l.wrap.b32 %r6604, %r6602, %r6602, 4; sub.s32 %r6605, %r6603, %r6604; xor.b32 %r6606, %r6605, %r6599; shf.l.wrap.b32 %r6607, %r6605, %r6605, 14; sub.s32 %r6608, %r6606, %r6607; xor.b32 %r6609, %r6608, %r6602; shf.l.wrap.b32 %r6610, %r6608, %r6608, 24; sub.s32 %r6611, %r6609, %r6610; and.b32 %r6612, %r6611, 31; setp.lt.u32 %p1113, %r6612, 24; selp.f32 %f3464, %f3322, %f3373, %p1113; setp.lt.u32 %p1114, %r6612, 16; selp.f32 %f3465, %f3373, %f3326, %p1114; setp.lt.u32 %p1115, %r6612, 8; selp.f32 %f3466, %f3326, %f3441, %p1115; and.b32 %r6613, %r6611, 1; setp.eq.b32 %p1116, %r6613, 1; neg.ftz.f32 %f3467, %f3464; selp.f32 %f3468, %f3467, %f3464, %p1116; and.b32 %r6614, %r6611, 2; setp.eq.s32 %p1117, %r6614, 0; neg.ftz.f32 %f3469, %f3465; selp.f32 %f3470, %f3465, %f3469, %p1117; add.ftz.f32 %f3471, %f3468, %f3470; and.b32 %r6615, %r6611, 4; setp.eq.s32 %p1118, %r6615, 0; neg.ftz.f32 %f3472, %f3466; selp.f32 %f3473, %f3466, %f3472, %p1118; add.ftz.f32 %f3474, %f3473, %f3471; add.s32 %r6616, %r6328, %r6549; xor.b32 %r6617, %r6336, %r6616; sub.s32 %r6618, %r6617, %r6338; xor.b32 %r6619, %r6618, %r6331; shf.l.wrap.b32 %r6620, %r6618, %r6618, 25; sub.s32 %r6621, %r6619, %r6620; xor.b32 %r6622, %r6621, %r6336; shf.l.wrap.b32 %r6623, %r6621, %r6621, 16; sub.s32 %r6624, %r6622, %r6623; xor.b32 %r6625, %r6624, %r6618; shf.l.wrap.b32 %r6626, %r6624, %r6624, 4; sub.s32 %r6627, %r6625, %r6626; xor.b32 %r6628, %r6627, %r6621; shf.l.wrap.b32 %r6629, %r6627, %r6627, 14; sub.s32 %r6630, %r6628, %r6629; xor.b32 %r6631, %r6630, %r6624; shf.l.wrap.b32 %r6632, %r6630, %r6630, 24; sub.s32 %r6633, %r6631, %r6632; and.b32 %r6634, %r6633, 31; setp.lt.u32 %p1119, %r6634, 24; selp.f32 %f3475, %f3360, %f3373, %p1119; setp.lt.u32 %p1120, %r6634, 16; selp.f32 %f3476, %f3373, %f3326, %p1120; setp.lt.u32 %p1121, %r6634, 8; selp.f32 %f3477, %f3326, %f3441, %p1121; and.b32 %r6635, %r6633, 1; setp.eq.b32 %p1122, %r6635, 1; neg.ftz.f32 %f3478, %f3475; selp.f32 %f3479, %f3478, %f3475, %p1122; and.b32 %r6636, %r6633, 2; setp.eq.s32 %p1123, %r6636, 0; neg.ftz.f32 %f3480, %f3476; selp.f32 %f3481, %f3476, %f3480, %p1123; add.ftz.f32 %f3482, %f3479, %f3481; and.b32 %r6637, %r6633, 4; setp.eq.s32 %p1124, %r6637, 0; neg.ftz.f32 %f3483, %f3477; selp.f32 %f3484, %f3477, %f3483, %p1124; add.ftz.f32 %f3485, %f3484, %f3482; add.s32 %r6638, %r6380, %r6549; xor.b32 %r6639, %r6388, %r6638; sub.s32 %r6640, %r6639, %r6390; xor.b32 %r6641, %r6640, %r6383; shf.l.wrap.b32 %r6642, %r6640, %r6640, 25; sub.s32 %r6643, %r6641, %r6642; xor.b32 %r6644, %r6643, %r6388; shf.l.wrap.b32 %r6645, %r6643, %r6643, 16; sub.s32 %r6646, %r6644, %r6645; xor.b32 %r6647, %r6646, %r6640; shf.l.wrap.b32 %r6648, %r6646, %r6646, 4; sub.s32 %r6649, %r6647, %r6648; xor.b32 %r6650, %r6649, %r6643; shf.l.wrap.b32 %r6651, %r6649, %r6649, 14; sub.s32 %r6652, %r6650, %r6651; xor.b32 %r6653, %r6652, %r6646; shf.l.wrap.b32 %r6654, %r6652, %r6652, 24; sub.s32 %r6655, %r6653, %r6654; and.b32 %r6656, %r6655, 31; setp.lt.u32 %p1125, %r6656, 24; selp.f32 %f3486, %f3322, %f3324, %p1125; setp.lt.u32 %p1126, %r6656, 16; selp.f32 %f3487, %f3324, %f3396, %p1126; setp.lt.u32 %p1127, %r6656, 8; selp.f32 %f3488, %f3396, %f3441, %p1127; and.b32 %r6657, %r6655, 1; setp.eq.b32 %p1128, %r6657, 1; neg.ftz.f32 %f3489, %f3486; selp.f32 %f3490, %f3489, %f3486, %p1128; and.b32 %r6658, %r6655, 2; setp.eq.s32 %p1129, %r6658, 0; neg.ftz.f32 %f3491, %f3487; selp.f32 %f3492, %f3487, %f3491, %p1129; add.ftz.f32 %f3493, %f3490, %f3492; and.b32 %r6659, %r6655, 4; setp.eq.s32 %p1130, %r6659, 0; neg.ftz.f32 %f3494, %f3488; selp.f32 %f3495, %f3488, %f3494, %p1130; add.ftz.f32 %f3496, %f3495, %f3493; add.s32 %r6660, %r6427, %r6549; xor.b32 %r6661, %r6435, %r6660; sub.s32 %r6662, %r6661, %r6437; xor.b32 %r6663, %r6662, %r6430; shf.l.wrap.b32 %r6664, %r6662, %r6662, 25; sub.s32 %r6665, %r6663, %r6664; xor.b32 %r6666, %r6665, %r6435; shf.l.wrap.b32 %r6667, %r6665, %r6665, 16; sub.s32 %r6668, %r6666, %r6667; xor.b32 %r6669, %r6668, %r6662; shf.l.wrap.b32 %r6670, %r6668, %r6668, 4; sub.s32 %r6671, %r6669, %r6670; xor.b32 %r6672, %r6671, %r6665; shf.l.wrap.b32 %r6673, %r6671, %r6671, 14; sub.s32 %r6674, %r6672, %r6673; xor.b32 %r6675, %r6674, %r6668; shf.l.wrap.b32 %r6676, %r6674, %r6674, 24; sub.s32 %r6677, %r6675, %r6676; and.b32 %r6678, %r6677, 31; setp.lt.u32 %p1131, %r6678, 24; selp.f32 %f3497, %f3360, %f3324, %p1131; setp.lt.u32 %p1132, %r6678, 16; selp.f32 %f3498, %f3324, %f3396, %p1132; setp.lt.u32 %p1133, %r6678, 8; selp.f32 %f3499, %f3396, %f3441, %p1133; and.b32 %r6679, %r6677, 1; setp.eq.b32 %p1134, %r6679, 1; neg.ftz.f32 %f3500, %f3497; selp.f32 %f3501, %f3500, %f3497, %p1134; and.b32 %r6680, %r6677, 2; setp.eq.s32 %p1135, %r6680, 0; neg.ftz.f32 %f3502, %f3498; selp.f32 %f3503, %f3498, %f3502, %p1135; add.ftz.f32 %f3504, %f3501, %f3503; and.b32 %r6681, %r6677, 4; setp.eq.s32 %p1136, %r6681, 0; neg.ftz.f32 %f3505, %f3499; selp.f32 %f3506, %f3499, %f3505, %p1136; add.ftz.f32 %f3507, %f3506, %f3504; add.s32 %r6682, %r6473, %r6549; xor.b32 %r6683, %r6481, %r6682; sub.s32 %r6684, %r6683, %r6483; xor.b32 %r6685, %r6684, %r6476; shf.l.wrap.b32 %r6686, %r6684, %r6684, 25; sub.s32 %r6687, %r6685, %r6686; xor.b32 %r6688, %r6687, %r6481; shf.l.wrap.b32 %r6689, %r6687, %r6687, 16; sub.s32 %r6690, %r6688, %r6689; xor.b32 %r6691, %r6690, %r6684; shf.l.wrap.b32 %r6692, %r6690, %r6690, 4; sub.s32 %r6693, %r6691, %r6692; xor.b32 %r6694, %r6693, %r6687; shf.l.wrap.b32 %r6695, %r6693, %r6693, 14; sub.s32 %r6696, %r6694, %r6695; xor.b32 %r6697, %r6696, %r6690; shf.l.wrap.b32 %r6698, %r6696, %r6696, 24; sub.s32 %r6699, %r6697, %r6698; and.b32 %r6700, %r6699, 31; setp.lt.u32 %p1137, %r6700, 24; selp.f32 %f3508, %f3322, %f3373, %p1137; setp.lt.u32 %p1138, %r6700, 16; selp.f32 %f3509, %f3373, %f3396, %p1138; setp.lt.u32 %p1139, %r6700, 8; selp.f32 %f3510, %f3396, %f3441, %p1139; and.b32 %r6701, %r6699, 1; setp.eq.b32 %p1140, %r6701, 1; neg.ftz.f32 %f3511, %f3508; selp.f32 %f3512, %f3511, %f3508, %p1140; and.b32 %r6702, %r6699, 2; setp.eq.s32 %p1141, %r6702, 0; neg.ftz.f32 %f3513, %f3509; selp.f32 %f3514, %f3509, %f3513, %p1141; add.ftz.f32 %f3515, %f3512, %f3514; and.b32 %r6703, %r6699, 4; setp.eq.s32 %p1142, %r6703, 0; neg.ftz.f32 %f3516, %f3510; selp.f32 %f3517, %f3510, %f3516, %p1142; add.ftz.f32 %f3518, %f3517, %f3515; add.s32 %r6704, %r6518, %r6549; xor.b32 %r6705, %r6526, %r6704; sub.s32 %r6706, %r6705, %r6528; xor.b32 %r6707, %r6706, %r6521; shf.l.wrap.b32 %r6708, %r6706, %r6706, 25; sub.s32 %r6709, %r6707, %r6708; xor.b32 %r6710, %r6709, %r6526; shf.l.wrap.b32 %r6711, %r6709, %r6709, 16; sub.s32 %r6712, %r6710, %r6711; xor.b32 %r6713, %r6712, %r6706; shf.l.wrap.b32 %r6714, %r6712, %r6712, 4; sub.s32 %r6715, %r6713, %r6714; xor.b32 %r6716, %r6715, %r6709; shf.l.wrap.b32 %r6717, %r6715, %r6715, 14; sub.s32 %r6718, %r6716, %r6717; xor.b32 %r6719, %r6718, %r6712; shf.l.wrap.b32 %r6720, %r6718, %r6718, 24; sub.s32 %r6721, %r6719, %r6720; and.b32 %r6722, %r6721, 31; setp.lt.u32 %p1143, %r6722, 24; selp.f32 %f3519, %f3360, %f3373, %p1143; setp.lt.u32 %p1144, %r6722, 16; selp.f32 %f3520, %f3373, %f3396, %p1144; setp.lt.u32 %p1145, %r6722, 8; selp.f32 %f3521, %f3396, %f3441, %p1145; and.b32 %r6723, %r6721, 1; setp.eq.b32 %p1146, %r6723, 1; neg.ftz.f32 %f3522, %f3519; selp.f32 %f3523, %f3522, %f3519, %p1146; and.b32 %r6724, %r6721, 2; setp.eq.s32 %p1147, %r6724, 0; neg.ftz.f32 %f3524, %f3520; selp.f32 %f3525, %f3520, %f3524, %p1147; add.ftz.f32 %f3526, %f3523, %f3525; and.b32 %r6725, %r6721, 4; setp.eq.s32 %p1148, %r6725, 0; neg.ftz.f32 %f3527, %f3521; selp.f32 %f3528, %f3521, %f3527, %p1148; add.ftz.f32 %f3529, %f3528, %f3526; sub.ftz.f32 %f3530, %f3361, %f3333; sub.ftz.f32 %f3531, %f3361, %f3338; sub.ftz.f32 %f3532, %f3361, %f3343; mul.ftz.f32 %f3533, %f3530, %f3359; fma.rn.ftz.f32 %f3534, %f3333, %f3372, %f3533; mul.ftz.f32 %f3535, %f3530, %f3384; fma.rn.ftz.f32 %f3536, %f3333, %f3395, %f3535; mul.ftz.f32 %f3537, %f3338, %f3536; fma.rn.ftz.f32 %f3538, %f3531, %f3534, %f3537; mul.ftz.f32 %f3539, %f3530, %f3407; fma.rn.ftz.f32 %f3540, %f3333, %f3418, %f3539; mul.ftz.f32 %f3541, %f3530, %f3429; fma.rn.ftz.f32 %f3542, %f3333, %f3440, %f3541; mul.ftz.f32 %f3543, %f3338, %f3542; fma.rn.ftz.f32 %f3544, %f3531, %f3540, %f3543; mul.ftz.f32 %f3545, %f3343, %f3544; fma.rn.ftz.f32 %f3546, %f3532, %f3538, %f3545; mul.ftz.f32 %f3547, %f3530, %f3452; fma.rn.ftz.f32 %f3548, %f3333, %f3463, %f3547; mul.ftz.f32 %f3549, %f3530, %f3474; fma.rn.ftz.f32 %f3550, %f3333, %f3485, %f3549; mul.ftz.f32 %f3551, %f3338, %f3550; fma.rn.ftz.f32 %f3552, %f3531, %f3548, %f3551; mul.ftz.f32 %f3553, %f3530, %f3496; fma.rn.ftz.f32 %f3554, %f3333, %f3507, %f3553; mul.ftz.f32 %f3555, %f3530, %f3518; fma.rn.ftz.f32 %f3556, %f3333, %f3529, %f3555; mul.ftz.f32 %f3557, %f3338, %f3556; fma.rn.ftz.f32 %f3558, %f3531, %f3554, %f3557; mul.ftz.f32 %f3559, %f3343, %f3558; fma.rn.ftz.f32 %f3560, %f3532, %f3552, %f3559; sub.ftz.f32 %f3561, %f3560, %f3546; fma.rn.ftz.f32 %f170, %f3348, %f3561, %f3546; mov.b32 %r41, %f170; setp.nan.ftz.f32 %p1149, %f170, %f170; mov.f32 %f5588, 0f00000000; @%p1149 bra $L__BB74_101; or.b32 %r6726, %r41, -2147483648; setp.ne.s32 %p1150, %r6726, -2147483648; add.ftz.f32 %f3563, %f170, %f170; setp.eq.ftz.f32 %p1151, %f3563, %f170; and.pred %p1152, %p1151, %p1150; @%p1152 bra $L__BB74_101; shl.b32 %r6727, %r41, 1; setp.lt.u32 %p1153, %r6727, -16777215; mul.ftz.f32 %f3564, %f170, 0f3F559B3D; selp.f32 %f5588, %f3564, 0f00000000, %p1153; $L__BB74_101: fma.rn.ftz.f32 %f3565, %f169, %f5588, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f3565; bra.uni $L__BB74_450; $L__BB74_439: cvt.rzi.ftz.s32.f32 %r479, %f26; setp.lt.s32 %p1974, %r479, 1; mov.f32 %f5827, 0f3F800000; mov.f32 %f5825, %f5827; @%p1974 bra $L__BB74_445; lg2.approx.ftz.f32 %f5466, %f27; mul.ftz.f32 %f5467, %f25, %f5466; neg.ftz.f32 %f5468, %f5467; ex2.approx.ftz.f32 %f795, %f5468; mov.f32 %f5825, 0f3F800000; mov.u32 %r10037, 0; mov.f32 %f5827, %f5825; $L__BB74_441: cvt.rmi.ftz.f32.f32 %f5470, %f5815; cvt.rzi.ftz.s32.f32 %r9924, %f5470; sub.ftz.f32 %f5471, %f5815, %f5470; mul.ftz.f32 %f5472, %f5471, %f5471; mul.ftz.f32 %f5473, %f5471, %f5472; fma.rn.ftz.f32 %f5474, %f5471, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5475, %f5471, %f5474, 0f41200000; mul.ftz.f32 %f5476, %f5473, %f5475; add.s32 %r9925, %r9924, -559038720; xor.b32 %r9926, %r9925, -1874868139; add.s32 %r9927, %r9926, 29184895; xor.b32 %r9928, %r9927, -559038720; shf.l.wrap.b32 %r9929, %r9927, %r9927, 25; sub.s32 %r9930, %r9928, %r9929; xor.b32 %r9931, %r9930, -1874868139; shf.l.wrap.b32 %r9932, %r9930, %r9930, 16; sub.s32 %r9933, %r9931, %r9932; xor.b32 %r9934, %r9933, %r9927; shf.l.wrap.b32 %r9935, %r9933, %r9933, 4; sub.s32 %r9936, %r9934, %r9935; xor.b32 %r9937, %r9936, %r9930; shf.l.wrap.b32 %r9938, %r9936, %r9936, 14; sub.s32 %r9939, %r9937, %r9938; xor.b32 %r9940, %r9939, %r9933; shf.l.wrap.b32 %r9941, %r9939, %r9939, 24; sub.s32 %r9942, %r9940, %r9941; and.b32 %r9943, %r9942, 7; add.s32 %r9944, %r9943, 1; cvt.rn.f32.s32 %f5477, %r9944; and.b32 %r9945, %r9942, 8; setp.eq.s32 %p1975, %r9945, 0; neg.ftz.f32 %f5478, %f5477; selp.f32 %f5479, %f5477, %f5478, %p1975; mul.ftz.f32 %f5480, %f5471, %f5479; add.s32 %r9946, %r9924, -559038719; xor.b32 %r9947, %r9946, -1874868139; add.s32 %r9948, %r9947, 29184895; xor.b32 %r9949, %r9948, -559038720; shf.l.wrap.b32 %r9950, %r9948, %r9948, 25; sub.s32 %r9951, %r9949, %r9950; xor.b32 %r9952, %r9951, -1874868139; shf.l.wrap.b32 %r9953, %r9951, %r9951, 16; sub.s32 %r9954, %r9952, %r9953; xor.b32 %r9955, %r9954, %r9948; shf.l.wrap.b32 %r9956, %r9954, %r9954, 4; sub.s32 %r9957, %r9955, %r9956; xor.b32 %r9958, %r9957, %r9951; shf.l.wrap.b32 %r9959, %r9957, %r9957, 14; sub.s32 %r9960, %r9958, %r9959; xor.b32 %r9961, %r9960, %r9954; shf.l.wrap.b32 %r9962, %r9960, %r9960, 24; sub.s32 %r9963, %r9961, %r9962; add.ftz.f32 %f5481, %f5471, 0fBF800000; and.b32 %r9964, %r9963, 7; add.s32 %r9965, %r9964, 1; cvt.rn.f32.s32 %f5482, %r9965; and.b32 %r9966, %r9963, 8; setp.eq.s32 %p1976, %r9966, 0; neg.ftz.f32 %f5483, %f5482; selp.f32 %f5484, %f5482, %f5483, %p1976; mul.ftz.f32 %f5485, %f5481, %f5484; sub.ftz.f32 %f5486, %f5485, %f5480; fma.rn.ftz.f32 %f799, %f5476, %f5486, %f5480; mov.b32 %r481, %f799; setp.nan.ftz.f32 %p1977, %f799, %f799; mov.f32 %f5822, 0f00000000; @%p1977 bra $L__BB74_444; or.b32 %r9967, %r481, -2147483648; setp.ne.s32 %p1978, %r9967, -2147483648; add.ftz.f32 %f5488, %f799, %f799; setp.eq.ftz.f32 %p1979, %f5488, %f799; and.pred %p1980, %p1979, %p1978; @%p1980 bra $L__BB74_444; shl.b32 %r9968, %r481, 1; setp.lt.u32 %p1981, %r9968, -16777215; selp.f32 %f5822, %f799, 0f00000000, %p1981; $L__BB74_444: mul.ftz.f32 %f5489, %f5822, 0f3E800000; fma.rn.ftz.f32 %f5490, %f5825, %f5489, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f5490; mul.ftz.f32 %f5825, %f795, %f5825; mul.ftz.f32 %f5815, %f27, %f5815; add.s32 %r10037, %r10037, 1; setp.lt.s32 %p1982, %r10037, %r479; @%p1982 bra $L__BB74_441; $L__BB74_445: cvt.rmi.ftz.f32.f32 %f5491, %f26; sub.ftz.f32 %f808, %f26, %f5491; setp.eq.ftz.f32 %p1983, %f808, 0f00000000; @%p1983 bra $L__BB74_450; mul.ftz.f32 %f809, %f5825, %f808; cvt.rmi.ftz.f32.f32 %f5493, %f5815; cvt.rzi.ftz.s32.f32 %r9969, %f5493; sub.ftz.f32 %f5494, %f5815, %f5493; mul.ftz.f32 %f5495, %f5494, %f5494; mul.ftz.f32 %f5496, %f5494, %f5495; fma.rn.ftz.f32 %f5497, %f5494, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f5498, %f5494, %f5497, 0f41200000; mul.ftz.f32 %f5499, %f5496, %f5498; add.s32 %r9970, %r9969, -559038720; xor.b32 %r9971, %r9970, -1874868139; add.s32 %r9972, %r9971, 29184895; xor.b32 %r9973, %r9972, -559038720; shf.l.wrap.b32 %r9974, %r9972, %r9972, 25; sub.s32 %r9975, %r9973, %r9974; xor.b32 %r9976, %r9975, -1874868139; shf.l.wrap.b32 %r9977, %r9975, %r9975, 16; sub.s32 %r9978, %r9976, %r9977; xor.b32 %r9979, %r9978, %r9972; shf.l.wrap.b32 %r9980, %r9978, %r9978, 4; sub.s32 %r9981, %r9979, %r9980; xor.b32 %r9982, %r9981, %r9975; shf.l.wrap.b32 %r9983, %r9981, %r9981, 14; sub.s32 %r9984, %r9982, %r9983; xor.b32 %r9985, %r9984, %r9978; shf.l.wrap.b32 %r9986, %r9984, %r9984, 24; sub.s32 %r9987, %r9985, %r9986; and.b32 %r9988, %r9987, 7; add.s32 %r9989, %r9988, 1; cvt.rn.f32.s32 %f5500, %r9989; and.b32 %r9990, %r9987, 8; setp.eq.s32 %p1984, %r9990, 0; neg.ftz.f32 %f5501, %f5500; selp.f32 %f5502, %f5500, %f5501, %p1984; mul.ftz.f32 %f5503, %f5494, %f5502; add.s32 %r9991, %r9969, -559038719; xor.b32 %r9992, %r9991, -1874868139; add.s32 %r9993, %r9992, 29184895; xor.b32 %r9994, %r9993, -559038720; shf.l.wrap.b32 %r9995, %r9993, %r9993, 25; sub.s32 %r9996, %r9994, %r9995; xor.b32 %r9997, %r9996, -1874868139; shf.l.wrap.b32 %r9998, %r9996, %r9996, 16; sub.s32 %r9999, %r9997, %r9998; xor.b32 %r10000, %r9999, %r9993; shf.l.wrap.b32 %r10001, %r9999, %r9999, 4; sub.s32 %r10002, %r10000, %r10001; xor.b32 %r10003, %r10002, %r9996; shf.l.wrap.b32 %r10004, %r10002, %r10002, 14; sub.s32 %r10005, %r10003, %r10004; xor.b32 %r10006, %r10005, %r9999; shf.l.wrap.b32 %r10007, %r10005, %r10005, 24; sub.s32 %r10008, %r10006, %r10007; add.ftz.f32 %f5504, %f5494, 0fBF800000; and.b32 %r10009, %r10008, 7; add.s32 %r10010, %r10009, 1; cvt.rn.f32.s32 %f5505, %r10010; and.b32 %r10011, %r10008, 8; setp.eq.s32 %p1985, %r10011, 0; neg.ftz.f32 %f5506, %f5505; selp.f32 %f5507, %f5505, %f5506, %p1985; mul.ftz.f32 %f5508, %f5504, %f5507; sub.ftz.f32 %f5509, %f5508, %f5503; fma.rn.ftz.f32 %f810, %f5499, %f5509, %f5503; mov.b32 %r483, %f810; setp.nan.ftz.f32 %p1986, %f810, %f810; mov.f32 %f5826, 0f00000000; @%p1986 bra $L__BB74_449; or.b32 %r10012, %r483, -2147483648; setp.ne.s32 %p1987, %r10012, -2147483648; add.ftz.f32 %f5511, %f810, %f810; setp.eq.ftz.f32 %p1988, %f5511, %f810; and.pred %p1989, %p1988, %p1987; @%p1989 bra $L__BB74_449; shl.b32 %r10013, %r483, 1; setp.lt.u32 %p1990, %r10013, -16777215; mul.ftz.f32 %f5512, %f810, 0f3E800000; selp.f32 %f5826, %f5512, 0f00000000, %p1990; $L__BB74_449: fma.rn.ftz.f32 %f5513, %f809, %f5826, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f5513; bra.uni $L__BB74_450; $L__BB74_297: cvt.rzi.ftz.s32.f32 %r357, %f26; setp.lt.s32 %p1582, %r357, 1; mov.f32 %f5827, 0f3F800000; mov.f32 %f5726, %f5827; @%p1582 bra $L__BB74_319; lg2.approx.ftz.f32 %f4349, %f27; mul.ftz.f32 %f4350, %f25, %f4349; neg.ftz.f32 %f4351, %f4350; ex2.approx.ftz.f32 %f523, %f4351; mov.f32 %f5726, 0f3F800000; mov.u32 %r10027, 0; mov.f32 %f5827, %f5726; $L__BB74_299: cvt.rmi.ftz.f32.f32 %f4352, %f5694; cvt.rzi.ftz.s32.f32 %r359, %f4352; sub.ftz.f32 %f529, %f5694, %f4352; cvt.rmi.ftz.f32.f32 %f4353, %f5695; cvt.rzi.ftz.s32.f32 %r360, %f4353; sub.ftz.f32 %f530, %f5695, %f4353; cvt.rmi.ftz.f32.f32 %f4354, %f5696; cvt.rzi.ftz.s32.f32 %r361, %f4354; sub.ftz.f32 %f531, %f5696, %f4354; fma.rn.ftz.f32 %f4355, %f529, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f532, %f529, %f4355, 0f41200000; fma.rn.ftz.f32 %f4356, %f530, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f533, %f530, %f4356, 0f41200000; fma.rn.ftz.f32 %f4357, %f531, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f534, %f531, %f4357, 0f41200000; add.s32 %r362, %r361, -559038712; add.s32 %r363, %r360, -559038712; add.s32 %r364, %r359, -559038712; xor.b32 %r8153, %r362, %r363; shf.l.wrap.b32 %r365, %r363, %r363, 14; sub.s32 %r366, %r8153, %r365; xor.b32 %r8154, %r366, %r364; shf.l.wrap.b32 %r367, %r366, %r366, 11; sub.s32 %r8155, %r8154, %r367; xor.b32 %r8156, %r8155, %r363; shf.l.wrap.b32 %r8157, %r8155, %r8155, 25; sub.s32 %r8158, %r8156, %r8157; xor.b32 %r8159, %r8158, %r366; shf.l.wrap.b32 %r8160, %r8158, %r8158, 16; sub.s32 %r8161, %r8159, %r8160; xor.b32 %r8162, %r8161, %r8155; shf.l.wrap.b32 %r8163, %r8161, %r8161, 4; sub.s32 %r8164, %r8162, %r8163; xor.b32 %r8165, %r8164, %r8158; shf.l.wrap.b32 %r8166, %r8164, %r8164, 14; sub.s32 %r8167, %r8165, %r8166; xor.b32 %r8168, %r8167, %r8161; shf.l.wrap.b32 %r8169, %r8167, %r8167, 24; sub.s32 %r368, %r8168, %r8169; and.b32 %r369, %r368, 15; cvt.u16.u32 %rs223, %r368; or.b16 %rs224, %rs223, 2; and.b16 %rs225, %rs224, 15; setp.eq.s16 %p1583, %rs225, 14; mov.f32 %f5713, %f529; @%p1583 bra $L__BB74_301; mov.f32 %f5713, %f531; $L__BB74_301: mul.ftz.f32 %f4358, %f529, %f529; mul.ftz.f32 %f4359, %f529, %f4358; mul.ftz.f32 %f536, %f4359, %f532; mul.ftz.f32 %f4360, %f530, %f530; mul.ftz.f32 %f4361, %f530, %f4360; mul.ftz.f32 %f537, %f4361, %f533; mul.ftz.f32 %f4362, %f531, %f531; mul.ftz.f32 %f4363, %f531, %f4362; mul.ftz.f32 %f538, %f4363, %f534; setp.lt.u32 %p1584, %r369, 8; selp.f32 %f4364, %f529, %f530, %p1584; setp.lt.u32 %p1585, %r369, 4; selp.f32 %f4365, %f530, %f5713, %p1585; and.b32 %r8170, %r368, 1; setp.eq.b32 %p1586, %r8170, 1; neg.ftz.f32 %f4366, %f4364; selp.f32 %f4367, %f4366, %f4364, %p1586; and.b32 %r8171, %r368, 2; setp.eq.s32 %p1587, %r8171, 0; neg.ftz.f32 %f4368, %f4365; selp.f32 %f4369, %f4365, %f4368, %p1587; add.ftz.f32 %f539, %f4367, %f4369; add.s32 %r370, %r359, -559038711; xor.b32 %r8172, %r366, %r370; sub.s32 %r8173, %r8172, %r367; xor.b32 %r8174, %r8173, %r363; shf.l.wrap.b32 %r8175, %r8173, %r8173, 25; sub.s32 %r8176, %r8174, %r8175; xor.b32 %r8177, %r8176, %r366; shf.l.wrap.b32 %r8178, %r8176, %r8176, 16; sub.s32 %r8179, %r8177, %r8178; xor.b32 %r8180, %r8179, %r8173; shf.l.wrap.b32 %r8181, %r8179, %r8179, 4; sub.s32 %r8182, %r8180, %r8181; xor.b32 %r8183, %r8182, %r8176; shf.l.wrap.b32 %r8184, %r8182, %r8182, 14; sub.s32 %r8185, %r8183, %r8184; xor.b32 %r8186, %r8185, %r8179; shf.l.wrap.b32 %r8187, %r8185, %r8185, 24; sub.s32 %r371, %r8186, %r8187; and.b32 %r372, %r371, 15; add.ftz.f32 %f540, %f529, 0fBF800000; cvt.u16.u32 %rs226, %r371; or.b16 %rs227, %rs226, 2; and.b16 %rs228, %rs227, 15; setp.eq.s16 %p1588, %rs228, 14; mov.f32 %f5714, %f540; @%p1588 bra $L__BB74_303; mov.f32 %f5714, %f531; $L__BB74_303: setp.lt.u32 %p1589, %r372, 8; selp.f32 %f4370, %f540, %f530, %p1589; setp.lt.u32 %p1590, %r372, 4; selp.f32 %f4371, %f530, %f5714, %p1590; and.b32 %r8188, %r371, 1; setp.eq.b32 %p1591, %r8188, 1; neg.ftz.f32 %f4372, %f4370; selp.f32 %f4373, %f4372, %f4370, %p1591; and.b32 %r8189, %r371, 2; setp.eq.s32 %p1592, %r8189, 0; neg.ftz.f32 %f4374, %f4371; selp.f32 %f4375, %f4371, %f4374, %p1592; add.ftz.f32 %f542, %f4373, %f4375; add.s32 %r373, %r360, -559038711; xor.b32 %r8190, %r362, %r373; shf.l.wrap.b32 %r374, %r373, %r373, 14; sub.s32 %r375, %r8190, %r374; xor.b32 %r8191, %r375, %r364; shf.l.wrap.b32 %r376, %r375, %r375, 11; sub.s32 %r8192, %r8191, %r376; xor.b32 %r8193, %r8192, %r373; shf.l.wrap.b32 %r8194, %r8192, %r8192, 25; sub.s32 %r8195, %r8193, %r8194; xor.b32 %r8196, %r8195, %r375; shf.l.wrap.b32 %r8197, %r8195, %r8195, 16; sub.s32 %r8198, %r8196, %r8197; xor.b32 %r8199, %r8198, %r8192; shf.l.wrap.b32 %r8200, %r8198, %r8198, 4; sub.s32 %r8201, %r8199, %r8200; xor.b32 %r8202, %r8201, %r8195; shf.l.wrap.b32 %r8203, %r8201, %r8201, 14; sub.s32 %r8204, %r8202, %r8203; xor.b32 %r8205, %r8204, %r8198; shf.l.wrap.b32 %r8206, %r8204, %r8204, 24; sub.s32 %r377, %r8205, %r8206; and.b32 %r378, %r377, 15; add.ftz.f32 %f543, %f530, 0fBF800000; cvt.u16.u32 %rs229, %r377; or.b16 %rs230, %rs229, 2; and.b16 %rs231, %rs230, 15; setp.eq.s16 %p1593, %rs231, 14; mov.f32 %f5715, %f529; @%p1593 bra $L__BB74_305; mov.f32 %f5715, %f531; $L__BB74_305: setp.lt.u32 %p1594, %r378, 8; selp.f32 %f4376, %f529, %f543, %p1594; setp.lt.u32 %p1595, %r378, 4; selp.f32 %f4377, %f543, %f5715, %p1595; and.b32 %r8207, %r377, 1; setp.eq.b32 %p1596, %r8207, 1; neg.ftz.f32 %f4378, %f4376; selp.f32 %f4379, %f4378, %f4376, %p1596; and.b32 %r8208, %r377, 2; setp.eq.s32 %p1597, %r8208, 0; neg.ftz.f32 %f4380, %f4377; selp.f32 %f4381, %f4377, %f4380, %p1597; add.ftz.f32 %f545, %f4379, %f4381; xor.b32 %r8209, %r375, %r370; sub.s32 %r8210, %r8209, %r376; xor.b32 %r8211, %r8210, %r373; shf.l.wrap.b32 %r8212, %r8210, %r8210, 25; sub.s32 %r8213, %r8211, %r8212; xor.b32 %r8214, %r8213, %r375; shf.l.wrap.b32 %r8215, %r8213, %r8213, 16; sub.s32 %r8216, %r8214, %r8215; xor.b32 %r8217, %r8216, %r8210; shf.l.wrap.b32 %r8218, %r8216, %r8216, 4; sub.s32 %r8219, %r8217, %r8218; xor.b32 %r8220, %r8219, %r8213; shf.l.wrap.b32 %r8221, %r8219, %r8219, 14; sub.s32 %r8222, %r8220, %r8221; xor.b32 %r8223, %r8222, %r8216; shf.l.wrap.b32 %r8224, %r8222, %r8222, 24; sub.s32 %r379, %r8223, %r8224; and.b32 %r380, %r379, 15; cvt.u16.u32 %rs232, %r379; or.b16 %rs233, %rs232, 2; and.b16 %rs234, %rs233, 15; setp.eq.s16 %p1598, %rs234, 14; mov.f32 %f5716, %f540; @%p1598 bra $L__BB74_307; mov.f32 %f5716, %f531; $L__BB74_307: setp.lt.u32 %p1599, %r380, 8; selp.f32 %f4382, %f540, %f543, %p1599; setp.lt.u32 %p1600, %r380, 4; selp.f32 %f4383, %f543, %f5716, %p1600; and.b32 %r8225, %r379, 1; setp.eq.b32 %p1601, %r8225, 1; neg.ftz.f32 %f4384, %f4382; selp.f32 %f4385, %f4384, %f4382, %p1601; and.b32 %r8226, %r379, 2; setp.eq.s32 %p1602, %r8226, 0; neg.ftz.f32 %f4386, %f4383; selp.f32 %f4387, %f4383, %f4386, %p1602; add.ftz.f32 %f547, %f4385, %f4387; add.s32 %r381, %r361, -559038711; xor.b32 %r8227, %r381, %r363; sub.s32 %r382, %r8227, %r365; xor.b32 %r8228, %r382, %r364; shf.l.wrap.b32 %r383, %r382, %r382, 11; sub.s32 %r8229, %r8228, %r383; xor.b32 %r8230, %r8229, %r363; shf.l.wrap.b32 %r8231, %r8229, %r8229, 25; sub.s32 %r8232, %r8230, %r8231; xor.b32 %r8233, %r8232, %r382; shf.l.wrap.b32 %r8234, %r8232, %r8232, 16; sub.s32 %r8235, %r8233, %r8234; xor.b32 %r8236, %r8235, %r8229; shf.l.wrap.b32 %r8237, %r8235, %r8235, 4; sub.s32 %r8238, %r8236, %r8237; xor.b32 %r8239, %r8238, %r8232; shf.l.wrap.b32 %r8240, %r8238, %r8238, 14; sub.s32 %r8241, %r8239, %r8240; xor.b32 %r8242, %r8241, %r8235; shf.l.wrap.b32 %r8243, %r8241, %r8241, 24; sub.s32 %r384, %r8242, %r8243; add.ftz.f32 %f548, %f531, 0fBF800000; and.b32 %r385, %r384, 15; cvt.u16.u32 %rs235, %r384; or.b16 %rs236, %rs235, 2; and.b16 %rs237, %rs236, 15; setp.eq.s16 %p1603, %rs237, 14; mov.f32 %f5717, %f529; @%p1603 bra $L__BB74_309; mov.f32 %f5717, %f548; $L__BB74_309: setp.lt.u32 %p1604, %r385, 8; selp.f32 %f4388, %f529, %f530, %p1604; setp.lt.u32 %p1605, %r385, 4; selp.f32 %f4389, %f530, %f5717, %p1605; and.b32 %r8244, %r384, 1; setp.eq.b32 %p1606, %r8244, 1; neg.ftz.f32 %f4390, %f4388; selp.f32 %f4391, %f4390, %f4388, %p1606; and.b32 %r8245, %r384, 2; setp.eq.s32 %p1607, %r8245, 0; neg.ftz.f32 %f4392, %f4389; selp.f32 %f4393, %f4389, %f4392, %p1607; add.ftz.f32 %f550, %f4391, %f4393; xor.b32 %r8246, %r382, %r370; sub.s32 %r8247, %r8246, %r383; xor.b32 %r8248, %r8247, %r363; shf.l.wrap.b32 %r8249, %r8247, %r8247, 25; sub.s32 %r8250, %r8248, %r8249; xor.b32 %r8251, %r8250, %r382; shf.l.wrap.b32 %r8252, %r8250, %r8250, 16; sub.s32 %r8253, %r8251, %r8252; xor.b32 %r8254, %r8253, %r8247; shf.l.wrap.b32 %r8255, %r8253, %r8253, 4; sub.s32 %r8256, %r8254, %r8255; xor.b32 %r8257, %r8256, %r8250; shf.l.wrap.b32 %r8258, %r8256, %r8256, 14; sub.s32 %r8259, %r8257, %r8258; xor.b32 %r8260, %r8259, %r8253; shf.l.wrap.b32 %r8261, %r8259, %r8259, 24; sub.s32 %r386, %r8260, %r8261; and.b32 %r387, %r386, 15; cvt.u16.u32 %rs238, %r386; or.b16 %rs239, %rs238, 2; and.b16 %rs240, %rs239, 15; setp.eq.s16 %p1608, %rs240, 14; mov.f32 %f5718, %f540; @%p1608 bra $L__BB74_311; mov.f32 %f5718, %f548; $L__BB74_311: setp.lt.u32 %p1609, %r387, 8; selp.f32 %f4394, %f540, %f530, %p1609; setp.lt.u32 %p1610, %r387, 4; selp.f32 %f4395, %f530, %f5718, %p1610; and.b32 %r8262, %r386, 1; setp.eq.b32 %p1611, %r8262, 1; neg.ftz.f32 %f4396, %f4394; selp.f32 %f4397, %f4396, %f4394, %p1611; and.b32 %r8263, %r386, 2; setp.eq.s32 %p1612, %r8263, 0; neg.ftz.f32 %f4398, %f4395; selp.f32 %f4399, %f4395, %f4398, %p1612; add.ftz.f32 %f552, %f4397, %f4399; xor.b32 %r8264, %r381, %r373; sub.s32 %r388, %r8264, %r374; xor.b32 %r8265, %r388, %r364; shf.l.wrap.b32 %r389, %r388, %r388, 11; sub.s32 %r8266, %r8265, %r389; xor.b32 %r8267, %r8266, %r373; shf.l.wrap.b32 %r8268, %r8266, %r8266, 25; sub.s32 %r8269, %r8267, %r8268; xor.b32 %r8270, %r8269, %r388; shf.l.wrap.b32 %r8271, %r8269, %r8269, 16; sub.s32 %r8272, %r8270, %r8271; xor.b32 %r8273, %r8272, %r8266; shf.l.wrap.b32 %r8274, %r8272, %r8272, 4; sub.s32 %r8275, %r8273, %r8274; xor.b32 %r8276, %r8275, %r8269; shf.l.wrap.b32 %r8277, %r8275, %r8275, 14; sub.s32 %r8278, %r8276, %r8277; xor.b32 %r8279, %r8278, %r8272; shf.l.wrap.b32 %r8280, %r8278, %r8278, 24; sub.s32 %r390, %r8279, %r8280; and.b32 %r391, %r390, 15; cvt.u16.u32 %rs241, %r390; or.b16 %rs242, %rs241, 2; and.b16 %rs243, %rs242, 15; setp.eq.s16 %p1613, %rs243, 14; mov.f32 %f5719, %f529; @%p1613 bra $L__BB74_313; mov.f32 %f5719, %f548; $L__BB74_313: setp.lt.u32 %p1614, %r391, 8; selp.f32 %f4400, %f529, %f543, %p1614; setp.lt.u32 %p1615, %r391, 4; selp.f32 %f4401, %f543, %f5719, %p1615; and.b32 %r8281, %r390, 1; setp.eq.b32 %p1616, %r8281, 1; neg.ftz.f32 %f4402, %f4400; selp.f32 %f4403, %f4402, %f4400, %p1616; and.b32 %r8282, %r390, 2; setp.eq.s32 %p1617, %r8282, 0; neg.ftz.f32 %f4404, %f4401; selp.f32 %f4405, %f4401, %f4404, %p1617; add.ftz.f32 %f554, %f4403, %f4405; xor.b32 %r8283, %r388, %r370; sub.s32 %r8284, %r8283, %r389; xor.b32 %r8285, %r8284, %r373; shf.l.wrap.b32 %r8286, %r8284, %r8284, 25; sub.s32 %r8287, %r8285, %r8286; xor.b32 %r8288, %r8287, %r388; shf.l.wrap.b32 %r8289, %r8287, %r8287, 16; sub.s32 %r8290, %r8288, %r8289; xor.b32 %r8291, %r8290, %r8284; shf.l.wrap.b32 %r8292, %r8290, %r8290, 4; sub.s32 %r8293, %r8291, %r8292; xor.b32 %r8294, %r8293, %r8287; shf.l.wrap.b32 %r8295, %r8293, %r8293, 14; sub.s32 %r8296, %r8294, %r8295; xor.b32 %r8297, %r8296, %r8290; shf.l.wrap.b32 %r8298, %r8296, %r8296, 24; sub.s32 %r392, %r8297, %r8298; and.b32 %r393, %r392, 15; cvt.u16.u32 %rs244, %r392; or.b16 %rs245, %rs244, 2; and.b16 %rs246, %rs245, 15; setp.eq.s16 %p1618, %rs246, 14; mov.f32 %f5720, %f540; @%p1618 bra $L__BB74_315; mov.f32 %f5720, %f548; $L__BB74_315: setp.lt.u32 %p1619, %r393, 8; selp.f32 %f4407, %f540, %f543, %p1619; setp.lt.u32 %p1620, %r393, 4; selp.f32 %f4408, %f543, %f5720, %p1620; and.b32 %r8299, %r392, 1; setp.eq.b32 %p1621, %r8299, 1; neg.ftz.f32 %f4409, %f4407; selp.f32 %f4410, %f4409, %f4407, %p1621; and.b32 %r8300, %r392, 2; setp.eq.s32 %p1622, %r8300, 0; neg.ftz.f32 %f4411, %f4408; selp.f32 %f4412, %f4408, %f4411, %p1622; add.ftz.f32 %f4413, %f4410, %f4412; mov.f32 %f4414, 0f3F800000; sub.ftz.f32 %f4415, %f4414, %f536; mul.ftz.f32 %f4416, %f536, %f542; fma.rn.ftz.f32 %f4417, %f4415, %f539, %f4416; sub.ftz.f32 %f4418, %f4414, %f537; mul.ftz.f32 %f4419, %f536, %f547; fma.rn.ftz.f32 %f4420, %f4415, %f545, %f4419; mul.ftz.f32 %f4421, %f537, %f4420; fma.rn.ftz.f32 %f4422, %f4418, %f4417, %f4421; sub.ftz.f32 %f4423, %f4414, %f538; mul.ftz.f32 %f4424, %f536, %f552; fma.rn.ftz.f32 %f4425, %f4415, %f550, %f4424; mul.ftz.f32 %f4426, %f536, %f4413; fma.rn.ftz.f32 %f4427, %f4415, %f554, %f4426; mul.ftz.f32 %f4428, %f537, %f4427; fma.rn.ftz.f32 %f4429, %f4418, %f4425, %f4428; mul.ftz.f32 %f4430, %f538, %f4429; fma.rn.ftz.f32 %f556, %f4423, %f4422, %f4430; mov.b32 %r394, %f556; setp.nan.ftz.f32 %p1623, %f556, %f556; mov.f32 %f5721, 0f00000000; @%p1623 bra $L__BB74_318; or.b32 %r8301, %r394, -2147483648; setp.ne.s32 %p1624, %r8301, -2147483648; add.ftz.f32 %f4432, %f556, %f556; setp.eq.ftz.f32 %p1625, %f4432, %f556; and.pred %p1626, %p1625, %p1624; @%p1626 bra $L__BB74_318; shl.b32 %r8302, %r394, 1; setp.lt.u32 %p1627, %r8302, -16777215; selp.f32 %f5721, %f556, 0f00000000, %p1627; $L__BB74_318: mul.ftz.f32 %f4433, %f5721, 0f3F7B645A; fma.rn.ftz.f32 %f4434, %f5726, %f4433, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f4434; mul.ftz.f32 %f5726, %f523, %f5726; mul.ftz.f32 %f5694, %f27, %f5694; mul.ftz.f32 %f5695, %f27, %f5695; mul.ftz.f32 %f5696, %f27, %f5696; add.s32 %r10027, %r10027, 1; setp.lt.s32 %p1628, %r10027, %r357; @%p1628 bra $L__BB74_299; $L__BB74_319: cvt.rmi.ftz.f32.f32 %f4435, %f26; sub.ftz.f32 %f569, %f26, %f4435; setp.eq.ftz.f32 %p1629, %f569, 0f00000000; @%p1629 bra $L__BB74_450; mul.ftz.f32 %f570, %f5726, %f569; cvt.rmi.ftz.f32.f32 %f4436, %f5694; cvt.rzi.ftz.s32.f32 %r396, %f4436; sub.ftz.f32 %f571, %f5694, %f4436; cvt.rmi.ftz.f32.f32 %f4437, %f5695; cvt.rzi.ftz.s32.f32 %r397, %f4437; sub.ftz.f32 %f572, %f5695, %f4437; cvt.rmi.ftz.f32.f32 %f4438, %f5696; cvt.rzi.ftz.s32.f32 %r398, %f4438; sub.ftz.f32 %f573, %f5696, %f4438; fma.rn.ftz.f32 %f4439, %f571, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f574, %f571, %f4439, 0f41200000; fma.rn.ftz.f32 %f4440, %f572, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f575, %f572, %f4440, 0f41200000; fma.rn.ftz.f32 %f4441, %f573, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f576, %f573, %f4441, 0f41200000; add.s32 %r399, %r398, -559038712; add.s32 %r400, %r397, -559038712; add.s32 %r401, %r396, -559038712; xor.b32 %r8303, %r399, %r400; shf.l.wrap.b32 %r402, %r400, %r400, 14; sub.s32 %r403, %r8303, %r402; xor.b32 %r8304, %r403, %r401; shf.l.wrap.b32 %r404, %r403, %r403, 11; sub.s32 %r8305, %r8304, %r404; xor.b32 %r8306, %r8305, %r400; shf.l.wrap.b32 %r8307, %r8305, %r8305, 25; sub.s32 %r8308, %r8306, %r8307; xor.b32 %r8309, %r8308, %r403; shf.l.wrap.b32 %r8310, %r8308, %r8308, 16; sub.s32 %r8311, %r8309, %r8310; xor.b32 %r8312, %r8311, %r8305; shf.l.wrap.b32 %r8313, %r8311, %r8311, 4; sub.s32 %r8314, %r8312, %r8313; xor.b32 %r8315, %r8314, %r8308; shf.l.wrap.b32 %r8316, %r8314, %r8314, 14; sub.s32 %r8317, %r8315, %r8316; xor.b32 %r8318, %r8317, %r8311; shf.l.wrap.b32 %r8319, %r8317, %r8317, 24; sub.s32 %r405, %r8318, %r8319; and.b32 %r406, %r405, 15; cvt.u16.u32 %rs247, %r405; or.b16 %rs248, %rs247, 2; and.b16 %rs249, %rs248, 15; setp.eq.s16 %p1630, %rs249, 14; mov.f32 %f5727, %f571; @%p1630 bra $L__BB74_322; mov.f32 %f5727, %f573; $L__BB74_322: mul.ftz.f32 %f4442, %f571, %f571; mul.ftz.f32 %f4443, %f571, %f4442; mul.ftz.f32 %f578, %f4443, %f574; mul.ftz.f32 %f4444, %f572, %f572; mul.ftz.f32 %f4445, %f572, %f4444; mul.ftz.f32 %f579, %f4445, %f575; mul.ftz.f32 %f4446, %f573, %f573; mul.ftz.f32 %f4447, %f573, %f4446; mul.ftz.f32 %f580, %f4447, %f576; setp.lt.u32 %p1631, %r406, 8; selp.f32 %f4448, %f571, %f572, %p1631; setp.lt.u32 %p1632, %r406, 4; selp.f32 %f4449, %f572, %f5727, %p1632; and.b32 %r8320, %r405, 1; setp.eq.b32 %p1633, %r8320, 1; neg.ftz.f32 %f4450, %f4448; selp.f32 %f4451, %f4450, %f4448, %p1633; and.b32 %r8321, %r405, 2; setp.eq.s32 %p1634, %r8321, 0; neg.ftz.f32 %f4452, %f4449; selp.f32 %f4453, %f4449, %f4452, %p1634; add.ftz.f32 %f581, %f4451, %f4453; add.s32 %r407, %r396, -559038711; xor.b32 %r8322, %r403, %r407; sub.s32 %r8323, %r8322, %r404; xor.b32 %r8324, %r8323, %r400; shf.l.wrap.b32 %r8325, %r8323, %r8323, 25; sub.s32 %r8326, %r8324, %r8325; xor.b32 %r8327, %r8326, %r403; shf.l.wrap.b32 %r8328, %r8326, %r8326, 16; sub.s32 %r8329, %r8327, %r8328; xor.b32 %r8330, %r8329, %r8323; shf.l.wrap.b32 %r8331, %r8329, %r8329, 4; sub.s32 %r8332, %r8330, %r8331; xor.b32 %r8333, %r8332, %r8326; shf.l.wrap.b32 %r8334, %r8332, %r8332, 14; sub.s32 %r8335, %r8333, %r8334; xor.b32 %r8336, %r8335, %r8329; shf.l.wrap.b32 %r8337, %r8335, %r8335, 24; sub.s32 %r408, %r8336, %r8337; and.b32 %r409, %r408, 15; add.ftz.f32 %f582, %f571, 0fBF800000; cvt.u16.u32 %rs250, %r408; or.b16 %rs251, %rs250, 2; and.b16 %rs252, %rs251, 15; setp.eq.s16 %p1635, %rs252, 14; mov.f32 %f5728, %f582; @%p1635 bra $L__BB74_324; mov.f32 %f5728, %f573; $L__BB74_324: setp.lt.u32 %p1636, %r409, 8; selp.f32 %f4454, %f582, %f572, %p1636; setp.lt.u32 %p1637, %r409, 4; selp.f32 %f4455, %f572, %f5728, %p1637; and.b32 %r8338, %r408, 1; setp.eq.b32 %p1638, %r8338, 1; neg.ftz.f32 %f4456, %f4454; selp.f32 %f4457, %f4456, %f4454, %p1638; and.b32 %r8339, %r408, 2; setp.eq.s32 %p1639, %r8339, 0; neg.ftz.f32 %f4458, %f4455; selp.f32 %f4459, %f4455, %f4458, %p1639; add.ftz.f32 %f584, %f4457, %f4459; add.s32 %r410, %r397, -559038711; xor.b32 %r8340, %r399, %r410; shf.l.wrap.b32 %r411, %r410, %r410, 14; sub.s32 %r412, %r8340, %r411; xor.b32 %r8341, %r412, %r401; shf.l.wrap.b32 %r413, %r412, %r412, 11; sub.s32 %r8342, %r8341, %r413; xor.b32 %r8343, %r8342, %r410; shf.l.wrap.b32 %r8344, %r8342, %r8342, 25; sub.s32 %r8345, %r8343, %r8344; xor.b32 %r8346, %r8345, %r412; shf.l.wrap.b32 %r8347, %r8345, %r8345, 16; sub.s32 %r8348, %r8346, %r8347; xor.b32 %r8349, %r8348, %r8342; shf.l.wrap.b32 %r8350, %r8348, %r8348, 4; sub.s32 %r8351, %r8349, %r8350; xor.b32 %r8352, %r8351, %r8345; shf.l.wrap.b32 %r8353, %r8351, %r8351, 14; sub.s32 %r8354, %r8352, %r8353; xor.b32 %r8355, %r8354, %r8348; shf.l.wrap.b32 %r8356, %r8354, %r8354, 24; sub.s32 %r414, %r8355, %r8356; and.b32 %r415, %r414, 15; add.ftz.f32 %f585, %f572, 0fBF800000; cvt.u16.u32 %rs253, %r414; or.b16 %rs254, %rs253, 2; and.b16 %rs255, %rs254, 15; setp.eq.s16 %p1640, %rs255, 14; mov.f32 %f5729, %f571; @%p1640 bra $L__BB74_326; mov.f32 %f5729, %f573; $L__BB74_326: setp.lt.u32 %p1641, %r415, 8; selp.f32 %f4460, %f571, %f585, %p1641; setp.lt.u32 %p1642, %r415, 4; selp.f32 %f4461, %f585, %f5729, %p1642; and.b32 %r8357, %r414, 1; setp.eq.b32 %p1643, %r8357, 1; neg.ftz.f32 %f4462, %f4460; selp.f32 %f4463, %f4462, %f4460, %p1643; and.b32 %r8358, %r414, 2; setp.eq.s32 %p1644, %r8358, 0; neg.ftz.f32 %f4464, %f4461; selp.f32 %f4465, %f4461, %f4464, %p1644; add.ftz.f32 %f587, %f4463, %f4465; xor.b32 %r8359, %r412, %r407; sub.s32 %r8360, %r8359, %r413; xor.b32 %r8361, %r8360, %r410; shf.l.wrap.b32 %r8362, %r8360, %r8360, 25; sub.s32 %r8363, %r8361, %r8362; xor.b32 %r8364, %r8363, %r412; shf.l.wrap.b32 %r8365, %r8363, %r8363, 16; sub.s32 %r8366, %r8364, %r8365; xor.b32 %r8367, %r8366, %r8360; shf.l.wrap.b32 %r8368, %r8366, %r8366, 4; sub.s32 %r8369, %r8367, %r8368; xor.b32 %r8370, %r8369, %r8363; shf.l.wrap.b32 %r8371, %r8369, %r8369, 14; sub.s32 %r8372, %r8370, %r8371; xor.b32 %r8373, %r8372, %r8366; shf.l.wrap.b32 %r8374, %r8372, %r8372, 24; sub.s32 %r416, %r8373, %r8374; and.b32 %r417, %r416, 15; cvt.u16.u32 %rs256, %r416; or.b16 %rs257, %rs256, 2; and.b16 %rs258, %rs257, 15; setp.eq.s16 %p1645, %rs258, 14; mov.f32 %f5730, %f582; @%p1645 bra $L__BB74_328; mov.f32 %f5730, %f573; $L__BB74_328: setp.lt.u32 %p1646, %r417, 8; selp.f32 %f4466, %f582, %f585, %p1646; setp.lt.u32 %p1647, %r417, 4; selp.f32 %f4467, %f585, %f5730, %p1647; and.b32 %r8375, %r416, 1; setp.eq.b32 %p1648, %r8375, 1; neg.ftz.f32 %f4468, %f4466; selp.f32 %f4469, %f4468, %f4466, %p1648; and.b32 %r8376, %r416, 2; setp.eq.s32 %p1649, %r8376, 0; neg.ftz.f32 %f4470, %f4467; selp.f32 %f4471, %f4467, %f4470, %p1649; add.ftz.f32 %f589, %f4469, %f4471; add.s32 %r418, %r398, -559038711; xor.b32 %r8377, %r418, %r400; sub.s32 %r419, %r8377, %r402; xor.b32 %r8378, %r419, %r401; shf.l.wrap.b32 %r420, %r419, %r419, 11; sub.s32 %r8379, %r8378, %r420; xor.b32 %r8380, %r8379, %r400; shf.l.wrap.b32 %r8381, %r8379, %r8379, 25; sub.s32 %r8382, %r8380, %r8381; xor.b32 %r8383, %r8382, %r419; shf.l.wrap.b32 %r8384, %r8382, %r8382, 16; sub.s32 %r8385, %r8383, %r8384; xor.b32 %r8386, %r8385, %r8379; shf.l.wrap.b32 %r8387, %r8385, %r8385, 4; sub.s32 %r8388, %r8386, %r8387; xor.b32 %r8389, %r8388, %r8382; shf.l.wrap.b32 %r8390, %r8388, %r8388, 14; sub.s32 %r8391, %r8389, %r8390; xor.b32 %r8392, %r8391, %r8385; shf.l.wrap.b32 %r8393, %r8391, %r8391, 24; sub.s32 %r421, %r8392, %r8393; add.ftz.f32 %f590, %f573, 0fBF800000; and.b32 %r422, %r421, 15; cvt.u16.u32 %rs259, %r421; or.b16 %rs260, %rs259, 2; and.b16 %rs261, %rs260, 15; setp.eq.s16 %p1650, %rs261, 14; mov.f32 %f5731, %f571; @%p1650 bra $L__BB74_330; mov.f32 %f5731, %f590; $L__BB74_330: setp.lt.u32 %p1651, %r422, 8; selp.f32 %f4472, %f571, %f572, %p1651; setp.lt.u32 %p1652, %r422, 4; selp.f32 %f4473, %f572, %f5731, %p1652; and.b32 %r8394, %r421, 1; setp.eq.b32 %p1653, %r8394, 1; neg.ftz.f32 %f4474, %f4472; selp.f32 %f4475, %f4474, %f4472, %p1653; and.b32 %r8395, %r421, 2; setp.eq.s32 %p1654, %r8395, 0; neg.ftz.f32 %f4476, %f4473; selp.f32 %f4477, %f4473, %f4476, %p1654; add.ftz.f32 %f592, %f4475, %f4477; xor.b32 %r8396, %r419, %r407; sub.s32 %r8397, %r8396, %r420; xor.b32 %r8398, %r8397, %r400; shf.l.wrap.b32 %r8399, %r8397, %r8397, 25; sub.s32 %r8400, %r8398, %r8399; xor.b32 %r8401, %r8400, %r419; shf.l.wrap.b32 %r8402, %r8400, %r8400, 16; sub.s32 %r8403, %r8401, %r8402; xor.b32 %r8404, %r8403, %r8397; shf.l.wrap.b32 %r8405, %r8403, %r8403, 4; sub.s32 %r8406, %r8404, %r8405; xor.b32 %r8407, %r8406, %r8400; shf.l.wrap.b32 %r8408, %r8406, %r8406, 14; sub.s32 %r8409, %r8407, %r8408; xor.b32 %r8410, %r8409, %r8403; shf.l.wrap.b32 %r8411, %r8409, %r8409, 24; sub.s32 %r423, %r8410, %r8411; and.b32 %r424, %r423, 15; cvt.u16.u32 %rs262, %r423; or.b16 %rs263, %rs262, 2; and.b16 %rs264, %rs263, 15; setp.eq.s16 %p1655, %rs264, 14; mov.f32 %f5732, %f582; @%p1655 bra $L__BB74_332; mov.f32 %f5732, %f590; $L__BB74_332: setp.lt.u32 %p1656, %r424, 8; selp.f32 %f4478, %f582, %f572, %p1656; setp.lt.u32 %p1657, %r424, 4; selp.f32 %f4479, %f572, %f5732, %p1657; and.b32 %r8412, %r423, 1; setp.eq.b32 %p1658, %r8412, 1; neg.ftz.f32 %f4480, %f4478; selp.f32 %f4481, %f4480, %f4478, %p1658; and.b32 %r8413, %r423, 2; setp.eq.s32 %p1659, %r8413, 0; neg.ftz.f32 %f4482, %f4479; selp.f32 %f4483, %f4479, %f4482, %p1659; add.ftz.f32 %f594, %f4481, %f4483; xor.b32 %r8414, %r418, %r410; sub.s32 %r425, %r8414, %r411; xor.b32 %r8415, %r425, %r401; shf.l.wrap.b32 %r426, %r425, %r425, 11; sub.s32 %r8416, %r8415, %r426; xor.b32 %r8417, %r8416, %r410; shf.l.wrap.b32 %r8418, %r8416, %r8416, 25; sub.s32 %r8419, %r8417, %r8418; xor.b32 %r8420, %r8419, %r425; shf.l.wrap.b32 %r8421, %r8419, %r8419, 16; sub.s32 %r8422, %r8420, %r8421; xor.b32 %r8423, %r8422, %r8416; shf.l.wrap.b32 %r8424, %r8422, %r8422, 4; sub.s32 %r8425, %r8423, %r8424; xor.b32 %r8426, %r8425, %r8419; shf.l.wrap.b32 %r8427, %r8425, %r8425, 14; sub.s32 %r8428, %r8426, %r8427; xor.b32 %r8429, %r8428, %r8422; shf.l.wrap.b32 %r8430, %r8428, %r8428, 24; sub.s32 %r427, %r8429, %r8430; and.b32 %r428, %r427, 15; cvt.u16.u32 %rs265, %r427; or.b16 %rs266, %rs265, 2; and.b16 %rs267, %rs266, 15; setp.eq.s16 %p1660, %rs267, 14; mov.f32 %f5733, %f571; @%p1660 bra $L__BB74_334; mov.f32 %f5733, %f590; $L__BB74_334: setp.lt.u32 %p1661, %r428, 8; selp.f32 %f4484, %f571, %f585, %p1661; setp.lt.u32 %p1662, %r428, 4; selp.f32 %f4485, %f585, %f5733, %p1662; and.b32 %r8431, %r427, 1; setp.eq.b32 %p1663, %r8431, 1; neg.ftz.f32 %f4486, %f4484; selp.f32 %f4487, %f4486, %f4484, %p1663; and.b32 %r8432, %r427, 2; setp.eq.s32 %p1664, %r8432, 0; neg.ftz.f32 %f4488, %f4485; selp.f32 %f4489, %f4485, %f4488, %p1664; add.ftz.f32 %f596, %f4487, %f4489; xor.b32 %r8433, %r425, %r407; sub.s32 %r8434, %r8433, %r426; xor.b32 %r8435, %r8434, %r410; shf.l.wrap.b32 %r8436, %r8434, %r8434, 25; sub.s32 %r8437, %r8435, %r8436; xor.b32 %r8438, %r8437, %r425; shf.l.wrap.b32 %r8439, %r8437, %r8437, 16; sub.s32 %r8440, %r8438, %r8439; xor.b32 %r8441, %r8440, %r8434; shf.l.wrap.b32 %r8442, %r8440, %r8440, 4; sub.s32 %r8443, %r8441, %r8442; xor.b32 %r8444, %r8443, %r8437; shf.l.wrap.b32 %r8445, %r8443, %r8443, 14; sub.s32 %r8446, %r8444, %r8445; xor.b32 %r8447, %r8446, %r8440; shf.l.wrap.b32 %r8448, %r8446, %r8446, 24; sub.s32 %r429, %r8447, %r8448; and.b32 %r430, %r429, 15; cvt.u16.u32 %rs268, %r429; or.b16 %rs269, %rs268, 2; and.b16 %rs270, %rs269, 15; setp.eq.s16 %p1665, %rs270, 14; mov.f32 %f5734, %f582; @%p1665 bra $L__BB74_336; mov.f32 %f5734, %f590; $L__BB74_336: setp.lt.u32 %p1666, %r430, 8; selp.f32 %f4491, %f582, %f585, %p1666; setp.lt.u32 %p1667, %r430, 4; selp.f32 %f4492, %f585, %f5734, %p1667; and.b32 %r8449, %r429, 1; setp.eq.b32 %p1668, %r8449, 1; neg.ftz.f32 %f4493, %f4491; selp.f32 %f4494, %f4493, %f4491, %p1668; and.b32 %r8450, %r429, 2; setp.eq.s32 %p1669, %r8450, 0; neg.ftz.f32 %f4495, %f4492; selp.f32 %f4496, %f4492, %f4495, %p1669; add.ftz.f32 %f4497, %f4494, %f4496; mov.f32 %f4498, 0f3F800000; sub.ftz.f32 %f4499, %f4498, %f578; mul.ftz.f32 %f4500, %f578, %f584; fma.rn.ftz.f32 %f4501, %f4499, %f581, %f4500; sub.ftz.f32 %f4502, %f4498, %f579; mul.ftz.f32 %f4503, %f578, %f589; fma.rn.ftz.f32 %f4504, %f4499, %f587, %f4503; mul.ftz.f32 %f4505, %f579, %f4504; fma.rn.ftz.f32 %f4506, %f4502, %f4501, %f4505; sub.ftz.f32 %f4507, %f4498, %f580; mul.ftz.f32 %f4508, %f578, %f594; fma.rn.ftz.f32 %f4509, %f4499, %f592, %f4508; mul.ftz.f32 %f4510, %f578, %f4497; fma.rn.ftz.f32 %f4511, %f4499, %f596, %f4510; mul.ftz.f32 %f4512, %f579, %f4511; fma.rn.ftz.f32 %f4513, %f4502, %f4509, %f4512; mul.ftz.f32 %f4514, %f580, %f4513; fma.rn.ftz.f32 %f598, %f4507, %f4506, %f4514; mov.b32 %r431, %f598; setp.nan.ftz.f32 %p1670, %f598, %f598; mov.f32 %f5735, 0f00000000; @%p1670 bra $L__BB74_339; or.b32 %r8451, %r431, -2147483648; setp.ne.s32 %p1671, %r8451, -2147483648; add.ftz.f32 %f4516, %f598, %f598; setp.eq.ftz.f32 %p1672, %f4516, %f598; and.pred %p1673, %p1672, %p1671; @%p1673 bra $L__BB74_339; shl.b32 %r8452, %r431, 1; setp.lt.u32 %p1674, %r8452, -16777215; mul.ftz.f32 %f4517, %f598, 0f3F7B645A; selp.f32 %f5735, %f4517, 0f00000000, %p1674; $L__BB74_339: fma.rn.ftz.f32 %f4518, %f570, %f5735, 0f3F800000; mul.ftz.f32 %f5827, %f5827, %f4518; $L__BB74_450: ld.param.u32 %r10017, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; ld.param.u32 %r10016, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u64 %rd28, [_Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; cvta.to.local.u64 %rd27, %rd28; shr.u32 %r10014, %r10016, 14; cvt.u64.u32 %rd24, %r10014; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd27, %rd25; st.local.f32 [%rd26], %f5827; add.s32 %r10015, %r10017, 2; st.param.b32 [func_retval0+0], %r10015; ret; } .func (.param .b32 func_retval0) _Z16fractal_noise_1dfff( .param .b32 _Z16fractal_noise_1dfff_param_0, .param .b32 _Z16fractal_noise_1dfff_param_1, .param .b32 _Z16fractal_noise_1dfff_param_2 ) { .reg .pred %p<18>; .reg .f32 %f<103>; .reg .b32 %r<98>; ld.param.f32 %f25, [_Z16fractal_noise_1dfff_param_0]; ld.param.f32 %f31, [_Z16fractal_noise_1dfff_param_1]; ld.param.f32 %f26, [_Z16fractal_noise_1dfff_param_2]; mov.f32 %f97, 0f00000000; max.ftz.f32 %f32, %f31, %f97; mov.f32 %f33, 0f41700000; min.ftz.f32 %f1, %f32, %f33; cvt.rzi.ftz.s32.f32 %r1, %f1; setp.lt.s32 %p1, %r1, 0; mov.f32 %f98, 0f3F800000; mov.f32 %f99, %f98; mov.f32 %f100, %f97; @%p1 bra $L__BB75_6; mov.f32 %f37, 0f00000000; max.ftz.f32 %f38, %f26, %f37; mov.f32 %f99, 0f3F800000; min.ftz.f32 %f2, %f38, %f99; mov.u32 %r97, -1; mov.f32 %f100, %f37; mov.f32 %f98, %f99; mov.f32 %f97, %f37; $L__BB75_2: mul.ftz.f32 %f40, %f99, %f25; cvt.rmi.ftz.f32.f32 %f41, %f40; cvt.rzi.ftz.s32.f32 %r7, %f41; sub.ftz.f32 %f42, %f40, %f41; mul.ftz.f32 %f43, %f42, %f42; mul.ftz.f32 %f44, %f42, %f43; fma.rn.ftz.f32 %f45, %f42, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f46, %f42, %f45, 0f41200000; mul.ftz.f32 %f47, %f44, %f46; add.s32 %r8, %r7, -559038720; xor.b32 %r9, %r8, -1874868139; add.s32 %r10, %r9, 29184895; xor.b32 %r11, %r10, -559038720; shf.l.wrap.b32 %r12, %r10, %r10, 25; sub.s32 %r13, %r11, %r12; xor.b32 %r14, %r13, -1874868139; shf.l.wrap.b32 %r15, %r13, %r13, 16; sub.s32 %r16, %r14, %r15; xor.b32 %r17, %r16, %r10; shf.l.wrap.b32 %r18, %r16, %r16, 4; sub.s32 %r19, %r17, %r18; xor.b32 %r20, %r19, %r13; shf.l.wrap.b32 %r21, %r19, %r19, 14; sub.s32 %r22, %r20, %r21; xor.b32 %r23, %r22, %r16; shf.l.wrap.b32 %r24, %r22, %r22, 24; sub.s32 %r25, %r23, %r24; and.b32 %r26, %r25, 7; add.s32 %r27, %r26, 1; cvt.rn.f32.s32 %f48, %r27; and.b32 %r28, %r25, 8; setp.eq.s32 %p2, %r28, 0; neg.ftz.f32 %f49, %f48; selp.f32 %f50, %f48, %f49, %p2; mul.ftz.f32 %f51, %f42, %f50; add.s32 %r29, %r7, -559038719; xor.b32 %r30, %r29, -1874868139; add.s32 %r31, %r30, 29184895; xor.b32 %r32, %r31, -559038720; shf.l.wrap.b32 %r33, %r31, %r31, 25; sub.s32 %r34, %r32, %r33; xor.b32 %r35, %r34, -1874868139; shf.l.wrap.b32 %r36, %r34, %r34, 16; sub.s32 %r37, %r35, %r36; xor.b32 %r38, %r37, %r31; shf.l.wrap.b32 %r39, %r37, %r37, 4; sub.s32 %r40, %r38, %r39; xor.b32 %r41, %r40, %r34; shf.l.wrap.b32 %r42, %r40, %r40, 14; sub.s32 %r43, %r41, %r42; xor.b32 %r44, %r43, %r37; shf.l.wrap.b32 %r45, %r43, %r43, 24; sub.s32 %r46, %r44, %r45; add.ftz.f32 %f52, %f42, 0fBF800000; and.b32 %r47, %r46, 7; add.s32 %r48, %r47, 1; cvt.rn.f32.s32 %f53, %r48; and.b32 %r49, %r46, 8; setp.eq.s32 %p3, %r49, 0; neg.ftz.f32 %f54, %f53; selp.f32 %f55, %f53, %f54, %p3; mul.ftz.f32 %f56, %f52, %f55; sub.ftz.f32 %f57, %f56, %f51; fma.rn.ftz.f32 %f7, %f47, %f57, %f51; mov.b32 %r3, %f7; setp.nan.ftz.f32 %p4, %f7, %f7; mov.f32 %f96, %f37; @%p4 bra $L__BB75_5; or.b32 %r50, %r3, -2147483648; setp.ne.s32 %p5, %r50, -2147483648; add.ftz.f32 %f59, %f7, %f7; setp.eq.ftz.f32 %p6, %f59, %f7; and.pred %p7, %p6, %p5; mov.f32 %f96, %f37; @%p7 bra $L__BB75_5; shl.b32 %r51, %r3, 1; setp.lt.u32 %p8, %r51, -16777215; selp.f32 %f96, %f7, 0f00000000, %p8; $L__BB75_5: mul.ftz.f32 %f60, %f96, 0f3E800000; fma.rn.ftz.f32 %f61, %f60, 0f3F000000, 0f3F000000; fma.rn.ftz.f32 %f100, %f98, %f61, %f100; add.ftz.f32 %f97, %f97, %f98; mul.ftz.f32 %f98, %f98, %f2; add.ftz.f32 %f99, %f99, %f99; add.s32 %r97, %r97, 1; setp.lt.s32 %p9, %r97, %r1; @%p9 bra $L__BB75_2; $L__BB75_6: cvt.rmi.ftz.f32.f32 %f62, %f1; sub.ftz.f32 %f18, %f1, %f62; setp.eq.ftz.f32 %p10, %f18, 0f00000000; div.approx.ftz.f32 %f102, %f100, %f97; @%p10 bra $L__BB75_11; mul.ftz.f32 %f64, %f99, %f25; cvt.rmi.ftz.f32.f32 %f65, %f64; cvt.rzi.ftz.s32.f32 %r52, %f65; sub.ftz.f32 %f66, %f64, %f65; mul.ftz.f32 %f67, %f66, %f66; mul.ftz.f32 %f68, %f66, %f67; fma.rn.ftz.f32 %f69, %f66, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f70, %f66, %f69, 0f41200000; mul.ftz.f32 %f71, %f68, %f70; add.s32 %r53, %r52, -559038720; xor.b32 %r54, %r53, -1874868139; add.s32 %r55, %r54, 29184895; xor.b32 %r56, %r55, -559038720; shf.l.wrap.b32 %r57, %r55, %r55, 25; sub.s32 %r58, %r56, %r57; xor.b32 %r59, %r58, -1874868139; shf.l.wrap.b32 %r60, %r58, %r58, 16; sub.s32 %r61, %r59, %r60; xor.b32 %r62, %r61, %r55; shf.l.wrap.b32 %r63, %r61, %r61, 4; sub.s32 %r64, %r62, %r63; xor.b32 %r65, %r64, %r58; shf.l.wrap.b32 %r66, %r64, %r64, 14; sub.s32 %r67, %r65, %r66; xor.b32 %r68, %r67, %r61; shf.l.wrap.b32 %r69, %r67, %r67, 24; sub.s32 %r70, %r68, %r69; and.b32 %r71, %r70, 7; add.s32 %r72, %r71, 1; cvt.rn.f32.s32 %f72, %r72; and.b32 %r73, %r70, 8; setp.eq.s32 %p11, %r73, 0; neg.ftz.f32 %f73, %f72; selp.f32 %f74, %f72, %f73, %p11; mul.ftz.f32 %f75, %f66, %f74; add.s32 %r74, %r52, -559038719; xor.b32 %r75, %r74, -1874868139; add.s32 %r76, %r75, 29184895; xor.b32 %r77, %r76, -559038720; shf.l.wrap.b32 %r78, %r76, %r76, 25; sub.s32 %r79, %r77, %r78; xor.b32 %r80, %r79, -1874868139; shf.l.wrap.b32 %r81, %r79, %r79, 16; sub.s32 %r82, %r80, %r81; xor.b32 %r83, %r82, %r76; shf.l.wrap.b32 %r84, %r82, %r82, 4; sub.s32 %r85, %r83, %r84; xor.b32 %r86, %r85, %r79; shf.l.wrap.b32 %r87, %r85, %r85, 14; sub.s32 %r88, %r86, %r87; xor.b32 %r89, %r88, %r82; shf.l.wrap.b32 %r90, %r88, %r88, 24; sub.s32 %r91, %r89, %r90; add.ftz.f32 %f76, %f66, 0fBF800000; and.b32 %r92, %r91, 7; add.s32 %r93, %r92, 1; cvt.rn.f32.s32 %f77, %r93; and.b32 %r94, %r91, 8; setp.eq.s32 %p12, %r94, 0; neg.ftz.f32 %f78, %f77; selp.f32 %f79, %f77, %f78, %p12; mul.ftz.f32 %f80, %f76, %f79; sub.ftz.f32 %f81, %f80, %f75; fma.rn.ftz.f32 %f20, %f71, %f81, %f75; mov.b32 %r5, %f20; setp.nan.ftz.f32 %p13, %f20, %f20; mov.f32 %f101, 0f3F000000; @%p13 bra $L__BB75_10; or.b32 %r95, %r5, -2147483648; setp.ne.s32 %p14, %r95, -2147483648; add.ftz.f32 %f83, %f20, %f20; setp.eq.ftz.f32 %p15, %f83, %f20; and.pred %p16, %p15, %p14; @%p16 bra $L__BB75_10; shl.b32 %r96, %r5, 1; setp.lt.u32 %p17, %r96, -16777215; mul.ftz.f32 %f84, %f20, 0f3E800000; fma.rn.ftz.f32 %f85, %f84, 0f3F000000, 0f3F000000; selp.f32 %f101, %f85, 0f3F000000, %p17; $L__BB75_10: fma.rn.ftz.f32 %f86, %f98, %f101, %f100; add.ftz.f32 %f87, %f97, %f98; div.approx.ftz.f32 %f88, %f86, %f87; mov.f32 %f89, 0f3F800000; sub.ftz.f32 %f90, %f89, %f18; mul.ftz.f32 %f91, %f18, %f88; fma.rn.ftz.f32 %f102, %f102, %f90, %f91; $L__BB75_11: st.param.f32 [func_retval0+0], %f102; ret; } .func (.param .b32 func_retval0) _Z16fractal_noise_2d6float2ff( .param .align 8 .b8 _Z16fractal_noise_2d6float2ff_param_0[8], .param .b32 _Z16fractal_noise_2d6float2ff_param_1, .param .b32 _Z16fractal_noise_2d6float2ff_param_2 ) { .reg .pred %p<38>; .reg .f32 %f<184>; .reg .b32 %r<200>; ld.param.f32 %f32, [_Z16fractal_noise_2d6float2ff_param_1]; ld.param.f32 %f27, [_Z16fractal_noise_2d6float2ff_param_2]; ld.param.f32 %f1, [_Z16fractal_noise_2d6float2ff_param_0]; ld.param.f32 %f2, [_Z16fractal_noise_2d6float2ff_param_0+4]; mov.f32 %f178, 0f00000000; max.ftz.f32 %f33, %f32, %f178; mov.f32 %f34, 0f41700000; min.ftz.f32 %f3, %f33, %f34; cvt.rzi.ftz.s32.f32 %r1, %f3; setp.lt.s32 %p1, %r1, 0; mov.f32 %f180, 0f3F800000; mov.f32 %f179, %f178; mov.f32 %f181, %f180; @%p1 bra $L__BB76_6; mov.f32 %f38, 0f00000000; max.ftz.f32 %f39, %f27, %f38; mov.f32 %f36, 0f3F800000; min.ftz.f32 %f4, %f39, %f36; mov.u32 %r199, -1; mov.f32 %f181, %f36; mov.f32 %f180, %f36; mov.f32 %f179, %f38; mov.f32 %f178, %f38; $L__BB76_2: mul.ftz.f32 %f41, %f1, %f181; cvt.rmi.ftz.f32.f32 %f42, %f41; cvt.rzi.ftz.s32.f32 %r7, %f42; sub.ftz.f32 %f43, %f41, %f42; mul.ftz.f32 %f44, %f2, %f181; cvt.rmi.ftz.f32.f32 %f45, %f44; cvt.rzi.ftz.s32.f32 %r8, %f45; sub.ftz.f32 %f46, %f44, %f45; mul.ftz.f32 %f47, %f43, %f43; mul.ftz.f32 %f48, %f43, %f47; fma.rn.ftz.f32 %f49, %f43, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f50, %f43, %f49, 0f41200000; mul.ftz.f32 %f51, %f48, %f50; mul.ftz.f32 %f52, %f46, %f46; mul.ftz.f32 %f53, %f46, %f52; fma.rn.ftz.f32 %f54, %f46, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f55, %f46, %f54, 0f41200000; mul.ftz.f32 %f56, %f53, %f55; add.s32 %r9, %r8, -559038716; add.s32 %r10, %r7, -559038716; xor.b32 %r11, %r9, -559038716; shf.l.wrap.b32 %r12, %r9, %r9, 14; sub.s32 %r13, %r11, %r12; xor.b32 %r14, %r13, %r10; shf.l.wrap.b32 %r15, %r13, %r13, 11; sub.s32 %r16, %r14, %r15; xor.b32 %r17, %r16, %r9; shf.l.wrap.b32 %r18, %r16, %r16, 25; sub.s32 %r19, %r17, %r18; xor.b32 %r20, %r19, %r13; shf.l.wrap.b32 %r21, %r19, %r19, 16; sub.s32 %r22, %r20, %r21; xor.b32 %r23, %r22, %r16; shf.l.wrap.b32 %r24, %r22, %r22, 4; sub.s32 %r25, %r23, %r24; xor.b32 %r26, %r25, %r19; shf.l.wrap.b32 %r27, %r25, %r25, 14; sub.s32 %r28, %r26, %r27; xor.b32 %r29, %r28, %r22; shf.l.wrap.b32 %r30, %r28, %r28, 24; sub.s32 %r31, %r29, %r30; and.b32 %r32, %r31, 4; setp.eq.s32 %p2, %r32, 0; selp.f32 %f57, %f43, %f46, %p2; selp.f32 %f58, %f46, %f43, %p2; add.ftz.f32 %f59, %f58, %f58; and.b32 %r33, %r31, 1; setp.eq.b32 %p3, %r33, 1; neg.ftz.f32 %f60, %f57; selp.f32 %f61, %f60, %f57, %p3; and.b32 %r34, %r31, 2; setp.eq.s32 %p4, %r34, 0; neg.ftz.f32 %f62, %f59; selp.f32 %f63, %f59, %f62, %p4; add.ftz.f32 %f64, %f61, %f63; add.s32 %r35, %r7, -559038715; xor.b32 %r36, %r13, %r35; sub.s32 %r37, %r36, %r15; xor.b32 %r38, %r37, %r9; shf.l.wrap.b32 %r39, %r37, %r37, 25; sub.s32 %r40, %r38, %r39; xor.b32 %r41, %r40, %r13; shf.l.wrap.b32 %r42, %r40, %r40, 16; sub.s32 %r43, %r41, %r42; xor.b32 %r44, %r43, %r37; shf.l.wrap.b32 %r45, %r43, %r43, 4; sub.s32 %r46, %r44, %r45; xor.b32 %r47, %r46, %r40; shf.l.wrap.b32 %r48, %r46, %r46, 14; sub.s32 %r49, %r47, %r48; xor.b32 %r50, %r49, %r43; shf.l.wrap.b32 %r51, %r49, %r49, 24; sub.s32 %r52, %r50, %r51; add.ftz.f32 %f65, %f43, 0fBF800000; and.b32 %r53, %r52, 4; setp.eq.s32 %p5, %r53, 0; selp.f32 %f67, %f65, %f46, %p5; selp.f32 %f68, %f46, %f65, %p5; add.ftz.f32 %f69, %f68, %f68; and.b32 %r54, %r52, 1; setp.eq.b32 %p6, %r54, 1; neg.ftz.f32 %f70, %f67; selp.f32 %f71, %f70, %f67, %p6; and.b32 %r55, %r52, 2; setp.eq.s32 %p7, %r55, 0; neg.ftz.f32 %f72, %f69; selp.f32 %f73, %f69, %f72, %p7; add.ftz.f32 %f74, %f71, %f73; add.s32 %r56, %r8, -559038715; xor.b32 %r57, %r56, -559038716; shf.l.wrap.b32 %r58, %r56, %r56, 14; sub.s32 %r59, %r57, %r58; xor.b32 %r60, %r59, %r10; shf.l.wrap.b32 %r61, %r59, %r59, 11; sub.s32 %r62, %r60, %r61; xor.b32 %r63, %r62, %r56; shf.l.wrap.b32 %r64, %r62, %r62, 25; sub.s32 %r65, %r63, %r64; xor.b32 %r66, %r65, %r59; shf.l.wrap.b32 %r67, %r65, %r65, 16; sub.s32 %r68, %r66, %r67; xor.b32 %r69, %r68, %r62; shf.l.wrap.b32 %r70, %r68, %r68, 4; sub.s32 %r71, %r69, %r70; xor.b32 %r72, %r71, %r65; shf.l.wrap.b32 %r73, %r71, %r71, 14; sub.s32 %r74, %r72, %r73; xor.b32 %r75, %r74, %r68; shf.l.wrap.b32 %r76, %r74, %r74, 24; sub.s32 %r77, %r75, %r76; add.ftz.f32 %f75, %f46, 0fBF800000; and.b32 %r78, %r77, 4; setp.eq.s32 %p8, %r78, 0; selp.f32 %f76, %f43, %f75, %p8; selp.f32 %f77, %f75, %f43, %p8; add.ftz.f32 %f78, %f77, %f77; and.b32 %r79, %r77, 1; setp.eq.b32 %p9, %r79, 1; neg.ftz.f32 %f79, %f76; selp.f32 %f80, %f79, %f76, %p9; and.b32 %r80, %r77, 2; setp.eq.s32 %p10, %r80, 0; neg.ftz.f32 %f81, %f78; selp.f32 %f82, %f78, %f81, %p10; add.ftz.f32 %f83, %f80, %f82; xor.b32 %r81, %r59, %r35; sub.s32 %r82, %r81, %r61; xor.b32 %r83, %r82, %r56; shf.l.wrap.b32 %r84, %r82, %r82, 25; sub.s32 %r85, %r83, %r84; xor.b32 %r86, %r85, %r59; shf.l.wrap.b32 %r87, %r85, %r85, 16; sub.s32 %r88, %r86, %r87; xor.b32 %r89, %r88, %r82; shf.l.wrap.b32 %r90, %r88, %r88, 4; sub.s32 %r91, %r89, %r90; xor.b32 %r92, %r91, %r85; shf.l.wrap.b32 %r93, %r91, %r91, 14; sub.s32 %r94, %r92, %r93; xor.b32 %r95, %r94, %r88; shf.l.wrap.b32 %r96, %r94, %r94, 24; sub.s32 %r97, %r95, %r96; and.b32 %r98, %r97, 4; setp.eq.s32 %p11, %r98, 0; selp.f32 %f84, %f65, %f75, %p11; selp.f32 %f85, %f75, %f65, %p11; add.ftz.f32 %f86, %f85, %f85; and.b32 %r99, %r97, 1; setp.eq.b32 %p12, %r99, 1; neg.ftz.f32 %f87, %f84; selp.f32 %f88, %f87, %f84, %p12; and.b32 %r100, %r97, 2; setp.eq.s32 %p13, %r100, 0; neg.ftz.f32 %f89, %f86; selp.f32 %f90, %f86, %f89, %p13; add.ftz.f32 %f91, %f88, %f90; sub.ftz.f32 %f92, %f36, %f51; sub.ftz.f32 %f93, %f36, %f56; mul.ftz.f32 %f94, %f92, %f64; fma.rn.ftz.f32 %f95, %f51, %f74, %f94; mul.ftz.f32 %f96, %f92, %f83; fma.rn.ftz.f32 %f97, %f51, %f91, %f96; mul.ftz.f32 %f98, %f56, %f97; fma.rn.ftz.f32 %f9, %f93, %f95, %f98; mov.b32 %r3, %f9; setp.nan.ftz.f32 %p14, %f9, %f9; mov.f32 %f177, %f38; @%p14 bra $L__BB76_5; or.b32 %r101, %r3, -2147483648; setp.ne.s32 %p15, %r101, -2147483648; add.ftz.f32 %f100, %f9, %f9; setp.eq.ftz.f32 %p16, %f100, %f9; and.pred %p17, %p16, %p15; mov.f32 %f177, %f38; @%p17 bra $L__BB76_5; shl.b32 %r102, %r3, 1; setp.lt.u32 %p18, %r102, -16777215; selp.f32 %f177, %f9, 0f00000000, %p18; $L__BB76_5: mul.ftz.f32 %f101, %f177, 0f3F295E9E; fma.rn.ftz.f32 %f102, %f101, 0f3F000000, 0f3F000000; fma.rn.ftz.f32 %f178, %f180, %f102, %f178; add.ftz.f32 %f179, %f179, %f180; mul.ftz.f32 %f180, %f180, %f4; add.ftz.f32 %f181, %f181, %f181; add.s32 %r199, %r199, 1; setp.lt.s32 %p19, %r199, %r1; @%p19 bra $L__BB76_2; $L__BB76_6: cvt.rmi.ftz.f32.f32 %f103, %f3; sub.ftz.f32 %f20, %f3, %f103; setp.eq.ftz.f32 %p20, %f20, 0f00000000; div.approx.ftz.f32 %f183, %f178, %f179; @%p20 bra $L__BB76_11; mul.ftz.f32 %f105, %f1, %f181; cvt.rmi.ftz.f32.f32 %f106, %f105; cvt.rzi.ftz.s32.f32 %r103, %f106; sub.ftz.f32 %f107, %f105, %f106; mul.ftz.f32 %f108, %f2, %f181; cvt.rmi.ftz.f32.f32 %f109, %f108; cvt.rzi.ftz.s32.f32 %r104, %f109; sub.ftz.f32 %f110, %f108, %f109; mul.ftz.f32 %f111, %f107, %f107; mul.ftz.f32 %f112, %f107, %f111; fma.rn.ftz.f32 %f113, %f107, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f114, %f107, %f113, 0f41200000; mul.ftz.f32 %f115, %f112, %f114; mul.ftz.f32 %f116, %f110, %f110; mul.ftz.f32 %f117, %f110, %f116; fma.rn.ftz.f32 %f118, %f110, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f119, %f110, %f118, 0f41200000; mul.ftz.f32 %f120, %f117, %f119; add.s32 %r105, %r104, -559038716; add.s32 %r106, %r103, -559038716; xor.b32 %r107, %r105, -559038716; shf.l.wrap.b32 %r108, %r105, %r105, 14; sub.s32 %r109, %r107, %r108; xor.b32 %r110, %r109, %r106; shf.l.wrap.b32 %r111, %r109, %r109, 11; sub.s32 %r112, %r110, %r111; xor.b32 %r113, %r112, %r105; shf.l.wrap.b32 %r114, %r112, %r112, 25; sub.s32 %r115, %r113, %r114; xor.b32 %r116, %r115, %r109; shf.l.wrap.b32 %r117, %r115, %r115, 16; sub.s32 %r118, %r116, %r117; xor.b32 %r119, %r118, %r112; shf.l.wrap.b32 %r120, %r118, %r118, 4; sub.s32 %r121, %r119, %r120; xor.b32 %r122, %r121, %r115; shf.l.wrap.b32 %r123, %r121, %r121, 14; sub.s32 %r124, %r122, %r123; xor.b32 %r125, %r124, %r118; shf.l.wrap.b32 %r126, %r124, %r124, 24; sub.s32 %r127, %r125, %r126; and.b32 %r128, %r127, 4; setp.eq.s32 %p21, %r128, 0; selp.f32 %f121, %f107, %f110, %p21; selp.f32 %f122, %f110, %f107, %p21; add.ftz.f32 %f123, %f122, %f122; and.b32 %r129, %r127, 1; setp.eq.b32 %p22, %r129, 1; neg.ftz.f32 %f124, %f121; selp.f32 %f125, %f124, %f121, %p22; and.b32 %r130, %r127, 2; setp.eq.s32 %p23, %r130, 0; neg.ftz.f32 %f126, %f123; selp.f32 %f127, %f123, %f126, %p23; add.ftz.f32 %f128, %f125, %f127; add.s32 %r131, %r103, -559038715; xor.b32 %r132, %r109, %r131; sub.s32 %r133, %r132, %r111; xor.b32 %r134, %r133, %r105; shf.l.wrap.b32 %r135, %r133, %r133, 25; sub.s32 %r136, %r134, %r135; xor.b32 %r137, %r136, %r109; shf.l.wrap.b32 %r138, %r136, %r136, 16; sub.s32 %r139, %r137, %r138; xor.b32 %r140, %r139, %r133; shf.l.wrap.b32 %r141, %r139, %r139, 4; sub.s32 %r142, %r140, %r141; xor.b32 %r143, %r142, %r136; shf.l.wrap.b32 %r144, %r142, %r142, 14; sub.s32 %r145, %r143, %r144; xor.b32 %r146, %r145, %r139; shf.l.wrap.b32 %r147, %r145, %r145, 24; sub.s32 %r148, %r146, %r147; add.ftz.f32 %f129, %f107, 0fBF800000; mov.f32 %f130, 0f3F800000; and.b32 %r149, %r148, 4; setp.eq.s32 %p24, %r149, 0; selp.f32 %f131, %f129, %f110, %p24; selp.f32 %f132, %f110, %f129, %p24; add.ftz.f32 %f133, %f132, %f132; and.b32 %r150, %r148, 1; setp.eq.b32 %p25, %r150, 1; neg.ftz.f32 %f134, %f131; selp.f32 %f135, %f134, %f131, %p25; and.b32 %r151, %r148, 2; setp.eq.s32 %p26, %r151, 0; neg.ftz.f32 %f136, %f133; selp.f32 %f137, %f133, %f136, %p26; add.ftz.f32 %f138, %f135, %f137; add.s32 %r152, %r104, -559038715; xor.b32 %r153, %r152, -559038716; shf.l.wrap.b32 %r154, %r152, %r152, 14; sub.s32 %r155, %r153, %r154; xor.b32 %r156, %r155, %r106; shf.l.wrap.b32 %r157, %r155, %r155, 11; sub.s32 %r158, %r156, %r157; xor.b32 %r159, %r158, %r152; shf.l.wrap.b32 %r160, %r158, %r158, 25; sub.s32 %r161, %r159, %r160; xor.b32 %r162, %r161, %r155; shf.l.wrap.b32 %r163, %r161, %r161, 16; sub.s32 %r164, %r162, %r163; xor.b32 %r165, %r164, %r158; shf.l.wrap.b32 %r166, %r164, %r164, 4; sub.s32 %r167, %r165, %r166; xor.b32 %r168, %r167, %r161; shf.l.wrap.b32 %r169, %r167, %r167, 14; sub.s32 %r170, %r168, %r169; xor.b32 %r171, %r170, %r164; shf.l.wrap.b32 %r172, %r170, %r170, 24; sub.s32 %r173, %r171, %r172; add.ftz.f32 %f139, %f110, 0fBF800000; and.b32 %r174, %r173, 4; setp.eq.s32 %p27, %r174, 0; selp.f32 %f140, %f107, %f139, %p27; selp.f32 %f141, %f139, %f107, %p27; add.ftz.f32 %f142, %f141, %f141; and.b32 %r175, %r173, 1; setp.eq.b32 %p28, %r175, 1; neg.ftz.f32 %f143, %f140; selp.f32 %f144, %f143, %f140, %p28; and.b32 %r176, %r173, 2; setp.eq.s32 %p29, %r176, 0; neg.ftz.f32 %f145, %f142; selp.f32 %f146, %f142, %f145, %p29; add.ftz.f32 %f147, %f144, %f146; xor.b32 %r177, %r155, %r131; sub.s32 %r178, %r177, %r157; xor.b32 %r179, %r178, %r152; shf.l.wrap.b32 %r180, %r178, %r178, 25; sub.s32 %r181, %r179, %r180; xor.b32 %r182, %r181, %r155; shf.l.wrap.b32 %r183, %r181, %r181, 16; sub.s32 %r184, %r182, %r183; xor.b32 %r185, %r184, %r178; shf.l.wrap.b32 %r186, %r184, %r184, 4; sub.s32 %r187, %r185, %r186; xor.b32 %r188, %r187, %r181; shf.l.wrap.b32 %r189, %r187, %r187, 14; sub.s32 %r190, %r188, %r189; xor.b32 %r191, %r190, %r184; shf.l.wrap.b32 %r192, %r190, %r190, 24; sub.s32 %r193, %r191, %r192; and.b32 %r194, %r193, 4; setp.eq.s32 %p30, %r194, 0; selp.f32 %f148, %f129, %f139, %p30; selp.f32 %f149, %f139, %f129, %p30; add.ftz.f32 %f150, %f149, %f149; and.b32 %r195, %r193, 1; setp.eq.b32 %p31, %r195, 1; neg.ftz.f32 %f151, %f148; selp.f32 %f152, %f151, %f148, %p31; and.b32 %r196, %r193, 2; setp.eq.s32 %p32, %r196, 0; neg.ftz.f32 %f153, %f150; selp.f32 %f154, %f150, %f153, %p32; add.ftz.f32 %f155, %f152, %f154; sub.ftz.f32 %f156, %f130, %f115; sub.ftz.f32 %f157, %f130, %f120; mul.ftz.f32 %f158, %f156, %f128; fma.rn.ftz.f32 %f159, %f115, %f138, %f158; mul.ftz.f32 %f160, %f156, %f147; fma.rn.ftz.f32 %f161, %f115, %f155, %f160; mul.ftz.f32 %f162, %f120, %f161; fma.rn.ftz.f32 %f22, %f157, %f159, %f162; mov.b32 %r5, %f22; setp.nan.ftz.f32 %p33, %f22, %f22; mov.f32 %f182, 0f3F000000; @%p33 bra $L__BB76_10; or.b32 %r197, %r5, -2147483648; setp.ne.s32 %p34, %r197, -2147483648; add.ftz.f32 %f164, %f22, %f22; setp.eq.ftz.f32 %p35, %f164, %f22; and.pred %p36, %p35, %p34; @%p36 bra $L__BB76_10; shl.b32 %r198, %r5, 1; setp.lt.u32 %p37, %r198, -16777215; mul.ftz.f32 %f165, %f22, 0f3F295E9E; fma.rn.ftz.f32 %f166, %f165, 0f3F000000, 0f3F000000; selp.f32 %f182, %f166, 0f3F000000, %p37; $L__BB76_10: fma.rn.ftz.f32 %f167, %f180, %f182, %f178; add.ftz.f32 %f168, %f179, %f180; div.approx.ftz.f32 %f169, %f167, %f168; sub.ftz.f32 %f171, %f130, %f20; mul.ftz.f32 %f172, %f20, %f169; fma.rn.ftz.f32 %f183, %f183, %f171, %f172; $L__BB76_11: st.param.f32 [func_retval0+0], %f183; ret; } .func (.param .b32 func_retval0) _Z16fractal_noise_3d6float3ff( .param .align 4 .b8 _Z16fractal_noise_3d6float3ff_param_0[12], .param .b32 _Z16fractal_noise_3d6float3ff_param_1, .param .b32 _Z16fractal_noise_3d6float3ff_param_2 ) { .reg .pred %p<94>; .reg .b16 %rs<49>; .reg .f32 %f<301>; .reg .b32 %r<392>; ld.param.f32 %f87, [_Z16fractal_noise_3d6float3ff_param_1]; ld.param.f32 %f82, [_Z16fractal_noise_3d6float3ff_param_2]; ld.param.f32 %f1, [_Z16fractal_noise_3d6float3ff_param_0]; ld.param.f32 %f2, [_Z16fractal_noise_3d6float3ff_param_0+4]; ld.param.f32 %f3, [_Z16fractal_noise_3d6float3ff_param_0+8]; mov.f32 %f287, 0f00000000; max.ftz.f32 %f88, %f87, %f287; mov.f32 %f89, 0f41700000; min.ftz.f32 %f4, %f88, %f89; cvt.rzi.ftz.s32.f32 %r1, %f4; setp.lt.s32 %p1, %r1, 0; mov.f32 %f289, 0f3F800000; mov.f32 %f288, %f287; mov.f32 %f290, %f289; @%p1 bra $L__BB77_22; mov.f32 %f288, 0f00000000; max.ftz.f32 %f94, %f82, %f288; mov.f32 %f290, 0f3F800000; min.ftz.f32 %f5, %f94, %f290; mov.u32 %r391, -1; mov.f32 %f289, %f290; mov.f32 %f287, %f288; $L__BB77_2: mul.ftz.f32 %f95, %f1, %f290; cvt.rmi.ftz.f32.f32 %f96, %f95; cvt.rzi.ftz.s32.f32 %r3, %f96; sub.ftz.f32 %f10, %f95, %f96; mul.ftz.f32 %f97, %f2, %f290; cvt.rmi.ftz.f32.f32 %f98, %f97; cvt.rzi.ftz.s32.f32 %r4, %f98; sub.ftz.f32 %f11, %f97, %f98; mul.ftz.f32 %f99, %f3, %f290; cvt.rmi.ftz.f32.f32 %f100, %f99; cvt.rzi.ftz.s32.f32 %r5, %f100; sub.ftz.f32 %f12, %f99, %f100; fma.rn.ftz.f32 %f101, %f10, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f13, %f10, %f101, 0f41200000; fma.rn.ftz.f32 %f102, %f11, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f14, %f11, %f102, 0f41200000; fma.rn.ftz.f32 %f103, %f12, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f15, %f12, %f103, 0f41200000; add.s32 %r71, %r4, -559038712; add.s32 %r72, %r5, -559038712; xor.b32 %r73, %r72, %r71; shf.l.wrap.b32 %r6, %r71, %r71, 14; sub.s32 %r7, %r73, %r6; add.s32 %r74, %r3, -559038712; xor.b32 %r75, %r7, %r74; shf.l.wrap.b32 %r8, %r7, %r7, 11; sub.s32 %r76, %r75, %r8; xor.b32 %r77, %r76, %r71; shf.l.wrap.b32 %r78, %r76, %r76, 25; sub.s32 %r79, %r77, %r78; xor.b32 %r80, %r79, %r7; shf.l.wrap.b32 %r81, %r79, %r79, 16; sub.s32 %r82, %r80, %r81; xor.b32 %r83, %r82, %r76; shf.l.wrap.b32 %r84, %r82, %r82, 4; sub.s32 %r85, %r83, %r84; xor.b32 %r86, %r85, %r79; shf.l.wrap.b32 %r87, %r85, %r85, 14; sub.s32 %r88, %r86, %r87; xor.b32 %r89, %r88, %r82; shf.l.wrap.b32 %r90, %r88, %r88, 24; sub.s32 %r9, %r89, %r90; and.b32 %r10, %r9, 15; cvt.u16.u32 %rs1, %r9; or.b16 %rs2, %rs1, 2; and.b16 %rs3, %rs2, 15; setp.eq.s16 %p2, %rs3, 14; mov.f32 %f278, %f10; @%p2 bra $L__BB77_4; mov.f32 %f278, %f12; $L__BB77_4: mul.ftz.f32 %f104, %f10, %f10; mul.ftz.f32 %f105, %f10, %f104; mul.ftz.f32 %f17, %f105, %f13; mul.ftz.f32 %f106, %f11, %f11; mul.ftz.f32 %f107, %f11, %f106; mul.ftz.f32 %f18, %f107, %f14; mul.ftz.f32 %f108, %f12, %f12; mul.ftz.f32 %f109, %f12, %f108; mul.ftz.f32 %f19, %f109, %f15; setp.lt.u32 %p3, %r10, 8; selp.f32 %f110, %f10, %f11, %p3; setp.lt.u32 %p4, %r10, 4; selp.f32 %f111, %f11, %f278, %p4; and.b32 %r91, %r9, 1; setp.eq.b32 %p5, %r91, 1; neg.ftz.f32 %f112, %f110; selp.f32 %f113, %f112, %f110, %p5; and.b32 %r92, %r9, 2; setp.eq.s32 %p6, %r92, 0; neg.ftz.f32 %f114, %f111; selp.f32 %f115, %f111, %f114, %p6; add.ftz.f32 %f20, %f113, %f115; add.s32 %r11, %r3, -559038711; xor.b32 %r93, %r7, %r11; sub.s32 %r94, %r93, %r8; xor.b32 %r96, %r94, %r71; shf.l.wrap.b32 %r97, %r94, %r94, 25; sub.s32 %r98, %r96, %r97; xor.b32 %r99, %r98, %r7; shf.l.wrap.b32 %r100, %r98, %r98, 16; sub.s32 %r101, %r99, %r100; xor.b32 %r102, %r101, %r94; shf.l.wrap.b32 %r103, %r101, %r101, 4; sub.s32 %r104, %r102, %r103; xor.b32 %r105, %r104, %r98; shf.l.wrap.b32 %r106, %r104, %r104, 14; sub.s32 %r107, %r105, %r106; xor.b32 %r108, %r107, %r101; shf.l.wrap.b32 %r109, %r107, %r107, 24; sub.s32 %r12, %r108, %r109; and.b32 %r13, %r12, 15; add.ftz.f32 %f21, %f10, 0fBF800000; cvt.u16.u32 %rs4, %r12; or.b16 %rs5, %rs4, 2; and.b16 %rs6, %rs5, 15; setp.eq.s16 %p7, %rs6, 14; mov.f32 %f279, %f21; @%p7 bra $L__BB77_6; mov.f32 %f279, %f12; $L__BB77_6: setp.lt.u32 %p8, %r13, 8; selp.f32 %f116, %f21, %f11, %p8; setp.lt.u32 %p9, %r13, 4; selp.f32 %f117, %f11, %f279, %p9; and.b32 %r110, %r12, 1; setp.eq.b32 %p10, %r110, 1; neg.ftz.f32 %f118, %f116; selp.f32 %f119, %f118, %f116, %p10; and.b32 %r111, %r12, 2; setp.eq.s32 %p11, %r111, 0; neg.ftz.f32 %f120, %f117; selp.f32 %f121, %f117, %f120, %p11; add.ftz.f32 %f23, %f119, %f121; add.s32 %r14, %r4, -559038711; xor.b32 %r113, %r72, %r14; shf.l.wrap.b32 %r15, %r14, %r14, 14; sub.s32 %r16, %r113, %r15; xor.b32 %r115, %r16, %r74; shf.l.wrap.b32 %r17, %r16, %r16, 11; sub.s32 %r116, %r115, %r17; xor.b32 %r117, %r116, %r14; shf.l.wrap.b32 %r118, %r116, %r116, 25; sub.s32 %r119, %r117, %r118; xor.b32 %r120, %r119, %r16; shf.l.wrap.b32 %r121, %r119, %r119, 16; sub.s32 %r122, %r120, %r121; xor.b32 %r123, %r122, %r116; shf.l.wrap.b32 %r124, %r122, %r122, 4; sub.s32 %r125, %r123, %r124; xor.b32 %r126, %r125, %r119; shf.l.wrap.b32 %r127, %r125, %r125, 14; sub.s32 %r128, %r126, %r127; xor.b32 %r129, %r128, %r122; shf.l.wrap.b32 %r130, %r128, %r128, 24; sub.s32 %r18, %r129, %r130; and.b32 %r19, %r18, 15; add.ftz.f32 %f24, %f11, 0fBF800000; cvt.u16.u32 %rs7, %r18; or.b16 %rs8, %rs7, 2; and.b16 %rs9, %rs8, 15; setp.eq.s16 %p12, %rs9, 14; mov.f32 %f280, %f10; @%p12 bra $L__BB77_8; mov.f32 %f280, %f12; $L__BB77_8: setp.lt.u32 %p13, %r19, 8; selp.f32 %f122, %f10, %f24, %p13; setp.lt.u32 %p14, %r19, 4; selp.f32 %f123, %f24, %f280, %p14; and.b32 %r131, %r18, 1; setp.eq.b32 %p15, %r131, 1; neg.ftz.f32 %f124, %f122; selp.f32 %f125, %f124, %f122, %p15; and.b32 %r132, %r18, 2; setp.eq.s32 %p16, %r132, 0; neg.ftz.f32 %f126, %f123; selp.f32 %f127, %f123, %f126, %p16; add.ftz.f32 %f26, %f125, %f127; xor.b32 %r133, %r16, %r11; sub.s32 %r134, %r133, %r17; xor.b32 %r135, %r134, %r14; shf.l.wrap.b32 %r136, %r134, %r134, 25; sub.s32 %r137, %r135, %r136; xor.b32 %r138, %r137, %r16; shf.l.wrap.b32 %r139, %r137, %r137, 16; sub.s32 %r140, %r138, %r139; xor.b32 %r141, %r140, %r134; shf.l.wrap.b32 %r142, %r140, %r140, 4; sub.s32 %r143, %r141, %r142; xor.b32 %r144, %r143, %r137; shf.l.wrap.b32 %r145, %r143, %r143, 14; sub.s32 %r146, %r144, %r145; xor.b32 %r147, %r146, %r140; shf.l.wrap.b32 %r148, %r146, %r146, 24; sub.s32 %r20, %r147, %r148; and.b32 %r21, %r20, 15; cvt.u16.u32 %rs10, %r20; or.b16 %rs11, %rs10, 2; and.b16 %rs12, %rs11, 15; setp.eq.s16 %p17, %rs12, 14; mov.f32 %f281, %f21; @%p17 bra $L__BB77_10; mov.f32 %f281, %f12; $L__BB77_10: setp.lt.u32 %p18, %r21, 8; selp.f32 %f128, %f21, %f24, %p18; setp.lt.u32 %p19, %r21, 4; selp.f32 %f129, %f24, %f281, %p19; and.b32 %r149, %r20, 1; setp.eq.b32 %p20, %r149, 1; neg.ftz.f32 %f130, %f128; selp.f32 %f131, %f130, %f128, %p20; and.b32 %r150, %r20, 2; setp.eq.s32 %p21, %r150, 0; neg.ftz.f32 %f132, %f129; selp.f32 %f133, %f129, %f132, %p21; add.ftz.f32 %f28, %f131, %f133; add.s32 %r22, %r5, -559038711; xor.b32 %r152, %r22, %r71; sub.s32 %r23, %r152, %r6; xor.b32 %r154, %r23, %r74; shf.l.wrap.b32 %r24, %r23, %r23, 11; sub.s32 %r155, %r154, %r24; xor.b32 %r156, %r155, %r71; shf.l.wrap.b32 %r157, %r155, %r155, 25; sub.s32 %r158, %r156, %r157; xor.b32 %r159, %r158, %r23; shf.l.wrap.b32 %r160, %r158, %r158, 16; sub.s32 %r161, %r159, %r160; xor.b32 %r162, %r161, %r155; shf.l.wrap.b32 %r163, %r161, %r161, 4; sub.s32 %r164, %r162, %r163; xor.b32 %r165, %r164, %r158; shf.l.wrap.b32 %r166, %r164, %r164, 14; sub.s32 %r167, %r165, %r166; xor.b32 %r168, %r167, %r161; shf.l.wrap.b32 %r169, %r167, %r167, 24; sub.s32 %r25, %r168, %r169; add.ftz.f32 %f29, %f12, 0fBF800000; and.b32 %r26, %r25, 15; cvt.u16.u32 %rs13, %r25; or.b16 %rs14, %rs13, 2; and.b16 %rs15, %rs14, 15; setp.eq.s16 %p22, %rs15, 14; mov.f32 %f282, %f10; @%p22 bra $L__BB77_12; mov.f32 %f282, %f29; $L__BB77_12: setp.lt.u32 %p23, %r26, 8; selp.f32 %f134, %f10, %f11, %p23; setp.lt.u32 %p24, %r26, 4; selp.f32 %f135, %f11, %f282, %p24; and.b32 %r170, %r25, 1; setp.eq.b32 %p25, %r170, 1; neg.ftz.f32 %f136, %f134; selp.f32 %f137, %f136, %f134, %p25; and.b32 %r171, %r25, 2; setp.eq.s32 %p26, %r171, 0; neg.ftz.f32 %f138, %f135; selp.f32 %f139, %f135, %f138, %p26; add.ftz.f32 %f31, %f137, %f139; xor.b32 %r172, %r23, %r11; sub.s32 %r173, %r172, %r24; xor.b32 %r175, %r173, %r71; shf.l.wrap.b32 %r176, %r173, %r173, 25; sub.s32 %r177, %r175, %r176; xor.b32 %r178, %r177, %r23; shf.l.wrap.b32 %r179, %r177, %r177, 16; sub.s32 %r180, %r178, %r179; xor.b32 %r181, %r180, %r173; shf.l.wrap.b32 %r182, %r180, %r180, 4; sub.s32 %r183, %r181, %r182; xor.b32 %r184, %r183, %r177; shf.l.wrap.b32 %r185, %r183, %r183, 14; sub.s32 %r186, %r184, %r185; xor.b32 %r187, %r186, %r180; shf.l.wrap.b32 %r188, %r186, %r186, 24; sub.s32 %r27, %r187, %r188; and.b32 %r28, %r27, 15; cvt.u16.u32 %rs16, %r27; or.b16 %rs17, %rs16, 2; and.b16 %rs18, %rs17, 15; setp.eq.s16 %p27, %rs18, 14; mov.f32 %f283, %f21; @%p27 bra $L__BB77_14; mov.f32 %f283, %f29; $L__BB77_14: setp.lt.u32 %p28, %r28, 8; selp.f32 %f140, %f21, %f11, %p28; setp.lt.u32 %p29, %r28, 4; selp.f32 %f141, %f11, %f283, %p29; and.b32 %r189, %r27, 1; setp.eq.b32 %p30, %r189, 1; neg.ftz.f32 %f142, %f140; selp.f32 %f143, %f142, %f140, %p30; and.b32 %r190, %r27, 2; setp.eq.s32 %p31, %r190, 0; neg.ftz.f32 %f144, %f141; selp.f32 %f145, %f141, %f144, %p31; add.ftz.f32 %f33, %f143, %f145; xor.b32 %r191, %r22, %r14; sub.s32 %r29, %r191, %r15; xor.b32 %r193, %r29, %r74; shf.l.wrap.b32 %r30, %r29, %r29, 11; sub.s32 %r194, %r193, %r30; xor.b32 %r195, %r194, %r14; shf.l.wrap.b32 %r196, %r194, %r194, 25; sub.s32 %r197, %r195, %r196; xor.b32 %r198, %r197, %r29; shf.l.wrap.b32 %r199, %r197, %r197, 16; sub.s32 %r200, %r198, %r199; xor.b32 %r201, %r200, %r194; shf.l.wrap.b32 %r202, %r200, %r200, 4; sub.s32 %r203, %r201, %r202; xor.b32 %r204, %r203, %r197; shf.l.wrap.b32 %r205, %r203, %r203, 14; sub.s32 %r206, %r204, %r205; xor.b32 %r207, %r206, %r200; shf.l.wrap.b32 %r208, %r206, %r206, 24; sub.s32 %r31, %r207, %r208; and.b32 %r32, %r31, 15; cvt.u16.u32 %rs19, %r31; or.b16 %rs20, %rs19, 2; and.b16 %rs21, %rs20, 15; setp.eq.s16 %p32, %rs21, 14; mov.f32 %f284, %f10; @%p32 bra $L__BB77_16; mov.f32 %f284, %f29; $L__BB77_16: setp.lt.u32 %p33, %r32, 8; selp.f32 %f146, %f10, %f24, %p33; setp.lt.u32 %p34, %r32, 4; selp.f32 %f147, %f24, %f284, %p34; and.b32 %r209, %r31, 1; setp.eq.b32 %p35, %r209, 1; neg.ftz.f32 %f148, %f146; selp.f32 %f149, %f148, %f146, %p35; and.b32 %r210, %r31, 2; setp.eq.s32 %p36, %r210, 0; neg.ftz.f32 %f150, %f147; selp.f32 %f151, %f147, %f150, %p36; add.ftz.f32 %f35, %f149, %f151; xor.b32 %r211, %r29, %r11; sub.s32 %r212, %r211, %r30; xor.b32 %r213, %r212, %r14; shf.l.wrap.b32 %r214, %r212, %r212, 25; sub.s32 %r215, %r213, %r214; xor.b32 %r216, %r215, %r29; shf.l.wrap.b32 %r217, %r215, %r215, 16; sub.s32 %r218, %r216, %r217; xor.b32 %r219, %r218, %r212; shf.l.wrap.b32 %r220, %r218, %r218, 4; sub.s32 %r221, %r219, %r220; xor.b32 %r222, %r221, %r215; shf.l.wrap.b32 %r223, %r221, %r221, 14; sub.s32 %r224, %r222, %r223; xor.b32 %r225, %r224, %r218; shf.l.wrap.b32 %r226, %r224, %r224, 24; sub.s32 %r33, %r225, %r226; and.b32 %r34, %r33, 15; cvt.u16.u32 %rs22, %r33; or.b16 %rs23, %rs22, 2; and.b16 %rs24, %rs23, 15; setp.eq.s16 %p37, %rs24, 14; mov.f32 %f285, %f21; @%p37 bra $L__BB77_18; mov.f32 %f285, %f29; $L__BB77_18: setp.lt.u32 %p38, %r34, 8; selp.f32 %f153, %f21, %f24, %p38; setp.lt.u32 %p39, %r34, 4; selp.f32 %f154, %f24, %f285, %p39; and.b32 %r227, %r33, 1; setp.eq.b32 %p40, %r227, 1; neg.ftz.f32 %f155, %f153; selp.f32 %f156, %f155, %f153, %p40; and.b32 %r228, %r33, 2; setp.eq.s32 %p41, %r228, 0; neg.ftz.f32 %f157, %f154; selp.f32 %f158, %f154, %f157, %p41; add.ftz.f32 %f159, %f156, %f158; mov.f32 %f160, 0f3F800000; sub.ftz.f32 %f161, %f160, %f17; mul.ftz.f32 %f162, %f17, %f23; fma.rn.ftz.f32 %f163, %f161, %f20, %f162; sub.ftz.f32 %f164, %f160, %f18; mul.ftz.f32 %f165, %f17, %f28; fma.rn.ftz.f32 %f166, %f161, %f26, %f165; mul.ftz.f32 %f167, %f18, %f166; fma.rn.ftz.f32 %f168, %f164, %f163, %f167; sub.ftz.f32 %f169, %f160, %f19; mul.ftz.f32 %f170, %f17, %f33; fma.rn.ftz.f32 %f171, %f161, %f31, %f170; mul.ftz.f32 %f172, %f17, %f159; fma.rn.ftz.f32 %f173, %f161, %f35, %f172; mul.ftz.f32 %f174, %f18, %f173; fma.rn.ftz.f32 %f175, %f164, %f171, %f174; mul.ftz.f32 %f176, %f19, %f175; fma.rn.ftz.f32 %f37, %f169, %f168, %f176; mov.b32 %r35, %f37; setp.nan.ftz.f32 %p42, %f37, %f37; mov.f32 %f286, 0f00000000; @%p42 bra $L__BB77_21; or.b32 %r229, %r35, -2147483648; setp.ne.s32 %p43, %r229, -2147483648; add.ftz.f32 %f178, %f37, %f37; setp.eq.ftz.f32 %p44, %f178, %f37; and.pred %p45, %p44, %p43; @%p45 bra $L__BB77_21; shl.b32 %r230, %r35, 1; setp.lt.u32 %p46, %r230, -16777215; selp.f32 %f286, %f37, 0f00000000, %p46; $L__BB77_21: mul.ftz.f32 %f179, %f286, 0f3F7B645A; fma.rn.ftz.f32 %f180, %f179, 0f3F000000, 0f3F000000; fma.rn.ftz.f32 %f287, %f289, %f180, %f287; add.ftz.f32 %f288, %f288, %f289; mul.ftz.f32 %f289, %f289, %f5; add.ftz.f32 %f290, %f290, %f290; add.s32 %r391, %r391, 1; setp.lt.s32 %p47, %r391, %r1; @%p47 bra $L__BB77_2; $L__BB77_22: cvt.rmi.ftz.f32.f32 %f181, %f4; sub.ftz.f32 %f48, %f4, %f181; setp.eq.ftz.f32 %p48, %f48, 0f00000000; div.approx.ftz.f32 %f300, %f287, %f288; @%p48 bra $L__BB77_43; mul.ftz.f32 %f182, %f1, %f290; cvt.rmi.ftz.f32.f32 %f183, %f182; cvt.rzi.ftz.s32.f32 %r37, %f183; sub.ftz.f32 %f50, %f182, %f183; mul.ftz.f32 %f184, %f2, %f290; cvt.rmi.ftz.f32.f32 %f185, %f184; cvt.rzi.ftz.s32.f32 %r38, %f185; sub.ftz.f32 %f51, %f184, %f185; mul.ftz.f32 %f186, %f3, %f290; cvt.rmi.ftz.f32.f32 %f187, %f186; cvt.rzi.ftz.s32.f32 %r39, %f187; sub.ftz.f32 %f52, %f186, %f187; fma.rn.ftz.f32 %f188, %f50, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f53, %f50, %f188, 0f41200000; fma.rn.ftz.f32 %f189, %f51, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f54, %f51, %f189, 0f41200000; fma.rn.ftz.f32 %f190, %f52, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f55, %f52, %f190, 0f41200000; add.s32 %r231, %r38, -559038712; add.s32 %r232, %r39, -559038712; xor.b32 %r233, %r232, %r231; shf.l.wrap.b32 %r40, %r231, %r231, 14; sub.s32 %r41, %r233, %r40; add.s32 %r234, %r37, -559038712; xor.b32 %r235, %r41, %r234; shf.l.wrap.b32 %r42, %r41, %r41, 11; sub.s32 %r236, %r235, %r42; xor.b32 %r237, %r236, %r231; shf.l.wrap.b32 %r238, %r236, %r236, 25; sub.s32 %r239, %r237, %r238; xor.b32 %r240, %r239, %r41; shf.l.wrap.b32 %r241, %r239, %r239, 16; sub.s32 %r242, %r240, %r241; xor.b32 %r243, %r242, %r236; shf.l.wrap.b32 %r244, %r242, %r242, 4; sub.s32 %r245, %r243, %r244; xor.b32 %r246, %r245, %r239; shf.l.wrap.b32 %r247, %r245, %r245, 14; sub.s32 %r248, %r246, %r247; xor.b32 %r249, %r248, %r242; shf.l.wrap.b32 %r250, %r248, %r248, 24; sub.s32 %r43, %r249, %r250; and.b32 %r44, %r43, 15; cvt.u16.u32 %rs25, %r43; or.b16 %rs26, %rs25, 2; and.b16 %rs27, %rs26, 15; setp.eq.s16 %p49, %rs27, 14; mov.f32 %f291, %f50; @%p49 bra $L__BB77_25; mov.f32 %f291, %f52; $L__BB77_25: mul.ftz.f32 %f191, %f50, %f50; mul.ftz.f32 %f192, %f50, %f191; mul.ftz.f32 %f57, %f192, %f53; mul.ftz.f32 %f193, %f51, %f51; mul.ftz.f32 %f194, %f51, %f193; mul.ftz.f32 %f58, %f194, %f54; mul.ftz.f32 %f195, %f52, %f52; mul.ftz.f32 %f196, %f52, %f195; mul.ftz.f32 %f59, %f196, %f55; setp.lt.u32 %p50, %r44, 8; selp.f32 %f197, %f50, %f51, %p50; setp.lt.u32 %p51, %r44, 4; selp.f32 %f198, %f51, %f291, %p51; and.b32 %r251, %r43, 1; setp.eq.b32 %p52, %r251, 1; neg.ftz.f32 %f199, %f197; selp.f32 %f200, %f199, %f197, %p52; and.b32 %r252, %r43, 2; setp.eq.s32 %p53, %r252, 0; neg.ftz.f32 %f201, %f198; selp.f32 %f202, %f198, %f201, %p53; add.ftz.f32 %f60, %f200, %f202; add.s32 %r45, %r37, -559038711; xor.b32 %r253, %r41, %r45; sub.s32 %r254, %r253, %r42; xor.b32 %r256, %r254, %r231; shf.l.wrap.b32 %r257, %r254, %r254, 25; sub.s32 %r258, %r256, %r257; xor.b32 %r259, %r258, %r41; shf.l.wrap.b32 %r260, %r258, %r258, 16; sub.s32 %r261, %r259, %r260; xor.b32 %r262, %r261, %r254; shf.l.wrap.b32 %r263, %r261, %r261, 4; sub.s32 %r264, %r262, %r263; xor.b32 %r265, %r264, %r258; shf.l.wrap.b32 %r266, %r264, %r264, 14; sub.s32 %r267, %r265, %r266; xor.b32 %r268, %r267, %r261; shf.l.wrap.b32 %r269, %r267, %r267, 24; sub.s32 %r46, %r268, %r269; and.b32 %r47, %r46, 15; add.ftz.f32 %f61, %f50, 0fBF800000; cvt.u16.u32 %rs28, %r46; or.b16 %rs29, %rs28, 2; and.b16 %rs30, %rs29, 15; setp.eq.s16 %p54, %rs30, 14; mov.f32 %f292, %f61; @%p54 bra $L__BB77_27; mov.f32 %f292, %f52; $L__BB77_27: setp.lt.u32 %p55, %r47, 8; selp.f32 %f203, %f61, %f51, %p55; setp.lt.u32 %p56, %r47, 4; selp.f32 %f204, %f51, %f292, %p56; and.b32 %r270, %r46, 1; setp.eq.b32 %p57, %r270, 1; neg.ftz.f32 %f205, %f203; selp.f32 %f206, %f205, %f203, %p57; and.b32 %r271, %r46, 2; setp.eq.s32 %p58, %r271, 0; neg.ftz.f32 %f207, %f204; selp.f32 %f208, %f204, %f207, %p58; add.ftz.f32 %f63, %f206, %f208; add.s32 %r48, %r38, -559038711; xor.b32 %r273, %r232, %r48; shf.l.wrap.b32 %r49, %r48, %r48, 14; sub.s32 %r50, %r273, %r49; xor.b32 %r275, %r50, %r234; shf.l.wrap.b32 %r51, %r50, %r50, 11; sub.s32 %r276, %r275, %r51; xor.b32 %r277, %r276, %r48; shf.l.wrap.b32 %r278, %r276, %r276, 25; sub.s32 %r279, %r277, %r278; xor.b32 %r280, %r279, %r50; shf.l.wrap.b32 %r281, %r279, %r279, 16; sub.s32 %r282, %r280, %r281; xor.b32 %r283, %r282, %r276; shf.l.wrap.b32 %r284, %r282, %r282, 4; sub.s32 %r285, %r283, %r284; xor.b32 %r286, %r285, %r279; shf.l.wrap.b32 %r287, %r285, %r285, 14; sub.s32 %r288, %r286, %r287; xor.b32 %r289, %r288, %r282; shf.l.wrap.b32 %r290, %r288, %r288, 24; sub.s32 %r52, %r289, %r290; and.b32 %r53, %r52, 15; add.ftz.f32 %f64, %f51, 0fBF800000; cvt.u16.u32 %rs31, %r52; or.b16 %rs32, %rs31, 2; and.b16 %rs33, %rs32, 15; setp.eq.s16 %p59, %rs33, 14; mov.f32 %f293, %f50; @%p59 bra $L__BB77_29; mov.f32 %f293, %f52; $L__BB77_29: setp.lt.u32 %p60, %r53, 8; selp.f32 %f209, %f50, %f64, %p60; setp.lt.u32 %p61, %r53, 4; selp.f32 %f210, %f64, %f293, %p61; and.b32 %r291, %r52, 1; setp.eq.b32 %p62, %r291, 1; neg.ftz.f32 %f211, %f209; selp.f32 %f212, %f211, %f209, %p62; and.b32 %r292, %r52, 2; setp.eq.s32 %p63, %r292, 0; neg.ftz.f32 %f213, %f210; selp.f32 %f214, %f210, %f213, %p63; add.ftz.f32 %f66, %f212, %f214; xor.b32 %r293, %r50, %r45; sub.s32 %r294, %r293, %r51; xor.b32 %r295, %r294, %r48; shf.l.wrap.b32 %r296, %r294, %r294, 25; sub.s32 %r297, %r295, %r296; xor.b32 %r298, %r297, %r50; shf.l.wrap.b32 %r299, %r297, %r297, 16; sub.s32 %r300, %r298, %r299; xor.b32 %r301, %r300, %r294; shf.l.wrap.b32 %r302, %r300, %r300, 4; sub.s32 %r303, %r301, %r302; xor.b32 %r304, %r303, %r297; shf.l.wrap.b32 %r305, %r303, %r303, 14; sub.s32 %r306, %r304, %r305; xor.b32 %r307, %r306, %r300; shf.l.wrap.b32 %r308, %r306, %r306, 24; sub.s32 %r54, %r307, %r308; and.b32 %r55, %r54, 15; cvt.u16.u32 %rs34, %r54; or.b16 %rs35, %rs34, 2; and.b16 %rs36, %rs35, 15; setp.eq.s16 %p64, %rs36, 14; mov.f32 %f294, %f61; @%p64 bra $L__BB77_31; mov.f32 %f294, %f52; $L__BB77_31: setp.lt.u32 %p65, %r55, 8; selp.f32 %f215, %f61, %f64, %p65; setp.lt.u32 %p66, %r55, 4; selp.f32 %f216, %f64, %f294, %p66; and.b32 %r309, %r54, 1; setp.eq.b32 %p67, %r309, 1; neg.ftz.f32 %f217, %f215; selp.f32 %f218, %f217, %f215, %p67; and.b32 %r310, %r54, 2; setp.eq.s32 %p68, %r310, 0; neg.ftz.f32 %f219, %f216; selp.f32 %f220, %f216, %f219, %p68; add.ftz.f32 %f68, %f218, %f220; add.s32 %r56, %r39, -559038711; xor.b32 %r312, %r56, %r231; sub.s32 %r57, %r312, %r40; xor.b32 %r314, %r57, %r234; shf.l.wrap.b32 %r58, %r57, %r57, 11; sub.s32 %r315, %r314, %r58; xor.b32 %r316, %r315, %r231; shf.l.wrap.b32 %r317, %r315, %r315, 25; sub.s32 %r318, %r316, %r317; xor.b32 %r319, %r318, %r57; shf.l.wrap.b32 %r320, %r318, %r318, 16; sub.s32 %r321, %r319, %r320; xor.b32 %r322, %r321, %r315; shf.l.wrap.b32 %r323, %r321, %r321, 4; sub.s32 %r324, %r322, %r323; xor.b32 %r325, %r324, %r318; shf.l.wrap.b32 %r326, %r324, %r324, 14; sub.s32 %r327, %r325, %r326; xor.b32 %r328, %r327, %r321; shf.l.wrap.b32 %r329, %r327, %r327, 24; sub.s32 %r59, %r328, %r329; add.ftz.f32 %f69, %f52, 0fBF800000; and.b32 %r60, %r59, 15; cvt.u16.u32 %rs37, %r59; or.b16 %rs38, %rs37, 2; and.b16 %rs39, %rs38, 15; setp.eq.s16 %p69, %rs39, 14; mov.f32 %f295, %f50; @%p69 bra $L__BB77_33; mov.f32 %f295, %f69; $L__BB77_33: setp.lt.u32 %p70, %r60, 8; selp.f32 %f221, %f50, %f51, %p70; setp.lt.u32 %p71, %r60, 4; selp.f32 %f222, %f51, %f295, %p71; and.b32 %r330, %r59, 1; setp.eq.b32 %p72, %r330, 1; neg.ftz.f32 %f223, %f221; selp.f32 %f224, %f223, %f221, %p72; and.b32 %r331, %r59, 2; setp.eq.s32 %p73, %r331, 0; neg.ftz.f32 %f225, %f222; selp.f32 %f226, %f222, %f225, %p73; add.ftz.f32 %f71, %f224, %f226; xor.b32 %r332, %r57, %r45; sub.s32 %r333, %r332, %r58; xor.b32 %r335, %r333, %r231; shf.l.wrap.b32 %r336, %r333, %r333, 25; sub.s32 %r337, %r335, %r336; xor.b32 %r338, %r337, %r57; shf.l.wrap.b32 %r339, %r337, %r337, 16; sub.s32 %r340, %r338, %r339; xor.b32 %r341, %r340, %r333; shf.l.wrap.b32 %r342, %r340, %r340, 4; sub.s32 %r343, %r341, %r342; xor.b32 %r344, %r343, %r337; shf.l.wrap.b32 %r345, %r343, %r343, 14; sub.s32 %r346, %r344, %r345; xor.b32 %r347, %r346, %r340; shf.l.wrap.b32 %r348, %r346, %r346, 24; sub.s32 %r61, %r347, %r348; and.b32 %r62, %r61, 15; cvt.u16.u32 %rs40, %r61; or.b16 %rs41, %rs40, 2; and.b16 %rs42, %rs41, 15; setp.eq.s16 %p74, %rs42, 14; mov.f32 %f296, %f61; @%p74 bra $L__BB77_35; mov.f32 %f296, %f69; $L__BB77_35: setp.lt.u32 %p75, %r62, 8; selp.f32 %f227, %f61, %f51, %p75; setp.lt.u32 %p76, %r62, 4; selp.f32 %f228, %f51, %f296, %p76; and.b32 %r349, %r61, 1; setp.eq.b32 %p77, %r349, 1; neg.ftz.f32 %f229, %f227; selp.f32 %f230, %f229, %f227, %p77; and.b32 %r350, %r61, 2; setp.eq.s32 %p78, %r350, 0; neg.ftz.f32 %f231, %f228; selp.f32 %f232, %f228, %f231, %p78; add.ftz.f32 %f73, %f230, %f232; xor.b32 %r351, %r56, %r48; sub.s32 %r63, %r351, %r49; xor.b32 %r353, %r63, %r234; shf.l.wrap.b32 %r64, %r63, %r63, 11; sub.s32 %r354, %r353, %r64; xor.b32 %r355, %r354, %r48; shf.l.wrap.b32 %r356, %r354, %r354, 25; sub.s32 %r357, %r355, %r356; xor.b32 %r358, %r357, %r63; shf.l.wrap.b32 %r359, %r357, %r357, 16; sub.s32 %r360, %r358, %r359; xor.b32 %r361, %r360, %r354; shf.l.wrap.b32 %r362, %r360, %r360, 4; sub.s32 %r363, %r361, %r362; xor.b32 %r364, %r363, %r357; shf.l.wrap.b32 %r365, %r363, %r363, 14; sub.s32 %r366, %r364, %r365; xor.b32 %r367, %r366, %r360; shf.l.wrap.b32 %r368, %r366, %r366, 24; sub.s32 %r65, %r367, %r368; and.b32 %r66, %r65, 15; cvt.u16.u32 %rs43, %r65; or.b16 %rs44, %rs43, 2; and.b16 %rs45, %rs44, 15; setp.eq.s16 %p79, %rs45, 14; mov.f32 %f297, %f50; @%p79 bra $L__BB77_37; mov.f32 %f297, %f69; $L__BB77_37: setp.lt.u32 %p80, %r66, 8; selp.f32 %f233, %f50, %f64, %p80; setp.lt.u32 %p81, %r66, 4; selp.f32 %f234, %f64, %f297, %p81; and.b32 %r369, %r65, 1; setp.eq.b32 %p82, %r369, 1; neg.ftz.f32 %f235, %f233; selp.f32 %f236, %f235, %f233, %p82; and.b32 %r370, %r65, 2; setp.eq.s32 %p83, %r370, 0; neg.ftz.f32 %f237, %f234; selp.f32 %f238, %f234, %f237, %p83; add.ftz.f32 %f75, %f236, %f238; xor.b32 %r371, %r63, %r45; sub.s32 %r372, %r371, %r64; xor.b32 %r373, %r372, %r48; shf.l.wrap.b32 %r374, %r372, %r372, 25; sub.s32 %r375, %r373, %r374; xor.b32 %r376, %r375, %r63; shf.l.wrap.b32 %r377, %r375, %r375, 16; sub.s32 %r378, %r376, %r377; xor.b32 %r379, %r378, %r372; shf.l.wrap.b32 %r380, %r378, %r378, 4; sub.s32 %r381, %r379, %r380; xor.b32 %r382, %r381, %r375; shf.l.wrap.b32 %r383, %r381, %r381, 14; sub.s32 %r384, %r382, %r383; xor.b32 %r385, %r384, %r378; shf.l.wrap.b32 %r386, %r384, %r384, 24; sub.s32 %r67, %r385, %r386; and.b32 %r68, %r67, 15; cvt.u16.u32 %rs46, %r67; or.b16 %rs47, %rs46, 2; and.b16 %rs48, %rs47, 15; setp.eq.s16 %p84, %rs48, 14; mov.f32 %f298, %f61; @%p84 bra $L__BB77_39; mov.f32 %f298, %f69; $L__BB77_39: setp.lt.u32 %p85, %r68, 8; selp.f32 %f240, %f61, %f64, %p85; setp.lt.u32 %p86, %r68, 4; selp.f32 %f241, %f64, %f298, %p86; and.b32 %r387, %r67, 1; setp.eq.b32 %p87, %r387, 1; neg.ftz.f32 %f242, %f240; selp.f32 %f243, %f242, %f240, %p87; and.b32 %r388, %r67, 2; setp.eq.s32 %p88, %r388, 0; neg.ftz.f32 %f244, %f241; selp.f32 %f245, %f241, %f244, %p88; add.ftz.f32 %f246, %f243, %f245; mov.f32 %f247, 0f3F800000; sub.ftz.f32 %f248, %f247, %f57; mul.ftz.f32 %f249, %f57, %f63; fma.rn.ftz.f32 %f250, %f248, %f60, %f249; sub.ftz.f32 %f251, %f247, %f58; mul.ftz.f32 %f252, %f57, %f68; fma.rn.ftz.f32 %f253, %f248, %f66, %f252; mul.ftz.f32 %f254, %f58, %f253; fma.rn.ftz.f32 %f255, %f251, %f250, %f254; sub.ftz.f32 %f256, %f247, %f59; mul.ftz.f32 %f257, %f57, %f73; fma.rn.ftz.f32 %f258, %f248, %f71, %f257; mul.ftz.f32 %f259, %f57, %f246; fma.rn.ftz.f32 %f260, %f248, %f75, %f259; mul.ftz.f32 %f261, %f58, %f260; fma.rn.ftz.f32 %f262, %f251, %f258, %f261; mul.ftz.f32 %f263, %f59, %f262; fma.rn.ftz.f32 %f77, %f256, %f255, %f263; mov.b32 %r69, %f77; setp.nan.ftz.f32 %p89, %f77, %f77; mov.f32 %f299, 0f3F000000; @%p89 bra $L__BB77_42; or.b32 %r389, %r69, -2147483648; setp.ne.s32 %p90, %r389, -2147483648; add.ftz.f32 %f265, %f77, %f77; setp.eq.ftz.f32 %p91, %f265, %f77; and.pred %p92, %p91, %p90; @%p92 bra $L__BB77_42; shl.b32 %r390, %r69, 1; setp.lt.u32 %p93, %r390, -16777215; mul.ftz.f32 %f266, %f77, 0f3F7B645A; fma.rn.ftz.f32 %f267, %f266, 0f3F000000, 0f3F000000; selp.f32 %f299, %f267, 0f3F000000, %p93; $L__BB77_42: fma.rn.ftz.f32 %f268, %f289, %f299, %f287; add.ftz.f32 %f269, %f288, %f289; div.approx.ftz.f32 %f270, %f268, %f269; sub.ftz.f32 %f272, %f247, %f48; mul.ftz.f32 %f273, %f48, %f270; fma.rn.ftz.f32 %f300, %f300, %f272, %f273; $L__BB77_43: st.param.f32 [func_retval0+0], %f300; ret; } .func (.param .b32 func_retval0) _Z16fractal_noise_4d6float4ff( .param .align 16 .b8 _Z16fractal_noise_4d6float4ff_param_0[16], .param .b32 _Z16fractal_noise_4d6float4ff_param_1, .param .b32 _Z16fractal_noise_4d6float4ff_param_2 ) { .reg .pred %p<206>; .reg .f32 %f<565>; .reg .b32 %r<1140>; ld.param.f32 %f34, [_Z16fractal_noise_4d6float4ff_param_1]; ld.param.f32 %f29, [_Z16fractal_noise_4d6float4ff_param_2]; ld.param.f32 %f1, [_Z16fractal_noise_4d6float4ff_param_0]; ld.param.f32 %f2, [_Z16fractal_noise_4d6float4ff_param_0+4]; ld.param.f32 %f3, [_Z16fractal_noise_4d6float4ff_param_0+8]; ld.param.f32 %f4, [_Z16fractal_noise_4d6float4ff_param_0+12]; mov.f32 %f559, 0f00000000; max.ftz.f32 %f35, %f34, %f559; mov.f32 %f36, 0f41700000; min.ftz.f32 %f5, %f35, %f36; cvt.rzi.ftz.s32.f32 %r1, %f5; setp.lt.s32 %p1, %r1, 0; mov.f32 %f561, 0f3F800000; mov.f32 %f560, %f559; mov.f32 %f562, %f561; @%p1 bra $L__BB78_6; mov.f32 %f40, 0f00000000; max.ftz.f32 %f41, %f29, %f40; mov.f32 %f38, 0f3F800000; min.ftz.f32 %f6, %f41, %f38; mov.u32 %r1139, -1; mov.f32 %f562, %f38; mov.f32 %f561, %f38; mov.f32 %f560, %f40; mov.f32 %f559, %f40; $L__BB78_2: mul.ftz.f32 %f43, %f1, %f562; cvt.rmi.ftz.f32.f32 %f44, %f43; cvt.rzi.ftz.s32.f32 %r7, %f44; sub.ftz.f32 %f45, %f43, %f44; mul.ftz.f32 %f46, %f2, %f562; cvt.rmi.ftz.f32.f32 %f47, %f46; cvt.rzi.ftz.s32.f32 %r8, %f47; sub.ftz.f32 %f48, %f46, %f47; mul.ftz.f32 %f49, %f3, %f562; cvt.rmi.ftz.f32.f32 %f50, %f49; cvt.rzi.ftz.s32.f32 %r9, %f50; sub.ftz.f32 %f51, %f49, %f50; mul.ftz.f32 %f52, %f4, %f562; cvt.rmi.ftz.f32.f32 %f53, %f52; cvt.rzi.ftz.s32.f32 %r10, %f53; sub.ftz.f32 %f54, %f52, %f53; mul.ftz.f32 %f55, %f45, %f45; mul.ftz.f32 %f56, %f45, %f55; fma.rn.ftz.f32 %f57, %f45, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f58, %f45, %f57, 0f41200000; mul.ftz.f32 %f59, %f56, %f58; mul.ftz.f32 %f60, %f48, %f48; mul.ftz.f32 %f61, %f48, %f60; fma.rn.ftz.f32 %f62, %f48, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f63, %f48, %f62, 0f41200000; mul.ftz.f32 %f64, %f61, %f63; mul.ftz.f32 %f65, %f51, %f51; mul.ftz.f32 %f66, %f51, %f65; fma.rn.ftz.f32 %f67, %f51, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f68, %f51, %f67, 0f41200000; mul.ftz.f32 %f69, %f66, %f68; mul.ftz.f32 %f70, %f54, %f54; mul.ftz.f32 %f71, %f54, %f70; fma.rn.ftz.f32 %f72, %f54, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f73, %f54, %f72, 0f41200000; mul.ftz.f32 %f74, %f71, %f73; add.s32 %r11, %r8, -559038708; add.s32 %r12, %r9, -559038708; sub.s32 %r13, %r7, %r9; shf.l.wrap.b32 %r14, %r12, %r12, 4; xor.b32 %r15, %r14, %r13; add.s32 %r16, %r12, %r11; sub.s32 %r17, %r11, %r15; shf.l.wrap.b32 %r18, %r15, %r15, 6; xor.b32 %r19, %r18, %r17; add.s32 %r20, %r15, %r16; sub.s32 %r21, %r16, %r19; shf.l.wrap.b32 %r22, %r19, %r19, 8; xor.b32 %r23, %r22, %r21; add.s32 %r24, %r19, %r20; sub.s32 %r25, %r20, %r23; shf.l.wrap.b32 %r26, %r23, %r23, 16; xor.b32 %r27, %r26, %r25; add.s32 %r28, %r23, %r24; sub.s32 %r29, %r24, %r27; shf.l.wrap.b32 %r30, %r27, %r27, 19; xor.b32 %r31, %r30, %r29; add.s32 %r32, %r27, %r28; sub.s32 %r33, %r28, %r31; shf.l.wrap.b32 %r34, %r31, %r31, 4; add.s32 %r35, %r31, %r32; add.s32 %r36, %r32, %r10; xor.b32 %r37, %r35, %r33; xor.b32 %r38, %r37, %r34; shf.l.wrap.b32 %r39, %r35, %r35, 14; sub.s32 %r40, %r38, %r39; xor.b32 %r41, %r40, %r36; shf.l.wrap.b32 %r42, %r40, %r40, 11; sub.s32 %r43, %r41, %r42; xor.b32 %r44, %r43, %r35; shf.l.wrap.b32 %r45, %r43, %r43, 25; sub.s32 %r46, %r44, %r45; xor.b32 %r47, %r46, %r40; shf.l.wrap.b32 %r48, %r46, %r46, 16; sub.s32 %r49, %r47, %r48; xor.b32 %r50, %r49, %r43; shf.l.wrap.b32 %r51, %r49, %r49, 4; sub.s32 %r52, %r50, %r51; xor.b32 %r53, %r52, %r46; shf.l.wrap.b32 %r54, %r52, %r52, 14; sub.s32 %r55, %r53, %r54; xor.b32 %r56, %r55, %r49; shf.l.wrap.b32 %r57, %r55, %r55, 24; sub.s32 %r58, %r56, %r57; and.b32 %r59, %r58, 31; setp.lt.u32 %p2, %r59, 24; selp.f32 %f75, %f45, %f48, %p2; setp.lt.u32 %p3, %r59, 16; selp.f32 %f76, %f48, %f51, %p3; setp.lt.u32 %p4, %r59, 8; selp.f32 %f77, %f51, %f54, %p4; and.b32 %r60, %r58, 1; setp.eq.b32 %p5, %r60, 1; neg.ftz.f32 %f78, %f75; selp.f32 %f79, %f78, %f75, %p5; and.b32 %r61, %r58, 2; setp.eq.s32 %p6, %r61, 0; neg.ftz.f32 %f80, %f76; selp.f32 %f81, %f76, %f80, %p6; add.ftz.f32 %f82, %f79, %f81; and.b32 %r62, %r58, 4; setp.eq.s32 %p7, %r62, 0; neg.ftz.f32 %f83, %f77; selp.f32 %f84, %f77, %f83, %p7; add.ftz.f32 %f85, %f84, %f82; add.s32 %r63, %r7, 1; sub.s32 %r64, %r63, %r9; xor.b32 %r65, %r14, %r64; sub.s32 %r66, %r11, %r65; shf.l.wrap.b32 %r67, %r65, %r65, 6; xor.b32 %r68, %r67, %r66; add.s32 %r69, %r65, %r16; sub.s32 %r70, %r16, %r68; shf.l.wrap.b32 %r71, %r68, %r68, 8; xor.b32 %r72, %r71, %r70; add.s32 %r73, %r68, %r69; sub.s32 %r74, %r69, %r72; shf.l.wrap.b32 %r75, %r72, %r72, 16; xor.b32 %r76, %r75, %r74; add.s32 %r77, %r72, %r73; sub.s32 %r78, %r73, %r76; shf.l.wrap.b32 %r79, %r76, %r76, 19; xor.b32 %r80, %r79, %r78; add.s32 %r81, %r76, %r77; sub.s32 %r82, %r77, %r80; shf.l.wrap.b32 %r83, %r80, %r80, 4; add.s32 %r84, %r80, %r81; add.s32 %r85, %r81, %r10; xor.b32 %r86, %r84, %r82; xor.b32 %r87, %r86, %r83; shf.l.wrap.b32 %r88, %r84, %r84, 14; sub.s32 %r89, %r87, %r88; xor.b32 %r90, %r89, %r85; shf.l.wrap.b32 %r91, %r89, %r89, 11; sub.s32 %r92, %r90, %r91; xor.b32 %r93, %r92, %r84; shf.l.wrap.b32 %r94, %r92, %r92, 25; sub.s32 %r95, %r93, %r94; xor.b32 %r96, %r95, %r89; shf.l.wrap.b32 %r97, %r95, %r95, 16; sub.s32 %r98, %r96, %r97; xor.b32 %r99, %r98, %r92; shf.l.wrap.b32 %r100, %r98, %r98, 4; sub.s32 %r101, %r99, %r100; xor.b32 %r102, %r101, %r95; shf.l.wrap.b32 %r103, %r101, %r101, 14; sub.s32 %r104, %r102, %r103; xor.b32 %r105, %r104, %r98; shf.l.wrap.b32 %r106, %r104, %r104, 24; sub.s32 %r107, %r105, %r106; add.ftz.f32 %f86, %f45, 0fBF800000; and.b32 %r108, %r107, 31; setp.lt.u32 %p8, %r108, 24; selp.f32 %f88, %f86, %f48, %p8; setp.lt.u32 %p9, %r108, 16; selp.f32 %f89, %f48, %f51, %p9; setp.lt.u32 %p10, %r108, 8; selp.f32 %f90, %f51, %f54, %p10; and.b32 %r109, %r107, 1; setp.eq.b32 %p11, %r109, 1; neg.ftz.f32 %f91, %f88; selp.f32 %f92, %f91, %f88, %p11; and.b32 %r110, %r107, 2; setp.eq.s32 %p12, %r110, 0; neg.ftz.f32 %f93, %f89; selp.f32 %f94, %f89, %f93, %p12; add.ftz.f32 %f95, %f92, %f94; and.b32 %r111, %r107, 4; setp.eq.s32 %p13, %r111, 0; neg.ftz.f32 %f96, %f90; selp.f32 %f97, %f90, %f96, %p13; add.ftz.f32 %f98, %f97, %f95; add.s32 %r112, %r8, -559038707; add.s32 %r113, %r12, %r112; sub.s32 %r114, %r112, %r15; xor.b32 %r115, %r18, %r114; add.s32 %r116, %r15, %r113; sub.s32 %r117, %r113, %r115; shf.l.wrap.b32 %r118, %r115, %r115, 8; xor.b32 %r119, %r118, %r117; add.s32 %r120, %r115, %r116; sub.s32 %r121, %r116, %r119; shf.l.wrap.b32 %r122, %r119, %r119, 16; xor.b32 %r123, %r122, %r121; add.s32 %r124, %r119, %r120; sub.s32 %r125, %r120, %r123; shf.l.wrap.b32 %r126, %r123, %r123, 19; xor.b32 %r127, %r126, %r125; add.s32 %r128, %r123, %r124; sub.s32 %r129, %r124, %r127; shf.l.wrap.b32 %r130, %r127, %r127, 4; add.s32 %r131, %r127, %r128; add.s32 %r132, %r128, %r10; xor.b32 %r133, %r131, %r129; xor.b32 %r134, %r133, %r130; shf.l.wrap.b32 %r135, %r131, %r131, 14; sub.s32 %r136, %r134, %r135; xor.b32 %r137, %r136, %r132; shf.l.wrap.b32 %r138, %r136, %r136, 11; sub.s32 %r139, %r137, %r138; xor.b32 %r140, %r139, %r131; shf.l.wrap.b32 %r141, %r139, %r139, 25; sub.s32 %r142, %r140, %r141; xor.b32 %r143, %r142, %r136; shf.l.wrap.b32 %r144, %r142, %r142, 16; sub.s32 %r145, %r143, %r144; xor.b32 %r146, %r145, %r139; shf.l.wrap.b32 %r147, %r145, %r145, 4; sub.s32 %r148, %r146, %r147; xor.b32 %r149, %r148, %r142; shf.l.wrap.b32 %r150, %r148, %r148, 14; sub.s32 %r151, %r149, %r150; xor.b32 %r152, %r151, %r145; shf.l.wrap.b32 %r153, %r151, %r151, 24; sub.s32 %r154, %r152, %r153; add.ftz.f32 %f99, %f48, 0fBF800000; and.b32 %r155, %r154, 31; setp.lt.u32 %p14, %r155, 24; selp.f32 %f100, %f45, %f99, %p14; setp.lt.u32 %p15, %r155, 16; selp.f32 %f101, %f99, %f51, %p15; setp.lt.u32 %p16, %r155, 8; selp.f32 %f102, %f51, %f54, %p16; and.b32 %r156, %r154, 1; setp.eq.b32 %p17, %r156, 1; neg.ftz.f32 %f103, %f100; selp.f32 %f104, %f103, %f100, %p17; and.b32 %r157, %r154, 2; setp.eq.s32 %p18, %r157, 0; neg.ftz.f32 %f105, %f101; selp.f32 %f106, %f101, %f105, %p18; add.ftz.f32 %f107, %f104, %f106; and.b32 %r158, %r154, 4; setp.eq.s32 %p19, %r158, 0; neg.ftz.f32 %f108, %f102; selp.f32 %f109, %f102, %f108, %p19; add.ftz.f32 %f110, %f109, %f107; sub.s32 %r159, %r112, %r65; xor.b32 %r160, %r67, %r159; add.s32 %r161, %r65, %r113; sub.s32 %r162, %r113, %r160; shf.l.wrap.b32 %r163, %r160, %r160, 8; xor.b32 %r164, %r163, %r162; add.s32 %r165, %r160, %r161; sub.s32 %r166, %r161, %r164; shf.l.wrap.b32 %r167, %r164, %r164, 16; xor.b32 %r168, %r167, %r166; add.s32 %r169, %r164, %r165; sub.s32 %r170, %r165, %r168; shf.l.wrap.b32 %r171, %r168, %r168, 19; xor.b32 %r172, %r171, %r170; add.s32 %r173, %r168, %r169; sub.s32 %r174, %r169, %r172; shf.l.wrap.b32 %r175, %r172, %r172, 4; add.s32 %r176, %r172, %r173; add.s32 %r177, %r173, %r10; xor.b32 %r178, %r176, %r174; xor.b32 %r179, %r178, %r175; shf.l.wrap.b32 %r180, %r176, %r176, 14; sub.s32 %r181, %r179, %r180; xor.b32 %r182, %r181, %r177; shf.l.wrap.b32 %r183, %r181, %r181, 11; sub.s32 %r184, %r182, %r183; xor.b32 %r185, %r184, %r176; shf.l.wrap.b32 %r186, %r184, %r184, 25; sub.s32 %r187, %r185, %r186; xor.b32 %r188, %r187, %r181; shf.l.wrap.b32 %r189, %r187, %r187, 16; sub.s32 %r190, %r188, %r189; xor.b32 %r191, %r190, %r184; shf.l.wrap.b32 %r192, %r190, %r190, 4; sub.s32 %r193, %r191, %r192; xor.b32 %r194, %r193, %r187; shf.l.wrap.b32 %r195, %r193, %r193, 14; sub.s32 %r196, %r194, %r195; xor.b32 %r197, %r196, %r190; shf.l.wrap.b32 %r198, %r196, %r196, 24; sub.s32 %r199, %r197, %r198; and.b32 %r200, %r199, 31; setp.lt.u32 %p20, %r200, 24; selp.f32 %f111, %f86, %f99, %p20; setp.lt.u32 %p21, %r200, 16; selp.f32 %f112, %f99, %f51, %p21; setp.lt.u32 %p22, %r200, 8; selp.f32 %f113, %f51, %f54, %p22; and.b32 %r201, %r199, 1; setp.eq.b32 %p23, %r201, 1; neg.ftz.f32 %f114, %f111; selp.f32 %f115, %f114, %f111, %p23; and.b32 %r202, %r199, 2; setp.eq.s32 %p24, %r202, 0; neg.ftz.f32 %f116, %f112; selp.f32 %f117, %f112, %f116, %p24; add.ftz.f32 %f118, %f115, %f117; and.b32 %r203, %r199, 4; setp.eq.s32 %p25, %r203, 0; neg.ftz.f32 %f119, %f113; selp.f32 %f120, %f113, %f119, %p25; add.ftz.f32 %f121, %f120, %f118; not.b32 %r204, %r9; add.s32 %r205, %r9, -559038707; add.s32 %r206, %r7, %r204; shf.l.wrap.b32 %r207, %r205, %r205, 4; xor.b32 %r208, %r207, %r206; add.s32 %r209, %r205, %r11; sub.s32 %r210, %r11, %r208; shf.l.wrap.b32 %r211, %r208, %r208, 6; xor.b32 %r212, %r211, %r210; add.s32 %r213, %r208, %r209; sub.s32 %r214, %r209, %r212; shf.l.wrap.b32 %r215, %r212, %r212, 8; xor.b32 %r216, %r215, %r214; add.s32 %r217, %r212, %r213; sub.s32 %r218, %r213, %r216; shf.l.wrap.b32 %r219, %r216, %r216, 16; xor.b32 %r220, %r219, %r218; add.s32 %r221, %r216, %r217; sub.s32 %r222, %r217, %r220; shf.l.wrap.b32 %r223, %r220, %r220, 19; xor.b32 %r224, %r223, %r222; add.s32 %r225, %r220, %r221; sub.s32 %r226, %r221, %r224; shf.l.wrap.b32 %r227, %r224, %r224, 4; add.s32 %r228, %r224, %r225; add.s32 %r229, %r225, %r10; xor.b32 %r230, %r228, %r226; xor.b32 %r231, %r230, %r227; shf.l.wrap.b32 %r232, %r228, %r228, 14; sub.s32 %r233, %r231, %r232; xor.b32 %r234, %r233, %r229; shf.l.wrap.b32 %r235, %r233, %r233, 11; sub.s32 %r236, %r234, %r235; xor.b32 %r237, %r236, %r228; shf.l.wrap.b32 %r238, %r236, %r236, 25; sub.s32 %r239, %r237, %r238; xor.b32 %r240, %r239, %r233; shf.l.wrap.b32 %r241, %r239, %r239, 16; sub.s32 %r242, %r240, %r241; xor.b32 %r243, %r242, %r236; shf.l.wrap.b32 %r244, %r242, %r242, 4; sub.s32 %r245, %r243, %r244; xor.b32 %r246, %r245, %r239; shf.l.wrap.b32 %r247, %r245, %r245, 14; sub.s32 %r248, %r246, %r247; xor.b32 %r249, %r248, %r242; shf.l.wrap.b32 %r250, %r248, %r248, 24; sub.s32 %r251, %r249, %r250; add.ftz.f32 %f122, %f51, 0fBF800000; and.b32 %r252, %r251, 31; setp.lt.u32 %p26, %r252, 24; selp.f32 %f123, %f45, %f48, %p26; setp.lt.u32 %p27, %r252, 16; selp.f32 %f124, %f48, %f122, %p27; setp.lt.u32 %p28, %r252, 8; selp.f32 %f125, %f122, %f54, %p28; and.b32 %r253, %r251, 1; setp.eq.b32 %p29, %r253, 1; neg.ftz.f32 %f126, %f123; selp.f32 %f127, %f126, %f123, %p29; and.b32 %r254, %r251, 2; setp.eq.s32 %p30, %r254, 0; neg.ftz.f32 %f128, %f124; selp.f32 %f129, %f124, %f128, %p30; add.ftz.f32 %f130, %f127, %f129; and.b32 %r255, %r251, 4; setp.eq.s32 %p31, %r255, 0; neg.ftz.f32 %f131, %f125; selp.f32 %f132, %f125, %f131, %p31; add.ftz.f32 %f133, %f132, %f130; xor.b32 %r256, %r207, %r13; sub.s32 %r257, %r11, %r256; shf.l.wrap.b32 %r258, %r256, %r256, 6; xor.b32 %r259, %r258, %r257; add.s32 %r260, %r256, %r209; sub.s32 %r261, %r209, %r259; shf.l.wrap.b32 %r262, %r259, %r259, 8; xor.b32 %r263, %r262, %r261; add.s32 %r264, %r259, %r260; sub.s32 %r265, %r260, %r263; shf.l.wrap.b32 %r266, %r263, %r263, 16; xor.b32 %r267, %r266, %r265; add.s32 %r268, %r263, %r264; sub.s32 %r269, %r264, %r267; shf.l.wrap.b32 %r270, %r267, %r267, 19; xor.b32 %r271, %r270, %r269; add.s32 %r272, %r267, %r268; sub.s32 %r273, %r268, %r271; shf.l.wrap.b32 %r274, %r271, %r271, 4; add.s32 %r275, %r271, %r272; add.s32 %r276, %r272, %r10; xor.b32 %r277, %r275, %r273; xor.b32 %r278, %r277, %r274; shf.l.wrap.b32 %r279, %r275, %r275, 14; sub.s32 %r280, %r278, %r279; xor.b32 %r281, %r280, %r276; shf.l.wrap.b32 %r282, %r280, %r280, 11; sub.s32 %r283, %r281, %r282; xor.b32 %r284, %r283, %r275; shf.l.wrap.b32 %r285, %r283, %r283, 25; sub.s32 %r286, %r284, %r285; xor.b32 %r287, %r286, %r280; shf.l.wrap.b32 %r288, %r286, %r286, 16; sub.s32 %r289, %r287, %r288; xor.b32 %r290, %r289, %r283; shf.l.wrap.b32 %r291, %r289, %r289, 4; sub.s32 %r292, %r290, %r291; xor.b32 %r293, %r292, %r286; shf.l.wrap.b32 %r294, %r292, %r292, 14; sub.s32 %r295, %r293, %r294; xor.b32 %r296, %r295, %r289; shf.l.wrap.b32 %r297, %r295, %r295, 24; sub.s32 %r298, %r296, %r297; and.b32 %r299, %r298, 31; setp.lt.u32 %p32, %r299, 24; selp.f32 %f134, %f86, %f48, %p32; setp.lt.u32 %p33, %r299, 16; selp.f32 %f135, %f48, %f122, %p33; setp.lt.u32 %p34, %r299, 8; selp.f32 %f136, %f122, %f54, %p34; and.b32 %r300, %r298, 1; setp.eq.b32 %p35, %r300, 1; neg.ftz.f32 %f137, %f134; selp.f32 %f138, %f137, %f134, %p35; and.b32 %r301, %r298, 2; setp.eq.s32 %p36, %r301, 0; neg.ftz.f32 %f139, %f135; selp.f32 %f140, %f135, %f139, %p36; add.ftz.f32 %f141, %f138, %f140; and.b32 %r302, %r298, 4; setp.eq.s32 %p37, %r302, 0; neg.ftz.f32 %f142, %f136; selp.f32 %f143, %f136, %f142, %p37; add.ftz.f32 %f144, %f143, %f141; add.s32 %r303, %r205, %r112; sub.s32 %r304, %r112, %r208; xor.b32 %r305, %r211, %r304; add.s32 %r306, %r208, %r303; sub.s32 %r307, %r303, %r305; shf.l.wrap.b32 %r308, %r305, %r305, 8; xor.b32 %r309, %r308, %r307; add.s32 %r310, %r305, %r306; sub.s32 %r311, %r306, %r309; shf.l.wrap.b32 %r312, %r309, %r309, 16; xor.b32 %r313, %r312, %r311; add.s32 %r314, %r309, %r310; sub.s32 %r315, %r310, %r313; shf.l.wrap.b32 %r316, %r313, %r313, 19; xor.b32 %r317, %r316, %r315; add.s32 %r318, %r313, %r314; sub.s32 %r319, %r314, %r317; shf.l.wrap.b32 %r320, %r317, %r317, 4; add.s32 %r321, %r317, %r318; add.s32 %r322, %r318, %r10; xor.b32 %r323, %r321, %r319; xor.b32 %r324, %r323, %r320; shf.l.wrap.b32 %r325, %r321, %r321, 14; sub.s32 %r326, %r324, %r325; xor.b32 %r327, %r326, %r322; shf.l.wrap.b32 %r328, %r326, %r326, 11; sub.s32 %r329, %r327, %r328; xor.b32 %r330, %r329, %r321; shf.l.wrap.b32 %r331, %r329, %r329, 25; sub.s32 %r332, %r330, %r331; xor.b32 %r333, %r332, %r326; shf.l.wrap.b32 %r334, %r332, %r332, 16; sub.s32 %r335, %r333, %r334; xor.b32 %r336, %r335, %r329; shf.l.wrap.b32 %r337, %r335, %r335, 4; sub.s32 %r338, %r336, %r337; xor.b32 %r339, %r338, %r332; shf.l.wrap.b32 %r340, %r338, %r338, 14; sub.s32 %r341, %r339, %r340; xor.b32 %r342, %r341, %r335; shf.l.wrap.b32 %r343, %r341, %r341, 24; sub.s32 %r344, %r342, %r343; and.b32 %r345, %r344, 31; setp.lt.u32 %p38, %r345, 24; selp.f32 %f145, %f45, %f99, %p38; setp.lt.u32 %p39, %r345, 16; selp.f32 %f146, %f99, %f122, %p39; setp.lt.u32 %p40, %r345, 8; selp.f32 %f147, %f122, %f54, %p40; and.b32 %r346, %r344, 1; setp.eq.b32 %p41, %r346, 1; neg.ftz.f32 %f148, %f145; selp.f32 %f149, %f148, %f145, %p41; and.b32 %r347, %r344, 2; setp.eq.s32 %p42, %r347, 0; neg.ftz.f32 %f150, %f146; selp.f32 %f151, %f146, %f150, %p42; add.ftz.f32 %f152, %f149, %f151; and.b32 %r348, %r344, 4; setp.eq.s32 %p43, %r348, 0; neg.ftz.f32 %f153, %f147; selp.f32 %f154, %f147, %f153, %p43; add.ftz.f32 %f155, %f154, %f152; sub.s32 %r349, %r112, %r256; xor.b32 %r350, %r258, %r349; add.s32 %r351, %r256, %r303; sub.s32 %r352, %r303, %r350; shf.l.wrap.b32 %r353, %r350, %r350, 8; xor.b32 %r354, %r353, %r352; add.s32 %r355, %r350, %r351; sub.s32 %r356, %r351, %r354; shf.l.wrap.b32 %r357, %r354, %r354, 16; xor.b32 %r358, %r357, %r356; add.s32 %r359, %r354, %r355; sub.s32 %r360, %r355, %r358; shf.l.wrap.b32 %r361, %r358, %r358, 19; xor.b32 %r362, %r361, %r360; add.s32 %r363, %r358, %r359; sub.s32 %r364, %r359, %r362; shf.l.wrap.b32 %r365, %r362, %r362, 4; add.s32 %r366, %r362, %r363; add.s32 %r367, %r363, %r10; xor.b32 %r368, %r366, %r364; xor.b32 %r369, %r368, %r365; shf.l.wrap.b32 %r370, %r366, %r366, 14; sub.s32 %r371, %r369, %r370; xor.b32 %r372, %r371, %r367; shf.l.wrap.b32 %r373, %r371, %r371, 11; sub.s32 %r374, %r372, %r373; xor.b32 %r375, %r374, %r366; shf.l.wrap.b32 %r376, %r374, %r374, 25; sub.s32 %r377, %r375, %r376; xor.b32 %r378, %r377, %r371; shf.l.wrap.b32 %r379, %r377, %r377, 16; sub.s32 %r380, %r378, %r379; xor.b32 %r381, %r380, %r374; shf.l.wrap.b32 %r382, %r380, %r380, 4; sub.s32 %r383, %r381, %r382; xor.b32 %r384, %r383, %r377; shf.l.wrap.b32 %r385, %r383, %r383, 14; sub.s32 %r386, %r384, %r385; xor.b32 %r387, %r386, %r380; shf.l.wrap.b32 %r388, %r386, %r386, 24; sub.s32 %r389, %r387, %r388; and.b32 %r390, %r389, 31; setp.lt.u32 %p44, %r390, 24; selp.f32 %f156, %f86, %f99, %p44; setp.lt.u32 %p45, %r390, 16; selp.f32 %f157, %f99, %f122, %p45; setp.lt.u32 %p46, %r390, 8; selp.f32 %f158, %f122, %f54, %p46; and.b32 %r391, %r389, 1; setp.eq.b32 %p47, %r391, 1; neg.ftz.f32 %f159, %f156; selp.f32 %f160, %f159, %f156, %p47; and.b32 %r392, %r389, 2; setp.eq.s32 %p48, %r392, 0; neg.ftz.f32 %f161, %f157; selp.f32 %f162, %f157, %f161, %p48; add.ftz.f32 %f163, %f160, %f162; and.b32 %r393, %r389, 4; setp.eq.s32 %p49, %r393, 0; neg.ftz.f32 %f164, %f158; selp.f32 %f165, %f158, %f164, %p49; add.ftz.f32 %f166, %f165, %f163; add.s32 %r394, %r10, 1; add.s32 %r395, %r32, %r394; xor.b32 %r396, %r40, %r395; sub.s32 %r397, %r396, %r42; xor.b32 %r398, %r397, %r35; shf.l.wrap.b32 %r399, %r397, %r397, 25; sub.s32 %r400, %r398, %r399; xor.b32 %r401, %r400, %r40; shf.l.wrap.b32 %r402, %r400, %r400, 16; sub.s32 %r403, %r401, %r402; xor.b32 %r404, %r403, %r397; shf.l.wrap.b32 %r405, %r403, %r403, 4; sub.s32 %r406, %r404, %r405; xor.b32 %r407, %r406, %r400; shf.l.wrap.b32 %r408, %r406, %r406, 14; sub.s32 %r409, %r407, %r408; xor.b32 %r410, %r409, %r403; shf.l.wrap.b32 %r411, %r409, %r409, 24; sub.s32 %r412, %r410, %r411; add.ftz.f32 %f167, %f54, 0fBF800000; and.b32 %r413, %r412, 31; setp.lt.u32 %p50, %r413, 24; selp.f32 %f168, %f45, %f48, %p50; setp.lt.u32 %p51, %r413, 16; selp.f32 %f169, %f48, %f51, %p51; setp.lt.u32 %p52, %r413, 8; selp.f32 %f170, %f51, %f167, %p52; and.b32 %r414, %r412, 1; setp.eq.b32 %p53, %r414, 1; neg.ftz.f32 %f171, %f168; selp.f32 %f172, %f171, %f168, %p53; and.b32 %r415, %r412, 2; setp.eq.s32 %p54, %r415, 0; neg.ftz.f32 %f173, %f169; selp.f32 %f174, %f169, %f173, %p54; add.ftz.f32 %f175, %f172, %f174; and.b32 %r416, %r412, 4; setp.eq.s32 %p55, %r416, 0; neg.ftz.f32 %f176, %f170; selp.f32 %f177, %f170, %f176, %p55; add.ftz.f32 %f178, %f177, %f175; add.s32 %r417, %r81, %r394; xor.b32 %r418, %r89, %r417; sub.s32 %r419, %r418, %r91; xor.b32 %r420, %r419, %r84; shf.l.wrap.b32 %r421, %r419, %r419, 25; sub.s32 %r422, %r420, %r421; xor.b32 %r423, %r422, %r89; shf.l.wrap.b32 %r424, %r422, %r422, 16; sub.s32 %r425, %r423, %r424; xor.b32 %r426, %r425, %r419; shf.l.wrap.b32 %r427, %r425, %r425, 4; sub.s32 %r428, %r426, %r427; xor.b32 %r429, %r428, %r422; shf.l.wrap.b32 %r430, %r428, %r428, 14; sub.s32 %r431, %r429, %r430; xor.b32 %r432, %r431, %r425; shf.l.wrap.b32 %r433, %r431, %r431, 24; sub.s32 %r434, %r432, %r433; and.b32 %r435, %r434, 31; setp.lt.u32 %p56, %r435, 24; selp.f32 %f179, %f86, %f48, %p56; setp.lt.u32 %p57, %r435, 16; selp.f32 %f180, %f48, %f51, %p57; setp.lt.u32 %p58, %r435, 8; selp.f32 %f181, %f51, %f167, %p58; and.b32 %r436, %r434, 1; setp.eq.b32 %p59, %r436, 1; neg.ftz.f32 %f182, %f179; selp.f32 %f183, %f182, %f179, %p59; and.b32 %r437, %r434, 2; setp.eq.s32 %p60, %r437, 0; neg.ftz.f32 %f184, %f180; selp.f32 %f185, %f180, %f184, %p60; add.ftz.f32 %f186, %f183, %f185; and.b32 %r438, %r434, 4; setp.eq.s32 %p61, %r438, 0; neg.ftz.f32 %f187, %f181; selp.f32 %f188, %f181, %f187, %p61; add.ftz.f32 %f189, %f188, %f186; add.s32 %r439, %r128, %r394; xor.b32 %r440, %r136, %r439; sub.s32 %r441, %r440, %r138; xor.b32 %r442, %r441, %r131; shf.l.wrap.b32 %r443, %r441, %r441, 25; sub.s32 %r444, %r442, %r443; xor.b32 %r445, %r444, %r136; shf.l.wrap.b32 %r446, %r444, %r444, 16; sub.s32 %r447, %r445, %r446; xor.b32 %r448, %r447, %r441; shf.l.wrap.b32 %r449, %r447, %r447, 4; sub.s32 %r450, %r448, %r449; xor.b32 %r451, %r450, %r444; shf.l.wrap.b32 %r452, %r450, %r450, 14; sub.s32 %r453, %r451, %r452; xor.b32 %r454, %r453, %r447; shf.l.wrap.b32 %r455, %r453, %r453, 24; sub.s32 %r456, %r454, %r455; and.b32 %r457, %r456, 31; setp.lt.u32 %p62, %r457, 24; selp.f32 %f190, %f45, %f99, %p62; setp.lt.u32 %p63, %r457, 16; selp.f32 %f191, %f99, %f51, %p63; setp.lt.u32 %p64, %r457, 8; selp.f32 %f192, %f51, %f167, %p64; and.b32 %r458, %r456, 1; setp.eq.b32 %p65, %r458, 1; neg.ftz.f32 %f193, %f190; selp.f32 %f194, %f193, %f190, %p65; and.b32 %r459, %r456, 2; setp.eq.s32 %p66, %r459, 0; neg.ftz.f32 %f195, %f191; selp.f32 %f196, %f191, %f195, %p66; add.ftz.f32 %f197, %f194, %f196; and.b32 %r460, %r456, 4; setp.eq.s32 %p67, %r460, 0; neg.ftz.f32 %f198, %f192; selp.f32 %f199, %f192, %f198, %p67; add.ftz.f32 %f200, %f199, %f197; add.s32 %r461, %r173, %r394; xor.b32 %r462, %r181, %r461; sub.s32 %r463, %r462, %r183; xor.b32 %r464, %r463, %r176; shf.l.wrap.b32 %r465, %r463, %r463, 25; sub.s32 %r466, %r464, %r465; xor.b32 %r467, %r466, %r181; shf.l.wrap.b32 %r468, %r466, %r466, 16; sub.s32 %r469, %r467, %r468; xor.b32 %r470, %r469, %r463; shf.l.wrap.b32 %r471, %r469, %r469, 4; sub.s32 %r472, %r470, %r471; xor.b32 %r473, %r472, %r466; shf.l.wrap.b32 %r474, %r472, %r472, 14; sub.s32 %r475, %r473, %r474; xor.b32 %r476, %r475, %r469; shf.l.wrap.b32 %r477, %r475, %r475, 24; sub.s32 %r478, %r476, %r477; and.b32 %r479, %r478, 31; setp.lt.u32 %p68, %r479, 24; selp.f32 %f201, %f86, %f99, %p68; setp.lt.u32 %p69, %r479, 16; selp.f32 %f202, %f99, %f51, %p69; setp.lt.u32 %p70, %r479, 8; selp.f32 %f203, %f51, %f167, %p70; and.b32 %r480, %r478, 1; setp.eq.b32 %p71, %r480, 1; neg.ftz.f32 %f204, %f201; selp.f32 %f205, %f204, %f201, %p71; and.b32 %r481, %r478, 2; setp.eq.s32 %p72, %r481, 0; neg.ftz.f32 %f206, %f202; selp.f32 %f207, %f202, %f206, %p72; add.ftz.f32 %f208, %f205, %f207; and.b32 %r482, %r478, 4; setp.eq.s32 %p73, %r482, 0; neg.ftz.f32 %f209, %f203; selp.f32 %f210, %f203, %f209, %p73; add.ftz.f32 %f211, %f210, %f208; add.s32 %r483, %r225, %r394; xor.b32 %r484, %r233, %r483; sub.s32 %r485, %r484, %r235; xor.b32 %r486, %r485, %r228; shf.l.wrap.b32 %r487, %r485, %r485, 25; sub.s32 %r488, %r486, %r487; xor.b32 %r489, %r488, %r233; shf.l.wrap.b32 %r490, %r488, %r488, 16; sub.s32 %r491, %r489, %r490; xor.b32 %r492, %r491, %r485; shf.l.wrap.b32 %r493, %r491, %r491, 4; sub.s32 %r494, %r492, %r493; xor.b32 %r495, %r494, %r488; shf.l.wrap.b32 %r496, %r494, %r494, 14; sub.s32 %r497, %r495, %r496; xor.b32 %r498, %r497, %r491; shf.l.wrap.b32 %r499, %r497, %r497, 24; sub.s32 %r500, %r498, %r499; and.b32 %r501, %r500, 31; setp.lt.u32 %p74, %r501, 24; selp.f32 %f212, %f45, %f48, %p74; setp.lt.u32 %p75, %r501, 16; selp.f32 %f213, %f48, %f122, %p75; setp.lt.u32 %p76, %r501, 8; selp.f32 %f214, %f122, %f167, %p76; and.b32 %r502, %r500, 1; setp.eq.b32 %p77, %r502, 1; neg.ftz.f32 %f215, %f212; selp.f32 %f216, %f215, %f212, %p77; and.b32 %r503, %r500, 2; setp.eq.s32 %p78, %r503, 0; neg.ftz.f32 %f217, %f213; selp.f32 %f218, %f213, %f217, %p78; add.ftz.f32 %f219, %f216, %f218; and.b32 %r504, %r500, 4; setp.eq.s32 %p79, %r504, 0; neg.ftz.f32 %f220, %f214; selp.f32 %f221, %f214, %f220, %p79; add.ftz.f32 %f222, %f221, %f219; add.s32 %r505, %r272, %r394; xor.b32 %r506, %r280, %r505; sub.s32 %r507, %r506, %r282; xor.b32 %r508, %r507, %r275; shf.l.wrap.b32 %r509, %r507, %r507, 25; sub.s32 %r510, %r508, %r509; xor.b32 %r511, %r510, %r280; shf.l.wrap.b32 %r512, %r510, %r510, 16; sub.s32 %r513, %r511, %r512; xor.b32 %r514, %r513, %r507; shf.l.wrap.b32 %r515, %r513, %r513, 4; sub.s32 %r516, %r514, %r515; xor.b32 %r517, %r516, %r510; shf.l.wrap.b32 %r518, %r516, %r516, 14; sub.s32 %r519, %r517, %r518; xor.b32 %r520, %r519, %r513; shf.l.wrap.b32 %r521, %r519, %r519, 24; sub.s32 %r522, %r520, %r521; and.b32 %r523, %r522, 31; setp.lt.u32 %p80, %r523, 24; selp.f32 %f223, %f86, %f48, %p80; setp.lt.u32 %p81, %r523, 16; selp.f32 %f224, %f48, %f122, %p81; setp.lt.u32 %p82, %r523, 8; selp.f32 %f225, %f122, %f167, %p82; and.b32 %r524, %r522, 1; setp.eq.b32 %p83, %r524, 1; neg.ftz.f32 %f226, %f223; selp.f32 %f227, %f226, %f223, %p83; and.b32 %r525, %r522, 2; setp.eq.s32 %p84, %r525, 0; neg.ftz.f32 %f228, %f224; selp.f32 %f229, %f224, %f228, %p84; add.ftz.f32 %f230, %f227, %f229; and.b32 %r526, %r522, 4; setp.eq.s32 %p85, %r526, 0; neg.ftz.f32 %f231, %f225; selp.f32 %f232, %f225, %f231, %p85; add.ftz.f32 %f233, %f232, %f230; add.s32 %r527, %r318, %r394; xor.b32 %r528, %r326, %r527; sub.s32 %r529, %r528, %r328; xor.b32 %r530, %r529, %r321; shf.l.wrap.b32 %r531, %r529, %r529, 25; sub.s32 %r532, %r530, %r531; xor.b32 %r533, %r532, %r326; shf.l.wrap.b32 %r534, %r532, %r532, 16; sub.s32 %r535, %r533, %r534; xor.b32 %r536, %r535, %r529; shf.l.wrap.b32 %r537, %r535, %r535, 4; sub.s32 %r538, %r536, %r537; xor.b32 %r539, %r538, %r532; shf.l.wrap.b32 %r540, %r538, %r538, 14; sub.s32 %r541, %r539, %r540; xor.b32 %r542, %r541, %r535; shf.l.wrap.b32 %r543, %r541, %r541, 24; sub.s32 %r544, %r542, %r543; and.b32 %r545, %r544, 31; setp.lt.u32 %p86, %r545, 24; selp.f32 %f234, %f45, %f99, %p86; setp.lt.u32 %p87, %r545, 16; selp.f32 %f235, %f99, %f122, %p87; setp.lt.u32 %p88, %r545, 8; selp.f32 %f236, %f122, %f167, %p88; and.b32 %r546, %r544, 1; setp.eq.b32 %p89, %r546, 1; neg.ftz.f32 %f237, %f234; selp.f32 %f238, %f237, %f234, %p89; and.b32 %r547, %r544, 2; setp.eq.s32 %p90, %r547, 0; neg.ftz.f32 %f239, %f235; selp.f32 %f240, %f235, %f239, %p90; add.ftz.f32 %f241, %f238, %f240; and.b32 %r548, %r544, 4; setp.eq.s32 %p91, %r548, 0; neg.ftz.f32 %f242, %f236; selp.f32 %f243, %f236, %f242, %p91; add.ftz.f32 %f244, %f243, %f241; add.s32 %r549, %r363, %r394; xor.b32 %r550, %r371, %r549; sub.s32 %r551, %r550, %r373; xor.b32 %r552, %r551, %r366; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r371; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; and.b32 %r567, %r566, 31; setp.lt.u32 %p92, %r567, 24; selp.f32 %f245, %f86, %f99, %p92; setp.lt.u32 %p93, %r567, 16; selp.f32 %f246, %f99, %f122, %p93; setp.lt.u32 %p94, %r567, 8; selp.f32 %f247, %f122, %f167, %p94; and.b32 %r568, %r566, 1; setp.eq.b32 %p95, %r568, 1; neg.ftz.f32 %f248, %f245; selp.f32 %f249, %f248, %f245, %p95; and.b32 %r569, %r566, 2; setp.eq.s32 %p96, %r569, 0; neg.ftz.f32 %f250, %f246; selp.f32 %f251, %f246, %f250, %p96; add.ftz.f32 %f252, %f249, %f251; and.b32 %r570, %r566, 4; setp.eq.s32 %p97, %r570, 0; neg.ftz.f32 %f253, %f247; selp.f32 %f254, %f247, %f253, %p97; add.ftz.f32 %f255, %f254, %f252; sub.ftz.f32 %f256, %f38, %f59; sub.ftz.f32 %f257, %f38, %f64; sub.ftz.f32 %f258, %f38, %f69; mul.ftz.f32 %f259, %f256, %f85; fma.rn.ftz.f32 %f260, %f59, %f98, %f259; mul.ftz.f32 %f261, %f256, %f110; fma.rn.ftz.f32 %f262, %f59, %f121, %f261; mul.ftz.f32 %f263, %f64, %f262; fma.rn.ftz.f32 %f264, %f257, %f260, %f263; mul.ftz.f32 %f265, %f256, %f133; fma.rn.ftz.f32 %f266, %f59, %f144, %f265; mul.ftz.f32 %f267, %f256, %f155; fma.rn.ftz.f32 %f268, %f59, %f166, %f267; mul.ftz.f32 %f269, %f64, %f268; fma.rn.ftz.f32 %f270, %f257, %f266, %f269; mul.ftz.f32 %f271, %f69, %f270; fma.rn.ftz.f32 %f272, %f258, %f264, %f271; mul.ftz.f32 %f273, %f256, %f178; fma.rn.ftz.f32 %f274, %f59, %f189, %f273; mul.ftz.f32 %f275, %f256, %f200; fma.rn.ftz.f32 %f276, %f59, %f211, %f275; mul.ftz.f32 %f277, %f64, %f276; fma.rn.ftz.f32 %f278, %f257, %f274, %f277; mul.ftz.f32 %f279, %f256, %f222; fma.rn.ftz.f32 %f280, %f59, %f233, %f279; mul.ftz.f32 %f281, %f256, %f244; fma.rn.ftz.f32 %f282, %f59, %f255, %f281; mul.ftz.f32 %f283, %f64, %f282; fma.rn.ftz.f32 %f284, %f257, %f280, %f283; mul.ftz.f32 %f285, %f69, %f284; fma.rn.ftz.f32 %f286, %f258, %f278, %f285; sub.ftz.f32 %f287, %f286, %f272; fma.rn.ftz.f32 %f11, %f74, %f287, %f272; mov.b32 %r3, %f11; setp.nan.ftz.f32 %p98, %f11, %f11; mov.f32 %f558, %f40; @%p98 bra $L__BB78_5; or.b32 %r571, %r3, -2147483648; setp.ne.s32 %p99, %r571, -2147483648; add.ftz.f32 %f289, %f11, %f11; setp.eq.ftz.f32 %p100, %f289, %f11; and.pred %p101, %p100, %p99; mov.f32 %f558, %f40; @%p101 bra $L__BB78_5; shl.b32 %r572, %r3, 1; setp.lt.u32 %p102, %r572, -16777215; selp.f32 %f558, %f11, 0f00000000, %p102; $L__BB78_5: mul.ftz.f32 %f290, %f558, 0f3F559B3D; fma.rn.ftz.f32 %f291, %f290, 0f3F000000, 0f3F000000; fma.rn.ftz.f32 %f559, %f561, %f291, %f559; add.ftz.f32 %f560, %f560, %f561; mul.ftz.f32 %f561, %f561, %f6; add.ftz.f32 %f562, %f562, %f562; add.s32 %r1139, %r1139, 1; setp.lt.s32 %p103, %r1139, %r1; @%p103 bra $L__BB78_2; $L__BB78_6: mov.f32 %f553, 0f00000000; ld.param.f32 %f552, [_Z16fractal_noise_4d6float4ff_param_1]; mov.f32 %f551, 0f41700000; max.ftz.f32 %f550, %f552, %f553; min.ftz.f32 %f549, %f550, %f551; cvt.rmi.ftz.f32.f32 %f292, %f549; sub.ftz.f32 %f22, %f549, %f292; setp.eq.ftz.f32 %p104, %f22, 0f00000000; div.approx.ftz.f32 %f564, %f559, %f560; @%p104 bra $L__BB78_11; mul.ftz.f32 %f294, %f1, %f562; cvt.rmi.ftz.f32.f32 %f295, %f294; cvt.rzi.ftz.s32.f32 %r573, %f295; sub.ftz.f32 %f296, %f294, %f295; mul.ftz.f32 %f297, %f2, %f562; cvt.rmi.ftz.f32.f32 %f298, %f297; cvt.rzi.ftz.s32.f32 %r574, %f298; sub.ftz.f32 %f299, %f297, %f298; mul.ftz.f32 %f300, %f3, %f562; cvt.rmi.ftz.f32.f32 %f301, %f300; cvt.rzi.ftz.s32.f32 %r575, %f301; sub.ftz.f32 %f302, %f300, %f301; mul.ftz.f32 %f303, %f4, %f562; cvt.rmi.ftz.f32.f32 %f304, %f303; cvt.rzi.ftz.s32.f32 %r576, %f304; sub.ftz.f32 %f305, %f303, %f304; mul.ftz.f32 %f306, %f296, %f296; mul.ftz.f32 %f307, %f296, %f306; fma.rn.ftz.f32 %f308, %f296, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f309, %f296, %f308, 0f41200000; mul.ftz.f32 %f310, %f307, %f309; mul.ftz.f32 %f311, %f299, %f299; mul.ftz.f32 %f312, %f299, %f311; fma.rn.ftz.f32 %f313, %f299, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f314, %f299, %f313, 0f41200000; mul.ftz.f32 %f315, %f312, %f314; mul.ftz.f32 %f316, %f302, %f302; mul.ftz.f32 %f317, %f302, %f316; fma.rn.ftz.f32 %f318, %f302, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f319, %f302, %f318, 0f41200000; mul.ftz.f32 %f320, %f317, %f319; mul.ftz.f32 %f321, %f305, %f305; mul.ftz.f32 %f322, %f305, %f321; fma.rn.ftz.f32 %f323, %f305, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f324, %f305, %f323, 0f41200000; mul.ftz.f32 %f325, %f322, %f324; add.s32 %r577, %r574, -559038708; add.s32 %r578, %r575, -559038708; sub.s32 %r579, %r573, %r575; shf.l.wrap.b32 %r580, %r578, %r578, 4; xor.b32 %r581, %r580, %r579; add.s32 %r582, %r578, %r577; sub.s32 %r583, %r577, %r581; shf.l.wrap.b32 %r584, %r581, %r581, 6; xor.b32 %r585, %r584, %r583; add.s32 %r586, %r581, %r582; sub.s32 %r587, %r582, %r585; shf.l.wrap.b32 %r588, %r585, %r585, 8; xor.b32 %r589, %r588, %r587; add.s32 %r590, %r585, %r586; sub.s32 %r591, %r586, %r589; shf.l.wrap.b32 %r592, %r589, %r589, 16; xor.b32 %r593, %r592, %r591; add.s32 %r594, %r589, %r590; sub.s32 %r595, %r590, %r593; shf.l.wrap.b32 %r596, %r593, %r593, 19; xor.b32 %r597, %r596, %r595; add.s32 %r598, %r593, %r594; sub.s32 %r599, %r594, %r597; shf.l.wrap.b32 %r600, %r597, %r597, 4; add.s32 %r601, %r597, %r598; add.s32 %r602, %r598, %r576; xor.b32 %r603, %r601, %r599; xor.b32 %r604, %r603, %r600; shf.l.wrap.b32 %r605, %r601, %r601, 14; sub.s32 %r606, %r604, %r605; xor.b32 %r607, %r606, %r602; shf.l.wrap.b32 %r608, %r606, %r606, 11; sub.s32 %r609, %r607, %r608; xor.b32 %r610, %r609, %r601; shf.l.wrap.b32 %r611, %r609, %r609, 25; sub.s32 %r612, %r610, %r611; xor.b32 %r613, %r612, %r606; shf.l.wrap.b32 %r614, %r612, %r612, 16; sub.s32 %r615, %r613, %r614; xor.b32 %r616, %r615, %r609; shf.l.wrap.b32 %r617, %r615, %r615, 4; sub.s32 %r618, %r616, %r617; xor.b32 %r619, %r618, %r612; shf.l.wrap.b32 %r620, %r618, %r618, 14; sub.s32 %r621, %r619, %r620; xor.b32 %r622, %r621, %r615; shf.l.wrap.b32 %r623, %r621, %r621, 24; sub.s32 %r624, %r622, %r623; and.b32 %r625, %r624, 31; setp.lt.u32 %p105, %r625, 24; selp.f32 %f326, %f296, %f299, %p105; setp.lt.u32 %p106, %r625, 16; selp.f32 %f327, %f299, %f302, %p106; setp.lt.u32 %p107, %r625, 8; selp.f32 %f328, %f302, %f305, %p107; and.b32 %r626, %r624, 1; setp.eq.b32 %p108, %r626, 1; neg.ftz.f32 %f329, %f326; selp.f32 %f330, %f329, %f326, %p108; and.b32 %r627, %r624, 2; setp.eq.s32 %p109, %r627, 0; neg.ftz.f32 %f331, %f327; selp.f32 %f332, %f327, %f331, %p109; add.ftz.f32 %f333, %f330, %f332; and.b32 %r628, %r624, 4; setp.eq.s32 %p110, %r628, 0; neg.ftz.f32 %f334, %f328; selp.f32 %f335, %f328, %f334, %p110; add.ftz.f32 %f336, %f335, %f333; add.s32 %r629, %r573, 1; sub.s32 %r630, %r629, %r575; xor.b32 %r631, %r580, %r630; sub.s32 %r632, %r577, %r631; shf.l.wrap.b32 %r633, %r631, %r631, 6; xor.b32 %r634, %r633, %r632; add.s32 %r635, %r631, %r582; sub.s32 %r636, %r582, %r634; shf.l.wrap.b32 %r637, %r634, %r634, 8; xor.b32 %r638, %r637, %r636; add.s32 %r639, %r634, %r635; sub.s32 %r640, %r635, %r638; shf.l.wrap.b32 %r641, %r638, %r638, 16; xor.b32 %r642, %r641, %r640; add.s32 %r643, %r638, %r639; sub.s32 %r644, %r639, %r642; shf.l.wrap.b32 %r645, %r642, %r642, 19; xor.b32 %r646, %r645, %r644; add.s32 %r647, %r642, %r643; sub.s32 %r648, %r643, %r646; shf.l.wrap.b32 %r649, %r646, %r646, 4; add.s32 %r650, %r646, %r647; add.s32 %r651, %r647, %r576; xor.b32 %r652, %r650, %r648; xor.b32 %r653, %r652, %r649; shf.l.wrap.b32 %r654, %r650, %r650, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r651; shf.l.wrap.b32 %r657, %r655, %r655, 11; sub.s32 %r658, %r656, %r657; xor.b32 %r659, %r658, %r650; shf.l.wrap.b32 %r660, %r658, %r658, 25; sub.s32 %r661, %r659, %r660; xor.b32 %r662, %r661, %r655; shf.l.wrap.b32 %r663, %r661, %r661, 16; sub.s32 %r664, %r662, %r663; xor.b32 %r665, %r664, %r658; shf.l.wrap.b32 %r666, %r664, %r664, 4; sub.s32 %r667, %r665, %r666; xor.b32 %r668, %r667, %r661; shf.l.wrap.b32 %r669, %r667, %r667, 14; sub.s32 %r670, %r668, %r669; xor.b32 %r671, %r670, %r664; shf.l.wrap.b32 %r672, %r670, %r670, 24; sub.s32 %r673, %r671, %r672; add.ftz.f32 %f337, %f296, 0fBF800000; mov.f32 %f338, 0f3F800000; and.b32 %r674, %r673, 31; setp.lt.u32 %p111, %r674, 24; selp.f32 %f339, %f337, %f299, %p111; setp.lt.u32 %p112, %r674, 16; selp.f32 %f340, %f299, %f302, %p112; setp.lt.u32 %p113, %r674, 8; selp.f32 %f341, %f302, %f305, %p113; and.b32 %r675, %r673, 1; setp.eq.b32 %p114, %r675, 1; neg.ftz.f32 %f342, %f339; selp.f32 %f343, %f342, %f339, %p114; and.b32 %r676, %r673, 2; setp.eq.s32 %p115, %r676, 0; neg.ftz.f32 %f344, %f340; selp.f32 %f345, %f340, %f344, %p115; add.ftz.f32 %f346, %f343, %f345; and.b32 %r677, %r673, 4; setp.eq.s32 %p116, %r677, 0; neg.ftz.f32 %f347, %f341; selp.f32 %f348, %f341, %f347, %p116; add.ftz.f32 %f349, %f348, %f346; add.s32 %r678, %r574, -559038707; add.s32 %r679, %r578, %r678; sub.s32 %r680, %r678, %r581; xor.b32 %r681, %r584, %r680; add.s32 %r682, %r581, %r679; sub.s32 %r683, %r679, %r681; shf.l.wrap.b32 %r684, %r681, %r681, 8; xor.b32 %r685, %r684, %r683; add.s32 %r686, %r681, %r682; sub.s32 %r687, %r682, %r685; shf.l.wrap.b32 %r688, %r685, %r685, 16; xor.b32 %r689, %r688, %r687; add.s32 %r690, %r685, %r686; sub.s32 %r691, %r686, %r689; shf.l.wrap.b32 %r692, %r689, %r689, 19; xor.b32 %r693, %r692, %r691; add.s32 %r694, %r689, %r690; sub.s32 %r695, %r690, %r693; shf.l.wrap.b32 %r696, %r693, %r693, 4; add.s32 %r697, %r693, %r694; add.s32 %r698, %r694, %r576; xor.b32 %r699, %r697, %r695; xor.b32 %r700, %r699, %r696; shf.l.wrap.b32 %r701, %r697, %r697, 14; sub.s32 %r702, %r700, %r701; xor.b32 %r703, %r702, %r698; shf.l.wrap.b32 %r704, %r702, %r702, 11; sub.s32 %r705, %r703, %r704; xor.b32 %r706, %r705, %r697; shf.l.wrap.b32 %r707, %r705, %r705, 25; sub.s32 %r708, %r706, %r707; xor.b32 %r709, %r708, %r702; shf.l.wrap.b32 %r710, %r708, %r708, 16; sub.s32 %r711, %r709, %r710; xor.b32 %r712, %r711, %r705; shf.l.wrap.b32 %r713, %r711, %r711, 4; sub.s32 %r714, %r712, %r713; xor.b32 %r715, %r714, %r708; shf.l.wrap.b32 %r716, %r714, %r714, 14; sub.s32 %r717, %r715, %r716; xor.b32 %r718, %r717, %r711; shf.l.wrap.b32 %r719, %r717, %r717, 24; sub.s32 %r720, %r718, %r719; add.ftz.f32 %f350, %f299, 0fBF800000; and.b32 %r721, %r720, 31; setp.lt.u32 %p117, %r721, 24; selp.f32 %f351, %f296, %f350, %p117; setp.lt.u32 %p118, %r721, 16; selp.f32 %f352, %f350, %f302, %p118; setp.lt.u32 %p119, %r721, 8; selp.f32 %f353, %f302, %f305, %p119; and.b32 %r722, %r720, 1; setp.eq.b32 %p120, %r722, 1; neg.ftz.f32 %f354, %f351; selp.f32 %f355, %f354, %f351, %p120; and.b32 %r723, %r720, 2; setp.eq.s32 %p121, %r723, 0; neg.ftz.f32 %f356, %f352; selp.f32 %f357, %f352, %f356, %p121; add.ftz.f32 %f358, %f355, %f357; and.b32 %r724, %r720, 4; setp.eq.s32 %p122, %r724, 0; neg.ftz.f32 %f359, %f353; selp.f32 %f360, %f353, %f359, %p122; add.ftz.f32 %f361, %f360, %f358; sub.s32 %r725, %r678, %r631; xor.b32 %r726, %r633, %r725; add.s32 %r727, %r631, %r679; sub.s32 %r728, %r679, %r726; shf.l.wrap.b32 %r729, %r726, %r726, 8; xor.b32 %r730, %r729, %r728; add.s32 %r731, %r726, %r727; sub.s32 %r732, %r727, %r730; shf.l.wrap.b32 %r733, %r730, %r730, 16; xor.b32 %r734, %r733, %r732; add.s32 %r735, %r730, %r731; sub.s32 %r736, %r731, %r734; shf.l.wrap.b32 %r737, %r734, %r734, 19; xor.b32 %r738, %r737, %r736; add.s32 %r739, %r734, %r735; sub.s32 %r740, %r735, %r738; shf.l.wrap.b32 %r741, %r738, %r738, 4; add.s32 %r742, %r738, %r739; add.s32 %r743, %r739, %r576; xor.b32 %r744, %r742, %r740; xor.b32 %r745, %r744, %r741; shf.l.wrap.b32 %r746, %r742, %r742, 14; sub.s32 %r747, %r745, %r746; xor.b32 %r748, %r747, %r743; shf.l.wrap.b32 %r749, %r747, %r747, 11; sub.s32 %r750, %r748, %r749; xor.b32 %r751, %r750, %r742; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r747; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; and.b32 %r766, %r765, 31; setp.lt.u32 %p123, %r766, 24; selp.f32 %f362, %f337, %f350, %p123; setp.lt.u32 %p124, %r766, 16; selp.f32 %f363, %f350, %f302, %p124; setp.lt.u32 %p125, %r766, 8; selp.f32 %f364, %f302, %f305, %p125; and.b32 %r767, %r765, 1; setp.eq.b32 %p126, %r767, 1; neg.ftz.f32 %f365, %f362; selp.f32 %f366, %f365, %f362, %p126; and.b32 %r768, %r765, 2; setp.eq.s32 %p127, %r768, 0; neg.ftz.f32 %f367, %f363; selp.f32 %f368, %f363, %f367, %p127; add.ftz.f32 %f369, %f366, %f368; and.b32 %r769, %r765, 4; setp.eq.s32 %p128, %r769, 0; neg.ftz.f32 %f370, %f364; selp.f32 %f371, %f364, %f370, %p128; add.ftz.f32 %f372, %f371, %f369; not.b32 %r770, %r575; add.s32 %r771, %r575, -559038707; add.s32 %r772, %r573, %r770; shf.l.wrap.b32 %r773, %r771, %r771, 4; xor.b32 %r774, %r773, %r772; add.s32 %r775, %r771, %r577; sub.s32 %r776, %r577, %r774; shf.l.wrap.b32 %r777, %r774, %r774, 6; xor.b32 %r778, %r777, %r776; add.s32 %r779, %r774, %r775; sub.s32 %r780, %r775, %r778; shf.l.wrap.b32 %r781, %r778, %r778, 8; xor.b32 %r782, %r781, %r780; add.s32 %r783, %r778, %r779; sub.s32 %r784, %r779, %r782; shf.l.wrap.b32 %r785, %r782, %r782, 16; xor.b32 %r786, %r785, %r784; add.s32 %r787, %r782, %r783; sub.s32 %r788, %r783, %r786; shf.l.wrap.b32 %r789, %r786, %r786, 19; xor.b32 %r790, %r789, %r788; add.s32 %r791, %r786, %r787; sub.s32 %r792, %r787, %r790; shf.l.wrap.b32 %r793, %r790, %r790, 4; add.s32 %r794, %r790, %r791; add.s32 %r795, %r791, %r576; xor.b32 %r796, %r794, %r792; xor.b32 %r797, %r796, %r793; shf.l.wrap.b32 %r798, %r794, %r794, 14; sub.s32 %r799, %r797, %r798; xor.b32 %r800, %r799, %r795; shf.l.wrap.b32 %r801, %r799, %r799, 11; sub.s32 %r802, %r800, %r801; xor.b32 %r803, %r802, %r794; shf.l.wrap.b32 %r804, %r802, %r802, 25; sub.s32 %r805, %r803, %r804; xor.b32 %r806, %r805, %r799; shf.l.wrap.b32 %r807, %r805, %r805, 16; sub.s32 %r808, %r806, %r807; xor.b32 %r809, %r808, %r802; shf.l.wrap.b32 %r810, %r808, %r808, 4; sub.s32 %r811, %r809, %r810; xor.b32 %r812, %r811, %r805; shf.l.wrap.b32 %r813, %r811, %r811, 14; sub.s32 %r814, %r812, %r813; xor.b32 %r815, %r814, %r808; shf.l.wrap.b32 %r816, %r814, %r814, 24; sub.s32 %r817, %r815, %r816; add.ftz.f32 %f373, %f302, 0fBF800000; and.b32 %r818, %r817, 31; setp.lt.u32 %p129, %r818, 24; selp.f32 %f374, %f296, %f299, %p129; setp.lt.u32 %p130, %r818, 16; selp.f32 %f375, %f299, %f373, %p130; setp.lt.u32 %p131, %r818, 8; selp.f32 %f376, %f373, %f305, %p131; and.b32 %r819, %r817, 1; setp.eq.b32 %p132, %r819, 1; neg.ftz.f32 %f377, %f374; selp.f32 %f378, %f377, %f374, %p132; and.b32 %r820, %r817, 2; setp.eq.s32 %p133, %r820, 0; neg.ftz.f32 %f379, %f375; selp.f32 %f380, %f375, %f379, %p133; add.ftz.f32 %f381, %f378, %f380; and.b32 %r821, %r817, 4; setp.eq.s32 %p134, %r821, 0; neg.ftz.f32 %f382, %f376; selp.f32 %f383, %f376, %f382, %p134; add.ftz.f32 %f384, %f383, %f381; xor.b32 %r822, %r773, %r579; sub.s32 %r823, %r577, %r822; shf.l.wrap.b32 %r824, %r822, %r822, 6; xor.b32 %r825, %r824, %r823; add.s32 %r826, %r822, %r775; sub.s32 %r827, %r775, %r825; shf.l.wrap.b32 %r828, %r825, %r825, 8; xor.b32 %r829, %r828, %r827; add.s32 %r830, %r825, %r826; sub.s32 %r831, %r826, %r829; shf.l.wrap.b32 %r832, %r829, %r829, 16; xor.b32 %r833, %r832, %r831; add.s32 %r834, %r829, %r830; sub.s32 %r835, %r830, %r833; shf.l.wrap.b32 %r836, %r833, %r833, 19; xor.b32 %r837, %r836, %r835; add.s32 %r838, %r833, %r834; sub.s32 %r839, %r834, %r837; shf.l.wrap.b32 %r840, %r837, %r837, 4; add.s32 %r841, %r837, %r838; add.s32 %r842, %r838, %r576; xor.b32 %r843, %r841, %r839; xor.b32 %r844, %r843, %r840; shf.l.wrap.b32 %r845, %r841, %r841, 14; sub.s32 %r846, %r844, %r845; xor.b32 %r847, %r846, %r842; shf.l.wrap.b32 %r848, %r846, %r846, 11; sub.s32 %r849, %r847, %r848; xor.b32 %r850, %r849, %r841; shf.l.wrap.b32 %r851, %r849, %r849, 25; sub.s32 %r852, %r850, %r851; xor.b32 %r853, %r852, %r846; shf.l.wrap.b32 %r854, %r852, %r852, 16; sub.s32 %r855, %r853, %r854; xor.b32 %r856, %r855, %r849; shf.l.wrap.b32 %r857, %r855, %r855, 4; sub.s32 %r858, %r856, %r857; xor.b32 %r859, %r858, %r852; shf.l.wrap.b32 %r860, %r858, %r858, 14; sub.s32 %r861, %r859, %r860; xor.b32 %r862, %r861, %r855; shf.l.wrap.b32 %r863, %r861, %r861, 24; sub.s32 %r864, %r862, %r863; and.b32 %r865, %r864, 31; setp.lt.u32 %p135, %r865, 24; selp.f32 %f385, %f337, %f299, %p135; setp.lt.u32 %p136, %r865, 16; selp.f32 %f386, %f299, %f373, %p136; setp.lt.u32 %p137, %r865, 8; selp.f32 %f387, %f373, %f305, %p137; and.b32 %r866, %r864, 1; setp.eq.b32 %p138, %r866, 1; neg.ftz.f32 %f388, %f385; selp.f32 %f389, %f388, %f385, %p138; and.b32 %r867, %r864, 2; setp.eq.s32 %p139, %r867, 0; neg.ftz.f32 %f390, %f386; selp.f32 %f391, %f386, %f390, %p139; add.ftz.f32 %f392, %f389, %f391; and.b32 %r868, %r864, 4; setp.eq.s32 %p140, %r868, 0; neg.ftz.f32 %f393, %f387; selp.f32 %f394, %f387, %f393, %p140; add.ftz.f32 %f395, %f394, %f392; add.s32 %r869, %r771, %r678; sub.s32 %r870, %r678, %r774; xor.b32 %r871, %r777, %r870; add.s32 %r872, %r774, %r869; sub.s32 %r873, %r869, %r871; shf.l.wrap.b32 %r874, %r871, %r871, 8; xor.b32 %r875, %r874, %r873; add.s32 %r876, %r871, %r872; sub.s32 %r877, %r872, %r875; shf.l.wrap.b32 %r878, %r875, %r875, 16; xor.b32 %r879, %r878, %r877; add.s32 %r880, %r875, %r876; sub.s32 %r881, %r876, %r879; shf.l.wrap.b32 %r882, %r879, %r879, 19; xor.b32 %r883, %r882, %r881; add.s32 %r884, %r879, %r880; sub.s32 %r885, %r880, %r883; shf.l.wrap.b32 %r886, %r883, %r883, 4; add.s32 %r887, %r883, %r884; add.s32 %r888, %r884, %r576; xor.b32 %r889, %r887, %r885; xor.b32 %r890, %r889, %r886; shf.l.wrap.b32 %r891, %r887, %r887, 14; sub.s32 %r892, %r890, %r891; xor.b32 %r893, %r892, %r888; shf.l.wrap.b32 %r894, %r892, %r892, 11; sub.s32 %r895, %r893, %r894; xor.b32 %r896, %r895, %r887; shf.l.wrap.b32 %r897, %r895, %r895, 25; sub.s32 %r898, %r896, %r897; xor.b32 %r899, %r898, %r892; shf.l.wrap.b32 %r900, %r898, %r898, 16; sub.s32 %r901, %r899, %r900; xor.b32 %r902, %r901, %r895; shf.l.wrap.b32 %r903, %r901, %r901, 4; sub.s32 %r904, %r902, %r903; xor.b32 %r905, %r904, %r898; shf.l.wrap.b32 %r906, %r904, %r904, 14; sub.s32 %r907, %r905, %r906; xor.b32 %r908, %r907, %r901; shf.l.wrap.b32 %r909, %r907, %r907, 24; sub.s32 %r910, %r908, %r909; and.b32 %r911, %r910, 31; setp.lt.u32 %p141, %r911, 24; selp.f32 %f396, %f296, %f350, %p141; setp.lt.u32 %p142, %r911, 16; selp.f32 %f397, %f350, %f373, %p142; setp.lt.u32 %p143, %r911, 8; selp.f32 %f398, %f373, %f305, %p143; and.b32 %r912, %r910, 1; setp.eq.b32 %p144, %r912, 1; neg.ftz.f32 %f399, %f396; selp.f32 %f400, %f399, %f396, %p144; and.b32 %r913, %r910, 2; setp.eq.s32 %p145, %r913, 0; neg.ftz.f32 %f401, %f397; selp.f32 %f402, %f397, %f401, %p145; add.ftz.f32 %f403, %f400, %f402; and.b32 %r914, %r910, 4; setp.eq.s32 %p146, %r914, 0; neg.ftz.f32 %f404, %f398; selp.f32 %f405, %f398, %f404, %p146; add.ftz.f32 %f406, %f405, %f403; sub.s32 %r915, %r678, %r822; xor.b32 %r916, %r824, %r915; add.s32 %r917, %r822, %r869; sub.s32 %r918, %r869, %r916; shf.l.wrap.b32 %r919, %r916, %r916, 8; xor.b32 %r920, %r919, %r918; add.s32 %r921, %r916, %r917; sub.s32 %r922, %r917, %r920; shf.l.wrap.b32 %r923, %r920, %r920, 16; xor.b32 %r924, %r923, %r922; add.s32 %r925, %r920, %r921; sub.s32 %r926, %r921, %r924; shf.l.wrap.b32 %r927, %r924, %r924, 19; xor.b32 %r928, %r927, %r926; add.s32 %r929, %r924, %r925; sub.s32 %r930, %r925, %r928; shf.l.wrap.b32 %r931, %r928, %r928, 4; add.s32 %r932, %r928, %r929; add.s32 %r933, %r929, %r576; xor.b32 %r934, %r932, %r930; xor.b32 %r935, %r934, %r931; shf.l.wrap.b32 %r936, %r932, %r932, 14; sub.s32 %r937, %r935, %r936; xor.b32 %r938, %r937, %r933; shf.l.wrap.b32 %r939, %r937, %r937, 11; sub.s32 %r940, %r938, %r939; xor.b32 %r941, %r940, %r932; shf.l.wrap.b32 %r942, %r940, %r940, 25; sub.s32 %r943, %r941, %r942; xor.b32 %r944, %r943, %r937; shf.l.wrap.b32 %r945, %r943, %r943, 16; sub.s32 %r946, %r944, %r945; xor.b32 %r947, %r946, %r940; shf.l.wrap.b32 %r948, %r946, %r946, 4; sub.s32 %r949, %r947, %r948; xor.b32 %r950, %r949, %r943; shf.l.wrap.b32 %r951, %r949, %r949, 14; sub.s32 %r952, %r950, %r951; xor.b32 %r953, %r952, %r946; shf.l.wrap.b32 %r954, %r952, %r952, 24; sub.s32 %r955, %r953, %r954; and.b32 %r956, %r955, 31; setp.lt.u32 %p147, %r956, 24; selp.f32 %f407, %f337, %f350, %p147; setp.lt.u32 %p148, %r956, 16; selp.f32 %f408, %f350, %f373, %p148; setp.lt.u32 %p149, %r956, 8; selp.f32 %f409, %f373, %f305, %p149; and.b32 %r957, %r955, 1; setp.eq.b32 %p150, %r957, 1; neg.ftz.f32 %f410, %f407; selp.f32 %f411, %f410, %f407, %p150; and.b32 %r958, %r955, 2; setp.eq.s32 %p151, %r958, 0; neg.ftz.f32 %f412, %f408; selp.f32 %f413, %f408, %f412, %p151; add.ftz.f32 %f414, %f411, %f413; and.b32 %r959, %r955, 4; setp.eq.s32 %p152, %r959, 0; neg.ftz.f32 %f415, %f409; selp.f32 %f416, %f409, %f415, %p152; add.ftz.f32 %f417, %f416, %f414; add.s32 %r960, %r576, 1; add.s32 %r961, %r598, %r960; xor.b32 %r962, %r606, %r961; sub.s32 %r963, %r962, %r608; xor.b32 %r964, %r963, %r601; shf.l.wrap.b32 %r965, %r963, %r963, 25; sub.s32 %r966, %r964, %r965; xor.b32 %r967, %r966, %r606; shf.l.wrap.b32 %r968, %r966, %r966, 16; sub.s32 %r969, %r967, %r968; xor.b32 %r970, %r969, %r963; shf.l.wrap.b32 %r971, %r969, %r969, 4; sub.s32 %r972, %r970, %r971; xor.b32 %r973, %r972, %r966; shf.l.wrap.b32 %r974, %r972, %r972, 14; sub.s32 %r975, %r973, %r974; xor.b32 %r976, %r975, %r969; shf.l.wrap.b32 %r977, %r975, %r975, 24; sub.s32 %r978, %r976, %r977; add.ftz.f32 %f418, %f305, 0fBF800000; and.b32 %r979, %r978, 31; setp.lt.u32 %p153, %r979, 24; selp.f32 %f419, %f296, %f299, %p153; setp.lt.u32 %p154, %r979, 16; selp.f32 %f420, %f299, %f302, %p154; setp.lt.u32 %p155, %r979, 8; selp.f32 %f421, %f302, %f418, %p155; and.b32 %r980, %r978, 1; setp.eq.b32 %p156, %r980, 1; neg.ftz.f32 %f422, %f419; selp.f32 %f423, %f422, %f419, %p156; and.b32 %r981, %r978, 2; setp.eq.s32 %p157, %r981, 0; neg.ftz.f32 %f424, %f420; selp.f32 %f425, %f420, %f424, %p157; add.ftz.f32 %f426, %f423, %f425; and.b32 %r982, %r978, 4; setp.eq.s32 %p158, %r982, 0; neg.ftz.f32 %f427, %f421; selp.f32 %f428, %f421, %f427, %p158; add.ftz.f32 %f429, %f428, %f426; add.s32 %r983, %r647, %r960; xor.b32 %r984, %r655, %r983; sub.s32 %r985, %r984, %r657; xor.b32 %r986, %r985, %r650; shf.l.wrap.b32 %r987, %r985, %r985, 25; sub.s32 %r988, %r986, %r987; xor.b32 %r989, %r988, %r655; shf.l.wrap.b32 %r990, %r988, %r988, 16; sub.s32 %r991, %r989, %r990; xor.b32 %r992, %r991, %r985; shf.l.wrap.b32 %r993, %r991, %r991, 4; sub.s32 %r994, %r992, %r993; xor.b32 %r995, %r994, %r988; shf.l.wrap.b32 %r996, %r994, %r994, 14; sub.s32 %r997, %r995, %r996; xor.b32 %r998, %r997, %r991; shf.l.wrap.b32 %r999, %r997, %r997, 24; sub.s32 %r1000, %r998, %r999; and.b32 %r1001, %r1000, 31; setp.lt.u32 %p159, %r1001, 24; selp.f32 %f430, %f337, %f299, %p159; setp.lt.u32 %p160, %r1001, 16; selp.f32 %f431, %f299, %f302, %p160; setp.lt.u32 %p161, %r1001, 8; selp.f32 %f432, %f302, %f418, %p161; and.b32 %r1002, %r1000, 1; setp.eq.b32 %p162, %r1002, 1; neg.ftz.f32 %f433, %f430; selp.f32 %f434, %f433, %f430, %p162; and.b32 %r1003, %r1000, 2; setp.eq.s32 %p163, %r1003, 0; neg.ftz.f32 %f435, %f431; selp.f32 %f436, %f431, %f435, %p163; add.ftz.f32 %f437, %f434, %f436; and.b32 %r1004, %r1000, 4; setp.eq.s32 %p164, %r1004, 0; neg.ftz.f32 %f438, %f432; selp.f32 %f439, %f432, %f438, %p164; add.ftz.f32 %f440, %f439, %f437; add.s32 %r1005, %r694, %r960; xor.b32 %r1006, %r702, %r1005; sub.s32 %r1007, %r1006, %r704; xor.b32 %r1008, %r1007, %r697; shf.l.wrap.b32 %r1009, %r1007, %r1007, 25; sub.s32 %r1010, %r1008, %r1009; xor.b32 %r1011, %r1010, %r702; shf.l.wrap.b32 %r1012, %r1010, %r1010, 16; sub.s32 %r1013, %r1011, %r1012; xor.b32 %r1014, %r1013, %r1007; shf.l.wrap.b32 %r1015, %r1013, %r1013, 4; sub.s32 %r1016, %r1014, %r1015; xor.b32 %r1017, %r1016, %r1010; shf.l.wrap.b32 %r1018, %r1016, %r1016, 14; sub.s32 %r1019, %r1017, %r1018; xor.b32 %r1020, %r1019, %r1013; shf.l.wrap.b32 %r1021, %r1019, %r1019, 24; sub.s32 %r1022, %r1020, %r1021; and.b32 %r1023, %r1022, 31; setp.lt.u32 %p165, %r1023, 24; selp.f32 %f441, %f296, %f350, %p165; setp.lt.u32 %p166, %r1023, 16; selp.f32 %f442, %f350, %f302, %p166; setp.lt.u32 %p167, %r1023, 8; selp.f32 %f443, %f302, %f418, %p167; and.b32 %r1024, %r1022, 1; setp.eq.b32 %p168, %r1024, 1; neg.ftz.f32 %f444, %f441; selp.f32 %f445, %f444, %f441, %p168; and.b32 %r1025, %r1022, 2; setp.eq.s32 %p169, %r1025, 0; neg.ftz.f32 %f446, %f442; selp.f32 %f447, %f442, %f446, %p169; add.ftz.f32 %f448, %f445, %f447; and.b32 %r1026, %r1022, 4; setp.eq.s32 %p170, %r1026, 0; neg.ftz.f32 %f449, %f443; selp.f32 %f450, %f443, %f449, %p170; add.ftz.f32 %f451, %f450, %f448; add.s32 %r1027, %r739, %r960; xor.b32 %r1028, %r747, %r1027; sub.s32 %r1029, %r1028, %r749; xor.b32 %r1030, %r1029, %r742; shf.l.wrap.b32 %r1031, %r1029, %r1029, 25; sub.s32 %r1032, %r1030, %r1031; xor.b32 %r1033, %r1032, %r747; shf.l.wrap.b32 %r1034, %r1032, %r1032, 16; sub.s32 %r1035, %r1033, %r1034; xor.b32 %r1036, %r1035, %r1029; shf.l.wrap.b32 %r1037, %r1035, %r1035, 4; sub.s32 %r1038, %r1036, %r1037; xor.b32 %r1039, %r1038, %r1032; shf.l.wrap.b32 %r1040, %r1038, %r1038, 14; sub.s32 %r1041, %r1039, %r1040; xor.b32 %r1042, %r1041, %r1035; shf.l.wrap.b32 %r1043, %r1041, %r1041, 24; sub.s32 %r1044, %r1042, %r1043; and.b32 %r1045, %r1044, 31; setp.lt.u32 %p171, %r1045, 24; selp.f32 %f452, %f337, %f350, %p171; setp.lt.u32 %p172, %r1045, 16; selp.f32 %f453, %f350, %f302, %p172; setp.lt.u32 %p173, %r1045, 8; selp.f32 %f454, %f302, %f418, %p173; and.b32 %r1046, %r1044, 1; setp.eq.b32 %p174, %r1046, 1; neg.ftz.f32 %f455, %f452; selp.f32 %f456, %f455, %f452, %p174; and.b32 %r1047, %r1044, 2; setp.eq.s32 %p175, %r1047, 0; neg.ftz.f32 %f457, %f453; selp.f32 %f458, %f453, %f457, %p175; add.ftz.f32 %f459, %f456, %f458; and.b32 %r1048, %r1044, 4; setp.eq.s32 %p176, %r1048, 0; neg.ftz.f32 %f460, %f454; selp.f32 %f461, %f454, %f460, %p176; add.ftz.f32 %f462, %f461, %f459; add.s32 %r1049, %r791, %r960; xor.b32 %r1050, %r799, %r1049; sub.s32 %r1051, %r1050, %r801; xor.b32 %r1052, %r1051, %r794; shf.l.wrap.b32 %r1053, %r1051, %r1051, 25; sub.s32 %r1054, %r1052, %r1053; xor.b32 %r1055, %r1054, %r799; shf.l.wrap.b32 %r1056, %r1054, %r1054, 16; sub.s32 %r1057, %r1055, %r1056; xor.b32 %r1058, %r1057, %r1051; shf.l.wrap.b32 %r1059, %r1057, %r1057, 4; sub.s32 %r1060, %r1058, %r1059; xor.b32 %r1061, %r1060, %r1054; shf.l.wrap.b32 %r1062, %r1060, %r1060, 14; sub.s32 %r1063, %r1061, %r1062; xor.b32 %r1064, %r1063, %r1057; shf.l.wrap.b32 %r1065, %r1063, %r1063, 24; sub.s32 %r1066, %r1064, %r1065; and.b32 %r1067, %r1066, 31; setp.lt.u32 %p177, %r1067, 24; selp.f32 %f463, %f296, %f299, %p177; setp.lt.u32 %p178, %r1067, 16; selp.f32 %f464, %f299, %f373, %p178; setp.lt.u32 %p179, %r1067, 8; selp.f32 %f465, %f373, %f418, %p179; and.b32 %r1068, %r1066, 1; setp.eq.b32 %p180, %r1068, 1; neg.ftz.f32 %f466, %f463; selp.f32 %f467, %f466, %f463, %p180; and.b32 %r1069, %r1066, 2; setp.eq.s32 %p181, %r1069, 0; neg.ftz.f32 %f468, %f464; selp.f32 %f469, %f464, %f468, %p181; add.ftz.f32 %f470, %f467, %f469; and.b32 %r1070, %r1066, 4; setp.eq.s32 %p182, %r1070, 0; neg.ftz.f32 %f471, %f465; selp.f32 %f472, %f465, %f471, %p182; add.ftz.f32 %f473, %f472, %f470; add.s32 %r1071, %r838, %r960; xor.b32 %r1072, %r846, %r1071; sub.s32 %r1073, %r1072, %r848; xor.b32 %r1074, %r1073, %r841; shf.l.wrap.b32 %r1075, %r1073, %r1073, 25; sub.s32 %r1076, %r1074, %r1075; xor.b32 %r1077, %r1076, %r846; shf.l.wrap.b32 %r1078, %r1076, %r1076, 16; sub.s32 %r1079, %r1077, %r1078; xor.b32 %r1080, %r1079, %r1073; shf.l.wrap.b32 %r1081, %r1079, %r1079, 4; sub.s32 %r1082, %r1080, %r1081; xor.b32 %r1083, %r1082, %r1076; shf.l.wrap.b32 %r1084, %r1082, %r1082, 14; sub.s32 %r1085, %r1083, %r1084; xor.b32 %r1086, %r1085, %r1079; shf.l.wrap.b32 %r1087, %r1085, %r1085, 24; sub.s32 %r1088, %r1086, %r1087; and.b32 %r1089, %r1088, 31; setp.lt.u32 %p183, %r1089, 24; selp.f32 %f474, %f337, %f299, %p183; setp.lt.u32 %p184, %r1089, 16; selp.f32 %f475, %f299, %f373, %p184; setp.lt.u32 %p185, %r1089, 8; selp.f32 %f476, %f373, %f418, %p185; and.b32 %r1090, %r1088, 1; setp.eq.b32 %p186, %r1090, 1; neg.ftz.f32 %f477, %f474; selp.f32 %f478, %f477, %f474, %p186; and.b32 %r1091, %r1088, 2; setp.eq.s32 %p187, %r1091, 0; neg.ftz.f32 %f479, %f475; selp.f32 %f480, %f475, %f479, %p187; add.ftz.f32 %f481, %f478, %f480; and.b32 %r1092, %r1088, 4; setp.eq.s32 %p188, %r1092, 0; neg.ftz.f32 %f482, %f476; selp.f32 %f483, %f476, %f482, %p188; add.ftz.f32 %f484, %f483, %f481; add.s32 %r1093, %r884, %r960; xor.b32 %r1094, %r892, %r1093; sub.s32 %r1095, %r1094, %r894; xor.b32 %r1096, %r1095, %r887; shf.l.wrap.b32 %r1097, %r1095, %r1095, 25; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r892; shf.l.wrap.b32 %r1100, %r1098, %r1098, 16; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 4; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 14; sub.s32 %r1107, %r1105, %r1106; xor.b32 %r1108, %r1107, %r1101; shf.l.wrap.b32 %r1109, %r1107, %r1107, 24; sub.s32 %r1110, %r1108, %r1109; and.b32 %r1111, %r1110, 31; setp.lt.u32 %p189, %r1111, 24; selp.f32 %f485, %f296, %f350, %p189; setp.lt.u32 %p190, %r1111, 16; selp.f32 %f486, %f350, %f373, %p190; setp.lt.u32 %p191, %r1111, 8; selp.f32 %f487, %f373, %f418, %p191; and.b32 %r1112, %r1110, 1; setp.eq.b32 %p192, %r1112, 1; neg.ftz.f32 %f488, %f485; selp.f32 %f489, %f488, %f485, %p192; and.b32 %r1113, %r1110, 2; setp.eq.s32 %p193, %r1113, 0; neg.ftz.f32 %f490, %f486; selp.f32 %f491, %f486, %f490, %p193; add.ftz.f32 %f492, %f489, %f491; and.b32 %r1114, %r1110, 4; setp.eq.s32 %p194, %r1114, 0; neg.ftz.f32 %f493, %f487; selp.f32 %f494, %f487, %f493, %p194; add.ftz.f32 %f495, %f494, %f492; add.s32 %r1115, %r929, %r960; xor.b32 %r1116, %r937, %r1115; sub.s32 %r1117, %r1116, %r939; xor.b32 %r1118, %r1117, %r932; shf.l.wrap.b32 %r1119, %r1117, %r1117, 25; sub.s32 %r1120, %r1118, %r1119; xor.b32 %r1121, %r1120, %r937; shf.l.wrap.b32 %r1122, %r1120, %r1120, 16; sub.s32 %r1123, %r1121, %r1122; xor.b32 %r1124, %r1123, %r1117; shf.l.wrap.b32 %r1125, %r1123, %r1123, 4; sub.s32 %r1126, %r1124, %r1125; xor.b32 %r1127, %r1126, %r1120; shf.l.wrap.b32 %r1128, %r1126, %r1126, 14; sub.s32 %r1129, %r1127, %r1128; xor.b32 %r1130, %r1129, %r1123; shf.l.wrap.b32 %r1131, %r1129, %r1129, 24; sub.s32 %r1132, %r1130, %r1131; and.b32 %r1133, %r1132, 31; setp.lt.u32 %p195, %r1133, 24; selp.f32 %f496, %f337, %f350, %p195; setp.lt.u32 %p196, %r1133, 16; selp.f32 %f497, %f350, %f373, %p196; setp.lt.u32 %p197, %r1133, 8; selp.f32 %f498, %f373, %f418, %p197; and.b32 %r1134, %r1132, 1; setp.eq.b32 %p198, %r1134, 1; neg.ftz.f32 %f499, %f496; selp.f32 %f500, %f499, %f496, %p198; and.b32 %r1135, %r1132, 2; setp.eq.s32 %p199, %r1135, 0; neg.ftz.f32 %f501, %f497; selp.f32 %f502, %f497, %f501, %p199; add.ftz.f32 %f503, %f500, %f502; and.b32 %r1136, %r1132, 4; setp.eq.s32 %p200, %r1136, 0; neg.ftz.f32 %f504, %f498; selp.f32 %f505, %f498, %f504, %p200; add.ftz.f32 %f506, %f505, %f503; sub.ftz.f32 %f507, %f338, %f310; sub.ftz.f32 %f508, %f338, %f315; sub.ftz.f32 %f509, %f338, %f320; mul.ftz.f32 %f510, %f507, %f336; fma.rn.ftz.f32 %f511, %f310, %f349, %f510; mul.ftz.f32 %f512, %f507, %f361; fma.rn.ftz.f32 %f513, %f310, %f372, %f512; mul.ftz.f32 %f514, %f315, %f513; fma.rn.ftz.f32 %f515, %f508, %f511, %f514; mul.ftz.f32 %f516, %f507, %f384; fma.rn.ftz.f32 %f517, %f310, %f395, %f516; mul.ftz.f32 %f518, %f507, %f406; fma.rn.ftz.f32 %f519, %f310, %f417, %f518; mul.ftz.f32 %f520, %f315, %f519; fma.rn.ftz.f32 %f521, %f508, %f517, %f520; mul.ftz.f32 %f522, %f320, %f521; fma.rn.ftz.f32 %f523, %f509, %f515, %f522; mul.ftz.f32 %f524, %f507, %f429; fma.rn.ftz.f32 %f525, %f310, %f440, %f524; mul.ftz.f32 %f526, %f507, %f451; fma.rn.ftz.f32 %f527, %f310, %f462, %f526; mul.ftz.f32 %f528, %f315, %f527; fma.rn.ftz.f32 %f529, %f508, %f525, %f528; mul.ftz.f32 %f530, %f507, %f473; fma.rn.ftz.f32 %f531, %f310, %f484, %f530; mul.ftz.f32 %f532, %f507, %f495; fma.rn.ftz.f32 %f533, %f310, %f506, %f532; mul.ftz.f32 %f534, %f315, %f533; fma.rn.ftz.f32 %f535, %f508, %f531, %f534; mul.ftz.f32 %f536, %f320, %f535; fma.rn.ftz.f32 %f537, %f509, %f529, %f536; sub.ftz.f32 %f538, %f537, %f523; fma.rn.ftz.f32 %f24, %f325, %f538, %f523; mov.b32 %r5, %f24; setp.nan.ftz.f32 %p201, %f24, %f24; mov.f32 %f563, 0f3F000000; @%p201 bra $L__BB78_10; or.b32 %r1137, %r5, -2147483648; setp.ne.s32 %p202, %r1137, -2147483648; add.ftz.f32 %f540, %f24, %f24; setp.eq.ftz.f32 %p203, %f540, %f24; and.pred %p204, %p203, %p202; @%p204 bra $L__BB78_10; shl.b32 %r1138, %r5, 1; setp.lt.u32 %p205, %r1138, -16777215; mul.ftz.f32 %f541, %f24, 0f3F559B3D; fma.rn.ftz.f32 %f542, %f541, 0f3F000000, 0f3F000000; selp.f32 %f563, %f542, 0f3F000000, %p205; $L__BB78_10: fma.rn.ftz.f32 %f543, %f561, %f563, %f559; add.ftz.f32 %f544, %f560, %f561; div.approx.ftz.f32 %f545, %f543, %f544; sub.ftz.f32 %f547, %f338, %f22; mul.ftz.f32 %f548, %f22, %f545; fma.rn.ftz.f32 %f564, %f564, %f547, %f548; $L__BB78_11: st.param.f32 [func_retval0+0], %f564; ret; } .func (.param .b32 func_retval0) _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<601>; .reg .b16 %rs<73>; .reg .f32 %f<1645>; .reg .b32 %r<3110>; .reg .b64 %rd<24>; ld.param.u64 %rd2, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r118, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r120, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r121, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r119, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; bfe.u32 %r1, %r120, 8, 8; bfe.u32 %r2, %r120, 16, 8; shr.u32 %r3, %r120, 24; and.b32 %r4, %r121, 255; bfe.u32 %r5, %r121, 8, 8; bfe.u32 %r6, %r121, 16, 8; shr.u32 %r7, %r121, 24; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r119, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r122, %r123, %r124, %r125}, [%rd6]; ld.global.f32 %f18, [%rd6+16]; shl.b32 %r126, %r120, 2; cvt.u64.u32 %rd7, %r126; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f2, [%rd9]; ld.local.f32 %f3, [%rd9+4]; ld.local.f32 %f4, [%rd9+8]; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB79_2; bra.uni $L__BB79_1; $L__BB79_2: mov.b32 %f1592, %r122; bra.uni $L__BB79_3; $L__BB79_1: mul.wide.u32 %rd10, %r1, 4; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f1592, [%rd11]; $L__BB79_3: setp.eq.s32 %p2, %r2, 255; @%p2 bra $L__BB79_5; bra.uni $L__BB79_4; $L__BB79_5: mov.b32 %f1593, %r123; bra.uni $L__BB79_6; $L__BB79_4: mul.wide.u32 %rd12, %r2, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f1593, [%rd13]; $L__BB79_6: setp.eq.s32 %p3, %r3, 255; @%p3 bra $L__BB79_8; bra.uni $L__BB79_7; $L__BB79_8: mov.b32 %f13, %r124; bra.uni $L__BB79_9; $L__BB79_7: mul.wide.u32 %rd14, %r3, 4; add.s64 %rd15, %rd1, %rd14; ld.local.f32 %f13, [%rd15]; $L__BB79_9: setp.eq.s32 %p4, %r4, 255; @%p4 bra $L__BB79_11; bra.uni $L__BB79_10; $L__BB79_11: mov.b32 %f16, %r125; bra.uni $L__BB79_12; $L__BB79_10: mul.wide.u32 %rd16, %r4, 4; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f16, [%rd17]; $L__BB79_12: setp.eq.s32 %p5, %r5, 255; @%p5 bra $L__BB79_14; mul.wide.u32 %rd18, %r5, 4; add.s64 %rd19, %rd1, %rd18; ld.local.f32 %f18, [%rd19]; $L__BB79_14: mul.ftz.f32 %f1601, %f2, %f1593; mul.ftz.f32 %f1602, %f3, %f1593; mul.ftz.f32 %f1603, %f4, %f1593; mul.ftz.f32 %f1604, %f1592, %f1593; setp.gt.s32 %p6, %r118, 2; @%p6 bra $L__BB79_27; setp.eq.s32 %p9, %r118, 1; @%p9 bra $L__BB79_106; setp.eq.s32 %p10, %r118, 2; @%p10 bra $L__BB79_17; bra.uni $L__BB79_113; $L__BB79_17: setp.eq.ftz.f32 %p554, %f18, 0f00000000; @%p554 bra $L__BB79_25; add.ftz.f32 %f1439, %f1601, 0f433A0802; cvt.rmi.ftz.f32.f32 %f1440, %f1439; cvt.rzi.ftz.s32.f32 %r2871, %f1440; sub.ftz.f32 %f1441, %f1439, %f1440; add.ftz.f32 %f1442, %f1602, 0f42E5E973; cvt.rmi.ftz.f32.f32 %f1443, %f1442; cvt.rzi.ftz.s32.f32 %r2872, %f1443; sub.ftz.f32 %f1444, %f1442, %f1443; mul.ftz.f32 %f1445, %f1441, %f1441; mul.ftz.f32 %f1446, %f1441, %f1445; fma.rn.ftz.f32 %f1447, %f1441, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1448, %f1441, %f1447, 0f41200000; mul.ftz.f32 %f1449, %f1446, %f1448; mul.ftz.f32 %f1450, %f1444, %f1444; mul.ftz.f32 %f1451, %f1444, %f1450; fma.rn.ftz.f32 %f1452, %f1444, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1453, %f1444, %f1452, 0f41200000; mul.ftz.f32 %f1454, %f1451, %f1453; add.s32 %r2873, %r2872, -559038716; add.s32 %r2874, %r2871, -559038716; xor.b32 %r2875, %r2873, -559038716; shf.l.wrap.b32 %r2876, %r2873, %r2873, 14; sub.s32 %r2877, %r2875, %r2876; xor.b32 %r2878, %r2877, %r2874; shf.l.wrap.b32 %r2879, %r2877, %r2877, 11; sub.s32 %r2880, %r2878, %r2879; xor.b32 %r2881, %r2880, %r2873; shf.l.wrap.b32 %r2882, %r2880, %r2880, 25; sub.s32 %r2883, %r2881, %r2882; xor.b32 %r2884, %r2883, %r2877; shf.l.wrap.b32 %r2885, %r2883, %r2883, 16; sub.s32 %r2886, %r2884, %r2885; xor.b32 %r2887, %r2886, %r2880; shf.l.wrap.b32 %r2888, %r2886, %r2886, 4; sub.s32 %r2889, %r2887, %r2888; xor.b32 %r2890, %r2889, %r2883; shf.l.wrap.b32 %r2891, %r2889, %r2889, 14; sub.s32 %r2892, %r2890, %r2891; xor.b32 %r2893, %r2892, %r2886; shf.l.wrap.b32 %r2894, %r2892, %r2892, 24; sub.s32 %r2895, %r2893, %r2894; and.b32 %r2896, %r2895, 4; setp.eq.s32 %p555, %r2896, 0; selp.f32 %f1455, %f1441, %f1444, %p555; selp.f32 %f1456, %f1444, %f1441, %p555; add.ftz.f32 %f1457, %f1456, %f1456; and.b32 %r2897, %r2895, 1; setp.eq.b32 %p556, %r2897, 1; neg.ftz.f32 %f1458, %f1455; selp.f32 %f1459, %f1458, %f1455, %p556; and.b32 %r2898, %r2895, 2; setp.eq.s32 %p557, %r2898, 0; neg.ftz.f32 %f1460, %f1457; selp.f32 %f1461, %f1457, %f1460, %p557; add.ftz.f32 %f1462, %f1459, %f1461; add.s32 %r2899, %r2871, -559038715; xor.b32 %r2900, %r2877, %r2899; sub.s32 %r2901, %r2900, %r2879; xor.b32 %r2902, %r2901, %r2873; shf.l.wrap.b32 %r2903, %r2901, %r2901, 25; sub.s32 %r2904, %r2902, %r2903; xor.b32 %r2905, %r2904, %r2877; shf.l.wrap.b32 %r2906, %r2904, %r2904, 16; sub.s32 %r2907, %r2905, %r2906; xor.b32 %r2908, %r2907, %r2901; shf.l.wrap.b32 %r2909, %r2907, %r2907, 4; sub.s32 %r2910, %r2908, %r2909; xor.b32 %r2911, %r2910, %r2904; shf.l.wrap.b32 %r2912, %r2910, %r2910, 14; sub.s32 %r2913, %r2911, %r2912; xor.b32 %r2914, %r2913, %r2907; shf.l.wrap.b32 %r2915, %r2913, %r2913, 24; sub.s32 %r2916, %r2914, %r2915; add.ftz.f32 %f1463, %f1441, 0fBF800000; mov.f32 %f1464, 0f3F800000; and.b32 %r2917, %r2916, 4; setp.eq.s32 %p558, %r2917, 0; selp.f32 %f1465, %f1463, %f1444, %p558; selp.f32 %f1466, %f1444, %f1463, %p558; add.ftz.f32 %f1467, %f1466, %f1466; and.b32 %r2918, %r2916, 1; setp.eq.b32 %p559, %r2918, 1; neg.ftz.f32 %f1468, %f1465; selp.f32 %f1469, %f1468, %f1465, %p559; and.b32 %r2919, %r2916, 2; setp.eq.s32 %p560, %r2919, 0; neg.ftz.f32 %f1470, %f1467; selp.f32 %f1471, %f1467, %f1470, %p560; add.ftz.f32 %f1472, %f1469, %f1471; add.s32 %r2920, %r2872, -559038715; xor.b32 %r2921, %r2920, -559038716; shf.l.wrap.b32 %r2922, %r2920, %r2920, 14; sub.s32 %r2923, %r2921, %r2922; xor.b32 %r2924, %r2923, %r2874; shf.l.wrap.b32 %r2925, %r2923, %r2923, 11; sub.s32 %r2926, %r2924, %r2925; xor.b32 %r2927, %r2926, %r2920; shf.l.wrap.b32 %r2928, %r2926, %r2926, 25; sub.s32 %r2929, %r2927, %r2928; xor.b32 %r2930, %r2929, %r2923; shf.l.wrap.b32 %r2931, %r2929, %r2929, 16; sub.s32 %r2932, %r2930, %r2931; xor.b32 %r2933, %r2932, %r2926; shf.l.wrap.b32 %r2934, %r2932, %r2932, 4; sub.s32 %r2935, %r2933, %r2934; xor.b32 %r2936, %r2935, %r2929; shf.l.wrap.b32 %r2937, %r2935, %r2935, 14; sub.s32 %r2938, %r2936, %r2937; xor.b32 %r2939, %r2938, %r2932; shf.l.wrap.b32 %r2940, %r2938, %r2938, 24; sub.s32 %r2941, %r2939, %r2940; add.ftz.f32 %f1473, %f1444, 0fBF800000; and.b32 %r2942, %r2941, 4; setp.eq.s32 %p561, %r2942, 0; selp.f32 %f1474, %f1441, %f1473, %p561; selp.f32 %f1475, %f1473, %f1441, %p561; add.ftz.f32 %f1476, %f1475, %f1475; and.b32 %r2943, %r2941, 1; setp.eq.b32 %p562, %r2943, 1; neg.ftz.f32 %f1477, %f1474; selp.f32 %f1478, %f1477, %f1474, %p562; and.b32 %r2944, %r2941, 2; setp.eq.s32 %p563, %r2944, 0; neg.ftz.f32 %f1479, %f1476; selp.f32 %f1480, %f1476, %f1479, %p563; add.ftz.f32 %f1481, %f1478, %f1480; xor.b32 %r2945, %r2923, %r2899; sub.s32 %r2946, %r2945, %r2925; xor.b32 %r2947, %r2946, %r2920; shf.l.wrap.b32 %r2948, %r2946, %r2946, 25; sub.s32 %r2949, %r2947, %r2948; xor.b32 %r2950, %r2949, %r2923; shf.l.wrap.b32 %r2951, %r2949, %r2949, 16; sub.s32 %r2952, %r2950, %r2951; xor.b32 %r2953, %r2952, %r2946; shf.l.wrap.b32 %r2954, %r2952, %r2952, 4; sub.s32 %r2955, %r2953, %r2954; xor.b32 %r2956, %r2955, %r2949; shf.l.wrap.b32 %r2957, %r2955, %r2955, 14; sub.s32 %r2958, %r2956, %r2957; xor.b32 %r2959, %r2958, %r2952; shf.l.wrap.b32 %r2960, %r2958, %r2958, 24; sub.s32 %r2961, %r2959, %r2960; and.b32 %r2962, %r2961, 4; setp.eq.s32 %p564, %r2962, 0; selp.f32 %f1482, %f1463, %f1473, %p564; selp.f32 %f1483, %f1473, %f1463, %p564; add.ftz.f32 %f1484, %f1483, %f1483; and.b32 %r2963, %r2961, 1; setp.eq.b32 %p565, %r2963, 1; neg.ftz.f32 %f1485, %f1482; selp.f32 %f1486, %f1485, %f1482, %p565; and.b32 %r2964, %r2961, 2; setp.eq.s32 %p566, %r2964, 0; neg.ftz.f32 %f1487, %f1484; selp.f32 %f1488, %f1484, %f1487, %p566; add.ftz.f32 %f1489, %f1486, %f1488; sub.ftz.f32 %f1490, %f1464, %f1449; sub.ftz.f32 %f1491, %f1464, %f1454; mul.ftz.f32 %f1492, %f1490, %f1462; fma.rn.ftz.f32 %f1493, %f1449, %f1472, %f1492; mul.ftz.f32 %f1494, %f1490, %f1481; fma.rn.ftz.f32 %f1495, %f1449, %f1489, %f1494; mul.ftz.f32 %f1496, %f1454, %f1495; fma.rn.ftz.f32 %f145, %f1491, %f1493, %f1496; mov.b32 %r115, %f145; setp.nan.ftz.f32 %p567, %f145, %f145; mov.f32 %f1636, 0f00000000; mov.f32 %f1635, %f1636; @%p567 bra $L__BB79_21; or.b32 %r2965, %r115, -2147483648; setp.ne.s32 %p568, %r2965, -2147483648; add.ftz.f32 %f1498, %f145, %f145; setp.eq.ftz.f32 %p569, %f1498, %f145; and.pred %p570, %p569, %p568; mov.f32 %f1635, %f1636; @%p570 bra $L__BB79_21; shl.b32 %r2966, %r115, 1; setp.lt.u32 %p571, %r2966, -16777215; mul.ftz.f32 %f1499, %f145, 0f3F295E9E; selp.f32 %f1635, %f1499, 0f00000000, %p571; $L__BB79_21: add.ftz.f32 %f1501, %f1601, 0f4347D70A; cvt.rmi.ftz.f32.f32 %f1502, %f1501; cvt.rzi.ftz.s32.f32 %r2967, %f1502; sub.ftz.f32 %f1503, %f1501, %f1502; add.ftz.f32 %f1504, %f1602, 0f43224AE8; cvt.rmi.ftz.f32.f32 %f1505, %f1504; cvt.rzi.ftz.s32.f32 %r2968, %f1505; sub.ftz.f32 %f1506, %f1504, %f1505; mul.ftz.f32 %f1507, %f1503, %f1503; mul.ftz.f32 %f1508, %f1503, %f1507; fma.rn.ftz.f32 %f1509, %f1503, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1510, %f1503, %f1509, 0f41200000; mul.ftz.f32 %f1511, %f1508, %f1510; mul.ftz.f32 %f1512, %f1506, %f1506; mul.ftz.f32 %f1513, %f1506, %f1512; fma.rn.ftz.f32 %f1514, %f1506, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1515, %f1506, %f1514, 0f41200000; mul.ftz.f32 %f1516, %f1513, %f1515; add.s32 %r2969, %r2968, -559038716; add.s32 %r2970, %r2967, -559038716; xor.b32 %r2971, %r2969, -559038716; shf.l.wrap.b32 %r2972, %r2969, %r2969, 14; sub.s32 %r2973, %r2971, %r2972; xor.b32 %r2974, %r2973, %r2970; shf.l.wrap.b32 %r2975, %r2973, %r2973, 11; sub.s32 %r2976, %r2974, %r2975; xor.b32 %r2977, %r2976, %r2969; shf.l.wrap.b32 %r2978, %r2976, %r2976, 25; sub.s32 %r2979, %r2977, %r2978; xor.b32 %r2980, %r2979, %r2973; shf.l.wrap.b32 %r2981, %r2979, %r2979, 16; sub.s32 %r2982, %r2980, %r2981; xor.b32 %r2983, %r2982, %r2976; shf.l.wrap.b32 %r2984, %r2982, %r2982, 4; sub.s32 %r2985, %r2983, %r2984; xor.b32 %r2986, %r2985, %r2979; shf.l.wrap.b32 %r2987, %r2985, %r2985, 14; sub.s32 %r2988, %r2986, %r2987; xor.b32 %r2989, %r2988, %r2982; shf.l.wrap.b32 %r2990, %r2988, %r2988, 24; sub.s32 %r2991, %r2989, %r2990; and.b32 %r2992, %r2991, 4; setp.eq.s32 %p572, %r2992, 0; selp.f32 %f1517, %f1503, %f1506, %p572; selp.f32 %f1518, %f1506, %f1503, %p572; add.ftz.f32 %f1519, %f1518, %f1518; and.b32 %r2993, %r2991, 1; setp.eq.b32 %p573, %r2993, 1; neg.ftz.f32 %f1520, %f1517; selp.f32 %f1521, %f1520, %f1517, %p573; and.b32 %r2994, %r2991, 2; setp.eq.s32 %p574, %r2994, 0; neg.ftz.f32 %f1522, %f1519; selp.f32 %f1523, %f1519, %f1522, %p574; add.ftz.f32 %f1524, %f1521, %f1523; add.s32 %r2995, %r2967, -559038715; xor.b32 %r2996, %r2973, %r2995; sub.s32 %r2997, %r2996, %r2975; xor.b32 %r2998, %r2997, %r2969; shf.l.wrap.b32 %r2999, %r2997, %r2997, 25; sub.s32 %r3000, %r2998, %r2999; xor.b32 %r3001, %r3000, %r2973; shf.l.wrap.b32 %r3002, %r3000, %r3000, 16; sub.s32 %r3003, %r3001, %r3002; xor.b32 %r3004, %r3003, %r2997; shf.l.wrap.b32 %r3005, %r3003, %r3003, 4; sub.s32 %r3006, %r3004, %r3005; xor.b32 %r3007, %r3006, %r3000; shf.l.wrap.b32 %r3008, %r3006, %r3006, 14; sub.s32 %r3009, %r3007, %r3008; xor.b32 %r3010, %r3009, %r3003; shf.l.wrap.b32 %r3011, %r3009, %r3009, 24; sub.s32 %r3012, %r3010, %r3011; add.ftz.f32 %f1525, %f1503, 0fBF800000; and.b32 %r3013, %r3012, 4; setp.eq.s32 %p575, %r3013, 0; selp.f32 %f1527, %f1525, %f1506, %p575; selp.f32 %f1528, %f1506, %f1525, %p575; add.ftz.f32 %f1529, %f1528, %f1528; and.b32 %r3014, %r3012, 1; setp.eq.b32 %p576, %r3014, 1; neg.ftz.f32 %f1530, %f1527; selp.f32 %f1531, %f1530, %f1527, %p576; and.b32 %r3015, %r3012, 2; setp.eq.s32 %p577, %r3015, 0; neg.ftz.f32 %f1532, %f1529; selp.f32 %f1533, %f1529, %f1532, %p577; add.ftz.f32 %f1534, %f1531, %f1533; add.s32 %r3016, %r2968, -559038715; xor.b32 %r3017, %r3016, -559038716; shf.l.wrap.b32 %r3018, %r3016, %r3016, 14; sub.s32 %r3019, %r3017, %r3018; xor.b32 %r3020, %r3019, %r2970; shf.l.wrap.b32 %r3021, %r3019, %r3019, 11; sub.s32 %r3022, %r3020, %r3021; xor.b32 %r3023, %r3022, %r3016; shf.l.wrap.b32 %r3024, %r3022, %r3022, 25; sub.s32 %r3025, %r3023, %r3024; xor.b32 %r3026, %r3025, %r3019; shf.l.wrap.b32 %r3027, %r3025, %r3025, 16; sub.s32 %r3028, %r3026, %r3027; xor.b32 %r3029, %r3028, %r3022; shf.l.wrap.b32 %r3030, %r3028, %r3028, 4; sub.s32 %r3031, %r3029, %r3030; xor.b32 %r3032, %r3031, %r3025; shf.l.wrap.b32 %r3033, %r3031, %r3031, 14; sub.s32 %r3034, %r3032, %r3033; xor.b32 %r3035, %r3034, %r3028; shf.l.wrap.b32 %r3036, %r3034, %r3034, 24; sub.s32 %r3037, %r3035, %r3036; add.ftz.f32 %f1535, %f1506, 0fBF800000; and.b32 %r3038, %r3037, 4; setp.eq.s32 %p578, %r3038, 0; selp.f32 %f1536, %f1503, %f1535, %p578; selp.f32 %f1537, %f1535, %f1503, %p578; add.ftz.f32 %f1538, %f1537, %f1537; and.b32 %r3039, %r3037, 1; setp.eq.b32 %p579, %r3039, 1; neg.ftz.f32 %f1539, %f1536; selp.f32 %f1540, %f1539, %f1536, %p579; and.b32 %r3040, %r3037, 2; setp.eq.s32 %p580, %r3040, 0; neg.ftz.f32 %f1541, %f1538; selp.f32 %f1542, %f1538, %f1541, %p580; add.ftz.f32 %f1543, %f1540, %f1542; xor.b32 %r3041, %r3019, %r2995; sub.s32 %r3042, %r3041, %r3021; xor.b32 %r3043, %r3042, %r3016; shf.l.wrap.b32 %r3044, %r3042, %r3042, 25; sub.s32 %r3045, %r3043, %r3044; xor.b32 %r3046, %r3045, %r3019; shf.l.wrap.b32 %r3047, %r3045, %r3045, 16; sub.s32 %r3048, %r3046, %r3047; xor.b32 %r3049, %r3048, %r3042; shf.l.wrap.b32 %r3050, %r3048, %r3048, 4; sub.s32 %r3051, %r3049, %r3050; xor.b32 %r3052, %r3051, %r3045; shf.l.wrap.b32 %r3053, %r3051, %r3051, 14; sub.s32 %r3054, %r3052, %r3053; xor.b32 %r3055, %r3054, %r3048; shf.l.wrap.b32 %r3056, %r3054, %r3054, 24; sub.s32 %r3057, %r3055, %r3056; and.b32 %r3058, %r3057, 4; setp.eq.s32 %p581, %r3058, 0; selp.f32 %f1544, %f1525, %f1535, %p581; selp.f32 %f1545, %f1535, %f1525, %p581; add.ftz.f32 %f1546, %f1545, %f1545; and.b32 %r3059, %r3057, 1; setp.eq.b32 %p582, %r3059, 1; neg.ftz.f32 %f1547, %f1544; selp.f32 %f1548, %f1547, %f1544, %p582; and.b32 %r3060, %r3057, 2; setp.eq.s32 %p583, %r3060, 0; neg.ftz.f32 %f1549, %f1546; selp.f32 %f1550, %f1546, %f1549, %p583; add.ftz.f32 %f1551, %f1548, %f1550; sub.ftz.f32 %f1552, %f1464, %f1511; sub.ftz.f32 %f1553, %f1464, %f1516; mul.ftz.f32 %f1554, %f1552, %f1524; fma.rn.ftz.f32 %f1555, %f1511, %f1534, %f1554; mul.ftz.f32 %f1556, %f1552, %f1543; fma.rn.ftz.f32 %f1557, %f1511, %f1551, %f1556; mul.ftz.f32 %f1558, %f1516, %f1557; fma.rn.ftz.f32 %f148, %f1553, %f1555, %f1558; mov.b32 %r116, %f148; setp.nan.ftz.f32 %p584, %f148, %f148; @%p584 bra $L__BB79_24; or.b32 %r3061, %r116, -2147483648; setp.ne.s32 %p585, %r3061, -2147483648; add.ftz.f32 %f1560, %f148, %f148; setp.eq.ftz.f32 %p586, %f1560, %f148; and.pred %p587, %p586, %p585; @%p587 bra $L__BB79_24; shl.b32 %r3062, %r116, 1; setp.lt.u32 %p588, %r3062, -16777215; mul.ftz.f32 %f1561, %f148, 0f3F295E9E; selp.f32 %f1636, %f1561, 0f00000000, %p588; $L__BB79_24: fma.rn.ftz.f32 %f1601, %f18, %f1635, %f1601; fma.rn.ftz.f32 %f1602, %f18, %f1636, %f1602; $L__BB79_25: { // callseq 54, 0 .reg .b32 temp_param_reg; .param .align 8 .b8 param0[8]; st.param.v2.f32 [param0+0], {%f1601, %f1602}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_2d6float2ff, ( param0, param1, param2 ); ld.param.f32 %f1644, [retval0+0]; } // callseq 54 setp.eq.s32 %p589, %r7, 255; @%p589 bra $L__BB79_113; add.ftz.f32 %f1563, %f1602, 0f431D5E90; add.ftz.f32 %f1564, %f1601, 0f42DF4487; { // callseq 55, 0 .reg .b32 temp_param_reg; .param .align 8 .b8 param0[8]; st.param.v2.f32 [param0+0], {%f1564, %f1563}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_2d6float2ff, ( param0, param1, param2 ); ld.param.f32 %f1643, [retval0+0]; } // callseq 55 add.ftz.f32 %f1565, %f1602, 0f4322C4AA; add.ftz.f32 %f1566, %f1601, 0f432571FE; { // callseq 56, 0 .reg .b32 temp_param_reg; .param .align 8 .b8 param0[8]; st.param.v2.f32 [param0+0], {%f1566, %f1565}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_2d6float2ff, ( param0, param1, param2 ); ld.param.f32 %f1642, [retval0+0]; } // callseq 56 bra.uni $L__BB79_113; $L__BB79_27: setp.eq.s32 %p7, %r118, 3; @%p7 bra $L__BB79_45; setp.ne.s32 %p8, %r118, 4; @%p8 bra $L__BB79_113; setp.eq.ftz.f32 %p11, %f18, 0f00000000; @%p11 bra $L__BB79_43; add.ftz.f32 %f172, %f1601, 0f433A0802; cvt.rmi.ftz.f32.f32 %f173, %f172; cvt.rzi.ftz.s32.f32 %r127, %f173; sub.ftz.f32 %f174, %f172, %f173; add.ftz.f32 %f175, %f1602, 0f42E5E973; cvt.rmi.ftz.f32.f32 %f176, %f175; cvt.rzi.ftz.s32.f32 %r128, %f176; sub.ftz.f32 %f177, %f175, %f176; add.ftz.f32 %f178, %f1603, 0f431A7290; cvt.rmi.ftz.f32.f32 %f179, %f178; cvt.rzi.ftz.s32.f32 %r129, %f179; sub.ftz.f32 %f180, %f178, %f179; add.ftz.f32 %f181, %f1604, 0f430E8142; cvt.rmi.ftz.f32.f32 %f182, %f181; cvt.rzi.ftz.s32.f32 %r130, %f182; sub.ftz.f32 %f183, %f181, %f182; mul.ftz.f32 %f184, %f174, %f174; mul.ftz.f32 %f185, %f174, %f184; fma.rn.ftz.f32 %f186, %f174, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f187, %f174, %f186, 0f41200000; mul.ftz.f32 %f188, %f185, %f187; mul.ftz.f32 %f189, %f177, %f177; mul.ftz.f32 %f190, %f177, %f189; fma.rn.ftz.f32 %f191, %f177, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f192, %f177, %f191, 0f41200000; mul.ftz.f32 %f193, %f190, %f192; mul.ftz.f32 %f194, %f180, %f180; mul.ftz.f32 %f195, %f180, %f194; fma.rn.ftz.f32 %f196, %f180, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f197, %f180, %f196, 0f41200000; mul.ftz.f32 %f198, %f195, %f197; mul.ftz.f32 %f199, %f183, %f183; mul.ftz.f32 %f200, %f183, %f199; fma.rn.ftz.f32 %f201, %f183, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f202, %f183, %f201, 0f41200000; mul.ftz.f32 %f203, %f200, %f202; add.s32 %r131, %r128, -559038708; add.s32 %r132, %r129, -559038708; sub.s32 %r133, %r127, %r129; shf.l.wrap.b32 %r134, %r132, %r132, 4; xor.b32 %r135, %r134, %r133; add.s32 %r136, %r132, %r131; sub.s32 %r137, %r131, %r135; shf.l.wrap.b32 %r138, %r135, %r135, 6; xor.b32 %r139, %r138, %r137; add.s32 %r140, %r135, %r136; sub.s32 %r141, %r136, %r139; shf.l.wrap.b32 %r142, %r139, %r139, 8; xor.b32 %r143, %r142, %r141; add.s32 %r144, %r139, %r140; sub.s32 %r145, %r140, %r143; shf.l.wrap.b32 %r146, %r143, %r143, 16; xor.b32 %r147, %r146, %r145; add.s32 %r148, %r143, %r144; sub.s32 %r149, %r144, %r147; shf.l.wrap.b32 %r150, %r147, %r147, 19; xor.b32 %r151, %r150, %r149; add.s32 %r152, %r147, %r148; sub.s32 %r153, %r148, %r151; shf.l.wrap.b32 %r154, %r151, %r151, 4; add.s32 %r155, %r151, %r152; add.s32 %r156, %r152, %r130; xor.b32 %r157, %r155, %r153; xor.b32 %r158, %r157, %r154; shf.l.wrap.b32 %r159, %r155, %r155, 14; sub.s32 %r160, %r158, %r159; xor.b32 %r161, %r160, %r156; shf.l.wrap.b32 %r162, %r160, %r160, 11; sub.s32 %r163, %r161, %r162; xor.b32 %r164, %r163, %r155; shf.l.wrap.b32 %r165, %r163, %r163, 25; sub.s32 %r166, %r164, %r165; xor.b32 %r167, %r166, %r160; shf.l.wrap.b32 %r168, %r166, %r166, 16; sub.s32 %r169, %r167, %r168; xor.b32 %r170, %r169, %r163; shf.l.wrap.b32 %r171, %r169, %r169, 4; sub.s32 %r172, %r170, %r171; xor.b32 %r173, %r172, %r166; shf.l.wrap.b32 %r174, %r172, %r172, 14; sub.s32 %r175, %r173, %r174; xor.b32 %r176, %r175, %r169; shf.l.wrap.b32 %r177, %r175, %r175, 24; sub.s32 %r178, %r176, %r177; and.b32 %r179, %r178, 31; setp.lt.u32 %p12, %r179, 24; selp.f32 %f204, %f174, %f177, %p12; setp.lt.u32 %p13, %r179, 16; selp.f32 %f205, %f177, %f180, %p13; setp.lt.u32 %p14, %r179, 8; selp.f32 %f206, %f180, %f183, %p14; and.b32 %r180, %r178, 1; setp.eq.b32 %p15, %r180, 1; neg.ftz.f32 %f207, %f204; selp.f32 %f208, %f207, %f204, %p15; and.b32 %r181, %r178, 2; setp.eq.s32 %p16, %r181, 0; neg.ftz.f32 %f209, %f205; selp.f32 %f210, %f205, %f209, %p16; add.ftz.f32 %f211, %f208, %f210; and.b32 %r182, %r178, 4; setp.eq.s32 %p17, %r182, 0; neg.ftz.f32 %f212, %f206; selp.f32 %f213, %f206, %f212, %p17; add.ftz.f32 %f214, %f213, %f211; add.s32 %r183, %r127, 1; sub.s32 %r184, %r183, %r129; xor.b32 %r185, %r134, %r184; sub.s32 %r186, %r131, %r185; shf.l.wrap.b32 %r187, %r185, %r185, 6; xor.b32 %r188, %r187, %r186; add.s32 %r189, %r185, %r136; sub.s32 %r190, %r136, %r188; shf.l.wrap.b32 %r191, %r188, %r188, 8; xor.b32 %r192, %r191, %r190; add.s32 %r193, %r188, %r189; sub.s32 %r194, %r189, %r192; shf.l.wrap.b32 %r195, %r192, %r192, 16; xor.b32 %r196, %r195, %r194; add.s32 %r197, %r192, %r193; sub.s32 %r198, %r193, %r196; shf.l.wrap.b32 %r199, %r196, %r196, 19; xor.b32 %r200, %r199, %r198; add.s32 %r201, %r196, %r197; sub.s32 %r202, %r197, %r200; shf.l.wrap.b32 %r203, %r200, %r200, 4; add.s32 %r204, %r200, %r201; add.s32 %r205, %r201, %r130; xor.b32 %r206, %r204, %r202; xor.b32 %r207, %r206, %r203; shf.l.wrap.b32 %r208, %r204, %r204, 14; sub.s32 %r209, %r207, %r208; xor.b32 %r210, %r209, %r205; shf.l.wrap.b32 %r211, %r209, %r209, 11; sub.s32 %r212, %r210, %r211; xor.b32 %r213, %r212, %r204; shf.l.wrap.b32 %r214, %r212, %r212, 25; sub.s32 %r215, %r213, %r214; xor.b32 %r216, %r215, %r209; shf.l.wrap.b32 %r217, %r215, %r215, 16; sub.s32 %r218, %r216, %r217; xor.b32 %r219, %r218, %r212; shf.l.wrap.b32 %r220, %r218, %r218, 4; sub.s32 %r221, %r219, %r220; xor.b32 %r222, %r221, %r215; shf.l.wrap.b32 %r223, %r221, %r221, 14; sub.s32 %r224, %r222, %r223; xor.b32 %r225, %r224, %r218; shf.l.wrap.b32 %r226, %r224, %r224, 24; sub.s32 %r227, %r225, %r226; add.ftz.f32 %f215, %f174, 0fBF800000; mov.f32 %f216, 0f3F800000; and.b32 %r228, %r227, 31; setp.lt.u32 %p18, %r228, 24; selp.f32 %f217, %f215, %f177, %p18; setp.lt.u32 %p19, %r228, 16; selp.f32 %f218, %f177, %f180, %p19; setp.lt.u32 %p20, %r228, 8; selp.f32 %f219, %f180, %f183, %p20; and.b32 %r229, %r227, 1; setp.eq.b32 %p21, %r229, 1; neg.ftz.f32 %f220, %f217; selp.f32 %f221, %f220, %f217, %p21; and.b32 %r230, %r227, 2; setp.eq.s32 %p22, %r230, 0; neg.ftz.f32 %f222, %f218; selp.f32 %f223, %f218, %f222, %p22; add.ftz.f32 %f224, %f221, %f223; and.b32 %r231, %r227, 4; setp.eq.s32 %p23, %r231, 0; neg.ftz.f32 %f225, %f219; selp.f32 %f226, %f219, %f225, %p23; add.ftz.f32 %f227, %f226, %f224; add.s32 %r232, %r128, -559038707; add.s32 %r233, %r132, %r232; sub.s32 %r234, %r232, %r135; xor.b32 %r235, %r138, %r234; add.s32 %r236, %r135, %r233; sub.s32 %r237, %r233, %r235; shf.l.wrap.b32 %r238, %r235, %r235, 8; xor.b32 %r239, %r238, %r237; add.s32 %r240, %r235, %r236; sub.s32 %r241, %r236, %r239; shf.l.wrap.b32 %r242, %r239, %r239, 16; xor.b32 %r243, %r242, %r241; add.s32 %r244, %r239, %r240; sub.s32 %r245, %r240, %r243; shf.l.wrap.b32 %r246, %r243, %r243, 19; xor.b32 %r247, %r246, %r245; add.s32 %r248, %r243, %r244; sub.s32 %r249, %r244, %r247; shf.l.wrap.b32 %r250, %r247, %r247, 4; add.s32 %r251, %r247, %r248; add.s32 %r252, %r248, %r130; xor.b32 %r253, %r251, %r249; xor.b32 %r254, %r253, %r250; shf.l.wrap.b32 %r255, %r251, %r251, 14; sub.s32 %r256, %r254, %r255; xor.b32 %r257, %r256, %r252; shf.l.wrap.b32 %r258, %r256, %r256, 11; sub.s32 %r259, %r257, %r258; xor.b32 %r260, %r259, %r251; shf.l.wrap.b32 %r261, %r259, %r259, 25; sub.s32 %r262, %r260, %r261; xor.b32 %r263, %r262, %r256; shf.l.wrap.b32 %r264, %r262, %r262, 16; sub.s32 %r265, %r263, %r264; xor.b32 %r266, %r265, %r259; shf.l.wrap.b32 %r267, %r265, %r265, 4; sub.s32 %r268, %r266, %r267; xor.b32 %r269, %r268, %r262; shf.l.wrap.b32 %r270, %r268, %r268, 14; sub.s32 %r271, %r269, %r270; xor.b32 %r272, %r271, %r265; shf.l.wrap.b32 %r273, %r271, %r271, 24; sub.s32 %r274, %r272, %r273; add.ftz.f32 %f228, %f177, 0fBF800000; and.b32 %r275, %r274, 31; setp.lt.u32 %p24, %r275, 24; selp.f32 %f229, %f174, %f228, %p24; setp.lt.u32 %p25, %r275, 16; selp.f32 %f230, %f228, %f180, %p25; setp.lt.u32 %p26, %r275, 8; selp.f32 %f231, %f180, %f183, %p26; and.b32 %r276, %r274, 1; setp.eq.b32 %p27, %r276, 1; neg.ftz.f32 %f232, %f229; selp.f32 %f233, %f232, %f229, %p27; and.b32 %r277, %r274, 2; setp.eq.s32 %p28, %r277, 0; neg.ftz.f32 %f234, %f230; selp.f32 %f235, %f230, %f234, %p28; add.ftz.f32 %f236, %f233, %f235; and.b32 %r278, %r274, 4; setp.eq.s32 %p29, %r278, 0; neg.ftz.f32 %f237, %f231; selp.f32 %f238, %f231, %f237, %p29; add.ftz.f32 %f239, %f238, %f236; sub.s32 %r279, %r232, %r185; xor.b32 %r280, %r187, %r279; add.s32 %r281, %r185, %r233; sub.s32 %r282, %r233, %r280; shf.l.wrap.b32 %r283, %r280, %r280, 8; xor.b32 %r284, %r283, %r282; add.s32 %r285, %r280, %r281; sub.s32 %r286, %r281, %r284; shf.l.wrap.b32 %r287, %r284, %r284, 16; xor.b32 %r288, %r287, %r286; add.s32 %r289, %r284, %r285; sub.s32 %r290, %r285, %r288; shf.l.wrap.b32 %r291, %r288, %r288, 19; xor.b32 %r292, %r291, %r290; add.s32 %r293, %r288, %r289; sub.s32 %r294, %r289, %r292; shf.l.wrap.b32 %r295, %r292, %r292, 4; add.s32 %r296, %r292, %r293; add.s32 %r297, %r293, %r130; xor.b32 %r298, %r296, %r294; xor.b32 %r299, %r298, %r295; shf.l.wrap.b32 %r300, %r296, %r296, 14; sub.s32 %r301, %r299, %r300; xor.b32 %r302, %r301, %r297; shf.l.wrap.b32 %r303, %r301, %r301, 11; sub.s32 %r304, %r302, %r303; xor.b32 %r305, %r304, %r296; shf.l.wrap.b32 %r306, %r304, %r304, 25; sub.s32 %r307, %r305, %r306; xor.b32 %r308, %r307, %r301; shf.l.wrap.b32 %r309, %r307, %r307, 16; sub.s32 %r310, %r308, %r309; xor.b32 %r311, %r310, %r304; shf.l.wrap.b32 %r312, %r310, %r310, 4; sub.s32 %r313, %r311, %r312; xor.b32 %r314, %r313, %r307; shf.l.wrap.b32 %r315, %r313, %r313, 14; sub.s32 %r316, %r314, %r315; xor.b32 %r317, %r316, %r310; shf.l.wrap.b32 %r318, %r316, %r316, 24; sub.s32 %r319, %r317, %r318; and.b32 %r320, %r319, 31; setp.lt.u32 %p30, %r320, 24; selp.f32 %f240, %f215, %f228, %p30; setp.lt.u32 %p31, %r320, 16; selp.f32 %f241, %f228, %f180, %p31; setp.lt.u32 %p32, %r320, 8; selp.f32 %f242, %f180, %f183, %p32; and.b32 %r321, %r319, 1; setp.eq.b32 %p33, %r321, 1; neg.ftz.f32 %f243, %f240; selp.f32 %f244, %f243, %f240, %p33; and.b32 %r322, %r319, 2; setp.eq.s32 %p34, %r322, 0; neg.ftz.f32 %f245, %f241; selp.f32 %f246, %f241, %f245, %p34; add.ftz.f32 %f247, %f244, %f246; and.b32 %r323, %r319, 4; setp.eq.s32 %p35, %r323, 0; neg.ftz.f32 %f248, %f242; selp.f32 %f249, %f242, %f248, %p35; add.ftz.f32 %f250, %f249, %f247; not.b32 %r324, %r129; add.s32 %r325, %r129, -559038707; add.s32 %r326, %r127, %r324; shf.l.wrap.b32 %r327, %r325, %r325, 4; xor.b32 %r328, %r327, %r326; add.s32 %r329, %r325, %r131; sub.s32 %r330, %r131, %r328; shf.l.wrap.b32 %r331, %r328, %r328, 6; xor.b32 %r332, %r331, %r330; add.s32 %r333, %r328, %r329; sub.s32 %r334, %r329, %r332; shf.l.wrap.b32 %r335, %r332, %r332, 8; xor.b32 %r336, %r335, %r334; add.s32 %r337, %r332, %r333; sub.s32 %r338, %r333, %r336; shf.l.wrap.b32 %r339, %r336, %r336, 16; xor.b32 %r340, %r339, %r338; add.s32 %r341, %r336, %r337; sub.s32 %r342, %r337, %r340; shf.l.wrap.b32 %r343, %r340, %r340, 19; xor.b32 %r344, %r343, %r342; add.s32 %r345, %r340, %r341; sub.s32 %r346, %r341, %r344; shf.l.wrap.b32 %r347, %r344, %r344, 4; add.s32 %r348, %r344, %r345; add.s32 %r349, %r345, %r130; xor.b32 %r350, %r348, %r346; xor.b32 %r351, %r350, %r347; shf.l.wrap.b32 %r352, %r348, %r348, 14; sub.s32 %r353, %r351, %r352; xor.b32 %r354, %r353, %r349; shf.l.wrap.b32 %r355, %r353, %r353, 11; sub.s32 %r356, %r354, %r355; xor.b32 %r357, %r356, %r348; shf.l.wrap.b32 %r358, %r356, %r356, 25; sub.s32 %r359, %r357, %r358; xor.b32 %r360, %r359, %r353; shf.l.wrap.b32 %r361, %r359, %r359, 16; sub.s32 %r362, %r360, %r361; xor.b32 %r363, %r362, %r356; shf.l.wrap.b32 %r364, %r362, %r362, 4; sub.s32 %r365, %r363, %r364; xor.b32 %r366, %r365, %r359; shf.l.wrap.b32 %r367, %r365, %r365, 14; sub.s32 %r368, %r366, %r367; xor.b32 %r369, %r368, %r362; shf.l.wrap.b32 %r370, %r368, %r368, 24; sub.s32 %r371, %r369, %r370; add.ftz.f32 %f251, %f180, 0fBF800000; and.b32 %r372, %r371, 31; setp.lt.u32 %p36, %r372, 24; selp.f32 %f252, %f174, %f177, %p36; setp.lt.u32 %p37, %r372, 16; selp.f32 %f253, %f177, %f251, %p37; setp.lt.u32 %p38, %r372, 8; selp.f32 %f254, %f251, %f183, %p38; and.b32 %r373, %r371, 1; setp.eq.b32 %p39, %r373, 1; neg.ftz.f32 %f255, %f252; selp.f32 %f256, %f255, %f252, %p39; and.b32 %r374, %r371, 2; setp.eq.s32 %p40, %r374, 0; neg.ftz.f32 %f257, %f253; selp.f32 %f258, %f253, %f257, %p40; add.ftz.f32 %f259, %f256, %f258; and.b32 %r375, %r371, 4; setp.eq.s32 %p41, %r375, 0; neg.ftz.f32 %f260, %f254; selp.f32 %f261, %f254, %f260, %p41; add.ftz.f32 %f262, %f261, %f259; xor.b32 %r376, %r327, %r133; sub.s32 %r377, %r131, %r376; shf.l.wrap.b32 %r378, %r376, %r376, 6; xor.b32 %r379, %r378, %r377; add.s32 %r380, %r376, %r329; sub.s32 %r381, %r329, %r379; shf.l.wrap.b32 %r382, %r379, %r379, 8; xor.b32 %r383, %r382, %r381; add.s32 %r384, %r379, %r380; sub.s32 %r385, %r380, %r383; shf.l.wrap.b32 %r386, %r383, %r383, 16; xor.b32 %r387, %r386, %r385; add.s32 %r388, %r383, %r384; sub.s32 %r389, %r384, %r387; shf.l.wrap.b32 %r390, %r387, %r387, 19; xor.b32 %r391, %r390, %r389; add.s32 %r392, %r387, %r388; sub.s32 %r393, %r388, %r391; shf.l.wrap.b32 %r394, %r391, %r391, 4; add.s32 %r395, %r391, %r392; add.s32 %r396, %r392, %r130; xor.b32 %r397, %r395, %r393; xor.b32 %r398, %r397, %r394; shf.l.wrap.b32 %r399, %r395, %r395, 14; sub.s32 %r400, %r398, %r399; xor.b32 %r401, %r400, %r396; shf.l.wrap.b32 %r402, %r400, %r400, 11; sub.s32 %r403, %r401, %r402; xor.b32 %r404, %r403, %r395; shf.l.wrap.b32 %r405, %r403, %r403, 25; sub.s32 %r406, %r404, %r405; xor.b32 %r407, %r406, %r400; shf.l.wrap.b32 %r408, %r406, %r406, 16; sub.s32 %r409, %r407, %r408; xor.b32 %r410, %r409, %r403; shf.l.wrap.b32 %r411, %r409, %r409, 4; sub.s32 %r412, %r410, %r411; xor.b32 %r413, %r412, %r406; shf.l.wrap.b32 %r414, %r412, %r412, 14; sub.s32 %r415, %r413, %r414; xor.b32 %r416, %r415, %r409; shf.l.wrap.b32 %r417, %r415, %r415, 24; sub.s32 %r418, %r416, %r417; and.b32 %r419, %r418, 31; setp.lt.u32 %p42, %r419, 24; selp.f32 %f263, %f215, %f177, %p42; setp.lt.u32 %p43, %r419, 16; selp.f32 %f264, %f177, %f251, %p43; setp.lt.u32 %p44, %r419, 8; selp.f32 %f265, %f251, %f183, %p44; and.b32 %r420, %r418, 1; setp.eq.b32 %p45, %r420, 1; neg.ftz.f32 %f266, %f263; selp.f32 %f267, %f266, %f263, %p45; and.b32 %r421, %r418, 2; setp.eq.s32 %p46, %r421, 0; neg.ftz.f32 %f268, %f264; selp.f32 %f269, %f264, %f268, %p46; add.ftz.f32 %f270, %f267, %f269; and.b32 %r422, %r418, 4; setp.eq.s32 %p47, %r422, 0; neg.ftz.f32 %f271, %f265; selp.f32 %f272, %f265, %f271, %p47; add.ftz.f32 %f273, %f272, %f270; add.s32 %r423, %r325, %r232; sub.s32 %r424, %r232, %r328; xor.b32 %r425, %r331, %r424; add.s32 %r426, %r328, %r423; sub.s32 %r427, %r423, %r425; shf.l.wrap.b32 %r428, %r425, %r425, 8; xor.b32 %r429, %r428, %r427; add.s32 %r430, %r425, %r426; sub.s32 %r431, %r426, %r429; shf.l.wrap.b32 %r432, %r429, %r429, 16; xor.b32 %r433, %r432, %r431; add.s32 %r434, %r429, %r430; sub.s32 %r435, %r430, %r433; shf.l.wrap.b32 %r436, %r433, %r433, 19; xor.b32 %r437, %r436, %r435; add.s32 %r438, %r433, %r434; sub.s32 %r439, %r434, %r437; shf.l.wrap.b32 %r440, %r437, %r437, 4; add.s32 %r441, %r437, %r438; add.s32 %r442, %r438, %r130; xor.b32 %r443, %r441, %r439; xor.b32 %r444, %r443, %r440; shf.l.wrap.b32 %r445, %r441, %r441, 14; sub.s32 %r446, %r444, %r445; xor.b32 %r447, %r446, %r442; shf.l.wrap.b32 %r448, %r446, %r446, 11; sub.s32 %r449, %r447, %r448; xor.b32 %r450, %r449, %r441; shf.l.wrap.b32 %r451, %r449, %r449, 25; sub.s32 %r452, %r450, %r451; xor.b32 %r453, %r452, %r446; shf.l.wrap.b32 %r454, %r452, %r452, 16; sub.s32 %r455, %r453, %r454; xor.b32 %r456, %r455, %r449; shf.l.wrap.b32 %r457, %r455, %r455, 4; sub.s32 %r458, %r456, %r457; xor.b32 %r459, %r458, %r452; shf.l.wrap.b32 %r460, %r458, %r458, 14; sub.s32 %r461, %r459, %r460; xor.b32 %r462, %r461, %r455; shf.l.wrap.b32 %r463, %r461, %r461, 24; sub.s32 %r464, %r462, %r463; and.b32 %r465, %r464, 31; setp.lt.u32 %p48, %r465, 24; selp.f32 %f274, %f174, %f228, %p48; setp.lt.u32 %p49, %r465, 16; selp.f32 %f275, %f228, %f251, %p49; setp.lt.u32 %p50, %r465, 8; selp.f32 %f276, %f251, %f183, %p50; and.b32 %r466, %r464, 1; setp.eq.b32 %p51, %r466, 1; neg.ftz.f32 %f277, %f274; selp.f32 %f278, %f277, %f274, %p51; and.b32 %r467, %r464, 2; setp.eq.s32 %p52, %r467, 0; neg.ftz.f32 %f279, %f275; selp.f32 %f280, %f275, %f279, %p52; add.ftz.f32 %f281, %f278, %f280; and.b32 %r468, %r464, 4; setp.eq.s32 %p53, %r468, 0; neg.ftz.f32 %f282, %f276; selp.f32 %f283, %f276, %f282, %p53; add.ftz.f32 %f284, %f283, %f281; sub.s32 %r469, %r232, %r376; xor.b32 %r470, %r378, %r469; add.s32 %r471, %r376, %r423; sub.s32 %r472, %r423, %r470; shf.l.wrap.b32 %r473, %r470, %r470, 8; xor.b32 %r474, %r473, %r472; add.s32 %r475, %r470, %r471; sub.s32 %r476, %r471, %r474; shf.l.wrap.b32 %r477, %r474, %r474, 16; xor.b32 %r478, %r477, %r476; add.s32 %r479, %r474, %r475; sub.s32 %r480, %r475, %r478; shf.l.wrap.b32 %r481, %r478, %r478, 19; xor.b32 %r482, %r481, %r480; add.s32 %r483, %r478, %r479; sub.s32 %r484, %r479, %r482; shf.l.wrap.b32 %r485, %r482, %r482, 4; add.s32 %r486, %r482, %r483; add.s32 %r487, %r483, %r130; xor.b32 %r488, %r486, %r484; xor.b32 %r489, %r488, %r485; shf.l.wrap.b32 %r490, %r486, %r486, 14; sub.s32 %r491, %r489, %r490; xor.b32 %r492, %r491, %r487; shf.l.wrap.b32 %r493, %r491, %r491, 11; sub.s32 %r494, %r492, %r493; xor.b32 %r495, %r494, %r486; shf.l.wrap.b32 %r496, %r494, %r494, 25; sub.s32 %r497, %r495, %r496; xor.b32 %r498, %r497, %r491; shf.l.wrap.b32 %r499, %r497, %r497, 16; sub.s32 %r500, %r498, %r499; xor.b32 %r501, %r500, %r494; shf.l.wrap.b32 %r502, %r500, %r500, 4; sub.s32 %r503, %r501, %r502; xor.b32 %r504, %r503, %r497; shf.l.wrap.b32 %r505, %r503, %r503, 14; sub.s32 %r506, %r504, %r505; xor.b32 %r507, %r506, %r500; shf.l.wrap.b32 %r508, %r506, %r506, 24; sub.s32 %r509, %r507, %r508; and.b32 %r510, %r509, 31; setp.lt.u32 %p54, %r510, 24; selp.f32 %f285, %f215, %f228, %p54; setp.lt.u32 %p55, %r510, 16; selp.f32 %f286, %f228, %f251, %p55; setp.lt.u32 %p56, %r510, 8; selp.f32 %f287, %f251, %f183, %p56; and.b32 %r511, %r509, 1; setp.eq.b32 %p57, %r511, 1; neg.ftz.f32 %f288, %f285; selp.f32 %f289, %f288, %f285, %p57; and.b32 %r512, %r509, 2; setp.eq.s32 %p58, %r512, 0; neg.ftz.f32 %f290, %f286; selp.f32 %f291, %f286, %f290, %p58; add.ftz.f32 %f292, %f289, %f291; and.b32 %r513, %r509, 4; setp.eq.s32 %p59, %r513, 0; neg.ftz.f32 %f293, %f287; selp.f32 %f294, %f287, %f293, %p59; add.ftz.f32 %f295, %f294, %f292; add.s32 %r514, %r130, 1; add.s32 %r515, %r152, %r514; xor.b32 %r516, %r160, %r515; sub.s32 %r517, %r516, %r162; xor.b32 %r518, %r517, %r155; shf.l.wrap.b32 %r519, %r517, %r517, 25; sub.s32 %r520, %r518, %r519; xor.b32 %r521, %r520, %r160; shf.l.wrap.b32 %r522, %r520, %r520, 16; sub.s32 %r523, %r521, %r522; xor.b32 %r524, %r523, %r517; shf.l.wrap.b32 %r525, %r523, %r523, 4; sub.s32 %r526, %r524, %r525; xor.b32 %r527, %r526, %r520; shf.l.wrap.b32 %r528, %r526, %r526, 14; sub.s32 %r529, %r527, %r528; xor.b32 %r530, %r529, %r523; shf.l.wrap.b32 %r531, %r529, %r529, 24; sub.s32 %r532, %r530, %r531; add.ftz.f32 %f296, %f183, 0fBF800000; and.b32 %r533, %r532, 31; setp.lt.u32 %p60, %r533, 24; selp.f32 %f297, %f174, %f177, %p60; setp.lt.u32 %p61, %r533, 16; selp.f32 %f298, %f177, %f180, %p61; setp.lt.u32 %p62, %r533, 8; selp.f32 %f299, %f180, %f296, %p62; and.b32 %r534, %r532, 1; setp.eq.b32 %p63, %r534, 1; neg.ftz.f32 %f300, %f297; selp.f32 %f301, %f300, %f297, %p63; and.b32 %r535, %r532, 2; setp.eq.s32 %p64, %r535, 0; neg.ftz.f32 %f302, %f298; selp.f32 %f303, %f298, %f302, %p64; add.ftz.f32 %f304, %f301, %f303; and.b32 %r536, %r532, 4; setp.eq.s32 %p65, %r536, 0; neg.ftz.f32 %f305, %f299; selp.f32 %f306, %f299, %f305, %p65; add.ftz.f32 %f307, %f306, %f304; add.s32 %r537, %r201, %r514; xor.b32 %r538, %r209, %r537; sub.s32 %r539, %r538, %r211; xor.b32 %r540, %r539, %r204; shf.l.wrap.b32 %r541, %r539, %r539, 25; sub.s32 %r542, %r540, %r541; xor.b32 %r543, %r542, %r209; shf.l.wrap.b32 %r544, %r542, %r542, 16; sub.s32 %r545, %r543, %r544; xor.b32 %r546, %r545, %r539; shf.l.wrap.b32 %r547, %r545, %r545, 4; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r542; shf.l.wrap.b32 %r550, %r548, %r548, 14; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r545; shf.l.wrap.b32 %r553, %r551, %r551, 24; sub.s32 %r554, %r552, %r553; and.b32 %r555, %r554, 31; setp.lt.u32 %p66, %r555, 24; selp.f32 %f308, %f215, %f177, %p66; setp.lt.u32 %p67, %r555, 16; selp.f32 %f309, %f177, %f180, %p67; setp.lt.u32 %p68, %r555, 8; selp.f32 %f310, %f180, %f296, %p68; and.b32 %r556, %r554, 1; setp.eq.b32 %p69, %r556, 1; neg.ftz.f32 %f311, %f308; selp.f32 %f312, %f311, %f308, %p69; and.b32 %r557, %r554, 2; setp.eq.s32 %p70, %r557, 0; neg.ftz.f32 %f313, %f309; selp.f32 %f314, %f309, %f313, %p70; add.ftz.f32 %f315, %f312, %f314; and.b32 %r558, %r554, 4; setp.eq.s32 %p71, %r558, 0; neg.ftz.f32 %f316, %f310; selp.f32 %f317, %f310, %f316, %p71; add.ftz.f32 %f318, %f317, %f315; add.s32 %r559, %r248, %r514; xor.b32 %r560, %r256, %r559; sub.s32 %r561, %r560, %r258; xor.b32 %r562, %r561, %r251; shf.l.wrap.b32 %r563, %r561, %r561, 25; sub.s32 %r564, %r562, %r563; xor.b32 %r565, %r564, %r256; shf.l.wrap.b32 %r566, %r564, %r564, 16; sub.s32 %r567, %r565, %r566; xor.b32 %r568, %r567, %r561; shf.l.wrap.b32 %r569, %r567, %r567, 4; sub.s32 %r570, %r568, %r569; xor.b32 %r571, %r570, %r564; shf.l.wrap.b32 %r572, %r570, %r570, 14; sub.s32 %r573, %r571, %r572; xor.b32 %r574, %r573, %r567; shf.l.wrap.b32 %r575, %r573, %r573, 24; sub.s32 %r576, %r574, %r575; and.b32 %r577, %r576, 31; setp.lt.u32 %p72, %r577, 24; selp.f32 %f319, %f174, %f228, %p72; setp.lt.u32 %p73, %r577, 16; selp.f32 %f320, %f228, %f180, %p73; setp.lt.u32 %p74, %r577, 8; selp.f32 %f321, %f180, %f296, %p74; and.b32 %r578, %r576, 1; setp.eq.b32 %p75, %r578, 1; neg.ftz.f32 %f322, %f319; selp.f32 %f323, %f322, %f319, %p75; and.b32 %r579, %r576, 2; setp.eq.s32 %p76, %r579, 0; neg.ftz.f32 %f324, %f320; selp.f32 %f325, %f320, %f324, %p76; add.ftz.f32 %f326, %f323, %f325; and.b32 %r580, %r576, 4; setp.eq.s32 %p77, %r580, 0; neg.ftz.f32 %f327, %f321; selp.f32 %f328, %f321, %f327, %p77; add.ftz.f32 %f329, %f328, %f326; add.s32 %r581, %r293, %r514; xor.b32 %r582, %r301, %r581; sub.s32 %r583, %r582, %r303; xor.b32 %r584, %r583, %r296; shf.l.wrap.b32 %r585, %r583, %r583, 25; sub.s32 %r586, %r584, %r585; xor.b32 %r587, %r586, %r301; shf.l.wrap.b32 %r588, %r586, %r586, 16; sub.s32 %r589, %r587, %r588; xor.b32 %r590, %r589, %r583; shf.l.wrap.b32 %r591, %r589, %r589, 4; sub.s32 %r592, %r590, %r591; xor.b32 %r593, %r592, %r586; shf.l.wrap.b32 %r594, %r592, %r592, 14; sub.s32 %r595, %r593, %r594; xor.b32 %r596, %r595, %r589; shf.l.wrap.b32 %r597, %r595, %r595, 24; sub.s32 %r598, %r596, %r597; and.b32 %r599, %r598, 31; setp.lt.u32 %p78, %r599, 24; selp.f32 %f330, %f215, %f228, %p78; setp.lt.u32 %p79, %r599, 16; selp.f32 %f331, %f228, %f180, %p79; setp.lt.u32 %p80, %r599, 8; selp.f32 %f332, %f180, %f296, %p80; and.b32 %r600, %r598, 1; setp.eq.b32 %p81, %r600, 1; neg.ftz.f32 %f333, %f330; selp.f32 %f334, %f333, %f330, %p81; and.b32 %r601, %r598, 2; setp.eq.s32 %p82, %r601, 0; neg.ftz.f32 %f335, %f331; selp.f32 %f336, %f331, %f335, %p82; add.ftz.f32 %f337, %f334, %f336; and.b32 %r602, %r598, 4; setp.eq.s32 %p83, %r602, 0; neg.ftz.f32 %f338, %f332; selp.f32 %f339, %f332, %f338, %p83; add.ftz.f32 %f340, %f339, %f337; add.s32 %r603, %r345, %r514; xor.b32 %r604, %r353, %r603; sub.s32 %r605, %r604, %r355; xor.b32 %r606, %r605, %r348; shf.l.wrap.b32 %r607, %r605, %r605, 25; sub.s32 %r608, %r606, %r607; xor.b32 %r609, %r608, %r353; shf.l.wrap.b32 %r610, %r608, %r608, 16; sub.s32 %r611, %r609, %r610; xor.b32 %r612, %r611, %r605; shf.l.wrap.b32 %r613, %r611, %r611, 4; sub.s32 %r614, %r612, %r613; xor.b32 %r615, %r614, %r608; shf.l.wrap.b32 %r616, %r614, %r614, 14; sub.s32 %r617, %r615, %r616; xor.b32 %r618, %r617, %r611; shf.l.wrap.b32 %r619, %r617, %r617, 24; sub.s32 %r620, %r618, %r619; and.b32 %r621, %r620, 31; setp.lt.u32 %p84, %r621, 24; selp.f32 %f341, %f174, %f177, %p84; setp.lt.u32 %p85, %r621, 16; selp.f32 %f342, %f177, %f251, %p85; setp.lt.u32 %p86, %r621, 8; selp.f32 %f343, %f251, %f296, %p86; and.b32 %r622, %r620, 1; setp.eq.b32 %p87, %r622, 1; neg.ftz.f32 %f344, %f341; selp.f32 %f345, %f344, %f341, %p87; and.b32 %r623, %r620, 2; setp.eq.s32 %p88, %r623, 0; neg.ftz.f32 %f346, %f342; selp.f32 %f347, %f342, %f346, %p88; add.ftz.f32 %f348, %f345, %f347; and.b32 %r624, %r620, 4; setp.eq.s32 %p89, %r624, 0; neg.ftz.f32 %f349, %f343; selp.f32 %f350, %f343, %f349, %p89; add.ftz.f32 %f351, %f350, %f348; add.s32 %r625, %r392, %r514; xor.b32 %r626, %r400, %r625; sub.s32 %r627, %r626, %r402; xor.b32 %r628, %r627, %r395; shf.l.wrap.b32 %r629, %r627, %r627, 25; sub.s32 %r630, %r628, %r629; xor.b32 %r631, %r630, %r400; shf.l.wrap.b32 %r632, %r630, %r630, 16; sub.s32 %r633, %r631, %r632; xor.b32 %r634, %r633, %r627; shf.l.wrap.b32 %r635, %r633, %r633, 4; sub.s32 %r636, %r634, %r635; xor.b32 %r637, %r636, %r630; shf.l.wrap.b32 %r638, %r636, %r636, 14; sub.s32 %r639, %r637, %r638; xor.b32 %r640, %r639, %r633; shf.l.wrap.b32 %r641, %r639, %r639, 24; sub.s32 %r642, %r640, %r641; and.b32 %r643, %r642, 31; setp.lt.u32 %p90, %r643, 24; selp.f32 %f352, %f215, %f177, %p90; setp.lt.u32 %p91, %r643, 16; selp.f32 %f353, %f177, %f251, %p91; setp.lt.u32 %p92, %r643, 8; selp.f32 %f354, %f251, %f296, %p92; and.b32 %r644, %r642, 1; setp.eq.b32 %p93, %r644, 1; neg.ftz.f32 %f355, %f352; selp.f32 %f356, %f355, %f352, %p93; and.b32 %r645, %r642, 2; setp.eq.s32 %p94, %r645, 0; neg.ftz.f32 %f357, %f353; selp.f32 %f358, %f353, %f357, %p94; add.ftz.f32 %f359, %f356, %f358; and.b32 %r646, %r642, 4; setp.eq.s32 %p95, %r646, 0; neg.ftz.f32 %f360, %f354; selp.f32 %f361, %f354, %f360, %p95; add.ftz.f32 %f362, %f361, %f359; add.s32 %r647, %r438, %r514; xor.b32 %r648, %r446, %r647; sub.s32 %r649, %r648, %r448; xor.b32 %r650, %r649, %r441; shf.l.wrap.b32 %r651, %r649, %r649, 25; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r446; shf.l.wrap.b32 %r654, %r652, %r652, 16; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 4; sub.s32 %r658, %r656, %r657; xor.b32 %r659, %r658, %r652; shf.l.wrap.b32 %r660, %r658, %r658, 14; sub.s32 %r661, %r659, %r660; xor.b32 %r662, %r661, %r655; shf.l.wrap.b32 %r663, %r661, %r661, 24; sub.s32 %r664, %r662, %r663; and.b32 %r665, %r664, 31; setp.lt.u32 %p96, %r665, 24; selp.f32 %f363, %f174, %f228, %p96; setp.lt.u32 %p97, %r665, 16; selp.f32 %f364, %f228, %f251, %p97; setp.lt.u32 %p98, %r665, 8; selp.f32 %f365, %f251, %f296, %p98; and.b32 %r666, %r664, 1; setp.eq.b32 %p99, %r666, 1; neg.ftz.f32 %f366, %f363; selp.f32 %f367, %f366, %f363, %p99; and.b32 %r667, %r664, 2; setp.eq.s32 %p100, %r667, 0; neg.ftz.f32 %f368, %f364; selp.f32 %f369, %f364, %f368, %p100; add.ftz.f32 %f370, %f367, %f369; and.b32 %r668, %r664, 4; setp.eq.s32 %p101, %r668, 0; neg.ftz.f32 %f371, %f365; selp.f32 %f372, %f365, %f371, %p101; add.ftz.f32 %f373, %f372, %f370; add.s32 %r669, %r483, %r514; xor.b32 %r670, %r491, %r669; sub.s32 %r671, %r670, %r493; xor.b32 %r672, %r671, %r486; shf.l.wrap.b32 %r673, %r671, %r671, 25; sub.s32 %r674, %r672, %r673; xor.b32 %r675, %r674, %r491; shf.l.wrap.b32 %r676, %r674, %r674, 16; sub.s32 %r677, %r675, %r676; xor.b32 %r678, %r677, %r671; shf.l.wrap.b32 %r679, %r677, %r677, 4; sub.s32 %r680, %r678, %r679; xor.b32 %r681, %r680, %r674; shf.l.wrap.b32 %r682, %r680, %r680, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r677; shf.l.wrap.b32 %r685, %r683, %r683, 24; sub.s32 %r686, %r684, %r685; and.b32 %r687, %r686, 31; setp.lt.u32 %p102, %r687, 24; selp.f32 %f374, %f215, %f228, %p102; setp.lt.u32 %p103, %r687, 16; selp.f32 %f375, %f228, %f251, %p103; setp.lt.u32 %p104, %r687, 8; selp.f32 %f376, %f251, %f296, %p104; and.b32 %r688, %r686, 1; setp.eq.b32 %p105, %r688, 1; neg.ftz.f32 %f377, %f374; selp.f32 %f378, %f377, %f374, %p105; and.b32 %r689, %r686, 2; setp.eq.s32 %p106, %r689, 0; neg.ftz.f32 %f379, %f375; selp.f32 %f380, %f375, %f379, %p106; add.ftz.f32 %f381, %f378, %f380; and.b32 %r690, %r686, 4; setp.eq.s32 %p107, %r690, 0; neg.ftz.f32 %f382, %f376; selp.f32 %f383, %f376, %f382, %p107; add.ftz.f32 %f384, %f383, %f381; sub.ftz.f32 %f385, %f216, %f188; sub.ftz.f32 %f386, %f216, %f193; sub.ftz.f32 %f387, %f216, %f198; mul.ftz.f32 %f388, %f385, %f214; fma.rn.ftz.f32 %f389, %f188, %f227, %f388; mul.ftz.f32 %f390, %f385, %f239; fma.rn.ftz.f32 %f391, %f188, %f250, %f390; mul.ftz.f32 %f392, %f193, %f391; fma.rn.ftz.f32 %f393, %f386, %f389, %f392; mul.ftz.f32 %f394, %f385, %f262; fma.rn.ftz.f32 %f395, %f188, %f273, %f394; mul.ftz.f32 %f396, %f385, %f284; fma.rn.ftz.f32 %f397, %f188, %f295, %f396; mul.ftz.f32 %f398, %f193, %f397; fma.rn.ftz.f32 %f399, %f386, %f395, %f398; mul.ftz.f32 %f400, %f198, %f399; fma.rn.ftz.f32 %f401, %f387, %f393, %f400; mul.ftz.f32 %f402, %f385, %f307; fma.rn.ftz.f32 %f403, %f188, %f318, %f402; mul.ftz.f32 %f404, %f385, %f329; fma.rn.ftz.f32 %f405, %f188, %f340, %f404; mul.ftz.f32 %f406, %f193, %f405; fma.rn.ftz.f32 %f407, %f386, %f403, %f406; mul.ftz.f32 %f408, %f385, %f351; fma.rn.ftz.f32 %f409, %f188, %f362, %f408; mul.ftz.f32 %f410, %f385, %f373; fma.rn.ftz.f32 %f411, %f188, %f384, %f410; mul.ftz.f32 %f412, %f193, %f411; fma.rn.ftz.f32 %f413, %f386, %f409, %f412; mul.ftz.f32 %f414, %f198, %f413; fma.rn.ftz.f32 %f415, %f387, %f407, %f414; sub.ftz.f32 %f416, %f415, %f401; fma.rn.ftz.f32 %f23, %f203, %f416, %f401; mov.b32 %r12, %f23; setp.nan.ftz.f32 %p108, %f23, %f23; mov.f32 %f1600, 0f00000000; mov.f32 %f1597, %f1600; @%p108 bra $L__BB79_33; or.b32 %r691, %r12, -2147483648; setp.ne.s32 %p109, %r691, -2147483648; add.ftz.f32 %f418, %f23, %f23; setp.eq.ftz.f32 %p110, %f418, %f23; and.pred %p111, %p110, %p109; mov.f32 %f1597, %f1600; @%p111 bra $L__BB79_33; shl.b32 %r692, %r12, 1; setp.lt.u32 %p112, %r692, -16777215; mul.ftz.f32 %f419, %f23, 0f3F559B3D; selp.f32 %f1597, %f419, 0f00000000, %p112; $L__BB79_33: add.ftz.f32 %f421, %f1601, 0f4347D70A; cvt.rmi.ftz.f32.f32 %f422, %f421; cvt.rzi.ftz.s32.f32 %r693, %f422; sub.ftz.f32 %f423, %f421, %f422; add.ftz.f32 %f424, %f1602, 0f43224AE8; cvt.rmi.ftz.f32.f32 %f425, %f424; cvt.rzi.ftz.s32.f32 %r694, %f425; sub.ftz.f32 %f426, %f424, %f425; add.ftz.f32 %f427, %f1603, 0f431A0C59; cvt.rmi.ftz.f32.f32 %f428, %f427; cvt.rzi.ftz.s32.f32 %r695, %f428; sub.ftz.f32 %f429, %f427, %f428; add.ftz.f32 %f430, %f1604, 0f430E4482; cvt.rmi.ftz.f32.f32 %f431, %f430; cvt.rzi.ftz.s32.f32 %r696, %f431; sub.ftz.f32 %f432, %f430, %f431; mul.ftz.f32 %f433, %f423, %f423; mul.ftz.f32 %f434, %f423, %f433; fma.rn.ftz.f32 %f435, %f423, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f436, %f423, %f435, 0f41200000; mul.ftz.f32 %f437, %f434, %f436; mul.ftz.f32 %f438, %f426, %f426; mul.ftz.f32 %f439, %f426, %f438; fma.rn.ftz.f32 %f440, %f426, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f441, %f426, %f440, 0f41200000; mul.ftz.f32 %f442, %f439, %f441; mul.ftz.f32 %f443, %f429, %f429; mul.ftz.f32 %f444, %f429, %f443; fma.rn.ftz.f32 %f445, %f429, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f446, %f429, %f445, 0f41200000; mul.ftz.f32 %f447, %f444, %f446; mul.ftz.f32 %f448, %f432, %f432; mul.ftz.f32 %f449, %f432, %f448; fma.rn.ftz.f32 %f450, %f432, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f451, %f432, %f450, 0f41200000; mul.ftz.f32 %f452, %f449, %f451; add.s32 %r697, %r694, -559038708; add.s32 %r698, %r695, -559038708; sub.s32 %r699, %r693, %r695; shf.l.wrap.b32 %r700, %r698, %r698, 4; xor.b32 %r701, %r700, %r699; add.s32 %r702, %r698, %r697; sub.s32 %r703, %r697, %r701; shf.l.wrap.b32 %r704, %r701, %r701, 6; xor.b32 %r705, %r704, %r703; add.s32 %r706, %r701, %r702; sub.s32 %r707, %r702, %r705; shf.l.wrap.b32 %r708, %r705, %r705, 8; xor.b32 %r709, %r708, %r707; add.s32 %r710, %r705, %r706; sub.s32 %r711, %r706, %r709; shf.l.wrap.b32 %r712, %r709, %r709, 16; xor.b32 %r713, %r712, %r711; add.s32 %r714, %r709, %r710; sub.s32 %r715, %r710, %r713; shf.l.wrap.b32 %r716, %r713, %r713, 19; xor.b32 %r717, %r716, %r715; add.s32 %r718, %r713, %r714; sub.s32 %r719, %r714, %r717; shf.l.wrap.b32 %r720, %r717, %r717, 4; add.s32 %r721, %r717, %r718; add.s32 %r722, %r718, %r696; xor.b32 %r723, %r721, %r719; xor.b32 %r724, %r723, %r720; shf.l.wrap.b32 %r725, %r721, %r721, 14; sub.s32 %r726, %r724, %r725; xor.b32 %r727, %r726, %r722; shf.l.wrap.b32 %r728, %r726, %r726, 11; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r721; shf.l.wrap.b32 %r731, %r729, %r729, 25; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r726; shf.l.wrap.b32 %r734, %r732, %r732, 16; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 4; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 14; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 24; sub.s32 %r744, %r742, %r743; and.b32 %r745, %r744, 31; setp.lt.u32 %p113, %r745, 24; selp.f32 %f453, %f423, %f426, %p113; setp.lt.u32 %p114, %r745, 16; selp.f32 %f454, %f426, %f429, %p114; setp.lt.u32 %p115, %r745, 8; selp.f32 %f455, %f429, %f432, %p115; and.b32 %r746, %r744, 1; setp.eq.b32 %p116, %r746, 1; neg.ftz.f32 %f456, %f453; selp.f32 %f457, %f456, %f453, %p116; and.b32 %r747, %r744, 2; setp.eq.s32 %p117, %r747, 0; neg.ftz.f32 %f458, %f454; selp.f32 %f459, %f454, %f458, %p117; add.ftz.f32 %f460, %f457, %f459; and.b32 %r748, %r744, 4; setp.eq.s32 %p118, %r748, 0; neg.ftz.f32 %f461, %f455; selp.f32 %f462, %f455, %f461, %p118; add.ftz.f32 %f463, %f462, %f460; add.s32 %r749, %r693, 1; sub.s32 %r750, %r749, %r695; xor.b32 %r751, %r700, %r750; sub.s32 %r752, %r697, %r751; shf.l.wrap.b32 %r753, %r751, %r751, 6; xor.b32 %r754, %r753, %r752; add.s32 %r755, %r751, %r702; sub.s32 %r756, %r702, %r754; shf.l.wrap.b32 %r757, %r754, %r754, 8; xor.b32 %r758, %r757, %r756; add.s32 %r759, %r754, %r755; sub.s32 %r760, %r755, %r758; shf.l.wrap.b32 %r761, %r758, %r758, 16; xor.b32 %r762, %r761, %r760; add.s32 %r763, %r758, %r759; sub.s32 %r764, %r759, %r762; shf.l.wrap.b32 %r765, %r762, %r762, 19; xor.b32 %r766, %r765, %r764; add.s32 %r767, %r762, %r763; sub.s32 %r768, %r763, %r766; shf.l.wrap.b32 %r769, %r766, %r766, 4; add.s32 %r770, %r766, %r767; add.s32 %r771, %r767, %r696; xor.b32 %r772, %r770, %r768; xor.b32 %r773, %r772, %r769; shf.l.wrap.b32 %r774, %r770, %r770, 14; sub.s32 %r775, %r773, %r774; xor.b32 %r776, %r775, %r771; shf.l.wrap.b32 %r777, %r775, %r775, 11; sub.s32 %r778, %r776, %r777; xor.b32 %r779, %r778, %r770; shf.l.wrap.b32 %r780, %r778, %r778, 25; sub.s32 %r781, %r779, %r780; xor.b32 %r782, %r781, %r775; shf.l.wrap.b32 %r783, %r781, %r781, 16; sub.s32 %r784, %r782, %r783; xor.b32 %r785, %r784, %r778; shf.l.wrap.b32 %r786, %r784, %r784, 4; sub.s32 %r787, %r785, %r786; xor.b32 %r788, %r787, %r781; shf.l.wrap.b32 %r789, %r787, %r787, 14; sub.s32 %r790, %r788, %r789; xor.b32 %r791, %r790, %r784; shf.l.wrap.b32 %r792, %r790, %r790, 24; sub.s32 %r793, %r791, %r792; add.ftz.f32 %f464, %f423, 0fBF800000; and.b32 %r794, %r793, 31; setp.lt.u32 %p119, %r794, 24; selp.f32 %f466, %f464, %f426, %p119; setp.lt.u32 %p120, %r794, 16; selp.f32 %f467, %f426, %f429, %p120; setp.lt.u32 %p121, %r794, 8; selp.f32 %f468, %f429, %f432, %p121; and.b32 %r795, %r793, 1; setp.eq.b32 %p122, %r795, 1; neg.ftz.f32 %f469, %f466; selp.f32 %f470, %f469, %f466, %p122; and.b32 %r796, %r793, 2; setp.eq.s32 %p123, %r796, 0; neg.ftz.f32 %f471, %f467; selp.f32 %f472, %f467, %f471, %p123; add.ftz.f32 %f473, %f470, %f472; and.b32 %r797, %r793, 4; setp.eq.s32 %p124, %r797, 0; neg.ftz.f32 %f474, %f468; selp.f32 %f475, %f468, %f474, %p124; add.ftz.f32 %f476, %f475, %f473; add.s32 %r798, %r694, -559038707; add.s32 %r799, %r698, %r798; sub.s32 %r800, %r798, %r701; xor.b32 %r801, %r704, %r800; add.s32 %r802, %r701, %r799; sub.s32 %r803, %r799, %r801; shf.l.wrap.b32 %r804, %r801, %r801, 8; xor.b32 %r805, %r804, %r803; add.s32 %r806, %r801, %r802; sub.s32 %r807, %r802, %r805; shf.l.wrap.b32 %r808, %r805, %r805, 16; xor.b32 %r809, %r808, %r807; add.s32 %r810, %r805, %r806; sub.s32 %r811, %r806, %r809; shf.l.wrap.b32 %r812, %r809, %r809, 19; xor.b32 %r813, %r812, %r811; add.s32 %r814, %r809, %r810; sub.s32 %r815, %r810, %r813; shf.l.wrap.b32 %r816, %r813, %r813, 4; add.s32 %r817, %r813, %r814; add.s32 %r818, %r814, %r696; xor.b32 %r819, %r817, %r815; xor.b32 %r820, %r819, %r816; shf.l.wrap.b32 %r821, %r817, %r817, 14; sub.s32 %r822, %r820, %r821; xor.b32 %r823, %r822, %r818; shf.l.wrap.b32 %r824, %r822, %r822, 11; sub.s32 %r825, %r823, %r824; xor.b32 %r826, %r825, %r817; shf.l.wrap.b32 %r827, %r825, %r825, 25; sub.s32 %r828, %r826, %r827; xor.b32 %r829, %r828, %r822; shf.l.wrap.b32 %r830, %r828, %r828, 16; sub.s32 %r831, %r829, %r830; xor.b32 %r832, %r831, %r825; shf.l.wrap.b32 %r833, %r831, %r831, 4; sub.s32 %r834, %r832, %r833; xor.b32 %r835, %r834, %r828; shf.l.wrap.b32 %r836, %r834, %r834, 14; sub.s32 %r837, %r835, %r836; xor.b32 %r838, %r837, %r831; shf.l.wrap.b32 %r839, %r837, %r837, 24; sub.s32 %r840, %r838, %r839; add.ftz.f32 %f477, %f426, 0fBF800000; and.b32 %r841, %r840, 31; setp.lt.u32 %p125, %r841, 24; selp.f32 %f478, %f423, %f477, %p125; setp.lt.u32 %p126, %r841, 16; selp.f32 %f479, %f477, %f429, %p126; setp.lt.u32 %p127, %r841, 8; selp.f32 %f480, %f429, %f432, %p127; and.b32 %r842, %r840, 1; setp.eq.b32 %p128, %r842, 1; neg.ftz.f32 %f481, %f478; selp.f32 %f482, %f481, %f478, %p128; and.b32 %r843, %r840, 2; setp.eq.s32 %p129, %r843, 0; neg.ftz.f32 %f483, %f479; selp.f32 %f484, %f479, %f483, %p129; add.ftz.f32 %f485, %f482, %f484; and.b32 %r844, %r840, 4; setp.eq.s32 %p130, %r844, 0; neg.ftz.f32 %f486, %f480; selp.f32 %f487, %f480, %f486, %p130; add.ftz.f32 %f488, %f487, %f485; sub.s32 %r845, %r798, %r751; xor.b32 %r846, %r753, %r845; add.s32 %r847, %r751, %r799; sub.s32 %r848, %r799, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 8; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 16; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 19; xor.b32 %r858, %r857, %r856; add.s32 %r859, %r854, %r855; sub.s32 %r860, %r855, %r858; shf.l.wrap.b32 %r861, %r858, %r858, 4; add.s32 %r862, %r858, %r859; add.s32 %r863, %r859, %r696; xor.b32 %r864, %r862, %r860; xor.b32 %r865, %r864, %r861; shf.l.wrap.b32 %r866, %r862, %r862, 14; sub.s32 %r867, %r865, %r866; xor.b32 %r868, %r867, %r863; shf.l.wrap.b32 %r869, %r867, %r867, 11; sub.s32 %r870, %r868, %r869; xor.b32 %r871, %r870, %r862; shf.l.wrap.b32 %r872, %r870, %r870, 25; sub.s32 %r873, %r871, %r872; xor.b32 %r874, %r873, %r867; shf.l.wrap.b32 %r875, %r873, %r873, 16; sub.s32 %r876, %r874, %r875; xor.b32 %r877, %r876, %r870; shf.l.wrap.b32 %r878, %r876, %r876, 4; sub.s32 %r879, %r877, %r878; xor.b32 %r880, %r879, %r873; shf.l.wrap.b32 %r881, %r879, %r879, 14; sub.s32 %r882, %r880, %r881; xor.b32 %r883, %r882, %r876; shf.l.wrap.b32 %r884, %r882, %r882, 24; sub.s32 %r885, %r883, %r884; and.b32 %r886, %r885, 31; setp.lt.u32 %p131, %r886, 24; selp.f32 %f489, %f464, %f477, %p131; setp.lt.u32 %p132, %r886, 16; selp.f32 %f490, %f477, %f429, %p132; setp.lt.u32 %p133, %r886, 8; selp.f32 %f491, %f429, %f432, %p133; and.b32 %r887, %r885, 1; setp.eq.b32 %p134, %r887, 1; neg.ftz.f32 %f492, %f489; selp.f32 %f493, %f492, %f489, %p134; and.b32 %r888, %r885, 2; setp.eq.s32 %p135, %r888, 0; neg.ftz.f32 %f494, %f490; selp.f32 %f495, %f490, %f494, %p135; add.ftz.f32 %f496, %f493, %f495; and.b32 %r889, %r885, 4; setp.eq.s32 %p136, %r889, 0; neg.ftz.f32 %f497, %f491; selp.f32 %f498, %f491, %f497, %p136; add.ftz.f32 %f499, %f498, %f496; not.b32 %r890, %r695; add.s32 %r891, %r695, -559038707; add.s32 %r892, %r693, %r890; shf.l.wrap.b32 %r893, %r891, %r891, 4; xor.b32 %r894, %r893, %r892; add.s32 %r895, %r891, %r697; sub.s32 %r896, %r697, %r894; shf.l.wrap.b32 %r897, %r894, %r894, 6; xor.b32 %r898, %r897, %r896; add.s32 %r899, %r894, %r895; sub.s32 %r900, %r895, %r898; shf.l.wrap.b32 %r901, %r898, %r898, 8; xor.b32 %r902, %r901, %r900; add.s32 %r903, %r898, %r899; sub.s32 %r904, %r899, %r902; shf.l.wrap.b32 %r905, %r902, %r902, 16; xor.b32 %r906, %r905, %r904; add.s32 %r907, %r902, %r903; sub.s32 %r908, %r903, %r906; shf.l.wrap.b32 %r909, %r906, %r906, 19; xor.b32 %r910, %r909, %r908; add.s32 %r911, %r906, %r907; sub.s32 %r912, %r907, %r910; shf.l.wrap.b32 %r913, %r910, %r910, 4; add.s32 %r914, %r910, %r911; add.s32 %r915, %r911, %r696; xor.b32 %r916, %r914, %r912; xor.b32 %r917, %r916, %r913; shf.l.wrap.b32 %r918, %r914, %r914, 14; sub.s32 %r919, %r917, %r918; xor.b32 %r920, %r919, %r915; shf.l.wrap.b32 %r921, %r919, %r919, 11; sub.s32 %r922, %r920, %r921; xor.b32 %r923, %r922, %r914; shf.l.wrap.b32 %r924, %r922, %r922, 25; sub.s32 %r925, %r923, %r924; xor.b32 %r926, %r925, %r919; shf.l.wrap.b32 %r927, %r925, %r925, 16; sub.s32 %r928, %r926, %r927; xor.b32 %r929, %r928, %r922; shf.l.wrap.b32 %r930, %r928, %r928, 4; sub.s32 %r931, %r929, %r930; xor.b32 %r932, %r931, %r925; shf.l.wrap.b32 %r933, %r931, %r931, 14; sub.s32 %r934, %r932, %r933; xor.b32 %r935, %r934, %r928; shf.l.wrap.b32 %r936, %r934, %r934, 24; sub.s32 %r937, %r935, %r936; add.ftz.f32 %f500, %f429, 0fBF800000; and.b32 %r938, %r937, 31; setp.lt.u32 %p137, %r938, 24; selp.f32 %f501, %f423, %f426, %p137; setp.lt.u32 %p138, %r938, 16; selp.f32 %f502, %f426, %f500, %p138; setp.lt.u32 %p139, %r938, 8; selp.f32 %f503, %f500, %f432, %p139; and.b32 %r939, %r937, 1; setp.eq.b32 %p140, %r939, 1; neg.ftz.f32 %f504, %f501; selp.f32 %f505, %f504, %f501, %p140; and.b32 %r940, %r937, 2; setp.eq.s32 %p141, %r940, 0; neg.ftz.f32 %f506, %f502; selp.f32 %f507, %f502, %f506, %p141; add.ftz.f32 %f508, %f505, %f507; and.b32 %r941, %r937, 4; setp.eq.s32 %p142, %r941, 0; neg.ftz.f32 %f509, %f503; selp.f32 %f510, %f503, %f509, %p142; add.ftz.f32 %f511, %f510, %f508; xor.b32 %r942, %r893, %r699; sub.s32 %r943, %r697, %r942; shf.l.wrap.b32 %r944, %r942, %r942, 6; xor.b32 %r945, %r944, %r943; add.s32 %r946, %r942, %r895; sub.s32 %r947, %r895, %r945; shf.l.wrap.b32 %r948, %r945, %r945, 8; xor.b32 %r949, %r948, %r947; add.s32 %r950, %r945, %r946; sub.s32 %r951, %r946, %r949; shf.l.wrap.b32 %r952, %r949, %r949, 16; xor.b32 %r953, %r952, %r951; add.s32 %r954, %r949, %r950; sub.s32 %r955, %r950, %r953; shf.l.wrap.b32 %r956, %r953, %r953, 19; xor.b32 %r957, %r956, %r955; add.s32 %r958, %r953, %r954; sub.s32 %r959, %r954, %r957; shf.l.wrap.b32 %r960, %r957, %r957, 4; add.s32 %r961, %r957, %r958; add.s32 %r962, %r958, %r696; xor.b32 %r963, %r961, %r959; xor.b32 %r964, %r963, %r960; shf.l.wrap.b32 %r965, %r961, %r961, 14; sub.s32 %r966, %r964, %r965; xor.b32 %r967, %r966, %r962; shf.l.wrap.b32 %r968, %r966, %r966, 11; sub.s32 %r969, %r967, %r968; xor.b32 %r970, %r969, %r961; shf.l.wrap.b32 %r971, %r969, %r969, 25; sub.s32 %r972, %r970, %r971; xor.b32 %r973, %r972, %r966; shf.l.wrap.b32 %r974, %r972, %r972, 16; sub.s32 %r975, %r973, %r974; xor.b32 %r976, %r975, %r969; shf.l.wrap.b32 %r977, %r975, %r975, 4; sub.s32 %r978, %r976, %r977; xor.b32 %r979, %r978, %r972; shf.l.wrap.b32 %r980, %r978, %r978, 14; sub.s32 %r981, %r979, %r980; xor.b32 %r982, %r981, %r975; shf.l.wrap.b32 %r983, %r981, %r981, 24; sub.s32 %r984, %r982, %r983; and.b32 %r985, %r984, 31; setp.lt.u32 %p143, %r985, 24; selp.f32 %f512, %f464, %f426, %p143; setp.lt.u32 %p144, %r985, 16; selp.f32 %f513, %f426, %f500, %p144; setp.lt.u32 %p145, %r985, 8; selp.f32 %f514, %f500, %f432, %p145; and.b32 %r986, %r984, 1; setp.eq.b32 %p146, %r986, 1; neg.ftz.f32 %f515, %f512; selp.f32 %f516, %f515, %f512, %p146; and.b32 %r987, %r984, 2; setp.eq.s32 %p147, %r987, 0; neg.ftz.f32 %f517, %f513; selp.f32 %f518, %f513, %f517, %p147; add.ftz.f32 %f519, %f516, %f518; and.b32 %r988, %r984, 4; setp.eq.s32 %p148, %r988, 0; neg.ftz.f32 %f520, %f514; selp.f32 %f521, %f514, %f520, %p148; add.ftz.f32 %f522, %f521, %f519; add.s32 %r989, %r891, %r798; sub.s32 %r990, %r798, %r894; xor.b32 %r991, %r897, %r990; add.s32 %r992, %r894, %r989; sub.s32 %r993, %r989, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 8; xor.b32 %r995, %r994, %r993; add.s32 %r996, %r991, %r992; sub.s32 %r997, %r992, %r995; shf.l.wrap.b32 %r998, %r995, %r995, 16; xor.b32 %r999, %r998, %r997; add.s32 %r1000, %r995, %r996; sub.s32 %r1001, %r996, %r999; shf.l.wrap.b32 %r1002, %r999, %r999, 19; xor.b32 %r1003, %r1002, %r1001; add.s32 %r1004, %r999, %r1000; sub.s32 %r1005, %r1000, %r1003; shf.l.wrap.b32 %r1006, %r1003, %r1003, 4; add.s32 %r1007, %r1003, %r1004; add.s32 %r1008, %r1004, %r696; xor.b32 %r1009, %r1007, %r1005; xor.b32 %r1010, %r1009, %r1006; shf.l.wrap.b32 %r1011, %r1007, %r1007, 14; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1008; shf.l.wrap.b32 %r1014, %r1012, %r1012, 11; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1007; shf.l.wrap.b32 %r1017, %r1015, %r1015, 25; sub.s32 %r1018, %r1016, %r1017; xor.b32 %r1019, %r1018, %r1012; shf.l.wrap.b32 %r1020, %r1018, %r1018, 16; sub.s32 %r1021, %r1019, %r1020; xor.b32 %r1022, %r1021, %r1015; shf.l.wrap.b32 %r1023, %r1021, %r1021, 4; sub.s32 %r1024, %r1022, %r1023; xor.b32 %r1025, %r1024, %r1018; shf.l.wrap.b32 %r1026, %r1024, %r1024, 14; sub.s32 %r1027, %r1025, %r1026; xor.b32 %r1028, %r1027, %r1021; shf.l.wrap.b32 %r1029, %r1027, %r1027, 24; sub.s32 %r1030, %r1028, %r1029; and.b32 %r1031, %r1030, 31; setp.lt.u32 %p149, %r1031, 24; selp.f32 %f523, %f423, %f477, %p149; setp.lt.u32 %p150, %r1031, 16; selp.f32 %f524, %f477, %f500, %p150; setp.lt.u32 %p151, %r1031, 8; selp.f32 %f525, %f500, %f432, %p151; and.b32 %r1032, %r1030, 1; setp.eq.b32 %p152, %r1032, 1; neg.ftz.f32 %f526, %f523; selp.f32 %f527, %f526, %f523, %p152; and.b32 %r1033, %r1030, 2; setp.eq.s32 %p153, %r1033, 0; neg.ftz.f32 %f528, %f524; selp.f32 %f529, %f524, %f528, %p153; add.ftz.f32 %f530, %f527, %f529; and.b32 %r1034, %r1030, 4; setp.eq.s32 %p154, %r1034, 0; neg.ftz.f32 %f531, %f525; selp.f32 %f532, %f525, %f531, %p154; add.ftz.f32 %f533, %f532, %f530; sub.s32 %r1035, %r798, %r942; xor.b32 %r1036, %r944, %r1035; add.s32 %r1037, %r942, %r989; sub.s32 %r1038, %r989, %r1036; shf.l.wrap.b32 %r1039, %r1036, %r1036, 8; xor.b32 %r1040, %r1039, %r1038; add.s32 %r1041, %r1036, %r1037; sub.s32 %r1042, %r1037, %r1040; shf.l.wrap.b32 %r1043, %r1040, %r1040, 16; xor.b32 %r1044, %r1043, %r1042; add.s32 %r1045, %r1040, %r1041; sub.s32 %r1046, %r1041, %r1044; shf.l.wrap.b32 %r1047, %r1044, %r1044, 19; xor.b32 %r1048, %r1047, %r1046; add.s32 %r1049, %r1044, %r1045; sub.s32 %r1050, %r1045, %r1048; shf.l.wrap.b32 %r1051, %r1048, %r1048, 4; add.s32 %r1052, %r1048, %r1049; add.s32 %r1053, %r1049, %r696; xor.b32 %r1054, %r1052, %r1050; xor.b32 %r1055, %r1054, %r1051; shf.l.wrap.b32 %r1056, %r1052, %r1052, 14; sub.s32 %r1057, %r1055, %r1056; xor.b32 %r1058, %r1057, %r1053; shf.l.wrap.b32 %r1059, %r1057, %r1057, 11; sub.s32 %r1060, %r1058, %r1059; xor.b32 %r1061, %r1060, %r1052; shf.l.wrap.b32 %r1062, %r1060, %r1060, 25; sub.s32 %r1063, %r1061, %r1062; xor.b32 %r1064, %r1063, %r1057; shf.l.wrap.b32 %r1065, %r1063, %r1063, 16; sub.s32 %r1066, %r1064, %r1065; xor.b32 %r1067, %r1066, %r1060; shf.l.wrap.b32 %r1068, %r1066, %r1066, 4; sub.s32 %r1069, %r1067, %r1068; xor.b32 %r1070, %r1069, %r1063; shf.l.wrap.b32 %r1071, %r1069, %r1069, 14; sub.s32 %r1072, %r1070, %r1071; xor.b32 %r1073, %r1072, %r1066; shf.l.wrap.b32 %r1074, %r1072, %r1072, 24; sub.s32 %r1075, %r1073, %r1074; and.b32 %r1076, %r1075, 31; setp.lt.u32 %p155, %r1076, 24; selp.f32 %f534, %f464, %f477, %p155; setp.lt.u32 %p156, %r1076, 16; selp.f32 %f535, %f477, %f500, %p156; setp.lt.u32 %p157, %r1076, 8; selp.f32 %f536, %f500, %f432, %p157; and.b32 %r1077, %r1075, 1; setp.eq.b32 %p158, %r1077, 1; neg.ftz.f32 %f537, %f534; selp.f32 %f538, %f537, %f534, %p158; and.b32 %r1078, %r1075, 2; setp.eq.s32 %p159, %r1078, 0; neg.ftz.f32 %f539, %f535; selp.f32 %f540, %f535, %f539, %p159; add.ftz.f32 %f541, %f538, %f540; and.b32 %r1079, %r1075, 4; setp.eq.s32 %p160, %r1079, 0; neg.ftz.f32 %f542, %f536; selp.f32 %f543, %f536, %f542, %p160; add.ftz.f32 %f544, %f543, %f541; add.s32 %r1080, %r696, 1; add.s32 %r1081, %r718, %r1080; xor.b32 %r1082, %r726, %r1081; sub.s32 %r1083, %r1082, %r728; xor.b32 %r1084, %r1083, %r721; shf.l.wrap.b32 %r1085, %r1083, %r1083, 25; sub.s32 %r1086, %r1084, %r1085; xor.b32 %r1087, %r1086, %r726; shf.l.wrap.b32 %r1088, %r1086, %r1086, 16; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1083; shf.l.wrap.b32 %r1091, %r1089, %r1089, 4; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1086; shf.l.wrap.b32 %r1094, %r1092, %r1092, 14; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 24; sub.s32 %r1098, %r1096, %r1097; add.ftz.f32 %f545, %f432, 0fBF800000; and.b32 %r1099, %r1098, 31; setp.lt.u32 %p161, %r1099, 24; selp.f32 %f546, %f423, %f426, %p161; setp.lt.u32 %p162, %r1099, 16; selp.f32 %f547, %f426, %f429, %p162; setp.lt.u32 %p163, %r1099, 8; selp.f32 %f548, %f429, %f545, %p163; and.b32 %r1100, %r1098, 1; setp.eq.b32 %p164, %r1100, 1; neg.ftz.f32 %f549, %f546; selp.f32 %f550, %f549, %f546, %p164; and.b32 %r1101, %r1098, 2; setp.eq.s32 %p165, %r1101, 0; neg.ftz.f32 %f551, %f547; selp.f32 %f552, %f547, %f551, %p165; add.ftz.f32 %f553, %f550, %f552; and.b32 %r1102, %r1098, 4; setp.eq.s32 %p166, %r1102, 0; neg.ftz.f32 %f554, %f548; selp.f32 %f555, %f548, %f554, %p166; add.ftz.f32 %f556, %f555, %f553; add.s32 %r1103, %r767, %r1080; xor.b32 %r1104, %r775, %r1103; sub.s32 %r1105, %r1104, %r777; xor.b32 %r1106, %r1105, %r770; shf.l.wrap.b32 %r1107, %r1105, %r1105, 25; sub.s32 %r1108, %r1106, %r1107; xor.b32 %r1109, %r1108, %r775; shf.l.wrap.b32 %r1110, %r1108, %r1108, 16; sub.s32 %r1111, %r1109, %r1110; xor.b32 %r1112, %r1111, %r1105; shf.l.wrap.b32 %r1113, %r1111, %r1111, 4; sub.s32 %r1114, %r1112, %r1113; xor.b32 %r1115, %r1114, %r1108; shf.l.wrap.b32 %r1116, %r1114, %r1114, 14; sub.s32 %r1117, %r1115, %r1116; xor.b32 %r1118, %r1117, %r1111; shf.l.wrap.b32 %r1119, %r1117, %r1117, 24; sub.s32 %r1120, %r1118, %r1119; and.b32 %r1121, %r1120, 31; setp.lt.u32 %p167, %r1121, 24; selp.f32 %f557, %f464, %f426, %p167; setp.lt.u32 %p168, %r1121, 16; selp.f32 %f558, %f426, %f429, %p168; setp.lt.u32 %p169, %r1121, 8; selp.f32 %f559, %f429, %f545, %p169; and.b32 %r1122, %r1120, 1; setp.eq.b32 %p170, %r1122, 1; neg.ftz.f32 %f560, %f557; selp.f32 %f561, %f560, %f557, %p170; and.b32 %r1123, %r1120, 2; setp.eq.s32 %p171, %r1123, 0; neg.ftz.f32 %f562, %f558; selp.f32 %f563, %f558, %f562, %p171; add.ftz.f32 %f564, %f561, %f563; and.b32 %r1124, %r1120, 4; setp.eq.s32 %p172, %r1124, 0; neg.ftz.f32 %f565, %f559; selp.f32 %f566, %f559, %f565, %p172; add.ftz.f32 %f567, %f566, %f564; add.s32 %r1125, %r814, %r1080; xor.b32 %r1126, %r822, %r1125; sub.s32 %r1127, %r1126, %r824; xor.b32 %r1128, %r1127, %r817; shf.l.wrap.b32 %r1129, %r1127, %r1127, 25; sub.s32 %r1130, %r1128, %r1129; xor.b32 %r1131, %r1130, %r822; shf.l.wrap.b32 %r1132, %r1130, %r1130, 16; sub.s32 %r1133, %r1131, %r1132; xor.b32 %r1134, %r1133, %r1127; shf.l.wrap.b32 %r1135, %r1133, %r1133, 4; sub.s32 %r1136, %r1134, %r1135; xor.b32 %r1137, %r1136, %r1130; shf.l.wrap.b32 %r1138, %r1136, %r1136, 14; sub.s32 %r1139, %r1137, %r1138; xor.b32 %r1140, %r1139, %r1133; shf.l.wrap.b32 %r1141, %r1139, %r1139, 24; sub.s32 %r1142, %r1140, %r1141; and.b32 %r1143, %r1142, 31; setp.lt.u32 %p173, %r1143, 24; selp.f32 %f568, %f423, %f477, %p173; setp.lt.u32 %p174, %r1143, 16; selp.f32 %f569, %f477, %f429, %p174; setp.lt.u32 %p175, %r1143, 8; selp.f32 %f570, %f429, %f545, %p175; and.b32 %r1144, %r1142, 1; setp.eq.b32 %p176, %r1144, 1; neg.ftz.f32 %f571, %f568; selp.f32 %f572, %f571, %f568, %p176; and.b32 %r1145, %r1142, 2; setp.eq.s32 %p177, %r1145, 0; neg.ftz.f32 %f573, %f569; selp.f32 %f574, %f569, %f573, %p177; add.ftz.f32 %f575, %f572, %f574; and.b32 %r1146, %r1142, 4; setp.eq.s32 %p178, %r1146, 0; neg.ftz.f32 %f576, %f570; selp.f32 %f577, %f570, %f576, %p178; add.ftz.f32 %f578, %f577, %f575; add.s32 %r1147, %r859, %r1080; xor.b32 %r1148, %r867, %r1147; sub.s32 %r1149, %r1148, %r869; xor.b32 %r1150, %r1149, %r862; shf.l.wrap.b32 %r1151, %r1149, %r1149, 25; sub.s32 %r1152, %r1150, %r1151; xor.b32 %r1153, %r1152, %r867; shf.l.wrap.b32 %r1154, %r1152, %r1152, 16; sub.s32 %r1155, %r1153, %r1154; xor.b32 %r1156, %r1155, %r1149; shf.l.wrap.b32 %r1157, %r1155, %r1155, 4; sub.s32 %r1158, %r1156, %r1157; xor.b32 %r1159, %r1158, %r1152; shf.l.wrap.b32 %r1160, %r1158, %r1158, 14; sub.s32 %r1161, %r1159, %r1160; xor.b32 %r1162, %r1161, %r1155; shf.l.wrap.b32 %r1163, %r1161, %r1161, 24; sub.s32 %r1164, %r1162, %r1163; and.b32 %r1165, %r1164, 31; setp.lt.u32 %p179, %r1165, 24; selp.f32 %f579, %f464, %f477, %p179; setp.lt.u32 %p180, %r1165, 16; selp.f32 %f580, %f477, %f429, %p180; setp.lt.u32 %p181, %r1165, 8; selp.f32 %f581, %f429, %f545, %p181; and.b32 %r1166, %r1164, 1; setp.eq.b32 %p182, %r1166, 1; neg.ftz.f32 %f582, %f579; selp.f32 %f583, %f582, %f579, %p182; and.b32 %r1167, %r1164, 2; setp.eq.s32 %p183, %r1167, 0; neg.ftz.f32 %f584, %f580; selp.f32 %f585, %f580, %f584, %p183; add.ftz.f32 %f586, %f583, %f585; and.b32 %r1168, %r1164, 4; setp.eq.s32 %p184, %r1168, 0; neg.ftz.f32 %f587, %f581; selp.f32 %f588, %f581, %f587, %p184; add.ftz.f32 %f589, %f588, %f586; add.s32 %r1169, %r911, %r1080; xor.b32 %r1170, %r919, %r1169; sub.s32 %r1171, %r1170, %r921; xor.b32 %r1172, %r1171, %r914; shf.l.wrap.b32 %r1173, %r1171, %r1171, 25; sub.s32 %r1174, %r1172, %r1173; xor.b32 %r1175, %r1174, %r919; shf.l.wrap.b32 %r1176, %r1174, %r1174, 16; sub.s32 %r1177, %r1175, %r1176; xor.b32 %r1178, %r1177, %r1171; shf.l.wrap.b32 %r1179, %r1177, %r1177, 4; sub.s32 %r1180, %r1178, %r1179; xor.b32 %r1181, %r1180, %r1174; shf.l.wrap.b32 %r1182, %r1180, %r1180, 14; sub.s32 %r1183, %r1181, %r1182; xor.b32 %r1184, %r1183, %r1177; shf.l.wrap.b32 %r1185, %r1183, %r1183, 24; sub.s32 %r1186, %r1184, %r1185; and.b32 %r1187, %r1186, 31; setp.lt.u32 %p185, %r1187, 24; selp.f32 %f590, %f423, %f426, %p185; setp.lt.u32 %p186, %r1187, 16; selp.f32 %f591, %f426, %f500, %p186; setp.lt.u32 %p187, %r1187, 8; selp.f32 %f592, %f500, %f545, %p187; and.b32 %r1188, %r1186, 1; setp.eq.b32 %p188, %r1188, 1; neg.ftz.f32 %f593, %f590; selp.f32 %f594, %f593, %f590, %p188; and.b32 %r1189, %r1186, 2; setp.eq.s32 %p189, %r1189, 0; neg.ftz.f32 %f595, %f591; selp.f32 %f596, %f591, %f595, %p189; add.ftz.f32 %f597, %f594, %f596; and.b32 %r1190, %r1186, 4; setp.eq.s32 %p190, %r1190, 0; neg.ftz.f32 %f598, %f592; selp.f32 %f599, %f592, %f598, %p190; add.ftz.f32 %f600, %f599, %f597; add.s32 %r1191, %r958, %r1080; xor.b32 %r1192, %r966, %r1191; sub.s32 %r1193, %r1192, %r968; xor.b32 %r1194, %r1193, %r961; shf.l.wrap.b32 %r1195, %r1193, %r1193, 25; sub.s32 %r1196, %r1194, %r1195; xor.b32 %r1197, %r1196, %r966; shf.l.wrap.b32 %r1198, %r1196, %r1196, 16; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1193; shf.l.wrap.b32 %r1201, %r1199, %r1199, 4; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1196; shf.l.wrap.b32 %r1204, %r1202, %r1202, 14; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 24; sub.s32 %r1208, %r1206, %r1207; and.b32 %r1209, %r1208, 31; setp.lt.u32 %p191, %r1209, 24; selp.f32 %f601, %f464, %f426, %p191; setp.lt.u32 %p192, %r1209, 16; selp.f32 %f602, %f426, %f500, %p192; setp.lt.u32 %p193, %r1209, 8; selp.f32 %f603, %f500, %f545, %p193; and.b32 %r1210, %r1208, 1; setp.eq.b32 %p194, %r1210, 1; neg.ftz.f32 %f604, %f601; selp.f32 %f605, %f604, %f601, %p194; and.b32 %r1211, %r1208, 2; setp.eq.s32 %p195, %r1211, 0; neg.ftz.f32 %f606, %f602; selp.f32 %f607, %f602, %f606, %p195; add.ftz.f32 %f608, %f605, %f607; and.b32 %r1212, %r1208, 4; setp.eq.s32 %p196, %r1212, 0; neg.ftz.f32 %f609, %f603; selp.f32 %f610, %f603, %f609, %p196; add.ftz.f32 %f611, %f610, %f608; add.s32 %r1213, %r1004, %r1080; xor.b32 %r1214, %r1012, %r1213; sub.s32 %r1215, %r1214, %r1014; xor.b32 %r1216, %r1215, %r1007; shf.l.wrap.b32 %r1217, %r1215, %r1215, 25; sub.s32 %r1218, %r1216, %r1217; xor.b32 %r1219, %r1218, %r1012; shf.l.wrap.b32 %r1220, %r1218, %r1218, 16; sub.s32 %r1221, %r1219, %r1220; xor.b32 %r1222, %r1221, %r1215; shf.l.wrap.b32 %r1223, %r1221, %r1221, 4; sub.s32 %r1224, %r1222, %r1223; xor.b32 %r1225, %r1224, %r1218; shf.l.wrap.b32 %r1226, %r1224, %r1224, 14; sub.s32 %r1227, %r1225, %r1226; xor.b32 %r1228, %r1227, %r1221; shf.l.wrap.b32 %r1229, %r1227, %r1227, 24; sub.s32 %r1230, %r1228, %r1229; and.b32 %r1231, %r1230, 31; setp.lt.u32 %p197, %r1231, 24; selp.f32 %f612, %f423, %f477, %p197; setp.lt.u32 %p198, %r1231, 16; selp.f32 %f613, %f477, %f500, %p198; setp.lt.u32 %p199, %r1231, 8; selp.f32 %f614, %f500, %f545, %p199; and.b32 %r1232, %r1230, 1; setp.eq.b32 %p200, %r1232, 1; neg.ftz.f32 %f615, %f612; selp.f32 %f616, %f615, %f612, %p200; and.b32 %r1233, %r1230, 2; setp.eq.s32 %p201, %r1233, 0; neg.ftz.f32 %f617, %f613; selp.f32 %f618, %f613, %f617, %p201; add.ftz.f32 %f619, %f616, %f618; and.b32 %r1234, %r1230, 4; setp.eq.s32 %p202, %r1234, 0; neg.ftz.f32 %f620, %f614; selp.f32 %f621, %f614, %f620, %p202; add.ftz.f32 %f622, %f621, %f619; add.s32 %r1235, %r1049, %r1080; xor.b32 %r1236, %r1057, %r1235; sub.s32 %r1237, %r1236, %r1059; xor.b32 %r1238, %r1237, %r1052; shf.l.wrap.b32 %r1239, %r1237, %r1237, 25; sub.s32 %r1240, %r1238, %r1239; xor.b32 %r1241, %r1240, %r1057; shf.l.wrap.b32 %r1242, %r1240, %r1240, 16; sub.s32 %r1243, %r1241, %r1242; xor.b32 %r1244, %r1243, %r1237; shf.l.wrap.b32 %r1245, %r1243, %r1243, 4; sub.s32 %r1246, %r1244, %r1245; xor.b32 %r1247, %r1246, %r1240; shf.l.wrap.b32 %r1248, %r1246, %r1246, 14; sub.s32 %r1249, %r1247, %r1248; xor.b32 %r1250, %r1249, %r1243; shf.l.wrap.b32 %r1251, %r1249, %r1249, 24; sub.s32 %r1252, %r1250, %r1251; and.b32 %r1253, %r1252, 31; setp.lt.u32 %p203, %r1253, 24; selp.f32 %f623, %f464, %f477, %p203; setp.lt.u32 %p204, %r1253, 16; selp.f32 %f624, %f477, %f500, %p204; setp.lt.u32 %p205, %r1253, 8; selp.f32 %f625, %f500, %f545, %p205; and.b32 %r1254, %r1252, 1; setp.eq.b32 %p206, %r1254, 1; neg.ftz.f32 %f626, %f623; selp.f32 %f627, %f626, %f623, %p206; and.b32 %r1255, %r1252, 2; setp.eq.s32 %p207, %r1255, 0; neg.ftz.f32 %f628, %f624; selp.f32 %f629, %f624, %f628, %p207; add.ftz.f32 %f630, %f627, %f629; and.b32 %r1256, %r1252, 4; setp.eq.s32 %p208, %r1256, 0; neg.ftz.f32 %f631, %f625; selp.f32 %f632, %f625, %f631, %p208; add.ftz.f32 %f633, %f632, %f630; sub.ftz.f32 %f634, %f216, %f437; sub.ftz.f32 %f635, %f216, %f442; sub.ftz.f32 %f636, %f216, %f447; mul.ftz.f32 %f637, %f634, %f463; fma.rn.ftz.f32 %f638, %f437, %f476, %f637; mul.ftz.f32 %f639, %f634, %f488; fma.rn.ftz.f32 %f640, %f437, %f499, %f639; mul.ftz.f32 %f641, %f442, %f640; fma.rn.ftz.f32 %f642, %f635, %f638, %f641; mul.ftz.f32 %f643, %f634, %f511; fma.rn.ftz.f32 %f644, %f437, %f522, %f643; mul.ftz.f32 %f645, %f634, %f533; fma.rn.ftz.f32 %f646, %f437, %f544, %f645; mul.ftz.f32 %f647, %f442, %f646; fma.rn.ftz.f32 %f648, %f635, %f644, %f647; mul.ftz.f32 %f649, %f447, %f648; fma.rn.ftz.f32 %f650, %f636, %f642, %f649; mul.ftz.f32 %f651, %f634, %f556; fma.rn.ftz.f32 %f652, %f437, %f567, %f651; mul.ftz.f32 %f653, %f634, %f578; fma.rn.ftz.f32 %f654, %f437, %f589, %f653; mul.ftz.f32 %f655, %f442, %f654; fma.rn.ftz.f32 %f656, %f635, %f652, %f655; mul.ftz.f32 %f657, %f634, %f600; fma.rn.ftz.f32 %f658, %f437, %f611, %f657; mul.ftz.f32 %f659, %f634, %f622; fma.rn.ftz.f32 %f660, %f437, %f633, %f659; mul.ftz.f32 %f661, %f442, %f660; fma.rn.ftz.f32 %f662, %f635, %f658, %f661; mul.ftz.f32 %f663, %f447, %f662; fma.rn.ftz.f32 %f664, %f636, %f656, %f663; sub.ftz.f32 %f665, %f664, %f650; fma.rn.ftz.f32 %f26, %f452, %f665, %f650; mov.b32 %r13, %f26; setp.nan.ftz.f32 %p209, %f26, %f26; mov.f32 %f1598, %f1600; @%p209 bra $L__BB79_36; or.b32 %r1257, %r13, -2147483648; setp.ne.s32 %p210, %r1257, -2147483648; add.ftz.f32 %f667, %f26, %f26; setp.eq.ftz.f32 %p211, %f667, %f26; and.pred %p212, %p211, %p210; mov.f32 %f1598, %f1600; @%p212 bra $L__BB79_36; shl.b32 %r1258, %r13, 1; setp.lt.u32 %p213, %r1258, -16777215; mul.ftz.f32 %f668, %f26, 0f3F559B3D; selp.f32 %f1598, %f668, 0f00000000, %p213; $L__BB79_36: add.ftz.f32 %f670, %f1601, 0f42DF4487; cvt.rmi.ftz.f32.f32 %f671, %f670; cvt.rzi.ftz.s32.f32 %r1259, %f671; sub.ftz.f32 %f672, %f670, %f671; add.ftz.f32 %f673, %f1602, 0f431D5E90; cvt.rmi.ftz.f32.f32 %f674, %f673; cvt.rzi.ftz.s32.f32 %r1260, %f674; sub.ftz.f32 %f675, %f673, %f674; add.ftz.f32 %f676, %f1603, 0f4347168E; cvt.rmi.ftz.f32.f32 %f677, %f676; cvt.rzi.ftz.s32.f32 %r1261, %f677; sub.ftz.f32 %f678, %f676, %f677; add.ftz.f32 %f679, %f1604, 0f43326FA6; cvt.rmi.ftz.f32.f32 %f680, %f679; cvt.rzi.ftz.s32.f32 %r1262, %f680; sub.ftz.f32 %f681, %f679, %f680; mul.ftz.f32 %f682, %f672, %f672; mul.ftz.f32 %f683, %f672, %f682; fma.rn.ftz.f32 %f684, %f672, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f685, %f672, %f684, 0f41200000; mul.ftz.f32 %f686, %f683, %f685; mul.ftz.f32 %f687, %f675, %f675; mul.ftz.f32 %f688, %f675, %f687; fma.rn.ftz.f32 %f689, %f675, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f690, %f675, %f689, 0f41200000; mul.ftz.f32 %f691, %f688, %f690; mul.ftz.f32 %f692, %f678, %f678; mul.ftz.f32 %f693, %f678, %f692; fma.rn.ftz.f32 %f694, %f678, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f695, %f678, %f694, 0f41200000; mul.ftz.f32 %f696, %f693, %f695; mul.ftz.f32 %f697, %f681, %f681; mul.ftz.f32 %f698, %f681, %f697; fma.rn.ftz.f32 %f699, %f681, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f700, %f681, %f699, 0f41200000; mul.ftz.f32 %f701, %f698, %f700; add.s32 %r1263, %r1260, -559038708; add.s32 %r1264, %r1261, -559038708; sub.s32 %r1265, %r1259, %r1261; shf.l.wrap.b32 %r1266, %r1264, %r1264, 4; xor.b32 %r1267, %r1266, %r1265; add.s32 %r1268, %r1264, %r1263; sub.s32 %r1269, %r1263, %r1267; shf.l.wrap.b32 %r1270, %r1267, %r1267, 6; xor.b32 %r1271, %r1270, %r1269; add.s32 %r1272, %r1267, %r1268; sub.s32 %r1273, %r1268, %r1271; shf.l.wrap.b32 %r1274, %r1271, %r1271, 8; xor.b32 %r1275, %r1274, %r1273; add.s32 %r1276, %r1271, %r1272; sub.s32 %r1277, %r1272, %r1275; shf.l.wrap.b32 %r1278, %r1275, %r1275, 16; xor.b32 %r1279, %r1278, %r1277; add.s32 %r1280, %r1275, %r1276; sub.s32 %r1281, %r1276, %r1279; shf.l.wrap.b32 %r1282, %r1279, %r1279, 19; xor.b32 %r1283, %r1282, %r1281; add.s32 %r1284, %r1279, %r1280; sub.s32 %r1285, %r1280, %r1283; shf.l.wrap.b32 %r1286, %r1283, %r1283, 4; add.s32 %r1287, %r1283, %r1284; add.s32 %r1288, %r1284, %r1262; xor.b32 %r1289, %r1287, %r1285; xor.b32 %r1290, %r1289, %r1286; shf.l.wrap.b32 %r1291, %r1287, %r1287, 14; sub.s32 %r1292, %r1290, %r1291; xor.b32 %r1293, %r1292, %r1288; shf.l.wrap.b32 %r1294, %r1292, %r1292, 11; sub.s32 %r1295, %r1293, %r1294; xor.b32 %r1296, %r1295, %r1287; shf.l.wrap.b32 %r1297, %r1295, %r1295, 25; sub.s32 %r1298, %r1296, %r1297; xor.b32 %r1299, %r1298, %r1292; shf.l.wrap.b32 %r1300, %r1298, %r1298, 16; sub.s32 %r1301, %r1299, %r1300; xor.b32 %r1302, %r1301, %r1295; shf.l.wrap.b32 %r1303, %r1301, %r1301, 4; sub.s32 %r1304, %r1302, %r1303; xor.b32 %r1305, %r1304, %r1298; shf.l.wrap.b32 %r1306, %r1304, %r1304, 14; sub.s32 %r1307, %r1305, %r1306; xor.b32 %r1308, %r1307, %r1301; shf.l.wrap.b32 %r1309, %r1307, %r1307, 24; sub.s32 %r1310, %r1308, %r1309; and.b32 %r1311, %r1310, 31; setp.lt.u32 %p214, %r1311, 24; selp.f32 %f702, %f672, %f675, %p214; setp.lt.u32 %p215, %r1311, 16; selp.f32 %f703, %f675, %f678, %p215; setp.lt.u32 %p216, %r1311, 8; selp.f32 %f704, %f678, %f681, %p216; and.b32 %r1312, %r1310, 1; setp.eq.b32 %p217, %r1312, 1; neg.ftz.f32 %f705, %f702; selp.f32 %f706, %f705, %f702, %p217; and.b32 %r1313, %r1310, 2; setp.eq.s32 %p218, %r1313, 0; neg.ftz.f32 %f707, %f703; selp.f32 %f708, %f703, %f707, %p218; add.ftz.f32 %f709, %f706, %f708; and.b32 %r1314, %r1310, 4; setp.eq.s32 %p219, %r1314, 0; neg.ftz.f32 %f710, %f704; selp.f32 %f711, %f704, %f710, %p219; add.ftz.f32 %f712, %f711, %f709; add.s32 %r1315, %r1259, 1; sub.s32 %r1316, %r1315, %r1261; xor.b32 %r1317, %r1266, %r1316; sub.s32 %r1318, %r1263, %r1317; shf.l.wrap.b32 %r1319, %r1317, %r1317, 6; xor.b32 %r1320, %r1319, %r1318; add.s32 %r1321, %r1317, %r1268; sub.s32 %r1322, %r1268, %r1320; shf.l.wrap.b32 %r1323, %r1320, %r1320, 8; xor.b32 %r1324, %r1323, %r1322; add.s32 %r1325, %r1320, %r1321; sub.s32 %r1326, %r1321, %r1324; shf.l.wrap.b32 %r1327, %r1324, %r1324, 16; xor.b32 %r1328, %r1327, %r1326; add.s32 %r1329, %r1324, %r1325; sub.s32 %r1330, %r1325, %r1328; shf.l.wrap.b32 %r1331, %r1328, %r1328, 19; xor.b32 %r1332, %r1331, %r1330; add.s32 %r1333, %r1328, %r1329; sub.s32 %r1334, %r1329, %r1332; shf.l.wrap.b32 %r1335, %r1332, %r1332, 4; add.s32 %r1336, %r1332, %r1333; add.s32 %r1337, %r1333, %r1262; xor.b32 %r1338, %r1336, %r1334; xor.b32 %r1339, %r1338, %r1335; shf.l.wrap.b32 %r1340, %r1336, %r1336, 14; sub.s32 %r1341, %r1339, %r1340; xor.b32 %r1342, %r1341, %r1337; shf.l.wrap.b32 %r1343, %r1341, %r1341, 11; sub.s32 %r1344, %r1342, %r1343; xor.b32 %r1345, %r1344, %r1336; shf.l.wrap.b32 %r1346, %r1344, %r1344, 25; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1341; shf.l.wrap.b32 %r1349, %r1347, %r1347, 16; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1344; shf.l.wrap.b32 %r1352, %r1350, %r1350, 4; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 14; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 24; sub.s32 %r1359, %r1357, %r1358; add.ftz.f32 %f713, %f672, 0fBF800000; and.b32 %r1360, %r1359, 31; setp.lt.u32 %p220, %r1360, 24; selp.f32 %f715, %f713, %f675, %p220; setp.lt.u32 %p221, %r1360, 16; selp.f32 %f716, %f675, %f678, %p221; setp.lt.u32 %p222, %r1360, 8; selp.f32 %f717, %f678, %f681, %p222; and.b32 %r1361, %r1359, 1; setp.eq.b32 %p223, %r1361, 1; neg.ftz.f32 %f718, %f715; selp.f32 %f719, %f718, %f715, %p223; and.b32 %r1362, %r1359, 2; setp.eq.s32 %p224, %r1362, 0; neg.ftz.f32 %f720, %f716; selp.f32 %f721, %f716, %f720, %p224; add.ftz.f32 %f722, %f719, %f721; and.b32 %r1363, %r1359, 4; setp.eq.s32 %p225, %r1363, 0; neg.ftz.f32 %f723, %f717; selp.f32 %f724, %f717, %f723, %p225; add.ftz.f32 %f725, %f724, %f722; add.s32 %r1364, %r1260, -559038707; add.s32 %r1365, %r1264, %r1364; sub.s32 %r1366, %r1364, %r1267; xor.b32 %r1367, %r1270, %r1366; add.s32 %r1368, %r1267, %r1365; sub.s32 %r1369, %r1365, %r1367; shf.l.wrap.b32 %r1370, %r1367, %r1367, 8; xor.b32 %r1371, %r1370, %r1369; add.s32 %r1372, %r1367, %r1368; sub.s32 %r1373, %r1368, %r1371; shf.l.wrap.b32 %r1374, %r1371, %r1371, 16; xor.b32 %r1375, %r1374, %r1373; add.s32 %r1376, %r1371, %r1372; sub.s32 %r1377, %r1372, %r1375; shf.l.wrap.b32 %r1378, %r1375, %r1375, 19; xor.b32 %r1379, %r1378, %r1377; add.s32 %r1380, %r1375, %r1376; sub.s32 %r1381, %r1376, %r1379; shf.l.wrap.b32 %r1382, %r1379, %r1379, 4; add.s32 %r1383, %r1379, %r1380; add.s32 %r1384, %r1380, %r1262; xor.b32 %r1385, %r1383, %r1381; xor.b32 %r1386, %r1385, %r1382; shf.l.wrap.b32 %r1387, %r1383, %r1383, 14; sub.s32 %r1388, %r1386, %r1387; xor.b32 %r1389, %r1388, %r1384; shf.l.wrap.b32 %r1390, %r1388, %r1388, 11; sub.s32 %r1391, %r1389, %r1390; xor.b32 %r1392, %r1391, %r1383; shf.l.wrap.b32 %r1393, %r1391, %r1391, 25; sub.s32 %r1394, %r1392, %r1393; xor.b32 %r1395, %r1394, %r1388; shf.l.wrap.b32 %r1396, %r1394, %r1394, 16; sub.s32 %r1397, %r1395, %r1396; xor.b32 %r1398, %r1397, %r1391; shf.l.wrap.b32 %r1399, %r1397, %r1397, 4; sub.s32 %r1400, %r1398, %r1399; xor.b32 %r1401, %r1400, %r1394; shf.l.wrap.b32 %r1402, %r1400, %r1400, 14; sub.s32 %r1403, %r1401, %r1402; xor.b32 %r1404, %r1403, %r1397; shf.l.wrap.b32 %r1405, %r1403, %r1403, 24; sub.s32 %r1406, %r1404, %r1405; add.ftz.f32 %f726, %f675, 0fBF800000; and.b32 %r1407, %r1406, 31; setp.lt.u32 %p226, %r1407, 24; selp.f32 %f727, %f672, %f726, %p226; setp.lt.u32 %p227, %r1407, 16; selp.f32 %f728, %f726, %f678, %p227; setp.lt.u32 %p228, %r1407, 8; selp.f32 %f729, %f678, %f681, %p228; and.b32 %r1408, %r1406, 1; setp.eq.b32 %p229, %r1408, 1; neg.ftz.f32 %f730, %f727; selp.f32 %f731, %f730, %f727, %p229; and.b32 %r1409, %r1406, 2; setp.eq.s32 %p230, %r1409, 0; neg.ftz.f32 %f732, %f728; selp.f32 %f733, %f728, %f732, %p230; add.ftz.f32 %f734, %f731, %f733; and.b32 %r1410, %r1406, 4; setp.eq.s32 %p231, %r1410, 0; neg.ftz.f32 %f735, %f729; selp.f32 %f736, %f729, %f735, %p231; add.ftz.f32 %f737, %f736, %f734; sub.s32 %r1411, %r1364, %r1317; xor.b32 %r1412, %r1319, %r1411; add.s32 %r1413, %r1317, %r1365; sub.s32 %r1414, %r1365, %r1412; shf.l.wrap.b32 %r1415, %r1412, %r1412, 8; xor.b32 %r1416, %r1415, %r1414; add.s32 %r1417, %r1412, %r1413; sub.s32 %r1418, %r1413, %r1416; shf.l.wrap.b32 %r1419, %r1416, %r1416, 16; xor.b32 %r1420, %r1419, %r1418; add.s32 %r1421, %r1416, %r1417; sub.s32 %r1422, %r1417, %r1420; shf.l.wrap.b32 %r1423, %r1420, %r1420, 19; xor.b32 %r1424, %r1423, %r1422; add.s32 %r1425, %r1420, %r1421; sub.s32 %r1426, %r1421, %r1424; shf.l.wrap.b32 %r1427, %r1424, %r1424, 4; add.s32 %r1428, %r1424, %r1425; add.s32 %r1429, %r1425, %r1262; xor.b32 %r1430, %r1428, %r1426; xor.b32 %r1431, %r1430, %r1427; shf.l.wrap.b32 %r1432, %r1428, %r1428, 14; sub.s32 %r1433, %r1431, %r1432; xor.b32 %r1434, %r1433, %r1429; shf.l.wrap.b32 %r1435, %r1433, %r1433, 11; sub.s32 %r1436, %r1434, %r1435; xor.b32 %r1437, %r1436, %r1428; shf.l.wrap.b32 %r1438, %r1436, %r1436, 25; sub.s32 %r1439, %r1437, %r1438; xor.b32 %r1440, %r1439, %r1433; shf.l.wrap.b32 %r1441, %r1439, %r1439, 16; sub.s32 %r1442, %r1440, %r1441; xor.b32 %r1443, %r1442, %r1436; shf.l.wrap.b32 %r1444, %r1442, %r1442, 4; sub.s32 %r1445, %r1443, %r1444; xor.b32 %r1446, %r1445, %r1439; shf.l.wrap.b32 %r1447, %r1445, %r1445, 14; sub.s32 %r1448, %r1446, %r1447; xor.b32 %r1449, %r1448, %r1442; shf.l.wrap.b32 %r1450, %r1448, %r1448, 24; sub.s32 %r1451, %r1449, %r1450; and.b32 %r1452, %r1451, 31; setp.lt.u32 %p232, %r1452, 24; selp.f32 %f738, %f713, %f726, %p232; setp.lt.u32 %p233, %r1452, 16; selp.f32 %f739, %f726, %f678, %p233; setp.lt.u32 %p234, %r1452, 8; selp.f32 %f740, %f678, %f681, %p234; and.b32 %r1453, %r1451, 1; setp.eq.b32 %p235, %r1453, 1; neg.ftz.f32 %f741, %f738; selp.f32 %f742, %f741, %f738, %p235; and.b32 %r1454, %r1451, 2; setp.eq.s32 %p236, %r1454, 0; neg.ftz.f32 %f743, %f739; selp.f32 %f744, %f739, %f743, %p236; add.ftz.f32 %f745, %f742, %f744; and.b32 %r1455, %r1451, 4; setp.eq.s32 %p237, %r1455, 0; neg.ftz.f32 %f746, %f740; selp.f32 %f747, %f740, %f746, %p237; add.ftz.f32 %f748, %f747, %f745; not.b32 %r1456, %r1261; add.s32 %r1457, %r1261, -559038707; add.s32 %r1458, %r1259, %r1456; shf.l.wrap.b32 %r1459, %r1457, %r1457, 4; xor.b32 %r1460, %r1459, %r1458; add.s32 %r1461, %r1457, %r1263; sub.s32 %r1462, %r1263, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 6; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 8; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 16; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 19; xor.b32 %r1476, %r1475, %r1474; add.s32 %r1477, %r1472, %r1473; sub.s32 %r1478, %r1473, %r1476; shf.l.wrap.b32 %r1479, %r1476, %r1476, 4; add.s32 %r1480, %r1476, %r1477; add.s32 %r1481, %r1477, %r1262; xor.b32 %r1482, %r1480, %r1478; xor.b32 %r1483, %r1482, %r1479; shf.l.wrap.b32 %r1484, %r1480, %r1480, 14; sub.s32 %r1485, %r1483, %r1484; xor.b32 %r1486, %r1485, %r1481; shf.l.wrap.b32 %r1487, %r1485, %r1485, 11; sub.s32 %r1488, %r1486, %r1487; xor.b32 %r1489, %r1488, %r1480; shf.l.wrap.b32 %r1490, %r1488, %r1488, 25; sub.s32 %r1491, %r1489, %r1490; xor.b32 %r1492, %r1491, %r1485; shf.l.wrap.b32 %r1493, %r1491, %r1491, 16; sub.s32 %r1494, %r1492, %r1493; xor.b32 %r1495, %r1494, %r1488; shf.l.wrap.b32 %r1496, %r1494, %r1494, 4; sub.s32 %r1497, %r1495, %r1496; xor.b32 %r1498, %r1497, %r1491; shf.l.wrap.b32 %r1499, %r1497, %r1497, 14; sub.s32 %r1500, %r1498, %r1499; xor.b32 %r1501, %r1500, %r1494; shf.l.wrap.b32 %r1502, %r1500, %r1500, 24; sub.s32 %r1503, %r1501, %r1502; add.ftz.f32 %f749, %f678, 0fBF800000; and.b32 %r1504, %r1503, 31; setp.lt.u32 %p238, %r1504, 24; selp.f32 %f750, %f672, %f675, %p238; setp.lt.u32 %p239, %r1504, 16; selp.f32 %f751, %f675, %f749, %p239; setp.lt.u32 %p240, %r1504, 8; selp.f32 %f752, %f749, %f681, %p240; and.b32 %r1505, %r1503, 1; setp.eq.b32 %p241, %r1505, 1; neg.ftz.f32 %f753, %f750; selp.f32 %f754, %f753, %f750, %p241; and.b32 %r1506, %r1503, 2; setp.eq.s32 %p242, %r1506, 0; neg.ftz.f32 %f755, %f751; selp.f32 %f756, %f751, %f755, %p242; add.ftz.f32 %f757, %f754, %f756; and.b32 %r1507, %r1503, 4; setp.eq.s32 %p243, %r1507, 0; neg.ftz.f32 %f758, %f752; selp.f32 %f759, %f752, %f758, %p243; add.ftz.f32 %f760, %f759, %f757; xor.b32 %r1508, %r1459, %r1265; sub.s32 %r1509, %r1263, %r1508; shf.l.wrap.b32 %r1510, %r1508, %r1508, 6; xor.b32 %r1511, %r1510, %r1509; add.s32 %r1512, %r1508, %r1461; sub.s32 %r1513, %r1461, %r1511; shf.l.wrap.b32 %r1514, %r1511, %r1511, 8; xor.b32 %r1515, %r1514, %r1513; add.s32 %r1516, %r1511, %r1512; sub.s32 %r1517, %r1512, %r1515; shf.l.wrap.b32 %r1518, %r1515, %r1515, 16; xor.b32 %r1519, %r1518, %r1517; add.s32 %r1520, %r1515, %r1516; sub.s32 %r1521, %r1516, %r1519; shf.l.wrap.b32 %r1522, %r1519, %r1519, 19; xor.b32 %r1523, %r1522, %r1521; add.s32 %r1524, %r1519, %r1520; sub.s32 %r1525, %r1520, %r1523; shf.l.wrap.b32 %r1526, %r1523, %r1523, 4; add.s32 %r1527, %r1523, %r1524; add.s32 %r1528, %r1524, %r1262; xor.b32 %r1529, %r1527, %r1525; xor.b32 %r1530, %r1529, %r1526; shf.l.wrap.b32 %r1531, %r1527, %r1527, 14; sub.s32 %r1532, %r1530, %r1531; xor.b32 %r1533, %r1532, %r1528; shf.l.wrap.b32 %r1534, %r1532, %r1532, 11; sub.s32 %r1535, %r1533, %r1534; xor.b32 %r1536, %r1535, %r1527; shf.l.wrap.b32 %r1537, %r1535, %r1535, 25; sub.s32 %r1538, %r1536, %r1537; xor.b32 %r1539, %r1538, %r1532; shf.l.wrap.b32 %r1540, %r1538, %r1538, 16; sub.s32 %r1541, %r1539, %r1540; xor.b32 %r1542, %r1541, %r1535; shf.l.wrap.b32 %r1543, %r1541, %r1541, 4; sub.s32 %r1544, %r1542, %r1543; xor.b32 %r1545, %r1544, %r1538; shf.l.wrap.b32 %r1546, %r1544, %r1544, 14; sub.s32 %r1547, %r1545, %r1546; xor.b32 %r1548, %r1547, %r1541; shf.l.wrap.b32 %r1549, %r1547, %r1547, 24; sub.s32 %r1550, %r1548, %r1549; and.b32 %r1551, %r1550, 31; setp.lt.u32 %p244, %r1551, 24; selp.f32 %f761, %f713, %f675, %p244; setp.lt.u32 %p245, %r1551, 16; selp.f32 %f762, %f675, %f749, %p245; setp.lt.u32 %p246, %r1551, 8; selp.f32 %f763, %f749, %f681, %p246; and.b32 %r1552, %r1550, 1; setp.eq.b32 %p247, %r1552, 1; neg.ftz.f32 %f764, %f761; selp.f32 %f765, %f764, %f761, %p247; and.b32 %r1553, %r1550, 2; setp.eq.s32 %p248, %r1553, 0; neg.ftz.f32 %f766, %f762; selp.f32 %f767, %f762, %f766, %p248; add.ftz.f32 %f768, %f765, %f767; and.b32 %r1554, %r1550, 4; setp.eq.s32 %p249, %r1554, 0; neg.ftz.f32 %f769, %f763; selp.f32 %f770, %f763, %f769, %p249; add.ftz.f32 %f771, %f770, %f768; add.s32 %r1555, %r1457, %r1364; sub.s32 %r1556, %r1364, %r1460; xor.b32 %r1557, %r1463, %r1556; add.s32 %r1558, %r1460, %r1555; sub.s32 %r1559, %r1555, %r1557; shf.l.wrap.b32 %r1560, %r1557, %r1557, 8; xor.b32 %r1561, %r1560, %r1559; add.s32 %r1562, %r1557, %r1558; sub.s32 %r1563, %r1558, %r1561; shf.l.wrap.b32 %r1564, %r1561, %r1561, 16; xor.b32 %r1565, %r1564, %r1563; add.s32 %r1566, %r1561, %r1562; sub.s32 %r1567, %r1562, %r1565; shf.l.wrap.b32 %r1568, %r1565, %r1565, 19; xor.b32 %r1569, %r1568, %r1567; add.s32 %r1570, %r1565, %r1566; sub.s32 %r1571, %r1566, %r1569; shf.l.wrap.b32 %r1572, %r1569, %r1569, 4; add.s32 %r1573, %r1569, %r1570; add.s32 %r1574, %r1570, %r1262; xor.b32 %r1575, %r1573, %r1571; xor.b32 %r1576, %r1575, %r1572; shf.l.wrap.b32 %r1577, %r1573, %r1573, 14; sub.s32 %r1578, %r1576, %r1577; xor.b32 %r1579, %r1578, %r1574; shf.l.wrap.b32 %r1580, %r1578, %r1578, 11; sub.s32 %r1581, %r1579, %r1580; xor.b32 %r1582, %r1581, %r1573; shf.l.wrap.b32 %r1583, %r1581, %r1581, 25; sub.s32 %r1584, %r1582, %r1583; xor.b32 %r1585, %r1584, %r1578; shf.l.wrap.b32 %r1586, %r1584, %r1584, 16; sub.s32 %r1587, %r1585, %r1586; xor.b32 %r1588, %r1587, %r1581; shf.l.wrap.b32 %r1589, %r1587, %r1587, 4; sub.s32 %r1590, %r1588, %r1589; xor.b32 %r1591, %r1590, %r1584; shf.l.wrap.b32 %r1592, %r1590, %r1590, 14; sub.s32 %r1593, %r1591, %r1592; xor.b32 %r1594, %r1593, %r1587; shf.l.wrap.b32 %r1595, %r1593, %r1593, 24; sub.s32 %r1596, %r1594, %r1595; and.b32 %r1597, %r1596, 31; setp.lt.u32 %p250, %r1597, 24; selp.f32 %f772, %f672, %f726, %p250; setp.lt.u32 %p251, %r1597, 16; selp.f32 %f773, %f726, %f749, %p251; setp.lt.u32 %p252, %r1597, 8; selp.f32 %f774, %f749, %f681, %p252; and.b32 %r1598, %r1596, 1; setp.eq.b32 %p253, %r1598, 1; neg.ftz.f32 %f775, %f772; selp.f32 %f776, %f775, %f772, %p253; and.b32 %r1599, %r1596, 2; setp.eq.s32 %p254, %r1599, 0; neg.ftz.f32 %f777, %f773; selp.f32 %f778, %f773, %f777, %p254; add.ftz.f32 %f779, %f776, %f778; and.b32 %r1600, %r1596, 4; setp.eq.s32 %p255, %r1600, 0; neg.ftz.f32 %f780, %f774; selp.f32 %f781, %f774, %f780, %p255; add.ftz.f32 %f782, %f781, %f779; sub.s32 %r1601, %r1364, %r1508; xor.b32 %r1602, %r1510, %r1601; add.s32 %r1603, %r1508, %r1555; sub.s32 %r1604, %r1555, %r1602; shf.l.wrap.b32 %r1605, %r1602, %r1602, 8; xor.b32 %r1606, %r1605, %r1604; add.s32 %r1607, %r1602, %r1603; sub.s32 %r1608, %r1603, %r1606; shf.l.wrap.b32 %r1609, %r1606, %r1606, 16; xor.b32 %r1610, %r1609, %r1608; add.s32 %r1611, %r1606, %r1607; sub.s32 %r1612, %r1607, %r1610; shf.l.wrap.b32 %r1613, %r1610, %r1610, 19; xor.b32 %r1614, %r1613, %r1612; add.s32 %r1615, %r1610, %r1611; sub.s32 %r1616, %r1611, %r1614; shf.l.wrap.b32 %r1617, %r1614, %r1614, 4; add.s32 %r1618, %r1614, %r1615; add.s32 %r1619, %r1615, %r1262; xor.b32 %r1620, %r1618, %r1616; xor.b32 %r1621, %r1620, %r1617; shf.l.wrap.b32 %r1622, %r1618, %r1618, 14; sub.s32 %r1623, %r1621, %r1622; xor.b32 %r1624, %r1623, %r1619; shf.l.wrap.b32 %r1625, %r1623, %r1623, 11; sub.s32 %r1626, %r1624, %r1625; xor.b32 %r1627, %r1626, %r1618; shf.l.wrap.b32 %r1628, %r1626, %r1626, 25; sub.s32 %r1629, %r1627, %r1628; xor.b32 %r1630, %r1629, %r1623; shf.l.wrap.b32 %r1631, %r1629, %r1629, 16; sub.s32 %r1632, %r1630, %r1631; xor.b32 %r1633, %r1632, %r1626; shf.l.wrap.b32 %r1634, %r1632, %r1632, 4; sub.s32 %r1635, %r1633, %r1634; xor.b32 %r1636, %r1635, %r1629; shf.l.wrap.b32 %r1637, %r1635, %r1635, 14; sub.s32 %r1638, %r1636, %r1637; xor.b32 %r1639, %r1638, %r1632; shf.l.wrap.b32 %r1640, %r1638, %r1638, 24; sub.s32 %r1641, %r1639, %r1640; and.b32 %r1642, %r1641, 31; setp.lt.u32 %p256, %r1642, 24; selp.f32 %f783, %f713, %f726, %p256; setp.lt.u32 %p257, %r1642, 16; selp.f32 %f784, %f726, %f749, %p257; setp.lt.u32 %p258, %r1642, 8; selp.f32 %f785, %f749, %f681, %p258; and.b32 %r1643, %r1641, 1; setp.eq.b32 %p259, %r1643, 1; neg.ftz.f32 %f786, %f783; selp.f32 %f787, %f786, %f783, %p259; and.b32 %r1644, %r1641, 2; setp.eq.s32 %p260, %r1644, 0; neg.ftz.f32 %f788, %f784; selp.f32 %f789, %f784, %f788, %p260; add.ftz.f32 %f790, %f787, %f789; and.b32 %r1645, %r1641, 4; setp.eq.s32 %p261, %r1645, 0; neg.ftz.f32 %f791, %f785; selp.f32 %f792, %f785, %f791, %p261; add.ftz.f32 %f793, %f792, %f790; add.s32 %r1646, %r1262, 1; add.s32 %r1647, %r1284, %r1646; xor.b32 %r1648, %r1292, %r1647; sub.s32 %r1649, %r1648, %r1294; xor.b32 %r1650, %r1649, %r1287; shf.l.wrap.b32 %r1651, %r1649, %r1649, 25; sub.s32 %r1652, %r1650, %r1651; xor.b32 %r1653, %r1652, %r1292; shf.l.wrap.b32 %r1654, %r1652, %r1652, 16; sub.s32 %r1655, %r1653, %r1654; xor.b32 %r1656, %r1655, %r1649; shf.l.wrap.b32 %r1657, %r1655, %r1655, 4; sub.s32 %r1658, %r1656, %r1657; xor.b32 %r1659, %r1658, %r1652; shf.l.wrap.b32 %r1660, %r1658, %r1658, 14; sub.s32 %r1661, %r1659, %r1660; xor.b32 %r1662, %r1661, %r1655; shf.l.wrap.b32 %r1663, %r1661, %r1661, 24; sub.s32 %r1664, %r1662, %r1663; add.ftz.f32 %f794, %f681, 0fBF800000; and.b32 %r1665, %r1664, 31; setp.lt.u32 %p262, %r1665, 24; selp.f32 %f795, %f672, %f675, %p262; setp.lt.u32 %p263, %r1665, 16; selp.f32 %f796, %f675, %f678, %p263; setp.lt.u32 %p264, %r1665, 8; selp.f32 %f797, %f678, %f794, %p264; and.b32 %r1666, %r1664, 1; setp.eq.b32 %p265, %r1666, 1; neg.ftz.f32 %f798, %f795; selp.f32 %f799, %f798, %f795, %p265; and.b32 %r1667, %r1664, 2; setp.eq.s32 %p266, %r1667, 0; neg.ftz.f32 %f800, %f796; selp.f32 %f801, %f796, %f800, %p266; add.ftz.f32 %f802, %f799, %f801; and.b32 %r1668, %r1664, 4; setp.eq.s32 %p267, %r1668, 0; neg.ftz.f32 %f803, %f797; selp.f32 %f804, %f797, %f803, %p267; add.ftz.f32 %f805, %f804, %f802; add.s32 %r1669, %r1333, %r1646; xor.b32 %r1670, %r1341, %r1669; sub.s32 %r1671, %r1670, %r1343; xor.b32 %r1672, %r1671, %r1336; shf.l.wrap.b32 %r1673, %r1671, %r1671, 25; sub.s32 %r1674, %r1672, %r1673; xor.b32 %r1675, %r1674, %r1341; shf.l.wrap.b32 %r1676, %r1674, %r1674, 16; sub.s32 %r1677, %r1675, %r1676; xor.b32 %r1678, %r1677, %r1671; shf.l.wrap.b32 %r1679, %r1677, %r1677, 4; sub.s32 %r1680, %r1678, %r1679; xor.b32 %r1681, %r1680, %r1674; shf.l.wrap.b32 %r1682, %r1680, %r1680, 14; sub.s32 %r1683, %r1681, %r1682; xor.b32 %r1684, %r1683, %r1677; shf.l.wrap.b32 %r1685, %r1683, %r1683, 24; sub.s32 %r1686, %r1684, %r1685; and.b32 %r1687, %r1686, 31; setp.lt.u32 %p268, %r1687, 24; selp.f32 %f806, %f713, %f675, %p268; setp.lt.u32 %p269, %r1687, 16; selp.f32 %f807, %f675, %f678, %p269; setp.lt.u32 %p270, %r1687, 8; selp.f32 %f808, %f678, %f794, %p270; and.b32 %r1688, %r1686, 1; setp.eq.b32 %p271, %r1688, 1; neg.ftz.f32 %f809, %f806; selp.f32 %f810, %f809, %f806, %p271; and.b32 %r1689, %r1686, 2; setp.eq.s32 %p272, %r1689, 0; neg.ftz.f32 %f811, %f807; selp.f32 %f812, %f807, %f811, %p272; add.ftz.f32 %f813, %f810, %f812; and.b32 %r1690, %r1686, 4; setp.eq.s32 %p273, %r1690, 0; neg.ftz.f32 %f814, %f808; selp.f32 %f815, %f808, %f814, %p273; add.ftz.f32 %f816, %f815, %f813; add.s32 %r1691, %r1380, %r1646; xor.b32 %r1692, %r1388, %r1691; sub.s32 %r1693, %r1692, %r1390; xor.b32 %r1694, %r1693, %r1383; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1388; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; and.b32 %r1709, %r1708, 31; setp.lt.u32 %p274, %r1709, 24; selp.f32 %f817, %f672, %f726, %p274; setp.lt.u32 %p275, %r1709, 16; selp.f32 %f818, %f726, %f678, %p275; setp.lt.u32 %p276, %r1709, 8; selp.f32 %f819, %f678, %f794, %p276; and.b32 %r1710, %r1708, 1; setp.eq.b32 %p277, %r1710, 1; neg.ftz.f32 %f820, %f817; selp.f32 %f821, %f820, %f817, %p277; and.b32 %r1711, %r1708, 2; setp.eq.s32 %p278, %r1711, 0; neg.ftz.f32 %f822, %f818; selp.f32 %f823, %f818, %f822, %p278; add.ftz.f32 %f824, %f821, %f823; and.b32 %r1712, %r1708, 4; setp.eq.s32 %p279, %r1712, 0; neg.ftz.f32 %f825, %f819; selp.f32 %f826, %f819, %f825, %p279; add.ftz.f32 %f827, %f826, %f824; add.s32 %r1713, %r1425, %r1646; xor.b32 %r1714, %r1433, %r1713; sub.s32 %r1715, %r1714, %r1435; xor.b32 %r1716, %r1715, %r1428; shf.l.wrap.b32 %r1717, %r1715, %r1715, 25; sub.s32 %r1718, %r1716, %r1717; xor.b32 %r1719, %r1718, %r1433; shf.l.wrap.b32 %r1720, %r1718, %r1718, 16; sub.s32 %r1721, %r1719, %r1720; xor.b32 %r1722, %r1721, %r1715; shf.l.wrap.b32 %r1723, %r1721, %r1721, 4; sub.s32 %r1724, %r1722, %r1723; xor.b32 %r1725, %r1724, %r1718; shf.l.wrap.b32 %r1726, %r1724, %r1724, 14; sub.s32 %r1727, %r1725, %r1726; xor.b32 %r1728, %r1727, %r1721; shf.l.wrap.b32 %r1729, %r1727, %r1727, 24; sub.s32 %r1730, %r1728, %r1729; and.b32 %r1731, %r1730, 31; setp.lt.u32 %p280, %r1731, 24; selp.f32 %f828, %f713, %f726, %p280; setp.lt.u32 %p281, %r1731, 16; selp.f32 %f829, %f726, %f678, %p281; setp.lt.u32 %p282, %r1731, 8; selp.f32 %f830, %f678, %f794, %p282; and.b32 %r1732, %r1730, 1; setp.eq.b32 %p283, %r1732, 1; neg.ftz.f32 %f831, %f828; selp.f32 %f832, %f831, %f828, %p283; and.b32 %r1733, %r1730, 2; setp.eq.s32 %p284, %r1733, 0; neg.ftz.f32 %f833, %f829; selp.f32 %f834, %f829, %f833, %p284; add.ftz.f32 %f835, %f832, %f834; and.b32 %r1734, %r1730, 4; setp.eq.s32 %p285, %r1734, 0; neg.ftz.f32 %f836, %f830; selp.f32 %f837, %f830, %f836, %p285; add.ftz.f32 %f838, %f837, %f835; add.s32 %r1735, %r1477, %r1646; xor.b32 %r1736, %r1485, %r1735; sub.s32 %r1737, %r1736, %r1487; xor.b32 %r1738, %r1737, %r1480; shf.l.wrap.b32 %r1739, %r1737, %r1737, 25; sub.s32 %r1740, %r1738, %r1739; xor.b32 %r1741, %r1740, %r1485; shf.l.wrap.b32 %r1742, %r1740, %r1740, 16; sub.s32 %r1743, %r1741, %r1742; xor.b32 %r1744, %r1743, %r1737; shf.l.wrap.b32 %r1745, %r1743, %r1743, 4; sub.s32 %r1746, %r1744, %r1745; xor.b32 %r1747, %r1746, %r1740; shf.l.wrap.b32 %r1748, %r1746, %r1746, 14; sub.s32 %r1749, %r1747, %r1748; xor.b32 %r1750, %r1749, %r1743; shf.l.wrap.b32 %r1751, %r1749, %r1749, 24; sub.s32 %r1752, %r1750, %r1751; and.b32 %r1753, %r1752, 31; setp.lt.u32 %p286, %r1753, 24; selp.f32 %f839, %f672, %f675, %p286; setp.lt.u32 %p287, %r1753, 16; selp.f32 %f840, %f675, %f749, %p287; setp.lt.u32 %p288, %r1753, 8; selp.f32 %f841, %f749, %f794, %p288; and.b32 %r1754, %r1752, 1; setp.eq.b32 %p289, %r1754, 1; neg.ftz.f32 %f842, %f839; selp.f32 %f843, %f842, %f839, %p289; and.b32 %r1755, %r1752, 2; setp.eq.s32 %p290, %r1755, 0; neg.ftz.f32 %f844, %f840; selp.f32 %f845, %f840, %f844, %p290; add.ftz.f32 %f846, %f843, %f845; and.b32 %r1756, %r1752, 4; setp.eq.s32 %p291, %r1756, 0; neg.ftz.f32 %f847, %f841; selp.f32 %f848, %f841, %f847, %p291; add.ftz.f32 %f849, %f848, %f846; add.s32 %r1757, %r1524, %r1646; xor.b32 %r1758, %r1532, %r1757; sub.s32 %r1759, %r1758, %r1534; xor.b32 %r1760, %r1759, %r1527; shf.l.wrap.b32 %r1761, %r1759, %r1759, 25; sub.s32 %r1762, %r1760, %r1761; xor.b32 %r1763, %r1762, %r1532; shf.l.wrap.b32 %r1764, %r1762, %r1762, 16; sub.s32 %r1765, %r1763, %r1764; xor.b32 %r1766, %r1765, %r1759; shf.l.wrap.b32 %r1767, %r1765, %r1765, 4; sub.s32 %r1768, %r1766, %r1767; xor.b32 %r1769, %r1768, %r1762; shf.l.wrap.b32 %r1770, %r1768, %r1768, 14; sub.s32 %r1771, %r1769, %r1770; xor.b32 %r1772, %r1771, %r1765; shf.l.wrap.b32 %r1773, %r1771, %r1771, 24; sub.s32 %r1774, %r1772, %r1773; and.b32 %r1775, %r1774, 31; setp.lt.u32 %p292, %r1775, 24; selp.f32 %f850, %f713, %f675, %p292; setp.lt.u32 %p293, %r1775, 16; selp.f32 %f851, %f675, %f749, %p293; setp.lt.u32 %p294, %r1775, 8; selp.f32 %f852, %f749, %f794, %p294; and.b32 %r1776, %r1774, 1; setp.eq.b32 %p295, %r1776, 1; neg.ftz.f32 %f853, %f850; selp.f32 %f854, %f853, %f850, %p295; and.b32 %r1777, %r1774, 2; setp.eq.s32 %p296, %r1777, 0; neg.ftz.f32 %f855, %f851; selp.f32 %f856, %f851, %f855, %p296; add.ftz.f32 %f857, %f854, %f856; and.b32 %r1778, %r1774, 4; setp.eq.s32 %p297, %r1778, 0; neg.ftz.f32 %f858, %f852; selp.f32 %f859, %f852, %f858, %p297; add.ftz.f32 %f860, %f859, %f857; add.s32 %r1779, %r1570, %r1646; xor.b32 %r1780, %r1578, %r1779; sub.s32 %r1781, %r1780, %r1580; xor.b32 %r1782, %r1781, %r1573; shf.l.wrap.b32 %r1783, %r1781, %r1781, 25; sub.s32 %r1784, %r1782, %r1783; xor.b32 %r1785, %r1784, %r1578; shf.l.wrap.b32 %r1786, %r1784, %r1784, 16; sub.s32 %r1787, %r1785, %r1786; xor.b32 %r1788, %r1787, %r1781; shf.l.wrap.b32 %r1789, %r1787, %r1787, 4; sub.s32 %r1790, %r1788, %r1789; xor.b32 %r1791, %r1790, %r1784; shf.l.wrap.b32 %r1792, %r1790, %r1790, 14; sub.s32 %r1793, %r1791, %r1792; xor.b32 %r1794, %r1793, %r1787; shf.l.wrap.b32 %r1795, %r1793, %r1793, 24; sub.s32 %r1796, %r1794, %r1795; and.b32 %r1797, %r1796, 31; setp.lt.u32 %p298, %r1797, 24; selp.f32 %f861, %f672, %f726, %p298; setp.lt.u32 %p299, %r1797, 16; selp.f32 %f862, %f726, %f749, %p299; setp.lt.u32 %p300, %r1797, 8; selp.f32 %f863, %f749, %f794, %p300; and.b32 %r1798, %r1796, 1; setp.eq.b32 %p301, %r1798, 1; neg.ftz.f32 %f864, %f861; selp.f32 %f865, %f864, %f861, %p301; and.b32 %r1799, %r1796, 2; setp.eq.s32 %p302, %r1799, 0; neg.ftz.f32 %f866, %f862; selp.f32 %f867, %f862, %f866, %p302; add.ftz.f32 %f868, %f865, %f867; and.b32 %r1800, %r1796, 4; setp.eq.s32 %p303, %r1800, 0; neg.ftz.f32 %f869, %f863; selp.f32 %f870, %f863, %f869, %p303; add.ftz.f32 %f871, %f870, %f868; add.s32 %r1801, %r1615, %r1646; xor.b32 %r1802, %r1623, %r1801; sub.s32 %r1803, %r1802, %r1625; xor.b32 %r1804, %r1803, %r1618; shf.l.wrap.b32 %r1805, %r1803, %r1803, 25; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1623; shf.l.wrap.b32 %r1808, %r1806, %r1806, 16; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1803; shf.l.wrap.b32 %r1811, %r1809, %r1809, 4; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 14; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 24; sub.s32 %r1818, %r1816, %r1817; and.b32 %r1819, %r1818, 31; setp.lt.u32 %p304, %r1819, 24; selp.f32 %f872, %f713, %f726, %p304; setp.lt.u32 %p305, %r1819, 16; selp.f32 %f873, %f726, %f749, %p305; setp.lt.u32 %p306, %r1819, 8; selp.f32 %f874, %f749, %f794, %p306; and.b32 %r1820, %r1818, 1; setp.eq.b32 %p307, %r1820, 1; neg.ftz.f32 %f875, %f872; selp.f32 %f876, %f875, %f872, %p307; and.b32 %r1821, %r1818, 2; setp.eq.s32 %p308, %r1821, 0; neg.ftz.f32 %f877, %f873; selp.f32 %f878, %f873, %f877, %p308; add.ftz.f32 %f879, %f876, %f878; and.b32 %r1822, %r1818, 4; setp.eq.s32 %p309, %r1822, 0; neg.ftz.f32 %f880, %f874; selp.f32 %f881, %f874, %f880, %p309; add.ftz.f32 %f882, %f881, %f879; sub.ftz.f32 %f883, %f216, %f686; sub.ftz.f32 %f884, %f216, %f691; sub.ftz.f32 %f885, %f216, %f696; mul.ftz.f32 %f886, %f883, %f712; fma.rn.ftz.f32 %f887, %f686, %f725, %f886; mul.ftz.f32 %f888, %f883, %f737; fma.rn.ftz.f32 %f889, %f686, %f748, %f888; mul.ftz.f32 %f890, %f691, %f889; fma.rn.ftz.f32 %f891, %f884, %f887, %f890; mul.ftz.f32 %f892, %f883, %f760; fma.rn.ftz.f32 %f893, %f686, %f771, %f892; mul.ftz.f32 %f894, %f883, %f782; fma.rn.ftz.f32 %f895, %f686, %f793, %f894; mul.ftz.f32 %f896, %f691, %f895; fma.rn.ftz.f32 %f897, %f884, %f893, %f896; mul.ftz.f32 %f898, %f696, %f897; fma.rn.ftz.f32 %f899, %f885, %f891, %f898; mul.ftz.f32 %f900, %f883, %f805; fma.rn.ftz.f32 %f901, %f686, %f816, %f900; mul.ftz.f32 %f902, %f883, %f827; fma.rn.ftz.f32 %f903, %f686, %f838, %f902; mul.ftz.f32 %f904, %f691, %f903; fma.rn.ftz.f32 %f905, %f884, %f901, %f904; mul.ftz.f32 %f906, %f883, %f849; fma.rn.ftz.f32 %f907, %f686, %f860, %f906; mul.ftz.f32 %f908, %f883, %f871; fma.rn.ftz.f32 %f909, %f686, %f882, %f908; mul.ftz.f32 %f910, %f691, %f909; fma.rn.ftz.f32 %f911, %f884, %f907, %f910; mul.ftz.f32 %f912, %f696, %f911; fma.rn.ftz.f32 %f913, %f885, %f905, %f912; sub.ftz.f32 %f914, %f913, %f899; fma.rn.ftz.f32 %f29, %f701, %f914, %f899; mov.b32 %r14, %f29; setp.nan.ftz.f32 %p310, %f29, %f29; mov.f32 %f1599, %f1600; @%p310 bra $L__BB79_39; or.b32 %r1823, %r14, -2147483648; setp.ne.s32 %p311, %r1823, -2147483648; add.ftz.f32 %f916, %f29, %f29; setp.eq.ftz.f32 %p312, %f916, %f29; and.pred %p313, %p312, %p311; mov.f32 %f1599, %f1600; @%p313 bra $L__BB79_39; shl.b32 %r1824, %r14, 1; setp.lt.u32 %p314, %r1824, -16777215; mul.ftz.f32 %f917, %f29, 0f3F559B3D; selp.f32 %f1599, %f917, 0f00000000, %p314; $L__BB79_39: add.ftz.f32 %f919, %f1601, 0f432571FE; cvt.rmi.ftz.f32.f32 %f920, %f919; cvt.rzi.ftz.s32.f32 %r1825, %f920; sub.ftz.f32 %f921, %f919, %f920; add.ftz.f32 %f922, %f1602, 0f4322C4AA; cvt.rmi.ftz.f32.f32 %f923, %f922; cvt.rzi.ftz.s32.f32 %r1826, %f923; sub.ftz.f32 %f924, %f922, %f923; add.ftz.f32 %f925, %f1603, 0f4341688E; cvt.rmi.ftz.f32.f32 %f926, %f925; cvt.rzi.ftz.s32.f32 %r1827, %f926; sub.ftz.f32 %f927, %f925, %f926; add.ftz.f32 %f928, %f1604, 0f43114169; cvt.rmi.ftz.f32.f32 %f929, %f928; cvt.rzi.ftz.s32.f32 %r1828, %f929; sub.ftz.f32 %f930, %f928, %f929; mul.ftz.f32 %f931, %f921, %f921; mul.ftz.f32 %f932, %f921, %f931; fma.rn.ftz.f32 %f933, %f921, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f934, %f921, %f933, 0f41200000; mul.ftz.f32 %f935, %f932, %f934; mul.ftz.f32 %f936, %f924, %f924; mul.ftz.f32 %f937, %f924, %f936; fma.rn.ftz.f32 %f938, %f924, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f939, %f924, %f938, 0f41200000; mul.ftz.f32 %f940, %f937, %f939; mul.ftz.f32 %f941, %f927, %f927; mul.ftz.f32 %f942, %f927, %f941; fma.rn.ftz.f32 %f943, %f927, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f944, %f927, %f943, 0f41200000; mul.ftz.f32 %f945, %f942, %f944; mul.ftz.f32 %f946, %f930, %f930; mul.ftz.f32 %f947, %f930, %f946; fma.rn.ftz.f32 %f948, %f930, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f949, %f930, %f948, 0f41200000; mul.ftz.f32 %f950, %f947, %f949; add.s32 %r1829, %r1826, -559038708; add.s32 %r1830, %r1827, -559038708; sub.s32 %r1831, %r1825, %r1827; shf.l.wrap.b32 %r1832, %r1830, %r1830, 4; xor.b32 %r1833, %r1832, %r1831; add.s32 %r1834, %r1830, %r1829; sub.s32 %r1835, %r1829, %r1833; shf.l.wrap.b32 %r1836, %r1833, %r1833, 6; xor.b32 %r1837, %r1836, %r1835; add.s32 %r1838, %r1833, %r1834; sub.s32 %r1839, %r1834, %r1837; shf.l.wrap.b32 %r1840, %r1837, %r1837, 8; xor.b32 %r1841, %r1840, %r1839; add.s32 %r1842, %r1837, %r1838; sub.s32 %r1843, %r1838, %r1841; shf.l.wrap.b32 %r1844, %r1841, %r1841, 16; xor.b32 %r1845, %r1844, %r1843; add.s32 %r1846, %r1841, %r1842; sub.s32 %r1847, %r1842, %r1845; shf.l.wrap.b32 %r1848, %r1845, %r1845, 19; xor.b32 %r1849, %r1848, %r1847; add.s32 %r1850, %r1845, %r1846; sub.s32 %r1851, %r1846, %r1849; shf.l.wrap.b32 %r1852, %r1849, %r1849, 4; add.s32 %r1853, %r1849, %r1850; add.s32 %r1854, %r1850, %r1828; xor.b32 %r1855, %r1853, %r1851; xor.b32 %r1856, %r1855, %r1852; shf.l.wrap.b32 %r1857, %r1853, %r1853, 14; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1854; shf.l.wrap.b32 %r1860, %r1858, %r1858, 11; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1853; shf.l.wrap.b32 %r1863, %r1861, %r1861, 25; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 16; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 4; sub.s32 %r1870, %r1868, %r1869; xor.b32 %r1871, %r1870, %r1864; shf.l.wrap.b32 %r1872, %r1870, %r1870, 14; sub.s32 %r1873, %r1871, %r1872; xor.b32 %r1874, %r1873, %r1867; shf.l.wrap.b32 %r1875, %r1873, %r1873, 24; sub.s32 %r1876, %r1874, %r1875; and.b32 %r1877, %r1876, 31; setp.lt.u32 %p315, %r1877, 24; selp.f32 %f951, %f921, %f924, %p315; setp.lt.u32 %p316, %r1877, 16; selp.f32 %f952, %f924, %f927, %p316; setp.lt.u32 %p317, %r1877, 8; selp.f32 %f953, %f927, %f930, %p317; and.b32 %r1878, %r1876, 1; setp.eq.b32 %p318, %r1878, 1; neg.ftz.f32 %f954, %f951; selp.f32 %f955, %f954, %f951, %p318; and.b32 %r1879, %r1876, 2; setp.eq.s32 %p319, %r1879, 0; neg.ftz.f32 %f956, %f952; selp.f32 %f957, %f952, %f956, %p319; add.ftz.f32 %f958, %f955, %f957; and.b32 %r1880, %r1876, 4; setp.eq.s32 %p320, %r1880, 0; neg.ftz.f32 %f959, %f953; selp.f32 %f960, %f953, %f959, %p320; add.ftz.f32 %f961, %f960, %f958; add.s32 %r1881, %r1825, 1; sub.s32 %r1882, %r1881, %r1827; xor.b32 %r1883, %r1832, %r1882; sub.s32 %r1884, %r1829, %r1883; shf.l.wrap.b32 %r1885, %r1883, %r1883, 6; xor.b32 %r1886, %r1885, %r1884; add.s32 %r1887, %r1883, %r1834; sub.s32 %r1888, %r1834, %r1886; shf.l.wrap.b32 %r1889, %r1886, %r1886, 8; xor.b32 %r1890, %r1889, %r1888; add.s32 %r1891, %r1886, %r1887; sub.s32 %r1892, %r1887, %r1890; shf.l.wrap.b32 %r1893, %r1890, %r1890, 16; xor.b32 %r1894, %r1893, %r1892; add.s32 %r1895, %r1890, %r1891; sub.s32 %r1896, %r1891, %r1894; shf.l.wrap.b32 %r1897, %r1894, %r1894, 19; xor.b32 %r1898, %r1897, %r1896; add.s32 %r1899, %r1894, %r1895; sub.s32 %r1900, %r1895, %r1898; shf.l.wrap.b32 %r1901, %r1898, %r1898, 4; add.s32 %r1902, %r1898, %r1899; add.s32 %r1903, %r1899, %r1828; xor.b32 %r1904, %r1902, %r1900; xor.b32 %r1905, %r1904, %r1901; shf.l.wrap.b32 %r1906, %r1902, %r1902, 14; sub.s32 %r1907, %r1905, %r1906; xor.b32 %r1908, %r1907, %r1903; shf.l.wrap.b32 %r1909, %r1907, %r1907, 11; sub.s32 %r1910, %r1908, %r1909; xor.b32 %r1911, %r1910, %r1902; shf.l.wrap.b32 %r1912, %r1910, %r1910, 25; sub.s32 %r1913, %r1911, %r1912; xor.b32 %r1914, %r1913, %r1907; shf.l.wrap.b32 %r1915, %r1913, %r1913, 16; sub.s32 %r1916, %r1914, %r1915; xor.b32 %r1917, %r1916, %r1910; shf.l.wrap.b32 %r1918, %r1916, %r1916, 4; sub.s32 %r1919, %r1917, %r1918; xor.b32 %r1920, %r1919, %r1913; shf.l.wrap.b32 %r1921, %r1919, %r1919, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1916; shf.l.wrap.b32 %r1924, %r1922, %r1922, 24; sub.s32 %r1925, %r1923, %r1924; add.ftz.f32 %f962, %f921, 0fBF800000; and.b32 %r1926, %r1925, 31; setp.lt.u32 %p321, %r1926, 24; selp.f32 %f964, %f962, %f924, %p321; setp.lt.u32 %p322, %r1926, 16; selp.f32 %f965, %f924, %f927, %p322; setp.lt.u32 %p323, %r1926, 8; selp.f32 %f966, %f927, %f930, %p323; and.b32 %r1927, %r1925, 1; setp.eq.b32 %p324, %r1927, 1; neg.ftz.f32 %f967, %f964; selp.f32 %f968, %f967, %f964, %p324; and.b32 %r1928, %r1925, 2; setp.eq.s32 %p325, %r1928, 0; neg.ftz.f32 %f969, %f965; selp.f32 %f970, %f965, %f969, %p325; add.ftz.f32 %f971, %f968, %f970; and.b32 %r1929, %r1925, 4; setp.eq.s32 %p326, %r1929, 0; neg.ftz.f32 %f972, %f966; selp.f32 %f973, %f966, %f972, %p326; add.ftz.f32 %f974, %f973, %f971; add.s32 %r1930, %r1826, -559038707; add.s32 %r1931, %r1830, %r1930; sub.s32 %r1932, %r1930, %r1833; xor.b32 %r1933, %r1836, %r1932; add.s32 %r1934, %r1833, %r1931; sub.s32 %r1935, %r1931, %r1933; shf.l.wrap.b32 %r1936, %r1933, %r1933, 8; xor.b32 %r1937, %r1936, %r1935; add.s32 %r1938, %r1933, %r1934; sub.s32 %r1939, %r1934, %r1937; shf.l.wrap.b32 %r1940, %r1937, %r1937, 16; xor.b32 %r1941, %r1940, %r1939; add.s32 %r1942, %r1937, %r1938; sub.s32 %r1943, %r1938, %r1941; shf.l.wrap.b32 %r1944, %r1941, %r1941, 19; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1941, %r1942; sub.s32 %r1947, %r1942, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 4; add.s32 %r1949, %r1945, %r1946; add.s32 %r1950, %r1946, %r1828; xor.b32 %r1951, %r1949, %r1947; xor.b32 %r1952, %r1951, %r1948; shf.l.wrap.b32 %r1953, %r1949, %r1949, 14; sub.s32 %r1954, %r1952, %r1953; xor.b32 %r1955, %r1954, %r1950; shf.l.wrap.b32 %r1956, %r1954, %r1954, 11; sub.s32 %r1957, %r1955, %r1956; xor.b32 %r1958, %r1957, %r1949; shf.l.wrap.b32 %r1959, %r1957, %r1957, 25; sub.s32 %r1960, %r1958, %r1959; xor.b32 %r1961, %r1960, %r1954; shf.l.wrap.b32 %r1962, %r1960, %r1960, 16; sub.s32 %r1963, %r1961, %r1962; xor.b32 %r1964, %r1963, %r1957; shf.l.wrap.b32 %r1965, %r1963, %r1963, 4; sub.s32 %r1966, %r1964, %r1965; xor.b32 %r1967, %r1966, %r1960; shf.l.wrap.b32 %r1968, %r1966, %r1966, 14; sub.s32 %r1969, %r1967, %r1968; xor.b32 %r1970, %r1969, %r1963; shf.l.wrap.b32 %r1971, %r1969, %r1969, 24; sub.s32 %r1972, %r1970, %r1971; add.ftz.f32 %f975, %f924, 0fBF800000; and.b32 %r1973, %r1972, 31; setp.lt.u32 %p327, %r1973, 24; selp.f32 %f976, %f921, %f975, %p327; setp.lt.u32 %p328, %r1973, 16; selp.f32 %f977, %f975, %f927, %p328; setp.lt.u32 %p329, %r1973, 8; selp.f32 %f978, %f927, %f930, %p329; and.b32 %r1974, %r1972, 1; setp.eq.b32 %p330, %r1974, 1; neg.ftz.f32 %f979, %f976; selp.f32 %f980, %f979, %f976, %p330; and.b32 %r1975, %r1972, 2; setp.eq.s32 %p331, %r1975, 0; neg.ftz.f32 %f981, %f977; selp.f32 %f982, %f977, %f981, %p331; add.ftz.f32 %f983, %f980, %f982; and.b32 %r1976, %r1972, 4; setp.eq.s32 %p332, %r1976, 0; neg.ftz.f32 %f984, %f978; selp.f32 %f985, %f978, %f984, %p332; add.ftz.f32 %f986, %f985, %f983; sub.s32 %r1977, %r1930, %r1883; xor.b32 %r1978, %r1885, %r1977; add.s32 %r1979, %r1883, %r1931; sub.s32 %r1980, %r1931, %r1978; shf.l.wrap.b32 %r1981, %r1978, %r1978, 8; xor.b32 %r1982, %r1981, %r1980; add.s32 %r1983, %r1978, %r1979; sub.s32 %r1984, %r1979, %r1982; shf.l.wrap.b32 %r1985, %r1982, %r1982, 16; xor.b32 %r1986, %r1985, %r1984; add.s32 %r1987, %r1982, %r1983; sub.s32 %r1988, %r1983, %r1986; shf.l.wrap.b32 %r1989, %r1986, %r1986, 19; xor.b32 %r1990, %r1989, %r1988; add.s32 %r1991, %r1986, %r1987; sub.s32 %r1992, %r1987, %r1990; shf.l.wrap.b32 %r1993, %r1990, %r1990, 4; add.s32 %r1994, %r1990, %r1991; add.s32 %r1995, %r1991, %r1828; xor.b32 %r1996, %r1994, %r1992; xor.b32 %r1997, %r1996, %r1993; shf.l.wrap.b32 %r1998, %r1994, %r1994, 14; sub.s32 %r1999, %r1997, %r1998; xor.b32 %r2000, %r1999, %r1995; shf.l.wrap.b32 %r2001, %r1999, %r1999, 11; sub.s32 %r2002, %r2000, %r2001; xor.b32 %r2003, %r2002, %r1994; shf.l.wrap.b32 %r2004, %r2002, %r2002, 25; sub.s32 %r2005, %r2003, %r2004; xor.b32 %r2006, %r2005, %r1999; shf.l.wrap.b32 %r2007, %r2005, %r2005, 16; sub.s32 %r2008, %r2006, %r2007; xor.b32 %r2009, %r2008, %r2002; shf.l.wrap.b32 %r2010, %r2008, %r2008, 4; sub.s32 %r2011, %r2009, %r2010; xor.b32 %r2012, %r2011, %r2005; shf.l.wrap.b32 %r2013, %r2011, %r2011, 14; sub.s32 %r2014, %r2012, %r2013; xor.b32 %r2015, %r2014, %r2008; shf.l.wrap.b32 %r2016, %r2014, %r2014, 24; sub.s32 %r2017, %r2015, %r2016; and.b32 %r2018, %r2017, 31; setp.lt.u32 %p333, %r2018, 24; selp.f32 %f987, %f962, %f975, %p333; setp.lt.u32 %p334, %r2018, 16; selp.f32 %f988, %f975, %f927, %p334; setp.lt.u32 %p335, %r2018, 8; selp.f32 %f989, %f927, %f930, %p335; and.b32 %r2019, %r2017, 1; setp.eq.b32 %p336, %r2019, 1; neg.ftz.f32 %f990, %f987; selp.f32 %f991, %f990, %f987, %p336; and.b32 %r2020, %r2017, 2; setp.eq.s32 %p337, %r2020, 0; neg.ftz.f32 %f992, %f988; selp.f32 %f993, %f988, %f992, %p337; add.ftz.f32 %f994, %f991, %f993; and.b32 %r2021, %r2017, 4; setp.eq.s32 %p338, %r2021, 0; neg.ftz.f32 %f995, %f989; selp.f32 %f996, %f989, %f995, %p338; add.ftz.f32 %f997, %f996, %f994; not.b32 %r2022, %r1827; add.s32 %r2023, %r1827, -559038707; add.s32 %r2024, %r1825, %r2022; shf.l.wrap.b32 %r2025, %r2023, %r2023, 4; xor.b32 %r2026, %r2025, %r2024; add.s32 %r2027, %r2023, %r1829; sub.s32 %r2028, %r1829, %r2026; shf.l.wrap.b32 %r2029, %r2026, %r2026, 6; xor.b32 %r2030, %r2029, %r2028; add.s32 %r2031, %r2026, %r2027; sub.s32 %r2032, %r2027, %r2030; shf.l.wrap.b32 %r2033, %r2030, %r2030, 8; xor.b32 %r2034, %r2033, %r2032; add.s32 %r2035, %r2030, %r2031; sub.s32 %r2036, %r2031, %r2034; shf.l.wrap.b32 %r2037, %r2034, %r2034, 16; xor.b32 %r2038, %r2037, %r2036; add.s32 %r2039, %r2034, %r2035; sub.s32 %r2040, %r2035, %r2038; shf.l.wrap.b32 %r2041, %r2038, %r2038, 19; xor.b32 %r2042, %r2041, %r2040; add.s32 %r2043, %r2038, %r2039; sub.s32 %r2044, %r2039, %r2042; shf.l.wrap.b32 %r2045, %r2042, %r2042, 4; add.s32 %r2046, %r2042, %r2043; add.s32 %r2047, %r2043, %r1828; xor.b32 %r2048, %r2046, %r2044; xor.b32 %r2049, %r2048, %r2045; shf.l.wrap.b32 %r2050, %r2046, %r2046, 14; sub.s32 %r2051, %r2049, %r2050; xor.b32 %r2052, %r2051, %r2047; shf.l.wrap.b32 %r2053, %r2051, %r2051, 11; sub.s32 %r2054, %r2052, %r2053; xor.b32 %r2055, %r2054, %r2046; shf.l.wrap.b32 %r2056, %r2054, %r2054, 25; sub.s32 %r2057, %r2055, %r2056; xor.b32 %r2058, %r2057, %r2051; shf.l.wrap.b32 %r2059, %r2057, %r2057, 16; sub.s32 %r2060, %r2058, %r2059; xor.b32 %r2061, %r2060, %r2054; shf.l.wrap.b32 %r2062, %r2060, %r2060, 4; sub.s32 %r2063, %r2061, %r2062; xor.b32 %r2064, %r2063, %r2057; shf.l.wrap.b32 %r2065, %r2063, %r2063, 14; sub.s32 %r2066, %r2064, %r2065; xor.b32 %r2067, %r2066, %r2060; shf.l.wrap.b32 %r2068, %r2066, %r2066, 24; sub.s32 %r2069, %r2067, %r2068; add.ftz.f32 %f998, %f927, 0fBF800000; and.b32 %r2070, %r2069, 31; setp.lt.u32 %p339, %r2070, 24; selp.f32 %f999, %f921, %f924, %p339; setp.lt.u32 %p340, %r2070, 16; selp.f32 %f1000, %f924, %f998, %p340; setp.lt.u32 %p341, %r2070, 8; selp.f32 %f1001, %f998, %f930, %p341; and.b32 %r2071, %r2069, 1; setp.eq.b32 %p342, %r2071, 1; neg.ftz.f32 %f1002, %f999; selp.f32 %f1003, %f1002, %f999, %p342; and.b32 %r2072, %r2069, 2; setp.eq.s32 %p343, %r2072, 0; neg.ftz.f32 %f1004, %f1000; selp.f32 %f1005, %f1000, %f1004, %p343; add.ftz.f32 %f1006, %f1003, %f1005; and.b32 %r2073, %r2069, 4; setp.eq.s32 %p344, %r2073, 0; neg.ftz.f32 %f1007, %f1001; selp.f32 %f1008, %f1001, %f1007, %p344; add.ftz.f32 %f1009, %f1008, %f1006; xor.b32 %r2074, %r2025, %r1831; sub.s32 %r2075, %r1829, %r2074; shf.l.wrap.b32 %r2076, %r2074, %r2074, 6; xor.b32 %r2077, %r2076, %r2075; add.s32 %r2078, %r2074, %r2027; sub.s32 %r2079, %r2027, %r2077; shf.l.wrap.b32 %r2080, %r2077, %r2077, 8; xor.b32 %r2081, %r2080, %r2079; add.s32 %r2082, %r2077, %r2078; sub.s32 %r2083, %r2078, %r2081; shf.l.wrap.b32 %r2084, %r2081, %r2081, 16; xor.b32 %r2085, %r2084, %r2083; add.s32 %r2086, %r2081, %r2082; sub.s32 %r2087, %r2082, %r2085; shf.l.wrap.b32 %r2088, %r2085, %r2085, 19; xor.b32 %r2089, %r2088, %r2087; add.s32 %r2090, %r2085, %r2086; sub.s32 %r2091, %r2086, %r2089; shf.l.wrap.b32 %r2092, %r2089, %r2089, 4; add.s32 %r2093, %r2089, %r2090; add.s32 %r2094, %r2090, %r1828; xor.b32 %r2095, %r2093, %r2091; xor.b32 %r2096, %r2095, %r2092; shf.l.wrap.b32 %r2097, %r2093, %r2093, 14; sub.s32 %r2098, %r2096, %r2097; xor.b32 %r2099, %r2098, %r2094; shf.l.wrap.b32 %r2100, %r2098, %r2098, 11; sub.s32 %r2101, %r2099, %r2100; xor.b32 %r2102, %r2101, %r2093; shf.l.wrap.b32 %r2103, %r2101, %r2101, 25; sub.s32 %r2104, %r2102, %r2103; xor.b32 %r2105, %r2104, %r2098; shf.l.wrap.b32 %r2106, %r2104, %r2104, 16; sub.s32 %r2107, %r2105, %r2106; xor.b32 %r2108, %r2107, %r2101; shf.l.wrap.b32 %r2109, %r2107, %r2107, 4; sub.s32 %r2110, %r2108, %r2109; xor.b32 %r2111, %r2110, %r2104; shf.l.wrap.b32 %r2112, %r2110, %r2110, 14; sub.s32 %r2113, %r2111, %r2112; xor.b32 %r2114, %r2113, %r2107; shf.l.wrap.b32 %r2115, %r2113, %r2113, 24; sub.s32 %r2116, %r2114, %r2115; and.b32 %r2117, %r2116, 31; setp.lt.u32 %p345, %r2117, 24; selp.f32 %f1010, %f962, %f924, %p345; setp.lt.u32 %p346, %r2117, 16; selp.f32 %f1011, %f924, %f998, %p346; setp.lt.u32 %p347, %r2117, 8; selp.f32 %f1012, %f998, %f930, %p347; and.b32 %r2118, %r2116, 1; setp.eq.b32 %p348, %r2118, 1; neg.ftz.f32 %f1013, %f1010; selp.f32 %f1014, %f1013, %f1010, %p348; and.b32 %r2119, %r2116, 2; setp.eq.s32 %p349, %r2119, 0; neg.ftz.f32 %f1015, %f1011; selp.f32 %f1016, %f1011, %f1015, %p349; add.ftz.f32 %f1017, %f1014, %f1016; and.b32 %r2120, %r2116, 4; setp.eq.s32 %p350, %r2120, 0; neg.ftz.f32 %f1018, %f1012; selp.f32 %f1019, %f1012, %f1018, %p350; add.ftz.f32 %f1020, %f1019, %f1017; add.s32 %r2121, %r2023, %r1930; sub.s32 %r2122, %r1930, %r2026; xor.b32 %r2123, %r2029, %r2122; add.s32 %r2124, %r2026, %r2121; sub.s32 %r2125, %r2121, %r2123; shf.l.wrap.b32 %r2126, %r2123, %r2123, 8; xor.b32 %r2127, %r2126, %r2125; add.s32 %r2128, %r2123, %r2124; sub.s32 %r2129, %r2124, %r2127; shf.l.wrap.b32 %r2130, %r2127, %r2127, 16; xor.b32 %r2131, %r2130, %r2129; add.s32 %r2132, %r2127, %r2128; sub.s32 %r2133, %r2128, %r2131; shf.l.wrap.b32 %r2134, %r2131, %r2131, 19; xor.b32 %r2135, %r2134, %r2133; add.s32 %r2136, %r2131, %r2132; sub.s32 %r2137, %r2132, %r2135; shf.l.wrap.b32 %r2138, %r2135, %r2135, 4; add.s32 %r2139, %r2135, %r2136; add.s32 %r2140, %r2136, %r1828; xor.b32 %r2141, %r2139, %r2137; xor.b32 %r2142, %r2141, %r2138; shf.l.wrap.b32 %r2143, %r2139, %r2139, 14; sub.s32 %r2144, %r2142, %r2143; xor.b32 %r2145, %r2144, %r2140; shf.l.wrap.b32 %r2146, %r2144, %r2144, 11; sub.s32 %r2147, %r2145, %r2146; xor.b32 %r2148, %r2147, %r2139; shf.l.wrap.b32 %r2149, %r2147, %r2147, 25; sub.s32 %r2150, %r2148, %r2149; xor.b32 %r2151, %r2150, %r2144; shf.l.wrap.b32 %r2152, %r2150, %r2150, 16; sub.s32 %r2153, %r2151, %r2152; xor.b32 %r2154, %r2153, %r2147; shf.l.wrap.b32 %r2155, %r2153, %r2153, 4; sub.s32 %r2156, %r2154, %r2155; xor.b32 %r2157, %r2156, %r2150; shf.l.wrap.b32 %r2158, %r2156, %r2156, 14; sub.s32 %r2159, %r2157, %r2158; xor.b32 %r2160, %r2159, %r2153; shf.l.wrap.b32 %r2161, %r2159, %r2159, 24; sub.s32 %r2162, %r2160, %r2161; and.b32 %r2163, %r2162, 31; setp.lt.u32 %p351, %r2163, 24; selp.f32 %f1021, %f921, %f975, %p351; setp.lt.u32 %p352, %r2163, 16; selp.f32 %f1022, %f975, %f998, %p352; setp.lt.u32 %p353, %r2163, 8; selp.f32 %f1023, %f998, %f930, %p353; and.b32 %r2164, %r2162, 1; setp.eq.b32 %p354, %r2164, 1; neg.ftz.f32 %f1024, %f1021; selp.f32 %f1025, %f1024, %f1021, %p354; and.b32 %r2165, %r2162, 2; setp.eq.s32 %p355, %r2165, 0; neg.ftz.f32 %f1026, %f1022; selp.f32 %f1027, %f1022, %f1026, %p355; add.ftz.f32 %f1028, %f1025, %f1027; and.b32 %r2166, %r2162, 4; setp.eq.s32 %p356, %r2166, 0; neg.ftz.f32 %f1029, %f1023; selp.f32 %f1030, %f1023, %f1029, %p356; add.ftz.f32 %f1031, %f1030, %f1028; sub.s32 %r2167, %r1930, %r2074; xor.b32 %r2168, %r2076, %r2167; add.s32 %r2169, %r2074, %r2121; sub.s32 %r2170, %r2121, %r2168; shf.l.wrap.b32 %r2171, %r2168, %r2168, 8; xor.b32 %r2172, %r2171, %r2170; add.s32 %r2173, %r2168, %r2169; sub.s32 %r2174, %r2169, %r2172; shf.l.wrap.b32 %r2175, %r2172, %r2172, 16; xor.b32 %r2176, %r2175, %r2174; add.s32 %r2177, %r2172, %r2173; sub.s32 %r2178, %r2173, %r2176; shf.l.wrap.b32 %r2179, %r2176, %r2176, 19; xor.b32 %r2180, %r2179, %r2178; add.s32 %r2181, %r2176, %r2177; sub.s32 %r2182, %r2177, %r2180; shf.l.wrap.b32 %r2183, %r2180, %r2180, 4; add.s32 %r2184, %r2180, %r2181; add.s32 %r2185, %r2181, %r1828; xor.b32 %r2186, %r2184, %r2182; xor.b32 %r2187, %r2186, %r2183; shf.l.wrap.b32 %r2188, %r2184, %r2184, 14; sub.s32 %r2189, %r2187, %r2188; xor.b32 %r2190, %r2189, %r2185; shf.l.wrap.b32 %r2191, %r2189, %r2189, 11; sub.s32 %r2192, %r2190, %r2191; xor.b32 %r2193, %r2192, %r2184; shf.l.wrap.b32 %r2194, %r2192, %r2192, 25; sub.s32 %r2195, %r2193, %r2194; xor.b32 %r2196, %r2195, %r2189; shf.l.wrap.b32 %r2197, %r2195, %r2195, 16; sub.s32 %r2198, %r2196, %r2197; xor.b32 %r2199, %r2198, %r2192; shf.l.wrap.b32 %r2200, %r2198, %r2198, 4; sub.s32 %r2201, %r2199, %r2200; xor.b32 %r2202, %r2201, %r2195; shf.l.wrap.b32 %r2203, %r2201, %r2201, 14; sub.s32 %r2204, %r2202, %r2203; xor.b32 %r2205, %r2204, %r2198; shf.l.wrap.b32 %r2206, %r2204, %r2204, 24; sub.s32 %r2207, %r2205, %r2206; and.b32 %r2208, %r2207, 31; setp.lt.u32 %p357, %r2208, 24; selp.f32 %f1032, %f962, %f975, %p357; setp.lt.u32 %p358, %r2208, 16; selp.f32 %f1033, %f975, %f998, %p358; setp.lt.u32 %p359, %r2208, 8; selp.f32 %f1034, %f998, %f930, %p359; and.b32 %r2209, %r2207, 1; setp.eq.b32 %p360, %r2209, 1; neg.ftz.f32 %f1035, %f1032; selp.f32 %f1036, %f1035, %f1032, %p360; and.b32 %r2210, %r2207, 2; setp.eq.s32 %p361, %r2210, 0; neg.ftz.f32 %f1037, %f1033; selp.f32 %f1038, %f1033, %f1037, %p361; add.ftz.f32 %f1039, %f1036, %f1038; and.b32 %r2211, %r2207, 4; setp.eq.s32 %p362, %r2211, 0; neg.ftz.f32 %f1040, %f1034; selp.f32 %f1041, %f1034, %f1040, %p362; add.ftz.f32 %f1042, %f1041, %f1039; add.s32 %r2212, %r1828, 1; add.s32 %r2213, %r1850, %r2212; xor.b32 %r2214, %r1858, %r2213; sub.s32 %r2215, %r2214, %r1860; xor.b32 %r2216, %r2215, %r1853; shf.l.wrap.b32 %r2217, %r2215, %r2215, 25; sub.s32 %r2218, %r2216, %r2217; xor.b32 %r2219, %r2218, %r1858; shf.l.wrap.b32 %r2220, %r2218, %r2218, 16; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r2215; shf.l.wrap.b32 %r2223, %r2221, %r2221, 4; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 14; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 24; sub.s32 %r2230, %r2228, %r2229; add.ftz.f32 %f1043, %f930, 0fBF800000; and.b32 %r2231, %r2230, 31; setp.lt.u32 %p363, %r2231, 24; selp.f32 %f1044, %f921, %f924, %p363; setp.lt.u32 %p364, %r2231, 16; selp.f32 %f1045, %f924, %f927, %p364; setp.lt.u32 %p365, %r2231, 8; selp.f32 %f1046, %f927, %f1043, %p365; and.b32 %r2232, %r2230, 1; setp.eq.b32 %p366, %r2232, 1; neg.ftz.f32 %f1047, %f1044; selp.f32 %f1048, %f1047, %f1044, %p366; and.b32 %r2233, %r2230, 2; setp.eq.s32 %p367, %r2233, 0; neg.ftz.f32 %f1049, %f1045; selp.f32 %f1050, %f1045, %f1049, %p367; add.ftz.f32 %f1051, %f1048, %f1050; and.b32 %r2234, %r2230, 4; setp.eq.s32 %p368, %r2234, 0; neg.ftz.f32 %f1052, %f1046; selp.f32 %f1053, %f1046, %f1052, %p368; add.ftz.f32 %f1054, %f1053, %f1051; add.s32 %r2235, %r1899, %r2212; xor.b32 %r2236, %r1907, %r2235; sub.s32 %r2237, %r2236, %r1909; xor.b32 %r2238, %r2237, %r1902; shf.l.wrap.b32 %r2239, %r2237, %r2237, 25; sub.s32 %r2240, %r2238, %r2239; xor.b32 %r2241, %r2240, %r1907; shf.l.wrap.b32 %r2242, %r2240, %r2240, 16; sub.s32 %r2243, %r2241, %r2242; xor.b32 %r2244, %r2243, %r2237; shf.l.wrap.b32 %r2245, %r2243, %r2243, 4; sub.s32 %r2246, %r2244, %r2245; xor.b32 %r2247, %r2246, %r2240; shf.l.wrap.b32 %r2248, %r2246, %r2246, 14; sub.s32 %r2249, %r2247, %r2248; xor.b32 %r2250, %r2249, %r2243; shf.l.wrap.b32 %r2251, %r2249, %r2249, 24; sub.s32 %r2252, %r2250, %r2251; and.b32 %r2253, %r2252, 31; setp.lt.u32 %p369, %r2253, 24; selp.f32 %f1055, %f962, %f924, %p369; setp.lt.u32 %p370, %r2253, 16; selp.f32 %f1056, %f924, %f927, %p370; setp.lt.u32 %p371, %r2253, 8; selp.f32 %f1057, %f927, %f1043, %p371; and.b32 %r2254, %r2252, 1; setp.eq.b32 %p372, %r2254, 1; neg.ftz.f32 %f1058, %f1055; selp.f32 %f1059, %f1058, %f1055, %p372; and.b32 %r2255, %r2252, 2; setp.eq.s32 %p373, %r2255, 0; neg.ftz.f32 %f1060, %f1056; selp.f32 %f1061, %f1056, %f1060, %p373; add.ftz.f32 %f1062, %f1059, %f1061; and.b32 %r2256, %r2252, 4; setp.eq.s32 %p374, %r2256, 0; neg.ftz.f32 %f1063, %f1057; selp.f32 %f1064, %f1057, %f1063, %p374; add.ftz.f32 %f1065, %f1064, %f1062; add.s32 %r2257, %r1946, %r2212; xor.b32 %r2258, %r1954, %r2257; sub.s32 %r2259, %r2258, %r1956; xor.b32 %r2260, %r2259, %r1949; shf.l.wrap.b32 %r2261, %r2259, %r2259, 25; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r1954; shf.l.wrap.b32 %r2264, %r2262, %r2262, 16; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 4; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 14; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 24; sub.s32 %r2274, %r2272, %r2273; and.b32 %r2275, %r2274, 31; setp.lt.u32 %p375, %r2275, 24; selp.f32 %f1066, %f921, %f975, %p375; setp.lt.u32 %p376, %r2275, 16; selp.f32 %f1067, %f975, %f927, %p376; setp.lt.u32 %p377, %r2275, 8; selp.f32 %f1068, %f927, %f1043, %p377; and.b32 %r2276, %r2274, 1; setp.eq.b32 %p378, %r2276, 1; neg.ftz.f32 %f1069, %f1066; selp.f32 %f1070, %f1069, %f1066, %p378; and.b32 %r2277, %r2274, 2; setp.eq.s32 %p379, %r2277, 0; neg.ftz.f32 %f1071, %f1067; selp.f32 %f1072, %f1067, %f1071, %p379; add.ftz.f32 %f1073, %f1070, %f1072; and.b32 %r2278, %r2274, 4; setp.eq.s32 %p380, %r2278, 0; neg.ftz.f32 %f1074, %f1068; selp.f32 %f1075, %f1068, %f1074, %p380; add.ftz.f32 %f1076, %f1075, %f1073; add.s32 %r2279, %r1991, %r2212; xor.b32 %r2280, %r1999, %r2279; sub.s32 %r2281, %r2280, %r2001; xor.b32 %r2282, %r2281, %r1994; shf.l.wrap.b32 %r2283, %r2281, %r2281, 25; sub.s32 %r2284, %r2282, %r2283; xor.b32 %r2285, %r2284, %r1999; shf.l.wrap.b32 %r2286, %r2284, %r2284, 16; sub.s32 %r2287, %r2285, %r2286; xor.b32 %r2288, %r2287, %r2281; shf.l.wrap.b32 %r2289, %r2287, %r2287, 4; sub.s32 %r2290, %r2288, %r2289; xor.b32 %r2291, %r2290, %r2284; shf.l.wrap.b32 %r2292, %r2290, %r2290, 14; sub.s32 %r2293, %r2291, %r2292; xor.b32 %r2294, %r2293, %r2287; shf.l.wrap.b32 %r2295, %r2293, %r2293, 24; sub.s32 %r2296, %r2294, %r2295; and.b32 %r2297, %r2296, 31; setp.lt.u32 %p381, %r2297, 24; selp.f32 %f1077, %f962, %f975, %p381; setp.lt.u32 %p382, %r2297, 16; selp.f32 %f1078, %f975, %f927, %p382; setp.lt.u32 %p383, %r2297, 8; selp.f32 %f1079, %f927, %f1043, %p383; and.b32 %r2298, %r2296, 1; setp.eq.b32 %p384, %r2298, 1; neg.ftz.f32 %f1080, %f1077; selp.f32 %f1081, %f1080, %f1077, %p384; and.b32 %r2299, %r2296, 2; setp.eq.s32 %p385, %r2299, 0; neg.ftz.f32 %f1082, %f1078; selp.f32 %f1083, %f1078, %f1082, %p385; add.ftz.f32 %f1084, %f1081, %f1083; and.b32 %r2300, %r2296, 4; setp.eq.s32 %p386, %r2300, 0; neg.ftz.f32 %f1085, %f1079; selp.f32 %f1086, %f1079, %f1085, %p386; add.ftz.f32 %f1087, %f1086, %f1084; add.s32 %r2301, %r2043, %r2212; xor.b32 %r2302, %r2051, %r2301; sub.s32 %r2303, %r2302, %r2053; xor.b32 %r2304, %r2303, %r2046; shf.l.wrap.b32 %r2305, %r2303, %r2303, 25; sub.s32 %r2306, %r2304, %r2305; xor.b32 %r2307, %r2306, %r2051; shf.l.wrap.b32 %r2308, %r2306, %r2306, 16; sub.s32 %r2309, %r2307, %r2308; xor.b32 %r2310, %r2309, %r2303; shf.l.wrap.b32 %r2311, %r2309, %r2309, 4; sub.s32 %r2312, %r2310, %r2311; xor.b32 %r2313, %r2312, %r2306; shf.l.wrap.b32 %r2314, %r2312, %r2312, 14; sub.s32 %r2315, %r2313, %r2314; xor.b32 %r2316, %r2315, %r2309; shf.l.wrap.b32 %r2317, %r2315, %r2315, 24; sub.s32 %r2318, %r2316, %r2317; and.b32 %r2319, %r2318, 31; setp.lt.u32 %p387, %r2319, 24; selp.f32 %f1088, %f921, %f924, %p387; setp.lt.u32 %p388, %r2319, 16; selp.f32 %f1089, %f924, %f998, %p388; setp.lt.u32 %p389, %r2319, 8; selp.f32 %f1090, %f998, %f1043, %p389; and.b32 %r2320, %r2318, 1; setp.eq.b32 %p390, %r2320, 1; neg.ftz.f32 %f1091, %f1088; selp.f32 %f1092, %f1091, %f1088, %p390; and.b32 %r2321, %r2318, 2; setp.eq.s32 %p391, %r2321, 0; neg.ftz.f32 %f1093, %f1089; selp.f32 %f1094, %f1089, %f1093, %p391; add.ftz.f32 %f1095, %f1092, %f1094; and.b32 %r2322, %r2318, 4; setp.eq.s32 %p392, %r2322, 0; neg.ftz.f32 %f1096, %f1090; selp.f32 %f1097, %f1090, %f1096, %p392; add.ftz.f32 %f1098, %f1097, %f1095; add.s32 %r2323, %r2090, %r2212; xor.b32 %r2324, %r2098, %r2323; sub.s32 %r2325, %r2324, %r2100; xor.b32 %r2326, %r2325, %r2093; shf.l.wrap.b32 %r2327, %r2325, %r2325, 25; sub.s32 %r2328, %r2326, %r2327; xor.b32 %r2329, %r2328, %r2098; shf.l.wrap.b32 %r2330, %r2328, %r2328, 16; sub.s32 %r2331, %r2329, %r2330; xor.b32 %r2332, %r2331, %r2325; shf.l.wrap.b32 %r2333, %r2331, %r2331, 4; sub.s32 %r2334, %r2332, %r2333; xor.b32 %r2335, %r2334, %r2328; shf.l.wrap.b32 %r2336, %r2334, %r2334, 14; sub.s32 %r2337, %r2335, %r2336; xor.b32 %r2338, %r2337, %r2331; shf.l.wrap.b32 %r2339, %r2337, %r2337, 24; sub.s32 %r2340, %r2338, %r2339; and.b32 %r2341, %r2340, 31; setp.lt.u32 %p393, %r2341, 24; selp.f32 %f1099, %f962, %f924, %p393; setp.lt.u32 %p394, %r2341, 16; selp.f32 %f1100, %f924, %f998, %p394; setp.lt.u32 %p395, %r2341, 8; selp.f32 %f1101, %f998, %f1043, %p395; and.b32 %r2342, %r2340, 1; setp.eq.b32 %p396, %r2342, 1; neg.ftz.f32 %f1102, %f1099; selp.f32 %f1103, %f1102, %f1099, %p396; and.b32 %r2343, %r2340, 2; setp.eq.s32 %p397, %r2343, 0; neg.ftz.f32 %f1104, %f1100; selp.f32 %f1105, %f1100, %f1104, %p397; add.ftz.f32 %f1106, %f1103, %f1105; and.b32 %r2344, %r2340, 4; setp.eq.s32 %p398, %r2344, 0; neg.ftz.f32 %f1107, %f1101; selp.f32 %f1108, %f1101, %f1107, %p398; add.ftz.f32 %f1109, %f1108, %f1106; add.s32 %r2345, %r2136, %r2212; xor.b32 %r2346, %r2144, %r2345; sub.s32 %r2347, %r2346, %r2146; xor.b32 %r2348, %r2347, %r2139; shf.l.wrap.b32 %r2349, %r2347, %r2347, 25; sub.s32 %r2350, %r2348, %r2349; xor.b32 %r2351, %r2350, %r2144; shf.l.wrap.b32 %r2352, %r2350, %r2350, 16; sub.s32 %r2353, %r2351, %r2352; xor.b32 %r2354, %r2353, %r2347; shf.l.wrap.b32 %r2355, %r2353, %r2353, 4; sub.s32 %r2356, %r2354, %r2355; xor.b32 %r2357, %r2356, %r2350; shf.l.wrap.b32 %r2358, %r2356, %r2356, 14; sub.s32 %r2359, %r2357, %r2358; xor.b32 %r2360, %r2359, %r2353; shf.l.wrap.b32 %r2361, %r2359, %r2359, 24; sub.s32 %r2362, %r2360, %r2361; and.b32 %r2363, %r2362, 31; setp.lt.u32 %p399, %r2363, 24; selp.f32 %f1110, %f921, %f975, %p399; setp.lt.u32 %p400, %r2363, 16; selp.f32 %f1111, %f975, %f998, %p400; setp.lt.u32 %p401, %r2363, 8; selp.f32 %f1112, %f998, %f1043, %p401; and.b32 %r2364, %r2362, 1; setp.eq.b32 %p402, %r2364, 1; neg.ftz.f32 %f1113, %f1110; selp.f32 %f1114, %f1113, %f1110, %p402; and.b32 %r2365, %r2362, 2; setp.eq.s32 %p403, %r2365, 0; neg.ftz.f32 %f1115, %f1111; selp.f32 %f1116, %f1111, %f1115, %p403; add.ftz.f32 %f1117, %f1114, %f1116; and.b32 %r2366, %r2362, 4; setp.eq.s32 %p404, %r2366, 0; neg.ftz.f32 %f1118, %f1112; selp.f32 %f1119, %f1112, %f1118, %p404; add.ftz.f32 %f1120, %f1119, %f1117; add.s32 %r2367, %r2181, %r2212; xor.b32 %r2368, %r2189, %r2367; sub.s32 %r2369, %r2368, %r2191; xor.b32 %r2370, %r2369, %r2184; shf.l.wrap.b32 %r2371, %r2369, %r2369, 25; sub.s32 %r2372, %r2370, %r2371; xor.b32 %r2373, %r2372, %r2189; shf.l.wrap.b32 %r2374, %r2372, %r2372, 16; sub.s32 %r2375, %r2373, %r2374; xor.b32 %r2376, %r2375, %r2369; shf.l.wrap.b32 %r2377, %r2375, %r2375, 4; sub.s32 %r2378, %r2376, %r2377; xor.b32 %r2379, %r2378, %r2372; shf.l.wrap.b32 %r2380, %r2378, %r2378, 14; sub.s32 %r2381, %r2379, %r2380; xor.b32 %r2382, %r2381, %r2375; shf.l.wrap.b32 %r2383, %r2381, %r2381, 24; sub.s32 %r2384, %r2382, %r2383; and.b32 %r2385, %r2384, 31; setp.lt.u32 %p405, %r2385, 24; selp.f32 %f1121, %f962, %f975, %p405; setp.lt.u32 %p406, %r2385, 16; selp.f32 %f1122, %f975, %f998, %p406; setp.lt.u32 %p407, %r2385, 8; selp.f32 %f1123, %f998, %f1043, %p407; and.b32 %r2386, %r2384, 1; setp.eq.b32 %p408, %r2386, 1; neg.ftz.f32 %f1124, %f1121; selp.f32 %f1125, %f1124, %f1121, %p408; and.b32 %r2387, %r2384, 2; setp.eq.s32 %p409, %r2387, 0; neg.ftz.f32 %f1126, %f1122; selp.f32 %f1127, %f1122, %f1126, %p409; add.ftz.f32 %f1128, %f1125, %f1127; and.b32 %r2388, %r2384, 4; setp.eq.s32 %p410, %r2388, 0; neg.ftz.f32 %f1129, %f1123; selp.f32 %f1130, %f1123, %f1129, %p410; add.ftz.f32 %f1131, %f1130, %f1128; sub.ftz.f32 %f1132, %f216, %f935; sub.ftz.f32 %f1133, %f216, %f940; sub.ftz.f32 %f1134, %f216, %f945; mul.ftz.f32 %f1135, %f1132, %f961; fma.rn.ftz.f32 %f1136, %f935, %f974, %f1135; mul.ftz.f32 %f1137, %f1132, %f986; fma.rn.ftz.f32 %f1138, %f935, %f997, %f1137; mul.ftz.f32 %f1139, %f940, %f1138; fma.rn.ftz.f32 %f1140, %f1133, %f1136, %f1139; mul.ftz.f32 %f1141, %f1132, %f1009; fma.rn.ftz.f32 %f1142, %f935, %f1020, %f1141; mul.ftz.f32 %f1143, %f1132, %f1031; fma.rn.ftz.f32 %f1144, %f935, %f1042, %f1143; mul.ftz.f32 %f1145, %f940, %f1144; fma.rn.ftz.f32 %f1146, %f1133, %f1142, %f1145; mul.ftz.f32 %f1147, %f945, %f1146; fma.rn.ftz.f32 %f1148, %f1134, %f1140, %f1147; mul.ftz.f32 %f1149, %f1132, %f1054; fma.rn.ftz.f32 %f1150, %f935, %f1065, %f1149; mul.ftz.f32 %f1151, %f1132, %f1076; fma.rn.ftz.f32 %f1152, %f935, %f1087, %f1151; mul.ftz.f32 %f1153, %f940, %f1152; fma.rn.ftz.f32 %f1154, %f1133, %f1150, %f1153; mul.ftz.f32 %f1155, %f1132, %f1098; fma.rn.ftz.f32 %f1156, %f935, %f1109, %f1155; mul.ftz.f32 %f1157, %f1132, %f1120; fma.rn.ftz.f32 %f1158, %f935, %f1131, %f1157; mul.ftz.f32 %f1159, %f940, %f1158; fma.rn.ftz.f32 %f1160, %f1133, %f1156, %f1159; mul.ftz.f32 %f1161, %f945, %f1160; fma.rn.ftz.f32 %f1162, %f1134, %f1154, %f1161; sub.ftz.f32 %f1163, %f1162, %f1148; fma.rn.ftz.f32 %f32, %f950, %f1163, %f1148; mov.b32 %r15, %f32; setp.nan.ftz.f32 %p411, %f32, %f32; @%p411 bra $L__BB79_42; or.b32 %r2389, %r15, -2147483648; setp.ne.s32 %p412, %r2389, -2147483648; add.ftz.f32 %f1165, %f32, %f32; setp.eq.ftz.f32 %p413, %f1165, %f32; and.pred %p414, %p413, %p412; @%p414 bra $L__BB79_42; shl.b32 %r2390, %r15, 1; setp.lt.u32 %p415, %r2390, -16777215; mul.ftz.f32 %f1166, %f32, 0f3F559B3D; selp.f32 %f1600, %f1166, 0f00000000, %p415; $L__BB79_42: fma.rn.ftz.f32 %f1601, %f18, %f1597, %f1601; fma.rn.ftz.f32 %f1602, %f18, %f1598, %f1602; fma.rn.ftz.f32 %f1603, %f18, %f1599, %f1603; fma.rn.ftz.f32 %f1604, %f18, %f1600, %f1604; $L__BB79_43: { // callseq 48, 0 .reg .b32 temp_param_reg; .param .align 16 .b8 param0[16]; st.param.v4.f32 [param0+0], {%f1601, %f1602, %f1603, %f1604}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_4d6float4ff, ( param0, param1, param2 ); ld.param.f32 %f1644, [retval0+0]; } // callseq 48 setp.eq.s32 %p416, %r7, 255; @%p416 bra $L__BB79_113; add.ftz.f32 %f1168, %f1604, 0f43342F8C; add.ftz.f32 %f1169, %f1603, 0f430A23D9; add.ftz.f32 %f1170, %f1602, 0f43312796; add.ftz.f32 %f1171, %f1601, 0f431B94B7; { // callseq 49, 0 .reg .b32 temp_param_reg; .param .align 16 .b8 param0[16]; st.param.v4.f32 [param0+0], {%f1171, %f1170, %f1169, %f1168}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_4d6float4ff, ( param0, param1, param2 ); ld.param.f32 %f1643, [retval0+0]; } // callseq 49 add.ftz.f32 %f1172, %f1604, 0f433944A6; add.ftz.f32 %f1173, %f1603, 0f43220D5A; add.ftz.f32 %f1174, %f1602, 0f4304AD7D; add.ftz.f32 %f1175, %f1601, 0f433CDA92; { // callseq 50, 0 .reg .b32 temp_param_reg; .param .align 16 .b8 param0[16]; st.param.v4.f32 [param0+0], {%f1175, %f1174, %f1173, %f1172}; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_4d6float4ff, ( param0, param1, param2 ); ld.param.f32 %f1642, [retval0+0]; } // callseq 50 bra.uni $L__BB79_113; $L__BB79_106: setp.eq.ftz.f32 %p590, %f18, 0f00000000; @%p590 bra $L__BB79_111; add.ftz.f32 %f1568, %f1604, 0f431E3E1A; cvt.rmi.ftz.f32.f32 %f1569, %f1568; cvt.rzi.ftz.s32.f32 %r3063, %f1569; sub.ftz.f32 %f1570, %f1568, %f1569; mul.ftz.f32 %f1571, %f1570, %f1570; mul.ftz.f32 %f1572, %f1570, %f1571; fma.rn.ftz.f32 %f1573, %f1570, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f1574, %f1570, %f1573, 0f41200000; mul.ftz.f32 %f1575, %f1572, %f1574; add.s32 %r3064, %r3063, -559038720; xor.b32 %r3065, %r3064, -1874868139; add.s32 %r3066, %r3065, 29184895; xor.b32 %r3067, %r3066, -559038720; shf.l.wrap.b32 %r3068, %r3066, %r3066, 25; sub.s32 %r3069, %r3067, %r3068; xor.b32 %r3070, %r3069, -1874868139; shf.l.wrap.b32 %r3071, %r3069, %r3069, 16; sub.s32 %r3072, %r3070, %r3071; xor.b32 %r3073, %r3072, %r3066; shf.l.wrap.b32 %r3074, %r3072, %r3072, 4; sub.s32 %r3075, %r3073, %r3074; xor.b32 %r3076, %r3075, %r3069; shf.l.wrap.b32 %r3077, %r3075, %r3075, 14; sub.s32 %r3078, %r3076, %r3077; xor.b32 %r3079, %r3078, %r3072; shf.l.wrap.b32 %r3080, %r3078, %r3078, 24; sub.s32 %r3081, %r3079, %r3080; and.b32 %r3082, %r3081, 7; add.s32 %r3083, %r3082, 1; cvt.rn.f32.s32 %f1576, %r3083; and.b32 %r3084, %r3081, 8; setp.eq.s32 %p591, %r3084, 0; neg.ftz.f32 %f1577, %f1576; selp.f32 %f1578, %f1576, %f1577, %p591; mul.ftz.f32 %f1579, %f1570, %f1578; add.s32 %r3085, %r3063, -559038719; xor.b32 %r3086, %r3085, -1874868139; add.s32 %r3087, %r3086, 29184895; xor.b32 %r3088, %r3087, -559038720; shf.l.wrap.b32 %r3089, %r3087, %r3087, 25; sub.s32 %r3090, %r3088, %r3089; xor.b32 %r3091, %r3090, -1874868139; shf.l.wrap.b32 %r3092, %r3090, %r3090, 16; sub.s32 %r3093, %r3091, %r3092; xor.b32 %r3094, %r3093, %r3087; shf.l.wrap.b32 %r3095, %r3093, %r3093, 4; sub.s32 %r3096, %r3094, %r3095; xor.b32 %r3097, %r3096, %r3090; shf.l.wrap.b32 %r3098, %r3096, %r3096, 14; sub.s32 %r3099, %r3097, %r3098; xor.b32 %r3100, %r3099, %r3093; shf.l.wrap.b32 %r3101, %r3099, %r3099, 24; sub.s32 %r3102, %r3100, %r3101; add.ftz.f32 %f1580, %f1570, 0fBF800000; and.b32 %r3103, %r3102, 7; add.s32 %r3104, %r3103, 1; cvt.rn.f32.s32 %f1581, %r3104; and.b32 %r3105, %r3102, 8; setp.eq.s32 %p592, %r3105, 0; neg.ftz.f32 %f1582, %f1581; selp.f32 %f1583, %f1581, %f1582, %p592; mul.ftz.f32 %f1584, %f1580, %f1583; sub.ftz.f32 %f1585, %f1584, %f1579; fma.rn.ftz.f32 %f158, %f1575, %f1585, %f1579; mov.b32 %r117, %f158; setp.nan.ftz.f32 %p593, %f158, %f158; mov.f32 %f1639, 0f00000000; @%p593 bra $L__BB79_110; or.b32 %r3106, %r117, -2147483648; setp.ne.s32 %p594, %r3106, -2147483648; add.ftz.f32 %f1587, %f158, %f158; setp.eq.ftz.f32 %p595, %f1587, %f158; and.pred %p596, %p595, %p594; @%p596 bra $L__BB79_110; shl.b32 %r3107, %r117, 1; setp.lt.u32 %p597, %r3107, -16777215; mul.ftz.f32 %f1588, %f158, 0f3E800000; selp.f32 %f1639, %f1588, 0f00000000, %p597; $L__BB79_110: fma.rn.ftz.f32 %f1604, %f18, %f1639, %f1604; $L__BB79_111: { // callseq 57, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1604; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_1dfff, ( param0, param1, param2 ); ld.param.f32 %f1644, [retval0+0]; } // callseq 57 setp.eq.s32 %p598, %r7, 255; @%p598 bra $L__BB79_113; add.ftz.f32 %f1590, %f1604, 0f42E5AC0F; { // callseq 58, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1590; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_1dfff, ( param0, param1, param2 ); ld.param.f32 %f1643, [retval0+0]; } // callseq 58 add.ftz.f32 %f1591, %f1604, 0f433C7D1B; { // callseq 59, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1591; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_1dfff, ( param0, param1, param2 ); ld.param.f32 %f1642, [retval0+0]; } // callseq 59 bra.uni $L__BB79_113; $L__BB79_45: setp.eq.ftz.f32 %p417, %f18, 0f00000000; @%p417 bra $L__BB79_104; add.ftz.f32 %f1176, %f1601, 0f433A0802; cvt.rmi.ftz.f32.f32 %f1177, %f1176; cvt.rzi.ftz.s32.f32 %r16, %f1177; sub.ftz.f32 %f46, %f1176, %f1177; add.ftz.f32 %f1178, %f1602, 0f42E5E973; cvt.rmi.ftz.f32.f32 %f1179, %f1178; cvt.rzi.ftz.s32.f32 %r17, %f1179; sub.ftz.f32 %f47, %f1178, %f1179; add.ftz.f32 %f1180, %f1603, 0f431A7290; cvt.rmi.ftz.f32.f32 %f1181, %f1180; cvt.rzi.ftz.s32.f32 %r18, %f1181; sub.ftz.f32 %f48, %f1180, %f1181; fma.rn.ftz.f32 %f1182, %f46, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f49, %f46, %f1182, 0f41200000; fma.rn.ftz.f32 %f1183, %f47, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f50, %f47, %f1183, 0f41200000; fma.rn.ftz.f32 %f1184, %f48, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f51, %f48, %f1184, 0f41200000; add.s32 %r2391, %r17, -559038712; add.s32 %r2392, %r18, -559038712; xor.b32 %r2393, %r2392, %r2391; shf.l.wrap.b32 %r19, %r2391, %r2391, 14; sub.s32 %r20, %r2393, %r19; add.s32 %r2394, %r16, -559038712; xor.b32 %r2395, %r20, %r2394; shf.l.wrap.b32 %r21, %r20, %r20, 11; sub.s32 %r2396, %r2395, %r21; xor.b32 %r2397, %r2396, %r2391; shf.l.wrap.b32 %r2398, %r2396, %r2396, 25; sub.s32 %r2399, %r2397, %r2398; xor.b32 %r2400, %r2399, %r20; shf.l.wrap.b32 %r2401, %r2399, %r2399, 16; sub.s32 %r2402, %r2400, %r2401; xor.b32 %r2403, %r2402, %r2396; shf.l.wrap.b32 %r2404, %r2402, %r2402, 4; sub.s32 %r2405, %r2403, %r2404; xor.b32 %r2406, %r2405, %r2399; shf.l.wrap.b32 %r2407, %r2405, %r2405, 14; sub.s32 %r2408, %r2406, %r2407; xor.b32 %r2409, %r2408, %r2402; shf.l.wrap.b32 %r2410, %r2408, %r2408, 24; sub.s32 %r22, %r2409, %r2410; and.b32 %r23, %r22, 15; cvt.u16.u32 %rs1, %r22; or.b16 %rs2, %rs1, 2; and.b16 %rs3, %rs2, 15; setp.eq.s16 %p418, %rs3, 14; mov.f32 %f1605, %f46; @%p418 bra $L__BB79_48; mov.f32 %f1605, %f48; $L__BB79_48: mul.ftz.f32 %f1185, %f46, %f46; mul.ftz.f32 %f1186, %f46, %f1185; mul.ftz.f32 %f53, %f1186, %f49; mul.ftz.f32 %f1187, %f47, %f47; mul.ftz.f32 %f1188, %f47, %f1187; mul.ftz.f32 %f54, %f1188, %f50; mul.ftz.f32 %f1189, %f48, %f48; mul.ftz.f32 %f1190, %f48, %f1189; mul.ftz.f32 %f55, %f1190, %f51; setp.lt.u32 %p419, %r23, 8; selp.f32 %f1191, %f46, %f47, %p419; setp.lt.u32 %p420, %r23, 4; selp.f32 %f1192, %f47, %f1605, %p420; and.b32 %r2411, %r22, 1; setp.eq.b32 %p421, %r2411, 1; neg.ftz.f32 %f1193, %f1191; selp.f32 %f1194, %f1193, %f1191, %p421; and.b32 %r2412, %r22, 2; setp.eq.s32 %p422, %r2412, 0; neg.ftz.f32 %f1195, %f1192; selp.f32 %f1196, %f1192, %f1195, %p422; add.ftz.f32 %f56, %f1194, %f1196; add.s32 %r24, %r16, -559038711; xor.b32 %r2413, %r20, %r24; sub.s32 %r2414, %r2413, %r21; xor.b32 %r2416, %r2414, %r2391; shf.l.wrap.b32 %r2417, %r2414, %r2414, 25; sub.s32 %r2418, %r2416, %r2417; xor.b32 %r2419, %r2418, %r20; shf.l.wrap.b32 %r2420, %r2418, %r2418, 16; sub.s32 %r2421, %r2419, %r2420; xor.b32 %r2422, %r2421, %r2414; shf.l.wrap.b32 %r2423, %r2421, %r2421, 4; sub.s32 %r2424, %r2422, %r2423; xor.b32 %r2425, %r2424, %r2418; shf.l.wrap.b32 %r2426, %r2424, %r2424, 14; sub.s32 %r2427, %r2425, %r2426; xor.b32 %r2428, %r2427, %r2421; shf.l.wrap.b32 %r2429, %r2427, %r2427, 24; sub.s32 %r25, %r2428, %r2429; and.b32 %r26, %r25, 15; add.ftz.f32 %f57, %f46, 0fBF800000; cvt.u16.u32 %rs4, %r25; or.b16 %rs5, %rs4, 2; and.b16 %rs6, %rs5, 15; setp.eq.s16 %p423, %rs6, 14; mov.f32 %f1606, %f57; @%p423 bra $L__BB79_50; mov.f32 %f1606, %f48; $L__BB79_50: setp.lt.u32 %p424, %r26, 8; selp.f32 %f1197, %f57, %f47, %p424; setp.lt.u32 %p425, %r26, 4; selp.f32 %f1198, %f47, %f1606, %p425; and.b32 %r2430, %r25, 1; setp.eq.b32 %p426, %r2430, 1; neg.ftz.f32 %f1199, %f1197; selp.f32 %f1200, %f1199, %f1197, %p426; and.b32 %r2431, %r25, 2; setp.eq.s32 %p427, %r2431, 0; neg.ftz.f32 %f1201, %f1198; selp.f32 %f1202, %f1198, %f1201, %p427; add.ftz.f32 %f59, %f1200, %f1202; add.s32 %r27, %r17, -559038711; xor.b32 %r2433, %r2392, %r27; shf.l.wrap.b32 %r28, %r27, %r27, 14; sub.s32 %r29, %r2433, %r28; xor.b32 %r2435, %r29, %r2394; shf.l.wrap.b32 %r30, %r29, %r29, 11; sub.s32 %r2436, %r2435, %r30; xor.b32 %r2437, %r2436, %r27; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r29; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r31, %r2449, %r2450; and.b32 %r32, %r31, 15; add.ftz.f32 %f60, %f47, 0fBF800000; cvt.u16.u32 %rs7, %r31; or.b16 %rs8, %rs7, 2; and.b16 %rs9, %rs8, 15; setp.eq.s16 %p428, %rs9, 14; mov.f32 %f1607, %f46; @%p428 bra $L__BB79_52; mov.f32 %f1607, %f48; $L__BB79_52: setp.lt.u32 %p429, %r32, 8; selp.f32 %f1203, %f46, %f60, %p429; setp.lt.u32 %p430, %r32, 4; selp.f32 %f1204, %f60, %f1607, %p430; and.b32 %r2451, %r31, 1; setp.eq.b32 %p431, %r2451, 1; neg.ftz.f32 %f1205, %f1203; selp.f32 %f1206, %f1205, %f1203, %p431; and.b32 %r2452, %r31, 2; setp.eq.s32 %p432, %r2452, 0; neg.ftz.f32 %f1207, %f1204; selp.f32 %f1208, %f1204, %f1207, %p432; add.ftz.f32 %f62, %f1206, %f1208; xor.b32 %r2453, %r29, %r24; sub.s32 %r2454, %r2453, %r30; xor.b32 %r2455, %r2454, %r27; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r29; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r33, %r2467, %r2468; and.b32 %r34, %r33, 15; cvt.u16.u32 %rs10, %r33; or.b16 %rs11, %rs10, 2; and.b16 %rs12, %rs11, 15; setp.eq.s16 %p433, %rs12, 14; mov.f32 %f1608, %f57; @%p433 bra $L__BB79_54; mov.f32 %f1608, %f48; $L__BB79_54: setp.lt.u32 %p434, %r34, 8; selp.f32 %f1209, %f57, %f60, %p434; setp.lt.u32 %p435, %r34, 4; selp.f32 %f1210, %f60, %f1608, %p435; and.b32 %r2469, %r33, 1; setp.eq.b32 %p436, %r2469, 1; neg.ftz.f32 %f1211, %f1209; selp.f32 %f1212, %f1211, %f1209, %p436; and.b32 %r2470, %r33, 2; setp.eq.s32 %p437, %r2470, 0; neg.ftz.f32 %f1213, %f1210; selp.f32 %f1214, %f1210, %f1213, %p437; add.ftz.f32 %f64, %f1212, %f1214; add.s32 %r35, %r18, -559038711; xor.b32 %r2472, %r35, %r2391; sub.s32 %r36, %r2472, %r19; xor.b32 %r2474, %r36, %r2394; shf.l.wrap.b32 %r37, %r36, %r36, 11; sub.s32 %r2475, %r2474, %r37; xor.b32 %r2476, %r2475, %r2391; shf.l.wrap.b32 %r2477, %r2475, %r2475, 25; sub.s32 %r2478, %r2476, %r2477; xor.b32 %r2479, %r2478, %r36; shf.l.wrap.b32 %r2480, %r2478, %r2478, 16; sub.s32 %r2481, %r2479, %r2480; xor.b32 %r2482, %r2481, %r2475; shf.l.wrap.b32 %r2483, %r2481, %r2481, 4; sub.s32 %r2484, %r2482, %r2483; xor.b32 %r2485, %r2484, %r2478; shf.l.wrap.b32 %r2486, %r2484, %r2484, 14; sub.s32 %r2487, %r2485, %r2486; xor.b32 %r2488, %r2487, %r2481; shf.l.wrap.b32 %r2489, %r2487, %r2487, 24; sub.s32 %r38, %r2488, %r2489; add.ftz.f32 %f65, %f48, 0fBF800000; and.b32 %r39, %r38, 15; cvt.u16.u32 %rs13, %r38; or.b16 %rs14, %rs13, 2; and.b16 %rs15, %rs14, 15; setp.eq.s16 %p438, %rs15, 14; mov.f32 %f1609, %f46; @%p438 bra $L__BB79_56; mov.f32 %f1609, %f65; $L__BB79_56: setp.lt.u32 %p439, %r39, 8; selp.f32 %f1215, %f46, %f47, %p439; setp.lt.u32 %p440, %r39, 4; selp.f32 %f1216, %f47, %f1609, %p440; and.b32 %r2490, %r38, 1; setp.eq.b32 %p441, %r2490, 1; neg.ftz.f32 %f1217, %f1215; selp.f32 %f1218, %f1217, %f1215, %p441; and.b32 %r2491, %r38, 2; setp.eq.s32 %p442, %r2491, 0; neg.ftz.f32 %f1219, %f1216; selp.f32 %f1220, %f1216, %f1219, %p442; add.ftz.f32 %f67, %f1218, %f1220; xor.b32 %r2492, %r36, %r24; sub.s32 %r2493, %r2492, %r37; xor.b32 %r2495, %r2493, %r2391; shf.l.wrap.b32 %r2496, %r2493, %r2493, 25; sub.s32 %r2497, %r2495, %r2496; xor.b32 %r2498, %r2497, %r36; shf.l.wrap.b32 %r2499, %r2497, %r2497, 16; sub.s32 %r2500, %r2498, %r2499; xor.b32 %r2501, %r2500, %r2493; shf.l.wrap.b32 %r2502, %r2500, %r2500, 4; sub.s32 %r2503, %r2501, %r2502; xor.b32 %r2504, %r2503, %r2497; shf.l.wrap.b32 %r2505, %r2503, %r2503, 14; sub.s32 %r2506, %r2504, %r2505; xor.b32 %r2507, %r2506, %r2500; shf.l.wrap.b32 %r2508, %r2506, %r2506, 24; sub.s32 %r40, %r2507, %r2508; and.b32 %r41, %r40, 15; cvt.u16.u32 %rs16, %r40; or.b16 %rs17, %rs16, 2; and.b16 %rs18, %rs17, 15; setp.eq.s16 %p443, %rs18, 14; mov.f32 %f1610, %f57; @%p443 bra $L__BB79_58; mov.f32 %f1610, %f65; $L__BB79_58: setp.lt.u32 %p444, %r41, 8; selp.f32 %f1221, %f57, %f47, %p444; setp.lt.u32 %p445, %r41, 4; selp.f32 %f1222, %f47, %f1610, %p445; and.b32 %r2509, %r40, 1; setp.eq.b32 %p446, %r2509, 1; neg.ftz.f32 %f1223, %f1221; selp.f32 %f1224, %f1223, %f1221, %p446; and.b32 %r2510, %r40, 2; setp.eq.s32 %p447, %r2510, 0; neg.ftz.f32 %f1225, %f1222; selp.f32 %f1226, %f1222, %f1225, %p447; add.ftz.f32 %f69, %f1224, %f1226; xor.b32 %r2511, %r35, %r27; sub.s32 %r42, %r2511, %r28; xor.b32 %r2513, %r42, %r2394; shf.l.wrap.b32 %r43, %r42, %r42, 11; sub.s32 %r2514, %r2513, %r43; xor.b32 %r2515, %r2514, %r27; shf.l.wrap.b32 %r2516, %r2514, %r2514, 25; sub.s32 %r2517, %r2515, %r2516; xor.b32 %r2518, %r2517, %r42; shf.l.wrap.b32 %r2519, %r2517, %r2517, 16; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2514; shf.l.wrap.b32 %r2522, %r2520, %r2520, 4; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2517; shf.l.wrap.b32 %r2525, %r2523, %r2523, 14; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 24; sub.s32 %r44, %r2527, %r2528; and.b32 %r45, %r44, 15; cvt.u16.u32 %rs19, %r44; or.b16 %rs20, %rs19, 2; and.b16 %rs21, %rs20, 15; setp.eq.s16 %p448, %rs21, 14; mov.f32 %f1611, %f46; @%p448 bra $L__BB79_60; mov.f32 %f1611, %f65; $L__BB79_60: setp.lt.u32 %p449, %r45, 8; selp.f32 %f1227, %f46, %f60, %p449; setp.lt.u32 %p450, %r45, 4; selp.f32 %f1228, %f60, %f1611, %p450; and.b32 %r2529, %r44, 1; setp.eq.b32 %p451, %r2529, 1; neg.ftz.f32 %f1229, %f1227; selp.f32 %f1230, %f1229, %f1227, %p451; and.b32 %r2530, %r44, 2; setp.eq.s32 %p452, %r2530, 0; neg.ftz.f32 %f1231, %f1228; selp.f32 %f1232, %f1228, %f1231, %p452; add.ftz.f32 %f71, %f1230, %f1232; xor.b32 %r2531, %r42, %r24; sub.s32 %r2532, %r2531, %r43; xor.b32 %r2533, %r2532, %r27; shf.l.wrap.b32 %r2534, %r2532, %r2532, 25; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r42; shf.l.wrap.b32 %r2537, %r2535, %r2535, 16; sub.s32 %r2538, %r2536, %r2537; xor.b32 %r2539, %r2538, %r2532; shf.l.wrap.b32 %r2540, %r2538, %r2538, 4; sub.s32 %r2541, %r2539, %r2540; xor.b32 %r2542, %r2541, %r2535; shf.l.wrap.b32 %r2543, %r2541, %r2541, 14; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2538; shf.l.wrap.b32 %r2546, %r2544, %r2544, 24; sub.s32 %r46, %r2545, %r2546; and.b32 %r47, %r46, 15; cvt.u16.u32 %rs22, %r46; or.b16 %rs23, %rs22, 2; and.b16 %rs24, %rs23, 15; setp.eq.s16 %p453, %rs24, 14; mov.f32 %f1612, %f57; @%p453 bra $L__BB79_62; mov.f32 %f1612, %f65; $L__BB79_62: setp.lt.u32 %p454, %r47, 8; selp.f32 %f1234, %f57, %f60, %p454; setp.lt.u32 %p455, %r47, 4; selp.f32 %f1235, %f60, %f1612, %p455; and.b32 %r2547, %r46, 1; setp.eq.b32 %p456, %r2547, 1; neg.ftz.f32 %f1236, %f1234; selp.f32 %f1237, %f1236, %f1234, %p456; and.b32 %r2548, %r46, 2; setp.eq.s32 %p457, %r2548, 0; neg.ftz.f32 %f1238, %f1235; selp.f32 %f1239, %f1235, %f1238, %p457; add.ftz.f32 %f1240, %f1237, %f1239; mov.f32 %f1241, 0f3F800000; sub.ftz.f32 %f1242, %f1241, %f53; mul.ftz.f32 %f1243, %f53, %f59; fma.rn.ftz.f32 %f1244, %f1242, %f56, %f1243; sub.ftz.f32 %f1245, %f1241, %f54; mul.ftz.f32 %f1246, %f53, %f64; fma.rn.ftz.f32 %f1247, %f1242, %f62, %f1246; mul.ftz.f32 %f1248, %f54, %f1247; fma.rn.ftz.f32 %f1249, %f1245, %f1244, %f1248; sub.ftz.f32 %f1250, %f1241, %f55; mul.ftz.f32 %f1251, %f53, %f69; fma.rn.ftz.f32 %f1252, %f1242, %f67, %f1251; mul.ftz.f32 %f1253, %f53, %f1240; fma.rn.ftz.f32 %f1254, %f1242, %f71, %f1253; mul.ftz.f32 %f1255, %f54, %f1254; fma.rn.ftz.f32 %f1256, %f1245, %f1252, %f1255; mul.ftz.f32 %f1257, %f55, %f1256; fma.rn.ftz.f32 %f73, %f1250, %f1249, %f1257; mov.b32 %r48, %f73; setp.nan.ftz.f32 %p458, %f73, %f73; mov.f32 %f1613, 0f00000000; @%p458 bra $L__BB79_65; or.b32 %r2549, %r48, -2147483648; setp.ne.s32 %p459, %r2549, -2147483648; add.ftz.f32 %f1259, %f73, %f73; setp.eq.ftz.f32 %p460, %f1259, %f73; and.pred %p461, %p460, %p459; @%p461 bra $L__BB79_65; shl.b32 %r2550, %r48, 1; setp.lt.u32 %p462, %r2550, -16777215; selp.f32 %f1613, %f73, 0f00000000, %p462; $L__BB79_65: add.ftz.f32 %f1260, %f1601, 0f4347D70A; cvt.rmi.ftz.f32.f32 %f1261, %f1260; cvt.rzi.ftz.s32.f32 %r49, %f1261; sub.ftz.f32 %f76, %f1260, %f1261; add.ftz.f32 %f1262, %f1602, 0f43224AE8; cvt.rmi.ftz.f32.f32 %f1263, %f1262; cvt.rzi.ftz.s32.f32 %r50, %f1263; sub.ftz.f32 %f77, %f1262, %f1263; add.ftz.f32 %f1264, %f1603, 0f431A0C59; cvt.rmi.ftz.f32.f32 %f1265, %f1264; cvt.rzi.ftz.s32.f32 %r51, %f1265; sub.ftz.f32 %f78, %f1264, %f1265; fma.rn.ftz.f32 %f1266, %f76, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f79, %f76, %f1266, 0f41200000; fma.rn.ftz.f32 %f1267, %f77, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f80, %f77, %f1267, 0f41200000; fma.rn.ftz.f32 %f1268, %f78, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f81, %f78, %f1268, 0f41200000; add.s32 %r2551, %r50, -559038712; add.s32 %r2552, %r51, -559038712; xor.b32 %r2553, %r2552, %r2551; shf.l.wrap.b32 %r52, %r2551, %r2551, 14; sub.s32 %r53, %r2553, %r52; add.s32 %r2554, %r49, -559038712; xor.b32 %r2555, %r53, %r2554; shf.l.wrap.b32 %r54, %r53, %r53, 11; sub.s32 %r2556, %r2555, %r54; xor.b32 %r2557, %r2556, %r2551; shf.l.wrap.b32 %r2558, %r2556, %r2556, 25; sub.s32 %r2559, %r2557, %r2558; xor.b32 %r2560, %r2559, %r53; shf.l.wrap.b32 %r2561, %r2559, %r2559, 16; sub.s32 %r2562, %r2560, %r2561; xor.b32 %r2563, %r2562, %r2556; shf.l.wrap.b32 %r2564, %r2562, %r2562, 4; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2559; shf.l.wrap.b32 %r2567, %r2565, %r2565, 14; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2562; shf.l.wrap.b32 %r2570, %r2568, %r2568, 24; sub.s32 %r55, %r2569, %r2570; and.b32 %r56, %r55, 15; cvt.u16.u32 %rs25, %r55; or.b16 %rs26, %rs25, 2; and.b16 %rs27, %rs26, 15; setp.eq.s16 %p463, %rs27, 14; mov.f32 %f1614, %f76; @%p463 bra $L__BB79_67; mov.f32 %f1614, %f78; $L__BB79_67: mul.ftz.f32 %f1269, %f76, %f76; mul.ftz.f32 %f1270, %f76, %f1269; mul.ftz.f32 %f83, %f1270, %f79; mul.ftz.f32 %f1271, %f77, %f77; mul.ftz.f32 %f1272, %f77, %f1271; mul.ftz.f32 %f84, %f1272, %f80; mul.ftz.f32 %f1273, %f78, %f78; mul.ftz.f32 %f1274, %f78, %f1273; mul.ftz.f32 %f85, %f1274, %f81; setp.lt.u32 %p464, %r56, 8; selp.f32 %f1275, %f76, %f77, %p464; setp.lt.u32 %p465, %r56, 4; selp.f32 %f1276, %f77, %f1614, %p465; and.b32 %r2571, %r55, 1; setp.eq.b32 %p466, %r2571, 1; neg.ftz.f32 %f1277, %f1275; selp.f32 %f1278, %f1277, %f1275, %p466; and.b32 %r2572, %r55, 2; setp.eq.s32 %p467, %r2572, 0; neg.ftz.f32 %f1279, %f1276; selp.f32 %f1280, %f1276, %f1279, %p467; add.ftz.f32 %f86, %f1278, %f1280; add.s32 %r57, %r49, -559038711; xor.b32 %r2573, %r53, %r57; sub.s32 %r2574, %r2573, %r54; xor.b32 %r2576, %r2574, %r2551; shf.l.wrap.b32 %r2577, %r2574, %r2574, 25; sub.s32 %r2578, %r2576, %r2577; xor.b32 %r2579, %r2578, %r53; shf.l.wrap.b32 %r2580, %r2578, %r2578, 16; sub.s32 %r2581, %r2579, %r2580; xor.b32 %r2582, %r2581, %r2574; shf.l.wrap.b32 %r2583, %r2581, %r2581, 4; sub.s32 %r2584, %r2582, %r2583; xor.b32 %r2585, %r2584, %r2578; shf.l.wrap.b32 %r2586, %r2584, %r2584, 14; sub.s32 %r2587, %r2585, %r2586; xor.b32 %r2588, %r2587, %r2581; shf.l.wrap.b32 %r2589, %r2587, %r2587, 24; sub.s32 %r58, %r2588, %r2589; and.b32 %r59, %r58, 15; add.ftz.f32 %f87, %f76, 0fBF800000; cvt.u16.u32 %rs28, %r58; or.b16 %rs29, %rs28, 2; and.b16 %rs30, %rs29, 15; setp.eq.s16 %p468, %rs30, 14; mov.f32 %f1615, %f87; @%p468 bra $L__BB79_69; mov.f32 %f1615, %f78; $L__BB79_69: setp.lt.u32 %p469, %r59, 8; selp.f32 %f1281, %f87, %f77, %p469; setp.lt.u32 %p470, %r59, 4; selp.f32 %f1282, %f77, %f1615, %p470; and.b32 %r2590, %r58, 1; setp.eq.b32 %p471, %r2590, 1; neg.ftz.f32 %f1283, %f1281; selp.f32 %f1284, %f1283, %f1281, %p471; and.b32 %r2591, %r58, 2; setp.eq.s32 %p472, %r2591, 0; neg.ftz.f32 %f1285, %f1282; selp.f32 %f1286, %f1282, %f1285, %p472; add.ftz.f32 %f89, %f1284, %f1286; add.s32 %r60, %r50, -559038711; xor.b32 %r2593, %r2552, %r60; shf.l.wrap.b32 %r61, %r60, %r60, 14; sub.s32 %r62, %r2593, %r61; xor.b32 %r2595, %r62, %r2554; shf.l.wrap.b32 %r63, %r62, %r62, 11; sub.s32 %r2596, %r2595, %r63; xor.b32 %r2597, %r2596, %r60; shf.l.wrap.b32 %r2598, %r2596, %r2596, 25; sub.s32 %r2599, %r2597, %r2598; xor.b32 %r2600, %r2599, %r62; shf.l.wrap.b32 %r2601, %r2599, %r2599, 16; sub.s32 %r2602, %r2600, %r2601; xor.b32 %r2603, %r2602, %r2596; shf.l.wrap.b32 %r2604, %r2602, %r2602, 4; sub.s32 %r2605, %r2603, %r2604; xor.b32 %r2606, %r2605, %r2599; shf.l.wrap.b32 %r2607, %r2605, %r2605, 14; sub.s32 %r2608, %r2606, %r2607; xor.b32 %r2609, %r2608, %r2602; shf.l.wrap.b32 %r2610, %r2608, %r2608, 24; sub.s32 %r64, %r2609, %r2610; and.b32 %r65, %r64, 15; add.ftz.f32 %f90, %f77, 0fBF800000; cvt.u16.u32 %rs31, %r64; or.b16 %rs32, %rs31, 2; and.b16 %rs33, %rs32, 15; setp.eq.s16 %p473, %rs33, 14; mov.f32 %f1616, %f76; @%p473 bra $L__BB79_71; mov.f32 %f1616, %f78; $L__BB79_71: setp.lt.u32 %p474, %r65, 8; selp.f32 %f1287, %f76, %f90, %p474; setp.lt.u32 %p475, %r65, 4; selp.f32 %f1288, %f90, %f1616, %p475; and.b32 %r2611, %r64, 1; setp.eq.b32 %p476, %r2611, 1; neg.ftz.f32 %f1289, %f1287; selp.f32 %f1290, %f1289, %f1287, %p476; and.b32 %r2612, %r64, 2; setp.eq.s32 %p477, %r2612, 0; neg.ftz.f32 %f1291, %f1288; selp.f32 %f1292, %f1288, %f1291, %p477; add.ftz.f32 %f92, %f1290, %f1292; xor.b32 %r2613, %r62, %r57; sub.s32 %r2614, %r2613, %r63; xor.b32 %r2615, %r2614, %r60; shf.l.wrap.b32 %r2616, %r2614, %r2614, 25; sub.s32 %r2617, %r2615, %r2616; xor.b32 %r2618, %r2617, %r62; shf.l.wrap.b32 %r2619, %r2617, %r2617, 16; sub.s32 %r2620, %r2618, %r2619; xor.b32 %r2621, %r2620, %r2614; shf.l.wrap.b32 %r2622, %r2620, %r2620, 4; sub.s32 %r2623, %r2621, %r2622; xor.b32 %r2624, %r2623, %r2617; shf.l.wrap.b32 %r2625, %r2623, %r2623, 14; sub.s32 %r2626, %r2624, %r2625; xor.b32 %r2627, %r2626, %r2620; shf.l.wrap.b32 %r2628, %r2626, %r2626, 24; sub.s32 %r66, %r2627, %r2628; and.b32 %r67, %r66, 15; cvt.u16.u32 %rs34, %r66; or.b16 %rs35, %rs34, 2; and.b16 %rs36, %rs35, 15; setp.eq.s16 %p478, %rs36, 14; mov.f32 %f1617, %f87; @%p478 bra $L__BB79_73; mov.f32 %f1617, %f78; $L__BB79_73: setp.lt.u32 %p479, %r67, 8; selp.f32 %f1293, %f87, %f90, %p479; setp.lt.u32 %p480, %r67, 4; selp.f32 %f1294, %f90, %f1617, %p480; and.b32 %r2629, %r66, 1; setp.eq.b32 %p481, %r2629, 1; neg.ftz.f32 %f1295, %f1293; selp.f32 %f1296, %f1295, %f1293, %p481; and.b32 %r2630, %r66, 2; setp.eq.s32 %p482, %r2630, 0; neg.ftz.f32 %f1297, %f1294; selp.f32 %f1298, %f1294, %f1297, %p482; add.ftz.f32 %f94, %f1296, %f1298; add.s32 %r68, %r51, -559038711; xor.b32 %r2632, %r68, %r2551; sub.s32 %r69, %r2632, %r52; xor.b32 %r2634, %r69, %r2554; shf.l.wrap.b32 %r70, %r69, %r69, 11; sub.s32 %r2635, %r2634, %r70; xor.b32 %r2636, %r2635, %r2551; shf.l.wrap.b32 %r2637, %r2635, %r2635, 25; sub.s32 %r2638, %r2636, %r2637; xor.b32 %r2639, %r2638, %r69; shf.l.wrap.b32 %r2640, %r2638, %r2638, 16; sub.s32 %r2641, %r2639, %r2640; xor.b32 %r2642, %r2641, %r2635; shf.l.wrap.b32 %r2643, %r2641, %r2641, 4; sub.s32 %r2644, %r2642, %r2643; xor.b32 %r2645, %r2644, %r2638; shf.l.wrap.b32 %r2646, %r2644, %r2644, 14; sub.s32 %r2647, %r2645, %r2646; xor.b32 %r2648, %r2647, %r2641; shf.l.wrap.b32 %r2649, %r2647, %r2647, 24; sub.s32 %r71, %r2648, %r2649; add.ftz.f32 %f95, %f78, 0fBF800000; and.b32 %r72, %r71, 15; cvt.u16.u32 %rs37, %r71; or.b16 %rs38, %rs37, 2; and.b16 %rs39, %rs38, 15; setp.eq.s16 %p483, %rs39, 14; mov.f32 %f1618, %f76; @%p483 bra $L__BB79_75; mov.f32 %f1618, %f95; $L__BB79_75: setp.lt.u32 %p484, %r72, 8; selp.f32 %f1299, %f76, %f77, %p484; setp.lt.u32 %p485, %r72, 4; selp.f32 %f1300, %f77, %f1618, %p485; and.b32 %r2650, %r71, 1; setp.eq.b32 %p486, %r2650, 1; neg.ftz.f32 %f1301, %f1299; selp.f32 %f1302, %f1301, %f1299, %p486; and.b32 %r2651, %r71, 2; setp.eq.s32 %p487, %r2651, 0; neg.ftz.f32 %f1303, %f1300; selp.f32 %f1304, %f1300, %f1303, %p487; add.ftz.f32 %f97, %f1302, %f1304; xor.b32 %r2652, %r69, %r57; sub.s32 %r2653, %r2652, %r70; xor.b32 %r2655, %r2653, %r2551; shf.l.wrap.b32 %r2656, %r2653, %r2653, 25; sub.s32 %r2657, %r2655, %r2656; xor.b32 %r2658, %r2657, %r69; shf.l.wrap.b32 %r2659, %r2657, %r2657, 16; sub.s32 %r2660, %r2658, %r2659; xor.b32 %r2661, %r2660, %r2653; shf.l.wrap.b32 %r2662, %r2660, %r2660, 4; sub.s32 %r2663, %r2661, %r2662; xor.b32 %r2664, %r2663, %r2657; shf.l.wrap.b32 %r2665, %r2663, %r2663, 14; sub.s32 %r2666, %r2664, %r2665; xor.b32 %r2667, %r2666, %r2660; shf.l.wrap.b32 %r2668, %r2666, %r2666, 24; sub.s32 %r73, %r2667, %r2668; and.b32 %r74, %r73, 15; cvt.u16.u32 %rs40, %r73; or.b16 %rs41, %rs40, 2; and.b16 %rs42, %rs41, 15; setp.eq.s16 %p488, %rs42, 14; mov.f32 %f1619, %f87; @%p488 bra $L__BB79_77; mov.f32 %f1619, %f95; $L__BB79_77: setp.lt.u32 %p489, %r74, 8; selp.f32 %f1305, %f87, %f77, %p489; setp.lt.u32 %p490, %r74, 4; selp.f32 %f1306, %f77, %f1619, %p490; and.b32 %r2669, %r73, 1; setp.eq.b32 %p491, %r2669, 1; neg.ftz.f32 %f1307, %f1305; selp.f32 %f1308, %f1307, %f1305, %p491; and.b32 %r2670, %r73, 2; setp.eq.s32 %p492, %r2670, 0; neg.ftz.f32 %f1309, %f1306; selp.f32 %f1310, %f1306, %f1309, %p492; add.ftz.f32 %f99, %f1308, %f1310; xor.b32 %r2671, %r68, %r60; sub.s32 %r75, %r2671, %r61; xor.b32 %r2673, %r75, %r2554; shf.l.wrap.b32 %r76, %r75, %r75, 11; sub.s32 %r2674, %r2673, %r76; xor.b32 %r2675, %r2674, %r60; shf.l.wrap.b32 %r2676, %r2674, %r2674, 25; sub.s32 %r2677, %r2675, %r2676; xor.b32 %r2678, %r2677, %r75; shf.l.wrap.b32 %r2679, %r2677, %r2677, 16; sub.s32 %r2680, %r2678, %r2679; xor.b32 %r2681, %r2680, %r2674; shf.l.wrap.b32 %r2682, %r2680, %r2680, 4; sub.s32 %r2683, %r2681, %r2682; xor.b32 %r2684, %r2683, %r2677; shf.l.wrap.b32 %r2685, %r2683, %r2683, 14; sub.s32 %r2686, %r2684, %r2685; xor.b32 %r2687, %r2686, %r2680; shf.l.wrap.b32 %r2688, %r2686, %r2686, 24; sub.s32 %r77, %r2687, %r2688; and.b32 %r78, %r77, 15; cvt.u16.u32 %rs43, %r77; or.b16 %rs44, %rs43, 2; and.b16 %rs45, %rs44, 15; setp.eq.s16 %p493, %rs45, 14; mov.f32 %f1620, %f76; @%p493 bra $L__BB79_79; mov.f32 %f1620, %f95; $L__BB79_79: setp.lt.u32 %p494, %r78, 8; selp.f32 %f1311, %f76, %f90, %p494; setp.lt.u32 %p495, %r78, 4; selp.f32 %f1312, %f90, %f1620, %p495; and.b32 %r2689, %r77, 1; setp.eq.b32 %p496, %r2689, 1; neg.ftz.f32 %f1313, %f1311; selp.f32 %f1314, %f1313, %f1311, %p496; and.b32 %r2690, %r77, 2; setp.eq.s32 %p497, %r2690, 0; neg.ftz.f32 %f1315, %f1312; selp.f32 %f1316, %f1312, %f1315, %p497; add.ftz.f32 %f101, %f1314, %f1316; xor.b32 %r2691, %r75, %r57; sub.s32 %r2692, %r2691, %r76; xor.b32 %r2693, %r2692, %r60; shf.l.wrap.b32 %r2694, %r2692, %r2692, 25; sub.s32 %r2695, %r2693, %r2694; xor.b32 %r2696, %r2695, %r75; shf.l.wrap.b32 %r2697, %r2695, %r2695, 16; sub.s32 %r2698, %r2696, %r2697; xor.b32 %r2699, %r2698, %r2692; shf.l.wrap.b32 %r2700, %r2698, %r2698, 4; sub.s32 %r2701, %r2699, %r2700; xor.b32 %r2702, %r2701, %r2695; shf.l.wrap.b32 %r2703, %r2701, %r2701, 14; sub.s32 %r2704, %r2702, %r2703; xor.b32 %r2705, %r2704, %r2698; shf.l.wrap.b32 %r2706, %r2704, %r2704, 24; sub.s32 %r79, %r2705, %r2706; and.b32 %r80, %r79, 15; cvt.u16.u32 %rs46, %r79; or.b16 %rs47, %rs46, 2; and.b16 %rs48, %rs47, 15; setp.eq.s16 %p498, %rs48, 14; mov.f32 %f1621, %f87; @%p498 bra $L__BB79_81; mov.f32 %f1621, %f95; $L__BB79_81: setp.lt.u32 %p499, %r80, 8; selp.f32 %f1318, %f87, %f90, %p499; setp.lt.u32 %p500, %r80, 4; selp.f32 %f1319, %f90, %f1621, %p500; and.b32 %r2707, %r79, 1; setp.eq.b32 %p501, %r2707, 1; neg.ftz.f32 %f1320, %f1318; selp.f32 %f1321, %f1320, %f1318, %p501; and.b32 %r2708, %r79, 2; setp.eq.s32 %p502, %r2708, 0; neg.ftz.f32 %f1322, %f1319; selp.f32 %f1323, %f1319, %f1322, %p502; add.ftz.f32 %f1324, %f1321, %f1323; sub.ftz.f32 %f1326, %f1241, %f83; mul.ftz.f32 %f1327, %f83, %f89; fma.rn.ftz.f32 %f1328, %f1326, %f86, %f1327; sub.ftz.f32 %f1329, %f1241, %f84; mul.ftz.f32 %f1330, %f83, %f94; fma.rn.ftz.f32 %f1331, %f1326, %f92, %f1330; mul.ftz.f32 %f1332, %f84, %f1331; fma.rn.ftz.f32 %f1333, %f1329, %f1328, %f1332; sub.ftz.f32 %f1334, %f1241, %f85; mul.ftz.f32 %f1335, %f83, %f99; fma.rn.ftz.f32 %f1336, %f1326, %f97, %f1335; mul.ftz.f32 %f1337, %f83, %f1324; fma.rn.ftz.f32 %f1338, %f1326, %f101, %f1337; mul.ftz.f32 %f1339, %f84, %f1338; fma.rn.ftz.f32 %f1340, %f1329, %f1336, %f1339; mul.ftz.f32 %f1341, %f85, %f1340; fma.rn.ftz.f32 %f103, %f1334, %f1333, %f1341; mov.b32 %r81, %f103; setp.nan.ftz.f32 %p503, %f103, %f103; mov.f32 %f1622, 0f00000000; @%p503 bra $L__BB79_84; or.b32 %r2709, %r81, -2147483648; setp.ne.s32 %p504, %r2709, -2147483648; add.ftz.f32 %f1343, %f103, %f103; setp.eq.ftz.f32 %p505, %f1343, %f103; and.pred %p506, %p505, %p504; @%p506 bra $L__BB79_84; shl.b32 %r2710, %r81, 1; setp.lt.u32 %p507, %r2710, -16777215; mul.ftz.f32 %f1344, %f103, 0f3F7B645A; selp.f32 %f1622, %f1344, 0f00000000, %p507; $L__BB79_84: add.ftz.f32 %f1345, %f1601, 0f42DF4487; cvt.rmi.ftz.f32.f32 %f1346, %f1345; cvt.rzi.ftz.s32.f32 %r82, %f1346; sub.ftz.f32 %f106, %f1345, %f1346; add.ftz.f32 %f1347, %f1602, 0f431D5E90; cvt.rmi.ftz.f32.f32 %f1348, %f1347; cvt.rzi.ftz.s32.f32 %r83, %f1348; sub.ftz.f32 %f107, %f1347, %f1348; add.ftz.f32 %f1349, %f1603, 0f4347168E; cvt.rmi.ftz.f32.f32 %f1350, %f1349; cvt.rzi.ftz.s32.f32 %r84, %f1350; sub.ftz.f32 %f108, %f1349, %f1350; fma.rn.ftz.f32 %f1351, %f106, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f109, %f106, %f1351, 0f41200000; fma.rn.ftz.f32 %f1352, %f107, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f110, %f107, %f1352, 0f41200000; fma.rn.ftz.f32 %f1353, %f108, 0f40C00000, 0fC1700000; fma.rn.ftz.f32 %f111, %f108, %f1353, 0f41200000; add.s32 %r2711, %r83, -559038712; add.s32 %r2712, %r84, -559038712; xor.b32 %r2713, %r2712, %r2711; shf.l.wrap.b32 %r85, %r2711, %r2711, 14; sub.s32 %r86, %r2713, %r85; add.s32 %r2714, %r82, -559038712; xor.b32 %r2715, %r86, %r2714; shf.l.wrap.b32 %r87, %r86, %r86, 11; sub.s32 %r2716, %r2715, %r87; xor.b32 %r2717, %r2716, %r2711; shf.l.wrap.b32 %r2718, %r2716, %r2716, 25; sub.s32 %r2719, %r2717, %r2718; xor.b32 %r2720, %r2719, %r86; shf.l.wrap.b32 %r2721, %r2719, %r2719, 16; sub.s32 %r2722, %r2720, %r2721; xor.b32 %r2723, %r2722, %r2716; shf.l.wrap.b32 %r2724, %r2722, %r2722, 4; sub.s32 %r2725, %r2723, %r2724; xor.b32 %r2726, %r2725, %r2719; shf.l.wrap.b32 %r2727, %r2725, %r2725, 14; sub.s32 %r2728, %r2726, %r2727; xor.b32 %r2729, %r2728, %r2722; shf.l.wrap.b32 %r2730, %r2728, %r2728, 24; sub.s32 %r88, %r2729, %r2730; and.b32 %r89, %r88, 15; cvt.u16.u32 %rs49, %r88; or.b16 %rs50, %rs49, 2; and.b16 %rs51, %rs50, 15; setp.eq.s16 %p508, %rs51, 14; mov.f32 %f1623, %f106; @%p508 bra $L__BB79_86; mov.f32 %f1623, %f108; $L__BB79_86: mul.ftz.f32 %f1354, %f106, %f106; mul.ftz.f32 %f1355, %f106, %f1354; mul.ftz.f32 %f113, %f1355, %f109; mul.ftz.f32 %f1356, %f107, %f107; mul.ftz.f32 %f1357, %f107, %f1356; mul.ftz.f32 %f114, %f1357, %f110; mul.ftz.f32 %f1358, %f108, %f108; mul.ftz.f32 %f1359, %f108, %f1358; mul.ftz.f32 %f115, %f1359, %f111; setp.lt.u32 %p509, %r89, 8; selp.f32 %f1360, %f106, %f107, %p509; setp.lt.u32 %p510, %r89, 4; selp.f32 %f1361, %f107, %f1623, %p510; and.b32 %r2731, %r88, 1; setp.eq.b32 %p511, %r2731, 1; neg.ftz.f32 %f1362, %f1360; selp.f32 %f1363, %f1362, %f1360, %p511; and.b32 %r2732, %r88, 2; setp.eq.s32 %p512, %r2732, 0; neg.ftz.f32 %f1364, %f1361; selp.f32 %f1365, %f1361, %f1364, %p512; add.ftz.f32 %f116, %f1363, %f1365; add.s32 %r90, %r82, -559038711; xor.b32 %r2733, %r86, %r90; sub.s32 %r2734, %r2733, %r87; xor.b32 %r2736, %r2734, %r2711; shf.l.wrap.b32 %r2737, %r2734, %r2734, 25; sub.s32 %r2738, %r2736, %r2737; xor.b32 %r2739, %r2738, %r86; shf.l.wrap.b32 %r2740, %r2738, %r2738, 16; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r2734; shf.l.wrap.b32 %r2743, %r2741, %r2741, 4; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 14; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 24; sub.s32 %r91, %r2748, %r2749; and.b32 %r92, %r91, 15; add.ftz.f32 %f117, %f106, 0fBF800000; cvt.u16.u32 %rs52, %r91; or.b16 %rs53, %rs52, 2; and.b16 %rs54, %rs53, 15; setp.eq.s16 %p513, %rs54, 14; mov.f32 %f1624, %f117; @%p513 bra $L__BB79_88; mov.f32 %f1624, %f108; $L__BB79_88: setp.lt.u32 %p514, %r92, 8; selp.f32 %f1366, %f117, %f107, %p514; setp.lt.u32 %p515, %r92, 4; selp.f32 %f1367, %f107, %f1624, %p515; and.b32 %r2750, %r91, 1; setp.eq.b32 %p516, %r2750, 1; neg.ftz.f32 %f1368, %f1366; selp.f32 %f1369, %f1368, %f1366, %p516; and.b32 %r2751, %r91, 2; setp.eq.s32 %p517, %r2751, 0; neg.ftz.f32 %f1370, %f1367; selp.f32 %f1371, %f1367, %f1370, %p517; add.ftz.f32 %f119, %f1369, %f1371; add.s32 %r93, %r83, -559038711; xor.b32 %r2753, %r2712, %r93; shf.l.wrap.b32 %r94, %r93, %r93, 14; sub.s32 %r95, %r2753, %r94; xor.b32 %r2755, %r95, %r2714; shf.l.wrap.b32 %r96, %r95, %r95, 11; sub.s32 %r2756, %r2755, %r96; xor.b32 %r2757, %r2756, %r93; shf.l.wrap.b32 %r2758, %r2756, %r2756, 25; sub.s32 %r2759, %r2757, %r2758; xor.b32 %r2760, %r2759, %r95; shf.l.wrap.b32 %r2761, %r2759, %r2759, 16; sub.s32 %r2762, %r2760, %r2761; xor.b32 %r2763, %r2762, %r2756; shf.l.wrap.b32 %r2764, %r2762, %r2762, 4; sub.s32 %r2765, %r2763, %r2764; xor.b32 %r2766, %r2765, %r2759; shf.l.wrap.b32 %r2767, %r2765, %r2765, 14; sub.s32 %r2768, %r2766, %r2767; xor.b32 %r2769, %r2768, %r2762; shf.l.wrap.b32 %r2770, %r2768, %r2768, 24; sub.s32 %r97, %r2769, %r2770; and.b32 %r98, %r97, 15; add.ftz.f32 %f120, %f107, 0fBF800000; cvt.u16.u32 %rs55, %r97; or.b16 %rs56, %rs55, 2; and.b16 %rs57, %rs56, 15; setp.eq.s16 %p518, %rs57, 14; mov.f32 %f1625, %f106; @%p518 bra $L__BB79_90; mov.f32 %f1625, %f108; $L__BB79_90: setp.lt.u32 %p519, %r98, 8; selp.f32 %f1372, %f106, %f120, %p519; setp.lt.u32 %p520, %r98, 4; selp.f32 %f1373, %f120, %f1625, %p520; and.b32 %r2771, %r97, 1; setp.eq.b32 %p521, %r2771, 1; neg.ftz.f32 %f1374, %f1372; selp.f32 %f1375, %f1374, %f1372, %p521; and.b32 %r2772, %r97, 2; setp.eq.s32 %p522, %r2772, 0; neg.ftz.f32 %f1376, %f1373; selp.f32 %f1377, %f1373, %f1376, %p522; add.ftz.f32 %f122, %f1375, %f1377; xor.b32 %r2773, %r95, %r90; sub.s32 %r2774, %r2773, %r96; xor.b32 %r2775, %r2774, %r93; shf.l.wrap.b32 %r2776, %r2774, %r2774, 25; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r95; shf.l.wrap.b32 %r2779, %r2777, %r2777, 16; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2774; shf.l.wrap.b32 %r2782, %r2780, %r2780, 4; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 14; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 24; sub.s32 %r99, %r2787, %r2788; and.b32 %r100, %r99, 15; cvt.u16.u32 %rs58, %r99; or.b16 %rs59, %rs58, 2; and.b16 %rs60, %rs59, 15; setp.eq.s16 %p523, %rs60, 14; mov.f32 %f1626, %f117; @%p523 bra $L__BB79_92; mov.f32 %f1626, %f108; $L__BB79_92: setp.lt.u32 %p524, %r100, 8; selp.f32 %f1378, %f117, %f120, %p524; setp.lt.u32 %p525, %r100, 4; selp.f32 %f1379, %f120, %f1626, %p525; and.b32 %r2789, %r99, 1; setp.eq.b32 %p526, %r2789, 1; neg.ftz.f32 %f1380, %f1378; selp.f32 %f1381, %f1380, %f1378, %p526; and.b32 %r2790, %r99, 2; setp.eq.s32 %p527, %r2790, 0; neg.ftz.f32 %f1382, %f1379; selp.f32 %f1383, %f1379, %f1382, %p527; add.ftz.f32 %f124, %f1381, %f1383; add.s32 %r101, %r84, -559038711; xor.b32 %r2792, %r101, %r2711; sub.s32 %r102, %r2792, %r85; xor.b32 %r2794, %r102, %r2714; shf.l.wrap.b32 %r103, %r102, %r102, 11; sub.s32 %r2795, %r2794, %r103; xor.b32 %r2796, %r2795, %r2711; shf.l.wrap.b32 %r2797, %r2795, %r2795, 25; sub.s32 %r2798, %r2796, %r2797; xor.b32 %r2799, %r2798, %r102; shf.l.wrap.b32 %r2800, %r2798, %r2798, 16; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2795; shf.l.wrap.b32 %r2803, %r2801, %r2801, 4; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 14; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 24; sub.s32 %r104, %r2808, %r2809; add.ftz.f32 %f125, %f108, 0fBF800000; and.b32 %r105, %r104, 15; cvt.u16.u32 %rs61, %r104; or.b16 %rs62, %rs61, 2; and.b16 %rs63, %rs62, 15; setp.eq.s16 %p528, %rs63, 14; mov.f32 %f1627, %f106; @%p528 bra $L__BB79_94; mov.f32 %f1627, %f125; $L__BB79_94: setp.lt.u32 %p529, %r105, 8; selp.f32 %f1384, %f106, %f107, %p529; setp.lt.u32 %p530, %r105, 4; selp.f32 %f1385, %f107, %f1627, %p530; and.b32 %r2810, %r104, 1; setp.eq.b32 %p531, %r2810, 1; neg.ftz.f32 %f1386, %f1384; selp.f32 %f1387, %f1386, %f1384, %p531; and.b32 %r2811, %r104, 2; setp.eq.s32 %p532, %r2811, 0; neg.ftz.f32 %f1388, %f1385; selp.f32 %f1389, %f1385, %f1388, %p532; add.ftz.f32 %f127, %f1387, %f1389; xor.b32 %r2812, %r102, %r90; sub.s32 %r2813, %r2812, %r103; xor.b32 %r2815, %r2813, %r2711; shf.l.wrap.b32 %r2816, %r2813, %r2813, 25; sub.s32 %r2817, %r2815, %r2816; xor.b32 %r2818, %r2817, %r102; shf.l.wrap.b32 %r2819, %r2817, %r2817, 16; sub.s32 %r2820, %r2818, %r2819; xor.b32 %r2821, %r2820, %r2813; shf.l.wrap.b32 %r2822, %r2820, %r2820, 4; sub.s32 %r2823, %r2821, %r2822; xor.b32 %r2824, %r2823, %r2817; shf.l.wrap.b32 %r2825, %r2823, %r2823, 14; sub.s32 %r2826, %r2824, %r2825; xor.b32 %r2827, %r2826, %r2820; shf.l.wrap.b32 %r2828, %r2826, %r2826, 24; sub.s32 %r106, %r2827, %r2828; and.b32 %r107, %r106, 15; cvt.u16.u32 %rs64, %r106; or.b16 %rs65, %rs64, 2; and.b16 %rs66, %rs65, 15; setp.eq.s16 %p533, %rs66, 14; mov.f32 %f1628, %f117; @%p533 bra $L__BB79_96; mov.f32 %f1628, %f125; $L__BB79_96: setp.lt.u32 %p534, %r107, 8; selp.f32 %f1390, %f117, %f107, %p534; setp.lt.u32 %p535, %r107, 4; selp.f32 %f1391, %f107, %f1628, %p535; and.b32 %r2829, %r106, 1; setp.eq.b32 %p536, %r2829, 1; neg.ftz.f32 %f1392, %f1390; selp.f32 %f1393, %f1392, %f1390, %p536; and.b32 %r2830, %r106, 2; setp.eq.s32 %p537, %r2830, 0; neg.ftz.f32 %f1394, %f1391; selp.f32 %f1395, %f1391, %f1394, %p537; add.ftz.f32 %f129, %f1393, %f1395; xor.b32 %r2831, %r101, %r93; sub.s32 %r108, %r2831, %r94; xor.b32 %r2833, %r108, %r2714; shf.l.wrap.b32 %r109, %r108, %r108, 11; sub.s32 %r2834, %r2833, %r109; xor.b32 %r2835, %r2834, %r93; shf.l.wrap.b32 %r2836, %r2834, %r2834, 25; sub.s32 %r2837, %r2835, %r2836; xor.b32 %r2838, %r2837, %r108; shf.l.wrap.b32 %r2839, %r2837, %r2837, 16; sub.s32 %r2840, %r2838, %r2839; xor.b32 %r2841, %r2840, %r2834; shf.l.wrap.b32 %r2842, %r2840, %r2840, 4; sub.s32 %r2843, %r2841, %r2842; xor.b32 %r2844, %r2843, %r2837; shf.l.wrap.b32 %r2845, %r2843, %r2843, 14; sub.s32 %r2846, %r2844, %r2845; xor.b32 %r2847, %r2846, %r2840; shf.l.wrap.b32 %r2848, %r2846, %r2846, 24; sub.s32 %r110, %r2847, %r2848; and.b32 %r111, %r110, 15; cvt.u16.u32 %rs67, %r110; or.b16 %rs68, %rs67, 2; and.b16 %rs69, %rs68, 15; setp.eq.s16 %p538, %rs69, 14; mov.f32 %f1629, %f106; @%p538 bra $L__BB79_98; mov.f32 %f1629, %f125; $L__BB79_98: setp.lt.u32 %p539, %r111, 8; selp.f32 %f1396, %f106, %f120, %p539; setp.lt.u32 %p540, %r111, 4; selp.f32 %f1397, %f120, %f1629, %p540; and.b32 %r2849, %r110, 1; setp.eq.b32 %p541, %r2849, 1; neg.ftz.f32 %f1398, %f1396; selp.f32 %f1399, %f1398, %f1396, %p541; and.b32 %r2850, %r110, 2; setp.eq.s32 %p542, %r2850, 0; neg.ftz.f32 %f1400, %f1397; selp.f32 %f1401, %f1397, %f1400, %p542; add.ftz.f32 %f131, %f1399, %f1401; xor.b32 %r2851, %r108, %r90; sub.s32 %r2852, %r2851, %r109; xor.b32 %r2853, %r2852, %r93; shf.l.wrap.b32 %r2854, %r2852, %r2852, 25; sub.s32 %r2855, %r2853, %r2854; xor.b32 %r2856, %r2855, %r108; shf.l.wrap.b32 %r2857, %r2855, %r2855, 16; sub.s32 %r2858, %r2856, %r2857; xor.b32 %r2859, %r2858, %r2852; shf.l.wrap.b32 %r2860, %r2858, %r2858, 4; sub.s32 %r2861, %r2859, %r2860; xor.b32 %r2862, %r2861, %r2855; shf.l.wrap.b32 %r2863, %r2861, %r2861, 14; sub.s32 %r2864, %r2862, %r2863; xor.b32 %r2865, %r2864, %r2858; shf.l.wrap.b32 %r2866, %r2864, %r2864, 24; sub.s32 %r112, %r2865, %r2866; and.b32 %r113, %r112, 15; cvt.u16.u32 %rs70, %r112; or.b16 %rs71, %rs70, 2; and.b16 %rs72, %rs71, 15; setp.eq.s16 %p543, %rs72, 14; mov.f32 %f1630, %f117; @%p543 bra $L__BB79_100; mov.f32 %f1630, %f125; $L__BB79_100: setp.lt.u32 %p544, %r113, 8; selp.f32 %f1403, %f117, %f120, %p544; setp.lt.u32 %p545, %r113, 4; selp.f32 %f1404, %f120, %f1630, %p545; and.b32 %r2867, %r112, 1; setp.eq.b32 %p546, %r2867, 1; neg.ftz.f32 %f1405, %f1403; selp.f32 %f1406, %f1405, %f1403, %p546; and.b32 %r2868, %r112, 2; setp.eq.s32 %p547, %r2868, 0; neg.ftz.f32 %f1407, %f1404; selp.f32 %f1408, %f1404, %f1407, %p547; add.ftz.f32 %f1409, %f1406, %f1408; sub.ftz.f32 %f1411, %f1241, %f113; mul.ftz.f32 %f1412, %f113, %f119; fma.rn.ftz.f32 %f1413, %f1411, %f116, %f1412; sub.ftz.f32 %f1414, %f1241, %f114; mul.ftz.f32 %f1415, %f113, %f124; fma.rn.ftz.f32 %f1416, %f1411, %f122, %f1415; mul.ftz.f32 %f1417, %f114, %f1416; fma.rn.ftz.f32 %f1418, %f1414, %f1413, %f1417; sub.ftz.f32 %f1419, %f1241, %f115; mul.ftz.f32 %f1420, %f113, %f129; fma.rn.ftz.f32 %f1421, %f1411, %f127, %f1420; mul.ftz.f32 %f1422, %f113, %f1409; fma.rn.ftz.f32 %f1423, %f1411, %f131, %f1422; mul.ftz.f32 %f1424, %f114, %f1423; fma.rn.ftz.f32 %f1425, %f1414, %f1421, %f1424; mul.ftz.f32 %f1426, %f115, %f1425; fma.rn.ftz.f32 %f133, %f1419, %f1418, %f1426; mov.b32 %r114, %f133; setp.nan.ftz.f32 %p548, %f133, %f133; mov.f32 %f1631, 0f00000000; @%p548 bra $L__BB79_103; or.b32 %r2869, %r114, -2147483648; setp.ne.s32 %p549, %r2869, -2147483648; add.ftz.f32 %f1428, %f133, %f133; setp.eq.ftz.f32 %p550, %f1428, %f133; and.pred %p551, %p550, %p549; @%p551 bra $L__BB79_103; shl.b32 %r2870, %r114, 1; setp.lt.u32 %p552, %r2870, -16777215; mul.ftz.f32 %f1429, %f133, 0f3F7B645A; selp.f32 %f1631, %f1429, 0f00000000, %p552; $L__BB79_103: mul.ftz.f32 %f1430, %f1613, 0f3F7B645A; fma.rn.ftz.f32 %f1601, %f18, %f1430, %f1601; fma.rn.ftz.f32 %f1602, %f18, %f1622, %f1602; fma.rn.ftz.f32 %f1603, %f18, %f1631, %f1603; $L__BB79_104: { // callseq 51, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1601; st.param.f32 [param0+4], %f1602; st.param.f32 [param0+8], %f1603; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_3d6float3ff, ( param0, param1, param2 ); ld.param.f32 %f1644, [retval0+0]; } // callseq 51 setp.eq.s32 %p553, %r7, 255; @%p553 bra $L__BB79_113; add.ftz.f32 %f1432, %f1601, 0f432571FE; add.ftz.f32 %f1433, %f1602, 0f4322C4AA; add.ftz.f32 %f1434, %f1603, 0f4341688E; { // callseq 52, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1432; st.param.f32 [param0+4], %f1433; st.param.f32 [param0+8], %f1434; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_3d6float3ff, ( param0, param1, param2 ); ld.param.f32 %f1643, [retval0+0]; } // callseq 52 add.ftz.f32 %f1435, %f1601, 0f431B94B7; add.ftz.f32 %f1436, %f1602, 0f43312796; add.ftz.f32 %f1437, %f1603, 0f430A23D9; { // callseq 53, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1435; st.param.f32 [param0+4], %f1436; st.param.f32 [param0+8], %f1437; .param .b32 param1; st.param.f32 [param1+0], %f13; .param .b32 param2; st.param.f32 [param2+0], %f16; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_3d6float3ff, ( param0, param1, param2 ); ld.param.f32 %f1642, [retval0+0]; } // callseq 53 $L__BB79_113: setp.eq.s32 %p599, %r6, 255; @%p599 bra $L__BB79_115; mul.wide.u32 %rd20, %r6, 4; add.s64 %rd21, %rd1, %rd20; st.local.f32 [%rd21], %f1644; $L__BB79_115: setp.eq.s32 %p600, %r7, 255; @%p600 bra $L__BB79_117; mul.wide.u32 %rd22, %r7, 4; add.s64 %rd23, %rd1, %rd22; st.local.f32 [%rd23], %f1644; st.local.f32 [%rd23+4], %f1643; st.local.f32 [%rd23+8], %f1642; $L__BB79_117: ld.param.u32 %r3109, [_Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r3108, %r3109, 2; st.param.b32 [func_retval0+0], %r3108; ret; } .func (.param .b32 func_retval0) _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<3>; .reg .f32 %f<28>; .reg .b32 %r<6>; .reg .b64 %rd<13>; ld.param.u64 %rd2, [_Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r4, [_Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r1, [_Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r2, [_Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r3, [_Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r3, 16; add.s64 %rd6, %rd4, %rd5; mul.wide.u32 %rd7, %r4, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f1, [%rd8]; ld.local.f32 %f2, [%rd8+4]; ld.local.f32 %f3, [%rd8+8]; ld.global.v4.f32 {%f7, %f8, %f9, %f10}, [%rd6]; mul.ftz.f32 %f14, %f8, %f8; fma.rn.ftz.f32 %f15, %f7, %f7, %f14; fma.rn.ftz.f32 %f16, %f9, %f9, %f15; sqrt.approx.ftz.f32 %f17, %f16; div.approx.ftz.f32 %f4, %f7, %f17; div.approx.ftz.f32 %f5, %f8, %f17; div.approx.ftz.f32 %f6, %f9, %f17; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB80_2; mul.wide.u32 %rd9, %r1, 4; add.s64 %rd10, %rd1, %rd9; st.local.f32 [%rd10], %f4; st.local.f32 [%rd10+4], %f5; st.local.f32 [%rd10+8], %f6; $L__BB80_2: setp.eq.s32 %p2, %r2, 255; @%p2 bra $L__BB80_4; mul.ftz.f32 %f18, %f2, %f2; fma.rn.ftz.f32 %f19, %f1, %f1, %f18; fma.rn.ftz.f32 %f20, %f3, %f3, %f19; sqrt.approx.ftz.f32 %f21, %f20; div.approx.ftz.f32 %f22, %f1, %f21; div.approx.ftz.f32 %f23, %f2, %f21; div.approx.ftz.f32 %f24, %f3, %f21; mul.ftz.f32 %f25, %f5, %f23; fma.rn.ftz.f32 %f26, %f4, %f22, %f25; fma.rn.ftz.f32 %f27, %f6, %f24, %f26; mul.wide.u32 %rd11, %r2, 4; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f27; $L__BB80_4: add.s32 %r5, %r3, 1; st.param.b32 [func_retval0+0], %r5; ret; } .func (.param .b32 func_retval0) _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<6>; .reg .f32 %f<41>; .reg .b32 %r<15>; .reg .b64 %rd<16>; ld.param.u64 %rd3, [_Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r5, [_Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r6, [_Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r7, [_Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd3; bfe.u32 %r1, %r7, 8, 8; bfe.u32 %r2, %r7, 16, 8; ld.const.u64 %rd4, [kernel_params+2400]; cvta.to.global.u64 %rd5, %rd4; mul.wide.s32 %rd6, %r5, 16; add.s64 %rd7, %rd5, %rd6; shl.b32 %r8, %r7, 2; cvt.u64.u32 %rd8, %r8; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; setp.eq.s32 %p1, %r6, 0; ld.local.f32 %f14, [%rd10]; cvt.ftz.sat.f32.f32 %f15, %f14; ld.global.u32 %r3, [%rd7]; add.s32 %r9, %r3, -1; cvt.rn.f32.s32 %f16, %r9; mul.ftz.f32 %f17, %f15, %f16; cvt.rzi.ftz.s32.f32 %r10, %f17; max.s32 %r11, %r10, 0; min.s32 %r12, %r11, %r9; cvt.rn.f32.s32 %f18, %r12; sub.ftz.f32 %f1, %f17, %f18; add.s32 %r4, %r5, 1; add.s32 %r13, %r12, %r4; mul.wide.s32 %rd11, %r13, 16; add.s64 %rd2, %rd5, %rd11; ld.global.v4.f32 {%f37, %f38, %f39, %f40}, [%rd2]; setp.leu.ftz.f32 %p2, %f1, 0f00000000; or.pred %p3, %p1, %p2; @%p3 bra $L__BB81_2; mov.f32 %f23, 0f3F800000; sub.ftz.f32 %f24, %f23, %f1; ld.global.v4.f32 {%f25, %f26, %f27, %f28}, [%rd2+16]; mul.ftz.f32 %f33, %f1, %f25; mul.ftz.f32 %f34, %f1, %f26; mul.ftz.f32 %f35, %f1, %f27; mul.ftz.f32 %f36, %f1, %f28; fma.rn.ftz.f32 %f37, %f24, %f37, %f33; fma.rn.ftz.f32 %f38, %f24, %f38, %f34; fma.rn.ftz.f32 %f39, %f24, %f39, %f35; fma.rn.ftz.f32 %f40, %f24, %f40, %f36; $L__BB81_2: setp.eq.s32 %p4, %r1, 255; @%p4 bra $L__BB81_4; mul.wide.u32 %rd12, %r1, 4; add.s64 %rd13, %rd1, %rd12; st.local.f32 [%rd13], %f37; st.local.f32 [%rd13+4], %f38; st.local.f32 [%rd13+8], %f39; $L__BB81_4: setp.eq.s32 %p5, %r2, 255; @%p5 bra $L__BB81_6; mul.wide.u32 %rd14, %r2, 4; add.s64 %rd15, %rd1, %rd14; st.local.f32 [%rd15], %f40; $L__BB81_6: add.s32 %r14, %r3, %r4; st.param.b32 [func_retval0+0], %r14; ret; } .func (.param .b32 func_retval0) _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<22>; .reg .f32 %f<346>; .reg .b32 %r<28>; .reg .b64 %rd<24>; ld.param.u64 %rd8, [_Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r4, [_Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r1, [_Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd8; ld.const.u64 %rd9, [kernel_params+2400]; cvta.to.global.u64 %rd2, %rd9; mul.wide.s32 %rd10, %r4, 16; add.s64 %rd3, %rd2, %rd10; add.s32 %r2, %r4, 1; shl.b32 %r5, %r1, 2; cvt.u64.u32 %rd11, %r5; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f1, [%rd13]; shr.u32 %r6, %r1, 6; cvt.u64.u32 %rd14, %r6; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.param.f32 %f164, [_Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.f32 %f165, [_Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; sub.ftz.f32 %f166, %f165, %f164; ld.local.f32 %f2, [%rd16]; sub.ftz.f32 %f167, %f2, %f164; ld.local.f32 %f3, [%rd16+4]; sub.ftz.f32 %f168, %f3, %f164; ld.local.f32 %f4, [%rd16+8]; sub.ftz.f32 %f169, %f4, %f164; div.approx.ftz.f32 %f5, %f167, %f166; div.approx.ftz.f32 %f6, %f168, %f166; div.approx.ftz.f32 %f7, %f169, %f166; setp.gt.u32 %p1, %r1, 16777215; setp.lt.ftz.f32 %p2, %f5, 0f00000000; setp.gt.ftz.f32 %p3, %f5, 0f3F800000; or.pred %p4, %p2, %p3; and.pred %p5, %p1, %p4; ld.global.u32 %r3, [%rd3]; add.s32 %r7, %r3, %r4; mul.wide.s32 %rd17, %r7, 16; add.s64 %rd4, %rd2, %rd17; @%p5 bra $L__BB82_3; bra.uni $L__BB82_1; $L__BB82_3: @%p2 bra $L__BB82_5; bra.uni $L__BB82_4; $L__BB82_5: ld.global.v4.f32 {%f333, %f205, %f206, %f207}, [%rd3+16]; ld.global.v4.f32 {%f208, %f209, %f210, %f211}, [%rd3+32]; sub.ftz.f32 %f332, %f333, %f208; neg.ftz.f32 %f334, %f5; bra.uni $L__BB82_6; $L__BB82_1: cvt.ftz.sat.f32.f32 %f170, %f5; add.s32 %r8, %r3, -1; cvt.rn.f32.s32 %f171, %r8; mul.ftz.f32 %f172, %f170, %f171; cvt.rzi.ftz.s32.f32 %r9, %f172; max.s32 %r10, %r9, 0; min.s32 %r11, %r10, %r8; cvt.rn.f32.s32 %f173, %r11; sub.ftz.f32 %f8, %f172, %f173; add.s32 %r12, %r11, %r2; mul.wide.s32 %rd18, %r12, 16; add.s64 %rd5, %rd2, %rd18; ld.global.v4.f32 {%f335, %f175, %f176, %f177}, [%rd5]; setp.leu.ftz.f32 %p6, %f8, 0f00000000; @%p6 bra $L__BB82_7; mov.f32 %f178, 0f3F800000; sub.ftz.f32 %f179, %f178, %f8; ld.global.v4.f32 {%f180, %f181, %f182, %f183}, [%rd5+16]; mul.ftz.f32 %f188, %f8, %f180; fma.rn.ftz.f32 %f335, %f179, %f335, %f188; bra.uni $L__BB82_7; $L__BB82_4: ld.global.v4.f32 {%f333, %f193, %f194, %f195}, [%rd4]; ld.global.v4.f32 {%f196, %f197, %f198, %f199}, [%rd4+-16]; sub.ftz.f32 %f332, %f333, %f196; add.ftz.f32 %f334, %f5, 0fBF800000; $L__BB82_6: add.s32 %r13, %r3, -1; cvt.rn.f32.s32 %f216, %r13; mul.ftz.f32 %f217, %f332, %f334; fma.rn.ftz.f32 %f335, %f217, %f216, %f333; $L__BB82_7: setp.gt.ftz.f32 %p8, %f6, 0f3F800000; setp.lt.ftz.f32 %p9, %f6, 0f00000000; or.pred %p10, %p9, %p8; and.pred %p12, %p1, %p10; @%p12 bra $L__BB82_10; bra.uni $L__BB82_8; $L__BB82_10: @%p9 bra $L__BB82_12; bra.uni $L__BB82_11; $L__BB82_12: ld.global.v4.f32 {%f255, %f338, %f257, %f258}, [%rd3+16]; ld.global.v4.f32 {%f259, %f260, %f261, %f262}, [%rd3+32]; sub.ftz.f32 %f337, %f338, %f260; neg.ftz.f32 %f339, %f6; bra.uni $L__BB82_13; $L__BB82_8: cvt.ftz.sat.f32.f32 %f221, %f6; add.s32 %r14, %r3, -1; cvt.rn.f32.s32 %f222, %r14; mul.ftz.f32 %f223, %f221, %f222; cvt.rzi.ftz.s32.f32 %r15, %f223; max.s32 %r16, %r15, 0; min.s32 %r17, %r16, %r14; cvt.rn.f32.s32 %f224, %r17; sub.ftz.f32 %f60, %f223, %f224; add.s32 %r18, %r17, %r2; mul.wide.s32 %rd19, %r18, 16; add.s64 %rd6, %rd2, %rd19; ld.global.v4.f32 {%f225, %f340, %f227, %f228}, [%rd6]; setp.leu.ftz.f32 %p13, %f60, 0f00000000; @%p13 bra $L__BB82_14; mov.f32 %f229, 0f3F800000; sub.ftz.f32 %f230, %f229, %f60; ld.global.v4.f32 {%f231, %f232, %f233, %f234}, [%rd6+16]; mul.ftz.f32 %f240, %f60, %f232; fma.rn.ftz.f32 %f340, %f230, %f340, %f240; bra.uni $L__BB82_14; $L__BB82_11: ld.global.v4.f32 {%f243, %f338, %f245, %f246}, [%rd4]; ld.global.v4.f32 {%f247, %f248, %f249, %f250}, [%rd4+-16]; sub.ftz.f32 %f337, %f338, %f248; add.ftz.f32 %f339, %f6, 0fBF800000; $L__BB82_13: add.s32 %r19, %r3, -1; cvt.rn.f32.s32 %f267, %r19; mul.ftz.f32 %f269, %f337, %f339; fma.rn.ftz.f32 %f340, %f269, %f267, %f338; $L__BB82_14: setp.gt.ftz.f32 %p15, %f7, 0f3F800000; setp.lt.ftz.f32 %p16, %f7, 0f00000000; or.pred %p17, %p16, %p15; and.pred %p19, %p1, %p17; @%p19 bra $L__BB82_17; bra.uni $L__BB82_15; $L__BB82_17: @%p16 bra $L__BB82_19; bra.uni $L__BB82_18; $L__BB82_19: ld.global.v4.f32 {%f306, %f307, %f343, %f309}, [%rd3+16]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd3+32]; sub.ftz.f32 %f342, %f343, %f312; neg.ftz.f32 %f344, %f7; bra.uni $L__BB82_20; $L__BB82_15: cvt.ftz.sat.f32.f32 %f272, %f7; add.s32 %r20, %r3, -1; cvt.rn.f32.s32 %f273, %r20; mul.ftz.f32 %f274, %f272, %f273; cvt.rzi.ftz.s32.f32 %r21, %f274; max.s32 %r22, %r21, 0; min.s32 %r23, %r22, %r20; cvt.rn.f32.s32 %f275, %r23; sub.ftz.f32 %f112, %f274, %f275; add.s32 %r24, %r23, %r2; mul.wide.s32 %rd20, %r24, 16; add.s64 %rd7, %rd2, %rd20; ld.global.v4.f32 {%f276, %f277, %f345, %f279}, [%rd7]; setp.leu.ftz.f32 %p20, %f112, 0f00000000; @%p20 bra $L__BB82_21; mov.f32 %f280, 0f3F800000; sub.ftz.f32 %f281, %f280, %f112; ld.global.v4.f32 {%f282, %f283, %f284, %f285}, [%rd7+16]; mul.ftz.f32 %f292, %f112, %f284; fma.rn.ftz.f32 %f345, %f281, %f345, %f292; bra.uni $L__BB82_21; $L__BB82_18: ld.global.v4.f32 {%f294, %f295, %f343, %f297}, [%rd4]; ld.global.v4.f32 {%f298, %f299, %f300, %f301}, [%rd4+-16]; sub.ftz.f32 %f342, %f343, %f300; add.ftz.f32 %f344, %f7, 0fBF800000; $L__BB82_20: add.s32 %r25, %r3, -1; cvt.rn.f32.s32 %f318, %r25; mul.ftz.f32 %f321, %f342, %f344; fma.rn.ftz.f32 %f345, %f321, %f318, %f343; $L__BB82_21: mov.f32 %f323, 0f3F800000; sub.ftz.f32 %f324, %f323, %f1; mul.ftz.f32 %f325, %f1, %f345; mul.ftz.f32 %f326, %f1, %f335; fma.rn.ftz.f32 %f327, %f2, %f324, %f326; mul.ftz.f32 %f328, %f1, %f340; fma.rn.ftz.f32 %f329, %f324, %f3, %f328; fma.rn.ftz.f32 %f330, %f324, %f4, %f325; shr.u32 %r26, %r1, 14; cvt.u64.u32 %rd21, %r26; and.b64 %rd22, %rd21, 1020; add.s64 %rd23, %rd1, %rd22; st.local.f32 [%rd23], %f327; st.local.f32 [%rd23+4], %f329; st.local.f32 [%rd23+8], %f330; add.s32 %r27, %r3, %r2; st.param.b32 [func_retval0+0], %r27; ret; } .func (.param .b32 func_retval0) _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<8>; .reg .f32 %f<42>; .reg .b32 %r<16>; .reg .b64 %rd<20>; ld.param.u64 %rd5, [_Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r4, [_Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r1, [_Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd5; ld.const.u64 %rd6, [kernel_params+2400]; cvta.to.global.u64 %rd2, %rd6; mul.wide.s32 %rd7, %r4, 16; add.s64 %rd3, %rd2, %rd7; ld.global.u32 %r2, [%rd3]; add.s32 %r3, %r4, 1; shl.b32 %r5, %r1, 2; cvt.u64.u32 %rd8, %r5; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r6, %r1, 6; cvt.u64.u32 %rd11, %r6; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.param.f32 %f18, [_Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.f32 %f19, [_Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; sub.ftz.f32 %f20, %f19, %f18; ld.local.f32 %f2, [%rd13]; sub.ftz.f32 %f21, %f2, %f18; div.approx.ftz.f32 %f3, %f21, %f20; setp.gt.u32 %p1, %r1, 16777215; setp.lt.ftz.f32 %p2, %f3, 0f00000000; setp.gt.ftz.f32 %p3, %f3, 0f3F800000; or.pred %p4, %p2, %p3; and.pred %p5, %p1, %p4; @%p5 bra $L__BB83_3; bra.uni $L__BB83_1; $L__BB83_3: @%p2 bra $L__BB83_5; bra.uni $L__BB83_4; $L__BB83_5: ld.global.f32 %f31, [%rd3+32]; ld.global.f32 %f39, [%rd3+16]; sub.ftz.f32 %f40, %f39, %f31; neg.ftz.f32 %f38, %f3; bra.uni $L__BB83_6; $L__BB83_1: cvt.ftz.sat.f32.f32 %f22, %f3; add.s32 %r7, %r2, -1; cvt.rn.f32.s32 %f23, %r7; mul.ftz.f32 %f24, %f22, %f23; cvt.rzi.ftz.s32.f32 %r8, %f24; max.s32 %r9, %r8, 0; min.s32 %r10, %r9, %r7; cvt.rn.f32.s32 %f25, %r10; sub.ftz.f32 %f4, %f24, %f25; add.s32 %r11, %r10, %r3; mul.wide.s32 %rd14, %r11, 16; add.s64 %rd4, %rd2, %rd14; ld.global.f32 %f41, [%rd4]; setp.leu.ftz.f32 %p6, %f4, 0f00000000; @%p6 bra $L__BB83_7; mov.f32 %f26, 0f3F800000; sub.ftz.f32 %f27, %f26, %f4; ld.global.f32 %f28, [%rd4+16]; mul.ftz.f32 %f29, %f4, %f28; fma.rn.ftz.f32 %f41, %f41, %f27, %f29; bra.uni $L__BB83_7; $L__BB83_4: add.s32 %r12, %r2, %r4; mul.wide.s32 %rd15, %r12, 16; add.s64 %rd16, %rd2, %rd15; ld.global.f32 %f30, [%rd16+-16]; ld.global.f32 %f39, [%rd16]; sub.ftz.f32 %f40, %f39, %f30; add.ftz.f32 %f38, %f3, 0fBF800000; $L__BB83_6: add.s32 %r13, %r2, -1; cvt.rn.f32.s32 %f32, %r13; mul.ftz.f32 %f33, %f38, %f40; fma.rn.ftz.f32 %f41, %f33, %f32, %f39; $L__BB83_7: mov.f32 %f34, 0f3F800000; sub.ftz.f32 %f35, %f34, %f1; mul.ftz.f32 %f36, %f1, %f41; fma.rn.ftz.f32 %f37, %f35, %f2, %f36; shr.u32 %r14, %r1, 14; cvt.u64.u32 %rd17, %r14; and.b64 %rd18, %rd17, 1020; add.s64 %rd19, %rd1, %rd18; st.local.f32 [%rd19], %f37; add.s32 %r15, %r2, %r3; st.param.b32 [func_retval0+0], %r15; ret; } .func _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<11>; .reg .f32 %f<85>; .reg .b32 %r<7>; .reg .b64 %rd<14>; ld.param.u64 %rd2, [_Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r1, [_Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r3, [_Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; shl.b32 %r4, %r3, 2; cvt.u64.u32 %rd3, %r4; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; shr.u32 %r5, %r3, 6; cvt.u64.u32 %rd6, %r5; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; shr.u32 %r6, %r3, 14; cvt.u64.u32 %rd9, %r6; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f1, [%rd5]; ld.local.f32 %f2, [%rd8]; ld.local.f32 %f3, [%rd11]; setp.eq.s32 %p1, %r1, 2; @%p1 bra $L__BB84_10; setp.ne.s32 %p2, %r1, 1; mov.f32 %f79, %f1; mov.f32 %f80, %f2; mov.f32 %f81, %f3; @%p2 bra $L__BB84_11; setp.neu.ftz.f32 %p3, %f2, 0f00000000; mov.f32 %f79, %f3; mov.f32 %f80, %f3; mov.f32 %f81, %f3; @%p3 bra $L__BB84_3; bra.uni $L__BB84_11; $L__BB84_3: setp.eq.ftz.f32 %p4, %f1, 0f3F800000; mov.f32 %f44, 0f3F800000; mul.ftz.f32 %f45, %f1, 0f40C00000; selp.f32 %f46, 0f00000000, %f45, %p4; cvt.rmi.ftz.f32.f32 %f10, %f46; sub.ftz.f32 %f47, %f46, %f10; sub.ftz.f32 %f48, %f44, %f2; mul.ftz.f32 %f11, %f3, %f48; mul.ftz.f32 %f49, %f2, %f47; sub.ftz.f32 %f50, %f44, %f49; mul.ftz.f32 %f12, %f3, %f50; sub.ftz.f32 %f51, %f44, %f47; mul.ftz.f32 %f52, %f2, %f51; sub.ftz.f32 %f53, %f44, %f52; mul.ftz.f32 %f13, %f3, %f53; setp.eq.ftz.f32 %p5, %f10, 0f00000000; @%p5 bra $L__BB84_9; bra.uni $L__BB84_4; $L__BB84_9: mov.f32 %f79, %f3; mov.f32 %f80, %f13; mov.f32 %f81, %f11; bra.uni $L__BB84_11; $L__BB84_10: fma.rn.ftz.f32 %f54, %f1, 0f40C00000, 0fC0400000; abs.ftz.f32 %f55, %f54; add.ftz.f32 %f56, %f55, 0fBF800000; mov.f32 %f57, 0f3F800000; fma.rn.ftz.f32 %f58, %f1, 0f40C00000, 0fC0000000; mov.f32 %f59, 0f40000000; abs.ftz.f32 %f60, %f58; sub.ftz.f32 %f61, %f59, %f60; fma.rn.ftz.f32 %f62, %f1, 0f40C00000, 0fC0800000; abs.ftz.f32 %f63, %f62; sub.ftz.f32 %f64, %f59, %f63; mov.f32 %f65, 0f00000000; max.ftz.f32 %f66, %f56, %f65; min.ftz.f32 %f67, %f66, %f57; max.ftz.f32 %f68, %f64, %f65; min.ftz.f32 %f69, %f68, %f57; max.ftz.f32 %f70, %f61, %f65; min.ftz.f32 %f71, %f70, %f57; fma.rn.ftz.f32 %f72, %f3, 0f40000000, 0fBF800000; abs.ftz.f32 %f73, %f72; sub.ftz.f32 %f74, %f57, %f73; mul.ftz.f32 %f75, %f2, %f74; add.ftz.f32 %f76, %f67, 0fBF000000; add.ftz.f32 %f77, %f71, 0fBF000000; add.ftz.f32 %f78, %f69, 0fBF000000; fma.rn.ftz.f32 %f81, %f78, %f75, %f3; fma.rn.ftz.f32 %f80, %f77, %f75, %f3; fma.rn.ftz.f32 %f79, %f76, %f75, %f3; $L__BB84_11: setp.eq.s32 %p10, %r2, 255; @%p10 bra $L__BB84_13; mul.wide.u32 %rd12, %r2, 4; add.s64 %rd13, %rd1, %rd12; st.local.f32 [%rd13], %f79; st.local.f32 [%rd13+4], %f80; st.local.f32 [%rd13+8], %f81; $L__BB84_13: ret; $L__BB84_4: setp.eq.ftz.f32 %p6, %f10, 0f3F800000; mov.f32 %f79, %f12; mov.f32 %f80, %f3; mov.f32 %f81, %f11; @%p6 bra $L__BB84_11; setp.eq.ftz.f32 %p7, %f10, 0f40000000; mov.f32 %f79, %f11; mov.f32 %f80, %f3; mov.f32 %f81, %f13; @%p7 bra $L__BB84_11; setp.eq.ftz.f32 %p8, %f10, 0f40400000; mov.f32 %f79, %f11; mov.f32 %f80, %f12; mov.f32 %f81, %f3; @%p8 bra $L__BB84_11; setp.eq.ftz.f32 %p9, %f10, 0f40800000; mov.f32 %f79, %f3; mov.f32 %f80, %f11; mov.f32 %f81, %f12; @%p9 bra $L__BB84_8; bra.uni $L__BB84_11; $L__BB84_8: mov.f32 %f79, %f13; mov.f32 %f80, %f11; mov.f32 %f81, %f3; bra.uni $L__BB84_11; } .func _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<16>; .reg .f32 %f<82>; .reg .b32 %r<7>; .reg .b64 %rd<11>; ld.param.u64 %rd2, [_Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r4, [_Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r6, [_Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r5, [_Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; mul.wide.u32 %rd3, %r6, 4; add.s64 %rd4, %rd1, %rd3; ld.local.f32 %f79, [%rd4]; ld.local.f32 %f2, [%rd4+4]; ld.local.f32 %f3, [%rd4+8]; setp.eq.s32 %p1, %r4, 2; @%p1 bra $L__BB85_12; setp.ne.s32 %p2, %r4, 1; mov.f32 %f78, %f2; mov.f32 %f81, %f3; @%p2 bra $L__BB85_22; max.ftz.f32 %f43, %f2, %f3; max.ftz.f32 %f81, %f79, %f43; min.ftz.f32 %f44, %f2, %f3; min.ftz.f32 %f45, %f79, %f44; sub.ftz.f32 %f8, %f81, %f45; setp.eq.ftz.f32 %p3, %f81, 0f00000000; mov.f32 %f75, 0f00000000; mov.f32 %f78, %f75; @%p3 bra $L__BB85_4; div.approx.ftz.f32 %f78, %f8, %f81; $L__BB85_4: setp.eq.ftz.f32 %p4, %f78, 0f00000000; @%p4 bra $L__BB85_11; sub.ftz.f32 %f47, %f81, %f79; div.approx.ftz.f32 %f11, %f47, %f8; sub.ftz.f32 %f48, %f81, %f2; div.approx.ftz.f32 %f12, %f48, %f8; sub.ftz.f32 %f49, %f81, %f3; div.approx.ftz.f32 %f13, %f49, %f8; setp.eq.ftz.f32 %p5, %f79, %f81; @%p5 bra $L__BB85_9; bra.uni $L__BB85_6; $L__BB85_9: sub.ftz.f32 %f74, %f13, %f12; bra.uni $L__BB85_10; $L__BB85_12: max.ftz.f32 %f57, %f2, %f3; max.ftz.f32 %f23, %f79, %f57; min.ftz.f32 %f58, %f2, %f3; min.ftz.f32 %f24, %f79, %f58; add.ftz.f32 %f25, %f23, %f24; mov.f32 %f59, 0f40000000; div.approx.ftz.f32 %f60, %f25, %f59; mov.f32 %f61, 0f3F800000; min.ftz.f32 %f81, %f61, %f60; setp.eq.ftz.f32 %p8, %f23, %f24; mov.f32 %f77, 0f00000000; mov.f32 %f78, %f77; @%p8 bra $L__BB85_21; sub.ftz.f32 %f27, %f23, %f24; setp.gt.ftz.f32 %p9, %f81, 0f3F000000; @%p9 bra $L__BB85_15; bra.uni $L__BB85_14; $L__BB85_15: sub.ftz.f32 %f63, %f59, %f23; sub.ftz.f32 %f64, %f63, %f24; div.approx.ftz.f32 %f78, %f27, %f64; bra.uni $L__BB85_16; $L__BB85_6: setp.eq.ftz.f32 %p6, %f2, %f81; @%p6 bra $L__BB85_8; bra.uni $L__BB85_7; $L__BB85_8: add.ftz.f32 %f51, %f11, 0f40000000; sub.ftz.f32 %f74, %f51, %f13; bra.uni $L__BB85_10; $L__BB85_14: div.approx.ftz.f32 %f78, %f27, %f25; $L__BB85_16: setp.eq.ftz.f32 %p10, %f23, %f79; @%p10 bra $L__BB85_20; bra.uni $L__BB85_17; $L__BB85_20: sub.ftz.f32 %f69, %f2, %f3; div.approx.ftz.f32 %f70, %f69, %f27; setp.lt.ftz.f32 %p12, %f2, %f3; selp.f32 %f71, 0f40C00000, 0f00000000, %p12; add.ftz.f32 %f77, %f71, %f70; bra.uni $L__BB85_21; $L__BB85_17: setp.eq.ftz.f32 %p11, %f23, %f2; @%p11 bra $L__BB85_19; bra.uni $L__BB85_18; $L__BB85_19: sub.ftz.f32 %f67, %f3, %f79; div.approx.ftz.f32 %f68, %f67, %f27; add.ftz.f32 %f77, %f68, 0f40000000; bra.uni $L__BB85_21; $L__BB85_7: add.ftz.f32 %f50, %f12, 0f40800000; sub.ftz.f32 %f74, %f50, %f11; $L__BB85_10: mov.f32 %f52, 0f40C00000; div.approx.ftz.f32 %f53, %f74, %f52; setp.lt.ftz.f32 %p7, %f53, 0f00000000; add.ftz.f32 %f54, %f53, 0f3F800000; selp.f32 %f75, %f54, %f53, %p7; $L__BB85_11: mov.f32 %f79, %f75; bra.uni $L__BB85_22; $L__BB85_18: sub.ftz.f32 %f65, %f79, %f2; div.approx.ftz.f32 %f66, %f65, %f27; add.ftz.f32 %f77, %f66, 0f40800000; $L__BB85_21: mov.f32 %f72, 0f40C00000; div.approx.ftz.f32 %f79, %f77, %f72; $L__BB85_22: bfe.u32 %r1, %r5, 8, 8; bfe.u32 %r2, %r5, 16, 8; and.b32 %r3, %r5, 255; setp.eq.s32 %p13, %r3, 255; @%p13 bra $L__BB85_24; mul.wide.u32 %rd5, %r3, 4; add.s64 %rd6, %rd1, %rd5; st.local.f32 [%rd6], %f79; $L__BB85_24: setp.eq.s32 %p14, %r1, 255; @%p14 bra $L__BB85_26; mul.wide.u32 %rd7, %r1, 4; add.s64 %rd8, %rd1, %rd7; st.local.f32 [%rd8], %f78; $L__BB85_26: setp.eq.s32 %p15, %r2, 255; @%p15 bra $L__BB85_28; mul.wide.u32 %rd9, %r2, 4; add.s64 %rd10, %rd1, %rd9; st.local.f32 [%rd10], %f81; $L__BB85_28: ret; } .func (.param .b32 func_retval0) _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<9>; .reg .f32 %f<45>; .reg .b32 %r<7>; .reg .b64 %rd<15>; ld.param.u64 %rd2, [_Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r2, [_Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r4, [_Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r5, [_Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r3, [_Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd2; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r3, 16; add.s64 %rd6, %rd4, %rd5; ld.global.u32 %r1, [%rd6+4]; mul.wide.u32 %rd7, %r4, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f1, [%rd8]; mul.wide.u32 %rd9, %r5, 4; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f2, [%rd10]; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB86_10; setp.neu.ftz.f32 %p2, %f1, 0f00000000; mov.f32 %f42, %f2; mov.f32 %f43, %f2; mov.f32 %f44, %f2; @%p2 bra $L__BB86_2; bra.uni $L__BB86_9; $L__BB86_2: mul.wide.u32 %rd11, %r2, 4; add.s64 %rd12, %rd1, %rd11; ld.local.f32 %f31, [%rd12]; setp.eq.ftz.f32 %p3, %f31, 0f3F800000; mov.f32 %f32, 0f3F800000; mul.ftz.f32 %f33, %f31, 0f40C00000; selp.f32 %f34, 0f00000000, %f33, %p3; cvt.rmi.ftz.f32.f32 %f6, %f34; sub.ftz.f32 %f35, %f34, %f6; sub.ftz.f32 %f36, %f32, %f1; mul.ftz.f32 %f7, %f2, %f36; mul.ftz.f32 %f37, %f1, %f35; sub.ftz.f32 %f38, %f32, %f37; mul.ftz.f32 %f8, %f2, %f38; sub.ftz.f32 %f39, %f32, %f35; mul.ftz.f32 %f40, %f1, %f39; sub.ftz.f32 %f41, %f32, %f40; mul.ftz.f32 %f9, %f2, %f41; setp.eq.ftz.f32 %p4, %f6, 0f00000000; @%p4 bra $L__BB86_8; bra.uni $L__BB86_3; $L__BB86_8: mov.f32 %f42, %f2; mov.f32 %f43, %f9; mov.f32 %f44, %f7; bra.uni $L__BB86_9; $L__BB86_3: setp.eq.ftz.f32 %p5, %f6, 0f3F800000; mov.f32 %f42, %f8; mov.f32 %f43, %f2; mov.f32 %f44, %f7; @%p5 bra $L__BB86_9; setp.eq.ftz.f32 %p6, %f6, 0f40000000; mov.f32 %f42, %f7; mov.f32 %f43, %f2; mov.f32 %f44, %f9; @%p6 bra $L__BB86_9; setp.eq.ftz.f32 %p7, %f6, 0f40400000; mov.f32 %f42, %f7; mov.f32 %f43, %f8; mov.f32 %f44, %f2; @%p7 bra $L__BB86_9; setp.eq.ftz.f32 %p8, %f6, 0f40800000; mov.f32 %f42, %f2; mov.f32 %f43, %f7; mov.f32 %f44, %f8; @%p8 bra $L__BB86_7; bra.uni $L__BB86_9; $L__BB86_7: mov.f32 %f42, %f9; mov.f32 %f43, %f7; mov.f32 %f44, %f2; $L__BB86_9: mul.wide.u32 %rd13, %r1, 4; add.s64 %rd14, %rd1, %rd13; st.local.f32 [%rd14], %f42; st.local.f32 [%rd14+4], %f43; st.local.f32 [%rd14+8], %f44; $L__BB86_10: add.s32 %r6, %r3, 1; st.param.b32 [func_retval0+0], %r6; ret; } .func (.param .b32 func_retval0) _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<9>; .reg .f32 %f<33>; .reg .b32 %r<7>; .reg .b64 %rd<16>; ld.param.u64 %rd3, [_Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r4, [_Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; ld.param.u32 %r1, [_Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r2, [_Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; ld.param.u32 %r3, [_Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd1, %rd3; ld.const.u64 %rd4, [kernel_params+2400]; cvta.to.global.u64 %rd5, %rd4; mul.wide.s32 %rd6, %r3, 16; add.s64 %rd7, %rd5, %rd6; ld.global.u32 %rd2, [%rd7+4]; mul.wide.u32 %rd8, %r4, 4; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f1, [%rd9+8]; ld.local.f32 %f2, [%rd9+4]; max.ftz.f32 %f18, %f2, %f1; ld.local.f32 %f3, [%rd9]; max.ftz.f32 %f4, %f3, %f18; min.ftz.f32 %f19, %f2, %f1; min.ftz.f32 %f20, %f3, %f19; sub.ftz.f32 %f5, %f4, %f20; setp.eq.ftz.f32 %p1, %f4, 0f00000000; mov.f32 %f32, 0f00000000; mov.f32 %f30, %f32; @%p1 bra $L__BB87_2; div.approx.ftz.f32 %f30, %f5, %f4; $L__BB87_2: setp.eq.ftz.f32 %p2, %f30, 0f00000000; @%p2 bra $L__BB87_9; sub.ftz.f32 %f22, %f4, %f3; div.approx.ftz.f32 %f8, %f22, %f5; sub.ftz.f32 %f23, %f4, %f2; div.approx.ftz.f32 %f9, %f23, %f5; sub.ftz.f32 %f24, %f4, %f1; div.approx.ftz.f32 %f10, %f24, %f5; setp.eq.ftz.f32 %p3, %f3, %f4; @%p3 bra $L__BB87_7; bra.uni $L__BB87_4; $L__BB87_7: sub.ftz.f32 %f31, %f10, %f9; bra.uni $L__BB87_8; $L__BB87_4: setp.eq.ftz.f32 %p4, %f2, %f4; @%p4 bra $L__BB87_6; bra.uni $L__BB87_5; $L__BB87_6: add.ftz.f32 %f26, %f8, 0f40000000; sub.ftz.f32 %f31, %f26, %f10; bra.uni $L__BB87_8; $L__BB87_5: add.ftz.f32 %f25, %f9, 0f40800000; sub.ftz.f32 %f31, %f25, %f8; $L__BB87_8: mov.f32 %f27, 0f40C00000; div.approx.ftz.f32 %f28, %f31, %f27; setp.lt.ftz.f32 %p5, %f28, 0f00000000; add.ftz.f32 %f29, %f28, 0f3F800000; selp.f32 %f32, %f29, %f28, %p5; $L__BB87_9: setp.eq.s32 %p6, %r1, 255; @%p6 bra $L__BB87_11; mul.wide.u32 %rd10, %r1, 4; add.s64 %rd11, %rd1, %rd10; st.local.f32 [%rd11], %f32; $L__BB87_11: setp.eq.s32 %p7, %r2, 255; @%p7 bra $L__BB87_13; mul.wide.u32 %rd12, %r2, 4; add.s64 %rd13, %rd1, %rd12; st.local.f32 [%rd13], %f30; $L__BB87_13: cvt.u32.u64 %r5, %rd2; setp.eq.s32 %p8, %r5, 255; @%p8 bra $L__BB87_15; shl.b64 %rd14, %rd2, 2; add.s64 %rd15, %rd1, %rd14; st.local.f32 [%rd15], %f4; $L__BB87_15: add.s32 %r6, %r3, 1; st.param.b32 [func_retval0+0], %r6; ret; } .func (.param .b32 func_retval0) _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<49>; .reg .f32 %f<602>; .reg .b32 %r<65>; .reg .b64 %rd<30>; ld.param.u64 %rd6, [_Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r16, [_Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r15, [_Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; ld.param.u32 %r14, [_Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r13, [_Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd6; mul.wide.u32 %rd7, %r13, 4; add.s64 %rd8, %rd1, %rd7; ld.const.u64 %rd9, [kernel_params+2400]; cvta.to.global.u64 %rd10, %rd9; mul.wide.s32 %rd11, %r16, 16; add.s64 %rd2, %rd10, %rd11; ld.global.v4.f32 {%f142, %f143, %f144, %f145}, [%rd2]; ld.local.f32 %f5, [%rd8+8]; mov.f32 %f146, 0fBF800000; max.ftz.f32 %f147, %f5, %f146; mov.f32 %f148, 0f3F800000; min.ftz.f32 %f149, %f147, %f148; abs.ftz.f32 %f150, %f149; sub.ftz.f32 %f151, %f148, %f150; mul.ftz.f32 %f152, %f151, 0f3F000000; sqrt.approx.ftz.f32 %f153, %f152; setp.gt.ftz.f32 %p1, %f150, 0f3F11EB85; selp.f32 %f154, %f153, %f150, %p1; mul.ftz.f32 %f155, %f154, %f154; mov.f32 %f156, 0f3C94D2E9; mov.f32 %f157, 0f3D53F941; fma.rn.ftz.f32 %f158, %f157, %f155, %f156; mov.f32 %f159, 0f3D3F841F; fma.rn.ftz.f32 %f160, %f158, %f155, %f159; mov.f32 %f161, 0f3D994929; fma.rn.ftz.f32 %f162, %f160, %f155, %f161; mov.f32 %f163, 0f3E2AAB94; fma.rn.ftz.f32 %f164, %f162, %f155, %f163; mul.ftz.f32 %f165, %f155, %f164; fma.rn.ftz.f32 %f166, %f165, %f154, %f154; add.ftz.f32 %f167, %f166, %f166; mov.f32 %f168, 0f3FC90FDB; sub.ftz.f32 %f169, %f168, %f166; selp.f32 %f170, %f167, %f169, %p1; setp.lt.ftz.f32 %p2, %f149, 0f00000000; mov.f32 %f171, 0f40490FDB; sub.ftz.f32 %f172, %f171, %f170; selp.f32 %f6, %f172, %f170, %p2; ld.local.f32 %f7, [%rd8+4]; abs.ftz.f32 %f8, %f7; ld.local.f32 %f9, [%rd8]; abs.ftz.f32 %f10, %f9; setp.eq.ftz.f32 %p3, %f8, 0f00000000; setp.eq.ftz.f32 %p4, %f10, 0f00000000; and.pred %p5, %p3, %p4; mov.b32 %r1, %f7; mov.b32 %r17, %f9; and.b32 %r2, %r17, -2147483648; @%p5 bra $L__BB88_4; bra.uni $L__BB88_1; $L__BB88_4: shr.s32 %r22, %r1, 31; and.b32 %r23, %r22, 1078530011; or.b32 %r24, %r23, %r2; mov.b32 %f584, %r24; bra.uni $L__BB88_5; $L__BB88_1: setp.eq.ftz.f32 %p6, %f8, 0f7F800000; setp.eq.ftz.f32 %p7, %f10, 0f7F800000; and.pred %p8, %p6, %p7; @%p8 bra $L__BB88_3; bra.uni $L__BB88_2; $L__BB88_3: setp.lt.s32 %p12, %r1, 0; selp.b32 %r20, 1075235812, 1061752795, %p12; or.b32 %r21, %r20, %r2; mov.b32 %f584, %r21; bra.uni $L__BB88_5; $L__BB88_2: setp.lt.s32 %p9, %r1, 0; min.ftz.f32 %f173, %f10, %f8; max.ftz.f32 %f174, %f10, %f8; div.full.ftz.f32 %f175, %f173, %f174; mul.rn.ftz.f32 %f176, %f175, %f175; mov.f32 %f177, 0fC0B59883; mov.f32 %f178, 0fBF52C7EA; fma.rn.ftz.f32 %f179, %f176, %f178, %f177; mov.f32 %f180, 0fC0D21907; fma.rn.ftz.f32 %f181, %f179, %f176, %f180; mul.ftz.f32 %f182, %f176, %f181; mul.ftz.f32 %f183, %f175, %f182; add.ftz.f32 %f184, %f176, 0f41355DC0; mov.f32 %f185, 0f41E6BD60; fma.rn.ftz.f32 %f186, %f184, %f176, %f185; mov.f32 %f187, 0f419D92C8; fma.rn.ftz.f32 %f188, %f186, %f176, %f187; rcp.approx.ftz.f32 %f189, %f188; fma.rn.ftz.f32 %f190, %f183, %f189, %f175; sub.ftz.f32 %f192, %f168, %f190; setp.gt.ftz.f32 %p10, %f10, %f8; selp.f32 %f193, %f192, %f190, %p10; sub.ftz.f32 %f195, %f171, %f193; selp.f32 %f196, %f195, %f193, %p9; mov.b32 %r18, %f196; or.b32 %r19, %r2, %r18; mov.b32 %f197, %r19; add.ftz.f32 %f198, %f8, %f10; setp.le.ftz.f32 %p11, %f198, 0f7F800000; selp.f32 %f584, %f197, %f198, %p11; $L__BB88_5: ld.global.v4.f32 {%f199, %f200, %f201, %f202}, [%rd2+16]; ld.global.v4.u32 {%r25, %r26, %r27, %r28}, [%rd2+32]; mov.b32 %f22, %r28; mov.b32 %f20, %r26; mov.b32 %f19, %r25; setp.lt.u32 %p13, %r15, 2; @%p13 bra $L__BB88_40; bra.uni $L__BB88_6; $L__BB88_40: add.s64 %rd5, %rd2, 48; ld.global.v4.f32 {%f338, %f339, %f340, %f341}, [%rd2+48]; ld.global.v4.f32 {%f342, %f343, %f344, %f345}, [%rd2+64]; ld.global.v4.f32 {%f346, %f347, %f348, %f349}, [%rd2+80]; ld.global.v4.f32 {%f350, %f351, %f352, %f353}, [%rd2+96]; add.s32 %r64, %r16, 8; sin.approx.ftz.f32 %f354, %f143; sin.approx.ftz.f32 %f355, %f6; mul.ftz.f32 %f356, %f355, %f354; sub.ftz.f32 %f357, %f142, %f584; cos.approx.ftz.f32 %f358, %f357; cos.approx.ftz.f32 %f359, %f143; cos.approx.ftz.f32 %f360, %f6; mul.ftz.f32 %f361, %f360, %f359; fma.rn.ftz.f32 %f362, %f356, %f358, %f361; max.ftz.f32 %f364, %f362, %f146; min.ftz.f32 %f366, %f364, %f148; abs.ftz.f32 %f367, %f366; sub.ftz.f32 %f368, %f148, %f367; mul.ftz.f32 %f369, %f368, 0f3F000000; sqrt.approx.ftz.f32 %f370, %f369; setp.gt.ftz.f32 %p42, %f367, 0f3F11EB85; selp.f32 %f371, %f370, %f367, %p42; mul.ftz.f32 %f372, %f371, %f371; fma.rn.ftz.f32 %f375, %f157, %f372, %f156; fma.rn.ftz.f32 %f377, %f375, %f372, %f159; fma.rn.ftz.f32 %f379, %f377, %f372, %f161; fma.rn.ftz.f32 %f381, %f379, %f372, %f163; mul.ftz.f32 %f382, %f372, %f381; fma.rn.ftz.f32 %f383, %f382, %f371, %f371; add.ftz.f32 %f384, %f383, %f383; sub.ftz.f32 %f386, %f168, %f383; selp.f32 %f387, %f384, %f386, %p42; setp.lt.ftz.f32 %p43, %f366, 0f00000000; sub.ftz.f32 %f389, %f171, %f387; selp.f32 %f123, %f389, %f387, %p43; mov.f32 %f390, 0f3FC8EF16; min.ftz.f32 %f391, %f6, %f390; cos.approx.ftz.f32 %f124, %f391; cos.approx.ftz.f32 %f125, %f123; setp.eq.s32 %p44, %r15, 0; @%p44 bra $L__BB88_42; ld.global.v4.f32 {%f392, %f393, %f394, %f395}, [%rd5+64]; mul.ftz.f32 %f398, %f123, %f20; mul.ftz.f32 %f399, %f398, 0f3FB8AA3B; ex2.approx.ftz.f32 %f400, %f399; mul.ftz.f32 %f401, %f125, %f125; add.ftz.f32 %f402, %f401, 0f3F800000; fma.rn.ftz.f32 %f403, %f339, %f339, 0f3F800000; mul.ftz.f32 %f404, %f339, 0fC0000000; fma.rn.ftz.f32 %f405, %f404, %f125, %f403; lg2.approx.ftz.f32 %f406, %f405; mul.ftz.f32 %f407, %f406, 0f3FC00000; ex2.approx.ftz.f32 %f408, %f407; div.approx.ftz.f32 %f409, %f402, %f408; add.ftz.f32 %f410, %f124, 0f3C23D70A; div.approx.ftz.f32 %f411, %f201, %f410; mul.ftz.f32 %f412, %f411, 0f3FB8AA3B; ex2.approx.ftz.f32 %f413, %f412; fma.rn.ftz.f32 %f414, %f413, %f200, 0f3F800000; fma.rn.ftz.f32 %f415, %f400, %f19, %f202; mov.b32 %f416, %r27; fma.rn.ftz.f32 %f417, %f401, %f416, %f415; fma.rn.ftz.f32 %f418, %f409, %f22, %f417; sqrt.approx.ftz.f32 %f419, %f124; fma.rn.ftz.f32 %f420, %f419, %f338, %f418; mul.ftz.f32 %f421, %f420, %f414; mul.ftz.f32 %f422, %f421, %f144; mul.ftz.f32 %f423, %f123, %f344; mul.ftz.f32 %f424, %f423, 0f3FB8AA3B; ex2.approx.ftz.f32 %f425, %f424; fma.rn.ftz.f32 %f426, %f348, %f348, 0f3F800000; add.ftz.f32 %f427, %f348, %f348; mul.ftz.f32 %f428, %f427, %f125; sub.ftz.f32 %f429, %f426, %f428; lg2.approx.ftz.f32 %f430, %f429; mul.ftz.f32 %f431, %f430, 0f3FC00000; ex2.approx.ftz.f32 %f432, %f431; div.approx.ftz.f32 %f433, %f402, %f432; div.approx.ftz.f32 %f434, %f341, %f410; mul.ftz.f32 %f435, %f434, 0f3FB8AA3B; ex2.approx.ftz.f32 %f436, %f435; fma.rn.ftz.f32 %f437, %f436, %f340, 0f3F800000; fma.rn.ftz.f32 %f438, %f425, %f343, %f342; fma.rn.ftz.f32 %f439, %f401, %f345, %f438; fma.rn.ftz.f32 %f440, %f433, %f346, %f439; fma.rn.ftz.f32 %f441, %f419, %f347, %f440; mul.ftz.f32 %f442, %f441, %f437; mul.ftz.f32 %f443, %f442, %f145; mul.ftz.f32 %f444, %f123, %f353; mul.ftz.f32 %f445, %f444, 0f3FB8AA3B; ex2.approx.ftz.f32 %f446, %f445; fma.rn.ftz.f32 %f447, %f395, %f395, 0f3F800000; add.ftz.f32 %f448, %f395, %f395; mul.ftz.f32 %f449, %f448, %f125; sub.ftz.f32 %f450, %f447, %f449; lg2.approx.ftz.f32 %f451, %f450; mul.ftz.f32 %f452, %f451, 0f3FC00000; ex2.approx.ftz.f32 %f453, %f452; div.approx.ftz.f32 %f454, %f402, %f453; div.approx.ftz.f32 %f455, %f350, %f410; mul.ftz.f32 %f456, %f455, 0f3FB8AA3B; ex2.approx.ftz.f32 %f457, %f456; fma.rn.ftz.f32 %f458, %f457, %f349, 0f3F800000; fma.rn.ftz.f32 %f459, %f446, %f352, %f351; fma.rn.ftz.f32 %f460, %f401, %f392, %f459; fma.rn.ftz.f32 %f462, %f454, %f393, %f460; fma.rn.ftz.f32 %f464, %f419, %f394, %f462; mul.ftz.f32 %f465, %f464, %f458; mul.ftz.f32 %f466, %f465, %f199; ld.const.v4.f32 {%f467, %f468, %f469, %f470}, [kernel_params+1136]; mul.ftz.f32 %f474, %f443, %f468; fma.rn.ftz.f32 %f475, %f422, %f467, %f474; fma.rn.ftz.f32 %f476, %f469, %f466, %f475; ld.const.v4.f32 {%f477, %f478, %f479, %f480}, [kernel_params+1152]; mul.ftz.f32 %f484, %f443, %f478; fma.rn.ftz.f32 %f485, %f422, %f477, %f484; fma.rn.ftz.f32 %f486, %f479, %f466, %f485; ld.const.v4.f32 {%f487, %f488, %f489, %f490}, [kernel_params+1168]; mul.ftz.f32 %f494, %f443, %f488; fma.rn.ftz.f32 %f495, %f422, %f487, %f494; fma.rn.ftz.f32 %f496, %f489, %f466, %f495; mov.f32 %f497, 0f00000000; max.ftz.f32 %f498, %f476, %f497; max.ftz.f32 %f499, %f486, %f497; max.ftz.f32 %f500, %f496, %f497; mul.ftz.f32 %f601, %f500, 0f3C16B90D; mul.ftz.f32 %f600, %f499, 0f3C16B90D; mul.ftz.f32 %f599, %f498, 0f3C16B90D; bra.uni $L__BB88_47; $L__BB88_6: setp.ltu.ftz.f32 %p14, %f5, 0f00000000; @%p14 bra $L__BB88_28; bra.uni $L__BB88_7; $L__BB88_28: setp.lt.ftz.f32 %p36, %f5, 0fBECCCCCD; mov.f32 %f594, 0f00000000; mov.f32 %f595, %f594; mov.f32 %f596, %f594; @%p36 bra $L__BB88_39; fma.rn.ftz.f32 %f71, %f5, 0f40200000, 0f3F800000; add.ftz.f32 %f293, %f584, 0f40490FDB; add.ftz.f32 %f294, %f293, %f202; mov.f32 %f295, 0f40C90FDB; div.approx.ftz.f32 %f296, %f294, %f295; setp.gt.ftz.f32 %p37, %f296, 0f3F800000; add.ftz.f32 %f297, %f296, 0fBF800000; selp.f32 %f72, %f297, %f296, %p37; ld.const.u64 %rd20, [kernel_params+2432]; cvta.to.global.u64 %rd21, %rd20; mul.wide.s32 %rd22, %r27, 96; add.s64 %rd23, %rd21, %rd22; add.s64 %rd4, %rd23, 8; ld.global.v2.u32 {%r53, %r54}, [%rd23+8]; setp.gt.u32 %p38, %r53, 6; @%p38 bra $L__BB88_34; mov.u32 %r55, 1; shl.b32 %r56, %r55, %r53; and.b32 %r57, %r56, 71; setp.ne.s32 %p39, %r57, 0; @%p39 bra $L__BB88_31; bra.uni $L__BB88_34; $L__BB88_31: and.b32 %r58, %r54, -2; setp.eq.s32 %p40, %r58, 2; @%p40 bra $L__BB88_33; bra.uni $L__BB88_32; $L__BB88_33: ld.global.u64 %rd25, [%rd4+-8]; ld.global.u32 %r59, [%rd4+12]; ld.global.u32 %r60, [%rd4+16]; mov.f32 %f299, 0fBF000000; { // callseq 62, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd25; .param .b32 param1; st.param.b32 [param1+0], %r59; .param .b32 param2; st.param.b32 [param2+0], %r60; .param .b32 param3; st.param.f32 [param3+0], %f72; .param .b32 param4; st.param.f32 [param4+0], %f299; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f591, [retval0+0]; ld.param.f32 %f592, [retval0+4]; ld.param.f32 %f593, [retval0+8]; ld.param.f32 %f80, [retval0+12]; } // callseq 62 bra.uni $L__BB88_38; $L__BB88_7: add.ftz.f32 %f203, %f202, 0f3FC90FDB; cos.approx.ftz.f32 %f204, %f203; cos.approx.ftz.f32 %f205, %f201; mul.ftz.f32 %f206, %f205, %f204; sin.approx.ftz.f32 %f207, %f203; mul.ftz.f32 %f208, %f205, %f207; sub.ftz.f32 %f209, %f9, %f206; sub.ftz.f32 %f210, %f7, %f208; sin.approx.ftz.f32 %f211, %f201; sub.ftz.f32 %f212, %f5, %f211; mul.ftz.f32 %f213, %f210, %f210; fma.rn.ftz.f32 %f214, %f209, %f209, %f213; fma.rn.ftz.f32 %f215, %f212, %f212, %f214; sqrt.approx.ftz.f32 %f216, %f215; add.ftz.f32 %f217, %f9, %f206; add.ftz.f32 %f218, %f7, %f208; add.ftz.f32 %f219, %f5, %f211; mul.ftz.f32 %f220, %f218, %f218; fma.rn.ftz.f32 %f221, %f217, %f217, %f220; fma.rn.ftz.f32 %f222, %f219, %f219, %f221; sqrt.approx.ftz.f32 %f223, %f222; abs.ftz.f32 %f25, %f223; abs.ftz.f32 %f26, %f216; setp.eq.ftz.f32 %p15, %f25, 0f00000000; setp.eq.ftz.f32 %p16, %f26, 0f00000000; and.pred %p17, %p15, %p16; mov.b32 %r5, %f223; mov.b32 %r32, %f216; and.b32 %r6, %r32, -2147483648; @%p17 bra $L__BB88_11; bra.uni $L__BB88_8; $L__BB88_11: shr.s32 %r37, %r5, 31; and.b32 %r38, %r37, 1078530011; or.b32 %r39, %r38, %r6; mov.b32 %f585, %r39; bra.uni $L__BB88_12; $L__BB88_42: div.approx.ftz.f32 %f502, %f341, %f124; mul.ftz.f32 %f503, %f502, 0f3FB8AA3B; ex2.approx.ftz.f32 %f504, %f503; fma.rn.ftz.f32 %f505, %f504, %f340, 0f3F800000; mul.ftz.f32 %f506, %f123, %f343; mul.ftz.f32 %f507, %f506, 0f3FB8AA3B; ex2.approx.ftz.f32 %f508, %f507; fma.rn.ftz.f32 %f509, %f508, %f342, 0f3F800000; mul.ftz.f32 %f510, %f125, %f344; fma.rn.ftz.f32 %f511, %f125, %f510, %f509; mul.ftz.f32 %f512, %f505, %f511; mul.ftz.f32 %f129, %f512, %f145; div.approx.ftz.f32 %f513, %f350, %f124; mul.ftz.f32 %f514, %f513, 0f3FB8AA3B; ex2.approx.ftz.f32 %f515, %f514; fma.rn.ftz.f32 %f516, %f515, %f349, 0f3F800000; mul.ftz.f32 %f517, %f123, %f352; mul.ftz.f32 %f518, %f517, 0f3FB8AA3B; ex2.approx.ftz.f32 %f519, %f518; fma.rn.ftz.f32 %f520, %f519, %f351, 0f3F800000; mul.ftz.f32 %f521, %f125, %f353; fma.rn.ftz.f32 %f522, %f125, %f521, %f520; mul.ftz.f32 %f523, %f516, %f522; mul.ftz.f32 %f130, %f523, %f199; div.approx.ftz.f32 %f524, %f201, %f124; mul.ftz.f32 %f525, %f524, 0f3FB8AA3B; ex2.approx.ftz.f32 %f526, %f525; fma.rn.ftz.f32 %f527, %f526, %f200, 0f3F800000; mul.ftz.f32 %f528, %f123, %f19; mul.ftz.f32 %f529, %f528, 0f3FB8AA3B; ex2.approx.ftz.f32 %f530, %f529; fma.rn.ftz.f32 %f531, %f530, %f202, 0f3F800000; mul.ftz.f32 %f532, %f125, %f20; fma.rn.ftz.f32 %f533, %f125, %f532, %f531; mul.ftz.f32 %f534, %f527, %f533; mul.ftz.f32 %f131, %f534, %f144; setp.eq.ftz.f32 %p45, %f130, 0f00000000; mov.f32 %f501, 0f00000000; mov.f32 %f597, %f501; @%p45 bra $L__BB88_44; div.approx.ftz.f32 %f535, %f129, %f130; mul.ftz.f32 %f597, %f131, %f535; $L__BB88_44: setp.eq.ftz.f32 %p46, %f131, 0f00000000; or.pred %p48, %p45, %p46; mov.f32 %f598, %f501; @%p48 bra $L__BB88_46; sub.ftz.f32 %f538, %f148, %f129; sub.ftz.f32 %f539, %f538, %f130; div.approx.ftz.f32 %f540, %f539, %f130; mul.ftz.f32 %f598, %f131, %f540; $L__BB88_46: ld.const.v4.f32 {%f541, %f542, %f543, %f544}, [kernel_params+1136]; mul.ftz.f32 %f548, %f131, %f542; fma.rn.ftz.f32 %f549, %f597, %f541, %f548; fma.rn.ftz.f32 %f550, %f598, %f543, %f549; ld.const.v4.f32 {%f551, %f552, %f553, %f554}, [kernel_params+1152]; mul.ftz.f32 %f558, %f131, %f552; fma.rn.ftz.f32 %f559, %f597, %f551, %f558; fma.rn.ftz.f32 %f560, %f598, %f553, %f559; ld.const.v4.f32 {%f561, %f562, %f563, %f564}, [kernel_params+1168]; mul.ftz.f32 %f568, %f131, %f562; fma.rn.ftz.f32 %f569, %f597, %f561, %f568; fma.rn.ftz.f32 %f570, %f598, %f563, %f569; max.ftz.f32 %f601, %f570, %f501; max.ftz.f32 %f600, %f560, %f501; max.ftz.f32 %f599, %f550, %f501; bra.uni $L__BB88_47; $L__BB88_8: setp.eq.ftz.f32 %p18, %f25, 0f7F800000; setp.eq.ftz.f32 %p19, %f26, 0f7F800000; and.pred %p20, %p18, %p19; @%p20 bra $L__BB88_10; bra.uni $L__BB88_9; $L__BB88_10: setp.lt.s32 %p24, %r5, 0; selp.b32 %r35, 1075235812, 1061752795, %p24; or.b32 %r36, %r35, %r6; mov.b32 %f585, %r36; bra.uni $L__BB88_12; $L__BB88_9: setp.lt.s32 %p21, %r5, 0; min.ftz.f32 %f224, %f26, %f25; max.ftz.f32 %f225, %f26, %f25; div.full.ftz.f32 %f226, %f224, %f225; mul.rn.ftz.f32 %f227, %f226, %f226; mov.f32 %f228, 0fC0B59883; mov.f32 %f229, 0fBF52C7EA; fma.rn.ftz.f32 %f230, %f227, %f229, %f228; mov.f32 %f231, 0fC0D21907; fma.rn.ftz.f32 %f232, %f230, %f227, %f231; mul.ftz.f32 %f233, %f227, %f232; mul.ftz.f32 %f234, %f226, %f233; add.ftz.f32 %f235, %f227, 0f41355DC0; mov.f32 %f236, 0f41E6BD60; fma.rn.ftz.f32 %f237, %f235, %f227, %f236; mov.f32 %f238, 0f419D92C8; fma.rn.ftz.f32 %f239, %f237, %f227, %f238; rcp.approx.ftz.f32 %f240, %f239; fma.rn.ftz.f32 %f241, %f234, %f240, %f226; sub.ftz.f32 %f243, %f168, %f241; setp.gt.ftz.f32 %p22, %f26, %f25; selp.f32 %f244, %f243, %f241, %p22; sub.ftz.f32 %f246, %f171, %f244; selp.f32 %f247, %f246, %f244, %p21; mov.b32 %r33, %f247; or.b32 %r34, %r6, %r33; mov.b32 %f248, %r34; add.ftz.f32 %f249, %f25, %f26; setp.le.ftz.f32 %p23, %f249, 0f7F800000; selp.f32 %f585, %f248, %f249, %p23; $L__BB88_12: sub.ftz.f32 %f31, %f168, %f6; mov.f32 %f251, 0f40000000; div.approx.ftz.f32 %f32, %f19, %f251; add.ftz.f32 %f33, %f585, %f585; setp.lt.ftz.f32 %p25, %f33, %f32; setp.ge.ftz.f32 %p26, %f19, 0f00000000; and.pred %p27, %p26, %p25; @%p27 bra $L__BB88_22; bra.uni $L__BB88_13; $L__BB88_22: sub.ftz.f32 %f264, %f201, %f32; setp.gt.ftz.f32 %p33, %f264, 0f00000000; add.ftz.f32 %f58, %f32, %f201; @%p33 bra $L__BB88_25; bra.uni $L__BB88_23; $L__BB88_25: setp.leu.ftz.f32 %p35, %f58, 0f00000000; @%p35 bra $L__BB88_27; sub.ftz.f32 %f274, %f31, %f201; div.approx.ftz.f32 %f275, %f274, %f19; add.ftz.f32 %f276, %f275, 0f3F000000; sub.ftz.f32 %f277, %f145, %f142; sub.ftz.f32 %f278, %f199, %f143; sub.ftz.f32 %f279, %f200, %f144; fma.rn.ftz.f32 %f280, %f277, %f276, %f142; fma.rn.ftz.f32 %f281, %f278, %f276, %f143; fma.rn.ftz.f32 %f282, %f279, %f276, %f144; mul.ftz.f32 %f587, %f280, %f20; mul.ftz.f32 %f588, %f281, %f20; mul.ftz.f32 %f589, %f282, %f20; bra.uni $L__BB88_27; $L__BB88_13: add.ftz.f32 %f252, %f584, 0f40490FDB; add.ftz.f32 %f253, %f252, %f202; mov.f32 %f254, 0f40C90FDB; div.approx.ftz.f32 %f255, %f253, %f254; div.approx.ftz.f32 %f257, %f31, %f168; mov.f32 %f258, 0f00000000; max.ftz.f32 %f259, %f257, %f258; sqrt.approx.ftz.f32 %f34, %f259; setp.gt.ftz.f32 %p28, %f255, 0f3F800000; add.ftz.f32 %f260, %f255, 0fBF800000; selp.f32 %f35, %f260, %f255, %p28; ld.const.u64 %rd12, [kernel_params+2432]; cvta.to.global.u64 %rd13, %rd12; mul.wide.s32 %rd14, %r27, 96; add.s64 %rd15, %rd13, %rd14; add.s64 %rd3, %rd15, 8; ld.global.v2.u32 {%r41, %r42}, [%rd15+8]; setp.gt.u32 %p29, %r41, 6; @%p29 bra $L__BB88_18; mov.u32 %r43, 1; shl.b32 %r44, %r43, %r41; and.b32 %r45, %r44, 71; setp.ne.s32 %p30, %r45, 0; @%p30 bra $L__BB88_15; bra.uni $L__BB88_18; $L__BB88_15: and.b32 %r46, %r42, -2; setp.eq.s32 %p31, %r46, 2; @%p31 bra $L__BB88_17; bra.uni $L__BB88_16; $L__BB88_17: ld.global.u64 %rd17, [%rd3+-8]; ld.global.u32 %r47, [%rd3+12]; ld.global.u32 %r48, [%rd3+16]; { // callseq 60, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd17; .param .b32 param1; st.param.b32 [param1+0], %r47; .param .b32 param2; st.param.b32 [param2+0], %r48; .param .b32 param3; st.param.f32 [param3+0], %f35; .param .b32 param4; st.param.f32 [param4+0], %f34; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicI6float4ET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f594, [retval0+0]; ld.param.f32 %f595, [retval0+4]; ld.param.f32 %f596, [retval0+8]; ld.param.f32 %f43, [retval0+12]; } // callseq 60 bra.uni $L__BB88_39; $L__BB88_23: setp.leu.ftz.f32 %p34, %f58, 0f00000000; @%p34 bra $L__BB88_27; div.approx.ftz.f32 %f266, %f31, %f58; sub.ftz.f32 %f267, %f145, %f142; sub.ftz.f32 %f268, %f199, %f143; sub.ftz.f32 %f269, %f200, %f144; fma.rn.ftz.f32 %f270, %f267, %f266, %f142; fma.rn.ftz.f32 %f271, %f268, %f266, %f143; fma.rn.ftz.f32 %f272, %f269, %f266, %f144; mul.ftz.f32 %f587, %f270, %f20; mul.ftz.f32 %f588, %f271, %f20; mul.ftz.f32 %f589, %f272, %f20; $L__BB88_27: div.approx.ftz.f32 %f283, %f33, %f32; mul.ftz.f32 %f284, %f283, %f283; sub.ftz.f32 %f286, %f148, %f284; sqrt.approx.ftz.f32 %f287, %f286; sub.ftz.f32 %f288, %f148, %f287; fma.rn.ftz.f32 %f289, %f288, 0fBF19999A, 0f3F800000; mul.ftz.f32 %f594, %f587, %f289; mul.ftz.f32 %f595, %f588, %f289; mul.ftz.f32 %f596, %f589, %f289; bra.uni $L__BB88_39; $L__BB88_32: ld.global.u64 %rd24, [%rd4+-8]; mov.f32 %f298, 0fBF000000; tex.2d.v4.f32.f32 {%f591, %f592, %f593, %f76}, [%rd24, {%f72, %f298}]; bra.uni $L__BB88_38; $L__BB88_16: ld.global.u64 %rd16, [%rd3+-8]; tex.2d.v4.f32.f32 {%f594, %f595, %f596, %f39}, [%rd16, {%f35, %f34}]; bra.uni $L__BB88_39; $L__BB88_34: and.b32 %r61, %r54, -2; setp.eq.s32 %p41, %r61, 2; @%p41 bra $L__BB88_36; bra.uni $L__BB88_35; $L__BB88_36: ld.global.u64 %rd27, [%rd4+-8]; ld.global.u32 %r62, [%rd4+12]; ld.global.u32 %r63, [%rd4+16]; mov.f32 %f304, 0fBF000000; { // callseq 63, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd27; .param .b32 param1; st.param.b32 [param1+0], %r62; .param .b32 param2; st.param.b32 [param2+0], %r63; .param .b32 param3; st.param.f32 [param3+0], %f72; .param .b32 param4; st.param.f32 [param4+0], %f304; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f591, [retval0+0]; } // callseq 63 bra.uni $L__BB88_37; $L__BB88_18: and.b32 %r49, %r42, -2; setp.eq.s32 %p32, %r49, 2; @%p32 bra $L__BB88_20; bra.uni $L__BB88_19; $L__BB88_20: ld.global.u64 %rd19, [%rd3+-8]; ld.global.u32 %r50, [%rd3+12]; ld.global.u32 %r51, [%rd3+16]; { // callseq 61, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.b32 [param1+0], %r50; .param .b32 param2; st.param.b32 [param2+0], %r51; .param .b32 param3; st.param.f32 [param3+0], %f35; .param .b32 param4; st.param.f32 [param4+0], %f34; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_bicubicIfET_RK11TextureInfoff, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f594, [retval0+0]; } // callseq 61 bra.uni $L__BB88_21; $L__BB88_35: ld.global.u64 %rd26, [%rd4+-8]; mov.f32 %f300, 0fBF000000; tex.2d.v4.f32.f32 {%f591, %f301, %f302, %f303}, [%rd26, {%f72, %f300}]; $L__BB88_37: mov.f32 %f592, %f591; mov.f32 %f593, %f591; $L__BB88_38: mul.ftz.f32 %f305, %f71, %f71; mul.ftz.f32 %f306, %f71, %f305; mul.ftz.f32 %f594, %f306, %f591; mul.ftz.f32 %f595, %f306, %f592; mul.ftz.f32 %f596, %f306, %f593; $L__BB88_39: ld.const.v4.f32 {%f307, %f308, %f309, %f310}, [kernel_params+1136]; mul.ftz.f32 %f314, %f595, %f308; fma.rn.ftz.f32 %f315, %f594, %f307, %f314; fma.rn.ftz.f32 %f316, %f596, %f309, %f315; ld.const.v4.f32 {%f317, %f318, %f319, %f320}, [kernel_params+1152]; mul.ftz.f32 %f324, %f595, %f318; fma.rn.ftz.f32 %f325, %f594, %f317, %f324; fma.rn.ftz.f32 %f326, %f596, %f319, %f325; ld.const.v4.f32 {%f327, %f328, %f329, %f330}, [kernel_params+1168]; mul.ftz.f32 %f334, %f595, %f328; fma.rn.ftz.f32 %f335, %f594, %f327, %f334; fma.rn.ftz.f32 %f336, %f596, %f329, %f335; mov.f32 %f337, 0f00000000; max.ftz.f32 %f601, %f336, %f337; max.ftz.f32 %f600, %f326, %f337; max.ftz.f32 %f599, %f316, %f337; add.s32 %r64, %r16, 3; $L__BB88_47: mul.wide.u32 %rd28, %r14, 4; add.s64 %rd29, %rd1, %rd28; st.local.f32 [%rd29], %f599; st.local.f32 [%rd29+4], %f600; st.local.f32 [%rd29+8], %f601; st.param.b32 [func_retval0+0], %r64; ret; $L__BB88_19: ld.global.u64 %rd18, [%rd3+-8]; tex.2d.v4.f32.f32 {%f594, %f261, %f262, %f263}, [%rd18, {%f35, %f34}]; $L__BB88_21: mov.f32 %f595, %f594; mov.f32 %f596, %f594; bra.uni $L__BB88_39; } .func (.param .b32 func_retval0) _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i( .param .b64 _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0, .param .b32 _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1, .param .b64 _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2, .param .align 4 .b8 _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3[16], .param .b32 _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4 ) { .reg .pred %p<134>; .reg .b16 %rs<16>; .reg .f32 %f<1303>; .reg .b32 %r<133>; .reg .b64 %rd<73>; ld.param.u64 %rd19, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0]; ld.param.u32 %r38, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1]; ld.param.u32 %r132, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; ld.param.u32 %r42, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+12]; ld.param.u32 %r40, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+4]; cvta.to.local.u64 %rd1, %rd19; setp.gt.s32 %p2, %r40, 3; @%p2 bra $L__BB89_12; setp.gt.s32 %p8, %r40, 1; @%p8 bra $L__BB89_6; setp.eq.s32 %p11, %r40, 0; @%p11 bra $L__BB89_108; setp.eq.s32 %p12, %r40, 1; @%p12 bra $L__BB89_4; bra.uni $L__BB89_116; $L__BB89_4: ld.local.v4.f32 {%f1300, %f1301, %f1302, %f1064}, [%rd1]; setp.eq.s32 %p126, %r42, 0; @%p126 bra $L__BB89_104; ld.const.u64 %rd52, [kernel_params+2400]; cvta.to.global.u64 %rd53, %rd52; mul.wide.s32 %rd54, %r132, 16; add.s64 %rd55, %rd53, %rd54; ld.global.v4.f32 {%f1065, %f1066, %f1067, %f1068}, [%rd55]; ld.global.v4.f32 {%f1073, %f1074, %f1075, %f1076}, [%rd55+16]; ld.global.v4.f32 {%f1081, %f1082, %f1083, %f1084}, [%rd55+32]; add.s32 %r132, %r132, 3; mul.ftz.f32 %f1089, %f1301, %f1066; fma.rn.ftz.f32 %f1090, %f1300, %f1065, %f1089; fma.rn.ftz.f32 %f1091, %f1302, %f1067, %f1090; add.ftz.f32 %f187, %f1091, %f1068; mul.ftz.f32 %f1092, %f1301, %f1074; fma.rn.ftz.f32 %f1093, %f1300, %f1073, %f1092; fma.rn.ftz.f32 %f1094, %f1302, %f1075, %f1093; add.ftz.f32 %f188, %f1094, %f1076; mul.ftz.f32 %f1095, %f1301, %f1082; fma.rn.ftz.f32 %f1096, %f1300, %f1081, %f1095; fma.rn.ftz.f32 %f1097, %f1302, %f1083, %f1096; add.ftz.f32 %f1302, %f1097, %f1084; mov.f32 %f1300, %f187; mov.f32 %f1301, %f188; bra.uni $L__BB89_116; $L__BB89_12: setp.gt.s32 %p3, %r40, 5; @%p3 bra $L__BB89_17; setp.eq.s32 %p6, %r40, 4; @%p6 bra $L__BB89_38; setp.eq.s32 %p7, %r40, 5; @%p7 bra $L__BB89_15; bra.uni $L__BB89_116; $L__BB89_15: ld.local.u32 %r23, [%rd1+76]; setp.eq.s32 %p30, %r23, -1; mov.f32 %f1300, 0f00000000; mov.f32 %f1301, %f1300; mov.f32 %f1302, %f1300; @%p30 bra $L__BB89_116; ld.const.u64 %rd45, [kernel_params+2160]; cvta.to.global.u64 %rd46, %rd45; mul.wide.s32 %rd47, %r23, 208; add.s64 %rd48, %rd46, %rd47; ld.global.v4.f32 {%f1300, %f1301, %f1302, %f324}, [%rd48+128]; bra.uni $L__BB89_116; $L__BB89_6: setp.eq.s32 %p9, %r40, 2; @%p9 bra $L__BB89_101; setp.eq.s32 %p10, %r40, 3; @%p10 bra $L__BB89_8; bra.uni $L__BB89_116; $L__BB89_8: and.b32 %r61, %r38, 1; setp.eq.b32 %p32, %r61, 1; mov.pred %p33, 0; xor.pred %p34, %p32, %p33; not.pred %p35, %p34; @%p35 bra $L__BB89_41; bra.uni $L__BB89_9; $L__BB89_41: ld.const.u32 %r131, [kernel_params+16]; bra.uni $L__BB89_42; $L__BB89_17: setp.eq.s32 %p4, %r40, 6; @%p4 bra $L__BB89_36; setp.ne.s32 %p5, %r40, 7; @%p5 bra $L__BB89_116; ld.local.v4.f32 {%f1300, %f1301, %f1302, %f215}, [%rd1]; ld.local.u32 %r1, [%rd1+76]; setp.eq.s32 %p13, %r1, -1; @%p13 bra $L__BB89_116; ld.local.s32 %rd3, [%rd1+60]; ld.const.u64 %rd22, [kernel_params+2160]; cvta.to.global.u64 %rd23, %rd22; mul.wide.s32 %rd24, %r1, 208; add.s64 %rd25, %rd23, %rd24; add.s64 %rd4, %rd25, 164; ld.global.u32 %r2, [%rd25+164]; ld.local.u32 %r46, [%rd1+64]; and.b32 %r47, %r46, 1; setp.eq.b32 %p14, %r47, 1; not.pred %p15, %p14; mov.u32 %r125, 0; setp.eq.s64 %p16, %rd3, -1; or.pred %p17, %p16, %p15; @%p17 bra $L__BB89_22; ld.const.u64 %rd26, [kernel_params+2240]; cvta.to.global.u64 %rd27, %rd26; shl.b64 %rd28, %rd3, 2; add.s64 %rd29, %rd27, %rd28; ld.global.u32 %r48, [%rd29]; setp.ne.s32 %p18, %r48, -1; selp.u32 %r125, 1, 0, %p18; $L__BB89_22: ld.const.u64 %rd30, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd30; add.s32 %r127, %r125, %r2; mul.wide.u32 %rd31, %r127, 16; add.s64 %rd32, %rd5, %rd31; ld.global.u16 %rs14, [%rd32+12]; ld.global.u32 %r126, [%rd32+8]; ld.global.u64 %rd72, [%rd32]; setp.eq.s64 %p19, %rd72, 8; @%p19 bra $L__BB89_27; $L__BB89_23: setp.eq.s64 %p20, %rd72, 0; @%p20 bra $L__BB89_25; add.s32 %r127, %r127, 2; bra.uni $L__BB89_26; $L__BB89_25: setp.eq.s16 %p21, %rs14, 0; mov.u32 %r130, -1; mov.u32 %r127, %r126; @%p21 bra $L__BB89_31; $L__BB89_26: mul.wide.u32 %rd33, %r127, 16; add.s64 %rd34, %rd5, %rd33; ld.global.u16 %rs14, [%rd34+12]; ld.global.u32 %r126, [%rd34+8]; ld.global.u64 %rd72, [%rd34]; setp.ne.s64 %p22, %rd72, 8; @%p22 bra $L__BB89_23; $L__BB89_27: cvt.u32.u64 %r53, %rd3; setp.ne.s32 %p23, %r53, -1; @%p23 bra $L__BB89_30; mov.u32 %r130, -1; add.s16 %rs7, %rs14, -1; setp.lt.u16 %p24, %rs7, 2; @%p24 bra $L__BB89_30; setp.ne.s16 %p25, %rs14, 1024; @%p25 bra $L__BB89_31; $L__BB89_30: setp.eq.s16 %p26, %rs14, 0; selp.b32 %r130, -1, %r126, %p26; $L__BB89_31: ld.local.u8 %rs8, [%rd1+56]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p27, %rs9, 0; @%p27 bra $L__BB89_33; bra.uni $L__BB89_32; $L__BB89_33: ld.global.v4.f32 {%f249, %f250, %f251, %f252}, [%rd4+-116]; ld.global.v4.f32 {%f257, %f258, %f259, %f260}, [%rd4+-100]; ld.global.v4.f32 {%f265, %f266, %f267, %f268}, [%rd4+-84]; mul.ftz.f32 %f273, %f1301, %f250; fma.rn.ftz.f32 %f274, %f1300, %f249, %f273; fma.rn.ftz.f32 %f275, %f1302, %f251, %f274; add.ftz.f32 %f7, %f252, %f275; mul.ftz.f32 %f276, %f1301, %f258; fma.rn.ftz.f32 %f277, %f1300, %f257, %f276; fma.rn.ftz.f32 %f278, %f1302, %f259, %f277; add.ftz.f32 %f8, %f260, %f278; mul.ftz.f32 %f279, %f1301, %f266; fma.rn.ftz.f32 %f280, %f1300, %f265, %f279; fma.rn.ftz.f32 %f281, %f1302, %f267, %f280; add.ftz.f32 %f1302, %f268, %f281; mov.f32 %f1300, %f7; mov.f32 %f1301, %f8; bra.uni $L__BB89_34; $L__BB89_108: ld.local.f32 %f1300, [%rd1+12]; ld.local.f32 %f1301, [%rd1+16]; ld.local.f32 %f1302, [%rd1+20]; ld.local.u8 %rs12, [%rd1+56]; and.b16 %rs13, %rs12, 2; setp.eq.s16 %p129, %rs13, 0; ld.local.s32 %rd18, [%rd1+76]; @%p129 bra $L__BB89_112; bra.uni $L__BB89_109; $L__BB89_112: cvt.u32.u64 %r120, %rd18; setp.ne.s32 %p132, %r120, -1; @%p132 bra $L__BB89_115; bra.uni $L__BB89_113; $L__BB89_115: ld.const.u64 %rd64, [kernel_params+2160]; cvta.to.global.u64 %rd65, %rd64; mul.lo.s64 %rd66, %rd18, 208; add.s64 %rd67, %rd65, %rd66; ld.global.v4.f32 {%f1232, %f1233, %f1234, %f1235}, [%rd67]; ld.global.v4.f32 {%f1239, %f1240, %f1241, %f1242}, [%rd67+16]; ld.global.v4.f32 {%f1246, %f1247, %f1248, %f1249}, [%rd67+32]; mul.ftz.f32 %f1253, %f1301, %f1239; fma.rn.ftz.f32 %f1254, %f1300, %f1232, %f1253; fma.rn.ftz.f32 %f1255, %f1302, %f1246, %f1254; mul.ftz.f32 %f1256, %f1301, %f1240; fma.rn.ftz.f32 %f1257, %f1300, %f1233, %f1256; fma.rn.ftz.f32 %f1258, %f1302, %f1247, %f1257; mul.ftz.f32 %f1259, %f1301, %f1241; fma.rn.ftz.f32 %f1260, %f1300, %f1234, %f1259; fma.rn.ftz.f32 %f1261, %f1302, %f1248, %f1260; mul.ftz.f32 %f1262, %f1258, %f1258; fma.rn.ftz.f32 %f1263, %f1255, %f1255, %f1262; fma.rn.ftz.f32 %f1264, %f1261, %f1261, %f1263; sqrt.approx.ftz.f32 %f1265, %f1264; div.approx.ftz.f32 %f1300, %f1255, %f1265; div.approx.ftz.f32 %f1301, %f1258, %f1265; div.approx.ftz.f32 %f1302, %f1261, %f1265; bra.uni $L__BB89_116; $L__BB89_38: ld.local.u32 %r60, [%rd1+76]; setp.eq.s32 %p31, %r60, -1; @%p31 bra $L__BB89_40; ld.local.f32 %f325, [%rd1+36]; ld.local.f32 %f326, [%rd1+12]; ld.local.f32 %f327, [%rd1+40]; ld.local.f32 %f328, [%rd1+16]; mul.ftz.f32 %f329, %f328, %f327; fma.rn.ftz.f32 %f330, %f326, %f325, %f329; ld.local.f32 %f331, [%rd1+44]; ld.local.f32 %f332, [%rd1+20]; fma.rn.ftz.f32 %f333, %f332, %f331, %f330; add.ftz.f32 %f334, %f333, %f333; mul.ftz.f32 %f335, %f326, %f334; mul.ftz.f32 %f336, %f328, %f334; mul.ftz.f32 %f337, %f332, %f334; sub.ftz.f32 %f1300, %f335, %f325; sub.ftz.f32 %f1301, %f336, %f327; sub.ftz.f32 %f1302, %f337, %f331; bra.uni $L__BB89_116; $L__BB89_101: ld.const.v4.f32 {%f1003, %f1004, %f1005, %f1006}, [kernel_params+688]; ld.const.v4.f32 {%f1007, %f1008, %f1009, %f1010}, [kernel_params+704]; ld.const.v4.f32 {%f1011, %f1012, %f1013, %f1014}, [kernel_params+720]; ld.local.u32 %r117, [%rd1+76]; setp.eq.s32 %p125, %r117, -1; @%p125 bra $L__BB89_103; ld.local.v4.f32 {%f1015, %f1016, %f1017, %f1018}, [%rd1]; mul.ftz.f32 %f1022, %f1004, %f1016; fma.rn.ftz.f32 %f1023, %f1003, %f1015, %f1022; fma.rn.ftz.f32 %f1024, %f1005, %f1017, %f1023; add.ftz.f32 %f1300, %f1006, %f1024; mul.ftz.f32 %f1025, %f1008, %f1016; fma.rn.ftz.f32 %f1026, %f1007, %f1015, %f1025; fma.rn.ftz.f32 %f1027, %f1009, %f1017, %f1026; add.ftz.f32 %f1301, %f1010, %f1027; mul.ftz.f32 %f1028, %f1012, %f1016; fma.rn.ftz.f32 %f1029, %f1011, %f1015, %f1028; fma.rn.ftz.f32 %f1030, %f1013, %f1017, %f1029; add.ftz.f32 %f1302, %f1014, %f1030; bra.uni $L__BB89_116; $L__BB89_36: ld.local.u32 %r22, [%rd1+76]; setp.eq.s32 %p29, %r22, -1; mov.f32 %f1302, 0f00000000; mov.f32 %f1300, %f1302; mov.f32 %f1301, %f1302; @%p29 bra $L__BB89_116; ld.const.u64 %rd41, [kernel_params+2160]; cvta.to.global.u64 %rd42, %rd41; mul.wide.s32 %rd43, %r22, 208; add.s64 %rd44, %rd42, %rd43; ld.global.f32 %f1301, [%rd44+144]; ld.global.f32 %f1300, [%rd44+140]; bra.uni $L__BB89_116; $L__BB89_109: cvt.u32.u64 %r118, %rd18; setp.ne.s32 %p130, %r118, -1; @%p130 bra $L__BB89_111; ld.local.u32 %r119, [%rd1+64]; setp.ne.s32 %p131, %r119, 32; @%p131 bra $L__BB89_116; $L__BB89_111: ld.local.v4.f32 {%f1164, %f1165, %f1166, %f1167}, [%rd1+144]; ld.local.v4.f32 {%f1171, %f1172, %f1173, %f1174}, [%rd1+160]; ld.local.v4.f32 {%f1178, %f1179, %f1180, %f1181}, [%rd1+176]; mul.ftz.f32 %f1185, %f1301, %f1171; fma.rn.ftz.f32 %f1186, %f1300, %f1164, %f1185; fma.rn.ftz.f32 %f1187, %f1302, %f1178, %f1186; mul.ftz.f32 %f1188, %f1301, %f1172; fma.rn.ftz.f32 %f1189, %f1300, %f1165, %f1188; fma.rn.ftz.f32 %f1190, %f1302, %f1179, %f1189; mul.ftz.f32 %f1191, %f1301, %f1173; fma.rn.ftz.f32 %f1192, %f1300, %f1166, %f1191; fma.rn.ftz.f32 %f1193, %f1302, %f1180, %f1192; mul.ftz.f32 %f1194, %f1190, %f1190; fma.rn.ftz.f32 %f1195, %f1187, %f1187, %f1194; fma.rn.ftz.f32 %f1196, %f1193, %f1193, %f1195; sqrt.approx.ftz.f32 %f1197, %f1196; div.approx.ftz.f32 %f1300, %f1187, %f1197; div.approx.ftz.f32 %f1301, %f1190, %f1197; div.approx.ftz.f32 %f1302, %f1193, %f1197; bra.uni $L__BB89_116; $L__BB89_9: ld.local.u32 %r62, [%rd1+76]; setp.ne.s32 %p36, %r62, -1; ld.const.u32 %r131, [kernel_params+16]; setp.ne.s32 %p37, %r131, 1; or.pred %p38, %p36, %p37; @%p38 bra $L__BB89_42; ld.local.v4.f32 {%f341, %f342, %f343, %f344}, [%rd1+240]; ld.const.v4.f32 {%f348, %f349, %f350, %f351}, [kernel_params+624]; ld.const.v4.f32 {%f356, %f357, %f358, %f359}, [kernel_params+640]; ld.const.v4.f32 {%f372, %f373, %f374, %f375}, [kernel_params+672]; mul.ftz.f32 %f380, %f342, %f349; fma.rn.ftz.f32 %f381, %f341, %f348, %f380; fma.rn.ftz.f32 %f382, %f343, %f350, %f351; add.ftz.f32 %f40, %f381, %f382; mul.ftz.f32 %f383, %f342, %f357; fma.rn.ftz.f32 %f384, %f341, %f356, %f383; fma.rn.ftz.f32 %f385, %f343, %f358, %f359; add.ftz.f32 %f41, %f384, %f385; mul.ftz.f32 %f389, %f342, %f373; fma.rn.ftz.f32 %f390, %f341, %f372, %f389; fma.rn.ftz.f32 %f391, %f343, %f374, %f375; add.ftz.f32 %f43, %f390, %f391; setp.eq.ftz.f32 %p39, %f43, 0f00000000; mov.f32 %f1302, 0f00000000; mov.f32 %f1300, %f1302; mov.f32 %f1301, %f1302; @%p39 bra $L__BB89_116; div.approx.ftz.f32 %f1301, %f41, %f43; div.approx.ftz.f32 %f1300, %f40, %f43; bra.uni $L__BB89_116; $L__BB89_42: ld.local.v4.f32 {%f1281, %f1282, %f1283, %f396}, [%rd1]; setp.eq.s32 %p40, %r131, 2; @%p40 bra $L__BB89_47; bra.uni $L__BB89_43; $L__BB89_47: ld.const.v4.f32 {%f459, %f460, %f461, %f462}, [kernel_params+688]; ld.const.v4.f32 {%f463, %f464, %f465, %f466}, [kernel_params+704]; ld.const.v4.f32 {%f467, %f468, %f469, %f470}, [kernel_params+720]; ld.local.u32 %r64, [%rd1+76]; setp.eq.s32 %p45, %r64, -1; mul.ftz.f32 %f471, %f1282, %f460; fma.rn.ftz.f32 %f472, %f1281, %f459, %f471; fma.rn.ftz.f32 %f80, %f1283, %f461, %f472; mul.ftz.f32 %f473, %f1282, %f464; fma.rn.ftz.f32 %f474, %f1281, %f463, %f473; fma.rn.ftz.f32 %f81, %f1283, %f465, %f474; mul.ftz.f32 %f475, %f1282, %f468; fma.rn.ftz.f32 %f476, %f1281, %f467, %f475; fma.rn.ftz.f32 %f82, %f1283, %f469, %f476; @%p45 bra $L__BB89_49; add.ftz.f32 %f477, %f462, %f80; add.ftz.f32 %f478, %f466, %f81; add.ftz.f32 %f479, %f470, %f82; mul.ftz.f32 %f480, %f478, %f478; fma.rn.ftz.f32 %f481, %f477, %f477, %f480; fma.rn.ftz.f32 %f482, %f479, %f479, %f481; sqrt.approx.ftz.f32 %f483, %f482; div.approx.ftz.f32 %f89, %f477, %f483; div.approx.ftz.f32 %f90, %f478, %f483; div.approx.ftz.f32 %f91, %f479, %f483; bra.uni $L__BB89_50; $L__BB89_40: ld.local.f32 %f1300, [%rd1+36]; ld.local.f32 %f1301, [%rd1+40]; ld.local.f32 %f1302, [%rd1+44]; bra.uni $L__BB89_116; $L__BB89_103: ld.const.v4.f32 {%f1031, %f1032, %f1033, %f1034}, [kernel_params+96]; ld.const.v4.f32 {%f1036, %f1037, %f1038, %f1039}, [kernel_params+112]; ld.const.v4.f32 {%f1041, %f1042, %f1043, %f1044}, [kernel_params+128]; ld.local.f32 %f1046, [%rd1]; add.ftz.f32 %f1047, %f1046, %f1034; ld.local.f32 %f1048, [%rd1+4]; add.ftz.f32 %f1049, %f1039, %f1048; ld.local.f32 %f1050, [%rd1+8]; add.ftz.f32 %f1051, %f1044, %f1050; mul.ftz.f32 %f1052, %f1004, %f1049; fma.rn.ftz.f32 %f1053, %f1003, %f1047, %f1052; fma.rn.ftz.f32 %f1054, %f1005, %f1051, %f1053; add.ftz.f32 %f1300, %f1006, %f1054; mul.ftz.f32 %f1055, %f1008, %f1049; fma.rn.ftz.f32 %f1056, %f1007, %f1047, %f1055; fma.rn.ftz.f32 %f1057, %f1009, %f1051, %f1056; add.ftz.f32 %f1301, %f1010, %f1057; mul.ftz.f32 %f1058, %f1012, %f1049; fma.rn.ftz.f32 %f1059, %f1011, %f1047, %f1058; fma.rn.ftz.f32 %f1060, %f1013, %f1051, %f1059; add.ftz.f32 %f1302, %f1014, %f1060; bra.uni $L__BB89_116; $L__BB89_104: ld.local.u32 %r36, [%rd1+76]; setp.eq.s32 %p127, %r36, -1; @%p127 bra $L__BB89_116; ld.local.u8 %rs10, [%rd1+56]; and.b16 %rs11, %rs10, 2; setp.eq.s16 %p128, %rs11, 0; @%p128 bra $L__BB89_107; bra.uni $L__BB89_106; $L__BB89_107: ld.const.u64 %rd56, [kernel_params+2160]; cvta.to.global.u64 %rd57, %rd56; mul.wide.s32 %rd58, %r36, 208; add.s64 %rd59, %rd57, %rd58; ld.global.v4.f32 {%f1131, %f1132, %f1133, %f1134}, [%rd59+48]; ld.global.v4.f32 {%f1139, %f1140, %f1141, %f1142}, [%rd59+64]; ld.global.v4.f32 {%f1147, %f1148, %f1149, %f1150}, [%rd59+80]; mul.ftz.f32 %f1155, %f1301, %f1132; fma.rn.ftz.f32 %f1156, %f1300, %f1131, %f1155; fma.rn.ftz.f32 %f1157, %f1302, %f1133, %f1156; add.ftz.f32 %f193, %f1134, %f1157; mul.ftz.f32 %f1158, %f1301, %f1140; fma.rn.ftz.f32 %f1159, %f1300, %f1139, %f1158; fma.rn.ftz.f32 %f1160, %f1302, %f1141, %f1159; add.ftz.f32 %f194, %f1142, %f1160; mul.ftz.f32 %f1161, %f1301, %f1148; fma.rn.ftz.f32 %f1162, %f1300, %f1147, %f1161; fma.rn.ftz.f32 %f1163, %f1302, %f1149, %f1162; add.ftz.f32 %f1302, %f1150, %f1163; mov.f32 %f1300, %f193; mov.f32 %f1301, %f194; bra.uni $L__BB89_116; $L__BB89_43: ld.local.u32 %r63, [%rd1+76]; setp.ne.s32 %p41, %r63, -1; setp.ne.s32 %p42, %r131, 0; or.pred %p43, %p41, %p42; @%p43 bra $L__BB89_45; ld.const.v4.f32 {%f397, %f398, %f399, %f400}, [kernel_params+96]; ld.const.v4.f32 {%f402, %f403, %f404, %f405}, [kernel_params+112]; ld.const.v4.f32 {%f407, %f408, %f409, %f410}, [kernel_params+128]; add.ftz.f32 %f1281, %f1281, %f400; add.ftz.f32 %f1282, %f1282, %f405; add.ftz.f32 %f1283, %f1283, %f410; $L__BB89_45: ld.const.v4.f32 {%f415, %f416, %f417, %f418}, [kernel_params+624]; ld.const.v4.f32 {%f423, %f424, %f425, %f426}, [kernel_params+640]; ld.const.v4.f32 {%f439, %f440, %f441, %f442}, [kernel_params+672]; mul.ftz.f32 %f447, %f1282, %f416; fma.rn.ftz.f32 %f448, %f1281, %f415, %f447; fma.rn.ftz.f32 %f449, %f1283, %f417, %f418; add.ftz.f32 %f61, %f448, %f449; mul.ftz.f32 %f450, %f1282, %f424; fma.rn.ftz.f32 %f451, %f1281, %f423, %f450; fma.rn.ftz.f32 %f452, %f1283, %f425, %f426; add.ftz.f32 %f62, %f451, %f452; mul.ftz.f32 %f456, %f1282, %f440; fma.rn.ftz.f32 %f457, %f1281, %f439, %f456; fma.rn.ftz.f32 %f458, %f1283, %f441, %f442; add.ftz.f32 %f64, %f457, %f458; setp.eq.ftz.f32 %p44, %f64, 0f00000000; mov.f32 %f1300, 0f00000000; mov.f32 %f1301, %f1300; @%p44 bra $L__BB89_100; div.approx.ftz.f32 %f1301, %f62, %f64; div.approx.ftz.f32 %f1300, %f61, %f64; bra.uni $L__BB89_100; $L__BB89_32: ld.local.v4.f32 {%f216, %f217, %f218, %f219}, [%rd1+192]; mul.ftz.f32 %f224, %f1301, %f217; fma.rn.ftz.f32 %f225, %f1300, %f216, %f224; fma.rn.ftz.f32 %f226, %f1302, %f218, %f225; add.ftz.f32 %f4, %f219, %f226; ld.local.v4.f32 {%f227, %f228, %f229, %f230}, [%rd1+208]; mul.ftz.f32 %f235, %f1301, %f228; fma.rn.ftz.f32 %f236, %f1300, %f227, %f235; fma.rn.ftz.f32 %f237, %f1302, %f229, %f236; add.ftz.f32 %f5, %f230, %f237; ld.local.v4.f32 {%f238, %f239, %f240, %f241}, [%rd1+224]; mul.ftz.f32 %f246, %f1301, %f239; fma.rn.ftz.f32 %f247, %f1300, %f238, %f246; fma.rn.ftz.f32 %f248, %f1302, %f240, %f247; add.ftz.f32 %f1302, %f241, %f248; mov.f32 %f1300, %f4; mov.f32 %f1301, %f5; $L__BB89_34: setp.eq.s32 %p28, %r130, -1; @%p28 bra $L__BB89_116; ld.const.u64 %rd37, [kernel_params+2344]; cvta.to.global.u64 %rd38, %rd37; mul.wide.s32 %rd39, %r130, 16; add.s64 %rd40, %rd38, %rd39; ld.global.v4.f32 {%f282, %f283, %f284, %f285}, [%rd40]; ld.global.v4.f32 {%f290, %f291, %f292, %f293}, [%rd40+16]; ld.global.v4.f32 {%f298, %f299, %f300, %f301}, [%rd40+32]; mul.ftz.f32 %f306, %f1301, %f283; fma.rn.ftz.f32 %f307, %f1300, %f282, %f306; fma.rn.ftz.f32 %f308, %f1302, %f284, %f307; add.ftz.f32 %f13, %f285, %f308; mul.ftz.f32 %f309, %f1301, %f291; fma.rn.ftz.f32 %f310, %f1300, %f290, %f309; fma.rn.ftz.f32 %f311, %f1302, %f292, %f310; add.ftz.f32 %f14, %f293, %f311; mul.ftz.f32 %f312, %f1301, %f299; fma.rn.ftz.f32 %f313, %f1300, %f298, %f312; fma.rn.ftz.f32 %f314, %f1302, %f300, %f313; add.ftz.f32 %f1302, %f301, %f314; mov.f32 %f1300, %f13; mov.f32 %f1301, %f14; bra.uni $L__BB89_116; $L__BB89_113: ld.local.u32 %r121, [%rd1+64]; setp.ne.s32 %p133, %r121, 32; @%p133 bra $L__BB89_116; ld.const.u64 %rd60, [kernel_params+2368]; cvta.to.global.u64 %rd61, %rd60; ld.local.u32 %r122, [%rd1+80]; mul.wide.s32 %rd62, %r122, 192; add.s64 %rd63, %rd61, %rd62; ld.global.v4.f32 {%f1198, %f1199, %f1200, %f1201}, [%rd63+48]; ld.global.v4.f32 {%f1205, %f1206, %f1207, %f1208}, [%rd63+64]; ld.global.v4.f32 {%f1212, %f1213, %f1214, %f1215}, [%rd63+80]; mul.ftz.f32 %f1219, %f1301, %f1205; fma.rn.ftz.f32 %f1220, %f1300, %f1198, %f1219; fma.rn.ftz.f32 %f1221, %f1302, %f1212, %f1220; mul.ftz.f32 %f1222, %f1301, %f1206; fma.rn.ftz.f32 %f1223, %f1300, %f1199, %f1222; fma.rn.ftz.f32 %f1224, %f1302, %f1213, %f1223; mul.ftz.f32 %f1225, %f1301, %f1207; fma.rn.ftz.f32 %f1226, %f1300, %f1200, %f1225; fma.rn.ftz.f32 %f1227, %f1302, %f1214, %f1226; mul.ftz.f32 %f1228, %f1224, %f1224; fma.rn.ftz.f32 %f1229, %f1221, %f1221, %f1228; fma.rn.ftz.f32 %f1230, %f1227, %f1227, %f1229; sqrt.approx.ftz.f32 %f1231, %f1230; div.approx.ftz.f32 %f1300, %f1221, %f1231; div.approx.ftz.f32 %f1301, %f1224, %f1231; div.approx.ftz.f32 %f1302, %f1227, %f1231; bra.uni $L__BB89_116; $L__BB89_49: mul.ftz.f32 %f484, %f81, %f81; fma.rn.ftz.f32 %f485, %f80, %f80, %f484; fma.rn.ftz.f32 %f486, %f82, %f82, %f485; sqrt.approx.ftz.f32 %f487, %f486; div.approx.ftz.f32 %f89, %f80, %f487; div.approx.ftz.f32 %f90, %f81, %f487; div.approx.ftz.f32 %f91, %f82, %f487; $L__BB89_50: ld.const.u32 %r65, [kernel_params+20]; mov.f32 %f488, 0fBF800000; max.ftz.f32 %f489, %f89, %f488; mov.f32 %f490, 0f3F800000; min.ftz.f32 %f491, %f489, %f490; abs.ftz.f32 %f492, %f491; sub.ftz.f32 %f493, %f490, %f492; mul.ftz.f32 %f494, %f493, 0f3F000000; sqrt.approx.ftz.f32 %f495, %f494; setp.gt.ftz.f32 %p46, %f492, 0f3F11EB85; selp.f32 %f496, %f495, %f492, %p46; mul.ftz.f32 %f497, %f496, %f496; mov.f32 %f498, 0f3C94D2E9; mov.f32 %f499, 0f3D53F941; fma.rn.ftz.f32 %f500, %f499, %f497, %f498; mov.f32 %f501, 0f3D3F841F; fma.rn.ftz.f32 %f502, %f500, %f497, %f501; mov.f32 %f503, 0f3D994929; fma.rn.ftz.f32 %f504, %f502, %f497, %f503; mov.f32 %f505, 0f3E2AAB94; fma.rn.ftz.f32 %f506, %f504, %f497, %f505; mul.ftz.f32 %f507, %f497, %f506; fma.rn.ftz.f32 %f508, %f507, %f496, %f496; add.ftz.f32 %f509, %f508, %f508; mov.f32 %f510, 0f3FC90FDB; sub.ftz.f32 %f511, %f510, %f508; selp.f32 %f512, %f509, %f511, %p46; setp.lt.ftz.f32 %p47, %f491, 0f00000000; mov.f32 %f513, 0f40490FDB; sub.ftz.f32 %f514, %f513, %f512; selp.f32 %f92, %f514, %f512, %p47; abs.ftz.f32 %f93, %f90; setp.eq.ftz.f32 %p48, %f93, 0f00000000; abs.ftz.f32 %f94, %f91; setp.eq.ftz.f32 %p49, %f94, 0f00000000; and.pred %p1, %p48, %p49; mov.b32 %r27, %f90; mov.b32 %r66, %f91; and.b32 %r28, %r66, -2147483648; setp.gt.s32 %p50, %r65, 2; @%p50 bra $L__BB89_56; bra.uni $L__BB89_51; $L__BB89_56: setp.eq.s32 %p51, %r65, 3; @%p51 bra $L__BB89_65; setp.eq.s32 %p52, %r65, 5; @%p52 bra $L__BB89_68; setp.ne.s32 %p53, %r65, 4; @%p53 bra $L__BB89_94; ld.const.f32 %f515, [kernel_params+48]; ld.const.v4.f32 {%f516, %f517, %f518, %f519}, [kernel_params+64]; neg.ftz.f32 %f524, %f92; sub.ftz.f32 %f525, %f524, %f515; div.approx.ftz.f32 %f526, %f525, %f516; mul.ftz.f32 %f527, %f526, %f526; mul.ftz.f32 %f528, %f526, %f527; mul.ftz.f32 %f529, %f527, %f527; mul.ftz.f32 %f530, %f517, %f527; fma.rn.ftz.f32 %f531, %f516, %f526, %f530; mul.ftz.f32 %f532, %f519, %f529; fma.rn.ftz.f32 %f533, %f518, %f528, %f532; add.ftz.f32 %f534, %f531, %f533; add.ftz.f32 %f535, %f515, %f534; sub.ftz.f32 %f536, %f524, %f535; div.approx.ftz.f32 %f537, %f536, %f516; mul.ftz.f32 %f538, %f537, %f537; mul.ftz.f32 %f539, %f537, %f538; mul.ftz.f32 %f540, %f538, %f538; mul.ftz.f32 %f541, %f517, %f538; fma.rn.ftz.f32 %f542, %f516, %f537, %f541; mul.ftz.f32 %f543, %f519, %f540; fma.rn.ftz.f32 %f544, %f518, %f539, %f543; add.ftz.f32 %f545, %f542, %f544; add.ftz.f32 %f546, %f515, %f545; sub.ftz.f32 %f547, %f524, %f546; div.approx.ftz.f32 %f548, %f547, %f516; mul.ftz.f32 %f549, %f548, %f548; mul.ftz.f32 %f550, %f548, %f549; mul.ftz.f32 %f551, %f549, %f549; mul.ftz.f32 %f552, %f517, %f549; fma.rn.ftz.f32 %f553, %f516, %f548, %f552; mul.ftz.f32 %f554, %f519, %f551; fma.rn.ftz.f32 %f555, %f518, %f550, %f554; add.ftz.f32 %f556, %f553, %f555; add.ftz.f32 %f557, %f515, %f556; sub.ftz.f32 %f558, %f524, %f557; div.approx.ftz.f32 %f559, %f558, %f516; mul.ftz.f32 %f560, %f559, %f559; mul.ftz.f32 %f561, %f559, %f560; mul.ftz.f32 %f562, %f560, %f560; mul.ftz.f32 %f563, %f517, %f560; fma.rn.ftz.f32 %f564, %f516, %f559, %f563; mul.ftz.f32 %f565, %f519, %f562; fma.rn.ftz.f32 %f566, %f518, %f561, %f565; add.ftz.f32 %f567, %f564, %f566; add.ftz.f32 %f568, %f515, %f567; sub.ftz.f32 %f569, %f524, %f568; div.approx.ftz.f32 %f570, %f569, %f516; mul.ftz.f32 %f571, %f570, %f570; mul.ftz.f32 %f572, %f570, %f571; mul.ftz.f32 %f573, %f571, %f571; mul.ftz.f32 %f574, %f517, %f571; fma.rn.ftz.f32 %f575, %f516, %f570, %f574; mul.ftz.f32 %f576, %f519, %f573; fma.rn.ftz.f32 %f577, %f518, %f572, %f576; add.ftz.f32 %f578, %f575, %f577; add.ftz.f32 %f579, %f515, %f578; sub.ftz.f32 %f580, %f524, %f579; div.approx.ftz.f32 %f581, %f580, %f516; mul.ftz.f32 %f582, %f581, %f581; mul.ftz.f32 %f583, %f581, %f582; mul.ftz.f32 %f584, %f582, %f582; mul.ftz.f32 %f585, %f517, %f582; fma.rn.ftz.f32 %f586, %f516, %f581, %f585; mul.ftz.f32 %f587, %f519, %f584; fma.rn.ftz.f32 %f588, %f518, %f583, %f587; add.ftz.f32 %f589, %f586, %f588; add.ftz.f32 %f590, %f515, %f589; sub.ftz.f32 %f591, %f524, %f590; div.approx.ftz.f32 %f592, %f591, %f516; mul.ftz.f32 %f593, %f592, %f592; mul.ftz.f32 %f594, %f592, %f593; mul.ftz.f32 %f595, %f593, %f593; mul.ftz.f32 %f596, %f517, %f593; fma.rn.ftz.f32 %f597, %f516, %f592, %f596; mul.ftz.f32 %f598, %f519, %f595; fma.rn.ftz.f32 %f599, %f518, %f594, %f598; add.ftz.f32 %f600, %f597, %f599; add.ftz.f32 %f601, %f515, %f600; sub.ftz.f32 %f602, %f524, %f601; div.approx.ftz.f32 %f603, %f602, %f516; mul.ftz.f32 %f604, %f603, %f603; mul.ftz.f32 %f605, %f603, %f604; mul.ftz.f32 %f606, %f604, %f604; mul.ftz.f32 %f607, %f517, %f604; fma.rn.ftz.f32 %f608, %f516, %f603, %f607; mul.ftz.f32 %f609, %f519, %f606; fma.rn.ftz.f32 %f610, %f518, %f605, %f609; add.ftz.f32 %f611, %f608, %f610; add.ftz.f32 %f612, %f515, %f611; sub.ftz.f32 %f613, %f524, %f612; div.approx.ftz.f32 %f614, %f613, %f516; mul.ftz.f32 %f615, %f614, %f614; mul.ftz.f32 %f616, %f614, %f615; mul.ftz.f32 %f617, %f615, %f615; mul.ftz.f32 %f618, %f517, %f615; fma.rn.ftz.f32 %f619, %f516, %f614, %f618; mul.ftz.f32 %f620, %f519, %f617; fma.rn.ftz.f32 %f621, %f518, %f616, %f620; add.ftz.f32 %f622, %f619, %f621; add.ftz.f32 %f623, %f515, %f622; sub.ftz.f32 %f624, %f524, %f623; div.approx.ftz.f32 %f625, %f624, %f516; mul.ftz.f32 %f626, %f625, %f625; mul.ftz.f32 %f627, %f625, %f626; mul.ftz.f32 %f628, %f626, %f626; mul.ftz.f32 %f629, %f517, %f626; fma.rn.ftz.f32 %f630, %f516, %f625, %f629; mul.ftz.f32 %f631, %f519, %f628; fma.rn.ftz.f32 %f632, %f518, %f627, %f631; add.ftz.f32 %f633, %f630, %f632; add.ftz.f32 %f634, %f515, %f633; sub.ftz.f32 %f635, %f524, %f634; div.approx.ftz.f32 %f636, %f635, %f516; mul.ftz.f32 %f637, %f636, %f636; mul.ftz.f32 %f638, %f636, %f637; mul.ftz.f32 %f639, %f637, %f637; mul.ftz.f32 %f640, %f517, %f637; fma.rn.ftz.f32 %f641, %f516, %f636, %f640; mul.ftz.f32 %f642, %f519, %f639; fma.rn.ftz.f32 %f643, %f518, %f638, %f642; add.ftz.f32 %f644, %f641, %f643; add.ftz.f32 %f645, %f515, %f644; sub.ftz.f32 %f646, %f524, %f645; div.approx.ftz.f32 %f647, %f646, %f516; mul.ftz.f32 %f648, %f647, %f647; mul.ftz.f32 %f649, %f647, %f648; mul.ftz.f32 %f650, %f648, %f648; mul.ftz.f32 %f651, %f517, %f648; fma.rn.ftz.f32 %f652, %f516, %f647, %f651; mul.ftz.f32 %f653, %f519, %f650; fma.rn.ftz.f32 %f654, %f518, %f649, %f653; add.ftz.f32 %f655, %f652, %f654; add.ftz.f32 %f656, %f515, %f655; sub.ftz.f32 %f657, %f524, %f656; div.approx.ftz.f32 %f658, %f657, %f516; mul.ftz.f32 %f659, %f658, %f658; mul.ftz.f32 %f660, %f658, %f659; mul.ftz.f32 %f661, %f659, %f659; mul.ftz.f32 %f662, %f517, %f659; fma.rn.ftz.f32 %f663, %f516, %f658, %f662; mul.ftz.f32 %f664, %f519, %f661; fma.rn.ftz.f32 %f665, %f518, %f660, %f664; add.ftz.f32 %f666, %f663, %f665; add.ftz.f32 %f667, %f515, %f666; sub.ftz.f32 %f668, %f524, %f667; div.approx.ftz.f32 %f669, %f668, %f516; mul.ftz.f32 %f670, %f669, %f669; mul.ftz.f32 %f671, %f669, %f670; mul.ftz.f32 %f672, %f670, %f670; mul.ftz.f32 %f673, %f517, %f670; fma.rn.ftz.f32 %f674, %f516, %f669, %f673; mul.ftz.f32 %f675, %f519, %f672; fma.rn.ftz.f32 %f676, %f518, %f671, %f675; add.ftz.f32 %f677, %f674, %f676; add.ftz.f32 %f678, %f515, %f677; sub.ftz.f32 %f679, %f524, %f678; div.approx.ftz.f32 %f680, %f679, %f516; mul.ftz.f32 %f681, %f680, %f680; mul.ftz.f32 %f682, %f680, %f681; mul.ftz.f32 %f683, %f681, %f681; mul.ftz.f32 %f684, %f517, %f681; fma.rn.ftz.f32 %f685, %f516, %f680, %f684; mul.ftz.f32 %f686, %f519, %f683; fma.rn.ftz.f32 %f687, %f518, %f682, %f686; add.ftz.f32 %f688, %f685, %f687; add.ftz.f32 %f689, %f515, %f688; sub.ftz.f32 %f690, %f524, %f689; div.approx.ftz.f32 %f691, %f690, %f516; mul.ftz.f32 %f692, %f691, %f691; mul.ftz.f32 %f693, %f691, %f692; mul.ftz.f32 %f694, %f692, %f692; mul.ftz.f32 %f695, %f517, %f692; fma.rn.ftz.f32 %f696, %f516, %f691, %f695; mul.ftz.f32 %f697, %f519, %f694; fma.rn.ftz.f32 %f698, %f518, %f693, %f697; add.ftz.f32 %f699, %f696, %f698; add.ftz.f32 %f700, %f515, %f699; sub.ftz.f32 %f701, %f524, %f700; div.approx.ftz.f32 %f702, %f701, %f516; mul.ftz.f32 %f703, %f702, %f702; mul.ftz.f32 %f704, %f702, %f703; mul.ftz.f32 %f705, %f703, %f703; mul.ftz.f32 %f706, %f517, %f703; fma.rn.ftz.f32 %f707, %f516, %f702, %f706; mul.ftz.f32 %f708, %f519, %f705; fma.rn.ftz.f32 %f709, %f518, %f704, %f708; add.ftz.f32 %f710, %f707, %f709; add.ftz.f32 %f711, %f515, %f710; sub.ftz.f32 %f712, %f524, %f711; div.approx.ftz.f32 %f713, %f712, %f516; mul.ftz.f32 %f714, %f713, %f713; mul.ftz.f32 %f715, %f713, %f714; mul.ftz.f32 %f716, %f714, %f714; mul.ftz.f32 %f717, %f517, %f714; fma.rn.ftz.f32 %f718, %f516, %f713, %f717; mul.ftz.f32 %f719, %f519, %f716; fma.rn.ftz.f32 %f720, %f518, %f715, %f719; add.ftz.f32 %f721, %f718, %f720; add.ftz.f32 %f722, %f515, %f721; sub.ftz.f32 %f723, %f524, %f722; div.approx.ftz.f32 %f724, %f723, %f516; mul.ftz.f32 %f725, %f724, %f724; mul.ftz.f32 %f726, %f724, %f725; mul.ftz.f32 %f727, %f725, %f725; mul.ftz.f32 %f728, %f517, %f725; fma.rn.ftz.f32 %f729, %f516, %f724, %f728; mul.ftz.f32 %f730, %f519, %f727; fma.rn.ftz.f32 %f731, %f518, %f726, %f730; add.ftz.f32 %f732, %f729, %f731; add.ftz.f32 %f733, %f515, %f732; sub.ftz.f32 %f734, %f524, %f733; div.approx.ftz.f32 %f735, %f734, %f516; mul.ftz.f32 %f736, %f735, %f735; mul.ftz.f32 %f737, %f735, %f736; mul.ftz.f32 %f738, %f736, %f736; mul.ftz.f32 %f739, %f517, %f736; fma.rn.ftz.f32 %f740, %f516, %f735, %f739; mul.ftz.f32 %f741, %f519, %f738; fma.rn.ftz.f32 %f742, %f518, %f737, %f741; add.ftz.f32 %f743, %f740, %f742; add.ftz.f32 %f744, %f515, %f743; sub.ftz.f32 %f745, %f524, %f744; div.approx.ftz.f32 %f97, %f745, %f516; @%p1 bra $L__BB89_63; bra.uni $L__BB89_60; $L__BB89_63: shr.s32 %r71, %r27, 31; and.b32 %r72, %r71, 1078530011; or.b32 %r73, %r72, %r28; mov.b32 %f1287, %r73; bra.uni $L__BB89_64; $L__BB89_51: setp.eq.s32 %p54, %r65, 0; @%p54 bra $L__BB89_87; setp.eq.s32 %p55, %r65, 1; @%p55 bra $L__BB89_53; bra.uni $L__BB89_94; $L__BB89_53: ld.const.f32 %f122, [kernel_params+24]; mul.ftz.f32 %f845, %f91, %f91; fma.rn.ftz.f32 %f846, %f90, %f90, %f845; sqrt.approx.ftz.f32 %f847, %f846; abs.ftz.f32 %f123, %f847; abs.ftz.f32 %f124, %f89; setp.eq.ftz.f32 %p84, %f124, 0f00000000; setp.eq.ftz.f32 %p85, %f123, 0f00000000; and.pred %p86, %p84, %p85; mov.b32 %r31, %f89; mov.b32 %r88, %f847; and.b32 %r32, %r88, -2147483648; @%p86 bra $L__BB89_80; bra.uni $L__BB89_54; $L__BB89_80: shr.s32 %r93, %r31, 31; and.b32 %r94, %r93, 1078530011; or.b32 %r95, %r32, %r94; mov.b32 %f1292, %r95; bra.uni $L__BB89_81; $L__BB89_106: ld.local.v4.f32 {%f1098, %f1099, %f1100, %f1101}, [%rd1+192]; mul.ftz.f32 %f1106, %f1301, %f1099; fma.rn.ftz.f32 %f1107, %f1300, %f1098, %f1106; fma.rn.ftz.f32 %f1108, %f1302, %f1100, %f1107; add.ftz.f32 %f190, %f1101, %f1108; ld.local.v4.f32 {%f1109, %f1110, %f1111, %f1112}, [%rd1+208]; mul.ftz.f32 %f1117, %f1301, %f1110; fma.rn.ftz.f32 %f1118, %f1300, %f1109, %f1117; fma.rn.ftz.f32 %f1119, %f1302, %f1111, %f1118; add.ftz.f32 %f191, %f1112, %f1119; ld.local.v4.f32 {%f1120, %f1121, %f1122, %f1123}, [%rd1+224]; mul.ftz.f32 %f1128, %f1301, %f1121; fma.rn.ftz.f32 %f1129, %f1300, %f1120, %f1128; fma.rn.ftz.f32 %f1130, %f1302, %f1122, %f1129; add.ftz.f32 %f1302, %f1123, %f1130; mov.f32 %f1300, %f190; mov.f32 %f1301, %f191; bra.uni $L__BB89_116; $L__BB89_65: add.ftz.f32 %f778, %f90, 0fBF800000; mul.ftz.f32 %f779, %f778, 0fBF000000; mov.f32 %f780, 0f00000000; max.ftz.f32 %f781, %f779, %f780; sqrt.approx.ftz.f32 %f782, %f781; add.ftz.f32 %f104, %f782, %f782; setp.leu.ftz.f32 %p63, %f104, 0f00000000; @%p63 bra $L__BB89_67; rcp.approx.ftz.f32 %f783, %f104; mul.ftz.f32 %f89, %f89, %f783; mul.ftz.f32 %f91, %f91, %f783; $L__BB89_67: add.ftz.f32 %f784, %f89, 0f3F800000; add.ftz.f32 %f785, %f91, 0f3F800000; mul.ftz.f32 %f1301, %f785, 0f3F000000; mul.ftz.f32 %f1300, %f784, 0f3F000000; bra.uni $L__BB89_100; $L__BB89_68: abs.ftz.f32 %f111, %f89; setp.eq.ftz.f32 %p65, %f111, 0f00000000; and.pred %p66, %p65, %p48; mov.b32 %r29, %f89; and.b32 %r30, %r27, -2147483648; @%p66 bra $L__BB89_72; bra.uni $L__BB89_69; $L__BB89_72: shr.s32 %r78, %r29, 31; and.b32 %r79, %r78, 1078530011; or.b32 %r80, %r79, %r30; mov.b32 %f1290, %r80; bra.uni $L__BB89_73; $L__BB89_87: setp.eq.ftz.f32 %p101, %f90, 0f00000000; setp.eq.ftz.f32 %p102, %f89, 0f00000000; and.pred %p103, %p102, %p101; setp.eq.ftz.f32 %p104, %f91, 0f00000000; and.pred %p105, %p104, %p103; mov.f32 %f1300, 0f00000000; mov.f32 %f1301, %f1300; @%p105 bra $L__BB89_100; ld.const.v4.f32 {%f904, %f905, %f906, %f907}, [kernel_params+32]; abs.ftz.f32 %f140, %f89; setp.eq.ftz.f32 %p107, %f140, 0f00000000; and.pred %p108, %p107, %p48; mov.b32 %r33, %f89; and.b32 %r34, %r27, -2147483648; @%p108 bra $L__BB89_92; bra.uni $L__BB89_89; $L__BB89_92: shr.s32 %r107, %r33, 31; and.b32 %r108, %r107, 1078530011; or.b32 %r109, %r108, %r34; mov.b32 %f1294, %r109; bra.uni $L__BB89_93; $L__BB89_94: ld.const.f32 %f966, [kernel_params+28]; ld.const.f32 %f147, [kernel_params+276]; ld.const.f32 %f148, [kernel_params+280]; add.ftz.f32 %f967, %f966, %f966; mul.ftz.f32 %f968, %f92, 0f3F000000; sin.approx.ftz.f32 %f969, %f968; mul.ftz.f32 %f149, %f967, %f969; @%p1 bra $L__BB89_98; bra.uni $L__BB89_95; $L__BB89_98: shr.s32 %r114, %r27, 31; and.b32 %r115, %r114, 1078530011; or.b32 %r116, %r115, %r28; mov.b32 %f1295, %r116; bra.uni $L__BB89_99; $L__BB89_69: setp.eq.ftz.f32 %p67, %f111, 0f7F800000; setp.eq.ftz.f32 %p68, %f93, 0f7F800000; and.pred %p69, %p67, %p68; @%p69 bra $L__BB89_71; bra.uni $L__BB89_70; $L__BB89_71: setp.lt.s32 %p73, %r29, 0; selp.b32 %r76, 1075235812, 1061752795, %p73; or.b32 %r77, %r76, %r30; mov.b32 %f1290, %r77; bra.uni $L__BB89_73; $L__BB89_60: setp.eq.ftz.f32 %p56, %f93, 0f7F800000; setp.eq.ftz.f32 %p57, %f94, 0f7F800000; and.pred %p58, %p56, %p57; @%p58 bra $L__BB89_62; bra.uni $L__BB89_61; $L__BB89_62: setp.lt.s32 %p62, %r27, 0; selp.b32 %r69, 1075235812, 1061752795, %p62; or.b32 %r70, %r69, %r28; mov.b32 %f1287, %r70; bra.uni $L__BB89_64; $L__BB89_54: setp.eq.ftz.f32 %p87, %f124, 0f7F800000; setp.eq.ftz.f32 %p88, %f123, 0f7F800000; and.pred %p89, %p87, %p88; @%p89 bra $L__BB89_79; bra.uni $L__BB89_55; $L__BB89_79: setp.lt.s32 %p93, %r31, 0; selp.b32 %r91, 1075235812, 1061752795, %p93; or.b32 %r92, %r32, %r91; mov.b32 %f1292, %r92; bra.uni $L__BB89_81; $L__BB89_95: setp.eq.ftz.f32 %p118, %f93, 0f7F800000; setp.eq.ftz.f32 %p119, %f94, 0f7F800000; and.pred %p120, %p118, %p119; @%p120 bra $L__BB89_97; bra.uni $L__BB89_96; $L__BB89_97: setp.lt.s32 %p124, %r27, 0; selp.b32 %r112, 1075235812, 1061752795, %p124; or.b32 %r113, %r112, %r28; mov.b32 %f1295, %r113; bra.uni $L__BB89_99; $L__BB89_70: setp.lt.s32 %p70, %r29, 0; min.ftz.f32 %f786, %f93, %f111; max.ftz.f32 %f787, %f93, %f111; div.full.ftz.f32 %f788, %f786, %f787; mul.rn.ftz.f32 %f789, %f788, %f788; mov.f32 %f790, 0fC0B59883; mov.f32 %f791, 0fBF52C7EA; fma.rn.ftz.f32 %f792, %f789, %f791, %f790; mov.f32 %f793, 0fC0D21907; fma.rn.ftz.f32 %f794, %f792, %f789, %f793; mul.ftz.f32 %f795, %f789, %f794; mul.ftz.f32 %f796, %f788, %f795; add.ftz.f32 %f797, %f789, 0f41355DC0; mov.f32 %f798, 0f41E6BD60; fma.rn.ftz.f32 %f799, %f797, %f789, %f798; mov.f32 %f800, 0f419D92C8; fma.rn.ftz.f32 %f801, %f799, %f789, %f800; rcp.approx.ftz.f32 %f802, %f801; fma.rn.ftz.f32 %f803, %f796, %f802, %f788; sub.ftz.f32 %f805, %f510, %f803; setp.gt.ftz.f32 %p71, %f93, %f111; selp.f32 %f806, %f805, %f803, %p71; sub.ftz.f32 %f808, %f513, %f806; selp.f32 %f809, %f808, %f806, %p70; mov.b32 %r74, %f809; or.b32 %r75, %r30, %r74; mov.b32 %f810, %r75; add.ftz.f32 %f811, %f111, %f93; setp.le.ftz.f32 %p72, %f811, 0f7F800000; selp.f32 %f1290, %f810, %f811, %p72; $L__BB89_73: and.pred %p76, %p65, %p49; @%p76 bra $L__BB89_77; bra.uni $L__BB89_74; $L__BB89_77: shr.s32 %r85, %r29, 31; and.b32 %r86, %r85, 1078530011; or.b32 %r87, %r28, %r86; mov.b32 %f1291, %r87; bra.uni $L__BB89_78; $L__BB89_74: setp.eq.ftz.f32 %p77, %f111, 0f7F800000; setp.eq.ftz.f32 %p78, %f94, 0f7F800000; and.pred %p79, %p77, %p78; @%p79 bra $L__BB89_76; bra.uni $L__BB89_75; $L__BB89_76: setp.lt.s32 %p83, %r29, 0; selp.b32 %r83, 1075235812, 1061752795, %p83; or.b32 %r84, %r28, %r83; mov.b32 %f1291, %r84; bra.uni $L__BB89_78; $L__BB89_75: setp.lt.s32 %p80, %r29, 0; min.ftz.f32 %f812, %f94, %f111; max.ftz.f32 %f813, %f94, %f111; div.full.ftz.f32 %f814, %f812, %f813; mul.rn.ftz.f32 %f815, %f814, %f814; mov.f32 %f816, 0fC0B59883; mov.f32 %f817, 0fBF52C7EA; fma.rn.ftz.f32 %f818, %f815, %f817, %f816; mov.f32 %f819, 0fC0D21907; fma.rn.ftz.f32 %f820, %f818, %f815, %f819; mul.ftz.f32 %f821, %f815, %f820; mul.ftz.f32 %f822, %f814, %f821; add.ftz.f32 %f823, %f815, 0f41355DC0; mov.f32 %f824, 0f41E6BD60; fma.rn.ftz.f32 %f825, %f823, %f815, %f824; mov.f32 %f826, 0f419D92C8; fma.rn.ftz.f32 %f827, %f825, %f815, %f826; rcp.approx.ftz.f32 %f828, %f827; fma.rn.ftz.f32 %f829, %f822, %f828, %f814; sub.ftz.f32 %f831, %f510, %f829; setp.gt.ftz.f32 %p81, %f94, %f111; selp.f32 %f832, %f831, %f829, %p81; sub.ftz.f32 %f834, %f513, %f832; selp.f32 %f835, %f834, %f832, %p80; mov.b32 %r81, %f835; or.b32 %r82, %r28, %r81; mov.b32 %f836, %r82; add.ftz.f32 %f837, %f111, %f94; setp.le.ftz.f32 %p82, %f837, 0f7F800000; selp.f32 %f1291, %f836, %f837, %p82; $L__BB89_78: add.ftz.f32 %f838, %f1291, %f1291; div.approx.ftz.f32 %f840, %f838, %f513; add.ftz.f32 %f841, %f1290, %f1290; div.approx.ftz.f32 %f842, %f841, %f513; add.ftz.f32 %f843, %f842, 0f3F000000; sub.ftz.f32 %f1300, %f490, %f843; add.ftz.f32 %f1301, %f840, 0f3F000000; bra.uni $L__BB89_100; $L__BB89_89: setp.eq.ftz.f32 %p109, %f140, 0f7F800000; setp.eq.ftz.f32 %p110, %f93, 0f7F800000; and.pred %p111, %p109, %p110; @%p111 bra $L__BB89_91; bra.uni $L__BB89_90; $L__BB89_91: setp.lt.s32 %p115, %r33, 0; selp.b32 %r105, 1075235812, 1061752795, %p115; or.b32 %r106, %r105, %r34; mov.b32 %f1294, %r106; bra.uni $L__BB89_93; $L__BB89_61: mov.f32 %f1267, 0f40490FDB; mov.f32 %f1266, 0f3FC90FDB; setp.lt.s32 %p59, %r27, 0; min.ftz.f32 %f746, %f94, %f93; max.ftz.f32 %f747, %f94, %f93; div.full.ftz.f32 %f748, %f746, %f747; mul.rn.ftz.f32 %f749, %f748, %f748; mov.f32 %f750, 0fC0B59883; mov.f32 %f751, 0fBF52C7EA; fma.rn.ftz.f32 %f752, %f749, %f751, %f750; mov.f32 %f753, 0fC0D21907; fma.rn.ftz.f32 %f754, %f752, %f749, %f753; mul.ftz.f32 %f755, %f749, %f754; mul.ftz.f32 %f756, %f748, %f755; add.ftz.f32 %f757, %f749, 0f41355DC0; mov.f32 %f758, 0f41E6BD60; fma.rn.ftz.f32 %f759, %f757, %f749, %f758; mov.f32 %f760, 0f419D92C8; fma.rn.ftz.f32 %f761, %f759, %f749, %f760; rcp.approx.ftz.f32 %f762, %f761; fma.rn.ftz.f32 %f763, %f756, %f762, %f748; sub.ftz.f32 %f765, %f1266, %f763; setp.gt.ftz.f32 %p60, %f94, %f93; selp.f32 %f766, %f765, %f763, %p60; sub.ftz.f32 %f768, %f1267, %f766; selp.f32 %f769, %f768, %f766, %p59; mov.b32 %r67, %f769; or.b32 %r68, %r28, %r67; mov.b32 %f770, %r68; add.ftz.f32 %f771, %f93, %f94; setp.le.ftz.f32 %p61, %f771, 0f7F800000; selp.f32 %f1287, %f770, %f771, %p61; $L__BB89_64: ld.const.f32 %f1269, [kernel_params+280]; ld.const.f32 %f1268, [kernel_params+276]; cos.approx.ftz.f32 %f772, %f1287; mul.ftz.f32 %f773, %f97, %f772; div.approx.ftz.f32 %f774, %f773, %f1268; sin.approx.ftz.f32 %f775, %f1287; mul.ftz.f32 %f776, %f97, %f775; div.approx.ftz.f32 %f777, %f776, %f1269; add.ftz.f32 %f1301, %f777, 0f3F000000; add.ftz.f32 %f1300, %f774, 0f3F000000; bra.uni $L__BB89_100; $L__BB89_55: setp.lt.s32 %p90, %r31, 0; min.ftz.f32 %f848, %f123, %f124; max.ftz.f32 %f849, %f123, %f124; div.full.ftz.f32 %f850, %f848, %f849; mul.rn.ftz.f32 %f851, %f850, %f850; mov.f32 %f852, 0fC0B59883; mov.f32 %f853, 0fBF52C7EA; fma.rn.ftz.f32 %f854, %f851, %f853, %f852; mov.f32 %f855, 0fC0D21907; fma.rn.ftz.f32 %f856, %f854, %f851, %f855; mul.ftz.f32 %f857, %f851, %f856; mul.ftz.f32 %f858, %f850, %f857; add.ftz.f32 %f859, %f851, 0f41355DC0; mov.f32 %f860, 0f41E6BD60; fma.rn.ftz.f32 %f861, %f859, %f851, %f860; mov.f32 %f862, 0f419D92C8; fma.rn.ftz.f32 %f863, %f861, %f851, %f862; rcp.approx.ftz.f32 %f864, %f863; fma.rn.ftz.f32 %f865, %f858, %f864, %f850; sub.ftz.f32 %f867, %f510, %f865; setp.gt.ftz.f32 %p91, %f123, %f124; selp.f32 %f868, %f867, %f865, %p91; sub.ftz.f32 %f870, %f513, %f868; selp.f32 %f871, %f870, %f868, %p90; mov.b32 %r89, %f871; or.b32 %r90, %r32, %r89; mov.b32 %f872, %r90; add.ftz.f32 %f873, %f124, %f123; setp.le.ftz.f32 %p92, %f873, 0f7F800000; selp.f32 %f1292, %f872, %f873, %p92; $L__BB89_81: div.approx.ftz.f32 %f129, %f1292, %f122; @%p1 bra $L__BB89_85; bra.uni $L__BB89_82; $L__BB89_85: shr.s32 %r100, %r27, 31; and.b32 %r101, %r100, 1078530011; or.b32 %r102, %r101, %r28; mov.b32 %f1293, %r102; bra.uni $L__BB89_86; $L__BB89_82: setp.eq.ftz.f32 %p94, %f93, 0f7F800000; setp.eq.ftz.f32 %p95, %f94, 0f7F800000; and.pred %p96, %p94, %p95; @%p96 bra $L__BB89_84; bra.uni $L__BB89_83; $L__BB89_84: setp.lt.s32 %p100, %r27, 0; selp.b32 %r98, 1075235812, 1061752795, %p100; or.b32 %r99, %r98, %r28; mov.b32 %f1293, %r99; bra.uni $L__BB89_86; $L__BB89_96: setp.lt.s32 %p121, %r27, 0; min.ftz.f32 %f970, %f94, %f93; max.ftz.f32 %f971, %f94, %f93; div.full.ftz.f32 %f972, %f970, %f971; mul.rn.ftz.f32 %f973, %f972, %f972; mov.f32 %f974, 0fC0B59883; mov.f32 %f975, 0fBF52C7EA; fma.rn.ftz.f32 %f976, %f973, %f975, %f974; mov.f32 %f977, 0fC0D21907; fma.rn.ftz.f32 %f978, %f976, %f973, %f977; mul.ftz.f32 %f979, %f973, %f978; mul.ftz.f32 %f980, %f972, %f979; add.ftz.f32 %f981, %f973, 0f41355DC0; mov.f32 %f982, 0f41E6BD60; fma.rn.ftz.f32 %f983, %f981, %f973, %f982; mov.f32 %f984, 0f419D92C8; fma.rn.ftz.f32 %f985, %f983, %f973, %f984; rcp.approx.ftz.f32 %f986, %f985; fma.rn.ftz.f32 %f987, %f980, %f986, %f972; sub.ftz.f32 %f989, %f510, %f987; setp.gt.ftz.f32 %p122, %f94, %f93; selp.f32 %f990, %f989, %f987, %p122; sub.ftz.f32 %f992, %f513, %f990; selp.f32 %f993, %f992, %f990, %p121; mov.b32 %r110, %f993; or.b32 %r111, %r28, %r110; mov.b32 %f994, %r111; add.ftz.f32 %f995, %f93, %f94; setp.le.ftz.f32 %p123, %f995, 0f7F800000; selp.f32 %f1295, %f994, %f995, %p123; $L__BB89_99: cos.approx.ftz.f32 %f996, %f1295; mul.ftz.f32 %f997, %f149, %f996; div.approx.ftz.f32 %f998, %f997, %f147; sin.approx.ftz.f32 %f999, %f1295; mul.ftz.f32 %f1000, %f149, %f999; div.approx.ftz.f32 %f1001, %f1000, %f148; add.ftz.f32 %f1301, %f1001, 0f3F000000; add.ftz.f32 %f1300, %f998, 0f3F000000; bra.uni $L__BB89_100; $L__BB89_83: setp.lt.s32 %p97, %r27, 0; min.ftz.f32 %f874, %f94, %f93; max.ftz.f32 %f875, %f94, %f93; div.full.ftz.f32 %f876, %f874, %f875; mul.rn.ftz.f32 %f877, %f876, %f876; mov.f32 %f878, 0fC0B59883; mov.f32 %f879, 0fBF52C7EA; fma.rn.ftz.f32 %f880, %f877, %f879, %f878; mov.f32 %f881, 0fC0D21907; fma.rn.ftz.f32 %f882, %f880, %f877, %f881; mul.ftz.f32 %f883, %f877, %f882; mul.ftz.f32 %f884, %f876, %f883; add.ftz.f32 %f885, %f877, 0f41355DC0; mov.f32 %f886, 0f41E6BD60; fma.rn.ftz.f32 %f887, %f885, %f877, %f886; mov.f32 %f888, 0f419D92C8; fma.rn.ftz.f32 %f889, %f887, %f877, %f888; rcp.approx.ftz.f32 %f890, %f889; fma.rn.ftz.f32 %f891, %f884, %f890, %f876; sub.ftz.f32 %f893, %f510, %f891; setp.gt.ftz.f32 %p98, %f94, %f93; selp.f32 %f894, %f893, %f891, %p98; sub.ftz.f32 %f896, %f513, %f894; selp.f32 %f897, %f896, %f894, %p97; mov.b32 %r96, %f897; or.b32 %r97, %r28, %r96; mov.b32 %f898, %r97; add.ftz.f32 %f899, %f93, %f94; setp.le.ftz.f32 %p99, %f899, 0f7F800000; selp.f32 %f1293, %f898, %f899, %p99; $L__BB89_86: cos.approx.ftz.f32 %f900, %f1293; sin.approx.ftz.f32 %f901, %f1293; fma.rn.ftz.f32 %f1301, %f129, %f901, 0f3F000000; fma.rn.ftz.f32 %f1300, %f129, %f900, 0f3F000000; bra.uni $L__BB89_100; $L__BB89_90: setp.lt.s32 %p112, %r33, 0; min.ftz.f32 %f908, %f93, %f140; max.ftz.f32 %f909, %f93, %f140; div.full.ftz.f32 %f910, %f908, %f909; mul.rn.ftz.f32 %f911, %f910, %f910; mov.f32 %f912, 0fC0B59883; mov.f32 %f913, 0fBF52C7EA; fma.rn.ftz.f32 %f914, %f911, %f913, %f912; mov.f32 %f915, 0fC0D21907; fma.rn.ftz.f32 %f916, %f914, %f911, %f915; mul.ftz.f32 %f917, %f911, %f916; mul.ftz.f32 %f918, %f910, %f917; add.ftz.f32 %f919, %f911, 0f41355DC0; mov.f32 %f920, 0f41E6BD60; fma.rn.ftz.f32 %f921, %f919, %f911, %f920; mov.f32 %f922, 0f419D92C8; fma.rn.ftz.f32 %f923, %f921, %f911, %f922; rcp.approx.ftz.f32 %f924, %f923; fma.rn.ftz.f32 %f925, %f918, %f924, %f910; sub.ftz.f32 %f927, %f510, %f925; setp.gt.ftz.f32 %p113, %f93, %f140; selp.f32 %f928, %f927, %f925, %p113; sub.ftz.f32 %f930, %f513, %f928; selp.f32 %f931, %f930, %f928, %p112; mov.b32 %r103, %f931; or.b32 %r104, %r34, %r103; mov.b32 %f932, %r104; add.ftz.f32 %f933, %f140, %f93; setp.le.ftz.f32 %p114, %f933, 0f7F800000; selp.f32 %f1294, %f932, %f933, %p114; $L__BB89_93: sub.ftz.f32 %f934, %f1294, %f905; mul.ftz.f32 %f935, %f90, %f90; fma.rn.ftz.f32 %f936, %f89, %f89, %f935; fma.rn.ftz.f32 %f937, %f91, %f91, %f936; sqrt.approx.ftz.f32 %f938, %f937; div.approx.ftz.f32 %f939, %f91, %f938; abs.ftz.f32 %f940, %f939; sub.ftz.f32 %f942, %f490, %f940; mul.ftz.f32 %f943, %f942, 0f3F000000; sqrt.approx.ftz.f32 %f944, %f943; setp.gt.ftz.f32 %p116, %f940, 0f3F11EB85; selp.f32 %f945, %f944, %f940, %p116; mul.ftz.f32 %f946, %f945, %f945; fma.rn.ftz.f32 %f949, %f499, %f946, %f498; fma.rn.ftz.f32 %f951, %f949, %f946, %f501; fma.rn.ftz.f32 %f953, %f951, %f946, %f503; fma.rn.ftz.f32 %f955, %f953, %f946, %f505; mul.ftz.f32 %f956, %f946, %f955; fma.rn.ftz.f32 %f957, %f956, %f945, %f945; add.ftz.f32 %f958, %f957, %f957; sub.ftz.f32 %f960, %f510, %f957; selp.f32 %f961, %f958, %f960, %p116; setp.lt.ftz.f32 %p117, %f939, 0f00000000; sub.ftz.f32 %f963, %f513, %f961; selp.f32 %f964, %f963, %f961, %p117; sub.ftz.f32 %f965, %f964, %f907; div.approx.ftz.f32 %f1301, %f965, %f906; div.approx.ftz.f32 %f1300, %f934, %f904; $L__BB89_100: ld.param.u32 %r132, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; mov.f32 %f1302, 0f00000000; $L__BB89_116: ld.param.u32 %r124, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+8]; ld.param.u64 %rd71, [_Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2]; cvta.to.local.u64 %rd68, %rd71; mul.wide.u32 %rd69, %r124, 4; add.s64 %rd70, %rd68, %rd69; st.local.f32 [%rd70], %f1300; st.local.f32 [%rd70+4], %f1301; st.local.f32 [%rd70+8], %f1302; st.param.b32 [func_retval0+0], %r132; ret; } .func (.param .b32 func_retval0) _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i( .param .b64 _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0, .param .b32 _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1, .param .b64 _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2, .param .align 4 .b8 _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3[16], .param .b32 _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4 ) { .reg .pred %p<146>; .reg .b16 %rs<16>; .reg .f32 %f<1368>; .reg .b32 %r<133>; .reg .b64 %rd<71>; ld.param.u64 %rd22, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0]; ld.param.u32 %r38, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1]; ld.param.u32 %r132, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; ld.param.u32 %r42, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+12]; ld.param.u32 %r40, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+4]; cvta.to.local.u64 %rd1, %rd22; setp.gt.s32 %p2, %r40, 3; @%p2 bra $L__BB90_12; setp.gt.s32 %p8, %r40, 1; @%p8 bra $L__BB90_6; setp.eq.s32 %p11, %r40, 0; @%p11 bra $L__BB90_108; setp.eq.s32 %p12, %r40, 1; @%p12 bra $L__BB90_4; bra.uni $L__BB90_116; $L__BB90_4: ld.local.f32 %f1109, [%rd1+28]; ld.local.f32 %f1110, [%rd1+24]; setp.neu.ftz.f32 %p135, %f1110, %f1109; ld.local.f32 %f1111, [%rd1+32]; setp.neu.ftz.f32 %p136, %f1110, %f1111; or.pred %p137, %p135, %p136; sub.ftz.f32 %f1112, %f1111, %f1109; neg.ftz.f32 %f1113, %f1109; neg.ftz.f32 %f1114, %f1111; selp.f32 %f1115, %f1114, %f1111, %p137; add.ftz.f32 %f1116, %f1110, %f1115; selp.f32 %f1117, %f1109, %f1113, %p137; sub.ftz.f32 %f1118, %f1117, %f1110; mul.ftz.f32 %f1119, %f1116, %f1116; fma.rn.ftz.f32 %f1120, %f1112, %f1112, %f1119; fma.rn.ftz.f32 %f1121, %f1118, %f1118, %f1120; sqrt.approx.ftz.f32 %f1122, %f1121; div.approx.ftz.f32 %f1123, %f1112, %f1122; div.approx.ftz.f32 %f1124, %f1116, %f1122; div.approx.ftz.f32 %f1125, %f1118, %f1122; ld.local.f32 %f1126, [%rd1+92]; ld.local.f32 %f1127, [%rd1]; fma.rn.ftz.f32 %f1367, %f1126, %f1123, %f1127; ld.local.f32 %f1128, [%rd1+4]; fma.rn.ftz.f32 %f1366, %f1126, %f1124, %f1128; ld.local.f32 %f1129, [%rd1+8]; fma.rn.ftz.f32 %f1365, %f1126, %f1125, %f1129; setp.eq.s32 %p138, %r42, 0; @%p138 bra $L__BB90_104; ld.const.u64 %rd50, [kernel_params+2400]; cvta.to.global.u64 %rd51, %rd50; mul.wide.s32 %rd52, %r132, 16; add.s64 %rd53, %rd51, %rd52; ld.global.v4.f32 {%f1130, %f1131, %f1132, %f1133}, [%rd53]; ld.global.v4.f32 {%f1138, %f1139, %f1140, %f1141}, [%rd53+16]; ld.global.v4.f32 {%f1146, %f1147, %f1148, %f1149}, [%rd53+32]; add.s32 %r132, %r132, 3; mul.ftz.f32 %f1154, %f1366, %f1131; fma.rn.ftz.f32 %f1155, %f1367, %f1130, %f1154; fma.rn.ftz.f32 %f1156, %f1365, %f1132, %f1155; add.ftz.f32 %f190, %f1156, %f1133; mul.ftz.f32 %f1157, %f1366, %f1139; fma.rn.ftz.f32 %f1158, %f1367, %f1138, %f1157; fma.rn.ftz.f32 %f1159, %f1365, %f1140, %f1158; add.ftz.f32 %f191, %f1159, %f1141; mul.ftz.f32 %f1160, %f1366, %f1147; fma.rn.ftz.f32 %f1161, %f1367, %f1146, %f1160; fma.rn.ftz.f32 %f1162, %f1365, %f1148, %f1161; add.ftz.f32 %f1365, %f1162, %f1149; mov.f32 %f1366, %f191; mov.f32 %f1367, %f190; bra.uni $L__BB90_116; $L__BB90_12: setp.gt.s32 %p3, %r40, 5; @%p3 bra $L__BB90_17; setp.eq.s32 %p6, %r40, 4; @%p6 bra $L__BB90_38; setp.eq.s32 %p7, %r40, 5; @%p7 bra $L__BB90_15; bra.uni $L__BB90_116; $L__BB90_15: ld.local.u32 %r23, [%rd1+76]; setp.eq.s32 %p33, %r23, -1; mov.f32 %f1367, 0f00000000; mov.f32 %f1366, %f1367; mov.f32 %f1365, %f1367; @%p33 bra $L__BB90_116; ld.const.u64 %rd46, [kernel_params+2160]; cvta.to.global.u64 %rd47, %rd46; mul.wide.s32 %rd48, %r23, 208; add.s64 %rd49, %rd47, %rd48; ld.global.v4.f32 {%f1367, %f1366, %f1365, %f344}, [%rd49+128]; bra.uni $L__BB90_116; $L__BB90_6: setp.eq.s32 %p9, %r40, 2; @%p9 bra $L__BB90_101; setp.eq.s32 %p10, %r40, 3; @%p10 bra $L__BB90_8; bra.uni $L__BB90_116; $L__BB90_8: and.b32 %r61, %r38, 1; setp.eq.b32 %p35, %r61, 1; mov.pred %p36, 0; xor.pred %p37, %p35, %p36; not.pred %p38, %p37; @%p38 bra $L__BB90_41; bra.uni $L__BB90_9; $L__BB90_41: ld.const.u32 %r131, [kernel_params+16]; bra.uni $L__BB90_42; $L__BB90_17: setp.eq.s32 %p4, %r40, 6; @%p4 bra $L__BB90_36; setp.ne.s32 %p5, %r40, 7; @%p5 bra $L__BB90_116; ld.local.f32 %f215, [%rd1+28]; ld.local.f32 %f216, [%rd1+24]; setp.neu.ftz.f32 %p13, %f216, %f215; ld.local.f32 %f217, [%rd1+32]; setp.neu.ftz.f32 %p14, %f216, %f217; or.pred %p15, %p13, %p14; sub.ftz.f32 %f218, %f217, %f215; neg.ftz.f32 %f219, %f215; neg.ftz.f32 %f220, %f217; selp.f32 %f221, %f220, %f217, %p15; add.ftz.f32 %f222, %f216, %f221; selp.f32 %f223, %f215, %f219, %p15; sub.ftz.f32 %f224, %f223, %f216; mul.ftz.f32 %f225, %f222, %f222; fma.rn.ftz.f32 %f226, %f218, %f218, %f225; fma.rn.ftz.f32 %f227, %f224, %f224, %f226; sqrt.approx.ftz.f32 %f228, %f227; div.approx.ftz.f32 %f229, %f218, %f228; div.approx.ftz.f32 %f230, %f222, %f228; div.approx.ftz.f32 %f231, %f224, %f228; ld.local.f32 %f232, [%rd1+92]; ld.local.f32 %f233, [%rd1]; fma.rn.ftz.f32 %f1367, %f232, %f229, %f233; ld.local.f32 %f234, [%rd1+4]; fma.rn.ftz.f32 %f1366, %f232, %f230, %f234; ld.local.f32 %f235, [%rd1+8]; fma.rn.ftz.f32 %f1365, %f232, %f231, %f235; ld.local.u32 %r1, [%rd1+76]; setp.eq.s32 %p16, %r1, -1; @%p16 bra $L__BB90_116; ld.local.s32 %rd3, [%rd1+60]; ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.wide.s32 %rd25, %r1, 208; add.s64 %rd26, %rd24, %rd25; add.s64 %rd4, %rd26, 164; ld.global.u32 %r2, [%rd26+164]; ld.local.u32 %r46, [%rd1+64]; and.b32 %r47, %r46, 1; setp.eq.b32 %p17, %r47, 1; not.pred %p18, %p17; mov.u32 %r125, 0; setp.eq.s64 %p19, %rd3, -1; or.pred %p20, %p19, %p18; @%p20 bra $L__BB90_22; ld.const.u64 %rd27, [kernel_params+2240]; cvta.to.global.u64 %rd28, %rd27; shl.b64 %rd29, %rd3, 2; add.s64 %rd30, %rd28, %rd29; ld.global.u32 %r48, [%rd30]; setp.ne.s32 %p21, %r48, -1; selp.u32 %r125, 1, 0, %p21; $L__BB90_22: ld.const.u64 %rd31, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd31; add.s32 %r127, %r125, %r2; mul.wide.u32 %rd32, %r127, 16; add.s64 %rd33, %rd5, %rd32; ld.global.u16 %rs14, [%rd33+12]; ld.global.u32 %r126, [%rd33+8]; ld.global.u64 %rd70, [%rd33]; setp.eq.s64 %p22, %rd70, 8; @%p22 bra $L__BB90_27; $L__BB90_23: setp.eq.s64 %p23, %rd70, 0; @%p23 bra $L__BB90_25; add.s32 %r127, %r127, 2; bra.uni $L__BB90_26; $L__BB90_25: setp.eq.s16 %p24, %rs14, 0; mov.u32 %r130, -1; mov.u32 %r127, %r126; @%p24 bra $L__BB90_31; $L__BB90_26: mul.wide.u32 %rd34, %r127, 16; add.s64 %rd35, %rd5, %rd34; ld.global.u16 %rs14, [%rd35+12]; ld.global.u32 %r126, [%rd35+8]; ld.global.u64 %rd70, [%rd35]; setp.ne.s64 %p25, %rd70, 8; @%p25 bra $L__BB90_23; $L__BB90_27: cvt.u32.u64 %r53, %rd3; setp.ne.s32 %p26, %r53, -1; @%p26 bra $L__BB90_30; mov.u32 %r130, -1; add.s16 %rs7, %rs14, -1; setp.lt.u16 %p27, %rs7, 2; @%p27 bra $L__BB90_30; setp.ne.s16 %p28, %rs14, 1024; @%p28 bra $L__BB90_31; $L__BB90_30: setp.eq.s16 %p29, %rs14, 0; selp.b32 %r130, -1, %r126, %p29; $L__BB90_31: ld.local.u8 %rs8, [%rd1+56]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p30, %rs9, 0; @%p30 bra $L__BB90_33; bra.uni $L__BB90_32; $L__BB90_33: ld.global.v4.f32 {%f269, %f270, %f271, %f272}, [%rd4+-116]; ld.global.v4.f32 {%f277, %f278, %f279, %f280}, [%rd4+-100]; ld.global.v4.f32 {%f285, %f286, %f287, %f288}, [%rd4+-84]; mul.ftz.f32 %f293, %f1366, %f270; fma.rn.ftz.f32 %f294, %f1367, %f269, %f293; fma.rn.ftz.f32 %f295, %f1365, %f271, %f294; add.ftz.f32 %f7, %f272, %f295; mul.ftz.f32 %f296, %f1366, %f278; fma.rn.ftz.f32 %f297, %f1367, %f277, %f296; fma.rn.ftz.f32 %f298, %f1365, %f279, %f297; add.ftz.f32 %f8, %f280, %f298; mul.ftz.f32 %f299, %f1366, %f286; fma.rn.ftz.f32 %f300, %f1367, %f285, %f299; fma.rn.ftz.f32 %f301, %f1365, %f287, %f300; add.ftz.f32 %f1365, %f288, %f301; mov.f32 %f1367, %f7; mov.f32 %f1366, %f8; bra.uni $L__BB90_34; $L__BB90_108: ld.local.f32 %f1367, [%rd1+12]; ld.local.f32 %f1366, [%rd1+16]; ld.local.f32 %f1365, [%rd1+20]; ld.local.u8 %rs12, [%rd1+56]; and.b16 %rs13, %rs12, 2; setp.eq.s16 %p141, %rs13, 0; ld.local.s32 %rd20, [%rd1+76]; @%p141 bra $L__BB90_112; bra.uni $L__BB90_109; $L__BB90_112: cvt.u32.u64 %r120, %rd20; setp.ne.s32 %p144, %r120, -1; @%p144 bra $L__BB90_115; bra.uni $L__BB90_113; $L__BB90_115: ld.const.u64 %rd62, [kernel_params+2160]; cvta.to.global.u64 %rd63, %rd62; mul.lo.s64 %rd64, %rd20, 208; add.s64 %rd65, %rd63, %rd64; ld.global.v4.f32 {%f1297, %f1298, %f1299, %f1300}, [%rd65]; ld.global.v4.f32 {%f1304, %f1305, %f1306, %f1307}, [%rd65+16]; ld.global.v4.f32 {%f1311, %f1312, %f1313, %f1314}, [%rd65+32]; mul.ftz.f32 %f1318, %f1366, %f1304; fma.rn.ftz.f32 %f1319, %f1367, %f1297, %f1318; fma.rn.ftz.f32 %f1320, %f1365, %f1311, %f1319; mul.ftz.f32 %f1321, %f1366, %f1305; fma.rn.ftz.f32 %f1322, %f1367, %f1298, %f1321; fma.rn.ftz.f32 %f1323, %f1365, %f1312, %f1322; mul.ftz.f32 %f1324, %f1366, %f1306; fma.rn.ftz.f32 %f1325, %f1367, %f1299, %f1324; fma.rn.ftz.f32 %f1326, %f1365, %f1313, %f1325; mul.ftz.f32 %f1327, %f1323, %f1323; fma.rn.ftz.f32 %f1328, %f1320, %f1320, %f1327; fma.rn.ftz.f32 %f1329, %f1326, %f1326, %f1328; sqrt.approx.ftz.f32 %f1330, %f1329; div.approx.ftz.f32 %f1367, %f1320, %f1330; div.approx.ftz.f32 %f1366, %f1323, %f1330; div.approx.ftz.f32 %f1365, %f1326, %f1330; bra.uni $L__BB90_116; $L__BB90_38: ld.local.u32 %r60, [%rd1+76]; setp.eq.s32 %p34, %r60, -1; @%p34 bra $L__BB90_40; ld.local.f32 %f345, [%rd1+36]; ld.local.f32 %f346, [%rd1+12]; ld.local.f32 %f347, [%rd1+40]; ld.local.f32 %f348, [%rd1+16]; mul.ftz.f32 %f349, %f348, %f347; fma.rn.ftz.f32 %f350, %f346, %f345, %f349; ld.local.f32 %f351, [%rd1+44]; ld.local.f32 %f352, [%rd1+20]; fma.rn.ftz.f32 %f353, %f352, %f351, %f350; add.ftz.f32 %f354, %f353, %f353; mul.ftz.f32 %f355, %f346, %f354; mul.ftz.f32 %f356, %f348, %f354; mul.ftz.f32 %f357, %f352, %f354; sub.ftz.f32 %f1367, %f355, %f345; sub.ftz.f32 %f1366, %f356, %f347; sub.ftz.f32 %f1365, %f357, %f351; bra.uni $L__BB90_116; $L__BB90_101: ld.const.v4.f32 {%f1040, %f1041, %f1042, %f1043}, [kernel_params+688]; ld.const.v4.f32 {%f1044, %f1045, %f1046, %f1047}, [kernel_params+704]; ld.const.v4.f32 {%f1048, %f1049, %f1050, %f1051}, [kernel_params+720]; ld.local.u32 %r117, [%rd1+76]; setp.eq.s32 %p131, %r117, -1; ld.local.f32 %f1052, [%rd1+28]; ld.local.f32 %f1053, [%rd1+24]; setp.neu.ftz.f32 %p132, %f1053, %f1052; ld.local.f32 %f1054, [%rd1+32]; setp.neu.ftz.f32 %p133, %f1053, %f1054; or.pred %p134, %p132, %p133; sub.ftz.f32 %f1055, %f1054, %f1052; neg.ftz.f32 %f1056, %f1052; neg.ftz.f32 %f1057, %f1054; selp.f32 %f1058, %f1057, %f1054, %p134; add.ftz.f32 %f1059, %f1053, %f1058; selp.f32 %f1060, %f1052, %f1056, %p134; sub.ftz.f32 %f1061, %f1060, %f1053; mul.ftz.f32 %f1062, %f1059, %f1059; fma.rn.ftz.f32 %f1063, %f1055, %f1055, %f1062; fma.rn.ftz.f32 %f1064, %f1061, %f1061, %f1063; sqrt.approx.ftz.f32 %f1065, %f1064; div.approx.ftz.f32 %f1066, %f1055, %f1065; div.approx.ftz.f32 %f1067, %f1059, %f1065; div.approx.ftz.f32 %f1068, %f1061, %f1065; ld.local.f32 %f1069, [%rd1+92]; ld.local.f32 %f1070, [%rd1]; fma.rn.ftz.f32 %f178, %f1069, %f1066, %f1070; ld.local.f32 %f1071, [%rd1+4]; fma.rn.ftz.f32 %f179, %f1069, %f1067, %f1071; ld.local.f32 %f1072, [%rd1+8]; fma.rn.ftz.f32 %f180, %f1069, %f1068, %f1072; @%p131 bra $L__BB90_103; mul.ftz.f32 %f1073, %f1041, %f179; fma.rn.ftz.f32 %f1074, %f1040, %f178, %f1073; fma.rn.ftz.f32 %f1075, %f1042, %f180, %f1074; add.ftz.f32 %f1367, %f1043, %f1075; mul.ftz.f32 %f1076, %f1045, %f179; fma.rn.ftz.f32 %f1077, %f1044, %f178, %f1076; fma.rn.ftz.f32 %f1078, %f1046, %f180, %f1077; add.ftz.f32 %f1366, %f1047, %f1078; mul.ftz.f32 %f1079, %f1049, %f179; fma.rn.ftz.f32 %f1080, %f1048, %f178, %f1079; fma.rn.ftz.f32 %f1081, %f1050, %f180, %f1080; add.ftz.f32 %f1365, %f1051, %f1081; bra.uni $L__BB90_116; $L__BB90_36: ld.local.u32 %r22, [%rd1+76]; setp.eq.s32 %p32, %r22, -1; mov.f32 %f1365, 0f00000000; mov.f32 %f1367, %f1365; mov.f32 %f1366, %f1365; @%p32 bra $L__BB90_116; ld.const.u64 %rd42, [kernel_params+2160]; cvta.to.global.u64 %rd43, %rd42; mul.wide.s32 %rd44, %r22, 208; add.s64 %rd45, %rd43, %rd44; ld.global.f32 %f1366, [%rd45+144]; ld.global.f32 %f1367, [%rd45+140]; bra.uni $L__BB90_116; $L__BB90_109: cvt.u32.u64 %r118, %rd20; setp.ne.s32 %p142, %r118, -1; @%p142 bra $L__BB90_111; ld.local.u32 %r119, [%rd1+64]; setp.ne.s32 %p143, %r119, 32; @%p143 bra $L__BB90_116; $L__BB90_111: ld.local.v4.f32 {%f1229, %f1230, %f1231, %f1232}, [%rd1+144]; ld.local.v4.f32 {%f1236, %f1237, %f1238, %f1239}, [%rd1+160]; ld.local.v4.f32 {%f1243, %f1244, %f1245, %f1246}, [%rd1+176]; mul.ftz.f32 %f1250, %f1366, %f1236; fma.rn.ftz.f32 %f1251, %f1367, %f1229, %f1250; fma.rn.ftz.f32 %f1252, %f1365, %f1243, %f1251; mul.ftz.f32 %f1253, %f1366, %f1237; fma.rn.ftz.f32 %f1254, %f1367, %f1230, %f1253; fma.rn.ftz.f32 %f1255, %f1365, %f1244, %f1254; mul.ftz.f32 %f1256, %f1366, %f1238; fma.rn.ftz.f32 %f1257, %f1367, %f1231, %f1256; fma.rn.ftz.f32 %f1258, %f1365, %f1245, %f1257; mul.ftz.f32 %f1259, %f1255, %f1255; fma.rn.ftz.f32 %f1260, %f1252, %f1252, %f1259; fma.rn.ftz.f32 %f1261, %f1258, %f1258, %f1260; sqrt.approx.ftz.f32 %f1262, %f1261; div.approx.ftz.f32 %f1367, %f1252, %f1262; div.approx.ftz.f32 %f1366, %f1255, %f1262; div.approx.ftz.f32 %f1365, %f1258, %f1262; bra.uni $L__BB90_116; $L__BB90_9: ld.local.u32 %r62, [%rd1+76]; setp.ne.s32 %p39, %r62, -1; ld.const.u32 %r131, [kernel_params+16]; setp.ne.s32 %p40, %r131, 1; or.pred %p41, %p39, %p40; @%p41 bra $L__BB90_42; ld.local.v4.f32 {%f361, %f362, %f363, %f364}, [%rd1+240]; ld.const.v4.f32 {%f368, %f369, %f370, %f371}, [kernel_params+624]; ld.const.v4.f32 {%f376, %f377, %f378, %f379}, [kernel_params+640]; ld.const.v4.f32 {%f392, %f393, %f394, %f395}, [kernel_params+672]; mul.ftz.f32 %f400, %f362, %f369; fma.rn.ftz.f32 %f401, %f361, %f368, %f400; fma.rn.ftz.f32 %f402, %f363, %f370, %f371; add.ftz.f32 %f40, %f401, %f402; mul.ftz.f32 %f403, %f362, %f377; fma.rn.ftz.f32 %f404, %f361, %f376, %f403; fma.rn.ftz.f32 %f405, %f363, %f378, %f379; add.ftz.f32 %f41, %f404, %f405; mul.ftz.f32 %f409, %f362, %f393; fma.rn.ftz.f32 %f410, %f361, %f392, %f409; fma.rn.ftz.f32 %f411, %f363, %f394, %f395; add.ftz.f32 %f43, %f410, %f411; setp.eq.ftz.f32 %p42, %f43, 0f00000000; mov.f32 %f1365, 0f00000000; mov.f32 %f1367, %f1365; mov.f32 %f1366, %f1365; @%p42 bra $L__BB90_116; div.approx.ftz.f32 %f1366, %f41, %f43; div.approx.ftz.f32 %f1367, %f40, %f43; bra.uni $L__BB90_116; $L__BB90_42: ld.local.f32 %f413, [%rd1+28]; ld.local.f32 %f414, [%rd1+24]; setp.neu.ftz.f32 %p43, %f414, %f413; ld.local.f32 %f415, [%rd1+32]; setp.neu.ftz.f32 %p44, %f414, %f415; or.pred %p45, %p43, %p44; sub.ftz.f32 %f416, %f415, %f413; neg.ftz.f32 %f417, %f413; neg.ftz.f32 %f418, %f415; selp.f32 %f419, %f418, %f415, %p45; add.ftz.f32 %f420, %f414, %f419; selp.f32 %f421, %f413, %f417, %p45; sub.ftz.f32 %f422, %f421, %f414; mul.ftz.f32 %f423, %f420, %f420; fma.rn.ftz.f32 %f424, %f416, %f416, %f423; fma.rn.ftz.f32 %f425, %f422, %f422, %f424; sqrt.approx.ftz.f32 %f426, %f425; div.approx.ftz.f32 %f427, %f416, %f426; div.approx.ftz.f32 %f428, %f420, %f426; div.approx.ftz.f32 %f429, %f422, %f426; ld.local.f32 %f430, [%rd1+92]; ld.local.f32 %f431, [%rd1]; fma.rn.ftz.f32 %f1346, %f430, %f427, %f431; ld.local.f32 %f432, [%rd1+4]; fma.rn.ftz.f32 %f1347, %f430, %f428, %f432; ld.local.f32 %f433, [%rd1+8]; fma.rn.ftz.f32 %f1348, %f430, %f429, %f433; setp.eq.s32 %p46, %r131, 2; @%p46 bra $L__BB90_47; bra.uni $L__BB90_43; $L__BB90_47: ld.const.v4.f32 {%f496, %f497, %f498, %f499}, [kernel_params+688]; ld.const.v4.f32 {%f500, %f501, %f502, %f503}, [kernel_params+704]; ld.const.v4.f32 {%f504, %f505, %f506, %f507}, [kernel_params+720]; ld.local.u32 %r64, [%rd1+76]; setp.eq.s32 %p51, %r64, -1; mul.ftz.f32 %f508, %f1347, %f497; fma.rn.ftz.f32 %f509, %f1346, %f496, %f508; fma.rn.ftz.f32 %f80, %f1348, %f498, %f509; mul.ftz.f32 %f510, %f1347, %f501; fma.rn.ftz.f32 %f511, %f1346, %f500, %f510; fma.rn.ftz.f32 %f81, %f1348, %f502, %f511; mul.ftz.f32 %f512, %f1347, %f505; fma.rn.ftz.f32 %f513, %f1346, %f504, %f512; fma.rn.ftz.f32 %f82, %f1348, %f506, %f513; @%p51 bra $L__BB90_49; add.ftz.f32 %f514, %f499, %f80; add.ftz.f32 %f515, %f503, %f81; add.ftz.f32 %f516, %f507, %f82; mul.ftz.f32 %f517, %f515, %f515; fma.rn.ftz.f32 %f518, %f514, %f514, %f517; fma.rn.ftz.f32 %f519, %f516, %f516, %f518; sqrt.approx.ftz.f32 %f520, %f519; div.approx.ftz.f32 %f89, %f514, %f520; div.approx.ftz.f32 %f90, %f515, %f520; div.approx.ftz.f32 %f91, %f516, %f520; bra.uni $L__BB90_50; $L__BB90_40: ld.local.f32 %f1367, [%rd1+36]; ld.local.f32 %f1366, [%rd1+40]; ld.local.f32 %f1365, [%rd1+44]; bra.uni $L__BB90_116; $L__BB90_103: ld.const.v4.f32 {%f1082, %f1083, %f1084, %f1085}, [kernel_params+96]; ld.const.v4.f32 {%f1087, %f1088, %f1089, %f1090}, [kernel_params+112]; ld.const.v4.f32 {%f1092, %f1093, %f1094, %f1095}, [kernel_params+128]; add.ftz.f32 %f1097, %f178, %f1085; add.ftz.f32 %f1098, %f179, %f1090; add.ftz.f32 %f1099, %f180, %f1095; mul.ftz.f32 %f1100, %f1041, %f1098; fma.rn.ftz.f32 %f1101, %f1040, %f1097, %f1100; fma.rn.ftz.f32 %f1102, %f1042, %f1099, %f1101; add.ftz.f32 %f1367, %f1043, %f1102; mul.ftz.f32 %f1103, %f1045, %f1098; fma.rn.ftz.f32 %f1104, %f1044, %f1097, %f1103; fma.rn.ftz.f32 %f1105, %f1046, %f1099, %f1104; add.ftz.f32 %f1366, %f1047, %f1105; mul.ftz.f32 %f1106, %f1049, %f1098; fma.rn.ftz.f32 %f1107, %f1048, %f1097, %f1106; fma.rn.ftz.f32 %f1108, %f1050, %f1099, %f1107; add.ftz.f32 %f1365, %f1051, %f1108; bra.uni $L__BB90_116; $L__BB90_104: ld.local.u32 %r36, [%rd1+76]; setp.eq.s32 %p139, %r36, -1; @%p139 bra $L__BB90_116; ld.local.u8 %rs10, [%rd1+56]; and.b16 %rs11, %rs10, 2; setp.eq.s16 %p140, %rs11, 0; @%p140 bra $L__BB90_107; bra.uni $L__BB90_106; $L__BB90_107: ld.const.u64 %rd54, [kernel_params+2160]; cvta.to.global.u64 %rd55, %rd54; mul.wide.s32 %rd56, %r36, 208; add.s64 %rd57, %rd55, %rd56; ld.global.v4.f32 {%f1196, %f1197, %f1198, %f1199}, [%rd57+48]; ld.global.v4.f32 {%f1204, %f1205, %f1206, %f1207}, [%rd57+64]; ld.global.v4.f32 {%f1212, %f1213, %f1214, %f1215}, [%rd57+80]; mul.ftz.f32 %f1220, %f1366, %f1197; fma.rn.ftz.f32 %f1221, %f1367, %f1196, %f1220; fma.rn.ftz.f32 %f1222, %f1365, %f1198, %f1221; add.ftz.f32 %f196, %f1199, %f1222; mul.ftz.f32 %f1223, %f1366, %f1205; fma.rn.ftz.f32 %f1224, %f1367, %f1204, %f1223; fma.rn.ftz.f32 %f1225, %f1365, %f1206, %f1224; add.ftz.f32 %f197, %f1207, %f1225; mul.ftz.f32 %f1226, %f1366, %f1213; fma.rn.ftz.f32 %f1227, %f1367, %f1212, %f1226; fma.rn.ftz.f32 %f1228, %f1365, %f1214, %f1227; add.ftz.f32 %f1365, %f1215, %f1228; mov.f32 %f1366, %f197; mov.f32 %f1367, %f196; bra.uni $L__BB90_116; $L__BB90_43: ld.local.u32 %r63, [%rd1+76]; setp.ne.s32 %p47, %r63, -1; setp.ne.s32 %p48, %r131, 0; or.pred %p49, %p47, %p48; @%p49 bra $L__BB90_45; ld.const.v4.f32 {%f434, %f435, %f436, %f437}, [kernel_params+96]; ld.const.v4.f32 {%f439, %f440, %f441, %f442}, [kernel_params+112]; ld.const.v4.f32 {%f444, %f445, %f446, %f447}, [kernel_params+128]; add.ftz.f32 %f1346, %f1346, %f437; add.ftz.f32 %f1347, %f1347, %f442; add.ftz.f32 %f1348, %f1348, %f447; $L__BB90_45: ld.const.v4.f32 {%f452, %f453, %f454, %f455}, [kernel_params+624]; ld.const.v4.f32 {%f460, %f461, %f462, %f463}, [kernel_params+640]; ld.const.v4.f32 {%f476, %f477, %f478, %f479}, [kernel_params+672]; mul.ftz.f32 %f484, %f1347, %f453; fma.rn.ftz.f32 %f485, %f1346, %f452, %f484; fma.rn.ftz.f32 %f486, %f1348, %f454, %f455; add.ftz.f32 %f61, %f485, %f486; mul.ftz.f32 %f487, %f1347, %f461; fma.rn.ftz.f32 %f488, %f1346, %f460, %f487; fma.rn.ftz.f32 %f489, %f1348, %f462, %f463; add.ftz.f32 %f62, %f488, %f489; mul.ftz.f32 %f493, %f1347, %f477; fma.rn.ftz.f32 %f494, %f1346, %f476, %f493; fma.rn.ftz.f32 %f495, %f1348, %f478, %f479; add.ftz.f32 %f64, %f494, %f495; setp.eq.ftz.f32 %p50, %f64, 0f00000000; mov.f32 %f1367, 0f00000000; mov.f32 %f1366, %f1367; @%p50 bra $L__BB90_100; div.approx.ftz.f32 %f1366, %f62, %f64; div.approx.ftz.f32 %f1367, %f61, %f64; bra.uni $L__BB90_100; $L__BB90_32: ld.local.v4.f32 {%f236, %f237, %f238, %f239}, [%rd1+192]; mul.ftz.f32 %f244, %f1366, %f237; fma.rn.ftz.f32 %f245, %f1367, %f236, %f244; fma.rn.ftz.f32 %f246, %f1365, %f238, %f245; add.ftz.f32 %f4, %f239, %f246; ld.local.v4.f32 {%f247, %f248, %f249, %f250}, [%rd1+208]; mul.ftz.f32 %f255, %f1366, %f248; fma.rn.ftz.f32 %f256, %f1367, %f247, %f255; fma.rn.ftz.f32 %f257, %f1365, %f249, %f256; add.ftz.f32 %f5, %f250, %f257; ld.local.v4.f32 {%f258, %f259, %f260, %f261}, [%rd1+224]; mul.ftz.f32 %f266, %f1366, %f259; fma.rn.ftz.f32 %f267, %f1367, %f258, %f266; fma.rn.ftz.f32 %f268, %f1365, %f260, %f267; add.ftz.f32 %f1365, %f261, %f268; mov.f32 %f1367, %f4; mov.f32 %f1366, %f5; $L__BB90_34: setp.eq.s32 %p31, %r130, -1; @%p31 bra $L__BB90_116; ld.const.u64 %rd38, [kernel_params+2344]; cvta.to.global.u64 %rd39, %rd38; mul.wide.s32 %rd40, %r130, 16; add.s64 %rd41, %rd39, %rd40; ld.global.v4.f32 {%f302, %f303, %f304, %f305}, [%rd41]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd41+16]; ld.global.v4.f32 {%f318, %f319, %f320, %f321}, [%rd41+32]; mul.ftz.f32 %f326, %f1366, %f303; fma.rn.ftz.f32 %f327, %f1367, %f302, %f326; fma.rn.ftz.f32 %f328, %f1365, %f304, %f327; add.ftz.f32 %f13, %f305, %f328; mul.ftz.f32 %f329, %f1366, %f311; fma.rn.ftz.f32 %f330, %f1367, %f310, %f329; fma.rn.ftz.f32 %f331, %f1365, %f312, %f330; add.ftz.f32 %f14, %f313, %f331; mul.ftz.f32 %f332, %f1366, %f319; fma.rn.ftz.f32 %f333, %f1367, %f318, %f332; fma.rn.ftz.f32 %f334, %f1365, %f320, %f333; add.ftz.f32 %f1365, %f321, %f334; mov.f32 %f1366, %f14; mov.f32 %f1367, %f13; bra.uni $L__BB90_116; $L__BB90_113: ld.local.u32 %r121, [%rd1+64]; setp.ne.s32 %p145, %r121, 32; @%p145 bra $L__BB90_116; ld.const.u64 %rd58, [kernel_params+2368]; cvta.to.global.u64 %rd59, %rd58; ld.local.u32 %r122, [%rd1+80]; mul.wide.s32 %rd60, %r122, 192; add.s64 %rd61, %rd59, %rd60; ld.global.v4.f32 {%f1263, %f1264, %f1265, %f1266}, [%rd61+48]; ld.global.v4.f32 {%f1270, %f1271, %f1272, %f1273}, [%rd61+64]; ld.global.v4.f32 {%f1277, %f1278, %f1279, %f1280}, [%rd61+80]; mul.ftz.f32 %f1284, %f1366, %f1270; fma.rn.ftz.f32 %f1285, %f1367, %f1263, %f1284; fma.rn.ftz.f32 %f1286, %f1365, %f1277, %f1285; mul.ftz.f32 %f1287, %f1366, %f1271; fma.rn.ftz.f32 %f1288, %f1367, %f1264, %f1287; fma.rn.ftz.f32 %f1289, %f1365, %f1278, %f1288; mul.ftz.f32 %f1290, %f1366, %f1272; fma.rn.ftz.f32 %f1291, %f1367, %f1265, %f1290; fma.rn.ftz.f32 %f1292, %f1365, %f1279, %f1291; mul.ftz.f32 %f1293, %f1289, %f1289; fma.rn.ftz.f32 %f1294, %f1286, %f1286, %f1293; fma.rn.ftz.f32 %f1295, %f1292, %f1292, %f1294; sqrt.approx.ftz.f32 %f1296, %f1295; div.approx.ftz.f32 %f1367, %f1286, %f1296; div.approx.ftz.f32 %f1366, %f1289, %f1296; div.approx.ftz.f32 %f1365, %f1292, %f1296; bra.uni $L__BB90_116; $L__BB90_49: mul.ftz.f32 %f521, %f81, %f81; fma.rn.ftz.f32 %f522, %f80, %f80, %f521; fma.rn.ftz.f32 %f523, %f82, %f82, %f522; sqrt.approx.ftz.f32 %f524, %f523; div.approx.ftz.f32 %f89, %f80, %f524; div.approx.ftz.f32 %f90, %f81, %f524; div.approx.ftz.f32 %f91, %f82, %f524; $L__BB90_50: ld.const.u32 %r65, [kernel_params+20]; mov.f32 %f525, 0fBF800000; max.ftz.f32 %f526, %f89, %f525; mov.f32 %f527, 0f3F800000; min.ftz.f32 %f528, %f526, %f527; abs.ftz.f32 %f529, %f528; sub.ftz.f32 %f530, %f527, %f529; mul.ftz.f32 %f531, %f530, 0f3F000000; sqrt.approx.ftz.f32 %f532, %f531; setp.gt.ftz.f32 %p52, %f529, 0f3F11EB85; selp.f32 %f533, %f532, %f529, %p52; mul.ftz.f32 %f534, %f533, %f533; mov.f32 %f535, 0f3C94D2E9; mov.f32 %f536, 0f3D53F941; fma.rn.ftz.f32 %f537, %f536, %f534, %f535; mov.f32 %f538, 0f3D3F841F; fma.rn.ftz.f32 %f539, %f537, %f534, %f538; mov.f32 %f540, 0f3D994929; fma.rn.ftz.f32 %f541, %f539, %f534, %f540; mov.f32 %f542, 0f3E2AAB94; fma.rn.ftz.f32 %f543, %f541, %f534, %f542; mul.ftz.f32 %f544, %f534, %f543; fma.rn.ftz.f32 %f545, %f544, %f533, %f533; add.ftz.f32 %f546, %f545, %f545; mov.f32 %f547, 0f3FC90FDB; sub.ftz.f32 %f548, %f547, %f545; selp.f32 %f549, %f546, %f548, %p52; setp.lt.ftz.f32 %p53, %f528, 0f00000000; mov.f32 %f550, 0f40490FDB; sub.ftz.f32 %f551, %f550, %f549; selp.f32 %f92, %f551, %f549, %p53; abs.ftz.f32 %f93, %f90; setp.eq.ftz.f32 %p54, %f93, 0f00000000; abs.ftz.f32 %f94, %f91; setp.eq.ftz.f32 %p55, %f94, 0f00000000; and.pred %p1, %p54, %p55; mov.b32 %r27, %f90; mov.b32 %r66, %f91; and.b32 %r28, %r66, -2147483648; setp.gt.s32 %p56, %r65, 2; @%p56 bra $L__BB90_56; bra.uni $L__BB90_51; $L__BB90_56: setp.eq.s32 %p57, %r65, 3; @%p57 bra $L__BB90_65; setp.eq.s32 %p58, %r65, 5; @%p58 bra $L__BB90_68; setp.ne.s32 %p59, %r65, 4; @%p59 bra $L__BB90_94; ld.const.f32 %f552, [kernel_params+48]; ld.const.v4.f32 {%f553, %f554, %f555, %f556}, [kernel_params+64]; neg.ftz.f32 %f561, %f92; sub.ftz.f32 %f562, %f561, %f552; div.approx.ftz.f32 %f563, %f562, %f553; mul.ftz.f32 %f564, %f563, %f563; mul.ftz.f32 %f565, %f563, %f564; mul.ftz.f32 %f566, %f564, %f564; mul.ftz.f32 %f567, %f554, %f564; fma.rn.ftz.f32 %f568, %f553, %f563, %f567; mul.ftz.f32 %f569, %f556, %f566; fma.rn.ftz.f32 %f570, %f555, %f565, %f569; add.ftz.f32 %f571, %f568, %f570; add.ftz.f32 %f572, %f552, %f571; sub.ftz.f32 %f573, %f561, %f572; div.approx.ftz.f32 %f574, %f573, %f553; mul.ftz.f32 %f575, %f574, %f574; mul.ftz.f32 %f576, %f574, %f575; mul.ftz.f32 %f577, %f575, %f575; mul.ftz.f32 %f578, %f554, %f575; fma.rn.ftz.f32 %f579, %f553, %f574, %f578; mul.ftz.f32 %f580, %f556, %f577; fma.rn.ftz.f32 %f581, %f555, %f576, %f580; add.ftz.f32 %f582, %f579, %f581; add.ftz.f32 %f583, %f552, %f582; sub.ftz.f32 %f584, %f561, %f583; div.approx.ftz.f32 %f585, %f584, %f553; mul.ftz.f32 %f586, %f585, %f585; mul.ftz.f32 %f587, %f585, %f586; mul.ftz.f32 %f588, %f586, %f586; mul.ftz.f32 %f589, %f554, %f586; fma.rn.ftz.f32 %f590, %f553, %f585, %f589; mul.ftz.f32 %f591, %f556, %f588; fma.rn.ftz.f32 %f592, %f555, %f587, %f591; add.ftz.f32 %f593, %f590, %f592; add.ftz.f32 %f594, %f552, %f593; sub.ftz.f32 %f595, %f561, %f594; div.approx.ftz.f32 %f596, %f595, %f553; mul.ftz.f32 %f597, %f596, %f596; mul.ftz.f32 %f598, %f596, %f597; mul.ftz.f32 %f599, %f597, %f597; mul.ftz.f32 %f600, %f554, %f597; fma.rn.ftz.f32 %f601, %f553, %f596, %f600; mul.ftz.f32 %f602, %f556, %f599; fma.rn.ftz.f32 %f603, %f555, %f598, %f602; add.ftz.f32 %f604, %f601, %f603; add.ftz.f32 %f605, %f552, %f604; sub.ftz.f32 %f606, %f561, %f605; div.approx.ftz.f32 %f607, %f606, %f553; mul.ftz.f32 %f608, %f607, %f607; mul.ftz.f32 %f609, %f607, %f608; mul.ftz.f32 %f610, %f608, %f608; mul.ftz.f32 %f611, %f554, %f608; fma.rn.ftz.f32 %f612, %f553, %f607, %f611; mul.ftz.f32 %f613, %f556, %f610; fma.rn.ftz.f32 %f614, %f555, %f609, %f613; add.ftz.f32 %f615, %f612, %f614; add.ftz.f32 %f616, %f552, %f615; sub.ftz.f32 %f617, %f561, %f616; div.approx.ftz.f32 %f618, %f617, %f553; mul.ftz.f32 %f619, %f618, %f618; mul.ftz.f32 %f620, %f618, %f619; mul.ftz.f32 %f621, %f619, %f619; mul.ftz.f32 %f622, %f554, %f619; fma.rn.ftz.f32 %f623, %f553, %f618, %f622; mul.ftz.f32 %f624, %f556, %f621; fma.rn.ftz.f32 %f625, %f555, %f620, %f624; add.ftz.f32 %f626, %f623, %f625; add.ftz.f32 %f627, %f552, %f626; sub.ftz.f32 %f628, %f561, %f627; div.approx.ftz.f32 %f629, %f628, %f553; mul.ftz.f32 %f630, %f629, %f629; mul.ftz.f32 %f631, %f629, %f630; mul.ftz.f32 %f632, %f630, %f630; mul.ftz.f32 %f633, %f554, %f630; fma.rn.ftz.f32 %f634, %f553, %f629, %f633; mul.ftz.f32 %f635, %f556, %f632; fma.rn.ftz.f32 %f636, %f555, %f631, %f635; add.ftz.f32 %f637, %f634, %f636; add.ftz.f32 %f638, %f552, %f637; sub.ftz.f32 %f639, %f561, %f638; div.approx.ftz.f32 %f640, %f639, %f553; mul.ftz.f32 %f641, %f640, %f640; mul.ftz.f32 %f642, %f640, %f641; mul.ftz.f32 %f643, %f641, %f641; mul.ftz.f32 %f644, %f554, %f641; fma.rn.ftz.f32 %f645, %f553, %f640, %f644; mul.ftz.f32 %f646, %f556, %f643; fma.rn.ftz.f32 %f647, %f555, %f642, %f646; add.ftz.f32 %f648, %f645, %f647; add.ftz.f32 %f649, %f552, %f648; sub.ftz.f32 %f650, %f561, %f649; div.approx.ftz.f32 %f651, %f650, %f553; mul.ftz.f32 %f652, %f651, %f651; mul.ftz.f32 %f653, %f651, %f652; mul.ftz.f32 %f654, %f652, %f652; mul.ftz.f32 %f655, %f554, %f652; fma.rn.ftz.f32 %f656, %f553, %f651, %f655; mul.ftz.f32 %f657, %f556, %f654; fma.rn.ftz.f32 %f658, %f555, %f653, %f657; add.ftz.f32 %f659, %f656, %f658; add.ftz.f32 %f660, %f552, %f659; sub.ftz.f32 %f661, %f561, %f660; div.approx.ftz.f32 %f662, %f661, %f553; mul.ftz.f32 %f663, %f662, %f662; mul.ftz.f32 %f664, %f662, %f663; mul.ftz.f32 %f665, %f663, %f663; mul.ftz.f32 %f666, %f554, %f663; fma.rn.ftz.f32 %f667, %f553, %f662, %f666; mul.ftz.f32 %f668, %f556, %f665; fma.rn.ftz.f32 %f669, %f555, %f664, %f668; add.ftz.f32 %f670, %f667, %f669; add.ftz.f32 %f671, %f552, %f670; sub.ftz.f32 %f672, %f561, %f671; div.approx.ftz.f32 %f673, %f672, %f553; mul.ftz.f32 %f674, %f673, %f673; mul.ftz.f32 %f675, %f673, %f674; mul.ftz.f32 %f676, %f674, %f674; mul.ftz.f32 %f677, %f554, %f674; fma.rn.ftz.f32 %f678, %f553, %f673, %f677; mul.ftz.f32 %f679, %f556, %f676; fma.rn.ftz.f32 %f680, %f555, %f675, %f679; add.ftz.f32 %f681, %f678, %f680; add.ftz.f32 %f682, %f552, %f681; sub.ftz.f32 %f683, %f561, %f682; div.approx.ftz.f32 %f684, %f683, %f553; mul.ftz.f32 %f685, %f684, %f684; mul.ftz.f32 %f686, %f684, %f685; mul.ftz.f32 %f687, %f685, %f685; mul.ftz.f32 %f688, %f554, %f685; fma.rn.ftz.f32 %f689, %f553, %f684, %f688; mul.ftz.f32 %f690, %f556, %f687; fma.rn.ftz.f32 %f691, %f555, %f686, %f690; add.ftz.f32 %f692, %f689, %f691; add.ftz.f32 %f693, %f552, %f692; sub.ftz.f32 %f694, %f561, %f693; div.approx.ftz.f32 %f695, %f694, %f553; mul.ftz.f32 %f696, %f695, %f695; mul.ftz.f32 %f697, %f695, %f696; mul.ftz.f32 %f698, %f696, %f696; mul.ftz.f32 %f699, %f554, %f696; fma.rn.ftz.f32 %f700, %f553, %f695, %f699; mul.ftz.f32 %f701, %f556, %f698; fma.rn.ftz.f32 %f702, %f555, %f697, %f701; add.ftz.f32 %f703, %f700, %f702; add.ftz.f32 %f704, %f552, %f703; sub.ftz.f32 %f705, %f561, %f704; div.approx.ftz.f32 %f706, %f705, %f553; mul.ftz.f32 %f707, %f706, %f706; mul.ftz.f32 %f708, %f706, %f707; mul.ftz.f32 %f709, %f707, %f707; mul.ftz.f32 %f710, %f554, %f707; fma.rn.ftz.f32 %f711, %f553, %f706, %f710; mul.ftz.f32 %f712, %f556, %f709; fma.rn.ftz.f32 %f713, %f555, %f708, %f712; add.ftz.f32 %f714, %f711, %f713; add.ftz.f32 %f715, %f552, %f714; sub.ftz.f32 %f716, %f561, %f715; div.approx.ftz.f32 %f717, %f716, %f553; mul.ftz.f32 %f718, %f717, %f717; mul.ftz.f32 %f719, %f717, %f718; mul.ftz.f32 %f720, %f718, %f718; mul.ftz.f32 %f721, %f554, %f718; fma.rn.ftz.f32 %f722, %f553, %f717, %f721; mul.ftz.f32 %f723, %f556, %f720; fma.rn.ftz.f32 %f724, %f555, %f719, %f723; add.ftz.f32 %f725, %f722, %f724; add.ftz.f32 %f726, %f552, %f725; sub.ftz.f32 %f727, %f561, %f726; div.approx.ftz.f32 %f728, %f727, %f553; mul.ftz.f32 %f729, %f728, %f728; mul.ftz.f32 %f730, %f728, %f729; mul.ftz.f32 %f731, %f729, %f729; mul.ftz.f32 %f732, %f554, %f729; fma.rn.ftz.f32 %f733, %f553, %f728, %f732; mul.ftz.f32 %f734, %f556, %f731; fma.rn.ftz.f32 %f735, %f555, %f730, %f734; add.ftz.f32 %f736, %f733, %f735; add.ftz.f32 %f737, %f552, %f736; sub.ftz.f32 %f738, %f561, %f737; div.approx.ftz.f32 %f739, %f738, %f553; mul.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f739, %f740; mul.ftz.f32 %f742, %f740, %f740; mul.ftz.f32 %f743, %f554, %f740; fma.rn.ftz.f32 %f744, %f553, %f739, %f743; mul.ftz.f32 %f745, %f556, %f742; fma.rn.ftz.f32 %f746, %f555, %f741, %f745; add.ftz.f32 %f747, %f744, %f746; add.ftz.f32 %f748, %f552, %f747; sub.ftz.f32 %f749, %f561, %f748; div.approx.ftz.f32 %f750, %f749, %f553; mul.ftz.f32 %f751, %f750, %f750; mul.ftz.f32 %f752, %f750, %f751; mul.ftz.f32 %f753, %f751, %f751; mul.ftz.f32 %f754, %f554, %f751; fma.rn.ftz.f32 %f755, %f553, %f750, %f754; mul.ftz.f32 %f756, %f556, %f753; fma.rn.ftz.f32 %f757, %f555, %f752, %f756; add.ftz.f32 %f758, %f755, %f757; add.ftz.f32 %f759, %f552, %f758; sub.ftz.f32 %f760, %f561, %f759; div.approx.ftz.f32 %f761, %f760, %f553; mul.ftz.f32 %f762, %f761, %f761; mul.ftz.f32 %f763, %f761, %f762; mul.ftz.f32 %f764, %f762, %f762; mul.ftz.f32 %f765, %f554, %f762; fma.rn.ftz.f32 %f766, %f553, %f761, %f765; mul.ftz.f32 %f767, %f556, %f764; fma.rn.ftz.f32 %f768, %f555, %f763, %f767; add.ftz.f32 %f769, %f766, %f768; add.ftz.f32 %f770, %f552, %f769; sub.ftz.f32 %f771, %f561, %f770; div.approx.ftz.f32 %f772, %f771, %f553; mul.ftz.f32 %f773, %f772, %f772; mul.ftz.f32 %f774, %f772, %f773; mul.ftz.f32 %f775, %f773, %f773; mul.ftz.f32 %f776, %f554, %f773; fma.rn.ftz.f32 %f777, %f553, %f772, %f776; mul.ftz.f32 %f778, %f556, %f775; fma.rn.ftz.f32 %f779, %f555, %f774, %f778; add.ftz.f32 %f780, %f777, %f779; add.ftz.f32 %f781, %f552, %f780; sub.ftz.f32 %f782, %f561, %f781; div.approx.ftz.f32 %f97, %f782, %f553; @%p1 bra $L__BB90_63; bra.uni $L__BB90_60; $L__BB90_63: shr.s32 %r71, %r27, 31; and.b32 %r72, %r71, 1078530011; or.b32 %r73, %r72, %r28; mov.b32 %f1352, %r73; bra.uni $L__BB90_64; $L__BB90_51: setp.eq.s32 %p60, %r65, 0; @%p60 bra $L__BB90_87; setp.eq.s32 %p61, %r65, 1; @%p61 bra $L__BB90_53; bra.uni $L__BB90_94; $L__BB90_53: ld.const.f32 %f122, [kernel_params+24]; mul.ftz.f32 %f882, %f91, %f91; fma.rn.ftz.f32 %f883, %f90, %f90, %f882; sqrt.approx.ftz.f32 %f884, %f883; abs.ftz.f32 %f123, %f884; abs.ftz.f32 %f124, %f89; setp.eq.ftz.f32 %p90, %f124, 0f00000000; setp.eq.ftz.f32 %p91, %f123, 0f00000000; and.pred %p92, %p90, %p91; mov.b32 %r31, %f89; mov.b32 %r88, %f884; and.b32 %r32, %r88, -2147483648; @%p92 bra $L__BB90_80; bra.uni $L__BB90_54; $L__BB90_80: shr.s32 %r93, %r31, 31; and.b32 %r94, %r93, 1078530011; or.b32 %r95, %r32, %r94; mov.b32 %f1357, %r95; bra.uni $L__BB90_81; $L__BB90_106: ld.local.v4.f32 {%f1163, %f1164, %f1165, %f1166}, [%rd1+192]; mul.ftz.f32 %f1171, %f1366, %f1164; fma.rn.ftz.f32 %f1172, %f1367, %f1163, %f1171; fma.rn.ftz.f32 %f1173, %f1365, %f1165, %f1172; add.ftz.f32 %f193, %f1166, %f1173; ld.local.v4.f32 {%f1174, %f1175, %f1176, %f1177}, [%rd1+208]; mul.ftz.f32 %f1182, %f1366, %f1175; fma.rn.ftz.f32 %f1183, %f1367, %f1174, %f1182; fma.rn.ftz.f32 %f1184, %f1365, %f1176, %f1183; add.ftz.f32 %f194, %f1177, %f1184; ld.local.v4.f32 {%f1185, %f1186, %f1187, %f1188}, [%rd1+224]; mul.ftz.f32 %f1193, %f1366, %f1186; fma.rn.ftz.f32 %f1194, %f1367, %f1185, %f1193; fma.rn.ftz.f32 %f1195, %f1365, %f1187, %f1194; add.ftz.f32 %f1365, %f1188, %f1195; mov.f32 %f1366, %f194; mov.f32 %f1367, %f193; bra.uni $L__BB90_116; $L__BB90_65: add.ftz.f32 %f815, %f90, 0fBF800000; mul.ftz.f32 %f816, %f815, 0fBF000000; mov.f32 %f817, 0f00000000; max.ftz.f32 %f818, %f816, %f817; sqrt.approx.ftz.f32 %f819, %f818; add.ftz.f32 %f104, %f819, %f819; setp.leu.ftz.f32 %p69, %f104, 0f00000000; @%p69 bra $L__BB90_67; rcp.approx.ftz.f32 %f820, %f104; mul.ftz.f32 %f89, %f89, %f820; mul.ftz.f32 %f91, %f91, %f820; $L__BB90_67: add.ftz.f32 %f821, %f89, 0f3F800000; add.ftz.f32 %f822, %f91, 0f3F800000; mul.ftz.f32 %f1366, %f822, 0f3F000000; mul.ftz.f32 %f1367, %f821, 0f3F000000; bra.uni $L__BB90_100; $L__BB90_68: abs.ftz.f32 %f111, %f89; setp.eq.ftz.f32 %p71, %f111, 0f00000000; and.pred %p72, %p71, %p54; mov.b32 %r29, %f89; and.b32 %r30, %r27, -2147483648; @%p72 bra $L__BB90_72; bra.uni $L__BB90_69; $L__BB90_72: shr.s32 %r78, %r29, 31; and.b32 %r79, %r78, 1078530011; or.b32 %r80, %r79, %r30; mov.b32 %f1355, %r80; bra.uni $L__BB90_73; $L__BB90_87: setp.eq.ftz.f32 %p107, %f90, 0f00000000; setp.eq.ftz.f32 %p108, %f89, 0f00000000; and.pred %p109, %p108, %p107; setp.eq.ftz.f32 %p110, %f91, 0f00000000; and.pred %p111, %p110, %p109; mov.f32 %f1367, 0f00000000; mov.f32 %f1366, %f1367; @%p111 bra $L__BB90_100; ld.const.v4.f32 {%f941, %f942, %f943, %f944}, [kernel_params+32]; abs.ftz.f32 %f140, %f89; setp.eq.ftz.f32 %p113, %f140, 0f00000000; and.pred %p114, %p113, %p54; mov.b32 %r33, %f89; and.b32 %r34, %r27, -2147483648; @%p114 bra $L__BB90_92; bra.uni $L__BB90_89; $L__BB90_92: shr.s32 %r107, %r33, 31; and.b32 %r108, %r107, 1078530011; or.b32 %r109, %r108, %r34; mov.b32 %f1359, %r109; bra.uni $L__BB90_93; $L__BB90_94: ld.const.f32 %f1003, [kernel_params+28]; ld.const.f32 %f147, [kernel_params+276]; ld.const.f32 %f148, [kernel_params+280]; add.ftz.f32 %f1004, %f1003, %f1003; mul.ftz.f32 %f1005, %f92, 0f3F000000; sin.approx.ftz.f32 %f1006, %f1005; mul.ftz.f32 %f149, %f1004, %f1006; @%p1 bra $L__BB90_98; bra.uni $L__BB90_95; $L__BB90_98: shr.s32 %r114, %r27, 31; and.b32 %r115, %r114, 1078530011; or.b32 %r116, %r115, %r28; mov.b32 %f1360, %r116; bra.uni $L__BB90_99; $L__BB90_69: setp.eq.ftz.f32 %p73, %f111, 0f7F800000; setp.eq.ftz.f32 %p74, %f93, 0f7F800000; and.pred %p75, %p73, %p74; @%p75 bra $L__BB90_71; bra.uni $L__BB90_70; $L__BB90_71: setp.lt.s32 %p79, %r29, 0; selp.b32 %r76, 1075235812, 1061752795, %p79; or.b32 %r77, %r76, %r30; mov.b32 %f1355, %r77; bra.uni $L__BB90_73; $L__BB90_60: setp.eq.ftz.f32 %p62, %f93, 0f7F800000; setp.eq.ftz.f32 %p63, %f94, 0f7F800000; and.pred %p64, %p62, %p63; @%p64 bra $L__BB90_62; bra.uni $L__BB90_61; $L__BB90_62: setp.lt.s32 %p68, %r27, 0; selp.b32 %r69, 1075235812, 1061752795, %p68; or.b32 %r70, %r69, %r28; mov.b32 %f1352, %r70; bra.uni $L__BB90_64; $L__BB90_54: setp.eq.ftz.f32 %p93, %f124, 0f7F800000; setp.eq.ftz.f32 %p94, %f123, 0f7F800000; and.pred %p95, %p93, %p94; @%p95 bra $L__BB90_79; bra.uni $L__BB90_55; $L__BB90_79: setp.lt.s32 %p99, %r31, 0; selp.b32 %r91, 1075235812, 1061752795, %p99; or.b32 %r92, %r32, %r91; mov.b32 %f1357, %r92; bra.uni $L__BB90_81; $L__BB90_95: setp.eq.ftz.f32 %p124, %f93, 0f7F800000; setp.eq.ftz.f32 %p125, %f94, 0f7F800000; and.pred %p126, %p124, %p125; @%p126 bra $L__BB90_97; bra.uni $L__BB90_96; $L__BB90_97: setp.lt.s32 %p130, %r27, 0; selp.b32 %r112, 1075235812, 1061752795, %p130; or.b32 %r113, %r112, %r28; mov.b32 %f1360, %r113; bra.uni $L__BB90_99; $L__BB90_70: setp.lt.s32 %p76, %r29, 0; min.ftz.f32 %f823, %f93, %f111; max.ftz.f32 %f824, %f93, %f111; div.full.ftz.f32 %f825, %f823, %f824; mul.rn.ftz.f32 %f826, %f825, %f825; mov.f32 %f827, 0fC0B59883; mov.f32 %f828, 0fBF52C7EA; fma.rn.ftz.f32 %f829, %f826, %f828, %f827; mov.f32 %f830, 0fC0D21907; fma.rn.ftz.f32 %f831, %f829, %f826, %f830; mul.ftz.f32 %f832, %f826, %f831; mul.ftz.f32 %f833, %f825, %f832; add.ftz.f32 %f834, %f826, 0f41355DC0; mov.f32 %f835, 0f41E6BD60; fma.rn.ftz.f32 %f836, %f834, %f826, %f835; mov.f32 %f837, 0f419D92C8; fma.rn.ftz.f32 %f838, %f836, %f826, %f837; rcp.approx.ftz.f32 %f839, %f838; fma.rn.ftz.f32 %f840, %f833, %f839, %f825; sub.ftz.f32 %f842, %f547, %f840; setp.gt.ftz.f32 %p77, %f93, %f111; selp.f32 %f843, %f842, %f840, %p77; sub.ftz.f32 %f845, %f550, %f843; selp.f32 %f846, %f845, %f843, %p76; mov.b32 %r74, %f846; or.b32 %r75, %r30, %r74; mov.b32 %f847, %r75; add.ftz.f32 %f848, %f111, %f93; setp.le.ftz.f32 %p78, %f848, 0f7F800000; selp.f32 %f1355, %f847, %f848, %p78; $L__BB90_73: and.pred %p82, %p71, %p55; @%p82 bra $L__BB90_77; bra.uni $L__BB90_74; $L__BB90_77: shr.s32 %r85, %r29, 31; and.b32 %r86, %r85, 1078530011; or.b32 %r87, %r28, %r86; mov.b32 %f1356, %r87; bra.uni $L__BB90_78; $L__BB90_74: setp.eq.ftz.f32 %p83, %f111, 0f7F800000; setp.eq.ftz.f32 %p84, %f94, 0f7F800000; and.pred %p85, %p83, %p84; @%p85 bra $L__BB90_76; bra.uni $L__BB90_75; $L__BB90_76: setp.lt.s32 %p89, %r29, 0; selp.b32 %r83, 1075235812, 1061752795, %p89; or.b32 %r84, %r28, %r83; mov.b32 %f1356, %r84; bra.uni $L__BB90_78; $L__BB90_75: setp.lt.s32 %p86, %r29, 0; min.ftz.f32 %f849, %f94, %f111; max.ftz.f32 %f850, %f94, %f111; div.full.ftz.f32 %f851, %f849, %f850; mul.rn.ftz.f32 %f852, %f851, %f851; mov.f32 %f853, 0fC0B59883; mov.f32 %f854, 0fBF52C7EA; fma.rn.ftz.f32 %f855, %f852, %f854, %f853; mov.f32 %f856, 0fC0D21907; fma.rn.ftz.f32 %f857, %f855, %f852, %f856; mul.ftz.f32 %f858, %f852, %f857; mul.ftz.f32 %f859, %f851, %f858; add.ftz.f32 %f860, %f852, 0f41355DC0; mov.f32 %f861, 0f41E6BD60; fma.rn.ftz.f32 %f862, %f860, %f852, %f861; mov.f32 %f863, 0f419D92C8; fma.rn.ftz.f32 %f864, %f862, %f852, %f863; rcp.approx.ftz.f32 %f865, %f864; fma.rn.ftz.f32 %f866, %f859, %f865, %f851; sub.ftz.f32 %f868, %f547, %f866; setp.gt.ftz.f32 %p87, %f94, %f111; selp.f32 %f869, %f868, %f866, %p87; sub.ftz.f32 %f871, %f550, %f869; selp.f32 %f872, %f871, %f869, %p86; mov.b32 %r81, %f872; or.b32 %r82, %r28, %r81; mov.b32 %f873, %r82; add.ftz.f32 %f874, %f111, %f94; setp.le.ftz.f32 %p88, %f874, 0f7F800000; selp.f32 %f1356, %f873, %f874, %p88; $L__BB90_78: add.ftz.f32 %f875, %f1356, %f1356; div.approx.ftz.f32 %f877, %f875, %f550; add.ftz.f32 %f878, %f1355, %f1355; div.approx.ftz.f32 %f879, %f878, %f550; add.ftz.f32 %f880, %f879, 0f3F000000; sub.ftz.f32 %f1367, %f527, %f880; add.ftz.f32 %f1366, %f877, 0f3F000000; bra.uni $L__BB90_100; $L__BB90_89: setp.eq.ftz.f32 %p115, %f140, 0f7F800000; setp.eq.ftz.f32 %p116, %f93, 0f7F800000; and.pred %p117, %p115, %p116; @%p117 bra $L__BB90_91; bra.uni $L__BB90_90; $L__BB90_91: setp.lt.s32 %p121, %r33, 0; selp.b32 %r105, 1075235812, 1061752795, %p121; or.b32 %r106, %r105, %r34; mov.b32 %f1359, %r106; bra.uni $L__BB90_93; $L__BB90_61: mov.f32 %f1332, 0f40490FDB; mov.f32 %f1331, 0f3FC90FDB; setp.lt.s32 %p65, %r27, 0; min.ftz.f32 %f783, %f94, %f93; max.ftz.f32 %f784, %f94, %f93; div.full.ftz.f32 %f785, %f783, %f784; mul.rn.ftz.f32 %f786, %f785, %f785; mov.f32 %f787, 0fC0B59883; mov.f32 %f788, 0fBF52C7EA; fma.rn.ftz.f32 %f789, %f786, %f788, %f787; mov.f32 %f790, 0fC0D21907; fma.rn.ftz.f32 %f791, %f789, %f786, %f790; mul.ftz.f32 %f792, %f786, %f791; mul.ftz.f32 %f793, %f785, %f792; add.ftz.f32 %f794, %f786, 0f41355DC0; mov.f32 %f795, 0f41E6BD60; fma.rn.ftz.f32 %f796, %f794, %f786, %f795; mov.f32 %f797, 0f419D92C8; fma.rn.ftz.f32 %f798, %f796, %f786, %f797; rcp.approx.ftz.f32 %f799, %f798; fma.rn.ftz.f32 %f800, %f793, %f799, %f785; sub.ftz.f32 %f802, %f1331, %f800; setp.gt.ftz.f32 %p66, %f94, %f93; selp.f32 %f803, %f802, %f800, %p66; sub.ftz.f32 %f805, %f1332, %f803; selp.f32 %f806, %f805, %f803, %p65; mov.b32 %r67, %f806; or.b32 %r68, %r28, %r67; mov.b32 %f807, %r68; add.ftz.f32 %f808, %f93, %f94; setp.le.ftz.f32 %p67, %f808, 0f7F800000; selp.f32 %f1352, %f807, %f808, %p67; $L__BB90_64: ld.const.f32 %f1334, [kernel_params+280]; ld.const.f32 %f1333, [kernel_params+276]; cos.approx.ftz.f32 %f809, %f1352; mul.ftz.f32 %f810, %f97, %f809; div.approx.ftz.f32 %f811, %f810, %f1333; sin.approx.ftz.f32 %f812, %f1352; mul.ftz.f32 %f813, %f97, %f812; div.approx.ftz.f32 %f814, %f813, %f1334; add.ftz.f32 %f1366, %f814, 0f3F000000; add.ftz.f32 %f1367, %f811, 0f3F000000; bra.uni $L__BB90_100; $L__BB90_55: setp.lt.s32 %p96, %r31, 0; min.ftz.f32 %f885, %f123, %f124; max.ftz.f32 %f886, %f123, %f124; div.full.ftz.f32 %f887, %f885, %f886; mul.rn.ftz.f32 %f888, %f887, %f887; mov.f32 %f889, 0fC0B59883; mov.f32 %f890, 0fBF52C7EA; fma.rn.ftz.f32 %f891, %f888, %f890, %f889; mov.f32 %f892, 0fC0D21907; fma.rn.ftz.f32 %f893, %f891, %f888, %f892; mul.ftz.f32 %f894, %f888, %f893; mul.ftz.f32 %f895, %f887, %f894; add.ftz.f32 %f896, %f888, 0f41355DC0; mov.f32 %f897, 0f41E6BD60; fma.rn.ftz.f32 %f898, %f896, %f888, %f897; mov.f32 %f899, 0f419D92C8; fma.rn.ftz.f32 %f900, %f898, %f888, %f899; rcp.approx.ftz.f32 %f901, %f900; fma.rn.ftz.f32 %f902, %f895, %f901, %f887; sub.ftz.f32 %f904, %f547, %f902; setp.gt.ftz.f32 %p97, %f123, %f124; selp.f32 %f905, %f904, %f902, %p97; sub.ftz.f32 %f907, %f550, %f905; selp.f32 %f908, %f907, %f905, %p96; mov.b32 %r89, %f908; or.b32 %r90, %r32, %r89; mov.b32 %f909, %r90; add.ftz.f32 %f910, %f124, %f123; setp.le.ftz.f32 %p98, %f910, 0f7F800000; selp.f32 %f1357, %f909, %f910, %p98; $L__BB90_81: div.approx.ftz.f32 %f129, %f1357, %f122; @%p1 bra $L__BB90_85; bra.uni $L__BB90_82; $L__BB90_85: shr.s32 %r100, %r27, 31; and.b32 %r101, %r100, 1078530011; or.b32 %r102, %r101, %r28; mov.b32 %f1358, %r102; bra.uni $L__BB90_86; $L__BB90_82: setp.eq.ftz.f32 %p100, %f93, 0f7F800000; setp.eq.ftz.f32 %p101, %f94, 0f7F800000; and.pred %p102, %p100, %p101; @%p102 bra $L__BB90_84; bra.uni $L__BB90_83; $L__BB90_84: setp.lt.s32 %p106, %r27, 0; selp.b32 %r98, 1075235812, 1061752795, %p106; or.b32 %r99, %r98, %r28; mov.b32 %f1358, %r99; bra.uni $L__BB90_86; $L__BB90_96: setp.lt.s32 %p127, %r27, 0; min.ftz.f32 %f1007, %f94, %f93; max.ftz.f32 %f1008, %f94, %f93; div.full.ftz.f32 %f1009, %f1007, %f1008; mul.rn.ftz.f32 %f1010, %f1009, %f1009; mov.f32 %f1011, 0fC0B59883; mov.f32 %f1012, 0fBF52C7EA; fma.rn.ftz.f32 %f1013, %f1010, %f1012, %f1011; mov.f32 %f1014, 0fC0D21907; fma.rn.ftz.f32 %f1015, %f1013, %f1010, %f1014; mul.ftz.f32 %f1016, %f1010, %f1015; mul.ftz.f32 %f1017, %f1009, %f1016; add.ftz.f32 %f1018, %f1010, 0f41355DC0; mov.f32 %f1019, 0f41E6BD60; fma.rn.ftz.f32 %f1020, %f1018, %f1010, %f1019; mov.f32 %f1021, 0f419D92C8; fma.rn.ftz.f32 %f1022, %f1020, %f1010, %f1021; rcp.approx.ftz.f32 %f1023, %f1022; fma.rn.ftz.f32 %f1024, %f1017, %f1023, %f1009; sub.ftz.f32 %f1026, %f547, %f1024; setp.gt.ftz.f32 %p128, %f94, %f93; selp.f32 %f1027, %f1026, %f1024, %p128; sub.ftz.f32 %f1029, %f550, %f1027; selp.f32 %f1030, %f1029, %f1027, %p127; mov.b32 %r110, %f1030; or.b32 %r111, %r28, %r110; mov.b32 %f1031, %r111; add.ftz.f32 %f1032, %f93, %f94; setp.le.ftz.f32 %p129, %f1032, 0f7F800000; selp.f32 %f1360, %f1031, %f1032, %p129; $L__BB90_99: cos.approx.ftz.f32 %f1033, %f1360; mul.ftz.f32 %f1034, %f149, %f1033; div.approx.ftz.f32 %f1035, %f1034, %f147; sin.approx.ftz.f32 %f1036, %f1360; mul.ftz.f32 %f1037, %f149, %f1036; div.approx.ftz.f32 %f1038, %f1037, %f148; add.ftz.f32 %f1366, %f1038, 0f3F000000; add.ftz.f32 %f1367, %f1035, 0f3F000000; bra.uni $L__BB90_100; $L__BB90_83: setp.lt.s32 %p103, %r27, 0; min.ftz.f32 %f911, %f94, %f93; max.ftz.f32 %f912, %f94, %f93; div.full.ftz.f32 %f913, %f911, %f912; mul.rn.ftz.f32 %f914, %f913, %f913; mov.f32 %f915, 0fC0B59883; mov.f32 %f916, 0fBF52C7EA; fma.rn.ftz.f32 %f917, %f914, %f916, %f915; mov.f32 %f918, 0fC0D21907; fma.rn.ftz.f32 %f919, %f917, %f914, %f918; mul.ftz.f32 %f920, %f914, %f919; mul.ftz.f32 %f921, %f913, %f920; add.ftz.f32 %f922, %f914, 0f41355DC0; mov.f32 %f923, 0f41E6BD60; fma.rn.ftz.f32 %f924, %f922, %f914, %f923; mov.f32 %f925, 0f419D92C8; fma.rn.ftz.f32 %f926, %f924, %f914, %f925; rcp.approx.ftz.f32 %f927, %f926; fma.rn.ftz.f32 %f928, %f921, %f927, %f913; sub.ftz.f32 %f930, %f547, %f928; setp.gt.ftz.f32 %p104, %f94, %f93; selp.f32 %f931, %f930, %f928, %p104; sub.ftz.f32 %f933, %f550, %f931; selp.f32 %f934, %f933, %f931, %p103; mov.b32 %r96, %f934; or.b32 %r97, %r28, %r96; mov.b32 %f935, %r97; add.ftz.f32 %f936, %f93, %f94; setp.le.ftz.f32 %p105, %f936, 0f7F800000; selp.f32 %f1358, %f935, %f936, %p105; $L__BB90_86: cos.approx.ftz.f32 %f937, %f1358; sin.approx.ftz.f32 %f938, %f1358; fma.rn.ftz.f32 %f1366, %f129, %f938, 0f3F000000; fma.rn.ftz.f32 %f1367, %f129, %f937, 0f3F000000; bra.uni $L__BB90_100; $L__BB90_90: setp.lt.s32 %p118, %r33, 0; min.ftz.f32 %f945, %f93, %f140; max.ftz.f32 %f946, %f93, %f140; div.full.ftz.f32 %f947, %f945, %f946; mul.rn.ftz.f32 %f948, %f947, %f947; mov.f32 %f949, 0fC0B59883; mov.f32 %f950, 0fBF52C7EA; fma.rn.ftz.f32 %f951, %f948, %f950, %f949; mov.f32 %f952, 0fC0D21907; fma.rn.ftz.f32 %f953, %f951, %f948, %f952; mul.ftz.f32 %f954, %f948, %f953; mul.ftz.f32 %f955, %f947, %f954; add.ftz.f32 %f956, %f948, 0f41355DC0; mov.f32 %f957, 0f41E6BD60; fma.rn.ftz.f32 %f958, %f956, %f948, %f957; mov.f32 %f959, 0f419D92C8; fma.rn.ftz.f32 %f960, %f958, %f948, %f959; rcp.approx.ftz.f32 %f961, %f960; fma.rn.ftz.f32 %f962, %f955, %f961, %f947; sub.ftz.f32 %f964, %f547, %f962; setp.gt.ftz.f32 %p119, %f93, %f140; selp.f32 %f965, %f964, %f962, %p119; sub.ftz.f32 %f967, %f550, %f965; selp.f32 %f968, %f967, %f965, %p118; mov.b32 %r103, %f968; or.b32 %r104, %r34, %r103; mov.b32 %f969, %r104; add.ftz.f32 %f970, %f140, %f93; setp.le.ftz.f32 %p120, %f970, 0f7F800000; selp.f32 %f1359, %f969, %f970, %p120; $L__BB90_93: sub.ftz.f32 %f971, %f1359, %f942; mul.ftz.f32 %f972, %f90, %f90; fma.rn.ftz.f32 %f973, %f89, %f89, %f972; fma.rn.ftz.f32 %f974, %f91, %f91, %f973; sqrt.approx.ftz.f32 %f975, %f974; div.approx.ftz.f32 %f976, %f91, %f975; abs.ftz.f32 %f977, %f976; sub.ftz.f32 %f979, %f527, %f977; mul.ftz.f32 %f980, %f979, 0f3F000000; sqrt.approx.ftz.f32 %f981, %f980; setp.gt.ftz.f32 %p122, %f977, 0f3F11EB85; selp.f32 %f982, %f981, %f977, %p122; mul.ftz.f32 %f983, %f982, %f982; fma.rn.ftz.f32 %f986, %f536, %f983, %f535; fma.rn.ftz.f32 %f988, %f986, %f983, %f538; fma.rn.ftz.f32 %f990, %f988, %f983, %f540; fma.rn.ftz.f32 %f992, %f990, %f983, %f542; mul.ftz.f32 %f993, %f983, %f992; fma.rn.ftz.f32 %f994, %f993, %f982, %f982; add.ftz.f32 %f995, %f994, %f994; sub.ftz.f32 %f997, %f547, %f994; selp.f32 %f998, %f995, %f997, %p122; setp.lt.ftz.f32 %p123, %f976, 0f00000000; sub.ftz.f32 %f1000, %f550, %f998; selp.f32 %f1001, %f1000, %f998, %p123; sub.ftz.f32 %f1002, %f1001, %f944; div.approx.ftz.f32 %f1366, %f1002, %f943; div.approx.ftz.f32 %f1367, %f971, %f941; $L__BB90_100: ld.param.u32 %r132, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; mov.f32 %f1365, 0f00000000; $L__BB90_116: ld.param.u32 %r124, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+8]; ld.param.u64 %rd69, [_Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2]; cvta.to.local.u64 %rd66, %rd69; mul.wide.u32 %rd67, %r124, 4; add.s64 %rd68, %rd66, %rd67; st.local.f32 [%rd68], %f1367; st.local.f32 [%rd68+4], %f1366; st.local.f32 [%rd68+8], %f1365; st.param.b32 [func_retval0+0], %r132; ret; } .func (.param .b32 func_retval0) _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i( .param .b64 _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0, .param .b32 _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1, .param .b64 _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2, .param .align 4 .b8 _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3[16], .param .b32 _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4 ) { .reg .pred %p<146>; .reg .b16 %rs<16>; .reg .f32 %f<1404>; .reg .b32 %r<133>; .reg .b64 %rd<71>; ld.param.u64 %rd22, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_0]; ld.param.u32 %r38, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_1]; ld.param.u32 %r132, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; ld.param.u32 %r42, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+12]; ld.param.u32 %r40, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+4]; cvta.to.local.u64 %rd1, %rd22; setp.gt.s32 %p2, %r40, 3; @%p2 bra $L__BB91_12; setp.gt.s32 %p8, %r40, 1; @%p8 bra $L__BB91_6; setp.eq.s32 %p11, %r40, 0; @%p11 bra $L__BB91_108; setp.eq.s32 %p12, %r40, 1; @%p12 bra $L__BB91_4; bra.uni $L__BB91_116; $L__BB91_4: ld.local.f32 %f1136, [%rd1+28]; ld.local.f32 %f1137, [%rd1+24]; setp.neu.ftz.f32 %p135, %f1137, %f1136; ld.local.f32 %f1138, [%rd1+32]; setp.neu.ftz.f32 %p136, %f1137, %f1138; or.pred %p137, %p135, %p136; sub.ftz.f32 %f1139, %f1138, %f1136; neg.ftz.f32 %f1140, %f1136; neg.ftz.f32 %f1141, %f1138; selp.f32 %f1142, %f1141, %f1138, %p137; add.ftz.f32 %f1143, %f1137, %f1142; selp.f32 %f1144, %f1136, %f1140, %p137; sub.ftz.f32 %f1145, %f1144, %f1137; mul.ftz.f32 %f1146, %f1143, %f1143; fma.rn.ftz.f32 %f1147, %f1139, %f1139, %f1146; fma.rn.ftz.f32 %f1148, %f1145, %f1145, %f1147; sqrt.approx.ftz.f32 %f1149, %f1148; div.approx.ftz.f32 %f1150, %f1139, %f1149; div.approx.ftz.f32 %f1151, %f1143, %f1149; div.approx.ftz.f32 %f1152, %f1145, %f1149; mul.ftz.f32 %f1153, %f1136, %f1152; mul.ftz.f32 %f1154, %f1138, %f1151; sub.ftz.f32 %f1155, %f1153, %f1154; mul.ftz.f32 %f1156, %f1138, %f1150; mul.ftz.f32 %f1157, %f1137, %f1152; sub.ftz.f32 %f1158, %f1156, %f1157; mul.ftz.f32 %f1159, %f1137, %f1151; mul.ftz.f32 %f1160, %f1136, %f1150; sub.ftz.f32 %f1161, %f1159, %f1160; ld.local.f32 %f1162, [%rd1+92]; ld.local.f32 %f1163, [%rd1]; fma.rn.ftz.f32 %f1403, %f1162, %f1155, %f1163; ld.local.f32 %f1164, [%rd1+4]; fma.rn.ftz.f32 %f1402, %f1162, %f1158, %f1164; ld.local.f32 %f1165, [%rd1+8]; fma.rn.ftz.f32 %f1401, %f1162, %f1161, %f1165; setp.eq.s32 %p138, %r42, 0; @%p138 bra $L__BB91_104; ld.const.u64 %rd50, [kernel_params+2400]; cvta.to.global.u64 %rd51, %rd50; mul.wide.s32 %rd52, %r132, 16; add.s64 %rd53, %rd51, %rd52; ld.global.v4.f32 {%f1166, %f1167, %f1168, %f1169}, [%rd53]; ld.global.v4.f32 {%f1174, %f1175, %f1176, %f1177}, [%rd53+16]; ld.global.v4.f32 {%f1182, %f1183, %f1184, %f1185}, [%rd53+32]; add.s32 %r132, %r132, 3; mul.ftz.f32 %f1190, %f1402, %f1167; fma.rn.ftz.f32 %f1191, %f1403, %f1166, %f1190; fma.rn.ftz.f32 %f1192, %f1401, %f1168, %f1191; add.ftz.f32 %f190, %f1192, %f1169; mul.ftz.f32 %f1193, %f1402, %f1175; fma.rn.ftz.f32 %f1194, %f1403, %f1174, %f1193; fma.rn.ftz.f32 %f1195, %f1401, %f1176, %f1194; add.ftz.f32 %f191, %f1195, %f1177; mul.ftz.f32 %f1196, %f1402, %f1183; fma.rn.ftz.f32 %f1197, %f1403, %f1182, %f1196; fma.rn.ftz.f32 %f1198, %f1401, %f1184, %f1197; add.ftz.f32 %f1401, %f1198, %f1185; mov.f32 %f1402, %f191; mov.f32 %f1403, %f190; bra.uni $L__BB91_116; $L__BB91_12: setp.gt.s32 %p3, %r40, 5; @%p3 bra $L__BB91_17; setp.eq.s32 %p6, %r40, 4; @%p6 bra $L__BB91_38; setp.eq.s32 %p7, %r40, 5; @%p7 bra $L__BB91_15; bra.uni $L__BB91_116; $L__BB91_15: ld.local.u32 %r23, [%rd1+76]; setp.eq.s32 %p33, %r23, -1; mov.f32 %f1403, 0f00000000; mov.f32 %f1402, %f1403; mov.f32 %f1401, %f1403; @%p33 bra $L__BB91_116; ld.const.u64 %rd46, [kernel_params+2160]; cvta.to.global.u64 %rd47, %rd46; mul.wide.s32 %rd48, %r23, 208; add.s64 %rd49, %rd47, %rd48; ld.global.v4.f32 {%f1403, %f1402, %f1401, %f353}, [%rd49+128]; bra.uni $L__BB91_116; $L__BB91_6: setp.eq.s32 %p9, %r40, 2; @%p9 bra $L__BB91_101; setp.eq.s32 %p10, %r40, 3; @%p10 bra $L__BB91_8; bra.uni $L__BB91_116; $L__BB91_8: and.b32 %r61, %r38, 1; setp.eq.b32 %p35, %r61, 1; mov.pred %p36, 0; xor.pred %p37, %p35, %p36; not.pred %p38, %p37; @%p38 bra $L__BB91_41; bra.uni $L__BB91_9; $L__BB91_41: ld.const.u32 %r131, [kernel_params+16]; bra.uni $L__BB91_42; $L__BB91_17: setp.eq.s32 %p4, %r40, 6; @%p4 bra $L__BB91_36; setp.ne.s32 %p5, %r40, 7; @%p5 bra $L__BB91_116; ld.local.f32 %f215, [%rd1+28]; ld.local.f32 %f216, [%rd1+24]; setp.neu.ftz.f32 %p13, %f216, %f215; ld.local.f32 %f217, [%rd1+32]; setp.neu.ftz.f32 %p14, %f216, %f217; or.pred %p15, %p13, %p14; sub.ftz.f32 %f218, %f217, %f215; neg.ftz.f32 %f219, %f215; neg.ftz.f32 %f220, %f217; selp.f32 %f221, %f220, %f217, %p15; add.ftz.f32 %f222, %f216, %f221; selp.f32 %f223, %f215, %f219, %p15; sub.ftz.f32 %f224, %f223, %f216; mul.ftz.f32 %f225, %f222, %f222; fma.rn.ftz.f32 %f226, %f218, %f218, %f225; fma.rn.ftz.f32 %f227, %f224, %f224, %f226; sqrt.approx.ftz.f32 %f228, %f227; div.approx.ftz.f32 %f229, %f218, %f228; div.approx.ftz.f32 %f230, %f222, %f228; div.approx.ftz.f32 %f231, %f224, %f228; mul.ftz.f32 %f232, %f215, %f231; mul.ftz.f32 %f233, %f217, %f230; sub.ftz.f32 %f234, %f232, %f233; mul.ftz.f32 %f235, %f217, %f229; mul.ftz.f32 %f236, %f216, %f231; sub.ftz.f32 %f237, %f235, %f236; mul.ftz.f32 %f238, %f216, %f230; mul.ftz.f32 %f239, %f215, %f229; sub.ftz.f32 %f240, %f238, %f239; ld.local.f32 %f241, [%rd1+92]; ld.local.f32 %f242, [%rd1]; fma.rn.ftz.f32 %f1403, %f241, %f234, %f242; ld.local.f32 %f243, [%rd1+4]; fma.rn.ftz.f32 %f1402, %f241, %f237, %f243; ld.local.f32 %f244, [%rd1+8]; fma.rn.ftz.f32 %f1401, %f241, %f240, %f244; ld.local.u32 %r1, [%rd1+76]; setp.eq.s32 %p16, %r1, -1; @%p16 bra $L__BB91_116; ld.local.s32 %rd3, [%rd1+60]; ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.wide.s32 %rd25, %r1, 208; add.s64 %rd26, %rd24, %rd25; add.s64 %rd4, %rd26, 164; ld.global.u32 %r2, [%rd26+164]; ld.local.u32 %r46, [%rd1+64]; and.b32 %r47, %r46, 1; setp.eq.b32 %p17, %r47, 1; not.pred %p18, %p17; mov.u32 %r125, 0; setp.eq.s64 %p19, %rd3, -1; or.pred %p20, %p19, %p18; @%p20 bra $L__BB91_22; ld.const.u64 %rd27, [kernel_params+2240]; cvta.to.global.u64 %rd28, %rd27; shl.b64 %rd29, %rd3, 2; add.s64 %rd30, %rd28, %rd29; ld.global.u32 %r48, [%rd30]; setp.ne.s32 %p21, %r48, -1; selp.u32 %r125, 1, 0, %p21; $L__BB91_22: ld.const.u64 %rd31, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd31; add.s32 %r127, %r125, %r2; mul.wide.u32 %rd32, %r127, 16; add.s64 %rd33, %rd5, %rd32; ld.global.u16 %rs14, [%rd33+12]; ld.global.u32 %r126, [%rd33+8]; ld.global.u64 %rd70, [%rd33]; setp.eq.s64 %p22, %rd70, 8; @%p22 bra $L__BB91_27; $L__BB91_23: setp.eq.s64 %p23, %rd70, 0; @%p23 bra $L__BB91_25; add.s32 %r127, %r127, 2; bra.uni $L__BB91_26; $L__BB91_25: setp.eq.s16 %p24, %rs14, 0; mov.u32 %r130, -1; mov.u32 %r127, %r126; @%p24 bra $L__BB91_31; $L__BB91_26: mul.wide.u32 %rd34, %r127, 16; add.s64 %rd35, %rd5, %rd34; ld.global.u16 %rs14, [%rd35+12]; ld.global.u32 %r126, [%rd35+8]; ld.global.u64 %rd70, [%rd35]; setp.ne.s64 %p25, %rd70, 8; @%p25 bra $L__BB91_23; $L__BB91_27: cvt.u32.u64 %r53, %rd3; setp.ne.s32 %p26, %r53, -1; @%p26 bra $L__BB91_30; mov.u32 %r130, -1; add.s16 %rs7, %rs14, -1; setp.lt.u16 %p27, %rs7, 2; @%p27 bra $L__BB91_30; setp.ne.s16 %p28, %rs14, 1024; @%p28 bra $L__BB91_31; $L__BB91_30: setp.eq.s16 %p29, %rs14, 0; selp.b32 %r130, -1, %r126, %p29; $L__BB91_31: ld.local.u8 %rs8, [%rd1+56]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p30, %rs9, 0; @%p30 bra $L__BB91_33; bra.uni $L__BB91_32; $L__BB91_33: ld.global.v4.f32 {%f278, %f279, %f280, %f281}, [%rd4+-116]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd4+-100]; ld.global.v4.f32 {%f294, %f295, %f296, %f297}, [%rd4+-84]; mul.ftz.f32 %f302, %f1402, %f279; fma.rn.ftz.f32 %f303, %f1403, %f278, %f302; fma.rn.ftz.f32 %f304, %f1401, %f280, %f303; add.ftz.f32 %f7, %f281, %f304; mul.ftz.f32 %f305, %f1402, %f287; fma.rn.ftz.f32 %f306, %f1403, %f286, %f305; fma.rn.ftz.f32 %f307, %f1401, %f288, %f306; add.ftz.f32 %f8, %f289, %f307; mul.ftz.f32 %f308, %f1402, %f295; fma.rn.ftz.f32 %f309, %f1403, %f294, %f308; fma.rn.ftz.f32 %f310, %f1401, %f296, %f309; add.ftz.f32 %f1401, %f297, %f310; mov.f32 %f1403, %f7; mov.f32 %f1402, %f8; bra.uni $L__BB91_34; $L__BB91_108: ld.local.f32 %f1403, [%rd1+12]; ld.local.f32 %f1402, [%rd1+16]; ld.local.f32 %f1401, [%rd1+20]; ld.local.u8 %rs12, [%rd1+56]; and.b16 %rs13, %rs12, 2; setp.eq.s16 %p141, %rs13, 0; ld.local.s32 %rd20, [%rd1+76]; @%p141 bra $L__BB91_112; bra.uni $L__BB91_109; $L__BB91_112: cvt.u32.u64 %r120, %rd20; setp.ne.s32 %p144, %r120, -1; @%p144 bra $L__BB91_115; bra.uni $L__BB91_113; $L__BB91_115: ld.const.u64 %rd62, [kernel_params+2160]; cvta.to.global.u64 %rd63, %rd62; mul.lo.s64 %rd64, %rd20, 208; add.s64 %rd65, %rd63, %rd64; ld.global.v4.f32 {%f1333, %f1334, %f1335, %f1336}, [%rd65]; ld.global.v4.f32 {%f1340, %f1341, %f1342, %f1343}, [%rd65+16]; ld.global.v4.f32 {%f1347, %f1348, %f1349, %f1350}, [%rd65+32]; mul.ftz.f32 %f1354, %f1402, %f1340; fma.rn.ftz.f32 %f1355, %f1403, %f1333, %f1354; fma.rn.ftz.f32 %f1356, %f1401, %f1347, %f1355; mul.ftz.f32 %f1357, %f1402, %f1341; fma.rn.ftz.f32 %f1358, %f1403, %f1334, %f1357; fma.rn.ftz.f32 %f1359, %f1401, %f1348, %f1358; mul.ftz.f32 %f1360, %f1402, %f1342; fma.rn.ftz.f32 %f1361, %f1403, %f1335, %f1360; fma.rn.ftz.f32 %f1362, %f1401, %f1349, %f1361; mul.ftz.f32 %f1363, %f1359, %f1359; fma.rn.ftz.f32 %f1364, %f1356, %f1356, %f1363; fma.rn.ftz.f32 %f1365, %f1362, %f1362, %f1364; sqrt.approx.ftz.f32 %f1366, %f1365; div.approx.ftz.f32 %f1403, %f1356, %f1366; div.approx.ftz.f32 %f1402, %f1359, %f1366; div.approx.ftz.f32 %f1401, %f1362, %f1366; bra.uni $L__BB91_116; $L__BB91_38: ld.local.u32 %r60, [%rd1+76]; setp.eq.s32 %p34, %r60, -1; @%p34 bra $L__BB91_40; ld.local.f32 %f354, [%rd1+36]; ld.local.f32 %f355, [%rd1+12]; ld.local.f32 %f356, [%rd1+40]; ld.local.f32 %f357, [%rd1+16]; mul.ftz.f32 %f358, %f357, %f356; fma.rn.ftz.f32 %f359, %f355, %f354, %f358; ld.local.f32 %f360, [%rd1+44]; ld.local.f32 %f361, [%rd1+20]; fma.rn.ftz.f32 %f362, %f361, %f360, %f359; add.ftz.f32 %f363, %f362, %f362; mul.ftz.f32 %f364, %f355, %f363; mul.ftz.f32 %f365, %f357, %f363; mul.ftz.f32 %f366, %f361, %f363; sub.ftz.f32 %f1403, %f364, %f354; sub.ftz.f32 %f1402, %f365, %f356; sub.ftz.f32 %f1401, %f366, %f360; bra.uni $L__BB91_116; $L__BB91_101: ld.const.v4.f32 {%f1058, %f1059, %f1060, %f1061}, [kernel_params+688]; ld.const.v4.f32 {%f1062, %f1063, %f1064, %f1065}, [kernel_params+704]; ld.const.v4.f32 {%f1066, %f1067, %f1068, %f1069}, [kernel_params+720]; ld.local.u32 %r117, [%rd1+76]; setp.eq.s32 %p131, %r117, -1; ld.local.f32 %f1070, [%rd1+28]; ld.local.f32 %f1071, [%rd1+24]; setp.neu.ftz.f32 %p132, %f1071, %f1070; ld.local.f32 %f1072, [%rd1+32]; setp.neu.ftz.f32 %p133, %f1071, %f1072; or.pred %p134, %p132, %p133; sub.ftz.f32 %f1073, %f1072, %f1070; neg.ftz.f32 %f1074, %f1070; neg.ftz.f32 %f1075, %f1072; selp.f32 %f1076, %f1075, %f1072, %p134; add.ftz.f32 %f1077, %f1071, %f1076; selp.f32 %f1078, %f1070, %f1074, %p134; sub.ftz.f32 %f1079, %f1078, %f1071; mul.ftz.f32 %f1080, %f1077, %f1077; fma.rn.ftz.f32 %f1081, %f1073, %f1073, %f1080; fma.rn.ftz.f32 %f1082, %f1079, %f1079, %f1081; sqrt.approx.ftz.f32 %f1083, %f1082; div.approx.ftz.f32 %f1084, %f1073, %f1083; div.approx.ftz.f32 %f1085, %f1077, %f1083; div.approx.ftz.f32 %f1086, %f1079, %f1083; mul.ftz.f32 %f1087, %f1070, %f1086; mul.ftz.f32 %f1088, %f1072, %f1085; sub.ftz.f32 %f1089, %f1087, %f1088; mul.ftz.f32 %f1090, %f1072, %f1084; mul.ftz.f32 %f1091, %f1071, %f1086; sub.ftz.f32 %f1092, %f1090, %f1091; mul.ftz.f32 %f1093, %f1071, %f1085; mul.ftz.f32 %f1094, %f1070, %f1084; sub.ftz.f32 %f1095, %f1093, %f1094; ld.local.f32 %f1096, [%rd1+92]; ld.local.f32 %f1097, [%rd1]; fma.rn.ftz.f32 %f178, %f1096, %f1089, %f1097; ld.local.f32 %f1098, [%rd1+4]; fma.rn.ftz.f32 %f179, %f1096, %f1092, %f1098; ld.local.f32 %f1099, [%rd1+8]; fma.rn.ftz.f32 %f180, %f1096, %f1095, %f1099; @%p131 bra $L__BB91_103; mul.ftz.f32 %f1100, %f1059, %f179; fma.rn.ftz.f32 %f1101, %f1058, %f178, %f1100; fma.rn.ftz.f32 %f1102, %f1060, %f180, %f1101; add.ftz.f32 %f1403, %f1061, %f1102; mul.ftz.f32 %f1103, %f1063, %f179; fma.rn.ftz.f32 %f1104, %f1062, %f178, %f1103; fma.rn.ftz.f32 %f1105, %f1064, %f180, %f1104; add.ftz.f32 %f1402, %f1065, %f1105; mul.ftz.f32 %f1106, %f1067, %f179; fma.rn.ftz.f32 %f1107, %f1066, %f178, %f1106; fma.rn.ftz.f32 %f1108, %f1068, %f180, %f1107; add.ftz.f32 %f1401, %f1069, %f1108; bra.uni $L__BB91_116; $L__BB91_36: ld.local.u32 %r22, [%rd1+76]; setp.eq.s32 %p32, %r22, -1; mov.f32 %f1401, 0f00000000; mov.f32 %f1403, %f1401; mov.f32 %f1402, %f1401; @%p32 bra $L__BB91_116; ld.const.u64 %rd42, [kernel_params+2160]; cvta.to.global.u64 %rd43, %rd42; mul.wide.s32 %rd44, %r22, 208; add.s64 %rd45, %rd43, %rd44; ld.global.f32 %f1402, [%rd45+144]; ld.global.f32 %f1403, [%rd45+140]; bra.uni $L__BB91_116; $L__BB91_109: cvt.u32.u64 %r118, %rd20; setp.ne.s32 %p142, %r118, -1; @%p142 bra $L__BB91_111; ld.local.u32 %r119, [%rd1+64]; setp.ne.s32 %p143, %r119, 32; @%p143 bra $L__BB91_116; $L__BB91_111: ld.local.v4.f32 {%f1265, %f1266, %f1267, %f1268}, [%rd1+144]; ld.local.v4.f32 {%f1272, %f1273, %f1274, %f1275}, [%rd1+160]; ld.local.v4.f32 {%f1279, %f1280, %f1281, %f1282}, [%rd1+176]; mul.ftz.f32 %f1286, %f1402, %f1272; fma.rn.ftz.f32 %f1287, %f1403, %f1265, %f1286; fma.rn.ftz.f32 %f1288, %f1401, %f1279, %f1287; mul.ftz.f32 %f1289, %f1402, %f1273; fma.rn.ftz.f32 %f1290, %f1403, %f1266, %f1289; fma.rn.ftz.f32 %f1291, %f1401, %f1280, %f1290; mul.ftz.f32 %f1292, %f1402, %f1274; fma.rn.ftz.f32 %f1293, %f1403, %f1267, %f1292; fma.rn.ftz.f32 %f1294, %f1401, %f1281, %f1293; mul.ftz.f32 %f1295, %f1291, %f1291; fma.rn.ftz.f32 %f1296, %f1288, %f1288, %f1295; fma.rn.ftz.f32 %f1297, %f1294, %f1294, %f1296; sqrt.approx.ftz.f32 %f1298, %f1297; div.approx.ftz.f32 %f1403, %f1288, %f1298; div.approx.ftz.f32 %f1402, %f1291, %f1298; div.approx.ftz.f32 %f1401, %f1294, %f1298; bra.uni $L__BB91_116; $L__BB91_9: ld.local.u32 %r62, [%rd1+76]; setp.ne.s32 %p39, %r62, -1; ld.const.u32 %r131, [kernel_params+16]; setp.ne.s32 %p40, %r131, 1; or.pred %p41, %p39, %p40; @%p41 bra $L__BB91_42; ld.local.v4.f32 {%f370, %f371, %f372, %f373}, [%rd1+240]; ld.const.v4.f32 {%f377, %f378, %f379, %f380}, [kernel_params+624]; ld.const.v4.f32 {%f385, %f386, %f387, %f388}, [kernel_params+640]; ld.const.v4.f32 {%f401, %f402, %f403, %f404}, [kernel_params+672]; mul.ftz.f32 %f409, %f371, %f378; fma.rn.ftz.f32 %f410, %f370, %f377, %f409; fma.rn.ftz.f32 %f411, %f372, %f379, %f380; add.ftz.f32 %f40, %f410, %f411; mul.ftz.f32 %f412, %f371, %f386; fma.rn.ftz.f32 %f413, %f370, %f385, %f412; fma.rn.ftz.f32 %f414, %f372, %f387, %f388; add.ftz.f32 %f41, %f413, %f414; mul.ftz.f32 %f418, %f371, %f402; fma.rn.ftz.f32 %f419, %f370, %f401, %f418; fma.rn.ftz.f32 %f420, %f372, %f403, %f404; add.ftz.f32 %f43, %f419, %f420; setp.eq.ftz.f32 %p42, %f43, 0f00000000; mov.f32 %f1401, 0f00000000; mov.f32 %f1403, %f1401; mov.f32 %f1402, %f1401; @%p42 bra $L__BB91_116; div.approx.ftz.f32 %f1402, %f41, %f43; div.approx.ftz.f32 %f1403, %f40, %f43; bra.uni $L__BB91_116; $L__BB91_42: ld.local.f32 %f422, [%rd1+28]; ld.local.f32 %f423, [%rd1+24]; setp.neu.ftz.f32 %p43, %f423, %f422; ld.local.f32 %f424, [%rd1+32]; setp.neu.ftz.f32 %p44, %f423, %f424; or.pred %p45, %p43, %p44; sub.ftz.f32 %f425, %f424, %f422; neg.ftz.f32 %f426, %f422; neg.ftz.f32 %f427, %f424; selp.f32 %f428, %f427, %f424, %p45; add.ftz.f32 %f429, %f423, %f428; selp.f32 %f430, %f422, %f426, %p45; sub.ftz.f32 %f431, %f430, %f423; mul.ftz.f32 %f432, %f429, %f429; fma.rn.ftz.f32 %f433, %f425, %f425, %f432; fma.rn.ftz.f32 %f434, %f431, %f431, %f433; sqrt.approx.ftz.f32 %f435, %f434; div.approx.ftz.f32 %f436, %f425, %f435; div.approx.ftz.f32 %f437, %f429, %f435; div.approx.ftz.f32 %f438, %f431, %f435; mul.ftz.f32 %f439, %f422, %f438; mul.ftz.f32 %f440, %f424, %f437; sub.ftz.f32 %f441, %f439, %f440; mul.ftz.f32 %f442, %f424, %f436; mul.ftz.f32 %f443, %f423, %f438; sub.ftz.f32 %f444, %f442, %f443; mul.ftz.f32 %f445, %f423, %f437; mul.ftz.f32 %f446, %f422, %f436; sub.ftz.f32 %f447, %f445, %f446; ld.local.f32 %f448, [%rd1+92]; ld.local.f32 %f449, [%rd1]; fma.rn.ftz.f32 %f1382, %f448, %f441, %f449; ld.local.f32 %f450, [%rd1+4]; fma.rn.ftz.f32 %f1383, %f448, %f444, %f450; ld.local.f32 %f451, [%rd1+8]; fma.rn.ftz.f32 %f1384, %f448, %f447, %f451; setp.eq.s32 %p46, %r131, 2; @%p46 bra $L__BB91_47; bra.uni $L__BB91_43; $L__BB91_47: ld.const.v4.f32 {%f514, %f515, %f516, %f517}, [kernel_params+688]; ld.const.v4.f32 {%f518, %f519, %f520, %f521}, [kernel_params+704]; ld.const.v4.f32 {%f522, %f523, %f524, %f525}, [kernel_params+720]; ld.local.u32 %r64, [%rd1+76]; setp.eq.s32 %p51, %r64, -1; mul.ftz.f32 %f526, %f1383, %f515; fma.rn.ftz.f32 %f527, %f1382, %f514, %f526; fma.rn.ftz.f32 %f80, %f1384, %f516, %f527; mul.ftz.f32 %f528, %f1383, %f519; fma.rn.ftz.f32 %f529, %f1382, %f518, %f528; fma.rn.ftz.f32 %f81, %f1384, %f520, %f529; mul.ftz.f32 %f530, %f1383, %f523; fma.rn.ftz.f32 %f531, %f1382, %f522, %f530; fma.rn.ftz.f32 %f82, %f1384, %f524, %f531; @%p51 bra $L__BB91_49; add.ftz.f32 %f532, %f517, %f80; add.ftz.f32 %f533, %f521, %f81; add.ftz.f32 %f534, %f525, %f82; mul.ftz.f32 %f535, %f533, %f533; fma.rn.ftz.f32 %f536, %f532, %f532, %f535; fma.rn.ftz.f32 %f537, %f534, %f534, %f536; sqrt.approx.ftz.f32 %f538, %f537; div.approx.ftz.f32 %f89, %f532, %f538; div.approx.ftz.f32 %f90, %f533, %f538; div.approx.ftz.f32 %f91, %f534, %f538; bra.uni $L__BB91_50; $L__BB91_40: ld.local.f32 %f1403, [%rd1+36]; ld.local.f32 %f1402, [%rd1+40]; ld.local.f32 %f1401, [%rd1+44]; bra.uni $L__BB91_116; $L__BB91_103: ld.const.v4.f32 {%f1109, %f1110, %f1111, %f1112}, [kernel_params+96]; ld.const.v4.f32 {%f1114, %f1115, %f1116, %f1117}, [kernel_params+112]; ld.const.v4.f32 {%f1119, %f1120, %f1121, %f1122}, [kernel_params+128]; add.ftz.f32 %f1124, %f178, %f1112; add.ftz.f32 %f1125, %f179, %f1117; add.ftz.f32 %f1126, %f180, %f1122; mul.ftz.f32 %f1127, %f1059, %f1125; fma.rn.ftz.f32 %f1128, %f1058, %f1124, %f1127; fma.rn.ftz.f32 %f1129, %f1060, %f1126, %f1128; add.ftz.f32 %f1403, %f1061, %f1129; mul.ftz.f32 %f1130, %f1063, %f1125; fma.rn.ftz.f32 %f1131, %f1062, %f1124, %f1130; fma.rn.ftz.f32 %f1132, %f1064, %f1126, %f1131; add.ftz.f32 %f1402, %f1065, %f1132; mul.ftz.f32 %f1133, %f1067, %f1125; fma.rn.ftz.f32 %f1134, %f1066, %f1124, %f1133; fma.rn.ftz.f32 %f1135, %f1068, %f1126, %f1134; add.ftz.f32 %f1401, %f1069, %f1135; bra.uni $L__BB91_116; $L__BB91_104: ld.local.u32 %r36, [%rd1+76]; setp.eq.s32 %p139, %r36, -1; @%p139 bra $L__BB91_116; ld.local.u8 %rs10, [%rd1+56]; and.b16 %rs11, %rs10, 2; setp.eq.s16 %p140, %rs11, 0; @%p140 bra $L__BB91_107; bra.uni $L__BB91_106; $L__BB91_107: ld.const.u64 %rd54, [kernel_params+2160]; cvta.to.global.u64 %rd55, %rd54; mul.wide.s32 %rd56, %r36, 208; add.s64 %rd57, %rd55, %rd56; ld.global.v4.f32 {%f1232, %f1233, %f1234, %f1235}, [%rd57+48]; ld.global.v4.f32 {%f1240, %f1241, %f1242, %f1243}, [%rd57+64]; ld.global.v4.f32 {%f1248, %f1249, %f1250, %f1251}, [%rd57+80]; mul.ftz.f32 %f1256, %f1402, %f1233; fma.rn.ftz.f32 %f1257, %f1403, %f1232, %f1256; fma.rn.ftz.f32 %f1258, %f1401, %f1234, %f1257; add.ftz.f32 %f196, %f1235, %f1258; mul.ftz.f32 %f1259, %f1402, %f1241; fma.rn.ftz.f32 %f1260, %f1403, %f1240, %f1259; fma.rn.ftz.f32 %f1261, %f1401, %f1242, %f1260; add.ftz.f32 %f197, %f1243, %f1261; mul.ftz.f32 %f1262, %f1402, %f1249; fma.rn.ftz.f32 %f1263, %f1403, %f1248, %f1262; fma.rn.ftz.f32 %f1264, %f1401, %f1250, %f1263; add.ftz.f32 %f1401, %f1251, %f1264; mov.f32 %f1402, %f197; mov.f32 %f1403, %f196; bra.uni $L__BB91_116; $L__BB91_43: ld.local.u32 %r63, [%rd1+76]; setp.ne.s32 %p47, %r63, -1; setp.ne.s32 %p48, %r131, 0; or.pred %p49, %p47, %p48; @%p49 bra $L__BB91_45; ld.const.v4.f32 {%f452, %f453, %f454, %f455}, [kernel_params+96]; ld.const.v4.f32 {%f457, %f458, %f459, %f460}, [kernel_params+112]; ld.const.v4.f32 {%f462, %f463, %f464, %f465}, [kernel_params+128]; add.ftz.f32 %f1382, %f1382, %f455; add.ftz.f32 %f1383, %f1383, %f460; add.ftz.f32 %f1384, %f1384, %f465; $L__BB91_45: ld.const.v4.f32 {%f470, %f471, %f472, %f473}, [kernel_params+624]; ld.const.v4.f32 {%f478, %f479, %f480, %f481}, [kernel_params+640]; ld.const.v4.f32 {%f494, %f495, %f496, %f497}, [kernel_params+672]; mul.ftz.f32 %f502, %f1383, %f471; fma.rn.ftz.f32 %f503, %f1382, %f470, %f502; fma.rn.ftz.f32 %f504, %f1384, %f472, %f473; add.ftz.f32 %f61, %f503, %f504; mul.ftz.f32 %f505, %f1383, %f479; fma.rn.ftz.f32 %f506, %f1382, %f478, %f505; fma.rn.ftz.f32 %f507, %f1384, %f480, %f481; add.ftz.f32 %f62, %f506, %f507; mul.ftz.f32 %f511, %f1383, %f495; fma.rn.ftz.f32 %f512, %f1382, %f494, %f511; fma.rn.ftz.f32 %f513, %f1384, %f496, %f497; add.ftz.f32 %f64, %f512, %f513; setp.eq.ftz.f32 %p50, %f64, 0f00000000; mov.f32 %f1403, 0f00000000; mov.f32 %f1402, %f1403; @%p50 bra $L__BB91_100; div.approx.ftz.f32 %f1402, %f62, %f64; div.approx.ftz.f32 %f1403, %f61, %f64; bra.uni $L__BB91_100; $L__BB91_32: ld.local.v4.f32 {%f245, %f246, %f247, %f248}, [%rd1+192]; mul.ftz.f32 %f253, %f1402, %f246; fma.rn.ftz.f32 %f254, %f1403, %f245, %f253; fma.rn.ftz.f32 %f255, %f1401, %f247, %f254; add.ftz.f32 %f4, %f248, %f255; ld.local.v4.f32 {%f256, %f257, %f258, %f259}, [%rd1+208]; mul.ftz.f32 %f264, %f1402, %f257; fma.rn.ftz.f32 %f265, %f1403, %f256, %f264; fma.rn.ftz.f32 %f266, %f1401, %f258, %f265; add.ftz.f32 %f5, %f259, %f266; ld.local.v4.f32 {%f267, %f268, %f269, %f270}, [%rd1+224]; mul.ftz.f32 %f275, %f1402, %f268; fma.rn.ftz.f32 %f276, %f1403, %f267, %f275; fma.rn.ftz.f32 %f277, %f1401, %f269, %f276; add.ftz.f32 %f1401, %f270, %f277; mov.f32 %f1403, %f4; mov.f32 %f1402, %f5; $L__BB91_34: setp.eq.s32 %p31, %r130, -1; @%p31 bra $L__BB91_116; ld.const.u64 %rd38, [kernel_params+2344]; cvta.to.global.u64 %rd39, %rd38; mul.wide.s32 %rd40, %r130, 16; add.s64 %rd41, %rd39, %rd40; ld.global.v4.f32 {%f311, %f312, %f313, %f314}, [%rd41]; ld.global.v4.f32 {%f319, %f320, %f321, %f322}, [%rd41+16]; ld.global.v4.f32 {%f327, %f328, %f329, %f330}, [%rd41+32]; mul.ftz.f32 %f335, %f1402, %f312; fma.rn.ftz.f32 %f336, %f1403, %f311, %f335; fma.rn.ftz.f32 %f337, %f1401, %f313, %f336; add.ftz.f32 %f13, %f314, %f337; mul.ftz.f32 %f338, %f1402, %f320; fma.rn.ftz.f32 %f339, %f1403, %f319, %f338; fma.rn.ftz.f32 %f340, %f1401, %f321, %f339; add.ftz.f32 %f14, %f322, %f340; mul.ftz.f32 %f341, %f1402, %f328; fma.rn.ftz.f32 %f342, %f1403, %f327, %f341; fma.rn.ftz.f32 %f343, %f1401, %f329, %f342; add.ftz.f32 %f1401, %f330, %f343; mov.f32 %f1402, %f14; mov.f32 %f1403, %f13; bra.uni $L__BB91_116; $L__BB91_113: ld.local.u32 %r121, [%rd1+64]; setp.ne.s32 %p145, %r121, 32; @%p145 bra $L__BB91_116; ld.const.u64 %rd58, [kernel_params+2368]; cvta.to.global.u64 %rd59, %rd58; ld.local.u32 %r122, [%rd1+80]; mul.wide.s32 %rd60, %r122, 192; add.s64 %rd61, %rd59, %rd60; ld.global.v4.f32 {%f1299, %f1300, %f1301, %f1302}, [%rd61+48]; ld.global.v4.f32 {%f1306, %f1307, %f1308, %f1309}, [%rd61+64]; ld.global.v4.f32 {%f1313, %f1314, %f1315, %f1316}, [%rd61+80]; mul.ftz.f32 %f1320, %f1402, %f1306; fma.rn.ftz.f32 %f1321, %f1403, %f1299, %f1320; fma.rn.ftz.f32 %f1322, %f1401, %f1313, %f1321; mul.ftz.f32 %f1323, %f1402, %f1307; fma.rn.ftz.f32 %f1324, %f1403, %f1300, %f1323; fma.rn.ftz.f32 %f1325, %f1401, %f1314, %f1324; mul.ftz.f32 %f1326, %f1402, %f1308; fma.rn.ftz.f32 %f1327, %f1403, %f1301, %f1326; fma.rn.ftz.f32 %f1328, %f1401, %f1315, %f1327; mul.ftz.f32 %f1329, %f1325, %f1325; fma.rn.ftz.f32 %f1330, %f1322, %f1322, %f1329; fma.rn.ftz.f32 %f1331, %f1328, %f1328, %f1330; sqrt.approx.ftz.f32 %f1332, %f1331; div.approx.ftz.f32 %f1403, %f1322, %f1332; div.approx.ftz.f32 %f1402, %f1325, %f1332; div.approx.ftz.f32 %f1401, %f1328, %f1332; bra.uni $L__BB91_116; $L__BB91_49: mul.ftz.f32 %f539, %f81, %f81; fma.rn.ftz.f32 %f540, %f80, %f80, %f539; fma.rn.ftz.f32 %f541, %f82, %f82, %f540; sqrt.approx.ftz.f32 %f542, %f541; div.approx.ftz.f32 %f89, %f80, %f542; div.approx.ftz.f32 %f90, %f81, %f542; div.approx.ftz.f32 %f91, %f82, %f542; $L__BB91_50: ld.const.u32 %r65, [kernel_params+20]; mov.f32 %f543, 0fBF800000; max.ftz.f32 %f544, %f89, %f543; mov.f32 %f545, 0f3F800000; min.ftz.f32 %f546, %f544, %f545; abs.ftz.f32 %f547, %f546; sub.ftz.f32 %f548, %f545, %f547; mul.ftz.f32 %f549, %f548, 0f3F000000; sqrt.approx.ftz.f32 %f550, %f549; setp.gt.ftz.f32 %p52, %f547, 0f3F11EB85; selp.f32 %f551, %f550, %f547, %p52; mul.ftz.f32 %f552, %f551, %f551; mov.f32 %f553, 0f3C94D2E9; mov.f32 %f554, 0f3D53F941; fma.rn.ftz.f32 %f555, %f554, %f552, %f553; mov.f32 %f556, 0f3D3F841F; fma.rn.ftz.f32 %f557, %f555, %f552, %f556; mov.f32 %f558, 0f3D994929; fma.rn.ftz.f32 %f559, %f557, %f552, %f558; mov.f32 %f560, 0f3E2AAB94; fma.rn.ftz.f32 %f561, %f559, %f552, %f560; mul.ftz.f32 %f562, %f552, %f561; fma.rn.ftz.f32 %f563, %f562, %f551, %f551; add.ftz.f32 %f564, %f563, %f563; mov.f32 %f565, 0f3FC90FDB; sub.ftz.f32 %f566, %f565, %f563; selp.f32 %f567, %f564, %f566, %p52; setp.lt.ftz.f32 %p53, %f546, 0f00000000; mov.f32 %f568, 0f40490FDB; sub.ftz.f32 %f569, %f568, %f567; selp.f32 %f92, %f569, %f567, %p53; abs.ftz.f32 %f93, %f90; setp.eq.ftz.f32 %p54, %f93, 0f00000000; abs.ftz.f32 %f94, %f91; setp.eq.ftz.f32 %p55, %f94, 0f00000000; and.pred %p1, %p54, %p55; mov.b32 %r27, %f90; mov.b32 %r66, %f91; and.b32 %r28, %r66, -2147483648; setp.gt.s32 %p56, %r65, 2; @%p56 bra $L__BB91_56; bra.uni $L__BB91_51; $L__BB91_56: setp.eq.s32 %p57, %r65, 3; @%p57 bra $L__BB91_65; setp.eq.s32 %p58, %r65, 5; @%p58 bra $L__BB91_68; setp.ne.s32 %p59, %r65, 4; @%p59 bra $L__BB91_94; ld.const.f32 %f570, [kernel_params+48]; ld.const.v4.f32 {%f571, %f572, %f573, %f574}, [kernel_params+64]; neg.ftz.f32 %f579, %f92; sub.ftz.f32 %f580, %f579, %f570; div.approx.ftz.f32 %f581, %f580, %f571; mul.ftz.f32 %f582, %f581, %f581; mul.ftz.f32 %f583, %f581, %f582; mul.ftz.f32 %f584, %f582, %f582; mul.ftz.f32 %f585, %f572, %f582; fma.rn.ftz.f32 %f586, %f571, %f581, %f585; mul.ftz.f32 %f587, %f574, %f584; fma.rn.ftz.f32 %f588, %f573, %f583, %f587; add.ftz.f32 %f589, %f586, %f588; add.ftz.f32 %f590, %f570, %f589; sub.ftz.f32 %f591, %f579, %f590; div.approx.ftz.f32 %f592, %f591, %f571; mul.ftz.f32 %f593, %f592, %f592; mul.ftz.f32 %f594, %f592, %f593; mul.ftz.f32 %f595, %f593, %f593; mul.ftz.f32 %f596, %f572, %f593; fma.rn.ftz.f32 %f597, %f571, %f592, %f596; mul.ftz.f32 %f598, %f574, %f595; fma.rn.ftz.f32 %f599, %f573, %f594, %f598; add.ftz.f32 %f600, %f597, %f599; add.ftz.f32 %f601, %f570, %f600; sub.ftz.f32 %f602, %f579, %f601; div.approx.ftz.f32 %f603, %f602, %f571; mul.ftz.f32 %f604, %f603, %f603; mul.ftz.f32 %f605, %f603, %f604; mul.ftz.f32 %f606, %f604, %f604; mul.ftz.f32 %f607, %f572, %f604; fma.rn.ftz.f32 %f608, %f571, %f603, %f607; mul.ftz.f32 %f609, %f574, %f606; fma.rn.ftz.f32 %f610, %f573, %f605, %f609; add.ftz.f32 %f611, %f608, %f610; add.ftz.f32 %f612, %f570, %f611; sub.ftz.f32 %f613, %f579, %f612; div.approx.ftz.f32 %f614, %f613, %f571; mul.ftz.f32 %f615, %f614, %f614; mul.ftz.f32 %f616, %f614, %f615; mul.ftz.f32 %f617, %f615, %f615; mul.ftz.f32 %f618, %f572, %f615; fma.rn.ftz.f32 %f619, %f571, %f614, %f618; mul.ftz.f32 %f620, %f574, %f617; fma.rn.ftz.f32 %f621, %f573, %f616, %f620; add.ftz.f32 %f622, %f619, %f621; add.ftz.f32 %f623, %f570, %f622; sub.ftz.f32 %f624, %f579, %f623; div.approx.ftz.f32 %f625, %f624, %f571; mul.ftz.f32 %f626, %f625, %f625; mul.ftz.f32 %f627, %f625, %f626; mul.ftz.f32 %f628, %f626, %f626; mul.ftz.f32 %f629, %f572, %f626; fma.rn.ftz.f32 %f630, %f571, %f625, %f629; mul.ftz.f32 %f631, %f574, %f628; fma.rn.ftz.f32 %f632, %f573, %f627, %f631; add.ftz.f32 %f633, %f630, %f632; add.ftz.f32 %f634, %f570, %f633; sub.ftz.f32 %f635, %f579, %f634; div.approx.ftz.f32 %f636, %f635, %f571; mul.ftz.f32 %f637, %f636, %f636; mul.ftz.f32 %f638, %f636, %f637; mul.ftz.f32 %f639, %f637, %f637; mul.ftz.f32 %f640, %f572, %f637; fma.rn.ftz.f32 %f641, %f571, %f636, %f640; mul.ftz.f32 %f642, %f574, %f639; fma.rn.ftz.f32 %f643, %f573, %f638, %f642; add.ftz.f32 %f644, %f641, %f643; add.ftz.f32 %f645, %f570, %f644; sub.ftz.f32 %f646, %f579, %f645; div.approx.ftz.f32 %f647, %f646, %f571; mul.ftz.f32 %f648, %f647, %f647; mul.ftz.f32 %f649, %f647, %f648; mul.ftz.f32 %f650, %f648, %f648; mul.ftz.f32 %f651, %f572, %f648; fma.rn.ftz.f32 %f652, %f571, %f647, %f651; mul.ftz.f32 %f653, %f574, %f650; fma.rn.ftz.f32 %f654, %f573, %f649, %f653; add.ftz.f32 %f655, %f652, %f654; add.ftz.f32 %f656, %f570, %f655; sub.ftz.f32 %f657, %f579, %f656; div.approx.ftz.f32 %f658, %f657, %f571; mul.ftz.f32 %f659, %f658, %f658; mul.ftz.f32 %f660, %f658, %f659; mul.ftz.f32 %f661, %f659, %f659; mul.ftz.f32 %f662, %f572, %f659; fma.rn.ftz.f32 %f663, %f571, %f658, %f662; mul.ftz.f32 %f664, %f574, %f661; fma.rn.ftz.f32 %f665, %f573, %f660, %f664; add.ftz.f32 %f666, %f663, %f665; add.ftz.f32 %f667, %f570, %f666; sub.ftz.f32 %f668, %f579, %f667; div.approx.ftz.f32 %f669, %f668, %f571; mul.ftz.f32 %f670, %f669, %f669; mul.ftz.f32 %f671, %f669, %f670; mul.ftz.f32 %f672, %f670, %f670; mul.ftz.f32 %f673, %f572, %f670; fma.rn.ftz.f32 %f674, %f571, %f669, %f673; mul.ftz.f32 %f675, %f574, %f672; fma.rn.ftz.f32 %f676, %f573, %f671, %f675; add.ftz.f32 %f677, %f674, %f676; add.ftz.f32 %f678, %f570, %f677; sub.ftz.f32 %f679, %f579, %f678; div.approx.ftz.f32 %f680, %f679, %f571; mul.ftz.f32 %f681, %f680, %f680; mul.ftz.f32 %f682, %f680, %f681; mul.ftz.f32 %f683, %f681, %f681; mul.ftz.f32 %f684, %f572, %f681; fma.rn.ftz.f32 %f685, %f571, %f680, %f684; mul.ftz.f32 %f686, %f574, %f683; fma.rn.ftz.f32 %f687, %f573, %f682, %f686; add.ftz.f32 %f688, %f685, %f687; add.ftz.f32 %f689, %f570, %f688; sub.ftz.f32 %f690, %f579, %f689; div.approx.ftz.f32 %f691, %f690, %f571; mul.ftz.f32 %f692, %f691, %f691; mul.ftz.f32 %f693, %f691, %f692; mul.ftz.f32 %f694, %f692, %f692; mul.ftz.f32 %f695, %f572, %f692; fma.rn.ftz.f32 %f696, %f571, %f691, %f695; mul.ftz.f32 %f697, %f574, %f694; fma.rn.ftz.f32 %f698, %f573, %f693, %f697; add.ftz.f32 %f699, %f696, %f698; add.ftz.f32 %f700, %f570, %f699; sub.ftz.f32 %f701, %f579, %f700; div.approx.ftz.f32 %f702, %f701, %f571; mul.ftz.f32 %f703, %f702, %f702; mul.ftz.f32 %f704, %f702, %f703; mul.ftz.f32 %f705, %f703, %f703; mul.ftz.f32 %f706, %f572, %f703; fma.rn.ftz.f32 %f707, %f571, %f702, %f706; mul.ftz.f32 %f708, %f574, %f705; fma.rn.ftz.f32 %f709, %f573, %f704, %f708; add.ftz.f32 %f710, %f707, %f709; add.ftz.f32 %f711, %f570, %f710; sub.ftz.f32 %f712, %f579, %f711; div.approx.ftz.f32 %f713, %f712, %f571; mul.ftz.f32 %f714, %f713, %f713; mul.ftz.f32 %f715, %f713, %f714; mul.ftz.f32 %f716, %f714, %f714; mul.ftz.f32 %f717, %f572, %f714; fma.rn.ftz.f32 %f718, %f571, %f713, %f717; mul.ftz.f32 %f719, %f574, %f716; fma.rn.ftz.f32 %f720, %f573, %f715, %f719; add.ftz.f32 %f721, %f718, %f720; add.ftz.f32 %f722, %f570, %f721; sub.ftz.f32 %f723, %f579, %f722; div.approx.ftz.f32 %f724, %f723, %f571; mul.ftz.f32 %f725, %f724, %f724; mul.ftz.f32 %f726, %f724, %f725; mul.ftz.f32 %f727, %f725, %f725; mul.ftz.f32 %f728, %f572, %f725; fma.rn.ftz.f32 %f729, %f571, %f724, %f728; mul.ftz.f32 %f730, %f574, %f727; fma.rn.ftz.f32 %f731, %f573, %f726, %f730; add.ftz.f32 %f732, %f729, %f731; add.ftz.f32 %f733, %f570, %f732; sub.ftz.f32 %f734, %f579, %f733; div.approx.ftz.f32 %f735, %f734, %f571; mul.ftz.f32 %f736, %f735, %f735; mul.ftz.f32 %f737, %f735, %f736; mul.ftz.f32 %f738, %f736, %f736; mul.ftz.f32 %f739, %f572, %f736; fma.rn.ftz.f32 %f740, %f571, %f735, %f739; mul.ftz.f32 %f741, %f574, %f738; fma.rn.ftz.f32 %f742, %f573, %f737, %f741; add.ftz.f32 %f743, %f740, %f742; add.ftz.f32 %f744, %f570, %f743; sub.ftz.f32 %f745, %f579, %f744; div.approx.ftz.f32 %f746, %f745, %f571; mul.ftz.f32 %f747, %f746, %f746; mul.ftz.f32 %f748, %f746, %f747; mul.ftz.f32 %f749, %f747, %f747; mul.ftz.f32 %f750, %f572, %f747; fma.rn.ftz.f32 %f751, %f571, %f746, %f750; mul.ftz.f32 %f752, %f574, %f749; fma.rn.ftz.f32 %f753, %f573, %f748, %f752; add.ftz.f32 %f754, %f751, %f753; add.ftz.f32 %f755, %f570, %f754; sub.ftz.f32 %f756, %f579, %f755; div.approx.ftz.f32 %f757, %f756, %f571; mul.ftz.f32 %f758, %f757, %f757; mul.ftz.f32 %f759, %f757, %f758; mul.ftz.f32 %f760, %f758, %f758; mul.ftz.f32 %f761, %f572, %f758; fma.rn.ftz.f32 %f762, %f571, %f757, %f761; mul.ftz.f32 %f763, %f574, %f760; fma.rn.ftz.f32 %f764, %f573, %f759, %f763; add.ftz.f32 %f765, %f762, %f764; add.ftz.f32 %f766, %f570, %f765; sub.ftz.f32 %f767, %f579, %f766; div.approx.ftz.f32 %f768, %f767, %f571; mul.ftz.f32 %f769, %f768, %f768; mul.ftz.f32 %f770, %f768, %f769; mul.ftz.f32 %f771, %f769, %f769; mul.ftz.f32 %f772, %f572, %f769; fma.rn.ftz.f32 %f773, %f571, %f768, %f772; mul.ftz.f32 %f774, %f574, %f771; fma.rn.ftz.f32 %f775, %f573, %f770, %f774; add.ftz.f32 %f776, %f773, %f775; add.ftz.f32 %f777, %f570, %f776; sub.ftz.f32 %f778, %f579, %f777; div.approx.ftz.f32 %f779, %f778, %f571; mul.ftz.f32 %f780, %f779, %f779; mul.ftz.f32 %f781, %f779, %f780; mul.ftz.f32 %f782, %f780, %f780; mul.ftz.f32 %f783, %f572, %f780; fma.rn.ftz.f32 %f784, %f571, %f779, %f783; mul.ftz.f32 %f785, %f574, %f782; fma.rn.ftz.f32 %f786, %f573, %f781, %f785; add.ftz.f32 %f787, %f784, %f786; add.ftz.f32 %f788, %f570, %f787; sub.ftz.f32 %f789, %f579, %f788; div.approx.ftz.f32 %f790, %f789, %f571; mul.ftz.f32 %f791, %f790, %f790; mul.ftz.f32 %f792, %f790, %f791; mul.ftz.f32 %f793, %f791, %f791; mul.ftz.f32 %f794, %f572, %f791; fma.rn.ftz.f32 %f795, %f571, %f790, %f794; mul.ftz.f32 %f796, %f574, %f793; fma.rn.ftz.f32 %f797, %f573, %f792, %f796; add.ftz.f32 %f798, %f795, %f797; add.ftz.f32 %f799, %f570, %f798; sub.ftz.f32 %f800, %f579, %f799; div.approx.ftz.f32 %f97, %f800, %f571; @%p1 bra $L__BB91_63; bra.uni $L__BB91_60; $L__BB91_63: shr.s32 %r71, %r27, 31; and.b32 %r72, %r71, 1078530011; or.b32 %r73, %r72, %r28; mov.b32 %f1388, %r73; bra.uni $L__BB91_64; $L__BB91_51: setp.eq.s32 %p60, %r65, 0; @%p60 bra $L__BB91_87; setp.eq.s32 %p61, %r65, 1; @%p61 bra $L__BB91_53; bra.uni $L__BB91_94; $L__BB91_53: ld.const.f32 %f122, [kernel_params+24]; mul.ftz.f32 %f900, %f91, %f91; fma.rn.ftz.f32 %f901, %f90, %f90, %f900; sqrt.approx.ftz.f32 %f902, %f901; abs.ftz.f32 %f123, %f902; abs.ftz.f32 %f124, %f89; setp.eq.ftz.f32 %p90, %f124, 0f00000000; setp.eq.ftz.f32 %p91, %f123, 0f00000000; and.pred %p92, %p90, %p91; mov.b32 %r31, %f89; mov.b32 %r88, %f902; and.b32 %r32, %r88, -2147483648; @%p92 bra $L__BB91_80; bra.uni $L__BB91_54; $L__BB91_80: shr.s32 %r93, %r31, 31; and.b32 %r94, %r93, 1078530011; or.b32 %r95, %r32, %r94; mov.b32 %f1393, %r95; bra.uni $L__BB91_81; $L__BB91_106: ld.local.v4.f32 {%f1199, %f1200, %f1201, %f1202}, [%rd1+192]; mul.ftz.f32 %f1207, %f1402, %f1200; fma.rn.ftz.f32 %f1208, %f1403, %f1199, %f1207; fma.rn.ftz.f32 %f1209, %f1401, %f1201, %f1208; add.ftz.f32 %f193, %f1202, %f1209; ld.local.v4.f32 {%f1210, %f1211, %f1212, %f1213}, [%rd1+208]; mul.ftz.f32 %f1218, %f1402, %f1211; fma.rn.ftz.f32 %f1219, %f1403, %f1210, %f1218; fma.rn.ftz.f32 %f1220, %f1401, %f1212, %f1219; add.ftz.f32 %f194, %f1213, %f1220; ld.local.v4.f32 {%f1221, %f1222, %f1223, %f1224}, [%rd1+224]; mul.ftz.f32 %f1229, %f1402, %f1222; fma.rn.ftz.f32 %f1230, %f1403, %f1221, %f1229; fma.rn.ftz.f32 %f1231, %f1401, %f1223, %f1230; add.ftz.f32 %f1401, %f1224, %f1231; mov.f32 %f1402, %f194; mov.f32 %f1403, %f193; bra.uni $L__BB91_116; $L__BB91_65: add.ftz.f32 %f833, %f90, 0fBF800000; mul.ftz.f32 %f834, %f833, 0fBF000000; mov.f32 %f835, 0f00000000; max.ftz.f32 %f836, %f834, %f835; sqrt.approx.ftz.f32 %f837, %f836; add.ftz.f32 %f104, %f837, %f837; setp.leu.ftz.f32 %p69, %f104, 0f00000000; @%p69 bra $L__BB91_67; rcp.approx.ftz.f32 %f838, %f104; mul.ftz.f32 %f89, %f89, %f838; mul.ftz.f32 %f91, %f91, %f838; $L__BB91_67: add.ftz.f32 %f839, %f89, 0f3F800000; add.ftz.f32 %f840, %f91, 0f3F800000; mul.ftz.f32 %f1402, %f840, 0f3F000000; mul.ftz.f32 %f1403, %f839, 0f3F000000; bra.uni $L__BB91_100; $L__BB91_68: abs.ftz.f32 %f111, %f89; setp.eq.ftz.f32 %p71, %f111, 0f00000000; and.pred %p72, %p71, %p54; mov.b32 %r29, %f89; and.b32 %r30, %r27, -2147483648; @%p72 bra $L__BB91_72; bra.uni $L__BB91_69; $L__BB91_72: shr.s32 %r78, %r29, 31; and.b32 %r79, %r78, 1078530011; or.b32 %r80, %r79, %r30; mov.b32 %f1391, %r80; bra.uni $L__BB91_73; $L__BB91_87: setp.eq.ftz.f32 %p107, %f90, 0f00000000; setp.eq.ftz.f32 %p108, %f89, 0f00000000; and.pred %p109, %p108, %p107; setp.eq.ftz.f32 %p110, %f91, 0f00000000; and.pred %p111, %p110, %p109; mov.f32 %f1403, 0f00000000; mov.f32 %f1402, %f1403; @%p111 bra $L__BB91_100; ld.const.v4.f32 {%f959, %f960, %f961, %f962}, [kernel_params+32]; abs.ftz.f32 %f140, %f89; setp.eq.ftz.f32 %p113, %f140, 0f00000000; and.pred %p114, %p113, %p54; mov.b32 %r33, %f89; and.b32 %r34, %r27, -2147483648; @%p114 bra $L__BB91_92; bra.uni $L__BB91_89; $L__BB91_92: shr.s32 %r107, %r33, 31; and.b32 %r108, %r107, 1078530011; or.b32 %r109, %r108, %r34; mov.b32 %f1395, %r109; bra.uni $L__BB91_93; $L__BB91_94: ld.const.f32 %f1021, [kernel_params+28]; ld.const.f32 %f147, [kernel_params+276]; ld.const.f32 %f148, [kernel_params+280]; add.ftz.f32 %f1022, %f1021, %f1021; mul.ftz.f32 %f1023, %f92, 0f3F000000; sin.approx.ftz.f32 %f1024, %f1023; mul.ftz.f32 %f149, %f1022, %f1024; @%p1 bra $L__BB91_98; bra.uni $L__BB91_95; $L__BB91_98: shr.s32 %r114, %r27, 31; and.b32 %r115, %r114, 1078530011; or.b32 %r116, %r115, %r28; mov.b32 %f1396, %r116; bra.uni $L__BB91_99; $L__BB91_69: setp.eq.ftz.f32 %p73, %f111, 0f7F800000; setp.eq.ftz.f32 %p74, %f93, 0f7F800000; and.pred %p75, %p73, %p74; @%p75 bra $L__BB91_71; bra.uni $L__BB91_70; $L__BB91_71: setp.lt.s32 %p79, %r29, 0; selp.b32 %r76, 1075235812, 1061752795, %p79; or.b32 %r77, %r76, %r30; mov.b32 %f1391, %r77; bra.uni $L__BB91_73; $L__BB91_60: setp.eq.ftz.f32 %p62, %f93, 0f7F800000; setp.eq.ftz.f32 %p63, %f94, 0f7F800000; and.pred %p64, %p62, %p63; @%p64 bra $L__BB91_62; bra.uni $L__BB91_61; $L__BB91_62: setp.lt.s32 %p68, %r27, 0; selp.b32 %r69, 1075235812, 1061752795, %p68; or.b32 %r70, %r69, %r28; mov.b32 %f1388, %r70; bra.uni $L__BB91_64; $L__BB91_54: setp.eq.ftz.f32 %p93, %f124, 0f7F800000; setp.eq.ftz.f32 %p94, %f123, 0f7F800000; and.pred %p95, %p93, %p94; @%p95 bra $L__BB91_79; bra.uni $L__BB91_55; $L__BB91_79: setp.lt.s32 %p99, %r31, 0; selp.b32 %r91, 1075235812, 1061752795, %p99; or.b32 %r92, %r32, %r91; mov.b32 %f1393, %r92; bra.uni $L__BB91_81; $L__BB91_95: setp.eq.ftz.f32 %p124, %f93, 0f7F800000; setp.eq.ftz.f32 %p125, %f94, 0f7F800000; and.pred %p126, %p124, %p125; @%p126 bra $L__BB91_97; bra.uni $L__BB91_96; $L__BB91_97: setp.lt.s32 %p130, %r27, 0; selp.b32 %r112, 1075235812, 1061752795, %p130; or.b32 %r113, %r112, %r28; mov.b32 %f1396, %r113; bra.uni $L__BB91_99; $L__BB91_70: setp.lt.s32 %p76, %r29, 0; min.ftz.f32 %f841, %f93, %f111; max.ftz.f32 %f842, %f93, %f111; div.full.ftz.f32 %f843, %f841, %f842; mul.rn.ftz.f32 %f844, %f843, %f843; mov.f32 %f845, 0fC0B59883; mov.f32 %f846, 0fBF52C7EA; fma.rn.ftz.f32 %f847, %f844, %f846, %f845; mov.f32 %f848, 0fC0D21907; fma.rn.ftz.f32 %f849, %f847, %f844, %f848; mul.ftz.f32 %f850, %f844, %f849; mul.ftz.f32 %f851, %f843, %f850; add.ftz.f32 %f852, %f844, 0f41355DC0; mov.f32 %f853, 0f41E6BD60; fma.rn.ftz.f32 %f854, %f852, %f844, %f853; mov.f32 %f855, 0f419D92C8; fma.rn.ftz.f32 %f856, %f854, %f844, %f855; rcp.approx.ftz.f32 %f857, %f856; fma.rn.ftz.f32 %f858, %f851, %f857, %f843; sub.ftz.f32 %f860, %f565, %f858; setp.gt.ftz.f32 %p77, %f93, %f111; selp.f32 %f861, %f860, %f858, %p77; sub.ftz.f32 %f863, %f568, %f861; selp.f32 %f864, %f863, %f861, %p76; mov.b32 %r74, %f864; or.b32 %r75, %r30, %r74; mov.b32 %f865, %r75; add.ftz.f32 %f866, %f111, %f93; setp.le.ftz.f32 %p78, %f866, 0f7F800000; selp.f32 %f1391, %f865, %f866, %p78; $L__BB91_73: and.pred %p82, %p71, %p55; @%p82 bra $L__BB91_77; bra.uni $L__BB91_74; $L__BB91_77: shr.s32 %r85, %r29, 31; and.b32 %r86, %r85, 1078530011; or.b32 %r87, %r28, %r86; mov.b32 %f1392, %r87; bra.uni $L__BB91_78; $L__BB91_74: setp.eq.ftz.f32 %p83, %f111, 0f7F800000; setp.eq.ftz.f32 %p84, %f94, 0f7F800000; and.pred %p85, %p83, %p84; @%p85 bra $L__BB91_76; bra.uni $L__BB91_75; $L__BB91_76: setp.lt.s32 %p89, %r29, 0; selp.b32 %r83, 1075235812, 1061752795, %p89; or.b32 %r84, %r28, %r83; mov.b32 %f1392, %r84; bra.uni $L__BB91_78; $L__BB91_75: setp.lt.s32 %p86, %r29, 0; min.ftz.f32 %f867, %f94, %f111; max.ftz.f32 %f868, %f94, %f111; div.full.ftz.f32 %f869, %f867, %f868; mul.rn.ftz.f32 %f870, %f869, %f869; mov.f32 %f871, 0fC0B59883; mov.f32 %f872, 0fBF52C7EA; fma.rn.ftz.f32 %f873, %f870, %f872, %f871; mov.f32 %f874, 0fC0D21907; fma.rn.ftz.f32 %f875, %f873, %f870, %f874; mul.ftz.f32 %f876, %f870, %f875; mul.ftz.f32 %f877, %f869, %f876; add.ftz.f32 %f878, %f870, 0f41355DC0; mov.f32 %f879, 0f41E6BD60; fma.rn.ftz.f32 %f880, %f878, %f870, %f879; mov.f32 %f881, 0f419D92C8; fma.rn.ftz.f32 %f882, %f880, %f870, %f881; rcp.approx.ftz.f32 %f883, %f882; fma.rn.ftz.f32 %f884, %f877, %f883, %f869; sub.ftz.f32 %f886, %f565, %f884; setp.gt.ftz.f32 %p87, %f94, %f111; selp.f32 %f887, %f886, %f884, %p87; sub.ftz.f32 %f889, %f568, %f887; selp.f32 %f890, %f889, %f887, %p86; mov.b32 %r81, %f890; or.b32 %r82, %r28, %r81; mov.b32 %f891, %r82; add.ftz.f32 %f892, %f111, %f94; setp.le.ftz.f32 %p88, %f892, 0f7F800000; selp.f32 %f1392, %f891, %f892, %p88; $L__BB91_78: add.ftz.f32 %f893, %f1392, %f1392; div.approx.ftz.f32 %f895, %f893, %f568; add.ftz.f32 %f896, %f1391, %f1391; div.approx.ftz.f32 %f897, %f896, %f568; add.ftz.f32 %f898, %f897, 0f3F000000; sub.ftz.f32 %f1403, %f545, %f898; add.ftz.f32 %f1402, %f895, 0f3F000000; bra.uni $L__BB91_100; $L__BB91_89: setp.eq.ftz.f32 %p115, %f140, 0f7F800000; setp.eq.ftz.f32 %p116, %f93, 0f7F800000; and.pred %p117, %p115, %p116; @%p117 bra $L__BB91_91; bra.uni $L__BB91_90; $L__BB91_91: setp.lt.s32 %p121, %r33, 0; selp.b32 %r105, 1075235812, 1061752795, %p121; or.b32 %r106, %r105, %r34; mov.b32 %f1395, %r106; bra.uni $L__BB91_93; $L__BB91_61: mov.f32 %f1368, 0f40490FDB; mov.f32 %f1367, 0f3FC90FDB; setp.lt.s32 %p65, %r27, 0; min.ftz.f32 %f801, %f94, %f93; max.ftz.f32 %f802, %f94, %f93; div.full.ftz.f32 %f803, %f801, %f802; mul.rn.ftz.f32 %f804, %f803, %f803; mov.f32 %f805, 0fC0B59883; mov.f32 %f806, 0fBF52C7EA; fma.rn.ftz.f32 %f807, %f804, %f806, %f805; mov.f32 %f808, 0fC0D21907; fma.rn.ftz.f32 %f809, %f807, %f804, %f808; mul.ftz.f32 %f810, %f804, %f809; mul.ftz.f32 %f811, %f803, %f810; add.ftz.f32 %f812, %f804, 0f41355DC0; mov.f32 %f813, 0f41E6BD60; fma.rn.ftz.f32 %f814, %f812, %f804, %f813; mov.f32 %f815, 0f419D92C8; fma.rn.ftz.f32 %f816, %f814, %f804, %f815; rcp.approx.ftz.f32 %f817, %f816; fma.rn.ftz.f32 %f818, %f811, %f817, %f803; sub.ftz.f32 %f820, %f1367, %f818; setp.gt.ftz.f32 %p66, %f94, %f93; selp.f32 %f821, %f820, %f818, %p66; sub.ftz.f32 %f823, %f1368, %f821; selp.f32 %f824, %f823, %f821, %p65; mov.b32 %r67, %f824; or.b32 %r68, %r28, %r67; mov.b32 %f825, %r68; add.ftz.f32 %f826, %f93, %f94; setp.le.ftz.f32 %p67, %f826, 0f7F800000; selp.f32 %f1388, %f825, %f826, %p67; $L__BB91_64: ld.const.f32 %f1370, [kernel_params+280]; ld.const.f32 %f1369, [kernel_params+276]; cos.approx.ftz.f32 %f827, %f1388; mul.ftz.f32 %f828, %f97, %f827; div.approx.ftz.f32 %f829, %f828, %f1369; sin.approx.ftz.f32 %f830, %f1388; mul.ftz.f32 %f831, %f97, %f830; div.approx.ftz.f32 %f832, %f831, %f1370; add.ftz.f32 %f1402, %f832, 0f3F000000; add.ftz.f32 %f1403, %f829, 0f3F000000; bra.uni $L__BB91_100; $L__BB91_55: setp.lt.s32 %p96, %r31, 0; min.ftz.f32 %f903, %f123, %f124; max.ftz.f32 %f904, %f123, %f124; div.full.ftz.f32 %f905, %f903, %f904; mul.rn.ftz.f32 %f906, %f905, %f905; mov.f32 %f907, 0fC0B59883; mov.f32 %f908, 0fBF52C7EA; fma.rn.ftz.f32 %f909, %f906, %f908, %f907; mov.f32 %f910, 0fC0D21907; fma.rn.ftz.f32 %f911, %f909, %f906, %f910; mul.ftz.f32 %f912, %f906, %f911; mul.ftz.f32 %f913, %f905, %f912; add.ftz.f32 %f914, %f906, 0f41355DC0; mov.f32 %f915, 0f41E6BD60; fma.rn.ftz.f32 %f916, %f914, %f906, %f915; mov.f32 %f917, 0f419D92C8; fma.rn.ftz.f32 %f918, %f916, %f906, %f917; rcp.approx.ftz.f32 %f919, %f918; fma.rn.ftz.f32 %f920, %f913, %f919, %f905; sub.ftz.f32 %f922, %f565, %f920; setp.gt.ftz.f32 %p97, %f123, %f124; selp.f32 %f923, %f922, %f920, %p97; sub.ftz.f32 %f925, %f568, %f923; selp.f32 %f926, %f925, %f923, %p96; mov.b32 %r89, %f926; or.b32 %r90, %r32, %r89; mov.b32 %f927, %r90; add.ftz.f32 %f928, %f124, %f123; setp.le.ftz.f32 %p98, %f928, 0f7F800000; selp.f32 %f1393, %f927, %f928, %p98; $L__BB91_81: div.approx.ftz.f32 %f129, %f1393, %f122; @%p1 bra $L__BB91_85; bra.uni $L__BB91_82; $L__BB91_85: shr.s32 %r100, %r27, 31; and.b32 %r101, %r100, 1078530011; or.b32 %r102, %r101, %r28; mov.b32 %f1394, %r102; bra.uni $L__BB91_86; $L__BB91_82: setp.eq.ftz.f32 %p100, %f93, 0f7F800000; setp.eq.ftz.f32 %p101, %f94, 0f7F800000; and.pred %p102, %p100, %p101; @%p102 bra $L__BB91_84; bra.uni $L__BB91_83; $L__BB91_84: setp.lt.s32 %p106, %r27, 0; selp.b32 %r98, 1075235812, 1061752795, %p106; or.b32 %r99, %r98, %r28; mov.b32 %f1394, %r99; bra.uni $L__BB91_86; $L__BB91_96: setp.lt.s32 %p127, %r27, 0; min.ftz.f32 %f1025, %f94, %f93; max.ftz.f32 %f1026, %f94, %f93; div.full.ftz.f32 %f1027, %f1025, %f1026; mul.rn.ftz.f32 %f1028, %f1027, %f1027; mov.f32 %f1029, 0fC0B59883; mov.f32 %f1030, 0fBF52C7EA; fma.rn.ftz.f32 %f1031, %f1028, %f1030, %f1029; mov.f32 %f1032, 0fC0D21907; fma.rn.ftz.f32 %f1033, %f1031, %f1028, %f1032; mul.ftz.f32 %f1034, %f1028, %f1033; mul.ftz.f32 %f1035, %f1027, %f1034; add.ftz.f32 %f1036, %f1028, 0f41355DC0; mov.f32 %f1037, 0f41E6BD60; fma.rn.ftz.f32 %f1038, %f1036, %f1028, %f1037; mov.f32 %f1039, 0f419D92C8; fma.rn.ftz.f32 %f1040, %f1038, %f1028, %f1039; rcp.approx.ftz.f32 %f1041, %f1040; fma.rn.ftz.f32 %f1042, %f1035, %f1041, %f1027; sub.ftz.f32 %f1044, %f565, %f1042; setp.gt.ftz.f32 %p128, %f94, %f93; selp.f32 %f1045, %f1044, %f1042, %p128; sub.ftz.f32 %f1047, %f568, %f1045; selp.f32 %f1048, %f1047, %f1045, %p127; mov.b32 %r110, %f1048; or.b32 %r111, %r28, %r110; mov.b32 %f1049, %r111; add.ftz.f32 %f1050, %f93, %f94; setp.le.ftz.f32 %p129, %f1050, 0f7F800000; selp.f32 %f1396, %f1049, %f1050, %p129; $L__BB91_99: cos.approx.ftz.f32 %f1051, %f1396; mul.ftz.f32 %f1052, %f149, %f1051; div.approx.ftz.f32 %f1053, %f1052, %f147; sin.approx.ftz.f32 %f1054, %f1396; mul.ftz.f32 %f1055, %f149, %f1054; div.approx.ftz.f32 %f1056, %f1055, %f148; add.ftz.f32 %f1402, %f1056, 0f3F000000; add.ftz.f32 %f1403, %f1053, 0f3F000000; bra.uni $L__BB91_100; $L__BB91_83: setp.lt.s32 %p103, %r27, 0; min.ftz.f32 %f929, %f94, %f93; max.ftz.f32 %f930, %f94, %f93; div.full.ftz.f32 %f931, %f929, %f930; mul.rn.ftz.f32 %f932, %f931, %f931; mov.f32 %f933, 0fC0B59883; mov.f32 %f934, 0fBF52C7EA; fma.rn.ftz.f32 %f935, %f932, %f934, %f933; mov.f32 %f936, 0fC0D21907; fma.rn.ftz.f32 %f937, %f935, %f932, %f936; mul.ftz.f32 %f938, %f932, %f937; mul.ftz.f32 %f939, %f931, %f938; add.ftz.f32 %f940, %f932, 0f41355DC0; mov.f32 %f941, 0f41E6BD60; fma.rn.ftz.f32 %f942, %f940, %f932, %f941; mov.f32 %f943, 0f419D92C8; fma.rn.ftz.f32 %f944, %f942, %f932, %f943; rcp.approx.ftz.f32 %f945, %f944; fma.rn.ftz.f32 %f946, %f939, %f945, %f931; sub.ftz.f32 %f948, %f565, %f946; setp.gt.ftz.f32 %p104, %f94, %f93; selp.f32 %f949, %f948, %f946, %p104; sub.ftz.f32 %f951, %f568, %f949; selp.f32 %f952, %f951, %f949, %p103; mov.b32 %r96, %f952; or.b32 %r97, %r28, %r96; mov.b32 %f953, %r97; add.ftz.f32 %f954, %f93, %f94; setp.le.ftz.f32 %p105, %f954, 0f7F800000; selp.f32 %f1394, %f953, %f954, %p105; $L__BB91_86: cos.approx.ftz.f32 %f955, %f1394; sin.approx.ftz.f32 %f956, %f1394; fma.rn.ftz.f32 %f1402, %f129, %f956, 0f3F000000; fma.rn.ftz.f32 %f1403, %f129, %f955, 0f3F000000; bra.uni $L__BB91_100; $L__BB91_90: setp.lt.s32 %p118, %r33, 0; min.ftz.f32 %f963, %f93, %f140; max.ftz.f32 %f964, %f93, %f140; div.full.ftz.f32 %f965, %f963, %f964; mul.rn.ftz.f32 %f966, %f965, %f965; mov.f32 %f967, 0fC0B59883; mov.f32 %f968, 0fBF52C7EA; fma.rn.ftz.f32 %f969, %f966, %f968, %f967; mov.f32 %f970, 0fC0D21907; fma.rn.ftz.f32 %f971, %f969, %f966, %f970; mul.ftz.f32 %f972, %f966, %f971; mul.ftz.f32 %f973, %f965, %f972; add.ftz.f32 %f974, %f966, 0f41355DC0; mov.f32 %f975, 0f41E6BD60; fma.rn.ftz.f32 %f976, %f974, %f966, %f975; mov.f32 %f977, 0f419D92C8; fma.rn.ftz.f32 %f978, %f976, %f966, %f977; rcp.approx.ftz.f32 %f979, %f978; fma.rn.ftz.f32 %f980, %f973, %f979, %f965; sub.ftz.f32 %f982, %f565, %f980; setp.gt.ftz.f32 %p119, %f93, %f140; selp.f32 %f983, %f982, %f980, %p119; sub.ftz.f32 %f985, %f568, %f983; selp.f32 %f986, %f985, %f983, %p118; mov.b32 %r103, %f986; or.b32 %r104, %r34, %r103; mov.b32 %f987, %r104; add.ftz.f32 %f988, %f140, %f93; setp.le.ftz.f32 %p120, %f988, 0f7F800000; selp.f32 %f1395, %f987, %f988, %p120; $L__BB91_93: sub.ftz.f32 %f989, %f1395, %f960; mul.ftz.f32 %f990, %f90, %f90; fma.rn.ftz.f32 %f991, %f89, %f89, %f990; fma.rn.ftz.f32 %f992, %f91, %f91, %f991; sqrt.approx.ftz.f32 %f993, %f992; div.approx.ftz.f32 %f994, %f91, %f993; abs.ftz.f32 %f995, %f994; sub.ftz.f32 %f997, %f545, %f995; mul.ftz.f32 %f998, %f997, 0f3F000000; sqrt.approx.ftz.f32 %f999, %f998; setp.gt.ftz.f32 %p122, %f995, 0f3F11EB85; selp.f32 %f1000, %f999, %f995, %p122; mul.ftz.f32 %f1001, %f1000, %f1000; fma.rn.ftz.f32 %f1004, %f554, %f1001, %f553; fma.rn.ftz.f32 %f1006, %f1004, %f1001, %f556; fma.rn.ftz.f32 %f1008, %f1006, %f1001, %f558; fma.rn.ftz.f32 %f1010, %f1008, %f1001, %f560; mul.ftz.f32 %f1011, %f1001, %f1010; fma.rn.ftz.f32 %f1012, %f1011, %f1000, %f1000; add.ftz.f32 %f1013, %f1012, %f1012; sub.ftz.f32 %f1015, %f565, %f1012; selp.f32 %f1016, %f1013, %f1015, %p122; setp.lt.ftz.f32 %p123, %f994, 0f00000000; sub.ftz.f32 %f1018, %f568, %f1016; selp.f32 %f1019, %f1018, %f1016, %p123; sub.ftz.f32 %f1020, %f1019, %f962; div.approx.ftz.f32 %f1402, %f1020, %f961; div.approx.ftz.f32 %f1403, %f989, %f959; $L__BB91_100: ld.param.u32 %r132, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_4]; mov.f32 %f1401, 0f00000000; $L__BB91_116: ld.param.u32 %r124, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_3+8]; ld.param.u64 %rd69, [_Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i_param_2]; cvta.to.local.u64 %rd66, %rd69; mul.wide.u32 %rd67, %r124, 4; add.s64 %rd68, %rd66, %rd67; st.local.f32 [%rd68], %f1403; st.local.f32 [%rd68+4], %f1402; st.local.f32 [%rd68+8], %f1401; st.param.b32 [func_retval0+0], %r132; ret; } .func _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<95>; .reg .b16 %rs<21>; .reg .f32 %f<873>; .reg .b32 %r<212>; .reg .b64 %rd<222>; ld.param.u64 %rd69, [_Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd70, [_Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r62, [_Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r61, [_Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r60, [_Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd71, %rd69; shr.u32 %r2, %r60, 24; cvta.to.local.u64 %rd1, %rd70; shl.b32 %r63, %r60, 2; cvt.u64.u32 %rd72, %r63; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; ld.local.f32 %f175, [%rd74]; add.ftz.f32 %f176, %f175, 0fBF000000; ld.local.f32 %f177, [%rd74+4]; add.ftz.f32 %f178, %f177, 0fBF000000; ld.local.f32 %f179, [%rd74+8]; add.ftz.f32 %f180, %f179, 0fBF000000; add.ftz.f32 %f861, %f176, %f176; add.ftz.f32 %f2, %f178, %f178; add.ftz.f32 %f3, %f180, %f180; add.s64 %rd2, %rd71, 52; ld.local.u32 %r64, [%rd71+52]; and.b32 %r3, %r64, 1; setp.eq.s32 %p1, %r2, 0; shr.u32 %r65, %r60, 14; cvt.u64.u32 %rd75, %r65; and.b64 %rd76, %rd75, 1020; add.s64 %rd3, %rd1, %rd76; @%p1 bra $L__BB92_7; add.s32 %r66, %r2, -3; setp.lt.u32 %p2, %r66, 2; neg.ftz.f32 %f181, %f2; selp.f32 %f862, %f181, %f2, %p2; neg.ftz.f32 %f182, %f3; selp.f32 %f863, %f182, %f3, %p2; or.b32 %r67, %r2, 2; setp.eq.s32 %p3, %r67, 3; @%p3 bra $L__BB92_4; bra.uni $L__BB92_2; $L__BB92_4: ld.local.u8 %rs13, [%rd2+4]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p5, %rs14, 0; @%p5 bra $L__BB92_6; bra.uni $L__BB92_5; $L__BB92_6: ld.const.u64 %rd77, [kernel_params+2160]; cvta.to.global.u64 %rd78, %rd77; ld.local.u32 %r68, [%rd2+24]; mul.wide.s32 %rd79, %r68, 208; add.s64 %rd80, %rd78, %rd79; ld.global.v4.f32 {%f221, %f222, %f223, %f224}, [%rd80+48]; ld.global.v4.f32 {%f228, %f229, %f230, %f231}, [%rd80+64]; ld.global.v4.f32 {%f235, %f236, %f237, %f238}, [%rd80+80]; mul.ftz.f32 %f242, %f862, %f228; fma.rn.ftz.f32 %f243, %f861, %f221, %f242; fma.rn.ftz.f32 %f244, %f863, %f235, %f243; mul.ftz.f32 %f245, %f862, %f229; fma.rn.ftz.f32 %f246, %f861, %f222, %f245; fma.rn.ftz.f32 %f247, %f863, %f236, %f246; mul.ftz.f32 %f248, %f862, %f230; fma.rn.ftz.f32 %f249, %f861, %f223, %f248; fma.rn.ftz.f32 %f250, %f863, %f237, %f249; mul.ftz.f32 %f251, %f247, %f247; fma.rn.ftz.f32 %f252, %f244, %f244, %f251; fma.rn.ftz.f32 %f253, %f250, %f250, %f252; sqrt.approx.ftz.f32 %f254, %f253; div.approx.ftz.f32 %f861, %f244, %f254; div.approx.ftz.f32 %f862, %f247, %f254; div.approx.ftz.f32 %f863, %f250, %f254; bra.uni $L__BB92_106; $L__BB92_7: ld.local.u32 %r4, [%rd2+24]; setp.eq.s32 %p6, %r4, -1; @%p6 bra $L__BB92_113; ld.local.u32 %r208, [%rd2+12]; and.b32 %r69, %r208, 1; setp.eq.b32 %p7, %r69, 1; mov.pred %p8, 0; xor.pred %p9, %p7, %p8; not.pred %p10, %p9; @%p10 bra $L__BB92_113; bra.uni $L__BB92_9; $L__BB92_113: ld.local.f32 %f830, [%rd2+-40]; ld.local.f32 %f831, [%rd2+-36]; ld.local.f32 %f832, [%rd2+-32]; st.local.f32 [%rd3], %f830; st.local.f32 [%rd3+4], %f831; st.local.f32 [%rd3+8], %f832; bra.uni $L__BB92_114; $L__BB92_2: mul.ftz.f32 %f183, %f862, %f862; fma.rn.ftz.f32 %f184, %f861, %f861, %f183; fma.rn.ftz.f32 %f185, %f863, %f863, %f184; sqrt.approx.ftz.f32 %f6, %f185; setp.neu.ftz.f32 %p4, %f6, 0f00000000; @%p4 bra $L__BB92_3; bra.uni $L__BB92_106; $L__BB92_3: rcp.approx.ftz.f32 %f186, %f6; mul.ftz.f32 %f863, %f863, %f186; mul.ftz.f32 %f862, %f862, %f186; mul.ftz.f32 %f861, %f861, %f186; bra.uni $L__BB92_106; $L__BB92_5: ld.local.v4.f32 {%f187, %f188, %f189, %f190}, [%rd2+140]; ld.local.v4.f32 {%f194, %f195, %f196, %f197}, [%rd2+156]; ld.local.v4.f32 {%f201, %f202, %f203, %f204}, [%rd2+172]; mul.ftz.f32 %f208, %f862, %f194; fma.rn.ftz.f32 %f209, %f861, %f187, %f208; fma.rn.ftz.f32 %f210, %f863, %f201, %f209; mul.ftz.f32 %f211, %f862, %f195; fma.rn.ftz.f32 %f212, %f861, %f188, %f211; fma.rn.ftz.f32 %f213, %f863, %f202, %f212; mul.ftz.f32 %f214, %f862, %f196; fma.rn.ftz.f32 %f215, %f861, %f189, %f214; fma.rn.ftz.f32 %f216, %f863, %f203, %f215; mul.ftz.f32 %f217, %f213, %f213; fma.rn.ftz.f32 %f218, %f210, %f210, %f217; fma.rn.ftz.f32 %f219, %f216, %f216, %f218; sqrt.approx.ftz.f32 %f220, %f219; div.approx.ftz.f32 %f861, %f210, %f220; div.approx.ftz.f32 %f862, %f213, %f220; div.approx.ftz.f32 %f863, %f216, %f220; $L__BB92_106: neg.ftz.f32 %f812, %f861; setp.ne.s32 %p87, %r3, 0; selp.f32 %f870, %f812, %f861, %p87; neg.ftz.f32 %f813, %f862; selp.f32 %f871, %f813, %f862, %p87; neg.ftz.f32 %f814, %f863; selp.f32 %f872, %f814, %f863, %p87; shr.u32 %r188, %r60, 6; cvt.u64.u32 %rd201, %r188; and.b64 %rd202, %rd201, 1020; add.s64 %rd203, %rd1, %rd202; ld.local.f32 %f149, [%rd203]; setp.eq.ftz.f32 %p88, %f149, 0f3F800000; @%p88 bra $L__BB92_109; mov.f32 %f815, 0f00000000; max.ftz.f32 %f816, %f149, %f815; ld.local.f32 %f817, [%rd2+-40]; sub.ftz.f32 %f818, %f870, %f817; ld.local.f32 %f819, [%rd2+-36]; sub.ftz.f32 %f820, %f871, %f819; ld.local.f32 %f821, [%rd2+-32]; sub.ftz.f32 %f822, %f872, %f821; fma.rn.ftz.f32 %f870, %f816, %f818, %f817; fma.rn.ftz.f32 %f871, %f816, %f820, %f819; fma.rn.ftz.f32 %f872, %f816, %f822, %f821; mul.ftz.f32 %f823, %f871, %f871; fma.rn.ftz.f32 %f824, %f870, %f870, %f823; fma.rn.ftz.f32 %f825, %f872, %f872, %f824; sqrt.approx.ftz.f32 %f153, %f825; setp.neu.ftz.f32 %p89, %f153, 0f00000000; @%p89 bra $L__BB92_108; bra.uni $L__BB92_109; $L__BB92_108: rcp.approx.ftz.f32 %f826, %f153; mul.ftz.f32 %f872, %f872, %f826; mul.ftz.f32 %f871, %f871, %f826; mul.ftz.f32 %f870, %f870, %f826; $L__BB92_109: setp.neu.ftz.f32 %p90, %f871, 0f00000000; setp.neu.ftz.f32 %p91, %f870, 0f00000000; or.pred %p92, %p91, %p90; setp.neu.ftz.f32 %p93, %f872, 0f00000000; or.pred %p94, %p93, %p92; @%p94 bra $L__BB92_111; ld.local.f32 %f870, [%rd2+-40]; ld.local.f32 %f871, [%rd2+-36]; ld.local.f32 %f872, [%rd2+-32]; $L__BB92_111: st.local.f32 [%rd3], %f870; st.local.f32 [%rd3+4], %f871; st.local.f32 [%rd3+8], %f872; $L__BB92_114: ret; $L__BB92_9: cvt.u64.u32 %rd4, %r61; ld.local.s32 %rd5, [%rd2+8]; ld.const.u64 %rd81, [kernel_params+2160]; cvta.to.global.u64 %rd6, %rd81; mul.wide.s32 %rd82, %r4, 208; add.s64 %rd83, %rd6, %rd82; ld.global.u32 %r6, [%rd83+164]; setp.eq.s64 %p11, %rd5, -1; mov.u32 %r189, -1; @%p11 bra $L__BB92_11; ld.const.u64 %rd84, [kernel_params+2240]; cvta.to.global.u64 %rd85, %rd84; shl.b64 %rd86, %rd5, 2; add.s64 %rd87, %rd85, %rd86; ld.global.u32 %r189, [%rd87]; $L__BB92_11: setp.ne.s32 %p12, %r189, -1; selp.u32 %r71, 1, 0, %p12; add.s32 %r191, %r6, %r71; ld.const.u64 %rd88, [kernel_params+2312]; cvta.to.global.u64 %rd7, %rd88; mul.wide.u32 %rd89, %r191, 16; add.s64 %rd90, %rd7, %rd89; add.s64 %rd206, %rd90, 14; ld.global.u16 %rs17, [%rd90+12]; ld.global.u32 %r190, [%rd90+8]; ld.global.u64 %rd204, [%rd90]; setp.eq.s64 %p13, %rd204, %rd4; add.s64 %rd205, %rd90, 15; @%p13 bra $L__BB92_17; $L__BB92_12: setp.eq.s64 %p14, %rd204, 0; @%p14 bra $L__BB92_14; add.s32 %r191, %r191, 2; bra.uni $L__BB92_15; $L__BB92_14: setp.eq.s16 %p15, %rs17, 0; mov.u32 %r24, -1; mov.u32 %r194, 0; mov.u32 %r191, %r190; mov.u32 %r195, %r194; mov.u32 %r196, %r194; @%p15 bra $L__BB92_21; $L__BB92_15: cvt.u64.u32 %rd12, %r191; mul.wide.u32 %rd91, %r191, 16; add.s64 %rd92, %rd7, %rd91; ld.global.u16 %rs17, [%rd92+12]; ld.global.u32 %r190, [%rd92+8]; ld.global.u64 %rd204, [%rd92]; setp.ne.s64 %p16, %rd204, %rd4; @%p16 bra $L__BB92_12; shl.b64 %rd93, %rd12, 4; add.s64 %rd94, %rd7, %rd93; add.s64 %rd206, %rd94, 14; add.s64 %rd205, %rd94, 15; $L__BB92_17: cvt.u32.u64 %r76, %rd5; setp.ne.s32 %p17, %r76, -1; ld.global.u8 %rs5, [%rd206]; ld.global.u8 %rs6, [%rd205]; @%p17 bra $L__BB92_20; mov.u32 %r24, -1; mov.u32 %r194, 0; add.s16 %rs15, %rs17, -1; setp.lt.u16 %p18, %rs15, 2; @%p18 bra $L__BB92_20; setp.ne.s16 %p19, %rs17, 1024; mov.u32 %r195, %r194; mov.u32 %r196, %r194; @%p19 bra $L__BB92_21; $L__BB92_20: setp.eq.s16 %p20, %rs17, 0; selp.b32 %r24, -1, %r190, %p20; cvt.u32.u16 %r194, %rs17; cvt.u32.u16 %r81, %rs6; and.b32 %r196, %r81, 255; cvt.u32.u16 %r82, %rs5; and.b32 %r195, %r82, 255; $L__BB92_21: cvt.u32.u64 %r84, %rd5; setp.eq.s32 %p21, %r84, -1; mov.u32 %r198, -1; cvt.u64.u32 %rd18, %r62; @%p21 bra $L__BB92_23; ld.const.u64 %rd95, [kernel_params+2240]; cvta.to.global.u64 %rd96, %rd95; shl.b64 %rd97, %rd5, 2; add.s64 %rd98, %rd96, %rd97; ld.global.u32 %r198, [%rd98]; $L__BB92_23: setp.ne.s32 %p22, %r198, -1; selp.u32 %r85, 1, 0, %p22; add.s32 %r200, %r6, %r85; mul.wide.u32 %rd99, %r200, 16; add.s64 %rd100, %rd7, %rd99; add.s64 %rd209, %rd100, 14; ld.global.u16 %rs19, [%rd100+12]; ld.global.u32 %r199, [%rd100+8]; ld.global.u64 %rd207, [%rd100]; setp.eq.s64 %p23, %rd207, %rd18; add.s64 %rd208, %rd100, 15; @%p23 bra $L__BB92_29; $L__BB92_24: setp.eq.s64 %p24, %rd207, 0; @%p24 bra $L__BB92_26; add.s32 %r200, %r200, 2; bra.uni $L__BB92_27; $L__BB92_26: setp.eq.s16 %p25, %rs19, 0; mov.u32 %r42, -1; mov.u32 %r39, 0; mov.u32 %r200, %r199; mov.u32 %r204, %r39; mov.u32 %r205, %r39; @%p25 bra $L__BB92_33; $L__BB92_27: cvt.u64.u32 %rd23, %r200; mul.wide.u32 %rd101, %r200, 16; add.s64 %rd102, %rd7, %rd101; ld.global.u16 %rs19, [%rd102+12]; ld.global.u32 %r199, [%rd102+8]; ld.global.u64 %rd207, [%rd102]; setp.ne.s64 %p26, %rd207, %rd18; @%p26 bra $L__BB92_24; shl.b64 %rd103, %rd23, 4; add.s64 %rd104, %rd7, %rd103; add.s64 %rd209, %rd104, 14; add.s64 %rd208, %rd104, 15; $L__BB92_29: setp.ne.s32 %p27, %r84, -1; ld.global.u8 %rs11, [%rd209]; ld.global.u8 %rs12, [%rd208]; @%p27 bra $L__BB92_32; mov.u32 %r42, -1; mov.u32 %r39, 0; add.s16 %rs16, %rs19, -1; setp.lt.u16 %p28, %rs16, 2; @%p28 bra $L__BB92_32; setp.ne.s16 %p29, %rs19, 1024; mov.u32 %r204, %r39; mov.u32 %r205, %r39; @%p29 bra $L__BB92_33; $L__BB92_32: setp.eq.s16 %p30, %rs19, 0; selp.b32 %r42, -1, %r199, %p30; cvt.u32.u16 %r39, %rs19; cvt.u32.u16 %r95, %rs12; and.b32 %r205, %r95, 255; cvt.u32.u16 %r96, %rs11; and.b32 %r204, %r96, 255; $L__BB92_33: setp.eq.s32 %p31, %r42, -1; setp.eq.s32 %p32, %r24, -1; or.pred %p33, %p32, %p31; @%p33 bra $L__BB92_112; bra.uni $L__BB92_34; $L__BB92_112: ld.local.f32 %f827, [%rd2+-40]; ld.local.f32 %f828, [%rd2+-36]; ld.local.f32 %f829, [%rd2+-32]; st.local.f32 [%rd3], %f827; st.local.f32 [%rd3+4], %f828; st.local.f32 [%rd3+8], %f829; bra.uni $L__BB92_114; $L__BB92_34: mov.u32 %r207, -1; @%p21 bra $L__BB92_36; ld.const.u64 %rd105, [kernel_params+2240]; cvta.to.global.u64 %rd106, %rd105; shl.b64 %rd107, %rd5, 2; add.s64 %rd108, %rd106, %rd107; ld.global.u32 %r207, [%rd108]; $L__BB92_36: setp.eq.s32 %p35, %r207, -1; @%p35 bra $L__BB92_38; mov.u64 %rd109, 0; { // callseq 64, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd69; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r194; st.param.b32 [param1+4], %r195; st.param.b32 [param1+8], %r196; st.param.b32 [param1+12], %r24; .param .b64 param2; st.param.b64 [param2+0], %rd109; .param .b64 param3; st.param.b64 [param3+0], %rd109; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f836, [retval0+0]; ld.param.f32 %f837, [retval0+4]; ld.param.f32 %f838, [retval0+8]; } // callseq 64 ld.local.u32 %r208, [%rd2+12]; bra.uni $L__BB92_45; $L__BB92_38: and.b32 %r99, %r194, 56; setp.eq.s32 %p36, %r99, 0; @%p36 bra $L__BB92_43; and.b32 %r100, %r194, 24; setp.eq.s32 %p37, %r100, 0; @%p37 bra $L__BB92_41; ld.const.u64 %rd110, [kernel_params+2232]; cvta.to.global.u64 %rd111, %rd110; shl.b64 %rd112, %rd5, 4; add.s64 %rd113, %rd111, %rd112; ld.global.v4.u32 {%r101, %r102, %r103, %r104}, [%rd113]; ld.const.u64 %rd114, [kernel_params+2336]; cvta.to.global.u64 %rd115, %rd114; add.s32 %r108, %r101, %r24; mul.wide.u32 %rd116, %r108, 12; add.s64 %rd216, %rd115, %rd116; add.s64 %rd217, %rd216, 4; add.s64 %rd218, %rd216, 8; add.s32 %r109, %r102, %r24; mul.wide.u32 %rd117, %r109, 12; add.s64 %rd213, %rd115, %rd117; add.s64 %rd214, %rd213, 4; add.s64 %rd215, %rd213, 8; add.s32 %r110, %r103, %r24; mul.wide.u32 %rd118, %r110, 12; add.s64 %rd210, %rd115, %rd118; add.s64 %rd211, %rd210, 4; add.s64 %rd212, %rd210, 8; bra.uni $L__BB92_42; $L__BB92_43: and.b32 %r115, %r194, 7; setp.eq.s32 %p38, %r115, 0; mov.f32 %f836, 0f00000000; mov.f32 %f837, %f836; mov.f32 %f838, %f836; @%p38 bra $L__BB92_45; setp.eq.s32 %p39, %r194, 4; selp.b32 %r117, %r84, 0, %p39; add.s32 %r118, %r24, %r117; ld.const.u64 %rd124, [kernel_params+2336]; cvta.to.global.u64 %rd125, %rd124; mul.wide.s32 %rd126, %r118, 12; add.s64 %rd127, %rd125, %rd126; ld.global.f32 %f838, [%rd127+8]; ld.global.f32 %f837, [%rd127+4]; ld.global.f32 %f836, [%rd127]; bra.uni $L__BB92_45; $L__BB92_41: mad.lo.s32 %r112, %r84, 3, %r24; ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r112, 12; add.s64 %rd216, %rd120, %rd121; add.s64 %rd217, %rd216, 4; add.s64 %rd218, %rd216, 8; add.s32 %r113, %r112, 1; mul.wide.s32 %rd122, %r113, 12; add.s64 %rd213, %rd120, %rd122; add.s64 %rd214, %rd213, 4; add.s64 %rd215, %rd213, 8; add.s32 %r114, %r112, 2; mul.wide.s32 %rd123, %r114, 12; add.s64 %rd210, %rd120, %rd123; add.s64 %rd211, %rd210, 4; add.s64 %rd212, %rd210, 8; $L__BB92_42: ld.local.f32 %f255, [%rd2+16]; ld.global.f32 %f256, [%rd213]; ld.global.f32 %f257, [%rd214]; ld.global.f32 %f258, [%rd215]; ld.local.f32 %f259, [%rd2+20]; ld.global.f32 %f260, [%rd210]; mul.ftz.f32 %f261, %f260, %f259; ld.global.f32 %f262, [%rd211]; mul.ftz.f32 %f263, %f262, %f259; ld.global.f32 %f264, [%rd212]; mul.ftz.f32 %f265, %f264, %f259; fma.rn.ftz.f32 %f266, %f256, %f255, %f261; fma.rn.ftz.f32 %f267, %f257, %f255, %f263; fma.rn.ftz.f32 %f268, %f258, %f255, %f265; mov.f32 %f269, 0f3F800000; sub.ftz.f32 %f270, %f269, %f255; sub.ftz.f32 %f271, %f270, %f259; ld.global.f32 %f272, [%rd216]; ld.global.f32 %f273, [%rd217]; ld.global.f32 %f274, [%rd218]; fma.rn.ftz.f32 %f838, %f274, %f271, %f268; fma.rn.ftz.f32 %f837, %f273, %f271, %f267; fma.rn.ftz.f32 %f836, %f272, %f271, %f266; $L__BB92_45: and.b32 %r119, %r208, 1; setp.eq.b32 %p40, %r119, 1; xor.pred %p42, %p40, %p8; not.pred %p43, %p42; @%p43 bra $L__BB92_57; bra.uni $L__BB92_46; $L__BB92_57: and.b32 %r140, %r208, 6; setp.eq.s32 %p50, %r140, 0; @%p50 bra $L__BB92_64; and.b32 %r141, %r39, 768; setp.eq.s32 %p51, %r141, 0; @%p51 bra $L__BB92_60; ld.const.u64 %rd150, [kernel_params+2264]; cvta.to.global.u64 %rd151, %rd150; ld.local.u32 %r142, [%rd2+8]; mul.wide.s32 %rd152, %r142, 16; add.s64 %rd153, %rd151, %rd152; shr.s32 %r143, %r208, 7; ld.global.u32 %r144, [%rd153+4]; add.s32 %r145, %r143, %r144; ld.const.u64 %rd154, [kernel_params+2320]; cvta.to.global.u64 %rd155, %rd154; add.s32 %r146, %r145, %r42; mul.wide.s32 %rd156, %r146, 4; add.s64 %rd157, %rd155, %rd156; ld.local.f32 %f289, [%rd2+16]; mov.f32 %f290, 0f3F800000; sub.ftz.f32 %f291, %f290, %f289; ld.global.f32 %f292, [%rd157]; mul.ftz.f32 %f293, %f292, %f291; ld.global.f32 %f294, [%rd157+4]; fma.rn.ftz.f32 %f839, %f294, %f289, %f293; bra.uni $L__BB92_66; $L__BB92_46: ld.local.s32 %rd56, [%rd2+8]; setp.eq.s64 %p44, %rd56, -1; mov.u32 %r209, -1; @%p44 bra $L__BB92_48; ld.const.u64 %rd128, [kernel_params+2240]; cvta.to.global.u64 %rd129, %rd128; shl.b64 %rd130, %rd56, 2; add.s64 %rd131, %rd129, %rd130; ld.global.u32 %r209, [%rd131]; $L__BB92_48: setp.eq.s32 %p45, %r209, -1; @%p45 bra $L__BB92_50; mov.u64 %rd132, 0; { // callseq 65, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd69; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r39; st.param.b32 [param1+4], %r204; st.param.b32 [param1+8], %r205; st.param.b32 [param1+12], %r42; .param .b64 param2; st.param.b64 [param2+0], %rd132; .param .b64 param3; st.param.b64 [param3+0], %rd132; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f839, [retval0+0]; } // callseq 65 bra.uni $L__BB92_66; $L__BB92_64: and.b32 %r149, %r208, 8; setp.eq.s32 %p54, %r149, 0; setp.ne.s32 %p55, %r39, 8; mov.f32 %f839, 0f00000000; or.pred %p56, %p55, %p54; @%p56 bra $L__BB92_66; ld.const.u64 %rd162, [kernel_params+2320]; cvta.to.global.u64 %rd163, %rd162; ld.local.u32 %r150, [%rd2+8]; add.s32 %r151, %r150, %r42; mul.wide.s32 %rd164, %r151, 4; add.s64 %rd165, %rd163, %rd164; ld.global.f32 %f839, [%rd165]; bra.uni $L__BB92_66; $L__BB92_50: and.b32 %r121, %r39, 56; setp.eq.s32 %p46, %r121, 0; @%p46 bra $L__BB92_55; and.b32 %r122, %r39, 24; setp.eq.s32 %p47, %r122, 0; @%p47 bra $L__BB92_53; ld.const.u64 %rd133, [kernel_params+2232]; cvta.to.global.u64 %rd134, %rd133; shl.b64 %rd135, %rd56, 4; add.s64 %rd136, %rd134, %rd135; ld.global.v4.u32 {%r123, %r124, %r125, %r126}, [%rd136]; ld.const.u64 %rd137, [kernel_params+2320]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r130, %r123, %r42; mul.wide.u32 %rd139, %r130, 4; add.s64 %rd219, %rd138, %rd139; add.s32 %r131, %r124, %r42; mul.wide.u32 %rd140, %r131, 4; add.s64 %rd220, %rd138, %rd140; add.s32 %r132, %r125, %r42; mul.wide.u32 %rd141, %r132, 4; add.s64 %rd221, %rd138, %rd141; bra.uni $L__BB92_54; $L__BB92_60: and.b32 %r147, %r39, 131; setp.eq.s32 %p52, %r147, 0; mov.f32 %f839, 0f00000000; @%p52 bra $L__BB92_66; setp.ne.s32 %p53, %r39, 128; @%p53 bra $L__BB92_63; ld.local.u32 %r148, [%rd2+8]; add.s32 %r42, %r148, %r42; $L__BB92_63: ld.const.u64 %rd158, [kernel_params+2320]; cvta.to.global.u64 %rd159, %rd158; mul.wide.s32 %rd160, %r42, 4; add.s64 %rd161, %rd159, %rd160; ld.global.f32 %f839, [%rd161]; bra.uni $L__BB92_66; $L__BB92_55: and.b32 %r136, %r39, 7; setp.eq.s32 %p48, %r136, 0; mov.f32 %f839, 0f00000000; @%p48 bra $L__BB92_66; cvt.u32.u64 %r137, %rd56; setp.eq.s32 %p49, %r39, 4; selp.b32 %r138, %r137, 0, %p49; add.s32 %r139, %r42, %r138; ld.const.u64 %rd146, [kernel_params+2320]; cvta.to.global.u64 %rd147, %rd146; mul.wide.s32 %rd148, %r139, 4; add.s64 %rd149, %rd147, %rd148; ld.global.f32 %f839, [%rd149]; bra.uni $L__BB92_66; $L__BB92_53: cvt.u32.u64 %r133, %rd56; mad.lo.s32 %r134, %r133, 3, %r42; ld.const.u64 %rd142, [kernel_params+2320]; cvta.to.global.u64 %rd143, %rd142; mul.wide.s32 %rd144, %r134, 4; add.s64 %rd219, %rd143, %rd144; add.s64 %rd220, %rd219, 4; add.s32 %r135, %r134, 2; mul.wide.s32 %rd145, %r135, 4; add.s64 %rd221, %rd143, %rd145; $L__BB92_54: ld.local.f32 %f278, [%rd2+16]; ld.global.f32 %f279, [%rd220]; ld.local.f32 %f280, [%rd2+20]; ld.global.f32 %f281, [%rd221]; mul.ftz.f32 %f282, %f281, %f280; fma.rn.ftz.f32 %f283, %f279, %f278, %f282; mov.f32 %f284, 0f3F800000; sub.ftz.f32 %f285, %f284, %f278; sub.ftz.f32 %f286, %f285, %f280; ld.global.f32 %f287, [%rd219]; fma.rn.ftz.f32 %f839, %f287, %f286, %f283; $L__BB92_66: ld.local.u32 %r152, [%rd2+-4]; setp.lt.s32 %p57, %r152, 0; ld.local.v2.f32 {%f297, %f298}, [%rd2+-28]; ld.local.f32 %f46, [%rd2+-20]; @%p57 bra $L__BB92_75; setp.ne.s32 %p58, %r3, 0; neg.ftz.f32 %f299, %f46; selp.f32 %f855, %f299, %f46, %p58; neg.ftz.f32 %f300, %f298; selp.f32 %f856, %f300, %f298, %p58; neg.ftz.f32 %f301, %f297; selp.f32 %f857, %f301, %f297, %p58; ld.local.u32 %r211, [%rd2+4]; and.b32 %r153, %r211, 2; setp.eq.s32 %p59, %r153, 0; ld.local.s32 %rd66, [%rd2+24]; @%p59 bra $L__BB92_71; bra.uni $L__BB92_68; $L__BB92_71: cvt.u32.u64 %r156, %rd66; setp.ne.s32 %p62, %r156, -1; @%p62 bra $L__BB92_74; bra.uni $L__BB92_72; $L__BB92_74: mul.lo.s64 %rd170, %rd66, 208; add.s64 %rd171, %rd6, %rd170; ld.global.v4.f32 {%f370, %f371, %f372, %f373}, [%rd171]; ld.global.v4.f32 {%f377, %f378, %f379, %f380}, [%rd171+16]; ld.global.v4.f32 {%f384, %f385, %f386, %f387}, [%rd171+32]; mul.ftz.f32 %f391, %f856, %f377; fma.rn.ftz.f32 %f392, %f857, %f370, %f391; fma.rn.ftz.f32 %f393, %f855, %f384, %f392; mul.ftz.f32 %f394, %f856, %f378; fma.rn.ftz.f32 %f395, %f857, %f371, %f394; fma.rn.ftz.f32 %f396, %f855, %f385, %f395; mul.ftz.f32 %f397, %f856, %f379; fma.rn.ftz.f32 %f398, %f857, %f372, %f397; fma.rn.ftz.f32 %f399, %f855, %f386, %f398; mul.ftz.f32 %f400, %f396, %f396; fma.rn.ftz.f32 %f401, %f393, %f393, %f400; fma.rn.ftz.f32 %f402, %f399, %f399, %f401; sqrt.approx.ftz.f32 %f403, %f402; div.approx.ftz.f32 %f857, %f393, %f403; div.approx.ftz.f32 %f856, %f396, %f403; div.approx.ftz.f32 %f855, %f399, %f403; bra.uni $L__BB92_101; $L__BB92_75: ld.local.v2.u32 {%r211, %r160}, [%rd2+4]; ld.local.f32 %f59, [%rd2+16]; ld.local.f32 %f60, [%rd2+20]; ld.const.v2.u64 {%rd172, %rd173}, [kernel_params+2224]; cvta.to.global.u64 %rd176, %rd173; mul.wide.s32 %rd177, %r160, 16; add.s64 %rd178, %rd176, %rd177; ld.global.v4.u32 {%r162, %r163, %r164, %r165}, [%rd178]; cvta.to.global.u64 %rd179, %rd172; mul.wide.u32 %rd180, %r162, 12; add.s64 %rd181, %rd179, %rd180; ld.global.f32 %f852, [%rd181]; ld.global.f32 %f853, [%rd181+4]; ld.global.f32 %f854, [%rd181+8]; mul.wide.u32 %rd182, %r163, 12; add.s64 %rd183, %rd179, %rd182; ld.global.f32 %f849, [%rd183]; ld.global.f32 %f850, [%rd183+4]; ld.global.f32 %f851, [%rd183+8]; mul.wide.u32 %rd184, %r164, 12; add.s64 %rd185, %rd179, %rd184; ld.global.f32 %f846, [%rd185]; ld.global.f32 %f847, [%rd185+4]; ld.global.f32 %f848, [%rd185+8]; and.b32 %r169, %r211, 4; setp.eq.s32 %p64, %r169, 0; @%p64 bra $L__BB92_100; and.b32 %r57, %r211, 2; setp.eq.s32 %p65, %r57, 0; ld.local.u32 %r170, [%rd2+24]; cvt.s64.s32 %rd67, %r170; mul.wide.s32 %rd186, %r170, 208; add.s64 %rd68, %rd6, %rd186; @%p65 bra $L__BB92_80; bra.uni $L__BB92_77; $L__BB92_80: cvt.u32.u64 %r173, %rd67; setp.ne.s32 %p68, %r173, -1; @%p68 bra $L__BB92_83; bra.uni $L__BB92_81; $L__BB92_83: ld.global.v4.f32 {%f472, %f473, %f474, %f475}, [%rd68]; ld.global.v4.f32 {%f479, %f480, %f481, %f482}, [%rd68+16]; ld.global.v4.f32 {%f486, %f487, %f488, %f489}, [%rd68+32]; mul.ftz.f32 %f493, %f853, %f479; fma.rn.ftz.f32 %f494, %f852, %f472, %f493; fma.rn.ftz.f32 %f495, %f854, %f486, %f494; mul.ftz.f32 %f496, %f853, %f480; fma.rn.ftz.f32 %f497, %f852, %f473, %f496; fma.rn.ftz.f32 %f498, %f854, %f487, %f497; mul.ftz.f32 %f499, %f853, %f481; fma.rn.ftz.f32 %f500, %f852, %f474, %f499; fma.rn.ftz.f32 %f501, %f854, %f488, %f500; mul.ftz.f32 %f502, %f498, %f498; fma.rn.ftz.f32 %f503, %f495, %f495, %f502; fma.rn.ftz.f32 %f504, %f501, %f501, %f503; sqrt.approx.ftz.f32 %f505, %f504; div.approx.ftz.f32 %f852, %f495, %f505; div.approx.ftz.f32 %f853, %f498, %f505; div.approx.ftz.f32 %f854, %f501, %f505; bra.uni $L__BB92_84; $L__BB92_68: cvt.u32.u64 %r154, %rd66; setp.ne.s32 %p60, %r154, -1; @%p60 bra $L__BB92_70; ld.local.u32 %r155, [%rd2+12]; setp.ne.s32 %p61, %r155, 32; @%p61 bra $L__BB92_101; $L__BB92_70: ld.local.v4.f32 {%f302, %f303, %f304, %f305}, [%rd2+92]; ld.local.v4.f32 {%f309, %f310, %f311, %f312}, [%rd2+108]; ld.local.v4.f32 {%f316, %f317, %f318, %f319}, [%rd2+124]; mul.ftz.f32 %f323, %f856, %f309; fma.rn.ftz.f32 %f324, %f857, %f302, %f323; fma.rn.ftz.f32 %f325, %f855, %f316, %f324; mul.ftz.f32 %f326, %f856, %f310; fma.rn.ftz.f32 %f327, %f857, %f303, %f326; fma.rn.ftz.f32 %f328, %f855, %f317, %f327; mul.ftz.f32 %f329, %f856, %f311; fma.rn.ftz.f32 %f330, %f857, %f304, %f329; fma.rn.ftz.f32 %f331, %f855, %f318, %f330; mul.ftz.f32 %f332, %f328, %f328; fma.rn.ftz.f32 %f333, %f325, %f325, %f332; fma.rn.ftz.f32 %f334, %f331, %f331, %f333; sqrt.approx.ftz.f32 %f335, %f334; div.approx.ftz.f32 %f857, %f325, %f335; div.approx.ftz.f32 %f856, %f328, %f335; div.approx.ftz.f32 %f855, %f331, %f335; bra.uni $L__BB92_101; $L__BB92_72: ld.local.u32 %r157, [%rd2+12]; setp.ne.s32 %p63, %r157, 32; @%p63 bra $L__BB92_101; ld.const.u64 %rd166, [kernel_params+2368]; cvta.to.global.u64 %rd167, %rd166; ld.local.u32 %r158, [%rd2+28]; mul.wide.s32 %rd168, %r158, 192; add.s64 %rd169, %rd167, %rd168; ld.global.v4.f32 {%f336, %f337, %f338, %f339}, [%rd169+48]; ld.global.v4.f32 {%f343, %f344, %f345, %f346}, [%rd169+64]; ld.global.v4.f32 {%f350, %f351, %f352, %f353}, [%rd169+80]; mul.ftz.f32 %f357, %f856, %f343; fma.rn.ftz.f32 %f358, %f857, %f336, %f357; fma.rn.ftz.f32 %f359, %f855, %f350, %f358; mul.ftz.f32 %f360, %f856, %f344; fma.rn.ftz.f32 %f361, %f857, %f337, %f360; fma.rn.ftz.f32 %f362, %f855, %f351, %f361; mul.ftz.f32 %f363, %f856, %f345; fma.rn.ftz.f32 %f364, %f857, %f338, %f363; fma.rn.ftz.f32 %f365, %f855, %f352, %f364; mul.ftz.f32 %f366, %f362, %f362; fma.rn.ftz.f32 %f367, %f359, %f359, %f366; fma.rn.ftz.f32 %f368, %f365, %f365, %f367; sqrt.approx.ftz.f32 %f369, %f368; div.approx.ftz.f32 %f857, %f359, %f369; div.approx.ftz.f32 %f856, %f362, %f369; div.approx.ftz.f32 %f855, %f365, %f369; bra.uni $L__BB92_101; $L__BB92_77: cvt.u32.u64 %r171, %rd67; setp.ne.s32 %p66, %r171, -1; @%p66 bra $L__BB92_79; ld.local.u32 %r172, [%rd2+12]; setp.ne.s32 %p67, %r172, 32; @%p67 bra $L__BB92_84; $L__BB92_79: ld.local.v4.f32 {%f404, %f405, %f406, %f407}, [%rd2+92]; ld.local.v4.f32 {%f411, %f412, %f413, %f414}, [%rd2+108]; ld.local.v4.f32 {%f418, %f419, %f420, %f421}, [%rd2+124]; mul.ftz.f32 %f425, %f853, %f411; fma.rn.ftz.f32 %f426, %f852, %f404, %f425; fma.rn.ftz.f32 %f427, %f854, %f418, %f426; mul.ftz.f32 %f428, %f853, %f412; fma.rn.ftz.f32 %f429, %f852, %f405, %f428; fma.rn.ftz.f32 %f430, %f854, %f419, %f429; mul.ftz.f32 %f431, %f853, %f413; fma.rn.ftz.f32 %f432, %f852, %f406, %f431; fma.rn.ftz.f32 %f433, %f854, %f420, %f432; mul.ftz.f32 %f434, %f430, %f430; fma.rn.ftz.f32 %f435, %f427, %f427, %f434; fma.rn.ftz.f32 %f436, %f433, %f433, %f435; sqrt.approx.ftz.f32 %f437, %f436; div.approx.ftz.f32 %f852, %f427, %f437; div.approx.ftz.f32 %f853, %f430, %f437; div.approx.ftz.f32 %f854, %f433, %f437; bra.uni $L__BB92_84; $L__BB92_81: ld.local.u32 %r174, [%rd2+12]; setp.ne.s32 %p69, %r174, 32; @%p69 bra $L__BB92_84; ld.const.u64 %rd187, [kernel_params+2368]; cvta.to.global.u64 %rd188, %rd187; ld.local.u32 %r175, [%rd2+28]; mul.wide.s32 %rd189, %r175, 192; add.s64 %rd190, %rd188, %rd189; ld.global.v4.f32 {%f438, %f439, %f440, %f441}, [%rd190+48]; ld.global.v4.f32 {%f445, %f446, %f447, %f448}, [%rd190+64]; ld.global.v4.f32 {%f452, %f453, %f454, %f455}, [%rd190+80]; mul.ftz.f32 %f459, %f853, %f445; fma.rn.ftz.f32 %f460, %f852, %f438, %f459; fma.rn.ftz.f32 %f461, %f854, %f452, %f460; mul.ftz.f32 %f462, %f853, %f446; fma.rn.ftz.f32 %f463, %f852, %f439, %f462; fma.rn.ftz.f32 %f464, %f854, %f453, %f463; mul.ftz.f32 %f465, %f853, %f447; fma.rn.ftz.f32 %f466, %f852, %f440, %f465; fma.rn.ftz.f32 %f467, %f854, %f454, %f466; mul.ftz.f32 %f468, %f464, %f464; fma.rn.ftz.f32 %f469, %f461, %f461, %f468; fma.rn.ftz.f32 %f470, %f467, %f467, %f469; sqrt.approx.ftz.f32 %f471, %f470; div.approx.ftz.f32 %f852, %f461, %f471; div.approx.ftz.f32 %f853, %f464, %f471; div.approx.ftz.f32 %f854, %f467, %f471; $L__BB92_84: @%p65 bra $L__BB92_88; bra.uni $L__BB92_85; $L__BB92_88: cvt.u32.u64 %r178, %rd67; setp.ne.s32 %p73, %r178, -1; @%p73 bra $L__BB92_91; bra.uni $L__BB92_89; $L__BB92_91: ld.global.v4.f32 {%f574, %f575, %f576, %f577}, [%rd68]; ld.global.v4.f32 {%f581, %f582, %f583, %f584}, [%rd68+16]; ld.global.v4.f32 {%f588, %f589, %f590, %f591}, [%rd68+32]; mul.ftz.f32 %f595, %f850, %f581; fma.rn.ftz.f32 %f596, %f849, %f574, %f595; fma.rn.ftz.f32 %f597, %f851, %f588, %f596; mul.ftz.f32 %f598, %f850, %f582; fma.rn.ftz.f32 %f599, %f849, %f575, %f598; fma.rn.ftz.f32 %f600, %f851, %f589, %f599; mul.ftz.f32 %f601, %f850, %f583; fma.rn.ftz.f32 %f602, %f849, %f576, %f601; fma.rn.ftz.f32 %f603, %f851, %f590, %f602; mul.ftz.f32 %f604, %f600, %f600; fma.rn.ftz.f32 %f605, %f597, %f597, %f604; fma.rn.ftz.f32 %f606, %f603, %f603, %f605; sqrt.approx.ftz.f32 %f607, %f606; div.approx.ftz.f32 %f849, %f597, %f607; div.approx.ftz.f32 %f850, %f600, %f607; div.approx.ftz.f32 %f851, %f603, %f607; bra.uni $L__BB92_92; $L__BB92_85: cvt.u32.u64 %r176, %rd67; setp.ne.s32 %p71, %r176, -1; @%p71 bra $L__BB92_87; ld.local.u32 %r177, [%rd2+12]; setp.ne.s32 %p72, %r177, 32; @%p72 bra $L__BB92_92; $L__BB92_87: ld.local.v4.f32 {%f506, %f507, %f508, %f509}, [%rd2+92]; ld.local.v4.f32 {%f513, %f514, %f515, %f516}, [%rd2+108]; ld.local.v4.f32 {%f520, %f521, %f522, %f523}, [%rd2+124]; mul.ftz.f32 %f527, %f850, %f513; fma.rn.ftz.f32 %f528, %f849, %f506, %f527; fma.rn.ftz.f32 %f529, %f851, %f520, %f528; mul.ftz.f32 %f530, %f850, %f514; fma.rn.ftz.f32 %f531, %f849, %f507, %f530; fma.rn.ftz.f32 %f532, %f851, %f521, %f531; mul.ftz.f32 %f533, %f850, %f515; fma.rn.ftz.f32 %f534, %f849, %f508, %f533; fma.rn.ftz.f32 %f535, %f851, %f522, %f534; mul.ftz.f32 %f536, %f532, %f532; fma.rn.ftz.f32 %f537, %f529, %f529, %f536; fma.rn.ftz.f32 %f538, %f535, %f535, %f537; sqrt.approx.ftz.f32 %f539, %f538; div.approx.ftz.f32 %f849, %f529, %f539; div.approx.ftz.f32 %f850, %f532, %f539; div.approx.ftz.f32 %f851, %f535, %f539; bra.uni $L__BB92_92; $L__BB92_89: ld.local.u32 %r179, [%rd2+12]; setp.ne.s32 %p74, %r179, 32; @%p74 bra $L__BB92_92; ld.const.u64 %rd191, [kernel_params+2368]; cvta.to.global.u64 %rd192, %rd191; ld.local.u32 %r180, [%rd2+28]; mul.wide.s32 %rd193, %r180, 192; add.s64 %rd194, %rd192, %rd193; ld.global.v4.f32 {%f540, %f541, %f542, %f543}, [%rd194+48]; ld.global.v4.f32 {%f547, %f548, %f549, %f550}, [%rd194+64]; ld.global.v4.f32 {%f554, %f555, %f556, %f557}, [%rd194+80]; mul.ftz.f32 %f561, %f850, %f547; fma.rn.ftz.f32 %f562, %f849, %f540, %f561; fma.rn.ftz.f32 %f563, %f851, %f554, %f562; mul.ftz.f32 %f564, %f850, %f548; fma.rn.ftz.f32 %f565, %f849, %f541, %f564; fma.rn.ftz.f32 %f566, %f851, %f555, %f565; mul.ftz.f32 %f567, %f850, %f549; fma.rn.ftz.f32 %f568, %f849, %f542, %f567; fma.rn.ftz.f32 %f569, %f851, %f556, %f568; mul.ftz.f32 %f570, %f566, %f566; fma.rn.ftz.f32 %f571, %f563, %f563, %f570; fma.rn.ftz.f32 %f572, %f569, %f569, %f571; sqrt.approx.ftz.f32 %f573, %f572; div.approx.ftz.f32 %f849, %f563, %f573; div.approx.ftz.f32 %f850, %f566, %f573; div.approx.ftz.f32 %f851, %f569, %f573; $L__BB92_92: @%p65 bra $L__BB92_96; bra.uni $L__BB92_93; $L__BB92_96: cvt.u32.u64 %r183, %rd67; setp.ne.s32 %p78, %r183, -1; @%p78 bra $L__BB92_99; bra.uni $L__BB92_97; $L__BB92_99: ld.global.v4.f32 {%f676, %f677, %f678, %f679}, [%rd68]; ld.global.v4.f32 {%f683, %f684, %f685, %f686}, [%rd68+16]; ld.global.v4.f32 {%f690, %f691, %f692, %f693}, [%rd68+32]; mul.ftz.f32 %f697, %f847, %f683; fma.rn.ftz.f32 %f698, %f846, %f676, %f697; fma.rn.ftz.f32 %f699, %f848, %f690, %f698; mul.ftz.f32 %f700, %f847, %f684; fma.rn.ftz.f32 %f701, %f846, %f677, %f700; fma.rn.ftz.f32 %f702, %f848, %f691, %f701; mul.ftz.f32 %f703, %f847, %f685; fma.rn.ftz.f32 %f704, %f846, %f678, %f703; fma.rn.ftz.f32 %f705, %f848, %f692, %f704; mul.ftz.f32 %f706, %f702, %f702; fma.rn.ftz.f32 %f707, %f699, %f699, %f706; fma.rn.ftz.f32 %f708, %f705, %f705, %f707; sqrt.approx.ftz.f32 %f709, %f708; div.approx.ftz.f32 %f846, %f699, %f709; div.approx.ftz.f32 %f847, %f702, %f709; div.approx.ftz.f32 %f848, %f705, %f709; bra.uni $L__BB92_100; $L__BB92_93: cvt.u32.u64 %r181, %rd67; setp.ne.s32 %p76, %r181, -1; @%p76 bra $L__BB92_95; ld.local.u32 %r182, [%rd2+12]; setp.ne.s32 %p77, %r182, 32; @%p77 bra $L__BB92_100; $L__BB92_95: ld.local.v4.f32 {%f608, %f609, %f610, %f611}, [%rd2+92]; ld.local.v4.f32 {%f615, %f616, %f617, %f618}, [%rd2+108]; ld.local.v4.f32 {%f622, %f623, %f624, %f625}, [%rd2+124]; mul.ftz.f32 %f629, %f847, %f615; fma.rn.ftz.f32 %f630, %f846, %f608, %f629; fma.rn.ftz.f32 %f631, %f848, %f622, %f630; mul.ftz.f32 %f632, %f847, %f616; fma.rn.ftz.f32 %f633, %f846, %f609, %f632; fma.rn.ftz.f32 %f634, %f848, %f623, %f633; mul.ftz.f32 %f635, %f847, %f617; fma.rn.ftz.f32 %f636, %f846, %f610, %f635; fma.rn.ftz.f32 %f637, %f848, %f624, %f636; mul.ftz.f32 %f638, %f634, %f634; fma.rn.ftz.f32 %f639, %f631, %f631, %f638; fma.rn.ftz.f32 %f640, %f637, %f637, %f639; sqrt.approx.ftz.f32 %f641, %f640; div.approx.ftz.f32 %f846, %f631, %f641; div.approx.ftz.f32 %f847, %f634, %f641; div.approx.ftz.f32 %f848, %f637, %f641; bra.uni $L__BB92_100; $L__BB92_97: ld.local.u32 %r184, [%rd2+12]; setp.ne.s32 %p79, %r184, 32; @%p79 bra $L__BB92_100; ld.const.u64 %rd195, [kernel_params+2368]; cvta.to.global.u64 %rd196, %rd195; ld.local.u32 %r185, [%rd2+28]; mul.wide.s32 %rd197, %r185, 192; add.s64 %rd198, %rd196, %rd197; ld.global.v4.f32 {%f642, %f643, %f644, %f645}, [%rd198+48]; ld.global.v4.f32 {%f649, %f650, %f651, %f652}, [%rd198+64]; ld.global.v4.f32 {%f656, %f657, %f658, %f659}, [%rd198+80]; mul.ftz.f32 %f663, %f847, %f649; fma.rn.ftz.f32 %f664, %f846, %f642, %f663; fma.rn.ftz.f32 %f665, %f848, %f656, %f664; mul.ftz.f32 %f666, %f847, %f650; fma.rn.ftz.f32 %f667, %f846, %f643, %f666; fma.rn.ftz.f32 %f668, %f848, %f657, %f667; mul.ftz.f32 %f669, %f847, %f651; fma.rn.ftz.f32 %f670, %f846, %f644, %f669; fma.rn.ftz.f32 %f671, %f848, %f658, %f670; mul.ftz.f32 %f672, %f668, %f668; fma.rn.ftz.f32 %f673, %f665, %f665, %f672; fma.rn.ftz.f32 %f674, %f671, %f671, %f673; sqrt.approx.ftz.f32 %f675, %f674; div.approx.ftz.f32 %f846, %f665, %f675; div.approx.ftz.f32 %f847, %f668, %f675; div.approx.ftz.f32 %f848, %f671, %f675; $L__BB92_100: mov.f32 %f710, 0f3F800000; sub.ftz.f32 %f711, %f710, %f59; sub.ftz.f32 %f712, %f711, %f60; mul.ftz.f32 %f713, %f712, %f852; mul.ftz.f32 %f714, %f712, %f853; mul.ftz.f32 %f715, %f712, %f854; fma.rn.ftz.f32 %f716, %f59, %f849, %f713; fma.rn.ftz.f32 %f717, %f59, %f850, %f714; fma.rn.ftz.f32 %f718, %f59, %f851, %f715; fma.rn.ftz.f32 %f719, %f60, %f848, %f718; fma.rn.ftz.f32 %f720, %f60, %f847, %f717; fma.rn.ftz.f32 %f721, %f60, %f846, %f716; setp.eq.ftz.f32 %p80, %f721, 0f00000000; setp.eq.ftz.f32 %p81, %f720, 0f00000000; and.pred %p82, %p80, %p81; setp.eq.ftz.f32 %p83, %f719, 0f00000000; and.pred %p84, %p83, %p82; selp.f32 %f857, %f297, %f721, %p84; selp.f32 %f856, %f298, %f720, %p84; selp.f32 %f855, %f46, %f719, %p84; $L__BB92_101: mul.ftz.f32 %f722, %f837, %f855; mul.ftz.f32 %f723, %f838, %f856; sub.ftz.f32 %f724, %f723, %f722; mul.ftz.f32 %f725, %f838, %f857; mul.ftz.f32 %f726, %f836, %f855; sub.ftz.f32 %f727, %f726, %f725; mul.ftz.f32 %f728, %f836, %f856; mul.ftz.f32 %f729, %f837, %f857; sub.ftz.f32 %f730, %f729, %f728; mul.ftz.f32 %f731, %f839, %f724; mul.ftz.f32 %f732, %f839, %f727; mul.ftz.f32 %f733, %f839, %f730; mul.ftz.f32 %f734, %f2, %f731; mul.ftz.f32 %f735, %f2, %f732; mul.ftz.f32 %f736, %f2, %f733; fma.rn.ftz.f32 %f737, %f861, %f836, %f734; fma.rn.ftz.f32 %f738, %f861, %f837, %f735; fma.rn.ftz.f32 %f739, %f861, %f838, %f736; fma.rn.ftz.f32 %f858, %f3, %f857, %f737; fma.rn.ftz.f32 %f859, %f3, %f856, %f738; fma.rn.ftz.f32 %f860, %f3, %f855, %f739; mul.ftz.f32 %f740, %f859, %f859; fma.rn.ftz.f32 %f741, %f858, %f858, %f740; fma.rn.ftz.f32 %f742, %f860, %f860, %f741; sqrt.approx.ftz.f32 %f124, %f742; setp.neu.ftz.f32 %p85, %f124, 0f00000000; @%p85 bra $L__BB92_102; bra.uni $L__BB92_103; $L__BB92_102: rcp.approx.ftz.f32 %f743, %f124; mul.ftz.f32 %f860, %f743, %f860; mul.ftz.f32 %f859, %f743, %f859; mul.ftz.f32 %f858, %f743, %f858; $L__BB92_103: and.b32 %r186, %r211, 2; setp.eq.s32 %p86, %r186, 0; @%p86 bra $L__BB92_105; bra.uni $L__BB92_104; $L__BB92_105: ld.local.u32 %r187, [%rd2+24]; mul.wide.s32 %rd199, %r187, 208; add.s64 %rd200, %rd6, %rd199; ld.global.v4.f32 {%f778, %f779, %f780, %f781}, [%rd200+48]; ld.global.v4.f32 {%f785, %f786, %f787, %f788}, [%rd200+64]; ld.global.v4.f32 {%f792, %f793, %f794, %f795}, [%rd200+80]; mul.ftz.f32 %f799, %f859, %f785; fma.rn.ftz.f32 %f800, %f858, %f778, %f799; fma.rn.ftz.f32 %f801, %f860, %f792, %f800; mul.ftz.f32 %f802, %f859, %f786; fma.rn.ftz.f32 %f803, %f858, %f779, %f802; fma.rn.ftz.f32 %f804, %f860, %f793, %f803; mul.ftz.f32 %f805, %f859, %f787; fma.rn.ftz.f32 %f806, %f858, %f780, %f805; fma.rn.ftz.f32 %f807, %f860, %f794, %f806; mul.ftz.f32 %f808, %f804, %f804; fma.rn.ftz.f32 %f809, %f801, %f801, %f808; fma.rn.ftz.f32 %f810, %f807, %f807, %f809; sqrt.approx.ftz.f32 %f811, %f810; div.approx.ftz.f32 %f861, %f801, %f811; div.approx.ftz.f32 %f862, %f804, %f811; div.approx.ftz.f32 %f863, %f807, %f811; bra.uni $L__BB92_106; $L__BB92_104: ld.local.v4.f32 {%f744, %f745, %f746, %f747}, [%rd2+140]; ld.local.v4.f32 {%f751, %f752, %f753, %f754}, [%rd2+156]; ld.local.v4.f32 {%f758, %f759, %f760, %f761}, [%rd2+172]; mul.ftz.f32 %f765, %f859, %f751; fma.rn.ftz.f32 %f766, %f858, %f744, %f765; fma.rn.ftz.f32 %f767, %f860, %f758, %f766; mul.ftz.f32 %f768, %f859, %f752; fma.rn.ftz.f32 %f769, %f858, %f745, %f768; fma.rn.ftz.f32 %f770, %f860, %f759, %f769; mul.ftz.f32 %f771, %f859, %f753; fma.rn.ftz.f32 %f772, %f858, %f746, %f771; fma.rn.ftz.f32 %f773, %f860, %f760, %f772; mul.ftz.f32 %f774, %f770, %f770; fma.rn.ftz.f32 %f775, %f767, %f767, %f774; fma.rn.ftz.f32 %f776, %f773, %f773, %f775; sqrt.approx.ftz.f32 %f777, %f776; div.approx.ftz.f32 %f861, %f767, %f777; div.approx.ftz.f32 %f862, %f770, %f777; div.approx.ftz.f32 %f863, %f773, %f777; bra.uni $L__BB92_106; } .func _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<52>; .reg .b16 %rs<12>; .reg .f32 %f<303>; .reg .b32 %r<135>; .reg .b64 %rd<167>; ld.param.u64 %rd49, [_Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd50, [_Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %rd1, [_Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r1, [_Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; bfe.u32 %r2, %r1, 16, 8; cvta.to.local.u64 %rd51, %rd49; add.s64 %rd2, %rd51, 76; ld.local.u32 %r3, [%rd51+76]; ld.local.s32 %rd9, [%rd51+60]; setp.eq.s32 %p1, %r3, -1; mov.u32 %r134, -1; mov.u32 %r21, 0; mov.u32 %r132, %r21; mov.u32 %r133, %r21; @%p1 bra $L__BB93_15; ld.const.u64 %rd52, [kernel_params+2160]; cvta.to.global.u64 %rd53, %rd52; mul.wide.s32 %rd54, %r3, 208; add.s64 %rd55, %rd53, %rd54; ld.global.u32 %r4, [%rd55+164]; ld.local.u32 %r38, [%rd2+-12]; and.b32 %r39, %r38, 1; setp.eq.b32 %p2, %r39, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r126, 0; @%p5 bra $L__BB93_5; cvt.u32.u64 %r41, %rd9; setp.eq.s32 %p6, %r41, -1; mov.u32 %r125, -1; @%p6 bra $L__BB93_4; ld.const.u64 %rd56, [kernel_params+2240]; cvta.to.global.u64 %rd57, %rd56; shl.b64 %rd58, %rd9, 2; add.s64 %rd59, %rd57, %rd58; ld.global.u32 %r125, [%rd59]; $L__BB93_4: setp.ne.s32 %p7, %r125, -1; selp.u32 %r126, 1, 0, %p7; $L__BB93_5: ld.const.u64 %rd60, [kernel_params+2312]; cvta.to.global.u64 %rd10, %rd60; add.s32 %r128, %r126, %r4; mul.wide.u32 %rd61, %r128, 16; add.s64 %rd62, %rd10, %rd61; add.s64 %rd157, %rd62, 14; add.s64 %rd156, %rd62, 15; ld.global.u16 %rs10, [%rd62+12]; ld.global.u32 %r127, [%rd62+8]; ld.global.u64 %rd155, [%rd62]; setp.eq.s64 %p8, %rd155, %rd1; @%p8 bra $L__BB93_11; $L__BB93_6: setp.eq.s64 %p9, %rd155, 0; @%p9 bra $L__BB93_8; add.s32 %r128, %r128, 2; bra.uni $L__BB93_9; $L__BB93_8: setp.eq.s16 %p10, %rs10, 0; mov.u32 %r21, 0; mov.u32 %r128, %r127; mov.u32 %r132, %r21; mov.u32 %r133, %r21; @%p10 bra $L__BB93_15; $L__BB93_9: cvt.u64.u32 %rd15, %r128; mul.wide.u32 %rd63, %r128, 16; add.s64 %rd64, %rd10, %rd63; ld.global.u16 %rs10, [%rd64+12]; ld.global.u32 %r127, [%rd64+8]; ld.global.u64 %rd155, [%rd64]; setp.ne.s64 %p11, %rd155, %rd1; @%p11 bra $L__BB93_6; shl.b64 %rd65, %rd15, 4; add.s64 %rd66, %rd10, %rd65; add.s64 %rd157, %rd66, 14; add.s64 %rd156, %rd66, 15; $L__BB93_11: ld.global.u8 %rs5, [%rd157]; ld.global.u8 %rs6, [%rd156]; cvt.u32.u64 %r46, %rd9; setp.ne.s32 %p12, %r46, -1; @%p12 bra $L__BB93_14; mov.u32 %r21, 0; add.s16 %rs7, %rs10, -1; setp.lt.u16 %p13, %rs7, 2; @%p13 bra $L__BB93_14; setp.ne.s16 %p14, %rs10, 1024; mov.u32 %r132, %r21; mov.u32 %r133, %r21; @%p14 bra $L__BB93_15; $L__BB93_14: setp.eq.s16 %p15, %rs10, 0; selp.b32 %r134, -1, %r127, %p15; cvt.u32.u16 %r21, %rs10; cvt.u32.u16 %r51, %rs6; and.b32 %r133, %r51, 255; cvt.u32.u16 %r52, %rs5; and.b32 %r132, %r52, 255; $L__BB93_15: setp.eq.s32 %p16, %r134, -1; @%p16 bra $L__BB93_54; ld.local.u32 %r31, [%rd2+-12]; and.b32 %r32, %r31, 1; setp.eq.s32 %p17, %r132, 1; @%p17 bra $L__BB93_35; bra.uni $L__BB93_17; $L__BB93_35: setp.eq.s32 %p32, %r32, 0; @%p32 bra $L__BB93_46; bra.uni $L__BB93_36; $L__BB93_46: and.b32 %r108, %r31, 6; setp.eq.s32 %p39, %r108, 0; @%p39 bra $L__BB93_51; and.b32 %r109, %r21, 768; setp.eq.s32 %p40, %r109, 0; @%p40 bra $L__BB93_49; ld.const.u64 %rd131, [kernel_params+2264]; cvta.to.global.u64 %rd132, %rd131; shl.b64 %rd133, %rd9, 4; add.s64 %rd134, %rd132, %rd133; shr.s32 %r110, %r31, 7; ld.global.u32 %r111, [%rd134+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd135, [kernel_params+2328]; cvta.to.global.u64 %rd136, %rd135; add.s32 %r113, %r112, %r134; mul.wide.s32 %rd137, %r113, 8; add.s64 %rd138, %rd136, %rd137; ld.global.v2.f32 {%f152, %f153}, [%rd138]; ld.global.v2.f32 {%f156, %f157}, [%rd138+8]; ld.local.f32 %f160, [%rd2+-8]; mov.f32 %f161, 0f3F800000; sub.ftz.f32 %f162, %f161, %f160; mul.ftz.f32 %f163, %f152, %f162; mul.ftz.f32 %f164, %f153, %f162; fma.rn.ftz.f32 %f281, %f157, %f160, %f164; fma.rn.ftz.f32 %f280, %f156, %f160, %f163; bra.uni $L__BB93_53; $L__BB93_17: setp.eq.s32 %p18, %r32, 0; @%p18 bra $L__BB93_28; bra.uni $L__BB93_18; $L__BB93_28: and.b32 %r75, %r31, 6; setp.eq.s32 %p25, %r75, 0; @%p25 bra $L__BB93_33; and.b32 %r76, %r21, 768; setp.eq.s32 %p26, %r76, 0; @%p26 bra $L__BB93_31; ld.const.u64 %rd90, [kernel_params+2264]; cvta.to.global.u64 %rd91, %rd90; shl.b64 %rd92, %rd9, 4; add.s64 %rd93, %rd91, %rd92; shr.s32 %r77, %r31, 7; ld.global.u32 %r78, [%rd93+4]; add.s32 %r79, %r77, %r78; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; add.s32 %r80, %r79, %r134; mul.wide.s32 %rd96, %r80, 12; add.s64 %rd97, %rd95, %rd96; ld.local.f32 %f109, [%rd2+-8]; mov.f32 %f110, 0f3F800000; sub.ftz.f32 %f111, %f110, %f109; ld.global.f32 %f112, [%rd97]; mul.ftz.f32 %f113, %f112, %f111; ld.global.f32 %f114, [%rd97+4]; mul.ftz.f32 %f115, %f114, %f111; ld.global.f32 %f116, [%rd97+8]; mul.ftz.f32 %f117, %f116, %f111; ld.global.f32 %f118, [%rd97+12]; ld.global.f32 %f119, [%rd97+16]; ld.global.f32 %f120, [%rd97+20]; fma.rn.ftz.f32 %f282, %f120, %f109, %f117; fma.rn.ftz.f32 %f281, %f119, %f109, %f115; fma.rn.ftz.f32 %f280, %f118, %f109, %f113; bra.uni $L__BB93_54; $L__BB93_36: cvt.u32.u64 %r88, %rd9; setp.eq.s32 %p33, %r88, -1; @%p33 bra $L__BB93_39; ld.const.u64 %rd106, [kernel_params+2240]; cvta.to.global.u64 %rd107, %rd106; shl.b64 %rd108, %rd9, 2; add.s64 %rd109, %rd107, %rd108; ld.global.u32 %r89, [%rd109]; setp.eq.s32 %p34, %r89, -1; @%p34 bra $L__BB93_39; mov.u64 %rd110, 0; { // callseq 67, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd49; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r21; st.param.b32 [param1+4], %r132; st.param.b32 [param1+8], %r133; st.param.b32 [param1+12], %r134; .param .b64 param2; st.param.b64 [param2+0], %rd110; .param .b64 param3; st.param.b64 [param3+0], %rd110; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f280, [retval0+0]; ld.param.f32 %f281, [retval0+4]; } // callseq 67 bra.uni $L__BB93_53; $L__BB93_18: cvt.u32.u64 %r53, %rd9; setp.eq.s32 %p19, %r53, -1; @%p19 bra $L__BB93_21; ld.const.u64 %rd67, [kernel_params+2240]; cvta.to.global.u64 %rd68, %rd67; shl.b64 %rd69, %rd9, 2; add.s64 %rd70, %rd68, %rd69; ld.global.u32 %r54, [%rd70]; setp.eq.s32 %p20, %r54, -1; @%p20 bra $L__BB93_21; mov.u64 %rd71, 0; { // callseq 66, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd49; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r21; st.param.b32 [param1+4], %r132; st.param.b32 [param1+8], %r133; st.param.b32 [param1+12], %r134; .param .b64 param2; st.param.b64 [param2+0], %rd71; .param .b64 param3; st.param.b64 [param3+0], %rd71; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f280, [retval0+0]; ld.param.f32 %f281, [retval0+4]; ld.param.f32 %f282, [retval0+8]; } // callseq 66 bra.uni $L__BB93_54; $L__BB93_39: and.b32 %r90, %r21, 56; setp.eq.s32 %p35, %r90, 0; @%p35 bra $L__BB93_44; and.b32 %r91, %r21, 24; setp.eq.s32 %p36, %r91, 0; @%p36 bra $L__BB93_42; ld.const.u64 %rd111, [kernel_params+2232]; cvta.to.global.u64 %rd112, %rd111; shl.b64 %rd113, %rd9, 4; add.s64 %rd114, %rd112, %rd113; ld.global.v4.u32 {%r92, %r93, %r94, %r95}, [%rd114]; ld.const.u64 %rd115, [kernel_params+2328]; cvta.to.global.u64 %rd116, %rd115; add.s32 %r99, %r92, %r134; mul.wide.u32 %rd117, %r99, 8; add.s64 %rd118, %rd116, %rd117; ld.global.v2.f32 {%f287, %f288}, [%rd118]; add.s32 %r100, %r93, %r134; mul.wide.u32 %rd119, %r100, 8; add.s64 %rd120, %rd116, %rd119; ld.global.v2.f32 {%f285, %f286}, [%rd120]; add.s32 %r101, %r94, %r134; mul.wide.u32 %rd121, %r101, 8; add.s64 %rd122, %rd116, %rd121; ld.global.v2.f32 {%f283, %f284}, [%rd122]; bra.uni $L__BB93_43; $L__BB93_21: and.b32 %r55, %r21, 56; setp.eq.s32 %p21, %r55, 0; @%p21 bra $L__BB93_26; and.b32 %r56, %r21, 24; setp.eq.s32 %p22, %r56, 0; @%p22 bra $L__BB93_24; ld.const.u64 %rd72, [kernel_params+2232]; cvta.to.global.u64 %rd73, %rd72; shl.b64 %rd74, %rd9, 4; add.s64 %rd75, %rd73, %rd74; ld.global.v4.u32 {%r57, %r58, %r59, %r60}, [%rd75]; ld.const.u64 %rd76, [kernel_params+2336]; cvta.to.global.u64 %rd77, %rd76; add.s32 %r64, %r57, %r134; mul.wide.u32 %rd78, %r64, 12; add.s64 %rd164, %rd77, %rd78; add.s64 %rd165, %rd164, 4; add.s64 %rd166, %rd164, 8; add.s32 %r65, %r58, %r134; mul.wide.u32 %rd79, %r65, 12; add.s64 %rd161, %rd77, %rd79; add.s64 %rd162, %rd161, 4; add.s64 %rd163, %rd161, 8; add.s32 %r66, %r59, %r134; mul.wide.u32 %rd80, %r66, 12; add.s64 %rd158, %rd77, %rd80; add.s64 %rd159, %rd158, 4; add.s64 %rd160, %rd158, 8; bra.uni $L__BB93_25; $L__BB93_51: and.b32 %r118, %r31, 8; setp.eq.s32 %p43, %r118, 0; setp.ne.s32 %p44, %r21, 8; mov.f32 %f280, 0f00000000; or.pred %p45, %p44, %p43; mov.f32 %f281, %f280; @%p45 bra $L__BB93_53; cvt.u32.u64 %r119, %rd9; ld.const.u64 %rd143, [kernel_params+2328]; cvta.to.global.u64 %rd144, %rd143; add.s32 %r120, %r119, %r134; mul.wide.s32 %rd145, %r120, 8; add.s64 %rd146, %rd144, %rd145; ld.global.v2.f32 {%f280, %f281}, [%rd146]; bra.uni $L__BB93_53; $L__BB93_33: and.b32 %r85, %r31, 8; setp.eq.s32 %p29, %r85, 0; setp.ne.s32 %p30, %r21, 8; mov.f32 %f280, 0f00000000; or.pred %p31, %p30, %p29; mov.f32 %f281, %f280; mov.f32 %f282, %f280; @%p31 bra $L__BB93_54; cvt.u32.u64 %r86, %rd9; ld.const.u64 %rd102, [kernel_params+2336]; cvta.to.global.u64 %rd103, %rd102; add.s32 %r87, %r86, %r134; mul.wide.s32 %rd104, %r87, 12; add.s64 %rd105, %rd103, %rd104; ld.global.f32 %f282, [%rd105+8]; ld.global.f32 %f281, [%rd105+4]; ld.global.f32 %f280, [%rd105]; bra.uni $L__BB93_54; $L__BB93_49: and.b32 %r114, %r21, 131; setp.eq.s32 %p41, %r114, 0; mov.f32 %f280, 0f00000000; mov.f32 %f281, %f280; @%p41 bra $L__BB93_53; cvt.u32.u64 %r115, %rd9; setp.eq.s32 %p42, %r21, 128; selp.b32 %r116, %r115, 0, %p42; add.s32 %r117, %r134, %r116; ld.const.u64 %rd139, [kernel_params+2328]; cvta.to.global.u64 %rd140, %rd139; mul.wide.s32 %rd141, %r117, 8; add.s64 %rd142, %rd140, %rd141; ld.global.v2.f32 {%f280, %f281}, [%rd142]; bra.uni $L__BB93_53; $L__BB93_31: and.b32 %r81, %r21, 131; setp.eq.s32 %p27, %r81, 0; mov.f32 %f280, 0f00000000; mov.f32 %f281, %f280; mov.f32 %f282, %f280; @%p27 bra $L__BB93_54; cvt.u32.u64 %r82, %rd9; setp.eq.s32 %p28, %r21, 128; selp.b32 %r83, %r82, 0, %p28; add.s32 %r84, %r134, %r83; ld.const.u64 %rd98, [kernel_params+2336]; cvta.to.global.u64 %rd99, %rd98; mul.wide.s32 %rd100, %r84, 12; add.s64 %rd101, %rd99, %rd100; ld.global.f32 %f282, [%rd101+8]; ld.global.f32 %f281, [%rd101+4]; ld.global.f32 %f280, [%rd101]; bra.uni $L__BB93_54; $L__BB93_44: and.b32 %r104, %r21, 7; setp.eq.s32 %p37, %r104, 0; mov.f32 %f280, 0f00000000; mov.f32 %f281, %f280; @%p37 bra $L__BB93_53; setp.eq.s32 %p38, %r21, 4; selp.b32 %r106, %r88, 0, %p38; add.s32 %r107, %r134, %r106; ld.const.u64 %rd127, [kernel_params+2328]; cvta.to.global.u64 %rd128, %rd127; mul.wide.s32 %rd129, %r107, 8; add.s64 %rd130, %rd128, %rd129; ld.global.v2.f32 {%f280, %f281}, [%rd130]; bra.uni $L__BB93_53; $L__BB93_26: and.b32 %r71, %r21, 7; setp.eq.s32 %p23, %r71, 0; mov.f32 %f280, 0f00000000; mov.f32 %f281, %f280; mov.f32 %f282, %f280; @%p23 bra $L__BB93_54; setp.eq.s32 %p24, %r21, 4; selp.b32 %r73, %r53, 0, %p24; add.s32 %r74, %r134, %r73; ld.const.u64 %rd86, [kernel_params+2336]; cvta.to.global.u64 %rd87, %rd86; mul.wide.s32 %rd88, %r74, 12; add.s64 %rd89, %rd87, %rd88; ld.global.f32 %f282, [%rd89+8]; ld.global.f32 %f281, [%rd89+4]; ld.global.f32 %f280, [%rd89]; bra.uni $L__BB93_54; $L__BB93_42: mad.lo.s32 %r103, %r88, 3, %r134; ld.const.u64 %rd123, [kernel_params+2328]; cvta.to.global.u64 %rd124, %rd123; mul.wide.s32 %rd125, %r103, 8; add.s64 %rd126, %rd124, %rd125; ld.global.v2.f32 {%f287, %f288}, [%rd126]; ld.global.v2.f32 {%f285, %f286}, [%rd126+8]; ld.global.v2.f32 {%f283, %f284}, [%rd126+16]; $L__BB93_43: ld.local.f32 %f139, [%rd2+-8]; ld.local.f32 %f140, [%rd2+-4]; mul.ftz.f32 %f141, %f283, %f140; mul.ftz.f32 %f142, %f284, %f140; fma.rn.ftz.f32 %f143, %f285, %f139, %f141; fma.rn.ftz.f32 %f144, %f286, %f139, %f142; mov.f32 %f145, 0f3F800000; sub.ftz.f32 %f146, %f145, %f139; sub.ftz.f32 %f147, %f146, %f140; fma.rn.ftz.f32 %f281, %f288, %f147, %f144; fma.rn.ftz.f32 %f280, %f287, %f147, %f143; $L__BB93_53: mov.f32 %f282, 0f00000000; $L__BB93_54: cvta.to.local.u64 %rd147, %rd50; shl.b32 %r121, %r1, 2; cvt.u64.u32 %rd148, %r121; and.b64 %rd149, %rd148, 1020; add.s64 %rd48, %rd147, %rd149; and.b32 %r122, %r1, 65280; setp.eq.s32 %p46, %r122, 256; @%p46 bra $L__BB93_62; bra.uni $L__BB93_55; $L__BB93_62: @%p16 bra $L__BB93_67; bra.uni $L__BB93_63; $L__BB93_67: mov.u32 %r124, 0; st.local.u32 [%rd48], %r124; st.local.u32 [%rd48+4], %r124; st.local.u32 [%rd48+8], %r124; bra.uni $L__BB93_68; $L__BB93_55: setp.ne.s32 %p47, %r134, -1; @%p47 bra $L__BB93_57; ld.local.v4.f32 {%f280, %f281, %f282, %f177}, [%rd51]; $L__BB93_57: setp.eq.s32 %p48, %r2, 0; @%p48 bra $L__BB93_61; setp.eq.s32 %p49, %r2, 1; add.ftz.f32 %f68, %f280, 0fBF000000; @%p49 bra $L__BB93_60; bra.uni $L__BB93_59; $L__BB93_60: add.ftz.f32 %f181, %f282, 0fBF000000; neg.ftz.f32 %f280, %f181; mov.f32 %f281, 0f00000000; mov.f32 %f282, %f68; bra.uni $L__BB93_63; $L__BB93_61: add.ftz.f32 %f183, %f282, 0fBF000000; neg.ftz.f32 %f71, %f183; add.ftz.f32 %f282, %f281, 0fBF000000; mov.f32 %f280, 0f00000000; mov.f32 %f281, %f71; bra.uni $L__BB93_63; $L__BB93_59: add.ftz.f32 %f179, %f281, 0fBF000000; neg.ftz.f32 %f280, %f179; mov.f32 %f282, 0f00000000; mov.f32 %f281, %f68; $L__BB93_63: ld.local.u8 %rs8, [%rd2+-20]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p51, %rs9, 0; @%p51 bra $L__BB93_65; bra.uni $L__BB93_64; $L__BB93_65: ld.local.u32 %r123, [%rd2]; ld.const.u64 %rd151, [kernel_params+2160]; cvta.to.global.u64 %rd152, %rd151; mul.wide.s32 %rd153, %r123, 208; add.s64 %rd154, %rd152, %rd153; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd154+48]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd154+64]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd154+80]; mul.ftz.f32 %f239, %f281, %f225; fma.rn.ftz.f32 %f240, %f280, %f218, %f239; fma.rn.ftz.f32 %f241, %f282, %f232, %f240; mul.ftz.f32 %f242, %f281, %f226; fma.rn.ftz.f32 %f243, %f280, %f219, %f242; fma.rn.ftz.f32 %f244, %f282, %f233, %f243; mul.ftz.f32 %f245, %f281, %f227; fma.rn.ftz.f32 %f246, %f280, %f220, %f245; fma.rn.ftz.f32 %f247, %f282, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f302, %f241, %f251; div.approx.ftz.f32 %f301, %f244, %f251; div.approx.ftz.f32 %f300, %f247, %f251; bra.uni $L__BB93_66; $L__BB93_64: ld.local.v4.f32 {%f184, %f185, %f186, %f187}, [%rd2+116]; ld.local.v4.f32 {%f191, %f192, %f193, %f194}, [%rd2+132]; ld.local.v4.f32 {%f198, %f199, %f200, %f201}, [%rd2+148]; mul.ftz.f32 %f205, %f281, %f191; fma.rn.ftz.f32 %f206, %f280, %f184, %f205; fma.rn.ftz.f32 %f207, %f282, %f198, %f206; mul.ftz.f32 %f208, %f281, %f192; fma.rn.ftz.f32 %f209, %f280, %f185, %f208; fma.rn.ftz.f32 %f210, %f282, %f199, %f209; mul.ftz.f32 %f211, %f281, %f193; fma.rn.ftz.f32 %f212, %f280, %f186, %f211; fma.rn.ftz.f32 %f213, %f282, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f302, %f207, %f217; div.approx.ftz.f32 %f301, %f210, %f217; div.approx.ftz.f32 %f300, %f213, %f217; $L__BB93_66: ld.local.f32 %f252, [%rd2+-56]; mul.ftz.f32 %f253, %f301, %f252; ld.local.f32 %f254, [%rd2+-60]; mul.ftz.f32 %f255, %f300, %f254; sub.ftz.f32 %f256, %f253, %f255; ld.local.f32 %f257, [%rd2+-64]; mul.ftz.f32 %f258, %f300, %f257; mul.ftz.f32 %f259, %f302, %f252; sub.ftz.f32 %f260, %f258, %f259; mul.ftz.f32 %f261, %f302, %f254; mul.ftz.f32 %f262, %f301, %f257; sub.ftz.f32 %f263, %f261, %f262; mul.ftz.f32 %f264, %f260, %f260; fma.rn.ftz.f32 %f265, %f256, %f256, %f264; fma.rn.ftz.f32 %f266, %f263, %f263, %f265; sqrt.approx.ftz.f32 %f267, %f266; div.approx.ftz.f32 %f268, %f256, %f267; div.approx.ftz.f32 %f269, %f260, %f267; div.approx.ftz.f32 %f270, %f263, %f267; mul.ftz.f32 %f271, %f270, %f254; mul.ftz.f32 %f272, %f269, %f252; sub.ftz.f32 %f273, %f271, %f272; mul.ftz.f32 %f274, %f268, %f252; mul.ftz.f32 %f275, %f270, %f257; sub.ftz.f32 %f276, %f274, %f275; mul.ftz.f32 %f277, %f269, %f257; mul.ftz.f32 %f278, %f268, %f254; sub.ftz.f32 %f279, %f277, %f278; st.local.f32 [%rd48], %f273; st.local.f32 [%rd48+4], %f276; st.local.f32 [%rd48+8], %f279; $L__BB93_68: ret; $L__BB93_24: mad.lo.s32 %r68, %r53, 3, %r134; ld.const.u64 %rd81, [kernel_params+2336]; cvta.to.global.u64 %rd82, %rd81; mul.wide.s32 %rd83, %r68, 12; add.s64 %rd164, %rd82, %rd83; add.s64 %rd165, %rd164, 4; add.s64 %rd166, %rd164, 8; add.s32 %r69, %r68, 1; mul.wide.s32 %rd84, %r69, 12; add.s64 %rd161, %rd82, %rd84; add.s64 %rd162, %rd161, 4; add.s64 %rd163, %rd161, 8; add.s32 %r70, %r68, 2; mul.wide.s32 %rd85, %r70, 12; add.s64 %rd158, %rd82, %rd85; add.s64 %rd159, %rd158, 4; add.s64 %rd160, %rd158, 8; $L__BB93_25: ld.local.f32 %f86, [%rd2+-8]; ld.global.f32 %f87, [%rd161]; ld.global.f32 %f88, [%rd162]; ld.global.f32 %f89, [%rd163]; ld.local.f32 %f90, [%rd2+-4]; ld.global.f32 %f91, [%rd158]; mul.ftz.f32 %f92, %f91, %f90; ld.global.f32 %f93, [%rd159]; mul.ftz.f32 %f94, %f93, %f90; ld.global.f32 %f95, [%rd160]; mul.ftz.f32 %f96, %f95, %f90; fma.rn.ftz.f32 %f97, %f87, %f86, %f92; fma.rn.ftz.f32 %f98, %f88, %f86, %f94; fma.rn.ftz.f32 %f99, %f89, %f86, %f96; mov.f32 %f100, 0f3F800000; sub.ftz.f32 %f101, %f100, %f86; sub.ftz.f32 %f102, %f101, %f90; ld.global.f32 %f103, [%rd164]; ld.global.f32 %f104, [%rd165]; ld.global.f32 %f105, [%rd166]; fma.rn.ftz.f32 %f282, %f105, %f102, %f99; fma.rn.ftz.f32 %f281, %f104, %f102, %f98; fma.rn.ftz.f32 %f280, %f103, %f102, %f97; bra.uni $L__BB93_54; } .func _Z22svm_node_vector_rotateP10ShaderDataPfjjj( .param .b64 _Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_0, .param .b32 _Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_1, .param .b32 _Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_2, .param .b32 _Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<9>; .reg .b16 %rs<3>; .reg .f32 %f<138>; .reg .b32 %r<11>; .reg .b64 %rd<20>; ld.param.u64 %rd2, [_Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_1]; ld.param.u32 %r3, [_Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_2]; ld.param.u32 %r4, [_Z22svm_node_vector_rotateP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r2, 24; setp.eq.s32 %p1, %r4, 255; @%p1 bra $L__BB94_14; and.b32 %r5, %r2, 255; shr.u32 %r6, %r2, 6; cvt.u64.u32 %rd3, %r6; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f135, [%rd5]; ld.local.f32 %f136, [%rd5+4]; ld.local.f32 %f137, [%rd5+8]; shl.b32 %r7, %r3, 2; cvt.u64.u32 %rd6, %r7; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f4, [%rd8]; ld.local.f32 %f5, [%rd8+4]; ld.local.f32 %f6, [%rd8+8]; setp.eq.s32 %p2, %r5, 4; @%p2 bra $L__BB94_10; bra.uni $L__BB94_2; $L__BB94_10: shr.u32 %r10, %r2, 14; cvt.u64.u32 %rd15, %r10; and.b64 %rd16, %rd15, 1020; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f97, [%rd17]; cos.approx.ftz.f32 %f98, %f97; ld.local.f32 %f99, [%rd17+4]; cos.approx.ftz.f32 %f100, %f99; ld.local.f32 %f101, [%rd17+8]; cos.approx.ftz.f32 %f102, %f101; sin.approx.ftz.f32 %f103, %f97; sin.approx.ftz.f32 %f104, %f99; sin.approx.ftz.f32 %f105, %f101; mul.ftz.f32 %f106, %f100, %f102; mul.ftz.f32 %f21, %f100, %f105; neg.ftz.f32 %f22, %f104; mul.ftz.f32 %f107, %f103, %f104; mul.ftz.f32 %f108, %f102, %f107; mul.ftz.f32 %f109, %f98, %f105; sub.ftz.f32 %f23, %f108, %f109; mul.ftz.f32 %f110, %f107, %f105; fma.rn.ftz.f32 %f111, %f98, %f102, %f110; mul.ftz.f32 %f24, %f100, %f103; mul.ftz.f32 %f112, %f98, %f104; mul.ftz.f32 %f113, %f103, %f105; fma.rn.ftz.f32 %f25, %f102, %f112, %f113; mul.ftz.f32 %f114, %f112, %f105; mul.ftz.f32 %f115, %f102, %f103; sub.ftz.f32 %f26, %f114, %f115; mul.ftz.f32 %f116, %f98, %f100; sub.ftz.f32 %f27, %f135, %f4; mul.ftz.f32 %f28, %f27, %f106; sub.ftz.f32 %f29, %f136, %f5; mul.ftz.f32 %f30, %f29, %f111; sub.ftz.f32 %f31, %f137, %f6; mul.ftz.f32 %f32, %f31, %f116; setp.eq.s32 %p8, %r1, 0; @%p8 bra $L__BB94_12; fma.rn.ftz.f32 %f117, %f29, %f21, %f28; fma.rn.ftz.f32 %f118, %f31, %f22, %f117; fma.rn.ftz.f32 %f119, %f27, %f23, %f30; fma.rn.ftz.f32 %f120, %f31, %f24, %f119; mul.ftz.f32 %f121, %f29, %f26; fma.rn.ftz.f32 %f122, %f27, %f25, %f121; add.ftz.f32 %f123, %f32, %f122; add.ftz.f32 %f136, %f5, %f120; add.ftz.f32 %f135, %f4, %f118; add.ftz.f32 %f137, %f6, %f123; bra.uni $L__BB94_13; $L__BB94_2: cvt.u16.u32 %rs2, %r2; and.b16 %rs1, %rs2, 255; mov.f32 %f133, 0f3F800000; mov.f32 %f131, 0f00000000; setp.eq.s16 %p3, %rs1, 1; mov.f32 %f132, %f131; mov.f32 %f134, %f133; @%p3 bra $L__BB94_8; setp.eq.s16 %p4, %rs1, 2; @%p4 bra $L__BB94_6; setp.ne.s16 %p5, %rs1, 3; @%p5 bra $L__BB94_7; mov.f32 %f131, 0f3F800000; mov.f32 %f132, 0f00000000; mov.f32 %f133, %f132; mov.f32 %f134, %f131; bra.uni $L__BB94_8; $L__BB94_12: fma.rn.ftz.f32 %f124, %f29, %f23, %f28; fma.rn.ftz.f32 %f125, %f31, %f25, %f124; fma.rn.ftz.f32 %f126, %f27, %f21, %f30; fma.rn.ftz.f32 %f127, %f31, %f26, %f126; mul.ftz.f32 %f128, %f29, %f24; fma.rn.ftz.f32 %f129, %f27, %f22, %f128; add.ftz.f32 %f130, %f32, %f129; add.ftz.f32 %f136, %f5, %f127; add.ftz.f32 %f135, %f4, %f125; add.ftz.f32 %f137, %f6, %f130; bra.uni $L__BB94_13; $L__BB94_6: mov.f32 %f132, 0f3F800000; mov.f32 %f131, 0f00000000; mov.f32 %f133, %f131; mov.f32 %f134, %f132; bra.uni $L__BB94_8; $L__BB94_7: shr.u32 %r8, %r3, 6; cvt.u64.u32 %rd9, %r8; and.b64 %rd10, %rd9, 1020; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f133, [%rd11]; ld.local.f32 %f132, [%rd11+4]; mul.ftz.f32 %f54, %f132, %f132; fma.rn.ftz.f32 %f55, %f133, %f133, %f54; ld.local.f32 %f131, [%rd11+8]; fma.rn.ftz.f32 %f56, %f131, %f131, %f55; sqrt.approx.ftz.f32 %f134, %f56; $L__BB94_8: setp.neu.ftz.f32 %p6, %f134, 0f00000000; @%p6 bra $L__BB94_9; bra.uni $L__BB94_13; $L__BB94_9: shr.u32 %r9, %r3, 14; cvt.u64.u32 %rd12, %r9; and.b64 %rd13, %rd12, 1020; add.s64 %rd14, %rd1, %rd13; ld.local.f32 %f57, [%rd14]; setp.eq.s32 %p7, %r1, 0; neg.ftz.f32 %f58, %f57; selp.f32 %f59, %f57, %f58, %p7; sub.ftz.f32 %f60, %f135, %f4; cos.approx.ftz.f32 %f61, %f59; mov.f32 %f62, 0f3F800000; sub.ftz.f32 %f63, %f62, %f61; div.approx.ftz.f32 %f64, %f133, %f134; mul.ftz.f32 %f65, %f64, %f63; fma.rn.ftz.f32 %f66, %f64, %f65, %f61; div.approx.ftz.f32 %f67, %f132, %f134; mul.ftz.f32 %f68, %f67, %f65; sin.approx.ftz.f32 %f69, %f59; div.approx.ftz.f32 %f70, %f131, %f134; mul.ftz.f32 %f71, %f70, %f69; sub.ftz.f32 %f72, %f68, %f71; sub.ftz.f32 %f73, %f136, %f5; mul.ftz.f32 %f74, %f73, %f72; fma.rn.ftz.f32 %f75, %f60, %f66, %f74; mul.ftz.f32 %f76, %f70, %f65; mul.ftz.f32 %f77, %f67, %f69; add.ftz.f32 %f78, %f77, %f76; sub.ftz.f32 %f79, %f137, %f6; fma.rn.ftz.f32 %f80, %f79, %f78, %f75; add.ftz.f32 %f81, %f71, %f68; mul.ftz.f32 %f82, %f67, %f63; fma.rn.ftz.f32 %f83, %f67, %f82, %f61; mul.ftz.f32 %f84, %f73, %f83; fma.rn.ftz.f32 %f85, %f60, %f81, %f84; mul.ftz.f32 %f86, %f70, %f82; mul.ftz.f32 %f87, %f64, %f69; sub.ftz.f32 %f88, %f86, %f87; fma.rn.ftz.f32 %f89, %f79, %f88, %f85; sub.ftz.f32 %f90, %f76, %f77; add.ftz.f32 %f91, %f87, %f86; mul.ftz.f32 %f92, %f73, %f91; fma.rn.ftz.f32 %f93, %f60, %f90, %f92; mul.ftz.f32 %f94, %f70, %f63; fma.rn.ftz.f32 %f95, %f70, %f94, %f61; fma.rn.ftz.f32 %f96, %f79, %f95, %f93; add.ftz.f32 %f136, %f5, %f89; add.ftz.f32 %f137, %f6, %f96; add.ftz.f32 %f135, %f4, %f80; $L__BB94_13: mul.wide.u32 %rd18, %r4, 4; add.s64 %rd19, %rd1, %rd18; st.local.f32 [%rd19], %f135; st.local.f32 [%rd19+4], %f136; st.local.f32 [%rd19+8], %f137; $L__BB94_14: ret; } .func _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<30>; .reg .f32 %f<539>; .reg .b32 %r<20>; .reg .b64 %rd<36>; ld.param.u64 %rd4, [_Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r9, [_Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r10, [_Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; and.b32 %r1, %r10, 255; bfe.u32 %r2, %r10, 8, 8; bfe.u32 %r3, %r10, 16, 8; bfe.u32 %r4, %r9, 8, 8; cvta.to.local.u64 %rd1, %rd5; shl.b32 %r11, %r9, 2; cvt.u64.u32 %rd7, %r11; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f535, [%rd9]; ld.local.f32 %f534, [%rd9+4]; ld.local.f32 %f533, [%rd9+8]; add.s64 %rd2, %rd6, 76; ld.local.s32 %rd3, [%rd6+76]; or.b32 %r5, %r1, 2; setp.eq.s32 %p1, %r2, 0; @%p1 bra $L__BB95_25; setp.eq.s32 %p2, %r2, 2; @%p2 bra $L__BB95_14; bra.uni $L__BB95_2; $L__BB95_14: setp.gt.u32 %p12, %r3, 1; @%p12 bra $L__BB95_17; setp.eq.s32 %p13, %r5, 2; ld.const.v4.f32 {%f247, %f248, %f249, %f250}, [kernel_params+96]; ld.const.v4.f32 {%f251, %f252, %f253, %f254}, [kernel_params+112]; ld.const.v4.f32 {%f255, %f256, %f257, %f258}, [kernel_params+128]; mul.ftz.f32 %f259, %f534, %f248; fma.rn.ftz.f32 %f260, %f535, %f247, %f259; fma.rn.ftz.f32 %f49, %f533, %f249, %f260; mul.ftz.f32 %f261, %f534, %f252; fma.rn.ftz.f32 %f262, %f535, %f251, %f261; fma.rn.ftz.f32 %f50, %f533, %f253, %f262; mul.ftz.f32 %f263, %f534, %f256; fma.rn.ftz.f32 %f264, %f535, %f255, %f263; fma.rn.ftz.f32 %f533, %f533, %f257, %f264; mov.f32 %f534, %f50; mov.f32 %f535, %f49; @%p13 bra $L__BB95_17; add.ftz.f32 %f535, %f250, %f49; add.ftz.f32 %f534, %f254, %f50; add.ftz.f32 %f533, %f258, %f533; $L__BB95_17: cvt.u32.u64 %r16, %rd3; setp.eq.s32 %p14, %r16, -1; setp.ne.s32 %p15, %r3, 1; or.pred %p16, %p15, %p14; @%p16 bra $L__BB95_36; setp.eq.s32 %p17, %r5, 2; ld.local.u32 %r17, [%rd2+-20]; and.b32 %r7, %r17, 2; @%p17 bra $L__BB95_22; bra.uni $L__BB95_19; $L__BB95_22: setp.eq.s32 %p19, %r7, 0; @%p19 bra $L__BB95_24; bra.uni $L__BB95_23; $L__BB95_24: ld.const.u64 %rd22, [kernel_params+2160]; cvta.to.global.u64 %rd23, %rd22; mul.lo.s64 %rd24, %rd3, 208; add.s64 %rd25, %rd23, %rd24; ld.global.v4.f32 {%f358, %f359, %f360, %f361}, [%rd25+48]; ld.global.v4.f32 {%f365, %f366, %f367, %f368}, [%rd25+64]; ld.global.v4.f32 {%f372, %f373, %f374, %f375}, [%rd25+80]; mul.ftz.f32 %f379, %f534, %f359; fma.rn.ftz.f32 %f380, %f535, %f358, %f379; fma.rn.ftz.f32 %f67, %f533, %f360, %f380; mul.ftz.f32 %f381, %f534, %f366; fma.rn.ftz.f32 %f382, %f535, %f365, %f381; fma.rn.ftz.f32 %f68, %f533, %f367, %f382; mul.ftz.f32 %f383, %f534, %f373; fma.rn.ftz.f32 %f384, %f535, %f372, %f383; fma.rn.ftz.f32 %f533, %f533, %f374, %f384; mov.f32 %f534, %f68; mov.f32 %f535, %f67; bra.uni $L__BB95_36; $L__BB95_25: setp.eq.s32 %p20, %r3, 2; @%p20 bra $L__BB95_34; bra.uni $L__BB95_26; $L__BB95_34: setp.eq.s32 %p27, %r5, 2; ld.const.v4.f32 {%f505, %f506, %f507, %f508}, [kernel_params+688]; ld.const.v4.f32 {%f509, %f510, %f511, %f512}, [kernel_params+704]; ld.const.v4.f32 {%f513, %f514, %f515, %f516}, [kernel_params+720]; mul.ftz.f32 %f517, %f534, %f506; fma.rn.ftz.f32 %f518, %f535, %f505, %f517; fma.rn.ftz.f32 %f94, %f533, %f507, %f518; mul.ftz.f32 %f519, %f534, %f510; fma.rn.ftz.f32 %f520, %f535, %f509, %f519; fma.rn.ftz.f32 %f95, %f533, %f511, %f520; mul.ftz.f32 %f521, %f534, %f514; fma.rn.ftz.f32 %f522, %f535, %f513, %f521; fma.rn.ftz.f32 %f533, %f533, %f515, %f522; mov.f32 %f534, %f95; mov.f32 %f535, %f94; @%p27 bra $L__BB95_36; add.ftz.f32 %f535, %f508, %f94; add.ftz.f32 %f534, %f512, %f95; add.ftz.f32 %f533, %f516, %f533; bra.uni $L__BB95_36; $L__BB95_2: setp.ne.s32 %p3, %r2, 1; @%p3 bra $L__BB95_36; cvt.u32.u64 %r12, %rd3; setp.eq.s32 %p4, %r12, -1; or.b32 %r13, %r3, 2; setp.ne.s32 %p5, %r13, 2; or.pred %p6, %p5, %p4; @%p6 bra $L__BB95_11; setp.eq.s32 %p7, %r5, 2; ld.local.u32 %r14, [%rd2+-20]; and.b32 %r6, %r14, 2; @%p7 bra $L__BB95_8; bra.uni $L__BB95_5; $L__BB95_8: setp.eq.s32 %p9, %r6, 0; @%p9 bra $L__BB95_10; bra.uni $L__BB95_9; $L__BB95_10: ld.const.u64 %rd14, [kernel_params+2160]; cvta.to.global.u64 %rd15, %rd14; mul.wide.s32 %rd16, %r12, 208; add.s64 %rd17, %rd15, %rd16; ld.global.v4.f32 {%f202, %f203, %f204, %f205}, [%rd17]; ld.global.v4.f32 {%f209, %f210, %f211, %f212}, [%rd17+16]; ld.global.v4.f32 {%f216, %f217, %f218, %f219}, [%rd17+32]; mul.ftz.f32 %f223, %f534, %f203; fma.rn.ftz.f32 %f224, %f535, %f202, %f223; fma.rn.ftz.f32 %f13, %f533, %f204, %f224; mul.ftz.f32 %f225, %f534, %f210; fma.rn.ftz.f32 %f226, %f535, %f209, %f225; fma.rn.ftz.f32 %f14, %f533, %f211, %f226; mul.ftz.f32 %f227, %f534, %f217; fma.rn.ftz.f32 %f228, %f535, %f216, %f227; fma.rn.ftz.f32 %f533, %f533, %f218, %f228; mov.f32 %f534, %f14; mov.f32 %f535, %f13; bra.uni $L__BB95_11; $L__BB95_26: cvt.u32.u64 %r18, %rd3; setp.eq.s32 %p21, %r18, -1; setp.ne.s32 %p22, %r3, 1; or.pred %p23, %p22, %p21; @%p23 bra $L__BB95_36; setp.eq.s32 %p24, %r5, 2; ld.local.u32 %r19, [%rd2+-20]; and.b32 %r8, %r19, 2; @%p24 bra $L__BB95_31; bra.uni $L__BB95_28; $L__BB95_31: setp.eq.s32 %p26, %r8, 0; @%p26 bra $L__BB95_33; bra.uni $L__BB95_32; $L__BB95_33: ld.const.u64 %rd30, [kernel_params+2160]; cvta.to.global.u64 %rd31, %rd30; mul.lo.s64 %rd32, %rd3, 208; add.s64 %rd33, %rd31, %rd32; ld.global.v4.f32 {%f478, %f479, %f480, %f481}, [%rd33+48]; ld.global.v4.f32 {%f485, %f486, %f487, %f488}, [%rd33+64]; ld.global.v4.f32 {%f492, %f493, %f494, %f495}, [%rd33+80]; mul.ftz.f32 %f499, %f534, %f479; fma.rn.ftz.f32 %f500, %f535, %f478, %f499; fma.rn.ftz.f32 %f79, %f533, %f480, %f500; mul.ftz.f32 %f501, %f534, %f486; fma.rn.ftz.f32 %f502, %f535, %f485, %f501; fma.rn.ftz.f32 %f80, %f533, %f487, %f502; mul.ftz.f32 %f503, %f534, %f493; fma.rn.ftz.f32 %f504, %f535, %f492, %f503; fma.rn.ftz.f32 %f533, %f533, %f494, %f504; mov.f32 %f534, %f80; mov.f32 %f535, %f79; bra.uni $L__BB95_36; $L__BB95_19: setp.eq.s32 %p18, %r7, 0; @%p18 bra $L__BB95_21; bra.uni $L__BB95_20; $L__BB95_21: ld.const.u64 %rd18, [kernel_params+2160]; cvta.to.global.u64 %rd19, %rd18; mul.lo.s64 %rd20, %rd3, 208; add.s64 %rd21, %rd19, %rd20; ld.global.v4.f32 {%f298, %f299, %f300, %f301}, [%rd21+48]; ld.global.v4.f32 {%f306, %f307, %f308, %f309}, [%rd21+64]; ld.global.v4.f32 {%f314, %f315, %f316, %f317}, [%rd21+80]; mul.ftz.f32 %f322, %f534, %f299; fma.rn.ftz.f32 %f323, %f535, %f298, %f322; fma.rn.ftz.f32 %f324, %f533, %f300, %f323; add.ftz.f32 %f61, %f301, %f324; mul.ftz.f32 %f325, %f534, %f307; fma.rn.ftz.f32 %f326, %f535, %f306, %f325; fma.rn.ftz.f32 %f327, %f533, %f308, %f326; add.ftz.f32 %f62, %f309, %f327; mul.ftz.f32 %f328, %f534, %f315; fma.rn.ftz.f32 %f329, %f535, %f314, %f328; fma.rn.ftz.f32 %f330, %f533, %f316, %f329; add.ftz.f32 %f533, %f317, %f330; mov.f32 %f534, %f62; mov.f32 %f535, %f61; bra.uni $L__BB95_36; $L__BB95_28: setp.eq.s32 %p25, %r8, 0; @%p25 bra $L__BB95_30; bra.uni $L__BB95_29; $L__BB95_30: ld.const.u64 %rd26, [kernel_params+2160]; cvta.to.global.u64 %rd27, %rd26; mul.lo.s64 %rd28, %rd3, 208; add.s64 %rd29, %rd27, %rd28; ld.global.v4.f32 {%f418, %f419, %f420, %f421}, [%rd29+48]; ld.global.v4.f32 {%f426, %f427, %f428, %f429}, [%rd29+64]; ld.global.v4.f32 {%f434, %f435, %f436, %f437}, [%rd29+80]; mul.ftz.f32 %f442, %f534, %f419; fma.rn.ftz.f32 %f443, %f535, %f418, %f442; fma.rn.ftz.f32 %f444, %f533, %f420, %f443; add.ftz.f32 %f73, %f421, %f444; mul.ftz.f32 %f445, %f534, %f427; fma.rn.ftz.f32 %f446, %f535, %f426, %f445; fma.rn.ftz.f32 %f447, %f533, %f428, %f446; add.ftz.f32 %f74, %f429, %f447; mul.ftz.f32 %f448, %f534, %f435; fma.rn.ftz.f32 %f449, %f535, %f434, %f448; fma.rn.ftz.f32 %f450, %f533, %f436, %f449; add.ftz.f32 %f533, %f437, %f450; mov.f32 %f534, %f74; mov.f32 %f535, %f73; bra.uni $L__BB95_36; $L__BB95_23: ld.local.v4.f32 {%f331, %f332, %f333, %f334}, [%rd2+116]; mul.ftz.f32 %f338, %f534, %f332; fma.rn.ftz.f32 %f339, %f535, %f331, %f338; fma.rn.ftz.f32 %f64, %f533, %f333, %f339; ld.local.v4.f32 {%f340, %f341, %f342, %f343}, [%rd2+132]; mul.ftz.f32 %f347, %f534, %f341; fma.rn.ftz.f32 %f348, %f535, %f340, %f347; fma.rn.ftz.f32 %f65, %f533, %f342, %f348; ld.local.v4.f32 {%f349, %f350, %f351, %f352}, [%rd2+148]; mul.ftz.f32 %f356, %f534, %f350; fma.rn.ftz.f32 %f357, %f535, %f349, %f356; fma.rn.ftz.f32 %f533, %f533, %f351, %f357; mov.f32 %f534, %f65; mov.f32 %f535, %f64; bra.uni $L__BB95_36; $L__BB95_5: setp.eq.s32 %p8, %r6, 0; @%p8 bra $L__BB95_7; bra.uni $L__BB95_6; $L__BB95_7: ld.const.u64 %rd10, [kernel_params+2160]; cvta.to.global.u64 %rd11, %rd10; mul.lo.s64 %rd12, %rd3, 208; add.s64 %rd13, %rd11, %rd12; ld.global.v4.f32 {%f142, %f143, %f144, %f145}, [%rd13]; ld.global.v4.f32 {%f150, %f151, %f152, %f153}, [%rd13+16]; ld.global.v4.f32 {%f158, %f159, %f160, %f161}, [%rd13+32]; mul.ftz.f32 %f166, %f534, %f143; fma.rn.ftz.f32 %f167, %f535, %f142, %f166; fma.rn.ftz.f32 %f168, %f533, %f144, %f167; add.ftz.f32 %f7, %f145, %f168; mul.ftz.f32 %f169, %f534, %f151; fma.rn.ftz.f32 %f170, %f535, %f150, %f169; fma.rn.ftz.f32 %f171, %f533, %f152, %f170; add.ftz.f32 %f8, %f153, %f171; mul.ftz.f32 %f172, %f534, %f159; fma.rn.ftz.f32 %f173, %f535, %f158, %f172; fma.rn.ftz.f32 %f174, %f533, %f160, %f173; add.ftz.f32 %f533, %f161, %f174; mov.f32 %f534, %f8; mov.f32 %f535, %f7; bra.uni $L__BB95_11; $L__BB95_20: ld.local.v4.f32 {%f265, %f266, %f267, %f268}, [%rd2+116]; mul.ftz.f32 %f273, %f534, %f266; fma.rn.ftz.f32 %f274, %f535, %f265, %f273; fma.rn.ftz.f32 %f275, %f533, %f267, %f274; add.ftz.f32 %f58, %f268, %f275; ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f284, %f534, %f277; fma.rn.ftz.f32 %f285, %f535, %f276, %f284; fma.rn.ftz.f32 %f286, %f533, %f278, %f285; add.ftz.f32 %f59, %f279, %f286; ld.local.v4.f32 {%f287, %f288, %f289, %f290}, [%rd2+148]; mul.ftz.f32 %f295, %f534, %f288; fma.rn.ftz.f32 %f296, %f535, %f287, %f295; fma.rn.ftz.f32 %f297, %f533, %f289, %f296; add.ftz.f32 %f533, %f290, %f297; mov.f32 %f534, %f59; mov.f32 %f535, %f58; bra.uni $L__BB95_36; $L__BB95_32: ld.local.v4.f32 {%f451, %f452, %f453, %f454}, [%rd2+116]; mul.ftz.f32 %f458, %f534, %f452; fma.rn.ftz.f32 %f459, %f535, %f451, %f458; fma.rn.ftz.f32 %f76, %f533, %f453, %f459; ld.local.v4.f32 {%f460, %f461, %f462, %f463}, [%rd2+132]; mul.ftz.f32 %f467, %f534, %f461; fma.rn.ftz.f32 %f468, %f535, %f460, %f467; fma.rn.ftz.f32 %f77, %f533, %f462, %f468; ld.local.v4.f32 {%f469, %f470, %f471, %f472}, [%rd2+148]; mul.ftz.f32 %f476, %f534, %f470; fma.rn.ftz.f32 %f477, %f535, %f469, %f476; fma.rn.ftz.f32 %f533, %f533, %f471, %f477; mov.f32 %f534, %f77; mov.f32 %f535, %f76; bra.uni $L__BB95_36; $L__BB95_29: ld.local.v4.f32 {%f385, %f386, %f387, %f388}, [%rd2+116]; mul.ftz.f32 %f393, %f534, %f386; fma.rn.ftz.f32 %f394, %f535, %f385, %f393; fma.rn.ftz.f32 %f395, %f533, %f387, %f394; add.ftz.f32 %f70, %f388, %f395; ld.local.v4.f32 {%f396, %f397, %f398, %f399}, [%rd2+132]; mul.ftz.f32 %f404, %f534, %f397; fma.rn.ftz.f32 %f405, %f535, %f396, %f404; fma.rn.ftz.f32 %f406, %f533, %f398, %f405; add.ftz.f32 %f71, %f399, %f406; ld.local.v4.f32 {%f407, %f408, %f409, %f410}, [%rd2+148]; mul.ftz.f32 %f415, %f534, %f408; fma.rn.ftz.f32 %f416, %f535, %f407, %f415; fma.rn.ftz.f32 %f417, %f533, %f409, %f416; add.ftz.f32 %f533, %f410, %f417; mov.f32 %f534, %f71; mov.f32 %f535, %f70; bra.uni $L__BB95_36; $L__BB95_9: ld.local.v4.f32 {%f175, %f176, %f177, %f178}, [%rd2+68]; mul.ftz.f32 %f182, %f534, %f176; fma.rn.ftz.f32 %f183, %f535, %f175, %f182; fma.rn.ftz.f32 %f10, %f533, %f177, %f183; ld.local.v4.f32 {%f184, %f185, %f186, %f187}, [%rd2+84]; mul.ftz.f32 %f191, %f534, %f185; fma.rn.ftz.f32 %f192, %f535, %f184, %f191; fma.rn.ftz.f32 %f11, %f533, %f186, %f192; ld.local.v4.f32 {%f193, %f194, %f195, %f196}, [%rd2+100]; mul.ftz.f32 %f200, %f534, %f194; fma.rn.ftz.f32 %f201, %f535, %f193, %f200; fma.rn.ftz.f32 %f533, %f533, %f195, %f201; mov.f32 %f534, %f11; mov.f32 %f535, %f10; bra.uni $L__BB95_11; $L__BB95_6: ld.local.v4.f32 {%f109, %f110, %f111, %f112}, [%rd2+68]; mul.ftz.f32 %f117, %f534, %f110; fma.rn.ftz.f32 %f118, %f535, %f109, %f117; fma.rn.ftz.f32 %f119, %f533, %f111, %f118; add.ftz.f32 %f4, %f112, %f119; ld.local.v4.f32 {%f120, %f121, %f122, %f123}, [%rd2+84]; mul.ftz.f32 %f128, %f534, %f121; fma.rn.ftz.f32 %f129, %f535, %f120, %f128; fma.rn.ftz.f32 %f130, %f533, %f122, %f129; add.ftz.f32 %f5, %f123, %f130; ld.local.v4.f32 {%f131, %f132, %f133, %f134}, [%rd2+100]; mul.ftz.f32 %f139, %f534, %f132; fma.rn.ftz.f32 %f140, %f535, %f131, %f139; fma.rn.ftz.f32 %f141, %f533, %f133, %f140; add.ftz.f32 %f533, %f134, %f141; mov.f32 %f534, %f5; mov.f32 %f535, %f4; $L__BB95_11: setp.ne.s32 %p10, %r3, 2; @%p10 bra $L__BB95_36; setp.eq.s32 %p11, %r5, 2; ld.const.v4.f32 {%f229, %f230, %f231, %f232}, [kernel_params+688]; ld.const.v4.f32 {%f233, %f234, %f235, %f236}, [kernel_params+704]; ld.const.v4.f32 {%f237, %f238, %f239, %f240}, [kernel_params+720]; mul.ftz.f32 %f241, %f534, %f230; fma.rn.ftz.f32 %f242, %f535, %f229, %f241; fma.rn.ftz.f32 %f31, %f533, %f231, %f242; mul.ftz.f32 %f243, %f534, %f234; fma.rn.ftz.f32 %f244, %f535, %f233, %f243; fma.rn.ftz.f32 %f32, %f533, %f235, %f244; mul.ftz.f32 %f245, %f534, %f238; fma.rn.ftz.f32 %f246, %f535, %f237, %f245; fma.rn.ftz.f32 %f533, %f533, %f239, %f246; mov.f32 %f534, %f32; mov.f32 %f535, %f31; @%p11 bra $L__BB95_36; add.ftz.f32 %f535, %f232, %f31; add.ftz.f32 %f534, %f236, %f32; add.ftz.f32 %f533, %f240, %f533; $L__BB95_36: setp.ne.s32 %p28, %r1, 2; @%p28 bra $L__BB95_38; mul.ftz.f32 %f523, %f535, %f535; fma.rn.ftz.f32 %f524, %f534, %f534, %f523; fma.rn.ftz.f32 %f525, %f533, %f533, %f524; sqrt.approx.ftz.f32 %f526, %f525; div.approx.ftz.f32 %f535, %f535, %f526; div.approx.ftz.f32 %f534, %f534, %f526; div.approx.ftz.f32 %f533, %f533, %f526; $L__BB95_38: setp.eq.s32 %p29, %r4, 255; @%p29 bra $L__BB95_40; mul.wide.u32 %rd34, %r4, 4; add.s64 %rd35, %rd1, %rd34; st.local.f32 [%rd35], %f535; st.local.f32 [%rd35+4], %f534; st.local.f32 [%rd35+8], %f533; $L__BB95_40: ret; } .func _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b64 _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3, .param .b32 _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4 ) { .reg .pred %p<62>; .reg .b16 %rs<37>; .reg .f32 %f<324>; .reg .b32 %r<136>; .reg .b64 %rd<176>; ld.param.u64 %rd50, [_Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u64 %rd51, [_Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r31, [_Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r32, [_Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; ld.param.u32 %r33, [_Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4]; cvta.to.local.u64 %rd52, %rd50; add.s64 %rd1, %rd52, 76; ld.local.u32 %r1, [%rd52+76]; ld.local.u32 %r2, [%rd52+60]; cvt.s64.s32 %rd7, %r2; setp.eq.s32 %p1, %r1, -1; mov.u32 %r23, -1; mov.u32 %r20, 0; mov.u32 %r133, %r20; mov.u32 %r134, %r20; @%p1 bra $L__BB96_15; ld.const.u64 %rd53, [kernel_params+2160]; cvta.to.global.u64 %rd54, %rd53; mul.wide.s32 %rd55, %r1, 208; add.s64 %rd56, %rd54, %rd55; ld.global.u32 %r3, [%rd56+164]; ld.local.u32 %r39, [%rd1+-12]; and.b32 %r40, %r39, 1; setp.eq.b32 %p2, %r40, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r127, 0; @%p5 bra $L__BB96_5; cvt.u32.u64 %r42, %rd7; setp.eq.s32 %p6, %r42, -1; mov.u32 %r126, -1; @%p6 bra $L__BB96_4; ld.const.u64 %rd57, [kernel_params+2240]; cvta.to.global.u64 %rd58, %rd57; mul.wide.s32 %rd59, %r2, 4; add.s64 %rd60, %rd58, %rd59; ld.global.u32 %r126, [%rd60]; $L__BB96_4: setp.ne.s32 %p7, %r126, -1; selp.u32 %r127, 1, 0, %p7; $L__BB96_5: ld.const.u64 %rd61, [kernel_params+2312]; cvta.to.global.u64 %rd8, %rd61; add.s32 %r129, %r127, %r3; mul.wide.u32 %rd62, %r129, 16; add.s64 %rd63, %rd8, %rd62; add.s64 %rd165, %rd63, 14; add.s64 %rd166, %rd63, 15; ld.global.u32 %r128, [%rd63+8]; ld.global.u16 %rs35, [%rd63+12]; ld.global.u64 %rd164, [%rd63]; cvt.u64.u32 %rd64, %r31; setp.eq.s64 %p8, %rd164, %rd64; @%p8 bra $L__BB96_11; $L__BB96_6: setp.eq.s64 %p9, %rd164, 0; @%p9 bra $L__BB96_8; add.s32 %r129, %r129, 2; bra.uni $L__BB96_9; $L__BB96_8: setp.eq.s16 %p10, %rs35, 0; mov.u32 %r20, 0; mov.u32 %r129, %r128; mov.u32 %r133, %r20; mov.u32 %r134, %r20; @%p10 bra $L__BB96_15; $L__BB96_9: cvt.u64.u32 %rd13, %r129; mul.wide.u32 %rd65, %r129, 16; add.s64 %rd66, %rd8, %rd65; ld.global.u32 %r128, [%rd66+8]; ld.global.u16 %rs35, [%rd66+12]; ld.global.u64 %rd164, [%rd66]; setp.ne.s64 %p11, %rd164, %rd64; @%p11 bra $L__BB96_6; shl.b64 %rd68, %rd13, 4; add.s64 %rd69, %rd8, %rd68; add.s64 %rd165, %rd69, 14; add.s64 %rd166, %rd69, 15; $L__BB96_11: ld.global.u8 %rs5, [%rd166]; ld.global.u8 %rs6, [%rd165]; cvt.u32.u64 %r47, %rd7; setp.ne.s32 %p12, %r47, -1; @%p12 bra $L__BB96_14; mov.u32 %r20, 0; add.s16 %rs19, %rs35, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB96_14; setp.ne.s16 %p14, %rs35, 1024; mov.u32 %r133, %r20; mov.u32 %r134, %r20; @%p14 bra $L__BB96_15; $L__BB96_14: setp.eq.s16 %p15, %rs35, 0; selp.b32 %r23, -1, %r128, %p15; cvt.u32.u16 %r20, %rs35; cvt.u32.u16 %r52, %rs5; and.b32 %r134, %r52, 255; cvt.u32.u16 %r53, %rs6; and.b32 %r133, %r53, 255; $L__BB96_15: setp.eq.s32 %p16, %r23, -1; cvta.to.local.u64 %rd70, %rd51; mul.wide.u32 %rd71, %r32, 4; add.s64 %rd19, %rd70, %rd71; mul.wide.u32 %rd72, %r33, 4; add.s64 %rd20, %rd70, %rd72; @%p16 bra $L__BB96_83; add.s32 %r54, %r133, -3; setp.lt.u32 %p17, %r54, 2; ld.local.u32 %r30, [%rd52+64]; @%p17 bra $L__BB96_36; bra.uni $L__BB96_17; $L__BB96_36: and.b32 %r92, %r30, 1; setp.eq.b32 %p35, %r92, 1; mov.pred %p36, 0; xor.pred %p37, %p35, %p36; not.pred %p38, %p37; @%p38 bra $L__BB96_75; bra.uni $L__BB96_37; $L__BB96_75: and.b32 %r113, %r30, 6; setp.eq.s32 %p55, %r113, 0; @%p55 bra $L__BB96_80; and.b32 %r114, %r20, 768; setp.eq.s32 %p56, %r114, 0; @%p56 bra $L__BB96_78; ld.const.u64 %rd144, [kernel_params+2264]; cvta.to.global.u64 %rd145, %rd144; shl.b64 %rd146, %rd7, 4; add.s64 %rd147, %rd145, %rd146; shr.s32 %r115, %r30, 7; ld.global.u32 %r116, [%rd147+4]; add.s32 %r117, %r115, %r116; ld.const.u64 %rd148, [kernel_params+2344]; cvta.to.global.u64 %rd149, %rd148; add.s32 %r118, %r117, %r23; mul.wide.s32 %rd150, %r118, 16; add.s64 %rd151, %rd149, %rd150; ld.global.v4.f32 {%f258, %f259, %f260, %f261}, [%rd151]; ld.global.v4.f32 {%f266, %f267, %f268, %f269}, [%rd151+16]; ld.local.f32 %f274, [%rd52+68]; mov.f32 %f275, 0f3F800000; sub.ftz.f32 %f276, %f275, %f274; mul.ftz.f32 %f277, %f258, %f276; mul.ftz.f32 %f278, %f259, %f276; mul.ftz.f32 %f279, %f260, %f276; mul.ftz.f32 %f280, %f261, %f276; fma.rn.ftz.f32 %f323, %f269, %f274, %f280; fma.rn.ftz.f32 %f322, %f268, %f274, %f279; fma.rn.ftz.f32 %f321, %f267, %f274, %f278; fma.rn.ftz.f32 %f320, %f266, %f274, %f277; bra.uni $L__BB96_82; $L__BB96_83: mov.u32 %r125, 0; st.local.u32 [%rd19], %r125; st.local.u32 [%rd19+4], %r125; st.local.u32 [%rd19+8], %r125; st.local.u32 [%rd20], %r125; bra.uni $L__BB96_84; $L__BB96_17: and.b32 %r55, %r30, 1; setp.eq.b32 %p18, %r55, 1; mov.pred %p19, 0; xor.pred %p20, %p18, %p19; not.pred %p21, %p20; @%p21 bra $L__BB96_28; bra.uni $L__BB96_18; $L__BB96_28: and.b32 %r78, %r30, 6; setp.eq.s32 %p28, %r78, 0; @%p28 bra $L__BB96_33; and.b32 %r79, %r20, 768; setp.eq.s32 %p29, %r79, 0; @%p29 bra $L__BB96_31; ld.const.u64 %rd98, [kernel_params+2264]; cvta.to.global.u64 %rd99, %rd98; shl.b64 %rd100, %rd7, 4; add.s64 %rd101, %rd99, %rd100; shr.s32 %r80, %r30, 7; ld.global.u32 %r81, [%rd101+4]; add.s32 %r82, %r80, %r81; ld.const.u64 %rd102, [kernel_params+2336]; cvta.to.global.u64 %rd103, %rd102; add.s32 %r83, %r82, %r23; mul.wide.s32 %rd104, %r83, 12; add.s64 %rd105, %rd103, %rd104; ld.local.f32 %f147, [%rd52+68]; mov.f32 %f148, 0f3F800000; sub.ftz.f32 %f149, %f148, %f147; ld.global.f32 %f150, [%rd105]; mul.ftz.f32 %f151, %f150, %f149; ld.global.f32 %f152, [%rd105+4]; mul.ftz.f32 %f153, %f152, %f149; ld.global.f32 %f154, [%rd105+8]; mul.ftz.f32 %f155, %f154, %f149; ld.global.f32 %f156, [%rd105+12]; ld.global.f32 %f157, [%rd105+16]; ld.global.f32 %f158, [%rd105+20]; fma.rn.ftz.f32 %f299, %f158, %f147, %f155; fma.rn.ftz.f32 %f298, %f157, %f147, %f153; fma.rn.ftz.f32 %f297, %f156, %f147, %f151; bra.uni $L__BB96_35; $L__BB96_37: cvt.u32.u64 %r93, %rd7; setp.eq.s32 %p39, %r93, -1; @%p39 bra $L__BB96_40; ld.const.u64 %rd115, [kernel_params+2240]; cvta.to.global.u64 %rd116, %rd115; shl.b64 %rd117, %rd7, 2; add.s64 %rd118, %rd116, %rd117; ld.global.u32 %r94, [%rd118]; setp.eq.s32 %p40, %r94, -1; @%p40 bra $L__BB96_40; mov.u64 %rd119, 0; { // callseq 69, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd50; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r20; st.param.b32 [param1+4], %r133; st.param.b32 [param1+8], %r134; st.param.b32 [param1+12], %r23; .param .b64 param2; st.param.b64 [param2+0], %rd119; .param .b64 param3; st.param.b64 [param3+0], %rd119; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f320, [retval0+0]; ld.param.f32 %f321, [retval0+4]; ld.param.f32 %f322, [retval0+8]; ld.param.f32 %f323, [retval0+12]; } // callseq 69 bra.uni $L__BB96_82; $L__BB96_18: cvt.u32.u64 %r56, %rd7; setp.eq.s32 %p22, %r56, -1; @%p22 bra $L__BB96_21; ld.const.u64 %rd74, [kernel_params+2240]; cvta.to.global.u64 %rd75, %rd74; shl.b64 %rd76, %rd7, 2; add.s64 %rd77, %rd75, %rd76; ld.global.u32 %r57, [%rd77]; setp.eq.s32 %p23, %r57, -1; @%p23 bra $L__BB96_21; mov.u64 %rd78, 0; { // callseq 68, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd50; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r20; st.param.b32 [param1+4], %r133; st.param.b32 [param1+8], %r134; st.param.b32 [param1+12], %r23; .param .b64 param2; st.param.b64 [param2+0], %rd78; .param .b64 param3; st.param.b64 [param3+0], %rd78; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f297, [retval0+0]; ld.param.f32 %f298, [retval0+4]; ld.param.f32 %f299, [retval0+8]; } // callseq 68 bra.uni $L__BB96_35; $L__BB96_40: and.b32 %r95, %r20, 120; setp.eq.s32 %p41, %r95, 0; @%p41 bra $L__BB96_73; and.b32 %r96, %r20, 24; setp.eq.s32 %p42, %r96, 0; @%p42 bra $L__BB96_43; ld.const.u64 %rd120, [kernel_params+2232]; cvta.to.global.u64 %rd121, %rd120; shl.b64 %rd122, %rd7, 4; add.s64 %rd123, %rd121, %rd122; ld.global.v4.u32 {%r97, %r98, %r99, %r100}, [%rd123]; ld.const.u64 %rd124, [kernel_params+2344]; cvta.to.global.u64 %rd125, %rd124; add.s32 %r104, %r97, %r23; mul.wide.u32 %rd126, %r104, 16; add.s64 %rd127, %rd125, %rd126; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd127]; add.s32 %r105, %r98, %r23; mul.wide.u32 %rd128, %r105, 16; add.s64 %rd129, %rd125, %rd128; ld.global.v4.f32 {%f312, %f313, %f314, %f315}, [%rd129]; add.s32 %r106, %r99, %r23; mul.wide.u32 %rd130, %r106, 16; add.s64 %rd131, %rd125, %rd130; ld.global.v4.f32 {%f308, %f309, %f310, %f311}, [%rd131]; bra.uni $L__BB96_72; $L__BB96_21: and.b32 %r58, %r20, 56; setp.eq.s32 %p24, %r58, 0; @%p24 bra $L__BB96_26; and.b32 %r59, %r20, 24; setp.eq.s32 %p25, %r59, 0; @%p25 bra $L__BB96_24; ld.const.u64 %rd79, [kernel_params+2232]; cvta.to.global.u64 %rd80, %rd79; shl.b64 %rd81, %rd7, 4; add.s64 %rd82, %rd80, %rd81; ld.global.v4.u32 {%r60, %r61, %r62, %r63}, [%rd82]; ld.const.u64 %rd83, [kernel_params+2336]; cvta.to.global.u64 %rd84, %rd83; add.s32 %r67, %r60, %r23; mul.wide.u32 %rd85, %r67, 12; add.s64 %rd173, %rd84, %rd85; add.s64 %rd174, %rd173, 4; add.s64 %rd175, %rd173, 8; add.s32 %r68, %r61, %r23; mul.wide.u32 %rd86, %r68, 12; add.s64 %rd170, %rd84, %rd86; add.s64 %rd171, %rd170, 4; add.s64 %rd172, %rd170, 8; add.s32 %r69, %r62, %r23; mul.wide.u32 %rd87, %r69, 12; add.s64 %rd167, %rd84, %rd87; add.s64 %rd168, %rd167, 4; add.s64 %rd169, %rd167, 8; bra.uni $L__BB96_25; $L__BB96_80: and.b32 %r122, %r30, 8; setp.eq.s32 %p59, %r122, 0; setp.ne.s32 %p60, %r20, 8; mov.f32 %f320, 0f00000000; or.pred %p61, %p60, %p59; mov.f32 %f321, %f320; mov.f32 %f322, %f320; mov.f32 %f323, %f320; @%p61 bra $L__BB96_82; cvt.u32.u64 %r123, %rd7; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r124, %r123, %r23; mul.wide.s32 %rd159, %r124, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f320, %f321, %f322, %f323}, [%rd160]; bra.uni $L__BB96_82; $L__BB96_33: and.b32 %r88, %r30, 8; setp.eq.s32 %p32, %r88, 0; setp.ne.s32 %p33, %r20, 8; mov.f32 %f297, 0f00000000; or.pred %p34, %p33, %p32; mov.f32 %f298, %f297; mov.f32 %f299, %f297; @%p34 bra $L__BB96_35; cvt.u32.u64 %r89, %rd7; ld.const.u64 %rd111, [kernel_params+2336]; cvta.to.global.u64 %rd112, %rd111; add.s32 %r90, %r89, %r23; mul.wide.s32 %rd113, %r90, 12; add.s64 %rd114, %rd112, %rd113; ld.global.f32 %f299, [%rd114+8]; ld.global.f32 %f298, [%rd114+4]; ld.global.f32 %f297, [%rd114]; bra.uni $L__BB96_35; $L__BB96_78: and.b32 %r119, %r20, 131; setp.eq.s32 %p57, %r119, 0; mov.f32 %f320, 0f00000000; mov.f32 %f321, %f320; mov.f32 %f322, %f320; mov.f32 %f323, %f320; @%p57 bra $L__BB96_82; setp.eq.s32 %p58, %r20, 128; selp.b32 %r120, %r2, 0, %p58; add.s32 %r121, %r23, %r120; ld.const.u64 %rd153, [kernel_params+2344]; cvta.to.global.u64 %rd154, %rd153; mul.wide.s32 %rd155, %r121, 16; add.s64 %rd156, %rd154, %rd155; ld.global.v4.f32 {%f320, %f321, %f322, %f323}, [%rd156]; bra.uni $L__BB96_82; $L__BB96_31: and.b32 %r84, %r20, 131; setp.eq.s32 %p30, %r84, 0; mov.f32 %f297, 0f00000000; mov.f32 %f298, %f297; mov.f32 %f299, %f297; @%p30 bra $L__BB96_35; cvt.u32.u64 %r85, %rd7; setp.eq.s32 %p31, %r20, 128; selp.b32 %r86, %r85, 0, %p31; add.s32 %r87, %r23, %r86; ld.const.u64 %rd107, [kernel_params+2336]; cvta.to.global.u64 %rd108, %rd107; mul.wide.s32 %rd109, %r87, 12; add.s64 %rd110, %rd108, %rd109; ld.global.f32 %f299, [%rd110+8]; ld.global.f32 %f298, [%rd110+4]; ld.global.f32 %f297, [%rd110]; bra.uni $L__BB96_35; $L__BB96_73: and.b32 %r109, %r20, 7; setp.eq.s32 %p53, %r109, 0; mov.f32 %f320, 0f00000000; mov.f32 %f321, %f320; mov.f32 %f322, %f320; mov.f32 %f323, %f320; @%p53 bra $L__BB96_82; setp.eq.s32 %p54, %r20, 4; selp.b32 %r111, %r93, 0, %p54; add.s32 %r112, %r23, %r111; ld.const.u64 %rd140, [kernel_params+2344]; cvta.to.global.u64 %rd141, %rd140; mul.wide.s32 %rd142, %r112, 16; add.s64 %rd143, %rd141, %rd142; ld.global.v4.f32 {%f320, %f321, %f322, %f323}, [%rd143]; bra.uni $L__BB96_82; $L__BB96_26: and.b32 %r74, %r20, 7; setp.eq.s32 %p26, %r74, 0; mov.f32 %f297, 0f00000000; mov.f32 %f298, %f297; mov.f32 %f299, %f297; @%p26 bra $L__BB96_35; setp.eq.s32 %p27, %r20, 4; selp.b32 %r76, %r56, 0, %p27; add.s32 %r77, %r23, %r76; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; mul.wide.s32 %rd96, %r77, 12; add.s64 %rd97, %rd95, %rd96; ld.global.f32 %f299, [%rd97+8]; ld.global.f32 %f298, [%rd97+4]; ld.global.f32 %f297, [%rd97]; bra.uni $L__BB96_35; $L__BB96_43: mad.lo.s32 %r108, %r93, 3, %r23; cvt.s64.s32 %rd48, %r108; setp.eq.s32 %p43, %r20, 32; @%p43 bra $L__BB96_71; bra.uni $L__BB96_44; $L__BB96_71: ld.const.u64 %rd135, [kernel_params+2344]; cvta.to.global.u64 %rd136, %rd135; shl.b64 %rd137, %rd48, 4; add.s64 %rd138, %rd136, %rd137; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd138]; ld.global.v4.f32 {%f312, %f313, %f314, %f315}, [%rd138+16]; ld.global.v4.f32 {%f308, %f309, %f310, %f311}, [%rd138+32]; bra.uni $L__BB96_72; $L__BB96_24: mad.lo.s32 %r71, %r56, 3, %r23; ld.const.u64 %rd88, [kernel_params+2336]; cvta.to.global.u64 %rd89, %rd88; mul.wide.s32 %rd90, %r71, 12; add.s64 %rd173, %rd89, %rd90; add.s64 %rd174, %rd173, 4; add.s64 %rd175, %rd173, 8; add.s32 %r72, %r71, 1; mul.wide.s32 %rd91, %r72, 12; add.s64 %rd170, %rd89, %rd91; add.s64 %rd171, %rd170, 4; add.s64 %rd172, %rd170, 8; add.s32 %r73, %r71, 2; mul.wide.s32 %rd92, %r73, 12; add.s64 %rd167, %rd89, %rd92; add.s64 %rd168, %rd167, 4; add.s64 %rd169, %rd167, 8; $L__BB96_25: ld.local.f32 %f124, [%rd52+68]; ld.global.f32 %f125, [%rd170]; ld.global.f32 %f126, [%rd171]; ld.global.f32 %f127, [%rd172]; ld.local.f32 %f128, [%rd52+72]; ld.global.f32 %f129, [%rd167]; mul.ftz.f32 %f130, %f129, %f128; ld.global.f32 %f131, [%rd168]; mul.ftz.f32 %f132, %f131, %f128; ld.global.f32 %f133, [%rd169]; mul.ftz.f32 %f134, %f133, %f128; fma.rn.ftz.f32 %f135, %f125, %f124, %f130; fma.rn.ftz.f32 %f136, %f126, %f124, %f132; fma.rn.ftz.f32 %f137, %f127, %f124, %f134; mov.f32 %f138, 0f3F800000; sub.ftz.f32 %f139, %f138, %f124; sub.ftz.f32 %f140, %f139, %f128; ld.global.f32 %f141, [%rd173]; ld.global.f32 %f142, [%rd174]; ld.global.f32 %f143, [%rd175]; fma.rn.ftz.f32 %f299, %f143, %f140, %f137; fma.rn.ftz.f32 %f298, %f142, %f140, %f136; fma.rn.ftz.f32 %f297, %f141, %f140, %f135; $L__BB96_35: st.local.f32 [%rd19], %f297; st.local.f32 [%rd19+4], %f298; st.local.f32 [%rd19+8], %f299; mov.u32 %r91, 1065353216; st.local.u32 [%rd20], %r91; bra.uni $L__BB96_84; $L__BB96_44: ld.const.u64 %rd132, [kernel_params+2352]; cvta.to.global.u64 %rd133, %rd132; shl.b64 %rd134, %rd48, 2; add.s64 %rd49, %rd133, %rd134; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd49]; cvt.rn.f32.u16 %f177, %rs20; mul.ftz.f32 %f38, %f177, 0f3B808081; cvt.rn.f32.u16 %f178, %rs21; mul.ftz.f32 %f39, %f178, 0f3B808081; cvt.rn.f32.u16 %f179, %rs22; mul.ftz.f32 %f40, %f179, 0f3B808081; setp.lt.ftz.f32 %p44, %f38, 0f3D25AEE6; @%p44 bra $L__BB96_46; bra.uni $L__BB96_45; $L__BB96_46: mul.ftz.f32 %f316, %f38, 0f3D9E8391; bra.uni $L__BB96_47; $L__BB96_45: add.ftz.f32 %f180, %f38, 0f3D6147AE; mul.ftz.f32 %f181, %f180, 0f3F72A76F; lg2.approx.ftz.f32 %f182, %f181; mul.ftz.f32 %f183, %f182, 0f4019999A; ex2.approx.ftz.f32 %f316, %f183; $L__BB96_47: setp.lt.ftz.f32 %p45, %f39, 0f3D25AEE6; @%p45 bra $L__BB96_49; bra.uni $L__BB96_48; $L__BB96_49: mul.ftz.f32 %f317, %f39, 0f3D9E8391; bra.uni $L__BB96_50; $L__BB96_48: add.ftz.f32 %f184, %f39, 0f3D6147AE; mul.ftz.f32 %f185, %f184, 0f3F72A76F; lg2.approx.ftz.f32 %f186, %f185; mul.ftz.f32 %f187, %f186, 0f4019999A; ex2.approx.ftz.f32 %f317, %f187; $L__BB96_50: setp.lt.ftz.f32 %p46, %f40, 0f3D25AEE6; @%p46 bra $L__BB96_52; bra.uni $L__BB96_51; $L__BB96_52: mul.ftz.f32 %f318, %f40, 0f3D9E8391; bra.uni $L__BB96_53; $L__BB96_51: add.ftz.f32 %f188, %f40, 0f3D6147AE; mul.ftz.f32 %f189, %f188, 0f3F72A76F; lg2.approx.ftz.f32 %f190, %f189; mul.ftz.f32 %f191, %f190, 0f4019999A; ex2.approx.ftz.f32 %f318, %f191; $L__BB96_53: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd49+4]; cvt.rn.f32.u16 %f192, %rs24; mul.ftz.f32 %f50, %f192, 0f3B808081; cvt.rn.f32.u16 %f193, %rs25; mul.ftz.f32 %f51, %f193, 0f3B808081; cvt.rn.f32.u16 %f194, %rs26; mul.ftz.f32 %f52, %f194, 0f3B808081; setp.lt.ftz.f32 %p47, %f50, 0f3D25AEE6; @%p47 bra $L__BB96_55; bra.uni $L__BB96_54; $L__BB96_55: mul.ftz.f32 %f312, %f50, 0f3D9E8391; bra.uni $L__BB96_56; $L__BB96_54: add.ftz.f32 %f195, %f50, 0f3D6147AE; mul.ftz.f32 %f196, %f195, 0f3F72A76F; lg2.approx.ftz.f32 %f197, %f196; mul.ftz.f32 %f198, %f197, 0f4019999A; ex2.approx.ftz.f32 %f312, %f198; $L__BB96_56: setp.lt.ftz.f32 %p48, %f51, 0f3D25AEE6; @%p48 bra $L__BB96_58; bra.uni $L__BB96_57; $L__BB96_58: mul.ftz.f32 %f313, %f51, 0f3D9E8391; bra.uni $L__BB96_59; $L__BB96_57: add.ftz.f32 %f199, %f51, 0f3D6147AE; mul.ftz.f32 %f200, %f199, 0f3F72A76F; lg2.approx.ftz.f32 %f201, %f200; mul.ftz.f32 %f202, %f201, 0f4019999A; ex2.approx.ftz.f32 %f313, %f202; $L__BB96_59: setp.lt.ftz.f32 %p49, %f52, 0f3D25AEE6; @%p49 bra $L__BB96_61; bra.uni $L__BB96_60; $L__BB96_61: mul.ftz.f32 %f314, %f52, 0f3D9E8391; bra.uni $L__BB96_62; $L__BB96_60: add.ftz.f32 %f203, %f52, 0f3D6147AE; mul.ftz.f32 %f204, %f203, 0f3F72A76F; lg2.approx.ftz.f32 %f205, %f204; mul.ftz.f32 %f206, %f205, 0f4019999A; ex2.approx.ftz.f32 %f314, %f206; $L__BB96_62: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd49+8]; cvt.rn.f32.u16 %f207, %rs28; mul.ftz.f32 %f62, %f207, 0f3B808081; cvt.rn.f32.u16 %f208, %rs29; mul.ftz.f32 %f63, %f208, 0f3B808081; cvt.rn.f32.u16 %f209, %rs30; mul.ftz.f32 %f64, %f209, 0f3B808081; setp.lt.ftz.f32 %p50, %f62, 0f3D25AEE6; @%p50 bra $L__BB96_64; bra.uni $L__BB96_63; $L__BB96_64: mul.ftz.f32 %f308, %f62, 0f3D9E8391; bra.uni $L__BB96_65; $L__BB96_63: add.ftz.f32 %f210, %f62, 0f3D6147AE; mul.ftz.f32 %f211, %f210, 0f3F72A76F; lg2.approx.ftz.f32 %f212, %f211; mul.ftz.f32 %f213, %f212, 0f4019999A; ex2.approx.ftz.f32 %f308, %f213; $L__BB96_65: setp.lt.ftz.f32 %p51, %f63, 0f3D25AEE6; @%p51 bra $L__BB96_67; bra.uni $L__BB96_66; $L__BB96_67: mul.ftz.f32 %f309, %f63, 0f3D9E8391; bra.uni $L__BB96_68; $L__BB96_66: add.ftz.f32 %f214, %f63, 0f3D6147AE; mul.ftz.f32 %f215, %f214, 0f3F72A76F; lg2.approx.ftz.f32 %f216, %f215; mul.ftz.f32 %f217, %f216, 0f4019999A; ex2.approx.ftz.f32 %f309, %f217; $L__BB96_68: cvt.rn.f32.u16 %f218, %rs23; mul.ftz.f32 %f319, %f218, 0f3B808081; cvt.rn.f32.u16 %f219, %rs27; mul.ftz.f32 %f315, %f219, 0f3B808081; cvt.rn.f32.u16 %f220, %rs31; mul.ftz.f32 %f311, %f220, 0f3B808081; setp.lt.ftz.f32 %p52, %f64, 0f3D25AEE6; @%p52 bra $L__BB96_70; bra.uni $L__BB96_69; $L__BB96_70: mul.ftz.f32 %f310, %f64, 0f3D9E8391; bra.uni $L__BB96_72; $L__BB96_69: add.ftz.f32 %f221, %f64, 0f3D6147AE; mul.ftz.f32 %f222, %f221, 0f3F72A76F; lg2.approx.ftz.f32 %f223, %f222; mul.ftz.f32 %f224, %f223, 0f4019999A; ex2.approx.ftz.f32 %f310, %f224; $L__BB96_72: ld.local.f32 %f237, [%rd52+68]; ld.local.f32 %f238, [%rd52+72]; mul.ftz.f32 %f239, %f308, %f238; mul.ftz.f32 %f240, %f309, %f238; mul.ftz.f32 %f241, %f310, %f238; mul.ftz.f32 %f242, %f311, %f238; fma.rn.ftz.f32 %f243, %f312, %f237, %f239; fma.rn.ftz.f32 %f244, %f313, %f237, %f240; fma.rn.ftz.f32 %f245, %f314, %f237, %f241; fma.rn.ftz.f32 %f246, %f315, %f237, %f242; mov.f32 %f247, 0f3F800000; sub.ftz.f32 %f248, %f247, %f237; sub.ftz.f32 %f249, %f248, %f238; fma.rn.ftz.f32 %f323, %f319, %f249, %f246; fma.rn.ftz.f32 %f322, %f318, %f249, %f245; fma.rn.ftz.f32 %f321, %f317, %f249, %f244; fma.rn.ftz.f32 %f320, %f316, %f249, %f243; $L__BB96_82: st.local.f32 [%rd19], %f320; st.local.f32 [%rd19+4], %f321; st.local.f32 [%rd19+8], %f322; st.local.f32 [%rd20], %f323; $L__BB96_84: ret; } .func _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b64 _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3, .param .b32 _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4 ) { .local .align 16 .b8 __local_depot97[32]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<56>; .reg .b16 %rs<37>; .reg .f32 %f<448>; .reg .b32 %r<137>; .reg .b64 %rd<183>; mov.u64 %SPL, __local_depot97; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd53, [_Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u64 %rd54, [_Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %rd1, [_Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r31, [_Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; ld.param.u32 %r32, [_Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4]; cvta.to.local.u64 %rd55, %rd53; add.s64 %rd2, %rd55, 76; ld.local.u32 %r1, [%rd55+76]; ld.local.s32 %rd9, [%rd55+60]; setp.eq.s32 %p1, %r1, -1; mov.u32 %r22, -1; mov.u32 %r19, 0; mov.u32 %r134, %r19; mov.u32 %r135, %r19; @%p1 bra $L__BB97_15; ld.const.u64 %rd56, [kernel_params+2160]; cvta.to.global.u64 %rd57, %rd56; mul.wide.s32 %rd58, %r1, 208; add.s64 %rd59, %rd57, %rd58; ld.global.u32 %r2, [%rd59+164]; ld.local.u32 %r38, [%rd2+-12]; and.b32 %r39, %r38, 1; setp.eq.b32 %p2, %r39, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r128, 0; @%p5 bra $L__BB97_5; cvt.u32.u64 %r41, %rd9; setp.eq.s32 %p6, %r41, -1; mov.u32 %r127, -1; @%p6 bra $L__BB97_4; ld.const.u64 %rd60, [kernel_params+2240]; cvta.to.global.u64 %rd61, %rd60; shl.b64 %rd62, %rd9, 2; add.s64 %rd63, %rd61, %rd62; ld.global.u32 %r127, [%rd63]; $L__BB97_4: setp.ne.s32 %p7, %r127, -1; selp.u32 %r128, 1, 0, %p7; $L__BB97_5: ld.const.u64 %rd64, [kernel_params+2312]; cvta.to.global.u64 %rd10, %rd64; add.s32 %r130, %r128, %r2; mul.wide.u32 %rd65, %r130, 16; add.s64 %rd66, %rd10, %rd65; add.s64 %rd172, %rd66, 14; add.s64 %rd173, %rd66, 15; ld.global.u32 %r129, [%rd66+8]; ld.global.u16 %rs35, [%rd66+12]; ld.global.u64 %rd171, [%rd66]; setp.eq.s64 %p8, %rd171, %rd1; @%p8 bra $L__BB97_11; $L__BB97_6: setp.eq.s64 %p9, %rd171, 0; @%p9 bra $L__BB97_8; add.s32 %r130, %r130, 2; bra.uni $L__BB97_9; $L__BB97_8: setp.eq.s16 %p10, %rs35, 0; mov.u32 %r19, 0; mov.u32 %r130, %r129; mov.u32 %r134, %r19; mov.u32 %r135, %r19; @%p10 bra $L__BB97_15; $L__BB97_9: cvt.u64.u32 %rd15, %r130; mul.wide.u32 %rd67, %r130, 16; add.s64 %rd68, %rd10, %rd67; ld.global.u32 %r129, [%rd68+8]; ld.global.u16 %rs35, [%rd68+12]; ld.global.u64 %rd171, [%rd68]; setp.ne.s64 %p11, %rd171, %rd1; @%p11 bra $L__BB97_6; shl.b64 %rd69, %rd15, 4; add.s64 %rd70, %rd10, %rd69; add.s64 %rd172, %rd70, 14; add.s64 %rd173, %rd70, 15; $L__BB97_11: ld.global.u8 %rs5, [%rd173]; ld.global.u8 %rs6, [%rd172]; cvt.u32.u64 %r46, %rd9; setp.ne.s32 %p12, %r46, -1; @%p12 bra $L__BB97_14; mov.u32 %r19, 0; add.s16 %rs19, %rs35, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB97_14; setp.ne.s16 %p14, %rs35, 1024; mov.u32 %r134, %r19; mov.u32 %r135, %r19; @%p14 bra $L__BB97_15; $L__BB97_14: setp.eq.s16 %p15, %rs35, 0; selp.b32 %r22, -1, %r129, %p15; cvt.u32.u16 %r19, %rs35; cvt.u32.u16 %r51, %rs5; and.b32 %r135, %r51, 255; cvt.u32.u16 %r52, %rs6; and.b32 %r134, %r52, 255; $L__BB97_15: cvta.to.local.u64 %rd71, %rd54; setp.eq.s32 %p16, %r22, -1; mul.wide.u32 %rd72, %r31, 4; add.s64 %rd21, %rd71, %rd72; mul.wide.u32 %rd73, %r32, 4; add.s64 %rd22, %rd71, %rd73; @%p16 bra $L__BB97_83; add.s32 %r53, %r134, -3; setp.lt.u32 %p17, %r53, 2; ld.local.u32 %r29, [%rd2+-12]; and.b32 %r30, %r29, 1; @%p17 bra $L__BB97_36; bra.uni $L__BB97_17; $L__BB97_36: setp.eq.s32 %p32, %r30, 0; @%p32 bra $L__BB97_75; bra.uni $L__BB97_37; $L__BB97_75: and.b32 %r113, %r29, 6; setp.eq.s32 %p49, %r113, 0; @%p49 bra $L__BB97_80; and.b32 %r114, %r19, 768; setp.eq.s32 %p50, %r114, 0; @%p50 bra $L__BB97_78; ld.const.u64 %rd149, [kernel_params+2264]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd9, 4; add.s64 %rd152, %rd150, %rd151; shr.s32 %r115, %r29, 7; ld.global.u32 %r116, [%rd152+4]; add.s32 %r117, %r115, %r116; ld.const.u64 %rd153, [kernel_params+2344]; cvta.to.global.u64 %rd154, %rd153; add.s32 %r118, %r117, %r22; mul.wide.s32 %rd155, %r118, 16; add.s64 %rd156, %rd154, %rd155; ld.global.v4.f32 {%f350, %f351, %f352, %f353}, [%rd156]; ld.global.v4.f32 {%f358, %f359, %f360, %f361}, [%rd156+16]; sub.ftz.f32 %f366, %f358, %f350; sub.ftz.f32 %f367, %f359, %f351; sub.ftz.f32 %f368, %f360, %f352; sub.ftz.f32 %f369, %f361, %f353; ld.local.f32 %f370, [%rd2+24]; mul.ftz.f32 %f443, %f370, %f366; mul.ftz.f32 %f442, %f370, %f367; mul.ftz.f32 %f441, %f370, %f368; mul.ftz.f32 %f440, %f370, %f369; add.u64 %rd158, %SPL, 0; st.local.v4.f32 [%rd158], {%f443, %f442, %f441, %f440}; ld.local.f32 %f371, [%rd2+-8]; mov.f32 %f372, 0f3F800000; sub.ftz.f32 %f373, %f372, %f371; mul.ftz.f32 %f374, %f350, %f373; mul.ftz.f32 %f375, %f351, %f373; mul.ftz.f32 %f376, %f352, %f373; mul.ftz.f32 %f377, %f353, %f373; fma.rn.ftz.f32 %f447, %f361, %f371, %f377; fma.rn.ftz.f32 %f446, %f360, %f371, %f376; fma.rn.ftz.f32 %f445, %f359, %f371, %f375; fma.rn.ftz.f32 %f444, %f358, %f371, %f374; bra.uni $L__BB97_82; $L__BB97_83: mov.u32 %r126, 0; st.local.u32 [%rd21], %r126; st.local.u32 [%rd21+4], %r126; st.local.u32 [%rd21+8], %r126; st.local.u32 [%rd22], %r126; bra.uni $L__BB97_84; $L__BB97_17: setp.eq.s32 %p18, %r30, 0; add.u64 %rd74, %SP, 16; add.u64 %rd23, %SPL, 16; @%p18 bra $L__BB97_28; bra.uni $L__BB97_18; $L__BB97_28: and.b32 %r77, %r29, 6; setp.eq.s32 %p25, %r77, 0; @%p25 bra $L__BB97_33; and.b32 %r78, %r19, 768; setp.eq.s32 %p26, %r78, 0; @%p26 bra $L__BB97_31; ld.const.u64 %rd99, [kernel_params+2264]; cvta.to.global.u64 %rd100, %rd99; shl.b64 %rd101, %rd9, 4; add.s64 %rd102, %rd100, %rd101; shr.s32 %r79, %r29, 7; ld.global.u32 %r80, [%rd102+4]; add.s32 %r81, %r79, %r80; ld.const.u64 %rd103, [kernel_params+2336]; cvta.to.global.u64 %rd104, %rd103; add.s32 %r82, %r81, %r22; mul.wide.s32 %rd105, %r82, 12; add.s64 %rd106, %rd104, %rd105; ld.global.f32 %f193, [%rd106+12]; ld.global.f32 %f194, [%rd106]; sub.ftz.f32 %f195, %f193, %f194; ld.global.f32 %f196, [%rd106+16]; ld.global.f32 %f197, [%rd106+4]; sub.ftz.f32 %f198, %f196, %f197; ld.global.f32 %f199, [%rd106+20]; ld.global.f32 %f200, [%rd106+8]; sub.ftz.f32 %f201, %f199, %f200; ld.local.f32 %f202, [%rd2+24]; mul.ftz.f32 %f416, %f195, %f202; mul.ftz.f32 %f415, %f198, %f202; mul.ftz.f32 %f414, %f201, %f202; st.local.f32 [%rd23], %f416; st.local.f32 [%rd23+4], %f415; st.local.f32 [%rd23+8], %f414; ld.local.f32 %f203, [%rd2+-8]; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f203; mul.ftz.f32 %f206, %f194, %f205; mul.ftz.f32 %f207, %f197, %f205; mul.ftz.f32 %f208, %f200, %f205; fma.rn.ftz.f32 %f419, %f199, %f203, %f208; fma.rn.ftz.f32 %f418, %f196, %f203, %f207; fma.rn.ftz.f32 %f417, %f193, %f203, %f206; bra.uni $L__BB97_35; $L__BB97_37: cvt.u32.u64 %r93, %rd9; setp.eq.s32 %p33, %r93, -1; @%p33 bra $L__BB97_40; ld.const.u64 %rd115, [kernel_params+2240]; cvta.to.global.u64 %rd116, %rd115; shl.b64 %rd117, %rd9, 2; add.s64 %rd118, %rd116, %rd117; ld.global.u32 %r94, [%rd118]; setp.eq.s32 %p34, %r94, -1; @%p34 bra $L__BB97_40; mov.u64 %rd119, 0; add.u64 %rd120, %SP, 0; { // callseq 71, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd53; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r19; st.param.b32 [param1+4], %r134; st.param.b32 [param1+8], %r135; st.param.b32 [param1+12], %r22; .param .b64 param2; st.param.b64 [param2+0], %rd120; .param .b64 param3; st.param.b64 [param3+0], %rd119; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f444, [retval0+0]; ld.param.f32 %f445, [retval0+4]; ld.param.f32 %f446, [retval0+8]; ld.param.f32 %f447, [retval0+12]; } // callseq 71 add.u64 %rd121, %SPL, 0; ld.local.v4.f32 {%f443, %f442, %f441, %f440}, [%rd121]; bra.uni $L__BB97_82; $L__BB97_18: cvt.u32.u64 %r54, %rd9; setp.eq.s32 %p19, %r54, -1; @%p19 bra $L__BB97_21; ld.const.u64 %rd75, [kernel_params+2240]; cvta.to.global.u64 %rd76, %rd75; shl.b64 %rd77, %rd9, 2; add.s64 %rd78, %rd76, %rd77; ld.global.u32 %r55, [%rd78]; setp.eq.s32 %p20, %r55, -1; @%p20 bra $L__BB97_21; mov.u64 %rd79, 0; { // callseq 70, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd53; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r19; st.param.b32 [param1+4], %r134; st.param.b32 [param1+8], %r135; st.param.b32 [param1+12], %r22; .param .b64 param2; st.param.b64 [param2+0], %rd74; .param .b64 param3; st.param.b64 [param3+0], %rd79; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f417, [retval0+0]; ld.param.f32 %f418, [retval0+4]; ld.param.f32 %f419, [retval0+8]; } // callseq 70 ld.local.f32 %f416, [%rd23]; ld.local.f32 %f415, [%rd23+4]; ld.local.f32 %f414, [%rd23+8]; bra.uni $L__BB97_35; $L__BB97_40: and.b32 %r95, %r19, 120; setp.eq.s32 %p35, %r95, 0; @%p35 bra $L__BB97_73; and.b32 %r96, %r19, 24; setp.eq.s32 %p36, %r96, 0; @%p36 bra $L__BB97_43; ld.const.u64 %rd122, [kernel_params+2232]; cvta.to.global.u64 %rd123, %rd122; shl.b64 %rd124, %rd9, 4; add.s64 %rd125, %rd123, %rd124; ld.global.v4.u32 {%r97, %r98, %r99, %r100}, [%rd125]; ld.const.u64 %rd126, [kernel_params+2344]; cvta.to.global.u64 %rd127, %rd126; add.s32 %r104, %r97, %r22; mul.wide.u32 %rd128, %r104, 16; add.s64 %rd129, %rd127, %rd128; ld.global.v4.f32 {%f436, %f437, %f438, %f439}, [%rd129]; add.s32 %r105, %r98, %r22; mul.wide.u32 %rd130, %r105, 16; add.s64 %rd131, %rd127, %rd130; ld.global.v4.f32 {%f432, %f433, %f434, %f435}, [%rd131]; add.s32 %r106, %r99, %r22; mul.wide.u32 %rd132, %r106, 16; add.s64 %rd133, %rd127, %rd132; ld.global.v4.f32 {%f428, %f429, %f430, %f431}, [%rd133]; bra.uni $L__BB97_72; $L__BB97_21: and.b32 %r56, %r19, 56; setp.eq.s32 %p21, %r56, 0; @%p21 bra $L__BB97_26; and.b32 %r57, %r19, 24; setp.eq.s32 %p22, %r57, 0; @%p22 bra $L__BB97_24; ld.const.u64 %rd81, [kernel_params+2232]; cvta.to.global.u64 %rd82, %rd81; shl.b64 %rd83, %rd9, 4; add.s64 %rd84, %rd82, %rd83; ld.global.v4.u32 {%r58, %r59, %r60, %r61}, [%rd84]; ld.const.u64 %rd85, [kernel_params+2336]; cvta.to.global.u64 %rd86, %rd85; add.s32 %r65, %r58, %r22; mul.wide.u32 %rd87, %r65, 12; add.s64 %rd180, %rd86, %rd87; add.s64 %rd181, %rd180, 4; add.s64 %rd182, %rd180, 8; add.s32 %r66, %r59, %r22; mul.wide.u32 %rd88, %r66, 12; add.s64 %rd177, %rd86, %rd88; add.s64 %rd178, %rd177, 4; add.s64 %rd179, %rd177, 8; add.s32 %r67, %r60, %r22; mul.wide.u32 %rd89, %r67, 12; add.s64 %rd174, %rd86, %rd89; add.s64 %rd175, %rd174, 4; add.s64 %rd176, %rd174, 8; bra.uni $L__BB97_25; $L__BB97_80: and.b32 %r123, %r29, 8; setp.eq.s32 %p53, %r123, 0; add.u64 %rd166, %SPL, 0; mov.f32 %f440, 0f00000000; st.local.v4.f32 [%rd166], {%f440, %f440, %f440, %f440}; setp.ne.s32 %p54, %r19, 8; or.pred %p55, %p54, %p53; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; mov.f32 %f444, %f440; mov.f32 %f445, %f440; mov.f32 %f446, %f440; mov.f32 %f447, %f440; @%p55 bra $L__BB97_82; cvt.u32.u64 %r124, %rd9; ld.const.u64 %rd167, [kernel_params+2344]; cvta.to.global.u64 %rd168, %rd167; add.s32 %r125, %r124, %r22; mul.wide.s32 %rd169, %r125, 16; add.s64 %rd170, %rd168, %rd169; ld.global.v4.f32 {%f444, %f445, %f446, %f447}, [%rd170]; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; bra.uni $L__BB97_82; $L__BB97_33: and.b32 %r88, %r29, 8; setp.eq.s32 %p29, %r88, 0; mov.u32 %r89, 0; st.local.u32 [%rd23], %r89; st.local.u32 [%rd23+4], %r89; st.local.u32 [%rd23+8], %r89; setp.ne.s32 %p30, %r19, 8; mov.f32 %f414, 0f00000000; or.pred %p31, %p30, %p29; mov.f32 %f415, %f414; mov.f32 %f416, %f414; mov.f32 %f417, %f414; mov.f32 %f418, %f414; mov.f32 %f419, %f414; @%p31 bra $L__BB97_35; cvt.u32.u64 %r90, %rd9; ld.const.u64 %rd111, [kernel_params+2336]; cvta.to.global.u64 %rd112, %rd111; add.s32 %r91, %r90, %r22; mul.wide.s32 %rd113, %r91, 12; add.s64 %rd114, %rd112, %rd113; ld.global.f32 %f419, [%rd114+8]; ld.global.f32 %f418, [%rd114+4]; ld.global.f32 %f417, [%rd114]; mov.f32 %f415, %f414; mov.f32 %f416, %f414; bra.uni $L__BB97_35; $L__BB97_78: add.u64 %rd160, %SPL, 0; mov.f32 %f440, 0f00000000; st.local.v4.f32 [%rd160], {%f440, %f440, %f440, %f440}; and.b32 %r119, %r19, 131; setp.eq.s32 %p51, %r119, 0; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; mov.f32 %f444, %f440; mov.f32 %f445, %f440; mov.f32 %f446, %f440; mov.f32 %f447, %f440; @%p51 bra $L__BB97_82; cvt.u32.u64 %r120, %rd9; setp.eq.s32 %p52, %r19, 128; selp.b32 %r121, %r120, 0, %p52; add.s32 %r122, %r22, %r121; ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r122, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f444, %f445, %f446, %f447}, [%rd164]; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; bra.uni $L__BB97_82; $L__BB97_31: mov.u32 %r83, 0; st.local.u32 [%rd23], %r83; st.local.u32 [%rd23+4], %r83; st.local.u32 [%rd23+8], %r83; and.b32 %r84, %r19, 131; setp.eq.s32 %p27, %r84, 0; mov.f32 %f414, 0f00000000; mov.f32 %f415, %f414; mov.f32 %f416, %f414; mov.f32 %f417, %f414; mov.f32 %f418, %f414; mov.f32 %f419, %f414; @%p27 bra $L__BB97_35; cvt.u32.u64 %r85, %rd9; setp.eq.s32 %p28, %r19, 128; selp.b32 %r86, %r85, 0, %p28; add.s32 %r87, %r22, %r86; ld.const.u64 %rd107, [kernel_params+2336]; cvta.to.global.u64 %rd108, %rd107; mul.wide.s32 %rd109, %r87, 12; add.s64 %rd110, %rd108, %rd109; ld.global.f32 %f419, [%rd110+8]; ld.global.f32 %f418, [%rd110+4]; ld.global.f32 %f417, [%rd110]; mov.f32 %f415, %f414; mov.f32 %f416, %f414; bra.uni $L__BB97_35; $L__BB97_73: add.u64 %rd144, %SPL, 0; mov.f32 %f440, 0f00000000; st.local.v4.f32 [%rd144], {%f440, %f440, %f440, %f440}; and.b32 %r109, %r19, 7; setp.eq.s32 %p47, %r109, 0; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; mov.f32 %f444, %f440; mov.f32 %f445, %f440; mov.f32 %f446, %f440; mov.f32 %f447, %f440; @%p47 bra $L__BB97_82; setp.eq.s32 %p48, %r19, 4; selp.b32 %r111, %r93, 0, %p48; add.s32 %r112, %r22, %r111; ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r112, 16; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f444, %f445, %f446, %f447}, [%rd148]; mov.f32 %f441, %f440; mov.f32 %f442, %f440; mov.f32 %f443, %f440; bra.uni $L__BB97_82; $L__BB97_26: mov.u32 %r72, 0; st.local.u32 [%rd23], %r72; st.local.u32 [%rd23+4], %r72; st.local.u32 [%rd23+8], %r72; and.b32 %r73, %r19, 7; setp.eq.s32 %p23, %r73, 0; mov.f32 %f414, 0f00000000; mov.f32 %f415, %f414; mov.f32 %f416, %f414; mov.f32 %f417, %f414; mov.f32 %f418, %f414; mov.f32 %f419, %f414; @%p23 bra $L__BB97_35; setp.eq.s32 %p24, %r19, 4; selp.b32 %r75, %r54, 0, %p24; add.s32 %r76, %r22, %r75; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; mul.wide.s32 %rd97, %r76, 12; add.s64 %rd98, %rd96, %rd97; ld.global.f32 %f419, [%rd98+8]; ld.global.f32 %f418, [%rd98+4]; ld.global.f32 %f417, [%rd98]; mov.f32 %f415, %f414; mov.f32 %f416, %f414; bra.uni $L__BB97_35; $L__BB97_43: mad.lo.s32 %r108, %r93, 3, %r22; cvt.s64.s32 %rd51, %r108; setp.eq.s32 %p37, %r19, 32; @%p37 bra $L__BB97_71; bra.uni $L__BB97_44; $L__BB97_71: ld.const.u64 %rd137, [kernel_params+2344]; cvta.to.global.u64 %rd138, %rd137; shl.b64 %rd139, %rd51, 4; add.s64 %rd140, %rd138, %rd139; ld.global.v4.f32 {%f436, %f437, %f438, %f439}, [%rd140]; ld.global.v4.f32 {%f432, %f433, %f434, %f435}, [%rd140+16]; ld.global.v4.f32 {%f428, %f429, %f430, %f431}, [%rd140+32]; bra.uni $L__BB97_72; $L__BB97_24: mad.lo.s32 %r69, %r54, 3, %r22; ld.const.u64 %rd90, [kernel_params+2336]; cvta.to.global.u64 %rd91, %rd90; mul.wide.s32 %rd92, %r69, 12; add.s64 %rd180, %rd91, %rd92; add.s64 %rd181, %rd180, 4; add.s64 %rd182, %rd180, 8; add.s32 %r70, %r69, 1; mul.wide.s32 %rd93, %r70, 12; add.s64 %rd177, %rd91, %rd93; add.s64 %rd178, %rd177, 4; add.s64 %rd179, %rd177, 8; add.s32 %r71, %r69, 2; mul.wide.s32 %rd94, %r71, 12; add.s64 %rd174, %rd91, %rd94; add.s64 %rd175, %rd174, 4; add.s64 %rd176, %rd174, 8; $L__BB97_25: ld.local.f32 %f152, [%rd2+24]; ld.global.f32 %f153, [%rd177]; ld.global.f32 %f154, [%rd178]; ld.global.f32 %f155, [%rd179]; ld.local.f32 %f156, [%rd2+32]; ld.global.f32 %f157, [%rd174]; mul.ftz.f32 %f158, %f157, %f156; ld.global.f32 %f159, [%rd175]; mul.ftz.f32 %f160, %f159, %f156; ld.global.f32 %f161, [%rd176]; mul.ftz.f32 %f162, %f161, %f156; fma.rn.ftz.f32 %f163, %f153, %f152, %f158; fma.rn.ftz.f32 %f164, %f154, %f152, %f160; fma.rn.ftz.f32 %f165, %f155, %f152, %f162; add.ftz.f32 %f166, %f152, %f156; ld.global.f32 %f167, [%rd180]; mul.ftz.f32 %f168, %f167, %f166; ld.global.f32 %f169, [%rd181]; mul.ftz.f32 %f170, %f169, %f166; ld.global.f32 %f171, [%rd182]; mul.ftz.f32 %f172, %f171, %f166; sub.ftz.f32 %f416, %f163, %f168; sub.ftz.f32 %f415, %f164, %f170; sub.ftz.f32 %f414, %f165, %f172; st.local.f32 [%rd23], %f416; st.local.f32 [%rd23+4], %f415; st.local.f32 [%rd23+8], %f414; ld.local.f32 %f173, [%rd2+-8]; ld.local.f32 %f174, [%rd2+-4]; mul.ftz.f32 %f175, %f157, %f174; mul.ftz.f32 %f176, %f159, %f174; mul.ftz.f32 %f177, %f161, %f174; fma.rn.ftz.f32 %f178, %f153, %f173, %f175; fma.rn.ftz.f32 %f179, %f154, %f173, %f176; fma.rn.ftz.f32 %f180, %f155, %f173, %f177; mov.f32 %f181, 0f3F800000; sub.ftz.f32 %f182, %f181, %f173; sub.ftz.f32 %f183, %f182, %f174; fma.rn.ftz.f32 %f419, %f171, %f183, %f180; fma.rn.ftz.f32 %f418, %f169, %f183, %f179; fma.rn.ftz.f32 %f417, %f167, %f183, %f178; $L__BB97_35: add.ftz.f32 %f227, %f417, %f416; add.ftz.f32 %f228, %f418, %f415; add.ftz.f32 %f229, %f419, %f414; st.local.f32 [%rd21], %f227; st.local.f32 [%rd21+4], %f228; st.local.f32 [%rd21+8], %f229; mov.u32 %r92, 1065353216; st.local.u32 [%rd22], %r92; bra.uni $L__BB97_84; $L__BB97_44: ld.const.u64 %rd134, [kernel_params+2352]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd51, 2; add.s64 %rd52, %rd135, %rd136; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd52]; cvt.rn.f32.u16 %f246, %rs20; mul.ftz.f32 %f54, %f246, 0f3B808081; cvt.rn.f32.u16 %f247, %rs21; mul.ftz.f32 %f55, %f247, 0f3B808081; cvt.rn.f32.u16 %f248, %rs22; mul.ftz.f32 %f56, %f248, 0f3B808081; setp.lt.ftz.f32 %p38, %f54, 0f3D25AEE6; @%p38 bra $L__BB97_46; bra.uni $L__BB97_45; $L__BB97_46: mul.ftz.f32 %f436, %f54, 0f3D9E8391; bra.uni $L__BB97_47; $L__BB97_45: add.ftz.f32 %f249, %f54, 0f3D6147AE; mul.ftz.f32 %f250, %f249, 0f3F72A76F; lg2.approx.ftz.f32 %f251, %f250; mul.ftz.f32 %f252, %f251, 0f4019999A; ex2.approx.ftz.f32 %f436, %f252; $L__BB97_47: setp.lt.ftz.f32 %p39, %f55, 0f3D25AEE6; @%p39 bra $L__BB97_49; bra.uni $L__BB97_48; $L__BB97_49: mul.ftz.f32 %f437, %f55, 0f3D9E8391; bra.uni $L__BB97_50; $L__BB97_48: add.ftz.f32 %f253, %f55, 0f3D6147AE; mul.ftz.f32 %f254, %f253, 0f3F72A76F; lg2.approx.ftz.f32 %f255, %f254; mul.ftz.f32 %f256, %f255, 0f4019999A; ex2.approx.ftz.f32 %f437, %f256; $L__BB97_50: setp.lt.ftz.f32 %p40, %f56, 0f3D25AEE6; @%p40 bra $L__BB97_52; bra.uni $L__BB97_51; $L__BB97_52: mul.ftz.f32 %f438, %f56, 0f3D9E8391; bra.uni $L__BB97_53; $L__BB97_51: add.ftz.f32 %f257, %f56, 0f3D6147AE; mul.ftz.f32 %f258, %f257, 0f3F72A76F; lg2.approx.ftz.f32 %f259, %f258; mul.ftz.f32 %f260, %f259, 0f4019999A; ex2.approx.ftz.f32 %f438, %f260; $L__BB97_53: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd52+4]; cvt.rn.f32.u16 %f261, %rs24; mul.ftz.f32 %f66, %f261, 0f3B808081; cvt.rn.f32.u16 %f262, %rs25; mul.ftz.f32 %f67, %f262, 0f3B808081; cvt.rn.f32.u16 %f263, %rs26; mul.ftz.f32 %f68, %f263, 0f3B808081; setp.lt.ftz.f32 %p41, %f66, 0f3D25AEE6; @%p41 bra $L__BB97_55; bra.uni $L__BB97_54; $L__BB97_55: mul.ftz.f32 %f432, %f66, 0f3D9E8391; bra.uni $L__BB97_56; $L__BB97_54: add.ftz.f32 %f264, %f66, 0f3D6147AE; mul.ftz.f32 %f265, %f264, 0f3F72A76F; lg2.approx.ftz.f32 %f266, %f265; mul.ftz.f32 %f267, %f266, 0f4019999A; ex2.approx.ftz.f32 %f432, %f267; $L__BB97_56: setp.lt.ftz.f32 %p42, %f67, 0f3D25AEE6; @%p42 bra $L__BB97_58; bra.uni $L__BB97_57; $L__BB97_58: mul.ftz.f32 %f433, %f67, 0f3D9E8391; bra.uni $L__BB97_59; $L__BB97_57: add.ftz.f32 %f268, %f67, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f433, %f271; $L__BB97_59: setp.lt.ftz.f32 %p43, %f68, 0f3D25AEE6; @%p43 bra $L__BB97_61; bra.uni $L__BB97_60; $L__BB97_61: mul.ftz.f32 %f434, %f68, 0f3D9E8391; bra.uni $L__BB97_62; $L__BB97_60: add.ftz.f32 %f272, %f68, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f434, %f275; $L__BB97_62: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd52+8]; cvt.rn.f32.u16 %f276, %rs28; mul.ftz.f32 %f78, %f276, 0f3B808081; cvt.rn.f32.u16 %f277, %rs29; mul.ftz.f32 %f79, %f277, 0f3B808081; cvt.rn.f32.u16 %f278, %rs30; mul.ftz.f32 %f80, %f278, 0f3B808081; setp.lt.ftz.f32 %p44, %f78, 0f3D25AEE6; @%p44 bra $L__BB97_64; bra.uni $L__BB97_63; $L__BB97_64: mul.ftz.f32 %f428, %f78, 0f3D9E8391; bra.uni $L__BB97_65; $L__BB97_63: add.ftz.f32 %f279, %f78, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f428, %f282; $L__BB97_65: setp.lt.ftz.f32 %p45, %f79, 0f3D25AEE6; @%p45 bra $L__BB97_67; bra.uni $L__BB97_66; $L__BB97_67: mul.ftz.f32 %f429, %f79, 0f3D9E8391; bra.uni $L__BB97_68; $L__BB97_66: add.ftz.f32 %f283, %f79, 0f3D6147AE; mul.ftz.f32 %f284, %f283, 0f3F72A76F; lg2.approx.ftz.f32 %f285, %f284; mul.ftz.f32 %f286, %f285, 0f4019999A; ex2.approx.ftz.f32 %f429, %f286; $L__BB97_68: cvt.rn.f32.u16 %f287, %rs23; mul.ftz.f32 %f439, %f287, 0f3B808081; cvt.rn.f32.u16 %f288, %rs27; mul.ftz.f32 %f435, %f288, 0f3B808081; cvt.rn.f32.u16 %f289, %rs31; mul.ftz.f32 %f431, %f289, 0f3B808081; setp.lt.ftz.f32 %p46, %f80, 0f3D25AEE6; @%p46 bra $L__BB97_70; bra.uni $L__BB97_69; $L__BB97_70: mul.ftz.f32 %f430, %f80, 0f3D9E8391; bra.uni $L__BB97_72; $L__BB97_69: add.ftz.f32 %f290, %f80, 0f3D6147AE; mul.ftz.f32 %f291, %f290, 0f3F72A76F; lg2.approx.ftz.f32 %f292, %f291; mul.ftz.f32 %f293, %f292, 0f4019999A; ex2.approx.ftz.f32 %f430, %f293; $L__BB97_72: ld.local.f32 %f306, [%rd2+24]; ld.local.f32 %f307, [%rd2+32]; mul.ftz.f32 %f308, %f428, %f307; mul.ftz.f32 %f309, %f429, %f307; mul.ftz.f32 %f310, %f430, %f307; mul.ftz.f32 %f311, %f431, %f307; fma.rn.ftz.f32 %f312, %f432, %f306, %f308; fma.rn.ftz.f32 %f313, %f433, %f306, %f309; fma.rn.ftz.f32 %f314, %f434, %f306, %f310; fma.rn.ftz.f32 %f315, %f435, %f306, %f311; add.ftz.f32 %f316, %f306, %f307; mul.ftz.f32 %f317, %f436, %f316; mul.ftz.f32 %f318, %f437, %f316; mul.ftz.f32 %f319, %f438, %f316; mul.ftz.f32 %f320, %f439, %f316; sub.ftz.f32 %f443, %f312, %f317; sub.ftz.f32 %f442, %f313, %f318; sub.ftz.f32 %f441, %f314, %f319; sub.ftz.f32 %f440, %f315, %f320; add.u64 %rd142, %SPL, 0; st.local.v4.f32 [%rd142], {%f443, %f442, %f441, %f440}; ld.local.f32 %f321, [%rd2+-8]; ld.local.f32 %f322, [%rd2+-4]; mul.ftz.f32 %f323, %f428, %f322; mul.ftz.f32 %f324, %f429, %f322; mul.ftz.f32 %f325, %f430, %f322; mul.ftz.f32 %f326, %f431, %f322; fma.rn.ftz.f32 %f327, %f432, %f321, %f323; fma.rn.ftz.f32 %f328, %f433, %f321, %f324; fma.rn.ftz.f32 %f329, %f434, %f321, %f325; fma.rn.ftz.f32 %f330, %f435, %f321, %f326; mov.f32 %f331, 0f3F800000; sub.ftz.f32 %f332, %f331, %f321; sub.ftz.f32 %f333, %f332, %f322; fma.rn.ftz.f32 %f447, %f439, %f333, %f330; fma.rn.ftz.f32 %f446, %f438, %f333, %f329; fma.rn.ftz.f32 %f445, %f437, %f333, %f328; fma.rn.ftz.f32 %f444, %f436, %f333, %f327; $L__BB97_82: add.ftz.f32 %f410, %f444, %f443; add.ftz.f32 %f411, %f445, %f442; add.ftz.f32 %f412, %f446, %f441; add.ftz.f32 %f413, %f447, %f440; st.local.f32 [%rd21], %f410; st.local.f32 [%rd21+4], %f411; st.local.f32 [%rd21+8], %f412; st.local.f32 [%rd22], %f413; $L__BB97_84: ret; } .func _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b64 _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3, .param .b32 _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4 ) { .local .align 16 .b8 __local_depot98[32]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<56>; .reg .b16 %rs<37>; .reg .f32 %f<439>; .reg .b32 %r<138>; .reg .b64 %rd<183>; mov.u64 %SPL, __local_depot98; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd53, [_Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u64 %rd54, [_Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %rd1, [_Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r31, [_Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; ld.param.u32 %r32, [_Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj_param_4]; cvta.to.local.u64 %rd55, %rd53; add.s64 %rd2, %rd55, 76; ld.local.u32 %r1, [%rd55+76]; ld.local.s32 %rd9, [%rd55+60]; setp.eq.s32 %p1, %r1, -1; mov.u32 %r22, -1; mov.u32 %r19, 0; mov.u32 %r135, %r19; mov.u32 %r136, %r19; @%p1 bra $L__BB98_15; ld.const.u64 %rd56, [kernel_params+2160]; cvta.to.global.u64 %rd57, %rd56; mul.wide.s32 %rd58, %r1, 208; add.s64 %rd59, %rd57, %rd58; ld.global.u32 %r2, [%rd59+164]; ld.local.u32 %r38, [%rd2+-12]; and.b32 %r39, %r38, 1; setp.eq.b32 %p2, %r39, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r129, 0; @%p5 bra $L__BB98_5; cvt.u32.u64 %r41, %rd9; setp.eq.s32 %p6, %r41, -1; mov.u32 %r128, -1; @%p6 bra $L__BB98_4; ld.const.u64 %rd60, [kernel_params+2240]; cvta.to.global.u64 %rd61, %rd60; shl.b64 %rd62, %rd9, 2; add.s64 %rd63, %rd61, %rd62; ld.global.u32 %r128, [%rd63]; $L__BB98_4: setp.ne.s32 %p7, %r128, -1; selp.u32 %r129, 1, 0, %p7; $L__BB98_5: ld.const.u64 %rd64, [kernel_params+2312]; cvta.to.global.u64 %rd10, %rd64; add.s32 %r131, %r129, %r2; mul.wide.u32 %rd65, %r131, 16; add.s64 %rd66, %rd10, %rd65; add.s64 %rd172, %rd66, 14; add.s64 %rd173, %rd66, 15; ld.global.u32 %r130, [%rd66+8]; ld.global.u16 %rs35, [%rd66+12]; ld.global.u64 %rd171, [%rd66]; setp.eq.s64 %p8, %rd171, %rd1; @%p8 bra $L__BB98_11; $L__BB98_6: setp.eq.s64 %p9, %rd171, 0; @%p9 bra $L__BB98_8; add.s32 %r131, %r131, 2; bra.uni $L__BB98_9; $L__BB98_8: setp.eq.s16 %p10, %rs35, 0; mov.u32 %r19, 0; mov.u32 %r131, %r130; mov.u32 %r135, %r19; mov.u32 %r136, %r19; @%p10 bra $L__BB98_15; $L__BB98_9: cvt.u64.u32 %rd15, %r131; mul.wide.u32 %rd67, %r131, 16; add.s64 %rd68, %rd10, %rd67; ld.global.u32 %r130, [%rd68+8]; ld.global.u16 %rs35, [%rd68+12]; ld.global.u64 %rd171, [%rd68]; setp.ne.s64 %p11, %rd171, %rd1; @%p11 bra $L__BB98_6; shl.b64 %rd69, %rd15, 4; add.s64 %rd70, %rd10, %rd69; add.s64 %rd172, %rd70, 14; add.s64 %rd173, %rd70, 15; $L__BB98_11: ld.global.u8 %rs5, [%rd173]; ld.global.u8 %rs6, [%rd172]; cvt.u32.u64 %r46, %rd9; setp.ne.s32 %p12, %r46, -1; @%p12 bra $L__BB98_14; mov.u32 %r19, 0; add.s16 %rs19, %rs35, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB98_14; setp.ne.s16 %p14, %rs35, 1024; mov.u32 %r135, %r19; mov.u32 %r136, %r19; @%p14 bra $L__BB98_15; $L__BB98_14: setp.eq.s16 %p15, %rs35, 0; selp.b32 %r22, -1, %r130, %p15; cvt.u32.u16 %r19, %rs35; cvt.u32.u16 %r51, %rs5; and.b32 %r136, %r51, 255; cvt.u32.u16 %r52, %rs6; and.b32 %r135, %r52, 255; $L__BB98_15: cvta.to.local.u64 %rd71, %rd54; setp.eq.s32 %p16, %r22, -1; mul.wide.u32 %rd72, %r31, 4; add.s64 %rd21, %rd71, %rd72; mul.wide.u32 %rd73, %r32, 4; add.s64 %rd22, %rd71, %rd73; @%p16 bra $L__BB98_83; add.s32 %r53, %r135, -3; setp.lt.u32 %p17, %r53, 2; ld.local.u32 %r29, [%rd2+-12]; and.b32 %r30, %r29, 1; @%p17 bra $L__BB98_36; bra.uni $L__BB98_17; $L__BB98_36: setp.eq.s32 %p32, %r30, 0; @%p32 bra $L__BB98_75; bra.uni $L__BB98_37; $L__BB98_75: and.b32 %r114, %r29, 6; setp.eq.s32 %p49, %r114, 0; @%p49 bra $L__BB98_80; and.b32 %r115, %r19, 768; setp.eq.s32 %p50, %r115, 0; @%p50 bra $L__BB98_78; ld.const.u64 %rd149, [kernel_params+2264]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd9, 4; add.s64 %rd152, %rd150, %rd151; shr.s32 %r116, %r29, 7; ld.global.u32 %r117, [%rd152+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd153, [kernel_params+2344]; cvta.to.global.u64 %rd154, %rd153; add.s32 %r119, %r118, %r22; mul.wide.s32 %rd155, %r119, 16; add.s64 %rd156, %rd154, %rd155; ld.global.v4.f32 {%f346, %f347, %f348, %f349}, [%rd156]; ld.global.v4.f32 {%f354, %f355, %f356, %f357}, [%rd156+16]; add.u64 %rd158, %SPL, 0; mov.f32 %f431, 0f00000000; st.local.v4.f32 [%rd158], {%f431, %f431, %f431, %f431}; ld.local.f32 %f362, [%rd2+-8]; mov.f32 %f363, 0f3F800000; sub.ftz.f32 %f364, %f363, %f362; mul.ftz.f32 %f365, %f346, %f364; mul.ftz.f32 %f366, %f347, %f364; mul.ftz.f32 %f367, %f348, %f364; mul.ftz.f32 %f368, %f349, %f364; fma.rn.ftz.f32 %f438, %f357, %f362, %f368; fma.rn.ftz.f32 %f437, %f356, %f362, %f367; fma.rn.ftz.f32 %f436, %f355, %f362, %f366; fma.rn.ftz.f32 %f435, %f354, %f362, %f365; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; bra.uni $L__BB98_82; $L__BB98_83: mov.u32 %r127, 0; st.local.u32 [%rd21], %r127; st.local.u32 [%rd21+4], %r127; st.local.u32 [%rd21+8], %r127; st.local.u32 [%rd22], %r127; bra.uni $L__BB98_84; $L__BB98_17: setp.eq.s32 %p18, %r30, 0; add.u64 %rd74, %SP, 16; add.u64 %rd23, %SPL, 16; @%p18 bra $L__BB98_28; bra.uni $L__BB98_18; $L__BB98_28: and.b32 %r77, %r29, 6; setp.eq.s32 %p25, %r77, 0; @%p25 bra $L__BB98_33; and.b32 %r78, %r19, 768; setp.eq.s32 %p26, %r78, 0; @%p26 bra $L__BB98_31; ld.const.u64 %rd99, [kernel_params+2264]; cvta.to.global.u64 %rd100, %rd99; shl.b64 %rd101, %rd9, 4; add.s64 %rd102, %rd100, %rd101; shr.s32 %r79, %r29, 7; ld.global.u32 %r80, [%rd102+4]; add.s32 %r81, %r79, %r80; ld.const.u64 %rd103, [kernel_params+2336]; cvta.to.global.u64 %rd104, %rd103; add.s32 %r82, %r81, %r22; mul.wide.s32 %rd105, %r82, 12; add.s64 %rd106, %rd104, %rd105; ld.global.f32 %f189, [%rd106]; ld.global.f32 %f190, [%rd106+4]; ld.global.f32 %f191, [%rd106+8]; ld.global.f32 %f192, [%rd106+12]; ld.global.f32 %f193, [%rd106+16]; ld.global.f32 %f194, [%rd106+20]; mov.f32 %f405, 0f00000000; mov.u32 %r83, 0; st.local.u32 [%rd23], %r83; st.local.u32 [%rd23+4], %r83; st.local.u32 [%rd23+8], %r83; ld.local.f32 %f195, [%rd2+-8]; mov.f32 %f196, 0f3F800000; sub.ftz.f32 %f197, %f196, %f195; mul.ftz.f32 %f198, %f189, %f197; mul.ftz.f32 %f199, %f190, %f197; mul.ftz.f32 %f200, %f191, %f197; fma.rn.ftz.f32 %f410, %f194, %f195, %f200; fma.rn.ftz.f32 %f409, %f193, %f195, %f199; fma.rn.ftz.f32 %f408, %f192, %f195, %f198; mov.f32 %f406, %f405; mov.f32 %f407, %f405; bra.uni $L__BB98_35; $L__BB98_37: cvt.u32.u64 %r94, %rd9; setp.eq.s32 %p33, %r94, -1; @%p33 bra $L__BB98_40; ld.const.u64 %rd115, [kernel_params+2240]; cvta.to.global.u64 %rd116, %rd115; shl.b64 %rd117, %rd9, 2; add.s64 %rd118, %rd116, %rd117; ld.global.u32 %r95, [%rd118]; setp.eq.s32 %p34, %r95, -1; @%p34 bra $L__BB98_40; mov.u64 %rd119, 0; add.u64 %rd120, %SP, 0; { // callseq 73, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd53; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r19; st.param.b32 [param1+4], %r135; st.param.b32 [param1+8], %r136; st.param.b32 [param1+12], %r22; .param .b64 param2; st.param.b64 [param2+0], %rd119; .param .b64 param3; st.param.b64 [param3+0], %rd120; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f435, [retval0+0]; ld.param.f32 %f436, [retval0+4]; ld.param.f32 %f437, [retval0+8]; ld.param.f32 %f438, [retval0+12]; } // callseq 73 add.u64 %rd121, %SPL, 0; ld.local.v4.f32 {%f434, %f433, %f432, %f431}, [%rd121]; bra.uni $L__BB98_82; $L__BB98_18: cvt.u32.u64 %r54, %rd9; setp.eq.s32 %p19, %r54, -1; @%p19 bra $L__BB98_21; ld.const.u64 %rd75, [kernel_params+2240]; cvta.to.global.u64 %rd76, %rd75; shl.b64 %rd77, %rd9, 2; add.s64 %rd78, %rd76, %rd77; ld.global.u32 %r55, [%rd78]; setp.eq.s32 %p20, %r55, -1; @%p20 bra $L__BB98_21; mov.u64 %rd79, 0; { // callseq 72, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd53; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r19; st.param.b32 [param1+4], %r135; st.param.b32 [param1+8], %r136; st.param.b32 [param1+12], %r22; .param .b64 param2; st.param.b64 [param2+0], %rd79; .param .b64 param3; st.param.b64 [param3+0], %rd74; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f408, [retval0+0]; ld.param.f32 %f409, [retval0+4]; ld.param.f32 %f410, [retval0+8]; } // callseq 72 ld.local.f32 %f407, [%rd23]; ld.local.f32 %f406, [%rd23+4]; ld.local.f32 %f405, [%rd23+8]; bra.uni $L__BB98_35; $L__BB98_40: and.b32 %r96, %r19, 120; setp.eq.s32 %p35, %r96, 0; @%p35 bra $L__BB98_73; and.b32 %r97, %r19, 24; setp.eq.s32 %p36, %r97, 0; @%p36 bra $L__BB98_43; ld.const.u64 %rd122, [kernel_params+2232]; cvta.to.global.u64 %rd123, %rd122; shl.b64 %rd124, %rd9, 4; add.s64 %rd125, %rd123, %rd124; ld.global.v4.u32 {%r98, %r99, %r100, %r101}, [%rd125]; ld.const.u64 %rd126, [kernel_params+2344]; cvta.to.global.u64 %rd127, %rd126; add.s32 %r105, %r98, %r22; mul.wide.u32 %rd128, %r105, 16; add.s64 %rd129, %rd127, %rd128; ld.global.v4.f32 {%f427, %f428, %f429, %f430}, [%rd129]; add.s32 %r106, %r99, %r22; mul.wide.u32 %rd130, %r106, 16; add.s64 %rd131, %rd127, %rd130; ld.global.v4.f32 {%f423, %f424, %f425, %f426}, [%rd131]; add.s32 %r107, %r100, %r22; mul.wide.u32 %rd132, %r107, 16; add.s64 %rd133, %rd127, %rd132; ld.global.v4.f32 {%f419, %f420, %f421, %f422}, [%rd133]; bra.uni $L__BB98_72; $L__BB98_21: and.b32 %r56, %r19, 56; setp.eq.s32 %p21, %r56, 0; @%p21 bra $L__BB98_26; and.b32 %r57, %r19, 24; setp.eq.s32 %p22, %r57, 0; @%p22 bra $L__BB98_24; ld.const.u64 %rd81, [kernel_params+2232]; cvta.to.global.u64 %rd82, %rd81; shl.b64 %rd83, %rd9, 4; add.s64 %rd84, %rd82, %rd83; ld.global.v4.u32 {%r58, %r59, %r60, %r61}, [%rd84]; ld.const.u64 %rd85, [kernel_params+2336]; cvta.to.global.u64 %rd86, %rd85; add.s32 %r65, %r58, %r22; mul.wide.u32 %rd87, %r65, 12; add.s64 %rd180, %rd86, %rd87; add.s64 %rd181, %rd180, 4; add.s64 %rd182, %rd180, 8; add.s32 %r66, %r59, %r22; mul.wide.u32 %rd88, %r66, 12; add.s64 %rd177, %rd86, %rd88; add.s64 %rd178, %rd177, 4; add.s64 %rd179, %rd177, 8; add.s32 %r67, %r60, %r22; mul.wide.u32 %rd89, %r67, 12; add.s64 %rd174, %rd86, %rd89; add.s64 %rd175, %rd174, 4; add.s64 %rd176, %rd174, 8; bra.uni $L__BB98_25; $L__BB98_80: and.b32 %r124, %r29, 8; setp.eq.s32 %p53, %r124, 0; add.u64 %rd166, %SPL, 0; mov.f32 %f431, 0f00000000; st.local.v4.f32 [%rd166], {%f431, %f431, %f431, %f431}; setp.ne.s32 %p54, %r19, 8; or.pred %p55, %p54, %p53; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; mov.f32 %f435, %f431; mov.f32 %f436, %f431; mov.f32 %f437, %f431; mov.f32 %f438, %f431; @%p55 bra $L__BB98_82; cvt.u32.u64 %r125, %rd9; ld.const.u64 %rd167, [kernel_params+2344]; cvta.to.global.u64 %rd168, %rd167; add.s32 %r126, %r125, %r22; mul.wide.s32 %rd169, %r126, 16; add.s64 %rd170, %rd168, %rd169; ld.global.v4.f32 {%f435, %f436, %f437, %f438}, [%rd170]; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; bra.uni $L__BB98_82; $L__BB98_33: and.b32 %r89, %r29, 8; setp.eq.s32 %p29, %r89, 0; mov.u32 %r90, 0; st.local.u32 [%rd23], %r90; st.local.u32 [%rd23+4], %r90; st.local.u32 [%rd23+8], %r90; setp.ne.s32 %p30, %r19, 8; mov.f32 %f405, 0f00000000; or.pred %p31, %p30, %p29; mov.f32 %f406, %f405; mov.f32 %f407, %f405; mov.f32 %f408, %f405; mov.f32 %f409, %f405; mov.f32 %f410, %f405; @%p31 bra $L__BB98_35; cvt.u32.u64 %r91, %rd9; ld.const.u64 %rd111, [kernel_params+2336]; cvta.to.global.u64 %rd112, %rd111; add.s32 %r92, %r91, %r22; mul.wide.s32 %rd113, %r92, 12; add.s64 %rd114, %rd112, %rd113; ld.global.f32 %f410, [%rd114+8]; ld.global.f32 %f409, [%rd114+4]; ld.global.f32 %f408, [%rd114]; mov.f32 %f406, %f405; mov.f32 %f407, %f405; bra.uni $L__BB98_35; $L__BB98_78: add.u64 %rd160, %SPL, 0; mov.f32 %f431, 0f00000000; st.local.v4.f32 [%rd160], {%f431, %f431, %f431, %f431}; and.b32 %r120, %r19, 131; setp.eq.s32 %p51, %r120, 0; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; mov.f32 %f435, %f431; mov.f32 %f436, %f431; mov.f32 %f437, %f431; mov.f32 %f438, %f431; @%p51 bra $L__BB98_82; cvt.u32.u64 %r121, %rd9; setp.eq.s32 %p52, %r19, 128; selp.b32 %r122, %r121, 0, %p52; add.s32 %r123, %r22, %r122; ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r123, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f435, %f436, %f437, %f438}, [%rd164]; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; bra.uni $L__BB98_82; $L__BB98_31: mov.u32 %r84, 0; st.local.u32 [%rd23], %r84; st.local.u32 [%rd23+4], %r84; st.local.u32 [%rd23+8], %r84; and.b32 %r85, %r19, 131; setp.eq.s32 %p27, %r85, 0; mov.f32 %f405, 0f00000000; mov.f32 %f406, %f405; mov.f32 %f407, %f405; mov.f32 %f408, %f405; mov.f32 %f409, %f405; mov.f32 %f410, %f405; @%p27 bra $L__BB98_35; cvt.u32.u64 %r86, %rd9; setp.eq.s32 %p28, %r19, 128; selp.b32 %r87, %r86, 0, %p28; add.s32 %r88, %r22, %r87; ld.const.u64 %rd107, [kernel_params+2336]; cvta.to.global.u64 %rd108, %rd107; mul.wide.s32 %rd109, %r88, 12; add.s64 %rd110, %rd108, %rd109; ld.global.f32 %f410, [%rd110+8]; ld.global.f32 %f409, [%rd110+4]; ld.global.f32 %f408, [%rd110]; mov.f32 %f406, %f405; mov.f32 %f407, %f405; bra.uni $L__BB98_35; $L__BB98_73: add.u64 %rd144, %SPL, 0; mov.f32 %f431, 0f00000000; st.local.v4.f32 [%rd144], {%f431, %f431, %f431, %f431}; and.b32 %r110, %r19, 7; setp.eq.s32 %p47, %r110, 0; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; mov.f32 %f435, %f431; mov.f32 %f436, %f431; mov.f32 %f437, %f431; mov.f32 %f438, %f431; @%p47 bra $L__BB98_82; setp.eq.s32 %p48, %r19, 4; selp.b32 %r112, %r94, 0, %p48; add.s32 %r113, %r22, %r112; ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r113, 16; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f435, %f436, %f437, %f438}, [%rd148]; mov.f32 %f432, %f431; mov.f32 %f433, %f431; mov.f32 %f434, %f431; bra.uni $L__BB98_82; $L__BB98_26: mov.u32 %r72, 0; st.local.u32 [%rd23], %r72; st.local.u32 [%rd23+4], %r72; st.local.u32 [%rd23+8], %r72; and.b32 %r73, %r19, 7; setp.eq.s32 %p23, %r73, 0; mov.f32 %f405, 0f00000000; mov.f32 %f406, %f405; mov.f32 %f407, %f405; mov.f32 %f408, %f405; mov.f32 %f409, %f405; mov.f32 %f410, %f405; @%p23 bra $L__BB98_35; setp.eq.s32 %p24, %r19, 4; selp.b32 %r75, %r54, 0, %p24; add.s32 %r76, %r22, %r75; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; mul.wide.s32 %rd97, %r76, 12; add.s64 %rd98, %rd96, %rd97; ld.global.f32 %f410, [%rd98+8]; ld.global.f32 %f409, [%rd98+4]; ld.global.f32 %f408, [%rd98]; mov.f32 %f406, %f405; mov.f32 %f407, %f405; bra.uni $L__BB98_35; $L__BB98_43: mad.lo.s32 %r109, %r94, 3, %r22; cvt.s64.s32 %rd51, %r109; setp.eq.s32 %p37, %r19, 32; @%p37 bra $L__BB98_71; bra.uni $L__BB98_44; $L__BB98_71: ld.const.u64 %rd137, [kernel_params+2344]; cvta.to.global.u64 %rd138, %rd137; shl.b64 %rd139, %rd51, 4; add.s64 %rd140, %rd138, %rd139; ld.global.v4.f32 {%f427, %f428, %f429, %f430}, [%rd140]; ld.global.v4.f32 {%f423, %f424, %f425, %f426}, [%rd140+16]; ld.global.v4.f32 {%f419, %f420, %f421, %f422}, [%rd140+32]; bra.uni $L__BB98_72; $L__BB98_24: mad.lo.s32 %r69, %r54, 3, %r22; ld.const.u64 %rd90, [kernel_params+2336]; cvta.to.global.u64 %rd91, %rd90; mul.wide.s32 %rd92, %r69, 12; add.s64 %rd180, %rd91, %rd92; add.s64 %rd181, %rd180, 4; add.s64 %rd182, %rd180, 8; add.s32 %r70, %r69, 1; mul.wide.s32 %rd93, %r70, 12; add.s64 %rd177, %rd91, %rd93; add.s64 %rd178, %rd177, 4; add.s64 %rd179, %rd177, 8; add.s32 %r71, %r69, 2; mul.wide.s32 %rd94, %r71, 12; add.s64 %rd174, %rd91, %rd94; add.s64 %rd175, %rd174, 4; add.s64 %rd176, %rd174, 8; $L__BB98_25: ld.local.f32 %f145, [%rd2+28]; ld.global.f32 %f146, [%rd177]; ld.global.f32 %f147, [%rd178]; ld.global.f32 %f148, [%rd179]; ld.local.f32 %f149, [%rd2+36]; ld.global.f32 %f150, [%rd174]; mul.ftz.f32 %f151, %f150, %f149; ld.global.f32 %f152, [%rd175]; mul.ftz.f32 %f153, %f152, %f149; ld.global.f32 %f154, [%rd176]; mul.ftz.f32 %f155, %f154, %f149; fma.rn.ftz.f32 %f156, %f146, %f145, %f151; fma.rn.ftz.f32 %f157, %f147, %f145, %f153; fma.rn.ftz.f32 %f158, %f148, %f145, %f155; add.ftz.f32 %f159, %f145, %f149; ld.global.f32 %f160, [%rd180]; mul.ftz.f32 %f161, %f160, %f159; ld.global.f32 %f162, [%rd181]; mul.ftz.f32 %f163, %f162, %f159; ld.global.f32 %f164, [%rd182]; mul.ftz.f32 %f165, %f164, %f159; sub.ftz.f32 %f407, %f156, %f161; sub.ftz.f32 %f406, %f157, %f163; sub.ftz.f32 %f405, %f158, %f165; st.local.f32 [%rd23], %f407; st.local.f32 [%rd23+4], %f406; st.local.f32 [%rd23+8], %f405; ld.local.f32 %f166, [%rd2+-8]; ld.local.f32 %f167, [%rd2+-4]; mul.ftz.f32 %f168, %f150, %f167; mul.ftz.f32 %f169, %f152, %f167; mul.ftz.f32 %f170, %f154, %f167; fma.rn.ftz.f32 %f171, %f146, %f166, %f168; fma.rn.ftz.f32 %f172, %f147, %f166, %f169; fma.rn.ftz.f32 %f173, %f148, %f166, %f170; mov.f32 %f174, 0f3F800000; sub.ftz.f32 %f175, %f174, %f166; sub.ftz.f32 %f176, %f175, %f167; fma.rn.ftz.f32 %f410, %f164, %f176, %f173; fma.rn.ftz.f32 %f409, %f162, %f176, %f172; fma.rn.ftz.f32 %f408, %f160, %f176, %f171; $L__BB98_35: add.ftz.f32 %f219, %f408, %f407; add.ftz.f32 %f220, %f409, %f406; add.ftz.f32 %f221, %f410, %f405; st.local.f32 [%rd21], %f219; st.local.f32 [%rd21+4], %f220; st.local.f32 [%rd21+8], %f221; mov.u32 %r93, 1065353216; st.local.u32 [%rd22], %r93; bra.uni $L__BB98_84; $L__BB98_44: ld.const.u64 %rd134, [kernel_params+2352]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd51, 2; add.s64 %rd52, %rd135, %rd136; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd52]; cvt.rn.f32.u16 %f238, %rs20; mul.ftz.f32 %f51, %f238, 0f3B808081; cvt.rn.f32.u16 %f239, %rs21; mul.ftz.f32 %f52, %f239, 0f3B808081; cvt.rn.f32.u16 %f240, %rs22; mul.ftz.f32 %f53, %f240, 0f3B808081; setp.lt.ftz.f32 %p38, %f51, 0f3D25AEE6; @%p38 bra $L__BB98_46; bra.uni $L__BB98_45; $L__BB98_46: mul.ftz.f32 %f427, %f51, 0f3D9E8391; bra.uni $L__BB98_47; $L__BB98_45: add.ftz.f32 %f241, %f51, 0f3D6147AE; mul.ftz.f32 %f242, %f241, 0f3F72A76F; lg2.approx.ftz.f32 %f243, %f242; mul.ftz.f32 %f244, %f243, 0f4019999A; ex2.approx.ftz.f32 %f427, %f244; $L__BB98_47: setp.lt.ftz.f32 %p39, %f52, 0f3D25AEE6; @%p39 bra $L__BB98_49; bra.uni $L__BB98_48; $L__BB98_49: mul.ftz.f32 %f428, %f52, 0f3D9E8391; bra.uni $L__BB98_50; $L__BB98_48: add.ftz.f32 %f245, %f52, 0f3D6147AE; mul.ftz.f32 %f246, %f245, 0f3F72A76F; lg2.approx.ftz.f32 %f247, %f246; mul.ftz.f32 %f248, %f247, 0f4019999A; ex2.approx.ftz.f32 %f428, %f248; $L__BB98_50: setp.lt.ftz.f32 %p40, %f53, 0f3D25AEE6; @%p40 bra $L__BB98_52; bra.uni $L__BB98_51; $L__BB98_52: mul.ftz.f32 %f429, %f53, 0f3D9E8391; bra.uni $L__BB98_53; $L__BB98_51: add.ftz.f32 %f249, %f53, 0f3D6147AE; mul.ftz.f32 %f250, %f249, 0f3F72A76F; lg2.approx.ftz.f32 %f251, %f250; mul.ftz.f32 %f252, %f251, 0f4019999A; ex2.approx.ftz.f32 %f429, %f252; $L__BB98_53: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd52+4]; cvt.rn.f32.u16 %f253, %rs24; mul.ftz.f32 %f63, %f253, 0f3B808081; cvt.rn.f32.u16 %f254, %rs25; mul.ftz.f32 %f64, %f254, 0f3B808081; cvt.rn.f32.u16 %f255, %rs26; mul.ftz.f32 %f65, %f255, 0f3B808081; setp.lt.ftz.f32 %p41, %f63, 0f3D25AEE6; @%p41 bra $L__BB98_55; bra.uni $L__BB98_54; $L__BB98_55: mul.ftz.f32 %f423, %f63, 0f3D9E8391; bra.uni $L__BB98_56; $L__BB98_54: add.ftz.f32 %f256, %f63, 0f3D6147AE; mul.ftz.f32 %f257, %f256, 0f3F72A76F; lg2.approx.ftz.f32 %f258, %f257; mul.ftz.f32 %f259, %f258, 0f4019999A; ex2.approx.ftz.f32 %f423, %f259; $L__BB98_56: setp.lt.ftz.f32 %p42, %f64, 0f3D25AEE6; @%p42 bra $L__BB98_58; bra.uni $L__BB98_57; $L__BB98_58: mul.ftz.f32 %f424, %f64, 0f3D9E8391; bra.uni $L__BB98_59; $L__BB98_57: add.ftz.f32 %f260, %f64, 0f3D6147AE; mul.ftz.f32 %f261, %f260, 0f3F72A76F; lg2.approx.ftz.f32 %f262, %f261; mul.ftz.f32 %f263, %f262, 0f4019999A; ex2.approx.ftz.f32 %f424, %f263; $L__BB98_59: setp.lt.ftz.f32 %p43, %f65, 0f3D25AEE6; @%p43 bra $L__BB98_61; bra.uni $L__BB98_60; $L__BB98_61: mul.ftz.f32 %f425, %f65, 0f3D9E8391; bra.uni $L__BB98_62; $L__BB98_60: add.ftz.f32 %f264, %f65, 0f3D6147AE; mul.ftz.f32 %f265, %f264, 0f3F72A76F; lg2.approx.ftz.f32 %f266, %f265; mul.ftz.f32 %f267, %f266, 0f4019999A; ex2.approx.ftz.f32 %f425, %f267; $L__BB98_62: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd52+8]; cvt.rn.f32.u16 %f268, %rs28; mul.ftz.f32 %f75, %f268, 0f3B808081; cvt.rn.f32.u16 %f269, %rs29; mul.ftz.f32 %f76, %f269, 0f3B808081; cvt.rn.f32.u16 %f270, %rs30; mul.ftz.f32 %f77, %f270, 0f3B808081; setp.lt.ftz.f32 %p44, %f75, 0f3D25AEE6; @%p44 bra $L__BB98_64; bra.uni $L__BB98_63; $L__BB98_64: mul.ftz.f32 %f419, %f75, 0f3D9E8391; bra.uni $L__BB98_65; $L__BB98_63: add.ftz.f32 %f271, %f75, 0f3D6147AE; mul.ftz.f32 %f272, %f271, 0f3F72A76F; lg2.approx.ftz.f32 %f273, %f272; mul.ftz.f32 %f274, %f273, 0f4019999A; ex2.approx.ftz.f32 %f419, %f274; $L__BB98_65: setp.lt.ftz.f32 %p45, %f76, 0f3D25AEE6; @%p45 bra $L__BB98_67; bra.uni $L__BB98_66; $L__BB98_67: mul.ftz.f32 %f420, %f76, 0f3D9E8391; bra.uni $L__BB98_68; $L__BB98_66: add.ftz.f32 %f275, %f76, 0f3D6147AE; mul.ftz.f32 %f276, %f275, 0f3F72A76F; lg2.approx.ftz.f32 %f277, %f276; mul.ftz.f32 %f278, %f277, 0f4019999A; ex2.approx.ftz.f32 %f420, %f278; $L__BB98_68: cvt.rn.f32.u16 %f279, %rs23; mul.ftz.f32 %f430, %f279, 0f3B808081; cvt.rn.f32.u16 %f280, %rs27; mul.ftz.f32 %f426, %f280, 0f3B808081; cvt.rn.f32.u16 %f281, %rs31; mul.ftz.f32 %f422, %f281, 0f3B808081; setp.lt.ftz.f32 %p46, %f77, 0f3D25AEE6; @%p46 bra $L__BB98_70; bra.uni $L__BB98_69; $L__BB98_70: mul.ftz.f32 %f421, %f77, 0f3D9E8391; bra.uni $L__BB98_72; $L__BB98_69: add.ftz.f32 %f282, %f77, 0f3D6147AE; mul.ftz.f32 %f283, %f282, 0f3F72A76F; lg2.approx.ftz.f32 %f284, %f283; mul.ftz.f32 %f285, %f284, 0f4019999A; ex2.approx.ftz.f32 %f421, %f285; $L__BB98_72: ld.local.f32 %f298, [%rd2+28]; ld.local.f32 %f299, [%rd2+36]; mul.ftz.f32 %f300, %f419, %f299; mul.ftz.f32 %f301, %f420, %f299; mul.ftz.f32 %f302, %f421, %f299; mul.ftz.f32 %f303, %f422, %f299; fma.rn.ftz.f32 %f304, %f423, %f298, %f300; fma.rn.ftz.f32 %f305, %f424, %f298, %f301; fma.rn.ftz.f32 %f306, %f425, %f298, %f302; fma.rn.ftz.f32 %f307, %f426, %f298, %f303; add.ftz.f32 %f308, %f298, %f299; mul.ftz.f32 %f309, %f427, %f308; mul.ftz.f32 %f310, %f428, %f308; mul.ftz.f32 %f311, %f429, %f308; mul.ftz.f32 %f312, %f430, %f308; sub.ftz.f32 %f434, %f304, %f309; sub.ftz.f32 %f433, %f305, %f310; sub.ftz.f32 %f432, %f306, %f311; sub.ftz.f32 %f431, %f307, %f312; add.u64 %rd142, %SPL, 0; st.local.v4.f32 [%rd142], {%f434, %f433, %f432, %f431}; ld.local.f32 %f313, [%rd2+-8]; ld.local.f32 %f314, [%rd2+-4]; mul.ftz.f32 %f315, %f419, %f314; mul.ftz.f32 %f316, %f420, %f314; mul.ftz.f32 %f317, %f421, %f314; mul.ftz.f32 %f318, %f422, %f314; fma.rn.ftz.f32 %f319, %f423, %f313, %f315; fma.rn.ftz.f32 %f320, %f424, %f313, %f316; fma.rn.ftz.f32 %f321, %f425, %f313, %f317; fma.rn.ftz.f32 %f322, %f426, %f313, %f318; mov.f32 %f323, 0f3F800000; sub.ftz.f32 %f324, %f323, %f313; sub.ftz.f32 %f325, %f324, %f314; fma.rn.ftz.f32 %f438, %f430, %f325, %f322; fma.rn.ftz.f32 %f437, %f429, %f325, %f321; fma.rn.ftz.f32 %f436, %f428, %f325, %f320; fma.rn.ftz.f32 %f435, %f427, %f325, %f319; $L__BB98_82: add.ftz.f32 %f401, %f435, %f434; add.ftz.f32 %f402, %f436, %f433; add.ftz.f32 %f403, %f437, %f432; add.ftz.f32 %f404, %f438, %f431; st.local.f32 [%rd21], %f401; st.local.f32 [%rd21+4], %f402; st.local.f32 [%rd21+8], %f403; st.local.f32 [%rd22], %f404; $L__BB98_84: ret; } .func (.param .b32 func_retval0) _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<30>; .reg .b16 %rs<12>; .reg .f32 %f<260>; .reg .b32 %r<72>; .reg .b64 %rd<62>; ld.param.u64 %rd18, [_Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd19, [_Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r71, [_Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r32, [_Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r31, [_Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; bfe.u32 %r1, %r32, 8, 8; bfe.u32 %r2, %r32, 16, 8; cvta.to.local.u64 %rd1, %rd19; shl.b32 %r35, %r32, 2; cvt.u64.u32 %rd20, %r35; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd1, %rd21; ld.local.f32 %f1, [%rd22]; ld.local.f32 %f2, [%rd22+4]; ld.local.f32 %f3, [%rd22+8]; setp.lt.u32 %p1, %r32, 16777216; @%p1 bra $L__BB99_2; bra.uni $L__BB99_1; $L__BB99_2: cvta.to.local.u64 %rd27, %rd18; add.s64 %rd2, %rd27, 76; ld.local.s32 %rd3, [%rd27+76]; ld.local.s32 %rd4, [%rd27+60]; setp.eq.s64 %p2, %rd3, -1; mov.u32 %r70, -1; @%p2 bra $L__BB99_16; ld.const.u64 %rd28, [kernel_params+2160]; cvta.to.global.u64 %rd29, %rd28; mul.lo.s64 %rd30, %rd3, 208; add.s64 %rd31, %rd29, %rd30; ld.global.u32 %r4, [%rd31+164]; ld.local.u32 %r41, [%rd2+-12]; and.b32 %r42, %r41, 1; setp.eq.b32 %p3, %r42, 1; mov.pred %p4, 0; xor.pred %p5, %p3, %p4; not.pred %p6, %p5; mov.u32 %r65, 0; @%p6 bra $L__BB99_7; cvt.u32.u64 %r44, %rd4; setp.eq.s32 %p7, %r44, -1; mov.u32 %r64, -1; @%p7 bra $L__BB99_6; ld.const.u64 %rd32, [kernel_params+2240]; cvta.to.global.u64 %rd33, %rd32; shl.b64 %rd34, %rd4, 2; add.s64 %rd35, %rd33, %rd34; ld.global.u32 %r64, [%rd35]; $L__BB99_6: setp.ne.s32 %p8, %r64, -1; selp.u32 %r65, 1, 0, %p8; $L__BB99_7: ld.const.u64 %rd36, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd36; add.s32 %r67, %r65, %r4; mul.wide.u32 %rd37, %r67, 16; add.s64 %rd38, %rd5, %rd37; ld.global.u16 %rs10, [%rd38+12]; ld.global.u32 %r66, [%rd38+8]; ld.global.u64 %rd61, [%rd38]; setp.eq.s64 %p9, %rd61, 8; @%p9 bra $L__BB99_12; $L__BB99_8: setp.eq.s64 %p10, %rd61, 0; @%p10 bra $L__BB99_10; add.s32 %r67, %r67, 2; bra.uni $L__BB99_11; $L__BB99_10: setp.eq.s16 %p11, %rs10, 0; mov.u32 %r67, %r66; @%p11 bra $L__BB99_16; $L__BB99_11: mul.wide.u32 %rd39, %r67, 16; add.s64 %rd40, %rd5, %rd39; ld.global.u16 %rs10, [%rd40+12]; ld.global.u32 %r66, [%rd40+8]; ld.global.u64 %rd61, [%rd40]; setp.ne.s64 %p12, %rd61, 8; @%p12 bra $L__BB99_8; $L__BB99_12: cvt.u32.u64 %r49, %rd4; setp.ne.s32 %p13, %r49, -1; @%p13 bra $L__BB99_15; add.s16 %rs7, %rs10, -1; setp.lt.u16 %p14, %rs7, 2; @%p14 bra $L__BB99_15; setp.ne.s16 %p15, %rs10, 1024; @%p15 bra $L__BB99_16; $L__BB99_15: setp.eq.s16 %p16, %rs10, 0; selp.b32 %r70, -1, %r66, %p16; $L__BB99_16: ld.local.u8 %rs8, [%rd2+-20]; and.b16 %rs9, %rs8, 2; setp.eq.s16 %p17, %rs9, 0; @%p17 bra $L__BB99_18; bra.uni $L__BB99_17; $L__BB99_18: cvt.u32.u64 %r56, %rd3; ld.const.u64 %rd43, [kernel_params+2160]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r56, 208; add.s64 %rd46, %rd44, %rd45; ld.global.v4.f32 {%f138, %f139, %f140, %f141}, [%rd46+48]; ld.global.v4.f32 {%f146, %f147, %f148, %f149}, [%rd46+64]; ld.global.v4.f32 {%f154, %f155, %f156, %f157}, [%rd46+80]; mul.ftz.f32 %f162, %f2, %f139; fma.rn.ftz.f32 %f163, %f1, %f138, %f162; fma.rn.ftz.f32 %f164, %f3, %f140, %f163; add.ftz.f32 %f34, %f141, %f164; mul.ftz.f32 %f165, %f2, %f147; fma.rn.ftz.f32 %f166, %f1, %f146, %f165; fma.rn.ftz.f32 %f167, %f3, %f148, %f166; add.ftz.f32 %f35, %f149, %f167; mul.ftz.f32 %f168, %f2, %f155; fma.rn.ftz.f32 %f169, %f1, %f154, %f168; fma.rn.ftz.f32 %f170, %f3, %f156, %f169; add.ftz.f32 %f36, %f157, %f170; bra.uni $L__BB99_19; $L__BB99_1: ld.const.u64 %rd23, [kernel_params+2400]; cvta.to.global.u64 %rd24, %rd23; mul.wide.s32 %rd25, %r71, 16; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f72, %f73, %f74, %f75}, [%rd26]; ld.global.v4.f32 {%f80, %f81, %f82, %f83}, [%rd26+16]; ld.global.v4.f32 {%f88, %f89, %f90, %f91}, [%rd26+32]; add.s32 %r71, %r71, 3; mul.ftz.f32 %f96, %f2, %f73; fma.rn.ftz.f32 %f97, %f1, %f72, %f96; fma.rn.ftz.f32 %f98, %f3, %f74, %f97; mul.ftz.f32 %f99, %f2, %f81; fma.rn.ftz.f32 %f100, %f1, %f80, %f99; fma.rn.ftz.f32 %f101, %f3, %f82, %f100; mul.ftz.f32 %f102, %f2, %f89; fma.rn.ftz.f32 %f103, %f1, %f88, %f102; fma.rn.ftz.f32 %f104, %f3, %f90, %f103; add.ftz.f32 %f36, %f104, %f91; add.ftz.f32 %f35, %f101, %f83; add.ftz.f32 %f34, %f98, %f75; bra.uni $L__BB99_21; $L__BB99_17: ld.local.v4.f32 {%f105, %f106, %f107, %f108}, [%rd2+116]; mul.ftz.f32 %f113, %f2, %f106; fma.rn.ftz.f32 %f114, %f1, %f105, %f113; fma.rn.ftz.f32 %f115, %f3, %f107, %f114; add.ftz.f32 %f34, %f108, %f115; ld.local.v4.f32 {%f116, %f117, %f118, %f119}, [%rd2+132]; mul.ftz.f32 %f124, %f2, %f117; fma.rn.ftz.f32 %f125, %f1, %f116, %f124; fma.rn.ftz.f32 %f126, %f3, %f118, %f125; add.ftz.f32 %f35, %f119, %f126; ld.local.v4.f32 {%f127, %f128, %f129, %f130}, [%rd2+148]; mul.ftz.f32 %f135, %f2, %f128; fma.rn.ftz.f32 %f136, %f1, %f127, %f135; fma.rn.ftz.f32 %f137, %f3, %f129, %f136; add.ftz.f32 %f36, %f130, %f137; $L__BB99_19: setp.eq.s32 %p18, %r70, -1; @%p18 bra $L__BB99_21; ld.const.u64 %rd47, [kernel_params+2344]; cvta.to.global.u64 %rd48, %rd47; mul.wide.s32 %rd49, %r70, 16; add.s64 %rd50, %rd48, %rd49; ld.global.v4.f32 {%f171, %f172, %f173, %f174}, [%rd50]; ld.global.v4.f32 {%f179, %f180, %f181, %f182}, [%rd50+16]; ld.global.v4.f32 {%f187, %f188, %f189, %f190}, [%rd50+32]; mul.ftz.f32 %f195, %f35, %f172; fma.rn.ftz.f32 %f196, %f34, %f171, %f195; fma.rn.ftz.f32 %f197, %f36, %f173, %f196; add.ftz.f32 %f16, %f174, %f197; mul.ftz.f32 %f198, %f35, %f180; fma.rn.ftz.f32 %f199, %f34, %f179, %f198; fma.rn.ftz.f32 %f200, %f36, %f181, %f199; add.ftz.f32 %f17, %f182, %f200; mul.ftz.f32 %f201, %f35, %f188; fma.rn.ftz.f32 %f202, %f34, %f187, %f201; fma.rn.ftz.f32 %f203, %f36, %f189, %f202; add.ftz.f32 %f36, %f190, %f203; mov.f32 %f34, %f16; mov.f32 %f35, %f17; $L__BB99_21: ld.const.u64 %rd51, [kernel_params+2432]; cvta.to.global.u64 %rd52, %rd51; mul.wide.s32 %rd53, %r31, 96; add.s64 %rd16, %rd51, %rd53; add.s64 %rd54, %rd52, %rd53; add.s64 %rd17, %rd54, 32; ld.global.u32 %r57, [%rd54+32]; setp.eq.s32 %p19, %r57, 0; @%p19 bra $L__BB99_23; ld.global.v4.f32 {%f204, %f205, %f206, %f207}, [%rd17+16]; mul.ftz.f32 %f212, %f35, %f205; fma.rn.ftz.f32 %f213, %f34, %f204, %f212; fma.rn.ftz.f32 %f214, %f36, %f206, %f213; add.ftz.f32 %f31, %f207, %f214; ld.global.v4.f32 {%f215, %f216, %f217, %f218}, [%rd17+32]; mul.ftz.f32 %f223, %f35, %f216; fma.rn.ftz.f32 %f224, %f34, %f215, %f223; fma.rn.ftz.f32 %f225, %f36, %f217, %f224; add.ftz.f32 %f32, %f218, %f225; ld.global.v4.f32 {%f226, %f227, %f228, %f229}, [%rd17+48]; mul.ftz.f32 %f234, %f35, %f227; fma.rn.ftz.f32 %f235, %f34, %f226, %f234; fma.rn.ftz.f32 %f236, %f36, %f228, %f235; add.ftz.f32 %f36, %f229, %f236; mov.f32 %f34, %f31; mov.f32 %f35, %f32; $L__BB99_23: ld.global.v2.u32 {%r58, %r59}, [%rd17+-24]; setp.eq.s32 %p20, %r58, 8; @%p20 bra $L__BB99_38; bra.uni $L__BB99_24; $L__BB99_38: { // callseq 79, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .b32 param4; st.param.b32 [param4+0], %r59; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f256, [retval0+0]; } // callseq 79 mov.f32 %f259, 0f3F800000; mov.f32 %f257, %f256; mov.f32 %f258, %f256; bra.uni $L__BB99_39; $L__BB99_24: setp.eq.s32 %p21, %r58, 9; @%p21 bra $L__BB99_37; bra.uni $L__BB99_25; $L__BB99_37: { // callseq 78, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .b32 param4; st.param.b32 [param4+0], %r59; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f256, [retval0+0]; ld.param.f32 %f257, [retval0+4]; ld.param.f32 %f258, [retval0+8]; } // callseq 78 mov.f32 %f259, 0f3F800000; bra.uni $L__BB99_39; $L__BB99_25: setp.eq.s32 %p22, %r58, 10; @%p22 bra $L__BB99_36; bra.uni $L__BB99_26; $L__BB99_36: { // callseq 77, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .b32 param4; st.param.b32 [param4+0], %r59; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f256, [retval0+0]; } // callseq 77 mov.f32 %f259, 0f3F800000; mov.f32 %f257, %f256; mov.f32 %f258, %f256; bra.uni $L__BB99_39; $L__BB99_26: setp.eq.s32 %p23, %r58, 11; @%p23 bra $L__BB99_35; bra.uni $L__BB99_27; $L__BB99_35: { // callseq 76, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .b32 param4; st.param.b32 [param4+0], %r59; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f256, [retval0+0]; } // callseq 76 mov.f32 %f259, 0f3F800000; mov.f32 %f257, %f256; mov.f32 %f258, %f256; bra.uni $L__BB99_39; $L__BB99_27: or.b32 %r29, %r59, 1; setp.gt.u32 %p24, %r58, 6; @%p24 bra $L__BB99_32; mov.u32 %r61, 1; shl.b32 %r62, %r61, %r58; and.b32 %r63, %r62, 71; setp.ne.s32 %p25, %r63, 0; @%p25 bra $L__BB99_29; bra.uni $L__BB99_32; $L__BB99_29: setp.eq.s32 %p26, %r29, 3; @%p26 bra $L__BB99_31; bra.uni $L__BB99_30; $L__BB99_31: { // callseq 74, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f256, [retval0+0]; ld.param.f32 %f257, [retval0+4]; ld.param.f32 %f258, [retval0+8]; ld.param.f32 %f259, [retval0+12]; } // callseq 74 bra.uni $L__BB99_39; $L__BB99_30: ld.global.u64 %rd55, [%rd17+-32]; tex.3d.v4.f32.f32 {%f256, %f257, %f258, %f259}, [%rd55, {%f34, %f35, %f36, %f36}]; $L__BB99_39: setp.eq.s32 %p28, %r1, 255; @%p28 bra $L__BB99_41; mul.wide.u32 %rd57, %r1, 4; add.s64 %rd58, %rd1, %rd57; st.local.f32 [%rd58], %f259; $L__BB99_41: setp.eq.s32 %p29, %r2, 255; @%p29 bra $L__BB99_43; mul.wide.u32 %rd59, %r2, 4; add.s64 %rd60, %rd1, %rd59; st.local.f32 [%rd60], %f256; st.local.f32 [%rd60+4], %f257; st.local.f32 [%rd60+8], %f258; $L__BB99_43: st.param.b32 [func_retval0+0], %r71; ret; $L__BB99_32: setp.eq.s32 %p27, %r29, 3; @%p27 bra $L__BB99_34; bra.uni $L__BB99_33; $L__BB99_34: { // callseq 75, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd16; .param .b32 param1; st.param.f32 [param1+0], %f34; .param .b32 param2; st.param.f32 [param2+0], %f35; .param .b32 param3; st.param.f32 [param3+0], %f36; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f256, [retval0+0]; } // callseq 75 mov.f32 %f259, 0f3F800000; mov.f32 %f257, %f256; mov.f32 %f258, %f256; bra.uni $L__BB99_39; $L__BB99_33: ld.global.u64 %rd56, [%rd17+-32]; tex.3d.v4.f32.f32 {%f256, %f237, %f238, %f239}, [%rd56, {%f34, %f35, %f36, %f36}]; mov.f32 %f259, 0f3F800000; mov.f32 %f257, %f256; mov.f32 %f258, %f256; bra.uni $L__BB99_39; } .func (.param .b32 func_retval0) _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .pred %p<19>; .reg .f32 %f<80>; .reg .b32 %r<35>; .reg .b64 %rd<26>; ld.param.u64 %rd2, [_Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r21, [_Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; ld.param.u32 %r22, [_Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+12]; ld.param.u32 %r23, [_Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+8]; ld.param.u32 %r9, [_Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+4]; cvta.to.local.u64 %rd1, %rd2; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r21, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r24, %r25, %r26, %r27}, [%rd6]; ld.global.v4.u32 {%r28, %r29, %r30, %r31}, [%rd6+16]; bfe.u32 %r10, %r9, 8, 8; bfe.u32 %r11, %r9, 16, 8; shr.u32 %r12, %r9, 24; bfe.u32 %r13, %r23, 8, 8; bfe.u32 %r14, %r23, 16, 8; and.b32 %r15, %r22, 255; bfe.u32 %r16, %r22, 8, 8; bfe.u32 %r17, %r22, 16, 8; shr.u32 %r18, %r22, 24; and.b32 %r19, %r24, 255; bfe.u32 %r20, %r24, 8, 8; shl.b32 %r32, %r23, 2; cvt.u64.u32 %rd7, %r32; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd1, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r13, 255; @%p1 bra $L__BB100_2; bra.uni $L__BB100_1; $L__BB100_2: mov.b32 %f68, %r25; bra.uni $L__BB100_3; $L__BB100_1: mul.wide.u32 %rd10, %r13, 4; add.s64 %rd11, %rd1, %rd10; ld.local.f32 %f68, [%rd11]; $L__BB100_3: setp.eq.s32 %p2, %r14, 255; @%p2 bra $L__BB100_5; bra.uni $L__BB100_4; $L__BB100_5: mov.b32 %f69, %r26; bra.uni $L__BB100_6; $L__BB100_4: mul.wide.u32 %rd12, %r14, 4; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f69, [%rd13]; $L__BB100_6: setp.eq.s32 %p3, %r15, 255; @%p3 bra $L__BB100_8; bra.uni $L__BB100_7; $L__BB100_8: mov.b32 %f70, %r27; bra.uni $L__BB100_9; $L__BB100_7: mul.wide.u32 %rd14, %r15, 4; add.s64 %rd15, %rd1, %rd14; ld.local.f32 %f70, [%rd15]; $L__BB100_9: setp.eq.s32 %p4, %r16, 255; @%p4 bra $L__BB100_11; bra.uni $L__BB100_10; $L__BB100_11: mov.b32 %f71, %r28; bra.uni $L__BB100_12; $L__BB100_10: mul.wide.u32 %rd16, %r16, 4; add.s64 %rd17, %rd1, %rd16; ld.local.f32 %f71, [%rd17]; $L__BB100_12: setp.eq.s32 %p5, %r17, 255; @%p5 bra $L__BB100_14; bra.uni $L__BB100_13; $L__BB100_14: mov.b32 %f72, %r29; bra.uni $L__BB100_15; $L__BB100_13: mul.wide.u32 %rd18, %r17, 4; add.s64 %rd19, %rd1, %rd18; ld.local.f32 %f72, [%rd19]; $L__BB100_15: setp.eq.s32 %p6, %r18, 255; @%p6 bra $L__BB100_17; bra.uni $L__BB100_16; $L__BB100_17: mov.b32 %f73, %r30; bra.uni $L__BB100_18; $L__BB100_16: mul.wide.u32 %rd20, %r18, 4; add.s64 %rd21, %rd1, %rd20; ld.local.f32 %f73, [%rd21]; $L__BB100_18: fma.rn.ftz.f32 %f45, %f1, %f68, 0f358637BD; fma.rn.ftz.f32 %f46, %f2, %f68, 0f358637BD; fma.rn.ftz.f32 %f47, %f3, %f68, 0f358637BD; mul.ftz.f32 %f22, %f45, 0f3F7FFFEF; mul.ftz.f32 %f23, %f46, 0f3F7FFFEF; mul.ftz.f32 %f24, %f47, 0f3F7FFFEF; and.b32 %r33, %r9, 255; setp.eq.s32 %p7, %r33, 0; @%p7 bra $L__BB100_25; setp.eq.s32 %p8, %r11, 0; @%p8 bra $L__BB100_23; setp.eq.s32 %p9, %r11, 1; @%p9 bra $L__BB100_22; bra.uni $L__BB100_21; $L__BB100_22: mul.ftz.f32 %f75, %f23, 0f00000000; mov.f32 %f74, %f22; mov.f32 %f76, %f24; bra.uni $L__BB100_24; $L__BB100_25: setp.eq.s32 %p11, %r10, 0; @%p11 bra $L__BB100_31; setp.eq.s32 %p12, %r10, 1; @%p12 bra $L__BB100_30; bra.uni $L__BB100_27; $L__BB100_30: mul.ftz.f32 %f77, %f23, 0f41A00000; bra.uni $L__BB100_32; $L__BB100_23: mul.ftz.f32 %f74, %f22, 0f00000000; mov.f32 %f75, %f23; mov.f32 %f76, %f24; bra.uni $L__BB100_24; $L__BB100_21: setp.eq.s32 %p10, %r11, 2; mul.ftz.f32 %f48, %f24, 0f00000000; selp.f32 %f76, %f48, %f24, %p10; mov.f32 %f74, %f22; mov.f32 %f75, %f23; $L__BB100_24: mul.ftz.f32 %f49, %f75, %f75; fma.rn.ftz.f32 %f50, %f74, %f74, %f49; fma.rn.ftz.f32 %f51, %f76, %f76, %f50; sqrt.approx.ftz.f32 %f52, %f51; mul.ftz.f32 %f77, %f52, 0f41A00000; $L__BB100_32: add.ftz.f32 %f78, %f73, %f77; setp.eq.ftz.f32 %p14, %f69, 0f00000000; @%p14 bra $L__BB100_34; mul.ftz.f32 %f55, %f22, %f71; mul.ftz.f32 %f56, %f23, %f71; mul.ftz.f32 %f57, %f24, %f71; { // callseq 80, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f55; st.param.f32 [param0+4], %f56; st.param.f32 [param0+8], %f57; .param .b32 param1; st.param.f32 [param1+0], %f70; .param .b32 param2; st.param.f32 [param2+0], %f72; .param .b32 retval0; call.uni (retval0), _Z16fractal_noise_3d6float3ff, ( param0, param1, param2 ); ld.param.f32 %f58, [retval0+0]; } // callseq 80 fma.rn.ftz.f32 %f59, %f58, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f78, %f69, %f59, %f78; $L__BB100_34: setp.eq.s32 %p15, %r12, 0; @%p15 bra $L__BB100_38; setp.eq.s32 %p16, %r12, 1; mov.f32 %f60, 0f40C90FDB; div.approx.ftz.f32 %f40, %f78, %f60; @%p16 bra $L__BB100_37; bra.uni $L__BB100_36; $L__BB100_37: cvt.rmi.ftz.f32.f32 %f65, %f40; sub.ftz.f32 %f79, %f40, %f65; bra.uni $L__BB100_39; $L__BB100_38: add.ftz.f32 %f66, %f78, 0fBFC90FDB; sin.approx.ftz.f32 %f67, %f66; fma.rn.ftz.f32 %f79, %f67, 0f3F000000, 0f3F000000; bra.uni $L__BB100_39; $L__BB100_36: add.ftz.f32 %f61, %f40, 0f3F000000; cvt.rmi.ftz.f32.f32 %f62, %f61; sub.ftz.f32 %f63, %f40, %f62; abs.ftz.f32 %f64, %f63; add.ftz.f32 %f79, %f64, %f64; $L__BB100_39: setp.eq.s32 %p17, %r20, 255; @%p17 bra $L__BB100_41; mul.wide.u32 %rd22, %r20, 4; add.s64 %rd23, %rd1, %rd22; st.local.f32 [%rd23], %f79; $L__BB100_41: setp.eq.s32 %p18, %r19, 255; @%p18 bra $L__BB100_43; mul.wide.u32 %rd24, %r19, 4; add.s64 %rd25, %rd1, %rd24; st.local.f32 [%rd25], %f79; st.local.f32 [%rd25+4], %f79; st.local.f32 [%rd25+8], %f79; $L__BB100_43: add.s32 %r34, %r21, 2; st.param.b32 [func_retval0+0], %r34; ret; $L__BB100_31: mul.ftz.f32 %f77, %f22, 0f41A00000; bra.uni $L__BB100_32; $L__BB100_27: setp.eq.s32 %p13, %r10, 2; @%p13 bra $L__BB100_29; bra.uni $L__BB100_28; $L__BB100_29: mul.ftz.f32 %f77, %f24, 0f41A00000; bra.uni $L__BB100_32; $L__BB100_28: add.ftz.f32 %f53, %f22, %f23; add.ftz.f32 %f54, %f24, %f53; mul.ftz.f32 %f77, %f54, 0f41200000; bra.uni $L__BB100_32; } .func _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj( .param .b64 _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0, .param .b32 _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1, .param .b32 _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2 ) { .reg .pred %p<2>; .reg .f32 %f<64>; .reg .b32 %r<4>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_0]; ld.param.u32 %r3, [_Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_1]; ld.param.u32 %r2, [_Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj_param_2]; cvta.to.local.u64 %rd1, %rd2; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd1, %rd3; ld.local.f32 %f11, [%rd4]; add.ftz.f32 %f12, %f11, 0fC3BE0000; mul.ftz.f32 %f1, %f12, 0f3E4CCCCD; cvt.rzi.ftz.s32.f32 %r1, %f1; setp.gt.u32 %p1, %r1, 79; mov.f32 %f10, 0f00000000; mov.f32 %f61, %f10; mov.f32 %f62, %f10; mov.f32 %f63, %f10; @%p1 bra $L__BB101_2; cvt.rn.f32.s32 %f13, %r1; sub.ftz.f32 %f14, %f1, %f13; mul.wide.s32 %rd5, %r1, 12; mov.u64 %rd6, cie_colour_match; add.s64 %rd7, %rd6, %rd5; ld.const.f32 %f15, [%rd7+12]; ld.const.f32 %f16, [%rd7]; sub.ftz.f32 %f17, %f15, %f16; ld.const.f32 %f18, [%rd7+16]; ld.const.f32 %f19, [%rd7+4]; sub.ftz.f32 %f20, %f18, %f19; ld.const.f32 %f21, [%rd7+20]; ld.const.f32 %f22, [%rd7+8]; sub.ftz.f32 %f23, %f21, %f22; fma.rn.ftz.f32 %f63, %f14, %f17, %f16; fma.rn.ftz.f32 %f62, %f14, %f20, %f19; fma.rn.ftz.f32 %f61, %f14, %f23, %f22; $L__BB101_2: ld.const.v4.f32 {%f24, %f25, %f26, %f27}, [kernel_params+1136]; mul.ftz.f32 %f31, %f62, %f25; fma.rn.ftz.f32 %f32, %f63, %f24, %f31; fma.rn.ftz.f32 %f33, %f61, %f26, %f32; ld.const.v4.f32 {%f34, %f35, %f36, %f37}, [kernel_params+1152]; mul.ftz.f32 %f41, %f62, %f35; fma.rn.ftz.f32 %f42, %f63, %f34, %f41; fma.rn.ftz.f32 %f43, %f61, %f36, %f42; ld.const.v4.f32 {%f44, %f45, %f46, %f47}, [kernel_params+1168]; mul.ftz.f32 %f51, %f62, %f45; fma.rn.ftz.f32 %f52, %f63, %f44, %f51; fma.rn.ftz.f32 %f53, %f61, %f46, %f52; mul.ftz.f32 %f54, %f33, 0f3ECB2CB3; mul.ftz.f32 %f55, %f43, 0f3ECB2CB3; mul.ftz.f32 %f56, %f53, 0f3ECB2CB3; max.ftz.f32 %f58, %f54, %f10; max.ftz.f32 %f59, %f55, %f10; max.ftz.f32 %f60, %f56, %f10; mul.wide.u32 %rd8, %r2, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f58; st.local.f32 [%rd9+4], %f59; st.local.f32 [%rd9+8], %f60; ret; } .func _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj( .param .b64 _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0, .param .b32 _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1, .param .b32 _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2, .param .b32 _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3 ) { .reg .pred %p<13>; .reg .f32 %f<49>; .reg .b32 %r<491>; .reg .b64 %rd<13>; ld.param.u64 %rd2, [_Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_0]; ld.param.u32 %r3, [_Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_1]; ld.param.u32 %r4, [_Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_2]; ld.param.u32 %r5, [_Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj_param_3]; cvta.to.local.u64 %rd1, %rd2; and.b32 %r1, %r5, 255; bfe.u32 %r2, %r5, 8, 8; shl.b32 %r6, %r4, 2; cvt.u64.u32 %rd3, %r6; and.b64 %rd4, %rd3, 1020; add.s64 %rd5, %rd1, %rd4; ld.local.f32 %f1, [%rd5]; ld.local.f32 %f2, [%rd5+4]; ld.local.f32 %f3, [%rd5+8]; shr.u32 %r7, %r4, 6; cvt.u64.u32 %rd6, %r7; and.b64 %rd7, %rd6, 1020; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f4, [%rd8]; setp.eq.s32 %p1, %r2, 255; @%p1 bra $L__BB102_11; mov.f32 %f45, 0f3F800000; mov.f32 %f46, 0f00000000; setp.gt.s32 %p2, %r3, 2; @%p2 bra $L__BB102_5; setp.eq.s32 %p5, %r3, 1; @%p5 bra $L__BB102_9; setp.eq.s32 %p6, %r3, 2; mov.f32 %f47, %f45; @%p6 bra $L__BB102_4; bra.uni $L__BB102_10; $L__BB102_4: mov.b32 %r246, %f1; mov.b32 %r247, %f2; add.s32 %r248, %r247, -559038716; add.s32 %r249, %r246, -559038716; xor.b32 %r250, %r248, -559038716; shf.l.wrap.b32 %r251, %r248, %r248, 14; sub.s32 %r252, %r250, %r251; xor.b32 %r253, %r252, %r249; shf.l.wrap.b32 %r254, %r252, %r252, 11; sub.s32 %r255, %r253, %r254; xor.b32 %r256, %r255, %r248; shf.l.wrap.b32 %r257, %r255, %r255, 25; sub.s32 %r258, %r256, %r257; xor.b32 %r259, %r258, %r252; shf.l.wrap.b32 %r260, %r258, %r258, 16; sub.s32 %r261, %r259, %r260; xor.b32 %r262, %r261, %r255; shf.l.wrap.b32 %r263, %r261, %r261, 4; sub.s32 %r264, %r262, %r263; xor.b32 %r265, %r264, %r258; shf.l.wrap.b32 %r266, %r264, %r264, 14; sub.s32 %r267, %r265, %r266; xor.b32 %r268, %r267, %r261; shf.l.wrap.b32 %r269, %r267, %r267, 24; sub.s32 %r270, %r268, %r269; cvt.rn.f32.u32 %f34, %r270; add.s32 %r271, %r247, -559038712; add.s32 %r272, %r246, -559038712; xor.b32 %r273, %r271, 506314504; shf.l.wrap.b32 %r274, %r271, %r271, 14; sub.s32 %r275, %r273, %r274; xor.b32 %r276, %r275, %r272; shf.l.wrap.b32 %r277, %r275, %r275, 11; sub.s32 %r278, %r276, %r277; xor.b32 %r279, %r278, %r271; shf.l.wrap.b32 %r280, %r278, %r278, 25; sub.s32 %r281, %r279, %r280; xor.b32 %r282, %r281, %r275; shf.l.wrap.b32 %r283, %r281, %r281, 16; sub.s32 %r284, %r282, %r283; xor.b32 %r285, %r284, %r278; shf.l.wrap.b32 %r286, %r284, %r284, 4; sub.s32 %r287, %r285, %r286; xor.b32 %r288, %r287, %r281; shf.l.wrap.b32 %r289, %r287, %r287, 14; sub.s32 %r290, %r288, %r289; xor.b32 %r291, %r290, %r284; shf.l.wrap.b32 %r292, %r290, %r290, 24; sub.s32 %r293, %r291, %r292; cvt.rn.f32.u32 %f35, %r293; xor.b32 %r294, %r271, 514703112; sub.s32 %r295, %r294, %r274; xor.b32 %r296, %r295, %r272; shf.l.wrap.b32 %r297, %r295, %r295, 11; sub.s32 %r298, %r296, %r297; xor.b32 %r299, %r298, %r271; shf.l.wrap.b32 %r300, %r298, %r298, 25; sub.s32 %r301, %r299, %r300; xor.b32 %r302, %r301, %r295; shf.l.wrap.b32 %r303, %r301, %r301, 16; sub.s32 %r304, %r302, %r303; xor.b32 %r305, %r304, %r298; shf.l.wrap.b32 %r306, %r304, %r304, 4; sub.s32 %r307, %r305, %r306; xor.b32 %r308, %r307, %r301; shf.l.wrap.b32 %r309, %r307, %r307, 14; sub.s32 %r310, %r308, %r309; xor.b32 %r311, %r310, %r304; shf.l.wrap.b32 %r312, %r310, %r310, 24; sub.s32 %r313, %r311, %r312; cvt.rn.f32.u32 %f36, %r313; mul.ftz.f32 %f47, %f36, 0f2F800000; mul.ftz.f32 %f46, %f35, 0f2F800000; mul.ftz.f32 %f45, %f34, 0f2F800000; bra.uni $L__BB102_10; $L__BB102_5: setp.eq.s32 %p3, %r3, 3; @%p3 bra $L__BB102_8; setp.ne.s32 %p4, %r3, 4; mov.f32 %f47, %f45; @%p4 bra $L__BB102_10; mov.b32 %r8, %f1; mov.b32 %r9, %f2; add.s32 %r10, %r9, -559038708; mov.b32 %r11, %f3; add.s32 %r12, %r11, -559038708; sub.s32 %r13, %r8, %r11; shf.l.wrap.b32 %r14, %r12, %r12, 4; xor.b32 %r15, %r14, %r13; add.s32 %r16, %r12, %r10; sub.s32 %r17, %r10, %r15; shf.l.wrap.b32 %r18, %r15, %r15, 6; xor.b32 %r19, %r18, %r17; add.s32 %r20, %r15, %r16; sub.s32 %r21, %r16, %r19; shf.l.wrap.b32 %r22, %r19, %r19, 8; xor.b32 %r23, %r22, %r21; add.s32 %r24, %r19, %r20; sub.s32 %r25, %r20, %r23; shf.l.wrap.b32 %r26, %r23, %r23, 16; xor.b32 %r27, %r26, %r25; add.s32 %r28, %r23, %r24; sub.s32 %r29, %r24, %r27; shf.l.wrap.b32 %r30, %r27, %r27, 19; xor.b32 %r31, %r30, %r29; add.s32 %r32, %r27, %r28; sub.s32 %r33, %r28, %r31; shf.l.wrap.b32 %r34, %r31, %r31, 4; add.s32 %r35, %r31, %r32; mov.b32 %r36, %f4; add.s32 %r37, %r32, %r36; xor.b32 %r38, %r35, %r33; xor.b32 %r39, %r38, %r34; shf.l.wrap.b32 %r40, %r35, %r35, 14; sub.s32 %r41, %r39, %r40; xor.b32 %r42, %r41, %r37; shf.l.wrap.b32 %r43, %r41, %r41, 11; sub.s32 %r44, %r42, %r43; xor.b32 %r45, %r44, %r35; shf.l.wrap.b32 %r46, %r44, %r44, 25; sub.s32 %r47, %r45, %r46; xor.b32 %r48, %r47, %r41; shf.l.wrap.b32 %r49, %r47, %r47, 16; sub.s32 %r50, %r48, %r49; xor.b32 %r51, %r50, %r44; shf.l.wrap.b32 %r52, %r50, %r50, 4; sub.s32 %r53, %r51, %r52; xor.b32 %r54, %r53, %r47; shf.l.wrap.b32 %r55, %r53, %r53, 14; sub.s32 %r56, %r54, %r55; xor.b32 %r57, %r56, %r50; shf.l.wrap.b32 %r58, %r56, %r56, 24; sub.s32 %r59, %r57, %r58; cvt.rn.f32.u32 %f28, %r59; add.s32 %r60, %r8, -559038708; add.s32 %r61, %r36, -559038708; sub.s32 %r62, %r11, %r36; shf.l.wrap.b32 %r63, %r61, %r61, 4; xor.b32 %r64, %r63, %r62; add.s32 %r65, %r61, %r60; sub.s32 %r66, %r60, %r64; shf.l.wrap.b32 %r67, %r64, %r64, 6; xor.b32 %r68, %r67, %r66; add.s32 %r69, %r64, %r65; sub.s32 %r70, %r65, %r68; shf.l.wrap.b32 %r71, %r68, %r68, 8; xor.b32 %r72, %r71, %r70; add.s32 %r73, %r68, %r69; sub.s32 %r74, %r69, %r72; shf.l.wrap.b32 %r75, %r72, %r72, 16; xor.b32 %r76, %r75, %r74; add.s32 %r77, %r72, %r73; sub.s32 %r78, %r73, %r76; shf.l.wrap.b32 %r79, %r76, %r76, 19; xor.b32 %r80, %r79, %r78; add.s32 %r81, %r76, %r77; sub.s32 %r82, %r77, %r80; shf.l.wrap.b32 %r83, %r80, %r80, 4; add.s32 %r84, %r80, %r81; add.s32 %r85, %r81, %r9; xor.b32 %r86, %r84, %r82; xor.b32 %r87, %r86, %r83; shf.l.wrap.b32 %r88, %r84, %r84, 14; sub.s32 %r89, %r87, %r88; xor.b32 %r90, %r89, %r85; shf.l.wrap.b32 %r91, %r89, %r89, 11; sub.s32 %r92, %r90, %r91; xor.b32 %r93, %r92, %r84; shf.l.wrap.b32 %r94, %r92, %r92, 25; sub.s32 %r95, %r93, %r94; xor.b32 %r96, %r95, %r89; shf.l.wrap.b32 %r97, %r95, %r95, 16; sub.s32 %r98, %r96, %r97; xor.b32 %r99, %r98, %r92; shf.l.wrap.b32 %r100, %r98, %r98, 4; sub.s32 %r101, %r99, %r100; xor.b32 %r102, %r101, %r95; shf.l.wrap.b32 %r103, %r101, %r101, 14; sub.s32 %r104, %r102, %r103; xor.b32 %r105, %r104, %r98; shf.l.wrap.b32 %r106, %r104, %r104, 24; sub.s32 %r107, %r105, %r106; cvt.rn.f32.u32 %f29, %r107; sub.s32 %r108, %r36, %r9; shf.l.wrap.b32 %r109, %r10, %r10, 4; xor.b32 %r110, %r109, %r108; sub.s32 %r111, %r12, %r110; shf.l.wrap.b32 %r112, %r110, %r110, 6; xor.b32 %r113, %r112, %r111; add.s32 %r114, %r110, %r16; sub.s32 %r115, %r16, %r113; shf.l.wrap.b32 %r116, %r113, %r113, 8; xor.b32 %r117, %r116, %r115; add.s32 %r118, %r113, %r114; sub.s32 %r119, %r114, %r117; shf.l.wrap.b32 %r120, %r117, %r117, 16; xor.b32 %r121, %r120, %r119; add.s32 %r122, %r117, %r118; sub.s32 %r123, %r118, %r121; shf.l.wrap.b32 %r124, %r121, %r121, 19; xor.b32 %r125, %r124, %r123; add.s32 %r126, %r121, %r122; sub.s32 %r127, %r122, %r125; shf.l.wrap.b32 %r128, %r125, %r125, 4; add.s32 %r129, %r125, %r126; add.s32 %r130, %r126, %r8; xor.b32 %r131, %r129, %r127; xor.b32 %r132, %r131, %r128; shf.l.wrap.b32 %r133, %r129, %r129, 14; sub.s32 %r134, %r132, %r133; xor.b32 %r135, %r134, %r130; shf.l.wrap.b32 %r136, %r134, %r134, 11; sub.s32 %r137, %r135, %r136; xor.b32 %r138, %r137, %r129; shf.l.wrap.b32 %r139, %r137, %r137, 25; sub.s32 %r140, %r138, %r139; xor.b32 %r141, %r140, %r134; shf.l.wrap.b32 %r142, %r140, %r140, 16; sub.s32 %r143, %r141, %r142; xor.b32 %r144, %r143, %r137; shf.l.wrap.b32 %r145, %r143, %r143, 4; sub.s32 %r146, %r144, %r145; xor.b32 %r147, %r146, %r140; shf.l.wrap.b32 %r148, %r146, %r146, 14; sub.s32 %r149, %r147, %r148; xor.b32 %r150, %r149, %r143; shf.l.wrap.b32 %r151, %r149, %r149, 24; sub.s32 %r152, %r150, %r151; cvt.rn.f32.u32 %f30, %r152; mul.ftz.f32 %f47, %f30, 0f2F800000; mul.ftz.f32 %f46, %f29, 0f2F800000; mul.ftz.f32 %f45, %f28, 0f2F800000; bra.uni $L__BB102_10; $L__BB102_9: mov.b32 %r314, %f4; add.s32 %r315, %r314, -559038720; xor.b32 %r316, %r315, -1874868139; add.s32 %r317, %r316, 29184895; xor.b32 %r318, %r317, -559038720; shf.l.wrap.b32 %r319, %r317, %r317, 25; sub.s32 %r320, %r318, %r319; xor.b32 %r321, %r320, -1874868139; shf.l.wrap.b32 %r322, %r320, %r320, 16; sub.s32 %r323, %r321, %r322; xor.b32 %r324, %r323, %r317; shf.l.wrap.b32 %r325, %r323, %r323, 4; sub.s32 %r326, %r324, %r325; xor.b32 %r327, %r326, %r320; shf.l.wrap.b32 %r328, %r326, %r326, 14; sub.s32 %r329, %r327, %r328; xor.b32 %r330, %r329, %r323; shf.l.wrap.b32 %r331, %r329, %r329, 24; sub.s32 %r332, %r330, %r331; cvt.rn.f32.u32 %f37, %r332; add.s32 %r333, %r314, -559038716; xor.b32 %r334, %r333, 1354692725; add.s32 %r335, %r334, 138171771; xor.b32 %r336, %r335, 506314500; shf.l.wrap.b32 %r337, %r335, %r335, 25; sub.s32 %r338, %r336, %r337; xor.b32 %r339, %r338, 1354692725; shf.l.wrap.b32 %r340, %r338, %r338, 16; sub.s32 %r341, %r339, %r340; xor.b32 %r342, %r341, %r335; shf.l.wrap.b32 %r343, %r341, %r341, 4; sub.s32 %r344, %r342, %r343; xor.b32 %r345, %r344, %r338; shf.l.wrap.b32 %r346, %r344, %r344, 14; sub.s32 %r347, %r345, %r346; xor.b32 %r348, %r347, %r341; shf.l.wrap.b32 %r349, %r347, %r347, 24; sub.s32 %r350, %r348, %r349; cvt.rn.f32.u32 %f38, %r350; xor.b32 %r351, %r333, 1346304085; add.s32 %r352, %r351, 138237311; xor.b32 %r353, %r352, 514703108; shf.l.wrap.b32 %r354, %r352, %r352, 25; sub.s32 %r355, %r353, %r354; xor.b32 %r356, %r355, 1346304085; shf.l.wrap.b32 %r357, %r355, %r355, 16; sub.s32 %r358, %r356, %r357; xor.b32 %r359, %r358, %r352; shf.l.wrap.b32 %r360, %r358, %r358, 4; sub.s32 %r361, %r359, %r360; xor.b32 %r362, %r361, %r355; shf.l.wrap.b32 %r363, %r361, %r361, 14; sub.s32 %r364, %r362, %r363; xor.b32 %r365, %r364, %r358; shf.l.wrap.b32 %r366, %r364, %r364, 24; sub.s32 %r367, %r365, %r366; cvt.rn.f32.u32 %f39, %r367; mul.ftz.f32 %f47, %f39, 0f2F800000; mul.ftz.f32 %f46, %f38, 0f2F800000; mul.ftz.f32 %f45, %f37, 0f2F800000; bra.uni $L__BB102_10; $L__BB102_8: mov.b32 %r153, %f1; mov.b32 %r154, %f3; add.s32 %r155, %r154, -559038712; mov.b32 %r156, %f2; add.s32 %r157, %r156, -559038712; add.s32 %r158, %r153, -559038712; xor.b32 %r159, %r155, %r157; shf.l.wrap.b32 %r160, %r157, %r157, 14; sub.s32 %r161, %r159, %r160; xor.b32 %r162, %r161, %r158; shf.l.wrap.b32 %r163, %r161, %r161, 11; sub.s32 %r164, %r162, %r163; xor.b32 %r165, %r164, %r157; shf.l.wrap.b32 %r166, %r164, %r164, 25; sub.s32 %r167, %r165, %r166; xor.b32 %r168, %r167, %r161; shf.l.wrap.b32 %r169, %r167, %r167, 16; sub.s32 %r170, %r168, %r169; xor.b32 %r171, %r170, %r164; shf.l.wrap.b32 %r172, %r170, %r170, 4; sub.s32 %r173, %r171, %r172; xor.b32 %r174, %r173, %r167; shf.l.wrap.b32 %r175, %r173, %r173, 14; sub.s32 %r176, %r174, %r175; xor.b32 %r177, %r176, %r170; shf.l.wrap.b32 %r178, %r176, %r176, 24; sub.s32 %r179, %r177, %r178; cvt.rn.f32.u32 %f31, %r179; add.s32 %r180, %r156, -559038708; add.s32 %r181, %r154, -559038708; sub.s32 %r182, %r153, %r154; shf.l.wrap.b32 %r183, %r181, %r181, 4; xor.b32 %r184, %r183, %r182; add.s32 %r185, %r181, %r180; sub.s32 %r186, %r180, %r184; shf.l.wrap.b32 %r187, %r184, %r184, 6; xor.b32 %r188, %r187, %r186; add.s32 %r189, %r184, %r185; sub.s32 %r190, %r185, %r188; shf.l.wrap.b32 %r191, %r188, %r188, 8; xor.b32 %r192, %r191, %r190; add.s32 %r193, %r188, %r189; sub.s32 %r194, %r189, %r192; shf.l.wrap.b32 %r195, %r192, %r192, 16; xor.b32 %r196, %r195, %r194; add.s32 %r197, %r192, %r193; sub.s32 %r198, %r193, %r196; shf.l.wrap.b32 %r199, %r196, %r196, 19; xor.b32 %r200, %r199, %r198; add.s32 %r201, %r196, %r197; sub.s32 %r202, %r197, %r200; shf.l.wrap.b32 %r203, %r200, %r200, 4; add.s32 %r204, %r200, %r201; add.s32 %r205, %r201, 1065353216; xor.b32 %r206, %r204, %r202; xor.b32 %r207, %r206, %r203; shf.l.wrap.b32 %r208, %r204, %r204, 14; sub.s32 %r209, %r207, %r208; xor.b32 %r210, %r209, %r205; shf.l.wrap.b32 %r211, %r209, %r209, 11; sub.s32 %r212, %r210, %r211; xor.b32 %r213, %r212, %r204; shf.l.wrap.b32 %r214, %r212, %r212, 25; sub.s32 %r215, %r213, %r214; xor.b32 %r216, %r215, %r209; shf.l.wrap.b32 %r217, %r215, %r215, 16; sub.s32 %r218, %r216, %r217; xor.b32 %r219, %r218, %r212; shf.l.wrap.b32 %r220, %r218, %r218, 4; sub.s32 %r221, %r219, %r220; xor.b32 %r222, %r221, %r215; shf.l.wrap.b32 %r223, %r221, %r221, 14; sub.s32 %r224, %r222, %r223; xor.b32 %r225, %r224, %r218; shf.l.wrap.b32 %r226, %r224, %r224, 24; sub.s32 %r227, %r225, %r226; cvt.rn.f32.u32 %f32, %r227; add.s32 %r228, %r201, 1073741824; xor.b32 %r229, %r209, %r228; sub.s32 %r230, %r229, %r211; xor.b32 %r231, %r230, %r204; shf.l.wrap.b32 %r232, %r230, %r230, 25; sub.s32 %r233, %r231, %r232; xor.b32 %r234, %r233, %r209; shf.l.wrap.b32 %r235, %r233, %r233, 16; sub.s32 %r236, %r234, %r235; xor.b32 %r237, %r236, %r230; shf.l.wrap.b32 %r238, %r236, %r236, 4; sub.s32 %r239, %r237, %r238; xor.b32 %r240, %r239, %r233; shf.l.wrap.b32 %r241, %r239, %r239, 14; sub.s32 %r242, %r240, %r241; xor.b32 %r243, %r242, %r236; shf.l.wrap.b32 %r244, %r242, %r242, 24; sub.s32 %r245, %r243, %r244; cvt.rn.f32.u32 %f33, %r245; mul.ftz.f32 %f47, %f33, 0f2F800000; mul.ftz.f32 %f46, %f32, 0f2F800000; mul.ftz.f32 %f45, %f31, 0f2F800000; $L__BB102_10: mul.wide.u32 %rd9, %r2, 4; add.s64 %rd10, %rd1, %rd9; st.local.f32 [%rd10], %f45; st.local.f32 [%rd10+4], %f46; st.local.f32 [%rd10+8], %f47; $L__BB102_11: setp.eq.s32 %p7, %r1, 255; @%p7 bra $L__BB102_22; mov.f32 %f48, 0f00000000; setp.gt.s32 %p8, %r3, 2; @%p8 bra $L__BB102_16; setp.eq.s32 %p11, %r3, 1; @%p11 bra $L__BB102_20; setp.eq.s32 %p12, %r3, 2; @%p12 bra $L__BB102_15; bra.uni $L__BB102_21; $L__BB102_15: mov.b32 %r447, %f1; mov.b32 %r448, %f2; add.s32 %r449, %r448, -559038716; add.s32 %r450, %r447, -559038716; xor.b32 %r451, %r449, -559038716; shf.l.wrap.b32 %r452, %r449, %r449, 14; sub.s32 %r453, %r451, %r452; xor.b32 %r454, %r453, %r450; shf.l.wrap.b32 %r455, %r453, %r453, 11; sub.s32 %r456, %r454, %r455; xor.b32 %r457, %r456, %r449; shf.l.wrap.b32 %r458, %r456, %r456, 25; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r453; shf.l.wrap.b32 %r461, %r459, %r459, 16; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r456; shf.l.wrap.b32 %r464, %r462, %r462, 4; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 14; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 24; sub.s32 %r471, %r469, %r470; cvt.rn.f32.u32 %f43, %r471; mul.ftz.f32 %f48, %f43, 0f2F800000; bra.uni $L__BB102_21; $L__BB102_16: setp.eq.s32 %p9, %r3, 3; @%p9 bra $L__BB102_19; setp.ne.s32 %p10, %r3, 4; @%p10 bra $L__BB102_21; mov.b32 %r368, %f1; mov.b32 %r369, %f2; add.s32 %r370, %r369, -559038708; mov.b32 %r371, %f3; add.s32 %r372, %r371, -559038708; sub.s32 %r373, %r368, %r371; shf.l.wrap.b32 %r374, %r372, %r372, 4; xor.b32 %r375, %r374, %r373; add.s32 %r376, %r372, %r370; sub.s32 %r377, %r370, %r375; shf.l.wrap.b32 %r378, %r375, %r375, 6; xor.b32 %r379, %r378, %r377; add.s32 %r380, %r375, %r376; sub.s32 %r381, %r376, %r379; shf.l.wrap.b32 %r382, %r379, %r379, 8; xor.b32 %r383, %r382, %r381; add.s32 %r384, %r379, %r380; sub.s32 %r385, %r380, %r383; shf.l.wrap.b32 %r386, %r383, %r383, 16; xor.b32 %r387, %r386, %r385; add.s32 %r388, %r383, %r384; sub.s32 %r389, %r384, %r387; shf.l.wrap.b32 %r390, %r387, %r387, 19; xor.b32 %r391, %r390, %r389; add.s32 %r392, %r387, %r388; sub.s32 %r393, %r388, %r391; shf.l.wrap.b32 %r394, %r391, %r391, 4; add.s32 %r395, %r391, %r392; mov.b32 %r396, %f4; add.s32 %r397, %r392, %r396; xor.b32 %r398, %r395, %r393; xor.b32 %r399, %r398, %r394; shf.l.wrap.b32 %r400, %r395, %r395, 14; sub.s32 %r401, %r399, %r400; xor.b32 %r402, %r401, %r397; shf.l.wrap.b32 %r403, %r401, %r401, 11; sub.s32 %r404, %r402, %r403; xor.b32 %r405, %r404, %r395; shf.l.wrap.b32 %r406, %r404, %r404, 25; sub.s32 %r407, %r405, %r406; xor.b32 %r408, %r407, %r401; shf.l.wrap.b32 %r409, %r407, %r407, 16; sub.s32 %r410, %r408, %r409; xor.b32 %r411, %r410, %r404; shf.l.wrap.b32 %r412, %r410, %r410, 4; sub.s32 %r413, %r411, %r412; xor.b32 %r414, %r413, %r407; shf.l.wrap.b32 %r415, %r413, %r413, 14; sub.s32 %r416, %r414, %r415; xor.b32 %r417, %r416, %r410; shf.l.wrap.b32 %r418, %r416, %r416, 24; sub.s32 %r419, %r417, %r418; cvt.rn.f32.u32 %f41, %r419; mul.ftz.f32 %f48, %f41, 0f2F800000; bra.uni $L__BB102_21; $L__BB102_20: mov.b32 %r472, %f4; add.s32 %r473, %r472, -559038720; xor.b32 %r474, %r473, -1874868139; add.s32 %r475, %r474, 29184895; xor.b32 %r476, %r475, -559038720; shf.l.wrap.b32 %r477, %r475, %r475, 25; sub.s32 %r478, %r476, %r477; xor.b32 %r479, %r478, -1874868139; shf.l.wrap.b32 %r480, %r478, %r478, 16; sub.s32 %r481, %r479, %r480; xor.b32 %r482, %r481, %r475; shf.l.wrap.b32 %r483, %r481, %r481, 4; sub.s32 %r484, %r482, %r483; xor.b32 %r485, %r484, %r478; shf.l.wrap.b32 %r486, %r484, %r484, 14; sub.s32 %r487, %r485, %r486; xor.b32 %r488, %r487, %r481; shf.l.wrap.b32 %r489, %r487, %r487, 24; sub.s32 %r490, %r488, %r489; cvt.rn.f32.u32 %f44, %r490; mul.ftz.f32 %f48, %f44, 0f2F800000; bra.uni $L__BB102_21; $L__BB102_19: mov.b32 %r420, %f1; mov.b32 %r421, %f3; add.s32 %r422, %r421, -559038712; mov.b32 %r423, %f2; add.s32 %r424, %r423, -559038712; add.s32 %r425, %r420, -559038712; xor.b32 %r426, %r422, %r424; shf.l.wrap.b32 %r427, %r424, %r424, 14; sub.s32 %r428, %r426, %r427; xor.b32 %r429, %r428, %r425; shf.l.wrap.b32 %r430, %r428, %r428, 11; sub.s32 %r431, %r429, %r430; xor.b32 %r432, %r431, %r424; shf.l.wrap.b32 %r433, %r431, %r431, 25; sub.s32 %r434, %r432, %r433; xor.b32 %r435, %r434, %r428; shf.l.wrap.b32 %r436, %r434, %r434, 16; sub.s32 %r437, %r435, %r436; xor.b32 %r438, %r437, %r431; shf.l.wrap.b32 %r439, %r437, %r437, 4; sub.s32 %r440, %r438, %r439; xor.b32 %r441, %r440, %r434; shf.l.wrap.b32 %r442, %r440, %r440, 14; sub.s32 %r443, %r441, %r442; xor.b32 %r444, %r443, %r437; shf.l.wrap.b32 %r445, %r443, %r443, 24; sub.s32 %r446, %r444, %r445; cvt.rn.f32.u32 %f42, %r446; mul.ftz.f32 %f48, %f42, 0f2F800000; $L__BB102_21: mul.wide.u32 %rd11, %r1, 4; add.s64 %rd12, %rd1, %rd11; st.local.f32 [%rd12], %f48; $L__BB102_22: ret; } .func _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<76>; .reg .b16 %rs<19>; .reg .f32 %f<1350>; .reg .b32 %r<192>; .reg .b64 %rd<413>; ld.param.u64 %rd178, [_Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd179, [_Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r70, [_Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r69, [_Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r68, [_Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd180, %rd178; and.b32 %r1, %r70, 255; bfe.u32 %r2, %r70, 8, 8; cvta.to.local.u64 %rd1, %rd179; mul.wide.u32 %rd181, %r68, 4; add.s64 %rd182, %rd1, %rd181; ld.local.f32 %f1, [%rd182]; add.s64 %rd2, %rd180, 24; ld.local.v2.f32 {%f243, %f244}, [%rd180+24]; setp.neu.ftz.f32 %p1, %f243, %f244; ld.local.f32 %f247, [%rd180+32]; setp.neu.ftz.f32 %p2, %f243, %f247; or.pred %p3, %p1, %p2; sub.ftz.f32 %f248, %f247, %f244; neg.ftz.f32 %f249, %f244; neg.ftz.f32 %f250, %f247; selp.f32 %f251, %f250, %f247, %p3; add.ftz.f32 %f252, %f243, %f251; selp.f32 %f253, %f244, %f249, %p3; sub.ftz.f32 %f254, %f253, %f243; mul.ftz.f32 %f255, %f252, %f252; fma.rn.ftz.f32 %f256, %f248, %f248, %f255; fma.rn.ftz.f32 %f257, %f254, %f254, %f256; sqrt.approx.ftz.f32 %f258, %f257; div.approx.ftz.f32 %f259, %f248, %f258; div.approx.ftz.f32 %f260, %f252, %f258; div.approx.ftz.f32 %f261, %f254, %f258; mul.ftz.f32 %f262, %f244, %f261; mul.ftz.f32 %f263, %f247, %f260; sub.ftz.f32 %f264, %f262, %f263; mul.ftz.f32 %f265, %f247, %f259; mul.ftz.f32 %f266, %f243, %f261; sub.ftz.f32 %f267, %f265, %f266; mul.ftz.f32 %f268, %f243, %f260; mul.ftz.f32 %f269, %f244, %f259; sub.ftz.f32 %f270, %f268, %f269; ld.local.f32 %f271, [%rd180+92]; mul.ftz.f32 %f2, %f271, %f259; mul.ftz.f32 %f3, %f271, %f260; mul.ftz.f32 %f4, %f271, %f261; mul.ftz.f32 %f5, %f271, %f264; mul.ftz.f32 %f6, %f271, %f267; mul.ftz.f32 %f7, %f271, %f270; ld.local.s32 %rd3, [%rd180+60]; setp.eq.s64 %p4, %rd3, -1; mov.f32 %f1296, 0f00000000; @%p4 bra $L__BB103_33; ld.local.u32 %r3, [%rd2+40]; and.b32 %r71, %r3, 1; setp.eq.b32 %p5, %r71, 1; mov.pred %p6, 0; xor.pred %p7, %p5, %p6; not.pred %p8, %p7; @%p8 bra $L__BB103_33; and.b32 %r72, %r3, 64; setp.eq.s32 %p9, %r72, 0; @%p9 bra $L__BB103_16; bra.uni $L__BB103_3; $L__BB103_16: ld.const.u64 %rd216, [kernel_params+2232]; cvta.to.global.u64 %rd217, %rd216; shl.b64 %rd218, %rd3, 4; add.s64 %rd219, %rd217, %rd218; ld.const.u64 %rd220, [kernel_params+2256]; cvta.to.global.u64 %rd221, %rd220; ld.global.u32 %r98, [%rd219+12]; mul.wide.u32 %rd222, %r98, 12; add.s64 %rd223, %rd221, %rd222; ld.global.f32 %f1294, [%rd223]; ld.global.f32 %f1293, [%rd223+4]; ld.global.f32 %f1292, [%rd223+8]; add.s32 %r99, %r98, 1; mul.wide.u32 %rd224, %r99, 12; add.s64 %rd225, %rd221, %rd224; ld.global.f32 %f1291, [%rd225]; ld.global.f32 %f1290, [%rd225+4]; ld.global.f32 %f1289, [%rd225+8]; add.s32 %r100, %r98, 2; mul.wide.u32 %rd226, %r100, 12; add.s64 %rd227, %rd221, %rd226; ld.global.f32 %f1288, [%rd227]; ld.global.f32 %f1287, [%rd227+4]; ld.global.f32 %f1286, [%rd227+8]; bra.uni $L__BB103_17; $L__BB103_3: ld.const.u64 %rd183, [kernel_params+2160]; cvta.to.global.u64 %rd184, %rd183; ld.local.u32 %r73, [%rd2+52]; mul.wide.s32 %rd185, %r73, 208; add.s64 %rd186, %rd184, %rd185; ld.global.u32 %r4, [%rd186+156]; ld.global.u32 %r5, [%rd186+152]; shl.b32 %r74, %r5, 1; cvt.rn.f32.s32 %f273, %r74; ld.local.f32 %f274, [%rd2+60]; mul.ftz.f32 %f8, %f274, %f273; cvt.rzi.ftz.s32.f32 %r75, %f8; add.s32 %r76, %r74, -1; min.s32 %r6, %r75, %r76; ld.const.u64 %rd187, [kernel_params+2312]; cvta.to.global.u64 %rd4, %rd187; ld.global.u32 %r178, [%rd186+164]; mul.wide.u32 %rd188, %r178, 16; add.s64 %rd189, %rd4, %rd188; ld.global.u16 %rs13, [%rd189+12]; ld.global.u32 %r177, [%rd189+8]; ld.global.u64 %rd356, [%rd189]; setp.eq.s64 %p10, %rd356, 11; @%p10 bra $L__BB103_8; $L__BB103_4: setp.eq.s64 %p11, %rd356, 0; @%p11 bra $L__BB103_6; add.s32 %r178, %r178, 2; bra.uni $L__BB103_7; $L__BB103_6: setp.eq.s16 %p12, %rs13, 0; mov.u32 %r181, -1; mov.u32 %r178, %r177; @%p12 bra $L__BB103_9; $L__BB103_7: mul.wide.u32 %rd190, %r178, 16; add.s64 %rd191, %rd4, %rd190; ld.global.u16 %rs13, [%rd191+12]; ld.global.u32 %r177, [%rd191+8]; ld.global.u64 %rd356, [%rd191]; setp.ne.s64 %p13, %rd356, 11; @%p13 bra $L__BB103_4; $L__BB103_8: setp.eq.s16 %p14, %rs13, 0; selp.b32 %r181, -1, %r177, %p14; $L__BB103_9: ld.const.u64 %rd192, [kernel_params+2232]; cvta.to.global.u64 %rd193, %rd192; shl.b64 %rd194, %rd3, 4; add.s64 %rd195, %rd193, %rd194; ld.global.v4.u32 {%r78, %r79, %r80, %r81}, [%rd195]; setp.eq.s32 %p15, %r6, %r5; @%p15 bra $L__BB103_11; bra.uni $L__BB103_10; $L__BB103_11: ld.const.u64 %rd201, [kernel_params+2256]; cvta.to.global.u64 %rd202, %rd201; mul.wide.u32 %rd203, %r81, 12; add.s64 %rd362, %rd202, %rd203; add.s32 %r88, %r81, 1; mul.wide.u32 %rd204, %r88, 12; add.s64 %rd359, %rd202, %rd204; add.s32 %r89, %r81, 2; mul.wide.u32 %rd205, %r89, 12; add.s64 %rd363, %rd202, %rd205; add.s64 %rd361, %rd362, 4; add.s64 %rd360, %rd362, 8; add.s64 %rd358, %rd359, 4; add.s64 %rd357, %rd359, 8; add.s64 %rd364, %rd363, 4; add.s64 %rd365, %rd363, 8; bra.uni $L__BB103_12; $L__BB103_10: setp.gt.s32 %p16, %r6, %r5; selp.b32 %r82, -1, 0, %p16; add.s32 %r83, %r6, %r82; mad.lo.s32 %r84, %r83, %r4, %r181; ld.const.u64 %rd196, [kernel_params+2336]; cvta.to.global.u64 %rd197, %rd196; add.s32 %r85, %r78, %r84; mul.wide.u32 %rd198, %r85, 12; add.s64 %rd362, %rd197, %rd198; add.s32 %r86, %r79, %r84; mul.wide.u32 %rd199, %r86, 12; add.s64 %rd359, %rd197, %rd199; add.s32 %r87, %r80, %r84; mul.wide.u32 %rd200, %r87, 12; add.s64 %rd363, %rd197, %rd200; add.s64 %rd361, %rd362, 4; add.s64 %rd360, %rd362, 8; add.s64 %rd358, %rd359, 4; add.s64 %rd357, %rd359, 8; add.s64 %rd364, %rd363, 4; add.s64 %rd365, %rd363, 8; $L__BB103_12: ld.global.f32 %f9, [%rd362]; ld.global.f32 %f10, [%rd361]; ld.global.f32 %f11, [%rd360]; ld.global.f32 %f12, [%rd359]; ld.global.f32 %f13, [%rd358]; ld.global.f32 %f14, [%rd357]; ld.global.f32 %f15, [%rd365]; ld.global.f32 %f16, [%rd364]; ld.global.f32 %f17, [%rd363]; add.s32 %r21, %r6, 1; setp.eq.s32 %p17, %r21, %r5; @%p17 bra $L__BB103_14; bra.uni $L__BB103_13; $L__BB103_14: ld.const.u64 %rd211, [kernel_params+2256]; cvta.to.global.u64 %rd212, %rd211; mul.wide.u32 %rd213, %r81, 12; add.s64 %rd366, %rd212, %rd213; add.s32 %r96, %r81, 1; mul.wide.u32 %rd214, %r96, 12; add.s64 %rd369, %rd212, %rd214; add.s32 %r97, %r81, 2; mul.wide.u32 %rd215, %r97, 12; add.s64 %rd372, %rd212, %rd215; add.s64 %rd367, %rd366, 4; add.s64 %rd368, %rd366, 8; add.s64 %rd370, %rd369, 4; add.s64 %rd371, %rd369, 8; add.s64 %rd373, %rd372, 4; add.s64 %rd374, %rd372, 8; bra.uni $L__BB103_15; $L__BB103_13: setp.ge.s32 %p18, %r6, %r5; selp.b32 %r90, -1, 0, %p18; add.s32 %r91, %r21, %r90; mad.lo.s32 %r92, %r91, %r4, %r181; ld.const.u64 %rd206, [kernel_params+2336]; cvta.to.global.u64 %rd207, %rd206; add.s32 %r93, %r78, %r92; mul.wide.u32 %rd208, %r93, 12; add.s64 %rd366, %rd207, %rd208; add.s32 %r94, %r79, %r92; mul.wide.u32 %rd209, %r94, 12; add.s64 %rd369, %rd207, %rd209; add.s32 %r95, %r80, %r92; mul.wide.u32 %rd210, %r95, 12; add.s64 %rd372, %rd207, %rd210; add.s64 %rd367, %rd366, 4; add.s64 %rd368, %rd366, 8; add.s64 %rd370, %rd369, 4; add.s64 %rd371, %rd369, 8; add.s64 %rd373, %rd372, 4; add.s64 %rd374, %rd372, 8; $L__BB103_15: cvt.rn.f32.s32 %f275, %r6; sub.ftz.f32 %f276, %f8, %f275; mov.f32 %f277, 0f3F800000; sub.ftz.f32 %f278, %f277, %f276; mul.ftz.f32 %f279, %f278, %f9; mul.ftz.f32 %f280, %f278, %f10; mul.ftz.f32 %f281, %f278, %f11; ld.global.f32 %f282, [%rd366]; ld.global.f32 %f283, [%rd367]; ld.global.f32 %f284, [%rd368]; fma.rn.ftz.f32 %f1294, %f276, %f282, %f279; fma.rn.ftz.f32 %f1293, %f276, %f283, %f280; fma.rn.ftz.f32 %f1292, %f276, %f284, %f281; mul.ftz.f32 %f285, %f278, %f12; mul.ftz.f32 %f286, %f278, %f13; mul.ftz.f32 %f287, %f278, %f14; ld.global.f32 %f288, [%rd369]; ld.global.f32 %f289, [%rd370]; ld.global.f32 %f290, [%rd371]; fma.rn.ftz.f32 %f1291, %f276, %f288, %f285; fma.rn.ftz.f32 %f1290, %f276, %f289, %f286; fma.rn.ftz.f32 %f1289, %f276, %f290, %f287; mul.ftz.f32 %f291, %f278, %f17; mul.ftz.f32 %f292, %f278, %f16; mul.ftz.f32 %f293, %f278, %f15; ld.global.f32 %f294, [%rd372]; ld.global.f32 %f295, [%rd373]; ld.global.f32 %f296, [%rd374]; fma.rn.ftz.f32 %f1288, %f276, %f294, %f291; fma.rn.ftz.f32 %f1287, %f276, %f295, %f292; fma.rn.ftz.f32 %f1286, %f276, %f296, %f293; $L__BB103_17: ld.local.u32 %r22, [%rd2+32]; and.b32 %r101, %r22, 4; setp.ne.s32 %p19, %r101, 0; @%p19 bra $L__BB103_27; and.b32 %r23, %r22, 2; setp.eq.s32 %p20, %r23, 0; @%p20 bra $L__BB103_20; bra.uni $L__BB103_19; $L__BB103_20: ld.const.u64 %rd228, [kernel_params+2160]; cvta.to.global.u64 %rd229, %rd228; ld.local.u32 %r102, [%rd2+52]; mul.wide.s32 %rd230, %r102, 208; add.s64 %rd231, %rd229, %rd230; ld.global.v4.f32 {%f330, %f331, %f332, %f333}, [%rd231]; ld.global.v4.f32 {%f338, %f339, %f340, %f341}, [%rd231+16]; ld.global.v4.f32 {%f346, %f347, %f348, %f349}, [%rd231+32]; mul.ftz.f32 %f354, %f331, %f1293; fma.rn.ftz.f32 %f355, %f1294, %f330, %f354; fma.rn.ftz.f32 %f356, %f332, %f1292, %f355; add.ftz.f32 %f1280, %f333, %f356; mul.ftz.f32 %f357, %f339, %f1293; fma.rn.ftz.f32 %f358, %f1294, %f338, %f357; fma.rn.ftz.f32 %f359, %f340, %f1292, %f358; add.ftz.f32 %f1281, %f341, %f359; mul.ftz.f32 %f360, %f347, %f1293; fma.rn.ftz.f32 %f361, %f1294, %f346, %f360; fma.rn.ftz.f32 %f362, %f348, %f1292, %f361; add.ftz.f32 %f1292, %f349, %f362; bra.uni $L__BB103_21; $L__BB103_19: ld.local.v4.f32 {%f297, %f298, %f299, %f300}, [%rd2+120]; mul.ftz.f32 %f305, %f1293, %f298; fma.rn.ftz.f32 %f306, %f1294, %f297, %f305; fma.rn.ftz.f32 %f307, %f1292, %f299, %f306; add.ftz.f32 %f1280, %f300, %f307; ld.local.v4.f32 {%f308, %f309, %f310, %f311}, [%rd2+136]; mul.ftz.f32 %f316, %f1293, %f309; fma.rn.ftz.f32 %f317, %f1294, %f308, %f316; fma.rn.ftz.f32 %f318, %f1292, %f310, %f317; add.ftz.f32 %f1281, %f311, %f318; ld.local.v4.f32 {%f319, %f320, %f321, %f322}, [%rd2+152]; mul.ftz.f32 %f327, %f1293, %f320; fma.rn.ftz.f32 %f328, %f1294, %f319, %f327; fma.rn.ftz.f32 %f329, %f1292, %f321, %f328; add.ftz.f32 %f1292, %f322, %f329; $L__BB103_21: @%p20 bra $L__BB103_23; bra.uni $L__BB103_22; $L__BB103_23: ld.const.u64 %rd232, [kernel_params+2160]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r103, [%rd2+52]; mul.wide.s32 %rd234, %r103, 208; add.s64 %rd235, %rd233, %rd234; ld.global.v4.f32 {%f396, %f397, %f398, %f399}, [%rd235]; ld.global.v4.f32 {%f404, %f405, %f406, %f407}, [%rd235+16]; ld.global.v4.f32 {%f412, %f413, %f414, %f415}, [%rd235+32]; mul.ftz.f32 %f420, %f397, %f1290; fma.rn.ftz.f32 %f421, %f1291, %f396, %f420; fma.rn.ftz.f32 %f422, %f398, %f1289, %f421; add.ftz.f32 %f1283, %f399, %f422; mul.ftz.f32 %f423, %f405, %f1290; fma.rn.ftz.f32 %f424, %f1291, %f404, %f423; fma.rn.ftz.f32 %f425, %f406, %f1289, %f424; add.ftz.f32 %f1284, %f407, %f425; mul.ftz.f32 %f426, %f413, %f1290; fma.rn.ftz.f32 %f427, %f1291, %f412, %f426; fma.rn.ftz.f32 %f428, %f414, %f1289, %f427; add.ftz.f32 %f1289, %f415, %f428; bra.uni $L__BB103_24; $L__BB103_22: ld.local.v4.f32 {%f363, %f364, %f365, %f366}, [%rd2+120]; mul.ftz.f32 %f371, %f1290, %f364; fma.rn.ftz.f32 %f372, %f1291, %f363, %f371; fma.rn.ftz.f32 %f373, %f1289, %f365, %f372; add.ftz.f32 %f1283, %f366, %f373; ld.local.v4.f32 {%f374, %f375, %f376, %f377}, [%rd2+136]; mul.ftz.f32 %f382, %f1290, %f375; fma.rn.ftz.f32 %f383, %f1291, %f374, %f382; fma.rn.ftz.f32 %f384, %f1289, %f376, %f383; add.ftz.f32 %f1284, %f377, %f384; ld.local.v4.f32 {%f385, %f386, %f387, %f388}, [%rd2+152]; mul.ftz.f32 %f393, %f1290, %f386; fma.rn.ftz.f32 %f394, %f1291, %f385, %f393; fma.rn.ftz.f32 %f395, %f1289, %f387, %f394; add.ftz.f32 %f1289, %f388, %f395; $L__BB103_24: @%p20 bra $L__BB103_26; bra.uni $L__BB103_25; $L__BB103_26: ld.const.u64 %rd236, [kernel_params+2160]; cvta.to.global.u64 %rd237, %rd236; ld.local.u32 %r104, [%rd2+52]; mul.wide.s32 %rd238, %r104, 208; add.s64 %rd239, %rd237, %rd238; ld.global.v4.f32 {%f462, %f463, %f464, %f465}, [%rd239]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd239+16]; ld.global.v4.f32 {%f478, %f479, %f480, %f481}, [%rd239+32]; mul.ftz.f32 %f486, %f463, %f1287; fma.rn.ftz.f32 %f487, %f1288, %f462, %f486; fma.rn.ftz.f32 %f488, %f464, %f1286, %f487; add.ftz.f32 %f66, %f465, %f488; mul.ftz.f32 %f489, %f471, %f1287; fma.rn.ftz.f32 %f490, %f1288, %f470, %f489; fma.rn.ftz.f32 %f491, %f472, %f1286, %f490; add.ftz.f32 %f67, %f473, %f491; mul.ftz.f32 %f492, %f479, %f1287; fma.rn.ftz.f32 %f493, %f1288, %f478, %f492; fma.rn.ftz.f32 %f494, %f480, %f1286, %f493; add.ftz.f32 %f1286, %f481, %f494; mov.f32 %f1287, %f67; mov.f32 %f1288, %f66; mov.f32 %f1290, %f1284; mov.f32 %f1291, %f1283; mov.f32 %f1293, %f1281; mov.f32 %f1294, %f1280; bra.uni $L__BB103_27; $L__BB103_25: ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+120]; mul.ftz.f32 %f437, %f1287, %f430; fma.rn.ftz.f32 %f438, %f1288, %f429, %f437; fma.rn.ftz.f32 %f439, %f1286, %f431, %f438; add.ftz.f32 %f63, %f432, %f439; ld.local.v4.f32 {%f440, %f441, %f442, %f443}, [%rd2+136]; mul.ftz.f32 %f448, %f1287, %f441; fma.rn.ftz.f32 %f449, %f1288, %f440, %f448; fma.rn.ftz.f32 %f450, %f1286, %f442, %f449; add.ftz.f32 %f64, %f443, %f450; ld.local.v4.f32 {%f451, %f452, %f453, %f454}, [%rd2+152]; mul.ftz.f32 %f459, %f1287, %f452; fma.rn.ftz.f32 %f460, %f1288, %f451, %f459; fma.rn.ftz.f32 %f461, %f1286, %f453, %f460; add.ftz.f32 %f1286, %f454, %f461; mov.f32 %f1287, %f64; mov.f32 %f1288, %f63; mov.f32 %f1290, %f1284; mov.f32 %f1291, %f1283; mov.f32 %f1293, %f1281; mov.f32 %f1294, %f1280; $L__BB103_27: setp.eq.s32 %p23, %r1, 0; mov.f32 %f1296, 0f3F800000; mov.f32 %f1295, %f1296; @%p23 bra $L__BB103_29; ld.local.f32 %f496, [%rd2+12]; ld.local.f32 %f497, [%rd2+16]; mul.ftz.f32 %f498, %f3, %f497; fma.rn.ftz.f32 %f499, %f2, %f496, %f498; ld.local.f32 %f500, [%rd2+20]; fma.rn.ftz.f32 %f501, %f4, %f500, %f499; mul.ftz.f32 %f502, %f496, %f501; mul.ftz.f32 %f503, %f497, %f501; mul.ftz.f32 %f504, %f501, %f500; sub.ftz.f32 %f505, %f2, %f502; sub.ftz.f32 %f506, %f3, %f503; sub.ftz.f32 %f507, %f4, %f504; mul.ftz.f32 %f508, %f506, %f506; fma.rn.ftz.f32 %f509, %f505, %f505, %f508; fma.rn.ftz.f32 %f510, %f507, %f507, %f509; sqrt.approx.ftz.f32 %f511, %f510; mul.ftz.f32 %f512, %f6, %f497; fma.rn.ftz.f32 %f513, %f5, %f496, %f512; fma.rn.ftz.f32 %f514, %f7, %f500, %f513; mul.ftz.f32 %f515, %f496, %f514; mul.ftz.f32 %f516, %f497, %f514; mul.ftz.f32 %f517, %f514, %f500; sub.ftz.f32 %f518, %f5, %f515; sub.ftz.f32 %f519, %f6, %f516; sub.ftz.f32 %f520, %f7, %f517; mul.ftz.f32 %f521, %f519, %f519; fma.rn.ftz.f32 %f522, %f518, %f518, %f521; fma.rn.ftz.f32 %f523, %f520, %f520, %f522; sqrt.approx.ftz.f32 %f524, %f523; add.ftz.f32 %f525, %f511, %f524; mul.ftz.f32 %f1295, %f525, 0f3F000000; $L__BB103_29: mul.ftz.f32 %f527, %f1, 0f3F000000; mul.ftz.f32 %f528, %f527, %f1295; mul.ftz.f32 %f80, %f528, %f528; ld.local.f32 %f81, [%rd2+-24]; sub.ftz.f32 %f529, %f81, %f1294; ld.local.f32 %f82, [%rd2+-20]; sub.ftz.f32 %f530, %f82, %f1293; ld.local.f32 %f83, [%rd2+-16]; sub.ftz.f32 %f531, %f83, %f1292; sub.ftz.f32 %f532, %f1293, %f1287; mul.ftz.f32 %f533, %f531, %f532; sub.ftz.f32 %f534, %f1292, %f1286; mul.ftz.f32 %f535, %f530, %f534; sub.ftz.f32 %f536, %f533, %f535; mul.ftz.f32 %f537, %f529, %f534; sub.ftz.f32 %f538, %f1294, %f1288; mul.ftz.f32 %f539, %f531, %f538; sub.ftz.f32 %f540, %f537, %f539; mul.ftz.f32 %f541, %f530, %f538; mul.ftz.f32 %f542, %f529, %f532; sub.ftz.f32 %f543, %f541, %f542; mul.ftz.f32 %f544, %f540, %f540; fma.rn.ftz.f32 %f545, %f536, %f536, %f544; fma.rn.ftz.f32 %f546, %f543, %f543, %f545; mul.ftz.f32 %f547, %f532, %f532; fma.rn.ftz.f32 %f548, %f538, %f538, %f547; fma.rn.ftz.f32 %f549, %f534, %f534, %f548; mul.ftz.f32 %f550, %f80, %f549; setp.lt.ftz.f32 %p24, %f546, %f550; @%p24 bra $L__BB103_33; sub.ftz.f32 %f552, %f81, %f1291; sub.ftz.f32 %f553, %f1290, %f1293; sub.ftz.f32 %f554, %f83, %f1289; mul.ftz.f32 %f555, %f554, %f553; sub.ftz.f32 %f556, %f1289, %f1292; sub.ftz.f32 %f557, %f82, %f1290; mul.ftz.f32 %f558, %f557, %f556; sub.ftz.f32 %f559, %f555, %f558; mul.ftz.f32 %f560, %f552, %f556; sub.ftz.f32 %f561, %f1291, %f1294; mul.ftz.f32 %f562, %f554, %f561; sub.ftz.f32 %f563, %f560, %f562; mul.ftz.f32 %f564, %f557, %f561; mul.ftz.f32 %f565, %f552, %f553; sub.ftz.f32 %f566, %f564, %f565; mul.ftz.f32 %f567, %f563, %f563; fma.rn.ftz.f32 %f568, %f559, %f559, %f567; fma.rn.ftz.f32 %f569, %f566, %f566, %f568; mul.ftz.f32 %f570, %f553, %f553; fma.rn.ftz.f32 %f571, %f561, %f561, %f570; fma.rn.ftz.f32 %f572, %f556, %f556, %f571; mul.ftz.f32 %f573, %f80, %f572; setp.lt.ftz.f32 %p25, %f569, %f573; @%p25 bra $L__BB103_33; sub.ftz.f32 %f575, %f81, %f1288; sub.ftz.f32 %f576, %f1287, %f1290; sub.ftz.f32 %f577, %f83, %f1286; mul.ftz.f32 %f578, %f577, %f576; sub.ftz.f32 %f579, %f1286, %f1289; sub.ftz.f32 %f580, %f82, %f1287; mul.ftz.f32 %f581, %f580, %f579; sub.ftz.f32 %f582, %f578, %f581; mul.ftz.f32 %f583, %f575, %f579; sub.ftz.f32 %f584, %f1288, %f1291; mul.ftz.f32 %f585, %f577, %f584; sub.ftz.f32 %f586, %f583, %f585; mul.ftz.f32 %f587, %f580, %f584; mul.ftz.f32 %f588, %f575, %f576; sub.ftz.f32 %f589, %f587, %f588; mul.ftz.f32 %f590, %f586, %f586; fma.rn.ftz.f32 %f591, %f582, %f582, %f590; fma.rn.ftz.f32 %f592, %f589, %f589, %f591; mul.ftz.f32 %f593, %f576, %f576; fma.rn.ftz.f32 %f594, %f584, %f584, %f593; fma.rn.ftz.f32 %f595, %f579, %f579, %f594; mul.ftz.f32 %f596, %f80, %f595; setp.lt.ftz.f32 %p26, %f592, %f596; @%p26 bra $L__BB103_33; mov.f32 %f1296, 0f00000000; $L__BB103_33: setp.eq.s32 %p27, %r2, 1; @%p27 bra $L__BB103_69; bra.uni $L__BB103_34; $L__BB103_69: cvt.u32.u64 %r141, %rd3; setp.eq.s32 %p52, %r141, -1; ld.local.f32 %f935, [%rd2+-24]; sub.ftz.f32 %f163, %f935, %f2; ld.local.f32 %f936, [%rd2+-20]; sub.ftz.f32 %f164, %f936, %f3; ld.local.f32 %f937, [%rd2+-16]; sub.ftz.f32 %f165, %f937, %f4; mov.f32 %f934, 0f00000000; mov.f32 %f1348, %f934; @%p52 bra $L__BB103_102; ld.local.u32 %r45, [%rd2+40]; and.b32 %r142, %r45, 1; setp.eq.b32 %p53, %r142, 1; mov.pred %p54, 0; xor.pred %p55, %p53, %p54; not.pred %p56, %p55; mov.f32 %f1348, %f934; @%p56 bra $L__BB103_102; and.b32 %r143, %r45, 64; setp.eq.s32 %p57, %r143, 0; @%p57 bra $L__BB103_85; bra.uni $L__BB103_72; $L__BB103_85: ld.const.u64 %rd330, [kernel_params+2232]; cvta.to.global.u64 %rd331, %rd330; mul.wide.s32 %rd332, %r141, 16; add.s64 %rd333, %rd331, %rd332; ld.const.u64 %rd334, [kernel_params+2256]; cvta.to.global.u64 %rd335, %rd334; ld.global.u32 %r170, [%rd333+12]; mul.wide.u32 %rd336, %r170, 12; add.s64 %rd337, %rd335, %rd336; ld.global.f32 %f1346, [%rd337]; ld.global.f32 %f1345, [%rd337+4]; ld.global.f32 %f1344, [%rd337+8]; add.s32 %r171, %r170, 1; mul.wide.u32 %rd338, %r171, 12; add.s64 %rd339, %rd335, %rd338; ld.global.f32 %f1343, [%rd339]; ld.global.f32 %f1342, [%rd339+4]; ld.global.f32 %f1341, [%rd339+8]; add.s32 %r172, %r170, 2; mul.wide.u32 %rd340, %r172, 12; add.s64 %rd341, %rd335, %rd340; ld.global.f32 %f1340, [%rd341]; ld.global.f32 %f1339, [%rd341+4]; ld.global.f32 %f1338, [%rd341+8]; bra.uni $L__BB103_86; $L__BB103_34: setp.ne.s32 %p28, %r2, 2; @%p28 bra $L__BB103_103; cvt.u32.u64 %r105, %rd3; setp.eq.s32 %p29, %r105, -1; ld.local.f32 %f599, [%rd2+-24]; sub.ftz.f32 %f85, %f599, %f5; ld.local.f32 %f600, [%rd2+-20]; sub.ftz.f32 %f86, %f600, %f6; ld.local.f32 %f601, [%rd2+-16]; sub.ftz.f32 %f87, %f601, %f7; mov.f32 %f598, 0f00000000; mov.f32 %f1322, %f598; @%p29 bra $L__BB103_68; ld.local.u32 %r24, [%rd2+40]; and.b32 %r106, %r24, 1; setp.eq.b32 %p30, %r106, 1; mov.pred %p31, 0; xor.pred %p32, %p30, %p31; not.pred %p33, %p32; mov.f32 %f1322, %f598; @%p33 bra $L__BB103_68; and.b32 %r107, %r24, 64; setp.eq.s32 %p34, %r107, 0; @%p34 bra $L__BB103_51; bra.uni $L__BB103_38; $L__BB103_51: ld.const.u64 %rd273, [kernel_params+2232]; cvta.to.global.u64 %rd274, %rd273; mul.wide.s32 %rd275, %r105, 16; add.s64 %rd276, %rd274, %rd275; ld.const.u64 %rd277, [kernel_params+2256]; cvta.to.global.u64 %rd278, %rd277; ld.global.u32 %r134, [%rd276+12]; mul.wide.u32 %rd279, %r134, 12; add.s64 %rd280, %rd278, %rd279; ld.global.f32 %f1320, [%rd280]; ld.global.f32 %f1319, [%rd280+4]; ld.global.f32 %f1318, [%rd280+8]; add.s32 %r135, %r134, 1; mul.wide.u32 %rd281, %r135, 12; add.s64 %rd282, %rd278, %rd281; ld.global.f32 %f1317, [%rd282]; ld.global.f32 %f1316, [%rd282+4]; ld.global.f32 %f1315, [%rd282+8]; add.s32 %r136, %r134, 2; mul.wide.u32 %rd283, %r136, 12; add.s64 %rd284, %rd278, %rd283; ld.global.f32 %f1314, [%rd284]; ld.global.f32 %f1313, [%rd284+4]; ld.global.f32 %f1312, [%rd284+8]; bra.uni $L__BB103_52; $L__BB103_72: ld.const.u64 %rd297, [kernel_params+2160]; cvta.to.global.u64 %rd298, %rd297; ld.local.u32 %r144, [%rd2+52]; mul.wide.s32 %rd299, %r144, 208; add.s64 %rd300, %rd298, %rd299; ld.global.u32 %r46, [%rd300+156]; ld.global.u32 %r47, [%rd300+152]; shl.b32 %r145, %r47, 1; cvt.rn.f32.s32 %f939, %r145; ld.local.f32 %f940, [%rd2+60]; mul.ftz.f32 %f166, %f940, %f939; cvt.rzi.ftz.s32.f32 %r146, %f166; add.s32 %r147, %r145, -1; min.s32 %r48, %r146, %r147; ld.const.u64 %rd301, [kernel_params+2312]; cvta.to.global.u64 %rd120, %rd301; ld.global.u32 %r188, [%rd300+164]; mul.wide.u32 %rd302, %r188, 16; add.s64 %rd303, %rd120, %rd302; ld.global.u16 %rs17, [%rd303+12]; ld.global.u32 %r187, [%rd303+8]; ld.global.u64 %rd394, [%rd303]; setp.eq.s64 %p58, %rd394, 11; @%p58 bra $L__BB103_77; $L__BB103_73: setp.eq.s64 %p59, %rd394, 0; @%p59 bra $L__BB103_75; add.s32 %r188, %r188, 2; bra.uni $L__BB103_76; $L__BB103_75: setp.eq.s16 %p60, %rs17, 0; mov.u32 %r191, -1; mov.u32 %r188, %r187; @%p60 bra $L__BB103_78; $L__BB103_76: mul.wide.u32 %rd304, %r188, 16; add.s64 %rd305, %rd120, %rd304; ld.global.u16 %rs17, [%rd305+12]; ld.global.u32 %r187, [%rd305+8]; ld.global.u64 %rd394, [%rd305]; setp.ne.s64 %p61, %rd394, 11; @%p61 bra $L__BB103_73; $L__BB103_77: setp.eq.s16 %p62, %rs17, 0; selp.b32 %r191, -1, %r187, %p62; $L__BB103_78: ld.const.u64 %rd306, [kernel_params+2232]; cvta.to.global.u64 %rd307, %rd306; shl.b64 %rd308, %rd3, 4; add.s64 %rd309, %rd307, %rd308; ld.global.v4.u32 {%r149, %r150, %r151, %r152}, [%rd309]; setp.eq.s32 %p63, %r48, %r47; @%p63 bra $L__BB103_80; bra.uni $L__BB103_79; $L__BB103_80: ld.const.u64 %rd315, [kernel_params+2256]; cvta.to.global.u64 %rd316, %rd315; mul.wide.u32 %rd317, %r152, 12; add.s64 %rd400, %rd316, %rd317; add.s32 %r159, %r152, 1; mul.wide.u32 %rd318, %r159, 12; add.s64 %rd397, %rd316, %rd318; add.s32 %r160, %r152, 2; mul.wide.u32 %rd319, %r160, 12; add.s64 %rd401, %rd316, %rd319; add.s64 %rd399, %rd400, 4; add.s64 %rd398, %rd400, 8; add.s64 %rd396, %rd397, 4; add.s64 %rd395, %rd397, 8; add.s64 %rd402, %rd401, 4; add.s64 %rd403, %rd401, 8; bra.uni $L__BB103_81; $L__BB103_38: ld.const.u64 %rd240, [kernel_params+2160]; cvta.to.global.u64 %rd241, %rd240; ld.local.u32 %r108, [%rd2+52]; mul.wide.s32 %rd242, %r108, 208; add.s64 %rd243, %rd241, %rd242; ld.global.u32 %r25, [%rd243+156]; ld.global.u32 %r26, [%rd243+152]; shl.b32 %r109, %r26, 1; cvt.rn.f32.s32 %f603, %r109; ld.local.f32 %f604, [%rd2+60]; mul.ftz.f32 %f88, %f604, %f603; cvt.rzi.ftz.s32.f32 %r110, %f88; add.s32 %r111, %r109, -1; min.s32 %r27, %r110, %r111; ld.const.u64 %rd244, [kernel_params+2312]; cvta.to.global.u64 %rd62, %rd244; ld.global.u32 %r183, [%rd243+164]; mul.wide.u32 %rd245, %r183, 16; add.s64 %rd246, %rd62, %rd245; ld.global.u16 %rs15, [%rd246+12]; ld.global.u32 %r182, [%rd246+8]; ld.global.u64 %rd375, [%rd246]; setp.eq.s64 %p35, %rd375, 11; @%p35 bra $L__BB103_43; $L__BB103_39: setp.eq.s64 %p36, %rd375, 0; @%p36 bra $L__BB103_41; add.s32 %r183, %r183, 2; bra.uni $L__BB103_42; $L__BB103_41: setp.eq.s16 %p37, %rs15, 0; mov.u32 %r186, -1; mov.u32 %r183, %r182; @%p37 bra $L__BB103_44; $L__BB103_42: mul.wide.u32 %rd247, %r183, 16; add.s64 %rd248, %rd62, %rd247; ld.global.u16 %rs15, [%rd248+12]; ld.global.u32 %r182, [%rd248+8]; ld.global.u64 %rd375, [%rd248]; setp.ne.s64 %p38, %rd375, 11; @%p38 bra $L__BB103_39; $L__BB103_43: setp.eq.s16 %p39, %rs15, 0; selp.b32 %r186, -1, %r182, %p39; $L__BB103_44: ld.const.u64 %rd249, [kernel_params+2232]; cvta.to.global.u64 %rd250, %rd249; shl.b64 %rd251, %rd3, 4; add.s64 %rd252, %rd250, %rd251; ld.global.v4.u32 {%r113, %r114, %r115, %r116}, [%rd252]; setp.eq.s32 %p40, %r27, %r26; @%p40 bra $L__BB103_46; bra.uni $L__BB103_45; $L__BB103_46: ld.const.u64 %rd258, [kernel_params+2256]; cvta.to.global.u64 %rd259, %rd258; mul.wide.u32 %rd260, %r116, 12; add.s64 %rd381, %rd259, %rd260; add.s32 %r123, %r116, 1; mul.wide.u32 %rd261, %r123, 12; add.s64 %rd378, %rd259, %rd261; add.s32 %r124, %r116, 2; mul.wide.u32 %rd262, %r124, 12; add.s64 %rd382, %rd259, %rd262; add.s64 %rd380, %rd381, 4; add.s64 %rd379, %rd381, 8; add.s64 %rd377, %rd378, 4; add.s64 %rd376, %rd378, 8; add.s64 %rd383, %rd382, 4; add.s64 %rd384, %rd382, 8; bra.uni $L__BB103_47; $L__BB103_79: setp.gt.s32 %p64, %r48, %r47; selp.b32 %r153, -1, 0, %p64; add.s32 %r154, %r48, %r153; mad.lo.s32 %r155, %r154, %r46, %r191; ld.const.u64 %rd310, [kernel_params+2336]; cvta.to.global.u64 %rd311, %rd310; add.s32 %r156, %r149, %r155; mul.wide.u32 %rd312, %r156, 12; add.s64 %rd400, %rd311, %rd312; add.s32 %r157, %r150, %r155; mul.wide.u32 %rd313, %r157, 12; add.s64 %rd397, %rd311, %rd313; add.s32 %r158, %r151, %r155; mul.wide.u32 %rd314, %r158, 12; add.s64 %rd401, %rd311, %rd314; add.s64 %rd399, %rd400, 4; add.s64 %rd398, %rd400, 8; add.s64 %rd396, %rd397, 4; add.s64 %rd395, %rd397, 8; add.s64 %rd402, %rd401, 4; add.s64 %rd403, %rd401, 8; $L__BB103_81: ld.global.f32 %f167, [%rd400]; ld.global.f32 %f168, [%rd399]; ld.global.f32 %f169, [%rd398]; ld.global.f32 %f170, [%rd397]; ld.global.f32 %f171, [%rd396]; ld.global.f32 %f172, [%rd395]; ld.global.f32 %f173, [%rd403]; ld.global.f32 %f174, [%rd402]; ld.global.f32 %f175, [%rd401]; add.s32 %r63, %r48, 1; setp.eq.s32 %p65, %r63, %r47; @%p65 bra $L__BB103_83; bra.uni $L__BB103_82; $L__BB103_83: ld.const.u64 %rd325, [kernel_params+2256]; cvta.to.global.u64 %rd326, %rd325; mul.wide.u32 %rd327, %r152, 12; add.s64 %rd404, %rd326, %rd327; add.s32 %r167, %r152, 1; mul.wide.u32 %rd328, %r167, 12; add.s64 %rd407, %rd326, %rd328; add.s32 %r168, %r152, 2; mul.wide.u32 %rd329, %r168, 12; add.s64 %rd410, %rd326, %rd329; add.s64 %rd405, %rd404, 4; add.s64 %rd406, %rd404, 8; add.s64 %rd408, %rd407, 4; add.s64 %rd409, %rd407, 8; add.s64 %rd411, %rd410, 4; add.s64 %rd412, %rd410, 8; bra.uni $L__BB103_84; $L__BB103_82: setp.ge.s32 %p66, %r48, %r47; selp.b32 %r161, -1, 0, %p66; add.s32 %r162, %r63, %r161; mad.lo.s32 %r163, %r162, %r46, %r191; ld.const.u64 %rd320, [kernel_params+2336]; cvta.to.global.u64 %rd321, %rd320; add.s32 %r164, %r149, %r163; mul.wide.u32 %rd322, %r164, 12; add.s64 %rd404, %rd321, %rd322; add.s32 %r165, %r150, %r163; mul.wide.u32 %rd323, %r165, 12; add.s64 %rd407, %rd321, %rd323; add.s32 %r166, %r151, %r163; mul.wide.u32 %rd324, %r166, 12; add.s64 %rd410, %rd321, %rd324; add.s64 %rd405, %rd404, 4; add.s64 %rd406, %rd404, 8; add.s64 %rd408, %rd407, 4; add.s64 %rd409, %rd407, 8; add.s64 %rd411, %rd410, 4; add.s64 %rd412, %rd410, 8; $L__BB103_84: cvt.rn.f32.s32 %f941, %r48; sub.ftz.f32 %f942, %f166, %f941; mov.f32 %f943, 0f3F800000; sub.ftz.f32 %f944, %f943, %f942; mul.ftz.f32 %f945, %f944, %f167; mul.ftz.f32 %f946, %f944, %f168; mul.ftz.f32 %f947, %f944, %f169; ld.global.f32 %f948, [%rd404]; ld.global.f32 %f949, [%rd405]; ld.global.f32 %f950, [%rd406]; fma.rn.ftz.f32 %f1346, %f942, %f948, %f945; fma.rn.ftz.f32 %f1345, %f942, %f949, %f946; fma.rn.ftz.f32 %f1344, %f942, %f950, %f947; mul.ftz.f32 %f951, %f944, %f170; mul.ftz.f32 %f952, %f944, %f171; mul.ftz.f32 %f953, %f944, %f172; ld.global.f32 %f954, [%rd407]; ld.global.f32 %f955, [%rd408]; ld.global.f32 %f956, [%rd409]; fma.rn.ftz.f32 %f1343, %f942, %f954, %f951; fma.rn.ftz.f32 %f1342, %f942, %f955, %f952; fma.rn.ftz.f32 %f1341, %f942, %f956, %f953; mul.ftz.f32 %f957, %f944, %f175; mul.ftz.f32 %f958, %f944, %f174; mul.ftz.f32 %f959, %f944, %f173; ld.global.f32 %f960, [%rd410]; ld.global.f32 %f961, [%rd411]; ld.global.f32 %f962, [%rd412]; fma.rn.ftz.f32 %f1340, %f942, %f960, %f957; fma.rn.ftz.f32 %f1339, %f942, %f961, %f958; fma.rn.ftz.f32 %f1338, %f942, %f962, %f959; $L__BB103_86: ld.local.u32 %r64, [%rd2+32]; and.b32 %r173, %r64, 4; setp.ne.s32 %p67, %r173, 0; @%p67 bra $L__BB103_96; and.b32 %r65, %r64, 2; setp.eq.s32 %p68, %r65, 0; @%p68 bra $L__BB103_89; bra.uni $L__BB103_88; $L__BB103_89: ld.const.u64 %rd342, [kernel_params+2160]; cvta.to.global.u64 %rd343, %rd342; ld.local.u32 %r174, [%rd2+52]; mul.wide.s32 %rd344, %r174, 208; add.s64 %rd345, %rd343, %rd344; ld.global.v4.f32 {%f996, %f997, %f998, %f999}, [%rd345]; ld.global.v4.f32 {%f1004, %f1005, %f1006, %f1007}, [%rd345+16]; ld.global.v4.f32 {%f1012, %f1013, %f1014, %f1015}, [%rd345+32]; mul.ftz.f32 %f1020, %f997, %f1345; fma.rn.ftz.f32 %f1021, %f1346, %f996, %f1020; fma.rn.ftz.f32 %f1022, %f998, %f1344, %f1021; add.ftz.f32 %f1332, %f999, %f1022; mul.ftz.f32 %f1023, %f1005, %f1345; fma.rn.ftz.f32 %f1024, %f1346, %f1004, %f1023; fma.rn.ftz.f32 %f1025, %f1006, %f1344, %f1024; add.ftz.f32 %f1333, %f1007, %f1025; mul.ftz.f32 %f1026, %f1013, %f1345; fma.rn.ftz.f32 %f1027, %f1346, %f1012, %f1026; fma.rn.ftz.f32 %f1028, %f1014, %f1344, %f1027; add.ftz.f32 %f1344, %f1015, %f1028; bra.uni $L__BB103_90; $L__BB103_88: ld.local.v4.f32 {%f963, %f964, %f965, %f966}, [%rd2+120]; mul.ftz.f32 %f971, %f1345, %f964; fma.rn.ftz.f32 %f972, %f1346, %f963, %f971; fma.rn.ftz.f32 %f973, %f1344, %f965, %f972; add.ftz.f32 %f1332, %f966, %f973; ld.local.v4.f32 {%f974, %f975, %f976, %f977}, [%rd2+136]; mul.ftz.f32 %f982, %f1345, %f975; fma.rn.ftz.f32 %f983, %f1346, %f974, %f982; fma.rn.ftz.f32 %f984, %f1344, %f976, %f983; add.ftz.f32 %f1333, %f977, %f984; ld.local.v4.f32 {%f985, %f986, %f987, %f988}, [%rd2+152]; mul.ftz.f32 %f993, %f1345, %f986; fma.rn.ftz.f32 %f994, %f1346, %f985, %f993; fma.rn.ftz.f32 %f995, %f1344, %f987, %f994; add.ftz.f32 %f1344, %f988, %f995; $L__BB103_90: @%p68 bra $L__BB103_92; bra.uni $L__BB103_91; $L__BB103_92: ld.const.u64 %rd346, [kernel_params+2160]; cvta.to.global.u64 %rd347, %rd346; ld.local.u32 %r175, [%rd2+52]; mul.wide.s32 %rd348, %r175, 208; add.s64 %rd349, %rd347, %rd348; ld.global.v4.f32 {%f1062, %f1063, %f1064, %f1065}, [%rd349]; ld.global.v4.f32 {%f1070, %f1071, %f1072, %f1073}, [%rd349+16]; ld.global.v4.f32 {%f1078, %f1079, %f1080, %f1081}, [%rd349+32]; mul.ftz.f32 %f1086, %f1063, %f1342; fma.rn.ftz.f32 %f1087, %f1343, %f1062, %f1086; fma.rn.ftz.f32 %f1088, %f1064, %f1341, %f1087; add.ftz.f32 %f1335, %f1065, %f1088; mul.ftz.f32 %f1089, %f1071, %f1342; fma.rn.ftz.f32 %f1090, %f1343, %f1070, %f1089; fma.rn.ftz.f32 %f1091, %f1072, %f1341, %f1090; add.ftz.f32 %f1336, %f1073, %f1091; mul.ftz.f32 %f1092, %f1079, %f1342; fma.rn.ftz.f32 %f1093, %f1343, %f1078, %f1092; fma.rn.ftz.f32 %f1094, %f1080, %f1341, %f1093; add.ftz.f32 %f1341, %f1081, %f1094; bra.uni $L__BB103_93; $L__BB103_91: ld.local.v4.f32 {%f1029, %f1030, %f1031, %f1032}, [%rd2+120]; mul.ftz.f32 %f1037, %f1342, %f1030; fma.rn.ftz.f32 %f1038, %f1343, %f1029, %f1037; fma.rn.ftz.f32 %f1039, %f1341, %f1031, %f1038; add.ftz.f32 %f1335, %f1032, %f1039; ld.local.v4.f32 {%f1040, %f1041, %f1042, %f1043}, [%rd2+136]; mul.ftz.f32 %f1048, %f1342, %f1041; fma.rn.ftz.f32 %f1049, %f1343, %f1040, %f1048; fma.rn.ftz.f32 %f1050, %f1341, %f1042, %f1049; add.ftz.f32 %f1336, %f1043, %f1050; ld.local.v4.f32 {%f1051, %f1052, %f1053, %f1054}, [%rd2+152]; mul.ftz.f32 %f1059, %f1342, %f1052; fma.rn.ftz.f32 %f1060, %f1343, %f1051, %f1059; fma.rn.ftz.f32 %f1061, %f1341, %f1053, %f1060; add.ftz.f32 %f1341, %f1054, %f1061; $L__BB103_93: @%p68 bra $L__BB103_95; bra.uni $L__BB103_94; $L__BB103_95: ld.const.u64 %rd350, [kernel_params+2160]; cvta.to.global.u64 %rd351, %rd350; ld.local.u32 %r176, [%rd2+52]; mul.wide.s32 %rd352, %r176, 208; add.s64 %rd353, %rd351, %rd352; ld.global.v4.f32 {%f1128, %f1129, %f1130, %f1131}, [%rd353]; ld.global.v4.f32 {%f1136, %f1137, %f1138, %f1139}, [%rd353+16]; ld.global.v4.f32 {%f1144, %f1145, %f1146, %f1147}, [%rd353+32]; mul.ftz.f32 %f1152, %f1129, %f1339; fma.rn.ftz.f32 %f1153, %f1340, %f1128, %f1152; fma.rn.ftz.f32 %f1154, %f1130, %f1338, %f1153; add.ftz.f32 %f224, %f1131, %f1154; mul.ftz.f32 %f1155, %f1137, %f1339; fma.rn.ftz.f32 %f1156, %f1340, %f1136, %f1155; fma.rn.ftz.f32 %f1157, %f1138, %f1338, %f1156; add.ftz.f32 %f225, %f1139, %f1157; mul.ftz.f32 %f1158, %f1145, %f1339; fma.rn.ftz.f32 %f1159, %f1340, %f1144, %f1158; fma.rn.ftz.f32 %f1160, %f1146, %f1338, %f1159; add.ftz.f32 %f1338, %f1147, %f1160; mov.f32 %f1339, %f225; mov.f32 %f1340, %f224; mov.f32 %f1342, %f1336; mov.f32 %f1343, %f1335; mov.f32 %f1345, %f1333; mov.f32 %f1346, %f1332; bra.uni $L__BB103_96; $L__BB103_94: ld.local.v4.f32 {%f1095, %f1096, %f1097, %f1098}, [%rd2+120]; mul.ftz.f32 %f1103, %f1339, %f1096; fma.rn.ftz.f32 %f1104, %f1340, %f1095, %f1103; fma.rn.ftz.f32 %f1105, %f1338, %f1097, %f1104; add.ftz.f32 %f221, %f1098, %f1105; ld.local.v4.f32 {%f1106, %f1107, %f1108, %f1109}, [%rd2+136]; mul.ftz.f32 %f1114, %f1339, %f1107; fma.rn.ftz.f32 %f1115, %f1340, %f1106, %f1114; fma.rn.ftz.f32 %f1116, %f1338, %f1108, %f1115; add.ftz.f32 %f222, %f1109, %f1116; ld.local.v4.f32 {%f1117, %f1118, %f1119, %f1120}, [%rd2+152]; mul.ftz.f32 %f1125, %f1339, %f1118; fma.rn.ftz.f32 %f1126, %f1340, %f1117, %f1125; fma.rn.ftz.f32 %f1127, %f1338, %f1119, %f1126; add.ftz.f32 %f1338, %f1120, %f1127; mov.f32 %f1339, %f222; mov.f32 %f1340, %f221; mov.f32 %f1342, %f1336; mov.f32 %f1343, %f1335; mov.f32 %f1345, %f1333; mov.f32 %f1346, %f1332; $L__BB103_96: setp.eq.s32 %p71, %r1, 0; mov.f32 %f1348, 0f3F800000; mov.f32 %f1347, %f1348; @%p71 bra $L__BB103_98; ld.local.f32 %f1162, [%rd2+12]; ld.local.f32 %f1163, [%rd2+16]; mul.ftz.f32 %f1164, %f3, %f1163; fma.rn.ftz.f32 %f1165, %f2, %f1162, %f1164; ld.local.f32 %f1166, [%rd2+20]; fma.rn.ftz.f32 %f1167, %f4, %f1166, %f1165; mul.ftz.f32 %f1168, %f1162, %f1167; mul.ftz.f32 %f1169, %f1163, %f1167; mul.ftz.f32 %f1170, %f1167, %f1166; sub.ftz.f32 %f1171, %f2, %f1168; sub.ftz.f32 %f1172, %f3, %f1169; sub.ftz.f32 %f1173, %f4, %f1170; mul.ftz.f32 %f1174, %f1172, %f1172; fma.rn.ftz.f32 %f1175, %f1171, %f1171, %f1174; fma.rn.ftz.f32 %f1176, %f1173, %f1173, %f1175; sqrt.approx.ftz.f32 %f1177, %f1176; mul.ftz.f32 %f1178, %f6, %f1163; fma.rn.ftz.f32 %f1179, %f5, %f1162, %f1178; fma.rn.ftz.f32 %f1180, %f7, %f1166, %f1179; mul.ftz.f32 %f1181, %f1162, %f1180; mul.ftz.f32 %f1182, %f1163, %f1180; mul.ftz.f32 %f1183, %f1180, %f1166; sub.ftz.f32 %f1184, %f5, %f1181; sub.ftz.f32 %f1185, %f6, %f1182; sub.ftz.f32 %f1186, %f7, %f1183; mul.ftz.f32 %f1187, %f1185, %f1185; fma.rn.ftz.f32 %f1188, %f1184, %f1184, %f1187; fma.rn.ftz.f32 %f1189, %f1186, %f1186, %f1188; sqrt.approx.ftz.f32 %f1190, %f1189; add.ftz.f32 %f1191, %f1177, %f1190; mul.ftz.f32 %f1347, %f1191, 0f3F000000; $L__BB103_98: mul.ftz.f32 %f1193, %f1, 0f3F000000; mul.ftz.f32 %f1194, %f1193, %f1347; mul.ftz.f32 %f238, %f1194, %f1194; sub.ftz.f32 %f1195, %f1345, %f1339; sub.ftz.f32 %f1196, %f165, %f1344; mul.ftz.f32 %f1197, %f1196, %f1195; sub.ftz.f32 %f1198, %f1344, %f1338; sub.ftz.f32 %f1199, %f164, %f1345; mul.ftz.f32 %f1200, %f1199, %f1198; sub.ftz.f32 %f1201, %f1197, %f1200; sub.ftz.f32 %f1202, %f163, %f1346; mul.ftz.f32 %f1203, %f1202, %f1198; sub.ftz.f32 %f1204, %f1346, %f1340; mul.ftz.f32 %f1205, %f1196, %f1204; sub.ftz.f32 %f1206, %f1203, %f1205; mul.ftz.f32 %f1207, %f1199, %f1204; mul.ftz.f32 %f1208, %f1202, %f1195; sub.ftz.f32 %f1209, %f1207, %f1208; mul.ftz.f32 %f1210, %f1206, %f1206; fma.rn.ftz.f32 %f1211, %f1201, %f1201, %f1210; fma.rn.ftz.f32 %f1212, %f1209, %f1209, %f1211; mul.ftz.f32 %f1213, %f1195, %f1195; fma.rn.ftz.f32 %f1214, %f1204, %f1204, %f1213; fma.rn.ftz.f32 %f1215, %f1198, %f1198, %f1214; mul.ftz.f32 %f1216, %f238, %f1215; setp.lt.ftz.f32 %p72, %f1212, %f1216; @%p72 bra $L__BB103_102; sub.ftz.f32 %f1218, %f163, %f1343; sub.ftz.f32 %f1219, %f1342, %f1345; sub.ftz.f32 %f1220, %f165, %f1341; mul.ftz.f32 %f1221, %f1220, %f1219; sub.ftz.f32 %f1222, %f1341, %f1344; sub.ftz.f32 %f1223, %f164, %f1342; mul.ftz.f32 %f1224, %f1223, %f1222; sub.ftz.f32 %f1225, %f1221, %f1224; mul.ftz.f32 %f1226, %f1218, %f1222; sub.ftz.f32 %f1227, %f1343, %f1346; mul.ftz.f32 %f1228, %f1220, %f1227; sub.ftz.f32 %f1229, %f1226, %f1228; mul.ftz.f32 %f1230, %f1223, %f1227; mul.ftz.f32 %f1231, %f1218, %f1219; sub.ftz.f32 %f1232, %f1230, %f1231; mul.ftz.f32 %f1233, %f1229, %f1229; fma.rn.ftz.f32 %f1234, %f1225, %f1225, %f1233; fma.rn.ftz.f32 %f1235, %f1232, %f1232, %f1234; mul.ftz.f32 %f1236, %f1219, %f1219; fma.rn.ftz.f32 %f1237, %f1227, %f1227, %f1236; fma.rn.ftz.f32 %f1238, %f1222, %f1222, %f1237; mul.ftz.f32 %f1239, %f238, %f1238; setp.lt.ftz.f32 %p73, %f1235, %f1239; @%p73 bra $L__BB103_102; sub.ftz.f32 %f1241, %f163, %f1340; sub.ftz.f32 %f1242, %f1339, %f1342; sub.ftz.f32 %f1243, %f165, %f1338; mul.ftz.f32 %f1244, %f1243, %f1242; sub.ftz.f32 %f1245, %f1338, %f1341; sub.ftz.f32 %f1246, %f164, %f1339; mul.ftz.f32 %f1247, %f1246, %f1245; sub.ftz.f32 %f1248, %f1244, %f1247; mul.ftz.f32 %f1249, %f1241, %f1245; sub.ftz.f32 %f1250, %f1340, %f1343; mul.ftz.f32 %f1251, %f1243, %f1250; sub.ftz.f32 %f1252, %f1249, %f1251; mul.ftz.f32 %f1253, %f1246, %f1250; mul.ftz.f32 %f1254, %f1241, %f1242; sub.ftz.f32 %f1255, %f1253, %f1254; mul.ftz.f32 %f1256, %f1252, %f1252; fma.rn.ftz.f32 %f1257, %f1248, %f1248, %f1256; fma.rn.ftz.f32 %f1258, %f1255, %f1255, %f1257; mul.ftz.f32 %f1259, %f1242, %f1242; fma.rn.ftz.f32 %f1260, %f1250, %f1250, %f1259; fma.rn.ftz.f32 %f1261, %f1245, %f1245, %f1260; mul.ftz.f32 %f1262, %f238, %f1261; setp.lt.ftz.f32 %p74, %f1258, %f1262; @%p74 bra $L__BB103_102; mov.f32 %f1348, %f934; $L__BB103_102: mul.ftz.f32 %f1264, %f3, %f3; fma.rn.ftz.f32 %f1265, %f2, %f2, %f1264; fma.rn.ftz.f32 %f1266, %f4, %f4, %f1265; sqrt.approx.ftz.f32 %f1267, %f1266; sub.ftz.f32 %f1268, %f1296, %f1348; div.approx.ftz.f32 %f1269, %f1268, %f1267; add.ftz.f32 %f1296, %f1296, %f1269; bra.uni $L__BB103_103; $L__BB103_45: setp.gt.s32 %p41, %r27, %r26; selp.b32 %r117, -1, 0, %p41; add.s32 %r118, %r27, %r117; mad.lo.s32 %r119, %r118, %r25, %r186; ld.const.u64 %rd253, [kernel_params+2336]; cvta.to.global.u64 %rd254, %rd253; add.s32 %r120, %r113, %r119; mul.wide.u32 %rd255, %r120, 12; add.s64 %rd381, %rd254, %rd255; add.s32 %r121, %r114, %r119; mul.wide.u32 %rd256, %r121, 12; add.s64 %rd378, %rd254, %rd256; add.s32 %r122, %r115, %r119; mul.wide.u32 %rd257, %r122, 12; add.s64 %rd382, %rd254, %rd257; add.s64 %rd380, %rd381, 4; add.s64 %rd379, %rd381, 8; add.s64 %rd377, %rd378, 4; add.s64 %rd376, %rd378, 8; add.s64 %rd383, %rd382, 4; add.s64 %rd384, %rd382, 8; $L__BB103_47: ld.global.f32 %f89, [%rd381]; ld.global.f32 %f90, [%rd380]; ld.global.f32 %f91, [%rd379]; ld.global.f32 %f92, [%rd378]; ld.global.f32 %f93, [%rd377]; ld.global.f32 %f94, [%rd376]; ld.global.f32 %f95, [%rd384]; ld.global.f32 %f96, [%rd383]; ld.global.f32 %f97, [%rd382]; add.s32 %r42, %r27, 1; setp.eq.s32 %p42, %r42, %r26; @%p42 bra $L__BB103_49; bra.uni $L__BB103_48; $L__BB103_49: ld.const.u64 %rd268, [kernel_params+2256]; cvta.to.global.u64 %rd269, %rd268; mul.wide.u32 %rd270, %r116, 12; add.s64 %rd385, %rd269, %rd270; add.s32 %r131, %r116, 1; mul.wide.u32 %rd271, %r131, 12; add.s64 %rd388, %rd269, %rd271; add.s32 %r132, %r116, 2; mul.wide.u32 %rd272, %r132, 12; add.s64 %rd391, %rd269, %rd272; add.s64 %rd386, %rd385, 4; add.s64 %rd387, %rd385, 8; add.s64 %rd389, %rd388, 4; add.s64 %rd390, %rd388, 8; add.s64 %rd392, %rd391, 4; add.s64 %rd393, %rd391, 8; bra.uni $L__BB103_50; $L__BB103_48: setp.ge.s32 %p43, %r27, %r26; selp.b32 %r125, -1, 0, %p43; add.s32 %r126, %r42, %r125; mad.lo.s32 %r127, %r126, %r25, %r186; ld.const.u64 %rd263, [kernel_params+2336]; cvta.to.global.u64 %rd264, %rd263; add.s32 %r128, %r113, %r127; mul.wide.u32 %rd265, %r128, 12; add.s64 %rd385, %rd264, %rd265; add.s32 %r129, %r114, %r127; mul.wide.u32 %rd266, %r129, 12; add.s64 %rd388, %rd264, %rd266; add.s32 %r130, %r115, %r127; mul.wide.u32 %rd267, %r130, 12; add.s64 %rd391, %rd264, %rd267; add.s64 %rd386, %rd385, 4; add.s64 %rd387, %rd385, 8; add.s64 %rd389, %rd388, 4; add.s64 %rd390, %rd388, 8; add.s64 %rd392, %rd391, 4; add.s64 %rd393, %rd391, 8; $L__BB103_50: cvt.rn.f32.s32 %f605, %r27; sub.ftz.f32 %f606, %f88, %f605; mov.f32 %f607, 0f3F800000; sub.ftz.f32 %f608, %f607, %f606; mul.ftz.f32 %f609, %f608, %f89; mul.ftz.f32 %f610, %f608, %f90; mul.ftz.f32 %f611, %f608, %f91; ld.global.f32 %f612, [%rd385]; ld.global.f32 %f613, [%rd386]; ld.global.f32 %f614, [%rd387]; fma.rn.ftz.f32 %f1320, %f606, %f612, %f609; fma.rn.ftz.f32 %f1319, %f606, %f613, %f610; fma.rn.ftz.f32 %f1318, %f606, %f614, %f611; mul.ftz.f32 %f615, %f608, %f92; mul.ftz.f32 %f616, %f608, %f93; mul.ftz.f32 %f617, %f608, %f94; ld.global.f32 %f618, [%rd388]; ld.global.f32 %f619, [%rd389]; ld.global.f32 %f620, [%rd390]; fma.rn.ftz.f32 %f1317, %f606, %f618, %f615; fma.rn.ftz.f32 %f1316, %f606, %f619, %f616; fma.rn.ftz.f32 %f1315, %f606, %f620, %f617; mul.ftz.f32 %f621, %f608, %f97; mul.ftz.f32 %f622, %f608, %f96; mul.ftz.f32 %f623, %f608, %f95; ld.global.f32 %f624, [%rd391]; ld.global.f32 %f625, [%rd392]; ld.global.f32 %f626, [%rd393]; fma.rn.ftz.f32 %f1314, %f606, %f624, %f621; fma.rn.ftz.f32 %f1313, %f606, %f625, %f622; fma.rn.ftz.f32 %f1312, %f606, %f626, %f623; $L__BB103_52: ld.local.u32 %r43, [%rd2+32]; and.b32 %r137, %r43, 4; setp.ne.s32 %p44, %r137, 0; @%p44 bra $L__BB103_62; and.b32 %r44, %r43, 2; setp.eq.s32 %p45, %r44, 0; @%p45 bra $L__BB103_55; bra.uni $L__BB103_54; $L__BB103_55: ld.const.u64 %rd285, [kernel_params+2160]; cvta.to.global.u64 %rd286, %rd285; ld.local.u32 %r138, [%rd2+52]; mul.wide.s32 %rd287, %r138, 208; add.s64 %rd288, %rd286, %rd287; ld.global.v4.f32 {%f660, %f661, %f662, %f663}, [%rd288]; ld.global.v4.f32 {%f668, %f669, %f670, %f671}, [%rd288+16]; ld.global.v4.f32 {%f676, %f677, %f678, %f679}, [%rd288+32]; mul.ftz.f32 %f684, %f661, %f1319; fma.rn.ftz.f32 %f685, %f1320, %f660, %f684; fma.rn.ftz.f32 %f686, %f662, %f1318, %f685; add.ftz.f32 %f1306, %f663, %f686; mul.ftz.f32 %f687, %f669, %f1319; fma.rn.ftz.f32 %f688, %f1320, %f668, %f687; fma.rn.ftz.f32 %f689, %f670, %f1318, %f688; add.ftz.f32 %f1307, %f671, %f689; mul.ftz.f32 %f690, %f677, %f1319; fma.rn.ftz.f32 %f691, %f1320, %f676, %f690; fma.rn.ftz.f32 %f692, %f678, %f1318, %f691; add.ftz.f32 %f1318, %f679, %f692; bra.uni $L__BB103_56; $L__BB103_54: ld.local.v4.f32 {%f627, %f628, %f629, %f630}, [%rd2+120]; mul.ftz.f32 %f635, %f1319, %f628; fma.rn.ftz.f32 %f636, %f1320, %f627, %f635; fma.rn.ftz.f32 %f637, %f1318, %f629, %f636; add.ftz.f32 %f1306, %f630, %f637; ld.local.v4.f32 {%f638, %f639, %f640, %f641}, [%rd2+136]; mul.ftz.f32 %f646, %f1319, %f639; fma.rn.ftz.f32 %f647, %f1320, %f638, %f646; fma.rn.ftz.f32 %f648, %f1318, %f640, %f647; add.ftz.f32 %f1307, %f641, %f648; ld.local.v4.f32 {%f649, %f650, %f651, %f652}, [%rd2+152]; mul.ftz.f32 %f657, %f1319, %f650; fma.rn.ftz.f32 %f658, %f1320, %f649, %f657; fma.rn.ftz.f32 %f659, %f1318, %f651, %f658; add.ftz.f32 %f1318, %f652, %f659; $L__BB103_56: @%p45 bra $L__BB103_58; bra.uni $L__BB103_57; $L__BB103_58: ld.const.u64 %rd289, [kernel_params+2160]; cvta.to.global.u64 %rd290, %rd289; ld.local.u32 %r139, [%rd2+52]; mul.wide.s32 %rd291, %r139, 208; add.s64 %rd292, %rd290, %rd291; ld.global.v4.f32 {%f726, %f727, %f728, %f729}, [%rd292]; ld.global.v4.f32 {%f734, %f735, %f736, %f737}, [%rd292+16]; ld.global.v4.f32 {%f742, %f743, %f744, %f745}, [%rd292+32]; mul.ftz.f32 %f750, %f727, %f1316; fma.rn.ftz.f32 %f751, %f1317, %f726, %f750; fma.rn.ftz.f32 %f752, %f728, %f1315, %f751; add.ftz.f32 %f1309, %f729, %f752; mul.ftz.f32 %f753, %f735, %f1316; fma.rn.ftz.f32 %f754, %f1317, %f734, %f753; fma.rn.ftz.f32 %f755, %f736, %f1315, %f754; add.ftz.f32 %f1310, %f737, %f755; mul.ftz.f32 %f756, %f743, %f1316; fma.rn.ftz.f32 %f757, %f1317, %f742, %f756; fma.rn.ftz.f32 %f758, %f744, %f1315, %f757; add.ftz.f32 %f1315, %f745, %f758; bra.uni $L__BB103_59; $L__BB103_57: ld.local.v4.f32 {%f693, %f694, %f695, %f696}, [%rd2+120]; mul.ftz.f32 %f701, %f1316, %f694; fma.rn.ftz.f32 %f702, %f1317, %f693, %f701; fma.rn.ftz.f32 %f703, %f1315, %f695, %f702; add.ftz.f32 %f1309, %f696, %f703; ld.local.v4.f32 {%f704, %f705, %f706, %f707}, [%rd2+136]; mul.ftz.f32 %f712, %f1316, %f705; fma.rn.ftz.f32 %f713, %f1317, %f704, %f712; fma.rn.ftz.f32 %f714, %f1315, %f706, %f713; add.ftz.f32 %f1310, %f707, %f714; ld.local.v4.f32 {%f715, %f716, %f717, %f718}, [%rd2+152]; mul.ftz.f32 %f723, %f1316, %f716; fma.rn.ftz.f32 %f724, %f1317, %f715, %f723; fma.rn.ftz.f32 %f725, %f1315, %f717, %f724; add.ftz.f32 %f1315, %f718, %f725; $L__BB103_59: @%p45 bra $L__BB103_61; bra.uni $L__BB103_60; $L__BB103_61: ld.const.u64 %rd293, [kernel_params+2160]; cvta.to.global.u64 %rd294, %rd293; ld.local.u32 %r140, [%rd2+52]; mul.wide.s32 %rd295, %r140, 208; add.s64 %rd296, %rd294, %rd295; ld.global.v4.f32 {%f792, %f793, %f794, %f795}, [%rd296]; ld.global.v4.f32 {%f800, %f801, %f802, %f803}, [%rd296+16]; ld.global.v4.f32 {%f808, %f809, %f810, %f811}, [%rd296+32]; mul.ftz.f32 %f816, %f793, %f1313; fma.rn.ftz.f32 %f817, %f1314, %f792, %f816; fma.rn.ftz.f32 %f818, %f794, %f1312, %f817; add.ftz.f32 %f146, %f795, %f818; mul.ftz.f32 %f819, %f801, %f1313; fma.rn.ftz.f32 %f820, %f1314, %f800, %f819; fma.rn.ftz.f32 %f821, %f802, %f1312, %f820; add.ftz.f32 %f147, %f803, %f821; mul.ftz.f32 %f822, %f809, %f1313; fma.rn.ftz.f32 %f823, %f1314, %f808, %f822; fma.rn.ftz.f32 %f824, %f810, %f1312, %f823; add.ftz.f32 %f1312, %f811, %f824; mov.f32 %f1313, %f147; mov.f32 %f1314, %f146; mov.f32 %f1316, %f1310; mov.f32 %f1317, %f1309; mov.f32 %f1319, %f1307; mov.f32 %f1320, %f1306; bra.uni $L__BB103_62; $L__BB103_60: ld.local.v4.f32 {%f759, %f760, %f761, %f762}, [%rd2+120]; mul.ftz.f32 %f767, %f1313, %f760; fma.rn.ftz.f32 %f768, %f1314, %f759, %f767; fma.rn.ftz.f32 %f769, %f1312, %f761, %f768; add.ftz.f32 %f143, %f762, %f769; ld.local.v4.f32 {%f770, %f771, %f772, %f773}, [%rd2+136]; mul.ftz.f32 %f778, %f1313, %f771; fma.rn.ftz.f32 %f779, %f1314, %f770, %f778; fma.rn.ftz.f32 %f780, %f1312, %f772, %f779; add.ftz.f32 %f144, %f773, %f780; ld.local.v4.f32 {%f781, %f782, %f783, %f784}, [%rd2+152]; mul.ftz.f32 %f789, %f1313, %f782; fma.rn.ftz.f32 %f790, %f1314, %f781, %f789; fma.rn.ftz.f32 %f791, %f1312, %f783, %f790; add.ftz.f32 %f1312, %f784, %f791; mov.f32 %f1313, %f144; mov.f32 %f1314, %f143; mov.f32 %f1316, %f1310; mov.f32 %f1317, %f1309; mov.f32 %f1319, %f1307; mov.f32 %f1320, %f1306; $L__BB103_62: setp.eq.s32 %p48, %r1, 0; mov.f32 %f1322, 0f3F800000; mov.f32 %f1321, %f1322; @%p48 bra $L__BB103_64; ld.local.f32 %f826, [%rd2+12]; ld.local.f32 %f827, [%rd2+16]; mul.ftz.f32 %f828, %f3, %f827; fma.rn.ftz.f32 %f829, %f2, %f826, %f828; ld.local.f32 %f830, [%rd2+20]; fma.rn.ftz.f32 %f831, %f4, %f830, %f829; mul.ftz.f32 %f832, %f826, %f831; mul.ftz.f32 %f833, %f827, %f831; mul.ftz.f32 %f834, %f831, %f830; sub.ftz.f32 %f835, %f2, %f832; sub.ftz.f32 %f836, %f3, %f833; sub.ftz.f32 %f837, %f4, %f834; mul.ftz.f32 %f838, %f836, %f836; fma.rn.ftz.f32 %f839, %f835, %f835, %f838; fma.rn.ftz.f32 %f840, %f837, %f837, %f839; sqrt.approx.ftz.f32 %f841, %f840; mul.ftz.f32 %f842, %f6, %f827; fma.rn.ftz.f32 %f843, %f5, %f826, %f842; fma.rn.ftz.f32 %f844, %f7, %f830, %f843; mul.ftz.f32 %f845, %f826, %f844; mul.ftz.f32 %f846, %f827, %f844; mul.ftz.f32 %f847, %f844, %f830; sub.ftz.f32 %f848, %f5, %f845; sub.ftz.f32 %f849, %f6, %f846; sub.ftz.f32 %f850, %f7, %f847; mul.ftz.f32 %f851, %f849, %f849; fma.rn.ftz.f32 %f852, %f848, %f848, %f851; fma.rn.ftz.f32 %f853, %f850, %f850, %f852; sqrt.approx.ftz.f32 %f854, %f853; add.ftz.f32 %f855, %f841, %f854; mul.ftz.f32 %f1321, %f855, 0f3F000000; $L__BB103_64: mul.ftz.f32 %f857, %f1, 0f3F000000; mul.ftz.f32 %f858, %f857, %f1321; mul.ftz.f32 %f160, %f858, %f858; sub.ftz.f32 %f859, %f1319, %f1313; sub.ftz.f32 %f860, %f87, %f1318; mul.ftz.f32 %f861, %f860, %f859; sub.ftz.f32 %f862, %f1318, %f1312; sub.ftz.f32 %f863, %f86, %f1319; mul.ftz.f32 %f864, %f863, %f862; sub.ftz.f32 %f865, %f861, %f864; sub.ftz.f32 %f866, %f85, %f1320; mul.ftz.f32 %f867, %f866, %f862; sub.ftz.f32 %f868, %f1320, %f1314; mul.ftz.f32 %f869, %f860, %f868; sub.ftz.f32 %f870, %f867, %f869; mul.ftz.f32 %f871, %f863, %f868; mul.ftz.f32 %f872, %f866, %f859; sub.ftz.f32 %f873, %f871, %f872; mul.ftz.f32 %f874, %f870, %f870; fma.rn.ftz.f32 %f875, %f865, %f865, %f874; fma.rn.ftz.f32 %f876, %f873, %f873, %f875; mul.ftz.f32 %f877, %f859, %f859; fma.rn.ftz.f32 %f878, %f868, %f868, %f877; fma.rn.ftz.f32 %f879, %f862, %f862, %f878; mul.ftz.f32 %f880, %f160, %f879; setp.lt.ftz.f32 %p49, %f876, %f880; @%p49 bra $L__BB103_68; sub.ftz.f32 %f882, %f85, %f1317; sub.ftz.f32 %f883, %f1316, %f1319; sub.ftz.f32 %f884, %f87, %f1315; mul.ftz.f32 %f885, %f884, %f883; sub.ftz.f32 %f886, %f1315, %f1318; sub.ftz.f32 %f887, %f86, %f1316; mul.ftz.f32 %f888, %f887, %f886; sub.ftz.f32 %f889, %f885, %f888; mul.ftz.f32 %f890, %f882, %f886; sub.ftz.f32 %f891, %f1317, %f1320; mul.ftz.f32 %f892, %f884, %f891; sub.ftz.f32 %f893, %f890, %f892; mul.ftz.f32 %f894, %f887, %f891; mul.ftz.f32 %f895, %f882, %f883; sub.ftz.f32 %f896, %f894, %f895; mul.ftz.f32 %f897, %f893, %f893; fma.rn.ftz.f32 %f898, %f889, %f889, %f897; fma.rn.ftz.f32 %f899, %f896, %f896, %f898; mul.ftz.f32 %f900, %f883, %f883; fma.rn.ftz.f32 %f901, %f891, %f891, %f900; fma.rn.ftz.f32 %f902, %f886, %f886, %f901; mul.ftz.f32 %f903, %f160, %f902; setp.lt.ftz.f32 %p50, %f899, %f903; @%p50 bra $L__BB103_68; sub.ftz.f32 %f905, %f85, %f1314; sub.ftz.f32 %f906, %f1313, %f1316; sub.ftz.f32 %f907, %f87, %f1312; mul.ftz.f32 %f908, %f907, %f906; sub.ftz.f32 %f909, %f1312, %f1315; sub.ftz.f32 %f910, %f86, %f1313; mul.ftz.f32 %f911, %f910, %f909; sub.ftz.f32 %f912, %f908, %f911; mul.ftz.f32 %f913, %f905, %f909; sub.ftz.f32 %f914, %f1314, %f1317; mul.ftz.f32 %f915, %f907, %f914; sub.ftz.f32 %f916, %f913, %f915; mul.ftz.f32 %f917, %f910, %f914; mul.ftz.f32 %f918, %f905, %f906; sub.ftz.f32 %f919, %f917, %f918; mul.ftz.f32 %f920, %f916, %f916; fma.rn.ftz.f32 %f921, %f912, %f912, %f920; fma.rn.ftz.f32 %f922, %f919, %f919, %f921; mul.ftz.f32 %f923, %f906, %f906; fma.rn.ftz.f32 %f924, %f914, %f914, %f923; fma.rn.ftz.f32 %f925, %f909, %f909, %f924; mul.ftz.f32 %f926, %f160, %f925; setp.lt.ftz.f32 %p51, %f922, %f926; @%p51 bra $L__BB103_68; mov.f32 %f1322, %f598; $L__BB103_68: mul.ftz.f32 %f928, %f6, %f6; fma.rn.ftz.f32 %f929, %f5, %f5, %f928; fma.rn.ftz.f32 %f930, %f7, %f7, %f929; sqrt.approx.ftz.f32 %f931, %f930; sub.ftz.f32 %f932, %f1296, %f1322; div.approx.ftz.f32 %f933, %f932, %f931; add.ftz.f32 %f1296, %f1296, %f933; $L__BB103_103: setp.eq.s32 %p75, %r69, 255; @%p75 bra $L__BB103_105; mul.wide.u32 %rd354, %r69, 4; add.s64 %rd355, %rd1, %rd354; st.local.f32 [%rd355], %f1296; $L__BB103_105: ret; } .func (.param .b32 func_retval0) _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj( .param .b64 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_0, .param .align 4 .b8 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_1[12], .param .b64 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_2, .param .b64 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_3, .param .b32 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_4, .param .b32 _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_5 ) { .reg .pred %p<531>; .reg .b16 %rs<37>; .reg .f32 %f<5306>; .reg .b32 %r<159>; .reg .b64 %rd<90>; ld.param.u64 %rd8, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_0]; ld.param.u64 %rd9, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_2]; ld.param.u64 %rd10, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_3]; ld.param.f32 %f5301, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_4]; ld.param.f32 %f5302, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_5]; ld.param.f32 %f1406, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_1+8]; ld.param.f32 %f1405, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_1+4]; ld.param.f32 %f1404, [_Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj_param_1]; cvta.to.local.u64 %rd11, %rd8; add.s64 %rd1, %rd11, 260; ld.local.u32 %r157, [%rd11+260]; setp.lt.s32 %p7, %r157, 1; @%p7 bra $L__BB104_541; ld.const.u64 %rd12, [kernel_params+2160]; cvta.to.global.u64 %rd2, %rd12; ld.const.v4.f32 {%f1409, %f1410, %f1411, %f1412}, [kernel_params+1184]; mov.u32 %r148, 0; cvta.to.local.u64 %rd17, %rd9; mov.u16 %rs23, 0; cvta.to.local.u64 %rd7, %rd10; mov.u16 %rs22, 1; mov.pred %p368, -1; mov.pred %p379, 0; bra.uni $L__BB104_2; $L__BB104_397: neg.ftz.f32 %f3650, %f940; div.approx.ftz.f32 %f3651, %f3650, %f3443; neg.ftz.f32 %f3652, %f941; div.approx.ftz.f32 %f3653, %f3652, %f3444; fma.rn.ftz.f32 %f3654, %f3651, %f3651, %f943; fma.rn.ftz.f32 %f3655, %f3653, %f3653, %f3654; mul.ftz.f32 %f3656, %f3655, 0f40490FDB; mul.ftz.f32 %f3657, %f3655, %f3656; mul.ftz.f32 %f3658, %f3443, %f3657; mul.ftz.f32 %f3659, %f3444, %f3658; mov.f32 %f3660, 0f33D6BF95; max.ftz.f32 %f3661, %f3659, %f3660; rcp.approx.ftz.f32 %f5186, %f3661; $L__BB104_399: add.ftz.f32 %f3671, %f5235, 0f3F800000; mov.f32 %f5211, 0f3F800000; sub.ftz.f32 %f3672, %f5211, %f3671; add.ftz.f32 %f3673, %f3672, %f5185; rcp.approx.ftz.f32 %f3674, %f3673; mul.ftz.f32 %f3675, %f3674, 0f3E800000; div.approx.ftz.f32 %f3676, %f3675, %f5183; mul.ftz.f32 %f5290, %f3676, %f5186; fma.rn.ftz.f32 %f948, %f924, %f924, 0fBF800000; abs.ftz.f32 %f949, %f5211; fma.rn.ftz.f32 %f950, %f949, %f949, %f948; setp.leu.ftz.f32 %p364, %f950, 0f00000000; mov.f32 %f5187, %f5211; @%p364 bra $L__BB104_401; sqrt.approx.ftz.f32 %f3677, %f950; sub.ftz.f32 %f3678, %f3677, %f949; add.ftz.f32 %f3679, %f949, %f3677; div.approx.ftz.f32 %f3680, %f3678, %f3679; fma.rn.ftz.f32 %f3681, %f949, %f3679, 0fBF800000; fma.rn.ftz.f32 %f3682, %f949, %f3678, 0f3F800000; div.approx.ftz.f32 %f3683, %f3681, %f3682; mul.ftz.f32 %f3684, %f3680, 0f3F000000; mul.ftz.f32 %f3685, %f3680, %f3684; fma.rn.ftz.f32 %f3686, %f3683, %f3683, 0f3F800000; mul.ftz.f32 %f5187, %f3685, %f3686; $L__BB104_401: sub.ftz.f32 %f5188, %f5211, %f5187; rcp.approx.ftz.f32 %f954, %f5188; setp.ne.s32 %p365, %r5, 15; mov.f32 %f5232, %f5211; mov.f32 %f5233, %f5211; mov.f32 %f5234, %f5211; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p365 bra $L__BB104_405; mul.ftz.f32 %f3690, %f5182, %f941; fma.rn.ftz.f32 %f3691, %f5181, %f940, %f3690; fma.rn.ftz.f32 %f3692, %f5183, %f942, %f3691; abs.ftz.f32 %f955, %f3692; fma.rn.ftz.f32 %f956, %f955, %f955, %f948; setp.leu.ftz.f32 %p366, %f956, 0f00000000; @%p366 bra $L__BB104_404; sqrt.approx.ftz.f32 %f3693, %f956; sub.ftz.f32 %f3694, %f3693, %f955; add.ftz.f32 %f3695, %f955, %f3693; div.approx.ftz.f32 %f3696, %f3694, %f3695; fma.rn.ftz.f32 %f3697, %f955, %f3695, 0fBF800000; fma.rn.ftz.f32 %f3698, %f955, %f3694, 0f3F800000; div.approx.ftz.f32 %f3699, %f3697, %f3698; mul.ftz.f32 %f3700, %f3696, 0f3F000000; mul.ftz.f32 %f3701, %f3696, %f3700; fma.rn.ftz.f32 %f3702, %f3699, %f3699, 0f3F800000; mul.ftz.f32 %f3703, %f3701, %f3702; sub.ftz.f32 %f5188, %f3703, %f5187; $L__BB104_404: mul.ftz.f32 %f3704, %f954, %f5188; mov.f32 %f3705, 0f3F800000; sub.ftz.f32 %f3706, %f3705, %f3704; fma.rn.ftz.f32 %f5232, %f925, %f3706, %f3704; fma.rn.ftz.f32 %f5233, %f926, %f3706, %f3704; fma.rn.ftz.f32 %f5234, %f927, %f3706, %f3704; mul.ftz.f32 %f962, %f5290, %f5232; mul.ftz.f32 %f5289, %f5290, %f5233; mul.ftz.f32 %f5288, %f5290, %f5234; mov.f32 %f5290, %f962; $L__BB104_405: ld.local.u32 %r155, [%rd1+-4]; mov.f32 %f5209, 0f00000000; neg.ftz.f32 %f5231, %f5183; mov.u32 %r154, 0; mov.f32 %f5210, %f5211; $L__BB104_406: mov.u32 %r23, %r154; mad.lo.s32 %r155, %r155, 1103515245, 12345; cvt.rn.f32.u32 %f3710, %r155; mul.ftz.f32 %f987, %f3710, 0f2F800000; setp.gt.ftz.f32 %p369, %f5231, 0f3F7FF972; mov.pred %p530, %p368; @%p369 bra $L__BB104_416; bra.uni $L__BB104_407; $L__BB104_416: @%p530 bra $L__BB104_449; mad.lo.s32 %r121, %r155, 1103515245, 12345; cvt.rn.f32.u32 %f3730, %r121; mul.ftz.f32 %f998, %f3730, 0f2F800000; mad.lo.s32 %r155, %r121, 1103515245, 12345; cvt.rn.f32.u32 %f3731, %r155; mul.ftz.f32 %f999, %f3731, 0f2F800000; neg.ftz.f32 %f1000, %f5229; mul.ftz.f32 %f3732, %f3443, %f1000; neg.ftz.f32 %f1001, %f5230; mul.ftz.f32 %f3733, %f3444, %f1001; mul.ftz.f32 %f3734, %f3733, %f3733; fma.rn.ftz.f32 %f3735, %f3732, %f3732, %f3734; fma.rn.ftz.f32 %f3736, %f5231, %f5231, %f3735; sqrt.approx.ftz.f32 %f3737, %f3736; div.approx.ftz.f32 %f5214, %f3732, %f3737; div.approx.ftz.f32 %f5215, %f3733, %f3737; neg.ftz.f32 %f1004, %f5231; div.approx.ftz.f32 %f1005, %f1004, %f3737; setp.gt.ftz.f32 %p382, %f1005, 0f3F7FF972; @%p382 bra $L__BB104_423; neg.ftz.f32 %f4966, %f5231; div.approx.ftz.f32 %f4965, %f4966, %f3737; abs.ftz.f32 %f3738, %f4965; setp.lt.ftz.f32 %p383, %f3738, 0f358637BD; @%p383 bra $L__BB104_423; bra.uni $L__BB104_419; $L__BB104_423: mov.f32 %f3782, 0f3F800000; sub.ftz.f32 %f3783, %f3782, %f999; mov.f32 %f3784, 0f33D6BF95; max.ftz.f32 %f3785, %f3783, %f3784; div.approx.ftz.f32 %f3786, %f999, %f3785; sqrt.approx.ftz.f32 %f3787, %f3786; mul.ftz.f32 %f3788, %f998, 0f40C90FDB; cos.approx.ftz.f32 %f3789, %f3788; sin.approx.ftz.f32 %f3790, %f3788; mul.ftz.f32 %f5213, %f3787, %f3790; mul.ftz.f32 %f5212, %f3787, %f3789; $L__BB104_424: mul.ftz.f32 %f3791, %f5215, %f5215; fma.rn.ftz.f32 %f3792, %f5214, %f5214, %f3791; add.ftz.f32 %f3793, %f3792, 0f00000000; sqrt.approx.ftz.f32 %f1021, %f3793; setp.neu.ftz.f32 %p391, %f1021, 0f00000000; @%p391 bra $L__BB104_425; bra.uni $L__BB104_426; $L__BB104_425: rcp.approx.ftz.f32 %f3794, %f1021; mul.ftz.f32 %f5215, %f5215, %f3794; mul.ftz.f32 %f5214, %f5214, %f3794; $L__BB104_426: mul.ftz.f32 %f3795, %f5212, %f5214; mul.ftz.f32 %f3796, %f5213, %f5215; sub.ftz.f32 %f3797, %f3795, %f3796; mul.ftz.f32 %f3798, %f3443, %f3797; mul.ftz.f32 %f3799, %f5213, %f5214; fma.rn.ftz.f32 %f3800, %f5212, %f5215, %f3799; mul.ftz.f32 %f3801, %f3444, %f3800; neg.ftz.f32 %f3802, %f3798; neg.ftz.f32 %f3803, %f3801; mul.ftz.f32 %f3804, %f3801, %f3801; fma.rn.ftz.f32 %f3805, %f3798, %f3798, %f3804; add.ftz.f32 %f3806, %f3805, 0f3F800000; mov.f32 %f3807, 0f3F800000; sqrt.approx.ftz.f32 %f3808, %f3806; div.approx.ftz.f32 %f1031, %f3802, %f3808; div.approx.ftz.f32 %f1032, %f3803, %f3808; div.approx.ftz.f32 %f1033, %f3807, %f3808; setp.eq.s32 %p392, %r23, 0; @%p392 bra $L__BB104_437; setp.gt.ftz.f32 %p523, %f5231, 0f3F7FF972; mov.f32 %f5217, 0f00000000; @%p523 bra $L__BB104_434; sub.ftz.f32 %f3815, %f5178, %f5229; sub.ftz.f32 %f3816, %f5179, %f5230; mul.ftz.f32 %f3817, %f3816, %f3816; fma.rn.ftz.f32 %f3818, %f3815, %f3815, %f3817; sub.ftz.f32 %f3819, %f5180, %f5231; fma.rn.ftz.f32 %f3820, %f3819, %f3819, %f3818; sqrt.approx.ftz.f32 %f3821, %f3820; div.approx.ftz.f32 %f1036, %f3819, %f3821; setp.lt.ftz.f32 %p394, %f1036, 0f00000000; @%p394 bra $L__BB104_434; sub.ftz.f32 %f4991, %f5180, %f5231; div.approx.ftz.f32 %f4990, %f4991, %f3821; sub.ftz.f32 %f4989, %f5178, %f5229; div.approx.ftz.f32 %f4988, %f4989, %f3821; sub.ftz.f32 %f4987, %f5179, %f5230; div.approx.ftz.f32 %f4986, %f4987, %f3821; mov.f32 %f5217, 0f00000000; neg.ftz.f32 %f4975, %f5231; neg.ftz.f32 %f4974, %f5229; neg.ftz.f32 %f4973, %f5230; mul.ftz.f32 %f3825, %f4986, %f4973; fma.rn.ftz.f32 %f3826, %f4988, %f4974, %f3825; fma.rn.ftz.f32 %f1037, %f4990, %f4975, %f3826; setp.lt.ftz.f32 %p395, %f1037, 0f00000000; @%p395 bra $L__BB104_434; sub.ftz.f32 %f4993, %f5180, %f5231; div.approx.ftz.f32 %f4992, %f4993, %f3821; setp.eq.ftz.f32 %p524, %f3443, %f3444; mul.ftz.f32 %f1038, %f4992, %f4992; @%p524 bra $L__BB104_432; bra.uni $L__BB104_431; $L__BB104_432: mov.f32 %f4983, 0f3F800000; mul.ftz.f32 %f4976, %f3443, %f3443; sub.ftz.f32 %f3840, %f4983, %f1038; fma.rn.ftz.f32 %f3841, %f4976, %f1038, %f3840; mul.ftz.f32 %f3842, %f3841, 0f40490FDB; mul.ftz.f32 %f3843, %f3841, %f3842; mov.f32 %f3844, 0f33D6BF95; max.ftz.f32 %f3845, %f3843, %f3844; div.approx.ftz.f32 %f5216, %f4976, %f3845; bra.uni $L__BB104_433; $L__BB104_407: setp.lt.ftz.f32 %p370, %f5231, 0fBF7FF972; @%p370 bra $L__BB104_414; bra.uni $L__BB104_408; $L__BB104_414: mul.ftz.f32 %f5210, %f5210, %f987; cvt.ftz.sat.f32.f32 %f3727, %f5210; fma.rn.ftz.f32 %f5211, %f3727, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p380, %f5231, 0f3727C5AC; mov.f32 %f5209, 0f00000000; mov.pred %p530, %p379; @%p380 bra $L__BB104_416; lg2.approx.ftz.f32 %f3728, %f5210; mul.ftz.f32 %f3729, %f5235, %f3728; ex2.approx.ftz.f32 %f5209, %f3729; mov.pred %p530, %p379; bra.uni $L__BB104_416; $L__BB104_408: abs.ftz.f32 %f3711, %f5231; setp.ltu.ftz.f32 %p372, %f3711, 0f38D1B717; mov.pred %p530, %p379; @%p372 bra $L__BB104_416; mov.f32 %f5208, 0f3F800000; sub.ftz.f32 %f3713, %f5208, %f5209; setp.lt.ftz.f32 %p374, %f3713, %f987; mov.pred %p530, %p368; @%p374 bra $L__BB104_416; setp.ge.ftz.f32 %p375, %f5235, 0f00000000; @%p375 bra $L__BB104_412; mov.f32 %f3715, 0f3F800000; sub.ftz.f32 %f3716, %f3715, %f987; lg2.approx.ftz.f32 %f3717, %f3716; mov.f32 %f3718, 0fBF800000; div.approx.ftz.f32 %f3719, %f3718, %f5235; mul.ftz.f32 %f3720, %f3719, %f3717; ex2.approx.ftz.f32 %f3721, %f3720; mul.ftz.f32 %f5208, %f5210, %f3721; $L__BB104_412: cvt.ftz.sat.f32.f32 %f3723, %f5208; fma.rn.ftz.f32 %f5211, %f3723, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p377, %f5231, 0f3727C5AC; mov.f32 %f5209, 0f00000000; mov.f32 %f5210, %f5208; mov.pred %p530, %p379; @%p377 bra $L__BB104_416; lg2.approx.ftz.f32 %f3724, %f5208; mul.ftz.f32 %f3725, %f5235, %f3724; ex2.approx.ftz.f32 %f5209, %f3725; mov.f32 %f5210, %f5208; mov.pred %p530, %p379; bra.uni $L__BB104_416; $L__BB104_419: neg.ftz.f32 %f4968, %f5231; div.approx.ftz.f32 %f4967, %f4968, %f3737; add.ftz.f32 %f3741, %f4967, 0f3F800000; mul.ftz.f32 %f1006, %f3741, 0f3F000000; setp.lt.ftz.f32 %p384, %f1006, 0f38D1B717; mov.f32 %f5212, 0f00000000; mov.f32 %f5213, %f5212; @%p384 bra $L__BB104_424; neg.ftz.f32 %f4970, %f5231; div.approx.ftz.f32 %f4969, %f4970, %f3737; add.ftz.f32 %f3744, %f999, %f999; mul.ftz.f32 %f3745, %f3744, %f1006; div.approx.ftz.f32 %f3746, %f3745, %f4969; add.ftz.f32 %f1007, %f3746, 0fBF800000; mul.ftz.f32 %f1008, %f1007, %f1007; add.ftz.f32 %f1009, %f1008, 0fBF800000; abs.ftz.f32 %f3747, %f1009; setp.lt.ftz.f32 %p385, %f3747, 0f33D6BF95; mov.f32 %f5213, %f5212; @%p385 bra $L__BB104_424; neg.ftz.f32 %f4972, %f5231; div.approx.ftz.f32 %f4971, %f4972, %f3737; mul.ftz.f32 %f3748, %f4971, %f4971; mov.f32 %f3749, 0f3F800000; sub.ftz.f32 %f3750, %f3749, %f3748; mov.f32 %f3751, 0f00000000; max.ftz.f32 %f3752, %f3750, %f3751; sqrt.approx.ftz.f32 %f3753, %f3752; div.approx.ftz.f32 %f3754, %f3753, %f4971; rcp.approx.ftz.f32 %f3755, %f1009; mul.ftz.f32 %f3756, %f3754, %f3754; mul.ftz.f32 %f3757, %f3756, %f3755; mul.ftz.f32 %f3758, %f3755, %f3757; sub.ftz.f32 %f3759, %f1008, %f3756; mul.ftz.f32 %f3760, %f3759, %f3755; sub.ftz.f32 %f3761, %f3758, %f3760; max.ftz.f32 %f3762, %f3761, %f3751; sqrt.approx.ftz.f32 %f3763, %f3762; mul.ftz.f32 %f3764, %f3754, %f3755; add.ftz.f32 %f3765, %f3764, %f3763; rcp.approx.ftz.f32 %f3766, %f3754; setp.gt.ftz.f32 %p386, %f3765, %f3766; setp.lt.ftz.f32 %p387, %f1007, 0f00000000; or.pred %p388, %p387, %p386; sub.ftz.f32 %f3767, %f3764, %f3763; mov.f32 %f3768, 0f3F000000; sub.ftz.f32 %f3769, %f3768, %f998; add.ftz.f32 %f3770, %f998, 0fBF000000; setp.ge.ftz.f32 %p389, %f998, 0f3F000000; selp.f32 %f3771, %f3770, %f3769, %p389; add.ftz.f32 %f3772, %f3771, %f3771; fma.rn.ftz.f32 %f3773, %f3772, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f3774, %f3772, %f3773, 0f3EED4413; mul.ftz.f32 %f3775, %f3772, %f3774; fma.rn.ftz.f32 %f3776, %f3772, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f3777, %f3772, %f3776, 0fBF800000; fma.rn.ftz.f32 %f3778, %f3772, %f3777, 0f3F191676; div.approx.ftz.f32 %f3779, %f3775, %f3778; selp.f32 %f5212, %f3767, %f3765, %p388; fma.rn.ftz.f32 %f3780, %f5212, %f5212, 0f3F800000; sqrt.approx.ftz.f32 %f3781, %f3780; mul.ftz.f32 %f5213, %f3779, %f3781; setp.ltu.ftz.f32 %p390, %f998, 0f3F000000; @%p390 bra $L__BB104_422; bra.uni $L__BB104_424; $L__BB104_422: neg.ftz.f32 %f5213, %f5213; bra.uni $L__BB104_424; $L__BB104_431: sub.ftz.f32 %f4997, %f5178, %f5229; div.approx.ftz.f32 %f4996, %f4997, %f3821; sub.ftz.f32 %f4995, %f5179, %f5230; div.approx.ftz.f32 %f4994, %f4995, %f3821; neg.ftz.f32 %f3827, %f4996; div.approx.ftz.f32 %f3828, %f3827, %f3443; neg.ftz.f32 %f3829, %f4994; div.approx.ftz.f32 %f3830, %f3829, %f3444; fma.rn.ftz.f32 %f3831, %f3828, %f3828, %f1038; fma.rn.ftz.f32 %f3832, %f3830, %f3830, %f3831; mul.ftz.f32 %f3833, %f3832, 0f40490FDB; mul.ftz.f32 %f3834, %f3832, %f3833; mul.ftz.f32 %f3835, %f3443, %f3834; mul.ftz.f32 %f3836, %f3444, %f3835; mov.f32 %f3837, 0f33D6BF95; max.ftz.f32 %f3838, %f3836, %f3837; rcp.approx.ftz.f32 %f5216, %f3838; $L__BB104_433: setp.lt.ftz.f32 %p397, %f5231, 0fBF7FF972; mul.ftz.f32 %f3846, %f5231, %f5235; selp.f32 %f3847, 0f3F800000, %f3846, %p397; mul.ftz.f32 %f3848, %f3847, %f1037; mov.f32 %f3849, 0f33D6BF95; max.ftz.f32 %f3850, %f3848, %f3849; mov.f32 %f3851, 0f00000000; max.ftz.f32 %f3852, %f3851, %f1037; mul.ftz.f32 %f3853, %f3852, 0f3E800000; div.approx.ftz.f32 %f3854, %f3853, %f3850; mul.ftz.f32 %f5217, %f3854, %f5216; $L__BB104_434: setp.gt.ftz.f32 %p525, %f5180, 0f3F7FF972; selp.f32 %f5218, 0f3F800000, 0f00000000, %p525; @%p525 bra $L__BB104_436; add.ftz.f32 %f3855, %f5211, 0f3F800000; mul.ftz.f32 %f3856, %f3855, 0f3F000000; cvt.ftz.sat.f32.f32 %f3857, %f3856; lg2.approx.ftz.f32 %f3858, %f3857; mul.ftz.f32 %f3859, %f5185, %f3858; ex2.approx.ftz.f32 %f5218, %f3859; $L__BB104_436: mul.ftz.f32 %f3860, %f5232, %f5217; mul.ftz.f32 %f3861, %f5232, %f3860; fma.rn.ftz.f32 %f5290, %f3861, %f5218, %f5290; mul.ftz.f32 %f3862, %f5233, %f5217; mul.ftz.f32 %f3863, %f5233, %f3862; fma.rn.ftz.f32 %f5289, %f3863, %f5218, %f5289; mul.ftz.f32 %f3864, %f5234, %f5217; mul.ftz.f32 %f3865, %f5234, %f3864; fma.rn.ftz.f32 %f5288, %f3865, %f5218, %f5288; $L__BB104_437: add.s32 %r154, %r23, 1; setp.gt.u32 %p399, %r154, 9; @%p399 bra $L__BB104_448; setp.eq.s32 %p527, %r23, 0; setp.ne.s32 %p526, %r5, 15; neg.ftz.f32 %f4980, %f5231; neg.ftz.f32 %f4979, %f5229; neg.ftz.f32 %f4978, %f5230; mul.ftz.f32 %f3866, %f1032, %f4978; fma.rn.ftz.f32 %f3867, %f1031, %f4979, %f3866; fma.rn.ftz.f32 %f1056, %f1033, %f4980, %f3867; or.pred %p402, %p526, %p527; @%p402 bra $L__BB104_442; mov.f32 %f4982, 0f3F800000; sub.ftz.f32 %f5222, %f4982, %f5187; abs.ftz.f32 %f1057, %f1056; fma.rn.ftz.f32 %f1058, %f1057, %f1057, %f948; setp.leu.ftz.f32 %p403, %f1058, 0f00000000; @%p403 bra $L__BB104_441; sqrt.approx.ftz.f32 %f3868, %f1058; sub.ftz.f32 %f3869, %f3868, %f1057; add.ftz.f32 %f3870, %f1057, %f3868; div.approx.ftz.f32 %f3871, %f3869, %f3870; fma.rn.ftz.f32 %f3872, %f1057, %f3870, 0fBF800000; fma.rn.ftz.f32 %f3873, %f1057, %f3869, 0f3F800000; div.approx.ftz.f32 %f3874, %f3872, %f3873; mul.ftz.f32 %f3875, %f3871, 0f3F000000; mul.ftz.f32 %f3876, %f3871, %f3875; fma.rn.ftz.f32 %f3877, %f3874, %f3874, 0f3F800000; mul.ftz.f32 %f3878, %f3876, %f3877; sub.ftz.f32 %f5222, %f3878, %f5187; $L__BB104_441: mul.ftz.f32 %f3879, %f954, %f5222; mov.f32 %f3880, 0f3F800000; sub.ftz.f32 %f3881, %f3880, %f3879; fma.rn.ftz.f32 %f3882, %f925, %f3881, %f3879; fma.rn.ftz.f32 %f3883, %f926, %f3881, %f3879; fma.rn.ftz.f32 %f3884, %f927, %f3881, %f3879; mul.ftz.f32 %f5232, %f5232, %f3882; mul.ftz.f32 %f5233, %f5233, %f3883; mul.ftz.f32 %f5234, %f5234, %f3884; $L__BB104_442: mov.f32 %f4999, 0f3F800000; div.approx.ftz.f32 %f4998, %f4999, %f3808; add.ftz.f32 %f3886, %f1031, %f1031; add.ftz.f32 %f3887, %f1032, %f1032; add.ftz.f32 %f3888, %f4998, %f4998; fma.rn.ftz.f32 %f5229, %f3886, %f1056, %f5229; fma.rn.ftz.f32 %f5230, %f3887, %f1056, %f5230; fma.rn.ftz.f32 %f5231, %f3888, %f1056, %f5231; setp.gt.ftz.f32 %p404, %f5231, 0f3F7FF972; mov.f32 %f5235, 0f00000000; @%p404 bra $L__BB104_445; setp.lt.ftz.f32 %p405, %f5231, 0fBF7FF972; mov.f32 %f5235, 0fBF7FF972; @%p405 bra $L__BB104_445; mul.ftz.f32 %f3890, %f5231, %f5231; mov.f32 %f3891, 0f33D6BF95; max.ftz.f32 %f3892, %f3890, %f3891; rcp.approx.ftz.f32 %f3893, %f3892; mul.ftz.f32 %f3894, %f3443, %f5229; mul.ftz.f32 %f3895, %f3444, %f5230; mul.ftz.f32 %f3896, %f3895, %f3895; fma.rn.ftz.f32 %f3897, %f3894, %f3894, %f3896; fma.rn.ftz.f32 %f3898, %f3897, %f3893, 0f3F800000; sqrt.approx.ftz.f32 %f3899, %f3898; neg.ftz.f32 %f3900, %f3899; setp.le.ftz.f32 %p406, %f5231, 0f00000000; selp.f32 %f3901, %f3900, %f3899, %p406; add.ftz.f32 %f3902, %f3901, 0fBF800000; mul.ftz.f32 %f5235, %f3902, 0f3F000000; $L__BB104_445: mov.f32 %f5209, 0f3F800000; mul.ftz.f32 %f3904, %f921, %f5232; setp.eq.s32 %p407, %r5, 15; selp.f32 %f5232, %f5232, %f3904, %p407; mul.ftz.f32 %f3905, %f922, %f5233; selp.f32 %f5233, %f5233, %f3905, %p407; mul.ftz.f32 %f3906, %f923, %f5234; selp.f32 %f5234, %f5234, %f3906, %p407; add.ftz.f32 %f3907, %f5211, 0f3F800000; mul.ftz.f32 %f3908, %f3907, 0f3F000000; cvt.ftz.sat.f32.f32 %f5210, %f3908; @%p404 bra $L__BB104_448; setp.lt.ftz.f32 %p409, %f5231, 0f3727C5AC; mov.f32 %f5209, 0f00000000; @%p409 bra $L__BB104_448; lg2.approx.ftz.f32 %f3910, %f5210; mul.ftz.f32 %f3911, %f5235, %f3910; ex2.approx.ftz.f32 %f5209, %f3911; $L__BB104_448: setp.lt.u32 %p410, %r154, 10; @%p410 bra $L__BB104_406; $L__BB104_449: st.local.u32 [%rd1+-4], %r155; setp.eq.s16 %p411, %rs35, 0; @%p411 bra $L__BB104_451; div.approx.ftz.f32 %f3912, %f5183, %f5180; abs.ftz.f32 %f3913, %f3912; mul.ftz.f32 %f5290, %f5290, %f3913; mul.ftz.f32 %f5289, %f5289, %f3913; mul.ftz.f32 %f5288, %f5288, %f3913; $L__BB104_451: mov.f32 %f5291, %f5177; bra.uni $L__BB104_519; $L__BB104_2: mul.wide.s32 %rd14, %r148, 80; add.s64 %rd15, %rd11, %rd14; add.s64 %rd16, %rd15, 304; setp.eq.s64 %p8, %rd16, %rd17; @%p8 bra $L__BB104_540; add.s64 %rd3, %rd15, 316; ld.local.u32 %r5, [%rd15+316]; setp.gt.s32 %p9, %r5, 34; @%p9 bra $L__BB104_540; setp.gt.s32 %p10, %r5, 31; @%p10 bra $L__BB104_539; mov.f32 %f5288, 0f00000000; setp.gt.s32 %p11, %r5, 13; @%p11 bra $L__BB104_36; bra.uni $L__BB104_6; $L__BB104_36: setp.gt.s32 %p12, %r5, 21; @%p12 bra $L__BB104_56; setp.gt.s32 %p20, %r5, 16; @%p20 bra $L__BB104_47; add.s32 %r38, %r5, -14; setp.lt.u32 %p24, %r38, 2; @%p24 bra $L__BB104_370; bra.uni $L__BB104_39; $L__BB104_370: ld.local.f32 %f876, [%rd1+-224]; ld.local.f32 %f877, [%rd1+-220]; ld.local.f32 %f878, [%rd1+-216]; ld.local.f32 %f3433, [%rd1+-248]; ld.local.f32 %f3434, [%rd1+-244]; mul.ftz.f32 %f3435, %f1405, %f3434; fma.rn.ftz.f32 %f3436, %f1404, %f3433, %f3435; ld.local.f32 %f3437, [%rd1+-240]; fma.rn.ftz.f32 %f3438, %f1406, %f3437, %f3436; setp.lt.ftz.f32 %p339, %f3438, 0f00000000; mov.f32 %f5290, 0f00000000; mov.f32 %f5291, %f5290; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p339 bra $L__BB104_519; ld.local.v2.f32 {%f3443, %f3444}, [%rd3+20]; mul.ftz.f32 %f881, %f3443, %f3444; setp.lt.ftz.f32 %p340, %f881, 0f33D6BF95; mov.f32 %f5291, %f5290; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p340 bra $L__BB104_519; ld.local.f32 %f882, [%rd3+8]; ld.local.f32 %f883, [%rd3+12]; mul.ftz.f32 %f3449, %f877, %f883; fma.rn.ftz.f32 %f3450, %f876, %f882, %f3449; ld.local.f32 %f884, [%rd3+16]; fma.rn.ftz.f32 %f5183, %f878, %f884, %f3450; setp.le.ftz.f32 %p341, %f5183, 0f00000000; mov.f32 %f5291, %f5290; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p341 bra $L__BB104_519; mul.ftz.f32 %f3455, %f1405, %f883; fma.rn.ftz.f32 %f3456, %f1404, %f882, %f3455; fma.rn.ftz.f32 %f886, %f1406, %f884, %f3456; setp.le.ftz.f32 %p342, %f886, 0f00000000; mov.f32 %f5291, %f5290; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p342 bra $L__BB104_519; setp.neu.ftz.f32 %p343, %f3443, %f3444; @%p343 bra $L__BB104_376; bra.uni $L__BB104_375; $L__BB104_376: ld.local.f32 %f3474, [%rd3+52]; mul.ftz.f32 %f3475, %f883, %f3474; ld.local.f32 %f3476, [%rd3+48]; mul.ftz.f32 %f3477, %f884, %f3476; sub.ftz.f32 %f3478, %f3475, %f3477; ld.local.f32 %f3479, [%rd3+44]; mul.ftz.f32 %f3480, %f884, %f3479; mul.ftz.f32 %f3481, %f882, %f3474; sub.ftz.f32 %f3482, %f3480, %f3481; mul.ftz.f32 %f3483, %f882, %f3476; mul.ftz.f32 %f3484, %f883, %f3479; sub.ftz.f32 %f3485, %f3483, %f3484; mul.ftz.f32 %f3486, %f3482, %f3482; fma.rn.ftz.f32 %f3487, %f3478, %f3478, %f3486; fma.rn.ftz.f32 %f3488, %f3485, %f3485, %f3487; sqrt.approx.ftz.f32 %f3489, %f3488; div.approx.ftz.f32 %f5169, %f3478, %f3489; div.approx.ftz.f32 %f5170, %f3482, %f3489; div.approx.ftz.f32 %f5171, %f3485, %f3489; mul.ftz.f32 %f3490, %f884, %f5170; mul.ftz.f32 %f3491, %f883, %f5171; sub.ftz.f32 %f5172, %f3490, %f3491; mul.ftz.f32 %f3492, %f882, %f5171; mul.ftz.f32 %f3493, %f884, %f5169; sub.ftz.f32 %f5173, %f3492, %f3493; mul.ftz.f32 %f3494, %f883, %f5169; mul.ftz.f32 %f3495, %f882, %f5170; sub.ftz.f32 %f5174, %f3494, %f3495; bra.uni $L__BB104_377; $L__BB104_6: setp.gt.s32 %p26, %r5, 7; @%p26 bra $L__BB104_25; setp.gt.s32 %p30, %r5, 4; @%p30 bra $L__BB104_15; bra.uni $L__BB104_8; $L__BB104_15: setp.eq.s32 %p31, %r5, 5; @%p31 bra $L__BB104_509; setp.eq.s32 %p32, %r5, 6; @%p32 bra $L__BB104_505; setp.eq.s32 %p33, %r5, 7; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p33 bra $L__BB104_18; bra.uni $L__BB104_519; $L__BB104_18: ld.local.f32 %f4596, [%rd3+8]; ld.local.f32 %f4597, [%rd3+12]; mul.ftz.f32 %f4598, %f1405, %f4597; fma.rn.ftz.f32 %f4599, %f1404, %f4596, %f4598; ld.local.f32 %f4600, [%rd3+16]; fma.rn.ftz.f32 %f1283, %f1406, %f4600, %f4599; setp.ltu.ftz.f32 %p472, %f1283, 0f00000000; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p472 bra $L__BB104_24; ld.local.f32 %f4602, [%rd3+20]; mul.ftz.f32 %f1284, %f4602, 0f3FC90FDB; mov.f32 %f4603, 0f3FC90FDB; ld.local.f32 %f4604, [%rd3+24]; mul.ftz.f32 %f1285, %f4604, 0f3FC90FDB; mov.f32 %f5291, 0f00000000; max.ftz.f32 %f4606, %f1283, %f5291; mov.f32 %f4607, 0fBF800000; max.ftz.f32 %f4608, %f4606, %f4607; mov.f32 %f5276, 0f3F800000; min.ftz.f32 %f4609, %f4608, %f5276; abs.ftz.f32 %f4610, %f4609; sub.ftz.f32 %f4611, %f5276, %f4610; mul.ftz.f32 %f4612, %f4611, 0f3F000000; sqrt.approx.ftz.f32 %f4613, %f4612; setp.gt.ftz.f32 %p473, %f4610, 0f3F11EB85; selp.f32 %f4614, %f4613, %f4610, %p473; mul.ftz.f32 %f4615, %f4614, %f4614; mov.f32 %f4616, 0f3C94D2E9; mov.f32 %f4617, 0f3D53F941; fma.rn.ftz.f32 %f4618, %f4617, %f4615, %f4616; mov.f32 %f4619, 0f3D3F841F; fma.rn.ftz.f32 %f4620, %f4618, %f4615, %f4619; mov.f32 %f4621, 0f3D994929; fma.rn.ftz.f32 %f4622, %f4620, %f4615, %f4621; mov.f32 %f4623, 0f3E2AAB94; fma.rn.ftz.f32 %f4624, %f4622, %f4615, %f4623; mul.ftz.f32 %f4625, %f4615, %f4624; fma.rn.ftz.f32 %f4626, %f4625, %f4614, %f4614; add.ftz.f32 %f4627, %f4626, %f4626; sub.ftz.f32 %f4628, %f4603, %f4626; selp.f32 %f4629, %f4627, %f4628, %p473; setp.lt.ftz.f32 %p474, %f4609, 0f00000000; mov.f32 %f4630, 0f40490FDB; sub.ftz.f32 %f4631, %f4630, %f4629; selp.f32 %f1286, %f4631, %f4629, %p474; setp.lt.ftz.f32 %p475, %f1286, %f1284; @%p475 bra $L__BB104_22; add.ftz.f32 %f4633, %f1284, %f1285; setp.leu.ftz.f32 %p476, %f4633, %f1286; setp.eq.ftz.f32 %p477, %f1285, 0f00000000; or.pred %p478, %p477, %p476; mov.f32 %f5276, %f5291; @%p478 bra $L__BB104_22; sub.ftz.f32 %f4634, %f1286, %f1284; div.approx.ftz.f32 %f4635, %f4634, %f1285; mov.f32 %f4636, 0f3F800000; sub.ftz.f32 %f5276, %f4636, %f4635; $L__BB104_22: setp.leu.ftz.f32 %p479, %f5276, 0f00000000; mov.f32 %f5288, %f5291; @%p479 bra $L__BB104_24; add.ftz.f32 %f4641, %f1284, %f1285; min.ftz.f32 %f4643, %f4641, %f4603; cos.approx.ftz.f32 %f4644, %f4643; mov.f32 %f4645, 0f3F800000; sub.ftz.f32 %f4646, %f4645, %f4644; mov.f32 %f4647, 0f3E22F983; div.approx.ftz.f32 %f5291, %f4647, %f4646; mul.ftz.f32 %f5288, %f5276, %f5291; $L__BB104_24: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_56: setp.gt.s32 %p13, %r5, 26; @%p13 bra $L__BB104_60; setp.eq.s32 %p17, %r5, 22; @%p17 bra $L__BB104_28; setp.eq.s32 %p18, %r5, 23; @%p18 bra $L__BB104_464; setp.eq.s32 %p19, %r5, 24; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p19 bra $L__BB104_63; bra.uni $L__BB104_519; $L__BB104_25: add.s32 %r39, %r5, -10; setp.lt.u32 %p27, %r39, 3; @%p27 bra $L__BB104_464; bra.uni $L__BB104_26; $L__BB104_464: ld.local.f32 %f1173, [%rd1+-224]; ld.local.f32 %f1174, [%rd1+-220]; ld.local.f32 %f1175, [%rd1+-216]; ld.local.v2.f32 {%f4134, %f4135}, [%rd3+20]; ld.local.f32 %f4136, [%rd1+-248]; ld.local.f32 %f4137, [%rd1+-244]; mul.ftz.f32 %f4138, %f1405, %f4137; fma.rn.ftz.f32 %f4139, %f1404, %f4136, %f4138; ld.local.f32 %f4140, [%rd1+-240]; fma.rn.ftz.f32 %f1178, %f1406, %f4140, %f4139; setp.lt.ftz.f32 %p431, %f1178, 0f00000000; setp.eq.s32 %p432, %r5, 23; xor.pred %p433, %p431, %p432; mul.ftz.f32 %f5255, %f4134, %f4135; setp.le.ftz.f32 %p434, %f5255, 0f33D6BF95; or.pred %p435, %p433, %p434; mov.f32 %f5291, 0f00000000; mov.f32 %f5290, %f5291; mov.f32 %f5289, %f5291; mov.f32 %f5288, %f5291; @%p435 bra $L__BB104_519; ld.local.f32 %f1180, [%rd3+8]; ld.local.f32 %f1181, [%rd3+12]; mul.ftz.f32 %f4141, %f1174, %f1181; fma.rn.ftz.f32 %f4142, %f1173, %f1180, %f4141; ld.local.f32 %f1182, [%rd3+16]; fma.rn.ftz.f32 %f1183, %f1175, %f1182, %f4142; mul.ftz.f32 %f4143, %f1405, %f1181; fma.rn.ftz.f32 %f4144, %f1404, %f1180, %f4143; fma.rn.ftz.f32 %f1184, %f1406, %f1182, %f4144; @%p431 bra $L__BB104_483; bra.uni $L__BB104_466; $L__BB104_483: setp.le.ftz.f32 %p447, %f1183, 0f00000000; setp.ge.ftz.f32 %p448, %f1184, 0f00000000; or.pred %p449, %p447, %p448; mov.f32 %f5290, %f5291; mov.f32 %f5289, %f5291; mov.f32 %f5288, %f5291; @%p449 bra $L__BB104_519; ld.local.f32 %f4328, [%rd3+28]; fma.rn.ftz.f32 %f4329, %f1404, %f4328, %f1173; fma.rn.ftz.f32 %f4330, %f1405, %f4328, %f1174; fma.rn.ftz.f32 %f4331, %f1406, %f4328, %f1175; neg.ftz.f32 %f4332, %f4329; neg.ftz.f32 %f4333, %f4330; neg.ftz.f32 %f4334, %f4331; mul.ftz.f32 %f4335, %f4330, %f4330; fma.rn.ftz.f32 %f4336, %f4329, %f4329, %f4335; fma.rn.ftz.f32 %f4337, %f4331, %f4331, %f4336; sqrt.approx.ftz.f32 %f4338, %f4337; div.approx.ftz.f32 %f4339, %f4332, %f4338; div.approx.ftz.f32 %f4340, %f4333, %f4338; div.approx.ftz.f32 %f4341, %f4334, %f4338; mul.ftz.f32 %f4342, %f1174, %f4340; fma.rn.ftz.f32 %f4343, %f1173, %f4339, %f4342; fma.rn.ftz.f32 %f4344, %f1175, %f4341, %f4343; mul.ftz.f32 %f4345, %f1405, %f4340; fma.rn.ftz.f32 %f4346, %f1404, %f4339, %f4345; fma.rn.ftz.f32 %f4347, %f1406, %f4341, %f4346; mul.ftz.f32 %f4348, %f1181, %f4340; fma.rn.ftz.f32 %f4349, %f1180, %f4339, %f4348; fma.rn.ftz.f32 %f4350, %f1182, %f4341, %f4349; mul.ftz.f32 %f4351, %f4350, %f4350; mov.f32 %f4352, 0f3F800000; sub.ftz.f32 %f4353, %f4352, %f4351; div.approx.ftz.f32 %f4354, %f4353, %f4351; mul.ftz.f32 %f4355, %f4351, %f4351; mul.ftz.f32 %f4356, %f4355, 0f40490FDB; add.ftz.f32 %f4357, %f5255, %f4354; mul.ftz.f32 %f4358, %f4357, %f4356; mul.ftz.f32 %f4359, %f4357, %f4358; div.approx.ftz.f32 %f4360, %f5255, %f4359; mul.ftz.f32 %f4361, %f1183, %f1183; sub.ftz.f32 %f4362, %f4352, %f4361; mul.ftz.f32 %f4363, %f5255, %f4362; div.approx.ftz.f32 %f4364, %f4363, %f4361; add.ftz.f32 %f4365, %f4364, 0f3F800000; mov.f32 %f4366, 0f00000000; max.ftz.f32 %f4367, %f4365, %f4366; sqrt.approx.ftz.f32 %f4368, %f4367; add.ftz.f32 %f4369, %f4368, 0f3F800000; mov.f32 %f4370, 0f40000000; div.approx.ftz.f32 %f4371, %f4370, %f4369; mul.ftz.f32 %f4372, %f1184, %f1184; sub.ftz.f32 %f4373, %f4352, %f4372; mul.ftz.f32 %f4374, %f5255, %f4373; div.approx.ftz.f32 %f4375, %f4374, %f4372; add.ftz.f32 %f4376, %f4375, 0f3F800000; max.ftz.f32 %f4377, %f4376, %f4366; sqrt.approx.ftz.f32 %f4378, %f4377; add.ftz.f32 %f4379, %f4378, 0f3F800000; div.approx.ftz.f32 %f4380, %f4370, %f4379; mul.ftz.f32 %f4381, %f4371, %f4380; mul.ftz.f32 %f4382, %f4328, %f4328; mul.ftz.f32 %f4383, %f4382, %f4360; mul.ftz.f32 %f4384, %f1183, %f4337; div.approx.ftz.f32 %f4385, %f4383, %f4384; mul.ftz.f32 %f4386, %f4347, %f4344; abs.ftz.f32 %f4387, %f4386; mul.ftz.f32 %f4388, %f4381, %f4387; mul.ftz.f32 %f4389, %f4371, %f4387; mul.ftz.f32 %f5291, %f4385, %f4389; mul.ftz.f32 %f5290, %f4385, %f4388; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; bra.uni $L__BB104_519; $L__BB104_47: setp.eq.s32 %p21, %r5, 17; @%p21 bra $L__BB104_363; setp.eq.s32 %p22, %r5, 19; @%p22 bra $L__BB104_357; setp.eq.s32 %p23, %r5, 20; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p23 bra $L__BB104_50; bra.uni $L__BB104_519; $L__BB104_50: ld.local.f32 %f3021, [%rd3+8]; ld.local.f32 %f3022, [%rd3+12]; mul.ftz.f32 %f3023, %f1405, %f3022; fma.rn.ftz.f32 %f3024, %f1404, %f3021, %f3023; ld.local.f32 %f3025, [%rd3+16]; fma.rn.ftz.f32 %f3026, %f1406, %f3025, %f3024; setp.lt.ftz.f32 %p295, %f3026, 0f00000000; mov.f32 %f5288, 0f00000000; @%p295 bra $L__BB104_356; ld.local.f32 %f3031, [%rd1+-224]; ld.local.f32 %f3032, [%rd1+-220]; ld.local.f32 %f3033, [%rd1+-216]; ld.local.f32 %f762, [%rd3+40]; ld.local.f32 %f763, [%rd3+32]; ld.local.f32 %f3034, [%rd3+20]; ld.local.f32 %f3035, [%rd3+24]; mul.ftz.f32 %f3036, %f3032, %f3035; fma.rn.ftz.f32 %f3037, %f3031, %f3034, %f3036; ld.local.f32 %f3038, [%rd3+28]; fma.rn.ftz.f32 %f3039, %f3033, %f3038, %f3037; mul.ftz.f32 %f3040, %f3034, %f3039; mul.ftz.f32 %f3041, %f3035, %f3039; mul.ftz.f32 %f3042, %f3038, %f3039; sub.ftz.f32 %f3043, %f3031, %f3040; sub.ftz.f32 %f3044, %f3032, %f3041; sub.ftz.f32 %f3045, %f3033, %f3042; mul.ftz.f32 %f3046, %f3044, %f3044; fma.rn.ftz.f32 %f3047, %f3043, %f3043, %f3046; fma.rn.ftz.f32 %f3048, %f3045, %f3045, %f3047; sqrt.approx.ftz.f32 %f3049, %f3048; div.approx.ftz.f32 %f3050, %f3043, %f3049; div.approx.ftz.f32 %f3051, %f3044, %f3049; div.approx.ftz.f32 %f3052, %f3045, %f3049; abs.ftz.f32 %f3053, %f3039; setp.lt.ftz.f32 %p296, %f3053, 0f3F800000; mov.f32 %f3054, 0f3F800000; add.ftz.f32 %f3055, %f3053, 0fBF800000; add.ftz.f32 %f3056, %f3055, 0f3F800000; selp.f32 %f3057, %f3056, 0f3F800000, %p296; sub.ftz.f32 %f3058, %f3054, %f3057; sqrt.approx.ftz.f32 %f3059, %f3058; mul.ftz.f32 %f3060, %f3057, 0f3CB1485D; mov.f32 %f3061, 0f3D9FB43B; sub.ftz.f32 %f3062, %f3061, %f3060; fma.rn.ftz.f32 %f3063, %f3057, %f3062, 0fBE5A6B93; fma.rn.ftz.f32 %f3064, %f3057, %f3063, 0f3FC90FDB; mov.f32 %f3065, 0f3FC90FDB; mul.ftz.f32 %f3066, %f3059, %f3064; setp.lt.ftz.f32 %p297, %f3039, 0f00000000; mov.f32 %f3067, 0f40490FDB; sub.ftz.f32 %f3068, %f3067, %f3066; selp.f32 %f3069, %f3068, %f3066, %p297; sub.ftz.f32 %f764, %f3065, %f3069; mul.ftz.f32 %f3070, %f1405, %f3035; fma.rn.ftz.f32 %f3071, %f1404, %f3034, %f3070; fma.rn.ftz.f32 %f3072, %f1406, %f3038, %f3071; mul.ftz.f32 %f3073, %f3034, %f3072; mul.ftz.f32 %f3074, %f3035, %f3072; mul.ftz.f32 %f3075, %f3038, %f3072; sub.ftz.f32 %f3076, %f1404, %f3073; sub.ftz.f32 %f3077, %f1405, %f3074; sub.ftz.f32 %f3078, %f1406, %f3075; mul.ftz.f32 %f3079, %f3077, %f3077; fma.rn.ftz.f32 %f3080, %f3076, %f3076, %f3079; fma.rn.ftz.f32 %f3081, %f3078, %f3078, %f3080; sqrt.approx.ftz.f32 %f3082, %f3081; div.approx.ftz.f32 %f3083, %f3076, %f3082; div.approx.ftz.f32 %f3084, %f3077, %f3082; div.approx.ftz.f32 %f3085, %f3078, %f3082; abs.ftz.f32 %f3086, %f3072; setp.lt.ftz.f32 %p298, %f3086, 0f3F800000; add.ftz.f32 %f3087, %f3086, 0fBF800000; add.ftz.f32 %f3088, %f3087, 0f3F800000; selp.f32 %f3089, %f3088, 0f3F800000, %p298; sub.ftz.f32 %f3090, %f3054, %f3089; sqrt.approx.ftz.f32 %f3091, %f3090; mul.ftz.f32 %f3092, %f3089, 0f3CB1485D; sub.ftz.f32 %f3093, %f3061, %f3092; fma.rn.ftz.f32 %f3094, %f3089, %f3093, 0fBE5A6B93; fma.rn.ftz.f32 %f3095, %f3089, %f3094, 0f3FC90FDB; mul.ftz.f32 %f3096, %f3091, %f3095; setp.lt.ftz.f32 %p299, %f3072, 0f00000000; sub.ftz.f32 %f3097, %f3067, %f3096; selp.f32 %f3098, %f3097, %f3096, %p299; sub.ftz.f32 %f765, %f3065, %f3098; mul.ftz.f32 %f3099, %f3051, %f3084; fma.rn.ftz.f32 %f3100, %f3050, %f3083, %f3099; fma.rn.ftz.f32 %f766, %f3052, %f3085, %f3100; abs.ftz.f32 %f3101, %f765; sub.ftz.f32 %f3102, %f3065, %f3101; setp.lt.ftz.f32 %p300, %f3102, 0f3A83126F; setp.lt.ftz.f32 %p301, %f766, 0f00000000; or.pred %p302, %p301, %p300; @%p302 bra $L__BB104_356; ld.local.f32 %f3104, [%rd3+36]; rcp.approx.ftz.f32 %f767, %f3104; abs.ftz.f32 %f3106, %f766; setp.lt.ftz.f32 %p303, %f3106, 0f3F800000; add.ftz.f32 %f3107, %f3106, 0fBF800000; add.ftz.f32 %f3108, %f3107, 0f3F800000; selp.f32 %f3109, %f3108, 0f3F800000, %p303; sub.ftz.f32 %f3110, %f3054, %f3109; sqrt.approx.ftz.f32 %f3111, %f3110; fma.rn.ftz.f32 %f3112, %f3109, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f3113, %f3109, %f3112, 0fBE5A6B93; fma.rn.ftz.f32 %f3114, %f3109, %f3113, 0f3FC90FDB; mul.ftz.f32 %f768, %f3111, %f3114; mul.ftz.f32 %f3115, %f765, 0f3EA2F983; mov.b32 %r102, %f3115; and.b32 %r103, %r102, -2147483648; or.b32 %r104, %r103, 1056964608; mov.b32 %f3116, %r104; add.ftz.f32 %f3117, %f3115, %f3116; cvt.rzi.ftz.s32.f32 %r19, %f3117; cvt.rn.f32.s32 %f3118, %r19; fma.rn.ftz.f32 %f3119, %f3118, 0fC0490000, %f765; fma.rn.ftz.f32 %f3120, %f3118, 0fBA7DA000, %f3119; fma.rn.ftz.f32 %f3121, %f3118, 0fB4222000, %f3120; fma.rn.ftz.f32 %f3122, %f3118, 0fACB4611A, %f3121; add.ftz.f32 %f3123, %f3122, 0fBFC90FDB; add.ftz.f32 %f3124, %f3123, 0f3FC90FDB; mul.ftz.f32 %f3125, %f3124, %f3124; fma.rn.ftz.f32 %f3126, %f3125, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f3127, %f3125, %f3126, 0fBAB60B58; fma.rn.ftz.f32 %f3128, %f3125, %f3127, 0f3D2AAAAA; fma.rn.ftz.f32 %f3129, %f3125, %f3128, 0fBF000000; fma.rn.ftz.f32 %f769, %f3125, %f3129, 0f3F800000; add.ftz.f32 %f3130, %f764, 0f3FC90FDB; mul.ftz.f32 %f3131, %f3130, 0f3F000000; sub.ftz.f32 %f3132, %f3131, %f762; rcp.approx.ftz.f32 %f770, %f763; mul.ftz.f32 %f771, %f770, %f3132; abs.ftz.f32 %f772, %f771; setp.eq.ftz.f32 %p304, %f772, 0f00000000; mov.f32 %f5157, 0f00000000; mov.f32 %f5156, %f5157; @%p304 bra $L__BB104_350; setp.eq.ftz.f32 %p305, %f772, 0f3F800000; mov.f32 %f5156, %f3054; @%p305 bra $L__BB104_350; setp.gt.ftz.f32 %p306, %f772, 0f3F800000; @%p306 bra $L__BB104_349; bra.uni $L__BB104_55; $L__BB104_349: mov.f32 %f3135, 0f3F800000; div.approx.ftz.f32 %f5156, %f3135, %f772; bra.uni $L__BB104_350; $L__BB104_60: setp.eq.s32 %p14, %r5, 30; @%p14 bra $L__BB104_485; setp.eq.s32 %p15, %r5, 29; @%p15 bra $L__BB104_204; setp.ne.s32 %p16, %r5, 27; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p16 bra $L__BB104_519; $L__BB104_63: ld.local.v2.f32 {%f1421, %f1422}, [%rd3+20]; mul.ftz.f32 %f1423, %f1421, %f1422; setp.lt.ftz.f32 %p36, %f1423, 0f33D6BF95; mov.f32 %f5291, 0f00000000; mov.f32 %f5290, %f5291; mov.f32 %f5289, %f5291; mov.f32 %f5288, %f5291; @%p36 bra $L__BB104_519; ld.local.f32 %f1425, [%rd1+-224]; ld.local.f32 %f1426, [%rd1+-220]; ld.local.f32 %f1427, [%rd1+-216]; ld.local.f32 %f1428, [%rd3+12]; ld.local.f32 %f1429, [%rd3+8]; setp.neu.ftz.f32 %p37, %f1429, %f1428; ld.local.f32 %f1430, [%rd3+16]; setp.neu.ftz.f32 %p38, %f1429, %f1430; or.pred %p39, %p37, %p38; sub.ftz.f32 %f1431, %f1430, %f1428; neg.ftz.f32 %f1432, %f1428; neg.ftz.f32 %f1433, %f1430; selp.f32 %f1434, %f1433, %f1430, %p39; add.ftz.f32 %f1435, %f1429, %f1434; selp.f32 %f1436, %f1428, %f1432, %p39; sub.ftz.f32 %f1437, %f1436, %f1429; mul.ftz.f32 %f1438, %f1435, %f1435; fma.rn.ftz.f32 %f1439, %f1431, %f1431, %f1438; fma.rn.ftz.f32 %f1440, %f1437, %f1437, %f1439; sqrt.approx.ftz.f32 %f1441, %f1440; div.approx.ftz.f32 %f1442, %f1431, %f1441; div.approx.ftz.f32 %f1443, %f1435, %f1441; div.approx.ftz.f32 %f1444, %f1437, %f1441; mul.ftz.f32 %f1445, %f1428, %f1444; mul.ftz.f32 %f1446, %f1430, %f1443; sub.ftz.f32 %f1447, %f1445, %f1446; mul.ftz.f32 %f1448, %f1430, %f1442; mul.ftz.f32 %f1449, %f1429, %f1444; sub.ftz.f32 %f1450, %f1448, %f1449; mul.ftz.f32 %f1451, %f1429, %f1443; mul.ftz.f32 %f1452, %f1428, %f1442; sub.ftz.f32 %f1453, %f1451, %f1452; mul.ftz.f32 %f1454, %f1426, %f1443; fma.rn.ftz.f32 %f1455, %f1425, %f1442, %f1454; fma.rn.ftz.f32 %f5012, %f1427, %f1444, %f1455; mul.ftz.f32 %f1456, %f1426, %f1450; fma.rn.ftz.f32 %f1457, %f1425, %f1447, %f1456; fma.rn.ftz.f32 %f5013, %f1427, %f1453, %f1457; mul.ftz.f32 %f1458, %f1426, %f1428; fma.rn.ftz.f32 %f1459, %f1425, %f1429, %f1458; fma.rn.ftz.f32 %f5014, %f1427, %f1430, %f1459; mul.ftz.f32 %f1460, %f1405, %f1443; fma.rn.ftz.f32 %f1461, %f1404, %f1442, %f1460; fma.rn.ftz.f32 %f16, %f1406, %f1444, %f1461; mul.ftz.f32 %f1462, %f1405, %f1450; fma.rn.ftz.f32 %f1463, %f1404, %f1447, %f1462; fma.rn.ftz.f32 %f17, %f1406, %f1453, %f1463; mul.ftz.f32 %f1464, %f1405, %f1428; fma.rn.ftz.f32 %f1465, %f1404, %f1429, %f1464; fma.rn.ftz.f32 %f18, %f1406, %f1430, %f1465; mov.f32 %f5002, 0f00000000; setp.geu.ftz.f32 %p1, %f18, 0f00000000; mul.ftz.f32 %f19, %f5014, %f18; setp.gt.ftz.f32 %p41, %f19, 0f00000000; ld.local.f32 %f20, [%rd3+28]; mul.ftz.f32 %f1466, %f16, %f20; mul.ftz.f32 %f1467, %f17, %f20; mul.ftz.f32 %f1468, %f18, %f20; selp.f32 %f1469, %f16, %f1466, %p41; selp.f32 %f1470, %f17, %f1467, %p41; selp.f32 %f1471, %f18, %f1468, %p41; add.ftz.f32 %f1472, %f5012, %f1469; add.ftz.f32 %f1473, %f5013, %f1470; add.ftz.f32 %f1474, %f5014, %f1471; mul.ftz.f32 %f1475, %f1473, %f1473; fma.rn.ftz.f32 %f1476, %f1472, %f1472, %f1475; fma.rn.ftz.f32 %f1477, %f1474, %f1474, %f1476; sqrt.approx.ftz.f32 %f21, %f1477; rsqrt.approx.ftz.f32 %f1478, %f1477; mul.ftz.f32 %f1479, %f1478, %f1472; mul.ftz.f32 %f1480, %f1478, %f1473; mul.ftz.f32 %f1481, %f1478, %f1474; setp.lt.ftz.f32 %p42, %f1481, 0f00000000; neg.ftz.f32 %f1482, %f1479; neg.ftz.f32 %f1483, %f1480; neg.ftz.f32 %f1484, %f1481; selp.f32 %f22, %f1482, %f1479, %p42; selp.f32 %f23, %f1483, %f1480, %p42; selp.f32 %f24, %f1484, %f1481, %p42; setp.lt.ftz.f32 %p43, %f5014, 0f00000000; neg.ftz.f32 %f25, %f5012; neg.ftz.f32 %f26, %f5013; neg.ftz.f32 %f27, %f5014; selp.f32 %f28, %f25, %f5012, %p43; selp.f32 %f29, %f26, %f5013, %p43; selp.f32 %f30, %f27, %f5014, %p43; setp.gt.ftz.f32 %p44, %f30, 0f3F7FF972; @%p44 bra $L__BB104_67; setp.lt.ftz.f32 %p45, %f30, 0fBF7FF972; mov.f32 %f5002, 0fBF7FF972; @%p45 bra $L__BB104_67; mul.ftz.f32 %f1486, %f30, %f30; mov.f32 %f1487, 0f33D6BF95; max.ftz.f32 %f1488, %f1486, %f1487; rcp.approx.ftz.f32 %f1489, %f1488; mul.ftz.f32 %f1490, %f28, %f1421; mul.ftz.f32 %f1491, %f29, %f1421; mul.ftz.f32 %f1492, %f1491, %f1491; fma.rn.ftz.f32 %f1493, %f1490, %f1490, %f1492; fma.rn.ftz.f32 %f1494, %f1493, %f1489, 0f3F800000; sqrt.approx.ftz.f32 %f1495, %f1494; neg.ftz.f32 %f1496, %f1495; setp.le.ftz.f32 %p46, %f30, 0f00000000; selp.f32 %f1497, %f1496, %f1495, %p46; add.ftz.f32 %f1498, %f1497, 0fBF800000; mul.ftz.f32 %f5002, %f1498, 0f3F000000; $L__BB104_67: mul.ftz.f32 %f1500, %f24, %f24; mov.f32 %f5003, 0f3F800000; sub.ftz.f32 %f1501, %f5003, %f1500; mul.ftz.f32 %f33, %f1421, %f1421; fma.rn.ftz.f32 %f1502, %f33, %f1500, %f1501; mul.ftz.f32 %f1503, %f1502, 0f40490FDB; mul.ftz.f32 %f1504, %f1502, %f1503; mov.f32 %f1505, 0f33D6BF95; max.ftz.f32 %f1506, %f1504, %f1505; div.approx.ftz.f32 %f34, %f33, %f1506; mul.ftz.f32 %f1507, %f29, %f23; fma.rn.ftz.f32 %f1508, %f28, %f22, %f1507; fma.rn.ftz.f32 %f35, %f30, %f24, %f1508; abs.ftz.f32 %f36, %f35; fma.rn.ftz.f32 %f37, %f20, %f20, 0fBF800000; fma.rn.ftz.f32 %f38, %f36, %f36, %f37; setp.leu.ftz.f32 %p47, %f38, 0f00000000; @%p47 bra $L__BB104_69; sqrt.approx.ftz.f32 %f1509, %f38; sub.ftz.f32 %f1510, %f1509, %f36; add.ftz.f32 %f1511, %f36, %f1509; div.approx.ftz.f32 %f1512, %f1510, %f1511; fma.rn.ftz.f32 %f1513, %f36, %f1511, 0fBF800000; fma.rn.ftz.f32 %f1514, %f36, %f1510, 0f3F800000; div.approx.ftz.f32 %f1515, %f1513, %f1514; mul.ftz.f32 %f1516, %f1512, 0f3F000000; mul.ftz.f32 %f1517, %f1512, %f1516; fma.rn.ftz.f32 %f1518, %f1515, %f1515, 0f3F800000; mul.ftz.f32 %f5003, %f1517, %f1518; $L__BB104_69: setp.gt.ftz.f32 %p522, %f19, 0f00000000; mul.ftz.f32 %f1519, %f18, 0f3EA2F983; abs.ftz.f32 %f41, %f1519; add.ftz.f32 %f1520, %f5002, 0f3F800000; mul.ftz.f32 %f42, %f30, %f1520; @%p522 bra $L__BB104_73; bra.uni $L__BB104_70; $L__BB104_73: mul.ftz.f32 %f1578, %f34, 0f3E800000; max.ftz.f32 %f1580, %f42, %f1505; div.approx.ftz.f32 %f1581, %f1578, %f1580; mul.ftz.f32 %f1582, %f1421, 0fBFFE59F3; mul.ftz.f32 %f1583, %f1421, %f1582; mul.ftz.f32 %f1584, %f1583, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1585, %f1584; fma.rn.ftz.f32 %f1586, %f1585, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f1587, %f1421, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f1588, %f1421, %f1587, 0f417C8B44; fma.rn.ftz.f32 %f1589, %f1421, %f1588, 0fC1665E35; fma.rn.ftz.f32 %f1590, %f1421, %f1589, 0f40CE7881; fma.rn.ftz.f32 %f1591, %f1421, %f1590, 0fBF993E81; fma.rn.ftz.f32 %f1592, %f1421, %f1591, 0f3CE3C31E; fma.rn.ftz.f32 %f1593, %f1421, %f1592, 0f3BBA5D49; sub.ftz.f32 %f1594, %f1586, %f1593; cvt.ftz.sat.f32.f32 %f1595, %f1594; mov.f32 %f1596, 0f3F800000; sub.ftz.f32 %f1597, %f1596, %f1595; mul.ftz.f32 %f1598, %f41, %f1597; fma.rn.ftz.f32 %f1599, %f1581, %f1595, %f1598; mul.ftz.f32 %f5291, %f5003, %f1599; bra.uni $L__BB104_74; $L__BB104_26: setp.eq.s32 %p28, %r5, 8; @%p28 bra $L__BB104_504; setp.eq.s32 %p29, %r5, 13; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p29 bra $L__BB104_28; bra.uni $L__BB104_519; $L__BB104_28: ld.local.f32 %f1105, [%rd1+-224]; ld.local.f32 %f1106, [%rd1+-220]; ld.local.f32 %f1107, [%rd1+-216]; ld.local.v2.f32 {%f3918, %f3919}, [%rd3+20]; ld.local.f32 %f3920, [%rd1+-248]; ld.local.f32 %f3921, [%rd1+-244]; mul.ftz.f32 %f3922, %f1405, %f3921; fma.rn.ftz.f32 %f3923, %f1404, %f3920, %f3922; ld.local.f32 %f3924, [%rd1+-240]; fma.rn.ftz.f32 %f3925, %f1406, %f3924, %f3923; setp.lt.ftz.f32 %p412, %f3925, 0f00000000; setp.eq.s32 %p413, %r5, 22; xor.pred %p414, %p412, %p413; mul.ftz.f32 %f1110, %f3918, %f3919; setp.le.ftz.f32 %p415, %f1110, 0f33D6BF95; or.pred %p416, %p414, %p415; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p416 bra $L__BB104_463; bra.uni $L__BB104_29; $L__BB104_463: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_8: setp.eq.s32 %p34, %r5, 2; @%p34 bra $L__BB104_518; setp.eq.s32 %p35, %r5, 3; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p35 bra $L__BB104_10; bra.uni $L__BB104_519; $L__BB104_10: ld.local.f32 %f1344, [%rd3+8]; ld.local.f32 %f1345, [%rd3+12]; mul.ftz.f32 %f4725, %f1405, %f1345; fma.rn.ftz.f32 %f4726, %f1404, %f1344, %f4725; ld.local.f32 %f1346, [%rd3+16]; fma.rn.ftz.f32 %f1347, %f1406, %f1346, %f4726; setp.leu.ftz.f32 %p489, %f1347, 0f00000000; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p489 bra $L__BB104_14; ld.local.f32 %f4727, [%rd1+-224]; ld.local.f32 %f4728, [%rd1+-220]; ld.local.f32 %f4729, [%rd1+-216]; mov.f32 %f4730, 0f00000000; max.ftz.f32 %f1348, %f1347, %f4730; mul.ftz.f32 %f4731, %f4728, %f1345; fma.rn.ftz.f32 %f4732, %f4727, %f1344, %f4731; fma.rn.ftz.f32 %f4733, %f4729, %f1346, %f4732; max.ftz.f32 %f1349, %f4733, %f4730; mul.ftz.f32 %f4734, %f1405, %f4728; fma.rn.ftz.f32 %f4735, %f1404, %f4727, %f4734; fma.rn.ftz.f32 %f4736, %f1406, %f4729, %f4735; mul.ftz.f32 %f4737, %f1348, %f1349; sub.ftz.f32 %f5285, %f4736, %f4737; setp.leu.ftz.f32 %p490, %f5285, 0f00000000; @%p490 bra $L__BB104_13; max.ftz.f32 %f4738, %f1348, %f1349; add.ftz.f32 %f4739, %f4738, 0f00800000; div.approx.ftz.f32 %f5285, %f5285, %f4739; $L__BB104_13: ld.local.f32 %f4741, [%rd3+28]; ld.local.f32 %f4742, [%rd3+24]; fma.rn.ftz.f32 %f4743, %f5285, %f4741, %f4742; mul.ftz.f32 %f5288, %f1348, %f4743; mov.f32 %f5291, 0f3E22F983; $L__BB104_14: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_39: setp.eq.s32 %p25, %r5, 16; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; @%p25 bra $L__BB104_40; bra.uni $L__BB104_519; $L__BB104_40: ld.local.f32 %f3323, [%rd1+-248]; ld.local.f32 %f3324, [%rd1+-244]; mul.ftz.f32 %f3325, %f1405, %f3324; fma.rn.ftz.f32 %f3326, %f1404, %f3323, %f3325; ld.local.f32 %f3327, [%rd1+-240]; fma.rn.ftz.f32 %f3328, %f1406, %f3327, %f3326; ld.local.f32 %f838, [%rd3+8]; ld.local.f32 %f839, [%rd1+-224]; ld.local.f32 %f840, [%rd3+12]; ld.local.f32 %f841, [%rd1+-220]; mul.ftz.f32 %f3329, %f841, %f840; fma.rn.ftz.f32 %f3330, %f839, %f838, %f3329; ld.local.f32 %f842, [%rd3+16]; ld.local.f32 %f843, [%rd1+-216]; fma.rn.ftz.f32 %f844, %f843, %f842, %f3330; mul.ftz.f32 %f3331, %f1405, %f840; fma.rn.ftz.f32 %f3332, %f1404, %f838, %f3331; fma.rn.ftz.f32 %f845, %f1406, %f842, %f3332; setp.lt.ftz.f32 %p332, %f3328, 0f00000000; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p332 bra $L__BB104_369; ld.local.v2.f32 {%f3337, %f3338}, [%rd3+20]; max.ftz.f32 %f3339, %f3337, %f3338; setp.le.ftz.f32 %p333, %f3339, 0f38D1B717; mov.f32 %f5288, %f5291; @%p333 bra $L__BB104_369; setp.leu.ftz.f32 %p334, %f844, 0f00000000; setp.leu.ftz.f32 %p335, %f845, 0f00000000; or.pred %p336, %p335, %p334; mov.f32 %f5288, %f5291; @%p336 bra $L__BB104_369; add.ftz.f32 %f3344, %f1404, %f839; add.ftz.f32 %f3345, %f1405, %f841; mul.ftz.f32 %f3346, %f3345, %f3345; fma.rn.ftz.f32 %f3347, %f3344, %f3344, %f3346; add.ftz.f32 %f3348, %f1406, %f843; fma.rn.ftz.f32 %f3349, %f3348, %f3348, %f3347; sqrt.approx.ftz.f32 %f3350, %f3349; div.approx.ftz.f32 %f848, %f3344, %f3350; div.approx.ftz.f32 %f849, %f3345, %f3350; div.approx.ftz.f32 %f850, %f3348, %f3350; mul.ftz.f32 %f3351, %f841, %f849; fma.rn.ftz.f32 %f3352, %f839, %f848, %f3351; fma.rn.ftz.f32 %f3353, %f843, %f850, %f3352; abs.ftz.f32 %f3354, %f3353; mov.f32 %f3355, 0f358637BD; max.ftz.f32 %f851, %f3354, %f3355; mul.ftz.f32 %f3356, %f840, %f849; fma.rn.ftz.f32 %f3357, %f838, %f848, %f3356; fma.rn.ftz.f32 %f3358, %f842, %f850, %f3357; max.ftz.f32 %f852, %f3358, %f3355; max.ftz.f32 %f853, %f845, %f3355; mul.ftz.f32 %f3359, %f3337, %f3337; mov.f32 %f3360, 0f40000000; div.approx.ftz.f32 %f3361, %f3360, %f3359; add.ftz.f32 %f854, %f3361, 0fC0000000; mul.ftz.f32 %f3362, %f3338, %f3338; div.approx.ftz.f32 %f3363, %f3360, %f3362; add.ftz.f32 %f855, %f3363, 0fC0000000; setp.eq.ftz.f32 %p337, %f854, %f855; @%p337 bra $L__BB104_367; bra.uni $L__BB104_44; $L__BB104_367: lg2.approx.ftz.f32 %f3415, %f852; mul.ftz.f32 %f3416, %f854, %f3415; ex2.approx.ftz.f32 %f3417, %f3416; add.ftz.f32 %f3418, %f854, 0f3F800000; mov.f32 %f3419, 0f41C90FDB; div.approx.ftz.f32 %f3420, %f3418, %f3419; mul.ftz.f32 %f3421, %f853, %f3420; mul.ftz.f32 %f5166, %f3417, %f3421; mul.ftz.f32 %f3422, %f3417, %f3420; div.approx.ftz.f32 %f5291, %f3422, %f851; bra.uni $L__BB104_368; $L__BB104_509: ld.local.f32 %f1318, [%rd1+-224]; ld.local.f32 %f1319, [%rd1+-220]; ld.local.f32 %f1320, [%rd1+-216]; ld.local.f32 %f1321, [%rd3+8]; ld.local.f32 %f1322, [%rd3+12]; mul.ftz.f32 %f4687, %f1405, %f1322; fma.rn.ftz.f32 %f4688, %f1404, %f1321, %f4687; ld.local.f32 %f1323, [%rd3+16]; fma.rn.ftz.f32 %f1324, %f1406, %f1323, %f4688; setp.leu.ftz.f32 %p484, %f1324, 0f00000000; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p484 bra $L__BB104_517; mov.f32 %f5288, 0f00000000; max.ftz.f32 %f4692, %f1324, %f5288; mul.ftz.f32 %f5291, %f4692, 0f3EA2F983; setp.le.ftz.f32 %p485, %f1324, 0f00000000; @%p485 bra $L__BB104_517; mul.ftz.f32 %f4693, %f1319, %f1322; fma.rn.ftz.f32 %f4694, %f1318, %f1321, %f4693; fma.rn.ftz.f32 %f4695, %f1320, %f1323, %f4694; mov.f32 %f4696, 0f3F800000; sub.ftz.f32 %f4697, %f4696, %f4695; mov.f32 %f4698, 0f00000000; max.ftz.f32 %f4699, %f4697, %f4698; min.ftz.f32 %f4700, %f4699, %f4696; mul.ftz.f32 %f4701, %f4700, %f4700; mul.ftz.f32 %f4702, %f4701, %f4701; mul.ftz.f32 %f1326, %f4700, %f4702; sub.ftz.f32 %f4703, %f4696, %f1324; max.ftz.f32 %f4704, %f4703, %f4698; min.ftz.f32 %f4705, %f4704, %f4696; mul.ftz.f32 %f4706, %f4705, %f4705; mul.ftz.f32 %f4707, %f4706, %f4706; mul.ftz.f32 %f1327, %f4705, %f4707; ld.local.u32 %r30, [%rd3+24]; and.b32 %r141, %r30, 3; setp.eq.s32 %p486, %r141, 0; fma.rn.ftz.f32 %f1328, %f1327, 0fBF000000, 0f3F800000; @%p486 bra $L__BB104_513; fma.rn.ftz.f32 %f4708, %f1326, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f5282, %f4708, %f1328, 0f00000000; bra.uni $L__BB104_514; $L__BB104_505: ld.local.f32 %f1300, [%rd3+8]; ld.local.f32 %f1301, [%rd3+12]; mul.ftz.f32 %f4652, %f1405, %f1301; fma.rn.ftz.f32 %f4653, %f1404, %f1300, %f4652; ld.local.f32 %f1302, [%rd3+16]; fma.rn.ftz.f32 %f1303, %f1406, %f1302, %f4653; setp.leu.ftz.f32 %p480, %f1303, 0f00000000; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p480 bra $L__BB104_508; ld.local.f32 %f4658, [%rd1+-224]; ld.local.f32 %f4659, [%rd1+-220]; ld.local.f32 %f4660, [%rd1+-216]; add.ftz.f32 %f4661, %f1404, %f4658; add.ftz.f32 %f4662, %f1405, %f4659; mul.ftz.f32 %f4663, %f4662, %f4662; fma.rn.ftz.f32 %f4664, %f4661, %f4661, %f4663; add.ftz.f32 %f4665, %f1406, %f4660; fma.rn.ftz.f32 %f4666, %f4665, %f4665, %f4664; sqrt.approx.ftz.f32 %f4667, %f4666; div.approx.ftz.f32 %f1304, %f4661, %f4667; div.approx.ftz.f32 %f1305, %f4662, %f4667; div.approx.ftz.f32 %f1306, %f4665, %f4667; mul.ftz.f32 %f4668, %f4659, %f1301; fma.rn.ftz.f32 %f4669, %f4658, %f1300, %f4668; fma.rn.ftz.f32 %f4670, %f4660, %f1302, %f4669; setp.lt.ftz.f32 %p481, %f4670, 0f00000000; setp.lt.ftz.f32 %p482, %f1303, 0f00000000; or.pred %p483, %p482, %p481; mov.f32 %f5288, %f5291; @%p483 bra $L__BB104_508; mov.f32 %f4671, 0f00000000; max.ftz.f32 %f4672, %f1303, %f4671; mul.ftz.f32 %f5291, %f4672, 0f3EA2F983; mul.ftz.f32 %f4673, %f1405, %f1305; fma.rn.ftz.f32 %f4674, %f1404, %f1304, %f4673; fma.rn.ftz.f32 %f4675, %f1406, %f1306, %f4674; mov.f32 %f4676, 0f3F800000; sub.ftz.f32 %f4677, %f4676, %f4675; max.ftz.f32 %f4678, %f4677, %f4671; min.ftz.f32 %f4679, %f4678, %f4676; mul.ftz.f32 %f4680, %f4679, %f4679; mul.ftz.f32 %f4681, %f4680, %f4680; mul.ftz.f32 %f4682, %f4679, %f4681; mul.ftz.f32 %f5288, %f1303, %f4682; $L__BB104_508: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_466: setp.leu.ftz.f32 %p437, %f1184, 0f00000000; setp.leu.ftz.f32 %p438, %f1183, 0f00000000; or.pred %p439, %p438, %p437; mov.f32 %f5290, %f5291; mov.f32 %f5289, %f5291; mov.f32 %f5288, %f5291; @%p439 bra $L__BB104_519; add.ftz.f32 %f4149, %f1404, %f1173; add.ftz.f32 %f4150, %f1405, %f1174; mul.ftz.f32 %f4151, %f4150, %f4150; fma.rn.ftz.f32 %f4152, %f4149, %f4149, %f4151; add.ftz.f32 %f4153, %f1406, %f1175; fma.rn.ftz.f32 %f4154, %f4153, %f4153, %f4152; sqrt.approx.ftz.f32 %f4155, %f4154; div.approx.ftz.f32 %f1185, %f4149, %f4155; div.approx.ftz.f32 %f1186, %f4150, %f4155; div.approx.ftz.f32 %f1187, %f4153, %f4155; mul.ftz.f32 %f4156, %f1181, %f1186; fma.rn.ftz.f32 %f4157, %f1180, %f1185, %f4156; fma.rn.ftz.f32 %f1188, %f1182, %f1187, %f4157; mul.ftz.f32 %f1189, %f1188, %f1188; setp.eq.ftz.f32 %p440, %f4134, %f4135; @%p440 bra $L__BB104_469; bra.uni $L__BB104_468; $L__BB104_469: setp.eq.s32 %p441, %r5, 12; @%p441 bra $L__BB104_471; bra.uni $L__BB104_470; $L__BB104_471: setp.ge.ftz.f32 %p442, %f4134, 0f3F800000; mov.f32 %f5256, 0f3EA2F983; mov.f32 %f5255, 0f3D800000; @%p442 bra $L__BB104_473; mul.ftz.f32 %f4255, %f4134, %f4134; add.ftz.f32 %f4256, %f4255, 0fBF800000; mul.ftz.f32 %f4257, %f1188, %f4256; fma.rn.ftz.f32 %f4258, %f1188, %f4257, 0f3F800000; lg2.approx.ftz.f32 %f4259, %f4255; mul.ftz.f32 %f4260, %f4259, 0f3F317218; mul.ftz.f32 %f4261, %f4260, 0f40490FDB; mul.ftz.f32 %f4262, %f4258, %f4261; div.approx.ftz.f32 %f5256, %f4256, %f4262; bra.uni $L__BB104_473; $L__BB104_363: ld.local.f32 %f818, [%rd3+8]; ld.local.f32 %f819, [%rd1+-224]; ld.local.f32 %f820, [%rd3+12]; ld.local.f32 %f821, [%rd1+-220]; mul.ftz.f32 %f3269, %f821, %f820; fma.rn.ftz.f32 %f3270, %f819, %f818, %f3269; ld.local.f32 %f822, [%rd3+16]; ld.local.f32 %f823, [%rd1+-216]; fma.rn.ftz.f32 %f824, %f823, %f822, %f3270; mul.ftz.f32 %f3271, %f1405, %f820; fma.rn.ftz.f32 %f3272, %f1404, %f818, %f3271; fma.rn.ftz.f32 %f825, %f1406, %f822, %f3272; setp.leu.ftz.f32 %p326, %f824, 0f00000000; setp.leu.ftz.f32 %p327, %f825, 0f00000000; or.pred %p328, %p327, %p326; mov.f32 %f5291, 0f00000000; mov.f32 %f5288, %f5291; @%p328 bra $L__BB104_366; add.ftz.f32 %f3277, %f1404, %f819; add.ftz.f32 %f3278, %f1405, %f821; mul.ftz.f32 %f3279, %f3278, %f3278; fma.rn.ftz.f32 %f3280, %f3277, %f3277, %f3279; add.ftz.f32 %f3281, %f1406, %f823; fma.rn.ftz.f32 %f3282, %f3281, %f3281, %f3280; sqrt.approx.ftz.f32 %f3283, %f3282; div.approx.ftz.f32 %f3284, %f3277, %f3283; div.approx.ftz.f32 %f3285, %f3278, %f3283; div.approx.ftz.f32 %f3286, %f3281, %f3283; mul.ftz.f32 %f3287, %f820, %f3285; fma.rn.ftz.f32 %f3288, %f818, %f3284, %f3287; fma.rn.ftz.f32 %f826, %f822, %f3286, %f3288; mul.ftz.f32 %f3289, %f821, %f3285; fma.rn.ftz.f32 %f3290, %f819, %f3284, %f3289; fma.rn.ftz.f32 %f3291, %f823, %f3286, %f3290; abs.ftz.f32 %f827, %f3291; abs.ftz.f32 %f3292, %f826; setp.geu.ftz.f32 %p329, %f3292, 0f3F7FFF58; setp.leu.ftz.f32 %p330, %f827, 0f3727C5AC; or.pred %p331, %p330, %p329; mov.f32 %f5288, %f5291; @%p331 bra $L__BB104_366; ld.local.f32 %f3294, [%rd3+24]; div.approx.ftz.f32 %f3295, %f826, %f827; mov.f32 %f3296, 0f3727C5AC; max.ftz.f32 %f3297, %f3295, %f3296; mul.ftz.f32 %f3298, %f824, %f3297; abs.ftz.f32 %f3299, %f3298; add.ftz.f32 %f3300, %f3299, %f3299; mul.ftz.f32 %f3301, %f825, %f3297; abs.ftz.f32 %f3302, %f3301; add.ftz.f32 %f3303, %f3302, %f3302; mul.ftz.f32 %f3304, %f826, %f826; mov.f32 %f3305, 0f3F800000; sub.ftz.f32 %f3306, %f3305, %f3304; mul.ftz.f32 %f3307, %f3306, %f3306; div.approx.ftz.f32 %f3308, %f3304, %f3306; mul.ftz.f32 %f3309, %f3294, %f3308; mul.ftz.f32 %f3310, %f3309, 0fBFB8AA3B; ex2.approx.ftz.f32 %f3311, %f3310; mul.ftz.f32 %f3312, %f3294, %f3311; mul.ftz.f32 %f3313, %f3312, 0f3EA2F983; div.approx.ftz.f32 %f3314, %f3313, %f3307; min.ftz.f32 %f3315, %f3300, %f3303; min.ftz.f32 %f3316, %f3305, %f3315; mul.ftz.f32 %f3317, %f3314, %f3316; mul.ftz.f32 %f3318, %f3317, 0f3E800000; div.approx.ftz.f32 %f5288, %f3318, %f824; mov.f32 %f5291, 0f3E22F983; $L__BB104_366: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_357: ld.local.f32 %f3209, [%rd3+20]; mul.ftz.f32 %f794, %f3209, 0f3FC90FDB; ld.local.f32 %f3210, [%rd3+24]; mul.ftz.f32 %f795, %f3210, 0f3FC90FDB; ld.local.f32 %f796, [%rd3+8]; ld.local.f32 %f797, [%rd3+12]; mul.ftz.f32 %f3211, %f1405, %f797; fma.rn.ftz.f32 %f3212, %f1404, %f796, %f3211; ld.local.f32 %f798, [%rd3+16]; fma.rn.ftz.f32 %f3213, %f1406, %f798, %f3212; ld.local.f32 %f799, [%rd1+-224]; ld.local.f32 %f800, [%rd1+-220]; mul.ftz.f32 %f3214, %f800, %f797; fma.rn.ftz.f32 %f3215, %f799, %f796, %f3214; ld.local.f32 %f801, [%rd1+-216]; fma.rn.ftz.f32 %f802, %f801, %f798, %f3215; setp.leu.ftz.f32 %p317, %f3213, 0f00000000; setp.leu.ftz.f32 %p318, %f802, 0f00000000; mov.f32 %f5291, 0f00000000; or.pred %p319, %p317, %p318; mov.f32 %f5288, %f5291; @%p319 bra $L__BB104_362; add.ftz.f32 %f3217, %f802, %f802; mul.ftz.f32 %f3218, %f3217, %f796; mul.ftz.f32 %f3219, %f3217, %f797; mul.ftz.f32 %f3220, %f3217, %f798; sub.ftz.f32 %f3221, %f3218, %f799; sub.ftz.f32 %f3222, %f3219, %f800; sub.ftz.f32 %f3223, %f3220, %f801; mul.ftz.f32 %f3224, %f1405, %f3222; fma.rn.ftz.f32 %f3225, %f1404, %f3221, %f3224; fma.rn.ftz.f32 %f3226, %f1406, %f3223, %f3225; mov.f32 %f3227, 0f00000000; max.ftz.f32 %f3228, %f3226, %f3227; mov.f32 %f3229, 0fBF800000; max.ftz.f32 %f3230, %f3228, %f3229; mov.f32 %f3216, 0f3F800000; min.ftz.f32 %f3231, %f3230, %f3216; abs.ftz.f32 %f3232, %f3231; sub.ftz.f32 %f3233, %f3216, %f3232; mul.ftz.f32 %f3234, %f3233, 0f3F000000; sqrt.approx.ftz.f32 %f3235, %f3234; setp.gt.ftz.f32 %p320, %f3232, 0f3F11EB85; selp.f32 %f3236, %f3235, %f3232, %p320; mul.ftz.f32 %f3237, %f3236, %f3236; mov.f32 %f3238, 0f3C94D2E9; mov.f32 %f3239, 0f3D53F941; fma.rn.ftz.f32 %f3240, %f3239, %f3237, %f3238; mov.f32 %f3241, 0f3D3F841F; fma.rn.ftz.f32 %f3242, %f3240, %f3237, %f3241; mov.f32 %f3243, 0f3D994929; fma.rn.ftz.f32 %f3244, %f3242, %f3237, %f3243; mov.f32 %f3245, 0f3E2AAB94; fma.rn.ftz.f32 %f3246, %f3244, %f3237, %f3245; mul.ftz.f32 %f3247, %f3237, %f3246; fma.rn.ftz.f32 %f3248, %f3247, %f3236, %f3236; add.ftz.f32 %f3249, %f3248, %f3248; mov.f32 %f3250, 0f3FC90FDB; sub.ftz.f32 %f3251, %f3250, %f3248; selp.f32 %f3252, %f3249, %f3251, %p320; setp.lt.ftz.f32 %p321, %f3231, 0f00000000; mov.f32 %f3253, 0f40490FDB; sub.ftz.f32 %f3254, %f3253, %f3252; selp.f32 %f803, %f3254, %f3252, %p321; setp.lt.ftz.f32 %p322, %f803, %f794; add.ftz.f32 %f804, %f794, %f795; mov.f32 %f5159, %f3216; @%p322 bra $L__BB104_361; setp.leu.ftz.f32 %p323, %f804, %f803; setp.eq.ftz.f32 %p324, %f795, 0f00000000; or.pred %p325, %p324, %p323; mov.f32 %f5159, %f3227; @%p325 bra $L__BB104_361; sub.ftz.f32 %f3256, %f803, %f794; div.approx.ftz.f32 %f3257, %f3256, %f795; mov.f32 %f3258, 0f3F800000; sub.ftz.f32 %f5159, %f3258, %f3257; $L__BB104_361: min.ftz.f32 %f3260, %f804, %f3250; cos.approx.ftz.f32 %f3261, %f3260; sub.ftz.f32 %f3263, %f3216, %f3261; mov.f32 %f3264, 0f3E22F983; div.approx.ftz.f32 %f5291, %f3264, %f3263; mul.ftz.f32 %f5288, %f5159, %f5291; $L__BB104_362: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_485: ld.local.f32 %f4394, [%rd3+8]; ld.local.f32 %f4395, [%rd3+12]; mul.ftz.f32 %f4396, %f1405, %f4395; fma.rn.ftz.f32 %f4397, %f1404, %f4394, %f4396; ld.local.f32 %f4398, [%rd3+16]; fma.rn.ftz.f32 %f4399, %f1406, %f4398, %f4397; setp.ge.ftz.f32 %p450, %f4399, 0f00000000; mov.f32 %f5288, 0f00000000; @%p450 bra $L__BB104_503; ld.local.f32 %f4404, [%rd1+-224]; ld.local.f32 %f4405, [%rd1+-220]; ld.local.f32 %f4406, [%rd1+-216]; ld.local.f32 %f1247, [%rd3+40]; ld.local.f32 %f1248, [%rd3+32]; ld.local.f32 %f1249, [%rd3+36]; ld.local.f32 %f4407, [%rd3+20]; ld.local.f32 %f4408, [%rd3+24]; mul.ftz.f32 %f4409, %f4405, %f4408; fma.rn.ftz.f32 %f4410, %f4404, %f4407, %f4409; ld.local.f32 %f4411, [%rd3+28]; fma.rn.ftz.f32 %f4412, %f4406, %f4411, %f4410; mul.ftz.f32 %f4413, %f4407, %f4412; mul.ftz.f32 %f4414, %f4408, %f4412; mul.ftz.f32 %f4415, %f4411, %f4412; sub.ftz.f32 %f4416, %f4404, %f4413; sub.ftz.f32 %f4417, %f4405, %f4414; sub.ftz.f32 %f4418, %f4406, %f4415; mul.ftz.f32 %f4419, %f4417, %f4417; fma.rn.ftz.f32 %f4420, %f4416, %f4416, %f4419; fma.rn.ftz.f32 %f4421, %f4418, %f4418, %f4420; sqrt.approx.ftz.f32 %f4422, %f4421; div.approx.ftz.f32 %f4423, %f4416, %f4422; div.approx.ftz.f32 %f4424, %f4417, %f4422; div.approx.ftz.f32 %f4425, %f4418, %f4422; abs.ftz.f32 %f4426, %f4412; setp.lt.ftz.f32 %p451, %f4426, 0f3F800000; mov.f32 %f4427, 0f3F800000; add.ftz.f32 %f4428, %f4426, 0fBF800000; add.ftz.f32 %f4429, %f4428, 0f3F800000; selp.f32 %f4430, %f4429, 0f3F800000, %p451; sub.ftz.f32 %f4431, %f4427, %f4430; sqrt.approx.ftz.f32 %f4432, %f4431; mul.ftz.f32 %f4433, %f4430, 0f3CB1485D; mov.f32 %f4434, 0f3D9FB43B; sub.ftz.f32 %f4435, %f4434, %f4433; fma.rn.ftz.f32 %f4436, %f4430, %f4435, 0fBE5A6B93; fma.rn.ftz.f32 %f4437, %f4430, %f4436, 0f3FC90FDB; mov.f32 %f4438, 0f3FC90FDB; mul.ftz.f32 %f4439, %f4432, %f4437; setp.lt.ftz.f32 %p452, %f4412, 0f00000000; mov.f32 %f4440, 0f40490FDB; sub.ftz.f32 %f4441, %f4440, %f4439; selp.f32 %f4442, %f4441, %f4439, %p452; sub.ftz.f32 %f1250, %f4438, %f4442; mul.ftz.f32 %f4443, %f1405, %f4408; fma.rn.ftz.f32 %f4444, %f1404, %f4407, %f4443; fma.rn.ftz.f32 %f4445, %f1406, %f4411, %f4444; mul.ftz.f32 %f4446, %f4407, %f4445; mul.ftz.f32 %f4447, %f4408, %f4445; mul.ftz.f32 %f4448, %f4411, %f4445; sub.ftz.f32 %f4449, %f1404, %f4446; sub.ftz.f32 %f4450, %f1405, %f4447; sub.ftz.f32 %f4451, %f1406, %f4448; mul.ftz.f32 %f4452, %f4450, %f4450; fma.rn.ftz.f32 %f4453, %f4449, %f4449, %f4452; fma.rn.ftz.f32 %f4454, %f4451, %f4451, %f4453; sqrt.approx.ftz.f32 %f4455, %f4454; div.approx.ftz.f32 %f4456, %f4449, %f4455; div.approx.ftz.f32 %f4457, %f4450, %f4455; div.approx.ftz.f32 %f4458, %f4451, %f4455; abs.ftz.f32 %f4459, %f4445; setp.lt.ftz.f32 %p453, %f4459, 0f3F800000; add.ftz.f32 %f4460, %f4459, 0fBF800000; add.ftz.f32 %f4461, %f4460, 0f3F800000; selp.f32 %f4462, %f4461, 0f3F800000, %p453; sub.ftz.f32 %f4463, %f4427, %f4462; sqrt.approx.ftz.f32 %f4464, %f4463; mul.ftz.f32 %f4465, %f4462, 0f3CB1485D; sub.ftz.f32 %f4466, %f4434, %f4465; fma.rn.ftz.f32 %f4467, %f4462, %f4466, 0fBE5A6B93; fma.rn.ftz.f32 %f4468, %f4462, %f4467, 0f3FC90FDB; mul.ftz.f32 %f4469, %f4464, %f4468; setp.lt.ftz.f32 %p454, %f4445, 0f00000000; sub.ftz.f32 %f4470, %f4440, %f4469; selp.f32 %f4471, %f4470, %f4469, %p454; sub.ftz.f32 %f1251, %f4438, %f4471; mul.ftz.f32 %f4472, %f4424, %f4457; fma.rn.ftz.f32 %f4473, %f4423, %f4456, %f4472; fma.rn.ftz.f32 %f1252, %f4425, %f4458, %f4473; abs.ftz.f32 %f4474, %f1251; sub.ftz.f32 %f4475, %f4438, %f4474; setp.lt.ftz.f32 %p455, %f4475, 0f3A83126F; @%p455 bra $L__BB104_503; abs.ftz.f32 %f4477, %f1252; setp.lt.ftz.f32 %p456, %f4477, 0f3F800000; add.ftz.f32 %f4479, %f4477, 0fBF800000; add.ftz.f32 %f4480, %f4479, 0f3F800000; selp.f32 %f4481, %f4480, 0f3F800000, %p456; sub.ftz.f32 %f4482, %f4427, %f4481; sqrt.approx.ftz.f32 %f4483, %f4482; fma.rn.ftz.f32 %f4484, %f4481, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f4485, %f4481, %f4484, 0fBE5A6B93; fma.rn.ftz.f32 %f4486, %f4481, %f4485, 0f3FC90FDB; mul.ftz.f32 %f4487, %f4483, %f4486; sub.ftz.f32 %f4489, %f4440, %f4487; setp.lt.ftz.f32 %p457, %f1252, 0f00000000; mov.f32 %f5274, 0f00000000; selp.f32 %f1253, %f4489, %f4487, %p457; mul.ftz.f32 %f4490, %f1251, 0f3EA2F983; mov.b32 %r123, %f4490; and.b32 %r124, %r123, -2147483648; or.b32 %r125, %r124, 1056964608; mov.b32 %f4491, %r125; add.ftz.f32 %f4492, %f4490, %f4491; cvt.rzi.ftz.s32.f32 %r126, %f4492; cvt.rn.f32.s32 %f4493, %r126; fma.rn.ftz.f32 %f4494, %f4493, 0fC0490000, %f1251; fma.rn.ftz.f32 %f4495, %f4493, 0fBA7DA000, %f4494; fma.rn.ftz.f32 %f4496, %f4493, 0fB4222000, %f4495; fma.rn.ftz.f32 %f4497, %f4493, 0fACB4611A, %f4496; add.ftz.f32 %f4498, %f4497, 0fBFC90FDB; add.ftz.f32 %f4499, %f4498, 0f3FC90FDB; mul.ftz.f32 %f4500, %f4499, %f4499; fma.rn.ftz.f32 %f4501, %f4500, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f4502, %f4500, %f4501, 0fBAB60B58; fma.rn.ftz.f32 %f4503, %f4500, %f4502, 0f3D2AAAAA; fma.rn.ftz.f32 %f4504, %f4500, %f4503, 0fBF000000; fma.rn.ftz.f32 %f4505, %f4500, %f4504, 0f3F800000; and.b32 %r127, %r126, 1; setp.eq.b32 %p458, %r127, 1; neg.ftz.f32 %f4506, %f4505; selp.f32 %f1254, %f4506, %f4505, %p458; add.ftz.f32 %f4507, %f1250, 0f3FC90FDB; mov.f32 %f4508, 0f40000000; div.approx.ftz.f32 %f4509, %f4507, %f4508; sub.ftz.f32 %f4510, %f4509, %f1247; rcp.approx.ftz.f32 %f1255, %f1248; mul.ftz.f32 %f1256, %f1255, %f4510; abs.ftz.f32 %f1257, %f1256; setp.eq.ftz.f32 %p459, %f1257, 0f00000000; mov.f32 %f5272, %f5274; @%p459 bra $L__BB104_492; setp.eq.ftz.f32 %p460, %f1257, 0f3F800000; mov.f32 %f5272, %f4427; @%p460 bra $L__BB104_492; setp.gt.ftz.f32 %p461, %f1257, 0f3F800000; @%p461 bra $L__BB104_491; bra.uni $L__BB104_490; $L__BB104_491: mov.f32 %f4513, 0f3F800000; div.approx.ftz.f32 %f5272, %f4513, %f1257; bra.uni $L__BB104_492; $L__BB104_204: ld.local.u64 %rd5, [%rd3+52]; ld.v4.f32 {%f2361, %f2362, %f2363, %f2364}, [%rd5]; ld.local.f32 %f5117, [%rd1+-144]; ld.local.f32 %f5118, [%rd1+-140]; mul.ftz.f32 %f2365, %f5118, %f5118; fma.rn.ftz.f32 %f2366, %f5117, %f5117, %f2365; ld.local.f32 %f5119, [%rd1+-136]; fma.rn.ftz.f32 %f2367, %f5119, %f5119, %f2366; sqrt.approx.ftz.f32 %f435, %f2367; setp.neu.ftz.f32 %p174, %f435, 0f00000000; @%p174 bra $L__BB104_205; bra.uni $L__BB104_206; $L__BB104_205: rcp.approx.ftz.f32 %f2368, %f435; mul.ftz.f32 %f5119, %f5119, %f2368; mul.ftz.f32 %f5118, %f5118, %f2368; mul.ftz.f32 %f5117, %f5117, %f2368; $L__BB104_206: mul.ftz.f32 %f2369, %f2363, %f5118; mul.ftz.f32 %f2370, %f2362, %f5119; sub.ftz.f32 %f5120, %f2369, %f2370; mul.ftz.f32 %f2371, %f2361, %f5119; mul.ftz.f32 %f2372, %f2363, %f5117; sub.ftz.f32 %f5121, %f2371, %f2372; mul.ftz.f32 %f2373, %f2362, %f5117; mul.ftz.f32 %f2374, %f2361, %f5118; sub.ftz.f32 %f5122, %f2373, %f2374; mul.ftz.f32 %f2375, %f5121, %f5121; fma.rn.ftz.f32 %f2376, %f5120, %f5120, %f2375; fma.rn.ftz.f32 %f2377, %f5122, %f5122, %f2376; sqrt.approx.ftz.f32 %f451, %f2377; setp.neu.ftz.f32 %p175, %f451, 0f00000000; @%p175 bra $L__BB104_207; bra.uni $L__BB104_208; $L__BB104_207: rcp.approx.ftz.f32 %f2378, %f451; mul.ftz.f32 %f5122, %f2378, %f5122; mul.ftz.f32 %f5121, %f2378, %f5121; mul.ftz.f32 %f5120, %f2378, %f5120; $L__BB104_208: ld.local.f32 %f2379, [%rd1+-224]; ld.local.f32 %f2380, [%rd1+-220]; mul.ftz.f32 %f2381, %f5118, %f2380; fma.rn.ftz.f32 %f2382, %f5117, %f2379, %f2381; ld.local.f32 %f2383, [%rd1+-216]; fma.rn.ftz.f32 %f461, %f5119, %f2383, %f2382; mul.ftz.f32 %f2384, %f2362, %f2380; fma.rn.ftz.f32 %f2385, %f2361, %f2379, %f2384; fma.rn.ftz.f32 %f2386, %f2363, %f2383, %f2385; mul.ftz.f32 %f2387, %f5121, %f2380; fma.rn.ftz.f32 %f2388, %f5120, %f2379, %f2387; fma.rn.ftz.f32 %f2389, %f5122, %f2383, %f2388; mul.ftz.f32 %f2390, %f1405, %f5118; fma.rn.ftz.f32 %f2391, %f1404, %f5117, %f2390; fma.rn.ftz.f32 %f462, %f1406, %f5119, %f2391; mul.ftz.f32 %f2392, %f1405, %f2362; fma.rn.ftz.f32 %f2393, %f1404, %f2361, %f2392; fma.rn.ftz.f32 %f463, %f1406, %f2363, %f2393; mul.ftz.f32 %f2394, %f1405, %f5121; fma.rn.ftz.f32 %f2395, %f1404, %f5120, %f2394; fma.rn.ftz.f32 %f464, %f1406, %f5122, %f2395; mul.ftz.f32 %f465, %f461, %f461; mov.f32 %f5124, 0f3F800000; sub.ftz.f32 %f2397, %f5124, %f465; mov.f32 %f2398, 0f00000000; max.ftz.f32 %f2399, %f2397, %f2398; sqrt.approx.ftz.f32 %f466, %f2399; abs.ftz.f32 %f467, %f2386; abs.ftz.f32 %f468, %f2389; setp.eq.ftz.f32 %p176, %f467, 0f00000000; setp.eq.ftz.f32 %p177, %f468, 0f00000000; and.pred %p178, %p176, %p177; mov.b32 %r15, %f2386; mov.b32 %r42, %f2389; and.b32 %r16, %r42, -2147483648; @%p178 bra $L__BB104_212; bra.uni $L__BB104_209; $L__BB104_212: shr.s32 %r47, %r15, 31; and.b32 %r48, %r47, 1078530011; or.b32 %r49, %r16, %r48; mov.b32 %f5123, %r49; bra.uni $L__BB104_213; $L__BB104_504: ld.local.f32 %f4583, [%rd3+8]; ld.local.f32 %f4584, [%rd3+12]; mul.ftz.f32 %f4585, %f1405, %f4584; fma.rn.ftz.f32 %f4586, %f1404, %f4583, %f4585; ld.local.f32 %f4587, [%rd3+16]; fma.rn.ftz.f32 %f4588, %f1406, %f4587, %f4586; neg.ftz.f32 %f4589, %f4588; mov.f32 %f4590, 0f00000000; max.ftz.f32 %f4591, %f4589, %f4590; mul.ftz.f32 %f5288, %f4591, 0f3EA2F983; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; bra.uni $L__BB104_519; $L__BB104_29: ld.local.f32 %f1111, [%rd3+8]; ld.local.f32 %f1112, [%rd3+12]; mul.ftz.f32 %f3926, %f1106, %f1112; fma.rn.ftz.f32 %f3927, %f1105, %f1111, %f3926; ld.local.f32 %f1113, [%rd3+16]; fma.rn.ftz.f32 %f1114, %f1107, %f1113, %f3927; mul.ftz.f32 %f3928, %f1405, %f1112; fma.rn.ftz.f32 %f3929, %f1404, %f1111, %f3928; fma.rn.ftz.f32 %f1115, %f1406, %f1113, %f3929; setp.lt.ftz.f32 %p417, %f1115, 0f00000000; @%p417 bra $L__BB104_457; bra.uni $L__BB104_30; $L__BB104_457: setp.le.ftz.f32 %p426, %f1114, 0f00000000; setp.ge.ftz.f32 %p427, %f1115, 0f00000000; or.pred %p428, %p426, %p427; mov.f32 %f5288, %f5291; @%p428 bra $L__BB104_463; ld.local.f32 %f1151, [%rd3+28]; fma.rn.ftz.f32 %f4068, %f1404, %f1151, %f1105; fma.rn.ftz.f32 %f4069, %f1405, %f1151, %f1106; fma.rn.ftz.f32 %f4070, %f1406, %f1151, %f1107; neg.ftz.f32 %f4071, %f4068; neg.ftz.f32 %f4072, %f4069; neg.ftz.f32 %f4073, %f4070; mul.ftz.f32 %f4074, %f4069, %f4069; fma.rn.ftz.f32 %f4075, %f4068, %f4068, %f4074; fma.rn.ftz.f32 %f1152, %f4070, %f4070, %f4075; sqrt.approx.ftz.f32 %f4076, %f1152; div.approx.ftz.f32 %f4077, %f4071, %f4076; div.approx.ftz.f32 %f4078, %f4072, %f4076; div.approx.ftz.f32 %f4079, %f4073, %f4076; mul.ftz.f32 %f4080, %f1106, %f4078; fma.rn.ftz.f32 %f4081, %f1105, %f4077, %f4080; fma.rn.ftz.f32 %f1153, %f1107, %f4079, %f4081; mul.ftz.f32 %f4082, %f1405, %f4078; fma.rn.ftz.f32 %f4083, %f1404, %f4077, %f4082; fma.rn.ftz.f32 %f1154, %f1406, %f4079, %f4083; mul.ftz.f32 %f4084, %f1112, %f4078; fma.rn.ftz.f32 %f4085, %f1111, %f4077, %f4084; fma.rn.ftz.f32 %f4086, %f1113, %f4079, %f4085; mov.f32 %f5252, 0f3F800000; min.ftz.f32 %f4087, %f4086, %f5252; mul.ftz.f32 %f4088, %f4087, %f4087; sub.ftz.f32 %f4089, %f5252, %f4088; div.approx.ftz.f32 %f4090, %f4089, %f4088; mul.ftz.f32 %f4091, %f4088, %f4088; neg.ftz.f32 %f4092, %f4090; div.approx.ftz.f32 %f4093, %f4092, %f1110; mul.ftz.f32 %f4094, %f4093, 0f3FB8AA3B; ex2.approx.ftz.f32 %f4095, %f4094; mul.ftz.f32 %f4096, %f1110, 0f40490FDB; mul.ftz.f32 %f4097, %f4096, %f4091; div.approx.ftz.f32 %f1155, %f4095, %f4097; mul.ftz.f32 %f4098, %f1114, %f1114; sub.ftz.f32 %f4099, %f5252, %f4098; div.approx.ftz.f32 %f4100, %f4099, %f4098; mov.f32 %f4101, 0f00000000; max.ftz.f32 %f4102, %f4100, %f4101; sqrt.approx.ftz.f32 %f4103, %f4102; mul.ftz.f32 %f1156, %f3918, %f4103; setp.lt.ftz.f32 %p429, %f1156, 0f3F200000; mov.f32 %f5251, %f5252; @%p429 bra $L__BB104_460; rcp.approx.ftz.f32 %f4104, %f1156; fma.rn.ftz.f32 %f4105, %f4104, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4106, %f4104, %f4105; fma.rn.ftz.f32 %f4107, %f4104, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4108, %f4104, %f4107, 0f3F800000; div.approx.ftz.f32 %f5251, %f4106, %f4108; $L__BB104_460: mul.ftz.f32 %f4110, %f1115, %f1115; sub.ftz.f32 %f4111, %f5252, %f4110; div.approx.ftz.f32 %f4112, %f4111, %f4110; max.ftz.f32 %f4114, %f4112, %f4101; sqrt.approx.ftz.f32 %f4115, %f4114; mul.ftz.f32 %f1159, %f3918, %f4115; setp.lt.ftz.f32 %p430, %f1159, 0f3F200000; @%p430 bra $L__BB104_462; rcp.approx.ftz.f32 %f4116, %f1159; fma.rn.ftz.f32 %f4117, %f4116, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4118, %f4116, %f4117; fma.rn.ftz.f32 %f4119, %f4116, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4120, %f4116, %f4119, 0f3F800000; div.approx.ftz.f32 %f5252, %f4118, %f4120; $L__BB104_462: mul.ftz.f32 %f4121, %f1151, %f1151; mul.ftz.f32 %f4122, %f4121, %f1155; mul.ftz.f32 %f4123, %f1114, %f1152; div.approx.ftz.f32 %f4124, %f4122, %f4123; mul.ftz.f32 %f4125, %f1154, %f1153; abs.ftz.f32 %f4126, %f4125; mul.ftz.f32 %f4127, %f5251, %f5252; mul.ftz.f32 %f4128, %f4127, %f4126; mul.ftz.f32 %f4129, %f5251, %f4126; mul.ftz.f32 %f5291, %f4124, %f4129; mul.ftz.f32 %f5288, %f4124, %f4128; bra.uni $L__BB104_463; $L__BB104_518: ld.local.f32 %f4744, [%rd3+8]; ld.local.f32 %f4745, [%rd3+12]; mul.ftz.f32 %f4746, %f1405, %f4745; fma.rn.ftz.f32 %f4747, %f1404, %f4744, %f4746; ld.local.f32 %f4748, [%rd3+16]; fma.rn.ftz.f32 %f4749, %f1406, %f4748, %f4747; mov.f32 %f4750, 0f00000000; max.ftz.f32 %f4751, %f4749, %f4750; mul.ftz.f32 %f5288, %f4751, 0f3EA2F983; mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; bra.uni $L__BB104_519; $L__BB104_70: mul.ftz.f32 %f1521, %f17, %f23; fma.rn.ftz.f32 %f1522, %f16, %f22, %f1521; fma.rn.ftz.f32 %f1523, %f18, %f24, %f1522; mul.ftz.f32 %f1524, %f35, %f1523; mul.ftz.f32 %f1525, %f1524, %f34; mul.ftz.f32 %f1526, %f20, %f1525; mul.ftz.f32 %f1527, %f20, %f1526; mul.ftz.f32 %f1528, %f21, %f42; mul.ftz.f32 %f1529, %f21, %f1528; max.ftz.f32 %f1531, %f1529, %f1505; div.approx.ftz.f32 %f43, %f1527, %f1531; setp.geu.ftz.f32 %p49, %f20, 0f3F800000; mov.f32 %f5004, %f20; @%p49 bra $L__BB104_72; rcp.approx.ftz.f32 %f5004, %f20; $L__BB104_72: mov.f32 %f1532, 0f3F800000; max.ftz.f32 %f1533, %f5004, %f1532; mov.f32 %f1534, 0f40400000; min.ftz.f32 %f1535, %f1533, %f1534; add.ftz.f32 %f1536, %f1535, 0fBF2814F9; mul.ftz.f32 %f1537, %f1536, 0fBF7A7547; mul.ftz.f32 %f1538, %f1536, %f1537; mul.ftz.f32 %f1539, %f1538, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1540, %f1539; mul.ftz.f32 %f1541, %f1535, 0fBD0A43BB; fma.rn.ftz.f32 %f1542, %f1540, 0f3D435661, %f1541; add.ftz.f32 %f1543, %f1542, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f1544, %f1421; fma.rn.ftz.f32 %f1545, %f1544, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f1546, %f1544, %f1545, 0f3D4D5C09; fma.rn.ftz.f32 %f1547, %f1544, %f1546, 0fBB86D80B; fma.rn.ftz.f32 %f1548, %f1544, %f1547, 0f3F80029F; fma.rn.ftz.f32 %f1549, %f1535, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f1550, %f1535, %f1549, 0fC2FFCFDF; fma.rn.ftz.f32 %f1551, %f1535, %f1550, 0f43927AA0; fma.rn.ftz.f32 %f1552, %f1535, %f1551, 0fC38FF917; mul.ftz.f32 %f1553, %f1535, %f1535; mov.f32 %f1554, 0f4347CD91; div.approx.ftz.f32 %f1555, %f1554, %f1553; add.ftz.f32 %f1556, %f1555, %f1552; mul.ftz.f32 %f1557, %f1535, %f1553; mov.f32 %f1558, 0f42CB5604; div.approx.ftz.f32 %f1559, %f1558, %f1557; sub.ftz.f32 %f1560, %f1556, %f1559; fma.rn.ftz.f32 %f1561, %f1544, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f1562, %f1544, %f1561, 0f41B5F319; fma.rn.ftz.f32 %f1563, %f1544, %f1562, 0fC05A14A5; fma.rn.ftz.f32 %f1564, %f1544, %f1563, 0f3DC9FFD6; fma.rn.ftz.f32 %f1565, %f1544, %f1564, 0f3BA1B61E; mul.ftz.f32 %f1566, %f1565, %f1560; fma.rn.ftz.f32 %f1567, %f1566, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f1568, %f1532, %f1543; sub.ftz.f32 %f1569, %f1532, %f1548; mul.ftz.f32 %f1570, %f1568, %f1569; fma.rn.ftz.f32 %f1571, %f1570, 0fC11520C5, %f1567; cvt.ftz.sat.f32.f32 %f1572, %f1571; abs.ftz.f32 %f1573, %f43; sub.ftz.f32 %f1574, %f1532, %f1572; mul.ftz.f32 %f1575, %f41, %f1574; fma.rn.ftz.f32 %f1576, %f1573, %f1572, %f1575; sub.ftz.f32 %f1577, %f1532, %f5003; mul.ftz.f32 %f5291, %f1577, %f1576; $L__BB104_74: ld.local.u64 %rd4, [%rd3+36]; ld.f32 %f49, [%rd4]; ld.f32 %f50, [%rd4+4]; ld.f32 %f51, [%rd4+8]; setp.lt.ftz.f32 %p50, %f18, 0f00000000; @%p50 bra $L__BB104_76; bra.uni $L__BB104_75; $L__BB104_76: mov.f32 %f5006, %f49; mov.f32 %f5007, %f50; mov.f32 %f5008, %f51; bra.uni $L__BB104_77; $L__BB104_209: setp.eq.ftz.f32 %p179, %f467, 0f7F800000; setp.eq.ftz.f32 %p180, %f468, 0f7F800000; and.pred %p181, %p179, %p180; @%p181 bra $L__BB104_211; bra.uni $L__BB104_210; $L__BB104_211: setp.lt.s32 %p185, %r15, 0; selp.b32 %r45, 1075235812, 1061752795, %p185; or.b32 %r46, %r16, %r45; mov.b32 %f5123, %r46; bra.uni $L__BB104_213; $L__BB104_30: setp.leu.ftz.f32 %p418, %f1114, 0f00000000; setp.leu.ftz.f32 %p419, %f1115, 0f00000000; or.pred %p420, %p419, %p418; mov.f32 %f5288, %f5291; @%p420 bra $L__BB104_463; add.ftz.f32 %f3934, %f1404, %f1105; add.ftz.f32 %f3935, %f1405, %f1106; mul.ftz.f32 %f3936, %f3935, %f3935; fma.rn.ftz.f32 %f3937, %f3934, %f3934, %f3936; add.ftz.f32 %f3938, %f1406, %f1107; fma.rn.ftz.f32 %f3939, %f3938, %f3938, %f3937; sqrt.approx.ftz.f32 %f3940, %f3939; div.approx.ftz.f32 %f1116, %f3934, %f3940; div.approx.ftz.f32 %f1117, %f3935, %f3940; div.approx.ftz.f32 %f1118, %f3938, %f3940; mul.ftz.f32 %f3941, %f1112, %f1117; fma.rn.ftz.f32 %f3942, %f1111, %f1116, %f3941; fma.rn.ftz.f32 %f1119, %f1113, %f1118, %f3942; mul.ftz.f32 %f1120, %f1119, %f1119; mul.ftz.f32 %f3943, %f1120, %f1120; mul.ftz.f32 %f3944, %f1110, 0f40490FDB; mul.ftz.f32 %f1121, %f3944, %f3943; mul.ftz.f32 %f1122, %f1114, %f1114; mov.f32 %f5250, 0f3F800000; sub.ftz.f32 %f1123, %f5250, %f1122; setp.eq.ftz.f32 %p421, %f3918, %f3919; @%p421 bra $L__BB104_452; bra.uni $L__BB104_32; $L__BB104_452: mov.f32 %f5250, 0f3F800000; sub.ftz.f32 %f4033, %f5250, %f1120; div.approx.ftz.f32 %f4034, %f4033, %f1120; neg.ftz.f32 %f4035, %f4034; div.approx.ftz.f32 %f4036, %f4035, %f1110; mul.ftz.f32 %f4037, %f4036, 0f3FB8AA3B; ex2.approx.ftz.f32 %f4038, %f4037; div.approx.ftz.f32 %f5248, %f4038, %f1121; div.approx.ftz.f32 %f4039, %f1123, %f1122; mov.f32 %f4040, 0f00000000; max.ftz.f32 %f4041, %f4039, %f4040; sqrt.approx.ftz.f32 %f4042, %f4041; mul.ftz.f32 %f1139, %f3918, %f4042; setp.lt.ftz.f32 %p424, %f1139, 0f3F200000; mov.f32 %f5249, %f5250; @%p424 bra $L__BB104_454; rcp.approx.ftz.f32 %f4043, %f1139; fma.rn.ftz.f32 %f4044, %f4043, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4045, %f4043, %f4044; fma.rn.ftz.f32 %f4046, %f4043, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4047, %f4043, %f4046, 0f3F800000; div.approx.ftz.f32 %f5249, %f4045, %f4047; $L__BB104_454: mul.ftz.f32 %f4049, %f1115, %f1115; sub.ftz.f32 %f4050, %f5250, %f4049; div.approx.ftz.f32 %f4051, %f4050, %f4049; max.ftz.f32 %f4053, %f4051, %f4040; sqrt.approx.ftz.f32 %f4054, %f4053; mul.ftz.f32 %f1142, %f3918, %f4054; setp.lt.ftz.f32 %p425, %f1142, 0f3F200000; @%p425 bra $L__BB104_456; rcp.approx.ftz.f32 %f4055, %f1142; fma.rn.ftz.f32 %f4056, %f4055, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4057, %f4055, %f4056; fma.rn.ftz.f32 %f4058, %f4055, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4059, %f4055, %f4058, 0f3F800000; div.approx.ftz.f32 %f5250, %f4057, %f4059; bra.uni $L__BB104_456; $L__BB104_75: ld.f32 %f5008, [%rd4+20]; ld.f32 %f5007, [%rd4+16]; ld.f32 %f5006, [%rd4+12]; $L__BB104_77: mul.ftz.f32 %f4947, %f5014, %f18; setp.lt.ftz.f32 %p51, %f4947, 0f00000000; @%p51 bra $L__BB104_80; bra.uni $L__BB104_78; $L__BB104_80: neg.ftz.f32 %f61, %f18; setp.leu.ftz.f32 %p53, %f5014, %f61; mov.f32 %f5009, %f16; mov.f32 %f5010, %f17; mov.f32 %f5011, %f18; mov.u16 %rs31, %rs23; @%p53 bra $L__BB104_82; neg.ftz.f32 %f5011, %f5014; neg.ftz.f32 %f5010, %f5013; neg.ftz.f32 %f5009, %f5012; neg.ftz.f32 %f5012, %f16; neg.ftz.f32 %f5013, %f17; mov.f32 %f5014, %f61; mov.u16 %rs31, %rs22; bra.uni $L__BB104_82; $L__BB104_78: setp.geu.ftz.f32 %p52, %f18, %f5014; mov.f32 %f5009, %f16; mov.f32 %f5010, %f17; mov.f32 %f5011, %f18; mov.u16 %rs31, %rs23; @%p52 bra $L__BB104_82; mov.f32 %f5009, %f5012; mov.f32 %f5010, %f5013; mov.f32 %f5011, %f5014; mov.f32 %f5012, %f16; mov.f32 %f5013, %f17; mov.f32 %f5014, %f18; mov.u16 %rs31, %rs22; $L__BB104_82: setp.lt.ftz.f32 %p54, %f5011, 0f3727C5AC; and.pred %p55, %p54, %p1; setp.lt.ftz.f32 %p56, %f5014, 0f3727C5AC; or.pred %p57, %p55, %p56; setp.gt.ftz.f32 %p58, %f5011, 0fB727C5AC; and.pred %p60, %p50, %p58; or.pred %p61, %p60, %p57; mov.f32 %f5290, 0f00000000; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p61 bra $L__BB104_519; neg.ftz.f32 %f70, %f5012; neg.ftz.f32 %f71, %f5013; neg.ftz.f32 %f72, %f5014; setp.lt.ftz.f32 %p62, %f5014, 0fBF7FF972; mov.f32 %f5106, 0f00000000; @%p62 bra $L__BB104_86; setp.gt.ftz.f32 %p63, %f5014, 0f3F7FF972; mov.f32 %f5106, 0fBF7FF972; @%p63 bra $L__BB104_86; mul.ftz.f32 %f1605, %f5014, %f5014; max.ftz.f32 %f1607, %f1605, %f1505; rcp.approx.ftz.f32 %f1608, %f1607; mul.ftz.f32 %f1609, %f1422, %f71; mul.ftz.f32 %f1610, %f1421, %f70; mul.ftz.f32 %f1611, %f1609, %f1609; fma.rn.ftz.f32 %f1612, %f1610, %f1610, %f1611; fma.rn.ftz.f32 %f1613, %f1612, %f1608, 0f3F800000; sqrt.approx.ftz.f32 %f1614, %f1613; neg.ftz.f32 %f1615, %f1614; setp.ge.ftz.f32 %p64, %f5014, 0f80000000; selp.f32 %f1616, %f1615, %f1614, %p64; add.ftz.f32 %f1617, %f1616, 0fBF800000; mul.ftz.f32 %f5106, %f1617, 0f3F000000; $L__BB104_86: mov.f32 %f5016, %f5009; mov.f32 %f5017, %f5010; mov.f32 %f5018, %f5011; @%p50 bra $L__BB104_87; bra.uni $L__BB104_88; $L__BB104_87: neg.ftz.f32 %f5018, %f5011; neg.ftz.f32 %f5017, %f5010; neg.ftz.f32 %f5016, %f5009; $L__BB104_88: setp.gt.ftz.f32 %p66, %f5018, 0f3F7FF972; mov.f32 %f5023, 0f00000000; mov.f32 %f5019, %f5023; @%p66 bra $L__BB104_91; setp.lt.ftz.f32 %p67, %f5018, 0fBF7FF972; mov.f32 %f5019, 0fBF7FF972; @%p67 bra $L__BB104_91; mul.ftz.f32 %f1620, %f5018, %f5018; max.ftz.f32 %f1622, %f1620, %f1505; rcp.approx.ftz.f32 %f1623, %f1622; mul.ftz.f32 %f1624, %f1422, %f5017; mul.ftz.f32 %f1625, %f1421, %f5016; mul.ftz.f32 %f1626, %f1624, %f1624; fma.rn.ftz.f32 %f1627, %f1625, %f1625, %f1626; fma.rn.ftz.f32 %f1628, %f1627, %f1623, 0f3F800000; sqrt.approx.ftz.f32 %f1629, %f1628; neg.ftz.f32 %f1630, %f1629; setp.le.ftz.f32 %p68, %f5018, 0f00000000; selp.f32 %f1631, %f1630, %f1629, %p68; add.ftz.f32 %f1632, %f1631, 0fBF800000; mul.ftz.f32 %f5019, %f1632, 0f3F000000; $L__BB104_91: add.ftz.f32 %f1636, %f5009, %f5012; add.ftz.f32 %f1637, %f5010, %f5013; mul.ftz.f32 %f1638, %f1637, %f1637; fma.rn.ftz.f32 %f1639, %f1636, %f1636, %f1638; add.ftz.f32 %f1640, %f5011, %f5014; fma.rn.ftz.f32 %f1641, %f1640, %f1640, %f1639; sqrt.approx.ftz.f32 %f1642, %f1641; div.approx.ftz.f32 %f87, %f1636, %f1642; div.approx.ftz.f32 %f88, %f1637, %f1642; div.approx.ftz.f32 %f89, %f1640, %f1642; @%p62 bra $L__BB104_101; mul.ftz.f32 %f1643, %f5106, %f72; setp.gt.ftz.f32 %p71, %f5014, 0f3F7FF972; selp.f32 %f90, 0f3F800000, %f1643, %p71; @%p50 bra $L__BB104_97; bra.uni $L__BB104_93; $L__BB104_97: fma.rn.ftz.f32 %f1678, %f20, %f5009, %f5012; fma.rn.ftz.f32 %f1679, %f20, %f5010, %f5013; fma.rn.ftz.f32 %f1680, %f20, %f5011, %f5014; mul.ftz.f32 %f1681, %f1679, %f1679; fma.rn.ftz.f32 %f1682, %f1678, %f1678, %f1681; fma.rn.ftz.f32 %f1683, %f1680, %f1680, %f1682; sqrt.approx.ftz.f32 %f1684, %f1683; div.approx.ftz.f32 %f1685, %f1678, %f1684; div.approx.ftz.f32 %f1686, %f1679, %f1684; div.approx.ftz.f32 %f1687, %f1680, %f1684; setp.lt.ftz.f32 %p74, %f1687, 0f00000000; neg.ftz.f32 %f1688, %f1685; neg.ftz.f32 %f1689, %f1686; neg.ftz.f32 %f1690, %f1687; selp.f32 %f1691, %f1688, %f1685, %p74; selp.f32 %f1692, %f1689, %f1686, %p74; selp.f32 %f97, %f1690, %f1687, %p74; mul.ftz.f32 %f1693, %f5013, %f1692; fma.rn.ftz.f32 %f1694, %f5012, %f1691, %f1693; fma.rn.ftz.f32 %f98, %f5014, %f97, %f1694; mul.ftz.f32 %f1695, %f5010, %f1692; fma.rn.ftz.f32 %f1696, %f5009, %f1691, %f1695; fma.rn.ftz.f32 %f99, %f5011, %f97, %f1696; setp.lt.ftz.f32 %p75, %f98, 0f00000000; @%p75 bra $L__BB104_101; fma.rn.ftz.f32 %f100, %f20, %f99, %f98; abs.ftz.f32 %f101, %f98; fma.rn.ftz.f32 %f102, %f101, %f101, %f37; setp.leu.ftz.f32 %p76, %f102, 0f00000000; mov.f32 %f1697, 0f3F800000; mov.f32 %f5021, %f1697; @%p76 bra $L__BB104_100; sqrt.approx.ftz.f32 %f1698, %f102; sub.ftz.f32 %f1699, %f1698, %f101; add.ftz.f32 %f1700, %f101, %f1698; div.approx.ftz.f32 %f1701, %f1699, %f1700; fma.rn.ftz.f32 %f1702, %f101, %f1700, 0fBF800000; fma.rn.ftz.f32 %f1703, %f101, %f1699, 0f3F800000; div.approx.ftz.f32 %f1704, %f1702, %f1703; mul.ftz.f32 %f1705, %f1701, 0f3F000000; mul.ftz.f32 %f1706, %f1701, %f1705; fma.rn.ftz.f32 %f1707, %f1704, %f1704, 0f3F800000; mul.ftz.f32 %f5021, %f1706, %f1707; $L__BB104_100: mov.f32 %f1708, 0f00000000; max.ftz.f32 %f1709, %f1708, %f98; sub.ftz.f32 %f1711, %f1697, %f5021; mul.ftz.f32 %f1712, %f1711, %f1709; neg.ftz.f32 %f1713, %f99; max.ftz.f32 %f1714, %f1708, %f1713; mul.ftz.f32 %f1715, %f1712, %f1714; mul.ftz.f32 %f1716, %f97, %f97; sub.ftz.f32 %f1717, %f1697, %f1716; fma.rn.ftz.f32 %f1718, %f33, %f1716, %f1717; mul.ftz.f32 %f1719, %f1718, 0f40490FDB; mul.ftz.f32 %f1720, %f1718, %f1719; max.ftz.f32 %f1722, %f1720, %f1505; div.approx.ftz.f32 %f1723, %f33, %f1722; mul.ftz.f32 %f1724, %f1715, %f1723; mul.ftz.f32 %f1725, %f90, %f100; mul.ftz.f32 %f1726, %f100, %f1725; div.approx.ftz.f32 %f5023, %f1724, %f1726; bra.uni $L__BB104_101; $L__BB104_210: setp.lt.s32 %p182, %r15, 0; min.ftz.f32 %f2400, %f468, %f467; max.ftz.f32 %f2401, %f468, %f467; div.full.ftz.f32 %f2402, %f2400, %f2401; mul.rn.ftz.f32 %f2403, %f2402, %f2402; mov.f32 %f2404, 0fC0B59883; mov.f32 %f2405, 0fBF52C7EA; fma.rn.ftz.f32 %f2406, %f2403, %f2405, %f2404; mov.f32 %f2407, 0fC0D21907; fma.rn.ftz.f32 %f2408, %f2406, %f2403, %f2407; mul.ftz.f32 %f2409, %f2403, %f2408; mul.ftz.f32 %f2410, %f2402, %f2409; add.ftz.f32 %f2411, %f2403, 0f41355DC0; mov.f32 %f2412, 0f41E6BD60; fma.rn.ftz.f32 %f2413, %f2411, %f2403, %f2412; mov.f32 %f2414, 0f419D92C8; fma.rn.ftz.f32 %f2415, %f2413, %f2403, %f2414; rcp.approx.ftz.f32 %f2416, %f2415; fma.rn.ftz.f32 %f2417, %f2410, %f2416, %f2402; mov.f32 %f2418, 0f3FC90FDB; sub.ftz.f32 %f2419, %f2418, %f2417; setp.gt.ftz.f32 %p183, %f468, %f467; selp.f32 %f2420, %f2419, %f2417, %p183; mov.f32 %f2421, 0f40490FDB; sub.ftz.f32 %f2422, %f2421, %f2420; selp.f32 %f2423, %f2422, %f2420, %p182; mov.b32 %r43, %f2423; or.b32 %r44, %r16, %r43; mov.b32 %f2424, %r44; add.ftz.f32 %f2425, %f467, %f468; setp.le.ftz.f32 %p184, %f2425, 0f7F800000; selp.f32 %f5123, %f2424, %f2425, %p184; $L__BB104_213: ld.local.f32 %f2427, [%rd3+44]; div.approx.ftz.f32 %f2428, %f461, %f2427; mul.ftz.f32 %f2429, %f2428, %f2428; sub.ftz.f32 %f2430, %f5124, %f2429; max.ftz.f32 %f2432, %f2430, %f2398; sqrt.approx.ftz.f32 %f2433, %f2432; ld.f32 %f2434, [%rd5+12]; mul.ftz.f32 %f2435, %f2434, %f2434; sub.ftz.f32 %f2436, %f5124, %f2435; max.ftz.f32 %f2437, %f2436, %f2398; sqrt.approx.ftz.f32 %f2438, %f2437; mov.f32 %f2439, 0fBF800000; max.ftz.f32 %f2440, %f2434, %f2439; min.ftz.f32 %f2441, %f2440, %f5124; abs.ftz.f32 %f2442, %f2441; sub.ftz.f32 %f2443, %f5124, %f2442; mul.ftz.f32 %f2444, %f2443, 0f3F000000; sqrt.approx.ftz.f32 %f2445, %f2444; setp.gt.ftz.f32 %p186, %f2442, 0f3F11EB85; selp.f32 %f2446, %f2445, %f2442, %p186; mul.ftz.f32 %f2447, %f2446, %f2446; mov.f32 %f2448, 0f3C94D2E9; mov.f32 %f2449, 0f3D53F941; fma.rn.ftz.f32 %f2450, %f2449, %f2447, %f2448; mov.f32 %f2451, 0f3D3F841F; fma.rn.ftz.f32 %f2452, %f2450, %f2447, %f2451; mov.f32 %f2453, 0f3D994929; fma.rn.ftz.f32 %f2454, %f2452, %f2447, %f2453; mov.f32 %f2455, 0f3E2AAB94; fma.rn.ftz.f32 %f2456, %f2454, %f2447, %f2455; mul.ftz.f32 %f2457, %f2447, %f2456; fma.rn.ftz.f32 %f2458, %f2457, %f2446, %f2446; mov.f32 %f2459, 0f3FC90FDB; mov.f32 %f2460, 0fC0000000; fma.rn.ftz.f32 %f2461, %f2460, %f2458, %f2459; selp.f32 %f2462, %f2461, %f2458, %p186; setp.le.ftz.f32 %p187, %f2462, 0f7F800000; mov.b32 %r50, %f2462; mov.b32 %r51, %f2441; and.b32 %r52, %r51, -2147483648; or.b32 %r53, %r52, %r50; mov.b32 %f2463, %r53; selp.f32 %f473, %f2463, %f2462, %p187; mul.ftz.f32 %f2464, %f466, %f2434; mul.ftz.f32 %f2465, %f2427, %f2427; sub.ftz.f32 %f2466, %f2465, %f465; sqrt.approx.ftz.f32 %f2467, %f2466; div.approx.ftz.f32 %f2468, %f2464, %f2467; mul.ftz.f32 %f2469, %f2468, %f2468; sub.ftz.f32 %f2470, %f5124, %f2469; max.ftz.f32 %f2471, %f2470, %f2398; sqrt.approx.ftz.f32 %f2472, %f2471; max.ftz.f32 %f2473, %f2468, %f2439; min.ftz.f32 %f2474, %f2473, %f5124; abs.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f5124, %f2475; mul.ftz.f32 %f2477, %f2476, 0f3F000000; sqrt.approx.ftz.f32 %f2478, %f2477; setp.gt.ftz.f32 %p188, %f2475, 0f3F11EB85; selp.f32 %f2479, %f2478, %f2475, %p188; mul.ftz.f32 %f2480, %f2479, %f2479; fma.rn.ftz.f32 %f2481, %f2449, %f2480, %f2448; fma.rn.ftz.f32 %f2482, %f2481, %f2480, %f2451; fma.rn.ftz.f32 %f2483, %f2482, %f2480, %f2453; fma.rn.ftz.f32 %f2484, %f2483, %f2480, %f2455; mul.ftz.f32 %f2485, %f2480, %f2484; fma.rn.ftz.f32 %f2486, %f2485, %f2479, %f2479; fma.rn.ftz.f32 %f2487, %f2460, %f2486, %f2459; selp.f32 %f2488, %f2487, %f2486, %p188; setp.le.ftz.f32 %p189, %f2488, 0f7F800000; mov.b32 %r54, %f2488; mov.b32 %r55, %f2474; and.b32 %r56, %r55, -2147483648; or.b32 %r57, %r56, %r54; mov.b32 %f2489, %r57; selp.f32 %f474, %f2489, %f2488, %p189; add.ftz.f32 %f2490, %f2472, %f2472; div.approx.ftz.f32 %f2491, %f2490, %f2433; ld.local.f32 %f2492, [%rd3+20]; mul.ftz.f32 %f475, %f2492, %f2491; ld.local.f32 %f2493, [%rd3+24]; mul.ftz.f32 %f476, %f2493, %f2491; ld.local.f32 %f2494, [%rd3+28]; mul.ftz.f32 %f477, %f2494, %f2491; mul.ftz.f32 %f2495, %f466, %f2438; abs.ftz.f32 %f478, %f2495; add.ftz.f32 %f2496, %f2465, 0fBF800000; fma.rn.ftz.f32 %f479, %f478, %f478, %f2496; setp.leu.ftz.f32 %p190, %f479, 0f00000000; @%p190 bra $L__BB104_215; sqrt.approx.ftz.f32 %f2497, %f479; sub.ftz.f32 %f2498, %f2497, %f478; add.ftz.f32 %f2499, %f478, %f2497; div.approx.ftz.f32 %f2500, %f2498, %f2499; fma.rn.ftz.f32 %f2501, %f478, %f2499, 0fBF800000; fma.rn.ftz.f32 %f2502, %f478, %f2498, 0f3F800000; div.approx.ftz.f32 %f2503, %f2501, %f2502; mul.ftz.f32 %f2504, %f2500, 0f3F000000; mul.ftz.f32 %f2505, %f2500, %f2504; fma.rn.ftz.f32 %f2506, %f2503, %f2503, 0f3F800000; mul.ftz.f32 %f5124, %f2505, %f2506; $L__BB104_215: mov.f32 %f5125, 0f00000000; mov.f32 %f4958, 0f3F800000; mul.ftz.f32 %f2508, %f475, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f476, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2511, %f2510; mul.ftz.f32 %f2512, %f477, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2513, %f2512; sub.ftz.f32 %f2515, %f4958, %f5124; mul.ftz.f32 %f2516, %f2515, %f2515; mul.ftz.f32 %f482, %f2509, %f2516; mul.ftz.f32 %f483, %f2511, %f2516; mul.ftz.f32 %f484, %f2513, %f2516; mul.ftz.f32 %f2517, %f483, %f1410; fma.rn.ftz.f32 %f2518, %f482, %f1409, %f2517; fma.rn.ftz.f32 %f485, %f484, %f1411, %f2518; mul.ftz.f32 %f486, %f2509, %f5124; mul.ftz.f32 %f487, %f486, %f482; mul.ftz.f32 %f488, %f2511, %f5124; mul.ftz.f32 %f489, %f488, %f483; mul.ftz.f32 %f490, %f2513, %f5124; mul.ftz.f32 %f491, %f490, %f484; mul.ftz.f32 %f2519, %f489, %f1410; fma.rn.ftz.f32 %f2520, %f487, %f1409, %f2519; fma.rn.ftz.f32 %f492, %f491, %f1411, %f2520; sub.ftz.f32 %f493, %f4958, %f486; sub.ftz.f32 %f494, %f4958, %f488; sub.ftz.f32 %f495, %f4958, %f490; setp.eq.ftz.f32 %p191, %f493, 0f00000000; @%p191 bra $L__BB104_217; div.approx.ftz.f32 %f5125, %f486, %f493; $L__BB104_217: mov.f32 %f5126, 0f00000000; setp.eq.ftz.f32 %p192, %f494, 0f00000000; @%p192 bra $L__BB104_219; div.approx.ftz.f32 %f5126, %f488, %f494; $L__BB104_219: mov.f32 %f5127, 0f00000000; setp.eq.ftz.f32 %p193, %f495, 0f00000000; @%p193 bra $L__BB104_221; div.approx.ftz.f32 %f5127, %f490, %f495; $L__BB104_221: mov.f32 %f5128, 0f00000000; mul.ftz.f32 %f502, %f487, %f5125; mul.ftz.f32 %f503, %f489, %f5126; mul.ftz.f32 %f2524, %f503, %f1410; fma.rn.ftz.f32 %f2525, %f502, %f1409, %f2524; mul.ftz.f32 %f504, %f491, %f5127; fma.rn.ftz.f32 %f505, %f504, %f1411, %f2525; add.ftz.f32 %f2526, %f5124, %f485; add.ftz.f32 %f2527, %f2526, %f492; add.ftz.f32 %f506, %f505, %f2527; setp.eq.ftz.f32 %p194, %f506, 0f00000000; @%p194 bra $L__BB104_223; mov.f32 %f2528, 0f3F800000; div.approx.ftz.f32 %f5128, %f2528, %f506; $L__BB104_223: mov.f32 %f4963, 0f00000000; mul.ftz.f32 %f2529, %f462, %f462; mov.f32 %f2530, 0f3F800000; sub.ftz.f32 %f2531, %f2530, %f2529; max.ftz.f32 %f2533, %f2531, %f4963; sqrt.approx.ftz.f32 %f509, %f2533; abs.ftz.f32 %f510, %f463; setp.eq.ftz.f32 %p195, %f510, 0f00000000; abs.ftz.f32 %f511, %f464; setp.eq.ftz.f32 %p196, %f511, 0f00000000; and.pred %p197, %p195, %p196; mov.b32 %r17, %f463; mov.b32 %r58, %f464; and.b32 %r18, %r58, -2147483648; @%p197 bra $L__BB104_227; bra.uni $L__BB104_224; $L__BB104_227: shr.s32 %r63, %r17, 31; and.b32 %r64, %r63, 1078530011; or.b32 %r65, %r18, %r64; mov.b32 %f5129, %r65; bra.uni $L__BB104_228; $L__BB104_224: setp.eq.ftz.f32 %p198, %f510, 0f7F800000; setp.eq.ftz.f32 %p199, %f511, 0f7F800000; and.pred %p200, %p198, %p199; @%p200 bra $L__BB104_226; bra.uni $L__BB104_225; $L__BB104_226: setp.lt.s32 %p204, %r17, 0; selp.b32 %r61, 1075235812, 1061752795, %p204; or.b32 %r62, %r18, %r61; mov.b32 %f5129, %r62; bra.uni $L__BB104_228; $L__BB104_225: mov.f32 %f4964, 0f3FC90FDB; setp.lt.s32 %p201, %r17, 0; min.ftz.f32 %f2534, %f511, %f510; max.ftz.f32 %f2535, %f511, %f510; div.full.ftz.f32 %f2536, %f2534, %f2535; mul.rn.ftz.f32 %f2537, %f2536, %f2536; mov.f32 %f2538, 0fC0B59883; mov.f32 %f2539, 0fBF52C7EA; fma.rn.ftz.f32 %f2540, %f2537, %f2539, %f2538; mov.f32 %f2541, 0fC0D21907; fma.rn.ftz.f32 %f2542, %f2540, %f2537, %f2541; mul.ftz.f32 %f2543, %f2537, %f2542; mul.ftz.f32 %f2544, %f2536, %f2543; add.ftz.f32 %f2545, %f2537, 0f41355DC0; mov.f32 %f2546, 0f41E6BD60; fma.rn.ftz.f32 %f2547, %f2545, %f2537, %f2546; mov.f32 %f2548, 0f419D92C8; fma.rn.ftz.f32 %f2549, %f2547, %f2537, %f2548; rcp.approx.ftz.f32 %f2550, %f2549; fma.rn.ftz.f32 %f2551, %f2544, %f2550, %f2536; sub.ftz.f32 %f2553, %f4964, %f2551; setp.gt.ftz.f32 %p202, %f511, %f510; selp.f32 %f2554, %f2553, %f2551, %p202; mov.f32 %f2555, 0f40490FDB; sub.ftz.f32 %f2556, %f2555, %f2554; selp.f32 %f2557, %f2556, %f2554, %p201; mov.b32 %r59, %f2557; or.b32 %r60, %r18, %r59; mov.b32 %f2558, %r60; add.ftz.f32 %f2559, %f510, %f511; setp.le.ftz.f32 %p203, %f2559, 0f7F800000; selp.f32 %f5129, %f2558, %f2559, %p203; $L__BB104_228: sub.ftz.f32 %f516, %f5129, %f5123; ld.local.v2.f32 {%f2560, %f2561}, [%rd3+36]; sin.approx.ftz.f32 %f2563, %f2561; mul.ftz.f32 %f2564, %f2563, %f2563; sub.ftz.f32 %f2566, %f2530, %f2564; mov.f32 %f5136, 0f00000000; max.ftz.f32 %f2568, %f2566, %f5136; sqrt.approx.ftz.f32 %f2569, %f2568; add.ftz.f32 %f2570, %f2563, %f2563; mul.ftz.f32 %f2571, %f2570, %f2569; mul.ftz.f32 %f2572, %f2569, %f2569; sub.ftz.f32 %f2573, %f2572, %f2564; add.ftz.f32 %f2574, %f2571, %f2571; mul.ftz.f32 %f2575, %f2574, %f2573; mul.ftz.f32 %f2576, %f2573, %f2573; mul.ftz.f32 %f2577, %f2571, %f2571; sub.ftz.f32 %f2578, %f2576, %f2577; mul.ftz.f32 %f2579, %f462, %f2573; fma.rn.ftz.f32 %f2580, %f509, %f2571, %f2579; mul.ftz.f32 %f2581, %f509, %f2573; mul.ftz.f32 %f2582, %f462, %f2571; sub.ftz.f32 %f2583, %f2581, %f2582; abs.ftz.f32 %f2584, %f2583; mul.ftz.f32 %f2585, %f462, %f2569; mul.ftz.f32 %f2586, %f509, %f2563; sub.ftz.f32 %f518, %f2585, %f2586; mul.ftz.f32 %f2587, %f509, %f2569; fma.rn.ftz.f32 %f2588, %f462, %f2563, %f2587; abs.ftz.f32 %f519, %f2588; mul.ftz.f32 %f2589, %f462, %f2578; mul.ftz.f32 %f2590, %f509, %f2575; sub.ftz.f32 %f520, %f2589, %f2590; mul.ftz.f32 %f2591, %f509, %f2578; fma.rn.ftz.f32 %f521, %f462, %f2575, %f2591; add.ftz.f32 %f522, %f473, %f473; mov.f32 %f2592, 0f40490FDB; div.approx.ftz.f32 %f523, %f2592, %f2560; mul.ftz.f32 %f524, %f523, 0f3FB8AA3B; ld.local.f32 %f525, [%rd3+48]; rcp.approx.ftz.f32 %f526, %f525; mul.ftz.f32 %f2593, %f466, %f2584; mul.ftz.f32 %f527, %f2593, %f526; mul.ftz.f32 %f2594, %f461, %f2580; mul.ftz.f32 %f528, %f2594, %f526; setp.gtu.ftz.f32 %p205, %f525, 0f3DCCCCCD; @%p205 bra $L__BB104_242; bra.uni $L__BB104_229; $L__BB104_242: mul.ftz.f32 %f551, %f527, %f527; fma.rn.ftz.f32 %f5132, %f551, 0f3E800000, 0f3F800000; mul.ftz.f32 %f553, %f551, %f551; mov.u64 %rd29, 64; cvt.rn.f32.u64 %f2625, %rd29; div.approx.ftz.f32 %f2626, %f553, %f2625; add.ftz.f32 %f554, %f5132, %f2626; setp.eq.ftz.f32 %p215, %f5132, %f554; @%p215 bra $L__BB104_251; mul.ftz.f32 %f555, %f551, %f553; mov.u64 %rd30, 2304; cvt.rn.f32.u64 %f2627, %rd30; div.approx.ftz.f32 %f2628, %f555, %f2627; add.ftz.f32 %f556, %f554, %f2628; setp.eq.ftz.f32 %p216, %f554, %f556; mov.f32 %f5132, %f554; @%p216 bra $L__BB104_251; mul.ftz.f32 %f557, %f551, %f555; mov.u64 %rd31, 147456; cvt.rn.f32.u64 %f2629, %rd31; div.approx.ftz.f32 %f2630, %f557, %f2629; add.ftz.f32 %f558, %f556, %f2630; setp.eq.ftz.f32 %p217, %f556, %f558; mov.f32 %f5132, %f556; @%p217 bra $L__BB104_251; mul.ftz.f32 %f559, %f551, %f557; mov.u64 %rd32, 14745600; cvt.rn.f32.u64 %f2631, %rd32; div.approx.ftz.f32 %f2632, %f559, %f2631; add.ftz.f32 %f560, %f558, %f2632; setp.eq.ftz.f32 %p218, %f558, %f560; mov.f32 %f5132, %f558; @%p218 bra $L__BB104_251; mul.ftz.f32 %f561, %f551, %f559; mov.u64 %rd33, 2123366400; cvt.rn.f32.u64 %f2633, %rd33; div.approx.ftz.f32 %f2634, %f561, %f2633; add.ftz.f32 %f562, %f560, %f2634; setp.eq.ftz.f32 %p219, %f560, %f562; mov.f32 %f5132, %f560; @%p219 bra $L__BB104_251; mul.ftz.f32 %f563, %f551, %f561; mov.u64 %rd34, 416179814400; cvt.rn.f32.u64 %f2635, %rd34; div.approx.ftz.f32 %f2636, %f563, %f2635; add.ftz.f32 %f564, %f562, %f2636; setp.eq.ftz.f32 %p220, %f562, %f564; mov.f32 %f5132, %f562; @%p220 bra $L__BB104_251; mul.ftz.f32 %f565, %f551, %f563; mov.u64 %rd35, 106542032486400; cvt.rn.f32.u64 %f2637, %rd35; div.approx.ftz.f32 %f2638, %f565, %f2637; add.ftz.f32 %f566, %f564, %f2638; setp.eq.ftz.f32 %p221, %f564, %f566; mov.f32 %f5132, %f564; @%p221 bra $L__BB104_251; mul.ftz.f32 %f2639, %f551, %f565; mov.u64 %rd36, 34519618525593600; cvt.rn.f32.u64 %f2640, %rd36; div.approx.ftz.f32 %f2641, %f2639, %f2640; add.ftz.f32 %f567, %f566, %f2641; setp.eq.ftz.f32 %p222, %f566, %f567; mov.f32 %f5132, %f566; @%p222 bra $L__BB104_251; mov.f32 %f5132, %f567; $L__BB104_251: mul.ftz.f32 %f2642, %f528, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2643, %f2642; mul.ftz.f32 %f569, %f5132, %f2643; abs.ftz.f32 %f570, %f526; setp.ltu.ftz.f32 %p223, %f570, 0f3F800000; @%p223 bra $L__BB104_253; bra.uni $L__BB104_252; $L__BB104_253: mul.ftz.f32 %f2665, %f526, %f526; mov.f32 %f2666, 0f394FFF49; mov.f32 %f2667, 0f363D0ADA; fma.rn.ftz.f32 %f2668, %f2667, %f2665, %f2666; mov.f32 %f2669, 0f3C08889A; fma.rn.ftz.f32 %f2670, %f2668, %f2665, %f2669; mov.f32 %f2671, 0f3E2AAAAB; fma.rn.ftz.f32 %f2672, %f2670, %f2665, %f2671; mul.ftz.f32 %f2673, %f2665, %f2672; fma.rn.ftz.f32 %f5133, %f2673, %f526, %f526; bra.uni $L__BB104_254; $L__BB104_229: setp.gt.ftz.f32 %p206, %f527, 0f41400000; @%p206 bra $L__BB104_240; bra.uni $L__BB104_230; $L__BB104_240: mul.ftz.f32 %f2613, %f527, 0f41000000; rcp.approx.ftz.f32 %f2614, %f2613; add.ftz.f32 %f2615, %f2614, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2616, %f527; fma.rn.ftz.f32 %f2617, %f2616, 0fBF317218, %f2615; fma.rn.ftz.f32 %f5131, %f2617, 0f3F000000, %f527; bra.uni $L__BB104_241; $L__BB104_252: mov.f32 %f2644, 0f3FB8AA3B; mul.rn.ftz.f32 %f2645, %f570, %f2644; cvt.rzi.ftz.f32.f32 %f2646, %f2645; abs.ftz.f32 %f2647, %f2646; setp.gt.ftz.f32 %p224, %f2647, 0f42FC0000; mov.b32 %r66, %f2646; and.b32 %r67, %r66, -2147483648; or.b32 %r68, %r67, 1123811328; mov.b32 %f2648, %r68; selp.f32 %f2649, %f2648, %f2646, %p224; mov.f32 %f2650, 0fBF317218; fma.rn.ftz.f32 %f2651, %f2649, %f2650, %f570; mov.f32 %f2652, 0f3102E308; fma.rn.ftz.f32 %f2653, %f2649, %f2652, %f2651; mul.ftz.f32 %f2654, %f2653, 0f3FB8AA3B; add.ftz.f32 %f2655, %f2649, 0f4B40007D; mov.b32 %r69, %f2655; shl.b32 %r70, %r69, 23; mov.b32 %f2656, %r70; ex2.approx.ftz.f32 %f2657, %f2654; mul.ftz.f32 %f2658, %f2657, %f2656; mov.f32 %f2659, 0f3E000000; div.approx.ftz.f32 %f2660, %f2659, %f2658; neg.ftz.f32 %f2661, %f2660; mov.f32 %f2662, 0f40000000; fma.rn.ftz.f32 %f2663, %f2662, %f2658, %f2661; setp.ge.ftz.f32 %p225, %f570, 0f42B40000; selp.f32 %f2664, 0f7F800000, %f2663, %p225; mov.b32 %r71, %f2664; mov.b32 %r72, %f526; and.b32 %r73, %r72, -2147483648; or.b32 %r74, %r73, %r71; mov.b32 %f5133, %r74; $L__BB104_254: add.ftz.f32 %f2674, %f5133, %f5133; mul.ftz.f32 %f2675, %f525, %f2674; div.approx.ftz.f32 %f5134, %f569, %f2675; bra.uni $L__BB104_255; $L__BB104_230: mul.ftz.f32 %f529, %f527, %f527; fma.rn.ftz.f32 %f5130, %f529, 0f3E800000, 0f3F800000; mul.ftz.f32 %f531, %f529, %f529; mov.u64 %rd21, 64; cvt.rn.f32.u64 %f2595, %rd21; div.approx.ftz.f32 %f2596, %f531, %f2595; add.ftz.f32 %f532, %f5130, %f2596; setp.eq.ftz.f32 %p207, %f5130, %f532; @%p207 bra $L__BB104_239; mul.ftz.f32 %f533, %f529, %f531; mov.u64 %rd22, 2304; cvt.rn.f32.u64 %f2597, %rd22; div.approx.ftz.f32 %f2598, %f533, %f2597; add.ftz.f32 %f534, %f532, %f2598; setp.eq.ftz.f32 %p208, %f532, %f534; mov.f32 %f5130, %f532; @%p208 bra $L__BB104_239; mul.ftz.f32 %f535, %f529, %f533; mov.u64 %rd23, 147456; cvt.rn.f32.u64 %f2599, %rd23; div.approx.ftz.f32 %f2600, %f535, %f2599; add.ftz.f32 %f536, %f534, %f2600; setp.eq.ftz.f32 %p209, %f534, %f536; mov.f32 %f5130, %f534; @%p209 bra $L__BB104_239; mul.ftz.f32 %f537, %f529, %f535; mov.u64 %rd24, 14745600; cvt.rn.f32.u64 %f2601, %rd24; div.approx.ftz.f32 %f2602, %f537, %f2601; add.ftz.f32 %f538, %f536, %f2602; setp.eq.ftz.f32 %p210, %f536, %f538; mov.f32 %f5130, %f536; @%p210 bra $L__BB104_239; mul.ftz.f32 %f539, %f529, %f537; mov.u64 %rd25, 2123366400; cvt.rn.f32.u64 %f2603, %rd25; div.approx.ftz.f32 %f2604, %f539, %f2603; add.ftz.f32 %f540, %f538, %f2604; setp.eq.ftz.f32 %p211, %f538, %f540; mov.f32 %f5130, %f538; @%p211 bra $L__BB104_239; mul.ftz.f32 %f541, %f529, %f539; mov.u64 %rd26, 416179814400; cvt.rn.f32.u64 %f2605, %rd26; div.approx.ftz.f32 %f2606, %f541, %f2605; add.ftz.f32 %f542, %f540, %f2606; setp.eq.ftz.f32 %p212, %f540, %f542; mov.f32 %f5130, %f540; @%p212 bra $L__BB104_239; mul.ftz.f32 %f543, %f529, %f541; mov.u64 %rd27, 106542032486400; cvt.rn.f32.u64 %f2607, %rd27; div.approx.ftz.f32 %f2608, %f543, %f2607; add.ftz.f32 %f544, %f542, %f2608; setp.eq.ftz.f32 %p213, %f542, %f544; mov.f32 %f5130, %f542; @%p213 bra $L__BB104_239; mul.ftz.f32 %f2609, %f529, %f543; mov.u64 %rd28, 34519618525593600; cvt.rn.f32.u64 %f2610, %rd28; div.approx.ftz.f32 %f2611, %f2609, %f2610; add.ftz.f32 %f545, %f544, %f2611; setp.eq.ftz.f32 %p214, %f544, %f545; mov.f32 %f5130, %f544; @%p214 bra $L__BB104_239; mov.f32 %f5130, %f545; $L__BB104_239: lg2.approx.ftz.f32 %f2612, %f5130; mul.ftz.f32 %f5131, %f2612, 0f3F317218; $L__BB104_241: sub.ftz.f32 %f2618, %f5131, %f528; sub.ftz.f32 %f2619, %f2618, %f526; add.ftz.f32 %f2620, %f2619, 0f3F316F00; mul.ftz.f32 %f2621, %f526, 0f3F000000; lg2.approx.ftz.f32 %f2622, %f2621; fma.rn.ftz.f32 %f2623, %f2622, 0f3F317218, %f2620; mul.ftz.f32 %f2624, %f2623, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5134, %f2624; $L__BB104_255: setp.gt.ftz.f32 %p226, %f523, 0f42B00000; mul.ftz.f32 %f576, %f5128, %f485; mul.ftz.f32 %f577, %f5128, %f492; mul.ftz.f32 %f578, %f5128, %f505; mov.f32 %f5135, %f5136; @%p226 bra $L__BB104_257; ex2.approx.ftz.f32 %f2677, %f524; add.ftz.f32 %f2678, %f2677, 0f3F800000; rcp.approx.ftz.f32 %f2679, %f2678; add.ftz.f32 %f5135, %f2679, %f2679; $L__BB104_257: sub.ftz.f32 %f581, %f2530, %f5135; setp.eq.ftz.f32 %p227, %f581, 0f00000000; @%p227 bra $L__BB104_259; mul.ftz.f32 %f2682, %f474, 0f00000000; sub.ftz.f32 %f2683, %f2682, %f522; add.ftz.f32 %f2684, %f2683, 0f00000000; sub.ftz.f32 %f2685, %f516, %f2684; add.ftz.f32 %f2686, %f2685, 0f40490FDB; mov.f32 %f2687, 0f40C90FDB; div.approx.ftz.f32 %f2688, %f2686, %f2687; cvt.rmi.ftz.f32.f32 %f2689, %f2688; mul.ftz.f32 %f2690, %f2689, 0f40C90FDB; sub.ftz.f32 %f2691, %f2686, %f2690; add.ftz.f32 %f2692, %f2691, 0fC0490FDB; abs.ftz.f32 %f2693, %f2692; neg.ftz.f32 %f2694, %f2693; div.approx.ftz.f32 %f2695, %f2694, %f2560; mul.ftz.f32 %f2696, %f2695, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2697, %f2696; add.ftz.f32 %f2698, %f2697, 0f3F800000; mul.ftz.f32 %f2699, %f2698, %f2698; mul.ftz.f32 %f2700, %f2560, %f2699; div.approx.ftz.f32 %f2701, %f2697, %f2700; div.approx.ftz.f32 %f5136, %f2701, %f581; $L__BB104_259: mul.ftz.f32 %f2702, %f5134, %f5124; fma.rn.ftz.f32 %f584, %f5136, %f2702, 0f00000000; mul.ftz.f32 %f2703, %f5128, %f5124; mul.ftz.f32 %f2704, %f5134, %f2703; fma.rn.ftz.f32 %f585, %f5136, %f2704, 0f00000000; ld.local.f32 %f586, [%rd3+32]; mul.ftz.f32 %f587, %f586, 0f3E800000; rcp.approx.ftz.f32 %f588, %f587; mul.ftz.f32 %f2705, %f466, %f519; mul.ftz.f32 %f589, %f2705, %f588; mul.ftz.f32 %f2706, %f461, %f518; mul.ftz.f32 %f590, %f2706, %f588; setp.gtu.ftz.f32 %p228, %f587, 0f3DCCCCCD; @%p228 bra $L__BB104_273; bra.uni $L__BB104_260; $L__BB104_273: mul.ftz.f32 %f613, %f589, %f589; fma.rn.ftz.f32 %f5139, %f613, 0f3E800000, 0f3F800000; mul.ftz.f32 %f615, %f613, %f613; mov.u64 %rd45, 64; cvt.rn.f32.u64 %f2737, %rd45; div.approx.ftz.f32 %f2738, %f615, %f2737; add.ftz.f32 %f616, %f5139, %f2738; setp.eq.ftz.f32 %p238, %f5139, %f616; @%p238 bra $L__BB104_282; mul.ftz.f32 %f617, %f613, %f615; mov.u64 %rd46, 2304; cvt.rn.f32.u64 %f2739, %rd46; div.approx.ftz.f32 %f2740, %f617, %f2739; add.ftz.f32 %f618, %f616, %f2740; setp.eq.ftz.f32 %p239, %f616, %f618; mov.f32 %f5139, %f616; @%p239 bra $L__BB104_282; mul.ftz.f32 %f619, %f613, %f617; mov.u64 %rd47, 147456; cvt.rn.f32.u64 %f2741, %rd47; div.approx.ftz.f32 %f2742, %f619, %f2741; add.ftz.f32 %f620, %f618, %f2742; setp.eq.ftz.f32 %p240, %f618, %f620; mov.f32 %f5139, %f618; @%p240 bra $L__BB104_282; mul.ftz.f32 %f621, %f613, %f619; mov.u64 %rd48, 14745600; cvt.rn.f32.u64 %f2743, %rd48; div.approx.ftz.f32 %f2744, %f621, %f2743; add.ftz.f32 %f622, %f620, %f2744; setp.eq.ftz.f32 %p241, %f620, %f622; mov.f32 %f5139, %f620; @%p241 bra $L__BB104_282; mul.ftz.f32 %f623, %f613, %f621; mov.u64 %rd49, 2123366400; cvt.rn.f32.u64 %f2745, %rd49; div.approx.ftz.f32 %f2746, %f623, %f2745; add.ftz.f32 %f624, %f622, %f2746; setp.eq.ftz.f32 %p242, %f622, %f624; mov.f32 %f5139, %f622; @%p242 bra $L__BB104_282; mul.ftz.f32 %f625, %f613, %f623; mov.u64 %rd50, 416179814400; cvt.rn.f32.u64 %f2747, %rd50; div.approx.ftz.f32 %f2748, %f625, %f2747; add.ftz.f32 %f626, %f624, %f2748; setp.eq.ftz.f32 %p243, %f624, %f626; mov.f32 %f5139, %f624; @%p243 bra $L__BB104_282; mul.ftz.f32 %f627, %f613, %f625; mov.u64 %rd51, 106542032486400; cvt.rn.f32.u64 %f2749, %rd51; div.approx.ftz.f32 %f2750, %f627, %f2749; add.ftz.f32 %f628, %f626, %f2750; setp.eq.ftz.f32 %p244, %f626, %f628; mov.f32 %f5139, %f626; @%p244 bra $L__BB104_282; mul.ftz.f32 %f2751, %f613, %f627; mov.u64 %rd52, 34519618525593600; cvt.rn.f32.u64 %f2752, %rd52; div.approx.ftz.f32 %f2753, %f2751, %f2752; add.ftz.f32 %f629, %f628, %f2753; setp.eq.ftz.f32 %p245, %f628, %f629; mov.f32 %f5139, %f628; @%p245 bra $L__BB104_282; mov.f32 %f5139, %f629; $L__BB104_282: mul.ftz.f32 %f2754, %f590, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2755, %f2754; mul.ftz.f32 %f631, %f5139, %f2755; abs.ftz.f32 %f632, %f588; setp.ltu.ftz.f32 %p246, %f632, 0f3F800000; @%p246 bra $L__BB104_284; bra.uni $L__BB104_283; $L__BB104_284: mul.ftz.f32 %f2777, %f588, %f588; mov.f32 %f2778, 0f394FFF49; mov.f32 %f2779, 0f363D0ADA; fma.rn.ftz.f32 %f2780, %f2779, %f2777, %f2778; mov.f32 %f2781, 0f3C08889A; fma.rn.ftz.f32 %f2782, %f2780, %f2777, %f2781; mov.f32 %f2783, 0f3E2AAAAB; fma.rn.ftz.f32 %f2784, %f2782, %f2777, %f2783; mul.ftz.f32 %f2785, %f2777, %f2784; fma.rn.ftz.f32 %f5140, %f2785, %f588, %f588; bra.uni $L__BB104_285; $L__BB104_260: setp.gt.ftz.f32 %p229, %f589, 0f41400000; @%p229 bra $L__BB104_271; bra.uni $L__BB104_261; $L__BB104_271: mul.ftz.f32 %f2725, %f589, 0f41000000; rcp.approx.ftz.f32 %f2726, %f2725; add.ftz.f32 %f2727, %f2726, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2728, %f589; fma.rn.ftz.f32 %f2729, %f2728, 0fBF317218, %f2727; fma.rn.ftz.f32 %f5138, %f2729, 0f3F000000, %f589; bra.uni $L__BB104_272; $L__BB104_283: mov.f32 %f2756, 0f3FB8AA3B; mul.rn.ftz.f32 %f2757, %f632, %f2756; cvt.rzi.ftz.f32.f32 %f2758, %f2757; abs.ftz.f32 %f2759, %f2758; setp.gt.ftz.f32 %p247, %f2759, 0f42FC0000; mov.b32 %r75, %f2758; and.b32 %r76, %r75, -2147483648; or.b32 %r77, %r76, 1123811328; mov.b32 %f2760, %r77; selp.f32 %f2761, %f2760, %f2758, %p247; mov.f32 %f2762, 0fBF317218; fma.rn.ftz.f32 %f2763, %f2761, %f2762, %f632; mov.f32 %f2764, 0f3102E308; fma.rn.ftz.f32 %f2765, %f2761, %f2764, %f2763; mul.ftz.f32 %f2766, %f2765, 0f3FB8AA3B; add.ftz.f32 %f2767, %f2761, 0f4B40007D; mov.b32 %r78, %f2767; shl.b32 %r79, %r78, 23; mov.b32 %f2768, %r79; ex2.approx.ftz.f32 %f2769, %f2766; mul.ftz.f32 %f2770, %f2769, %f2768; mov.f32 %f2771, 0f3E000000; div.approx.ftz.f32 %f2772, %f2771, %f2770; neg.ftz.f32 %f2773, %f2772; mov.f32 %f2774, 0f40000000; fma.rn.ftz.f32 %f2775, %f2774, %f2770, %f2773; setp.ge.ftz.f32 %p248, %f632, 0f42B40000; selp.f32 %f2776, 0f7F800000, %f2775, %p248; mov.b32 %r80, %f2776; mov.b32 %r81, %f588; and.b32 %r82, %r81, -2147483648; or.b32 %r83, %r82, %r80; mov.b32 %f5140, %r83; $L__BB104_285: add.ftz.f32 %f2786, %f5140, %f5140; mul.ftz.f32 %f2787, %f587, %f2786; div.approx.ftz.f32 %f5141, %f631, %f2787; bra.uni $L__BB104_286; $L__BB104_261: mul.ftz.f32 %f591, %f589, %f589; fma.rn.ftz.f32 %f5137, %f591, 0f3E800000, 0f3F800000; mul.ftz.f32 %f593, %f591, %f591; mov.u64 %rd37, 64; cvt.rn.f32.u64 %f2707, %rd37; div.approx.ftz.f32 %f2708, %f593, %f2707; add.ftz.f32 %f594, %f5137, %f2708; setp.eq.ftz.f32 %p230, %f5137, %f594; @%p230 bra $L__BB104_270; mul.ftz.f32 %f595, %f591, %f593; mov.u64 %rd38, 2304; cvt.rn.f32.u64 %f2709, %rd38; div.approx.ftz.f32 %f2710, %f595, %f2709; add.ftz.f32 %f596, %f594, %f2710; setp.eq.ftz.f32 %p231, %f594, %f596; mov.f32 %f5137, %f594; @%p231 bra $L__BB104_270; mul.ftz.f32 %f597, %f591, %f595; mov.u64 %rd39, 147456; cvt.rn.f32.u64 %f2711, %rd39; div.approx.ftz.f32 %f2712, %f597, %f2711; add.ftz.f32 %f598, %f596, %f2712; setp.eq.ftz.f32 %p232, %f596, %f598; mov.f32 %f5137, %f596; @%p232 bra $L__BB104_270; mul.ftz.f32 %f599, %f591, %f597; mov.u64 %rd40, 14745600; cvt.rn.f32.u64 %f2713, %rd40; div.approx.ftz.f32 %f2714, %f599, %f2713; add.ftz.f32 %f600, %f598, %f2714; setp.eq.ftz.f32 %p233, %f598, %f600; mov.f32 %f5137, %f598; @%p233 bra $L__BB104_270; mul.ftz.f32 %f601, %f591, %f599; mov.u64 %rd41, 2123366400; cvt.rn.f32.u64 %f2715, %rd41; div.approx.ftz.f32 %f2716, %f601, %f2715; add.ftz.f32 %f602, %f600, %f2716; setp.eq.ftz.f32 %p234, %f600, %f602; mov.f32 %f5137, %f600; @%p234 bra $L__BB104_270; mul.ftz.f32 %f603, %f591, %f601; mov.u64 %rd42, 416179814400; cvt.rn.f32.u64 %f2717, %rd42; div.approx.ftz.f32 %f2718, %f603, %f2717; add.ftz.f32 %f604, %f602, %f2718; setp.eq.ftz.f32 %p235, %f602, %f604; mov.f32 %f5137, %f602; @%p235 bra $L__BB104_270; mul.ftz.f32 %f605, %f591, %f603; mov.u64 %rd43, 106542032486400; cvt.rn.f32.u64 %f2719, %rd43; div.approx.ftz.f32 %f2720, %f605, %f2719; add.ftz.f32 %f606, %f604, %f2720; setp.eq.ftz.f32 %p236, %f604, %f606; mov.f32 %f5137, %f604; @%p236 bra $L__BB104_270; mul.ftz.f32 %f2721, %f591, %f605; mov.u64 %rd44, 34519618525593600; cvt.rn.f32.u64 %f2722, %rd44; div.approx.ftz.f32 %f2723, %f2721, %f2722; add.ftz.f32 %f607, %f606, %f2723; setp.eq.ftz.f32 %p237, %f606, %f607; mov.f32 %f5137, %f606; @%p237 bra $L__BB104_270; mov.f32 %f5137, %f607; $L__BB104_270: lg2.approx.ftz.f32 %f2724, %f5137; mul.ftz.f32 %f5138, %f2724, 0f3F317218; $L__BB104_272: sub.ftz.f32 %f2730, %f5138, %f590; sub.ftz.f32 %f2731, %f2730, %f588; add.ftz.f32 %f2732, %f2731, 0f3F316F00; mul.ftz.f32 %f2733, %f588, 0f3F000000; lg2.approx.ftz.f32 %f2734, %f2733; fma.rn.ftz.f32 %f2735, %f2734, 0f3F317218, %f2732; mul.ftz.f32 %f2736, %f2735, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5141, %f2736; $L__BB104_286: mov.f32 %f5143, 0f00000000; mov.f32 %f5142, %f5143; @%p226 bra $L__BB104_288; ex2.approx.ftz.f32 %f2789, %f524; add.ftz.f32 %f2790, %f2789, 0f3F800000; rcp.approx.ftz.f32 %f2791, %f2790; add.ftz.f32 %f5142, %f2791, %f2791; $L__BB104_288: sub.ftz.f32 %f640, %f2530, %f5142; setp.eq.ftz.f32 %p250, %f640, 0f00000000; @%p250 bra $L__BB104_290; add.ftz.f32 %f2794, %f474, %f474; sub.ftz.f32 %f2795, %f2794, %f522; add.ftz.f32 %f2796, %f2795, 0f40490FDB; sub.ftz.f32 %f2797, %f516, %f2796; add.ftz.f32 %f2798, %f2797, 0f40490FDB; mov.f32 %f2799, 0f40C90FDB; div.approx.ftz.f32 %f2800, %f2798, %f2799; cvt.rmi.ftz.f32.f32 %f2801, %f2800; mul.ftz.f32 %f2802, %f2801, 0f40C90FDB; sub.ftz.f32 %f2803, %f2798, %f2802; add.ftz.f32 %f2804, %f2803, 0fC0490FDB; abs.ftz.f32 %f2805, %f2804; neg.ftz.f32 %f2806, %f2805; div.approx.ftz.f32 %f2807, %f2806, %f2560; mul.ftz.f32 %f2808, %f2807, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2809, %f2808; add.ftz.f32 %f2810, %f2809, 0f3F800000; mul.ftz.f32 %f2811, %f2810, %f2810; mul.ftz.f32 %f2812, %f2560, %f2811; div.approx.ftz.f32 %f2813, %f2809, %f2812; div.approx.ftz.f32 %f5143, %f2813, %f640; $L__BB104_290: mul.ftz.f32 %f2814, %f5141, %f482; mul.ftz.f32 %f2815, %f5141, %f483; mul.ftz.f32 %f2816, %f5141, %f484; fma.rn.ftz.f32 %f643, %f5143, %f2814, %f584; fma.rn.ftz.f32 %f644, %f5143, %f2815, %f584; fma.rn.ftz.f32 %f645, %f5143, %f2816, %f584; mul.ftz.f32 %f2817, %f5141, %f576; fma.rn.ftz.f32 %f646, %f5143, %f2817, %f585; mul.ftz.f32 %f647, %f586, 0f40800000; rcp.approx.ftz.f32 %f648, %f647; abs.ftz.f32 %f2818, %f521; mul.ftz.f32 %f2819, %f466, %f2818; mul.ftz.f32 %f649, %f2819, %f648; mul.ftz.f32 %f2820, %f461, %f520; mul.ftz.f32 %f650, %f2820, %f648; setp.gtu.ftz.f32 %p251, %f647, 0f3DCCCCCD; @%p251 bra $L__BB104_304; bra.uni $L__BB104_291; $L__BB104_304: mul.ftz.f32 %f673, %f649, %f649; fma.rn.ftz.f32 %f5146, %f673, 0f3E800000, 0f3F800000; mul.ftz.f32 %f675, %f673, %f673; mov.u64 %rd61, 64; cvt.rn.f32.u64 %f2851, %rd61; div.approx.ftz.f32 %f2852, %f675, %f2851; add.ftz.f32 %f676, %f5146, %f2852; setp.eq.ftz.f32 %p261, %f5146, %f676; @%p261 bra $L__BB104_313; mul.ftz.f32 %f677, %f673, %f675; mov.u64 %rd62, 2304; cvt.rn.f32.u64 %f2853, %rd62; div.approx.ftz.f32 %f2854, %f677, %f2853; add.ftz.f32 %f678, %f676, %f2854; setp.eq.ftz.f32 %p262, %f676, %f678; mov.f32 %f5146, %f676; @%p262 bra $L__BB104_313; mul.ftz.f32 %f679, %f673, %f677; mov.u64 %rd63, 147456; cvt.rn.f32.u64 %f2855, %rd63; div.approx.ftz.f32 %f2856, %f679, %f2855; add.ftz.f32 %f680, %f678, %f2856; setp.eq.ftz.f32 %p263, %f678, %f680; mov.f32 %f5146, %f678; @%p263 bra $L__BB104_313; mul.ftz.f32 %f681, %f673, %f679; mov.u64 %rd64, 14745600; cvt.rn.f32.u64 %f2857, %rd64; div.approx.ftz.f32 %f2858, %f681, %f2857; add.ftz.f32 %f682, %f680, %f2858; setp.eq.ftz.f32 %p264, %f680, %f682; mov.f32 %f5146, %f680; @%p264 bra $L__BB104_313; mul.ftz.f32 %f683, %f673, %f681; mov.u64 %rd65, 2123366400; cvt.rn.f32.u64 %f2859, %rd65; div.approx.ftz.f32 %f2860, %f683, %f2859; add.ftz.f32 %f684, %f682, %f2860; setp.eq.ftz.f32 %p265, %f682, %f684; mov.f32 %f5146, %f682; @%p265 bra $L__BB104_313; mul.ftz.f32 %f685, %f673, %f683; mov.u64 %rd66, 416179814400; cvt.rn.f32.u64 %f2861, %rd66; div.approx.ftz.f32 %f2862, %f685, %f2861; add.ftz.f32 %f686, %f684, %f2862; setp.eq.ftz.f32 %p266, %f684, %f686; mov.f32 %f5146, %f684; @%p266 bra $L__BB104_313; mul.ftz.f32 %f687, %f673, %f685; mov.u64 %rd67, 106542032486400; cvt.rn.f32.u64 %f2863, %rd67; div.approx.ftz.f32 %f2864, %f687, %f2863; add.ftz.f32 %f688, %f686, %f2864; setp.eq.ftz.f32 %p267, %f686, %f688; mov.f32 %f5146, %f686; @%p267 bra $L__BB104_313; mul.ftz.f32 %f2865, %f673, %f687; mov.u64 %rd68, 34519618525593600; cvt.rn.f32.u64 %f2866, %rd68; div.approx.ftz.f32 %f2867, %f2865, %f2866; add.ftz.f32 %f689, %f688, %f2867; setp.eq.ftz.f32 %p268, %f688, %f689; mov.f32 %f5146, %f688; @%p268 bra $L__BB104_313; mov.f32 %f5146, %f689; $L__BB104_313: mul.ftz.f32 %f2868, %f650, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2869, %f2868; mul.ftz.f32 %f691, %f5146, %f2869; abs.ftz.f32 %f692, %f648; setp.ltu.ftz.f32 %p269, %f692, 0f3F800000; @%p269 bra $L__BB104_315; bra.uni $L__BB104_314; $L__BB104_315: mul.ftz.f32 %f2891, %f648, %f648; mov.f32 %f2892, 0f394FFF49; mov.f32 %f2893, 0f363D0ADA; fma.rn.ftz.f32 %f2894, %f2893, %f2891, %f2892; mov.f32 %f2895, 0f3C08889A; fma.rn.ftz.f32 %f2896, %f2894, %f2891, %f2895; mov.f32 %f2897, 0f3E2AAAAB; fma.rn.ftz.f32 %f2898, %f2896, %f2891, %f2897; mul.ftz.f32 %f2899, %f2891, %f2898; fma.rn.ftz.f32 %f5147, %f2899, %f648, %f648; bra.uni $L__BB104_316; $L__BB104_291: setp.gt.ftz.f32 %p252, %f649, 0f41400000; @%p252 bra $L__BB104_302; bra.uni $L__BB104_292; $L__BB104_302: mul.ftz.f32 %f2839, %f649, 0f41000000; rcp.approx.ftz.f32 %f2840, %f2839; add.ftz.f32 %f2841, %f2840, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2842, %f649; fma.rn.ftz.f32 %f2843, %f2842, 0fBF317218, %f2841; fma.rn.ftz.f32 %f5145, %f2843, 0f3F000000, %f649; bra.uni $L__BB104_303; $L__BB104_314: mov.f32 %f2870, 0f3FB8AA3B; mul.rn.ftz.f32 %f2871, %f692, %f2870; cvt.rzi.ftz.f32.f32 %f2872, %f2871; abs.ftz.f32 %f2873, %f2872; setp.gt.ftz.f32 %p270, %f2873, 0f42FC0000; mov.b32 %r84, %f2872; and.b32 %r85, %r84, -2147483648; or.b32 %r86, %r85, 1123811328; mov.b32 %f2874, %r86; selp.f32 %f2875, %f2874, %f2872, %p270; mov.f32 %f2876, 0fBF317218; fma.rn.ftz.f32 %f2877, %f2875, %f2876, %f692; mov.f32 %f2878, 0f3102E308; fma.rn.ftz.f32 %f2879, %f2875, %f2878, %f2877; mul.ftz.f32 %f2880, %f2879, 0f3FB8AA3B; add.ftz.f32 %f2881, %f2875, 0f4B40007D; mov.b32 %r87, %f2881; shl.b32 %r88, %r87, 23; mov.b32 %f2882, %r88; ex2.approx.ftz.f32 %f2883, %f2880; mul.ftz.f32 %f2884, %f2883, %f2882; mov.f32 %f2885, 0f3E000000; div.approx.ftz.f32 %f2886, %f2885, %f2884; neg.ftz.f32 %f2887, %f2886; mov.f32 %f2888, 0f40000000; fma.rn.ftz.f32 %f2889, %f2888, %f2884, %f2887; setp.ge.ftz.f32 %p271, %f692, 0f42B40000; selp.f32 %f2890, 0f7F800000, %f2889, %p271; mov.b32 %r89, %f2890; mov.b32 %r90, %f648; and.b32 %r91, %r90, -2147483648; or.b32 %r92, %r91, %r89; mov.b32 %f5147, %r92; $L__BB104_316: add.ftz.f32 %f2900, %f5147, %f5147; mul.ftz.f32 %f2901, %f647, %f2900; div.approx.ftz.f32 %f5148, %f691, %f2901; bra.uni $L__BB104_317; $L__BB104_292: mul.ftz.f32 %f651, %f649, %f649; fma.rn.ftz.f32 %f5144, %f651, 0f3E800000, 0f3F800000; mul.ftz.f32 %f653, %f651, %f651; mov.u64 %rd53, 64; cvt.rn.f32.u64 %f2821, %rd53; div.approx.ftz.f32 %f2822, %f653, %f2821; add.ftz.f32 %f654, %f5144, %f2822; setp.eq.ftz.f32 %p253, %f5144, %f654; @%p253 bra $L__BB104_301; mul.ftz.f32 %f655, %f651, %f653; mov.u64 %rd54, 2304; cvt.rn.f32.u64 %f2823, %rd54; div.approx.ftz.f32 %f2824, %f655, %f2823; add.ftz.f32 %f656, %f654, %f2824; setp.eq.ftz.f32 %p254, %f654, %f656; mov.f32 %f5144, %f654; @%p254 bra $L__BB104_301; mul.ftz.f32 %f657, %f651, %f655; mov.u64 %rd55, 147456; cvt.rn.f32.u64 %f2825, %rd55; div.approx.ftz.f32 %f2826, %f657, %f2825; add.ftz.f32 %f658, %f656, %f2826; setp.eq.ftz.f32 %p255, %f656, %f658; mov.f32 %f5144, %f656; @%p255 bra $L__BB104_301; mul.ftz.f32 %f659, %f651, %f657; mov.u64 %rd56, 14745600; cvt.rn.f32.u64 %f2827, %rd56; div.approx.ftz.f32 %f2828, %f659, %f2827; add.ftz.f32 %f660, %f658, %f2828; setp.eq.ftz.f32 %p256, %f658, %f660; mov.f32 %f5144, %f658; @%p256 bra $L__BB104_301; mul.ftz.f32 %f661, %f651, %f659; mov.u64 %rd57, 2123366400; cvt.rn.f32.u64 %f2829, %rd57; div.approx.ftz.f32 %f2830, %f661, %f2829; add.ftz.f32 %f662, %f660, %f2830; setp.eq.ftz.f32 %p257, %f660, %f662; mov.f32 %f5144, %f660; @%p257 bra $L__BB104_301; mul.ftz.f32 %f663, %f651, %f661; mov.u64 %rd58, 416179814400; cvt.rn.f32.u64 %f2831, %rd58; div.approx.ftz.f32 %f2832, %f663, %f2831; add.ftz.f32 %f664, %f662, %f2832; setp.eq.ftz.f32 %p258, %f662, %f664; mov.f32 %f5144, %f662; @%p258 bra $L__BB104_301; mul.ftz.f32 %f665, %f651, %f663; mov.u64 %rd59, 106542032486400; cvt.rn.f32.u64 %f2833, %rd59; div.approx.ftz.f32 %f2834, %f665, %f2833; add.ftz.f32 %f666, %f664, %f2834; setp.eq.ftz.f32 %p259, %f664, %f666; mov.f32 %f5144, %f664; @%p259 bra $L__BB104_301; mul.ftz.f32 %f2835, %f651, %f665; mov.u64 %rd60, 34519618525593600; cvt.rn.f32.u64 %f2836, %rd60; div.approx.ftz.f32 %f2837, %f2835, %f2836; add.ftz.f32 %f667, %f666, %f2837; setp.eq.ftz.f32 %p260, %f666, %f667; mov.f32 %f5144, %f666; @%p260 bra $L__BB104_301; mov.f32 %f5144, %f667; $L__BB104_301: lg2.approx.ftz.f32 %f2838, %f5144; mul.ftz.f32 %f5145, %f2838, 0f3F317218; $L__BB104_303: sub.ftz.f32 %f2844, %f5145, %f650; sub.ftz.f32 %f2845, %f2844, %f648; add.ftz.f32 %f2846, %f2845, 0f3F316F00; mul.ftz.f32 %f2847, %f648, 0f3F000000; lg2.approx.ftz.f32 %f2848, %f2847; fma.rn.ftz.f32 %f2849, %f2848, 0f3F317218, %f2846; mul.ftz.f32 %f2850, %f2849, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5148, %f2850; $L__BB104_317: mul.ftz.f32 %f2903, %f474, 0f40800000; sub.ftz.f32 %f2904, %f2903, %f522; add.ftz.f32 %f2905, %f2904, 0f40C90FDB; sub.ftz.f32 %f698, %f516, %f2905; mov.f32 %f5150, 0f00000000; mov.f32 %f5149, %f5150; @%p226 bra $L__BB104_319; ex2.approx.ftz.f32 %f2906, %f524; add.ftz.f32 %f2907, %f2906, 0f3F800000; rcp.approx.ftz.f32 %f2908, %f2907; add.ftz.f32 %f5149, %f2908, %f2908; $L__BB104_319: add.ftz.f32 %f2910, %f698, 0f40490FDB; mov.f32 %f2911, 0f40C90FDB; div.approx.ftz.f32 %f2912, %f2910, %f2911; cvt.rmi.ftz.f32.f32 %f2913, %f2912; mul.ftz.f32 %f2914, %f2913, 0f40C90FDB; sub.ftz.f32 %f2915, %f2910, %f2914; add.ftz.f32 %f2916, %f2915, 0fC0490FDB; abs.ftz.f32 %f2917, %f2916; neg.ftz.f32 %f2918, %f2917; div.approx.ftz.f32 %f2919, %f2918, %f2560; mul.ftz.f32 %f2920, %f2919, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2921, %f2920; add.ftz.f32 %f2922, %f2921, 0f3F800000; mul.ftz.f32 %f2924, %f2922, %f2922; mul.ftz.f32 %f2925, %f2560, %f2924; div.approx.ftz.f32 %f701, %f2921, %f2925; sub.ftz.f32 %f702, %f2530, %f5149; setp.eq.ftz.f32 %p273, %f702, 0f00000000; @%p273 bra $L__BB104_321; div.approx.ftz.f32 %f5150, %f701, %f702; $L__BB104_321: mul.ftz.f32 %f2926, %f5148, %f487; mul.ftz.f32 %f2927, %f5148, %f489; mul.ftz.f32 %f2928, %f5148, %f491; fma.rn.ftz.f32 %f705, %f5150, %f2926, %f643; fma.rn.ftz.f32 %f706, %f5150, %f2927, %f644; fma.rn.ftz.f32 %f707, %f5150, %f2928, %f645; mul.ftz.f32 %f2929, %f5148, %f577; fma.rn.ftz.f32 %f708, %f5150, %f2929, %f646; mul.ftz.f32 %f2930, %f466, %f509; mul.ftz.f32 %f709, %f2930, %f648; mul.ftz.f32 %f2931, %f462, %f461; mul.ftz.f32 %f710, %f2931, %f648; @%p251 bra $L__BB104_335; bra.uni $L__BB104_322; $L__BB104_335: mul.ftz.f32 %f733, %f709, %f709; fma.rn.ftz.f32 %f5153, %f733, 0f3E800000, 0f3F800000; mul.ftz.f32 %f735, %f733, %f733; mov.u64 %rd77, 64; cvt.rn.f32.u64 %f2962, %rd77; div.approx.ftz.f32 %f2963, %f735, %f2962; add.ftz.f32 %f736, %f5153, %f2963; setp.eq.ftz.f32 %p284, %f5153, %f736; @%p284 bra $L__BB104_344; mul.ftz.f32 %f737, %f733, %f735; mov.u64 %rd78, 2304; cvt.rn.f32.u64 %f2964, %rd78; div.approx.ftz.f32 %f2965, %f737, %f2964; add.ftz.f32 %f738, %f736, %f2965; setp.eq.ftz.f32 %p285, %f736, %f738; mov.f32 %f5153, %f736; @%p285 bra $L__BB104_344; mul.ftz.f32 %f739, %f733, %f737; mov.u64 %rd79, 147456; cvt.rn.f32.u64 %f2966, %rd79; div.approx.ftz.f32 %f2967, %f739, %f2966; add.ftz.f32 %f740, %f738, %f2967; setp.eq.ftz.f32 %p286, %f738, %f740; mov.f32 %f5153, %f738; @%p286 bra $L__BB104_344; mul.ftz.f32 %f741, %f733, %f739; mov.u64 %rd80, 14745600; cvt.rn.f32.u64 %f2968, %rd80; div.approx.ftz.f32 %f2969, %f741, %f2968; add.ftz.f32 %f742, %f740, %f2969; setp.eq.ftz.f32 %p287, %f740, %f742; mov.f32 %f5153, %f740; @%p287 bra $L__BB104_344; mul.ftz.f32 %f743, %f733, %f741; mov.u64 %rd81, 2123366400; cvt.rn.f32.u64 %f2970, %rd81; div.approx.ftz.f32 %f2971, %f743, %f2970; add.ftz.f32 %f744, %f742, %f2971; setp.eq.ftz.f32 %p288, %f742, %f744; mov.f32 %f5153, %f742; @%p288 bra $L__BB104_344; mul.ftz.f32 %f745, %f733, %f743; mov.u64 %rd82, 416179814400; cvt.rn.f32.u64 %f2972, %rd82; div.approx.ftz.f32 %f2973, %f745, %f2972; add.ftz.f32 %f746, %f744, %f2973; setp.eq.ftz.f32 %p289, %f744, %f746; mov.f32 %f5153, %f744; @%p289 bra $L__BB104_344; mul.ftz.f32 %f747, %f733, %f745; mov.u64 %rd83, 106542032486400; cvt.rn.f32.u64 %f2974, %rd83; div.approx.ftz.f32 %f2975, %f747, %f2974; add.ftz.f32 %f748, %f746, %f2975; setp.eq.ftz.f32 %p290, %f746, %f748; mov.f32 %f5153, %f746; @%p290 bra $L__BB104_344; mul.ftz.f32 %f2976, %f733, %f747; mov.u64 %rd84, 34519618525593600; cvt.rn.f32.u64 %f2977, %rd84; div.approx.ftz.f32 %f2978, %f2976, %f2977; add.ftz.f32 %f749, %f748, %f2978; setp.eq.ftz.f32 %p291, %f748, %f749; mov.f32 %f5153, %f748; @%p291 bra $L__BB104_344; mov.f32 %f5153, %f749; $L__BB104_344: mul.ftz.f32 %f2979, %f710, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2980, %f2979; mul.ftz.f32 %f751, %f5153, %f2980; abs.ftz.f32 %f752, %f648; setp.ltu.ftz.f32 %p292, %f752, 0f3F800000; @%p292 bra $L__BB104_346; bra.uni $L__BB104_345; $L__BB104_346: mul.ftz.f32 %f3002, %f648, %f648; mov.f32 %f3003, 0f394FFF49; mov.f32 %f3004, 0f363D0ADA; fma.rn.ftz.f32 %f3005, %f3004, %f3002, %f3003; mov.f32 %f3006, 0f3C08889A; fma.rn.ftz.f32 %f3007, %f3005, %f3002, %f3006; mov.f32 %f3008, 0f3E2AAAAB; fma.rn.ftz.f32 %f3009, %f3007, %f3002, %f3008; mul.ftz.f32 %f3010, %f3002, %f3009; fma.rn.ftz.f32 %f5154, %f3010, %f648, %f648; bra.uni $L__BB104_347; $L__BB104_322: setp.gt.ftz.f32 %p275, %f709, 0f41400000; @%p275 bra $L__BB104_333; bra.uni $L__BB104_323; $L__BB104_333: mul.ftz.f32 %f2950, %f709, 0f41000000; rcp.approx.ftz.f32 %f2951, %f2950; add.ftz.f32 %f2952, %f2951, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2953, %f709; fma.rn.ftz.f32 %f2954, %f2953, 0fBF317218, %f2952; fma.rn.ftz.f32 %f5152, %f2954, 0f3F000000, %f709; bra.uni $L__BB104_334; $L__BB104_345: mov.f32 %f2981, 0f3FB8AA3B; mul.rn.ftz.f32 %f2982, %f752, %f2981; cvt.rzi.ftz.f32.f32 %f2983, %f2982; abs.ftz.f32 %f2984, %f2983; setp.gt.ftz.f32 %p293, %f2984, 0f42FC0000; mov.b32 %r93, %f2983; and.b32 %r94, %r93, -2147483648; or.b32 %r95, %r94, 1123811328; mov.b32 %f2985, %r95; selp.f32 %f2986, %f2985, %f2983, %p293; mov.f32 %f2987, 0fBF317218; fma.rn.ftz.f32 %f2988, %f2986, %f2987, %f752; mov.f32 %f2989, 0f3102E308; fma.rn.ftz.f32 %f2990, %f2986, %f2989, %f2988; mul.ftz.f32 %f2991, %f2990, 0f3FB8AA3B; add.ftz.f32 %f2992, %f2986, 0f4B40007D; mov.b32 %r96, %f2992; shl.b32 %r97, %r96, 23; mov.b32 %f2993, %r97; ex2.approx.ftz.f32 %f2994, %f2991; mul.ftz.f32 %f2995, %f2994, %f2993; mov.f32 %f2996, 0f3E000000; div.approx.ftz.f32 %f2997, %f2996, %f2995; neg.ftz.f32 %f2998, %f2997; mov.f32 %f2999, 0f40000000; fma.rn.ftz.f32 %f3000, %f2999, %f2995, %f2998; setp.ge.ftz.f32 %p294, %f752, 0f42B40000; selp.f32 %f3001, 0f7F800000, %f3000, %p294; mov.b32 %r98, %f3001; mov.b32 %r99, %f648; and.b32 %r100, %r99, -2147483648; or.b32 %r101, %r100, %r98; mov.b32 %f5154, %r101; $L__BB104_347: add.ftz.f32 %f3011, %f5154, %f5154; mul.ftz.f32 %f3012, %f647, %f3011; div.approx.ftz.f32 %f5155, %f751, %f3012; bra.uni $L__BB104_348; $L__BB104_323: mul.ftz.f32 %f711, %f709, %f709; fma.rn.ftz.f32 %f5151, %f711, 0f3E800000, 0f3F800000; mul.ftz.f32 %f713, %f711, %f711; mov.u64 %rd69, 64; cvt.rn.f32.u64 %f2932, %rd69; div.approx.ftz.f32 %f2933, %f713, %f2932; add.ftz.f32 %f714, %f5151, %f2933; setp.eq.ftz.f32 %p276, %f5151, %f714; @%p276 bra $L__BB104_332; mul.ftz.f32 %f715, %f711, %f713; mov.u64 %rd70, 2304; cvt.rn.f32.u64 %f2934, %rd70; div.approx.ftz.f32 %f2935, %f715, %f2934; add.ftz.f32 %f716, %f714, %f2935; setp.eq.ftz.f32 %p277, %f714, %f716; mov.f32 %f5151, %f714; @%p277 bra $L__BB104_332; mul.ftz.f32 %f717, %f711, %f715; mov.u64 %rd71, 147456; cvt.rn.f32.u64 %f2936, %rd71; div.approx.ftz.f32 %f2937, %f717, %f2936; add.ftz.f32 %f718, %f716, %f2937; setp.eq.ftz.f32 %p278, %f716, %f718; mov.f32 %f5151, %f716; @%p278 bra $L__BB104_332; mul.ftz.f32 %f719, %f711, %f717; mov.u64 %rd72, 14745600; cvt.rn.f32.u64 %f2938, %rd72; div.approx.ftz.f32 %f2939, %f719, %f2938; add.ftz.f32 %f720, %f718, %f2939; setp.eq.ftz.f32 %p279, %f718, %f720; mov.f32 %f5151, %f718; @%p279 bra $L__BB104_332; mul.ftz.f32 %f721, %f711, %f719; mov.u64 %rd73, 2123366400; cvt.rn.f32.u64 %f2940, %rd73; div.approx.ftz.f32 %f2941, %f721, %f2940; add.ftz.f32 %f722, %f720, %f2941; setp.eq.ftz.f32 %p280, %f720, %f722; mov.f32 %f5151, %f720; @%p280 bra $L__BB104_332; mul.ftz.f32 %f723, %f711, %f721; mov.u64 %rd74, 416179814400; cvt.rn.f32.u64 %f2942, %rd74; div.approx.ftz.f32 %f2943, %f723, %f2942; add.ftz.f32 %f724, %f722, %f2943; setp.eq.ftz.f32 %p281, %f722, %f724; mov.f32 %f5151, %f722; @%p281 bra $L__BB104_332; mul.ftz.f32 %f725, %f711, %f723; mov.u64 %rd75, 106542032486400; cvt.rn.f32.u64 %f2944, %rd75; div.approx.ftz.f32 %f2945, %f725, %f2944; add.ftz.f32 %f726, %f724, %f2945; setp.eq.ftz.f32 %p282, %f724, %f726; mov.f32 %f5151, %f724; @%p282 bra $L__BB104_332; mul.ftz.f32 %f2946, %f711, %f725; mov.u64 %rd76, 34519618525593600; cvt.rn.f32.u64 %f2947, %rd76; div.approx.ftz.f32 %f2948, %f2946, %f2947; add.ftz.f32 %f727, %f726, %f2948; setp.eq.ftz.f32 %p283, %f726, %f727; mov.f32 %f5151, %f726; @%p283 bra $L__BB104_332; mov.f32 %f5151, %f727; $L__BB104_332: lg2.approx.ftz.f32 %f2949, %f5151; mul.ftz.f32 %f5152, %f2949, 0f3F317218; $L__BB104_334: sub.ftz.f32 %f2955, %f5152, %f710; sub.ftz.f32 %f2956, %f2955, %f648; add.ftz.f32 %f2957, %f2956, 0f3F316F00; mul.ftz.f32 %f2958, %f648, 0f3F000000; lg2.approx.ftz.f32 %f2959, %f2958; fma.rn.ftz.f32 %f2960, %f2959, 0f3F317218, %f2957; mul.ftz.f32 %f2961, %f2960, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5155, %f2961; $L__BB104_348: mul.ftz.f32 %f3013, %f5155, %f502; mul.ftz.f32 %f3014, %f5155, %f503; mul.ftz.f32 %f3015, %f5155, %f504; fma.rn.ftz.f32 %f5290, %f3013, 0f3E22F983, %f705; fma.rn.ftz.f32 %f5289, %f3014, 0f3E22F983, %f706; fma.rn.ftz.f32 %f5288, %f3015, 0f3E22F983, %f707; mul.ftz.f32 %f3016, %f5155, %f578; fma.rn.ftz.f32 %f5291, %f3016, 0f3E22F983, %f708; bra.uni $L__BB104_519; $L__BB104_468: ld.local.f32 %f4158, [%rd3+52]; mul.ftz.f32 %f4159, %f1181, %f4158; ld.local.f32 %f4160, [%rd3+48]; mul.ftz.f32 %f4161, %f1182, %f4160; sub.ftz.f32 %f4162, %f4159, %f4161; ld.local.f32 %f4163, [%rd3+44]; mul.ftz.f32 %f4164, %f1182, %f4163; mul.ftz.f32 %f4165, %f1180, %f4158; sub.ftz.f32 %f4166, %f4164, %f4165; mul.ftz.f32 %f4167, %f1180, %f4160; mul.ftz.f32 %f4168, %f1181, %f4163; sub.ftz.f32 %f4169, %f4167, %f4168; mul.ftz.f32 %f4170, %f4166, %f4166; fma.rn.ftz.f32 %f4171, %f4162, %f4162, %f4170; fma.rn.ftz.f32 %f4172, %f4169, %f4169, %f4171; sqrt.approx.ftz.f32 %f4173, %f4172; div.approx.ftz.f32 %f4174, %f4162, %f4173; div.approx.ftz.f32 %f4175, %f4166, %f4173; div.approx.ftz.f32 %f4176, %f4169, %f4173; mul.ftz.f32 %f4177, %f1182, %f4175; mul.ftz.f32 %f4178, %f1181, %f4176; sub.ftz.f32 %f4179, %f4177, %f4178; mul.ftz.f32 %f4180, %f1180, %f4176; mul.ftz.f32 %f4181, %f1182, %f4174; sub.ftz.f32 %f4182, %f4180, %f4181; mul.ftz.f32 %f4183, %f1181, %f4174; mul.ftz.f32 %f4184, %f1180, %f4175; sub.ftz.f32 %f4185, %f4183, %f4184; mul.ftz.f32 %f4186, %f1186, %f4182; fma.rn.ftz.f32 %f4187, %f1185, %f4179, %f4186; fma.rn.ftz.f32 %f4188, %f1187, %f4185, %f4187; mul.ftz.f32 %f4189, %f1186, %f4175; fma.rn.ftz.f32 %f4190, %f1185, %f4174, %f4189; fma.rn.ftz.f32 %f4191, %f1187, %f4176, %f4190; neg.ftz.f32 %f4192, %f4188; mul.ftz.f32 %f4193, %f4134, %f1188; div.approx.ftz.f32 %f4194, %f4192, %f4193; neg.ftz.f32 %f4195, %f4191; mul.ftz.f32 %f4196, %f4135, %f1188; div.approx.ftz.f32 %f4197, %f4195, %f4196; fma.rn.ftz.f32 %f4198, %f4194, %f4194, 0f3F800000; mov.f32 %f4199, 0f3F800000; fma.rn.ftz.f32 %f4200, %f4197, %f4197, %f4198; mul.ftz.f32 %f4201, %f4200, %f4200; mul.ftz.f32 %f4202, %f4201, 0f40490FDB; mul.ftz.f32 %f4203, %f5255, %f4202; mul.ftz.f32 %f4204, %f1189, %f1189; mul.ftz.f32 %f4205, %f4204, %f4203; rcp.approx.ftz.f32 %f5256, %f4205; mul.ftz.f32 %f4206, %f1183, %f1183; sub.ftz.f32 %f4207, %f4199, %f4206; div.approx.ftz.f32 %f4208, %f4207, %f4206; mul.ftz.f32 %f4209, %f1174, %f4182; fma.rn.ftz.f32 %f4210, %f1173, %f4179, %f4209; fma.rn.ftz.f32 %f4211, %f1175, %f4185, %f4210; mul.ftz.f32 %f4212, %f1174, %f4175; fma.rn.ftz.f32 %f4213, %f1173, %f4174, %f4212; fma.rn.ftz.f32 %f4214, %f1175, %f4176, %f4213; mul.ftz.f32 %f4215, %f4211, %f4211; mul.ftz.f32 %f4216, %f4134, %f4134; mul.ftz.f32 %f4217, %f4216, %f4215; mul.ftz.f32 %f4218, %f4214, %f4214; mul.ftz.f32 %f4219, %f4135, %f4135; fma.rn.ftz.f32 %f4220, %f4219, %f4218, %f4217; add.ftz.f32 %f4221, %f4218, %f4215; div.approx.ftz.f32 %f4222, %f4220, %f4221; fma.rn.ftz.f32 %f4223, %f4208, %f4222, 0f3F800000; mov.f32 %f4224, 0f00000000; max.ftz.f32 %f4225, %f4223, %f4224; sqrt.approx.ftz.f32 %f4226, %f4225; add.ftz.f32 %f4227, %f4226, 0f3F800000; mov.f32 %f4228, 0f40000000; div.approx.ftz.f32 %f5258, %f4228, %f4227; mul.ftz.f32 %f4229, %f1184, %f1184; sub.ftz.f32 %f4230, %f4199, %f4229; div.approx.ftz.f32 %f4231, %f4230, %f4229; mul.ftz.f32 %f4232, %f1405, %f4182; fma.rn.ftz.f32 %f4233, %f1404, %f4179, %f4232; fma.rn.ftz.f32 %f4234, %f1406, %f4185, %f4233; mul.ftz.f32 %f4235, %f1405, %f4175; fma.rn.ftz.f32 %f4236, %f1404, %f4174, %f4235; fma.rn.ftz.f32 %f4237, %f1406, %f4176, %f4236; mul.ftz.f32 %f4238, %f4234, %f4234; mul.ftz.f32 %f4239, %f4216, %f4238; mul.ftz.f32 %f4240, %f4237, %f4237; fma.rn.ftz.f32 %f4241, %f4219, %f4240, %f4239; add.ftz.f32 %f4242, %f4240, %f4238; div.approx.ftz.f32 %f4243, %f4241, %f4242; mul.ftz.f32 %f5259, %f4231, %f4243; bra.uni $L__BB104_474; $L__BB104_470: mul.ftz.f32 %f4244, %f1189, %f1189; mul.ftz.f32 %f4245, %f4244, 0f40490FDB; mov.f32 %f4246, 0f3F800000; sub.ftz.f32 %f4247, %f4246, %f1189; div.approx.ftz.f32 %f4248, %f4247, %f1189; add.ftz.f32 %f4249, %f5255, %f4248; mul.ftz.f32 %f4250, %f4245, %f4249; mul.ftz.f32 %f4251, %f4249, %f4250; div.approx.ftz.f32 %f5256, %f5255, %f4251; $L__BB104_473: mul.ftz.f32 %f4263, %f1183, %f1183; mov.f32 %f4264, 0f3F800000; sub.ftz.f32 %f4265, %f4264, %f4263; mul.ftz.f32 %f4266, %f4265, %f5255; div.approx.ftz.f32 %f4267, %f4266, %f4263; add.ftz.f32 %f4268, %f4267, 0f3F800000; mov.f32 %f4269, 0f00000000; max.ftz.f32 %f4270, %f4268, %f4269; sqrt.approx.ftz.f32 %f4271, %f4270; add.ftz.f32 %f4272, %f4271, 0f3F800000; mov.f32 %f4273, 0f40000000; div.approx.ftz.f32 %f5258, %f4273, %f4272; mul.ftz.f32 %f4274, %f1184, %f1184; sub.ftz.f32 %f4275, %f4264, %f4274; mul.ftz.f32 %f4276, %f4275, %f5255; div.approx.ftz.f32 %f5259, %f4276, %f4274; $L__BB104_474: add.ftz.f32 %f4280, %f5259, 0f3F800000; mov.f32 %f4281, 0f00000000; max.ftz.f32 %f4282, %f4280, %f4281; sqrt.approx.ftz.f32 %f4283, %f4282; add.ftz.f32 %f4284, %f4283, 0f3F800000; mov.f32 %f4285, 0f40000000; div.approx.ftz.f32 %f4286, %f4285, %f4284; mul.ftz.f32 %f1202, %f5258, %f4286; mul.ftz.f32 %f4287, %f5256, 0f3E800000; div.approx.ftz.f32 %f1203, %f4287, %f1183; add.s32 %r122, %r5, -11; setp.gt.u32 %p443, %r122, 1; mov.f32 %f5267, 0f3F800000; mov.f32 %f5266, %f5267; mov.f32 %f5265, %f5267; @%p443 bra $L__BB104_480; ld.local.f32 %f4289, [%rd3+28]; fma.rn.ftz.f32 %f1204, %f4289, %f4289, 0fBF800000; mov.f32 %f4288, 0f3F800000; abs.ftz.f32 %f1205, %f4288; fma.rn.ftz.f32 %f1206, %f1205, %f1205, %f1204; setp.leu.ftz.f32 %p444, %f1206, 0f00000000; mov.f32 %f5260, %f4288; @%p444 bra $L__BB104_477; sqrt.approx.ftz.f32 %f4290, %f1206; sub.ftz.f32 %f4291, %f4290, %f1205; add.ftz.f32 %f4292, %f1205, %f4290; div.approx.ftz.f32 %f4293, %f4291, %f4292; fma.rn.ftz.f32 %f4294, %f1205, %f4292, 0fBF800000; fma.rn.ftz.f32 %f4295, %f1205, %f4291, 0f3F800000; div.approx.ftz.f32 %f4296, %f4294, %f4295; mul.ftz.f32 %f4297, %f4293, 0f3F000000; mul.ftz.f32 %f4298, %f4293, %f4297; fma.rn.ftz.f32 %f4299, %f4296, %f4296, 0f3F800000; mul.ftz.f32 %f5260, %f4298, %f4299; $L__BB104_477: ld.local.u64 %rd86, [%rd3+36]; ld.f32 %f1209, [%rd86+12]; ld.f32 %f1210, [%rd86+16]; ld.f32 %f1211, [%rd86+20]; sub.ftz.f32 %f1212, %f4288, %f5260; mul.ftz.f32 %f4301, %f1405, %f1186; fma.rn.ftz.f32 %f4302, %f1404, %f1185, %f4301; fma.rn.ftz.f32 %f4303, %f1406, %f1187, %f4302; abs.ftz.f32 %f1213, %f4303; fma.rn.ftz.f32 %f1214, %f1213, %f1213, %f1204; setp.leu.ftz.f32 %p445, %f1214, 0f00000000; mov.f32 %f5261, %f1212; @%p445 bra $L__BB104_479; sqrt.approx.ftz.f32 %f4304, %f1214; sub.ftz.f32 %f4305, %f4304, %f1213; add.ftz.f32 %f4306, %f1213, %f4304; div.approx.ftz.f32 %f4307, %f4305, %f4306; fma.rn.ftz.f32 %f4308, %f1213, %f4306, 0fBF800000; fma.rn.ftz.f32 %f4309, %f1213, %f4305, 0f3F800000; div.approx.ftz.f32 %f4310, %f4308, %f4309; mul.ftz.f32 %f4311, %f4307, 0f3F000000; mul.ftz.f32 %f4312, %f4307, %f4311; fma.rn.ftz.f32 %f4313, %f4310, %f4310, 0f3F800000; mul.ftz.f32 %f4314, %f4312, %f4313; sub.ftz.f32 %f5261, %f4314, %f5260; $L__BB104_479: rcp.approx.ftz.f32 %f4315, %f1212; mov.f32 %f4316, 0f3F800000; mul.ftz.f32 %f4317, %f4315, %f5261; sub.ftz.f32 %f4318, %f4316, %f4317; fma.rn.ftz.f32 %f5265, %f1211, %f4318, %f4317; fma.rn.ftz.f32 %f5266, %f1210, %f4318, %f4317; fma.rn.ftz.f32 %f5267, %f1209, %f4318, %f4317; $L__BB104_480: setp.ne.s32 %p446, %r5, 12; @%p446 bra $L__BB104_482; ld.local.u64 %rd87, [%rd3+36]; ld.f32 %f4319, [%rd87+36]; mul.ftz.f32 %f4320, %f4319, 0f3E800000; mul.ftz.f32 %f5267, %f5267, %f4320; mul.ftz.f32 %f5266, %f5266, %f4320; mul.ftz.f32 %f5265, %f5265, %f4320; $L__BB104_482: mul.ftz.f32 %f4321, %f1202, %f5267; mul.ftz.f32 %f4322, %f1202, %f5266; mul.ftz.f32 %f4323, %f1202, %f5265; mul.ftz.f32 %f5288, %f1203, %f4323; mul.ftz.f32 %f5289, %f1203, %f4322; mul.ftz.f32 %f5290, %f1203, %f4321; mul.ftz.f32 %f5291, %f5258, %f1203; bra.uni $L__BB104_519; $L__BB104_93: setp.lt.ftz.f32 %p72, %f89, 0f00000000; @%p72 bra $L__BB104_101; mul.ftz.f32 %f1648, %f5013, %f88; fma.rn.ftz.f32 %f1649, %f5012, %f87, %f1648; fma.rn.ftz.f32 %f91, %f5014, %f89, %f1649; abs.ftz.f32 %f92, %f91; fma.rn.ftz.f32 %f93, %f92, %f92, %f37; setp.leu.ftz.f32 %p73, %f93, 0f00000000; mov.f32 %f1647, 0f3F800000; mov.f32 %f5020, %f1647; @%p73 bra $L__BB104_96; sqrt.approx.ftz.f32 %f1650, %f93; sub.ftz.f32 %f1651, %f1650, %f92; add.ftz.f32 %f1652, %f92, %f1650; div.approx.ftz.f32 %f1653, %f1651, %f1652; fma.rn.ftz.f32 %f1654, %f92, %f1652, 0fBF800000; fma.rn.ftz.f32 %f1655, %f92, %f1651, 0f3F800000; div.approx.ftz.f32 %f1656, %f1654, %f1655; mul.ftz.f32 %f1657, %f1653, 0f3F000000; mul.ftz.f32 %f1658, %f1653, %f1657; fma.rn.ftz.f32 %f1659, %f1656, %f1656, 0f3F800000; mul.ftz.f32 %f5020, %f1658, %f1659; $L__BB104_96: mov.f32 %f1660, 0f00000000; max.ftz.f32 %f1661, %f1660, %f91; mul.ftz.f32 %f1662, %f5020, %f1661; mul.ftz.f32 %f1663, %f89, %f89; sub.ftz.f32 %f1665, %f1647, %f1663; fma.rn.ftz.f32 %f1666, %f33, %f1663, %f1665; mul.ftz.f32 %f1667, %f1666, 0f40490FDB; mul.ftz.f32 %f1668, %f1666, %f1667; max.ftz.f32 %f1670, %f1668, %f1505; div.approx.ftz.f32 %f1671, %f33, %f1670; mul.ftz.f32 %f1672, %f1662, %f1671; mul.ftz.f32 %f1673, %f1672, 0f3E800000; mul.ftz.f32 %f1674, %f90, %f91; div.approx.ftz.f32 %f5023, %f1673, %f1674; $L__BB104_101: neg.ftz.f32 %f113, %f5106; @%p50 bra $L__BB104_103; bra.uni $L__BB104_102; $L__BB104_103: add.ftz.f32 %f1728, %f5019, 0f3F800000; { // callseq 81, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f113; .param .b32 param1; st.param.f32 [param1+0], %f1728; .param .b32 retval0; call.uni (retval0), _Z4betaff, ( param0, param1 ); ld.param.f32 %f1729, [retval0+0]; } // callseq 81 mul.ftz.f32 %f5024, %f1729, %f113; bra.uni $L__BB104_104; $L__BB104_102: sub.ftz.f32 %f1727, %f5019, %f5106; div.approx.ftz.f32 %f5024, %f113, %f1727; $L__BB104_104: mul.ftz.f32 %f5289, %f5023, %f5024; mul.ftz.f32 %f5290, %f5023, %f5024; mul.ftz.f32 %f5288, %f5023, %f5024; mov.f32 %f5049, 0f3F800000; abs.ftz.f32 %f120, %f5049; fma.rn.ftz.f32 %f121, %f120, %f120, %f37; setp.leu.ftz.f32 %p78, %f121, 0f00000000; mov.f32 %f5025, %f5049; @%p78 bra $L__BB104_106; sqrt.approx.ftz.f32 %f1731, %f121; sub.ftz.f32 %f1732, %f1731, %f120; add.ftz.f32 %f1733, %f120, %f1731; div.approx.ftz.f32 %f1734, %f1732, %f1733; fma.rn.ftz.f32 %f1735, %f120, %f1733, 0fBF800000; fma.rn.ftz.f32 %f1736, %f120, %f1732, 0f3F800000; div.approx.ftz.f32 %f1737, %f1735, %f1736; mul.ftz.f32 %f1738, %f1734, 0f3F000000; mul.ftz.f32 %f1739, %f1734, %f1738; fma.rn.ftz.f32 %f1740, %f1737, %f1737, 0f3F800000; mul.ftz.f32 %f5025, %f1739, %f1740; $L__BB104_106: setp.ne.s32 %p79, %r5, 27; or.pred %p3, %p50, %p79; sub.ftz.f32 %f5026, %f5049, %f5025; rcp.approx.ftz.f32 %f128, %f5026; mov.f32 %f5103, %f5049; mov.f32 %f5104, %f5049; mov.f32 %f5105, %f5049; @%p3 bra $L__BB104_110; add.ftz.f32 %f4954, %f5009, %f5012; div.approx.ftz.f32 %f4953, %f4954, %f1642; add.ftz.f32 %f4952, %f5010, %f5013; div.approx.ftz.f32 %f4951, %f4952, %f1642; mul.ftz.f32 %f1744, %f5013, %f4951; fma.rn.ftz.f32 %f1745, %f5012, %f4953, %f1744; fma.rn.ftz.f32 %f1746, %f5014, %f89, %f1745; abs.ftz.f32 %f129, %f1746; fma.rn.ftz.f32 %f130, %f129, %f129, %f37; setp.leu.ftz.f32 %p81, %f130, 0f00000000; @%p81 bra $L__BB104_109; sqrt.approx.ftz.f32 %f1747, %f130; sub.ftz.f32 %f1748, %f1747, %f129; add.ftz.f32 %f1749, %f129, %f1747; div.approx.ftz.f32 %f1750, %f1748, %f1749; fma.rn.ftz.f32 %f1751, %f129, %f1749, 0fBF800000; fma.rn.ftz.f32 %f1752, %f129, %f1748, 0f3F800000; div.approx.ftz.f32 %f1753, %f1751, %f1752; mul.ftz.f32 %f1754, %f1750, 0f3F000000; mul.ftz.f32 %f1755, %f1750, %f1754; fma.rn.ftz.f32 %f1756, %f1753, %f1753, 0f3F800000; mul.ftz.f32 %f1757, %f1755, %f1756; sub.ftz.f32 %f5026, %f1757, %f5025; $L__BB104_109: mul.ftz.f32 %f1758, %f128, %f5026; mov.f32 %f1759, 0f3F800000; sub.ftz.f32 %f1760, %f1759, %f1758; fma.rn.ftz.f32 %f5103, %f5006, %f1760, %f1758; fma.rn.ftz.f32 %f5104, %f5007, %f1760, %f1758; fma.rn.ftz.f32 %f5105, %f5008, %f1760, %f1758; mul.ftz.f32 %f5290, %f5290, %f5103; mul.ftz.f32 %f5289, %f5289, %f5104; mul.ftz.f32 %f5288, %f5288, %f5105; $L__BB104_110: neg.ftz.f32 %f5102, %f5014; neg.ftz.f32 %f5100, %f5012; neg.ftz.f32 %f5101, %f5013; rcp.approx.ftz.f32 %f145, %f20; fma.rn.ftz.f32 %f152, %f145, %f145, 0fBF800000; ld.local.u32 %r151, [%rd1+-4]; mov.f32 %f5047, 0f00000000; mov.u32 %r150, 0; mov.u16 %rs34, %rs22; mov.f32 %f5048, %f5049; $L__BB104_111: mov.u32 %r8, %r150; mad.lo.s32 %r151, %r151, 1103515245, 12345; cvt.rn.f32.u32 %f1767, %r151; mul.ftz.f32 %f167, %f1767, 0f2F800000; setp.gt.ftz.f32 %p84, %f5102, 0f3F7FF972; mov.pred %p83, -1; mov.pred %p528, %p83; @%p84 bra $L__BB104_121; bra.uni $L__BB104_112; $L__BB104_121: @%p528 bra $L__BB104_202; mad.lo.s32 %r41, %r151, 1103515245, 12345; cvt.rn.f32.u32 %f1787, %r41; mul.ftz.f32 %f178, %f1787, 0f2F800000; mad.lo.s32 %r151, %r41, 1103515245, 12345; cvt.rn.f32.u32 %f1788, %r151; mul.ftz.f32 %f179, %f1788, 0f2F800000; neg.ftz.f32 %f180, %f5100; mul.ftz.f32 %f1789, %f1421, %f180; neg.ftz.f32 %f181, %f5101; mul.ftz.f32 %f1790, %f1422, %f181; mul.ftz.f32 %f1791, %f1790, %f1790; fma.rn.ftz.f32 %f1792, %f1789, %f1789, %f1791; fma.rn.ftz.f32 %f1793, %f5102, %f5102, %f1792; sqrt.approx.ftz.f32 %f1794, %f1793; div.approx.ftz.f32 %f5052, %f1789, %f1794; div.approx.ftz.f32 %f5053, %f1790, %f1794; neg.ftz.f32 %f184, %f5102; div.approx.ftz.f32 %f185, %f184, %f1794; setp.gt.ftz.f32 %p97, %f185, 0f3F7FF972; @%p97 bra $L__BB104_128; neg.ftz.f32 %f4890, %f5102; div.approx.ftz.f32 %f4889, %f4890, %f1794; abs.ftz.f32 %f1795, %f4889; setp.lt.ftz.f32 %p98, %f1795, 0f358637BD; @%p98 bra $L__BB104_128; bra.uni $L__BB104_124; $L__BB104_128: mov.f32 %f1839, 0f3F800000; sub.ftz.f32 %f1840, %f1839, %f179; max.ftz.f32 %f1842, %f1840, %f1505; div.approx.ftz.f32 %f1843, %f179, %f1842; sqrt.approx.ftz.f32 %f1844, %f1843; mul.ftz.f32 %f1845, %f178, 0f40C90FDB; cos.approx.ftz.f32 %f1846, %f1845; sin.approx.ftz.f32 %f1847, %f1845; mul.ftz.f32 %f5051, %f1844, %f1847; mul.ftz.f32 %f5050, %f1844, %f1846; $L__BB104_129: mul.ftz.f32 %f1848, %f5053, %f5053; fma.rn.ftz.f32 %f1849, %f5052, %f5052, %f1848; add.ftz.f32 %f1850, %f1849, 0f00000000; sqrt.approx.ftz.f32 %f201, %f1850; setp.neu.ftz.f32 %p106, %f201, 0f00000000; @%p106 bra $L__BB104_130; bra.uni $L__BB104_131; $L__BB104_130: rcp.approx.ftz.f32 %f1851, %f201; mul.ftz.f32 %f5053, %f5053, %f1851; mul.ftz.f32 %f5052, %f5052, %f1851; $L__BB104_131: setp.eq.s32 %p512, %r5, 27; and.pred %p511, %p512, %p1; mul.ftz.f32 %f1852, %f5050, %f5052; mul.ftz.f32 %f1853, %f5051, %f5053; sub.ftz.f32 %f1854, %f1852, %f1853; mul.ftz.f32 %f1855, %f1421, %f1854; mul.ftz.f32 %f1856, %f5051, %f5052; fma.rn.ftz.f32 %f1857, %f5050, %f5053, %f1856; mul.ftz.f32 %f1858, %f1422, %f1857; neg.ftz.f32 %f1859, %f1855; neg.ftz.f32 %f1860, %f1858; mul.ftz.f32 %f1861, %f1858, %f1858; fma.rn.ftz.f32 %f1862, %f1855, %f1855, %f1861; add.ftz.f32 %f1863, %f1862, 0f3F800000; mov.f32 %f1864, 0f3F800000; sqrt.approx.ftz.f32 %f1865, %f1863; div.approx.ftz.f32 %f211, %f1859, %f1865; div.approx.ftz.f32 %f212, %f1860, %f1865; div.approx.ftz.f32 %f213, %f1864, %f1865; setp.ne.s32 %p107, %r8, 0; not.pred %p108, %p511; or.pred %p109, %p108, %p107; @%p109 bra $L__BB104_156; and.b16 %rs15, %rs34, 255; mov.f32 %f5057, 0f00000000; setp.eq.s16 %p110, %rs15, 0; @%p110 bra $L__BB104_143; @%p84 bra $L__BB104_153; mul.ftz.f32 %f1869, %f5102, %f5106; setp.lt.ftz.f32 %p113, %f5102, 0fBF7FF972; selp.f32 %f214, 0f3F800000, %f1869, %p113; @%p50 bra $L__BB104_139; bra.uni $L__BB104_135; $L__BB104_139: mul.ftz.f32 %f4864, %f20, %f5011; mul.ftz.f32 %f4863, %f20, %f5010; mul.ftz.f32 %f4862, %f20, %f5009; neg.ftz.f32 %f4861, %f5102; neg.ftz.f32 %f4860, %f5100; neg.ftz.f32 %f4859, %f5101; sub.ftz.f32 %f1911, %f4862, %f5100; sub.ftz.f32 %f1912, %f4863, %f5101; mul.ftz.f32 %f1913, %f1912, %f1912; fma.rn.ftz.f32 %f1914, %f1911, %f1911, %f1913; sub.ftz.f32 %f1915, %f4864, %f5102; fma.rn.ftz.f32 %f1916, %f1915, %f1915, %f1914; sqrt.approx.ftz.f32 %f1917, %f1916; div.approx.ftz.f32 %f1918, %f1911, %f1917; div.approx.ftz.f32 %f1919, %f1912, %f1917; div.approx.ftz.f32 %f1920, %f1915, %f1917; setp.lt.ftz.f32 %p116, %f1920, 0f00000000; neg.ftz.f32 %f1921, %f1918; neg.ftz.f32 %f1922, %f1919; neg.ftz.f32 %f1923, %f1920; selp.f32 %f1924, %f1921, %f1918, %p116; selp.f32 %f1925, %f1922, %f1919, %p116; selp.f32 %f224, %f1923, %f1920, %p116; mul.ftz.f32 %f1926, %f1925, %f4859; fma.rn.ftz.f32 %f1927, %f1924, %f4860, %f1926; fma.rn.ftz.f32 %f225, %f224, %f4861, %f1927; mul.ftz.f32 %f1928, %f5010, %f1925; fma.rn.ftz.f32 %f1929, %f5009, %f1924, %f1928; fma.rn.ftz.f32 %f226, %f5011, %f224, %f1929; setp.lt.ftz.f32 %p117, %f225, 0f00000000; @%p117 bra $L__BB104_153; abs.ftz.f32 %f228, %f225; fma.rn.ftz.f32 %f229, %f228, %f228, %f37; setp.leu.ftz.f32 %p118, %f229, 0f00000000; mov.f32 %f1930, 0f3F800000; mov.f32 %f5055, %f1930; @%p118 bra $L__BB104_142; sqrt.approx.ftz.f32 %f1931, %f229; sub.ftz.f32 %f1932, %f1931, %f228; add.ftz.f32 %f1933, %f228, %f1931; div.approx.ftz.f32 %f1934, %f1932, %f1933; fma.rn.ftz.f32 %f1935, %f228, %f1933, 0fBF800000; fma.rn.ftz.f32 %f1936, %f228, %f1932, 0f3F800000; div.approx.ftz.f32 %f1937, %f1935, %f1936; mul.ftz.f32 %f1938, %f1934, 0f3F000000; mul.ftz.f32 %f1939, %f1934, %f1938; fma.rn.ftz.f32 %f1940, %f1937, %f1937, 0f3F800000; mul.ftz.f32 %f5055, %f1939, %f1940; $L__BB104_142: fma.rn.ftz.f32 %f4905, %f20, %f226, %f225; mov.f32 %f1941, 0f00000000; max.ftz.f32 %f1942, %f1941, %f225; sub.ftz.f32 %f1944, %f1930, %f5055; mul.ftz.f32 %f1945, %f1944, %f1942; neg.ftz.f32 %f1946, %f226; max.ftz.f32 %f1947, %f1941, %f1946; mul.ftz.f32 %f1948, %f1945, %f1947; mul.ftz.f32 %f1949, %f224, %f224; sub.ftz.f32 %f1950, %f1930, %f1949; fma.rn.ftz.f32 %f1951, %f33, %f1949, %f1950; mul.ftz.f32 %f1952, %f1951, 0f40490FDB; mul.ftz.f32 %f1953, %f1951, %f1952; max.ftz.f32 %f1955, %f1953, %f1505; div.approx.ftz.f32 %f1956, %f33, %f1955; mul.ftz.f32 %f1957, %f1948, %f1956; mul.ftz.f32 %f1958, %f214, %f4905; mul.ftz.f32 %f1959, %f4905, %f1958; div.approx.ftz.f32 %f5057, %f1957, %f1959; bra.uni $L__BB104_153; $L__BB104_112: setp.lt.ftz.f32 %p85, %f5102, 0fBF7FF972; @%p85 bra $L__BB104_119; bra.uni $L__BB104_113; $L__BB104_119: mul.ftz.f32 %f5048, %f5048, %f167; cvt.ftz.sat.f32.f32 %f1784, %f5048; fma.rn.ftz.f32 %f5049, %f1784, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p95, %f5102, 0f3727C5AC; mov.pred %p528, 0; mov.f32 %f5047, 0f00000000; @%p95 bra $L__BB104_121; lg2.approx.ftz.f32 %f1785, %f5048; mul.ftz.f32 %f1786, %f5106, %f1785; ex2.approx.ftz.f32 %f5047, %f1786; bra.uni $L__BB104_121; $L__BB104_113: abs.ftz.f32 %f1768, %f5102; setp.ltu.ftz.f32 %p87, %f1768, 0f38D1B717; mov.pred %p86, 0; mov.pred %p528, %p86; @%p87 bra $L__BB104_121; mov.f32 %f5046, 0f3F800000; sub.ftz.f32 %f1770, %f5046, %f5047; setp.lt.ftz.f32 %p89, %f1770, %f167; mov.pred %p528, %p83; @%p89 bra $L__BB104_121; setp.ge.ftz.f32 %p90, %f5106, 0f00000000; @%p90 bra $L__BB104_117; mov.f32 %f1772, 0f3F800000; sub.ftz.f32 %f1773, %f1772, %f167; lg2.approx.ftz.f32 %f1774, %f1773; mov.f32 %f1775, 0fBF800000; div.approx.ftz.f32 %f1776, %f1775, %f5106; mul.ftz.f32 %f1777, %f1776, %f1774; ex2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f5046, %f5048, %f1778; $L__BB104_117: cvt.ftz.sat.f32.f32 %f1780, %f5046; fma.rn.ftz.f32 %f5049, %f1780, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p92, %f5102, 0f3727C5AC; mov.f32 %f5047, 0f00000000; mov.f32 %f5048, %f5046; mov.pred %p528, %p86; @%p92 bra $L__BB104_121; lg2.approx.ftz.f32 %f1781, %f5046; mul.ftz.f32 %f1782, %f5106, %f1781; ex2.approx.ftz.f32 %f5047, %f1782; mov.f32 %f5048, %f5046; mov.pred %p528, %p86; bra.uni $L__BB104_121; $L__BB104_124: neg.ftz.f32 %f4892, %f5102; div.approx.ftz.f32 %f4891, %f4892, %f1794; add.ftz.f32 %f1798, %f4891, 0f3F800000; mul.ftz.f32 %f186, %f1798, 0f3F000000; setp.lt.ftz.f32 %p99, %f186, 0f38D1B717; mov.f32 %f5050, 0f00000000; mov.f32 %f5051, %f5050; @%p99 bra $L__BB104_129; neg.ftz.f32 %f4894, %f5102; div.approx.ftz.f32 %f4893, %f4894, %f1794; add.ftz.f32 %f1801, %f179, %f179; mul.ftz.f32 %f1802, %f1801, %f186; div.approx.ftz.f32 %f1803, %f1802, %f4893; add.ftz.f32 %f187, %f1803, 0fBF800000; mul.ftz.f32 %f188, %f187, %f187; add.ftz.f32 %f189, %f188, 0fBF800000; abs.ftz.f32 %f1804, %f189; setp.lt.ftz.f32 %p100, %f1804, 0f33D6BF95; mov.f32 %f5051, %f5050; @%p100 bra $L__BB104_129; neg.ftz.f32 %f4896, %f5102; div.approx.ftz.f32 %f4895, %f4896, %f1794; mul.ftz.f32 %f1805, %f4895, %f4895; mov.f32 %f1806, 0f3F800000; sub.ftz.f32 %f1807, %f1806, %f1805; mov.f32 %f1808, 0f00000000; max.ftz.f32 %f1809, %f1807, %f1808; sqrt.approx.ftz.f32 %f1810, %f1809; div.approx.ftz.f32 %f1811, %f1810, %f4895; rcp.approx.ftz.f32 %f1812, %f189; mul.ftz.f32 %f1813, %f1811, %f1811; mul.ftz.f32 %f1814, %f1813, %f1812; mul.ftz.f32 %f1815, %f1812, %f1814; sub.ftz.f32 %f1816, %f188, %f1813; mul.ftz.f32 %f1817, %f1816, %f1812; sub.ftz.f32 %f1818, %f1815, %f1817; max.ftz.f32 %f1819, %f1818, %f1808; sqrt.approx.ftz.f32 %f1820, %f1819; mul.ftz.f32 %f1821, %f1811, %f1812; add.ftz.f32 %f1822, %f1821, %f1820; rcp.approx.ftz.f32 %f1823, %f1811; setp.gt.ftz.f32 %p101, %f1822, %f1823; setp.lt.ftz.f32 %p102, %f187, 0f00000000; or.pred %p103, %p102, %p101; sub.ftz.f32 %f1824, %f1821, %f1820; mov.f32 %f1825, 0f3F000000; sub.ftz.f32 %f1826, %f1825, %f178; add.ftz.f32 %f1827, %f178, 0fBF000000; setp.ge.ftz.f32 %p104, %f178, 0f3F000000; selp.f32 %f1828, %f1827, %f1826, %p104; add.ftz.f32 %f1829, %f1828, %f1828; fma.rn.ftz.f32 %f1830, %f1829, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f1831, %f1829, %f1830, 0f3EED4413; mul.ftz.f32 %f1832, %f1829, %f1831; fma.rn.ftz.f32 %f1833, %f1829, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f1834, %f1829, %f1833, 0fBF800000; fma.rn.ftz.f32 %f1835, %f1829, %f1834, 0f3F191676; div.approx.ftz.f32 %f1836, %f1832, %f1835; selp.f32 %f5050, %f1824, %f1822, %p103; fma.rn.ftz.f32 %f1837, %f5050, %f5050, 0f3F800000; sqrt.approx.ftz.f32 %f1838, %f1837; mul.ftz.f32 %f5051, %f1836, %f1838; setp.ltu.ftz.f32 %p105, %f178, 0f3F000000; @%p105 bra $L__BB104_127; bra.uni $L__BB104_129; $L__BB104_127: neg.ftz.f32 %f5051, %f5051; bra.uni $L__BB104_129; $L__BB104_143: @%p84 bra $L__BB104_153; mul.ftz.f32 %f1963, %f5102, %f5106; setp.lt.ftz.f32 %p121, %f5102, 0fBF7FF972; selp.f32 %f243, 0f3F800000, %f1963, %p121; @%p50 bra $L__BB104_149; bra.uni $L__BB104_145; $L__BB104_149: neg.ftz.f32 %f4877, %f5011; neg.ftz.f32 %f4876, %f5010; neg.ftz.f32 %f4875, %f5009; sub.ftz.f32 %f2023, %f4875, %f5100; sub.ftz.f32 %f2025, %f4876, %f5101; sub.ftz.f32 %f2027, %f4877, %f5102; mul.ftz.f32 %f2028, %f2025, %f2025; fma.rn.ftz.f32 %f2029, %f2023, %f2023, %f2028; fma.rn.ftz.f32 %f2030, %f2027, %f2027, %f2029; sqrt.approx.ftz.f32 %f2031, %f2030; div.approx.ftz.f32 %f255, %f2027, %f2031; setp.lt.ftz.f32 %p125, %f255, 0f00000000; @%p125 bra $L__BB104_153; neg.ftz.f32 %f4915, %f5011; sub.ftz.f32 %f4914, %f4915, %f5102; div.approx.ftz.f32 %f4913, %f4914, %f2031; neg.ftz.f32 %f4912, %f5009; sub.ftz.f32 %f4911, %f4912, %f5100; div.approx.ftz.f32 %f4910, %f4911, %f2031; neg.ftz.f32 %f4909, %f5010; sub.ftz.f32 %f4908, %f4909, %f5101; div.approx.ftz.f32 %f4907, %f4908, %f2031; neg.ftz.f32 %f4880, %f5102; neg.ftz.f32 %f4879, %f5100; neg.ftz.f32 %f4878, %f5101; mul.ftz.f32 %f2033, %f4907, %f4878; fma.rn.ftz.f32 %f2034, %f4910, %f4879, %f2033; fma.rn.ftz.f32 %f256, %f4913, %f4880, %f2034; abs.ftz.f32 %f257, %f256; fma.rn.ftz.f32 %f258, %f257, %f257, %f152; setp.leu.ftz.f32 %p126, %f258, 0f00000000; mov.f32 %f2032, 0f3F800000; mov.f32 %f5059, %f2032; @%p126 bra $L__BB104_152; sqrt.approx.ftz.f32 %f2035, %f258; sub.ftz.f32 %f2036, %f2035, %f257; add.ftz.f32 %f2037, %f257, %f2035; div.approx.ftz.f32 %f2038, %f2036, %f2037; fma.rn.ftz.f32 %f2039, %f257, %f2037, 0fBF800000; fma.rn.ftz.f32 %f2040, %f257, %f2036, 0f3F800000; div.approx.ftz.f32 %f2041, %f2039, %f2040; mul.ftz.f32 %f2042, %f2038, 0f3F000000; mul.ftz.f32 %f2043, %f2038, %f2042; fma.rn.ftz.f32 %f2044, %f2041, %f2041, 0f3F800000; mul.ftz.f32 %f5059, %f2043, %f2044; $L__BB104_152: neg.ftz.f32 %f4918, %f5011; sub.ftz.f32 %f4917, %f4918, %f5102; div.approx.ftz.f32 %f4916, %f4917, %f2031; mov.f32 %f2045, 0f00000000; max.ftz.f32 %f2046, %f2045, %f256; mul.ftz.f32 %f2047, %f5059, %f2046; mul.ftz.f32 %f2048, %f4916, %f4916; sub.ftz.f32 %f2050, %f2032, %f2048; fma.rn.ftz.f32 %f2051, %f33, %f2048, %f2050; mul.ftz.f32 %f2052, %f2051, 0f40490FDB; mul.ftz.f32 %f2053, %f2051, %f2052; max.ftz.f32 %f2055, %f2053, %f1505; div.approx.ftz.f32 %f2056, %f33, %f2055; mul.ftz.f32 %f2057, %f2047, %f2056; mul.ftz.f32 %f2058, %f2057, 0f3E800000; mul.ftz.f32 %f2059, %f243, %f256; div.approx.ftz.f32 %f5057, %f2058, %f2059; bra.uni $L__BB104_153; $L__BB104_135: sub.ftz.f32 %f1873, %f5009, %f5100; sub.ftz.f32 %f1874, %f5010, %f5101; mul.ftz.f32 %f1875, %f1874, %f1874; fma.rn.ftz.f32 %f1876, %f1873, %f1873, %f1875; sub.ftz.f32 %f1877, %f5011, %f5102; fma.rn.ftz.f32 %f1878, %f1877, %f1877, %f1876; sqrt.approx.ftz.f32 %f1879, %f1878; div.approx.ftz.f32 %f217, %f1877, %f1879; setp.lt.ftz.f32 %p114, %f217, 0f00000000; @%p114 bra $L__BB104_153; sub.ftz.f32 %f4902, %f5011, %f5102; div.approx.ftz.f32 %f4901, %f4902, %f1879; sub.ftz.f32 %f4900, %f5009, %f5100; div.approx.ftz.f32 %f4899, %f4900, %f1879; sub.ftz.f32 %f4898, %f5010, %f5101; div.approx.ftz.f32 %f4897, %f4898, %f1879; neg.ftz.f32 %f4858, %f5102; neg.ftz.f32 %f4857, %f5100; neg.ftz.f32 %f4856, %f5101; mul.ftz.f32 %f1881, %f4897, %f4856; fma.rn.ftz.f32 %f1882, %f4899, %f4857, %f1881; fma.rn.ftz.f32 %f218, %f4901, %f4858, %f1882; abs.ftz.f32 %f219, %f218; fma.rn.ftz.f32 %f220, %f219, %f219, %f37; setp.leu.ftz.f32 %p115, %f220, 0f00000000; mov.f32 %f1880, 0f3F800000; mov.f32 %f5054, %f1880; @%p115 bra $L__BB104_138; sqrt.approx.ftz.f32 %f1883, %f220; sub.ftz.f32 %f1884, %f1883, %f219; add.ftz.f32 %f1885, %f219, %f1883; div.approx.ftz.f32 %f1886, %f1884, %f1885; fma.rn.ftz.f32 %f1887, %f219, %f1885, 0fBF800000; fma.rn.ftz.f32 %f1888, %f219, %f1884, 0f3F800000; div.approx.ftz.f32 %f1889, %f1887, %f1888; mul.ftz.f32 %f1890, %f1886, 0f3F000000; mul.ftz.f32 %f1891, %f1886, %f1890; fma.rn.ftz.f32 %f1892, %f1889, %f1889, 0f3F800000; mul.ftz.f32 %f5054, %f1891, %f1892; $L__BB104_138: sub.ftz.f32 %f4904, %f5011, %f5102; div.approx.ftz.f32 %f4903, %f4904, %f1879; mov.f32 %f1893, 0f00000000; max.ftz.f32 %f1894, %f1893, %f218; mul.ftz.f32 %f1895, %f5054, %f1894; mul.ftz.f32 %f1896, %f4903, %f4903; sub.ftz.f32 %f1898, %f1880, %f1896; fma.rn.ftz.f32 %f1899, %f33, %f1896, %f1898; mul.ftz.f32 %f1900, %f1899, 0f40490FDB; mul.ftz.f32 %f1901, %f1899, %f1900; max.ftz.f32 %f1903, %f1901, %f1505; div.approx.ftz.f32 %f1904, %f33, %f1903; mul.ftz.f32 %f1905, %f1895, %f1904; mul.ftz.f32 %f1906, %f1905, 0f3E800000; mul.ftz.f32 %f1907, %f214, %f218; div.approx.ftz.f32 %f5057, %f1906, %f1907; bra.uni $L__BB104_153; $L__BB104_145: neg.ftz.f32 %f4874, %f5009; neg.ftz.f32 %f4873, %f5011; mul.ftz.f32 %f4872, %f145, %f4873; neg.ftz.f32 %f4871, %f5010; mul.ftz.f32 %f4870, %f145, %f4871; neg.ftz.f32 %f4869, %f5009; mul.ftz.f32 %f4868, %f145, %f4869; neg.ftz.f32 %f4867, %f5102; neg.ftz.f32 %f4866, %f5100; neg.ftz.f32 %f4865, %f5101; sub.ftz.f32 %f1967, %f4868, %f5100; sub.ftz.f32 %f1968, %f4870, %f5101; mul.ftz.f32 %f1969, %f1968, %f1968; fma.rn.ftz.f32 %f1970, %f1967, %f1967, %f1969; sub.ftz.f32 %f1971, %f4872, %f5102; fma.rn.ftz.f32 %f1972, %f1971, %f1971, %f1970; sqrt.approx.ftz.f32 %f1973, %f1972; div.approx.ftz.f32 %f1974, %f1967, %f1973; div.approx.ftz.f32 %f1975, %f1968, %f1973; div.approx.ftz.f32 %f1976, %f1971, %f1973; setp.lt.ftz.f32 %p122, %f1976, 0f00000000; neg.ftz.f32 %f1977, %f1974; neg.ftz.f32 %f1978, %f1975; neg.ftz.f32 %f1979, %f1976; selp.f32 %f1980, %f1977, %f1974, %p122; selp.f32 %f1981, %f1978, %f1975, %p122; selp.f32 %f244, %f1979, %f1976, %p122; mul.ftz.f32 %f1982, %f1981, %f4865; fma.rn.ftz.f32 %f1983, %f1980, %f4866, %f1982; fma.rn.ftz.f32 %f245, %f244, %f4867, %f1983; mul.ftz.f32 %f1985, %f1980, %f4869; mul.ftz.f32 %f1986, %f1981, %f5010; sub.ftz.f32 %f1987, %f1985, %f1986; mul.ftz.f32 %f1988, %f244, %f5011; sub.ftz.f32 %f246, %f1987, %f1988; setp.lt.ftz.f32 %p123, %f245, 0f00000000; @%p123 bra $L__BB104_153; abs.ftz.f32 %f248, %f245; fma.rn.ftz.f32 %f249, %f248, %f248, %f152; setp.leu.ftz.f32 %p124, %f249, 0f00000000; mov.f32 %f1989, 0f3F800000; mov.f32 %f5058, %f1989; @%p124 bra $L__BB104_148; sqrt.approx.ftz.f32 %f1990, %f249; sub.ftz.f32 %f1991, %f1990, %f248; add.ftz.f32 %f1992, %f248, %f1990; div.approx.ftz.f32 %f1993, %f1991, %f1992; fma.rn.ftz.f32 %f1994, %f248, %f1992, 0fBF800000; fma.rn.ftz.f32 %f1995, %f248, %f1991, 0f3F800000; div.approx.ftz.f32 %f1996, %f1994, %f1995; mul.ftz.f32 %f1997, %f1993, 0f3F000000; mul.ftz.f32 %f1998, %f1993, %f1997; fma.rn.ftz.f32 %f1999, %f1996, %f1996, 0f3F800000; mul.ftz.f32 %f5058, %f1998, %f1999; $L__BB104_148: fma.rn.ftz.f32 %f4906, %f145, %f246, %f245; mov.f32 %f2000, 0f00000000; max.ftz.f32 %f2001, %f2000, %f245; sub.ftz.f32 %f2003, %f1989, %f5058; mul.ftz.f32 %f2004, %f2003, %f2001; neg.ftz.f32 %f2005, %f246; max.ftz.f32 %f2006, %f2000, %f2005; mul.ftz.f32 %f2007, %f2004, %f2006; mul.ftz.f32 %f2008, %f244, %f244; sub.ftz.f32 %f2009, %f1989, %f2008; fma.rn.ftz.f32 %f2010, %f33, %f2008, %f2009; mul.ftz.f32 %f2011, %f2010, 0f40490FDB; mul.ftz.f32 %f2012, %f2010, %f2011; max.ftz.f32 %f2014, %f2012, %f1505; div.approx.ftz.f32 %f2015, %f33, %f2014; mul.ftz.f32 %f2016, %f2007, %f2015; mul.ftz.f32 %f2017, %f243, %f4906; mul.ftz.f32 %f2018, %f4906, %f2017; div.approx.ftz.f32 %f5057, %f2016, %f2018; $L__BB104_153: neg.ftz.f32 %f4826, %f5011; selp.f32 %f4825, %f4826, %f5011, %p50; setp.gt.ftz.f32 %p127, %f4825, 0f3F7FF972; setp.lt.ftz.f32 %p128, %f4825, 0f3727C5AC; or.pred %p129, %p127, %p128; selp.f32 %f5065, 0f3F800000, 0f00000000, %p127; @%p129 bra $L__BB104_155; selp.u16 %rs26, 1, 0, %p1; and.b16 %rs25, %rs34, 255; setp.eq.s16 %p130, %rs25, %rs26; neg.ftz.f32 %f2060, %f5049; selp.f32 %f2061, %f5049, %f2060, %p130; add.ftz.f32 %f2062, %f2061, 0f3F800000; mul.ftz.f32 %f2063, %f2062, 0f3F000000; cvt.ftz.sat.f32.f32 %f2064, %f2063; lg2.approx.ftz.f32 %f2065, %f2064; mul.ftz.f32 %f2066, %f5019, %f2065; ex2.approx.ftz.f32 %f5065, %f2066; $L__BB104_155: mul.ftz.f32 %f2067, %f5103, %f5057; mul.ftz.f32 %f5290, %f2067, %f5065; mul.ftz.f32 %f2068, %f5104, %f5057; mul.ftz.f32 %f5289, %f2068, %f5065; mul.ftz.f32 %f2069, %f5105, %f5057; mul.ftz.f32 %f5288, %f2069, %f5065; $L__BB104_156: setp.eq.s32 %p131, %r8, 0; @%p131 bra $L__BB104_181; and.b16 %rs17, %rs34, 255; mov.f32 %f5072, 0f00000000; setp.eq.s16 %p132, %rs17, 0; @%p132 bra $L__BB104_168; @%p84 bra $L__BB104_178; mul.ftz.f32 %f2073, %f5102, %f5106; setp.lt.ftz.f32 %p135, %f5102, 0fBF7FF972; selp.f32 %f284, 0f3F800000, %f2073, %p135; @%p50 bra $L__BB104_164; bra.uni $L__BB104_160; $L__BB104_164: mul.ftz.f32 %f4837, %f20, %f5011; mul.ftz.f32 %f4836, %f20, %f5010; mul.ftz.f32 %f4835, %f20, %f5009; neg.ftz.f32 %f4834, %f5102; neg.ftz.f32 %f4833, %f5100; neg.ftz.f32 %f4832, %f5101; sub.ftz.f32 %f2115, %f4835, %f5100; sub.ftz.f32 %f2116, %f4836, %f5101; mul.ftz.f32 %f2117, %f2116, %f2116; fma.rn.ftz.f32 %f2118, %f2115, %f2115, %f2117; sub.ftz.f32 %f2119, %f4837, %f5102; fma.rn.ftz.f32 %f2120, %f2119, %f2119, %f2118; sqrt.approx.ftz.f32 %f2121, %f2120; div.approx.ftz.f32 %f2122, %f2115, %f2121; div.approx.ftz.f32 %f2123, %f2116, %f2121; div.approx.ftz.f32 %f2124, %f2119, %f2121; setp.lt.ftz.f32 %p138, %f2124, 0f00000000; neg.ftz.f32 %f2125, %f2122; neg.ftz.f32 %f2126, %f2123; neg.ftz.f32 %f2127, %f2124; selp.f32 %f2128, %f2125, %f2122, %p138; selp.f32 %f2129, %f2126, %f2123, %p138; selp.f32 %f294, %f2127, %f2124, %p138; mul.ftz.f32 %f2130, %f2129, %f4832; fma.rn.ftz.f32 %f2131, %f2128, %f4833, %f2130; fma.rn.ftz.f32 %f295, %f294, %f4834, %f2131; mul.ftz.f32 %f2132, %f5010, %f2129; fma.rn.ftz.f32 %f2133, %f5009, %f2128, %f2132; fma.rn.ftz.f32 %f296, %f5011, %f294, %f2133; setp.lt.ftz.f32 %p139, %f295, 0f00000000; @%p139 bra $L__BB104_178; abs.ftz.f32 %f298, %f295; fma.rn.ftz.f32 %f299, %f298, %f298, %f37; setp.leu.ftz.f32 %p140, %f299, 0f00000000; mov.f32 %f2134, 0f3F800000; mov.f32 %f5070, %f2134; @%p140 bra $L__BB104_167; sqrt.approx.ftz.f32 %f2135, %f299; sub.ftz.f32 %f2136, %f2135, %f298; add.ftz.f32 %f2137, %f298, %f2135; div.approx.ftz.f32 %f2138, %f2136, %f2137; fma.rn.ftz.f32 %f2139, %f298, %f2137, 0fBF800000; fma.rn.ftz.f32 %f2140, %f298, %f2136, 0f3F800000; div.approx.ftz.f32 %f2141, %f2139, %f2140; mul.ftz.f32 %f2142, %f2138, 0f3F000000; mul.ftz.f32 %f2143, %f2138, %f2142; fma.rn.ftz.f32 %f2144, %f2141, %f2141, 0f3F800000; mul.ftz.f32 %f5070, %f2143, %f2144; $L__BB104_167: fma.rn.ftz.f32 %f4927, %f20, %f296, %f295; mov.f32 %f2145, 0f00000000; max.ftz.f32 %f2146, %f2145, %f295; sub.ftz.f32 %f2148, %f2134, %f5070; mul.ftz.f32 %f2149, %f2148, %f2146; neg.ftz.f32 %f2150, %f296; max.ftz.f32 %f2151, %f2145, %f2150; mul.ftz.f32 %f2152, %f2149, %f2151; mul.ftz.f32 %f2153, %f294, %f294; sub.ftz.f32 %f2154, %f2134, %f2153; fma.rn.ftz.f32 %f2155, %f33, %f2153, %f2154; mul.ftz.f32 %f2156, %f2155, 0f40490FDB; mul.ftz.f32 %f2157, %f2155, %f2156; max.ftz.f32 %f2159, %f2157, %f1505; div.approx.ftz.f32 %f2160, %f33, %f2159; mul.ftz.f32 %f2161, %f2152, %f2160; mul.ftz.f32 %f2162, %f284, %f4927; mul.ftz.f32 %f2163, %f4927, %f2162; div.approx.ftz.f32 %f5072, %f2161, %f2163; bra.uni $L__BB104_178; $L__BB104_168: @%p84 bra $L__BB104_178; mul.ftz.f32 %f2167, %f5102, %f5106; setp.lt.ftz.f32 %p143, %f5102, 0fBF7FF972; selp.f32 %f313, 0f3F800000, %f2167, %p143; @%p50 bra $L__BB104_174; bra.uni $L__BB104_170; $L__BB104_174: neg.ftz.f32 %f4852, %f5011; neg.ftz.f32 %f4851, %f5010; neg.ftz.f32 %f4850, %f5009; sub.ftz.f32 %f2227, %f4850, %f5100; sub.ftz.f32 %f2229, %f4851, %f5101; sub.ftz.f32 %f2231, %f4852, %f5102; mul.ftz.f32 %f2232, %f2229, %f2229; fma.rn.ftz.f32 %f2233, %f2227, %f2227, %f2232; fma.rn.ftz.f32 %f2234, %f2231, %f2231, %f2233; sqrt.approx.ftz.f32 %f2235, %f2234; div.approx.ftz.f32 %f325, %f2231, %f2235; setp.lt.ftz.f32 %p147, %f325, 0f00000000; @%p147 bra $L__BB104_178; neg.ftz.f32 %f4937, %f5011; sub.ftz.f32 %f4936, %f4937, %f5102; div.approx.ftz.f32 %f4935, %f4936, %f2235; neg.ftz.f32 %f4934, %f5009; sub.ftz.f32 %f4933, %f4934, %f5100; div.approx.ftz.f32 %f4932, %f4933, %f2235; neg.ftz.f32 %f4931, %f5010; sub.ftz.f32 %f4930, %f4931, %f5101; div.approx.ftz.f32 %f4929, %f4930, %f2235; neg.ftz.f32 %f4855, %f5102; neg.ftz.f32 %f4854, %f5100; neg.ftz.f32 %f4853, %f5101; mul.ftz.f32 %f2237, %f4929, %f4853; fma.rn.ftz.f32 %f2238, %f4932, %f4854, %f2237; fma.rn.ftz.f32 %f326, %f4935, %f4855, %f2238; abs.ftz.f32 %f327, %f326; fma.rn.ftz.f32 %f328, %f327, %f327, %f152; setp.leu.ftz.f32 %p148, %f328, 0f00000000; mov.f32 %f2236, 0f3F800000; mov.f32 %f5074, %f2236; @%p148 bra $L__BB104_177; sqrt.approx.ftz.f32 %f2239, %f328; sub.ftz.f32 %f2240, %f2239, %f327; add.ftz.f32 %f2241, %f327, %f2239; div.approx.ftz.f32 %f2242, %f2240, %f2241; fma.rn.ftz.f32 %f2243, %f327, %f2241, 0fBF800000; fma.rn.ftz.f32 %f2244, %f327, %f2240, 0f3F800000; div.approx.ftz.f32 %f2245, %f2243, %f2244; mul.ftz.f32 %f2246, %f2242, 0f3F000000; mul.ftz.f32 %f2247, %f2242, %f2246; fma.rn.ftz.f32 %f2248, %f2245, %f2245, 0f3F800000; mul.ftz.f32 %f5074, %f2247, %f2248; $L__BB104_177: neg.ftz.f32 %f4940, %f5011; sub.ftz.f32 %f4939, %f4940, %f5102; div.approx.ftz.f32 %f4938, %f4939, %f2235; mov.f32 %f2249, 0f00000000; max.ftz.f32 %f2250, %f2249, %f326; mul.ftz.f32 %f2251, %f5074, %f2250; mul.ftz.f32 %f2252, %f4938, %f4938; sub.ftz.f32 %f2254, %f2236, %f2252; fma.rn.ftz.f32 %f2255, %f33, %f2252, %f2254; mul.ftz.f32 %f2256, %f2255, 0f40490FDB; mul.ftz.f32 %f2257, %f2255, %f2256; max.ftz.f32 %f2259, %f2257, %f1505; div.approx.ftz.f32 %f2260, %f33, %f2259; mul.ftz.f32 %f2261, %f2251, %f2260; mul.ftz.f32 %f2262, %f2261, 0f3E800000; mul.ftz.f32 %f2263, %f313, %f326; div.approx.ftz.f32 %f5072, %f2262, %f2263; bra.uni $L__BB104_178; $L__BB104_160: sub.ftz.f32 %f2077, %f5009, %f5100; sub.ftz.f32 %f2078, %f5010, %f5101; mul.ftz.f32 %f2079, %f2078, %f2078; fma.rn.ftz.f32 %f2080, %f2077, %f2077, %f2079; sub.ftz.f32 %f2081, %f5011, %f5102; fma.rn.ftz.f32 %f2082, %f2081, %f2081, %f2080; sqrt.approx.ftz.f32 %f2083, %f2082; div.approx.ftz.f32 %f287, %f2081, %f2083; setp.lt.ftz.f32 %p136, %f287, 0f00000000; @%p136 bra $L__BB104_178; sub.ftz.f32 %f4924, %f5011, %f5102; div.approx.ftz.f32 %f4923, %f4924, %f2083; sub.ftz.f32 %f4922, %f5009, %f5100; div.approx.ftz.f32 %f4921, %f4922, %f2083; sub.ftz.f32 %f4920, %f5010, %f5101; div.approx.ftz.f32 %f4919, %f4920, %f2083; mov.f32 %f5069, 0f3F800000; neg.ftz.f32 %f4829, %f5102; neg.ftz.f32 %f4828, %f5100; neg.ftz.f32 %f4827, %f5101; mul.ftz.f32 %f2085, %f4919, %f4827; fma.rn.ftz.f32 %f2086, %f4921, %f4828, %f2085; fma.rn.ftz.f32 %f288, %f4923, %f4829, %f2086; abs.ftz.f32 %f289, %f288; fma.rn.ftz.f32 %f290, %f289, %f289, %f37; setp.leu.ftz.f32 %p137, %f290, 0f00000000; @%p137 bra $L__BB104_163; sqrt.approx.ftz.f32 %f2087, %f290; sub.ftz.f32 %f2088, %f2087, %f289; add.ftz.f32 %f2089, %f289, %f2087; div.approx.ftz.f32 %f2090, %f2088, %f2089; fma.rn.ftz.f32 %f2091, %f289, %f2089, 0fBF800000; fma.rn.ftz.f32 %f2092, %f289, %f2088, 0f3F800000; div.approx.ftz.f32 %f2093, %f2091, %f2092; mul.ftz.f32 %f2094, %f2090, 0f3F000000; mul.ftz.f32 %f2095, %f2090, %f2094; fma.rn.ftz.f32 %f2096, %f2093, %f2093, 0f3F800000; mul.ftz.f32 %f5069, %f2095, %f2096; $L__BB104_163: sub.ftz.f32 %f4926, %f5011, %f5102; div.approx.ftz.f32 %f4925, %f4926, %f2083; mov.f32 %f4831, 0f3F800000; mov.f32 %f2097, 0f00000000; max.ftz.f32 %f2098, %f2097, %f288; mul.ftz.f32 %f2099, %f5069, %f2098; mul.ftz.f32 %f2100, %f4925, %f4925; sub.ftz.f32 %f2102, %f4831, %f2100; fma.rn.ftz.f32 %f2103, %f33, %f2100, %f2102; mul.ftz.f32 %f2104, %f2103, 0f40490FDB; mul.ftz.f32 %f2105, %f2103, %f2104; max.ftz.f32 %f2107, %f2105, %f1505; div.approx.ftz.f32 %f2108, %f33, %f2107; mul.ftz.f32 %f2109, %f2099, %f2108; mul.ftz.f32 %f2110, %f2109, 0f3E800000; mul.ftz.f32 %f2111, %f284, %f288; div.approx.ftz.f32 %f5072, %f2110, %f2111; bra.uni $L__BB104_178; $L__BB104_170: neg.ftz.f32 %f4849, %f5009; neg.ftz.f32 %f4848, %f5011; mul.ftz.f32 %f4847, %f145, %f4848; neg.ftz.f32 %f4846, %f5010; mul.ftz.f32 %f4845, %f145, %f4846; neg.ftz.f32 %f4844, %f5009; mul.ftz.f32 %f4843, %f145, %f4844; neg.ftz.f32 %f4842, %f5102; neg.ftz.f32 %f4841, %f5100; neg.ftz.f32 %f4840, %f5101; sub.ftz.f32 %f2171, %f4843, %f5100; sub.ftz.f32 %f2172, %f4845, %f5101; mul.ftz.f32 %f2173, %f2172, %f2172; fma.rn.ftz.f32 %f2174, %f2171, %f2171, %f2173; sub.ftz.f32 %f2175, %f4847, %f5102; fma.rn.ftz.f32 %f2176, %f2175, %f2175, %f2174; sqrt.approx.ftz.f32 %f2177, %f2176; div.approx.ftz.f32 %f2178, %f2171, %f2177; div.approx.ftz.f32 %f2179, %f2172, %f2177; div.approx.ftz.f32 %f2180, %f2175, %f2177; setp.lt.ftz.f32 %p144, %f2180, 0f00000000; neg.ftz.f32 %f2181, %f2178; neg.ftz.f32 %f2182, %f2179; neg.ftz.f32 %f2183, %f2180; selp.f32 %f2184, %f2181, %f2178, %p144; selp.f32 %f2185, %f2182, %f2179, %p144; selp.f32 %f314, %f2183, %f2180, %p144; mul.ftz.f32 %f2186, %f2185, %f4840; fma.rn.ftz.f32 %f2187, %f2184, %f4841, %f2186; fma.rn.ftz.f32 %f315, %f314, %f4842, %f2187; mul.ftz.f32 %f2189, %f2184, %f4844; mul.ftz.f32 %f2190, %f2185, %f5010; sub.ftz.f32 %f2191, %f2189, %f2190; mul.ftz.f32 %f2192, %f314, %f5011; sub.ftz.f32 %f316, %f2191, %f2192; setp.lt.ftz.f32 %p145, %f315, 0f00000000; @%p145 bra $L__BB104_178; abs.ftz.f32 %f318, %f315; fma.rn.ftz.f32 %f319, %f318, %f318, %f152; setp.leu.ftz.f32 %p146, %f319, 0f00000000; mov.f32 %f2193, 0f3F800000; mov.f32 %f5073, %f2193; @%p146 bra $L__BB104_173; sqrt.approx.ftz.f32 %f2194, %f319; sub.ftz.f32 %f2195, %f2194, %f318; add.ftz.f32 %f2196, %f318, %f2194; div.approx.ftz.f32 %f2197, %f2195, %f2196; fma.rn.ftz.f32 %f2198, %f318, %f2196, 0fBF800000; fma.rn.ftz.f32 %f2199, %f318, %f2195, 0f3F800000; div.approx.ftz.f32 %f2200, %f2198, %f2199; mul.ftz.f32 %f2201, %f2197, 0f3F000000; mul.ftz.f32 %f2202, %f2197, %f2201; fma.rn.ftz.f32 %f2203, %f2200, %f2200, 0f3F800000; mul.ftz.f32 %f5073, %f2202, %f2203; $L__BB104_173: fma.rn.ftz.f32 %f4928, %f145, %f316, %f315; mov.f32 %f2204, 0f00000000; max.ftz.f32 %f2205, %f2204, %f315; sub.ftz.f32 %f2207, %f2193, %f5073; mul.ftz.f32 %f2208, %f2207, %f2205; neg.ftz.f32 %f2209, %f316; max.ftz.f32 %f2210, %f2204, %f2209; mul.ftz.f32 %f2211, %f2208, %f2210; mul.ftz.f32 %f2212, %f314, %f314; sub.ftz.f32 %f2213, %f2193, %f2212; fma.rn.ftz.f32 %f2214, %f33, %f2212, %f2213; mul.ftz.f32 %f2215, %f2214, 0f40490FDB; mul.ftz.f32 %f2216, %f2214, %f2215; max.ftz.f32 %f2218, %f2216, %f1505; div.approx.ftz.f32 %f2219, %f33, %f2218; mul.ftz.f32 %f2220, %f2211, %f2219; mul.ftz.f32 %f2221, %f313, %f4928; mul.ftz.f32 %f2222, %f4928, %f2221; div.approx.ftz.f32 %f5072, %f2220, %f2222; $L__BB104_178: neg.ftz.f32 %f4839, %f5011; selp.f32 %f4838, %f4839, %f5011, %p50; setp.gt.ftz.f32 %p149, %f4838, 0f3F7FF972; setp.lt.ftz.f32 %p150, %f4838, 0f3727C5AC; or.pred %p151, %p149, %p150; selp.f32 %f5080, 0f3F800000, 0f00000000, %p149; @%p151 bra $L__BB104_180; and.b16 %rs28, %rs34, 255; selp.u16 %rs27, 1, 0, %p1; setp.eq.s16 %p152, %rs28, %rs27; neg.ftz.f32 %f2264, %f5049; selp.f32 %f2265, %f5049, %f2264, %p152; add.ftz.f32 %f2266, %f2265, 0f3F800000; mul.ftz.f32 %f2267, %f2266, 0f3F000000; cvt.ftz.sat.f32.f32 %f2268, %f2267; lg2.approx.ftz.f32 %f2269, %f2268; mul.ftz.f32 %f2270, %f5019, %f2269; ex2.approx.ftz.f32 %f5080, %f2270; $L__BB104_180: mul.ftz.f32 %f2271, %f5103, %f5072; fma.rn.ftz.f32 %f5290, %f2271, %f5080, %f5290; mul.ftz.f32 %f2272, %f5104, %f5072; fma.rn.ftz.f32 %f5289, %f2272, %f5080, %f5289; mul.ftz.f32 %f2273, %f5105, %f5072; fma.rn.ftz.f32 %f5288, %f2273, %f5080, %f5288; $L__BB104_181: add.s32 %r150, %r8, 1; setp.gt.u32 %p153, %r150, 9; @%p153 bra $L__BB104_201; neg.ftz.f32 %f4883, %f5102; neg.ftz.f32 %f4882, %f5100; neg.ftz.f32 %f4881, %f5101; mad.lo.s32 %r151, %r151, 1103515245, 12345; and.b16 %rs19, %rs34, 255; setp.eq.s16 %p154, %rs19, 0; selp.f32 %f354, %f145, %f20, %p154; mul.ftz.f32 %f2275, %f212, %f4881; fma.rn.ftz.f32 %f2276, %f211, %f4882, %f2275; fma.rn.ftz.f32 %f355, %f213, %f4883, %f2276; abs.ftz.f32 %f356, %f355; fma.rn.ftz.f32 %f2277, %f354, %f354, 0fBF800000; mov.f32 %f5084, 0f3F800000; mul.ftz.f32 %f357, %f356, %f356; add.ftz.f32 %f358, %f2277, %f357; setp.leu.ftz.f32 %p155, %f358, 0f00000000; @%p155 bra $L__BB104_184; sqrt.approx.ftz.f32 %f2278, %f358; sub.ftz.f32 %f2279, %f2278, %f356; add.ftz.f32 %f2280, %f356, %f2278; div.approx.ftz.f32 %f2281, %f2279, %f2280; fma.rn.ftz.f32 %f2282, %f356, %f2280, 0fBF800000; fma.rn.ftz.f32 %f2283, %f356, %f2279, 0f3F800000; div.approx.ftz.f32 %f2284, %f2282, %f2283; mul.ftz.f32 %f2285, %f2281, 0f3F000000; mul.ftz.f32 %f2286, %f2281, %f2285; fma.rn.ftz.f32 %f2287, %f2284, %f2284, 0f3F800000; mul.ftz.f32 %f5084, %f2286, %f2287; $L__BB104_184: cvt.rn.f32.u32 %f2288, %r151; mul.ftz.f32 %f2289, %f2288, 0f2F800000; setp.gt.ftz.f32 %p156, %f5084, %f2289; @%p156 bra $L__BB104_186; bra.uni $L__BB104_185; $L__BB104_186: mov.f32 %f4946, 0f3F800000; div.approx.ftz.f32 %f4945, %f4946, %f1865; add.ftz.f32 %f2315, %f211, %f211; add.ftz.f32 %f2316, %f212, %f212; add.ftz.f32 %f2317, %f4945, %f4945; fma.rn.ftz.f32 %f5102, %f2317, %f355, %f5102; fma.rn.ftz.f32 %f5101, %f2316, %f355, %f5101; fma.rn.ftz.f32 %f5100, %f2315, %f355, %f5100; mov.pred %p529, 0; bra.uni $L__BB104_187; $L__BB104_185: mov.f32 %f4943, 0f3F800000; div.approx.ftz.f32 %f4942, %f4943, %f1865; and.b16 %rs29, %rs34, 255; setp.eq.s16 %p520, %rs29, 0; selp.f32 %f4941, %f145, %f20, %p520; neg.ftz.f32 %f4886, %f5102; neg.ftz.f32 %f4885, %f5100; neg.ftz.f32 %f4884, %f5101; rcp.approx.ftz.f32 %f2290, %f4941; mov.f32 %f2291, 0f3F800000; mul.ftz.f32 %f2292, %f355, %f355; sub.ftz.f32 %f2293, %f2291, %f2292; mul.ftz.f32 %f2294, %f2293, %f2290; mul.ftz.f32 %f2295, %f2290, %f2294; sub.ftz.f32 %f2296, %f2291, %f2295; mov.f32 %f2297, 0f00000000; max.ftz.f32 %f2298, %f2296, %f2297; sqrt.approx.ftz.f32 %f2299, %f2298; mul.ftz.f32 %f2300, %f355, %f2290; sub.ftz.f32 %f2301, %f2300, %f2299; mul.ftz.f32 %f2302, %f211, %f2301; mul.ftz.f32 %f2303, %f212, %f2301; mul.ftz.f32 %f2304, %f4942, %f2301; mul.ftz.f32 %f2305, %f2290, %f4885; mul.ftz.f32 %f2306, %f2290, %f4884; mul.ftz.f32 %f2307, %f2290, %f4886; sub.ftz.f32 %f2308, %f2302, %f2305; sub.ftz.f32 %f2309, %f2303, %f2306; sub.ftz.f32 %f2310, %f2304, %f2307; mul.ftz.f32 %f2311, %f2309, %f2309; fma.rn.ftz.f32 %f2312, %f2308, %f2308, %f2311; fma.rn.ftz.f32 %f2313, %f2310, %f2310, %f2312; sqrt.approx.ftz.f32 %f2314, %f2313; div.approx.ftz.f32 %f5102, %f2310, %f2314; div.approx.ftz.f32 %f5101, %f2309, %f2314; div.approx.ftz.f32 %f5100, %f2308, %f2314; mov.pred %p529, -1; $L__BB104_187: not.pred %p159, %p529; @%p159 bra $L__BB104_189; and.b16 %rs30, %rs34, 255; setp.eq.s16 %p521, %rs30, 0; selp.u16 %rs34, 1, 0, %p521; neg.ftz.f32 %f5100, %f5100; neg.ftz.f32 %f5101, %f5101; neg.ftz.f32 %f5102, %f5102; neg.ftz.f32 %f5049, %f5049; $L__BB104_189: setp.ne.s32 %p514, %r5, 27; or.pred %p513, %p50, %p514; or.pred %p162, %p513, %p159; @%p162 bra $L__BB104_191; bra.uni $L__BB104_190; $L__BB104_191: setp.eq.s32 %p519, %r8, 0; setp.eq.s32 %p518, %r5, 27; not.pred %p517, %p511; or.pred %p165, %p517, %p519; @%p165 bra $L__BB104_195; mul.ftz.f32 %f4944, %f356, %f356; mov.f32 %f4888, 0f3F800000; sub.ftz.f32 %f5092, %f4888, %f5025; add.ftz.f32 %f384, %f37, %f4944; setp.leu.ftz.f32 %p166, %f384, 0f00000000; @%p166 bra $L__BB104_194; sqrt.approx.ftz.f32 %f2318, %f384; sub.ftz.f32 %f2319, %f2318, %f356; add.ftz.f32 %f2320, %f356, %f2318; div.approx.ftz.f32 %f2321, %f2319, %f2320; fma.rn.ftz.f32 %f2322, %f356, %f2320, 0fBF800000; fma.rn.ftz.f32 %f2323, %f356, %f2319, 0f3F800000; div.approx.ftz.f32 %f2324, %f2322, %f2323; mul.ftz.f32 %f2325, %f2321, 0f3F000000; mul.ftz.f32 %f2326, %f2321, %f2325; fma.rn.ftz.f32 %f2327, %f2324, %f2324, 0f3F800000; mul.ftz.f32 %f2328, %f2326, %f2327; sub.ftz.f32 %f5092, %f2328, %f5025; $L__BB104_194: mul.ftz.f32 %f2329, %f128, %f5092; mov.f32 %f2330, 0f3F800000; sub.ftz.f32 %f2331, %f2330, %f2329; fma.rn.ftz.f32 %f2332, %f5006, %f2331, %f2329; fma.rn.ftz.f32 %f2333, %f5007, %f2331, %f2329; fma.rn.ftz.f32 %f2334, %f5008, %f2331, %f2329; mul.ftz.f32 %f5103, %f5103, %f2332; mul.ftz.f32 %f5104, %f5104, %f2333; mul.ftz.f32 %f5105, %f5105, %f2334; bra.uni $L__BB104_195; $L__BB104_190: mul.ftz.f32 %f5103, %f49, %f5103; mul.ftz.f32 %f5104, %f50, %f5104; mul.ftz.f32 %f5105, %f51, %f5105; $L__BB104_195: setp.gt.ftz.f32 %p167, %f5102, 0f3F7FF972; mov.f32 %f5106, 0f00000000; @%p167 bra $L__BB104_198; setp.lt.ftz.f32 %p168, %f5102, 0fBF7FF972; mov.f32 %f5106, 0fBF7FF972; @%p168 bra $L__BB104_198; mul.ftz.f32 %f2337, %f5102, %f5102; max.ftz.f32 %f2339, %f2337, %f1505; rcp.approx.ftz.f32 %f2340, %f2339; mul.ftz.f32 %f2341, %f1422, %f5101; mul.ftz.f32 %f2342, %f1421, %f5100; mul.ftz.f32 %f2343, %f2341, %f2341; fma.rn.ftz.f32 %f2344, %f2342, %f2342, %f2343; fma.rn.ftz.f32 %f2345, %f2344, %f2340, 0f3F800000; sqrt.approx.ftz.f32 %f2346, %f2345; neg.ftz.f32 %f2347, %f2346; setp.le.ftz.f32 %p169, %f5102, 0f00000000; selp.f32 %f2348, %f2347, %f2346, %p169; add.ftz.f32 %f2349, %f2348, 0fBF800000; mul.ftz.f32 %f5106, %f2349, 0f3F000000; $L__BB104_198: setp.ne.s32 %p516, %r5, 27; or.pred %p515, %p50, %p516; mul.ftz.f32 %f2351, %f49, %f5103; selp.f32 %f5103, %f2351, %f5103, %p515; mul.ftz.f32 %f2352, %f50, %f5104; selp.f32 %f5104, %f2352, %f5104, %p515; mul.ftz.f32 %f2353, %f51, %f5105; selp.f32 %f5105, %f2353, %f5105, %p515; add.ftz.f32 %f2354, %f5049, 0f3F800000; mov.f32 %f5047, 0f3F800000; mul.ftz.f32 %f2355, %f2354, 0f3F000000; cvt.ftz.sat.f32.f32 %f5048, %f2355; @%p167 bra $L__BB104_201; setp.lt.ftz.f32 %p171, %f5102, 0f3727C5AC; mov.f32 %f5047, 0f00000000; @%p171 bra $L__BB104_201; lg2.approx.ftz.f32 %f2357, %f5048; mul.ftz.f32 %f2358, %f5106, %f2357; ex2.approx.ftz.f32 %f5047, %f2358; $L__BB104_201: setp.lt.u32 %p172, %r150, 10; @%p172 bra $L__BB104_111; $L__BB104_202: st.local.u32 [%rd1+-4], %r151; setp.eq.s16 %p173, %rs31, 0; @%p173 bra $L__BB104_519; div.approx.ftz.f32 %f2359, %f5014, %f5011; abs.ftz.f32 %f2360, %f2359; mul.ftz.f32 %f5290, %f5290, %f2360; mul.ftz.f32 %f5289, %f5289, %f2360; mul.ftz.f32 %f5288, %f5288, %f2360; bra.uni $L__BB104_519; $L__BB104_513: and.b32 %r142, %r30, 4; setp.eq.s32 %p487, %r142, 0; add.ftz.f32 %f4709, %f1328, 0f00000000; selp.f32 %f5282, 0f00000000, %f4709, %p487; $L__BB104_514: and.b32 %r143, %r30, 9; setp.eq.s32 %p488, %r143, 0; @%p488 bra $L__BB104_516; mul.ftz.f32 %f4710, %f1405, %f1319; fma.rn.ftz.f32 %f4711, %f1404, %f1318, %f4710; fma.rn.ftz.f32 %f4712, %f1406, %f1320, %f4711; add.ftz.f32 %f4713, %f4712, 0f3F800000; ld.local.f32 %f4714, [%rd3+20]; mul.ftz.f32 %f4715, %f4713, %f4714; add.ftz.f32 %f4716, %f4715, 0fBF800000; mul.ftz.f32 %f4717, %f1326, %f1327; add.ftz.f32 %f4718, %f1326, %f1327; fma.rn.ftz.f32 %f4719, %f4717, %f4716, %f4718; fma.rn.ftz.f32 %f5282, %f4715, %f4719, %f5282; $L__BB104_516: mul.ftz.f32 %f4720, %f1324, 0f3EA2F983; mul.ftz.f32 %f5288, %f4720, %f5282; $L__BB104_517: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_32: ld.local.f32 %f3947, [%rd3+52]; mul.ftz.f32 %f3948, %f1112, %f3947; ld.local.f32 %f3949, [%rd3+48]; mul.ftz.f32 %f3950, %f1113, %f3949; sub.ftz.f32 %f3951, %f3948, %f3950; ld.local.f32 %f3952, [%rd3+44]; mul.ftz.f32 %f3953, %f1113, %f3952; mul.ftz.f32 %f3954, %f1111, %f3947; sub.ftz.f32 %f3955, %f3953, %f3954; mul.ftz.f32 %f3956, %f1111, %f3949; mul.ftz.f32 %f3957, %f1112, %f3952; sub.ftz.f32 %f3958, %f3956, %f3957; mul.ftz.f32 %f3959, %f3955, %f3955; fma.rn.ftz.f32 %f3960, %f3951, %f3951, %f3959; fma.rn.ftz.f32 %f3961, %f3958, %f3958, %f3960; sqrt.approx.ftz.f32 %f3962, %f3961; div.approx.ftz.f32 %f1124, %f3951, %f3962; div.approx.ftz.f32 %f1125, %f3955, %f3962; div.approx.ftz.f32 %f1126, %f3958, %f3962; mul.ftz.f32 %f3963, %f1113, %f1125; mul.ftz.f32 %f3964, %f1112, %f1126; sub.ftz.f32 %f1127, %f3963, %f3964; mul.ftz.f32 %f3965, %f1111, %f1126; mul.ftz.f32 %f3966, %f1113, %f1124; sub.ftz.f32 %f1128, %f3965, %f3966; mul.ftz.f32 %f3967, %f1112, %f1124; mul.ftz.f32 %f3968, %f1111, %f1125; sub.ftz.f32 %f1129, %f3967, %f3968; mul.ftz.f32 %f3969, %f1117, %f1128; fma.rn.ftz.f32 %f3970, %f1116, %f1127, %f3969; fma.rn.ftz.f32 %f3971, %f1118, %f1129, %f3970; mul.ftz.f32 %f3972, %f1117, %f1125; fma.rn.ftz.f32 %f3973, %f1116, %f1124, %f3972; fma.rn.ftz.f32 %f3974, %f1118, %f1126, %f3973; neg.ftz.f32 %f3975, %f3971; mul.ftz.f32 %f3976, %f3918, %f1119; div.approx.ftz.f32 %f3977, %f3975, %f3976; neg.ftz.f32 %f3978, %f3974; mul.ftz.f32 %f3979, %f3919, %f1119; div.approx.ftz.f32 %f3980, %f3978, %f3979; mul.ftz.f32 %f3981, %f3977, %f3977; neg.ftz.f32 %f3982, %f3981; mul.ftz.f32 %f3983, %f3980, %f3980; sub.ftz.f32 %f3984, %f3982, %f3983; mul.ftz.f32 %f3985, %f3984, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3986, %f3985; div.approx.ftz.f32 %f5248, %f3986, %f1121; mul.ftz.f32 %f3987, %f1106, %f1128; fma.rn.ftz.f32 %f3988, %f1105, %f1127, %f3987; fma.rn.ftz.f32 %f3989, %f1107, %f1129, %f3988; mul.ftz.f32 %f3990, %f1106, %f1125; fma.rn.ftz.f32 %f3991, %f1105, %f1124, %f3990; fma.rn.ftz.f32 %f3992, %f1107, %f1126, %f3991; mul.ftz.f32 %f3993, %f3992, %f3992; mul.ftz.f32 %f3994, %f3989, %f3989; mul.ftz.f32 %f1131, %f3918, %f3918; mul.ftz.f32 %f3995, %f1131, %f3994; mul.ftz.f32 %f1132, %f3919, %f3919; fma.rn.ftz.f32 %f3996, %f1132, %f3993, %f3995; add.ftz.f32 %f3997, %f3993, %f3994; div.approx.ftz.f32 %f3998, %f3996, %f3997; mul.ftz.f32 %f3999, %f1123, %f3998; div.approx.ftz.f32 %f4000, %f3999, %f1122; mov.f32 %f4001, 0f00000000; max.ftz.f32 %f4002, %f4000, %f4001; sqrt.approx.ftz.f32 %f1133, %f4002; setp.lt.ftz.f32 %p422, %f1133, 0f3F200000; mov.f32 %f5249, %f5250; @%p422 bra $L__BB104_34; rcp.approx.ftz.f32 %f4003, %f1133; fma.rn.ftz.f32 %f4004, %f4003, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4005, %f4003, %f4004; fma.rn.ftz.f32 %f4006, %f4003, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4007, %f4003, %f4006, 0f3F800000; div.approx.ftz.f32 %f5249, %f4005, %f4007; $L__BB104_34: mul.ftz.f32 %f4009, %f1405, %f1128; fma.rn.ftz.f32 %f4010, %f1404, %f1127, %f4009; fma.rn.ftz.f32 %f4011, %f1406, %f1129, %f4010; mul.ftz.f32 %f4012, %f1405, %f1125; fma.rn.ftz.f32 %f4013, %f1404, %f1124, %f4012; fma.rn.ftz.f32 %f4014, %f1406, %f1126, %f4013; mul.ftz.f32 %f4015, %f4014, %f4014; mul.ftz.f32 %f4016, %f4011, %f4011; mul.ftz.f32 %f4017, %f1131, %f4016; fma.rn.ftz.f32 %f4018, %f1132, %f4015, %f4017; add.ftz.f32 %f4019, %f4015, %f4016; div.approx.ftz.f32 %f4020, %f4018, %f4019; mul.ftz.f32 %f4021, %f1115, %f1115; sub.ftz.f32 %f4022, %f5250, %f4021; mul.ftz.f32 %f4023, %f4022, %f4020; div.approx.ftz.f32 %f4024, %f4023, %f4021; max.ftz.f32 %f4026, %f4024, %f4001; sqrt.approx.ftz.f32 %f1136, %f4026; setp.lt.ftz.f32 %p423, %f1136, 0f3F200000; @%p423 bra $L__BB104_456; rcp.approx.ftz.f32 %f4027, %f1136; fma.rn.ftz.f32 %f4028, %f4027, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4029, %f4027, %f4028; fma.rn.ftz.f32 %f4030, %f4027, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4031, %f4027, %f4030, 0f3F800000; div.approx.ftz.f32 %f5250, %f4029, %f4031; $L__BB104_456: mul.ftz.f32 %f4060, %f5248, 0f3E800000; div.approx.ftz.f32 %f4061, %f4060, %f1114; mul.ftz.f32 %f4062, %f5249, %f5250; mul.ftz.f32 %f5291, %f5249, %f4061; mul.ftz.f32 %f5288, %f4062, %f4061; bra.uni $L__BB104_463; $L__BB104_375: setp.neu.ftz.f32 %p344, %f882, %f883; setp.neu.ftz.f32 %p345, %f882, %f884; or.pred %p346, %p344, %p345; neg.ftz.f32 %f3457, %f884; selp.f32 %f3458, %f3457, %f884, %p346; add.ftz.f32 %f3459, %f882, %f3458; neg.ftz.f32 %f3460, %f883; selp.f32 %f3461, %f883, %f3460, %p346; sub.ftz.f32 %f3462, %f3461, %f882; sub.ftz.f32 %f3463, %f884, %f883; mul.ftz.f32 %f3464, %f3459, %f3459; fma.rn.ftz.f32 %f3465, %f3463, %f3463, %f3464; fma.rn.ftz.f32 %f3466, %f3462, %f3462, %f3465; sqrt.approx.ftz.f32 %f3467, %f3466; div.approx.ftz.f32 %f5172, %f3463, %f3467; div.approx.ftz.f32 %f5173, %f3459, %f3467; div.approx.ftz.f32 %f5174, %f3462, %f3467; mul.ftz.f32 %f3468, %f883, %f5174; mul.ftz.f32 %f3469, %f884, %f5173; sub.ftz.f32 %f5169, %f3468, %f3469; mul.ftz.f32 %f3470, %f884, %f5172; mul.ftz.f32 %f3471, %f882, %f5174; sub.ftz.f32 %f5170, %f3470, %f3471; mul.ftz.f32 %f3472, %f882, %f5173; mul.ftz.f32 %f3473, %f883, %f5172; sub.ftz.f32 %f5171, %f3472, %f3473; $L__BB104_377: mul.ftz.f32 %f3496, %f877, %f5173; fma.rn.ftz.f32 %f3497, %f876, %f5172, %f3496; fma.rn.ftz.f32 %f5181, %f878, %f5174, %f3497; mul.ftz.f32 %f3498, %f877, %f5170; fma.rn.ftz.f32 %f3499, %f876, %f5169, %f3498; fma.rn.ftz.f32 %f5182, %f878, %f5171, %f3499; mul.ftz.f32 %f3500, %f1405, %f5173; fma.rn.ftz.f32 %f3501, %f1404, %f5172, %f3500; fma.rn.ftz.f32 %f907, %f1406, %f5174, %f3501; mul.ftz.f32 %f3502, %f1405, %f5170; fma.rn.ftz.f32 %f3503, %f1404, %f5169, %f3502; fma.rn.ftz.f32 %f908, %f1406, %f5171, %f3503; add.ftz.f32 %f3504, %f5181, %f907; add.ftz.f32 %f3505, %f5182, %f908; mul.ftz.f32 %f3506, %f3504, %f3504; fma.rn.ftz.f32 %f3507, %f3505, %f3505, %f3506; add.ftz.f32 %f3508, %f5183, %f886; fma.rn.ftz.f32 %f3509, %f3508, %f3508, %f3507; sqrt.approx.ftz.f32 %f3510, %f3509; div.approx.ftz.f32 %f909, %f3504, %f3510; div.approx.ftz.f32 %f910, %f3505, %f3510; div.approx.ftz.f32 %f3511, %f3508, %f3510; mul.ftz.f32 %f911, %f3511, %f3511; @%p343 bra $L__BB104_382; bra.uni $L__BB104_378; $L__BB104_382: setp.gt.ftz.f32 %p350, %f5183, 0f3F7FF972; neg.ftz.f32 %f3558, %f909; div.approx.ftz.f32 %f3559, %f3558, %f3443; neg.ftz.f32 %f3560, %f910; div.approx.ftz.f32 %f3561, %f3560, %f3444; fma.rn.ftz.f32 %f3562, %f3559, %f3559, %f911; fma.rn.ftz.f32 %f916, %f3561, %f3561, %f3562; mov.f32 %f5176, 0f00000000; @%p350 bra $L__BB104_385; setp.lt.ftz.f32 %p351, %f5183, 0fBF7FF972; mov.f32 %f5176, 0fBF7FF972; @%p351 bra $L__BB104_385; mul.ftz.f32 %f3564, %f5183, %f5183; mov.f32 %f3565, 0f33D6BF95; max.ftz.f32 %f3566, %f3564, %f3565; rcp.approx.ftz.f32 %f3567, %f3566; mul.ftz.f32 %f3568, %f3443, %f5181; mul.ftz.f32 %f3569, %f5182, %f3444; mul.ftz.f32 %f3570, %f3569, %f3569; fma.rn.ftz.f32 %f3571, %f3568, %f3568, %f3570; fma.rn.ftz.f32 %f3572, %f3571, %f3567, 0f3F800000; sqrt.approx.ftz.f32 %f3573, %f3572; add.ftz.f32 %f3574, %f3573, 0fBF800000; mul.ftz.f32 %f5176, %f3574, 0f3F000000; $L__BB104_385: add.ftz.f32 %f3575, %f5176, 0f3F800000; mov.f32 %f3576, 0f3F800000; mul.ftz.f32 %f3577, %f5183, %f3575; mov.f32 %f3578, 0f33D6BF95; max.ftz.f32 %f3579, %f3577, %f3578; mul.ftz.f32 %f3580, %f916, 0f40490FDB; mul.ftz.f32 %f3581, %f916, %f3580; mul.ftz.f32 %f3582, %f3443, %f3581; mul.ftz.f32 %f3583, %f3444, %f3582; max.ftz.f32 %f3584, %f3583, %f3578; rcp.approx.ftz.f32 %f3585, %f3584; mul.ftz.f32 %f3586, %f3585, 0f3E800000; div.approx.ftz.f32 %f3587, %f3586, %f3579; sqrt.approx.ftz.f32 %f3588, %f881; mul.ftz.f32 %f3589, %f3588, 0fBFFE59F3; mul.ftz.f32 %f3590, %f3588, %f3589; mul.ftz.f32 %f3591, %f3590, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3592, %f3591; fma.rn.ftz.f32 %f3593, %f3592, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f3594, %f3588, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f3595, %f3588, %f3594, 0f417C8B44; fma.rn.ftz.f32 %f3596, %f3588, %f3595, 0fC1665E35; fma.rn.ftz.f32 %f3597, %f3588, %f3596, 0f40CE7881; fma.rn.ftz.f32 %f3598, %f3588, %f3597, 0fBF993E81; fma.rn.ftz.f32 %f3599, %f3588, %f3598, 0f3CE3C31E; fma.rn.ftz.f32 %f3600, %f3588, %f3599, 0f3BBA5D49; sub.ftz.f32 %f3601, %f3593, %f3600; cvt.ftz.sat.f32.f32 %f3602, %f3601; sub.ftz.f32 %f3603, %f3576, %f3602; mul.ftz.f32 %f3604, %f886, 0f3EA2F983; mul.ftz.f32 %f3605, %f3604, %f3603; fma.rn.ftz.f32 %f5177, %f3587, %f3602, %f3605; bra.uni $L__BB104_386; $L__BB104_44: ld.local.f32 %f3365, [%rd3+52]; mul.ftz.f32 %f3366, %f840, %f3365; ld.local.f32 %f3367, [%rd3+48]; mul.ftz.f32 %f3368, %f842, %f3367; sub.ftz.f32 %f3369, %f3366, %f3368; ld.local.f32 %f3370, [%rd3+44]; mul.ftz.f32 %f3371, %f842, %f3370; mul.ftz.f32 %f3372, %f838, %f3365; sub.ftz.f32 %f3373, %f3371, %f3372; mul.ftz.f32 %f3374, %f838, %f3367; mul.ftz.f32 %f3375, %f840, %f3370; sub.ftz.f32 %f3376, %f3374, %f3375; mul.ftz.f32 %f3377, %f3373, %f3373; fma.rn.ftz.f32 %f3378, %f3369, %f3369, %f3377; fma.rn.ftz.f32 %f3379, %f3376, %f3376, %f3378; sqrt.approx.ftz.f32 %f3380, %f3379; div.approx.ftz.f32 %f3381, %f3369, %f3380; div.approx.ftz.f32 %f3382, %f3373, %f3380; div.approx.ftz.f32 %f3383, %f3376, %f3380; mul.ftz.f32 %f3384, %f842, %f3382; mul.ftz.f32 %f3385, %f840, %f3383; sub.ftz.f32 %f3386, %f3384, %f3385; mul.ftz.f32 %f3387, %f838, %f3383; mul.ftz.f32 %f3388, %f842, %f3381; sub.ftz.f32 %f3389, %f3387, %f3388; mul.ftz.f32 %f3390, %f840, %f3381; mul.ftz.f32 %f3391, %f838, %f3382; sub.ftz.f32 %f3392, %f3390, %f3391; mul.ftz.f32 %f3393, %f849, %f3389; fma.rn.ftz.f32 %f3394, %f848, %f3386, %f3393; fma.rn.ftz.f32 %f856, %f850, %f3392, %f3394; mul.ftz.f32 %f3395, %f849, %f3382; fma.rn.ftz.f32 %f3396, %f848, %f3381, %f3395; fma.rn.ftz.f32 %f857, %f850, %f3383, %f3396; setp.geu.ftz.f32 %p338, %f852, 0f3F800000; mov.f32 %f5164, 0f3F800000; @%p338 bra $L__BB104_46; mul.ftz.f32 %f3397, %f854, %f856; mul.ftz.f32 %f3398, %f856, %f3397; mul.ftz.f32 %f3399, %f855, %f857; fma.rn.ftz.f32 %f3400, %f857, %f3399, %f3398; mul.ftz.f32 %f3401, %f852, %f852; mov.f32 %f3402, 0f3F800000; sub.ftz.f32 %f3403, %f3402, %f3401; div.approx.ftz.f32 %f3404, %f3400, %f3403; lg2.approx.ftz.f32 %f3405, %f852; mul.ftz.f32 %f3406, %f3404, %f3405; ex2.approx.ftz.f32 %f5164, %f3406; $L__BB104_46: add.ftz.f32 %f3407, %f855, 0f3F800000; add.ftz.f32 %f3408, %f854, 0f3F800000; mul.ftz.f32 %f3409, %f3408, %f3407; sqrt.approx.ftz.f32 %f3410, %f3409; mov.f32 %f3411, 0f41C90FDB; div.approx.ftz.f32 %f3412, %f3410, %f3411; mul.ftz.f32 %f3413, %f853, %f3412; mul.ftz.f32 %f5166, %f5164, %f3413; mul.ftz.f32 %f3414, %f5164, %f3412; div.approx.ftz.f32 %f5291, %f3414, %f851; $L__BB104_368: max.ftz.f32 %f3424, %f844, %f3355; max.ftz.f32 %f3425, %f853, %f3424; mul.ftz.f32 %f3426, %f851, %f3425; max.ftz.f32 %f3427, %f3355, %f3426; rcp.approx.ftz.f32 %f3428, %f3427; mul.ftz.f32 %f5288, %f3428, %f5166; $L__BB104_369: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; bra.uni $L__BB104_519; $L__BB104_378: setp.gt.ftz.f32 %p348, %f5183, 0f3F7FF972; mul.ftz.f32 %f3513, %f3443, %f3443; mov.f32 %f3514, 0f3F800000; sub.ftz.f32 %f3515, %f3514, %f911; fma.rn.ftz.f32 %f3516, %f3513, %f911, %f3515; mul.ftz.f32 %f3517, %f3516, 0f40490FDB; mul.ftz.f32 %f3518, %f3516, %f3517; mov.f32 %f3519, 0f33D6BF95; max.ftz.f32 %f3520, %f3518, %f3519; div.approx.ftz.f32 %f912, %f3513, %f3520; mov.f32 %f5175, 0f00000000; @%p348 bra $L__BB104_381; setp.lt.ftz.f32 %p349, %f5183, 0fBF7FF972; mov.f32 %f5175, 0fBF7FF972; @%p349 bra $L__BB104_381; mul.ftz.f32 %f3522, %f5183, %f5183; max.ftz.f32 %f3524, %f3522, %f3519; rcp.approx.ftz.f32 %f3525, %f3524; mul.ftz.f32 %f3526, %f3443, %f5181; mul.ftz.f32 %f3527, %f3526, %f3526; mul.ftz.f32 %f3528, %f5182, %f3443; fma.rn.ftz.f32 %f3529, %f3528, %f3528, %f3527; fma.rn.ftz.f32 %f3530, %f3529, %f3525, 0f3F800000; sqrt.approx.ftz.f32 %f3531, %f3530; add.ftz.f32 %f3532, %f3531, 0fBF800000; mul.ftz.f32 %f5175, %f3532, 0f3F000000; $L__BB104_381: add.ftz.f32 %f3533, %f5175, 0f3F800000; mul.ftz.f32 %f3535, %f5183, %f3533; max.ftz.f32 %f3537, %f3535, %f3519; mul.ftz.f32 %f3538, %f912, 0f3E800000; div.approx.ftz.f32 %f3539, %f3538, %f3537; mul.ftz.f32 %f3540, %f3443, 0fBFFE59F3; mul.ftz.f32 %f3541, %f3443, %f3540; mul.ftz.f32 %f3542, %f3541, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3543, %f3542; fma.rn.ftz.f32 %f3544, %f3543, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f3545, %f3443, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f3546, %f3443, %f3545, 0f417C8B44; fma.rn.ftz.f32 %f3547, %f3443, %f3546, 0fC1665E35; fma.rn.ftz.f32 %f3548, %f3443, %f3547, 0f40CE7881; fma.rn.ftz.f32 %f3549, %f3443, %f3548, 0fBF993E81; fma.rn.ftz.f32 %f3550, %f3443, %f3549, 0f3CE3C31E; fma.rn.ftz.f32 %f3551, %f3443, %f3550, 0f3BBA5D49; sub.ftz.f32 %f3552, %f3544, %f3551; cvt.ftz.sat.f32.f32 %f3553, %f3552; sub.ftz.f32 %f3554, %f3514, %f3553; mul.ftz.f32 %f3555, %f886, 0f3EA2F983; mul.ftz.f32 %f3556, %f3555, %f3554; fma.rn.ftz.f32 %f5177, %f3539, %f3553, %f3556; $L__BB104_386: setp.le.ftz.f32 %p352, %f5177, 0f00000000; mov.f32 %f5291, %f5290; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p352 bra $L__BB104_519; ld.local.u64 %rd85, [%rd3+36]; ld.f32 %f921, [%rd85]; ld.f32 %f922, [%rd85+4]; ld.f32 %f923, [%rd85+8]; ld.local.f32 %f924, [%rd3+28]; ld.f32 %f925, [%rd85+12]; ld.f32 %f926, [%rd85+16]; ld.f32 %f927, [%rd85+20]; setp.geu.ftz.f32 %p353, %f886, %f5183; mov.f32 %f5178, %f907; mov.f32 %f5179, %f908; mov.f32 %f5180, %f886; mov.u16 %rs35, %rs23; @%p353 bra $L__BB104_389; mov.f32 %f5178, %f5181; mov.f32 %f5179, %f5182; mov.f32 %f5180, %f5183; mov.f32 %f5181, %f907; mov.f32 %f5182, %f908; mov.f32 %f5183, %f886; mov.u16 %rs35, %rs22; $L__BB104_389: setp.lt.ftz.f32 %p354, %f5180, 0f3727C5AC; setp.lt.ftz.f32 %p355, %f5183, 0f3727C5AC; or.pred %p356, %p354, %p355; mov.f32 %f5291, %f5177; mov.f32 %f5289, %f5290; mov.f32 %f5288, %f5290; @%p356 bra $L__BB104_519; neg.ftz.f32 %f5229, %f5181; neg.ftz.f32 %f5230, %f5182; setp.lt.ftz.f32 %p357, %f5183, 0fBF7FF972; mov.f32 %f5185, 0f00000000; mov.f32 %f5235, %f5185; @%p357 bra $L__BB104_393; setp.gt.ftz.f32 %p358, %f5183, 0f3F7FF972; mov.f32 %f5235, 0fBF7FF972; @%p358 bra $L__BB104_393; mul.ftz.f32 %f3615, %f5183, %f5183; mov.f32 %f3616, 0f33D6BF95; max.ftz.f32 %f3617, %f3615, %f3616; rcp.approx.ftz.f32 %f3618, %f3617; mul.ftz.f32 %f3619, %f3443, %f5229; mul.ftz.f32 %f3620, %f3444, %f5230; mul.ftz.f32 %f3621, %f3620, %f3620; fma.rn.ftz.f32 %f3622, %f3619, %f3619, %f3621; fma.rn.ftz.f32 %f3623, %f3622, %f3618, 0f3F800000; sqrt.approx.ftz.f32 %f3624, %f3623; neg.ftz.f32 %f3625, %f3624; setp.ge.ftz.f32 %p359, %f5183, 0f80000000; selp.f32 %f3626, %f3625, %f3624, %p359; add.ftz.f32 %f3627, %f3626, 0fBF800000; mul.ftz.f32 %f5235, %f3627, 0f3F000000; $L__BB104_393: setp.gt.ftz.f32 %p360, %f5180, 0f3F7FF972; @%p360 bra $L__BB104_396; setp.lt.ftz.f32 %p361, %f5180, 0fBF7FF972; mov.f32 %f5185, 0fBF7FF972; @%p361 bra $L__BB104_396; mul.ftz.f32 %f3630, %f5180, %f5180; mov.f32 %f3631, 0f33D6BF95; max.ftz.f32 %f3632, %f3630, %f3631; rcp.approx.ftz.f32 %f3633, %f3632; mul.ftz.f32 %f3634, %f3443, %f5178; mul.ftz.f32 %f3635, %f3444, %f5179; mul.ftz.f32 %f3636, %f3635, %f3635; fma.rn.ftz.f32 %f3637, %f3634, %f3634, %f3636; fma.rn.ftz.f32 %f3638, %f3637, %f3633, 0f3F800000; sqrt.approx.ftz.f32 %f3639, %f3638; neg.ftz.f32 %f3640, %f3639; setp.le.ftz.f32 %p362, %f5180, 0f00000000; selp.f32 %f3641, %f3640, %f3639, %p362; add.ftz.f32 %f3642, %f3641, 0fBF800000; mul.ftz.f32 %f5185, %f3642, 0f3F000000; $L__BB104_396: add.ftz.f32 %f3643, %f5178, %f5181; add.ftz.f32 %f3644, %f5179, %f5182; mul.ftz.f32 %f3645, %f3644, %f3644; fma.rn.ftz.f32 %f3646, %f3643, %f3643, %f3645; add.ftz.f32 %f3647, %f5180, %f5183; fma.rn.ftz.f32 %f3648, %f3647, %f3647, %f3646; sqrt.approx.ftz.f32 %f3649, %f3648; div.approx.ftz.f32 %f940, %f3643, %f3649; div.approx.ftz.f32 %f941, %f3644, %f3649; div.approx.ftz.f32 %f942, %f3647, %f3649; mul.ftz.f32 %f943, %f942, %f942; setp.eq.ftz.f32 %p363, %f3443, %f3444; @%p363 bra $L__BB104_398; bra.uni $L__BB104_397; $L__BB104_398: mul.ftz.f32 %f3662, %f3443, %f3443; mov.f32 %f3663, 0f3F800000; sub.ftz.f32 %f3664, %f3663, %f943; fma.rn.ftz.f32 %f3665, %f3662, %f943, %f3664; mul.ftz.f32 %f3666, %f3665, 0f40490FDB; mul.ftz.f32 %f3667, %f3665, %f3666; mov.f32 %f3668, 0f33D6BF95; max.ftz.f32 %f3669, %f3667, %f3668; div.approx.ftz.f32 %f5186, %f3662, %f3669; bra.uni $L__BB104_399; $L__BB104_55: mov.f32 %f3134, 0f3F800000; div.approx.ftz.f32 %f5156, %f772, %f3134; $L__BB104_350: mul.ftz.f32 %f776, %f767, %f768; neg.ftz.f32 %f3137, %f769; and.b32 %r105, %r19, 1; setp.eq.b32 %p307, %r105, 1; selp.f32 %f777, %f3137, %f769, %p307; add.ftz.f32 %f3138, %f5156, 0fBF800000; add.ftz.f32 %f3139, %f3138, 0f3F800000; mul.ftz.f32 %f3140, %f3139, %f3139; fma.rn.ftz.f32 %f3141, %f3140, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f3142, %f3139, %f3141; fma.rn.ftz.f32 %f3143, %f3140, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f3144, %f3140, %f3143, 0f3F800000; div.approx.ftz.f32 %f3145, %f3142, %f3144; sub.ftz.f32 %f3147, %f3065, %f3145; setp.gt.ftz.f32 %p308, %f772, 0f3F800000; selp.f32 %f3148, %f3147, %f3145, %p308; mov.b32 %r106, %f771; and.b32 %r107, %r106, -2147483648; mov.b32 %r108, %f3148; and.b32 %r109, %r108, 2147483647; or.b32 %r20, %r109, %r107; add.ftz.f32 %f3149, %f764, 0fBFC90FDB; mul.ftz.f32 %f3150, %f3149, 0f3F000000; sub.ftz.f32 %f3151, %f3150, %f762; mul.ftz.f32 %f778, %f770, %f3151; abs.ftz.f32 %f779, %f778; setp.eq.ftz.f32 %p309, %f779, 0f00000000; @%p309 bra $L__BB104_355; setp.eq.ftz.f32 %p310, %f779, 0f3F800000; mov.f32 %f5157, 0f3F800000; @%p310 bra $L__BB104_355; setp.gt.ftz.f32 %p311, %f779, 0f3F800000; @%p311 bra $L__BB104_354; bra.uni $L__BB104_353; $L__BB104_354: mov.f32 %f3154, 0f3F800000; div.approx.ftz.f32 %f5157, %f3154, %f779; bra.uni $L__BB104_355; $L__BB104_490: mov.f32 %f4512, 0f3F800000; div.approx.ftz.f32 %f5272, %f1257, %f4512; $L__BB104_492: add.ftz.f32 %f4515, %f5272, 0fBF800000; add.ftz.f32 %f4516, %f4515, 0f3F800000; mul.ftz.f32 %f4517, %f4516, %f4516; fma.rn.ftz.f32 %f4518, %f4517, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4519, %f4516, %f4518; fma.rn.ftz.f32 %f4520, %f4517, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4521, %f4517, %f4520, 0f3F800000; div.approx.ftz.f32 %f4522, %f4519, %f4521; sub.ftz.f32 %f4524, %f4438, %f4522; setp.gt.ftz.f32 %p462, %f1257, 0f3F800000; selp.f32 %f4525, %f4524, %f4522, %p462; mov.b32 %r128, %f1256; and.b32 %r129, %r128, -2147483648; mov.b32 %r130, %f4525; and.b32 %r131, %r130, 2147483647; or.b32 %r28, %r131, %r129; add.ftz.f32 %f4526, %f1250, 0fBFC90FDB; div.approx.ftz.f32 %f4528, %f4526, %f4508; sub.ftz.f32 %f4529, %f4528, %f1247; mul.ftz.f32 %f1261, %f1255, %f4529; abs.ftz.f32 %f1262, %f1261; setp.eq.ftz.f32 %p463, %f1262, 0f00000000; mov.f32 %f5273, %f5274; @%p463 bra $L__BB104_497; setp.eq.ftz.f32 %p464, %f1262, 0f3F800000; mov.f32 %f5273, 0f3F800000; @%p464 bra $L__BB104_497; setp.gt.ftz.f32 %p465, %f1262, 0f3F800000; @%p465 bra $L__BB104_496; bra.uni $L__BB104_495; $L__BB104_496: mov.f32 %f4532, 0f3F800000; div.approx.ftz.f32 %f5273, %f4532, %f1262; bra.uni $L__BB104_497; $L__BB104_353: mov.f32 %f3153, 0f3F800000; div.approx.ftz.f32 %f5157, %f779, %f3153; $L__BB104_355: add.ftz.f32 %f3155, %f5157, 0fBF800000; add.ftz.f32 %f3156, %f3155, 0f3F800000; mul.ftz.f32 %f3157, %f3156, %f3156; fma.rn.ftz.f32 %f3158, %f3157, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f3159, %f3156, %f3158; fma.rn.ftz.f32 %f3160, %f3157, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f3161, %f3157, %f3160, 0f3F800000; div.approx.ftz.f32 %f3162, %f3159, %f3161; sub.ftz.f32 %f3164, %f3065, %f3162; setp.gt.ftz.f32 %p312, %f779, 0f3F800000; selp.f32 %f3165, %f3164, %f3162, %p312; mov.b32 %r110, %f778; and.b32 %r111, %r110, -2147483648; mov.b32 %r112, %f3165; and.b32 %r113, %r112, 2147483647; or.b32 %r114, %r113, %r111; mov.b32 %f3166, %r114; add.ftz.f32 %f3167, %f764, %f765; mul.ftz.f32 %f3168, %f3167, 0f3F000000; sub.ftz.f32 %f3169, %f3168, %f762; mul.ftz.f32 %f3170, %f776, 0f3F000000; abs.ftz.f32 %f3171, %f776; setp.lt.ftz.f32 %p313, %f3171, 0f40490FDB; selp.f32 %f3172, %f3170, 0f3FC90FDB, %p313; mul.ftz.f32 %f3173, %f3172, 0f3EA2F983; mov.b32 %r115, %f3173; and.b32 %r116, %r115, -2147483648; or.b32 %r117, %r116, 1056964608; mov.b32 %f3174, %r117; add.ftz.f32 %f3175, %f3173, %f3174; cvt.rzi.ftz.s32.f32 %r118, %f3175; cvt.rn.f32.s32 %f3176, %r118; fma.rn.ftz.f32 %f3177, %f3176, 0fC0490000, %f3172; fma.rn.ftz.f32 %f3178, %f3176, 0fBA7DA000, %f3177; fma.rn.ftz.f32 %f3179, %f3176, 0fB4222000, %f3178; fma.rn.ftz.f32 %f3180, %f3176, 0fACB4611A, %f3179; add.ftz.f32 %f3181, %f3180, 0fBFC90FDB; add.ftz.f32 %f3182, %f3181, 0f3FC90FDB; mul.ftz.f32 %f3183, %f3182, %f3182; fma.rn.ftz.f32 %f3184, %f3183, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f3185, %f3183, %f3184, 0fBAB60B58; fma.rn.ftz.f32 %f3186, %f3183, %f3185, 0f3D2AAAAA; fma.rn.ftz.f32 %f3187, %f3183, %f3186, 0fBF000000; fma.rn.ftz.f32 %f3188, %f3183, %f3187, 0f3F800000; and.b32 %r119, %r118, 1; setp.eq.b32 %p314, %r119, 1; neg.ftz.f32 %f3189, %f3188; selp.f32 %f3190, %f3189, %f3188, %p314; abs.ftz.f32 %f3191, %f3190; setp.gt.ftz.f32 %p315, %f3191, 0f3F800000; mul.ftz.f32 %f3192, %f3190, 0f3E800000; selp.f32 %f3193, 0f00000000, %f3192, %p315; mul.ftz.f32 %f3194, %f767, %f3193; mul.ftz.f32 %f3195, %f3169, %f3169; fma.rn.ftz.f32 %f3196, %f763, %f763, %f3195; add.ftz.f32 %f3197, %f3196, %f3196; mov.b32 %f3198, %r20; sub.ftz.f32 %f3199, %f3198, %f3166; mul.ftz.f32 %f3200, %f3197, %f3199; abs.ftz.f32 %f3201, %f777; setp.gt.ftz.f32 %p316, %f3201, 0f3F800000; selp.f32 %f3202, 0f00000000, %f777, %p316; mul.ftz.f32 %f3203, %f3202, %f3200; div.approx.ftz.f32 %f3204, %f763, %f3203; mul.ftz.f32 %f5288, %f3204, %f3194; $L__BB104_356: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; bra.uni $L__BB104_519; $L__BB104_495: mov.f32 %f4531, 0f3F800000; div.approx.ftz.f32 %f5273, %f1262, %f4531; $L__BB104_497: add.ftz.f32 %f4534, %f5273, 0fBF800000; add.ftz.f32 %f4535, %f4534, 0f3F800000; mul.ftz.f32 %f4536, %f4535, %f4535; fma.rn.ftz.f32 %f4537, %f4536, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4538, %f4535, %f4537; fma.rn.ftz.f32 %f4539, %f4536, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4540, %f4536, %f4539, 0f3F800000; div.approx.ftz.f32 %f4541, %f4538, %f4540; sub.ftz.f32 %f4543, %f4438, %f4541; setp.gt.ftz.f32 %p466, %f1262, 0f3F800000; selp.f32 %f4544, %f4543, %f4541, %p466; mov.b32 %r132, %f1261; and.b32 %r133, %r132, -2147483648; mov.b32 %r134, %f4544; and.b32 %r135, %r134, 2147483647; or.b32 %r29, %r135, %r133; div.approx.ftz.f32 %f1266, %f4438, %f1249; abs.ftz.f32 %f1267, %f1266; setp.eq.ftz.f32 %p467, %f1267, 0f00000000; @%p467 bra $L__BB104_502; setp.eq.ftz.f32 %p468, %f1267, 0f3F800000; mov.f32 %f5274, 0f3F800000; @%p468 bra $L__BB104_502; setp.gt.ftz.f32 %p469, %f1267, 0f3F800000; @%p469 bra $L__BB104_501; bra.uni $L__BB104_500; $L__BB104_501: mov.f32 %f4547, 0f3F800000; div.approx.ftz.f32 %f5274, %f4547, %f1267; bra.uni $L__BB104_502; $L__BB104_500: mov.f32 %f4546, 0f3F800000; div.approx.ftz.f32 %f5274, %f1267, %f4546; $L__BB104_502: add.ftz.f32 %f4548, %f5274, 0fBF800000; add.ftz.f32 %f4549, %f4548, 0f3F800000; mul.ftz.f32 %f4550, %f4549, %f4549; fma.rn.ftz.f32 %f4551, %f4550, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4552, %f4549, %f4551; fma.rn.ftz.f32 %f4553, %f4550, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4554, %f4550, %f4553, 0f3F800000; div.approx.ftz.f32 %f4555, %f4552, %f4554; sub.ftz.f32 %f4557, %f4438, %f4555; setp.gt.ftz.f32 %p470, %f1267, 0f3F800000; selp.f32 %f4558, %f4557, %f4555, %p470; mov.b32 %r136, %f1266; and.b32 %r137, %r136, -2147483648; mov.b32 %r138, %f4558; and.b32 %r139, %r138, 2147483647; or.b32 %r140, %r139, %r137; mov.b32 %f4559, %r140; add.ftz.f32 %f4560, %f4559, %f4559; add.ftz.f32 %f4562, %f1250, %f1251; div.approx.ftz.f32 %f4563, %f4562, %f4508; sub.ftz.f32 %f4564, %f4563, %f1247; abs.ftz.f32 %f4565, %f1253; sub.ftz.f32 %f4567, %f4440, %f4565; mul.ftz.f32 %f4568, %f4564, %f4564; fma.rn.ftz.f32 %f4569, %f1248, %f1248, %f4568; add.ftz.f32 %f4570, %f4569, %f4569; mov.b32 %f4571, %r28; mov.b32 %f4572, %r29; sub.ftz.f32 %f4573, %f4571, %f4572; mul.ftz.f32 %f4574, %f4573, %f4570; abs.ftz.f32 %f4575, %f1254; setp.gt.ftz.f32 %p471, %f4575, 0f3F800000; selp.f32 %f4576, 0f00000000, %f1254, %p471; mul.ftz.f32 %f4577, %f4576, %f4574; div.approx.ftz.f32 %f4578, %f1248, %f4577; mul.ftz.f32 %f4579, %f4567, %f4567; fma.rn.ftz.f32 %f4580, %f1249, %f1249, %f4579; mul.ftz.f32 %f4581, %f4580, %f4560; div.approx.ftz.f32 %f4582, %f1249, %f4581; mul.ftz.f32 %f5288, %f4578, %f4582; $L__BB104_503: mov.f32 %f5289, %f5288; mov.f32 %f5290, %f5288; mov.f32 %f5291, %f5288; $L__BB104_519: add.s32 %r31, %r5, -2; setp.gt.u32 %p491, %r31, 6; @%p491 bra $L__BB104_530; ld.local.f32 %f1368, [%rd3+16]; ld.local.f32 %f1369, [%rd1+-240]; ld.local.f32 %f1370, [%rd1+-248]; ld.local.f32 %f1371, [%rd3+8]; setp.neu.ftz.f32 %p492, %f1371, %f1370; mov.u16 %rs36, %rs23; @%p492 bra $L__BB104_523; ld.local.f32 %f4752, [%rd3+12]; ld.local.f32 %f4753, [%rd1+-244]; setp.neu.ftz.f32 %p493, %f4752, %f4753; mov.u16 %rs36, %rs23; @%p493 bra $L__BB104_523; setp.eq.ftz.f32 %p494, %f1368, %f1369; selp.u16 %rs36, 1, 0, %p494; $L__BB104_523: setp.ne.s16 %p495, %rs36, 0; @%p495 bra $L__BB104_530; ld.local.f32 %f4755, [%rd3+12]; mul.ftz.f32 %f4756, %f1405, %f4755; fma.rn.ftz.f32 %f4757, %f1404, %f1371, %f4756; fma.rn.ftz.f32 %f4758, %f1406, %f1368, %f4757; setp.lt.ftz.f32 %p496, %f4758, 0f00000000; mov.f32 %f5292, 0f00000000; ld.local.f32 %f4759, [%rd1+-244]; neg.ftz.f32 %f4760, %f4759; neg.ftz.f32 %f4761, %f1370; selp.f32 %f4762, %f4761, %f1370, %p496; selp.f32 %f4763, %f4760, %f4759, %p496; neg.ftz.f32 %f4764, %f1369; selp.f32 %f4765, %f4764, %f1369, %p496; mul.ftz.f32 %f4766, %f1405, %f4763; fma.rn.ftz.f32 %f4767, %f1404, %f4762, %f4766; fma.rn.ftz.f32 %f1372, %f1406, %f4765, %f4767; mul.ftz.f32 %f4768, %f4755, %f4763; fma.rn.ftz.f32 %f4769, %f1371, %f4762, %f4768; fma.rn.ftz.f32 %f4770, %f1368, %f4765, %f4769; mul.ftz.f32 %f1373, %f4758, %f4770; setp.eq.ftz.f32 %p497, %f1373, 0f00000000; @%p497 bra $L__BB104_526; div.approx.ftz.f32 %f5292, %f1372, %f1373; $L__BB104_526: setp.ge.ftz.f32 %p498, %f5292, 0f3F800000; mov.f32 %f5293, 0f3F800000; @%p498 bra $L__BB104_529; setp.lt.ftz.f32 %p499, %f5292, 0f00000000; mov.f32 %f5293, 0f00000000; @%p499 bra $L__BB104_529; mul.ftz.f32 %f4773, %f5292, %f5292; mul.ftz.f32 %f4774, %f5292, %f4773; sub.ftz.f32 %f4775, %f4773, %f4774; add.ftz.f32 %f5293, %f5292, %f4775; $L__BB104_529: mul.ftz.f32 %f5290, %f5290, %f5293; mul.ftz.f32 %f5289, %f5289, %f5293; mul.ftz.f32 %f5288, %f5288, %f5293; $L__BB104_530: ld.local.u32 %r144, [%rd1+-184]; mul.wide.s32 %rd88, %r144, 208; add.s64 %rd89, %rd2, %rd88; ld.global.f32 %f1384, [%rd89+180]; setp.leu.ftz.f32 %p500, %f1384, 0f3F800000; @%p500 bra $L__BB104_533; ld.local.f32 %f4776, [%rd3+8]; ld.local.f32 %f4777, [%rd3+12]; mul.ftz.f32 %f4778, %f1405, %f4777; fma.rn.ftz.f32 %f4779, %f1404, %f4776, %f4778; ld.local.f32 %f4780, [%rd3+16]; fma.rn.ftz.f32 %f1385, %f1406, %f4780, %f4779; setp.ltu.ftz.f32 %p501, %f1385, 0f00000000; @%p501 bra $L__BB104_533; mov.f32 %f4781, 0f3F800000; min.ftz.f32 %f4782, %f1385, %f4781; abs.ftz.f32 %f4783, %f4782; setp.lt.ftz.f32 %p502, %f4783, 0f3F800000; add.ftz.f32 %f4784, %f4783, 0fBF800000; add.ftz.f32 %f4785, %f4784, 0f3F800000; selp.f32 %f4786, %f4785, 0f3F800000, %p502; sub.ftz.f32 %f4787, %f4781, %f4786; sqrt.approx.ftz.f32 %f4788, %f4787; fma.rn.ftz.f32 %f4789, %f4786, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f4790, %f4786, %f4789, 0fBE5A6B93; fma.rn.ftz.f32 %f4791, %f4786, %f4790, 0f3FC90FDB; mul.ftz.f32 %f4792, %f4788, %f4791; setp.lt.ftz.f32 %p503, %f4782, 0f00000000; mov.f32 %f4793, 0f00000000; mov.f32 %f4794, 0f40490FDB; sub.ftz.f32 %f4795, %f4794, %f4792; selp.f32 %f4796, %f4795, %f4792, %p503; mul.ftz.f32 %f4797, %f1384, %f4796; cos.approx.ftz.f32 %f4798, %f4797; max.ftz.f32 %f4799, %f4798, %f4793; div.approx.ftz.f32 %f4800, %f4799, %f4782; mul.ftz.f32 %f5290, %f5290, %f4800; mul.ftz.f32 %f5289, %f5289, %f4800; mul.ftz.f32 %f5288, %f5288, %f4800; $L__BB104_533: setp.eq.ftz.f32 %p504, %f5291, 0f00000000; @%p504 bra $L__BB104_539; setp.lt.u32 %p505, %r31, 7; ld.local.f32 %f4801, [%rd3+-12]; mul.ftz.f32 %f1392, %f5290, %f4801; ld.local.f32 %f4802, [%rd3+-8]; mul.ftz.f32 %f1393, %f5289, %f4802; ld.local.f32 %f4803, [%rd3+-4]; mul.ftz.f32 %f1394, %f5288, %f4803; @%p505 bra $L__BB104_537; bra.uni $L__BB104_535; $L__BB104_537: ld.local.f32 %f4810, [%rd7]; add.ftz.f32 %f4811, %f1392, %f4810; ld.local.f32 %f4812, [%rd7+4]; add.ftz.f32 %f4813, %f1393, %f4812; ld.local.f32 %f4814, [%rd7+8]; add.ftz.f32 %f4815, %f1394, %f4814; st.local.f32 [%rd7], %f4811; st.local.f32 [%rd7+4], %f4813; st.local.f32 [%rd7+8], %f4815; bra.uni $L__BB104_538; $L__BB104_535: add.s32 %r145, %r5, -9; setp.gt.u32 %p506, %r145, 11; setp.ne.s32 %p507, %r5, 29; and.pred %p508, %p507, %p506; @%p508 bra $L__BB104_538; ld.local.f32 %f4804, [%rd7+12]; add.ftz.f32 %f4805, %f1392, %f4804; ld.local.f32 %f4806, [%rd7+16]; add.ftz.f32 %f4807, %f1393, %f4806; ld.local.f32 %f4808, [%rd7+20]; add.ftz.f32 %f4809, %f1394, %f4808; st.local.f32 [%rd7+12], %f4805; st.local.f32 [%rd7+16], %f4807; st.local.f32 [%rd7+20], %f4809; $L__BB104_538: ld.local.f32 %f4816, [%rd7+24]; add.ftz.f32 %f4817, %f1392, %f4816; ld.local.f32 %f4818, [%rd7+28]; add.ftz.f32 %f4819, %f1393, %f4818; ld.local.f32 %f4820, [%rd7+32]; add.ftz.f32 %f4821, %f1394, %f4820; st.local.f32 [%rd7+24], %f4817; st.local.f32 [%rd7+28], %f4819; st.local.f32 [%rd7+32], %f4821; ld.local.f32 %f4822, [%rd3+4]; fma.rn.ftz.f32 %f5301, %f5291, %f4822, %f5301; ld.local.u32 %r157, [%rd1]; $L__BB104_539: ld.local.f32 %f4823, [%rd3+4]; add.ftz.f32 %f5302, %f5302, %f4823; $L__BB104_540: add.s32 %r148, %r148, 1; setp.lt.s32 %p509, %r148, %r157; @%p509 bra $L__BB104_2; $L__BB104_541: setp.leu.ftz.f32 %p510, %f5302, 0f00000000; mov.f32 %f5305, 0f00000000; @%p510 bra $L__BB104_543; div.approx.ftz.f32 %f5305, %f5301, %f5302; $L__BB104_543: st.param.f32 [func_retval0+0], %f5305; ret; } .func (.param .b32 func_retval0) _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj( .param .b64 _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_0, .param .align 4 .b8 _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_1[12], .param .b64 _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_2, .param .b32 _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_3 ) { .reg .pred %p<542>; .reg .b16 %rs<43>; .reg .f32 %f<5312>; .reg .b32 %r<170>; .reg .b64 %rd<87>; ld.param.u64 %rd7, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_0]; ld.param.u64 %rd8, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_2]; ld.param.u32 %r37, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_3]; ld.param.f32 %f1406, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_1+8]; ld.param.f32 %f1405, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_1+4]; ld.param.f32 %f1404, [_Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj_param_1]; cvta.to.local.u64 %rd9, %rd7; cvta.to.local.u64 %rd10, %rd8; mov.f32 %f5307, 0f00000000; mov.u32 %r159, 0; st.local.u32 [%rd10], %r159; st.local.u32 [%rd10+4], %r159; st.local.u32 [%rd10+8], %r159; st.local.u32 [%rd10+12], %r159; st.local.u32 [%rd10+16], %r159; st.local.u32 [%rd10+20], %r159; st.local.u32 [%rd10+24], %r159; st.local.u32 [%rd10+28], %r159; st.local.u32 [%rd10+32], %r159; add.s64 %rd1, %rd9, 260; ld.local.u32 %r168, [%rd9+260]; setp.lt.s32 %p7, %r168, 1; mov.f32 %f5308, %f5307; @%p7 bra $L__BB105_546; ld.const.u64 %rd11, [kernel_params+2160]; cvta.to.global.u64 %rd2, %rd11; ld.const.v4.f32 {%f1411, %f1412, %f1413, %f1414}, [kernel_params+1184]; mov.f32 %f5308, 0f00000000; mov.u16 %rs12, 0; mov.u16 %rs27, 1; mov.pred %p379, -1; mov.pred %p390, 0; mov.f32 %f5307, %f5308; bra.uni $L__BB105_2; $L__BB105_402: neg.ftz.f32 %f3652, %f940; div.approx.ftz.f32 %f3653, %f3652, %f3445; neg.ftz.f32 %f3654, %f941; div.approx.ftz.f32 %f3655, %f3654, %f3446; fma.rn.ftz.f32 %f3656, %f3653, %f3653, %f943; fma.rn.ftz.f32 %f3657, %f3655, %f3655, %f3656; mul.ftz.f32 %f3658, %f3657, 0f40490FDB; mul.ftz.f32 %f3659, %f3657, %f3658; mul.ftz.f32 %f3660, %f3445, %f3659; mul.ftz.f32 %f3661, %f3446, %f3660; mov.f32 %f3662, 0f33D6BF95; max.ftz.f32 %f3663, %f3661, %f3662; rcp.approx.ftz.f32 %f5192, %f3663; $L__BB105_404: add.ftz.f32 %f3673, %f5241, 0f3F800000; mov.f32 %f5217, 0f3F800000; sub.ftz.f32 %f3674, %f5217, %f3673; add.ftz.f32 %f3675, %f3674, %f5191; rcp.approx.ftz.f32 %f3676, %f3675; mul.ftz.f32 %f3677, %f3676, 0f3E800000; div.approx.ftz.f32 %f3678, %f3677, %f5189; mul.ftz.f32 %f5296, %f3678, %f5192; fma.rn.ftz.f32 %f948, %f924, %f924, 0fBF800000; abs.ftz.f32 %f949, %f5217; fma.rn.ftz.f32 %f950, %f949, %f949, %f948; setp.leu.ftz.f32 %p375, %f950, 0f00000000; mov.f32 %f5193, %f5217; @%p375 bra $L__BB105_406; sqrt.approx.ftz.f32 %f3679, %f950; sub.ftz.f32 %f3680, %f3679, %f949; add.ftz.f32 %f3681, %f949, %f3679; div.approx.ftz.f32 %f3682, %f3680, %f3681; fma.rn.ftz.f32 %f3683, %f949, %f3681, 0fBF800000; fma.rn.ftz.f32 %f3684, %f949, %f3680, 0f3F800000; div.approx.ftz.f32 %f3685, %f3683, %f3684; mul.ftz.f32 %f3686, %f3682, 0f3F000000; mul.ftz.f32 %f3687, %f3682, %f3686; fma.rn.ftz.f32 %f3688, %f3685, %f3685, 0f3F800000; mul.ftz.f32 %f5193, %f3687, %f3688; $L__BB105_406: sub.ftz.f32 %f5194, %f5217, %f5193; rcp.approx.ftz.f32 %f954, %f5194; setp.ne.s32 %p376, %r5, 15; mov.f32 %f5238, %f5217; mov.f32 %f5239, %f5217; mov.f32 %f5240, %f5217; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p376 bra $L__BB105_410; mul.ftz.f32 %f3692, %f5188, %f941; fma.rn.ftz.f32 %f3693, %f5187, %f940, %f3692; fma.rn.ftz.f32 %f3694, %f5189, %f942, %f3693; abs.ftz.f32 %f955, %f3694; fma.rn.ftz.f32 %f956, %f955, %f955, %f948; setp.leu.ftz.f32 %p377, %f956, 0f00000000; @%p377 bra $L__BB105_409; sqrt.approx.ftz.f32 %f3695, %f956; sub.ftz.f32 %f3696, %f3695, %f955; add.ftz.f32 %f3697, %f955, %f3695; div.approx.ftz.f32 %f3698, %f3696, %f3697; fma.rn.ftz.f32 %f3699, %f955, %f3697, 0fBF800000; fma.rn.ftz.f32 %f3700, %f955, %f3696, 0f3F800000; div.approx.ftz.f32 %f3701, %f3699, %f3700; mul.ftz.f32 %f3702, %f3698, 0f3F000000; mul.ftz.f32 %f3703, %f3698, %f3702; fma.rn.ftz.f32 %f3704, %f3701, %f3701, 0f3F800000; mul.ftz.f32 %f3705, %f3703, %f3704; sub.ftz.f32 %f5194, %f3705, %f5193; $L__BB105_409: mul.ftz.f32 %f3706, %f954, %f5194; mov.f32 %f3707, 0f3F800000; sub.ftz.f32 %f3708, %f3707, %f3706; fma.rn.ftz.f32 %f5238, %f925, %f3708, %f3706; fma.rn.ftz.f32 %f5239, %f926, %f3708, %f3706; fma.rn.ftz.f32 %f5240, %f927, %f3708, %f3706; mul.ftz.f32 %f962, %f5296, %f5238; mul.ftz.f32 %f5295, %f5296, %f5239; mul.ftz.f32 %f5294, %f5296, %f5240; mov.f32 %f5296, %f962; $L__BB105_410: ld.local.u32 %r166, [%rd1+-4]; mov.f32 %f5215, 0f00000000; neg.ftz.f32 %f5237, %f5189; mov.u32 %r165, 0; mov.f32 %f5216, %f5217; $L__BB105_411: mov.u32 %r23, %r165; mad.lo.s32 %r166, %r166, 1103515245, 12345; cvt.rn.f32.u32 %f3712, %r166; mul.ftz.f32 %f987, %f3712, 0f2F800000; setp.gt.ftz.f32 %p380, %f5237, 0f3F7FF972; mov.pred %p541, %p379; @%p380 bra $L__BB105_421; bra.uni $L__BB105_412; $L__BB105_421: @%p541 bra $L__BB105_454; mad.lo.s32 %r132, %r166, 1103515245, 12345; cvt.rn.f32.u32 %f3732, %r132; mul.ftz.f32 %f998, %f3732, 0f2F800000; mad.lo.s32 %r166, %r132, 1103515245, 12345; cvt.rn.f32.u32 %f3733, %r166; mul.ftz.f32 %f999, %f3733, 0f2F800000; neg.ftz.f32 %f1000, %f5235; mul.ftz.f32 %f3734, %f3445, %f1000; neg.ftz.f32 %f1001, %f5236; mul.ftz.f32 %f3735, %f3446, %f1001; mul.ftz.f32 %f3736, %f3735, %f3735; fma.rn.ftz.f32 %f3737, %f3734, %f3734, %f3736; fma.rn.ftz.f32 %f3738, %f5237, %f5237, %f3737; sqrt.approx.ftz.f32 %f3739, %f3738; div.approx.ftz.f32 %f5220, %f3734, %f3739; div.approx.ftz.f32 %f5221, %f3735, %f3739; neg.ftz.f32 %f1004, %f5237; div.approx.ftz.f32 %f1005, %f1004, %f3739; setp.gt.ftz.f32 %p393, %f1005, 0f3F7FF972; @%p393 bra $L__BB105_428; neg.ftz.f32 %f4972, %f5237; div.approx.ftz.f32 %f4971, %f4972, %f3739; abs.ftz.f32 %f3740, %f4971; setp.lt.ftz.f32 %p394, %f3740, 0f358637BD; @%p394 bra $L__BB105_428; bra.uni $L__BB105_424; $L__BB105_428: mov.f32 %f3784, 0f3F800000; sub.ftz.f32 %f3785, %f3784, %f999; mov.f32 %f3786, 0f33D6BF95; max.ftz.f32 %f3787, %f3785, %f3786; div.approx.ftz.f32 %f3788, %f999, %f3787; sqrt.approx.ftz.f32 %f3789, %f3788; mul.ftz.f32 %f3790, %f998, 0f40C90FDB; cos.approx.ftz.f32 %f3791, %f3790; sin.approx.ftz.f32 %f3792, %f3790; mul.ftz.f32 %f5219, %f3789, %f3792; mul.ftz.f32 %f5218, %f3789, %f3791; $L__BB105_429: mul.ftz.f32 %f3793, %f5221, %f5221; fma.rn.ftz.f32 %f3794, %f5220, %f5220, %f3793; add.ftz.f32 %f3795, %f3794, 0f00000000; sqrt.approx.ftz.f32 %f1021, %f3795; setp.neu.ftz.f32 %p402, %f1021, 0f00000000; @%p402 bra $L__BB105_430; bra.uni $L__BB105_431; $L__BB105_430: rcp.approx.ftz.f32 %f3796, %f1021; mul.ftz.f32 %f5221, %f5221, %f3796; mul.ftz.f32 %f5220, %f5220, %f3796; $L__BB105_431: mul.ftz.f32 %f3797, %f5218, %f5220; mul.ftz.f32 %f3798, %f5219, %f5221; sub.ftz.f32 %f3799, %f3797, %f3798; mul.ftz.f32 %f3800, %f3445, %f3799; mul.ftz.f32 %f3801, %f5219, %f5220; fma.rn.ftz.f32 %f3802, %f5218, %f5221, %f3801; mul.ftz.f32 %f3803, %f3446, %f3802; neg.ftz.f32 %f3804, %f3800; neg.ftz.f32 %f3805, %f3803; mul.ftz.f32 %f3806, %f3803, %f3803; fma.rn.ftz.f32 %f3807, %f3800, %f3800, %f3806; add.ftz.f32 %f3808, %f3807, 0f3F800000; mov.f32 %f3809, 0f3F800000; sqrt.approx.ftz.f32 %f3810, %f3808; div.approx.ftz.f32 %f1031, %f3804, %f3810; div.approx.ftz.f32 %f1032, %f3805, %f3810; div.approx.ftz.f32 %f1033, %f3809, %f3810; setp.eq.s32 %p403, %r23, 0; @%p403 bra $L__BB105_442; setp.gt.ftz.f32 %p534, %f5237, 0f3F7FF972; mov.f32 %f5223, 0f00000000; @%p534 bra $L__BB105_439; sub.ftz.f32 %f3817, %f5184, %f5235; sub.ftz.f32 %f3818, %f5185, %f5236; mul.ftz.f32 %f3819, %f3818, %f3818; fma.rn.ftz.f32 %f3820, %f3817, %f3817, %f3819; sub.ftz.f32 %f3821, %f5186, %f5237; fma.rn.ftz.f32 %f3822, %f3821, %f3821, %f3820; sqrt.approx.ftz.f32 %f3823, %f3822; div.approx.ftz.f32 %f1036, %f3821, %f3823; setp.lt.ftz.f32 %p405, %f1036, 0f00000000; @%p405 bra $L__BB105_439; sub.ftz.f32 %f4997, %f5186, %f5237; div.approx.ftz.f32 %f4996, %f4997, %f3823; sub.ftz.f32 %f4995, %f5184, %f5235; div.approx.ftz.f32 %f4994, %f4995, %f3823; sub.ftz.f32 %f4993, %f5185, %f5236; div.approx.ftz.f32 %f4992, %f4993, %f3823; mov.f32 %f5223, 0f00000000; neg.ftz.f32 %f4981, %f5237; neg.ftz.f32 %f4980, %f5235; neg.ftz.f32 %f4979, %f5236; mul.ftz.f32 %f3827, %f4992, %f4979; fma.rn.ftz.f32 %f3828, %f4994, %f4980, %f3827; fma.rn.ftz.f32 %f1037, %f4996, %f4981, %f3828; setp.lt.ftz.f32 %p406, %f1037, 0f00000000; @%p406 bra $L__BB105_439; sub.ftz.f32 %f4999, %f5186, %f5237; div.approx.ftz.f32 %f4998, %f4999, %f3823; setp.eq.ftz.f32 %p535, %f3445, %f3446; mul.ftz.f32 %f1038, %f4998, %f4998; @%p535 bra $L__BB105_437; bra.uni $L__BB105_436; $L__BB105_437: mov.f32 %f4989, 0f3F800000; mul.ftz.f32 %f4982, %f3445, %f3445; sub.ftz.f32 %f3842, %f4989, %f1038; fma.rn.ftz.f32 %f3843, %f4982, %f1038, %f3842; mul.ftz.f32 %f3844, %f3843, 0f40490FDB; mul.ftz.f32 %f3845, %f3843, %f3844; mov.f32 %f3846, 0f33D6BF95; max.ftz.f32 %f3847, %f3845, %f3846; div.approx.ftz.f32 %f5222, %f4982, %f3847; bra.uni $L__BB105_438; $L__BB105_412: setp.lt.ftz.f32 %p381, %f5237, 0fBF7FF972; @%p381 bra $L__BB105_419; bra.uni $L__BB105_413; $L__BB105_419: mul.ftz.f32 %f5216, %f5216, %f987; cvt.ftz.sat.f32.f32 %f3729, %f5216; fma.rn.ftz.f32 %f5217, %f3729, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p391, %f5237, 0f3727C5AC; mov.f32 %f5215, 0f00000000; mov.pred %p541, %p390; @%p391 bra $L__BB105_421; lg2.approx.ftz.f32 %f3730, %f5216; mul.ftz.f32 %f3731, %f5241, %f3730; ex2.approx.ftz.f32 %f5215, %f3731; mov.pred %p541, %p390; bra.uni $L__BB105_421; $L__BB105_413: abs.ftz.f32 %f3713, %f5237; setp.ltu.ftz.f32 %p383, %f3713, 0f38D1B717; mov.pred %p541, %p390; @%p383 bra $L__BB105_421; mov.f32 %f5214, 0f3F800000; sub.ftz.f32 %f3715, %f5214, %f5215; setp.lt.ftz.f32 %p385, %f3715, %f987; mov.pred %p541, %p379; @%p385 bra $L__BB105_421; setp.ge.ftz.f32 %p386, %f5241, 0f00000000; @%p386 bra $L__BB105_417; mov.f32 %f3717, 0f3F800000; sub.ftz.f32 %f3718, %f3717, %f987; lg2.approx.ftz.f32 %f3719, %f3718; mov.f32 %f3720, 0fBF800000; div.approx.ftz.f32 %f3721, %f3720, %f5241; mul.ftz.f32 %f3722, %f3721, %f3719; ex2.approx.ftz.f32 %f3723, %f3722; mul.ftz.f32 %f5214, %f5216, %f3723; $L__BB105_417: cvt.ftz.sat.f32.f32 %f3725, %f5214; fma.rn.ftz.f32 %f5217, %f3725, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p388, %f5237, 0f3727C5AC; mov.f32 %f5215, 0f00000000; mov.f32 %f5216, %f5214; mov.pred %p541, %p390; @%p388 bra $L__BB105_421; lg2.approx.ftz.f32 %f3726, %f5214; mul.ftz.f32 %f3727, %f5241, %f3726; ex2.approx.ftz.f32 %f5215, %f3727; mov.f32 %f5216, %f5214; mov.pred %p541, %p390; bra.uni $L__BB105_421; $L__BB105_424: neg.ftz.f32 %f4974, %f5237; div.approx.ftz.f32 %f4973, %f4974, %f3739; add.ftz.f32 %f3743, %f4973, 0f3F800000; mul.ftz.f32 %f1006, %f3743, 0f3F000000; setp.lt.ftz.f32 %p395, %f1006, 0f38D1B717; mov.f32 %f5218, 0f00000000; mov.f32 %f5219, %f5218; @%p395 bra $L__BB105_429; neg.ftz.f32 %f4976, %f5237; div.approx.ftz.f32 %f4975, %f4976, %f3739; add.ftz.f32 %f3746, %f999, %f999; mul.ftz.f32 %f3747, %f3746, %f1006; div.approx.ftz.f32 %f3748, %f3747, %f4975; add.ftz.f32 %f1007, %f3748, 0fBF800000; mul.ftz.f32 %f1008, %f1007, %f1007; add.ftz.f32 %f1009, %f1008, 0fBF800000; abs.ftz.f32 %f3749, %f1009; setp.lt.ftz.f32 %p396, %f3749, 0f33D6BF95; mov.f32 %f5219, %f5218; @%p396 bra $L__BB105_429; neg.ftz.f32 %f4978, %f5237; div.approx.ftz.f32 %f4977, %f4978, %f3739; mul.ftz.f32 %f3750, %f4977, %f4977; mov.f32 %f3751, 0f3F800000; sub.ftz.f32 %f3752, %f3751, %f3750; mov.f32 %f3753, 0f00000000; max.ftz.f32 %f3754, %f3752, %f3753; sqrt.approx.ftz.f32 %f3755, %f3754; div.approx.ftz.f32 %f3756, %f3755, %f4977; rcp.approx.ftz.f32 %f3757, %f1009; mul.ftz.f32 %f3758, %f3756, %f3756; mul.ftz.f32 %f3759, %f3758, %f3757; mul.ftz.f32 %f3760, %f3757, %f3759; sub.ftz.f32 %f3761, %f1008, %f3758; mul.ftz.f32 %f3762, %f3761, %f3757; sub.ftz.f32 %f3763, %f3760, %f3762; max.ftz.f32 %f3764, %f3763, %f3753; sqrt.approx.ftz.f32 %f3765, %f3764; mul.ftz.f32 %f3766, %f3756, %f3757; add.ftz.f32 %f3767, %f3766, %f3765; rcp.approx.ftz.f32 %f3768, %f3756; setp.gt.ftz.f32 %p397, %f3767, %f3768; setp.lt.ftz.f32 %p398, %f1007, 0f00000000; or.pred %p399, %p398, %p397; sub.ftz.f32 %f3769, %f3766, %f3765; mov.f32 %f3770, 0f3F000000; sub.ftz.f32 %f3771, %f3770, %f998; add.ftz.f32 %f3772, %f998, 0fBF000000; setp.ge.ftz.f32 %p400, %f998, 0f3F000000; selp.f32 %f3773, %f3772, %f3771, %p400; add.ftz.f32 %f3774, %f3773, %f3773; fma.rn.ftz.f32 %f3775, %f3774, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f3776, %f3774, %f3775, 0f3EED4413; mul.ftz.f32 %f3777, %f3774, %f3776; fma.rn.ftz.f32 %f3778, %f3774, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f3779, %f3774, %f3778, 0fBF800000; fma.rn.ftz.f32 %f3780, %f3774, %f3779, 0f3F191676; div.approx.ftz.f32 %f3781, %f3777, %f3780; selp.f32 %f5218, %f3769, %f3767, %p399; fma.rn.ftz.f32 %f3782, %f5218, %f5218, 0f3F800000; sqrt.approx.ftz.f32 %f3783, %f3782; mul.ftz.f32 %f5219, %f3781, %f3783; setp.ltu.ftz.f32 %p401, %f998, 0f3F000000; @%p401 bra $L__BB105_427; bra.uni $L__BB105_429; $L__BB105_427: neg.ftz.f32 %f5219, %f5219; bra.uni $L__BB105_429; $L__BB105_436: sub.ftz.f32 %f5003, %f5184, %f5235; div.approx.ftz.f32 %f5002, %f5003, %f3823; sub.ftz.f32 %f5001, %f5185, %f5236; div.approx.ftz.f32 %f5000, %f5001, %f3823; neg.ftz.f32 %f3829, %f5002; div.approx.ftz.f32 %f3830, %f3829, %f3445; neg.ftz.f32 %f3831, %f5000; div.approx.ftz.f32 %f3832, %f3831, %f3446; fma.rn.ftz.f32 %f3833, %f3830, %f3830, %f1038; fma.rn.ftz.f32 %f3834, %f3832, %f3832, %f3833; mul.ftz.f32 %f3835, %f3834, 0f40490FDB; mul.ftz.f32 %f3836, %f3834, %f3835; mul.ftz.f32 %f3837, %f3445, %f3836; mul.ftz.f32 %f3838, %f3446, %f3837; mov.f32 %f3839, 0f33D6BF95; max.ftz.f32 %f3840, %f3838, %f3839; rcp.approx.ftz.f32 %f5222, %f3840; $L__BB105_438: setp.lt.ftz.f32 %p408, %f5237, 0fBF7FF972; mul.ftz.f32 %f3848, %f5237, %f5241; selp.f32 %f3849, 0f3F800000, %f3848, %p408; mul.ftz.f32 %f3850, %f3849, %f1037; mov.f32 %f3851, 0f33D6BF95; max.ftz.f32 %f3852, %f3850, %f3851; mov.f32 %f3853, 0f00000000; max.ftz.f32 %f3854, %f3853, %f1037; mul.ftz.f32 %f3855, %f3854, 0f3E800000; div.approx.ftz.f32 %f3856, %f3855, %f3852; mul.ftz.f32 %f5223, %f3856, %f5222; $L__BB105_439: setp.gt.ftz.f32 %p536, %f5186, 0f3F7FF972; selp.f32 %f5224, 0f3F800000, 0f00000000, %p536; @%p536 bra $L__BB105_441; add.ftz.f32 %f3857, %f5217, 0f3F800000; mul.ftz.f32 %f3858, %f3857, 0f3F000000; cvt.ftz.sat.f32.f32 %f3859, %f3858; lg2.approx.ftz.f32 %f3860, %f3859; mul.ftz.f32 %f3861, %f5191, %f3860; ex2.approx.ftz.f32 %f5224, %f3861; $L__BB105_441: mul.ftz.f32 %f3862, %f5238, %f5223; mul.ftz.f32 %f3863, %f5238, %f3862; fma.rn.ftz.f32 %f5296, %f3863, %f5224, %f5296; mul.ftz.f32 %f3864, %f5239, %f5223; mul.ftz.f32 %f3865, %f5239, %f3864; fma.rn.ftz.f32 %f5295, %f3865, %f5224, %f5295; mul.ftz.f32 %f3866, %f5240, %f5223; mul.ftz.f32 %f3867, %f5240, %f3866; fma.rn.ftz.f32 %f5294, %f3867, %f5224, %f5294; $L__BB105_442: add.s32 %r165, %r23, 1; setp.gt.u32 %p410, %r165, 9; @%p410 bra $L__BB105_453; setp.eq.s32 %p538, %r23, 0; setp.ne.s32 %p537, %r5, 15; neg.ftz.f32 %f4986, %f5237; neg.ftz.f32 %f4985, %f5235; neg.ftz.f32 %f4984, %f5236; mul.ftz.f32 %f3868, %f1032, %f4984; fma.rn.ftz.f32 %f3869, %f1031, %f4985, %f3868; fma.rn.ftz.f32 %f1056, %f1033, %f4986, %f3869; or.pred %p413, %p537, %p538; @%p413 bra $L__BB105_447; mov.f32 %f4988, 0f3F800000; sub.ftz.f32 %f5228, %f4988, %f5193; abs.ftz.f32 %f1057, %f1056; fma.rn.ftz.f32 %f1058, %f1057, %f1057, %f948; setp.leu.ftz.f32 %p414, %f1058, 0f00000000; @%p414 bra $L__BB105_446; sqrt.approx.ftz.f32 %f3870, %f1058; sub.ftz.f32 %f3871, %f3870, %f1057; add.ftz.f32 %f3872, %f1057, %f3870; div.approx.ftz.f32 %f3873, %f3871, %f3872; fma.rn.ftz.f32 %f3874, %f1057, %f3872, 0fBF800000; fma.rn.ftz.f32 %f3875, %f1057, %f3871, 0f3F800000; div.approx.ftz.f32 %f3876, %f3874, %f3875; mul.ftz.f32 %f3877, %f3873, 0f3F000000; mul.ftz.f32 %f3878, %f3873, %f3877; fma.rn.ftz.f32 %f3879, %f3876, %f3876, 0f3F800000; mul.ftz.f32 %f3880, %f3878, %f3879; sub.ftz.f32 %f5228, %f3880, %f5193; $L__BB105_446: mul.ftz.f32 %f3881, %f954, %f5228; mov.f32 %f3882, 0f3F800000; sub.ftz.f32 %f3883, %f3882, %f3881; fma.rn.ftz.f32 %f3884, %f925, %f3883, %f3881; fma.rn.ftz.f32 %f3885, %f926, %f3883, %f3881; fma.rn.ftz.f32 %f3886, %f927, %f3883, %f3881; mul.ftz.f32 %f5238, %f5238, %f3884; mul.ftz.f32 %f5239, %f5239, %f3885; mul.ftz.f32 %f5240, %f5240, %f3886; $L__BB105_447: mov.f32 %f5005, 0f3F800000; div.approx.ftz.f32 %f5004, %f5005, %f3810; add.ftz.f32 %f3888, %f1031, %f1031; add.ftz.f32 %f3889, %f1032, %f1032; add.ftz.f32 %f3890, %f5004, %f5004; fma.rn.ftz.f32 %f5235, %f3888, %f1056, %f5235; fma.rn.ftz.f32 %f5236, %f3889, %f1056, %f5236; fma.rn.ftz.f32 %f5237, %f3890, %f1056, %f5237; setp.gt.ftz.f32 %p415, %f5237, 0f3F7FF972; mov.f32 %f5241, 0f00000000; @%p415 bra $L__BB105_450; setp.lt.ftz.f32 %p416, %f5237, 0fBF7FF972; mov.f32 %f5241, 0fBF7FF972; @%p416 bra $L__BB105_450; mul.ftz.f32 %f3892, %f5237, %f5237; mov.f32 %f3893, 0f33D6BF95; max.ftz.f32 %f3894, %f3892, %f3893; rcp.approx.ftz.f32 %f3895, %f3894; mul.ftz.f32 %f3896, %f3445, %f5235; mul.ftz.f32 %f3897, %f3446, %f5236; mul.ftz.f32 %f3898, %f3897, %f3897; fma.rn.ftz.f32 %f3899, %f3896, %f3896, %f3898; fma.rn.ftz.f32 %f3900, %f3899, %f3895, 0f3F800000; sqrt.approx.ftz.f32 %f3901, %f3900; neg.ftz.f32 %f3902, %f3901; setp.le.ftz.f32 %p417, %f5237, 0f00000000; selp.f32 %f3903, %f3902, %f3901, %p417; add.ftz.f32 %f3904, %f3903, 0fBF800000; mul.ftz.f32 %f5241, %f3904, 0f3F000000; $L__BB105_450: mov.f32 %f5215, 0f3F800000; mul.ftz.f32 %f3906, %f921, %f5238; setp.eq.s32 %p418, %r5, 15; selp.f32 %f5238, %f5238, %f3906, %p418; mul.ftz.f32 %f3907, %f922, %f5239; selp.f32 %f5239, %f5239, %f3907, %p418; mul.ftz.f32 %f3908, %f923, %f5240; selp.f32 %f5240, %f5240, %f3908, %p418; add.ftz.f32 %f3909, %f5217, 0f3F800000; mul.ftz.f32 %f3910, %f3909, 0f3F000000; cvt.ftz.sat.f32.f32 %f5216, %f3910; @%p415 bra $L__BB105_453; setp.lt.ftz.f32 %p420, %f5237, 0f3727C5AC; mov.f32 %f5215, 0f00000000; @%p420 bra $L__BB105_453; lg2.approx.ftz.f32 %f3912, %f5216; mul.ftz.f32 %f3913, %f5241, %f3912; ex2.approx.ftz.f32 %f5215, %f3913; $L__BB105_453: setp.lt.u32 %p421, %r165, 10; @%p421 bra $L__BB105_411; $L__BB105_454: st.local.u32 [%rd1+-4], %r166; setp.eq.s16 %p422, %rs41, 0; @%p422 bra $L__BB105_456; div.approx.ftz.f32 %f3914, %f5189, %f5186; abs.ftz.f32 %f3915, %f3914; mul.ftz.f32 %f5296, %f5296, %f3915; mul.ftz.f32 %f5295, %f5295, %f3915; mul.ftz.f32 %f5294, %f5294, %f3915; $L__BB105_456: mov.f32 %f5297, %f5183; bra.uni $L__BB105_524; $L__BB105_2: mul.wide.s32 %rd13, %r159, 80; add.s64 %rd14, %rd9, %rd13; add.s64 %rd3, %rd14, 316; ld.local.u32 %r5, [%rd14+316]; setp.gt.s32 %p8, %r5, 34; @%p8 bra $L__BB105_545; setp.gt.s32 %p9, %r5, 31; @%p9 bra $L__BB105_544; and.b32 %r40, %r37, 264241152; setp.eq.s32 %p10, %r40, 0; mov.u16 %rs36, %rs12; @%p10 bra $L__BB105_9; shr.u32 %r41, %r37, 27; and.b32 %r42, %r41, 1; setp.eq.b32 %p11, %r42, 1; add.s32 %r43, %r5, -2; setp.lt.u32 %p12, %r43, 7; and.pred %p13, %p11, %p12; mov.u16 %rs36, 1; @%p13 bra $L__BB105_9; and.b32 %r44, %r37, 67108864; setp.eq.s32 %p14, %r44, 0; @%p14 bra $L__BB105_8; add.s32 %r45, %r5, -9; setp.lt.u32 %p15, %r45, 12; setp.eq.s32 %p16, %r5, 29; or.pred %p17, %p16, %p15; @%p17 bra $L__BB105_9; $L__BB105_8: shr.u32 %r46, %r37, 25; and.b32 %r47, %r46, 1; setp.eq.b32 %p18, %r47, 1; add.s32 %r48, %r5, -21; setp.lt.u32 %p19, %r48, 10; and.pred %p20, %p18, %p19; selp.u16 %rs36, 1, 0, %p20; $L__BB105_9: setp.ne.s16 %p21, %rs36, 0; @%p21 bra $L__BB105_544; mov.f32 %f5294, 0f00000000; setp.gt.s32 %p22, %r5, 13; @%p22 bra $L__BB105_41; bra.uni $L__BB105_11; $L__BB105_41: setp.gt.s32 %p23, %r5, 21; @%p23 bra $L__BB105_61; setp.gt.s32 %p31, %r5, 16; @%p31 bra $L__BB105_52; add.s32 %r49, %r5, -14; setp.lt.u32 %p35, %r49, 2; @%p35 bra $L__BB105_375; bra.uni $L__BB105_44; $L__BB105_375: ld.local.f32 %f876, [%rd1+-224]; ld.local.f32 %f877, [%rd1+-220]; ld.local.f32 %f878, [%rd1+-216]; ld.local.f32 %f3435, [%rd1+-248]; ld.local.f32 %f3436, [%rd1+-244]; mul.ftz.f32 %f3437, %f1405, %f3436; fma.rn.ftz.f32 %f3438, %f1404, %f3435, %f3437; ld.local.f32 %f3439, [%rd1+-240]; fma.rn.ftz.f32 %f3440, %f1406, %f3439, %f3438; setp.lt.ftz.f32 %p350, %f3440, 0f00000000; mov.f32 %f5296, 0f00000000; mov.f32 %f5297, %f5296; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p350 bra $L__BB105_524; ld.local.v2.f32 {%f3445, %f3446}, [%rd3+20]; mul.ftz.f32 %f881, %f3445, %f3446; setp.lt.ftz.f32 %p351, %f881, 0f33D6BF95; mov.f32 %f5297, %f5296; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p351 bra $L__BB105_524; ld.local.f32 %f882, [%rd3+8]; ld.local.f32 %f883, [%rd3+12]; mul.ftz.f32 %f3451, %f877, %f883; fma.rn.ftz.f32 %f3452, %f876, %f882, %f3451; ld.local.f32 %f884, [%rd3+16]; fma.rn.ftz.f32 %f5189, %f878, %f884, %f3452; setp.le.ftz.f32 %p352, %f5189, 0f00000000; mov.f32 %f5297, %f5296; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p352 bra $L__BB105_524; mul.ftz.f32 %f3457, %f1405, %f883; fma.rn.ftz.f32 %f3458, %f1404, %f882, %f3457; fma.rn.ftz.f32 %f886, %f1406, %f884, %f3458; setp.le.ftz.f32 %p353, %f886, 0f00000000; mov.f32 %f5297, %f5296; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p353 bra $L__BB105_524; setp.neu.ftz.f32 %p354, %f3445, %f3446; @%p354 bra $L__BB105_381; bra.uni $L__BB105_380; $L__BB105_381: ld.local.f32 %f3476, [%rd3+52]; mul.ftz.f32 %f3477, %f883, %f3476; ld.local.f32 %f3478, [%rd3+48]; mul.ftz.f32 %f3479, %f884, %f3478; sub.ftz.f32 %f3480, %f3477, %f3479; ld.local.f32 %f3481, [%rd3+44]; mul.ftz.f32 %f3482, %f884, %f3481; mul.ftz.f32 %f3483, %f882, %f3476; sub.ftz.f32 %f3484, %f3482, %f3483; mul.ftz.f32 %f3485, %f882, %f3478; mul.ftz.f32 %f3486, %f883, %f3481; sub.ftz.f32 %f3487, %f3485, %f3486; mul.ftz.f32 %f3488, %f3484, %f3484; fma.rn.ftz.f32 %f3489, %f3480, %f3480, %f3488; fma.rn.ftz.f32 %f3490, %f3487, %f3487, %f3489; sqrt.approx.ftz.f32 %f3491, %f3490; div.approx.ftz.f32 %f5175, %f3480, %f3491; div.approx.ftz.f32 %f5176, %f3484, %f3491; div.approx.ftz.f32 %f5177, %f3487, %f3491; mul.ftz.f32 %f3492, %f884, %f5176; mul.ftz.f32 %f3493, %f883, %f5177; sub.ftz.f32 %f5178, %f3492, %f3493; mul.ftz.f32 %f3494, %f882, %f5177; mul.ftz.f32 %f3495, %f884, %f5175; sub.ftz.f32 %f5179, %f3494, %f3495; mul.ftz.f32 %f3496, %f883, %f5175; mul.ftz.f32 %f3497, %f882, %f5176; sub.ftz.f32 %f5180, %f3496, %f3497; bra.uni $L__BB105_382; $L__BB105_11: setp.gt.s32 %p37, %r5, 7; @%p37 bra $L__BB105_30; setp.gt.s32 %p41, %r5, 4; @%p41 bra $L__BB105_20; bra.uni $L__BB105_13; $L__BB105_20: setp.eq.s32 %p42, %r5, 5; @%p42 bra $L__BB105_514; setp.eq.s32 %p43, %r5, 6; @%p43 bra $L__BB105_510; setp.eq.s32 %p44, %r5, 7; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p44 bra $L__BB105_23; bra.uni $L__BB105_524; $L__BB105_23: ld.local.f32 %f4598, [%rd3+8]; ld.local.f32 %f4599, [%rd3+12]; mul.ftz.f32 %f4600, %f1405, %f4599; fma.rn.ftz.f32 %f4601, %f1404, %f4598, %f4600; ld.local.f32 %f4602, [%rd3+16]; fma.rn.ftz.f32 %f1283, %f1406, %f4602, %f4601; setp.ltu.ftz.f32 %p483, %f1283, 0f00000000; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p483 bra $L__BB105_29; ld.local.f32 %f4604, [%rd3+20]; mul.ftz.f32 %f1284, %f4604, 0f3FC90FDB; mov.f32 %f4605, 0f3FC90FDB; ld.local.f32 %f4606, [%rd3+24]; mul.ftz.f32 %f1285, %f4606, 0f3FC90FDB; mov.f32 %f5297, 0f00000000; max.ftz.f32 %f4608, %f1283, %f5297; mov.f32 %f4609, 0fBF800000; max.ftz.f32 %f4610, %f4608, %f4609; mov.f32 %f5282, 0f3F800000; min.ftz.f32 %f4611, %f4610, %f5282; abs.ftz.f32 %f4612, %f4611; sub.ftz.f32 %f4613, %f5282, %f4612; mul.ftz.f32 %f4614, %f4613, 0f3F000000; sqrt.approx.ftz.f32 %f4615, %f4614; setp.gt.ftz.f32 %p484, %f4612, 0f3F11EB85; selp.f32 %f4616, %f4615, %f4612, %p484; mul.ftz.f32 %f4617, %f4616, %f4616; mov.f32 %f4618, 0f3C94D2E9; mov.f32 %f4619, 0f3D53F941; fma.rn.ftz.f32 %f4620, %f4619, %f4617, %f4618; mov.f32 %f4621, 0f3D3F841F; fma.rn.ftz.f32 %f4622, %f4620, %f4617, %f4621; mov.f32 %f4623, 0f3D994929; fma.rn.ftz.f32 %f4624, %f4622, %f4617, %f4623; mov.f32 %f4625, 0f3E2AAB94; fma.rn.ftz.f32 %f4626, %f4624, %f4617, %f4625; mul.ftz.f32 %f4627, %f4617, %f4626; fma.rn.ftz.f32 %f4628, %f4627, %f4616, %f4616; add.ftz.f32 %f4629, %f4628, %f4628; sub.ftz.f32 %f4630, %f4605, %f4628; selp.f32 %f4631, %f4629, %f4630, %p484; setp.lt.ftz.f32 %p485, %f4611, 0f00000000; mov.f32 %f4632, 0f40490FDB; sub.ftz.f32 %f4633, %f4632, %f4631; selp.f32 %f1286, %f4633, %f4631, %p485; setp.lt.ftz.f32 %p486, %f1286, %f1284; @%p486 bra $L__BB105_27; add.ftz.f32 %f4635, %f1284, %f1285; setp.leu.ftz.f32 %p487, %f4635, %f1286; setp.eq.ftz.f32 %p488, %f1285, 0f00000000; or.pred %p489, %p488, %p487; mov.f32 %f5282, %f5297; @%p489 bra $L__BB105_27; sub.ftz.f32 %f4636, %f1286, %f1284; div.approx.ftz.f32 %f4637, %f4636, %f1285; mov.f32 %f4638, 0f3F800000; sub.ftz.f32 %f5282, %f4638, %f4637; $L__BB105_27: setp.leu.ftz.f32 %p490, %f5282, 0f00000000; mov.f32 %f5294, %f5297; @%p490 bra $L__BB105_29; add.ftz.f32 %f4643, %f1284, %f1285; min.ftz.f32 %f4645, %f4643, %f4605; cos.approx.ftz.f32 %f4646, %f4645; mov.f32 %f4647, 0f3F800000; sub.ftz.f32 %f4648, %f4647, %f4646; mov.f32 %f4649, 0f3E22F983; div.approx.ftz.f32 %f5297, %f4649, %f4648; mul.ftz.f32 %f5294, %f5282, %f5297; $L__BB105_29: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_61: setp.gt.s32 %p24, %r5, 26; @%p24 bra $L__BB105_65; setp.eq.s32 %p28, %r5, 22; @%p28 bra $L__BB105_33; setp.eq.s32 %p29, %r5, 23; @%p29 bra $L__BB105_469; setp.eq.s32 %p30, %r5, 24; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p30 bra $L__BB105_68; bra.uni $L__BB105_524; $L__BB105_30: add.s32 %r50, %r5, -10; setp.lt.u32 %p38, %r50, 3; @%p38 bra $L__BB105_469; bra.uni $L__BB105_31; $L__BB105_469: ld.local.f32 %f1173, [%rd1+-224]; ld.local.f32 %f1174, [%rd1+-220]; ld.local.f32 %f1175, [%rd1+-216]; ld.local.v2.f32 {%f4136, %f4137}, [%rd3+20]; ld.local.f32 %f4138, [%rd1+-248]; ld.local.f32 %f4139, [%rd1+-244]; mul.ftz.f32 %f4140, %f1405, %f4139; fma.rn.ftz.f32 %f4141, %f1404, %f4138, %f4140; ld.local.f32 %f4142, [%rd1+-240]; fma.rn.ftz.f32 %f1178, %f1406, %f4142, %f4141; setp.lt.ftz.f32 %p442, %f1178, 0f00000000; setp.eq.s32 %p443, %r5, 23; xor.pred %p444, %p442, %p443; mul.ftz.f32 %f5261, %f4136, %f4137; setp.le.ftz.f32 %p445, %f5261, 0f33D6BF95; or.pred %p446, %p444, %p445; mov.f32 %f5297, 0f00000000; mov.f32 %f5296, %f5297; mov.f32 %f5295, %f5297; mov.f32 %f5294, %f5297; @%p446 bra $L__BB105_524; ld.local.f32 %f1180, [%rd3+8]; ld.local.f32 %f1181, [%rd3+12]; mul.ftz.f32 %f4143, %f1174, %f1181; fma.rn.ftz.f32 %f4144, %f1173, %f1180, %f4143; ld.local.f32 %f1182, [%rd3+16]; fma.rn.ftz.f32 %f1183, %f1175, %f1182, %f4144; mul.ftz.f32 %f4145, %f1405, %f1181; fma.rn.ftz.f32 %f4146, %f1404, %f1180, %f4145; fma.rn.ftz.f32 %f1184, %f1406, %f1182, %f4146; @%p442 bra $L__BB105_488; bra.uni $L__BB105_471; $L__BB105_488: setp.le.ftz.f32 %p458, %f1183, 0f00000000; setp.ge.ftz.f32 %p459, %f1184, 0f00000000; or.pred %p460, %p458, %p459; mov.f32 %f5296, %f5297; mov.f32 %f5295, %f5297; mov.f32 %f5294, %f5297; @%p460 bra $L__BB105_524; ld.local.f32 %f4330, [%rd3+28]; fma.rn.ftz.f32 %f4331, %f1404, %f4330, %f1173; fma.rn.ftz.f32 %f4332, %f1405, %f4330, %f1174; fma.rn.ftz.f32 %f4333, %f1406, %f4330, %f1175; neg.ftz.f32 %f4334, %f4331; neg.ftz.f32 %f4335, %f4332; neg.ftz.f32 %f4336, %f4333; mul.ftz.f32 %f4337, %f4332, %f4332; fma.rn.ftz.f32 %f4338, %f4331, %f4331, %f4337; fma.rn.ftz.f32 %f4339, %f4333, %f4333, %f4338; sqrt.approx.ftz.f32 %f4340, %f4339; div.approx.ftz.f32 %f4341, %f4334, %f4340; div.approx.ftz.f32 %f4342, %f4335, %f4340; div.approx.ftz.f32 %f4343, %f4336, %f4340; mul.ftz.f32 %f4344, %f1174, %f4342; fma.rn.ftz.f32 %f4345, %f1173, %f4341, %f4344; fma.rn.ftz.f32 %f4346, %f1175, %f4343, %f4345; mul.ftz.f32 %f4347, %f1405, %f4342; fma.rn.ftz.f32 %f4348, %f1404, %f4341, %f4347; fma.rn.ftz.f32 %f4349, %f1406, %f4343, %f4348; mul.ftz.f32 %f4350, %f1181, %f4342; fma.rn.ftz.f32 %f4351, %f1180, %f4341, %f4350; fma.rn.ftz.f32 %f4352, %f1182, %f4343, %f4351; mul.ftz.f32 %f4353, %f4352, %f4352; mov.f32 %f4354, 0f3F800000; sub.ftz.f32 %f4355, %f4354, %f4353; div.approx.ftz.f32 %f4356, %f4355, %f4353; mul.ftz.f32 %f4357, %f4353, %f4353; mul.ftz.f32 %f4358, %f4357, 0f40490FDB; add.ftz.f32 %f4359, %f5261, %f4356; mul.ftz.f32 %f4360, %f4359, %f4358; mul.ftz.f32 %f4361, %f4359, %f4360; div.approx.ftz.f32 %f4362, %f5261, %f4361; mul.ftz.f32 %f4363, %f1183, %f1183; sub.ftz.f32 %f4364, %f4354, %f4363; mul.ftz.f32 %f4365, %f5261, %f4364; div.approx.ftz.f32 %f4366, %f4365, %f4363; add.ftz.f32 %f4367, %f4366, 0f3F800000; mov.f32 %f4368, 0f00000000; max.ftz.f32 %f4369, %f4367, %f4368; sqrt.approx.ftz.f32 %f4370, %f4369; add.ftz.f32 %f4371, %f4370, 0f3F800000; mov.f32 %f4372, 0f40000000; div.approx.ftz.f32 %f4373, %f4372, %f4371; mul.ftz.f32 %f4374, %f1184, %f1184; sub.ftz.f32 %f4375, %f4354, %f4374; mul.ftz.f32 %f4376, %f5261, %f4375; div.approx.ftz.f32 %f4377, %f4376, %f4374; add.ftz.f32 %f4378, %f4377, 0f3F800000; max.ftz.f32 %f4379, %f4378, %f4368; sqrt.approx.ftz.f32 %f4380, %f4379; add.ftz.f32 %f4381, %f4380, 0f3F800000; div.approx.ftz.f32 %f4382, %f4372, %f4381; mul.ftz.f32 %f4383, %f4373, %f4382; mul.ftz.f32 %f4384, %f4330, %f4330; mul.ftz.f32 %f4385, %f4384, %f4362; mul.ftz.f32 %f4386, %f1183, %f4339; div.approx.ftz.f32 %f4387, %f4385, %f4386; mul.ftz.f32 %f4388, %f4349, %f4346; abs.ftz.f32 %f4389, %f4388; mul.ftz.f32 %f4390, %f4383, %f4389; mul.ftz.f32 %f4391, %f4373, %f4389; mul.ftz.f32 %f5297, %f4387, %f4391; mul.ftz.f32 %f5296, %f4387, %f4390; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; bra.uni $L__BB105_524; $L__BB105_52: setp.eq.s32 %p32, %r5, 17; @%p32 bra $L__BB105_368; setp.eq.s32 %p33, %r5, 19; @%p33 bra $L__BB105_362; setp.eq.s32 %p34, %r5, 20; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p34 bra $L__BB105_55; bra.uni $L__BB105_524; $L__BB105_55: ld.local.f32 %f3023, [%rd3+8]; ld.local.f32 %f3024, [%rd3+12]; mul.ftz.f32 %f3025, %f1405, %f3024; fma.rn.ftz.f32 %f3026, %f1404, %f3023, %f3025; ld.local.f32 %f3027, [%rd3+16]; fma.rn.ftz.f32 %f3028, %f1406, %f3027, %f3026; setp.lt.ftz.f32 %p306, %f3028, 0f00000000; mov.f32 %f5294, 0f00000000; @%p306 bra $L__BB105_361; ld.local.f32 %f3033, [%rd1+-224]; ld.local.f32 %f3034, [%rd1+-220]; ld.local.f32 %f3035, [%rd1+-216]; ld.local.f32 %f762, [%rd3+40]; ld.local.f32 %f763, [%rd3+32]; ld.local.f32 %f3036, [%rd3+20]; ld.local.f32 %f3037, [%rd3+24]; mul.ftz.f32 %f3038, %f3034, %f3037; fma.rn.ftz.f32 %f3039, %f3033, %f3036, %f3038; ld.local.f32 %f3040, [%rd3+28]; fma.rn.ftz.f32 %f3041, %f3035, %f3040, %f3039; mul.ftz.f32 %f3042, %f3036, %f3041; mul.ftz.f32 %f3043, %f3037, %f3041; mul.ftz.f32 %f3044, %f3040, %f3041; sub.ftz.f32 %f3045, %f3033, %f3042; sub.ftz.f32 %f3046, %f3034, %f3043; sub.ftz.f32 %f3047, %f3035, %f3044; mul.ftz.f32 %f3048, %f3046, %f3046; fma.rn.ftz.f32 %f3049, %f3045, %f3045, %f3048; fma.rn.ftz.f32 %f3050, %f3047, %f3047, %f3049; sqrt.approx.ftz.f32 %f3051, %f3050; div.approx.ftz.f32 %f3052, %f3045, %f3051; div.approx.ftz.f32 %f3053, %f3046, %f3051; div.approx.ftz.f32 %f3054, %f3047, %f3051; abs.ftz.f32 %f3055, %f3041; setp.lt.ftz.f32 %p307, %f3055, 0f3F800000; mov.f32 %f3056, 0f3F800000; add.ftz.f32 %f3057, %f3055, 0fBF800000; add.ftz.f32 %f3058, %f3057, 0f3F800000; selp.f32 %f3059, %f3058, 0f3F800000, %p307; sub.ftz.f32 %f3060, %f3056, %f3059; sqrt.approx.ftz.f32 %f3061, %f3060; mul.ftz.f32 %f3062, %f3059, 0f3CB1485D; mov.f32 %f3063, 0f3D9FB43B; sub.ftz.f32 %f3064, %f3063, %f3062; fma.rn.ftz.f32 %f3065, %f3059, %f3064, 0fBE5A6B93; fma.rn.ftz.f32 %f3066, %f3059, %f3065, 0f3FC90FDB; mov.f32 %f3067, 0f3FC90FDB; mul.ftz.f32 %f3068, %f3061, %f3066; setp.lt.ftz.f32 %p308, %f3041, 0f00000000; mov.f32 %f3069, 0f40490FDB; sub.ftz.f32 %f3070, %f3069, %f3068; selp.f32 %f3071, %f3070, %f3068, %p308; sub.ftz.f32 %f764, %f3067, %f3071; mul.ftz.f32 %f3072, %f1405, %f3037; fma.rn.ftz.f32 %f3073, %f1404, %f3036, %f3072; fma.rn.ftz.f32 %f3074, %f1406, %f3040, %f3073; mul.ftz.f32 %f3075, %f3036, %f3074; mul.ftz.f32 %f3076, %f3037, %f3074; mul.ftz.f32 %f3077, %f3040, %f3074; sub.ftz.f32 %f3078, %f1404, %f3075; sub.ftz.f32 %f3079, %f1405, %f3076; sub.ftz.f32 %f3080, %f1406, %f3077; mul.ftz.f32 %f3081, %f3079, %f3079; fma.rn.ftz.f32 %f3082, %f3078, %f3078, %f3081; fma.rn.ftz.f32 %f3083, %f3080, %f3080, %f3082; sqrt.approx.ftz.f32 %f3084, %f3083; div.approx.ftz.f32 %f3085, %f3078, %f3084; div.approx.ftz.f32 %f3086, %f3079, %f3084; div.approx.ftz.f32 %f3087, %f3080, %f3084; abs.ftz.f32 %f3088, %f3074; setp.lt.ftz.f32 %p309, %f3088, 0f3F800000; add.ftz.f32 %f3089, %f3088, 0fBF800000; add.ftz.f32 %f3090, %f3089, 0f3F800000; selp.f32 %f3091, %f3090, 0f3F800000, %p309; sub.ftz.f32 %f3092, %f3056, %f3091; sqrt.approx.ftz.f32 %f3093, %f3092; mul.ftz.f32 %f3094, %f3091, 0f3CB1485D; sub.ftz.f32 %f3095, %f3063, %f3094; fma.rn.ftz.f32 %f3096, %f3091, %f3095, 0fBE5A6B93; fma.rn.ftz.f32 %f3097, %f3091, %f3096, 0f3FC90FDB; mul.ftz.f32 %f3098, %f3093, %f3097; setp.lt.ftz.f32 %p310, %f3074, 0f00000000; sub.ftz.f32 %f3099, %f3069, %f3098; selp.f32 %f3100, %f3099, %f3098, %p310; sub.ftz.f32 %f765, %f3067, %f3100; mul.ftz.f32 %f3101, %f3053, %f3086; fma.rn.ftz.f32 %f3102, %f3052, %f3085, %f3101; fma.rn.ftz.f32 %f766, %f3054, %f3087, %f3102; abs.ftz.f32 %f3103, %f765; sub.ftz.f32 %f3104, %f3067, %f3103; setp.lt.ftz.f32 %p311, %f3104, 0f3A83126F; setp.lt.ftz.f32 %p312, %f766, 0f00000000; or.pred %p313, %p312, %p311; @%p313 bra $L__BB105_361; ld.local.f32 %f3106, [%rd3+36]; rcp.approx.ftz.f32 %f767, %f3106; abs.ftz.f32 %f3108, %f766; setp.lt.ftz.f32 %p314, %f3108, 0f3F800000; add.ftz.f32 %f3109, %f3108, 0fBF800000; add.ftz.f32 %f3110, %f3109, 0f3F800000; selp.f32 %f3111, %f3110, 0f3F800000, %p314; sub.ftz.f32 %f3112, %f3056, %f3111; sqrt.approx.ftz.f32 %f3113, %f3112; fma.rn.ftz.f32 %f3114, %f3111, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f3115, %f3111, %f3114, 0fBE5A6B93; fma.rn.ftz.f32 %f3116, %f3111, %f3115, 0f3FC90FDB; mul.ftz.f32 %f768, %f3113, %f3116; mul.ftz.f32 %f3117, %f765, 0f3EA2F983; mov.b32 %r113, %f3117; and.b32 %r114, %r113, -2147483648; or.b32 %r115, %r114, 1056964608; mov.b32 %f3118, %r115; add.ftz.f32 %f3119, %f3117, %f3118; cvt.rzi.ftz.s32.f32 %r19, %f3119; cvt.rn.f32.s32 %f3120, %r19; fma.rn.ftz.f32 %f3121, %f3120, 0fC0490000, %f765; fma.rn.ftz.f32 %f3122, %f3120, 0fBA7DA000, %f3121; fma.rn.ftz.f32 %f3123, %f3120, 0fB4222000, %f3122; fma.rn.ftz.f32 %f3124, %f3120, 0fACB4611A, %f3123; add.ftz.f32 %f3125, %f3124, 0fBFC90FDB; add.ftz.f32 %f3126, %f3125, 0f3FC90FDB; mul.ftz.f32 %f3127, %f3126, %f3126; fma.rn.ftz.f32 %f3128, %f3127, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f3129, %f3127, %f3128, 0fBAB60B58; fma.rn.ftz.f32 %f3130, %f3127, %f3129, 0f3D2AAAAA; fma.rn.ftz.f32 %f3131, %f3127, %f3130, 0fBF000000; fma.rn.ftz.f32 %f769, %f3127, %f3131, 0f3F800000; add.ftz.f32 %f3132, %f764, 0f3FC90FDB; mul.ftz.f32 %f3133, %f3132, 0f3F000000; sub.ftz.f32 %f3134, %f3133, %f762; rcp.approx.ftz.f32 %f770, %f763; mul.ftz.f32 %f771, %f770, %f3134; abs.ftz.f32 %f772, %f771; setp.eq.ftz.f32 %p315, %f772, 0f00000000; mov.f32 %f5163, 0f00000000; mov.f32 %f5162, %f5163; @%p315 bra $L__BB105_355; setp.eq.ftz.f32 %p316, %f772, 0f3F800000; mov.f32 %f5162, %f3056; @%p316 bra $L__BB105_355; setp.gt.ftz.f32 %p317, %f772, 0f3F800000; @%p317 bra $L__BB105_354; bra.uni $L__BB105_60; $L__BB105_354: mov.f32 %f3137, 0f3F800000; div.approx.ftz.f32 %f5162, %f3137, %f772; bra.uni $L__BB105_355; $L__BB105_65: setp.eq.s32 %p25, %r5, 30; @%p25 bra $L__BB105_490; setp.eq.s32 %p26, %r5, 29; @%p26 bra $L__BB105_209; setp.ne.s32 %p27, %r5, 27; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p27 bra $L__BB105_524; $L__BB105_68: ld.local.v2.f32 {%f1423, %f1424}, [%rd3+20]; mul.ftz.f32 %f1425, %f1423, %f1424; setp.lt.ftz.f32 %p47, %f1425, 0f33D6BF95; mov.f32 %f5297, 0f00000000; mov.f32 %f5296, %f5297; mov.f32 %f5295, %f5297; mov.f32 %f5294, %f5297; @%p47 bra $L__BB105_524; ld.local.f32 %f1427, [%rd1+-224]; ld.local.f32 %f1428, [%rd1+-220]; ld.local.f32 %f1429, [%rd1+-216]; ld.local.f32 %f1430, [%rd3+12]; ld.local.f32 %f1431, [%rd3+8]; setp.neu.ftz.f32 %p48, %f1431, %f1430; ld.local.f32 %f1432, [%rd3+16]; setp.neu.ftz.f32 %p49, %f1431, %f1432; or.pred %p50, %p48, %p49; sub.ftz.f32 %f1433, %f1432, %f1430; neg.ftz.f32 %f1434, %f1430; neg.ftz.f32 %f1435, %f1432; selp.f32 %f1436, %f1435, %f1432, %p50; add.ftz.f32 %f1437, %f1431, %f1436; selp.f32 %f1438, %f1430, %f1434, %p50; sub.ftz.f32 %f1439, %f1438, %f1431; mul.ftz.f32 %f1440, %f1437, %f1437; fma.rn.ftz.f32 %f1441, %f1433, %f1433, %f1440; fma.rn.ftz.f32 %f1442, %f1439, %f1439, %f1441; sqrt.approx.ftz.f32 %f1443, %f1442; div.approx.ftz.f32 %f1444, %f1433, %f1443; div.approx.ftz.f32 %f1445, %f1437, %f1443; div.approx.ftz.f32 %f1446, %f1439, %f1443; mul.ftz.f32 %f1447, %f1430, %f1446; mul.ftz.f32 %f1448, %f1432, %f1445; sub.ftz.f32 %f1449, %f1447, %f1448; mul.ftz.f32 %f1450, %f1432, %f1444; mul.ftz.f32 %f1451, %f1431, %f1446; sub.ftz.f32 %f1452, %f1450, %f1451; mul.ftz.f32 %f1453, %f1431, %f1445; mul.ftz.f32 %f1454, %f1430, %f1444; sub.ftz.f32 %f1455, %f1453, %f1454; mul.ftz.f32 %f1456, %f1428, %f1445; fma.rn.ftz.f32 %f1457, %f1427, %f1444, %f1456; fma.rn.ftz.f32 %f5018, %f1429, %f1446, %f1457; mul.ftz.f32 %f1458, %f1428, %f1452; fma.rn.ftz.f32 %f1459, %f1427, %f1449, %f1458; fma.rn.ftz.f32 %f5019, %f1429, %f1455, %f1459; mul.ftz.f32 %f1460, %f1428, %f1430; fma.rn.ftz.f32 %f1461, %f1427, %f1431, %f1460; fma.rn.ftz.f32 %f5020, %f1429, %f1432, %f1461; mul.ftz.f32 %f1462, %f1405, %f1445; fma.rn.ftz.f32 %f1463, %f1404, %f1444, %f1462; fma.rn.ftz.f32 %f16, %f1406, %f1446, %f1463; mul.ftz.f32 %f1464, %f1405, %f1452; fma.rn.ftz.f32 %f1465, %f1404, %f1449, %f1464; fma.rn.ftz.f32 %f17, %f1406, %f1455, %f1465; mul.ftz.f32 %f1466, %f1405, %f1430; fma.rn.ftz.f32 %f1467, %f1404, %f1431, %f1466; fma.rn.ftz.f32 %f18, %f1406, %f1432, %f1467; mov.f32 %f5008, 0f00000000; setp.geu.ftz.f32 %p1, %f18, 0f00000000; mul.ftz.f32 %f19, %f5020, %f18; setp.gt.ftz.f32 %p52, %f19, 0f00000000; ld.local.f32 %f20, [%rd3+28]; mul.ftz.f32 %f1468, %f16, %f20; mul.ftz.f32 %f1469, %f17, %f20; mul.ftz.f32 %f1470, %f18, %f20; selp.f32 %f1471, %f16, %f1468, %p52; selp.f32 %f1472, %f17, %f1469, %p52; selp.f32 %f1473, %f18, %f1470, %p52; add.ftz.f32 %f1474, %f5018, %f1471; add.ftz.f32 %f1475, %f5019, %f1472; add.ftz.f32 %f1476, %f5020, %f1473; mul.ftz.f32 %f1477, %f1475, %f1475; fma.rn.ftz.f32 %f1478, %f1474, %f1474, %f1477; fma.rn.ftz.f32 %f1479, %f1476, %f1476, %f1478; sqrt.approx.ftz.f32 %f21, %f1479; rsqrt.approx.ftz.f32 %f1480, %f1479; mul.ftz.f32 %f1481, %f1480, %f1474; mul.ftz.f32 %f1482, %f1480, %f1475; mul.ftz.f32 %f1483, %f1480, %f1476; setp.lt.ftz.f32 %p53, %f1483, 0f00000000; neg.ftz.f32 %f1484, %f1481; neg.ftz.f32 %f1485, %f1482; neg.ftz.f32 %f1486, %f1483; selp.f32 %f22, %f1484, %f1481, %p53; selp.f32 %f23, %f1485, %f1482, %p53; selp.f32 %f24, %f1486, %f1483, %p53; setp.lt.ftz.f32 %p54, %f5020, 0f00000000; neg.ftz.f32 %f25, %f5018; neg.ftz.f32 %f26, %f5019; neg.ftz.f32 %f27, %f5020; selp.f32 %f28, %f25, %f5018, %p54; selp.f32 %f29, %f26, %f5019, %p54; selp.f32 %f30, %f27, %f5020, %p54; setp.gt.ftz.f32 %p55, %f30, 0f3F7FF972; @%p55 bra $L__BB105_72; setp.lt.ftz.f32 %p56, %f30, 0fBF7FF972; mov.f32 %f5008, 0fBF7FF972; @%p56 bra $L__BB105_72; mul.ftz.f32 %f1488, %f30, %f30; mov.f32 %f1489, 0f33D6BF95; max.ftz.f32 %f1490, %f1488, %f1489; rcp.approx.ftz.f32 %f1491, %f1490; mul.ftz.f32 %f1492, %f28, %f1423; mul.ftz.f32 %f1493, %f29, %f1423; mul.ftz.f32 %f1494, %f1493, %f1493; fma.rn.ftz.f32 %f1495, %f1492, %f1492, %f1494; fma.rn.ftz.f32 %f1496, %f1495, %f1491, 0f3F800000; sqrt.approx.ftz.f32 %f1497, %f1496; neg.ftz.f32 %f1498, %f1497; setp.le.ftz.f32 %p57, %f30, 0f00000000; selp.f32 %f1499, %f1498, %f1497, %p57; add.ftz.f32 %f1500, %f1499, 0fBF800000; mul.ftz.f32 %f5008, %f1500, 0f3F000000; $L__BB105_72: mul.ftz.f32 %f1502, %f24, %f24; mov.f32 %f5009, 0f3F800000; sub.ftz.f32 %f1503, %f5009, %f1502; mul.ftz.f32 %f33, %f1423, %f1423; fma.rn.ftz.f32 %f1504, %f33, %f1502, %f1503; mul.ftz.f32 %f1505, %f1504, 0f40490FDB; mul.ftz.f32 %f1506, %f1504, %f1505; mov.f32 %f1507, 0f33D6BF95; max.ftz.f32 %f1508, %f1506, %f1507; div.approx.ftz.f32 %f34, %f33, %f1508; mul.ftz.f32 %f1509, %f29, %f23; fma.rn.ftz.f32 %f1510, %f28, %f22, %f1509; fma.rn.ftz.f32 %f35, %f30, %f24, %f1510; abs.ftz.f32 %f36, %f35; fma.rn.ftz.f32 %f37, %f20, %f20, 0fBF800000; fma.rn.ftz.f32 %f38, %f36, %f36, %f37; setp.leu.ftz.f32 %p58, %f38, 0f00000000; @%p58 bra $L__BB105_74; sqrt.approx.ftz.f32 %f1511, %f38; sub.ftz.f32 %f1512, %f1511, %f36; add.ftz.f32 %f1513, %f36, %f1511; div.approx.ftz.f32 %f1514, %f1512, %f1513; fma.rn.ftz.f32 %f1515, %f36, %f1513, 0fBF800000; fma.rn.ftz.f32 %f1516, %f36, %f1512, 0f3F800000; div.approx.ftz.f32 %f1517, %f1515, %f1516; mul.ftz.f32 %f1518, %f1514, 0f3F000000; mul.ftz.f32 %f1519, %f1514, %f1518; fma.rn.ftz.f32 %f1520, %f1517, %f1517, 0f3F800000; mul.ftz.f32 %f5009, %f1519, %f1520; $L__BB105_74: setp.gt.ftz.f32 %p533, %f19, 0f00000000; mul.ftz.f32 %f1521, %f18, 0f3EA2F983; abs.ftz.f32 %f41, %f1521; add.ftz.f32 %f1522, %f5008, 0f3F800000; mul.ftz.f32 %f42, %f30, %f1522; @%p533 bra $L__BB105_78; bra.uni $L__BB105_75; $L__BB105_78: mul.ftz.f32 %f1580, %f34, 0f3E800000; max.ftz.f32 %f1582, %f42, %f1507; div.approx.ftz.f32 %f1583, %f1580, %f1582; mul.ftz.f32 %f1584, %f1423, 0fBFFE59F3; mul.ftz.f32 %f1585, %f1423, %f1584; mul.ftz.f32 %f1586, %f1585, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1587, %f1586; fma.rn.ftz.f32 %f1588, %f1587, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f1589, %f1423, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f1590, %f1423, %f1589, 0f417C8B44; fma.rn.ftz.f32 %f1591, %f1423, %f1590, 0fC1665E35; fma.rn.ftz.f32 %f1592, %f1423, %f1591, 0f40CE7881; fma.rn.ftz.f32 %f1593, %f1423, %f1592, 0fBF993E81; fma.rn.ftz.f32 %f1594, %f1423, %f1593, 0f3CE3C31E; fma.rn.ftz.f32 %f1595, %f1423, %f1594, 0f3BBA5D49; sub.ftz.f32 %f1596, %f1588, %f1595; cvt.ftz.sat.f32.f32 %f1597, %f1596; mov.f32 %f1598, 0f3F800000; sub.ftz.f32 %f1599, %f1598, %f1597; mul.ftz.f32 %f1600, %f41, %f1599; fma.rn.ftz.f32 %f1601, %f1583, %f1597, %f1600; mul.ftz.f32 %f5297, %f5009, %f1601; bra.uni $L__BB105_79; $L__BB105_31: setp.eq.s32 %p39, %r5, 8; @%p39 bra $L__BB105_509; setp.eq.s32 %p40, %r5, 13; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p40 bra $L__BB105_33; bra.uni $L__BB105_524; $L__BB105_33: ld.local.f32 %f1105, [%rd1+-224]; ld.local.f32 %f1106, [%rd1+-220]; ld.local.f32 %f1107, [%rd1+-216]; ld.local.v2.f32 {%f3920, %f3921}, [%rd3+20]; ld.local.f32 %f3922, [%rd1+-248]; ld.local.f32 %f3923, [%rd1+-244]; mul.ftz.f32 %f3924, %f1405, %f3923; fma.rn.ftz.f32 %f3925, %f1404, %f3922, %f3924; ld.local.f32 %f3926, [%rd1+-240]; fma.rn.ftz.f32 %f3927, %f1406, %f3926, %f3925; setp.lt.ftz.f32 %p423, %f3927, 0f00000000; setp.eq.s32 %p424, %r5, 22; xor.pred %p425, %p423, %p424; mul.ftz.f32 %f1110, %f3920, %f3921; setp.le.ftz.f32 %p426, %f1110, 0f33D6BF95; or.pred %p427, %p425, %p426; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p427 bra $L__BB105_468; bra.uni $L__BB105_34; $L__BB105_468: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_13: setp.eq.s32 %p45, %r5, 2; @%p45 bra $L__BB105_523; setp.eq.s32 %p46, %r5, 3; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p46 bra $L__BB105_15; bra.uni $L__BB105_524; $L__BB105_15: ld.local.f32 %f1344, [%rd3+8]; ld.local.f32 %f1345, [%rd3+12]; mul.ftz.f32 %f4727, %f1405, %f1345; fma.rn.ftz.f32 %f4728, %f1404, %f1344, %f4727; ld.local.f32 %f1346, [%rd3+16]; fma.rn.ftz.f32 %f1347, %f1406, %f1346, %f4728; setp.leu.ftz.f32 %p500, %f1347, 0f00000000; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p500 bra $L__BB105_19; ld.local.f32 %f4729, [%rd1+-224]; ld.local.f32 %f4730, [%rd1+-220]; ld.local.f32 %f4731, [%rd1+-216]; mov.f32 %f4732, 0f00000000; max.ftz.f32 %f1348, %f1347, %f4732; mul.ftz.f32 %f4733, %f4730, %f1345; fma.rn.ftz.f32 %f4734, %f4729, %f1344, %f4733; fma.rn.ftz.f32 %f4735, %f4731, %f1346, %f4734; max.ftz.f32 %f1349, %f4735, %f4732; mul.ftz.f32 %f4736, %f1405, %f4730; fma.rn.ftz.f32 %f4737, %f1404, %f4729, %f4736; fma.rn.ftz.f32 %f4738, %f1406, %f4731, %f4737; mul.ftz.f32 %f4739, %f1348, %f1349; sub.ftz.f32 %f5291, %f4738, %f4739; setp.leu.ftz.f32 %p501, %f5291, 0f00000000; @%p501 bra $L__BB105_18; max.ftz.f32 %f4740, %f1348, %f1349; add.ftz.f32 %f4741, %f4740, 0f00800000; div.approx.ftz.f32 %f5291, %f5291, %f4741; $L__BB105_18: ld.local.f32 %f4743, [%rd3+28]; ld.local.f32 %f4744, [%rd3+24]; fma.rn.ftz.f32 %f4745, %f5291, %f4743, %f4744; mul.ftz.f32 %f5294, %f1348, %f4745; mov.f32 %f5297, 0f3E22F983; $L__BB105_19: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_44: setp.eq.s32 %p36, %r5, 16; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; @%p36 bra $L__BB105_45; bra.uni $L__BB105_524; $L__BB105_45: ld.local.f32 %f3325, [%rd1+-248]; ld.local.f32 %f3326, [%rd1+-244]; mul.ftz.f32 %f3327, %f1405, %f3326; fma.rn.ftz.f32 %f3328, %f1404, %f3325, %f3327; ld.local.f32 %f3329, [%rd1+-240]; fma.rn.ftz.f32 %f3330, %f1406, %f3329, %f3328; ld.local.f32 %f838, [%rd3+8]; ld.local.f32 %f839, [%rd1+-224]; ld.local.f32 %f840, [%rd3+12]; ld.local.f32 %f841, [%rd1+-220]; mul.ftz.f32 %f3331, %f841, %f840; fma.rn.ftz.f32 %f3332, %f839, %f838, %f3331; ld.local.f32 %f842, [%rd3+16]; ld.local.f32 %f843, [%rd1+-216]; fma.rn.ftz.f32 %f844, %f843, %f842, %f3332; mul.ftz.f32 %f3333, %f1405, %f840; fma.rn.ftz.f32 %f3334, %f1404, %f838, %f3333; fma.rn.ftz.f32 %f845, %f1406, %f842, %f3334; setp.lt.ftz.f32 %p343, %f3330, 0f00000000; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p343 bra $L__BB105_374; ld.local.v2.f32 {%f3339, %f3340}, [%rd3+20]; max.ftz.f32 %f3341, %f3339, %f3340; setp.le.ftz.f32 %p344, %f3341, 0f38D1B717; mov.f32 %f5294, %f5297; @%p344 bra $L__BB105_374; setp.leu.ftz.f32 %p345, %f844, 0f00000000; setp.leu.ftz.f32 %p346, %f845, 0f00000000; or.pred %p347, %p346, %p345; mov.f32 %f5294, %f5297; @%p347 bra $L__BB105_374; add.ftz.f32 %f3346, %f1404, %f839; add.ftz.f32 %f3347, %f1405, %f841; mul.ftz.f32 %f3348, %f3347, %f3347; fma.rn.ftz.f32 %f3349, %f3346, %f3346, %f3348; add.ftz.f32 %f3350, %f1406, %f843; fma.rn.ftz.f32 %f3351, %f3350, %f3350, %f3349; sqrt.approx.ftz.f32 %f3352, %f3351; div.approx.ftz.f32 %f848, %f3346, %f3352; div.approx.ftz.f32 %f849, %f3347, %f3352; div.approx.ftz.f32 %f850, %f3350, %f3352; mul.ftz.f32 %f3353, %f841, %f849; fma.rn.ftz.f32 %f3354, %f839, %f848, %f3353; fma.rn.ftz.f32 %f3355, %f843, %f850, %f3354; abs.ftz.f32 %f3356, %f3355; mov.f32 %f3357, 0f358637BD; max.ftz.f32 %f851, %f3356, %f3357; mul.ftz.f32 %f3358, %f840, %f849; fma.rn.ftz.f32 %f3359, %f838, %f848, %f3358; fma.rn.ftz.f32 %f3360, %f842, %f850, %f3359; max.ftz.f32 %f852, %f3360, %f3357; max.ftz.f32 %f853, %f845, %f3357; mul.ftz.f32 %f3361, %f3339, %f3339; mov.f32 %f3362, 0f40000000; div.approx.ftz.f32 %f3363, %f3362, %f3361; add.ftz.f32 %f854, %f3363, 0fC0000000; mul.ftz.f32 %f3364, %f3340, %f3340; div.approx.ftz.f32 %f3365, %f3362, %f3364; add.ftz.f32 %f855, %f3365, 0fC0000000; setp.eq.ftz.f32 %p348, %f854, %f855; @%p348 bra $L__BB105_372; bra.uni $L__BB105_49; $L__BB105_372: lg2.approx.ftz.f32 %f3417, %f852; mul.ftz.f32 %f3418, %f854, %f3417; ex2.approx.ftz.f32 %f3419, %f3418; add.ftz.f32 %f3420, %f854, 0f3F800000; mov.f32 %f3421, 0f41C90FDB; div.approx.ftz.f32 %f3422, %f3420, %f3421; mul.ftz.f32 %f3423, %f853, %f3422; mul.ftz.f32 %f5172, %f3419, %f3423; mul.ftz.f32 %f3424, %f3419, %f3422; div.approx.ftz.f32 %f5297, %f3424, %f851; bra.uni $L__BB105_373; $L__BB105_514: ld.local.f32 %f1318, [%rd1+-224]; ld.local.f32 %f1319, [%rd1+-220]; ld.local.f32 %f1320, [%rd1+-216]; ld.local.f32 %f1321, [%rd3+8]; ld.local.f32 %f1322, [%rd3+12]; mul.ftz.f32 %f4689, %f1405, %f1322; fma.rn.ftz.f32 %f4690, %f1404, %f1321, %f4689; ld.local.f32 %f1323, [%rd3+16]; fma.rn.ftz.f32 %f1324, %f1406, %f1323, %f4690; setp.leu.ftz.f32 %p495, %f1324, 0f00000000; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p495 bra $L__BB105_522; mov.f32 %f5294, 0f00000000; max.ftz.f32 %f4694, %f1324, %f5294; mul.ftz.f32 %f5297, %f4694, 0f3EA2F983; setp.le.ftz.f32 %p496, %f1324, 0f00000000; @%p496 bra $L__BB105_522; mul.ftz.f32 %f4695, %f1319, %f1322; fma.rn.ftz.f32 %f4696, %f1318, %f1321, %f4695; fma.rn.ftz.f32 %f4697, %f1320, %f1323, %f4696; mov.f32 %f4698, 0f3F800000; sub.ftz.f32 %f4699, %f4698, %f4697; mov.f32 %f4700, 0f00000000; max.ftz.f32 %f4701, %f4699, %f4700; min.ftz.f32 %f4702, %f4701, %f4698; mul.ftz.f32 %f4703, %f4702, %f4702; mul.ftz.f32 %f4704, %f4703, %f4703; mul.ftz.f32 %f1326, %f4702, %f4704; sub.ftz.f32 %f4705, %f4698, %f1324; max.ftz.f32 %f4706, %f4705, %f4700; min.ftz.f32 %f4707, %f4706, %f4698; mul.ftz.f32 %f4708, %f4707, %f4707; mul.ftz.f32 %f4709, %f4708, %f4708; mul.ftz.f32 %f1327, %f4707, %f4709; ld.local.u32 %r30, [%rd3+24]; and.b32 %r152, %r30, 3; setp.eq.s32 %p497, %r152, 0; fma.rn.ftz.f32 %f1328, %f1327, 0fBF000000, 0f3F800000; @%p497 bra $L__BB105_518; fma.rn.ftz.f32 %f4710, %f1326, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f5288, %f4710, %f1328, 0f00000000; bra.uni $L__BB105_519; $L__BB105_510: ld.local.f32 %f1300, [%rd3+8]; ld.local.f32 %f1301, [%rd3+12]; mul.ftz.f32 %f4654, %f1405, %f1301; fma.rn.ftz.f32 %f4655, %f1404, %f1300, %f4654; ld.local.f32 %f1302, [%rd3+16]; fma.rn.ftz.f32 %f1303, %f1406, %f1302, %f4655; setp.leu.ftz.f32 %p491, %f1303, 0f00000000; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p491 bra $L__BB105_513; ld.local.f32 %f4660, [%rd1+-224]; ld.local.f32 %f4661, [%rd1+-220]; ld.local.f32 %f4662, [%rd1+-216]; add.ftz.f32 %f4663, %f1404, %f4660; add.ftz.f32 %f4664, %f1405, %f4661; mul.ftz.f32 %f4665, %f4664, %f4664; fma.rn.ftz.f32 %f4666, %f4663, %f4663, %f4665; add.ftz.f32 %f4667, %f1406, %f4662; fma.rn.ftz.f32 %f4668, %f4667, %f4667, %f4666; sqrt.approx.ftz.f32 %f4669, %f4668; div.approx.ftz.f32 %f1304, %f4663, %f4669; div.approx.ftz.f32 %f1305, %f4664, %f4669; div.approx.ftz.f32 %f1306, %f4667, %f4669; mul.ftz.f32 %f4670, %f4661, %f1301; fma.rn.ftz.f32 %f4671, %f4660, %f1300, %f4670; fma.rn.ftz.f32 %f4672, %f4662, %f1302, %f4671; setp.lt.ftz.f32 %p492, %f4672, 0f00000000; setp.lt.ftz.f32 %p493, %f1303, 0f00000000; or.pred %p494, %p493, %p492; mov.f32 %f5294, %f5297; @%p494 bra $L__BB105_513; mov.f32 %f4673, 0f00000000; max.ftz.f32 %f4674, %f1303, %f4673; mul.ftz.f32 %f5297, %f4674, 0f3EA2F983; mul.ftz.f32 %f4675, %f1405, %f1305; fma.rn.ftz.f32 %f4676, %f1404, %f1304, %f4675; fma.rn.ftz.f32 %f4677, %f1406, %f1306, %f4676; mov.f32 %f4678, 0f3F800000; sub.ftz.f32 %f4679, %f4678, %f4677; max.ftz.f32 %f4680, %f4679, %f4673; min.ftz.f32 %f4681, %f4680, %f4678; mul.ftz.f32 %f4682, %f4681, %f4681; mul.ftz.f32 %f4683, %f4682, %f4682; mul.ftz.f32 %f4684, %f4681, %f4683; mul.ftz.f32 %f5294, %f1303, %f4684; $L__BB105_513: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_471: setp.leu.ftz.f32 %p448, %f1184, 0f00000000; setp.leu.ftz.f32 %p449, %f1183, 0f00000000; or.pred %p450, %p449, %p448; mov.f32 %f5296, %f5297; mov.f32 %f5295, %f5297; mov.f32 %f5294, %f5297; @%p450 bra $L__BB105_524; add.ftz.f32 %f4151, %f1404, %f1173; add.ftz.f32 %f4152, %f1405, %f1174; mul.ftz.f32 %f4153, %f4152, %f4152; fma.rn.ftz.f32 %f4154, %f4151, %f4151, %f4153; add.ftz.f32 %f4155, %f1406, %f1175; fma.rn.ftz.f32 %f4156, %f4155, %f4155, %f4154; sqrt.approx.ftz.f32 %f4157, %f4156; div.approx.ftz.f32 %f1185, %f4151, %f4157; div.approx.ftz.f32 %f1186, %f4152, %f4157; div.approx.ftz.f32 %f1187, %f4155, %f4157; mul.ftz.f32 %f4158, %f1181, %f1186; fma.rn.ftz.f32 %f4159, %f1180, %f1185, %f4158; fma.rn.ftz.f32 %f1188, %f1182, %f1187, %f4159; mul.ftz.f32 %f1189, %f1188, %f1188; setp.eq.ftz.f32 %p451, %f4136, %f4137; @%p451 bra $L__BB105_474; bra.uni $L__BB105_473; $L__BB105_474: setp.eq.s32 %p452, %r5, 12; @%p452 bra $L__BB105_476; bra.uni $L__BB105_475; $L__BB105_476: setp.ge.ftz.f32 %p453, %f4136, 0f3F800000; mov.f32 %f5262, 0f3EA2F983; mov.f32 %f5261, 0f3D800000; @%p453 bra $L__BB105_478; mul.ftz.f32 %f4257, %f4136, %f4136; add.ftz.f32 %f4258, %f4257, 0fBF800000; mul.ftz.f32 %f4259, %f1188, %f4258; fma.rn.ftz.f32 %f4260, %f1188, %f4259, 0f3F800000; lg2.approx.ftz.f32 %f4261, %f4257; mul.ftz.f32 %f4262, %f4261, 0f3F317218; mul.ftz.f32 %f4263, %f4262, 0f40490FDB; mul.ftz.f32 %f4264, %f4260, %f4263; div.approx.ftz.f32 %f5262, %f4258, %f4264; bra.uni $L__BB105_478; $L__BB105_368: ld.local.f32 %f818, [%rd3+8]; ld.local.f32 %f819, [%rd1+-224]; ld.local.f32 %f820, [%rd3+12]; ld.local.f32 %f821, [%rd1+-220]; mul.ftz.f32 %f3271, %f821, %f820; fma.rn.ftz.f32 %f3272, %f819, %f818, %f3271; ld.local.f32 %f822, [%rd3+16]; ld.local.f32 %f823, [%rd1+-216]; fma.rn.ftz.f32 %f824, %f823, %f822, %f3272; mul.ftz.f32 %f3273, %f1405, %f820; fma.rn.ftz.f32 %f3274, %f1404, %f818, %f3273; fma.rn.ftz.f32 %f825, %f1406, %f822, %f3274; setp.leu.ftz.f32 %p337, %f824, 0f00000000; setp.leu.ftz.f32 %p338, %f825, 0f00000000; or.pred %p339, %p338, %p337; mov.f32 %f5297, 0f00000000; mov.f32 %f5294, %f5297; @%p339 bra $L__BB105_371; add.ftz.f32 %f3279, %f1404, %f819; add.ftz.f32 %f3280, %f1405, %f821; mul.ftz.f32 %f3281, %f3280, %f3280; fma.rn.ftz.f32 %f3282, %f3279, %f3279, %f3281; add.ftz.f32 %f3283, %f1406, %f823; fma.rn.ftz.f32 %f3284, %f3283, %f3283, %f3282; sqrt.approx.ftz.f32 %f3285, %f3284; div.approx.ftz.f32 %f3286, %f3279, %f3285; div.approx.ftz.f32 %f3287, %f3280, %f3285; div.approx.ftz.f32 %f3288, %f3283, %f3285; mul.ftz.f32 %f3289, %f820, %f3287; fma.rn.ftz.f32 %f3290, %f818, %f3286, %f3289; fma.rn.ftz.f32 %f826, %f822, %f3288, %f3290; mul.ftz.f32 %f3291, %f821, %f3287; fma.rn.ftz.f32 %f3292, %f819, %f3286, %f3291; fma.rn.ftz.f32 %f3293, %f823, %f3288, %f3292; abs.ftz.f32 %f827, %f3293; abs.ftz.f32 %f3294, %f826; setp.geu.ftz.f32 %p340, %f3294, 0f3F7FFF58; setp.leu.ftz.f32 %p341, %f827, 0f3727C5AC; or.pred %p342, %p341, %p340; mov.f32 %f5294, %f5297; @%p342 bra $L__BB105_371; ld.local.f32 %f3296, [%rd3+24]; div.approx.ftz.f32 %f3297, %f826, %f827; mov.f32 %f3298, 0f3727C5AC; max.ftz.f32 %f3299, %f3297, %f3298; mul.ftz.f32 %f3300, %f824, %f3299; abs.ftz.f32 %f3301, %f3300; add.ftz.f32 %f3302, %f3301, %f3301; mul.ftz.f32 %f3303, %f825, %f3299; abs.ftz.f32 %f3304, %f3303; add.ftz.f32 %f3305, %f3304, %f3304; mul.ftz.f32 %f3306, %f826, %f826; mov.f32 %f3307, 0f3F800000; sub.ftz.f32 %f3308, %f3307, %f3306; mul.ftz.f32 %f3309, %f3308, %f3308; div.approx.ftz.f32 %f3310, %f3306, %f3308; mul.ftz.f32 %f3311, %f3296, %f3310; mul.ftz.f32 %f3312, %f3311, 0fBFB8AA3B; ex2.approx.ftz.f32 %f3313, %f3312; mul.ftz.f32 %f3314, %f3296, %f3313; mul.ftz.f32 %f3315, %f3314, 0f3EA2F983; div.approx.ftz.f32 %f3316, %f3315, %f3309; min.ftz.f32 %f3317, %f3302, %f3305; min.ftz.f32 %f3318, %f3307, %f3317; mul.ftz.f32 %f3319, %f3316, %f3318; mul.ftz.f32 %f3320, %f3319, 0f3E800000; div.approx.ftz.f32 %f5294, %f3320, %f824; mov.f32 %f5297, 0f3E22F983; $L__BB105_371: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_362: ld.local.f32 %f3211, [%rd3+20]; mul.ftz.f32 %f794, %f3211, 0f3FC90FDB; ld.local.f32 %f3212, [%rd3+24]; mul.ftz.f32 %f795, %f3212, 0f3FC90FDB; ld.local.f32 %f796, [%rd3+8]; ld.local.f32 %f797, [%rd3+12]; mul.ftz.f32 %f3213, %f1405, %f797; fma.rn.ftz.f32 %f3214, %f1404, %f796, %f3213; ld.local.f32 %f798, [%rd3+16]; fma.rn.ftz.f32 %f3215, %f1406, %f798, %f3214; ld.local.f32 %f799, [%rd1+-224]; ld.local.f32 %f800, [%rd1+-220]; mul.ftz.f32 %f3216, %f800, %f797; fma.rn.ftz.f32 %f3217, %f799, %f796, %f3216; ld.local.f32 %f801, [%rd1+-216]; fma.rn.ftz.f32 %f802, %f801, %f798, %f3217; setp.leu.ftz.f32 %p328, %f3215, 0f00000000; setp.leu.ftz.f32 %p329, %f802, 0f00000000; mov.f32 %f5297, 0f00000000; or.pred %p330, %p328, %p329; mov.f32 %f5294, %f5297; @%p330 bra $L__BB105_367; add.ftz.f32 %f3219, %f802, %f802; mul.ftz.f32 %f3220, %f3219, %f796; mul.ftz.f32 %f3221, %f3219, %f797; mul.ftz.f32 %f3222, %f3219, %f798; sub.ftz.f32 %f3223, %f3220, %f799; sub.ftz.f32 %f3224, %f3221, %f800; sub.ftz.f32 %f3225, %f3222, %f801; mul.ftz.f32 %f3226, %f1405, %f3224; fma.rn.ftz.f32 %f3227, %f1404, %f3223, %f3226; fma.rn.ftz.f32 %f3228, %f1406, %f3225, %f3227; mov.f32 %f3229, 0f00000000; max.ftz.f32 %f3230, %f3228, %f3229; mov.f32 %f3231, 0fBF800000; max.ftz.f32 %f3232, %f3230, %f3231; mov.f32 %f3218, 0f3F800000; min.ftz.f32 %f3233, %f3232, %f3218; abs.ftz.f32 %f3234, %f3233; sub.ftz.f32 %f3235, %f3218, %f3234; mul.ftz.f32 %f3236, %f3235, 0f3F000000; sqrt.approx.ftz.f32 %f3237, %f3236; setp.gt.ftz.f32 %p331, %f3234, 0f3F11EB85; selp.f32 %f3238, %f3237, %f3234, %p331; mul.ftz.f32 %f3239, %f3238, %f3238; mov.f32 %f3240, 0f3C94D2E9; mov.f32 %f3241, 0f3D53F941; fma.rn.ftz.f32 %f3242, %f3241, %f3239, %f3240; mov.f32 %f3243, 0f3D3F841F; fma.rn.ftz.f32 %f3244, %f3242, %f3239, %f3243; mov.f32 %f3245, 0f3D994929; fma.rn.ftz.f32 %f3246, %f3244, %f3239, %f3245; mov.f32 %f3247, 0f3E2AAB94; fma.rn.ftz.f32 %f3248, %f3246, %f3239, %f3247; mul.ftz.f32 %f3249, %f3239, %f3248; fma.rn.ftz.f32 %f3250, %f3249, %f3238, %f3238; add.ftz.f32 %f3251, %f3250, %f3250; mov.f32 %f3252, 0f3FC90FDB; sub.ftz.f32 %f3253, %f3252, %f3250; selp.f32 %f3254, %f3251, %f3253, %p331; setp.lt.ftz.f32 %p332, %f3233, 0f00000000; mov.f32 %f3255, 0f40490FDB; sub.ftz.f32 %f3256, %f3255, %f3254; selp.f32 %f803, %f3256, %f3254, %p332; setp.lt.ftz.f32 %p333, %f803, %f794; add.ftz.f32 %f804, %f794, %f795; mov.f32 %f5165, %f3218; @%p333 bra $L__BB105_366; setp.leu.ftz.f32 %p334, %f804, %f803; setp.eq.ftz.f32 %p335, %f795, 0f00000000; or.pred %p336, %p335, %p334; mov.f32 %f5165, %f3229; @%p336 bra $L__BB105_366; sub.ftz.f32 %f3258, %f803, %f794; div.approx.ftz.f32 %f3259, %f3258, %f795; mov.f32 %f3260, 0f3F800000; sub.ftz.f32 %f5165, %f3260, %f3259; $L__BB105_366: min.ftz.f32 %f3262, %f804, %f3252; cos.approx.ftz.f32 %f3263, %f3262; sub.ftz.f32 %f3265, %f3218, %f3263; mov.f32 %f3266, 0f3E22F983; div.approx.ftz.f32 %f5297, %f3266, %f3265; mul.ftz.f32 %f5294, %f5165, %f5297; $L__BB105_367: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_490: ld.local.f32 %f4396, [%rd3+8]; ld.local.f32 %f4397, [%rd3+12]; mul.ftz.f32 %f4398, %f1405, %f4397; fma.rn.ftz.f32 %f4399, %f1404, %f4396, %f4398; ld.local.f32 %f4400, [%rd3+16]; fma.rn.ftz.f32 %f4401, %f1406, %f4400, %f4399; setp.ge.ftz.f32 %p461, %f4401, 0f00000000; mov.f32 %f5294, 0f00000000; @%p461 bra $L__BB105_508; ld.local.f32 %f4406, [%rd1+-224]; ld.local.f32 %f4407, [%rd1+-220]; ld.local.f32 %f4408, [%rd1+-216]; ld.local.f32 %f1247, [%rd3+40]; ld.local.f32 %f1248, [%rd3+32]; ld.local.f32 %f1249, [%rd3+36]; ld.local.f32 %f4409, [%rd3+20]; ld.local.f32 %f4410, [%rd3+24]; mul.ftz.f32 %f4411, %f4407, %f4410; fma.rn.ftz.f32 %f4412, %f4406, %f4409, %f4411; ld.local.f32 %f4413, [%rd3+28]; fma.rn.ftz.f32 %f4414, %f4408, %f4413, %f4412; mul.ftz.f32 %f4415, %f4409, %f4414; mul.ftz.f32 %f4416, %f4410, %f4414; mul.ftz.f32 %f4417, %f4413, %f4414; sub.ftz.f32 %f4418, %f4406, %f4415; sub.ftz.f32 %f4419, %f4407, %f4416; sub.ftz.f32 %f4420, %f4408, %f4417; mul.ftz.f32 %f4421, %f4419, %f4419; fma.rn.ftz.f32 %f4422, %f4418, %f4418, %f4421; fma.rn.ftz.f32 %f4423, %f4420, %f4420, %f4422; sqrt.approx.ftz.f32 %f4424, %f4423; div.approx.ftz.f32 %f4425, %f4418, %f4424; div.approx.ftz.f32 %f4426, %f4419, %f4424; div.approx.ftz.f32 %f4427, %f4420, %f4424; abs.ftz.f32 %f4428, %f4414; setp.lt.ftz.f32 %p462, %f4428, 0f3F800000; mov.f32 %f4429, 0f3F800000; add.ftz.f32 %f4430, %f4428, 0fBF800000; add.ftz.f32 %f4431, %f4430, 0f3F800000; selp.f32 %f4432, %f4431, 0f3F800000, %p462; sub.ftz.f32 %f4433, %f4429, %f4432; sqrt.approx.ftz.f32 %f4434, %f4433; mul.ftz.f32 %f4435, %f4432, 0f3CB1485D; mov.f32 %f4436, 0f3D9FB43B; sub.ftz.f32 %f4437, %f4436, %f4435; fma.rn.ftz.f32 %f4438, %f4432, %f4437, 0fBE5A6B93; fma.rn.ftz.f32 %f4439, %f4432, %f4438, 0f3FC90FDB; mov.f32 %f4440, 0f3FC90FDB; mul.ftz.f32 %f4441, %f4434, %f4439; setp.lt.ftz.f32 %p463, %f4414, 0f00000000; mov.f32 %f4442, 0f40490FDB; sub.ftz.f32 %f4443, %f4442, %f4441; selp.f32 %f4444, %f4443, %f4441, %p463; sub.ftz.f32 %f1250, %f4440, %f4444; mul.ftz.f32 %f4445, %f1405, %f4410; fma.rn.ftz.f32 %f4446, %f1404, %f4409, %f4445; fma.rn.ftz.f32 %f4447, %f1406, %f4413, %f4446; mul.ftz.f32 %f4448, %f4409, %f4447; mul.ftz.f32 %f4449, %f4410, %f4447; mul.ftz.f32 %f4450, %f4413, %f4447; sub.ftz.f32 %f4451, %f1404, %f4448; sub.ftz.f32 %f4452, %f1405, %f4449; sub.ftz.f32 %f4453, %f1406, %f4450; mul.ftz.f32 %f4454, %f4452, %f4452; fma.rn.ftz.f32 %f4455, %f4451, %f4451, %f4454; fma.rn.ftz.f32 %f4456, %f4453, %f4453, %f4455; sqrt.approx.ftz.f32 %f4457, %f4456; div.approx.ftz.f32 %f4458, %f4451, %f4457; div.approx.ftz.f32 %f4459, %f4452, %f4457; div.approx.ftz.f32 %f4460, %f4453, %f4457; abs.ftz.f32 %f4461, %f4447; setp.lt.ftz.f32 %p464, %f4461, 0f3F800000; add.ftz.f32 %f4462, %f4461, 0fBF800000; add.ftz.f32 %f4463, %f4462, 0f3F800000; selp.f32 %f4464, %f4463, 0f3F800000, %p464; sub.ftz.f32 %f4465, %f4429, %f4464; sqrt.approx.ftz.f32 %f4466, %f4465; mul.ftz.f32 %f4467, %f4464, 0f3CB1485D; sub.ftz.f32 %f4468, %f4436, %f4467; fma.rn.ftz.f32 %f4469, %f4464, %f4468, 0fBE5A6B93; fma.rn.ftz.f32 %f4470, %f4464, %f4469, 0f3FC90FDB; mul.ftz.f32 %f4471, %f4466, %f4470; setp.lt.ftz.f32 %p465, %f4447, 0f00000000; sub.ftz.f32 %f4472, %f4442, %f4471; selp.f32 %f4473, %f4472, %f4471, %p465; sub.ftz.f32 %f1251, %f4440, %f4473; mul.ftz.f32 %f4474, %f4426, %f4459; fma.rn.ftz.f32 %f4475, %f4425, %f4458, %f4474; fma.rn.ftz.f32 %f1252, %f4427, %f4460, %f4475; abs.ftz.f32 %f4476, %f1251; sub.ftz.f32 %f4477, %f4440, %f4476; setp.lt.ftz.f32 %p466, %f4477, 0f3A83126F; @%p466 bra $L__BB105_508; abs.ftz.f32 %f4479, %f1252; setp.lt.ftz.f32 %p467, %f4479, 0f3F800000; add.ftz.f32 %f4481, %f4479, 0fBF800000; add.ftz.f32 %f4482, %f4481, 0f3F800000; selp.f32 %f4483, %f4482, 0f3F800000, %p467; sub.ftz.f32 %f4484, %f4429, %f4483; sqrt.approx.ftz.f32 %f4485, %f4484; fma.rn.ftz.f32 %f4486, %f4483, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f4487, %f4483, %f4486, 0fBE5A6B93; fma.rn.ftz.f32 %f4488, %f4483, %f4487, 0f3FC90FDB; mul.ftz.f32 %f4489, %f4485, %f4488; sub.ftz.f32 %f4491, %f4442, %f4489; setp.lt.ftz.f32 %p468, %f1252, 0f00000000; mov.f32 %f5280, 0f00000000; selp.f32 %f1253, %f4491, %f4489, %p468; mul.ftz.f32 %f4492, %f1251, 0f3EA2F983; mov.b32 %r134, %f4492; and.b32 %r135, %r134, -2147483648; or.b32 %r136, %r135, 1056964608; mov.b32 %f4493, %r136; add.ftz.f32 %f4494, %f4492, %f4493; cvt.rzi.ftz.s32.f32 %r137, %f4494; cvt.rn.f32.s32 %f4495, %r137; fma.rn.ftz.f32 %f4496, %f4495, 0fC0490000, %f1251; fma.rn.ftz.f32 %f4497, %f4495, 0fBA7DA000, %f4496; fma.rn.ftz.f32 %f4498, %f4495, 0fB4222000, %f4497; fma.rn.ftz.f32 %f4499, %f4495, 0fACB4611A, %f4498; add.ftz.f32 %f4500, %f4499, 0fBFC90FDB; add.ftz.f32 %f4501, %f4500, 0f3FC90FDB; mul.ftz.f32 %f4502, %f4501, %f4501; fma.rn.ftz.f32 %f4503, %f4502, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f4504, %f4502, %f4503, 0fBAB60B58; fma.rn.ftz.f32 %f4505, %f4502, %f4504, 0f3D2AAAAA; fma.rn.ftz.f32 %f4506, %f4502, %f4505, 0fBF000000; fma.rn.ftz.f32 %f4507, %f4502, %f4506, 0f3F800000; and.b32 %r138, %r137, 1; setp.eq.b32 %p469, %r138, 1; neg.ftz.f32 %f4508, %f4507; selp.f32 %f1254, %f4508, %f4507, %p469; add.ftz.f32 %f4509, %f1250, 0f3FC90FDB; mov.f32 %f4510, 0f40000000; div.approx.ftz.f32 %f4511, %f4509, %f4510; sub.ftz.f32 %f4512, %f4511, %f1247; rcp.approx.ftz.f32 %f1255, %f1248; mul.ftz.f32 %f1256, %f1255, %f4512; abs.ftz.f32 %f1257, %f1256; setp.eq.ftz.f32 %p470, %f1257, 0f00000000; mov.f32 %f5278, %f5280; @%p470 bra $L__BB105_497; setp.eq.ftz.f32 %p471, %f1257, 0f3F800000; mov.f32 %f5278, %f4429; @%p471 bra $L__BB105_497; setp.gt.ftz.f32 %p472, %f1257, 0f3F800000; @%p472 bra $L__BB105_496; bra.uni $L__BB105_495; $L__BB105_496: mov.f32 %f4515, 0f3F800000; div.approx.ftz.f32 %f5278, %f4515, %f1257; bra.uni $L__BB105_497; $L__BB105_209: ld.local.u64 %rd5, [%rd3+52]; ld.v4.f32 {%f2363, %f2364, %f2365, %f2366}, [%rd5]; ld.local.f32 %f5123, [%rd1+-144]; ld.local.f32 %f5124, [%rd1+-140]; mul.ftz.f32 %f2367, %f5124, %f5124; fma.rn.ftz.f32 %f2368, %f5123, %f5123, %f2367; ld.local.f32 %f5125, [%rd1+-136]; fma.rn.ftz.f32 %f2369, %f5125, %f5125, %f2368; sqrt.approx.ftz.f32 %f435, %f2369; setp.neu.ftz.f32 %p185, %f435, 0f00000000; @%p185 bra $L__BB105_210; bra.uni $L__BB105_211; $L__BB105_210: rcp.approx.ftz.f32 %f2370, %f435; mul.ftz.f32 %f5125, %f5125, %f2370; mul.ftz.f32 %f5124, %f5124, %f2370; mul.ftz.f32 %f5123, %f5123, %f2370; $L__BB105_211: mul.ftz.f32 %f2371, %f2365, %f5124; mul.ftz.f32 %f2372, %f2364, %f5125; sub.ftz.f32 %f5126, %f2371, %f2372; mul.ftz.f32 %f2373, %f2363, %f5125; mul.ftz.f32 %f2374, %f2365, %f5123; sub.ftz.f32 %f5127, %f2373, %f2374; mul.ftz.f32 %f2375, %f2364, %f5123; mul.ftz.f32 %f2376, %f2363, %f5124; sub.ftz.f32 %f5128, %f2375, %f2376; mul.ftz.f32 %f2377, %f5127, %f5127; fma.rn.ftz.f32 %f2378, %f5126, %f5126, %f2377; fma.rn.ftz.f32 %f2379, %f5128, %f5128, %f2378; sqrt.approx.ftz.f32 %f451, %f2379; setp.neu.ftz.f32 %p186, %f451, 0f00000000; @%p186 bra $L__BB105_212; bra.uni $L__BB105_213; $L__BB105_212: rcp.approx.ftz.f32 %f2380, %f451; mul.ftz.f32 %f5128, %f2380, %f5128; mul.ftz.f32 %f5127, %f2380, %f5127; mul.ftz.f32 %f5126, %f2380, %f5126; $L__BB105_213: ld.local.f32 %f2381, [%rd1+-224]; ld.local.f32 %f2382, [%rd1+-220]; mul.ftz.f32 %f2383, %f5124, %f2382; fma.rn.ftz.f32 %f2384, %f5123, %f2381, %f2383; ld.local.f32 %f2385, [%rd1+-216]; fma.rn.ftz.f32 %f461, %f5125, %f2385, %f2384; mul.ftz.f32 %f2386, %f2364, %f2382; fma.rn.ftz.f32 %f2387, %f2363, %f2381, %f2386; fma.rn.ftz.f32 %f2388, %f2365, %f2385, %f2387; mul.ftz.f32 %f2389, %f5127, %f2382; fma.rn.ftz.f32 %f2390, %f5126, %f2381, %f2389; fma.rn.ftz.f32 %f2391, %f5128, %f2385, %f2390; mul.ftz.f32 %f2392, %f1405, %f5124; fma.rn.ftz.f32 %f2393, %f1404, %f5123, %f2392; fma.rn.ftz.f32 %f462, %f1406, %f5125, %f2393; mul.ftz.f32 %f2394, %f1405, %f2364; fma.rn.ftz.f32 %f2395, %f1404, %f2363, %f2394; fma.rn.ftz.f32 %f463, %f1406, %f2365, %f2395; mul.ftz.f32 %f2396, %f1405, %f5127; fma.rn.ftz.f32 %f2397, %f1404, %f5126, %f2396; fma.rn.ftz.f32 %f464, %f1406, %f5128, %f2397; mul.ftz.f32 %f465, %f461, %f461; mov.f32 %f5130, 0f3F800000; sub.ftz.f32 %f2399, %f5130, %f465; mov.f32 %f2400, 0f00000000; max.ftz.f32 %f2401, %f2399, %f2400; sqrt.approx.ftz.f32 %f466, %f2401; abs.ftz.f32 %f467, %f2388; abs.ftz.f32 %f468, %f2391; setp.eq.ftz.f32 %p187, %f467, 0f00000000; setp.eq.ftz.f32 %p188, %f468, 0f00000000; and.pred %p189, %p187, %p188; mov.b32 %r15, %f2388; mov.b32 %r53, %f2391; and.b32 %r16, %r53, -2147483648; @%p189 bra $L__BB105_217; bra.uni $L__BB105_214; $L__BB105_217: shr.s32 %r58, %r15, 31; and.b32 %r59, %r58, 1078530011; or.b32 %r60, %r16, %r59; mov.b32 %f5129, %r60; bra.uni $L__BB105_218; $L__BB105_509: ld.local.f32 %f4585, [%rd3+8]; ld.local.f32 %f4586, [%rd3+12]; mul.ftz.f32 %f4587, %f1405, %f4586; fma.rn.ftz.f32 %f4588, %f1404, %f4585, %f4587; ld.local.f32 %f4589, [%rd3+16]; fma.rn.ftz.f32 %f4590, %f1406, %f4589, %f4588; neg.ftz.f32 %f4591, %f4590; mov.f32 %f4592, 0f00000000; max.ftz.f32 %f4593, %f4591, %f4592; mul.ftz.f32 %f5294, %f4593, 0f3EA2F983; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; bra.uni $L__BB105_524; $L__BB105_34: ld.local.f32 %f1111, [%rd3+8]; ld.local.f32 %f1112, [%rd3+12]; mul.ftz.f32 %f3928, %f1106, %f1112; fma.rn.ftz.f32 %f3929, %f1105, %f1111, %f3928; ld.local.f32 %f1113, [%rd3+16]; fma.rn.ftz.f32 %f1114, %f1107, %f1113, %f3929; mul.ftz.f32 %f3930, %f1405, %f1112; fma.rn.ftz.f32 %f3931, %f1404, %f1111, %f3930; fma.rn.ftz.f32 %f1115, %f1406, %f1113, %f3931; setp.lt.ftz.f32 %p428, %f1115, 0f00000000; @%p428 bra $L__BB105_462; bra.uni $L__BB105_35; $L__BB105_462: setp.le.ftz.f32 %p437, %f1114, 0f00000000; setp.ge.ftz.f32 %p438, %f1115, 0f00000000; or.pred %p439, %p437, %p438; mov.f32 %f5294, %f5297; @%p439 bra $L__BB105_468; ld.local.f32 %f1151, [%rd3+28]; fma.rn.ftz.f32 %f4070, %f1404, %f1151, %f1105; fma.rn.ftz.f32 %f4071, %f1405, %f1151, %f1106; fma.rn.ftz.f32 %f4072, %f1406, %f1151, %f1107; neg.ftz.f32 %f4073, %f4070; neg.ftz.f32 %f4074, %f4071; neg.ftz.f32 %f4075, %f4072; mul.ftz.f32 %f4076, %f4071, %f4071; fma.rn.ftz.f32 %f4077, %f4070, %f4070, %f4076; fma.rn.ftz.f32 %f1152, %f4072, %f4072, %f4077; sqrt.approx.ftz.f32 %f4078, %f1152; div.approx.ftz.f32 %f4079, %f4073, %f4078; div.approx.ftz.f32 %f4080, %f4074, %f4078; div.approx.ftz.f32 %f4081, %f4075, %f4078; mul.ftz.f32 %f4082, %f1106, %f4080; fma.rn.ftz.f32 %f4083, %f1105, %f4079, %f4082; fma.rn.ftz.f32 %f1153, %f1107, %f4081, %f4083; mul.ftz.f32 %f4084, %f1405, %f4080; fma.rn.ftz.f32 %f4085, %f1404, %f4079, %f4084; fma.rn.ftz.f32 %f1154, %f1406, %f4081, %f4085; mul.ftz.f32 %f4086, %f1112, %f4080; fma.rn.ftz.f32 %f4087, %f1111, %f4079, %f4086; fma.rn.ftz.f32 %f4088, %f1113, %f4081, %f4087; mov.f32 %f5258, 0f3F800000; min.ftz.f32 %f4089, %f4088, %f5258; mul.ftz.f32 %f4090, %f4089, %f4089; sub.ftz.f32 %f4091, %f5258, %f4090; div.approx.ftz.f32 %f4092, %f4091, %f4090; mul.ftz.f32 %f4093, %f4090, %f4090; neg.ftz.f32 %f4094, %f4092; div.approx.ftz.f32 %f4095, %f4094, %f1110; mul.ftz.f32 %f4096, %f4095, 0f3FB8AA3B; ex2.approx.ftz.f32 %f4097, %f4096; mul.ftz.f32 %f4098, %f1110, 0f40490FDB; mul.ftz.f32 %f4099, %f4098, %f4093; div.approx.ftz.f32 %f1155, %f4097, %f4099; mul.ftz.f32 %f4100, %f1114, %f1114; sub.ftz.f32 %f4101, %f5258, %f4100; div.approx.ftz.f32 %f4102, %f4101, %f4100; mov.f32 %f4103, 0f00000000; max.ftz.f32 %f4104, %f4102, %f4103; sqrt.approx.ftz.f32 %f4105, %f4104; mul.ftz.f32 %f1156, %f3920, %f4105; setp.lt.ftz.f32 %p440, %f1156, 0f3F200000; mov.f32 %f5257, %f5258; @%p440 bra $L__BB105_465; rcp.approx.ftz.f32 %f4106, %f1156; fma.rn.ftz.f32 %f4107, %f4106, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4108, %f4106, %f4107; fma.rn.ftz.f32 %f4109, %f4106, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4110, %f4106, %f4109, 0f3F800000; div.approx.ftz.f32 %f5257, %f4108, %f4110; $L__BB105_465: mul.ftz.f32 %f4112, %f1115, %f1115; sub.ftz.f32 %f4113, %f5258, %f4112; div.approx.ftz.f32 %f4114, %f4113, %f4112; max.ftz.f32 %f4116, %f4114, %f4103; sqrt.approx.ftz.f32 %f4117, %f4116; mul.ftz.f32 %f1159, %f3920, %f4117; setp.lt.ftz.f32 %p441, %f1159, 0f3F200000; @%p441 bra $L__BB105_467; rcp.approx.ftz.f32 %f4118, %f1159; fma.rn.ftz.f32 %f4119, %f4118, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4120, %f4118, %f4119; fma.rn.ftz.f32 %f4121, %f4118, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4122, %f4118, %f4121, 0f3F800000; div.approx.ftz.f32 %f5258, %f4120, %f4122; $L__BB105_467: mul.ftz.f32 %f4123, %f1151, %f1151; mul.ftz.f32 %f4124, %f4123, %f1155; mul.ftz.f32 %f4125, %f1114, %f1152; div.approx.ftz.f32 %f4126, %f4124, %f4125; mul.ftz.f32 %f4127, %f1154, %f1153; abs.ftz.f32 %f4128, %f4127; mul.ftz.f32 %f4129, %f5257, %f5258; mul.ftz.f32 %f4130, %f4129, %f4128; mul.ftz.f32 %f4131, %f5257, %f4128; mul.ftz.f32 %f5297, %f4126, %f4131; mul.ftz.f32 %f5294, %f4126, %f4130; bra.uni $L__BB105_468; $L__BB105_523: ld.local.f32 %f4746, [%rd3+8]; ld.local.f32 %f4747, [%rd3+12]; mul.ftz.f32 %f4748, %f1405, %f4747; fma.rn.ftz.f32 %f4749, %f1404, %f4746, %f4748; ld.local.f32 %f4750, [%rd3+16]; fma.rn.ftz.f32 %f4751, %f1406, %f4750, %f4749; mov.f32 %f4752, 0f00000000; max.ftz.f32 %f4753, %f4751, %f4752; mul.ftz.f32 %f5294, %f4753, 0f3EA2F983; mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; bra.uni $L__BB105_524; $L__BB105_75: mul.ftz.f32 %f1523, %f17, %f23; fma.rn.ftz.f32 %f1524, %f16, %f22, %f1523; fma.rn.ftz.f32 %f1525, %f18, %f24, %f1524; mul.ftz.f32 %f1526, %f35, %f1525; mul.ftz.f32 %f1527, %f1526, %f34; mul.ftz.f32 %f1528, %f20, %f1527; mul.ftz.f32 %f1529, %f20, %f1528; mul.ftz.f32 %f1530, %f21, %f42; mul.ftz.f32 %f1531, %f21, %f1530; max.ftz.f32 %f1533, %f1531, %f1507; div.approx.ftz.f32 %f43, %f1529, %f1533; setp.geu.ftz.f32 %p60, %f20, 0f3F800000; mov.f32 %f5010, %f20; @%p60 bra $L__BB105_77; rcp.approx.ftz.f32 %f5010, %f20; $L__BB105_77: mov.f32 %f1534, 0f3F800000; max.ftz.f32 %f1535, %f5010, %f1534; mov.f32 %f1536, 0f40400000; min.ftz.f32 %f1537, %f1535, %f1536; add.ftz.f32 %f1538, %f1537, 0fBF2814F9; mul.ftz.f32 %f1539, %f1538, 0fBF7A7547; mul.ftz.f32 %f1540, %f1538, %f1539; mul.ftz.f32 %f1541, %f1540, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1542, %f1541; mul.ftz.f32 %f1543, %f1537, 0fBD0A43BB; fma.rn.ftz.f32 %f1544, %f1542, 0f3D435661, %f1543; add.ftz.f32 %f1545, %f1544, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f1546, %f1423; fma.rn.ftz.f32 %f1547, %f1546, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f1548, %f1546, %f1547, 0f3D4D5C09; fma.rn.ftz.f32 %f1549, %f1546, %f1548, 0fBB86D80B; fma.rn.ftz.f32 %f1550, %f1546, %f1549, 0f3F80029F; fma.rn.ftz.f32 %f1551, %f1537, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f1552, %f1537, %f1551, 0fC2FFCFDF; fma.rn.ftz.f32 %f1553, %f1537, %f1552, 0f43927AA0; fma.rn.ftz.f32 %f1554, %f1537, %f1553, 0fC38FF917; mul.ftz.f32 %f1555, %f1537, %f1537; mov.f32 %f1556, 0f4347CD91; div.approx.ftz.f32 %f1557, %f1556, %f1555; add.ftz.f32 %f1558, %f1557, %f1554; mul.ftz.f32 %f1559, %f1537, %f1555; mov.f32 %f1560, 0f42CB5604; div.approx.ftz.f32 %f1561, %f1560, %f1559; sub.ftz.f32 %f1562, %f1558, %f1561; fma.rn.ftz.f32 %f1563, %f1546, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f1564, %f1546, %f1563, 0f41B5F319; fma.rn.ftz.f32 %f1565, %f1546, %f1564, 0fC05A14A5; fma.rn.ftz.f32 %f1566, %f1546, %f1565, 0f3DC9FFD6; fma.rn.ftz.f32 %f1567, %f1546, %f1566, 0f3BA1B61E; mul.ftz.f32 %f1568, %f1567, %f1562; fma.rn.ftz.f32 %f1569, %f1568, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f1570, %f1534, %f1545; sub.ftz.f32 %f1571, %f1534, %f1550; mul.ftz.f32 %f1572, %f1570, %f1571; fma.rn.ftz.f32 %f1573, %f1572, 0fC11520C5, %f1569; cvt.ftz.sat.f32.f32 %f1574, %f1573; abs.ftz.f32 %f1575, %f43; sub.ftz.f32 %f1576, %f1534, %f1574; mul.ftz.f32 %f1577, %f41, %f1576; fma.rn.ftz.f32 %f1578, %f1575, %f1574, %f1577; sub.ftz.f32 %f1579, %f1534, %f5009; mul.ftz.f32 %f5297, %f1579, %f1578; $L__BB105_79: ld.local.u64 %rd4, [%rd3+36]; ld.f32 %f49, [%rd4]; ld.f32 %f50, [%rd4+4]; ld.f32 %f51, [%rd4+8]; setp.lt.ftz.f32 %p61, %f18, 0f00000000; @%p61 bra $L__BB105_81; bra.uni $L__BB105_80; $L__BB105_81: mov.f32 %f5012, %f49; mov.f32 %f5013, %f50; mov.f32 %f5014, %f51; bra.uni $L__BB105_82; $L__BB105_214: setp.eq.ftz.f32 %p190, %f467, 0f7F800000; setp.eq.ftz.f32 %p191, %f468, 0f7F800000; and.pred %p192, %p190, %p191; @%p192 bra $L__BB105_216; bra.uni $L__BB105_215; $L__BB105_216: setp.lt.s32 %p196, %r15, 0; selp.b32 %r56, 1075235812, 1061752795, %p196; or.b32 %r57, %r16, %r56; mov.b32 %f5129, %r57; bra.uni $L__BB105_218; $L__BB105_35: setp.leu.ftz.f32 %p429, %f1114, 0f00000000; setp.leu.ftz.f32 %p430, %f1115, 0f00000000; or.pred %p431, %p430, %p429; mov.f32 %f5294, %f5297; @%p431 bra $L__BB105_468; add.ftz.f32 %f3936, %f1404, %f1105; add.ftz.f32 %f3937, %f1405, %f1106; mul.ftz.f32 %f3938, %f3937, %f3937; fma.rn.ftz.f32 %f3939, %f3936, %f3936, %f3938; add.ftz.f32 %f3940, %f1406, %f1107; fma.rn.ftz.f32 %f3941, %f3940, %f3940, %f3939; sqrt.approx.ftz.f32 %f3942, %f3941; div.approx.ftz.f32 %f1116, %f3936, %f3942; div.approx.ftz.f32 %f1117, %f3937, %f3942; div.approx.ftz.f32 %f1118, %f3940, %f3942; mul.ftz.f32 %f3943, %f1112, %f1117; fma.rn.ftz.f32 %f3944, %f1111, %f1116, %f3943; fma.rn.ftz.f32 %f1119, %f1113, %f1118, %f3944; mul.ftz.f32 %f1120, %f1119, %f1119; mul.ftz.f32 %f3945, %f1120, %f1120; mul.ftz.f32 %f3946, %f1110, 0f40490FDB; mul.ftz.f32 %f1121, %f3946, %f3945; mul.ftz.f32 %f1122, %f1114, %f1114; mov.f32 %f5256, 0f3F800000; sub.ftz.f32 %f1123, %f5256, %f1122; setp.eq.ftz.f32 %p432, %f3920, %f3921; @%p432 bra $L__BB105_457; bra.uni $L__BB105_37; $L__BB105_457: mov.f32 %f5256, 0f3F800000; sub.ftz.f32 %f4035, %f5256, %f1120; div.approx.ftz.f32 %f4036, %f4035, %f1120; neg.ftz.f32 %f4037, %f4036; div.approx.ftz.f32 %f4038, %f4037, %f1110; mul.ftz.f32 %f4039, %f4038, 0f3FB8AA3B; ex2.approx.ftz.f32 %f4040, %f4039; div.approx.ftz.f32 %f5254, %f4040, %f1121; div.approx.ftz.f32 %f4041, %f1123, %f1122; mov.f32 %f4042, 0f00000000; max.ftz.f32 %f4043, %f4041, %f4042; sqrt.approx.ftz.f32 %f4044, %f4043; mul.ftz.f32 %f1139, %f3920, %f4044; setp.lt.ftz.f32 %p435, %f1139, 0f3F200000; mov.f32 %f5255, %f5256; @%p435 bra $L__BB105_459; rcp.approx.ftz.f32 %f4045, %f1139; fma.rn.ftz.f32 %f4046, %f4045, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4047, %f4045, %f4046; fma.rn.ftz.f32 %f4048, %f4045, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4049, %f4045, %f4048, 0f3F800000; div.approx.ftz.f32 %f5255, %f4047, %f4049; $L__BB105_459: mul.ftz.f32 %f4051, %f1115, %f1115; sub.ftz.f32 %f4052, %f5256, %f4051; div.approx.ftz.f32 %f4053, %f4052, %f4051; max.ftz.f32 %f4055, %f4053, %f4042; sqrt.approx.ftz.f32 %f4056, %f4055; mul.ftz.f32 %f1142, %f3920, %f4056; setp.lt.ftz.f32 %p436, %f1142, 0f3F200000; @%p436 bra $L__BB105_461; rcp.approx.ftz.f32 %f4057, %f1142; fma.rn.ftz.f32 %f4058, %f4057, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4059, %f4057, %f4058; fma.rn.ftz.f32 %f4060, %f4057, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4061, %f4057, %f4060, 0f3F800000; div.approx.ftz.f32 %f5256, %f4059, %f4061; bra.uni $L__BB105_461; $L__BB105_80: ld.f32 %f5014, [%rd4+20]; ld.f32 %f5013, [%rd4+16]; ld.f32 %f5012, [%rd4+12]; $L__BB105_82: mul.ftz.f32 %f4953, %f5020, %f18; setp.lt.ftz.f32 %p62, %f4953, 0f00000000; @%p62 bra $L__BB105_85; bra.uni $L__BB105_83; $L__BB105_85: neg.ftz.f32 %f61, %f18; setp.leu.ftz.f32 %p64, %f5020, %f61; mov.f32 %f5015, %f16; mov.f32 %f5016, %f17; mov.f32 %f5017, %f18; mov.u16 %rs37, %rs12; @%p64 bra $L__BB105_87; neg.ftz.f32 %f5017, %f5020; neg.ftz.f32 %f5016, %f5019; neg.ftz.f32 %f5015, %f5018; neg.ftz.f32 %f5018, %f16; neg.ftz.f32 %f5019, %f17; mov.f32 %f5020, %f61; mov.u16 %rs37, %rs27; bra.uni $L__BB105_87; $L__BB105_83: setp.geu.ftz.f32 %p63, %f18, %f5020; mov.f32 %f5015, %f16; mov.f32 %f5016, %f17; mov.f32 %f5017, %f18; mov.u16 %rs37, %rs12; @%p63 bra $L__BB105_87; mov.f32 %f5015, %f5018; mov.f32 %f5016, %f5019; mov.f32 %f5017, %f5020; mov.f32 %f5018, %f16; mov.f32 %f5019, %f17; mov.f32 %f5020, %f18; mov.u16 %rs37, %rs27; $L__BB105_87: setp.lt.ftz.f32 %p65, %f5017, 0f3727C5AC; and.pred %p66, %p65, %p1; setp.lt.ftz.f32 %p67, %f5020, 0f3727C5AC; or.pred %p68, %p66, %p67; setp.gt.ftz.f32 %p69, %f5017, 0fB727C5AC; and.pred %p71, %p61, %p69; or.pred %p72, %p71, %p68; mov.f32 %f5296, 0f00000000; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p72 bra $L__BB105_524; neg.ftz.f32 %f70, %f5018; neg.ftz.f32 %f71, %f5019; neg.ftz.f32 %f72, %f5020; setp.lt.ftz.f32 %p73, %f5020, 0fBF7FF972; mov.f32 %f5112, 0f00000000; @%p73 bra $L__BB105_91; setp.gt.ftz.f32 %p74, %f5020, 0f3F7FF972; mov.f32 %f5112, 0fBF7FF972; @%p74 bra $L__BB105_91; mul.ftz.f32 %f1607, %f5020, %f5020; max.ftz.f32 %f1609, %f1607, %f1507; rcp.approx.ftz.f32 %f1610, %f1609; mul.ftz.f32 %f1611, %f1424, %f71; mul.ftz.f32 %f1612, %f1423, %f70; mul.ftz.f32 %f1613, %f1611, %f1611; fma.rn.ftz.f32 %f1614, %f1612, %f1612, %f1613; fma.rn.ftz.f32 %f1615, %f1614, %f1610, 0f3F800000; sqrt.approx.ftz.f32 %f1616, %f1615; neg.ftz.f32 %f1617, %f1616; setp.ge.ftz.f32 %p75, %f5020, 0f80000000; selp.f32 %f1618, %f1617, %f1616, %p75; add.ftz.f32 %f1619, %f1618, 0fBF800000; mul.ftz.f32 %f5112, %f1619, 0f3F000000; $L__BB105_91: mov.f32 %f5022, %f5015; mov.f32 %f5023, %f5016; mov.f32 %f5024, %f5017; @%p61 bra $L__BB105_92; bra.uni $L__BB105_93; $L__BB105_92: neg.ftz.f32 %f5024, %f5017; neg.ftz.f32 %f5023, %f5016; neg.ftz.f32 %f5022, %f5015; $L__BB105_93: setp.gt.ftz.f32 %p77, %f5024, 0f3F7FF972; mov.f32 %f5029, 0f00000000; mov.f32 %f5025, %f5029; @%p77 bra $L__BB105_96; setp.lt.ftz.f32 %p78, %f5024, 0fBF7FF972; mov.f32 %f5025, 0fBF7FF972; @%p78 bra $L__BB105_96; mul.ftz.f32 %f1622, %f5024, %f5024; max.ftz.f32 %f1624, %f1622, %f1507; rcp.approx.ftz.f32 %f1625, %f1624; mul.ftz.f32 %f1626, %f1424, %f5023; mul.ftz.f32 %f1627, %f1423, %f5022; mul.ftz.f32 %f1628, %f1626, %f1626; fma.rn.ftz.f32 %f1629, %f1627, %f1627, %f1628; fma.rn.ftz.f32 %f1630, %f1629, %f1625, 0f3F800000; sqrt.approx.ftz.f32 %f1631, %f1630; neg.ftz.f32 %f1632, %f1631; setp.le.ftz.f32 %p79, %f5024, 0f00000000; selp.f32 %f1633, %f1632, %f1631, %p79; add.ftz.f32 %f1634, %f1633, 0fBF800000; mul.ftz.f32 %f5025, %f1634, 0f3F000000; $L__BB105_96: add.ftz.f32 %f1638, %f5015, %f5018; add.ftz.f32 %f1639, %f5016, %f5019; mul.ftz.f32 %f1640, %f1639, %f1639; fma.rn.ftz.f32 %f1641, %f1638, %f1638, %f1640; add.ftz.f32 %f1642, %f5017, %f5020; fma.rn.ftz.f32 %f1643, %f1642, %f1642, %f1641; sqrt.approx.ftz.f32 %f1644, %f1643; div.approx.ftz.f32 %f87, %f1638, %f1644; div.approx.ftz.f32 %f88, %f1639, %f1644; div.approx.ftz.f32 %f89, %f1642, %f1644; @%p73 bra $L__BB105_106; mul.ftz.f32 %f1645, %f5112, %f72; setp.gt.ftz.f32 %p82, %f5020, 0f3F7FF972; selp.f32 %f90, 0f3F800000, %f1645, %p82; @%p61 bra $L__BB105_102; bra.uni $L__BB105_98; $L__BB105_102: fma.rn.ftz.f32 %f1680, %f20, %f5015, %f5018; fma.rn.ftz.f32 %f1681, %f20, %f5016, %f5019; fma.rn.ftz.f32 %f1682, %f20, %f5017, %f5020; mul.ftz.f32 %f1683, %f1681, %f1681; fma.rn.ftz.f32 %f1684, %f1680, %f1680, %f1683; fma.rn.ftz.f32 %f1685, %f1682, %f1682, %f1684; sqrt.approx.ftz.f32 %f1686, %f1685; div.approx.ftz.f32 %f1687, %f1680, %f1686; div.approx.ftz.f32 %f1688, %f1681, %f1686; div.approx.ftz.f32 %f1689, %f1682, %f1686; setp.lt.ftz.f32 %p85, %f1689, 0f00000000; neg.ftz.f32 %f1690, %f1687; neg.ftz.f32 %f1691, %f1688; neg.ftz.f32 %f1692, %f1689; selp.f32 %f1693, %f1690, %f1687, %p85; selp.f32 %f1694, %f1691, %f1688, %p85; selp.f32 %f97, %f1692, %f1689, %p85; mul.ftz.f32 %f1695, %f5019, %f1694; fma.rn.ftz.f32 %f1696, %f5018, %f1693, %f1695; fma.rn.ftz.f32 %f98, %f5020, %f97, %f1696; mul.ftz.f32 %f1697, %f5016, %f1694; fma.rn.ftz.f32 %f1698, %f5015, %f1693, %f1697; fma.rn.ftz.f32 %f99, %f5017, %f97, %f1698; setp.lt.ftz.f32 %p86, %f98, 0f00000000; @%p86 bra $L__BB105_106; fma.rn.ftz.f32 %f100, %f20, %f99, %f98; abs.ftz.f32 %f101, %f98; fma.rn.ftz.f32 %f102, %f101, %f101, %f37; setp.leu.ftz.f32 %p87, %f102, 0f00000000; mov.f32 %f1699, 0f3F800000; mov.f32 %f5027, %f1699; @%p87 bra $L__BB105_105; sqrt.approx.ftz.f32 %f1700, %f102; sub.ftz.f32 %f1701, %f1700, %f101; add.ftz.f32 %f1702, %f101, %f1700; div.approx.ftz.f32 %f1703, %f1701, %f1702; fma.rn.ftz.f32 %f1704, %f101, %f1702, 0fBF800000; fma.rn.ftz.f32 %f1705, %f101, %f1701, 0f3F800000; div.approx.ftz.f32 %f1706, %f1704, %f1705; mul.ftz.f32 %f1707, %f1703, 0f3F000000; mul.ftz.f32 %f1708, %f1703, %f1707; fma.rn.ftz.f32 %f1709, %f1706, %f1706, 0f3F800000; mul.ftz.f32 %f5027, %f1708, %f1709; $L__BB105_105: mov.f32 %f1710, 0f00000000; max.ftz.f32 %f1711, %f1710, %f98; sub.ftz.f32 %f1713, %f1699, %f5027; mul.ftz.f32 %f1714, %f1713, %f1711; neg.ftz.f32 %f1715, %f99; max.ftz.f32 %f1716, %f1710, %f1715; mul.ftz.f32 %f1717, %f1714, %f1716; mul.ftz.f32 %f1718, %f97, %f97; sub.ftz.f32 %f1719, %f1699, %f1718; fma.rn.ftz.f32 %f1720, %f33, %f1718, %f1719; mul.ftz.f32 %f1721, %f1720, 0f40490FDB; mul.ftz.f32 %f1722, %f1720, %f1721; max.ftz.f32 %f1724, %f1722, %f1507; div.approx.ftz.f32 %f1725, %f33, %f1724; mul.ftz.f32 %f1726, %f1717, %f1725; mul.ftz.f32 %f1727, %f90, %f100; mul.ftz.f32 %f1728, %f100, %f1727; div.approx.ftz.f32 %f5029, %f1726, %f1728; bra.uni $L__BB105_106; $L__BB105_215: setp.lt.s32 %p193, %r15, 0; min.ftz.f32 %f2402, %f468, %f467; max.ftz.f32 %f2403, %f468, %f467; div.full.ftz.f32 %f2404, %f2402, %f2403; mul.rn.ftz.f32 %f2405, %f2404, %f2404; mov.f32 %f2406, 0fC0B59883; mov.f32 %f2407, 0fBF52C7EA; fma.rn.ftz.f32 %f2408, %f2405, %f2407, %f2406; mov.f32 %f2409, 0fC0D21907; fma.rn.ftz.f32 %f2410, %f2408, %f2405, %f2409; mul.ftz.f32 %f2411, %f2405, %f2410; mul.ftz.f32 %f2412, %f2404, %f2411; add.ftz.f32 %f2413, %f2405, 0f41355DC0; mov.f32 %f2414, 0f41E6BD60; fma.rn.ftz.f32 %f2415, %f2413, %f2405, %f2414; mov.f32 %f2416, 0f419D92C8; fma.rn.ftz.f32 %f2417, %f2415, %f2405, %f2416; rcp.approx.ftz.f32 %f2418, %f2417; fma.rn.ftz.f32 %f2419, %f2412, %f2418, %f2404; mov.f32 %f2420, 0f3FC90FDB; sub.ftz.f32 %f2421, %f2420, %f2419; setp.gt.ftz.f32 %p194, %f468, %f467; selp.f32 %f2422, %f2421, %f2419, %p194; mov.f32 %f2423, 0f40490FDB; sub.ftz.f32 %f2424, %f2423, %f2422; selp.f32 %f2425, %f2424, %f2422, %p193; mov.b32 %r54, %f2425; or.b32 %r55, %r16, %r54; mov.b32 %f2426, %r55; add.ftz.f32 %f2427, %f467, %f468; setp.le.ftz.f32 %p195, %f2427, 0f7F800000; selp.f32 %f5129, %f2426, %f2427, %p195; $L__BB105_218: ld.local.f32 %f2429, [%rd3+44]; div.approx.ftz.f32 %f2430, %f461, %f2429; mul.ftz.f32 %f2431, %f2430, %f2430; sub.ftz.f32 %f2432, %f5130, %f2431; max.ftz.f32 %f2434, %f2432, %f2400; sqrt.approx.ftz.f32 %f2435, %f2434; ld.f32 %f2436, [%rd5+12]; mul.ftz.f32 %f2437, %f2436, %f2436; sub.ftz.f32 %f2438, %f5130, %f2437; max.ftz.f32 %f2439, %f2438, %f2400; sqrt.approx.ftz.f32 %f2440, %f2439; mov.f32 %f2441, 0fBF800000; max.ftz.f32 %f2442, %f2436, %f2441; min.ftz.f32 %f2443, %f2442, %f5130; abs.ftz.f32 %f2444, %f2443; sub.ftz.f32 %f2445, %f5130, %f2444; mul.ftz.f32 %f2446, %f2445, 0f3F000000; sqrt.approx.ftz.f32 %f2447, %f2446; setp.gt.ftz.f32 %p197, %f2444, 0f3F11EB85; selp.f32 %f2448, %f2447, %f2444, %p197; mul.ftz.f32 %f2449, %f2448, %f2448; mov.f32 %f2450, 0f3C94D2E9; mov.f32 %f2451, 0f3D53F941; fma.rn.ftz.f32 %f2452, %f2451, %f2449, %f2450; mov.f32 %f2453, 0f3D3F841F; fma.rn.ftz.f32 %f2454, %f2452, %f2449, %f2453; mov.f32 %f2455, 0f3D994929; fma.rn.ftz.f32 %f2456, %f2454, %f2449, %f2455; mov.f32 %f2457, 0f3E2AAB94; fma.rn.ftz.f32 %f2458, %f2456, %f2449, %f2457; mul.ftz.f32 %f2459, %f2449, %f2458; fma.rn.ftz.f32 %f2460, %f2459, %f2448, %f2448; mov.f32 %f2461, 0f3FC90FDB; mov.f32 %f2462, 0fC0000000; fma.rn.ftz.f32 %f2463, %f2462, %f2460, %f2461; selp.f32 %f2464, %f2463, %f2460, %p197; setp.le.ftz.f32 %p198, %f2464, 0f7F800000; mov.b32 %r61, %f2464; mov.b32 %r62, %f2443; and.b32 %r63, %r62, -2147483648; or.b32 %r64, %r63, %r61; mov.b32 %f2465, %r64; selp.f32 %f473, %f2465, %f2464, %p198; mul.ftz.f32 %f2466, %f466, %f2436; mul.ftz.f32 %f2467, %f2429, %f2429; sub.ftz.f32 %f2468, %f2467, %f465; sqrt.approx.ftz.f32 %f2469, %f2468; div.approx.ftz.f32 %f2470, %f2466, %f2469; mul.ftz.f32 %f2471, %f2470, %f2470; sub.ftz.f32 %f2472, %f5130, %f2471; max.ftz.f32 %f2473, %f2472, %f2400; sqrt.approx.ftz.f32 %f2474, %f2473; max.ftz.f32 %f2475, %f2470, %f2441; min.ftz.f32 %f2476, %f2475, %f5130; abs.ftz.f32 %f2477, %f2476; sub.ftz.f32 %f2478, %f5130, %f2477; mul.ftz.f32 %f2479, %f2478, 0f3F000000; sqrt.approx.ftz.f32 %f2480, %f2479; setp.gt.ftz.f32 %p199, %f2477, 0f3F11EB85; selp.f32 %f2481, %f2480, %f2477, %p199; mul.ftz.f32 %f2482, %f2481, %f2481; fma.rn.ftz.f32 %f2483, %f2451, %f2482, %f2450; fma.rn.ftz.f32 %f2484, %f2483, %f2482, %f2453; fma.rn.ftz.f32 %f2485, %f2484, %f2482, %f2455; fma.rn.ftz.f32 %f2486, %f2485, %f2482, %f2457; mul.ftz.f32 %f2487, %f2482, %f2486; fma.rn.ftz.f32 %f2488, %f2487, %f2481, %f2481; fma.rn.ftz.f32 %f2489, %f2462, %f2488, %f2461; selp.f32 %f2490, %f2489, %f2488, %p199; setp.le.ftz.f32 %p200, %f2490, 0f7F800000; mov.b32 %r65, %f2490; mov.b32 %r66, %f2476; and.b32 %r67, %r66, -2147483648; or.b32 %r68, %r67, %r65; mov.b32 %f2491, %r68; selp.f32 %f474, %f2491, %f2490, %p200; add.ftz.f32 %f2492, %f2474, %f2474; div.approx.ftz.f32 %f2493, %f2492, %f2435; ld.local.f32 %f2494, [%rd3+20]; mul.ftz.f32 %f475, %f2494, %f2493; ld.local.f32 %f2495, [%rd3+24]; mul.ftz.f32 %f476, %f2495, %f2493; ld.local.f32 %f2496, [%rd3+28]; mul.ftz.f32 %f477, %f2496, %f2493; mul.ftz.f32 %f2497, %f466, %f2440; abs.ftz.f32 %f478, %f2497; add.ftz.f32 %f2498, %f2467, 0fBF800000; fma.rn.ftz.f32 %f479, %f478, %f478, %f2498; setp.leu.ftz.f32 %p201, %f479, 0f00000000; @%p201 bra $L__BB105_220; sqrt.approx.ftz.f32 %f2499, %f479; sub.ftz.f32 %f2500, %f2499, %f478; add.ftz.f32 %f2501, %f478, %f2499; div.approx.ftz.f32 %f2502, %f2500, %f2501; fma.rn.ftz.f32 %f2503, %f478, %f2501, 0fBF800000; fma.rn.ftz.f32 %f2504, %f478, %f2500, 0f3F800000; div.approx.ftz.f32 %f2505, %f2503, %f2504; mul.ftz.f32 %f2506, %f2502, 0f3F000000; mul.ftz.f32 %f2507, %f2502, %f2506; fma.rn.ftz.f32 %f2508, %f2505, %f2505, 0f3F800000; mul.ftz.f32 %f5130, %f2507, %f2508; $L__BB105_220: mov.f32 %f5131, 0f00000000; mov.f32 %f4964, 0f3F800000; mul.ftz.f32 %f2510, %f475, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2511, %f2510; mul.ftz.f32 %f2512, %f476, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f477, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2515, %f2514; sub.ftz.f32 %f2517, %f4964, %f5130; mul.ftz.f32 %f2518, %f2517, %f2517; mul.ftz.f32 %f482, %f2511, %f2518; mul.ftz.f32 %f483, %f2513, %f2518; mul.ftz.f32 %f484, %f2515, %f2518; mul.ftz.f32 %f2519, %f483, %f1412; fma.rn.ftz.f32 %f2520, %f482, %f1411, %f2519; fma.rn.ftz.f32 %f485, %f484, %f1413, %f2520; mul.ftz.f32 %f486, %f2511, %f5130; mul.ftz.f32 %f487, %f486, %f482; mul.ftz.f32 %f488, %f2513, %f5130; mul.ftz.f32 %f489, %f488, %f483; mul.ftz.f32 %f490, %f2515, %f5130; mul.ftz.f32 %f491, %f490, %f484; mul.ftz.f32 %f2521, %f489, %f1412; fma.rn.ftz.f32 %f2522, %f487, %f1411, %f2521; fma.rn.ftz.f32 %f492, %f491, %f1413, %f2522; sub.ftz.f32 %f493, %f4964, %f486; sub.ftz.f32 %f494, %f4964, %f488; sub.ftz.f32 %f495, %f4964, %f490; setp.eq.ftz.f32 %p202, %f493, 0f00000000; @%p202 bra $L__BB105_222; div.approx.ftz.f32 %f5131, %f486, %f493; $L__BB105_222: mov.f32 %f5132, 0f00000000; setp.eq.ftz.f32 %p203, %f494, 0f00000000; @%p203 bra $L__BB105_224; div.approx.ftz.f32 %f5132, %f488, %f494; $L__BB105_224: mov.f32 %f5133, 0f00000000; setp.eq.ftz.f32 %p204, %f495, 0f00000000; @%p204 bra $L__BB105_226; div.approx.ftz.f32 %f5133, %f490, %f495; $L__BB105_226: mov.f32 %f5134, 0f00000000; mul.ftz.f32 %f502, %f487, %f5131; mul.ftz.f32 %f503, %f489, %f5132; mul.ftz.f32 %f2526, %f503, %f1412; fma.rn.ftz.f32 %f2527, %f502, %f1411, %f2526; mul.ftz.f32 %f504, %f491, %f5133; fma.rn.ftz.f32 %f505, %f504, %f1413, %f2527; add.ftz.f32 %f2528, %f5130, %f485; add.ftz.f32 %f2529, %f2528, %f492; add.ftz.f32 %f506, %f505, %f2529; setp.eq.ftz.f32 %p205, %f506, 0f00000000; @%p205 bra $L__BB105_228; mov.f32 %f2530, 0f3F800000; div.approx.ftz.f32 %f5134, %f2530, %f506; $L__BB105_228: mov.f32 %f4969, 0f00000000; mul.ftz.f32 %f2531, %f462, %f462; mov.f32 %f2532, 0f3F800000; sub.ftz.f32 %f2533, %f2532, %f2531; max.ftz.f32 %f2535, %f2533, %f4969; sqrt.approx.ftz.f32 %f509, %f2535; abs.ftz.f32 %f510, %f463; setp.eq.ftz.f32 %p206, %f510, 0f00000000; abs.ftz.f32 %f511, %f464; setp.eq.ftz.f32 %p207, %f511, 0f00000000; and.pred %p208, %p206, %p207; mov.b32 %r17, %f463; mov.b32 %r69, %f464; and.b32 %r18, %r69, -2147483648; @%p208 bra $L__BB105_232; bra.uni $L__BB105_229; $L__BB105_232: shr.s32 %r74, %r17, 31; and.b32 %r75, %r74, 1078530011; or.b32 %r76, %r18, %r75; mov.b32 %f5135, %r76; bra.uni $L__BB105_233; $L__BB105_229: setp.eq.ftz.f32 %p209, %f510, 0f7F800000; setp.eq.ftz.f32 %p210, %f511, 0f7F800000; and.pred %p211, %p209, %p210; @%p211 bra $L__BB105_231; bra.uni $L__BB105_230; $L__BB105_231: setp.lt.s32 %p215, %r17, 0; selp.b32 %r72, 1075235812, 1061752795, %p215; or.b32 %r73, %r18, %r72; mov.b32 %f5135, %r73; bra.uni $L__BB105_233; $L__BB105_230: mov.f32 %f4970, 0f3FC90FDB; setp.lt.s32 %p212, %r17, 0; min.ftz.f32 %f2536, %f511, %f510; max.ftz.f32 %f2537, %f511, %f510; div.full.ftz.f32 %f2538, %f2536, %f2537; mul.rn.ftz.f32 %f2539, %f2538, %f2538; mov.f32 %f2540, 0fC0B59883; mov.f32 %f2541, 0fBF52C7EA; fma.rn.ftz.f32 %f2542, %f2539, %f2541, %f2540; mov.f32 %f2543, 0fC0D21907; fma.rn.ftz.f32 %f2544, %f2542, %f2539, %f2543; mul.ftz.f32 %f2545, %f2539, %f2544; mul.ftz.f32 %f2546, %f2538, %f2545; add.ftz.f32 %f2547, %f2539, 0f41355DC0; mov.f32 %f2548, 0f41E6BD60; fma.rn.ftz.f32 %f2549, %f2547, %f2539, %f2548; mov.f32 %f2550, 0f419D92C8; fma.rn.ftz.f32 %f2551, %f2549, %f2539, %f2550; rcp.approx.ftz.f32 %f2552, %f2551; fma.rn.ftz.f32 %f2553, %f2546, %f2552, %f2538; sub.ftz.f32 %f2555, %f4970, %f2553; setp.gt.ftz.f32 %p213, %f511, %f510; selp.f32 %f2556, %f2555, %f2553, %p213; mov.f32 %f2557, 0f40490FDB; sub.ftz.f32 %f2558, %f2557, %f2556; selp.f32 %f2559, %f2558, %f2556, %p212; mov.b32 %r70, %f2559; or.b32 %r71, %r18, %r70; mov.b32 %f2560, %r71; add.ftz.f32 %f2561, %f510, %f511; setp.le.ftz.f32 %p214, %f2561, 0f7F800000; selp.f32 %f5135, %f2560, %f2561, %p214; $L__BB105_233: sub.ftz.f32 %f516, %f5135, %f5129; ld.local.v2.f32 {%f2562, %f2563}, [%rd3+36]; sin.approx.ftz.f32 %f2565, %f2563; mul.ftz.f32 %f2566, %f2565, %f2565; sub.ftz.f32 %f2568, %f2532, %f2566; mov.f32 %f5142, 0f00000000; max.ftz.f32 %f2570, %f2568, %f5142; sqrt.approx.ftz.f32 %f2571, %f2570; add.ftz.f32 %f2572, %f2565, %f2565; mul.ftz.f32 %f2573, %f2572, %f2571; mul.ftz.f32 %f2574, %f2571, %f2571; sub.ftz.f32 %f2575, %f2574, %f2566; add.ftz.f32 %f2576, %f2573, %f2573; mul.ftz.f32 %f2577, %f2576, %f2575; mul.ftz.f32 %f2578, %f2575, %f2575; mul.ftz.f32 %f2579, %f2573, %f2573; sub.ftz.f32 %f2580, %f2578, %f2579; mul.ftz.f32 %f2581, %f462, %f2575; fma.rn.ftz.f32 %f2582, %f509, %f2573, %f2581; mul.ftz.f32 %f2583, %f509, %f2575; mul.ftz.f32 %f2584, %f462, %f2573; sub.ftz.f32 %f2585, %f2583, %f2584; abs.ftz.f32 %f2586, %f2585; mul.ftz.f32 %f2587, %f462, %f2571; mul.ftz.f32 %f2588, %f509, %f2565; sub.ftz.f32 %f518, %f2587, %f2588; mul.ftz.f32 %f2589, %f509, %f2571; fma.rn.ftz.f32 %f2590, %f462, %f2565, %f2589; abs.ftz.f32 %f519, %f2590; mul.ftz.f32 %f2591, %f462, %f2580; mul.ftz.f32 %f2592, %f509, %f2577; sub.ftz.f32 %f520, %f2591, %f2592; mul.ftz.f32 %f2593, %f509, %f2580; fma.rn.ftz.f32 %f521, %f462, %f2577, %f2593; add.ftz.f32 %f522, %f473, %f473; mov.f32 %f2594, 0f40490FDB; div.approx.ftz.f32 %f523, %f2594, %f2562; mul.ftz.f32 %f524, %f523, 0f3FB8AA3B; ld.local.f32 %f525, [%rd3+48]; rcp.approx.ftz.f32 %f526, %f525; mul.ftz.f32 %f2595, %f466, %f2586; mul.ftz.f32 %f527, %f2595, %f526; mul.ftz.f32 %f2596, %f461, %f2582; mul.ftz.f32 %f528, %f2596, %f526; setp.gtu.ftz.f32 %p216, %f525, 0f3DCCCCCD; @%p216 bra $L__BB105_247; bra.uni $L__BB105_234; $L__BB105_247: mul.ftz.f32 %f551, %f527, %f527; fma.rn.ftz.f32 %f5138, %f551, 0f3E800000, 0f3F800000; mul.ftz.f32 %f553, %f551, %f551; mov.u64 %rd23, 64; cvt.rn.f32.u64 %f2627, %rd23; div.approx.ftz.f32 %f2628, %f553, %f2627; add.ftz.f32 %f554, %f5138, %f2628; setp.eq.ftz.f32 %p226, %f5138, %f554; @%p226 bra $L__BB105_256; mul.ftz.f32 %f555, %f551, %f553; mov.u64 %rd24, 2304; cvt.rn.f32.u64 %f2629, %rd24; div.approx.ftz.f32 %f2630, %f555, %f2629; add.ftz.f32 %f556, %f554, %f2630; setp.eq.ftz.f32 %p227, %f554, %f556; mov.f32 %f5138, %f554; @%p227 bra $L__BB105_256; mul.ftz.f32 %f557, %f551, %f555; mov.u64 %rd25, 147456; cvt.rn.f32.u64 %f2631, %rd25; div.approx.ftz.f32 %f2632, %f557, %f2631; add.ftz.f32 %f558, %f556, %f2632; setp.eq.ftz.f32 %p228, %f556, %f558; mov.f32 %f5138, %f556; @%p228 bra $L__BB105_256; mul.ftz.f32 %f559, %f551, %f557; mov.u64 %rd26, 14745600; cvt.rn.f32.u64 %f2633, %rd26; div.approx.ftz.f32 %f2634, %f559, %f2633; add.ftz.f32 %f560, %f558, %f2634; setp.eq.ftz.f32 %p229, %f558, %f560; mov.f32 %f5138, %f558; @%p229 bra $L__BB105_256; mul.ftz.f32 %f561, %f551, %f559; mov.u64 %rd27, 2123366400; cvt.rn.f32.u64 %f2635, %rd27; div.approx.ftz.f32 %f2636, %f561, %f2635; add.ftz.f32 %f562, %f560, %f2636; setp.eq.ftz.f32 %p230, %f560, %f562; mov.f32 %f5138, %f560; @%p230 bra $L__BB105_256; mul.ftz.f32 %f563, %f551, %f561; mov.u64 %rd28, 416179814400; cvt.rn.f32.u64 %f2637, %rd28; div.approx.ftz.f32 %f2638, %f563, %f2637; add.ftz.f32 %f564, %f562, %f2638; setp.eq.ftz.f32 %p231, %f562, %f564; mov.f32 %f5138, %f562; @%p231 bra $L__BB105_256; mul.ftz.f32 %f565, %f551, %f563; mov.u64 %rd29, 106542032486400; cvt.rn.f32.u64 %f2639, %rd29; div.approx.ftz.f32 %f2640, %f565, %f2639; add.ftz.f32 %f566, %f564, %f2640; setp.eq.ftz.f32 %p232, %f564, %f566; mov.f32 %f5138, %f564; @%p232 bra $L__BB105_256; mul.ftz.f32 %f2641, %f551, %f565; mov.u64 %rd30, 34519618525593600; cvt.rn.f32.u64 %f2642, %rd30; div.approx.ftz.f32 %f2643, %f2641, %f2642; add.ftz.f32 %f567, %f566, %f2643; setp.eq.ftz.f32 %p233, %f566, %f567; mov.f32 %f5138, %f566; @%p233 bra $L__BB105_256; mov.f32 %f5138, %f567; $L__BB105_256: mul.ftz.f32 %f2644, %f528, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2645, %f2644; mul.ftz.f32 %f569, %f5138, %f2645; abs.ftz.f32 %f570, %f526; setp.ltu.ftz.f32 %p234, %f570, 0f3F800000; @%p234 bra $L__BB105_258; bra.uni $L__BB105_257; $L__BB105_258: mul.ftz.f32 %f2667, %f526, %f526; mov.f32 %f2668, 0f394FFF49; mov.f32 %f2669, 0f363D0ADA; fma.rn.ftz.f32 %f2670, %f2669, %f2667, %f2668; mov.f32 %f2671, 0f3C08889A; fma.rn.ftz.f32 %f2672, %f2670, %f2667, %f2671; mov.f32 %f2673, 0f3E2AAAAB; fma.rn.ftz.f32 %f2674, %f2672, %f2667, %f2673; mul.ftz.f32 %f2675, %f2667, %f2674; fma.rn.ftz.f32 %f5139, %f2675, %f526, %f526; bra.uni $L__BB105_259; $L__BB105_234: setp.gt.ftz.f32 %p217, %f527, 0f41400000; @%p217 bra $L__BB105_245; bra.uni $L__BB105_235; $L__BB105_245: mul.ftz.f32 %f2615, %f527, 0f41000000; rcp.approx.ftz.f32 %f2616, %f2615; add.ftz.f32 %f2617, %f2616, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2618, %f527; fma.rn.ftz.f32 %f2619, %f2618, 0fBF317218, %f2617; fma.rn.ftz.f32 %f5137, %f2619, 0f3F000000, %f527; bra.uni $L__BB105_246; $L__BB105_257: mov.f32 %f2646, 0f3FB8AA3B; mul.rn.ftz.f32 %f2647, %f570, %f2646; cvt.rzi.ftz.f32.f32 %f2648, %f2647; abs.ftz.f32 %f2649, %f2648; setp.gt.ftz.f32 %p235, %f2649, 0f42FC0000; mov.b32 %r77, %f2648; and.b32 %r78, %r77, -2147483648; or.b32 %r79, %r78, 1123811328; mov.b32 %f2650, %r79; selp.f32 %f2651, %f2650, %f2648, %p235; mov.f32 %f2652, 0fBF317218; fma.rn.ftz.f32 %f2653, %f2651, %f2652, %f570; mov.f32 %f2654, 0f3102E308; fma.rn.ftz.f32 %f2655, %f2651, %f2654, %f2653; mul.ftz.f32 %f2656, %f2655, 0f3FB8AA3B; add.ftz.f32 %f2657, %f2651, 0f4B40007D; mov.b32 %r80, %f2657; shl.b32 %r81, %r80, 23; mov.b32 %f2658, %r81; ex2.approx.ftz.f32 %f2659, %f2656; mul.ftz.f32 %f2660, %f2659, %f2658; mov.f32 %f2661, 0f3E000000; div.approx.ftz.f32 %f2662, %f2661, %f2660; neg.ftz.f32 %f2663, %f2662; mov.f32 %f2664, 0f40000000; fma.rn.ftz.f32 %f2665, %f2664, %f2660, %f2663; setp.ge.ftz.f32 %p236, %f570, 0f42B40000; selp.f32 %f2666, 0f7F800000, %f2665, %p236; mov.b32 %r82, %f2666; mov.b32 %r83, %f526; and.b32 %r84, %r83, -2147483648; or.b32 %r85, %r84, %r82; mov.b32 %f5139, %r85; $L__BB105_259: add.ftz.f32 %f2676, %f5139, %f5139; mul.ftz.f32 %f2677, %f525, %f2676; div.approx.ftz.f32 %f5140, %f569, %f2677; bra.uni $L__BB105_260; $L__BB105_235: mul.ftz.f32 %f529, %f527, %f527; fma.rn.ftz.f32 %f5136, %f529, 0f3E800000, 0f3F800000; mul.ftz.f32 %f531, %f529, %f529; mov.u64 %rd15, 64; cvt.rn.f32.u64 %f2597, %rd15; div.approx.ftz.f32 %f2598, %f531, %f2597; add.ftz.f32 %f532, %f5136, %f2598; setp.eq.ftz.f32 %p218, %f5136, %f532; @%p218 bra $L__BB105_244; mul.ftz.f32 %f533, %f529, %f531; mov.u64 %rd16, 2304; cvt.rn.f32.u64 %f2599, %rd16; div.approx.ftz.f32 %f2600, %f533, %f2599; add.ftz.f32 %f534, %f532, %f2600; setp.eq.ftz.f32 %p219, %f532, %f534; mov.f32 %f5136, %f532; @%p219 bra $L__BB105_244; mul.ftz.f32 %f535, %f529, %f533; mov.u64 %rd17, 147456; cvt.rn.f32.u64 %f2601, %rd17; div.approx.ftz.f32 %f2602, %f535, %f2601; add.ftz.f32 %f536, %f534, %f2602; setp.eq.ftz.f32 %p220, %f534, %f536; mov.f32 %f5136, %f534; @%p220 bra $L__BB105_244; mul.ftz.f32 %f537, %f529, %f535; mov.u64 %rd18, 14745600; cvt.rn.f32.u64 %f2603, %rd18; div.approx.ftz.f32 %f2604, %f537, %f2603; add.ftz.f32 %f538, %f536, %f2604; setp.eq.ftz.f32 %p221, %f536, %f538; mov.f32 %f5136, %f536; @%p221 bra $L__BB105_244; mul.ftz.f32 %f539, %f529, %f537; mov.u64 %rd19, 2123366400; cvt.rn.f32.u64 %f2605, %rd19; div.approx.ftz.f32 %f2606, %f539, %f2605; add.ftz.f32 %f540, %f538, %f2606; setp.eq.ftz.f32 %p222, %f538, %f540; mov.f32 %f5136, %f538; @%p222 bra $L__BB105_244; mul.ftz.f32 %f541, %f529, %f539; mov.u64 %rd20, 416179814400; cvt.rn.f32.u64 %f2607, %rd20; div.approx.ftz.f32 %f2608, %f541, %f2607; add.ftz.f32 %f542, %f540, %f2608; setp.eq.ftz.f32 %p223, %f540, %f542; mov.f32 %f5136, %f540; @%p223 bra $L__BB105_244; mul.ftz.f32 %f543, %f529, %f541; mov.u64 %rd21, 106542032486400; cvt.rn.f32.u64 %f2609, %rd21; div.approx.ftz.f32 %f2610, %f543, %f2609; add.ftz.f32 %f544, %f542, %f2610; setp.eq.ftz.f32 %p224, %f542, %f544; mov.f32 %f5136, %f542; @%p224 bra $L__BB105_244; mul.ftz.f32 %f2611, %f529, %f543; mov.u64 %rd22, 34519618525593600; cvt.rn.f32.u64 %f2612, %rd22; div.approx.ftz.f32 %f2613, %f2611, %f2612; add.ftz.f32 %f545, %f544, %f2613; setp.eq.ftz.f32 %p225, %f544, %f545; mov.f32 %f5136, %f544; @%p225 bra $L__BB105_244; mov.f32 %f5136, %f545; $L__BB105_244: lg2.approx.ftz.f32 %f2614, %f5136; mul.ftz.f32 %f5137, %f2614, 0f3F317218; $L__BB105_246: sub.ftz.f32 %f2620, %f5137, %f528; sub.ftz.f32 %f2621, %f2620, %f526; add.ftz.f32 %f2622, %f2621, 0f3F316F00; mul.ftz.f32 %f2623, %f526, 0f3F000000; lg2.approx.ftz.f32 %f2624, %f2623; fma.rn.ftz.f32 %f2625, %f2624, 0f3F317218, %f2622; mul.ftz.f32 %f2626, %f2625, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5140, %f2626; $L__BB105_260: setp.gt.ftz.f32 %p237, %f523, 0f42B00000; mul.ftz.f32 %f576, %f5134, %f485; mul.ftz.f32 %f577, %f5134, %f492; mul.ftz.f32 %f578, %f5134, %f505; mov.f32 %f5141, %f5142; @%p237 bra $L__BB105_262; ex2.approx.ftz.f32 %f2679, %f524; add.ftz.f32 %f2680, %f2679, 0f3F800000; rcp.approx.ftz.f32 %f2681, %f2680; add.ftz.f32 %f5141, %f2681, %f2681; $L__BB105_262: sub.ftz.f32 %f581, %f2532, %f5141; setp.eq.ftz.f32 %p238, %f581, 0f00000000; @%p238 bra $L__BB105_264; mul.ftz.f32 %f2684, %f474, 0f00000000; sub.ftz.f32 %f2685, %f2684, %f522; add.ftz.f32 %f2686, %f2685, 0f00000000; sub.ftz.f32 %f2687, %f516, %f2686; add.ftz.f32 %f2688, %f2687, 0f40490FDB; mov.f32 %f2689, 0f40C90FDB; div.approx.ftz.f32 %f2690, %f2688, %f2689; cvt.rmi.ftz.f32.f32 %f2691, %f2690; mul.ftz.f32 %f2692, %f2691, 0f40C90FDB; sub.ftz.f32 %f2693, %f2688, %f2692; add.ftz.f32 %f2694, %f2693, 0fC0490FDB; abs.ftz.f32 %f2695, %f2694; neg.ftz.f32 %f2696, %f2695; div.approx.ftz.f32 %f2697, %f2696, %f2562; mul.ftz.f32 %f2698, %f2697, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2699, %f2698; add.ftz.f32 %f2700, %f2699, 0f3F800000; mul.ftz.f32 %f2701, %f2700, %f2700; mul.ftz.f32 %f2702, %f2562, %f2701; div.approx.ftz.f32 %f2703, %f2699, %f2702; div.approx.ftz.f32 %f5142, %f2703, %f581; $L__BB105_264: mul.ftz.f32 %f2704, %f5140, %f5130; fma.rn.ftz.f32 %f584, %f5142, %f2704, 0f00000000; mul.ftz.f32 %f2705, %f5134, %f5130; mul.ftz.f32 %f2706, %f5140, %f2705; fma.rn.ftz.f32 %f585, %f5142, %f2706, 0f00000000; ld.local.f32 %f586, [%rd3+32]; mul.ftz.f32 %f587, %f586, 0f3E800000; rcp.approx.ftz.f32 %f588, %f587; mul.ftz.f32 %f2707, %f466, %f519; mul.ftz.f32 %f589, %f2707, %f588; mul.ftz.f32 %f2708, %f461, %f518; mul.ftz.f32 %f590, %f2708, %f588; setp.gtu.ftz.f32 %p239, %f587, 0f3DCCCCCD; @%p239 bra $L__BB105_278; bra.uni $L__BB105_265; $L__BB105_278: mul.ftz.f32 %f613, %f589, %f589; fma.rn.ftz.f32 %f5145, %f613, 0f3E800000, 0f3F800000; mul.ftz.f32 %f615, %f613, %f613; mov.u64 %rd39, 64; cvt.rn.f32.u64 %f2739, %rd39; div.approx.ftz.f32 %f2740, %f615, %f2739; add.ftz.f32 %f616, %f5145, %f2740; setp.eq.ftz.f32 %p249, %f5145, %f616; @%p249 bra $L__BB105_287; mul.ftz.f32 %f617, %f613, %f615; mov.u64 %rd40, 2304; cvt.rn.f32.u64 %f2741, %rd40; div.approx.ftz.f32 %f2742, %f617, %f2741; add.ftz.f32 %f618, %f616, %f2742; setp.eq.ftz.f32 %p250, %f616, %f618; mov.f32 %f5145, %f616; @%p250 bra $L__BB105_287; mul.ftz.f32 %f619, %f613, %f617; mov.u64 %rd41, 147456; cvt.rn.f32.u64 %f2743, %rd41; div.approx.ftz.f32 %f2744, %f619, %f2743; add.ftz.f32 %f620, %f618, %f2744; setp.eq.ftz.f32 %p251, %f618, %f620; mov.f32 %f5145, %f618; @%p251 bra $L__BB105_287; mul.ftz.f32 %f621, %f613, %f619; mov.u64 %rd42, 14745600; cvt.rn.f32.u64 %f2745, %rd42; div.approx.ftz.f32 %f2746, %f621, %f2745; add.ftz.f32 %f622, %f620, %f2746; setp.eq.ftz.f32 %p252, %f620, %f622; mov.f32 %f5145, %f620; @%p252 bra $L__BB105_287; mul.ftz.f32 %f623, %f613, %f621; mov.u64 %rd43, 2123366400; cvt.rn.f32.u64 %f2747, %rd43; div.approx.ftz.f32 %f2748, %f623, %f2747; add.ftz.f32 %f624, %f622, %f2748; setp.eq.ftz.f32 %p253, %f622, %f624; mov.f32 %f5145, %f622; @%p253 bra $L__BB105_287; mul.ftz.f32 %f625, %f613, %f623; mov.u64 %rd44, 416179814400; cvt.rn.f32.u64 %f2749, %rd44; div.approx.ftz.f32 %f2750, %f625, %f2749; add.ftz.f32 %f626, %f624, %f2750; setp.eq.ftz.f32 %p254, %f624, %f626; mov.f32 %f5145, %f624; @%p254 bra $L__BB105_287; mul.ftz.f32 %f627, %f613, %f625; mov.u64 %rd45, 106542032486400; cvt.rn.f32.u64 %f2751, %rd45; div.approx.ftz.f32 %f2752, %f627, %f2751; add.ftz.f32 %f628, %f626, %f2752; setp.eq.ftz.f32 %p255, %f626, %f628; mov.f32 %f5145, %f626; @%p255 bra $L__BB105_287; mul.ftz.f32 %f2753, %f613, %f627; mov.u64 %rd46, 34519618525593600; cvt.rn.f32.u64 %f2754, %rd46; div.approx.ftz.f32 %f2755, %f2753, %f2754; add.ftz.f32 %f629, %f628, %f2755; setp.eq.ftz.f32 %p256, %f628, %f629; mov.f32 %f5145, %f628; @%p256 bra $L__BB105_287; mov.f32 %f5145, %f629; $L__BB105_287: mul.ftz.f32 %f2756, %f590, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2757, %f2756; mul.ftz.f32 %f631, %f5145, %f2757; abs.ftz.f32 %f632, %f588; setp.ltu.ftz.f32 %p257, %f632, 0f3F800000; @%p257 bra $L__BB105_289; bra.uni $L__BB105_288; $L__BB105_289: mul.ftz.f32 %f2779, %f588, %f588; mov.f32 %f2780, 0f394FFF49; mov.f32 %f2781, 0f363D0ADA; fma.rn.ftz.f32 %f2782, %f2781, %f2779, %f2780; mov.f32 %f2783, 0f3C08889A; fma.rn.ftz.f32 %f2784, %f2782, %f2779, %f2783; mov.f32 %f2785, 0f3E2AAAAB; fma.rn.ftz.f32 %f2786, %f2784, %f2779, %f2785; mul.ftz.f32 %f2787, %f2779, %f2786; fma.rn.ftz.f32 %f5146, %f2787, %f588, %f588; bra.uni $L__BB105_290; $L__BB105_265: setp.gt.ftz.f32 %p240, %f589, 0f41400000; @%p240 bra $L__BB105_276; bra.uni $L__BB105_266; $L__BB105_276: mul.ftz.f32 %f2727, %f589, 0f41000000; rcp.approx.ftz.f32 %f2728, %f2727; add.ftz.f32 %f2729, %f2728, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2730, %f589; fma.rn.ftz.f32 %f2731, %f2730, 0fBF317218, %f2729; fma.rn.ftz.f32 %f5144, %f2731, 0f3F000000, %f589; bra.uni $L__BB105_277; $L__BB105_288: mov.f32 %f2758, 0f3FB8AA3B; mul.rn.ftz.f32 %f2759, %f632, %f2758; cvt.rzi.ftz.f32.f32 %f2760, %f2759; abs.ftz.f32 %f2761, %f2760; setp.gt.ftz.f32 %p258, %f2761, 0f42FC0000; mov.b32 %r86, %f2760; and.b32 %r87, %r86, -2147483648; or.b32 %r88, %r87, 1123811328; mov.b32 %f2762, %r88; selp.f32 %f2763, %f2762, %f2760, %p258; mov.f32 %f2764, 0fBF317218; fma.rn.ftz.f32 %f2765, %f2763, %f2764, %f632; mov.f32 %f2766, 0f3102E308; fma.rn.ftz.f32 %f2767, %f2763, %f2766, %f2765; mul.ftz.f32 %f2768, %f2767, 0f3FB8AA3B; add.ftz.f32 %f2769, %f2763, 0f4B40007D; mov.b32 %r89, %f2769; shl.b32 %r90, %r89, 23; mov.b32 %f2770, %r90; ex2.approx.ftz.f32 %f2771, %f2768; mul.ftz.f32 %f2772, %f2771, %f2770; mov.f32 %f2773, 0f3E000000; div.approx.ftz.f32 %f2774, %f2773, %f2772; neg.ftz.f32 %f2775, %f2774; mov.f32 %f2776, 0f40000000; fma.rn.ftz.f32 %f2777, %f2776, %f2772, %f2775; setp.ge.ftz.f32 %p259, %f632, 0f42B40000; selp.f32 %f2778, 0f7F800000, %f2777, %p259; mov.b32 %r91, %f2778; mov.b32 %r92, %f588; and.b32 %r93, %r92, -2147483648; or.b32 %r94, %r93, %r91; mov.b32 %f5146, %r94; $L__BB105_290: add.ftz.f32 %f2788, %f5146, %f5146; mul.ftz.f32 %f2789, %f587, %f2788; div.approx.ftz.f32 %f5147, %f631, %f2789; bra.uni $L__BB105_291; $L__BB105_266: mul.ftz.f32 %f591, %f589, %f589; fma.rn.ftz.f32 %f5143, %f591, 0f3E800000, 0f3F800000; mul.ftz.f32 %f593, %f591, %f591; mov.u64 %rd31, 64; cvt.rn.f32.u64 %f2709, %rd31; div.approx.ftz.f32 %f2710, %f593, %f2709; add.ftz.f32 %f594, %f5143, %f2710; setp.eq.ftz.f32 %p241, %f5143, %f594; @%p241 bra $L__BB105_275; mul.ftz.f32 %f595, %f591, %f593; mov.u64 %rd32, 2304; cvt.rn.f32.u64 %f2711, %rd32; div.approx.ftz.f32 %f2712, %f595, %f2711; add.ftz.f32 %f596, %f594, %f2712; setp.eq.ftz.f32 %p242, %f594, %f596; mov.f32 %f5143, %f594; @%p242 bra $L__BB105_275; mul.ftz.f32 %f597, %f591, %f595; mov.u64 %rd33, 147456; cvt.rn.f32.u64 %f2713, %rd33; div.approx.ftz.f32 %f2714, %f597, %f2713; add.ftz.f32 %f598, %f596, %f2714; setp.eq.ftz.f32 %p243, %f596, %f598; mov.f32 %f5143, %f596; @%p243 bra $L__BB105_275; mul.ftz.f32 %f599, %f591, %f597; mov.u64 %rd34, 14745600; cvt.rn.f32.u64 %f2715, %rd34; div.approx.ftz.f32 %f2716, %f599, %f2715; add.ftz.f32 %f600, %f598, %f2716; setp.eq.ftz.f32 %p244, %f598, %f600; mov.f32 %f5143, %f598; @%p244 bra $L__BB105_275; mul.ftz.f32 %f601, %f591, %f599; mov.u64 %rd35, 2123366400; cvt.rn.f32.u64 %f2717, %rd35; div.approx.ftz.f32 %f2718, %f601, %f2717; add.ftz.f32 %f602, %f600, %f2718; setp.eq.ftz.f32 %p245, %f600, %f602; mov.f32 %f5143, %f600; @%p245 bra $L__BB105_275; mul.ftz.f32 %f603, %f591, %f601; mov.u64 %rd36, 416179814400; cvt.rn.f32.u64 %f2719, %rd36; div.approx.ftz.f32 %f2720, %f603, %f2719; add.ftz.f32 %f604, %f602, %f2720; setp.eq.ftz.f32 %p246, %f602, %f604; mov.f32 %f5143, %f602; @%p246 bra $L__BB105_275; mul.ftz.f32 %f605, %f591, %f603; mov.u64 %rd37, 106542032486400; cvt.rn.f32.u64 %f2721, %rd37; div.approx.ftz.f32 %f2722, %f605, %f2721; add.ftz.f32 %f606, %f604, %f2722; setp.eq.ftz.f32 %p247, %f604, %f606; mov.f32 %f5143, %f604; @%p247 bra $L__BB105_275; mul.ftz.f32 %f2723, %f591, %f605; mov.u64 %rd38, 34519618525593600; cvt.rn.f32.u64 %f2724, %rd38; div.approx.ftz.f32 %f2725, %f2723, %f2724; add.ftz.f32 %f607, %f606, %f2725; setp.eq.ftz.f32 %p248, %f606, %f607; mov.f32 %f5143, %f606; @%p248 bra $L__BB105_275; mov.f32 %f5143, %f607; $L__BB105_275: lg2.approx.ftz.f32 %f2726, %f5143; mul.ftz.f32 %f5144, %f2726, 0f3F317218; $L__BB105_277: sub.ftz.f32 %f2732, %f5144, %f590; sub.ftz.f32 %f2733, %f2732, %f588; add.ftz.f32 %f2734, %f2733, 0f3F316F00; mul.ftz.f32 %f2735, %f588, 0f3F000000; lg2.approx.ftz.f32 %f2736, %f2735; fma.rn.ftz.f32 %f2737, %f2736, 0f3F317218, %f2734; mul.ftz.f32 %f2738, %f2737, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5147, %f2738; $L__BB105_291: mov.f32 %f5149, 0f00000000; mov.f32 %f5148, %f5149; @%p237 bra $L__BB105_293; ex2.approx.ftz.f32 %f2791, %f524; add.ftz.f32 %f2792, %f2791, 0f3F800000; rcp.approx.ftz.f32 %f2793, %f2792; add.ftz.f32 %f5148, %f2793, %f2793; $L__BB105_293: sub.ftz.f32 %f640, %f2532, %f5148; setp.eq.ftz.f32 %p261, %f640, 0f00000000; @%p261 bra $L__BB105_295; add.ftz.f32 %f2796, %f474, %f474; sub.ftz.f32 %f2797, %f2796, %f522; add.ftz.f32 %f2798, %f2797, 0f40490FDB; sub.ftz.f32 %f2799, %f516, %f2798; add.ftz.f32 %f2800, %f2799, 0f40490FDB; mov.f32 %f2801, 0f40C90FDB; div.approx.ftz.f32 %f2802, %f2800, %f2801; cvt.rmi.ftz.f32.f32 %f2803, %f2802; mul.ftz.f32 %f2804, %f2803, 0f40C90FDB; sub.ftz.f32 %f2805, %f2800, %f2804; add.ftz.f32 %f2806, %f2805, 0fC0490FDB; abs.ftz.f32 %f2807, %f2806; neg.ftz.f32 %f2808, %f2807; div.approx.ftz.f32 %f2809, %f2808, %f2562; mul.ftz.f32 %f2810, %f2809, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2811, %f2810; add.ftz.f32 %f2812, %f2811, 0f3F800000; mul.ftz.f32 %f2813, %f2812, %f2812; mul.ftz.f32 %f2814, %f2562, %f2813; div.approx.ftz.f32 %f2815, %f2811, %f2814; div.approx.ftz.f32 %f5149, %f2815, %f640; $L__BB105_295: mul.ftz.f32 %f2816, %f5147, %f482; mul.ftz.f32 %f2817, %f5147, %f483; mul.ftz.f32 %f2818, %f5147, %f484; fma.rn.ftz.f32 %f643, %f5149, %f2816, %f584; fma.rn.ftz.f32 %f644, %f5149, %f2817, %f584; fma.rn.ftz.f32 %f645, %f5149, %f2818, %f584; mul.ftz.f32 %f2819, %f5147, %f576; fma.rn.ftz.f32 %f646, %f5149, %f2819, %f585; mul.ftz.f32 %f647, %f586, 0f40800000; rcp.approx.ftz.f32 %f648, %f647; abs.ftz.f32 %f2820, %f521; mul.ftz.f32 %f2821, %f466, %f2820; mul.ftz.f32 %f649, %f2821, %f648; mul.ftz.f32 %f2822, %f461, %f520; mul.ftz.f32 %f650, %f2822, %f648; setp.gtu.ftz.f32 %p262, %f647, 0f3DCCCCCD; @%p262 bra $L__BB105_309; bra.uni $L__BB105_296; $L__BB105_309: mul.ftz.f32 %f673, %f649, %f649; fma.rn.ftz.f32 %f5152, %f673, 0f3E800000, 0f3F800000; mul.ftz.f32 %f675, %f673, %f673; mov.u64 %rd55, 64; cvt.rn.f32.u64 %f2853, %rd55; div.approx.ftz.f32 %f2854, %f675, %f2853; add.ftz.f32 %f676, %f5152, %f2854; setp.eq.ftz.f32 %p272, %f5152, %f676; @%p272 bra $L__BB105_318; mul.ftz.f32 %f677, %f673, %f675; mov.u64 %rd56, 2304; cvt.rn.f32.u64 %f2855, %rd56; div.approx.ftz.f32 %f2856, %f677, %f2855; add.ftz.f32 %f678, %f676, %f2856; setp.eq.ftz.f32 %p273, %f676, %f678; mov.f32 %f5152, %f676; @%p273 bra $L__BB105_318; mul.ftz.f32 %f679, %f673, %f677; mov.u64 %rd57, 147456; cvt.rn.f32.u64 %f2857, %rd57; div.approx.ftz.f32 %f2858, %f679, %f2857; add.ftz.f32 %f680, %f678, %f2858; setp.eq.ftz.f32 %p274, %f678, %f680; mov.f32 %f5152, %f678; @%p274 bra $L__BB105_318; mul.ftz.f32 %f681, %f673, %f679; mov.u64 %rd58, 14745600; cvt.rn.f32.u64 %f2859, %rd58; div.approx.ftz.f32 %f2860, %f681, %f2859; add.ftz.f32 %f682, %f680, %f2860; setp.eq.ftz.f32 %p275, %f680, %f682; mov.f32 %f5152, %f680; @%p275 bra $L__BB105_318; mul.ftz.f32 %f683, %f673, %f681; mov.u64 %rd59, 2123366400; cvt.rn.f32.u64 %f2861, %rd59; div.approx.ftz.f32 %f2862, %f683, %f2861; add.ftz.f32 %f684, %f682, %f2862; setp.eq.ftz.f32 %p276, %f682, %f684; mov.f32 %f5152, %f682; @%p276 bra $L__BB105_318; mul.ftz.f32 %f685, %f673, %f683; mov.u64 %rd60, 416179814400; cvt.rn.f32.u64 %f2863, %rd60; div.approx.ftz.f32 %f2864, %f685, %f2863; add.ftz.f32 %f686, %f684, %f2864; setp.eq.ftz.f32 %p277, %f684, %f686; mov.f32 %f5152, %f684; @%p277 bra $L__BB105_318; mul.ftz.f32 %f687, %f673, %f685; mov.u64 %rd61, 106542032486400; cvt.rn.f32.u64 %f2865, %rd61; div.approx.ftz.f32 %f2866, %f687, %f2865; add.ftz.f32 %f688, %f686, %f2866; setp.eq.ftz.f32 %p278, %f686, %f688; mov.f32 %f5152, %f686; @%p278 bra $L__BB105_318; mul.ftz.f32 %f2867, %f673, %f687; mov.u64 %rd62, 34519618525593600; cvt.rn.f32.u64 %f2868, %rd62; div.approx.ftz.f32 %f2869, %f2867, %f2868; add.ftz.f32 %f689, %f688, %f2869; setp.eq.ftz.f32 %p279, %f688, %f689; mov.f32 %f5152, %f688; @%p279 bra $L__BB105_318; mov.f32 %f5152, %f689; $L__BB105_318: mul.ftz.f32 %f2870, %f650, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2871, %f2870; mul.ftz.f32 %f691, %f5152, %f2871; abs.ftz.f32 %f692, %f648; setp.ltu.ftz.f32 %p280, %f692, 0f3F800000; @%p280 bra $L__BB105_320; bra.uni $L__BB105_319; $L__BB105_320: mul.ftz.f32 %f2893, %f648, %f648; mov.f32 %f2894, 0f394FFF49; mov.f32 %f2895, 0f363D0ADA; fma.rn.ftz.f32 %f2896, %f2895, %f2893, %f2894; mov.f32 %f2897, 0f3C08889A; fma.rn.ftz.f32 %f2898, %f2896, %f2893, %f2897; mov.f32 %f2899, 0f3E2AAAAB; fma.rn.ftz.f32 %f2900, %f2898, %f2893, %f2899; mul.ftz.f32 %f2901, %f2893, %f2900; fma.rn.ftz.f32 %f5153, %f2901, %f648, %f648; bra.uni $L__BB105_321; $L__BB105_296: setp.gt.ftz.f32 %p263, %f649, 0f41400000; @%p263 bra $L__BB105_307; bra.uni $L__BB105_297; $L__BB105_307: mul.ftz.f32 %f2841, %f649, 0f41000000; rcp.approx.ftz.f32 %f2842, %f2841; add.ftz.f32 %f2843, %f2842, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2844, %f649; fma.rn.ftz.f32 %f2845, %f2844, 0fBF317218, %f2843; fma.rn.ftz.f32 %f5151, %f2845, 0f3F000000, %f649; bra.uni $L__BB105_308; $L__BB105_319: mov.f32 %f2872, 0f3FB8AA3B; mul.rn.ftz.f32 %f2873, %f692, %f2872; cvt.rzi.ftz.f32.f32 %f2874, %f2873; abs.ftz.f32 %f2875, %f2874; setp.gt.ftz.f32 %p281, %f2875, 0f42FC0000; mov.b32 %r95, %f2874; and.b32 %r96, %r95, -2147483648; or.b32 %r97, %r96, 1123811328; mov.b32 %f2876, %r97; selp.f32 %f2877, %f2876, %f2874, %p281; mov.f32 %f2878, 0fBF317218; fma.rn.ftz.f32 %f2879, %f2877, %f2878, %f692; mov.f32 %f2880, 0f3102E308; fma.rn.ftz.f32 %f2881, %f2877, %f2880, %f2879; mul.ftz.f32 %f2882, %f2881, 0f3FB8AA3B; add.ftz.f32 %f2883, %f2877, 0f4B40007D; mov.b32 %r98, %f2883; shl.b32 %r99, %r98, 23; mov.b32 %f2884, %r99; ex2.approx.ftz.f32 %f2885, %f2882; mul.ftz.f32 %f2886, %f2885, %f2884; mov.f32 %f2887, 0f3E000000; div.approx.ftz.f32 %f2888, %f2887, %f2886; neg.ftz.f32 %f2889, %f2888; mov.f32 %f2890, 0f40000000; fma.rn.ftz.f32 %f2891, %f2890, %f2886, %f2889; setp.ge.ftz.f32 %p282, %f692, 0f42B40000; selp.f32 %f2892, 0f7F800000, %f2891, %p282; mov.b32 %r100, %f2892; mov.b32 %r101, %f648; and.b32 %r102, %r101, -2147483648; or.b32 %r103, %r102, %r100; mov.b32 %f5153, %r103; $L__BB105_321: add.ftz.f32 %f2902, %f5153, %f5153; mul.ftz.f32 %f2903, %f647, %f2902; div.approx.ftz.f32 %f5154, %f691, %f2903; bra.uni $L__BB105_322; $L__BB105_297: mul.ftz.f32 %f651, %f649, %f649; fma.rn.ftz.f32 %f5150, %f651, 0f3E800000, 0f3F800000; mul.ftz.f32 %f653, %f651, %f651; mov.u64 %rd47, 64; cvt.rn.f32.u64 %f2823, %rd47; div.approx.ftz.f32 %f2824, %f653, %f2823; add.ftz.f32 %f654, %f5150, %f2824; setp.eq.ftz.f32 %p264, %f5150, %f654; @%p264 bra $L__BB105_306; mul.ftz.f32 %f655, %f651, %f653; mov.u64 %rd48, 2304; cvt.rn.f32.u64 %f2825, %rd48; div.approx.ftz.f32 %f2826, %f655, %f2825; add.ftz.f32 %f656, %f654, %f2826; setp.eq.ftz.f32 %p265, %f654, %f656; mov.f32 %f5150, %f654; @%p265 bra $L__BB105_306; mul.ftz.f32 %f657, %f651, %f655; mov.u64 %rd49, 147456; cvt.rn.f32.u64 %f2827, %rd49; div.approx.ftz.f32 %f2828, %f657, %f2827; add.ftz.f32 %f658, %f656, %f2828; setp.eq.ftz.f32 %p266, %f656, %f658; mov.f32 %f5150, %f656; @%p266 bra $L__BB105_306; mul.ftz.f32 %f659, %f651, %f657; mov.u64 %rd50, 14745600; cvt.rn.f32.u64 %f2829, %rd50; div.approx.ftz.f32 %f2830, %f659, %f2829; add.ftz.f32 %f660, %f658, %f2830; setp.eq.ftz.f32 %p267, %f658, %f660; mov.f32 %f5150, %f658; @%p267 bra $L__BB105_306; mul.ftz.f32 %f661, %f651, %f659; mov.u64 %rd51, 2123366400; cvt.rn.f32.u64 %f2831, %rd51; div.approx.ftz.f32 %f2832, %f661, %f2831; add.ftz.f32 %f662, %f660, %f2832; setp.eq.ftz.f32 %p268, %f660, %f662; mov.f32 %f5150, %f660; @%p268 bra $L__BB105_306; mul.ftz.f32 %f663, %f651, %f661; mov.u64 %rd52, 416179814400; cvt.rn.f32.u64 %f2833, %rd52; div.approx.ftz.f32 %f2834, %f663, %f2833; add.ftz.f32 %f664, %f662, %f2834; setp.eq.ftz.f32 %p269, %f662, %f664; mov.f32 %f5150, %f662; @%p269 bra $L__BB105_306; mul.ftz.f32 %f665, %f651, %f663; mov.u64 %rd53, 106542032486400; cvt.rn.f32.u64 %f2835, %rd53; div.approx.ftz.f32 %f2836, %f665, %f2835; add.ftz.f32 %f666, %f664, %f2836; setp.eq.ftz.f32 %p270, %f664, %f666; mov.f32 %f5150, %f664; @%p270 bra $L__BB105_306; mul.ftz.f32 %f2837, %f651, %f665; mov.u64 %rd54, 34519618525593600; cvt.rn.f32.u64 %f2838, %rd54; div.approx.ftz.f32 %f2839, %f2837, %f2838; add.ftz.f32 %f667, %f666, %f2839; setp.eq.ftz.f32 %p271, %f666, %f667; mov.f32 %f5150, %f666; @%p271 bra $L__BB105_306; mov.f32 %f5150, %f667; $L__BB105_306: lg2.approx.ftz.f32 %f2840, %f5150; mul.ftz.f32 %f5151, %f2840, 0f3F317218; $L__BB105_308: sub.ftz.f32 %f2846, %f5151, %f650; sub.ftz.f32 %f2847, %f2846, %f648; add.ftz.f32 %f2848, %f2847, 0f3F316F00; mul.ftz.f32 %f2849, %f648, 0f3F000000; lg2.approx.ftz.f32 %f2850, %f2849; fma.rn.ftz.f32 %f2851, %f2850, 0f3F317218, %f2848; mul.ftz.f32 %f2852, %f2851, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5154, %f2852; $L__BB105_322: mul.ftz.f32 %f2905, %f474, 0f40800000; sub.ftz.f32 %f2906, %f2905, %f522; add.ftz.f32 %f2907, %f2906, 0f40C90FDB; sub.ftz.f32 %f698, %f516, %f2907; mov.f32 %f5156, 0f00000000; mov.f32 %f5155, %f5156; @%p237 bra $L__BB105_324; ex2.approx.ftz.f32 %f2908, %f524; add.ftz.f32 %f2909, %f2908, 0f3F800000; rcp.approx.ftz.f32 %f2910, %f2909; add.ftz.f32 %f5155, %f2910, %f2910; $L__BB105_324: add.ftz.f32 %f2912, %f698, 0f40490FDB; mov.f32 %f2913, 0f40C90FDB; div.approx.ftz.f32 %f2914, %f2912, %f2913; cvt.rmi.ftz.f32.f32 %f2915, %f2914; mul.ftz.f32 %f2916, %f2915, 0f40C90FDB; sub.ftz.f32 %f2917, %f2912, %f2916; add.ftz.f32 %f2918, %f2917, 0fC0490FDB; abs.ftz.f32 %f2919, %f2918; neg.ftz.f32 %f2920, %f2919; div.approx.ftz.f32 %f2921, %f2920, %f2562; mul.ftz.f32 %f2922, %f2921, 0f3FB8AA3B; ex2.approx.ftz.f32 %f2923, %f2922; add.ftz.f32 %f2924, %f2923, 0f3F800000; mul.ftz.f32 %f2926, %f2924, %f2924; mul.ftz.f32 %f2927, %f2562, %f2926; div.approx.ftz.f32 %f701, %f2923, %f2927; sub.ftz.f32 %f702, %f2532, %f5155; setp.eq.ftz.f32 %p284, %f702, 0f00000000; @%p284 bra $L__BB105_326; div.approx.ftz.f32 %f5156, %f701, %f702; $L__BB105_326: mul.ftz.f32 %f2928, %f5154, %f487; mul.ftz.f32 %f2929, %f5154, %f489; mul.ftz.f32 %f2930, %f5154, %f491; fma.rn.ftz.f32 %f705, %f5156, %f2928, %f643; fma.rn.ftz.f32 %f706, %f5156, %f2929, %f644; fma.rn.ftz.f32 %f707, %f5156, %f2930, %f645; mul.ftz.f32 %f2931, %f5154, %f577; fma.rn.ftz.f32 %f708, %f5156, %f2931, %f646; mul.ftz.f32 %f2932, %f466, %f509; mul.ftz.f32 %f709, %f2932, %f648; mul.ftz.f32 %f2933, %f462, %f461; mul.ftz.f32 %f710, %f2933, %f648; @%p262 bra $L__BB105_340; bra.uni $L__BB105_327; $L__BB105_340: mul.ftz.f32 %f733, %f709, %f709; fma.rn.ftz.f32 %f5159, %f733, 0f3E800000, 0f3F800000; mul.ftz.f32 %f735, %f733, %f733; mov.u64 %rd71, 64; cvt.rn.f32.u64 %f2964, %rd71; div.approx.ftz.f32 %f2965, %f735, %f2964; add.ftz.f32 %f736, %f5159, %f2965; setp.eq.ftz.f32 %p295, %f5159, %f736; @%p295 bra $L__BB105_349; mul.ftz.f32 %f737, %f733, %f735; mov.u64 %rd72, 2304; cvt.rn.f32.u64 %f2966, %rd72; div.approx.ftz.f32 %f2967, %f737, %f2966; add.ftz.f32 %f738, %f736, %f2967; setp.eq.ftz.f32 %p296, %f736, %f738; mov.f32 %f5159, %f736; @%p296 bra $L__BB105_349; mul.ftz.f32 %f739, %f733, %f737; mov.u64 %rd73, 147456; cvt.rn.f32.u64 %f2968, %rd73; div.approx.ftz.f32 %f2969, %f739, %f2968; add.ftz.f32 %f740, %f738, %f2969; setp.eq.ftz.f32 %p297, %f738, %f740; mov.f32 %f5159, %f738; @%p297 bra $L__BB105_349; mul.ftz.f32 %f741, %f733, %f739; mov.u64 %rd74, 14745600; cvt.rn.f32.u64 %f2970, %rd74; div.approx.ftz.f32 %f2971, %f741, %f2970; add.ftz.f32 %f742, %f740, %f2971; setp.eq.ftz.f32 %p298, %f740, %f742; mov.f32 %f5159, %f740; @%p298 bra $L__BB105_349; mul.ftz.f32 %f743, %f733, %f741; mov.u64 %rd75, 2123366400; cvt.rn.f32.u64 %f2972, %rd75; div.approx.ftz.f32 %f2973, %f743, %f2972; add.ftz.f32 %f744, %f742, %f2973; setp.eq.ftz.f32 %p299, %f742, %f744; mov.f32 %f5159, %f742; @%p299 bra $L__BB105_349; mul.ftz.f32 %f745, %f733, %f743; mov.u64 %rd76, 416179814400; cvt.rn.f32.u64 %f2974, %rd76; div.approx.ftz.f32 %f2975, %f745, %f2974; add.ftz.f32 %f746, %f744, %f2975; setp.eq.ftz.f32 %p300, %f744, %f746; mov.f32 %f5159, %f744; @%p300 bra $L__BB105_349; mul.ftz.f32 %f747, %f733, %f745; mov.u64 %rd77, 106542032486400; cvt.rn.f32.u64 %f2976, %rd77; div.approx.ftz.f32 %f2977, %f747, %f2976; add.ftz.f32 %f748, %f746, %f2977; setp.eq.ftz.f32 %p301, %f746, %f748; mov.f32 %f5159, %f746; @%p301 bra $L__BB105_349; mul.ftz.f32 %f2978, %f733, %f747; mov.u64 %rd78, 34519618525593600; cvt.rn.f32.u64 %f2979, %rd78; div.approx.ftz.f32 %f2980, %f2978, %f2979; add.ftz.f32 %f749, %f748, %f2980; setp.eq.ftz.f32 %p302, %f748, %f749; mov.f32 %f5159, %f748; @%p302 bra $L__BB105_349; mov.f32 %f5159, %f749; $L__BB105_349: mul.ftz.f32 %f2981, %f710, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2982, %f2981; mul.ftz.f32 %f751, %f5159, %f2982; abs.ftz.f32 %f752, %f648; setp.ltu.ftz.f32 %p303, %f752, 0f3F800000; @%p303 bra $L__BB105_351; bra.uni $L__BB105_350; $L__BB105_351: mul.ftz.f32 %f3004, %f648, %f648; mov.f32 %f3005, 0f394FFF49; mov.f32 %f3006, 0f363D0ADA; fma.rn.ftz.f32 %f3007, %f3006, %f3004, %f3005; mov.f32 %f3008, 0f3C08889A; fma.rn.ftz.f32 %f3009, %f3007, %f3004, %f3008; mov.f32 %f3010, 0f3E2AAAAB; fma.rn.ftz.f32 %f3011, %f3009, %f3004, %f3010; mul.ftz.f32 %f3012, %f3004, %f3011; fma.rn.ftz.f32 %f5160, %f3012, %f648, %f648; bra.uni $L__BB105_352; $L__BB105_327: setp.gt.ftz.f32 %p286, %f709, 0f41400000; @%p286 bra $L__BB105_338; bra.uni $L__BB105_328; $L__BB105_338: mul.ftz.f32 %f2952, %f709, 0f41000000; rcp.approx.ftz.f32 %f2953, %f2952; add.ftz.f32 %f2954, %f2953, 0fBFEB3F8E; lg2.approx.ftz.f32 %f2955, %f709; fma.rn.ftz.f32 %f2956, %f2955, 0fBF317218, %f2954; fma.rn.ftz.f32 %f5158, %f2956, 0f3F000000, %f709; bra.uni $L__BB105_339; $L__BB105_350: mov.f32 %f2983, 0f3FB8AA3B; mul.rn.ftz.f32 %f2984, %f752, %f2983; cvt.rzi.ftz.f32.f32 %f2985, %f2984; abs.ftz.f32 %f2986, %f2985; setp.gt.ftz.f32 %p304, %f2986, 0f42FC0000; mov.b32 %r104, %f2985; and.b32 %r105, %r104, -2147483648; or.b32 %r106, %r105, 1123811328; mov.b32 %f2987, %r106; selp.f32 %f2988, %f2987, %f2985, %p304; mov.f32 %f2989, 0fBF317218; fma.rn.ftz.f32 %f2990, %f2988, %f2989, %f752; mov.f32 %f2991, 0f3102E308; fma.rn.ftz.f32 %f2992, %f2988, %f2991, %f2990; mul.ftz.f32 %f2993, %f2992, 0f3FB8AA3B; add.ftz.f32 %f2994, %f2988, 0f4B40007D; mov.b32 %r107, %f2994; shl.b32 %r108, %r107, 23; mov.b32 %f2995, %r108; ex2.approx.ftz.f32 %f2996, %f2993; mul.ftz.f32 %f2997, %f2996, %f2995; mov.f32 %f2998, 0f3E000000; div.approx.ftz.f32 %f2999, %f2998, %f2997; neg.ftz.f32 %f3000, %f2999; mov.f32 %f3001, 0f40000000; fma.rn.ftz.f32 %f3002, %f3001, %f2997, %f3000; setp.ge.ftz.f32 %p305, %f752, 0f42B40000; selp.f32 %f3003, 0f7F800000, %f3002, %p305; mov.b32 %r109, %f3003; mov.b32 %r110, %f648; and.b32 %r111, %r110, -2147483648; or.b32 %r112, %r111, %r109; mov.b32 %f5160, %r112; $L__BB105_352: add.ftz.f32 %f3013, %f5160, %f5160; mul.ftz.f32 %f3014, %f647, %f3013; div.approx.ftz.f32 %f5161, %f751, %f3014; bra.uni $L__BB105_353; $L__BB105_328: mul.ftz.f32 %f711, %f709, %f709; fma.rn.ftz.f32 %f5157, %f711, 0f3E800000, 0f3F800000; mul.ftz.f32 %f713, %f711, %f711; mov.u64 %rd63, 64; cvt.rn.f32.u64 %f2934, %rd63; div.approx.ftz.f32 %f2935, %f713, %f2934; add.ftz.f32 %f714, %f5157, %f2935; setp.eq.ftz.f32 %p287, %f5157, %f714; @%p287 bra $L__BB105_337; mul.ftz.f32 %f715, %f711, %f713; mov.u64 %rd64, 2304; cvt.rn.f32.u64 %f2936, %rd64; div.approx.ftz.f32 %f2937, %f715, %f2936; add.ftz.f32 %f716, %f714, %f2937; setp.eq.ftz.f32 %p288, %f714, %f716; mov.f32 %f5157, %f714; @%p288 bra $L__BB105_337; mul.ftz.f32 %f717, %f711, %f715; mov.u64 %rd65, 147456; cvt.rn.f32.u64 %f2938, %rd65; div.approx.ftz.f32 %f2939, %f717, %f2938; add.ftz.f32 %f718, %f716, %f2939; setp.eq.ftz.f32 %p289, %f716, %f718; mov.f32 %f5157, %f716; @%p289 bra $L__BB105_337; mul.ftz.f32 %f719, %f711, %f717; mov.u64 %rd66, 14745600; cvt.rn.f32.u64 %f2940, %rd66; div.approx.ftz.f32 %f2941, %f719, %f2940; add.ftz.f32 %f720, %f718, %f2941; setp.eq.ftz.f32 %p290, %f718, %f720; mov.f32 %f5157, %f718; @%p290 bra $L__BB105_337; mul.ftz.f32 %f721, %f711, %f719; mov.u64 %rd67, 2123366400; cvt.rn.f32.u64 %f2942, %rd67; div.approx.ftz.f32 %f2943, %f721, %f2942; add.ftz.f32 %f722, %f720, %f2943; setp.eq.ftz.f32 %p291, %f720, %f722; mov.f32 %f5157, %f720; @%p291 bra $L__BB105_337; mul.ftz.f32 %f723, %f711, %f721; mov.u64 %rd68, 416179814400; cvt.rn.f32.u64 %f2944, %rd68; div.approx.ftz.f32 %f2945, %f723, %f2944; add.ftz.f32 %f724, %f722, %f2945; setp.eq.ftz.f32 %p292, %f722, %f724; mov.f32 %f5157, %f722; @%p292 bra $L__BB105_337; mul.ftz.f32 %f725, %f711, %f723; mov.u64 %rd69, 106542032486400; cvt.rn.f32.u64 %f2946, %rd69; div.approx.ftz.f32 %f2947, %f725, %f2946; add.ftz.f32 %f726, %f724, %f2947; setp.eq.ftz.f32 %p293, %f724, %f726; mov.f32 %f5157, %f724; @%p293 bra $L__BB105_337; mul.ftz.f32 %f2948, %f711, %f725; mov.u64 %rd70, 34519618525593600; cvt.rn.f32.u64 %f2949, %rd70; div.approx.ftz.f32 %f2950, %f2948, %f2949; add.ftz.f32 %f727, %f726, %f2950; setp.eq.ftz.f32 %p294, %f726, %f727; mov.f32 %f5157, %f726; @%p294 bra $L__BB105_337; mov.f32 %f5157, %f727; $L__BB105_337: lg2.approx.ftz.f32 %f2951, %f5157; mul.ftz.f32 %f5158, %f2951, 0f3F317218; $L__BB105_339: sub.ftz.f32 %f2957, %f5158, %f710; sub.ftz.f32 %f2958, %f2957, %f648; add.ftz.f32 %f2959, %f2958, 0f3F316F00; mul.ftz.f32 %f2960, %f648, 0f3F000000; lg2.approx.ftz.f32 %f2961, %f2960; fma.rn.ftz.f32 %f2962, %f2961, 0f3F317218, %f2959; mul.ftz.f32 %f2963, %f2962, 0f3FB8AA3B; ex2.approx.ftz.f32 %f5161, %f2963; $L__BB105_353: mul.ftz.f32 %f3015, %f5161, %f502; mul.ftz.f32 %f3016, %f5161, %f503; mul.ftz.f32 %f3017, %f5161, %f504; fma.rn.ftz.f32 %f5296, %f3015, 0f3E22F983, %f705; fma.rn.ftz.f32 %f5295, %f3016, 0f3E22F983, %f706; fma.rn.ftz.f32 %f5294, %f3017, 0f3E22F983, %f707; mul.ftz.f32 %f3018, %f5161, %f578; fma.rn.ftz.f32 %f5297, %f3018, 0f3E22F983, %f708; bra.uni $L__BB105_524; $L__BB105_473: ld.local.f32 %f4160, [%rd3+52]; mul.ftz.f32 %f4161, %f1181, %f4160; ld.local.f32 %f4162, [%rd3+48]; mul.ftz.f32 %f4163, %f1182, %f4162; sub.ftz.f32 %f4164, %f4161, %f4163; ld.local.f32 %f4165, [%rd3+44]; mul.ftz.f32 %f4166, %f1182, %f4165; mul.ftz.f32 %f4167, %f1180, %f4160; sub.ftz.f32 %f4168, %f4166, %f4167; mul.ftz.f32 %f4169, %f1180, %f4162; mul.ftz.f32 %f4170, %f1181, %f4165; sub.ftz.f32 %f4171, %f4169, %f4170; mul.ftz.f32 %f4172, %f4168, %f4168; fma.rn.ftz.f32 %f4173, %f4164, %f4164, %f4172; fma.rn.ftz.f32 %f4174, %f4171, %f4171, %f4173; sqrt.approx.ftz.f32 %f4175, %f4174; div.approx.ftz.f32 %f4176, %f4164, %f4175; div.approx.ftz.f32 %f4177, %f4168, %f4175; div.approx.ftz.f32 %f4178, %f4171, %f4175; mul.ftz.f32 %f4179, %f1182, %f4177; mul.ftz.f32 %f4180, %f1181, %f4178; sub.ftz.f32 %f4181, %f4179, %f4180; mul.ftz.f32 %f4182, %f1180, %f4178; mul.ftz.f32 %f4183, %f1182, %f4176; sub.ftz.f32 %f4184, %f4182, %f4183; mul.ftz.f32 %f4185, %f1181, %f4176; mul.ftz.f32 %f4186, %f1180, %f4177; sub.ftz.f32 %f4187, %f4185, %f4186; mul.ftz.f32 %f4188, %f1186, %f4184; fma.rn.ftz.f32 %f4189, %f1185, %f4181, %f4188; fma.rn.ftz.f32 %f4190, %f1187, %f4187, %f4189; mul.ftz.f32 %f4191, %f1186, %f4177; fma.rn.ftz.f32 %f4192, %f1185, %f4176, %f4191; fma.rn.ftz.f32 %f4193, %f1187, %f4178, %f4192; neg.ftz.f32 %f4194, %f4190; mul.ftz.f32 %f4195, %f4136, %f1188; div.approx.ftz.f32 %f4196, %f4194, %f4195; neg.ftz.f32 %f4197, %f4193; mul.ftz.f32 %f4198, %f4137, %f1188; div.approx.ftz.f32 %f4199, %f4197, %f4198; fma.rn.ftz.f32 %f4200, %f4196, %f4196, 0f3F800000; mov.f32 %f4201, 0f3F800000; fma.rn.ftz.f32 %f4202, %f4199, %f4199, %f4200; mul.ftz.f32 %f4203, %f4202, %f4202; mul.ftz.f32 %f4204, %f4203, 0f40490FDB; mul.ftz.f32 %f4205, %f5261, %f4204; mul.ftz.f32 %f4206, %f1189, %f1189; mul.ftz.f32 %f4207, %f4206, %f4205; rcp.approx.ftz.f32 %f5262, %f4207; mul.ftz.f32 %f4208, %f1183, %f1183; sub.ftz.f32 %f4209, %f4201, %f4208; div.approx.ftz.f32 %f4210, %f4209, %f4208; mul.ftz.f32 %f4211, %f1174, %f4184; fma.rn.ftz.f32 %f4212, %f1173, %f4181, %f4211; fma.rn.ftz.f32 %f4213, %f1175, %f4187, %f4212; mul.ftz.f32 %f4214, %f1174, %f4177; fma.rn.ftz.f32 %f4215, %f1173, %f4176, %f4214; fma.rn.ftz.f32 %f4216, %f1175, %f4178, %f4215; mul.ftz.f32 %f4217, %f4213, %f4213; mul.ftz.f32 %f4218, %f4136, %f4136; mul.ftz.f32 %f4219, %f4218, %f4217; mul.ftz.f32 %f4220, %f4216, %f4216; mul.ftz.f32 %f4221, %f4137, %f4137; fma.rn.ftz.f32 %f4222, %f4221, %f4220, %f4219; add.ftz.f32 %f4223, %f4220, %f4217; div.approx.ftz.f32 %f4224, %f4222, %f4223; fma.rn.ftz.f32 %f4225, %f4210, %f4224, 0f3F800000; mov.f32 %f4226, 0f00000000; max.ftz.f32 %f4227, %f4225, %f4226; sqrt.approx.ftz.f32 %f4228, %f4227; add.ftz.f32 %f4229, %f4228, 0f3F800000; mov.f32 %f4230, 0f40000000; div.approx.ftz.f32 %f5264, %f4230, %f4229; mul.ftz.f32 %f4231, %f1184, %f1184; sub.ftz.f32 %f4232, %f4201, %f4231; div.approx.ftz.f32 %f4233, %f4232, %f4231; mul.ftz.f32 %f4234, %f1405, %f4184; fma.rn.ftz.f32 %f4235, %f1404, %f4181, %f4234; fma.rn.ftz.f32 %f4236, %f1406, %f4187, %f4235; mul.ftz.f32 %f4237, %f1405, %f4177; fma.rn.ftz.f32 %f4238, %f1404, %f4176, %f4237; fma.rn.ftz.f32 %f4239, %f1406, %f4178, %f4238; mul.ftz.f32 %f4240, %f4236, %f4236; mul.ftz.f32 %f4241, %f4218, %f4240; mul.ftz.f32 %f4242, %f4239, %f4239; fma.rn.ftz.f32 %f4243, %f4221, %f4242, %f4241; add.ftz.f32 %f4244, %f4242, %f4240; div.approx.ftz.f32 %f4245, %f4243, %f4244; mul.ftz.f32 %f5265, %f4233, %f4245; bra.uni $L__BB105_479; $L__BB105_475: mul.ftz.f32 %f4246, %f1189, %f1189; mul.ftz.f32 %f4247, %f4246, 0f40490FDB; mov.f32 %f4248, 0f3F800000; sub.ftz.f32 %f4249, %f4248, %f1189; div.approx.ftz.f32 %f4250, %f4249, %f1189; add.ftz.f32 %f4251, %f5261, %f4250; mul.ftz.f32 %f4252, %f4247, %f4251; mul.ftz.f32 %f4253, %f4251, %f4252; div.approx.ftz.f32 %f5262, %f5261, %f4253; $L__BB105_478: mul.ftz.f32 %f4265, %f1183, %f1183; mov.f32 %f4266, 0f3F800000; sub.ftz.f32 %f4267, %f4266, %f4265; mul.ftz.f32 %f4268, %f4267, %f5261; div.approx.ftz.f32 %f4269, %f4268, %f4265; add.ftz.f32 %f4270, %f4269, 0f3F800000; mov.f32 %f4271, 0f00000000; max.ftz.f32 %f4272, %f4270, %f4271; sqrt.approx.ftz.f32 %f4273, %f4272; add.ftz.f32 %f4274, %f4273, 0f3F800000; mov.f32 %f4275, 0f40000000; div.approx.ftz.f32 %f5264, %f4275, %f4274; mul.ftz.f32 %f4276, %f1184, %f1184; sub.ftz.f32 %f4277, %f4266, %f4276; mul.ftz.f32 %f4278, %f4277, %f5261; div.approx.ftz.f32 %f5265, %f4278, %f4276; $L__BB105_479: add.ftz.f32 %f4282, %f5265, 0f3F800000; mov.f32 %f4283, 0f00000000; max.ftz.f32 %f4284, %f4282, %f4283; sqrt.approx.ftz.f32 %f4285, %f4284; add.ftz.f32 %f4286, %f4285, 0f3F800000; mov.f32 %f4287, 0f40000000; div.approx.ftz.f32 %f4288, %f4287, %f4286; mul.ftz.f32 %f1202, %f5264, %f4288; mul.ftz.f32 %f4289, %f5262, 0f3E800000; div.approx.ftz.f32 %f1203, %f4289, %f1183; add.s32 %r133, %r5, -11; setp.gt.u32 %p454, %r133, 1; mov.f32 %f5273, 0f3F800000; mov.f32 %f5272, %f5273; mov.f32 %f5271, %f5273; @%p454 bra $L__BB105_485; ld.local.f32 %f4291, [%rd3+28]; fma.rn.ftz.f32 %f1204, %f4291, %f4291, 0fBF800000; mov.f32 %f4290, 0f3F800000; abs.ftz.f32 %f1205, %f4290; fma.rn.ftz.f32 %f1206, %f1205, %f1205, %f1204; setp.leu.ftz.f32 %p455, %f1206, 0f00000000; mov.f32 %f5266, %f4290; @%p455 bra $L__BB105_482; sqrt.approx.ftz.f32 %f4292, %f1206; sub.ftz.f32 %f4293, %f4292, %f1205; add.ftz.f32 %f4294, %f1205, %f4292; div.approx.ftz.f32 %f4295, %f4293, %f4294; fma.rn.ftz.f32 %f4296, %f1205, %f4294, 0fBF800000; fma.rn.ftz.f32 %f4297, %f1205, %f4293, 0f3F800000; div.approx.ftz.f32 %f4298, %f4296, %f4297; mul.ftz.f32 %f4299, %f4295, 0f3F000000; mul.ftz.f32 %f4300, %f4295, %f4299; fma.rn.ftz.f32 %f4301, %f4298, %f4298, 0f3F800000; mul.ftz.f32 %f5266, %f4300, %f4301; $L__BB105_482: ld.local.u64 %rd80, [%rd3+36]; ld.f32 %f1209, [%rd80+12]; ld.f32 %f1210, [%rd80+16]; ld.f32 %f1211, [%rd80+20]; sub.ftz.f32 %f1212, %f4290, %f5266; mul.ftz.f32 %f4303, %f1405, %f1186; fma.rn.ftz.f32 %f4304, %f1404, %f1185, %f4303; fma.rn.ftz.f32 %f4305, %f1406, %f1187, %f4304; abs.ftz.f32 %f1213, %f4305; fma.rn.ftz.f32 %f1214, %f1213, %f1213, %f1204; setp.leu.ftz.f32 %p456, %f1214, 0f00000000; mov.f32 %f5267, %f1212; @%p456 bra $L__BB105_484; sqrt.approx.ftz.f32 %f4306, %f1214; sub.ftz.f32 %f4307, %f4306, %f1213; add.ftz.f32 %f4308, %f1213, %f4306; div.approx.ftz.f32 %f4309, %f4307, %f4308; fma.rn.ftz.f32 %f4310, %f1213, %f4308, 0fBF800000; fma.rn.ftz.f32 %f4311, %f1213, %f4307, 0f3F800000; div.approx.ftz.f32 %f4312, %f4310, %f4311; mul.ftz.f32 %f4313, %f4309, 0f3F000000; mul.ftz.f32 %f4314, %f4309, %f4313; fma.rn.ftz.f32 %f4315, %f4312, %f4312, 0f3F800000; mul.ftz.f32 %f4316, %f4314, %f4315; sub.ftz.f32 %f5267, %f4316, %f5266; $L__BB105_484: rcp.approx.ftz.f32 %f4317, %f1212; mov.f32 %f4318, 0f3F800000; mul.ftz.f32 %f4319, %f4317, %f5267; sub.ftz.f32 %f4320, %f4318, %f4319; fma.rn.ftz.f32 %f5271, %f1211, %f4320, %f4319; fma.rn.ftz.f32 %f5272, %f1210, %f4320, %f4319; fma.rn.ftz.f32 %f5273, %f1209, %f4320, %f4319; $L__BB105_485: setp.ne.s32 %p457, %r5, 12; @%p457 bra $L__BB105_487; ld.local.u64 %rd81, [%rd3+36]; ld.f32 %f4321, [%rd81+36]; mul.ftz.f32 %f4322, %f4321, 0f3E800000; mul.ftz.f32 %f5273, %f5273, %f4322; mul.ftz.f32 %f5272, %f5272, %f4322; mul.ftz.f32 %f5271, %f5271, %f4322; $L__BB105_487: mul.ftz.f32 %f4323, %f1202, %f5273; mul.ftz.f32 %f4324, %f1202, %f5272; mul.ftz.f32 %f4325, %f1202, %f5271; mul.ftz.f32 %f5294, %f1203, %f4325; mul.ftz.f32 %f5295, %f1203, %f4324; mul.ftz.f32 %f5296, %f1203, %f4323; mul.ftz.f32 %f5297, %f5264, %f1203; bra.uni $L__BB105_524; $L__BB105_98: setp.lt.ftz.f32 %p83, %f89, 0f00000000; @%p83 bra $L__BB105_106; mul.ftz.f32 %f1650, %f5019, %f88; fma.rn.ftz.f32 %f1651, %f5018, %f87, %f1650; fma.rn.ftz.f32 %f91, %f5020, %f89, %f1651; abs.ftz.f32 %f92, %f91; fma.rn.ftz.f32 %f93, %f92, %f92, %f37; setp.leu.ftz.f32 %p84, %f93, 0f00000000; mov.f32 %f1649, 0f3F800000; mov.f32 %f5026, %f1649; @%p84 bra $L__BB105_101; sqrt.approx.ftz.f32 %f1652, %f93; sub.ftz.f32 %f1653, %f1652, %f92; add.ftz.f32 %f1654, %f92, %f1652; div.approx.ftz.f32 %f1655, %f1653, %f1654; fma.rn.ftz.f32 %f1656, %f92, %f1654, 0fBF800000; fma.rn.ftz.f32 %f1657, %f92, %f1653, 0f3F800000; div.approx.ftz.f32 %f1658, %f1656, %f1657; mul.ftz.f32 %f1659, %f1655, 0f3F000000; mul.ftz.f32 %f1660, %f1655, %f1659; fma.rn.ftz.f32 %f1661, %f1658, %f1658, 0f3F800000; mul.ftz.f32 %f5026, %f1660, %f1661; $L__BB105_101: mov.f32 %f1662, 0f00000000; max.ftz.f32 %f1663, %f1662, %f91; mul.ftz.f32 %f1664, %f5026, %f1663; mul.ftz.f32 %f1665, %f89, %f89; sub.ftz.f32 %f1667, %f1649, %f1665; fma.rn.ftz.f32 %f1668, %f33, %f1665, %f1667; mul.ftz.f32 %f1669, %f1668, 0f40490FDB; mul.ftz.f32 %f1670, %f1668, %f1669; max.ftz.f32 %f1672, %f1670, %f1507; div.approx.ftz.f32 %f1673, %f33, %f1672; mul.ftz.f32 %f1674, %f1664, %f1673; mul.ftz.f32 %f1675, %f1674, 0f3E800000; mul.ftz.f32 %f1676, %f90, %f91; div.approx.ftz.f32 %f5029, %f1675, %f1676; $L__BB105_106: neg.ftz.f32 %f113, %f5112; @%p61 bra $L__BB105_108; bra.uni $L__BB105_107; $L__BB105_108: add.ftz.f32 %f1730, %f5025, 0f3F800000; { // callseq 82, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f113; .param .b32 param1; st.param.f32 [param1+0], %f1730; .param .b32 retval0; call.uni (retval0), _Z4betaff, ( param0, param1 ); ld.param.f32 %f1731, [retval0+0]; } // callseq 82 mul.ftz.f32 %f5030, %f1731, %f113; bra.uni $L__BB105_109; $L__BB105_107: sub.ftz.f32 %f1729, %f5025, %f5112; div.approx.ftz.f32 %f5030, %f113, %f1729; $L__BB105_109: mul.ftz.f32 %f5295, %f5029, %f5030; mul.ftz.f32 %f5296, %f5029, %f5030; mul.ftz.f32 %f5294, %f5029, %f5030; mov.f32 %f5055, 0f3F800000; abs.ftz.f32 %f120, %f5055; fma.rn.ftz.f32 %f121, %f120, %f120, %f37; setp.leu.ftz.f32 %p89, %f121, 0f00000000; mov.f32 %f5031, %f5055; @%p89 bra $L__BB105_111; sqrt.approx.ftz.f32 %f1733, %f121; sub.ftz.f32 %f1734, %f1733, %f120; add.ftz.f32 %f1735, %f120, %f1733; div.approx.ftz.f32 %f1736, %f1734, %f1735; fma.rn.ftz.f32 %f1737, %f120, %f1735, 0fBF800000; fma.rn.ftz.f32 %f1738, %f120, %f1734, 0f3F800000; div.approx.ftz.f32 %f1739, %f1737, %f1738; mul.ftz.f32 %f1740, %f1736, 0f3F000000; mul.ftz.f32 %f1741, %f1736, %f1740; fma.rn.ftz.f32 %f1742, %f1739, %f1739, 0f3F800000; mul.ftz.f32 %f5031, %f1741, %f1742; $L__BB105_111: mov.f32 %f4859, 0f3F800000; setp.ne.s32 %p90, %r5, 27; or.pred %p3, %p61, %p90; sub.ftz.f32 %f127, %f4859, %f5031; rcp.approx.ftz.f32 %f128, %f127; mov.f32 %f5109, %f4859; mov.f32 %f5110, %f4859; mov.f32 %f5111, %f4859; @%p3 bra $L__BB105_115; add.ftz.f32 %f4960, %f5015, %f5018; div.approx.ftz.f32 %f4959, %f4960, %f1644; add.ftz.f32 %f4958, %f5016, %f5019; div.approx.ftz.f32 %f4957, %f4958, %f1644; mov.f32 %f4861, 0f3F800000; sub.ftz.f32 %f5032, %f4859, %f5031; mul.ftz.f32 %f1746, %f5019, %f4957; fma.rn.ftz.f32 %f1747, %f5018, %f4959, %f1746; fma.rn.ftz.f32 %f1748, %f5020, %f89, %f1747; abs.ftz.f32 %f129, %f1748; fma.rn.ftz.f32 %f130, %f129, %f129, %f37; setp.leu.ftz.f32 %p92, %f130, 0f00000000; @%p92 bra $L__BB105_114; sqrt.approx.ftz.f32 %f1749, %f130; sub.ftz.f32 %f1750, %f1749, %f129; add.ftz.f32 %f1751, %f129, %f1749; div.approx.ftz.f32 %f1752, %f1750, %f1751; fma.rn.ftz.f32 %f1753, %f129, %f1751, 0fBF800000; fma.rn.ftz.f32 %f1754, %f129, %f1750, 0f3F800000; div.approx.ftz.f32 %f1755, %f1753, %f1754; mul.ftz.f32 %f1756, %f1752, 0f3F000000; mul.ftz.f32 %f1757, %f1752, %f1756; fma.rn.ftz.f32 %f1758, %f1755, %f1755, 0f3F800000; mul.ftz.f32 %f1759, %f1757, %f1758; sub.ftz.f32 %f5032, %f1759, %f5031; $L__BB105_114: mul.ftz.f32 %f1760, %f128, %f5032; mov.f32 %f1761, 0f3F800000; sub.ftz.f32 %f1762, %f1761, %f1760; fma.rn.ftz.f32 %f5109, %f5012, %f1762, %f1760; fma.rn.ftz.f32 %f5110, %f5013, %f1762, %f1760; fma.rn.ftz.f32 %f5111, %f5014, %f1762, %f1760; mul.ftz.f32 %f5296, %f5296, %f5109; mul.ftz.f32 %f5295, %f5295, %f5110; mul.ftz.f32 %f5294, %f5294, %f5111; $L__BB105_115: neg.ftz.f32 %f5108, %f5020; neg.ftz.f32 %f5106, %f5018; neg.ftz.f32 %f5107, %f5019; rcp.approx.ftz.f32 %f145, %f20; fma.rn.ftz.f32 %f152, %f145, %f145, 0fBF800000; ld.local.u32 %r162, [%rd1+-4]; mov.f32 %f5053, 0f00000000; mov.u32 %r161, 0; mov.u16 %rs40, %rs27; mov.f32 %f5054, %f5055; $L__BB105_116: mov.u32 %r8, %r161; mad.lo.s32 %r162, %r162, 1103515245, 12345; cvt.rn.f32.u32 %f1769, %r162; mul.ftz.f32 %f167, %f1769, 0f2F800000; setp.gt.ftz.f32 %p95, %f5108, 0f3F7FF972; mov.pred %p94, -1; mov.pred %p539, %p94; @%p95 bra $L__BB105_126; bra.uni $L__BB105_117; $L__BB105_126: @%p539 bra $L__BB105_207; mad.lo.s32 %r52, %r162, 1103515245, 12345; cvt.rn.f32.u32 %f1789, %r52; mul.ftz.f32 %f178, %f1789, 0f2F800000; mad.lo.s32 %r162, %r52, 1103515245, 12345; cvt.rn.f32.u32 %f1790, %r162; mul.ftz.f32 %f179, %f1790, 0f2F800000; neg.ftz.f32 %f180, %f5106; mul.ftz.f32 %f1791, %f1423, %f180; neg.ftz.f32 %f181, %f5107; mul.ftz.f32 %f1792, %f1424, %f181; mul.ftz.f32 %f1793, %f1792, %f1792; fma.rn.ftz.f32 %f1794, %f1791, %f1791, %f1793; fma.rn.ftz.f32 %f1795, %f5108, %f5108, %f1794; sqrt.approx.ftz.f32 %f1796, %f1795; div.approx.ftz.f32 %f5058, %f1791, %f1796; div.approx.ftz.f32 %f5059, %f1792, %f1796; neg.ftz.f32 %f184, %f5108; div.approx.ftz.f32 %f185, %f184, %f1796; setp.gt.ftz.f32 %p108, %f185, 0f3F7FF972; @%p108 bra $L__BB105_133; neg.ftz.f32 %f4896, %f5108; div.approx.ftz.f32 %f4895, %f4896, %f1796; abs.ftz.f32 %f1797, %f4895; setp.lt.ftz.f32 %p109, %f1797, 0f358637BD; @%p109 bra $L__BB105_133; bra.uni $L__BB105_129; $L__BB105_133: mov.f32 %f1841, 0f3F800000; sub.ftz.f32 %f1842, %f1841, %f179; max.ftz.f32 %f1844, %f1842, %f1507; div.approx.ftz.f32 %f1845, %f179, %f1844; sqrt.approx.ftz.f32 %f1846, %f1845; mul.ftz.f32 %f1847, %f178, 0f40C90FDB; cos.approx.ftz.f32 %f1848, %f1847; sin.approx.ftz.f32 %f1849, %f1847; mul.ftz.f32 %f5057, %f1846, %f1849; mul.ftz.f32 %f5056, %f1846, %f1848; $L__BB105_134: mul.ftz.f32 %f1850, %f5059, %f5059; fma.rn.ftz.f32 %f1851, %f5058, %f5058, %f1850; add.ftz.f32 %f1852, %f1851, 0f00000000; sqrt.approx.ftz.f32 %f201, %f1852; setp.neu.ftz.f32 %p117, %f201, 0f00000000; @%p117 bra $L__BB105_135; bra.uni $L__BB105_136; $L__BB105_135: rcp.approx.ftz.f32 %f1853, %f201; mul.ftz.f32 %f5059, %f5059, %f1853; mul.ftz.f32 %f5058, %f5058, %f1853; $L__BB105_136: setp.eq.s32 %p523, %r5, 27; and.pred %p522, %p523, %p1; mul.ftz.f32 %f1854, %f5056, %f5058; mul.ftz.f32 %f1855, %f5057, %f5059; sub.ftz.f32 %f1856, %f1854, %f1855; mul.ftz.f32 %f1857, %f1423, %f1856; mul.ftz.f32 %f1858, %f5057, %f5058; fma.rn.ftz.f32 %f1859, %f5056, %f5059, %f1858; mul.ftz.f32 %f1860, %f1424, %f1859; neg.ftz.f32 %f1861, %f1857; neg.ftz.f32 %f1862, %f1860; mul.ftz.f32 %f1863, %f1860, %f1860; fma.rn.ftz.f32 %f1864, %f1857, %f1857, %f1863; add.ftz.f32 %f1865, %f1864, 0f3F800000; mov.f32 %f1866, 0f3F800000; sqrt.approx.ftz.f32 %f1867, %f1865; div.approx.ftz.f32 %f211, %f1861, %f1867; div.approx.ftz.f32 %f212, %f1862, %f1867; div.approx.ftz.f32 %f213, %f1866, %f1867; setp.ne.s32 %p118, %r8, 0; not.pred %p119, %p522; or.pred %p120, %p119, %p118; @%p120 bra $L__BB105_161; and.b16 %rs20, %rs40, 255; mov.f32 %f5063, 0f00000000; setp.eq.s16 %p121, %rs20, 0; @%p121 bra $L__BB105_148; @%p95 bra $L__BB105_158; mul.ftz.f32 %f1871, %f5108, %f5112; setp.lt.ftz.f32 %p124, %f5108, 0fBF7FF972; selp.f32 %f214, 0f3F800000, %f1871, %p124; @%p61 bra $L__BB105_144; bra.uni $L__BB105_140; $L__BB105_144: mul.ftz.f32 %f4870, %f20, %f5017; mul.ftz.f32 %f4869, %f20, %f5016; mul.ftz.f32 %f4868, %f20, %f5015; neg.ftz.f32 %f4867, %f5108; neg.ftz.f32 %f4866, %f5106; neg.ftz.f32 %f4865, %f5107; sub.ftz.f32 %f1913, %f4868, %f5106; sub.ftz.f32 %f1914, %f4869, %f5107; mul.ftz.f32 %f1915, %f1914, %f1914; fma.rn.ftz.f32 %f1916, %f1913, %f1913, %f1915; sub.ftz.f32 %f1917, %f4870, %f5108; fma.rn.ftz.f32 %f1918, %f1917, %f1917, %f1916; sqrt.approx.ftz.f32 %f1919, %f1918; div.approx.ftz.f32 %f1920, %f1913, %f1919; div.approx.ftz.f32 %f1921, %f1914, %f1919; div.approx.ftz.f32 %f1922, %f1917, %f1919; setp.lt.ftz.f32 %p127, %f1922, 0f00000000; neg.ftz.f32 %f1923, %f1920; neg.ftz.f32 %f1924, %f1921; neg.ftz.f32 %f1925, %f1922; selp.f32 %f1926, %f1923, %f1920, %p127; selp.f32 %f1927, %f1924, %f1921, %p127; selp.f32 %f224, %f1925, %f1922, %p127; mul.ftz.f32 %f1928, %f1927, %f4865; fma.rn.ftz.f32 %f1929, %f1926, %f4866, %f1928; fma.rn.ftz.f32 %f225, %f224, %f4867, %f1929; mul.ftz.f32 %f1930, %f5016, %f1927; fma.rn.ftz.f32 %f1931, %f5015, %f1926, %f1930; fma.rn.ftz.f32 %f226, %f5017, %f224, %f1931; setp.lt.ftz.f32 %p128, %f225, 0f00000000; @%p128 bra $L__BB105_158; abs.ftz.f32 %f228, %f225; fma.rn.ftz.f32 %f229, %f228, %f228, %f37; setp.leu.ftz.f32 %p129, %f229, 0f00000000; mov.f32 %f1932, 0f3F800000; mov.f32 %f5061, %f1932; @%p129 bra $L__BB105_147; sqrt.approx.ftz.f32 %f1933, %f229; sub.ftz.f32 %f1934, %f1933, %f228; add.ftz.f32 %f1935, %f228, %f1933; div.approx.ftz.f32 %f1936, %f1934, %f1935; fma.rn.ftz.f32 %f1937, %f228, %f1935, 0fBF800000; fma.rn.ftz.f32 %f1938, %f228, %f1934, 0f3F800000; div.approx.ftz.f32 %f1939, %f1937, %f1938; mul.ftz.f32 %f1940, %f1936, 0f3F000000; mul.ftz.f32 %f1941, %f1936, %f1940; fma.rn.ftz.f32 %f1942, %f1939, %f1939, 0f3F800000; mul.ftz.f32 %f5061, %f1941, %f1942; $L__BB105_147: fma.rn.ftz.f32 %f4911, %f20, %f226, %f225; mov.f32 %f1943, 0f00000000; max.ftz.f32 %f1944, %f1943, %f225; sub.ftz.f32 %f1946, %f1932, %f5061; mul.ftz.f32 %f1947, %f1946, %f1944; neg.ftz.f32 %f1948, %f226; max.ftz.f32 %f1949, %f1943, %f1948; mul.ftz.f32 %f1950, %f1947, %f1949; mul.ftz.f32 %f1951, %f224, %f224; sub.ftz.f32 %f1952, %f1932, %f1951; fma.rn.ftz.f32 %f1953, %f33, %f1951, %f1952; mul.ftz.f32 %f1954, %f1953, 0f40490FDB; mul.ftz.f32 %f1955, %f1953, %f1954; max.ftz.f32 %f1957, %f1955, %f1507; div.approx.ftz.f32 %f1958, %f33, %f1957; mul.ftz.f32 %f1959, %f1950, %f1958; mul.ftz.f32 %f1960, %f214, %f4911; mul.ftz.f32 %f1961, %f4911, %f1960; div.approx.ftz.f32 %f5063, %f1959, %f1961; bra.uni $L__BB105_158; $L__BB105_117: setp.lt.ftz.f32 %p96, %f5108, 0fBF7FF972; @%p96 bra $L__BB105_124; bra.uni $L__BB105_118; $L__BB105_124: mul.ftz.f32 %f5054, %f5054, %f167; cvt.ftz.sat.f32.f32 %f1786, %f5054; fma.rn.ftz.f32 %f5055, %f1786, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p106, %f5108, 0f3727C5AC; mov.pred %p539, 0; mov.f32 %f5053, 0f00000000; @%p106 bra $L__BB105_126; lg2.approx.ftz.f32 %f1787, %f5054; mul.ftz.f32 %f1788, %f5112, %f1787; ex2.approx.ftz.f32 %f5053, %f1788; bra.uni $L__BB105_126; $L__BB105_118: abs.ftz.f32 %f1770, %f5108; setp.ltu.ftz.f32 %p98, %f1770, 0f38D1B717; mov.pred %p97, 0; mov.pred %p539, %p97; @%p98 bra $L__BB105_126; mov.f32 %f5052, 0f3F800000; sub.ftz.f32 %f1772, %f5052, %f5053; setp.lt.ftz.f32 %p100, %f1772, %f167; mov.pred %p539, %p94; @%p100 bra $L__BB105_126; setp.ge.ftz.f32 %p101, %f5112, 0f00000000; @%p101 bra $L__BB105_122; mov.f32 %f1774, 0f3F800000; sub.ftz.f32 %f1775, %f1774, %f167; lg2.approx.ftz.f32 %f1776, %f1775; mov.f32 %f1777, 0fBF800000; div.approx.ftz.f32 %f1778, %f1777, %f5112; mul.ftz.f32 %f1779, %f1778, %f1776; ex2.approx.ftz.f32 %f1780, %f1779; mul.ftz.f32 %f5052, %f5054, %f1780; $L__BB105_122: cvt.ftz.sat.f32.f32 %f1782, %f5052; fma.rn.ftz.f32 %f5055, %f1782, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p103, %f5108, 0f3727C5AC; mov.f32 %f5053, 0f00000000; mov.f32 %f5054, %f5052; mov.pred %p539, %p97; @%p103 bra $L__BB105_126; lg2.approx.ftz.f32 %f1783, %f5052; mul.ftz.f32 %f1784, %f5112, %f1783; ex2.approx.ftz.f32 %f5053, %f1784; mov.f32 %f5054, %f5052; mov.pred %p539, %p97; bra.uni $L__BB105_126; $L__BB105_129: neg.ftz.f32 %f4898, %f5108; div.approx.ftz.f32 %f4897, %f4898, %f1796; add.ftz.f32 %f1800, %f4897, 0f3F800000; mul.ftz.f32 %f186, %f1800, 0f3F000000; setp.lt.ftz.f32 %p110, %f186, 0f38D1B717; mov.f32 %f5056, 0f00000000; mov.f32 %f5057, %f5056; @%p110 bra $L__BB105_134; neg.ftz.f32 %f4900, %f5108; div.approx.ftz.f32 %f4899, %f4900, %f1796; add.ftz.f32 %f1803, %f179, %f179; mul.ftz.f32 %f1804, %f1803, %f186; div.approx.ftz.f32 %f1805, %f1804, %f4899; add.ftz.f32 %f187, %f1805, 0fBF800000; mul.ftz.f32 %f188, %f187, %f187; add.ftz.f32 %f189, %f188, 0fBF800000; abs.ftz.f32 %f1806, %f189; setp.lt.ftz.f32 %p111, %f1806, 0f33D6BF95; mov.f32 %f5057, %f5056; @%p111 bra $L__BB105_134; neg.ftz.f32 %f4902, %f5108; div.approx.ftz.f32 %f4901, %f4902, %f1796; mul.ftz.f32 %f1807, %f4901, %f4901; mov.f32 %f1808, 0f3F800000; sub.ftz.f32 %f1809, %f1808, %f1807; mov.f32 %f1810, 0f00000000; max.ftz.f32 %f1811, %f1809, %f1810; sqrt.approx.ftz.f32 %f1812, %f1811; div.approx.ftz.f32 %f1813, %f1812, %f4901; rcp.approx.ftz.f32 %f1814, %f189; mul.ftz.f32 %f1815, %f1813, %f1813; mul.ftz.f32 %f1816, %f1815, %f1814; mul.ftz.f32 %f1817, %f1814, %f1816; sub.ftz.f32 %f1818, %f188, %f1815; mul.ftz.f32 %f1819, %f1818, %f1814; sub.ftz.f32 %f1820, %f1817, %f1819; max.ftz.f32 %f1821, %f1820, %f1810; sqrt.approx.ftz.f32 %f1822, %f1821; mul.ftz.f32 %f1823, %f1813, %f1814; add.ftz.f32 %f1824, %f1823, %f1822; rcp.approx.ftz.f32 %f1825, %f1813; setp.gt.ftz.f32 %p112, %f1824, %f1825; setp.lt.ftz.f32 %p113, %f187, 0f00000000; or.pred %p114, %p113, %p112; sub.ftz.f32 %f1826, %f1823, %f1822; mov.f32 %f1827, 0f3F000000; sub.ftz.f32 %f1828, %f1827, %f178; add.ftz.f32 %f1829, %f178, 0fBF000000; setp.ge.ftz.f32 %p115, %f178, 0f3F000000; selp.f32 %f1830, %f1829, %f1828, %p115; add.ftz.f32 %f1831, %f1830, %f1830; fma.rn.ftz.f32 %f1832, %f1831, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f1833, %f1831, %f1832, 0f3EED4413; mul.ftz.f32 %f1834, %f1831, %f1833; fma.rn.ftz.f32 %f1835, %f1831, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f1836, %f1831, %f1835, 0fBF800000; fma.rn.ftz.f32 %f1837, %f1831, %f1836, 0f3F191676; div.approx.ftz.f32 %f1838, %f1834, %f1837; selp.f32 %f5056, %f1826, %f1824, %p114; fma.rn.ftz.f32 %f1839, %f5056, %f5056, 0f3F800000; sqrt.approx.ftz.f32 %f1840, %f1839; mul.ftz.f32 %f5057, %f1838, %f1840; setp.ltu.ftz.f32 %p116, %f178, 0f3F000000; @%p116 bra $L__BB105_132; bra.uni $L__BB105_134; $L__BB105_132: neg.ftz.f32 %f5057, %f5057; bra.uni $L__BB105_134; $L__BB105_148: @%p95 bra $L__BB105_158; mul.ftz.f32 %f1965, %f5108, %f5112; setp.lt.ftz.f32 %p132, %f5108, 0fBF7FF972; selp.f32 %f243, 0f3F800000, %f1965, %p132; @%p61 bra $L__BB105_154; bra.uni $L__BB105_150; $L__BB105_154: neg.ftz.f32 %f4883, %f5017; neg.ftz.f32 %f4882, %f5016; neg.ftz.f32 %f4881, %f5015; sub.ftz.f32 %f2025, %f4881, %f5106; sub.ftz.f32 %f2027, %f4882, %f5107; sub.ftz.f32 %f2029, %f4883, %f5108; mul.ftz.f32 %f2030, %f2027, %f2027; fma.rn.ftz.f32 %f2031, %f2025, %f2025, %f2030; fma.rn.ftz.f32 %f2032, %f2029, %f2029, %f2031; sqrt.approx.ftz.f32 %f2033, %f2032; div.approx.ftz.f32 %f255, %f2029, %f2033; setp.lt.ftz.f32 %p136, %f255, 0f00000000; @%p136 bra $L__BB105_158; neg.ftz.f32 %f4921, %f5017; sub.ftz.f32 %f4920, %f4921, %f5108; div.approx.ftz.f32 %f4919, %f4920, %f2033; neg.ftz.f32 %f4918, %f5015; sub.ftz.f32 %f4917, %f4918, %f5106; div.approx.ftz.f32 %f4916, %f4917, %f2033; neg.ftz.f32 %f4915, %f5016; sub.ftz.f32 %f4914, %f4915, %f5107; div.approx.ftz.f32 %f4913, %f4914, %f2033; neg.ftz.f32 %f4886, %f5108; neg.ftz.f32 %f4885, %f5106; neg.ftz.f32 %f4884, %f5107; mul.ftz.f32 %f2035, %f4913, %f4884; fma.rn.ftz.f32 %f2036, %f4916, %f4885, %f2035; fma.rn.ftz.f32 %f256, %f4919, %f4886, %f2036; abs.ftz.f32 %f257, %f256; fma.rn.ftz.f32 %f258, %f257, %f257, %f152; setp.leu.ftz.f32 %p137, %f258, 0f00000000; mov.f32 %f2034, 0f3F800000; mov.f32 %f5065, %f2034; @%p137 bra $L__BB105_157; sqrt.approx.ftz.f32 %f2037, %f258; sub.ftz.f32 %f2038, %f2037, %f257; add.ftz.f32 %f2039, %f257, %f2037; div.approx.ftz.f32 %f2040, %f2038, %f2039; fma.rn.ftz.f32 %f2041, %f257, %f2039, 0fBF800000; fma.rn.ftz.f32 %f2042, %f257, %f2038, 0f3F800000; div.approx.ftz.f32 %f2043, %f2041, %f2042; mul.ftz.f32 %f2044, %f2040, 0f3F000000; mul.ftz.f32 %f2045, %f2040, %f2044; fma.rn.ftz.f32 %f2046, %f2043, %f2043, 0f3F800000; mul.ftz.f32 %f5065, %f2045, %f2046; $L__BB105_157: neg.ftz.f32 %f4924, %f5017; sub.ftz.f32 %f4923, %f4924, %f5108; div.approx.ftz.f32 %f4922, %f4923, %f2033; mov.f32 %f2047, 0f00000000; max.ftz.f32 %f2048, %f2047, %f256; mul.ftz.f32 %f2049, %f5065, %f2048; mul.ftz.f32 %f2050, %f4922, %f4922; sub.ftz.f32 %f2052, %f2034, %f2050; fma.rn.ftz.f32 %f2053, %f33, %f2050, %f2052; mul.ftz.f32 %f2054, %f2053, 0f40490FDB; mul.ftz.f32 %f2055, %f2053, %f2054; max.ftz.f32 %f2057, %f2055, %f1507; div.approx.ftz.f32 %f2058, %f33, %f2057; mul.ftz.f32 %f2059, %f2049, %f2058; mul.ftz.f32 %f2060, %f2059, 0f3E800000; mul.ftz.f32 %f2061, %f243, %f256; div.approx.ftz.f32 %f5063, %f2060, %f2061; bra.uni $L__BB105_158; $L__BB105_140: sub.ftz.f32 %f1875, %f5015, %f5106; sub.ftz.f32 %f1876, %f5016, %f5107; mul.ftz.f32 %f1877, %f1876, %f1876; fma.rn.ftz.f32 %f1878, %f1875, %f1875, %f1877; sub.ftz.f32 %f1879, %f5017, %f5108; fma.rn.ftz.f32 %f1880, %f1879, %f1879, %f1878; sqrt.approx.ftz.f32 %f1881, %f1880; div.approx.ftz.f32 %f217, %f1879, %f1881; setp.lt.ftz.f32 %p125, %f217, 0f00000000; @%p125 bra $L__BB105_158; sub.ftz.f32 %f4908, %f5017, %f5108; div.approx.ftz.f32 %f4907, %f4908, %f1881; sub.ftz.f32 %f4906, %f5015, %f5106; div.approx.ftz.f32 %f4905, %f4906, %f1881; sub.ftz.f32 %f4904, %f5016, %f5107; div.approx.ftz.f32 %f4903, %f4904, %f1881; neg.ftz.f32 %f4864, %f5108; neg.ftz.f32 %f4863, %f5106; neg.ftz.f32 %f4862, %f5107; mul.ftz.f32 %f1883, %f4903, %f4862; fma.rn.ftz.f32 %f1884, %f4905, %f4863, %f1883; fma.rn.ftz.f32 %f218, %f4907, %f4864, %f1884; abs.ftz.f32 %f219, %f218; fma.rn.ftz.f32 %f220, %f219, %f219, %f37; setp.leu.ftz.f32 %p126, %f220, 0f00000000; mov.f32 %f1882, 0f3F800000; mov.f32 %f5060, %f1882; @%p126 bra $L__BB105_143; sqrt.approx.ftz.f32 %f1885, %f220; sub.ftz.f32 %f1886, %f1885, %f219; add.ftz.f32 %f1887, %f219, %f1885; div.approx.ftz.f32 %f1888, %f1886, %f1887; fma.rn.ftz.f32 %f1889, %f219, %f1887, 0fBF800000; fma.rn.ftz.f32 %f1890, %f219, %f1886, 0f3F800000; div.approx.ftz.f32 %f1891, %f1889, %f1890; mul.ftz.f32 %f1892, %f1888, 0f3F000000; mul.ftz.f32 %f1893, %f1888, %f1892; fma.rn.ftz.f32 %f1894, %f1891, %f1891, 0f3F800000; mul.ftz.f32 %f5060, %f1893, %f1894; $L__BB105_143: sub.ftz.f32 %f4910, %f5017, %f5108; div.approx.ftz.f32 %f4909, %f4910, %f1881; mov.f32 %f1895, 0f00000000; max.ftz.f32 %f1896, %f1895, %f218; mul.ftz.f32 %f1897, %f5060, %f1896; mul.ftz.f32 %f1898, %f4909, %f4909; sub.ftz.f32 %f1900, %f1882, %f1898; fma.rn.ftz.f32 %f1901, %f33, %f1898, %f1900; mul.ftz.f32 %f1902, %f1901, 0f40490FDB; mul.ftz.f32 %f1903, %f1901, %f1902; max.ftz.f32 %f1905, %f1903, %f1507; div.approx.ftz.f32 %f1906, %f33, %f1905; mul.ftz.f32 %f1907, %f1897, %f1906; mul.ftz.f32 %f1908, %f1907, 0f3E800000; mul.ftz.f32 %f1909, %f214, %f218; div.approx.ftz.f32 %f5063, %f1908, %f1909; bra.uni $L__BB105_158; $L__BB105_150: neg.ftz.f32 %f4880, %f5015; neg.ftz.f32 %f4879, %f5017; mul.ftz.f32 %f4878, %f145, %f4879; neg.ftz.f32 %f4877, %f5016; mul.ftz.f32 %f4876, %f145, %f4877; neg.ftz.f32 %f4875, %f5015; mul.ftz.f32 %f4874, %f145, %f4875; neg.ftz.f32 %f4873, %f5108; neg.ftz.f32 %f4872, %f5106; neg.ftz.f32 %f4871, %f5107; sub.ftz.f32 %f1969, %f4874, %f5106; sub.ftz.f32 %f1970, %f4876, %f5107; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f4878, %f5108; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f1975, %f1974; div.approx.ftz.f32 %f1976, %f1969, %f1975; div.approx.ftz.f32 %f1977, %f1970, %f1975; div.approx.ftz.f32 %f1978, %f1973, %f1975; setp.lt.ftz.f32 %p133, %f1978, 0f00000000; neg.ftz.f32 %f1979, %f1976; neg.ftz.f32 %f1980, %f1977; neg.ftz.f32 %f1981, %f1978; selp.f32 %f1982, %f1979, %f1976, %p133; selp.f32 %f1983, %f1980, %f1977, %p133; selp.f32 %f244, %f1981, %f1978, %p133; mul.ftz.f32 %f1984, %f1983, %f4871; fma.rn.ftz.f32 %f1985, %f1982, %f4872, %f1984; fma.rn.ftz.f32 %f245, %f244, %f4873, %f1985; mul.ftz.f32 %f1987, %f1982, %f4875; mul.ftz.f32 %f1988, %f1983, %f5016; sub.ftz.f32 %f1989, %f1987, %f1988; mul.ftz.f32 %f1990, %f244, %f5017; sub.ftz.f32 %f246, %f1989, %f1990; setp.lt.ftz.f32 %p134, %f245, 0f00000000; @%p134 bra $L__BB105_158; abs.ftz.f32 %f248, %f245; fma.rn.ftz.f32 %f249, %f248, %f248, %f152; setp.leu.ftz.f32 %p135, %f249, 0f00000000; mov.f32 %f1991, 0f3F800000; mov.f32 %f5064, %f1991; @%p135 bra $L__BB105_153; sqrt.approx.ftz.f32 %f1992, %f249; sub.ftz.f32 %f1993, %f1992, %f248; add.ftz.f32 %f1994, %f248, %f1992; div.approx.ftz.f32 %f1995, %f1993, %f1994; fma.rn.ftz.f32 %f1996, %f248, %f1994, 0fBF800000; fma.rn.ftz.f32 %f1997, %f248, %f1993, 0f3F800000; div.approx.ftz.f32 %f1998, %f1996, %f1997; mul.ftz.f32 %f1999, %f1995, 0f3F000000; mul.ftz.f32 %f2000, %f1995, %f1999; fma.rn.ftz.f32 %f2001, %f1998, %f1998, 0f3F800000; mul.ftz.f32 %f5064, %f2000, %f2001; $L__BB105_153: fma.rn.ftz.f32 %f4912, %f145, %f246, %f245; mov.f32 %f2002, 0f00000000; max.ftz.f32 %f2003, %f2002, %f245; sub.ftz.f32 %f2005, %f1991, %f5064; mul.ftz.f32 %f2006, %f2005, %f2003; neg.ftz.f32 %f2007, %f246; max.ftz.f32 %f2008, %f2002, %f2007; mul.ftz.f32 %f2009, %f2006, %f2008; mul.ftz.f32 %f2010, %f244, %f244; sub.ftz.f32 %f2011, %f1991, %f2010; fma.rn.ftz.f32 %f2012, %f33, %f2010, %f2011; mul.ftz.f32 %f2013, %f2012, 0f40490FDB; mul.ftz.f32 %f2014, %f2012, %f2013; max.ftz.f32 %f2016, %f2014, %f1507; div.approx.ftz.f32 %f2017, %f33, %f2016; mul.ftz.f32 %f2018, %f2009, %f2017; mul.ftz.f32 %f2019, %f243, %f4912; mul.ftz.f32 %f2020, %f4912, %f2019; div.approx.ftz.f32 %f5063, %f2018, %f2020; $L__BB105_158: neg.ftz.f32 %f4828, %f5017; selp.f32 %f4827, %f4828, %f5017, %p61; setp.gt.ftz.f32 %p138, %f4827, 0f3F7FF972; setp.lt.ftz.f32 %p139, %f4827, 0f3727C5AC; or.pred %p140, %p138, %p139; selp.f32 %f5071, 0f3F800000, 0f00000000, %p138; @%p140 bra $L__BB105_160; selp.u16 %rs31, 1, 0, %p1; and.b16 %rs30, %rs40, 255; setp.eq.s16 %p141, %rs30, %rs31; neg.ftz.f32 %f2062, %f5055; selp.f32 %f2063, %f5055, %f2062, %p141; add.ftz.f32 %f2064, %f2063, 0f3F800000; mul.ftz.f32 %f2065, %f2064, 0f3F000000; cvt.ftz.sat.f32.f32 %f2066, %f2065; lg2.approx.ftz.f32 %f2067, %f2066; mul.ftz.f32 %f2068, %f5025, %f2067; ex2.approx.ftz.f32 %f5071, %f2068; $L__BB105_160: mul.ftz.f32 %f2069, %f5109, %f5063; mul.ftz.f32 %f5296, %f2069, %f5071; mul.ftz.f32 %f2070, %f5110, %f5063; mul.ftz.f32 %f5295, %f2070, %f5071; mul.ftz.f32 %f2071, %f5111, %f5063; mul.ftz.f32 %f5294, %f2071, %f5071; $L__BB105_161: setp.eq.s32 %p142, %r8, 0; @%p142 bra $L__BB105_186; and.b16 %rs22, %rs40, 255; mov.f32 %f5078, 0f00000000; setp.eq.s16 %p143, %rs22, 0; @%p143 bra $L__BB105_173; @%p95 bra $L__BB105_183; mul.ftz.f32 %f2075, %f5108, %f5112; setp.lt.ftz.f32 %p146, %f5108, 0fBF7FF972; selp.f32 %f284, 0f3F800000, %f2075, %p146; @%p61 bra $L__BB105_169; bra.uni $L__BB105_165; $L__BB105_169: mul.ftz.f32 %f4839, %f20, %f5017; mul.ftz.f32 %f4838, %f20, %f5016; mul.ftz.f32 %f4837, %f20, %f5015; neg.ftz.f32 %f4836, %f5108; neg.ftz.f32 %f4835, %f5106; neg.ftz.f32 %f4834, %f5107; sub.ftz.f32 %f2117, %f4837, %f5106; sub.ftz.f32 %f2118, %f4838, %f5107; mul.ftz.f32 %f2119, %f2118, %f2118; fma.rn.ftz.f32 %f2120, %f2117, %f2117, %f2119; sub.ftz.f32 %f2121, %f4839, %f5108; fma.rn.ftz.f32 %f2122, %f2121, %f2121, %f2120; sqrt.approx.ftz.f32 %f2123, %f2122; div.approx.ftz.f32 %f2124, %f2117, %f2123; div.approx.ftz.f32 %f2125, %f2118, %f2123; div.approx.ftz.f32 %f2126, %f2121, %f2123; setp.lt.ftz.f32 %p149, %f2126, 0f00000000; neg.ftz.f32 %f2127, %f2124; neg.ftz.f32 %f2128, %f2125; neg.ftz.f32 %f2129, %f2126; selp.f32 %f2130, %f2127, %f2124, %p149; selp.f32 %f2131, %f2128, %f2125, %p149; selp.f32 %f294, %f2129, %f2126, %p149; mul.ftz.f32 %f2132, %f2131, %f4834; fma.rn.ftz.f32 %f2133, %f2130, %f4835, %f2132; fma.rn.ftz.f32 %f295, %f294, %f4836, %f2133; mul.ftz.f32 %f2134, %f5016, %f2131; fma.rn.ftz.f32 %f2135, %f5015, %f2130, %f2134; fma.rn.ftz.f32 %f296, %f5017, %f294, %f2135; setp.lt.ftz.f32 %p150, %f295, 0f00000000; @%p150 bra $L__BB105_183; abs.ftz.f32 %f298, %f295; fma.rn.ftz.f32 %f299, %f298, %f298, %f37; setp.leu.ftz.f32 %p151, %f299, 0f00000000; mov.f32 %f2136, 0f3F800000; mov.f32 %f5076, %f2136; @%p151 bra $L__BB105_172; sqrt.approx.ftz.f32 %f2137, %f299; sub.ftz.f32 %f2138, %f2137, %f298; add.ftz.f32 %f2139, %f298, %f2137; div.approx.ftz.f32 %f2140, %f2138, %f2139; fma.rn.ftz.f32 %f2141, %f298, %f2139, 0fBF800000; fma.rn.ftz.f32 %f2142, %f298, %f2138, 0f3F800000; div.approx.ftz.f32 %f2143, %f2141, %f2142; mul.ftz.f32 %f2144, %f2140, 0f3F000000; mul.ftz.f32 %f2145, %f2140, %f2144; fma.rn.ftz.f32 %f2146, %f2143, %f2143, 0f3F800000; mul.ftz.f32 %f5076, %f2145, %f2146; $L__BB105_172: fma.rn.ftz.f32 %f4933, %f20, %f296, %f295; mov.f32 %f2147, 0f00000000; max.ftz.f32 %f2148, %f2147, %f295; sub.ftz.f32 %f2150, %f2136, %f5076; mul.ftz.f32 %f2151, %f2150, %f2148; neg.ftz.f32 %f2152, %f296; max.ftz.f32 %f2153, %f2147, %f2152; mul.ftz.f32 %f2154, %f2151, %f2153; mul.ftz.f32 %f2155, %f294, %f294; sub.ftz.f32 %f2156, %f2136, %f2155; fma.rn.ftz.f32 %f2157, %f33, %f2155, %f2156; mul.ftz.f32 %f2158, %f2157, 0f40490FDB; mul.ftz.f32 %f2159, %f2157, %f2158; max.ftz.f32 %f2161, %f2159, %f1507; div.approx.ftz.f32 %f2162, %f33, %f2161; mul.ftz.f32 %f2163, %f2154, %f2162; mul.ftz.f32 %f2164, %f284, %f4933; mul.ftz.f32 %f2165, %f4933, %f2164; div.approx.ftz.f32 %f5078, %f2163, %f2165; bra.uni $L__BB105_183; $L__BB105_173: @%p95 bra $L__BB105_183; mul.ftz.f32 %f2169, %f5108, %f5112; setp.lt.ftz.f32 %p154, %f5108, 0fBF7FF972; selp.f32 %f313, 0f3F800000, %f2169, %p154; @%p61 bra $L__BB105_179; bra.uni $L__BB105_175; $L__BB105_179: neg.ftz.f32 %f4854, %f5017; neg.ftz.f32 %f4853, %f5016; neg.ftz.f32 %f4852, %f5015; sub.ftz.f32 %f2229, %f4852, %f5106; sub.ftz.f32 %f2231, %f4853, %f5107; sub.ftz.f32 %f2233, %f4854, %f5108; mul.ftz.f32 %f2234, %f2231, %f2231; fma.rn.ftz.f32 %f2235, %f2229, %f2229, %f2234; fma.rn.ftz.f32 %f2236, %f2233, %f2233, %f2235; sqrt.approx.ftz.f32 %f2237, %f2236; div.approx.ftz.f32 %f325, %f2233, %f2237; setp.lt.ftz.f32 %p158, %f325, 0f00000000; @%p158 bra $L__BB105_183; neg.ftz.f32 %f4943, %f5017; sub.ftz.f32 %f4942, %f4943, %f5108; div.approx.ftz.f32 %f4941, %f4942, %f2237; neg.ftz.f32 %f4940, %f5015; sub.ftz.f32 %f4939, %f4940, %f5106; div.approx.ftz.f32 %f4938, %f4939, %f2237; neg.ftz.f32 %f4937, %f5016; sub.ftz.f32 %f4936, %f4937, %f5107; div.approx.ftz.f32 %f4935, %f4936, %f2237; neg.ftz.f32 %f4857, %f5108; neg.ftz.f32 %f4856, %f5106; neg.ftz.f32 %f4855, %f5107; mul.ftz.f32 %f2239, %f4935, %f4855; fma.rn.ftz.f32 %f2240, %f4938, %f4856, %f2239; fma.rn.ftz.f32 %f326, %f4941, %f4857, %f2240; abs.ftz.f32 %f327, %f326; fma.rn.ftz.f32 %f328, %f327, %f327, %f152; setp.leu.ftz.f32 %p159, %f328, 0f00000000; mov.f32 %f2238, 0f3F800000; mov.f32 %f5080, %f2238; @%p159 bra $L__BB105_182; sqrt.approx.ftz.f32 %f2241, %f328; sub.ftz.f32 %f2242, %f2241, %f327; add.ftz.f32 %f2243, %f327, %f2241; div.approx.ftz.f32 %f2244, %f2242, %f2243; fma.rn.ftz.f32 %f2245, %f327, %f2243, 0fBF800000; fma.rn.ftz.f32 %f2246, %f327, %f2242, 0f3F800000; div.approx.ftz.f32 %f2247, %f2245, %f2246; mul.ftz.f32 %f2248, %f2244, 0f3F000000; mul.ftz.f32 %f2249, %f2244, %f2248; fma.rn.ftz.f32 %f2250, %f2247, %f2247, 0f3F800000; mul.ftz.f32 %f5080, %f2249, %f2250; $L__BB105_182: neg.ftz.f32 %f4946, %f5017; sub.ftz.f32 %f4945, %f4946, %f5108; div.approx.ftz.f32 %f4944, %f4945, %f2237; mov.f32 %f2251, 0f00000000; max.ftz.f32 %f2252, %f2251, %f326; mul.ftz.f32 %f2253, %f5080, %f2252; mul.ftz.f32 %f2254, %f4944, %f4944; sub.ftz.f32 %f2256, %f2238, %f2254; fma.rn.ftz.f32 %f2257, %f33, %f2254, %f2256; mul.ftz.f32 %f2258, %f2257, 0f40490FDB; mul.ftz.f32 %f2259, %f2257, %f2258; max.ftz.f32 %f2261, %f2259, %f1507; div.approx.ftz.f32 %f2262, %f33, %f2261; mul.ftz.f32 %f2263, %f2253, %f2262; mul.ftz.f32 %f2264, %f2263, 0f3E800000; mul.ftz.f32 %f2265, %f313, %f326; div.approx.ftz.f32 %f5078, %f2264, %f2265; bra.uni $L__BB105_183; $L__BB105_165: sub.ftz.f32 %f2079, %f5015, %f5106; sub.ftz.f32 %f2080, %f5016, %f5107; mul.ftz.f32 %f2081, %f2080, %f2080; fma.rn.ftz.f32 %f2082, %f2079, %f2079, %f2081; sub.ftz.f32 %f2083, %f5017, %f5108; fma.rn.ftz.f32 %f2084, %f2083, %f2083, %f2082; sqrt.approx.ftz.f32 %f2085, %f2084; div.approx.ftz.f32 %f287, %f2083, %f2085; setp.lt.ftz.f32 %p147, %f287, 0f00000000; @%p147 bra $L__BB105_183; sub.ftz.f32 %f4930, %f5017, %f5108; div.approx.ftz.f32 %f4929, %f4930, %f2085; sub.ftz.f32 %f4928, %f5015, %f5106; div.approx.ftz.f32 %f4927, %f4928, %f2085; sub.ftz.f32 %f4926, %f5016, %f5107; div.approx.ftz.f32 %f4925, %f4926, %f2085; mov.f32 %f5075, 0f3F800000; neg.ftz.f32 %f4831, %f5108; neg.ftz.f32 %f4830, %f5106; neg.ftz.f32 %f4829, %f5107; mul.ftz.f32 %f2087, %f4925, %f4829; fma.rn.ftz.f32 %f2088, %f4927, %f4830, %f2087; fma.rn.ftz.f32 %f288, %f4929, %f4831, %f2088; abs.ftz.f32 %f289, %f288; fma.rn.ftz.f32 %f290, %f289, %f289, %f37; setp.leu.ftz.f32 %p148, %f290, 0f00000000; @%p148 bra $L__BB105_168; sqrt.approx.ftz.f32 %f2089, %f290; sub.ftz.f32 %f2090, %f2089, %f289; add.ftz.f32 %f2091, %f289, %f2089; div.approx.ftz.f32 %f2092, %f2090, %f2091; fma.rn.ftz.f32 %f2093, %f289, %f2091, 0fBF800000; fma.rn.ftz.f32 %f2094, %f289, %f2090, 0f3F800000; div.approx.ftz.f32 %f2095, %f2093, %f2094; mul.ftz.f32 %f2096, %f2092, 0f3F000000; mul.ftz.f32 %f2097, %f2092, %f2096; fma.rn.ftz.f32 %f2098, %f2095, %f2095, 0f3F800000; mul.ftz.f32 %f5075, %f2097, %f2098; $L__BB105_168: sub.ftz.f32 %f4932, %f5017, %f5108; div.approx.ftz.f32 %f4931, %f4932, %f2085; mov.f32 %f4833, 0f3F800000; mov.f32 %f2099, 0f00000000; max.ftz.f32 %f2100, %f2099, %f288; mul.ftz.f32 %f2101, %f5075, %f2100; mul.ftz.f32 %f2102, %f4931, %f4931; sub.ftz.f32 %f2104, %f4833, %f2102; fma.rn.ftz.f32 %f2105, %f33, %f2102, %f2104; mul.ftz.f32 %f2106, %f2105, 0f40490FDB; mul.ftz.f32 %f2107, %f2105, %f2106; max.ftz.f32 %f2109, %f2107, %f1507; div.approx.ftz.f32 %f2110, %f33, %f2109; mul.ftz.f32 %f2111, %f2101, %f2110; mul.ftz.f32 %f2112, %f2111, 0f3E800000; mul.ftz.f32 %f2113, %f284, %f288; div.approx.ftz.f32 %f5078, %f2112, %f2113; bra.uni $L__BB105_183; $L__BB105_175: neg.ftz.f32 %f4851, %f5015; neg.ftz.f32 %f4850, %f5017; mul.ftz.f32 %f4849, %f145, %f4850; neg.ftz.f32 %f4848, %f5016; mul.ftz.f32 %f4847, %f145, %f4848; neg.ftz.f32 %f4846, %f5015; mul.ftz.f32 %f4845, %f145, %f4846; neg.ftz.f32 %f4844, %f5108; neg.ftz.f32 %f4843, %f5106; neg.ftz.f32 %f4842, %f5107; sub.ftz.f32 %f2173, %f4845, %f5106; sub.ftz.f32 %f2174, %f4847, %f5107; mul.ftz.f32 %f2175, %f2174, %f2174; fma.rn.ftz.f32 %f2176, %f2173, %f2173, %f2175; sub.ftz.f32 %f2177, %f4849, %f5108; fma.rn.ftz.f32 %f2178, %f2177, %f2177, %f2176; sqrt.approx.ftz.f32 %f2179, %f2178; div.approx.ftz.f32 %f2180, %f2173, %f2179; div.approx.ftz.f32 %f2181, %f2174, %f2179; div.approx.ftz.f32 %f2182, %f2177, %f2179; setp.lt.ftz.f32 %p155, %f2182, 0f00000000; neg.ftz.f32 %f2183, %f2180; neg.ftz.f32 %f2184, %f2181; neg.ftz.f32 %f2185, %f2182; selp.f32 %f2186, %f2183, %f2180, %p155; selp.f32 %f2187, %f2184, %f2181, %p155; selp.f32 %f314, %f2185, %f2182, %p155; mul.ftz.f32 %f2188, %f2187, %f4842; fma.rn.ftz.f32 %f2189, %f2186, %f4843, %f2188; fma.rn.ftz.f32 %f315, %f314, %f4844, %f2189; mul.ftz.f32 %f2191, %f2186, %f4846; mul.ftz.f32 %f2192, %f2187, %f5016; sub.ftz.f32 %f2193, %f2191, %f2192; mul.ftz.f32 %f2194, %f314, %f5017; sub.ftz.f32 %f316, %f2193, %f2194; setp.lt.ftz.f32 %p156, %f315, 0f00000000; @%p156 bra $L__BB105_183; abs.ftz.f32 %f318, %f315; fma.rn.ftz.f32 %f319, %f318, %f318, %f152; setp.leu.ftz.f32 %p157, %f319, 0f00000000; mov.f32 %f2195, 0f3F800000; mov.f32 %f5079, %f2195; @%p157 bra $L__BB105_178; sqrt.approx.ftz.f32 %f2196, %f319; sub.ftz.f32 %f2197, %f2196, %f318; add.ftz.f32 %f2198, %f318, %f2196; div.approx.ftz.f32 %f2199, %f2197, %f2198; fma.rn.ftz.f32 %f2200, %f318, %f2198, 0fBF800000; fma.rn.ftz.f32 %f2201, %f318, %f2197, 0f3F800000; div.approx.ftz.f32 %f2202, %f2200, %f2201; mul.ftz.f32 %f2203, %f2199, 0f3F000000; mul.ftz.f32 %f2204, %f2199, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, 0f3F800000; mul.ftz.f32 %f5079, %f2204, %f2205; $L__BB105_178: fma.rn.ftz.f32 %f4934, %f145, %f316, %f315; mov.f32 %f2206, 0f00000000; max.ftz.f32 %f2207, %f2206, %f315; sub.ftz.f32 %f2209, %f2195, %f5079; mul.ftz.f32 %f2210, %f2209, %f2207; neg.ftz.f32 %f2211, %f316; max.ftz.f32 %f2212, %f2206, %f2211; mul.ftz.f32 %f2213, %f2210, %f2212; mul.ftz.f32 %f2214, %f314, %f314; sub.ftz.f32 %f2215, %f2195, %f2214; fma.rn.ftz.f32 %f2216, %f33, %f2214, %f2215; mul.ftz.f32 %f2217, %f2216, 0f40490FDB; mul.ftz.f32 %f2218, %f2216, %f2217; max.ftz.f32 %f2220, %f2218, %f1507; div.approx.ftz.f32 %f2221, %f33, %f2220; mul.ftz.f32 %f2222, %f2213, %f2221; mul.ftz.f32 %f2223, %f313, %f4934; mul.ftz.f32 %f2224, %f4934, %f2223; div.approx.ftz.f32 %f5078, %f2222, %f2224; $L__BB105_183: neg.ftz.f32 %f4841, %f5017; selp.f32 %f4840, %f4841, %f5017, %p61; setp.gt.ftz.f32 %p160, %f4840, 0f3F7FF972; setp.lt.ftz.f32 %p161, %f4840, 0f3727C5AC; or.pred %p162, %p160, %p161; selp.f32 %f5086, 0f3F800000, 0f00000000, %p160; @%p162 bra $L__BB105_185; and.b16 %rs33, %rs40, 255; selp.u16 %rs32, 1, 0, %p1; setp.eq.s16 %p163, %rs33, %rs32; neg.ftz.f32 %f2266, %f5055; selp.f32 %f2267, %f5055, %f2266, %p163; add.ftz.f32 %f2268, %f2267, 0f3F800000; mul.ftz.f32 %f2269, %f2268, 0f3F000000; cvt.ftz.sat.f32.f32 %f2270, %f2269; lg2.approx.ftz.f32 %f2271, %f2270; mul.ftz.f32 %f2272, %f5025, %f2271; ex2.approx.ftz.f32 %f5086, %f2272; $L__BB105_185: mul.ftz.f32 %f2273, %f5109, %f5078; fma.rn.ftz.f32 %f5296, %f2273, %f5086, %f5296; mul.ftz.f32 %f2274, %f5110, %f5078; fma.rn.ftz.f32 %f5295, %f2274, %f5086, %f5295; mul.ftz.f32 %f2275, %f5111, %f5078; fma.rn.ftz.f32 %f5294, %f2275, %f5086, %f5294; $L__BB105_186: add.s32 %r161, %r8, 1; setp.gt.u32 %p164, %r161, 9; @%p164 bra $L__BB105_206; neg.ftz.f32 %f4889, %f5108; neg.ftz.f32 %f4888, %f5106; neg.ftz.f32 %f4887, %f5107; mad.lo.s32 %r162, %r162, 1103515245, 12345; and.b16 %rs24, %rs40, 255; setp.eq.s16 %p165, %rs24, 0; selp.f32 %f354, %f145, %f20, %p165; mul.ftz.f32 %f2277, %f212, %f4887; fma.rn.ftz.f32 %f2278, %f211, %f4888, %f2277; fma.rn.ftz.f32 %f355, %f213, %f4889, %f2278; abs.ftz.f32 %f356, %f355; fma.rn.ftz.f32 %f2279, %f354, %f354, 0fBF800000; mov.f32 %f5090, 0f3F800000; mul.ftz.f32 %f357, %f356, %f356; add.ftz.f32 %f358, %f2279, %f357; setp.leu.ftz.f32 %p166, %f358, 0f00000000; @%p166 bra $L__BB105_189; sqrt.approx.ftz.f32 %f2280, %f358; sub.ftz.f32 %f2281, %f2280, %f356; add.ftz.f32 %f2282, %f356, %f2280; div.approx.ftz.f32 %f2283, %f2281, %f2282; fma.rn.ftz.f32 %f2284, %f356, %f2282, 0fBF800000; fma.rn.ftz.f32 %f2285, %f356, %f2281, 0f3F800000; div.approx.ftz.f32 %f2286, %f2284, %f2285; mul.ftz.f32 %f2287, %f2283, 0f3F000000; mul.ftz.f32 %f2288, %f2283, %f2287; fma.rn.ftz.f32 %f2289, %f2286, %f2286, 0f3F800000; mul.ftz.f32 %f5090, %f2288, %f2289; $L__BB105_189: cvt.rn.f32.u32 %f2290, %r162; mul.ftz.f32 %f2291, %f2290, 0f2F800000; setp.gt.ftz.f32 %p167, %f5090, %f2291; @%p167 bra $L__BB105_191; bra.uni $L__BB105_190; $L__BB105_191: mov.f32 %f4952, 0f3F800000; div.approx.ftz.f32 %f4951, %f4952, %f1867; add.ftz.f32 %f2317, %f211, %f211; add.ftz.f32 %f2318, %f212, %f212; add.ftz.f32 %f2319, %f4951, %f4951; fma.rn.ftz.f32 %f5108, %f2319, %f355, %f5108; fma.rn.ftz.f32 %f5107, %f2318, %f355, %f5107; fma.rn.ftz.f32 %f5106, %f2317, %f355, %f5106; mov.pred %p540, 0; bra.uni $L__BB105_192; $L__BB105_190: mov.f32 %f4949, 0f3F800000; div.approx.ftz.f32 %f4948, %f4949, %f1867; and.b16 %rs34, %rs40, 255; setp.eq.s16 %p531, %rs34, 0; selp.f32 %f4947, %f145, %f20, %p531; neg.ftz.f32 %f4892, %f5108; neg.ftz.f32 %f4891, %f5106; neg.ftz.f32 %f4890, %f5107; rcp.approx.ftz.f32 %f2292, %f4947; mov.f32 %f2293, 0f3F800000; mul.ftz.f32 %f2294, %f355, %f355; sub.ftz.f32 %f2295, %f2293, %f2294; mul.ftz.f32 %f2296, %f2295, %f2292; mul.ftz.f32 %f2297, %f2292, %f2296; sub.ftz.f32 %f2298, %f2293, %f2297; mov.f32 %f2299, 0f00000000; max.ftz.f32 %f2300, %f2298, %f2299; sqrt.approx.ftz.f32 %f2301, %f2300; mul.ftz.f32 %f2302, %f355, %f2292; sub.ftz.f32 %f2303, %f2302, %f2301; mul.ftz.f32 %f2304, %f211, %f2303; mul.ftz.f32 %f2305, %f212, %f2303; mul.ftz.f32 %f2306, %f4948, %f2303; mul.ftz.f32 %f2307, %f2292, %f4891; mul.ftz.f32 %f2308, %f2292, %f4890; mul.ftz.f32 %f2309, %f2292, %f4892; sub.ftz.f32 %f2310, %f2304, %f2307; sub.ftz.f32 %f2311, %f2305, %f2308; sub.ftz.f32 %f2312, %f2306, %f2309; mul.ftz.f32 %f2313, %f2311, %f2311; fma.rn.ftz.f32 %f2314, %f2310, %f2310, %f2313; fma.rn.ftz.f32 %f2315, %f2312, %f2312, %f2314; sqrt.approx.ftz.f32 %f2316, %f2315; div.approx.ftz.f32 %f5108, %f2312, %f2316; div.approx.ftz.f32 %f5107, %f2311, %f2316; div.approx.ftz.f32 %f5106, %f2310, %f2316; mov.pred %p540, -1; $L__BB105_192: not.pred %p170, %p540; @%p170 bra $L__BB105_194; and.b16 %rs35, %rs40, 255; setp.eq.s16 %p532, %rs35, 0; selp.u16 %rs40, 1, 0, %p532; neg.ftz.f32 %f5106, %f5106; neg.ftz.f32 %f5107, %f5107; neg.ftz.f32 %f5108, %f5108; neg.ftz.f32 %f5055, %f5055; $L__BB105_194: setp.ne.s32 %p525, %r5, 27; or.pred %p524, %p61, %p525; or.pred %p173, %p524, %p170; @%p173 bra $L__BB105_196; bra.uni $L__BB105_195; $L__BB105_196: setp.eq.s32 %p530, %r8, 0; setp.eq.s32 %p529, %r5, 27; not.pred %p528, %p522; or.pred %p176, %p528, %p530; @%p176 bra $L__BB105_200; mul.ftz.f32 %f4950, %f356, %f356; mov.f32 %f4894, 0f3F800000; sub.ftz.f32 %f5098, %f4859, %f5031; add.ftz.f32 %f384, %f37, %f4950; setp.leu.ftz.f32 %p177, %f384, 0f00000000; @%p177 bra $L__BB105_199; sqrt.approx.ftz.f32 %f2320, %f384; sub.ftz.f32 %f2321, %f2320, %f356; add.ftz.f32 %f2322, %f356, %f2320; div.approx.ftz.f32 %f2323, %f2321, %f2322; fma.rn.ftz.f32 %f2324, %f356, %f2322, 0fBF800000; fma.rn.ftz.f32 %f2325, %f356, %f2321, 0f3F800000; div.approx.ftz.f32 %f2326, %f2324, %f2325; mul.ftz.f32 %f2327, %f2323, 0f3F000000; mul.ftz.f32 %f2328, %f2323, %f2327; fma.rn.ftz.f32 %f2329, %f2326, %f2326, 0f3F800000; mul.ftz.f32 %f2330, %f2328, %f2329; sub.ftz.f32 %f5098, %f2330, %f5031; $L__BB105_199: mul.ftz.f32 %f2331, %f128, %f5098; mov.f32 %f2332, 0f3F800000; sub.ftz.f32 %f2333, %f2332, %f2331; fma.rn.ftz.f32 %f2334, %f5012, %f2333, %f2331; fma.rn.ftz.f32 %f2335, %f5013, %f2333, %f2331; fma.rn.ftz.f32 %f2336, %f5014, %f2333, %f2331; mul.ftz.f32 %f5109, %f5109, %f2334; mul.ftz.f32 %f5110, %f5110, %f2335; mul.ftz.f32 %f5111, %f5111, %f2336; bra.uni $L__BB105_200; $L__BB105_195: mul.ftz.f32 %f5109, %f49, %f5109; mul.ftz.f32 %f5110, %f50, %f5110; mul.ftz.f32 %f5111, %f51, %f5111; $L__BB105_200: setp.gt.ftz.f32 %p178, %f5108, 0f3F7FF972; mov.f32 %f5112, 0f00000000; @%p178 bra $L__BB105_203; setp.lt.ftz.f32 %p179, %f5108, 0fBF7FF972; mov.f32 %f5112, 0fBF7FF972; @%p179 bra $L__BB105_203; mul.ftz.f32 %f2339, %f5108, %f5108; max.ftz.f32 %f2341, %f2339, %f1507; rcp.approx.ftz.f32 %f2342, %f2341; mul.ftz.f32 %f2343, %f1424, %f5107; mul.ftz.f32 %f2344, %f1423, %f5106; mul.ftz.f32 %f2345, %f2343, %f2343; fma.rn.ftz.f32 %f2346, %f2344, %f2344, %f2345; fma.rn.ftz.f32 %f2347, %f2346, %f2342, 0f3F800000; sqrt.approx.ftz.f32 %f2348, %f2347; neg.ftz.f32 %f2349, %f2348; setp.le.ftz.f32 %p180, %f5108, 0f00000000; selp.f32 %f2350, %f2349, %f2348, %p180; add.ftz.f32 %f2351, %f2350, 0fBF800000; mul.ftz.f32 %f5112, %f2351, 0f3F000000; $L__BB105_203: setp.ne.s32 %p527, %r5, 27; or.pred %p526, %p61, %p527; mul.ftz.f32 %f2353, %f49, %f5109; selp.f32 %f5109, %f2353, %f5109, %p526; mul.ftz.f32 %f2354, %f50, %f5110; selp.f32 %f5110, %f2354, %f5110, %p526; mul.ftz.f32 %f2355, %f51, %f5111; selp.f32 %f5111, %f2355, %f5111, %p526; add.ftz.f32 %f2356, %f5055, 0f3F800000; mov.f32 %f5053, 0f3F800000; mul.ftz.f32 %f2357, %f2356, 0f3F000000; cvt.ftz.sat.f32.f32 %f5054, %f2357; @%p178 bra $L__BB105_206; setp.lt.ftz.f32 %p182, %f5108, 0f3727C5AC; mov.f32 %f5053, 0f00000000; @%p182 bra $L__BB105_206; lg2.approx.ftz.f32 %f2359, %f5054; mul.ftz.f32 %f2360, %f5112, %f2359; ex2.approx.ftz.f32 %f5053, %f2360; $L__BB105_206: setp.lt.u32 %p183, %r161, 10; @%p183 bra $L__BB105_116; $L__BB105_207: st.local.u32 [%rd1+-4], %r162; setp.eq.s16 %p184, %rs37, 0; @%p184 bra $L__BB105_524; div.approx.ftz.f32 %f2361, %f5020, %f5017; abs.ftz.f32 %f2362, %f2361; mul.ftz.f32 %f5296, %f5296, %f2362; mul.ftz.f32 %f5295, %f5295, %f2362; mul.ftz.f32 %f5294, %f5294, %f2362; bra.uni $L__BB105_524; $L__BB105_518: and.b32 %r153, %r30, 4; setp.eq.s32 %p498, %r153, 0; add.ftz.f32 %f4711, %f1328, 0f00000000; selp.f32 %f5288, 0f00000000, %f4711, %p498; $L__BB105_519: and.b32 %r154, %r30, 9; setp.eq.s32 %p499, %r154, 0; @%p499 bra $L__BB105_521; mul.ftz.f32 %f4712, %f1405, %f1319; fma.rn.ftz.f32 %f4713, %f1404, %f1318, %f4712; fma.rn.ftz.f32 %f4714, %f1406, %f1320, %f4713; add.ftz.f32 %f4715, %f4714, 0f3F800000; ld.local.f32 %f4716, [%rd3+20]; mul.ftz.f32 %f4717, %f4715, %f4716; add.ftz.f32 %f4718, %f4717, 0fBF800000; mul.ftz.f32 %f4719, %f1326, %f1327; add.ftz.f32 %f4720, %f1326, %f1327; fma.rn.ftz.f32 %f4721, %f4719, %f4718, %f4720; fma.rn.ftz.f32 %f5288, %f4717, %f4721, %f5288; $L__BB105_521: mul.ftz.f32 %f4722, %f1324, 0f3EA2F983; mul.ftz.f32 %f5294, %f4722, %f5288; $L__BB105_522: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_37: ld.local.f32 %f3949, [%rd3+52]; mul.ftz.f32 %f3950, %f1112, %f3949; ld.local.f32 %f3951, [%rd3+48]; mul.ftz.f32 %f3952, %f1113, %f3951; sub.ftz.f32 %f3953, %f3950, %f3952; ld.local.f32 %f3954, [%rd3+44]; mul.ftz.f32 %f3955, %f1113, %f3954; mul.ftz.f32 %f3956, %f1111, %f3949; sub.ftz.f32 %f3957, %f3955, %f3956; mul.ftz.f32 %f3958, %f1111, %f3951; mul.ftz.f32 %f3959, %f1112, %f3954; sub.ftz.f32 %f3960, %f3958, %f3959; mul.ftz.f32 %f3961, %f3957, %f3957; fma.rn.ftz.f32 %f3962, %f3953, %f3953, %f3961; fma.rn.ftz.f32 %f3963, %f3960, %f3960, %f3962; sqrt.approx.ftz.f32 %f3964, %f3963; div.approx.ftz.f32 %f1124, %f3953, %f3964; div.approx.ftz.f32 %f1125, %f3957, %f3964; div.approx.ftz.f32 %f1126, %f3960, %f3964; mul.ftz.f32 %f3965, %f1113, %f1125; mul.ftz.f32 %f3966, %f1112, %f1126; sub.ftz.f32 %f1127, %f3965, %f3966; mul.ftz.f32 %f3967, %f1111, %f1126; mul.ftz.f32 %f3968, %f1113, %f1124; sub.ftz.f32 %f1128, %f3967, %f3968; mul.ftz.f32 %f3969, %f1112, %f1124; mul.ftz.f32 %f3970, %f1111, %f1125; sub.ftz.f32 %f1129, %f3969, %f3970; mul.ftz.f32 %f3971, %f1117, %f1128; fma.rn.ftz.f32 %f3972, %f1116, %f1127, %f3971; fma.rn.ftz.f32 %f3973, %f1118, %f1129, %f3972; mul.ftz.f32 %f3974, %f1117, %f1125; fma.rn.ftz.f32 %f3975, %f1116, %f1124, %f3974; fma.rn.ftz.f32 %f3976, %f1118, %f1126, %f3975; neg.ftz.f32 %f3977, %f3973; mul.ftz.f32 %f3978, %f3920, %f1119; div.approx.ftz.f32 %f3979, %f3977, %f3978; neg.ftz.f32 %f3980, %f3976; mul.ftz.f32 %f3981, %f3921, %f1119; div.approx.ftz.f32 %f3982, %f3980, %f3981; mul.ftz.f32 %f3983, %f3979, %f3979; neg.ftz.f32 %f3984, %f3983; mul.ftz.f32 %f3985, %f3982, %f3982; sub.ftz.f32 %f3986, %f3984, %f3985; mul.ftz.f32 %f3987, %f3986, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3988, %f3987; div.approx.ftz.f32 %f5254, %f3988, %f1121; mul.ftz.f32 %f3989, %f1106, %f1128; fma.rn.ftz.f32 %f3990, %f1105, %f1127, %f3989; fma.rn.ftz.f32 %f3991, %f1107, %f1129, %f3990; mul.ftz.f32 %f3992, %f1106, %f1125; fma.rn.ftz.f32 %f3993, %f1105, %f1124, %f3992; fma.rn.ftz.f32 %f3994, %f1107, %f1126, %f3993; mul.ftz.f32 %f3995, %f3994, %f3994; mul.ftz.f32 %f3996, %f3991, %f3991; mul.ftz.f32 %f1131, %f3920, %f3920; mul.ftz.f32 %f3997, %f1131, %f3996; mul.ftz.f32 %f1132, %f3921, %f3921; fma.rn.ftz.f32 %f3998, %f1132, %f3995, %f3997; add.ftz.f32 %f3999, %f3995, %f3996; div.approx.ftz.f32 %f4000, %f3998, %f3999; mul.ftz.f32 %f4001, %f1123, %f4000; div.approx.ftz.f32 %f4002, %f4001, %f1122; mov.f32 %f4003, 0f00000000; max.ftz.f32 %f4004, %f4002, %f4003; sqrt.approx.ftz.f32 %f1133, %f4004; setp.lt.ftz.f32 %p433, %f1133, 0f3F200000; mov.f32 %f5255, %f5256; @%p433 bra $L__BB105_39; rcp.approx.ftz.f32 %f4005, %f1133; fma.rn.ftz.f32 %f4006, %f4005, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4007, %f4005, %f4006; fma.rn.ftz.f32 %f4008, %f4005, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4009, %f4005, %f4008, 0f3F800000; div.approx.ftz.f32 %f5255, %f4007, %f4009; $L__BB105_39: mul.ftz.f32 %f4011, %f1405, %f1128; fma.rn.ftz.f32 %f4012, %f1404, %f1127, %f4011; fma.rn.ftz.f32 %f4013, %f1406, %f1129, %f4012; mul.ftz.f32 %f4014, %f1405, %f1125; fma.rn.ftz.f32 %f4015, %f1404, %f1124, %f4014; fma.rn.ftz.f32 %f4016, %f1406, %f1126, %f4015; mul.ftz.f32 %f4017, %f4016, %f4016; mul.ftz.f32 %f4018, %f4013, %f4013; mul.ftz.f32 %f4019, %f1131, %f4018; fma.rn.ftz.f32 %f4020, %f1132, %f4017, %f4019; add.ftz.f32 %f4021, %f4017, %f4018; div.approx.ftz.f32 %f4022, %f4020, %f4021; mul.ftz.f32 %f4023, %f1115, %f1115; sub.ftz.f32 %f4024, %f5256, %f4023; mul.ftz.f32 %f4025, %f4024, %f4022; div.approx.ftz.f32 %f4026, %f4025, %f4023; max.ftz.f32 %f4028, %f4026, %f4003; sqrt.approx.ftz.f32 %f1136, %f4028; setp.lt.ftz.f32 %p434, %f1136, 0f3F200000; @%p434 bra $L__BB105_461; rcp.approx.ftz.f32 %f4029, %f1136; fma.rn.ftz.f32 %f4030, %f4029, 0f400B9581, 0f40623D71; mul.ftz.f32 %f4031, %f4029, %f4030; fma.rn.ftz.f32 %f4032, %f4029, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f4033, %f4029, %f4032, 0f3F800000; div.approx.ftz.f32 %f5256, %f4031, %f4033; $L__BB105_461: mul.ftz.f32 %f4062, %f5254, 0f3E800000; div.approx.ftz.f32 %f4063, %f4062, %f1114; mul.ftz.f32 %f4064, %f5255, %f5256; mul.ftz.f32 %f5297, %f5255, %f4063; mul.ftz.f32 %f5294, %f4064, %f4063; bra.uni $L__BB105_468; $L__BB105_380: setp.neu.ftz.f32 %p355, %f882, %f883; setp.neu.ftz.f32 %p356, %f882, %f884; or.pred %p357, %p355, %p356; neg.ftz.f32 %f3459, %f884; selp.f32 %f3460, %f3459, %f884, %p357; add.ftz.f32 %f3461, %f882, %f3460; neg.ftz.f32 %f3462, %f883; selp.f32 %f3463, %f883, %f3462, %p357; sub.ftz.f32 %f3464, %f3463, %f882; sub.ftz.f32 %f3465, %f884, %f883; mul.ftz.f32 %f3466, %f3461, %f3461; fma.rn.ftz.f32 %f3467, %f3465, %f3465, %f3466; fma.rn.ftz.f32 %f3468, %f3464, %f3464, %f3467; sqrt.approx.ftz.f32 %f3469, %f3468; div.approx.ftz.f32 %f5178, %f3465, %f3469; div.approx.ftz.f32 %f5179, %f3461, %f3469; div.approx.ftz.f32 %f5180, %f3464, %f3469; mul.ftz.f32 %f3470, %f883, %f5180; mul.ftz.f32 %f3471, %f884, %f5179; sub.ftz.f32 %f5175, %f3470, %f3471; mul.ftz.f32 %f3472, %f884, %f5178; mul.ftz.f32 %f3473, %f882, %f5180; sub.ftz.f32 %f5176, %f3472, %f3473; mul.ftz.f32 %f3474, %f882, %f5179; mul.ftz.f32 %f3475, %f883, %f5178; sub.ftz.f32 %f5177, %f3474, %f3475; $L__BB105_382: mul.ftz.f32 %f3498, %f877, %f5179; fma.rn.ftz.f32 %f3499, %f876, %f5178, %f3498; fma.rn.ftz.f32 %f5187, %f878, %f5180, %f3499; mul.ftz.f32 %f3500, %f877, %f5176; fma.rn.ftz.f32 %f3501, %f876, %f5175, %f3500; fma.rn.ftz.f32 %f5188, %f878, %f5177, %f3501; mul.ftz.f32 %f3502, %f1405, %f5179; fma.rn.ftz.f32 %f3503, %f1404, %f5178, %f3502; fma.rn.ftz.f32 %f907, %f1406, %f5180, %f3503; mul.ftz.f32 %f3504, %f1405, %f5176; fma.rn.ftz.f32 %f3505, %f1404, %f5175, %f3504; fma.rn.ftz.f32 %f908, %f1406, %f5177, %f3505; add.ftz.f32 %f3506, %f5187, %f907; add.ftz.f32 %f3507, %f5188, %f908; mul.ftz.f32 %f3508, %f3506, %f3506; fma.rn.ftz.f32 %f3509, %f3507, %f3507, %f3508; add.ftz.f32 %f3510, %f5189, %f886; fma.rn.ftz.f32 %f3511, %f3510, %f3510, %f3509; sqrt.approx.ftz.f32 %f3512, %f3511; div.approx.ftz.f32 %f909, %f3506, %f3512; div.approx.ftz.f32 %f910, %f3507, %f3512; div.approx.ftz.f32 %f3513, %f3510, %f3512; mul.ftz.f32 %f911, %f3513, %f3513; @%p354 bra $L__BB105_387; bra.uni $L__BB105_383; $L__BB105_387: setp.gt.ftz.f32 %p361, %f5189, 0f3F7FF972; neg.ftz.f32 %f3560, %f909; div.approx.ftz.f32 %f3561, %f3560, %f3445; neg.ftz.f32 %f3562, %f910; div.approx.ftz.f32 %f3563, %f3562, %f3446; fma.rn.ftz.f32 %f3564, %f3561, %f3561, %f911; fma.rn.ftz.f32 %f916, %f3563, %f3563, %f3564; mov.f32 %f5182, 0f00000000; @%p361 bra $L__BB105_390; setp.lt.ftz.f32 %p362, %f5189, 0fBF7FF972; mov.f32 %f5182, 0fBF7FF972; @%p362 bra $L__BB105_390; mul.ftz.f32 %f3566, %f5189, %f5189; mov.f32 %f3567, 0f33D6BF95; max.ftz.f32 %f3568, %f3566, %f3567; rcp.approx.ftz.f32 %f3569, %f3568; mul.ftz.f32 %f3570, %f3445, %f5187; mul.ftz.f32 %f3571, %f5188, %f3446; mul.ftz.f32 %f3572, %f3571, %f3571; fma.rn.ftz.f32 %f3573, %f3570, %f3570, %f3572; fma.rn.ftz.f32 %f3574, %f3573, %f3569, 0f3F800000; sqrt.approx.ftz.f32 %f3575, %f3574; add.ftz.f32 %f3576, %f3575, 0fBF800000; mul.ftz.f32 %f5182, %f3576, 0f3F000000; $L__BB105_390: add.ftz.f32 %f3577, %f5182, 0f3F800000; mov.f32 %f3578, 0f3F800000; mul.ftz.f32 %f3579, %f5189, %f3577; mov.f32 %f3580, 0f33D6BF95; max.ftz.f32 %f3581, %f3579, %f3580; mul.ftz.f32 %f3582, %f916, 0f40490FDB; mul.ftz.f32 %f3583, %f916, %f3582; mul.ftz.f32 %f3584, %f3445, %f3583; mul.ftz.f32 %f3585, %f3446, %f3584; max.ftz.f32 %f3586, %f3585, %f3580; rcp.approx.ftz.f32 %f3587, %f3586; mul.ftz.f32 %f3588, %f3587, 0f3E800000; div.approx.ftz.f32 %f3589, %f3588, %f3581; sqrt.approx.ftz.f32 %f3590, %f881; mul.ftz.f32 %f3591, %f3590, 0fBFFE59F3; mul.ftz.f32 %f3592, %f3590, %f3591; mul.ftz.f32 %f3593, %f3592, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3594, %f3593; fma.rn.ftz.f32 %f3595, %f3594, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f3596, %f3590, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f3597, %f3590, %f3596, 0f417C8B44; fma.rn.ftz.f32 %f3598, %f3590, %f3597, 0fC1665E35; fma.rn.ftz.f32 %f3599, %f3590, %f3598, 0f40CE7881; fma.rn.ftz.f32 %f3600, %f3590, %f3599, 0fBF993E81; fma.rn.ftz.f32 %f3601, %f3590, %f3600, 0f3CE3C31E; fma.rn.ftz.f32 %f3602, %f3590, %f3601, 0f3BBA5D49; sub.ftz.f32 %f3603, %f3595, %f3602; cvt.ftz.sat.f32.f32 %f3604, %f3603; sub.ftz.f32 %f3605, %f3578, %f3604; mul.ftz.f32 %f3606, %f886, 0f3EA2F983; mul.ftz.f32 %f3607, %f3606, %f3605; fma.rn.ftz.f32 %f5183, %f3589, %f3604, %f3607; bra.uni $L__BB105_391; $L__BB105_49: ld.local.f32 %f3367, [%rd3+52]; mul.ftz.f32 %f3368, %f840, %f3367; ld.local.f32 %f3369, [%rd3+48]; mul.ftz.f32 %f3370, %f842, %f3369; sub.ftz.f32 %f3371, %f3368, %f3370; ld.local.f32 %f3372, [%rd3+44]; mul.ftz.f32 %f3373, %f842, %f3372; mul.ftz.f32 %f3374, %f838, %f3367; sub.ftz.f32 %f3375, %f3373, %f3374; mul.ftz.f32 %f3376, %f838, %f3369; mul.ftz.f32 %f3377, %f840, %f3372; sub.ftz.f32 %f3378, %f3376, %f3377; mul.ftz.f32 %f3379, %f3375, %f3375; fma.rn.ftz.f32 %f3380, %f3371, %f3371, %f3379; fma.rn.ftz.f32 %f3381, %f3378, %f3378, %f3380; sqrt.approx.ftz.f32 %f3382, %f3381; div.approx.ftz.f32 %f3383, %f3371, %f3382; div.approx.ftz.f32 %f3384, %f3375, %f3382; div.approx.ftz.f32 %f3385, %f3378, %f3382; mul.ftz.f32 %f3386, %f842, %f3384; mul.ftz.f32 %f3387, %f840, %f3385; sub.ftz.f32 %f3388, %f3386, %f3387; mul.ftz.f32 %f3389, %f838, %f3385; mul.ftz.f32 %f3390, %f842, %f3383; sub.ftz.f32 %f3391, %f3389, %f3390; mul.ftz.f32 %f3392, %f840, %f3383; mul.ftz.f32 %f3393, %f838, %f3384; sub.ftz.f32 %f3394, %f3392, %f3393; mul.ftz.f32 %f3395, %f849, %f3391; fma.rn.ftz.f32 %f3396, %f848, %f3388, %f3395; fma.rn.ftz.f32 %f856, %f850, %f3394, %f3396; mul.ftz.f32 %f3397, %f849, %f3384; fma.rn.ftz.f32 %f3398, %f848, %f3383, %f3397; fma.rn.ftz.f32 %f857, %f850, %f3385, %f3398; setp.geu.ftz.f32 %p349, %f852, 0f3F800000; mov.f32 %f5170, 0f3F800000; @%p349 bra $L__BB105_51; mul.ftz.f32 %f3399, %f854, %f856; mul.ftz.f32 %f3400, %f856, %f3399; mul.ftz.f32 %f3401, %f855, %f857; fma.rn.ftz.f32 %f3402, %f857, %f3401, %f3400; mul.ftz.f32 %f3403, %f852, %f852; mov.f32 %f3404, 0f3F800000; sub.ftz.f32 %f3405, %f3404, %f3403; div.approx.ftz.f32 %f3406, %f3402, %f3405; lg2.approx.ftz.f32 %f3407, %f852; mul.ftz.f32 %f3408, %f3406, %f3407; ex2.approx.ftz.f32 %f5170, %f3408; $L__BB105_51: add.ftz.f32 %f3409, %f855, 0f3F800000; add.ftz.f32 %f3410, %f854, 0f3F800000; mul.ftz.f32 %f3411, %f3410, %f3409; sqrt.approx.ftz.f32 %f3412, %f3411; mov.f32 %f3413, 0f41C90FDB; div.approx.ftz.f32 %f3414, %f3412, %f3413; mul.ftz.f32 %f3415, %f853, %f3414; mul.ftz.f32 %f5172, %f5170, %f3415; mul.ftz.f32 %f3416, %f5170, %f3414; div.approx.ftz.f32 %f5297, %f3416, %f851; $L__BB105_373: max.ftz.f32 %f3426, %f844, %f3357; max.ftz.f32 %f3427, %f853, %f3426; mul.ftz.f32 %f3428, %f851, %f3427; max.ftz.f32 %f3429, %f3357, %f3428; rcp.approx.ftz.f32 %f3430, %f3429; mul.ftz.f32 %f5294, %f3430, %f5172; $L__BB105_374: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; bra.uni $L__BB105_524; $L__BB105_383: setp.gt.ftz.f32 %p359, %f5189, 0f3F7FF972; mul.ftz.f32 %f3515, %f3445, %f3445; mov.f32 %f3516, 0f3F800000; sub.ftz.f32 %f3517, %f3516, %f911; fma.rn.ftz.f32 %f3518, %f3515, %f911, %f3517; mul.ftz.f32 %f3519, %f3518, 0f40490FDB; mul.ftz.f32 %f3520, %f3518, %f3519; mov.f32 %f3521, 0f33D6BF95; max.ftz.f32 %f3522, %f3520, %f3521; div.approx.ftz.f32 %f912, %f3515, %f3522; mov.f32 %f5181, 0f00000000; @%p359 bra $L__BB105_386; setp.lt.ftz.f32 %p360, %f5189, 0fBF7FF972; mov.f32 %f5181, 0fBF7FF972; @%p360 bra $L__BB105_386; mul.ftz.f32 %f3524, %f5189, %f5189; max.ftz.f32 %f3526, %f3524, %f3521; rcp.approx.ftz.f32 %f3527, %f3526; mul.ftz.f32 %f3528, %f3445, %f5187; mul.ftz.f32 %f3529, %f3528, %f3528; mul.ftz.f32 %f3530, %f5188, %f3445; fma.rn.ftz.f32 %f3531, %f3530, %f3530, %f3529; fma.rn.ftz.f32 %f3532, %f3531, %f3527, 0f3F800000; sqrt.approx.ftz.f32 %f3533, %f3532; add.ftz.f32 %f3534, %f3533, 0fBF800000; mul.ftz.f32 %f5181, %f3534, 0f3F000000; $L__BB105_386: add.ftz.f32 %f3535, %f5181, 0f3F800000; mul.ftz.f32 %f3537, %f5189, %f3535; max.ftz.f32 %f3539, %f3537, %f3521; mul.ftz.f32 %f3540, %f912, 0f3E800000; div.approx.ftz.f32 %f3541, %f3540, %f3539; mul.ftz.f32 %f3542, %f3445, 0fBFFE59F3; mul.ftz.f32 %f3543, %f3445, %f3542; mul.ftz.f32 %f3544, %f3543, 0f3FB8AA3B; ex2.approx.ftz.f32 %f3545, %f3544; fma.rn.ftz.f32 %f3546, %f3545, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f3547, %f3445, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f3548, %f3445, %f3547, 0f417C8B44; fma.rn.ftz.f32 %f3549, %f3445, %f3548, 0fC1665E35; fma.rn.ftz.f32 %f3550, %f3445, %f3549, 0f40CE7881; fma.rn.ftz.f32 %f3551, %f3445, %f3550, 0fBF993E81; fma.rn.ftz.f32 %f3552, %f3445, %f3551, 0f3CE3C31E; fma.rn.ftz.f32 %f3553, %f3445, %f3552, 0f3BBA5D49; sub.ftz.f32 %f3554, %f3546, %f3553; cvt.ftz.sat.f32.f32 %f3555, %f3554; sub.ftz.f32 %f3556, %f3516, %f3555; mul.ftz.f32 %f3557, %f886, 0f3EA2F983; mul.ftz.f32 %f3558, %f3557, %f3556; fma.rn.ftz.f32 %f5183, %f3541, %f3555, %f3558; $L__BB105_391: setp.le.ftz.f32 %p363, %f5183, 0f00000000; mov.f32 %f5297, %f5296; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p363 bra $L__BB105_524; ld.local.u64 %rd79, [%rd3+36]; ld.f32 %f921, [%rd79]; ld.f32 %f922, [%rd79+4]; ld.f32 %f923, [%rd79+8]; ld.local.f32 %f924, [%rd3+28]; ld.f32 %f925, [%rd79+12]; ld.f32 %f926, [%rd79+16]; ld.f32 %f927, [%rd79+20]; setp.geu.ftz.f32 %p364, %f886, %f5189; mov.f32 %f5184, %f907; mov.f32 %f5185, %f908; mov.f32 %f5186, %f886; mov.u16 %rs41, %rs12; @%p364 bra $L__BB105_394; mov.f32 %f5184, %f5187; mov.f32 %f5185, %f5188; mov.f32 %f5186, %f5189; mov.f32 %f5187, %f907; mov.f32 %f5188, %f908; mov.f32 %f5189, %f886; mov.u16 %rs41, %rs27; $L__BB105_394: setp.lt.ftz.f32 %p365, %f5186, 0f3727C5AC; setp.lt.ftz.f32 %p366, %f5189, 0f3727C5AC; or.pred %p367, %p365, %p366; mov.f32 %f5297, %f5183; mov.f32 %f5295, %f5296; mov.f32 %f5294, %f5296; @%p367 bra $L__BB105_524; neg.ftz.f32 %f5235, %f5187; neg.ftz.f32 %f5236, %f5188; setp.lt.ftz.f32 %p368, %f5189, 0fBF7FF972; mov.f32 %f5191, 0f00000000; mov.f32 %f5241, %f5191; @%p368 bra $L__BB105_398; setp.gt.ftz.f32 %p369, %f5189, 0f3F7FF972; mov.f32 %f5241, 0fBF7FF972; @%p369 bra $L__BB105_398; mul.ftz.f32 %f3617, %f5189, %f5189; mov.f32 %f3618, 0f33D6BF95; max.ftz.f32 %f3619, %f3617, %f3618; rcp.approx.ftz.f32 %f3620, %f3619; mul.ftz.f32 %f3621, %f3445, %f5235; mul.ftz.f32 %f3622, %f3446, %f5236; mul.ftz.f32 %f3623, %f3622, %f3622; fma.rn.ftz.f32 %f3624, %f3621, %f3621, %f3623; fma.rn.ftz.f32 %f3625, %f3624, %f3620, 0f3F800000; sqrt.approx.ftz.f32 %f3626, %f3625; neg.ftz.f32 %f3627, %f3626; setp.ge.ftz.f32 %p370, %f5189, 0f80000000; selp.f32 %f3628, %f3627, %f3626, %p370; add.ftz.f32 %f3629, %f3628, 0fBF800000; mul.ftz.f32 %f5241, %f3629, 0f3F000000; $L__BB105_398: setp.gt.ftz.f32 %p371, %f5186, 0f3F7FF972; @%p371 bra $L__BB105_401; setp.lt.ftz.f32 %p372, %f5186, 0fBF7FF972; mov.f32 %f5191, 0fBF7FF972; @%p372 bra $L__BB105_401; mul.ftz.f32 %f3632, %f5186, %f5186; mov.f32 %f3633, 0f33D6BF95; max.ftz.f32 %f3634, %f3632, %f3633; rcp.approx.ftz.f32 %f3635, %f3634; mul.ftz.f32 %f3636, %f3445, %f5184; mul.ftz.f32 %f3637, %f3446, %f5185; mul.ftz.f32 %f3638, %f3637, %f3637; fma.rn.ftz.f32 %f3639, %f3636, %f3636, %f3638; fma.rn.ftz.f32 %f3640, %f3639, %f3635, 0f3F800000; sqrt.approx.ftz.f32 %f3641, %f3640; neg.ftz.f32 %f3642, %f3641; setp.le.ftz.f32 %p373, %f5186, 0f00000000; selp.f32 %f3643, %f3642, %f3641, %p373; add.ftz.f32 %f3644, %f3643, 0fBF800000; mul.ftz.f32 %f5191, %f3644, 0f3F000000; $L__BB105_401: add.ftz.f32 %f3645, %f5184, %f5187; add.ftz.f32 %f3646, %f5185, %f5188; mul.ftz.f32 %f3647, %f3646, %f3646; fma.rn.ftz.f32 %f3648, %f3645, %f3645, %f3647; add.ftz.f32 %f3649, %f5186, %f5189; fma.rn.ftz.f32 %f3650, %f3649, %f3649, %f3648; sqrt.approx.ftz.f32 %f3651, %f3650; div.approx.ftz.f32 %f940, %f3645, %f3651; div.approx.ftz.f32 %f941, %f3646, %f3651; div.approx.ftz.f32 %f942, %f3649, %f3651; mul.ftz.f32 %f943, %f942, %f942; setp.eq.ftz.f32 %p374, %f3445, %f3446; @%p374 bra $L__BB105_403; bra.uni $L__BB105_402; $L__BB105_403: mul.ftz.f32 %f3664, %f3445, %f3445; mov.f32 %f3665, 0f3F800000; sub.ftz.f32 %f3666, %f3665, %f943; fma.rn.ftz.f32 %f3667, %f3664, %f943, %f3666; mul.ftz.f32 %f3668, %f3667, 0f40490FDB; mul.ftz.f32 %f3669, %f3667, %f3668; mov.f32 %f3670, 0f33D6BF95; max.ftz.f32 %f3671, %f3669, %f3670; div.approx.ftz.f32 %f5192, %f3664, %f3671; bra.uni $L__BB105_404; $L__BB105_60: mov.f32 %f3136, 0f3F800000; div.approx.ftz.f32 %f5162, %f772, %f3136; $L__BB105_355: mul.ftz.f32 %f776, %f767, %f768; neg.ftz.f32 %f3139, %f769; and.b32 %r116, %r19, 1; setp.eq.b32 %p318, %r116, 1; selp.f32 %f777, %f3139, %f769, %p318; add.ftz.f32 %f3140, %f5162, 0fBF800000; add.ftz.f32 %f3141, %f3140, 0f3F800000; mul.ftz.f32 %f3142, %f3141, %f3141; fma.rn.ftz.f32 %f3143, %f3142, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f3144, %f3141, %f3143; fma.rn.ftz.f32 %f3145, %f3142, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f3146, %f3142, %f3145, 0f3F800000; div.approx.ftz.f32 %f3147, %f3144, %f3146; sub.ftz.f32 %f3149, %f3067, %f3147; setp.gt.ftz.f32 %p319, %f772, 0f3F800000; selp.f32 %f3150, %f3149, %f3147, %p319; mov.b32 %r117, %f771; and.b32 %r118, %r117, -2147483648; mov.b32 %r119, %f3150; and.b32 %r120, %r119, 2147483647; or.b32 %r20, %r120, %r118; add.ftz.f32 %f3151, %f764, 0fBFC90FDB; mul.ftz.f32 %f3152, %f3151, 0f3F000000; sub.ftz.f32 %f3153, %f3152, %f762; mul.ftz.f32 %f778, %f770, %f3153; abs.ftz.f32 %f779, %f778; setp.eq.ftz.f32 %p320, %f779, 0f00000000; @%p320 bra $L__BB105_360; setp.eq.ftz.f32 %p321, %f779, 0f3F800000; mov.f32 %f5163, 0f3F800000; @%p321 bra $L__BB105_360; setp.gt.ftz.f32 %p322, %f779, 0f3F800000; @%p322 bra $L__BB105_359; bra.uni $L__BB105_358; $L__BB105_359: mov.f32 %f3156, 0f3F800000; div.approx.ftz.f32 %f5163, %f3156, %f779; bra.uni $L__BB105_360; $L__BB105_495: mov.f32 %f4514, 0f3F800000; div.approx.ftz.f32 %f5278, %f1257, %f4514; $L__BB105_497: add.ftz.f32 %f4517, %f5278, 0fBF800000; add.ftz.f32 %f4518, %f4517, 0f3F800000; mul.ftz.f32 %f4519, %f4518, %f4518; fma.rn.ftz.f32 %f4520, %f4519, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4521, %f4518, %f4520; fma.rn.ftz.f32 %f4522, %f4519, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4523, %f4519, %f4522, 0f3F800000; div.approx.ftz.f32 %f4524, %f4521, %f4523; sub.ftz.f32 %f4526, %f4440, %f4524; setp.gt.ftz.f32 %p473, %f1257, 0f3F800000; selp.f32 %f4527, %f4526, %f4524, %p473; mov.b32 %r139, %f1256; and.b32 %r140, %r139, -2147483648; mov.b32 %r141, %f4527; and.b32 %r142, %r141, 2147483647; or.b32 %r28, %r142, %r140; add.ftz.f32 %f4528, %f1250, 0fBFC90FDB; div.approx.ftz.f32 %f4530, %f4528, %f4510; sub.ftz.f32 %f4531, %f4530, %f1247; mul.ftz.f32 %f1261, %f1255, %f4531; abs.ftz.f32 %f1262, %f1261; setp.eq.ftz.f32 %p474, %f1262, 0f00000000; mov.f32 %f5279, %f5280; @%p474 bra $L__BB105_502; setp.eq.ftz.f32 %p475, %f1262, 0f3F800000; mov.f32 %f5279, 0f3F800000; @%p475 bra $L__BB105_502; setp.gt.ftz.f32 %p476, %f1262, 0f3F800000; @%p476 bra $L__BB105_501; bra.uni $L__BB105_500; $L__BB105_501: mov.f32 %f4534, 0f3F800000; div.approx.ftz.f32 %f5279, %f4534, %f1262; bra.uni $L__BB105_502; $L__BB105_358: mov.f32 %f3155, 0f3F800000; div.approx.ftz.f32 %f5163, %f779, %f3155; $L__BB105_360: add.ftz.f32 %f3157, %f5163, 0fBF800000; add.ftz.f32 %f3158, %f3157, 0f3F800000; mul.ftz.f32 %f3159, %f3158, %f3158; fma.rn.ftz.f32 %f3160, %f3159, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f3161, %f3158, %f3160; fma.rn.ftz.f32 %f3162, %f3159, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f3163, %f3159, %f3162, 0f3F800000; div.approx.ftz.f32 %f3164, %f3161, %f3163; sub.ftz.f32 %f3166, %f3067, %f3164; setp.gt.ftz.f32 %p323, %f779, 0f3F800000; selp.f32 %f3167, %f3166, %f3164, %p323; mov.b32 %r121, %f778; and.b32 %r122, %r121, -2147483648; mov.b32 %r123, %f3167; and.b32 %r124, %r123, 2147483647; or.b32 %r125, %r124, %r122; mov.b32 %f3168, %r125; add.ftz.f32 %f3169, %f764, %f765; mul.ftz.f32 %f3170, %f3169, 0f3F000000; sub.ftz.f32 %f3171, %f3170, %f762; mul.ftz.f32 %f3172, %f776, 0f3F000000; abs.ftz.f32 %f3173, %f776; setp.lt.ftz.f32 %p324, %f3173, 0f40490FDB; selp.f32 %f3174, %f3172, 0f3FC90FDB, %p324; mul.ftz.f32 %f3175, %f3174, 0f3EA2F983; mov.b32 %r126, %f3175; and.b32 %r127, %r126, -2147483648; or.b32 %r128, %r127, 1056964608; mov.b32 %f3176, %r128; add.ftz.f32 %f3177, %f3175, %f3176; cvt.rzi.ftz.s32.f32 %r129, %f3177; cvt.rn.f32.s32 %f3178, %r129; fma.rn.ftz.f32 %f3179, %f3178, 0fC0490000, %f3174; fma.rn.ftz.f32 %f3180, %f3178, 0fBA7DA000, %f3179; fma.rn.ftz.f32 %f3181, %f3178, 0fB4222000, %f3180; fma.rn.ftz.f32 %f3182, %f3178, 0fACB4611A, %f3181; add.ftz.f32 %f3183, %f3182, 0fBFC90FDB; add.ftz.f32 %f3184, %f3183, 0f3FC90FDB; mul.ftz.f32 %f3185, %f3184, %f3184; fma.rn.ftz.f32 %f3186, %f3185, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f3187, %f3185, %f3186, 0fBAB60B58; fma.rn.ftz.f32 %f3188, %f3185, %f3187, 0f3D2AAAAA; fma.rn.ftz.f32 %f3189, %f3185, %f3188, 0fBF000000; fma.rn.ftz.f32 %f3190, %f3185, %f3189, 0f3F800000; and.b32 %r130, %r129, 1; setp.eq.b32 %p325, %r130, 1; neg.ftz.f32 %f3191, %f3190; selp.f32 %f3192, %f3191, %f3190, %p325; abs.ftz.f32 %f3193, %f3192; setp.gt.ftz.f32 %p326, %f3193, 0f3F800000; mul.ftz.f32 %f3194, %f3192, 0f3E800000; selp.f32 %f3195, 0f00000000, %f3194, %p326; mul.ftz.f32 %f3196, %f767, %f3195; mul.ftz.f32 %f3197, %f3171, %f3171; fma.rn.ftz.f32 %f3198, %f763, %f763, %f3197; add.ftz.f32 %f3199, %f3198, %f3198; mov.b32 %f3200, %r20; sub.ftz.f32 %f3201, %f3200, %f3168; mul.ftz.f32 %f3202, %f3199, %f3201; abs.ftz.f32 %f3203, %f777; setp.gt.ftz.f32 %p327, %f3203, 0f3F800000; selp.f32 %f3204, 0f00000000, %f777, %p327; mul.ftz.f32 %f3205, %f3204, %f3202; div.approx.ftz.f32 %f3206, %f763, %f3205; mul.ftz.f32 %f5294, %f3206, %f3196; $L__BB105_361: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; bra.uni $L__BB105_524; $L__BB105_500: mov.f32 %f4533, 0f3F800000; div.approx.ftz.f32 %f5279, %f1262, %f4533; $L__BB105_502: add.ftz.f32 %f4536, %f5279, 0fBF800000; add.ftz.f32 %f4537, %f4536, 0f3F800000; mul.ftz.f32 %f4538, %f4537, %f4537; fma.rn.ftz.f32 %f4539, %f4538, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4540, %f4537, %f4539; fma.rn.ftz.f32 %f4541, %f4538, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4542, %f4538, %f4541, 0f3F800000; div.approx.ftz.f32 %f4543, %f4540, %f4542; sub.ftz.f32 %f4545, %f4440, %f4543; setp.gt.ftz.f32 %p477, %f1262, 0f3F800000; selp.f32 %f4546, %f4545, %f4543, %p477; mov.b32 %r143, %f1261; and.b32 %r144, %r143, -2147483648; mov.b32 %r145, %f4546; and.b32 %r146, %r145, 2147483647; or.b32 %r29, %r146, %r144; div.approx.ftz.f32 %f1266, %f4440, %f1249; abs.ftz.f32 %f1267, %f1266; setp.eq.ftz.f32 %p478, %f1267, 0f00000000; @%p478 bra $L__BB105_507; setp.eq.ftz.f32 %p479, %f1267, 0f3F800000; mov.f32 %f5280, 0f3F800000; @%p479 bra $L__BB105_507; setp.gt.ftz.f32 %p480, %f1267, 0f3F800000; @%p480 bra $L__BB105_506; bra.uni $L__BB105_505; $L__BB105_506: mov.f32 %f4549, 0f3F800000; div.approx.ftz.f32 %f5280, %f4549, %f1267; bra.uni $L__BB105_507; $L__BB105_505: mov.f32 %f4548, 0f3F800000; div.approx.ftz.f32 %f5280, %f1267, %f4548; $L__BB105_507: add.ftz.f32 %f4550, %f5280, 0fBF800000; add.ftz.f32 %f4551, %f4550, 0f3F800000; mul.ftz.f32 %f4552, %f4551, %f4551; fma.rn.ftz.f32 %f4553, %f4552, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f4554, %f4551, %f4553; fma.rn.ftz.f32 %f4555, %f4552, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f4556, %f4552, %f4555, 0f3F800000; div.approx.ftz.f32 %f4557, %f4554, %f4556; sub.ftz.f32 %f4559, %f4440, %f4557; setp.gt.ftz.f32 %p481, %f1267, 0f3F800000; selp.f32 %f4560, %f4559, %f4557, %p481; mov.b32 %r147, %f1266; and.b32 %r148, %r147, -2147483648; mov.b32 %r149, %f4560; and.b32 %r150, %r149, 2147483647; or.b32 %r151, %r150, %r148; mov.b32 %f4561, %r151; add.ftz.f32 %f4562, %f4561, %f4561; add.ftz.f32 %f4564, %f1250, %f1251; div.approx.ftz.f32 %f4565, %f4564, %f4510; sub.ftz.f32 %f4566, %f4565, %f1247; abs.ftz.f32 %f4567, %f1253; sub.ftz.f32 %f4569, %f4442, %f4567; mul.ftz.f32 %f4570, %f4566, %f4566; fma.rn.ftz.f32 %f4571, %f1248, %f1248, %f4570; add.ftz.f32 %f4572, %f4571, %f4571; mov.b32 %f4573, %r28; mov.b32 %f4574, %r29; sub.ftz.f32 %f4575, %f4573, %f4574; mul.ftz.f32 %f4576, %f4575, %f4572; abs.ftz.f32 %f4577, %f1254; setp.gt.ftz.f32 %p482, %f4577, 0f3F800000; selp.f32 %f4578, 0f00000000, %f1254, %p482; mul.ftz.f32 %f4579, %f4578, %f4576; div.approx.ftz.f32 %f4580, %f1248, %f4579; mul.ftz.f32 %f4581, %f4569, %f4569; fma.rn.ftz.f32 %f4582, %f1249, %f1249, %f4581; mul.ftz.f32 %f4583, %f4582, %f4562; div.approx.ftz.f32 %f4584, %f1249, %f4583; mul.ftz.f32 %f5294, %f4580, %f4584; $L__BB105_508: mov.f32 %f5295, %f5294; mov.f32 %f5296, %f5294; mov.f32 %f5297, %f5294; $L__BB105_524: add.s32 %r31, %r5, -2; setp.gt.u32 %p502, %r31, 6; @%p502 bra $L__BB105_535; ld.local.f32 %f1368, [%rd3+16]; ld.local.f32 %f1369, [%rd1+-240]; ld.local.f32 %f1370, [%rd1+-248]; ld.local.f32 %f1371, [%rd3+8]; setp.neu.ftz.f32 %p503, %f1371, %f1370; mov.u16 %rs42, %rs12; @%p503 bra $L__BB105_528; ld.local.f32 %f4754, [%rd3+12]; ld.local.f32 %f4755, [%rd1+-244]; setp.neu.ftz.f32 %p504, %f4754, %f4755; mov.u16 %rs42, %rs12; @%p504 bra $L__BB105_528; setp.eq.ftz.f32 %p505, %f1368, %f1369; selp.u16 %rs42, 1, 0, %p505; $L__BB105_528: setp.ne.s16 %p506, %rs42, 0; @%p506 bra $L__BB105_535; ld.local.f32 %f4757, [%rd3+12]; mul.ftz.f32 %f4758, %f1405, %f4757; fma.rn.ftz.f32 %f4759, %f1404, %f1371, %f4758; fma.rn.ftz.f32 %f4760, %f1406, %f1368, %f4759; setp.lt.ftz.f32 %p507, %f4760, 0f00000000; mov.f32 %f5298, 0f00000000; ld.local.f32 %f4761, [%rd1+-244]; neg.ftz.f32 %f4762, %f4761; neg.ftz.f32 %f4763, %f1370; selp.f32 %f4764, %f4763, %f1370, %p507; selp.f32 %f4765, %f4762, %f4761, %p507; neg.ftz.f32 %f4766, %f1369; selp.f32 %f4767, %f4766, %f1369, %p507; mul.ftz.f32 %f4768, %f1405, %f4765; fma.rn.ftz.f32 %f4769, %f1404, %f4764, %f4768; fma.rn.ftz.f32 %f1372, %f1406, %f4767, %f4769; mul.ftz.f32 %f4770, %f4757, %f4765; fma.rn.ftz.f32 %f4771, %f1371, %f4764, %f4770; fma.rn.ftz.f32 %f4772, %f1368, %f4767, %f4771; mul.ftz.f32 %f1373, %f4760, %f4772; setp.eq.ftz.f32 %p508, %f1373, 0f00000000; @%p508 bra $L__BB105_531; div.approx.ftz.f32 %f5298, %f1372, %f1373; $L__BB105_531: setp.ge.ftz.f32 %p509, %f5298, 0f3F800000; mov.f32 %f5299, 0f3F800000; @%p509 bra $L__BB105_534; setp.lt.ftz.f32 %p510, %f5298, 0f00000000; mov.f32 %f5299, 0f00000000; @%p510 bra $L__BB105_534; mul.ftz.f32 %f4775, %f5298, %f5298; mul.ftz.f32 %f4776, %f5298, %f4775; sub.ftz.f32 %f4777, %f4775, %f4776; add.ftz.f32 %f5299, %f5298, %f4777; $L__BB105_534: mul.ftz.f32 %f5296, %f5296, %f5299; mul.ftz.f32 %f5295, %f5295, %f5299; mul.ftz.f32 %f5294, %f5294, %f5299; $L__BB105_535: ld.local.u32 %r155, [%rd1+-184]; mul.wide.s32 %rd82, %r155, 208; add.s64 %rd83, %rd2, %rd82; ld.global.f32 %f1384, [%rd83+180]; setp.leu.ftz.f32 %p511, %f1384, 0f3F800000; @%p511 bra $L__BB105_538; ld.local.f32 %f4778, [%rd3+8]; ld.local.f32 %f4779, [%rd3+12]; mul.ftz.f32 %f4780, %f1405, %f4779; fma.rn.ftz.f32 %f4781, %f1404, %f4778, %f4780; ld.local.f32 %f4782, [%rd3+16]; fma.rn.ftz.f32 %f1385, %f1406, %f4782, %f4781; setp.ltu.ftz.f32 %p512, %f1385, 0f00000000; @%p512 bra $L__BB105_538; mov.f32 %f4783, 0f3F800000; min.ftz.f32 %f4784, %f1385, %f4783; abs.ftz.f32 %f4785, %f4784; setp.lt.ftz.f32 %p513, %f4785, 0f3F800000; add.ftz.f32 %f4786, %f4785, 0fBF800000; add.ftz.f32 %f4787, %f4786, 0f3F800000; selp.f32 %f4788, %f4787, 0f3F800000, %p513; sub.ftz.f32 %f4789, %f4783, %f4788; sqrt.approx.ftz.f32 %f4790, %f4789; fma.rn.ftz.f32 %f4791, %f4788, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f4792, %f4788, %f4791, 0fBE5A6B93; fma.rn.ftz.f32 %f4793, %f4788, %f4792, 0f3FC90FDB; mul.ftz.f32 %f4794, %f4790, %f4793; setp.lt.ftz.f32 %p514, %f4784, 0f00000000; mov.f32 %f4795, 0f00000000; mov.f32 %f4796, 0f40490FDB; sub.ftz.f32 %f4797, %f4796, %f4794; selp.f32 %f4798, %f4797, %f4794, %p514; mul.ftz.f32 %f4799, %f1384, %f4798; cos.approx.ftz.f32 %f4800, %f4799; max.ftz.f32 %f4801, %f4800, %f4795; div.approx.ftz.f32 %f4802, %f4801, %f4784; mul.ftz.f32 %f5296, %f5296, %f4802; mul.ftz.f32 %f5295, %f5295, %f4802; mul.ftz.f32 %f5294, %f5294, %f4802; $L__BB105_538: setp.eq.ftz.f32 %p515, %f5297, 0f00000000; @%p515 bra $L__BB105_544; setp.lt.u32 %p516, %r31, 7; ld.local.f32 %f4803, [%rd3+-12]; mul.ftz.f32 %f1392, %f5296, %f4803; ld.local.f32 %f4804, [%rd3+-8]; mul.ftz.f32 %f1393, %f5295, %f4804; ld.local.f32 %f4805, [%rd3+-4]; mul.ftz.f32 %f1394, %f5294, %f4805; @%p516 bra $L__BB105_542; bra.uni $L__BB105_540; $L__BB105_542: ld.local.f32 %f4812, [%rd10]; add.ftz.f32 %f4813, %f1392, %f4812; ld.local.f32 %f4814, [%rd10+4]; add.ftz.f32 %f4815, %f1393, %f4814; ld.local.f32 %f4816, [%rd10+8]; add.ftz.f32 %f4817, %f1394, %f4816; st.local.f32 [%rd10], %f4813; st.local.f32 [%rd10+4], %f4815; st.local.f32 [%rd10+8], %f4817; bra.uni $L__BB105_543; $L__BB105_540: add.s32 %r156, %r5, -9; setp.gt.u32 %p517, %r156, 11; setp.ne.s32 %p518, %r5, 29; and.pred %p519, %p518, %p517; @%p519 bra $L__BB105_543; ld.local.f32 %f4806, [%rd10+12]; add.ftz.f32 %f4807, %f1392, %f4806; ld.local.f32 %f4808, [%rd10+16]; add.ftz.f32 %f4809, %f1393, %f4808; ld.local.f32 %f4810, [%rd10+20]; add.ftz.f32 %f4811, %f1394, %f4810; st.local.f32 [%rd10+12], %f4807; st.local.f32 [%rd10+16], %f4809; st.local.f32 [%rd10+20], %f4811; $L__BB105_543: ld.local.f32 %f4818, [%rd10+24]; add.ftz.f32 %f4819, %f1392, %f4818; ld.local.f32 %f4820, [%rd10+28]; add.ftz.f32 %f4821, %f1393, %f4820; ld.local.f32 %f4822, [%rd10+32]; add.ftz.f32 %f4823, %f1394, %f4822; st.local.f32 [%rd10+24], %f4819; st.local.f32 [%rd10+28], %f4821; st.local.f32 [%rd10+32], %f4823; ld.local.f32 %f4824, [%rd3+4]; fma.rn.ftz.f32 %f5307, %f5297, %f4824, %f5307; ld.local.u32 %r168, [%rd1]; $L__BB105_544: ld.local.f32 %f4825, [%rd3+4]; add.ftz.f32 %f5308, %f5308, %f4825; $L__BB105_545: add.s32 %r159, %r159, 1; setp.lt.s32 %p520, %r159, %r168; @%p520 bra $L__BB105_2; $L__BB105_546: mov.f32 %f5311, 0f00000000; setp.leu.ftz.f32 %p521, %f5308, 0f00000000; @%p521 bra $L__BB105_548; div.approx.ftz.f32 %f5311, %f5307, %f5308; $L__BB105_548: st.param.f32 [func_retval0+0], %f5311; ret; } .func (.param .b32 func_retval0) _Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf( .param .b32 _Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf_param_0, .param .b64 _Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf_param_1 ) { .local .align 16 .b8 __local_depot106[7152]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<2190>; .reg .b16 %rs<162>; .reg .f32 %f<16792>; .reg .b32 %r<3302>; .reg .b64 %rd<3078>; mov.u64 %SPL, __local_depot106; cvta.local.u64 %SP, %SPL; ld.param.u32 %r953, [_Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf_param_0]; ld.param.u64 %rd618, [_Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf_param_1]; add.u64 %rd619, %SP, 1728; add.u64 %rd1, %SPL, 1728; ld.const.v2.u64 {%rd620, %rd621}, [kernel_params+2704]; mov.u32 %r954, 0; cvta.to.global.u64 %rd624, %rd621; cvt.s64.s32 %rd2, %r953; mul.wide.s32 %rd625, %r953, 4; add.s64 %rd626, %rd624, %rd625; ld.global.u32 %r955, [%rd626]; cvt.s64.s32 %rd3, %r955; ld.const.v2.u64 {%rd627, %rd628}, [kernel_params+2720]; cvta.to.global.u64 %rd631, %rd627; add.s64 %rd632, %rd631, %rd625; cvta.to.global.u64 %rd633, %rd628; add.s64 %rd634, %rd633, %rd625; ld.global.u32 %r1, [%rd634]; ld.const.v2.u64 {%rd635, %rd636}, [kernel_params+2688]; cvta.to.global.u64 %rd639, %rd636; add.s64 %rd640, %rd639, %rd625; ld.global.f32 %f1, [%rd640]; cvta.to.global.u64 %rd641, %rd620; add.s64 %rd642, %rd641, %rd625; ld.global.f32 %f2, [%rd642]; cvta.to.global.u64 %rd643, %rd635; add.s64 %rd644, %rd643, %rd625; ld.global.f32 %f15597, [%rd644]; ld.const.u64 %rd645, [kernel_params+2632]; cvta.to.global.u64 %rd646, %rd645; mul.wide.s32 %rd647, %r953, 12; add.s64 %rd4, %rd646, %rd647; ld.global.f32 %f15593, [%rd4]; ld.global.f32 %f15592, [%rd4+4]; ld.global.f32 %f15591, [%rd4+8]; ld.const.u64 %rd648, [kernel_params+2640]; cvta.to.global.u64 %rd649, %rd648; add.s64 %rd650, %rd649, %rd647; ld.global.f32 %f7, [%rd650]; ld.global.f32 %f8, [%rd650+4]; ld.global.f32 %f9, [%rd650+8]; ld.const.u64 %rd651, [kernel_params+2664]; cvta.to.global.u64 %rd652, %rd651; add.s64 %rd653, %rd652, %rd625; ld.global.f32 %f10, [%rd653]; ld.const.v2.u64 {%rd654, %rd655}, [kernel_params+2672]; cvta.to.global.u64 %rd656, %rd654; add.s64 %rd657, %rd656, %rd625; ld.global.f32 %f11, [%rd657]; cvta.to.global.u64 %rd658, %rd655; add.s64 %rd659, %rd658, %rd625; ld.global.f32 %f12, [%rd659]; add.s64 %rd7, %rd1, 88; ld.global.u32 %r956, [%rd632]; st.local.f32 [%rd1+88], %f15597; mov.b32 %r957, %f2; mov.b32 %r958, %f1; st.local.v4.u32 [%rd1+64], {%r1, %r958, %r957, %r956}; ld.const.u64 %rd660, [kernel_params+2184]; cvta.to.global.u64 %rd8, %rd660; cvt.s64.s32 %rd9, %r956; mul.wide.s32 %rd661, %r956, 4; add.s64 %rd662, %rd8, %rd661; ld.global.u32 %r3072, [%rd662]; st.local.u32 [%rd1+56], %r3072; st.local.u32 [%rd1+60], %r955; st.local.u32 [%rd1+52], %r954; mov.b32 %r959, %f10; mov.u32 %r3066, -1; st.local.v2.u32 [%rd1+80], {%r3066, %r959}; and.b32 %r3, %r3072, 2; setp.eq.s32 %p15, %r3, 0; @%p15 bra $L__BB106_7; ld.const.u64 %rd663, [kernel_params+2160]; cvta.to.global.u64 %rd664, %rd663; mul.lo.s64 %rd665, %rd9, 208; add.s64 %rd666, %rd664, %rd665; ld.const.u64 %rd667, [kernel_params+2176]; cvta.to.global.u64 %rd668, %rd667; ld.global.u32 %rd669, [%rd666+168]; ld.global.u32 %r961, [%rd666+152]; shl.b32 %r962, %r961, 1; cvt.rn.f32.s32 %f4315, %r962; mul.ftz.f32 %f4316, %f10, %f4315; cvt.rzi.ftz.s32.f32 %r963, %f4316; add.s32 %r964, %r962, -1; min.s32 %r965, %r963, %r964; cvt.rn.f32.s32 %f4317, %r965; sub.ftz.f32 %f13, %f4316, %f4317; cvt.s64.s32 %rd670, %r965; add.s64 %rd671, %rd670, %rd669; shl.b64 %rd672, %rd671, 6; add.s64 %rd10, %rd668, %rd672; ld.global.v4.f32 {%f4318, %f4319, %f4320, %f4321}, [%rd10]; ld.global.v4.f32 {%f4322, %f4323, %f4324, %f4325}, [%rd10+64]; mul.ftz.f32 %f4326, %f4319, %f4323; fma.rn.ftz.f32 %f4327, %f4318, %f4322, %f4326; mul.ftz.f32 %f4328, %f4321, %f4325; fma.rn.ftz.f32 %f4329, %f4320, %f4324, %f4328; add.ftz.f32 %f22, %f4327, %f4329; setp.gt.ftz.f32 %p16, %f22, 0f3F7FDF3B; @%p16 bra $L__BB106_3; bra.uni $L__BB106_2; $L__BB106_3: mov.f32 %f15510, 0f3F800000; sub.ftz.f32 %f15509, %f15510, %f13; mul.ftz.f32 %f4384, %f4318, %f15509; fma.rn.ftz.f32 %f4385, %f4322, %f13, %f4384; mul.ftz.f32 %f4386, %f4319, %f15509; fma.rn.ftz.f32 %f4387, %f4323, %f13, %f4386; mul.ftz.f32 %f4388, %f4320, %f15509; fma.rn.ftz.f32 %f4389, %f4324, %f13, %f4388; mul.ftz.f32 %f4390, %f4321, %f15509; fma.rn.ftz.f32 %f4391, %f4325, %f13, %f4390; mul.ftz.f32 %f4392, %f4387, %f4387; fma.rn.ftz.f32 %f4393, %f4385, %f4385, %f4392; mul.ftz.f32 %f4394, %f4391, %f4391; fma.rn.ftz.f32 %f4395, %f4389, %f4389, %f4394; add.ftz.f32 %f4396, %f4393, %f4395; rsqrt.approx.ftz.f32 %f4397, %f4396; mul.ftz.f32 %f15523, %f4391, %f4397; mul.ftz.f32 %f15522, %f4389, %f4397; mul.ftz.f32 %f15521, %f4387, %f4397; mul.ftz.f32 %f15520, %f4385, %f4397; bra.uni $L__BB106_4; $L__BB106_2: mov.f32 %f15506, 0f3F800000; mov.f32 %f4331, 0fBF800000; max.ftz.f32 %f4332, %f22, %f4331; min.ftz.f32 %f4334, %f4332, %f15506; abs.ftz.f32 %f4335, %f4334; sub.ftz.f32 %f4336, %f15506, %f4335; mul.ftz.f32 %f4337, %f4336, 0f3F000000; sqrt.approx.ftz.f32 %f4338, %f4337; setp.gt.ftz.f32 %p17, %f4335, 0f3F11EB85; selp.f32 %f4339, %f4338, %f4335, %p17; mul.ftz.f32 %f4340, %f4339, %f4339; mov.f32 %f4341, 0f3C94D2E9; mov.f32 %f4342, 0f3D53F941; fma.rn.ftz.f32 %f4343, %f4342, %f4340, %f4341; mov.f32 %f4344, 0f3D3F841F; fma.rn.ftz.f32 %f4345, %f4343, %f4340, %f4344; mov.f32 %f4346, 0f3D994929; fma.rn.ftz.f32 %f4347, %f4345, %f4340, %f4346; mov.f32 %f4348, 0f3E2AAB94; fma.rn.ftz.f32 %f4349, %f4347, %f4340, %f4348; mul.ftz.f32 %f4350, %f4340, %f4349; fma.rn.ftz.f32 %f4351, %f4350, %f4339, %f4339; add.ftz.f32 %f4352, %f4351, %f4351; mov.f32 %f4353, 0f3FC90FDB; sub.ftz.f32 %f4354, %f4353, %f4351; selp.f32 %f4355, %f4352, %f4354, %p17; setp.lt.ftz.f32 %p18, %f4334, 0f00000000; mov.f32 %f4356, 0f40490FDB; sub.ftz.f32 %f4357, %f4356, %f4355; selp.f32 %f4358, %f4357, %f4355, %p18; mul.ftz.f32 %f4359, %f4318, %f22; sub.ftz.f32 %f4360, %f4322, %f4359; mul.ftz.f32 %f4361, %f4319, %f22; sub.ftz.f32 %f4362, %f4323, %f4361; mul.ftz.f32 %f4363, %f4320, %f22; sub.ftz.f32 %f4364, %f4324, %f4363; mul.ftz.f32 %f4365, %f4321, %f22; sub.ftz.f32 %f4366, %f4325, %f4365; mul.ftz.f32 %f4367, %f4362, %f4362; fma.rn.ftz.f32 %f4368, %f4360, %f4360, %f4367; mul.ftz.f32 %f4369, %f4366, %f4366; fma.rn.ftz.f32 %f4370, %f4364, %f4364, %f4369; add.ftz.f32 %f4371, %f4368, %f4370; rsqrt.approx.ftz.f32 %f4372, %f4371; mul.ftz.f32 %f4373, %f4360, %f4372; mul.ftz.f32 %f4374, %f4362, %f4372; mul.ftz.f32 %f4375, %f4364, %f4372; mul.ftz.f32 %f4376, %f4366, %f4372; mul.ftz.f32 %f4377, %f13, %f4358; cos.approx.ftz.f32 %f4378, %f4377; sin.approx.ftz.f32 %f4379, %f4377; mul.ftz.f32 %f4380, %f4373, %f4379; mul.ftz.f32 %f4381, %f4374, %f4379; mul.ftz.f32 %f4382, %f4375, %f4379; mul.ftz.f32 %f4383, %f4376, %f4379; fma.rn.ftz.f32 %f15523, %f4321, %f4378, %f4383; fma.rn.ftz.f32 %f15522, %f4320, %f4378, %f4382; fma.rn.ftz.f32 %f15521, %f4319, %f4378, %f4381; fma.rn.ftz.f32 %f15520, %f4318, %f4378, %f4380; $L__BB106_4: mov.f32 %f15508, 0f3F800000; sub.ftz.f32 %f15507, %f15508, %f13; ld.global.v4.f32 {%f4398, %f4399, %f4400, %f4401}, [%rd10+16]; ld.global.v4.f32 {%f4406, %f4407, %f4408, %f4409}, [%rd10+80]; mul.ftz.f32 %f4414, %f13, %f4406; mul.ftz.f32 %f4415, %f13, %f4407; mul.ftz.f32 %f4416, %f13, %f4408; mul.ftz.f32 %f4417, %f13, %f4409; fma.rn.ftz.f32 %f144, %f15507, %f4398, %f4414; fma.rn.ftz.f32 %f148, %f15507, %f4399, %f4415; fma.rn.ftz.f32 %f152, %f15507, %f4400, %f4416; fma.rn.ftz.f32 %f4418, %f15507, %f4401, %f4417; ld.global.v4.f32 {%f4419, %f4420, %f4421, %f4422}, [%rd10+32]; ld.global.v4.f32 {%f4427, %f4428, %f4429, %f4430}, [%rd10+96]; mul.ftz.f32 %f4435, %f13, %f4427; mul.ftz.f32 %f4436, %f13, %f4428; mul.ftz.f32 %f4437, %f13, %f4429; mul.ftz.f32 %f4438, %f13, %f4430; fma.rn.ftz.f32 %f4439, %f15507, %f4419, %f4435; fma.rn.ftz.f32 %f4440, %f15507, %f4420, %f4436; fma.rn.ftz.f32 %f4441, %f15507, %f4421, %f4437; fma.rn.ftz.f32 %f4442, %f15507, %f4422, %f4438; ld.global.v4.f32 {%f4443, %f4444, %f4445, %f4446}, [%rd10+48]; ld.global.v4.f32 {%f4451, %f4452, %f4453, %f4454}, [%rd10+112]; mul.ftz.f32 %f4459, %f13, %f4451; mul.ftz.f32 %f4460, %f13, %f4452; mul.ftz.f32 %f4461, %f13, %f4453; mul.ftz.f32 %f4462, %f13, %f4454; fma.rn.ftz.f32 %f4463, %f15507, %f4443, %f4459; fma.rn.ftz.f32 %f4464, %f15507, %f4444, %f4460; fma.rn.ftz.f32 %f4465, %f15507, %f4445, %f4461; fma.rn.ftz.f32 %f4466, %f15507, %f4446, %f4462; mul.ftz.f32 %f4467, %f15523, 0f3FB504F3; mul.ftz.f32 %f4468, %f15520, 0f3FB504F3; mul.ftz.f32 %f4469, %f15521, 0f3FB504F3; mul.ftz.f32 %f4470, %f15522, 0f3FB504F3; mul.ftz.f32 %f4471, %f4467, %f4468; mul.ftz.f32 %f4472, %f4467, %f4469; mul.ftz.f32 %f4473, %f4467, %f4470; mul.ftz.f32 %f4474, %f4468, %f4468; mul.ftz.f32 %f4475, %f4468, %f4469; mul.ftz.f32 %f4476, %f4468, %f4470; mul.ftz.f32 %f4477, %f4469, %f4469; mul.ftz.f32 %f4478, %f4469, %f4470; mul.ftz.f32 %f4479, %f4470, %f4470; sub.ftz.f32 %f4481, %f15508, %f4477; sub.ftz.f32 %f4482, %f4481, %f4479; sub.ftz.f32 %f4483, %f4475, %f4473; add.ftz.f32 %f4484, %f4472, %f4476; add.ftz.f32 %f4485, %f4473, %f4475; sub.ftz.f32 %f4486, %f15508, %f4474; sub.ftz.f32 %f4487, %f4486, %f4479; sub.ftz.f32 %f4488, %f4478, %f4471; sub.ftz.f32 %f4489, %f4476, %f4472; add.ftz.f32 %f4490, %f4471, %f4478; sub.ftz.f32 %f4491, %f4486, %f4477; mul.ftz.f32 %f4492, %f4483, %f4441; fma.rn.ftz.f32 %f4493, %f4482, %f4418, %f4492; fma.rn.ftz.f32 %f141, %f4484, %f4464, %f4493; mul.ftz.f32 %f4494, %f4483, %f4442; fma.rn.ftz.f32 %f4495, %f4482, %f4439, %f4494; fma.rn.ftz.f32 %f142, %f4484, %f4465, %f4495; mul.ftz.f32 %f4496, %f4483, %f4463; fma.rn.ftz.f32 %f4497, %f4482, %f4440, %f4496; fma.rn.ftz.f32 %f143, %f4484, %f4466, %f4497; mul.ftz.f32 %f4498, %f4487, %f4441; fma.rn.ftz.f32 %f4499, %f4485, %f4418, %f4498; fma.rn.ftz.f32 %f145, %f4488, %f4464, %f4499; mul.ftz.f32 %f4500, %f4487, %f4442; fma.rn.ftz.f32 %f4501, %f4485, %f4439, %f4500; fma.rn.ftz.f32 %f146, %f4488, %f4465, %f4501; mul.ftz.f32 %f4502, %f4487, %f4463; fma.rn.ftz.f32 %f4503, %f4485, %f4440, %f4502; fma.rn.ftz.f32 %f147, %f4488, %f4466, %f4503; mul.ftz.f32 %f4504, %f4490, %f4441; fma.rn.ftz.f32 %f4505, %f4489, %f4418, %f4504; fma.rn.ftz.f32 %f149, %f4491, %f4464, %f4505; mul.ftz.f32 %f4506, %f4490, %f4442; fma.rn.ftz.f32 %f4507, %f4489, %f4439, %f4506; fma.rn.ftz.f32 %f150, %f4491, %f4465, %f4507; mul.ftz.f32 %f4508, %f4490, %f4463; fma.rn.ftz.f32 %f4509, %f4489, %f4440, %f4508; fma.rn.ftz.f32 %f151, %f4491, %f4466, %f4509; st.local.v4.f32 [%rd7+56], {%f141, %f142, %f143, %f144}; st.local.v4.f32 [%rd7+72], {%f145, %f146, %f147, %f148}; st.local.v4.f32 [%rd7+88], {%f149, %f150, %f151, %f152}; mul.ftz.f32 %f4510, %f146, %f151; mul.ftz.f32 %f60, %f150, %f147; sub.ftz.f32 %f15526, %f4510, %f60; mul.ftz.f32 %f62, %f150, %f143; mul.ftz.f32 %f4511, %f142, %f151; sub.ftz.f32 %f15525, %f62, %f4511; mul.ftz.f32 %f64, %f142, %f147; mul.ftz.f32 %f4512, %f146, %f143; sub.ftz.f32 %f15524, %f64, %f4512; mul.ftz.f32 %f4513, %f145, %f15525; fma.rn.ftz.f32 %f4514, %f141, %f15526, %f4513; fma.rn.ftz.f32 %f15530, %f149, %f15524, %f4514; setp.neu.ftz.f32 %p19, %f15530, 0f00000000; mov.f32 %f15527, %f141; mov.f32 %f15528, %f146; mov.f32 %f15529, %f151; @%p19 bra $L__BB106_6; add.ftz.f32 %f15527, %f141, 0f322BCC77; add.ftz.f32 %f15529, %f151, 0f322BCC77; add.ftz.f32 %f15528, %f146, 0f322BCC77; mul.ftz.f32 %f4515, %f15528, %f15529; sub.ftz.f32 %f15526, %f4515, %f60; mul.ftz.f32 %f4516, %f142, %f15529; sub.ftz.f32 %f15525, %f62, %f4516; mul.ftz.f32 %f4517, %f15528, %f143; sub.ftz.f32 %f15524, %f64, %f4517; mul.ftz.f32 %f4518, %f145, %f15525; fma.rn.ftz.f32 %f4519, %f15527, %f15526, %f4518; fma.rn.ftz.f32 %f4520, %f149, %f15524, %f4519; setp.eq.ftz.f32 %p20, %f4520, 0f00000000; selp.f32 %f15530, 0f7F7FFFFF, %f4520, %p20; $L__BB106_6: mul.ftz.f32 %f4521, %f145, %f15529; mul.ftz.f32 %f4522, %f149, %f147; sub.ftz.f32 %f4523, %f4522, %f4521; mul.ftz.f32 %f4524, %f149, %f143; mul.ftz.f32 %f4525, %f15527, %f15529; sub.ftz.f32 %f4526, %f4525, %f4524; mul.ftz.f32 %f4527, %f147, %f15527; mul.ftz.f32 %f4528, %f145, %f143; sub.ftz.f32 %f4529, %f4528, %f4527; mul.ftz.f32 %f4530, %f149, %f15528; mul.ftz.f32 %f4531, %f145, %f150; sub.ftz.f32 %f4532, %f4531, %f4530; mul.ftz.f32 %f4533, %f150, %f15527; mul.ftz.f32 %f4534, %f149, %f142; sub.ftz.f32 %f4535, %f4534, %f4533; mul.ftz.f32 %f4536, %f145, %f142; mul.ftz.f32 %f4537, %f15527, %f15528; sub.ftz.f32 %f4538, %f4537, %f4536; div.approx.ftz.f32 %f15565, %f15525, %f15530; mul.ftz.f32 %f4539, %f148, %f15565; div.approx.ftz.f32 %f15564, %f15526, %f15530; fma.rn.ftz.f32 %f4540, %f144, %f15564, %f4539; div.approx.ftz.f32 %f15566, %f15524, %f15530; fma.rn.ftz.f32 %f4541, %f152, %f15566, %f4540; div.approx.ftz.f32 %f15569, %f4526, %f15530; mul.ftz.f32 %f4542, %f148, %f15569; div.approx.ftz.f32 %f15568, %f4523, %f15530; fma.rn.ftz.f32 %f4543, %f144, %f15568, %f4542; div.approx.ftz.f32 %f15570, %f4529, %f15530; fma.rn.ftz.f32 %f4544, %f152, %f15570, %f4543; div.approx.ftz.f32 %f15573, %f4535, %f15530; mul.ftz.f32 %f4545, %f148, %f15573; div.approx.ftz.f32 %f15572, %f4532, %f15530; fma.rn.ftz.f32 %f4546, %f144, %f15572, %f4545; div.approx.ftz.f32 %f15574, %f4538, %f15530; fma.rn.ftz.f32 %f4547, %f152, %f15574, %f4546; neg.ftz.f32 %f15579, %f4541; neg.ftz.f32 %f15587, %f4547; neg.ftz.f32 %f15583, %f4544; st.local.v4.f32 [%rd7+104], {%f15564, %f15565, %f15566, %f15579}; st.local.v4.f32 [%rd7+120], {%f15568, %f15569, %f15570, %f15583}; st.local.v4.f32 [%rd7+136], {%f15572, %f15573, %f15574, %f15587}; $L__BB106_7: neg.ftz.f32 %f165, %f7; st.local.f32 [%rd7+-52], %f165; neg.ftz.f32 %f166, %f8; st.local.f32 [%rd7+-48], %f166; neg.ftz.f32 %f167, %f9; st.local.f32 [%rd7+-44], %f167; and.b32 %r966, %r1, 6; setp.eq.s32 %p21, %r966, 0; @%p21 bra $L__BB106_46; and.b32 %r4, %r3072, 4; setp.ne.s32 %p22, %r4, 0; mov.f32 %f15594, %f9; mov.f32 %f15595, %f8; mov.f32 %f15596, %f7; @%p22 bra $L__BB106_14; and.b32 %r3042, %r3072, 2; setp.ne.s32 %p23, %r3042, 0; mov.f32 %f15576, %f15564; mov.f32 %f15577, %f15565; mov.f32 %f15578, %f15566; mov.f32 %f15580, %f15568; mov.f32 %f15581, %f15569; mov.f32 %f15582, %f15570; mov.f32 %f15584, %f15572; mov.f32 %f15585, %f15573; mov.f32 %f15586, %f15574; @%p23 bra $L__BB106_11; ld.const.u64 %rd673, [kernel_params+2160]; cvta.to.global.u64 %rd674, %rd673; mul.lo.s64 %rd675, %rd9, 208; add.s64 %rd676, %rd674, %rd675; ld.global.v4.f32 {%f15576, %f15577, %f15578, %f15579}, [%rd676+48]; ld.global.v4.f32 {%f15580, %f15581, %f15582, %f15583}, [%rd676+64]; ld.global.v4.f32 {%f15584, %f15585, %f15586, %f15587}, [%rd676+80]; $L__BB106_11: mul.ftz.f32 %f4560, %f15592, %f15577; fma.rn.ftz.f32 %f4561, %f15593, %f15576, %f4560; fma.rn.ftz.f32 %f4562, %f15591, %f15578, %f4561; add.ftz.f32 %f192, %f15579, %f4562; mul.ftz.f32 %f4563, %f15592, %f15581; fma.rn.ftz.f32 %f4564, %f15593, %f15580, %f4563; fma.rn.ftz.f32 %f4565, %f15591, %f15582, %f4564; add.ftz.f32 %f193, %f15583, %f4565; mul.ftz.f32 %f4566, %f15592, %f15585; fma.rn.ftz.f32 %f4567, %f15593, %f15584, %f4566; fma.rn.ftz.f32 %f4568, %f15591, %f15586, %f4567; add.ftz.f32 %f15591, %f15587, %f4568; mul.ftz.f32 %f4569, %f15597, %f7; mul.ftz.f32 %f4570, %f15597, %f8; mul.ftz.f32 %f4571, %f4570, %f15577; fma.rn.ftz.f32 %f4572, %f4569, %f15576, %f4571; mul.ftz.f32 %f4573, %f15597, %f9; fma.rn.ftz.f32 %f15596, %f4573, %f15578, %f4572; mul.ftz.f32 %f4574, %f4570, %f15581; fma.rn.ftz.f32 %f4575, %f4569, %f15580, %f4574; fma.rn.ftz.f32 %f15595, %f4573, %f15582, %f4575; mul.ftz.f32 %f4576, %f4570, %f15585; fma.rn.ftz.f32 %f4577, %f4569, %f15584, %f4576; fma.rn.ftz.f32 %f15594, %f4573, %f15586, %f4577; mul.ftz.f32 %f4578, %f15595, %f15595; fma.rn.ftz.f32 %f4579, %f15596, %f15596, %f4578; fma.rn.ftz.f32 %f4580, %f15594, %f15594, %f4579; sqrt.approx.ftz.f32 %f15597, %f4580; setp.neu.ftz.f32 %p24, %f15597, 0f00000000; @%p24 bra $L__BB106_12; bra.uni $L__BB106_13; $L__BB106_12: div.approx.ftz.f32 %f15594, %f15594, %f15597; div.approx.ftz.f32 %f15595, %f15595, %f15597; div.approx.ftz.f32 %f15596, %f15596, %f15597; $L__BB106_13: mov.f32 %f15592, %f193; mov.f32 %f15593, %f192; $L__BB106_14: ld.const.u64 %rd677, [kernel_params+2264]; cvta.to.global.u64 %rd678, %rd677; shl.b64 %rd679, %rd3, 4; add.s64 %rd680, %rd678, %rd679; add.s64 %rd11, %rd680, 4; shr.s32 %r967, %r1, 7; ld.global.u32 %r968, [%rd680+4]; add.s32 %r5, %r967, %r968; add.s32 %r969, %r5, -1; max.s32 %r6, %r969, %r968; add.s32 %r970, %r5, 2; ld.global.u32 %r971, [%rd680+8]; add.s32 %r972, %r968, %r971; add.s32 %r973, %r972, -1; min.s32 %r7, %r970, %r973; and.b32 %r974, %r1, 64; setp.eq.s32 %p25, %r974, 0; @%p25 bra $L__BB106_28; bra.uni $L__BB106_15; $L__BB106_28: ld.const.u64 %rd722, [kernel_params+2272]; cvta.to.global.u64 %rd723, %rd722; mul.wide.s32 %rd724, %r6, 16; add.s64 %rd725, %rd723, %rd724; ld.global.v4.f32 {%f15633, %f15632, %f15631, %f4665}, [%rd725]; mul.wide.s32 %rd726, %r5, 16; add.s64 %rd727, %rd723, %rd726; ld.global.v4.f32 {%f15630, %f15629, %f15628, %f4669}, [%rd727]; ld.global.v4.f32 {%f15627, %f15626, %f15625, %f4673}, [%rd727+16]; mul.wide.s32 %rd728, %r7, 16; add.s64 %rd729, %rd723, %rd728; ld.global.v4.f32 {%f15624, %f15623, %f15622, %f4677}, [%rd729]; bra.uni $L__BB106_29; $L__BB106_46: and.b32 %r993, %r1, 8; setp.eq.s32 %p46, %r993, 0; @%p46 bra $L__BB106_67; bra.uni $L__BB106_47; $L__BB106_67: setp.eq.s32 %p59, %r1, 1; @%p59 bra $L__BB106_69; bra.uni $L__BB106_68; $L__BB106_69: cvt.u32.u64 %r1012, %rd3; ld.const.u64 %rd783, [kernel_params+2232]; cvta.to.global.u64 %rd784, %rd783; mul.wide.s32 %rd785, %r1012, 16; add.s64 %rd786, %rd784, %rd785; add.s64 %rd27, %rd786, 12; ld.const.u64 %rd787, [kernel_params+2256]; cvta.to.global.u64 %rd788, %rd787; ld.global.u32 %r1013, [%rd786+12]; mul.wide.u32 %rd789, %r1013, 12; add.s64 %rd790, %rd788, %rd789; add.s32 %r1014, %r1013, 1; mul.wide.u32 %rd791, %r1014, 12; add.s64 %rd792, %rd788, %rd791; add.s32 %r1015, %r1013, 2; mul.wide.u32 %rd793, %r1015, 12; add.s64 %rd794, %rd788, %rd793; and.b32 %r1016, %r3072, 8; setp.eq.s32 %p60, %r1016, 0; ld.global.f32 %f5011, [%rd794]; ld.global.f32 %f425, [%rd790]; sub.ftz.f32 %f426, %f5011, %f425; ld.global.f32 %f5012, [%rd794+4]; ld.global.f32 %f427, [%rd790+4]; sub.ftz.f32 %f428, %f5012, %f427; ld.global.f32 %f5013, [%rd794+8]; ld.global.f32 %f429, [%rd790+8]; sub.ftz.f32 %f430, %f5013, %f429; ld.global.f32 %f5014, [%rd792]; sub.ftz.f32 %f431, %f5014, %f425; ld.global.f32 %f5015, [%rd792+4]; sub.ftz.f32 %f432, %f5015, %f427; ld.global.f32 %f5016, [%rd792+8]; sub.ftz.f32 %f433, %f5016, %f429; mul.ftz.f32 %f434, %f433, %f428; mul.ftz.f32 %f435, %f432, %f430; mul.ftz.f32 %f436, %f431, %f430; mul.ftz.f32 %f437, %f433, %f426; mul.ftz.f32 %f438, %f432, %f426; mul.ftz.f32 %f439, %f431, %f428; @%p60 bra $L__BB106_71; bra.uni $L__BB106_70; $L__BB106_71: sub.ftz.f32 %f5024, %f435, %f434; sub.ftz.f32 %f5025, %f437, %f436; mul.ftz.f32 %f5026, %f5025, %f5025; fma.rn.ftz.f32 %f5027, %f5024, %f5024, %f5026; sub.ftz.f32 %f5028, %f439, %f438; fma.rn.ftz.f32 %f5029, %f5028, %f5028, %f5027; sqrt.approx.ftz.f32 %f5030, %f5029; div.approx.ftz.f32 %f15663, %f5028, %f5030; div.approx.ftz.f32 %f15662, %f5025, %f5030; div.approx.ftz.f32 %f15661, %f5024, %f5030; bra.uni $L__BB106_72; $L__BB106_15: ld.const.u64 %rd681, [kernel_params+2160]; cvta.to.global.u64 %rd682, %rd681; mul.lo.s64 %rd683, %rd9, 208; add.s64 %rd684, %rd682, %rd683; ld.global.u32 %r8, [%rd684+148]; ld.global.u32 %r9, [%rd684+152]; shl.b32 %r975, %r9, 1; cvt.rn.f32.s32 %f4581, %r975; mul.ftz.f32 %f218, %f10, %f4581; cvt.rzi.ftz.s32.f32 %r976, %f218; add.s32 %r977, %r975, -1; min.s32 %r10, %r976, %r977; ld.const.u64 %rd685, [kernel_params+2312]; cvta.to.global.u64 %rd12, %rd685; ld.global.u32 %r3063, [%rd684+164]; mul.wide.u32 %rd686, %r3063, 16; add.s64 %rd687, %rd12, %rd686; ld.global.u32 %r3062, [%rd687+8]; ld.global.u16 %rs133, [%rd687+12]; ld.global.u64 %rd2925, [%rd687]; setp.eq.s64 %p26, %rd2925, 11; @%p26 bra $L__BB106_20; $L__BB106_16: setp.eq.s64 %p27, %rd2925, 0; @%p27 bra $L__BB106_18; add.s32 %r3063, %r3063, 2; bra.uni $L__BB106_19; $L__BB106_18: setp.eq.s16 %p28, %rs133, 0; mov.u32 %r3063, %r3062; @%p28 bra $L__BB106_21; $L__BB106_19: mul.wide.u32 %rd688, %r3063, 16; add.s64 %rd689, %rd12, %rd688; ld.global.u32 %r3062, [%rd689+8]; ld.global.u16 %rs133, [%rd689+12]; ld.global.u64 %rd2925, [%rd689]; setp.ne.s64 %p29, %rd2925, 11; @%p29 bra $L__BB106_16; $L__BB106_20: setp.eq.s16 %p30, %rs133, 0; selp.b32 %r3066, -1, %r3062, %p30; $L__BB106_21: setp.eq.s32 %p31, %r10, %r9; @%p31 bra $L__BB106_23; bra.uni $L__BB106_22; $L__BB106_23: ld.const.u64 %rd698, [kernel_params+2272]; cvta.to.global.u64 %rd699, %rd698; mul.wide.s32 %rd700, %r6, 16; add.s64 %rd701, %rd699, %rd700; ld.global.v4.f32 {%f15609, %f15608, %f15607, %f4601}, [%rd701]; mul.wide.s32 %rd702, %r5, 16; add.s64 %rd703, %rd699, %rd702; ld.global.v4.f32 {%f15606, %f15605, %f15604, %f4605}, [%rd703]; ld.global.v4.f32 {%f15603, %f15602, %f15601, %f4609}, [%rd703+16]; mul.wide.s32 %rd704, %r7, 16; add.s64 %rd705, %rd699, %rd704; ld.global.v4.f32 {%f15600, %f15599, %f15598, %f4613}, [%rd705]; bra.uni $L__BB106_24; $L__BB106_47: ld.const.u64 %rd744, [kernel_params+2304]; cvta.to.global.u64 %rd745, %rd744; shl.b64 %rd746, %rd3, 2; add.s64 %rd747, %rd745, %rd746; ld.global.u32 %r994, [%rd747]; st.local.u32 [%rd7+-40], %r994; fma.rn.ftz.f32 %f4925, %f15597, %f9, %f15591; fma.rn.ftz.f32 %f4926, %f15597, %f8, %f15592; fma.rn.ftz.f32 %f4927, %f15597, %f7, %f15593; st.local.v2.f32 [%rd1], {%f4927, %f4926}; st.local.f32 [%rd7+-80], %f4925; and.b32 %r995, %r1, 64; setp.eq.s32 %p47, %r995, 0; @%p47 bra $L__BB106_61; bra.uni $L__BB106_48; $L__BB106_61: ld.const.u64 %rd771, [kernel_params+2296]; cvta.to.global.u64 %rd772, %rd771; shl.b64 %rd773, %rd3, 4; add.s64 %rd774, %rd772, %rd773; ld.global.v4.f32 {%f15658, %f15659, %f15660, %f4952}, [%rd774]; bra.uni $L__BB106_62; $L__BB106_22: setp.gt.s32 %p32, %r10, %r9; selp.b32 %r979, -1, 0, %p32; add.s32 %r980, %r10, %r979; mad.lo.s32 %r981, %r980, %r8, %r3066; ld.const.u64 %rd690, [kernel_params+2344]; cvta.to.global.u64 %rd691, %rd690; add.s32 %r982, %r981, %r6; mul.wide.s32 %rd692, %r982, 16; add.s64 %rd693, %rd691, %rd692; ld.global.v4.f32 {%f15609, %f15608, %f15607, %f4585}, [%rd693]; add.s32 %r983, %r981, %r5; mul.wide.s32 %rd694, %r983, 16; add.s64 %rd695, %rd691, %rd694; ld.global.v4.f32 {%f15606, %f15605, %f15604, %f4589}, [%rd695]; ld.global.v4.f32 {%f15603, %f15602, %f15601, %f4593}, [%rd695+16]; add.s32 %r984, %r981, %r7; mul.wide.s32 %rd696, %r984, 16; add.s64 %rd697, %rd691, %rd696; ld.global.v4.f32 {%f15600, %f15599, %f15598, %f4597}, [%rd697]; $L__BB106_24: add.s32 %r21, %r10, 1; setp.eq.s32 %p33, %r21, %r9; @%p33 bra $L__BB106_26; bra.uni $L__BB106_25; $L__BB106_26: ld.const.u64 %rd714, [kernel_params+2272]; cvta.to.global.u64 %rd715, %rd714; mul.wide.s32 %rd716, %r6, 16; add.s64 %rd717, %rd715, %rd716; ld.global.v4.f32 {%f15621, %f15620, %f15619, %f4633}, [%rd717]; mul.wide.s32 %rd718, %r5, 16; add.s64 %rd719, %rd715, %rd718; ld.global.v4.f32 {%f15618, %f15617, %f15616, %f4637}, [%rd719]; ld.global.v4.f32 {%f15615, %f15614, %f15613, %f4641}, [%rd719+16]; mul.wide.s32 %rd720, %r7, 16; add.s64 %rd721, %rd715, %rd720; ld.global.v4.f32 {%f15612, %f15611, %f15610, %f4645}, [%rd721]; bra.uni $L__BB106_27; $L__BB106_25: setp.ge.s32 %p34, %r10, %r9; selp.b32 %r985, -1, 0, %p34; add.s32 %r986, %r21, %r985; mad.lo.s32 %r987, %r986, %r8, %r3066; ld.const.u64 %rd706, [kernel_params+2344]; cvta.to.global.u64 %rd707, %rd706; add.s32 %r988, %r987, %r6; mul.wide.s32 %rd708, %r988, 16; add.s64 %rd709, %rd707, %rd708; ld.global.v4.f32 {%f15621, %f15620, %f15619, %f4617}, [%rd709]; add.s32 %r989, %r987, %r5; mul.wide.s32 %rd710, %r989, 16; add.s64 %rd711, %rd707, %rd710; ld.global.v4.f32 {%f15618, %f15617, %f15616, %f4621}, [%rd711]; ld.global.v4.f32 {%f15615, %f15614, %f15613, %f4625}, [%rd711+16]; add.s32 %r990, %r987, %r7; mul.wide.s32 %rd712, %r990, 16; add.s64 %rd713, %rd707, %rd712; ld.global.v4.f32 {%f15612, %f15611, %f15610, %f4629}, [%rd713]; $L__BB106_27: shl.b32 %r3044, %r9, 1; cvt.rn.f32.s32 %f15458, %r3044; mul.ftz.f32 %f15457, %f10, %f15458; cvt.rn.f32.s32 %f4646, %r10; sub.ftz.f32 %f4647, %f15457, %f4646; mov.f32 %f4648, 0f3F800000; sub.ftz.f32 %f4649, %f4648, %f4647; mul.ftz.f32 %f4650, %f4647, %f15621; fma.rn.ftz.f32 %f15633, %f4649, %f15609, %f4650; mul.ftz.f32 %f4651, %f4647, %f15620; fma.rn.ftz.f32 %f15632, %f4649, %f15608, %f4651; mul.ftz.f32 %f4652, %f4647, %f15619; fma.rn.ftz.f32 %f15631, %f4649, %f15607, %f4652; mul.ftz.f32 %f4653, %f4647, %f15618; fma.rn.ftz.f32 %f15630, %f4649, %f15606, %f4653; mul.ftz.f32 %f4654, %f4647, %f15617; fma.rn.ftz.f32 %f15629, %f4649, %f15605, %f4654; mul.ftz.f32 %f4655, %f4647, %f15616; fma.rn.ftz.f32 %f15628, %f4649, %f15604, %f4655; mul.ftz.f32 %f4656, %f4647, %f15615; fma.rn.ftz.f32 %f15627, %f4649, %f15603, %f4656; mul.ftz.f32 %f4657, %f4647, %f15614; fma.rn.ftz.f32 %f15626, %f4649, %f15602, %f4657; mul.ftz.f32 %f4658, %f4647, %f15613; fma.rn.ftz.f32 %f15625, %f4649, %f15601, %f4658; mul.ftz.f32 %f4659, %f4647, %f15612; fma.rn.ftz.f32 %f15624, %f4649, %f15600, %f4659; mul.ftz.f32 %f4660, %f4647, %f15611; fma.rn.ftz.f32 %f15623, %f4649, %f15599, %f4660; mul.ftz.f32 %f4661, %f4647, %f15610; fma.rn.ftz.f32 %f15622, %f4649, %f15598, %f4661; $L__BB106_29: fma.rn.ftz.f32 %f15654, %f15596, %f15597, %f15593; fma.rn.ftz.f32 %f15653, %f15595, %f15597, %f15592; fma.rn.ftz.f32 %f15652, %f15594, %f15597, %f15591; mov.f32 %f4678, 0f3F800000; sub.ftz.f32 %f330, %f4678, %f1; mul.ftz.f32 %f331, %f330, %f330; add.ftz.f32 %f4679, %f330, %f330; mul.ftz.f32 %f4680, %f1, %f4679; sub.ftz.f32 %f4681, %f4680, %f331; mul.ftz.f32 %f4682, %f1, 0f40400000; add.ftz.f32 %f332, %f4682, 0fC0A00000; add.ftz.f32 %f4683, %f1, %f1; mul.ftz.f32 %f4684, %f4683, %f332; fma.rn.ftz.f32 %f4685, %f1, %f4682, %f4684; add.ftz.f32 %f4686, %f4682, 0f40000000; mul.ftz.f32 %f4687, %f4679, %f4686; mul.ftz.f32 %f333, %f330, 0f40400000; mul.ftz.f32 %f4688, %f330, %f333; sub.ftz.f32 %f4689, %f4687, %f4688; mul.ftz.f32 %f4690, %f330, 0fC0000000; mul.ftz.f32 %f334, %f1, %f1; fma.rn.ftz.f32 %f4691, %f1, %f4690, %f334; mul.ftz.f32 %f4692, %f15630, %f4685; mul.ftz.f32 %f4693, %f15629, %f4685; mul.ftz.f32 %f4694, %f15628, %f4685; fma.rn.ftz.f32 %f4695, %f15633, %f4681, %f4692; fma.rn.ftz.f32 %f4696, %f15632, %f4681, %f4693; fma.rn.ftz.f32 %f4697, %f15631, %f4681, %f4694; fma.rn.ftz.f32 %f4698, %f15627, %f4689, %f4695; fma.rn.ftz.f32 %f4699, %f15626, %f4689, %f4696; fma.rn.ftz.f32 %f4700, %f15625, %f4689, %f4697; fma.rn.ftz.f32 %f4701, %f15624, %f4691, %f4698; fma.rn.ftz.f32 %f4702, %f15623, %f4691, %f4699; fma.rn.ftz.f32 %f4703, %f15622, %f4691, %f4700; mul.ftz.f32 %f15646, %f4701, 0f3F000000; mul.ftz.f32 %f15648, %f4702, 0f3F000000; mul.ftz.f32 %f15647, %f4703, 0f3F000000; and.b32 %r22, %r1, 4; setp.eq.s32 %p35, %r22, 0; @%p35 bra $L__BB106_31; bra.uni $L__BB106_30; $L__BB106_31: mul.ftz.f32 %f4762, %f1, %f330; mul.ftz.f32 %f4763, %f330, %f4762; fma.rn.ftz.f32 %f4764, %f334, %f332, 0f40000000; add.ftz.f32 %f4765, %f333, 0fC0A00000; fma.rn.ftz.f32 %f4766, %f331, %f4765, 0f40000000; mul.ftz.f32 %f4767, %f1, %f4762; mul.ftz.f32 %f4768, %f15633, %f4763; mul.ftz.f32 %f4769, %f15632, %f4763; mul.ftz.f32 %f4770, %f15631, %f4763; mul.ftz.f32 %f4771, %f15630, %f4764; mul.ftz.f32 %f4772, %f15629, %f4764; mul.ftz.f32 %f4773, %f15628, %f4764; sub.ftz.f32 %f4774, %f4771, %f4768; sub.ftz.f32 %f4775, %f4772, %f4769; sub.ftz.f32 %f4776, %f4773, %f4770; fma.rn.ftz.f32 %f4777, %f15627, %f4766, %f4774; fma.rn.ftz.f32 %f4778, %f15626, %f4766, %f4775; fma.rn.ftz.f32 %f4779, %f15625, %f4766, %f4776; mul.ftz.f32 %f4780, %f15624, %f4767; mul.ftz.f32 %f4781, %f15623, %f4767; mul.ftz.f32 %f4782, %f15622, %f4767; sub.ftz.f32 %f4783, %f4777, %f4780; sub.ftz.f32 %f4784, %f4778, %f4781; sub.ftz.f32 %f4785, %f4779, %f4782; mul.ftz.f32 %f341, %f4783, 0f3F000000; mul.ftz.f32 %f342, %f4784, 0f3F000000; mul.ftz.f32 %f343, %f4785, 0f3F000000; setp.eq.ftz.f32 %p36, %f15654, %f341; setp.eq.ftz.f32 %p37, %f15653, %f342; and.pred %p38, %p36, %p37; setp.eq.ftz.f32 %p39, %f15652, %f343; and.pred %p40, %p39, %p38; @%p40 bra $L__BB106_33; bra.uni $L__BB106_32; $L__BB106_33: mov.f32 %f15651, %f7; mov.f32 %f15650, %f8; mov.f32 %f15649, %f9; bra.uni $L__BB106_34; $L__BB106_30: mul.ftz.f32 %f4704, %f15648, %f15648; fma.rn.ftz.f32 %f4705, %f15646, %f15646, %f4704; fma.rn.ftz.f32 %f4706, %f15647, %f15647, %f4705; sqrt.approx.ftz.f32 %f4707, %f4706; div.approx.ftz.f32 %f4708, %f15646, %f4707; div.approx.ftz.f32 %f4709, %f15648, %f4707; div.approx.ftz.f32 %f4710, %f15647, %f4707; mul.ftz.f32 %f4711, %f4709, %f15594; mul.ftz.f32 %f4712, %f4710, %f15595; sub.ftz.f32 %f4713, %f4712, %f4711; mul.ftz.f32 %f4714, %f4710, %f15596; mul.ftz.f32 %f4715, %f4708, %f15594; sub.ftz.f32 %f4716, %f4715, %f4714; mul.ftz.f32 %f4717, %f4708, %f15595; mul.ftz.f32 %f4718, %f4709, %f15596; sub.ftz.f32 %f4719, %f4718, %f4717; mul.ftz.f32 %f4720, %f4716, %f4716; fma.rn.ftz.f32 %f4721, %f4713, %f4713, %f4720; fma.rn.ftz.f32 %f4722, %f4719, %f4719, %f4721; sqrt.approx.ftz.f32 %f4723, %f4722; div.approx.ftz.f32 %f4724, %f4713, %f4723; div.approx.ftz.f32 %f4725, %f4716, %f4723; div.approx.ftz.f32 %f4726, %f4719, %f4723; mul.ftz.f32 %f4727, %f2, %f2; sub.ftz.f32 %f4729, %f4678, %f4727; mov.f32 %f4730, 0f00000000; max.ftz.f32 %f4731, %f4729, %f4730; sqrt.approx.ftz.f32 %f4732, %f4731; mul.ftz.f32 %f4733, %f4724, %f2; mul.ftz.f32 %f4734, %f4725, %f2; mul.ftz.f32 %f4735, %f4726, %f2; mul.ftz.f32 %f4736, %f4709, %f4726; mul.ftz.f32 %f4737, %f4710, %f4725; sub.ftz.f32 %f4738, %f4736, %f4737; mul.ftz.f32 %f4739, %f4710, %f4724; mul.ftz.f32 %f4740, %f4708, %f4726; sub.ftz.f32 %f4741, %f4739, %f4740; mul.ftz.f32 %f4742, %f4708, %f4725; mul.ftz.f32 %f4743, %f4709, %f4724; sub.ftz.f32 %f4744, %f4742, %f4743; mul.ftz.f32 %f4745, %f4741, %f4741; fma.rn.ftz.f32 %f4746, %f4738, %f4738, %f4745; fma.rn.ftz.f32 %f4747, %f4744, %f4744, %f4746; sqrt.approx.ftz.f32 %f4748, %f4747; div.approx.ftz.f32 %f4749, %f4738, %f4748; div.approx.ftz.f32 %f4750, %f4741, %f4748; div.approx.ftz.f32 %f4751, %f4744, %f4748; mul.ftz.f32 %f4752, %f4732, %f4749; mul.ftz.f32 %f4753, %f4732, %f4750; mul.ftz.f32 %f4754, %f4732, %f4751; sub.ftz.f32 %f4755, %f4733, %f4752; sub.ftz.f32 %f4756, %f4734, %f4753; sub.ftz.f32 %f4757, %f4735, %f4754; mul.ftz.f32 %f4758, %f4756, %f4756; fma.rn.ftz.f32 %f4759, %f4755, %f4755, %f4758; fma.rn.ftz.f32 %f4760, %f4757, %f4757, %f4759; sqrt.approx.ftz.f32 %f4761, %f4760; div.approx.ftz.f32 %f15651, %f4755, %f4761; div.approx.ftz.f32 %f15650, %f4756, %f4761; div.approx.ftz.f32 %f15649, %f4757, %f4761; st.local.f32 [%rd7+-76], %f15651; st.local.f32 [%rd7+-72], %f15650; st.local.f32 [%rd7+-68], %f15649; bra.uni $L__BB106_35; $L__BB106_32: sub.ftz.f32 %f4786, %f15654, %f341; sub.ftz.f32 %f4787, %f15653, %f342; mul.ftz.f32 %f4788, %f4787, %f4787; fma.rn.ftz.f32 %f4789, %f4786, %f4786, %f4788; sub.ftz.f32 %f4790, %f15652, %f343; fma.rn.ftz.f32 %f4791, %f4790, %f4790, %f4789; sqrt.approx.ftz.f32 %f4792, %f4791; div.approx.ftz.f32 %f15649, %f4790, %f4792; div.approx.ftz.f32 %f15650, %f4787, %f4792; div.approx.ftz.f32 %f15651, %f4786, %f4792; $L__BB106_34: mov.u32 %r3041, 0; st.local.f32 [%rd7+-76], %f15651; st.local.f32 [%rd7+-72], %f15650; st.local.f32 [%rd7+-68], %f15649; st.local.u32 [%rd7+-16], %r3041; $L__BB106_35: and.b32 %r3045, %r3072, 4; setp.ne.s32 %p2181, %r3045, 0; st.local.f32 [%rd7+28], %f15646; st.local.f32 [%rd7+32], %f15648; st.local.f32 [%rd7+36], %f15647; @%p2181 bra $L__BB106_45; @%p15 bra $L__BB106_38; bra.uni $L__BB106_37; $L__BB106_38: ld.const.u64 %rd730, [kernel_params+2160]; cvta.to.global.u64 %rd731, %rd730; mul.lo.s64 %rd732, %rd9, 208; add.s64 %rd733, %rd731, %rd732; ld.global.v4.f32 {%f4802, %f4803, %f4804, %f4805}, [%rd733]; ld.global.v4.f32 {%f4810, %f4811, %f4812, %f4813}, [%rd733+16]; ld.global.v4.f32 {%f4818, %f4819, %f4820, %f4821}, [%rd733+32]; mul.ftz.f32 %f4826, %f15653, %f4803; fma.rn.ftz.f32 %f4827, %f15654, %f4802, %f4826; fma.rn.ftz.f32 %f4828, %f15652, %f4804, %f4827; add.ftz.f32 %f15642, %f4805, %f4828; mul.ftz.f32 %f4829, %f15653, %f4811; fma.rn.ftz.f32 %f4830, %f15654, %f4810, %f4829; fma.rn.ftz.f32 %f4831, %f15652, %f4812, %f4830; add.ftz.f32 %f15641, %f4813, %f4831; mul.ftz.f32 %f4832, %f15653, %f4819; fma.rn.ftz.f32 %f4833, %f15654, %f4818, %f4832; fma.rn.ftz.f32 %f4834, %f15652, %f4820, %f4833; add.ftz.f32 %f15652, %f4821, %f4834; bra.uni $L__BB106_39; $L__BB106_37: mul.ftz.f32 %f4793, %f15653, %f142; fma.rn.ftz.f32 %f4794, %f15654, %f141, %f4793; fma.rn.ftz.f32 %f4795, %f15652, %f143, %f4794; add.ftz.f32 %f15642, %f144, %f4795; mul.ftz.f32 %f4796, %f15653, %f146; fma.rn.ftz.f32 %f4797, %f15654, %f145, %f4796; fma.rn.ftz.f32 %f4798, %f15652, %f147, %f4797; add.ftz.f32 %f15641, %f148, %f4798; mul.ftz.f32 %f4799, %f15653, %f150; fma.rn.ftz.f32 %f4800, %f15654, %f149, %f4799; fma.rn.ftz.f32 %f4801, %f15652, %f151, %f4800; add.ftz.f32 %f15652, %f152, %f4801; $L__BB106_39: @%p15 bra $L__BB106_41; bra.uni $L__BB106_40; $L__BB106_41: ld.const.u64 %rd734, [kernel_params+2160]; cvta.to.global.u64 %rd735, %rd734; mul.lo.s64 %rd736, %rd9, 208; add.s64 %rd737, %rd735, %rd736; ld.global.v4.f32 {%f4848, %f4849, %f4850, %f4851}, [%rd737+48]; ld.global.v4.f32 {%f4855, %f4856, %f4857, %f4858}, [%rd737+64]; ld.global.v4.f32 {%f4862, %f4863, %f4864, %f4865}, [%rd737+80]; mul.ftz.f32 %f4869, %f4855, %f15650; fma.rn.ftz.f32 %f4870, %f4848, %f15651, %f4869; fma.rn.ftz.f32 %f4871, %f4862, %f15649, %f4870; mul.ftz.f32 %f4872, %f4856, %f15650; fma.rn.ftz.f32 %f4873, %f4849, %f15651, %f4872; fma.rn.ftz.f32 %f4874, %f4863, %f15649, %f4873; mul.ftz.f32 %f4875, %f4857, %f15650; fma.rn.ftz.f32 %f4876, %f4850, %f15651, %f4875; fma.rn.ftz.f32 %f4877, %f4864, %f15649, %f4876; mul.ftz.f32 %f4878, %f4874, %f4874; fma.rn.ftz.f32 %f4879, %f4871, %f4871, %f4878; fma.rn.ftz.f32 %f4880, %f4877, %f4877, %f4879; sqrt.approx.ftz.f32 %f4881, %f4880; div.approx.ftz.f32 %f15651, %f4871, %f4881; div.approx.ftz.f32 %f15650, %f4874, %f4881; div.approx.ftz.f32 %f15649, %f4877, %f4881; bra.uni $L__BB106_42; $L__BB106_40: mul.ftz.f32 %f4835, %f15650, %f15568; fma.rn.ftz.f32 %f4836, %f15651, %f15564, %f4835; fma.rn.ftz.f32 %f4837, %f15649, %f15572, %f4836; mul.ftz.f32 %f4838, %f15650, %f15569; fma.rn.ftz.f32 %f4839, %f15651, %f15565, %f4838; fma.rn.ftz.f32 %f4840, %f15649, %f15573, %f4839; mul.ftz.f32 %f4841, %f15650, %f15570; fma.rn.ftz.f32 %f4842, %f15651, %f15566, %f4841; fma.rn.ftz.f32 %f4843, %f15649, %f15574, %f4842; mul.ftz.f32 %f4844, %f4840, %f4840; fma.rn.ftz.f32 %f4845, %f4837, %f4837, %f4844; fma.rn.ftz.f32 %f4846, %f4843, %f4843, %f4845; sqrt.approx.ftz.f32 %f4847, %f4846; div.approx.ftz.f32 %f15651, %f4837, %f4847; div.approx.ftz.f32 %f15650, %f4840, %f4847; div.approx.ftz.f32 %f15649, %f4843, %f4847; $L__BB106_42: st.local.f32 [%rd7+-76], %f15651; st.local.f32 [%rd7+-72], %f15650; st.local.f32 [%rd7+-68], %f15649; @%p15 bra $L__BB106_44; bra.uni $L__BB106_43; $L__BB106_44: ld.const.u64 %rd738, [kernel_params+2160]; cvta.to.global.u64 %rd739, %rd738; mul.lo.s64 %rd740, %rd9, 208; add.s64 %rd741, %rd739, %rd740; ld.global.v4.f32 {%f4888, %f4889, %f4890, %f4891}, [%rd741]; ld.global.v4.f32 {%f4895, %f4896, %f4897, %f4898}, [%rd741+16]; ld.global.v4.f32 {%f4902, %f4903, %f4904, %f4905}, [%rd741+32]; mul.ftz.f32 %f4909, %f4889, %f15648; fma.rn.ftz.f32 %f4910, %f15646, %f4888, %f4909; fma.rn.ftz.f32 %f380, %f4890, %f15647, %f4910; mul.ftz.f32 %f4911, %f4896, %f15648; fma.rn.ftz.f32 %f4912, %f15646, %f4895, %f4911; fma.rn.ftz.f32 %f381, %f4897, %f15647, %f4912; mul.ftz.f32 %f4913, %f4903, %f15648; fma.rn.ftz.f32 %f4914, %f15646, %f4902, %f4913; fma.rn.ftz.f32 %f15647, %f4904, %f15647, %f4914; st.local.f32 [%rd7+28], %f380; st.local.f32 [%rd7+32], %f381; st.local.f32 [%rd7+36], %f15647; mov.f32 %f15646, %f380; mov.f32 %f15648, %f381; mov.f32 %f15653, %f15641; mov.f32 %f15654, %f15642; bra.uni $L__BB106_45; $L__BB106_43: mul.ftz.f32 %f4882, %f15648, %f142; fma.rn.ftz.f32 %f4883, %f15646, %f141, %f4882; fma.rn.ftz.f32 %f377, %f15647, %f143, %f4883; mul.ftz.f32 %f4884, %f15648, %f146; fma.rn.ftz.f32 %f4885, %f15646, %f145, %f4884; fma.rn.ftz.f32 %f378, %f15647, %f147, %f4885; mul.ftz.f32 %f4886, %f15648, %f150; fma.rn.ftz.f32 %f4887, %f15646, %f149, %f4886; fma.rn.ftz.f32 %f15647, %f15647, %f151, %f4887; st.local.f32 [%rd7+28], %f377; st.local.f32 [%rd7+32], %f378; st.local.f32 [%rd7+36], %f15647; mov.f32 %f15646, %f377; mov.f32 %f15648, %f378; mov.f32 %f15653, %f15641; mov.f32 %f15654, %f15642; $L__BB106_45: neg.ftz.f32 %f15456, %f7; neg.ftz.f32 %f15455, %f8; neg.ftz.f32 %f15454, %f9; st.local.v2.f32 [%rd1], {%f15654, %f15653}; st.local.f32 [%rd7+-80], %f15652; selp.f32 %f4915, %f15649, %f15454, %p35; selp.f32 %f15692, %f15650, %f15455, %p35; selp.f32 %f15691, %f15651, %f15456, %p35; st.local.v2.f32 [%rd7+-64], {%f15691, %f15692}; st.local.f32 [%rd7+-56], %f4915; mul.ftz.f32 %f4916, %f15648, %f4915; mul.ftz.f32 %f4917, %f15647, %f15692; mul.ftz.f32 %f4918, %f15647, %f15691; mul.ftz.f32 %f4919, %f15646, %f4915; mul.ftz.f32 %f4920, %f15646, %f15692; mul.ftz.f32 %f4921, %f15648, %f15691; sub.ftz.f32 %f4922, %f4920, %f4921; sub.ftz.f32 %f4923, %f4918, %f4919; sub.ftz.f32 %f4924, %f4916, %f4917; st.local.v2.f32 [%rd7+40], {%f4924, %f4923}; st.local.f32 [%rd7+48], %f4922; ld.global.u32 %r992, [%rd11+-4]; st.local.u32 [%rd7+-40], %r992; bra.uni $L__BB106_95; $L__BB106_68: { // callseq 83, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 83 ld.local.u32 %r3072, [%rd7+-32]; bra.uni $L__BB106_81; $L__BB106_48: ld.const.u64 %rd750, [kernel_params+2160]; cvta.to.global.u64 %rd751, %rd750; mul.lo.s64 %rd752, %rd9, 208; add.s64 %rd753, %rd751, %rd752; ld.global.u32 %r23, [%rd753+148]; ld.global.u32 %r24, [%rd753+152]; shl.b32 %r996, %r24, 1; cvt.rn.f32.s32 %f4928, %r996; mul.ftz.f32 %f394, %f10, %f4928; cvt.rzi.ftz.s32.f32 %r997, %f394; add.s32 %r998, %r996, -1; min.s32 %r25, %r997, %r998; ld.const.u64 %rd754, [kernel_params+2312]; cvta.to.global.u64 %rd16, %rd754; ld.global.u32 %r3068, [%rd753+164]; mul.wide.u32 %rd755, %r3068, 16; add.s64 %rd756, %rd16, %rd755; ld.global.u16 %rs135, [%rd756+12]; ld.global.u32 %r3067, [%rd756+8]; ld.global.u64 %rd2926, [%rd756]; setp.eq.s64 %p48, %rd2926, 11; @%p48 bra $L__BB106_53; $L__BB106_49: setp.eq.s64 %p49, %rd2926, 0; @%p49 bra $L__BB106_51; add.s32 %r3068, %r3068, 2; bra.uni $L__BB106_52; $L__BB106_51: setp.eq.s16 %p50, %rs135, 0; mov.u32 %r3071, -1; mov.u32 %r3068, %r3067; @%p50 bra $L__BB106_54; $L__BB106_52: mul.wide.u32 %rd757, %r3068, 16; add.s64 %rd758, %rd16, %rd757; ld.global.u16 %rs135, [%rd758+12]; ld.global.u32 %r3067, [%rd758+8]; ld.global.u64 %rd2926, [%rd758]; setp.ne.s64 %p51, %rd2926, 11; @%p51 bra $L__BB106_49; $L__BB106_53: setp.eq.s16 %p52, %rs135, 0; selp.b32 %r3071, -1, %r3067, %p52; $L__BB106_54: setp.eq.s32 %p53, %r25, %r24; @%p53 bra $L__BB106_56; bra.uni $L__BB106_55; $L__BB106_56: ld.const.u64 %rd762, [kernel_params+2296]; cvta.to.global.u64 %rd763, %rd762; shl.b64 %rd764, %rd3, 4; add.s64 %rd2927, %rd763, %rd764; bra.uni $L__BB106_57; $L__BB106_70: sub.ftz.f32 %f5017, %f434, %f435; sub.ftz.f32 %f5018, %f436, %f437; mul.ftz.f32 %f5019, %f5018, %f5018; fma.rn.ftz.f32 %f5020, %f5017, %f5017, %f5019; sub.ftz.f32 %f5021, %f438, %f439; fma.rn.ftz.f32 %f5022, %f5021, %f5021, %f5020; sqrt.approx.ftz.f32 %f5023, %f5022; div.approx.ftz.f32 %f15663, %f5021, %f5023; div.approx.ftz.f32 %f15662, %f5018, %f5023; div.approx.ftz.f32 %f15661, %f5017, %f5023; $L__BB106_72: ld.const.u64 %rd795, [kernel_params+2216]; cvta.to.global.u64 %rd796, %rd795; shl.b64 %rd797, %rd3, 2; add.s64 %rd798, %rd796, %rd797; ld.global.u32 %r38, [%rd798]; st.local.u32 [%rd7+-40], %r38; fma.rn.ftz.f32 %f5031, %f1, %f431, %f425; fma.rn.ftz.f32 %f5032, %f1, %f432, %f427; fma.rn.ftz.f32 %f5033, %f1, %f433, %f429; fma.rn.ftz.f32 %f15676, %f2, %f426, %f5031; fma.rn.ftz.f32 %f15677, %f2, %f428, %f5032; fma.rn.ftz.f32 %f15678, %f2, %f430, %f5033; and.b32 %r1017, %r3072, 4; setp.ne.s32 %p61, %r1017, 0; @%p61 bra $L__BB106_76; and.b32 %r3043, %r3072, 2; setp.ne.s32 %p62, %r3043, 0; @%p62 bra $L__BB106_75; ld.const.u64 %rd799, [kernel_params+2160]; cvta.to.global.u64 %rd800, %rd799; mul.lo.s64 %rd801, %rd9, 208; add.s64 %rd802, %rd800, %rd801; ld.global.v4.f32 {%f141, %f142, %f143, %f144}, [%rd802]; ld.global.v4.f32 {%f145, %f146, %f147, %f148}, [%rd802+16]; ld.global.v4.f32 {%f149, %f150, %f151, %f152}, [%rd802+32]; $L__BB106_75: mul.ftz.f32 %f5046, %f15677, %f142; fma.rn.ftz.f32 %f5047, %f15676, %f141, %f5046; fma.rn.ftz.f32 %f5048, %f15678, %f143, %f5047; mul.ftz.f32 %f5049, %f15677, %f146; fma.rn.ftz.f32 %f5050, %f15676, %f145, %f5049; fma.rn.ftz.f32 %f5051, %f15678, %f147, %f5050; mul.ftz.f32 %f5052, %f15677, %f150; fma.rn.ftz.f32 %f5053, %f15676, %f149, %f5052; fma.rn.ftz.f32 %f5054, %f15678, %f151, %f5053; add.ftz.f32 %f15676, %f144, %f5048; add.ftz.f32 %f15677, %f148, %f5051; add.ftz.f32 %f15678, %f152, %f5054; $L__BB106_76: st.local.f32 [%rd7+-56], %f15663; st.local.v4.f32 [%rd1], {%f15676, %f15677, %f15678, %f15661}; st.local.v4.f32 [%rd7+-72], {%f15662, %f15663, %f15661, %f15662}; setp.gt.s32 %p63, %r38, -1; @%p63 bra $L__BB106_80; ld.global.v4.u32 {%r1018, %r1019, %r1020, %r1021}, [%rd27+-12]; ld.const.u64 %rd805, [kernel_params+2224]; cvta.to.global.u64 %rd806, %rd805; mul.wide.u32 %rd807, %r1018, 12; add.s64 %rd808, %rd806, %rd807; mul.wide.u32 %rd809, %r1019, 12; add.s64 %rd810, %rd806, %rd809; mul.wide.u32 %rd811, %r1020, 12; add.s64 %rd812, %rd806, %rd811; mov.f32 %f5055, 0f3F800000; sub.ftz.f32 %f5056, %f5055, %f1; sub.ftz.f32 %f5057, %f5056, %f2; ld.global.f32 %f5058, [%rd808]; ld.global.f32 %f5059, [%rd808+4]; ld.global.f32 %f5060, [%rd808+8]; ld.global.f32 %f5061, [%rd810]; mul.ftz.f32 %f5062, %f1, %f5061; ld.global.f32 %f5063, [%rd810+4]; mul.ftz.f32 %f5064, %f1, %f5063; ld.global.f32 %f5065, [%rd810+8]; mul.ftz.f32 %f5066, %f1, %f5065; fma.rn.ftz.f32 %f5067, %f5057, %f5058, %f5062; fma.rn.ftz.f32 %f5068, %f5057, %f5059, %f5064; fma.rn.ftz.f32 %f5069, %f5057, %f5060, %f5066; ld.global.f32 %f5070, [%rd812]; ld.global.f32 %f5071, [%rd812+4]; ld.global.f32 %f5072, [%rd812+8]; fma.rn.ftz.f32 %f15679, %f2, %f5070, %f5067; fma.rn.ftz.f32 %f15680, %f2, %f5071, %f5068; fma.rn.ftz.f32 %f15681, %f2, %f5072, %f5069; mul.ftz.f32 %f5073, %f15680, %f15680; fma.rn.ftz.f32 %f5074, %f15679, %f15679, %f5073; fma.rn.ftz.f32 %f5075, %f15681, %f15681, %f5074; sqrt.approx.ftz.f32 %f488, %f5075; setp.neu.ftz.f32 %p64, %f488, 0f00000000; @%p64 bra $L__BB106_78; bra.uni $L__BB106_79; $L__BB106_78: rcp.approx.ftz.f32 %f5076, %f488; mul.ftz.f32 %f15681, %f15681, %f5076; mul.ftz.f32 %f15680, %f15680, %f5076; mul.ftz.f32 %f15679, %f15679, %f5076; $L__BB106_79: setp.eq.ftz.f32 %p65, %f15679, 0f00000000; setp.eq.ftz.f32 %p66, %f15680, 0f00000000; and.pred %p67, %p65, %p66; setp.eq.ftz.f32 %p68, %f15681, 0f00000000; and.pred %p69, %p68, %p67; selp.f32 %f5077, %f15663, %f15681, %p69; selp.f32 %f5078, %f15662, %f15680, %p69; selp.f32 %f5079, %f15661, %f15679, %p69; st.local.f32 [%rd7+-76], %f5079; st.local.f32 [%rd7+-72], %f5078; st.local.f32 [%rd7+-68], %f5077; $L__BB106_80: st.local.f32 [%rd7+28], %f431; st.local.f32 [%rd7+32], %f432; st.local.f32 [%rd7+36], %f433; st.local.v2.f32 [%rd7+40], {%f426, %f428}; st.local.f32 [%rd7+48], %f430; $L__BB106_81: and.b32 %r1025, %r3072, 4; setp.eq.s32 %p70, %r1025, 0; @%p70 bra $L__BB106_83; bra.uni $L__BB106_82; $L__BB106_83: and.b32 %r40, %r3072, 2; setp.eq.s32 %p71, %r40, 0; @%p71 bra $L__BB106_85; bra.uni $L__BB106_84; $L__BB106_85: ld.const.u64 %rd813, [kernel_params+2160]; cvta.to.global.u64 %rd814, %rd813; ld.local.u32 %r1026, [%rd7+-12]; mul.wide.s32 %rd815, %r1026, 208; add.s64 %rd816, %rd814, %rd815; ld.global.v4.f32 {%f5121, %f5122, %f5123, %f5124}, [%rd816+48]; ld.global.v4.f32 {%f5128, %f5129, %f5130, %f5131}, [%rd816+64]; ld.global.v4.f32 {%f5135, %f5136, %f5137, %f5138}, [%rd816+80]; ld.local.v2.f32 {%f5142, %f5143}, [%rd7+-72]; ld.local.f32 %f5146, [%rd7+-76]; mul.ftz.f32 %f5147, %f5128, %f5142; fma.rn.ftz.f32 %f5148, %f5121, %f5146, %f5147; fma.rn.ftz.f32 %f5149, %f5135, %f5143, %f5148; mul.ftz.f32 %f5150, %f5129, %f5142; fma.rn.ftz.f32 %f5151, %f5122, %f5146, %f5150; fma.rn.ftz.f32 %f5152, %f5136, %f5143, %f5151; mul.ftz.f32 %f5153, %f5130, %f5142; fma.rn.ftz.f32 %f5154, %f5123, %f5146, %f5153; fma.rn.ftz.f32 %f5155, %f5137, %f5143, %f5154; mul.ftz.f32 %f5156, %f5152, %f5152; fma.rn.ftz.f32 %f5157, %f5149, %f5149, %f5156; fma.rn.ftz.f32 %f5158, %f5155, %f5155, %f5157; sqrt.approx.ftz.f32 %f5159, %f5158; div.approx.ftz.f32 %f15682, %f5149, %f5159; div.approx.ftz.f32 %f15683, %f5152, %f5159; div.approx.ftz.f32 %f15684, %f5155, %f5159; bra.uni $L__BB106_86; $L__BB106_82: ld.local.v2.f32 {%f15691, %f15692}, [%rd7+-64]; bra.uni $L__BB106_95; $L__BB106_55: cvt.u32.u64 %r1000, %rd3; setp.gt.s32 %p54, %r25, %r24; selp.b32 %r1001, -1, 0, %p54; add.s32 %r1002, %r25, %r1001; ld.const.u64 %rd759, [kernel_params+2344]; cvta.to.global.u64 %rd760, %rd759; mad.lo.s32 %r1003, %r1002, %r23, %r1000; add.s32 %r1004, %r1003, %r3071; mul.wide.s32 %rd761, %r1004, 16; add.s64 %rd2927, %rd760, %rd761; $L__BB106_57: ld.global.v4.f32 {%f4929, %f4930, %f4931, %f4932}, [%rd2927]; add.s32 %r36, %r25, 1; setp.eq.s32 %p55, %r36, %r24; @%p55 bra $L__BB106_59; bra.uni $L__BB106_58; $L__BB106_59: ld.const.u64 %rd768, [kernel_params+2296]; cvta.to.global.u64 %rd769, %rd768; shl.b64 %rd770, %rd3, 4; add.s64 %rd2928, %rd769, %rd770; bra.uni $L__BB106_60; $L__BB106_84: ld.local.v2.f32 {%f5082, %f5083}, [%rd7+-72]; ld.local.v4.f32 {%f5086, %f5087, %f5088, %f5089}, [%rd7+104]; ld.local.v4.f32 {%f5093, %f5094, %f5095, %f5096}, [%rd7+120]; ld.local.v4.f32 {%f5100, %f5101, %f5102, %f5103}, [%rd7+136]; ld.local.f32 %f5107, [%rd7+-76]; mul.ftz.f32 %f5108, %f5082, %f5093; fma.rn.ftz.f32 %f5109, %f5107, %f5086, %f5108; fma.rn.ftz.f32 %f5110, %f5083, %f5100, %f5109; mul.ftz.f32 %f5111, %f5082, %f5094; fma.rn.ftz.f32 %f5112, %f5107, %f5087, %f5111; fma.rn.ftz.f32 %f5113, %f5083, %f5101, %f5112; mul.ftz.f32 %f5114, %f5082, %f5095; fma.rn.ftz.f32 %f5115, %f5107, %f5088, %f5114; fma.rn.ftz.f32 %f5116, %f5083, %f5102, %f5115; mul.ftz.f32 %f5117, %f5113, %f5113; fma.rn.ftz.f32 %f5118, %f5110, %f5110, %f5117; fma.rn.ftz.f32 %f5119, %f5116, %f5116, %f5118; sqrt.approx.ftz.f32 %f5120, %f5119; div.approx.ftz.f32 %f15682, %f5110, %f5120; div.approx.ftz.f32 %f15683, %f5113, %f5120; div.approx.ftz.f32 %f15684, %f5116, %f5120; $L__BB106_86: st.local.f32 [%rd7+-76], %f15682; st.local.f32 [%rd7+-72], %f15683; st.local.f32 [%rd7+-68], %f15684; @%p71 bra $L__BB106_88; bra.uni $L__BB106_87; $L__BB106_88: ld.const.u64 %rd817, [kernel_params+2160]; cvta.to.global.u64 %rd818, %rd817; ld.local.u32 %r1027, [%rd7+-12]; mul.wide.s32 %rd819, %r1027, 208; add.s64 %rd820, %rd818, %rd819; ld.global.v4.f32 {%f5199, %f5200, %f5201, %f5202}, [%rd820+48]; ld.global.v4.f32 {%f5206, %f5207, %f5208, %f5209}, [%rd820+64]; ld.global.v4.f32 {%f5213, %f5214, %f5215, %f5216}, [%rd820+80]; ld.local.v2.f32 {%f5220, %f5221}, [%rd7+-64]; mul.ftz.f32 %f5224, %f5206, %f5221; fma.rn.ftz.f32 %f5225, %f5199, %f5220, %f5224; ld.local.f32 %f5226, [%rd7+-56]; fma.rn.ftz.f32 %f5227, %f5213, %f5226, %f5225; mul.ftz.f32 %f5228, %f5207, %f5221; fma.rn.ftz.f32 %f5229, %f5200, %f5220, %f5228; fma.rn.ftz.f32 %f5230, %f5214, %f5226, %f5229; mul.ftz.f32 %f5231, %f5208, %f5221; fma.rn.ftz.f32 %f5232, %f5201, %f5220, %f5231; fma.rn.ftz.f32 %f5233, %f5215, %f5226, %f5232; mul.ftz.f32 %f5234, %f5230, %f5230; fma.rn.ftz.f32 %f5235, %f5227, %f5227, %f5234; fma.rn.ftz.f32 %f5236, %f5233, %f5233, %f5235; sqrt.approx.ftz.f32 %f5237, %f5236; div.approx.ftz.f32 %f15691, %f5227, %f5237; div.approx.ftz.f32 %f15692, %f5230, %f5237; div.approx.ftz.f32 %f15687, %f5233, %f5237; bra.uni $L__BB106_89; $L__BB106_58: cvt.u32.u64 %r1005, %rd3; setp.ge.s32 %p56, %r25, %r24; selp.b32 %r1006, -1, 0, %p56; add.s32 %r1007, %r36, %r1006; ld.const.u64 %rd765, [kernel_params+2344]; cvta.to.global.u64 %rd766, %rd765; mad.lo.s32 %r1008, %r1007, %r23, %r1005; add.s32 %r1009, %r1008, %r3071; mul.wide.s32 %rd767, %r1009, 16; add.s64 %rd2928, %rd766, %rd767; $L__BB106_60: cvt.rn.f32.s32 %f4933, %r25; sub.ftz.f32 %f4934, %f394, %f4933; ld.global.v4.f32 {%f4935, %f4936, %f4937, %f4938}, [%rd2928]; mov.f32 %f4943, 0f3F800000; sub.ftz.f32 %f4944, %f4943, %f4934; mul.ftz.f32 %f4945, %f4934, %f4935; mul.ftz.f32 %f4946, %f4934, %f4936; mul.ftz.f32 %f4947, %f4934, %f4937; fma.rn.ftz.f32 %f15660, %f4944, %f4931, %f4947; fma.rn.ftz.f32 %f15659, %f4944, %f4930, %f4946; fma.rn.ftz.f32 %f15658, %f4944, %f4929, %f4945; $L__BB106_62: and.b32 %r1010, %r3072, 4; setp.ne.s32 %p57, %r1010, 0; @%p57 bra $L__BB106_66; @%p15 bra $L__BB106_65; bra.uni $L__BB106_64; $L__BB106_65: ld.const.u64 %rd775, [kernel_params+2160]; cvta.to.global.u64 %rd776, %rd775; mul.lo.s64 %rd777, %rd9, 208; add.s64 %rd778, %rd776, %rd777; ld.global.v4.f32 {%f4962, %f4963, %f4964, %f4965}, [%rd778]; ld.global.v4.f32 {%f4970, %f4971, %f4972, %f4973}, [%rd778+16]; ld.global.v4.f32 {%f4978, %f4979, %f4980, %f4981}, [%rd778+32]; mul.ftz.f32 %f4986, %f15659, %f4963; fma.rn.ftz.f32 %f4987, %f15658, %f4962, %f4986; fma.rn.ftz.f32 %f4988, %f15660, %f4964, %f4987; add.ftz.f32 %f417, %f4965, %f4988; mul.ftz.f32 %f4989, %f15659, %f4971; fma.rn.ftz.f32 %f4990, %f15658, %f4970, %f4989; fma.rn.ftz.f32 %f4991, %f15660, %f4972, %f4990; add.ftz.f32 %f418, %f4973, %f4991; mul.ftz.f32 %f4992, %f15659, %f4979; fma.rn.ftz.f32 %f4993, %f15658, %f4978, %f4992; fma.rn.ftz.f32 %f4994, %f15660, %f4980, %f4993; add.ftz.f32 %f15660, %f4981, %f4994; mov.f32 %f15658, %f417; mov.f32 %f15659, %f418; bra.uni $L__BB106_66; $L__BB106_87: ld.local.v2.f32 {%f5160, %f5161}, [%rd7+-64]; ld.local.v4.f32 {%f5164, %f5165, %f5166, %f5167}, [%rd7+104]; ld.local.v4.f32 {%f5171, %f5172, %f5173, %f5174}, [%rd7+120]; ld.local.v4.f32 {%f5178, %f5179, %f5180, %f5181}, [%rd7+136]; mul.ftz.f32 %f5185, %f5161, %f5171; fma.rn.ftz.f32 %f5186, %f5160, %f5164, %f5185; ld.local.f32 %f5187, [%rd7+-56]; fma.rn.ftz.f32 %f5188, %f5187, %f5178, %f5186; mul.ftz.f32 %f5189, %f5161, %f5172; fma.rn.ftz.f32 %f5190, %f5160, %f5165, %f5189; fma.rn.ftz.f32 %f5191, %f5187, %f5179, %f5190; mul.ftz.f32 %f5192, %f5161, %f5173; fma.rn.ftz.f32 %f5193, %f5160, %f5166, %f5192; fma.rn.ftz.f32 %f5194, %f5187, %f5180, %f5193; mul.ftz.f32 %f5195, %f5191, %f5191; fma.rn.ftz.f32 %f5196, %f5188, %f5188, %f5195; fma.rn.ftz.f32 %f5197, %f5194, %f5194, %f5196; sqrt.approx.ftz.f32 %f5198, %f5197; div.approx.ftz.f32 %f15691, %f5188, %f5198; div.approx.ftz.f32 %f15692, %f5191, %f5198; div.approx.ftz.f32 %f15687, %f5194, %f5198; $L__BB106_89: st.local.v2.f32 [%rd7+-64], {%f15691, %f15692}; st.local.f32 [%rd7+-56], %f15687; @%p71 bra $L__BB106_91; bra.uni $L__BB106_90; $L__BB106_91: ld.const.u64 %rd821, [kernel_params+2160]; cvta.to.global.u64 %rd822, %rd821; ld.local.u32 %r1028, [%rd7+-12]; mul.wide.s32 %rd823, %r1028, 208; add.s64 %rd824, %rd822, %rd823; ld.global.v4.f32 {%f5270, %f5271, %f5272, %f5273}, [%rd824]; ld.global.v4.f32 {%f5277, %f5278, %f5279, %f5280}, [%rd824+16]; ld.global.v4.f32 {%f5284, %f5285, %f5286, %f5287}, [%rd824+32]; ld.local.v2.f32 {%f5291, %f5292}, [%rd7+32]; ld.local.f32 %f5295, [%rd7+28]; mul.ftz.f32 %f5296, %f5271, %f5291; fma.rn.ftz.f32 %f5297, %f5295, %f5270, %f5296; fma.rn.ftz.f32 %f15688, %f5272, %f5292, %f5297; mul.ftz.f32 %f5298, %f5278, %f5291; fma.rn.ftz.f32 %f5299, %f5295, %f5277, %f5298; fma.rn.ftz.f32 %f15689, %f5279, %f5292, %f5299; mul.ftz.f32 %f5300, %f5285, %f5291; fma.rn.ftz.f32 %f5301, %f5295, %f5284, %f5300; fma.rn.ftz.f32 %f15690, %f5286, %f5292, %f5301; bra.uni $L__BB106_92; $L__BB106_64: mul.ftz.f32 %f4953, %f15659, %f142; fma.rn.ftz.f32 %f4954, %f15658, %f141, %f4953; fma.rn.ftz.f32 %f4955, %f15660, %f143, %f4954; add.ftz.f32 %f414, %f144, %f4955; mul.ftz.f32 %f4956, %f15659, %f146; fma.rn.ftz.f32 %f4957, %f15658, %f145, %f4956; fma.rn.ftz.f32 %f4958, %f15660, %f147, %f4957; add.ftz.f32 %f415, %f148, %f4958; mul.ftz.f32 %f4959, %f15659, %f150; fma.rn.ftz.f32 %f4960, %f15658, %f149, %f4959; fma.rn.ftz.f32 %f4961, %f15660, %f151, %f4960; add.ftz.f32 %f15660, %f152, %f4961; mov.f32 %f15658, %f414; mov.f32 %f15659, %f415; $L__BB106_66: ld.local.v4.f32 {%f4995, %f4996, %f4997, %f4998}, [%rd1]; sub.ftz.f32 %f5002, %f4995, %f15658; sub.ftz.f32 %f5003, %f4996, %f15659; sub.ftz.f32 %f5004, %f4997, %f15660; mul.ftz.f32 %f5005, %f5003, %f5003; fma.rn.ftz.f32 %f5006, %f5002, %f5002, %f5005; fma.rn.ftz.f32 %f5007, %f5004, %f5004, %f5006; sqrt.approx.ftz.f32 %f5008, %f5007; div.approx.ftz.f32 %f5009, %f5004, %f5008; st.local.f32 [%rd7+-56], %f5009; div.approx.ftz.f32 %f15691, %f5002, %f5008; st.local.f32 [%rd7+-76], %f15691; div.approx.ftz.f32 %f15692, %f5003, %f5008; st.local.v4.f32 [%rd7+-72], {%f15692, %f5009, %f15691, %f15692}; mov.u32 %r1011, 0; st.local.u32 [%rd7+28], %r1011; mov.u64 %rd781, 0; st.local.u64 [%rd7+32], %rd781; mov.f32 %f5010, 0f00000000; st.local.v2.f32 [%rd7+40], {%f5010, %f5010}; st.local.u32 [%rd7+48], %r1011; $L__BB106_95: ld.const.u64 %rd829, [kernel_params+2408]; cvta.to.global.u64 %rd28, %rd829; ld.local.u32 %rd830, [%rd7+-40]; cvt.u32.u64 %r1030, %rd830; and.b32 %r3074, %r1030, 4194303; shl.b64 %rd831, %rd830, 5; and.b64 %rd832, %rd831, 134217696; add.s64 %rd833, %rd28, %rd832; ld.global.u32 %r3134, [%rd833+16]; st.local.u32 [%rd7+-36], %r3134; ld.local.v4.f32 {%f5376, %f5377, %f5378, %f5379}, [%rd7+-56]; mul.ftz.f32 %f5380, %f15692, %f5378; fma.rn.ftz.f32 %f5381, %f15691, %f5377, %f5380; fma.rn.ftz.f32 %f5382, %f5376, %f5379, %f5381; setp.lt.ftz.f32 %p75, %f5382, 0f00000000; @%p75 bra $L__BB106_97; bra.uni $L__BB106_96; $L__BB106_97: or.b32 %r3134, %r3134, 1; st.local.u32 [%rd7+-36], %r3134; neg.ftz.f32 %f15695, %f5376; neg.ftz.f32 %f15692, %f15692; neg.ftz.f32 %f15691, %f15691; st.local.f32 [%rd7+-56], %f15695; ld.local.f32 %f5387, [%rd7+-76]; neg.ftz.f32 %f5388, %f5387; ld.local.v2.f32 {%f5389, %f5390}, [%rd7+-72]; st.local.f32 [%rd7+-76], %f5388; neg.ftz.f32 %f5393, %f5390; neg.ftz.f32 %f5394, %f5389; st.local.v4.f32 [%rd7+-72], {%f5394, %f5393, %f15691, %f15692}; ld.local.f32 %f5395, [%rd7+28]; neg.ftz.f32 %f15707, %f5395; ld.local.v2.f32 {%f5396, %f5397}, [%rd7+32]; neg.ftz.f32 %f15708, %f5396; neg.ftz.f32 %f5400, %f5397; st.local.f32 [%rd7+28], %f15707; st.local.f32 [%rd7+32], %f15708; st.local.f32 [%rd7+36], %f5400; ld.local.v4.f32 {%f5401, %f5402, %f5403, %f5404}, [%rd7+40]; neg.ftz.f32 %f15698, %f5403; neg.ftz.f32 %f15706, %f5402; neg.ftz.f32 %f15705, %f5401; st.local.v2.f32 [%rd7+40], {%f15705, %f15706}; st.local.f32 [%rd7+48], %f15698; bra.uni $L__BB106_98; $L__BB106_96: ld.local.f32 %f15707, [%rd7+28]; ld.local.f32 %f15708, [%rd7+32]; ld.local.v4.f32 {%f15705, %f15706, %f15698, %f5386}, [%rd7+40]; mov.b32 %r1031, %f5376; mov.b64 %rd834, {%r1031, %r1032}; cvt.u32.u64 %r1033, %rd834; mov.b32 %f15695, %r1033; $L__BB106_98: ld.local.f32 %f564, [%rd7]; fma.rn.ftz.f32 %f5408, %f12, %f564, %f11; st.local.f32 [%rd7+4], %f5408; st.local.f32 [%rd7+8], %f12; setp.neu.ftz.f32 %p76, %f15691, %f15692; setp.neu.ftz.f32 %p77, %f15691, %f15695; or.pred %p78, %p76, %p77; sub.ftz.f32 %f5409, %f15695, %f15692; neg.ftz.f32 %f5410, %f15692; neg.ftz.f32 %f5411, %f15695; selp.f32 %f5412, %f5411, %f15695, %p78; add.ftz.f32 %f5413, %f15691, %f5412; selp.f32 %f5414, %f15692, %f5410, %p78; sub.ftz.f32 %f5415, %f5414, %f15691; mul.ftz.f32 %f5416, %f5413, %f5413; fma.rn.ftz.f32 %f5417, %f5409, %f5409, %f5416; fma.rn.ftz.f32 %f5418, %f5415, %f5415, %f5417; sqrt.approx.ftz.f32 %f5419, %f5418; div.approx.ftz.f32 %f5420, %f5409, %f5419; div.approx.ftz.f32 %f5421, %f5413, %f5419; div.approx.ftz.f32 %f5422, %f5415, %f5419; mul.ftz.f32 %f5423, %f15692, %f5422; mul.ftz.f32 %f5424, %f15695, %f5421; sub.ftz.f32 %f5425, %f5423, %f5424; mul.ftz.f32 %f5426, %f15695, %f5420; mul.ftz.f32 %f5427, %f15691, %f5422; sub.ftz.f32 %f5428, %f5426, %f5427; mul.ftz.f32 %f5429, %f15691, %f5421; mul.ftz.f32 %f5430, %f15692, %f5420; sub.ftz.f32 %f5431, %f5429, %f5430; mul.ftz.f32 %f15701, %f5408, %f5420; mul.ftz.f32 %f15702, %f5408, %f5421; mul.ftz.f32 %f567, %f5408, %f5422; mul.ftz.f32 %f15703, %f5408, %f5425; mul.ftz.f32 %f15704, %f5408, %f5428; mul.ftz.f32 %f570, %f5408, %f5431; abs.ftz.f32 %f573, %f15691; abs.ftz.f32 %f574, %f15692; abs.ftz.f32 %f575, %f15695; setp.geu.ftz.f32 %p79, %f575, %f573; setp.geu.ftz.f32 %p80, %f575, %f574; and.pred %p81, %p79, %p80; @%p81 bra $L__BB106_100; ld.local.f32 %f576, [%rd7+36]; setp.lt.ftz.f32 %p82, %f574, %f575; setp.lt.ftz.f32 %p83, %f574, %f573; or.pred %p84, %p83, %p82; selp.f32 %f15701, %f15702, %f15701, %p84; selp.f32 %f15703, %f15704, %f15703, %p84; selp.f32 %f15705, %f15706, %f15705, %p84; selp.f32 %f15707, %f15708, %f15707, %p84; mov.f32 %f15702, %f567; mov.f32 %f15704, %f570; mov.f32 %f15706, %f15698; mov.f32 %f15708, %f576; $L__BB106_100: mul.ftz.f32 %f5432, %f15705, %f15708; mul.ftz.f32 %f5433, %f15706, %f15707; sub.ftz.f32 %f15709, %f5433, %f5432; setp.eq.ftz.f32 %p85, %f15709, 0f00000000; @%p85 bra $L__BB106_102; rcp.approx.ftz.f32 %f15709, %f15709; $L__BB106_102: mul.ftz.f32 %f5434, %f15702, %f15705; mul.ftz.f32 %f5435, %f15701, %f15706; sub.ftz.f32 %f5436, %f5435, %f5434; mul.ftz.f32 %f5437, %f5436, %f15709; st.local.f32 [%rd7+12], %f5437; mul.ftz.f32 %f5438, %f15701, %f15708; mul.ftz.f32 %f5439, %f15702, %f15707; sub.ftz.f32 %f5440, %f5439, %f5438; mul.ftz.f32 %f5441, %f15704, %f15705; mul.ftz.f32 %f5442, %f15703, %f15706; sub.ftz.f32 %f5443, %f5442, %f5441; mul.ftz.f32 %f5444, %f5440, %f15709; mul.ftz.f32 %f5445, %f5443, %f15709; st.local.v2.f32 [%rd7+16], {%f5445, %f5444}; mul.ftz.f32 %f5446, %f15703, %f15708; mul.ftz.f32 %f5447, %f15704, %f15707; sub.ftz.f32 %f5448, %f5447, %f5446; mul.ftz.f32 %f5449, %f5448, %f15709; st.local.f32 [%rd7+24], %f5449; ld.const.u64 %rd835, [kernel_params+2544]; cvta.to.global.u64 %rd836, %rd835; shl.b64 %rd837, %rd2, 2; add.s64 %rd29, %rd836, %rd837; ld.global.u32 %r45, [%rd29]; and.b32 %r1034, %r3134, 524288; setp.eq.s32 %p86, %r1034, 0; @%p86 bra $L__BB106_110; bra.uni $L__BB106_103; $L__BB106_110: and.b32 %r46, %r45, 15728640; setp.ne.s32 %p92, %r46, 0; and.b32 %r1039, %r3134, 2097152; setp.eq.s32 %p93, %r1039, 0; and.pred %p94, %p93, %p92; @%p94 bra $L__BB106_116; and.b32 %r1040, %r45, 1016576; setp.eq.s32 %p95, %r1040, 0; mov.u32 %r1041, 0; ld.const.u32 %r1042, [kernel_params+4]; selp.b32 %r1043, %r1042, 0, %p95; st.local.u32 [%rd7+172], %r1041; st.local.u32 [%rd7+176], %r1043; ld.const.u64 %rd858, [kernel_params+2400]; cvta.to.global.u64 %rd30, %rd858; and.b32 %r1044, %r45, 24576; setp.ne.s32 %p96, %r1044, 16384; setp.eq.s64 %p97, %rd618, 0; or.pred %p1, %p97, %p96; ld.const.u64 %rd859, [kernel_params+2448]; cvta.to.global.u64 %rd860, %rd859; add.s64 %rd31, %rd860, %rd837; ld.const.s32 %rd32, [kernel_params+1264]; ld.const.s32 %rd33, [kernel_params+1436]; ld.const.s32 %rd34, [kernel_params+1440]; add.u64 %rd862, %SP, 704; add.u64 %rd35, %SPL, 704; bra.uni $L__BB106_112; $L__BB106_103: cvt.u16.u32 %rs60, %r45; setp.lt.s16 %p87, %rs60, 0; mov.f32 %f15710, 0f00000000; @%p87 bra $L__BB106_105; ld.const.u64 %rd838, [kernel_params+2576]; cvta.to.global.u64 %rd839, %rd838; add.s64 %rd841, %rd839, %rd837; ld.global.f32 %f15710, [%rd841]; $L__BB106_105: setp.eq.ftz.f32 %p88, %f15710, 0f00000000; mov.u16 %rs161, 0; @%p88 bra $L__BB106_2421; setp.eq.ftz.f32 %p89, %f15710, 0f3F800000; @%p89 bra $L__BB106_108; ld.const.u64 %rd842, [kernel_params+2584]; cvta.to.global.u64 %rd843, %rd842; mul.lo.s64 %rd844, %rd2, 12; add.s64 %rd845, %rd843, %rd844; ld.global.f32 %f5451, [%rd845]; rcp.approx.ftz.f32 %f5452, %f15710; mul.ftz.f32 %f5453, %f5452, %f5451; ld.global.f32 %f5454, [%rd845+4]; mul.ftz.f32 %f5455, %f5452, %f5454; ld.global.f32 %f5456, [%rd845+8]; mul.ftz.f32 %f5457, %f5452, %f5456; st.global.f32 [%rd845], %f5453; st.global.f32 [%rd845+4], %f5455; st.global.f32 [%rd845+8], %f5457; $L__BB106_108: mov.u32 %r3294, 0; ld.const.u64 %rd846, [kernel_params+2512]; cvta.to.global.u64 %rd847, %rd846; shl.b64 %rd848, %rd2, 1; add.s64 %rd849, %rd847, %rd848; ld.global.u16 %rs62, [%rd849]; add.s16 %rs63, %rs62, 1; st.global.u16 [%rd849], %rs63; setp.gt.u16 %p90, %rs62, 1023; @%p90 bra $L__BB106_2405; ld.const.u64 %rd850, [kernel_params+2536]; cvta.to.global.u64 %rd851, %rd850; add.s64 %rd853, %rd851, %rd848; ld.global.u16 %rs64, [%rd853]; add.s16 %rs65, %rs64, 16; st.global.u16 [%rd853], %rs65; mov.b32 %r1037, %f564; add.s32 %r1038, %r1037, 1; mov.b32 %f5458, %r1038; setp.eq.ftz.f32 %p91, %f564, 0f00000000; selp.f32 %f5459, 0f00000001, %f5458, %p91; ld.const.u64 %rd854, [kernel_params+2648]; cvta.to.global.u64 %rd855, %rd854; add.s64 %rd857, %rd855, %rd837; st.global.f32 [%rd857], %f5459; mov.u32 %r3294, 33; bra.uni $L__BB106_2405; $L__BB106_309: { // callseq 84, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 84 add.s32 %r3074, %r47, 1; $L__BB106_112: mov.u32 %r47, %r3074; mul.wide.s32 %rd863, %r47, 16; add.s64 %rd36, %rd30, %rd863; ld.global.v4.u32 {%r1046, %r1047, %r1048, %r1049}, [%rd36]; mul.wide.u32 %rd864, %r1047, 4; add.s64 %rd37, %rd35, %rd864; mul.wide.u32 %rd865, %r1048, 4; add.s64 %rd38, %rd35, %rd865; setp.gt.s32 %p98, %r1046, 1; @%p98 bra $L__BB106_144; setp.eq.s32 %p228, %r1046, 1; mov.u32 %r3074, %r1047; @%p228 bra $L__BB106_112; bra.uni $L__BB106_114; $L__BB106_144: setp.gt.s32 %p99, %r1046, 49; @%p99 bra $L__BB106_225; bra.uni $L__BB106_145; $L__BB106_225: setp.gt.s32 %p100, %r1046, 73; @%p100 bra $L__BB106_265; bra.uni $L__BB106_226; $L__BB106_265: setp.gt.s32 %p101, %r1046, 85; @%p101 bra $L__BB106_288; bra.uni $L__BB106_266; $L__BB106_288: setp.gt.s32 %p102, %r1046, 91; @%p102 bra $L__BB106_298; bra.uni $L__BB106_289; $L__BB106_298: setp.gt.s32 %p103, %r1046, 94; @%p103 bra $L__BB106_303; bra.uni $L__BB106_299; $L__BB106_303: setp.gt.s32 %p104, %r1046, 96; @%p104 bra $L__BB106_307; setp.eq.s32 %p107, %r1046, 95; @%p107 bra $L__BB106_310; setp.eq.s32 %p108, %r1046, 96; @%p108 bra $L__BB106_306; bra.uni $L__BB106_114; $L__BB106_306: { // callseq 85, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 85 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_145: setp.gt.s32 %p165, %r1046, 25; @%p165 bra $L__BB106_186; setp.gt.s32 %p197, %r1046, 13; @%p197 bra $L__BB106_167; setp.gt.s32 %p213, %r1046, 7; @%p213 bra $L__BB106_157; setp.gt.s32 %p221, %r1046, 4; @%p221 bra $L__BB106_153; setp.eq.s32 %p225, %r1046, 2; @%p225 bra $L__BB106_376; setp.eq.s32 %p226, %r1046, 3; @%p226 bra $L__BB106_375; setp.eq.s32 %p227, %r1046, 4; @%p227 bra $L__BB106_152; bra.uni $L__BB106_114; $L__BB106_152: mov.u32 %r1123, 4; { // callseq 163, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1123, %r1047, %r1048, %r1049}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 163 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_226: setp.gt.s32 %p134, %r1046, 61; @%p134 bra $L__BB106_246; setp.gt.s32 %p150, %r1046, 55; @%p150 bra $L__BB106_237; setp.gt.s32 %p158, %r1046, 52; @%p158 bra $L__BB106_233; setp.eq.s32 %p162, %r1046, 50; @%p162 bra $L__BB106_343; setp.eq.s32 %p163, %r1046, 51; @%p163 bra $L__BB106_342; setp.eq.s32 %p164, %r1046, 52; @%p164 bra $L__BB106_232; bra.uni $L__BB106_114; $L__BB106_232: add.s32 %r1086, %r47, 1; { // callseq 124, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 124 bra.uni $L__BB106_112; $L__BB106_186: setp.gt.s32 %p166, %r1046, 37; @%p166 bra $L__BB106_206; setp.gt.s32 %p182, %r1046, 31; @%p182 bra $L__BB106_197; setp.gt.s32 %p190, %r1046, 28; @%p190 bra $L__BB106_193; setp.eq.s32 %p194, %r1046, 26; @%p194 bra $L__BB106_359; setp.eq.s32 %p195, %r1046, 27; @%p195 bra $L__BB106_358; setp.eq.s32 %p196, %r1046, 28; @%p196 bra $L__BB106_192; bra.uni $L__BB106_114; $L__BB106_192: mov.u32 %r1099, 28; { // callseq 146, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1099; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 146 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_266: setp.gt.s32 %p119, %r1046, 79; @%p119 bra $L__BB106_279; setp.gt.s32 %p127, %r1046, 76; @%p127 bra $L__BB106_275; setp.eq.s32 %p131, %r1046, 74; @%p131 bra $L__BB106_327; setp.eq.s32 %p132, %r1046, 75; @%p132 bra $L__BB106_326; setp.eq.s32 %p133, %r1046, 76; @%p133 bra $L__BB106_271; bra.uni $L__BB106_114; $L__BB106_271: add.s32 %r3074, %r47, 1; setp.eq.s32 %p230, %r1049, 255; @%p230 bra $L__BB106_112; add.s64 %rd39, %rd37, 4; add.s64 %rd40, %rd37, 8; setp.eq.s32 %p231, %r1048, 0; mul.wide.u32 %rd913, %r1049, 4; add.s64 %rd41, %rd35, %rd913; @%p231 bra $L__BB106_325; setp.eq.s32 %p232, %r1048, 1; @%p232 bra $L__BB106_324; bra.uni $L__BB106_274; $L__BB106_324: ld.local.f32 %f5471, [%rd39]; st.local.f32 [%rd41], %f5471; bra.uni $L__BB106_112; $L__BB106_167: setp.gt.s32 %p198, %r1046, 19; @%p198 bra $L__BB106_177; setp.gt.s32 %p206, %r1046, 16; @%p206 bra $L__BB106_173; setp.eq.s32 %p210, %r1046, 14; @%p210 bra $L__BB106_367; setp.eq.s32 %p211, %r1046, 15; @%p211 bra $L__BB106_366; setp.eq.s32 %p212, %r1046, 16; @%p212 bra $L__BB106_172; bra.uni $L__BB106_114; $L__BB106_172: mov.u32 %r1109, 16; { // callseq 157, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1109; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 157 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_246: setp.gt.s32 %p135, %r1046, 67; @%p135 bra $L__BB106_256; setp.gt.s32 %p143, %r1046, 64; @%p143 bra $L__BB106_252; setp.eq.s32 %p147, %r1046, 62; @%p147 bra $L__BB106_335; setp.eq.s32 %p148, %r1046, 63; @%p148 bra $L__BB106_334; setp.eq.s32 %p149, %r1046, 64; @%p149 bra $L__BB106_251; bra.uni $L__BB106_114; $L__BB106_251: mov.u32 %r1072, 64; add.s32 %r1073, %r47, 1; { // callseq 112, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1072; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1073; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 112 bra.uni $L__BB106_112; $L__BB106_206: setp.gt.s32 %p167, %r1046, 43; @%p167 bra $L__BB106_216; setp.gt.s32 %p175, %r1046, 40; @%p175 bra $L__BB106_212; setp.eq.s32 %p179, %r1046, 38; @%p179 bra $L__BB106_351; setp.eq.s32 %p180, %r1046, 39; @%p180 bra $L__BB106_350; setp.eq.s32 %p181, %r1046, 40; @%p181 bra $L__BB106_211; bra.uni $L__BB106_114; $L__BB106_211: add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_289: setp.gt.s32 %p112, %r1046, 88; @%p112 bra $L__BB106_294; setp.eq.s32 %p116, %r1046, 86; @%p116 bra $L__BB106_316; setp.eq.s32 %p117, %r1046, 87; @%p117 bra $L__BB106_315; setp.eq.s32 %p118, %r1046, 88; @%p118 bra $L__BB106_293; bra.uni $L__BB106_114; $L__BB106_293: mov.u32 %r1055, 88; { // callseq 90, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b64 param1; st.param.b64 [param1+0], %rd619; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1055; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; call.uni _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3 ); } // callseq 90 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_157: setp.gt.s32 %p214, %r1046, 10; @%p214 bra $L__BB106_163; setp.eq.s32 %p218, %r1046, 8; @%p218 bra $L__BB106_372; setp.eq.s32 %p219, %r1046, 9; @%p219 bra $L__BB106_370; setp.eq.s32 %p220, %r1046, 10; @%p220 bra $L__BB106_161; bra.uni $L__BB106_114; $L__BB106_161: ld.local.f32 %f5497, [%rd38]; setp.ltu.ftz.f32 %p233, %f5497, 0f3F800000; add.s32 %r3074, %r47, 1; @%p233 bra $L__BB106_112; add.s32 %r1119, %r1047, %r47; add.s32 %r3074, %r1119, 1; bra.uni $L__BB106_112; $L__BB106_237: setp.gt.s32 %p151, %r1046, 58; @%p151 bra $L__BB106_242; setp.eq.s32 %p155, %r1046, 56; @%p155 bra $L__BB106_339; setp.eq.s32 %p156, %r1046, 57; @%p156 bra $L__BB106_338; setp.eq.s32 %p157, %r1046, 58; @%p157 bra $L__BB106_241; bra.uni $L__BB106_114; $L__BB106_241: mov.u32 %r1081, 58; { // callseq 118, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1081; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 118 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_197: setp.gt.s32 %p183, %r1046, 34; @%p183 bra $L__BB106_202; setp.eq.s32 %p187, %r1046, 32; @%p187 bra $L__BB106_355; setp.eq.s32 %p188, %r1046, 33; @%p188 bra $L__BB106_354; setp.eq.s32 %p189, %r1046, 34; @%p189 bra $L__BB106_201; bra.uni $L__BB106_114; $L__BB106_201: { // callseq 141, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 141 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_279: setp.gt.s32 %p120, %r1046, 82; @%p120 bra $L__BB106_284; setp.eq.s32 %p124, %r1046, 80; @%p124 bra $L__BB106_320; setp.eq.s32 %p125, %r1046, 81; @%p125 bra $L__BB106_319; setp.eq.s32 %p126, %r1046, 82; @%p126 bra $L__BB106_283; bra.uni $L__BB106_114; $L__BB106_283: mov.u32 %r1059, 82; { // callseq 96, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1059; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 96 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_177: setp.gt.s32 %p199, %r1046, 22; @%p199 bra $L__BB106_182; setp.eq.s32 %p203, %r1046, 20; @%p203 bra $L__BB106_363; setp.eq.s32 %p204, %r1046, 21; @%p204 bra $L__BB106_362; setp.eq.s32 %p205, %r1046, 22; @%p205 bra $L__BB106_181; bra.uni $L__BB106_114; $L__BB106_181: mov.u32 %r1106, 22; add.s32 %r1107, %r47, 1; { // callseq 152, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1106; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r1107; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 152 bra.uni $L__BB106_112; $L__BB106_256: setp.gt.s32 %p136, %r1046, 70; @%p136 bra $L__BB106_261; setp.eq.s32 %p140, %r1046, 68; @%p140 bra $L__BB106_331; setp.eq.s32 %p141, %r1046, 69; @%p141 bra $L__BB106_330; setp.eq.s32 %p142, %r1046, 70; @%p142 bra $L__BB106_260; bra.uni $L__BB106_114; $L__BB106_260: mov.u32 %r1066, 70; { // callseq 106, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1066; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 106 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_216: setp.gt.s32 %p168, %r1046, 46; @%p168 bra $L__BB106_221; setp.eq.s32 %p172, %r1046, 44; @%p172 bra $L__BB106_347; setp.eq.s32 %p173, %r1046, 45; @%p173 bra $L__BB106_346; setp.eq.s32 %p174, %r1046, 46; @%p174 bra $L__BB106_220; bra.uni $L__BB106_114; $L__BB106_220: { // callseq 130, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 130 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_299: setp.eq.s32 %p109, %r1046, 92; @%p109 bra $L__BB106_312; setp.eq.s32 %p110, %r1046, 93; @%p110 bra $L__BB106_311; setp.eq.s32 %p111, %r1046, 94; @%p111 bra $L__BB106_302; bra.uni $L__BB106_114; $L__BB106_302: mov.u32 %r1050, 94; add.s32 %r1051, %r47, 1; { // callseq 87, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1050; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1051; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 87 bra.uni $L__BB106_112; $L__BB106_153: setp.eq.s32 %p222, %r1046, 5; @%p222 bra $L__BB106_374; setp.eq.s32 %p223, %r1046, 6; @%p223 bra $L__BB106_373; setp.eq.s32 %p224, %r1046, 7; @%p224 bra $L__BB106_156; bra.uni $L__BB106_114; $L__BB106_156: mov.u32 %r1122, 7; { // callseq 162, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1122; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 162 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_233: setp.eq.s32 %p159, %r1046, 53; @%p159 bra $L__BB106_341; setp.eq.s32 %p160, %r1046, 54; @%p160 bra $L__BB106_340; setp.eq.s32 %p161, %r1046, 55; @%p161 bra $L__BB106_236; bra.uni $L__BB106_114; $L__BB106_236: ld.local.v4.f32 {%f5473, %f5474, %f5475, %f5476}, [%rd1]; { // callseq 121, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f5473; st.param.f32 [param0+4], %f5474; st.param.f32 [param0+8], %f5475; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 121 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_193: setp.eq.s32 %p191, %r1046, 29; @%p191 bra $L__BB106_357; setp.eq.s32 %p192, %r1046, 30; @%p192 bra $L__BB106_356; setp.eq.s32 %p193, %r1046, 31; @%p193 bra $L__BB106_196; bra.uni $L__BB106_114; $L__BB106_196: mov.u32 %r1097, 31; add.s32 %r1098, %r47, 1; { // callseq 143, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1097; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1098; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 143 bra.uni $L__BB106_112; $L__BB106_275: setp.eq.s32 %p128, %r1046, 77; @%p128 bra $L__BB106_322; setp.eq.s32 %p129, %r1046, 78; @%p129 bra $L__BB106_321; setp.eq.s32 %p130, %r1046, 79; @%p130 bra $L__BB106_278; bra.uni $L__BB106_114; $L__BB106_278: add.s32 %r1061, %r47, 1; { // callseq 99, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1061; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 99 bra.uni $L__BB106_112; $L__BB106_173: setp.eq.s32 %p207, %r1046, 17; @%p207 bra $L__BB106_365; setp.eq.s32 %p208, %r1046, 18; @%p208 bra $L__BB106_364; setp.eq.s32 %p209, %r1046, 19; @%p209 bra $L__BB106_176; bra.uni $L__BB106_114; $L__BB106_176: { // callseq 154, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 154 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_252: setp.eq.s32 %p144, %r1046, 65; @%p144 bra $L__BB106_333; setp.eq.s32 %p145, %r1046, 66; @%p145 bra $L__BB106_332; setp.eq.s32 %p146, %r1046, 67; @%p146 bra $L__BB106_255; bra.uni $L__BB106_114; $L__BB106_255: mov.u32 %r1070, 67; { // callseq 109, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1070, %r1047, %r1048, %r1049}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 109 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_212: setp.eq.s32 %p176, %r1046, 41; @%p176 bra $L__BB106_349; setp.eq.s32 %p177, %r1046, 42; @%p177 bra $L__BB106_348; setp.eq.s32 %p178, %r1046, 43; @%p178 bra $L__BB106_215; bra.uni $L__BB106_114; $L__BB106_215: add.s32 %r1090, %r47, 1; { // callseq 133, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1090; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 133 bra.uni $L__BB106_112; $L__BB106_294: setp.eq.s32 %p113, %r1046, 89; @%p113 bra $L__BB106_314; setp.eq.s32 %p114, %r1046, 90; @%p114 bra $L__BB106_313; setp.eq.s32 %p115, %r1046, 91; @%p115 bra $L__BB106_297; bra.uni $L__BB106_114; $L__BB106_297: add.s32 %r3074, %r47, 1; @%p1 bra $L__BB106_114; bra.uni $L__BB106_112; $L__BB106_163: setp.eq.s32 %p215, %r1046, 11; @%p215 bra $L__BB106_369; setp.eq.s32 %p216, %r1046, 12; @%p216 bra $L__BB106_368; setp.eq.s32 %p217, %r1046, 13; @%p217 bra $L__BB106_166; bra.uni $L__BB106_114; $L__BB106_166: mov.u32 %r1117, 13; add.s32 %r1118, %r47, 1; { // callseq 158, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1117; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1118; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 158 bra.uni $L__BB106_112; $L__BB106_242: setp.eq.s32 %p152, %r1046, 59; @%p152 bra $L__BB106_337; setp.eq.s32 %p153, %r1046, 60; @%p153 bra $L__BB106_336; setp.eq.s32 %p154, %r1046, 61; @%p154 bra $L__BB106_245; bra.uni $L__BB106_114; $L__BB106_245: mov.u32 %r1077, 61; add.s32 %r1078, %r47, 1; { // callseq 115, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1077; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1078; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 115 bra.uni $L__BB106_112; $L__BB106_202: setp.eq.s32 %p184, %r1046, 35; @%p184 bra $L__BB106_353; setp.eq.s32 %p185, %r1046, 36; @%p185 bra $L__BB106_352; setp.eq.s32 %p186, %r1046, 37; @%p186 bra $L__BB106_205; bra.uni $L__BB106_114; $L__BB106_205: mov.u32 %r1093, 37; { // callseq 138, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1093, %r1047, %r1048, %r1049}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 138 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_284: setp.eq.s32 %p121, %r1046, 83; @%p121 bra $L__BB106_318; setp.eq.s32 %p122, %r1046, 84; @%p122 bra $L__BB106_317; setp.eq.s32 %p123, %r1046, 85; @%p123 bra $L__BB106_287; bra.uni $L__BB106_114; $L__BB106_287: add.s32 %r1058, %r47, 1; { // callseq 93, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1058; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 93 bra.uni $L__BB106_112; $L__BB106_182: setp.eq.s32 %p200, %r1046, 23; @%p200 bra $L__BB106_361; setp.eq.s32 %p201, %r1046, 24; @%p201 bra $L__BB106_360; setp.eq.s32 %p202, %r1046, 25; @%p202 bra $L__BB106_185; bra.uni $L__BB106_114; $L__BB106_185: add.s32 %r1102, %r47, 1; { // callseq 149, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1102; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 149 bra.uni $L__BB106_112; $L__BB106_261: setp.eq.s32 %p137, %r1046, 71; @%p137 bra $L__BB106_329; setp.eq.s32 %p138, %r1046, 72; @%p138 bra $L__BB106_328; setp.eq.s32 %p139, %r1046, 73; @%p139 bra $L__BB106_264; bra.uni $L__BB106_114; $L__BB106_264: add.s32 %r1064, %r47, 1; { // callseq 103, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1064; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 103 bra.uni $L__BB106_112; $L__BB106_221: setp.eq.s32 %p169, %r1046, 47; @%p169 bra $L__BB106_345; setp.eq.s32 %p170, %r1046, 48; @%p170 bra $L__BB106_344; setp.eq.s32 %p171, %r1046, 49; @%p171 bra $L__BB106_224; bra.uni $L__BB106_114; $L__BB106_224: { // callseq 127, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 127 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_307: setp.eq.s32 %p105, %r1046, 98; @%p105 bra $L__BB106_377; setp.ne.s32 %p106, %r1046, 97; @%p106 bra $L__BB106_114; bra.uni $L__BB106_309; $L__BB106_310: { // callseq 86, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 86 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_376: mov.u32 %r1125, 2; add.s32 %r1126, %r47, 1; { // callseq 165, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1125; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r45; .param .b32 param4; st.param.b32 [param4+0], %r1126; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 165 bra.uni $L__BB106_112; $L__BB106_375: mov.u32 %r1124, 3; { // callseq 164, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1124, %r1047, %r1048, %r1049}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 164 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_343: { // callseq 126, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 126 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_342: { // callseq 125, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 125 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_359: mov.u32 %r1101, 26; { // callseq 148, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1101; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 148 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_358: mov.u32 %r1100, 27; { // callseq 147, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1100; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 147 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_327: { // callseq 102, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 102 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_326: { // callseq 101, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 101 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_367: st.local.u32 [%rd38], %r1047; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_366: ld.global.v4.u32 {%r1110, %r1111, %r1112, %r1113}, [%rd36+16]; add.s32 %r3074, %r47, 2; st.local.u32 [%rd37], %r1111; st.local.u32 [%rd37+4], %r1112; st.local.u32 [%rd37+8], %r1113; bra.uni $L__BB106_112; $L__BB106_335: mov.u32 %r1075, 62; add.s32 %r1076, %r47, 1; { // callseq 114, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1075; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1076; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 114 bra.uni $L__BB106_112; $L__BB106_334: mov.u32 %r1074, 63; { // callseq 113, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1074; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 113 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_351: { // callseq 137, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 137 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_350: mov.u32 %r1092, 39; { // callseq 136, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1092, %r1047, %r1048, %r1049}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 136 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_316: add.s32 %r1057, %r47, 1; { // callseq 92, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1057; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 92 bra.uni $L__BB106_112; $L__BB106_315: add.s32 %r1056, %r47, 1; { // callseq 91, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1056; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 91 bra.uni $L__BB106_112; $L__BB106_372: mov.u32 %r1121, 8; { // callseq 161, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1121; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 161 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_370: ld.local.f32 %f5498, [%rd38]; setp.gtu.ftz.f32 %p234, %f5498, 0f00000000; add.s32 %r3074, %r47, 1; @%p234 bra $L__BB106_112; add.s32 %r1120, %r1047, %r47; add.s32 %r3074, %r1120, 1; bra.uni $L__BB106_112; $L__BB106_339: mov.u32 %r1084, 56; { // callseq 120, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1084; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 120 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_338: mov.u32 %r1082, 57; add.s32 %r1083, %r47, 1; { // callseq 119, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1082; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1083; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 119 bra.uni $L__BB106_112; $L__BB106_355: mov.u32 %r1095, 32; add.s32 %r1096, %r47, 1; { // callseq 142, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1095; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1096; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 142 bra.uni $L__BB106_112; $L__BB106_354: ld.local.f32 %f5481, [%rd37]; ld.local.f32 %f5482, [%rd37+4]; ld.local.f32 %f5483, [%rd37+8]; st.local.f32 [%rd7+-76], %f5481; st.local.f32 [%rd7+-72], %f5482; st.local.f32 [%rd7+-68], %f5483; st.local.f32 [%rd38], %f5481; st.local.f32 [%rd38+4], %f5482; st.local.f32 [%rd38+8], %f5483; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_320: { // callseq 98, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 98 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_319: mov.u32 %r1060, 81; { // callseq 97, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1060; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 97 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_363: ld.local.f32 %f5484, [%rd37]; ld.local.v4.f32 {%f5485, %f5486, %f5487, %f5488}, [%rd1]; ld.local.f32 %f5492, [%rd37+4]; ld.local.f32 %f5493, [%rd37+8]; add.ftz.f32 %f5494, %f5493, %f5487; add.ftz.f32 %f5495, %f5492, %f5486; add.ftz.f32 %f5496, %f5484, %f5485; st.local.v2.f32 [%rd1], {%f5496, %f5495}; st.local.f32 [%rd7+-80], %f5494; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_362: mov.u32 %r1108, 21; { // callseq 153, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1108; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 153 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_331: mov.u32 %r1069, 68; { // callseq 108, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1069; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 108 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_330: mov.u32 %r1067, 69; add.s32 %r1068, %r47, 1; { // callseq 107, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1067; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1068; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 107 bra.uni $L__BB106_112; $L__BB106_347: mov.u32 %r1088, 44; add.s32 %r1089, %r47, 1; { // callseq 132, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1088; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1089; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 132 bra.uni $L__BB106_112; $L__BB106_346: { // callseq 131, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 131 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_312: ld.local.f32 %f5462, [%rd37]; ld.local.f32 %f5463, [%rd37+4]; ld.local.f32 %f5464, [%rd37+8]; ld.global.u32 %rd878, [%rd31]; mul.lo.s64 %rd879, %rd32, %rd878; cvt.s64.s32 %rd880, %r1048; add.s64 %rd881, %rd33, %rd880; add.s64 %rd882, %rd881, %rd879; cvta.to.global.u64 %rd883, %rd618; shl.b64 %rd884, %rd882, 2; add.s64 %rd885, %rd883, %rd884; add.s64 %rd886, %rd885, 4; add.s64 %rd887, %rd885, 8; add.s64 %rd888, %rd885, 12; atom.global.add.f32 %f5465, [%rd885], %f5462; atom.global.add.f32 %f5466, [%rd886], %f5463; atom.global.add.f32 %f5467, [%rd887], %f5464; atom.global.add.f32 %f5468, [%rd888], 0f3F800000; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_311: ld.local.f32 %f5460, [%rd37]; ld.global.u32 %rd870, [%rd31]; mul.lo.s64 %rd871, %rd32, %rd870; cvt.s64.s32 %rd872, %r1048; add.s64 %rd873, %rd34, %rd872; add.s64 %rd874, %rd873, %rd871; cvta.to.global.u64 %rd875, %rd618; shl.b64 %rd876, %rd874, 2; add.s64 %rd877, %rd875, %rd876; atom.global.add.f32 %f5461, [%rd877], %f5460; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_374: st.local.u32 [%rd7+180], %r1047; st.local.u32 [%rd7+184], %r1048; st.local.u32 [%rd7+188], %r1049; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_373: ld.local.f32 %f5499, [%rd37]; ld.local.f32 %f5500, [%rd37+4]; ld.local.f32 %f5501, [%rd37+8]; st.local.f32 [%rd7+180], %f5499; st.local.f32 [%rd7+184], %f5500; st.local.f32 [%rd7+188], %f5501; add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_341: { // callseq 123, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 123 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_340: add.s32 %r1085, %r47, 1; { // callseq 122, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1085; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 122 bra.uni $L__BB106_112; $L__BB106_357: { // callseq 145, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 145 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_356: { // callseq 144, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 144 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_322: setp.eq.s32 %p229, %r1049, 255; add.s32 %r3074, %r47, 1; @%p229 bra $L__BB106_112; ld.local.f32 %f5469, [%rd37]; add.s32 %r1063, %r1049, %r1048; mul.wide.u32 %rd909, %r1063, 4; add.s64 %rd910, %rd35, %rd909; st.local.f32 [%rd910], %f5469; bra.uni $L__BB106_112; $L__BB106_321: add.s32 %r1062, %r47, 1; { // callseq 100, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1062; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 100 bra.uni $L__BB106_112; $L__BB106_365: { // callseq 156, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 156 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_364: { // callseq 155, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 155 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_333: { // callseq 111, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 111 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_332: add.s32 %r1071, %r47, 1; { // callseq 110, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1071; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 110 bra.uni $L__BB106_112; $L__BB106_349: add.s32 %r1091, %r47, 1; { // callseq 135, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1091; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 135 bra.uni $L__BB106_112; $L__BB106_348: { // callseq 134, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 134 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_314: mov.u32 %r1054, 89; { // callseq 89, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b64 param1; st.param.b64 [param1+0], %rd619; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1054; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; call.uni _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3 ); } // callseq 89 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_313: mov.u32 %r1052, 90; add.s32 %r1053, %r47, 1; { // callseq 88, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1052; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r1053; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 88 bra.uni $L__BB106_112; $L__BB106_369: { // callseq 160, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 160 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_368: { // callseq 159, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 159 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_337: add.s32 %r1080, %r47, 1; { // callseq 117, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1080; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 117 bra.uni $L__BB106_112; $L__BB106_336: add.s32 %r1079, %r47, 1; { // callseq 116, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1079; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 116 bra.uni $L__BB106_112; $L__BB106_353: { // callseq 140, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 140 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_352: mov.u32 %r1094, 36; { // callseq 139, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1094; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 139 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_318: { // callseq 95, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 95 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_317: { // callseq 94, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 94 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_361: mov.u32 %r1104, 23; add.s32 %r1105, %r47, 1; { // callseq 151, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1104; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1105; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 151 bra.uni $L__BB106_112; $L__BB106_360: mov.u32 %r1103, 24; { // callseq 150, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1103; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 150 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_329: mov.u32 %r1065, 71; { // callseq 105, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1065; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 105 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_328: { // callseq 104, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 104 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_345: ld.local.u32 %r1087, [%rd7+-36]; ld.local.f32 %f5480, [%rd7]; { // callseq 129, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b32 param1; st.param.b32 [param1+0], %r1087; .param .b32 param2; st.param.f32 [param2+0], %f5480; .param .b64 param3; st.param.b64 [param3+0], %rd862; .param .b32 param4; st.param.b32 [param4+0], %r1047; .param .b32 param5; st.param.b32 [param5+0], %r1048; .param .b32 param6; st.param.b32 [param6+0], %r45; call.uni _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5, param6 ); } // callseq 129 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_344: { // callseq 128, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 128 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_377: { // callseq 166, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 166 add.s32 %r3074, %r47, 1; bra.uni $L__BB106_112; $L__BB106_325: ld.local.f32 %f5472, [%rd37]; st.local.f32 [%rd41], %f5472; bra.uni $L__BB106_112; $L__BB106_274: ld.local.f32 %f5470, [%rd40]; st.local.f32 [%rd41], %f5470; bra.uni $L__BB106_112; $L__BB106_114: ld.local.u32 %r3134, [%rd7+-36]; and.b32 %r1127, %r3134, 1024; setp.eq.s32 %p235, %r1127, 0; @%p235 bra $L__BB106_116; ld.const.v2.u64 {%rd1015, %rd1016}, [kernel_params+2528]; cvta.to.global.u64 %rd1019, %rd1015; add.s64 %rd1021, %rd1019, %rd837; cvta.to.global.u64 %rd1022, %rd1016; shl.b64 %rd1023, %rd2, 1; add.s64 %rd1024, %rd1022, %rd1023; ld.global.u16 %r1128, [%rd1024]; ld.const.u64 %rd1025, [kernel_params+2456]; cvta.to.global.u64 %rd1026, %rd1025; add.s64 %rd1027, %rd1026, %rd837; ld.global.u32 %r1129, [%rd1021]; add.s32 %r1130, %r1129, %r1128; ld.global.u32 %r1131, [%rd1027]; mad.lo.s32 %r1132, %r1131, -1262732973, %r1130; mad.lo.s32 %r1133, %r1132, 1103515245, 12345; st.local.u32 [%rd7+168], %r1133; $L__BB106_116: setp.eq.s32 %p236, %r46, 0; @%p236 bra $L__BB106_386; ld.local.v2.f32 {%f15742, %f15743}, [%rd7+-72]; ld.local.f32 %f15741, [%rd7+-76]; and.b32 %r1134, %r3134, 2097152; setp.eq.s32 %p237, %r1134, 0; @%p237 bra $L__BB106_136; ld.local.u32 %r156, [%rd7+172]; setp.lt.s32 %p238, %r156, 1; mov.f32 %f15714, 0f00000000; mov.f32 %f15715, %f15714; mov.f32 %f15716, %f15714; @%p238 bra $L__BB106_134; add.s32 %r1136, %r156, -1; and.b32 %r3080, %r156, 3; setp.lt.u32 %p239, %r1136, 3; mov.f32 %f15716, 0f00000000; mov.u32 %r3078, 0; mov.f32 %f15715, %f15716; mov.f32 %f15714, %f15716; @%p239 bra $L__BB106_130; sub.s32 %r3077, %r156, %r3080; mov.f32 %f15716, 0f00000000; mov.u32 %r3078, 0; mov.f32 %f15715, %f15716; mov.f32 %f15714, %f15716; $L__BB106_121: cvt.s64.s32 %rd42, %r3078; mul.wide.s32 %rd1028, %r3078, 80; add.s64 %rd1029, %rd1, %rd1028; add.s64 %rd43, %rd1029, 316; ld.local.u32 %r1138, [%rd1029+316]; add.s32 %r1139, %r1138, -32; setp.gt.u32 %p240, %r1139, 2; @%p240 bra $L__BB106_123; ld.local.v4.f32 {%f5514, %f5515, %f5516, %f5517}, [%rd43+-12]; add.ftz.f32 %f5521, %f5514, %f5515; add.ftz.f32 %f5522, %f5516, %f5521; mul.ftz.f32 %f5523, %f5522, 0f3EAAAAAB; abs.ftz.f32 %f5524, %f5523; ld.local.f32 %f5525, [%rd43+8]; ld.local.f32 %f5526, [%rd43+12]; ld.local.f32 %f5527, [%rd43+16]; fma.rn.ftz.f32 %f15714, %f5524, %f5525, %f15714; fma.rn.ftz.f32 %f15715, %f5524, %f5526, %f15715; fma.rn.ftz.f32 %f15716, %f5524, %f5527, %f15716; $L__BB106_123: ld.local.u32 %r1140, [%rd43+80]; add.s32 %r1141, %r1140, -32; setp.gt.u32 %p241, %r1141, 2; @%p241 bra $L__BB106_125; ld.local.v4.f32 {%f5528, %f5529, %f5530, %f5531}, [%rd43+68]; add.ftz.f32 %f5535, %f5528, %f5529; add.ftz.f32 %f5536, %f5530, %f5535; mul.ftz.f32 %f5537, %f5536, 0f3EAAAAAB; abs.ftz.f32 %f5538, %f5537; ld.local.f32 %f5539, [%rd43+88]; ld.local.f32 %f5540, [%rd43+92]; ld.local.f32 %f5541, [%rd43+96]; fma.rn.ftz.f32 %f15714, %f5538, %f5539, %f15714; fma.rn.ftz.f32 %f15715, %f5538, %f5540, %f15715; fma.rn.ftz.f32 %f15716, %f5538, %f5541, %f15716; $L__BB106_125: ld.local.u32 %r1142, [%rd43+160]; add.s32 %r1143, %r1142, -32; setp.gt.u32 %p242, %r1143, 2; @%p242 bra $L__BB106_127; ld.local.v4.f32 {%f5542, %f5543, %f5544, %f5545}, [%rd43+148]; add.ftz.f32 %f5549, %f5542, %f5543; add.ftz.f32 %f5550, %f5544, %f5549; mul.ftz.f32 %f5551, %f5550, 0f3EAAAAAB; abs.ftz.f32 %f5552, %f5551; ld.local.f32 %f5553, [%rd43+168]; ld.local.f32 %f5554, [%rd43+172]; ld.local.f32 %f5555, [%rd43+176]; fma.rn.ftz.f32 %f15714, %f5552, %f5553, %f15714; fma.rn.ftz.f32 %f15715, %f5552, %f5554, %f15715; fma.rn.ftz.f32 %f15716, %f5552, %f5555, %f15716; $L__BB106_127: ld.local.u32 %r1144, [%rd43+240]; add.s32 %r1145, %r1144, -32; setp.gt.u32 %p243, %r1145, 2; @%p243 bra $L__BB106_129; ld.local.v4.f32 {%f5556, %f5557, %f5558, %f5559}, [%rd43+228]; add.ftz.f32 %f5563, %f5556, %f5557; add.ftz.f32 %f5564, %f5558, %f5563; mul.ftz.f32 %f5565, %f5564, 0f3EAAAAAB; abs.ftz.f32 %f5566, %f5565; ld.local.f32 %f5567, [%rd43+248]; ld.local.f32 %f5568, [%rd43+252]; ld.local.f32 %f5569, [%rd43+256]; fma.rn.ftz.f32 %f15714, %f5566, %f5567, %f15714; fma.rn.ftz.f32 %f15715, %f5566, %f5568, %f15715; fma.rn.ftz.f32 %f15716, %f5566, %f5569, %f15716; $L__BB106_129: cvt.u32.u64 %r1146, %rd42; add.s32 %r3078, %r1146, 4; add.s32 %r3077, %r3077, -4; setp.ne.s32 %p244, %r3077, 0; @%p244 bra $L__BB106_121; $L__BB106_130: setp.eq.s32 %p245, %r3080, 0; @%p245 bra $L__BB106_134; $L__BB106_131: .pragma "nounroll"; cvt.s64.s32 %rd44, %r3078; mul.wide.s32 %rd1030, %r3078, 80; add.s64 %rd1031, %rd1, %rd1030; add.s64 %rd45, %rd1031, 316; ld.local.u32 %r1147, [%rd1031+316]; add.s32 %r1148, %r1147, -32; setp.gt.u32 %p246, %r1148, 2; @%p246 bra $L__BB106_133; ld.local.f32 %f5570, [%rd45+-8]; ld.local.f32 %f5571, [%rd45+-12]; add.ftz.f32 %f5572, %f5571, %f5570; ld.local.f32 %f5573, [%rd45+-4]; add.ftz.f32 %f5574, %f5573, %f5572; mul.ftz.f32 %f5575, %f5574, 0f3EAAAAAB; abs.ftz.f32 %f5576, %f5575; ld.local.f32 %f5577, [%rd45+8]; ld.local.f32 %f5578, [%rd45+12]; ld.local.f32 %f5579, [%rd45+16]; fma.rn.ftz.f32 %f15714, %f5576, %f5577, %f15714; fma.rn.ftz.f32 %f15715, %f5576, %f5578, %f15715; fma.rn.ftz.f32 %f15716, %f5576, %f5579, %f15716; $L__BB106_133: cvt.u32.u64 %r1149, %rd44; add.s32 %r3078, %r1149, 1; add.s32 %r3080, %r3080, -1; setp.ne.s32 %p247, %r3080, 0; @%p247 bra $L__BB106_131; $L__BB106_134: setp.eq.ftz.f32 %p248, %f15715, 0f00000000; setp.eq.ftz.f32 %p249, %f15714, 0f00000000; and.pred %p250, %p249, %p248; setp.eq.ftz.f32 %p251, %f15716, 0f00000000; and.pred %p252, %p251, %p250; @%p252 bra $L__BB106_136; mul.ftz.f32 %f5580, %f15715, %f15715; fma.rn.ftz.f32 %f5581, %f15714, %f15714, %f5580; fma.rn.ftz.f32 %f5582, %f15716, %f15716, %f5581; sqrt.approx.ftz.f32 %f5583, %f5582; div.approx.ftz.f32 %f15743, %f15716, %f5583; div.approx.ftz.f32 %f15742, %f15715, %f5583; div.approx.ftz.f32 %f15741, %f15714, %f5583; $L__BB106_136: and.b32 %r1150, %r3134, -1279; st.local.u32 [%rd7+-36], %r1150; mov.u32 %r1151, 0; st.local.u32 [%rd7+172], %r1151; ld.const.u32 %r168, [kernel_params+4]; st.local.u32 [%rd7+176], %r168; and.b32 %r1152, %r45, 4194304; setp.eq.s32 %p253, %r1152, 0; mov.f32 %f5584, 0f3F800000; abs.ftz.f32 %f649, %f5584; @%p253 bra $L__BB106_378; bra.uni $L__BB106_137; $L__BB106_378: setp.ltu.ftz.f32 %p258, %f649, 0f3727C5AC; mov.u64 %rd2935, 0; mov.u64 %rd2936, %rd2935; @%p258 bra $L__BB106_383; setp.eq.s32 %p259, %r168, 0; mov.u64 %rd2935, 0; mov.u64 %rd2933, %rd2935; mov.u64 %rd2934, %rd2935; @%p259 bra $L__BB106_381; add.s64 %rd2933, %rd1, 304; add.s64 %rd2934, %rd619, 304; mov.u64 %rd1046, 1065353216; st.local.u64 [%rd7+224], %rd1046; mov.u64 %rd1047, 4575657222473777152; st.local.u64 [%rd7+216], %rd1047; mov.u32 %r1160, 1; st.local.u32 [%rd7+172], %r1160; add.s32 %r1161, %r168, -1; st.local.u32 [%rd7+176], %r1161; $L__BB106_381: setp.eq.s64 %p260, %rd2934, 0; mov.u64 %rd2936, %rd2935; @%p260 bra $L__BB106_383; st.local.f32 [%rd2933+16], %f649; mov.u64 %rd2935, %rd2933; mov.u64 %rd2936, %rd2934; $L__BB106_383: setp.eq.s64 %p261, %rd2936, 0; @%p261 bra $L__BB106_385; st.local.f32 [%rd2935+20], %f15741; st.local.f32 [%rd2935+24], %f15742; st.local.f32 [%rd2935+28], %f15743; mov.u32 %r1162, 2; st.local.u32 [%rd2935+12], %r1162; ld.local.u32 %r1163, [%rd7+-36]; or.b32 %r1164, %r1163, 12; st.local.u32 [%rd7+-36], %r1164; bra.uni $L__BB106_385; $L__BB106_386: ld.const.u32 %r169, [kernel_params+1624]; setp.eq.s32 %p262, %r169, 0; @%p262 bra $L__BB106_520; and.b32 %r1167, %r169, 1; setp.eq.b32 %p263, %r1167, 1; mov.pred %p264, 0; xor.pred %p265, %p263, %p264; not.pred %p266, %p265; @%p266 bra $L__BB106_389; mov.u32 %r1168, 0; mov.f32 %f5585, 0f00000000; st.local.v2.f32 [%rd7+192], {%f5585, %f5585}; st.local.u32 [%rd7+200], %r1168; $L__BB106_389: and.b32 %r1169, %r169, 32; setp.eq.s32 %p267, %r1169, 0; @%p267 bra $L__BB106_391; and.b32 %r3134, %r3134, -9; st.local.u32 [%rd7+-36], %r3134; $L__BB106_391: and.b32 %r1170, %r45, 1; setp.eq.b32 %p268, %r1170, 1; xor.pred %p270, %p268, %p264; not.pred %p271, %p270; @%p271 bra $L__BB106_520; ld.local.u32 %r172, [%rd7+172]; setp.lt.s32 %p272, %r172, 1; @%p272 bra $L__BB106_520; and.b32 %r1171, %r169, 2; setp.eq.s32 %p273, %r1171, 0; and.b32 %r173, %r169, 4; and.b32 %r174, %r169, 8; and.b32 %r175, %r169, 16; add.s32 %r176, %r172, -1; and.b32 %r177, %r172, 3; @%p273 bra $L__BB106_462; bra.uni $L__BB106_394; $L__BB106_462: setp.eq.s32 %p368, %r173, 0; @%p368 bra $L__BB106_494; bra.uni $L__BB106_463; $L__BB106_494: setp.lt.u32 %p418, %r176, 3; mov.u32 %r3129, 0; @%p418 bra $L__BB106_513; sub.s32 %r3123, %r172, %r177; mov.u32 %r3129, 0; setp.eq.s32 %p420, %r174, 0; setp.eq.s32 %p422, %r175, 0; $L__BB106_496: cvt.s64.s32 %rd70, %r3129; mul.wide.s32 %rd1062, %r3129, 80; add.s64 %rd1063, %rd1, %rd1062; add.s64 %rd71, %rd1063, 316; ld.local.u32 %r266, [%rd1063+316]; add.s32 %r1266, %r266, -21; setp.gt.u32 %p419, %r1266, 9; or.pred %p421, %p419, %p420; @%p421 bra $L__BB106_498; bra.uni $L__BB106_497; $L__BB106_498: setp.ne.s32 %p423, %r266, 31; or.pred %p424, %p423, %p422; @%p424 bra $L__BB106_500; mov.u32 %r1268, 35; st.local.u32 [%rd71], %r1268; mov.u32 %r1269, 0; st.local.u32 [%rd71+4], %r1269; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_500; $L__BB106_497: mov.u32 %r1267, 0; st.local.u32 [%rd71], %r1267; st.local.u32 [%rd71+4], %r1267; $L__BB106_500: ld.local.u32 %r269, [%rd71+80]; add.s32 %r1270, %r269, -21; setp.gt.u32 %p425, %r1270, 9; or.pred %p427, %p425, %p420; @%p427 bra $L__BB106_502; bra.uni $L__BB106_501; $L__BB106_502: setp.ne.s32 %p429, %r269, 31; or.pred %p430, %p429, %p422; @%p430 bra $L__BB106_504; mov.u32 %r1272, 35; st.local.u32 [%rd71+80], %r1272; mov.u32 %r1273, 0; st.local.u32 [%rd71+84], %r1273; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_504; $L__BB106_501: mov.u32 %r1271, 0; st.local.u32 [%rd71+80], %r1271; st.local.u32 [%rd71+84], %r1271; $L__BB106_504: ld.local.u32 %r272, [%rd71+160]; add.s32 %r1274, %r272, -21; setp.gt.u32 %p431, %r1274, 9; or.pred %p433, %p431, %p420; @%p433 bra $L__BB106_506; bra.uni $L__BB106_505; $L__BB106_506: setp.ne.s32 %p435, %r272, 31; or.pred %p436, %p435, %p422; @%p436 bra $L__BB106_508; mov.u32 %r1276, 35; st.local.u32 [%rd71+160], %r1276; mov.u32 %r1277, 0; st.local.u32 [%rd71+164], %r1277; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_508; $L__BB106_505: mov.u32 %r1275, 0; st.local.u32 [%rd71+160], %r1275; st.local.u32 [%rd71+164], %r1275; $L__BB106_508: ld.local.u32 %r275, [%rd71+240]; add.s32 %r1278, %r275, -21; setp.gt.u32 %p437, %r1278, 9; or.pred %p439, %p437, %p420; @%p439 bra $L__BB106_510; bra.uni $L__BB106_509; $L__BB106_510: setp.ne.s32 %p441, %r275, 31; or.pred %p442, %p441, %p422; @%p442 bra $L__BB106_512; mov.u32 %r1280, 35; st.local.u32 [%rd71+240], %r1280; mov.u32 %r1281, 0; st.local.u32 [%rd71+244], %r1281; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_512; $L__BB106_509: mov.u32 %r1279, 0; st.local.u32 [%rd71+240], %r1279; st.local.u32 [%rd71+244], %r1279; $L__BB106_512: cvt.u32.u64 %r1282, %rd70; add.s32 %r3129, %r1282, 4; add.s32 %r3123, %r3123, -4; setp.ne.s32 %p443, %r3123, 0; @%p443 bra $L__BB106_496; $L__BB106_513: setp.eq.s32 %p444, %r177, 0; @%p444 bra $L__BB106_520; setp.eq.s32 %p446, %r174, 0; setp.eq.s32 %p448, %r175, 0; $L__BB106_515: .pragma "nounroll"; cvt.s64.s32 %rd72, %r3129; mul.wide.s32 %rd1064, %r3129, 80; add.s64 %rd1065, %rd1, %rd1064; add.s64 %rd73, %rd1065, 316; ld.local.u32 %r285, [%rd1065+316]; add.s32 %r1283, %r285, -21; setp.gt.u32 %p445, %r1283, 9; or.pred %p447, %p445, %p446; @%p447 bra $L__BB106_517; bra.uni $L__BB106_516; $L__BB106_517: setp.ne.s32 %p449, %r285, 31; or.pred %p450, %p449, %p448; @%p450 bra $L__BB106_519; mov.u32 %r1285, 35; st.local.u32 [%rd73], %r1285; mov.u32 %r1286, 0; st.local.u32 [%rd73+4], %r1286; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_519; $L__BB106_516: mov.u32 %r1284, 0; st.local.u32 [%rd73], %r1284; st.local.u32 [%rd73+4], %r1284; $L__BB106_519: cvt.u32.u64 %r1287, %rd72; add.s32 %r3129, %r1287, 1; add.s32 %r177, %r177, -1; setp.ne.s32 %p451, %r177, 0; @%p451 bra $L__BB106_515; bra.uni $L__BB106_520; $L__BB106_137: setp.ltu.ftz.f32 %p254, %f649, 0f3727C5AC; mov.u64 %rd2931, 0; mov.u64 %rd2932, %rd2931; @%p254 bra $L__BB106_142; setp.eq.s32 %p255, %r168, 0; mov.u64 %rd2931, 0; mov.u64 %rd2929, %rd2931; mov.u64 %rd2930, %rd2931; @%p255 bra $L__BB106_140; add.s64 %rd2929, %rd1, 304; add.s64 %rd2930, %rd619, 304; mov.u64 %rd1037, 1065353216; st.local.u64 [%rd7+224], %rd1037; mov.u64 %rd1038, 4575657222473777152; st.local.u64 [%rd7+216], %rd1038; mov.u32 %r1153, 1; st.local.u32 [%rd7+172], %r1153; add.s32 %r1154, %r168, -1; st.local.u32 [%rd7+176], %r1154; $L__BB106_140: setp.eq.s64 %p256, %rd2930, 0; mov.u64 %rd2932, %rd2931; @%p256 bra $L__BB106_142; st.local.f32 [%rd2929+16], %f649; mov.u64 %rd2931, %rd2929; mov.u64 %rd2932, %rd2930; $L__BB106_142: setp.eq.s64 %p257, %rd2932, 0; @%p257 bra $L__BB106_385; st.local.f32 [%rd2931+20], %f15741; st.local.f32 [%rd2931+24], %f15742; st.local.f32 [%rd2931+28], %f15743; mov.u32 %r1155, 2139095039; st.local.u32 [%rd2931+32], %r1155; mov.u32 %r1156, 5; st.local.u32 [%rd2931+12], %r1156; mov.u32 %r1157, 4; st.local.u32 [%rd2931+36], %r1157; ld.local.u32 %r1158, [%rd7+-36]; or.b32 %r1159, %r1158, 12; st.local.u32 [%rd7+-36], %r1159; $L__BB106_385: ld.global.u32 %r1165, [%rd29]; and.b32 %r1166, %r1165, -15728641; st.global.u32 [%rd29], %r1166; $L__BB106_1312: ld.const.v2.u64 {%rd1702, %rd1703}, [kernel_params+2528]; cvta.to.global.u64 %rd1706, %rd1702; add.s64 %rd277, %rd1706, %rd837; ld.global.u32 %r601, [%rd277]; cvta.to.global.u64 %rd1708, %rd1703; shl.b64 %rd1709, %rd2, 1; add.s64 %rd278, %rd1708, %rd1709; ld.global.u16 %rs28, [%rd278]; cvt.u32.u16 %r602, %rs28; ld.const.u64 %rd1710, [kernel_params+2456]; cvta.to.global.u64 %rd1711, %rd1710; add.s64 %rd279, %rd1711, %rd837; ld.global.u32 %r603, [%rd279]; ld.const.u32 %r1894, [kernel_params+1488]; setp.eq.s32 %p1237, %r1894, 0; ld.local.u8 %rs83, [%rd7+-36]; and.b16 %rs84, %rs83, 8; setp.eq.s16 %p1238, %rs84, 0; or.pred %p1239, %p1237, %p1238; @%p1239 bra $L__BB106_1685; ld.global.u32 %r604, [%rd29]; ld.const.u64 %rd1712, [kernel_params+2464]; cvta.to.global.u64 %rd1713, %rd1712; add.s64 %rd280, %rd1713, %rd1709; ld.global.u16 %rs29, [%rd280]; add.s32 %r605, %r602, 1; ld.const.u32 %r606, [kernel_params+1596]; setp.eq.s32 %p1240, %r606, 0; brev.b32 %r1895, %r603; mul.lo.s32 %r1896, %r1895, 1025551850; xor.b32 %r607, %r1896, %r1895; @%p1240 bra $L__BB106_1316; ld.const.v2.u32 {%r1897, %r1898}, [kernel_params+1600]; mov.b32 %f1922, %r1898; setp.geu.ftz.f32 %p1241, %f1922, 0f3F800000; setp.lt.ftz.f32 %p1242, %f1922, 0f3F800000; ld.const.u32 %r1901, [kernel_params+1576]; selp.b32 %r1902, %r1901, %r601, %p1242; add.s16 %rs85, %rs28, 1; and.b16 %rs86, %rs85, 255; cvt.u32.u16 %r1903, %rs86; xor.b32 %r1904, %r1902, %r1903; mul.lo.s32 %r1905, %r1904, -512718531; shr.u32 %r1906, %r1902, 16; xor.b32 %r1907, %r1905, %r1906; bfe.u32 %r1908, %r1907, 4, 4; shr.u32 %r1909, %r1902, 8; xor.b32 %r1910, %r1907, %r1909; xor.b32 %r1911, %r1910, %r1908; mul.lo.s32 %r1912, %r1911, 153742143; shr.u32 %r1913, %r1902, 23; xor.b32 %r1914, %r1912, %r1913; bfe.u32 %r1915, %r1914, 1, 7; xor.b32 %r1916, %r1915, %r1914; shr.u32 %r1917, %r1902, 27; or.b32 %r1918, %r1917, 1; mul.lo.s32 %r1919, %r1918, %r1916; mul.lo.s32 %r1920, %r1919, -1074439621; bfe.u32 %r1921, %r1920, 2, 6; xor.b32 %r1922, %r1921, %r1920; mul.lo.s32 %r1923, %r1922, -1638916925; bfe.u32 %r1924, %r1923, 2, 6; xor.b32 %r1925, %r1924, %r1923; mul.lo.s32 %r1926, %r1925, -933190689; and.b32 %r1927, %r1926, 255; bfe.u32 %r1928, %r1926, 5, 3; xor.b32 %r1929, %r1928, %r1927; add.s32 %r1930, %r1897, -1; xor.b32 %r608, %r605, 61; xor.b32 %r1931, %r1902, %r608; mul.lo.s32 %r1932, %r1931, 9; shr.u32 %r1933, %r1932, 4; xor.b32 %r1934, %r1933, %r1932; mul.lo.s32 %r1935, %r1934, 668265261; add.s32 %r1936, %r1935, %r607; shr.u32 %r1937, %r1935, 16; or.b32 %r1938, %r1937, 1; mul.lo.s32 %r1939, %r1938, %r1936; mul.lo.s32 %r1940, %r1939, 89287766; xor.b32 %r1941, %r1940, %r1939; mul.lo.s32 %r1942, %r1941, 1403136100; xor.b32 %r1943, %r1942, %r1941; brev.b32 %r1944, %r1943; neg.s32 %r1945, %r1897; and.b32 %r1946, %r603, %r1945; and.b32 %r1947, %r1944, %r1930; or.b32 %r1948, %r1947, %r1946; mad.lo.s32 %r1949, %r1929, %r1897, %r1948; shl.b32 %r1950, %r1897, 8; rem.u32 %r1951, %r1949, %r1950; ld.const.u64 %rd1715, [kernel_params+2424]; cvta.to.global.u64 %rd1716, %rd1715; shl.b32 %r1952, %r1951, 1; mul.wide.u32 %rd1717, %r1952, 4; add.s64 %rd1718, %rd1716, %rd1717; ld.global.f32 %f16177, [%rd1718]; ld.global.f32 %f16178, [%rd1718+4]; @%p1241 bra $L__BB106_1320; xor.b32 %r1953, %r608, %r601; mul.lo.s32 %r1954, %r1953, 9; shr.u32 %r1955, %r1954, 4; xor.b32 %r1956, %r1955, %r1954; mul.lo.s32 %r1957, %r1956, 668265261; cvt.rn.f32.u32 %f9676, %r1957; mul.ftz.f32 %f9677, %f9676, 0f2F7FFFFE; xor.b32 %r1958, %r601, %r605; xor.b32 %r1959, %r1958, -905047700; mul.lo.s32 %r1960, %r1959, 9; shr.u32 %r1961, %r1960, 4; xor.b32 %r1962, %r1961, %r1960; mul.lo.s32 %r1963, %r1962, 668265261; cvt.rn.f32.u32 %f9678, %r1963; mul.ftz.f32 %f9679, %f9678, 0f2F7FFFFE; fma.rn.ftz.f32 %f9680, %f9677, %f1922, %f16177; fma.rn.ftz.f32 %f9681, %f9679, %f1922, %f16178; cvt.rmi.ftz.f32.f32 %f9682, %f9680; sub.ftz.f32 %f16177, %f9680, %f9682; cvt.rmi.ftz.f32.f32 %f9683, %f9681; sub.ftz.f32 %f16178, %f9681, %f9683; bra.uni $L__BB106_1320; $L__BB106_1316: shr.u32 %r1965, %r605, 16; xor.b32 %r1966, %r1965, %r605; mul.lo.s32 %r1967, %r1966, 569420461; shr.u32 %r1968, %r1967, 15; xor.b32 %r1969, %r1968, %r1967; mul.lo.s32 %r1970, %r1969, -749064809; shr.u32 %r1971, %r1970, 15; xor.b32 %r1972, %r601, %r1970; xor.b32 %r1973, %r1972, %r1971; xor.b32 %r609, %r1973, -419546133; xor.b32 %r1974, %r1973, 508809841; add.s32 %r1975, %r1974, %r607; shr.u32 %r1976, %r1974, 16; or.b32 %r1977, %r1976, 1; mul.lo.s32 %r1978, %r1977, %r1975; mul.lo.s32 %r1979, %r1978, 89287766; xor.b32 %r1980, %r1979, %r1978; mul.lo.s32 %r1981, %r1980, 1403136100; xor.b32 %r610, %r1981, %r1980; setp.eq.s32 %p1243, %r610, 0; mov.u32 %r3225, 0; @%p1243 bra $L__BB106_1319; mov.u32 %r3222, 0; mov.u64 %rd1720, sobol_burley_table; mov.u32 %r3223, %r610; mov.u32 %r3225, %r3222; $L__BB106_1318: clz.b32 %r1984, %r3223; add.s32 %r1985, %r1984, %r3222; mul.wide.u32 %rd1719, %r1985, 4; add.s64 %rd1721, %rd1720, %rd1719; ld.const.u32 %r1986, [%rd1721+128]; xor.b32 %r3225, %r1986, %r3225; add.s32 %r3222, %r1985, 1; shl.b32 %r1987, %r3223, %r1984; shl.b32 %r3223, %r1987, 1; setp.ne.s32 %p1244, %r3223, 0; @%p1244 bra $L__BB106_1318; $L__BB106_1319: xor.b32 %r1988, %r609, -525684874; brev.b32 %r1989, %r610; shr.u32 %r1990, %r1988, 16; or.b32 %r1991, %r1990, 1; mul.lo.s32 %r1992, %r1989, 1025551850; xor.b32 %r1993, %r1992, %r1989; add.s32 %r1994, %r1993, %r1988; mul.lo.s32 %r1995, %r1994, %r1991; mul.lo.s32 %r1996, %r1995, 89287766; xor.b32 %r1997, %r1996, %r1995; xor.b32 %r1998, %r609, -1802089138; mul.lo.s32 %r1999, %r3225, 1025551850; xor.b32 %r2000, %r1999, %r3225; add.s32 %r2001, %r2000, %r1998; shr.u32 %r2002, %r1998, 16; or.b32 %r2003, %r2002, 1; mul.lo.s32 %r2004, %r2001, %r2003; mul.lo.s32 %r2005, %r2004, 89287766; xor.b32 %r2006, %r2005, %r2004; mul.lo.s32 %r2007, %r2006, 1403136100; xor.b32 %r2008, %r2007, %r2006; brev.b32 %r2009, %r2008; cvt.rn.f32.u32 %f9684, %r2009; mul.lo.s32 %r2010, %r1997, 1403136100; xor.b32 %r2011, %r2010, %r1997; brev.b32 %r2012, %r2011; cvt.rn.f32.u32 %f9685, %r2012; mul.ftz.f32 %f16178, %f9684, 0f2F7FFFFE; mul.ftz.f32 %f16177, %f9685, 0f2F7FFFFE; $L__BB106_1320: ld.local.f32 %f9686, [%rd7+-4]; ld.local.v4.f32 {%f9687, %f9688, %f9689, %f9690}, [%rd1]; cvt.u32.u16 %r2013, %rs29; add.u64 %rd1723, %SP, 0; { // callseq 170, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f16177; .param .b32 param1; st.param.f32 [param1+0], %f16178; .param .b32 param2; st.param.f32 [param2+0], %f9686; .param .align 4 .b8 param3[12]; st.param.f32 [param3+0], %f9687; st.param.f32 [param3+4], %f9688; st.param.f32 [param3+8], %f9689; .param .b32 param4; st.param.b32 [param4+0], %r2013; .param .b32 param5; st.param.b32 [param5+0], %r604; .param .b64 param6; st.param.b64 [param6+0], %rd1723; .param .b32 retval0; call.uni (retval0), _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample, ( param0, param1, param2, param3, param4, param5, param6 ); ld.param.b32 %r2014, [retval0+0]; } // callseq 170 cvt.u16.u32 %rs87, %r2014; setp.eq.s16 %p1245, %rs87, 0; @%p1245 bra $L__BB106_1685; ld.const.u64 %rd2920, [kernel_params+2408]; cvta.to.global.u64 %rd2919, %rd2920; cvta.to.local.u64 %rd1725, %rd1723; add.s64 %rd282, %rd1725, 24; ld.local.f32 %f9697, [%rd7+-76]; ld.local.f32 %f1935, [%rd1725+24]; ld.local.v2.f32 {%f9698, %f9699}, [%rd7+-72]; ld.local.f32 %f1936, [%rd1725+28]; mul.ftz.f32 %f9702, %f1936, %f9698; fma.rn.ftz.f32 %f9703, %f1935, %f9697, %f9702; ld.local.f32 %f1937, [%rd1725+32]; fma.rn.ftz.f32 %f1938, %f1937, %f9699, %f9703; ld.local.f32 %f1939, [%rd7+-4]; ld.local.u32 %r618, [%rd1725+64]; and.b32 %r3229, %r618, 4194303; shl.b32 %r2015, %r618, 5; cvt.u64.u32 %rd1726, %r2015; and.b64 %rd1727, %rd1726, 134217696; add.s64 %rd1728, %rd2919, %rd1727; add.s64 %rd283, %rd1728, 16; ld.global.u32 %r620, [%rd1728+16]; and.b32 %r2016, %r620, 134217728; setp.eq.s32 %p1247, %r2016, 0; mov.pred %p2184, -1; mov.f32 %f16220, 0f00000000; mov.f32 %f16221, %f16220; mov.f32 %f16222, %f16220; @%p1247 bra $L__BB106_1323; ld.global.f32 %f16220, [%rd283+-16]; ld.global.f32 %f16221, [%rd283+-12]; ld.global.f32 %f16222, [%rd283+-8]; mov.pred %p2184, 0; $L__BB106_1323: @%p2184 bra $L__BB106_1327; bra.uni $L__BB106_1324; $L__BB106_1327: ld.local.u32 %r621, [%rd282+52]; setp.eq.s32 %p1251, %r621, 2; ld.local.f32 %f1949, [%rd282+-24]; ld.local.f32 %f1950, [%rd282+-20]; ld.local.f32 %f1951, [%rd282+-16]; add.u64 %rd1729, %SP, 80; or.b64 %rd284, %rd1729, 12; neg.ftz.f32 %f1952, %f1935; neg.ftz.f32 %f1953, %f1936; neg.ftz.f32 %f1954, %f1937; @%p1251 bra $L__BB106_1365; bra.uni $L__BB106_1328; $L__BB106_1365: ld.const.u64 %rd2923, [kernel_params+2408]; cvta.to.global.u64 %rd2922, %rd2923; add.s64 %rd1888, %rd284, 20; add.s64 %rd1889, %rd284, 4; add.s64 %rd1891, %rd1729, 248; cvta.to.local.u64 %rd1892, %rd1891; add.s64 %rd1893, %rd1729, 240; cvta.to.local.u64 %rd1894, %rd1893; st.local.v2.f32 [%rd1894], {%f1949, %f1950}; st.local.f32 [%rd1892], %f1951; cvta.to.local.u64 %rd1895, %rd1729; st.local.v4.f32 [%rd1895], {%f1935, %f1936, %f1937, %f1952}; cvta.to.local.u64 %rd1896, %rd1889; st.local.v4.f32 [%rd1896], {%f1953, %f1954, %f1952, %f1953}; cvta.to.local.u64 %rd1897, %rd1888; st.local.v4.f32 [%rd1897], {%f1954, %f1952, %f1953, %f1954}; ld.const.u32 %r2048, [kernel_params+1024]; and.b32 %r3229, %r2048, 4194303; shl.b32 %r2049, %r2048, 5; cvt.u64.u32 %rd1898, %r2049; and.b64 %rd1899, %rd1898, 134217696; add.s64 %rd1900, %rd2922, %rd1899; add.s64 %rd1901, %rd1729, 76; cvta.to.local.u64 %rd1902, %rd1901; ld.global.u32 %r2050, [%rd1900+16]; mov.u32 %r2051, -1; st.local.u32 [%rd1902], %r2051; mov.u32 %r2052, 0; add.s64 %rd1903, %rd284, 36; cvta.to.local.u64 %rd1904, %rd1903; st.local.v4.u32 [%rd1904], {%r2048, %r2050, %r2052, %r2051}; add.s64 %rd1905, %rd1729, 68; cvta.to.local.u64 %rd1906, %rd1905; st.local.u32 [%rd1906], %r2052; add.s64 %rd1907, %rd1729, 72; cvta.to.local.u64 %rd1908, %rd1907; st.local.u32 [%rd1908], %r2052; add.s64 %rd1909, %rd1729, 136; cvta.to.local.u64 %rd1910, %rd1909; add.s64 %rd1911, %rd1729, 128; cvta.to.local.u64 %rd1912, %rd1911; mov.f32 %f10187, 0f00000000; st.local.v2.f32 [%rd1912], {%f10187, %f10187}; st.local.u32 [%rd1910], %r2052; mov.b32 %r2053, %f1939; add.s64 %rd1913, %rd1729, 80; cvta.to.local.u64 %rd1914, %rd1913; st.local.v4.u32 [%rd1914], {%r2051, %r2053, %r2052, %r2052}; add.s64 %rd1915, %rd1729, 96; cvta.to.local.u64 %rd1916, %rd1915; st.local.v4.f32 [%rd1916], {%f10187, %f10187, %f10187, %f10187}; add.s64 %rd1917, %rd1729, 112; cvta.to.local.u64 %rd1918, %rd1917; st.local.v4.f32 [%rd1918], {%f10187, %f10187, %f10187, %f10187}; bra.uni $L__BB106_1366; $L__BB106_1324: ld.local.u32 %r2017, [%rd282+36]; setp.eq.s32 %p1249, %r2017, -1; @%p1249 bra $L__BB106_1577; ld.local.f32 %f1946, [%rd282+-12]; ld.local.f32 %f1947, [%rd282+-8]; mul.ftz.f32 %f9704, %f1947, %f1936; fma.rn.ftz.f32 %f9705, %f1946, %f1935, %f9704; ld.local.f32 %f1948, [%rd282+-4]; fma.rn.ftz.f32 %f9706, %f1948, %f1937, %f9705; setp.leu.ftz.f32 %p1250, %f9706, 0f00000000; @%p1250 bra $L__BB106_1577; neg.ftz.f32 %f9707, %f1946; st.local.f32 [%rd282+-12], %f9707; neg.ftz.f32 %f9708, %f1947; st.local.f32 [%rd282+-8], %f9708; neg.ftz.f32 %f9709, %f1948; st.local.f32 [%rd282+-4], %f9709; bra.uni $L__BB106_1577; $L__BB106_90: ld.local.v2.f32 {%f5238, %f5239}, [%rd7+32]; ld.local.v4.f32 {%f5242, %f5243, %f5244, %f5245}, [%rd7+56]; ld.local.f32 %f5249, [%rd7+28]; mul.ftz.f32 %f5250, %f5238, %f5243; fma.rn.ftz.f32 %f5251, %f5249, %f5242, %f5250; fma.rn.ftz.f32 %f15688, %f5239, %f5244, %f5251; ld.local.v4.f32 {%f5252, %f5253, %f5254, %f5255}, [%rd7+72]; mul.ftz.f32 %f5259, %f5238, %f5253; fma.rn.ftz.f32 %f5260, %f5249, %f5252, %f5259; fma.rn.ftz.f32 %f15689, %f5239, %f5254, %f5260; ld.local.v4.f32 {%f5261, %f5262, %f5263, %f5264}, [%rd7+88]; mul.ftz.f32 %f5268, %f5238, %f5262; fma.rn.ftz.f32 %f5269, %f5249, %f5261, %f5268; fma.rn.ftz.f32 %f15690, %f5239, %f5263, %f5269; $L__BB106_92: st.local.f32 [%rd7+28], %f15688; st.local.f32 [%rd7+32], %f15689; st.local.f32 [%rd7+36], %f15690; @%p71 bra $L__BB106_94; bra.uni $L__BB106_93; $L__BB106_94: ld.const.u64 %rd825, [kernel_params+2160]; cvta.to.global.u64 %rd826, %rd825; ld.local.u32 %r1029, [%rd7+-12]; mul.wide.s32 %rd827, %r1029, 208; add.s64 %rd828, %rd826, %rd827; ld.global.v4.f32 {%f5339, %f5340, %f5341, %f5342}, [%rd828]; ld.global.v4.f32 {%f5346, %f5347, %f5348, %f5349}, [%rd828+16]; ld.global.v4.f32 {%f5353, %f5354, %f5355, %f5356}, [%rd828+32]; ld.local.v4.f32 {%f5360, %f5361, %f5362, %f5363}, [%rd7+40]; mul.ftz.f32 %f5367, %f5340, %f5361; fma.rn.ftz.f32 %f5368, %f5360, %f5339, %f5367; mul.ftz.f32 %f5369, %f5347, %f5361; fma.rn.ftz.f32 %f5370, %f5360, %f5346, %f5369; mul.ftz.f32 %f5371, %f5354, %f5361; fma.rn.ftz.f32 %f5372, %f5360, %f5353, %f5371; fma.rn.ftz.f32 %f5373, %f5355, %f5362, %f5372; fma.rn.ftz.f32 %f5374, %f5348, %f5362, %f5370; fma.rn.ftz.f32 %f5375, %f5341, %f5362, %f5368; st.local.v2.f32 [%rd7+40], {%f5375, %f5374}; st.local.f32 [%rd7+48], %f5373; bra.uni $L__BB106_95; $L__BB106_93: ld.local.v4.f32 {%f5302, %f5303, %f5304, %f5305}, [%rd7+40]; ld.local.v4.f32 {%f5309, %f5310, %f5311, %f5312}, [%rd7+56]; mul.ftz.f32 %f5316, %f5303, %f5310; fma.rn.ftz.f32 %f5317, %f5302, %f5309, %f5316; ld.local.v4.f32 {%f5318, %f5319, %f5320, %f5321}, [%rd7+72]; mul.ftz.f32 %f5325, %f5303, %f5319; fma.rn.ftz.f32 %f5326, %f5302, %f5318, %f5325; ld.local.v4.f32 {%f5327, %f5328, %f5329, %f5330}, [%rd7+88]; mul.ftz.f32 %f5334, %f5303, %f5328; fma.rn.ftz.f32 %f5335, %f5302, %f5327, %f5334; fma.rn.ftz.f32 %f5336, %f5304, %f5329, %f5335; fma.rn.ftz.f32 %f5337, %f5304, %f5320, %f5326; fma.rn.ftz.f32 %f5338, %f5304, %f5311, %f5317; st.local.v2.f32 [%rd7+40], {%f5338, %f5337}; st.local.f32 [%rd7+48], %f5336; bra.uni $L__BB106_95; $L__BB106_1328: add.s64 %rd1730, %rd284, 20; add.s64 %rd1731, %rd284, 4; add.s64 %rd1732, %rd284, 36; ld.local.f32 %f16212, [%rd282+-12]; ld.local.f32 %f16213, [%rd282+-8]; ld.local.f32 %f1957, [%rd282+-4]; ld.local.u32 %r622, [%rd282+32]; cvt.s64.s32 %rd285, %r622; ld.local.u32 %r623, [%rd282+36]; cvt.s64.s32 %rd286, %r623; ld.local.f32 %f1958, [%rd282+16]; ld.local.f32 %f1959, [%rd282+20]; ld.local.f32 %f1960, [%rd282+12]; ld.local.u32 %r624, [%rd282+44]; cvta.to.local.u64 %rd1734, %rd1729; st.local.v4.f32 [%rd1734], {%f1949, %f1950, %f1951, %f16212}; cvta.to.local.u64 %rd287, %rd1731; st.local.v4.f32 [%rd287], {%f16213, %f1957, %f16212, %f16213}; cvta.to.local.u64 %rd1735, %rd1730; st.local.v4.f32 [%rd1735], {%f1957, %f1952, %f1953, %f1954}; cvta.to.local.u64 %rd1736, %rd1732; st.local.u32 [%rd1736], %r618; setp.eq.s32 %p1252, %r623, -1; add.s64 %rd288, %rd1729, 64; @%p1252 bra $L__BB106_1330; cvta.to.local.u64 %rd1737, %rd288; mov.u32 %r3226, 1; st.local.u32 [%rd1737], %r3226; bra.uni $L__BB106_1333; $L__BB106_394: setp.eq.s32 %p274, %r173, 0; @%p274 bra $L__BB106_431; bra.uni $L__BB106_395; $L__BB106_431: setp.lt.u32 %p329, %r176, 3; mov.u32 %r3103, 0; @%p329 bra $L__BB106_454; sub.s32 %r3097, %r172, %r177; mov.u32 %r3103, 0; setp.ne.s32 %p352, %r174, 0; setp.eq.s32 %p355, %r175, 0; $L__BB106_433: cvt.s64.s32 %rd62, %r3103; mul.wide.s32 %rd1054, %r3103, 80; add.s64 %rd1055, %rd1, %rd1054; add.s64 %rd63, %rd1055, 316; ld.local.u32 %r210, [%rd1055+316]; add.s32 %r1208, %r210, -2; setp.lt.u32 %p330, %r1208, 7; @%p330 bra $L__BB106_437; add.s32 %r1209, %r210, -21; setp.lt.u32 %p332, %r1209, 10; and.pred %p333, %p332, %p352; @%p333 bra $L__BB106_437; setp.ne.s32 %p335, %r210, 31; or.pred %p336, %p335, %p355; @%p336 bra $L__BB106_438; mov.u32 %r1210, 35; st.local.u32 [%rd63], %r1210; mov.u32 %r1211, 0; st.local.u32 [%rd63+4], %r1211; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_438; $L__BB106_437: mov.u32 %r1212, 0; st.local.u32 [%rd63], %r1212; st.local.u32 [%rd63+4], %r1212; $L__BB106_438: ld.local.u32 %r213, [%rd63+80]; add.s32 %r1213, %r213, -2; setp.lt.u32 %p337, %r1213, 7; @%p337 bra $L__BB106_442; add.s32 %r1214, %r213, -21; setp.lt.u32 %p339, %r1214, 10; and.pred %p340, %p339, %p352; @%p340 bra $L__BB106_442; setp.ne.s32 %p342, %r213, 31; or.pred %p343, %p342, %p355; @%p343 bra $L__BB106_443; mov.u32 %r1215, 35; st.local.u32 [%rd63+80], %r1215; mov.u32 %r1216, 0; st.local.u32 [%rd63+84], %r1216; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_443; $L__BB106_442: mov.u32 %r1217, 0; st.local.u32 [%rd63+80], %r1217; st.local.u32 [%rd63+84], %r1217; $L__BB106_443: ld.local.u32 %r216, [%rd63+160]; add.s32 %r1218, %r216, -2; setp.lt.u32 %p344, %r1218, 7; @%p344 bra $L__BB106_447; add.s32 %r1219, %r216, -21; setp.lt.u32 %p346, %r1219, 10; and.pred %p347, %p346, %p352; @%p347 bra $L__BB106_447; setp.ne.s32 %p349, %r216, 31; or.pred %p350, %p349, %p355; @%p350 bra $L__BB106_448; mov.u32 %r1220, 35; st.local.u32 [%rd63+160], %r1220; mov.u32 %r1221, 0; st.local.u32 [%rd63+164], %r1221; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_448; $L__BB106_447: mov.u32 %r1222, 0; st.local.u32 [%rd63+160], %r1222; st.local.u32 [%rd63+164], %r1222; $L__BB106_448: ld.local.u32 %r219, [%rd63+240]; add.s32 %r1223, %r219, -2; setp.lt.u32 %p351, %r1223, 7; @%p351 bra $L__BB106_452; add.s32 %r1224, %r219, -21; setp.lt.u32 %p353, %r1224, 10; and.pred %p354, %p353, %p352; @%p354 bra $L__BB106_452; setp.ne.s32 %p356, %r219, 31; or.pred %p357, %p356, %p355; @%p357 bra $L__BB106_453; mov.u32 %r1225, 35; st.local.u32 [%rd63+240], %r1225; mov.u32 %r1226, 0; st.local.u32 [%rd63+244], %r1226; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_453; $L__BB106_452: mov.u32 %r1227, 0; st.local.u32 [%rd63+240], %r1227; st.local.u32 [%rd63+244], %r1227; $L__BB106_453: cvt.u32.u64 %r1228, %rd62; add.s32 %r3103, %r1228, 4; add.s32 %r3097, %r3097, -4; setp.ne.s32 %p358, %r3097, 0; @%p358 bra $L__BB106_433; $L__BB106_454: setp.eq.s32 %p359, %r177, 0; @%p359 bra $L__BB106_520; setp.ne.s32 %p361, %r174, 0; setp.eq.s32 %p364, %r175, 0; $L__BB106_456: .pragma "nounroll"; cvt.s64.s32 %rd64, %r3103; mul.wide.s32 %rd1056, %r3103, 80; add.s64 %rd1057, %rd1, %rd1056; add.s64 %rd65, %rd1057, 316; ld.local.u32 %r229, [%rd1057+316]; add.s32 %r1229, %r229, -2; setp.lt.u32 %p360, %r1229, 7; @%p360 bra $L__BB106_460; add.s32 %r1230, %r229, -21; setp.lt.u32 %p362, %r1230, 10; and.pred %p363, %p362, %p361; @%p363 bra $L__BB106_460; setp.ne.s32 %p365, %r229, 31; or.pred %p366, %p365, %p364; @%p366 bra $L__BB106_461; mov.u32 %r1231, 35; st.local.u32 [%rd65], %r1231; mov.u32 %r1232, 0; st.local.u32 [%rd65+4], %r1232; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_461; $L__BB106_460: mov.u32 %r1233, 0; st.local.u32 [%rd65], %r1233; st.local.u32 [%rd65+4], %r1233; $L__BB106_461: cvt.u32.u64 %r1234, %rd64; add.s32 %r3103, %r1234, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p367, %r177, 0; @%p367 bra $L__BB106_520; bra.uni $L__BB106_456; $L__BB106_1330: cvta.to.local.u64 %rd289, %rd288; setp.eq.s32 %p1253, %r624, -1; @%p1253 bra $L__BB106_1332; mov.u32 %r3226, 32; st.local.u32 [%rd289], %r3226; bra.uni $L__BB106_1333; $L__BB106_463: setp.lt.u32 %p369, %r176, 3; mov.u32 %r3116, 0; @%p369 bra $L__BB106_486; sub.s32 %r3110, %r172, %r177; mov.u32 %r3116, 0; setp.ne.s32 %p400, %r174, 0; setp.eq.s32 %p403, %r175, 0; $L__BB106_465: cvt.s64.s32 %rd66, %r3116; mul.wide.s32 %rd1058, %r3116, 80; add.s64 %rd1059, %rd1, %rd1058; add.s64 %rd67, %rd1059, 316; ld.local.u32 %r238, [%rd1059+316]; add.s32 %r1237, %r238, -9; setp.lt.u32 %p370, %r1237, 12; setp.eq.s32 %p371, %r238, 29; or.pred %p372, %p370, %p371; @%p372 bra $L__BB106_469; add.s32 %r1238, %r238, -21; setp.lt.u32 %p374, %r1238, 10; and.pred %p375, %p374, %p400; @%p375 bra $L__BB106_469; setp.ne.s32 %p377, %r238, 31; or.pred %p378, %p377, %p403; @%p378 bra $L__BB106_470; mov.u32 %r1239, 35; st.local.u32 [%rd67], %r1239; mov.u32 %r1240, 0; st.local.u32 [%rd67+4], %r1240; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_470; $L__BB106_469: mov.u32 %r1241, 0; st.local.u32 [%rd67], %r1241; st.local.u32 [%rd67+4], %r1241; $L__BB106_470: ld.local.u32 %r241, [%rd67+80]; add.s32 %r1242, %r241, -9; setp.lt.u32 %p379, %r1242, 12; setp.eq.s32 %p380, %r241, 29; or.pred %p381, %p379, %p380; @%p381 bra $L__BB106_474; add.s32 %r1243, %r241, -21; setp.lt.u32 %p383, %r1243, 10; and.pred %p384, %p383, %p400; @%p384 bra $L__BB106_474; setp.ne.s32 %p386, %r241, 31; or.pred %p387, %p386, %p403; @%p387 bra $L__BB106_475; mov.u32 %r1244, 35; st.local.u32 [%rd67+80], %r1244; mov.u32 %r1245, 0; st.local.u32 [%rd67+84], %r1245; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_475; $L__BB106_474: mov.u32 %r1246, 0; st.local.u32 [%rd67+80], %r1246; st.local.u32 [%rd67+84], %r1246; $L__BB106_475: ld.local.u32 %r244, [%rd67+160]; add.s32 %r1247, %r244, -9; setp.lt.u32 %p388, %r1247, 12; setp.eq.s32 %p389, %r244, 29; or.pred %p390, %p388, %p389; @%p390 bra $L__BB106_479; add.s32 %r1248, %r244, -21; setp.lt.u32 %p392, %r1248, 10; and.pred %p393, %p392, %p400; @%p393 bra $L__BB106_479; setp.ne.s32 %p395, %r244, 31; or.pred %p396, %p395, %p403; @%p396 bra $L__BB106_480; mov.u32 %r1249, 35; st.local.u32 [%rd67+160], %r1249; mov.u32 %r1250, 0; st.local.u32 [%rd67+164], %r1250; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_480; $L__BB106_479: mov.u32 %r1251, 0; st.local.u32 [%rd67+160], %r1251; st.local.u32 [%rd67+164], %r1251; $L__BB106_480: ld.local.u32 %r247, [%rd67+240]; add.s32 %r1252, %r247, -9; setp.lt.u32 %p397, %r1252, 12; setp.eq.s32 %p398, %r247, 29; or.pred %p399, %p397, %p398; @%p399 bra $L__BB106_484; add.s32 %r1253, %r247, -21; setp.lt.u32 %p401, %r1253, 10; and.pred %p402, %p401, %p400; @%p402 bra $L__BB106_484; setp.ne.s32 %p404, %r247, 31; or.pred %p405, %p404, %p403; @%p405 bra $L__BB106_485; mov.u32 %r1254, 35; st.local.u32 [%rd67+240], %r1254; mov.u32 %r1255, 0; st.local.u32 [%rd67+244], %r1255; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_485; $L__BB106_484: mov.u32 %r1256, 0; st.local.u32 [%rd67+240], %r1256; st.local.u32 [%rd67+244], %r1256; $L__BB106_485: cvt.u32.u64 %r1257, %rd66; add.s32 %r3116, %r1257, 4; add.s32 %r3110, %r3110, -4; setp.ne.s32 %p406, %r3110, 0; @%p406 bra $L__BB106_465; $L__BB106_486: setp.eq.s32 %p407, %r177, 0; @%p407 bra $L__BB106_520; setp.ne.s32 %p411, %r174, 0; setp.eq.s32 %p414, %r175, 0; $L__BB106_488: .pragma "nounroll"; cvt.s64.s32 %rd68, %r3116; mul.wide.s32 %rd1060, %r3116, 80; add.s64 %rd1061, %rd1, %rd1060; add.s64 %rd69, %rd1061, 316; ld.local.u32 %r257, [%rd1061+316]; add.s32 %r1258, %r257, -9; setp.lt.u32 %p408, %r1258, 12; setp.eq.s32 %p409, %r257, 29; or.pred %p410, %p408, %p409; @%p410 bra $L__BB106_492; add.s32 %r1259, %r257, -21; setp.lt.u32 %p412, %r1259, 10; and.pred %p413, %p412, %p411; @%p413 bra $L__BB106_492; setp.ne.s32 %p415, %r257, 31; or.pred %p416, %p415, %p414; @%p416 bra $L__BB106_493; mov.u32 %r1260, 35; st.local.u32 [%rd69], %r1260; mov.u32 %r1261, 0; st.local.u32 [%rd69+4], %r1261; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_493; $L__BB106_492: mov.u32 %r1262, 0; st.local.u32 [%rd69], %r1262; st.local.u32 [%rd69+4], %r1262; $L__BB106_493: cvt.u32.u64 %r1263, %rd68; add.s32 %r3116, %r1263, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p417, %r177, 0; @%p417 bra $L__BB106_520; bra.uni $L__BB106_488; $L__BB106_395: setp.lt.u32 %p275, %r176, 3; mov.u32 %r3090, 0; @%p275 bra $L__BB106_422; sub.s32 %r3084, %r172, %r177; mov.u32 %r3090, 0; setp.ne.s32 %p310, %r174, 0; setp.eq.s32 %p313, %r175, 0; $L__BB106_397: cvt.s64.s32 %rd58, %r3090; mul.wide.s32 %rd1050, %r3090, 80; add.s64 %rd1051, %rd1, %rd1050; add.s64 %rd59, %rd1051, 316; ld.local.u32 %r182, [%rd1051+316]; add.s32 %r1174, %r182, -2; setp.lt.u32 %p276, %r1174, 7; @%p276 bra $L__BB106_402; add.s32 %r1175, %r182, -9; setp.lt.u32 %p277, %r1175, 12; setp.eq.s32 %p278, %r182, 29; or.pred %p279, %p277, %p278; @%p279 bra $L__BB106_402; add.s32 %r1176, %r182, -21; setp.lt.u32 %p281, %r1176, 10; and.pred %p282, %p281, %p310; @%p282 bra $L__BB106_402; setp.ne.s32 %p284, %r182, 31; or.pred %p285, %p284, %p313; @%p285 bra $L__BB106_403; mov.u32 %r1177, 35; st.local.u32 [%rd59], %r1177; mov.u32 %r1178, 0; st.local.u32 [%rd59+4], %r1178; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_403; $L__BB106_402: mov.u32 %r1179, 0; st.local.u32 [%rd59], %r1179; st.local.u32 [%rd59+4], %r1179; $L__BB106_403: ld.local.u32 %r185, [%rd59+80]; add.s32 %r1180, %r185, -2; setp.lt.u32 %p286, %r1180, 7; @%p286 bra $L__BB106_408; add.s32 %r1181, %r185, -9; setp.lt.u32 %p287, %r1181, 12; setp.eq.s32 %p288, %r185, 29; or.pred %p289, %p287, %p288; @%p289 bra $L__BB106_408; add.s32 %r1182, %r185, -21; setp.lt.u32 %p291, %r1182, 10; and.pred %p292, %p291, %p310; @%p292 bra $L__BB106_408; setp.ne.s32 %p294, %r185, 31; or.pred %p295, %p294, %p313; @%p295 bra $L__BB106_409; mov.u32 %r1183, 35; st.local.u32 [%rd59+80], %r1183; mov.u32 %r1184, 0; st.local.u32 [%rd59+84], %r1184; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_409; $L__BB106_408: mov.u32 %r1185, 0; st.local.u32 [%rd59+80], %r1185; st.local.u32 [%rd59+84], %r1185; $L__BB106_409: ld.local.u32 %r188, [%rd59+160]; add.s32 %r1186, %r188, -2; setp.lt.u32 %p296, %r1186, 7; @%p296 bra $L__BB106_414; add.s32 %r1187, %r188, -9; setp.lt.u32 %p297, %r1187, 12; setp.eq.s32 %p298, %r188, 29; or.pred %p299, %p297, %p298; @%p299 bra $L__BB106_414; add.s32 %r1188, %r188, -21; setp.lt.u32 %p301, %r1188, 10; and.pred %p302, %p301, %p310; @%p302 bra $L__BB106_414; setp.ne.s32 %p304, %r188, 31; or.pred %p305, %p304, %p313; @%p305 bra $L__BB106_415; mov.u32 %r1189, 35; st.local.u32 [%rd59+160], %r1189; mov.u32 %r1190, 0; st.local.u32 [%rd59+164], %r1190; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_415; $L__BB106_414: mov.u32 %r1191, 0; st.local.u32 [%rd59+160], %r1191; st.local.u32 [%rd59+164], %r1191; $L__BB106_415: ld.local.u32 %r191, [%rd59+240]; add.s32 %r1192, %r191, -2; setp.lt.u32 %p306, %r1192, 7; @%p306 bra $L__BB106_420; add.s32 %r1193, %r191, -9; setp.lt.u32 %p307, %r1193, 12; setp.eq.s32 %p308, %r191, 29; or.pred %p309, %p307, %p308; @%p309 bra $L__BB106_420; add.s32 %r1194, %r191, -21; setp.lt.u32 %p311, %r1194, 10; and.pred %p312, %p311, %p310; @%p312 bra $L__BB106_420; setp.ne.s32 %p314, %r191, 31; or.pred %p315, %p314, %p313; @%p315 bra $L__BB106_421; mov.u32 %r1195, 35; st.local.u32 [%rd59+240], %r1195; mov.u32 %r1196, 0; st.local.u32 [%rd59+244], %r1196; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_421; $L__BB106_420: mov.u32 %r1197, 0; st.local.u32 [%rd59+240], %r1197; st.local.u32 [%rd59+244], %r1197; $L__BB106_421: cvt.u32.u64 %r1198, %rd58; add.s32 %r3090, %r1198, 4; add.s32 %r3084, %r3084, -4; setp.ne.s32 %p316, %r3084, 0; @%p316 bra $L__BB106_397; $L__BB106_422: setp.eq.s32 %p317, %r177, 0; @%p317 bra $L__BB106_520; setp.ne.s32 %p322, %r174, 0; setp.eq.s32 %p325, %r175, 0; $L__BB106_424: .pragma "nounroll"; cvt.s64.s32 %rd60, %r3090; mul.wide.s32 %rd1052, %r3090, 80; add.s64 %rd1053, %rd1, %rd1052; add.s64 %rd61, %rd1053, 316; ld.local.u32 %r201, [%rd1053+316]; add.s32 %r1199, %r201, -2; setp.lt.u32 %p318, %r1199, 7; @%p318 bra $L__BB106_429; add.s32 %r1200, %r201, -9; setp.lt.u32 %p319, %r1200, 12; setp.eq.s32 %p320, %r201, 29; or.pred %p321, %p319, %p320; @%p321 bra $L__BB106_429; add.s32 %r1201, %r201, -21; setp.lt.u32 %p323, %r1201, 10; and.pred %p324, %p323, %p322; @%p324 bra $L__BB106_429; setp.ne.s32 %p326, %r201, 31; or.pred %p327, %p326, %p325; @%p327 bra $L__BB106_430; mov.u32 %r1202, 35; st.local.u32 [%rd61], %r1202; mov.u32 %r1203, 0; st.local.u32 [%rd61+4], %r1203; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB106_430; $L__BB106_429: mov.u32 %r1204, 0; st.local.u32 [%rd61], %r1204; st.local.u32 [%rd61+4], %r1204; $L__BB106_430: cvt.u32.u64 %r1205, %rd60; add.s32 %r3090, %r1205, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p328, %r177, 0; @%p328 bra $L__BB106_520; bra.uni $L__BB106_424; $L__BB106_520: ld.const.v2.u64 {%rd1066, %rd1067}, [kernel_params+2464]; cvta.to.global.u64 %rd1070, %rd1066; shl.b64 %rd1071, %rd2, 1; add.s64 %rd74, %rd1070, %rd1071; ld.global.u16 %r1288, [%rd74]; cvta.to.global.u64 %rd1072, %rd1067; add.s64 %rd1073, %rd1072, %rd1071; ld.global.u16 %r1289, [%rd1073]; neg.s32 %r1290, %r1289; setp.ne.s32 %p452, %r1288, %r1290; ld.local.u32 %r291, [%rd7+172]; setp.lt.s32 %p453, %r291, 2; or.pred %p454, %p452, %p453; @%p454 bra $L__BB106_553; add.s32 %r292, %r291, -1; and.b32 %r3144, %r291, 3; setp.lt.u32 %p455, %r292, 3; mov.f32 %f15745, 0f00000000; mov.u32 %r3137, 0; @%p455 bra $L__BB106_532; sub.s32 %r3136, %r291, %r3144; mov.f32 %f15745, 0f00000000; mov.u32 %r3137, 0; $L__BB106_523: cvt.s64.s32 %rd75, %r3137; mul.wide.s32 %rd1074, %r3137, 80; add.s64 %rd1075, %rd1, %rd1074; add.s64 %rd76, %rd1075, 316; ld.local.u32 %r1293, [%rd1075+316]; setp.gt.s32 %p456, %r1293, 34; @%p456 bra $L__BB106_525; ld.local.f32 %f5589, [%rd76+4]; add.ftz.f32 %f15745, %f15745, %f5589; $L__BB106_525: ld.local.u32 %r1294, [%rd76+80]; setp.gt.s32 %p457, %r1294, 34; @%p457 bra $L__BB106_527; ld.local.f32 %f5590, [%rd76+84]; add.ftz.f32 %f15745, %f15745, %f5590; $L__BB106_527: ld.local.u32 %r1295, [%rd76+160]; setp.gt.s32 %p458, %r1295, 34; @%p458 bra $L__BB106_529; ld.local.f32 %f5591, [%rd76+164]; add.ftz.f32 %f15745, %f15745, %f5591; $L__BB106_529: ld.local.u32 %r1296, [%rd76+240]; setp.gt.s32 %p459, %r1296, 34; @%p459 bra $L__BB106_531; ld.local.f32 %f5592, [%rd76+244]; add.ftz.f32 %f15745, %f15745, %f5592; $L__BB106_531: cvt.u32.u64 %r1297, %rd75; add.s32 %r3137, %r1297, 4; add.s32 %r3136, %r3136, -4; setp.ne.s32 %p460, %r3136, 0; @%p460 bra $L__BB106_523; $L__BB106_532: setp.eq.s32 %p461, %r3144, 0; @%p461 bra $L__BB106_537; mov.u32 %r3139, %r3144; $L__BB106_534: .pragma "nounroll"; cvt.s64.s32 %rd77, %r3137; mul.wide.s32 %rd1076, %r3137, 80; add.s64 %rd1077, %rd1, %rd1076; add.s64 %rd78, %rd1077, 316; ld.local.u32 %r1298, [%rd1077+316]; setp.gt.s32 %p462, %r1298, 34; @%p462 bra $L__BB106_536; ld.local.f32 %f5593, [%rd78+4]; add.ftz.f32 %f15745, %f15745, %f5593; $L__BB106_536: cvt.u32.u64 %r1299, %rd77; add.s32 %r3137, %r1299, 1; add.s32 %r3139, %r3139, -1; setp.ne.s32 %p463, %r3139, 0; @%p463 bra $L__BB106_534; $L__BB106_537: setp.lt.s32 %p464, %r291, 1; @%p464 bra $L__BB106_553; mul.ftz.f32 %f665, %f15745, 0f3E000000; mov.u32 %r3142, 0; @%p455 bra $L__BB106_549; sub.s32 %r3141, %r291, %r3144; mov.u32 %r3142, 0; $L__BB106_540: cvt.s64.s32 %rd79, %r3142; mul.wide.s32 %rd1078, %r3142, 80; add.s64 %rd1079, %rd1, %rd1078; add.s64 %rd80, %rd1079, 316; ld.local.u32 %r1302, [%rd1079+316]; setp.gt.s32 %p466, %r1302, 34; @%p466 bra $L__BB106_542; ld.local.f32 %f5594, [%rd80+4]; max.ftz.f32 %f5595, %f5594, %f665; st.local.f32 [%rd80+4], %f5595; $L__BB106_542: ld.local.u32 %r1303, [%rd80+80]; setp.gt.s32 %p467, %r1303, 34; @%p467 bra $L__BB106_544; ld.local.f32 %f5596, [%rd80+84]; max.ftz.f32 %f5597, %f5596, %f665; st.local.f32 [%rd80+84], %f5597; $L__BB106_544: ld.local.u32 %r1304, [%rd80+160]; setp.gt.s32 %p468, %r1304, 34; @%p468 bra $L__BB106_546; ld.local.f32 %f5598, [%rd80+164]; max.ftz.f32 %f5599, %f5598, %f665; st.local.f32 [%rd80+164], %f5599; $L__BB106_546: ld.local.u32 %r1305, [%rd80+240]; setp.gt.s32 %p469, %r1305, 34; @%p469 bra $L__BB106_548; ld.local.f32 %f5600, [%rd80+244]; max.ftz.f32 %f5601, %f5600, %f665; st.local.f32 [%rd80+244], %f5601; $L__BB106_548: cvt.u32.u64 %r1306, %rd79; add.s32 %r3142, %r1306, 4; add.s32 %r3141, %r3141, -4; setp.ne.s32 %p470, %r3141, 0; @%p470 bra $L__BB106_540; $L__BB106_549: @%p461 bra $L__BB106_553; $L__BB106_550: .pragma "nounroll"; cvt.s64.s32 %rd81, %r3142; mul.wide.s32 %rd1080, %r3142, 80; add.s64 %rd1081, %rd1, %rd1080; add.s64 %rd82, %rd1081, 316; ld.local.u32 %r1307, [%rd1081+316]; setp.gt.s32 %p472, %r1307, 34; @%p472 bra $L__BB106_552; ld.local.f32 %f5602, [%rd82+4]; max.ftz.f32 %f5603, %f5602, %f665; st.local.f32 [%rd82+4], %f5603; $L__BB106_552: cvt.u32.u64 %r1308, %rd81; add.s32 %r3142, %r1308, 1; add.s32 %r3144, %r3144, -1; setp.ne.s32 %p473, %r3144, 0; @%p473 bra $L__BB106_550; $L__BB106_553: ld.const.f32 %f666, [kernel_params+1572]; setp.eq.ftz.f32 %p474, %f666, 0f7F7FFFFF; @%p474 bra $L__BB106_647; ld.const.u64 %rd1082, [kernel_params+2552]; cvta.to.global.u64 %rd1083, %rd1082; add.s64 %rd1084, %rd1083, %rd2; ld.global.u8 %rs66, [%rd1084]; and.b16 %rs67, %rs66, 1; setp.eq.b16 %p475, %rs67, 1; mov.pred %p476, 0; xor.pred %p477, %p475, %p476; @%p477 bra $L__BB106_647; ld.const.u64 %rd1085, [kernel_params+2568]; cvta.to.global.u64 %rd1086, %rd1085; add.s64 %rd1088, %rd1086, %rd837; ld.global.f32 %f5604, [%rd1088]; mul.ftz.f32 %f667, %f666, %f5604; setp.geu.ftz.f32 %p478, %f667, 0f3F800000; @%p478 bra $L__BB106_647; mov.f32 %f5605, 0f3F800000; sub.ftz.f32 %f5606, %f5605, %f667; sqrt.approx.ftz.f32 %f5607, %f5606; mul.ftz.f32 %f668, %f5607, 0f3F000000; setp.lt.s32 %p479, %r291, 1; @%p479 bra $L__BB106_647; max.s32 %r314, %r291, 1; add.s32 %r1310, %r314, -1; and.b32 %r3149, %r314, 3; setp.lt.u32 %p480, %r1310, 3; mov.u32 %r3147, 0; @%p480 bra $L__BB106_628; sub.s32 %r3146, %r314, %r3149; mov.u32 %r3147, 0; $L__BB106_559: cvt.s64.s32 %rd83, %r3147; mul.wide.s32 %rd1089, %r3147, 80; add.s64 %rd1090, %rd1, %rd1089; add.s64 %rd84, %rd1090, 316; ld.local.u32 %r319, [%rd1090+316]; setp.gt.s32 %p481, %r319, 31; @%p481 bra $L__BB106_576; setp.gt.s32 %p482, %r319, 15; @%p482 bra $L__BB106_565; add.s32 %r1312, %r319, -10; setp.lt.u32 %p490, %r1312, 3; @%p490 bra $L__BB106_569; add.s32 %r1313, %r319, -14; setp.lt.u32 %p491, %r1313, 2; @%p491 bra $L__BB106_575; setp.eq.s32 %p492, %r319, 13; @%p492 bra $L__BB106_564; bra.uni $L__BB106_576; $L__BB106_565: setp.gt.s32 %p483, %r319, 23; @%p483 bra $L__BB106_570; setp.eq.s32 %p487, %r319, 16; @%p487 bra $L__BB106_574; setp.eq.s32 %p488, %r319, 22; @%p488 bra $L__BB106_564; setp.eq.s32 %p489, %r319, 23; @%p489 bra $L__BB106_569; bra.uni $L__BB106_576; $L__BB106_569: ld.local.v2.f32 {%f5632, %f5633}, [%rd84+20]; max.ftz.f32 %f5636, %f668, %f5633; max.ftz.f32 %f5637, %f668, %f5632; st.local.v2.f32 [%rd84+20], {%f5637, %f5636}; bra.uni $L__BB106_576; $L__BB106_570: setp.eq.s32 %p484, %r319, 24; @%p484 bra $L__BB106_575; setp.eq.s32 %p485, %r319, 27; @%p485 bra $L__BB106_575; setp.ne.s32 %p486, %r319, 29; @%p486 bra $L__BB106_576; ld.local.f32 %f5608, [%rd84+32]; max.ftz.f32 %f5609, %f668, %f5608; st.local.f32 [%rd84+32], %f5609; ld.local.f32 %f5610, [%rd84+36]; max.ftz.f32 %f5611, %f668, %f5610; st.local.f32 [%rd84+36], %f5611; ld.local.f32 %f5612, [%rd84+48]; max.ftz.f32 %f5613, %f668, %f5612; st.local.f32 [%rd84+48], %f5613; bra.uni $L__BB106_576; $L__BB106_575: ld.local.v2.f32 {%f5620, %f5621}, [%rd84+20]; max.ftz.f32 %f5624, %f668, %f5621; max.ftz.f32 %f5625, %f668, %f5620; st.local.v2.f32 [%rd84+20], {%f5625, %f5624}; bra.uni $L__BB106_576; $L__BB106_564: ld.local.v2.f32 {%f5626, %f5627}, [%rd84+20]; max.ftz.f32 %f5630, %f668, %f5627; max.ftz.f32 %f5631, %f668, %f5626; st.local.v2.f32 [%rd84+20], {%f5631, %f5630}; bra.uni $L__BB106_576; $L__BB106_574: ld.local.v2.f32 {%f5614, %f5615}, [%rd84+20]; max.ftz.f32 %f5618, %f668, %f5615; max.ftz.f32 %f5619, %f668, %f5614; st.local.v2.f32 [%rd84+20], {%f5619, %f5618}; $L__BB106_576: ld.local.u32 %r320, [%rd84+80]; setp.gt.s32 %p493, %r320, 31; @%p493 bra $L__BB106_593; setp.gt.s32 %p494, %r320, 15; @%p494 bra $L__BB106_582; add.s32 %r1314, %r320, -10; setp.lt.u32 %p502, %r1314, 3; @%p502 bra $L__BB106_586; add.s32 %r1315, %r320, -14; setp.lt.u32 %p503, %r1315, 2; @%p503 bra $L__BB106_592; setp.eq.s32 %p504, %r320, 13; @%p504 bra $L__BB106_581; bra.uni $L__BB106_593; $L__BB106_582: setp.gt.s32 %p495, %r320, 23; @%p495 bra $L__BB106_587; setp.eq.s32 %p499, %r320, 16; @%p499 bra $L__BB106_591; setp.eq.s32 %p500, %r320, 22; @%p500 bra $L__BB106_581; setp.eq.s32 %p501, %r320, 23; @%p501 bra $L__BB106_586; bra.uni $L__BB106_593; $L__BB106_586: ld.local.v2.f32 {%f5662, %f5663}, [%rd84+100]; max.ftz.f32 %f5666, %f668, %f5663; max.ftz.f32 %f5667, %f668, %f5662; st.local.v2.f32 [%rd84+100], {%f5667, %f5666}; bra.uni $L__BB106_593; $L__BB106_587: setp.eq.s32 %p496, %r320, 24; @%p496 bra $L__BB106_592; setp.eq.s32 %p497, %r320, 27; @%p497 bra $L__BB106_592; setp.ne.s32 %p498, %r320, 29; @%p498 bra $L__BB106_593; ld.local.f32 %f5638, [%rd84+112]; max.ftz.f32 %f5639, %f668, %f5638; st.local.f32 [%rd84+112], %f5639; ld.local.f32 %f5640, [%rd84+116]; max.ftz.f32 %f5641, %f668, %f5640; st.local.f32 [%rd84+116], %f5641; ld.local.f32 %f5642, [%rd84+128]; max.ftz.f32 %f5643, %f668, %f5642; st.local.f32 [%rd84+128], %f5643; bra.uni $L__BB106_593; $L__BB106_592: ld.local.v2.f32 {%f5650, %f5651}, [%rd84+100]; max.ftz.f32 %f5654, %f668, %f5651; max.ftz.f32 %f5655, %f668, %f5650; st.local.v2.f32 [%rd84+100], {%f5655, %f5654}; bra.uni $L__BB106_593; $L__BB106_581: ld.local.v2.f32 {%f5656, %f5657}, [%rd84+100]; max.ftz.f32 %f5660, %f668, %f5657; max.ftz.f32 %f5661, %f668, %f5656; st.local.v2.f32 [%rd84+100], {%f5661, %f5660}; bra.uni $L__BB106_593; $L__BB106_591: ld.local.v2.f32 {%f5644, %f5645}, [%rd84+100]; max.ftz.f32 %f5648, %f668, %f5645; max.ftz.f32 %f5649, %f668, %f5644; st.local.v2.f32 [%rd84+100], {%f5649, %f5648}; $L__BB106_593: ld.local.u32 %r321, [%rd84+160]; setp.gt.s32 %p505, %r321, 31; @%p505 bra $L__BB106_610; setp.gt.s32 %p506, %r321, 15; @%p506 bra $L__BB106_599; add.s32 %r1316, %r321, -10; setp.lt.u32 %p514, %r1316, 3; @%p514 bra $L__BB106_603; add.s32 %r1317, %r321, -14; setp.lt.u32 %p515, %r1317, 2; @%p515 bra $L__BB106_609; setp.eq.s32 %p516, %r321, 13; @%p516 bra $L__BB106_598; bra.uni $L__BB106_610; $L__BB106_599: setp.gt.s32 %p507, %r321, 23; @%p507 bra $L__BB106_604; setp.eq.s32 %p511, %r321, 16; @%p511 bra $L__BB106_608; setp.eq.s32 %p512, %r321, 22; @%p512 bra $L__BB106_598; setp.eq.s32 %p513, %r321, 23; @%p513 bra $L__BB106_603; bra.uni $L__BB106_610; $L__BB106_603: ld.local.v2.f32 {%f5692, %f5693}, [%rd84+180]; max.ftz.f32 %f5696, %f668, %f5693; max.ftz.f32 %f5697, %f668, %f5692; st.local.v2.f32 [%rd84+180], {%f5697, %f5696}; bra.uni $L__BB106_610; $L__BB106_604: setp.eq.s32 %p508, %r321, 24; @%p508 bra $L__BB106_609; setp.eq.s32 %p509, %r321, 27; @%p509 bra $L__BB106_609; setp.ne.s32 %p510, %r321, 29; @%p510 bra $L__BB106_610; ld.local.f32 %f5668, [%rd84+192]; max.ftz.f32 %f5669, %f668, %f5668; st.local.f32 [%rd84+192], %f5669; ld.local.f32 %f5670, [%rd84+196]; max.ftz.f32 %f5671, %f668, %f5670; st.local.f32 [%rd84+196], %f5671; ld.local.f32 %f5672, [%rd84+208]; max.ftz.f32 %f5673, %f668, %f5672; st.local.f32 [%rd84+208], %f5673; bra.uni $L__BB106_610; $L__BB106_609: ld.local.v2.f32 {%f5680, %f5681}, [%rd84+180]; max.ftz.f32 %f5684, %f668, %f5681; max.ftz.f32 %f5685, %f668, %f5680; st.local.v2.f32 [%rd84+180], {%f5685, %f5684}; bra.uni $L__BB106_610; $L__BB106_598: ld.local.v2.f32 {%f5686, %f5687}, [%rd84+180]; max.ftz.f32 %f5690, %f668, %f5687; max.ftz.f32 %f5691, %f668, %f5686; st.local.v2.f32 [%rd84+180], {%f5691, %f5690}; bra.uni $L__BB106_610; $L__BB106_608: ld.local.v2.f32 {%f5674, %f5675}, [%rd84+180]; max.ftz.f32 %f5678, %f668, %f5675; max.ftz.f32 %f5679, %f668, %f5674; st.local.v2.f32 [%rd84+180], {%f5679, %f5678}; $L__BB106_610: ld.local.u32 %r322, [%rd84+240]; setp.gt.s32 %p517, %r322, 31; @%p517 bra $L__BB106_627; setp.gt.s32 %p518, %r322, 15; @%p518 bra $L__BB106_616; add.s32 %r1318, %r322, -10; setp.lt.u32 %p526, %r1318, 3; @%p526 bra $L__BB106_620; add.s32 %r1319, %r322, -14; setp.lt.u32 %p527, %r1319, 2; @%p527 bra $L__BB106_626; setp.eq.s32 %p528, %r322, 13; @%p528 bra $L__BB106_615; bra.uni $L__BB106_627; $L__BB106_616: setp.gt.s32 %p519, %r322, 23; @%p519 bra $L__BB106_621; setp.eq.s32 %p523, %r322, 16; @%p523 bra $L__BB106_625; setp.eq.s32 %p524, %r322, 22; @%p524 bra $L__BB106_615; setp.eq.s32 %p525, %r322, 23; @%p525 bra $L__BB106_620; bra.uni $L__BB106_627; $L__BB106_620: ld.local.v2.f32 {%f5722, %f5723}, [%rd84+260]; max.ftz.f32 %f5726, %f668, %f5723; max.ftz.f32 %f5727, %f668, %f5722; st.local.v2.f32 [%rd84+260], {%f5727, %f5726}; bra.uni $L__BB106_627; $L__BB106_621: setp.eq.s32 %p520, %r322, 24; @%p520 bra $L__BB106_626; setp.eq.s32 %p521, %r322, 27; @%p521 bra $L__BB106_626; setp.ne.s32 %p522, %r322, 29; @%p522 bra $L__BB106_627; ld.local.f32 %f5698, [%rd84+272]; max.ftz.f32 %f5699, %f668, %f5698; st.local.f32 [%rd84+272], %f5699; ld.local.f32 %f5700, [%rd84+276]; max.ftz.f32 %f5701, %f668, %f5700; st.local.f32 [%rd84+276], %f5701; ld.local.f32 %f5702, [%rd84+288]; max.ftz.f32 %f5703, %f668, %f5702; st.local.f32 [%rd84+288], %f5703; bra.uni $L__BB106_627; $L__BB106_626: ld.local.v2.f32 {%f5710, %f5711}, [%rd84+260]; max.ftz.f32 %f5714, %f668, %f5711; max.ftz.f32 %f5715, %f668, %f5710; st.local.v2.f32 [%rd84+260], {%f5715, %f5714}; bra.uni $L__BB106_627; $L__BB106_615: ld.local.v2.f32 {%f5716, %f5717}, [%rd84+260]; max.ftz.f32 %f5720, %f668, %f5717; max.ftz.f32 %f5721, %f668, %f5716; st.local.v2.f32 [%rd84+260], {%f5721, %f5720}; bra.uni $L__BB106_627; $L__BB106_625: ld.local.v2.f32 {%f5704, %f5705}, [%rd84+260]; max.ftz.f32 %f5708, %f668, %f5705; max.ftz.f32 %f5709, %f668, %f5704; st.local.v2.f32 [%rd84+260], {%f5709, %f5708}; $L__BB106_627: cvt.u32.u64 %r1320, %rd83; add.s32 %r3147, %r1320, 4; add.s32 %r3146, %r3146, -4; setp.ne.s32 %p529, %r3146, 0; @%p529 bra $L__BB106_559; $L__BB106_628: setp.eq.s32 %p530, %r3149, 0; @%p530 bra $L__BB106_647; $L__BB106_629: .pragma "nounroll"; cvt.s64.s32 %rd85, %r3147; mul.wide.s32 %rd1091, %r3147, 80; add.s64 %rd1092, %rd1, %rd1091; add.s64 %rd86, %rd1092, 316; ld.local.u32 %r328, [%rd1092+316]; setp.gt.s32 %p531, %r328, 31; @%p531 bra $L__BB106_646; setp.gt.s32 %p532, %r328, 15; @%p532 bra $L__BB106_635; add.s32 %r1321, %r328, -10; setp.lt.u32 %p540, %r1321, 3; @%p540 bra $L__BB106_639; add.s32 %r1322, %r328, -14; setp.lt.u32 %p541, %r1322, 2; @%p541 bra $L__BB106_645; setp.eq.s32 %p542, %r328, 13; @%p542 bra $L__BB106_634; bra.uni $L__BB106_646; $L__BB106_635: setp.gt.s32 %p533, %r328, 23; @%p533 bra $L__BB106_640; setp.eq.s32 %p537, %r328, 16; @%p537 bra $L__BB106_644; setp.eq.s32 %p538, %r328, 22; @%p538 bra $L__BB106_634; setp.eq.s32 %p539, %r328, 23; @%p539 bra $L__BB106_639; bra.uni $L__BB106_646; $L__BB106_639: ld.local.v2.f32 {%f5752, %f5753}, [%rd86+20]; max.ftz.f32 %f5756, %f668, %f5753; max.ftz.f32 %f5757, %f668, %f5752; st.local.v2.f32 [%rd86+20], {%f5757, %f5756}; bra.uni $L__BB106_646; $L__BB106_640: setp.eq.s32 %p534, %r328, 24; @%p534 bra $L__BB106_645; setp.eq.s32 %p535, %r328, 27; @%p535 bra $L__BB106_645; setp.ne.s32 %p536, %r328, 29; @%p536 bra $L__BB106_646; ld.local.f32 %f5728, [%rd86+32]; max.ftz.f32 %f5729, %f668, %f5728; st.local.f32 [%rd86+32], %f5729; ld.local.f32 %f5730, [%rd86+36]; max.ftz.f32 %f5731, %f668, %f5730; st.local.f32 [%rd86+36], %f5731; ld.local.f32 %f5732, [%rd86+48]; max.ftz.f32 %f5733, %f668, %f5732; st.local.f32 [%rd86+48], %f5733; bra.uni $L__BB106_646; $L__BB106_645: ld.local.v2.f32 {%f5740, %f5741}, [%rd86+20]; max.ftz.f32 %f5744, %f668, %f5741; max.ftz.f32 %f5745, %f668, %f5740; st.local.v2.f32 [%rd86+20], {%f5745, %f5744}; bra.uni $L__BB106_646; $L__BB106_634: ld.local.v2.f32 {%f5746, %f5747}, [%rd86+20]; max.ftz.f32 %f5750, %f668, %f5747; max.ftz.f32 %f5751, %f668, %f5746; st.local.v2.f32 [%rd86+20], {%f5751, %f5750}; bra.uni $L__BB106_646; $L__BB106_644: ld.local.v2.f32 {%f5734, %f5735}, [%rd86+20]; max.ftz.f32 %f5738, %f668, %f5735; max.ftz.f32 %f5739, %f668, %f5734; st.local.v2.f32 [%rd86+20], {%f5739, %f5738}; $L__BB106_646: cvt.u32.u64 %r1323, %rd85; add.s32 %r3147, %r1323, 1; add.s32 %r3149, %r3149, -1; setp.ne.s32 %p543, %r3149, 0; @%p543 bra $L__BB106_629; $L__BB106_647: ld.global.u32 %r331, [%rd29]; and.b32 %r1324, %r3134, 32; setp.eq.s32 %p544, %r1324, 0; and.b32 %r332, %r331, 16384; @%p544 bra $L__BB106_649; bra.uni $L__BB106_648; $L__BB106_649: ld.local.u32 %r1325, [%rd7+-32]; and.b32 %r1326, %r1325, 1; setp.eq.b32 %p546, %r1326, 1; not.pred %p547, %p546; setp.eq.s32 %p548, %r332, 0; or.pred %p549, %p548, %p547; @%p549 bra $L__BB106_690; bra.uni $L__BB106_650; $L__BB106_648: setp.eq.s32 %p545, %r332, 0; @%p545 bra $L__BB106_690; $L__BB106_650: ld.local.u32 %r1327, [%rd7+-32]; and.b32 %r1328, %r1327, 1; setp.eq.b32 %p550, %r1328, 1; mov.pred %p551, 0; xor.pred %p552, %p550, %p551; not.pred %p553, %p552; @%p553 bra $L__BB106_669; bra.uni $L__BB106_651; $L__BB106_669: ld.local.u32 %r346, [%rd7+172]; setp.lt.s32 %p565, %r346, 1; mov.f32 %f15757, 0f00000000; mov.f32 %f15758, %f15757; mov.f32 %f15759, %f15757; @%p565 bra $L__BB106_685; add.s32 %r1344, %r346, -1; and.b32 %r3159, %r346, 3; setp.lt.u32 %p566, %r1344, 3; mov.f32 %f15759, 0f00000000; mov.u32 %r3157, 0; mov.f32 %f15758, %f15759; mov.f32 %f15757, %f15759; @%p566 bra $L__BB106_681; sub.s32 %r3156, %r346, %r3159; mov.f32 %f15759, 0f00000000; mov.u32 %r3157, 0; mov.f32 %f15758, %f15759; mov.f32 %f15757, %f15759; $L__BB106_672: cvt.s64.s32 %rd91, %r3157; mul.wide.s32 %rd1095, %r3157, 80; add.s64 %rd1096, %rd1, %rd1095; add.s64 %rd92, %rd1096, 316; ld.local.u32 %r1346, [%rd1096+316]; setp.ne.s32 %p567, %r1346, 35; @%p567 bra $L__BB106_674; ld.local.v4.f32 {%f5774, %f5775, %f5776, %f5777}, [%rd92+-12]; add.ftz.f32 %f15757, %f15757, %f5774; add.ftz.f32 %f15758, %f15758, %f5775; add.ftz.f32 %f15759, %f15759, %f5776; $L__BB106_674: ld.local.u32 %r1347, [%rd92+80]; setp.ne.s32 %p568, %r1347, 35; @%p568 bra $L__BB106_676; ld.local.v4.f32 {%f5781, %f5782, %f5783, %f5784}, [%rd92+68]; add.ftz.f32 %f15757, %f15757, %f5781; add.ftz.f32 %f15758, %f15758, %f5782; add.ftz.f32 %f15759, %f15759, %f5783; $L__BB106_676: ld.local.u32 %r1348, [%rd92+160]; setp.ne.s32 %p569, %r1348, 35; @%p569 bra $L__BB106_678; ld.local.v4.f32 {%f5788, %f5789, %f5790, %f5791}, [%rd92+148]; add.ftz.f32 %f15757, %f15757, %f5788; add.ftz.f32 %f15758, %f15758, %f5789; add.ftz.f32 %f15759, %f15759, %f5790; $L__BB106_678: ld.local.u32 %r1349, [%rd92+240]; setp.ne.s32 %p570, %r1349, 35; @%p570 bra $L__BB106_680; ld.local.v4.f32 {%f5795, %f5796, %f5797, %f5798}, [%rd92+228]; add.ftz.f32 %f15757, %f15757, %f5795; add.ftz.f32 %f15758, %f15758, %f5796; add.ftz.f32 %f15759, %f15759, %f5797; $L__BB106_680: cvt.u32.u64 %r1350, %rd91; add.s32 %r3157, %r1350, 4; add.s32 %r3156, %r3156, -4; setp.ne.s32 %p571, %r3156, 0; @%p571 bra $L__BB106_672; $L__BB106_681: setp.eq.s32 %p572, %r3159, 0; @%p572 bra $L__BB106_685; $L__BB106_682: .pragma "nounroll"; cvt.s64.s32 %rd93, %r3157; mul.wide.s32 %rd1097, %r3157, 80; add.s64 %rd1098, %rd1, %rd1097; add.s64 %rd94, %rd1098, 316; ld.local.u32 %r1351, [%rd1098+316]; setp.ne.s32 %p573, %r1351, 35; @%p573 bra $L__BB106_684; ld.local.f32 %f5802, [%rd94+-12]; add.ftz.f32 %f15757, %f15757, %f5802; ld.local.f32 %f5803, [%rd94+-8]; add.ftz.f32 %f15758, %f15758, %f5803; ld.local.f32 %f5804, [%rd94+-4]; add.ftz.f32 %f15759, %f15759, %f5804; $L__BB106_684: cvt.u32.u64 %r1352, %rd93; add.s32 %r3157, %r1352, 1; add.s32 %r3159, %r3159, -1; setp.ne.s32 %p574, %r3159, 0; @%p574 bra $L__BB106_682; bra.uni $L__BB106_685; $L__BB106_651: and.b32 %r1329, %r3134, 524800; setp.ne.s32 %p554, %r1329, 512; mov.f32 %f15757, 0f3F800000; mov.f32 %f15758, %f15757; mov.f32 %f15759, %f15757; @%p554 bra $L__BB106_685; ld.local.f32 %f669, [%rd7+204]; ld.local.v2.f32 {%f5761, %f5762}, [%rd7+208]; ld.local.u32 %r333, [%rd7+172]; setp.lt.s32 %p555, %r333, 1; @%p555 bra $L__BB106_668; add.s32 %r1331, %r333, -1; and.b32 %r3154, %r333, 3; setp.lt.u32 %p556, %r1331, 3; mov.u32 %r3152, 0; @%p556 bra $L__BB106_664; sub.s32 %r3151, %r333, %r3154; add.s64 %rd2937, %rd1, 316; mov.u32 %r3152, 0; $L__BB106_655: ld.local.u32 %r1333, [%rd2937]; setp.eq.s32 %p557, %r1333, 31; @%p557 bra $L__BB106_657; mov.u32 %r1334, 40; st.local.u32 [%rd2937], %r1334; $L__BB106_657: ld.local.u32 %r1335, [%rd2937+80]; setp.eq.s32 %p558, %r1335, 31; @%p558 bra $L__BB106_659; mov.u32 %r1336, 40; st.local.u32 [%rd2937+80], %r1336; $L__BB106_659: ld.local.u32 %r1337, [%rd2937+160]; setp.eq.s32 %p559, %r1337, 31; @%p559 bra $L__BB106_661; mov.u32 %r1338, 40; st.local.u32 [%rd2937+160], %r1338; $L__BB106_661: ld.local.u32 %r1339, [%rd2937+240]; setp.eq.s32 %p560, %r1339, 31; @%p560 bra $L__BB106_663; mov.u32 %r1340, 40; st.local.u32 [%rd2937+240], %r1340; $L__BB106_663: add.s32 %r3152, %r3152, 4; add.s32 %r3151, %r3151, -4; setp.ne.s32 %p561, %r3151, 0; add.s64 %rd2937, %rd2937, 320; @%p561 bra $L__BB106_655; $L__BB106_664: setp.eq.s32 %p562, %r3154, 0; @%p562 bra $L__BB106_668; $L__BB106_665: .pragma "nounroll"; mul.wide.s32 %rd1093, %r3152, 80; add.s64 %rd1094, %rd1, %rd1093; add.s64 %rd90, %rd1094, 316; ld.local.u32 %r1341, [%rd1094+316]; setp.eq.s32 %p563, %r1341, 31; @%p563 bra $L__BB106_667; mov.u32 %r1342, 40; st.local.u32 [%rd90], %r1342; $L__BB106_667: add.s32 %r3152, %r3152, 1; add.s32 %r3154, %r3154, -1; setp.ne.s32 %p564, %r3154, 0; @%p564 bra $L__BB106_665; $L__BB106_668: and.b32 %r3134, %r3134, -763; st.local.u32 [%rd7+-36], %r3134; mov.f32 %f5763, 0f3F800000; sub.ftz.f32 %f15759, %f5763, %f5762; sub.ftz.f32 %f15758, %f5763, %f5761; sub.ftz.f32 %f15757, %f5763, %f669; $L__BB106_685: ld.const.u64 %rd1099, [kernel_params+2584]; cvta.to.global.u64 %rd1100, %rd1099; mul.lo.s64 %rd1101, %rd2, 12; add.s64 %rd1102, %rd1100, %rd1101; ld.global.f32 %f5805, [%rd1102]; ld.global.f32 %f5806, [%rd1102+4]; mul.ftz.f32 %f5807, %f15758, %f5806; ld.global.f32 %f5808, [%rd1102+8]; fma.rn.ftz.f32 %f5809, %f15757, %f5805, %f5807; fma.rn.ftz.f32 %f5810, %f15759, %f5808, %f5809; mul.ftz.f32 %f720, %f5810, 0f3EAAAAAB; ld.const.u64 %rd1103, [kernel_params+2448]; cvta.to.global.u64 %rd1104, %rd1103; add.s64 %rd1106, %rd1104, %rd837; ld.global.u32 %rd1107, [%rd1106]; ld.const.s32 %rd1108, [kernel_params+1264]; mul.lo.s64 %rd95, %rd1108, %rd1107; ld.const.u8 %rs68, [kernel_params+1260]; and.b16 %rs69, %rs68, 2; setp.eq.s16 %p575, %rs69, 0; @%p575 bra $L__BB106_687; ld.const.s32 %rd1109, [kernel_params+1268]; cvta.to.global.u64 %rd1110, %rd618; add.s64 %rd1111, %rd1109, %rd95; shl.b64 %rd1112, %rd1111, 2; add.s64 %rd1113, %rd1110, %rd1112; add.s64 %rd1114, %rd1113, 12; atom.global.add.f32 %f5811, [%rd1114], %f720; $L__BB106_687: ld.const.u32 %r1353, [kernel_params+1620]; setp.eq.s32 %p576, %r1353, 0; shr.u32 %r1354, %r331, 29; and.b32 %r1355, %r1354, 1; setp.eq.b32 %p577, %r1355, 1; or.pred %p578, %p577, %p576; @%p578 bra $L__BB106_689; ld.const.s32 %rd1115, [kernel_params+1384]; cvta.to.global.u64 %rd1116, %rd618; add.s64 %rd1117, %rd1115, %rd95; shl.b64 %rd1118, %rd1117, 2; add.s64 %rd1119, %rd1116, %rd1118; add.s64 %rd1120, %rd1119, 12; atom.global.add.f32 %f5812, [%rd1120], %f720; $L__BB106_689: setp.eq.ftz.f32 %p579, %f15757, 0f3F800000; setp.eq.ftz.f32 %p580, %f15758, 0f3F800000; and.pred %p581, %p579, %p580; setp.eq.ftz.f32 %p582, %f15759, 0f3F800000; and.pred %p583, %p581, %p582; mov.u16 %rs161, 0; @%p583 bra $L__BB106_2421; $L__BB106_690: and.b32 %r1356, %r3134, 2; setp.eq.s32 %p584, %r1356, 0; @%p584 bra $L__BB106_794; ld.global.u32 %r360, [%rd29]; ld.local.v2.f32 {%f5814, %f5815}, [%rd7+-64]; ld.local.v4.f32 {%f5816, %f5817, %f5818, %f5819}, [%rd7+-56]; mul.ftz.f32 %f5820, %f5815, %f5818; fma.rn.ftz.f32 %f5821, %f5814, %f5817, %f5820; fma.rn.ftz.f32 %f727, %f5816, %f5819, %f5821; ld.local.v2.f32 {%f5822, %f5823}, [%rd7+192]; ld.local.f32 %f730, [%rd7+200]; shr.u32 %r1357, %r360, 11; and.b32 %r1358, %r1357, 1; setp.eq.b32 %p585, %r1358, 1; and.b32 %r1359, %r3134, 65536; setp.eq.s32 %p586, %r1359, 0; or.pred %p587, %p585, %p586; mov.f32 %f15880, 0f3F800000; @%p587 bra $L__BB106_750; ld.local.u32 %r1360, [%rd7+-24]; and.b32 %r1361, %r1360, 1; setp.eq.b32 %p588, %r1361, 1; mov.pred %p589, 0; xor.pred %p590, %p588, %p589; not.pred %p591, %p590; @%p591 bra $L__BB106_750; ld.const.u64 %rd2916, [kernel_params+2184]; cvta.to.global.u64 %rd2915, %rd2916; ld.const.u64 %rd1121, [kernel_params+2560]; cvta.to.global.u64 %rd1122, %rd1121; add.s64 %rd1124, %rd1122, %rd837; ld.global.f32 %f731, [%rd1124]; ld.local.f32 %f732, [%rd7]; ld.local.s32 %rd96, [%rd7+-28]; ld.local.u32 %r1362, [%rd7+-12]; cvt.s64.s32 %rd97, %r1362; mul.wide.s32 %rd1125, %r1362, 4; add.s64 %rd1126, %rd2915, %rd1125; ld.global.u32 %r361, [%rd1126]; shr.u32 %r1363, %r361, 6; and.b32 %r1364, %r1363, 1; setp.eq.b32 %p592, %r1364, 1; ld.local.f32 %f733, [%rd7+-4]; setp.ge.ftz.f32 %p593, %f733, 0f00000000; and.pred %p594, %p593, %p592; @%p594 bra $L__BB106_695; bra.uni $L__BB106_694; $L__BB106_695: ld.const.u64 %rd1138, [kernel_params+2160]; cvta.to.global.u64 %rd1139, %rd1138; mul.lo.s64 %rd1140, %rd97, 208; add.s64 %rd1141, %rd1139, %rd1140; ld.global.u32 %r362, [%rd1141+156]; ld.global.u32 %r363, [%rd1141+152]; shl.b32 %r1368, %r363, 1; cvt.rn.f32.s32 %f5825, %r1368; mul.ftz.f32 %f743, %f733, %f5825; cvt.rzi.ftz.s32.f32 %r1369, %f743; add.s32 %r1370, %r1368, -1; min.s32 %r364, %r1369, %r1370; ld.const.u64 %rd1142, [kernel_params+2312]; cvta.to.global.u64 %rd99, %rd1142; ld.global.u32 %r3163, [%rd1141+164]; mul.wide.u32 %rd1143, %r3163, 16; add.s64 %rd1144, %rd99, %rd1143; ld.global.u16 %rs137, [%rd1144+12]; ld.global.u32 %r3162, [%rd1144+8]; ld.global.u64 %rd2938, [%rd1144]; setp.eq.s64 %p595, %rd2938, 11; @%p595 bra $L__BB106_700; $L__BB106_696: setp.eq.s64 %p596, %rd2938, 0; @%p596 bra $L__BB106_698; add.s32 %r3163, %r3163, 2; bra.uni $L__BB106_699; $L__BB106_698: setp.eq.s16 %p597, %rs137, 0; mov.u32 %r3166, -1; mov.u32 %r3163, %r3162; @%p597 bra $L__BB106_701; $L__BB106_699: mul.wide.u32 %rd1145, %r3163, 16; add.s64 %rd1146, %rd99, %rd1145; ld.global.u16 %rs137, [%rd1146+12]; ld.global.u32 %r3162, [%rd1146+8]; ld.global.u64 %rd2938, [%rd1146]; setp.ne.s64 %p598, %rd2938, 11; @%p598 bra $L__BB106_696; $L__BB106_700: setp.eq.s16 %p599, %rs137, 0; selp.b32 %r3166, -1, %r3162, %p599; $L__BB106_701: ld.const.u64 %rd2957, [kernel_params+2232]; cvta.to.global.u64 %rd1147, %rd2957; shl.b64 %rd1148, %rd96, 4; add.s64 %rd1149, %rd1147, %rd1148; ld.global.v4.u32 {%r1372, %r1373, %r1374, %r1375}, [%rd1149]; setp.eq.s32 %p600, %r364, %r363; @%p600 bra $L__BB106_703; bra.uni $L__BB106_702; $L__BB106_703: ld.const.u64 %rd1155, [kernel_params+2256]; cvta.to.global.u64 %rd1156, %rd1155; mul.wide.u32 %rd1157, %r1375, 12; add.s64 %rd2939, %rd1156, %rd1157; add.s32 %r1382, %r1375, 1; mul.wide.u32 %rd1158, %r1382, 12; add.s64 %rd2942, %rd1156, %rd1158; add.s32 %r1383, %r1375, 2; mul.wide.u32 %rd1159, %r1383, 12; add.s64 %rd2945, %rd1156, %rd1159; add.s64 %rd2940, %rd2939, 4; add.s64 %rd2941, %rd2939, 8; add.s64 %rd2943, %rd2942, 4; add.s64 %rd2944, %rd2942, 8; add.s64 %rd2946, %rd2945, 4; add.s64 %rd2947, %rd2945, 8; bra.uni $L__BB106_704; $L__BB106_694: ld.const.u64 %rd2957, [kernel_params+2232]; cvta.to.global.u64 %rd1127, %rd2957; shl.b64 %rd1128, %rd96, 4; add.s64 %rd1129, %rd1127, %rd1128; ld.const.u64 %rd1130, [kernel_params+2256]; cvta.to.global.u64 %rd1131, %rd1130; ld.global.u32 %r1365, [%rd1129+12]; mul.wide.u32 %rd1132, %r1365, 12; add.s64 %rd1133, %rd1131, %rd1132; ld.global.f32 %f15809, [%rd1133]; ld.global.f32 %f15810, [%rd1133+4]; ld.global.f32 %f15811, [%rd1133+8]; add.s32 %r1366, %r1365, 1; mul.wide.u32 %rd1134, %r1366, 12; add.s64 %rd1135, %rd1131, %rd1134; ld.global.f32 %f15812, [%rd1135]; ld.global.f32 %f15813, [%rd1135+4]; ld.global.f32 %f15814, [%rd1135+8]; add.s32 %r1367, %r1365, 2; mul.wide.u32 %rd1136, %r1367, 12; add.s64 %rd1137, %rd1131, %rd1136; ld.global.f32 %f15815, [%rd1137]; ld.global.f32 %f15816, [%rd1137+4]; ld.global.f32 %f15817, [%rd1137+8]; mov.u16 %rs140, 0; bra.uni $L__BB106_708; $L__BB106_702: setp.gt.s32 %p601, %r364, %r363; selp.b32 %r1376, -1, 0, %p601; add.s32 %r1377, %r364, %r1376; mad.lo.s32 %r1378, %r1377, %r362, %r3166; ld.const.u64 %rd1150, [kernel_params+2336]; cvta.to.global.u64 %rd1151, %rd1150; add.s32 %r1379, %r1372, %r1378; mul.wide.u32 %rd1152, %r1379, 12; add.s64 %rd2939, %rd1151, %rd1152; add.s32 %r1380, %r1373, %r1378; mul.wide.u32 %rd1153, %r1380, 12; add.s64 %rd2942, %rd1151, %rd1153; add.s32 %r1381, %r1374, %r1378; mul.wide.u32 %rd1154, %r1381, 12; add.s64 %rd2945, %rd1151, %rd1154; add.s64 %rd2940, %rd2939, 4; add.s64 %rd2941, %rd2939, 8; add.s64 %rd2943, %rd2942, 4; add.s64 %rd2944, %rd2942, 8; add.s64 %rd2946, %rd2945, 4; add.s64 %rd2947, %rd2945, 8; $L__BB106_704: ld.global.f32 %f744, [%rd2947]; ld.global.f32 %f745, [%rd2946]; ld.global.f32 %f746, [%rd2945]; ld.global.f32 %f747, [%rd2944]; ld.global.f32 %f748, [%rd2943]; ld.global.f32 %f749, [%rd2942]; ld.global.f32 %f750, [%rd2941]; ld.global.f32 %f751, [%rd2940]; ld.global.f32 %f752, [%rd2939]; add.s32 %r379, %r364, 1; setp.eq.s32 %p602, %r379, %r363; @%p602 bra $L__BB106_706; bra.uni $L__BB106_705; $L__BB106_706: ld.const.u64 %rd1165, [kernel_params+2256]; cvta.to.global.u64 %rd1166, %rd1165; mul.wide.u32 %rd1167, %r1375, 12; add.s64 %rd2948, %rd1166, %rd1167; add.s32 %r1390, %r1375, 1; mul.wide.u32 %rd1168, %r1390, 12; add.s64 %rd2951, %rd1166, %rd1168; add.s32 %r1391, %r1375, 2; mul.wide.u32 %rd1169, %r1391, 12; add.s64 %rd2954, %rd1166, %rd1169; add.s64 %rd2949, %rd2948, 4; add.s64 %rd2950, %rd2948, 8; add.s64 %rd2952, %rd2951, 4; add.s64 %rd2953, %rd2951, 8; add.s64 %rd2955, %rd2954, 4; add.s64 %rd2956, %rd2954, 8; bra.uni $L__BB106_707; $L__BB106_705: setp.ge.s32 %p603, %r364, %r363; selp.b32 %r1384, -1, 0, %p603; add.s32 %r1385, %r379, %r1384; mad.lo.s32 %r1386, %r1385, %r362, %r3166; ld.const.u64 %rd1160, [kernel_params+2336]; cvta.to.global.u64 %rd1161, %rd1160; add.s32 %r1387, %r1372, %r1386; mul.wide.u32 %rd1162, %r1387, 12; add.s64 %rd2948, %rd1161, %rd1162; add.s32 %r1388, %r1373, %r1386; mul.wide.u32 %rd1163, %r1388, 12; add.s64 %rd2951, %rd1161, %rd1163; add.s32 %r1389, %r1374, %r1386; mul.wide.u32 %rd1164, %r1389, 12; add.s64 %rd2954, %rd1161, %rd1164; add.s64 %rd2949, %rd2948, 4; add.s64 %rd2950, %rd2948, 8; add.s64 %rd2952, %rd2951, 4; add.s64 %rd2953, %rd2951, 8; add.s64 %rd2955, %rd2954, 4; add.s64 %rd2956, %rd2954, 8; $L__BB106_707: cvt.rn.f32.s32 %f5826, %r364; sub.ftz.f32 %f5827, %f743, %f5826; mov.f32 %f5828, 0f3F800000; sub.ftz.f32 %f5829, %f5828, %f5827; ld.global.f32 %f5830, [%rd2948]; mul.ftz.f32 %f5831, %f5827, %f5830; ld.global.f32 %f5832, [%rd2949]; mul.ftz.f32 %f5833, %f5827, %f5832; ld.global.f32 %f5834, [%rd2950]; mul.ftz.f32 %f5835, %f5827, %f5834; fma.rn.ftz.f32 %f15809, %f5829, %f752, %f5831; fma.rn.ftz.f32 %f15810, %f5829, %f751, %f5833; fma.rn.ftz.f32 %f15811, %f5829, %f750, %f5835; ld.global.f32 %f5836, [%rd2951]; mul.ftz.f32 %f5837, %f5827, %f5836; ld.global.f32 %f5838, [%rd2952]; mul.ftz.f32 %f5839, %f5827, %f5838; ld.global.f32 %f5840, [%rd2953]; mul.ftz.f32 %f5841, %f5827, %f5840; fma.rn.ftz.f32 %f15812, %f5829, %f749, %f5837; fma.rn.ftz.f32 %f15813, %f5829, %f748, %f5839; fma.rn.ftz.f32 %f15814, %f5829, %f747, %f5841; ld.global.f32 %f5842, [%rd2954]; mul.ftz.f32 %f5843, %f5827, %f5842; ld.global.f32 %f5844, [%rd2955]; mul.ftz.f32 %f5845, %f5827, %f5844; ld.global.f32 %f5846, [%rd2956]; mul.ftz.f32 %f5847, %f5827, %f5846; fma.rn.ftz.f32 %f15815, %f5829, %f746, %f5843; fma.rn.ftz.f32 %f15816, %f5829, %f745, %f5845; fma.rn.ftz.f32 %f15817, %f5829, %f744, %f5847; mov.u16 %rs140, 1; $L__BB106_708: and.b32 %r380, %r361, 4; setp.ne.s32 %p604, %r380, 0; @%p604 bra $L__BB106_716; and.b32 %r1392, %r361, 2; setp.eq.s32 %p605, %r1392, 0; ld.const.u64 %rd1170, [kernel_params+2160]; cvta.to.global.u64 %rd1171, %rd1170; mul.lo.s64 %rd1172, %rd97, 208; add.s64 %rd159, %rd1171, %rd1172; @%p605 bra $L__BB106_714; bra.uni $L__BB106_710; $L__BB106_714: ld.global.v4.f32 {%f15797, %f15798, %f15799, %f15800}, [%rd159]; ld.global.v4.f32 {%f15801, %f15802, %f15803, %f15804}, [%rd159+16]; ld.global.v4.f32 {%f15805, %f15806, %f15807, %f15808}, [%rd159+32]; bra.uni $L__BB106_715; $L__BB106_710: ld.const.u64 %rd1173, [kernel_params+2176]; cvta.to.global.u64 %rd1174, %rd1173; ld.global.u32 %rd1175, [%rd159+168]; ld.global.u32 %r1393, [%rd159+152]; shl.b32 %r1394, %r1393, 1; cvt.rn.f32.s32 %f5848, %r1394; selp.f32 %f5849, %f733, 0f3F000000, %p593; mul.ftz.f32 %f5850, %f5849, %f5848; cvt.rzi.ftz.s32.f32 %r1395, %f5850; add.s32 %r1396, %r1394, -1; min.s32 %r1397, %r1395, %r1396; cvt.rn.f32.s32 %f5851, %r1397; sub.ftz.f32 %f771, %f5850, %f5851; cvt.s64.s32 %rd1176, %r1397; add.s64 %rd1177, %rd1176, %rd1175; shl.b64 %rd1178, %rd1177, 6; add.s64 %rd160, %rd1174, %rd1178; ld.global.v4.f32 {%f5852, %f5853, %f5854, %f5855}, [%rd160]; ld.global.v4.f32 {%f5856, %f5857, %f5858, %f5859}, [%rd160+64]; mul.ftz.f32 %f5860, %f5853, %f5857; fma.rn.ftz.f32 %f5861, %f5852, %f5856, %f5860; mul.ftz.f32 %f5862, %f5855, %f5859; fma.rn.ftz.f32 %f5863, %f5854, %f5858, %f5862; add.ftz.f32 %f780, %f5861, %f5863; setp.gt.ftz.f32 %p607, %f780, 0f3F7FDF3B; mov.f32 %f5864, 0f3F800000; sub.ftz.f32 %f781, %f5864, %f771; @%p607 bra $L__BB106_712; bra.uni $L__BB106_711; $L__BB106_712: mul.ftz.f32 %f5918, %f5852, %f781; fma.rn.ftz.f32 %f5919, %f5856, %f771, %f5918; mul.ftz.f32 %f5920, %f5853, %f781; fma.rn.ftz.f32 %f5921, %f5857, %f771, %f5920; mul.ftz.f32 %f5922, %f5854, %f781; fma.rn.ftz.f32 %f5923, %f5858, %f771, %f5922; mul.ftz.f32 %f5924, %f5855, %f781; fma.rn.ftz.f32 %f5925, %f5859, %f771, %f5924; mul.ftz.f32 %f5926, %f5921, %f5921; fma.rn.ftz.f32 %f5927, %f5919, %f5919, %f5926; mul.ftz.f32 %f5928, %f5925, %f5925; fma.rn.ftz.f32 %f5929, %f5923, %f5923, %f5928; add.ftz.f32 %f5930, %f5927, %f5929; rsqrt.approx.ftz.f32 %f5931, %f5930; mul.ftz.f32 %f15796, %f5925, %f5931; mul.ftz.f32 %f15795, %f5923, %f5931; mul.ftz.f32 %f15794, %f5921, %f5931; mul.ftz.f32 %f15793, %f5919, %f5931; bra.uni $L__BB106_713; $L__BB106_1332: mov.u32 %r3226, 0; st.local.u32 [%rd289], %r3226; $L__BB106_1333: add.s64 %rd1739, %rd1729, 76; cvta.to.local.u64 %rd1740, %rd1739; st.local.u32 [%rd1740], %r622; add.s64 %rd1741, %rd1729, 60; cvta.to.local.u64 %rd1742, %rd1741; st.local.u32 [%rd1742], %r623; add.s64 %rd1743, %rd1729, 68; cvta.to.local.u64 %rd1744, %rd1743; st.local.f32 [%rd1744], %f1958; add.s64 %rd1745, %rd1729, 72; cvta.to.local.u64 %rd1746, %rd1745; st.local.f32 [%rd1746], %f1959; mov.b32 %r2021, %f1939; mov.u32 %r2022, -1; add.s64 %rd290, %rd1729, 80; cvta.to.local.u64 %rd1747, %rd290; st.local.v2.u32 [%rd1747], {%r2022, %r2021}; add.s64 %rd1748, %rd1729, 88; cvta.to.local.u64 %rd1749, %rd1748; st.local.f32 [%rd1749], %f1960; add.s64 %rd1750, %rd1729, 52; cvta.to.local.u64 %rd291, %rd1750; st.local.u32 [%rd291], %r620; add.s64 %rd1751, %rd1729, 56; cvta.to.local.u64 %rd292, %rd1751; mov.u32 %r2023, 0; st.local.u32 [%rd292], %r2023; setp.eq.s32 %p1254, %r622, -1; @%p1254 bra $L__BB106_1358; ld.const.u64 %rd2914, [kernel_params+2184]; cvta.to.global.u64 %rd2913, %rd2914; shl.b64 %rd1752, %rd285, 2; add.s64 %rd1753, %rd2913, %rd1752; ld.global.u32 %r626, [%rd1753]; st.local.u32 [%rd292], %r626; and.b32 %r627, %r626, 2; setp.eq.s32 %p1255, %r627, 0; @%p1255 bra $L__BB106_1341; ld.const.u64 %rd1754, [kernel_params+2160]; cvta.to.global.u64 %rd1755, %rd1754; mul.lo.s64 %rd1756, %rd285, 208; add.s64 %rd1757, %rd1755, %rd1756; ld.const.u64 %rd1758, [kernel_params+2176]; cvta.to.global.u64 %rd1759, %rd1758; ld.global.u32 %rd1760, [%rd1757+168]; ld.global.u32 %r2024, [%rd1757+152]; shl.b32 %r2025, %r2024, 1; cvt.rn.f32.s32 %f9710, %r2025; mul.ftz.f32 %f9711, %f1939, %f9710; cvt.rzi.ftz.s32.f32 %r2026, %f9711; add.s32 %r2027, %r2025, -1; min.s32 %r2028, %r2026, %r2027; cvt.rn.f32.s32 %f9712, %r2028; sub.ftz.f32 %f1969, %f9711, %f9712; cvt.s64.s32 %rd1761, %r2028; add.s64 %rd1762, %rd1761, %rd1760; shl.b64 %rd1763, %rd1762, 6; add.s64 %rd293, %rd1759, %rd1763; ld.global.v4.f32 {%f9713, %f9714, %f9715, %f9716}, [%rd293]; ld.global.v4.f32 {%f9717, %f9718, %f9719, %f9720}, [%rd293+64]; mul.ftz.f32 %f9721, %f9714, %f9718; fma.rn.ftz.f32 %f9722, %f9713, %f9717, %f9721; mul.ftz.f32 %f9723, %f9716, %f9720; fma.rn.ftz.f32 %f9724, %f9715, %f9719, %f9723; add.ftz.f32 %f1978, %f9722, %f9724; setp.gt.ftz.f32 %p1256, %f1978, 0f3F7FDF3B; mov.f32 %f9725, 0f3F800000; sub.ftz.f32 %f1979, %f9725, %f1969; @%p1256 bra $L__BB106_1337; bra.uni $L__BB106_1336; $L__BB106_1337: mul.ftz.f32 %f9779, %f9713, %f1979; fma.rn.ftz.f32 %f9780, %f9717, %f1969, %f9779; mul.ftz.f32 %f9781, %f9714, %f1979; fma.rn.ftz.f32 %f9782, %f9718, %f1969, %f9781; mul.ftz.f32 %f9783, %f9715, %f1979; fma.rn.ftz.f32 %f9784, %f9719, %f1969, %f9783; mul.ftz.f32 %f9785, %f9716, %f1979; fma.rn.ftz.f32 %f9786, %f9720, %f1969, %f9785; mul.ftz.f32 %f9787, %f9782, %f9782; fma.rn.ftz.f32 %f9788, %f9780, %f9780, %f9787; mul.ftz.f32 %f9789, %f9786, %f9786; fma.rn.ftz.f32 %f9790, %f9784, %f9784, %f9789; add.ftz.f32 %f9791, %f9788, %f9790; rsqrt.approx.ftz.f32 %f9792, %f9791; mul.ftz.f32 %f16185, %f9786, %f9792; mul.ftz.f32 %f16184, %f9784, %f9792; mul.ftz.f32 %f16183, %f9782, %f9792; mul.ftz.f32 %f16182, %f9780, %f9792; bra.uni $L__BB106_1338; $L__BB106_1358: setp.eq.s32 %p1274, %r624, -1; @%p1274 bra $L__BB106_1360; st.local.u32 [%rd1747], %r624; $L__BB106_1360: add.s64 %rd1856, %rd1729, 116; cvta.to.local.u64 %rd1857, %rd1856; mov.u32 %r2042, 0; st.local.u32 [%rd1857], %r2042; add.s64 %rd1858, %rd1729, 120; cvta.to.local.u64 %rd1859, %rd1858; st.local.u32 [%rd1859], %r2042; add.s64 %rd1860, %rd1729, 124; cvta.to.local.u64 %rd1861, %rd1860; st.local.u32 [%rd1861], %r2042; add.s64 %rd1862, %rd1729, 136; cvta.to.local.u64 %rd1863, %rd1862; add.s64 %rd1864, %rd1729, 128; cvta.to.local.u64 %rd1865, %rd1864; mov.f32 %f16202, 0f00000000; st.local.v2.f32 [%rd1865], {%f16202, %f16202}; st.local.u32 [%rd1863], %r2042; mov.f32 %f16203, %f16202; mov.f32 %f16204, %f16202; mov.f32 %f16205, %f16202; mov.f32 %f16206, %f16202; mov.f32 %f16207, %f16202; mov.f32 %f16208, %f1957; mov.f32 %f16209, %f16213; mov.f32 %f16210, %f16212; bra.uni $L__BB106_1361; $L__BB106_1336: mov.f32 %f9726, 0fBF800000; max.ftz.f32 %f9727, %f1978, %f9726; min.ftz.f32 %f9729, %f9727, %f9725; abs.ftz.f32 %f9730, %f9729; sub.ftz.f32 %f9731, %f9725, %f9730; mul.ftz.f32 %f9732, %f9731, 0f3F000000; sqrt.approx.ftz.f32 %f9733, %f9732; setp.gt.ftz.f32 %p1257, %f9730, 0f3F11EB85; selp.f32 %f9734, %f9733, %f9730, %p1257; mul.ftz.f32 %f9735, %f9734, %f9734; mov.f32 %f9736, 0f3C94D2E9; mov.f32 %f9737, 0f3D53F941; fma.rn.ftz.f32 %f9738, %f9737, %f9735, %f9736; mov.f32 %f9739, 0f3D3F841F; fma.rn.ftz.f32 %f9740, %f9738, %f9735, %f9739; mov.f32 %f9741, 0f3D994929; fma.rn.ftz.f32 %f9742, %f9740, %f9735, %f9741; mov.f32 %f9743, 0f3E2AAB94; fma.rn.ftz.f32 %f9744, %f9742, %f9735, %f9743; mul.ftz.f32 %f9745, %f9735, %f9744; fma.rn.ftz.f32 %f9746, %f9745, %f9734, %f9734; add.ftz.f32 %f9747, %f9746, %f9746; mov.f32 %f9748, 0f3FC90FDB; sub.ftz.f32 %f9749, %f9748, %f9746; selp.f32 %f9750, %f9747, %f9749, %p1257; setp.lt.ftz.f32 %p1258, %f9729, 0f00000000; mov.f32 %f9751, 0f40490FDB; sub.ftz.f32 %f9752, %f9751, %f9750; selp.f32 %f9753, %f9752, %f9750, %p1258; mul.ftz.f32 %f9754, %f9713, %f1978; sub.ftz.f32 %f9755, %f9717, %f9754; mul.ftz.f32 %f9756, %f9714, %f1978; sub.ftz.f32 %f9757, %f9718, %f9756; mul.ftz.f32 %f9758, %f9715, %f1978; sub.ftz.f32 %f9759, %f9719, %f9758; mul.ftz.f32 %f9760, %f9716, %f1978; sub.ftz.f32 %f9761, %f9720, %f9760; mul.ftz.f32 %f9762, %f9757, %f9757; fma.rn.ftz.f32 %f9763, %f9755, %f9755, %f9762; mul.ftz.f32 %f9764, %f9761, %f9761; fma.rn.ftz.f32 %f9765, %f9759, %f9759, %f9764; add.ftz.f32 %f9766, %f9763, %f9765; rsqrt.approx.ftz.f32 %f9767, %f9766; mul.ftz.f32 %f9768, %f9755, %f9767; mul.ftz.f32 %f9769, %f9757, %f9767; mul.ftz.f32 %f9770, %f9759, %f9767; mul.ftz.f32 %f9771, %f9761, %f9767; mul.ftz.f32 %f9772, %f1969, %f9753; cos.approx.ftz.f32 %f9773, %f9772; sin.approx.ftz.f32 %f9774, %f9772; mul.ftz.f32 %f9775, %f9768, %f9774; mul.ftz.f32 %f9776, %f9769, %f9774; mul.ftz.f32 %f9777, %f9770, %f9774; mul.ftz.f32 %f9778, %f9771, %f9774; fma.rn.ftz.f32 %f16185, %f9716, %f9773, %f9778; fma.rn.ftz.f32 %f16184, %f9715, %f9773, %f9777; fma.rn.ftz.f32 %f16183, %f9714, %f9773, %f9776; fma.rn.ftz.f32 %f16182, %f9713, %f9773, %f9775; $L__BB106_1338: ld.global.v4.f32 {%f9793, %f9794, %f9795, %f9796}, [%rd293+16]; ld.global.v4.f32 {%f9801, %f9802, %f9803, %f9804}, [%rd293+80]; mul.ftz.f32 %f9809, %f1969, %f9801; mul.ftz.f32 %f9810, %f1969, %f9802; mul.ftz.f32 %f9811, %f1969, %f9803; mul.ftz.f32 %f9812, %f1969, %f9804; fma.rn.ftz.f32 %f1992, %f1979, %f9793, %f9809; fma.rn.ftz.f32 %f1993, %f1979, %f9794, %f9810; fma.rn.ftz.f32 %f1994, %f1979, %f9795, %f9811; fma.rn.ftz.f32 %f9813, %f1979, %f9796, %f9812; ld.global.v4.f32 {%f9814, %f9815, %f9816, %f9817}, [%rd293+32]; ld.global.v4.f32 {%f9822, %f9823, %f9824, %f9825}, [%rd293+96]; mul.ftz.f32 %f9830, %f1969, %f9822; mul.ftz.f32 %f9831, %f1969, %f9823; mul.ftz.f32 %f9832, %f1969, %f9824; mul.ftz.f32 %f9833, %f1969, %f9825; fma.rn.ftz.f32 %f9834, %f1979, %f9814, %f9830; fma.rn.ftz.f32 %f9835, %f1979, %f9815, %f9831; fma.rn.ftz.f32 %f9836, %f1979, %f9816, %f9832; fma.rn.ftz.f32 %f9837, %f1979, %f9817, %f9833; ld.global.v4.f32 {%f9838, %f9839, %f9840, %f9841}, [%rd293+48]; ld.global.v4.f32 {%f9846, %f9847, %f9848, %f9849}, [%rd293+112]; mul.ftz.f32 %f9854, %f1969, %f9846; mul.ftz.f32 %f9855, %f1969, %f9847; mul.ftz.f32 %f9856, %f1969, %f9848; mul.ftz.f32 %f9857, %f1969, %f9849; fma.rn.ftz.f32 %f9858, %f1979, %f9838, %f9854; fma.rn.ftz.f32 %f9859, %f1979, %f9839, %f9855; fma.rn.ftz.f32 %f9860, %f1979, %f9840, %f9856; fma.rn.ftz.f32 %f9861, %f1979, %f9841, %f9857; mul.ftz.f32 %f9862, %f16185, 0f3FB504F3; mul.ftz.f32 %f9863, %f16182, 0f3FB504F3; mul.ftz.f32 %f9864, %f16183, 0f3FB504F3; mul.ftz.f32 %f9865, %f16184, 0f3FB504F3; mul.ftz.f32 %f9866, %f9862, %f9863; mul.ftz.f32 %f9867, %f9862, %f9864; mul.ftz.f32 %f9868, %f9862, %f9865; mul.ftz.f32 %f9869, %f9863, %f9863; mul.ftz.f32 %f9870, %f9863, %f9864; mul.ftz.f32 %f9871, %f9863, %f9865; mul.ftz.f32 %f9872, %f9864, %f9864; mul.ftz.f32 %f9873, %f9864, %f9865; mul.ftz.f32 %f9874, %f9865, %f9865; sub.ftz.f32 %f9876, %f9725, %f9872; sub.ftz.f32 %f9877, %f9876, %f9874; sub.ftz.f32 %f9878, %f9870, %f9868; add.ftz.f32 %f9879, %f9867, %f9871; add.ftz.f32 %f9880, %f9868, %f9870; sub.ftz.f32 %f9881, %f9725, %f9869; sub.ftz.f32 %f9882, %f9881, %f9874; sub.ftz.f32 %f9883, %f9873, %f9866; sub.ftz.f32 %f9884, %f9871, %f9867; add.ftz.f32 %f9885, %f9866, %f9873; sub.ftz.f32 %f9886, %f9881, %f9872; mul.ftz.f32 %f9887, %f9878, %f9836; fma.rn.ftz.f32 %f9888, %f9877, %f9813, %f9887; fma.rn.ftz.f32 %f16191, %f9879, %f9859, %f9888; mul.ftz.f32 %f9889, %f9878, %f9837; fma.rn.ftz.f32 %f9890, %f9877, %f9834, %f9889; fma.rn.ftz.f32 %f1996, %f9879, %f9860, %f9890; mul.ftz.f32 %f9891, %f9878, %f9858; fma.rn.ftz.f32 %f9892, %f9877, %f9835, %f9891; fma.rn.ftz.f32 %f1997, %f9879, %f9861, %f9892; mul.ftz.f32 %f9893, %f9882, %f9836; fma.rn.ftz.f32 %f9894, %f9880, %f9813, %f9893; fma.rn.ftz.f32 %f1998, %f9883, %f9859, %f9894; mul.ftz.f32 %f9895, %f9882, %f9837; fma.rn.ftz.f32 %f9896, %f9880, %f9834, %f9895; fma.rn.ftz.f32 %f16190, %f9883, %f9860, %f9896; mul.ftz.f32 %f9897, %f9882, %f9858; fma.rn.ftz.f32 %f9898, %f9880, %f9835, %f9897; fma.rn.ftz.f32 %f2000, %f9883, %f9861, %f9898; mul.ftz.f32 %f9899, %f9885, %f9836; fma.rn.ftz.f32 %f9900, %f9884, %f9813, %f9899; fma.rn.ftz.f32 %f2001, %f9886, %f9859, %f9900; mul.ftz.f32 %f9901, %f9885, %f9837; fma.rn.ftz.f32 %f9902, %f9884, %f9834, %f9901; fma.rn.ftz.f32 %f2002, %f9886, %f9860, %f9902; mul.ftz.f32 %f9903, %f9885, %f9858; fma.rn.ftz.f32 %f9904, %f9884, %f9835, %f9903; fma.rn.ftz.f32 %f16189, %f9886, %f9861, %f9904; add.s64 %rd1765, %rd1729, 144; cvta.to.local.u64 %rd1766, %rd1765; st.local.v4.f32 [%rd1766], {%f16191, %f1996, %f1997, %f1992}; add.s64 %rd1767, %rd1729, 160; cvta.to.local.u64 %rd1768, %rd1767; st.local.v4.f32 [%rd1768], {%f1998, %f16190, %f2000, %f1993}; add.s64 %rd1769, %rd1729, 176; cvta.to.local.u64 %rd1770, %rd1769; st.local.v4.f32 [%rd1770], {%f2001, %f2002, %f16189, %f1994}; mul.ftz.f32 %f9905, %f16190, %f16189; mul.ftz.f32 %f2004, %f2002, %f2000; sub.ftz.f32 %f16188, %f9905, %f2004; mul.ftz.f32 %f2006, %f2002, %f1997; mul.ftz.f32 %f9906, %f1996, %f16189; sub.ftz.f32 %f16187, %f2006, %f9906; mul.ftz.f32 %f2008, %f1996, %f2000; mul.ftz.f32 %f9907, %f16190, %f1997; sub.ftz.f32 %f16186, %f2008, %f9907; mul.ftz.f32 %f9908, %f1998, %f16187; fma.rn.ftz.f32 %f9909, %f16191, %f16188, %f9908; fma.rn.ftz.f32 %f16192, %f2001, %f16186, %f9909; setp.neu.ftz.f32 %p1259, %f16192, 0f00000000; @%p1259 bra $L__BB106_1340; add.ftz.f32 %f16191, %f16191, 0f322BCC77; add.ftz.f32 %f16189, %f16189, 0f322BCC77; add.ftz.f32 %f16190, %f16190, 0f322BCC77; mul.ftz.f32 %f9910, %f16190, %f16189; sub.ftz.f32 %f16188, %f9910, %f2004; mul.ftz.f32 %f9911, %f1996, %f16189; sub.ftz.f32 %f16187, %f2006, %f9911; mul.ftz.f32 %f9912, %f16190, %f1997; sub.ftz.f32 %f16186, %f2008, %f9912; mul.ftz.f32 %f9913, %f1998, %f16187; fma.rn.ftz.f32 %f9914, %f16191, %f16188, %f9913; fma.rn.ftz.f32 %f9915, %f2001, %f16186, %f9914; setp.eq.ftz.f32 %p1260, %f9915, 0f00000000; selp.f32 %f16192, 0f7F7FFFFF, %f9915, %p1260; $L__BB106_1340: mul.ftz.f32 %f9916, %f1998, %f16189; mul.ftz.f32 %f9917, %f2001, %f2000; sub.ftz.f32 %f9918, %f9917, %f9916; mul.ftz.f32 %f9919, %f2001, %f1997; mul.ftz.f32 %f9920, %f16189, %f16191; sub.ftz.f32 %f9921, %f9920, %f9919; mul.ftz.f32 %f9922, %f2000, %f16191; mul.ftz.f32 %f9923, %f1998, %f1997; sub.ftz.f32 %f9924, %f9923, %f9922; mul.ftz.f32 %f9925, %f2001, %f16190; mul.ftz.f32 %f9926, %f1998, %f2002; sub.ftz.f32 %f9927, %f9926, %f9925; mul.ftz.f32 %f9928, %f2002, %f16191; mul.ftz.f32 %f9929, %f2001, %f1996; sub.ftz.f32 %f9930, %f9929, %f9928; mul.ftz.f32 %f9931, %f1998, %f1996; mul.ftz.f32 %f9932, %f16190, %f16191; sub.ftz.f32 %f9933, %f9932, %f9931; div.approx.ftz.f32 %f9934, %f16187, %f16192; mul.ftz.f32 %f9935, %f1993, %f9934; div.approx.ftz.f32 %f9936, %f16188, %f16192; fma.rn.ftz.f32 %f9937, %f1992, %f9936, %f9935; div.approx.ftz.f32 %f9938, %f16186, %f16192; fma.rn.ftz.f32 %f9939, %f1994, %f9938, %f9937; div.approx.ftz.f32 %f9940, %f9921, %f16192; mul.ftz.f32 %f9941, %f1993, %f9940; div.approx.ftz.f32 %f9942, %f9918, %f16192; fma.rn.ftz.f32 %f9943, %f1992, %f9942, %f9941; div.approx.ftz.f32 %f9944, %f9924, %f16192; fma.rn.ftz.f32 %f9945, %f1994, %f9944, %f9943; div.approx.ftz.f32 %f9946, %f9930, %f16192; mul.ftz.f32 %f9947, %f1993, %f9946; div.approx.ftz.f32 %f9948, %f9927, %f16192; fma.rn.ftz.f32 %f9949, %f1992, %f9948, %f9947; div.approx.ftz.f32 %f9950, %f9933, %f16192; fma.rn.ftz.f32 %f9951, %f1994, %f9950, %f9949; neg.ftz.f32 %f9952, %f9939; add.s64 %rd1772, %rd1729, 192; cvta.to.local.u64 %rd1773, %rd1772; st.local.v4.f32 [%rd1773], {%f9936, %f9934, %f9938, %f9952}; neg.ftz.f32 %f9953, %f9945; add.s64 %rd1774, %rd1729, 208; cvta.to.local.u64 %rd1775, %rd1774; st.local.v4.f32 [%rd1775], {%f9942, %f9940, %f9944, %f9953}; neg.ftz.f32 %f9954, %f9951; add.s64 %rd1776, %rd1729, 224; cvta.to.local.u64 %rd1777, %rd1776; st.local.v4.f32 [%rd1777], {%f9948, %f9946, %f9950, %f9954}; $L__BB106_1341: setp.eq.s32 %p1261, %r3226, 1; @%p1261 bra $L__BB106_1343; bra.uni $L__BB106_1342; $L__BB106_1343: setp.lt.s32 %p1262, %r618, 0; @%p1262 bra $L__BB106_1345; ld.const.u64 %rd2984, [kernel_params+2232]; and.b32 %r3227, %r626, 4; mov.f32 %f16208, %f1957; mov.f32 %f16209, %f16213; mov.f32 %f16210, %f16212; bra.uni $L__BB106_1351; $L__BB106_1342: add.s64 %rd1779, %rd1729, 116; cvta.to.local.u64 %rd1780, %rd1779; st.local.u32 [%rd1780], %r2023; add.s64 %rd1781, %rd1729, 120; cvta.to.local.u64 %rd1782, %rd1781; st.local.u32 [%rd1782], %r2023; add.s64 %rd1783, %rd1729, 124; cvta.to.local.u64 %rd1784, %rd1783; st.local.u32 [%rd1784], %r2023; add.s64 %rd1785, %rd1729, 136; cvta.to.local.u64 %rd1786, %rd1785; add.s64 %rd1787, %rd1729, 128; cvta.to.local.u64 %rd1788, %rd1787; mov.f32 %f16202, 0f00000000; st.local.v2.f32 [%rd1788], {%f16202, %f16202}; st.local.u32 [%rd1786], %r2023; mov.f32 %f16203, %f16202; mov.f32 %f16204, %f16202; mov.f32 %f16205, %f16202; mov.f32 %f16206, %f16202; mov.f32 %f16207, %f16202; mov.f32 %f16208, %f1957; mov.f32 %f16209, %f16213; mov.f32 %f16210, %f16212; bra.uni $L__BB106_1361; $L__BB106_1345: ld.const.v2.u64 {%rd1789, %rd2984}, [kernel_params+2224]; cvta.to.global.u64 %rd1792, %rd2984; shl.b64 %rd1793, %rd286, 4; add.s64 %rd1794, %rd1792, %rd1793; ld.global.v4.u32 {%r2030, %r2031, %r2032, %r2033}, [%rd1794]; cvta.to.global.u64 %rd1795, %rd1789; mul.wide.u32 %rd1796, %r2030, 12; add.s64 %rd1797, %rd1795, %rd1796; mul.wide.u32 %rd1798, %r2031, 12; add.s64 %rd1799, %rd1795, %rd1798; mul.wide.u32 %rd1800, %r2032, 12; add.s64 %rd1801, %rd1795, %rd1800; mov.f32 %f9961, 0f3F800000; sub.ftz.f32 %f9962, %f9961, %f1958; sub.ftz.f32 %f9963, %f9962, %f1959; ld.global.f32 %f9964, [%rd1797]; ld.global.f32 %f9965, [%rd1797+4]; ld.global.f32 %f9966, [%rd1797+8]; ld.global.f32 %f9967, [%rd1799]; mul.ftz.f32 %f9968, %f1958, %f9967; ld.global.f32 %f9969, [%rd1799+4]; mul.ftz.f32 %f9970, %f1958, %f9969; ld.global.f32 %f9971, [%rd1799+8]; mul.ftz.f32 %f9972, %f1958, %f9971; fma.rn.ftz.f32 %f9973, %f9963, %f9964, %f9968; fma.rn.ftz.f32 %f9974, %f9963, %f9965, %f9970; fma.rn.ftz.f32 %f9975, %f9963, %f9966, %f9972; ld.global.f32 %f9976, [%rd1801]; ld.global.f32 %f9977, [%rd1801+4]; ld.global.f32 %f9978, [%rd1801+8]; fma.rn.ftz.f32 %f16193, %f1959, %f9976, %f9973; fma.rn.ftz.f32 %f16194, %f1959, %f9977, %f9974; fma.rn.ftz.f32 %f16195, %f1959, %f9978, %f9975; mul.ftz.f32 %f9979, %f16194, %f16194; fma.rn.ftz.f32 %f9980, %f16193, %f16193, %f9979; fma.rn.ftz.f32 %f9981, %f16195, %f16195, %f9980; sqrt.approx.ftz.f32 %f2028, %f9981; setp.neu.ftz.f32 %p1263, %f2028, 0f00000000; @%p1263 bra $L__BB106_1346; bra.uni $L__BB106_1347; $L__BB106_1346: rcp.approx.ftz.f32 %f9982, %f2028; mul.ftz.f32 %f16195, %f16195, %f9982; mul.ftz.f32 %f16194, %f16194, %f9982; mul.ftz.f32 %f16193, %f16193, %f9982; $L__BB106_1347: add.s64 %rd1802, %rd284, 8; setp.eq.ftz.f32 %p1264, %f16193, 0f00000000; setp.eq.ftz.f32 %p1265, %f16194, 0f00000000; and.pred %p1266, %p1264, %p1265; setp.eq.ftz.f32 %p1267, %f16195, 0f00000000; and.pred %p1268, %p1267, %p1266; selp.f32 %f16208, %f1957, %f16195, %p1268; selp.f32 %f16209, %f16213, %f16194, %p1268; selp.f32 %f16210, %f16212, %f16193, %p1268; cvta.to.local.u64 %rd297, %rd284; st.local.f32 [%rd297], %f16210; st.local.f32 [%rd287], %f16209; cvta.to.local.u64 %rd299, %rd1802; st.local.f32 [%rd299], %f16208; and.b32 %r3227, %r626, 4; setp.ne.s32 %p1269, %r3227, 0; @%p1269 bra $L__BB106_1351; @%p1255 bra $L__BB106_1350; bra.uni $L__BB106_1349; $L__BB106_1350: ld.const.u64 %rd1811, [kernel_params+2160]; cvta.to.global.u64 %rd1812, %rd1811; mul.lo.s64 %rd1813, %rd285, 208; add.s64 %rd1814, %rd1812, %rd1813; ld.global.v4.f32 {%f10017, %f10018, %f10019, %f10020}, [%rd1814+48]; ld.global.v4.f32 {%f10024, %f10025, %f10026, %f10027}, [%rd1814+64]; ld.global.v4.f32 {%f10031, %f10032, %f10033, %f10034}, [%rd1814+80]; mul.ftz.f32 %f10038, %f10024, %f16209; fma.rn.ftz.f32 %f10039, %f10017, %f16210, %f10038; fma.rn.ftz.f32 %f10040, %f10031, %f16208, %f10039; mul.ftz.f32 %f10041, %f10025, %f16209; fma.rn.ftz.f32 %f10042, %f10018, %f16210, %f10041; fma.rn.ftz.f32 %f10043, %f10032, %f16208, %f10042; mul.ftz.f32 %f10044, %f10026, %f16209; fma.rn.ftz.f32 %f10045, %f10019, %f16210, %f10044; fma.rn.ftz.f32 %f10046, %f10033, %f16208, %f10045; mul.ftz.f32 %f10047, %f10043, %f10043; fma.rn.ftz.f32 %f10048, %f10040, %f10040, %f10047; fma.rn.ftz.f32 %f10049, %f10046, %f10046, %f10048; sqrt.approx.ftz.f32 %f10050, %f10049; div.approx.ftz.f32 %f16210, %f10040, %f10050; div.approx.ftz.f32 %f16209, %f10043, %f10050; div.approx.ftz.f32 %f16208, %f10046, %f10050; st.local.f32 [%rd297], %f16210; st.local.f32 [%rd287], %f16209; st.local.f32 [%rd299], %f16208; mov.u32 %r3227, %r2023; bra.uni $L__BB106_1351; $L__BB106_711: mov.f32 %f5865, 0fBF800000; max.ftz.f32 %f5866, %f780, %f5865; min.ftz.f32 %f5868, %f5866, %f5864; abs.ftz.f32 %f5869, %f5868; sub.ftz.f32 %f5870, %f5864, %f5869; mul.ftz.f32 %f5871, %f5870, 0f3F000000; sqrt.approx.ftz.f32 %f5872, %f5871; setp.gt.ftz.f32 %p608, %f5869, 0f3F11EB85; selp.f32 %f5873, %f5872, %f5869, %p608; mul.ftz.f32 %f5874, %f5873, %f5873; mov.f32 %f5875, 0f3C94D2E9; mov.f32 %f5876, 0f3D53F941; fma.rn.ftz.f32 %f5877, %f5876, %f5874, %f5875; mov.f32 %f5878, 0f3D3F841F; fma.rn.ftz.f32 %f5879, %f5877, %f5874, %f5878; mov.f32 %f5880, 0f3D994929; fma.rn.ftz.f32 %f5881, %f5879, %f5874, %f5880; mov.f32 %f5882, 0f3E2AAB94; fma.rn.ftz.f32 %f5883, %f5881, %f5874, %f5882; mul.ftz.f32 %f5884, %f5874, %f5883; fma.rn.ftz.f32 %f5885, %f5884, %f5873, %f5873; add.ftz.f32 %f5886, %f5885, %f5885; mov.f32 %f5887, 0f3FC90FDB; sub.ftz.f32 %f5888, %f5887, %f5885; selp.f32 %f5889, %f5886, %f5888, %p608; setp.lt.ftz.f32 %p609, %f5868, 0f00000000; mov.f32 %f5890, 0f40490FDB; sub.ftz.f32 %f5891, %f5890, %f5889; selp.f32 %f5892, %f5891, %f5889, %p609; mul.ftz.f32 %f5893, %f5852, %f780; sub.ftz.f32 %f5894, %f5856, %f5893; mul.ftz.f32 %f5895, %f5853, %f780; sub.ftz.f32 %f5896, %f5857, %f5895; mul.ftz.f32 %f5897, %f5854, %f780; sub.ftz.f32 %f5898, %f5858, %f5897; mul.ftz.f32 %f5899, %f5855, %f780; sub.ftz.f32 %f5900, %f5859, %f5899; mul.ftz.f32 %f5901, %f5896, %f5896; fma.rn.ftz.f32 %f5902, %f5894, %f5894, %f5901; mul.ftz.f32 %f5903, %f5900, %f5900; fma.rn.ftz.f32 %f5904, %f5898, %f5898, %f5903; add.ftz.f32 %f5905, %f5902, %f5904; rsqrt.approx.ftz.f32 %f5906, %f5905; mul.ftz.f32 %f5907, %f5894, %f5906; mul.ftz.f32 %f5908, %f5896, %f5906; mul.ftz.f32 %f5909, %f5898, %f5906; mul.ftz.f32 %f5910, %f5900, %f5906; mul.ftz.f32 %f5911, %f771, %f5892; cos.approx.ftz.f32 %f5912, %f5911; sin.approx.ftz.f32 %f5913, %f5911; mul.ftz.f32 %f5914, %f5907, %f5913; mul.ftz.f32 %f5915, %f5908, %f5913; mul.ftz.f32 %f5916, %f5909, %f5913; mul.ftz.f32 %f5917, %f5910, %f5913; fma.rn.ftz.f32 %f15796, %f5855, %f5912, %f5917; fma.rn.ftz.f32 %f15795, %f5854, %f5912, %f5916; fma.rn.ftz.f32 %f15794, %f5853, %f5912, %f5915; fma.rn.ftz.f32 %f15793, %f5852, %f5912, %f5914; $L__BB106_713: ld.global.v4.f32 {%f5932, %f5933, %f5934, %f5935}, [%rd160+16]; ld.global.v4.f32 {%f5940, %f5941, %f5942, %f5943}, [%rd160+80]; mul.ftz.f32 %f5948, %f771, %f5940; mul.ftz.f32 %f5949, %f771, %f5941; mul.ftz.f32 %f5950, %f771, %f5942; mul.ftz.f32 %f5951, %f771, %f5943; fma.rn.ftz.f32 %f5952, %f781, %f5935, %f5951; ld.global.v4.f32 {%f5953, %f5954, %f5955, %f5956}, [%rd160+32]; ld.global.v4.f32 {%f5961, %f5962, %f5963, %f5964}, [%rd160+96]; mul.ftz.f32 %f5969, %f771, %f5961; mul.ftz.f32 %f5970, %f771, %f5962; mul.ftz.f32 %f5971, %f771, %f5963; mul.ftz.f32 %f5972, %f771, %f5964; fma.rn.ftz.f32 %f5973, %f781, %f5953, %f5969; fma.rn.ftz.f32 %f5974, %f781, %f5954, %f5970; fma.rn.ftz.f32 %f5975, %f781, %f5955, %f5971; fma.rn.ftz.f32 %f5976, %f781, %f5956, %f5972; ld.global.v4.f32 {%f5977, %f5978, %f5979, %f5980}, [%rd160+48]; ld.global.v4.f32 {%f5985, %f5986, %f5987, %f5988}, [%rd160+112]; mul.ftz.f32 %f5993, %f771, %f5985; mul.ftz.f32 %f5994, %f771, %f5986; mul.ftz.f32 %f5995, %f771, %f5987; mul.ftz.f32 %f5996, %f771, %f5988; fma.rn.ftz.f32 %f5997, %f781, %f5977, %f5993; fma.rn.ftz.f32 %f5998, %f781, %f5978, %f5994; fma.rn.ftz.f32 %f5999, %f781, %f5979, %f5995; fma.rn.ftz.f32 %f6000, %f781, %f5980, %f5996; mul.ftz.f32 %f6001, %f15796, 0f3FB504F3; mul.ftz.f32 %f6002, %f15793, 0f3FB504F3; mul.ftz.f32 %f6003, %f15794, 0f3FB504F3; mul.ftz.f32 %f6004, %f15795, 0f3FB504F3; mul.ftz.f32 %f6005, %f6001, %f6002; mul.ftz.f32 %f6006, %f6001, %f6003; mul.ftz.f32 %f6007, %f6001, %f6004; mul.ftz.f32 %f6008, %f6002, %f6002; mul.ftz.f32 %f6009, %f6002, %f6003; mul.ftz.f32 %f6010, %f6002, %f6004; mul.ftz.f32 %f6011, %f6003, %f6003; mul.ftz.f32 %f6012, %f6003, %f6004; mul.ftz.f32 %f6013, %f6004, %f6004; sub.ftz.f32 %f6015, %f5864, %f6011; sub.ftz.f32 %f6016, %f6015, %f6013; sub.ftz.f32 %f6017, %f6009, %f6007; add.ftz.f32 %f6018, %f6006, %f6010; add.ftz.f32 %f6019, %f6007, %f6009; sub.ftz.f32 %f6020, %f5864, %f6008; sub.ftz.f32 %f6021, %f6020, %f6013; sub.ftz.f32 %f6022, %f6012, %f6005; sub.ftz.f32 %f6023, %f6010, %f6006; add.ftz.f32 %f6024, %f6005, %f6012; sub.ftz.f32 %f6025, %f6020, %f6011; mul.ftz.f32 %f6026, %f6017, %f5975; fma.rn.ftz.f32 %f6027, %f6016, %f5952, %f6026; mul.ftz.f32 %f6028, %f6017, %f5976; fma.rn.ftz.f32 %f6029, %f6016, %f5973, %f6028; mul.ftz.f32 %f6030, %f6017, %f5997; fma.rn.ftz.f32 %f6031, %f6016, %f5974, %f6030; mul.ftz.f32 %f6032, %f6021, %f5975; fma.rn.ftz.f32 %f6033, %f6019, %f5952, %f6032; mul.ftz.f32 %f6034, %f6021, %f5976; fma.rn.ftz.f32 %f6035, %f6019, %f5973, %f6034; mul.ftz.f32 %f6036, %f6021, %f5997; fma.rn.ftz.f32 %f6037, %f6019, %f5974, %f6036; mul.ftz.f32 %f6038, %f6024, %f5975; fma.rn.ftz.f32 %f6039, %f6023, %f5952, %f6038; mul.ftz.f32 %f6040, %f6024, %f5976; fma.rn.ftz.f32 %f6041, %f6023, %f5973, %f6040; mul.ftz.f32 %f6042, %f6024, %f5997; fma.rn.ftz.f32 %f6043, %f6023, %f5974, %f6042; fma.rn.ftz.f32 %f15803, %f6022, %f6000, %f6037; fma.rn.ftz.f32 %f15802, %f6022, %f5999, %f6035; fma.rn.ftz.f32 %f15801, %f6022, %f5998, %f6033; fma.rn.ftz.f32 %f15807, %f6025, %f6000, %f6043; fma.rn.ftz.f32 %f15806, %f6025, %f5999, %f6041; fma.rn.ftz.f32 %f15805, %f6025, %f5998, %f6039; fma.rn.ftz.f32 %f15799, %f6018, %f6000, %f6031; fma.rn.ftz.f32 %f15798, %f6018, %f5999, %f6029; fma.rn.ftz.f32 %f15797, %f6018, %f5998, %f6027; fma.rn.ftz.f32 %f15808, %f781, %f5934, %f5950; fma.rn.ftz.f32 %f15804, %f781, %f5933, %f5949; fma.rn.ftz.f32 %f15800, %f781, %f5932, %f5948; $L__BB106_715: mul.ftz.f32 %f6056, %f15810, %f15798; fma.rn.ftz.f32 %f6057, %f15809, %f15797, %f6056; fma.rn.ftz.f32 %f6058, %f15811, %f15799, %f6057; add.ftz.f32 %f830, %f15800, %f6058; mul.ftz.f32 %f6059, %f15810, %f15802; fma.rn.ftz.f32 %f6060, %f15809, %f15801, %f6059; fma.rn.ftz.f32 %f6061, %f15811, %f15803, %f6060; add.ftz.f32 %f831, %f15804, %f6061; mul.ftz.f32 %f6062, %f15810, %f15806; fma.rn.ftz.f32 %f6063, %f15809, %f15805, %f6062; fma.rn.ftz.f32 %f6064, %f15811, %f15807, %f6063; add.ftz.f32 %f15811, %f15808, %f6064; mul.ftz.f32 %f6065, %f15813, %f15798; fma.rn.ftz.f32 %f6066, %f15812, %f15797, %f6065; fma.rn.ftz.f32 %f6067, %f15814, %f15799, %f6066; add.ftz.f32 %f833, %f15800, %f6067; mul.ftz.f32 %f6068, %f15813, %f15802; fma.rn.ftz.f32 %f6069, %f15812, %f15801, %f6068; fma.rn.ftz.f32 %f6070, %f15814, %f15803, %f6069; add.ftz.f32 %f834, %f15804, %f6070; mul.ftz.f32 %f6071, %f15813, %f15806; fma.rn.ftz.f32 %f6072, %f15812, %f15805, %f6071; fma.rn.ftz.f32 %f6073, %f15814, %f15807, %f6072; add.ftz.f32 %f15814, %f15808, %f6073; mul.ftz.f32 %f6074, %f15816, %f15798; fma.rn.ftz.f32 %f6075, %f15815, %f15797, %f6074; fma.rn.ftz.f32 %f6076, %f15817, %f15799, %f6075; add.ftz.f32 %f836, %f15800, %f6076; mul.ftz.f32 %f6077, %f15816, %f15802; fma.rn.ftz.f32 %f6078, %f15815, %f15801, %f6077; fma.rn.ftz.f32 %f6079, %f15817, %f15803, %f6078; add.ftz.f32 %f837, %f15804, %f6079; mul.ftz.f32 %f6080, %f15816, %f15806; fma.rn.ftz.f32 %f6081, %f15815, %f15805, %f6080; fma.rn.ftz.f32 %f6082, %f15817, %f15807, %f6081; add.ftz.f32 %f15817, %f15808, %f6082; mov.u16 %rs140, 1; mov.f32 %f15809, %f830; mov.f32 %f15810, %f831; mov.f32 %f15812, %f833; mov.f32 %f15813, %f834; mov.f32 %f15815, %f836; mov.f32 %f15816, %f837; $L__BB106_716: sub.ftz.f32 %f6083, %f15812, %f15809; sub.ftz.f32 %f6084, %f15813, %f15810; mul.ftz.f32 %f6085, %f6084, %f6084; fma.rn.ftz.f32 %f6086, %f6083, %f6083, %f6085; sub.ftz.f32 %f6087, %f15814, %f15811; fma.rn.ftz.f32 %f6088, %f6087, %f6087, %f6086; sub.ftz.f32 %f6089, %f15815, %f15809; sub.ftz.f32 %f6090, %f15816, %f15810; mul.ftz.f32 %f6091, %f6090, %f6090; fma.rn.ftz.f32 %f6092, %f6089, %f6089, %f6091; sub.ftz.f32 %f6093, %f15817, %f15811; fma.rn.ftz.f32 %f6094, %f6093, %f6093, %f6092; sub.ftz.f32 %f6095, %f15815, %f15812; sub.ftz.f32 %f6096, %f15816, %f15813; mul.ftz.f32 %f6097, %f6096, %f6096; fma.rn.ftz.f32 %f6098, %f6095, %f6095, %f6097; sub.ftz.f32 %f6099, %f15817, %f15814; fma.rn.ftz.f32 %f6100, %f6099, %f6099, %f6098; max.ftz.f32 %f6101, %f6094, %f6100; max.ftz.f32 %f6102, %f6088, %f6101; mul.ftz.f32 %f6103, %f6084, %f6093; mul.ftz.f32 %f6104, %f6087, %f6090; sub.ftz.f32 %f6105, %f6103, %f6104; mul.ftz.f32 %f6106, %f6087, %f6089; mul.ftz.f32 %f6107, %f6083, %f6093; sub.ftz.f32 %f6108, %f6106, %f6107; mul.ftz.f32 %f6109, %f6083, %f6090; mul.ftz.f32 %f6110, %f6084, %f6089; sub.ftz.f32 %f6111, %f6109, %f6110; ld.local.f32 %f848, [%rd7+-52]; mul.ftz.f32 %f849, %f732, %f848; ld.local.v2.f32 {%f6112, %f6113}, [%rd7+-48]; mul.ftz.f32 %f852, %f732, %f6112; mul.ftz.f32 %f853, %f732, %f6113; mul.ftz.f32 %f6114, %f6108, %f852; fma.rn.ftz.f32 %f6115, %f6105, %f849, %f6114; fma.rn.ftz.f32 %f6116, %f6111, %f853, %f6115; abs.ftz.f32 %f6117, %f6116; mul.ftz.f32 %f6118, %f6108, %f6108; fma.rn.ftz.f32 %f6119, %f6105, %f6105, %f6118; fma.rn.ftz.f32 %f854, %f6111, %f6111, %f6119; div.approx.ftz.f32 %f6120, %f6117, %f854; mul.ftz.f32 %f6121, %f6120, %f6120; setp.gt.ftz.f32 %p610, %f6102, %f6121; @%p610 bra $L__BB106_730; bra.uni $L__BB106_717; $L__BB106_730: ld.local.v4.f32 {%f6387, %f6388, %f6389, %f6390}, [%rd1]; add.ftz.f32 %f6394, %f849, %f6387; add.ftz.f32 %f6395, %f852, %f6388; add.ftz.f32 %f6396, %f853, %f6389; sub.ftz.f32 %f946, %f15809, %f6394; sub.ftz.f32 %f947, %f15810, %f6395; sub.ftz.f32 %f948, %f15811, %f6396; sub.ftz.f32 %f949, %f15812, %f6394; sub.ftz.f32 %f950, %f15813, %f6395; sub.ftz.f32 %f951, %f15814, %f6396; sub.ftz.f32 %f952, %f15815, %f6394; sub.ftz.f32 %f953, %f15816, %f6395; sub.ftz.f32 %f954, %f15817, %f6396; mul.ftz.f32 %f6397, %f947, %f951; mul.ftz.f32 %f6398, %f950, %f948; sub.ftz.f32 %f15844, %f6397, %f6398; mul.ftz.f32 %f6399, %f949, %f948; mul.ftz.f32 %f6400, %f946, %f951; sub.ftz.f32 %f15845, %f6399, %f6400; mul.ftz.f32 %f6401, %f946, %f950; mul.ftz.f32 %f6402, %f949, %f947; sub.ftz.f32 %f15846, %f6401, %f6402; mul.ftz.f32 %f6403, %f15845, %f15845; fma.rn.ftz.f32 %f6404, %f15844, %f15844, %f6403; fma.rn.ftz.f32 %f6405, %f15846, %f15846, %f6404; sqrt.approx.ftz.f32 %f958, %f6405; setp.neu.ftz.f32 %p619, %f958, 0f00000000; @%p619 bra $L__BB106_731; bra.uni $L__BB106_732; $L__BB106_731: rcp.approx.ftz.f32 %f6406, %f958; mul.ftz.f32 %f15846, %f6406, %f15846; mul.ftz.f32 %f15845, %f6406, %f15845; mul.ftz.f32 %f15844, %f6406, %f15844; $L__BB106_732: mul.ftz.f32 %f6407, %f953, %f948; mul.ftz.f32 %f6408, %f947, %f954; sub.ftz.f32 %f15847, %f6408, %f6407; mul.ftz.f32 %f6409, %f946, %f954; mul.ftz.f32 %f6410, %f952, %f948; sub.ftz.f32 %f15848, %f6410, %f6409; mul.ftz.f32 %f6411, %f952, %f947; mul.ftz.f32 %f6412, %f946, %f953; sub.ftz.f32 %f15849, %f6412, %f6411; mul.ftz.f32 %f6413, %f15848, %f15848; fma.rn.ftz.f32 %f6414, %f15847, %f15847, %f6413; fma.rn.ftz.f32 %f6415, %f15849, %f15849, %f6414; sqrt.approx.ftz.f32 %f974, %f6415; setp.neu.ftz.f32 %p620, %f974, 0f00000000; @%p620 bra $L__BB106_733; bra.uni $L__BB106_734; $L__BB106_733: rcp.approx.ftz.f32 %f6416, %f974; mul.ftz.f32 %f15849, %f15849, %f6416; mul.ftz.f32 %f15848, %f15848, %f6416; mul.ftz.f32 %f15847, %f15847, %f6416; $L__BB106_734: mul.ftz.f32 %f6417, %f953, %f951; mul.ftz.f32 %f6418, %f950, %f954; sub.ftz.f32 %f15850, %f6418, %f6417; mul.ftz.f32 %f6419, %f949, %f954; mul.ftz.f32 %f6420, %f952, %f951; sub.ftz.f32 %f15851, %f6420, %f6419; mul.ftz.f32 %f6421, %f952, %f950; mul.ftz.f32 %f6422, %f949, %f953; sub.ftz.f32 %f15852, %f6422, %f6421; mul.ftz.f32 %f6423, %f15851, %f15851; fma.rn.ftz.f32 %f6424, %f15850, %f15850, %f6423; fma.rn.ftz.f32 %f6425, %f15852, %f15852, %f6424; sqrt.approx.ftz.f32 %f990, %f6425; setp.neu.ftz.f32 %p621, %f990, 0f00000000; @%p621 bra $L__BB106_735; bra.uni $L__BB106_736; $L__BB106_735: rcp.approx.ftz.f32 %f6426, %f990; mul.ftz.f32 %f15852, %f15852, %f6426; mul.ftz.f32 %f15851, %f15851, %f6426; mul.ftz.f32 %f15850, %f15850, %f6426; $L__BB106_736: mul.ftz.f32 %f6428, %f15845, %f15848; fma.rn.ftz.f32 %f6429, %f15844, %f15847, %f6428; fma.rn.ftz.f32 %f6430, %f15846, %f15849, %f6429; abs.ftz.f32 %f6431, %f6430; setp.lt.ftz.f32 %p622, %f6431, 0f3F800000; mov.f32 %f6432, 0f3F800000; add.ftz.f32 %f6433, %f6431, 0fBF800000; add.ftz.f32 %f6434, %f6433, 0f3F800000; selp.f32 %f6435, %f6434, 0f3F800000, %p622; sub.ftz.f32 %f6436, %f6432, %f6435; sqrt.approx.ftz.f32 %f6437, %f6436; mul.ftz.f32 %f6438, %f6435, 0f3CB1485D; mov.f32 %f6439, 0f3D9FB43B; sub.ftz.f32 %f6440, %f6439, %f6438; fma.rn.ftz.f32 %f6441, %f6435, %f6440, 0fBE5A6B93; fma.rn.ftz.f32 %f6442, %f6435, %f6441, 0f3FC90FDB; mul.ftz.f32 %f6443, %f6437, %f6442; setp.lt.ftz.f32 %p623, %f6430, 0f00000000; mov.f32 %f15879, 0f00000000; mov.f32 %f6444, 0f40490FDB; sub.ftz.f32 %f6445, %f6444, %f6443; selp.f32 %f6446, %f6445, %f6443, %p623; mul.ftz.f32 %f6447, %f15845, %f15851; fma.rn.ftz.f32 %f6448, %f15844, %f15850, %f6447; fma.rn.ftz.f32 %f6449, %f15846, %f15852, %f6448; neg.ftz.f32 %f6450, %f6449; abs.ftz.f32 %f6451, %f6450; setp.lt.ftz.f32 %p624, %f6451, 0f3F800000; add.ftz.f32 %f6452, %f6451, 0fBF800000; add.ftz.f32 %f6453, %f6452, 0f3F800000; selp.f32 %f6454, %f6453, 0f3F800000, %p624; sub.ftz.f32 %f6455, %f6432, %f6454; sqrt.approx.ftz.f32 %f6456, %f6455; mul.ftz.f32 %f6457, %f6454, 0f3CB1485D; sub.ftz.f32 %f6458, %f6439, %f6457; fma.rn.ftz.f32 %f6459, %f6454, %f6458, 0fBE5A6B93; fma.rn.ftz.f32 %f6460, %f6454, %f6459, 0f3FC90FDB; mul.ftz.f32 %f6461, %f6456, %f6460; setp.gt.ftz.f32 %p625, %f6449, 0f80000000; sub.ftz.f32 %f6462, %f6444, %f6461; selp.f32 %f6463, %f6462, %f6461, %p625; mul.ftz.f32 %f6464, %f15848, %f15851; fma.rn.ftz.f32 %f6465, %f15847, %f15850, %f6464; fma.rn.ftz.f32 %f6466, %f15849, %f15852, %f6465; abs.ftz.f32 %f6467, %f6466; setp.lt.ftz.f32 %p626, %f6467, 0f3F800000; add.ftz.f32 %f6468, %f6467, 0fBF800000; add.ftz.f32 %f6469, %f6468, 0f3F800000; selp.f32 %f6470, %f6469, 0f3F800000, %p626; sub.ftz.f32 %f6471, %f6432, %f6470; sqrt.approx.ftz.f32 %f6472, %f6471; mul.ftz.f32 %f6473, %f6470, 0f3CB1485D; sub.ftz.f32 %f6474, %f6439, %f6473; fma.rn.ftz.f32 %f6475, %f6470, %f6474, 0fBE5A6B93; fma.rn.ftz.f32 %f6476, %f6470, %f6475, 0f3FC90FDB; mul.ftz.f32 %f6477, %f6472, %f6476; setp.lt.ftz.f32 %p627, %f6466, 0f00000000; sub.ftz.f32 %f6478, %f6444, %f6477; selp.f32 %f6479, %f6478, %f6477, %p627; add.ftz.f32 %f6480, %f6446, %f6463; add.ftz.f32 %f6481, %f6480, %f6479; add.ftz.f32 %f1000, %f6481, 0fC0490FDB; setp.eq.ftz.f32 %p628, %f1000, 0f00000000; @%p628 bra $L__BB106_749; setp.eq.s16 %p629, %rs140, 0; @%p629 bra $L__BB106_747; cvt.u32.u64 %r1411, %rd96; cvta.to.global.u64 %rd1202, %rd2957; mul.wide.s32 %rd1203, %r1411, 16; add.s64 %rd1204, %rd1202, %rd1203; ld.const.u64 %rd1205, [kernel_params+2256]; cvta.to.global.u64 %rd1206, %rd1205; ld.global.u32 %r1412, [%rd1204+12]; mul.wide.u32 %rd1207, %r1412, 12; add.s64 %rd1208, %rd1206, %rd1207; ld.global.f32 %f15869, [%rd1208]; ld.global.f32 %f15870, [%rd1208+4]; ld.global.f32 %f15871, [%rd1208+8]; add.s32 %r1413, %r1412, 1; mul.wide.u32 %rd1209, %r1413, 12; add.s64 %rd1210, %rd1206, %rd1209; ld.global.f32 %f15872, [%rd1210]; ld.global.f32 %f15873, [%rd1210+4]; ld.global.f32 %f15874, [%rd1210+8]; add.s32 %r1414, %r1412, 2; mul.wide.u32 %rd1211, %r1414, 12; add.s64 %rd1212, %rd1206, %rd1211; ld.global.f32 %f15875, [%rd1212]; ld.global.f32 %f15876, [%rd1212+4]; ld.global.f32 %f15877, [%rd1212+8]; @%p604 bra $L__BB106_746; and.b32 %r1415, %r361, 2; setp.eq.s32 %p631, %r1415, 0; ld.const.u64 %rd1213, [kernel_params+2160]; cvta.to.global.u64 %rd1214, %rd1213; mul.lo.s64 %rd1215, %rd97, 208; add.s64 %rd163, %rd1214, %rd1215; @%p631 bra $L__BB106_744; bra.uni $L__BB106_740; $L__BB106_744: ld.global.v4.f32 {%f15857, %f15858, %f15859, %f15860}, [%rd163]; ld.global.v4.f32 {%f15861, %f15862, %f15863, %f15864}, [%rd163+16]; ld.global.v4.f32 {%f15865, %f15866, %f15867, %f15868}, [%rd163+32]; bra.uni $L__BB106_745; $L__BB106_717: mul.ftz.f32 %f6123, %f5815, %f6112; fma.rn.ftz.f32 %f6124, %f5814, %f848, %f6123; mov.b32 %r1398, %f5816; mov.b64 %rd1179, {%r1398, %r1399}; cvt.u32.u64 %r1400, %rd1179; mov.b32 %f6125, %r1400; fma.rn.ftz.f32 %f6126, %f6125, %f6113, %f6124; abs.ftz.f32 %f855, %f6126; setp.eq.ftz.f32 %p611, %f855, 0f00000000; mov.f32 %f15818, 0f00000000; @%p611 bra $L__BB106_719; ld.const.f32 %f6127, [kernel_params+1500]; mul.ftz.f32 %f6128, %f732, %f732; mul.ftz.f32 %f6129, %f6128, %f6127; div.approx.ftz.f32 %f15818, %f6129, %f855; $L__BB106_719: setp.eq.s16 %p612, %rs140, 0; mov.f32 %f15879, %f15818; @%p612 bra $L__BB106_749; sqrt.approx.ftz.f32 %f6131, %f854; mul.ftz.f32 %f858, %f6131, 0f3F000000; setp.eq.ftz.f32 %p613, %f858, 0f00000000; mov.f32 %f15879, 0f00000000; @%p613 bra $L__BB106_749; cvt.u32.u64 %r1401, %rd96; cvta.to.global.u64 %rd1180, %rd2957; mul.wide.s32 %rd1181, %r1401, 16; add.s64 %rd1182, %rd1180, %rd1181; ld.const.u64 %rd1183, [kernel_params+2256]; cvta.to.global.u64 %rd1184, %rd1183; ld.global.u32 %r1402, [%rd1182+12]; mul.wide.u32 %rd1185, %r1402, 12; add.s64 %rd1186, %rd1184, %rd1185; ld.global.f32 %f15835, [%rd1186]; ld.global.f32 %f15836, [%rd1186+4]; ld.global.f32 %f15837, [%rd1186+8]; add.s32 %r1403, %r1402, 1; mul.wide.u32 %rd1187, %r1403, 12; add.s64 %rd1188, %rd1184, %rd1187; ld.global.f32 %f15838, [%rd1188]; ld.global.f32 %f15839, [%rd1188+4]; ld.global.f32 %f15840, [%rd1188+8]; add.s32 %r1404, %r1402, 2; mul.wide.u32 %rd1189, %r1404, 12; add.s64 %rd1190, %rd1184, %rd1189; ld.global.f32 %f15841, [%rd1190]; ld.global.f32 %f15842, [%rd1190+4]; ld.global.f32 %f15843, [%rd1190+8]; @%p604 bra $L__BB106_729; and.b32 %r1405, %r361, 2; setp.eq.s32 %p615, %r1405, 0; ld.const.u64 %rd1191, [kernel_params+2160]; cvta.to.global.u64 %rd1192, %rd1191; mul.lo.s64 %rd1193, %rd97, 208; add.s64 %rd161, %rd1192, %rd1193; @%p615 bra $L__BB106_727; bra.uni $L__BB106_723; $L__BB106_727: ld.global.v4.f32 {%f15823, %f15824, %f15825, %f15826}, [%rd161]; ld.global.v4.f32 {%f15827, %f15828, %f15829, %f15830}, [%rd161+16]; ld.global.v4.f32 {%f15831, %f15832, %f15833, %f15834}, [%rd161+32]; bra.uni $L__BB106_728; $L__BB106_747: sqrt.approx.ftz.f32 %f15878, %f854; bra.uni $L__BB106_748; $L__BB106_1349: add.s64 %rd1805, %rd1729, 192; cvta.to.local.u64 %rd1806, %rd1805; ld.local.v4.f32 {%f9983, %f9984, %f9985, %f9986}, [%rd1806]; add.s64 %rd1807, %rd1729, 208; cvta.to.local.u64 %rd1808, %rd1807; ld.local.v4.f32 {%f9990, %f9991, %f9992, %f9993}, [%rd1808]; add.s64 %rd1809, %rd1729, 224; cvta.to.local.u64 %rd1810, %rd1809; ld.local.v4.f32 {%f9997, %f9998, %f9999, %f10000}, [%rd1810]; mul.ftz.f32 %f10004, %f16209, %f9990; fma.rn.ftz.f32 %f10005, %f16210, %f9983, %f10004; fma.rn.ftz.f32 %f10006, %f16208, %f9997, %f10005; mul.ftz.f32 %f10007, %f16209, %f9991; fma.rn.ftz.f32 %f10008, %f16210, %f9984, %f10007; fma.rn.ftz.f32 %f10009, %f16208, %f9998, %f10008; mul.ftz.f32 %f10010, %f16209, %f9992; fma.rn.ftz.f32 %f10011, %f16210, %f9985, %f10010; fma.rn.ftz.f32 %f10012, %f16208, %f9999, %f10011; mul.ftz.f32 %f10013, %f10009, %f10009; fma.rn.ftz.f32 %f10014, %f10006, %f10006, %f10013; fma.rn.ftz.f32 %f10015, %f10012, %f10012, %f10014; sqrt.approx.ftz.f32 %f10016, %f10015; div.approx.ftz.f32 %f16210, %f10006, %f10016; div.approx.ftz.f32 %f16209, %f10009, %f10016; div.approx.ftz.f32 %f16208, %f10012, %f10016; st.local.f32 [%rd297], %f16210; st.local.f32 [%rd287], %f16209; st.local.f32 [%rd299], %f16208; mov.u32 %r3227, %r2023; $L__BB106_1351: cvta.to.global.u64 %rd1815, %rd2984; shl.b64 %rd1816, %rd286, 4; add.s64 %rd1817, %rd1815, %rd1816; ld.const.u64 %rd1818, [kernel_params+2256]; cvta.to.global.u64 %rd1819, %rd1818; ld.global.u32 %r2039, [%rd1817+12]; mul.wide.u32 %rd1820, %r2039, 12; add.s64 %rd1821, %rd1819, %rd1820; add.s32 %r2040, %r2039, 1; mul.wide.u32 %rd1822, %r2040, 12; add.s64 %rd1823, %rd1819, %rd1822; add.s32 %r2041, %r2039, 2; mul.wide.u32 %rd1824, %r2041, 12; add.s64 %rd1825, %rd1819, %rd1824; ld.global.f32 %f10051, [%rd1825]; ld.global.f32 %f10052, [%rd1825+4]; ld.global.f32 %f10053, [%rd1825+8]; ld.global.f32 %f10054, [%rd1823]; ld.global.f32 %f10055, [%rd1821]; sub.ftz.f32 %f16207, %f10054, %f10055; ld.global.f32 %f10056, [%rd1823+4]; ld.global.f32 %f10057, [%rd1821+4]; sub.ftz.f32 %f16206, %f10056, %f10057; ld.global.f32 %f10058, [%rd1823+8]; ld.global.f32 %f10059, [%rd1821+8]; sub.ftz.f32 %f16205, %f10058, %f10059; add.s64 %rd1827, %rd1729, 116; cvta.to.local.u64 %rd303, %rd1827; st.local.f32 [%rd303], %f16207; add.s64 %rd1828, %rd1729, 120; cvta.to.local.u64 %rd304, %rd1828; st.local.f32 [%rd304], %f16206; add.s64 %rd1829, %rd1729, 124; cvta.to.local.u64 %rd305, %rd1829; st.local.f32 [%rd305], %f16205; sub.ftz.f32 %f16204, %f10051, %f10055; sub.ftz.f32 %f16203, %f10052, %f10057; sub.ftz.f32 %f16202, %f10053, %f10059; add.s64 %rd1830, %rd1729, 128; cvta.to.local.u64 %rd306, %rd1830; st.local.v2.f32 [%rd306], {%f16204, %f16203}; add.s64 %rd1831, %rd1729, 136; cvta.to.local.u64 %rd307, %rd1831; st.local.f32 [%rd307], %f16202; setp.ne.s32 %p1271, %r3227, 0; @%p1271 bra $L__BB106_1361; @%p1255 bra $L__BB106_1354; bra.uni $L__BB106_1353; $L__BB106_1354: ld.const.u64 %rd1839, [kernel_params+2160]; cvta.to.global.u64 %rd1840, %rd1839; mul.lo.s64 %rd1841, %rd285, 208; add.s64 %rd1842, %rd1840, %rd1841; ld.global.v4.f32 {%f10087, %f10088, %f10089, %f10090}, [%rd1842]; ld.global.v4.f32 {%f10094, %f10095, %f10096, %f10097}, [%rd1842+16]; ld.global.v4.f32 {%f10101, %f10102, %f10103, %f10104}, [%rd1842+32]; mul.ftz.f32 %f10108, %f10088, %f16206; fma.rn.ftz.f32 %f10109, %f16207, %f10087, %f10108; fma.rn.ftz.f32 %f16201, %f10089, %f16205, %f10109; mul.ftz.f32 %f10110, %f10095, %f16206; fma.rn.ftz.f32 %f10111, %f16207, %f10094, %f10110; fma.rn.ftz.f32 %f16200, %f10096, %f16205, %f10111; mul.ftz.f32 %f10112, %f10102, %f16206; fma.rn.ftz.f32 %f10113, %f16207, %f10101, %f10112; fma.rn.ftz.f32 %f16205, %f10103, %f16205, %f10113; bra.uni $L__BB106_1355; $L__BB106_1353: add.s64 %rd1833, %rd1729, 144; cvta.to.local.u64 %rd1834, %rd1833; ld.local.v4.f32 {%f10060, %f10061, %f10062, %f10063}, [%rd1834]; mul.ftz.f32 %f10067, %f16206, %f10061; fma.rn.ftz.f32 %f10068, %f16207, %f10060, %f10067; fma.rn.ftz.f32 %f16201, %f16205, %f10062, %f10068; add.s64 %rd1835, %rd1729, 160; cvta.to.local.u64 %rd1836, %rd1835; ld.local.v4.f32 {%f10069, %f10070, %f10071, %f10072}, [%rd1836]; mul.ftz.f32 %f10076, %f16206, %f10070; fma.rn.ftz.f32 %f10077, %f16207, %f10069, %f10076; fma.rn.ftz.f32 %f16200, %f16205, %f10071, %f10077; add.s64 %rd1837, %rd1729, 176; cvta.to.local.u64 %rd1838, %rd1837; ld.local.v4.f32 {%f10078, %f10079, %f10080, %f10081}, [%rd1838]; mul.ftz.f32 %f10085, %f16206, %f10079; fma.rn.ftz.f32 %f10086, %f16207, %f10078, %f10085; fma.rn.ftz.f32 %f16205, %f16205, %f10080, %f10086; $L__BB106_1355: st.local.f32 [%rd303], %f16201; st.local.f32 [%rd304], %f16200; st.local.f32 [%rd305], %f16205; @%p1255 bra $L__BB106_1357; bra.uni $L__BB106_1356; $L__BB106_1357: ld.const.u64 %rd1850, [kernel_params+2160]; cvta.to.global.u64 %rd1851, %rd1850; mul.lo.s64 %rd1852, %rd285, 208; add.s64 %rd1853, %rd1851, %rd1852; ld.global.v4.f32 {%f10141, %f10142, %f10143, %f10144}, [%rd1853]; ld.global.v4.f32 {%f10148, %f10149, %f10150, %f10151}, [%rd1853+16]; ld.global.v4.f32 {%f10155, %f10156, %f10157, %f10158}, [%rd1853+32]; mul.ftz.f32 %f10162, %f10142, %f16203; fma.rn.ftz.f32 %f10163, %f16204, %f10141, %f10162; fma.rn.ftz.f32 %f2068, %f10143, %f16202, %f10163; mul.ftz.f32 %f10164, %f10149, %f16203; fma.rn.ftz.f32 %f10165, %f16204, %f10148, %f10164; fma.rn.ftz.f32 %f2069, %f10150, %f16202, %f10165; mul.ftz.f32 %f10166, %f10156, %f16203; fma.rn.ftz.f32 %f10167, %f16204, %f10155, %f10166; fma.rn.ftz.f32 %f16202, %f10157, %f16202, %f10167; st.local.v2.f32 [%rd306], {%f2068, %f2069}; st.local.f32 [%rd307], %f16202; mov.f32 %f16203, %f2069; mov.f32 %f16204, %f2068; mov.f32 %f16206, %f16200; mov.f32 %f16207, %f16201; bra.uni $L__BB106_1361; $L__BB106_1356: add.s64 %rd1844, %rd1729, 144; cvta.to.local.u64 %rd1845, %rd1844; ld.local.v4.f32 {%f10114, %f10115, %f10116, %f10117}, [%rd1845]; mul.ftz.f32 %f10121, %f16203, %f10115; fma.rn.ftz.f32 %f10122, %f16204, %f10114, %f10121; fma.rn.ftz.f32 %f2065, %f16202, %f10116, %f10122; add.s64 %rd1846, %rd1729, 160; cvta.to.local.u64 %rd1847, %rd1846; ld.local.v4.f32 {%f10123, %f10124, %f10125, %f10126}, [%rd1847]; mul.ftz.f32 %f10130, %f16203, %f10124; fma.rn.ftz.f32 %f10131, %f16204, %f10123, %f10130; fma.rn.ftz.f32 %f2066, %f16202, %f10125, %f10131; add.s64 %rd1848, %rd1729, 176; cvta.to.local.u64 %rd1849, %rd1848; ld.local.v4.f32 {%f10132, %f10133, %f10134, %f10135}, [%rd1849]; mul.ftz.f32 %f10139, %f16203, %f10133; fma.rn.ftz.f32 %f10140, %f16204, %f10132, %f10139; fma.rn.ftz.f32 %f16202, %f16202, %f10134, %f10140; st.local.v2.f32 [%rd306], {%f2065, %f2066}; st.local.f32 [%rd307], %f16202; mov.f32 %f16203, %f2066; mov.f32 %f16204, %f2065; mov.f32 %f16206, %f16200; mov.f32 %f16207, %f16201; $L__BB106_1361: mov.b32 %r2043, %f1957; mov.b64 %rd1866, {%r2043, %r2044}; cvt.u32.u64 %r2045, %rd1866; mov.b32 %f16211, %r2045; @%p1252 bra $L__BB106_1364; mul.ftz.f32 %f10174, %f16213, %f1953; fma.rn.ftz.f32 %f10175, %f16212, %f1952, %f10174; fma.rn.ftz.f32 %f10176, %f1957, %f1954, %f10175; setp.geu.ftz.f32 %p1276, %f10176, 0f00000000; @%p1276 bra $L__BB106_1364; or.b32 %r2046, %r620, 1; st.local.u32 [%rd291], %r2046; neg.ftz.f32 %f16211, %f1957; st.local.f32 [%rd1735], %f16211; cvta.to.local.u64 %rd1869, %rd284; neg.ftz.f32 %f10177, %f16210; st.local.f32 [%rd1869], %f10177; neg.ftz.f32 %f10178, %f16208; neg.ftz.f32 %f10179, %f16209; neg.ftz.f32 %f16213, %f16213; neg.ftz.f32 %f16212, %f16212; st.local.v4.f32 [%rd287], {%f10179, %f10178, %f16212, %f16213}; add.s64 %rd1871, %rd1729, 116; add.s64 %rd1872, %rd1729, 120; add.s64 %rd1873, %rd1729, 124; cvta.to.local.u64 %rd1874, %rd1871; neg.ftz.f32 %f10180, %f16207; st.local.f32 [%rd1874], %f10180; cvta.to.local.u64 %rd1875, %rd1872; neg.ftz.f32 %f10181, %f16206; st.local.f32 [%rd1875], %f10181; cvta.to.local.u64 %rd1876, %rd1873; neg.ftz.f32 %f10182, %f16205; st.local.f32 [%rd1876], %f10182; add.s64 %rd1877, %rd1729, 136; cvta.to.local.u64 %rd1878, %rd1877; neg.ftz.f32 %f10183, %f16203; neg.ftz.f32 %f10184, %f16204; add.s64 %rd1879, %rd1729, 128; cvta.to.local.u64 %rd1880, %rd1879; st.local.v2.f32 [%rd1880], {%f10184, %f10183}; neg.ftz.f32 %f10185, %f16202; st.local.f32 [%rd1878], %f10185; $L__BB106_1364: add.s64 %rd1882, %rd1729, 92; cvta.to.local.u64 %rd1883, %rd1882; mov.u32 %r2047, 0; st.local.u32 [%rd1883], %r2047; add.s64 %rd1884, %rd1729, 96; cvta.to.local.u64 %rd1885, %rd1884; mov.f32 %f10186, 0f00000000; st.local.v4.f32 [%rd1885], {%f10186, %f10186, %f10186, %f10186}; add.s64 %rd1886, %rd1729, 112; cvta.to.local.u64 %rd1887, %rd1886; st.local.u32 [%rd1887], %r2047; st.local.f32 [%rd282+-12], %f16212; st.local.f32 [%rd282+-8], %f16213; st.local.f32 [%rd282+-4], %f16211; $L__BB106_1366: add.s64 %rd1920, %rd1729, 260; cvta.to.local.u64 %rd1921, %rd1920; mov.u32 %r2054, 0; st.local.u32 [%rd1921], %r2054; add.s64 %rd1922, %rd1729, 264; cvta.to.local.u64 %rd1923, %rd1922; st.local.u32 [%rd1923], %r2054; ld.const.u64 %rd1924, [kernel_params+2400]; cvta.to.global.u64 %rd308, %rd1924; add.s64 %rd1925, %rd1729, 268; cvta.to.local.u64 %rd312, %rd1925; add.s64 %rd1926, %rd1729, 272; cvta.to.local.u64 %rd313, %rd1926; add.s64 %rd1927, %rd1729, 276; cvta.to.local.u64 %rd314, %rd1927; add.u64 %rd1928, %SP, 704; add.u64 %rd315, %SPL, 704; bra.uni $L__BB106_1367; $L__BB106_1519: { // callseq 171, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 171 $L__BB106_1367: mov.u32 %r633, %r3229; mul.wide.s32 %rd1929, %r633, 16; add.s64 %rd316, %rd308, %rd1929; ld.global.v4.u32 {%r2056, %r2057, %r2058, %r2059}, [%rd316]; mul.wide.u32 %rd1930, %r2057, 4; add.s64 %rd317, %rd315, %rd1930; mul.wide.u32 %rd1931, %r2058, 4; add.s64 %rd318, %rd315, %rd1931; add.s32 %r3229, %r633, 1; setp.gt.s32 %p1277, %r2056, 36; @%p1277 bra $L__BB106_1416; setp.gt.s32 %p1359, %r2056, 25; @%p1359 bra $L__BB106_1412; bra.uni $L__BB106_1369; $L__BB106_1412: add.s32 %r2061, %r2056, -27; setp.lt.u32 %p1360, %r2061, 9; @%p1360 bra $L__BB106_1367; setp.eq.s32 %p1361, %r2056, 26; @%p1361 bra $L__BB106_1558; setp.eq.s32 %p1362, %r2056, 36; @%p1362 bra $L__BB106_1415; bra.uni $L__BB106_1572; $L__BB106_1415: mov.u32 %r2102, 36; { // callseq 224, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2102; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 224 bra.uni $L__BB106_1367; $L__BB106_1416: setp.gt.s32 %p1278, %r2056, 68; @%p1278 bra $L__BB106_1471; bra.uni $L__BB106_1417; $L__BB106_1471: setp.gt.s32 %p1279, %r2056, 88; @%p1279 bra $L__BB106_1507; bra.uni $L__BB106_1472; $L__BB106_1507: setp.gt.s32 %p1280, %r2056, 93; @%p1280 bra $L__BB106_1512; bra.uni $L__BB106_1508; $L__BB106_1512: setp.gt.s32 %p1281, %r2056, 95; @%p1281 bra $L__BB106_1516; bra.uni $L__BB106_1513; $L__BB106_1516: setp.eq.s32 %p1282, %r2056, 98; @%p1282 bra $L__BB106_1571; setp.eq.s32 %p1283, %r2056, 97; @%p1283 bra $L__BB106_1521; setp.ne.s32 %p1284, %r2056, 96; @%p1284 bra $L__BB106_1572; bra.uni $L__BB106_1519; $L__BB106_1417: setp.gt.s32 %p1317, %r2056, 48; @%p1317 bra $L__BB106_1436; bra.uni $L__BB106_1418; $L__BB106_1436: setp.gt.s32 %p1318, %r2056, 58; @%p1318 bra $L__BB106_1454; setp.gt.s32 %p1332, %r2056, 53; @%p1332 bra $L__BB106_1446; setp.gt.s32 %p1339, %r2056, 50; @%p1339 bra $L__BB106_1442; bra.uni $L__BB106_1439; $L__BB106_1442: setp.eq.s32 %p1340, %r2056, 51; @%p1340 bra $L__BB106_1550; setp.eq.s32 %p1341, %r2056, 52; @%p1341 bra $L__BB106_1549; setp.eq.s32 %p1342, %r2056, 53; @%p1342 bra $L__BB106_1445; bra.uni $L__BB106_1572; $L__BB106_1445: { // callseq 209, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 209 bra.uni $L__BB106_1367; $L__BB106_1369: setp.gt.s32 %p1363, %r2056, 17; @%p1363 bra $L__BB106_1403; bra.uni $L__BB106_1370; $L__BB106_1403: setp.lt.s32 %p1364, %r2056, 21; @%p1364 bra $L__BB106_1367; setp.gt.s32 %p1365, %r2056, 22; @%p1365 bra $L__BB106_1408; bra.uni $L__BB106_1405; $L__BB106_1408: setp.eq.s32 %p1366, %r2056, 23; @%p1366 bra $L__BB106_1560; setp.eq.s32 %p1367, %r2056, 24; @%p1367 bra $L__BB106_1559; setp.eq.s32 %p1368, %r2056, 25; @%p1368 bra $L__BB106_1411; bra.uni $L__BB106_1572; $L__BB106_1411: add.s32 %r2104, %r633, 1; { // callseq 226, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2104; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 226 bra.uni $L__BB106_1367; $L__BB106_1472: setp.gt.s32 %p1290, %r2056, 78; @%p1290 bra $L__BB106_1490; setp.gt.s32 %p1304, %r2056, 73; @%p1304 bra $L__BB106_1482; setp.gt.s32 %p1311, %r2056, 70; @%p1311 bra $L__BB106_1478; bra.uni $L__BB106_1475; $L__BB106_1478: setp.eq.s32 %p1312, %r2056, 71; @%p1312 bra $L__BB106_1538; setp.eq.s32 %p1313, %r2056, 72; @%p1313 bra $L__BB106_1537; setp.eq.s32 %p1314, %r2056, 73; @%p1314 bra $L__BB106_1481; bra.uni $L__BB106_1572; $L__BB106_1481: add.s32 %r2074, %r633, 1; { // callseq 189, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2074; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 189 bra.uni $L__BB106_1367; $L__BB106_1418: setp.gt.s32 %p1345, %r2056, 41; @%p1345 bra $L__BB106_1424; setp.gt.s32 %p1355, %r2056, 39; @%p1355 bra $L__BB106_1367; setp.eq.s32 %p1356, %r2056, 37; @%p1356 bra $L__BB106_1557; setp.eq.s32 %p1357, %r2056, 38; @%p1357 bra $L__BB106_1556; setp.eq.s32 %p1358, %r2056, 39; @%p1358 bra $L__BB106_1423; bra.uni $L__BB106_1572; $L__BB106_1423: mov.u32 %r2100, 39; { // callseq 221, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2100, %r2057, %r2058, %r2059}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 221 bra.uni $L__BB106_1367; $L__BB106_1370: setp.gt.s32 %p1371, %r2056, 8; @%p1371 bra $L__BB106_1386; bra.uni $L__BB106_1371; $L__BB106_1386: setp.gt.s32 %p1372, %r2056, 12; @%p1372 bra $L__BB106_1395; bra.uni $L__BB106_1387; $L__BB106_1395: setp.gt.s32 %p1373, %r2056, 14; @%p1373 bra $L__BB106_1399; bra.uni $L__BB106_1396; $L__BB106_1399: setp.eq.s32 %p1374, %r2056, 15; @%p1374 bra $L__BB106_1563; setp.eq.s32 %p1375, %r2056, 16; @%p1375 bra $L__BB106_1562; setp.eq.s32 %p1376, %r2056, 17; @%p1376 bra $L__BB106_1402; bra.uni $L__BB106_1572; $L__BB106_1402: { // callseq 231, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 231 bra.uni $L__BB106_1367; $L__BB106_1454: setp.gt.s32 %p1319, %r2056, 63; @%p1319 bra $L__BB106_1463; setp.gt.s32 %p1326, %r2056, 60; @%p1326 bra $L__BB106_1459; bra.uni $L__BB106_1456; $L__BB106_1459: setp.eq.s32 %p1327, %r2056, 61; @%p1327 bra $L__BB106_1544; setp.eq.s32 %p1328, %r2056, 62; @%p1328 bra $L__BB106_1543; setp.eq.s32 %p1329, %r2056, 63; @%p1329 bra $L__BB106_1462; bra.uni $L__BB106_1572; $L__BB106_1462: mov.u32 %r2084, 63; { // callseq 199, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2084; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 199 bra.uni $L__BB106_1367; $L__BB106_1490: setp.gt.s32 %p1291, %r2056, 83; @%p1291 bra $L__BB106_1499; setp.gt.s32 %p1298, %r2056, 80; @%p1298 bra $L__BB106_1495; bra.uni $L__BB106_1492; $L__BB106_1495: setp.eq.s32 %p1299, %r2056, 81; @%p1299 bra $L__BB106_1526; setp.eq.s32 %p1300, %r2056, 82; @%p1300 bra $L__BB106_1525; setp.eq.s32 %p1301, %r2056, 83; @%p1301 bra $L__BB106_1498; bra.uni $L__BB106_1572; $L__BB106_1498: { // callseq 181, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 181 bra.uni $L__BB106_1367; $L__BB106_1371: setp.gt.s32 %p1384, %r2056, 4; @%p1384 bra $L__BB106_1379; setp.gt.s32 %p1390, %r2056, 2; @%p1390 bra $L__BB106_1376; setp.eq.s32 %p1393, %r2056, 1; mov.u32 %r3229, %r2057; @%p1393 bra $L__BB106_1367; setp.eq.s32 %p1394, %r2056, 2; @%p1394 bra $L__BB106_1375; bra.uni $L__BB106_1572; $L__BB106_1375: mov.u32 %r2128, 2; add.s32 %r2129, %r633, 1; { // callseq 240, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r2128; st.param.b32 [param0+4], %r2057; st.param.b32 [param0+8], %r2058; st.param.b32 [param0+12], %r2059; .param .b32 param1; st.param.b32 [param1+0], %r2129; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 240 bra.uni $L__BB106_1367; $L__BB106_1424: setp.gt.s32 %p1346, %r2056, 44; @%p1346 bra $L__BB106_1429; bra.uni $L__BB106_1425; $L__BB106_1429: setp.gt.s32 %p1347, %r2056, 46; @%p1347 bra $L__BB106_1433; setp.eq.s32 %p1350, %r2056, 45; @%p1350 bra $L__BB106_1553; setp.eq.s32 %p1351, %r2056, 46; @%p1351 bra $L__BB106_1432; bra.uni $L__BB106_1572; $L__BB106_1432: { // callseq 216, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 216 bra.uni $L__BB106_1367; $L__BB106_1513: setp.eq.s32 %p1285, %r2056, 94; @%p1285 bra $L__BB106_1520; setp.eq.s32 %p1286, %r2056, 95; @%p1286 bra $L__BB106_1515; bra.uni $L__BB106_1572; $L__BB106_1515: { // callseq 172, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 172 bra.uni $L__BB106_1367; $L__BB106_1508: add.s32 %r2060, %r2056, -92; setp.lt.u32 %p1287, %r2060, 2; @%p1287 bra $L__BB106_1367; setp.eq.s32 %p1288, %r2056, 90; @%p1288 bra $L__BB106_1367; setp.eq.s32 %p1289, %r2056, 89; @%p1289 bra $L__BB106_1511; bra.uni $L__BB106_1572; $L__BB106_1511: mov.u32 %r2064, 89; { // callseq 175, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2064; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 175 bra.uni $L__BB106_1367; $L__BB106_1446: setp.gt.s32 %p1333, %r2056, 55; @%p1333 bra $L__BB106_1450; bra.uni $L__BB106_1447; $L__BB106_1450: setp.eq.s32 %p1334, %r2056, 56; @%p1334 bra $L__BB106_1547; setp.eq.s32 %p1335, %r2056, 57; @%p1335 bra $L__BB106_1546; setp.eq.s32 %p1336, %r2056, 58; @%p1336 bra $L__BB106_1453; bra.uni $L__BB106_1572; $L__BB106_1453: mov.u32 %r2091, 58; { // callseq 204, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2091; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 204 bra.uni $L__BB106_1367; $L__BB106_1482: setp.gt.s32 %p1305, %r2056, 75; @%p1305 bra $L__BB106_1486; bra.uni $L__BB106_1483; $L__BB106_1486: setp.eq.s32 %p1306, %r2056, 76; @%p1306 bra $L__BB106_1530; setp.eq.s32 %p1307, %r2056, 77; @%p1307 bra $L__BB106_1528; setp.eq.s32 %p1308, %r2056, 78; @%p1308 bra $L__BB106_1489; bra.uni $L__BB106_1572; $L__BB106_1489: add.s32 %r2072, %r633, 1; { // callseq 186, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2072; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 186 bra.uni $L__BB106_1367; $L__BB106_1463: setp.gt.s32 %p1320, %r2056, 65; @%p1320 bra $L__BB106_1467; bra.uni $L__BB106_1464; $L__BB106_1467: setp.eq.s32 %p1321, %r2056, 66; @%p1321 bra $L__BB106_1541; setp.eq.s32 %p1322, %r2056, 67; @%p1322 bra $L__BB106_1540; setp.eq.s32 %p1323, %r2056, 68; @%p1323 bra $L__BB106_1470; bra.uni $L__BB106_1572; $L__BB106_1470: mov.u32 %r2079, 68; { // callseq 194, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2079; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 194 bra.uni $L__BB106_1367; $L__BB106_1499: setp.gt.s32 %p1292, %r2056, 85; @%p1292 bra $L__BB106_1503; bra.uni $L__BB106_1500; $L__BB106_1503: setp.eq.s32 %p1293, %r2056, 86; @%p1293 bra $L__BB106_1523; setp.eq.s32 %p1294, %r2056, 87; @%p1294 bra $L__BB106_1522; setp.eq.s32 %p1295, %r2056, 88; @%p1295 bra $L__BB106_1506; bra.uni $L__BB106_1572; $L__BB106_1506: mov.u32 %r2065, 88; cvta.to.local.u64 %rd1938, %rd1729; ld.local.f32 %f10188, [%rd1938+12]; ld.local.v2.f32 {%f10189, %f10190}, [%rd1938+16]; { // callseq 176, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f10188; .param .b32 param1; st.param.f32 [param1+0], %f10189; .param .b32 param2; st.param.f32 [param2+0], %f10190; .param .b64 param3; st.param.b64 [param3+0], %rd1928; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r2065; st.param.b32 [param4+4], %r2057; st.param.b32 [param4+8], %r2058; st.param.b32 [param4+12], %r2059; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 176 bra.uni $L__BB106_1367; $L__BB106_1387: setp.gt.s32 %p1379, %r2056, 10; @%p1379 bra $L__BB106_1392; setp.eq.s32 %p1382, %r2056, 9; @%p1382 bra $L__BB106_1566; setp.eq.s32 %p1383, %r2056, 10; @%p1383 bra $L__BB106_1390; bra.uni $L__BB106_1572; $L__BB106_1390: ld.local.f32 %f10204, [%rd318]; setp.ltu.ftz.f32 %p1399, %f10204, 0f3F800000; @%p1399 bra $L__BB106_1367; add.s32 %r2122, %r2057, %r633; add.s32 %r3229, %r2122, 1; bra.uni $L__BB106_1367; $L__BB106_1379: setp.gt.s32 %p1385, %r2056, 6; @%p1385 bra $L__BB106_1383; setp.eq.s32 %p1388, %r2056, 5; @%p1388 bra $L__BB106_1569; setp.eq.s32 %p1389, %r2056, 6; @%p1389 bra $L__BB106_1382; bra.uni $L__BB106_1572; $L__BB106_1382: ld.local.f32 %f10206, [%rd317]; ld.local.f32 %f10207, [%rd317+4]; ld.local.f32 %f10208, [%rd317+8]; st.local.f32 [%rd312], %f10206; st.local.f32 [%rd313], %f10207; st.local.f32 [%rd314], %f10208; bra.uni $L__BB106_1367; $L__BB106_1425: setp.eq.s32 %p1352, %r2056, 42; @%p1352 bra $L__BB106_1555; setp.eq.s32 %p1353, %r2056, 43; @%p1353 bra $L__BB106_1554; setp.eq.s32 %p1354, %r2056, 44; @%p1354 bra $L__BB106_1428; bra.uni $L__BB106_1572; $L__BB106_1428: mov.u32 %r2097, 44; add.s32 %r2098, %r633, 1; { // callseq 218, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2097; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2098; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 218 bra.uni $L__BB106_1367; $L__BB106_1571: { // callseq 241, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 241 bra.uni $L__BB106_1367; $L__BB106_1521: { // callseq 174, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 174 bra.uni $L__BB106_1367; $L__BB106_1520: mov.u32 %r2062, 94; add.s32 %r2063, %r633, 1; { // callseq 173, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2062; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2063; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 173 bra.uni $L__BB106_1367; $L__BB106_1405: setp.eq.s32 %p1369, %r2056, 21; @%p1369 bra $L__BB106_1561; setp.eq.s32 %p1370, %r2056, 22; @%p1370 bra $L__BB106_1407; bra.uni $L__BB106_1572; $L__BB106_1407: mov.u32 %r2108, 22; add.s32 %r2109, %r633, 1; { // callseq 229, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2108; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2109; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 229 bra.uni $L__BB106_1367; $L__BB106_1396: setp.eq.s32 %p1377, %r2056, 13; @%p1377 bra $L__BB106_1564; setp.eq.s32 %p1378, %r2056, 14; @%p1378 bra $L__BB106_1398; bra.uni $L__BB106_1572; $L__BB106_1398: st.local.u32 [%rd318], %r2057; bra.uni $L__BB106_1367; $L__BB106_1376: setp.eq.s32 %p1391, %r2056, 3; @%p1391 bra $L__BB106_1570; setp.eq.s32 %p1392, %r2056, 4; @%p1392 bra $L__BB106_1378; bra.uni $L__BB106_1572; $L__BB106_1378: mov.u32 %r2126, 4; { // callseq 238, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2126, %r2057, %r2058, %r2059}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 238 bra.uni $L__BB106_1367; $L__BB106_1392: setp.eq.s32 %p1380, %r2056, 11; @%p1380 bra $L__BB106_1565; setp.eq.s32 %p1381, %r2056, 12; @%p1381 bra $L__BB106_1394; bra.uni $L__BB106_1572; $L__BB106_1394: { // callseq 234, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 234 bra.uni $L__BB106_1367; $L__BB106_1383: setp.eq.s32 %p1386, %r2056, 7; @%p1386 bra $L__BB106_1568; setp.eq.s32 %p1387, %r2056, 8; @%p1387 bra $L__BB106_1385; bra.uni $L__BB106_1572; $L__BB106_1385: mov.u32 %r2124, 8; { // callseq 236, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2124; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 236 bra.uni $L__BB106_1367; $L__BB106_1439: setp.eq.s32 %p1343, %r2056, 49; @%p1343 bra $L__BB106_1551; setp.eq.s32 %p1344, %r2056, 50; @%p1344 bra $L__BB106_1441; bra.uni $L__BB106_1572; $L__BB106_1441: { // callseq 212, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 212 bra.uni $L__BB106_1367; $L__BB106_1475: setp.eq.s32 %p1315, %r2056, 69; @%p1315 bra $L__BB106_1539; setp.eq.s32 %p1316, %r2056, 70; @%p1316 bra $L__BB106_1477; bra.uni $L__BB106_1572; $L__BB106_1477: mov.u32 %r2076, 70; { // callseq 192, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2076; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 192 bra.uni $L__BB106_1367; $L__BB106_1456: setp.eq.s32 %p1330, %r2056, 59; @%p1330 bra $L__BB106_1545; setp.eq.s32 %p1331, %r2056, 60; @%p1331 bra $L__BB106_1458; bra.uni $L__BB106_1572; $L__BB106_1458: add.s32 %r2089, %r633, 1; { // callseq 202, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2089; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 202 bra.uni $L__BB106_1367; $L__BB106_1492: setp.eq.s32 %p1302, %r2056, 79; @%p1302 bra $L__BB106_1527; setp.eq.s32 %p1303, %r2056, 80; @%p1303 bra $L__BB106_1494; bra.uni $L__BB106_1572; $L__BB106_1494: { // callseq 184, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 184 bra.uni $L__BB106_1367; $L__BB106_1433: setp.eq.s32 %p1348, %r2056, 47; @%p1348 bra $L__BB106_1552; setp.eq.s32 %p1349, %r2056, 48; @%p1349 bra $L__BB106_1435; bra.uni $L__BB106_1572; $L__BB106_1435: { // callseq 214, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 214 bra.uni $L__BB106_1367; $L__BB106_1447: setp.eq.s32 %p1337, %r2056, 54; @%p1337 bra $L__BB106_1548; setp.eq.s32 %p1338, %r2056, 55; @%p1338 bra $L__BB106_1449; bra.uni $L__BB106_1572; $L__BB106_1449: cvta.to.local.u64 %rd1983, %rd1729; ld.local.v4.f32 {%f10197, %f10198, %f10199, %f10200}, [%rd1983]; { // callseq 207, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f10197; st.param.f32 [param0+4], %f10198; st.param.f32 [param0+8], %f10199; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 207 bra.uni $L__BB106_1367; $L__BB106_1483: setp.eq.s32 %p1309, %r2056, 74; @%p1309 bra $L__BB106_1536; setp.eq.s32 %p1310, %r2056, 75; @%p1310 bra $L__BB106_1485; bra.uni $L__BB106_1572; $L__BB106_1485: { // callseq 187, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 187 bra.uni $L__BB106_1367; $L__BB106_1464: setp.eq.s32 %p1324, %r2056, 64; @%p1324 bra $L__BB106_1542; setp.eq.s32 %p1325, %r2056, 65; @%p1325 bra $L__BB106_1466; bra.uni $L__BB106_1572; $L__BB106_1466: { // callseq 197, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 197 bra.uni $L__BB106_1367; $L__BB106_1500: setp.eq.s32 %p1296, %r2056, 84; @%p1296 bra $L__BB106_1524; setp.eq.s32 %p1297, %r2056, 85; @%p1297 bra $L__BB106_1502; bra.uni $L__BB106_1572; $L__BB106_1502: add.s32 %r2068, %r633, 1; { // callseq 179, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2068; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 179 bra.uni $L__BB106_1367; $L__BB106_1560: mov.u32 %r2106, 23; add.s32 %r2107, %r633, 1; { // callseq 228, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2106; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2107; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 228 bra.uni $L__BB106_1367; $L__BB106_1559: mov.u32 %r2105, 24; { // callseq 227, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2105; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 227 bra.uni $L__BB106_1367; $L__BB106_1563: ld.global.v4.u32 {%r2112, %r2113, %r2114, %r2115}, [%rd316+16]; add.s32 %r3229, %r633, 2; st.local.u32 [%rd317], %r2113; st.local.u32 [%rd317+4], %r2114; st.local.u32 [%rd317+8], %r2115; bra.uni $L__BB106_1367; $L__BB106_1562: mov.u32 %r2111, 16; { // callseq 232, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2111; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 232 bra.uni $L__BB106_1367; $L__BB106_1566: ld.local.f32 %f10205, [%rd318]; setp.gtu.ftz.f32 %p1400, %f10205, 0f00000000; @%p1400 bra $L__BB106_1367; add.s32 %r2123, %r2057, %r633; add.s32 %r3229, %r2123, 1; bra.uni $L__BB106_1367; $L__BB106_1569: st.local.u32 [%rd312], %r2057; st.local.u32 [%rd313], %r2058; st.local.u32 [%rd314], %r2059; bra.uni $L__BB106_1367; $L__BB106_1561: mov.u32 %r2110, 21; { // callseq 230, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2110; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 230 bra.uni $L__BB106_1367; $L__BB106_1564: mov.u32 %r2119, 13; add.s32 %r2120, %r633, 1; mov.u32 %r2121, 524288; { // callseq 233, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b32 param1; st.param.b32 [param1+0], %r2121; .param .b64 param2; st.param.b64 [param2+0], %rd1928; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r2119; st.param.b32 [param3+4], %r2057; st.param.b32 [param3+8], %r2058; st.param.b32 [param3+12], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2120; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 233 bra.uni $L__BB106_1367; $L__BB106_1570: mov.u32 %r2127, 3; { // callseq 239, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2127, %r2057, %r2058, %r2059}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 239 bra.uni $L__BB106_1367; $L__BB106_1565: { // callseq 235, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 235 bra.uni $L__BB106_1367; $L__BB106_1568: mov.u32 %r2125, 7; { // callseq 237, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2125; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 237 bra.uni $L__BB106_1367; $L__BB106_1550: { // callseq 211, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 211 bra.uni $L__BB106_1367; $L__BB106_1549: add.s32 %r2096, %r633, 1; { // callseq 210, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2096; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 210 bra.uni $L__BB106_1367; $L__BB106_1538: mov.u32 %r2075, 71; { // callseq 191, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2075; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 191 bra.uni $L__BB106_1367; $L__BB106_1537: { // callseq 190, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 190 bra.uni $L__BB106_1367; $L__BB106_1557: mov.u32 %r2101, 37; { // callseq 223, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2101, %r2057, %r2058, %r2059}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 223 bra.uni $L__BB106_1367; $L__BB106_1556: { // callseq 222, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 222 bra.uni $L__BB106_1367; $L__BB106_1544: mov.u32 %r2087, 61; add.s32 %r2088, %r633, 1; { // callseq 201, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2087; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2088; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 201 bra.uni $L__BB106_1367; $L__BB106_1543: mov.u32 %r2085, 62; add.s32 %r2086, %r633, 1; { // callseq 200, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2085; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2086; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 200 bra.uni $L__BB106_1367; $L__BB106_1526: mov.u32 %r2070, 81; { // callseq 183, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2070; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 183 bra.uni $L__BB106_1367; $L__BB106_1525: mov.u32 %r2069, 82; { // callseq 182, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2069; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 182 bra.uni $L__BB106_1367; $L__BB106_1553: { // callseq 217, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 217 bra.uni $L__BB106_1367; $L__BB106_1547: mov.u32 %r2094, 56; { // callseq 206, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2094; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 206 bra.uni $L__BB106_1367; $L__BB106_1546: mov.u32 %r2092, 57; add.s32 %r2093, %r633, 1; { // callseq 205, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2092; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2093; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 205 bra.uni $L__BB106_1367; $L__BB106_1530: setp.eq.s32 %p1396, %r2059, 255; @%p1396 bra $L__BB106_1367; add.s64 %rd319, %rd317, 4; add.s64 %rd320, %rd317, 8; setp.eq.s32 %p1397, %r2058, 0; mul.wide.u32 %rd1958, %r2059, 4; add.s64 %rd321, %rd315, %rd1958; @%p1397 bra $L__BB106_1535; setp.eq.s32 %p1398, %r2058, 1; @%p1398 bra $L__BB106_1534; bra.uni $L__BB106_1533; $L__BB106_1534: ld.local.f32 %f10195, [%rd319]; st.local.f32 [%rd321], %f10195; bra.uni $L__BB106_1367; $L__BB106_1528: setp.eq.s32 %p1395, %r2059, 255; @%p1395 bra $L__BB106_1367; ld.local.f32 %f10193, [%rd317]; add.s32 %r2073, %r2059, %r2058; mul.wide.u32 %rd1954, %r2073, 4; add.s64 %rd1955, %rd315, %rd1954; st.local.f32 [%rd1955], %f10193; bra.uni $L__BB106_1367; $L__BB106_1541: add.s32 %r2081, %r633, 1; { // callseq 196, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2081; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 196 bra.uni $L__BB106_1367; $L__BB106_1540: mov.u32 %r2080, 67; { // callseq 195, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2080, %r2057, %r2058, %r2059}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 195 bra.uni $L__BB106_1367; $L__BB106_1523: add.s32 %r2067, %r633, 1; { // callseq 178, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2067; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 178 bra.uni $L__BB106_1367; $L__BB106_1522: add.s32 %r2066, %r633, 1; { // callseq 177, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2066; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 177 bra.uni $L__BB106_1367; $L__BB106_1555: { // callseq 220, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 220 bra.uni $L__BB106_1367; $L__BB106_1554: add.s32 %r2099, %r633, 1; { // callseq 219, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2099; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 219 bra.uni $L__BB106_1367; $L__BB106_1551: { // callseq 213, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1729; .param .b64 param1; st.param.b64 [param1+0], %rd1928; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 213 bra.uni $L__BB106_1367; $L__BB106_1539: mov.u32 %r2077, 69; add.s32 %r2078, %r633, 1; { // callseq 193, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2077; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2078; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 193 bra.uni $L__BB106_1367; $L__BB106_1545: add.s32 %r2090, %r633, 1; { // callseq 203, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2090; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 203 bra.uni $L__BB106_1367; $L__BB106_1527: add.s32 %r2071, %r633, 1; { // callseq 185, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2071; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 185 bra.uni $L__BB106_1367; $L__BB106_1552: { // callseq 215, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b64 param1; st.param.b64 [param1+0], %rd1729; .param .b64 param2; st.param.b64 [param2+0], %rd1928; .param .b32 param3; st.param.b32 [param3+0], %r2057; .param .b32 param4; st.param.b32 [param4+0], %r2058; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 215 bra.uni $L__BB106_1367; $L__BB106_1548: add.s32 %r2095, %r633, 1; { // callseq 208, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2095; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 208 bra.uni $L__BB106_1367; $L__BB106_1536: { // callseq 188, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 188 bra.uni $L__BB106_1367; $L__BB106_1542: mov.u32 %r2082, 64; add.s32 %r2083, %r633, 1; { // callseq 198, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2082; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2083; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 198 bra.uni $L__BB106_1367; $L__BB106_1524: { // callseq 180, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 180 bra.uni $L__BB106_1367; $L__BB106_1558: mov.u32 %r2103, 26; { // callseq 225, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1928; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2103; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 225 bra.uni $L__BB106_1367; $L__BB106_1535: ld.local.f32 %f10196, [%rd317]; st.local.f32 [%rd321], %f10196; bra.uni $L__BB106_1367; $L__BB106_1533: ld.local.f32 %f10194, [%rd320]; st.local.f32 [%rd321], %f10194; bra.uni $L__BB106_1367; $L__BB106_1572: add.s64 %rd2035, %rd1729, 52; cvta.to.local.u64 %rd2036, %rd2035; ld.local.u32 %r2130, [%rd2036]; and.b32 %r723, %r2130, 2; @%p1251 bra $L__BB106_1575; bra.uni $L__BB106_1573; $L__BB106_1575: setp.eq.s32 %p1404, %r723, 0; mov.f32 %f16220, 0f00000000; mov.f32 %f16221, %f16220; mov.f32 %f16222, %f16220; @%p1404 bra $L__BB106_1577; add.s64 %rd2047, %rd1729, 280; cvta.to.local.u64 %rd2048, %rd2047; ld.local.v2.f32 {%f16220, %f16221}, [%rd2048]; add.s64 %rd2049, %rd1729, 288; cvta.to.local.u64 %rd2050, %rd2049; ld.local.f32 %f16222, [%rd2050]; bra.uni $L__BB106_1577; $L__BB106_1573: setp.eq.s32 %p1402, %r723, 0; mov.f32 %f16220, 0f00000000; mov.f32 %f16221, %f16220; mov.f32 %f16222, %f16220; @%p1402 bra $L__BB106_1577; add.s64 %rd2037, %rd284, 20; add.s64 %rd2038, %rd284, 12; cvta.to.local.u64 %rd2039, %rd2038; ld.local.v2.f32 {%f10212, %f10213}, [%rd2039]; cvta.to.local.u64 %rd2040, %rd2037; ld.local.v4.f32 {%f10216, %f10217, %f10218, %f10219}, [%rd2040]; mul.ftz.f32 %f10224, %f10213, %f10218; fma.rn.ftz.f32 %f10225, %f10212, %f10217, %f10224; fma.rn.ftz.f32 %f10226, %f10216, %f10219, %f10225; abs.ftz.f32 %f10227, %f10226; setp.gt.ftz.f32 %p1403, %f10227, 0f00000000; selp.f32 %f10228, 0f3F800000, 0f00000000, %p1403; add.s64 %rd2042, %rd1729, 280; cvta.to.local.u64 %rd2043, %rd2042; ld.local.v2.f32 {%f10229, %f10230}, [%rd2043]; add.s64 %rd2044, %rd1729, 288; cvta.to.local.u64 %rd2045, %rd2044; ld.local.f32 %f10233, [%rd2045]; mul.ftz.f32 %f16222, %f10228, %f10233; mul.ftz.f32 %f16221, %f10228, %f10230; mul.ftz.f32 %f16220, %f10229, %f10228; $L__BB106_1577: ld.local.f32 %f10239, [%rd282+28]; mul.ftz.f32 %f16223, %f16220, %f10239; mul.ftz.f32 %f16224, %f16221, %f10239; mul.ftz.f32 %f16225, %f16222, %f10239; ld.local.u32 %r724, [%rd282+44]; setp.eq.s32 %p1405, %r724, -1; @%p1405 bra $L__BB106_1579; ld.const.u64 %rd2051, [kernel_params+2368]; cvta.to.global.u64 %rd2052, %rd2051; mul.wide.s32 %rd2053, %r724, 192; add.s64 %rd2054, %rd2052, %rd2053; ld.global.f32 %f10240, [%rd2054+28]; mul.ftz.f32 %f16223, %f16223, %f10240; ld.global.f32 %f10241, [%rd2054+32]; mul.ftz.f32 %f16224, %f16224, %f10241; ld.global.f32 %f10242, [%rd2054+36]; mul.ftz.f32 %f16225, %f16225, %f10242; $L__BB106_1579: setp.eq.ftz.f32 %p1406, %f16224, 0f00000000; setp.eq.ftz.f32 %p1407, %f16223, 0f00000000; and.pred %p1408, %p1407, %p1406; setp.eq.ftz.f32 %p1409, %f16225, 0f00000000; and.pred %p1410, %p1409, %p1408; @%p1410 bra $L__BB106_1685; add.u64 %rd2921, %SP, 1728; add.u64 %rd2056, %SP, 704; { // callseq 242, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2921; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f1935; st.param.f32 [param1+4], %f1936; st.param.f32 [param1+8], %f1937; .param .b64 param2; st.param.b64 [param2+0], %rd2056; .param .b32 param3; st.param.b32 [param3+0], %r618; .param .b32 retval0; call.uni (retval0), _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj, ( param0, param1, param2, param3 ); ld.param.f32 %f2119, [retval0+0]; } // callseq 242 ld.local.f32 %f2120, [%rd282+24]; div.approx.ftz.f32 %f10243, %f16223, %f2120; div.approx.ftz.f32 %f10244, %f16224, %f2120; div.approx.ftz.f32 %f10245, %f16225, %f2120; add.u64 %rd322, %SPL, 704; ld.local.f32 %f10246, [%rd322]; mul.ftz.f32 %f16239, %f10243, %f10246; ld.local.f32 %f10247, [%rd322+4]; mul.ftz.f32 %f16241, %f10244, %f10247; ld.local.f32 %f10248, [%rd322+8]; mul.ftz.f32 %f16240, %f10245, %f10248; st.local.f32 [%rd322], %f16239; st.local.f32 [%rd322+4], %f16241; st.local.f32 [%rd322+8], %f16240; ld.local.f32 %f10249, [%rd322+12]; mul.ftz.f32 %f16236, %f10243, %f10249; ld.local.f32 %f10250, [%rd322+16]; mul.ftz.f32 %f16238, %f10244, %f10250; ld.local.f32 %f10251, [%rd322+20]; mul.ftz.f32 %f16237, %f10245, %f10251; st.local.f32 [%rd322+12], %f16236; st.local.f32 [%rd322+16], %f16238; st.local.f32 [%rd322+20], %f16237; ld.local.f32 %f10252, [%rd322+24]; mul.ftz.f32 %f16244, %f10243, %f10252; ld.local.f32 %f10253, [%rd322+28]; mul.ftz.f32 %f16243, %f10244, %f10253; ld.local.f32 %f10254, [%rd322+32]; mul.ftz.f32 %f16242, %f10245, %f10254; st.local.f32 [%rd322+24], %f16244; st.local.f32 [%rd322+28], %f16243; st.local.f32 [%rd322+32], %f16242; and.b32 %r2131, %r618, 268435456; setp.eq.s32 %p1411, %r2131, 0; @%p1411 bra $L__BB106_1582; mul.ftz.f32 %f10255, %f2120, %f2120; fma.rn.ftz.f32 %f10256, %f2119, %f2119, %f10255; div.approx.ftz.f32 %f10257, %f10255, %f10256; mul.ftz.f32 %f16239, %f10257, %f16239; mul.ftz.f32 %f16241, %f10257, %f16241; mul.ftz.f32 %f16240, %f10257, %f16240; st.local.f32 [%rd322], %f16239; st.local.f32 [%rd322+4], %f16241; st.local.f32 [%rd322+8], %f16240; mul.ftz.f32 %f16236, %f10257, %f16236; mul.ftz.f32 %f16238, %f10257, %f16238; mul.ftz.f32 %f16237, %f10257, %f16237; st.local.f32 [%rd322+12], %f16236; st.local.f32 [%rd322+16], %f16238; st.local.f32 [%rd322+20], %f16237; mul.ftz.f32 %f16244, %f10257, %f16244; mul.ftz.f32 %f16243, %f10257, %f16243; mul.ftz.f32 %f16242, %f10257, %f16242; st.local.f32 [%rd322+24], %f16244; st.local.f32 [%rd322+28], %f16243; st.local.f32 [%rd322+32], %f16242; $L__BB106_1582: ld.const.f32 %f2148, [kernel_params+1508]; setp.leu.ftz.f32 %p1412, %f2148, 0f00000000; mov.f32 %f16235, 0f00000000; @%p1412 bra $L__BB106_1587; add.s32 %r725, %r602, 2; @%p1240 bra $L__BB106_1586; ld.const.v2.u32 {%r2132, %r2133}, [kernel_params+1600]; mov.b32 %f2149, %r2133; setp.geu.ftz.f32 %p1414, %f2149, 0f3F800000; setp.lt.ftz.f32 %p1415, %f2149, 0f3F800000; ld.const.u32 %r2136, [kernel_params+1576]; selp.b32 %r2137, %r2136, %r601, %p1415; add.s16 %rs88, %rs28, 2; and.b16 %rs89, %rs88, 255; cvt.u32.u16 %r2138, %rs89; xor.b32 %r2139, %r2137, %r2138; mul.lo.s32 %r2140, %r2139, -512718531; shr.u32 %r2141, %r2137, 16; xor.b32 %r2142, %r2140, %r2141; bfe.u32 %r2143, %r2142, 4, 4; shr.u32 %r2144, %r2137, 8; xor.b32 %r2145, %r2142, %r2144; xor.b32 %r2146, %r2145, %r2143; mul.lo.s32 %r2147, %r2146, 153742143; shr.u32 %r2148, %r2137, 23; xor.b32 %r2149, %r2147, %r2148; bfe.u32 %r2150, %r2149, 1, 7; xor.b32 %r2151, %r2150, %r2149; shr.u32 %r2152, %r2137, 27; or.b32 %r2153, %r2152, 1; mul.lo.s32 %r2154, %r2153, %r2151; mul.lo.s32 %r2155, %r2154, -1074439621; bfe.u32 %r2156, %r2155, 2, 6; xor.b32 %r2157, %r2156, %r2155; mul.lo.s32 %r2158, %r2157, -1638916925; bfe.u32 %r2159, %r2158, 2, 6; xor.b32 %r2160, %r2159, %r2158; mul.lo.s32 %r2161, %r2160, -933190689; and.b32 %r2162, %r2161, 255; bfe.u32 %r2163, %r2161, 5, 3; xor.b32 %r2164, %r2163, %r2162; add.s32 %r2165, %r2132, -1; xor.b32 %r726, %r725, 61; xor.b32 %r2166, %r2137, %r726; mul.lo.s32 %r2167, %r2166, 9; shr.u32 %r2168, %r2167, 4; xor.b32 %r2169, %r2168, %r2167; mul.lo.s32 %r2170, %r2169, 668265261; add.s32 %r2171, %r2170, %r607; shr.u32 %r2172, %r2170, 16; or.b32 %r2173, %r2172, 1; mul.lo.s32 %r2174, %r2173, %r2171; mul.lo.s32 %r2175, %r2174, 89287766; xor.b32 %r2176, %r2175, %r2174; mul.lo.s32 %r2177, %r2176, 1403136100; xor.b32 %r2178, %r2177, %r2176; brev.b32 %r2179, %r2178; neg.s32 %r2180, %r2132; and.b32 %r2181, %r603, %r2180; and.b32 %r2182, %r2179, %r2165; or.b32 %r2183, %r2182, %r2181; mad.lo.s32 %r2184, %r2164, %r2132, %r2183; shl.b32 %r2185, %r2132, 8; rem.u32 %r2186, %r2184, %r2185; ld.const.u64 %rd2057, [kernel_params+2424]; cvta.to.global.u64 %rd2058, %rd2057; shl.b32 %r2187, %r2186, 1; mul.wide.u32 %rd2059, %r2187, 4; add.s64 %rd2060, %rd2058, %rd2059; ld.global.f32 %f16235, [%rd2060]; @%p1414 bra $L__BB106_1587; xor.b32 %r2188, %r726, %r601; mul.lo.s32 %r2189, %r2188, 9; shr.u32 %r2190, %r2189, 4; xor.b32 %r2191, %r2190, %r2189; mul.lo.s32 %r2192, %r2191, 668265261; cvt.rn.f32.u32 %f10259, %r2192; mul.ftz.f32 %f10260, %f10259, 0f2F7FFFFE; fma.rn.ftz.f32 %f10261, %f10260, %f2149, %f16235; cvt.rmi.ftz.f32.f32 %f10262, %f10261; sub.ftz.f32 %f16235, %f10261, %f10262; bra.uni $L__BB106_1587; $L__BB106_1586: shr.u32 %r2193, %r725, 16; xor.b32 %r2194, %r2193, %r725; mul.lo.s32 %r2195, %r2194, 569420461; shr.u32 %r2196, %r2195, 15; xor.b32 %r2197, %r2196, %r2195; mul.lo.s32 %r2198, %r2197, -749064809; shr.u32 %r2199, %r2198, 15; xor.b32 %r2200, %r601, %r2198; xor.b32 %r2201, %r2200, %r2199; xor.b32 %r2202, %r2201, 1493655573; add.s32 %r2203, %r2202, %r607; shr.u32 %r2204, %r2202, 16; or.b32 %r2205, %r2204, 1; mul.lo.s32 %r2206, %r2205, %r2203; mul.lo.s32 %r2207, %r2206, 89287766; xor.b32 %r2208, %r2207, %r2206; mul.lo.s32 %r2209, %r2208, 1403136100; xor.b32 %r2210, %r2209, %r2208; xor.b32 %r2211, %r2201, -2052961194; brev.b32 %r2212, %r2210; mul.lo.s32 %r2213, %r2212, 1025551850; xor.b32 %r2214, %r2213, %r2212; add.s32 %r2215, %r2214, %r2211; shr.u32 %r2216, %r2211, 16; or.b32 %r2217, %r2216, 1; mul.lo.s32 %r2218, %r2215, %r2217; mul.lo.s32 %r2219, %r2218, 89287766; xor.b32 %r2220, %r2219, %r2218; mul.lo.s32 %r2221, %r2220, 1403136100; xor.b32 %r2222, %r2221, %r2220; brev.b32 %r2223, %r2222; cvt.rn.f32.u32 %f10263, %r2223; mul.ftz.f32 %f16235, %f10263, 0f2F7FFFFE; $L__BB106_1587: setp.eq.ftz.f32 %p1416, %f16243, 0f00000000; setp.eq.ftz.f32 %p1417, %f16244, 0f00000000; and.pred %p1418, %p1417, %p1416; setp.eq.ftz.f32 %p1419, %f16242, 0f00000000; and.pred %p1420, %p1419, %p1418; @%p1420 bra $L__BB106_1685; @%p1412 bra $L__BB106_1592; abs.ftz.f32 %f10264, %f16244; abs.ftz.f32 %f10265, %f16243; max.ftz.f32 %f10266, %f10264, %f10265; abs.ftz.f32 %f10267, %f16242; max.ftz.f32 %f10268, %f10266, %f10267; mul.ftz.f32 %f2154, %f10268, %f2148; setp.geu.ftz.f32 %p1422, %f2154, 0f3F800000; @%p1422 bra $L__BB106_1592; setp.le.ftz.f32 %p1423, %f2154, %f16235; @%p1423 bra $L__BB106_1685; rcp.approx.ftz.f32 %f10269, %f2154; mul.ftz.f32 %f16239, %f10269, %f16239; mul.ftz.f32 %f16241, %f10269, %f16241; mul.ftz.f32 %f16240, %f10269, %f16240; st.local.f32 [%rd322], %f16239; st.local.f32 [%rd322+4], %f16241; st.local.f32 [%rd322+8], %f16240; mul.ftz.f32 %f16236, %f10269, %f16236; mul.ftz.f32 %f16238, %f10269, %f16238; mul.ftz.f32 %f16237, %f10269, %f16237; st.local.f32 [%rd322+12], %f16236; st.local.f32 [%rd322+16], %f16238; st.local.f32 [%rd322+20], %f16237; mul.ftz.f32 %f16244, %f10269, %f16244; mul.ftz.f32 %f16243, %f10269, %f16243; mul.ftz.f32 %f16242, %f10269, %f16242; st.local.f32 [%rd322+24], %f16244; st.local.f32 [%rd322+28], %f16243; st.local.f32 [%rd322+32], %f16242; $L__BB106_1592: ld.local.v4.f32 {%f16270, %f16271, %f16272, %f10273}, [%rd1]; ld.local.u32 %r727, [%rd7+-24]; and.b32 %r728, %r727, 1; setp.eq.s32 %p1424, %r728, 0; ld.local.u32 %r729, [%rd7+-40]; setp.gt.s32 %p1425, %r729, -1; mov.u16 %rs149, 1; or.pred %p1426, %p1424, %p1425; @%p1426 bra $L__BB106_1635; ld.const.u64 %rd2061, [kernel_params+2160]; cvta.to.global.u64 %rd2062, %rd2061; ld.local.u32 %r2224, [%rd7+-12]; mul.wide.s32 %rd2063, %r2224, 208; add.s64 %rd2064, %rd2062, %rd2063; add.s64 %rd323, %rd2064, 184; ld.global.f32 %f2176, [%rd2064+184]; setp.leu.ftz.f32 %p1427, %f2176, 0f00000000; @%p1427 bra $L__BB106_1635; ld.local.f32 %f10274, [%rd7+-76]; ld.local.v2.f32 {%f10275, %f10276}, [%rd7+-72]; mul.ftz.f32 %f10279, %f1936, %f10275; fma.rn.ftz.f32 %f10280, %f1935, %f10274, %f10279; fma.rn.ftz.f32 %f10281, %f1937, %f10276, %f10280; setp.lt.ftz.f32 %p1428, %f10281, 0f00000000; neg.ftz.f32 %f10282, %f10281; selp.f32 %f2177, %f10282, %f10281, %p1428; @%p1428 bra $L__BB106_1596; bra.uni $L__BB106_1595; $L__BB106_1596: ld.local.v2.f32 {%f10285, %f10286}, [%rd7+-64]; neg.ftz.f32 %f16245, %f10285; neg.ftz.f32 %f16246, %f10286; ld.local.f32 %f10289, [%rd7+-56]; neg.ftz.f32 %f16247, %f10289; bra.uni $L__BB106_1597; $L__BB106_1595: ld.local.v2.f32 {%f16245, %f16246}, [%rd7+-64]; ld.local.f32 %f16247, [%rd7+-56]; $L__BB106_1597: mul.ftz.f32 %f10290, %f1936, %f16246; fma.rn.ftz.f32 %f10291, %f1935, %f16245, %f10290; fma.rn.ftz.f32 %f2187, %f1937, %f16247, %f10291; add.ftz.f32 %f10292, %f2177, %f2187; setp.lt.ftz.f32 %p1429, %f2177, %f2176; selp.f32 %f10293, 0f40000000, 0f3F800000, %p1429; mov.f32 %f10294, 0f3F800000; selp.f32 %f10295, %f10292, %f2187, %p1429; div.approx.ftz.f32 %f10296, %f10295, %f2176; sub.ftz.f32 %f10297, %f10293, %f10296; mov.f32 %f10298, 0f00000000; max.ftz.f32 %f10299, %f10297, %f10298; min.ftz.f32 %f2188, %f10299, %f10294; setp.leu.ftz.f32 %p1430, %f2188, 0f00000000; @%p1430 bra $L__BB106_1635; setp.eq.s32 %p1431, %r727, 65; ld.local.s32 %rd324, [%rd7+-28]; @%p1431 bra $L__BB106_1600; bra.uni $L__BB106_1599; $L__BB106_1600: ld.global.u32 %r730, [%rd323+-28]; ld.global.u32 %r731, [%rd323+-32]; shl.b32 %r2236, %r731, 1; cvt.rn.f32.s32 %f10300, %r2236; ld.local.f32 %f10301, [%rd7+-4]; mul.ftz.f32 %f10302, %f10301, %f10300; cvt.rzi.ftz.s32.f32 %r2237, %f10302; add.s32 %r2238, %r2236, -1; min.s32 %r732, %r2237, %r2238; cvt.rn.f32.s32 %f10303, %r732; sub.ftz.f32 %f2207, %f10302, %f10303; ld.const.u64 %rd2087, [kernel_params+2312]; cvta.to.global.u64 %rd325, %rd2087; ld.global.u32 %r3236, [%rd323+-20]; mul.wide.u32 %rd2088, %r3236, 16; add.s64 %rd2089, %rd325, %rd2088; ld.global.u16 %rs147, [%rd2089+12]; ld.global.u32 %r3235, [%rd2089+8]; ld.global.u64 %rd3004, [%rd2089]; setp.eq.s64 %p1432, %rd3004, 11; mov.u16 %rs145, %rs147; mov.u32 %r3230, %r3235; @%p1432 bra $L__BB106_1606; mov.u64 %rd2985, %rd3004; mov.u32 %r3230, %r3235; mov.u16 %rs145, %rs147; mov.u32 %r3231, %r3236; $L__BB106_1602: setp.eq.s64 %p1433, %rd2985, 0; @%p1433 bra $L__BB106_1604; add.s32 %r3231, %r3231, 2; bra.uni $L__BB106_1605; $L__BB106_1604: setp.eq.s16 %p1434, %rs145, 0; mov.u32 %r3234, -1; mov.u32 %r3231, %r3230; @%p1434 bra $L__BB106_1607; $L__BB106_1605: mul.wide.u32 %rd2090, %r3231, 16; add.s64 %rd2091, %rd325, %rd2090; ld.global.u16 %rs145, [%rd2091+12]; ld.global.u32 %r3230, [%rd2091+8]; ld.global.u64 %rd2985, [%rd2091]; setp.ne.s64 %p1435, %rd2985, 11; @%p1435 bra $L__BB106_1602; $L__BB106_1606: setp.eq.s16 %p1436, %rs145, 0; selp.b32 %r3234, -1, %r3230, %p1436; $L__BB106_1607: ld.const.u64 %rd2092, [kernel_params+2232]; cvta.to.global.u64 %rd2093, %rd2092; shl.b64 %rd2094, %rd324, 4; add.s64 %rd2095, %rd2093, %rd2094; ld.global.v4.u32 {%r2240, %r2241, %r2242, %r2243}, [%rd2095]; setp.eq.s32 %p1437, %r732, %r731; @%p1437 bra $L__BB106_1609; bra.uni $L__BB106_1608; $L__BB106_1609: ld.const.u64 %rd2101, [kernel_params+2256]; cvta.to.global.u64 %rd2102, %rd2101; mul.wide.u32 %rd2103, %r2243, 12; add.s64 %rd2986, %rd2102, %rd2103; add.s32 %r2250, %r2243, 1; mul.wide.u32 %rd2104, %r2250, 12; add.s64 %rd2989, %rd2102, %rd2104; add.s32 %r2251, %r2243, 2; mul.wide.u32 %rd2105, %r2251, 12; add.s64 %rd2992, %rd2102, %rd2105; add.s64 %rd2987, %rd2986, 4; add.s64 %rd2988, %rd2986, 8; add.s64 %rd2990, %rd2989, 4; add.s64 %rd2991, %rd2989, 8; add.s64 %rd2993, %rd2992, 4; add.s64 %rd2994, %rd2992, 8; bra.uni $L__BB106_1610; $L__BB106_1599: cvt.u32.u64 %r2225, %rd324; ld.const.v2.u64 {%rd2065, %rd2066}, [kernel_params+2224]; cvta.to.global.u64 %rd2069, %rd2066; mul.wide.s32 %rd2070, %r2225, 16; add.s64 %rd2071, %rd2069, %rd2070; ld.global.v4.u32 {%r2226, %r2227, %r2228, %r2229}, [%rd2071]; ld.const.u64 %rd2072, [kernel_params+2256]; cvta.to.global.u64 %rd2073, %rd2072; mul.wide.u32 %rd2074, %r2229, 12; add.s64 %rd2075, %rd2073, %rd2074; ld.global.f32 %f16257, [%rd2075]; ld.global.f32 %f16258, [%rd2075+4]; ld.global.f32 %f16259, [%rd2075+8]; add.s32 %r2234, %r2229, 1; mul.wide.u32 %rd2076, %r2234, 12; add.s64 %rd2077, %rd2073, %rd2076; ld.global.f32 %f16260, [%rd2077]; ld.global.f32 %f16261, [%rd2077+4]; ld.global.f32 %f16262, [%rd2077+8]; add.s32 %r2235, %r2229, 2; mul.wide.u32 %rd2078, %r2235, 12; add.s64 %rd2079, %rd2073, %rd2078; ld.global.f32 %f16263, [%rd2079]; ld.global.f32 %f16264, [%rd2079+4]; ld.global.f32 %f16265, [%rd2079+8]; cvta.to.global.u64 %rd2080, %rd2065; mul.wide.u32 %rd2081, %r2226, 12; add.s64 %rd2082, %rd2080, %rd2081; ld.global.f32 %f16248, [%rd2082]; ld.global.f32 %f16249, [%rd2082+4]; ld.global.f32 %f16250, [%rd2082+8]; mul.wide.u32 %rd2083, %r2227, 12; add.s64 %rd2084, %rd2080, %rd2083; ld.global.f32 %f16251, [%rd2084]; ld.global.f32 %f16252, [%rd2084+4]; ld.global.f32 %f16253, [%rd2084+8]; mul.wide.u32 %rd2085, %r2228, 12; add.s64 %rd2086, %rd2080, %rd2085; ld.global.f32 %f16254, [%rd2086]; ld.global.f32 %f16255, [%rd2086+4]; ld.global.f32 %f16256, [%rd2086+8]; bra.uni $L__BB106_1627; $L__BB106_1608: setp.gt.s32 %p1438, %r732, %r731; selp.b32 %r2244, -1, 0, %p1438; add.s32 %r2245, %r732, %r2244; mad.lo.s32 %r2246, %r2245, %r730, %r3234; ld.const.u64 %rd2096, [kernel_params+2336]; cvta.to.global.u64 %rd2097, %rd2096; add.s32 %r2247, %r2240, %r2246; mul.wide.u32 %rd2098, %r2247, 12; add.s64 %rd2986, %rd2097, %rd2098; add.s32 %r2248, %r2241, %r2246; mul.wide.u32 %rd2099, %r2248, 12; add.s64 %rd2989, %rd2097, %rd2099; add.s32 %r2249, %r2242, %r2246; mul.wide.u32 %rd2100, %r2249, 12; add.s64 %rd2992, %rd2097, %rd2100; add.s64 %rd2987, %rd2986, 4; add.s64 %rd2988, %rd2986, 8; add.s64 %rd2990, %rd2989, 4; add.s64 %rd2991, %rd2989, 8; add.s64 %rd2993, %rd2992, 4; add.s64 %rd2994, %rd2992, 8; $L__BB106_1610: ld.global.f32 %f2208, [%rd2994]; ld.global.f32 %f2209, [%rd2993]; ld.global.f32 %f2210, [%rd2992]; ld.global.f32 %f2211, [%rd2991]; ld.global.f32 %f2212, [%rd2990]; ld.global.f32 %f2213, [%rd2989]; ld.global.f32 %f2214, [%rd2988]; ld.global.f32 %f2215, [%rd2987]; ld.global.f32 %f2216, [%rd2986]; add.s32 %r747, %r732, 1; setp.eq.s32 %p1439, %r747, %r731; @%p1439 bra $L__BB106_1612; bra.uni $L__BB106_1611; $L__BB106_1612: ld.const.u64 %rd2111, [kernel_params+2256]; cvta.to.global.u64 %rd2112, %rd2111; mul.wide.u32 %rd2113, %r2243, 12; add.s64 %rd2995, %rd2112, %rd2113; add.s32 %r2258, %r2243, 1; mul.wide.u32 %rd2114, %r2258, 12; add.s64 %rd2998, %rd2112, %rd2114; add.s32 %r2259, %r2243, 2; mul.wide.u32 %rd2115, %r2259, 12; add.s64 %rd3001, %rd2112, %rd2115; add.s64 %rd2996, %rd2995, 4; add.s64 %rd2997, %rd2995, 8; add.s64 %rd2999, %rd2998, 4; add.s64 %rd3000, %rd2998, 8; add.s64 %rd3002, %rd3001, 4; add.s64 %rd3003, %rd3001, 8; bra.uni $L__BB106_1613; $L__BB106_1611: setp.ge.s32 %p1440, %r732, %r731; selp.b32 %r2252, -1, 0, %p1440; add.s32 %r2253, %r747, %r2252; mad.lo.s32 %r2254, %r2253, %r730, %r3234; ld.const.u64 %rd2106, [kernel_params+2336]; cvta.to.global.u64 %rd2107, %rd2106; add.s32 %r2255, %r2240, %r2254; mul.wide.u32 %rd2108, %r2255, 12; add.s64 %rd2995, %rd2107, %rd2108; add.s32 %r2256, %r2241, %r2254; mul.wide.u32 %rd2109, %r2256, 12; add.s64 %rd2998, %rd2107, %rd2109; add.s32 %r2257, %r2242, %r2254; mul.wide.u32 %rd2110, %r2257, 12; add.s64 %rd3001, %rd2107, %rd2110; add.s64 %rd2996, %rd2995, 4; add.s64 %rd2997, %rd2995, 8; add.s64 %rd2999, %rd2998, 4; add.s64 %rd3000, %rd2998, 8; add.s64 %rd3002, %rd3001, 4; add.s64 %rd3003, %rd3001, 8; $L__BB106_1613: sub.ftz.f32 %f2217, %f10294, %f2207; ld.global.f32 %f10305, [%rd2995]; mul.ftz.f32 %f10306, %f2207, %f10305; ld.global.f32 %f10307, [%rd2996]; mul.ftz.f32 %f10308, %f2207, %f10307; ld.global.f32 %f10309, [%rd2997]; mul.ftz.f32 %f10310, %f2207, %f10309; fma.rn.ftz.f32 %f16257, %f2217, %f2216, %f10306; fma.rn.ftz.f32 %f16258, %f2217, %f2215, %f10308; fma.rn.ftz.f32 %f16259, %f2217, %f2214, %f10310; ld.global.f32 %f10311, [%rd2998]; mul.ftz.f32 %f10312, %f2207, %f10311; ld.global.f32 %f10313, [%rd2999]; mul.ftz.f32 %f10314, %f2207, %f10313; ld.global.f32 %f10315, [%rd3000]; mul.ftz.f32 %f10316, %f2207, %f10315; fma.rn.ftz.f32 %f16260, %f2217, %f2213, %f10312; fma.rn.ftz.f32 %f16261, %f2217, %f2212, %f10314; fma.rn.ftz.f32 %f16262, %f2217, %f2211, %f10316; ld.global.f32 %f10317, [%rd3001]; mul.ftz.f32 %f10318, %f2207, %f10317; ld.global.f32 %f10319, [%rd3002]; mul.ftz.f32 %f10320, %f2207, %f10319; ld.global.f32 %f10321, [%rd3003]; mul.ftz.f32 %f10322, %f2207, %f10321; fma.rn.ftz.f32 %f16263, %f2217, %f2210, %f10318; fma.rn.ftz.f32 %f16264, %f2217, %f2209, %f10320; fma.rn.ftz.f32 %f16265, %f2217, %f2208, %f10322; setp.eq.s64 %p1441, %rd3004, 12; @%p1441 bra $L__BB106_1619; mov.u32 %r3239, -1; $L__BB106_1615: setp.eq.s64 %p1442, %rd3004, 0; @%p1442 bra $L__BB106_1617; add.s32 %r3236, %r3236, 2; bra.uni $L__BB106_1618; $L__BB106_1617: setp.eq.s16 %p1443, %rs147, 0; mov.u32 %r3236, %r3235; @%p1443 bra $L__BB106_1620; $L__BB106_1618: mul.wide.u32 %rd2116, %r3236, 16; add.s64 %rd2117, %rd325, %rd2116; ld.global.u16 %rs147, [%rd2117+12]; ld.global.u32 %r3235, [%rd2117+8]; ld.global.u64 %rd3004, [%rd2117]; setp.ne.s64 %p1444, %rd3004, 12; @%p1444 bra $L__BB106_1615; $L__BB106_1619: setp.eq.s16 %p1445, %rs147, 0; selp.b32 %r3239, -1, %r3235, %p1445; $L__BB106_1620: @%p1437 bra $L__BB106_1622; bra.uni $L__BB106_1621; $L__BB106_1622: ld.const.u64 %rd2123, [kernel_params+2224]; cvta.to.global.u64 %rd2124, %rd2123; mul.wide.u32 %rd2125, %r2240, 12; add.s64 %rd3005, %rd2124, %rd2125; mul.wide.u32 %rd2126, %r2241, 12; add.s64 %rd3008, %rd2124, %rd2126; mul.wide.u32 %rd2127, %r2242, 12; add.s64 %rd3011, %rd2124, %rd2127; add.s64 %rd3006, %rd3005, 4; add.s64 %rd3007, %rd3005, 8; add.s64 %rd3009, %rd3008, 4; add.s64 %rd3010, %rd3008, 8; add.s64 %rd3012, %rd3011, 4; add.s64 %rd3013, %rd3011, 8; bra.uni $L__BB106_1623; $L__BB106_1621: setp.gt.s32 %p1447, %r732, %r731; selp.b32 %r2261, -1, 0, %p1447; add.s32 %r2262, %r732, %r2261; mad.lo.s32 %r2263, %r2262, %r730, %r3239; ld.const.u64 %rd2118, [kernel_params+2336]; cvta.to.global.u64 %rd2119, %rd2118; add.s32 %r2264, %r2263, %r2240; mul.wide.u32 %rd2120, %r2264, 12; add.s64 %rd3005, %rd2119, %rd2120; add.s32 %r2265, %r2263, %r2241; mul.wide.u32 %rd2121, %r2265, 12; add.s64 %rd3008, %rd2119, %rd2121; add.s32 %r2266, %r2263, %r2242; mul.wide.u32 %rd2122, %r2266, 12; add.s64 %rd3011, %rd2119, %rd2122; add.s64 %rd3006, %rd3005, 4; add.s64 %rd3007, %rd3005, 8; add.s64 %rd3009, %rd3008, 4; add.s64 %rd3010, %rd3008, 8; add.s64 %rd3012, %rd3011, 4; add.s64 %rd3013, %rd3011, 8; $L__BB106_1623: ld.global.f32 %f2227, [%rd3013]; ld.global.f32 %f2228, [%rd3012]; ld.global.f32 %f2229, [%rd3011]; ld.global.f32 %f2230, [%rd3010]; ld.global.f32 %f2231, [%rd3009]; ld.global.f32 %f2232, [%rd3008]; ld.global.f32 %f2233, [%rd3007]; ld.global.f32 %f2234, [%rd3006]; ld.global.f32 %f2235, [%rd3005]; @%p1439 bra $L__BB106_1625; bra.uni $L__BB106_1624; $L__BB106_1625: ld.const.u64 %rd2133, [kernel_params+2224]; cvta.to.global.u64 %rd2134, %rd2133; mul.wide.u32 %rd2135, %r2240, 12; add.s64 %rd3014, %rd2134, %rd2135; mul.wide.u32 %rd2136, %r2241, 12; add.s64 %rd3017, %rd2134, %rd2136; mul.wide.u32 %rd2137, %r2242, 12; add.s64 %rd3020, %rd2134, %rd2137; add.s64 %rd3015, %rd3014, 4; add.s64 %rd3016, %rd3014, 8; add.s64 %rd3018, %rd3017, 4; add.s64 %rd3019, %rd3017, 8; add.s64 %rd3021, %rd3020, 4; add.s64 %rd3022, %rd3020, 8; bra.uni $L__BB106_1626; $L__BB106_1624: setp.ge.s32 %p1449, %r732, %r731; selp.b32 %r2267, -1, 0, %p1449; add.s32 %r2268, %r747, %r2267; mad.lo.s32 %r2269, %r2268, %r730, %r3239; ld.const.u64 %rd2128, [kernel_params+2336]; cvta.to.global.u64 %rd2129, %rd2128; add.s32 %r2270, %r2269, %r2240; mul.wide.u32 %rd2130, %r2270, 12; add.s64 %rd3014, %rd2129, %rd2130; add.s32 %r2271, %r2269, %r2241; mul.wide.u32 %rd2131, %r2271, 12; add.s64 %rd3017, %rd2129, %rd2131; add.s32 %r2272, %r2269, %r2242; mul.wide.u32 %rd2132, %r2272, 12; add.s64 %rd3020, %rd2129, %rd2132; add.s64 %rd3015, %rd3014, 4; add.s64 %rd3016, %rd3014, 8; add.s64 %rd3018, %rd3017, 4; add.s64 %rd3019, %rd3017, 8; add.s64 %rd3021, %rd3020, 4; add.s64 %rd3022, %rd3020, 8; $L__BB106_1626: ld.global.f32 %f10323, [%rd3014]; mul.ftz.f32 %f10324, %f2207, %f10323; ld.global.f32 %f10325, [%rd3015]; mul.ftz.f32 %f10326, %f2207, %f10325; ld.global.f32 %f10327, [%rd3016]; mul.ftz.f32 %f10328, %f2207, %f10327; fma.rn.ftz.f32 %f16248, %f2217, %f2235, %f10324; fma.rn.ftz.f32 %f16249, %f2217, %f2234, %f10326; fma.rn.ftz.f32 %f16250, %f2217, %f2233, %f10328; ld.global.f32 %f10329, [%rd3017]; mul.ftz.f32 %f10330, %f2207, %f10329; ld.global.f32 %f10331, [%rd3018]; mul.ftz.f32 %f10332, %f2207, %f10331; ld.global.f32 %f10333, [%rd3019]; mul.ftz.f32 %f10334, %f2207, %f10333; fma.rn.ftz.f32 %f16251, %f2217, %f2232, %f10330; fma.rn.ftz.f32 %f16252, %f2217, %f2231, %f10332; fma.rn.ftz.f32 %f16253, %f2217, %f2230, %f10334; ld.global.f32 %f10335, [%rd3020]; mul.ftz.f32 %f10336, %f2207, %f10335; ld.global.f32 %f10337, [%rd3021]; mul.ftz.f32 %f10338, %f2207, %f10337; ld.global.f32 %f10339, [%rd3022]; mul.ftz.f32 %f10340, %f2207, %f10339; fma.rn.ftz.f32 %f16254, %f2217, %f2229, %f10336; fma.rn.ftz.f32 %f16255, %f2217, %f2228, %f10338; fma.rn.ftz.f32 %f16256, %f2217, %f2227, %f10340; $L__BB106_1627: ld.local.f32 %f2263, [%rd7+-20]; sub.ftz.f32 %f10342, %f10294, %f2263; ld.local.f32 %f10343, [%rd7+-16]; sub.ftz.f32 %f2264, %f10342, %f10343; mul.ftz.f32 %f10344, %f16257, %f2264; mul.ftz.f32 %f10345, %f16258, %f2264; mul.ftz.f32 %f10346, %f16259, %f2264; fma.rn.ftz.f32 %f10347, %f16260, %f2263, %f10344; fma.rn.ftz.f32 %f10348, %f16261, %f2263, %f10345; fma.rn.ftz.f32 %f10349, %f16262, %f2263, %f10346; fma.rn.ftz.f32 %f2265, %f16263, %f10343, %f10347; fma.rn.ftz.f32 %f2266, %f16264, %f10343, %f10348; fma.rn.ftz.f32 %f2267, %f16265, %f10343, %f10349; mul.ftz.f32 %f10350, %f16248, %f2264; mul.ftz.f32 %f10351, %f16249, %f2264; mul.ftz.f32 %f10352, %f16250, %f2264; fma.rn.ftz.f32 %f10353, %f16251, %f2263, %f10350; fma.rn.ftz.f32 %f10354, %f16252, %f2263, %f10351; fma.rn.ftz.f32 %f10355, %f16253, %f2263, %f10352; fma.rn.ftz.f32 %f16266, %f16254, %f10343, %f10353; fma.rn.ftz.f32 %f16267, %f16255, %f10343, %f10354; fma.rn.ftz.f32 %f16268, %f16256, %f10343, %f10355; ld.local.u32 %r756, [%rd7+-32]; and.b32 %r2273, %r756, 4; setp.ne.s32 %p1450, %r2273, 0; @%p1450 bra $L__BB106_1631; and.b32 %r2274, %r756, 2; setp.eq.s32 %p1451, %r2274, 0; @%p1451 bra $L__BB106_1630; bra.uni $L__BB106_1629; $L__BB106_1630: ld.global.v4.f32 {%f10383, %f10384, %f10385, %f10386}, [%rd323+-184]; ld.global.v4.f32 {%f10390, %f10391, %f10392, %f10393}, [%rd323+-168]; ld.global.v4.f32 {%f10397, %f10398, %f10399, %f10400}, [%rd323+-152]; mul.ftz.f32 %f10404, %f16267, %f10384; fma.rn.ftz.f32 %f10405, %f16266, %f10383, %f10404; fma.rn.ftz.f32 %f2274, %f16268, %f10385, %f10405; mul.ftz.f32 %f10406, %f16267, %f10391; fma.rn.ftz.f32 %f10407, %f16266, %f10390, %f10406; fma.rn.ftz.f32 %f2275, %f16268, %f10392, %f10407; mul.ftz.f32 %f10408, %f16267, %f10398; fma.rn.ftz.f32 %f10409, %f16266, %f10397, %f10408; fma.rn.ftz.f32 %f16268, %f16268, %f10399, %f10409; mov.f32 %f16266, %f2274; mov.f32 %f16267, %f2275; bra.uni $L__BB106_1631; $L__BB106_1629: ld.local.v4.f32 {%f10356, %f10357, %f10358, %f10359}, [%rd7+56]; mul.ftz.f32 %f10363, %f16267, %f10357; fma.rn.ftz.f32 %f10364, %f16266, %f10356, %f10363; fma.rn.ftz.f32 %f2271, %f16268, %f10358, %f10364; ld.local.v4.f32 {%f10365, %f10366, %f10367, %f10368}, [%rd7+72]; mul.ftz.f32 %f10372, %f16267, %f10366; fma.rn.ftz.f32 %f10373, %f16266, %f10365, %f10372; fma.rn.ftz.f32 %f2272, %f16268, %f10367, %f10373; ld.local.v4.f32 {%f10374, %f10375, %f10376, %f10377}, [%rd7+88]; mul.ftz.f32 %f10381, %f16267, %f10375; fma.rn.ftz.f32 %f10382, %f16266, %f10374, %f10381; fma.rn.ftz.f32 %f16268, %f16268, %f10376, %f10382; mov.f32 %f16266, %f2271; mov.f32 %f16267, %f2272; $L__BB106_1631: sub.ftz.f32 %f2280, %f16257, %f16263; sub.ftz.f32 %f10410, %f16254, %f16248; sub.ftz.f32 %f2281, %f16258, %f16264; sub.ftz.f32 %f10411, %f16255, %f16249; mul.ftz.f32 %f10412, %f10411, %f2281; fma.rn.ftz.f32 %f10413, %f10410, %f2280, %f10412; sub.ftz.f32 %f2282, %f16259, %f16265; sub.ftz.f32 %f10414, %f16256, %f16250; fma.rn.ftz.f32 %f10415, %f10414, %f2282, %f10413; sub.ftz.f32 %f2283, %f16260, %f16263; sub.ftz.f32 %f10416, %f16254, %f16251; sub.ftz.f32 %f2284, %f16261, %f16264; sub.ftz.f32 %f10417, %f16255, %f16252; mul.ftz.f32 %f10418, %f10417, %f2284; fma.rn.ftz.f32 %f10419, %f10416, %f2283, %f10418; sub.ftz.f32 %f2285, %f16262, %f16265; sub.ftz.f32 %f10420, %f16256, %f16253; fma.rn.ftz.f32 %f10421, %f10420, %f2285, %f10419; sub.ftz.f32 %f2286, %f16260, %f16257; sub.ftz.f32 %f10422, %f16251, %f16248; sub.ftz.f32 %f2287, %f16261, %f16258; sub.ftz.f32 %f10423, %f16252, %f16249; mul.ftz.f32 %f10424, %f10423, %f2287; fma.rn.ftz.f32 %f10425, %f10422, %f2286, %f10424; sub.ftz.f32 %f2288, %f16262, %f16259; sub.ftz.f32 %f10426, %f16253, %f16250; fma.rn.ftz.f32 %f10427, %f10426, %f2288, %f10425; mul.ftz.f32 %f10428, %f10415, %f2264; add.ftz.f32 %f10429, %f2264, 0fBF800000; add.ftz.f32 %f10430, %f10415, %f10421; add.ftz.f32 %f10431, %f10427, %f10430; mul.ftz.f32 %f10432, %f2264, %f10431; mul.ftz.f32 %f10433, %f2263, %f10432; fma.rn.ftz.f32 %f10434, %f10428, %f10429, %f10433; mul.ftz.f32 %f10435, %f2263, %f10421; add.ftz.f32 %f10436, %f2263, 0fBF800000; fma.rn.ftz.f32 %f10437, %f10436, %f10435, %f10434; mul.ftz.f32 %f10438, %f16246, %f16267; fma.rn.ftz.f32 %f10439, %f16245, %f16266, %f10438; fma.rn.ftz.f32 %f10440, %f16247, %f16268, %f10439; setp.gt.ftz.f32 %p1452, %f10440, 0f00000000; sub.ftz.f32 %f2289, %f16257, %f16260; sub.ftz.f32 %f2290, %f16258, %f16261; sub.ftz.f32 %f2291, %f16259, %f16262; sub.ftz.f32 %f2292, %f16263, %f16257; sub.ftz.f32 %f2293, %f16264, %f16258; sub.ftz.f32 %f2294, %f16265, %f16259; sub.ftz.f32 %f2295, %f16263, %f16260; sub.ftz.f32 %f2296, %f16264, %f16261; sub.ftz.f32 %f2297, %f16265, %f16262; mul.ftz.f32 %f2298, %f10437, 0f3F000000; @%p1452 bra $L__BB106_1633; bra.uni $L__BB106_1632; $L__BB106_1633: mul.ftz.f32 %f10493, %f16249, %f2287; fma.rn.ftz.f32 %f10494, %f16248, %f2286, %f10493; fma.rn.ftz.f32 %f10495, %f16250, %f2288, %f10494; mul.ftz.f32 %f10496, %f16249, %f2293; fma.rn.ftz.f32 %f10497, %f16248, %f2292, %f10496; fma.rn.ftz.f32 %f10498, %f16250, %f2294, %f10497; max.ftz.f32 %f10499, %f10495, %f10498; max.ftz.f32 %f10501, %f10499, %f10298; mul.ftz.f32 %f10502, %f16252, %f2290; fma.rn.ftz.f32 %f10503, %f16251, %f2289, %f10502; fma.rn.ftz.f32 %f10504, %f16253, %f2291, %f10503; mul.ftz.f32 %f10505, %f16252, %f2296; fma.rn.ftz.f32 %f10506, %f16251, %f2295, %f10505; fma.rn.ftz.f32 %f10507, %f16253, %f2297, %f10506; max.ftz.f32 %f10508, %f10504, %f10507; max.ftz.f32 %f10509, %f10508, %f10298; mul.ftz.f32 %f10510, %f16255, %f2281; fma.rn.ftz.f32 %f10511, %f16254, %f2280, %f10510; fma.rn.ftz.f32 %f10512, %f16256, %f2282, %f10511; mul.ftz.f32 %f10513, %f16255, %f2284; fma.rn.ftz.f32 %f10514, %f16254, %f2283, %f10513; fma.rn.ftz.f32 %f10515, %f16256, %f2285, %f10514; max.ftz.f32 %f10516, %f10512, %f10515; max.ftz.f32 %f10517, %f10516, %f10298; sub.ftz.f32 %f10518, %f16257, %f2265; sub.ftz.f32 %f10519, %f16258, %f2266; mul.ftz.f32 %f10520, %f16249, %f10519; fma.rn.ftz.f32 %f10521, %f16248, %f10518, %f10520; sub.ftz.f32 %f10522, %f16259, %f2267; fma.rn.ftz.f32 %f10523, %f16250, %f10522, %f10521; add.ftz.f32 %f10524, %f10523, %f10501; max.ftz.f32 %f10525, %f10524, %f10298; sub.ftz.f32 %f10526, %f16260, %f2265; sub.ftz.f32 %f10527, %f16261, %f2266; mul.ftz.f32 %f10528, %f16252, %f10527; fma.rn.ftz.f32 %f10529, %f16251, %f10526, %f10528; sub.ftz.f32 %f10530, %f16262, %f2267; fma.rn.ftz.f32 %f10531, %f16253, %f10530, %f10529; add.ftz.f32 %f10532, %f10531, %f10509; max.ftz.f32 %f10533, %f10532, %f10298; sub.ftz.f32 %f10534, %f16263, %f2265; sub.ftz.f32 %f10535, %f16264, %f2266; mul.ftz.f32 %f10536, %f16255, %f10535; fma.rn.ftz.f32 %f10537, %f16254, %f10534, %f10536; sub.ftz.f32 %f10538, %f16265, %f2267; fma.rn.ftz.f32 %f10539, %f16256, %f10538, %f10537; add.ftz.f32 %f10540, %f10539, %f10517; max.ftz.f32 %f10541, %f10540, %f10298; min.ftz.f32 %f10542, %f10525, %f10533; min.ftz.f32 %f10543, %f10542, %f10541; max.ftz.f32 %f16269, %f10543, %f2298; bra.uni $L__BB106_1634; $L__BB106_1632: mul.ftz.f32 %f10441, %f16249, %f2290; fma.rn.ftz.f32 %f10442, %f16248, %f2289, %f10441; fma.rn.ftz.f32 %f10443, %f16250, %f2291, %f10442; mul.ftz.f32 %f10444, %f16249, %f2281; fma.rn.ftz.f32 %f10445, %f16248, %f2280, %f10444; fma.rn.ftz.f32 %f10446, %f16250, %f2282, %f10445; max.ftz.f32 %f10447, %f10443, %f10446; max.ftz.f32 %f10449, %f10447, %f10298; mul.ftz.f32 %f10450, %f16252, %f2287; fma.rn.ftz.f32 %f10451, %f16251, %f2286, %f10450; fma.rn.ftz.f32 %f10452, %f16253, %f2288, %f10451; mul.ftz.f32 %f10453, %f16252, %f2284; fma.rn.ftz.f32 %f10454, %f16251, %f2283, %f10453; fma.rn.ftz.f32 %f10455, %f16253, %f2285, %f10454; max.ftz.f32 %f10456, %f10452, %f10455; max.ftz.f32 %f10457, %f10456, %f10298; mul.ftz.f32 %f10458, %f16255, %f2293; fma.rn.ftz.f32 %f10459, %f16254, %f2292, %f10458; fma.rn.ftz.f32 %f10460, %f16256, %f2294, %f10459; mul.ftz.f32 %f10461, %f16255, %f2296; fma.rn.ftz.f32 %f10462, %f16254, %f2295, %f10461; fma.rn.ftz.f32 %f10463, %f16256, %f2297, %f10462; max.ftz.f32 %f10464, %f10460, %f10463; max.ftz.f32 %f10465, %f10464, %f10298; sub.ftz.f32 %f10466, %f2265, %f16257; sub.ftz.f32 %f10467, %f2266, %f16258; mul.ftz.f32 %f10468, %f16249, %f10467; fma.rn.ftz.f32 %f10469, %f16248, %f10466, %f10468; sub.ftz.f32 %f10470, %f2267, %f16259; fma.rn.ftz.f32 %f10471, %f16250, %f10470, %f10469; add.ftz.f32 %f10472, %f10471, %f10449; max.ftz.f32 %f10473, %f10472, %f10298; sub.ftz.f32 %f10474, %f2265, %f16260; sub.ftz.f32 %f10475, %f2266, %f16261; mul.ftz.f32 %f10476, %f16252, %f10475; fma.rn.ftz.f32 %f10477, %f16251, %f10474, %f10476; sub.ftz.f32 %f10478, %f2267, %f16262; fma.rn.ftz.f32 %f10479, %f16253, %f10478, %f10477; add.ftz.f32 %f10480, %f10479, %f10457; max.ftz.f32 %f10481, %f10480, %f10298; sub.ftz.f32 %f10482, %f2265, %f16263; sub.ftz.f32 %f10483, %f2266, %f16264; mul.ftz.f32 %f10484, %f16255, %f10483; fma.rn.ftz.f32 %f10485, %f16254, %f10482, %f10484; sub.ftz.f32 %f10486, %f2267, %f16265; fma.rn.ftz.f32 %f10487, %f16256, %f10486, %f10485; add.ftz.f32 %f10488, %f10487, %f10465; max.ftz.f32 %f10489, %f10488, %f10298; min.ftz.f32 %f10490, %f10473, %f10481; min.ftz.f32 %f10491, %f10490, %f10489; neg.ftz.f32 %f10492, %f10491; min.ftz.f32 %f16269, %f10492, %f2298; $L__BB106_1634: mul.ftz.f32 %f10544, %f16266, %f16269; mul.ftz.f32 %f10545, %f16267, %f16269; mul.ftz.f32 %f10546, %f16268, %f16269; fma.rn.ftz.f32 %f16270, %f2188, %f10544, %f16270; fma.rn.ftz.f32 %f16271, %f2188, %f10545, %f16271; fma.rn.ftz.f32 %f16272, %f2188, %f10546, %f16272; setp.gt.ftz.f32 %p1453, %f2187, 0f00000000; selp.u16 %rs149, 1, 0, %p1453; $L__BB106_1635: and.b32 %r2275, %r618, 1073741824; setp.eq.s32 %p1454, %r2275, 0; mov.f32 %f16273, 0f00000000; mov.f32 %f16274, %f16273; mov.f32 %f16275, %f16273; mov.f32 %f16276, %f16273; mov.f32 %f16277, %f16273; mov.f32 %f16278, %f16273; mov.f32 %f16279, %f16273; @%p1454 bra $L__BB106_1638; ld.local.f32 %f10555, [%rd282+12]; setp.eq.ftz.f32 %p1455, %f10555, 0f7F7FFFFF; mov.f32 %f16279, 0f7F7FFFFF; mov.f32 %f16273, %f16270; mov.f32 %f16274, %f16271; mov.f32 %f16275, %f16272; mov.f32 %f16276, %f1935; mov.f32 %f16277, %f1936; mov.f32 %f16278, %f1937; @%p1455 bra $L__BB106_1638; ld.local.f32 %f10556, [%rd282+-24]; sub.ftz.f32 %f10557, %f10556, %f16270; ld.local.f32 %f10558, [%rd282+-20]; sub.ftz.f32 %f10559, %f10558, %f16271; ld.local.f32 %f10560, [%rd282+-16]; sub.ftz.f32 %f10561, %f10560, %f16272; mul.ftz.f32 %f10562, %f10559, %f10559; fma.rn.ftz.f32 %f10563, %f10557, %f10557, %f10562; fma.rn.ftz.f32 %f10564, %f10561, %f10561, %f10563; sqrt.approx.ftz.f32 %f16279, %f10564; rsqrt.approx.ftz.f32 %f10565, %f10564; mul.ftz.f32 %f16276, %f10557, %f10565; mul.ftz.f32 %f16277, %f10559, %f10565; mul.ftz.f32 %f16278, %f10561, %f10565; mov.f32 %f16273, %f16270; mov.f32 %f16274, %f16271; mov.f32 %f16275, %f16272; $L__BB106_1638: ld.local.f32 %f2319, [%rd7+4]; ld.local.f32 %f2320, [%rd7+-4]; ld.local.u32 %r757, [%rd7+-12]; setp.eq.s16 %p6, %rs149, 0; selp.b32 %r758, -1, %r757, %p6; ld.local.u32 %r759, [%rd7+-28]; ld.local.u32 %r760, [%rd282+32]; ld.local.u32 %r761, [%rd282+36]; ld.local.u32 %r762, [%rd282+52]; ld.const.u64 %rd2138, [kernel_params+4376]; cvta.to.global.u64 %rd2139, %rd2138; atom.global.add.u32 %r2276, [%rd2139], 1; ld.const.u64 %rd2140, [kernel_params+4256]; cvta.to.global.u64 %rd2141, %rd2140; add.s64 %rd2142, %rd2141, 12; atom.global.add.u32 %r2277, [%rd2142], 1; ld.const.u64 %rd2143, [kernel_params+3416]; cvta.to.global.u64 %rd2144, %rd2143; cvt.s64.s32 %rd439, %r2276; mul.wide.s32 %rd2145, %r2276, 2; add.s64 %rd2146, %rd2144, %rd2145; mov.u16 %rs93, 3; st.global.u16 [%rd2146], %rs93; ld.const.u32 %r763, [kernel_params]; and.b32 %r2278, %r763, 131072; setp.eq.s32 %p1456, %r2278, 0; @%p1456 bra $L__BB106_1641; mov.u32 %r3240, 0; mov.u64 %rd2148, kernel_params; $L__BB106_1640: mul.wide.s32 %rd2147, %r3240, 16; add.s64 %rd2149, %rd2148, %rd2147; ld.const.v2.u64 {%rd2150, %rd2151}, [%rd2149+2768]; cvta.to.global.u64 %rd2154, %rd2151; add.s64 %rd2156, %rd2154, %rd837; ld.global.u32 %r2280, [%rd2156]; cvta.to.global.u64 %rd2157, %rd2150; add.s64 %rd2158, %rd2157, %rd837; ld.global.u32 %r2281, [%rd2158]; ld.const.v2.u64 {%rd2159, %rd2160}, [%rd2149+3744]; cvta.to.global.u64 %rd2163, %rd2159; shl.b64 %rd2164, %rd439, 2; add.s64 %rd2165, %rd2163, %rd2164; st.global.u32 [%rd2165], %r2281; cvta.to.global.u64 %rd2166, %rd2160; add.s64 %rd2167, %rd2166, %rd2164; st.global.u32 [%rd2167], %r2280; add.s32 %r3240, %r3240, 1; setp.ne.s32 %p1457, %r2280, -1; @%p1457 bra $L__BB106_1640; $L__BB106_1641: setp.geu.ftz.f32 %p1458, %f1938, 0f00000000; @%p1458 bra $L__BB106_1656; ld.local.u32 %r766, [%rd7+-36]; and.b32 %r2282, %r766, 262144; setp.eq.s32 %p1459, %r2282, 0; @%p1459 bra $L__BB106_1656; bra.uni $L__BB106_1643; $L__BB106_1656: setp.eq.s32 %p1476, %r758, -1; selp.b32 %r780, -1, %r759, %p6; @%p1476 bra $L__BB106_1665; @%p1424 bra $L__BB106_1665; ld.const.u64 %rd2231, [kernel_params+2232]; cvta.to.global.u64 %rd2232, %rd2231; mul.wide.s32 %rd2233, %r759, 16; add.s64 %rd2234, %rd2232, %rd2233; ld.const.u64 %rd2235, [kernel_params+2256]; cvta.to.global.u64 %rd2236, %rd2235; ld.global.u32 %r2295, [%rd2234+12]; mul.wide.u32 %rd2237, %r2295, 12; add.s64 %rd2238, %rd2236, %rd2237; ld.global.f32 %f2324, [%rd2238]; ld.global.f32 %f2325, [%rd2238+4]; ld.global.f32 %f2326, [%rd2238+8]; add.s32 %r2296, %r2295, 1; mul.wide.u32 %rd2239, %r2296, 12; add.s64 %rd2240, %rd2236, %rd2239; ld.global.f32 %f2327, [%rd2240]; ld.global.f32 %f2328, [%rd2240+4]; ld.global.f32 %f2329, [%rd2240+8]; add.s32 %r2297, %r2295, 2; mul.wide.u32 %rd2241, %r2297, 12; add.s64 %rd2242, %rd2236, %rd2241; ld.global.f32 %f2330, [%rd2242]; ld.global.f32 %f2331, [%rd2242+4]; ld.global.f32 %f2332, [%rd2242+8]; ld.local.u32 %r781, [%rd7+-32]; and.b32 %r2298, %r781, 4; setp.ne.s32 %p1478, %r2298, 0; mov.f32 %f16292, %f16273; mov.f32 %f16293, %f16274; mov.f32 %f16294, %f16275; mov.f32 %f16295, %f16276; mov.f32 %f16296, %f16277; mov.f32 %f16297, %f16278; @%p1478 bra $L__BB106_1663; and.b32 %r2299, %r781, 2; setp.eq.s32 %p1479, %r2299, 0; @%p1479 bra $L__BB106_1661; bra.uni $L__BB106_1660; $L__BB106_1661: ld.const.u64 %rd2243, [kernel_params+2160]; cvta.to.global.u64 %rd2244, %rd2243; mul.wide.s32 %rd2245, %r757, 208; add.s64 %rd2246, %rd2244, %rd2245; ld.global.v4.f32 {%f16280, %f16281, %f16282, %f16283}, [%rd2246+48]; ld.global.v4.f32 {%f16284, %f16285, %f16286, %f16287}, [%rd2246+64]; ld.global.v4.f32 {%f16288, %f16289, %f16290, %f16291}, [%rd2246+80]; bra.uni $L__BB106_1662; $L__BB106_1643: and.b32 %r2283, %r766, 1; setp.eq.b32 %p1460, %r2283, 1; mov.pred %p1461, 0; xor.pred %p1462, %p1460, %p1461; not.pred %p1463, %p1462; ld.const.u64 %rd3023, [kernel_params+3752]; cvta.to.global.u64 %rd2168, %rd3023; shl.b64 %rd2169, %rd439, 2; add.s64 %rd2170, %rd2168, %rd2169; ld.global.u32 %r3241, [%rd2170]; @%p1463 bra $L__BB106_1650; bra.uni $L__BB106_1644; $L__BB106_1650: mov.u64 %rd2201, kernel_params; add.s64 %rd3029, %rd2201, 3744; setp.eq.s32 %p1470, %r3241, -1; mov.u32 %r3246, 0; @%p1470 bra $L__BB106_1654; add.s64 %rd3029, %rd2201, 3744; mov.u32 %r3246, 0; $L__BB106_1652: ld.const.u64 %rd2204, [%rd3029]; cvta.to.global.u64 %rd2205, %rd2204; add.s64 %rd2207, %rd2205, %rd2169; ld.global.u32 %r2290, [%rd2207]; setp.eq.s32 %p1471, %r2290, %r757; setp.eq.s32 %p1472, %r3241, %r729; and.pred %p1473, %p1471, %p1472; @%p1473 bra $L__BB106_1656; add.s32 %r3246, %r3246, 1; mul.wide.s32 %rd2208, %r3246, 16; add.s64 %rd2210, %rd2201, %rd2208; add.s64 %rd3029, %rd2210, 3744; ld.const.u64 %rd3023, [%rd2210+3752]; cvta.to.global.u64 %rd2211, %rd3023; add.s64 %rd2213, %rd2211, %rd2169; ld.global.u32 %r3241, [%rd2213]; setp.ne.s32 %p1474, %r3241, -1; @%p1474 bra $L__BB106_1652; $L__BB106_1654: ld.const.u32 %r2291, [kernel_params+12]; add.s32 %r2292, %r2291, -1; setp.ge.u32 %p1475, %r3246, %r2292; @%p1475 bra $L__BB106_1656; ld.const.u64 %rd2214, [%rd3029]; cvta.to.global.u64 %rd2215, %rd2214; add.s64 %rd2217, %rd2215, %rd2169; st.global.u32 [%rd2217], %r757; cvta.to.global.u64 %rd2218, %rd3023; add.s64 %rd2219, %rd2218, %rd2169; st.global.u32 [%rd2219], %r729; add.s32 %r2293, %r3246, 1; mul.wide.s32 %rd2220, %r2293, 16; add.s64 %rd2222, %rd2201, %rd2220; ld.const.v2.u64 {%rd2223, %rd2224}, [%rd2222+3744]; cvta.to.global.u64 %rd2227, %rd2223; add.s64 %rd2228, %rd2227, %rd2169; mov.u32 %r2294, -1; st.global.u32 [%rd2228], %r2294; cvta.to.global.u64 %rd2229, %rd2224; add.s64 %rd2230, %rd2229, %rd2169; st.global.u32 [%rd2230], %r2294; bra.uni $L__BB106_1656; $L__BB106_1644: setp.eq.s32 %p1464, %r3241, -1; @%p1464 bra $L__BB106_1656; mov.u64 %rd2172, kernel_params; add.s64 %rd3024, %rd2172, 3744; mov.u32 %r3242, 0; $L__BB106_1646: ld.const.u64 %rd2173, [%rd3024]; cvta.to.global.u64 %rd2174, %rd2173; add.s64 %rd2176, %rd2174, %rd2169; ld.global.u32 %r2285, [%rd2176]; setp.eq.s32 %p1465, %r2285, %r757; setp.eq.s32 %p1466, %r3241, %r729; and.pred %p1467, %p1465, %p1466; add.s32 %r770, %r3242, 1; mul.wide.s32 %rd2177, %r770, 16; add.s64 %rd2179, %rd2172, %rd2177; add.s64 %rd443, %rd2179, 3752; @%p1467 bra $L__BB106_1648; add.s64 %rd3024, %rd2179, 3744; ld.const.u64 %rd3023, [%rd443]; cvta.to.global.u64 %rd2183, %rd3023; add.s64 %rd2185, %rd2183, %rd2169; ld.global.u32 %r3241, [%rd2185]; setp.eq.s32 %p1468, %r3241, -1; mov.u32 %r3242, %r770; @%p1468 bra $L__BB106_1656; bra.uni $L__BB106_1646; $L__BB106_1660: ld.local.v4.f32 {%f16280, %f16281, %f16282, %f16283}, [%rd7+104]; ld.local.v4.f32 {%f16284, %f16285, %f16286, %f16287}, [%rd7+120]; ld.local.v4.f32 {%f16288, %f16289, %f16290, %f16291}, [%rd7+136]; $L__BB106_1662: mul.ftz.f32 %f10590, %f16274, %f16281; fma.rn.ftz.f32 %f10591, %f16273, %f16280, %f10590; fma.rn.ftz.f32 %f10592, %f16275, %f16282, %f10591; mul.ftz.f32 %f10593, %f16274, %f16285; fma.rn.ftz.f32 %f10594, %f16273, %f16284, %f10593; fma.rn.ftz.f32 %f10595, %f16275, %f16286, %f10594; mul.ftz.f32 %f10596, %f16274, %f16289; fma.rn.ftz.f32 %f10597, %f16273, %f16288, %f10596; fma.rn.ftz.f32 %f10598, %f16275, %f16290, %f10597; add.ftz.f32 %f16292, %f16283, %f10592; add.ftz.f32 %f16293, %f16287, %f10595; add.ftz.f32 %f16294, %f16291, %f10598; mul.ftz.f32 %f10599, %f16277, %f16281; fma.rn.ftz.f32 %f10600, %f16276, %f16280, %f10599; mul.ftz.f32 %f10601, %f16277, %f16285; fma.rn.ftz.f32 %f10602, %f16276, %f16284, %f10601; mul.ftz.f32 %f10603, %f16277, %f16289; fma.rn.ftz.f32 %f10604, %f16276, %f16288, %f10603; fma.rn.ftz.f32 %f16295, %f16278, %f16282, %f10600; fma.rn.ftz.f32 %f16296, %f16278, %f16286, %f10602; fma.rn.ftz.f32 %f16297, %f16278, %f16290, %f10604; $L__BB106_1663: sub.ftz.f32 %f10605, %f2324, %f16292; sub.ftz.f32 %f10606, %f2325, %f16293; sub.ftz.f32 %f10607, %f2326, %f16294; sub.ftz.f32 %f10608, %f2327, %f16292; sub.ftz.f32 %f10609, %f2328, %f16293; sub.ftz.f32 %f10610, %f2329, %f16294; sub.ftz.f32 %f10611, %f2330, %f16292; sub.ftz.f32 %f10612, %f2331, %f16293; sub.ftz.f32 %f10613, %f2332, %f16294; sub.ftz.f32 %f10614, %f10611, %f10605; sub.ftz.f32 %f10615, %f10612, %f10606; sub.ftz.f32 %f10616, %f10613, %f10607; sub.ftz.f32 %f10617, %f10605, %f10608; sub.ftz.f32 %f10618, %f10606, %f10609; sub.ftz.f32 %f10619, %f10607, %f10610; sub.ftz.f32 %f10620, %f10608, %f10611; sub.ftz.f32 %f10621, %f10609, %f10612; sub.ftz.f32 %f10622, %f10610, %f10613; add.ftz.f32 %f10623, %f10611, %f10605; add.ftz.f32 %f10624, %f10612, %f10606; add.ftz.f32 %f10625, %f10613, %f10607; mul.ftz.f32 %f10626, %f10624, %f10616; mul.ftz.f32 %f10627, %f10625, %f10615; sub.ftz.f32 %f10628, %f10626, %f10627; mul.ftz.f32 %f10629, %f10625, %f10614; mul.ftz.f32 %f10630, %f10623, %f10616; sub.ftz.f32 %f10631, %f10629, %f10630; mul.ftz.f32 %f10632, %f10623, %f10615; mul.ftz.f32 %f10633, %f10624, %f10614; sub.ftz.f32 %f10634, %f10632, %f10633; mul.ftz.f32 %f10635, %f16296, %f10631; fma.rn.ftz.f32 %f10636, %f16295, %f10628, %f10635; fma.rn.ftz.f32 %f10637, %f16297, %f10634, %f10636; add.ftz.f32 %f10638, %f10605, %f10608; add.ftz.f32 %f10639, %f10606, %f10609; add.ftz.f32 %f10640, %f10607, %f10610; mul.ftz.f32 %f10641, %f10639, %f10619; mul.ftz.f32 %f10642, %f10640, %f10618; sub.ftz.f32 %f10643, %f10641, %f10642; mul.ftz.f32 %f10644, %f10640, %f10617; mul.ftz.f32 %f10645, %f10638, %f10619; sub.ftz.f32 %f10646, %f10644, %f10645; mul.ftz.f32 %f10647, %f10638, %f10618; mul.ftz.f32 %f10648, %f10639, %f10617; sub.ftz.f32 %f10649, %f10647, %f10648; mul.ftz.f32 %f10650, %f16296, %f10646; fma.rn.ftz.f32 %f10651, %f16295, %f10643, %f10650; fma.rn.ftz.f32 %f10652, %f16297, %f10649, %f10651; add.ftz.f32 %f10653, %f10608, %f10611; add.ftz.f32 %f10654, %f10609, %f10612; add.ftz.f32 %f10655, %f10610, %f10613; mul.ftz.f32 %f10656, %f10654, %f10622; mul.ftz.f32 %f10657, %f10655, %f10621; sub.ftz.f32 %f10658, %f10656, %f10657; mul.ftz.f32 %f10659, %f10655, %f10620; mul.ftz.f32 %f10660, %f10653, %f10622; sub.ftz.f32 %f10661, %f10659, %f10660; mul.ftz.f32 %f10662, %f10653, %f10621; mul.ftz.f32 %f10663, %f10654, %f10620; sub.ftz.f32 %f10664, %f10662, %f10663; mul.ftz.f32 %f10665, %f16296, %f10661; fma.rn.ftz.f32 %f10666, %f16295, %f10658, %f10665; fma.rn.ftz.f32 %f10667, %f16297, %f10664, %f10666; add.ftz.f32 %f10668, %f10637, %f10652; add.ftz.f32 %f10669, %f10667, %f10668; abs.ftz.f32 %f10670, %f10669; mul.ftz.f32 %f10671, %f10670, 0f34000000; min.ftz.f32 %f10672, %f10652, %f10667; min.ftz.f32 %f10673, %f10637, %f10672; max.ftz.f32 %f10674, %f10652, %f10667; max.ftz.f32 %f10675, %f10637, %f10674; setp.ge.ftz.f32 %p1480, %f10673, %f10671; mul.ftz.f32 %f10676, %f10670, 0fB4000000; setp.le.ftz.f32 %p1481, %f10675, %f10676; or.pred %p1482, %p1480, %p1481; @%p1482 bra $L__BB106_1665; ld.local.v2.f32 {%f10677, %f10678}, [%rd7+-64]; mul.ftz.f32 %f10681, %f10677, 0f43800000; cvt.rzi.ftz.s32.f32 %r2300, %f10681; mul.ftz.f32 %f10682, %f10678, 0f43800000; cvt.rzi.ftz.s32.f32 %r2301, %f10682; ld.local.f32 %f10683, [%rd7+-56]; mul.ftz.f32 %f10684, %f10683, 0f43800000; cvt.rzi.ftz.s32.f32 %r2302, %f10684; neg.s32 %r2303, %r2300; setp.lt.ftz.f32 %p1483, %f16273, 0f00000000; selp.b32 %r2304, %r2303, %r2300, %p1483; mov.b32 %r2305, %f16273; add.s32 %r2306, %r2304, %r2305; mov.b32 %f10685, %r2306; neg.s32 %r2307, %r2301; setp.lt.ftz.f32 %p1484, %f16274, 0f00000000; selp.b32 %r2308, %r2307, %r2301, %p1484; mov.b32 %r2309, %f16274; add.s32 %r2310, %r2308, %r2309; mov.b32 %f10686, %r2310; neg.s32 %r2311, %r2302; setp.lt.ftz.f32 %p1485, %f16275, 0f00000000; selp.b32 %r2312, %r2311, %r2302, %p1485; mov.b32 %r2313, %f16275; add.s32 %r2314, %r2312, %r2313; mov.b32 %f10687, %r2314; abs.ftz.f32 %f10688, %f16273; setp.lt.ftz.f32 %p1486, %f10688, 0f3D000000; fma.rn.ftz.f32 %f10689, %f10677, 0f37800000, %f16273; abs.ftz.f32 %f10690, %f16274; setp.lt.ftz.f32 %p1487, %f10690, 0f3D000000; fma.rn.ftz.f32 %f10691, %f10678, 0f37800000, %f16274; abs.ftz.f32 %f10692, %f16275; setp.lt.ftz.f32 %p1488, %f10692, 0f3D000000; fma.rn.ftz.f32 %f10693, %f10683, 0f37800000, %f16275; selp.f32 %f16275, %f10693, %f10687, %p1488; selp.f32 %f16274, %f10691, %f10686, %p1487; selp.f32 %f16273, %f10689, %f10685, %p1486; $L__BB106_1665: ld.const.u64 %rd2247, [kernel_params+3496]; cvta.to.global.u64 %rd2248, %rd2247; mul.lo.s64 %rd2249, %rd439, 12; add.s64 %rd2250, %rd2248, %rd2249; st.global.f32 [%rd2250], %f16273; st.global.f32 [%rd2250+4], %f16274; st.global.f32 [%rd2250+8], %f16275; ld.const.v2.u64 {%rd2251, %rd2252}, [kernel_params+3504]; mov.u32 %r2315, 0; cvta.to.global.u64 %rd2255, %rd2251; add.s64 %rd2256, %rd2255, %rd2249; st.global.f32 [%rd2256], %f16276; st.global.f32 [%rd2256+4], %f16277; st.global.f32 [%rd2256+8], %f16278; cvta.to.global.u64 %rd2257, %rd2252; shl.b64 %rd2258, %rd439, 2; add.s64 %rd2259, %rd2257, %rd2258; st.global.u32 [%rd2259], %r2315; ld.const.v2.u64 {%rd2260, %rd2261}, [kernel_params+3520]; cvta.to.global.u64 %rd2264, %rd2260; add.s64 %rd2265, %rd2264, %rd2258; st.global.f32 [%rd2265], %f16279; cvta.to.global.u64 %rd2266, %rd2261; add.s64 %rd2267, %rd2266, %rd2258; st.global.f32 [%rd2267], %f2320; ld.const.u64 %rd2268, [kernel_params+3536]; cvta.to.global.u64 %rd2269, %rd2268; add.s64 %rd2270, %rd2269, %rd2258; st.global.f32 [%rd2270], %f2319; ld.const.u64 %rd2271, [kernel_params+3584]; cvta.to.global.u64 %rd2272, %rd2271; add.s64 %rd2273, %rd2272, %rd2258; st.global.u32 [%rd2273], %r758; ld.const.u64 %rd2274, [kernel_params+3576]; cvta.to.global.u64 %rd2275, %rd2274; add.s64 %rd2276, %rd2275, %rd2258; st.global.u32 [%rd2276], %r780; ld.const.u64 %rd2277, [kernel_params+3632]; cvta.to.global.u64 %rd2278, %rd2277; add.s64 %rd2279, %rd2278, %rd2258; st.global.u32 [%rd2279], %r760; ld.const.u64 %rd2280, [kernel_params+3624]; cvta.to.global.u64 %rd2281, %rd2280; add.s64 %rd2282, %rd2281, %rd2258; st.global.u32 [%rd2282], %r761; setp.ne.s32 %p1489, %r761, -1; setp.eq.s32 %p1490, %r762, 2; or.pred %p1491, %p1489, %p1490; selp.b32 %r2316, 0, 134217728, %p1491; ld.global.u32 %r782, [%rd29]; or.b32 %r3248, %r782, %r2316; ld.const.u64 %rd2283, [kernel_params+2584]; cvta.to.global.u64 %rd2284, %rd2283; mul.lo.s64 %rd2285, %rd2, 12; add.s64 %rd2286, %rd2284, %rd2285; ld.global.f32 %f10694, [%rd2286]; mul.ftz.f32 %f2396, %f10694, %f16244; ld.global.f32 %f10695, [%rd2286+4]; mul.ftz.f32 %f2397, %f10695, %f16243; ld.global.f32 %f10696, [%rd2286+8]; mul.ftz.f32 %f2398, %f10696, %f16242; and.b32 %r2317, %r763, 2097152; setp.eq.s32 %p1492, %r2317, 0; @%p1492 bra $L__BB106_1682; and.b32 %r2318, %r782, 100663296; setp.eq.s32 %p1493, %r2318, 0; @%p1493 bra $L__BB106_1668; ld.const.u64 %rd2287, [kernel_params+2600]; cvta.to.global.u64 %rd2288, %rd2287; add.s64 %rd2290, %rd2288, %rd2285; ld.global.f32 %f16312, [%rd2290+8]; ld.global.f32 %f16311, [%rd2290+4]; ld.global.f32 %f16310, [%rd2290]; ld.const.u64 %rd2291, [kernel_params+2608]; cvta.to.global.u64 %rd2292, %rd2291; add.s64 %rd2293, %rd2292, %rd2285; ld.global.f32 %f16315, [%rd2293+8]; ld.global.f32 %f16314, [%rd2293+4]; ld.global.f32 %f16313, [%rd2293]; bra.uni $L__BB106_1681; $L__BB106_1668: setp.eq.ftz.f32 %p1494, %f16244, 0f00000000; mov.f32 %f16315, 0f00000000; mov.f32 %f16310, %f16315; @%p1494 bra $L__BB106_1670; div.approx.ftz.f32 %f16310, %f16239, %f16244; $L__BB106_1670: setp.eq.ftz.f32 %p1495, %f16243, 0f00000000; mov.f32 %f16311, %f16315; @%p1495 bra $L__BB106_1672; div.approx.ftz.f32 %f16311, %f16241, %f16243; $L__BB106_1672: setp.eq.ftz.f32 %p1496, %f16242, 0f00000000; mov.f32 %f16312, %f16315; @%p1496 bra $L__BB106_1674; div.approx.ftz.f32 %f16312, %f16240, %f16242; $L__BB106_1674: mov.f32 %f16313, %f16315; @%p1494 bra $L__BB106_1676; div.approx.ftz.f32 %f16313, %f16236, %f16244; $L__BB106_1676: mov.f32 %f16314, %f16315; @%p1495 bra $L__BB106_1678; div.approx.ftz.f32 %f16314, %f16238, %f16243; $L__BB106_1678: @%p1496 bra $L__BB106_1680; div.approx.ftz.f32 %f16315, %f16237, %f16242; $L__BB106_1680: or.b32 %r3248, %r3248, 33554432; $L__BB106_1681: ld.const.u64 %rd2294, [kernel_params+3448]; cvta.to.global.u64 %rd2295, %rd2294; add.s64 %rd2297, %rd2295, %rd2249; st.global.f32 [%rd2297], %f16310; st.global.f32 [%rd2297+4], %f16311; st.global.f32 [%rd2297+8], %f16312; ld.const.u64 %rd2298, [kernel_params+3456]; cvta.to.global.u64 %rd2299, %rd2298; add.s64 %rd2300, %rd2299, %rd2249; st.global.f32 [%rd2300], %f16313; st.global.f32 [%rd2300+4], %f16314; st.global.f32 [%rd2300+8], %f16315; $L__BB106_1682: ld.const.u64 %rd2301, [kernel_params+2448]; cvta.to.global.u64 %rd2302, %rd2301; add.s64 %rd2304, %rd2302, %rd837; ld.global.u32 %r2319, [%rd2304]; ld.const.v2.u64 {%rd2305, %rd2306}, [kernel_params+3344]; cvta.to.global.u64 %rd2309, %rd2305; add.s64 %rd2311, %rd2309, %rd2258; st.global.u32 [%rd2311], %r2319; ld.global.u16 %rs94, [%rd278]; ld.const.v2.u64 {%rd2312, %rd2313}, [kernel_params+3360]; cvta.to.global.u64 %rd2316, %rd2313; shl.b64 %rd2317, %rd439, 1; add.s64 %rd2318, %rd2316, %rd2317; st.global.u16 [%rd2318], %rs94; ld.global.u32 %r2320, [%rd277]; cvta.to.global.u64 %rd2319, %rd2312; add.s64 %rd2320, %rd2319, %rd2258; st.global.u32 [%rd2320], %r2320; ld.global.u32 %r2321, [%rd279]; cvta.to.global.u64 %rd2321, %rd2306; add.s64 %rd2322, %rd2321, %rd2258; st.global.u32 [%rd2322], %r2321; ld.const.v2.u64 {%rd2323, %rd2324}, [kernel_params+3424]; cvta.to.global.u64 %rd2327, %rd2323; add.s64 %rd2328, %rd2327, %rd2258; st.global.u32 [%rd2328], %r3248; ld.const.u64 %rd2329, [kernel_params+2472]; cvta.to.global.u64 %rd2330, %rd2329; add.s64 %rd2332, %rd2330, %rd1709; ld.global.u16 %rs95, [%rd2332]; ld.const.u64 %rd2333, [kernel_params+3384]; cvta.to.global.u64 %rd2334, %rd2333; add.s64 %rd2335, %rd2334, %rd2317; st.global.u16 [%rd2335], %rs95; ld.const.v2.u64 {%rd2336, %rd2337}, [kernel_params+2480]; cvta.to.global.u64 %rd2340, %rd2337; add.s64 %rd2341, %rd2340, %rd1709; ld.global.u16 %rs96, [%rd2341]; ld.const.v2.u64 {%rd2342, %rd2343}, [kernel_params+3392]; cvta.to.global.u64 %rd2346, %rd2343; add.s64 %rd2347, %rd2346, %rd2317; st.global.u16 [%rd2347], %rs96; ld.const.u64 %rd2348, [kernel_params+2496]; cvta.to.global.u64 %rd2349, %rd2348; add.s64 %rd2350, %rd2349, %rd1709; ld.global.u16 %rs97, [%rd2350]; ld.const.u64 %rd2351, [kernel_params+3408]; cvta.to.global.u64 %rd2352, %rd2351; add.s64 %rd2353, %rd2352, %rd2317; st.global.u16 [%rd2353], %rs97; cvta.to.global.u64 %rd2354, %rd2336; add.s64 %rd2355, %rd2354, %rd1709; ld.global.u16 %rs98, [%rd2355]; ld.const.v2.u64 {%rd2356, %rd2357}, [kernel_params+3376]; cvta.to.global.u64 %rd2359, %rd2342; add.s64 %rd2360, %rd2359, %rd2317; st.global.u16 [%rd2360], %rs98; ld.global.u16 %rs99, [%rd280]; cvta.to.global.u64 %rd2361, %rd2356; add.s64 %rd2362, %rd2361, %rd2317; st.global.u16 [%rd2362], %rs99; cvta.to.global.u64 %rd2363, %rd2324; add.s64 %rd2365, %rd2363, %rd2249; st.global.f32 [%rd2365], %f2396; st.global.f32 [%rd2365+4], %f2397; st.global.f32 [%rd2365+8], %f2398; and.b32 %r2322, %r763, 4194304; setp.eq.s32 %p1500, %r2322, 0; @%p1500 bra $L__BB106_1684; ld.const.u64 %rd2366, [kernel_params+3440]; cvta.to.global.u64 %rd2367, %rd2366; add.s64 %rd2369, %rd2367, %rd2249; st.global.f32 [%rd2369], %f2396; st.global.f32 [%rd2369+4], %f2397; st.global.f32 [%rd2369+8], %f2398; $L__BB106_1684: ld.const.u32 %r2323, [kernel_params+1076]; ld.local.u32 %r2324, [%rd282+48]; selp.b32 %r2325, %r2323, %r2324, %p1490; cvt.u16.u32 %rs100, %r2325; add.s16 %rs101, %rs100, 1; ld.const.u64 %rd2370, [kernel_params+3472]; cvta.to.global.u64 %rd2371, %rd2370; add.s64 %rd2372, %rd2371, %rd439; st.global.u8 [%rd2372], %rs101; $L__BB106_1685: ld.const.u32 %r787, [kernel_params]; and.b32 %r2326, %r787, 25165824; setp.eq.s32 %p1502, %r2326, 0; @%p1502 bra $L__BB106_1775; and.b32 %r788, %r787, 16777216; setp.ne.s32 %p1503, %r788, 0; @%p1503 bra $L__BB106_1688; ld.global.u32 %r2327, [%rd29]; and.b32 %r2328, %r2327, 1; setp.eq.b32 %p1504, %r2328, 1; mov.pred %p1505, 0; xor.pred %p1506, %p1504, %p1505; not.pred %p1507, %p1506; @%p1507 bra $L__BB106_1775; $L__BB106_1688: add.s32 %r789, %r602, 3; ld.const.u32 %r2329, [kernel_params+1596]; setp.eq.s32 %p1508, %r2329, 0; brev.b32 %r2330, %r603; mul.lo.s32 %r2331, %r2330, 1025551850; xor.b32 %r790, %r2331, %r2330; @%p1508 bra $L__BB106_1691; ld.const.v2.u32 {%r2332, %r2333}, [kernel_params+1600]; mov.b32 %f2429, %r2333; setp.geu.ftz.f32 %p1509, %f2429, 0f3F800000; setp.lt.ftz.f32 %p1510, %f2429, 0f3F800000; ld.const.u32 %r2336, [kernel_params+1576]; selp.b32 %r2337, %r2336, %r601, %p1510; add.s16 %rs102, %rs28, 3; and.b16 %rs103, %rs102, 255; cvt.u32.u16 %r2338, %rs103; xor.b32 %r2339, %r2337, %r2338; mul.lo.s32 %r2340, %r2339, -512718531; shr.u32 %r2341, %r2337, 16; xor.b32 %r2342, %r2340, %r2341; bfe.u32 %r2343, %r2342, 4, 4; shr.u32 %r2344, %r2337, 8; xor.b32 %r2345, %r2342, %r2344; xor.b32 %r2346, %r2345, %r2343; mul.lo.s32 %r2347, %r2346, 153742143; shr.u32 %r2348, %r2337, 23; xor.b32 %r2349, %r2347, %r2348; bfe.u32 %r2350, %r2349, 1, 7; xor.b32 %r2351, %r2350, %r2349; shr.u32 %r2352, %r2337, 27; or.b32 %r2353, %r2352, 1; mul.lo.s32 %r2354, %r2353, %r2351; mul.lo.s32 %r2355, %r2354, -1074439621; bfe.u32 %r2356, %r2355, 2, 6; xor.b32 %r2357, %r2356, %r2355; mul.lo.s32 %r2358, %r2357, -1638916925; bfe.u32 %r2359, %r2358, 2, 6; xor.b32 %r2360, %r2359, %r2358; mul.lo.s32 %r2361, %r2360, -933190689; and.b32 %r2362, %r2361, 255; bfe.u32 %r2363, %r2361, 5, 3; xor.b32 %r2364, %r2363, %r2362; add.s32 %r2365, %r2332, -1; xor.b32 %r791, %r789, 61; xor.b32 %r2366, %r2337, %r791; mul.lo.s32 %r2367, %r2366, 9; shr.u32 %r2368, %r2367, 4; xor.b32 %r2369, %r2368, %r2367; mul.lo.s32 %r2370, %r2369, 668265261; add.s32 %r2371, %r2370, %r790; shr.u32 %r2372, %r2370, 16; or.b32 %r2373, %r2372, 1; mul.lo.s32 %r2374, %r2373, %r2371; mul.lo.s32 %r2375, %r2374, 89287766; xor.b32 %r2376, %r2375, %r2374; mul.lo.s32 %r2377, %r2376, 1403136100; xor.b32 %r2378, %r2377, %r2376; brev.b32 %r2379, %r2378; neg.s32 %r2380, %r2332; and.b32 %r2381, %r603, %r2380; and.b32 %r2382, %r2379, %r2365; or.b32 %r2383, %r2382, %r2381; mad.lo.s32 %r2384, %r2364, %r2332, %r2383; shl.b32 %r2385, %r2332, 8; rem.u32 %r2386, %r2384, %r2385; ld.const.u64 %rd2373, [kernel_params+2424]; cvta.to.global.u64 %rd2374, %rd2373; shl.b32 %r2387, %r2386, 1; mul.wide.u32 %rd2375, %r2387, 4; add.s64 %rd2376, %rd2374, %rd2375; ld.global.f32 %f16318, [%rd2376]; ld.global.f32 %f16319, [%rd2376+4]; @%p1509 bra $L__BB106_1695; xor.b32 %r2388, %r791, %r601; mul.lo.s32 %r2389, %r2388, 9; shr.u32 %r2390, %r2389, 4; xor.b32 %r2391, %r2390, %r2389; mul.lo.s32 %r2392, %r2391, 668265261; cvt.rn.f32.u32 %f10703, %r2392; mul.ftz.f32 %f10704, %f10703, 0f2F7FFFFE; xor.b32 %r2393, %r601, %r789; xor.b32 %r2394, %r2393, -905047700; mul.lo.s32 %r2395, %r2394, 9; shr.u32 %r2396, %r2395, 4; xor.b32 %r2397, %r2396, %r2395; mul.lo.s32 %r2398, %r2397, 668265261; cvt.rn.f32.u32 %f10705, %r2398; mul.ftz.f32 %f10706, %f10705, 0f2F7FFFFE; fma.rn.ftz.f32 %f10707, %f10704, %f2429, %f16318; fma.rn.ftz.f32 %f10708, %f10706, %f2429, %f16319; cvt.rmi.ftz.f32.f32 %f10709, %f10707; sub.ftz.f32 %f16318, %f10707, %f10709; cvt.rmi.ftz.f32.f32 %f10710, %f10708; sub.ftz.f32 %f16319, %f10708, %f10710; bra.uni $L__BB106_1695; $L__BB106_1691: shr.u32 %r2400, %r789, 16; xor.b32 %r2401, %r2400, %r789; mul.lo.s32 %r2402, %r2401, 569420461; shr.u32 %r2403, %r2402, 15; xor.b32 %r2404, %r2403, %r2402; mul.lo.s32 %r2405, %r2404, -749064809; shr.u32 %r2406, %r2405, 15; xor.b32 %r2407, %r601, %r2405; xor.b32 %r2408, %r2407, %r2406; xor.b32 %r792, %r2408, -419546133; xor.b32 %r2409, %r2408, 508809841; add.s32 %r2410, %r2409, %r790; shr.u32 %r2411, %r2409, 16; or.b32 %r2412, %r2411, 1; mul.lo.s32 %r2413, %r2412, %r2410; mul.lo.s32 %r2414, %r2413, 89287766; xor.b32 %r2415, %r2414, %r2413; mul.lo.s32 %r2416, %r2415, 1403136100; xor.b32 %r793, %r2416, %r2415; setp.eq.s32 %p1511, %r793, 0; mov.u32 %r3252, 0; @%p1511 bra $L__BB106_1694; mov.u32 %r3249, 0; mov.u64 %rd2378, sobol_burley_table; mov.u32 %r3250, %r793; mov.u32 %r3252, %r3249; $L__BB106_1693: clz.b32 %r2419, %r3250; add.s32 %r2420, %r2419, %r3249; mul.wide.u32 %rd2377, %r2420, 4; add.s64 %rd2379, %rd2378, %rd2377; ld.const.u32 %r2421, [%rd2379+128]; xor.b32 %r3252, %r2421, %r3252; add.s32 %r3249, %r2420, 1; shl.b32 %r2422, %r3250, %r2419; shl.b32 %r3250, %r2422, 1; setp.ne.s32 %p1512, %r3250, 0; @%p1512 bra $L__BB106_1693; $L__BB106_1694: xor.b32 %r2423, %r792, -525684874; brev.b32 %r2424, %r793; shr.u32 %r2425, %r2423, 16; or.b32 %r2426, %r2425, 1; mul.lo.s32 %r2427, %r2424, 1025551850; xor.b32 %r2428, %r2427, %r2424; add.s32 %r2429, %r2428, %r2423; mul.lo.s32 %r2430, %r2429, %r2426; mul.lo.s32 %r2431, %r2430, 89287766; xor.b32 %r2432, %r2431, %r2430; xor.b32 %r2433, %r792, -1802089138; mul.lo.s32 %r2434, %r3252, 1025551850; xor.b32 %r2435, %r2434, %r3252; add.s32 %r2436, %r2435, %r2433; shr.u32 %r2437, %r2433, 16; or.b32 %r2438, %r2437, 1; mul.lo.s32 %r2439, %r2436, %r2438; mul.lo.s32 %r2440, %r2439, 89287766; xor.b32 %r2441, %r2440, %r2439; mul.lo.s32 %r2442, %r2441, 1403136100; xor.b32 %r2443, %r2442, %r2441; brev.b32 %r2444, %r2443; cvt.rn.f32.u32 %f10711, %r2444; mul.lo.s32 %r2445, %r2432, 1403136100; xor.b32 %r2446, %r2445, %r2432; brev.b32 %r2447, %r2446; cvt.rn.f32.u32 %f10712, %r2447; mul.ftz.f32 %f16319, %f10711, 0f2F7FFFFE; mul.ftz.f32 %f16318, %f10712, 0f2F7FFFFE; $L__BB106_1695: ld.const.f32 %f2442, [kernel_params+1548]; ld.local.u32 %r801, [%rd7+172]; setp.lt.s32 %p1513, %r801, 1; mov.f32 %f16326, 0f00000000; mov.f32 %f16327, %f16326; mov.f32 %f16328, %f16326; mov.f32 %f16329, %f16326; mov.f32 %f16330, %f16326; mov.f32 %f16331, %f16326; @%p1513 bra $L__BB106_1712; add.s32 %r2449, %r801, -1; and.b32 %r3256, %r801, 3; setp.lt.u32 %p1514, %r2449, 3; mov.f32 %f16331, 0f00000000; mov.u32 %r3255, 0; mov.f32 %f16330, %f16331; mov.f32 %f16329, %f16331; mov.f32 %f16328, %f16331; mov.f32 %f16327, %f16331; mov.f32 %f16326, %f16331; @%p1514 bra $L__BB106_1707; sub.s32 %r3254, %r801, %r3256; mov.f32 %f16331, 0f00000000; mov.u32 %r3255, 0; mov.f32 %f16330, %f16331; mov.f32 %f16329, %f16331; mov.f32 %f16328, %f16331; mov.f32 %f16327, %f16331; mov.f32 %f16326, %f16331; $L__BB106_1698: cvt.s64.s32 %rd456, %r3255; mul.wide.s32 %rd2380, %r3255, 80; add.s64 %rd2381, %rd1, %rd2380; add.s64 %rd457, %rd2381, 316; ld.local.u32 %r2451, [%rd2381+316]; add.s32 %r2452, %r2451, -2; setp.gt.u32 %p1515, %r2452, 6; @%p1515 bra $L__BB106_1700; ld.local.v4.f32 {%f10732, %f10733, %f10734, %f10735}, [%rd457+-12]; fma.rn.ftz.f32 %f16329, %f2442, %f10732, %f16329; fma.rn.ftz.f32 %f16330, %f2442, %f10733, %f16330; fma.rn.ftz.f32 %f16331, %f2442, %f10734, %f16331; add.ftz.f32 %f10739, %f10732, %f10733; add.ftz.f32 %f10740, %f10734, %f10739; mul.ftz.f32 %f10741, %f10740, 0f3EAAAAAB; abs.ftz.f32 %f10742, %f10741; ld.local.f32 %f10743, [%rd457+8]; ld.local.f32 %f10744, [%rd457+12]; ld.local.f32 %f10745, [%rd457+16]; fma.rn.ftz.f32 %f16326, %f10742, %f10743, %f16326; fma.rn.ftz.f32 %f16327, %f10742, %f10744, %f16327; fma.rn.ftz.f32 %f16328, %f10742, %f10745, %f16328; $L__BB106_1700: ld.local.u32 %r2453, [%rd457+80]; add.s32 %r2454, %r2453, -2; setp.gt.u32 %p1516, %r2454, 6; @%p1516 bra $L__BB106_1702; ld.local.v4.f32 {%f10746, %f10747, %f10748, %f10749}, [%rd457+68]; fma.rn.ftz.f32 %f16329, %f2442, %f10746, %f16329; fma.rn.ftz.f32 %f16330, %f2442, %f10747, %f16330; fma.rn.ftz.f32 %f16331, %f2442, %f10748, %f16331; add.ftz.f32 %f10753, %f10746, %f10747; add.ftz.f32 %f10754, %f10748, %f10753; mul.ftz.f32 %f10755, %f10754, 0f3EAAAAAB; abs.ftz.f32 %f10756, %f10755; ld.local.f32 %f10757, [%rd457+88]; ld.local.f32 %f10758, [%rd457+92]; ld.local.f32 %f10759, [%rd457+96]; fma.rn.ftz.f32 %f16326, %f10756, %f10757, %f16326; fma.rn.ftz.f32 %f16327, %f10756, %f10758, %f16327; fma.rn.ftz.f32 %f16328, %f10756, %f10759, %f16328; $L__BB106_1702: ld.local.u32 %r2455, [%rd457+160]; add.s32 %r2456, %r2455, -2; setp.gt.u32 %p1517, %r2456, 6; @%p1517 bra $L__BB106_1704; ld.local.v4.f32 {%f10760, %f10761, %f10762, %f10763}, [%rd457+148]; fma.rn.ftz.f32 %f16329, %f2442, %f10760, %f16329; fma.rn.ftz.f32 %f16330, %f2442, %f10761, %f16330; fma.rn.ftz.f32 %f16331, %f2442, %f10762, %f16331; add.ftz.f32 %f10767, %f10760, %f10761; add.ftz.f32 %f10768, %f10762, %f10767; mul.ftz.f32 %f10769, %f10768, 0f3EAAAAAB; abs.ftz.f32 %f10770, %f10769; ld.local.f32 %f10771, [%rd457+168]; ld.local.f32 %f10772, [%rd457+172]; ld.local.f32 %f10773, [%rd457+176]; fma.rn.ftz.f32 %f16326, %f10770, %f10771, %f16326; fma.rn.ftz.f32 %f16327, %f10770, %f10772, %f16327; fma.rn.ftz.f32 %f16328, %f10770, %f10773, %f16328; $L__BB106_1704: ld.local.u32 %r2457, [%rd457+240]; add.s32 %r2458, %r2457, -2; setp.gt.u32 %p1518, %r2458, 6; @%p1518 bra $L__BB106_1706; ld.local.v4.f32 {%f10774, %f10775, %f10776, %f10777}, [%rd457+228]; fma.rn.ftz.f32 %f16329, %f2442, %f10774, %f16329; fma.rn.ftz.f32 %f16330, %f2442, %f10775, %f16330; fma.rn.ftz.f32 %f16331, %f2442, %f10776, %f16331; add.ftz.f32 %f10781, %f10774, %f10775; add.ftz.f32 %f10782, %f10776, %f10781; mul.ftz.f32 %f10783, %f10782, 0f3EAAAAAB; abs.ftz.f32 %f10784, %f10783; ld.local.f32 %f10785, [%rd457+248]; ld.local.f32 %f10786, [%rd457+252]; ld.local.f32 %f10787, [%rd457+256]; fma.rn.ftz.f32 %f16326, %f10784, %f10785, %f16326; fma.rn.ftz.f32 %f16327, %f10784, %f10786, %f16327; fma.rn.ftz.f32 %f16328, %f10784, %f10787, %f16328; $L__BB106_1706: cvt.u32.u64 %r2459, %rd456; add.s32 %r3255, %r2459, 4; add.s32 %r3254, %r3254, -4; setp.ne.s32 %p1519, %r3254, 0; @%p1519 bra $L__BB106_1698; $L__BB106_1707: setp.eq.s32 %p1520, %r3256, 0; @%p1520 bra $L__BB106_1712; mul.wide.s32 %rd2382, %r3255, 80; add.s64 %rd2383, %rd1, %rd2382; add.s64 %rd3030, %rd2383, 332; $L__BB106_1709: .pragma "nounroll"; ld.local.u32 %r2460, [%rd3030+-16]; add.s32 %r2461, %r2460, -2; setp.gt.u32 %p1521, %r2461, 6; @%p1521 bra $L__BB106_1711; ld.local.f32 %f10788, [%rd3030+-28]; ld.local.f32 %f10789, [%rd3030+-24]; ld.local.f32 %f10790, [%rd3030+-20]; fma.rn.ftz.f32 %f16329, %f2442, %f10788, %f16329; fma.rn.ftz.f32 %f16330, %f2442, %f10789, %f16330; fma.rn.ftz.f32 %f16331, %f2442, %f10790, %f16331; add.ftz.f32 %f10791, %f10788, %f10789; add.ftz.f32 %f10792, %f10790, %f10791; mul.ftz.f32 %f10793, %f10792, 0f3EAAAAAB; abs.ftz.f32 %f10794, %f10793; ld.local.f32 %f10795, [%rd3030+-8]; ld.local.f32 %f10796, [%rd3030+-4]; ld.local.f32 %f10797, [%rd3030]; fma.rn.ftz.f32 %f16326, %f10794, %f10795, %f16326; fma.rn.ftz.f32 %f16327, %f10794, %f10796, %f16327; fma.rn.ftz.f32 %f16328, %f10794, %f10797, %f16328; $L__BB106_1711: add.s64 %rd3030, %rd3030, 80; add.s32 %r3256, %r3256, -1; setp.ne.s32 %p1522, %r3256, 0; @%p1522 bra $L__BB106_1709; $L__BB106_1712: setp.eq.ftz.f32 %p1523, %f16327, 0f00000000; setp.eq.ftz.f32 %p1524, %f16326, 0f00000000; and.pred %p1525, %p1524, %p1523; setp.eq.ftz.f32 %p1526, %f16328, 0f00000000; and.pred %p1527, %p1526, %p1525; @%p1527 bra $L__BB106_1714; bra.uni $L__BB106_1713; $L__BB106_1714: ld.local.v2.f32 {%f16381, %f16382}, [%rd7+-72]; ld.local.f32 %f16380, [%rd7+-76]; bra.uni $L__BB106_1715; $L__BB106_1713: mul.ftz.f32 %f10798, %f16327, %f16327; fma.rn.ftz.f32 %f10799, %f16326, %f16326, %f10798; fma.rn.ftz.f32 %f10800, %f16328, %f16328, %f10799; sqrt.approx.ftz.f32 %f10801, %f10800; div.approx.ftz.f32 %f16382, %f16328, %f10801; div.approx.ftz.f32 %f16381, %f16327, %f10801; div.approx.ftz.f32 %f16380, %f16326, %f10801; $L__BB106_1715: mul.ftz.f32 %f10804, %f16318, 0f40C90FDB; cos.approx.ftz.f32 %f10805, %f10804; sqrt.approx.ftz.f32 %f10806, %f16319; mul.ftz.f32 %f10807, %f10806, %f10805; sin.approx.ftz.f32 %f10808, %f10804; mul.ftz.f32 %f10809, %f10806, %f10808; mul.ftz.f32 %f10810, %f10807, %f10807; mov.f32 %f10811, 0f3F800000; sub.ftz.f32 %f10812, %f10811, %f10810; mul.ftz.f32 %f10813, %f10809, %f10809; sub.ftz.f32 %f10814, %f10812, %f10813; mov.f32 %f10815, 0f00000000; max.ftz.f32 %f10816, %f10814, %f10815; sqrt.approx.ftz.f32 %f10817, %f10816; setp.neu.ftz.f32 %p1528, %f16380, %f16381; setp.neu.ftz.f32 %p1529, %f16380, %f16382; or.pred %p1530, %p1528, %p1529; sub.ftz.f32 %f10818, %f16382, %f16381; neg.ftz.f32 %f10819, %f16381; neg.ftz.f32 %f10820, %f16382; selp.f32 %f10821, %f10820, %f16382, %p1530; add.ftz.f32 %f10822, %f16380, %f10821; selp.f32 %f10823, %f16381, %f10819, %p1530; sub.ftz.f32 %f10824, %f10823, %f16380; mul.ftz.f32 %f10825, %f10822, %f10822; fma.rn.ftz.f32 %f10826, %f10818, %f10818, %f10825; fma.rn.ftz.f32 %f10827, %f10824, %f10824, %f10826; sqrt.approx.ftz.f32 %f10828, %f10827; div.approx.ftz.f32 %f10829, %f10818, %f10828; div.approx.ftz.f32 %f10830, %f10822, %f10828; div.approx.ftz.f32 %f10831, %f10824, %f10828; mul.ftz.f32 %f10832, %f16381, %f10831; mul.ftz.f32 %f10833, %f16382, %f10830; sub.ftz.f32 %f10834, %f10832, %f10833; mul.ftz.f32 %f10835, %f16382, %f10829; mul.ftz.f32 %f10836, %f16380, %f10831; sub.ftz.f32 %f10837, %f10835, %f10836; mul.ftz.f32 %f10838, %f16380, %f10830; mul.ftz.f32 %f10839, %f16381, %f10829; sub.ftz.f32 %f10840, %f10838, %f10839; mul.ftz.f32 %f10841, %f10809, %f10834; mul.ftz.f32 %f10842, %f10809, %f10837; mul.ftz.f32 %f10843, %f10809, %f10840; fma.rn.ftz.f32 %f10844, %f10807, %f10829, %f10841; fma.rn.ftz.f32 %f10845, %f10807, %f10830, %f10842; fma.rn.ftz.f32 %f10846, %f10807, %f10831, %f10843; fma.rn.ftz.f32 %f2542, %f16380, %f10817, %f10844; fma.rn.ftz.f32 %f2543, %f16381, %f10817, %f10845; fma.rn.ftz.f32 %f2544, %f16382, %f10817, %f10846; ld.local.v4.f32 {%f16408, %f16409, %f16410, %f10850}, [%rd1]; ld.local.u32 %r811, [%rd7+-24]; and.b32 %r812, %r811, 1; setp.eq.s32 %p1531, %r812, 0; ld.local.u32 %r2462, [%rd7+-40]; setp.gt.s32 %p1532, %r2462, -1; mov.u16 %rs154, 1; or.pred %p1533, %p1531, %p1532; @%p1533 bra $L__BB106_1758; ld.const.u64 %rd2386, [kernel_params+2160]; cvta.to.global.u64 %rd2387, %rd2386; ld.local.u32 %r2463, [%rd7+-12]; mul.wide.s32 %rd2388, %r2463, 208; add.s64 %rd2389, %rd2387, %rd2388; add.s64 %rd461, %rd2389, 184; ld.global.f32 %f2548, [%rd2389+184]; setp.leu.ftz.f32 %p1534, %f2548, 0f00000000; @%p1534 bra $L__BB106_1758; ld.local.f32 %f10851, [%rd7+-76]; ld.local.v2.f32 {%f10852, %f10853}, [%rd7+-72]; mul.ftz.f32 %f10856, %f2543, %f10852; fma.rn.ftz.f32 %f10857, %f2542, %f10851, %f10856; fma.rn.ftz.f32 %f10858, %f2544, %f10853, %f10857; setp.lt.ftz.f32 %p1535, %f10858, 0f00000000; neg.ftz.f32 %f10859, %f10858; selp.f32 %f2549, %f10859, %f10858, %p1535; @%p1535 bra $L__BB106_1719; bra.uni $L__BB106_1718; $L__BB106_1719: ld.local.v2.f32 {%f10862, %f10863}, [%rd7+-64]; neg.ftz.f32 %f16383, %f10862; neg.ftz.f32 %f16384, %f10863; ld.local.f32 %f10866, [%rd7+-56]; neg.ftz.f32 %f16385, %f10866; bra.uni $L__BB106_1720; $L__BB106_1718: ld.local.v2.f32 {%f16383, %f16384}, [%rd7+-64]; ld.local.f32 %f16385, [%rd7+-56]; $L__BB106_1720: mul.ftz.f32 %f10867, %f2543, %f16384; fma.rn.ftz.f32 %f10868, %f2542, %f16383, %f10867; fma.rn.ftz.f32 %f2559, %f2544, %f16385, %f10868; add.ftz.f32 %f10869, %f2549, %f2559; setp.lt.ftz.f32 %p1536, %f2549, %f2548; selp.f32 %f10870, 0f40000000, 0f3F800000, %p1536; mov.f32 %f10871, 0f3F800000; selp.f32 %f10872, %f10869, %f2559, %p1536; div.approx.ftz.f32 %f10873, %f10872, %f2548; sub.ftz.f32 %f10874, %f10870, %f10873; mov.f32 %f10875, 0f00000000; max.ftz.f32 %f10876, %f10874, %f10875; min.ftz.f32 %f2560, %f10876, %f10871; setp.leu.ftz.f32 %p1537, %f2560, 0f00000000; @%p1537 bra $L__BB106_1758; setp.eq.s32 %p1538, %r811, 65; ld.local.s32 %rd462, [%rd7+-28]; @%p1538 bra $L__BB106_1723; bra.uni $L__BB106_1722; $L__BB106_1723: ld.global.u32 %r813, [%rd461+-28]; ld.global.u32 %r814, [%rd461+-32]; shl.b32 %r2475, %r814, 1; cvt.rn.f32.s32 %f10877, %r2475; ld.local.f32 %f10878, [%rd7+-4]; mul.ftz.f32 %f10879, %f10878, %f10877; cvt.rzi.ftz.s32.f32 %r2476, %f10879; add.s32 %r2477, %r2475, -1; min.s32 %r815, %r2476, %r2477; cvt.rn.f32.s32 %f10880, %r815; sub.ftz.f32 %f2579, %f10879, %f10880; ld.const.u64 %rd2412, [kernel_params+2312]; cvta.to.global.u64 %rd463, %rd2412; ld.global.u32 %r3263, [%rd461+-20]; mul.wide.u32 %rd2413, %r3263, 16; add.s64 %rd2414, %rd463, %rd2413; ld.global.u16 %rs152, [%rd2414+12]; ld.global.u32 %r3262, [%rd2414+8]; ld.global.u64 %rd3050, [%rd2414]; setp.eq.s64 %p1539, %rd3050, 11; mov.u16 %rs150, %rs152; mov.u32 %r3257, %r3262; @%p1539 bra $L__BB106_1729; mov.u64 %rd3031, %rd3050; mov.u32 %r3257, %r3262; mov.u16 %rs150, %rs152; mov.u32 %r3258, %r3263; $L__BB106_1725: setp.eq.s64 %p1540, %rd3031, 0; @%p1540 bra $L__BB106_1727; add.s32 %r3258, %r3258, 2; bra.uni $L__BB106_1728; $L__BB106_1727: setp.eq.s16 %p1541, %rs150, 0; mov.u32 %r3261, -1; mov.u32 %r3258, %r3257; @%p1541 bra $L__BB106_1730; $L__BB106_1728: mul.wide.u32 %rd2415, %r3258, 16; add.s64 %rd2416, %rd463, %rd2415; ld.global.u16 %rs150, [%rd2416+12]; ld.global.u32 %r3257, [%rd2416+8]; ld.global.u64 %rd3031, [%rd2416]; setp.ne.s64 %p1542, %rd3031, 11; @%p1542 bra $L__BB106_1725; $L__BB106_1729: setp.eq.s16 %p1543, %rs150, 0; selp.b32 %r3261, -1, %r3257, %p1543; $L__BB106_1730: ld.const.u64 %rd2417, [kernel_params+2232]; cvta.to.global.u64 %rd2418, %rd2417; shl.b64 %rd2419, %rd462, 4; add.s64 %rd2420, %rd2418, %rd2419; ld.global.v4.u32 {%r2479, %r2480, %r2481, %r2482}, [%rd2420]; setp.eq.s32 %p1544, %r815, %r814; @%p1544 bra $L__BB106_1732; bra.uni $L__BB106_1731; $L__BB106_1732: ld.const.u64 %rd2426, [kernel_params+2256]; cvta.to.global.u64 %rd2427, %rd2426; mul.wide.u32 %rd2428, %r2482, 12; add.s64 %rd3032, %rd2427, %rd2428; add.s32 %r2489, %r2482, 1; mul.wide.u32 %rd2429, %r2489, 12; add.s64 %rd3035, %rd2427, %rd2429; add.s32 %r2490, %r2482, 2; mul.wide.u32 %rd2430, %r2490, 12; add.s64 %rd3038, %rd2427, %rd2430; add.s64 %rd3033, %rd3032, 4; add.s64 %rd3034, %rd3032, 8; add.s64 %rd3036, %rd3035, 4; add.s64 %rd3037, %rd3035, 8; add.s64 %rd3039, %rd3038, 4; add.s64 %rd3040, %rd3038, 8; bra.uni $L__BB106_1733; $L__BB106_1722: cvt.u32.u64 %r2464, %rd462; ld.const.v2.u64 {%rd2390, %rd2391}, [kernel_params+2224]; cvta.to.global.u64 %rd2394, %rd2391; mul.wide.s32 %rd2395, %r2464, 16; add.s64 %rd2396, %rd2394, %rd2395; ld.global.v4.u32 {%r2465, %r2466, %r2467, %r2468}, [%rd2396]; ld.const.u64 %rd2397, [kernel_params+2256]; cvta.to.global.u64 %rd2398, %rd2397; mul.wide.u32 %rd2399, %r2468, 12; add.s64 %rd2400, %rd2398, %rd2399; ld.global.f32 %f16395, [%rd2400]; ld.global.f32 %f16396, [%rd2400+4]; ld.global.f32 %f16397, [%rd2400+8]; add.s32 %r2473, %r2468, 1; mul.wide.u32 %rd2401, %r2473, 12; add.s64 %rd2402, %rd2398, %rd2401; ld.global.f32 %f16398, [%rd2402]; ld.global.f32 %f16399, [%rd2402+4]; ld.global.f32 %f16400, [%rd2402+8]; add.s32 %r2474, %r2468, 2; mul.wide.u32 %rd2403, %r2474, 12; add.s64 %rd2404, %rd2398, %rd2403; ld.global.f32 %f16401, [%rd2404]; ld.global.f32 %f16402, [%rd2404+4]; ld.global.f32 %f16403, [%rd2404+8]; cvta.to.global.u64 %rd2405, %rd2390; mul.wide.u32 %rd2406, %r2465, 12; add.s64 %rd2407, %rd2405, %rd2406; ld.global.f32 %f16386, [%rd2407]; ld.global.f32 %f16387, [%rd2407+4]; ld.global.f32 %f16388, [%rd2407+8]; mul.wide.u32 %rd2408, %r2466, 12; add.s64 %rd2409, %rd2405, %rd2408; ld.global.f32 %f16389, [%rd2409]; ld.global.f32 %f16390, [%rd2409+4]; ld.global.f32 %f16391, [%rd2409+8]; mul.wide.u32 %rd2410, %r2467, 12; add.s64 %rd2411, %rd2405, %rd2410; ld.global.f32 %f16392, [%rd2411]; ld.global.f32 %f16393, [%rd2411+4]; ld.global.f32 %f16394, [%rd2411+8]; bra.uni $L__BB106_1750; $L__BB106_1731: setp.gt.s32 %p1545, %r815, %r814; selp.b32 %r2483, -1, 0, %p1545; add.s32 %r2484, %r815, %r2483; mad.lo.s32 %r2485, %r2484, %r813, %r3261; ld.const.u64 %rd2421, [kernel_params+2336]; cvta.to.global.u64 %rd2422, %rd2421; add.s32 %r2486, %r2479, %r2485; mul.wide.u32 %rd2423, %r2486, 12; add.s64 %rd3032, %rd2422, %rd2423; add.s32 %r2487, %r2480, %r2485; mul.wide.u32 %rd2424, %r2487, 12; add.s64 %rd3035, %rd2422, %rd2424; add.s32 %r2488, %r2481, %r2485; mul.wide.u32 %rd2425, %r2488, 12; add.s64 %rd3038, %rd2422, %rd2425; add.s64 %rd3033, %rd3032, 4; add.s64 %rd3034, %rd3032, 8; add.s64 %rd3036, %rd3035, 4; add.s64 %rd3037, %rd3035, 8; add.s64 %rd3039, %rd3038, 4; add.s64 %rd3040, %rd3038, 8; $L__BB106_1733: ld.global.f32 %f2580, [%rd3040]; ld.global.f32 %f2581, [%rd3039]; ld.global.f32 %f2582, [%rd3038]; ld.global.f32 %f2583, [%rd3037]; ld.global.f32 %f2584, [%rd3036]; ld.global.f32 %f2585, [%rd3035]; ld.global.f32 %f2586, [%rd3034]; ld.global.f32 %f2587, [%rd3033]; ld.global.f32 %f2588, [%rd3032]; add.s32 %r830, %r815, 1; setp.eq.s32 %p1546, %r830, %r814; @%p1546 bra $L__BB106_1735; bra.uni $L__BB106_1734; $L__BB106_1735: ld.const.u64 %rd2436, [kernel_params+2256]; cvta.to.global.u64 %rd2437, %rd2436; mul.wide.u32 %rd2438, %r2482, 12; add.s64 %rd3041, %rd2437, %rd2438; add.s32 %r2497, %r2482, 1; mul.wide.u32 %rd2439, %r2497, 12; add.s64 %rd3044, %rd2437, %rd2439; add.s32 %r2498, %r2482, 2; mul.wide.u32 %rd2440, %r2498, 12; add.s64 %rd3047, %rd2437, %rd2440; add.s64 %rd3042, %rd3041, 4; add.s64 %rd3043, %rd3041, 8; add.s64 %rd3045, %rd3044, 4; add.s64 %rd3046, %rd3044, 8; add.s64 %rd3048, %rd3047, 4; add.s64 %rd3049, %rd3047, 8; bra.uni $L__BB106_1736; $L__BB106_1734: setp.ge.s32 %p1547, %r815, %r814; selp.b32 %r2491, -1, 0, %p1547; add.s32 %r2492, %r830, %r2491; mad.lo.s32 %r2493, %r2492, %r813, %r3261; ld.const.u64 %rd2431, [kernel_params+2336]; cvta.to.global.u64 %rd2432, %rd2431; add.s32 %r2494, %r2479, %r2493; mul.wide.u32 %rd2433, %r2494, 12; add.s64 %rd3041, %rd2432, %rd2433; add.s32 %r2495, %r2480, %r2493; mul.wide.u32 %rd2434, %r2495, 12; add.s64 %rd3044, %rd2432, %rd2434; add.s32 %r2496, %r2481, %r2493; mul.wide.u32 %rd2435, %r2496, 12; add.s64 %rd3047, %rd2432, %rd2435; add.s64 %rd3042, %rd3041, 4; add.s64 %rd3043, %rd3041, 8; add.s64 %rd3045, %rd3044, 4; add.s64 %rd3046, %rd3044, 8; add.s64 %rd3048, %rd3047, 4; add.s64 %rd3049, %rd3047, 8; $L__BB106_1736: sub.ftz.f32 %f2589, %f10871, %f2579; ld.global.f32 %f10882, [%rd3041]; mul.ftz.f32 %f10883, %f2579, %f10882; ld.global.f32 %f10884, [%rd3042]; mul.ftz.f32 %f10885, %f2579, %f10884; ld.global.f32 %f10886, [%rd3043]; mul.ftz.f32 %f10887, %f2579, %f10886; fma.rn.ftz.f32 %f16395, %f2589, %f2588, %f10883; fma.rn.ftz.f32 %f16396, %f2589, %f2587, %f10885; fma.rn.ftz.f32 %f16397, %f2589, %f2586, %f10887; ld.global.f32 %f10888, [%rd3044]; mul.ftz.f32 %f10889, %f2579, %f10888; ld.global.f32 %f10890, [%rd3045]; mul.ftz.f32 %f10891, %f2579, %f10890; ld.global.f32 %f10892, [%rd3046]; mul.ftz.f32 %f10893, %f2579, %f10892; fma.rn.ftz.f32 %f16398, %f2589, %f2585, %f10889; fma.rn.ftz.f32 %f16399, %f2589, %f2584, %f10891; fma.rn.ftz.f32 %f16400, %f2589, %f2583, %f10893; ld.global.f32 %f10894, [%rd3047]; mul.ftz.f32 %f10895, %f2579, %f10894; ld.global.f32 %f10896, [%rd3048]; mul.ftz.f32 %f10897, %f2579, %f10896; ld.global.f32 %f10898, [%rd3049]; mul.ftz.f32 %f10899, %f2579, %f10898; fma.rn.ftz.f32 %f16401, %f2589, %f2582, %f10895; fma.rn.ftz.f32 %f16402, %f2589, %f2581, %f10897; fma.rn.ftz.f32 %f16403, %f2589, %f2580, %f10899; setp.eq.s64 %p1548, %rd3050, 12; @%p1548 bra $L__BB106_1742; mov.u32 %r3266, -1; $L__BB106_1738: setp.eq.s64 %p1549, %rd3050, 0; @%p1549 bra $L__BB106_1740; add.s32 %r3263, %r3263, 2; bra.uni $L__BB106_1741; $L__BB106_1740: setp.eq.s16 %p1550, %rs152, 0; mov.u32 %r3263, %r3262; @%p1550 bra $L__BB106_1743; $L__BB106_1741: mul.wide.u32 %rd2441, %r3263, 16; add.s64 %rd2442, %rd463, %rd2441; ld.global.u16 %rs152, [%rd2442+12]; ld.global.u32 %r3262, [%rd2442+8]; ld.global.u64 %rd3050, [%rd2442]; setp.ne.s64 %p1551, %rd3050, 12; @%p1551 bra $L__BB106_1738; $L__BB106_1742: setp.eq.s16 %p1552, %rs152, 0; selp.b32 %r3266, -1, %r3262, %p1552; $L__BB106_1743: @%p1544 bra $L__BB106_1745; bra.uni $L__BB106_1744; $L__BB106_1745: ld.const.u64 %rd2448, [kernel_params+2224]; cvta.to.global.u64 %rd2449, %rd2448; mul.wide.u32 %rd2450, %r2479, 12; add.s64 %rd3051, %rd2449, %rd2450; mul.wide.u32 %rd2451, %r2480, 12; add.s64 %rd3054, %rd2449, %rd2451; mul.wide.u32 %rd2452, %r2481, 12; add.s64 %rd3057, %rd2449, %rd2452; add.s64 %rd3052, %rd3051, 4; add.s64 %rd3053, %rd3051, 8; add.s64 %rd3055, %rd3054, 4; add.s64 %rd3056, %rd3054, 8; add.s64 %rd3058, %rd3057, 4; add.s64 %rd3059, %rd3057, 8; bra.uni $L__BB106_1746; $L__BB106_1744: setp.gt.s32 %p1554, %r815, %r814; selp.b32 %r2500, -1, 0, %p1554; add.s32 %r2501, %r815, %r2500; mad.lo.s32 %r2502, %r2501, %r813, %r3266; ld.const.u64 %rd2443, [kernel_params+2336]; cvta.to.global.u64 %rd2444, %rd2443; add.s32 %r2503, %r2502, %r2479; mul.wide.u32 %rd2445, %r2503, 12; add.s64 %rd3051, %rd2444, %rd2445; add.s32 %r2504, %r2502, %r2480; mul.wide.u32 %rd2446, %r2504, 12; add.s64 %rd3054, %rd2444, %rd2446; add.s32 %r2505, %r2502, %r2481; mul.wide.u32 %rd2447, %r2505, 12; add.s64 %rd3057, %rd2444, %rd2447; add.s64 %rd3052, %rd3051, 4; add.s64 %rd3053, %rd3051, 8; add.s64 %rd3055, %rd3054, 4; add.s64 %rd3056, %rd3054, 8; add.s64 %rd3058, %rd3057, 4; add.s64 %rd3059, %rd3057, 8; $L__BB106_1746: ld.global.f32 %f2599, [%rd3059]; ld.global.f32 %f2600, [%rd3058]; ld.global.f32 %f2601, [%rd3057]; ld.global.f32 %f2602, [%rd3056]; ld.global.f32 %f2603, [%rd3055]; ld.global.f32 %f2604, [%rd3054]; ld.global.f32 %f2605, [%rd3053]; ld.global.f32 %f2606, [%rd3052]; ld.global.f32 %f2607, [%rd3051]; @%p1546 bra $L__BB106_1748; bra.uni $L__BB106_1747; $L__BB106_1748: ld.const.u64 %rd2458, [kernel_params+2224]; cvta.to.global.u64 %rd2459, %rd2458; mul.wide.u32 %rd2460, %r2479, 12; add.s64 %rd3060, %rd2459, %rd2460; mul.wide.u32 %rd2461, %r2480, 12; add.s64 %rd3063, %rd2459, %rd2461; mul.wide.u32 %rd2462, %r2481, 12; add.s64 %rd3066, %rd2459, %rd2462; add.s64 %rd3061, %rd3060, 4; add.s64 %rd3062, %rd3060, 8; add.s64 %rd3064, %rd3063, 4; add.s64 %rd3065, %rd3063, 8; add.s64 %rd3067, %rd3066, 4; add.s64 %rd3068, %rd3066, 8; bra.uni $L__BB106_1749; $L__BB106_1747: setp.ge.s32 %p1556, %r815, %r814; selp.b32 %r2506, -1, 0, %p1556; add.s32 %r2507, %r830, %r2506; mad.lo.s32 %r2508, %r2507, %r813, %r3266; ld.const.u64 %rd2453, [kernel_params+2336]; cvta.to.global.u64 %rd2454, %rd2453; add.s32 %r2509, %r2508, %r2479; mul.wide.u32 %rd2455, %r2509, 12; add.s64 %rd3060, %rd2454, %rd2455; add.s32 %r2510, %r2508, %r2480; mul.wide.u32 %rd2456, %r2510, 12; add.s64 %rd3063, %rd2454, %rd2456; add.s32 %r2511, %r2508, %r2481; mul.wide.u32 %rd2457, %r2511, 12; add.s64 %rd3066, %rd2454, %rd2457; add.s64 %rd3061, %rd3060, 4; add.s64 %rd3062, %rd3060, 8; add.s64 %rd3064, %rd3063, 4; add.s64 %rd3065, %rd3063, 8; add.s64 %rd3067, %rd3066, 4; add.s64 %rd3068, %rd3066, 8; $L__BB106_1749: ld.global.f32 %f10900, [%rd3060]; mul.ftz.f32 %f10901, %f2579, %f10900; ld.global.f32 %f10902, [%rd3061]; mul.ftz.f32 %f10903, %f2579, %f10902; ld.global.f32 %f10904, [%rd3062]; mul.ftz.f32 %f10905, %f2579, %f10904; fma.rn.ftz.f32 %f16386, %f2589, %f2607, %f10901; fma.rn.ftz.f32 %f16387, %f2589, %f2606, %f10903; fma.rn.ftz.f32 %f16388, %f2589, %f2605, %f10905; ld.global.f32 %f10906, [%rd3063]; mul.ftz.f32 %f10907, %f2579, %f10906; ld.global.f32 %f10908, [%rd3064]; mul.ftz.f32 %f10909, %f2579, %f10908; ld.global.f32 %f10910, [%rd3065]; mul.ftz.f32 %f10911, %f2579, %f10910; fma.rn.ftz.f32 %f16389, %f2589, %f2604, %f10907; fma.rn.ftz.f32 %f16390, %f2589, %f2603, %f10909; fma.rn.ftz.f32 %f16391, %f2589, %f2602, %f10911; ld.global.f32 %f10912, [%rd3066]; mul.ftz.f32 %f10913, %f2579, %f10912; ld.global.f32 %f10914, [%rd3067]; mul.ftz.f32 %f10915, %f2579, %f10914; ld.global.f32 %f10916, [%rd3068]; mul.ftz.f32 %f10917, %f2579, %f10916; fma.rn.ftz.f32 %f16392, %f2589, %f2601, %f10913; fma.rn.ftz.f32 %f16393, %f2589, %f2600, %f10915; fma.rn.ftz.f32 %f16394, %f2589, %f2599, %f10917; $L__BB106_1750: ld.local.f32 %f2635, [%rd7+-20]; sub.ftz.f32 %f10919, %f10871, %f2635; ld.local.f32 %f10920, [%rd7+-16]; sub.ftz.f32 %f2636, %f10919, %f10920; mul.ftz.f32 %f10921, %f16395, %f2636; mul.ftz.f32 %f10922, %f16396, %f2636; mul.ftz.f32 %f10923, %f16397, %f2636; fma.rn.ftz.f32 %f10924, %f16398, %f2635, %f10921; fma.rn.ftz.f32 %f10925, %f16399, %f2635, %f10922; fma.rn.ftz.f32 %f10926, %f16400, %f2635, %f10923; fma.rn.ftz.f32 %f2637, %f16401, %f10920, %f10924; fma.rn.ftz.f32 %f2638, %f16402, %f10920, %f10925; fma.rn.ftz.f32 %f2639, %f16403, %f10920, %f10926; mul.ftz.f32 %f10927, %f16386, %f2636; mul.ftz.f32 %f10928, %f16387, %f2636; mul.ftz.f32 %f10929, %f16388, %f2636; fma.rn.ftz.f32 %f10930, %f16389, %f2635, %f10927; fma.rn.ftz.f32 %f10931, %f16390, %f2635, %f10928; fma.rn.ftz.f32 %f10932, %f16391, %f2635, %f10929; fma.rn.ftz.f32 %f16404, %f16392, %f10920, %f10930; fma.rn.ftz.f32 %f16405, %f16393, %f10920, %f10931; fma.rn.ftz.f32 %f16406, %f16394, %f10920, %f10932; ld.local.u32 %r839, [%rd7+-32]; and.b32 %r2512, %r839, 4; setp.ne.s32 %p1557, %r2512, 0; @%p1557 bra $L__BB106_1754; and.b32 %r2513, %r839, 2; setp.eq.s32 %p1558, %r2513, 0; @%p1558 bra $L__BB106_1753; bra.uni $L__BB106_1752; $L__BB106_1753: ld.global.v4.f32 {%f10960, %f10961, %f10962, %f10963}, [%rd461+-184]; ld.global.v4.f32 {%f10967, %f10968, %f10969, %f10970}, [%rd461+-168]; ld.global.v4.f32 {%f10974, %f10975, %f10976, %f10977}, [%rd461+-152]; mul.ftz.f32 %f10981, %f16405, %f10961; fma.rn.ftz.f32 %f10982, %f16404, %f10960, %f10981; fma.rn.ftz.f32 %f2646, %f16406, %f10962, %f10982; mul.ftz.f32 %f10983, %f16405, %f10968; fma.rn.ftz.f32 %f10984, %f16404, %f10967, %f10983; fma.rn.ftz.f32 %f2647, %f16406, %f10969, %f10984; mul.ftz.f32 %f10985, %f16405, %f10975; fma.rn.ftz.f32 %f10986, %f16404, %f10974, %f10985; fma.rn.ftz.f32 %f16406, %f16406, %f10976, %f10986; mov.f32 %f16404, %f2646; mov.f32 %f16405, %f2647; bra.uni $L__BB106_1754; $L__BB106_1752: ld.local.v4.f32 {%f10933, %f10934, %f10935, %f10936}, [%rd7+56]; mul.ftz.f32 %f10940, %f16405, %f10934; fma.rn.ftz.f32 %f10941, %f16404, %f10933, %f10940; fma.rn.ftz.f32 %f2643, %f16406, %f10935, %f10941; ld.local.v4.f32 {%f10942, %f10943, %f10944, %f10945}, [%rd7+72]; mul.ftz.f32 %f10949, %f16405, %f10943; fma.rn.ftz.f32 %f10950, %f16404, %f10942, %f10949; fma.rn.ftz.f32 %f2644, %f16406, %f10944, %f10950; ld.local.v4.f32 {%f10951, %f10952, %f10953, %f10954}, [%rd7+88]; mul.ftz.f32 %f10958, %f16405, %f10952; fma.rn.ftz.f32 %f10959, %f16404, %f10951, %f10958; fma.rn.ftz.f32 %f16406, %f16406, %f10953, %f10959; mov.f32 %f16404, %f2643; mov.f32 %f16405, %f2644; $L__BB106_1754: sub.ftz.f32 %f2652, %f16395, %f16401; sub.ftz.f32 %f10987, %f16392, %f16386; sub.ftz.f32 %f2653, %f16396, %f16402; sub.ftz.f32 %f10988, %f16393, %f16387; mul.ftz.f32 %f10989, %f10988, %f2653; fma.rn.ftz.f32 %f10990, %f10987, %f2652, %f10989; sub.ftz.f32 %f2654, %f16397, %f16403; sub.ftz.f32 %f10991, %f16394, %f16388; fma.rn.ftz.f32 %f10992, %f10991, %f2654, %f10990; sub.ftz.f32 %f2655, %f16398, %f16401; sub.ftz.f32 %f10993, %f16392, %f16389; sub.ftz.f32 %f2656, %f16399, %f16402; sub.ftz.f32 %f10994, %f16393, %f16390; mul.ftz.f32 %f10995, %f10994, %f2656; fma.rn.ftz.f32 %f10996, %f10993, %f2655, %f10995; sub.ftz.f32 %f2657, %f16400, %f16403; sub.ftz.f32 %f10997, %f16394, %f16391; fma.rn.ftz.f32 %f10998, %f10997, %f2657, %f10996; sub.ftz.f32 %f2658, %f16398, %f16395; sub.ftz.f32 %f10999, %f16389, %f16386; sub.ftz.f32 %f2659, %f16399, %f16396; sub.ftz.f32 %f11000, %f16390, %f16387; mul.ftz.f32 %f11001, %f11000, %f2659; fma.rn.ftz.f32 %f11002, %f10999, %f2658, %f11001; sub.ftz.f32 %f2660, %f16400, %f16397; sub.ftz.f32 %f11003, %f16391, %f16388; fma.rn.ftz.f32 %f11004, %f11003, %f2660, %f11002; mul.ftz.f32 %f11005, %f10992, %f2636; add.ftz.f32 %f11006, %f2636, 0fBF800000; add.ftz.f32 %f11007, %f10992, %f10998; add.ftz.f32 %f11008, %f11004, %f11007; mul.ftz.f32 %f11009, %f2636, %f11008; mul.ftz.f32 %f11010, %f2635, %f11009; fma.rn.ftz.f32 %f11011, %f11005, %f11006, %f11010; mul.ftz.f32 %f11012, %f2635, %f10998; add.ftz.f32 %f11013, %f2635, 0fBF800000; fma.rn.ftz.f32 %f11014, %f11013, %f11012, %f11011; mul.ftz.f32 %f11015, %f16384, %f16405; fma.rn.ftz.f32 %f11016, %f16383, %f16404, %f11015; fma.rn.ftz.f32 %f11017, %f16385, %f16406, %f11016; setp.gt.ftz.f32 %p1559, %f11017, 0f00000000; sub.ftz.f32 %f2661, %f16395, %f16398; sub.ftz.f32 %f2662, %f16396, %f16399; sub.ftz.f32 %f2663, %f16397, %f16400; sub.ftz.f32 %f2664, %f16401, %f16395; sub.ftz.f32 %f2665, %f16402, %f16396; sub.ftz.f32 %f2666, %f16403, %f16397; sub.ftz.f32 %f2667, %f16401, %f16398; sub.ftz.f32 %f2668, %f16402, %f16399; sub.ftz.f32 %f2669, %f16403, %f16400; mul.ftz.f32 %f2670, %f11014, 0f3F000000; @%p1559 bra $L__BB106_1756; bra.uni $L__BB106_1755; $L__BB106_1756: mul.ftz.f32 %f11070, %f16387, %f2659; fma.rn.ftz.f32 %f11071, %f16386, %f2658, %f11070; fma.rn.ftz.f32 %f11072, %f16388, %f2660, %f11071; mul.ftz.f32 %f11073, %f16387, %f2665; fma.rn.ftz.f32 %f11074, %f16386, %f2664, %f11073; fma.rn.ftz.f32 %f11075, %f16388, %f2666, %f11074; max.ftz.f32 %f11076, %f11072, %f11075; max.ftz.f32 %f11078, %f11076, %f10875; mul.ftz.f32 %f11079, %f16390, %f2662; fma.rn.ftz.f32 %f11080, %f16389, %f2661, %f11079; fma.rn.ftz.f32 %f11081, %f16391, %f2663, %f11080; mul.ftz.f32 %f11082, %f16390, %f2668; fma.rn.ftz.f32 %f11083, %f16389, %f2667, %f11082; fma.rn.ftz.f32 %f11084, %f16391, %f2669, %f11083; max.ftz.f32 %f11085, %f11081, %f11084; max.ftz.f32 %f11086, %f11085, %f10875; mul.ftz.f32 %f11087, %f16393, %f2653; fma.rn.ftz.f32 %f11088, %f16392, %f2652, %f11087; fma.rn.ftz.f32 %f11089, %f16394, %f2654, %f11088; mul.ftz.f32 %f11090, %f16393, %f2656; fma.rn.ftz.f32 %f11091, %f16392, %f2655, %f11090; fma.rn.ftz.f32 %f11092, %f16394, %f2657, %f11091; max.ftz.f32 %f11093, %f11089, %f11092; max.ftz.f32 %f11094, %f11093, %f10875; sub.ftz.f32 %f11095, %f16395, %f2637; sub.ftz.f32 %f11096, %f16396, %f2638; mul.ftz.f32 %f11097, %f16387, %f11096; fma.rn.ftz.f32 %f11098, %f16386, %f11095, %f11097; sub.ftz.f32 %f11099, %f16397, %f2639; fma.rn.ftz.f32 %f11100, %f16388, %f11099, %f11098; add.ftz.f32 %f11101, %f11100, %f11078; max.ftz.f32 %f11102, %f11101, %f10875; sub.ftz.f32 %f11103, %f16398, %f2637; sub.ftz.f32 %f11104, %f16399, %f2638; mul.ftz.f32 %f11105, %f16390, %f11104; fma.rn.ftz.f32 %f11106, %f16389, %f11103, %f11105; sub.ftz.f32 %f11107, %f16400, %f2639; fma.rn.ftz.f32 %f11108, %f16391, %f11107, %f11106; add.ftz.f32 %f11109, %f11108, %f11086; max.ftz.f32 %f11110, %f11109, %f10875; sub.ftz.f32 %f11111, %f16401, %f2637; sub.ftz.f32 %f11112, %f16402, %f2638; mul.ftz.f32 %f11113, %f16393, %f11112; fma.rn.ftz.f32 %f11114, %f16392, %f11111, %f11113; sub.ftz.f32 %f11115, %f16403, %f2639; fma.rn.ftz.f32 %f11116, %f16394, %f11115, %f11114; add.ftz.f32 %f11117, %f11116, %f11094; max.ftz.f32 %f11118, %f11117, %f10875; min.ftz.f32 %f11119, %f11102, %f11110; min.ftz.f32 %f11120, %f11119, %f11118; max.ftz.f32 %f16407, %f11120, %f2670; bra.uni $L__BB106_1757; $L__BB106_1755: mul.ftz.f32 %f11018, %f16387, %f2662; fma.rn.ftz.f32 %f11019, %f16386, %f2661, %f11018; fma.rn.ftz.f32 %f11020, %f16388, %f2663, %f11019; mul.ftz.f32 %f11021, %f16387, %f2653; fma.rn.ftz.f32 %f11022, %f16386, %f2652, %f11021; fma.rn.ftz.f32 %f11023, %f16388, %f2654, %f11022; max.ftz.f32 %f11024, %f11020, %f11023; max.ftz.f32 %f11026, %f11024, %f10875; mul.ftz.f32 %f11027, %f16390, %f2659; fma.rn.ftz.f32 %f11028, %f16389, %f2658, %f11027; fma.rn.ftz.f32 %f11029, %f16391, %f2660, %f11028; mul.ftz.f32 %f11030, %f16390, %f2656; fma.rn.ftz.f32 %f11031, %f16389, %f2655, %f11030; fma.rn.ftz.f32 %f11032, %f16391, %f2657, %f11031; max.ftz.f32 %f11033, %f11029, %f11032; max.ftz.f32 %f11034, %f11033, %f10875; mul.ftz.f32 %f11035, %f16393, %f2665; fma.rn.ftz.f32 %f11036, %f16392, %f2664, %f11035; fma.rn.ftz.f32 %f11037, %f16394, %f2666, %f11036; mul.ftz.f32 %f11038, %f16393, %f2668; fma.rn.ftz.f32 %f11039, %f16392, %f2667, %f11038; fma.rn.ftz.f32 %f11040, %f16394, %f2669, %f11039; max.ftz.f32 %f11041, %f11037, %f11040; max.ftz.f32 %f11042, %f11041, %f10875; sub.ftz.f32 %f11043, %f2637, %f16395; sub.ftz.f32 %f11044, %f2638, %f16396; mul.ftz.f32 %f11045, %f16387, %f11044; fma.rn.ftz.f32 %f11046, %f16386, %f11043, %f11045; sub.ftz.f32 %f11047, %f2639, %f16397; fma.rn.ftz.f32 %f11048, %f16388, %f11047, %f11046; add.ftz.f32 %f11049, %f11048, %f11026; max.ftz.f32 %f11050, %f11049, %f10875; sub.ftz.f32 %f11051, %f2637, %f16398; sub.ftz.f32 %f11052, %f2638, %f16399; mul.ftz.f32 %f11053, %f16390, %f11052; fma.rn.ftz.f32 %f11054, %f16389, %f11051, %f11053; sub.ftz.f32 %f11055, %f2639, %f16400; fma.rn.ftz.f32 %f11056, %f16391, %f11055, %f11054; add.ftz.f32 %f11057, %f11056, %f11034; max.ftz.f32 %f11058, %f11057, %f10875; sub.ftz.f32 %f11059, %f2637, %f16401; sub.ftz.f32 %f11060, %f2638, %f16402; mul.ftz.f32 %f11061, %f16393, %f11060; fma.rn.ftz.f32 %f11062, %f16392, %f11059, %f11061; sub.ftz.f32 %f11063, %f2639, %f16403; fma.rn.ftz.f32 %f11064, %f16394, %f11063, %f11062; add.ftz.f32 %f11065, %f11064, %f11042; max.ftz.f32 %f11066, %f11065, %f10875; min.ftz.f32 %f11067, %f11050, %f11058; min.ftz.f32 %f11068, %f11067, %f11066; neg.ftz.f32 %f11069, %f11068; min.ftz.f32 %f16407, %f11069, %f2670; $L__BB106_1757: mul.ftz.f32 %f11121, %f16404, %f16407; mul.ftz.f32 %f11122, %f16405, %f16407; mul.ftz.f32 %f11123, %f16406, %f16407; fma.rn.ftz.f32 %f16408, %f2560, %f11121, %f16408; fma.rn.ftz.f32 %f16409, %f2560, %f11122, %f16409; fma.rn.ftz.f32 %f16410, %f2560, %f11123, %f16410; setp.gt.ftz.f32 %p1560, %f2559, 0f00000000; selp.u16 %rs154, 1, 0, %p1560; $L__BB106_1758: setp.eq.s16 %p1561, %rs154, 0; @%p1561 bra $L__BB106_1767; @%p1531 bra $L__BB106_1767; ld.const.u64 %rd2463, [kernel_params+2232]; cvta.to.global.u64 %rd2464, %rd2463; ld.local.v2.u32 {%r2514, %r2515}, [%rd7+-32]; mul.wide.s32 %rd2465, %r2515, 16; add.s64 %rd2466, %rd2464, %rd2465; ld.const.u64 %rd2467, [kernel_params+2256]; cvta.to.global.u64 %rd2468, %rd2467; ld.global.u32 %r2517, [%rd2466+12]; mul.wide.u32 %rd2469, %r2517, 12; add.s64 %rd2470, %rd2468, %rd2469; ld.global.f32 %f2683, [%rd2470]; ld.global.f32 %f2684, [%rd2470+4]; ld.global.f32 %f2685, [%rd2470+8]; add.s32 %r2518, %r2517, 1; mul.wide.u32 %rd2471, %r2518, 12; add.s64 %rd2472, %rd2468, %rd2471; ld.global.f32 %f2686, [%rd2472]; ld.global.f32 %f2687, [%rd2472+4]; ld.global.f32 %f2688, [%rd2472+8]; add.s32 %r2519, %r2517, 2; mul.wide.u32 %rd2473, %r2519, 12; add.s64 %rd2474, %rd2468, %rd2473; ld.global.f32 %f2689, [%rd2474]; ld.global.f32 %f2690, [%rd2474+4]; ld.global.f32 %f2691, [%rd2474+8]; and.b32 %r2520, %r2514, 4; setp.ne.s32 %p1563, %r2520, 0; mov.f32 %f16423, %f16408; mov.f32 %f16424, %f16409; mov.f32 %f16425, %f16410; mov.f32 %f16426, %f2542; mov.f32 %f16427, %f2543; mov.f32 %f16428, %f2544; @%p1563 bra $L__BB106_1765; and.b32 %r2521, %r2514, 2; setp.eq.s32 %p1564, %r2521, 0; @%p1564 bra $L__BB106_1763; bra.uni $L__BB106_1762; $L__BB106_1763: ld.const.u64 %rd2475, [kernel_params+2160]; cvta.to.global.u64 %rd2476, %rd2475; ld.local.u32 %r2522, [%rd7+-12]; mul.wide.s32 %rd2477, %r2522, 208; add.s64 %rd2478, %rd2476, %rd2477; ld.global.v4.f32 {%f16411, %f16412, %f16413, %f16414}, [%rd2478+48]; ld.global.v4.f32 {%f16415, %f16416, %f16417, %f16418}, [%rd2478+64]; ld.global.v4.f32 {%f16419, %f16420, %f16421, %f16422}, [%rd2478+80]; bra.uni $L__BB106_1764; $L__BB106_1762: ld.local.v4.f32 {%f16411, %f16412, %f16413, %f16414}, [%rd7+104]; ld.local.v4.f32 {%f16415, %f16416, %f16417, %f16418}, [%rd7+120]; ld.local.v4.f32 {%f16419, %f16420, %f16421, %f16422}, [%rd7+136]; $L__BB106_1764: mul.ftz.f32 %f11148, %f16409, %f16412; fma.rn.ftz.f32 %f11149, %f16408, %f16411, %f11148; fma.rn.ftz.f32 %f11150, %f16410, %f16413, %f11149; mul.ftz.f32 %f11151, %f16409, %f16416; fma.rn.ftz.f32 %f11152, %f16408, %f16415, %f11151; fma.rn.ftz.f32 %f11153, %f16410, %f16417, %f11152; mul.ftz.f32 %f11154, %f16409, %f16420; fma.rn.ftz.f32 %f11155, %f16408, %f16419, %f11154; fma.rn.ftz.f32 %f11156, %f16410, %f16421, %f11155; add.ftz.f32 %f16423, %f16414, %f11150; add.ftz.f32 %f16424, %f16418, %f11153; add.ftz.f32 %f16425, %f16422, %f11156; mul.ftz.f32 %f11157, %f2543, %f16412; fma.rn.ftz.f32 %f11158, %f2542, %f16411, %f11157; mul.ftz.f32 %f11159, %f2543, %f16416; fma.rn.ftz.f32 %f11160, %f2542, %f16415, %f11159; mul.ftz.f32 %f11161, %f2543, %f16420; fma.rn.ftz.f32 %f11162, %f2542, %f16419, %f11161; fma.rn.ftz.f32 %f16426, %f2544, %f16413, %f11158; fma.rn.ftz.f32 %f16427, %f2544, %f16417, %f11160; fma.rn.ftz.f32 %f16428, %f2544, %f16421, %f11162; $L__BB106_1765: sub.ftz.f32 %f11163, %f2683, %f16423; sub.ftz.f32 %f11164, %f2684, %f16424; sub.ftz.f32 %f11165, %f2685, %f16425; sub.ftz.f32 %f11166, %f2686, %f16423; sub.ftz.f32 %f11167, %f2687, %f16424; sub.ftz.f32 %f11168, %f2688, %f16425; sub.ftz.f32 %f11169, %f2689, %f16423; sub.ftz.f32 %f11170, %f2690, %f16424; sub.ftz.f32 %f11171, %f2691, %f16425; sub.ftz.f32 %f11172, %f11169, %f11163; sub.ftz.f32 %f11173, %f11170, %f11164; sub.ftz.f32 %f11174, %f11171, %f11165; sub.ftz.f32 %f11175, %f11163, %f11166; sub.ftz.f32 %f11176, %f11164, %f11167; sub.ftz.f32 %f11177, %f11165, %f11168; sub.ftz.f32 %f11178, %f11166, %f11169; sub.ftz.f32 %f11179, %f11167, %f11170; sub.ftz.f32 %f11180, %f11168, %f11171; add.ftz.f32 %f11181, %f11169, %f11163; add.ftz.f32 %f11182, %f11170, %f11164; add.ftz.f32 %f11183, %f11171, %f11165; mul.ftz.f32 %f11184, %f11182, %f11174; mul.ftz.f32 %f11185, %f11183, %f11173; sub.ftz.f32 %f11186, %f11184, %f11185; mul.ftz.f32 %f11187, %f11183, %f11172; mul.ftz.f32 %f11188, %f11181, %f11174; sub.ftz.f32 %f11189, %f11187, %f11188; mul.ftz.f32 %f11190, %f11181, %f11173; mul.ftz.f32 %f11191, %f11182, %f11172; sub.ftz.f32 %f11192, %f11190, %f11191; mul.ftz.f32 %f11193, %f16427, %f11189; fma.rn.ftz.f32 %f11194, %f16426, %f11186, %f11193; fma.rn.ftz.f32 %f11195, %f16428, %f11192, %f11194; add.ftz.f32 %f11196, %f11163, %f11166; add.ftz.f32 %f11197, %f11164, %f11167; add.ftz.f32 %f11198, %f11165, %f11168; mul.ftz.f32 %f11199, %f11197, %f11177; mul.ftz.f32 %f11200, %f11198, %f11176; sub.ftz.f32 %f11201, %f11199, %f11200; mul.ftz.f32 %f11202, %f11198, %f11175; mul.ftz.f32 %f11203, %f11196, %f11177; sub.ftz.f32 %f11204, %f11202, %f11203; mul.ftz.f32 %f11205, %f11196, %f11176; mul.ftz.f32 %f11206, %f11197, %f11175; sub.ftz.f32 %f11207, %f11205, %f11206; mul.ftz.f32 %f11208, %f16427, %f11204; fma.rn.ftz.f32 %f11209, %f16426, %f11201, %f11208; fma.rn.ftz.f32 %f11210, %f16428, %f11207, %f11209; add.ftz.f32 %f11211, %f11166, %f11169; add.ftz.f32 %f11212, %f11167, %f11170; add.ftz.f32 %f11213, %f11168, %f11171; mul.ftz.f32 %f11214, %f11212, %f11180; mul.ftz.f32 %f11215, %f11213, %f11179; sub.ftz.f32 %f11216, %f11214, %f11215; mul.ftz.f32 %f11217, %f11213, %f11178; mul.ftz.f32 %f11218, %f11211, %f11180; sub.ftz.f32 %f11219, %f11217, %f11218; mul.ftz.f32 %f11220, %f11211, %f11179; mul.ftz.f32 %f11221, %f11212, %f11178; sub.ftz.f32 %f11222, %f11220, %f11221; mul.ftz.f32 %f11223, %f16427, %f11219; fma.rn.ftz.f32 %f11224, %f16426, %f11216, %f11223; fma.rn.ftz.f32 %f11225, %f16428, %f11222, %f11224; add.ftz.f32 %f11226, %f11195, %f11210; add.ftz.f32 %f11227, %f11225, %f11226; abs.ftz.f32 %f11228, %f11227; mul.ftz.f32 %f11229, %f11228, 0f34000000; min.ftz.f32 %f11230, %f11210, %f11225; min.ftz.f32 %f11231, %f11195, %f11230; max.ftz.f32 %f11232, %f11210, %f11225; max.ftz.f32 %f11233, %f11195, %f11232; setp.ge.ftz.f32 %p1565, %f11231, %f11229; mul.ftz.f32 %f11234, %f11228, 0fB4000000; setp.le.ftz.f32 %p1566, %f11233, %f11234; or.pred %p1567, %p1565, %p1566; @%p1567 bra $L__BB106_1767; ld.local.v2.f32 {%f11235, %f11236}, [%rd7+-64]; mul.ftz.f32 %f11239, %f11235, 0f43800000; cvt.rzi.ftz.s32.f32 %r2523, %f11239; mul.ftz.f32 %f11240, %f11236, 0f43800000; cvt.rzi.ftz.s32.f32 %r2524, %f11240; ld.local.f32 %f11241, [%rd7+-56]; mul.ftz.f32 %f11242, %f11241, 0f43800000; cvt.rzi.ftz.s32.f32 %r2525, %f11242; neg.s32 %r2526, %r2523; setp.lt.ftz.f32 %p1568, %f16408, 0f00000000; selp.b32 %r2527, %r2526, %r2523, %p1568; mov.b32 %r2528, %f16408; add.s32 %r2529, %r2527, %r2528; mov.b32 %f11243, %r2529; neg.s32 %r2530, %r2524; setp.lt.ftz.f32 %p1569, %f16409, 0f00000000; selp.b32 %r2531, %r2530, %r2524, %p1569; mov.b32 %r2532, %f16409; add.s32 %r2533, %r2531, %r2532; mov.b32 %f11244, %r2533; neg.s32 %r2534, %r2525; setp.lt.ftz.f32 %p1570, %f16410, 0f00000000; selp.b32 %r2535, %r2534, %r2525, %p1570; mov.b32 %r2536, %f16410; add.s32 %r2537, %r2535, %r2536; mov.b32 %f11245, %r2537; abs.ftz.f32 %f11246, %f16408; setp.lt.ftz.f32 %p1571, %f11246, 0f3D000000; fma.rn.ftz.f32 %f11247, %f11235, 0f37800000, %f16408; abs.ftz.f32 %f11248, %f16409; setp.lt.ftz.f32 %p1572, %f11248, 0f3D000000; fma.rn.ftz.f32 %f11249, %f11236, 0f37800000, %f16409; abs.ftz.f32 %f11250, %f16410; setp.lt.ftz.f32 %p1573, %f11250, 0f3D000000; fma.rn.ftz.f32 %f11251, %f11241, 0f37800000, %f16410; selp.f32 %f16410, %f11251, %f11245, %p1573; selp.f32 %f16409, %f11249, %f11244, %p1572; selp.f32 %f16408, %f11247, %f11243, %p1571; $L__BB106_1767: ld.const.f32 %f2755, [kernel_params+1540]; ld.local.f32 %f2756, [%rd7+-4]; ld.local.u32 %r2538, [%rd7+-12]; selp.b32 %r841, -1, %r2538, %p1561; ld.local.u32 %r2539, [%rd7+-28]; selp.b32 %r842, -1, %r2539, %p1561; ld.const.u64 %rd2479, [kernel_params+4376]; cvta.to.global.u64 %rd2480, %rd2479; atom.global.add.u32 %r2540, [%rd2480], 1; ld.const.u64 %rd2481, [kernel_params+4256]; cvta.to.global.u64 %rd2482, %rd2481; add.s64 %rd2483, %rd2482, 12; atom.global.add.u32 %r2541, [%rd2483], 1; ld.const.u64 %rd2484, [kernel_params+3416]; cvta.to.global.u64 %rd2485, %rd2484; cvt.s64.s32 %rd577, %r2540; mul.wide.s32 %rd2486, %r2540, 2; add.s64 %rd2487, %rd2485, %rd2486; mov.u16 %rs107, 3; st.global.u16 [%rd2487], %rs107; and.b32 %r2542, %r787, 131072; setp.eq.s32 %p1575, %r2542, 0; @%p1575 bra $L__BB106_1770; mov.u32 %r3267, 0; mov.u64 %rd2489, kernel_params; $L__BB106_1769: mul.wide.s32 %rd2488, %r3267, 16; add.s64 %rd2490, %rd2489, %rd2488; ld.const.v2.u64 {%rd2491, %rd2492}, [%rd2490+2768]; cvta.to.global.u64 %rd2495, %rd2492; add.s64 %rd2497, %rd2495, %rd837; ld.global.u32 %r2544, [%rd2497]; cvta.to.global.u64 %rd2498, %rd2491; add.s64 %rd2499, %rd2498, %rd837; ld.global.u32 %r2545, [%rd2499]; ld.const.v2.u64 {%rd2500, %rd2501}, [%rd2490+3744]; cvta.to.global.u64 %rd2504, %rd2500; shl.b64 %rd2505, %rd577, 2; add.s64 %rd2506, %rd2504, %rd2505; st.global.u32 [%rd2506], %r2545; cvta.to.global.u64 %rd2507, %rd2501; add.s64 %rd2508, %rd2507, %rd2505; st.global.u32 [%rd2508], %r2544; add.s32 %r3267, %r3267, 1; setp.ne.s32 %p1576, %r2544, -1; @%p1576 bra $L__BB106_1769; $L__BB106_1770: ld.const.u64 %rd2509, [kernel_params+3496]; cvta.to.global.u64 %rd2510, %rd2509; mul.lo.s64 %rd2511, %rd577, 12; add.s64 %rd2512, %rd2510, %rd2511; st.global.f32 [%rd2512], %f16408; st.global.f32 [%rd2512+4], %f16409; st.global.f32 [%rd2512+8], %f16410; ld.const.v2.u64 {%rd2513, %rd2514}, [kernel_params+3504]; mov.u32 %r2546, 0; cvta.to.global.u64 %rd2517, %rd2513; add.s64 %rd2518, %rd2517, %rd2511; st.global.f32 [%rd2518], %f2542; st.global.f32 [%rd2518+4], %f2543; st.global.f32 [%rd2518+8], %f2544; cvta.to.global.u64 %rd2519, %rd2514; shl.b64 %rd2520, %rd577, 2; add.s64 %rd2521, %rd2519, %rd2520; st.global.u32 [%rd2521], %r2546; ld.const.v2.u64 {%rd2522, %rd2523}, [kernel_params+3520]; cvta.to.global.u64 %rd2526, %rd2522; add.s64 %rd2527, %rd2526, %rd2520; st.global.f32 [%rd2527], %f2755; cvta.to.global.u64 %rd2528, %rd2523; add.s64 %rd2529, %rd2528, %rd2520; st.global.f32 [%rd2529], %f2756; ld.const.u64 %rd2530, [kernel_params+3536]; cvta.to.global.u64 %rd2531, %rd2530; add.s64 %rd2532, %rd2531, %rd2520; st.global.u32 [%rd2532], %r2546; ld.const.u64 %rd2533, [kernel_params+3584]; cvta.to.global.u64 %rd2534, %rd2533; add.s64 %rd2535, %rd2534, %rd2520; st.global.u32 [%rd2535], %r841; ld.const.u64 %rd2536, [kernel_params+3576]; cvta.to.global.u64 %rd2537, %rd2536; add.s64 %rd2538, %rd2537, %rd2520; st.global.u32 [%rd2538], %r842; ld.const.u64 %rd2539, [kernel_params+3632]; cvta.to.global.u64 %rd2540, %rd2539; add.s64 %rd2541, %rd2540, %rd2520; mov.u32 %r2547, -1; st.global.u32 [%rd2541], %r2547; ld.const.u64 %rd2542, [kernel_params+3624]; cvta.to.global.u64 %rd2543, %rd2542; add.s64 %rd2544, %rd2543, %rd2520; st.global.u32 [%rd2544], %r2547; ld.const.v2.u64 {%rd2545, %rd2546}, [kernel_params+2464]; cvta.to.global.u64 %rd2549, %rd2545; add.s64 %rd2551, %rd2549, %rd1709; ld.global.u16 %rs48, [%rd2551]; cvta.to.global.u64 %rd2552, %rd2546; add.s64 %rd2553, %rd2552, %rd1709; ld.global.u16 %rs49, [%rd2553]; ld.global.u32 %r845, [%rd29]; ld.const.u64 %rd578, [kernel_params+2584]; ld.local.u32 %r846, [%rd7+-36]; and.b32 %r2548, %r846, 524288; setp.ne.s32 %p1577, %r2548, 0; mov.f32 %f16435, %f10811; mov.f32 %f16436, %f10811; mov.f32 %f16437, %f10811; @%p1577 bra $L__BB106_1773; and.b32 %r2549, %r846, 512; setp.eq.s32 %p1578, %r2549, 0; mov.f32 %f16435, %f10815; mov.f32 %f16436, %f10815; mov.f32 %f16437, %f10815; @%p1578 bra $L__BB106_1773; ld.local.v2.f32 {%f16436, %f16437}, [%rd7+208]; ld.local.f32 %f16435, [%rd7+204]; $L__BB106_1773: sub.ftz.f32 %f11261, %f10811, %f16435; sub.ftz.f32 %f11262, %f10811, %f16436; sub.ftz.f32 %f11263, %f10811, %f16437; cvt.ftz.sat.f32.f32 %f11264, %f11261; cvt.ftz.sat.f32.f32 %f11265, %f11262; cvt.ftz.sat.f32.f32 %f11266, %f11263; cvta.to.global.u64 %rd2554, %rd578; mul.lo.s64 %rd2555, %rd2, 12; add.s64 %rd2556, %rd2554, %rd2555; ld.global.f32 %f11267, [%rd2556]; mul.ftz.f32 %f11268, %f11264, %f11267; ld.global.f32 %f11269, [%rd2556+4]; mul.ftz.f32 %f11270, %f11265, %f11269; ld.global.f32 %f11271, [%rd2556+8]; mul.ftz.f32 %f11272, %f11266, %f11271; ld.const.u64 %rd2557, [kernel_params+2448]; cvta.to.global.u64 %rd2558, %rd2557; add.s64 %rd2560, %rd2558, %rd837; ld.global.u32 %r2550, [%rd2560]; ld.const.v2.u64 {%rd2561, %rd2562}, [kernel_params+3344]; cvta.to.global.u64 %rd2565, %rd2561; add.s64 %rd2567, %rd2565, %rd2520; st.global.u32 [%rd2567], %r2550; ld.global.u16 %rs108, [%rd278]; ld.const.v2.u64 {%rd2568, %rd2569}, [kernel_params+3360]; cvta.to.global.u64 %rd2572, %rd2569; shl.b64 %rd2573, %rd577, 1; add.s64 %rd2574, %rd2572, %rd2573; st.global.u16 [%rd2574], %rs108; ld.global.u32 %r2551, [%rd277]; cvta.to.global.u64 %rd2575, %rd2568; add.s64 %rd2576, %rd2575, %rd2520; st.global.u32 [%rd2576], %r2551; ld.global.u32 %r2552, [%rd279]; cvta.to.global.u64 %rd2577, %rd2562; add.s64 %rd2578, %rd2577, %rd2520; st.global.u32 [%rd2578], %r2552; ld.const.v2.u64 {%rd2579, %rd2580}, [kernel_params+3424]; cvta.to.global.u64 %rd2583, %rd2579; add.s64 %rd2584, %rd2583, %rd2520; or.b32 %r2553, %r845, 268435456; st.global.u32 [%rd2584], %r2553; ld.const.v2.u64 {%rd2585, %rd2586}, [kernel_params+3376]; cvta.to.global.u64 %rd2589, %rd2585; add.s64 %rd2590, %rd2589, %rd2573; st.global.u16 [%rd2590], %rs48; cvta.to.global.u64 %rd2591, %rd2586; add.s64 %rd2592, %rd2591, %rd2573; st.global.u16 [%rd2592], %rs49; cvta.to.global.u64 %rd2593, %rd2580; add.s64 %rd2595, %rd2593, %rd2511; st.global.f32 [%rd2595], %f11268; st.global.f32 [%rd2595+4], %f11270; st.global.f32 [%rd2595+8], %f11272; setp.eq.s32 %p1579, %r788, 0; @%p1579 bra $L__BB106_1775; ld.const.u64 %rd2596, [kernel_params+3440]; cvta.to.global.u64 %rd2597, %rd2596; add.s64 %rd2599, %rd2597, %rd2511; st.global.f32 [%rd2599], %f16329; st.global.f32 [%rd2599+4], %f16330; st.global.f32 [%rd2599+8], %f16331; $L__BB106_1775: ld.local.u32 %r847, [%rd7+-36]; and.b32 %r2555, %r847, 20; setp.eq.s32 %p1580, %r2555, 0; mov.u32 %r3294, 0; @%p1580 bra $L__BB106_2405; add.s32 %r848, %r602, 3; ld.const.u32 %r2556, [kernel_params+1596]; setp.eq.s32 %p1581, %r2556, 0; brev.b32 %r2557, %r603; mul.lo.s32 %r2558, %r2557, 1025551850; xor.b32 %r849, %r2558, %r2557; @%p1581 bra $L__BB106_1779; ld.const.v2.u32 {%r2559, %r2560}, [kernel_params+1600]; mov.b32 %f2763, %r2560; setp.geu.ftz.f32 %p1582, %f2763, 0f3F800000; setp.lt.ftz.f32 %p1583, %f2763, 0f3F800000; ld.const.u32 %r2563, [kernel_params+1576]; selp.b32 %r2564, %r2563, %r601, %p1583; add.s16 %rs109, %rs28, 3; and.b16 %rs110, %rs109, 255; cvt.u32.u16 %r2565, %rs110; xor.b32 %r2566, %r2564, %r2565; mul.lo.s32 %r2567, %r2566, -512718531; shr.u32 %r2568, %r2564, 16; xor.b32 %r2569, %r2567, %r2568; bfe.u32 %r2570, %r2569, 4, 4; shr.u32 %r2571, %r2564, 8; xor.b32 %r2572, %r2569, %r2571; xor.b32 %r2573, %r2572, %r2570; mul.lo.s32 %r2574, %r2573, 153742143; shr.u32 %r2575, %r2564, 23; xor.b32 %r2576, %r2574, %r2575; bfe.u32 %r2577, %r2576, 1, 7; xor.b32 %r2578, %r2577, %r2576; shr.u32 %r2579, %r2564, 27; or.b32 %r2580, %r2579, 1; mul.lo.s32 %r2581, %r2580, %r2578; mul.lo.s32 %r2582, %r2581, -1074439621; bfe.u32 %r2583, %r2582, 2, 6; xor.b32 %r2584, %r2583, %r2582; mul.lo.s32 %r2585, %r2584, -1638916925; bfe.u32 %r2586, %r2585, 2, 6; xor.b32 %r2587, %r2586, %r2585; mul.lo.s32 %r2588, %r2587, -933190689; and.b32 %r2589, %r2588, 255; bfe.u32 %r2590, %r2588, 5, 3; xor.b32 %r2591, %r2590, %r2589; add.s32 %r2592, %r2559, -1; xor.b32 %r850, %r848, 61; xor.b32 %r2593, %r2564, %r850; mul.lo.s32 %r2594, %r2593, 9; shr.u32 %r2595, %r2594, 4; xor.b32 %r2596, %r2595, %r2594; mul.lo.s32 %r2597, %r2596, 668265261; add.s32 %r2598, %r2597, %r849; shr.u32 %r2599, %r2597, 16; or.b32 %r2600, %r2599, 1; mul.lo.s32 %r2601, %r2600, %r2598; mul.lo.s32 %r2602, %r2601, 89287766; xor.b32 %r2603, %r2602, %r2601; mul.lo.s32 %r2604, %r2603, 1403136100; xor.b32 %r2605, %r2604, %r2603; brev.b32 %r2606, %r2605; neg.s32 %r2607, %r2559; and.b32 %r2608, %r603, %r2607; and.b32 %r2609, %r2606, %r2592; or.b32 %r2610, %r2609, %r2608; mad.lo.s32 %r2611, %r2591, %r2559, %r2610; shl.b32 %r2612, %r2559, 8; rem.u32 %r2613, %r2611, %r2612; ld.const.u64 %rd2600, [kernel_params+2424]; cvta.to.global.u64 %rd2601, %rd2600; shl.b32 %r2614, %r2613, 1; mul.wide.u32 %rd2602, %r2614, 4; add.s64 %rd2603, %rd2601, %rd2602; ld.global.f32 %f2799, [%rd2603]; ld.global.f32 %f2775, [%rd2603+4]; @%p1582 bra $L__BB106_1783; xor.b32 %r2615, %r850, %r601; mul.lo.s32 %r2616, %r2615, 9; shr.u32 %r2617, %r2616, 4; xor.b32 %r2618, %r2617, %r2616; mul.lo.s32 %r2619, %r2618, 668265261; cvt.rn.f32.u32 %f11273, %r2619; mul.ftz.f32 %f11274, %f11273, 0f2F7FFFFE; xor.b32 %r2620, %r601, %r848; xor.b32 %r2621, %r2620, -905047700; mul.lo.s32 %r2622, %r2621, 9; shr.u32 %r2623, %r2622, 4; xor.b32 %r2624, %r2623, %r2622; mul.lo.s32 %r2625, %r2624, 668265261; cvt.rn.f32.u32 %f11275, %r2625; mul.ftz.f32 %f11276, %f11275, 0f2F7FFFFE; fma.rn.ftz.f32 %f11277, %f11274, %f2763, %f2799; fma.rn.ftz.f32 %f11278, %f11276, %f2763, %f2775; cvt.rmi.ftz.f32.f32 %f11279, %f11277; sub.ftz.f32 %f2799, %f11277, %f11279; cvt.rmi.ftz.f32.f32 %f11280, %f11278; sub.ftz.f32 %f2775, %f11278, %f11280; bra.uni $L__BB106_1783; $L__BB106_1779: shr.u32 %r2627, %r848, 16; xor.b32 %r2628, %r2627, %r848; mul.lo.s32 %r2629, %r2628, 569420461; shr.u32 %r2630, %r2629, 15; xor.b32 %r2631, %r2630, %r2629; mul.lo.s32 %r2632, %r2631, -749064809; shr.u32 %r2633, %r2632, 15; xor.b32 %r2634, %r601, %r2632; xor.b32 %r2635, %r2634, %r2633; xor.b32 %r851, %r2635, -419546133; xor.b32 %r2636, %r2635, 508809841; add.s32 %r2637, %r2636, %r849; shr.u32 %r2638, %r2636, 16; or.b32 %r2639, %r2638, 1; mul.lo.s32 %r2640, %r2639, %r2637; mul.lo.s32 %r2641, %r2640, 89287766; xor.b32 %r2642, %r2641, %r2640; mul.lo.s32 %r2643, %r2642, 1403136100; xor.b32 %r852, %r2643, %r2642; setp.eq.s32 %p1584, %r852, 0; mov.u32 %r3271, 0; @%p1584 bra $L__BB106_1782; mov.u32 %r3268, 0; mov.u64 %rd2605, sobol_burley_table; mov.u32 %r3269, %r852; mov.u32 %r3271, %r3268; $L__BB106_1781: clz.b32 %r2646, %r3269; add.s32 %r2647, %r2646, %r3268; mul.wide.u32 %rd2604, %r2647, 4; add.s64 %rd2606, %rd2605, %rd2604; ld.const.u32 %r2648, [%rd2606+128]; xor.b32 %r3271, %r2648, %r3271; add.s32 %r3268, %r2647, 1; shl.b32 %r2649, %r3269, %r2646; shl.b32 %r3269, %r2649, 1; setp.ne.s32 %p1585, %r3269, 0; @%p1585 bra $L__BB106_1781; $L__BB106_1782: xor.b32 %r2650, %r851, -525684874; brev.b32 %r2651, %r852; shr.u32 %r2652, %r2650, 16; or.b32 %r2653, %r2652, 1; mul.lo.s32 %r2654, %r2651, 1025551850; xor.b32 %r2655, %r2654, %r2651; add.s32 %r2656, %r2655, %r2650; mul.lo.s32 %r2657, %r2656, %r2653; mul.lo.s32 %r2658, %r2657, 89287766; xor.b32 %r2659, %r2658, %r2657; xor.b32 %r2660, %r851, -1802089138; mul.lo.s32 %r2661, %r3271, 1025551850; xor.b32 %r2662, %r2661, %r3271; add.s32 %r2663, %r2662, %r2660; shr.u32 %r2664, %r2660, 16; or.b32 %r2665, %r2664, 1; mul.lo.s32 %r2666, %r2663, %r2665; mul.lo.s32 %r2667, %r2666, 89287766; xor.b32 %r2668, %r2667, %r2666; mul.lo.s32 %r2669, %r2668, 1403136100; xor.b32 %r2670, %r2669, %r2668; brev.b32 %r2671, %r2670; cvt.rn.f32.u32 %f11281, %r2671; mul.lo.s32 %r2672, %r2659, 1403136100; xor.b32 %r2673, %r2672, %r2659; brev.b32 %r2674, %r2673; cvt.rn.f32.u32 %f11282, %r2674; mul.ftz.f32 %f2775, %f11281, 0f2F7FFFFE; mul.ftz.f32 %f2799, %f11282, 0f2F7FFFFE; $L__BB106_1783: ld.local.u32 %r860, [%rd7+172]; setp.lt.s32 %p1586, %r860, 2; mov.u32 %r3277, 0; @%p1586 bra $L__BB106_1805; add.s32 %r2677, %r860, -1; and.b32 %r3275, %r860, 3; setp.lt.u32 %p1587, %r2677, 3; mov.f32 %f16443, 0f00000000; mov.u32 %r3274, 0; @%p1587 bra $L__BB106_1795; sub.s32 %r3273, %r860, %r3275; mov.f32 %f16443, 0f00000000; mov.u32 %r3274, 0; $L__BB106_1786: cvt.s64.s32 %rd579, %r3274; mul.wide.s32 %rd2607, %r3274, 80; add.s64 %rd2608, %rd1, %rd2607; add.s64 %rd580, %rd2608, 316; ld.local.u32 %r2679, [%rd2608+316]; setp.gt.s32 %p1588, %r2679, 34; @%p1588 bra $L__BB106_1788; ld.local.f32 %f11286, [%rd580+4]; add.ftz.f32 %f16443, %f16443, %f11286; $L__BB106_1788: ld.local.u32 %r2680, [%rd580+80]; setp.gt.s32 %p1589, %r2680, 34; @%p1589 bra $L__BB106_1790; ld.local.f32 %f11287, [%rd580+84]; add.ftz.f32 %f16443, %f16443, %f11287; $L__BB106_1790: ld.local.u32 %r2681, [%rd580+160]; setp.gt.s32 %p1590, %r2681, 34; @%p1590 bra $L__BB106_1792; ld.local.f32 %f11288, [%rd580+164]; add.ftz.f32 %f16443, %f16443, %f11288; $L__BB106_1792: ld.local.u32 %r2682, [%rd580+240]; setp.gt.s32 %p1591, %r2682, 34; @%p1591 bra $L__BB106_1794; ld.local.f32 %f11289, [%rd580+244]; add.ftz.f32 %f16443, %f16443, %f11289; $L__BB106_1794: cvt.u32.u64 %r2683, %rd579; add.s32 %r3274, %r2683, 4; add.s32 %r3273, %r3273, -4; setp.ne.s32 %p1592, %r3273, 0; @%p1592 bra $L__BB106_1786; $L__BB106_1795: setp.eq.s32 %p1593, %r3275, 0; @%p1593 bra $L__BB106_1800; mul.wide.s32 %rd2609, %r3274, 80; add.s64 %rd2610, %rd1, %rd2609; add.s64 %rd3069, %rd2610, 320; $L__BB106_1797: .pragma "nounroll"; ld.local.u32 %r2684, [%rd3069+-4]; setp.gt.s32 %p1594, %r2684, 34; @%p1594 bra $L__BB106_1799; ld.local.f32 %f11290, [%rd3069]; add.ftz.f32 %f16443, %f16443, %f11290; $L__BB106_1799: add.s64 %rd3069, %rd3069, 80; add.s32 %r3275, %r3275, -1; setp.ne.s32 %p1595, %r3275, 0; @%p1595 bra $L__BB106_1797; $L__BB106_1800: mul.ftz.f32 %f2793, %f2799, %f16443; mov.f32 %f16452, 0f00000000; mov.u32 %r3277, 0; mov.u32 %r3276, %r3277; mov.f32 %f16453, %f16452; $L__BB106_1801: mul.wide.s32 %rd2611, %r3276, 80; add.s64 %rd2612, %rd1, %rd2611; add.s64 %rd585, %rd2612, 316; ld.local.u32 %r2686, [%rd2612+316]; setp.gt.s32 %p1596, %r2686, 34; @%p1596 bra $L__BB106_1803; ld.local.f32 %f2795, [%rd585+4]; add.ftz.f32 %f16453, %f16452, %f2795; setp.lt.ftz.f32 %p1597, %f2793, %f16453; @%p1597 bra $L__BB106_1804; $L__BB106_1803: add.s32 %r3276, %r3276, 1; setp.lt.s32 %p1598, %r3276, %r860; mov.f32 %f16452, %f16453; @%p1598 bra $L__BB106_1801; bra.uni $L__BB106_1805; $L__BB106_1804: sub.ftz.f32 %f11292, %f2793, %f16452; div.approx.ftz.f32 %f2799, %f11292, %f2795; mov.u32 %r3277, %r3276; $L__BB106_1805: cvt.s64.s32 %rd586, %r3277; mul.wide.s32 %rd2613, %r3277, 80; add.s64 %rd2614, %rd1, %rd2613; add.s64 %rd587, %rd2614, 316; ld.local.u32 %r873, [%rd2614+316]; add.s32 %r2688, %r873, -32; setp.lt.u32 %p1599, %r2688, 3; @%p1599 bra $L__BB106_2383; bra.uni $L__BB106_1806; $L__BB106_2383: ld.const.u64 %rd2912, [kernel_params+2632]; mul.wide.s32 %rd2911, %r953, 12; cvta.to.global.u64 %rd2910, %rd2912; add.s64 %rd2909, %rd2910, %rd2911; ld.local.v4.f32 {%f15414, %f15415, %f15416, %f15417}, [%rd1]; st.global.f32 [%rd2909], %f15414; st.global.f32 [%rd2909+4], %f15415; st.global.f32 [%rd2909+8], %f15416; ld.const.v2.u64 {%rd2792, %rd2793}, [kernel_params+2640]; cvta.to.global.u64 %rd2796, %rd2792; ld.local.f32 %f15421, [%rd587+8]; ld.local.f32 %f15422, [%rd587+12]; ld.local.f32 %f15423, [%rd587+16]; mul.lo.s64 %rd2797, %rd2, 12; add.s64 %rd2798, %rd2796, %rd2797; st.global.f32 [%rd2798], %f15421; st.global.f32 [%rd2798+4], %f15422; st.global.f32 [%rd2798+8], %f15423; cvta.to.global.u64 %rd2799, %rd2793; add.s64 %rd2801, %rd2799, %rd837; mov.u32 %r2999, 0; st.global.u32 [%rd2801], %r2999; ld.const.u64 %rd2802, [kernel_params+2656]; cvta.to.global.u64 %rd2803, %rd2802; add.s64 %rd2804, %rd2803, %rd837; mov.u32 %r3000, 2139095039; st.global.u32 [%rd2804], %r3000; ld.local.f32 %f15424, [%rd7+4]; add.s64 %rd2806, %rd656, %rd837; st.global.f32 [%rd2806], %f15424; add.s64 %rd2808, %rd658, %rd837; st.global.u32 [%rd2808], %r2999; ld.const.u64 %rd2809, [kernel_params+2760]; cvta.to.global.u64 %rd2810, %rd2809; ld.local.v2.f32 {%f15425, %f15426}, [%rd7+-64]; ld.local.f32 %f15429, [%rd7+-56]; add.s64 %rd2811, %rd2810, %rd2797; st.global.f32 [%rd2811], %f15425; st.global.f32 [%rd2811+4], %f15426; st.global.f32 [%rd2811+8], %f15429; ld.global.u32 %r3001, [%rd29]; and.b32 %r3002, %r3001, -2; setp.eq.s32 %p2146, %r873, 32; selp.b32 %r3003, 2097152, 1048576, %p2146; or.b32 %r924, %r3003, %r3002; ld.local.f32 %f16789, [%rd587+-12]; ld.local.f32 %f16790, [%rd587+-8]; ld.local.f32 %f16791, [%rd587+-4]; @%p1586 bra $L__BB106_2401; add.s32 %r3005, %r860, -1; and.b32 %r3293, %r860, 3; setp.lt.u32 %p2147, %r3005, 3; mov.f32 %f16780, 0f00000000; mov.u32 %r3292, %r2999; @%p2147 bra $L__BB106_2395; sub.s32 %r3291, %r860, %r3293; mov.f32 %f16780, 0f00000000; mov.u32 %r3292, 0; $L__BB106_2386: cvt.s64.s32 %rd596, %r3292; mul.wide.s32 %rd2812, %r3292, 80; add.s64 %rd2813, %rd1, %rd2812; add.s64 %rd597, %rd2813, 316; ld.local.u32 %r3007, [%rd2813+316]; setp.gt.s32 %p2148, %r3007, 34; @%p2148 bra $L__BB106_2388; ld.local.f32 %f15433, [%rd597+4]; add.ftz.f32 %f16780, %f16780, %f15433; $L__BB106_2388: ld.local.u32 %r3008, [%rd597+80]; setp.gt.s32 %p2149, %r3008, 34; @%p2149 bra $L__BB106_2390; ld.local.f32 %f15434, [%rd597+84]; add.ftz.f32 %f16780, %f16780, %f15434; $L__BB106_2390: ld.local.u32 %r3009, [%rd597+160]; setp.gt.s32 %p2150, %r3009, 34; @%p2150 bra $L__BB106_2392; ld.local.f32 %f15435, [%rd597+164]; add.ftz.f32 %f16780, %f16780, %f15435; $L__BB106_2392: ld.local.u32 %r3010, [%rd597+240]; setp.gt.s32 %p2151, %r3010, 34; @%p2151 bra $L__BB106_2394; ld.local.f32 %f15436, [%rd597+244]; add.ftz.f32 %f16780, %f16780, %f15436; $L__BB106_2394: cvt.u32.u64 %r3011, %rd596; add.s32 %r3292, %r3011, 4; add.s32 %r3291, %r3291, -4; setp.ne.s32 %p2152, %r3291, 0; @%p2152 bra $L__BB106_2386; $L__BB106_2395: setp.eq.s32 %p2153, %r3293, 0; @%p2153 bra $L__BB106_2400; mul.wide.s32 %rd2814, %r3292, 80; add.s64 %rd2815, %rd1, %rd2814; add.s64 %rd3070, %rd2815, 320; $L__BB106_2397: .pragma "nounroll"; add.s64 %rd600, %rd3070, -4; ld.local.u32 %r3012, [%rd3070+-4]; setp.gt.s32 %p2154, %r3012, 34; @%p2154 bra $L__BB106_2399; ld.local.f32 %f15437, [%rd600+4]; add.ftz.f32 %f16780, %f16780, %f15437; $L__BB106_2399: add.s64 %rd3070, %rd3070, 80; add.s32 %r3293, %r3293, -1; setp.ne.s32 %p2155, %r3293, 0; @%p2155 bra $L__BB106_2397; $L__BB106_2400: ld.local.f32 %f15438, [%rd587+4]; div.approx.ftz.f32 %f15439, %f16780, %f15438; mul.ftz.f32 %f16789, %f16789, %f15439; mul.ftz.f32 %f16790, %f16790, %f15439; mul.ftz.f32 %f16791, %f16791, %f15439; $L__BB106_2401: ld.local.f32 %f15440, [%rd587+44]; setp.neu.ftz.f32 %p2156, %f15440, 0f7F7FFFFF; or.b32 %r3013, %r924, 4194304; selp.b32 %r3014, %r3013, %r924, %p2156; shl.b32 %r3015, %r847, 23; and.b32 %r3016, %r3015, 8388608; or.b32 %r3017, %r3016, %r3014; ld.const.u64 %rd2816, [kernel_params+2584]; cvta.to.global.u64 %rd2817, %rd2816; add.s64 %rd2819, %rd2817, %rd2797; ld.global.f32 %f15441, [%rd2819]; mul.ftz.f32 %f15442, %f16789, %f15441; ld.global.f32 %f15443, [%rd2819+4]; mul.ftz.f32 %f15444, %f16790, %f15443; ld.global.f32 %f15445, [%rd2819+8]; mul.ftz.f32 %f15446, %f16791, %f15445; st.global.f32 [%rd2819], %f15442; st.global.f32 [%rd2819+4], %f15444; st.global.f32 [%rd2819+8], %f15446; st.global.u32 [%rd29], %r3017; ld.global.u16 %rs130, [%rd278]; add.s16 %rs131, %rs130, 16; st.global.u16 [%rd278], %rs131; and.b32 %r3018, %r787, 2097152; setp.eq.s32 %p2157, %r3018, 0; @%p2157 bra $L__BB106_2404; ld.const.u64 %rd2820, [kernel_params+2464]; cvta.to.global.u64 %rd2821, %rd2820; add.s64 %rd2823, %rd2821, %rd1709; ld.global.u16 %rs132, [%rd2823]; setp.ne.s16 %p2158, %rs132, 0; @%p2158 bra $L__BB106_2404; ld.const.u64 %rd2824, [kernel_params+2600]; cvta.to.global.u64 %rd2825, %rd2824; add.s64 %rd2827, %rd2825, %rd2797; mov.u32 %r3019, 1065353216; st.global.u32 [%rd2827], %r3019; st.global.u32 [%rd2827+4], %r3019; st.global.u32 [%rd2827+8], %r3019; ld.const.u64 %rd2828, [kernel_params+2608]; cvta.to.global.u64 %rd2829, %rd2828; add.s64 %rd2830, %rd2829, %rd2797; st.global.u32 [%rd2830], %r2999; st.global.u32 [%rd2830+4], %r2999; st.global.u32 [%rd2830+8], %r2999; $L__BB106_2404: ld.const.v2.u64 {%rd2831, %rd2832}, [kernel_params+2736]; cvta.to.global.u64 %rd2835, %rd2831; ld.local.f32 %f15447, [%rd587+32]; ld.local.f32 %f15448, [%rd587+36]; ld.local.f32 %f15449, [%rd587+40]; add.s64 %rd2837, %rd2835, %rd2797; st.global.f32 [%rd2837], %f15447; st.global.f32 [%rd2837+4], %f15448; st.global.f32 [%rd2837+8], %f15449; cvta.to.global.u64 %rd2838, %rd2832; ld.local.f32 %f15450, [%rd587+20]; ld.local.f32 %f15451, [%rd587+24]; ld.local.f32 %f15452, [%rd587+28]; add.s64 %rd2839, %rd2838, %rd2797; st.global.f32 [%rd2839], %f15450; st.global.f32 [%rd2839+4], %f15451; st.global.f32 [%rd2839+8], %f15452; ld.local.f32 %f15453, [%rd587+48]; ld.const.u64 %rd2840, [kernel_params+2752]; cvta.to.global.u64 %rd2841, %rd2840; add.s64 %rd2843, %rd2841, %rd837; st.global.f32 [%rd2843], %f15453; mov.u32 %r3294, 256; bra.uni $L__BB106_2405; $L__BB106_1806: ld.local.u8 %rs111, [%rd7+-24]; and.b16 %rs112, %rs111, 6; setp.eq.s16 %p1600, %rs112, 0; @%p1600 bra $L__BB106_1808; ld.local.f32 %f2808, [%rd587+16]; ld.local.f32 %f2807, [%rd587+12]; ld.local.f32 %f2806, [%rd587+8]; bra.uni $L__BB106_1809; $L__BB106_1808: ld.local.v2.f32 {%f2806, %f2807}, [%rd7+-64]; ld.local.f32 %f2808, [%rd7+-56]; $L__BB106_1809: mov.u32 %r3285, 0; mov.f32 %f16738, 0f00000000; setp.gt.s32 %p1601, %r873, 15; @%p1601 bra $L__BB106_1836; setp.gt.s32 %p1617, %r873, 9; @%p1617 bra $L__BB106_1833; setp.gt.s32 %p1621, %r873, 5; @%p1621 bra $L__BB106_1819; bra.uni $L__BB106_1812; $L__BB106_1819: setp.gt.s32 %p1622, %r873, 7; @%p1622 bra $L__BB106_1828; setp.eq.s32 %p1625, %r873, 6; @%p1625 bra $L__BB106_2294; setp.eq.s32 %p1626, %r873, 7; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1626 bra $L__BB106_1822; bra.uni $L__BB106_2309; $L__BB106_1822: ld.local.f32 %f14849, [%rd587+20]; mul.ftz.f32 %f4047, %f14849, 0f3FC90FDB; mov.f32 %f14850, 0f3FC90FDB; ld.local.f32 %f14851, [%rd587+24]; mul.ftz.f32 %f4048, %f14851, 0f3FC90FDB; add.ftz.f32 %f4049, %f4047, %f4048; min.ftz.f32 %f4050, %f4049, %f14850; mul.ftz.f32 %f4051, %f2799, %f4050; setp.leu.ftz.f32 %p2028, %f4050, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2028 bra $L__BB106_2309; cos.approx.ftz.f32 %f14856, %f4050; mul.ftz.f32 %f14857, %f2799, %f14856; sub.ftz.f32 %f14858, %f14856, %f14857; add.ftz.f32 %f14859, %f2799, %f14858; mul.ftz.f32 %f14860, %f14859, %f14859; mov.f32 %f16728, 0f3F800000; sub.ftz.f32 %f14862, %f16728, %f14860; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f14863, %f14862, %f16738; sqrt.approx.ftz.f32 %f14864, %f14863; mul.ftz.f32 %f14865, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f14866, %f14865; mul.ftz.f32 %f14867, %f14864, %f14866; sin.approx.ftz.f32 %f14868, %f14865; mul.ftz.f32 %f14869, %f14864, %f14868; ld.local.f32 %f14870, [%rd587+12]; ld.local.f32 %f14871, [%rd587+8]; setp.neu.ftz.f32 %p2029, %f14871, %f14870; ld.local.f32 %f14872, [%rd587+16]; setp.neu.ftz.f32 %p2030, %f14871, %f14872; or.pred %p2031, %p2029, %p2030; sub.ftz.f32 %f14873, %f14872, %f14870; neg.ftz.f32 %f14874, %f14870; neg.ftz.f32 %f14875, %f14872; selp.f32 %f14876, %f14875, %f14872, %p2031; add.ftz.f32 %f14877, %f14871, %f14876; selp.f32 %f14878, %f14870, %f14874, %p2031; sub.ftz.f32 %f14879, %f14878, %f14871; mul.ftz.f32 %f14880, %f14877, %f14877; fma.rn.ftz.f32 %f14881, %f14873, %f14873, %f14880; fma.rn.ftz.f32 %f14882, %f14879, %f14879, %f14881; sqrt.approx.ftz.f32 %f14883, %f14882; div.approx.ftz.f32 %f14884, %f14873, %f14883; div.approx.ftz.f32 %f14885, %f14877, %f14883; div.approx.ftz.f32 %f14886, %f14879, %f14883; mul.ftz.f32 %f14887, %f14870, %f14886; mul.ftz.f32 %f14888, %f14872, %f14885; sub.ftz.f32 %f14889, %f14887, %f14888; mul.ftz.f32 %f14890, %f14872, %f14884; mul.ftz.f32 %f14891, %f14871, %f14886; sub.ftz.f32 %f14892, %f14890, %f14891; mul.ftz.f32 %f14893, %f14871, %f14885; mul.ftz.f32 %f14894, %f14870, %f14884; sub.ftz.f32 %f14895, %f14893, %f14894; mul.ftz.f32 %f14896, %f14869, %f14889; mul.ftz.f32 %f14897, %f14869, %f14892; mul.ftz.f32 %f14898, %f14869, %f14895; fma.rn.ftz.f32 %f14899, %f14867, %f14884, %f14896; fma.rn.ftz.f32 %f14900, %f14867, %f14885, %f14897; fma.rn.ftz.f32 %f14901, %f14867, %f14886, %f14898; fma.rn.ftz.f32 %f16735, %f14871, %f14859, %f14899; fma.rn.ftz.f32 %f16736, %f14870, %f14859, %f14900; fma.rn.ftz.f32 %f16737, %f14872, %f14859, %f14901; sub.ftz.f32 %f14902, %f16728, %f14856; mov.f32 %f14903, 0f3E22F983; div.approx.ftz.f32 %f4055, %f14903, %f14902; mul.ftz.f32 %f14904, %f2807, %f16736; fma.rn.ftz.f32 %f14905, %f2806, %f16735, %f14904; fma.rn.ftz.f32 %f14906, %f2808, %f16737, %f14905; setp.leu.ftz.f32 %p2032, %f14906, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2032 bra $L__BB106_2309; setp.lt.ftz.f32 %p2033, %f4051, %f4047; @%p2033 bra $L__BB106_1827; setp.leu.ftz.f32 %p2034, %f4049, %f4051; setp.eq.ftz.f32 %p2035, %f4048, 0f00000000; mov.f32 %f16728, 0f00000000; or.pred %p2036, %p2035, %p2034; @%p2036 bra $L__BB106_1827; sub.ftz.f32 %f14909, %f4051, %f4047; div.approx.ftz.f32 %f14910, %f14909, %f4048; mov.f32 %f14911, 0f3F800000; sub.ftz.f32 %f16728, %f14911, %f14910; $L__BB106_1827: mul.ftz.f32 %f16739, %f4055, %f16728; mov.f32 %f16738, %f4055; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_1836: setp.gt.s32 %p1602, %r873, 22; @%p1602 bra $L__BB106_1853; setp.gt.s32 %p1610, %r873, 19; @%p1610 bra $L__BB106_1848; setp.eq.s32 %p1614, %r873, 16; @%p1614 bra $L__BB106_2113; setp.eq.s32 %p1615, %r873, 17; @%p1615 bra $L__BB106_2108; setp.eq.s32 %p1616, %r873, 19; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1616 bra $L__BB106_1841; bra.uni $L__BB106_2309; $L__BB106_1841: ld.local.v2.f32 {%f12949, %f12950}, [%rd7+-48]; ld.local.f32 %f12951, [%rd587+20]; mul.ftz.f32 %f3454, %f12951, 0f3FC90FDB; ld.local.f32 %f12952, [%rd587+24]; mul.ftz.f32 %f3455, %f12952, 0f3FC90FDB; ld.local.f32 %f3456, [%rd587+8]; ld.local.f32 %f3457, [%rd7+-52]; ld.local.f32 %f3458, [%rd587+12]; mul.ftz.f32 %f12953, %f12949, %f3458; fma.rn.ftz.f32 %f12954, %f3457, %f3456, %f12953; ld.local.f32 %f3459, [%rd587+16]; fma.rn.ftz.f32 %f3460, %f12950, %f3459, %f12954; setp.leu.ftz.f32 %p1861, %f3460, 0f00000000; mov.u32 %r3285, 10; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1861 bra $L__BB106_2309; add.ftz.f32 %f12959, %f3460, %f3460; mul.ftz.f32 %f12960, %f12959, %f3456; mul.ftz.f32 %f12961, %f12959, %f3458; mul.ftz.f32 %f12962, %f12959, %f3459; sub.ftz.f32 %f12963, %f12960, %f3457; sub.ftz.f32 %f12964, %f12961, %f12949; sub.ftz.f32 %f12965, %f12962, %f12950; add.ftz.f32 %f3461, %f3454, %f3455; mov.f32 %f12966, 0f3FC90FDB; min.ftz.f32 %f12967, %f3461, %f12966; mul.ftz.f32 %f3462, %f2799, %f12967; cos.approx.ftz.f32 %f12968, %f12967; mul.ftz.f32 %f12969, %f2799, %f12968; sub.ftz.f32 %f12970, %f12968, %f12969; add.ftz.f32 %f12971, %f2799, %f12970; mul.ftz.f32 %f12972, %f12971, %f12971; mov.f32 %f16566, 0f3F800000; sub.ftz.f32 %f12974, %f16566, %f12972; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f12975, %f12974, %f16738; sqrt.approx.ftz.f32 %f12976, %f12975; mul.ftz.f32 %f12977, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f12978, %f12977; mul.ftz.f32 %f12979, %f12976, %f12978; sin.approx.ftz.f32 %f12980, %f12977; mul.ftz.f32 %f12981, %f12976, %f12980; setp.neu.ftz.f32 %p1862, %f12963, %f12964; setp.neu.ftz.f32 %p1863, %f12963, %f12965; or.pred %p1864, %p1862, %p1863; sub.ftz.f32 %f12982, %f12965, %f12964; neg.ftz.f32 %f12983, %f12964; neg.ftz.f32 %f12984, %f12965; selp.f32 %f12985, %f12984, %f12965, %p1864; add.ftz.f32 %f12986, %f12963, %f12985; selp.f32 %f12987, %f12964, %f12983, %p1864; sub.ftz.f32 %f12988, %f12987, %f12963; mul.ftz.f32 %f12989, %f12986, %f12986; fma.rn.ftz.f32 %f12990, %f12982, %f12982, %f12989; fma.rn.ftz.f32 %f12991, %f12988, %f12988, %f12990; sqrt.approx.ftz.f32 %f12992, %f12991; div.approx.ftz.f32 %f12993, %f12982, %f12992; div.approx.ftz.f32 %f12994, %f12986, %f12992; div.approx.ftz.f32 %f12995, %f12988, %f12992; mul.ftz.f32 %f12996, %f12964, %f12995; mul.ftz.f32 %f12997, %f12965, %f12994; sub.ftz.f32 %f12998, %f12996, %f12997; mul.ftz.f32 %f12999, %f12965, %f12993; mul.ftz.f32 %f13000, %f12963, %f12995; sub.ftz.f32 %f13001, %f12999, %f13000; mul.ftz.f32 %f13002, %f12963, %f12994; mul.ftz.f32 %f13003, %f12964, %f12993; sub.ftz.f32 %f13004, %f13002, %f13003; mul.ftz.f32 %f13005, %f12981, %f12998; mul.ftz.f32 %f13006, %f12981, %f13001; mul.ftz.f32 %f13007, %f12981, %f13004; fma.rn.ftz.f32 %f13008, %f12979, %f12993, %f13005; fma.rn.ftz.f32 %f13009, %f12979, %f12994, %f13006; fma.rn.ftz.f32 %f13010, %f12979, %f12995, %f13007; fma.rn.ftz.f32 %f16735, %f12963, %f12971, %f13008; fma.rn.ftz.f32 %f16736, %f12964, %f12971, %f13009; fma.rn.ftz.f32 %f16737, %f12965, %f12971, %f13010; sub.ftz.f32 %f13011, %f16566, %f12968; mov.f32 %f13012, 0f3E22F983; div.approx.ftz.f32 %f3466, %f13012, %f13011; mul.ftz.f32 %f13013, %f2807, %f16736; fma.rn.ftz.f32 %f13014, %f2806, %f16735, %f13013; fma.rn.ftz.f32 %f13015, %f2808, %f16737, %f13014; setp.leu.ftz.f32 %p1865, %f13015, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1865 bra $L__BB106_2309; mul.ftz.f32 %f13020, %f16736, %f3458; fma.rn.ftz.f32 %f13021, %f16735, %f3456, %f13020; fma.rn.ftz.f32 %f13022, %f16737, %f3459, %f13021; setp.leu.ftz.f32 %p1866, %f13022, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1866 bra $L__BB106_2309; setp.lt.ftz.f32 %p1867, %f3462, %f3454; @%p1867 bra $L__BB106_1847; setp.leu.ftz.f32 %p1868, %f3461, %f3462; setp.eq.ftz.f32 %p1869, %f3455, 0f00000000; mov.f32 %f16566, 0f00000000; or.pred %p1870, %p1869, %p1868; @%p1870 bra $L__BB106_1847; sub.ftz.f32 %f13025, %f3462, %f3454; div.approx.ftz.f32 %f13026, %f13025, %f3455; mov.f32 %f13027, 0f3F800000; sub.ftz.f32 %f16566, %f13027, %f13026; $L__BB106_1847: mul.ftz.f32 %f16739, %f3466, %f16566; mov.f32 %f16738, %f3466; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_1833: add.s32 %r2690, %r873, -10; setp.lt.u32 %p1618, %r2690, 3; @%p1618 bra $L__BB106_2246; add.s32 %r2691, %r873, -14; setp.lt.u32 %p1619, %r2691, 2; @%p1619 bra $L__BB106_2143; bra.uni $L__BB106_1835; $L__BB106_2143: ld.local.v2.f32 {%f13461, %f13462}, [%rd7+-48]; ld.local.f32 %f3573, [%rd587+8]; ld.local.f32 %f3574, [%rd7+-52]; ld.local.f32 %f3575, [%rd587+12]; mul.ftz.f32 %f13463, %f13461, %f3575; fma.rn.ftz.f32 %f13464, %f3574, %f3573, %f13463; ld.local.f32 %f3576, [%rd587+16]; fma.rn.ftz.f32 %f3577, %f13462, %f3576, %f13464; setp.le.ftz.f32 %p1906, %f3577, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1906 bra $L__BB106_2309; ld.local.v2.f32 {%f13465, %f13466}, [%rd587+20]; mul.ftz.f32 %f13467, %f13465, %f13466; setp.lt.ftz.f32 %p1907, %f13467, 0f33D6BF95; @%p1907 bra $L__BB106_2196; bra.uni $L__BB106_2145; $L__BB106_2196: add.ftz.f32 %f13798, %f3577, %f3577; mul.ftz.f32 %f13799, %f3573, %f13798; mul.ftz.f32 %f13800, %f3575, %f13798; mul.ftz.f32 %f13801, %f3576, %f13798; sub.ftz.f32 %f16735, %f13799, %f3574; sub.ftz.f32 %f16736, %f13800, %f13461; sub.ftz.f32 %f16737, %f13801, %f13462; mul.ftz.f32 %f13802, %f2807, %f16736; fma.rn.ftz.f32 %f13803, %f2806, %f16735, %f13802; fma.rn.ftz.f32 %f13804, %f2808, %f16737, %f13803; setp.le.ftz.f32 %p1954, %f13804, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1954 bra $L__BB106_2309; mov.u32 %r3285, 18; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1853: setp.gt.s32 %p1603, %r873, 28; @%p1603 bra $L__BB106_1870; setp.eq.s32 %p1607, %r873, 23; @%p1607 bra $L__BB106_2246; setp.eq.s32 %p1608, %r873, 24; @%p1608 bra $L__BB106_1857; setp.eq.s32 %p1609, %r873, 27; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1609 bra $L__BB106_1857; bra.uni $L__BB106_2309; $L__BB106_1857: ld.local.v2.f32 {%f12249, %f12250}, [%rd7+-48]; ld.local.v2.f32 {%f12251, %f12252}, [%rd587+20]; mul.ftz.f32 %f12253, %f12251, %f12252; setp.lt.ftz.f32 %p1774, %f12253, 0f33D6BF95; ld.local.f32 %f16546, [%rd587+8]; ld.local.f32 %f3197, [%rd7+-52]; ld.local.f32 %f16547, [%rd587+12]; mul.ftz.f32 %f12254, %f12249, %f16547; fma.rn.ftz.f32 %f12255, %f3197, %f16546, %f12254; ld.local.f32 %f16548, [%rd587+16]; fma.rn.ftz.f32 %f16549, %f12250, %f16548, %f12255; @%p1774 bra $L__BB106_2085; bra.uni $L__BB106_1858; $L__BB106_2085: ld.local.f32 %f3376, [%rd587+28]; setp.gt.ftz.f32 %p1832, %f16549, 0f00000000; @%p1832 bra $L__BB106_2087; bra.uni $L__BB106_2086; $L__BB106_2087: rcp.approx.ftz.f32 %f16550, %f3376; bra.uni $L__BB106_2088; $L__BB106_2246: ld.local.v2.f32 {%f14360, %f14361}, [%rd7+-48]; ld.local.v2.f32 {%f14362, %f14363}, [%rd587+20]; ld.local.f32 %f3898, [%rd587+8]; ld.local.f32 %f3899, [%rd7+-52]; ld.local.f32 %f3900, [%rd587+12]; mul.ftz.f32 %f14364, %f14360, %f3900; fma.rn.ftz.f32 %f14365, %f3899, %f3898, %f14364; ld.local.f32 %f3901, [%rd587+16]; fma.rn.ftz.f32 %f3902, %f14361, %f3901, %f14365; setp.gt.ftz.f32 %p1989, %f3902, 0f00000000; @%p1989 bra $L__BB106_2248; bra.uni $L__BB106_2247; $L__BB106_2248: setp.eq.ftz.f32 %p1991, %f14362, %f14363; @%p1991 bra $L__BB106_2250; bra.uni $L__BB106_2249; $L__BB106_2250: setp.neu.ftz.f32 %p1992, %f3898, %f3900; setp.neu.ftz.f32 %p1993, %f3898, %f3901; or.pred %p1994, %p1992, %p1993; neg.ftz.f32 %f14393, %f3901; selp.f32 %f14394, %f14393, %f3901, %p1994; add.ftz.f32 %f14395, %f3898, %f14394; neg.ftz.f32 %f14396, %f3900; selp.f32 %f14397, %f3900, %f14396, %p1994; sub.ftz.f32 %f14398, %f14397, %f3898; sub.ftz.f32 %f14399, %f3901, %f3900; mul.ftz.f32 %f14400, %f14395, %f14395; fma.rn.ftz.f32 %f14401, %f14399, %f14399, %f14400; fma.rn.ftz.f32 %f14402, %f14398, %f14398, %f14401; sqrt.approx.ftz.f32 %f14403, %f14402; div.approx.ftz.f32 %f16691, %f14399, %f14403; div.approx.ftz.f32 %f16692, %f14395, %f14403; div.approx.ftz.f32 %f16693, %f14398, %f14403; mul.ftz.f32 %f14404, %f3900, %f16693; mul.ftz.f32 %f14405, %f3901, %f16692; sub.ftz.f32 %f16688, %f14404, %f14405; mul.ftz.f32 %f14406, %f3901, %f16691; mul.ftz.f32 %f14407, %f3898, %f16693; sub.ftz.f32 %f16689, %f14406, %f14407; mul.ftz.f32 %f14408, %f3898, %f16692; mul.ftz.f32 %f14409, %f3900, %f16691; sub.ftz.f32 %f16690, %f14408, %f14409; bra.uni $L__BB106_2251; $L__BB106_1812: setp.eq.s32 %p1627, %r873, 2; @%p1627 bra $L__BB106_2307; setp.eq.s32 %p1628, %r873, 3; @%p1628 bra $L__BB106_2303; setp.eq.s32 %p1629, %r873, 5; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1629 bra $L__BB106_1815; bra.uni $L__BB106_2309; $L__BB106_1815: ld.local.f32 %f4081, [%rd7+-52]; ld.local.v2.f32 {%f14996, %f14997}, [%rd7+-48]; mul.ftz.f32 %f14998, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f14999, %f14998; sqrt.approx.ftz.f32 %f15000, %f2775; mul.ftz.f32 %f15001, %f15000, %f14999; sin.approx.ftz.f32 %f15002, %f14998; mul.ftz.f32 %f15003, %f15000, %f15002; mul.ftz.f32 %f15004, %f15001, %f15001; mov.f32 %f15005, 0f3F800000; sub.ftz.f32 %f15006, %f15005, %f15004; mul.ftz.f32 %f15007, %f15003, %f15003; sub.ftz.f32 %f15008, %f15006, %f15007; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f15009, %f15008, %f16738; sqrt.approx.ftz.f32 %f4084, %f15009; ld.local.f32 %f4085, [%rd587+12]; ld.local.f32 %f4086, [%rd587+8]; setp.neu.ftz.f32 %p2044, %f4086, %f4085; ld.local.f32 %f4087, [%rd587+16]; setp.neu.ftz.f32 %p2045, %f4086, %f4087; or.pred %p2046, %p2044, %p2045; sub.ftz.f32 %f15010, %f4087, %f4085; neg.ftz.f32 %f15011, %f4085; neg.ftz.f32 %f15012, %f4087; selp.f32 %f15013, %f15012, %f4087, %p2046; add.ftz.f32 %f15014, %f4086, %f15013; selp.f32 %f15015, %f4085, %f15011, %p2046; sub.ftz.f32 %f15016, %f15015, %f4086; mul.ftz.f32 %f15017, %f15014, %f15014; fma.rn.ftz.f32 %f15018, %f15010, %f15010, %f15017; fma.rn.ftz.f32 %f15019, %f15016, %f15016, %f15018; sqrt.approx.ftz.f32 %f15020, %f15019; div.approx.ftz.f32 %f15021, %f15010, %f15020; div.approx.ftz.f32 %f15022, %f15014, %f15020; div.approx.ftz.f32 %f15023, %f15016, %f15020; mul.ftz.f32 %f15024, %f4085, %f15023; mul.ftz.f32 %f15025, %f4087, %f15022; sub.ftz.f32 %f15026, %f15024, %f15025; mul.ftz.f32 %f15027, %f4087, %f15021; mul.ftz.f32 %f15028, %f4086, %f15023; sub.ftz.f32 %f15029, %f15027, %f15028; mul.ftz.f32 %f15030, %f4086, %f15022; mul.ftz.f32 %f15031, %f4085, %f15021; sub.ftz.f32 %f15032, %f15030, %f15031; mul.ftz.f32 %f15033, %f15003, %f15026; mul.ftz.f32 %f15034, %f15003, %f15029; mul.ftz.f32 %f15035, %f15003, %f15032; fma.rn.ftz.f32 %f15036, %f15001, %f15021, %f15033; fma.rn.ftz.f32 %f15037, %f15001, %f15022, %f15034; fma.rn.ftz.f32 %f15038, %f15001, %f15023, %f15035; fma.rn.ftz.f32 %f16735, %f4086, %f4084, %f15036; fma.rn.ftz.f32 %f16736, %f4085, %f4084, %f15037; fma.rn.ftz.f32 %f16737, %f4087, %f4084, %f15038; mul.ftz.f32 %f15039, %f2807, %f16736; fma.rn.ftz.f32 %f15040, %f2806, %f16735, %f15039; fma.rn.ftz.f32 %f15041, %f2808, %f16737, %f15040; setp.leu.ftz.f32 %p2047, %f15041, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2047 bra $L__BB106_2309; mul.ftz.f32 %f15045, %f4085, %f16736; fma.rn.ftz.f32 %f15046, %f4086, %f16735, %f15045; fma.rn.ftz.f32 %f4091, %f4087, %f16737, %f15046; setp.le.ftz.f32 %p2048, %f4091, 0f00000000; mov.f32 %f16739, 0f00000000; @%p2048 bra $L__BB106_2302; mul.ftz.f32 %f15047, %f14996, %f4085; fma.rn.ftz.f32 %f15048, %f4081, %f4086, %f15047; fma.rn.ftz.f32 %f15049, %f14997, %f4087, %f15048; sub.ftz.f32 %f15051, %f15005, %f15049; mov.f32 %f15052, 0f00000000; max.ftz.f32 %f15053, %f15051, %f15052; min.ftz.f32 %f15054, %f15053, %f15005; mul.ftz.f32 %f15055, %f15054, %f15054; mul.ftz.f32 %f15056, %f15055, %f15055; mul.ftz.f32 %f4092, %f15054, %f15056; sub.ftz.f32 %f15057, %f15005, %f4091; max.ftz.f32 %f15058, %f15057, %f15052; min.ftz.f32 %f15059, %f15058, %f15005; mul.ftz.f32 %f15060, %f15059, %f15059; mul.ftz.f32 %f15061, %f15060, %f15060; mul.ftz.f32 %f4093, %f15059, %f15061; ld.local.u32 %r903, [%rd587+24]; and.b32 %r2895, %r903, 3; setp.eq.s32 %p2049, %r2895, 0; fma.rn.ftz.f32 %f4094, %f4093, 0fBF000000, 0f3F800000; @%p2049 bra $L__BB106_2298; fma.rn.ftz.f32 %f15062, %f4092, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f16732, %f15062, %f4094, 0f00000000; bra.uni $L__BB106_2299; $L__BB106_1848: setp.eq.s32 %p1611, %r873, 20; @%p1611 bra $L__BB106_2097; setp.eq.s32 %p1612, %r873, 21; @%p1612 bra $L__BB106_2091; setp.eq.s32 %p1613, %r873, 22; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1613 bra $L__BB106_1851; bra.uni $L__BB106_2309; $L__BB106_1870: setp.eq.s32 %p1604, %r873, 29; @%p1604 bra $L__BB106_1890; setp.eq.s32 %p1605, %r873, 30; @%p1605 bra $L__BB106_1874; setp.ne.s32 %p1606, %r873, 31; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1606 bra $L__BB106_2309; ld.local.v2.f32 {%f11304, %f11305}, [%rd7+-48]; ld.local.f32 %f11308, [%rd7+-52]; neg.ftz.f32 %f16735, %f11308; neg.ftz.f32 %f16736, %f11304; neg.ftz.f32 %f16737, %f11305; mov.u32 %r3285, 33; mov.f32 %f16738, 0f3F800000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1828: setp.eq.s32 %p1623, %r873, 8; @%p1623 bra $L__BB106_2292; setp.eq.s32 %p1624, %r873, 9; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1624 bra $L__BB106_1830; bra.uni $L__BB106_2309; $L__BB106_1830: ld.local.v2.f32 {%f14771, %f14772}, [%rd7+-48]; ld.local.f32 %f4034, [%rd587+8]; ld.local.f32 %f4035, [%rd7+-52]; ld.local.f32 %f4036, [%rd587+12]; mul.ftz.f32 %f14773, %f14771, %f4036; fma.rn.ftz.f32 %f14774, %f4035, %f4034, %f14773; ld.local.f32 %f4037, [%rd587+16]; fma.rn.ftz.f32 %f4038, %f14772, %f4037, %f14774; setp.leu.ftz.f32 %p2022, %f4038, 0f00000000; mov.u32 %r3285, 18; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2022 bra $L__BB106_2309; add.ftz.f32 %f14779, %f4038, %f4038; mul.ftz.f32 %f14780, %f4034, %f14779; mul.ftz.f32 %f14781, %f4036, %f14779; mul.ftz.f32 %f14782, %f4037, %f14779; sub.ftz.f32 %f16735, %f14780, %f4035; sub.ftz.f32 %f16736, %f14781, %f14771; sub.ftz.f32 %f16737, %f14782, %f14772; mul.ftz.f32 %f14783, %f2807, %f16736; fma.rn.ftz.f32 %f14784, %f2806, %f16735, %f14783; fma.rn.ftz.f32 %f14785, %f2808, %f16737, %f14784; setp.leu.ftz.f32 %p2023, %f14785, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2023 bra $L__BB106_2309; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_2247: setp.eq.s32 %p1990, %r873, 23; selp.b32 %r3285, 9, 10, %p1990; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1835: setp.eq.s32 %p1620, %r873, 13; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1620 bra $L__BB106_1851; bra.uni $L__BB106_2309; $L__BB106_1851: ld.local.v2.f32 {%f13809, %f13810}, [%rd7+-48]; ld.local.v2.f32 {%f13811, %f13812}, [%rd587+20]; ld.local.f32 %f3736, [%rd587+8]; ld.local.f32 %f3737, [%rd7+-52]; ld.local.f32 %f3738, [%rd587+12]; mul.ftz.f32 %f13813, %f13809, %f3738; fma.rn.ftz.f32 %f13814, %f3737, %f3736, %f13813; ld.local.f32 %f3739, [%rd587+16]; fma.rn.ftz.f32 %f3740, %f13810, %f3739, %f13814; setp.gt.ftz.f32 %p1955, %f3740, 0f00000000; @%p1955 bra $L__BB106_2198; bra.uni $L__BB106_1852; $L__BB106_2198: setp.eq.ftz.f32 %p1957, %f13811, %f13812; @%p1957 bra $L__BB106_2200; bra.uni $L__BB106_2199; $L__BB106_2200: setp.neu.ftz.f32 %p1958, %f3736, %f3738; setp.neu.ftz.f32 %p1959, %f3736, %f3739; or.pred %p1960, %p1958, %p1959; neg.ftz.f32 %f13842, %f3739; selp.f32 %f13843, %f13842, %f3739, %p1960; add.ftz.f32 %f13844, %f3736, %f13843; neg.ftz.f32 %f13845, %f3738; selp.f32 %f13846, %f3738, %f13845, %p1960; sub.ftz.f32 %f13847, %f13846, %f3736; sub.ftz.f32 %f13848, %f3739, %f3738; mul.ftz.f32 %f13849, %f13844, %f13844; fma.rn.ftz.f32 %f13850, %f13848, %f13848, %f13849; fma.rn.ftz.f32 %f13851, %f13847, %f13847, %f13850; sqrt.approx.ftz.f32 %f13852, %f13851; div.approx.ftz.f32 %f16624, %f13848, %f13852; div.approx.ftz.f32 %f16625, %f13844, %f13852; div.approx.ftz.f32 %f16626, %f13847, %f13852; mul.ftz.f32 %f13853, %f3738, %f16626; mul.ftz.f32 %f13854, %f3739, %f16625; sub.ftz.f32 %f16621, %f13853, %f13854; mul.ftz.f32 %f13855, %f3739, %f16624; mul.ftz.f32 %f13856, %f3736, %f16626; sub.ftz.f32 %f16622, %f13855, %f13856; mul.ftz.f32 %f13857, %f3736, %f16625; mul.ftz.f32 %f13858, %f3738, %f16624; sub.ftz.f32 %f16623, %f13857, %f13858; bra.uni $L__BB106_2201; $L__BB106_2294: mul.ftz.f32 %f14916, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f14917, %f14916; sqrt.approx.ftz.f32 %f14918, %f2775; mul.ftz.f32 %f14919, %f14918, %f14917; sin.approx.ftz.f32 %f14920, %f14916; mul.ftz.f32 %f14921, %f14918, %f14920; mul.ftz.f32 %f14922, %f14919, %f14919; mov.f32 %f14923, 0f3F800000; sub.ftz.f32 %f14924, %f14923, %f14922; mul.ftz.f32 %f14925, %f14921, %f14921; sub.ftz.f32 %f14926, %f14924, %f14925; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f14927, %f14926, %f16738; sqrt.approx.ftz.f32 %f4059, %f14927; ld.local.f32 %f4060, [%rd587+12]; ld.local.f32 %f4061, [%rd587+8]; setp.neu.ftz.f32 %p2037, %f4061, %f4060; ld.local.f32 %f4062, [%rd587+16]; setp.neu.ftz.f32 %p2038, %f4061, %f4062; or.pred %p2039, %p2037, %p2038; sub.ftz.f32 %f14928, %f4062, %f4060; neg.ftz.f32 %f14929, %f4060; neg.ftz.f32 %f14930, %f4062; selp.f32 %f14931, %f14930, %f4062, %p2039; add.ftz.f32 %f14932, %f4061, %f14931; selp.f32 %f14933, %f4060, %f14929, %p2039; sub.ftz.f32 %f14934, %f14933, %f4061; mul.ftz.f32 %f14935, %f14932, %f14932; fma.rn.ftz.f32 %f14936, %f14928, %f14928, %f14935; fma.rn.ftz.f32 %f14937, %f14934, %f14934, %f14936; sqrt.approx.ftz.f32 %f14938, %f14937; div.approx.ftz.f32 %f14939, %f14928, %f14938; div.approx.ftz.f32 %f14940, %f14932, %f14938; div.approx.ftz.f32 %f14941, %f14934, %f14938; mul.ftz.f32 %f14942, %f4060, %f14941; mul.ftz.f32 %f14943, %f4062, %f14940; sub.ftz.f32 %f14944, %f14942, %f14943; mul.ftz.f32 %f14945, %f4062, %f14939; mul.ftz.f32 %f14946, %f4061, %f14941; sub.ftz.f32 %f14947, %f14945, %f14946; mul.ftz.f32 %f14948, %f4061, %f14940; mul.ftz.f32 %f14949, %f4060, %f14939; sub.ftz.f32 %f14950, %f14948, %f14949; mul.ftz.f32 %f14951, %f14921, %f14944; mul.ftz.f32 %f14952, %f14921, %f14947; mul.ftz.f32 %f14953, %f14921, %f14950; fma.rn.ftz.f32 %f14954, %f14919, %f14939, %f14951; fma.rn.ftz.f32 %f14955, %f14919, %f14940, %f14952; fma.rn.ftz.f32 %f14956, %f14919, %f14941, %f14953; fma.rn.ftz.f32 %f16735, %f4061, %f4059, %f14954; fma.rn.ftz.f32 %f16736, %f4060, %f4059, %f14955; fma.rn.ftz.f32 %f16737, %f4062, %f4059, %f14956; mul.ftz.f32 %f14957, %f2807, %f16736; fma.rn.ftz.f32 %f14958, %f2806, %f16735, %f14957; fma.rn.ftz.f32 %f14959, %f2808, %f16737, %f14958; setp.leu.ftz.f32 %p2040, %f14959, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2040 bra $L__BB106_2309; ld.local.f32 %f14964, [%rd7+-52]; ld.local.v2.f32 {%f14965, %f14966}, [%rd7+-48]; add.ftz.f32 %f14969, %f14964, %f16735; add.ftz.f32 %f14970, %f14965, %f16736; mul.ftz.f32 %f14971, %f14970, %f14970; fma.rn.ftz.f32 %f14972, %f14969, %f14969, %f14971; add.ftz.f32 %f14973, %f14966, %f16737; fma.rn.ftz.f32 %f14974, %f14973, %f14973, %f14972; sqrt.approx.ftz.f32 %f14975, %f14974; div.approx.ftz.f32 %f4066, %f14969, %f14975; div.approx.ftz.f32 %f4067, %f14970, %f14975; div.approx.ftz.f32 %f4068, %f14973, %f14975; mul.ftz.f32 %f14976, %f4060, %f16736; fma.rn.ftz.f32 %f14977, %f4061, %f16735, %f14976; fma.rn.ftz.f32 %f4069, %f4062, %f16737, %f14977; mul.ftz.f32 %f14978, %f14965, %f4060; fma.rn.ftz.f32 %f14979, %f14964, %f4061, %f14978; fma.rn.ftz.f32 %f14980, %f14966, %f4062, %f14979; setp.lt.ftz.f32 %p2041, %f4069, 0f00000000; setp.lt.ftz.f32 %p2042, %f14980, 0f00000000; or.pred %p2043, %p2042, %p2041; mov.f32 %f16738, 0f00000000; mov.f32 %f16739, %f16738; @%p2043 bra $L__BB106_2297; mul.ftz.f32 %f14981, %f16736, %f4067; fma.rn.ftz.f32 %f14982, %f16735, %f4066, %f14981; fma.rn.ftz.f32 %f14983, %f16737, %f4068, %f14982; sub.ftz.f32 %f14985, %f14923, %f14983; mov.f32 %f14986, 0f00000000; max.ftz.f32 %f14987, %f14985, %f14986; min.ftz.f32 %f14988, %f14987, %f14923; mul.ftz.f32 %f14989, %f14988, %f14988; mul.ftz.f32 %f14990, %f14989, %f14989; mul.ftz.f32 %f14991, %f14988, %f14990; mul.ftz.f32 %f16739, %f4069, %f14991; mul.ftz.f32 %f16738, %f4059, 0f3EA2F983; $L__BB106_2297: mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_2113: ld.local.v2.f32 {%f13134, %f13135}, [%rd7+-48]; ld.local.f32 %f3484, [%rd587+8]; ld.local.f32 %f3485, [%rd7+-52]; ld.local.f32 %f3486, [%rd587+12]; mul.ftz.f32 %f13136, %f13134, %f3486; fma.rn.ftz.f32 %f13137, %f3485, %f3484, %f13136; ld.local.f32 %f3487, [%rd587+16]; fma.rn.ftz.f32 %f3488, %f13135, %f3487, %f13137; setp.leu.ftz.f32 %p1880, %f3488, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1880 bra $L__BB106_2309; ld.local.v2.f32 {%f13138, %f13139}, [%rd587+20]; mul.ftz.f32 %f13140, %f13138, %f13138; mov.f32 %f13141, 0f40000000; div.approx.ftz.f32 %f13142, %f13141, %f13140; add.ftz.f32 %f3491, %f13142, 0fC0000000; mul.ftz.f32 %f13143, %f13139, %f13139; div.approx.ftz.f32 %f13144, %f13141, %f13143; add.ftz.f32 %f3492, %f13144, 0fC0000000; setp.eq.ftz.f32 %p1881, %f3491, %f3492; @%p1881 bra $L__BB106_2116; bra.uni $L__BB106_2115; $L__BB106_2116: setp.neu.ftz.f32 %p1882, %f3484, %f3486; setp.neu.ftz.f32 %p1883, %f3484, %f3487; or.pred %p1884, %p1882, %p1883; neg.ftz.f32 %f13167, %f3487; selp.f32 %f13168, %f13167, %f3487, %p1884; add.ftz.f32 %f13169, %f3484, %f13168; neg.ftz.f32 %f13170, %f3486; selp.f32 %f13171, %f3486, %f13170, %p1884; sub.ftz.f32 %f13172, %f13171, %f3484; sub.ftz.f32 %f13173, %f3487, %f3486; mul.ftz.f32 %f13174, %f13169, %f13169; fma.rn.ftz.f32 %f13175, %f13173, %f13173, %f13174; fma.rn.ftz.f32 %f13176, %f13172, %f13172, %f13175; sqrt.approx.ftz.f32 %f13177, %f13176; div.approx.ftz.f32 %f16570, %f13173, %f13177; div.approx.ftz.f32 %f16571, %f13169, %f13177; div.approx.ftz.f32 %f16572, %f13172, %f13177; mul.ftz.f32 %f13178, %f3486, %f16572; mul.ftz.f32 %f13179, %f3487, %f16571; sub.ftz.f32 %f16567, %f13178, %f13179; mul.ftz.f32 %f13180, %f3487, %f16570; mul.ftz.f32 %f13181, %f3484, %f16572; sub.ftz.f32 %f16568, %f13180, %f13181; mul.ftz.f32 %f13182, %f3484, %f16571; mul.ftz.f32 %f13183, %f3486, %f16570; sub.ftz.f32 %f16569, %f13182, %f13183; bra.uni $L__BB106_2117; $L__BB106_2108: mul.ftz.f32 %f13032, %f2799, %f2799; mov.f32 %f13033, 0f3F800000; sub.ftz.f32 %f13034, %f13033, %f13032; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f13035, %f16738, %f13034; sqrt.approx.ftz.f32 %f13036, %f13035; mul.ftz.f32 %f13037, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f13038, %f13037; mul.ftz.f32 %f13039, %f13036, %f13038; sin.approx.ftz.f32 %f13040, %f13037; mul.ftz.f32 %f13041, %f13036, %f13040; ld.local.f32 %f3470, [%rd587+12]; ld.local.f32 %f3471, [%rd587+8]; setp.neu.ftz.f32 %p1871, %f3471, %f3470; ld.local.f32 %f3472, [%rd587+16]; setp.neu.ftz.f32 %p1872, %f3471, %f3472; or.pred %p1873, %p1871, %p1872; sub.ftz.f32 %f13042, %f3472, %f3470; neg.ftz.f32 %f13043, %f3470; neg.ftz.f32 %f13044, %f3472; selp.f32 %f13045, %f13044, %f3472, %p1873; add.ftz.f32 %f13046, %f3471, %f13045; selp.f32 %f13047, %f3470, %f13043, %p1873; sub.ftz.f32 %f13048, %f13047, %f3471; mul.ftz.f32 %f13049, %f13046, %f13046; fma.rn.ftz.f32 %f13050, %f13042, %f13042, %f13049; fma.rn.ftz.f32 %f13051, %f13048, %f13048, %f13050; sqrt.approx.ftz.f32 %f13052, %f13051; div.approx.ftz.f32 %f13053, %f13042, %f13052; div.approx.ftz.f32 %f13054, %f13046, %f13052; div.approx.ftz.f32 %f13055, %f13048, %f13052; mul.ftz.f32 %f13056, %f3470, %f13055; mul.ftz.f32 %f13057, %f3472, %f13054; sub.ftz.f32 %f13058, %f13056, %f13057; mul.ftz.f32 %f13059, %f3472, %f13053; mul.ftz.f32 %f13060, %f3471, %f13055; sub.ftz.f32 %f13061, %f13059, %f13060; mul.ftz.f32 %f13062, %f3471, %f13054; mul.ftz.f32 %f13063, %f3470, %f13053; sub.ftz.f32 %f13064, %f13062, %f13063; mul.ftz.f32 %f13065, %f13041, %f13058; mul.ftz.f32 %f13066, %f13041, %f13061; mul.ftz.f32 %f13067, %f13041, %f13064; fma.rn.ftz.f32 %f13068, %f13039, %f13053, %f13065; fma.rn.ftz.f32 %f13069, %f13039, %f13054, %f13066; fma.rn.ftz.f32 %f13070, %f13039, %f13055, %f13067; fma.rn.ftz.f32 %f16735, %f2799, %f3471, %f13068; fma.rn.ftz.f32 %f16736, %f2799, %f3470, %f13069; fma.rn.ftz.f32 %f16737, %f2799, %f3472, %f13070; mul.ftz.f32 %f13071, %f2807, %f16736; fma.rn.ftz.f32 %f13072, %f2806, %f16735, %f13071; fma.rn.ftz.f32 %f13073, %f2808, %f16737, %f13072; setp.leu.ftz.f32 %p1874, %f13073, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1874 bra $L__BB106_2309; ld.local.f32 %f13074, [%rd7+-52]; ld.local.v2.f32 {%f13075, %f13076}, [%rd7+-48]; ld.local.f32 %f3476, [%rd587+24]; add.ftz.f32 %f13079, %f13074, %f16735; add.ftz.f32 %f13080, %f13075, %f16736; mul.ftz.f32 %f13081, %f13080, %f13080; fma.rn.ftz.f32 %f13082, %f13079, %f13079, %f13081; add.ftz.f32 %f13083, %f13076, %f16737; fma.rn.ftz.f32 %f13084, %f13083, %f13083, %f13082; sqrt.approx.ftz.f32 %f13085, %f13084; div.approx.ftz.f32 %f13086, %f13079, %f13085; div.approx.ftz.f32 %f13087, %f13080, %f13085; div.approx.ftz.f32 %f13088, %f13083, %f13085; mul.ftz.f32 %f13089, %f3470, %f16736; fma.rn.ftz.f32 %f13090, %f3471, %f16735, %f13089; fma.rn.ftz.f32 %f3477, %f3472, %f16737, %f13090; mul.ftz.f32 %f13091, %f13075, %f3470; fma.rn.ftz.f32 %f13092, %f13074, %f3471, %f13091; fma.rn.ftz.f32 %f3478, %f13076, %f3472, %f13092; mul.ftz.f32 %f13093, %f3470, %f13087; fma.rn.ftz.f32 %f13094, %f3471, %f13086, %f13093; fma.rn.ftz.f32 %f3479, %f3472, %f13088, %f13094; mul.ftz.f32 %f13095, %f13075, %f13087; fma.rn.ftz.f32 %f13096, %f13074, %f13086, %f13095; fma.rn.ftz.f32 %f13097, %f13076, %f13088, %f13096; abs.ftz.f32 %f3480, %f13097; abs.ftz.f32 %f13098, %f3478; setp.leu.ftz.f32 %p1876, %f13098, 0f3727C5AC; mov.pred %p2188, 0; @%p1876 bra $L__BB106_2111; abs.ftz.f32 %f13099, %f3479; setp.lt.ftz.f32 %p2188, %f13099, 0f3F7FFF58; $L__BB106_2111: setp.leu.ftz.f32 %p1877, %f3480, 0f3727C5AC; not.pred %p1878, %p2188; or.pred %p1879, %p1877, %p1878; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1879 bra $L__BB106_2309; div.approx.ftz.f32 %f13105, %f3479, %f3480; mov.f32 %f13106, 0f3727C5AC; max.ftz.f32 %f13107, %f13105, %f13106; mul.ftz.f32 %f13108, %f3478, %f13107; abs.ftz.f32 %f13109, %f13108; add.ftz.f32 %f13110, %f13109, %f13109; mul.ftz.f32 %f13111, %f3477, %f13107; abs.ftz.f32 %f13112, %f13111; add.ftz.f32 %f13113, %f13112, %f13112; mul.ftz.f32 %f13114, %f3479, %f3479; sub.ftz.f32 %f13116, %f13033, %f13114; mul.ftz.f32 %f13117, %f13116, %f13116; div.approx.ftz.f32 %f13118, %f13114, %f13116; mul.ftz.f32 %f13119, %f3476, %f13118; mul.ftz.f32 %f13120, %f13119, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13121, %f13120; mul.ftz.f32 %f13122, %f3476, %f13121; mul.ftz.f32 %f13123, %f13122, 0f3EA2F983; div.approx.ftz.f32 %f13124, %f13123, %f13117; min.ftz.f32 %f13125, %f13110, %f13113; min.ftz.f32 %f13126, %f13033, %f13125; mul.ftz.f32 %f13127, %f13124, %f13126; mul.ftz.f32 %f13128, %f13127, 0f3E800000; div.approx.ftz.f32 %f16739, %f13128, %f3478; mov.u32 %r3285, 6; mov.f32 %f16738, 0f3E22F983; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_1858: setp.neu.ftz.f32 %p1775, %f16546, %f16547; setp.neu.ftz.f32 %p1776, %f16546, %f16548; or.pred %p1777, %p1775, %p1776; neg.ftz.f32 %f12257, %f16548; selp.f32 %f12258, %f12257, %f16548, %p1777; add.ftz.f32 %f12259, %f16546, %f12258; neg.ftz.f32 %f12260, %f16547; selp.f32 %f12261, %f16547, %f12260, %p1777; sub.ftz.f32 %f12262, %f12261, %f16546; sub.ftz.f32 %f12263, %f16548, %f16547; mul.ftz.f32 %f12264, %f12259, %f12259; fma.rn.ftz.f32 %f12265, %f12263, %f12263, %f12264; fma.rn.ftz.f32 %f12266, %f12262, %f12262, %f12265; sqrt.approx.ftz.f32 %f12267, %f12266; div.approx.ftz.f32 %f3201, %f12263, %f12267; div.approx.ftz.f32 %f3202, %f12259, %f12267; div.approx.ftz.f32 %f3203, %f12262, %f12267; mul.ftz.f32 %f12268, %f16547, %f3203; mul.ftz.f32 %f12269, %f16548, %f3202; sub.ftz.f32 %f3204, %f12268, %f12269; mul.ftz.f32 %f12270, %f16548, %f3201; mul.ftz.f32 %f12271, %f16546, %f3203; sub.ftz.f32 %f3205, %f12270, %f12271; mul.ftz.f32 %f12272, %f16546, %f3202; mul.ftz.f32 %f12273, %f16547, %f3201; sub.ftz.f32 %f3206, %f12272, %f12273; mul.ftz.f32 %f12274, %f12249, %f3202; fma.rn.ftz.f32 %f12275, %f3197, %f3201, %f12274; fma.rn.ftz.f32 %f3207, %f12250, %f3203, %f12275; mul.ftz.f32 %f12276, %f12249, %f3205; fma.rn.ftz.f32 %f12277, %f3197, %f3204, %f12276; fma.rn.ftz.f32 %f3208, %f12250, %f3206, %f12277; ld.local.u64 %rd2687, [%rd587+36]; ld.f32 %f3209, [%rd2687]; ld.f32 %f3210, [%rd2687+4]; ld.f32 %f3211, [%rd2687+8]; ld.local.f32 %f3212, [%rd587+28]; ld.f32 %f3213, [%rd2687+12]; ld.f32 %f3214, [%rd2687+16]; ld.f32 %f3215, [%rd2687+20]; neg.ftz.f32 %f3216, %f3207; neg.ftz.f32 %f3217, %f3208; neg.ftz.f32 %f3218, %f16549; setp.lt.ftz.f32 %p1778, %f16549, 0fBF7FF972; mov.f32 %f16532, 0f00000000; @%p1778 bra $L__BB106_1861; setp.gt.ftz.f32 %p1779, %f16549, 0f3F7FF972; mov.f32 %f16532, 0fBF7FF972; @%p1779 bra $L__BB106_1861; mul.ftz.f32 %f12279, %f16549, %f16549; mov.f32 %f12280, 0f33D6BF95; max.ftz.f32 %f12281, %f12279, %f12280; rcp.approx.ftz.f32 %f12282, %f12281; mul.ftz.f32 %f12283, %f12251, %f3216; mul.ftz.f32 %f12284, %f12252, %f3217; mul.ftz.f32 %f12285, %f12284, %f12284; fma.rn.ftz.f32 %f12286, %f12283, %f12283, %f12285; fma.rn.ftz.f32 %f12287, %f12286, %f12282, 0f3F800000; sqrt.approx.ftz.f32 %f12288, %f12287; neg.ftz.f32 %f12289, %f12288; setp.ge.ftz.f32 %p1780, %f16549, 0f80000000; selp.f32 %f12290, %f12289, %f12288, %p1780; add.ftz.f32 %f12291, %f12290, 0fBF800000; mul.ftz.f32 %f16532, %f12291, 0f3F000000; $L__BB106_1861: fma.rn.ftz.f32 %f3221, %f3212, %f3212, 0fBF800000; mov.f32 %f16531, 0f3F800000; abs.ftz.f32 %f3222, %f16531; fma.rn.ftz.f32 %f3223, %f3222, %f3222, %f3221; setp.leu.ftz.f32 %p1781, %f3223, 0f00000000; mov.f32 %f16506, %f16531; @%p1781 bra $L__BB106_1863; sqrt.approx.ftz.f32 %f12293, %f3223; sub.ftz.f32 %f12294, %f12293, %f3222; add.ftz.f32 %f12295, %f3222, %f12293; div.approx.ftz.f32 %f12296, %f12294, %f12295; fma.rn.ftz.f32 %f12297, %f3222, %f12295, 0fBF800000; fma.rn.ftz.f32 %f12298, %f3222, %f12294, 0f3F800000; div.approx.ftz.f32 %f12299, %f12297, %f12298; mul.ftz.f32 %f12300, %f12296, 0f3F000000; mul.ftz.f32 %f12301, %f12296, %f12300; fma.rn.ftz.f32 %f12302, %f12299, %f12299, 0f3F800000; mul.ftz.f32 %f16506, %f12301, %f12302; $L__BB106_1863: setp.ne.s32 %p8, %r873, 27; rcp.approx.ftz.f32 %f3226, %f3212; sub.ftz.f32 %f3227, %f16531, %f16506; rcp.approx.ftz.f32 %f3228, %f3227; ld.local.u32 %r3279, [%rd7+168]; mov.f32 %f16533, 0f00000000; mov.u16 %rs155, 1; mov.u32 %r3280, 0; mov.pred %p1782, -1; mov.pred %p1809, 0; mov.f32 %f16530, %f16531; mov.f32 %f16529, %f16531; mov.f32 %f3232, %f3218; mov.f32 %f3233, %f3217; mov.f32 %f3234, %f3216; mov.f32 %f16516, %f16531; $L__BB106_1864: mov.f32 %f3236, %f16516; mad.lo.s32 %r885, %r3279, 1103515245, 12345; cvt.rn.f32.u32 %f12308, %r885; mul.ftz.f32 %f3238, %f12308, 0f2F800000; setp.gt.ftz.f32 %p1783, %f3232, 0f3F7FF972; mov.pred %p2186, %p1782; @%p1783 bra $L__BB106_2042; setp.lt.ftz.f32 %p1784, %f3232, 0fBF7FF972; @%p1784 bra $L__BB106_2041; bra.uni $L__BB106_1866; $L__BB106_2041: mul.ftz.f32 %f16516, %f3236, %f3238; mov.pred %p2186, %p1809; bra.uni $L__BB106_2042; $L__BB106_1866: abs.ftz.f32 %f12309, %f3232; setp.ltu.ftz.f32 %p1786, %f12309, 0f38D1B717; mov.f32 %f16516, %f3236; mov.pred %p2186, %p1809; @%p1786 bra $L__BB106_2042; mov.f32 %f12310, 0f3F800000; sub.ftz.f32 %f12311, %f12310, %f16533; setp.lt.ftz.f32 %p1788, %f12311, %f3238; mov.f32 %f16516, %f3236; mov.pred %p2186, %p1782; @%p1788 bra $L__BB106_2042; setp.ge.ftz.f32 %p1790, %f16532, 0f00000000; mov.f32 %f16516, %f12310; mov.pred %p2186, %p1809; @%p1790 bra $L__BB106_2042; mov.f32 %f12313, 0f3F800000; sub.ftz.f32 %f12314, %f12313, %f3238; lg2.approx.ftz.f32 %f12315, %f12314; mov.f32 %f12316, 0fBF800000; div.approx.ftz.f32 %f12317, %f12316, %f16532; mul.ftz.f32 %f12318, %f12317, %f12315; ex2.approx.ftz.f32 %f12319, %f12318; mul.ftz.f32 %f16516, %f3236, %f12319; mov.pred %p2186, %p1809; $L__BB106_2042: @%p2186 bra $L__BB106_2074; mad.lo.s32 %r2776, %r885, 1103515245, 12345; cvt.rn.f32.u32 %f12320, %r2776; mul.ftz.f32 %f3242, %f12320, 0f2F800000; mad.lo.s32 %r886, %r2776, 1103515245, 12345; cvt.rn.f32.u32 %f12321, %r886; mul.ftz.f32 %f3243, %f12321, 0f2F800000; neg.ftz.f32 %f3244, %f3234; mul.ftz.f32 %f12322, %f12251, %f3244; neg.ftz.f32 %f3245, %f3233; mul.ftz.f32 %f12323, %f12252, %f3245; mul.ftz.f32 %f12324, %f12323, %f12323; fma.rn.ftz.f32 %f12325, %f12322, %f12322, %f12324; fma.rn.ftz.f32 %f12326, %f3232, %f3232, %f12325; sqrt.approx.ftz.f32 %f12327, %f12326; div.approx.ftz.f32 %f16519, %f12322, %f12327; div.approx.ftz.f32 %f16520, %f12323, %f12327; neg.ftz.f32 %f3248, %f3232; div.approx.ftz.f32 %f3249, %f3248, %f12327; setp.gt.ftz.f32 %p1793, %f3249, 0f3F7FF972; @%p1793 bra $L__BB106_2049; abs.ftz.f32 %f12328, %f3249; setp.lt.ftz.f32 %p1794, %f12328, 0f358637BD; @%p1794 bra $L__BB106_2049; bra.uni $L__BB106_2045; $L__BB106_2049: mov.f32 %f12372, 0f3F800000; sub.ftz.f32 %f12373, %f12372, %f3243; mov.f32 %f12374, 0f33D6BF95; max.ftz.f32 %f12375, %f12373, %f12374; div.approx.ftz.f32 %f12376, %f3243, %f12375; sqrt.approx.ftz.f32 %f12377, %f12376; mul.ftz.f32 %f12378, %f3242, 0f40C90FDB; cos.approx.ftz.f32 %f12379, %f12378; sin.approx.ftz.f32 %f12380, %f12378; mul.ftz.f32 %f16518, %f12377, %f12380; mul.ftz.f32 %f16517, %f12377, %f12379; $L__BB106_2050: mul.ftz.f32 %f12381, %f16520, %f16520; fma.rn.ftz.f32 %f12382, %f16519, %f16519, %f12381; add.ftz.f32 %f12383, %f12382, 0f00000000; sqrt.approx.ftz.f32 %f3267, %f12383; setp.neu.ftz.f32 %p1802, %f3267, 0f00000000; @%p1802 bra $L__BB106_2051; bra.uni $L__BB106_2052; $L__BB106_2051: rcp.approx.ftz.f32 %f12384, %f3267; mul.ftz.f32 %f16520, %f16520, %f12384; mul.ftz.f32 %f16519, %f16519, %f12384; $L__BB106_2052: mul.ftz.f32 %f12386, %f16517, %f16519; mul.ftz.f32 %f12387, %f16518, %f16520; sub.ftz.f32 %f12388, %f12386, %f12387; mul.ftz.f32 %f12389, %f12251, %f12388; mul.ftz.f32 %f12390, %f16518, %f16519; fma.rn.ftz.f32 %f12391, %f16517, %f16520, %f12390; mul.ftz.f32 %f12392, %f12252, %f12391; neg.ftz.f32 %f12393, %f12389; neg.ftz.f32 %f12394, %f12392; mul.ftz.f32 %f12395, %f12392, %f12392; fma.rn.ftz.f32 %f12396, %f12389, %f12389, %f12395; add.ftz.f32 %f12397, %f12396, 0f3F800000; mov.f32 %f16521, 0f3F800000; sqrt.approx.ftz.f32 %f12398, %f12397; div.approx.ftz.f32 %f3277, %f12393, %f12398; div.approx.ftz.f32 %f3278, %f12394, %f12398; div.approx.ftz.f32 %f3279, %f16521, %f12398; setp.ne.s32 %p1803, %r3280, 0; and.pred %p1804, %p1803, %p8; mul.ftz.f32 %f12399, %f3209, %f16529; selp.f32 %f3280, %f12399, %f16529, %p1804; mul.ftz.f32 %f12400, %f3210, %f16530; selp.f32 %f3281, %f12400, %f16530, %p1804; mul.ftz.f32 %f12401, %f3211, %f16531; selp.f32 %f3282, %f12401, %f16531, %p1804; mad.lo.s32 %r3279, %r886, 1103515245, 12345; and.b16 %rs114, %rs155, 255; setp.eq.s16 %p1805, %rs114, 0; selp.f32 %f3283, %f3226, %f3212, %p1805; mul.ftz.f32 %f12402, %f3278, %f3245; fma.rn.ftz.f32 %f12403, %f3277, %f3244, %f12402; fma.rn.ftz.f32 %f3284, %f3279, %f3248, %f12403; abs.ftz.f32 %f3285, %f3284; fma.rn.ftz.f32 %f12404, %f3283, %f3283, 0fBF800000; mul.ftz.f32 %f3286, %f3285, %f3285; add.ftz.f32 %f3287, %f12404, %f3286; setp.leu.ftz.f32 %p1806, %f3287, 0f00000000; @%p1806 bra $L__BB106_2054; sqrt.approx.ftz.f32 %f12405, %f3287; sub.ftz.f32 %f12406, %f12405, %f3285; add.ftz.f32 %f12407, %f3285, %f12405; div.approx.ftz.f32 %f12408, %f12406, %f12407; fma.rn.ftz.f32 %f12409, %f3285, %f12407, 0fBF800000; fma.rn.ftz.f32 %f12410, %f3285, %f12406, 0f3F800000; div.approx.ftz.f32 %f12411, %f12409, %f12410; mul.ftz.f32 %f12412, %f12408, 0f3F000000; mul.ftz.f32 %f12413, %f12408, %f12412; fma.rn.ftz.f32 %f12414, %f12411, %f12411, 0f3F800000; mul.ftz.f32 %f16521, %f12413, %f12414; $L__BB106_2054: cvt.rn.f32.u32 %f12415, %r3279; mul.ftz.f32 %f12416, %f12415, 0f2F800000; setp.gt.ftz.f32 %p1807, %f16521, %f12416; @%p1807 bra $L__BB106_2056; bra.uni $L__BB106_2055; $L__BB106_2056: add.ftz.f32 %f12442, %f3277, %f3277; add.ftz.f32 %f12443, %f3278, %f3278; add.ftz.f32 %f12444, %f3279, %f3279; fma.rn.ftz.f32 %f3232, %f12444, %f3284, %f3232; fma.rn.ftz.f32 %f3233, %f12443, %f3284, %f3233; fma.rn.ftz.f32 %f3234, %f12442, %f3284, %f3234; mov.pred %p2187, %p1809; bra.uni $L__BB106_2057; $L__BB106_2055: rcp.approx.ftz.f32 %f12417, %f3283; mov.f32 %f12418, 0f3F800000; mul.ftz.f32 %f12419, %f3284, %f3284; sub.ftz.f32 %f12420, %f12418, %f12419; mul.ftz.f32 %f12421, %f12420, %f12417; mul.ftz.f32 %f12422, %f12417, %f12421; sub.ftz.f32 %f12423, %f12418, %f12422; mov.f32 %f12424, 0f00000000; max.ftz.f32 %f12425, %f12423, %f12424; sqrt.approx.ftz.f32 %f12426, %f12425; mul.ftz.f32 %f12427, %f3284, %f12417; sub.ftz.f32 %f12428, %f12427, %f12426; mul.ftz.f32 %f12429, %f3277, %f12428; mul.ftz.f32 %f12430, %f3278, %f12428; mul.ftz.f32 %f12431, %f3279, %f12428; mul.ftz.f32 %f12432, %f12417, %f3244; mul.ftz.f32 %f12433, %f12417, %f3245; mul.ftz.f32 %f12434, %f12417, %f3248; sub.ftz.f32 %f12435, %f12429, %f12432; sub.ftz.f32 %f12436, %f12430, %f12433; sub.ftz.f32 %f12437, %f12431, %f12434; mul.ftz.f32 %f12438, %f12436, %f12436; fma.rn.ftz.f32 %f12439, %f12435, %f12435, %f12438; fma.rn.ftz.f32 %f12440, %f12437, %f12437, %f12439; sqrt.approx.ftz.f32 %f12441, %f12440; div.approx.ftz.f32 %f3232, %f12437, %f12441; div.approx.ftz.f32 %f3233, %f12436, %f12441; div.approx.ftz.f32 %f3234, %f12435, %f12441; mov.pred %p2187, %p1782; $L__BB106_2057: not.pred %p1810, %p2187; @%p1810 bra $L__BB106_2059; neg.ftz.f32 %f3234, %f3234; neg.ftz.f32 %f3233, %f3233; neg.ftz.f32 %f3232, %f3232; selp.u16 %rs155, 1, 0, %p1805; $L__BB106_2059: mov.f32 %f16529, %f3280; mov.f32 %f16530, %f3281; mov.f32 %f16531, %f3282; @%p8 bra $L__BB106_2066; @%p2187 bra $L__BB106_2065; bra.uni $L__BB106_2061; $L__BB106_2065: mul.ftz.f32 %f16529, %f3209, %f3280; mul.ftz.f32 %f16530, %f3210, %f3281; mul.ftz.f32 %f16531, %f3211, %f3282; bra.uni $L__BB106_2066; $L__BB106_2045: add.ftz.f32 %f12331, %f3249, 0f3F800000; mul.ftz.f32 %f3250, %f12331, 0f3F000000; setp.lt.ftz.f32 %p1795, %f3250, 0f38D1B717; mov.f32 %f16517, 0f00000000; mov.f32 %f16518, %f16517; @%p1795 bra $L__BB106_2050; add.ftz.f32 %f12334, %f3243, %f3243; mul.ftz.f32 %f12335, %f12334, %f3250; div.approx.ftz.f32 %f12336, %f12335, %f3249; add.ftz.f32 %f3251, %f12336, 0fBF800000; mul.ftz.f32 %f3252, %f3251, %f3251; add.ftz.f32 %f3253, %f3252, 0fBF800000; abs.ftz.f32 %f12337, %f3253; setp.lt.ftz.f32 %p1796, %f12337, 0f33D6BF95; mov.f32 %f16518, %f16517; @%p1796 bra $L__BB106_2050; mul.ftz.f32 %f12338, %f3249, %f3249; mov.f32 %f12339, 0f3F800000; sub.ftz.f32 %f12340, %f12339, %f12338; mov.f32 %f12341, 0f00000000; max.ftz.f32 %f12342, %f12340, %f12341; sqrt.approx.ftz.f32 %f12343, %f12342; div.approx.ftz.f32 %f12344, %f12343, %f3249; rcp.approx.ftz.f32 %f12345, %f3253; mul.ftz.f32 %f12346, %f12344, %f12344; mul.ftz.f32 %f12347, %f12346, %f12345; mul.ftz.f32 %f12348, %f12345, %f12347; sub.ftz.f32 %f12349, %f3252, %f12346; mul.ftz.f32 %f12350, %f12349, %f12345; sub.ftz.f32 %f12351, %f12348, %f12350; max.ftz.f32 %f12352, %f12351, %f12341; sqrt.approx.ftz.f32 %f12353, %f12352; mul.ftz.f32 %f12354, %f12344, %f12345; add.ftz.f32 %f12355, %f12354, %f12353; rcp.approx.ftz.f32 %f12356, %f12344; setp.gt.ftz.f32 %p1797, %f12355, %f12356; setp.lt.ftz.f32 %p1798, %f3251, 0f00000000; or.pred %p1799, %p1798, %p1797; sub.ftz.f32 %f12357, %f12354, %f12353; mov.f32 %f12358, 0f3F000000; sub.ftz.f32 %f12359, %f12358, %f3242; add.ftz.f32 %f12360, %f3242, 0fBF000000; setp.ge.ftz.f32 %p1800, %f3242, 0f3F000000; selp.f32 %f12361, %f12360, %f12359, %p1800; add.ftz.f32 %f12362, %f12361, %f12361; fma.rn.ftz.f32 %f12363, %f12362, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f12364, %f12362, %f12363, 0f3EED4413; mul.ftz.f32 %f12365, %f12362, %f12364; fma.rn.ftz.f32 %f12366, %f12362, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f12367, %f12362, %f12366, 0fBF800000; fma.rn.ftz.f32 %f12368, %f12362, %f12367, 0f3F191676; div.approx.ftz.f32 %f12369, %f12365, %f12368; selp.f32 %f16517, %f12357, %f12355, %p1799; fma.rn.ftz.f32 %f12370, %f16517, %f16517, 0f3F800000; sqrt.approx.ftz.f32 %f12371, %f12370; mul.ftz.f32 %f16518, %f12369, %f12371; setp.ltu.ftz.f32 %p1801, %f3242, 0f3F000000; @%p1801 bra $L__BB106_2048; bra.uni $L__BB106_2050; $L__BB106_2048: neg.ftz.f32 %f16518, %f16518; bra.uni $L__BB106_2050; $L__BB106_2061: add.ftz.f32 %f3308, %f3221, %f3286; setp.leu.ftz.f32 %p1813, %f3308, 0f00000000; mov.f32 %f16528, %f3227; @%p1813 bra $L__BB106_2063; sqrt.approx.ftz.f32 %f12445, %f3308; sub.ftz.f32 %f12446, %f12445, %f3285; add.ftz.f32 %f12447, %f3285, %f12445; div.approx.ftz.f32 %f12448, %f12446, %f12447; fma.rn.ftz.f32 %f12449, %f3285, %f12447, 0fBF800000; fma.rn.ftz.f32 %f12450, %f3285, %f12446, 0f3F800000; div.approx.ftz.f32 %f12451, %f12449, %f12450; mul.ftz.f32 %f12452, %f12448, 0f3F000000; mul.ftz.f32 %f12453, %f12448, %f12452; fma.rn.ftz.f32 %f12454, %f12451, %f12451, 0f3F800000; mul.ftz.f32 %f12455, %f12453, %f12454; sub.ftz.f32 %f16528, %f12455, %f16506; $L__BB106_2063: mul.ftz.f32 %f12456, %f3228, %f16528; mov.f32 %f12457, 0f3F800000; sub.ftz.f32 %f12458, %f12457, %f12456; fma.rn.ftz.f32 %f16529, %f3213, %f12458, %f12456; fma.rn.ftz.f32 %f16530, %f3214, %f12458, %f12456; fma.rn.ftz.f32 %f16531, %f3215, %f12458, %f12456; setp.eq.s32 %p1814, %r3280, 0; @%p1814 bra $L__BB106_2066; mul.ftz.f32 %f16529, %f3280, %f16529; mul.ftz.f32 %f16530, %f3281, %f16530; mul.ftz.f32 %f16531, %f3282, %f16531; $L__BB106_2066: setp.gt.ftz.f32 %p1815, %f3232, 0f3F7FF972; mov.f32 %f16532, 0f00000000; @%p1815 bra $L__BB106_2069; setp.lt.ftz.f32 %p1816, %f3232, 0fBF7FF972; mov.f32 %f16532, 0fBF7FF972; @%p1816 bra $L__BB106_2069; mul.ftz.f32 %f12461, %f3232, %f3232; mov.f32 %f12462, 0f33D6BF95; max.ftz.f32 %f12463, %f12461, %f12462; rcp.approx.ftz.f32 %f12464, %f12463; mul.ftz.f32 %f12465, %f12251, %f3234; mul.ftz.f32 %f12466, %f12252, %f3233; mul.ftz.f32 %f12467, %f12466, %f12466; fma.rn.ftz.f32 %f12468, %f12465, %f12465, %f12467; fma.rn.ftz.f32 %f12469, %f12468, %f12464, 0f3F800000; sqrt.approx.ftz.f32 %f12470, %f12469; neg.ftz.f32 %f12471, %f12470; setp.le.ftz.f32 %p1817, %f3232, 0f00000000; selp.f32 %f12472, %f12471, %f12470, %p1817; add.ftz.f32 %f12473, %f12472, 0fBF800000; mul.ftz.f32 %f16532, %f12473, 0f3F000000; $L__BB106_2069: mov.f32 %f16533, 0f3F800000; @%p1815 bra $L__BB106_2072; setp.lt.ftz.f32 %p1819, %f3232, 0f3727C5AC; mov.f32 %f16533, 0f00000000; @%p1819 bra $L__BB106_2072; lg2.approx.ftz.f32 %f12476, %f16516; mul.ftz.f32 %f12477, %f16532, %f12476; ex2.approx.ftz.f32 %f16533, %f12477; $L__BB106_2072: add.s32 %r3280, %r3280, 1; setp.lt.u32 %p1820, %r3280, 10; @%p1820 bra $L__BB106_1864; st.local.u32 [%rd7+168], %r3279; mov.f32 %f3234, 0f00000000; mov.f32 %f3232, 0f3F800000; mov.f32 %f3233, %f3234; mov.f32 %f16529, %f3234; mov.f32 %f16530, %f3234; mov.f32 %f16531, %f3234; bra.uni $L__BB106_2076; $L__BB106_2307: mul.ftz.f32 %f15147, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f15148, %f15147; sqrt.approx.ftz.f32 %f15149, %f2775; mul.ftz.f32 %f15150, %f15149, %f15148; sin.approx.ftz.f32 %f15151, %f15147; mul.ftz.f32 %f15152, %f15149, %f15151; mul.ftz.f32 %f15153, %f15150, %f15150; mov.f32 %f15154, 0f3F800000; sub.ftz.f32 %f15155, %f15154, %f15153; mul.ftz.f32 %f15156, %f15152, %f15152; sub.ftz.f32 %f15157, %f15155, %f15156; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f15158, %f15157, %f16738; sqrt.approx.ftz.f32 %f4122, %f15158; ld.local.f32 %f15159, [%rd587+12]; ld.local.f32 %f15160, [%rd587+8]; setp.neu.ftz.f32 %p2057, %f15160, %f15159; ld.local.f32 %f15161, [%rd587+16]; setp.neu.ftz.f32 %p2058, %f15160, %f15161; or.pred %p2059, %p2057, %p2058; sub.ftz.f32 %f15162, %f15161, %f15159; neg.ftz.f32 %f15163, %f15159; neg.ftz.f32 %f15164, %f15161; selp.f32 %f15165, %f15164, %f15161, %p2059; add.ftz.f32 %f15166, %f15160, %f15165; selp.f32 %f15167, %f15159, %f15163, %p2059; sub.ftz.f32 %f15168, %f15167, %f15160; mul.ftz.f32 %f15169, %f15166, %f15166; fma.rn.ftz.f32 %f15170, %f15162, %f15162, %f15169; fma.rn.ftz.f32 %f15171, %f15168, %f15168, %f15170; sqrt.approx.ftz.f32 %f15172, %f15171; div.approx.ftz.f32 %f15173, %f15162, %f15172; div.approx.ftz.f32 %f15174, %f15166, %f15172; div.approx.ftz.f32 %f15175, %f15168, %f15172; mul.ftz.f32 %f15176, %f15159, %f15175; mul.ftz.f32 %f15177, %f15161, %f15174; sub.ftz.f32 %f15178, %f15176, %f15177; mul.ftz.f32 %f15179, %f15161, %f15173; mul.ftz.f32 %f15180, %f15160, %f15175; sub.ftz.f32 %f15181, %f15179, %f15180; mul.ftz.f32 %f15182, %f15160, %f15174; mul.ftz.f32 %f15183, %f15159, %f15173; sub.ftz.f32 %f15184, %f15182, %f15183; mul.ftz.f32 %f15185, %f15152, %f15178; mul.ftz.f32 %f15186, %f15152, %f15181; mul.ftz.f32 %f15187, %f15152, %f15184; fma.rn.ftz.f32 %f15188, %f15150, %f15173, %f15185; fma.rn.ftz.f32 %f15189, %f15150, %f15174, %f15186; fma.rn.ftz.f32 %f15190, %f15150, %f15175, %f15187; fma.rn.ftz.f32 %f16735, %f15160, %f4122, %f15188; fma.rn.ftz.f32 %f16736, %f15159, %f4122, %f15189; fma.rn.ftz.f32 %f16737, %f15161, %f4122, %f15190; mul.ftz.f32 %f15191, %f2807, %f16736; fma.rn.ftz.f32 %f15192, %f2806, %f16735, %f15191; fma.rn.ftz.f32 %f15193, %f2808, %f16737, %f15192; setp.leu.ftz.f32 %p2060, %f15193, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2060 bra $L__BB106_2309; mul.ftz.f32 %f16738, %f4122, 0f3EA2F983; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_2303: mul.ftz.f32 %f15079, %f2799, %f2799; mov.f32 %f15080, 0f3F800000; sub.ftz.f32 %f15081, %f15080, %f15079; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f15082, %f16738, %f15081; sqrt.approx.ftz.f32 %f15083, %f15082; mul.ftz.f32 %f15084, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f15085, %f15084; mul.ftz.f32 %f15086, %f15083, %f15085; sin.approx.ftz.f32 %f15087, %f15084; mul.ftz.f32 %f15088, %f15083, %f15087; ld.local.f32 %f4110, [%rd587+12]; ld.local.f32 %f4111, [%rd587+8]; setp.neu.ftz.f32 %p2052, %f4111, %f4110; ld.local.f32 %f4112, [%rd587+16]; setp.neu.ftz.f32 %p2053, %f4111, %f4112; or.pred %p2054, %p2052, %p2053; sub.ftz.f32 %f15089, %f4112, %f4110; neg.ftz.f32 %f15090, %f4110; neg.ftz.f32 %f15091, %f4112; selp.f32 %f15092, %f15091, %f4112, %p2054; add.ftz.f32 %f15093, %f4111, %f15092; selp.f32 %f15094, %f4110, %f15090, %p2054; sub.ftz.f32 %f15095, %f15094, %f4111; mul.ftz.f32 %f15096, %f15093, %f15093; fma.rn.ftz.f32 %f15097, %f15089, %f15089, %f15096; fma.rn.ftz.f32 %f15098, %f15095, %f15095, %f15097; sqrt.approx.ftz.f32 %f15099, %f15098; div.approx.ftz.f32 %f15100, %f15089, %f15099; div.approx.ftz.f32 %f15101, %f15093, %f15099; div.approx.ftz.f32 %f15102, %f15095, %f15099; mul.ftz.f32 %f15103, %f4110, %f15102; mul.ftz.f32 %f15104, %f4112, %f15101; sub.ftz.f32 %f15105, %f15103, %f15104; mul.ftz.f32 %f15106, %f4112, %f15100; mul.ftz.f32 %f15107, %f4111, %f15102; sub.ftz.f32 %f15108, %f15106, %f15107; mul.ftz.f32 %f15109, %f4111, %f15101; mul.ftz.f32 %f15110, %f4110, %f15100; sub.ftz.f32 %f15111, %f15109, %f15110; mul.ftz.f32 %f15112, %f15088, %f15105; mul.ftz.f32 %f15113, %f15088, %f15108; mul.ftz.f32 %f15114, %f15088, %f15111; fma.rn.ftz.f32 %f15115, %f15086, %f15100, %f15112; fma.rn.ftz.f32 %f15116, %f15086, %f15101, %f15113; fma.rn.ftz.f32 %f15117, %f15086, %f15102, %f15114; fma.rn.ftz.f32 %f16735, %f2799, %f4111, %f15115; fma.rn.ftz.f32 %f16736, %f2799, %f4110, %f15116; fma.rn.ftz.f32 %f16737, %f2799, %f4112, %f15117; mul.ftz.f32 %f15118, %f2807, %f16736; fma.rn.ftz.f32 %f15119, %f2806, %f16735, %f15118; fma.rn.ftz.f32 %f15120, %f2808, %f16737, %f15119; setp.leu.ftz.f32 %p2055, %f15120, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2055 bra $L__BB106_2309; ld.local.f32 %f15121, [%rd7+-52]; ld.local.v2.f32 {%f15122, %f15123}, [%rd7+-48]; mul.ftz.f32 %f15126, %f16736, %f4110; fma.rn.ftz.f32 %f15127, %f16735, %f4111, %f15126; fma.rn.ftz.f32 %f15128, %f16737, %f4112, %f15127; mov.f32 %f15129, 0f00000000; max.ftz.f32 %f4116, %f15128, %f15129; mul.ftz.f32 %f15130, %f15122, %f4110; fma.rn.ftz.f32 %f15131, %f15121, %f4111, %f15130; fma.rn.ftz.f32 %f15132, %f15123, %f4112, %f15131; max.ftz.f32 %f4117, %f15132, %f15129; mul.ftz.f32 %f15133, %f15122, %f16736; fma.rn.ftz.f32 %f15134, %f15121, %f16735, %f15133; fma.rn.ftz.f32 %f15135, %f15123, %f16737, %f15134; mul.ftz.f32 %f15136, %f4116, %f4117; sub.ftz.f32 %f16734, %f15135, %f15136; setp.leu.ftz.f32 %p2056, %f16734, 0f00000000; @%p2056 bra $L__BB106_2306; max.ftz.f32 %f15137, %f4116, %f4117; add.ftz.f32 %f15138, %f15137, 0f00800000; div.approx.ftz.f32 %f16734, %f16734, %f15138; $L__BB106_2306: ld.local.f32 %f15140, [%rd587+28]; ld.local.f32 %f15141, [%rd587+24]; fma.rn.ftz.f32 %f15142, %f16734, %f15140, %f15141; mul.ftz.f32 %f16739, %f4116, %f15142; mov.f32 %f16738, 0f3E22F983; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_2097: ld.local.v2.f32 {%f12708, %f12709}, [%rd7+-48]; ld.local.f32 %f3427, [%rd587+36]; ld.local.f32 %f3428, [%rd587+20]; ld.local.f32 %f12712, [%rd7+-52]; ld.local.f32 %f3429, [%rd587+24]; mul.ftz.f32 %f12713, %f12708, %f3429; fma.rn.ftz.f32 %f12714, %f12712, %f3428, %f12713; ld.local.f32 %f3430, [%rd587+28]; fma.rn.ftz.f32 %f12715, %f12709, %f3430, %f12714; mul.ftz.f32 %f12716, %f3428, %f12715; mul.ftz.f32 %f12717, %f3429, %f12715; mul.ftz.f32 %f12718, %f3430, %f12715; sub.ftz.f32 %f12719, %f12712, %f12716; sub.ftz.f32 %f12720, %f12708, %f12717; sub.ftz.f32 %f12721, %f12709, %f12718; mul.ftz.f32 %f12722, %f12720, %f12720; fma.rn.ftz.f32 %f12723, %f12719, %f12719, %f12722; fma.rn.ftz.f32 %f12724, %f12721, %f12721, %f12723; sqrt.approx.ftz.f32 %f12725, %f12724; div.approx.ftz.f32 %f3431, %f12719, %f12725; div.approx.ftz.f32 %f3432, %f12720, %f12725; div.approx.ftz.f32 %f3433, %f12721, %f12725; abs.ftz.f32 %f12726, %f12715; setp.lt.ftz.f32 %p1840, %f12726, 0f3F800000; mov.f32 %f12727, 0f3F800000; add.ftz.f32 %f12728, %f12726, 0fBF800000; add.ftz.f32 %f12729, %f12728, 0f3F800000; selp.f32 %f12730, %f12729, 0f3F800000, %p1840; sub.ftz.f32 %f12731, %f12727, %f12730; sqrt.approx.ftz.f32 %f12732, %f12731; fma.rn.ftz.f32 %f12733, %f12730, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f12734, %f12730, %f12733, 0fBE5A6B93; fma.rn.ftz.f32 %f12735, %f12730, %f12734, 0f3FC90FDB; mov.f32 %f12736, 0f3FC90FDB; mul.ftz.f32 %f12737, %f12732, %f12735; setp.lt.ftz.f32 %p1841, %f12715, 0f00000000; mov.f32 %f16565, 0f00000000; mov.f32 %f12738, 0f40490FDB; sub.ftz.f32 %f12739, %f12738, %f12737; selp.f32 %f12740, %f12739, %f12737, %p1841; sub.ftz.f32 %f3434, %f12736, %f12740; ld.local.f32 %f3435, [%rd587+32]; rcp.approx.ftz.f32 %f3436, %f3435; add.ftz.f32 %f12741, %f3434, 0f3FC90FDB; mul.ftz.f32 %f12742, %f12741, 0f3F000000; ld.local.f32 %f3437, [%rd587+40]; sub.ftz.f32 %f12743, %f12742, %f3437; mul.ftz.f32 %f3438, %f3436, %f12743; abs.ftz.f32 %f3439, %f3438; setp.eq.ftz.f32 %p1842, %f3439, 0f00000000; mov.f32 %f16564, %f16565; @%p1842 bra $L__BB106_2102; setp.eq.ftz.f32 %p1843, %f3439, 0f3F800000; mov.f32 %f16564, %f12727; @%p1843 bra $L__BB106_2102; setp.gt.ftz.f32 %p1844, %f3439, 0f3F800000; @%p1844 bra $L__BB106_2101; bra.uni $L__BB106_2100; $L__BB106_2101: mov.f32 %f12746, 0f3F800000; div.approx.ftz.f32 %f16564, %f12746, %f3439; bra.uni $L__BB106_2102; $L__BB106_2091: ld.local.v2.f32 {%f12671, %f12672}, [%rd7+-48]; ld.local.f32 %f3401, [%rd587+28]; ld.local.f32 %f16555, [%rd587+8]; ld.local.f32 %f3403, [%rd7+-52]; ld.local.f32 %f16556, [%rd587+12]; mul.ftz.f32 %f12673, %f12671, %f16556; fma.rn.ftz.f32 %f12674, %f3403, %f16555, %f12673; ld.local.f32 %f16557, [%rd587+16]; fma.rn.ftz.f32 %f16558, %f12672, %f16557, %f12674; setp.gt.ftz.f32 %p1835, %f16558, 0f00000000; @%p1835 bra $L__BB106_2093; bra.uni $L__BB106_2092; $L__BB106_2093: rcp.approx.ftz.f32 %f16559, %f3401; mov.u16 %rs157, 0; bra.uni $L__BB106_2094; $L__BB106_1890: ld.local.u64 %rd588, [%rd587+52]; ld.v4.f32 {%f11518, %f11519, %f11520, %f11521}, [%rd588]; ld.local.v2.f32 {%f16462, %f16463}, [%rd7+32]; ld.local.f32 %f16461, [%rd7+28]; mul.ftz.f32 %f11524, %f16462, %f16462; fma.rn.ftz.f32 %f11525, %f16461, %f16461, %f11524; fma.rn.ftz.f32 %f11526, %f16463, %f16463, %f11525; sqrt.approx.ftz.f32 %f2848, %f11526; setp.neu.ftz.f32 %p1651, %f2848, 0f00000000; @%p1651 bra $L__BB106_1891; bra.uni $L__BB106_1892; $L__BB106_1891: rcp.approx.ftz.f32 %f11527, %f2848; mul.ftz.f32 %f16463, %f16463, %f11527; mul.ftz.f32 %f16462, %f16462, %f11527; mul.ftz.f32 %f16461, %f16461, %f11527; $L__BB106_1892: mul.ftz.f32 %f11528, %f11520, %f16462; mul.ftz.f32 %f11529, %f11519, %f16463; sub.ftz.f32 %f16464, %f11528, %f11529; mul.ftz.f32 %f11530, %f11518, %f16463; mul.ftz.f32 %f11531, %f11520, %f16461; sub.ftz.f32 %f16465, %f11530, %f11531; mul.ftz.f32 %f11532, %f11519, %f16461; mul.ftz.f32 %f11533, %f11518, %f16462; sub.ftz.f32 %f16466, %f11532, %f11533; mul.ftz.f32 %f11534, %f16465, %f16465; fma.rn.ftz.f32 %f11535, %f16464, %f16464, %f11534; fma.rn.ftz.f32 %f11536, %f16466, %f16466, %f11535; sqrt.approx.ftz.f32 %f2864, %f11536; setp.neu.ftz.f32 %p1652, %f2864, 0f00000000; @%p1652 bra $L__BB106_1893; bra.uni $L__BB106_1894; $L__BB106_1893: rcp.approx.ftz.f32 %f11537, %f2864; mul.ftz.f32 %f16466, %f11537, %f16466; mul.ftz.f32 %f16465, %f11537, %f16465; mul.ftz.f32 %f16464, %f11537, %f16464; $L__BB106_1894: ld.local.f32 %f11538, [%rd7+-52]; ld.local.v2.f32 {%f11539, %f11540}, [%rd7+-48]; mul.ftz.f32 %f11543, %f16462, %f11539; fma.rn.ftz.f32 %f11544, %f16461, %f11538, %f11543; fma.rn.ftz.f32 %f2874, %f16463, %f11540, %f11544; mul.ftz.f32 %f11545, %f11519, %f11539; fma.rn.ftz.f32 %f11546, %f11518, %f11538, %f11545; fma.rn.ftz.f32 %f11547, %f11520, %f11540, %f11546; mul.ftz.f32 %f11548, %f16465, %f11539; fma.rn.ftz.f32 %f11549, %f16464, %f11538, %f11548; fma.rn.ftz.f32 %f11550, %f16466, %f11540, %f11549; ld.local.u32 %r2717, [%rd7+168]; mad.lo.s32 %r876, %r2717, 1103515245, 12345; mad.lo.s32 %r877, %r876, 1103515245, 12345; st.local.u32 [%rd7+168], %r877; mul.ftz.f32 %f2878, %f2874, %f2874; mov.f32 %f11551, 0f3F800000; sub.ftz.f32 %f11552, %f11551, %f2878; mov.f32 %f16472, 0f00000000; max.ftz.f32 %f11554, %f11552, %f16472; sqrt.approx.ftz.f32 %f2879, %f11554; abs.ftz.f32 %f2880, %f11547; abs.ftz.f32 %f2881, %f11550; setp.eq.ftz.f32 %p1653, %f2880, 0f00000000; setp.eq.ftz.f32 %p1654, %f2881, 0f00000000; and.pred %p1655, %p1653, %p1654; mov.b32 %r878, %f11547; mov.b32 %r2718, %f11550; and.b32 %r879, %r2718, -2147483648; @%p1655 bra $L__BB106_1898; bra.uni $L__BB106_1895; $L__BB106_1898: shr.s32 %r2723, %r878, 31; and.b32 %r2724, %r2723, 1078530011; or.b32 %r2725, %r879, %r2724; mov.b32 %f16467, %r2725; bra.uni $L__BB106_1899; $L__BB106_1874: ld.local.v2.f32 {%f11310, %f11311}, [%rd7+-48]; ld.local.f32 %f2812, [%rd587+36]; ld.local.f32 %f2813, [%rd587+20]; ld.local.f32 %f11314, [%rd7+-52]; ld.local.f32 %f2814, [%rd587+24]; mul.ftz.f32 %f11315, %f11310, %f2814; fma.rn.ftz.f32 %f11316, %f11314, %f2813, %f11315; ld.local.f32 %f2815, [%rd587+28]; fma.rn.ftz.f32 %f11317, %f11311, %f2815, %f11316; mul.ftz.f32 %f11318, %f2813, %f11317; mul.ftz.f32 %f11319, %f2814, %f11317; mul.ftz.f32 %f11320, %f2815, %f11317; sub.ftz.f32 %f11321, %f11314, %f11318; sub.ftz.f32 %f11322, %f11310, %f11319; sub.ftz.f32 %f11323, %f11311, %f11320; mul.ftz.f32 %f11324, %f11322, %f11322; fma.rn.ftz.f32 %f11325, %f11321, %f11321, %f11324; fma.rn.ftz.f32 %f11326, %f11323, %f11323, %f11325; sqrt.approx.ftz.f32 %f11327, %f11326; div.approx.ftz.f32 %f2816, %f11321, %f11327; div.approx.ftz.f32 %f2817, %f11322, %f11327; div.approx.ftz.f32 %f2818, %f11323, %f11327; abs.ftz.f32 %f11328, %f11317; setp.lt.ftz.f32 %p1630, %f11328, 0f3F800000; mov.f32 %f11329, 0f3F800000; add.ftz.f32 %f11330, %f11328, 0fBF800000; add.ftz.f32 %f11331, %f11330, 0f3F800000; selp.f32 %f11332, %f11331, 0f3F800000, %p1630; sub.ftz.f32 %f11333, %f11329, %f11332; sqrt.approx.ftz.f32 %f11334, %f11333; fma.rn.ftz.f32 %f11335, %f11332, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f11336, %f11332, %f11335, 0fBE5A6B93; fma.rn.ftz.f32 %f11337, %f11332, %f11336, 0f3FC90FDB; mov.f32 %f11338, 0f3FC90FDB; mul.ftz.f32 %f11339, %f11334, %f11337; setp.lt.ftz.f32 %p1631, %f11317, 0f00000000; mov.f32 %f16460, 0f00000000; mov.f32 %f11340, 0f40490FDB; sub.ftz.f32 %f11341, %f11340, %f11339; selp.f32 %f11342, %f11341, %f11339, %p1631; sub.ftz.f32 %f2819, %f11338, %f11342; ld.local.f32 %f2820, [%rd587+32]; rcp.approx.ftz.f32 %f2821, %f2820; add.ftz.f32 %f11343, %f2819, 0f3FC90FDB; mov.f32 %f11344, 0f40000000; div.approx.ftz.f32 %f11345, %f11343, %f11344; ld.local.f32 %f2822, [%rd587+40]; sub.ftz.f32 %f11346, %f11345, %f2822; mul.ftz.f32 %f2823, %f2821, %f11346; abs.ftz.f32 %f2824, %f2823; setp.eq.ftz.f32 %p1632, %f2824, 0f00000000; mov.f32 %f16458, %f16460; @%p1632 bra $L__BB106_1879; setp.eq.ftz.f32 %p1633, %f2824, 0f3F800000; mov.f32 %f16458, %f11329; @%p1633 bra $L__BB106_1879; setp.gt.ftz.f32 %p1634, %f2824, 0f3F800000; @%p1634 bra $L__BB106_1878; bra.uni $L__BB106_1877; $L__BB106_1878: mov.f32 %f11349, 0f3F800000; div.approx.ftz.f32 %f16458, %f11349, %f2824; bra.uni $L__BB106_1879; $L__BB106_2292: ld.local.f32 %f14794, [%rd587+8]; ld.local.f32 %f14795, [%rd587+12]; neg.ftz.f32 %f14796, %f14795; ld.local.f32 %f14797, [%rd587+16]; neg.ftz.f32 %f14798, %f14797; mul.ftz.f32 %f14799, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f14800, %f14799; sqrt.approx.ftz.f32 %f14801, %f2775; mul.ftz.f32 %f14802, %f14801, %f14800; sin.approx.ftz.f32 %f14803, %f14799; mul.ftz.f32 %f14804, %f14801, %f14803; mul.ftz.f32 %f14805, %f14802, %f14802; mov.f32 %f14806, 0f3F800000; sub.ftz.f32 %f14807, %f14806, %f14805; mul.ftz.f32 %f14808, %f14804, %f14804; sub.ftz.f32 %f14809, %f14807, %f14808; mov.f32 %f16738, 0f00000000; max.ftz.f32 %f14810, %f14809, %f16738; sqrt.approx.ftz.f32 %f4042, %f14810; setp.neu.ftz.f32 %p2024, %f14794, %f14795; setp.neu.ftz.f32 %p2025, %f14794, %f14797; or.pred %p2026, %p2024, %p2025; sub.ftz.f32 %f14811, %f14795, %f14797; selp.f32 %f14812, %f14797, %f14798, %p2026; selp.f32 %f14813, %f14796, %f14795, %p2026; add.ftz.f32 %f14814, %f14794, %f14813; sub.ftz.f32 %f14815, %f14812, %f14794; mul.ftz.f32 %f14816, %f14815, %f14815; fma.rn.ftz.f32 %f14817, %f14811, %f14811, %f14816; fma.rn.ftz.f32 %f14818, %f14814, %f14814, %f14817; sqrt.approx.ftz.f32 %f14819, %f14818; div.approx.ftz.f32 %f14820, %f14811, %f14819; div.approx.ftz.f32 %f14821, %f14815, %f14819; div.approx.ftz.f32 %f14822, %f14814, %f14819; mul.ftz.f32 %f14823, %f14822, %f14795; mul.ftz.f32 %f14824, %f14821, %f14797; sub.ftz.f32 %f14825, %f14824, %f14823; mul.ftz.f32 %f14826, %f14820, %f14797; mul.ftz.f32 %f14827, %f14822, %f14794; sub.ftz.f32 %f14828, %f14827, %f14826; mul.ftz.f32 %f14829, %f14821, %f14794; mul.ftz.f32 %f14830, %f14820, %f14795; sub.ftz.f32 %f14831, %f14830, %f14829; mul.ftz.f32 %f14832, %f14804, %f14825; mul.ftz.f32 %f14833, %f14804, %f14828; mul.ftz.f32 %f14834, %f14804, %f14831; fma.rn.ftz.f32 %f14835, %f14802, %f14820, %f14832; fma.rn.ftz.f32 %f14836, %f14802, %f14821, %f14833; fma.rn.ftz.f32 %f14837, %f14802, %f14822, %f14834; mul.ftz.f32 %f14838, %f4042, %f14794; mul.ftz.f32 %f14839, %f4042, %f14795; mul.ftz.f32 %f14840, %f4042, %f14797; sub.ftz.f32 %f16735, %f14835, %f14838; sub.ftz.f32 %f16736, %f14836, %f14839; sub.ftz.f32 %f16737, %f14837, %f14840; mul.ftz.f32 %f14841, %f2807, %f16736; fma.rn.ftz.f32 %f14842, %f2806, %f16735, %f14841; fma.rn.ftz.f32 %f14843, %f2808, %f16737, %f14842; setp.geu.ftz.f32 %p2027, %f14843, 0f00000000; mov.u32 %r3285, 5; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2027 bra $L__BB106_2309; mul.ftz.f32 %f16738, %f4042, 0f3EA2F983; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1852: setp.eq.s32 %p1956, %r873, 22; selp.b32 %r3285, 9, 10, %p1956; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_2249: ld.local.f32 %f14371, [%rd587+52]; mul.ftz.f32 %f14372, %f3900, %f14371; ld.local.f32 %f14373, [%rd587+48]; mul.ftz.f32 %f14374, %f3901, %f14373; sub.ftz.f32 %f14375, %f14372, %f14374; ld.local.f32 %f14376, [%rd587+44]; mul.ftz.f32 %f14377, %f3901, %f14376; mul.ftz.f32 %f14378, %f3898, %f14371; sub.ftz.f32 %f14379, %f14377, %f14378; mul.ftz.f32 %f14380, %f3898, %f14373; mul.ftz.f32 %f14381, %f3900, %f14376; sub.ftz.f32 %f14382, %f14380, %f14381; mul.ftz.f32 %f14383, %f14379, %f14379; fma.rn.ftz.f32 %f14384, %f14375, %f14375, %f14383; fma.rn.ftz.f32 %f14385, %f14382, %f14382, %f14384; sqrt.approx.ftz.f32 %f14386, %f14385; div.approx.ftz.f32 %f16688, %f14375, %f14386; div.approx.ftz.f32 %f16689, %f14379, %f14386; div.approx.ftz.f32 %f16690, %f14382, %f14386; mul.ftz.f32 %f14387, %f3901, %f16689; mul.ftz.f32 %f14388, %f3900, %f16690; sub.ftz.f32 %f16691, %f14387, %f14388; mul.ftz.f32 %f14389, %f3898, %f16690; mul.ftz.f32 %f14390, %f3901, %f16688; sub.ftz.f32 %f16692, %f14389, %f14390; mul.ftz.f32 %f14391, %f3900, %f16688; mul.ftz.f32 %f14392, %f3898, %f16689; sub.ftz.f32 %f16693, %f14391, %f14392; $L__BB106_2251: mul.ftz.f32 %f14414, %f14360, %f16692; fma.rn.ftz.f32 %f14415, %f3899, %f16691, %f14414; fma.rn.ftz.f32 %f14416, %f14361, %f16693, %f14415; mul.ftz.f32 %f14417, %f14360, %f16689; fma.rn.ftz.f32 %f14418, %f3899, %f16688, %f14417; fma.rn.ftz.f32 %f14419, %f14361, %f16690, %f14418; mul.ftz.f32 %f14420, %f14362, %f14416; mul.ftz.f32 %f14421, %f14363, %f14419; mul.ftz.f32 %f14422, %f14420, %f14420; fma.rn.ftz.f32 %f14423, %f14421, %f14421, %f14422; mul.ftz.f32 %f3921, %f3902, %f3902; add.ftz.f32 %f14424, %f3921, %f14423; sqrt.approx.ftz.f32 %f14425, %f14424; div.approx.ftz.f32 %f3922, %f14420, %f14425; div.approx.ftz.f32 %f3923, %f14421, %f14425; div.approx.ftz.f32 %f3924, %f3902, %f14425; setp.geu.ftz.f32 %p1995, %f3924, 0f3F7FFF58; mov.f32 %f16695, 0f00000000; mov.f32 %f16694, 0f3F800000; mov.f32 %f16696, %f16694; mov.f32 %f16697, %f16695; @%p1995 bra $L__BB106_2253; mul.ftz.f32 %f14426, %f3924, %f3924; mov.f32 %f14427, 0f3F800000; sub.ftz.f32 %f14428, %f14427, %f14426; mov.f32 %f14429, 0f00000000; max.ftz.f32 %f14430, %f14428, %f14429; sqrt.approx.ftz.f32 %f16695, %f14430; rsqrt.approx.ftz.f32 %f14431, %f14430; mul.ftz.f32 %f16696, %f3922, %f14431; mul.ftz.f32 %f16697, %f3923, %f14431; mov.f32 %f16694, %f3924; $L__BB106_2253: setp.ltu.ftz.f32 %p1996, %f16694, 0f3F7FFF58; @%p1996 bra $L__BB106_2255; bra.uni $L__BB106_2254; $L__BB106_2255: div.approx.ftz.f32 %f14439, %f16695, %f16694; mul.ftz.f32 %f14440, %f14439, %f14439; add.ftz.f32 %f14441, %f14440, 0f3F800000; mov.f32 %f14442, 0f00000000; max.ftz.f32 %f14443, %f14441, %f14442; sqrt.approx.ftz.f32 %f14444, %f14443; add.ftz.f32 %f14445, %f14444, 0f3F800000; mul.ftz.f32 %f14446, %f14445, 0f3F000000; mov.f32 %f14447, 0f3F000000; rcp.approx.ftz.f32 %f3937, %f14446; add.ftz.f32 %f14448, %f2799, %f2799; fma.rn.ftz.f32 %f14449, %f14448, %f14446, 0fBF800000; mul.ftz.f32 %f14450, %f14449, %f14449; add.ftz.f32 %f14451, %f14450, 0fBF800000; rcp.approx.ftz.f32 %f14452, %f14451; mul.ftz.f32 %f14453, %f14452, %f14452; mul.ftz.f32 %f14454, %f14440, %f14453; sub.ftz.f32 %f14455, %f14450, %f14440; mul.ftz.f32 %f14456, %f14455, %f14452; sub.ftz.f32 %f14457, %f14454, %f14456; max.ftz.f32 %f14458, %f14457, %f14442; sqrt.approx.ftz.f32 %f14459, %f14458; mul.ftz.f32 %f14460, %f14439, %f14452; sub.ftz.f32 %f14461, %f14460, %f14459; add.ftz.f32 %f14462, %f14459, %f14460; setp.lt.ftz.f32 %p1997, %f14449, 0f00000000; mul.ftz.f32 %f14463, %f14439, %f14462; setp.gt.ftz.f32 %p1998, %f14463, 0f3F800000; or.pred %p1999, %p1997, %p1998; selp.f32 %f16700, %f14461, %f14462, %p1999; sub.ftz.f32 %f14464, %f14447, %f2775; add.ftz.f32 %f14465, %f2775, 0fBF000000; setp.gt.ftz.f32 %p2000, %f2775, 0f3F000000; selp.f32 %f14466, %f14465, %f14464, %p2000; add.ftz.f32 %f14467, %f14466, %f14466; selp.f32 %f14468, 0f3F800000, 0fBF800000, %p2000; fma.rn.ftz.f32 %f14469, %f14467, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f14470, %f14467, %f14469, 0f3EED4413; mul.ftz.f32 %f14471, %f14467, %f14470; fma.rn.ftz.f32 %f14472, %f14467, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f14473, %f14467, %f14472, 0fBF800000; fma.rn.ftz.f32 %f14474, %f14467, %f14473, 0f3F191676; div.approx.ftz.f32 %f14475, %f14471, %f14474; mul.ftz.f32 %f14476, %f14468, %f14475; fma.rn.ftz.f32 %f14477, %f16700, %f16700, 0f3F800000; max.ftz.f32 %f14478, %f14477, %f14442; sqrt.approx.ftz.f32 %f14479, %f14478; mul.ftz.f32 %f16699, %f14476, %f14479; bra.uni $L__BB106_2256; $L__BB106_2254: mov.f32 %f3937, 0f3F800000; sub.ftz.f32 %f14433, %f3937, %f2799; div.approx.ftz.f32 %f14434, %f2799, %f14433; sqrt.approx.ftz.f32 %f14435, %f14434; mul.ftz.f32 %f14436, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f14437, %f14436; mul.ftz.f32 %f16700, %f14435, %f14437; sin.approx.ftz.f32 %f14438, %f14436; mul.ftz.f32 %f16699, %f14435, %f14438; $L__BB106_2256: mul.ftz.f32 %f14480, %f16697, %f16699; mul.ftz.f32 %f14481, %f16696, %f16700; sub.ftz.f32 %f14482, %f14481, %f14480; mul.ftz.f32 %f14483, %f16697, %f16700; fma.rn.ftz.f32 %f14484, %f16696, %f16699, %f14483; mul.ftz.f32 %f14485, %f14362, %f14482; mul.ftz.f32 %f14486, %f14363, %f14484; neg.ftz.f32 %f14487, %f14485; neg.ftz.f32 %f14488, %f14486; mul.ftz.f32 %f14489, %f14486, %f14486; fma.rn.ftz.f32 %f14490, %f14485, %f14485, %f14489; add.ftz.f32 %f14491, %f14490, 0f3F800000; mov.f32 %f14492, 0f3F800000; sqrt.approx.ftz.f32 %f14493, %f14491; div.approx.ftz.f32 %f14494, %f14487, %f14493; div.approx.ftz.f32 %f14495, %f14488, %f14493; div.approx.ftz.f32 %f3940, %f14492, %f14493; mul.ftz.f32 %f14496, %f16688, %f14495; mul.ftz.f32 %f14497, %f16689, %f14495; mul.ftz.f32 %f14498, %f16690, %f14495; fma.rn.ftz.f32 %f14499, %f16691, %f14494, %f14496; fma.rn.ftz.f32 %f14500, %f16692, %f14494, %f14497; fma.rn.ftz.f32 %f14501, %f16693, %f14494, %f14498; fma.rn.ftz.f32 %f3941, %f3898, %f3940, %f14499; fma.rn.ftz.f32 %f3942, %f3900, %f3940, %f14500; fma.rn.ftz.f32 %f3943, %f3901, %f3940, %f14501; mul.ftz.f32 %f14502, %f14360, %f3942; fma.rn.ftz.f32 %f14503, %f3899, %f3941, %f14502; fma.rn.ftz.f32 %f3944, %f14361, %f3943, %f14503; setp.eq.s32 %p2001, %r873, 23; @%p2001 bra $L__BB106_2283; bra.uni $L__BB106_2257; $L__BB106_2283: setp.gt.ftz.f32 %p2015, %f3944, 0f00000000; ld.local.f32 %f4009, [%rd587+28]; @%p2015 bra $L__BB106_2285; bra.uni $L__BB106_2284; $L__BB106_2285: rcp.approx.ftz.f32 %f16723, %f4009; mov.u16 %rs159, 0; mov.f32 %f16719, %f3941; mov.f32 %f16720, %f3942; mov.f32 %f16721, %f3943; mov.f32 %f16722, %f3944; bra.uni $L__BB106_2286; $L__BB106_2257: setp.leu.ftz.f32 %p2002, %f3944, 0f00000000; mov.u32 %r3285, 10; mov.f32 %f16738, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2002 bra $L__BB106_2309; add.ftz.f32 %f14513, %f3944, %f3944; mul.ftz.f32 %f14514, %f3941, %f14513; mul.ftz.f32 %f14515, %f3942, %f14513; mul.ftz.f32 %f14516, %f3943, %f14513; sub.ftz.f32 %f16735, %f14514, %f3899; sub.ftz.f32 %f16736, %f14515, %f14360; sub.ftz.f32 %f16737, %f14516, %f14361; mul.ftz.f32 %f14517, %f2807, %f16736; fma.rn.ftz.f32 %f14518, %f2806, %f16735, %f14517; fma.rn.ftz.f32 %f14519, %f2808, %f16737, %f14518; setp.leu.ftz.f32 %p2003, %f14519, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2003 bra $L__BB106_2309; mul.ftz.f32 %f16705, %f14362, %f14363; setp.gtu.ftz.f32 %p2004, %f16705, 0f33D6BF95; @%p2004 bra $L__BB106_2266; bra.uni $L__BB106_2260; $L__BB106_2266: mul.ftz.f32 %f14559, %f3900, %f16736; fma.rn.ftz.f32 %f14560, %f3898, %f16735, %f14559; fma.rn.ftz.f32 %f3965, %f3901, %f16737, %f14560; @%p1991 bra $L__BB106_2268; bra.uni $L__BB106_2267; $L__BB106_2268: setp.eq.s32 %p2009, %r873, 12; @%p2009 bra $L__BB106_2270; bra.uni $L__BB106_2269; $L__BB106_2270: setp.ge.ftz.f32 %p2010, %f14362, 0f3F800000; mov.f32 %f16706, 0f3EA2F983; @%p2010 bra $L__BB106_2272; mul.ftz.f32 %f14611, %f14362, %f14362; add.ftz.f32 %f14612, %f14611, 0fBF800000; mul.ftz.f32 %f14613, %f3940, %f14612; fma.rn.ftz.f32 %f14614, %f3940, %f14613, 0f3F800000; lg2.approx.ftz.f32 %f14615, %f14611; mul.ftz.f32 %f14616, %f14615, 0f3F317218; mul.ftz.f32 %f14617, %f14616, 0f40490FDB; mul.ftz.f32 %f14618, %f14614, %f14617; div.approx.ftz.f32 %f16706, %f14612, %f14618; $L__BB106_2272: mov.f32 %f14620, 0f3F800000; sub.ftz.f32 %f14621, %f14620, %f3921; mul.ftz.f32 %f14622, %f14621, 0f3D800000; mov.f32 %f16705, 0f3D800000; div.approx.ftz.f32 %f14623, %f14622, %f3921; add.ftz.f32 %f14624, %f14623, 0f3F800000; mov.f32 %f14625, 0f00000000; max.ftz.f32 %f14626, %f14624, %f14625; sqrt.approx.ftz.f32 %f14627, %f14626; add.ftz.f32 %f14628, %f14627, 0f3F800000; mov.f32 %f14629, 0f40000000; div.approx.ftz.f32 %f3937, %f14629, %f14628; bra.uni $L__BB106_2273; $L__BB106_2074: st.local.u32 [%rd7+168], %r885; and.b16 %rs116, %rs155, 255; setp.eq.s16 %p1821, %rs116, 0; @%p1821 bra $L__BB106_2075; bra.uni $L__BB106_2076; $L__BB106_2075: neg.ftz.f32 %f3232, %f3232; neg.ftz.f32 %f3233, %f3233; neg.ftz.f32 %f3234, %f3234; $L__BB106_2076: ld.local.f32 %f3348, [%rd587+20]; mul.ftz.f32 %f3349, %f16549, %f3232; setp.gt.ftz.f32 %p1822, %f3349, 0f00000000; mov.f32 %f16543, 0f00000000; mul.ftz.f32 %f12485, %f3234, %f3212; selp.f32 %f12486, %f3234, %f12485, %p1822; mul.ftz.f32 %f12487, %f3233, %f3212; selp.f32 %f12488, %f3233, %f12487, %p1822; mul.ftz.f32 %f12489, %f3232, %f3212; selp.f32 %f12490, %f3232, %f12489, %p1822; add.ftz.f32 %f12491, %f3207, %f12486; add.ftz.f32 %f12492, %f3208, %f12488; add.ftz.f32 %f12493, %f16549, %f12490; mul.ftz.f32 %f12494, %f12492, %f12492; fma.rn.ftz.f32 %f12495, %f12491, %f12491, %f12494; fma.rn.ftz.f32 %f12496, %f12493, %f12493, %f12495; sqrt.approx.ftz.f32 %f3350, %f12496; rsqrt.approx.ftz.f32 %f12497, %f12496; mul.ftz.f32 %f12498, %f12497, %f12491; mul.ftz.f32 %f12499, %f12497, %f12492; mul.ftz.f32 %f12500, %f12497, %f12493; setp.lt.ftz.f32 %p1823, %f12500, 0f00000000; neg.ftz.f32 %f12501, %f12498; neg.ftz.f32 %f12502, %f12499; neg.ftz.f32 %f12503, %f12500; selp.f32 %f3351, %f12501, %f12498, %p1823; selp.f32 %f3352, %f12502, %f12499, %p1823; selp.f32 %f3353, %f12503, %f12500, %p1823; setp.lt.ftz.f32 %p1824, %f16549, 0f00000000; selp.f32 %f3354, %f3216, %f3207, %p1824; selp.f32 %f3355, %f3217, %f3208, %p1824; selp.f32 %f3356, %f3218, %f16549, %p1824; setp.gt.ftz.f32 %p1825, %f3356, 0f3F7FF972; @%p1825 bra $L__BB106_2079; setp.lt.ftz.f32 %p1826, %f3356, 0fBF7FF972; mov.f32 %f16543, 0fBF7FF972; @%p1826 bra $L__BB106_2079; mul.ftz.f32 %f12505, %f3356, %f3356; mov.f32 %f12506, 0f33D6BF95; max.ftz.f32 %f12507, %f12505, %f12506; rcp.approx.ftz.f32 %f12508, %f12507; mul.ftz.f32 %f12509, %f3354, %f3348; mul.ftz.f32 %f12510, %f3355, %f3348; mul.ftz.f32 %f12511, %f12510, %f12510; fma.rn.ftz.f32 %f12512, %f12509, %f12509, %f12511; fma.rn.ftz.f32 %f12513, %f12512, %f12508, 0f3F800000; sqrt.approx.ftz.f32 %f12514, %f12513; neg.ftz.f32 %f12515, %f12514; setp.le.ftz.f32 %p1827, %f3356, 0f00000000; selp.f32 %f12516, %f12515, %f12514, %p1827; add.ftz.f32 %f12517, %f12516, 0fBF800000; mul.ftz.f32 %f16543, %f12517, 0f3F000000; $L__BB106_2079: mul.ftz.f32 %f12519, %f3353, %f3353; mov.f32 %f16544, 0f3F800000; sub.ftz.f32 %f12520, %f16544, %f12519; mul.ftz.f32 %f12521, %f3348, %f3348; fma.rn.ftz.f32 %f12522, %f12521, %f12519, %f12520; mul.ftz.f32 %f12523, %f12522, 0f40490FDB; mul.ftz.f32 %f12524, %f12522, %f12523; mov.f32 %f12525, 0f33D6BF95; max.ftz.f32 %f12526, %f12524, %f12525; div.approx.ftz.f32 %f3359, %f12521, %f12526; mul.ftz.f32 %f12527, %f3355, %f3352; fma.rn.ftz.f32 %f12528, %f3354, %f3351, %f12527; fma.rn.ftz.f32 %f3360, %f3356, %f3353, %f12528; abs.ftz.f32 %f3361, %f3360; fma.rn.ftz.f32 %f3362, %f3361, %f3361, %f3221; setp.leu.ftz.f32 %p1828, %f3362, 0f00000000; @%p1828 bra $L__BB106_2081; sqrt.approx.ftz.f32 %f12529, %f3362; sub.ftz.f32 %f12530, %f12529, %f3361; add.ftz.f32 %f12531, %f3361, %f12529; div.approx.ftz.f32 %f12532, %f12530, %f12531; fma.rn.ftz.f32 %f12533, %f3361, %f12531, 0fBF800000; fma.rn.ftz.f32 %f12534, %f3361, %f12530, 0f3F800000; div.approx.ftz.f32 %f12535, %f12533, %f12534; mul.ftz.f32 %f12536, %f12532, 0f3F000000; mul.ftz.f32 %f12537, %f12532, %f12536; fma.rn.ftz.f32 %f12538, %f12535, %f12535, 0f3F800000; mul.ftz.f32 %f16544, %f12537, %f12538; $L__BB106_2081: mul.ftz.f32 %f12539, %f3232, 0f3EA2F983; abs.ftz.f32 %f3365, %f12539; add.ftz.f32 %f12540, %f16543, 0f3F800000; mul.ftz.f32 %f3366, %f3356, %f12540; @%p1822 bra $L__BB106_2083; bra.uni $L__BB106_2082; $L__BB106_2083: mul.ftz.f32 %f12600, %f3359, 0f3E800000; max.ftz.f32 %f12602, %f3366, %f12525; div.approx.ftz.f32 %f12603, %f12600, %f12602; mul.ftz.f32 %f12604, %f3348, 0fBFFE59F3; mul.ftz.f32 %f12605, %f3348, %f12604; mul.ftz.f32 %f12606, %f12605, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12607, %f12606; fma.rn.ftz.f32 %f12608, %f12607, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f12609, %f3348, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f12610, %f3348, %f12609, 0f417C8B44; fma.rn.ftz.f32 %f12611, %f3348, %f12610, 0fC1665E35; fma.rn.ftz.f32 %f12612, %f3348, %f12611, 0f40CE7881; fma.rn.ftz.f32 %f12613, %f3348, %f12612, 0fBF993E81; fma.rn.ftz.f32 %f12614, %f3348, %f12613, 0f3CE3C31E; fma.rn.ftz.f32 %f12615, %f3348, %f12614, 0f3BBA5D49; sub.ftz.f32 %f12616, %f12608, %f12615; cvt.ftz.sat.f32.f32 %f12617, %f12616; mov.f32 %f12618, 0f3F800000; sub.ftz.f32 %f12619, %f12618, %f12617; mul.ftz.f32 %f12620, %f3365, %f12619; fma.rn.ftz.f32 %f12621, %f12603, %f12617, %f12620; mul.ftz.f32 %f16738, %f16544, %f12621; bra.uni $L__BB106_2084; $L__BB106_2145: setp.neu.ftz.f32 %p1908, %f13465, %f13466; @%p1908 bra $L__BB106_2147; bra.uni $L__BB106_2146; $L__BB106_2147: ld.local.f32 %f13485, [%rd587+52]; mul.ftz.f32 %f13486, %f3575, %f13485; ld.local.f32 %f13487, [%rd587+48]; mul.ftz.f32 %f13488, %f3576, %f13487; sub.ftz.f32 %f13489, %f13486, %f13488; ld.local.f32 %f13490, [%rd587+44]; mul.ftz.f32 %f13491, %f3576, %f13490; mul.ftz.f32 %f13492, %f3573, %f13485; sub.ftz.f32 %f13493, %f13491, %f13492; mul.ftz.f32 %f13494, %f3573, %f13487; mul.ftz.f32 %f13495, %f3575, %f13490; sub.ftz.f32 %f13496, %f13494, %f13495; mul.ftz.f32 %f13497, %f13493, %f13493; fma.rn.ftz.f32 %f13498, %f13489, %f13489, %f13497; fma.rn.ftz.f32 %f13499, %f13496, %f13496, %f13498; sqrt.approx.ftz.f32 %f13500, %f13499; div.approx.ftz.f32 %f16584, %f13489, %f13500; div.approx.ftz.f32 %f16585, %f13493, %f13500; div.approx.ftz.f32 %f16586, %f13496, %f13500; mul.ftz.f32 %f13501, %f3576, %f16585; mul.ftz.f32 %f13502, %f3575, %f16586; sub.ftz.f32 %f16587, %f13501, %f13502; mul.ftz.f32 %f13503, %f3573, %f16586; mul.ftz.f32 %f13504, %f3576, %f16584; sub.ftz.f32 %f16588, %f13503, %f13504; mul.ftz.f32 %f13505, %f3575, %f16584; mul.ftz.f32 %f13506, %f3573, %f16585; sub.ftz.f32 %f16589, %f13505, %f13506; bra.uni $L__BB106_2148; $L__BB106_2086: neg.ftz.f32 %f16549, %f16549; neg.ftz.f32 %f16546, %f16546; neg.ftz.f32 %f16547, %f16547; neg.ftz.f32 %f16548, %f16548; mov.f32 %f16550, %f3376; $L__BB106_2088: mul.ftz.f32 %f12632, %f16549, %f16549; mov.f32 %f16554, 0f3F800000; sub.ftz.f32 %f12633, %f16554, %f12632; mul.ftz.f32 %f12634, %f16550, %f16550; mul.ftz.f32 %f12635, %f12634, %f12633; sub.ftz.f32 %f3387, %f16554, %f12635; setp.lt.ftz.f32 %p1833, %f3387, 0f00000000; mov.f32 %f16551, 0f00000000; mov.f32 %f16552, %f16551; mov.f32 %f16553, %f16551; @%p1833 bra $L__BB106_2090; sqrt.approx.ftz.f32 %f12636, %f3387; mov.f32 %f12637, 0f33D6BF95; max.ftz.f32 %f12638, %f12636, %f12637; mul.ftz.f32 %f12639, %f16549, %f16550; sub.ftz.f32 %f12640, %f12639, %f12638; mul.ftz.f32 %f12641, %f16546, %f12640; mul.ftz.f32 %f12642, %f16547, %f12640; mul.ftz.f32 %f12643, %f16548, %f12640; mul.ftz.f32 %f12644, %f3197, %f16550; sub.ftz.f32 %f16551, %f12641, %f12644; mul.ftz.f32 %f12645, %f12249, %f16550; sub.ftz.f32 %f16552, %f12642, %f12645; mul.ftz.f32 %f12646, %f12250, %f16550; sub.ftz.f32 %f16553, %f12643, %f12646; mul.ftz.f32 %f12647, %f16547, %f16552; fma.rn.ftz.f32 %f12648, %f16546, %f16551, %f12647; fma.rn.ftz.f32 %f12649, %f16548, %f16553, %f12648; mul.ftz.f32 %f12650, %f3376, %f12649; add.ftz.f32 %f12651, %f16549, %f12650; sub.ftz.f32 %f12652, %f16549, %f12650; div.approx.ftz.f32 %f12653, %f12651, %f12652; mul.ftz.f32 %f12654, %f3376, %f16549; add.ftz.f32 %f12655, %f12654, %f12649; sub.ftz.f32 %f12656, %f12654, %f12649; div.approx.ftz.f32 %f12657, %f12655, %f12656; mul.ftz.f32 %f12658, %f12657, %f12657; fma.rn.ftz.f32 %f12659, %f12653, %f12653, %f12658; mul.ftz.f32 %f16554, %f12659, 0f3F000000; $L__BB106_2090: add.ftz.f32 %f12664, %f16549, %f16549; setp.gt.ftz.f32 %p1834, %f16554, %f2799; mul.ftz.f32 %f12665, %f16546, %f12664; sub.ftz.f32 %f12666, %f12665, %f3197; selp.f32 %f16735, %f12666, %f16551, %p1834; mul.ftz.f32 %f12667, %f16547, %f12664; sub.ftz.f32 %f12668, %f12667, %f12249; selp.f32 %f16736, %f12668, %f16552, %p1834; mul.ftz.f32 %f12669, %f16548, %f12664; sub.ftz.f32 %f12670, %f12669, %f12250; selp.f32 %f16737, %f12670, %f16553, %p1834; selp.b32 %r3285, 18, 17, %p1834; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_2092: neg.ftz.f32 %f16558, %f16558; neg.ftz.f32 %f16555, %f16555; neg.ftz.f32 %f16556, %f16556; neg.ftz.f32 %f16557, %f16557; mov.u16 %rs157, 1; mov.f32 %f16559, %f3401; $L__BB106_2094: mul.ftz.f32 %f12679, %f16558, %f16558; mov.f32 %f16563, 0f3F800000; sub.ftz.f32 %f12680, %f16563, %f12679; mul.ftz.f32 %f12681, %f16559, %f16559; mul.ftz.f32 %f12682, %f12681, %f12680; sub.ftz.f32 %f3417, %f16563, %f12682; setp.lt.ftz.f32 %p1836, %f3417, 0f00000000; mov.f32 %f16735, 0f00000000; mov.f32 %f16736, %f16735; mov.f32 %f16737, %f16735; @%p1836 bra $L__BB106_2096; sqrt.approx.ftz.f32 %f12683, %f3417; mov.f32 %f12684, 0f33D6BF95; max.ftz.f32 %f12685, %f12683, %f12684; mul.ftz.f32 %f12686, %f16558, %f16559; sub.ftz.f32 %f12687, %f12686, %f12685; mul.ftz.f32 %f12688, %f16555, %f12687; mul.ftz.f32 %f12689, %f16556, %f12687; mul.ftz.f32 %f12690, %f16557, %f12687; mul.ftz.f32 %f12691, %f3403, %f16559; sub.ftz.f32 %f16735, %f12688, %f12691; mul.ftz.f32 %f12692, %f12671, %f16559; sub.ftz.f32 %f16736, %f12689, %f12692; mul.ftz.f32 %f12693, %f12672, %f16559; sub.ftz.f32 %f16737, %f12690, %f12693; mul.ftz.f32 %f12694, %f16556, %f16736; fma.rn.ftz.f32 %f12695, %f16555, %f16735, %f12694; fma.rn.ftz.f32 %f12696, %f16557, %f16737, %f12695; mul.ftz.f32 %f12697, %f3401, %f12696; add.ftz.f32 %f12698, %f16558, %f12697; sub.ftz.f32 %f12699, %f16558, %f12697; div.approx.ftz.f32 %f12700, %f12698, %f12699; mul.ftz.f32 %f12701, %f3401, %f16558; add.ftz.f32 %f12702, %f12701, %f12696; sub.ftz.f32 %f12703, %f12701, %f12696; div.approx.ftz.f32 %f12704, %f12702, %f12703; mul.ftz.f32 %f12705, %f12704, %f12704; fma.rn.ftz.f32 %f12706, %f12700, %f12700, %f12705; mul.ftz.f32 %f16563, %f12706, 0f3F000000; $L__BB106_2096: setp.neu.ftz.f32 %p1837, %f16563, 0f3F800000; setp.eq.s16 %p1838, %rs157, 0; and.pred %p1839, %p1838, %p1837; selp.f32 %f16738, 0f49742400, 0f00000000, %p1839; mov.u32 %r3285, 17; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1895: setp.eq.ftz.f32 %p1656, %f2880, 0f7F800000; setp.eq.ftz.f32 %p1657, %f2881, 0f7F800000; and.pred %p1658, %p1656, %p1657; @%p1658 bra $L__BB106_1897; bra.uni $L__BB106_1896; $L__BB106_1897: setp.lt.s32 %p1662, %r878, 0; selp.b32 %r2721, 1075235812, 1061752795, %p1662; or.b32 %r2722, %r879, %r2721; mov.b32 %f16467, %r2722; bra.uni $L__BB106_1899; $L__BB106_2284: neg.ftz.f32 %f16722, %f3944; neg.ftz.f32 %f16719, %f3941; neg.ftz.f32 %f16720, %f3942; neg.ftz.f32 %f16721, %f3943; mov.u16 %rs159, 1; mov.f32 %f16723, %f4009; $L__BB106_2286: mul.ftz.f32 %f14684, %f16722, %f16722; mov.f32 %f16727, 0f3F800000; sub.ftz.f32 %f14685, %f16727, %f14684; mul.ftz.f32 %f14686, %f16723, %f16723; mul.ftz.f32 %f14687, %f14686, %f14685; sub.ftz.f32 %f4020, %f16727, %f14687; setp.lt.ftz.f32 %p2016, %f4020, 0f00000000; mov.f32 %f16738, 0f00000000; mov.f32 %f16735, %f16738; mov.f32 %f16736, %f16738; mov.f32 %f16737, %f16738; @%p2016 bra $L__BB106_2288; sqrt.approx.ftz.f32 %f14688, %f4020; mov.f32 %f14689, 0f33D6BF95; max.ftz.f32 %f14690, %f14688, %f14689; mul.ftz.f32 %f14691, %f16722, %f16723; sub.ftz.f32 %f14692, %f14691, %f14690; mul.ftz.f32 %f14693, %f16719, %f14692; mul.ftz.f32 %f14694, %f16720, %f14692; mul.ftz.f32 %f14695, %f16721, %f14692; mul.ftz.f32 %f14696, %f3899, %f16723; sub.ftz.f32 %f16735, %f14693, %f14696; mul.ftz.f32 %f14697, %f14360, %f16723; sub.ftz.f32 %f16736, %f14694, %f14697; mul.ftz.f32 %f14698, %f14361, %f16723; sub.ftz.f32 %f16737, %f14695, %f14698; mul.ftz.f32 %f14699, %f16720, %f16736; fma.rn.ftz.f32 %f14700, %f16719, %f16735, %f14699; fma.rn.ftz.f32 %f14701, %f16721, %f16737, %f14700; mul.ftz.f32 %f14702, %f4009, %f14701; add.ftz.f32 %f14703, %f16722, %f14702; sub.ftz.f32 %f14704, %f16722, %f14702; div.approx.ftz.f32 %f14705, %f14703, %f14704; mul.ftz.f32 %f14706, %f4009, %f16722; add.ftz.f32 %f14707, %f14706, %f14701; sub.ftz.f32 %f14708, %f14706, %f14701; div.approx.ftz.f32 %f14709, %f14707, %f14708; mul.ftz.f32 %f14710, %f14709, %f14709; fma.rn.ftz.f32 %f14711, %f14705, %f14705, %f14710; mul.ftz.f32 %f16727, %f14711, 0f3F000000; $L__BB106_2288: setp.eq.ftz.f32 %p2017, %f16727, 0f3F800000; setp.ne.s16 %p2018, %rs159, 0; mov.u32 %r2880, 9; or.pred %p2019, %p2018, %p2017; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; mov.u32 %r3285, %r2880; @%p2019 bra $L__BB106_2309; mul.ftz.f32 %f4029, %f14362, %f14363; setp.le.ftz.f32 %p2020, %f4029, 0f33D6BF95; mov.u32 %r3285, 17; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2020 bra $L__BB106_2309; add.ftz.f32 %f14725, %f4009, 0fBF800000; abs.ftz.f32 %f14726, %f14725; setp.lt.ftz.f32 %p2021, %f14726, 0f38D1B717; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2021 bra $L__BB106_2309; mul.ftz.f32 %f14727, %f3940, %f3940; mul.ftz.f32 %f14728, %f14727, %f14727; rcp.approx.ftz.f32 %f14729, %f14727; mov.f32 %f14730, 0f3F800000; add.ftz.f32 %f14731, %f14729, 0fBF800000; mul.ftz.f32 %f14732, %f14728, 0f40490FDB; add.ftz.f32 %f14733, %f4029, %f14731; mul.ftz.f32 %f14734, %f14732, %f14733; mul.ftz.f32 %f14735, %f14733, %f14734; div.approx.ftz.f32 %f14736, %f4029, %f14735; mul.ftz.f32 %f14737, %f3900, %f16736; fma.rn.ftz.f32 %f14738, %f3898, %f16735, %f14737; fma.rn.ftz.f32 %f14739, %f3901, %f16737, %f14738; mul.ftz.f32 %f14740, %f14739, %f14739; sub.ftz.f32 %f14741, %f14730, %f14740; mul.ftz.f32 %f14742, %f4029, %f14741; div.approx.ftz.f32 %f14743, %f14742, %f14740; add.ftz.f32 %f14744, %f14743, 0f3F800000; mov.f32 %f14745, 0f00000000; max.ftz.f32 %f14746, %f14744, %f14745; sqrt.approx.ftz.f32 %f14747, %f14746; add.ftz.f32 %f14748, %f14747, 0f3F800000; mov.f32 %f14749, 0f40000000; div.approx.ftz.f32 %f14750, %f14749, %f14748; mul.ftz.f32 %f14751, %f3942, %f16736; fma.rn.ftz.f32 %f14752, %f3941, %f16735, %f14751; fma.rn.ftz.f32 %f14753, %f3943, %f16737, %f14752; fma.rn.ftz.f32 %f14754, %f4009, %f14753, %f3944; mul.ftz.f32 %f14755, %f14754, %f14754; mul.ftz.f32 %f14756, %f3937, %f14736; mul.ftz.f32 %f14757, %f4009, %f4009; mul.ftz.f32 %f14758, %f14757, %f14756; mul.ftz.f32 %f14759, %f3902, %f14755; div.approx.ftz.f32 %f14760, %f14758, %f14759; mul.ftz.f32 %f14761, %f3944, %f14753; abs.ftz.f32 %f14762, %f14761; mul.ftz.f32 %f14763, %f14750, %f14762; mul.ftz.f32 %f16739, %f14760, %f14763; abs.ftz.f32 %f14764, %f14753; mul.ftz.f32 %f14765, %f3944, %f14764; mul.ftz.f32 %f16738, %f14760, %f14765; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; mov.u32 %r3285, %r2880; bra.uni $L__BB106_2309; $L__BB106_2082: mul.ftz.f32 %f12541, %f3233, %f3352; fma.rn.ftz.f32 %f12542, %f3234, %f3351, %f12541; fma.rn.ftz.f32 %f12543, %f3232, %f3353, %f12542; mul.ftz.f32 %f12544, %f3360, %f12543; mul.ftz.f32 %f12545, %f12544, %f3359; mul.ftz.f32 %f12546, %f3212, %f12545; mul.ftz.f32 %f12547, %f3212, %f12546; mul.ftz.f32 %f12548, %f3350, %f3366; mul.ftz.f32 %f12549, %f3350, %f12548; max.ftz.f32 %f12551, %f12549, %f12525; div.approx.ftz.f32 %f12552, %f12547, %f12551; abs.ftz.f32 %f12553, %f12552; setp.lt.ftz.f32 %p1830, %f3212, 0f3F800000; mov.f32 %f12554, 0f3F800000; selp.f32 %f12555, %f3226, %f3212, %p1830; max.ftz.f32 %f12556, %f12555, %f12554; mov.f32 %f12557, 0f40400000; min.ftz.f32 %f12558, %f12556, %f12557; add.ftz.f32 %f12559, %f12558, 0fBF2814F9; mul.ftz.f32 %f12560, %f12559, 0fBF7A7547; mul.ftz.f32 %f12561, %f12559, %f12560; mul.ftz.f32 %f12562, %f12561, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12563, %f12562; mul.ftz.f32 %f12564, %f12558, 0fBD0A43BB; fma.rn.ftz.f32 %f12565, %f12563, 0f3D435661, %f12564; add.ftz.f32 %f12566, %f12565, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f12567, %f3348; fma.rn.ftz.f32 %f12568, %f12567, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f12569, %f12567, %f12568, 0f3D4D5C09; fma.rn.ftz.f32 %f12570, %f12567, %f12569, 0fBB86D80B; fma.rn.ftz.f32 %f12571, %f12567, %f12570, 0f3F80029F; fma.rn.ftz.f32 %f12572, %f12558, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f12573, %f12558, %f12572, 0fC2FFCFDF; fma.rn.ftz.f32 %f12574, %f12558, %f12573, 0f43927AA0; fma.rn.ftz.f32 %f12575, %f12558, %f12574, 0fC38FF917; mul.ftz.f32 %f12576, %f12558, %f12558; mov.f32 %f12577, 0f4347CD91; div.approx.ftz.f32 %f12578, %f12577, %f12576; add.ftz.f32 %f12579, %f12578, %f12575; mul.ftz.f32 %f12580, %f12558, %f12576; mov.f32 %f12581, 0f42CB5604; div.approx.ftz.f32 %f12582, %f12581, %f12580; sub.ftz.f32 %f12583, %f12579, %f12582; fma.rn.ftz.f32 %f12584, %f12567, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f12585, %f12567, %f12584, 0f41B5F319; fma.rn.ftz.f32 %f12586, %f12567, %f12585, 0fC05A14A5; fma.rn.ftz.f32 %f12587, %f12567, %f12586, 0f3DC9FFD6; fma.rn.ftz.f32 %f12588, %f12567, %f12587, 0f3BA1B61E; mul.ftz.f32 %f12589, %f12588, %f12583; fma.rn.ftz.f32 %f12590, %f12589, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f12591, %f12554, %f12566; sub.ftz.f32 %f12592, %f12554, %f12571; mul.ftz.f32 %f12593, %f12591, %f12592; fma.rn.ftz.f32 %f12594, %f12593, 0fC11520C5, %f12590; cvt.ftz.sat.f32.f32 %f12595, %f12594; sub.ftz.f32 %f12596, %f12554, %f12595; mul.ftz.f32 %f12597, %f3365, %f12596; fma.rn.ftz.f32 %f12598, %f12553, %f12595, %f12597; sub.ftz.f32 %f12599, %f12554, %f16544; mul.ftz.f32 %f16738, %f12599, %f12598; $L__BB106_2084: mul.ftz.f32 %f16739, %f16529, %f16738; mul.ftz.f32 %f16740, %f16530, %f16738; mul.ftz.f32 %f16741, %f16531, %f16738; mul.ftz.f32 %f12622, %f3204, %f3233; fma.rn.ftz.f32 %f12623, %f3201, %f3234, %f12622; mul.ftz.f32 %f12624, %f3205, %f3233; fma.rn.ftz.f32 %f12625, %f3202, %f3234, %f12624; mul.ftz.f32 %f12626, %f3206, %f3233; fma.rn.ftz.f32 %f12627, %f3203, %f3234, %f12626; fma.rn.ftz.f32 %f16735, %f16546, %f3232, %f12623; fma.rn.ftz.f32 %f16736, %f16547, %f3232, %f12625; fma.rn.ftz.f32 %f16737, %f16548, %f3232, %f12627; selp.b32 %r3285, 10, 9, %p1822; bra.uni $L__BB106_2309; $L__BB106_2199: ld.local.f32 %f13820, [%rd587+52]; mul.ftz.f32 %f13821, %f3738, %f13820; ld.local.f32 %f13822, [%rd587+48]; mul.ftz.f32 %f13823, %f3739, %f13822; sub.ftz.f32 %f13824, %f13821, %f13823; ld.local.f32 %f13825, [%rd587+44]; mul.ftz.f32 %f13826, %f3739, %f13825; mul.ftz.f32 %f13827, %f3736, %f13820; sub.ftz.f32 %f13828, %f13826, %f13827; mul.ftz.f32 %f13829, %f3736, %f13822; mul.ftz.f32 %f13830, %f3738, %f13825; sub.ftz.f32 %f13831, %f13829, %f13830; mul.ftz.f32 %f13832, %f13828, %f13828; fma.rn.ftz.f32 %f13833, %f13824, %f13824, %f13832; fma.rn.ftz.f32 %f13834, %f13831, %f13831, %f13833; sqrt.approx.ftz.f32 %f13835, %f13834; div.approx.ftz.f32 %f16621, %f13824, %f13835; div.approx.ftz.f32 %f16622, %f13828, %f13835; div.approx.ftz.f32 %f16623, %f13831, %f13835; mul.ftz.f32 %f13836, %f3739, %f16622; mul.ftz.f32 %f13837, %f3738, %f16623; sub.ftz.f32 %f16624, %f13836, %f13837; mul.ftz.f32 %f13838, %f3736, %f16623; mul.ftz.f32 %f13839, %f3739, %f16621; sub.ftz.f32 %f16625, %f13838, %f13839; mul.ftz.f32 %f13840, %f3738, %f16621; mul.ftz.f32 %f13841, %f3736, %f16622; sub.ftz.f32 %f16626, %f13840, %f13841; $L__BB106_2201: mul.ftz.f32 %f13863, %f13809, %f16625; fma.rn.ftz.f32 %f13864, %f3737, %f16624, %f13863; fma.rn.ftz.f32 %f13865, %f13810, %f16626, %f13864; mul.ftz.f32 %f13866, %f13809, %f16622; fma.rn.ftz.f32 %f13867, %f3737, %f16621, %f13866; fma.rn.ftz.f32 %f13868, %f13810, %f16623, %f13867; mul.ftz.f32 %f13869, %f13811, %f13865; mul.ftz.f32 %f13870, %f13811, %f13868; mul.ftz.f32 %f13871, %f13869, %f13869; fma.rn.ftz.f32 %f13872, %f13870, %f13870, %f13871; fma.rn.ftz.f32 %f13873, %f3740, %f3740, %f13872; sqrt.approx.ftz.f32 %f13874, %f13873; div.approx.ftz.f32 %f3759, %f13869, %f13874; div.approx.ftz.f32 %f3760, %f13870, %f13874; div.approx.ftz.f32 %f3761, %f3740, %f13874; setp.geu.ftz.f32 %p1961, %f3761, 0f3F7FFF58; mov.f32 %f16628, 0f00000000; mov.f32 %f16627, 0f3F800000; mov.f32 %f16629, %f16627; mov.f32 %f16630, %f16628; @%p1961 bra $L__BB106_2203; mul.ftz.f32 %f13875, %f3761, %f3761; mov.f32 %f13876, 0f3F800000; sub.ftz.f32 %f13877, %f13876, %f13875; mov.f32 %f13878, 0f00000000; max.ftz.f32 %f13879, %f13877, %f13878; sqrt.approx.ftz.f32 %f16628, %f13879; rsqrt.approx.ftz.f32 %f13880, %f13879; mul.ftz.f32 %f16629, %f3759, %f13880; mul.ftz.f32 %f16630, %f3760, %f13880; mov.f32 %f16627, %f3761; $L__BB106_2203: setp.ltu.ftz.f32 %p1962, %f16627, 0f3F7FFF58; @%p1962 bra $L__BB106_2205; bra.uni $L__BB106_2204; $L__BB106_2205: div.approx.ftz.f32 %f3771, %f16628, %f16627; rcp.approx.ftz.f32 %f3772, %f3771; abs.ftz.f32 %f3773, %f3772; setp.ge.ftz.f32 %p1963, %f3773, 0f4144CCCD; mov.u32 %r3283, 1065353216; @%p1963 bra $L__BB106_2207; add.ftz.f32 %f13888, %f3773, 0fBF800000; mov.f32 %f13889, 0f3F800000; add.ftz.f32 %f13890, %f13888, 0f3F800000; fma.rn.ftz.f32 %f13891, %f13890, 0f38349F67, 0f39910039; fma.rn.ftz.f32 %f13892, %f13890, %f13891, 0f391F6607; fma.rn.ftz.f32 %f13893, %f13890, %f13892, 0f3C17E369; fma.rn.ftz.f32 %f13894, %f13890, %f13893, 0f3D2D2FE7; fma.rn.ftz.f32 %f13895, %f13890, %f13894, 0f3D906E67; fma.rn.ftz.f32 %f13896, %f13890, %f13895, 0f3F800000; mul.ftz.f32 %f13897, %f13896, %f13896; mul.ftz.f32 %f13898, %f13897, %f13897; mul.ftz.f32 %f13899, %f13898, %f13898; mul.ftz.f32 %f13900, %f13899, %f13899; rcp.approx.ftz.f32 %f13901, %f13900; sub.ftz.f32 %f13902, %f13889, %f13901; mov.b32 %r2841, %f13902; and.b32 %r3283, %r2841, 2147483647; $L__BB106_2207: mov.b32 %r2842, %f3772; and.b32 %r2843, %r2842, -2147483648; or.b32 %r2844, %r3283, %r2843; mov.b32 %f13903, %r2844; mul.ftz.f32 %f13904, %f3772, %f3772; mul.ftz.f32 %f13905, %f13904, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13906, %f13905; add.ftz.f32 %f13907, %f13903, 0fBF800000; mov.f32 %f13908, 0f3F800000; mul.ftz.f32 %f13909, %f13907, 0f3F000000; mul.ftz.f32 %f13910, %f3771, %f13906; fma.rn.ftz.f32 %f3774, %f13910, 0f3E906EBB, %f13909; add.ftz.f32 %f13911, %f13903, 0f3F800000; mul.ftz.f32 %f13912, %f13903, %f13903; sub.ftz.f32 %f13913, %f13908, %f13912; mul.ftz.f32 %f3775, %f3771, 0f3F106EBB; fma.rn.ftz.f32 %f13914, %f3775, %f13913, %f13911; mul.ftz.f32 %f3776, %f2799, %f13914; fma.rn.ftz.f32 %f13915, %f3775, %f13906, %f13911; mul.ftz.f32 %f3777, %f2799, %f13915; setp.gt.ftz.f32 %p1964, %f3775, 0f00000000; @%p1964 bra $L__BB106_2209; bra.uni $L__BB106_2208; $L__BB106_2209: sub.ftz.f32 %f13916, %f3775, %f3776; add.ftz.f32 %f13917, %f13916, 0f3F800000; fma.rn.ftz.f32 %f13918, %f3775, %f13917, 0f3E800000; sqrt.approx.ftz.f32 %f13919, %f13918; mov.f32 %f13920, 0f3F000000; sub.ftz.f32 %f13921, %f13920, %f13919; div.approx.ftz.f32 %f16631, %f13921, %f3775; bra.uni $L__BB106_2210; $L__BB106_2204: lg2.approx.ftz.f32 %f13882, %f2799; mul.ftz.f32 %f13883, %f13882, 0fBF317218; sqrt.approx.ftz.f32 %f13884, %f13883; mul.ftz.f32 %f13885, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f13886, %f13885; mul.ftz.f32 %f16675, %f13884, %f13886; sin.approx.ftz.f32 %f13887, %f13885; mul.ftz.f32 %f16674, %f13884, %f13887; mov.f32 %f16673, 0f3F800000; bra.uni $L__BB106_2225; $L__BB106_1896: setp.lt.s32 %p1659, %r878, 0; min.ftz.f32 %f11555, %f2881, %f2880; max.ftz.f32 %f11556, %f2881, %f2880; div.full.ftz.f32 %f11557, %f11555, %f11556; mul.rn.ftz.f32 %f11558, %f11557, %f11557; mov.f32 %f11559, 0fC0B59883; mov.f32 %f11560, 0fBF52C7EA; fma.rn.ftz.f32 %f11561, %f11558, %f11560, %f11559; mov.f32 %f11562, 0fC0D21907; fma.rn.ftz.f32 %f11563, %f11561, %f11558, %f11562; mul.ftz.f32 %f11564, %f11558, %f11563; mul.ftz.f32 %f11565, %f11557, %f11564; add.ftz.f32 %f11566, %f11558, 0f41355DC0; mov.f32 %f11567, 0f41E6BD60; fma.rn.ftz.f32 %f11568, %f11566, %f11558, %f11567; mov.f32 %f11569, 0f419D92C8; fma.rn.ftz.f32 %f11570, %f11568, %f11558, %f11569; rcp.approx.ftz.f32 %f11571, %f11570; fma.rn.ftz.f32 %f11572, %f11565, %f11571, %f11557; mov.f32 %f11573, 0f3FC90FDB; sub.ftz.f32 %f11574, %f11573, %f11572; setp.gt.ftz.f32 %p1660, %f2881, %f2880; selp.f32 %f11575, %f11574, %f11572, %p1660; mov.f32 %f11576, 0f40490FDB; sub.ftz.f32 %f11577, %f11576, %f11575; selp.f32 %f11578, %f11577, %f11575, %p1659; mov.b32 %r2719, %f11578; or.b32 %r2720, %r879, %r2719; mov.b32 %f11579, %r2720; add.ftz.f32 %f11580, %f2880, %f2881; setp.le.ftz.f32 %p1661, %f11580, 0f7F800000; selp.f32 %f16467, %f11579, %f11580, %p1661; $L__BB106_1899: ld.local.f32 %f11582, [%rd587+44]; div.approx.ftz.f32 %f11583, %f2874, %f11582; mul.ftz.f32 %f11584, %f11583, %f11583; sub.ftz.f32 %f11585, %f11551, %f11584; max.ftz.f32 %f11587, %f11585, %f16472; sqrt.approx.ftz.f32 %f11588, %f11587; ld.f32 %f11589, [%rd588+12]; mul.ftz.f32 %f11590, %f11589, %f11589; sub.ftz.f32 %f11591, %f11551, %f11590; max.ftz.f32 %f11592, %f11591, %f16472; sqrt.approx.ftz.f32 %f11593, %f11592; mov.f32 %f11594, 0fBF800000; max.ftz.f32 %f11595, %f11589, %f11594; min.ftz.f32 %f11596, %f11595, %f11551; abs.ftz.f32 %f11597, %f11596; sub.ftz.f32 %f11598, %f11551, %f11597; mul.ftz.f32 %f11599, %f11598, 0f3F000000; sqrt.approx.ftz.f32 %f11600, %f11599; setp.gt.ftz.f32 %p1663, %f11597, 0f3F11EB85; selp.f32 %f11601, %f11600, %f11597, %p1663; mul.ftz.f32 %f11602, %f11601, %f11601; mov.f32 %f11603, 0f3C94D2E9; mov.f32 %f11604, 0f3D53F941; fma.rn.ftz.f32 %f11605, %f11604, %f11602, %f11603; mov.f32 %f11606, 0f3D3F841F; fma.rn.ftz.f32 %f11607, %f11605, %f11602, %f11606; mov.f32 %f11608, 0f3D994929; fma.rn.ftz.f32 %f11609, %f11607, %f11602, %f11608; mov.f32 %f11610, 0f3E2AAB94; fma.rn.ftz.f32 %f11611, %f11609, %f11602, %f11610; mul.ftz.f32 %f11612, %f11602, %f11611; fma.rn.ftz.f32 %f11613, %f11612, %f11601, %f11601; mov.f32 %f11614, 0f3FC90FDB; mov.f32 %f11615, 0fC0000000; fma.rn.ftz.f32 %f11616, %f11615, %f11613, %f11614; selp.f32 %f11617, %f11616, %f11613, %p1663; setp.le.ftz.f32 %p1664, %f11617, 0f7F800000; mov.b32 %r2726, %f11617; mov.b32 %r2727, %f11596; and.b32 %r2728, %r2727, -2147483648; or.b32 %r2729, %r2728, %r2726; mov.b32 %f11618, %r2729; selp.f32 %f2886, %f11618, %f11617, %p1664; mul.ftz.f32 %f11619, %f2879, %f11589; mul.ftz.f32 %f11620, %f11582, %f11582; sub.ftz.f32 %f11621, %f11620, %f2878; sqrt.approx.ftz.f32 %f11622, %f11621; div.approx.ftz.f32 %f11623, %f11619, %f11622; mul.ftz.f32 %f11624, %f11623, %f11623; sub.ftz.f32 %f11625, %f11551, %f11624; max.ftz.f32 %f11626, %f11625, %f16472; sqrt.approx.ftz.f32 %f11627, %f11626; max.ftz.f32 %f11628, %f11623, %f11594; min.ftz.f32 %f11629, %f11628, %f11551; abs.ftz.f32 %f11630, %f11629; sub.ftz.f32 %f11631, %f11551, %f11630; mul.ftz.f32 %f11632, %f11631, 0f3F000000; sqrt.approx.ftz.f32 %f11633, %f11632; setp.gt.ftz.f32 %p1665, %f11630, 0f3F11EB85; selp.f32 %f11634, %f11633, %f11630, %p1665; mul.ftz.f32 %f11635, %f11634, %f11634; fma.rn.ftz.f32 %f11636, %f11604, %f11635, %f11603; fma.rn.ftz.f32 %f11637, %f11636, %f11635, %f11606; fma.rn.ftz.f32 %f11638, %f11637, %f11635, %f11608; fma.rn.ftz.f32 %f11639, %f11638, %f11635, %f11610; mul.ftz.f32 %f11640, %f11635, %f11639; fma.rn.ftz.f32 %f11641, %f11640, %f11634, %f11634; fma.rn.ftz.f32 %f11642, %f11615, %f11641, %f11614; selp.f32 %f11643, %f11642, %f11641, %p1665; setp.le.ftz.f32 %p1666, %f11643, 0f7F800000; mov.b32 %r2730, %f11643; mov.b32 %r2731, %f11629; and.b32 %r2732, %r2731, -2147483648; or.b32 %r2733, %r2732, %r2730; mov.b32 %f11644, %r2733; selp.f32 %f2887, %f11644, %f11643, %p1666; add.ftz.f32 %f11645, %f11627, %f11627; div.approx.ftz.f32 %f11646, %f11645, %f11588; ld.local.f32 %f11647, [%rd587+20]; mul.ftz.f32 %f2888, %f11647, %f11646; ld.local.f32 %f11648, [%rd587+24]; mul.ftz.f32 %f2889, %f11648, %f11646; ld.local.f32 %f11649, [%rd587+28]; mul.ftz.f32 %f2890, %f11649, %f11646; mul.ftz.f32 %f11650, %f2879, %f11593; abs.ftz.f32 %f2891, %f11650; add.ftz.f32 %f11651, %f11620, 0fBF800000; fma.rn.ftz.f32 %f2892, %f2891, %f2891, %f11651; setp.leu.ftz.f32 %p1667, %f2892, 0f00000000; mov.f32 %f16468, %f11551; @%p1667 bra $L__BB106_1901; sqrt.approx.ftz.f32 %f11652, %f2892; sub.ftz.f32 %f11653, %f11652, %f2891; add.ftz.f32 %f11654, %f2891, %f11652; div.approx.ftz.f32 %f11655, %f11653, %f11654; fma.rn.ftz.f32 %f11656, %f2891, %f11654, 0fBF800000; fma.rn.ftz.f32 %f11657, %f2891, %f11653, 0f3F800000; div.approx.ftz.f32 %f11658, %f11656, %f11657; mul.ftz.f32 %f11659, %f11655, 0f3F000000; mul.ftz.f32 %f11660, %f11655, %f11659; fma.rn.ftz.f32 %f11661, %f11658, %f11658, 0f3F800000; mul.ftz.f32 %f16468, %f11660, %f11661; $L__BB106_1901: mul.ftz.f32 %f11663, %f2888, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11664, %f11663; mul.ftz.f32 %f11665, %f2889, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11666, %f11665; mul.ftz.f32 %f11667, %f2890, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11668, %f11667; sub.ftz.f32 %f11670, %f11551, %f16468; mul.ftz.f32 %f11671, %f11670, %f11670; mul.ftz.f32 %f2895, %f11664, %f11671; mul.ftz.f32 %f2896, %f11666, %f11671; mul.ftz.f32 %f2897, %f11668, %f11671; ld.const.v4.f32 {%f11672, %f11673, %f11674, %f11675}, [kernel_params+1184]; mul.ftz.f32 %f11676, %f2896, %f11673; fma.rn.ftz.f32 %f11677, %f2895, %f11672, %f11676; fma.rn.ftz.f32 %f2901, %f2897, %f11674, %f11677; mul.ftz.f32 %f2902, %f11664, %f16468; mul.ftz.f32 %f2903, %f2902, %f2895; mul.ftz.f32 %f2904, %f11666, %f16468; mul.ftz.f32 %f2905, %f2904, %f2896; mul.ftz.f32 %f2906, %f11668, %f16468; mul.ftz.f32 %f2907, %f2906, %f2897; mul.ftz.f32 %f11678, %f2905, %f11673; fma.rn.ftz.f32 %f11679, %f2903, %f11672, %f11678; fma.rn.ftz.f32 %f2908, %f2907, %f11674, %f11679; sub.ftz.f32 %f2909, %f11551, %f2902; sub.ftz.f32 %f2910, %f11551, %f2904; sub.ftz.f32 %f2911, %f11551, %f2906; setp.eq.ftz.f32 %p1668, %f2909, 0f00000000; mov.f32 %f16469, %f16472; @%p1668 bra $L__BB106_1903; div.approx.ftz.f32 %f16469, %f2902, %f2909; $L__BB106_1903: setp.eq.ftz.f32 %p1669, %f2910, 0f00000000; mov.f32 %f16470, %f16472; @%p1669 bra $L__BB106_1905; div.approx.ftz.f32 %f16470, %f2904, %f2910; $L__BB106_1905: setp.eq.ftz.f32 %p1670, %f2911, 0f00000000; mov.f32 %f16471, %f16472; @%p1670 bra $L__BB106_1907; div.approx.ftz.f32 %f16471, %f2906, %f2911; $L__BB106_1907: mul.ftz.f32 %f2918, %f2903, %f16469; mul.ftz.f32 %f2919, %f2905, %f16470; mul.ftz.f32 %f11683, %f2919, %f11673; fma.rn.ftz.f32 %f11684, %f2918, %f11672, %f11683; mul.ftz.f32 %f2920, %f2907, %f16471; fma.rn.ftz.f32 %f2921, %f2920, %f11674, %f11684; add.ftz.f32 %f11685, %f16468, %f2901; add.ftz.f32 %f11686, %f11685, %f2908; add.ftz.f32 %f2922, %f2921, %f11686; setp.eq.ftz.f32 %p1671, %f2922, 0f00000000; @%p1671 bra $L__BB106_1909; mov.f32 %f11687, 0f3F800000; div.approx.ftz.f32 %f16472, %f11687, %f2922; $L__BB106_1909: mul.ftz.f32 %f2925, %f16472, %f2901; mul.ftz.f32 %f2926, %f16472, %f2908; mul.ftz.f32 %f2927, %f16472, %f2921; mul.ftz.f32 %f2928, %f16472, %f16468; setp.lt.ftz.f32 %p1673, %f2799, %f2928; mov.pred %p2185, -1; mov.u32 %r3278, 0; @%p1673 bra $L__BB106_1913; sub.ftz.f32 %f2929, %f2799, %f2928; setp.lt.ftz.f32 %p1675, %f2929, %f2925; mov.u32 %r3278, 1; @%p1675 bra $L__BB106_1913; sub.ftz.f32 %f11688, %f2929, %f2925; setp.lt.ftz.f32 %p1677, %f11688, %f2926; mov.u32 %r3278, 2; @%p1677 bra $L__BB106_1913; mov.pred %p2185, 0; mov.u32 %r3278, 3; $L__BB106_1913: ld.local.f32 %f2930, [%rd587+32]; mul.ftz.f32 %f2931, %f2930, 0f3E800000; setp.eq.s32 %p1679, %r3278, 1; selp.f32 %f11689, %f2931, %f2930, %p1679; mul.ftz.f32 %f11690, %f11689, 0f40800000; setp.gt.u32 %p1680, %r3278, 1; selp.f32 %f11691, %f11690, %f11689, %p1680; cvt.rn.f32.u32 %f11692, %r876; mul.ftz.f32 %f11693, %f11692, 0f2F800000; mov.f32 %f11694, 0f3727C5AC; max.ftz.f32 %f11695, %f11693, %f11694; mov.f32 %f11696, 0f3F800000; sub.ftz.f32 %f11697, %f11696, %f11695; div.approx.ftz.f32 %f11699, %f11615, %f11691; mul.ftz.f32 %f11700, %f11699, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11701, %f11700; fma.rn.ftz.f32 %f11702, %f11697, %f11701, %f11695; lg2.approx.ftz.f32 %f11703, %f11702; mul.ftz.f32 %f11704, %f11703, 0f3F317218; fma.rn.ftz.f32 %f11705, %f11691, %f11704, 0f3F800000; mul.ftz.f32 %f11706, %f2874, %f11705; mul.ftz.f32 %f11707, %f11705, %f11705; sub.ftz.f32 %f11708, %f11696, %f11707; mov.f32 %f16485, 0f00000000; max.ftz.f32 %f11710, %f11708, %f16485; sqrt.approx.ftz.f32 %f11711, %f11710; cvt.rn.f32.u32 %f11712, %r877; mul.ftz.f32 %f11713, %f11712, 0f2F800000; mul.ftz.f32 %f11714, %f11713, 0f40C90FDB; cos.approx.ftz.f32 %f11715, %f11714; mul.ftz.f32 %f11716, %f11711, %f11715; mul.ftz.f32 %f11717, %f2879, %f11716; sub.ftz.f32 %f16473, %f11717, %f11706; mul.ftz.f32 %f11718, %f16473, %f16473; sub.ftz.f32 %f11719, %f11696, %f11718; max.ftz.f32 %f11720, %f11719, %f16485; sqrt.approx.ftz.f32 %f16474, %f11720; add.u64 %rd2616, %SPL, 704; add.s64 %rd589, %rd2616, 8; not.pred %p1681, %p2185; @%p1681 bra $L__BB106_1915; ld.local.f32 %f11721, [%rd587+40]; neg.ftz.f32 %f11722, %f11721; sin.approx.ftz.f32 %f11723, %f11722; mul.ftz.f32 %f11724, %f11723, %f11723; sub.ftz.f32 %f11726, %f11696, %f11724; mov.f32 %f11727, 0f00000000; max.ftz.f32 %f11728, %f11726, %f11727; sqrt.approx.ftz.f32 %f11729, %f11728; add.ftz.f32 %f11730, %f11723, %f11723; mul.ftz.f32 %f11731, %f11730, %f11729; mul.ftz.f32 %f11732, %f11729, %f11729; sub.ftz.f32 %f11733, %f11732, %f11724; add.ftz.f32 %f11734, %f11731, %f11731; mul.ftz.f32 %f11735, %f11734, %f11733; mul.ftz.f32 %f11736, %f11733, %f11733; mul.ftz.f32 %f11737, %f11731, %f11731; sub.ftz.f32 %f11738, %f11736, %f11737; mul.ftz.f32 %f11739, %f16473, %f11733; mul.ftz.f32 %f11740, %f16474, %f11733; mul.ftz.f32 %f11741, %f16473, %f11731; sub.ftz.f32 %f11742, %f11740, %f11741; fma.rn.ftz.f32 %f11743, %f16474, %f11731, %f11739; abs.ftz.f32 %f11744, %f11742; st.local.v2.f32 [%rd2616], {%f11743, %f11744}; mul.ftz.f32 %f11745, %f16473, %f11729; mul.ftz.f32 %f11746, %f16474, %f11723; mul.ftz.f32 %f11747, %f16474, %f11729; fma.rn.ftz.f32 %f11748, %f16473, %f11723, %f11747; sub.ftz.f32 %f11749, %f11745, %f11746; abs.ftz.f32 %f11750, %f11748; st.local.v2.f32 [%rd589], {%f11749, %f11750}; mul.ftz.f32 %f11751, %f16473, %f11738; mul.ftz.f32 %f11752, %f16474, %f11735; mul.ftz.f32 %f11753, %f16474, %f11738; fma.rn.ftz.f32 %f11754, %f16473, %f11735, %f11753; sub.ftz.f32 %f11755, %f11751, %f11752; abs.ftz.f32 %f11756, %f11754; st.local.v2.f32 [%rd589+8], {%f11755, %f11756}; shl.b32 %r2738, %r3278, 1; mul.wide.u32 %rd2619, %r2738, 4; add.s64 %rd2620, %rd2616, %rd2619; ld.local.v2.f32 {%f16473, %f16474}, [%rd2620]; $L__BB106_1915: add.ftz.f32 %f2938, %f2886, %f2886; @%p2185 bra $L__BB106_1917; bra.uni $L__BB106_1916; $L__BB106_1917: cvt.rn.f32.s32 %f11761, %r3278; add.ftz.f32 %f11762, %f11761, %f11761; mul.ftz.f32 %f11763, %f2887, %f11762; sub.ftz.f32 %f11764, %f11763, %f2938; mov.f32 %f11765, 0f40490FDB; fma.rn.ftz.f32 %f2942, %f11761, 0f40490FDB, %f11764; ld.local.f32 %f16477, [%rd587+36]; div.approx.ftz.f32 %f16476, %f11765, %f16477; setp.gt.ftz.f32 %p1682, %f16476, 0f42B00000; mov.f32 %f16475, 0f00000000; @%p1682 bra $L__BB106_1919; mul.ftz.f32 %f11766, %f16476, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11767, %f11766; add.ftz.f32 %f11768, %f11767, 0f3F800000; rcp.approx.ftz.f32 %f16475, %f11768; $L__BB106_1919: add.ftz.f32 %f11769, %f16475, %f16475; sub.ftz.f32 %f11771, %f11696, %f11769; fma.rn.ftz.f32 %f11772, %f2775, %f11771, %f16475; rcp.approx.ftz.f32 %f11773, %f11772; add.ftz.f32 %f11774, %f11773, 0fBF800000; lg2.approx.ftz.f32 %f11775, %f11774; mul.ftz.f32 %f11776, %f11775, 0fBF317218; mul.ftz.f32 %f11777, %f16477, %f11776; mov.f32 %f11778, 0fC0490FDB; max.ftz.f32 %f11779, %f11777, %f11778; min.ftz.f32 %f11781, %f11779, %f11765; add.ftz.f32 %f16478, %f2942, %f11781; bra.uni $L__BB106_1920; $L__BB106_1916: mul.ftz.f32 %f16478, %f2775, 0f40C90FDB; ld.local.f32 %f16477, [%rd587+36]; mov.f32 %f11759, 0f40490FDB; div.approx.ftz.f32 %f16476, %f11759, %f16477; $L__BB106_1920: ld.local.f32 %f11782, [%rd587+40]; sin.approx.ftz.f32 %f11783, %f11782; mul.ftz.f32 %f11784, %f11783, %f11783; sub.ftz.f32 %f11786, %f11696, %f11784; max.ftz.f32 %f11788, %f11786, %f16485; sqrt.approx.ftz.f32 %f11789, %f11788; add.ftz.f32 %f11790, %f11783, %f11783; mul.ftz.f32 %f11791, %f11790, %f11789; mul.ftz.f32 %f11792, %f11789, %f11789; sub.ftz.f32 %f11793, %f11792, %f11784; add.ftz.f32 %f11794, %f11791, %f11791; mul.ftz.f32 %f11795, %f11794, %f11793; mul.ftz.f32 %f11796, %f11793, %f11793; mul.ftz.f32 %f11797, %f11791, %f11791; sub.ftz.f32 %f11798, %f11796, %f11797; mul.ftz.f32 %f11799, %f16473, %f11793; mul.ftz.f32 %f11800, %f16474, %f11793; mul.ftz.f32 %f11801, %f16473, %f11791; sub.ftz.f32 %f11802, %f11800, %f11801; fma.rn.ftz.f32 %f11803, %f16474, %f11791, %f11799; abs.ftz.f32 %f11804, %f11802; st.local.v2.f32 [%rd2616], {%f11803, %f11804}; mul.ftz.f32 %f11805, %f16473, %f11789; mul.ftz.f32 %f11806, %f16474, %f11783; sub.ftz.f32 %f2951, %f11805, %f11806; mul.ftz.f32 %f11807, %f16474, %f11789; fma.rn.ftz.f32 %f11808, %f16473, %f11783, %f11807; abs.ftz.f32 %f2952, %f11808; st.local.v2.f32 [%rd589], {%f2951, %f2952}; mul.ftz.f32 %f11809, %f16473, %f11798; mul.ftz.f32 %f11810, %f16474, %f11795; sub.ftz.f32 %f2953, %f11809, %f11810; mul.ftz.f32 %f11811, %f16474, %f11798; fma.rn.ftz.f32 %f11812, %f16473, %f11795, %f11811; abs.ftz.f32 %f2954, %f11812; st.local.v2.f32 [%rd589+8], {%f2953, %f2954}; mul.ftz.f32 %f2955, %f16476, 0f3FB8AA3B; ld.local.f32 %f2956, [%rd587+48]; rcp.approx.ftz.f32 %f2957, %f2956; mul.ftz.f32 %f11813, %f2879, %f11804; mul.ftz.f32 %f2958, %f11813, %f2957; mul.ftz.f32 %f11814, %f2874, %f11803; mul.ftz.f32 %f2959, %f11814, %f2957; setp.gtu.ftz.f32 %p1683, %f2956, 0f3DCCCCCD; @%p1683 bra $L__BB106_1934; bra.uni $L__BB106_1921; $L__BB106_1934: mul.ftz.f32 %f2982, %f2958, %f2958; fma.rn.ftz.f32 %f16481, %f2982, 0f3E800000, 0f3F800000; mul.ftz.f32 %f2984, %f2982, %f2982; mov.u64 %rd2631, 64; cvt.rn.f32.u64 %f11845, %rd2631; div.approx.ftz.f32 %f11846, %f2984, %f11845; add.ftz.f32 %f2985, %f16481, %f11846; setp.eq.ftz.f32 %p1693, %f16481, %f2985; @%p1693 bra $L__BB106_1943; mul.ftz.f32 %f2986, %f2982, %f2984; mov.u64 %rd2632, 2304; cvt.rn.f32.u64 %f11847, %rd2632; div.approx.ftz.f32 %f11848, %f2986, %f11847; add.ftz.f32 %f2987, %f2985, %f11848; setp.eq.ftz.f32 %p1694, %f2985, %f2987; mov.f32 %f16481, %f2985; @%p1694 bra $L__BB106_1943; mul.ftz.f32 %f2988, %f2982, %f2986; mov.u64 %rd2633, 147456; cvt.rn.f32.u64 %f11849, %rd2633; div.approx.ftz.f32 %f11850, %f2988, %f11849; add.ftz.f32 %f2989, %f2987, %f11850; setp.eq.ftz.f32 %p1695, %f2987, %f2989; mov.f32 %f16481, %f2987; @%p1695 bra $L__BB106_1943; mul.ftz.f32 %f2990, %f2982, %f2988; mov.u64 %rd2634, 14745600; cvt.rn.f32.u64 %f11851, %rd2634; div.approx.ftz.f32 %f11852, %f2990, %f11851; add.ftz.f32 %f2991, %f2989, %f11852; setp.eq.ftz.f32 %p1696, %f2989, %f2991; mov.f32 %f16481, %f2989; @%p1696 bra $L__BB106_1943; mul.ftz.f32 %f2992, %f2982, %f2990; mov.u64 %rd2635, 2123366400; cvt.rn.f32.u64 %f11853, %rd2635; div.approx.ftz.f32 %f11854, %f2992, %f11853; add.ftz.f32 %f2993, %f2991, %f11854; setp.eq.ftz.f32 %p1697, %f2991, %f2993; mov.f32 %f16481, %f2991; @%p1697 bra $L__BB106_1943; mul.ftz.f32 %f2994, %f2982, %f2992; mov.u64 %rd2636, 416179814400; cvt.rn.f32.u64 %f11855, %rd2636; div.approx.ftz.f32 %f11856, %f2994, %f11855; add.ftz.f32 %f2995, %f2993, %f11856; setp.eq.ftz.f32 %p1698, %f2993, %f2995; mov.f32 %f16481, %f2993; @%p1698 bra $L__BB106_1943; mul.ftz.f32 %f2996, %f2982, %f2994; mov.u64 %rd2637, 106542032486400; cvt.rn.f32.u64 %f11857, %rd2637; div.approx.ftz.f32 %f11858, %f2996, %f11857; add.ftz.f32 %f2997, %f2995, %f11858; setp.eq.ftz.f32 %p1699, %f2995, %f2997; mov.f32 %f16481, %f2995; @%p1699 bra $L__BB106_1943; mul.ftz.f32 %f11859, %f2982, %f2996; mov.u64 %rd2638, 34519618525593600; cvt.rn.f32.u64 %f11860, %rd2638; div.approx.ftz.f32 %f11861, %f11859, %f11860; add.ftz.f32 %f2998, %f2997, %f11861; setp.eq.ftz.f32 %p1700, %f2997, %f2998; mov.f32 %f16481, %f2997; @%p1700 bra $L__BB106_1943; mov.f32 %f16481, %f2998; $L__BB106_1943: mul.ftz.f32 %f11862, %f2959, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11863, %f11862; mul.ftz.f32 %f3000, %f16481, %f11863; abs.ftz.f32 %f3001, %f2957; setp.ltu.ftz.f32 %p1701, %f3001, 0f3F800000; @%p1701 bra $L__BB106_1945; bra.uni $L__BB106_1944; $L__BB106_1945: mul.ftz.f32 %f11885, %f2957, %f2957; mov.f32 %f11886, 0f394FFF49; mov.f32 %f11887, 0f363D0ADA; fma.rn.ftz.f32 %f11888, %f11887, %f11885, %f11886; mov.f32 %f11889, 0f3C08889A; fma.rn.ftz.f32 %f11890, %f11888, %f11885, %f11889; mov.f32 %f11891, 0f3E2AAAAB; fma.rn.ftz.f32 %f11892, %f11890, %f11885, %f11891; mul.ftz.f32 %f11893, %f11885, %f11892; fma.rn.ftz.f32 %f16482, %f11893, %f2957, %f2957; bra.uni $L__BB106_1946; $L__BB106_1921: setp.gt.ftz.f32 %p1684, %f2958, 0f41400000; @%p1684 bra $L__BB106_1932; bra.uni $L__BB106_1922; $L__BB106_1932: mul.ftz.f32 %f11833, %f2958, 0f41000000; rcp.approx.ftz.f32 %f11834, %f11833; add.ftz.f32 %f11835, %f11834, 0fBFEB3F8E; lg2.approx.ftz.f32 %f11836, %f2958; fma.rn.ftz.f32 %f11837, %f11836, 0fBF317218, %f11835; fma.rn.ftz.f32 %f16480, %f11837, 0f3F000000, %f2958; bra.uni $L__BB106_1933; $L__BB106_2208: add.ftz.f32 %f16631, %f3776, 0fBF800000; $L__BB106_2210: abs.ftz.f32 %f13922, %f16631; setp.gt.ftz.f32 %p1965, %f13922, 0f3F7FFFFF; selp.f32 %f13923, 0f3F7FFFFF, %f13922, %p1965; sub.ftz.f32 %f13925, %f13908, %f13923; add.ftz.f32 %f13926, %f13923, 0f3F800000; mul.ftz.f32 %f13927, %f13925, %f13926; mov.f32 %f13928, 0f00800000; max.ftz.f32 %f13929, %f13927, %f13928; mov.f32 %f13930, 0f7F7FFFFF; min.ftz.f32 %f13931, %f13929, %f13930; mov.b32 %r2845, %f13931; shr.u32 %r2846, %r2845, 23; add.s32 %r2847, %r2846, -127; and.b32 %r2848, %r2845, 8388607; or.b32 %r2849, %r2848, 1065353216; mov.b32 %f13932, %r2849; add.ftz.f32 %f13933, %f13932, 0fBF800000; mul.ftz.f32 %f13934, %f13933, %f13933; mul.ftz.f32 %f13935, %f13934, %f13934; fma.rn.ftz.f32 %f13936, %f13933, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f13937, %f13933, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f13938, %f13933, %f13936, 0fBE0CD4FD; fma.rn.ftz.f32 %f13939, %f13933, %f13938, 0f3E77ADBD; fma.rn.ftz.f32 %f13940, %f13933, %f13939, 0fBEB1D206; fma.rn.ftz.f32 %f13941, %f13933, %f13937, 0f3FB8AA10; mul.ftz.f32 %f13942, %f13935, %f13940; fma.rn.ftz.f32 %f13943, %f13933, %f13941, %f13942; cvt.rn.f32.s32 %f13944, %r2847; add.ftz.f32 %f13945, %f13943, %f13944; mul.ftz.f32 %f3781, %f13945, 0f3F317218; setp.gt.ftz.f32 %p1966, %f3781, 0fC0A00000; @%p1966 bra $L__BB106_2212; bra.uni $L__BB106_2211; $L__BB106_2212: mov.f32 %f13966, 0fC0200000; sub.ftz.f32 %f16633, %f13966, %f3781; mov.f32 %f16641, 0f32F16588; mov.f32 %f16640, 0f34B84B36; mov.f32 %f16639, 0fB66C7357; mov.f32 %f16638, 0fB6935AC1; mov.f32 %f16637, 0f396532DB; mov.f32 %f16636, 0fBAA45408; mov.f32 %f16635, 0fBB88E4EF; mov.f32 %f16634, 0f3E7C8F63; mov.f32 %f16632, 0f3FC02E2F; bra.uni $L__BB106_2213; $L__BB106_1944: mov.f32 %f11864, 0f3FB8AA3B; mul.rn.ftz.f32 %f11865, %f3001, %f11864; cvt.rzi.ftz.f32.f32 %f11866, %f11865; abs.ftz.f32 %f11867, %f11866; setp.gt.ftz.f32 %p1702, %f11867, 0f42FC0000; mov.b32 %r2739, %f11866; and.b32 %r2740, %r2739, -2147483648; or.b32 %r2741, %r2740, 1123811328; mov.b32 %f11868, %r2741; selp.f32 %f11869, %f11868, %f11866, %p1702; mov.f32 %f11870, 0fBF317218; fma.rn.ftz.f32 %f11871, %f11869, %f11870, %f3001; mov.f32 %f11872, 0f3102E308; fma.rn.ftz.f32 %f11873, %f11869, %f11872, %f11871; mul.ftz.f32 %f11874, %f11873, 0f3FB8AA3B; add.ftz.f32 %f11875, %f11869, 0f4B40007D; mov.b32 %r2742, %f11875; shl.b32 %r2743, %r2742, 23; mov.b32 %f11876, %r2743; ex2.approx.ftz.f32 %f11877, %f11874; mul.ftz.f32 %f11878, %f11877, %f11876; mov.f32 %f11879, 0f3E000000; div.approx.ftz.f32 %f11880, %f11879, %f11878; neg.ftz.f32 %f11881, %f11880; mov.f32 %f11882, 0f40000000; fma.rn.ftz.f32 %f11883, %f11882, %f11878, %f11881; setp.ge.ftz.f32 %p1703, %f3001, 0f42B40000; selp.f32 %f11884, 0f7F800000, %f11883, %p1703; mov.b32 %r2744, %f11884; mov.b32 %r2745, %f2957; and.b32 %r2746, %r2745, -2147483648; or.b32 %r2747, %r2746, %r2744; mov.b32 %f16482, %r2747; $L__BB106_1946: add.ftz.f32 %f11894, %f16482, %f16482; mul.ftz.f32 %f11895, %f2956, %f11894; div.approx.ftz.f32 %f16483, %f3000, %f11895; bra.uni $L__BB106_1947; $L__BB106_1922: mul.ftz.f32 %f2960, %f2958, %f2958; fma.rn.ftz.f32 %f16479, %f2960, 0f3E800000, 0f3F800000; mul.ftz.f32 %f2962, %f2960, %f2960; mov.u64 %rd2623, 64; cvt.rn.f32.u64 %f11815, %rd2623; div.approx.ftz.f32 %f11816, %f2962, %f11815; add.ftz.f32 %f2963, %f16479, %f11816; setp.eq.ftz.f32 %p1685, %f16479, %f2963; @%p1685 bra $L__BB106_1931; mul.ftz.f32 %f2964, %f2960, %f2962; mov.u64 %rd2624, 2304; cvt.rn.f32.u64 %f11817, %rd2624; div.approx.ftz.f32 %f11818, %f2964, %f11817; add.ftz.f32 %f2965, %f2963, %f11818; setp.eq.ftz.f32 %p1686, %f2963, %f2965; mov.f32 %f16479, %f2963; @%p1686 bra $L__BB106_1931; mul.ftz.f32 %f2966, %f2960, %f2964; mov.u64 %rd2625, 147456; cvt.rn.f32.u64 %f11819, %rd2625; div.approx.ftz.f32 %f11820, %f2966, %f11819; add.ftz.f32 %f2967, %f2965, %f11820; setp.eq.ftz.f32 %p1687, %f2965, %f2967; mov.f32 %f16479, %f2965; @%p1687 bra $L__BB106_1931; mul.ftz.f32 %f2968, %f2960, %f2966; mov.u64 %rd2626, 14745600; cvt.rn.f32.u64 %f11821, %rd2626; div.approx.ftz.f32 %f11822, %f2968, %f11821; add.ftz.f32 %f2969, %f2967, %f11822; setp.eq.ftz.f32 %p1688, %f2967, %f2969; mov.f32 %f16479, %f2967; @%p1688 bra $L__BB106_1931; mul.ftz.f32 %f2970, %f2960, %f2968; mov.u64 %rd2627, 2123366400; cvt.rn.f32.u64 %f11823, %rd2627; div.approx.ftz.f32 %f11824, %f2970, %f11823; add.ftz.f32 %f2971, %f2969, %f11824; setp.eq.ftz.f32 %p1689, %f2969, %f2971; mov.f32 %f16479, %f2969; @%p1689 bra $L__BB106_1931; mul.ftz.f32 %f2972, %f2960, %f2970; mov.u64 %rd2628, 416179814400; cvt.rn.f32.u64 %f11825, %rd2628; div.approx.ftz.f32 %f11826, %f2972, %f11825; add.ftz.f32 %f2973, %f2971, %f11826; setp.eq.ftz.f32 %p1690, %f2971, %f2973; mov.f32 %f16479, %f2971; @%p1690 bra $L__BB106_1931; mul.ftz.f32 %f2974, %f2960, %f2972; mov.u64 %rd2629, 106542032486400; cvt.rn.f32.u64 %f11827, %rd2629; div.approx.ftz.f32 %f11828, %f2974, %f11827; add.ftz.f32 %f2975, %f2973, %f11828; setp.eq.ftz.f32 %p1691, %f2973, %f2975; mov.f32 %f16479, %f2973; @%p1691 bra $L__BB106_1931; mul.ftz.f32 %f11829, %f2960, %f2974; mov.u64 %rd2630, 34519618525593600; cvt.rn.f32.u64 %f11830, %rd2630; div.approx.ftz.f32 %f11831, %f11829, %f11830; add.ftz.f32 %f2976, %f2975, %f11831; setp.eq.ftz.f32 %p1692, %f2975, %f2976; mov.f32 %f16479, %f2975; @%p1692 bra $L__BB106_1931; mov.f32 %f16479, %f2976; $L__BB106_1931: lg2.approx.ftz.f32 %f11832, %f16479; mul.ftz.f32 %f16480, %f11832, 0f3F317218; $L__BB106_1933: sub.ftz.f32 %f11838, %f16480, %f2959; sub.ftz.f32 %f11839, %f11838, %f2957; add.ftz.f32 %f11840, %f11839, 0f3F316F00; mul.ftz.f32 %f11841, %f2957, 0f3F000000; lg2.approx.ftz.f32 %f11842, %f11841; fma.rn.ftz.f32 %f11843, %f11842, 0f3F317218, %f11840; mul.ftz.f32 %f11844, %f11843, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16483, %f11844; $L__BB106_1947: setp.gt.ftz.f32 %p1704, %f16476, 0f42B00000; mov.f32 %f16484, %f16485; @%p1704 bra $L__BB106_1949; ex2.approx.ftz.f32 %f11897, %f2955; add.ftz.f32 %f11898, %f11897, 0f3F800000; rcp.approx.ftz.f32 %f11899, %f11898; add.ftz.f32 %f16484, %f11899, %f11899; $L__BB106_1949: sub.ftz.f32 %f3009, %f11696, %f16484; setp.eq.ftz.f32 %p1705, %f3009, 0f00000000; @%p1705 bra $L__BB106_1951; mul.ftz.f32 %f11902, %f2887, 0f00000000; sub.ftz.f32 %f11903, %f11902, %f2938; add.ftz.f32 %f11904, %f11903, 0f00000000; sub.ftz.f32 %f11905, %f16478, %f11904; add.ftz.f32 %f11906, %f11905, 0f40490FDB; mov.f32 %f11907, 0f40C90FDB; div.approx.ftz.f32 %f11908, %f11906, %f11907; cvt.rmi.ftz.f32.f32 %f11909, %f11908; mul.ftz.f32 %f11910, %f11909, 0f40C90FDB; sub.ftz.f32 %f11911, %f11906, %f11910; add.ftz.f32 %f11912, %f11911, 0fC0490FDB; abs.ftz.f32 %f11913, %f11912; neg.ftz.f32 %f11914, %f11913; div.approx.ftz.f32 %f11915, %f11914, %f16477; mul.ftz.f32 %f11916, %f11915, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11917, %f11916; add.ftz.f32 %f11918, %f11917, 0f3F800000; mul.ftz.f32 %f11919, %f11918, %f11918; mul.ftz.f32 %f11920, %f16477, %f11919; div.approx.ftz.f32 %f11921, %f11917, %f11920; div.approx.ftz.f32 %f16485, %f11921, %f3009; $L__BB106_1951: mul.ftz.f32 %f11922, %f16483, %f16468; fma.rn.ftz.f32 %f3012, %f16485, %f11922, 0f00000000; mul.ftz.f32 %f11923, %f16483, %f2928; fma.rn.ftz.f32 %f3013, %f16485, %f11923, 0f00000000; mul.ftz.f32 %f11924, %f2879, %f2952; rcp.approx.ftz.f32 %f3014, %f2931; mul.ftz.f32 %f3015, %f11924, %f3014; mul.ftz.f32 %f11925, %f2874, %f2951; mul.ftz.f32 %f3016, %f11925, %f3014; setp.gtu.ftz.f32 %p1706, %f2931, 0f3DCCCCCD; @%p1706 bra $L__BB106_1965; bra.uni $L__BB106_1952; $L__BB106_1965: mul.ftz.f32 %f3039, %f3015, %f3015; fma.rn.ftz.f32 %f16488, %f3039, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3041, %f3039, %f3039; mov.u64 %rd2647, 64; cvt.rn.f32.u64 %f11956, %rd2647; div.approx.ftz.f32 %f11957, %f3041, %f11956; add.ftz.f32 %f3042, %f16488, %f11957; setp.eq.ftz.f32 %p1716, %f16488, %f3042; @%p1716 bra $L__BB106_1974; mul.ftz.f32 %f3043, %f3039, %f3041; mov.u64 %rd2648, 2304; cvt.rn.f32.u64 %f11958, %rd2648; div.approx.ftz.f32 %f11959, %f3043, %f11958; add.ftz.f32 %f3044, %f3042, %f11959; setp.eq.ftz.f32 %p1717, %f3042, %f3044; mov.f32 %f16488, %f3042; @%p1717 bra $L__BB106_1974; mul.ftz.f32 %f3045, %f3039, %f3043; mov.u64 %rd2649, 147456; cvt.rn.f32.u64 %f11960, %rd2649; div.approx.ftz.f32 %f11961, %f3045, %f11960; add.ftz.f32 %f3046, %f3044, %f11961; setp.eq.ftz.f32 %p1718, %f3044, %f3046; mov.f32 %f16488, %f3044; @%p1718 bra $L__BB106_1974; mul.ftz.f32 %f3047, %f3039, %f3045; mov.u64 %rd2650, 14745600; cvt.rn.f32.u64 %f11962, %rd2650; div.approx.ftz.f32 %f11963, %f3047, %f11962; add.ftz.f32 %f3048, %f3046, %f11963; setp.eq.ftz.f32 %p1719, %f3046, %f3048; mov.f32 %f16488, %f3046; @%p1719 bra $L__BB106_1974; mul.ftz.f32 %f3049, %f3039, %f3047; mov.u64 %rd2651, 2123366400; cvt.rn.f32.u64 %f11964, %rd2651; div.approx.ftz.f32 %f11965, %f3049, %f11964; add.ftz.f32 %f3050, %f3048, %f11965; setp.eq.ftz.f32 %p1720, %f3048, %f3050; mov.f32 %f16488, %f3048; @%p1720 bra $L__BB106_1974; mul.ftz.f32 %f3051, %f3039, %f3049; mov.u64 %rd2652, 416179814400; cvt.rn.f32.u64 %f11966, %rd2652; div.approx.ftz.f32 %f11967, %f3051, %f11966; add.ftz.f32 %f3052, %f3050, %f11967; setp.eq.ftz.f32 %p1721, %f3050, %f3052; mov.f32 %f16488, %f3050; @%p1721 bra $L__BB106_1974; mul.ftz.f32 %f3053, %f3039, %f3051; mov.u64 %rd2653, 106542032486400; cvt.rn.f32.u64 %f11968, %rd2653; div.approx.ftz.f32 %f11969, %f3053, %f11968; add.ftz.f32 %f3054, %f3052, %f11969; setp.eq.ftz.f32 %p1722, %f3052, %f3054; mov.f32 %f16488, %f3052; @%p1722 bra $L__BB106_1974; mul.ftz.f32 %f11970, %f3039, %f3053; mov.u64 %rd2654, 34519618525593600; cvt.rn.f32.u64 %f11971, %rd2654; div.approx.ftz.f32 %f11972, %f11970, %f11971; add.ftz.f32 %f3055, %f3054, %f11972; setp.eq.ftz.f32 %p1723, %f3054, %f3055; mov.f32 %f16488, %f3054; @%p1723 bra $L__BB106_1974; mov.f32 %f16488, %f3055; $L__BB106_1974: mul.ftz.f32 %f11973, %f3016, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11974, %f11973; mul.ftz.f32 %f3057, %f16488, %f11974; abs.ftz.f32 %f3058, %f3014; setp.ltu.ftz.f32 %p1724, %f3058, 0f3F800000; @%p1724 bra $L__BB106_1976; bra.uni $L__BB106_1975; $L__BB106_1976: mul.ftz.f32 %f11996, %f3014, %f3014; mov.f32 %f11997, 0f394FFF49; mov.f32 %f11998, 0f363D0ADA; fma.rn.ftz.f32 %f11999, %f11998, %f11996, %f11997; mov.f32 %f12000, 0f3C08889A; fma.rn.ftz.f32 %f12001, %f11999, %f11996, %f12000; mov.f32 %f12002, 0f3E2AAAAB; fma.rn.ftz.f32 %f12003, %f12001, %f11996, %f12002; mul.ftz.f32 %f12004, %f11996, %f12003; fma.rn.ftz.f32 %f16489, %f12004, %f3014, %f3014; bra.uni $L__BB106_1977; $L__BB106_1952: setp.gt.ftz.f32 %p1707, %f3015, 0f41400000; @%p1707 bra $L__BB106_1963; bra.uni $L__BB106_1953; $L__BB106_1963: mul.ftz.f32 %f11944, %f3015, 0f41000000; rcp.approx.ftz.f32 %f11945, %f11944; add.ftz.f32 %f11946, %f11945, 0fBFEB3F8E; lg2.approx.ftz.f32 %f11947, %f3015; fma.rn.ftz.f32 %f11948, %f11947, 0fBF317218, %f11946; fma.rn.ftz.f32 %f16487, %f11948, 0f3F000000, %f3015; bra.uni $L__BB106_1964; $L__BB106_2211: neg.ftz.f32 %f13955, %f3781; sqrt.approx.ftz.f32 %f13956, %f13955; add.ftz.f32 %f16633, %f13956, 0fC0400000; mov.f32 %f16641, 0fB951F09B; mov.f32 %f16640, 0f38D3B56B; mov.f32 %f16639, 0f3AB0DC72; mov.f32 %f16638, 0fBB70BDE7; mov.f32 %f16637, 0f3BBC127B; mov.f32 %f16636, 0fBBF9C5D7; mov.f32 %f16635, 0f3C1AA57E; mov.f32 %f16634, 0f3F8036DB; mov.f32 %f16632, 0f40354F7E; $L__BB106_2213: fma.rn.ftz.f32 %f13967, %f16633, %f16641, %f16640; fma.rn.ftz.f32 %f13968, %f16633, %f13967, %f16639; fma.rn.ftz.f32 %f13969, %f16633, %f13968, %f16638; fma.rn.ftz.f32 %f13970, %f16633, %f13969, %f16637; fma.rn.ftz.f32 %f13971, %f16633, %f13970, %f16636; fma.rn.ftz.f32 %f13972, %f16633, %f13971, %f16635; fma.rn.ftz.f32 %f13973, %f16633, %f13972, %f16634; fma.rn.ftz.f32 %f13974, %f16633, %f13973, %f16632; mul.ftz.f32 %f16675, %f16631, %f13974; mul.ftz.f32 %f13975, %f16675, %f16675; mul.ftz.f32 %f13976, %f13975, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13977, %f13976; add.ftz.f32 %f13978, %f16631, 0f3F800000; fma.rn.ftz.f32 %f13979, %f3775, %f13977, %f13978; sub.ftz.f32 %f3795, %f13979, %f3777; abs.ftz.f32 %f13980, %f3795; setp.leu.ftz.f32 %p1967, %f13980, 0f358637BD; @%p1967 bra $L__BB106_2221; mul.ftz.f32 %f13981, %f3771, %f16675; sub.ftz.f32 %f13983, %f13908, %f13981; div.approx.ftz.f32 %f13984, %f3795, %f13983; sub.ftz.f32 %f3796, %f16631, %f13984; abs.ftz.f32 %f13985, %f3796; setp.gt.ftz.f32 %p1968, %f13985, 0f3F7FFFFF; selp.f32 %f13986, 0f3F7FFFFF, %f13985, %p1968; sub.ftz.f32 %f13987, %f13908, %f13986; add.ftz.f32 %f13988, %f13986, 0f3F800000; mul.ftz.f32 %f13989, %f13987, %f13988; max.ftz.f32 %f13991, %f13989, %f13928; min.ftz.f32 %f13993, %f13991, %f13930; mov.b32 %r2850, %f13993; shr.u32 %r2851, %r2850, 23; add.s32 %r2852, %r2851, -127; and.b32 %r2853, %r2850, 8388607; or.b32 %r2854, %r2853, 1065353216; mov.b32 %f13994, %r2854; add.ftz.f32 %f13995, %f13994, 0fBF800000; mul.ftz.f32 %f13996, %f13995, %f13995; mul.ftz.f32 %f13997, %f13996, %f13996; fma.rn.ftz.f32 %f13998, %f13995, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f13999, %f13995, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14000, %f13995, %f13998, 0fBE0CD4FD; fma.rn.ftz.f32 %f14001, %f13995, %f14000, 0f3E77ADBD; fma.rn.ftz.f32 %f14002, %f13995, %f14001, 0fBEB1D206; fma.rn.ftz.f32 %f14003, %f13995, %f13999, 0f3FB8AA10; mul.ftz.f32 %f14004, %f13997, %f14002; fma.rn.ftz.f32 %f14005, %f13995, %f14003, %f14004; cvt.rn.f32.s32 %f14006, %r2852; add.ftz.f32 %f14007, %f14005, %f14006; mul.ftz.f32 %f3797, %f14007, 0f3F317218; setp.gt.ftz.f32 %p1969, %f3797, 0fC0A00000; @%p1969 bra $L__BB106_2216; bra.uni $L__BB106_2215; $L__BB106_2216: mov.f32 %f14028, 0fC0200000; sub.ftz.f32 %f16643, %f14028, %f3797; mov.f32 %f16651, 0f32F16588; mov.f32 %f16650, 0f34B84B36; mov.f32 %f16649, 0fB66C7357; mov.f32 %f16648, 0fB6935AC1; mov.f32 %f16647, 0f396532DB; mov.f32 %f16646, 0fBAA45408; mov.f32 %f16645, 0fBB88E4EF; mov.f32 %f16644, 0f3E7C8F63; mov.f32 %f16642, 0f3FC02E2F; bra.uni $L__BB106_2217; $L__BB106_1975: mov.f32 %f11975, 0f3FB8AA3B; mul.rn.ftz.f32 %f11976, %f3058, %f11975; cvt.rzi.ftz.f32.f32 %f11977, %f11976; abs.ftz.f32 %f11978, %f11977; setp.gt.ftz.f32 %p1725, %f11978, 0f42FC0000; mov.b32 %r2748, %f11977; and.b32 %r2749, %r2748, -2147483648; or.b32 %r2750, %r2749, 1123811328; mov.b32 %f11979, %r2750; selp.f32 %f11980, %f11979, %f11977, %p1725; mov.f32 %f11981, 0fBF317218; fma.rn.ftz.f32 %f11982, %f11980, %f11981, %f3058; mov.f32 %f11983, 0f3102E308; fma.rn.ftz.f32 %f11984, %f11980, %f11983, %f11982; mul.ftz.f32 %f11985, %f11984, 0f3FB8AA3B; add.ftz.f32 %f11986, %f11980, 0f4B40007D; mov.b32 %r2751, %f11986; shl.b32 %r2752, %r2751, 23; mov.b32 %f11987, %r2752; ex2.approx.ftz.f32 %f11988, %f11985; mul.ftz.f32 %f11989, %f11988, %f11987; mov.f32 %f11990, 0f3E000000; div.approx.ftz.f32 %f11991, %f11990, %f11989; neg.ftz.f32 %f11992, %f11991; mov.f32 %f11993, 0f40000000; fma.rn.ftz.f32 %f11994, %f11993, %f11989, %f11992; setp.ge.ftz.f32 %p1726, %f3058, 0f42B40000; selp.f32 %f11995, 0f7F800000, %f11994, %p1726; mov.b32 %r2753, %f11995; mov.b32 %r2754, %f3014; and.b32 %r2755, %r2754, -2147483648; or.b32 %r2756, %r2755, %r2753; mov.b32 %f16489, %r2756; $L__BB106_1977: add.ftz.f32 %f12005, %f16489, %f16489; mul.ftz.f32 %f12006, %f2931, %f12005; div.approx.ftz.f32 %f16490, %f3057, %f12006; bra.uni $L__BB106_1978; $L__BB106_1953: mul.ftz.f32 %f3017, %f3015, %f3015; fma.rn.ftz.f32 %f16486, %f3017, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3019, %f3017, %f3017; mov.u64 %rd2639, 64; cvt.rn.f32.u64 %f11926, %rd2639; div.approx.ftz.f32 %f11927, %f3019, %f11926; add.ftz.f32 %f3020, %f16486, %f11927; setp.eq.ftz.f32 %p1708, %f16486, %f3020; @%p1708 bra $L__BB106_1962; mul.ftz.f32 %f3021, %f3017, %f3019; mov.u64 %rd2640, 2304; cvt.rn.f32.u64 %f11928, %rd2640; div.approx.ftz.f32 %f11929, %f3021, %f11928; add.ftz.f32 %f3022, %f3020, %f11929; setp.eq.ftz.f32 %p1709, %f3020, %f3022; mov.f32 %f16486, %f3020; @%p1709 bra $L__BB106_1962; mul.ftz.f32 %f3023, %f3017, %f3021; mov.u64 %rd2641, 147456; cvt.rn.f32.u64 %f11930, %rd2641; div.approx.ftz.f32 %f11931, %f3023, %f11930; add.ftz.f32 %f3024, %f3022, %f11931; setp.eq.ftz.f32 %p1710, %f3022, %f3024; mov.f32 %f16486, %f3022; @%p1710 bra $L__BB106_1962; mul.ftz.f32 %f3025, %f3017, %f3023; mov.u64 %rd2642, 14745600; cvt.rn.f32.u64 %f11932, %rd2642; div.approx.ftz.f32 %f11933, %f3025, %f11932; add.ftz.f32 %f3026, %f3024, %f11933; setp.eq.ftz.f32 %p1711, %f3024, %f3026; mov.f32 %f16486, %f3024; @%p1711 bra $L__BB106_1962; mul.ftz.f32 %f3027, %f3017, %f3025; mov.u64 %rd2643, 2123366400; cvt.rn.f32.u64 %f11934, %rd2643; div.approx.ftz.f32 %f11935, %f3027, %f11934; add.ftz.f32 %f3028, %f3026, %f11935; setp.eq.ftz.f32 %p1712, %f3026, %f3028; mov.f32 %f16486, %f3026; @%p1712 bra $L__BB106_1962; mul.ftz.f32 %f3029, %f3017, %f3027; mov.u64 %rd2644, 416179814400; cvt.rn.f32.u64 %f11936, %rd2644; div.approx.ftz.f32 %f11937, %f3029, %f11936; add.ftz.f32 %f3030, %f3028, %f11937; setp.eq.ftz.f32 %p1713, %f3028, %f3030; mov.f32 %f16486, %f3028; @%p1713 bra $L__BB106_1962; mul.ftz.f32 %f3031, %f3017, %f3029; mov.u64 %rd2645, 106542032486400; cvt.rn.f32.u64 %f11938, %rd2645; div.approx.ftz.f32 %f11939, %f3031, %f11938; add.ftz.f32 %f3032, %f3030, %f11939; setp.eq.ftz.f32 %p1714, %f3030, %f3032; mov.f32 %f16486, %f3030; @%p1714 bra $L__BB106_1962; mul.ftz.f32 %f11940, %f3017, %f3031; mov.u64 %rd2646, 34519618525593600; cvt.rn.f32.u64 %f11941, %rd2646; div.approx.ftz.f32 %f11942, %f11940, %f11941; add.ftz.f32 %f3033, %f3032, %f11942; setp.eq.ftz.f32 %p1715, %f3032, %f3033; mov.f32 %f16486, %f3032; @%p1715 bra $L__BB106_1962; mov.f32 %f16486, %f3033; $L__BB106_1962: lg2.approx.ftz.f32 %f11943, %f16486; mul.ftz.f32 %f16487, %f11943, 0f3F317218; $L__BB106_1964: sub.ftz.f32 %f11949, %f16487, %f3016; sub.ftz.f32 %f11950, %f11949, %f3014; add.ftz.f32 %f11951, %f11950, 0f3F316F00; mul.ftz.f32 %f11952, %f3014, 0f3F000000; lg2.approx.ftz.f32 %f11953, %f11952; fma.rn.ftz.f32 %f11954, %f11953, 0f3F317218, %f11951; mul.ftz.f32 %f11955, %f11954, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16490, %f11955; $L__BB106_1978: mov.f32 %f16492, 0f00000000; mov.f32 %f16491, %f16492; @%p1704 bra $L__BB106_1980; ex2.approx.ftz.f32 %f12008, %f2955; add.ftz.f32 %f12009, %f12008, 0f3F800000; rcp.approx.ftz.f32 %f12010, %f12009; add.ftz.f32 %f16491, %f12010, %f12010; $L__BB106_1980: sub.ftz.f32 %f3066, %f11696, %f16491; setp.eq.ftz.f32 %p1728, %f3066, 0f00000000; @%p1728 bra $L__BB106_1982; add.ftz.f32 %f12013, %f2887, %f2887; sub.ftz.f32 %f12014, %f12013, %f2938; add.ftz.f32 %f12015, %f12014, 0f40490FDB; sub.ftz.f32 %f12016, %f16478, %f12015; add.ftz.f32 %f12017, %f12016, 0f40490FDB; mov.f32 %f12018, 0f40C90FDB; div.approx.ftz.f32 %f12019, %f12017, %f12018; cvt.rmi.ftz.f32.f32 %f12020, %f12019; mul.ftz.f32 %f12021, %f12020, 0f40C90FDB; sub.ftz.f32 %f12022, %f12017, %f12021; add.ftz.f32 %f12023, %f12022, 0fC0490FDB; abs.ftz.f32 %f12024, %f12023; neg.ftz.f32 %f12025, %f12024; div.approx.ftz.f32 %f12026, %f12025, %f16477; mul.ftz.f32 %f12027, %f12026, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12028, %f12027; add.ftz.f32 %f12029, %f12028, 0f3F800000; mul.ftz.f32 %f12030, %f12029, %f12029; mul.ftz.f32 %f12031, %f16477, %f12030; div.approx.ftz.f32 %f12032, %f12028, %f12031; div.approx.ftz.f32 %f16492, %f12032, %f3066; $L__BB106_1982: mul.ftz.f32 %f12033, %f16490, %f2895; mul.ftz.f32 %f12034, %f16490, %f2896; mul.ftz.f32 %f12035, %f16490, %f2897; fma.rn.ftz.f32 %f3069, %f16492, %f12033, %f3012; fma.rn.ftz.f32 %f3070, %f16492, %f12034, %f3012; fma.rn.ftz.f32 %f3071, %f16492, %f12035, %f3012; mul.ftz.f32 %f12036, %f16490, %f2925; fma.rn.ftz.f32 %f3072, %f16492, %f12036, %f3013; mul.ftz.f32 %f3073, %f2930, 0f40800000; rcp.approx.ftz.f32 %f3074, %f3073; mul.ftz.f32 %f12037, %f2879, %f2954; mul.ftz.f32 %f3075, %f12037, %f3074; mul.ftz.f32 %f12038, %f2874, %f2953; mul.ftz.f32 %f3076, %f12038, %f3074; setp.gtu.ftz.f32 %p1729, %f3073, 0f3DCCCCCD; @%p1729 bra $L__BB106_1996; bra.uni $L__BB106_1983; $L__BB106_1996: mul.ftz.f32 %f3099, %f3075, %f3075; fma.rn.ftz.f32 %f16495, %f3099, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3101, %f3099, %f3099; mov.u64 %rd2663, 64; cvt.rn.f32.u64 %f12069, %rd2663; div.approx.ftz.f32 %f12070, %f3101, %f12069; add.ftz.f32 %f3102, %f16495, %f12070; setp.eq.ftz.f32 %p1739, %f16495, %f3102; @%p1739 bra $L__BB106_2005; mul.ftz.f32 %f3103, %f3099, %f3101; mov.u64 %rd2664, 2304; cvt.rn.f32.u64 %f12071, %rd2664; div.approx.ftz.f32 %f12072, %f3103, %f12071; add.ftz.f32 %f3104, %f3102, %f12072; setp.eq.ftz.f32 %p1740, %f3102, %f3104; mov.f32 %f16495, %f3102; @%p1740 bra $L__BB106_2005; mul.ftz.f32 %f3105, %f3099, %f3103; mov.u64 %rd2665, 147456; cvt.rn.f32.u64 %f12073, %rd2665; div.approx.ftz.f32 %f12074, %f3105, %f12073; add.ftz.f32 %f3106, %f3104, %f12074; setp.eq.ftz.f32 %p1741, %f3104, %f3106; mov.f32 %f16495, %f3104; @%p1741 bra $L__BB106_2005; mul.ftz.f32 %f3107, %f3099, %f3105; mov.u64 %rd2666, 14745600; cvt.rn.f32.u64 %f12075, %rd2666; div.approx.ftz.f32 %f12076, %f3107, %f12075; add.ftz.f32 %f3108, %f3106, %f12076; setp.eq.ftz.f32 %p1742, %f3106, %f3108; mov.f32 %f16495, %f3106; @%p1742 bra $L__BB106_2005; mul.ftz.f32 %f3109, %f3099, %f3107; mov.u64 %rd2667, 2123366400; cvt.rn.f32.u64 %f12077, %rd2667; div.approx.ftz.f32 %f12078, %f3109, %f12077; add.ftz.f32 %f3110, %f3108, %f12078; setp.eq.ftz.f32 %p1743, %f3108, %f3110; mov.f32 %f16495, %f3108; @%p1743 bra $L__BB106_2005; mul.ftz.f32 %f3111, %f3099, %f3109; mov.u64 %rd2668, 416179814400; cvt.rn.f32.u64 %f12079, %rd2668; div.approx.ftz.f32 %f12080, %f3111, %f12079; add.ftz.f32 %f3112, %f3110, %f12080; setp.eq.ftz.f32 %p1744, %f3110, %f3112; mov.f32 %f16495, %f3110; @%p1744 bra $L__BB106_2005; mul.ftz.f32 %f3113, %f3099, %f3111; mov.u64 %rd2669, 106542032486400; cvt.rn.f32.u64 %f12081, %rd2669; div.approx.ftz.f32 %f12082, %f3113, %f12081; add.ftz.f32 %f3114, %f3112, %f12082; setp.eq.ftz.f32 %p1745, %f3112, %f3114; mov.f32 %f16495, %f3112; @%p1745 bra $L__BB106_2005; mul.ftz.f32 %f12083, %f3099, %f3113; mov.u64 %rd2670, 34519618525593600; cvt.rn.f32.u64 %f12084, %rd2670; div.approx.ftz.f32 %f12085, %f12083, %f12084; add.ftz.f32 %f3115, %f3114, %f12085; setp.eq.ftz.f32 %p1746, %f3114, %f3115; mov.f32 %f16495, %f3114; @%p1746 bra $L__BB106_2005; mov.f32 %f16495, %f3115; $L__BB106_2005: mul.ftz.f32 %f12086, %f3076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f12087, %f12086; mul.ftz.f32 %f3117, %f16495, %f12087; abs.ftz.f32 %f3118, %f3074; setp.ltu.ftz.f32 %p1747, %f3118, 0f3F800000; @%p1747 bra $L__BB106_2007; bra.uni $L__BB106_2006; $L__BB106_2007: mul.ftz.f32 %f12109, %f3074, %f3074; mov.f32 %f12110, 0f394FFF49; mov.f32 %f12111, 0f363D0ADA; fma.rn.ftz.f32 %f12112, %f12111, %f12109, %f12110; mov.f32 %f12113, 0f3C08889A; fma.rn.ftz.f32 %f12114, %f12112, %f12109, %f12113; mov.f32 %f12115, 0f3E2AAAAB; fma.rn.ftz.f32 %f12116, %f12114, %f12109, %f12115; mul.ftz.f32 %f12117, %f12109, %f12116; fma.rn.ftz.f32 %f16496, %f12117, %f3074, %f3074; bra.uni $L__BB106_2008; $L__BB106_1983: setp.gt.ftz.f32 %p1730, %f3075, 0f41400000; @%p1730 bra $L__BB106_1994; bra.uni $L__BB106_1984; $L__BB106_1994: mul.ftz.f32 %f12057, %f3075, 0f41000000; rcp.approx.ftz.f32 %f12058, %f12057; add.ftz.f32 %f12059, %f12058, 0fBFEB3F8E; lg2.approx.ftz.f32 %f12060, %f3075; fma.rn.ftz.f32 %f12061, %f12060, 0fBF317218, %f12059; fma.rn.ftz.f32 %f16494, %f12061, 0f3F000000, %f3075; bra.uni $L__BB106_1995; $L__BB106_2006: mov.f32 %f12088, 0f3FB8AA3B; mul.rn.ftz.f32 %f12089, %f3118, %f12088; cvt.rzi.ftz.f32.f32 %f12090, %f12089; abs.ftz.f32 %f12091, %f12090; setp.gt.ftz.f32 %p1748, %f12091, 0f42FC0000; mov.b32 %r2757, %f12090; and.b32 %r2758, %r2757, -2147483648; or.b32 %r2759, %r2758, 1123811328; mov.b32 %f12092, %r2759; selp.f32 %f12093, %f12092, %f12090, %p1748; mov.f32 %f12094, 0fBF317218; fma.rn.ftz.f32 %f12095, %f12093, %f12094, %f3118; mov.f32 %f12096, 0f3102E308; fma.rn.ftz.f32 %f12097, %f12093, %f12096, %f12095; mul.ftz.f32 %f12098, %f12097, 0f3FB8AA3B; add.ftz.f32 %f12099, %f12093, 0f4B40007D; mov.b32 %r2760, %f12099; shl.b32 %r2761, %r2760, 23; mov.b32 %f12100, %r2761; ex2.approx.ftz.f32 %f12101, %f12098; mul.ftz.f32 %f12102, %f12101, %f12100; mov.f32 %f12103, 0f3E000000; div.approx.ftz.f32 %f12104, %f12103, %f12102; neg.ftz.f32 %f12105, %f12104; mov.f32 %f12106, 0f40000000; fma.rn.ftz.f32 %f12107, %f12106, %f12102, %f12105; setp.ge.ftz.f32 %p1749, %f3118, 0f42B40000; selp.f32 %f12108, 0f7F800000, %f12107, %p1749; mov.b32 %r2762, %f12108; mov.b32 %r2763, %f3074; and.b32 %r2764, %r2763, -2147483648; or.b32 %r2765, %r2764, %r2762; mov.b32 %f16496, %r2765; $L__BB106_2008: add.ftz.f32 %f12118, %f16496, %f16496; mul.ftz.f32 %f12119, %f3073, %f12118; div.approx.ftz.f32 %f16497, %f3117, %f12119; bra.uni $L__BB106_2009; $L__BB106_1984: mul.ftz.f32 %f3077, %f3075, %f3075; fma.rn.ftz.f32 %f16493, %f3077, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3079, %f3077, %f3077; mov.u64 %rd2655, 64; cvt.rn.f32.u64 %f12039, %rd2655; div.approx.ftz.f32 %f12040, %f3079, %f12039; add.ftz.f32 %f3080, %f16493, %f12040; setp.eq.ftz.f32 %p1731, %f16493, %f3080; @%p1731 bra $L__BB106_1993; mul.ftz.f32 %f3081, %f3077, %f3079; mov.u64 %rd2656, 2304; cvt.rn.f32.u64 %f12041, %rd2656; div.approx.ftz.f32 %f12042, %f3081, %f12041; add.ftz.f32 %f3082, %f3080, %f12042; setp.eq.ftz.f32 %p1732, %f3080, %f3082; mov.f32 %f16493, %f3080; @%p1732 bra $L__BB106_1993; mul.ftz.f32 %f3083, %f3077, %f3081; mov.u64 %rd2657, 147456; cvt.rn.f32.u64 %f12043, %rd2657; div.approx.ftz.f32 %f12044, %f3083, %f12043; add.ftz.f32 %f3084, %f3082, %f12044; setp.eq.ftz.f32 %p1733, %f3082, %f3084; mov.f32 %f16493, %f3082; @%p1733 bra $L__BB106_1993; mul.ftz.f32 %f3085, %f3077, %f3083; mov.u64 %rd2658, 14745600; cvt.rn.f32.u64 %f12045, %rd2658; div.approx.ftz.f32 %f12046, %f3085, %f12045; add.ftz.f32 %f3086, %f3084, %f12046; setp.eq.ftz.f32 %p1734, %f3084, %f3086; mov.f32 %f16493, %f3084; @%p1734 bra $L__BB106_1993; mul.ftz.f32 %f3087, %f3077, %f3085; mov.u64 %rd2659, 2123366400; cvt.rn.f32.u64 %f12047, %rd2659; div.approx.ftz.f32 %f12048, %f3087, %f12047; add.ftz.f32 %f3088, %f3086, %f12048; setp.eq.ftz.f32 %p1735, %f3086, %f3088; mov.f32 %f16493, %f3086; @%p1735 bra $L__BB106_1993; mul.ftz.f32 %f3089, %f3077, %f3087; mov.u64 %rd2660, 416179814400; cvt.rn.f32.u64 %f12049, %rd2660; div.approx.ftz.f32 %f12050, %f3089, %f12049; add.ftz.f32 %f3090, %f3088, %f12050; setp.eq.ftz.f32 %p1736, %f3088, %f3090; mov.f32 %f16493, %f3088; @%p1736 bra $L__BB106_1993; mul.ftz.f32 %f3091, %f3077, %f3089; mov.u64 %rd2661, 106542032486400; cvt.rn.f32.u64 %f12051, %rd2661; div.approx.ftz.f32 %f12052, %f3091, %f12051; add.ftz.f32 %f3092, %f3090, %f12052; setp.eq.ftz.f32 %p1737, %f3090, %f3092; mov.f32 %f16493, %f3090; @%p1737 bra $L__BB106_1993; mul.ftz.f32 %f12053, %f3077, %f3091; mov.u64 %rd2662, 34519618525593600; cvt.rn.f32.u64 %f12054, %rd2662; div.approx.ftz.f32 %f12055, %f12053, %f12054; add.ftz.f32 %f3093, %f3092, %f12055; setp.eq.ftz.f32 %p1738, %f3092, %f3093; mov.f32 %f16493, %f3092; @%p1738 bra $L__BB106_1993; mov.f32 %f16493, %f3093; $L__BB106_1993: lg2.approx.ftz.f32 %f12056, %f16493; mul.ftz.f32 %f16494, %f12056, 0f3F317218; $L__BB106_1995: sub.ftz.f32 %f12062, %f16494, %f3076; sub.ftz.f32 %f12063, %f12062, %f3074; add.ftz.f32 %f12064, %f12063, 0f3F316F00; mul.ftz.f32 %f12065, %f3074, 0f3F000000; lg2.approx.ftz.f32 %f12066, %f12065; fma.rn.ftz.f32 %f12067, %f12066, 0f3F317218, %f12064; mul.ftz.f32 %f12068, %f12067, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16497, %f12068; $L__BB106_2009: mul.ftz.f32 %f12121, %f2887, 0f40800000; sub.ftz.f32 %f3124, %f12121, %f2938; mov.f32 %f16499, 0f00000000; mov.f32 %f16498, %f16499; @%p1704 bra $L__BB106_2011; ex2.approx.ftz.f32 %f12122, %f2955; add.ftz.f32 %f12123, %f12122, 0f3F800000; rcp.approx.ftz.f32 %f12124, %f12123; add.ftz.f32 %f16498, %f12124, %f12124; $L__BB106_2011: add.ftz.f32 %f12126, %f3124, 0f40C90FDB; mov.f32 %f12127, 0f40C90FDB; sub.ftz.f32 %f12128, %f16478, %f12126; add.ftz.f32 %f12129, %f12128, 0f40490FDB; div.approx.ftz.f32 %f12130, %f12129, %f12127; cvt.rmi.ftz.f32.f32 %f12131, %f12130; mul.ftz.f32 %f12132, %f12131, 0f40C90FDB; sub.ftz.f32 %f12133, %f12129, %f12132; add.ftz.f32 %f12134, %f12133, 0fC0490FDB; abs.ftz.f32 %f12135, %f12134; neg.ftz.f32 %f12136, %f12135; div.approx.ftz.f32 %f12137, %f12136, %f16477; mul.ftz.f32 %f12138, %f12137, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12139, %f12138; add.ftz.f32 %f12140, %f12139, 0f3F800000; mul.ftz.f32 %f12142, %f12140, %f12140; mul.ftz.f32 %f12143, %f16477, %f12142; div.approx.ftz.f32 %f3127, %f12139, %f12143; sub.ftz.f32 %f3128, %f11696, %f16498; setp.eq.ftz.f32 %p1751, %f3128, 0f00000000; @%p1751 bra $L__BB106_2013; div.approx.ftz.f32 %f16499, %f3127, %f3128; $L__BB106_2013: mul.ftz.f32 %f12144, %f16497, %f2903; mul.ftz.f32 %f12145, %f16497, %f2905; mul.ftz.f32 %f12146, %f16497, %f2907; fma.rn.ftz.f32 %f3131, %f16499, %f12144, %f3069; fma.rn.ftz.f32 %f3132, %f16499, %f12145, %f3070; fma.rn.ftz.f32 %f3133, %f16499, %f12146, %f3071; mul.ftz.f32 %f12147, %f16497, %f2926; fma.rn.ftz.f32 %f3134, %f16499, %f12147, %f3072; add.ftz.f32 %f3135, %f16467, %f16478; mul.ftz.f32 %f12148, %f2879, %f16474; mul.ftz.f32 %f3136, %f12148, %f3074; mul.ftz.f32 %f12149, %f2874, %f16473; mul.ftz.f32 %f3137, %f12149, %f3074; @%p1729 bra $L__BB106_2027; bra.uni $L__BB106_2014; $L__BB106_2027: mul.ftz.f32 %f3160, %f3136, %f3136; fma.rn.ftz.f32 %f16502, %f3160, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3162, %f3160, %f3160; mov.u64 %rd2679, 64; cvt.rn.f32.u64 %f12180, %rd2679; div.approx.ftz.f32 %f12181, %f3162, %f12180; add.ftz.f32 %f3163, %f16502, %f12181; setp.eq.ftz.f32 %p1762, %f16502, %f3163; @%p1762 bra $L__BB106_2036; mul.ftz.f32 %f3164, %f3160, %f3162; mov.u64 %rd2680, 2304; cvt.rn.f32.u64 %f12182, %rd2680; div.approx.ftz.f32 %f12183, %f3164, %f12182; add.ftz.f32 %f3165, %f3163, %f12183; setp.eq.ftz.f32 %p1763, %f3163, %f3165; mov.f32 %f16502, %f3163; @%p1763 bra $L__BB106_2036; mul.ftz.f32 %f3166, %f3160, %f3164; mov.u64 %rd2681, 147456; cvt.rn.f32.u64 %f12184, %rd2681; div.approx.ftz.f32 %f12185, %f3166, %f12184; add.ftz.f32 %f3167, %f3165, %f12185; setp.eq.ftz.f32 %p1764, %f3165, %f3167; mov.f32 %f16502, %f3165; @%p1764 bra $L__BB106_2036; mul.ftz.f32 %f3168, %f3160, %f3166; mov.u64 %rd2682, 14745600; cvt.rn.f32.u64 %f12186, %rd2682; div.approx.ftz.f32 %f12187, %f3168, %f12186; add.ftz.f32 %f3169, %f3167, %f12187; setp.eq.ftz.f32 %p1765, %f3167, %f3169; mov.f32 %f16502, %f3167; @%p1765 bra $L__BB106_2036; mul.ftz.f32 %f3170, %f3160, %f3168; mov.u64 %rd2683, 2123366400; cvt.rn.f32.u64 %f12188, %rd2683; div.approx.ftz.f32 %f12189, %f3170, %f12188; add.ftz.f32 %f3171, %f3169, %f12189; setp.eq.ftz.f32 %p1766, %f3169, %f3171; mov.f32 %f16502, %f3169; @%p1766 bra $L__BB106_2036; mul.ftz.f32 %f3172, %f3160, %f3170; mov.u64 %rd2684, 416179814400; cvt.rn.f32.u64 %f12190, %rd2684; div.approx.ftz.f32 %f12191, %f3172, %f12190; add.ftz.f32 %f3173, %f3171, %f12191; setp.eq.ftz.f32 %p1767, %f3171, %f3173; mov.f32 %f16502, %f3171; @%p1767 bra $L__BB106_2036; mul.ftz.f32 %f3174, %f3160, %f3172; mov.u64 %rd2685, 106542032486400; cvt.rn.f32.u64 %f12192, %rd2685; div.approx.ftz.f32 %f12193, %f3174, %f12192; add.ftz.f32 %f3175, %f3173, %f12193; setp.eq.ftz.f32 %p1768, %f3173, %f3175; mov.f32 %f16502, %f3173; @%p1768 bra $L__BB106_2036; mul.ftz.f32 %f12194, %f3160, %f3174; mov.u64 %rd2686, 34519618525593600; cvt.rn.f32.u64 %f12195, %rd2686; div.approx.ftz.f32 %f12196, %f12194, %f12195; add.ftz.f32 %f3176, %f3175, %f12196; setp.eq.ftz.f32 %p1769, %f3175, %f3176; mov.f32 %f16502, %f3175; @%p1769 bra $L__BB106_2036; mov.f32 %f16502, %f3176; $L__BB106_2036: mul.ftz.f32 %f12197, %f3137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f12198, %f12197; mul.ftz.f32 %f3178, %f16502, %f12198; abs.ftz.f32 %f3179, %f3074; setp.ltu.ftz.f32 %p1770, %f3179, 0f3F800000; @%p1770 bra $L__BB106_2038; bra.uni $L__BB106_2037; $L__BB106_2038: mul.ftz.f32 %f12220, %f3074, %f3074; mov.f32 %f12221, 0f394FFF49; mov.f32 %f12222, 0f363D0ADA; fma.rn.ftz.f32 %f12223, %f12222, %f12220, %f12221; mov.f32 %f12224, 0f3C08889A; fma.rn.ftz.f32 %f12225, %f12223, %f12220, %f12224; mov.f32 %f12226, 0f3E2AAAAB; fma.rn.ftz.f32 %f12227, %f12225, %f12220, %f12226; mul.ftz.f32 %f12228, %f12220, %f12227; fma.rn.ftz.f32 %f16503, %f12228, %f3074, %f3074; bra.uni $L__BB106_2039; $L__BB106_2014: setp.gt.ftz.f32 %p1753, %f3136, 0f41400000; @%p1753 bra $L__BB106_2025; bra.uni $L__BB106_2015; $L__BB106_2025: mul.ftz.f32 %f12168, %f3136, 0f41000000; rcp.approx.ftz.f32 %f12169, %f12168; add.ftz.f32 %f12170, %f12169, 0fBFEB3F8E; lg2.approx.ftz.f32 %f12171, %f3136; fma.rn.ftz.f32 %f12172, %f12171, 0fBF317218, %f12170; fma.rn.ftz.f32 %f16501, %f12172, 0f3F000000, %f3136; bra.uni $L__BB106_2026; $L__BB106_2037: mov.f32 %f12199, 0f3FB8AA3B; mul.rn.ftz.f32 %f12200, %f3179, %f12199; cvt.rzi.ftz.f32.f32 %f12201, %f12200; abs.ftz.f32 %f12202, %f12201; setp.gt.ftz.f32 %p1771, %f12202, 0f42FC0000; mov.b32 %r2766, %f12201; and.b32 %r2767, %r2766, -2147483648; or.b32 %r2768, %r2767, 1123811328; mov.b32 %f12203, %r2768; selp.f32 %f12204, %f12203, %f12201, %p1771; mov.f32 %f12205, 0fBF317218; fma.rn.ftz.f32 %f12206, %f12204, %f12205, %f3179; mov.f32 %f12207, 0f3102E308; fma.rn.ftz.f32 %f12208, %f12204, %f12207, %f12206; mul.ftz.f32 %f12209, %f12208, 0f3FB8AA3B; add.ftz.f32 %f12210, %f12204, 0f4B40007D; mov.b32 %r2769, %f12210; shl.b32 %r2770, %r2769, 23; mov.b32 %f12211, %r2770; ex2.approx.ftz.f32 %f12212, %f12209; mul.ftz.f32 %f12213, %f12212, %f12211; mov.f32 %f12214, 0f3E000000; div.approx.ftz.f32 %f12215, %f12214, %f12213; neg.ftz.f32 %f12216, %f12215; mov.f32 %f12217, 0f40000000; fma.rn.ftz.f32 %f12218, %f12217, %f12213, %f12216; setp.ge.ftz.f32 %p1772, %f3179, 0f42B40000; selp.f32 %f12219, 0f7F800000, %f12218, %p1772; mov.b32 %r2771, %f12219; mov.b32 %r2772, %f3074; and.b32 %r2773, %r2772, -2147483648; or.b32 %r2774, %r2773, %r2771; mov.b32 %f16503, %r2774; $L__BB106_2039: add.ftz.f32 %f12229, %f16503, %f16503; mul.ftz.f32 %f12230, %f3073, %f12229; div.approx.ftz.f32 %f16504, %f3178, %f12230; bra.uni $L__BB106_2040; $L__BB106_2015: mul.ftz.f32 %f3138, %f3136, %f3136; fma.rn.ftz.f32 %f16500, %f3138, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3140, %f3138, %f3138; mov.u64 %rd2671, 64; cvt.rn.f32.u64 %f12150, %rd2671; div.approx.ftz.f32 %f12151, %f3140, %f12150; add.ftz.f32 %f3141, %f16500, %f12151; setp.eq.ftz.f32 %p1754, %f16500, %f3141; @%p1754 bra $L__BB106_2024; mul.ftz.f32 %f3142, %f3138, %f3140; mov.u64 %rd2672, 2304; cvt.rn.f32.u64 %f12152, %rd2672; div.approx.ftz.f32 %f12153, %f3142, %f12152; add.ftz.f32 %f3143, %f3141, %f12153; setp.eq.ftz.f32 %p1755, %f3141, %f3143; mov.f32 %f16500, %f3141; @%p1755 bra $L__BB106_2024; mul.ftz.f32 %f3144, %f3138, %f3142; mov.u64 %rd2673, 147456; cvt.rn.f32.u64 %f12154, %rd2673; div.approx.ftz.f32 %f12155, %f3144, %f12154; add.ftz.f32 %f3145, %f3143, %f12155; setp.eq.ftz.f32 %p1756, %f3143, %f3145; mov.f32 %f16500, %f3143; @%p1756 bra $L__BB106_2024; mul.ftz.f32 %f3146, %f3138, %f3144; mov.u64 %rd2674, 14745600; cvt.rn.f32.u64 %f12156, %rd2674; div.approx.ftz.f32 %f12157, %f3146, %f12156; add.ftz.f32 %f3147, %f3145, %f12157; setp.eq.ftz.f32 %p1757, %f3145, %f3147; mov.f32 %f16500, %f3145; @%p1757 bra $L__BB106_2024; mul.ftz.f32 %f3148, %f3138, %f3146; mov.u64 %rd2675, 2123366400; cvt.rn.f32.u64 %f12158, %rd2675; div.approx.ftz.f32 %f12159, %f3148, %f12158; add.ftz.f32 %f3149, %f3147, %f12159; setp.eq.ftz.f32 %p1758, %f3147, %f3149; mov.f32 %f16500, %f3147; @%p1758 bra $L__BB106_2024; mul.ftz.f32 %f3150, %f3138, %f3148; mov.u64 %rd2676, 416179814400; cvt.rn.f32.u64 %f12160, %rd2676; div.approx.ftz.f32 %f12161, %f3150, %f12160; add.ftz.f32 %f3151, %f3149, %f12161; setp.eq.ftz.f32 %p1759, %f3149, %f3151; mov.f32 %f16500, %f3149; @%p1759 bra $L__BB106_2024; mul.ftz.f32 %f3152, %f3138, %f3150; mov.u64 %rd2677, 106542032486400; cvt.rn.f32.u64 %f12162, %rd2677; div.approx.ftz.f32 %f12163, %f3152, %f12162; add.ftz.f32 %f3153, %f3151, %f12163; setp.eq.ftz.f32 %p1760, %f3151, %f3153; mov.f32 %f16500, %f3151; @%p1760 bra $L__BB106_2024; mul.ftz.f32 %f12164, %f3138, %f3152; mov.u64 %rd2678, 34519618525593600; cvt.rn.f32.u64 %f12165, %rd2678; div.approx.ftz.f32 %f12166, %f12164, %f12165; add.ftz.f32 %f3154, %f3153, %f12166; setp.eq.ftz.f32 %p1761, %f3153, %f3154; mov.f32 %f16500, %f3153; @%p1761 bra $L__BB106_2024; mov.f32 %f16500, %f3154; $L__BB106_2024: lg2.approx.ftz.f32 %f12167, %f16500; mul.ftz.f32 %f16501, %f12167, 0f3F317218; $L__BB106_2026: sub.ftz.f32 %f12173, %f16501, %f3137; sub.ftz.f32 %f12174, %f12173, %f3074; add.ftz.f32 %f12175, %f12174, 0f3F316F00; mul.ftz.f32 %f12176, %f3074, 0f3F000000; lg2.approx.ftz.f32 %f12177, %f12176; fma.rn.ftz.f32 %f12178, %f12177, 0f3F317218, %f12175; mul.ftz.f32 %f12179, %f12178, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16504, %f12179; $L__BB106_2040: mul.ftz.f32 %f12231, %f16504, %f2918; mul.ftz.f32 %f12232, %f16504, %f2919; mul.ftz.f32 %f12233, %f16504, %f2920; fma.rn.ftz.f32 %f16739, %f12231, 0f3E22F983, %f3131; fma.rn.ftz.f32 %f16740, %f12232, 0f3E22F983, %f3132; fma.rn.ftz.f32 %f16741, %f12233, 0f3E22F983, %f3133; mul.ftz.f32 %f12234, %f16504, %f2927; fma.rn.ftz.f32 %f16738, %f12234, 0f3E22F983, %f3134; cos.approx.ftz.f32 %f12235, %f3135; mul.ftz.f32 %f12236, %f11518, %f16474; mul.ftz.f32 %f12237, %f12236, %f12235; mul.ftz.f32 %f12238, %f11519, %f16474; mul.ftz.f32 %f12239, %f12238, %f12235; mul.ftz.f32 %f12240, %f11520, %f16474; mul.ftz.f32 %f12241, %f12240, %f12235; fma.rn.ftz.f32 %f12242, %f16461, %f16473, %f12237; fma.rn.ftz.f32 %f12243, %f16462, %f16473, %f12239; fma.rn.ftz.f32 %f12244, %f16463, %f16473, %f12241; sin.approx.ftz.f32 %f12245, %f3135; mul.ftz.f32 %f12246, %f16464, %f16474; mul.ftz.f32 %f12247, %f16465, %f16474; mul.ftz.f32 %f12248, %f16466, %f16474; fma.rn.ftz.f32 %f16735, %f12246, %f12245, %f12242; fma.rn.ftz.f32 %f16736, %f12247, %f12245, %f12243; fma.rn.ftz.f32 %f16737, %f12248, %f12245, %f12244; setp.eq.s32 %p1773, %r3278, 0; selp.b32 %r3285, 10, 9, %p1773; bra.uni $L__BB106_2309; $L__BB106_1648: ld.const.u64 %rd3026, [%rd443+-24]; $L__BB106_1649: add.s32 %r3242, %r3242, 1; mul.wide.s32 %rd2186, %r3242, 16; add.s64 %rd2188, %rd2172, %rd2186; ld.const.v2.u64 {%rd2189, %rd2190}, [%rd2188+3744]; cvta.to.global.u64 %rd2191, %rd2189; add.s64 %rd2193, %rd2191, %rd2169; ld.global.u32 %r2286, [%rd2193]; cvta.to.global.u64 %rd2194, %rd2190; add.s64 %rd2195, %rd2194, %rd2169; ld.global.u32 %r2287, [%rd2195]; cvta.to.global.u64 %rd2196, %rd3026; add.s64 %rd2197, %rd2196, %rd2169; st.global.u32 [%rd2197], %r2286; cvta.to.global.u64 %rd2198, %rd3023; add.s64 %rd2199, %rd2198, %rd2169; st.global.u32 [%rd2199], %r2287; setp.eq.s32 %p1469, %r2287, -1; mov.u64 %rd3023, %rd2190; mov.u64 %rd3026, %rd2189; @%p1469 bra $L__BB106_1656; bra.uni $L__BB106_1649; $L__BB106_2115: ld.local.f32 %f13145, [%rd587+52]; mul.ftz.f32 %f13146, %f3486, %f13145; ld.local.f32 %f13147, [%rd587+48]; mul.ftz.f32 %f13148, %f3487, %f13147; sub.ftz.f32 %f13149, %f13146, %f13148; ld.local.f32 %f13150, [%rd587+44]; mul.ftz.f32 %f13151, %f3487, %f13150; mul.ftz.f32 %f13152, %f3484, %f13145; sub.ftz.f32 %f13153, %f13151, %f13152; mul.ftz.f32 %f13154, %f3484, %f13147; mul.ftz.f32 %f13155, %f3486, %f13150; sub.ftz.f32 %f13156, %f13154, %f13155; mul.ftz.f32 %f13157, %f13153, %f13153; fma.rn.ftz.f32 %f13158, %f13149, %f13149, %f13157; fma.rn.ftz.f32 %f13159, %f13156, %f13156, %f13158; sqrt.approx.ftz.f32 %f13160, %f13159; div.approx.ftz.f32 %f16567, %f13149, %f13160; div.approx.ftz.f32 %f16568, %f13153, %f13160; div.approx.ftz.f32 %f16569, %f13156, %f13160; mul.ftz.f32 %f13161, %f3487, %f16568; mul.ftz.f32 %f13162, %f3486, %f16569; sub.ftz.f32 %f16570, %f13161, %f13162; mul.ftz.f32 %f13163, %f3484, %f16569; mul.ftz.f32 %f13164, %f3487, %f16567; sub.ftz.f32 %f16571, %f13163, %f13164; mul.ftz.f32 %f13165, %f3486, %f16567; mul.ftz.f32 %f13166, %f3484, %f16568; sub.ftz.f32 %f16572, %f13165, %f13166; $L__BB106_2117: add.ftz.f32 %f3511, %f3491, 0f3F800000; @%p1881 bra $L__BB106_2133; bra.uni $L__BB106_2118; $L__BB106_2133: mul.ftz.f32 %f16578, %f2799, 0f40C90FDB; lg2.approx.ftz.f32 %f13332, %f2775; rcp.approx.ftz.f32 %f13333, %f3511; mul.ftz.f32 %f13334, %f13333, %f13332; ex2.approx.ftz.f32 %f16577, %f13334; bra.uni $L__BB106_2134; $L__BB106_2146: setp.neu.ftz.f32 %p1909, %f3573, %f3575; setp.neu.ftz.f32 %p1910, %f3573, %f3576; or.pred %p1911, %p1909, %p1910; neg.ftz.f32 %f13468, %f3576; selp.f32 %f13469, %f13468, %f3576, %p1911; add.ftz.f32 %f13470, %f3573, %f13469; neg.ftz.f32 %f13471, %f3575; selp.f32 %f13472, %f3575, %f13471, %p1911; sub.ftz.f32 %f13473, %f13472, %f3573; sub.ftz.f32 %f13474, %f3576, %f3575; mul.ftz.f32 %f13475, %f13470, %f13470; fma.rn.ftz.f32 %f13476, %f13474, %f13474, %f13475; fma.rn.ftz.f32 %f13477, %f13473, %f13473, %f13476; sqrt.approx.ftz.f32 %f13478, %f13477; div.approx.ftz.f32 %f16587, %f13474, %f13478; div.approx.ftz.f32 %f16588, %f13470, %f13478; div.approx.ftz.f32 %f16589, %f13473, %f13478; mul.ftz.f32 %f13479, %f3575, %f16589; mul.ftz.f32 %f13480, %f3576, %f16588; sub.ftz.f32 %f16584, %f13479, %f13480; mul.ftz.f32 %f13481, %f3576, %f16587; mul.ftz.f32 %f13482, %f3573, %f16589; sub.ftz.f32 %f16585, %f13481, %f13482; mul.ftz.f32 %f13483, %f3573, %f16588; mul.ftz.f32 %f13484, %f3575, %f16587; sub.ftz.f32 %f16586, %f13483, %f13484; $L__BB106_2148: mul.ftz.f32 %f13508, %f13461, %f16588; fma.rn.ftz.f32 %f13509, %f3574, %f16587, %f13508; fma.rn.ftz.f32 %f3601, %f13462, %f16589, %f13509; mul.ftz.f32 %f13510, %f13461, %f16585; fma.rn.ftz.f32 %f13511, %f3574, %f16584, %f13510; fma.rn.ftz.f32 %f3602, %f13462, %f16586, %f13511; ld.local.u64 %rd2688, [%rd587+36]; ld.f32 %f3603, [%rd2688]; ld.f32 %f3604, [%rd2688+4]; ld.f32 %f3605, [%rd2688+8]; ld.local.f32 %f3606, [%rd587+28]; ld.f32 %f3607, [%rd2688+12]; ld.f32 %f3608, [%rd2688+16]; ld.f32 %f3609, [%rd2688+20]; neg.ftz.f32 %f16597, %f3601; neg.ftz.f32 %f16596, %f3602; setp.lt.ftz.f32 %p1912, %f3577, 0fBF7FF972; mov.f32 %f16610, 0f00000000; @%p1912 bra $L__BB106_2151; setp.gt.ftz.f32 %p1913, %f3577, 0f3F7FF972; mov.f32 %f16610, 0fBF7FF972; @%p1913 bra $L__BB106_2151; mul.ftz.f32 %f13513, %f3577, %f3577; mov.f32 %f13514, 0f33D6BF95; max.ftz.f32 %f13515, %f13513, %f13514; rcp.approx.ftz.f32 %f13516, %f13515; mul.ftz.f32 %f13517, %f13465, %f16597; mul.ftz.f32 %f13518, %f13466, %f16596; mul.ftz.f32 %f13519, %f13518, %f13518; fma.rn.ftz.f32 %f13520, %f13517, %f13517, %f13519; fma.rn.ftz.f32 %f13521, %f13520, %f13516, 0f3F800000; sqrt.approx.ftz.f32 %f13522, %f13521; neg.ftz.f32 %f13523, %f13522; setp.ge.ftz.f32 %p1914, %f3577, 0f80000000; selp.f32 %f13524, %f13523, %f13522, %p1914; add.ftz.f32 %f13525, %f13524, 0fBF800000; mul.ftz.f32 %f16610, %f13525, 0f3F000000; $L__BB106_2151: fma.rn.ftz.f32 %f3614, %f3606, %f3606, 0fBF800000; mov.f32 %f16741, 0f3F800000; abs.ftz.f32 %f3615, %f16741; fma.rn.ftz.f32 %f3616, %f3615, %f3615, %f3614; setp.leu.ftz.f32 %p1915, %f3616, 0f00000000; mov.f32 %f16591, %f16741; @%p1915 bra $L__BB106_2153; sqrt.approx.ftz.f32 %f13527, %f3616; sub.ftz.f32 %f13528, %f13527, %f3615; add.ftz.f32 %f13529, %f3615, %f13527; div.approx.ftz.f32 %f13530, %f13528, %f13529; fma.rn.ftz.f32 %f13531, %f3615, %f13529, 0fBF800000; fma.rn.ftz.f32 %f13532, %f3615, %f13528, 0f3F800000; div.approx.ftz.f32 %f13533, %f13531, %f13532; mul.ftz.f32 %f13534, %f13530, 0f3F000000; mul.ftz.f32 %f13535, %f13530, %f13534; fma.rn.ftz.f32 %f13536, %f13533, %f13533, 0f3F800000; mul.ftz.f32 %f16591, %f13535, %f13536; $L__BB106_2153: setp.ne.s32 %p13, %r873, 15; sub.ftz.f32 %f3619, %f16741, %f16591; rcp.approx.ftz.f32 %f3620, %f3619; ld.local.u32 %r3281, [%rd7+168]; neg.ftz.f32 %f16614, %f3577; mov.f32 %f16611, 0f00000000; mov.u32 %r3282, 0; mov.pred %p1916, -1; mov.pred %p1926, 0; mov.f32 %f16740, %f16741; mov.f32 %f16739, %f16741; mov.f32 %f16601, %f16741; $L__BB106_2154: mov.f32 %f3629, %f16601; mad.lo.s32 %r895, %r3281, 1103515245, 12345; cvt.rn.f32.u32 %f13542, %r895; mul.ftz.f32 %f3631, %f13542, 0f2F800000; setp.gt.ftz.f32 %p1917, %f16614, 0f3F7FF972; mov.pred %p2189, %p1916; @%p1917 bra $L__BB106_2161; setp.lt.ftz.f32 %p1918, %f16614, 0fBF7FF972; @%p1918 bra $L__BB106_2160; bra.uni $L__BB106_2156; $L__BB106_2160: mul.ftz.f32 %f16601, %f3629, %f3631; mov.pred %p2189, %p1926; bra.uni $L__BB106_2161; $L__BB106_2156: abs.ftz.f32 %f13543, %f16614; setp.ltu.ftz.f32 %p1920, %f13543, 0f38D1B717; mov.f32 %f16601, %f3629; mov.pred %p2189, %p1926; @%p1920 bra $L__BB106_2161; mov.f32 %f13544, 0f3F800000; sub.ftz.f32 %f13545, %f13544, %f16611; setp.lt.ftz.f32 %p1922, %f13545, %f3631; mov.f32 %f16601, %f3629; mov.pred %p2189, %p1916; @%p1922 bra $L__BB106_2161; setp.ge.ftz.f32 %p1924, %f16610, 0f00000000; mov.f32 %f16601, %f13544; mov.pred %p2189, %p1926; @%p1924 bra $L__BB106_2161; mov.f32 %f13547, 0f3F800000; sub.ftz.f32 %f13548, %f13547, %f3631; lg2.approx.ftz.f32 %f13549, %f13548; mov.f32 %f13550, 0fBF800000; div.approx.ftz.f32 %f13551, %f13550, %f16610; mul.ftz.f32 %f13552, %f13551, %f13549; ex2.approx.ftz.f32 %f13553, %f13552; mul.ftz.f32 %f16601, %f3629, %f13553; mov.pred %p2189, %p1926; $L__BB106_2161: @%p2189 bra $L__BB106_2184; mad.lo.s32 %r2835, %r895, 1103515245, 12345; cvt.rn.f32.u32 %f13554, %r2835; mul.ftz.f32 %f3635, %f13554, 0f2F800000; mad.lo.s32 %r3281, %r2835, 1103515245, 12345; cvt.rn.f32.u32 %f13555, %r3281; mul.ftz.f32 %f3636, %f13555, 0f2F800000; neg.ftz.f32 %f3637, %f16597; mul.ftz.f32 %f13556, %f13465, %f3637; neg.ftz.f32 %f3638, %f16596; mul.ftz.f32 %f13557, %f13466, %f3638; mul.ftz.f32 %f13558, %f13557, %f13557; fma.rn.ftz.f32 %f13559, %f13556, %f13556, %f13558; fma.rn.ftz.f32 %f13560, %f16614, %f16614, %f13559; sqrt.approx.ftz.f32 %f13561, %f13560; div.approx.ftz.f32 %f16604, %f13556, %f13561; div.approx.ftz.f32 %f16605, %f13557, %f13561; neg.ftz.f32 %f3641, %f16614; div.approx.ftz.f32 %f3642, %f3641, %f13561; setp.gt.ftz.f32 %p1927, %f3642, 0f3F7FF972; @%p1927 bra $L__BB106_2168; abs.ftz.f32 %f13562, %f3642; setp.lt.ftz.f32 %p1928, %f13562, 0f358637BD; @%p1928 bra $L__BB106_2168; bra.uni $L__BB106_2164; $L__BB106_2168: mov.f32 %f13606, 0f3F800000; sub.ftz.f32 %f13607, %f13606, %f3636; mov.f32 %f13608, 0f33D6BF95; max.ftz.f32 %f13609, %f13607, %f13608; div.approx.ftz.f32 %f13610, %f3636, %f13609; sqrt.approx.ftz.f32 %f13611, %f13610; mul.ftz.f32 %f13612, %f3635, 0f40C90FDB; cos.approx.ftz.f32 %f13613, %f13612; sin.approx.ftz.f32 %f13614, %f13612; mul.ftz.f32 %f16603, %f13611, %f13614; mul.ftz.f32 %f16602, %f13611, %f13613; $L__BB106_2169: mul.ftz.f32 %f13615, %f16605, %f16605; fma.rn.ftz.f32 %f13616, %f16604, %f16604, %f13615; add.ftz.f32 %f13617, %f13616, 0f00000000; sqrt.approx.ftz.f32 %f3660, %f13617; setp.neu.ftz.f32 %p1936, %f3660, 0f00000000; @%p1936 bra $L__BB106_2170; bra.uni $L__BB106_2171; $L__BB106_2170: rcp.approx.ftz.f32 %f13618, %f3660; mul.ftz.f32 %f16605, %f16605, %f13618; mul.ftz.f32 %f16604, %f16604, %f13618; $L__BB106_2171: mul.ftz.f32 %f13619, %f16602, %f16604; mul.ftz.f32 %f13620, %f16603, %f16605; sub.ftz.f32 %f13621, %f13619, %f13620; mul.ftz.f32 %f13622, %f13465, %f13621; mul.ftz.f32 %f13623, %f16603, %f16604; fma.rn.ftz.f32 %f13624, %f16602, %f16605, %f13623; mul.ftz.f32 %f13625, %f13466, %f13624; neg.ftz.f32 %f13626, %f13622; neg.ftz.f32 %f13627, %f13625; mul.ftz.f32 %f13628, %f13625, %f13625; fma.rn.ftz.f32 %f13629, %f13622, %f13622, %f13628; add.ftz.f32 %f13630, %f13629, 0f3F800000; mov.f32 %f16611, 0f3F800000; sqrt.approx.ftz.f32 %f13632, %f13630; div.approx.ftz.f32 %f3670, %f13626, %f13632; div.approx.ftz.f32 %f3671, %f13627, %f13632; div.approx.ftz.f32 %f3672, %f16611, %f13632; setp.ne.s32 %p1937, %r3282, 0; and.pred %p1938, %p1937, %p13; mul.ftz.f32 %f13633, %f3603, %f16739; selp.f32 %f3673, %f13633, %f16739, %p1938; mul.ftz.f32 %f13634, %f3604, %f16740; selp.f32 %f3674, %f13634, %f16740, %p1938; mul.ftz.f32 %f13635, %f3605, %f16741; selp.f32 %f3675, %f13635, %f16741, %p1938; mul.ftz.f32 %f13636, %f3671, %f3638; fma.rn.ftz.f32 %f13637, %f3670, %f3637, %f13636; fma.rn.ftz.f32 %f3676, %f3672, %f3641, %f13637; mov.f32 %f16739, %f3673; mov.f32 %f16740, %f3674; mov.f32 %f16741, %f3675; @%p13 bra $L__BB106_2176; abs.ftz.f32 %f3677, %f3676; fma.rn.ftz.f32 %f3678, %f3677, %f3677, %f3614; setp.leu.ftz.f32 %p1940, %f3678, 0f00000000; mov.f32 %f16606, %f3619; @%p1940 bra $L__BB106_2174; sqrt.approx.ftz.f32 %f13638, %f3678; sub.ftz.f32 %f13639, %f13638, %f3677; add.ftz.f32 %f13640, %f3677, %f13638; div.approx.ftz.f32 %f13641, %f13639, %f13640; fma.rn.ftz.f32 %f13642, %f3677, %f13640, 0fBF800000; fma.rn.ftz.f32 %f13643, %f3677, %f13639, 0f3F800000; div.approx.ftz.f32 %f13644, %f13642, %f13643; mul.ftz.f32 %f13645, %f13641, 0f3F000000; mul.ftz.f32 %f13646, %f13641, %f13645; fma.rn.ftz.f32 %f13647, %f13644, %f13644, 0f3F800000; mul.ftz.f32 %f13648, %f13646, %f13647; sub.ftz.f32 %f16606, %f13648, %f16591; $L__BB106_2174: mul.ftz.f32 %f13649, %f3620, %f16606; mov.f32 %f13650, 0f3F800000; sub.ftz.f32 %f13651, %f13650, %f13649; fma.rn.ftz.f32 %f16739, %f3607, %f13651, %f13649; fma.rn.ftz.f32 %f16740, %f3608, %f13651, %f13649; fma.rn.ftz.f32 %f16741, %f3609, %f13651, %f13649; setp.eq.s32 %p1941, %r3282, 0; @%p1941 bra $L__BB106_2176; mul.ftz.f32 %f16739, %f3673, %f16739; mul.ftz.f32 %f16740, %f3674, %f16740; mul.ftz.f32 %f16741, %f3675, %f16741; $L__BB106_2176: add.ftz.f32 %f13653, %f3670, %f3670; add.ftz.f32 %f13654, %f3671, %f3671; add.ftz.f32 %f13655, %f3672, %f3672; fma.rn.ftz.f32 %f16597, %f13653, %f3676, %f16597; fma.rn.ftz.f32 %f16596, %f13654, %f3676, %f16596; fma.rn.ftz.f32 %f16614, %f13655, %f3676, %f16614; setp.gt.ftz.f32 %p1942, %f16614, 0f3F7FF972; mov.f32 %f16610, 0f00000000; @%p1942 bra $L__BB106_2179; setp.lt.ftz.f32 %p1943, %f16614, 0fBF7FF972; mov.f32 %f16610, 0fBF7FF972; @%p1943 bra $L__BB106_2179; mul.ftz.f32 %f13657, %f16614, %f16614; mov.f32 %f13658, 0f33D6BF95; max.ftz.f32 %f13659, %f13657, %f13658; rcp.approx.ftz.f32 %f13660, %f13659; mul.ftz.f32 %f13661, %f13465, %f16597; mul.ftz.f32 %f13662, %f13466, %f16596; mul.ftz.f32 %f13663, %f13662, %f13662; fma.rn.ftz.f32 %f13664, %f13661, %f13661, %f13663; fma.rn.ftz.f32 %f13665, %f13664, %f13660, 0f3F800000; sqrt.approx.ftz.f32 %f13666, %f13665; neg.ftz.f32 %f13667, %f13666; setp.le.ftz.f32 %p1944, %f16614, 0f00000000; selp.f32 %f13668, %f13667, %f13666, %p1944; add.ftz.f32 %f13669, %f13668, 0fBF800000; mul.ftz.f32 %f16610, %f13669, 0f3F000000; $L__BB106_2179: @%p1942 bra $L__BB106_2182; setp.lt.ftz.f32 %p1946, %f16614, 0f3727C5AC; mov.f32 %f16611, 0f00000000; @%p1946 bra $L__BB106_2182; lg2.approx.ftz.f32 %f13672, %f16601; mul.ftz.f32 %f13673, %f16610, %f13672; ex2.approx.ftz.f32 %f16611, %f13673; $L__BB106_2182: add.s32 %r3282, %r3282, 1; setp.lt.u32 %p1947, %r3282, 10; @%p1947 bra $L__BB106_2154; bra.uni $L__BB106_2183; $L__BB106_2164: add.ftz.f32 %f13565, %f3642, 0f3F800000; mul.ftz.f32 %f3643, %f13565, 0f3F000000; setp.lt.ftz.f32 %p1929, %f3643, 0f38D1B717; mov.f32 %f16602, 0f00000000; mov.f32 %f16603, %f16602; @%p1929 bra $L__BB106_2169; add.ftz.f32 %f13568, %f3636, %f3636; mul.ftz.f32 %f13569, %f13568, %f3643; div.approx.ftz.f32 %f13570, %f13569, %f3642; add.ftz.f32 %f3644, %f13570, 0fBF800000; mul.ftz.f32 %f3645, %f3644, %f3644; add.ftz.f32 %f3646, %f3645, 0fBF800000; abs.ftz.f32 %f13571, %f3646; setp.lt.ftz.f32 %p1930, %f13571, 0f33D6BF95; mov.f32 %f16603, %f16602; @%p1930 bra $L__BB106_2169; mul.ftz.f32 %f13572, %f3642, %f3642; mov.f32 %f13573, 0f3F800000; sub.ftz.f32 %f13574, %f13573, %f13572; mov.f32 %f13575, 0f00000000; max.ftz.f32 %f13576, %f13574, %f13575; sqrt.approx.ftz.f32 %f13577, %f13576; div.approx.ftz.f32 %f13578, %f13577, %f3642; rcp.approx.ftz.f32 %f13579, %f3646; mul.ftz.f32 %f13580, %f13578, %f13578; mul.ftz.f32 %f13581, %f13580, %f13579; mul.ftz.f32 %f13582, %f13579, %f13581; sub.ftz.f32 %f13583, %f3645, %f13580; mul.ftz.f32 %f13584, %f13583, %f13579; sub.ftz.f32 %f13585, %f13582, %f13584; max.ftz.f32 %f13586, %f13585, %f13575; sqrt.approx.ftz.f32 %f13587, %f13586; mul.ftz.f32 %f13588, %f13578, %f13579; add.ftz.f32 %f13589, %f13588, %f13587; rcp.approx.ftz.f32 %f13590, %f13578; setp.gt.ftz.f32 %p1931, %f13589, %f13590; setp.lt.ftz.f32 %p1932, %f3644, 0f00000000; or.pred %p1933, %p1932, %p1931; sub.ftz.f32 %f13591, %f13588, %f13587; mov.f32 %f13592, 0f3F000000; sub.ftz.f32 %f13593, %f13592, %f3635; add.ftz.f32 %f13594, %f3635, 0fBF000000; setp.ge.ftz.f32 %p1934, %f3635, 0f3F000000; selp.f32 %f13595, %f13594, %f13593, %p1934; add.ftz.f32 %f13596, %f13595, %f13595; fma.rn.ftz.f32 %f13597, %f13596, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f13598, %f13596, %f13597, 0f3EED4413; mul.ftz.f32 %f13599, %f13596, %f13598; fma.rn.ftz.f32 %f13600, %f13596, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f13601, %f13596, %f13600, 0fBF800000; fma.rn.ftz.f32 %f13602, %f13596, %f13601, 0f3F191676; div.approx.ftz.f32 %f13603, %f13599, %f13602; selp.f32 %f16602, %f13591, %f13589, %p1933; fma.rn.ftz.f32 %f13604, %f16602, %f16602, 0f3F800000; sqrt.approx.ftz.f32 %f13605, %f13604; mul.ftz.f32 %f16603, %f13603, %f13605; setp.ltu.ftz.f32 %p1935, %f3635, 0f3F000000; @%p1935 bra $L__BB106_2167; bra.uni $L__BB106_2169; $L__BB106_2167: neg.ftz.f32 %f16603, %f16603; bra.uni $L__BB106_2169; $L__BB106_2184: st.local.u32 [%rd7+168], %r895; bra.uni $L__BB106_2185; $L__BB106_2183: st.local.u32 [%rd7+168], %r3281; mov.f32 %f16597, 0f00000000; mov.f32 %f16614, 0f3F800000; mov.f32 %f16596, %f16597; mov.f32 %f16739, %f16597; mov.f32 %f16740, %f16597; mov.f32 %f16741, %f16597; $L__BB106_2185: mul.ftz.f32 %f13681, %f16584, %f16596; fma.rn.ftz.f32 %f13682, %f16587, %f16597, %f13681; mul.ftz.f32 %f13683, %f16585, %f16596; fma.rn.ftz.f32 %f13684, %f16588, %f16597, %f13683; mul.ftz.f32 %f13685, %f16586, %f16596; fma.rn.ftz.f32 %f13686, %f16589, %f16597, %f13685; fma.rn.ftz.f32 %f16735, %f3573, %f16614, %f13682; fma.rn.ftz.f32 %f16736, %f3575, %f16614, %f13684; fma.rn.ftz.f32 %f16737, %f3576, %f16614, %f13686; mul.ftz.f32 %f13687, %f2807, %f16736; fma.rn.ftz.f32 %f13688, %f2806, %f16735, %f13687; fma.rn.ftz.f32 %f13689, %f2808, %f16737, %f13688; setp.le.ftz.f32 %p1948, %f13689, 0f00000000; mov.u32 %r3285, 0; mov.f32 %f16738, 0f00000000; @%p1948 bra $L__BB106_2309; ld.local.f32 %f3712, [%rd587+20]; add.ftz.f32 %f13690, %f3601, %f16597; add.ftz.f32 %f13691, %f3602, %f16596; mul.ftz.f32 %f13692, %f13691, %f13691; fma.rn.ftz.f32 %f13693, %f13690, %f13690, %f13692; add.ftz.f32 %f13694, %f3577, %f16614; fma.rn.ftz.f32 %f13695, %f13694, %f13694, %f13693; sqrt.approx.ftz.f32 %f13696, %f13695; div.approx.ftz.f32 %f3713, %f13690, %f13696; div.approx.ftz.f32 %f3714, %f13691, %f13696; div.approx.ftz.f32 %f13697, %f13694, %f13696; mul.ftz.f32 %f3715, %f13697, %f13697; @%p1908 bra $L__BB106_2191; bra.uni $L__BB106_2187; $L__BB106_2191: setp.gt.ftz.f32 %p1952, %f3577, 0f3F7FF972; neg.ftz.f32 %f13744, %f3713; div.approx.ftz.f32 %f13745, %f13744, %f3712; neg.ftz.f32 %f13746, %f3714; div.approx.ftz.f32 %f13747, %f13746, %f13466; fma.rn.ftz.f32 %f13748, %f13745, %f13745, %f3715; fma.rn.ftz.f32 %f3720, %f13747, %f13747, %f13748; mov.f32 %f16619, 0f00000000; @%p1952 bra $L__BB106_2194; mov.f32 %f16619, 0fBF7FF972; @%p1912 bra $L__BB106_2194; mul.ftz.f32 %f13750, %f3577, %f3577; mov.f32 %f13751, 0f33D6BF95; max.ftz.f32 %f13752, %f13750, %f13751; rcp.approx.ftz.f32 %f13753, %f13752; mul.ftz.f32 %f13754, %f3601, %f3712; mul.ftz.f32 %f13755, %f3602, %f13466; mul.ftz.f32 %f13756, %f13755, %f13755; fma.rn.ftz.f32 %f13757, %f13754, %f13754, %f13756; fma.rn.ftz.f32 %f13758, %f13757, %f13753, 0f3F800000; sqrt.approx.ftz.f32 %f13759, %f13758; add.ftz.f32 %f13760, %f13759, 0fBF800000; mul.ftz.f32 %f16619, %f13760, 0f3F000000; $L__BB106_2194: add.ftz.f32 %f13761, %f16619, 0f3F800000; mov.f32 %f13762, 0f3F800000; mul.ftz.f32 %f13763, %f3577, %f13761; mov.f32 %f13764, 0f33D6BF95; max.ftz.f32 %f13765, %f13763, %f13764; mul.ftz.f32 %f13766, %f3720, 0f40490FDB; mul.ftz.f32 %f13767, %f3720, %f13766; mul.ftz.f32 %f13768, %f3712, %f13767; mul.ftz.f32 %f13769, %f13466, %f13768; max.ftz.f32 %f13770, %f13769, %f13764; rcp.approx.ftz.f32 %f13771, %f13770; mul.ftz.f32 %f13772, %f13771, 0f3E800000; div.approx.ftz.f32 %f13773, %f13772, %f13765; mul.ftz.f32 %f13774, %f3712, %f13466; sqrt.approx.ftz.f32 %f13775, %f13774; mul.ftz.f32 %f13776, %f13775, 0fBFFE59F3; mul.ftz.f32 %f13777, %f13775, %f13776; mul.ftz.f32 %f13778, %f13777, 0f3FB8AA3B; ex2.approx.ftz.f32 %f13779, %f13778; fma.rn.ftz.f32 %f13780, %f13779, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f13781, %f13775, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f13782, %f13775, %f13781, 0f417C8B44; fma.rn.ftz.f32 %f13783, %f13775, %f13782, 0fC1665E35; fma.rn.ftz.f32 %f13784, %f13775, %f13783, 0f40CE7881; fma.rn.ftz.f32 %f13785, %f13775, %f13784, 0fBF993E81; fma.rn.ftz.f32 %f13786, %f13775, %f13785, 0f3CE3C31E; fma.rn.ftz.f32 %f13787, %f13775, %f13786, 0f3BBA5D49; sub.ftz.f32 %f13788, %f13780, %f13787; cvt.ftz.sat.f32.f32 %f13789, %f13788; sub.ftz.f32 %f13790, %f13762, %f13789; mul.ftz.f32 %f13791, %f16614, 0f3EA2F983; mul.ftz.f32 %f13792, %f13791, %f13790; fma.rn.ftz.f32 %f16620, %f13773, %f13789, %f13792; bra.uni $L__BB106_2195; $L__BB106_2118: setp.lt.ftz.f32 %p1886, %f2799, 0f3E800000; add.ftz.f32 %f13184, %f3492, 0f3F800000; div.approx.ftz.f32 %f13185, %f3511, %f13184; sqrt.approx.ftz.f32 %f3512, %f13185; @%p1886 bra $L__BB106_2130; bra.uni $L__BB106_2119; $L__BB106_2130: mul.ftz.f32 %f13298, %f2799, 0f40800000; mul.ftz.f32 %f13299, %f13298, 0f3FC90FDB; sin.approx.ftz.f32 %f13300, %f13299; cos.approx.ftz.f32 %f13301, %f13299; div.approx.ftz.f32 %f13302, %f13300, %f13301; mul.ftz.f32 %f3531, %f3512, %f13302; abs.ftz.f32 %f3532, %f3531; setp.leu.ftz.f32 %p1898, %f3532, 0f3F800000; mov.f32 %f16576, %f3532; @%p1898 bra $L__BB106_2132; rcp.approx.ftz.f32 %f16576, %f3532; $L__BB106_2132: mul.rn.ftz.f32 %f13303, %f16576, %f16576; mov.f32 %f13304, 0fC0B59883; mov.f32 %f13305, 0fBF52C7EA; fma.rn.ftz.f32 %f13306, %f13303, %f13305, %f13304; mov.f32 %f13307, 0fC0D21907; fma.rn.ftz.f32 %f13308, %f13306, %f13303, %f13307; mul.ftz.f32 %f13309, %f13303, %f13308; mul.ftz.f32 %f13310, %f16576, %f13309; add.ftz.f32 %f13311, %f13303, 0f41355DC0; mov.f32 %f13312, 0f41E6BD60; fma.rn.ftz.f32 %f13313, %f13311, %f13303, %f13312; mov.f32 %f13314, 0f419D92C8; fma.rn.ftz.f32 %f13315, %f13313, %f13303, %f13314; rcp.approx.ftz.f32 %f13316, %f13315; fma.rn.ftz.f32 %f13317, %f13310, %f13316, %f16576; mov.f32 %f13318, 0f3FC90FDB; sub.ftz.f32 %f13319, %f13318, %f13317; setp.gt.ftz.f32 %p1899, %f3532, 0f3F800000; selp.f32 %f13320, %f13319, %f13317, %p1899; mov.b32 %r2827, %f13320; mov.b32 %r2828, %f3531; and.b32 %r2829, %r2828, -2147483648; or.b32 %r2830, %r2829, %r2827; mov.b32 %f13321, %r2830; setp.le.ftz.f32 %p1900, %f3532, 0f7F800000; selp.f32 %f16578, %f13321, %f13320, %p1900; cos.approx.ftz.f32 %f13322, %f16578; sin.approx.ftz.f32 %f13323, %f16578; mul.ftz.f32 %f13324, %f3491, %f13322; mul.ftz.f32 %f13325, %f3492, %f13323; mul.ftz.f32 %f13326, %f13323, %f13325; fma.rn.ftz.f32 %f13327, %f13322, %f13324, %f13326; add.ftz.f32 %f13328, %f13327, 0f3F800000; rcp.approx.ftz.f32 %f13329, %f13328; lg2.approx.ftz.f32 %f13330, %f2775; mul.ftz.f32 %f13331, %f13330, %f13329; ex2.approx.ftz.f32 %f16577, %f13331; bra.uni $L__BB106_2134; $L__BB106_2260: add.s32 %r2876, %r873, -11; setp.gt.u32 %p2005, %r2876, 1; mov.u32 %r3285, 18; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p2005 bra $L__BB106_2281; ld.local.f32 %f14525, [%rd587+28]; fma.rn.ftz.f32 %f3949, %f14525, %f14525, 0fBF800000; abs.ftz.f32 %f3950, %f14492; fma.rn.ftz.f32 %f3951, %f3950, %f3950, %f3949; setp.leu.ftz.f32 %p2006, %f3951, 0f00000000; mov.f32 %f16701, %f14492; @%p2006 bra $L__BB106_2263; sqrt.approx.ftz.f32 %f14526, %f3951; sub.ftz.f32 %f14527, %f14526, %f3950; add.ftz.f32 %f14528, %f3950, %f14526; div.approx.ftz.f32 %f14529, %f14527, %f14528; fma.rn.ftz.f32 %f14530, %f3950, %f14528, 0fBF800000; fma.rn.ftz.f32 %f14531, %f3950, %f14527, 0f3F800000; div.approx.ftz.f32 %f14532, %f14530, %f14531; mul.ftz.f32 %f14533, %f14529, 0f3F000000; mul.ftz.f32 %f14534, %f14529, %f14533; fma.rn.ftz.f32 %f14535, %f14532, %f14532, 0f3F800000; mul.ftz.f32 %f16701, %f14534, %f14535; $L__BB106_2263: ld.local.u64 %rd2689, [%rd587+36]; ld.f32 %f3954, [%rd2689+12]; ld.f32 %f3955, [%rd2689+16]; ld.f32 %f3956, [%rd2689+20]; sub.ftz.f32 %f3957, %f14492, %f16701; mul.ftz.f32 %f14537, %f3942, %f16736; fma.rn.ftz.f32 %f14538, %f3941, %f16735, %f14537; fma.rn.ftz.f32 %f14539, %f3943, %f16737, %f14538; abs.ftz.f32 %f3958, %f14539; fma.rn.ftz.f32 %f3959, %f3958, %f3958, %f3949; setp.leu.ftz.f32 %p2007, %f3959, 0f00000000; mov.f32 %f16702, %f3957; @%p2007 bra $L__BB106_2265; sqrt.approx.ftz.f32 %f14540, %f3959; sub.ftz.f32 %f14541, %f14540, %f3958; add.ftz.f32 %f14542, %f3958, %f14540; div.approx.ftz.f32 %f14543, %f14541, %f14542; fma.rn.ftz.f32 %f14544, %f3958, %f14542, 0fBF800000; fma.rn.ftz.f32 %f14545, %f3958, %f14541, 0f3F800000; div.approx.ftz.f32 %f14546, %f14544, %f14545; mul.ftz.f32 %f14547, %f14543, 0f3F000000; mul.ftz.f32 %f14548, %f14543, %f14547; fma.rn.ftz.f32 %f14549, %f14546, %f14546, 0f3F800000; mul.ftz.f32 %f14550, %f14548, %f14549; sub.ftz.f32 %f16702, %f14550, %f16701; $L__BB106_2265: rcp.approx.ftz.f32 %f14552, %f3957; mov.f32 %f14553, 0f3F800000; mul.ftz.f32 %f14554, %f14552, %f16702; sub.ftz.f32 %f14555, %f14553, %f14554; fma.rn.ftz.f32 %f14556, %f3956, %f14555, %f14554; fma.rn.ftz.f32 %f14557, %f3955, %f14555, %f14554; fma.rn.ftz.f32 %f14558, %f3954, %f14555, %f14554; mul.ftz.f32 %f16739, %f14558, 0f49742400; mul.ftz.f32 %f16740, %f14557, 0f49742400; mul.ftz.f32 %f16741, %f14556, 0f49742400; bra.uni $L__BB106_2281; $L__BB106_2215: neg.ftz.f32 %f14017, %f3797; sqrt.approx.ftz.f32 %f14018, %f14017; add.ftz.f32 %f16643, %f14018, 0fC0400000; mov.f32 %f16651, 0fB951F09B; mov.f32 %f16650, 0f38D3B56B; mov.f32 %f16649, 0f3AB0DC72; mov.f32 %f16648, 0fBB70BDE7; mov.f32 %f16647, 0f3BBC127B; mov.f32 %f16646, 0fBBF9C5D7; mov.f32 %f16645, 0f3C1AA57E; mov.f32 %f16644, 0f3F8036DB; mov.f32 %f16642, 0f40354F7E; $L__BB106_2217: fma.rn.ftz.f32 %f14029, %f16643, %f16651, %f16650; fma.rn.ftz.f32 %f14030, %f16643, %f14029, %f16649; fma.rn.ftz.f32 %f14031, %f16643, %f14030, %f16648; fma.rn.ftz.f32 %f14032, %f16643, %f14031, %f16647; fma.rn.ftz.f32 %f14033, %f16643, %f14032, %f16646; fma.rn.ftz.f32 %f14034, %f16643, %f14033, %f16645; fma.rn.ftz.f32 %f14035, %f16643, %f14034, %f16644; fma.rn.ftz.f32 %f14036, %f16643, %f14035, %f16642; mul.ftz.f32 %f14037, %f3796, %f14036; mul.ftz.f32 %f14038, %f14037, %f14037; mul.ftz.f32 %f14039, %f14038, 0fBFB8AA3B; ex2.approx.ftz.f32 %f14040, %f14039; add.ftz.f32 %f14041, %f3796, 0f3F800000; fma.rn.ftz.f32 %f14043, %f3775, %f14040, %f14041; sub.ftz.f32 %f14044, %f14043, %f3777; mul.ftz.f32 %f14045, %f3771, %f14037; sub.ftz.f32 %f14046, %f13908, %f14045; div.approx.ftz.f32 %f14047, %f14044, %f14046; sub.ftz.f32 %f3810, %f3796, %f14047; abs.ftz.f32 %f14048, %f3810; setp.gt.ftz.f32 %p1970, %f14048, 0f3F7FFFFF; selp.f32 %f14049, 0f3F7FFFFF, %f14048, %p1970; sub.ftz.f32 %f14050, %f13908, %f14049; add.ftz.f32 %f14051, %f14049, 0f3F800000; mul.ftz.f32 %f14052, %f14050, %f14051; max.ftz.f32 %f14054, %f14052, %f13928; min.ftz.f32 %f14056, %f14054, %f13930; mov.b32 %r2855, %f14056; shr.u32 %r2856, %r2855, 23; add.s32 %r2857, %r2856, -127; and.b32 %r2858, %r2855, 8388607; or.b32 %r2859, %r2858, 1065353216; mov.b32 %f14057, %r2859; add.ftz.f32 %f14058, %f14057, 0fBF800000; mul.ftz.f32 %f14059, %f14058, %f14058; mul.ftz.f32 %f14060, %f14059, %f14059; fma.rn.ftz.f32 %f14061, %f14058, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f14062, %f14058, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14063, %f14058, %f14061, 0fBE0CD4FD; fma.rn.ftz.f32 %f14064, %f14058, %f14063, 0f3E77ADBD; fma.rn.ftz.f32 %f14065, %f14058, %f14064, 0fBEB1D206; fma.rn.ftz.f32 %f14066, %f14058, %f14062, 0f3FB8AA10; mul.ftz.f32 %f14067, %f14060, %f14065; fma.rn.ftz.f32 %f14068, %f14058, %f14066, %f14067; cvt.rn.f32.s32 %f14069, %r2857; add.ftz.f32 %f14070, %f14068, %f14069; mul.ftz.f32 %f3811, %f14070, 0f3F317218; setp.gt.ftz.f32 %p1971, %f3811, 0fC0A00000; @%p1971 bra $L__BB106_2219; bra.uni $L__BB106_2218; $L__BB106_2219: mov.f32 %f14091, 0fC0200000; sub.ftz.f32 %f16653, %f14091, %f3811; mov.f32 %f16661, 0f32F16588; mov.f32 %f16660, 0f34B84B36; mov.f32 %f16659, 0fB66C7357; mov.f32 %f16658, 0fB6935AC1; mov.f32 %f16657, 0f396532DB; mov.f32 %f16656, 0fBAA45408; mov.f32 %f16655, 0fBB88E4EF; mov.f32 %f16654, 0f3E7C8F63; mov.f32 %f16652, 0f3FC02E2F; bra.uni $L__BB106_2220; $L__BB106_2218: neg.ftz.f32 %f14080, %f3811; sqrt.approx.ftz.f32 %f14081, %f14080; add.ftz.f32 %f16653, %f14081, 0fC0400000; mov.f32 %f16661, 0fB951F09B; mov.f32 %f16660, 0f38D3B56B; mov.f32 %f16659, 0f3AB0DC72; mov.f32 %f16658, 0fBB70BDE7; mov.f32 %f16657, 0f3BBC127B; mov.f32 %f16656, 0fBBF9C5D7; mov.f32 %f16655, 0f3C1AA57E; mov.f32 %f16654, 0f3F8036DB; mov.f32 %f16652, 0f40354F7E; $L__BB106_2220: fma.rn.ftz.f32 %f14092, %f16653, %f16661, %f16660; fma.rn.ftz.f32 %f14093, %f16653, %f14092, %f16659; fma.rn.ftz.f32 %f14094, %f16653, %f14093, %f16658; fma.rn.ftz.f32 %f14095, %f16653, %f14094, %f16657; fma.rn.ftz.f32 %f14096, %f16653, %f14095, %f16656; fma.rn.ftz.f32 %f14097, %f16653, %f14096, %f16655; fma.rn.ftz.f32 %f14098, %f16653, %f14097, %f16654; fma.rn.ftz.f32 %f14099, %f16653, %f14098, %f16652; mul.ftz.f32 %f16675, %f3810, %f14099; $L__BB106_2221: fma.rn.ftz.f32 %f3826, %f2775, 0f40000000, 0fBF800000; abs.ftz.f32 %f14101, %f3826; setp.gt.ftz.f32 %p1972, %f14101, 0f3F7FFFFF; selp.f32 %f14102, 0f3F7FFFFF, %f14101, %p1972; sub.ftz.f32 %f14103, %f13908, %f14102; add.ftz.f32 %f14104, %f14102, 0f3F800000; mul.ftz.f32 %f14105, %f14103, %f14104; max.ftz.f32 %f14107, %f14105, %f13928; min.ftz.f32 %f14109, %f14107, %f13930; mov.b32 %r2860, %f14109; shr.u32 %r2861, %r2860, 23; add.s32 %r2862, %r2861, -127; and.b32 %r2863, %r2860, 8388607; or.b32 %r2864, %r2863, 1065353216; mov.b32 %f14110, %r2864; add.ftz.f32 %f14111, %f14110, 0fBF800000; mul.ftz.f32 %f14112, %f14111, %f14111; mul.ftz.f32 %f14113, %f14112, %f14112; fma.rn.ftz.f32 %f14114, %f14111, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f14115, %f14111, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14116, %f14111, %f14114, 0fBE0CD4FD; fma.rn.ftz.f32 %f14117, %f14111, %f14116, 0f3E77ADBD; fma.rn.ftz.f32 %f14118, %f14111, %f14117, 0fBEB1D206; fma.rn.ftz.f32 %f14119, %f14111, %f14115, 0f3FB8AA10; mul.ftz.f32 %f14120, %f14113, %f14118; fma.rn.ftz.f32 %f14121, %f14111, %f14119, %f14120; cvt.rn.f32.s32 %f14122, %r2862; add.ftz.f32 %f14123, %f14121, %f14122; mul.ftz.f32 %f3827, %f14123, 0f3F317218; setp.gt.ftz.f32 %p1973, %f3827, 0fC0A00000; @%p1973 bra $L__BB106_2223; bra.uni $L__BB106_2222; $L__BB106_2223: mov.f32 %f14144, 0fC0200000; sub.ftz.f32 %f16664, %f14144, %f3827; mov.f32 %f16672, 0f32F16588; mov.f32 %f16671, 0f34B84B36; mov.f32 %f16670, 0fB66C7357; mov.f32 %f16669, 0fB6935AC1; mov.f32 %f16668, 0f396532DB; mov.f32 %f16667, 0fBAA45408; mov.f32 %f16666, 0fBB88E4EF; mov.f32 %f16665, 0f3E7C8F63; mov.f32 %f16663, 0f3FC02E2F; bra.uni $L__BB106_2224; $L__BB106_2222: neg.ftz.f32 %f14133, %f3827; sqrt.approx.ftz.f32 %f14134, %f14133; add.ftz.f32 %f16664, %f14134, 0fC0400000; mov.f32 %f16672, 0fB951F09B; mov.f32 %f16671, 0f38D3B56B; mov.f32 %f16670, 0f3AB0DC72; mov.f32 %f16669, 0fBB70BDE7; mov.f32 %f16668, 0f3BBC127B; mov.f32 %f16667, 0fBBF9C5D7; mov.f32 %f16666, 0f3C1AA57E; mov.f32 %f16665, 0f3F8036DB; mov.f32 %f16663, 0f40354F7E; $L__BB106_2224: fma.rn.ftz.f32 %f14145, %f16664, %f16672, %f16671; fma.rn.ftz.f32 %f14146, %f16664, %f14145, %f16670; fma.rn.ftz.f32 %f14147, %f16664, %f14146, %f16669; fma.rn.ftz.f32 %f14148, %f16664, %f14147, %f16668; fma.rn.ftz.f32 %f14149, %f16664, %f14148, %f16667; fma.rn.ftz.f32 %f14150, %f16664, %f14149, %f16666; fma.rn.ftz.f32 %f14151, %f16664, %f14150, %f16665; fma.rn.ftz.f32 %f14152, %f16664, %f14151, %f16663; mul.ftz.f32 %f16674, %f3826, %f14152; add.ftz.f32 %f14153, %f3774, 0f3F800000; rcp.approx.ftz.f32 %f16673, %f14153; $L__BB106_2225: mul.ftz.f32 %f14154, %f16630, %f16674; mul.ftz.f32 %f14155, %f16629, %f16675; sub.ftz.f32 %f14156, %f14155, %f14154; mul.ftz.f32 %f14157, %f16630, %f16675; fma.rn.ftz.f32 %f14158, %f16629, %f16674, %f14157; mul.ftz.f32 %f14159, %f13811, %f14156; mul.ftz.f32 %f14160, %f13811, %f14158; neg.ftz.f32 %f14161, %f14159; neg.ftz.f32 %f14162, %f14160; mul.ftz.f32 %f14163, %f14160, %f14160; fma.rn.ftz.f32 %f14164, %f14159, %f14159, %f14163; add.ftz.f32 %f14165, %f14164, 0f3F800000; mov.f32 %f16677, 0f3F800000; sqrt.approx.ftz.f32 %f14167, %f14165; div.approx.ftz.f32 %f14168, %f14161, %f14167; div.approx.ftz.f32 %f14169, %f14162, %f14167; div.approx.ftz.f32 %f3845, %f16677, %f14167; mul.ftz.f32 %f14170, %f16621, %f14169; mul.ftz.f32 %f14171, %f16622, %f14169; mul.ftz.f32 %f14172, %f16623, %f14169; fma.rn.ftz.f32 %f14173, %f16624, %f14168, %f14170; fma.rn.ftz.f32 %f14174, %f16625, %f14168, %f14171; fma.rn.ftz.f32 %f14175, %f16626, %f14168, %f14172; fma.rn.ftz.f32 %f3846, %f3736, %f3845, %f14173; fma.rn.ftz.f32 %f3847, %f3738, %f3845, %f14174; fma.rn.ftz.f32 %f3848, %f3739, %f3845, %f14175; mul.ftz.f32 %f14176, %f13809, %f3847; fma.rn.ftz.f32 %f14177, %f3737, %f3846, %f14176; fma.rn.ftz.f32 %f3849, %f13810, %f3848, %f14177; setp.eq.s32 %p1974, %r873, 22; @%p1974 bra $L__BB106_2235; bra.uni $L__BB106_2226; $L__BB106_2235: setp.gt.ftz.f32 %p1981, %f3849, 0f00000000; ld.local.f32 %f3867, [%rd587+28]; @%p1981 bra $L__BB106_2237; bra.uni $L__BB106_2236; $L__BB106_2237: rcp.approx.ftz.f32 %f16682, %f3867; mov.u16 %rs158, 0; mov.f32 %f16678, %f3846; mov.f32 %f16679, %f3847; mov.f32 %f16680, %f3848; mov.f32 %f16681, %f3849; bra.uni $L__BB106_2238; $L__BB106_2226: setp.leu.ftz.f32 %p1975, %f3849, 0f00000000; mov.u32 %r2865, 10; mov.f32 %f16738, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; mov.u32 %r3285, %r2865; @%p1975 bra $L__BB106_2309; add.ftz.f32 %f14187, %f3849, %f3849; mul.ftz.f32 %f14188, %f3846, %f14187; mul.ftz.f32 %f14189, %f3847, %f14187; mul.ftz.f32 %f14190, %f3848, %f14187; sub.ftz.f32 %f16735, %f14188, %f3737; sub.ftz.f32 %f16736, %f14189, %f13809; sub.ftz.f32 %f16737, %f14190, %f13810; mul.ftz.f32 %f14191, %f2807, %f16736; fma.rn.ftz.f32 %f14192, %f2806, %f16735, %f14191; fma.rn.ftz.f32 %f14193, %f2808, %f16737, %f14192; setp.leu.ftz.f32 %p1976, %f14193, 0f00000000; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; mov.u32 %r3285, %r2865; @%p1976 bra $L__BB106_2309; mul.ftz.f32 %f3853, %f13811, %f13812; setp.le.ftz.f32 %p1977, %f3853, 0f33D6BF95; mov.u32 %r3285, 18; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1977 bra $L__BB106_2309; mul.ftz.f32 %f3854, %f3853, 0f40490FDB; mul.ftz.f32 %f14198, %f3738, %f16736; fma.rn.ftz.f32 %f14199, %f3736, %f16735, %f14198; fma.rn.ftz.f32 %f14200, %f3739, %f16737, %f14199; mul.ftz.f32 %f3855, %f14200, %f14200; sub.ftz.f32 %f3856, %f16677, %f3855; @%p1957 bra $L__BB106_2232; bra.uni $L__BB106_2230; $L__BB106_2232: mul.ftz.f32 %f14251, %f3845, %f3845; mul.ftz.f32 %f14252, %f14251, %f14251; rcp.approx.ftz.f32 %f14253, %f14251; add.ftz.f32 %f14254, %f14253, 0fBF800000; neg.ftz.f32 %f14255, %f14254; div.approx.ftz.f32 %f14256, %f14255, %f3853; mul.ftz.f32 %f14257, %f14256, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14258, %f14257; mul.ftz.f32 %f14259, %f3854, %f14252; div.approx.ftz.f32 %f16676, %f14258, %f14259; div.approx.ftz.f32 %f14260, %f3856, %f3855; mov.f32 %f14261, 0f00000000; max.ftz.f32 %f14262, %f14260, %f14261; sqrt.approx.ftz.f32 %f14263, %f14262; mul.ftz.f32 %f3861, %f13811, %f14263; setp.lt.ftz.f32 %p1980, %f3861, 0f3F200000; @%p1980 bra $L__BB106_2234; rcp.approx.ftz.f32 %f14264, %f3861; fma.rn.ftz.f32 %f14265, %f14264, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14266, %f14264, %f14265; fma.rn.ftz.f32 %f14267, %f14264, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14268, %f14264, %f14267, 0f3F800000; div.approx.ftz.f32 %f16677, %f14266, %f14268; bra.uni $L__BB106_2234; $L__BB106_2236: neg.ftz.f32 %f16681, %f3849; neg.ftz.f32 %f16678, %f3846; neg.ftz.f32 %f16679, %f3847; neg.ftz.f32 %f16680, %f3848; mov.u16 %rs158, 1; mov.f32 %f16682, %f3867; $L__BB106_2238: mul.ftz.f32 %f14276, %f16681, %f16681; mov.f32 %f16686, 0f3F800000; sub.ftz.f32 %f14277, %f16686, %f14276; mul.ftz.f32 %f14278, %f16682, %f16682; mul.ftz.f32 %f14279, %f14278, %f14277; sub.ftz.f32 %f3878, %f16686, %f14279; setp.lt.ftz.f32 %p1982, %f3878, 0f00000000; mov.f32 %f16738, 0f00000000; mov.f32 %f16735, %f16738; mov.f32 %f16736, %f16738; mov.f32 %f16737, %f16738; @%p1982 bra $L__BB106_2240; sqrt.approx.ftz.f32 %f14280, %f3878; mov.f32 %f14281, 0f33D6BF95; max.ftz.f32 %f14282, %f14280, %f14281; mul.ftz.f32 %f14283, %f16681, %f16682; sub.ftz.f32 %f14284, %f14283, %f14282; mul.ftz.f32 %f14285, %f16678, %f14284; mul.ftz.f32 %f14286, %f16679, %f14284; mul.ftz.f32 %f14287, %f16680, %f14284; mul.ftz.f32 %f14288, %f3737, %f16682; sub.ftz.f32 %f16735, %f14285, %f14288; mul.ftz.f32 %f14289, %f13809, %f16682; sub.ftz.f32 %f16736, %f14286, %f14289; mul.ftz.f32 %f14290, %f13810, %f16682; sub.ftz.f32 %f16737, %f14287, %f14290; mul.ftz.f32 %f14291, %f16679, %f16736; fma.rn.ftz.f32 %f14292, %f16678, %f16735, %f14291; fma.rn.ftz.f32 %f14293, %f16680, %f16737, %f14292; mul.ftz.f32 %f14294, %f3867, %f14293; add.ftz.f32 %f14295, %f16681, %f14294; sub.ftz.f32 %f14296, %f16681, %f14294; div.approx.ftz.f32 %f14297, %f14295, %f14296; mul.ftz.f32 %f14298, %f3867, %f16681; add.ftz.f32 %f14299, %f14298, %f14293; sub.ftz.f32 %f14300, %f14298, %f14293; div.approx.ftz.f32 %f14301, %f14299, %f14300; mul.ftz.f32 %f14302, %f14301, %f14301; fma.rn.ftz.f32 %f14303, %f14297, %f14297, %f14302; mul.ftz.f32 %f16686, %f14303, 0f3F000000; $L__BB106_2240: setp.eq.ftz.f32 %p1983, %f16686, 0f3F800000; setp.ne.s16 %p1984, %rs158, 0; mov.u32 %r2869, 9; or.pred %p1985, %p1984, %p1983; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; mov.u32 %r3285, %r2869; @%p1985 bra $L__BB106_2309; mul.ftz.f32 %f3887, %f13811, %f13812; setp.le.ftz.f32 %p1986, %f3887, 0f33D6BF95; mov.u32 %r3285, 17; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1986 bra $L__BB106_2309; add.ftz.f32 %f14317, %f3867, 0fBF800000; abs.ftz.f32 %f14318, %f14317; setp.lt.ftz.f32 %p1987, %f14318, 0f38D1B717; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1987 bra $L__BB106_2309; mul.ftz.f32 %f14320, %f3845, %f3845; mul.ftz.f32 %f14321, %f14320, %f14320; rcp.approx.ftz.f32 %f14322, %f14320; mov.f32 %f16687, 0f3F800000; add.ftz.f32 %f14323, %f14322, 0fBF800000; neg.ftz.f32 %f14324, %f14323; div.approx.ftz.f32 %f14325, %f14324, %f3887; mul.ftz.f32 %f14326, %f14325, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14327, %f14326; mul.ftz.f32 %f14328, %f3887, 0f40490FDB; mul.ftz.f32 %f14329, %f14328, %f14321; div.approx.ftz.f32 %f3888, %f14327, %f14329; mul.ftz.f32 %f14330, %f3738, %f16736; fma.rn.ftz.f32 %f14331, %f3736, %f16735, %f14330; fma.rn.ftz.f32 %f14332, %f3739, %f16737, %f14331; mul.ftz.f32 %f14333, %f14332, %f14332; sub.ftz.f32 %f14334, %f16687, %f14333; div.approx.ftz.f32 %f14335, %f14334, %f14333; mov.f32 %f14336, 0f00000000; max.ftz.f32 %f14337, %f14335, %f14336; sqrt.approx.ftz.f32 %f14338, %f14337; mul.ftz.f32 %f3889, %f13811, %f14338; setp.lt.ftz.f32 %p1988, %f3889, 0f3F200000; @%p1988 bra $L__BB106_2245; rcp.approx.ftz.f32 %f14339, %f3889; fma.rn.ftz.f32 %f14340, %f14339, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14341, %f14339, %f14340; fma.rn.ftz.f32 %f14342, %f14339, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14343, %f14339, %f14342, 0f3F800000; div.approx.ftz.f32 %f16687, %f14341, %f14343; $L__BB106_2245: mul.ftz.f32 %f14344, %f3847, %f16736; fma.rn.ftz.f32 %f14345, %f3846, %f16735, %f14344; fma.rn.ftz.f32 %f14346, %f3848, %f16737, %f14345; fma.rn.ftz.f32 %f14347, %f3867, %f14346, %f3849; mul.ftz.f32 %f14348, %f14347, %f14347; mul.ftz.f32 %f14349, %f3867, %f3867; mul.ftz.f32 %f14350, %f14349, %f3888; mul.ftz.f32 %f14351, %f3740, %f14348; div.approx.ftz.f32 %f14352, %f14350, %f14351; mul.ftz.f32 %f14353, %f3849, %f14346; abs.ftz.f32 %f14354, %f14353; mul.ftz.f32 %f14355, %f16673, %f16687; mul.ftz.f32 %f14356, %f14355, %f14354; mul.ftz.f32 %f16739, %f14352, %f14356; abs.ftz.f32 %f14357, %f14346; mul.ftz.f32 %f14358, %f16673, %f3849; mul.ftz.f32 %f14359, %f14358, %f14357; mul.ftz.f32 %f16738, %f14352, %f14359; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; mov.u32 %r3285, %r2869; bra.uni $L__BB106_2309; $L__BB106_740: ld.const.u64 %rd1216, [kernel_params+2176]; cvta.to.global.u64 %rd1217, %rd1216; ld.global.u32 %rd1218, [%rd163+168]; ld.global.u32 %r1416, [%rd163+152]; shl.b32 %r1417, %r1416, 1; cvt.rn.f32.s32 %f6482, %r1417; mul.ftz.f32 %f6483, %f6482, 0f3F000000; cvt.rzi.ftz.s32.f32 %r1418, %f6483; add.s32 %r1419, %r1417, -1; min.s32 %r1420, %r1418, %r1419; cvt.rn.f32.s32 %f6484, %r1420; sub.ftz.f32 %f1010, %f6483, %f6484; cvt.s64.s32 %rd1219, %r1420; add.s64 %rd1220, %rd1219, %rd1218; shl.b64 %rd1221, %rd1220, 6; add.s64 %rd164, %rd1217, %rd1221; ld.global.v4.f32 {%f6485, %f6486, %f6487, %f6488}, [%rd164]; ld.global.v4.f32 {%f6489, %f6490, %f6491, %f6492}, [%rd164+64]; mul.ftz.f32 %f6493, %f6486, %f6490; fma.rn.ftz.f32 %f6494, %f6485, %f6489, %f6493; mul.ftz.f32 %f6495, %f6488, %f6492; fma.rn.ftz.f32 %f6496, %f6487, %f6491, %f6495; add.ftz.f32 %f1019, %f6494, %f6496; setp.gt.ftz.f32 %p632, %f1019, 0f3F7FDF3B; sub.ftz.f32 %f1020, %f6432, %f1010; @%p632 bra $L__BB106_742; bra.uni $L__BB106_741; $L__BB106_742: mul.ftz.f32 %f6551, %f6485, %f1020; fma.rn.ftz.f32 %f6552, %f6489, %f1010, %f6551; mul.ftz.f32 %f6553, %f6486, %f1020; fma.rn.ftz.f32 %f6554, %f6490, %f1010, %f6553; mul.ftz.f32 %f6555, %f6487, %f1020; fma.rn.ftz.f32 %f6556, %f6491, %f1010, %f6555; mul.ftz.f32 %f6557, %f6488, %f1020; fma.rn.ftz.f32 %f6558, %f6492, %f1010, %f6557; mul.ftz.f32 %f6559, %f6554, %f6554; fma.rn.ftz.f32 %f6560, %f6552, %f6552, %f6559; mul.ftz.f32 %f6561, %f6558, %f6558; fma.rn.ftz.f32 %f6562, %f6556, %f6556, %f6561; add.ftz.f32 %f6563, %f6560, %f6562; rsqrt.approx.ftz.f32 %f6564, %f6563; mul.ftz.f32 %f15856, %f6558, %f6564; mul.ftz.f32 %f15855, %f6556, %f6564; mul.ftz.f32 %f15854, %f6554, %f6564; mul.ftz.f32 %f15853, %f6552, %f6564; bra.uni $L__BB106_743; $L__BB106_723: ld.const.u64 %rd1194, [kernel_params+2176]; cvta.to.global.u64 %rd1195, %rd1194; ld.global.u32 %rd1196, [%rd161+168]; ld.global.u32 %r1406, [%rd161+152]; shl.b32 %r1407, %r1406, 1; cvt.rn.f32.s32 %f6132, %r1407; mul.ftz.f32 %f6133, %f6132, 0f3F000000; cvt.rzi.ftz.s32.f32 %r1408, %f6133; add.s32 %r1409, %r1407, -1; min.s32 %r1410, %r1408, %r1409; cvt.rn.f32.s32 %f6134, %r1410; sub.ftz.f32 %f868, %f6133, %f6134; cvt.s64.s32 %rd1197, %r1410; add.s64 %rd1198, %rd1197, %rd1196; shl.b64 %rd1199, %rd1198, 6; add.s64 %rd162, %rd1195, %rd1199; ld.global.v4.f32 {%f6135, %f6136, %f6137, %f6138}, [%rd162]; ld.global.v4.f32 {%f6139, %f6140, %f6141, %f6142}, [%rd162+64]; mul.ftz.f32 %f6143, %f6136, %f6140; fma.rn.ftz.f32 %f6144, %f6135, %f6139, %f6143; mul.ftz.f32 %f6145, %f6138, %f6142; fma.rn.ftz.f32 %f6146, %f6137, %f6141, %f6145; add.ftz.f32 %f877, %f6144, %f6146; setp.gt.ftz.f32 %p616, %f877, 0f3F7FDF3B; mov.f32 %f6147, 0f3F800000; sub.ftz.f32 %f878, %f6147, %f868; @%p616 bra $L__BB106_725; bra.uni $L__BB106_724; $L__BB106_725: mul.ftz.f32 %f6201, %f6135, %f878; fma.rn.ftz.f32 %f6202, %f6139, %f868, %f6201; mul.ftz.f32 %f6203, %f6136, %f878; fma.rn.ftz.f32 %f6204, %f6140, %f868, %f6203; mul.ftz.f32 %f6205, %f6137, %f878; fma.rn.ftz.f32 %f6206, %f6141, %f868, %f6205; mul.ftz.f32 %f6207, %f6138, %f878; fma.rn.ftz.f32 %f6208, %f6142, %f868, %f6207; mul.ftz.f32 %f6209, %f6204, %f6204; fma.rn.ftz.f32 %f6210, %f6202, %f6202, %f6209; mul.ftz.f32 %f6211, %f6208, %f6208; fma.rn.ftz.f32 %f6212, %f6206, %f6206, %f6211; add.ftz.f32 %f6213, %f6210, %f6212; rsqrt.approx.ftz.f32 %f6214, %f6213; mul.ftz.f32 %f15822, %f6208, %f6214; mul.ftz.f32 %f15821, %f6206, %f6214; mul.ftz.f32 %f15820, %f6204, %f6214; mul.ftz.f32 %f15819, %f6202, %f6214; bra.uni $L__BB106_726; $L__BB106_2298: and.b32 %r2896, %r903, 4; setp.eq.s32 %p2050, %r2896, 0; add.ftz.f32 %f15063, %f4094, 0f00000000; selp.f32 %f16732, 0f00000000, %f15063, %p2050; $L__BB106_2299: and.b32 %r2897, %r903, 9; setp.eq.s32 %p2051, %r2897, 0; @%p2051 bra $L__BB106_2301; mul.ftz.f32 %f15064, %f14996, %f16736; fma.rn.ftz.f32 %f15065, %f4081, %f16735, %f15064; fma.rn.ftz.f32 %f15066, %f14997, %f16737, %f15065; add.ftz.f32 %f15067, %f15066, 0f3F800000; ld.local.f32 %f15068, [%rd587+20]; mul.ftz.f32 %f15069, %f15067, %f15068; add.ftz.f32 %f15070, %f15069, 0fBF800000; mul.ftz.f32 %f15071, %f4092, %f4093; add.ftz.f32 %f15072, %f4092, %f4093; fma.rn.ftz.f32 %f15073, %f15071, %f15070, %f15072; fma.rn.ftz.f32 %f16732, %f15069, %f15073, %f16732; $L__BB106_2301: mul.ftz.f32 %f15074, %f4091, 0f3EA2F983; mul.ftz.f32 %f16739, %f15074, %f16732; $L__BB106_2302: mul.ftz.f32 %f16738, %f4084, 0f3EA2F983; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_2187: setp.gt.ftz.f32 %p1950, %f3577, 0f3F7FF972; mul.ftz.f32 %f13699, %f3712, %f3712; mov.f32 %f13700, 0f3F800000; sub.ftz.f32 %f13701, %f13700, %f3715; fma.rn.ftz.f32 %f13702, %f13699, %f3715, %f13701; mul.ftz.f32 %f13703, %f13702, 0f40490FDB; mul.ftz.f32 %f13704, %f13702, %f13703; mov.f32 %f13705, 0f33D6BF95; max.ftz.f32 %f13706, %f13704, %f13705; div.approx.ftz.f32 %f3716, %f13699, %f13706; mov.f32 %f16618, 0f00000000; @%p1950 bra $L__BB106_2190; mov.f32 %f16618, 0fBF7FF972; @%p1912 bra $L__BB106_2190; mul.ftz.f32 %f13708, %f3577, %f3577; max.ftz.f32 %f13710, %f13708, %f13705; rcp.approx.ftz.f32 %f13711, %f13710; mul.ftz.f32 %f13712, %f3601, %f3712; mul.ftz.f32 %f13713, %f3602, %f3712; mul.ftz.f32 %f13714, %f13713, %f13713; fma.rn.ftz.f32 %f13715, %f13712, %f13712, %f13714; fma.rn.ftz.f32 %f13716, %f13715, %f13711, 0f3F800000; sqrt.approx.ftz.f32 %f13717, %f13716; add.ftz.f32 %f13718, %f13717, 0fBF800000; mul.ftz.f32 %f16618, %f13718, 0f3F000000; $L__BB106_2190: add.ftz.f32 %f13719, %f16618, 0f3F800000; mul.ftz.f32 %f13721, %f3577, %f13719; max.ftz.f32 %f13723, %f13721, %f13705; mul.ftz.f32 %f13724, %f3716, 0f3E800000; div.approx.ftz.f32 %f13725, %f13724, %f13723; mul.ftz.f32 %f13726, %f3712, 0fBFFE59F3; mul.ftz.f32 %f13727, %f3712, %f13726; mul.ftz.f32 %f13728, %f13727, 0f3FB8AA3B; ex2.approx.ftz.f32 %f13729, %f13728; fma.rn.ftz.f32 %f13730, %f13729, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f13731, %f3712, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f13732, %f3712, %f13731, 0f417C8B44; fma.rn.ftz.f32 %f13733, %f3712, %f13732, 0fC1665E35; fma.rn.ftz.f32 %f13734, %f3712, %f13733, 0f40CE7881; fma.rn.ftz.f32 %f13735, %f3712, %f13734, 0fBF993E81; fma.rn.ftz.f32 %f13736, %f3712, %f13735, 0f3CE3C31E; fma.rn.ftz.f32 %f13737, %f3712, %f13736, 0f3BBA5D49; sub.ftz.f32 %f13738, %f13730, %f13737; cvt.ftz.sat.f32.f32 %f13739, %f13738; sub.ftz.f32 %f13740, %f13700, %f13739; mul.ftz.f32 %f13741, %f16614, 0f3EA2F983; mul.ftz.f32 %f13742, %f13741, %f13740; fma.rn.ftz.f32 %f16620, %f13725, %f13739, %f13742; $L__BB106_2195: mov.f32 %f13793, 0f00000000; max.ftz.f32 %f16738, %f13793, %f16620; mul.ftz.f32 %f16739, %f16739, %f16738; mul.ftz.f32 %f16740, %f16740, %f16738; mul.ftz.f32 %f16741, %f16741, %f16738; mov.u32 %r3285, 10; bra.uni $L__BB106_2309; $L__BB106_2119: setp.lt.ftz.f32 %p1887, %f2799, 0f3F000000; @%p1887 bra $L__BB106_2127; bra.uni $L__BB106_2120; $L__BB106_2127: mov.f32 %f13260, 0f3F000000; sub.ftz.f32 %f13261, %f13260, %f2799; mul.ftz.f32 %f13262, %f13261, 0f40800000; mul.ftz.f32 %f13263, %f13262, 0f3FC90FDB; sin.approx.ftz.f32 %f13264, %f13263; cos.approx.ftz.f32 %f13265, %f13263; div.approx.ftz.f32 %f13266, %f13264, %f13265; mul.ftz.f32 %f3525, %f3512, %f13266; abs.ftz.f32 %f3526, %f3525; setp.leu.ftz.f32 %p1895, %f3526, 0f3F800000; mov.f32 %f16575, %f3526; @%p1895 bra $L__BB106_2129; rcp.approx.ftz.f32 %f16575, %f3526; $L__BB106_2129: mul.rn.ftz.f32 %f13267, %f16575, %f16575; mov.f32 %f13268, 0fC0B59883; mov.f32 %f13269, 0fBF52C7EA; fma.rn.ftz.f32 %f13270, %f13267, %f13269, %f13268; mov.f32 %f13271, 0fC0D21907; fma.rn.ftz.f32 %f13272, %f13270, %f13267, %f13271; mul.ftz.f32 %f13273, %f13267, %f13272; mul.ftz.f32 %f13274, %f16575, %f13273; add.ftz.f32 %f13275, %f13267, 0f41355DC0; mov.f32 %f13276, 0f41E6BD60; fma.rn.ftz.f32 %f13277, %f13275, %f13267, %f13276; mov.f32 %f13278, 0f419D92C8; fma.rn.ftz.f32 %f13279, %f13277, %f13267, %f13278; rcp.approx.ftz.f32 %f13280, %f13279; fma.rn.ftz.f32 %f13281, %f13274, %f13280, %f16575; mov.f32 %f13282, 0f3FC90FDB; sub.ftz.f32 %f13283, %f13282, %f13281; setp.gt.ftz.f32 %p1896, %f3526, 0f3F800000; selp.f32 %f13284, %f13283, %f13281, %p1896; mov.b32 %r2823, %f13284; mov.b32 %r2824, %f3525; and.b32 %r2825, %r2824, -2147483648; or.b32 %r2826, %r2825, %r2823; mov.b32 %f13285, %r2826; setp.le.ftz.f32 %p1897, %f3526, 0f7F800000; selp.f32 %f13286, %f13285, %f13284, %p1897; cos.approx.ftz.f32 %f13287, %f13286; sin.approx.ftz.f32 %f13288, %f13286; mul.ftz.f32 %f13289, %f3491, %f13287; mul.ftz.f32 %f13290, %f3492, %f13288; mul.ftz.f32 %f13291, %f13288, %f13290; fma.rn.ftz.f32 %f13292, %f13287, %f13289, %f13291; add.ftz.f32 %f13293, %f13292, 0f3F800000; rcp.approx.ftz.f32 %f13294, %f13293; lg2.approx.ftz.f32 %f13295, %f2775; mul.ftz.f32 %f13296, %f13295, %f13294; ex2.approx.ftz.f32 %f16577, %f13296; mov.f32 %f13297, 0f40490FDB; sub.ftz.f32 %f16578, %f13297, %f13286; bra.uni $L__BB106_2134; $L__BB106_2100: mov.f32 %f12745, 0f3F800000; div.approx.ftz.f32 %f16564, %f3439, %f12745; $L__BB106_2102: add.ftz.f32 %f12748, %f16564, 0fBF800000; add.ftz.f32 %f12749, %f12748, 0f3F800000; mul.ftz.f32 %f12750, %f12749, %f12749; fma.rn.ftz.f32 %f12751, %f12750, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f12752, %f12749, %f12751; fma.rn.ftz.f32 %f12753, %f12750, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f12754, %f12750, %f12753, 0f3F800000; div.approx.ftz.f32 %f12755, %f12752, %f12754; sub.ftz.f32 %f12757, %f12736, %f12755; setp.gt.ftz.f32 %p1845, %f3439, 0f3F800000; selp.f32 %f12758, %f12757, %f12755, %p1845; mov.b32 %r2778, %f3438; and.b32 %r2779, %r2778, -2147483648; mov.b32 %r2780, %f12758; and.b32 %r2781, %r2780, 2147483647; or.b32 %r891, %r2781, %r2779; add.ftz.f32 %f12759, %f3434, 0fBFC90FDB; mul.ftz.f32 %f12760, %f12759, 0f3F000000; sub.ftz.f32 %f12761, %f12760, %f3437; mul.ftz.f32 %f3443, %f3436, %f12761; abs.ftz.f32 %f3444, %f3443; setp.eq.ftz.f32 %p1846, %f3444, 0f00000000; @%p1846 bra $L__BB106_2107; setp.eq.ftz.f32 %p1847, %f3444, 0f3F800000; mov.f32 %f16565, 0f3F800000; @%p1847 bra $L__BB106_2107; setp.gt.ftz.f32 %p1848, %f3444, 0f3F800000; @%p1848 bra $L__BB106_2106; bra.uni $L__BB106_2105; $L__BB106_2106: mov.f32 %f12764, 0f3F800000; div.approx.ftz.f32 %f16565, %f12764, %f3444; bra.uni $L__BB106_2107; $L__BB106_1877: mov.f32 %f11348, 0f3F800000; div.approx.ftz.f32 %f16458, %f2824, %f11348; $L__BB106_1879: add.ftz.f32 %f11351, %f16458, 0fBF800000; add.ftz.f32 %f11352, %f11351, 0f3F800000; mul.ftz.f32 %f11353, %f11352, %f11352; fma.rn.ftz.f32 %f11354, %f11353, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11355, %f11352, %f11354; fma.rn.ftz.f32 %f11356, %f11353, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11357, %f11353, %f11356, 0f3F800000; div.approx.ftz.f32 %f11358, %f11355, %f11357; sub.ftz.f32 %f11360, %f11338, %f11358; setp.gt.ftz.f32 %p1635, %f2824, 0f3F800000; selp.f32 %f11361, %f11360, %f11358, %p1635; mov.b32 %r2693, %f2823; and.b32 %r2694, %r2693, -2147483648; mov.b32 %r2695, %f11361; and.b32 %r2696, %r2695, 2147483647; or.b32 %r874, %r2696, %r2694; add.ftz.f32 %f11362, %f2819, 0fBFC90FDB; div.approx.ftz.f32 %f11364, %f11362, %f11344; sub.ftz.f32 %f11365, %f11364, %f2822; mul.ftz.f32 %f2828, %f2821, %f11365; abs.ftz.f32 %f2829, %f2828; setp.eq.ftz.f32 %p1636, %f2829, 0f00000000; mov.f32 %f16459, %f16460; @%p1636 bra $L__BB106_1884; setp.eq.ftz.f32 %p1637, %f2829, 0f3F800000; mov.f32 %f16459, 0f3F800000; @%p1637 bra $L__BB106_1884; setp.gt.ftz.f32 %p1638, %f2829, 0f3F800000; @%p1638 bra $L__BB106_1883; bra.uni $L__BB106_1882; $L__BB106_1883: mov.f32 %f11368, 0f3F800000; div.approx.ftz.f32 %f16459, %f11368, %f2829; bra.uni $L__BB106_1884; $L__BB106_2105: mov.f32 %f12763, 0f3F800000; div.approx.ftz.f32 %f16565, %f3444, %f12763; $L__BB106_2107: add.ftz.f32 %f12765, %f16565, 0fBF800000; mov.f32 %f12766, 0fBF800000; mov.f32 %f12767, 0f3F800000; add.ftz.f32 %f12768, %f12765, 0f3F800000; mul.ftz.f32 %f12769, %f12768, %f12768; fma.rn.ftz.f32 %f12770, %f12769, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f12771, %f12768, %f12770; fma.rn.ftz.f32 %f12772, %f12769, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f12773, %f12769, %f12772, 0f3F800000; div.approx.ftz.f32 %f12774, %f12771, %f12773; sub.ftz.f32 %f12776, %f12736, %f12774; setp.gt.ftz.f32 %p1849, %f3444, 0f3F800000; selp.f32 %f12777, %f12776, %f12774, %p1849; mov.b32 %r2783, %f3443; and.b32 %r2784, %r2783, -2147483648; mov.b32 %r2785, %f12777; and.b32 %r2786, %r2785, 2147483647; or.b32 %r2787, %r2786, %r2784; mov.b32 %f12778, %r2787; mov.b32 %f12779, %r891; sub.ftz.f32 %f12780, %f12779, %f12778; fma.rn.ftz.f32 %f12781, %f2799, %f12780, %f12778; sin.approx.ftz.f32 %f12782, %f12781; cos.approx.ftz.f32 %f12783, %f12781; div.approx.ftz.f32 %f12784, %f12782, %f12783; mul.ftz.f32 %f12785, %f3435, %f12784; add.ftz.f32 %f12786, %f3437, %f12785; add.ftz.f32 %f12787, %f12786, %f12786; sub.ftz.f32 %f12788, %f12787, %f3434; mul.ftz.f32 %f12789, %f12788, 0f3EA2F983; mov.b32 %r2788, %f12789; and.b32 %r2789, %r2788, -2147483648; or.b32 %r2790, %r2789, 1056964608; mov.b32 %f12790, %r2790; add.ftz.f32 %f12791, %f12789, %f12790; cvt.rzi.ftz.s32.f32 %r2791, %f12791; cvt.rn.f32.s32 %f12792, %r2791; mul.ftz.f32 %f12793, %f12792, 0f40490000; sub.ftz.f32 %f12794, %f12788, %f12793; mul.ftz.f32 %f12795, %f12792, 0f3A7DA000; sub.ftz.f32 %f12796, %f12794, %f12795; mul.ftz.f32 %f12797, %f12792, 0f34222000; sub.ftz.f32 %f12798, %f12796, %f12797; mul.ftz.f32 %f12799, %f12792, 0f2CB4611A; sub.ftz.f32 %f12800, %f12798, %f12799; add.ftz.f32 %f12801, %f12800, 0fBFC90FDB; add.ftz.f32 %f12802, %f12801, 0f3FC90FDB; mul.ftz.f32 %f12803, %f12802, %f12802; and.b32 %r2792, %r2791, 1; setp.eq.b32 %p1850, %r2792, 1; neg.ftz.f32 %f12804, %f12802; selp.f32 %f12805, %f12804, %f12802, %p1850; fma.rn.ftz.f32 %f12806, %f12803, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f12807, %f12803, %f12806, 0f3C088777; fma.rn.ftz.f32 %f12808, %f12803, %f12807, 0fBE2AAAA6; mul.ftz.f32 %f12809, %f12805, %f12808; fma.rn.ftz.f32 %f12810, %f12803, %f12809, %f12805; mul.ftz.f32 %f12811, %f12803, 0f3491ED89; mov.f32 %f12812, 0f37D0078B; sub.ftz.f32 %f12813, %f12812, %f12811; fma.rn.ftz.f32 %f12814, %f12803, %f12813, 0fBAB60B58; fma.rn.ftz.f32 %f12815, %f12803, %f12814, 0f3D2AAAAA; fma.rn.ftz.f32 %f12816, %f12803, %f12815, 0fBF000000; fma.rn.ftz.f32 %f12817, %f12803, %f12816, 0f3F800000; neg.ftz.f32 %f12818, %f12817; selp.f32 %f12819, %f12818, %f12817, %p1850; abs.ftz.f32 %f12820, %f12810; setp.gt.ftz.f32 %p1851, %f12820, 0f3F800000; selp.f32 %f12821, 0f00000000, %f12810, %p1851; abs.ftz.f32 %f12822, %f12819; setp.gt.ftz.f32 %p1852, %f12822, 0f3F800000; selp.f32 %f12823, 0f00000000, %f12819, %p1852; add.ftz.f32 %f12824, %f2775, %f2775; sub.ftz.f32 %f12825, %f12767, %f12824; max.ftz.f32 %f12826, %f12825, %f12766; min.ftz.f32 %f12827, %f12826, %f12767; abs.ftz.f32 %f12828, %f12827; sub.ftz.f32 %f12829, %f12767, %f12828; mul.ftz.f32 %f12830, %f12829, 0f3F000000; sqrt.approx.ftz.f32 %f12831, %f12830; setp.gt.ftz.f32 %p1853, %f12828, 0f3F11EB85; selp.f32 %f12832, %f12831, %f12828, %p1853; mul.ftz.f32 %f12833, %f12832, %f12832; mov.f32 %f12834, 0f3C94D2E9; mov.f32 %f12835, 0f3D53F941; fma.rn.ftz.f32 %f12836, %f12835, %f12833, %f12834; mov.f32 %f12837, 0f3D3F841F; fma.rn.ftz.f32 %f12838, %f12836, %f12833, %f12837; mov.f32 %f12839, 0f3D994929; fma.rn.ftz.f32 %f12840, %f12838, %f12833, %f12839; mov.f32 %f12841, 0f3E2AAB94; fma.rn.ftz.f32 %f12842, %f12840, %f12833, %f12841; mul.ftz.f32 %f12843, %f12833, %f12842; fma.rn.ftz.f32 %f12844, %f12843, %f12832, %f12832; mov.f32 %f12845, 0fC0000000; fma.rn.ftz.f32 %f12846, %f12845, %f12844, %f12736; selp.f32 %f12847, %f12846, %f12844, %p1853; setp.le.ftz.f32 %p1854, %f12847, 0f7F800000; mov.b32 %r2793, %f12847; mov.b32 %r2794, %f12827; and.b32 %r2795, %r2794, -2147483648; or.b32 %r2796, %r2795, %r2793; mov.b32 %f12848, %r2796; selp.f32 %f12849, %f12848, %f12847, %p1854; add.ftz.f32 %f12850, %f12849, %f12849; mul.ftz.f32 %f12851, %f3427, %f12850; mul.ftz.f32 %f12852, %f12851, 0f3F000000; mul.ftz.f32 %f12853, %f12852, 0f3EA2F983; mov.b32 %r2797, %f12853; and.b32 %r2798, %r2797, -2147483648; or.b32 %r2799, %r2798, 1056964608; mov.b32 %f12854, %r2799; add.ftz.f32 %f12855, %f12853, %f12854; cvt.rzi.ftz.s32.f32 %r2800, %f12855; cvt.rn.f32.s32 %f12856, %r2800; mul.ftz.f32 %f12857, %f12856, 0f40490000; sub.ftz.f32 %f12858, %f12852, %f12857; mul.ftz.f32 %f12859, %f12856, 0f3A7DA000; sub.ftz.f32 %f12860, %f12858, %f12859; mul.ftz.f32 %f12861, %f12856, 0f34222000; sub.ftz.f32 %f12862, %f12860, %f12861; mul.ftz.f32 %f12863, %f12856, 0f2CB4611A; sub.ftz.f32 %f12864, %f12862, %f12863; add.ftz.f32 %f12865, %f12864, 0fBFC90FDB; add.ftz.f32 %f12866, %f12865, 0f3FC90FDB; mul.ftz.f32 %f12867, %f12866, %f12866; mul.ftz.f32 %f12868, %f12867, 0f3491ED89; sub.ftz.f32 %f12869, %f12812, %f12868; fma.rn.ftz.f32 %f12870, %f12867, %f12869, 0fBAB60B58; fma.rn.ftz.f32 %f12871, %f12867, %f12870, 0f3D2AAAAA; fma.rn.ftz.f32 %f12872, %f12867, %f12871, 0fBF000000; fma.rn.ftz.f32 %f12873, %f12867, %f12872, 0f3F800000; and.b32 %r2801, %r2800, 1; setp.eq.b32 %p1855, %r2801, 1; neg.ftz.f32 %f12874, %f12873; selp.f32 %f12875, %f12874, %f12873, %p1855; abs.ftz.f32 %f12876, %f12875; setp.gt.ftz.f32 %p1856, %f12876, 0f3F800000; mul.ftz.f32 %f12877, %f12875, 0f3E800000; selp.f32 %f12878, 0f00000000, %f12877, %p1856; div.approx.ftz.f32 %f12879, %f12878, %f3427; mul.ftz.f32 %f12880, %f12785, %f12785; fma.rn.ftz.f32 %f12881, %f3435, %f3435, %f12880; add.ftz.f32 %f12882, %f12881, %f12881; mul.ftz.f32 %f12883, %f12882, %f12780; mul.ftz.f32 %f12884, %f12883, %f12823; div.approx.ftz.f32 %f12885, %f3435, %f12884; mul.ftz.f32 %f12886, %f12851, 0f3EA2F983; mov.b32 %r2802, %f12886; and.b32 %r2803, %r2802, -2147483648; or.b32 %r2804, %r2803, 1056964608; mov.b32 %f12887, %r2804; add.ftz.f32 %f12888, %f12886, %f12887; cvt.rzi.ftz.s32.f32 %r2805, %f12888; cvt.rn.f32.s32 %f12889, %r2805; mul.ftz.f32 %f12890, %f12889, 0f40490000; sub.ftz.f32 %f12891, %f12851, %f12890; mul.ftz.f32 %f12892, %f12889, 0f3A7DA000; sub.ftz.f32 %f12893, %f12891, %f12892; mul.ftz.f32 %f12894, %f12889, 0f34222000; sub.ftz.f32 %f12895, %f12893, %f12894; mul.ftz.f32 %f12896, %f12889, 0f2CB4611A; sub.ftz.f32 %f12897, %f12895, %f12896; add.ftz.f32 %f12898, %f12897, 0fBFC90FDB; add.ftz.f32 %f12899, %f12898, 0f3FC90FDB; mul.ftz.f32 %f12900, %f12899, %f12899; and.b32 %r2806, %r2805, 1; setp.eq.b32 %p1857, %r2806, 1; neg.ftz.f32 %f12901, %f12899; selp.f32 %f12902, %f12901, %f12899, %p1857; fma.rn.ftz.f32 %f12903, %f12900, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f12904, %f12900, %f12903, 0f3C088777; fma.rn.ftz.f32 %f12905, %f12900, %f12904, 0fBE2AAAA6; mul.ftz.f32 %f12906, %f12902, %f12905; fma.rn.ftz.f32 %f12907, %f12900, %f12906, %f12902; mul.ftz.f32 %f12908, %f12900, 0f3491ED89; sub.ftz.f32 %f12909, %f12812, %f12908; fma.rn.ftz.f32 %f12910, %f12900, %f12909, 0fBAB60B58; fma.rn.ftz.f32 %f12911, %f12900, %f12910, 0f3D2AAAAA; fma.rn.ftz.f32 %f12912, %f12900, %f12911, 0fBF000000; fma.rn.ftz.f32 %f12913, %f12900, %f12912, 0f3F800000; neg.ftz.f32 %f12914, %f12913; selp.f32 %f12915, %f12914, %f12913, %p1857; abs.ftz.f32 %f12916, %f12907; setp.gt.ftz.f32 %p1858, %f12916, 0f3F800000; selp.f32 %f12917, 0f00000000, %f12907, %p1858; abs.ftz.f32 %f12918, %f12915; setp.gt.ftz.f32 %p1859, %f12918, 0f3F800000; selp.f32 %f12919, 0f00000000, %f12915, %p1859; mul.ftz.f32 %f12920, %f12823, %f12919; mul.ftz.f32 %f12921, %f3431, %f12920; mul.ftz.f32 %f12922, %f3432, %f12920; mul.ftz.f32 %f12923, %f3433, %f12920; mul.ftz.f32 %f12924, %f12823, %f12917; mul.ftz.f32 %f12925, %f3430, %f3432; mul.ftz.f32 %f12926, %f3429, %f3433; sub.ftz.f32 %f12927, %f12925, %f12926; mul.ftz.f32 %f12928, %f12927, %f12924; mul.ftz.f32 %f12929, %f3428, %f3433; mul.ftz.f32 %f12930, %f3430, %f3431; sub.ftz.f32 %f12931, %f12929, %f12930; mul.ftz.f32 %f12932, %f12931, %f12924; mul.ftz.f32 %f12933, %f3429, %f3431; mul.ftz.f32 %f12934, %f3428, %f3432; sub.ftz.f32 %f12935, %f12933, %f12934; mul.ftz.f32 %f12936, %f12935, %f12924; sub.ftz.f32 %f12937, %f12921, %f12928; sub.ftz.f32 %f12938, %f12922, %f12932; sub.ftz.f32 %f12939, %f12923, %f12936; fma.rn.ftz.f32 %f16735, %f3428, %f12821, %f12937; fma.rn.ftz.f32 %f16736, %f3429, %f12821, %f12938; fma.rn.ftz.f32 %f16737, %f3430, %f12821, %f12939; mul.ftz.f32 %f12940, %f12879, %f12885; abs.ftz.f32 %f12941, %f12940; abs.ftz.f32 %f12942, %f12788; sub.ftz.f32 %f12943, %f12736, %f12942; setp.lt.ftz.f32 %p1860, %f12943, 0f3A83126F; selp.f32 %f16738, 0f00000000, %f12941, %p1860; mov.u32 %r3285, 10; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; bra.uni $L__BB106_2309; $L__BB106_1882: mov.f32 %f11367, 0f3F800000; div.approx.ftz.f32 %f16459, %f2829, %f11367; $L__BB106_1884: add.ftz.f32 %f11370, %f16459, 0fBF800000; add.ftz.f32 %f11371, %f11370, 0f3F800000; mul.ftz.f32 %f11372, %f11371, %f11371; fma.rn.ftz.f32 %f11373, %f11372, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11374, %f11371, %f11373; fma.rn.ftz.f32 %f11375, %f11372, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11376, %f11372, %f11375, 0f3F800000; div.approx.ftz.f32 %f11377, %f11374, %f11376; sub.ftz.f32 %f11379, %f11338, %f11377; setp.gt.ftz.f32 %p1639, %f2829, 0f3F800000; selp.f32 %f11380, %f11379, %f11377, %p1639; mov.b32 %r2697, %f2828; and.b32 %r2698, %r2697, -2147483648; mov.b32 %r2699, %f11380; and.b32 %r2700, %r2699, 2147483647; or.b32 %r875, %r2700, %r2698; div.approx.ftz.f32 %f2833, %f11338, %f2812; abs.ftz.f32 %f2834, %f2833; setp.eq.ftz.f32 %p1640, %f2834, 0f00000000; @%p1640 bra $L__BB106_1889; setp.eq.ftz.f32 %p1641, %f2834, 0f3F800000; mov.f32 %f16460, 0f3F800000; @%p1641 bra $L__BB106_1889; setp.gt.ftz.f32 %p1642, %f2834, 0f3F800000; @%p1642 bra $L__BB106_1888; bra.uni $L__BB106_1887; $L__BB106_1888: mov.f32 %f11383, 0f3F800000; div.approx.ftz.f32 %f16460, %f11383, %f2834; bra.uni $L__BB106_1889; $L__BB106_1887: mov.f32 %f11382, 0f3F800000; div.approx.ftz.f32 %f16460, %f2834, %f11382; $L__BB106_1889: mov.b32 %f11384, %r875; add.ftz.f32 %f11385, %f16460, 0fBF800000; add.ftz.f32 %f11386, %f11385, 0f3F800000; mul.ftz.f32 %f11387, %f11386, %f11386; fma.rn.ftz.f32 %f11388, %f11387, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11389, %f11386, %f11388; fma.rn.ftz.f32 %f11390, %f11387, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11391, %f11387, %f11390, 0f3F800000; div.approx.ftz.f32 %f11392, %f11389, %f11391; sub.ftz.f32 %f11394, %f11338, %f11392; setp.gt.ftz.f32 %p1643, %f2834, 0f3F800000; selp.f32 %f11395, %f11394, %f11392, %p1643; mov.b32 %r2702, %f2833; and.b32 %r2703, %r2702, -2147483648; mov.b32 %r2704, %f11395; and.b32 %r2705, %r2704, 2147483647; or.b32 %r2706, %r2705, %r2703; mov.b32 %f11396, %r2706; add.ftz.f32 %f11397, %f11396, %f11396; mov.b32 %f11398, %r874; sub.ftz.f32 %f11399, %f11398, %f11384; fma.rn.ftz.f32 %f11400, %f2799, %f11399, %f11384; sin.approx.ftz.f32 %f11401, %f11400; cos.approx.ftz.f32 %f11402, %f11400; div.approx.ftz.f32 %f11403, %f11401, %f11402; mul.ftz.f32 %f11404, %f2820, %f11403; add.ftz.f32 %f11405, %f2822, %f11404; add.ftz.f32 %f11406, %f11405, %f11405; sub.ftz.f32 %f11407, %f11406, %f2819; mul.ftz.f32 %f11408, %f11407, 0f3EA2F983; mov.b32 %r2707, %f11408; and.b32 %r2708, %r2707, -2147483648; or.b32 %r2709, %r2708, 1056964608; mov.b32 %f11409, %r2709; add.ftz.f32 %f11410, %f11408, %f11409; cvt.rzi.ftz.s32.f32 %r2710, %f11410; cvt.rn.f32.s32 %f11411, %r2710; mul.ftz.f32 %f11412, %f11411, 0f40490000; sub.ftz.f32 %f11413, %f11407, %f11412; mul.ftz.f32 %f11414, %f11411, 0f3A7DA000; sub.ftz.f32 %f11415, %f11413, %f11414; mul.ftz.f32 %f11416, %f11411, 0f34222000; sub.ftz.f32 %f11417, %f11415, %f11416; mul.ftz.f32 %f11418, %f11411, 0f2CB4611A; sub.ftz.f32 %f11419, %f11417, %f11418; add.ftz.f32 %f11420, %f11419, 0fBFC90FDB; add.ftz.f32 %f11421, %f11420, 0f3FC90FDB; mul.ftz.f32 %f11422, %f11421, %f11421; and.b32 %r2711, %r2710, 1; setp.eq.b32 %p1644, %r2711, 1; neg.ftz.f32 %f11423, %f11421; selp.f32 %f11424, %f11423, %f11421, %p1644; fma.rn.ftz.f32 %f11425, %f11422, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f11426, %f11422, %f11425, 0f3C088777; fma.rn.ftz.f32 %f11427, %f11422, %f11426, 0fBE2AAAA6; mul.ftz.f32 %f11428, %f11424, %f11427; fma.rn.ftz.f32 %f11429, %f11422, %f11428, %f11424; mul.ftz.f32 %f11430, %f11422, 0f3491ED89; mov.f32 %f11431, 0f37D0078B; sub.ftz.f32 %f11432, %f11431, %f11430; fma.rn.ftz.f32 %f11433, %f11422, %f11432, 0fBAB60B58; fma.rn.ftz.f32 %f11434, %f11422, %f11433, 0f3D2AAAAA; fma.rn.ftz.f32 %f11435, %f11422, %f11434, 0fBF000000; fma.rn.ftz.f32 %f11436, %f11422, %f11435, 0f3F800000; neg.ftz.f32 %f11437, %f11436; selp.f32 %f11438, %f11437, %f11436, %p1644; abs.ftz.f32 %f11439, %f11429; setp.gt.ftz.f32 %p1645, %f11439, 0f3F800000; selp.f32 %f11440, 0f00000000, %f11429, %p1645; abs.ftz.f32 %f11441, %f11438; setp.gt.ftz.f32 %p1646, %f11441, 0f3F800000; selp.f32 %f11442, 0f00000000, %f11438, %p1646; add.ftz.f32 %f11443, %f2775, 0fBF000000; mul.ftz.f32 %f11444, %f11443, %f11397; sin.approx.ftz.f32 %f11445, %f11444; cos.approx.ftz.f32 %f11446, %f11444; div.approx.ftz.f32 %f11447, %f11445, %f11446; mul.ftz.f32 %f11448, %f2812, %f11447; add.ftz.f32 %f11449, %f11448, 0f40490FDB; mul.ftz.f32 %f11450, %f11404, %f11404; fma.rn.ftz.f32 %f11451, %f2820, %f2820, %f11450; add.ftz.f32 %f11452, %f11451, %f11451; mul.ftz.f32 %f11453, %f11399, %f11452; mul.ftz.f32 %f11454, %f11453, %f11442; div.approx.ftz.f32 %f11455, %f2820, %f11454; mul.ftz.f32 %f11456, %f11448, %f11448; fma.rn.ftz.f32 %f11457, %f2812, %f2812, %f11456; mul.ftz.f32 %f11458, %f11397, %f11457; div.approx.ftz.f32 %f11459, %f2812, %f11458; mul.ftz.f32 %f11460, %f11449, 0f3EA2F983; mov.b32 %r2712, %f11460; and.b32 %r2713, %r2712, -2147483648; or.b32 %r2714, %r2713, 1056964608; mov.b32 %f11461, %r2714; add.ftz.f32 %f11462, %f11460, %f11461; cvt.rzi.ftz.s32.f32 %r2715, %f11462; cvt.rn.f32.s32 %f11463, %r2715; mul.ftz.f32 %f11464, %f11463, 0f40490000; sub.ftz.f32 %f11465, %f11449, %f11464; mul.ftz.f32 %f11466, %f11463, 0f3A7DA000; sub.ftz.f32 %f11467, %f11465, %f11466; mul.ftz.f32 %f11468, %f11463, 0f34222000; sub.ftz.f32 %f11469, %f11467, %f11468; mul.ftz.f32 %f11470, %f11463, 0f2CB4611A; sub.ftz.f32 %f11471, %f11469, %f11470; add.ftz.f32 %f11472, %f11471, 0fBFC90FDB; add.ftz.f32 %f11473, %f11472, 0f3FC90FDB; mul.ftz.f32 %f11474, %f11473, %f11473; and.b32 %r2716, %r2715, 1; setp.eq.b32 %p1647, %r2716, 1; neg.ftz.f32 %f11475, %f11473; selp.f32 %f11476, %f11475, %f11473, %p1647; fma.rn.ftz.f32 %f11477, %f11474, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f11478, %f11474, %f11477, 0f3C088777; fma.rn.ftz.f32 %f11479, %f11474, %f11478, 0fBE2AAAA6; mul.ftz.f32 %f11480, %f11476, %f11479; fma.rn.ftz.f32 %f11481, %f11474, %f11480, %f11476; mul.ftz.f32 %f11482, %f11474, 0f3491ED89; sub.ftz.f32 %f11483, %f11431, %f11482; fma.rn.ftz.f32 %f11484, %f11474, %f11483, 0fBAB60B58; fma.rn.ftz.f32 %f11485, %f11474, %f11484, 0f3D2AAAAA; fma.rn.ftz.f32 %f11486, %f11474, %f11485, 0fBF000000; fma.rn.ftz.f32 %f11487, %f11474, %f11486, 0f3F800000; neg.ftz.f32 %f11488, %f11487; selp.f32 %f11489, %f11488, %f11487, %p1647; abs.ftz.f32 %f11490, %f11481; setp.gt.ftz.f32 %p1648, %f11490, 0f3F800000; selp.f32 %f11491, 0f00000000, %f11481, %p1648; abs.ftz.f32 %f11492, %f11489; setp.gt.ftz.f32 %p1649, %f11492, 0f3F800000; selp.f32 %f11493, 0f00000000, %f11489, %p1649; mul.ftz.f32 %f11494, %f11442, %f11493; mul.ftz.f32 %f11495, %f2816, %f11494; mul.ftz.f32 %f11496, %f2817, %f11494; mul.ftz.f32 %f11497, %f2818, %f11494; mul.ftz.f32 %f11498, %f11442, %f11491; mul.ftz.f32 %f11499, %f2814, %f2818; mul.ftz.f32 %f11500, %f2815, %f2817; sub.ftz.f32 %f11501, %f11500, %f11499; mul.ftz.f32 %f11502, %f11501, %f11498; mul.ftz.f32 %f11503, %f2815, %f2816; mul.ftz.f32 %f11504, %f2813, %f2818; sub.ftz.f32 %f11505, %f11504, %f11503; mul.ftz.f32 %f11506, %f11505, %f11498; mul.ftz.f32 %f11507, %f2813, %f2817; mul.ftz.f32 %f11508, %f2814, %f2816; sub.ftz.f32 %f11509, %f11508, %f11507; mul.ftz.f32 %f11510, %f11509, %f11498; sub.ftz.f32 %f11511, %f11495, %f11502; sub.ftz.f32 %f11512, %f11496, %f11506; sub.ftz.f32 %f11513, %f11497, %f11510; fma.rn.ftz.f32 %f16735, %f2813, %f11440, %f11511; fma.rn.ftz.f32 %f16736, %f2814, %f11440, %f11512; fma.rn.ftz.f32 %f16737, %f2815, %f11440, %f11513; mul.ftz.f32 %f11514, %f11455, %f11459; abs.ftz.f32 %f11515, %f11514; abs.ftz.f32 %f11516, %f11407; sub.ftz.f32 %f11517, %f11338, %f11516; setp.lt.ftz.f32 %p1650, %f11517, 0f3A83126F; selp.f32 %f16738, 0f00000000, %f11515, %p1650; mov.u32 %r3285, 9; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; $L__BB106_2309: and.b32 %r2903, %r3285, 1; setp.eq.b32 %p2061, %r2903, 1; mov.pred %p2062, 0; xor.pred %p2063, %p2061, %p2062; not.pred %p2064, %p2063; @%p2064 bra $L__BB106_2317; bra.uni $L__BB106_2310; $L__BB106_2317: ld.const.u64 %rd2692, [kernel_params+2160]; cvta.to.global.u64 %rd2693, %rd2692; ld.local.u32 %r2913, [%rd7+-12]; mul.wide.s32 %rd2694, %r2913, 208; add.s64 %rd2695, %rd2693, %rd2694; ld.global.f32 %f4137, [%rd2695+180]; setp.leu.ftz.f32 %p2075, %f4137, 0f3F800000; @%p2075 bra $L__BB106_2319; ld.local.f32 %f15200, [%rd587+8]; ld.local.f32 %f15201, [%rd587+12]; mul.ftz.f32 %f15202, %f16736, %f15201; fma.rn.ftz.f32 %f15203, %f16735, %f15200, %f15202; ld.local.f32 %f15204, [%rd587+16]; fma.rn.ftz.f32 %f15205, %f16737, %f15204, %f15203; mov.f32 %f15206, 0f3F800000; min.ftz.f32 %f15207, %f15205, %f15206; abs.ftz.f32 %f15208, %f15207; setp.lt.ftz.f32 %p2076, %f15208, 0f3F800000; add.ftz.f32 %f15209, %f15208, 0fBF800000; add.ftz.f32 %f15210, %f15209, 0f3F800000; selp.f32 %f15211, %f15210, 0f3F800000, %p2076; sub.ftz.f32 %f15212, %f15206, %f15211; sqrt.approx.ftz.f32 %f15213, %f15212; fma.rn.ftz.f32 %f15214, %f15211, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f15215, %f15211, %f15214, 0fBE5A6B93; fma.rn.ftz.f32 %f15216, %f15211, %f15215, 0f3FC90FDB; mul.ftz.f32 %f15217, %f15213, %f15216; setp.lt.ftz.f32 %p2077, %f15207, 0f00000000; mov.f32 %f15218, 0f00000000; mov.f32 %f15219, 0f40490FDB; sub.ftz.f32 %f15220, %f15219, %f15217; selp.f32 %f15221, %f15220, %f15217, %p2077; mul.ftz.f32 %f15222, %f4137, %f15221; cos.approx.ftz.f32 %f15223, %f15222; max.ftz.f32 %f15224, %f15223, %f15218; div.approx.ftz.f32 %f15225, %f15224, %f15207; mul.ftz.f32 %f16739, %f16739, %f15225; mul.ftz.f32 %f16740, %f16740, %f15225; mul.ftz.f32 %f16741, %f16741, %f15225; $L__BB106_2319: and.b32 %r2914, %r3285, 4; setp.eq.s32 %p2078, %r2914, 0; @%p2078 bra $L__BB106_2330; ld.local.f32 %f4144, [%rd587+16]; ld.local.f32 %f4145, [%rd7+-68]; ld.local.f32 %f4146, [%rd7+-76]; ld.local.f32 %f4147, [%rd587+8]; setp.neu.ftz.f32 %p2079, %f4147, %f4146; mov.u16 %rs160, 0; @%p2079 bra $L__BB106_2323; ld.local.f32 %f15226, [%rd587+12]; ld.local.f32 %f15227, [%rd7+-72]; setp.neu.ftz.f32 %p2080, %f15226, %f15227; @%p2080 bra $L__BB106_2323; setp.eq.ftz.f32 %p2081, %f4144, %f4145; selp.u16 %rs160, 1, 0, %p2081; $L__BB106_2323: setp.ne.s16 %p2082, %rs160, 0; @%p2082 bra $L__BB106_2330; ld.local.f32 %f15229, [%rd587+12]; mul.ftz.f32 %f15230, %f16736, %f15229; fma.rn.ftz.f32 %f15231, %f16735, %f4147, %f15230; fma.rn.ftz.f32 %f15232, %f16737, %f4144, %f15231; setp.lt.ftz.f32 %p2083, %f15232, 0f00000000; mov.f32 %f16746, 0f00000000; ld.local.f32 %f15233, [%rd7+-72]; neg.ftz.f32 %f15234, %f15233; neg.ftz.f32 %f15235, %f4146; selp.f32 %f15236, %f15235, %f4146, %p2083; selp.f32 %f15237, %f15234, %f15233, %p2083; neg.ftz.f32 %f15238, %f4145; selp.f32 %f15239, %f15238, %f4145, %p2083; mul.ftz.f32 %f15240, %f16736, %f15237; fma.rn.ftz.f32 %f15241, %f16735, %f15236, %f15240; fma.rn.ftz.f32 %f4148, %f16737, %f15239, %f15241; mul.ftz.f32 %f15242, %f15229, %f15237; fma.rn.ftz.f32 %f15243, %f4147, %f15236, %f15242; fma.rn.ftz.f32 %f15244, %f4144, %f15239, %f15243; mul.ftz.f32 %f4149, %f15232, %f15244; setp.eq.ftz.f32 %p2084, %f4149, 0f00000000; @%p2084 bra $L__BB106_2326; div.approx.ftz.f32 %f16746, %f4148, %f4149; $L__BB106_2326: setp.ge.ftz.f32 %p2085, %f16746, 0f3F800000; mov.f32 %f16747, 0f3F800000; @%p2085 bra $L__BB106_2329; setp.lt.ftz.f32 %p2086, %f16746, 0f00000000; mov.f32 %f16747, 0f00000000; @%p2086 bra $L__BB106_2329; mul.ftz.f32 %f15247, %f16746, %f16746; mul.ftz.f32 %f15248, %f16746, %f15247; sub.ftz.f32 %f15249, %f15247, %f15248; add.ftz.f32 %f16747, %f16746, %f15249; $L__BB106_2329: mul.ftz.f32 %f16739, %f16739, %f16747; mul.ftz.f32 %f16740, %f16740, %f16747; mul.ftz.f32 %f16741, %f16741, %f16747; bra.uni $L__BB106_2330; $L__BB106_2310: ld.const.f32 %f4134, [kernel_params+1040]; setp.ltu.ftz.f32 %p2065, %f4134, 0f00000000; shr.u32 %r2904, %r3285, 2; and.b32 %r2905, %r2904, 1; setp.eq.b32 %p2066, %r2905, 1; or.pred %p2067, %p2066, %p2065; @%p2067 bra $L__BB106_2330; mov.f32 %f16742, 0f00000000; setp.gt.u32 %p2068, %r873, 31; @%p2068 bra $L__BB106_2313; mov.u32 %r2907, 1; shl.b32 %r2908, %r2907, %r873; and.b32 %r2909, %r2908, -2145385984; setp.ne.s32 %p2069, %r2909, 0; @%p2069 bra $L__BB106_2316; $L__BB106_2313: add.s32 %r2910, %r873, -10; setp.lt.u32 %p2070, %r2910, 7; @%p2070 bra $L__BB106_2315; add.s32 %r2911, %r873, -22; setp.gt.u32 %p2071, %r2911, 2; setp.ne.s32 %p2072, %r873, 27; mov.f32 %f16742, 0f3F800000; and.pred %p2073, %p2071, %p2072; @%p2073 bra $L__BB106_2316; $L__BB106_2315: ld.local.v2.f32 {%f15196, %f15197}, [%rd587+20]; mul.ftz.f32 %f16742, %f15196, %f15197; $L__BB106_2316: or.b32 %r2912, %r3285, 128; setp.le.ftz.f32 %p2074, %f16742, %f4134; selp.b32 %r3285, %r2912, %r3285, %p2074; $L__BB106_2330: add.s32 %r907, %r873, -2; add.u64 %rd2696, %SP, 704; add.u64 %rd591, %SPL, 704; setp.neu.ftz.f32 %p2087, %f16738, 0f00000000; @%p2087 bra $L__BB106_2336; bra.uni $L__BB106_2331; $L__BB106_2336: setp.lt.u32 %p2092, %r907, 7; ld.local.f32 %f15250, [%rd587+-12]; mul.ftz.f32 %f4160, %f16739, %f15250; ld.local.f32 %f15251, [%rd587+-8]; mul.ftz.f32 %f4161, %f16740, %f15251; ld.local.f32 %f15252, [%rd587+-4]; mul.ftz.f32 %f4162, %f16741, %f15252; mov.u32 %r2920, 0; st.local.u32 [%rd591], %r2920; st.local.u32 [%rd591+4], %r2920; st.local.u32 [%rd591+8], %r2920; st.local.u32 [%rd591+12], %r2920; st.local.u32 [%rd591+16], %r2920; st.local.u32 [%rd591+20], %r2920; @%p2092 bra $L__BB106_2339; bra.uni $L__BB106_2337; $L__BB106_2339: st.local.f32 [%rd591], %f4160; st.local.f32 [%rd591+4], %f4161; st.local.f32 [%rd591+8], %f4162; bra.uni $L__BB106_2340; $L__BB106_2331: setp.lt.u32 %p2088, %r907, 7; mov.u32 %r2915, 0; st.local.u32 [%rd591], %r2915; st.local.u32 [%rd591+4], %r2915; st.local.u32 [%rd591+8], %r2915; st.local.u32 [%rd591+12], %r2915; st.local.u32 [%rd591+16], %r2915; st.local.u32 [%rd591+20], %r2915; @%p2088 bra $L__BB106_2334; bra.uni $L__BB106_2332; $L__BB106_2334: st.local.u32 [%rd591], %r2915; st.local.u32 [%rd591+4], %r2915; st.local.u32 [%rd591+8], %r2915; bra.uni $L__BB106_2335; $L__BB106_2337: add.s32 %r2921, %r873, -9; setp.gt.u32 %p2093, %r2921, 11; setp.ne.s32 %p2094, %r873, 29; and.pred %p2095, %p2094, %p2093; @%p2095 bra $L__BB106_2340; st.local.f32 [%rd591+12], %f4160; st.local.f32 [%rd591+16], %f4161; st.local.f32 [%rd591+20], %f4162; $L__BB106_2340: st.local.f32 [%rd591+24], %f4160; st.local.f32 [%rd591+28], %f4161; st.local.f32 [%rd591+32], %f4162; @%p1586 bra $L__BB106_2342; add.u64 %rd2924, %SP, 1728; ld.local.f32 %f15253, [%rd587+4]; mul.ftz.f32 %f15254, %f16738, %f15253; mul.lo.s64 %rd2697, %rd586, 80; add.s64 %rd2699, %rd2924, %rd2697; add.s64 %rd2700, %rd2699, 304; { // callseq 243, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2924; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f16735; st.param.f32 [param1+4], %f16736; st.param.f32 [param1+8], %f16737; .param .b64 param2; st.param.b64 [param2+0], %rd2700; .param .b64 param3; st.param.b64 [param3+0], %rd2696; .param .b32 param4; st.param.f32 [param4+0], %f15254; .param .b32 param5; st.param.f32 [param5+0], %f15253; .param .b32 retval0; call.uni (retval0), _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj, ( param0, param1, param2, param3, param4, param5 ); ld.param.f32 %f16738, [retval0+0]; } // callseq 243 bra.uni $L__BB106_2342; $L__BB106_2332: add.s32 %r2916, %r873, -9; setp.gt.u32 %p2089, %r2916, 11; setp.ne.s32 %p2090, %r873, 29; and.pred %p2091, %p2090, %p2089; @%p2091 bra $L__BB106_2335; st.local.u32 [%rd591+12], %r2915; st.local.u32 [%rd591+16], %r2915; st.local.u32 [%rd591+20], %r2915; $L__BB106_2335: st.local.u32 [%rd591+24], %r2915; st.local.u32 [%rd591+28], %r2915; st.local.u32 [%rd591+32], %r2915; $L__BB106_2342: setp.eq.ftz.f32 %p2097, %f16738, 0f00000000; mov.u32 %r3294, 0; @%p2097 bra $L__BB106_2405; ld.local.f32 %f4165, [%rd591+24]; setp.eq.ftz.f32 %p2098, %f4165, 0f00000000; ld.local.f32 %f4166, [%rd591+28]; setp.eq.ftz.f32 %p2099, %f4166, 0f00000000; and.pred %p2100, %p2098, %p2099; ld.local.f32 %f4167, [%rd591+32]; setp.eq.ftz.f32 %p2101, %f4167, 0f00000000; and.pred %p2102, %p2101, %p2100; @%p2102 bra $L__BB106_2405; and.b32 %r908, %r3285, 32; setp.eq.s32 %p2103, %r908, 0; @%p2103 bra $L__BB106_2346; bra.uni $L__BB106_2345; $L__BB106_2346: mul.ftz.f32 %f15258, %f16736, %f16736; fma.rn.ftz.f32 %f15259, %f16735, %f16735, %f15258; fma.rn.ftz.f32 %f15260, %f16737, %f16737, %f15259; sqrt.approx.ftz.f32 %f15261, %f15260; div.approx.ftz.f32 %f4168, %f16735, %f15261; div.approx.ftz.f32 %f4169, %f16736, %f15261; div.approx.ftz.f32 %f4170, %f16737, %f15261; ld.local.v4.f32 {%f16770, %f16771, %f16772, %f15265}, [%rd1]; ld.local.u32 %r2926, [%rd7+-24]; and.b32 %r2927, %r2926, 1; setp.eq.b32 %p2105, %r2927, 1; xor.pred %p2107, %p2105, %p2062; not.pred %p2108, %p2107; @%p2108 bra $L__BB106_2354; ld.const.u64 %rd2708, [kernel_params+2232]; cvta.to.global.u64 %rd2709, %rd2708; ld.local.v2.u32 {%r2928, %r2929}, [%rd7+-32]; mul.wide.s32 %rd2710, %r2929, 16; add.s64 %rd2711, %rd2709, %rd2710; ld.const.u64 %rd2712, [kernel_params+2256]; cvta.to.global.u64 %rd2713, %rd2712; ld.global.u32 %r2931, [%rd2711+12]; mul.wide.u32 %rd2714, %r2931, 12; add.s64 %rd2715, %rd2713, %rd2714; ld.global.f32 %f4177, [%rd2715]; ld.global.f32 %f4178, [%rd2715+4]; ld.global.f32 %f4179, [%rd2715+8]; add.s32 %r2932, %r2931, 1; mul.wide.u32 %rd2716, %r2932, 12; add.s64 %rd2717, %rd2713, %rd2716; ld.global.f32 %f4180, [%rd2717]; ld.global.f32 %f4181, [%rd2717+4]; ld.global.f32 %f4182, [%rd2717+8]; add.s32 %r2933, %r2931, 2; mul.wide.u32 %rd2718, %r2933, 12; add.s64 %rd2719, %rd2713, %rd2718; ld.global.f32 %f4183, [%rd2719]; ld.global.f32 %f4184, [%rd2719+4]; ld.global.f32 %f4185, [%rd2719+8]; and.b32 %r2934, %r2928, 4; setp.ne.s32 %p2109, %r2934, 0; mov.f32 %f16764, %f16770; mov.f32 %f16765, %f16771; mov.f32 %f16766, %f16772; mov.f32 %f16767, %f4168; mov.f32 %f16768, %f4169; mov.f32 %f16769, %f4170; @%p2109 bra $L__BB106_2352; and.b32 %r2935, %r2928, 2; setp.eq.s32 %p2110, %r2935, 0; @%p2110 bra $L__BB106_2350; bra.uni $L__BB106_2349; $L__BB106_2350: ld.const.u64 %rd2720, [kernel_params+2160]; cvta.to.global.u64 %rd2721, %rd2720; ld.local.u32 %r2936, [%rd7+-12]; mul.wide.s32 %rd2722, %r2936, 208; add.s64 %rd2723, %rd2721, %rd2722; ld.global.v4.f32 {%f16752, %f16753, %f16754, %f16755}, [%rd2723+48]; ld.global.v4.f32 {%f16756, %f16757, %f16758, %f16759}, [%rd2723+64]; ld.global.v4.f32 {%f16760, %f16761, %f16762, %f16763}, [%rd2723+80]; bra.uni $L__BB106_2351; $L__BB106_2345: ld.local.f32 %f15255, [%rd7]; setp.eq.ftz.f32 %p2104, %f15255, 0f00000000; mov.b32 %r2924, %f15255; add.s32 %r2925, %r2924, 1; mov.b32 %f15256, %r2925; selp.f32 %f15257, 0f00000001, %f15256, %p2104; ld.const.u64 %rd2702, [kernel_params+2648]; cvta.to.global.u64 %rd2703, %rd2702; add.s64 %rd2705, %rd2703, %rd837; st.global.f32 [%rd2705], %f15257; bra.uni $L__BB106_2355; $L__BB106_2349: ld.local.v4.f32 {%f16752, %f16753, %f16754, %f16755}, [%rd7+104]; ld.local.v4.f32 {%f16756, %f16757, %f16758, %f16759}, [%rd7+120]; ld.local.v4.f32 {%f16760, %f16761, %f16762, %f16763}, [%rd7+136]; $L__BB106_2351: mul.ftz.f32 %f15290, %f16771, %f16753; fma.rn.ftz.f32 %f15291, %f16770, %f16752, %f15290; fma.rn.ftz.f32 %f15292, %f16772, %f16754, %f15291; mul.ftz.f32 %f15293, %f16771, %f16757; fma.rn.ftz.f32 %f15294, %f16770, %f16756, %f15293; fma.rn.ftz.f32 %f15295, %f16772, %f16758, %f15294; mul.ftz.f32 %f15296, %f16771, %f16761; fma.rn.ftz.f32 %f15297, %f16770, %f16760, %f15296; fma.rn.ftz.f32 %f15298, %f16772, %f16762, %f15297; add.ftz.f32 %f16764, %f16755, %f15292; add.ftz.f32 %f16765, %f16759, %f15295; add.ftz.f32 %f16766, %f16763, %f15298; mul.ftz.f32 %f15299, %f4169, %f16753; fma.rn.ftz.f32 %f15300, %f4168, %f16752, %f15299; mul.ftz.f32 %f15301, %f4169, %f16757; fma.rn.ftz.f32 %f15302, %f4168, %f16756, %f15301; mul.ftz.f32 %f15303, %f4169, %f16761; fma.rn.ftz.f32 %f15304, %f4168, %f16760, %f15303; fma.rn.ftz.f32 %f16767, %f4170, %f16754, %f15300; fma.rn.ftz.f32 %f16768, %f4170, %f16758, %f15302; fma.rn.ftz.f32 %f16769, %f4170, %f16762, %f15304; $L__BB106_2352: sub.ftz.f32 %f15305, %f4177, %f16764; sub.ftz.f32 %f15306, %f4178, %f16765; sub.ftz.f32 %f15307, %f4179, %f16766; sub.ftz.f32 %f15308, %f4180, %f16764; sub.ftz.f32 %f15309, %f4181, %f16765; sub.ftz.f32 %f15310, %f4182, %f16766; sub.ftz.f32 %f15311, %f4183, %f16764; sub.ftz.f32 %f15312, %f4184, %f16765; sub.ftz.f32 %f15313, %f4185, %f16766; sub.ftz.f32 %f15314, %f15311, %f15305; sub.ftz.f32 %f15315, %f15312, %f15306; sub.ftz.f32 %f15316, %f15313, %f15307; sub.ftz.f32 %f15317, %f15305, %f15308; sub.ftz.f32 %f15318, %f15306, %f15309; sub.ftz.f32 %f15319, %f15307, %f15310; sub.ftz.f32 %f15320, %f15308, %f15311; sub.ftz.f32 %f15321, %f15309, %f15312; sub.ftz.f32 %f15322, %f15310, %f15313; add.ftz.f32 %f15323, %f15311, %f15305; add.ftz.f32 %f15324, %f15312, %f15306; add.ftz.f32 %f15325, %f15313, %f15307; mul.ftz.f32 %f15326, %f15324, %f15316; mul.ftz.f32 %f15327, %f15325, %f15315; sub.ftz.f32 %f15328, %f15326, %f15327; mul.ftz.f32 %f15329, %f15325, %f15314; mul.ftz.f32 %f15330, %f15323, %f15316; sub.ftz.f32 %f15331, %f15329, %f15330; mul.ftz.f32 %f15332, %f15323, %f15315; mul.ftz.f32 %f15333, %f15324, %f15314; sub.ftz.f32 %f15334, %f15332, %f15333; mul.ftz.f32 %f15335, %f16768, %f15331; fma.rn.ftz.f32 %f15336, %f16767, %f15328, %f15335; fma.rn.ftz.f32 %f15337, %f16769, %f15334, %f15336; add.ftz.f32 %f15338, %f15305, %f15308; add.ftz.f32 %f15339, %f15306, %f15309; add.ftz.f32 %f15340, %f15307, %f15310; mul.ftz.f32 %f15341, %f15339, %f15319; mul.ftz.f32 %f15342, %f15340, %f15318; sub.ftz.f32 %f15343, %f15341, %f15342; mul.ftz.f32 %f15344, %f15340, %f15317; mul.ftz.f32 %f15345, %f15338, %f15319; sub.ftz.f32 %f15346, %f15344, %f15345; mul.ftz.f32 %f15347, %f15338, %f15318; mul.ftz.f32 %f15348, %f15339, %f15317; sub.ftz.f32 %f15349, %f15347, %f15348; mul.ftz.f32 %f15350, %f16768, %f15346; fma.rn.ftz.f32 %f15351, %f16767, %f15343, %f15350; fma.rn.ftz.f32 %f15352, %f16769, %f15349, %f15351; add.ftz.f32 %f15353, %f15308, %f15311; add.ftz.f32 %f15354, %f15309, %f15312; add.ftz.f32 %f15355, %f15310, %f15313; mul.ftz.f32 %f15356, %f15354, %f15322; mul.ftz.f32 %f15357, %f15355, %f15321; sub.ftz.f32 %f15358, %f15356, %f15357; mul.ftz.f32 %f15359, %f15355, %f15320; mul.ftz.f32 %f15360, %f15353, %f15322; sub.ftz.f32 %f15361, %f15359, %f15360; mul.ftz.f32 %f15362, %f15353, %f15321; mul.ftz.f32 %f15363, %f15354, %f15320; sub.ftz.f32 %f15364, %f15362, %f15363; mul.ftz.f32 %f15365, %f16768, %f15361; fma.rn.ftz.f32 %f15366, %f16767, %f15358, %f15365; fma.rn.ftz.f32 %f15367, %f16769, %f15364, %f15366; add.ftz.f32 %f15368, %f15337, %f15352; add.ftz.f32 %f15369, %f15367, %f15368; abs.ftz.f32 %f15370, %f15369; mul.ftz.f32 %f15371, %f15370, 0f34000000; min.ftz.f32 %f15372, %f15352, %f15367; min.ftz.f32 %f15373, %f15337, %f15372; max.ftz.f32 %f15374, %f15352, %f15367; max.ftz.f32 %f15375, %f15337, %f15374; setp.ge.ftz.f32 %p2111, %f15373, %f15371; mul.ftz.f32 %f15376, %f15370, 0fB4000000; setp.le.ftz.f32 %p2112, %f15375, %f15376; or.pred %p2113, %p2111, %p2112; @%p2113 bra $L__BB106_2354; ld.local.v2.f32 {%f15377, %f15378}, [%rd7+-64]; mul.ftz.f32 %f15381, %f15377, 0f43800000; cvt.rzi.ftz.s32.f32 %r2937, %f15381; mul.ftz.f32 %f15382, %f15378, 0f43800000; cvt.rzi.ftz.s32.f32 %r2938, %f15382; ld.local.f32 %f15383, [%rd7+-56]; mul.ftz.f32 %f15384, %f15383, 0f43800000; cvt.rzi.ftz.s32.f32 %r2939, %f15384; neg.s32 %r2940, %r2937; setp.lt.ftz.f32 %p2114, %f16770, 0f00000000; selp.b32 %r2941, %r2940, %r2937, %p2114; mov.b32 %r2942, %f16770; add.s32 %r2943, %r2941, %r2942; mov.b32 %f15385, %r2943; neg.s32 %r2944, %r2938; setp.lt.ftz.f32 %p2115, %f16771, 0f00000000; selp.b32 %r2945, %r2944, %r2938, %p2115; mov.b32 %r2946, %f16771; add.s32 %r2947, %r2945, %r2946; mov.b32 %f15386, %r2947; neg.s32 %r2948, %r2939; setp.lt.ftz.f32 %p2116, %f16772, 0f00000000; selp.b32 %r2949, %r2948, %r2939, %p2116; mov.b32 %r2950, %f16772; add.s32 %r2951, %r2949, %r2950; mov.b32 %f15387, %r2951; abs.ftz.f32 %f15388, %f16770; setp.lt.ftz.f32 %p2117, %f15388, 0f3D000000; fma.rn.ftz.f32 %f15389, %f15377, 0f37800000, %f16770; abs.ftz.f32 %f15390, %f16771; setp.lt.ftz.f32 %p2118, %f15390, 0f3D000000; fma.rn.ftz.f32 %f15391, %f15378, 0f37800000, %f16771; abs.ftz.f32 %f15392, %f16772; setp.lt.ftz.f32 %p2119, %f15392, 0f3D000000; fma.rn.ftz.f32 %f15393, %f15383, 0f37800000, %f16772; selp.f32 %f16772, %f15393, %f15387, %p2119; selp.f32 %f16771, %f15391, %f15386, %p2118; selp.f32 %f16770, %f15389, %f15385, %p2117; $L__BB106_2354: ld.const.u64 %rd2908, [kernel_params+2632]; cvta.to.global.u64 %rd2907, %rd2908; mul.lo.s64 %rd2725, %rd2, 12; add.s64 %rd2726, %rd2907, %rd2725; st.global.f32 [%rd2726], %f16770; st.global.f32 [%rd2726+4], %f16771; st.global.f32 [%rd2726+8], %f16772; ld.const.v2.u64 {%rd2727, %rd2728}, [kernel_params+2640]; cvta.to.global.u64 %rd2731, %rd2727; add.s64 %rd2732, %rd2731, %rd2725; st.global.f32 [%rd2732], %f4168; st.global.f32 [%rd2732+4], %f4169; st.global.f32 [%rd2732+8], %f4170; cvta.to.global.u64 %rd2733, %rd2728; add.s64 %rd2735, %rd2733, %rd837; mov.u32 %r2952, 0; st.global.u32 [%rd2735], %r2952; ld.const.v2.u64 {%rd2736, %rd2737}, [kernel_params+2656]; cvta.to.global.u64 %rd2739, %rd2736; add.s64 %rd2740, %rd2739, %rd837; mov.u32 %r2953, 2139095039; st.global.u32 [%rd2740], %r2953; ld.local.f32 %f15394, [%rd7+4]; ld.const.u64 %rd2741, [kernel_params+2672]; cvta.to.global.u64 %rd2742, %rd2741; add.s64 %rd2743, %rd2742, %rd837; st.global.f32 [%rd2743], %f15394; $L__BB106_2355: div.approx.ftz.f32 %f15395, %f4165, %f16738; ld.const.u64 %rd2744, [kernel_params+2584]; cvta.to.global.u64 %rd2745, %rd2744; mul.lo.s64 %rd2746, %rd2, 12; add.s64 %rd2747, %rd2745, %rd2746; ld.global.f32 %f15396, [%rd2747]; mul.ftz.f32 %f15397, %f15395, %f15396; ld.global.f32 %f15398, [%rd2747+4]; div.approx.ftz.f32 %f15399, %f4166, %f16738; mul.ftz.f32 %f15400, %f15399, %f15398; ld.global.f32 %f15401, [%rd2747+8]; div.approx.ftz.f32 %f15402, %f4167, %f16738; mul.ftz.f32 %f15403, %f15402, %f15401; st.global.f32 [%rd2747], %f15397; st.global.f32 [%rd2747+4], %f15400; st.global.f32 [%rd2747+8], %f15403; and.b32 %r2954, %r787, 2097152; setp.eq.s32 %p2120, %r2954, 0; @%p2120 bra $L__BB106_2370; ld.const.u64 %rd2748, [kernel_params+2464]; cvta.to.global.u64 %rd2749, %rd2748; add.s64 %rd2751, %rd2749, %rd1709; ld.global.u16 %rs125, [%rd2751]; setp.ne.s16 %p2121, %rs125, 0; @%p2121 bra $L__BB106_2370; ld.const.u64 %rd593, [kernel_params+2600]; ld.local.f32 %f4243, [%rd591+4]; ld.local.f32 %f4244, [%rd591+8]; mov.f32 %f16778, 0f00000000; mov.f32 %f16773, %f16778; @%p2098 bra $L__BB106_2359; ld.local.f32 %f15405, [%rd591]; div.approx.ftz.f32 %f16773, %f15405, %f4165; $L__BB106_2359: mov.f32 %f16774, %f16778; @%p2099 bra $L__BB106_2361; div.approx.ftz.f32 %f16774, %f4243, %f4166; $L__BB106_2361: mov.f32 %f16775, %f16778; @%p2101 bra $L__BB106_2363; div.approx.ftz.f32 %f16775, %f4244, %f4167; $L__BB106_2363: cvta.to.global.u64 %rd2752, %rd593; add.s64 %rd2754, %rd2752, %rd2746; st.global.f32 [%rd2754], %f16773; st.global.f32 [%rd2754+4], %f16774; st.global.f32 [%rd2754+8], %f16775; ld.const.u64 %rd594, [kernel_params+2608]; ld.local.f32 %f4251, [%rd591+16]; ld.local.f32 %f4252, [%rd591+20]; mov.f32 %f16776, %f16778; @%p2098 bra $L__BB106_2365; ld.local.f32 %f15409, [%rd591+12]; div.approx.ftz.f32 %f16776, %f15409, %f4165; $L__BB106_2365: mov.f32 %f16777, %f16778; @%p2099 bra $L__BB106_2367; div.approx.ftz.f32 %f16777, %f4251, %f4166; $L__BB106_2367: @%p2101 bra $L__BB106_2369; div.approx.ftz.f32 %f16778, %f4252, %f4167; $L__BB106_2369: cvta.to.global.u64 %rd2755, %rd594; add.s64 %rd2757, %rd2755, %rd2746; st.global.f32 [%rd2757], %f16776; st.global.f32 [%rd2757+4], %f16777; st.global.f32 [%rd2757+8], %f16778; $L__BB106_2370: setp.ne.s32 %p2128, %r908, 0; @%p2128 bra $L__BB106_2372; ld.const.v2.u64 {%rd2758, %rd2759}, [kernel_params+2560]; cvta.to.global.u64 %rd2762, %rd2758; add.s64 %rd2764, %rd2762, %rd837; st.global.f32 [%rd2764], %f16738; cvta.to.global.u64 %rd2765, %rd2759; add.s64 %rd2766, %rd2765, %rd837; ld.global.f32 %f15412, [%rd2766]; min.ftz.f32 %f15413, %f16738, %f15412; st.global.f32 [%rd2766], %f15413; $L__BB106_2372: ld.global.u32 %r910, [%rd29]; @%p2103 bra $L__BB106_2374; bra.uni $L__BB106_2373; $L__BB106_2374: ld.const.u64 %rd2771, [kernel_params+2464]; cvta.to.global.u64 %rd2772, %rd2771; add.s64 %rd595, %rd2772, %rd1709; ld.global.u16 %r2963, [%rd595]; add.s32 %r911, %r2963, 1; ld.const.u32 %r2964, [kernel_params+1516]; setp.lt.u32 %p2132, %r911, %r2964; or.b32 %r2965, %r910, 131072; selp.b32 %r912, %r910, %r2965, %p2132; and.b32 %r2966, %r3285, 64; setp.eq.s32 %p2133, %r2966, 0; @%p2133 bra $L__BB106_2376; bra.uni $L__BB106_2375; $L__BB106_2376: and.b32 %r2975, %r3285, 2; setp.eq.s32 %p2136, %r2975, 0; and.b32 %r3287, %r3285, 4; @%p2136 bra $L__BB106_2380; bra.uni $L__BB106_2377; $L__BB106_2380: and.b32 %r2984, %r912, -4096; or.b32 %r2985, %r2984, 4; and.b32 %r2986, %r3285, 128; setp.eq.s32 %p2140, %r2986, 0; and.b32 %r2987, %r2985, -20476; selp.b32 %r2988, %r2987, %r2985, %p2140; ld.const.u64 %rd2786, [kernel_params+2496]; cvta.to.global.u64 %rd2787, %rd2786; add.s64 %rd2789, %rd2787, %rd1709; ld.global.u16 %r2989, [%rd2789]; add.s32 %r2990, %r2989, 1; st.global.u16 [%rd2789], %r2990; or.b32 %r2991, %r2988, 131072; ld.const.u32 %r2992, [kernel_params+1528]; setp.lt.s32 %p2141, %r2990, %r2992; selp.b32 %r3288, %r2988, %r2991, %p2141; bra.uni $L__BB106_2381; $L__BB106_2373: ld.const.u64 %rd2767, [kernel_params+2472]; cvta.to.global.u64 %rd2768, %rd2767; add.s64 %rd2770, %rd2768, %rd1709; ld.global.u16 %r2955, [%rd2770]; add.s32 %r2956, %r2955, 1; ld.const.u32 %r2957, [kernel_params+1556]; setp.lt.u32 %p2130, %r2956, %r2957; selp.b32 %r2958, 64, 32832, %p2130; or.b32 %r2959, %r2958, %r910; ld.const.u32 %r2960, [kernel_params+1560]; setp.eq.s32 %p2131, %r2960, 0; or.b32 %r2961, %r2959, 2048; selp.b32 %r2962, %r2961, %r2959, %p2131; st.global.u32 [%rd29], %r2962; st.global.u16 [%rd2770], %r2956; ld.global.u16 %rs126, [%rd278]; add.s16 %rs127, %rs126, 16; st.global.u16 [%rd278], %rs127; mov.u32 %r3294, %r3285; bra.uni $L__BB106_2405; $L__BB106_2375: and.b32 %r2967, %r912, -20480; and.b32 %r2968, %r912, 100663296; setp.eq.s32 %p2134, %r2968, 0; selp.b32 %r2969, 67108992, 128, %p2134; or.b32 %r2970, %r2969, %r2967; ld.const.u64 %rd2774, [kernel_params+2504]; cvta.to.global.u64 %rd2775, %rd2774; add.s64 %rd2777, %rd2775, %rd1709; ld.global.u16 %r2971, [%rd2777]; add.s32 %r2972, %r2971, 1; st.global.u16 [%rd2777], %r2972; ld.const.u32 %r2973, [kernel_params+1532]; setp.lt.s32 %p2135, %r2972, %r2973; or.b32 %r2974, %r2970, 131072; selp.b32 %r3289, %r2970, %r2974, %p2135; bra.uni $L__BB106_2382; $L__BB106_2377: and.b32 %r2976, %r912, -20480; or.b32 %r915, %r2976, 2; or.b32 %r916, %r2976, 131074; setp.eq.s32 %p2137, %r3287, 0; @%p2137 bra $L__BB106_2379; bra.uni $L__BB106_2378; $L__BB106_2379: ld.const.u64 %rd2782, [kernel_params+2488]; cvta.to.global.u64 %rd2783, %rd2782; add.s64 %rd2785, %rd2783, %rd1709; ld.global.u16 %r2981, [%rd2785]; add.s32 %r2982, %r2981, 1; st.global.u16 [%rd2785], %r2982; ld.const.u32 %r2983, [kernel_params+1524]; setp.lt.s32 %p2139, %r2982, %r2983; selp.b32 %r3288, %r915, %r916, %p2139; mov.u32 %r3287, 0; bra.uni $L__BB106_2381; $L__BB106_2378: ld.const.u64 %rd2778, [kernel_params+2480]; cvta.to.global.u64 %rd2779, %rd2778; add.s64 %rd2781, %rd2779, %rd1709; ld.global.u16 %r2977, [%rd2781]; add.s32 %r2978, %r2977, 1; st.global.u16 [%rd2781], %r2978; ld.const.u32 %r2979, [kernel_params+1520]; setp.lt.s32 %p2138, %r2978, %r2979; selp.b32 %r3288, %r915, %r916, %p2138; $L__BB106_2381: and.b32 %r2993, %r3285, 8; setp.eq.s32 %p2142, %r2993, 0; selp.b32 %r2994, 2096, 16, %p2142; setp.eq.s32 %p2143, %r3287, 0; selp.b32 %r2995, %r2994, 4104, %p2143; or.b32 %r2996, %r2995, %r3288; and.b32 %r2997, %r3288, 100679680; setp.eq.s32 %p2144, %r2997, 0; or.b32 %r2998, %r2996, 33554432; selp.b32 %r3289, %r2998, %r2996, %p2144; $L__BB106_2382: st.global.u32 [%rd29], %r3289; add.s32 %r3039, %r2963, 1; st.global.u16 [%rd595], %r3039; ld.global.u16 %rs128, [%rd278]; add.s16 %rs129, %rs128, 16; st.global.u16 [%rd278], %rs129; mov.u32 %r3294, %r3285; $L__BB106_2405: and.b32 %r3022, %r3294, 1; setp.eq.b32 %p2159, %r3022, 1; mov.pred %p2160, 0; xor.pred %p2161, %p2159, %p2160; not.pred %p2162, %p2161; @%p2162 bra $L__BB106_2420; ld.local.u32 %r936, [%rd7+-36]; and.b32 %r3023, %r936, 262144; setp.eq.s32 %p2163, %r3023, 0; @%p2163 bra $L__BB106_2420; bra.uni $L__BB106_2407; $L__BB106_2420: setp.ne.s32 %p2180, %r3294, 0; selp.u16 %rs161, 1, 0, %p2180; $L__BB106_2421: cvt.u32.u16 %r3038, %rs161; st.param.b32 [func_retval0+0], %r3038; ret; $L__BB106_2407: and.b32 %r3024, %r936, 1; setp.eq.b32 %p2164, %r3024, 1; xor.pred %p2166, %p2164, %p2160; not.pred %p2167, %p2166; ld.const.u64 %rd3071, [kernel_params+2776]; cvta.to.global.u64 %rd2844, %rd3071; add.s64 %rd2846, %rd2844, %rd837; ld.global.u32 %r3296, [%rd2846]; @%p2167 bra $L__BB106_2414; bra.uni $L__BB106_2408; $L__BB106_2414: mov.u64 %rd2877, kernel_params; add.s64 %rd3077, %rd2877, 2768; setp.eq.s32 %p2174, %r3296, -1; mov.u32 %r3301, 0; @%p2174 bra $L__BB106_2418; ld.local.u32 %r946, [%rd7+-12]; add.s64 %rd3077, %rd2877, 2768; ld.local.u32 %r947, [%rd7+-40]; mov.u32 %r3301, 0; $L__BB106_2416: ld.const.u64 %rd2880, [%rd3077]; cvta.to.global.u64 %rd2881, %rd2880; add.s64 %rd2883, %rd2881, %rd837; ld.global.u32 %r3031, [%rd2883]; setp.eq.s32 %p2175, %r3031, %r946; setp.eq.s32 %p2176, %r3296, %r947; and.pred %p2177, %p2175, %p2176; @%p2177 bra $L__BB106_2420; add.s32 %r3301, %r3301, 1; mul.wide.s32 %rd2884, %r3301, 16; add.s64 %rd2886, %rd2877, %rd2884; add.s64 %rd3077, %rd2886, 2768; ld.const.u64 %rd3071, [%rd2886+2776]; cvta.to.global.u64 %rd2887, %rd3071; add.s64 %rd2889, %rd2887, %rd837; ld.global.u32 %r3296, [%rd2889]; setp.ne.s32 %p2178, %r3296, -1; @%p2178 bra $L__BB106_2416; $L__BB106_2418: ld.const.u32 %r3032, [kernel_params+12]; add.s32 %r3033, %r3032, -1; setp.ge.u32 %p2179, %r3301, %r3033; @%p2179 bra $L__BB106_2420; ld.local.u32 %r3034, [%rd7+-12]; ld.local.u32 %r3035, [%rd7+-40]; ld.const.u64 %rd2890, [%rd3077]; cvta.to.global.u64 %rd2891, %rd2890; add.s64 %rd2893, %rd2891, %rd837; st.global.u32 [%rd2893], %r3034; cvta.to.global.u64 %rd2894, %rd3071; add.s64 %rd2895, %rd2894, %rd837; st.global.u32 [%rd2895], %r3035; add.s32 %r3036, %r3301, 1; mul.wide.s32 %rd2896, %r3036, 16; add.s64 %rd2898, %rd2877, %rd2896; ld.const.v2.u64 {%rd2899, %rd2900}, [%rd2898+2768]; cvta.to.global.u64 %rd2903, %rd2899; add.s64 %rd2904, %rd2903, %rd837; mov.u32 %r3037, -1; st.global.u32 [%rd2904], %r3037; cvta.to.global.u64 %rd2905, %rd2900; add.s64 %rd2906, %rd2905, %rd837; st.global.u32 [%rd2906], %r3037; bra.uni $L__BB106_2420; $L__BB106_2408: setp.eq.s32 %p2168, %r3296, -1; @%p2168 bra $L__BB106_2420; ld.local.u32 %r938, [%rd7+-12]; mov.u64 %rd2848, kernel_params; add.s64 %rd3072, %rd2848, 2768; ld.local.u32 %r939, [%rd7+-40]; mov.u32 %r3297, 0; $L__BB106_2410: ld.const.u64 %rd2849, [%rd3072]; cvta.to.global.u64 %rd2850, %rd2849; add.s64 %rd2852, %rd2850, %rd837; ld.global.u32 %r3026, [%rd2852]; setp.eq.s32 %p2169, %r3026, %r938; setp.eq.s32 %p2170, %r3296, %r939; and.pred %p2171, %p2169, %p2170; add.s32 %r942, %r3297, 1; mul.wide.s32 %rd2853, %r942, 16; add.s64 %rd2855, %rd2848, %rd2853; add.s64 %rd605, %rd2855, 2776; @%p2171 bra $L__BB106_2412; add.s64 %rd3072, %rd2855, 2768; ld.const.u64 %rd3071, [%rd605]; cvta.to.global.u64 %rd2859, %rd3071; add.s64 %rd2861, %rd2859, %rd837; ld.global.u32 %r3296, [%rd2861]; setp.eq.s32 %p2172, %r3296, -1; mov.u32 %r3297, %r942; @%p2172 bra $L__BB106_2420; bra.uni $L__BB106_2410; $L__BB106_2412: ld.const.u64 %rd3074, [%rd605+-24]; $L__BB106_2413: add.s32 %r3297, %r3297, 1; mul.wide.s32 %rd2862, %r3297, 16; add.s64 %rd2864, %rd2848, %rd2862; ld.const.v2.u64 {%rd2865, %rd2866}, [%rd2864+2768]; cvta.to.global.u64 %rd2867, %rd2865; add.s64 %rd2869, %rd2867, %rd837; ld.global.u32 %r3027, [%rd2869]; cvta.to.global.u64 %rd2870, %rd2866; add.s64 %rd2871, %rd2870, %rd837; ld.global.u32 %r3028, [%rd2871]; cvta.to.global.u64 %rd2872, %rd3074; add.s64 %rd2873, %rd2872, %rd837; st.global.u32 [%rd2873], %r3027; cvta.to.global.u64 %rd2874, %rd3071; add.s64 %rd2875, %rd2874, %rd837; st.global.u32 [%rd2875], %r3028; setp.eq.s32 %p2173, %r3028, -1; mov.u64 %rd3071, %rd2866; mov.u64 %rd3074, %rd2865; @%p2173 bra $L__BB106_2420; bra.uni $L__BB106_2413; $L__BB106_741: mov.f32 %f6498, 0fBF800000; max.ftz.f32 %f6499, %f1019, %f6498; min.ftz.f32 %f6501, %f6499, %f6432; abs.ftz.f32 %f6502, %f6501; sub.ftz.f32 %f6503, %f6432, %f6502; mul.ftz.f32 %f6504, %f6503, 0f3F000000; sqrt.approx.ftz.f32 %f6505, %f6504; setp.gt.ftz.f32 %p633, %f6502, 0f3F11EB85; selp.f32 %f6506, %f6505, %f6502, %p633; mul.ftz.f32 %f6507, %f6506, %f6506; mov.f32 %f6508, 0f3C94D2E9; mov.f32 %f6509, 0f3D53F941; fma.rn.ftz.f32 %f6510, %f6509, %f6507, %f6508; mov.f32 %f6511, 0f3D3F841F; fma.rn.ftz.f32 %f6512, %f6510, %f6507, %f6511; mov.f32 %f6513, 0f3D994929; fma.rn.ftz.f32 %f6514, %f6512, %f6507, %f6513; mov.f32 %f6515, 0f3E2AAB94; fma.rn.ftz.f32 %f6516, %f6514, %f6507, %f6515; mul.ftz.f32 %f6517, %f6507, %f6516; fma.rn.ftz.f32 %f6518, %f6517, %f6506, %f6506; add.ftz.f32 %f6519, %f6518, %f6518; mov.f32 %f6520, 0f3FC90FDB; sub.ftz.f32 %f6521, %f6520, %f6518; selp.f32 %f6522, %f6519, %f6521, %p633; setp.lt.ftz.f32 %p634, %f6501, 0f00000000; sub.ftz.f32 %f6524, %f6444, %f6522; selp.f32 %f6525, %f6524, %f6522, %p634; mul.ftz.f32 %f6526, %f6485, %f1019; sub.ftz.f32 %f6527, %f6489, %f6526; mul.ftz.f32 %f6528, %f6486, %f1019; sub.ftz.f32 %f6529, %f6490, %f6528; mul.ftz.f32 %f6530, %f6487, %f1019; sub.ftz.f32 %f6531, %f6491, %f6530; mul.ftz.f32 %f6532, %f6488, %f1019; sub.ftz.f32 %f6533, %f6492, %f6532; mul.ftz.f32 %f6534, %f6529, %f6529; fma.rn.ftz.f32 %f6535, %f6527, %f6527, %f6534; mul.ftz.f32 %f6536, %f6533, %f6533; fma.rn.ftz.f32 %f6537, %f6531, %f6531, %f6536; add.ftz.f32 %f6538, %f6535, %f6537; rsqrt.approx.ftz.f32 %f6539, %f6538; mul.ftz.f32 %f6540, %f6527, %f6539; mul.ftz.f32 %f6541, %f6529, %f6539; mul.ftz.f32 %f6542, %f6531, %f6539; mul.ftz.f32 %f6543, %f6533, %f6539; mul.ftz.f32 %f6544, %f1010, %f6525; cos.approx.ftz.f32 %f6545, %f6544; sin.approx.ftz.f32 %f6546, %f6544; mul.ftz.f32 %f6547, %f6540, %f6546; mul.ftz.f32 %f6548, %f6541, %f6546; mul.ftz.f32 %f6549, %f6542, %f6546; mul.ftz.f32 %f6550, %f6543, %f6546; fma.rn.ftz.f32 %f15856, %f6488, %f6545, %f6550; fma.rn.ftz.f32 %f15855, %f6487, %f6545, %f6549; fma.rn.ftz.f32 %f15854, %f6486, %f6545, %f6548; fma.rn.ftz.f32 %f15853, %f6485, %f6545, %f6547; $L__BB106_743: ld.global.v4.f32 {%f6565, %f6566, %f6567, %f6568}, [%rd164+16]; ld.global.v4.f32 {%f6573, %f6574, %f6575, %f6576}, [%rd164+80]; mul.ftz.f32 %f6581, %f1010, %f6573; mul.ftz.f32 %f6582, %f1010, %f6574; mul.ftz.f32 %f6583, %f1010, %f6575; mul.ftz.f32 %f6584, %f1010, %f6576; fma.rn.ftz.f32 %f6585, %f1020, %f6568, %f6584; ld.global.v4.f32 {%f6586, %f6587, %f6588, %f6589}, [%rd164+32]; ld.global.v4.f32 {%f6594, %f6595, %f6596, %f6597}, [%rd164+96]; mul.ftz.f32 %f6602, %f1010, %f6594; mul.ftz.f32 %f6603, %f1010, %f6595; mul.ftz.f32 %f6604, %f1010, %f6596; mul.ftz.f32 %f6605, %f1010, %f6597; fma.rn.ftz.f32 %f6606, %f1020, %f6586, %f6602; fma.rn.ftz.f32 %f6607, %f1020, %f6587, %f6603; fma.rn.ftz.f32 %f6608, %f1020, %f6588, %f6604; fma.rn.ftz.f32 %f6609, %f1020, %f6589, %f6605; ld.global.v4.f32 {%f6610, %f6611, %f6612, %f6613}, [%rd164+48]; ld.global.v4.f32 {%f6618, %f6619, %f6620, %f6621}, [%rd164+112]; mul.ftz.f32 %f6626, %f1010, %f6618; mul.ftz.f32 %f6627, %f1010, %f6619; mul.ftz.f32 %f6628, %f1010, %f6620; mul.ftz.f32 %f6629, %f1010, %f6621; fma.rn.ftz.f32 %f6630, %f1020, %f6610, %f6626; fma.rn.ftz.f32 %f6631, %f1020, %f6611, %f6627; fma.rn.ftz.f32 %f6632, %f1020, %f6612, %f6628; fma.rn.ftz.f32 %f6633, %f1020, %f6613, %f6629; mul.ftz.f32 %f6634, %f15856, 0f3FB504F3; mul.ftz.f32 %f6635, %f15853, 0f3FB504F3; mul.ftz.f32 %f6636, %f15854, 0f3FB504F3; mul.ftz.f32 %f6637, %f15855, 0f3FB504F3; mul.ftz.f32 %f6638, %f6634, %f6635; mul.ftz.f32 %f6639, %f6634, %f6636; mul.ftz.f32 %f6640, %f6634, %f6637; mul.ftz.f32 %f6641, %f6635, %f6635; mul.ftz.f32 %f6642, %f6635, %f6636; mul.ftz.f32 %f6643, %f6635, %f6637; mul.ftz.f32 %f6644, %f6636, %f6636; mul.ftz.f32 %f6645, %f6636, %f6637; mul.ftz.f32 %f6646, %f6637, %f6637; sub.ftz.f32 %f6648, %f6432, %f6644; sub.ftz.f32 %f6649, %f6648, %f6646; sub.ftz.f32 %f6650, %f6642, %f6640; add.ftz.f32 %f6651, %f6639, %f6643; add.ftz.f32 %f6652, %f6640, %f6642; sub.ftz.f32 %f6653, %f6432, %f6641; sub.ftz.f32 %f6654, %f6653, %f6646; sub.ftz.f32 %f6655, %f6645, %f6638; sub.ftz.f32 %f6656, %f6643, %f6639; add.ftz.f32 %f6657, %f6638, %f6645; sub.ftz.f32 %f6658, %f6653, %f6644; mul.ftz.f32 %f6659, %f6650, %f6608; fma.rn.ftz.f32 %f6660, %f6649, %f6585, %f6659; mul.ftz.f32 %f6661, %f6650, %f6609; fma.rn.ftz.f32 %f6662, %f6649, %f6606, %f6661; mul.ftz.f32 %f6663, %f6650, %f6630; fma.rn.ftz.f32 %f6664, %f6649, %f6607, %f6663; mul.ftz.f32 %f6665, %f6654, %f6608; fma.rn.ftz.f32 %f6666, %f6652, %f6585, %f6665; mul.ftz.f32 %f6667, %f6654, %f6609; fma.rn.ftz.f32 %f6668, %f6652, %f6606, %f6667; mul.ftz.f32 %f6669, %f6654, %f6630; fma.rn.ftz.f32 %f6670, %f6652, %f6607, %f6669; mul.ftz.f32 %f6671, %f6657, %f6608; fma.rn.ftz.f32 %f6672, %f6656, %f6585, %f6671; mul.ftz.f32 %f6673, %f6657, %f6609; fma.rn.ftz.f32 %f6674, %f6656, %f6606, %f6673; mul.ftz.f32 %f6675, %f6657, %f6630; fma.rn.ftz.f32 %f6676, %f6656, %f6607, %f6675; fma.rn.ftz.f32 %f15863, %f6655, %f6633, %f6670; fma.rn.ftz.f32 %f15862, %f6655, %f6632, %f6668; fma.rn.ftz.f32 %f15861, %f6655, %f6631, %f6666; fma.rn.ftz.f32 %f15867, %f6658, %f6633, %f6676; fma.rn.ftz.f32 %f15866, %f6658, %f6632, %f6674; fma.rn.ftz.f32 %f15865, %f6658, %f6631, %f6672; fma.rn.ftz.f32 %f15859, %f6651, %f6633, %f6664; fma.rn.ftz.f32 %f15858, %f6651, %f6632, %f6662; fma.rn.ftz.f32 %f15857, %f6651, %f6631, %f6660; fma.rn.ftz.f32 %f15868, %f1020, %f6567, %f6583; fma.rn.ftz.f32 %f15864, %f1020, %f6566, %f6582; fma.rn.ftz.f32 %f15860, %f1020, %f6565, %f6581; $L__BB106_745: mul.ftz.f32 %f6689, %f15870, %f15858; fma.rn.ftz.f32 %f6690, %f15869, %f15857, %f6689; fma.rn.ftz.f32 %f6691, %f15871, %f15859, %f6690; add.ftz.f32 %f1069, %f15860, %f6691; mul.ftz.f32 %f6692, %f15870, %f15862; fma.rn.ftz.f32 %f6693, %f15869, %f15861, %f6692; fma.rn.ftz.f32 %f6694, %f15871, %f15863, %f6693; add.ftz.f32 %f1070, %f15864, %f6694; mul.ftz.f32 %f6695, %f15870, %f15866; fma.rn.ftz.f32 %f6696, %f15869, %f15865, %f6695; fma.rn.ftz.f32 %f6697, %f15871, %f15867, %f6696; add.ftz.f32 %f15871, %f15868, %f6697; mul.ftz.f32 %f6698, %f15873, %f15858; fma.rn.ftz.f32 %f6699, %f15872, %f15857, %f6698; fma.rn.ftz.f32 %f6700, %f15874, %f15859, %f6699; add.ftz.f32 %f1072, %f15860, %f6700; mul.ftz.f32 %f6701, %f15873, %f15862; fma.rn.ftz.f32 %f6702, %f15872, %f15861, %f6701; fma.rn.ftz.f32 %f6703, %f15874, %f15863, %f6702; add.ftz.f32 %f1073, %f15864, %f6703; mul.ftz.f32 %f6704, %f15873, %f15866; fma.rn.ftz.f32 %f6705, %f15872, %f15865, %f6704; fma.rn.ftz.f32 %f6706, %f15874, %f15867, %f6705; add.ftz.f32 %f15874, %f15868, %f6706; mul.ftz.f32 %f6707, %f15876, %f15858; fma.rn.ftz.f32 %f6708, %f15875, %f15857, %f6707; fma.rn.ftz.f32 %f6709, %f15877, %f15859, %f6708; add.ftz.f32 %f1075, %f15860, %f6709; mul.ftz.f32 %f6710, %f15876, %f15862; fma.rn.ftz.f32 %f6711, %f15875, %f15861, %f6710; fma.rn.ftz.f32 %f6712, %f15877, %f15863, %f6711; add.ftz.f32 %f1076, %f15864, %f6712; mul.ftz.f32 %f6713, %f15876, %f15866; fma.rn.ftz.f32 %f6714, %f15875, %f15865, %f6713; fma.rn.ftz.f32 %f6715, %f15877, %f15867, %f6714; add.ftz.f32 %f15877, %f15868, %f6715; mov.f32 %f15869, %f1069; mov.f32 %f15870, %f1070; mov.f32 %f15872, %f1072; mov.f32 %f15873, %f1073; mov.f32 %f15875, %f1075; mov.f32 %f15876, %f1076; $L__BB106_746: sub.ftz.f32 %f6716, %f15871, %f15874; sub.ftz.f32 %f6717, %f15876, %f15873; mul.ftz.f32 %f6718, %f6716, %f6717; sub.ftz.f32 %f6719, %f15870, %f15873; sub.ftz.f32 %f6720, %f15877, %f15874; mul.ftz.f32 %f6721, %f6719, %f6720; sub.ftz.f32 %f6722, %f6718, %f6721; sub.ftz.f32 %f6723, %f15869, %f15872; mul.ftz.f32 %f6724, %f6723, %f6720; sub.ftz.f32 %f6725, %f15875, %f15872; mul.ftz.f32 %f6726, %f6716, %f6725; sub.ftz.f32 %f6727, %f6724, %f6726; mul.ftz.f32 %f6728, %f6719, %f6725; mul.ftz.f32 %f6729, %f6723, %f6717; sub.ftz.f32 %f6730, %f6728, %f6729; mul.ftz.f32 %f6731, %f6727, %f6727; fma.rn.ftz.f32 %f6732, %f6722, %f6722, %f6731; fma.rn.ftz.f32 %f6733, %f6730, %f6730, %f6732; sqrt.approx.ftz.f32 %f15878, %f6733; $L__BB106_748: ld.const.f32 %f6734, [kernel_params+1500]; mul.ftz.f32 %f6735, %f15878, 0f3F000000; mul.ftz.f32 %f6736, %f6735, %f6734; div.approx.ftz.f32 %f15879, %f6736, %f1000; bra.uni $L__BB106_749; $L__BB106_724: mov.f32 %f6148, 0fBF800000; max.ftz.f32 %f6149, %f877, %f6148; min.ftz.f32 %f6151, %f6149, %f6147; abs.ftz.f32 %f6152, %f6151; sub.ftz.f32 %f6153, %f6147, %f6152; mul.ftz.f32 %f6154, %f6153, 0f3F000000; sqrt.approx.ftz.f32 %f6155, %f6154; setp.gt.ftz.f32 %p617, %f6152, 0f3F11EB85; selp.f32 %f6156, %f6155, %f6152, %p617; mul.ftz.f32 %f6157, %f6156, %f6156; mov.f32 %f6158, 0f3C94D2E9; mov.f32 %f6159, 0f3D53F941; fma.rn.ftz.f32 %f6160, %f6159, %f6157, %f6158; mov.f32 %f6161, 0f3D3F841F; fma.rn.ftz.f32 %f6162, %f6160, %f6157, %f6161; mov.f32 %f6163, 0f3D994929; fma.rn.ftz.f32 %f6164, %f6162, %f6157, %f6163; mov.f32 %f6165, 0f3E2AAB94; fma.rn.ftz.f32 %f6166, %f6164, %f6157, %f6165; mul.ftz.f32 %f6167, %f6157, %f6166; fma.rn.ftz.f32 %f6168, %f6167, %f6156, %f6156; add.ftz.f32 %f6169, %f6168, %f6168; mov.f32 %f6170, 0f3FC90FDB; sub.ftz.f32 %f6171, %f6170, %f6168; selp.f32 %f6172, %f6169, %f6171, %p617; setp.lt.ftz.f32 %p618, %f6151, 0f00000000; mov.f32 %f6173, 0f40490FDB; sub.ftz.f32 %f6174, %f6173, %f6172; selp.f32 %f6175, %f6174, %f6172, %p618; mul.ftz.f32 %f6176, %f6135, %f877; sub.ftz.f32 %f6177, %f6139, %f6176; mul.ftz.f32 %f6178, %f6136, %f877; sub.ftz.f32 %f6179, %f6140, %f6178; mul.ftz.f32 %f6180, %f6137, %f877; sub.ftz.f32 %f6181, %f6141, %f6180; mul.ftz.f32 %f6182, %f6138, %f877; sub.ftz.f32 %f6183, %f6142, %f6182; mul.ftz.f32 %f6184, %f6179, %f6179; fma.rn.ftz.f32 %f6185, %f6177, %f6177, %f6184; mul.ftz.f32 %f6186, %f6183, %f6183; fma.rn.ftz.f32 %f6187, %f6181, %f6181, %f6186; add.ftz.f32 %f6188, %f6185, %f6187; rsqrt.approx.ftz.f32 %f6189, %f6188; mul.ftz.f32 %f6190, %f6177, %f6189; mul.ftz.f32 %f6191, %f6179, %f6189; mul.ftz.f32 %f6192, %f6181, %f6189; mul.ftz.f32 %f6193, %f6183, %f6189; mul.ftz.f32 %f6194, %f868, %f6175; cos.approx.ftz.f32 %f6195, %f6194; sin.approx.ftz.f32 %f6196, %f6194; mul.ftz.f32 %f6197, %f6190, %f6196; mul.ftz.f32 %f6198, %f6191, %f6196; mul.ftz.f32 %f6199, %f6192, %f6196; mul.ftz.f32 %f6200, %f6193, %f6196; fma.rn.ftz.f32 %f15822, %f6138, %f6195, %f6200; fma.rn.ftz.f32 %f15821, %f6137, %f6195, %f6199; fma.rn.ftz.f32 %f15820, %f6136, %f6195, %f6198; fma.rn.ftz.f32 %f15819, %f6135, %f6195, %f6197; $L__BB106_726: ld.global.v4.f32 {%f6215, %f6216, %f6217, %f6218}, [%rd162+16]; ld.global.v4.f32 {%f6223, %f6224, %f6225, %f6226}, [%rd162+80]; mul.ftz.f32 %f6231, %f868, %f6223; mul.ftz.f32 %f6232, %f868, %f6224; mul.ftz.f32 %f6233, %f868, %f6225; mul.ftz.f32 %f6234, %f868, %f6226; fma.rn.ftz.f32 %f6235, %f878, %f6218, %f6234; ld.global.v4.f32 {%f6236, %f6237, %f6238, %f6239}, [%rd162+32]; ld.global.v4.f32 {%f6244, %f6245, %f6246, %f6247}, [%rd162+96]; mul.ftz.f32 %f6252, %f868, %f6244; mul.ftz.f32 %f6253, %f868, %f6245; mul.ftz.f32 %f6254, %f868, %f6246; mul.ftz.f32 %f6255, %f868, %f6247; fma.rn.ftz.f32 %f6256, %f878, %f6236, %f6252; fma.rn.ftz.f32 %f6257, %f878, %f6237, %f6253; fma.rn.ftz.f32 %f6258, %f878, %f6238, %f6254; fma.rn.ftz.f32 %f6259, %f878, %f6239, %f6255; ld.global.v4.f32 {%f6260, %f6261, %f6262, %f6263}, [%rd162+48]; ld.global.v4.f32 {%f6268, %f6269, %f6270, %f6271}, [%rd162+112]; mul.ftz.f32 %f6276, %f868, %f6268; mul.ftz.f32 %f6277, %f868, %f6269; mul.ftz.f32 %f6278, %f868, %f6270; mul.ftz.f32 %f6279, %f868, %f6271; fma.rn.ftz.f32 %f6280, %f878, %f6260, %f6276; fma.rn.ftz.f32 %f6281, %f878, %f6261, %f6277; fma.rn.ftz.f32 %f6282, %f878, %f6262, %f6278; fma.rn.ftz.f32 %f6283, %f878, %f6263, %f6279; mul.ftz.f32 %f6284, %f15822, 0f3FB504F3; mul.ftz.f32 %f6285, %f15819, 0f3FB504F3; mul.ftz.f32 %f6286, %f15820, 0f3FB504F3; mul.ftz.f32 %f6287, %f15821, 0f3FB504F3; mul.ftz.f32 %f6288, %f6284, %f6285; mul.ftz.f32 %f6289, %f6284, %f6286; mul.ftz.f32 %f6290, %f6284, %f6287; mul.ftz.f32 %f6291, %f6285, %f6285; mul.ftz.f32 %f6292, %f6285, %f6286; mul.ftz.f32 %f6293, %f6285, %f6287; mul.ftz.f32 %f6294, %f6286, %f6286; mul.ftz.f32 %f6295, %f6286, %f6287; mul.ftz.f32 %f6296, %f6287, %f6287; sub.ftz.f32 %f6298, %f6147, %f6294; sub.ftz.f32 %f6299, %f6298, %f6296; sub.ftz.f32 %f6300, %f6292, %f6290; add.ftz.f32 %f6301, %f6289, %f6293; add.ftz.f32 %f6302, %f6290, %f6292; sub.ftz.f32 %f6303, %f6147, %f6291; sub.ftz.f32 %f6304, %f6303, %f6296; sub.ftz.f32 %f6305, %f6295, %f6288; sub.ftz.f32 %f6306, %f6293, %f6289; add.ftz.f32 %f6307, %f6288, %f6295; sub.ftz.f32 %f6308, %f6303, %f6294; mul.ftz.f32 %f6309, %f6300, %f6258; fma.rn.ftz.f32 %f6310, %f6299, %f6235, %f6309; mul.ftz.f32 %f6311, %f6300, %f6259; fma.rn.ftz.f32 %f6312, %f6299, %f6256, %f6311; mul.ftz.f32 %f6313, %f6300, %f6280; fma.rn.ftz.f32 %f6314, %f6299, %f6257, %f6313; mul.ftz.f32 %f6315, %f6304, %f6258; fma.rn.ftz.f32 %f6316, %f6302, %f6235, %f6315; mul.ftz.f32 %f6317, %f6304, %f6259; fma.rn.ftz.f32 %f6318, %f6302, %f6256, %f6317; mul.ftz.f32 %f6319, %f6304, %f6280; fma.rn.ftz.f32 %f6320, %f6302, %f6257, %f6319; mul.ftz.f32 %f6321, %f6307, %f6258; fma.rn.ftz.f32 %f6322, %f6306, %f6235, %f6321; mul.ftz.f32 %f6323, %f6307, %f6259; fma.rn.ftz.f32 %f6324, %f6306, %f6256, %f6323; mul.ftz.f32 %f6325, %f6307, %f6280; fma.rn.ftz.f32 %f6326, %f6306, %f6257, %f6325; fma.rn.ftz.f32 %f15829, %f6305, %f6283, %f6320; fma.rn.ftz.f32 %f15828, %f6305, %f6282, %f6318; fma.rn.ftz.f32 %f15827, %f6305, %f6281, %f6316; fma.rn.ftz.f32 %f15833, %f6308, %f6283, %f6326; fma.rn.ftz.f32 %f15832, %f6308, %f6282, %f6324; fma.rn.ftz.f32 %f15831, %f6308, %f6281, %f6322; fma.rn.ftz.f32 %f15825, %f6301, %f6283, %f6314; fma.rn.ftz.f32 %f15824, %f6301, %f6282, %f6312; fma.rn.ftz.f32 %f15823, %f6301, %f6281, %f6310; fma.rn.ftz.f32 %f15834, %f878, %f6217, %f6233; fma.rn.ftz.f32 %f15830, %f878, %f6216, %f6232; fma.rn.ftz.f32 %f15826, %f878, %f6215, %f6231; $L__BB106_728: mul.ftz.f32 %f6339, %f15836, %f15824; fma.rn.ftz.f32 %f6340, %f15835, %f15823, %f6339; fma.rn.ftz.f32 %f6341, %f15837, %f15825, %f6340; add.ftz.f32 %f927, %f15826, %f6341; mul.ftz.f32 %f6342, %f15836, %f15828; fma.rn.ftz.f32 %f6343, %f15835, %f15827, %f6342; fma.rn.ftz.f32 %f6344, %f15837, %f15829, %f6343; add.ftz.f32 %f928, %f15830, %f6344; mul.ftz.f32 %f6345, %f15836, %f15832; fma.rn.ftz.f32 %f6346, %f15835, %f15831, %f6345; fma.rn.ftz.f32 %f6347, %f15837, %f15833, %f6346; add.ftz.f32 %f15837, %f15834, %f6347; mul.ftz.f32 %f6348, %f15839, %f15824; fma.rn.ftz.f32 %f6349, %f15838, %f15823, %f6348; fma.rn.ftz.f32 %f6350, %f15840, %f15825, %f6349; add.ftz.f32 %f930, %f15826, %f6350; mul.ftz.f32 %f6351, %f15839, %f15828; fma.rn.ftz.f32 %f6352, %f15838, %f15827, %f6351; fma.rn.ftz.f32 %f6353, %f15840, %f15829, %f6352; add.ftz.f32 %f931, %f15830, %f6353; mul.ftz.f32 %f6354, %f15839, %f15832; fma.rn.ftz.f32 %f6355, %f15838, %f15831, %f6354; fma.rn.ftz.f32 %f6356, %f15840, %f15833, %f6355; add.ftz.f32 %f15840, %f15834, %f6356; mul.ftz.f32 %f6357, %f15842, %f15824; fma.rn.ftz.f32 %f6358, %f15841, %f15823, %f6357; fma.rn.ftz.f32 %f6359, %f15843, %f15825, %f6358; add.ftz.f32 %f933, %f15826, %f6359; mul.ftz.f32 %f6360, %f15842, %f15828; fma.rn.ftz.f32 %f6361, %f15841, %f15827, %f6360; fma.rn.ftz.f32 %f6362, %f15843, %f15829, %f6361; add.ftz.f32 %f934, %f15830, %f6362; mul.ftz.f32 %f6363, %f15842, %f15832; fma.rn.ftz.f32 %f6364, %f15841, %f15831, %f6363; fma.rn.ftz.f32 %f6365, %f15843, %f15833, %f6364; add.ftz.f32 %f15843, %f15834, %f6365; mov.f32 %f15835, %f927; mov.f32 %f15836, %f928; mov.f32 %f15838, %f930; mov.f32 %f15839, %f931; mov.f32 %f15841, %f933; mov.f32 %f15842, %f934; $L__BB106_729: sub.ftz.f32 %f6366, %f15837, %f15840; sub.ftz.f32 %f6367, %f15842, %f15839; mul.ftz.f32 %f6368, %f6366, %f6367; sub.ftz.f32 %f6369, %f15836, %f15839; sub.ftz.f32 %f6370, %f15843, %f15840; mul.ftz.f32 %f6371, %f6369, %f6370; sub.ftz.f32 %f6372, %f6368, %f6371; sub.ftz.f32 %f6373, %f15835, %f15838; mul.ftz.f32 %f6374, %f6373, %f6370; sub.ftz.f32 %f6375, %f15841, %f15838; mul.ftz.f32 %f6376, %f6366, %f6375; sub.ftz.f32 %f6377, %f6374, %f6376; mul.ftz.f32 %f6378, %f6369, %f6375; mul.ftz.f32 %f6379, %f6373, %f6367; sub.ftz.f32 %f6380, %f6378, %f6379; mul.ftz.f32 %f6381, %f6377, %f6377; fma.rn.ftz.f32 %f6382, %f6372, %f6372, %f6381; fma.rn.ftz.f32 %f6383, %f6380, %f6380, %f6382; sqrt.approx.ftz.f32 %f6384, %f6383; mul.ftz.f32 %f6385, %f6384, 0f3F000000; mul.ftz.f32 %f6386, %f15818, %f6385; div.approx.ftz.f32 %f15879, %f6386, %f858; $L__BB106_749: mul.ftz.f32 %f6737, %f731, %f731; fma.rn.ftz.f32 %f6738, %f15879, %f15879, %f6737; div.approx.ftz.f32 %f15880, %f6737, %f6738; $L__BB106_750: ld.local.u32 %r381, [%rd7+-12]; setp.eq.s32 %p635, %r381, -1; mov.u32 %r3167, -1; @%p635 bra $L__BB106_752; ld.const.u64 %rd1222, [kernel_params+2160]; cvta.to.global.u64 %rd1223, %rd1222; mul.wide.s32 %rd1224, %r381, 208; add.s64 %rd1225, %rd1223, %rd1224; ld.global.u32 %r3167, [%rd1225+192]; $L__BB106_752: abs.ftz.f32 %f6740, %f727; setp.gt.ftz.f32 %p636, %f6740, 0f00000000; mov.f32 %f15889, 0f00000000; selp.f32 %f6741, 0f3F800000, 0f00000000, %p636; ld.const.u64 %rd1226, [kernel_params+2584]; cvta.to.global.u64 %rd1227, %rd1226; mul.lo.s64 %rd1228, %rd2, 12; add.s64 %rd1229, %rd1227, %rd1228; ld.global.f32 %f6742, [%rd1229]; mul.ftz.f32 %f6743, %f5822, %f6741; mul.ftz.f32 %f6744, %f6743, %f6742; ld.global.f32 %f6745, [%rd1229+4]; mul.ftz.f32 %f6746, %f6741, %f5823; mul.ftz.f32 %f6747, %f6746, %f6745; ld.global.f32 %f6748, [%rd1229+8]; mul.ftz.f32 %f6749, %f6741, %f730; mul.ftz.f32 %f6750, %f6749, %f6748; mul.ftz.f32 %f1094, %f15880, %f6744; mul.ftz.f32 %f1095, %f15880, %f6747; mul.ftz.f32 %f1096, %f15880, %f6750; ld.global.u16 %rs15, [%rd74]; mov.b32 %r384, %f1094; setp.nan.ftz.f32 %p637, %f1094, %f1094; mov.f32 %f15887, %f15889; @%p637 bra $L__BB106_755; or.b32 %r1422, %r384, -2147483648; setp.ne.s32 %p638, %r1422, -2147483648; add.ftz.f32 %f6752, %f1094, %f1094; setp.eq.ftz.f32 %p639, %f6752, %f1094; and.pred %p640, %p639, %p638; mov.f32 %f15887, %f15889; @%p640 bra $L__BB106_755; shl.b32 %r1423, %r384, 1; setp.lt.u32 %p641, %r1423, -16777215; selp.f32 %f15887, %f1094, 0f00000000, %p641; $L__BB106_755: mov.b32 %r385, %f1095; setp.nan.ftz.f32 %p642, %f1095, %f1095; mov.f32 %f15888, %f15889; @%p642 bra $L__BB106_758; or.b32 %r1424, %r385, -2147483648; setp.ne.s32 %p643, %r1424, -2147483648; add.ftz.f32 %f6755, %f1095, %f1095; setp.eq.ftz.f32 %p644, %f6755, %f1095; and.pred %p645, %p644, %p643; mov.f32 %f15888, %f15889; @%p645 bra $L__BB106_758; shl.b32 %r1425, %r385, 1; setp.lt.u32 %p646, %r1425, -16777215; selp.f32 %f15888, %f1095, 0f00000000, %p646; $L__BB106_758: mov.b32 %r386, %f1096; setp.nan.ftz.f32 %p647, %f1096, %f1096; @%p647 bra $L__BB106_761; or.b32 %r1426, %r386, -2147483648; setp.ne.s32 %p648, %r1426, -2147483648; add.ftz.f32 %f6758, %f1096, %f1096; setp.eq.ftz.f32 %p649, %f6758, %f1096; and.pred %p650, %p649, %p648; @%p650 bra $L__BB106_761; shl.b32 %r1427, %r386, 1; setp.lt.u32 %p651, %r1427, -16777215; selp.f32 %f15889, %f1096, 0f00000000, %p651; $L__BB106_761: ld.const.f32 %f6759, [kernel_params+1580]; ld.const.f32 %f6760, [kernel_params+1584]; setp.gt.u16 %p652, %rs15, 1; selp.f32 %f1103, %f6760, %f6759, %p652; abs.ftz.f32 %f6761, %f15888; abs.ftz.f32 %f6762, %f15887; add.ftz.f32 %f6763, %f6762, %f6761; abs.ftz.f32 %f6764, %f15889; add.ftz.f32 %f1104, %f6764, %f6763; setp.leu.ftz.f32 %p653, %f1104, %f1103; @%p653 bra $L__BB106_763; div.approx.ftz.f32 %f6765, %f1103, %f1104; mul.ftz.f32 %f15887, %f15887, %f6765; mul.ftz.f32 %f15888, %f15888, %f6765; mul.ftz.f32 %f15889, %f15889, %f6765; $L__BB106_763: ld.const.v2.u64 {%rd1230, %rd1231}, [kernel_params+2448]; cvta.to.global.u64 %rd1234, %rd1230; add.s64 %rd1236, %rd1234, %rd837; ld.global.u32 %rd1237, [%rd1236]; ld.const.s32 %rd1238, [kernel_params+1264]; mul.lo.s64 %rd165, %rd1238, %rd1237; cvta.to.global.u64 %rd1239, %rd1231; add.s64 %rd1240, %rd1239, %rd837; ld.global.u32 %r387, [%rd1240]; ld.const.u32 %r1428, [kernel_params+1620]; setp.eq.s32 %p654, %r1428, 0; @%p654 bra $L__BB106_768; and.b32 %r1429, %r360, 536870912; setp.ne.s32 %p655, %r1429, 0; @%p655 bra $L__BB106_766; ld.const.s32 %rd1241, [kernel_params+1384]; add.s64 %rd1242, %rd165, %rd1241; cvta.to.global.u64 %rd1243, %rd618; shl.b64 %rd1244, %rd1242, 2; add.s64 %rd1245, %rd1243, %rd1244; add.s64 %rd1246, %rd1245, 4; add.s64 %rd1247, %rd1245, 8; atom.global.add.f32 %f6766, [%rd1245], %f15887; atom.global.add.f32 %f6767, [%rd1246], %f15888; atom.global.add.f32 %f6768, [%rd1247], %f15889; $L__BB106_766: and.b32 %r1430, %r360, 1073741824; setp.eq.s32 %p656, %r1430, 0; @%p656 bra $L__BB106_768; bra.uni $L__BB106_767; $L__BB106_768: ld.const.u32 %r3168, [kernel_params+1260]; and.b32 %r1431, %r3168, 2; setp.eq.s32 %p657, %r1431, 0; @%p657 bra $L__BB106_770; ld.const.s32 %rd1255, [kernel_params+1268]; add.s64 %rd1256, %rd165, %rd1255; cvta.to.global.u64 %rd1257, %rd618; shl.b64 %rd1258, %rd1256, 2; add.s64 %rd1259, %rd1257, %rd1258; add.s64 %rd1260, %rd1259, 4; add.s64 %rd1261, %rd1259, 8; atom.global.add.f32 %f6772, [%rd1259], %f15887; atom.global.add.f32 %f6773, [%rd1260], %f15888; atom.global.add.f32 %f6774, [%rd1261], %f15889; $L__BB106_770: ld.const.u32 %r390, [kernel_params+1400]; setp.eq.s32 %p658, %r390, -1; @%p658 bra $L__BB106_773; and.b32 %r1432, %r387, -1431655766; popc.b32 %r1433, %r1432; and.b32 %r1434, %r1433, 1; setp.eq.b32 %p659, %r1434, 1; mov.pred %p660, 0; xor.pred %p661, %p659, %p660; not.pred %p662, %p661; @%p662 bra $L__BB106_773; cvt.s64.s32 %rd1262, %r390; add.s64 %rd1263, %rd165, %rd1262; cvta.to.global.u64 %rd1264, %rd618; shl.b64 %rd1265, %rd1263, 2; add.s64 %rd1266, %rd1264, %rd1265; add.s64 %rd1267, %rd1266, 4; add.s64 %rd1268, %rd1266, 8; add.s64 %rd1269, %rd1266, 12; add.ftz.f32 %f6775, %f15887, %f15887; atom.global.add.f32 %f6776, [%rd1266], %f6775; add.ftz.f32 %f6777, %f15888, %f15888; atom.global.add.f32 %f6778, [%rd1267], %f6777; add.ftz.f32 %f6779, %f15889, %f15889; atom.global.add.f32 %f6780, [%rd1268], %f6779; atom.global.add.f32 %f6781, [%rd1269], 0f00000000; bra.uni $L__BB106_773; $L__BB106_767: ld.const.s32 %rd1248, [kernel_params+1376]; add.s64 %rd1249, %rd165, %rd1248; cvta.to.global.u64 %rd1250, %rd618; shl.b64 %rd1251, %rd1249, 2; add.s64 %rd1252, %rd1250, %rd1251; add.s64 %rd1253, %rd1252, 4; add.s64 %rd1254, %rd1252, 8; atom.global.add.f32 %f6769, [%rd1252], %f15887; atom.global.add.f32 %f6770, [%rd1253], %f15888; atom.global.add.f32 %f6771, [%rd1254], %f15889; ld.const.u32 %r3168, [kernel_params+1260]; $L__BB106_773: setp.eq.s32 %p663, %r3168, 0; @%p663 bra $L__BB106_794; ld.const.u32 %r3169, [kernel_params+1352]; and.b32 %r1435, %r360, 16777216; setp.eq.s32 %p664, %r1435, 0; ld.const.u32 %r393, [kernel_params+1428]; setp.eq.s32 %p665, %r393, -1; or.pred %p666, %p664, %p665; @%p666 bra $L__BB106_776; ld.const.u64 %rd1270, [kernel_params+2616]; cvta.to.global.u64 %rd1271, %rd1270; add.s64 %rd1273, %rd1271, %rd1228; ld.global.f32 %f6782, [%rd1273]; mul.ftz.f32 %f6783, %f15887, %f6782; ld.global.f32 %f6784, [%rd1273+4]; mul.ftz.f32 %f6785, %f15888, %f6784; ld.global.f32 %f6786, [%rd1273+8]; mul.ftz.f32 %f6787, %f15889, %f6786; cvta.to.global.u64 %rd1274, %rd618; cvt.s64.s32 %rd1275, %r393; add.s64 %rd1276, %rd165, %rd1275; shl.b64 %rd1277, %rd1276, 2; add.s64 %rd1278, %rd1274, %rd1277; add.s64 %rd1279, %rd1278, 4; add.s64 %rd1280, %rd1278, 8; atom.global.add.f32 %f6788, [%rd1278], %f6783; atom.global.add.f32 %f6789, [%rd1279], %f6785; atom.global.add.f32 %f6790, [%rd1280], %f6787; $L__BB106_776: setp.eq.s32 %p667, %r3167, -1; ld.const.u32 %r394, [kernel_params+1444]; setp.eq.s32 %p668, %r394, -1; or.pred %p669, %p667, %p668; @%p669 bra $L__BB106_778; mul.lo.s32 %r1436, %r3167, 3; cvt.s64.s32 %rd1281, %r1436; add.s64 %rd1282, %rd165, %rd1281; cvt.s64.s32 %rd1283, %r394; add.s64 %rd1284, %rd1282, %rd1283; cvta.to.global.u64 %rd1285, %rd618; shl.b64 %rd1286, %rd1284, 2; add.s64 %rd1287, %rd1285, %rd1286; add.s64 %rd1288, %rd1287, 4; add.s64 %rd1289, %rd1287, 8; atom.global.add.f32 %f6791, [%rd1287], %f15887; atom.global.add.f32 %f6792, [%rd1288], %f15888; atom.global.add.f32 %f6793, [%rd1289], %f15889; $L__BB106_778: and.b32 %r1437, %r360, 100663296; setp.eq.s32 %p670, %r1437, 0; @%p670 bra $L__BB106_792; ld.const.u8 %rs74, [kernel_params+2]; and.b16 %rs75, %rs74, 32; setp.eq.s16 %p671, %rs75, 0; mov.u32 %r3169, -1; @%p671 bra $L__BB106_792; bra.uni $L__BB106_780; $L__BB106_792: setp.eq.s32 %p682, %r3169, -1; @%p682 bra $L__BB106_794; cvt.s64.s32 %rd1311, %r3169; add.s64 %rd1312, %rd165, %rd1311; cvta.to.global.u64 %rd1313, %rd618; shl.b64 %rd1314, %rd1312, 2; add.s64 %rd1315, %rd1313, %rd1314; add.s64 %rd1316, %rd1315, 4; add.s64 %rd1317, %rd1315, 8; atom.global.add.f32 %f6813, [%rd1315], %f15887; atom.global.add.f32 %f6814, [%rd1316], %f15888; atom.global.add.f32 %f6815, [%rd1317], %f15889; bra.uni $L__BB106_794; $L__BB106_780: and.b32 %r1439, %r360, 536870912; setp.ne.s32 %p672, %r1439, 0; @%p672 bra $L__BB106_794; and.b32 %r1440, %r360, 33554432; setp.eq.s32 %p673, %r1440, 0; @%p673 bra $L__BB106_788; bra.uni $L__BB106_782; $L__BB106_788: and.b32 %r1449, %r360, 67108864; setp.eq.s32 %p680, %r1449, 0; @%p680 bra $L__BB106_792; setp.eq.s16 %p681, %rs15, 1; @%p681 bra $L__BB106_791; bra.uni $L__BB106_790; $L__BB106_791: ld.const.u32 %r3169, [kernel_params+1348]; bra.uni $L__BB106_792; $L__BB106_794: cvt.u16.u32 %rs76, %r45; setp.lt.s16 %p683, %rs76, 0; mov.f32 %f15890, 0f00000000; @%p683 bra $L__BB106_796; ld.const.u64 %rd1318, [kernel_params+2576]; cvta.to.global.u64 %rd1319, %rd1318; add.s64 %rd1321, %rd1319, %rd837; ld.global.f32 %f15890, [%rd1321]; $L__BB106_796: setp.eq.ftz.f32 %p684, %f15890, 0f00000000; mov.u16 %rs161, 0; @%p684 bra $L__BB106_2421; setp.eq.ftz.f32 %p685, %f15890, 0f3F800000; @%p685 bra $L__BB106_799; ld.const.u64 %rd1322, [kernel_params+2584]; cvta.to.global.u64 %rd1323, %rd1322; mul.lo.s64 %rd1324, %rd2, 12; add.s64 %rd1325, %rd1323, %rd1324; ld.global.f32 %f6817, [%rd1325]; rcp.approx.ftz.f32 %f6818, %f15890; mul.ftz.f32 %f6819, %f6818, %f6817; ld.global.f32 %f6820, [%rd1325+4]; mul.ftz.f32 %f6821, %f6818, %f6820; ld.global.f32 %f6822, [%rd1325+8]; mul.ftz.f32 %f6823, %f6818, %f6822; st.global.f32 [%rd1325], %f6819; st.global.f32 [%rd1325+4], %f6821; st.global.f32 [%rd1325+8], %f6823; $L__BB106_799: ld.global.u32 %r401, [%rd29]; and.b32 %r1450, %r401, 16384; setp.eq.s32 %p686, %r1450, 0; @%p686 bra $L__BB106_1257; ld.const.u32 %r402, [kernel_params+1256]; setp.eq.s32 %p687, %r402, 0; @%p687 bra $L__BB106_1257; ld.const.u64 %rd1326, [kernel_params+2448]; cvta.to.global.u64 %rd1327, %rd1326; add.s64 %rd1329, %rd1327, %rd837; ld.global.u32 %rd1330, [%rd1329]; ld.const.s32 %rd1331, [kernel_params+1264]; mul.lo.s64 %rd166, %rd1331, %rd1330; and.b32 %r1451, %r401, 8192; setp.ne.s32 %p688, %r1451, 0; @%p688 bra $L__BB106_1146; ld.const.u64 %rd1332, [kernel_params+2456]; cvta.to.global.u64 %rd1333, %rd1332; add.s64 %rd1335, %rd1333, %rd837; ld.global.u32 %r1452, [%rd1335]; setp.ne.s32 %p689, %r1452, 0; @%p689 bra $L__BB106_816; and.b32 %r1453, %r402, 1; setp.eq.b32 %p690, %r1453, 1; mov.pred %p691, 0; xor.pred %p692, %p690, %p691; not.pred %p693, %p692; @%p693 bra $L__BB106_808; ld.local.v4.f32 {%f6824, %f6825, %f6826, %f6827}, [%rd1]; ld.const.u32 %r1454, [kernel_params+16]; setp.eq.s32 %p694, %r1454, 2; @%p694 bra $L__BB106_806; bra.uni $L__BB106_805; $L__BB106_806: ld.const.v4.f32 {%f6839, %f6840, %f6841, %f6842}, [kernel_params+96]; ld.const.v4.f32 {%f6844, %f6845, %f6846, %f6847}, [kernel_params+112]; ld.const.v4.f32 {%f6849, %f6850, %f6851, %f6852}, [kernel_params+128]; sub.ftz.f32 %f6854, %f6824, %f6842; sub.ftz.f32 %f6855, %f6825, %f6847; sub.ftz.f32 %f6856, %f6826, %f6852; mul.ftz.f32 %f6857, %f6855, %f6855; fma.rn.ftz.f32 %f6858, %f6854, %f6854, %f6857; fma.rn.ftz.f32 %f6859, %f6856, %f6856, %f6858; sqrt.approx.ftz.f32 %f15891, %f6859; bra.uni $L__BB106_807; $L__BB106_782: ld.const.u64 %rd1290, [kernel_params+2600]; cvta.to.global.u64 %rd1291, %rd1290; add.s64 %rd1293, %rd1291, %rd1228; ld.global.f32 %f1111, [%rd1293]; ld.global.f32 %f1112, [%rd1293+4]; ld.global.f32 %f1113, [%rd1293+8]; ld.const.u64 %rd1294, [kernel_params+2608]; cvta.to.global.u64 %rd1295, %rd1294; add.s64 %rd1296, %rd1295, %rd1228; ld.global.f32 %f1114, [%rd1296]; ld.global.f32 %f1115, [%rd1296+4]; ld.global.f32 %f1116, [%rd1296+8]; ld.const.u32 %r1441, [kernel_params+1324]; ld.const.u32 %r1442, [kernel_params+1340]; setp.eq.s16 %p674, %rs15, 1; selp.b32 %r395, %r1442, %r1441, %p674; setp.eq.s32 %p675, %r395, -1; @%p675 bra $L__BB106_784; cvt.s64.s32 %rd1297, %r395; add.s64 %rd1298, %rd165, %rd1297; cvta.to.global.u64 %rd1299, %rd618; shl.b64 %rd1300, %rd1298, 2; add.s64 %rd1301, %rd1299, %rd1300; add.s64 %rd1302, %rd1301, 4; add.s64 %rd1303, %rd1301, 8; mul.ftz.f32 %f6794, %f15887, %f1114; atom.global.add.f32 %f6795, [%rd1301], %f6794; mul.ftz.f32 %f6796, %f15888, %f1115; atom.global.add.f32 %f6797, [%rd1302], %f6796; mul.ftz.f32 %f6798, %f15889, %f1116; atom.global.add.f32 %f6799, [%rd1303], %f6798; $L__BB106_784: ld.const.u32 %r1443, [kernel_params+1328]; ld.const.u32 %r1444, [kernel_params+1344]; selp.b32 %r396, %r1444, %r1443, %p674; setp.eq.s32 %p677, %r396, -1; @%p677 bra $L__BB106_786; mov.f32 %f6800, 0f3F800000; sub.ftz.f32 %f6801, %f6800, %f1111; sub.ftz.f32 %f6802, %f6801, %f1114; sub.ftz.f32 %f6803, %f6800, %f1112; sub.ftz.f32 %f6804, %f6803, %f1115; sub.ftz.f32 %f6805, %f6800, %f1113; sub.ftz.f32 %f6806, %f6805, %f1116; cvt.s64.s32 %rd1304, %r396; add.s64 %rd1305, %rd165, %rd1304; cvta.to.global.u64 %rd1306, %rd618; shl.b64 %rd1307, %rd1305, 2; add.s64 %rd1308, %rd1306, %rd1307; mul.ftz.f32 %f6807, %f15887, %f6802; mul.ftz.f32 %f6808, %f15888, %f6804; mul.ftz.f32 %f6809, %f15889, %f6806; add.s64 %rd1309, %rd1308, 4; add.s64 %rd1310, %rd1308, 8; atom.global.add.f32 %f6810, [%rd1308], %f6807; atom.global.add.f32 %f6811, [%rd1309], %f6808; atom.global.add.f32 %f6812, [%rd1310], %f6809; $L__BB106_786: ld.const.u32 %r1446, [kernel_params+1320]; ld.const.u32 %r1447, [kernel_params+1336]; selp.b32 %r397, %r1447, %r1446, %p674; setp.eq.s32 %p679, %r397, -1; @%p679 bra $L__BB106_792; mul.ftz.f32 %f15887, %f15887, %f1111; mul.ftz.f32 %f15888, %f15888, %f1112; mul.ftz.f32 %f15889, %f15889, %f1113; mov.u32 %r3169, %r397; bra.uni $L__BB106_792; $L__BB106_790: ld.const.u32 %r3169, [kernel_params+1332]; bra.uni $L__BB106_792; $L__BB106_805: ld.const.v4.f32 {%f6828, %f6829, %f6830, %f6831}, [kernel_params+720]; mul.ftz.f32 %f6836, %f6825, %f6829; fma.rn.ftz.f32 %f6837, %f6824, %f6828, %f6836; fma.rn.ftz.f32 %f6838, %f6826, %f6830, %f6837; add.ftz.f32 %f15891, %f6831, %f6838; $L__BB106_807: ld.const.s32 %rd1338, [kernel_params+1272]; add.s64 %rd1339, %rd166, %rd1338; cvta.to.global.u64 %rd1340, %rd618; shl.b64 %rd1341, %rd1339, 2; add.s64 %rd1342, %rd1340, %rd1341; st.global.f32 [%rd1342], %f15891; $L__BB106_808: and.b32 %r1455, %r402, 32; setp.eq.s32 %p695, %r1455, 0; @%p695 bra $L__BB106_812; ld.local.u32 %r403, [%rd7+-12]; setp.eq.s32 %p696, %r403, -1; mov.f32 %f15892, 0f00000000; @%p696 bra $L__BB106_811; ld.const.u64 %rd1343, [kernel_params+2160]; cvta.to.global.u64 %rd1344, %rd1343; mul.wide.s32 %rd1345, %r403, 208; add.s64 %rd1346, %rd1344, %rd1345; ld.global.f32 %f15892, [%rd1346+100]; $L__BB106_811: ld.const.s32 %rd1347, [kernel_params+1300]; add.s64 %rd1348, %rd166, %rd1347; cvta.to.global.u64 %rd1349, %rd618; shl.b64 %rd1350, %rd1348, 2; add.s64 %rd1351, %rd1349, %rd1350; st.global.f32 [%rd1351], %f15892; $L__BB106_812: and.b32 %r1456, %r402, 64; setp.eq.s32 %p697, %r1456, 0; @%p697 bra $L__BB106_814; ld.local.u32 %rd1352, [%rd7+-40]; shl.b64 %rd1353, %rd1352, 5; and.b64 %rd1354, %rd1353, 134217696; add.s64 %rd1355, %rd28, %rd1354; ld.global.u32 %r1457, [%rd1355+20]; cvt.rn.f32.s32 %f6861, %r1457; ld.const.s32 %rd1356, [kernel_params+1304]; add.s64 %rd1357, %rd166, %rd1356; cvta.to.global.u64 %rd1358, %rd618; shl.b64 %rd1359, %rd1357, 2; add.s64 %rd1360, %rd1358, %rd1359; st.global.f32 [%rd1360], %f6861; $L__BB106_814: and.b32 %r1458, %r402, 2; setp.eq.s32 %p698, %r1458, 0; @%p698 bra $L__BB106_816; ld.local.v4.f32 {%f6862, %f6863, %f6864, %f6865}, [%rd1]; ld.const.s32 %rd1363, [kernel_params+1276]; add.s64 %rd1364, %rd166, %rd1363; cvta.to.global.u64 %rd1365, %rd618; shl.b64 %rd1366, %rd1364, 2; add.s64 %rd1367, %rd1365, %rd1366; st.global.f32 [%rd1367], %f6862; st.global.f32 [%rd1367+4], %f6863; st.global.f32 [%rd1367+8], %f6864; $L__BB106_816: and.b32 %r1459, %r3134, 512; setp.eq.s32 %p699, %r1459, 0; ld.const.f32 %f1133, [kernel_params+1364]; setp.eq.ftz.f32 %p700, %f1133, 0f00000000; or.pred %p701, %p699, %p700; @%p701 bra $L__BB106_820; and.b32 %r1460, %r3134, 524288; setp.ne.s32 %p702, %r1460, 0; mov.f32 %f6871, 0f3F800000; mov.f32 %f15893, %f6871; mov.f32 %f15894, %f6871; mov.f32 %f15895, %f6871; @%p702 bra $L__BB106_819; ld.local.v2.f32 {%f15894, %f15895}, [%rd7+208]; ld.local.f32 %f15893, [%rd7+204]; $L__BB106_819: sub.ftz.f32 %f6875, %f6871, %f15893; sub.ftz.f32 %f6876, %f6871, %f15894; sub.ftz.f32 %f6877, %f6871, %f15895; cvt.ftz.sat.f32.f32 %f6878, %f6875; cvt.ftz.sat.f32.f32 %f6879, %f6876; cvt.ftz.sat.f32.f32 %f6880, %f6877; add.ftz.f32 %f6881, %f6878, %f6879; add.ftz.f32 %f6882, %f6880, %f6881; mul.ftz.f32 %f6883, %f6882, 0f3EAAAAAB; setp.ltu.ftz.f32 %p703, %f6883, %f1133; @%p703 bra $L__BB106_1146; $L__BB106_820: and.b32 %r1461, %r402, 4; setp.eq.s32 %p704, %r1461, 0; @%p704 bra $L__BB106_841; ld.local.u32 %r404, [%rd7+172]; setp.lt.s32 %p705, %r404, 1; mov.f32 %f15899, 0f00000000; mov.f32 %f15900, %f15899; mov.f32 %f15901, %f15899; @%p705 bra $L__BB106_837; add.s32 %r1463, %r404, -1; and.b32 %r3174, %r404, 3; setp.lt.u32 %p706, %r1463, 3; mov.f32 %f15901, 0f00000000; mov.u32 %r3172, 0; mov.f32 %f15900, %f15901; mov.f32 %f15899, %f15901; @%p706 bra $L__BB106_833; sub.s32 %r3171, %r404, %r3174; mov.f32 %f15901, 0f00000000; mov.u32 %r3172, 0; mov.f32 %f15900, %f15901; mov.f32 %f15899, %f15901; $L__BB106_824: cvt.s64.s32 %rd167, %r3172; mul.wide.s32 %rd1368, %r3172, 80; add.s64 %rd1369, %rd1, %rd1368; add.s64 %rd168, %rd1369, 316; ld.local.u32 %r1465, [%rd1369+316]; setp.gt.s32 %p707, %r1465, 34; @%p707 bra $L__BB106_826; ld.local.v4.f32 {%f6894, %f6895, %f6896, %f6897}, [%rd168+-12]; add.ftz.f32 %f6901, %f6894, %f6895; add.ftz.f32 %f6902, %f6896, %f6901; mul.ftz.f32 %f6903, %f6902, 0f3EAAAAAB; abs.ftz.f32 %f6904, %f6903; ld.local.f32 %f6905, [%rd168+8]; ld.local.f32 %f6906, [%rd168+12]; ld.local.f32 %f6907, [%rd168+16]; fma.rn.ftz.f32 %f15899, %f6904, %f6905, %f15899; fma.rn.ftz.f32 %f15900, %f6904, %f6906, %f15900; fma.rn.ftz.f32 %f15901, %f6904, %f6907, %f15901; $L__BB106_826: ld.local.u32 %r1466, [%rd168+80]; setp.gt.s32 %p708, %r1466, 34; @%p708 bra $L__BB106_828; ld.local.v4.f32 {%f6908, %f6909, %f6910, %f6911}, [%rd168+68]; add.ftz.f32 %f6915, %f6908, %f6909; add.ftz.f32 %f6916, %f6910, %f6915; mul.ftz.f32 %f6917, %f6916, 0f3EAAAAAB; abs.ftz.f32 %f6918, %f6917; ld.local.f32 %f6919, [%rd168+88]; ld.local.f32 %f6920, [%rd168+92]; ld.local.f32 %f6921, [%rd168+96]; fma.rn.ftz.f32 %f15899, %f6918, %f6919, %f15899; fma.rn.ftz.f32 %f15900, %f6918, %f6920, %f15900; fma.rn.ftz.f32 %f15901, %f6918, %f6921, %f15901; $L__BB106_828: ld.local.u32 %r1467, [%rd168+160]; setp.gt.s32 %p709, %r1467, 34; @%p709 bra $L__BB106_830; ld.local.v4.f32 {%f6922, %f6923, %f6924, %f6925}, [%rd168+148]; add.ftz.f32 %f6929, %f6922, %f6923; add.ftz.f32 %f6930, %f6924, %f6929; mul.ftz.f32 %f6931, %f6930, 0f3EAAAAAB; abs.ftz.f32 %f6932, %f6931; ld.local.f32 %f6933, [%rd168+168]; ld.local.f32 %f6934, [%rd168+172]; ld.local.f32 %f6935, [%rd168+176]; fma.rn.ftz.f32 %f15899, %f6932, %f6933, %f15899; fma.rn.ftz.f32 %f15900, %f6932, %f6934, %f15900; fma.rn.ftz.f32 %f15901, %f6932, %f6935, %f15901; $L__BB106_830: ld.local.u32 %r1468, [%rd168+240]; setp.gt.s32 %p710, %r1468, 34; @%p710 bra $L__BB106_832; ld.local.v4.f32 {%f6936, %f6937, %f6938, %f6939}, [%rd168+228]; add.ftz.f32 %f6943, %f6936, %f6937; add.ftz.f32 %f6944, %f6938, %f6943; mul.ftz.f32 %f6945, %f6944, 0f3EAAAAAB; abs.ftz.f32 %f6946, %f6945; ld.local.f32 %f6947, [%rd168+248]; ld.local.f32 %f6948, [%rd168+252]; ld.local.f32 %f6949, [%rd168+256]; fma.rn.ftz.f32 %f15899, %f6946, %f6947, %f15899; fma.rn.ftz.f32 %f15900, %f6946, %f6948, %f15900; fma.rn.ftz.f32 %f15901, %f6946, %f6949, %f15901; $L__BB106_832: cvt.u32.u64 %r1469, %rd167; add.s32 %r3172, %r1469, 4; add.s32 %r3171, %r3171, -4; setp.ne.s32 %p711, %r3171, 0; @%p711 bra $L__BB106_824; $L__BB106_833: setp.eq.s32 %p712, %r3174, 0; @%p712 bra $L__BB106_837; $L__BB106_834: .pragma "nounroll"; cvt.s64.s32 %rd169, %r3172; mul.wide.s32 %rd1370, %r3172, 80; add.s64 %rd1371, %rd1, %rd1370; add.s64 %rd170, %rd1371, 316; ld.local.u32 %r1470, [%rd1371+316]; setp.gt.s32 %p713, %r1470, 34; @%p713 bra $L__BB106_836; ld.local.f32 %f6950, [%rd170+-8]; ld.local.f32 %f6951, [%rd170+-12]; add.ftz.f32 %f6952, %f6951, %f6950; ld.local.f32 %f6953, [%rd170+-4]; add.ftz.f32 %f6954, %f6953, %f6952; mul.ftz.f32 %f6955, %f6954, 0f3EAAAAAB; abs.ftz.f32 %f6956, %f6955; ld.local.f32 %f6957, [%rd170+8]; ld.local.f32 %f6958, [%rd170+12]; ld.local.f32 %f6959, [%rd170+16]; fma.rn.ftz.f32 %f15899, %f6956, %f6957, %f15899; fma.rn.ftz.f32 %f15900, %f6956, %f6958, %f15900; fma.rn.ftz.f32 %f15901, %f6956, %f6959, %f15901; $L__BB106_836: cvt.u32.u64 %r1471, %rd169; add.s32 %r3172, %r1471, 1; add.s32 %r3174, %r3174, -1; setp.ne.s32 %p714, %r3174, 0; @%p714 bra $L__BB106_834; $L__BB106_837: setp.eq.ftz.f32 %p715, %f15900, 0f00000000; setp.eq.ftz.f32 %p716, %f15899, 0f00000000; and.pred %p717, %p716, %p715; setp.eq.ftz.f32 %p718, %f15901, 0f00000000; and.pred %p719, %p718, %p717; @%p719 bra $L__BB106_839; bra.uni $L__BB106_838; $L__BB106_839: ld.local.v2.f32 {%f15927, %f15928}, [%rd7+-72]; ld.local.f32 %f15926, [%rd7+-76]; bra.uni $L__BB106_840; $L__BB106_838: mul.ftz.f32 %f6960, %f15900, %f15900; fma.rn.ftz.f32 %f6961, %f15899, %f15899, %f6960; fma.rn.ftz.f32 %f6962, %f15901, %f15901, %f6961; sqrt.approx.ftz.f32 %f6963, %f6962; div.approx.ftz.f32 %f15928, %f15901, %f6963; div.approx.ftz.f32 %f15927, %f15900, %f6963; div.approx.ftz.f32 %f15926, %f15899, %f6963; $L__BB106_840: ld.const.s32 %rd1372, [kernel_params+1280]; add.s64 %rd1373, %rd166, %rd1372; cvta.to.global.u64 %rd1374, %rd618; shl.b64 %rd1375, %rd1373, 2; add.s64 %rd1376, %rd1374, %rd1375; add.s64 %rd1377, %rd1376, 4; add.s64 %rd1378, %rd1376, 8; atom.global.add.f32 %f6966, [%rd1376], %f15926; atom.global.add.f32 %f6967, [%rd1377], %f15927; atom.global.add.f32 %f6968, [%rd1378], %f15928; $L__BB106_841: and.b32 %r1472, %r402, 8; setp.eq.s32 %p720, %r1472, 0; @%p720 bra $L__BB106_886; ld.local.u32 %r416, [%rd7+172]; setp.lt.s32 %p721, %r416, 1; mov.f32 %f15944, 0f00000000; mov.f32 %f15932, %f15944; mov.f32 %f15933, %f15944; @%p721 bra $L__BB106_883; and.b32 %r417, %r416, 1; setp.eq.s32 %p722, %r416, 1; mov.f32 %f15933, 0f00000000; mov.u32 %r3177, 0; mov.f32 %f15932, %f15933; @%p722 bra $L__BB106_870; sub.s32 %r3176, %r416, %r417; mov.f32 %f15933, 0f00000000; mov.u32 %r3177, 0; mov.f32 %f15932, %f15933; $L__BB106_845: cvt.s64.s32 %rd171, %r3177; mul.wide.s32 %rd1379, %r3177, 80; add.s64 %rd1380, %rd1, %rd1379; add.s64 %rd172, %rd1380, 316; ld.local.u32 %r421, [%rd1380+316]; setp.gt.s32 %p723, %r421, 31; @%p723 bra $L__BB106_857; ld.local.v4.f32 {%f6976, %f6977, %f6978, %f6979}, [%rd172+-12]; setp.eq.s32 %p724, %r421, 3; @%p724 bra $L__BB106_855; bra.uni $L__BB106_847; $L__BB106_855: ld.local.f32 %f6989, [%rd172+20]; mul.ftz.f32 %f6990, %f6989, %f6989; mul.ftz.f32 %f15931, %f6990, %f6990; bra.uni $L__BB106_856; $L__BB106_847: setp.eq.s32 %p725, %r421, 5; @%p725 bra $L__BB106_854; bra.uni $L__BB106_848; $L__BB106_854: ld.local.f32 %f6987, [%rd172+20]; mul.ftz.f32 %f6988, %f6987, %f6987; mul.ftz.f32 %f15931, %f6988, %f6988; bra.uni $L__BB106_856; $L__BB106_848: add.s32 %r1475, %r421, -2; setp.lt.u32 %p726, %r1475, 7; mov.f32 %f15931, 0f00000000; @%p726 bra $L__BB106_856; setp.gt.u32 %p727, %r421, 31; @%p727 bra $L__BB106_851; mov.u32 %r1477, 1; shl.b32 %r1478, %r1477, %r421; and.b32 %r1479, %r1478, -2145385984; setp.ne.s32 %p728, %r1479, 0; @%p728 bra $L__BB106_856; $L__BB106_851: add.s32 %r1480, %r421, -10; setp.lt.u32 %p729, %r1480, 7; @%p729 bra $L__BB106_853; add.s32 %r1481, %r421, -22; setp.gt.u32 %p730, %r1481, 2; setp.ne.s32 %p731, %r421, 27; mov.f32 %f15931, 0f3F800000; and.pred %p732, %p730, %p731; @%p732 bra $L__BB106_856; $L__BB106_853: ld.local.v2.f32 {%f6983, %f6984}, [%rd172+20]; mul.ftz.f32 %f15931, %f6983, %f6984; $L__BB106_856: add.ftz.f32 %f6991, %f6976, %f6977; add.ftz.f32 %f6992, %f6978, %f6991; mul.ftz.f32 %f6993, %f6992, 0f3EAAAAAB; abs.ftz.f32 %f6994, %f6993; mov.f32 %f6995, 0f00000000; max.ftz.f32 %f6996, %f15931, %f6995; sqrt.approx.ftz.f32 %f6997, %f6996; sqrt.approx.ftz.f32 %f6998, %f6997; fma.rn.ftz.f32 %f15932, %f6994, %f6998, %f15932; add.ftz.f32 %f15933, %f15933, %f6994; $L__BB106_857: ld.local.u32 %r422, [%rd172+80]; setp.gt.s32 %p733, %r422, 31; @%p733 bra $L__BB106_869; ld.local.v4.f32 {%f6999, %f7000, %f7001, %f7002}, [%rd172+68]; setp.eq.s32 %p734, %r422, 3; @%p734 bra $L__BB106_867; bra.uni $L__BB106_859; $L__BB106_867: ld.local.f32 %f7012, [%rd172+100]; mul.ftz.f32 %f7013, %f7012, %f7012; mul.ftz.f32 %f15934, %f7013, %f7013; bra.uni $L__BB106_868; $L__BB106_859: setp.eq.s32 %p735, %r422, 5; @%p735 bra $L__BB106_866; bra.uni $L__BB106_860; $L__BB106_866: ld.local.f32 %f7010, [%rd172+100]; mul.ftz.f32 %f7011, %f7010, %f7010; mul.ftz.f32 %f15934, %f7011, %f7011; bra.uni $L__BB106_868; $L__BB106_860: add.s32 %r1482, %r422, -2; setp.lt.u32 %p736, %r1482, 7; mov.f32 %f15934, 0f00000000; @%p736 bra $L__BB106_868; setp.gt.u32 %p737, %r422, 31; @%p737 bra $L__BB106_863; mov.u32 %r1484, 1; shl.b32 %r1485, %r1484, %r422; and.b32 %r1486, %r1485, -2145385984; setp.ne.s32 %p738, %r1486, 0; @%p738 bra $L__BB106_868; $L__BB106_863: add.s32 %r1487, %r422, -10; setp.lt.u32 %p739, %r1487, 7; @%p739 bra $L__BB106_865; add.s32 %r1488, %r422, -22; setp.gt.u32 %p740, %r1488, 2; setp.ne.s32 %p741, %r422, 27; mov.f32 %f15934, 0f3F800000; and.pred %p742, %p740, %p741; @%p742 bra $L__BB106_868; $L__BB106_865: ld.local.v2.f32 {%f7006, %f7007}, [%rd172+100]; mul.ftz.f32 %f15934, %f7006, %f7007; $L__BB106_868: add.ftz.f32 %f7014, %f6999, %f7000; add.ftz.f32 %f7015, %f7001, %f7014; mul.ftz.f32 %f7016, %f7015, 0f3EAAAAAB; abs.ftz.f32 %f7017, %f7016; mov.f32 %f7018, 0f00000000; max.ftz.f32 %f7019, %f15934, %f7018; sqrt.approx.ftz.f32 %f7020, %f7019; sqrt.approx.ftz.f32 %f7021, %f7020; fma.rn.ftz.f32 %f15932, %f7017, %f7021, %f15932; add.ftz.f32 %f15933, %f15933, %f7017; $L__BB106_869: cvt.u32.u64 %r1489, %rd171; add.s32 %r3177, %r1489, 2; add.s32 %r3176, %r3176, -2; setp.ne.s32 %p743, %r3176, 0; @%p743 bra $L__BB106_845; $L__BB106_870: setp.eq.s32 %p744, %r417, 0; @%p744 bra $L__BB106_883; mul.wide.s32 %rd1381, %r3177, 80; add.s64 %rd1382, %rd1, %rd1381; add.s64 %rd173, %rd1382, 316; ld.local.u32 %r426, [%rd1382+316]; setp.gt.s32 %p745, %r426, 31; @%p745 bra $L__BB106_883; ld.local.f32 %f7022, [%rd173+-8]; ld.local.f32 %f7023, [%rd173+-12]; add.ftz.f32 %f7024, %f7023, %f7022; ld.local.f32 %f7025, [%rd173+-4]; add.ftz.f32 %f1224, %f7025, %f7024; setp.eq.s32 %p746, %r426, 3; @%p746 bra $L__BB106_881; bra.uni $L__BB106_873; $L__BB106_881: ld.local.f32 %f7035, [%rd173+20]; mul.ftz.f32 %f7036, %f7035, %f7035; mul.ftz.f32 %f15941, %f7036, %f7036; bra.uni $L__BB106_882; $L__BB106_873: setp.eq.s32 %p747, %r426, 5; @%p747 bra $L__BB106_880; bra.uni $L__BB106_874; $L__BB106_880: ld.local.f32 %f7033, [%rd173+20]; mul.ftz.f32 %f7034, %f7033, %f7033; mul.ftz.f32 %f15941, %f7034, %f7034; bra.uni $L__BB106_882; $L__BB106_2267: mul.ftz.f32 %f14561, %f16692, %f3942; fma.rn.ftz.f32 %f14562, %f16691, %f3941, %f14561; fma.rn.ftz.f32 %f14563, %f16693, %f3943, %f14562; mul.ftz.f32 %f14564, %f16689, %f3942; fma.rn.ftz.f32 %f14565, %f16688, %f3941, %f14564; fma.rn.ftz.f32 %f14566, %f16690, %f3943, %f14565; mul.ftz.f32 %f14567, %f3900, %f3942; fma.rn.ftz.f32 %f14568, %f3898, %f3941, %f14567; fma.rn.ftz.f32 %f14569, %f3901, %f3943, %f14568; neg.ftz.f32 %f14570, %f14563; mul.ftz.f32 %f14571, %f14362, %f14569; div.approx.ftz.f32 %f14572, %f14570, %f14571; neg.ftz.f32 %f14573, %f14566; mul.ftz.f32 %f14574, %f14363, %f14569; div.approx.ftz.f32 %f14575, %f14573, %f14574; fma.rn.ftz.f32 %f14576, %f14572, %f14572, 0f3F800000; mov.f32 %f14577, 0f3F800000; fma.rn.ftz.f32 %f14578, %f14575, %f14575, %f14576; mul.ftz.f32 %f14579, %f14569, %f14569; mul.ftz.f32 %f14580, %f14579, %f14579; mul.ftz.f32 %f14581, %f14578, %f14578; mul.ftz.f32 %f14582, %f14581, 0f40490FDB; mul.ftz.f32 %f14583, %f16705, %f14582; mul.ftz.f32 %f14584, %f14580, %f14583; rcp.approx.ftz.f32 %f16706, %f14584; mul.ftz.f32 %f14585, %f3965, %f3965; sub.ftz.f32 %f14586, %f14577, %f14585; div.approx.ftz.f32 %f14587, %f14586, %f14585; mul.ftz.f32 %f14588, %f16692, %f16736; fma.rn.ftz.f32 %f14589, %f16691, %f16735, %f14588; fma.rn.ftz.f32 %f14590, %f16693, %f16737, %f14589; mul.ftz.f32 %f14591, %f16689, %f16736; fma.rn.ftz.f32 %f14592, %f16688, %f16735, %f14591; fma.rn.ftz.f32 %f14593, %f16690, %f16737, %f14592; mul.ftz.f32 %f14594, %f14590, %f14590; mul.ftz.f32 %f14595, %f14362, %f14362; mul.ftz.f32 %f14596, %f14593, %f14593; mul.ftz.f32 %f14597, %f14363, %f14363; mul.ftz.f32 %f14598, %f14597, %f14596; fma.rn.ftz.f32 %f14599, %f14595, %f14594, %f14598; add.ftz.f32 %f14600, %f14594, %f14596; div.approx.ftz.f32 %f14601, %f14599, %f14600; mul.ftz.f32 %f16709, %f14587, %f14601; bra.uni $L__BB106_2274; $L__BB106_2120: setp.lt.ftz.f32 %p1888, %f2799, 0f3F400000; @%p1888 bra $L__BB106_2124; bra.uni $L__BB106_2121; $L__BB106_2124: add.ftz.f32 %f13224, %f2799, 0fBF000000; mul.ftz.f32 %f13225, %f13224, 0f40800000; mul.ftz.f32 %f13226, %f13225, 0f3FC90FDB; sin.approx.ftz.f32 %f13227, %f13226; cos.approx.ftz.f32 %f13228, %f13226; div.approx.ftz.f32 %f13229, %f13227, %f13228; mul.ftz.f32 %f3519, %f3512, %f13229; abs.ftz.f32 %f3520, %f3519; setp.leu.ftz.f32 %p1892, %f3520, 0f3F800000; mov.f32 %f16574, %f3520; @%p1892 bra $L__BB106_2126; rcp.approx.ftz.f32 %f16574, %f3520; $L__BB106_2126: mul.rn.ftz.f32 %f13230, %f16574, %f16574; mov.f32 %f13231, 0fC0B59883; mov.f32 %f13232, 0fBF52C7EA; fma.rn.ftz.f32 %f13233, %f13230, %f13232, %f13231; mov.f32 %f13234, 0fC0D21907; fma.rn.ftz.f32 %f13235, %f13233, %f13230, %f13234; mul.ftz.f32 %f13236, %f13230, %f13235; mul.ftz.f32 %f13237, %f16574, %f13236; add.ftz.f32 %f13238, %f13230, 0f41355DC0; mov.f32 %f13239, 0f41E6BD60; fma.rn.ftz.f32 %f13240, %f13238, %f13230, %f13239; mov.f32 %f13241, 0f419D92C8; fma.rn.ftz.f32 %f13242, %f13240, %f13230, %f13241; rcp.approx.ftz.f32 %f13243, %f13242; fma.rn.ftz.f32 %f13244, %f13237, %f13243, %f16574; mov.f32 %f13245, 0f3FC90FDB; sub.ftz.f32 %f13246, %f13245, %f13244; setp.gt.ftz.f32 %p1893, %f3520, 0f3F800000; selp.f32 %f13247, %f13246, %f13244, %p1893; mov.b32 %r2819, %f13247; mov.b32 %r2820, %f3519; and.b32 %r2821, %r2820, -2147483648; or.b32 %r2822, %r2821, %r2819; mov.b32 %f13248, %r2822; setp.le.ftz.f32 %p1894, %f3520, 0f7F800000; selp.f32 %f13249, %f13248, %f13247, %p1894; cos.approx.ftz.f32 %f13250, %f13249; sin.approx.ftz.f32 %f13251, %f13249; mul.ftz.f32 %f13252, %f3491, %f13250; mul.ftz.f32 %f13253, %f3492, %f13251; mul.ftz.f32 %f13254, %f13251, %f13253; fma.rn.ftz.f32 %f13255, %f13250, %f13252, %f13254; add.ftz.f32 %f13256, %f13255, 0f3F800000; rcp.approx.ftz.f32 %f13257, %f13256; lg2.approx.ftz.f32 %f13258, %f2775; mul.ftz.f32 %f13259, %f13258, %f13257; ex2.approx.ftz.f32 %f16577, %f13259; add.ftz.f32 %f16578, %f13249, 0f40490FDB; bra.uni $L__BB106_2134; $L__BB106_2269: mul.ftz.f32 %f14602, %f3940, %f3940; rcp.approx.ftz.f32 %f14603, %f14602; add.ftz.f32 %f14604, %f14603, 0fBF800000; mul.ftz.f32 %f14605, %f14602, %f14602; mul.ftz.f32 %f14606, %f14605, 0f40490FDB; add.ftz.f32 %f14607, %f16705, %f14604; mul.ftz.f32 %f14608, %f14606, %f14607; mul.ftz.f32 %f14609, %f14607, %f14608; div.approx.ftz.f32 %f16706, %f16705, %f14609; $L__BB106_2273: mul.ftz.f32 %f14630, %f3965, %f3965; mov.f32 %f14631, 0f3F800000; sub.ftz.f32 %f14632, %f14631, %f14630; mul.ftz.f32 %f14633, %f14632, %f16705; div.approx.ftz.f32 %f16709, %f14633, %f14630; $L__BB106_2274: mul.ftz.f32 %f14637, %f3937, %f16706; mul.ftz.f32 %f14638, %f14637, 0f3E800000; div.approx.ftz.f32 %f16738, %f14638, %f3902; add.s32 %r2878, %r873, -11; setp.gt.u32 %p2011, %r2878, 1; mov.f32 %f16712, 0f3F800000; mov.f32 %f16713, %f16712; mov.f32 %f16714, %f16712; @%p2011 bra $L__BB106_2280; ld.local.f32 %f14640, [%rd587+28]; fma.rn.ftz.f32 %f3980, %f14640, %f14640, 0fBF800000; mov.f32 %f14639, 0f3F800000; abs.ftz.f32 %f3981, %f14639; fma.rn.ftz.f32 %f3982, %f3981, %f3981, %f3980; setp.leu.ftz.f32 %p2012, %f3982, 0f00000000; mov.f32 %f16710, %f14639; @%p2012 bra $L__BB106_2277; sqrt.approx.ftz.f32 %f14641, %f3982; sub.ftz.f32 %f14642, %f14641, %f3981; add.ftz.f32 %f14643, %f3981, %f14641; div.approx.ftz.f32 %f14644, %f14642, %f14643; fma.rn.ftz.f32 %f14645, %f3981, %f14643, 0fBF800000; fma.rn.ftz.f32 %f14646, %f3981, %f14642, 0f3F800000; div.approx.ftz.f32 %f14647, %f14645, %f14646; mul.ftz.f32 %f14648, %f14644, 0f3F000000; mul.ftz.f32 %f14649, %f14644, %f14648; fma.rn.ftz.f32 %f14650, %f14647, %f14647, 0f3F800000; mul.ftz.f32 %f16710, %f14649, %f14650; $L__BB106_2277: ld.local.u64 %rd2690, [%rd587+36]; ld.f32 %f3985, [%rd2690+12]; ld.f32 %f3986, [%rd2690+16]; ld.f32 %f3987, [%rd2690+20]; sub.ftz.f32 %f3988, %f14639, %f16710; mul.ftz.f32 %f14652, %f3942, %f16736; fma.rn.ftz.f32 %f14653, %f3941, %f16735, %f14652; fma.rn.ftz.f32 %f14654, %f3943, %f16737, %f14653; abs.ftz.f32 %f3989, %f14654; fma.rn.ftz.f32 %f3990, %f3989, %f3989, %f3980; setp.leu.ftz.f32 %p2013, %f3990, 0f00000000; mov.f32 %f16711, %f3988; @%p2013 bra $L__BB106_2279; sqrt.approx.ftz.f32 %f14655, %f3990; sub.ftz.f32 %f14656, %f14655, %f3989; add.ftz.f32 %f14657, %f3989, %f14655; div.approx.ftz.f32 %f14658, %f14656, %f14657; fma.rn.ftz.f32 %f14659, %f3989, %f14657, 0fBF800000; fma.rn.ftz.f32 %f14660, %f3989, %f14656, 0f3F800000; div.approx.ftz.f32 %f14661, %f14659, %f14660; mul.ftz.f32 %f14662, %f14658, 0f3F000000; mul.ftz.f32 %f14663, %f14658, %f14662; fma.rn.ftz.f32 %f14664, %f14661, %f14661, 0f3F800000; mul.ftz.f32 %f14665, %f14663, %f14664; sub.ftz.f32 %f16711, %f14665, %f16710; $L__BB106_2279: rcp.approx.ftz.f32 %f14666, %f3988; mov.f32 %f14667, 0f3F800000; mul.ftz.f32 %f14668, %f14666, %f16711; sub.ftz.f32 %f14669, %f14667, %f14668; fma.rn.ftz.f32 %f16714, %f3987, %f14669, %f14668; fma.rn.ftz.f32 %f16713, %f3986, %f14669, %f14668; fma.rn.ftz.f32 %f16712, %f3985, %f14669, %f14668; $L__BB106_2280: add.ftz.f32 %f14670, %f16709, 0f3F800000; mov.f32 %f14671, 0f00000000; max.ftz.f32 %f14672, %f14670, %f14671; sqrt.approx.ftz.f32 %f14673, %f14672; add.ftz.f32 %f14674, %f14673, 0f3F800000; mov.f32 %f14675, 0f40000000; div.approx.ftz.f32 %f14676, %f14675, %f14674; mul.ftz.f32 %f14677, %f14676, %f16738; mul.ftz.f32 %f16739, %f14677, %f16712; mul.ftz.f32 %f16740, %f14677, %f16713; mul.ftz.f32 %f16741, %f14677, %f16714; mov.u32 %r3285, 10; $L__BB106_2281: setp.ne.s32 %p2014, %r873, 12; @%p2014 bra $L__BB106_2309; ld.local.u64 %rd2691, [%rd587+36]; ld.f32 %f14678, [%rd2691+36]; mul.ftz.f32 %f14679, %f14678, 0f3E800000; mul.ftz.f32 %f16739, %f16739, %f14679; mul.ftz.f32 %f16740, %f16740, %f14679; mul.ftz.f32 %f16741, %f16741, %f14679; bra.uni $L__BB106_2309; $L__BB106_874: add.s32 %r1490, %r426, -2; setp.lt.u32 %p748, %r1490, 7; mov.f32 %f15941, 0f00000000; @%p748 bra $L__BB106_882; setp.gt.u32 %p749, %r426, 31; @%p749 bra $L__BB106_877; mov.u32 %r1492, 1; shl.b32 %r1493, %r1492, %r426; and.b32 %r1494, %r1493, -2145385984; setp.ne.s32 %p750, %r1494, 0; @%p750 bra $L__BB106_882; bra.uni $L__BB106_877; $L__BB106_882: mul.ftz.f32 %f7037, %f1224, 0f3EAAAAAB; abs.ftz.f32 %f7038, %f7037; mov.f32 %f7039, 0f00000000; max.ftz.f32 %f7040, %f15941, %f7039; sqrt.approx.ftz.f32 %f7041, %f7040; sqrt.approx.ftz.f32 %f7042, %f7041; fma.rn.ftz.f32 %f15932, %f7038, %f7042, %f15932; add.ftz.f32 %f15933, %f15933, %f7038; $L__BB106_883: setp.leu.ftz.f32 %p755, %f15933, 0f00000000; @%p755 bra $L__BB106_885; div.approx.ftz.f32 %f15944, %f15932, %f15933; $L__BB106_885: ld.const.s32 %rd1383, [kernel_params+1284]; add.s64 %rd1384, %rd166, %rd1383; cvta.to.global.u64 %rd1385, %rd618; shl.b64 %rd1386, %rd1384, 2; add.s64 %rd1387, %rd1385, %rd1386; atom.global.add.f32 %f7044, [%rd1387], %f15944; $L__BB106_886: and.b32 %r1497, %r402, 16; setp.eq.s32 %p756, %r1497, 0; @%p756 bra $L__BB106_924; ld.local.u32 %r427, [%rd7+-12]; ld.local.s32 %rd174, [%rd7+-28]; setp.eq.s32 %p757, %r427, -1; mov.u32 %r448, -1; mov.u32 %r445, 0; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p757 bra $L__BB106_902; ld.const.u64 %rd1388, [kernel_params+2160]; cvta.to.global.u64 %rd1389, %rd1388; mul.wide.s32 %rd1390, %r427, 208; add.s64 %rd1391, %rd1389, %rd1390; ld.global.u32 %r428, [%rd1391+164]; ld.local.u32 %r1503, [%rd7+-24]; and.b32 %r1504, %r1503, 1; setp.eq.b32 %p758, %r1504, 1; mov.pred %p759, 0; xor.pred %p760, %p758, %p759; not.pred %p761, %p760; mov.u32 %r3179, 0; @%p761 bra $L__BB106_892; cvt.u32.u64 %r1506, %rd174; setp.eq.s32 %p762, %r1506, -1; mov.u32 %r3178, -1; @%p762 bra $L__BB106_891; ld.const.u64 %rd1392, [kernel_params+2240]; cvta.to.global.u64 %rd1393, %rd1392; shl.b64 %rd1394, %rd174, 2; add.s64 %rd1395, %rd1393, %rd1394; ld.global.u32 %r3178, [%rd1395]; $L__BB106_891: setp.ne.s32 %p763, %r3178, -1; selp.u32 %r3179, 1, 0, %p763; $L__BB106_892: ld.const.u64 %rd1396, [kernel_params+2312]; cvta.to.global.u64 %rd175, %rd1396; add.s32 %r3181, %r3179, %r428; mul.wide.u32 %rd1397, %r3181, 16; add.s64 %rd1398, %rd175, %rd1397; add.s64 %rd2960, %rd1398, 14; ld.global.u16 %rs141, [%rd1398+12]; ld.global.u32 %r3180, [%rd1398+8]; ld.global.u64 %rd2958, [%rd1398]; setp.eq.s64 %p764, %rd2958, 3; add.s64 %rd2959, %rd1398, 15; @%p764 bra $L__BB106_898; $L__BB106_893: setp.eq.s64 %p765, %rd2958, 0; @%p765 bra $L__BB106_895; add.s32 %r3181, %r3181, 2; bra.uni $L__BB106_896; $L__BB106_895: setp.eq.s16 %p766, %rs141, 0; mov.u32 %r445, 0; mov.u32 %r3181, %r3180; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p766 bra $L__BB106_902; $L__BB106_896: cvt.u64.u32 %rd180, %r3181; mul.wide.u32 %rd1399, %r3181, 16; add.s64 %rd1400, %rd175, %rd1399; ld.global.u16 %rs141, [%rd1400+12]; ld.global.u32 %r3180, [%rd1400+8]; ld.global.u64 %rd2958, [%rd1400]; setp.ne.s64 %p767, %rd2958, 3; @%p767 bra $L__BB106_893; shl.b64 %rd1401, %rd180, 4; add.s64 %rd1402, %rd175, %rd1401; add.s64 %rd2960, %rd1402, 14; add.s64 %rd2959, %rd1402, 15; $L__BB106_898: ld.global.u8 %rs20, [%rd2960]; ld.global.u8 %rs21, [%rd2959]; cvt.u32.u64 %r1511, %rd174; setp.ne.s32 %p768, %r1511, -1; @%p768 bra $L__BB106_901; mov.u32 %r445, 0; add.s16 %rs78, %rs141, -1; setp.lt.u16 %p769, %rs78, 2; @%p769 bra $L__BB106_901; setp.ne.s16 %p770, %rs141, 1024; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p770 bra $L__BB106_902; $L__BB106_901: setp.eq.s16 %p771, %rs141, 0; selp.b32 %r448, -1, %r3180, %p771; cvt.u32.u16 %r445, %rs141; cvt.u32.u16 %r1516, %rs21; and.b32 %r3186, %r1516, 255; cvt.u32.u16 %r1517, %rs20; and.b32 %r3185, %r1517, 255; $L__BB106_902: setp.eq.s32 %p772, %r448, -1; mov.f32 %f15951, 0f00000000; mov.f32 %f15952, %f15951; mov.f32 %f15955, %f15951; @%p772 bra $L__BB106_923; ld.local.u32 %r451, [%rd7+-24]; and.b32 %r1518, %r451, 1; setp.eq.b32 %p773, %r1518, 1; mov.pred %p774, 0; xor.pred %p775, %p773, %p774; not.pred %p776, %p775; @%p776 bra $L__BB106_915; bra.uni $L__BB106_904; $L__BB106_915: and.b32 %r1539, %r451, 6; setp.eq.s32 %p783, %r1539, 0; @%p783 bra $L__BB106_920; and.b32 %r1540, %r445, 768; setp.eq.s32 %p784, %r1540, 0; @%p784 bra $L__BB106_918; cvt.u32.u64 %r1541, %rd174; ld.const.u64 %rd1429, [kernel_params+2264]; cvta.to.global.u64 %rd1430, %rd1429; mul.wide.s32 %rd1431, %r1541, 16; add.s64 %rd1432, %rd1430, %rd1431; shr.s32 %r1542, %r451, 7; ld.global.u32 %r1543, [%rd1432+4]; add.s32 %r1544, %r1542, %r1543; ld.const.u64 %rd1433, [kernel_params+2328]; cvta.to.global.u64 %rd1434, %rd1433; add.s32 %r1545, %r1544, %r448; mul.wide.s32 %rd1435, %r1545, 8; add.s64 %rd1436, %rd1434, %rd1435; ld.global.v2.f32 {%f7073, %f7074}, [%rd1436]; ld.global.v2.f32 {%f7077, %f7078}, [%rd1436+8]; ld.local.f32 %f7081, [%rd7+-20]; mov.f32 %f7082, 0f3F800000; sub.ftz.f32 %f7083, %f7082, %f7081; mul.ftz.f32 %f7084, %f7073, %f7083; mul.ftz.f32 %f7085, %f7074, %f7083; fma.rn.ftz.f32 %f15952, %f7078, %f7081, %f7085; fma.rn.ftz.f32 %f15951, %f7077, %f7081, %f7084; bra.uni $L__BB106_922; $L__BB106_904: cvt.u32.u64 %r1520, %rd174; setp.eq.s32 %p777, %r1520, -1; mov.u32 %r3188, -1; @%p777 bra $L__BB106_906; ld.const.u64 %rd1403, [kernel_params+2240]; cvta.to.global.u64 %rd1404, %rd1403; shl.b64 %rd1405, %rd174, 2; add.s64 %rd1406, %rd1404, %rd1405; ld.global.u32 %r3188, [%rd1406]; $L__BB106_906: setp.eq.s32 %p778, %r3188, -1; @%p778 bra $L__BB106_908; mov.u64 %rd1407, 0; { // callseq 167, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r445; st.param.b32 [param1+4], %r3185; st.param.b32 [param1+8], %r3186; st.param.b32 [param1+12], %r448; .param .b64 param2; st.param.b64 [param2+0], %rd1407; .param .b64 param3; st.param.b64 [param3+0], %rd1407; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15951, [retval0+0]; ld.param.f32 %f15952, [retval0+4]; } // callseq 167 bra.uni $L__BB106_922; $L__BB106_920: and.b32 %r1550, %r451, 8; setp.eq.s32 %p787, %r1550, 0; setp.ne.s32 %p788, %r445, 8; mov.f32 %f15951, 0f00000000; or.pred %p789, %p788, %p787; mov.f32 %f15952, %f15951; @%p789 bra $L__BB106_922; cvt.u32.u64 %r1551, %rd174; ld.const.u64 %rd1441, [kernel_params+2328]; cvta.to.global.u64 %rd1442, %rd1441; add.s32 %r1552, %r1551, %r448; mul.wide.s32 %rd1443, %r1552, 8; add.s64 %rd1444, %rd1442, %rd1443; ld.global.v2.f32 {%f15951, %f15952}, [%rd1444]; bra.uni $L__BB106_922; $L__BB106_908: and.b32 %r1521, %r445, 56; setp.eq.s32 %p779, %r1521, 0; @%p779 bra $L__BB106_913; and.b32 %r1522, %r445, 24; setp.eq.s32 %p780, %r1522, 0; @%p780 bra $L__BB106_911; ld.const.u64 %rd1409, [kernel_params+2232]; cvta.to.global.u64 %rd1410, %rd1409; shl.b64 %rd1411, %rd174, 4; add.s64 %rd1412, %rd1410, %rd1411; ld.global.v4.u32 {%r1523, %r1524, %r1525, %r1526}, [%rd1412]; ld.const.u64 %rd1413, [kernel_params+2328]; cvta.to.global.u64 %rd1414, %rd1413; add.s32 %r1530, %r1523, %r448; mul.wide.u32 %rd1415, %r1530, 8; add.s64 %rd1416, %rd1414, %rd1415; ld.global.v2.f32 {%f15949, %f15950}, [%rd1416]; add.s32 %r1531, %r1524, %r448; mul.wide.u32 %rd1417, %r1531, 8; add.s64 %rd1418, %rd1414, %rd1417; ld.global.v2.f32 {%f15947, %f15948}, [%rd1418]; add.s32 %r1532, %r1525, %r448; mul.wide.u32 %rd1419, %r1532, 8; add.s64 %rd1420, %rd1414, %rd1419; ld.global.v2.f32 {%f15945, %f15946}, [%rd1420]; bra.uni $L__BB106_912; $L__BB106_918: and.b32 %r1546, %r445, 131; setp.eq.s32 %p785, %r1546, 0; mov.f32 %f15951, 0f00000000; mov.f32 %f15952, %f15951; @%p785 bra $L__BB106_922; cvt.u32.u64 %r1547, %rd174; setp.eq.s32 %p786, %r445, 128; selp.b32 %r1548, %r1547, 0, %p786; add.s32 %r1549, %r1548, %r448; ld.const.u64 %rd1437, [kernel_params+2328]; cvta.to.global.u64 %rd1438, %rd1437; mul.wide.s32 %rd1439, %r1549, 8; add.s64 %rd1440, %rd1438, %rd1439; ld.global.v2.f32 {%f15951, %f15952}, [%rd1440]; bra.uni $L__BB106_922; $L__BB106_913: and.b32 %r1535, %r445, 7; setp.eq.s32 %p781, %r1535, 0; mov.f32 %f15951, 0f00000000; mov.f32 %f15952, %f15951; @%p781 bra $L__BB106_922; setp.eq.s32 %p782, %r445, 4; selp.b32 %r1537, %r1520, 0, %p782; add.s32 %r1538, %r1537, %r448; ld.const.u64 %rd1425, [kernel_params+2328]; cvta.to.global.u64 %rd1426, %rd1425; mul.wide.s32 %rd1427, %r1538, 8; add.s64 %rd1428, %rd1426, %rd1427; ld.global.v2.f32 {%f15951, %f15952}, [%rd1428]; bra.uni $L__BB106_922; $L__BB106_2230: mul.ftz.f32 %f14203, %f16625, %f3847; fma.rn.ftz.f32 %f14204, %f16624, %f3846, %f14203; fma.rn.ftz.f32 %f14205, %f16626, %f3848, %f14204; mul.ftz.f32 %f14206, %f16622, %f3847; fma.rn.ftz.f32 %f14207, %f16621, %f3846, %f14206; fma.rn.ftz.f32 %f14208, %f16623, %f3848, %f14207; mul.ftz.f32 %f14209, %f3738, %f3847; fma.rn.ftz.f32 %f14210, %f3736, %f3846, %f14209; fma.rn.ftz.f32 %f14211, %f3739, %f3848, %f14210; neg.ftz.f32 %f14212, %f14205; mul.ftz.f32 %f14213, %f13811, %f14211; div.approx.ftz.f32 %f14214, %f14212, %f14213; neg.ftz.f32 %f14215, %f14208; mul.ftz.f32 %f14216, %f13812, %f14211; div.approx.ftz.f32 %f14217, %f14215, %f14216; mul.ftz.f32 %f14218, %f14211, %f14211; mul.ftz.f32 %f14219, %f14218, %f14218; mul.ftz.f32 %f14220, %f14214, %f14214; neg.ftz.f32 %f14221, %f14220; mul.ftz.f32 %f14222, %f14217, %f14217; sub.ftz.f32 %f14223, %f14221, %f14222; mul.ftz.f32 %f14224, %f14223, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14225, %f14224; mul.ftz.f32 %f14226, %f3854, %f14219; div.approx.ftz.f32 %f16676, %f14225, %f14226; mul.ftz.f32 %f14227, %f16625, %f16736; fma.rn.ftz.f32 %f14228, %f16624, %f16735, %f14227; fma.rn.ftz.f32 %f14229, %f16626, %f16737, %f14228; mul.ftz.f32 %f14230, %f16622, %f16736; fma.rn.ftz.f32 %f14231, %f16621, %f16735, %f14230; fma.rn.ftz.f32 %f14232, %f16623, %f16737, %f14231; mul.ftz.f32 %f14233, %f14232, %f14232; mul.ftz.f32 %f14234, %f14229, %f14229; mul.ftz.f32 %f14235, %f13811, %f13811; mul.ftz.f32 %f14236, %f13812, %f13812; mul.ftz.f32 %f14237, %f14236, %f14233; fma.rn.ftz.f32 %f14238, %f14235, %f14234, %f14237; add.ftz.f32 %f14239, %f14234, %f14233; div.approx.ftz.f32 %f14240, %f14238, %f14239; mul.ftz.f32 %f14241, %f3856, %f14240; div.approx.ftz.f32 %f14242, %f14241, %f3855; mov.f32 %f14243, 0f00000000; max.ftz.f32 %f14244, %f14242, %f14243; sqrt.approx.ftz.f32 %f3858, %f14244; setp.lt.ftz.f32 %p1979, %f3858, 0f3F200000; @%p1979 bra $L__BB106_2234; rcp.approx.ftz.f32 %f14245, %f3858; fma.rn.ftz.f32 %f14246, %f14245, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14247, %f14245, %f14246; fma.rn.ftz.f32 %f14248, %f14245, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14249, %f14245, %f14248, 0f3F800000; div.approx.ftz.f32 %f16677, %f14247, %f14249; $L__BB106_2234: mul.ftz.f32 %f14269, %f16676, 0f3E800000; div.approx.ftz.f32 %f14270, %f14269, %f3740; mul.ftz.f32 %f14271, %f16673, %f16677; mul.ftz.f32 %f16739, %f14271, %f14270; mul.ftz.f32 %f16738, %f16673, %f14270; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; mov.u32 %r3285, %r2865; bra.uni $L__BB106_2309; $L__BB106_2121: mov.f32 %f13186, 0f3F800000; sub.ftz.f32 %f13187, %f13186, %f2799; mul.ftz.f32 %f13188, %f13187, 0f40800000; mul.ftz.f32 %f13189, %f13188, 0f3FC90FDB; sin.approx.ftz.f32 %f13190, %f13189; cos.approx.ftz.f32 %f13191, %f13189; div.approx.ftz.f32 %f13192, %f13190, %f13191; mul.ftz.f32 %f3513, %f3512, %f13192; abs.ftz.f32 %f3514, %f3513; setp.leu.ftz.f32 %p1889, %f3514, 0f3F800000; mov.f32 %f16573, %f3514; @%p1889 bra $L__BB106_2123; rcp.approx.ftz.f32 %f16573, %f3514; $L__BB106_2123: mul.rn.ftz.f32 %f13193, %f16573, %f16573; mov.f32 %f13194, 0fC0B59883; mov.f32 %f13195, 0fBF52C7EA; fma.rn.ftz.f32 %f13196, %f13193, %f13195, %f13194; mov.f32 %f13197, 0fC0D21907; fma.rn.ftz.f32 %f13198, %f13196, %f13193, %f13197; mul.ftz.f32 %f13199, %f13193, %f13198; mul.ftz.f32 %f13200, %f16573, %f13199; add.ftz.f32 %f13201, %f13193, 0f41355DC0; mov.f32 %f13202, 0f41E6BD60; fma.rn.ftz.f32 %f13203, %f13201, %f13193, %f13202; mov.f32 %f13204, 0f419D92C8; fma.rn.ftz.f32 %f13205, %f13203, %f13193, %f13204; rcp.approx.ftz.f32 %f13206, %f13205; fma.rn.ftz.f32 %f13207, %f13200, %f13206, %f16573; mov.f32 %f13208, 0f3FC90FDB; sub.ftz.f32 %f13209, %f13208, %f13207; setp.gt.ftz.f32 %p1890, %f3514, 0f3F800000; selp.f32 %f13210, %f13209, %f13207, %p1890; mov.b32 %r2815, %f13210; mov.b32 %r2816, %f3513; and.b32 %r2817, %r2816, -2147483648; or.b32 %r2818, %r2817, %r2815; mov.b32 %f13211, %r2818; setp.le.ftz.f32 %p1891, %f3514, 0f7F800000; selp.f32 %f13212, %f13211, %f13210, %p1891; cos.approx.ftz.f32 %f13213, %f13212; sin.approx.ftz.f32 %f13214, %f13212; mul.ftz.f32 %f13215, %f3491, %f13213; mul.ftz.f32 %f13216, %f3492, %f13214; mul.ftz.f32 %f13217, %f13214, %f13216; fma.rn.ftz.f32 %f13218, %f13213, %f13215, %f13217; add.ftz.f32 %f13219, %f13218, 0f3F800000; rcp.approx.ftz.f32 %f13220, %f13219; lg2.approx.ftz.f32 %f13221, %f2775; mul.ftz.f32 %f13222, %f13221, %f13220; ex2.approx.ftz.f32 %f16577, %f13222; mov.f32 %f13223, 0f40C90FDB; sub.ftz.f32 %f16578, %f13223, %f13212; $L__BB106_2134: mul.ftz.f32 %f13339, %f16577, %f16577; mov.f32 %f16579, 0f3F800000; sub.ftz.f32 %f13341, %f16579, %f13339; mov.f32 %f16582, 0f00000000; max.ftz.f32 %f13343, %f16582, %f13341; sqrt.approx.ftz.f32 %f13344, %f13343; cos.approx.ftz.f32 %f13345, %f16578; mul.ftz.f32 %f13346, %f13344, %f13345; sin.approx.ftz.f32 %f13347, %f16578; mul.ftz.f32 %f13348, %f13344, %f13347; mul.ftz.f32 %f13349, %f16567, %f13348; mul.ftz.f32 %f13350, %f16568, %f13348; mul.ftz.f32 %f13351, %f16569, %f13348; fma.rn.ftz.f32 %f13352, %f16570, %f13346, %f13349; fma.rn.ftz.f32 %f13353, %f16571, %f13346, %f13350; fma.rn.ftz.f32 %f13354, %f16572, %f13346, %f13351; fma.rn.ftz.f32 %f13355, %f3484, %f16577, %f13352; fma.rn.ftz.f32 %f13356, %f3486, %f16577, %f13353; fma.rn.ftz.f32 %f13357, %f3487, %f16577, %f13354; mul.ftz.f32 %f13358, %f13134, %f13356; fma.rn.ftz.f32 %f13359, %f3485, %f13355, %f13358; fma.rn.ftz.f32 %f13360, %f13135, %f13357, %f13359; setp.lt.ftz.f32 %p1901, %f13360, 0f00000000; neg.ftz.f32 %f13361, %f13355; neg.ftz.f32 %f13362, %f13356; neg.ftz.f32 %f13363, %f13357; selp.f32 %f13364, %f13361, %f13355, %p1901; selp.f32 %f13365, %f13362, %f13356, %p1901; selp.f32 %f13366, %f13363, %f13357, %p1901; add.ftz.f32 %f13367, %f13360, %f13360; mul.ftz.f32 %f13368, %f13367, %f13364; mul.ftz.f32 %f13369, %f13367, %f13365; mul.ftz.f32 %f13370, %f13367, %f13366; sub.ftz.f32 %f16735, %f13368, %f3485; sub.ftz.f32 %f16736, %f13369, %f13134; sub.ftz.f32 %f16737, %f13370, %f13135; max.ftz.f32 %f13371, %f13138, %f13139; setp.le.ftz.f32 %p1902, %f13371, 0f38D1B717; mov.u32 %r3285, 18; mov.f32 %f16738, 0f49742400; mov.f32 %f16739, %f16738; mov.f32 %f16740, %f16738; mov.f32 %f16741, %f16738; @%p1902 bra $L__BB106_2309; mul.ftz.f32 %f13376, %f3486, %f16736; fma.rn.ftz.f32 %f13377, %f3484, %f16735, %f13376; fma.rn.ftz.f32 %f3544, %f3487, %f16737, %f13377; setp.leu.ftz.f32 %p1903, %f3544, 0f00000000; mov.f32 %f16739, %f16582; @%p1903 bra $L__BB106_2142; add.ftz.f32 %f13378, %f3485, %f16735; add.ftz.f32 %f13379, %f13134, %f16736; mul.ftz.f32 %f13380, %f13379, %f13379; fma.rn.ftz.f32 %f13381, %f13378, %f13378, %f13380; add.ftz.f32 %f13382, %f13135, %f16737; fma.rn.ftz.f32 %f13383, %f13382, %f13382, %f13381; sqrt.approx.ftz.f32 %f13384, %f13383; div.approx.ftz.f32 %f3545, %f13378, %f13384; div.approx.ftz.f32 %f3546, %f13379, %f13384; div.approx.ftz.f32 %f3547, %f13382, %f13384; mul.ftz.f32 %f13385, %f13134, %f3546; fma.rn.ftz.f32 %f13386, %f3485, %f3545, %f13385; fma.rn.ftz.f32 %f13387, %f13135, %f3547, %f13386; abs.ftz.f32 %f13388, %f13387; mov.f32 %f13389, 0f358637BD; max.ftz.f32 %f3548, %f13388, %f13389; mul.ftz.f32 %f13390, %f3486, %f3546; fma.rn.ftz.f32 %f13391, %f3484, %f3545, %f13390; fma.rn.ftz.f32 %f13392, %f3487, %f3547, %f13391; max.ftz.f32 %f3549, %f13392, %f13389; max.ftz.f32 %f3550, %f3544, %f13389; @%p1881 bra $L__BB106_2140; bra.uni $L__BB106_2137; $L__BB106_2140: lg2.approx.ftz.f32 %f13443, %f3549; mul.ftz.f32 %f13444, %f3491, %f13443; ex2.approx.ftz.f32 %f13445, %f13444; mov.f32 %f13446, 0f41C90FDB; div.approx.ftz.f32 %f13447, %f3511, %f13446; mul.ftz.f32 %f13448, %f3550, %f13447; mul.ftz.f32 %f16581, %f13445, %f13448; mul.ftz.f32 %f13449, %f13445, %f13447; div.approx.ftz.f32 %f16582, %f13449, %f3548; bra.uni $L__BB106_2141; $L__BB106_2137: ld.local.f32 %f13394, [%rd587+52]; mul.ftz.f32 %f13395, %f3486, %f13394; ld.local.f32 %f13396, [%rd587+48]; mul.ftz.f32 %f13397, %f3487, %f13396; sub.ftz.f32 %f13398, %f13395, %f13397; ld.local.f32 %f13399, [%rd587+44]; mul.ftz.f32 %f13400, %f3487, %f13399; mul.ftz.f32 %f13401, %f3484, %f13394; sub.ftz.f32 %f13402, %f13400, %f13401; mul.ftz.f32 %f13403, %f3484, %f13396; mul.ftz.f32 %f13404, %f3486, %f13399; sub.ftz.f32 %f13405, %f13403, %f13404; mul.ftz.f32 %f13406, %f13402, %f13402; fma.rn.ftz.f32 %f13407, %f13398, %f13398, %f13406; fma.rn.ftz.f32 %f13408, %f13405, %f13405, %f13407; sqrt.approx.ftz.f32 %f13409, %f13408; div.approx.ftz.f32 %f13410, %f13398, %f13409; div.approx.ftz.f32 %f13411, %f13402, %f13409; div.approx.ftz.f32 %f13412, %f13405, %f13409; mul.ftz.f32 %f13413, %f3487, %f13411; mul.ftz.f32 %f13414, %f3486, %f13412; sub.ftz.f32 %f13415, %f13413, %f13414; mul.ftz.f32 %f13416, %f3484, %f13412; mul.ftz.f32 %f13417, %f3487, %f13410; sub.ftz.f32 %f13418, %f13416, %f13417; mul.ftz.f32 %f13419, %f3486, %f13410; mul.ftz.f32 %f13420, %f3484, %f13411; sub.ftz.f32 %f13421, %f13419, %f13420; mul.ftz.f32 %f13422, %f3546, %f13418; fma.rn.ftz.f32 %f13423, %f3545, %f13415, %f13422; fma.rn.ftz.f32 %f3551, %f3547, %f13421, %f13423; mul.ftz.f32 %f13424, %f3546, %f13411; fma.rn.ftz.f32 %f13425, %f3545, %f13410, %f13424; fma.rn.ftz.f32 %f3552, %f3547, %f13412, %f13425; setp.geu.ftz.f32 %p1905, %f3549, 0f3F800000; @%p1905 bra $L__BB106_2139; mul.ftz.f32 %f13426, %f3491, %f3551; mul.ftz.f32 %f13427, %f3551, %f13426; mul.ftz.f32 %f13428, %f3492, %f3552; fma.rn.ftz.f32 %f13429, %f3552, %f13428, %f13427; mul.ftz.f32 %f13430, %f3549, %f3549; mov.f32 %f13431, 0f3F800000; sub.ftz.f32 %f13432, %f13431, %f13430; div.approx.ftz.f32 %f13433, %f13429, %f13432; lg2.approx.ftz.f32 %f13434, %f3549; mul.ftz.f32 %f13435, %f13433, %f13434; ex2.approx.ftz.f32 %f16579, %f13435; $L__BB106_2139: add.ftz.f32 %f13436, %f3492, 0f3F800000; mul.ftz.f32 %f13437, %f3511, %f13436; sqrt.approx.ftz.f32 %f13438, %f13437; mov.f32 %f13439, 0f41C90FDB; div.approx.ftz.f32 %f13440, %f13438, %f13439; mul.ftz.f32 %f13441, %f3550, %f13440; mul.ftz.f32 %f16581, %f16579, %f13441; mul.ftz.f32 %f13442, %f16579, %f13440; div.approx.ftz.f32 %f16582, %f13442, %f3548; $L__BB106_2141: max.ftz.f32 %f13451, %f3488, %f13389; max.ftz.f32 %f13452, %f3550, %f13451; mul.ftz.f32 %f13453, %f3548, %f13452; max.ftz.f32 %f13454, %f13389, %f13453; rcp.approx.ftz.f32 %f13455, %f13454; mul.ftz.f32 %f16739, %f13455, %f16581; $L__BB106_2142: mov.u32 %r3285, 10; mov.f32 %f16738, %f16582; mov.f32 %f16740, %f16739; mov.f32 %f16741, %f16739; bra.uni $L__BB106_2309; $L__BB106_911: mad.lo.s32 %r1534, %r1520, 3, %r448; ld.const.u64 %rd1421, [kernel_params+2328]; cvta.to.global.u64 %rd1422, %rd1421; mul.wide.s32 %rd1423, %r1534, 8; add.s64 %rd1424, %rd1422, %rd1423; ld.global.v2.f32 {%f15949, %f15950}, [%rd1424]; ld.global.v2.f32 {%f15947, %f15948}, [%rd1424+8]; ld.global.v2.f32 {%f15945, %f15946}, [%rd1424+16]; $L__BB106_912: ld.local.f32 %f7060, [%rd7+-20]; ld.local.f32 %f7061, [%rd7+-16]; mul.ftz.f32 %f7062, %f15945, %f7061; mul.ftz.f32 %f7063, %f15946, %f7061; fma.rn.ftz.f32 %f7064, %f15947, %f7060, %f7062; fma.rn.ftz.f32 %f7065, %f15948, %f7060, %f7063; mov.f32 %f7066, 0f3F800000; sub.ftz.f32 %f7067, %f7066, %f7060; sub.ftz.f32 %f7068, %f7067, %f7061; fma.rn.ftz.f32 %f15952, %f15950, %f7068, %f7065; fma.rn.ftz.f32 %f15951, %f15949, %f7068, %f7064; $L__BB106_922: mov.f32 %f15955, 0f3F800000; $L__BB106_923: ld.const.s32 %rd1445, [kernel_params+1296]; add.s64 %rd1446, %rd166, %rd1445; cvta.to.global.u64 %rd1447, %rd618; shl.b64 %rd1448, %rd1446, 2; add.s64 %rd1449, %rd1447, %rd1448; add.s64 %rd1450, %rd1449, 4; add.s64 %rd1451, %rd1449, 8; atom.global.add.f32 %f7094, [%rd1449], %f15951; atom.global.add.f32 %f7095, [%rd1450], %f15952; atom.global.add.f32 %f7096, [%rd1451], %f15955; $L__BB106_924: and.b32 %r1553, %r402, 128; setp.eq.s32 %p790, %r1553, 0; @%p790 bra $L__BB106_1145; ld.local.u32 %r454, [%rd7+-24]; and.b32 %r455, %r454, 14; setp.ne.s32 %p791, %r455, 0; @%p791 bra $L__BB106_927; bra.uni $L__BB106_926; $L__BB106_927: and.b32 %r1554, %r454, 6; setp.eq.s32 %p792, %r1554, 0; @%p792 bra $L__BB106_929; ld.const.u64 %rd1454, [kernel_params+2264]; cvta.to.global.u64 %rd1455, %rd1454; ld.local.u32 %r1555, [%rd7+-28]; mul.wide.s32 %rd1456, %r1555, 16; add.s64 %rd1457, %rd1455, %rd1456; shr.s32 %r1556, %r454, 7; ld.global.u32 %r1557, [%rd1457+4]; add.s32 %r1558, %r1556, %r1557; ld.const.u64 %rd1458, [kernel_params+2272]; cvta.to.global.u64 %rd1459, %rd1458; mul.wide.s32 %rd1460, %r1558, 16; add.s64 %rd1461, %rd1459, %rd1460; ld.global.v4.f32 {%f7101, %f7102, %f7103, %f7104}, [%rd1461]; ld.global.v4.f32 {%f7108, %f7109, %f7110, %f7111}, [%rd1461+16]; ld.local.f32 %f7115, [%rd7+-20]; mov.f32 %f7116, 0f3F800000; sub.ftz.f32 %f7117, %f7116, %f7115; mul.ftz.f32 %f7118, %f7101, %f7117; mul.ftz.f32 %f7119, %f7102, %f7117; mul.ftz.f32 %f7120, %f7103, %f7117; fma.rn.ftz.f32 %f15959, %f7108, %f7115, %f7118; fma.rn.ftz.f32 %f15960, %f7109, %f7115, %f7119; fma.rn.ftz.f32 %f15961, %f7110, %f7115, %f7120; bra.uni $L__BB106_931; $L__BB106_926: ld.local.v4.f32 {%f15959, %f15960, %f15961, %f7100}, [%rd1]; bra.uni $L__BB106_935; $L__BB106_929: and.b32 %r1559, %r454, 8; setp.eq.s32 %p793, %r1559, 0; mov.f32 %f15959, 0f00000000; mov.f32 %f15960, %f15959; mov.f32 %f15961, %f15959; @%p793 bra $L__BB106_931; ld.const.u64 %rd1462, [kernel_params+2296]; cvta.to.global.u64 %rd1463, %rd1462; ld.local.u32 %r1560, [%rd7+-28]; mul.wide.s32 %rd1464, %r1560, 16; add.s64 %rd1465, %rd1463, %rd1464; ld.global.v4.f32 {%f15959, %f15960, %f15961, %f7127}, [%rd1465]; $L__BB106_931: ld.local.u32 %r456, [%rd7+-32]; and.b32 %r1561, %r456, 4; setp.ne.s32 %p794, %r1561, 0; @%p794 bra $L__BB106_935; and.b32 %r1562, %r456, 2; setp.eq.s32 %p795, %r1562, 0; @%p795 bra $L__BB106_934; bra.uni $L__BB106_933; $L__BB106_934: ld.const.u64 %rd1466, [kernel_params+2160]; cvta.to.global.u64 %rd1467, %rd1466; ld.local.u32 %r1563, [%rd7+-12]; mul.wide.s32 %rd1468, %r1563, 208; add.s64 %rd1469, %rd1467, %rd1468; ld.global.v4.f32 {%f7161, %f7162, %f7163, %f7164}, [%rd1469]; ld.global.v4.f32 {%f7169, %f7170, %f7171, %f7172}, [%rd1469+16]; ld.global.v4.f32 {%f7177, %f7178, %f7179, %f7180}, [%rd1469+32]; mul.ftz.f32 %f7185, %f15960, %f7162; fma.rn.ftz.f32 %f7186, %f15959, %f7161, %f7185; fma.rn.ftz.f32 %f7187, %f15961, %f7163, %f7186; add.ftz.f32 %f1288, %f7164, %f7187; mul.ftz.f32 %f7188, %f15960, %f7170; fma.rn.ftz.f32 %f7189, %f15959, %f7169, %f7188; fma.rn.ftz.f32 %f7190, %f15961, %f7171, %f7189; add.ftz.f32 %f1289, %f7172, %f7190; mul.ftz.f32 %f7191, %f15960, %f7178; fma.rn.ftz.f32 %f7192, %f15959, %f7177, %f7191; fma.rn.ftz.f32 %f7193, %f15961, %f7179, %f7192; add.ftz.f32 %f15961, %f7180, %f7193; mov.f32 %f15959, %f1288; mov.f32 %f15960, %f1289; bra.uni $L__BB106_935; $L__BB106_933: ld.local.v4.f32 {%f7128, %f7129, %f7130, %f7131}, [%rd7+56]; mul.ftz.f32 %f7136, %f15960, %f7129; fma.rn.ftz.f32 %f7137, %f15959, %f7128, %f7136; fma.rn.ftz.f32 %f7138, %f15961, %f7130, %f7137; add.ftz.f32 %f1285, %f7131, %f7138; ld.local.v4.f32 {%f7139, %f7140, %f7141, %f7142}, [%rd7+72]; mul.ftz.f32 %f7147, %f15960, %f7140; fma.rn.ftz.f32 %f7148, %f15959, %f7139, %f7147; fma.rn.ftz.f32 %f7149, %f15961, %f7141, %f7148; add.ftz.f32 %f1286, %f7142, %f7149; ld.local.v4.f32 {%f7150, %f7151, %f7152, %f7153}, [%rd7+88]; mul.ftz.f32 %f7158, %f15960, %f7151; fma.rn.ftz.f32 %f7159, %f15959, %f7150, %f7158; fma.rn.ftz.f32 %f7160, %f15961, %f7152, %f7159; add.ftz.f32 %f15961, %f7153, %f7160; mov.f32 %f15959, %f1285; mov.f32 %f15960, %f1286; $L__BB106_935: ld.local.u32 %r3200, [%rd7+-12]; cvt.s64.s32 %rd186, %r3200; ld.local.s32 %rd187, [%rd7+-28]; setp.eq.s32 %p796, %r3200, -1; mov.u32 %r478, -1; mov.u32 %r475, 0; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p796 bra $L__BB106_950; ld.const.u64 %rd1470, [kernel_params+2160]; cvta.to.global.u64 %rd1471, %rd1470; mul.lo.s64 %rd1472, %rd186, 208; add.s64 %rd1473, %rd1471, %rd1472; ld.global.u32 %r458, [%rd1473+164]; and.b32 %r1569, %r454, 1; setp.eq.b32 %p797, %r1569, 1; mov.pred %p798, 0; xor.pred %p799, %p797, %p798; not.pred %p800, %p799; mov.u32 %r3190, 0; @%p800 bra $L__BB106_940; cvt.u32.u64 %r1571, %rd187; setp.eq.s32 %p801, %r1571, -1; mov.u32 %r3189, -1; @%p801 bra $L__BB106_939; ld.const.u64 %rd1474, [kernel_params+2240]; cvta.to.global.u64 %rd1475, %rd1474; shl.b64 %rd1476, %rd187, 2; add.s64 %rd1477, %rd1475, %rd1476; ld.global.u32 %r3189, [%rd1477]; $L__BB106_939: setp.ne.s32 %p802, %r3189, -1; selp.u32 %r3190, 1, 0, %p802; $L__BB106_940: ld.const.u64 %rd1478, [kernel_params+2312]; cvta.to.global.u64 %rd188, %rd1478; add.s32 %r3192, %r3190, %r458; mul.wide.u32 %rd1479, %r3192, 16; add.s64 %rd1480, %rd188, %rd1479; add.s64 %rd2963, %rd1480, 14; ld.global.u16 %rs143, [%rd1480+12]; ld.global.u32 %r3191, [%rd1480+8]; ld.global.u64 %rd2961, [%rd1480]; setp.eq.s64 %p803, %rd2961, 11; add.s64 %rd2962, %rd1480, 15; @%p803 bra $L__BB106_946; $L__BB106_941: setp.eq.s64 %p804, %rd2961, 0; @%p804 bra $L__BB106_943; add.s32 %r3192, %r3192, 2; bra.uni $L__BB106_944; $L__BB106_943: setp.eq.s16 %p805, %rs143, 0; mov.u32 %r475, 0; mov.u32 %r3192, %r3191; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p805 bra $L__BB106_950; $L__BB106_944: cvt.u64.u32 %rd193, %r3192; mul.wide.u32 %rd1481, %r3192, 16; add.s64 %rd1482, %rd188, %rd1481; ld.global.u16 %rs143, [%rd1482+12]; ld.global.u32 %r3191, [%rd1482+8]; ld.global.u64 %rd2961, [%rd1482]; setp.ne.s64 %p806, %rd2961, 11; @%p806 bra $L__BB106_941; shl.b64 %rd1483, %rd193, 4; add.s64 %rd1484, %rd188, %rd1483; add.s64 %rd2963, %rd1484, 14; add.s64 %rd2962, %rd1484, 15; $L__BB106_946: ld.global.u8 %rs26, [%rd2963]; ld.global.u8 %rs27, [%rd2962]; cvt.u32.u64 %r1576, %rd187; setp.ne.s32 %p807, %r1576, -1; @%p807 bra $L__BB106_949; mov.u32 %r475, 0; add.s16 %rs79, %rs143, -1; setp.lt.u16 %p808, %rs79, 2; @%p808 bra $L__BB106_949; setp.ne.s16 %p809, %rs143, 1024; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p809 bra $L__BB106_950; $L__BB106_949: setp.eq.s16 %p810, %rs143, 0; selp.b32 %r478, -1, %r3191, %p810; cvt.u32.u16 %r475, %rs143; cvt.u32.u16 %r1581, %rs27; and.b32 %r3197, %r1581, 255; cvt.u32.u16 %r1582, %rs26; and.b32 %r3196, %r1582, 255; $L__BB106_950: setp.eq.s32 %p811, %r478, -1; mov.f32 %f15977, %f15959; mov.f32 %f15978, %f15960; mov.f32 %f15979, %f15961; mov.f32 %f15980, %f15959; mov.f32 %f15981, %f15960; mov.f32 %f15982, %f15961; @%p811 bra $L__BB106_986; ld.const.u64 %rd1485, [kernel_params+2160]; cvta.to.global.u64 %rd1486, %rd1485; mul.lo.s64 %rd1487, %rd186, 208; add.s64 %rd1488, %rd1486, %rd1487; add.s64 %rd199, %rd1488, 148; ld.global.u32 %r481, [%rd1488+156]; @%p791 bra $L__BB106_971; bra.uni $L__BB106_952; $L__BB106_971: ld.global.u32 %r487, [%rd199]; and.b32 %r488, %r475, 768; setp.eq.s32 %p827, %r488, 0; @%p827 bra $L__BB106_973; ld.const.u64 %rd1527, [kernel_params+2264]; cvta.to.global.u64 %rd1528, %rd1527; shl.b64 %rd1529, %rd187, 4; add.s64 %rd1530, %rd1528, %rd1529; shr.s32 %r1621, %r454, 7; ld.global.u32 %r1622, [%rd1530+4]; add.s32 %r1623, %r1621, %r1622; ld.const.u64 %rd1531, [kernel_params+2344]; cvta.to.global.u64 %rd1532, %rd1531; add.s32 %r1624, %r1623, %r478; mul.wide.s32 %rd1533, %r1624, 16; add.s64 %rd1534, %rd1532, %rd1533; ld.global.v4.f32 {%f7240, %f7241, %f7242, %f7243}, [%rd1534]; ld.global.v4.f32 {%f7248, %f7249, %f7250, %f7251}, [%rd1534+16]; ld.local.f32 %f7256, [%rd7+-20]; mov.f32 %f7257, 0f3F800000; sub.ftz.f32 %f7258, %f7257, %f7256; mul.ftz.f32 %f7259, %f7240, %f7258; mul.ftz.f32 %f7260, %f7241, %f7258; mul.ftz.f32 %f7261, %f7242, %f7258; fma.rn.ftz.f32 %f15982, %f7250, %f7256, %f7261; fma.rn.ftz.f32 %f15981, %f7249, %f7256, %f7260; fma.rn.ftz.f32 %f15980, %f7248, %f7256, %f7259; bra.uni $L__BB106_975; $L__BB106_952: and.b32 %r1583, %r454, 1; setp.eq.b32 %p813, %r1583, 1; mov.pred %p814, 0; xor.pred %p815, %p813, %p814; not.pred %p816, %p815; mov.f32 %f15977, %f15959; mov.f32 %f15978, %f15960; mov.f32 %f15979, %f15961; mov.f32 %f15980, %f15959; mov.f32 %f15981, %f15960; mov.f32 %f15982, %f15961; @%p816 bra $L__BB106_986; cvt.u32.u64 %r1585, %rd187; setp.eq.s32 %p817, %r1585, -1; mov.u32 %r3199, -1; @%p817 bra $L__BB106_955; ld.const.u64 %rd1489, [kernel_params+2240]; cvta.to.global.u64 %rd1490, %rd1489; shl.b64 %rd1491, %rd187, 2; add.s64 %rd1492, %rd1490, %rd1491; ld.global.u32 %r3199, [%rd1492]; $L__BB106_955: setp.eq.s32 %p818, %r3199, -1; @%p818 bra $L__BB106_957; mov.u64 %rd1493, 0; { // callseq 168, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r3196; st.param.b32 [param1+8], %r3197; st.param.b32 [param1+12], %r478; .param .b64 param2; st.param.b64 [param2+0], %rd1493; .param .b64 param3; st.param.b64 [param3+0], %rd1493; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15980, [retval0+0]; ld.param.f32 %f15981, [retval0+4]; ld.param.f32 %f15982, [retval0+8]; } // callseq 168 add.s32 %r1586, %r481, %r478; { // callseq 169, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r3196; st.param.b32 [param1+8], %r3197; st.param.b32 [param1+12], %r1586; .param .b64 param2; st.param.b64 [param2+0], %rd1493; .param .b64 param3; st.param.b64 [param3+0], %rd1493; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15977, [retval0+0]; ld.param.f32 %f15978, [retval0+4]; ld.param.f32 %f15979, [retval0+8]; } // callseq 169 ld.local.u32 %r3200, [%rd7+-12]; bra.uni $L__BB106_986; $L__BB106_973: and.b32 %r1625, %r475, 131; setp.eq.s32 %p828, %r1625, 0; mov.f32 %f15980, 0f00000000; mov.f32 %f15981, %f15980; mov.f32 %f15982, %f15980; @%p828 bra $L__BB106_975; cvt.u32.u64 %r1626, %rd187; setp.eq.s32 %p829, %r475, 128; selp.b32 %r1627, %r1626, 0, %p829; add.s32 %r1628, %r1627, %r478; ld.const.u64 %rd1535, [kernel_params+2344]; cvta.to.global.u64 %rd1536, %rd1535; mul.wide.s32 %rd1537, %r1628, 16; add.s64 %rd1538, %rd1536, %rd1537; ld.global.v4.f32 {%f15980, %f15981, %f15982, %f7270}, [%rd1538]; $L__BB106_975: add.s32 %r489, %r487, %r478; @%p827 bra $L__BB106_977; cvt.u32.u64 %r1629, %rd187; ld.const.u64 %rd1539, [kernel_params+2264]; cvta.to.global.u64 %rd1540, %rd1539; mul.wide.s32 %rd1541, %r1629, 16; add.s64 %rd1542, %rd1540, %rd1541; shr.s32 %r1630, %r454, 7; ld.global.u32 %r1631, [%rd1542+4]; add.s32 %r1632, %r1630, %r1631; ld.const.u64 %rd1543, [kernel_params+2344]; cvta.to.global.u64 %rd1544, %rd1543; add.s32 %r1633, %r1632, %r489; mul.wide.s32 %rd1545, %r1633, 16; add.s64 %rd1546, %rd1544, %rd1545; ld.global.v4.f32 {%f7271, %f7272, %f7273, %f7274}, [%rd1546]; ld.global.v4.f32 {%f7279, %f7280, %f7281, %f7282}, [%rd1546+16]; ld.local.f32 %f7287, [%rd7+-20]; mov.f32 %f7288, 0f3F800000; sub.ftz.f32 %f7289, %f7288, %f7287; mul.ftz.f32 %f7290, %f7271, %f7289; mul.ftz.f32 %f7291, %f7272, %f7289; mul.ftz.f32 %f7292, %f7273, %f7289; fma.rn.ftz.f32 %f15979, %f7281, %f7287, %f7292; fma.rn.ftz.f32 %f15978, %f7280, %f7287, %f7291; fma.rn.ftz.f32 %f15977, %f7279, %f7287, %f7290; bra.uni $L__BB106_979; $L__BB106_977: and.b32 %r1634, %r475, 131; setp.eq.s32 %p831, %r1634, 0; mov.f32 %f15977, 0f00000000; mov.f32 %f15978, %f15977; mov.f32 %f15979, %f15977; @%p831 bra $L__BB106_979; cvt.u32.u64 %r1635, %rd187; setp.eq.s32 %p832, %r475, 128; selp.b32 %r1636, %r1635, 0, %p832; add.s32 %r1637, %r1636, %r489; ld.const.u64 %rd1547, [kernel_params+2344]; cvta.to.global.u64 %rd1548, %rd1547; mul.wide.s32 %rd1549, %r1637, 16; add.s64 %rd1550, %rd1548, %rd1549; ld.global.v4.f32 {%f15977, %f15978, %f15979, %f7301}, [%rd1550]; $L__BB106_979: ld.local.u32 %r490, [%rd7+-32]; and.b32 %r1638, %r490, 64; setp.ne.s32 %p833, %r1638, 0; @%p833 bra $L__BB106_986; and.b32 %r491, %r490, 2; setp.eq.s32 %p834, %r491, 0; @%p834 bra $L__BB106_982; bra.uni $L__BB106_981; $L__BB106_982: ld.global.v4.f32 {%f7335, %f7336, %f7337, %f7338}, [%rd199+-148]; ld.global.v4.f32 {%f7343, %f7344, %f7345, %f7346}, [%rd199+-132]; ld.global.v4.f32 {%f7351, %f7352, %f7353, %f7354}, [%rd199+-116]; mul.ftz.f32 %f7359, %f15981, %f7336; fma.rn.ftz.f32 %f7360, %f15980, %f7335, %f7359; fma.rn.ftz.f32 %f7361, %f15982, %f7337, %f7360; add.ftz.f32 %f15974, %f7338, %f7361; mul.ftz.f32 %f7362, %f15981, %f7344; fma.rn.ftz.f32 %f7363, %f15980, %f7343, %f7362; fma.rn.ftz.f32 %f7364, %f15982, %f7345, %f7363; add.ftz.f32 %f15975, %f7346, %f7364; mul.ftz.f32 %f7365, %f15981, %f7352; fma.rn.ftz.f32 %f7366, %f15980, %f7351, %f7365; fma.rn.ftz.f32 %f7367, %f15982, %f7353, %f7366; add.ftz.f32 %f15982, %f7354, %f7367; bra.uni $L__BB106_983; $L__BB106_981: ld.local.v4.f32 {%f7302, %f7303, %f7304, %f7305}, [%rd7+56]; mul.ftz.f32 %f7310, %f15981, %f7303; fma.rn.ftz.f32 %f7311, %f15980, %f7302, %f7310; fma.rn.ftz.f32 %f7312, %f15982, %f7304, %f7311; add.ftz.f32 %f15974, %f7305, %f7312; ld.local.v4.f32 {%f7313, %f7314, %f7315, %f7316}, [%rd7+72]; mul.ftz.f32 %f7321, %f15981, %f7314; fma.rn.ftz.f32 %f7322, %f15980, %f7313, %f7321; fma.rn.ftz.f32 %f7323, %f15982, %f7315, %f7322; add.ftz.f32 %f15975, %f7316, %f7323; ld.local.v4.f32 {%f7324, %f7325, %f7326, %f7327}, [%rd7+88]; mul.ftz.f32 %f7332, %f15981, %f7325; fma.rn.ftz.f32 %f7333, %f15980, %f7324, %f7332; fma.rn.ftz.f32 %f7334, %f15982, %f7326, %f7333; add.ftz.f32 %f15982, %f7327, %f7334; $L__BB106_983: @%p834 bra $L__BB106_985; bra.uni $L__BB106_984; $L__BB106_985: ld.global.v4.f32 {%f7401, %f7402, %f7403, %f7404}, [%rd199+-148]; ld.global.v4.f32 {%f7409, %f7410, %f7411, %f7412}, [%rd199+-132]; ld.global.v4.f32 {%f7417, %f7418, %f7419, %f7420}, [%rd199+-116]; mul.ftz.f32 %f7425, %f15978, %f7402; fma.rn.ftz.f32 %f7426, %f15977, %f7401, %f7425; fma.rn.ftz.f32 %f7427, %f15979, %f7403, %f7426; add.ftz.f32 %f1366, %f7404, %f7427; mul.ftz.f32 %f7428, %f15978, %f7410; fma.rn.ftz.f32 %f7429, %f15977, %f7409, %f7428; fma.rn.ftz.f32 %f7430, %f15979, %f7411, %f7429; add.ftz.f32 %f1367, %f7412, %f7430; mul.ftz.f32 %f7431, %f15978, %f7418; fma.rn.ftz.f32 %f7432, %f15977, %f7417, %f7431; fma.rn.ftz.f32 %f7433, %f15979, %f7419, %f7432; add.ftz.f32 %f15979, %f7420, %f7433; mov.f32 %f15977, %f1366; mov.f32 %f15978, %f1367; mov.f32 %f15980, %f15974; mov.f32 %f15981, %f15975; bra.uni $L__BB106_986; $L__BB106_984: ld.local.v4.f32 {%f7368, %f7369, %f7370, %f7371}, [%rd7+56]; mul.ftz.f32 %f7376, %f15978, %f7369; fma.rn.ftz.f32 %f7377, %f15977, %f7368, %f7376; fma.rn.ftz.f32 %f7378, %f15979, %f7370, %f7377; add.ftz.f32 %f1363, %f7371, %f7378; ld.local.v4.f32 {%f7379, %f7380, %f7381, %f7382}, [%rd7+72]; mul.ftz.f32 %f7387, %f15978, %f7380; fma.rn.ftz.f32 %f7388, %f15977, %f7379, %f7387; fma.rn.ftz.f32 %f7389, %f15979, %f7381, %f7388; add.ftz.f32 %f1364, %f7382, %f7389; ld.local.v4.f32 {%f7390, %f7391, %f7392, %f7393}, [%rd7+88]; mul.ftz.f32 %f7398, %f15978, %f7391; fma.rn.ftz.f32 %f7399, %f15977, %f7390, %f7398; fma.rn.ftz.f32 %f7400, %f15979, %f7392, %f7399; add.ftz.f32 %f15979, %f7393, %f7400; mov.f32 %f15977, %f1363; mov.f32 %f15978, %f1364; mov.f32 %f15980, %f15974; mov.f32 %f15981, %f15975; bra.uni $L__BB106_986; $L__BB106_957: and.b32 %r485, %r475, 56; setp.eq.s32 %p819, %r485, 0; @%p819 bra $L__BB106_962; and.b32 %r1587, %r475, 24; setp.eq.s32 %p820, %r1587, 0; @%p820 bra $L__BB106_960; ld.const.u64 %rd1495, [kernel_params+2232]; cvta.to.global.u64 %rd1496, %rd1495; shl.b64 %rd1497, %rd187, 4; add.s64 %rd1498, %rd1496, %rd1497; ld.global.v4.u32 {%r1588, %r1589, %r1590, %r1591}, [%rd1498]; ld.const.u64 %rd1499, [kernel_params+2336]; cvta.to.global.u64 %rd1500, %rd1499; add.s32 %r1595, %r1588, %r478; mul.wide.u32 %rd1501, %r1595, 12; add.s64 %rd2970, %rd1500, %rd1501; add.s32 %r1596, %r1589, %r478; mul.wide.u32 %rd1502, %r1596, 12; add.s64 %rd2967, %rd1500, %rd1502; add.s32 %r1597, %r1590, %r478; mul.wide.u32 %rd1503, %r1597, 12; add.s64 %rd2964, %rd1500, %rd1503; add.s64 %rd2971, %rd2970, 4; add.s64 %rd2972, %rd2970, 8; add.s64 %rd2968, %rd2967, 4; add.s64 %rd2969, %rd2967, 8; add.s64 %rd2965, %rd2964, 4; add.s64 %rd2966, %rd2964, 8; bra.uni $L__BB106_961; $L__BB106_962: and.b32 %r1600, %r475, 7; setp.eq.s32 %p821, %r1600, 0; mov.f32 %f15980, 0f00000000; mov.f32 %f15981, %f15980; mov.f32 %f15982, %f15980; @%p821 bra $L__BB106_964; setp.eq.s32 %p822, %r475, 4; selp.b32 %r1602, %r1585, 0, %p822; add.s32 %r1603, %r1602, %r478; ld.const.u64 %rd1507, [kernel_params+2336]; cvta.to.global.u64 %rd1508, %rd1507; mul.wide.s32 %rd1509, %r1603, 12; add.s64 %rd1510, %rd1508, %rd1509; ld.global.f32 %f15982, [%rd1510+8]; ld.global.f32 %f15981, [%rd1510+4]; ld.global.f32 %f15980, [%rd1510]; bra.uni $L__BB106_964; $L__BB106_960: mad.lo.s32 %r1599, %r1585, 3, %r478; ld.const.u64 %rd1504, [kernel_params+2336]; cvta.to.global.u64 %rd1505, %rd1504; mul.wide.s32 %rd1506, %r1599, 12; add.s64 %rd2970, %rd1505, %rd1506; add.s64 %rd2971, %rd2970, 4; add.s64 %rd2972, %rd2970, 8; add.s64 %rd2967, %rd2970, 12; add.s64 %rd2968, %rd2970, 16; add.s64 %rd2969, %rd2970, 20; add.s64 %rd2964, %rd2970, 24; add.s64 %rd2965, %rd2970, 28; add.s64 %rd2966, %rd2970, 32; $L__BB106_961: ld.local.f32 %f7194, [%rd7+-20]; ld.global.f32 %f7195, [%rd2967]; ld.global.f32 %f7196, [%rd2968]; ld.global.f32 %f7197, [%rd2969]; ld.local.f32 %f7198, [%rd7+-16]; ld.global.f32 %f7199, [%rd2964]; mul.ftz.f32 %f7200, %f7199, %f7198; ld.global.f32 %f7201, [%rd2965]; mul.ftz.f32 %f7202, %f7201, %f7198; ld.global.f32 %f7203, [%rd2966]; mul.ftz.f32 %f7204, %f7203, %f7198; fma.rn.ftz.f32 %f7205, %f7195, %f7194, %f7200; fma.rn.ftz.f32 %f7206, %f7196, %f7194, %f7202; fma.rn.ftz.f32 %f7207, %f7197, %f7194, %f7204; mov.f32 %f7208, 0f3F800000; sub.ftz.f32 %f7209, %f7208, %f7194; sub.ftz.f32 %f7210, %f7209, %f7198; ld.global.f32 %f7211, [%rd2970]; ld.global.f32 %f7212, [%rd2971]; ld.global.f32 %f7213, [%rd2972]; fma.rn.ftz.f32 %f15982, %f7213, %f7210, %f7207; fma.rn.ftz.f32 %f15981, %f7212, %f7210, %f7206; fma.rn.ftz.f32 %f15980, %f7211, %f7210, %f7205; $L__BB106_964: add.s32 %r486, %r481, %r478; @%p819 bra $L__BB106_969; and.b32 %r1604, %r475, 24; setp.eq.s32 %p824, %r1604, 0; @%p824 bra $L__BB106_967; ld.const.u64 %rd1511, [kernel_params+2232]; cvta.to.global.u64 %rd1512, %rd1511; shl.b64 %rd1513, %rd187, 4; add.s64 %rd1514, %rd1512, %rd1513; ld.global.v4.u32 {%r1605, %r1606, %r1607, %r1608}, [%rd1514]; ld.const.u64 %rd1515, [kernel_params+2336]; cvta.to.global.u64 %rd1516, %rd1515; add.s32 %r1612, %r1605, %r486; mul.wide.u32 %rd1517, %r1612, 12; add.s64 %rd2979, %rd1516, %rd1517; add.s32 %r1613, %r1606, %r486; mul.wide.u32 %rd1518, %r1613, 12; add.s64 %rd2976, %rd1516, %rd1518; add.s32 %r1614, %r1607, %r486; mul.wide.u32 %rd1519, %r1614, 12; add.s64 %rd2973, %rd1516, %rd1519; add.s64 %rd2980, %rd2979, 4; add.s64 %rd2981, %rd2979, 8; add.s64 %rd2977, %rd2976, 4; add.s64 %rd2978, %rd2976, 8; add.s64 %rd2974, %rd2973, 4; add.s64 %rd2975, %rd2973, 8; bra.uni $L__BB106_968; $L__BB106_969: and.b32 %r1617, %r475, 7; setp.eq.s32 %p825, %r1617, 0; mov.f32 %f15977, 0f00000000; mov.f32 %f15978, %f15977; mov.f32 %f15979, %f15977; @%p825 bra $L__BB106_986; setp.eq.s32 %p826, %r475, 4; selp.b32 %r1619, %r1585, 0, %p826; add.s32 %r1620, %r1619, %r486; ld.const.u64 %rd1523, [kernel_params+2336]; cvta.to.global.u64 %rd1524, %rd1523; mul.wide.s32 %rd1525, %r1620, 12; add.s64 %rd1526, %rd1524, %rd1525; ld.global.f32 %f15979, [%rd1526+8]; ld.global.f32 %f15978, [%rd1526+4]; ld.global.f32 %f15977, [%rd1526]; bra.uni $L__BB106_986; $L__BB106_967: mad.lo.s32 %r1616, %r1585, 3, %r486; ld.const.u64 %rd1520, [kernel_params+2336]; cvta.to.global.u64 %rd1521, %rd1520; mul.wide.s32 %rd1522, %r1616, 12; add.s64 %rd2979, %rd1521, %rd1522; add.s64 %rd2980, %rd2979, 4; add.s64 %rd2981, %rd2979, 8; add.s64 %rd2976, %rd2979, 12; add.s64 %rd2977, %rd2979, 16; add.s64 %rd2978, %rd2979, 20; add.s64 %rd2973, %rd2979, 24; add.s64 %rd2974, %rd2979, 28; add.s64 %rd2975, %rd2979, 32; $L__BB106_968: ld.local.f32 %f7217, [%rd7+-20]; ld.global.f32 %f7218, [%rd2976]; ld.global.f32 %f7219, [%rd2977]; ld.global.f32 %f7220, [%rd2978]; ld.local.f32 %f7221, [%rd7+-16]; ld.global.f32 %f7222, [%rd2973]; mul.ftz.f32 %f7223, %f7222, %f7221; ld.global.f32 %f7224, [%rd2974]; mul.ftz.f32 %f7225, %f7224, %f7221; ld.global.f32 %f7226, [%rd2975]; mul.ftz.f32 %f7227, %f7226, %f7221; fma.rn.ftz.f32 %f7228, %f7218, %f7217, %f7223; fma.rn.ftz.f32 %f7229, %f7219, %f7217, %f7225; fma.rn.ftz.f32 %f7230, %f7220, %f7217, %f7227; mov.f32 %f7231, 0f3F800000; sub.ftz.f32 %f7232, %f7231, %f7217; sub.ftz.f32 %f7233, %f7232, %f7221; ld.global.f32 %f7234, [%rd2979]; ld.global.f32 %f7235, [%rd2980]; ld.global.f32 %f7236, [%rd2981]; fma.rn.ftz.f32 %f15979, %f7236, %f7233, %f7230; fma.rn.ftz.f32 %f15978, %f7235, %f7233, %f7229; fma.rn.ftz.f32 %f15977, %f7234, %f7233, %f7228; $L__BB106_986: ld.const.u64 %rd1551, [kernel_params+2168]; cvta.to.global.u64 %rd1552, %rd1551; shl.b32 %r1639, %r3200, 1; mul.wide.s32 %rd1553, %r1639, 48; add.s64 %rd1554, %rd1552, %rd1553; ld.global.v4.f32 {%f7434, %f7435, %f7436, %f7437}, [%rd1554]; ld.global.v4.f32 {%f7442, %f7443, %f7444, %f7445}, [%rd1554+16]; ld.global.v4.f32 {%f7450, %f7451, %f7452, %f7453}, [%rd1554+32]; mul.ftz.f32 %f7458, %f15981, %f7435; fma.rn.ftz.f32 %f7459, %f15980, %f7434, %f7458; fma.rn.ftz.f32 %f7460, %f15982, %f7436, %f7459; add.ftz.f32 %f1375, %f7437, %f7460; mul.ftz.f32 %f7461, %f15981, %f7443; fma.rn.ftz.f32 %f7462, %f15980, %f7442, %f7461; fma.rn.ftz.f32 %f7463, %f15982, %f7444, %f7462; add.ftz.f32 %f1376, %f7445, %f7463; mul.ftz.f32 %f7464, %f15981, %f7451; fma.rn.ftz.f32 %f7465, %f15980, %f7450, %f7464; fma.rn.ftz.f32 %f7466, %f15982, %f7452, %f7465; add.ftz.f32 %f1377, %f7453, %f7466; ld.global.v4.f32 {%f7467, %f7468, %f7469, %f7470}, [%rd1554+48]; ld.global.v4.f32 {%f7475, %f7476, %f7477, %f7478}, [%rd1554+64]; ld.global.v4.f32 {%f7483, %f7484, %f7485, %f7486}, [%rd1554+80]; mul.ftz.f32 %f7491, %f15978, %f7468; fma.rn.ftz.f32 %f7492, %f15977, %f7467, %f7491; fma.rn.ftz.f32 %f7493, %f15979, %f7469, %f7492; add.ftz.f32 %f1378, %f7470, %f7493; mul.ftz.f32 %f7494, %f15978, %f7476; fma.rn.ftz.f32 %f7495, %f15977, %f7475, %f7494; fma.rn.ftz.f32 %f7496, %f15979, %f7477, %f7495; add.ftz.f32 %f1379, %f7478, %f7496; mul.ftz.f32 %f7497, %f15978, %f7484; fma.rn.ftz.f32 %f7498, %f15977, %f7483, %f7497; fma.rn.ftz.f32 %f7499, %f15979, %f7485, %f7498; add.ftz.f32 %f1380, %f7486, %f7499; ld.const.u32 %r1640, [kernel_params+16]; setp.eq.s32 %p836, %r1640, 2; @%p836 bra $L__BB106_993; bra.uni $L__BB106_987; $L__BB106_993: ld.const.v4.f32 {%f7641, %f7642, %f7643, %f7644}, [kernel_params+688]; ld.const.v4.f32 {%f7649, %f7650, %f7651, %f7652}, [kernel_params+704]; ld.const.v4.f32 {%f7657, %f7658, %f7659, %f7660}, [kernel_params+720]; mul.ftz.f32 %f7665, %f15960, %f7642; fma.rn.ftz.f32 %f7666, %f15959, %f7641, %f7665; fma.rn.ftz.f32 %f7667, %f15961, %f7643, %f7666; add.ftz.f32 %f7668, %f7644, %f7667; mul.ftz.f32 %f7669, %f15960, %f7650; fma.rn.ftz.f32 %f7670, %f15959, %f7649, %f7669; fma.rn.ftz.f32 %f7671, %f15961, %f7651, %f7670; add.ftz.f32 %f7672, %f7652, %f7671; mul.ftz.f32 %f7673, %f15960, %f7658; fma.rn.ftz.f32 %f7674, %f15959, %f7657, %f7673; fma.rn.ftz.f32 %f7675, %f15961, %f7659, %f7674; add.ftz.f32 %f7676, %f7660, %f7675; mul.ftz.f32 %f7677, %f7672, %f7672; fma.rn.ftz.f32 %f7678, %f7668, %f7668, %f7677; fma.rn.ftz.f32 %f7679, %f7676, %f7676, %f7678; sqrt.approx.ftz.f32 %f7680, %f7679; div.approx.ftz.f32 %f1417, %f7668, %f7680; div.approx.ftz.f32 %f1418, %f7672, %f7680; div.approx.ftz.f32 %f1419, %f7676, %f7680; ld.const.u32 %r493, [kernel_params+20]; mov.f32 %f7681, 0fBF800000; max.ftz.f32 %f7682, %f1417, %f7681; mov.f32 %f7683, 0f3F800000; min.ftz.f32 %f7684, %f7682, %f7683; abs.ftz.f32 %f7685, %f7684; sub.ftz.f32 %f7686, %f7683, %f7685; mul.ftz.f32 %f7687, %f7686, 0f3F000000; sqrt.approx.ftz.f32 %f7688, %f7687; setp.gt.ftz.f32 %p840, %f7685, 0f3F11EB85; selp.f32 %f7689, %f7688, %f7685, %p840; mul.ftz.f32 %f7690, %f7689, %f7689; mov.f32 %f7691, 0f3C94D2E9; mov.f32 %f7692, 0f3D53F941; fma.rn.ftz.f32 %f7693, %f7692, %f7690, %f7691; mov.f32 %f7694, 0f3D3F841F; fma.rn.ftz.f32 %f7695, %f7693, %f7690, %f7694; mov.f32 %f7696, 0f3D994929; fma.rn.ftz.f32 %f7697, %f7695, %f7690, %f7696; mov.f32 %f7698, 0f3E2AAB94; fma.rn.ftz.f32 %f7699, %f7697, %f7690, %f7698; mul.ftz.f32 %f7700, %f7690, %f7699; fma.rn.ftz.f32 %f7701, %f7700, %f7689, %f7689; add.ftz.f32 %f7702, %f7701, %f7701; mov.f32 %f7703, 0f3FC90FDB; sub.ftz.f32 %f7704, %f7703, %f7701; selp.f32 %f7705, %f7702, %f7704, %p840; setp.lt.ftz.f32 %p841, %f7684, 0f00000000; mov.f32 %f7706, 0f40490FDB; sub.ftz.f32 %f7707, %f7706, %f7705; selp.f32 %f1420, %f7707, %f7705, %p841; abs.ftz.f32 %f1421, %f1418; abs.ftz.f32 %f1422, %f1419; setp.eq.ftz.f32 %p842, %f1421, 0f00000000; setp.eq.ftz.f32 %p843, %f1422, 0f00000000; and.pred %p2, %p842, %p843; mov.b32 %r494, %f1418; mov.b32 %r1641, %f1419; and.b32 %r495, %r1641, -2147483648; setp.gt.s32 %p844, %r493, 2; @%p844 bra $L__BB106_999; bra.uni $L__BB106_994; $L__BB106_999: setp.eq.s32 %p845, %r493, 3; @%p845 bra $L__BB106_1008; setp.eq.s32 %p846, %r493, 5; @%p846 bra $L__BB106_1011; setp.ne.s32 %p847, %r493, 4; @%p847 bra $L__BB106_1037; ld.const.f32 %f7708, [kernel_params+48]; ld.const.v4.f32 {%f7709, %f7710, %f7711, %f7712}, [kernel_params+64]; neg.ftz.f32 %f7717, %f1420; sub.ftz.f32 %f7718, %f7717, %f7708; div.approx.ftz.f32 %f7719, %f7718, %f7709; mul.ftz.f32 %f7720, %f7719, %f7719; mul.ftz.f32 %f7721, %f7719, %f7720; mul.ftz.f32 %f7722, %f7720, %f7720; mul.ftz.f32 %f7723, %f7710, %f7720; fma.rn.ftz.f32 %f7724, %f7709, %f7719, %f7723; mul.ftz.f32 %f7725, %f7712, %f7722; fma.rn.ftz.f32 %f7726, %f7711, %f7721, %f7725; add.ftz.f32 %f7727, %f7724, %f7726; add.ftz.f32 %f7728, %f7708, %f7727; sub.ftz.f32 %f7729, %f7717, %f7728; div.approx.ftz.f32 %f7730, %f7729, %f7709; mul.ftz.f32 %f7731, %f7730, %f7730; mul.ftz.f32 %f7732, %f7730, %f7731; mul.ftz.f32 %f7733, %f7731, %f7731; mul.ftz.f32 %f7734, %f7710, %f7731; fma.rn.ftz.f32 %f7735, %f7709, %f7730, %f7734; mul.ftz.f32 %f7736, %f7712, %f7733; fma.rn.ftz.f32 %f7737, %f7711, %f7732, %f7736; add.ftz.f32 %f7738, %f7735, %f7737; add.ftz.f32 %f7739, %f7708, %f7738; sub.ftz.f32 %f7740, %f7717, %f7739; div.approx.ftz.f32 %f7741, %f7740, %f7709; mul.ftz.f32 %f7742, %f7741, %f7741; mul.ftz.f32 %f7743, %f7741, %f7742; mul.ftz.f32 %f7744, %f7742, %f7742; mul.ftz.f32 %f7745, %f7710, %f7742; fma.rn.ftz.f32 %f7746, %f7709, %f7741, %f7745; mul.ftz.f32 %f7747, %f7712, %f7744; fma.rn.ftz.f32 %f7748, %f7711, %f7743, %f7747; add.ftz.f32 %f7749, %f7746, %f7748; add.ftz.f32 %f7750, %f7708, %f7749; sub.ftz.f32 %f7751, %f7717, %f7750; div.approx.ftz.f32 %f7752, %f7751, %f7709; mul.ftz.f32 %f7753, %f7752, %f7752; mul.ftz.f32 %f7754, %f7752, %f7753; mul.ftz.f32 %f7755, %f7753, %f7753; mul.ftz.f32 %f7756, %f7710, %f7753; fma.rn.ftz.f32 %f7757, %f7709, %f7752, %f7756; mul.ftz.f32 %f7758, %f7712, %f7755; fma.rn.ftz.f32 %f7759, %f7711, %f7754, %f7758; add.ftz.f32 %f7760, %f7757, %f7759; add.ftz.f32 %f7761, %f7708, %f7760; sub.ftz.f32 %f7762, %f7717, %f7761; div.approx.ftz.f32 %f7763, %f7762, %f7709; mul.ftz.f32 %f7764, %f7763, %f7763; mul.ftz.f32 %f7765, %f7763, %f7764; mul.ftz.f32 %f7766, %f7764, %f7764; mul.ftz.f32 %f7767, %f7710, %f7764; fma.rn.ftz.f32 %f7768, %f7709, %f7763, %f7767; mul.ftz.f32 %f7769, %f7712, %f7766; fma.rn.ftz.f32 %f7770, %f7711, %f7765, %f7769; add.ftz.f32 %f7771, %f7768, %f7770; add.ftz.f32 %f7772, %f7708, %f7771; sub.ftz.f32 %f7773, %f7717, %f7772; div.approx.ftz.f32 %f7774, %f7773, %f7709; mul.ftz.f32 %f7775, %f7774, %f7774; mul.ftz.f32 %f7776, %f7774, %f7775; mul.ftz.f32 %f7777, %f7775, %f7775; mul.ftz.f32 %f7778, %f7710, %f7775; fma.rn.ftz.f32 %f7779, %f7709, %f7774, %f7778; mul.ftz.f32 %f7780, %f7712, %f7777; fma.rn.ftz.f32 %f7781, %f7711, %f7776, %f7780; add.ftz.f32 %f7782, %f7779, %f7781; add.ftz.f32 %f7783, %f7708, %f7782; sub.ftz.f32 %f7784, %f7717, %f7783; div.approx.ftz.f32 %f7785, %f7784, %f7709; mul.ftz.f32 %f7786, %f7785, %f7785; mul.ftz.f32 %f7787, %f7785, %f7786; mul.ftz.f32 %f7788, %f7786, %f7786; mul.ftz.f32 %f7789, %f7710, %f7786; fma.rn.ftz.f32 %f7790, %f7709, %f7785, %f7789; mul.ftz.f32 %f7791, %f7712, %f7788; fma.rn.ftz.f32 %f7792, %f7711, %f7787, %f7791; add.ftz.f32 %f7793, %f7790, %f7792; add.ftz.f32 %f7794, %f7708, %f7793; sub.ftz.f32 %f7795, %f7717, %f7794; div.approx.ftz.f32 %f7796, %f7795, %f7709; mul.ftz.f32 %f7797, %f7796, %f7796; mul.ftz.f32 %f7798, %f7796, %f7797; mul.ftz.f32 %f7799, %f7797, %f7797; mul.ftz.f32 %f7800, %f7710, %f7797; fma.rn.ftz.f32 %f7801, %f7709, %f7796, %f7800; mul.ftz.f32 %f7802, %f7712, %f7799; fma.rn.ftz.f32 %f7803, %f7711, %f7798, %f7802; add.ftz.f32 %f7804, %f7801, %f7803; add.ftz.f32 %f7805, %f7708, %f7804; sub.ftz.f32 %f7806, %f7717, %f7805; div.approx.ftz.f32 %f7807, %f7806, %f7709; mul.ftz.f32 %f7808, %f7807, %f7807; mul.ftz.f32 %f7809, %f7807, %f7808; mul.ftz.f32 %f7810, %f7808, %f7808; mul.ftz.f32 %f7811, %f7710, %f7808; fma.rn.ftz.f32 %f7812, %f7709, %f7807, %f7811; mul.ftz.f32 %f7813, %f7712, %f7810; fma.rn.ftz.f32 %f7814, %f7711, %f7809, %f7813; add.ftz.f32 %f7815, %f7812, %f7814; add.ftz.f32 %f7816, %f7708, %f7815; sub.ftz.f32 %f7817, %f7717, %f7816; div.approx.ftz.f32 %f7818, %f7817, %f7709; mul.ftz.f32 %f7819, %f7818, %f7818; mul.ftz.f32 %f7820, %f7818, %f7819; mul.ftz.f32 %f7821, %f7819, %f7819; mul.ftz.f32 %f7822, %f7710, %f7819; fma.rn.ftz.f32 %f7823, %f7709, %f7818, %f7822; mul.ftz.f32 %f7824, %f7712, %f7821; fma.rn.ftz.f32 %f7825, %f7711, %f7820, %f7824; add.ftz.f32 %f7826, %f7823, %f7825; add.ftz.f32 %f7827, %f7708, %f7826; sub.ftz.f32 %f7828, %f7717, %f7827; div.approx.ftz.f32 %f7829, %f7828, %f7709; mul.ftz.f32 %f7830, %f7829, %f7829; mul.ftz.f32 %f7831, %f7829, %f7830; mul.ftz.f32 %f7832, %f7830, %f7830; mul.ftz.f32 %f7833, %f7710, %f7830; fma.rn.ftz.f32 %f7834, %f7709, %f7829, %f7833; mul.ftz.f32 %f7835, %f7712, %f7832; fma.rn.ftz.f32 %f7836, %f7711, %f7831, %f7835; add.ftz.f32 %f7837, %f7834, %f7836; add.ftz.f32 %f7838, %f7708, %f7837; sub.ftz.f32 %f7839, %f7717, %f7838; div.approx.ftz.f32 %f7840, %f7839, %f7709; mul.ftz.f32 %f7841, %f7840, %f7840; mul.ftz.f32 %f7842, %f7840, %f7841; mul.ftz.f32 %f7843, %f7841, %f7841; mul.ftz.f32 %f7844, %f7710, %f7841; fma.rn.ftz.f32 %f7845, %f7709, %f7840, %f7844; mul.ftz.f32 %f7846, %f7712, %f7843; fma.rn.ftz.f32 %f7847, %f7711, %f7842, %f7846; add.ftz.f32 %f7848, %f7845, %f7847; add.ftz.f32 %f7849, %f7708, %f7848; sub.ftz.f32 %f7850, %f7717, %f7849; div.approx.ftz.f32 %f7851, %f7850, %f7709; mul.ftz.f32 %f7852, %f7851, %f7851; mul.ftz.f32 %f7853, %f7851, %f7852; mul.ftz.f32 %f7854, %f7852, %f7852; mul.ftz.f32 %f7855, %f7710, %f7852; fma.rn.ftz.f32 %f7856, %f7709, %f7851, %f7855; mul.ftz.f32 %f7857, %f7712, %f7854; fma.rn.ftz.f32 %f7858, %f7711, %f7853, %f7857; add.ftz.f32 %f7859, %f7856, %f7858; add.ftz.f32 %f7860, %f7708, %f7859; sub.ftz.f32 %f7861, %f7717, %f7860; div.approx.ftz.f32 %f7862, %f7861, %f7709; mul.ftz.f32 %f7863, %f7862, %f7862; mul.ftz.f32 %f7864, %f7862, %f7863; mul.ftz.f32 %f7865, %f7863, %f7863; mul.ftz.f32 %f7866, %f7710, %f7863; fma.rn.ftz.f32 %f7867, %f7709, %f7862, %f7866; mul.ftz.f32 %f7868, %f7712, %f7865; fma.rn.ftz.f32 %f7869, %f7711, %f7864, %f7868; add.ftz.f32 %f7870, %f7867, %f7869; add.ftz.f32 %f7871, %f7708, %f7870; sub.ftz.f32 %f7872, %f7717, %f7871; div.approx.ftz.f32 %f7873, %f7872, %f7709; mul.ftz.f32 %f7874, %f7873, %f7873; mul.ftz.f32 %f7875, %f7873, %f7874; mul.ftz.f32 %f7876, %f7874, %f7874; mul.ftz.f32 %f7877, %f7710, %f7874; fma.rn.ftz.f32 %f7878, %f7709, %f7873, %f7877; mul.ftz.f32 %f7879, %f7712, %f7876; fma.rn.ftz.f32 %f7880, %f7711, %f7875, %f7879; add.ftz.f32 %f7881, %f7878, %f7880; add.ftz.f32 %f7882, %f7708, %f7881; sub.ftz.f32 %f7883, %f7717, %f7882; div.approx.ftz.f32 %f7884, %f7883, %f7709; mul.ftz.f32 %f7885, %f7884, %f7884; mul.ftz.f32 %f7886, %f7884, %f7885; mul.ftz.f32 %f7887, %f7885, %f7885; mul.ftz.f32 %f7888, %f7710, %f7885; fma.rn.ftz.f32 %f7889, %f7709, %f7884, %f7888; mul.ftz.f32 %f7890, %f7712, %f7887; fma.rn.ftz.f32 %f7891, %f7711, %f7886, %f7890; add.ftz.f32 %f7892, %f7889, %f7891; add.ftz.f32 %f7893, %f7708, %f7892; sub.ftz.f32 %f7894, %f7717, %f7893; div.approx.ftz.f32 %f7895, %f7894, %f7709; mul.ftz.f32 %f7896, %f7895, %f7895; mul.ftz.f32 %f7897, %f7895, %f7896; mul.ftz.f32 %f7898, %f7896, %f7896; mul.ftz.f32 %f7899, %f7710, %f7896; fma.rn.ftz.f32 %f7900, %f7709, %f7895, %f7899; mul.ftz.f32 %f7901, %f7712, %f7898; fma.rn.ftz.f32 %f7902, %f7711, %f7897, %f7901; add.ftz.f32 %f7903, %f7900, %f7902; add.ftz.f32 %f7904, %f7708, %f7903; sub.ftz.f32 %f7905, %f7717, %f7904; div.approx.ftz.f32 %f7906, %f7905, %f7709; mul.ftz.f32 %f7907, %f7906, %f7906; mul.ftz.f32 %f7908, %f7906, %f7907; mul.ftz.f32 %f7909, %f7907, %f7907; mul.ftz.f32 %f7910, %f7710, %f7907; fma.rn.ftz.f32 %f7911, %f7709, %f7906, %f7910; mul.ftz.f32 %f7912, %f7712, %f7909; fma.rn.ftz.f32 %f7913, %f7711, %f7908, %f7912; add.ftz.f32 %f7914, %f7911, %f7913; add.ftz.f32 %f7915, %f7708, %f7914; sub.ftz.f32 %f7916, %f7717, %f7915; div.approx.ftz.f32 %f7917, %f7916, %f7709; mul.ftz.f32 %f7918, %f7917, %f7917; mul.ftz.f32 %f7919, %f7917, %f7918; mul.ftz.f32 %f7920, %f7918, %f7918; mul.ftz.f32 %f7921, %f7710, %f7918; fma.rn.ftz.f32 %f7922, %f7709, %f7917, %f7921; mul.ftz.f32 %f7923, %f7712, %f7920; fma.rn.ftz.f32 %f7924, %f7711, %f7919, %f7923; add.ftz.f32 %f7925, %f7922, %f7924; add.ftz.f32 %f7926, %f7708, %f7925; sub.ftz.f32 %f7927, %f7717, %f7926; div.approx.ftz.f32 %f7928, %f7927, %f7709; mul.ftz.f32 %f7929, %f7928, %f7928; mul.ftz.f32 %f7930, %f7928, %f7929; mul.ftz.f32 %f7931, %f7929, %f7929; mul.ftz.f32 %f7932, %f7710, %f7929; fma.rn.ftz.f32 %f7933, %f7709, %f7928, %f7932; mul.ftz.f32 %f7934, %f7712, %f7931; fma.rn.ftz.f32 %f7935, %f7711, %f7930, %f7934; add.ftz.f32 %f7936, %f7933, %f7935; add.ftz.f32 %f7937, %f7708, %f7936; sub.ftz.f32 %f7938, %f7717, %f7937; div.approx.ftz.f32 %f1425, %f7938, %f7709; @%p2 bra $L__BB106_1006; bra.uni $L__BB106_1003; $L__BB106_1006: shr.s32 %r1646, %r494, 31; and.b32 %r1647, %r1646, 1078530011; or.b32 %r1648, %r1647, %r495; mov.b32 %f15989, %r1648; bra.uni $L__BB106_1007; $L__BB106_987: ld.const.v4.f32 {%f7503, %f7504, %f7505, %f7506}, [kernel_params+560]; ld.const.v4.f32 {%f7511, %f7512, %f7513, %f7514}, [kernel_params+576]; ld.const.v4.f32 {%f7527, %f7528, %f7529, %f7530}, [kernel_params+608]; mul.ftz.f32 %f7535, %f15960, %f7504; fma.rn.ftz.f32 %f7536, %f15959, %f7503, %f7535; fma.rn.ftz.f32 %f7537, %f15961, %f7505, %f7506; add.ftz.f32 %f1381, %f7536, %f7537; mul.ftz.f32 %f7538, %f15960, %f7512; fma.rn.ftz.f32 %f7539, %f15959, %f7511, %f7538; fma.rn.ftz.f32 %f7540, %f15961, %f7513, %f7514; add.ftz.f32 %f1382, %f7539, %f7540; mul.ftz.f32 %f7544, %f15960, %f7528; fma.rn.ftz.f32 %f7545, %f15959, %f7527, %f7544; fma.rn.ftz.f32 %f7546, %f15961, %f7529, %f7530; add.ftz.f32 %f1384, %f7545, %f7546; setp.eq.ftz.f32 %p837, %f1384, 0f00000000; mov.f32 %f16024, 0f00000000; mov.f32 %f16022, %f16024; mov.f32 %f16023, %f16024; @%p837 bra $L__BB106_989; div.approx.ftz.f32 %f16023, %f1382, %f1384; div.approx.ftz.f32 %f16022, %f1381, %f1384; $L__BB106_989: ld.const.v4.f32 {%f7550, %f7551, %f7552, %f7553}, [kernel_params+736]; ld.const.v4.f32 {%f7558, %f7559, %f7560, %f7561}, [kernel_params+752]; ld.const.v4.f32 {%f7574, %f7575, %f7576, %f7577}, [kernel_params+784]; mul.ftz.f32 %f7582, %f1376, %f7551; fma.rn.ftz.f32 %f7583, %f1375, %f7550, %f7582; fma.rn.ftz.f32 %f7584, %f1377, %f7552, %f7553; add.ftz.f32 %f1393, %f7583, %f7584; mul.ftz.f32 %f7585, %f1376, %f7559; fma.rn.ftz.f32 %f7586, %f1375, %f7558, %f7585; fma.rn.ftz.f32 %f7587, %f1377, %f7560, %f7561; add.ftz.f32 %f1394, %f7586, %f7587; mul.ftz.f32 %f7591, %f1376, %f7575; fma.rn.ftz.f32 %f7592, %f1375, %f7574, %f7591; fma.rn.ftz.f32 %f7593, %f1377, %f7576, %f7577; add.ftz.f32 %f1396, %f7592, %f7593; setp.eq.ftz.f32 %p838, %f1396, 0f00000000; mov.f32 %f16026, %f16024; mov.f32 %f16027, %f16024; @%p838 bra $L__BB106_991; div.approx.ftz.f32 %f16027, %f1394, %f1396; div.approx.ftz.f32 %f16026, %f1393, %f1396; $L__BB106_991: ld.const.v4.f32 {%f7597, %f7598, %f7599, %f7600}, [kernel_params+800]; ld.const.v4.f32 {%f7605, %f7606, %f7607, %f7608}, [kernel_params+816]; ld.const.v4.f32 {%f7621, %f7622, %f7623, %f7624}, [kernel_params+848]; mul.ftz.f32 %f7629, %f1379, %f7598; fma.rn.ftz.f32 %f7630, %f1378, %f7597, %f7629; fma.rn.ftz.f32 %f7631, %f1380, %f7599, %f7600; add.ftz.f32 %f1405, %f7630, %f7631; mul.ftz.f32 %f7632, %f1379, %f7606; fma.rn.ftz.f32 %f7633, %f1378, %f7605, %f7632; fma.rn.ftz.f32 %f7634, %f1380, %f7607, %f7608; add.ftz.f32 %f1406, %f7633, %f7634; mul.ftz.f32 %f7638, %f1379, %f7622; fma.rn.ftz.f32 %f7639, %f1378, %f7621, %f7638; fma.rn.ftz.f32 %f7640, %f1380, %f7623, %f7624; add.ftz.f32 %f1408, %f7639, %f7640; setp.eq.ftz.f32 %p839, %f1408, 0f00000000; mov.f32 %f16025, %f16024; @%p839 bra $L__BB106_1144; div.approx.ftz.f32 %f16025, %f1406, %f1408; div.approx.ftz.f32 %f16024, %f1405, %f1408; bra.uni $L__BB106_1144; $L__BB106_994: setp.eq.s32 %p848, %r493, 0; @%p848 bra $L__BB106_1030; setp.eq.s32 %p849, %r493, 1; @%p849 bra $L__BB106_996; bra.uni $L__BB106_1037; $L__BB106_996: ld.const.f32 %f1450, [kernel_params+24]; mul.ftz.f32 %f8038, %f1419, %f1419; fma.rn.ftz.f32 %f8039, %f1418, %f1418, %f8038; sqrt.approx.ftz.f32 %f8040, %f8039; abs.ftz.f32 %f1451, %f8040; abs.ftz.f32 %f1452, %f1417; setp.eq.ftz.f32 %p878, %f1452, 0f00000000; setp.eq.ftz.f32 %p879, %f1451, 0f00000000; and.pred %p880, %p878, %p879; mov.b32 %r498, %f1417; mov.b32 %r1663, %f8040; and.b32 %r499, %r1663, -2147483648; @%p880 bra $L__BB106_1023; bra.uni $L__BB106_997; $L__BB106_1023: shr.s32 %r1668, %r498, 31; and.b32 %r1669, %r1668, 1078530011; or.b32 %r1670, %r499, %r1669; mov.b32 %f15994, %r1670; bra.uni $L__BB106_1024; $L__BB106_1008: add.ftz.f32 %f7971, %f1418, 0fBF800000; mul.ftz.f32 %f7972, %f7971, 0fBF000000; mov.f32 %f7973, 0f00000000; max.ftz.f32 %f7974, %f7972, %f7973; sqrt.approx.ftz.f32 %f7975, %f7974; add.ftz.f32 %f1432, %f7975, %f7975; setp.leu.ftz.f32 %p857, %f1432, 0f00000000; @%p857 bra $L__BB106_1010; rcp.approx.ftz.f32 %f7976, %f1432; mul.ftz.f32 %f1417, %f1417, %f7976; mul.ftz.f32 %f1419, %f1419, %f7976; $L__BB106_1010: add.ftz.f32 %f7977, %f1417, 0f3F800000; add.ftz.f32 %f7978, %f1419, 0f3F800000; mul.ftz.f32 %f15999, %f7978, 0f3F000000; mul.ftz.f32 %f15998, %f7977, 0f3F000000; bra.uni $L__BB106_1043; $L__BB106_1011: abs.ftz.f32 %f1439, %f1417; setp.eq.ftz.f32 %p859, %f1439, 0f00000000; and.pred %p860, %p859, %p842; mov.b32 %r496, %f1417; and.b32 %r497, %r494, -2147483648; @%p860 bra $L__BB106_1015; bra.uni $L__BB106_1012; $L__BB106_1015: shr.s32 %r1653, %r496, 31; and.b32 %r1654, %r1653, 1078530011; or.b32 %r1655, %r1654, %r497; mov.b32 %f15992, %r1655; bra.uni $L__BB106_1016; $L__BB106_1030: setp.eq.ftz.f32 %p895, %f1418, 0f00000000; setp.eq.ftz.f32 %p896, %f1417, 0f00000000; and.pred %p897, %p896, %p895; setp.eq.ftz.f32 %p898, %f1419, 0f00000000; and.pred %p899, %p898, %p897; mov.f32 %f15998, 0f00000000; mov.f32 %f15999, %f15998; @%p899 bra $L__BB106_1043; ld.const.v4.f32 {%f8097, %f8098, %f8099, %f8100}, [kernel_params+32]; abs.ftz.f32 %f1468, %f1417; setp.eq.ftz.f32 %p901, %f1468, 0f00000000; and.pred %p902, %p901, %p842; mov.b32 %r500, %f1417; and.b32 %r501, %r494, -2147483648; @%p902 bra $L__BB106_1035; bra.uni $L__BB106_1032; $L__BB106_1035: shr.s32 %r1682, %r500, 31; and.b32 %r1683, %r1682, 1078530011; or.b32 %r1684, %r1683, %r501; mov.b32 %f15996, %r1684; bra.uni $L__BB106_1036; $L__BB106_1037: ld.const.f32 %f8159, [kernel_params+28]; ld.const.f32 %f1475, [kernel_params+276]; ld.const.f32 %f1476, [kernel_params+280]; add.ftz.f32 %f8160, %f8159, %f8159; mul.ftz.f32 %f8161, %f1420, 0f3F000000; sin.approx.ftz.f32 %f8162, %f8161; mul.ftz.f32 %f1477, %f8160, %f8162; @%p2 bra $L__BB106_1041; bra.uni $L__BB106_1038; $L__BB106_1041: shr.s32 %r1689, %r494, 31; and.b32 %r1690, %r1689, 1078530011; or.b32 %r1691, %r1690, %r495; mov.b32 %f15997, %r1691; bra.uni $L__BB106_1042; $L__BB106_1003: setp.eq.ftz.f32 %p850, %f1421, 0f7F800000; setp.eq.ftz.f32 %p851, %f1422, 0f7F800000; and.pred %p852, %p850, %p851; @%p852 bra $L__BB106_1005; bra.uni $L__BB106_1004; $L__BB106_1005: setp.lt.s32 %p856, %r494, 0; selp.b32 %r1644, 1075235812, 1061752795, %p856; or.b32 %r1645, %r1644, %r495; mov.b32 %f15989, %r1645; bra.uni $L__BB106_1007; $L__BB106_997: setp.eq.ftz.f32 %p881, %f1452, 0f7F800000; setp.eq.ftz.f32 %p882, %f1451, 0f7F800000; and.pred %p883, %p881, %p882; @%p883 bra $L__BB106_1022; bra.uni $L__BB106_998; $L__BB106_1022: setp.lt.s32 %p887, %r498, 0; selp.b32 %r1666, 1075235812, 1061752795, %p887; or.b32 %r1667, %r499, %r1666; mov.b32 %f15994, %r1667; bra.uni $L__BB106_1024; $L__BB106_1012: setp.eq.ftz.f32 %p861, %f1439, 0f7F800000; setp.eq.ftz.f32 %p862, %f1421, 0f7F800000; and.pred %p863, %p861, %p862; @%p863 bra $L__BB106_1014; bra.uni $L__BB106_1013; $L__BB106_1014: setp.lt.s32 %p867, %r496, 0; selp.b32 %r1651, 1075235812, 1061752795, %p867; or.b32 %r1652, %r1651, %r497; mov.b32 %f15992, %r1652; bra.uni $L__BB106_1016; $L__BB106_1038: setp.eq.ftz.f32 %p912, %f1421, 0f7F800000; setp.eq.ftz.f32 %p913, %f1422, 0f7F800000; and.pred %p914, %p912, %p913; @%p914 bra $L__BB106_1040; bra.uni $L__BB106_1039; $L__BB106_1040: setp.lt.s32 %p918, %r494, 0; selp.b32 %r1687, 1075235812, 1061752795, %p918; or.b32 %r1688, %r1687, %r495; mov.b32 %f15997, %r1688; bra.uni $L__BB106_1042; $L__BB106_1032: setp.eq.ftz.f32 %p903, %f1468, 0f7F800000; setp.eq.ftz.f32 %p904, %f1421, 0f7F800000; and.pred %p905, %p903, %p904; @%p905 bra $L__BB106_1034; bra.uni $L__BB106_1033; $L__BB106_1034: setp.lt.s32 %p909, %r500, 0; selp.b32 %r1680, 1075235812, 1061752795, %p909; or.b32 %r1681, %r1680, %r501; mov.b32 %f15996, %r1681; bra.uni $L__BB106_1036; $L__BB106_1004: mov.f32 %f15511, 0f3FC90FDB; setp.lt.s32 %p853, %r494, 0; min.ftz.f32 %f7939, %f1422, %f1421; max.ftz.f32 %f7940, %f1422, %f1421; div.full.ftz.f32 %f7941, %f7939, %f7940; mul.rn.ftz.f32 %f7942, %f7941, %f7941; mov.f32 %f7943, 0fC0B59883; mov.f32 %f7944, 0fBF52C7EA; fma.rn.ftz.f32 %f7945, %f7942, %f7944, %f7943; mov.f32 %f7946, 0fC0D21907; fma.rn.ftz.f32 %f7947, %f7945, %f7942, %f7946; mul.ftz.f32 %f7948, %f7942, %f7947; mul.ftz.f32 %f7949, %f7941, %f7948; add.ftz.f32 %f7950, %f7942, 0f41355DC0; mov.f32 %f7951, 0f41E6BD60; fma.rn.ftz.f32 %f7952, %f7950, %f7942, %f7951; mov.f32 %f7953, 0f419D92C8; fma.rn.ftz.f32 %f7954, %f7952, %f7942, %f7953; rcp.approx.ftz.f32 %f7955, %f7954; fma.rn.ftz.f32 %f7956, %f7949, %f7955, %f7941; sub.ftz.f32 %f7958, %f15511, %f7956; setp.gt.ftz.f32 %p854, %f1422, %f1421; selp.f32 %f7959, %f7958, %f7956, %p854; sub.ftz.f32 %f7961, %f7706, %f7959; selp.f32 %f7962, %f7961, %f7959, %p853; mov.b32 %r1642, %f7962; or.b32 %r1643, %r495, %r1642; mov.b32 %f7963, %r1643; add.ftz.f32 %f7964, %f1421, %f1422; setp.le.ftz.f32 %p855, %f7964, 0f7F800000; selp.f32 %f15989, %f7963, %f7964, %p855; $L__BB106_1007: ld.const.f32 %f15460, [kernel_params+280]; ld.const.f32 %f15459, [kernel_params+276]; cos.approx.ftz.f32 %f7965, %f15989; mul.ftz.f32 %f7966, %f1425, %f7965; div.approx.ftz.f32 %f7967, %f7966, %f15459; sin.approx.ftz.f32 %f7968, %f15989; mul.ftz.f32 %f7969, %f1425, %f7968; div.approx.ftz.f32 %f7970, %f7969, %f15460; add.ftz.f32 %f15999, %f7970, 0f3F000000; add.ftz.f32 %f15998, %f7967, 0f3F000000; bra.uni $L__BB106_1043; $L__BB106_998: setp.lt.s32 %p884, %r498, 0; min.ftz.f32 %f8041, %f1451, %f1452; max.ftz.f32 %f8042, %f1451, %f1452; div.full.ftz.f32 %f8043, %f8041, %f8042; mul.rn.ftz.f32 %f8044, %f8043, %f8043; mov.f32 %f8045, 0fC0B59883; mov.f32 %f8046, 0fBF52C7EA; fma.rn.ftz.f32 %f8047, %f8044, %f8046, %f8045; mov.f32 %f8048, 0fC0D21907; fma.rn.ftz.f32 %f8049, %f8047, %f8044, %f8048; mul.ftz.f32 %f8050, %f8044, %f8049; mul.ftz.f32 %f8051, %f8043, %f8050; add.ftz.f32 %f8052, %f8044, 0f41355DC0; mov.f32 %f8053, 0f41E6BD60; fma.rn.ftz.f32 %f8054, %f8052, %f8044, %f8053; mov.f32 %f8055, 0f419D92C8; fma.rn.ftz.f32 %f8056, %f8054, %f8044, %f8055; rcp.approx.ftz.f32 %f8057, %f8056; fma.rn.ftz.f32 %f8058, %f8051, %f8057, %f8043; sub.ftz.f32 %f8060, %f7703, %f8058; setp.gt.ftz.f32 %p885, %f1451, %f1452; selp.f32 %f8061, %f8060, %f8058, %p885; sub.ftz.f32 %f8063, %f7706, %f8061; selp.f32 %f8064, %f8063, %f8061, %p884; mov.b32 %r1664, %f8064; or.b32 %r1665, %r499, %r1664; mov.b32 %f8065, %r1665; add.ftz.f32 %f8066, %f1452, %f1451; setp.le.ftz.f32 %p886, %f8066, 0f7F800000; selp.f32 %f15994, %f8065, %f8066, %p886; $L__BB106_1024: div.approx.ftz.f32 %f1457, %f15994, %f1450; @%p2 bra $L__BB106_1028; bra.uni $L__BB106_1025; $L__BB106_1028: shr.s32 %r1675, %r494, 31; and.b32 %r1676, %r1675, 1078530011; or.b32 %r1677, %r1676, %r495; mov.b32 %f15995, %r1677; bra.uni $L__BB106_1029; $L__BB106_1025: setp.eq.ftz.f32 %p888, %f1421, 0f7F800000; setp.eq.ftz.f32 %p889, %f1422, 0f7F800000; and.pred %p890, %p888, %p889; @%p890 bra $L__BB106_1027; bra.uni $L__BB106_1026; $L__BB106_1027: setp.lt.s32 %p894, %r494, 0; selp.b32 %r1673, 1075235812, 1061752795, %p894; or.b32 %r1674, %r1673, %r495; mov.b32 %f15995, %r1674; bra.uni $L__BB106_1029; $L__BB106_1013: setp.lt.s32 %p864, %r496, 0; min.ftz.f32 %f7979, %f1421, %f1439; max.ftz.f32 %f7980, %f1421, %f1439; div.full.ftz.f32 %f7981, %f7979, %f7980; mul.rn.ftz.f32 %f7982, %f7981, %f7981; mov.f32 %f7983, 0fC0B59883; mov.f32 %f7984, 0fBF52C7EA; fma.rn.ftz.f32 %f7985, %f7982, %f7984, %f7983; mov.f32 %f7986, 0fC0D21907; fma.rn.ftz.f32 %f7987, %f7985, %f7982, %f7986; mul.ftz.f32 %f7988, %f7982, %f7987; mul.ftz.f32 %f7989, %f7981, %f7988; add.ftz.f32 %f7990, %f7982, 0f41355DC0; mov.f32 %f7991, 0f41E6BD60; fma.rn.ftz.f32 %f7992, %f7990, %f7982, %f7991; mov.f32 %f7993, 0f419D92C8; fma.rn.ftz.f32 %f7994, %f7992, %f7982, %f7993; rcp.approx.ftz.f32 %f7995, %f7994; fma.rn.ftz.f32 %f7996, %f7989, %f7995, %f7981; sub.ftz.f32 %f7998, %f7703, %f7996; setp.gt.ftz.f32 %p865, %f1421, %f1439; selp.f32 %f7999, %f7998, %f7996, %p865; sub.ftz.f32 %f8001, %f7706, %f7999; selp.f32 %f8002, %f8001, %f7999, %p864; mov.b32 %r1649, %f8002; or.b32 %r1650, %r497, %r1649; mov.b32 %f8003, %r1650; add.ftz.f32 %f8004, %f1439, %f1421; setp.le.ftz.f32 %p866, %f8004, 0f7F800000; selp.f32 %f15992, %f8003, %f8004, %p866; $L__BB106_1016: and.pred %p870, %p859, %p843; @%p870 bra $L__BB106_1020; bra.uni $L__BB106_1017; $L__BB106_1020: shr.s32 %r1660, %r496, 31; and.b32 %r1661, %r1660, 1078530011; or.b32 %r1662, %r495, %r1661; mov.b32 %f15993, %r1662; bra.uni $L__BB106_1021; $L__BB106_1017: setp.eq.ftz.f32 %p871, %f1439, 0f7F800000; setp.eq.ftz.f32 %p872, %f1422, 0f7F800000; and.pred %p873, %p871, %p872; @%p873 bra $L__BB106_1019; bra.uni $L__BB106_1018; $L__BB106_1019: setp.lt.s32 %p877, %r496, 0; selp.b32 %r1658, 1075235812, 1061752795, %p877; or.b32 %r1659, %r495, %r1658; mov.b32 %f15993, %r1659; bra.uni $L__BB106_1021; $L__BB106_1039: setp.lt.s32 %p915, %r494, 0; min.ftz.f32 %f8163, %f1422, %f1421; max.ftz.f32 %f8164, %f1422, %f1421; div.full.ftz.f32 %f8165, %f8163, %f8164; mul.rn.ftz.f32 %f8166, %f8165, %f8165; mov.f32 %f8167, 0fC0B59883; mov.f32 %f8168, 0fBF52C7EA; fma.rn.ftz.f32 %f8169, %f8166, %f8168, %f8167; mov.f32 %f8170, 0fC0D21907; fma.rn.ftz.f32 %f8171, %f8169, %f8166, %f8170; mul.ftz.f32 %f8172, %f8166, %f8171; mul.ftz.f32 %f8173, %f8165, %f8172; add.ftz.f32 %f8174, %f8166, 0f41355DC0; mov.f32 %f8175, 0f41E6BD60; fma.rn.ftz.f32 %f8176, %f8174, %f8166, %f8175; mov.f32 %f8177, 0f419D92C8; fma.rn.ftz.f32 %f8178, %f8176, %f8166, %f8177; rcp.approx.ftz.f32 %f8179, %f8178; fma.rn.ftz.f32 %f8180, %f8173, %f8179, %f8165; sub.ftz.f32 %f8182, %f7703, %f8180; setp.gt.ftz.f32 %p916, %f1422, %f1421; selp.f32 %f8183, %f8182, %f8180, %p916; sub.ftz.f32 %f8185, %f7706, %f8183; selp.f32 %f8186, %f8185, %f8183, %p915; mov.b32 %r1685, %f8186; or.b32 %r1686, %r495, %r1685; mov.b32 %f8187, %r1686; add.ftz.f32 %f8188, %f1421, %f1422; setp.le.ftz.f32 %p917, %f8188, 0f7F800000; selp.f32 %f15997, %f8187, %f8188, %p917; $L__BB106_1042: cos.approx.ftz.f32 %f8189, %f15997; mul.ftz.f32 %f8190, %f1477, %f8189; div.approx.ftz.f32 %f8191, %f8190, %f1475; sin.approx.ftz.f32 %f8192, %f15997; mul.ftz.f32 %f8193, %f1477, %f8192; div.approx.ftz.f32 %f8194, %f8193, %f1476; add.ftz.f32 %f15999, %f8194, 0f3F000000; add.ftz.f32 %f15998, %f8191, 0f3F000000; bra.uni $L__BB106_1043; $L__BB106_1026: setp.lt.s32 %p891, %r494, 0; min.ftz.f32 %f8067, %f1422, %f1421; max.ftz.f32 %f8068, %f1422, %f1421; div.full.ftz.f32 %f8069, %f8067, %f8068; mul.rn.ftz.f32 %f8070, %f8069, %f8069; mov.f32 %f8071, 0fC0B59883; mov.f32 %f8072, 0fBF52C7EA; fma.rn.ftz.f32 %f8073, %f8070, %f8072, %f8071; mov.f32 %f8074, 0fC0D21907; fma.rn.ftz.f32 %f8075, %f8073, %f8070, %f8074; mul.ftz.f32 %f8076, %f8070, %f8075; mul.ftz.f32 %f8077, %f8069, %f8076; add.ftz.f32 %f8078, %f8070, 0f41355DC0; mov.f32 %f8079, 0f41E6BD60; fma.rn.ftz.f32 %f8080, %f8078, %f8070, %f8079; mov.f32 %f8081, 0f419D92C8; fma.rn.ftz.f32 %f8082, %f8080, %f8070, %f8081; rcp.approx.ftz.f32 %f8083, %f8082; fma.rn.ftz.f32 %f8084, %f8077, %f8083, %f8069; sub.ftz.f32 %f8086, %f7703, %f8084; setp.gt.ftz.f32 %p892, %f1422, %f1421; selp.f32 %f8087, %f8086, %f8084, %p892; sub.ftz.f32 %f8089, %f7706, %f8087; selp.f32 %f8090, %f8089, %f8087, %p891; mov.b32 %r1671, %f8090; or.b32 %r1672, %r495, %r1671; mov.b32 %f8091, %r1672; add.ftz.f32 %f8092, %f1421, %f1422; setp.le.ftz.f32 %p893, %f8092, 0f7F800000; selp.f32 %f15995, %f8091, %f8092, %p893; $L__BB106_1029: cos.approx.ftz.f32 %f8093, %f15995; sin.approx.ftz.f32 %f8094, %f15995; fma.rn.ftz.f32 %f15999, %f1457, %f8094, 0f3F000000; fma.rn.ftz.f32 %f15998, %f1457, %f8093, 0f3F000000; bra.uni $L__BB106_1043; $L__BB106_1018: setp.lt.s32 %p874, %r496, 0; min.ftz.f32 %f8005, %f1422, %f1439; max.ftz.f32 %f8006, %f1422, %f1439; div.full.ftz.f32 %f8007, %f8005, %f8006; mul.rn.ftz.f32 %f8008, %f8007, %f8007; mov.f32 %f8009, 0fC0B59883; mov.f32 %f8010, 0fBF52C7EA; fma.rn.ftz.f32 %f8011, %f8008, %f8010, %f8009; mov.f32 %f8012, 0fC0D21907; fma.rn.ftz.f32 %f8013, %f8011, %f8008, %f8012; mul.ftz.f32 %f8014, %f8008, %f8013; mul.ftz.f32 %f8015, %f8007, %f8014; add.ftz.f32 %f8016, %f8008, 0f41355DC0; mov.f32 %f8017, 0f41E6BD60; fma.rn.ftz.f32 %f8018, %f8016, %f8008, %f8017; mov.f32 %f8019, 0f419D92C8; fma.rn.ftz.f32 %f8020, %f8018, %f8008, %f8019; rcp.approx.ftz.f32 %f8021, %f8020; fma.rn.ftz.f32 %f8022, %f8015, %f8021, %f8007; sub.ftz.f32 %f8024, %f7703, %f8022; setp.gt.ftz.f32 %p875, %f1422, %f1439; selp.f32 %f8025, %f8024, %f8022, %p875; sub.ftz.f32 %f8027, %f7706, %f8025; selp.f32 %f8028, %f8027, %f8025, %p874; mov.b32 %r1656, %f8028; or.b32 %r1657, %r495, %r1656; mov.b32 %f8029, %r1657; add.ftz.f32 %f8030, %f1439, %f1422; setp.le.ftz.f32 %p876, %f8030, 0f7F800000; selp.f32 %f15993, %f8029, %f8030, %p876; $L__BB106_1021: add.ftz.f32 %f8031, %f15993, %f15993; div.approx.ftz.f32 %f8033, %f8031, %f7706; add.ftz.f32 %f8034, %f15992, %f15992; div.approx.ftz.f32 %f8035, %f8034, %f7706; add.ftz.f32 %f8036, %f8035, 0f3F000000; sub.ftz.f32 %f15998, %f7683, %f8036; add.ftz.f32 %f15999, %f8033, 0f3F000000; bra.uni $L__BB106_1043; $L__BB106_1033: setp.lt.s32 %p906, %r500, 0; min.ftz.f32 %f8101, %f1421, %f1468; max.ftz.f32 %f8102, %f1421, %f1468; div.full.ftz.f32 %f8103, %f8101, %f8102; mul.rn.ftz.f32 %f8104, %f8103, %f8103; mov.f32 %f8105, 0fC0B59883; mov.f32 %f8106, 0fBF52C7EA; fma.rn.ftz.f32 %f8107, %f8104, %f8106, %f8105; mov.f32 %f8108, 0fC0D21907; fma.rn.ftz.f32 %f8109, %f8107, %f8104, %f8108; mul.ftz.f32 %f8110, %f8104, %f8109; mul.ftz.f32 %f8111, %f8103, %f8110; add.ftz.f32 %f8112, %f8104, 0f41355DC0; mov.f32 %f8113, 0f41E6BD60; fma.rn.ftz.f32 %f8114, %f8112, %f8104, %f8113; mov.f32 %f8115, 0f419D92C8; fma.rn.ftz.f32 %f8116, %f8114, %f8104, %f8115; rcp.approx.ftz.f32 %f8117, %f8116; fma.rn.ftz.f32 %f8118, %f8111, %f8117, %f8103; sub.ftz.f32 %f8120, %f7703, %f8118; setp.gt.ftz.f32 %p907, %f1421, %f1468; selp.f32 %f8121, %f8120, %f8118, %p907; sub.ftz.f32 %f8123, %f7706, %f8121; selp.f32 %f8124, %f8123, %f8121, %p906; mov.b32 %r1678, %f8124; or.b32 %r1679, %r501, %r1678; mov.b32 %f8125, %r1679; add.ftz.f32 %f8126, %f1468, %f1421; setp.le.ftz.f32 %p908, %f8126, 0f7F800000; selp.f32 %f15996, %f8125, %f8126, %p908; $L__BB106_1036: sub.ftz.f32 %f8127, %f15996, %f8098; mul.ftz.f32 %f8128, %f1418, %f1418; fma.rn.ftz.f32 %f8129, %f1417, %f1417, %f8128; fma.rn.ftz.f32 %f8130, %f1419, %f1419, %f8129; sqrt.approx.ftz.f32 %f8131, %f8130; div.approx.ftz.f32 %f8132, %f1419, %f8131; abs.ftz.f32 %f8133, %f8132; sub.ftz.f32 %f8135, %f7683, %f8133; mul.ftz.f32 %f8136, %f8135, 0f3F000000; sqrt.approx.ftz.f32 %f8137, %f8136; setp.gt.ftz.f32 %p910, %f8133, 0f3F11EB85; selp.f32 %f8138, %f8137, %f8133, %p910; mul.ftz.f32 %f8139, %f8138, %f8138; fma.rn.ftz.f32 %f8142, %f7692, %f8139, %f7691; fma.rn.ftz.f32 %f8144, %f8142, %f8139, %f7694; fma.rn.ftz.f32 %f8146, %f8144, %f8139, %f7696; fma.rn.ftz.f32 %f8148, %f8146, %f8139, %f7698; mul.ftz.f32 %f8149, %f8139, %f8148; fma.rn.ftz.f32 %f8150, %f8149, %f8138, %f8138; add.ftz.f32 %f8151, %f8150, %f8150; sub.ftz.f32 %f8153, %f7703, %f8150; selp.f32 %f8154, %f8151, %f8153, %p910; setp.lt.ftz.f32 %p911, %f8132, 0f00000000; sub.ftz.f32 %f8156, %f7706, %f8154; selp.f32 %f8157, %f8156, %f8154, %p911; sub.ftz.f32 %f8158, %f8157, %f8100; div.approx.ftz.f32 %f15999, %f8158, %f8099; div.approx.ftz.f32 %f15998, %f8127, %f8097; $L__BB106_1043: ld.const.u32 %r3061, [kernel_params+20]; setp.gt.s32 %p2183, %r3061, 2; mov.f32 %f15512, 0f3FC90FDB; mov.f32 %f15467, 0fBF800000; mov.f32 %f15466, 0f3E2AAB94; mov.f32 %f15465, 0f3D994929; mov.f32 %f15464, 0f3D3F841F; mov.f32 %f15463, 0f3C94D2E9; mov.f32 %f15462, 0f3D53F941; mov.f32 %f15461, 0f3F800000; ld.const.v4.f32 {%f8195, %f8196, %f8197, %f8198}, [kernel_params+864]; ld.const.v4.f32 {%f8203, %f8204, %f8205, %f8206}, [kernel_params+880]; ld.const.v4.f32 {%f8211, %f8212, %f8213, %f8214}, [kernel_params+896]; mul.ftz.f32 %f8219, %f1376, %f8196; fma.rn.ftz.f32 %f8220, %f1375, %f8195, %f8219; fma.rn.ftz.f32 %f8221, %f1377, %f8197, %f8220; add.ftz.f32 %f8222, %f8198, %f8221; mul.ftz.f32 %f8223, %f1376, %f8204; fma.rn.ftz.f32 %f8224, %f1375, %f8203, %f8223; fma.rn.ftz.f32 %f8225, %f1377, %f8205, %f8224; add.ftz.f32 %f8226, %f8206, %f8225; mul.ftz.f32 %f8227, %f1376, %f8212; fma.rn.ftz.f32 %f8228, %f1375, %f8211, %f8227; fma.rn.ftz.f32 %f8229, %f1377, %f8213, %f8228; add.ftz.f32 %f8230, %f8214, %f8229; mul.ftz.f32 %f8231, %f8226, %f8226; fma.rn.ftz.f32 %f8232, %f8222, %f8222, %f8231; fma.rn.ftz.f32 %f8233, %f8230, %f8230, %f8232; sqrt.approx.ftz.f32 %f8234, %f8233; div.approx.ftz.f32 %f1488, %f8222, %f8234; div.approx.ftz.f32 %f1489, %f8226, %f8234; div.approx.ftz.f32 %f1490, %f8230, %f8234; max.ftz.f32 %f8236, %f1488, %f15467; min.ftz.f32 %f8238, %f8236, %f15461; abs.ftz.f32 %f8239, %f8238; sub.ftz.f32 %f8240, %f15461, %f8239; mul.ftz.f32 %f8241, %f8240, 0f3F000000; sqrt.approx.ftz.f32 %f8242, %f8241; setp.gt.ftz.f32 %p919, %f8239, 0f3F11EB85; selp.f32 %f8243, %f8242, %f8239, %p919; mul.ftz.f32 %f8244, %f8243, %f8243; fma.rn.ftz.f32 %f8247, %f15462, %f8244, %f15463; fma.rn.ftz.f32 %f8249, %f8247, %f8244, %f15464; fma.rn.ftz.f32 %f8251, %f8249, %f8244, %f15465; fma.rn.ftz.f32 %f8253, %f8251, %f8244, %f15466; mul.ftz.f32 %f8254, %f8244, %f8253; fma.rn.ftz.f32 %f8255, %f8254, %f8243, %f8243; add.ftz.f32 %f8256, %f8255, %f8255; sub.ftz.f32 %f8258, %f15512, %f8255; selp.f32 %f8259, %f8256, %f8258, %p919; setp.lt.ftz.f32 %p920, %f8238, 0f00000000; sub.ftz.f32 %f8261, %f7706, %f8259; selp.f32 %f1491, %f8261, %f8259, %p920; abs.ftz.f32 %f1492, %f1489; abs.ftz.f32 %f1493, %f1490; setp.eq.ftz.f32 %p921, %f1492, 0f00000000; setp.eq.ftz.f32 %p922, %f1493, 0f00000000; and.pred %p3, %p921, %p922; mov.b32 %r502, %f1489; mov.b32 %r1692, %f1490; and.b32 %r503, %r1692, -2147483648; @%p2183 bra $L__BB106_1049; bra.uni $L__BB106_1044; $L__BB106_1049: ld.const.u32 %r3046, [kernel_params+20]; setp.eq.s32 %p924, %r3046, 3; @%p924 bra $L__BB106_1058; ld.const.u32 %r3047, [kernel_params+20]; setp.eq.s32 %p925, %r3047, 5; @%p925 bra $L__BB106_1061; ld.const.u32 %r3048, [kernel_params+20]; setp.ne.s32 %p926, %r3048, 4; @%p926 bra $L__BB106_1087; ld.const.f32 %f8262, [kernel_params+48]; ld.const.v4.f32 {%f8263, %f8264, %f8265, %f8266}, [kernel_params+64]; neg.ftz.f32 %f8271, %f1491; sub.ftz.f32 %f8272, %f8271, %f8262; div.approx.ftz.f32 %f8273, %f8272, %f8263; mul.ftz.f32 %f8274, %f8273, %f8273; mul.ftz.f32 %f8275, %f8273, %f8274; mul.ftz.f32 %f8276, %f8274, %f8274; mul.ftz.f32 %f8277, %f8264, %f8274; fma.rn.ftz.f32 %f8278, %f8263, %f8273, %f8277; mul.ftz.f32 %f8279, %f8266, %f8276; fma.rn.ftz.f32 %f8280, %f8265, %f8275, %f8279; add.ftz.f32 %f8281, %f8278, %f8280; add.ftz.f32 %f8282, %f8262, %f8281; sub.ftz.f32 %f8283, %f8271, %f8282; div.approx.ftz.f32 %f8284, %f8283, %f8263; mul.ftz.f32 %f8285, %f8284, %f8284; mul.ftz.f32 %f8286, %f8284, %f8285; mul.ftz.f32 %f8287, %f8285, %f8285; mul.ftz.f32 %f8288, %f8264, %f8285; fma.rn.ftz.f32 %f8289, %f8263, %f8284, %f8288; mul.ftz.f32 %f8290, %f8266, %f8287; fma.rn.ftz.f32 %f8291, %f8265, %f8286, %f8290; add.ftz.f32 %f8292, %f8289, %f8291; add.ftz.f32 %f8293, %f8262, %f8292; sub.ftz.f32 %f8294, %f8271, %f8293; div.approx.ftz.f32 %f8295, %f8294, %f8263; mul.ftz.f32 %f8296, %f8295, %f8295; mul.ftz.f32 %f8297, %f8295, %f8296; mul.ftz.f32 %f8298, %f8296, %f8296; mul.ftz.f32 %f8299, %f8264, %f8296; fma.rn.ftz.f32 %f8300, %f8263, %f8295, %f8299; mul.ftz.f32 %f8301, %f8266, %f8298; fma.rn.ftz.f32 %f8302, %f8265, %f8297, %f8301; add.ftz.f32 %f8303, %f8300, %f8302; add.ftz.f32 %f8304, %f8262, %f8303; sub.ftz.f32 %f8305, %f8271, %f8304; div.approx.ftz.f32 %f8306, %f8305, %f8263; mul.ftz.f32 %f8307, %f8306, %f8306; mul.ftz.f32 %f8308, %f8306, %f8307; mul.ftz.f32 %f8309, %f8307, %f8307; mul.ftz.f32 %f8310, %f8264, %f8307; fma.rn.ftz.f32 %f8311, %f8263, %f8306, %f8310; mul.ftz.f32 %f8312, %f8266, %f8309; fma.rn.ftz.f32 %f8313, %f8265, %f8308, %f8312; add.ftz.f32 %f8314, %f8311, %f8313; add.ftz.f32 %f8315, %f8262, %f8314; sub.ftz.f32 %f8316, %f8271, %f8315; div.approx.ftz.f32 %f8317, %f8316, %f8263; mul.ftz.f32 %f8318, %f8317, %f8317; mul.ftz.f32 %f8319, %f8317, %f8318; mul.ftz.f32 %f8320, %f8318, %f8318; mul.ftz.f32 %f8321, %f8264, %f8318; fma.rn.ftz.f32 %f8322, %f8263, %f8317, %f8321; mul.ftz.f32 %f8323, %f8266, %f8320; fma.rn.ftz.f32 %f8324, %f8265, %f8319, %f8323; add.ftz.f32 %f8325, %f8322, %f8324; add.ftz.f32 %f8326, %f8262, %f8325; sub.ftz.f32 %f8327, %f8271, %f8326; div.approx.ftz.f32 %f8328, %f8327, %f8263; mul.ftz.f32 %f8329, %f8328, %f8328; mul.ftz.f32 %f8330, %f8328, %f8329; mul.ftz.f32 %f8331, %f8329, %f8329; mul.ftz.f32 %f8332, %f8264, %f8329; fma.rn.ftz.f32 %f8333, %f8263, %f8328, %f8332; mul.ftz.f32 %f8334, %f8266, %f8331; fma.rn.ftz.f32 %f8335, %f8265, %f8330, %f8334; add.ftz.f32 %f8336, %f8333, %f8335; add.ftz.f32 %f8337, %f8262, %f8336; sub.ftz.f32 %f8338, %f8271, %f8337; div.approx.ftz.f32 %f8339, %f8338, %f8263; mul.ftz.f32 %f8340, %f8339, %f8339; mul.ftz.f32 %f8341, %f8339, %f8340; mul.ftz.f32 %f8342, %f8340, %f8340; mul.ftz.f32 %f8343, %f8264, %f8340; fma.rn.ftz.f32 %f8344, %f8263, %f8339, %f8343; mul.ftz.f32 %f8345, %f8266, %f8342; fma.rn.ftz.f32 %f8346, %f8265, %f8341, %f8345; add.ftz.f32 %f8347, %f8344, %f8346; add.ftz.f32 %f8348, %f8262, %f8347; sub.ftz.f32 %f8349, %f8271, %f8348; div.approx.ftz.f32 %f8350, %f8349, %f8263; mul.ftz.f32 %f8351, %f8350, %f8350; mul.ftz.f32 %f8352, %f8350, %f8351; mul.ftz.f32 %f8353, %f8351, %f8351; mul.ftz.f32 %f8354, %f8264, %f8351; fma.rn.ftz.f32 %f8355, %f8263, %f8350, %f8354; mul.ftz.f32 %f8356, %f8266, %f8353; fma.rn.ftz.f32 %f8357, %f8265, %f8352, %f8356; add.ftz.f32 %f8358, %f8355, %f8357; add.ftz.f32 %f8359, %f8262, %f8358; sub.ftz.f32 %f8360, %f8271, %f8359; div.approx.ftz.f32 %f8361, %f8360, %f8263; mul.ftz.f32 %f8362, %f8361, %f8361; mul.ftz.f32 %f8363, %f8361, %f8362; mul.ftz.f32 %f8364, %f8362, %f8362; mul.ftz.f32 %f8365, %f8264, %f8362; fma.rn.ftz.f32 %f8366, %f8263, %f8361, %f8365; mul.ftz.f32 %f8367, %f8266, %f8364; fma.rn.ftz.f32 %f8368, %f8265, %f8363, %f8367; add.ftz.f32 %f8369, %f8366, %f8368; add.ftz.f32 %f8370, %f8262, %f8369; sub.ftz.f32 %f8371, %f8271, %f8370; div.approx.ftz.f32 %f8372, %f8371, %f8263; mul.ftz.f32 %f8373, %f8372, %f8372; mul.ftz.f32 %f8374, %f8372, %f8373; mul.ftz.f32 %f8375, %f8373, %f8373; mul.ftz.f32 %f8376, %f8264, %f8373; fma.rn.ftz.f32 %f8377, %f8263, %f8372, %f8376; mul.ftz.f32 %f8378, %f8266, %f8375; fma.rn.ftz.f32 %f8379, %f8265, %f8374, %f8378; add.ftz.f32 %f8380, %f8377, %f8379; add.ftz.f32 %f8381, %f8262, %f8380; sub.ftz.f32 %f8382, %f8271, %f8381; div.approx.ftz.f32 %f8383, %f8382, %f8263; mul.ftz.f32 %f8384, %f8383, %f8383; mul.ftz.f32 %f8385, %f8383, %f8384; mul.ftz.f32 %f8386, %f8384, %f8384; mul.ftz.f32 %f8387, %f8264, %f8384; fma.rn.ftz.f32 %f8388, %f8263, %f8383, %f8387; mul.ftz.f32 %f8389, %f8266, %f8386; fma.rn.ftz.f32 %f8390, %f8265, %f8385, %f8389; add.ftz.f32 %f8391, %f8388, %f8390; add.ftz.f32 %f8392, %f8262, %f8391; sub.ftz.f32 %f8393, %f8271, %f8392; div.approx.ftz.f32 %f8394, %f8393, %f8263; mul.ftz.f32 %f8395, %f8394, %f8394; mul.ftz.f32 %f8396, %f8394, %f8395; mul.ftz.f32 %f8397, %f8395, %f8395; mul.ftz.f32 %f8398, %f8264, %f8395; fma.rn.ftz.f32 %f8399, %f8263, %f8394, %f8398; mul.ftz.f32 %f8400, %f8266, %f8397; fma.rn.ftz.f32 %f8401, %f8265, %f8396, %f8400; add.ftz.f32 %f8402, %f8399, %f8401; add.ftz.f32 %f8403, %f8262, %f8402; sub.ftz.f32 %f8404, %f8271, %f8403; div.approx.ftz.f32 %f8405, %f8404, %f8263; mul.ftz.f32 %f8406, %f8405, %f8405; mul.ftz.f32 %f8407, %f8405, %f8406; mul.ftz.f32 %f8408, %f8406, %f8406; mul.ftz.f32 %f8409, %f8264, %f8406; fma.rn.ftz.f32 %f8410, %f8263, %f8405, %f8409; mul.ftz.f32 %f8411, %f8266, %f8408; fma.rn.ftz.f32 %f8412, %f8265, %f8407, %f8411; add.ftz.f32 %f8413, %f8410, %f8412; add.ftz.f32 %f8414, %f8262, %f8413; sub.ftz.f32 %f8415, %f8271, %f8414; div.approx.ftz.f32 %f8416, %f8415, %f8263; mul.ftz.f32 %f8417, %f8416, %f8416; mul.ftz.f32 %f8418, %f8416, %f8417; mul.ftz.f32 %f8419, %f8417, %f8417; mul.ftz.f32 %f8420, %f8264, %f8417; fma.rn.ftz.f32 %f8421, %f8263, %f8416, %f8420; mul.ftz.f32 %f8422, %f8266, %f8419; fma.rn.ftz.f32 %f8423, %f8265, %f8418, %f8422; add.ftz.f32 %f8424, %f8421, %f8423; add.ftz.f32 %f8425, %f8262, %f8424; sub.ftz.f32 %f8426, %f8271, %f8425; div.approx.ftz.f32 %f8427, %f8426, %f8263; mul.ftz.f32 %f8428, %f8427, %f8427; mul.ftz.f32 %f8429, %f8427, %f8428; mul.ftz.f32 %f8430, %f8428, %f8428; mul.ftz.f32 %f8431, %f8264, %f8428; fma.rn.ftz.f32 %f8432, %f8263, %f8427, %f8431; mul.ftz.f32 %f8433, %f8266, %f8430; fma.rn.ftz.f32 %f8434, %f8265, %f8429, %f8433; add.ftz.f32 %f8435, %f8432, %f8434; add.ftz.f32 %f8436, %f8262, %f8435; sub.ftz.f32 %f8437, %f8271, %f8436; div.approx.ftz.f32 %f8438, %f8437, %f8263; mul.ftz.f32 %f8439, %f8438, %f8438; mul.ftz.f32 %f8440, %f8438, %f8439; mul.ftz.f32 %f8441, %f8439, %f8439; mul.ftz.f32 %f8442, %f8264, %f8439; fma.rn.ftz.f32 %f8443, %f8263, %f8438, %f8442; mul.ftz.f32 %f8444, %f8266, %f8441; fma.rn.ftz.f32 %f8445, %f8265, %f8440, %f8444; add.ftz.f32 %f8446, %f8443, %f8445; add.ftz.f32 %f8447, %f8262, %f8446; sub.ftz.f32 %f8448, %f8271, %f8447; div.approx.ftz.f32 %f8449, %f8448, %f8263; mul.ftz.f32 %f8450, %f8449, %f8449; mul.ftz.f32 %f8451, %f8449, %f8450; mul.ftz.f32 %f8452, %f8450, %f8450; mul.ftz.f32 %f8453, %f8264, %f8450; fma.rn.ftz.f32 %f8454, %f8263, %f8449, %f8453; mul.ftz.f32 %f8455, %f8266, %f8452; fma.rn.ftz.f32 %f8456, %f8265, %f8451, %f8455; add.ftz.f32 %f8457, %f8454, %f8456; add.ftz.f32 %f8458, %f8262, %f8457; sub.ftz.f32 %f8459, %f8271, %f8458; div.approx.ftz.f32 %f8460, %f8459, %f8263; mul.ftz.f32 %f8461, %f8460, %f8460; mul.ftz.f32 %f8462, %f8460, %f8461; mul.ftz.f32 %f8463, %f8461, %f8461; mul.ftz.f32 %f8464, %f8264, %f8461; fma.rn.ftz.f32 %f8465, %f8263, %f8460, %f8464; mul.ftz.f32 %f8466, %f8266, %f8463; fma.rn.ftz.f32 %f8467, %f8265, %f8462, %f8466; add.ftz.f32 %f8468, %f8465, %f8467; add.ftz.f32 %f8469, %f8262, %f8468; sub.ftz.f32 %f8470, %f8271, %f8469; div.approx.ftz.f32 %f8471, %f8470, %f8263; mul.ftz.f32 %f8472, %f8471, %f8471; mul.ftz.f32 %f8473, %f8471, %f8472; mul.ftz.f32 %f8474, %f8472, %f8472; mul.ftz.f32 %f8475, %f8264, %f8472; fma.rn.ftz.f32 %f8476, %f8263, %f8471, %f8475; mul.ftz.f32 %f8477, %f8266, %f8474; fma.rn.ftz.f32 %f8478, %f8265, %f8473, %f8477; add.ftz.f32 %f8479, %f8476, %f8478; add.ftz.f32 %f8480, %f8262, %f8479; sub.ftz.f32 %f8481, %f8271, %f8480; div.approx.ftz.f32 %f8482, %f8481, %f8263; mul.ftz.f32 %f8483, %f8482, %f8482; mul.ftz.f32 %f8484, %f8482, %f8483; mul.ftz.f32 %f8485, %f8483, %f8483; mul.ftz.f32 %f8486, %f8264, %f8483; fma.rn.ftz.f32 %f8487, %f8263, %f8482, %f8486; mul.ftz.f32 %f8488, %f8266, %f8485; fma.rn.ftz.f32 %f8489, %f8265, %f8484, %f8488; add.ftz.f32 %f8490, %f8487, %f8489; add.ftz.f32 %f8491, %f8262, %f8490; sub.ftz.f32 %f8492, %f8271, %f8491; div.approx.ftz.f32 %f1496, %f8492, %f8263; @%p3 bra $L__BB106_1056; bra.uni $L__BB106_1053; $L__BB106_1056: shr.s32 %r1697, %r502, 31; and.b32 %r1698, %r1697, 1078530011; or.b32 %r1699, %r1698, %r503; mov.b32 %f16000, %r1699; bra.uni $L__BB106_1057; $L__BB106_1044: ld.const.u32 %r3058, [kernel_params+20]; setp.eq.s32 %p927, %r3058, 0; @%p927 bra $L__BB106_1080; ld.const.u32 %r3059, [kernel_params+20]; setp.eq.s32 %p928, %r3059, 1; @%p928 bra $L__BB106_1046; bra.uni $L__BB106_1087; $L__BB106_1046: ld.const.f32 %f1521, [kernel_params+24]; mul.ftz.f32 %f8592, %f1490, %f1490; fma.rn.ftz.f32 %f8593, %f1489, %f1489, %f8592; sqrt.approx.ftz.f32 %f8594, %f8593; abs.ftz.f32 %f1522, %f8594; abs.ftz.f32 %f1523, %f1488; setp.eq.ftz.f32 %p957, %f1523, 0f00000000; setp.eq.ftz.f32 %p958, %f1522, 0f00000000; and.pred %p959, %p957, %p958; mov.b32 %r506, %f1488; mov.b32 %r1714, %f8594; and.b32 %r507, %r1714, -2147483648; @%p959 bra $L__BB106_1073; bra.uni $L__BB106_1047; $L__BB106_1073: shr.s32 %r1719, %r506, 31; and.b32 %r1720, %r1719, 1078530011; or.b32 %r1721, %r507, %r1720; mov.b32 %f16005, %r1721; bra.uni $L__BB106_1074; $L__BB106_1058: add.ftz.f32 %f8525, %f1489, 0fBF800000; mul.ftz.f32 %f8526, %f8525, 0fBF000000; mov.f32 %f8527, 0f00000000; max.ftz.f32 %f8528, %f8526, %f8527; sqrt.approx.ftz.f32 %f8529, %f8528; add.ftz.f32 %f1503, %f8529, %f8529; setp.leu.ftz.f32 %p936, %f1503, 0f00000000; @%p936 bra $L__BB106_1060; rcp.approx.ftz.f32 %f8530, %f1503; mul.ftz.f32 %f1488, %f1488, %f8530; mul.ftz.f32 %f1490, %f1490, %f8530; $L__BB106_1060: add.ftz.f32 %f8531, %f1488, 0f3F800000; add.ftz.f32 %f8532, %f1490, 0f3F800000; mul.ftz.f32 %f16010, %f8532, 0f3F000000; mul.ftz.f32 %f16009, %f8531, 0f3F000000; bra.uni $L__BB106_1093; $L__BB106_1061: abs.ftz.f32 %f1510, %f1488; setp.eq.ftz.f32 %p938, %f1510, 0f00000000; and.pred %p939, %p938, %p921; mov.b32 %r504, %f1488; and.b32 %r505, %r502, -2147483648; @%p939 bra $L__BB106_1065; bra.uni $L__BB106_1062; $L__BB106_1065: shr.s32 %r1704, %r504, 31; and.b32 %r1705, %r1704, 1078530011; or.b32 %r1706, %r1705, %r505; mov.b32 %f16003, %r1706; bra.uni $L__BB106_1066; $L__BB106_1080: setp.eq.ftz.f32 %p974, %f1489, 0f00000000; setp.eq.ftz.f32 %p975, %f1488, 0f00000000; and.pred %p976, %p975, %p974; setp.eq.ftz.f32 %p977, %f1490, 0f00000000; and.pred %p978, %p977, %p976; mov.f32 %f16009, 0f00000000; mov.f32 %f16010, %f16009; @%p978 bra $L__BB106_1093; ld.const.v4.f32 {%f8651, %f8652, %f8653, %f8654}, [kernel_params+32]; abs.ftz.f32 %f1539, %f1488; setp.eq.ftz.f32 %p980, %f1539, 0f00000000; and.pred %p981, %p980, %p921; mov.b32 %r508, %f1488; and.b32 %r509, %r502, -2147483648; @%p981 bra $L__BB106_1085; bra.uni $L__BB106_1082; $L__BB106_1085: shr.s32 %r1733, %r508, 31; and.b32 %r1734, %r1733, 1078530011; or.b32 %r1735, %r1734, %r509; mov.b32 %f16007, %r1735; bra.uni $L__BB106_1086; $L__BB106_1087: ld.const.f32 %f8713, [kernel_params+28]; ld.const.f32 %f1546, [kernel_params+276]; ld.const.f32 %f1547, [kernel_params+280]; add.ftz.f32 %f8714, %f8713, %f8713; mul.ftz.f32 %f8715, %f1491, 0f3F000000; sin.approx.ftz.f32 %f8716, %f8715; mul.ftz.f32 %f1548, %f8714, %f8716; @%p3 bra $L__BB106_1091; bra.uni $L__BB106_1088; $L__BB106_1091: shr.s32 %r1740, %r502, 31; and.b32 %r1741, %r1740, 1078530011; or.b32 %r1742, %r1741, %r503; mov.b32 %f16008, %r1742; bra.uni $L__BB106_1092; $L__BB106_1053: setp.eq.ftz.f32 %p929, %f1492, 0f7F800000; setp.eq.ftz.f32 %p930, %f1493, 0f7F800000; and.pred %p931, %p929, %p930; @%p931 bra $L__BB106_1055; bra.uni $L__BB106_1054; $L__BB106_1055: setp.lt.s32 %p935, %r502, 0; selp.b32 %r1695, 1075235812, 1061752795, %p935; or.b32 %r1696, %r1695, %r503; mov.b32 %f16000, %r1696; bra.uni $L__BB106_1057; $L__BB106_1047: setp.eq.ftz.f32 %p960, %f1523, 0f7F800000; setp.eq.ftz.f32 %p961, %f1522, 0f7F800000; and.pred %p962, %p960, %p961; @%p962 bra $L__BB106_1072; bra.uni $L__BB106_1048; $L__BB106_1072: setp.lt.s32 %p966, %r506, 0; selp.b32 %r1717, 1075235812, 1061752795, %p966; or.b32 %r1718, %r507, %r1717; mov.b32 %f16005, %r1718; bra.uni $L__BB106_1074; $L__BB106_1062: setp.eq.ftz.f32 %p940, %f1510, 0f7F800000; setp.eq.ftz.f32 %p941, %f1492, 0f7F800000; and.pred %p942, %p940, %p941; @%p942 bra $L__BB106_1064; bra.uni $L__BB106_1063; $L__BB106_1064: setp.lt.s32 %p946, %r504, 0; selp.b32 %r1702, 1075235812, 1061752795, %p946; or.b32 %r1703, %r1702, %r505; mov.b32 %f16003, %r1703; bra.uni $L__BB106_1066; $L__BB106_1088: setp.eq.ftz.f32 %p991, %f1492, 0f7F800000; setp.eq.ftz.f32 %p992, %f1493, 0f7F800000; and.pred %p993, %p991, %p992; @%p993 bra $L__BB106_1090; bra.uni $L__BB106_1089; $L__BB106_1090: setp.lt.s32 %p997, %r502, 0; selp.b32 %r1738, 1075235812, 1061752795, %p997; or.b32 %r1739, %r1738, %r503; mov.b32 %f16008, %r1739; bra.uni $L__BB106_1092; $L__BB106_1082: setp.eq.ftz.f32 %p982, %f1539, 0f7F800000; setp.eq.ftz.f32 %p983, %f1492, 0f7F800000; and.pred %p984, %p982, %p983; @%p984 bra $L__BB106_1084; bra.uni $L__BB106_1083; $L__BB106_1084: setp.lt.s32 %p988, %r508, 0; selp.b32 %r1731, 1075235812, 1061752795, %p988; or.b32 %r1732, %r1731, %r509; mov.b32 %f16007, %r1732; bra.uni $L__BB106_1086; $L__BB106_1054: mov.f32 %f15489, 0f3FC90FDB; setp.lt.s32 %p932, %r502, 0; min.ftz.f32 %f8493, %f1493, %f1492; max.ftz.f32 %f8494, %f1493, %f1492; div.full.ftz.f32 %f8495, %f8493, %f8494; mul.rn.ftz.f32 %f8496, %f8495, %f8495; mov.f32 %f8497, 0fC0B59883; mov.f32 %f8498, 0fBF52C7EA; fma.rn.ftz.f32 %f8499, %f8496, %f8498, %f8497; mov.f32 %f8500, 0fC0D21907; fma.rn.ftz.f32 %f8501, %f8499, %f8496, %f8500; mul.ftz.f32 %f8502, %f8496, %f8501; mul.ftz.f32 %f8503, %f8495, %f8502; add.ftz.f32 %f8504, %f8496, 0f41355DC0; mov.f32 %f8505, 0f41E6BD60; fma.rn.ftz.f32 %f8506, %f8504, %f8496, %f8505; mov.f32 %f8507, 0f419D92C8; fma.rn.ftz.f32 %f8508, %f8506, %f8496, %f8507; rcp.approx.ftz.f32 %f8509, %f8508; fma.rn.ftz.f32 %f8510, %f8503, %f8509, %f8495; sub.ftz.f32 %f8512, %f15489, %f8510; setp.gt.ftz.f32 %p933, %f1493, %f1492; selp.f32 %f8513, %f8512, %f8510, %p933; sub.ftz.f32 %f8515, %f7706, %f8513; selp.f32 %f8516, %f8515, %f8513, %p932; mov.b32 %r1693, %f8516; or.b32 %r1694, %r503, %r1693; mov.b32 %f8517, %r1694; add.ftz.f32 %f8518, %f1492, %f1493; setp.le.ftz.f32 %p934, %f8518, 0f7F800000; selp.f32 %f16000, %f8517, %f8518, %p934; $L__BB106_1057: ld.const.f32 %f15491, [kernel_params+280]; ld.const.f32 %f15490, [kernel_params+276]; cos.approx.ftz.f32 %f8519, %f16000; mul.ftz.f32 %f8520, %f1496, %f8519; div.approx.ftz.f32 %f8521, %f8520, %f15490; sin.approx.ftz.f32 %f8522, %f16000; mul.ftz.f32 %f8523, %f1496, %f8522; div.approx.ftz.f32 %f8524, %f8523, %f15491; add.ftz.f32 %f16010, %f8524, 0f3F000000; add.ftz.f32 %f16009, %f8521, 0f3F000000; bra.uni $L__BB106_1093; $L__BB106_1048: mov.f32 %f15516, 0f3FC90FDB; setp.lt.s32 %p963, %r506, 0; min.ftz.f32 %f8595, %f1522, %f1523; max.ftz.f32 %f8596, %f1522, %f1523; div.full.ftz.f32 %f8597, %f8595, %f8596; mul.rn.ftz.f32 %f8598, %f8597, %f8597; mov.f32 %f8599, 0fC0B59883; mov.f32 %f8600, 0fBF52C7EA; fma.rn.ftz.f32 %f8601, %f8598, %f8600, %f8599; mov.f32 %f8602, 0fC0D21907; fma.rn.ftz.f32 %f8603, %f8601, %f8598, %f8602; mul.ftz.f32 %f8604, %f8598, %f8603; mul.ftz.f32 %f8605, %f8597, %f8604; add.ftz.f32 %f8606, %f8598, 0f41355DC0; mov.f32 %f8607, 0f41E6BD60; fma.rn.ftz.f32 %f8608, %f8606, %f8598, %f8607; mov.f32 %f8609, 0f419D92C8; fma.rn.ftz.f32 %f8610, %f8608, %f8598, %f8609; rcp.approx.ftz.f32 %f8611, %f8610; fma.rn.ftz.f32 %f8612, %f8605, %f8611, %f8597; sub.ftz.f32 %f8614, %f15516, %f8612; setp.gt.ftz.f32 %p964, %f1522, %f1523; selp.f32 %f8615, %f8614, %f8612, %p964; sub.ftz.f32 %f8617, %f7706, %f8615; selp.f32 %f8618, %f8617, %f8615, %p963; mov.b32 %r1715, %f8618; or.b32 %r1716, %r507, %r1715; mov.b32 %f8619, %r1716; add.ftz.f32 %f8620, %f1523, %f1522; setp.le.ftz.f32 %p965, %f8620, 0f7F800000; selp.f32 %f16005, %f8619, %f8620, %p965; $L__BB106_1074: div.approx.ftz.f32 %f1528, %f16005, %f1521; @%p3 bra $L__BB106_1078; bra.uni $L__BB106_1075; $L__BB106_1078: shr.s32 %r1726, %r502, 31; and.b32 %r1727, %r1726, 1078530011; or.b32 %r1728, %r1727, %r503; mov.b32 %f16006, %r1728; bra.uni $L__BB106_1079; $L__BB106_1075: setp.eq.ftz.f32 %p967, %f1492, 0f7F800000; setp.eq.ftz.f32 %p968, %f1493, 0f7F800000; and.pred %p969, %p967, %p968; @%p969 bra $L__BB106_1077; bra.uni $L__BB106_1076; $L__BB106_1077: setp.lt.s32 %p973, %r502, 0; selp.b32 %r1724, 1075235812, 1061752795, %p973; or.b32 %r1725, %r1724, %r503; mov.b32 %f16006, %r1725; bra.uni $L__BB106_1079; $L__BB106_1063: mov.f32 %f15514, 0f3FC90FDB; setp.lt.s32 %p943, %r504, 0; min.ftz.f32 %f8533, %f1492, %f1510; max.ftz.f32 %f8534, %f1492, %f1510; div.full.ftz.f32 %f8535, %f8533, %f8534; mul.rn.ftz.f32 %f8536, %f8535, %f8535; mov.f32 %f8537, 0fC0B59883; mov.f32 %f8538, 0fBF52C7EA; fma.rn.ftz.f32 %f8539, %f8536, %f8538, %f8537; mov.f32 %f8540, 0fC0D21907; fma.rn.ftz.f32 %f8541, %f8539, %f8536, %f8540; mul.ftz.f32 %f8542, %f8536, %f8541; mul.ftz.f32 %f8543, %f8535, %f8542; add.ftz.f32 %f8544, %f8536, 0f41355DC0; mov.f32 %f8545, 0f41E6BD60; fma.rn.ftz.f32 %f8546, %f8544, %f8536, %f8545; mov.f32 %f8547, 0f419D92C8; fma.rn.ftz.f32 %f8548, %f8546, %f8536, %f8547; rcp.approx.ftz.f32 %f8549, %f8548; fma.rn.ftz.f32 %f8550, %f8543, %f8549, %f8535; sub.ftz.f32 %f8552, %f15514, %f8550; setp.gt.ftz.f32 %p944, %f1492, %f1510; selp.f32 %f8553, %f8552, %f8550, %p944; sub.ftz.f32 %f8555, %f7706, %f8553; selp.f32 %f8556, %f8555, %f8553, %p943; mov.b32 %r1700, %f8556; or.b32 %r1701, %r505, %r1700; mov.b32 %f8557, %r1701; add.ftz.f32 %f8558, %f1510, %f1492; setp.le.ftz.f32 %p945, %f8558, 0f7F800000; selp.f32 %f16003, %f8557, %f8558, %p945; $L__BB106_1066: and.pred %p949, %p938, %p922; @%p949 bra $L__BB106_1070; bra.uni $L__BB106_1067; $L__BB106_1070: shr.s32 %r1711, %r504, 31; and.b32 %r1712, %r1711, 1078530011; or.b32 %r1713, %r503, %r1712; mov.b32 %f16004, %r1713; bra.uni $L__BB106_1071; $L__BB106_1067: setp.eq.ftz.f32 %p950, %f1510, 0f7F800000; setp.eq.ftz.f32 %p951, %f1493, 0f7F800000; and.pred %p952, %p950, %p951; @%p952 bra $L__BB106_1069; bra.uni $L__BB106_1068; $L__BB106_1069: setp.lt.s32 %p956, %r504, 0; selp.b32 %r1709, 1075235812, 1061752795, %p956; or.b32 %r1710, %r503, %r1709; mov.b32 %f16004, %r1710; bra.uni $L__BB106_1071; $L__BB106_1089: mov.f32 %f15513, 0f3FC90FDB; setp.lt.s32 %p994, %r502, 0; min.ftz.f32 %f8717, %f1493, %f1492; max.ftz.f32 %f8718, %f1493, %f1492; div.full.ftz.f32 %f8719, %f8717, %f8718; mul.rn.ftz.f32 %f8720, %f8719, %f8719; mov.f32 %f8721, 0fC0B59883; mov.f32 %f8722, 0fBF52C7EA; fma.rn.ftz.f32 %f8723, %f8720, %f8722, %f8721; mov.f32 %f8724, 0fC0D21907; fma.rn.ftz.f32 %f8725, %f8723, %f8720, %f8724; mul.ftz.f32 %f8726, %f8720, %f8725; mul.ftz.f32 %f8727, %f8719, %f8726; add.ftz.f32 %f8728, %f8720, 0f41355DC0; mov.f32 %f8729, 0f41E6BD60; fma.rn.ftz.f32 %f8730, %f8728, %f8720, %f8729; mov.f32 %f8731, 0f419D92C8; fma.rn.ftz.f32 %f8732, %f8730, %f8720, %f8731; rcp.approx.ftz.f32 %f8733, %f8732; fma.rn.ftz.f32 %f8734, %f8727, %f8733, %f8719; sub.ftz.f32 %f8736, %f15513, %f8734; setp.gt.ftz.f32 %p995, %f1493, %f1492; selp.f32 %f8737, %f8736, %f8734, %p995; sub.ftz.f32 %f8739, %f7706, %f8737; selp.f32 %f8740, %f8739, %f8737, %p994; mov.b32 %r1736, %f8740; or.b32 %r1737, %r503, %r1736; mov.b32 %f8741, %r1737; add.ftz.f32 %f8742, %f1492, %f1493; setp.le.ftz.f32 %p996, %f8742, 0f7F800000; selp.f32 %f16008, %f8741, %f8742, %p996; $L__BB106_1092: cos.approx.ftz.f32 %f8743, %f16008; mul.ftz.f32 %f8744, %f1548, %f8743; div.approx.ftz.f32 %f8745, %f8744, %f1546; sin.approx.ftz.f32 %f8746, %f16008; mul.ftz.f32 %f8747, %f1548, %f8746; div.approx.ftz.f32 %f8748, %f8747, %f1547; add.ftz.f32 %f16010, %f8748, 0f3F000000; add.ftz.f32 %f16009, %f8745, 0f3F000000; bra.uni $L__BB106_1093; $L__BB106_1076: mov.f32 %f15517, 0f3FC90FDB; setp.lt.s32 %p970, %r502, 0; min.ftz.f32 %f8621, %f1493, %f1492; max.ftz.f32 %f8622, %f1493, %f1492; div.full.ftz.f32 %f8623, %f8621, %f8622; mul.rn.ftz.f32 %f8624, %f8623, %f8623; mov.f32 %f8625, 0fC0B59883; mov.f32 %f8626, 0fBF52C7EA; fma.rn.ftz.f32 %f8627, %f8624, %f8626, %f8625; mov.f32 %f8628, 0fC0D21907; fma.rn.ftz.f32 %f8629, %f8627, %f8624, %f8628; mul.ftz.f32 %f8630, %f8624, %f8629; mul.ftz.f32 %f8631, %f8623, %f8630; add.ftz.f32 %f8632, %f8624, 0f41355DC0; mov.f32 %f8633, 0f41E6BD60; fma.rn.ftz.f32 %f8634, %f8632, %f8624, %f8633; mov.f32 %f8635, 0f419D92C8; fma.rn.ftz.f32 %f8636, %f8634, %f8624, %f8635; rcp.approx.ftz.f32 %f8637, %f8636; fma.rn.ftz.f32 %f8638, %f8631, %f8637, %f8623; sub.ftz.f32 %f8640, %f15517, %f8638; setp.gt.ftz.f32 %p971, %f1493, %f1492; selp.f32 %f8641, %f8640, %f8638, %p971; sub.ftz.f32 %f8643, %f7706, %f8641; selp.f32 %f8644, %f8643, %f8641, %p970; mov.b32 %r1722, %f8644; or.b32 %r1723, %r503, %r1722; mov.b32 %f8645, %r1723; add.ftz.f32 %f8646, %f1492, %f1493; setp.le.ftz.f32 %p972, %f8646, 0f7F800000; selp.f32 %f16006, %f8645, %f8646, %p972; $L__BB106_1079: cos.approx.ftz.f32 %f8647, %f16006; sin.approx.ftz.f32 %f8648, %f16006; fma.rn.ftz.f32 %f16010, %f1528, %f8648, 0f3F000000; fma.rn.ftz.f32 %f16009, %f1528, %f8647, 0f3F000000; bra.uni $L__BB106_1093; $L__BB106_1068: mov.f32 %f15515, 0f3FC90FDB; setp.lt.s32 %p953, %r504, 0; min.ftz.f32 %f8559, %f1493, %f1510; max.ftz.f32 %f8560, %f1493, %f1510; div.full.ftz.f32 %f8561, %f8559, %f8560; mul.rn.ftz.f32 %f8562, %f8561, %f8561; mov.f32 %f8563, 0fC0B59883; mov.f32 %f8564, 0fBF52C7EA; fma.rn.ftz.f32 %f8565, %f8562, %f8564, %f8563; mov.f32 %f8566, 0fC0D21907; fma.rn.ftz.f32 %f8567, %f8565, %f8562, %f8566; mul.ftz.f32 %f8568, %f8562, %f8567; mul.ftz.f32 %f8569, %f8561, %f8568; add.ftz.f32 %f8570, %f8562, 0f41355DC0; mov.f32 %f8571, 0f41E6BD60; fma.rn.ftz.f32 %f8572, %f8570, %f8562, %f8571; mov.f32 %f8573, 0f419D92C8; fma.rn.ftz.f32 %f8574, %f8572, %f8562, %f8573; rcp.approx.ftz.f32 %f8575, %f8574; fma.rn.ftz.f32 %f8576, %f8569, %f8575, %f8561; sub.ftz.f32 %f8578, %f15515, %f8576; setp.gt.ftz.f32 %p954, %f1493, %f1510; selp.f32 %f8579, %f8578, %f8576, %p954; sub.ftz.f32 %f8581, %f7706, %f8579; selp.f32 %f8582, %f8581, %f8579, %p953; mov.b32 %r1707, %f8582; or.b32 %r1708, %r503, %r1707; mov.b32 %f8583, %r1708; add.ftz.f32 %f8584, %f1510, %f1493; setp.le.ftz.f32 %p955, %f8584, 0f7F800000; selp.f32 %f16004, %f8583, %f8584, %p955; $L__BB106_1071: mov.f32 %f15482, 0f3F800000; add.ftz.f32 %f8585, %f16004, %f16004; div.approx.ftz.f32 %f8587, %f8585, %f7706; add.ftz.f32 %f8588, %f16003, %f16003; div.approx.ftz.f32 %f8589, %f8588, %f7706; add.ftz.f32 %f8590, %f8589, 0f3F000000; sub.ftz.f32 %f16009, %f15482, %f8590; add.ftz.f32 %f16010, %f8587, 0f3F000000; bra.uni $L__BB106_1093; $L__BB106_1083: mov.f32 %f15518, 0f3FC90FDB; setp.lt.s32 %p985, %r508, 0; min.ftz.f32 %f8655, %f1492, %f1539; max.ftz.f32 %f8656, %f1492, %f1539; div.full.ftz.f32 %f8657, %f8655, %f8656; mul.rn.ftz.f32 %f8658, %f8657, %f8657; mov.f32 %f8659, 0fC0B59883; mov.f32 %f8660, 0fBF52C7EA; fma.rn.ftz.f32 %f8661, %f8658, %f8660, %f8659; mov.f32 %f8662, 0fC0D21907; fma.rn.ftz.f32 %f8663, %f8661, %f8658, %f8662; mul.ftz.f32 %f8664, %f8658, %f8663; mul.ftz.f32 %f8665, %f8657, %f8664; add.ftz.f32 %f8666, %f8658, 0f41355DC0; mov.f32 %f8667, 0f41E6BD60; fma.rn.ftz.f32 %f8668, %f8666, %f8658, %f8667; mov.f32 %f8669, 0f419D92C8; fma.rn.ftz.f32 %f8670, %f8668, %f8658, %f8669; rcp.approx.ftz.f32 %f8671, %f8670; fma.rn.ftz.f32 %f8672, %f8665, %f8671, %f8657; sub.ftz.f32 %f8674, %f15518, %f8672; setp.gt.ftz.f32 %p986, %f1492, %f1539; selp.f32 %f8675, %f8674, %f8672, %p986; sub.ftz.f32 %f8677, %f7706, %f8675; selp.f32 %f8678, %f8677, %f8675, %p985; mov.b32 %r1729, %f8678; or.b32 %r1730, %r509, %r1729; mov.b32 %f8679, %r1730; add.ftz.f32 %f8680, %f1539, %f1492; setp.le.ftz.f32 %p987, %f8680, 0f7F800000; selp.f32 %f16007, %f8679, %f8680, %p987; $L__BB106_1086: mov.f32 %f15519, 0f3FC90FDB; mov.f32 %f15488, 0f3E2AAB94; mov.f32 %f15487, 0f3D994929; mov.f32 %f15486, 0f3D3F841F; mov.f32 %f15485, 0f3C94D2E9; mov.f32 %f15484, 0f3D53F941; mov.f32 %f15483, 0f3F800000; sub.ftz.f32 %f8681, %f16007, %f8652; mul.ftz.f32 %f8682, %f1489, %f1489; fma.rn.ftz.f32 %f8683, %f1488, %f1488, %f8682; fma.rn.ftz.f32 %f8684, %f1490, %f1490, %f8683; sqrt.approx.ftz.f32 %f8685, %f8684; div.approx.ftz.f32 %f8686, %f1490, %f8685; abs.ftz.f32 %f8687, %f8686; sub.ftz.f32 %f8689, %f15483, %f8687; mul.ftz.f32 %f8690, %f8689, 0f3F000000; sqrt.approx.ftz.f32 %f8691, %f8690; setp.gt.ftz.f32 %p989, %f8687, 0f3F11EB85; selp.f32 %f8692, %f8691, %f8687, %p989; mul.ftz.f32 %f8693, %f8692, %f8692; fma.rn.ftz.f32 %f8696, %f15484, %f8693, %f15485; fma.rn.ftz.f32 %f8698, %f8696, %f8693, %f15486; fma.rn.ftz.f32 %f8700, %f8698, %f8693, %f15487; fma.rn.ftz.f32 %f8702, %f8700, %f8693, %f15488; mul.ftz.f32 %f8703, %f8693, %f8702; fma.rn.ftz.f32 %f8704, %f8703, %f8692, %f8692; add.ftz.f32 %f8705, %f8704, %f8704; sub.ftz.f32 %f8707, %f15519, %f8704; selp.f32 %f8708, %f8705, %f8707, %p989; setp.lt.ftz.f32 %p990, %f8686, 0f00000000; sub.ftz.f32 %f8710, %f7706, %f8708; selp.f32 %f8711, %f8710, %f8708, %p990; sub.ftz.f32 %f8712, %f8711, %f8654; div.approx.ftz.f32 %f16010, %f8712, %f8653; div.approx.ftz.f32 %f16009, %f8681, %f8651; $L__BB106_1093: ld.const.u32 %r3060, [kernel_params+20]; setp.gt.s32 %p2182, %r3060, 2; mov.f32 %f15492, 0f3FC90FDB; mov.f32 %f15474, 0fBF800000; mov.f32 %f15473, 0f3E2AAB94; mov.f32 %f15472, 0f3D994929; mov.f32 %f15471, 0f3D3F841F; mov.f32 %f15470, 0f3C94D2E9; mov.f32 %f15469, 0f3D53F941; mov.f32 %f15468, 0f3F800000; ld.const.v4.f32 {%f8749, %f8750, %f8751, %f8752}, [kernel_params+912]; ld.const.v4.f32 {%f8757, %f8758, %f8759, %f8760}, [kernel_params+928]; ld.const.v4.f32 {%f8765, %f8766, %f8767, %f8768}, [kernel_params+944]; mul.ftz.f32 %f8773, %f1379, %f8750; fma.rn.ftz.f32 %f8774, %f1378, %f8749, %f8773; fma.rn.ftz.f32 %f8775, %f1380, %f8751, %f8774; add.ftz.f32 %f8776, %f8752, %f8775; mul.ftz.f32 %f8777, %f1379, %f8758; fma.rn.ftz.f32 %f8778, %f1378, %f8757, %f8777; fma.rn.ftz.f32 %f8779, %f1380, %f8759, %f8778; add.ftz.f32 %f8780, %f8760, %f8779; mul.ftz.f32 %f8781, %f1379, %f8766; fma.rn.ftz.f32 %f8782, %f1378, %f8765, %f8781; fma.rn.ftz.f32 %f8783, %f1380, %f8767, %f8782; add.ftz.f32 %f8784, %f8768, %f8783; mul.ftz.f32 %f8785, %f8780, %f8780; fma.rn.ftz.f32 %f8786, %f8776, %f8776, %f8785; fma.rn.ftz.f32 %f8787, %f8784, %f8784, %f8786; sqrt.approx.ftz.f32 %f8788, %f8787; div.approx.ftz.f32 %f1557, %f8776, %f8788; div.approx.ftz.f32 %f1558, %f8780, %f8788; div.approx.ftz.f32 %f1559, %f8784, %f8788; max.ftz.f32 %f8790, %f1557, %f15474; min.ftz.f32 %f8792, %f8790, %f15468; abs.ftz.f32 %f8793, %f8792; sub.ftz.f32 %f8794, %f15468, %f8793; mul.ftz.f32 %f8795, %f8794, 0f3F000000; sqrt.approx.ftz.f32 %f8796, %f8795; setp.gt.ftz.f32 %p998, %f8793, 0f3F11EB85; selp.f32 %f8797, %f8796, %f8793, %p998; mul.ftz.f32 %f8798, %f8797, %f8797; fma.rn.ftz.f32 %f8801, %f15469, %f8798, %f15470; fma.rn.ftz.f32 %f8803, %f8801, %f8798, %f15471; fma.rn.ftz.f32 %f8805, %f8803, %f8798, %f15472; fma.rn.ftz.f32 %f8807, %f8805, %f8798, %f15473; mul.ftz.f32 %f8808, %f8798, %f8807; fma.rn.ftz.f32 %f8809, %f8808, %f8797, %f8797; add.ftz.f32 %f8810, %f8809, %f8809; sub.ftz.f32 %f8812, %f15492, %f8809; selp.f32 %f8813, %f8810, %f8812, %p998; setp.lt.ftz.f32 %p999, %f8792, 0f00000000; sub.ftz.f32 %f8815, %f7706, %f8813; selp.f32 %f1560, %f8815, %f8813, %p999; abs.ftz.f32 %f1561, %f1558; abs.ftz.f32 %f1562, %f1559; setp.eq.ftz.f32 %p1000, %f1561, 0f00000000; setp.eq.ftz.f32 %p1001, %f1562, 0f00000000; and.pred %p4, %p1000, %p1001; mov.b32 %r510, %f1558; mov.b32 %r1743, %f1559; and.b32 %r511, %r1743, -2147483648; @%p2182 bra $L__BB106_1099; bra.uni $L__BB106_1094; $L__BB106_1099: ld.const.u32 %r3049, [kernel_params+20]; setp.eq.s32 %p1003, %r3049, 3; @%p1003 bra $L__BB106_1108; ld.const.u32 %r3050, [kernel_params+20]; setp.eq.s32 %p1004, %r3050, 5; @%p1004 bra $L__BB106_1111; ld.const.u32 %r3051, [kernel_params+20]; setp.ne.s32 %p1005, %r3051, 4; @%p1005 bra $L__BB106_1137; ld.const.f32 %f8816, [kernel_params+48]; ld.const.v4.f32 {%f8817, %f8818, %f8819, %f8820}, [kernel_params+64]; neg.ftz.f32 %f8825, %f1560; sub.ftz.f32 %f8826, %f8825, %f8816; div.approx.ftz.f32 %f8827, %f8826, %f8817; mul.ftz.f32 %f8828, %f8827, %f8827; mul.ftz.f32 %f8829, %f8827, %f8828; mul.ftz.f32 %f8830, %f8828, %f8828; mul.ftz.f32 %f8831, %f8818, %f8828; fma.rn.ftz.f32 %f8832, %f8817, %f8827, %f8831; mul.ftz.f32 %f8833, %f8820, %f8830; fma.rn.ftz.f32 %f8834, %f8819, %f8829, %f8833; add.ftz.f32 %f8835, %f8832, %f8834; add.ftz.f32 %f8836, %f8816, %f8835; sub.ftz.f32 %f8837, %f8825, %f8836; div.approx.ftz.f32 %f8838, %f8837, %f8817; mul.ftz.f32 %f8839, %f8838, %f8838; mul.ftz.f32 %f8840, %f8838, %f8839; mul.ftz.f32 %f8841, %f8839, %f8839; mul.ftz.f32 %f8842, %f8818, %f8839; fma.rn.ftz.f32 %f8843, %f8817, %f8838, %f8842; mul.ftz.f32 %f8844, %f8820, %f8841; fma.rn.ftz.f32 %f8845, %f8819, %f8840, %f8844; add.ftz.f32 %f8846, %f8843, %f8845; add.ftz.f32 %f8847, %f8816, %f8846; sub.ftz.f32 %f8848, %f8825, %f8847; div.approx.ftz.f32 %f8849, %f8848, %f8817; mul.ftz.f32 %f8850, %f8849, %f8849; mul.ftz.f32 %f8851, %f8849, %f8850; mul.ftz.f32 %f8852, %f8850, %f8850; mul.ftz.f32 %f8853, %f8818, %f8850; fma.rn.ftz.f32 %f8854, %f8817, %f8849, %f8853; mul.ftz.f32 %f8855, %f8820, %f8852; fma.rn.ftz.f32 %f8856, %f8819, %f8851, %f8855; add.ftz.f32 %f8857, %f8854, %f8856; add.ftz.f32 %f8858, %f8816, %f8857; sub.ftz.f32 %f8859, %f8825, %f8858; div.approx.ftz.f32 %f8860, %f8859, %f8817; mul.ftz.f32 %f8861, %f8860, %f8860; mul.ftz.f32 %f8862, %f8860, %f8861; mul.ftz.f32 %f8863, %f8861, %f8861; mul.ftz.f32 %f8864, %f8818, %f8861; fma.rn.ftz.f32 %f8865, %f8817, %f8860, %f8864; mul.ftz.f32 %f8866, %f8820, %f8863; fma.rn.ftz.f32 %f8867, %f8819, %f8862, %f8866; add.ftz.f32 %f8868, %f8865, %f8867; add.ftz.f32 %f8869, %f8816, %f8868; sub.ftz.f32 %f8870, %f8825, %f8869; div.approx.ftz.f32 %f8871, %f8870, %f8817; mul.ftz.f32 %f8872, %f8871, %f8871; mul.ftz.f32 %f8873, %f8871, %f8872; mul.ftz.f32 %f8874, %f8872, %f8872; mul.ftz.f32 %f8875, %f8818, %f8872; fma.rn.ftz.f32 %f8876, %f8817, %f8871, %f8875; mul.ftz.f32 %f8877, %f8820, %f8874; fma.rn.ftz.f32 %f8878, %f8819, %f8873, %f8877; add.ftz.f32 %f8879, %f8876, %f8878; add.ftz.f32 %f8880, %f8816, %f8879; sub.ftz.f32 %f8881, %f8825, %f8880; div.approx.ftz.f32 %f8882, %f8881, %f8817; mul.ftz.f32 %f8883, %f8882, %f8882; mul.ftz.f32 %f8884, %f8882, %f8883; mul.ftz.f32 %f8885, %f8883, %f8883; mul.ftz.f32 %f8886, %f8818, %f8883; fma.rn.ftz.f32 %f8887, %f8817, %f8882, %f8886; mul.ftz.f32 %f8888, %f8820, %f8885; fma.rn.ftz.f32 %f8889, %f8819, %f8884, %f8888; add.ftz.f32 %f8890, %f8887, %f8889; add.ftz.f32 %f8891, %f8816, %f8890; sub.ftz.f32 %f8892, %f8825, %f8891; div.approx.ftz.f32 %f8893, %f8892, %f8817; mul.ftz.f32 %f8894, %f8893, %f8893; mul.ftz.f32 %f8895, %f8893, %f8894; mul.ftz.f32 %f8896, %f8894, %f8894; mul.ftz.f32 %f8897, %f8818, %f8894; fma.rn.ftz.f32 %f8898, %f8817, %f8893, %f8897; mul.ftz.f32 %f8899, %f8820, %f8896; fma.rn.ftz.f32 %f8900, %f8819, %f8895, %f8899; add.ftz.f32 %f8901, %f8898, %f8900; add.ftz.f32 %f8902, %f8816, %f8901; sub.ftz.f32 %f8903, %f8825, %f8902; div.approx.ftz.f32 %f8904, %f8903, %f8817; mul.ftz.f32 %f8905, %f8904, %f8904; mul.ftz.f32 %f8906, %f8904, %f8905; mul.ftz.f32 %f8907, %f8905, %f8905; mul.ftz.f32 %f8908, %f8818, %f8905; fma.rn.ftz.f32 %f8909, %f8817, %f8904, %f8908; mul.ftz.f32 %f8910, %f8820, %f8907; fma.rn.ftz.f32 %f8911, %f8819, %f8906, %f8910; add.ftz.f32 %f8912, %f8909, %f8911; add.ftz.f32 %f8913, %f8816, %f8912; sub.ftz.f32 %f8914, %f8825, %f8913; div.approx.ftz.f32 %f8915, %f8914, %f8817; mul.ftz.f32 %f8916, %f8915, %f8915; mul.ftz.f32 %f8917, %f8915, %f8916; mul.ftz.f32 %f8918, %f8916, %f8916; mul.ftz.f32 %f8919, %f8818, %f8916; fma.rn.ftz.f32 %f8920, %f8817, %f8915, %f8919; mul.ftz.f32 %f8921, %f8820, %f8918; fma.rn.ftz.f32 %f8922, %f8819, %f8917, %f8921; add.ftz.f32 %f8923, %f8920, %f8922; add.ftz.f32 %f8924, %f8816, %f8923; sub.ftz.f32 %f8925, %f8825, %f8924; div.approx.ftz.f32 %f8926, %f8925, %f8817; mul.ftz.f32 %f8927, %f8926, %f8926; mul.ftz.f32 %f8928, %f8926, %f8927; mul.ftz.f32 %f8929, %f8927, %f8927; mul.ftz.f32 %f8930, %f8818, %f8927; fma.rn.ftz.f32 %f8931, %f8817, %f8926, %f8930; mul.ftz.f32 %f8932, %f8820, %f8929; fma.rn.ftz.f32 %f8933, %f8819, %f8928, %f8932; add.ftz.f32 %f8934, %f8931, %f8933; add.ftz.f32 %f8935, %f8816, %f8934; sub.ftz.f32 %f8936, %f8825, %f8935; div.approx.ftz.f32 %f8937, %f8936, %f8817; mul.ftz.f32 %f8938, %f8937, %f8937; mul.ftz.f32 %f8939, %f8937, %f8938; mul.ftz.f32 %f8940, %f8938, %f8938; mul.ftz.f32 %f8941, %f8818, %f8938; fma.rn.ftz.f32 %f8942, %f8817, %f8937, %f8941; mul.ftz.f32 %f8943, %f8820, %f8940; fma.rn.ftz.f32 %f8944, %f8819, %f8939, %f8943; add.ftz.f32 %f8945, %f8942, %f8944; add.ftz.f32 %f8946, %f8816, %f8945; sub.ftz.f32 %f8947, %f8825, %f8946; div.approx.ftz.f32 %f8948, %f8947, %f8817; mul.ftz.f32 %f8949, %f8948, %f8948; mul.ftz.f32 %f8950, %f8948, %f8949; mul.ftz.f32 %f8951, %f8949, %f8949; mul.ftz.f32 %f8952, %f8818, %f8949; fma.rn.ftz.f32 %f8953, %f8817, %f8948, %f8952; mul.ftz.f32 %f8954, %f8820, %f8951; fma.rn.ftz.f32 %f8955, %f8819, %f8950, %f8954; add.ftz.f32 %f8956, %f8953, %f8955; add.ftz.f32 %f8957, %f8816, %f8956; sub.ftz.f32 %f8958, %f8825, %f8957; div.approx.ftz.f32 %f8959, %f8958, %f8817; mul.ftz.f32 %f8960, %f8959, %f8959; mul.ftz.f32 %f8961, %f8959, %f8960; mul.ftz.f32 %f8962, %f8960, %f8960; mul.ftz.f32 %f8963, %f8818, %f8960; fma.rn.ftz.f32 %f8964, %f8817, %f8959, %f8963; mul.ftz.f32 %f8965, %f8820, %f8962; fma.rn.ftz.f32 %f8966, %f8819, %f8961, %f8965; add.ftz.f32 %f8967, %f8964, %f8966; add.ftz.f32 %f8968, %f8816, %f8967; sub.ftz.f32 %f8969, %f8825, %f8968; div.approx.ftz.f32 %f8970, %f8969, %f8817; mul.ftz.f32 %f8971, %f8970, %f8970; mul.ftz.f32 %f8972, %f8970, %f8971; mul.ftz.f32 %f8973, %f8971, %f8971; mul.ftz.f32 %f8974, %f8818, %f8971; fma.rn.ftz.f32 %f8975, %f8817, %f8970, %f8974; mul.ftz.f32 %f8976, %f8820, %f8973; fma.rn.ftz.f32 %f8977, %f8819, %f8972, %f8976; add.ftz.f32 %f8978, %f8975, %f8977; add.ftz.f32 %f8979, %f8816, %f8978; sub.ftz.f32 %f8980, %f8825, %f8979; div.approx.ftz.f32 %f8981, %f8980, %f8817; mul.ftz.f32 %f8982, %f8981, %f8981; mul.ftz.f32 %f8983, %f8981, %f8982; mul.ftz.f32 %f8984, %f8982, %f8982; mul.ftz.f32 %f8985, %f8818, %f8982; fma.rn.ftz.f32 %f8986, %f8817, %f8981, %f8985; mul.ftz.f32 %f8987, %f8820, %f8984; fma.rn.ftz.f32 %f8988, %f8819, %f8983, %f8987; add.ftz.f32 %f8989, %f8986, %f8988; add.ftz.f32 %f8990, %f8816, %f8989; sub.ftz.f32 %f8991, %f8825, %f8990; div.approx.ftz.f32 %f8992, %f8991, %f8817; mul.ftz.f32 %f8993, %f8992, %f8992; mul.ftz.f32 %f8994, %f8992, %f8993; mul.ftz.f32 %f8995, %f8993, %f8993; mul.ftz.f32 %f8996, %f8818, %f8993; fma.rn.ftz.f32 %f8997, %f8817, %f8992, %f8996; mul.ftz.f32 %f8998, %f8820, %f8995; fma.rn.ftz.f32 %f8999, %f8819, %f8994, %f8998; add.ftz.f32 %f9000, %f8997, %f8999; add.ftz.f32 %f9001, %f8816, %f9000; sub.ftz.f32 %f9002, %f8825, %f9001; div.approx.ftz.f32 %f9003, %f9002, %f8817; mul.ftz.f32 %f9004, %f9003, %f9003; mul.ftz.f32 %f9005, %f9003, %f9004; mul.ftz.f32 %f9006, %f9004, %f9004; mul.ftz.f32 %f9007, %f8818, %f9004; fma.rn.ftz.f32 %f9008, %f8817, %f9003, %f9007; mul.ftz.f32 %f9009, %f8820, %f9006; fma.rn.ftz.f32 %f9010, %f8819, %f9005, %f9009; add.ftz.f32 %f9011, %f9008, %f9010; add.ftz.f32 %f9012, %f8816, %f9011; sub.ftz.f32 %f9013, %f8825, %f9012; div.approx.ftz.f32 %f9014, %f9013, %f8817; mul.ftz.f32 %f9015, %f9014, %f9014; mul.ftz.f32 %f9016, %f9014, %f9015; mul.ftz.f32 %f9017, %f9015, %f9015; mul.ftz.f32 %f9018, %f8818, %f9015; fma.rn.ftz.f32 %f9019, %f8817, %f9014, %f9018; mul.ftz.f32 %f9020, %f8820, %f9017; fma.rn.ftz.f32 %f9021, %f8819, %f9016, %f9020; add.ftz.f32 %f9022, %f9019, %f9021; add.ftz.f32 %f9023, %f8816, %f9022; sub.ftz.f32 %f9024, %f8825, %f9023; div.approx.ftz.f32 %f9025, %f9024, %f8817; mul.ftz.f32 %f9026, %f9025, %f9025; mul.ftz.f32 %f9027, %f9025, %f9026; mul.ftz.f32 %f9028, %f9026, %f9026; mul.ftz.f32 %f9029, %f8818, %f9026; fma.rn.ftz.f32 %f9030, %f8817, %f9025, %f9029; mul.ftz.f32 %f9031, %f8820, %f9028; fma.rn.ftz.f32 %f9032, %f8819, %f9027, %f9031; add.ftz.f32 %f9033, %f9030, %f9032; add.ftz.f32 %f9034, %f8816, %f9033; sub.ftz.f32 %f9035, %f8825, %f9034; div.approx.ftz.f32 %f9036, %f9035, %f8817; mul.ftz.f32 %f9037, %f9036, %f9036; mul.ftz.f32 %f9038, %f9036, %f9037; mul.ftz.f32 %f9039, %f9037, %f9037; mul.ftz.f32 %f9040, %f8818, %f9037; fma.rn.ftz.f32 %f9041, %f8817, %f9036, %f9040; mul.ftz.f32 %f9042, %f8820, %f9039; fma.rn.ftz.f32 %f9043, %f8819, %f9038, %f9042; add.ftz.f32 %f9044, %f9041, %f9043; add.ftz.f32 %f9045, %f8816, %f9044; sub.ftz.f32 %f9046, %f8825, %f9045; div.approx.ftz.f32 %f1565, %f9046, %f8817; @%p4 bra $L__BB106_1106; bra.uni $L__BB106_1103; $L__BB106_1106: shr.s32 %r1748, %r510, 31; and.b32 %r1749, %r1748, 1078530011; or.b32 %r1750, %r1749, %r511; mov.b32 %f16011, %r1750; bra.uni $L__BB106_1107; $L__BB106_1094: ld.const.u32 %r3056, [kernel_params+20]; setp.eq.s32 %p1006, %r3056, 0; @%p1006 bra $L__BB106_1130; ld.const.u32 %r3057, [kernel_params+20]; setp.eq.s32 %p1007, %r3057, 1; @%p1007 bra $L__BB106_1096; bra.uni $L__BB106_1137; $L__BB106_1096: ld.const.f32 %f1590, [kernel_params+24]; mul.ftz.f32 %f9146, %f1559, %f1559; fma.rn.ftz.f32 %f9147, %f1558, %f1558, %f9146; sqrt.approx.ftz.f32 %f9148, %f9147; abs.ftz.f32 %f1591, %f9148; abs.ftz.f32 %f1592, %f1557; setp.eq.ftz.f32 %p1036, %f1592, 0f00000000; setp.eq.ftz.f32 %p1037, %f1591, 0f00000000; and.pred %p1038, %p1036, %p1037; mov.b32 %r514, %f1557; mov.b32 %r1765, %f9148; and.b32 %r515, %r1765, -2147483648; @%p1038 bra $L__BB106_1123; bra.uni $L__BB106_1097; $L__BB106_1123: shr.s32 %r1770, %r514, 31; and.b32 %r1771, %r1770, 1078530011; or.b32 %r1772, %r515, %r1771; mov.b32 %f16016, %r1772; bra.uni $L__BB106_1124; $L__BB106_1108: add.ftz.f32 %f9079, %f1558, 0fBF800000; mul.ftz.f32 %f9080, %f9079, 0fBF000000; mov.f32 %f9081, 0f00000000; max.ftz.f32 %f9082, %f9080, %f9081; sqrt.approx.ftz.f32 %f9083, %f9082; add.ftz.f32 %f1572, %f9083, %f9083; setp.leu.ftz.f32 %p1015, %f1572, 0f00000000; @%p1015 bra $L__BB106_1110; rcp.approx.ftz.f32 %f9084, %f1572; mul.ftz.f32 %f1557, %f1557, %f9084; mul.ftz.f32 %f1559, %f1559, %f9084; $L__BB106_1110: add.ftz.f32 %f9085, %f1557, 0f3F800000; add.ftz.f32 %f9086, %f1559, 0f3F800000; mul.ftz.f32 %f16021, %f9086, 0f3F000000; mul.ftz.f32 %f16020, %f9085, 0f3F000000; bra.uni $L__BB106_1143; $L__BB106_1111: abs.ftz.f32 %f1579, %f1557; setp.eq.ftz.f32 %p1017, %f1579, 0f00000000; and.pred %p1018, %p1017, %p1000; mov.b32 %r512, %f1557; and.b32 %r513, %r510, -2147483648; @%p1018 bra $L__BB106_1115; bra.uni $L__BB106_1112; $L__BB106_1115: shr.s32 %r1755, %r512, 31; and.b32 %r1756, %r1755, 1078530011; or.b32 %r1757, %r1756, %r513; mov.b32 %f16014, %r1757; bra.uni $L__BB106_1116; $L__BB106_1130: setp.eq.ftz.f32 %p1053, %f1558, 0f00000000; setp.eq.ftz.f32 %p1054, %f1557, 0f00000000; and.pred %p1055, %p1054, %p1053; setp.eq.ftz.f32 %p1056, %f1559, 0f00000000; and.pred %p1057, %p1056, %p1055; mov.f32 %f16020, 0f00000000; mov.f32 %f16021, %f16020; @%p1057 bra $L__BB106_1143; ld.const.v4.f32 {%f9205, %f9206, %f9207, %f9208}, [kernel_params+32]; abs.ftz.f32 %f1608, %f1557; setp.eq.ftz.f32 %p1059, %f1608, 0f00000000; and.pred %p1060, %p1059, %p1000; mov.b32 %r516, %f1557; and.b32 %r517, %r510, -2147483648; @%p1060 bra $L__BB106_1135; bra.uni $L__BB106_1132; $L__BB106_1135: shr.s32 %r1784, %r516, 31; and.b32 %r1785, %r1784, 1078530011; or.b32 %r1786, %r1785, %r517; mov.b32 %f16018, %r1786; bra.uni $L__BB106_1136; $L__BB106_1137: ld.const.f32 %f9267, [kernel_params+28]; ld.const.f32 %f1615, [kernel_params+276]; ld.const.f32 %f1616, [kernel_params+280]; add.ftz.f32 %f9268, %f9267, %f9267; mul.ftz.f32 %f9269, %f1560, 0f3F000000; sin.approx.ftz.f32 %f9270, %f9269; mul.ftz.f32 %f1617, %f9268, %f9270; @%p4 bra $L__BB106_1141; bra.uni $L__BB106_1138; $L__BB106_1141: shr.s32 %r1791, %r510, 31; and.b32 %r1792, %r1791, 1078530011; or.b32 %r1793, %r1792, %r511; mov.b32 %f16019, %r1793; bra.uni $L__BB106_1142; $L__BB106_1103: setp.eq.ftz.f32 %p1008, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1009, %f1562, 0f7F800000; and.pred %p1010, %p1008, %p1009; @%p1010 bra $L__BB106_1105; bra.uni $L__BB106_1104; $L__BB106_1105: setp.lt.s32 %p1014, %r510, 0; selp.b32 %r1746, 1075235812, 1061752795, %p1014; or.b32 %r1747, %r1746, %r511; mov.b32 %f16011, %r1747; bra.uni $L__BB106_1107; $L__BB106_1097: setp.eq.ftz.f32 %p1039, %f1592, 0f7F800000; setp.eq.ftz.f32 %p1040, %f1591, 0f7F800000; and.pred %p1041, %p1039, %p1040; @%p1041 bra $L__BB106_1122; bra.uni $L__BB106_1098; $L__BB106_1122: setp.lt.s32 %p1045, %r514, 0; selp.b32 %r1768, 1075235812, 1061752795, %p1045; or.b32 %r1769, %r515, %r1768; mov.b32 %f16016, %r1769; bra.uni $L__BB106_1124; $L__BB106_1112: setp.eq.ftz.f32 %p1019, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1020, %f1561, 0f7F800000; and.pred %p1021, %p1019, %p1020; @%p1021 bra $L__BB106_1114; bra.uni $L__BB106_1113; $L__BB106_1114: setp.lt.s32 %p1025, %r512, 0; selp.b32 %r1753, 1075235812, 1061752795, %p1025; or.b32 %r1754, %r1753, %r513; mov.b32 %f16014, %r1754; bra.uni $L__BB106_1116; $L__BB106_1138: setp.eq.ftz.f32 %p1070, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1071, %f1562, 0f7F800000; and.pred %p1072, %p1070, %p1071; @%p1072 bra $L__BB106_1140; bra.uni $L__BB106_1139; $L__BB106_1140: setp.lt.s32 %p1076, %r510, 0; selp.b32 %r1789, 1075235812, 1061752795, %p1076; or.b32 %r1790, %r1789, %r511; mov.b32 %f16019, %r1790; bra.uni $L__BB106_1142; $L__BB106_1132: setp.eq.ftz.f32 %p1061, %f1608, 0f7F800000; setp.eq.ftz.f32 %p1062, %f1561, 0f7F800000; and.pred %p1063, %p1061, %p1062; @%p1063 bra $L__BB106_1134; bra.uni $L__BB106_1133; $L__BB106_1134: setp.lt.s32 %p1067, %r516, 0; selp.b32 %r1782, 1075235812, 1061752795, %p1067; or.b32 %r1783, %r1782, %r517; mov.b32 %f16018, %r1783; bra.uni $L__BB106_1136; $L__BB106_1104: mov.f32 %f15503, 0f40490FDB; mov.f32 %f15496, 0f3FC90FDB; setp.lt.s32 %p1011, %r510, 0; min.ftz.f32 %f9047, %f1562, %f1561; max.ftz.f32 %f9048, %f1562, %f1561; div.full.ftz.f32 %f9049, %f9047, %f9048; mul.rn.ftz.f32 %f9050, %f9049, %f9049; mov.f32 %f9051, 0fC0B59883; mov.f32 %f9052, 0fBF52C7EA; fma.rn.ftz.f32 %f9053, %f9050, %f9052, %f9051; mov.f32 %f9054, 0fC0D21907; fma.rn.ftz.f32 %f9055, %f9053, %f9050, %f9054; mul.ftz.f32 %f9056, %f9050, %f9055; mul.ftz.f32 %f9057, %f9049, %f9056; add.ftz.f32 %f9058, %f9050, 0f41355DC0; mov.f32 %f9059, 0f41E6BD60; fma.rn.ftz.f32 %f9060, %f9058, %f9050, %f9059; mov.f32 %f9061, 0f419D92C8; fma.rn.ftz.f32 %f9062, %f9060, %f9050, %f9061; rcp.approx.ftz.f32 %f9063, %f9062; fma.rn.ftz.f32 %f9064, %f9057, %f9063, %f9049; sub.ftz.f32 %f9066, %f15496, %f9064; setp.gt.ftz.f32 %p1012, %f1562, %f1561; selp.f32 %f9067, %f9066, %f9064, %p1012; sub.ftz.f32 %f9069, %f15503, %f9067; selp.f32 %f9070, %f9069, %f9067, %p1011; mov.b32 %r1744, %f9070; or.b32 %r1745, %r511, %r1744; mov.b32 %f9071, %r1745; add.ftz.f32 %f9072, %f1561, %f1562; setp.le.ftz.f32 %p1013, %f9072, 0f7F800000; selp.f32 %f16011, %f9071, %f9072, %p1013; $L__BB106_1107: ld.const.f32 %f15505, [kernel_params+280]; ld.const.f32 %f15504, [kernel_params+276]; cos.approx.ftz.f32 %f9073, %f16011; mul.ftz.f32 %f9074, %f1565, %f9073; div.approx.ftz.f32 %f9075, %f9074, %f15504; sin.approx.ftz.f32 %f9076, %f16011; mul.ftz.f32 %f9077, %f1565, %f9076; div.approx.ftz.f32 %f9078, %f9077, %f15505; add.ftz.f32 %f16021, %f9078, 0f3F000000; add.ftz.f32 %f16020, %f9075, 0f3F000000; bra.uni $L__BB106_1143; $L__BB106_1098: mov.f32 %f15499, 0f3FC90FDB; setp.lt.s32 %p1042, %r514, 0; min.ftz.f32 %f9149, %f1591, %f1592; max.ftz.f32 %f9150, %f1591, %f1592; div.full.ftz.f32 %f9151, %f9149, %f9150; mul.rn.ftz.f32 %f9152, %f9151, %f9151; mov.f32 %f9153, 0fC0B59883; mov.f32 %f9154, 0fBF52C7EA; fma.rn.ftz.f32 %f9155, %f9152, %f9154, %f9153; mov.f32 %f9156, 0fC0D21907; fma.rn.ftz.f32 %f9157, %f9155, %f9152, %f9156; mul.ftz.f32 %f9158, %f9152, %f9157; mul.ftz.f32 %f9159, %f9151, %f9158; add.ftz.f32 %f9160, %f9152, 0f41355DC0; mov.f32 %f9161, 0f41E6BD60; fma.rn.ftz.f32 %f9162, %f9160, %f9152, %f9161; mov.f32 %f9163, 0f419D92C8; fma.rn.ftz.f32 %f9164, %f9162, %f9152, %f9163; rcp.approx.ftz.f32 %f9165, %f9164; fma.rn.ftz.f32 %f9166, %f9159, %f9165, %f9151; sub.ftz.f32 %f9168, %f15499, %f9166; setp.gt.ftz.f32 %p1043, %f1591, %f1592; selp.f32 %f9169, %f9168, %f9166, %p1043; sub.ftz.f32 %f9171, %f7706, %f9169; selp.f32 %f9172, %f9171, %f9169, %p1042; mov.b32 %r1766, %f9172; or.b32 %r1767, %r515, %r1766; mov.b32 %f9173, %r1767; add.ftz.f32 %f9174, %f1592, %f1591; setp.le.ftz.f32 %p1044, %f9174, 0f7F800000; selp.f32 %f16016, %f9173, %f9174, %p1044; $L__BB106_1124: div.approx.ftz.f32 %f1597, %f16016, %f1590; @%p4 bra $L__BB106_1128; bra.uni $L__BB106_1125; $L__BB106_1128: shr.s32 %r1777, %r510, 31; and.b32 %r1778, %r1777, 1078530011; or.b32 %r1779, %r1778, %r511; mov.b32 %f16017, %r1779; bra.uni $L__BB106_1129; $L__BB106_1125: setp.eq.ftz.f32 %p1046, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1047, %f1562, 0f7F800000; and.pred %p1048, %p1046, %p1047; @%p1048 bra $L__BB106_1127; bra.uni $L__BB106_1126; $L__BB106_1127: setp.lt.s32 %p1052, %r510, 0; selp.b32 %r1775, 1075235812, 1061752795, %p1052; or.b32 %r1776, %r1775, %r511; mov.b32 %f16017, %r1776; bra.uni $L__BB106_1129; $L__BB106_1113: mov.f32 %f15497, 0f3FC90FDB; setp.lt.s32 %p1022, %r512, 0; min.ftz.f32 %f9087, %f1561, %f1579; max.ftz.f32 %f9088, %f1561, %f1579; div.full.ftz.f32 %f9089, %f9087, %f9088; mul.rn.ftz.f32 %f9090, %f9089, %f9089; mov.f32 %f9091, 0fC0B59883; mov.f32 %f9092, 0fBF52C7EA; fma.rn.ftz.f32 %f9093, %f9090, %f9092, %f9091; mov.f32 %f9094, 0fC0D21907; fma.rn.ftz.f32 %f9095, %f9093, %f9090, %f9094; mul.ftz.f32 %f9096, %f9090, %f9095; mul.ftz.f32 %f9097, %f9089, %f9096; add.ftz.f32 %f9098, %f9090, 0f41355DC0; mov.f32 %f9099, 0f41E6BD60; fma.rn.ftz.f32 %f9100, %f9098, %f9090, %f9099; mov.f32 %f9101, 0f419D92C8; fma.rn.ftz.f32 %f9102, %f9100, %f9090, %f9101; rcp.approx.ftz.f32 %f9103, %f9102; fma.rn.ftz.f32 %f9104, %f9097, %f9103, %f9089; sub.ftz.f32 %f9106, %f15497, %f9104; setp.gt.ftz.f32 %p1023, %f1561, %f1579; selp.f32 %f9107, %f9106, %f9104, %p1023; sub.ftz.f32 %f9109, %f7706, %f9107; selp.f32 %f9110, %f9109, %f9107, %p1022; mov.b32 %r1751, %f9110; or.b32 %r1752, %r513, %r1751; mov.b32 %f9111, %r1752; add.ftz.f32 %f9112, %f1579, %f1561; setp.le.ftz.f32 %p1024, %f9112, 0f7F800000; selp.f32 %f16014, %f9111, %f9112, %p1024; $L__BB106_1116: and.pred %p1028, %p1017, %p1001; @%p1028 bra $L__BB106_1120; bra.uni $L__BB106_1117; $L__BB106_1120: shr.s32 %r1762, %r512, 31; and.b32 %r1763, %r1762, 1078530011; or.b32 %r1764, %r511, %r1763; mov.b32 %f16015, %r1764; bra.uni $L__BB106_1121; $L__BB106_1117: setp.eq.ftz.f32 %p1029, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1030, %f1562, 0f7F800000; and.pred %p1031, %p1029, %p1030; @%p1031 bra $L__BB106_1119; bra.uni $L__BB106_1118; $L__BB106_1119: setp.lt.s32 %p1035, %r512, 0; selp.b32 %r1760, 1075235812, 1061752795, %p1035; or.b32 %r1761, %r511, %r1760; mov.b32 %f16015, %r1761; bra.uni $L__BB106_1121; $L__BB106_1139: mov.f32 %f15493, 0f3FC90FDB; setp.lt.s32 %p1073, %r510, 0; min.ftz.f32 %f9271, %f1562, %f1561; max.ftz.f32 %f9272, %f1562, %f1561; div.full.ftz.f32 %f9273, %f9271, %f9272; mul.rn.ftz.f32 %f9274, %f9273, %f9273; mov.f32 %f9275, 0fC0B59883; mov.f32 %f9276, 0fBF52C7EA; fma.rn.ftz.f32 %f9277, %f9274, %f9276, %f9275; mov.f32 %f9278, 0fC0D21907; fma.rn.ftz.f32 %f9279, %f9277, %f9274, %f9278; mul.ftz.f32 %f9280, %f9274, %f9279; mul.ftz.f32 %f9281, %f9273, %f9280; add.ftz.f32 %f9282, %f9274, 0f41355DC0; mov.f32 %f9283, 0f41E6BD60; fma.rn.ftz.f32 %f9284, %f9282, %f9274, %f9283; mov.f32 %f9285, 0f419D92C8; fma.rn.ftz.f32 %f9286, %f9284, %f9274, %f9285; rcp.approx.ftz.f32 %f9287, %f9286; fma.rn.ftz.f32 %f9288, %f9281, %f9287, %f9273; sub.ftz.f32 %f9290, %f15493, %f9288; setp.gt.ftz.f32 %p1074, %f1562, %f1561; selp.f32 %f9291, %f9290, %f9288, %p1074; sub.ftz.f32 %f9293, %f7706, %f9291; selp.f32 %f9294, %f9293, %f9291, %p1073; mov.b32 %r1787, %f9294; or.b32 %r1788, %r511, %r1787; mov.b32 %f9295, %r1788; add.ftz.f32 %f9296, %f1561, %f1562; setp.le.ftz.f32 %p1075, %f9296, 0f7F800000; selp.f32 %f16019, %f9295, %f9296, %p1075; $L__BB106_1142: cos.approx.ftz.f32 %f9297, %f16019; mul.ftz.f32 %f9298, %f1617, %f9297; div.approx.ftz.f32 %f9299, %f9298, %f1615; sin.approx.ftz.f32 %f9300, %f16019; mul.ftz.f32 %f9301, %f1617, %f9300; div.approx.ftz.f32 %f9302, %f9301, %f1616; add.ftz.f32 %f16021, %f9302, 0f3F000000; add.ftz.f32 %f16020, %f9299, 0f3F000000; bra.uni $L__BB106_1143; $L__BB106_1126: mov.f32 %f15500, 0f3FC90FDB; setp.lt.s32 %p1049, %r510, 0; min.ftz.f32 %f9175, %f1562, %f1561; max.ftz.f32 %f9176, %f1562, %f1561; div.full.ftz.f32 %f9177, %f9175, %f9176; mul.rn.ftz.f32 %f9178, %f9177, %f9177; mov.f32 %f9179, 0fC0B59883; mov.f32 %f9180, 0fBF52C7EA; fma.rn.ftz.f32 %f9181, %f9178, %f9180, %f9179; mov.f32 %f9182, 0fC0D21907; fma.rn.ftz.f32 %f9183, %f9181, %f9178, %f9182; mul.ftz.f32 %f9184, %f9178, %f9183; mul.ftz.f32 %f9185, %f9177, %f9184; add.ftz.f32 %f9186, %f9178, 0f41355DC0; mov.f32 %f9187, 0f41E6BD60; fma.rn.ftz.f32 %f9188, %f9186, %f9178, %f9187; mov.f32 %f9189, 0f419D92C8; fma.rn.ftz.f32 %f9190, %f9188, %f9178, %f9189; rcp.approx.ftz.f32 %f9191, %f9190; fma.rn.ftz.f32 %f9192, %f9185, %f9191, %f9177; sub.ftz.f32 %f9194, %f15500, %f9192; setp.gt.ftz.f32 %p1050, %f1562, %f1561; selp.f32 %f9195, %f9194, %f9192, %p1050; sub.ftz.f32 %f9197, %f7706, %f9195; selp.f32 %f9198, %f9197, %f9195, %p1049; mov.b32 %r1773, %f9198; or.b32 %r1774, %r511, %r1773; mov.b32 %f9199, %r1774; add.ftz.f32 %f9200, %f1561, %f1562; setp.le.ftz.f32 %p1051, %f9200, 0f7F800000; selp.f32 %f16017, %f9199, %f9200, %p1051; $L__BB106_1129: cos.approx.ftz.f32 %f9201, %f16017; sin.approx.ftz.f32 %f9202, %f16017; fma.rn.ftz.f32 %f16021, %f1597, %f9202, 0f3F000000; fma.rn.ftz.f32 %f16020, %f1597, %f9201, 0f3F000000; bra.uni $L__BB106_1143; $L__BB106_1118: mov.f32 %f15498, 0f3FC90FDB; setp.lt.s32 %p1032, %r512, 0; min.ftz.f32 %f9113, %f1562, %f1579; max.ftz.f32 %f9114, %f1562, %f1579; div.full.ftz.f32 %f9115, %f9113, %f9114; mul.rn.ftz.f32 %f9116, %f9115, %f9115; mov.f32 %f9117, 0fC0B59883; mov.f32 %f9118, 0fBF52C7EA; fma.rn.ftz.f32 %f9119, %f9116, %f9118, %f9117; mov.f32 %f9120, 0fC0D21907; fma.rn.ftz.f32 %f9121, %f9119, %f9116, %f9120; mul.ftz.f32 %f9122, %f9116, %f9121; mul.ftz.f32 %f9123, %f9115, %f9122; add.ftz.f32 %f9124, %f9116, 0f41355DC0; mov.f32 %f9125, 0f41E6BD60; fma.rn.ftz.f32 %f9126, %f9124, %f9116, %f9125; mov.f32 %f9127, 0f419D92C8; fma.rn.ftz.f32 %f9128, %f9126, %f9116, %f9127; rcp.approx.ftz.f32 %f9129, %f9128; fma.rn.ftz.f32 %f9130, %f9123, %f9129, %f9115; sub.ftz.f32 %f9132, %f15498, %f9130; setp.gt.ftz.f32 %p1033, %f1562, %f1579; selp.f32 %f9133, %f9132, %f9130, %p1033; sub.ftz.f32 %f9135, %f7706, %f9133; selp.f32 %f9136, %f9135, %f9133, %p1032; mov.b32 %r1758, %f9136; or.b32 %r1759, %r511, %r1758; mov.b32 %f9137, %r1759; add.ftz.f32 %f9138, %f1579, %f1562; setp.le.ftz.f32 %p1034, %f9138, 0f7F800000; selp.f32 %f16015, %f9137, %f9138, %p1034; $L__BB106_1121: mov.f32 %f15475, 0f3F800000; add.ftz.f32 %f9139, %f16015, %f16015; div.approx.ftz.f32 %f9141, %f9139, %f7706; add.ftz.f32 %f9142, %f16014, %f16014; div.approx.ftz.f32 %f9143, %f9142, %f7706; add.ftz.f32 %f9144, %f9143, 0f3F000000; sub.ftz.f32 %f16020, %f15475, %f9144; add.ftz.f32 %f16021, %f9141, 0f3F000000; bra.uni $L__BB106_1143; $L__BB106_1133: mov.f32 %f15501, 0f3FC90FDB; setp.lt.s32 %p1064, %r516, 0; min.ftz.f32 %f9209, %f1561, %f1608; max.ftz.f32 %f9210, %f1561, %f1608; div.full.ftz.f32 %f9211, %f9209, %f9210; mul.rn.ftz.f32 %f9212, %f9211, %f9211; mov.f32 %f9213, 0fC0B59883; mov.f32 %f9214, 0fBF52C7EA; fma.rn.ftz.f32 %f9215, %f9212, %f9214, %f9213; mov.f32 %f9216, 0fC0D21907; fma.rn.ftz.f32 %f9217, %f9215, %f9212, %f9216; mul.ftz.f32 %f9218, %f9212, %f9217; mul.ftz.f32 %f9219, %f9211, %f9218; add.ftz.f32 %f9220, %f9212, 0f41355DC0; mov.f32 %f9221, 0f41E6BD60; fma.rn.ftz.f32 %f9222, %f9220, %f9212, %f9221; mov.f32 %f9223, 0f419D92C8; fma.rn.ftz.f32 %f9224, %f9222, %f9212, %f9223; rcp.approx.ftz.f32 %f9225, %f9224; fma.rn.ftz.f32 %f9226, %f9219, %f9225, %f9211; sub.ftz.f32 %f9228, %f15501, %f9226; setp.gt.ftz.f32 %p1065, %f1561, %f1608; selp.f32 %f9229, %f9228, %f9226, %p1065; sub.ftz.f32 %f9231, %f7706, %f9229; selp.f32 %f9232, %f9231, %f9229, %p1064; mov.b32 %r1780, %f9232; or.b32 %r1781, %r517, %r1780; mov.b32 %f9233, %r1781; add.ftz.f32 %f9234, %f1608, %f1561; setp.le.ftz.f32 %p1066, %f9234, 0f7F800000; selp.f32 %f16018, %f9233, %f9234, %p1066; $L__BB106_1136: mov.f32 %f15502, 0f3FC90FDB; mov.f32 %f15481, 0f3E2AAB94; mov.f32 %f15480, 0f3D994929; mov.f32 %f15479, 0f3D3F841F; mov.f32 %f15478, 0f3C94D2E9; mov.f32 %f15477, 0f3D53F941; mov.f32 %f15476, 0f3F800000; sub.ftz.f32 %f9235, %f16018, %f9206; mul.ftz.f32 %f9236, %f1558, %f1558; fma.rn.ftz.f32 %f9237, %f1557, %f1557, %f9236; fma.rn.ftz.f32 %f9238, %f1559, %f1559, %f9237; sqrt.approx.ftz.f32 %f9239, %f9238; div.approx.ftz.f32 %f9240, %f1559, %f9239; abs.ftz.f32 %f9241, %f9240; sub.ftz.f32 %f9243, %f15476, %f9241; mul.ftz.f32 %f9244, %f9243, 0f3F000000; sqrt.approx.ftz.f32 %f9245, %f9244; setp.gt.ftz.f32 %p1068, %f9241, 0f3F11EB85; selp.f32 %f9246, %f9245, %f9241, %p1068; mul.ftz.f32 %f9247, %f9246, %f9246; fma.rn.ftz.f32 %f9250, %f15477, %f9247, %f15478; fma.rn.ftz.f32 %f9252, %f9250, %f9247, %f15479; fma.rn.ftz.f32 %f9254, %f9252, %f9247, %f15480; fma.rn.ftz.f32 %f9256, %f9254, %f9247, %f15481; mul.ftz.f32 %f9257, %f9247, %f9256; fma.rn.ftz.f32 %f9258, %f9257, %f9246, %f9246; add.ftz.f32 %f9259, %f9258, %f9258; sub.ftz.f32 %f9261, %f15502, %f9258; selp.f32 %f9262, %f9259, %f9261, %p1068; setp.lt.ftz.f32 %p1069, %f9240, 0f00000000; sub.ftz.f32 %f9264, %f7706, %f9262; selp.f32 %f9265, %f9264, %f9262, %p1069; sub.ftz.f32 %f9266, %f9265, %f9208; div.approx.ftz.f32 %f16021, %f9266, %f9207; div.approx.ftz.f32 %f16020, %f9235, %f9205; $L__BB106_1143: ld.const.f32 %f15495, [kernel_params+288]; ld.const.f32 %f15494, [kernel_params+284]; mul.ftz.f32 %f16024, %f16020, %f15494; mul.ftz.f32 %f16025, %f16021, %f15495; mul.ftz.f32 %f16022, %f15998, %f15494; mul.ftz.f32 %f16023, %f15999, %f15495; mul.ftz.f32 %f16027, %f16010, %f15495; mul.ftz.f32 %f16026, %f16009, %f15494; $L__BB106_1144: ld.const.s32 %rd1555, [kernel_params+1288]; add.s64 %rd1556, %rd166, %rd1555; cvta.to.global.u64 %rd1557, %rd618; shl.b64 %rd1558, %rd1556, 2; add.s64 %rd1559, %rd1557, %rd1558; add.s64 %rd1560, %rd1559, 4; add.s64 %rd1561, %rd1559, 8; add.s64 %rd1562, %rd1559, 12; sub.ftz.f32 %f9303, %f16026, %f16022; atom.global.add.f32 %f9304, [%rd1559], %f9303; sub.ftz.f32 %f9305, %f16027, %f16023; atom.global.add.f32 %f9306, [%rd1560], %f9305; sub.ftz.f32 %f9307, %f16022, %f16024; atom.global.add.f32 %f9308, [%rd1561], %f9307; sub.ftz.f32 %f9309, %f16023, %f16025; atom.global.add.f32 %f9310, [%rd1562], %f9309; ld.const.s32 %rd1563, [kernel_params+1292]; add.s64 %rd1564, %rd166, %rd1563; shl.b64 %rd1565, %rd1564, 2; add.s64 %rd1566, %rd1557, %rd1565; atom.global.add.f32 %f9311, [%rd1566], 0f3F800000; $L__BB106_1145: ld.global.u32 %r1794, [%rd29]; or.b32 %r1795, %r1794, 8192; st.global.u32 [%rd29], %r1795; $L__BB106_1146: ld.const.u32 %r518, [kernel_params+1388]; setp.eq.s32 %p1077, %r518, 0; @%p1077 bra $L__BB106_1184; ld.const.u64 %rd1567, [kernel_params+2584]; cvta.to.global.u64 %rd1568, %rd1567; mul.lo.s64 %rd1569, %rd2, 12; add.s64 %rd1570, %rd1568, %rd1569; ld.global.f32 %f9315, [%rd1570+4]; ld.global.f32 %f9316, [%rd1570]; add.ftz.f32 %f9317, %f9316, %f9315; ld.global.f32 %f9318, [%rd1570+8]; add.ftz.f32 %f1638, %f9318, %f9317; ld.local.u32 %r519, [%rd7+-36]; and.b32 %r1796, %r519, 524288; setp.ne.s32 %p1078, %r1796, 0; mov.f32 %f9314, 0f3F800000; mov.f32 %f16028, %f9314; mov.f32 %f16029, %f9314; mov.f32 %f16030, %f9314; @%p1078 bra $L__BB106_1150; and.b32 %r1797, %r519, 512; setp.eq.s32 %p1079, %r1797, 0; mov.f32 %f16028, 0f00000000; mov.f32 %f16029, %f16028; mov.f32 %f16030, %f16028; @%p1079 bra $L__BB106_1150; ld.local.v2.f32 {%f16029, %f16030}, [%rd7+208]; ld.local.f32 %f16028, [%rd7+204]; $L__BB106_1150: add.ftz.f32 %f9324, %f16028, %f16029; add.ftz.f32 %f9325, %f16030, %f9324; mul.ftz.f32 %f9326, %f9325, 0f3EAAAAAB; sub.ftz.f32 %f9328, %f9314, %f9326; mul.ftz.f32 %f9329, %f1638, 0f3EAAAAAB; mul.ftz.f32 %f1645, %f9329, %f9328; setp.leu.ftz.f32 %p1080, %f1645, 0f00000000; @%p1080 bra $L__BB106_1184; ld.const.s32 %rd1571, [kernel_params+1396]; add.s64 %rd1572, %rd166, %rd1571; cvta.to.global.u64 %rd1573, %rd618; shl.b64 %rd1574, %rd1572, 2; add.s64 %rd2983, %rd1573, %rd1574; and.b32 %r1798, %r518, 1; setp.eq.b32 %p1081, %r1798, 1; mov.pred %p1082, 0; xor.pred %p1083, %p1081, %p1082; not.pred %p1084, %p1083; @%p1084 bra $L__BB106_1163; ld.local.u32 %r520, [%rd7+-12]; setp.eq.s32 %p1085, %r520, -1; mov.f32 %f1647, 0f00000000; mov.u32 %r522, 0; @%p1085 bra $L__BB106_1154; ld.const.u64 %rd1575, [kernel_params+2160]; cvta.to.global.u64 %rd1576, %rd1575; mul.wide.s32 %rd1577, %r520, 208; add.s64 %rd1578, %rd1576, %rd1577; ld.global.f32 %f1647, [%rd1578+172]; mov.b32 %r522, %f1647; $L__BB106_1154: ld.const.u32 %r523, [kernel_params+1392]; shl.b32 %r524, %r523, 1; setp.lt.s32 %p1086, %r524, 1; setp.eq.ftz.f32 %p1087, %f1645, 0f00000000; or.pred %p1088, %p1087, %p1086; @%p1088 bra $L__BB106_1162; add.s32 %r525, %r524, -1; mov.u32 %r526, 0; $L__BB106_1156: mul.wide.s32 %rd1579, %r526, 8; add.s64 %rd255, %rd2983, %rd1579; ld.global.f32 %f1648, [%rd255]; setp.eq.ftz.f32 %p1089, %f1648, 0f00000000; @%p1089 bra $L__BB106_1159; bra.uni $L__BB106_1157; $L__BB106_1159: shl.b32 %r1801, %r526, 1; mul.wide.s32 %rd1581, %r1801, 4; add.s64 %rd1582, %rd2983, %rd1581; mov.u32 %r1802, 0; atom.global.cas.b32 %r1803, [%rd1582], %r1802, %r522; mov.b32 %f9332, %r1803; setp.neu.ftz.f32 %p1093, %f9332, 0f00000000; setp.neu.ftz.f32 %p1094, %f1647, %f9332; and.pred %p1095, %p1093, %p1094; @%p1095 bra $L__BB106_1161; bra.uni $L__BB106_1160; $L__BB106_1157: setp.neu.ftz.f32 %p1090, %f1648, %f1647; setp.ne.s32 %p1091, %r526, %r525; and.pred %p1092, %p1091, %p1090; @%p1092 bra $L__BB106_1161; bra.uni $L__BB106_1158; $L__BB106_1161: add.s32 %r526, %r526, 1; setp.lt.s32 %p1096, %r526, %r524; @%p1096 bra $L__BB106_1156; bra.uni $L__BB106_1162; $L__BB106_1160: add.s64 %rd1583, %rd255, 4; atom.global.add.f32 %f9333, [%rd1583], %f1645; bra.uni $L__BB106_1162; $L__BB106_1158: add.s64 %rd1580, %rd255, 4; atom.global.add.f32 %f9331, [%rd1580], %f1645; $L__BB106_1162: mul.wide.s32 %rd1584, %r523, 4; shl.b64 %rd1585, %rd1584, 2; add.s64 %rd2983, %rd2983, %rd1585; $L__BB106_1163: and.b32 %r1804, %r518, 2; setp.eq.s32 %p1097, %r1804, 0; @%p1097 bra $L__BB106_1173; ld.const.u64 %rd2918, [kernel_params+2408]; cvta.to.global.u64 %rd2917, %rd2918; ld.local.u32 %rd1586, [%rd7+-40]; shl.b64 %rd1587, %rd1586, 5; and.b64 %rd1588, %rd1587, 134217696; add.s64 %rd1589, %rd2917, %rd1588; ld.global.f32 %f1649, [%rd1589+12]; ld.const.u32 %r528, [kernel_params+1392]; shl.b32 %r529, %r528, 1; setp.lt.s32 %p1098, %r529, 1; setp.eq.ftz.f32 %p1099, %f1645, 0f00000000; or.pred %p1100, %p1099, %p1098; @%p1100 bra $L__BB106_1172; mov.b32 %r530, %f1649; add.s32 %r531, %r529, -1; mov.u32 %r532, 0; $L__BB106_1166: mul.wide.s32 %rd1590, %r532, 8; add.s64 %rd258, %rd2983, %rd1590; ld.global.f32 %f1650, [%rd258]; setp.eq.ftz.f32 %p1101, %f1650, 0f00000000; @%p1101 bra $L__BB106_1169; bra.uni $L__BB106_1167; $L__BB106_1169: mov.u32 %r1806, 0; atom.global.cas.b32 %r1807, [%rd258], %r1806, %r530; mov.b32 %f9335, %r1807; setp.neu.ftz.f32 %p1105, %f9335, 0f00000000; setp.neu.ftz.f32 %p1106, %f1649, %f9335; and.pred %p1107, %p1105, %p1106; @%p1107 bra $L__BB106_1171; bra.uni $L__BB106_1170; $L__BB106_1167: setp.neu.ftz.f32 %p1102, %f1650, %f1649; setp.ne.s32 %p1103, %r532, %r531; and.pred %p1104, %p1103, %p1102; @%p1104 bra $L__BB106_1171; bra.uni $L__BB106_1168; $L__BB106_1171: add.s32 %r532, %r532, 1; setp.lt.s32 %p1108, %r532, %r529; @%p1108 bra $L__BB106_1166; bra.uni $L__BB106_1172; $L__BB106_1170: add.s64 %rd1592, %rd258, 4; atom.global.add.f32 %f9336, [%rd1592], %f1645; bra.uni $L__BB106_1172; $L__BB106_1168: add.s64 %rd1591, %rd258, 4; atom.global.add.f32 %f9334, [%rd1591], %f1645; $L__BB106_1172: mul.wide.s32 %rd1593, %r528, 4; shl.b64 %rd1594, %rd1593, 2; add.s64 %rd2983, %rd2983, %rd1594; $L__BB106_1173: and.b32 %r1808, %r518, 4; setp.eq.s32 %p1109, %r1808, 0; @%p1109 bra $L__BB106_1184; ld.local.u32 %r534, [%rd7+-12]; setp.eq.s32 %p1110, %r534, -1; mov.f32 %f1652, 0f00000000; mov.u32 %r536, 0; @%p1110 bra $L__BB106_1176; ld.const.u64 %rd1595, [kernel_params+2160]; cvta.to.global.u64 %rd1596, %rd1595; mul.wide.s32 %rd1597, %r534, 208; add.s64 %rd1598, %rd1596, %rd1597; ld.global.f32 %f1652, [%rd1598+176]; mov.b32 %r536, %f1652; $L__BB106_1176: ld.const.u32 %r1810, [kernel_params+1392]; shl.b32 %r537, %r1810, 1; setp.lt.s32 %p1111, %r537, 1; setp.eq.ftz.f32 %p1112, %f1645, 0f00000000; or.pred %p1113, %p1112, %p1111; @%p1113 bra $L__BB106_1184; add.s32 %r538, %r537, -1; mov.u32 %r539, 0; $L__BB106_1178: mul.wide.s32 %rd1599, %r539, 8; add.s64 %rd261, %rd2983, %rd1599; ld.global.f32 %f1653, [%rd261]; setp.eq.ftz.f32 %p1114, %f1653, 0f00000000; @%p1114 bra $L__BB106_1181; bra.uni $L__BB106_1179; $L__BB106_1181: mov.u32 %r1812, 0; atom.global.cas.b32 %r1813, [%rd261], %r1812, %r536; mov.b32 %f9339, %r1813; setp.neu.ftz.f32 %p1118, %f9339, 0f00000000; setp.neu.ftz.f32 %p1119, %f1652, %f9339; and.pred %p1120, %p1118, %p1119; @%p1120 bra $L__BB106_1183; bra.uni $L__BB106_1182; $L__BB106_1179: setp.neu.ftz.f32 %p1115, %f1653, %f1652; setp.ne.s32 %p1116, %r539, %r538; and.pred %p1117, %p1116, %p1115; @%p1117 bra $L__BB106_1183; bra.uni $L__BB106_1180; $L__BB106_1183: add.s32 %r539, %r539, 1; setp.lt.s32 %p1121, %r539, %r537; @%p1121 bra $L__BB106_1178; bra.uni $L__BB106_1184; $L__BB106_1182: add.s64 %rd1601, %rd261, 4; atom.global.add.f32 %f9340, [%rd1601], %f1645; bra.uni $L__BB106_1184; $L__BB106_1180: add.s64 %rd1600, %rd261, 4; atom.global.add.f32 %f9338, [%rd1600], %f1645; $L__BB106_1184: ld.const.u32 %r3052, [kernel_params+1256]; and.b32 %r1814, %r3052, 16384; setp.eq.s32 %p1122, %r1814, 0; @%p1122 bra $L__BB106_1207; ld.const.u64 %rd1602, [kernel_params+2584]; cvta.to.global.u64 %rd1603, %rd1602; mul.lo.s64 %rd1604, %rd2, 12; add.s64 %rd1605, %rd1603, %rd1604; ld.global.f32 %f1654, [%rd1605]; ld.global.f32 %f1655, [%rd1605+4]; ld.global.f32 %f1656, [%rd1605+8]; ld.const.u32 %r541, [kernel_params+1308]; ld.local.u32 %r542, [%rd7+172]; setp.lt.s32 %p1123, %r542, 1; mov.f32 %f16035, 0f00000000; mov.f32 %f16034, %f16035; mov.f32 %f16033, %f16035; @%p1123 bra $L__BB106_1206; add.s32 %r1816, %r542, -1; and.b32 %r3210, %r542, 3; setp.lt.u32 %p1124, %r1816, 3; mov.f32 %f16033, 0f00000000; mov.u32 %r3208, 0; mov.f32 %f16034, %f16033; mov.f32 %f16035, %f16033; @%p1124 bra $L__BB106_1201; sub.s32 %r3207, %r542, %r3210; mov.f32 %f16033, 0f00000000; mov.u32 %r3208, 0; mov.f32 %f16034, %f16033; mov.f32 %f16035, %f16033; $L__BB106_1188: cvt.s64.s32 %rd262, %r3208; mul.wide.s32 %rd1606, %r3208, 80; add.s64 %rd1607, %rd1, %rd1606; add.s64 %rd263, %rd1607, 316; ld.local.u32 %r547, [%rd1607+316]; add.s32 %r1818, %r547, -2; setp.lt.u32 %p1125, %r1818, 7; @%p1125 bra $L__BB106_1190; add.s32 %r1819, %r547, -32; setp.gt.u32 %p1126, %r1819, 2; @%p1126 bra $L__BB106_1191; $L__BB106_1190: ld.local.v4.f32 {%f9351, %f9352, %f9353, %f9354}, [%rd263+-12]; add.ftz.f32 %f16035, %f16035, %f9351; add.ftz.f32 %f16034, %f16034, %f9352; add.ftz.f32 %f16033, %f16033, %f9353; $L__BB106_1191: ld.local.u32 %r548, [%rd263+80]; add.s32 %r1820, %r548, -2; setp.lt.u32 %p1127, %r1820, 7; @%p1127 bra $L__BB106_1193; add.s32 %r1821, %r548, -32; setp.gt.u32 %p1128, %r1821, 2; @%p1128 bra $L__BB106_1194; $L__BB106_1193: ld.local.v4.f32 {%f9358, %f9359, %f9360, %f9361}, [%rd263+68]; add.ftz.f32 %f16035, %f16035, %f9358; add.ftz.f32 %f16034, %f16034, %f9359; add.ftz.f32 %f16033, %f16033, %f9360; $L__BB106_1194: ld.local.u32 %r549, [%rd263+160]; add.s32 %r1822, %r549, -2; setp.lt.u32 %p1129, %r1822, 7; @%p1129 bra $L__BB106_1196; add.s32 %r1823, %r549, -32; setp.gt.u32 %p1130, %r1823, 2; @%p1130 bra $L__BB106_1197; $L__BB106_1196: ld.local.v4.f32 {%f9365, %f9366, %f9367, %f9368}, [%rd263+148]; add.ftz.f32 %f16035, %f16035, %f9365; add.ftz.f32 %f16034, %f16034, %f9366; add.ftz.f32 %f16033, %f16033, %f9367; $L__BB106_1197: ld.local.u32 %r550, [%rd263+240]; add.s32 %r1824, %r550, -2; setp.lt.u32 %p1131, %r1824, 7; @%p1131 bra $L__BB106_1199; add.s32 %r1825, %r550, -32; setp.gt.u32 %p1132, %r1825, 2; @%p1132 bra $L__BB106_1200; $L__BB106_1199: ld.local.v4.f32 {%f9372, %f9373, %f9374, %f9375}, [%rd263+228]; add.ftz.f32 %f16035, %f16035, %f9372; add.ftz.f32 %f16034, %f16034, %f9373; add.ftz.f32 %f16033, %f16033, %f9374; $L__BB106_1200: cvt.u32.u64 %r1826, %rd262; add.s32 %r3208, %r1826, 4; add.s32 %r3207, %r3207, -4; setp.ne.s32 %p1133, %r3207, 0; @%p1133 bra $L__BB106_1188; $L__BB106_1201: setp.eq.s32 %p1134, %r3210, 0; @%p1134 bra $L__BB106_1206; $L__BB106_1202: .pragma "nounroll"; cvt.s64.s32 %rd264, %r3208; mul.wide.s32 %rd1608, %r3208, 80; add.s64 %rd1609, %rd1, %rd1608; add.s64 %rd265, %rd1609, 316; ld.local.u32 %r556, [%rd1609+316]; add.s32 %r1827, %r556, -2; setp.lt.u32 %p1135, %r1827, 7; @%p1135 bra $L__BB106_1204; add.s32 %r1828, %r556, -32; setp.gt.u32 %p1136, %r1828, 2; @%p1136 bra $L__BB106_1205; $L__BB106_1204: ld.local.f32 %f9379, [%rd265+-12]; add.ftz.f32 %f16035, %f16035, %f9379; ld.local.f32 %f9380, [%rd265+-8]; add.ftz.f32 %f16034, %f16034, %f9380; ld.local.f32 %f9381, [%rd265+-4]; add.ftz.f32 %f16033, %f16033, %f9381; $L__BB106_1205: cvt.u32.u64 %r1829, %rd264; add.s32 %r3208, %r1829, 1; add.s32 %r3210, %r3210, -1; setp.ne.s32 %p1137, %r3210, 0; @%p1137 bra $L__BB106_1202; $L__BB106_1206: cvta.to.global.u64 %rd1610, %rd618; cvt.s64.s32 %rd1611, %r541; add.s64 %rd1612, %rd166, %rd1611; shl.b64 %rd1613, %rd1612, 2; add.s64 %rd1614, %rd1610, %rd1613; mul.ftz.f32 %f9382, %f1654, %f16035; atom.global.add.f32 %f9383, [%rd1614], %f9382; add.s64 %rd1615, %rd1614, 4; mul.ftz.f32 %f9384, %f1655, %f16034; atom.global.add.f32 %f9385, [%rd1615], %f9384; add.s64 %rd1616, %rd1614, 8; mul.ftz.f32 %f9386, %f1656, %f16033; atom.global.add.f32 %f9387, [%rd1616], %f9386; $L__BB106_1207: ld.const.u32 %r3053, [kernel_params+1256]; cvt.u16.u32 %rs80, %r3053; setp.gt.s16 %p1138, %rs80, -1; @%p1138 bra $L__BB106_1225; ld.const.u64 %rd1617, [kernel_params+2584]; cvta.to.global.u64 %rd1618, %rd1617; mul.lo.s64 %rd1619, %rd2, 12; add.s64 %rd1620, %rd1618, %rd1619; ld.global.f32 %f1702, [%rd1620]; ld.global.f32 %f1703, [%rd1620+4]; ld.global.f32 %f1704, [%rd1620+8]; ld.const.u32 %r559, [kernel_params+1312]; ld.local.u32 %r560, [%rd7+172]; setp.lt.s32 %p1139, %r560, 1; mov.f32 %f16066, 0f00000000; mov.f32 %f16067, %f16066; mov.f32 %f16068, %f16066; @%p1139 bra $L__BB106_1224; add.s32 %r1831, %r560, -1; and.b32 %r3215, %r560, 3; setp.lt.u32 %p1140, %r1831, 3; mov.f32 %f16068, 0f00000000; mov.u32 %r3213, 0; mov.f32 %f16067, %f16068; mov.f32 %f16066, %f16068; @%p1140 bra $L__BB106_1220; sub.s32 %r3212, %r560, %r3215; mov.f32 %f16068, 0f00000000; mov.u32 %r3213, 0; mov.f32 %f16067, %f16068; mov.f32 %f16066, %f16068; $L__BB106_1211: cvt.s64.s32 %rd266, %r3213; mul.wide.s32 %rd1621, %r3213, 80; add.s64 %rd1622, %rd1, %rd1621; add.s64 %rd267, %rd1622, 316; ld.local.u32 %r1833, [%rd1622+316]; add.s32 %r1834, %r1833, -9; setp.gt.u32 %p1141, %r1834, 11; setp.ne.s32 %p1142, %r1833, 29; and.pred %p1143, %p1141, %p1142; @%p1143 bra $L__BB106_1213; ld.local.v4.f32 {%f9398, %f9399, %f9400, %f9401}, [%rd267+-12]; add.ftz.f32 %f16066, %f16066, %f9398; add.ftz.f32 %f16067, %f16067, %f9399; add.ftz.f32 %f16068, %f16068, %f9400; $L__BB106_1213: ld.local.u32 %r1835, [%rd267+80]; add.s32 %r1836, %r1835, -9; setp.gt.u32 %p1144, %r1836, 11; setp.ne.s32 %p1145, %r1835, 29; and.pred %p1146, %p1144, %p1145; @%p1146 bra $L__BB106_1215; ld.local.v4.f32 {%f9405, %f9406, %f9407, %f9408}, [%rd267+68]; add.ftz.f32 %f16066, %f16066, %f9405; add.ftz.f32 %f16067, %f16067, %f9406; add.ftz.f32 %f16068, %f16068, %f9407; $L__BB106_1215: ld.local.u32 %r1837, [%rd267+160]; add.s32 %r1838, %r1837, -9; setp.gt.u32 %p1147, %r1838, 11; setp.ne.s32 %p1148, %r1837, 29; and.pred %p1149, %p1147, %p1148; @%p1149 bra $L__BB106_1217; ld.local.v4.f32 {%f9412, %f9413, %f9414, %f9415}, [%rd267+148]; add.ftz.f32 %f16066, %f16066, %f9412; add.ftz.f32 %f16067, %f16067, %f9413; add.ftz.f32 %f16068, %f16068, %f9414; $L__BB106_1217: ld.local.u32 %r1839, [%rd267+240]; add.s32 %r1840, %r1839, -9; setp.gt.u32 %p1150, %r1840, 11; setp.ne.s32 %p1151, %r1839, 29; and.pred %p1152, %p1150, %p1151; @%p1152 bra $L__BB106_1219; ld.local.v4.f32 {%f9419, %f9420, %f9421, %f9422}, [%rd267+228]; add.ftz.f32 %f16066, %f16066, %f9419; add.ftz.f32 %f16067, %f16067, %f9420; add.ftz.f32 %f16068, %f16068, %f9421; $L__BB106_1219: cvt.u32.u64 %r1841, %rd266; add.s32 %r3213, %r1841, 4; add.s32 %r3212, %r3212, -4; setp.ne.s32 %p1153, %r3212, 0; @%p1153 bra $L__BB106_1211; $L__BB106_1220: setp.eq.s32 %p1154, %r3215, 0; @%p1154 bra $L__BB106_1224; $L__BB106_1221: .pragma "nounroll"; cvt.s64.s32 %rd268, %r3213; mul.wide.s32 %rd1623, %r3213, 80; add.s64 %rd1624, %rd1, %rd1623; add.s64 %rd269, %rd1624, 316; ld.local.u32 %r1842, [%rd1624+316]; add.s32 %r1843, %r1842, -9; setp.gt.u32 %p1155, %r1843, 11; setp.ne.s32 %p1156, %r1842, 29; and.pred %p1157, %p1155, %p1156; @%p1157 bra $L__BB106_1223; ld.local.f32 %f9426, [%rd269+-12]; add.ftz.f32 %f16066, %f16066, %f9426; ld.local.f32 %f9427, [%rd269+-8]; add.ftz.f32 %f16067, %f16067, %f9427; ld.local.f32 %f9428, [%rd269+-4]; add.ftz.f32 %f16068, %f16068, %f9428; $L__BB106_1223: cvt.u32.u64 %r1844, %rd268; add.s32 %r3213, %r1844, 1; add.s32 %r3215, %r3215, -1; setp.ne.s32 %p1158, %r3215, 0; @%p1158 bra $L__BB106_1221; $L__BB106_1224: cvta.to.global.u64 %rd1625, %rd618; cvt.s64.s32 %rd1626, %r559; add.s64 %rd1627, %rd166, %rd1626; shl.b64 %rd1628, %rd1627, 2; add.s64 %rd1629, %rd1625, %rd1628; mul.ftz.f32 %f9429, %f1702, %f16066; atom.global.add.f32 %f9430, [%rd1629], %f9429; add.s64 %rd1630, %rd1629, 4; mul.ftz.f32 %f9431, %f1703, %f16067; atom.global.add.f32 %f9432, [%rd1630], %f9431; add.s64 %rd1631, %rd1629, 8; mul.ftz.f32 %f9433, %f1704, %f16068; atom.global.add.f32 %f9434, [%rd1631], %f9433; $L__BB106_1225: ld.const.u32 %r3054, [kernel_params+1256]; and.b32 %r1845, %r3054, 65536; setp.eq.s32 %p1159, %r1845, 0; @%p1159 bra $L__BB106_1243; ld.const.u64 %rd1632, [kernel_params+2584]; cvta.to.global.u64 %rd1633, %rd1632; mul.lo.s64 %rd1634, %rd2, 12; add.s64 %rd1635, %rd1633, %rd1634; ld.global.f32 %f1750, [%rd1635]; ld.global.f32 %f1751, [%rd1635+4]; ld.global.f32 %f1752, [%rd1635+8]; ld.const.u32 %r572, [kernel_params+1316]; ld.local.u32 %r573, [%rd7+172]; setp.lt.s32 %p1160, %r573, 1; mov.f32 %f16096, 0f00000000; mov.f32 %f16097, %f16096; mov.f32 %f16098, %f16096; @%p1160 bra $L__BB106_1242; add.s32 %r1847, %r573, -1; and.b32 %r3220, %r573, 3; setp.lt.u32 %p1161, %r1847, 3; mov.f32 %f16098, 0f00000000; mov.u32 %r3218, 0; mov.f32 %f16097, %f16098; mov.f32 %f16096, %f16098; @%p1161 bra $L__BB106_1238; sub.s32 %r3217, %r573, %r3220; mov.f32 %f16098, 0f00000000; mov.u32 %r3218, 0; mov.f32 %f16097, %f16098; mov.f32 %f16096, %f16098; $L__BB106_1229: cvt.s64.s32 %rd270, %r3218; mul.wide.s32 %rd1636, %r3218, 80; add.s64 %rd1637, %rd1, %rd1636; add.s64 %rd271, %rd1637, 316; ld.local.u32 %r1849, [%rd1637+316]; add.s32 %r1850, %r1849, -21; setp.gt.u32 %p1162, %r1850, 9; @%p1162 bra $L__BB106_1231; ld.local.v4.f32 {%f9445, %f9446, %f9447, %f9448}, [%rd271+-12]; add.ftz.f32 %f16096, %f16096, %f9445; add.ftz.f32 %f16097, %f16097, %f9446; add.ftz.f32 %f16098, %f16098, %f9447; $L__BB106_1231: ld.local.u32 %r1851, [%rd271+80]; add.s32 %r1852, %r1851, -21; setp.gt.u32 %p1163, %r1852, 9; @%p1163 bra $L__BB106_1233; ld.local.v4.f32 {%f9452, %f9453, %f9454, %f9455}, [%rd271+68]; add.ftz.f32 %f16096, %f16096, %f9452; add.ftz.f32 %f16097, %f16097, %f9453; add.ftz.f32 %f16098, %f16098, %f9454; $L__BB106_1233: ld.local.u32 %r1853, [%rd271+160]; add.s32 %r1854, %r1853, -21; setp.gt.u32 %p1164, %r1854, 9; @%p1164 bra $L__BB106_1235; ld.local.v4.f32 {%f9459, %f9460, %f9461, %f9462}, [%rd271+148]; add.ftz.f32 %f16096, %f16096, %f9459; add.ftz.f32 %f16097, %f16097, %f9460; add.ftz.f32 %f16098, %f16098, %f9461; $L__BB106_1235: ld.local.u32 %r1855, [%rd271+240]; add.s32 %r1856, %r1855, -21; setp.gt.u32 %p1165, %r1856, 9; @%p1165 bra $L__BB106_1237; ld.local.v4.f32 {%f9466, %f9467, %f9468, %f9469}, [%rd271+228]; add.ftz.f32 %f16096, %f16096, %f9466; add.ftz.f32 %f16097, %f16097, %f9467; add.ftz.f32 %f16098, %f16098, %f9468; $L__BB106_1237: cvt.u32.u64 %r1857, %rd270; add.s32 %r3218, %r1857, 4; add.s32 %r3217, %r3217, -4; setp.ne.s32 %p1166, %r3217, 0; @%p1166 bra $L__BB106_1229; $L__BB106_1238: setp.eq.s32 %p1167, %r3220, 0; @%p1167 bra $L__BB106_1242; $L__BB106_1239: .pragma "nounroll"; cvt.s64.s32 %rd272, %r3218; mul.wide.s32 %rd1638, %r3218, 80; add.s64 %rd1639, %rd1, %rd1638; add.s64 %rd273, %rd1639, 316; ld.local.u32 %r1858, [%rd1639+316]; add.s32 %r1859, %r1858, -21; setp.gt.u32 %p1168, %r1859, 9; @%p1168 bra $L__BB106_1241; ld.local.f32 %f9473, [%rd273+-12]; add.ftz.f32 %f16096, %f16096, %f9473; ld.local.f32 %f9474, [%rd273+-8]; add.ftz.f32 %f16097, %f16097, %f9474; ld.local.f32 %f9475, [%rd273+-4]; add.ftz.f32 %f16098, %f16098, %f9475; $L__BB106_1241: cvt.u32.u64 %r1860, %rd272; add.s32 %r3218, %r1860, 1; add.s32 %r3220, %r3220, -1; setp.ne.s32 %p1169, %r3220, 0; @%p1169 bra $L__BB106_1239; $L__BB106_1242: cvta.to.global.u64 %rd1640, %rd618; cvt.s64.s32 %rd1641, %r572; add.s64 %rd1642, %rd166, %rd1641; shl.b64 %rd1643, %rd1642, 2; add.s64 %rd1644, %rd1640, %rd1643; mul.ftz.f32 %f9476, %f1750, %f16096; atom.global.add.f32 %f9477, [%rd1644], %f9476; add.s64 %rd1645, %rd1644, 4; mul.ftz.f32 %f9478, %f1751, %f16097; atom.global.add.f32 %f9479, [%rd1645], %f9478; add.s64 %rd1646, %rd1644, 8; mul.ftz.f32 %f9480, %f1752, %f16098; atom.global.add.f32 %f9481, [%rd1646], %f9480; $L__BB106_1243: ld.const.u32 %r3055, [kernel_params+1256]; and.b32 %r1861, %r3055, 131072; setp.eq.s32 %p1170, %r1861, 0; @%p1170 bra $L__BB106_1257; ld.const.f32 %f1798, [kernel_params+1412]; ld.const.f32 %f1799, [kernel_params+1416]; ld.local.v4.f32 {%f9482, %f9483, %f9484, %f9485}, [%rd1]; ld.const.v4.f32 {%f9489, %f9490, %f9491, %f9492}, [kernel_params+96]; ld.const.v4.f32 {%f9493, %f9494, %f9495, %f9496}, [kernel_params+112]; ld.const.v4.f32 {%f9497, %f9498, %f9499, %f9500}, [kernel_params+128]; ld.const.u32 %r1862, [kernel_params+16]; setp.eq.s32 %p1171, %r1862, 1; sub.ftz.f32 %f1812, %f9482, %f9492; sub.ftz.f32 %f1813, %f9483, %f9496; sub.ftz.f32 %f1814, %f9484, %f9500; @%p1171 bra $L__BB106_1246; bra.uni $L__BB106_1245; $L__BB106_1246: mul.ftz.f32 %f9504, %f9495, %f1813; fma.rn.ftz.f32 %f9505, %f9491, %f1812, %f9504; fma.rn.ftz.f32 %f9506, %f9499, %f1814, %f9505; abs.ftz.f32 %f16123, %f9506; bra.uni $L__BB106_1247; $L__BB106_1245: mul.ftz.f32 %f9501, %f1813, %f1813; fma.rn.ftz.f32 %f9502, %f1812, %f1812, %f9501; fma.rn.ftz.f32 %f9503, %f1814, %f1814, %f9502; sqrt.approx.ftz.f32 %f16123, %f9503; $L__BB106_1247: sub.ftz.f32 %f9507, %f16123, %f1798; mul.ftz.f32 %f9508, %f1799, %f9507; cvt.ftz.sat.f32.f32 %f16124, %f9508; ld.const.f32 %f1819, [kernel_params+1420]; setp.eq.ftz.f32 %p1172, %f1819, 0f3F800000; @%p1172 bra $L__BB106_1253; setp.eq.ftz.f32 %p1173, %f1819, 0f40000000; @%p1173 bra $L__BB106_1252; bra.uni $L__BB106_1249; $L__BB106_1252: mul.ftz.f32 %f16124, %f16124, %f16124; bra.uni $L__BB106_1253; $L__BB106_1249: setp.eq.ftz.f32 %p1174, %f1819, 0f3F000000; @%p1174 bra $L__BB106_1251; bra.uni $L__BB106_1250; $L__BB106_1251: sqrt.approx.ftz.f32 %f16124, %f16124; bra.uni $L__BB106_1253; $L__BB106_1250: lg2.approx.ftz.f32 %f9509, %f16124; mul.ftz.f32 %f9510, %f1819, %f9509; ex2.approx.ftz.f32 %f16124, %f9510; $L__BB106_1253: ld.const.u64 %rd1649, [kernel_params+2584]; cvta.to.global.u64 %rd1650, %rd1649; mul.lo.s64 %rd1651, %rd2, 12; add.s64 %rd1652, %rd1650, %rd1651; ld.global.f32 %f1824, [%rd1652]; ld.global.f32 %f1825, [%rd1652+4]; ld.global.f32 %f1826, [%rd1652+8]; ld.local.u32 %r585, [%rd7+-36]; and.b32 %r1863, %r585, 524288; setp.ne.s32 %p1175, %r1863, 0; mov.f32 %f9513, 0f3F800000; mov.f32 %f16125, %f9513; mov.f32 %f16126, %f9513; mov.f32 %f16127, %f9513; @%p1175 bra $L__BB106_1256; and.b32 %r1864, %r585, 512; setp.eq.s32 %p1176, %r1864, 0; mov.f32 %f16125, 0f00000000; mov.f32 %f16126, %f16125; mov.f32 %f16127, %f16125; @%p1176 bra $L__BB106_1256; ld.local.v2.f32 {%f16126, %f16127}, [%rd7+208]; ld.local.f32 %f16125, [%rd7+204]; $L__BB106_1256: sub.ftz.f32 %f9520, %f9513, %f16125; sub.ftz.f32 %f9521, %f9513, %f16126; sub.ftz.f32 %f9522, %f9513, %f16127; cvt.ftz.sat.f32.f32 %f9523, %f9520; cvt.ftz.sat.f32.f32 %f9524, %f9521; cvt.ftz.sat.f32.f32 %f9525, %f9522; mul.ftz.f32 %f9526, %f1825, %f9524; fma.rn.ftz.f32 %f9527, %f1824, %f9523, %f9526; fma.rn.ftz.f32 %f9528, %f1826, %f9525, %f9527; mul.ftz.f32 %f9529, %f9528, 0f3EAAAAAB; sub.ftz.f32 %f9530, %f9513, %f16124; mul.ftz.f32 %f9531, %f9530, %f9529; ld.const.s32 %rd1653, [kernel_params+1408]; add.s64 %rd1654, %rd166, %rd1653; cvta.to.global.u64 %rd1655, %rd618; shl.b64 %rd1656, %rd1654, 2; add.s64 %rd1657, %rd1655, %rd1656; atom.global.add.f32 %f9532, [%rd1657], %f9531; $L__BB106_1257: ld.global.u32 %r586, [%rd29]; and.b32 %r1865, %r586, 16777216; setp.eq.s32 %p1177, %r1865, 0; @%p1177 bra $L__BB106_1312; ld.local.u8 %rs81, [%rd7+-34]; and.b16 %rs82, %rs81, 8; setp.ne.s16 %p1178, %rs82, 0; @%p1178 bra $L__BB106_1312; ld.const.u64 %rd1658, [kernel_params+2448]; cvta.to.global.u64 %rd1659, %rd1658; add.s64 %rd1661, %rd1659, %rd837; ld.global.u32 %rd1662, [%rd1661]; ld.const.s32 %rd1663, [kernel_params+1264]; mul.lo.s64 %rd274, %rd1663, %rd1662; ld.const.u32 %r587, [kernel_params+1432]; setp.eq.s32 %p1179, %r587, -1; @%p1179 bra $L__BB106_1264; ld.const.u64 %rd1664, [kernel_params+2616]; cvta.to.global.u64 %rd1665, %rd1664; mul.lo.s64 %rd1666, %rd2, 12; add.s64 %rd1667, %rd1665, %rd1666; ld.const.u64 %rd1668, [kernel_params+2648]; cvta.to.global.u64 %rd1669, %rd1668; add.s64 %rd1671, %rd1669, %rd837; ld.global.f32 %f9534, [%rd1671]; ld.local.f32 %f9535, [%rd7]; sub.ftz.f32 %f9536, %f9535, %f9534; ld.global.f32 %f9537, [%rd1667+4]; ld.global.f32 %f9538, [%rd1667]; add.ftz.f32 %f9539, %f9538, %f9537; ld.global.f32 %f9540, [%rd1667+8]; add.ftz.f32 %f9541, %f9540, %f9539; mul.ftz.f32 %f9542, %f9541, 0f3EAAAAAB; mul.ftz.f32 %f1833, %f9542, %f9536; mov.b32 %r588, %f1833; setp.nan.ftz.f32 %p1180, %f1833, %f1833; mov.f32 %f16128, 0f00000000; @%p1180 bra $L__BB106_1263; or.b32 %r1866, %r588, -2147483648; setp.ne.s32 %p1181, %r1866, -2147483648; add.ftz.f32 %f9544, %f1833, %f1833; setp.eq.ftz.f32 %p1182, %f9544, %f1833; and.pred %p1183, %p1182, %p1181; @%p1183 bra $L__BB106_1263; shl.b32 %r1867, %r588, 1; setp.lt.u32 %p1184, %r1867, -16777215; selp.f32 %f16128, %f1833, 0f00000000, %p1184; $L__BB106_1263: cvta.to.global.u64 %rd1672, %rd618; cvt.s64.s32 %rd1673, %r587; add.s64 %rd1674, %rd274, %rd1673; shl.b64 %rd1675, %rd1674, 2; add.s64 %rd1676, %rd1672, %rd1675; atom.global.add.f32 %f9545, [%rd1676], %f16128; $L__BB106_1264: ld.local.u32 %r589, [%rd7+172]; setp.lt.s32 %p1185, %r589, 1; mov.f32 %f16144, 0f00000000; mov.f32 %f16145, %f16144; mov.f32 %f16146, %f16144; mov.f32 %f16147, %f16144; mov.f32 %f16148, %f16144; mov.f32 %f16149, %f16144; mov.f32 %f16150, %f16144; mov.f32 %f16151, %f16144; mov.f32 %f16152, %f16144; mov.f32 %f16153, %f16144; mov.f32 %f16154, %f16144; @%p1185 bra $L__BB106_1285; mov.f32 %f16154, 0f00000000; mov.u32 %r3221, 0; mov.f32 %f16153, %f16154; mov.f32 %f16152, %f16154; mov.f32 %f16151, %f16154; mov.f32 %f16150, %f16154; mov.f32 %f16149, %f16154; mov.f32 %f16148, %f16154; mov.f32 %f16147, %f16154; mov.f32 %f16146, %f16154; mov.f32 %f16145, %f16154; mov.f32 %f16144, %f16154; bra.uni $L__BB106_1266; $L__BB106_1278: add.s32 %r1878, %r591, -10; setp.lt.u32 %p1195, %r1878, 7; @%p1195 bra $L__BB106_1280; add.s32 %r1879, %r591, -22; setp.gt.u32 %p1196, %r1879, 2; setp.ne.s32 %p1197, %r591, 27; mov.f32 %f16143, 0f3F800000; and.pred %p1198, %p1196, %p1197; @%p1198 bra $L__BB106_1281; $L__BB106_1280: ld.local.v2.f32 {%f9610, %f9611}, [%rd276+20]; mul.ftz.f32 %f16143, %f9610, %f9611; bra.uni $L__BB106_1281; $L__BB106_1270: setp.eq.s32 %p1189, %r591, 6; @%p1189 bra $L__BB106_1275; bra.uni $L__BB106_1271; $L__BB106_1275: ld.local.f32 %f9607, [%rd276+20]; mul.ftz.f32 %f16140, %f16140, %f9607; mul.ftz.f32 %f16141, %f16141, %f9607; mul.ftz.f32 %f16142, %f16142, %f9607; bra.uni $L__BB106_1276; $L__BB106_1271: setp.eq.s32 %p1190, %r591, 29; @%p1190 bra $L__BB106_1274; bra.uni $L__BB106_1272; $L__BB106_1274: ld.local.v4.f32 {%f9582, %f9583, %f9584, %f9585}, [%rd276+20]; sqrt.approx.ftz.f32 %f9590, %f9582; sqrt.approx.ftz.f32 %f9591, %f9583; sqrt.approx.ftz.f32 %f9592, %f9584; fma.rn.ftz.f32 %f9593, %f9585, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f9594, %f9585, %f9593, 0fC12BAE14; fma.rn.ftz.f32 %f9595, %f9585, %f9594, 0f40220C4A; fma.rn.ftz.f32 %f9596, %f9585, %f9595, 0fBE5C28F6; fma.rn.ftz.f32 %f9597, %f9585, %f9596, 0f40BF020C; mul.ftz.f32 %f9598, %f9590, %f9597; mul.ftz.f32 %f9599, %f9591, %f9597; mul.ftz.f32 %f9600, %f9592, %f9597; mul.ftz.f32 %f9601, %f9598, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9602, %f9601; mul.ftz.f32 %f9603, %f9599, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9604, %f9603; mul.ftz.f32 %f9605, %f9600, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9606, %f9605; mul.ftz.f32 %f16140, %f16140, %f9602; mul.ftz.f32 %f16141, %f16141, %f9604; mul.ftz.f32 %f16142, %f16142, %f9606; bra.uni $L__BB106_1276; $L__BB106_1272: setp.ne.s32 %p1191, %r591, 5; @%p1191 bra $L__BB106_1276; ld.local.u32 %r1873, [%rd276+24]; setp.eq.s32 %p1192, %r1873, 8; @%p1192 bra $L__BB106_1284; bra.uni $L__BB106_1276; $L__BB106_1266: cvt.s64.s32 %rd275, %r3221; mul.wide.s32 %rd1677, %r3221, 80; add.s64 %rd1678, %rd1, %rd1677; add.s64 %rd276, %rd1678, 316; ld.local.u32 %r591, [%rd1678+316]; setp.gt.s32 %p1186, %r591, 34; @%p1186 bra $L__BB106_1284; ld.local.v4.f32 {%f9568, %f9569, %f9570, %f9571}, [%rd276+4]; fma.rn.ftz.f32 %f16150, %f9568, %f9569, %f16150; fma.rn.ftz.f32 %f16151, %f9568, %f9570, %f16151; fma.rn.ftz.f32 %f16152, %f9568, %f9571, %f16152; add.ftz.f32 %f16153, %f16153, %f9568; ld.local.v4.f32 {%f16140, %f16141, %f16142, %f9578}, [%rd276+-12]; setp.gt.u32 %p1187, %r591, 27; @%p1187 bra $L__BB106_1270; mov.u32 %r1870, 1; shl.b32 %r1871, %r1870, %r591; and.b32 %r1872, %r1871, 134256640; setp.ne.s32 %p1188, %r1872, 0; @%p1188 bra $L__BB106_1269; bra.uni $L__BB106_1270; $L__BB106_1269: ld.local.u64 %rd1679, [%rd276+36]; ld.f32 %f9579, [%rd1679+24]; mul.ftz.f32 %f16140, %f16140, %f9579; ld.f32 %f9580, [%rd1679+28]; mul.ftz.f32 %f16141, %f16141, %f9580; ld.f32 %f9581, [%rd1679+32]; mul.ftz.f32 %f16142, %f16142, %f9581; $L__BB106_1276: mov.f32 %f16143, 0f00000000; setp.gt.u32 %p1193, %r591, 31; @%p1193 bra $L__BB106_1278; mov.u32 %r1875, 1; shl.b32 %r1876, %r1875, %r591; and.b32 %r1877, %r1876, -2145385984; setp.ne.s32 %p1194, %r1877, 0; @%p1194 bra $L__BB106_1281; bra.uni $L__BB106_1278; $L__BB106_1281: setp.gt.ftz.f32 %p1199, %f16143, 0f3BB851EC; @%p1199 bra $L__BB106_1283; bra.uni $L__BB106_1282; $L__BB106_1283: add.ftz.f32 %f16147, %f16147, %f16140; add.ftz.f32 %f16148, %f16148, %f16141; add.ftz.f32 %f16149, %f16149, %f16142; add.ftz.f32 %f16154, %f16154, %f9568; bra.uni $L__BB106_1284; $L__BB106_1282: add.ftz.f32 %f16144, %f16144, %f16140; add.ftz.f32 %f16145, %f16145, %f16141; add.ftz.f32 %f16146, %f16146, %f16142; $L__BB106_1284: cvt.u32.u64 %r1880, %rd275; add.s32 %r3221, %r1880, 1; setp.lt.s32 %p1200, %r3221, %r589; @%p1200 bra $L__BB106_1266; $L__BB106_1285: mul.ftz.f32 %f9614, %f16154, 0f40800000; setp.gt.ftz.f32 %p1201, %f9614, %f16153; setp.eq.ftz.f32 %p1202, %f16153, 0f00000000; or.pred %p1203, %p1202, %p1201; @%p1203 bra $L__BB106_1287; bra.uni $L__BB106_1286; $L__BB106_1287: @%p1202 bra $L__BB106_1289; rcp.approx.ftz.f32 %f9621, %f16153; mul.ftz.f32 %f16150, %f16150, %f9621; mul.ftz.f32 %f16151, %f16151, %f9621; mul.ftz.f32 %f16152, %f16152, %f9621; $L__BB106_1289: ld.const.u32 %r593, [kernel_params+1424]; setp.eq.s32 %p1205, %r593, -1; @%p1205 bra $L__BB106_1300; ld.const.v4.f32 {%f9623, %f9624, %f9625, %f9626}, [kernel_params+688]; ld.const.v4.f32 {%f9630, %f9631, %f9632, %f9633}, [kernel_params+704]; ld.const.v4.f32 {%f9637, %f9638, %f9639, %f9640}, [kernel_params+720]; mul.ftz.f32 %f9644, %f16151, %f9624; fma.rn.ftz.f32 %f9645, %f16150, %f9623, %f9644; fma.rn.ftz.f32 %f1904, %f16152, %f9625, %f9645; mul.ftz.f32 %f9646, %f16151, %f9631; fma.rn.ftz.f32 %f9647, %f16150, %f9630, %f9646; fma.rn.ftz.f32 %f1905, %f16152, %f9632, %f9647; mul.ftz.f32 %f9648, %f16151, %f9638; fma.rn.ftz.f32 %f9649, %f16150, %f9637, %f9648; fma.rn.ftz.f32 %f1906, %f16152, %f9639, %f9649; mov.b32 %r594, %f1904; setp.nan.ftz.f32 %p1206, %f1904, %f1904; mov.f32 %f16171, 0f00000000; mov.f32 %f16169, %f16171; @%p1206 bra $L__BB106_1293; or.b32 %r1881, %r594, -2147483648; setp.ne.s32 %p1207, %r1881, -2147483648; add.ftz.f32 %f9651, %f1904, %f1904; setp.eq.ftz.f32 %p1208, %f9651, %f1904; and.pred %p1209, %p1208, %p1207; mov.f32 %f16169, %f16171; @%p1209 bra $L__BB106_1293; shl.b32 %r1882, %r594, 1; setp.lt.u32 %p1210, %r1882, -16777215; selp.f32 %f16169, %f1904, 0f00000000, %p1210; $L__BB106_1293: mov.b32 %r595, %f1905; setp.nan.ftz.f32 %p1211, %f1905, %f1905; mov.f32 %f16170, %f16171; @%p1211 bra $L__BB106_1296; or.b32 %r1883, %r595, -2147483648; setp.ne.s32 %p1212, %r1883, -2147483648; add.ftz.f32 %f9654, %f1905, %f1905; setp.eq.ftz.f32 %p1213, %f9654, %f1905; and.pred %p1214, %p1213, %p1212; mov.f32 %f16170, %f16171; @%p1214 bra $L__BB106_1296; shl.b32 %r1884, %r595, 1; setp.lt.u32 %p1215, %r1884, -16777215; selp.f32 %f16170, %f1905, 0f00000000, %p1215; $L__BB106_1296: mov.b32 %r596, %f1906; setp.nan.ftz.f32 %p1216, %f1906, %f1906; @%p1216 bra $L__BB106_1299; or.b32 %r1885, %r596, -2147483648; setp.ne.s32 %p1217, %r1885, -2147483648; add.ftz.f32 %f9657, %f1906, %f1906; setp.eq.ftz.f32 %p1218, %f9657, %f1906; and.pred %p1219, %p1218, %p1217; @%p1219 bra $L__BB106_1299; shl.b32 %r1886, %r596, 1; setp.lt.u32 %p1220, %r1886, -16777215; selp.f32 %f16171, %f1906, 0f00000000, %p1220; $L__BB106_1299: cvta.to.global.u64 %rd1684, %rd618; cvt.s64.s32 %rd1685, %r593; add.s64 %rd1686, %rd274, %rd1685; shl.b64 %rd1687, %rd1686, 2; add.s64 %rd1688, %rd1684, %rd1687; add.s64 %rd1689, %rd1688, 4; add.s64 %rd1690, %rd1688, 8; atom.global.add.f32 %f9658, [%rd1688], %f16169; atom.global.add.f32 %f9659, [%rd1689], %f16170; atom.global.add.f32 %f9660, [%rd1690], %f16171; $L__BB106_1300: ld.const.u32 %r597, [kernel_params+1428]; setp.eq.s32 %p1221, %r597, -1; @%p1221 bra $L__BB106_1311; ld.const.u64 %rd1691, [kernel_params+2616]; cvta.to.global.u64 %rd1692, %rd1691; mul.lo.s64 %rd1693, %rd2, 12; add.s64 %rd1694, %rd1692, %rd1693; ld.global.f32 %f9662, [%rd1694]; mul.ftz.f32 %f1913, %f16147, %f9662; ld.global.f32 %f9663, [%rd1694+4]; mul.ftz.f32 %f1914, %f16148, %f9663; ld.global.f32 %f9664, [%rd1694+8]; mul.ftz.f32 %f1915, %f16149, %f9664; mov.b32 %r598, %f1913; setp.nan.ftz.f32 %p1222, %f1913, %f1913; mov.f32 %f16174, 0f00000000; mov.f32 %f16172, %f16174; @%p1222 bra $L__BB106_1304; or.b32 %r1887, %r598, -2147483648; setp.ne.s32 %p1223, %r1887, -2147483648; add.ftz.f32 %f9666, %f1913, %f1913; setp.eq.ftz.f32 %p1224, %f9666, %f1913; and.pred %p1225, %p1224, %p1223; mov.f32 %f16172, %f16174; @%p1225 bra $L__BB106_1304; shl.b32 %r1888, %r598, 1; setp.lt.u32 %p1226, %r1888, -16777215; selp.f32 %f16172, %f1913, 0f00000000, %p1226; $L__BB106_1304: mov.b32 %r599, %f1914; setp.nan.ftz.f32 %p1227, %f1914, %f1914; mov.f32 %f16173, %f16174; @%p1227 bra $L__BB106_1307; or.b32 %r1889, %r599, -2147483648; setp.ne.s32 %p1228, %r1889, -2147483648; add.ftz.f32 %f9669, %f1914, %f1914; setp.eq.ftz.f32 %p1229, %f9669, %f1914; and.pred %p1230, %p1229, %p1228; mov.f32 %f16173, %f16174; @%p1230 bra $L__BB106_1307; shl.b32 %r1890, %r599, 1; setp.lt.u32 %p1231, %r1890, -16777215; selp.f32 %f16173, %f1914, 0f00000000, %p1231; $L__BB106_1307: mov.b32 %r600, %f1915; setp.nan.ftz.f32 %p1232, %f1915, %f1915; @%p1232 bra $L__BB106_1310; or.b32 %r1891, %r600, -2147483648; setp.ne.s32 %p1233, %r1891, -2147483648; add.ftz.f32 %f9672, %f1915, %f1915; setp.eq.ftz.f32 %p1234, %f9672, %f1915; and.pred %p1235, %p1234, %p1233; @%p1235 bra $L__BB106_1310; shl.b32 %r1892, %r600, 1; setp.lt.u32 %p1236, %r1892, -16777215; selp.f32 %f16174, %f1915, 0f00000000, %p1236; $L__BB106_1310: cvta.to.global.u64 %rd1695, %rd618; cvt.s64.s32 %rd1696, %r597; add.s64 %rd1697, %rd274, %rd1696; shl.b64 %rd1698, %rd1697, 2; add.s64 %rd1699, %rd1695, %rd1698; add.s64 %rd1700, %rd1699, 4; add.s64 %rd1701, %rd1699, 8; atom.global.add.f32 %f9673, [%rd1699], %f16172; atom.global.add.f32 %f9674, [%rd1700], %f16173; atom.global.add.f32 %f9675, [%rd1701], %f16174; $L__BB106_1311: and.b32 %r1893, %r586, -16777217; st.global.u32 [%rd29], %r1893; bra.uni $L__BB106_1312; $L__BB106_1286: ld.const.u64 %rd1680, [kernel_params+2616]; cvta.to.global.u64 %rd1681, %rd1680; mul.lo.s64 %rd1682, %rd2, 12; add.s64 %rd1683, %rd1681, %rd1682; ld.global.f32 %f9615, [%rd1683]; mul.ftz.f32 %f9616, %f16144, %f9615; ld.global.f32 %f9617, [%rd1683+4]; mul.ftz.f32 %f9618, %f16145, %f9617; ld.global.f32 %f9619, [%rd1683+8]; mul.ftz.f32 %f9620, %f16146, %f9619; st.global.f32 [%rd1683], %f9616; st.global.f32 [%rd1683+4], %f9618; st.global.f32 [%rd1683+8], %f9620; bra.uni $L__BB106_1312; $L__BB106_877: add.s32 %r1495, %r426, -10; setp.lt.u32 %p751, %r1495, 7; @%p751 bra $L__BB106_879; add.s32 %r1496, %r426, -22; setp.gt.u32 %p752, %r1496, 2; setp.ne.s32 %p753, %r426, 27; mov.f32 %f15941, 0f3F800000; and.pred %p754, %p752, %p753; @%p754 bra $L__BB106_882; $L__BB106_879: ld.local.v2.f32 {%f7029, %f7030}, [%rd173+20]; mul.ftz.f32 %f15941, %f7029, %f7030; bra.uni $L__BB106_882; } .func (.param .b32 func_retval0) _Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf( .param .b32 _Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf_param_0, .param .b64 _Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf_param_1 ) { .local .align 16 .b8 __local_depot107[8848]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<3933>; .reg .b16 %rs<749>; .reg .f32 %f<30956>; .reg .b32 %r<5894>; .reg .b64 %rd<4970>; mov.u64 %SPL, __local_depot107; cvta.local.u64 %SP, %SPL; ld.param.u32 %r1602, [_Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf_param_0]; ld.param.u64 %rd1059, [_Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf_param_1]; add.u64 %rd1, %SPL, 1520; add.u64 %rd1061, %SP, 3424; add.u64 %rd2, %SPL, 3424; ld.const.v2.u64 {%rd1062, %rd1063}, [kernel_params+2704]; mov.u32 %r1603, 0; cvta.to.global.u64 %rd1066, %rd1063; cvt.s64.s32 %rd3, %r1602; mul.wide.s32 %rd1067, %r1602, 4; add.s64 %rd1068, %rd1066, %rd1067; ld.global.u32 %r1604, [%rd1068]; cvt.s64.s32 %rd4, %r1604; ld.const.v2.u64 {%rd1069, %rd1070}, [kernel_params+2720]; cvta.to.global.u64 %rd1073, %rd1069; add.s64 %rd1074, %rd1073, %rd1067; cvta.to.global.u64 %rd1075, %rd1070; add.s64 %rd1076, %rd1075, %rd1067; ld.global.u32 %r1, [%rd1076]; ld.const.v2.u64 {%rd1077, %rd1078}, [kernel_params+2688]; cvta.to.global.u64 %rd1081, %rd1078; add.s64 %rd1082, %rd1081, %rd1067; ld.global.f32 %f1, [%rd1082]; cvta.to.global.u64 %rd1083, %rd1062; add.s64 %rd1084, %rd1083, %rd1067; ld.global.f32 %f2, [%rd1084]; cvta.to.global.u64 %rd1085, %rd1077; add.s64 %rd1086, %rd1085, %rd1067; ld.global.f32 %f28854, [%rd1086]; ld.const.u64 %rd1087, [kernel_params+2632]; cvta.to.global.u64 %rd1088, %rd1087; mul.wide.s32 %rd1089, %r1602, 12; add.s64 %rd1090, %rd1088, %rd1089; ld.global.f32 %f28850, [%rd1090]; ld.global.f32 %f28849, [%rd1090+4]; ld.global.f32 %f28848, [%rd1090+8]; ld.const.u64 %rd1091, [kernel_params+2640]; cvta.to.global.u64 %rd1092, %rd1091; add.s64 %rd1093, %rd1092, %rd1089; ld.global.f32 %f7, [%rd1093]; ld.global.f32 %f8, [%rd1093+4]; ld.global.f32 %f9, [%rd1093+8]; ld.const.u64 %rd1094, [kernel_params+2664]; cvta.to.global.u64 %rd1095, %rd1094; add.s64 %rd1096, %rd1095, %rd1067; ld.global.f32 %f10, [%rd1096]; ld.const.v2.u64 {%rd1097, %rd1098}, [kernel_params+2672]; cvta.to.global.u64 %rd1101, %rd1097; add.s64 %rd1102, %rd1101, %rd1067; ld.global.f32 %f11, [%rd1102]; cvta.to.global.u64 %rd1103, %rd1098; add.s64 %rd1104, %rd1103, %rd1067; ld.global.f32 %f12, [%rd1104]; add.s64 %rd5, %rd2, 88; ld.global.u32 %r1605, [%rd1074]; st.local.f32 [%rd2+88], %f28854; mov.b32 %r1606, %f2; mov.b32 %r1607, %f1; st.local.v4.u32 [%rd2+64], {%r1, %r1607, %r1606, %r1605}; ld.const.u64 %rd1105, [kernel_params+2184]; cvta.to.global.u64 %rd6, %rd1105; cvt.s64.s32 %rd7, %r1605; mul.wide.s32 %rd1106, %r1605, 4; add.s64 %rd1107, %rd6, %rd1106; ld.global.u32 %r5556, [%rd1107]; st.local.u32 [%rd2+56], %r5556; st.local.u32 [%rd2+60], %r1604; st.local.u32 [%rd2+52], %r1603; mov.b32 %r1608, %f10; mov.u32 %r5550, -1; st.local.v2.u32 [%rd2+80], {%r5550, %r1608}; and.b32 %r3, %r5556, 2; setp.eq.s32 %p24, %r3, 0; @%p24 bra $L__BB107_7; ld.const.u64 %rd1108, [kernel_params+2160]; cvta.to.global.u64 %rd1109, %rd1108; mul.lo.s64 %rd1110, %rd7, 208; add.s64 %rd1111, %rd1109, %rd1110; ld.const.u64 %rd1112, [kernel_params+2176]; cvta.to.global.u64 %rd1113, %rd1112; ld.global.u32 %rd1114, [%rd1111+168]; ld.global.u32 %r1610, [%rd1111+152]; shl.b32 %r1611, %r1610, 1; cvt.rn.f32.s32 %f7742, %r1611; mul.ftz.f32 %f7743, %f10, %f7742; cvt.rzi.ftz.s32.f32 %r1612, %f7743; add.s32 %r1613, %r1611, -1; min.s32 %r1614, %r1612, %r1613; cvt.rn.f32.s32 %f7744, %r1614; sub.ftz.f32 %f13, %f7743, %f7744; cvt.s64.s32 %rd1115, %r1614; add.s64 %rd1116, %rd1115, %rd1114; shl.b64 %rd1117, %rd1116, 6; add.s64 %rd8, %rd1113, %rd1117; ld.global.v4.f32 {%f7745, %f7746, %f7747, %f7748}, [%rd8]; ld.global.v4.f32 {%f7749, %f7750, %f7751, %f7752}, [%rd8+64]; mul.ftz.f32 %f7753, %f7746, %f7750; fma.rn.ftz.f32 %f7754, %f7745, %f7749, %f7753; mul.ftz.f32 %f7755, %f7748, %f7752; fma.rn.ftz.f32 %f7756, %f7747, %f7751, %f7755; add.ftz.f32 %f22, %f7754, %f7756; setp.gt.ftz.f32 %p25, %f22, 0f3F7FDF3B; mov.f32 %f7757, 0f3F800000; sub.ftz.f32 %f23, %f7757, %f13; @%p25 bra $L__BB107_3; bra.uni $L__BB107_2; $L__BB107_3: mul.ftz.f32 %f7811, %f7745, %f23; fma.rn.ftz.f32 %f7812, %f7749, %f13, %f7811; mul.ftz.f32 %f7813, %f7746, %f23; fma.rn.ftz.f32 %f7814, %f7750, %f13, %f7813; mul.ftz.f32 %f7815, %f7747, %f23; fma.rn.ftz.f32 %f7816, %f7751, %f13, %f7815; mul.ftz.f32 %f7817, %f7748, %f23; fma.rn.ftz.f32 %f7818, %f7752, %f13, %f7817; mul.ftz.f32 %f7819, %f7814, %f7814; fma.rn.ftz.f32 %f7820, %f7812, %f7812, %f7819; mul.ftz.f32 %f7821, %f7818, %f7818; fma.rn.ftz.f32 %f7822, %f7816, %f7816, %f7821; add.ftz.f32 %f7823, %f7820, %f7822; rsqrt.approx.ftz.f32 %f7824, %f7823; mul.ftz.f32 %f28780, %f7818, %f7824; mul.ftz.f32 %f28779, %f7816, %f7824; mul.ftz.f32 %f28778, %f7814, %f7824; mul.ftz.f32 %f28777, %f7812, %f7824; bra.uni $L__BB107_4; $L__BB107_2: mov.f32 %f28764, 0f3F800000; mov.f32 %f7758, 0fBF800000; max.ftz.f32 %f7759, %f22, %f7758; min.ftz.f32 %f7761, %f7759, %f28764; abs.ftz.f32 %f7762, %f7761; sub.ftz.f32 %f7763, %f28764, %f7762; mul.ftz.f32 %f7764, %f7763, 0f3F000000; sqrt.approx.ftz.f32 %f7765, %f7764; setp.gt.ftz.f32 %p26, %f7762, 0f3F11EB85; selp.f32 %f7766, %f7765, %f7762, %p26; mul.ftz.f32 %f7767, %f7766, %f7766; mov.f32 %f7768, 0f3C94D2E9; mov.f32 %f7769, 0f3D53F941; fma.rn.ftz.f32 %f7770, %f7769, %f7767, %f7768; mov.f32 %f7771, 0f3D3F841F; fma.rn.ftz.f32 %f7772, %f7770, %f7767, %f7771; mov.f32 %f7773, 0f3D994929; fma.rn.ftz.f32 %f7774, %f7772, %f7767, %f7773; mov.f32 %f7775, 0f3E2AAB94; fma.rn.ftz.f32 %f7776, %f7774, %f7767, %f7775; mul.ftz.f32 %f7777, %f7767, %f7776; fma.rn.ftz.f32 %f7778, %f7777, %f7766, %f7766; add.ftz.f32 %f7779, %f7778, %f7778; mov.f32 %f7780, 0f3FC90FDB; sub.ftz.f32 %f7781, %f7780, %f7778; selp.f32 %f7782, %f7779, %f7781, %p26; setp.lt.ftz.f32 %p27, %f7761, 0f00000000; mov.f32 %f7783, 0f40490FDB; sub.ftz.f32 %f7784, %f7783, %f7782; selp.f32 %f7785, %f7784, %f7782, %p27; mul.ftz.f32 %f7786, %f7745, %f22; sub.ftz.f32 %f7787, %f7749, %f7786; mul.ftz.f32 %f7788, %f7746, %f22; sub.ftz.f32 %f7789, %f7750, %f7788; mul.ftz.f32 %f7790, %f7747, %f22; sub.ftz.f32 %f7791, %f7751, %f7790; mul.ftz.f32 %f7792, %f7748, %f22; sub.ftz.f32 %f7793, %f7752, %f7792; mul.ftz.f32 %f7794, %f7789, %f7789; fma.rn.ftz.f32 %f7795, %f7787, %f7787, %f7794; mul.ftz.f32 %f7796, %f7793, %f7793; fma.rn.ftz.f32 %f7797, %f7791, %f7791, %f7796; add.ftz.f32 %f7798, %f7795, %f7797; rsqrt.approx.ftz.f32 %f7799, %f7798; mul.ftz.f32 %f7800, %f7787, %f7799; mul.ftz.f32 %f7801, %f7789, %f7799; mul.ftz.f32 %f7802, %f7791, %f7799; mul.ftz.f32 %f7803, %f7793, %f7799; mul.ftz.f32 %f7804, %f13, %f7785; cos.approx.ftz.f32 %f7805, %f7804; sin.approx.ftz.f32 %f7806, %f7804; mul.ftz.f32 %f7807, %f7800, %f7806; mul.ftz.f32 %f7808, %f7801, %f7806; mul.ftz.f32 %f7809, %f7802, %f7806; mul.ftz.f32 %f7810, %f7803, %f7806; fma.rn.ftz.f32 %f28780, %f7748, %f7805, %f7810; fma.rn.ftz.f32 %f28779, %f7747, %f7805, %f7809; fma.rn.ftz.f32 %f28778, %f7746, %f7805, %f7808; fma.rn.ftz.f32 %f28777, %f7745, %f7805, %f7807; $L__BB107_4: mov.f32 %f28767, 0f3F800000; sub.ftz.f32 %f28766, %f28767, %f13; mov.f32 %f28765, 0f3F800000; ld.global.v4.f32 {%f7825, %f7826, %f7827, %f7828}, [%rd8+16]; ld.global.v4.f32 {%f7833, %f7834, %f7835, %f7836}, [%rd8+80]; mul.ftz.f32 %f7841, %f13, %f7833; mul.ftz.f32 %f7842, %f13, %f7834; mul.ftz.f32 %f7843, %f13, %f7835; mul.ftz.f32 %f7844, %f13, %f7836; fma.rn.ftz.f32 %f144, %f28766, %f7825, %f7841; fma.rn.ftz.f32 %f148, %f28766, %f7826, %f7842; fma.rn.ftz.f32 %f152, %f28766, %f7827, %f7843; fma.rn.ftz.f32 %f7845, %f28766, %f7828, %f7844; ld.global.v4.f32 {%f7846, %f7847, %f7848, %f7849}, [%rd8+32]; ld.global.v4.f32 {%f7854, %f7855, %f7856, %f7857}, [%rd8+96]; mul.ftz.f32 %f7862, %f13, %f7854; mul.ftz.f32 %f7863, %f13, %f7855; mul.ftz.f32 %f7864, %f13, %f7856; mul.ftz.f32 %f7865, %f13, %f7857; fma.rn.ftz.f32 %f7866, %f28766, %f7846, %f7862; fma.rn.ftz.f32 %f7867, %f28766, %f7847, %f7863; fma.rn.ftz.f32 %f7868, %f28766, %f7848, %f7864; fma.rn.ftz.f32 %f7869, %f28766, %f7849, %f7865; ld.global.v4.f32 {%f7870, %f7871, %f7872, %f7873}, [%rd8+48]; ld.global.v4.f32 {%f7878, %f7879, %f7880, %f7881}, [%rd8+112]; mul.ftz.f32 %f7886, %f13, %f7878; mul.ftz.f32 %f7887, %f13, %f7879; mul.ftz.f32 %f7888, %f13, %f7880; mul.ftz.f32 %f7889, %f13, %f7881; fma.rn.ftz.f32 %f7890, %f28766, %f7870, %f7886; fma.rn.ftz.f32 %f7891, %f28766, %f7871, %f7887; fma.rn.ftz.f32 %f7892, %f28766, %f7872, %f7888; fma.rn.ftz.f32 %f7893, %f28766, %f7873, %f7889; mul.ftz.f32 %f7894, %f28780, 0f3FB504F3; mul.ftz.f32 %f7895, %f28777, 0f3FB504F3; mul.ftz.f32 %f7896, %f28778, 0f3FB504F3; mul.ftz.f32 %f7897, %f28779, 0f3FB504F3; mul.ftz.f32 %f7898, %f7894, %f7895; mul.ftz.f32 %f7899, %f7894, %f7896; mul.ftz.f32 %f7900, %f7894, %f7897; mul.ftz.f32 %f7901, %f7895, %f7895; mul.ftz.f32 %f7902, %f7895, %f7896; mul.ftz.f32 %f7903, %f7895, %f7897; mul.ftz.f32 %f7904, %f7896, %f7896; mul.ftz.f32 %f7905, %f7896, %f7897; mul.ftz.f32 %f7906, %f7897, %f7897; sub.ftz.f32 %f7908, %f28765, %f7904; sub.ftz.f32 %f7909, %f7908, %f7906; sub.ftz.f32 %f7910, %f7902, %f7900; add.ftz.f32 %f7911, %f7899, %f7903; add.ftz.f32 %f7912, %f7900, %f7902; sub.ftz.f32 %f7913, %f28765, %f7901; sub.ftz.f32 %f7914, %f7913, %f7906; sub.ftz.f32 %f7915, %f7905, %f7898; sub.ftz.f32 %f7916, %f7903, %f7899; add.ftz.f32 %f7917, %f7898, %f7905; sub.ftz.f32 %f7918, %f7913, %f7904; mul.ftz.f32 %f7919, %f7910, %f7868; fma.rn.ftz.f32 %f7920, %f7909, %f7845, %f7919; fma.rn.ftz.f32 %f141, %f7911, %f7891, %f7920; mul.ftz.f32 %f7921, %f7910, %f7869; fma.rn.ftz.f32 %f7922, %f7909, %f7866, %f7921; fma.rn.ftz.f32 %f142, %f7911, %f7892, %f7922; mul.ftz.f32 %f7923, %f7910, %f7890; fma.rn.ftz.f32 %f7924, %f7909, %f7867, %f7923; fma.rn.ftz.f32 %f143, %f7911, %f7893, %f7924; mul.ftz.f32 %f7925, %f7914, %f7868; fma.rn.ftz.f32 %f7926, %f7912, %f7845, %f7925; fma.rn.ftz.f32 %f145, %f7915, %f7891, %f7926; mul.ftz.f32 %f7927, %f7914, %f7869; fma.rn.ftz.f32 %f7928, %f7912, %f7866, %f7927; fma.rn.ftz.f32 %f146, %f7915, %f7892, %f7928; mul.ftz.f32 %f7929, %f7914, %f7890; fma.rn.ftz.f32 %f7930, %f7912, %f7867, %f7929; fma.rn.ftz.f32 %f147, %f7915, %f7893, %f7930; mul.ftz.f32 %f7931, %f7917, %f7868; fma.rn.ftz.f32 %f7932, %f7916, %f7845, %f7931; fma.rn.ftz.f32 %f149, %f7918, %f7891, %f7932; mul.ftz.f32 %f7933, %f7917, %f7869; fma.rn.ftz.f32 %f7934, %f7916, %f7866, %f7933; fma.rn.ftz.f32 %f150, %f7918, %f7892, %f7934; mul.ftz.f32 %f7935, %f7917, %f7890; fma.rn.ftz.f32 %f7936, %f7916, %f7867, %f7935; fma.rn.ftz.f32 %f151, %f7918, %f7893, %f7936; st.local.v4.f32 [%rd5+56], {%f141, %f142, %f143, %f144}; st.local.v4.f32 [%rd5+72], {%f145, %f146, %f147, %f148}; st.local.v4.f32 [%rd5+88], {%f149, %f150, %f151, %f152}; mul.ftz.f32 %f7937, %f146, %f151; mul.ftz.f32 %f60, %f150, %f147; sub.ftz.f32 %f28783, %f7937, %f60; mul.ftz.f32 %f62, %f150, %f143; mul.ftz.f32 %f7938, %f142, %f151; sub.ftz.f32 %f28782, %f62, %f7938; mul.ftz.f32 %f64, %f142, %f147; mul.ftz.f32 %f7939, %f146, %f143; sub.ftz.f32 %f28781, %f64, %f7939; mul.ftz.f32 %f7940, %f145, %f28782; fma.rn.ftz.f32 %f7941, %f141, %f28783, %f7940; fma.rn.ftz.f32 %f28787, %f149, %f28781, %f7941; setp.neu.ftz.f32 %p28, %f28787, 0f00000000; mov.f32 %f28784, %f141; mov.f32 %f28785, %f146; mov.f32 %f28786, %f151; @%p28 bra $L__BB107_6; add.ftz.f32 %f28784, %f141, 0f322BCC77; add.ftz.f32 %f28786, %f151, 0f322BCC77; add.ftz.f32 %f28785, %f146, 0f322BCC77; mul.ftz.f32 %f7942, %f28785, %f28786; sub.ftz.f32 %f28783, %f7942, %f60; mul.ftz.f32 %f7943, %f142, %f28786; sub.ftz.f32 %f28782, %f62, %f7943; mul.ftz.f32 %f7944, %f28785, %f143; sub.ftz.f32 %f28781, %f64, %f7944; mul.ftz.f32 %f7945, %f145, %f28782; fma.rn.ftz.f32 %f7946, %f28784, %f28783, %f7945; fma.rn.ftz.f32 %f7947, %f149, %f28781, %f7946; setp.eq.ftz.f32 %p29, %f7947, 0f00000000; selp.f32 %f28787, 0f7F7FFFFF, %f7947, %p29; $L__BB107_6: mul.ftz.f32 %f7948, %f145, %f28786; mul.ftz.f32 %f7949, %f149, %f147; sub.ftz.f32 %f7950, %f7949, %f7948; mul.ftz.f32 %f7951, %f149, %f143; mul.ftz.f32 %f7952, %f28784, %f28786; sub.ftz.f32 %f7953, %f7952, %f7951; mul.ftz.f32 %f7954, %f147, %f28784; mul.ftz.f32 %f7955, %f145, %f143; sub.ftz.f32 %f7956, %f7955, %f7954; mul.ftz.f32 %f7957, %f149, %f28785; mul.ftz.f32 %f7958, %f145, %f150; sub.ftz.f32 %f7959, %f7958, %f7957; mul.ftz.f32 %f7960, %f150, %f28784; mul.ftz.f32 %f7961, %f149, %f142; sub.ftz.f32 %f7962, %f7961, %f7960; mul.ftz.f32 %f7963, %f145, %f142; mul.ftz.f32 %f7964, %f28784, %f28785; sub.ftz.f32 %f7965, %f7964, %f7963; div.approx.ftz.f32 %f28822, %f28782, %f28787; mul.ftz.f32 %f7966, %f148, %f28822; div.approx.ftz.f32 %f28821, %f28783, %f28787; fma.rn.ftz.f32 %f7967, %f144, %f28821, %f7966; div.approx.ftz.f32 %f28823, %f28781, %f28787; fma.rn.ftz.f32 %f7968, %f152, %f28823, %f7967; div.approx.ftz.f32 %f28826, %f7953, %f28787; mul.ftz.f32 %f7969, %f148, %f28826; div.approx.ftz.f32 %f28825, %f7950, %f28787; fma.rn.ftz.f32 %f7970, %f144, %f28825, %f7969; div.approx.ftz.f32 %f28827, %f7956, %f28787; fma.rn.ftz.f32 %f7971, %f152, %f28827, %f7970; div.approx.ftz.f32 %f28830, %f7962, %f28787; mul.ftz.f32 %f7972, %f148, %f28830; div.approx.ftz.f32 %f28829, %f7959, %f28787; fma.rn.ftz.f32 %f7973, %f144, %f28829, %f7972; div.approx.ftz.f32 %f28831, %f7965, %f28787; fma.rn.ftz.f32 %f7974, %f152, %f28831, %f7973; neg.ftz.f32 %f28836, %f7968; neg.ftz.f32 %f28844, %f7974; neg.ftz.f32 %f28840, %f7971; st.local.v4.f32 [%rd5+104], {%f28821, %f28822, %f28823, %f28836}; st.local.v4.f32 [%rd5+120], {%f28825, %f28826, %f28827, %f28840}; st.local.v4.f32 [%rd5+136], {%f28829, %f28830, %f28831, %f28844}; $L__BB107_7: neg.ftz.f32 %f165, %f7; st.local.f32 [%rd5+-52], %f165; neg.ftz.f32 %f166, %f8; st.local.f32 [%rd5+-48], %f166; neg.ftz.f32 %f167, %f9; st.local.f32 [%rd5+-44], %f167; and.b32 %r1615, %r1, 6; setp.eq.s32 %p30, %r1615, 0; @%p30 bra $L__BB107_46; and.b32 %r4, %r5556, 4; setp.ne.s32 %p31, %r4, 0; mov.f32 %f28851, %f9; mov.f32 %f28852, %f8; mov.f32 %f28853, %f7; @%p31 bra $L__BB107_14; and.b32 %r5526, %r5556, 2; setp.ne.s32 %p32, %r5526, 0; mov.f32 %f28833, %f28821; mov.f32 %f28834, %f28822; mov.f32 %f28835, %f28823; mov.f32 %f28837, %f28825; mov.f32 %f28838, %f28826; mov.f32 %f28839, %f28827; mov.f32 %f28841, %f28829; mov.f32 %f28842, %f28830; mov.f32 %f28843, %f28831; @%p32 bra $L__BB107_11; ld.const.u64 %rd1118, [kernel_params+2160]; cvta.to.global.u64 %rd1119, %rd1118; mul.lo.s64 %rd1120, %rd7, 208; add.s64 %rd1121, %rd1119, %rd1120; ld.global.v4.f32 {%f28833, %f28834, %f28835, %f28836}, [%rd1121+48]; ld.global.v4.f32 {%f28837, %f28838, %f28839, %f28840}, [%rd1121+64]; ld.global.v4.f32 {%f28841, %f28842, %f28843, %f28844}, [%rd1121+80]; $L__BB107_11: mul.ftz.f32 %f7987, %f28849, %f28834; fma.rn.ftz.f32 %f7988, %f28850, %f28833, %f7987; fma.rn.ftz.f32 %f7989, %f28848, %f28835, %f7988; add.ftz.f32 %f192, %f28836, %f7989; mul.ftz.f32 %f7990, %f28849, %f28838; fma.rn.ftz.f32 %f7991, %f28850, %f28837, %f7990; fma.rn.ftz.f32 %f7992, %f28848, %f28839, %f7991; add.ftz.f32 %f193, %f28840, %f7992; mul.ftz.f32 %f7993, %f28849, %f28842; fma.rn.ftz.f32 %f7994, %f28850, %f28841, %f7993; fma.rn.ftz.f32 %f7995, %f28848, %f28843, %f7994; add.ftz.f32 %f28848, %f28844, %f7995; mul.ftz.f32 %f7996, %f28854, %f7; mul.ftz.f32 %f7997, %f28854, %f8; mul.ftz.f32 %f7998, %f7997, %f28834; fma.rn.ftz.f32 %f7999, %f7996, %f28833, %f7998; mul.ftz.f32 %f8000, %f28854, %f9; fma.rn.ftz.f32 %f28853, %f8000, %f28835, %f7999; mul.ftz.f32 %f8001, %f7997, %f28838; fma.rn.ftz.f32 %f8002, %f7996, %f28837, %f8001; fma.rn.ftz.f32 %f28852, %f8000, %f28839, %f8002; mul.ftz.f32 %f8003, %f7997, %f28842; fma.rn.ftz.f32 %f8004, %f7996, %f28841, %f8003; fma.rn.ftz.f32 %f28851, %f8000, %f28843, %f8004; mul.ftz.f32 %f8005, %f28852, %f28852; fma.rn.ftz.f32 %f8006, %f28853, %f28853, %f8005; fma.rn.ftz.f32 %f8007, %f28851, %f28851, %f8006; sqrt.approx.ftz.f32 %f28854, %f8007; setp.neu.ftz.f32 %p33, %f28854, 0f00000000; @%p33 bra $L__BB107_12; bra.uni $L__BB107_13; $L__BB107_12: div.approx.ftz.f32 %f28851, %f28851, %f28854; div.approx.ftz.f32 %f28852, %f28852, %f28854; div.approx.ftz.f32 %f28853, %f28853, %f28854; $L__BB107_13: mov.f32 %f28849, %f193; mov.f32 %f28850, %f192; $L__BB107_14: ld.const.u64 %rd1122, [kernel_params+2264]; cvta.to.global.u64 %rd1123, %rd1122; shl.b64 %rd1124, %rd4, 4; add.s64 %rd1125, %rd1123, %rd1124; add.s64 %rd9, %rd1125, 4; shr.s32 %r1616, %r1, 7; ld.global.u32 %r1617, [%rd1125+4]; add.s32 %r5, %r1616, %r1617; add.s32 %r1618, %r5, -1; max.s32 %r6, %r1618, %r1617; add.s32 %r1619, %r5, 2; ld.global.u32 %r1620, [%rd1125+8]; add.s32 %r1621, %r1617, %r1620; add.s32 %r1622, %r1621, -1; min.s32 %r7, %r1619, %r1622; and.b32 %r1623, %r1, 64; setp.eq.s32 %p34, %r1623, 0; @%p34 bra $L__BB107_28; bra.uni $L__BB107_15; $L__BB107_28: ld.const.u64 %rd1167, [kernel_params+2272]; cvta.to.global.u64 %rd1168, %rd1167; mul.wide.s32 %rd1169, %r6, 16; add.s64 %rd1170, %rd1168, %rd1169; ld.global.v4.f32 {%f28890, %f28889, %f28888, %f8092}, [%rd1170]; mul.wide.s32 %rd1171, %r5, 16; add.s64 %rd1172, %rd1168, %rd1171; ld.global.v4.f32 {%f28887, %f28886, %f28885, %f8096}, [%rd1172]; ld.global.v4.f32 {%f28884, %f28883, %f28882, %f8100}, [%rd1172+16]; mul.wide.s32 %rd1173, %r7, 16; add.s64 %rd1174, %rd1168, %rd1173; ld.global.v4.f32 {%f28881, %f28880, %f28879, %f8104}, [%rd1174]; bra.uni $L__BB107_29; $L__BB107_46: and.b32 %r1642, %r1, 8; setp.eq.s32 %p55, %r1642, 0; @%p55 bra $L__BB107_67; bra.uni $L__BB107_47; $L__BB107_67: setp.eq.s32 %p68, %r1, 1; @%p68 bra $L__BB107_69; bra.uni $L__BB107_68; $L__BB107_69: cvt.u32.u64 %r1661, %rd4; ld.const.u64 %rd1228, [kernel_params+2232]; cvta.to.global.u64 %rd1229, %rd1228; mul.wide.s32 %rd1230, %r1661, 16; add.s64 %rd1231, %rd1229, %rd1230; add.s64 %rd25, %rd1231, 12; ld.const.u64 %rd1232, [kernel_params+2256]; cvta.to.global.u64 %rd1233, %rd1232; ld.global.u32 %r1662, [%rd1231+12]; mul.wide.u32 %rd1234, %r1662, 12; add.s64 %rd1235, %rd1233, %rd1234; add.s32 %r1663, %r1662, 1; mul.wide.u32 %rd1236, %r1663, 12; add.s64 %rd1237, %rd1233, %rd1236; add.s32 %r1664, %r1662, 2; mul.wide.u32 %rd1238, %r1664, 12; add.s64 %rd1239, %rd1233, %rd1238; and.b32 %r1665, %r5556, 8; setp.eq.s32 %p69, %r1665, 0; ld.global.f32 %f8438, [%rd1239]; ld.global.f32 %f425, [%rd1235]; sub.ftz.f32 %f426, %f8438, %f425; ld.global.f32 %f8439, [%rd1239+4]; ld.global.f32 %f427, [%rd1235+4]; sub.ftz.f32 %f428, %f8439, %f427; ld.global.f32 %f8440, [%rd1239+8]; ld.global.f32 %f429, [%rd1235+8]; sub.ftz.f32 %f430, %f8440, %f429; ld.global.f32 %f8441, [%rd1237]; sub.ftz.f32 %f431, %f8441, %f425; ld.global.f32 %f8442, [%rd1237+4]; sub.ftz.f32 %f432, %f8442, %f427; ld.global.f32 %f8443, [%rd1237+8]; sub.ftz.f32 %f433, %f8443, %f429; mul.ftz.f32 %f434, %f433, %f428; mul.ftz.f32 %f435, %f432, %f430; mul.ftz.f32 %f436, %f431, %f430; mul.ftz.f32 %f437, %f433, %f426; mul.ftz.f32 %f438, %f432, %f426; mul.ftz.f32 %f439, %f431, %f428; @%p69 bra $L__BB107_71; bra.uni $L__BB107_70; $L__BB107_71: sub.ftz.f32 %f8451, %f435, %f434; sub.ftz.f32 %f8452, %f437, %f436; mul.ftz.f32 %f8453, %f8452, %f8452; fma.rn.ftz.f32 %f8454, %f8451, %f8451, %f8453; sub.ftz.f32 %f8455, %f439, %f438; fma.rn.ftz.f32 %f8456, %f8455, %f8455, %f8454; sqrt.approx.ftz.f32 %f8457, %f8456; div.approx.ftz.f32 %f28920, %f8455, %f8457; div.approx.ftz.f32 %f28919, %f8452, %f8457; div.approx.ftz.f32 %f28918, %f8451, %f8457; bra.uni $L__BB107_72; $L__BB107_15: ld.const.u64 %rd1126, [kernel_params+2160]; cvta.to.global.u64 %rd1127, %rd1126; mul.lo.s64 %rd1128, %rd7, 208; add.s64 %rd1129, %rd1127, %rd1128; ld.global.u32 %r8, [%rd1129+148]; ld.global.u32 %r9, [%rd1129+152]; shl.b32 %r1624, %r9, 1; cvt.rn.f32.s32 %f8008, %r1624; mul.ftz.f32 %f218, %f10, %f8008; cvt.rzi.ftz.s32.f32 %r1625, %f218; add.s32 %r1626, %r1624, -1; min.s32 %r10, %r1625, %r1626; ld.const.u64 %rd1130, [kernel_params+2312]; cvta.to.global.u64 %rd10, %rd1130; ld.global.u32 %r5547, [%rd1129+164]; mul.wide.u32 %rd1131, %r5547, 16; add.s64 %rd1132, %rd10, %rd1131; ld.global.u32 %r5546, [%rd1132+8]; ld.global.u16 %rs608, [%rd1132+12]; ld.global.u64 %rd4731, [%rd1132]; setp.eq.s64 %p35, %rd4731, 11; @%p35 bra $L__BB107_20; $L__BB107_16: setp.eq.s64 %p36, %rd4731, 0; @%p36 bra $L__BB107_18; add.s32 %r5547, %r5547, 2; bra.uni $L__BB107_19; $L__BB107_18: setp.eq.s16 %p37, %rs608, 0; mov.u32 %r5547, %r5546; @%p37 bra $L__BB107_21; $L__BB107_19: mul.wide.u32 %rd1133, %r5547, 16; add.s64 %rd1134, %rd10, %rd1133; ld.global.u32 %r5546, [%rd1134+8]; ld.global.u16 %rs608, [%rd1134+12]; ld.global.u64 %rd4731, [%rd1134]; setp.ne.s64 %p38, %rd4731, 11; @%p38 bra $L__BB107_16; $L__BB107_20: setp.eq.s16 %p39, %rs608, 0; selp.b32 %r5550, -1, %r5546, %p39; $L__BB107_21: setp.eq.s32 %p40, %r10, %r9; @%p40 bra $L__BB107_23; bra.uni $L__BB107_22; $L__BB107_23: ld.const.u64 %rd1143, [kernel_params+2272]; cvta.to.global.u64 %rd1144, %rd1143; mul.wide.s32 %rd1145, %r6, 16; add.s64 %rd1146, %rd1144, %rd1145; ld.global.v4.f32 {%f28866, %f28865, %f28864, %f8028}, [%rd1146]; mul.wide.s32 %rd1147, %r5, 16; add.s64 %rd1148, %rd1144, %rd1147; ld.global.v4.f32 {%f28863, %f28862, %f28861, %f8032}, [%rd1148]; ld.global.v4.f32 {%f28860, %f28859, %f28858, %f8036}, [%rd1148+16]; mul.wide.s32 %rd1149, %r7, 16; add.s64 %rd1150, %rd1144, %rd1149; ld.global.v4.f32 {%f28857, %f28856, %f28855, %f8040}, [%rd1150]; bra.uni $L__BB107_24; $L__BB107_47: ld.const.u64 %rd1189, [kernel_params+2304]; cvta.to.global.u64 %rd1190, %rd1189; shl.b64 %rd1191, %rd4, 2; add.s64 %rd1192, %rd1190, %rd1191; ld.global.u32 %r1643, [%rd1192]; st.local.u32 [%rd5+-40], %r1643; fma.rn.ftz.f32 %f8352, %f28854, %f9, %f28848; fma.rn.ftz.f32 %f8353, %f28854, %f8, %f28849; fma.rn.ftz.f32 %f8354, %f28854, %f7, %f28850; cvta.to.local.u64 %rd1194, %rd1061; st.local.v2.f32 [%rd1194], {%f8354, %f8353}; st.local.f32 [%rd5+-80], %f8352; and.b32 %r1644, %r1, 64; setp.eq.s32 %p56, %r1644, 0; @%p56 bra $L__BB107_61; bra.uni $L__BB107_48; $L__BB107_61: ld.const.u64 %rd1216, [kernel_params+2296]; cvta.to.global.u64 %rd1217, %rd1216; shl.b64 %rd1218, %rd4, 4; add.s64 %rd1219, %rd1217, %rd1218; ld.global.v4.f32 {%f28917, %f28916, %f28915, %f8379}, [%rd1219]; bra.uni $L__BB107_62; $L__BB107_22: setp.gt.s32 %p41, %r10, %r9; selp.b32 %r1628, -1, 0, %p41; add.s32 %r1629, %r10, %r1628; mad.lo.s32 %r1630, %r1629, %r8, %r5550; ld.const.u64 %rd1135, [kernel_params+2344]; cvta.to.global.u64 %rd1136, %rd1135; add.s32 %r1631, %r1630, %r6; mul.wide.s32 %rd1137, %r1631, 16; add.s64 %rd1138, %rd1136, %rd1137; ld.global.v4.f32 {%f28866, %f28865, %f28864, %f8012}, [%rd1138]; add.s32 %r1632, %r1630, %r5; mul.wide.s32 %rd1139, %r1632, 16; add.s64 %rd1140, %rd1136, %rd1139; ld.global.v4.f32 {%f28863, %f28862, %f28861, %f8016}, [%rd1140]; ld.global.v4.f32 {%f28860, %f28859, %f28858, %f8020}, [%rd1140+16]; add.s32 %r1633, %r1630, %r7; mul.wide.s32 %rd1141, %r1633, 16; add.s64 %rd1142, %rd1136, %rd1141; ld.global.v4.f32 {%f28857, %f28856, %f28855, %f8024}, [%rd1142]; $L__BB107_24: add.s32 %r21, %r10, 1; setp.eq.s32 %p42, %r21, %r9; @%p42 bra $L__BB107_26; bra.uni $L__BB107_25; $L__BB107_26: ld.const.u64 %rd1159, [kernel_params+2272]; cvta.to.global.u64 %rd1160, %rd1159; mul.wide.s32 %rd1161, %r6, 16; add.s64 %rd1162, %rd1160, %rd1161; ld.global.v4.f32 {%f28878, %f28877, %f28876, %f8060}, [%rd1162]; mul.wide.s32 %rd1163, %r5, 16; add.s64 %rd1164, %rd1160, %rd1163; ld.global.v4.f32 {%f28875, %f28874, %f28873, %f8064}, [%rd1164]; ld.global.v4.f32 {%f28872, %f28871, %f28870, %f8068}, [%rd1164+16]; mul.wide.s32 %rd1165, %r7, 16; add.s64 %rd1166, %rd1160, %rd1165; ld.global.v4.f32 {%f28869, %f28868, %f28867, %f8072}, [%rd1166]; bra.uni $L__BB107_27; $L__BB107_25: setp.ge.s32 %p43, %r10, %r9; selp.b32 %r1634, -1, 0, %p43; add.s32 %r1635, %r21, %r1634; mad.lo.s32 %r1636, %r1635, %r8, %r5550; ld.const.u64 %rd1151, [kernel_params+2344]; cvta.to.global.u64 %rd1152, %rd1151; add.s32 %r1637, %r1636, %r6; mul.wide.s32 %rd1153, %r1637, 16; add.s64 %rd1154, %rd1152, %rd1153; ld.global.v4.f32 {%f28878, %f28877, %f28876, %f8044}, [%rd1154]; add.s32 %r1638, %r1636, %r5; mul.wide.s32 %rd1155, %r1638, 16; add.s64 %rd1156, %rd1152, %rd1155; ld.global.v4.f32 {%f28875, %f28874, %f28873, %f8048}, [%rd1156]; ld.global.v4.f32 {%f28872, %f28871, %f28870, %f8052}, [%rd1156+16]; add.s32 %r1639, %r1636, %r7; mul.wide.s32 %rd1157, %r1639, 16; add.s64 %rd1158, %rd1152, %rd1157; ld.global.v4.f32 {%f28869, %f28868, %f28867, %f8056}, [%rd1158]; $L__BB107_27: shl.b32 %r5528, %r9, 1; cvt.rn.f32.s32 %f28716, %r5528; mul.ftz.f32 %f28715, %f10, %f28716; cvt.rn.f32.s32 %f8073, %r10; sub.ftz.f32 %f8074, %f28715, %f8073; mov.f32 %f8075, 0f3F800000; sub.ftz.f32 %f8076, %f8075, %f8074; mul.ftz.f32 %f8077, %f8074, %f28878; fma.rn.ftz.f32 %f28890, %f8076, %f28866, %f8077; mul.ftz.f32 %f8078, %f8074, %f28877; fma.rn.ftz.f32 %f28889, %f8076, %f28865, %f8078; mul.ftz.f32 %f8079, %f8074, %f28876; fma.rn.ftz.f32 %f28888, %f8076, %f28864, %f8079; mul.ftz.f32 %f8080, %f8074, %f28875; fma.rn.ftz.f32 %f28887, %f8076, %f28863, %f8080; mul.ftz.f32 %f8081, %f8074, %f28874; fma.rn.ftz.f32 %f28886, %f8076, %f28862, %f8081; mul.ftz.f32 %f8082, %f8074, %f28873; fma.rn.ftz.f32 %f28885, %f8076, %f28861, %f8082; mul.ftz.f32 %f8083, %f8074, %f28872; fma.rn.ftz.f32 %f28884, %f8076, %f28860, %f8083; mul.ftz.f32 %f8084, %f8074, %f28871; fma.rn.ftz.f32 %f28883, %f8076, %f28859, %f8084; mul.ftz.f32 %f8085, %f8074, %f28870; fma.rn.ftz.f32 %f28882, %f8076, %f28858, %f8085; mul.ftz.f32 %f8086, %f8074, %f28869; fma.rn.ftz.f32 %f28881, %f8076, %f28857, %f8086; mul.ftz.f32 %f8087, %f8074, %f28868; fma.rn.ftz.f32 %f28880, %f8076, %f28856, %f8087; mul.ftz.f32 %f8088, %f8074, %f28867; fma.rn.ftz.f32 %f28879, %f8076, %f28855, %f8088; $L__BB107_29: fma.rn.ftz.f32 %f28911, %f28853, %f28854, %f28850; fma.rn.ftz.f32 %f28910, %f28852, %f28854, %f28849; fma.rn.ftz.f32 %f28909, %f28851, %f28854, %f28848; mov.f32 %f8105, 0f3F800000; sub.ftz.f32 %f330, %f8105, %f1; mul.ftz.f32 %f331, %f330, %f330; add.ftz.f32 %f8106, %f330, %f330; mul.ftz.f32 %f8107, %f1, %f8106; sub.ftz.f32 %f8108, %f8107, %f331; mul.ftz.f32 %f8109, %f1, 0f40400000; add.ftz.f32 %f332, %f8109, 0fC0A00000; add.ftz.f32 %f8110, %f1, %f1; mul.ftz.f32 %f8111, %f8110, %f332; fma.rn.ftz.f32 %f8112, %f1, %f8109, %f8111; add.ftz.f32 %f8113, %f8109, 0f40000000; mul.ftz.f32 %f8114, %f8106, %f8113; mul.ftz.f32 %f333, %f330, 0f40400000; mul.ftz.f32 %f8115, %f330, %f333; sub.ftz.f32 %f8116, %f8114, %f8115; mul.ftz.f32 %f8117, %f330, 0fC0000000; mul.ftz.f32 %f334, %f1, %f1; fma.rn.ftz.f32 %f8118, %f1, %f8117, %f334; mul.ftz.f32 %f8119, %f28887, %f8112; mul.ftz.f32 %f8120, %f28886, %f8112; mul.ftz.f32 %f8121, %f28885, %f8112; fma.rn.ftz.f32 %f8122, %f28890, %f8108, %f8119; fma.rn.ftz.f32 %f8123, %f28889, %f8108, %f8120; fma.rn.ftz.f32 %f8124, %f28888, %f8108, %f8121; fma.rn.ftz.f32 %f8125, %f28884, %f8116, %f8122; fma.rn.ftz.f32 %f8126, %f28883, %f8116, %f8123; fma.rn.ftz.f32 %f8127, %f28882, %f8116, %f8124; fma.rn.ftz.f32 %f8128, %f28881, %f8118, %f8125; fma.rn.ftz.f32 %f8129, %f28880, %f8118, %f8126; fma.rn.ftz.f32 %f8130, %f28879, %f8118, %f8127; mul.ftz.f32 %f28903, %f8128, 0f3F000000; mul.ftz.f32 %f28905, %f8129, 0f3F000000; mul.ftz.f32 %f28904, %f8130, 0f3F000000; and.b32 %r22, %r1, 4; setp.eq.s32 %p44, %r22, 0; @%p44 bra $L__BB107_31; bra.uni $L__BB107_30; $L__BB107_31: mul.ftz.f32 %f8189, %f1, %f330; mul.ftz.f32 %f8190, %f330, %f8189; fma.rn.ftz.f32 %f8191, %f334, %f332, 0f40000000; add.ftz.f32 %f8192, %f333, 0fC0A00000; fma.rn.ftz.f32 %f8193, %f331, %f8192, 0f40000000; mul.ftz.f32 %f8194, %f1, %f8189; mul.ftz.f32 %f8195, %f28890, %f8190; mul.ftz.f32 %f8196, %f28889, %f8190; mul.ftz.f32 %f8197, %f28888, %f8190; mul.ftz.f32 %f8198, %f28887, %f8191; mul.ftz.f32 %f8199, %f28886, %f8191; mul.ftz.f32 %f8200, %f28885, %f8191; sub.ftz.f32 %f8201, %f8198, %f8195; sub.ftz.f32 %f8202, %f8199, %f8196; sub.ftz.f32 %f8203, %f8200, %f8197; fma.rn.ftz.f32 %f8204, %f28884, %f8193, %f8201; fma.rn.ftz.f32 %f8205, %f28883, %f8193, %f8202; fma.rn.ftz.f32 %f8206, %f28882, %f8193, %f8203; mul.ftz.f32 %f8207, %f28881, %f8194; mul.ftz.f32 %f8208, %f28880, %f8194; mul.ftz.f32 %f8209, %f28879, %f8194; sub.ftz.f32 %f8210, %f8204, %f8207; sub.ftz.f32 %f8211, %f8205, %f8208; sub.ftz.f32 %f8212, %f8206, %f8209; mul.ftz.f32 %f341, %f8210, 0f3F000000; mul.ftz.f32 %f342, %f8211, 0f3F000000; mul.ftz.f32 %f343, %f8212, 0f3F000000; setp.eq.ftz.f32 %p45, %f28911, %f341; setp.eq.ftz.f32 %p46, %f28910, %f342; and.pred %p47, %p45, %p46; setp.eq.ftz.f32 %p48, %f28909, %f343; and.pred %p49, %p48, %p47; @%p49 bra $L__BB107_33; bra.uni $L__BB107_32; $L__BB107_33: mov.f32 %f28908, %f7; mov.f32 %f28907, %f8; mov.f32 %f28906, %f9; bra.uni $L__BB107_34; $L__BB107_30: mul.ftz.f32 %f8131, %f28905, %f28905; fma.rn.ftz.f32 %f8132, %f28903, %f28903, %f8131; fma.rn.ftz.f32 %f8133, %f28904, %f28904, %f8132; sqrt.approx.ftz.f32 %f8134, %f8133; div.approx.ftz.f32 %f8135, %f28903, %f8134; div.approx.ftz.f32 %f8136, %f28905, %f8134; div.approx.ftz.f32 %f8137, %f28904, %f8134; mul.ftz.f32 %f8138, %f8136, %f28851; mul.ftz.f32 %f8139, %f8137, %f28852; sub.ftz.f32 %f8140, %f8139, %f8138; mul.ftz.f32 %f8141, %f8137, %f28853; mul.ftz.f32 %f8142, %f8135, %f28851; sub.ftz.f32 %f8143, %f8142, %f8141; mul.ftz.f32 %f8144, %f8135, %f28852; mul.ftz.f32 %f8145, %f8136, %f28853; sub.ftz.f32 %f8146, %f8145, %f8144; mul.ftz.f32 %f8147, %f8143, %f8143; fma.rn.ftz.f32 %f8148, %f8140, %f8140, %f8147; fma.rn.ftz.f32 %f8149, %f8146, %f8146, %f8148; sqrt.approx.ftz.f32 %f8150, %f8149; div.approx.ftz.f32 %f8151, %f8140, %f8150; div.approx.ftz.f32 %f8152, %f8143, %f8150; div.approx.ftz.f32 %f8153, %f8146, %f8150; mul.ftz.f32 %f8154, %f2, %f2; sub.ftz.f32 %f8156, %f8105, %f8154; mov.f32 %f8157, 0f00000000; max.ftz.f32 %f8158, %f8156, %f8157; sqrt.approx.ftz.f32 %f8159, %f8158; mul.ftz.f32 %f8160, %f8151, %f2; mul.ftz.f32 %f8161, %f8152, %f2; mul.ftz.f32 %f8162, %f8153, %f2; mul.ftz.f32 %f8163, %f8136, %f8153; mul.ftz.f32 %f8164, %f8137, %f8152; sub.ftz.f32 %f8165, %f8163, %f8164; mul.ftz.f32 %f8166, %f8137, %f8151; mul.ftz.f32 %f8167, %f8135, %f8153; sub.ftz.f32 %f8168, %f8166, %f8167; mul.ftz.f32 %f8169, %f8135, %f8152; mul.ftz.f32 %f8170, %f8136, %f8151; sub.ftz.f32 %f8171, %f8169, %f8170; mul.ftz.f32 %f8172, %f8168, %f8168; fma.rn.ftz.f32 %f8173, %f8165, %f8165, %f8172; fma.rn.ftz.f32 %f8174, %f8171, %f8171, %f8173; sqrt.approx.ftz.f32 %f8175, %f8174; div.approx.ftz.f32 %f8176, %f8165, %f8175; div.approx.ftz.f32 %f8177, %f8168, %f8175; div.approx.ftz.f32 %f8178, %f8171, %f8175; mul.ftz.f32 %f8179, %f8159, %f8176; mul.ftz.f32 %f8180, %f8159, %f8177; mul.ftz.f32 %f8181, %f8159, %f8178; sub.ftz.f32 %f8182, %f8160, %f8179; sub.ftz.f32 %f8183, %f8161, %f8180; sub.ftz.f32 %f8184, %f8162, %f8181; mul.ftz.f32 %f8185, %f8183, %f8183; fma.rn.ftz.f32 %f8186, %f8182, %f8182, %f8185; fma.rn.ftz.f32 %f8187, %f8184, %f8184, %f8186; sqrt.approx.ftz.f32 %f8188, %f8187; div.approx.ftz.f32 %f28908, %f8182, %f8188; div.approx.ftz.f32 %f28907, %f8183, %f8188; div.approx.ftz.f32 %f28906, %f8184, %f8188; st.local.f32 [%rd5+-76], %f28908; st.local.f32 [%rd5+-72], %f28907; st.local.f32 [%rd5+-68], %f28906; bra.uni $L__BB107_35; $L__BB107_32: sub.ftz.f32 %f8213, %f28911, %f341; sub.ftz.f32 %f8214, %f28910, %f342; mul.ftz.f32 %f8215, %f8214, %f8214; fma.rn.ftz.f32 %f8216, %f8213, %f8213, %f8215; sub.ftz.f32 %f8217, %f28909, %f343; fma.rn.ftz.f32 %f8218, %f8217, %f8217, %f8216; sqrt.approx.ftz.f32 %f8219, %f8218; div.approx.ftz.f32 %f28906, %f8217, %f8219; div.approx.ftz.f32 %f28907, %f8214, %f8219; div.approx.ftz.f32 %f28908, %f8213, %f8219; $L__BB107_34: mov.u32 %r5525, 0; st.local.f32 [%rd5+-76], %f28908; st.local.f32 [%rd5+-72], %f28907; st.local.f32 [%rd5+-68], %f28906; st.local.u32 [%rd5+-16], %r5525; $L__BB107_35: and.b32 %r5529, %r5556, 4; setp.ne.s32 %p3917, %r5529, 0; st.local.f32 [%rd5+28], %f28903; st.local.f32 [%rd5+32], %f28905; st.local.f32 [%rd5+36], %f28904; @%p3917 bra $L__BB107_45; @%p24 bra $L__BB107_38; bra.uni $L__BB107_37; $L__BB107_38: ld.const.u64 %rd1175, [kernel_params+2160]; cvta.to.global.u64 %rd1176, %rd1175; mul.lo.s64 %rd1177, %rd7, 208; add.s64 %rd1178, %rd1176, %rd1177; ld.global.v4.f32 {%f8229, %f8230, %f8231, %f8232}, [%rd1178]; ld.global.v4.f32 {%f8237, %f8238, %f8239, %f8240}, [%rd1178+16]; ld.global.v4.f32 {%f8245, %f8246, %f8247, %f8248}, [%rd1178+32]; mul.ftz.f32 %f8253, %f28910, %f8230; fma.rn.ftz.f32 %f8254, %f28911, %f8229, %f8253; fma.rn.ftz.f32 %f8255, %f28909, %f8231, %f8254; add.ftz.f32 %f28899, %f8232, %f8255; mul.ftz.f32 %f8256, %f28910, %f8238; fma.rn.ftz.f32 %f8257, %f28911, %f8237, %f8256; fma.rn.ftz.f32 %f8258, %f28909, %f8239, %f8257; add.ftz.f32 %f28898, %f8240, %f8258; mul.ftz.f32 %f8259, %f28910, %f8246; fma.rn.ftz.f32 %f8260, %f28911, %f8245, %f8259; fma.rn.ftz.f32 %f8261, %f28909, %f8247, %f8260; add.ftz.f32 %f28909, %f8248, %f8261; bra.uni $L__BB107_39; $L__BB107_37: mul.ftz.f32 %f8220, %f28910, %f142; fma.rn.ftz.f32 %f8221, %f28911, %f141, %f8220; fma.rn.ftz.f32 %f8222, %f28909, %f143, %f8221; add.ftz.f32 %f28899, %f144, %f8222; mul.ftz.f32 %f8223, %f28910, %f146; fma.rn.ftz.f32 %f8224, %f28911, %f145, %f8223; fma.rn.ftz.f32 %f8225, %f28909, %f147, %f8224; add.ftz.f32 %f28898, %f148, %f8225; mul.ftz.f32 %f8226, %f28910, %f150; fma.rn.ftz.f32 %f8227, %f28911, %f149, %f8226; fma.rn.ftz.f32 %f8228, %f28909, %f151, %f8227; add.ftz.f32 %f28909, %f152, %f8228; $L__BB107_39: @%p24 bra $L__BB107_41; bra.uni $L__BB107_40; $L__BB107_41: ld.const.u64 %rd1179, [kernel_params+2160]; cvta.to.global.u64 %rd1180, %rd1179; mul.lo.s64 %rd1181, %rd7, 208; add.s64 %rd1182, %rd1180, %rd1181; ld.global.v4.f32 {%f8275, %f8276, %f8277, %f8278}, [%rd1182+48]; ld.global.v4.f32 {%f8282, %f8283, %f8284, %f8285}, [%rd1182+64]; ld.global.v4.f32 {%f8289, %f8290, %f8291, %f8292}, [%rd1182+80]; mul.ftz.f32 %f8296, %f8282, %f28907; fma.rn.ftz.f32 %f8297, %f8275, %f28908, %f8296; fma.rn.ftz.f32 %f8298, %f8289, %f28906, %f8297; mul.ftz.f32 %f8299, %f8283, %f28907; fma.rn.ftz.f32 %f8300, %f8276, %f28908, %f8299; fma.rn.ftz.f32 %f8301, %f8290, %f28906, %f8300; mul.ftz.f32 %f8302, %f8284, %f28907; fma.rn.ftz.f32 %f8303, %f8277, %f28908, %f8302; fma.rn.ftz.f32 %f8304, %f8291, %f28906, %f8303; mul.ftz.f32 %f8305, %f8301, %f8301; fma.rn.ftz.f32 %f8306, %f8298, %f8298, %f8305; fma.rn.ftz.f32 %f8307, %f8304, %f8304, %f8306; sqrt.approx.ftz.f32 %f8308, %f8307; div.approx.ftz.f32 %f28908, %f8298, %f8308; div.approx.ftz.f32 %f28907, %f8301, %f8308; div.approx.ftz.f32 %f28906, %f8304, %f8308; bra.uni $L__BB107_42; $L__BB107_40: mul.ftz.f32 %f8262, %f28907, %f28825; fma.rn.ftz.f32 %f8263, %f28908, %f28821, %f8262; fma.rn.ftz.f32 %f8264, %f28906, %f28829, %f8263; mul.ftz.f32 %f8265, %f28907, %f28826; fma.rn.ftz.f32 %f8266, %f28908, %f28822, %f8265; fma.rn.ftz.f32 %f8267, %f28906, %f28830, %f8266; mul.ftz.f32 %f8268, %f28907, %f28827; fma.rn.ftz.f32 %f8269, %f28908, %f28823, %f8268; fma.rn.ftz.f32 %f8270, %f28906, %f28831, %f8269; mul.ftz.f32 %f8271, %f8267, %f8267; fma.rn.ftz.f32 %f8272, %f8264, %f8264, %f8271; fma.rn.ftz.f32 %f8273, %f8270, %f8270, %f8272; sqrt.approx.ftz.f32 %f8274, %f8273; div.approx.ftz.f32 %f28908, %f8264, %f8274; div.approx.ftz.f32 %f28907, %f8267, %f8274; div.approx.ftz.f32 %f28906, %f8270, %f8274; $L__BB107_42: st.local.f32 [%rd5+-76], %f28908; st.local.f32 [%rd5+-72], %f28907; st.local.f32 [%rd5+-68], %f28906; @%p24 bra $L__BB107_44; bra.uni $L__BB107_43; $L__BB107_44: ld.const.u64 %rd1183, [kernel_params+2160]; cvta.to.global.u64 %rd1184, %rd1183; mul.lo.s64 %rd1185, %rd7, 208; add.s64 %rd1186, %rd1184, %rd1185; ld.global.v4.f32 {%f8315, %f8316, %f8317, %f8318}, [%rd1186]; ld.global.v4.f32 {%f8322, %f8323, %f8324, %f8325}, [%rd1186+16]; ld.global.v4.f32 {%f8329, %f8330, %f8331, %f8332}, [%rd1186+32]; mul.ftz.f32 %f8336, %f8316, %f28905; fma.rn.ftz.f32 %f8337, %f28903, %f8315, %f8336; fma.rn.ftz.f32 %f380, %f8317, %f28904, %f8337; mul.ftz.f32 %f8338, %f8323, %f28905; fma.rn.ftz.f32 %f8339, %f28903, %f8322, %f8338; fma.rn.ftz.f32 %f381, %f8324, %f28904, %f8339; mul.ftz.f32 %f8340, %f8330, %f28905; fma.rn.ftz.f32 %f8341, %f28903, %f8329, %f8340; fma.rn.ftz.f32 %f28904, %f8331, %f28904, %f8341; st.local.f32 [%rd5+28], %f380; st.local.f32 [%rd5+32], %f381; st.local.f32 [%rd5+36], %f28904; mov.f32 %f28903, %f380; mov.f32 %f28905, %f381; mov.f32 %f28910, %f28898; mov.f32 %f28911, %f28899; bra.uni $L__BB107_45; $L__BB107_43: mul.ftz.f32 %f8309, %f28905, %f142; fma.rn.ftz.f32 %f8310, %f28903, %f141, %f8309; fma.rn.ftz.f32 %f377, %f28904, %f143, %f8310; mul.ftz.f32 %f8311, %f28905, %f146; fma.rn.ftz.f32 %f8312, %f28903, %f145, %f8311; fma.rn.ftz.f32 %f378, %f28904, %f147, %f8312; mul.ftz.f32 %f8313, %f28905, %f150; fma.rn.ftz.f32 %f8314, %f28903, %f149, %f8313; fma.rn.ftz.f32 %f28904, %f28904, %f151, %f8314; st.local.f32 [%rd5+28], %f377; st.local.f32 [%rd5+32], %f378; st.local.f32 [%rd5+36], %f28904; mov.f32 %f28903, %f377; mov.f32 %f28905, %f378; mov.f32 %f28910, %f28898; mov.f32 %f28911, %f28899; $L__BB107_45: neg.ftz.f32 %f28714, %f7; neg.ftz.f32 %f28713, %f8; neg.ftz.f32 %f28712, %f9; cvta.to.local.u64 %rd1188, %rd1061; st.local.v2.f32 [%rd1188], {%f28911, %f28910}; st.local.f32 [%rd5+-80], %f28909; selp.f32 %f8342, %f28906, %f28712, %p44; selp.f32 %f28949, %f28907, %f28713, %p44; selp.f32 %f28948, %f28908, %f28714, %p44; st.local.v2.f32 [%rd5+-64], {%f28948, %f28949}; st.local.f32 [%rd5+-56], %f8342; mul.ftz.f32 %f8343, %f28905, %f8342; mul.ftz.f32 %f8344, %f28904, %f28949; mul.ftz.f32 %f8345, %f28904, %f28948; mul.ftz.f32 %f8346, %f28903, %f8342; mul.ftz.f32 %f8347, %f28903, %f28949; mul.ftz.f32 %f8348, %f28905, %f28948; sub.ftz.f32 %f8349, %f8347, %f8348; sub.ftz.f32 %f8350, %f8345, %f8346; sub.ftz.f32 %f8351, %f8343, %f8344; st.local.v2.f32 [%rd5+40], {%f8351, %f8350}; st.local.f32 [%rd5+48], %f8349; ld.global.u32 %r1641, [%rd9+-4]; st.local.u32 [%rd5+-40], %r1641; bra.uni $L__BB107_95; $L__BB107_68: { // callseq 244, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 244 ld.local.u32 %r5556, [%rd5+-32]; bra.uni $L__BB107_81; $L__BB107_48: ld.const.u64 %rd1195, [kernel_params+2160]; cvta.to.global.u64 %rd1196, %rd1195; mul.lo.s64 %rd1197, %rd7, 208; add.s64 %rd1198, %rd1196, %rd1197; ld.global.u32 %r23, [%rd1198+148]; ld.global.u32 %r24, [%rd1198+152]; shl.b32 %r1645, %r24, 1; cvt.rn.f32.s32 %f8355, %r1645; mul.ftz.f32 %f394, %f10, %f8355; cvt.rzi.ftz.s32.f32 %r1646, %f394; add.s32 %r1647, %r1645, -1; min.s32 %r25, %r1646, %r1647; ld.const.u64 %rd1199, [kernel_params+2312]; cvta.to.global.u64 %rd14, %rd1199; ld.global.u32 %r5552, [%rd1198+164]; mul.wide.u32 %rd1200, %r5552, 16; add.s64 %rd1201, %rd14, %rd1200; ld.global.u32 %r5551, [%rd1201+8]; ld.global.u16 %rs610, [%rd1201+12]; ld.global.u64 %rd4732, [%rd1201]; setp.eq.s64 %p57, %rd4732, 11; @%p57 bra $L__BB107_53; $L__BB107_49: setp.eq.s64 %p58, %rd4732, 0; @%p58 bra $L__BB107_51; add.s32 %r5552, %r5552, 2; bra.uni $L__BB107_52; $L__BB107_51: setp.eq.s16 %p59, %rs610, 0; mov.u32 %r5555, -1; mov.u32 %r5552, %r5551; @%p59 bra $L__BB107_54; $L__BB107_52: mul.wide.u32 %rd1202, %r5552, 16; add.s64 %rd1203, %rd14, %rd1202; ld.global.u32 %r5551, [%rd1203+8]; ld.global.u16 %rs610, [%rd1203+12]; ld.global.u64 %rd4732, [%rd1203]; setp.ne.s64 %p60, %rd4732, 11; @%p60 bra $L__BB107_49; $L__BB107_53: setp.eq.s16 %p61, %rs610, 0; selp.b32 %r5555, -1, %r5551, %p61; $L__BB107_54: setp.eq.s32 %p62, %r25, %r24; @%p62 bra $L__BB107_56; bra.uni $L__BB107_55; $L__BB107_56: ld.const.u64 %rd1207, [kernel_params+2296]; cvta.to.global.u64 %rd1208, %rd1207; shl.b64 %rd1209, %rd4, 4; add.s64 %rd4733, %rd1208, %rd1209; bra.uni $L__BB107_57; $L__BB107_70: sub.ftz.f32 %f8444, %f434, %f435; sub.ftz.f32 %f8445, %f436, %f437; mul.ftz.f32 %f8446, %f8445, %f8445; fma.rn.ftz.f32 %f8447, %f8444, %f8444, %f8446; sub.ftz.f32 %f8448, %f438, %f439; fma.rn.ftz.f32 %f8449, %f8448, %f8448, %f8447; sqrt.approx.ftz.f32 %f8450, %f8449; div.approx.ftz.f32 %f28920, %f8448, %f8450; div.approx.ftz.f32 %f28919, %f8445, %f8450; div.approx.ftz.f32 %f28918, %f8444, %f8450; $L__BB107_72: ld.const.u64 %rd1240, [kernel_params+2216]; cvta.to.global.u64 %rd1241, %rd1240; shl.b64 %rd1242, %rd4, 2; add.s64 %rd1243, %rd1241, %rd1242; ld.global.u32 %r38, [%rd1243]; st.local.u32 [%rd5+-40], %r38; fma.rn.ftz.f32 %f8458, %f1, %f431, %f425; fma.rn.ftz.f32 %f8459, %f1, %f432, %f427; fma.rn.ftz.f32 %f8460, %f1, %f433, %f429; fma.rn.ftz.f32 %f28933, %f2, %f426, %f8458; fma.rn.ftz.f32 %f28934, %f2, %f428, %f8459; fma.rn.ftz.f32 %f28935, %f2, %f430, %f8460; and.b32 %r1666, %r5556, 4; setp.ne.s32 %p70, %r1666, 0; @%p70 bra $L__BB107_76; and.b32 %r5527, %r5556, 2; setp.ne.s32 %p71, %r5527, 0; @%p71 bra $L__BB107_75; ld.const.u64 %rd1244, [kernel_params+2160]; cvta.to.global.u64 %rd1245, %rd1244; mul.lo.s64 %rd1246, %rd7, 208; add.s64 %rd1247, %rd1245, %rd1246; ld.global.v4.f32 {%f141, %f142, %f143, %f144}, [%rd1247]; ld.global.v4.f32 {%f145, %f146, %f147, %f148}, [%rd1247+16]; ld.global.v4.f32 {%f149, %f150, %f151, %f152}, [%rd1247+32]; $L__BB107_75: mul.ftz.f32 %f8473, %f28934, %f142; fma.rn.ftz.f32 %f8474, %f28933, %f141, %f8473; fma.rn.ftz.f32 %f8475, %f28935, %f143, %f8474; mul.ftz.f32 %f8476, %f28934, %f146; fma.rn.ftz.f32 %f8477, %f28933, %f145, %f8476; fma.rn.ftz.f32 %f8478, %f28935, %f147, %f8477; mul.ftz.f32 %f8479, %f28934, %f150; fma.rn.ftz.f32 %f8480, %f28933, %f149, %f8479; fma.rn.ftz.f32 %f8481, %f28935, %f151, %f8480; add.ftz.f32 %f28933, %f144, %f8475; add.ftz.f32 %f28934, %f148, %f8478; add.ftz.f32 %f28935, %f152, %f8481; $L__BB107_76: st.local.f32 [%rd5+-56], %f28920; cvta.to.local.u64 %rd1249, %rd1061; st.local.v4.f32 [%rd1249], {%f28933, %f28934, %f28935, %f28918}; st.local.v4.f32 [%rd5+-72], {%f28919, %f28920, %f28918, %f28919}; setp.gt.s32 %p72, %r38, -1; @%p72 bra $L__BB107_80; ld.global.v4.u32 {%r1667, %r1668, %r1669, %r1670}, [%rd25+-12]; ld.const.u64 %rd1250, [kernel_params+2224]; cvta.to.global.u64 %rd1251, %rd1250; mul.wide.u32 %rd1252, %r1667, 12; add.s64 %rd1253, %rd1251, %rd1252; mul.wide.u32 %rd1254, %r1668, 12; add.s64 %rd1255, %rd1251, %rd1254; mul.wide.u32 %rd1256, %r1669, 12; add.s64 %rd1257, %rd1251, %rd1256; mov.f32 %f8482, 0f3F800000; sub.ftz.f32 %f8483, %f8482, %f1; sub.ftz.f32 %f8484, %f8483, %f2; ld.global.f32 %f8485, [%rd1253]; ld.global.f32 %f8486, [%rd1253+4]; ld.global.f32 %f8487, [%rd1253+8]; ld.global.f32 %f8488, [%rd1255]; mul.ftz.f32 %f8489, %f1, %f8488; ld.global.f32 %f8490, [%rd1255+4]; mul.ftz.f32 %f8491, %f1, %f8490; ld.global.f32 %f8492, [%rd1255+8]; mul.ftz.f32 %f8493, %f1, %f8492; fma.rn.ftz.f32 %f8494, %f8484, %f8485, %f8489; fma.rn.ftz.f32 %f8495, %f8484, %f8486, %f8491; fma.rn.ftz.f32 %f8496, %f8484, %f8487, %f8493; ld.global.f32 %f8497, [%rd1257]; ld.global.f32 %f8498, [%rd1257+4]; ld.global.f32 %f8499, [%rd1257+8]; fma.rn.ftz.f32 %f28936, %f2, %f8497, %f8494; fma.rn.ftz.f32 %f28937, %f2, %f8498, %f8495; fma.rn.ftz.f32 %f28938, %f2, %f8499, %f8496; mul.ftz.f32 %f8500, %f28937, %f28937; fma.rn.ftz.f32 %f8501, %f28936, %f28936, %f8500; fma.rn.ftz.f32 %f8502, %f28938, %f28938, %f8501; sqrt.approx.ftz.f32 %f488, %f8502; setp.neu.ftz.f32 %p73, %f488, 0f00000000; @%p73 bra $L__BB107_78; bra.uni $L__BB107_79; $L__BB107_78: rcp.approx.ftz.f32 %f8503, %f488; mul.ftz.f32 %f28938, %f28938, %f8503; mul.ftz.f32 %f28937, %f28937, %f8503; mul.ftz.f32 %f28936, %f28936, %f8503; $L__BB107_79: setp.eq.ftz.f32 %p74, %f28936, 0f00000000; setp.eq.ftz.f32 %p75, %f28937, 0f00000000; and.pred %p76, %p74, %p75; setp.eq.ftz.f32 %p77, %f28938, 0f00000000; and.pred %p78, %p77, %p76; selp.f32 %f8504, %f28920, %f28938, %p78; selp.f32 %f8505, %f28919, %f28937, %p78; selp.f32 %f8506, %f28918, %f28936, %p78; st.local.f32 [%rd5+-76], %f8506; st.local.f32 [%rd5+-72], %f8505; st.local.f32 [%rd5+-68], %f8504; $L__BB107_80: st.local.f32 [%rd5+28], %f431; st.local.f32 [%rd5+32], %f432; st.local.f32 [%rd5+36], %f433; st.local.v2.f32 [%rd5+40], {%f426, %f428}; st.local.f32 [%rd5+48], %f430; $L__BB107_81: and.b32 %r1674, %r5556, 4; setp.eq.s32 %p79, %r1674, 0; @%p79 bra $L__BB107_83; bra.uni $L__BB107_82; $L__BB107_83: and.b32 %r40, %r5556, 2; setp.eq.s32 %p80, %r40, 0; @%p80 bra $L__BB107_85; bra.uni $L__BB107_84; $L__BB107_85: ld.const.u64 %rd1258, [kernel_params+2160]; cvta.to.global.u64 %rd1259, %rd1258; ld.local.u32 %r1675, [%rd5+-12]; mul.wide.s32 %rd1260, %r1675, 208; add.s64 %rd1261, %rd1259, %rd1260; ld.global.v4.f32 {%f8548, %f8549, %f8550, %f8551}, [%rd1261+48]; ld.global.v4.f32 {%f8555, %f8556, %f8557, %f8558}, [%rd1261+64]; ld.global.v4.f32 {%f8562, %f8563, %f8564, %f8565}, [%rd1261+80]; ld.local.v2.f32 {%f8569, %f8570}, [%rd5+-72]; ld.local.f32 %f8573, [%rd5+-76]; mul.ftz.f32 %f8574, %f8555, %f8569; fma.rn.ftz.f32 %f8575, %f8548, %f8573, %f8574; fma.rn.ftz.f32 %f8576, %f8562, %f8570, %f8575; mul.ftz.f32 %f8577, %f8556, %f8569; fma.rn.ftz.f32 %f8578, %f8549, %f8573, %f8577; fma.rn.ftz.f32 %f8579, %f8563, %f8570, %f8578; mul.ftz.f32 %f8580, %f8557, %f8569; fma.rn.ftz.f32 %f8581, %f8550, %f8573, %f8580; fma.rn.ftz.f32 %f8582, %f8564, %f8570, %f8581; mul.ftz.f32 %f8583, %f8579, %f8579; fma.rn.ftz.f32 %f8584, %f8576, %f8576, %f8583; fma.rn.ftz.f32 %f8585, %f8582, %f8582, %f8584; sqrt.approx.ftz.f32 %f8586, %f8585; div.approx.ftz.f32 %f28939, %f8576, %f8586; div.approx.ftz.f32 %f28940, %f8579, %f8586; div.approx.ftz.f32 %f28941, %f8582, %f8586; bra.uni $L__BB107_86; $L__BB107_82: ld.local.v2.f32 {%f28948, %f28949}, [%rd5+-64]; bra.uni $L__BB107_95; $L__BB107_55: cvt.u32.u64 %r1649, %rd4; setp.gt.s32 %p63, %r25, %r24; selp.b32 %r1650, -1, 0, %p63; add.s32 %r1651, %r25, %r1650; ld.const.u64 %rd1204, [kernel_params+2344]; cvta.to.global.u64 %rd1205, %rd1204; mad.lo.s32 %r1652, %r1651, %r23, %r1649; add.s32 %r1653, %r1652, %r5555; mul.wide.s32 %rd1206, %r1653, 16; add.s64 %rd4733, %rd1205, %rd1206; $L__BB107_57: ld.global.v4.f32 {%f8356, %f8357, %f8358, %f8359}, [%rd4733]; add.s32 %r36, %r25, 1; setp.eq.s32 %p64, %r36, %r24; @%p64 bra $L__BB107_59; bra.uni $L__BB107_58; $L__BB107_59: ld.const.u64 %rd1213, [kernel_params+2296]; cvta.to.global.u64 %rd1214, %rd1213; shl.b64 %rd1215, %rd4, 4; add.s64 %rd4734, %rd1214, %rd1215; bra.uni $L__BB107_60; $L__BB107_84: ld.local.v2.f32 {%f8509, %f8510}, [%rd5+-72]; ld.local.v4.f32 {%f8513, %f8514, %f8515, %f8516}, [%rd5+104]; ld.local.v4.f32 {%f8520, %f8521, %f8522, %f8523}, [%rd5+120]; ld.local.v4.f32 {%f8527, %f8528, %f8529, %f8530}, [%rd5+136]; ld.local.f32 %f8534, [%rd5+-76]; mul.ftz.f32 %f8535, %f8509, %f8520; fma.rn.ftz.f32 %f8536, %f8534, %f8513, %f8535; fma.rn.ftz.f32 %f8537, %f8510, %f8527, %f8536; mul.ftz.f32 %f8538, %f8509, %f8521; fma.rn.ftz.f32 %f8539, %f8534, %f8514, %f8538; fma.rn.ftz.f32 %f8540, %f8510, %f8528, %f8539; mul.ftz.f32 %f8541, %f8509, %f8522; fma.rn.ftz.f32 %f8542, %f8534, %f8515, %f8541; fma.rn.ftz.f32 %f8543, %f8510, %f8529, %f8542; mul.ftz.f32 %f8544, %f8540, %f8540; fma.rn.ftz.f32 %f8545, %f8537, %f8537, %f8544; fma.rn.ftz.f32 %f8546, %f8543, %f8543, %f8545; sqrt.approx.ftz.f32 %f8547, %f8546; div.approx.ftz.f32 %f28939, %f8537, %f8547; div.approx.ftz.f32 %f28940, %f8540, %f8547; div.approx.ftz.f32 %f28941, %f8543, %f8547; $L__BB107_86: st.local.f32 [%rd5+-76], %f28939; st.local.f32 [%rd5+-72], %f28940; st.local.f32 [%rd5+-68], %f28941; @%p80 bra $L__BB107_88; bra.uni $L__BB107_87; $L__BB107_88: ld.const.u64 %rd1262, [kernel_params+2160]; cvta.to.global.u64 %rd1263, %rd1262; ld.local.u32 %r1676, [%rd5+-12]; mul.wide.s32 %rd1264, %r1676, 208; add.s64 %rd1265, %rd1263, %rd1264; ld.global.v4.f32 {%f8626, %f8627, %f8628, %f8629}, [%rd1265+48]; ld.global.v4.f32 {%f8633, %f8634, %f8635, %f8636}, [%rd1265+64]; ld.global.v4.f32 {%f8640, %f8641, %f8642, %f8643}, [%rd1265+80]; ld.local.v2.f32 {%f8647, %f8648}, [%rd5+-64]; mul.ftz.f32 %f8651, %f8633, %f8648; fma.rn.ftz.f32 %f8652, %f8626, %f8647, %f8651; ld.local.f32 %f8653, [%rd5+-56]; fma.rn.ftz.f32 %f8654, %f8640, %f8653, %f8652; mul.ftz.f32 %f8655, %f8634, %f8648; fma.rn.ftz.f32 %f8656, %f8627, %f8647, %f8655; fma.rn.ftz.f32 %f8657, %f8641, %f8653, %f8656; mul.ftz.f32 %f8658, %f8635, %f8648; fma.rn.ftz.f32 %f8659, %f8628, %f8647, %f8658; fma.rn.ftz.f32 %f8660, %f8642, %f8653, %f8659; mul.ftz.f32 %f8661, %f8657, %f8657; fma.rn.ftz.f32 %f8662, %f8654, %f8654, %f8661; fma.rn.ftz.f32 %f8663, %f8660, %f8660, %f8662; sqrt.approx.ftz.f32 %f8664, %f8663; div.approx.ftz.f32 %f28948, %f8654, %f8664; div.approx.ftz.f32 %f28949, %f8657, %f8664; div.approx.ftz.f32 %f28944, %f8660, %f8664; bra.uni $L__BB107_89; $L__BB107_58: cvt.u32.u64 %r1654, %rd4; setp.ge.s32 %p65, %r25, %r24; selp.b32 %r1655, -1, 0, %p65; add.s32 %r1656, %r36, %r1655; ld.const.u64 %rd1210, [kernel_params+2344]; cvta.to.global.u64 %rd1211, %rd1210; mad.lo.s32 %r1657, %r1656, %r23, %r1654; add.s32 %r1658, %r1657, %r5555; mul.wide.s32 %rd1212, %r1658, 16; add.s64 %rd4734, %rd1211, %rd1212; $L__BB107_60: cvt.rn.f32.s32 %f8360, %r25; sub.ftz.f32 %f8361, %f394, %f8360; ld.global.v4.f32 {%f8362, %f8363, %f8364, %f8365}, [%rd4734]; mov.f32 %f8370, 0f3F800000; sub.ftz.f32 %f8371, %f8370, %f8361; mul.ftz.f32 %f8372, %f8361, %f8362; mul.ftz.f32 %f8373, %f8361, %f8363; mul.ftz.f32 %f8374, %f8361, %f8364; fma.rn.ftz.f32 %f28915, %f8371, %f8358, %f8374; fma.rn.ftz.f32 %f28916, %f8371, %f8357, %f8373; fma.rn.ftz.f32 %f28917, %f8371, %f8356, %f8372; $L__BB107_62: and.b32 %r1659, %r5556, 4; setp.ne.s32 %p66, %r1659, 0; @%p66 bra $L__BB107_66; @%p24 bra $L__BB107_65; bra.uni $L__BB107_64; $L__BB107_65: ld.const.u64 %rd1220, [kernel_params+2160]; cvta.to.global.u64 %rd1221, %rd1220; mul.lo.s64 %rd1222, %rd7, 208; add.s64 %rd1223, %rd1221, %rd1222; ld.global.v4.f32 {%f8389, %f8390, %f8391, %f8392}, [%rd1223]; ld.global.v4.f32 {%f8397, %f8398, %f8399, %f8400}, [%rd1223+16]; ld.global.v4.f32 {%f8405, %f8406, %f8407, %f8408}, [%rd1223+32]; mul.ftz.f32 %f8413, %f28916, %f8390; fma.rn.ftz.f32 %f8414, %f28917, %f8389, %f8413; fma.rn.ftz.f32 %f8415, %f28915, %f8391, %f8414; add.ftz.f32 %f417, %f8392, %f8415; mul.ftz.f32 %f8416, %f28916, %f8398; fma.rn.ftz.f32 %f8417, %f28917, %f8397, %f8416; fma.rn.ftz.f32 %f8418, %f28915, %f8399, %f8417; add.ftz.f32 %f418, %f8400, %f8418; mul.ftz.f32 %f8419, %f28916, %f8406; fma.rn.ftz.f32 %f8420, %f28917, %f8405, %f8419; fma.rn.ftz.f32 %f8421, %f28915, %f8407, %f8420; add.ftz.f32 %f28915, %f8408, %f8421; mov.f32 %f28916, %f418; mov.f32 %f28917, %f417; bra.uni $L__BB107_66; $L__BB107_87: ld.local.v2.f32 {%f8587, %f8588}, [%rd5+-64]; ld.local.v4.f32 {%f8591, %f8592, %f8593, %f8594}, [%rd5+104]; ld.local.v4.f32 {%f8598, %f8599, %f8600, %f8601}, [%rd5+120]; ld.local.v4.f32 {%f8605, %f8606, %f8607, %f8608}, [%rd5+136]; mul.ftz.f32 %f8612, %f8588, %f8598; fma.rn.ftz.f32 %f8613, %f8587, %f8591, %f8612; ld.local.f32 %f8614, [%rd5+-56]; fma.rn.ftz.f32 %f8615, %f8614, %f8605, %f8613; mul.ftz.f32 %f8616, %f8588, %f8599; fma.rn.ftz.f32 %f8617, %f8587, %f8592, %f8616; fma.rn.ftz.f32 %f8618, %f8614, %f8606, %f8617; mul.ftz.f32 %f8619, %f8588, %f8600; fma.rn.ftz.f32 %f8620, %f8587, %f8593, %f8619; fma.rn.ftz.f32 %f8621, %f8614, %f8607, %f8620; mul.ftz.f32 %f8622, %f8618, %f8618; fma.rn.ftz.f32 %f8623, %f8615, %f8615, %f8622; fma.rn.ftz.f32 %f8624, %f8621, %f8621, %f8623; sqrt.approx.ftz.f32 %f8625, %f8624; div.approx.ftz.f32 %f28948, %f8615, %f8625; div.approx.ftz.f32 %f28949, %f8618, %f8625; div.approx.ftz.f32 %f28944, %f8621, %f8625; $L__BB107_89: st.local.v2.f32 [%rd5+-64], {%f28948, %f28949}; st.local.f32 [%rd5+-56], %f28944; @%p80 bra $L__BB107_91; bra.uni $L__BB107_90; $L__BB107_91: ld.const.u64 %rd1266, [kernel_params+2160]; cvta.to.global.u64 %rd1267, %rd1266; ld.local.u32 %r1677, [%rd5+-12]; mul.wide.s32 %rd1268, %r1677, 208; add.s64 %rd1269, %rd1267, %rd1268; ld.global.v4.f32 {%f8697, %f8698, %f8699, %f8700}, [%rd1269]; ld.global.v4.f32 {%f8704, %f8705, %f8706, %f8707}, [%rd1269+16]; ld.global.v4.f32 {%f8711, %f8712, %f8713, %f8714}, [%rd1269+32]; ld.local.v2.f32 {%f8718, %f8719}, [%rd5+32]; ld.local.f32 %f8722, [%rd5+28]; mul.ftz.f32 %f8723, %f8698, %f8718; fma.rn.ftz.f32 %f8724, %f8722, %f8697, %f8723; fma.rn.ftz.f32 %f28945, %f8699, %f8719, %f8724; mul.ftz.f32 %f8725, %f8705, %f8718; fma.rn.ftz.f32 %f8726, %f8722, %f8704, %f8725; fma.rn.ftz.f32 %f28946, %f8706, %f8719, %f8726; mul.ftz.f32 %f8727, %f8712, %f8718; fma.rn.ftz.f32 %f8728, %f8722, %f8711, %f8727; fma.rn.ftz.f32 %f28947, %f8713, %f8719, %f8728; bra.uni $L__BB107_92; $L__BB107_64: mul.ftz.f32 %f8380, %f28916, %f142; fma.rn.ftz.f32 %f8381, %f28917, %f141, %f8380; fma.rn.ftz.f32 %f8382, %f28915, %f143, %f8381; add.ftz.f32 %f414, %f144, %f8382; mul.ftz.f32 %f8383, %f28916, %f146; fma.rn.ftz.f32 %f8384, %f28917, %f145, %f8383; fma.rn.ftz.f32 %f8385, %f28915, %f147, %f8384; add.ftz.f32 %f415, %f148, %f8385; mul.ftz.f32 %f8386, %f28916, %f150; fma.rn.ftz.f32 %f8387, %f28917, %f149, %f8386; fma.rn.ftz.f32 %f8388, %f28915, %f151, %f8387; add.ftz.f32 %f28915, %f152, %f8388; mov.f32 %f28916, %f415; mov.f32 %f28917, %f414; $L__BB107_66: ld.local.v4.f32 {%f8422, %f8423, %f8424, %f8425}, [%rd1194]; sub.ftz.f32 %f8429, %f8422, %f28917; sub.ftz.f32 %f8430, %f8423, %f28916; sub.ftz.f32 %f8431, %f8424, %f28915; mul.ftz.f32 %f8432, %f8430, %f8430; fma.rn.ftz.f32 %f8433, %f8429, %f8429, %f8432; fma.rn.ftz.f32 %f8434, %f8431, %f8431, %f8433; sqrt.approx.ftz.f32 %f8435, %f8434; div.approx.ftz.f32 %f8436, %f8431, %f8435; st.local.f32 [%rd5+-56], %f8436; div.approx.ftz.f32 %f28948, %f8429, %f8435; st.local.f32 [%rd5+-76], %f28948; div.approx.ftz.f32 %f28949, %f8430, %f8435; st.local.v4.f32 [%rd5+-72], {%f28949, %f8436, %f28948, %f28949}; mov.u32 %r1660, 0; st.local.u32 [%rd5+28], %r1660; mov.u64 %rd1226, 0; st.local.u64 [%rd5+32], %rd1226; mov.f32 %f8437, 0f00000000; st.local.v2.f32 [%rd5+40], {%f8437, %f8437}; st.local.u32 [%rd5+48], %r1660; $L__BB107_95: ld.const.u64 %rd1274, [kernel_params+2408]; cvta.to.global.u64 %rd26, %rd1274; ld.local.u32 %rd1275, [%rd5+-40]; cvt.u32.u64 %r1679, %rd1275; and.b32 %r5558, %r1679, 4194303; shl.b64 %rd1276, %rd1275, 5; and.b64 %rd1277, %rd1276, 134217696; add.s64 %rd1278, %rd26, %rd1277; ld.global.u32 %r5618, [%rd1278+16]; st.local.u32 [%rd5+-36], %r5618; ld.local.v4.f32 {%f8803, %f8804, %f8805, %f8806}, [%rd5+-56]; mul.ftz.f32 %f8807, %f28949, %f8805; fma.rn.ftz.f32 %f8808, %f28948, %f8804, %f8807; fma.rn.ftz.f32 %f8809, %f8803, %f8806, %f8808; setp.lt.ftz.f32 %p84, %f8809, 0f00000000; @%p84 bra $L__BB107_97; bra.uni $L__BB107_96; $L__BB107_97: or.b32 %r5618, %r5618, 1; st.local.u32 [%rd5+-36], %r5618; neg.ftz.f32 %f28952, %f8803; neg.ftz.f32 %f28949, %f28949; neg.ftz.f32 %f28948, %f28948; st.local.f32 [%rd5+-56], %f28952; ld.local.f32 %f8814, [%rd5+-76]; neg.ftz.f32 %f8815, %f8814; ld.local.v2.f32 {%f8816, %f8817}, [%rd5+-72]; st.local.f32 [%rd5+-76], %f8815; neg.ftz.f32 %f8820, %f8817; neg.ftz.f32 %f8821, %f8816; st.local.v4.f32 [%rd5+-72], {%f8821, %f8820, %f28948, %f28949}; ld.local.f32 %f8822, [%rd5+28]; neg.ftz.f32 %f28964, %f8822; ld.local.v2.f32 {%f8823, %f8824}, [%rd5+32]; neg.ftz.f32 %f28965, %f8823; neg.ftz.f32 %f8827, %f8824; st.local.f32 [%rd5+28], %f28964; st.local.f32 [%rd5+32], %f28965; st.local.f32 [%rd5+36], %f8827; ld.local.v4.f32 {%f8828, %f8829, %f8830, %f8831}, [%rd5+40]; neg.ftz.f32 %f28955, %f8830; neg.ftz.f32 %f28963, %f8829; neg.ftz.f32 %f28962, %f8828; st.local.v2.f32 [%rd5+40], {%f28962, %f28963}; st.local.f32 [%rd5+48], %f28955; bra.uni $L__BB107_98; $L__BB107_96: ld.local.f32 %f28964, [%rd5+28]; ld.local.f32 %f28965, [%rd5+32]; ld.local.v4.f32 {%f28962, %f28963, %f28955, %f8813}, [%rd5+40]; mov.b32 %r1680, %f8803; mov.b64 %rd1279, {%r1680, %r1681}; cvt.u32.u64 %r1682, %rd1279; mov.b32 %f28952, %r1682; $L__BB107_98: ld.local.f32 %f564, [%rd5]; fma.rn.ftz.f32 %f8835, %f12, %f564, %f11; st.local.f32 [%rd5+4], %f8835; st.local.f32 [%rd5+8], %f12; setp.neu.ftz.f32 %p85, %f28948, %f28949; setp.neu.ftz.f32 %p86, %f28948, %f28952; or.pred %p87, %p85, %p86; sub.ftz.f32 %f8836, %f28952, %f28949; neg.ftz.f32 %f8837, %f28949; neg.ftz.f32 %f8838, %f28952; selp.f32 %f8839, %f8838, %f28952, %p87; add.ftz.f32 %f8840, %f28948, %f8839; selp.f32 %f8841, %f28949, %f8837, %p87; sub.ftz.f32 %f8842, %f8841, %f28948; mul.ftz.f32 %f8843, %f8840, %f8840; fma.rn.ftz.f32 %f8844, %f8836, %f8836, %f8843; fma.rn.ftz.f32 %f8845, %f8842, %f8842, %f8844; sqrt.approx.ftz.f32 %f8846, %f8845; div.approx.ftz.f32 %f8847, %f8836, %f8846; div.approx.ftz.f32 %f8848, %f8840, %f8846; div.approx.ftz.f32 %f8849, %f8842, %f8846; mul.ftz.f32 %f8850, %f28949, %f8849; mul.ftz.f32 %f8851, %f28952, %f8848; sub.ftz.f32 %f8852, %f8850, %f8851; mul.ftz.f32 %f8853, %f28952, %f8847; mul.ftz.f32 %f8854, %f28948, %f8849; sub.ftz.f32 %f8855, %f8853, %f8854; mul.ftz.f32 %f8856, %f28948, %f8848; mul.ftz.f32 %f8857, %f28949, %f8847; sub.ftz.f32 %f8858, %f8856, %f8857; mul.ftz.f32 %f28958, %f8835, %f8847; mul.ftz.f32 %f28959, %f8835, %f8848; mul.ftz.f32 %f567, %f8835, %f8849; mul.ftz.f32 %f28960, %f8835, %f8852; mul.ftz.f32 %f28961, %f8835, %f8855; mul.ftz.f32 %f570, %f8835, %f8858; abs.ftz.f32 %f573, %f28948; abs.ftz.f32 %f574, %f28949; abs.ftz.f32 %f575, %f28952; setp.geu.ftz.f32 %p88, %f575, %f573; setp.geu.ftz.f32 %p89, %f575, %f574; and.pred %p90, %p88, %p89; @%p90 bra $L__BB107_100; ld.local.f32 %f576, [%rd5+36]; setp.lt.ftz.f32 %p91, %f574, %f575; setp.lt.ftz.f32 %p92, %f574, %f573; or.pred %p93, %p92, %p91; selp.f32 %f28958, %f28959, %f28958, %p93; selp.f32 %f28960, %f28961, %f28960, %p93; selp.f32 %f28962, %f28963, %f28962, %p93; selp.f32 %f28964, %f28965, %f28964, %p93; mov.f32 %f28959, %f567; mov.f32 %f28961, %f570; mov.f32 %f28963, %f28955; mov.f32 %f28965, %f576; $L__BB107_100: mul.ftz.f32 %f8859, %f28962, %f28965; mul.ftz.f32 %f8860, %f28963, %f28964; sub.ftz.f32 %f28966, %f8860, %f8859; setp.eq.ftz.f32 %p94, %f28966, 0f00000000; @%p94 bra $L__BB107_102; rcp.approx.ftz.f32 %f28966, %f28966; $L__BB107_102: mul.ftz.f32 %f8861, %f28959, %f28962; mul.ftz.f32 %f8862, %f28958, %f28963; sub.ftz.f32 %f8863, %f8862, %f8861; mul.ftz.f32 %f8864, %f8863, %f28966; st.local.f32 [%rd5+12], %f8864; mul.ftz.f32 %f8865, %f28958, %f28965; mul.ftz.f32 %f8866, %f28959, %f28964; sub.ftz.f32 %f8867, %f8866, %f8865; mul.ftz.f32 %f8868, %f28961, %f28962; mul.ftz.f32 %f8869, %f28960, %f28963; sub.ftz.f32 %f8870, %f8869, %f8868; mul.ftz.f32 %f8871, %f8867, %f28966; mul.ftz.f32 %f8872, %f8870, %f28966; st.local.v2.f32 [%rd5+16], {%f8872, %f8871}; mul.ftz.f32 %f8873, %f28960, %f28965; mul.ftz.f32 %f8874, %f28961, %f28964; sub.ftz.f32 %f8875, %f8874, %f8873; mul.ftz.f32 %f8876, %f8875, %f28966; st.local.f32 [%rd5+24], %f8876; ld.const.u64 %rd1280, [kernel_params+2544]; cvta.to.global.u64 %rd1281, %rd1280; shl.b64 %rd1282, %rd3, 2; add.s64 %rd27, %rd1281, %rd1282; ld.global.u32 %r45, [%rd27]; and.b32 %r1683, %r5618, 524288; setp.eq.s32 %p95, %r1683, 0; @%p95 bra $L__BB107_110; bra.uni $L__BB107_103; $L__BB107_110: and.b32 %r46, %r45, 15728640; setp.ne.s32 %p101, %r46, 0; and.b32 %r1688, %r5618, 2097152; setp.eq.s32 %p102, %r1688, 0; and.pred %p103, %p102, %p101; @%p103 bra $L__BB107_116; and.b32 %r1689, %r45, 1016576; setp.eq.s32 %p104, %r1689, 0; mov.u32 %r1690, 0; ld.const.u32 %r1691, [kernel_params+4]; selp.b32 %r1692, %r1691, 0, %p104; st.local.u32 [%rd5+172], %r1690; st.local.u32 [%rd5+176], %r1692; ld.const.u64 %rd1303, [kernel_params+2400]; cvta.to.global.u64 %rd28, %rd1303; and.b32 %r1693, %r45, 24576; setp.ne.s32 %p105, %r1693, 16384; setp.eq.s64 %p106, %rd1059, 0; or.pred %p1, %p106, %p105; ld.const.u64 %rd1304, [kernel_params+2448]; cvta.to.global.u64 %rd1305, %rd1304; add.s64 %rd29, %rd1305, %rd1282; ld.const.s32 %rd30, [kernel_params+1264]; ld.const.s32 %rd31, [kernel_params+1436]; ld.const.s32 %rd32, [kernel_params+1440]; add.u64 %rd1307, %SP, 192; add.u64 %rd33, %SPL, 192; bra.uni $L__BB107_112; $L__BB107_103: cvt.u16.u32 %rs306, %r45; setp.lt.s16 %p96, %rs306, 0; mov.f32 %f28967, 0f00000000; @%p96 bra $L__BB107_105; ld.const.u64 %rd1283, [kernel_params+2576]; cvta.to.global.u64 %rd1284, %rd1283; add.s64 %rd1286, %rd1284, %rd1282; ld.global.f32 %f28967, [%rd1286]; $L__BB107_105: setp.eq.ftz.f32 %p97, %f28967, 0f00000000; mov.u16 %rs748, 0; @%p97 bra $L__BB107_4545; setp.eq.ftz.f32 %p98, %f28967, 0f3F800000; @%p98 bra $L__BB107_108; ld.const.u64 %rd1287, [kernel_params+2584]; cvta.to.global.u64 %rd1288, %rd1287; mul.lo.s64 %rd1289, %rd3, 12; add.s64 %rd1290, %rd1288, %rd1289; ld.global.f32 %f8878, [%rd1290]; rcp.approx.ftz.f32 %f8879, %f28967; mul.ftz.f32 %f8880, %f8879, %f8878; ld.global.f32 %f8881, [%rd1290+4]; mul.ftz.f32 %f8882, %f8879, %f8881; ld.global.f32 %f8883, [%rd1290+8]; mul.ftz.f32 %f8884, %f8879, %f8883; st.global.f32 [%rd1290], %f8880; st.global.f32 [%rd1290+4], %f8882; st.global.f32 [%rd1290+8], %f8884; $L__BB107_108: mov.u32 %r5886, 0; ld.const.u64 %rd1291, [kernel_params+2512]; cvta.to.global.u64 %rd1292, %rd1291; shl.b64 %rd1293, %rd3, 1; add.s64 %rd1294, %rd1292, %rd1293; ld.global.u16 %rs308, [%rd1294]; add.s16 %rs309, %rs308, 1; st.global.u16 [%rd1294], %rs309; setp.gt.u16 %p99, %rs308, 1023; @%p99 bra $L__BB107_4529; ld.const.u64 %rd1295, [kernel_params+2536]; cvta.to.global.u64 %rd1296, %rd1295; add.s64 %rd1298, %rd1296, %rd1293; ld.global.u16 %rs310, [%rd1298]; add.s16 %rs311, %rs310, 16; st.global.u16 [%rd1298], %rs311; mov.b32 %r1686, %f564; add.s32 %r1687, %r1686, 1; mov.b32 %f8885, %r1687; setp.eq.ftz.f32 %p100, %f564, 0f00000000; selp.f32 %f8886, 0f00000001, %f8885, %p100; ld.const.u64 %rd1299, [kernel_params+2648]; cvta.to.global.u64 %rd1300, %rd1299; add.s64 %rd1302, %rd1300, %rd1282; st.global.f32 [%rd1302], %f8886; mov.u32 %r5886, 33; bra.uni $L__BB107_4529; $L__BB107_309: { // callseq 245, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 245 add.s32 %r5558, %r47, 1; $L__BB107_112: mov.u32 %r47, %r5558; mul.wide.s32 %rd1308, %r47, 16; add.s64 %rd34, %rd28, %rd1308; ld.global.v4.u32 {%r1695, %r1696, %r1697, %r1698}, [%rd34]; mul.wide.u32 %rd1309, %r1696, 4; add.s64 %rd35, %rd33, %rd1309; mul.wide.u32 %rd1310, %r1697, 4; add.s64 %rd36, %rd33, %rd1310; setp.gt.s32 %p107, %r1695, 1; @%p107 bra $L__BB107_144; setp.eq.s32 %p237, %r1695, 1; mov.u32 %r5558, %r1696; @%p237 bra $L__BB107_112; bra.uni $L__BB107_114; $L__BB107_144: setp.gt.s32 %p108, %r1695, 49; @%p108 bra $L__BB107_225; bra.uni $L__BB107_145; $L__BB107_225: setp.gt.s32 %p109, %r1695, 73; @%p109 bra $L__BB107_265; bra.uni $L__BB107_226; $L__BB107_265: setp.gt.s32 %p110, %r1695, 85; @%p110 bra $L__BB107_288; bra.uni $L__BB107_266; $L__BB107_288: setp.gt.s32 %p111, %r1695, 91; @%p111 bra $L__BB107_298; bra.uni $L__BB107_289; $L__BB107_298: setp.gt.s32 %p112, %r1695, 94; @%p112 bra $L__BB107_303; bra.uni $L__BB107_299; $L__BB107_303: setp.gt.s32 %p113, %r1695, 96; @%p113 bra $L__BB107_307; setp.eq.s32 %p116, %r1695, 95; @%p116 bra $L__BB107_310; setp.eq.s32 %p117, %r1695, 96; @%p117 bra $L__BB107_306; bra.uni $L__BB107_114; $L__BB107_306: { // callseq 246, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 246 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_145: setp.gt.s32 %p174, %r1695, 25; @%p174 bra $L__BB107_186; setp.gt.s32 %p206, %r1695, 13; @%p206 bra $L__BB107_167; setp.gt.s32 %p222, %r1695, 7; @%p222 bra $L__BB107_157; setp.gt.s32 %p230, %r1695, 4; @%p230 bra $L__BB107_153; setp.eq.s32 %p234, %r1695, 2; @%p234 bra $L__BB107_376; setp.eq.s32 %p235, %r1695, 3; @%p235 bra $L__BB107_375; setp.eq.s32 %p236, %r1695, 4; @%p236 bra $L__BB107_152; bra.uni $L__BB107_114; $L__BB107_152: mov.u32 %r1772, 4; { // callseq 324, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1772, %r1696, %r1697, %r1698}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 324 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_226: setp.gt.s32 %p143, %r1695, 61; @%p143 bra $L__BB107_246; setp.gt.s32 %p159, %r1695, 55; @%p159 bra $L__BB107_237; setp.gt.s32 %p167, %r1695, 52; @%p167 bra $L__BB107_233; setp.eq.s32 %p171, %r1695, 50; @%p171 bra $L__BB107_343; setp.eq.s32 %p172, %r1695, 51; @%p172 bra $L__BB107_342; setp.eq.s32 %p173, %r1695, 52; @%p173 bra $L__BB107_232; bra.uni $L__BB107_114; $L__BB107_232: add.s32 %r1735, %r47, 1; { // callseq 285, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1735; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 285 bra.uni $L__BB107_112; $L__BB107_186: setp.gt.s32 %p175, %r1695, 37; @%p175 bra $L__BB107_206; setp.gt.s32 %p191, %r1695, 31; @%p191 bra $L__BB107_197; setp.gt.s32 %p199, %r1695, 28; @%p199 bra $L__BB107_193; setp.eq.s32 %p203, %r1695, 26; @%p203 bra $L__BB107_359; setp.eq.s32 %p204, %r1695, 27; @%p204 bra $L__BB107_358; setp.eq.s32 %p205, %r1695, 28; @%p205 bra $L__BB107_192; bra.uni $L__BB107_114; $L__BB107_192: mov.u32 %r1748, 28; { // callseq 307, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1748; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 307 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_266: setp.gt.s32 %p128, %r1695, 79; @%p128 bra $L__BB107_279; setp.gt.s32 %p136, %r1695, 76; @%p136 bra $L__BB107_275; setp.eq.s32 %p140, %r1695, 74; @%p140 bra $L__BB107_327; setp.eq.s32 %p141, %r1695, 75; @%p141 bra $L__BB107_326; setp.eq.s32 %p142, %r1695, 76; @%p142 bra $L__BB107_271; bra.uni $L__BB107_114; $L__BB107_271: add.s32 %r5558, %r47, 1; setp.eq.s32 %p239, %r1698, 255; @%p239 bra $L__BB107_112; add.s64 %rd37, %rd35, 4; add.s64 %rd38, %rd35, 8; setp.eq.s32 %p240, %r1697, 0; mul.wide.u32 %rd1356, %r1698, 4; add.s64 %rd39, %rd33, %rd1356; @%p240 bra $L__BB107_325; setp.eq.s32 %p241, %r1697, 1; @%p241 bra $L__BB107_324; bra.uni $L__BB107_274; $L__BB107_324: ld.local.f32 %f8903, [%rd37]; st.local.f32 [%rd39], %f8903; bra.uni $L__BB107_112; $L__BB107_246: setp.gt.s32 %p144, %r1695, 67; @%p144 bra $L__BB107_256; setp.gt.s32 %p152, %r1695, 64; @%p152 bra $L__BB107_252; setp.eq.s32 %p156, %r1695, 62; @%p156 bra $L__BB107_335; setp.eq.s32 %p157, %r1695, 63; @%p157 bra $L__BB107_334; setp.eq.s32 %p158, %r1695, 64; @%p158 bra $L__BB107_251; bra.uni $L__BB107_114; $L__BB107_251: mov.u32 %r1721, 64; add.s32 %r1722, %r47, 1; { // callseq 273, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1721; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1722; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 273 bra.uni $L__BB107_112; $L__BB107_167: setp.gt.s32 %p207, %r1695, 19; @%p207 bra $L__BB107_177; setp.gt.s32 %p215, %r1695, 16; @%p215 bra $L__BB107_173; setp.eq.s32 %p219, %r1695, 14; @%p219 bra $L__BB107_367; setp.eq.s32 %p220, %r1695, 15; @%p220 bra $L__BB107_366; setp.eq.s32 %p221, %r1695, 16; @%p221 bra $L__BB107_172; bra.uni $L__BB107_114; $L__BB107_172: mov.u32 %r1758, 16; { // callseq 318, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1758; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 318 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_206: setp.gt.s32 %p176, %r1695, 43; @%p176 bra $L__BB107_216; setp.gt.s32 %p184, %r1695, 40; @%p184 bra $L__BB107_212; setp.eq.s32 %p188, %r1695, 38; @%p188 bra $L__BB107_351; setp.eq.s32 %p189, %r1695, 39; @%p189 bra $L__BB107_350; setp.eq.s32 %p190, %r1695, 40; @%p190 bra $L__BB107_211; bra.uni $L__BB107_114; $L__BB107_211: add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_279: setp.gt.s32 %p129, %r1695, 82; @%p129 bra $L__BB107_284; setp.eq.s32 %p133, %r1695, 80; @%p133 bra $L__BB107_320; setp.eq.s32 %p134, %r1695, 81; @%p134 bra $L__BB107_319; setp.eq.s32 %p135, %r1695, 82; @%p135 bra $L__BB107_283; bra.uni $L__BB107_114; $L__BB107_283: mov.u32 %r1708, 82; { // callseq 257, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1708; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 257 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_289: setp.gt.s32 %p121, %r1695, 88; @%p121 bra $L__BB107_294; setp.eq.s32 %p125, %r1695, 86; @%p125 bra $L__BB107_316; setp.eq.s32 %p126, %r1695, 87; @%p126 bra $L__BB107_315; setp.eq.s32 %p127, %r1695, 88; @%p127 bra $L__BB107_293; bra.uni $L__BB107_114; $L__BB107_293: mov.u32 %r1704, 88; ld.local.f32 %f8896, [%rd5+-76]; ld.local.v2.f32 {%f8897, %f8898}, [%rd5+-72]; { // callseq 251, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f8896; .param .b32 param1; st.param.f32 [param1+0], %f8897; .param .b32 param2; st.param.f32 [param2+0], %f8898; .param .b64 param3; st.param.b64 [param3+0], %rd1307; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r1704; st.param.b32 [param4+4], %r1696; st.param.b32 [param4+8], %r1697; st.param.b32 [param4+12], %r1698; call.uni _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 251 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_157: setp.gt.s32 %p223, %r1695, 10; @%p223 bra $L__BB107_163; setp.eq.s32 %p227, %r1695, 8; @%p227 bra $L__BB107_372; setp.eq.s32 %p228, %r1695, 9; @%p228 bra $L__BB107_370; setp.eq.s32 %p229, %r1695, 10; @%p229 bra $L__BB107_161; bra.uni $L__BB107_114; $L__BB107_161: ld.local.f32 %f8929, [%rd36]; setp.ltu.ftz.f32 %p242, %f8929, 0f3F800000; add.s32 %r5558, %r47, 1; @%p242 bra $L__BB107_112; add.s32 %r1768, %r1696, %r47; add.s32 %r5558, %r1768, 1; bra.uni $L__BB107_112; $L__BB107_237: setp.gt.s32 %p160, %r1695, 58; @%p160 bra $L__BB107_242; setp.eq.s32 %p164, %r1695, 56; @%p164 bra $L__BB107_339; setp.eq.s32 %p165, %r1695, 57; @%p165 bra $L__BB107_338; setp.eq.s32 %p166, %r1695, 58; @%p166 bra $L__BB107_241; bra.uni $L__BB107_114; $L__BB107_241: mov.u32 %r1730, 58; { // callseq 279, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1730; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 279 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_197: setp.gt.s32 %p192, %r1695, 34; @%p192 bra $L__BB107_202; setp.eq.s32 %p196, %r1695, 32; @%p196 bra $L__BB107_355; setp.eq.s32 %p197, %r1695, 33; @%p197 bra $L__BB107_354; setp.eq.s32 %p198, %r1695, 34; @%p198 bra $L__BB107_201; bra.uni $L__BB107_114; $L__BB107_201: { // callseq 302, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 302 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_256: setp.gt.s32 %p145, %r1695, 70; @%p145 bra $L__BB107_261; setp.eq.s32 %p149, %r1695, 68; @%p149 bra $L__BB107_331; setp.eq.s32 %p150, %r1695, 69; @%p150 bra $L__BB107_330; setp.eq.s32 %p151, %r1695, 70; @%p151 bra $L__BB107_260; bra.uni $L__BB107_114; $L__BB107_260: mov.u32 %r1715, 70; { // callseq 267, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1715; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 267 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_177: setp.gt.s32 %p208, %r1695, 22; @%p208 bra $L__BB107_182; setp.eq.s32 %p212, %r1695, 20; @%p212 bra $L__BB107_363; setp.eq.s32 %p213, %r1695, 21; @%p213 bra $L__BB107_362; setp.eq.s32 %p214, %r1695, 22; @%p214 bra $L__BB107_181; bra.uni $L__BB107_114; $L__BB107_181: mov.u32 %r1755, 22; add.s32 %r1756, %r47, 1; { // callseq 313, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1755; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; .param .b32 param3; st.param.b32 [param3+0], %r1756; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 313 bra.uni $L__BB107_112; $L__BB107_216: setp.gt.s32 %p177, %r1695, 46; @%p177 bra $L__BB107_221; setp.eq.s32 %p181, %r1695, 44; @%p181 bra $L__BB107_347; setp.eq.s32 %p182, %r1695, 45; @%p182 bra $L__BB107_346; setp.eq.s32 %p183, %r1695, 46; @%p183 bra $L__BB107_220; bra.uni $L__BB107_114; $L__BB107_220: { // callseq 291, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 291 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_275: setp.eq.s32 %p137, %r1695, 77; @%p137 bra $L__BB107_322; setp.eq.s32 %p138, %r1695, 78; @%p138 bra $L__BB107_321; setp.eq.s32 %p139, %r1695, 79; @%p139 bra $L__BB107_278; bra.uni $L__BB107_114; $L__BB107_278: add.s32 %r1710, %r47, 1; { // callseq 260, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1710; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 260 bra.uni $L__BB107_112; $L__BB107_284: setp.eq.s32 %p130, %r1695, 83; @%p130 bra $L__BB107_318; setp.eq.s32 %p131, %r1695, 84; @%p131 bra $L__BB107_317; setp.eq.s32 %p132, %r1695, 85; @%p132 bra $L__BB107_287; bra.uni $L__BB107_114; $L__BB107_287: add.s32 %r1707, %r47, 1; { // callseq 254, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1707; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 254 bra.uni $L__BB107_112; $L__BB107_153: setp.eq.s32 %p231, %r1695, 5; @%p231 bra $L__BB107_374; setp.eq.s32 %p232, %r1695, 6; @%p232 bra $L__BB107_373; setp.eq.s32 %p233, %r1695, 7; @%p233 bra $L__BB107_156; bra.uni $L__BB107_114; $L__BB107_156: mov.u32 %r1771, 7; { // callseq 323, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1771; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 323 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_233: setp.eq.s32 %p168, %r1695, 53; @%p168 bra $L__BB107_341; setp.eq.s32 %p169, %r1695, 54; @%p169 bra $L__BB107_340; setp.eq.s32 %p170, %r1695, 55; @%p170 bra $L__BB107_236; bra.uni $L__BB107_114; $L__BB107_236: cvta.to.local.u64 %rd1381, %rd1061; ld.local.v4.f32 {%f8905, %f8906, %f8907, %f8908}, [%rd1381]; { // callseq 282, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f8905; st.param.f32 [param0+4], %f8906; st.param.f32 [param0+8], %f8907; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; .param .b32 param4; st.param.b32 [param4+0], %r1698; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 282 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_193: setp.eq.s32 %p200, %r1695, 29; @%p200 bra $L__BB107_357; setp.eq.s32 %p201, %r1695, 30; @%p201 bra $L__BB107_356; setp.eq.s32 %p202, %r1695, 31; @%p202 bra $L__BB107_196; bra.uni $L__BB107_114; $L__BB107_196: mov.u32 %r1746, 31; add.s32 %r1747, %r47, 1; { // callseq 304, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd1307; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1746; st.param.b32 [param3+4], %r1696; st.param.b32 [param3+8], %r1697; st.param.b32 [param3+12], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1747; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 304 bra.uni $L__BB107_112; $L__BB107_252: setp.eq.s32 %p153, %r1695, 65; @%p153 bra $L__BB107_333; setp.eq.s32 %p154, %r1695, 66; @%p154 bra $L__BB107_332; setp.eq.s32 %p155, %r1695, 67; @%p155 bra $L__BB107_255; bra.uni $L__BB107_114; $L__BB107_255: mov.u32 %r1719, 67; { // callseq 270, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1719, %r1696, %r1697, %r1698}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 270 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_173: setp.eq.s32 %p216, %r1695, 17; @%p216 bra $L__BB107_365; setp.eq.s32 %p217, %r1695, 18; @%p217 bra $L__BB107_364; setp.eq.s32 %p218, %r1695, 19; @%p218 bra $L__BB107_176; bra.uni $L__BB107_114; $L__BB107_176: { // callseq 315, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 315 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_212: setp.eq.s32 %p185, %r1695, 41; @%p185 bra $L__BB107_349; setp.eq.s32 %p186, %r1695, 42; @%p186 bra $L__BB107_348; setp.eq.s32 %p187, %r1695, 43; @%p187 bra $L__BB107_215; bra.uni $L__BB107_114; $L__BB107_215: add.s32 %r1739, %r47, 1; { // callseq 294, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1739; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 294 bra.uni $L__BB107_112; $L__BB107_294: setp.eq.s32 %p122, %r1695, 89; @%p122 bra $L__BB107_314; setp.eq.s32 %p123, %r1695, 90; @%p123 bra $L__BB107_313; setp.eq.s32 %p124, %r1695, 91; @%p124 bra $L__BB107_297; bra.uni $L__BB107_114; $L__BB107_297: add.s32 %r5558, %r47, 1; @%p1 bra $L__BB107_114; bra.uni $L__BB107_112; $L__BB107_163: setp.eq.s32 %p224, %r1695, 11; @%p224 bra $L__BB107_369; setp.eq.s32 %p225, %r1695, 12; @%p225 bra $L__BB107_368; setp.eq.s32 %p226, %r1695, 13; @%p226 bra $L__BB107_166; bra.uni $L__BB107_114; $L__BB107_166: mov.u32 %r1766, 13; add.s32 %r1767, %r47, 1; { // callseq 319, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd1307; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1766; st.param.b32 [param3+4], %r1696; st.param.b32 [param3+8], %r1697; st.param.b32 [param3+12], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1767; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 319 bra.uni $L__BB107_112; $L__BB107_242: setp.eq.s32 %p161, %r1695, 59; @%p161 bra $L__BB107_337; setp.eq.s32 %p162, %r1695, 60; @%p162 bra $L__BB107_336; setp.eq.s32 %p163, %r1695, 61; @%p163 bra $L__BB107_245; bra.uni $L__BB107_114; $L__BB107_245: mov.u32 %r1726, 61; add.s32 %r1727, %r47, 1; { // callseq 276, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1726; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 276 bra.uni $L__BB107_112; $L__BB107_202: setp.eq.s32 %p193, %r1695, 35; @%p193 bra $L__BB107_353; setp.eq.s32 %p194, %r1695, 36; @%p194 bra $L__BB107_352; setp.eq.s32 %p195, %r1695, 37; @%p195 bra $L__BB107_205; bra.uni $L__BB107_114; $L__BB107_205: mov.u32 %r1742, 37; { // callseq 299, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1742, %r1696, %r1697, %r1698}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 299 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_261: setp.eq.s32 %p146, %r1695, 71; @%p146 bra $L__BB107_329; setp.eq.s32 %p147, %r1695, 72; @%p147 bra $L__BB107_328; setp.eq.s32 %p148, %r1695, 73; @%p148 bra $L__BB107_264; bra.uni $L__BB107_114; $L__BB107_264: add.s32 %r1713, %r47, 1; { // callseq 264, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1713; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 264 bra.uni $L__BB107_112; $L__BB107_182: setp.eq.s32 %p209, %r1695, 23; @%p209 bra $L__BB107_361; setp.eq.s32 %p210, %r1695, 24; @%p210 bra $L__BB107_360; setp.eq.s32 %p211, %r1695, 25; @%p211 bra $L__BB107_185; bra.uni $L__BB107_114; $L__BB107_185: add.s32 %r1751, %r47, 1; { // callseq 310, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1751; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 310 bra.uni $L__BB107_112; $L__BB107_221: setp.eq.s32 %p178, %r1695, 47; @%p178 bra $L__BB107_345; setp.eq.s32 %p179, %r1695, 48; @%p179 bra $L__BB107_344; setp.eq.s32 %p180, %r1695, 49; @%p180 bra $L__BB107_224; bra.uni $L__BB107_114; $L__BB107_224: { // callseq 288, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 288 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_299: setp.eq.s32 %p118, %r1695, 92; @%p118 bra $L__BB107_312; setp.eq.s32 %p119, %r1695, 93; @%p119 bra $L__BB107_311; setp.eq.s32 %p120, %r1695, 94; @%p120 bra $L__BB107_302; bra.uni $L__BB107_114; $L__BB107_302: mov.u32 %r1699, 94; add.s32 %r1700, %r47, 1; { // callseq 248, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1699; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1700; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 248 bra.uni $L__BB107_112; $L__BB107_307: setp.eq.s32 %p114, %r1695, 98; @%p114 bra $L__BB107_377; setp.ne.s32 %p115, %r1695, 97; @%p115 bra $L__BB107_114; bra.uni $L__BB107_309; $L__BB107_310: { // callseq 247, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 247 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_376: mov.u32 %r1774, 2; add.s32 %r1775, %r47, 1; { // callseq 326, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1774; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; .param .b32 param3; st.param.b32 [param3+0], %r45; .param .b32 param4; st.param.b32 [param4+0], %r1775; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 326 bra.uni $L__BB107_112; $L__BB107_375: mov.u32 %r1773, 3; { // callseq 325, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1773, %r1696, %r1697, %r1698}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 325 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_343: { // callseq 287, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 287 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_342: { // callseq 286, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 286 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_359: mov.u32 %r1750, 26; { // callseq 309, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1750; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 309 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_358: mov.u32 %r1749, 27; { // callseq 308, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1749; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 308 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_327: { // callseq 263, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 263 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_326: { // callseq 262, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 262 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_335: mov.u32 %r1724, 62; add.s32 %r1725, %r47, 1; { // callseq 275, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1724; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1725; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 275 bra.uni $L__BB107_112; $L__BB107_334: mov.u32 %r1723, 63; { // callseq 274, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1723; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 274 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_367: st.local.u32 [%rd36], %r1696; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_366: ld.global.v4.u32 {%r1759, %r1760, %r1761, %r1762}, [%rd34+16]; add.s32 %r5558, %r47, 2; st.local.u32 [%rd35], %r1760; st.local.u32 [%rd35+4], %r1761; st.local.u32 [%rd35+8], %r1762; bra.uni $L__BB107_112; $L__BB107_351: { // callseq 298, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; .param .b32 param4; st.param.b32 [param4+0], %r1698; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 298 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_350: mov.u32 %r1741, 39; { // callseq 297, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1741, %r1696, %r1697, %r1698}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 297 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_320: { // callseq 259, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 259 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_319: mov.u32 %r1709, 81; { // callseq 258, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1709; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 258 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_316: add.s32 %r1706, %r47, 1; { // callseq 253, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1706; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 253 bra.uni $L__BB107_112; $L__BB107_315: add.s32 %r1705, %r47, 1; { // callseq 252, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1705; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 252 bra.uni $L__BB107_112; $L__BB107_372: mov.u32 %r1770, 8; { // callseq 322, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1770; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 322 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_370: ld.local.f32 %f8930, [%rd36]; setp.gtu.ftz.f32 %p243, %f8930, 0f00000000; add.s32 %r5558, %r47, 1; @%p243 bra $L__BB107_112; add.s32 %r1769, %r1696, %r47; add.s32 %r5558, %r1769, 1; bra.uni $L__BB107_112; $L__BB107_339: mov.u32 %r1733, 56; { // callseq 281, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1733; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 281 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_338: mov.u32 %r1731, 57; add.s32 %r1732, %r47, 1; { // callseq 280, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1731; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1732; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 280 bra.uni $L__BB107_112; $L__BB107_355: mov.u32 %r1744, 32; add.s32 %r1745, %r47, 1; { // callseq 303, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd1307; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1744; st.param.b32 [param3+4], %r1696; st.param.b32 [param3+8], %r1697; st.param.b32 [param3+12], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1745; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 303 bra.uni $L__BB107_112; $L__BB107_354: ld.local.f32 %f8913, [%rd35]; ld.local.f32 %f8914, [%rd35+4]; ld.local.f32 %f8915, [%rd35+8]; st.local.f32 [%rd5+-76], %f8913; st.local.f32 [%rd5+-72], %f8914; st.local.f32 [%rd5+-68], %f8915; st.local.f32 [%rd36], %f8913; st.local.f32 [%rd36+4], %f8914; st.local.f32 [%rd36+8], %f8915; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_331: mov.u32 %r1718, 68; { // callseq 269, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1718; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 269 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_330: mov.u32 %r1716, 69; add.s32 %r1717, %r47, 1; { // callseq 268, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1716; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1717; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 268 bra.uni $L__BB107_112; $L__BB107_363: ld.local.f32 %f8916, [%rd35]; cvta.to.local.u64 %rd1434, %rd1061; ld.local.v4.f32 {%f8917, %f8918, %f8919, %f8920}, [%rd1434]; ld.local.f32 %f8924, [%rd35+4]; ld.local.f32 %f8925, [%rd35+8]; add.ftz.f32 %f8926, %f8925, %f8919; add.ftz.f32 %f8927, %f8924, %f8918; add.ftz.f32 %f8928, %f8916, %f8917; st.local.v2.f32 [%rd1434], {%f8928, %f8927}; st.local.f32 [%rd5+-80], %f8926; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_362: mov.u32 %r1757, 21; { // callseq 314, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1757; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 314 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_347: mov.u32 %r1737, 44; add.s32 %r1738, %r47, 1; { // callseq 293, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1737; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1738; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 293 bra.uni $L__BB107_112; $L__BB107_346: { // callseq 292, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 292 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_322: setp.eq.s32 %p238, %r1698, 255; add.s32 %r5558, %r47, 1; @%p238 bra $L__BB107_112; ld.local.f32 %f8901, [%rd35]; add.s32 %r1712, %r1698, %r1697; mul.wide.u32 %rd1352, %r1712, 4; add.s64 %rd1353, %rd33, %rd1352; st.local.f32 [%rd1353], %f8901; bra.uni $L__BB107_112; $L__BB107_321: add.s32 %r1711, %r47, 1; { // callseq 261, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1711; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 261 bra.uni $L__BB107_112; $L__BB107_318: { // callseq 256, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 256 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_317: { // callseq 255, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 255 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_374: st.local.u32 [%rd5+180], %r1696; st.local.u32 [%rd5+184], %r1697; st.local.u32 [%rd5+188], %r1698; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_373: ld.local.f32 %f8931, [%rd35]; ld.local.f32 %f8932, [%rd35+4]; ld.local.f32 %f8933, [%rd35+8]; st.local.f32 [%rd5+180], %f8931; st.local.f32 [%rd5+184], %f8932; st.local.f32 [%rd5+188], %f8933; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_341: { // callseq 284, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 284 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_340: add.s32 %r1734, %r47, 1; { // callseq 283, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1734; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 283 bra.uni $L__BB107_112; $L__BB107_357: { // callseq 306, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; .param .b32 param4; st.param.b32 [param4+0], %r1698; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 306 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_356: { // callseq 305, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; .param .b32 param4; st.param.b32 [param4+0], %r1698; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 305 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_333: { // callseq 272, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 272 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_332: add.s32 %r1720, %r47, 1; { // callseq 271, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1720; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 271 bra.uni $L__BB107_112; $L__BB107_365: { // callseq 317, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; .param .b32 param4; st.param.b32 [param4+0], %r1698; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 317 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_364: { // callseq 316, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 316 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_349: add.s32 %r1740, %r47, 1; { // callseq 296, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1740; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 296 bra.uni $L__BB107_112; $L__BB107_348: { // callseq 295, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 295 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_314: mov.u32 %r1703, 89; { // callseq 250, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1703; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 250 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_313: mov.u32 %r1701, 90; add.s32 %r1702, %r47, 1; { // callseq 249, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1701; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; .param .b32 param3; st.param.b32 [param3+0], %r1702; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 249 bra.uni $L__BB107_112; $L__BB107_369: { // callseq 321, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 321 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_368: { // callseq 320, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 320 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_337: add.s32 %r1729, %r47, 1; { // callseq 278, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1729; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 278 bra.uni $L__BB107_112; $L__BB107_336: add.s32 %r1728, %r47, 1; { // callseq 277, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; .param .b32 param4; st.param.b32 [param4+0], %r1728; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 277 bra.uni $L__BB107_112; $L__BB107_353: { // callseq 301, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 301 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_352: mov.u32 %r1743, 36; { // callseq 300, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1743; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 300 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_329: mov.u32 %r1714, 71; { // callseq 266, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1714; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 266 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_328: { // callseq 265, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; .param .b32 param3; st.param.b32 [param3+0], %r1698; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 265 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_361: mov.u32 %r1753, 23; add.s32 %r1754, %r47, 1; { // callseq 312, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1753; st.param.b32 [param1+4], %r1696; st.param.b32 [param1+8], %r1697; st.param.b32 [param1+12], %r1698; .param .b32 param2; st.param.b32 [param2+0], %r1754; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5558, [retval0+0]; } // callseq 312 bra.uni $L__BB107_112; $L__BB107_360: mov.u32 %r1752, 24; { // callseq 311, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1752; st.param.b32 [param2+4], %r1696; st.param.b32 [param2+8], %r1697; st.param.b32 [param2+12], %r1698; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 311 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_345: ld.local.u32 %r1736, [%rd5+-36]; ld.local.f32 %f8912, [%rd5]; { // callseq 290, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1602; .param .b32 param1; st.param.b32 [param1+0], %r1736; .param .b32 param2; st.param.f32 [param2+0], %f8912; .param .b64 param3; st.param.b64 [param3+0], %rd1307; .param .b32 param4; st.param.b32 [param4+0], %r1696; .param .b32 param5; st.param.b32 [param5+0], %r1697; .param .b32 param6; st.param.b32 [param6+0], %r45; call.uni _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5, param6 ); } // callseq 290 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_344: { // callseq 289, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .b64 param1; st.param.b64 [param1+0], %rd1307; .param .b32 param2; st.param.b32 [param2+0], %r1696; .param .b32 param3; st.param.b32 [param3+0], %r1697; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 289 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_311: ld.local.f32 %f8887, [%rd35]; ld.global.u32 %rd1315, [%rd29]; mul.lo.s64 %rd1316, %rd30, %rd1315; cvt.s64.s32 %rd1317, %r1697; add.s64 %rd1318, %rd32, %rd1317; add.s64 %rd1319, %rd1318, %rd1316; cvta.to.global.u64 %rd1320, %rd1059; shl.b64 %rd1321, %rd1319, 2; add.s64 %rd1322, %rd1320, %rd1321; atom.global.add.f32 %f8888, [%rd1322], %f8887; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_377: { // callseq 327, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1307; .param .b32 param1; st.param.b32 [param1+0], %r1696; .param .b32 param2; st.param.b32 [param2+0], %r1697; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 327 add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_312: ld.local.f32 %f8889, [%rd35]; ld.local.f32 %f8890, [%rd35+4]; ld.local.f32 %f8891, [%rd35+8]; ld.global.u32 %rd1323, [%rd29]; mul.lo.s64 %rd1324, %rd30, %rd1323; cvt.s64.s32 %rd1325, %r1697; add.s64 %rd1326, %rd31, %rd1325; add.s64 %rd1327, %rd1326, %rd1324; cvta.to.global.u64 %rd1328, %rd1059; shl.b64 %rd1329, %rd1327, 2; add.s64 %rd1330, %rd1328, %rd1329; add.s64 %rd1331, %rd1330, 4; add.s64 %rd1332, %rd1330, 8; add.s64 %rd1333, %rd1330, 12; atom.global.add.f32 %f8892, [%rd1330], %f8889; atom.global.add.f32 %f8893, [%rd1331], %f8890; atom.global.add.f32 %f8894, [%rd1332], %f8891; atom.global.add.f32 %f8895, [%rd1333], 0f3F800000; add.s32 %r5558, %r47, 1; bra.uni $L__BB107_112; $L__BB107_325: ld.local.f32 %f8904, [%rd35]; st.local.f32 [%rd39], %f8904; bra.uni $L__BB107_112; $L__BB107_274: ld.local.f32 %f8902, [%rd38]; st.local.f32 [%rd39], %f8902; bra.uni $L__BB107_112; $L__BB107_114: ld.local.u32 %r5618, [%rd5+-36]; and.b32 %r1776, %r5618, 1024; setp.eq.s32 %p244, %r1776, 0; @%p244 bra $L__BB107_116; ld.const.v2.u64 {%rd1458, %rd1459}, [kernel_params+2528]; cvta.to.global.u64 %rd1462, %rd1458; add.s64 %rd1464, %rd1462, %rd1282; cvta.to.global.u64 %rd1465, %rd1459; shl.b64 %rd1466, %rd3, 1; add.s64 %rd1467, %rd1465, %rd1466; ld.global.u16 %r1777, [%rd1467]; ld.const.u64 %rd1468, [kernel_params+2456]; cvta.to.global.u64 %rd1469, %rd1468; add.s64 %rd1470, %rd1469, %rd1282; ld.global.u32 %r1778, [%rd1464]; add.s32 %r1779, %r1778, %r1777; ld.global.u32 %r1780, [%rd1470]; mad.lo.s32 %r1781, %r1780, -1262732973, %r1779; mad.lo.s32 %r1782, %r1781, 1103515245, 12345; st.local.u32 [%rd5+168], %r1782; $L__BB107_116: setp.eq.s32 %p245, %r46, 0; @%p245 bra $L__BB107_386; ld.local.v2.f32 {%f28999, %f29000}, [%rd5+-72]; ld.local.f32 %f28998, [%rd5+-76]; and.b32 %r1783, %r5618, 2097152; setp.eq.s32 %p246, %r1783, 0; @%p246 bra $L__BB107_136; ld.local.u32 %r156, [%rd5+172]; setp.lt.s32 %p247, %r156, 1; mov.f32 %f28971, 0f00000000; mov.f32 %f28972, %f28971; mov.f32 %f28973, %f28971; @%p247 bra $L__BB107_134; add.s32 %r1785, %r156, -1; and.b32 %r5564, %r156, 3; setp.lt.u32 %p248, %r1785, 3; mov.f32 %f28973, 0f00000000; mov.u32 %r5562, 0; mov.f32 %f28972, %f28973; mov.f32 %f28971, %f28973; @%p248 bra $L__BB107_130; sub.s32 %r5561, %r156, %r5564; mov.f32 %f28973, 0f00000000; mov.u32 %r5562, 0; mov.f32 %f28972, %f28973; mov.f32 %f28971, %f28973; $L__BB107_121: cvt.s64.s32 %rd40, %r5562; mul.wide.s32 %rd1471, %r5562, 80; add.s64 %rd1472, %rd2, %rd1471; add.s64 %rd41, %rd1472, 316; ld.local.u32 %r1787, [%rd1472+316]; add.s32 %r1788, %r1787, -32; setp.gt.u32 %p249, %r1788, 2; @%p249 bra $L__BB107_123; ld.local.v4.f32 {%f8946, %f8947, %f8948, %f8949}, [%rd41+-12]; add.ftz.f32 %f8953, %f8946, %f8947; add.ftz.f32 %f8954, %f8948, %f8953; mul.ftz.f32 %f8955, %f8954, 0f3EAAAAAB; abs.ftz.f32 %f8956, %f8955; ld.local.f32 %f8957, [%rd41+8]; ld.local.f32 %f8958, [%rd41+12]; ld.local.f32 %f8959, [%rd41+16]; fma.rn.ftz.f32 %f28971, %f8956, %f8957, %f28971; fma.rn.ftz.f32 %f28972, %f8956, %f8958, %f28972; fma.rn.ftz.f32 %f28973, %f8956, %f8959, %f28973; $L__BB107_123: ld.local.u32 %r1789, [%rd41+80]; add.s32 %r1790, %r1789, -32; setp.gt.u32 %p250, %r1790, 2; @%p250 bra $L__BB107_125; ld.local.v4.f32 {%f8960, %f8961, %f8962, %f8963}, [%rd41+68]; add.ftz.f32 %f8967, %f8960, %f8961; add.ftz.f32 %f8968, %f8962, %f8967; mul.ftz.f32 %f8969, %f8968, 0f3EAAAAAB; abs.ftz.f32 %f8970, %f8969; ld.local.f32 %f8971, [%rd41+88]; ld.local.f32 %f8972, [%rd41+92]; ld.local.f32 %f8973, [%rd41+96]; fma.rn.ftz.f32 %f28971, %f8970, %f8971, %f28971; fma.rn.ftz.f32 %f28972, %f8970, %f8972, %f28972; fma.rn.ftz.f32 %f28973, %f8970, %f8973, %f28973; $L__BB107_125: ld.local.u32 %r1791, [%rd41+160]; add.s32 %r1792, %r1791, -32; setp.gt.u32 %p251, %r1792, 2; @%p251 bra $L__BB107_127; ld.local.v4.f32 {%f8974, %f8975, %f8976, %f8977}, [%rd41+148]; add.ftz.f32 %f8981, %f8974, %f8975; add.ftz.f32 %f8982, %f8976, %f8981; mul.ftz.f32 %f8983, %f8982, 0f3EAAAAAB; abs.ftz.f32 %f8984, %f8983; ld.local.f32 %f8985, [%rd41+168]; ld.local.f32 %f8986, [%rd41+172]; ld.local.f32 %f8987, [%rd41+176]; fma.rn.ftz.f32 %f28971, %f8984, %f8985, %f28971; fma.rn.ftz.f32 %f28972, %f8984, %f8986, %f28972; fma.rn.ftz.f32 %f28973, %f8984, %f8987, %f28973; $L__BB107_127: ld.local.u32 %r1793, [%rd41+240]; add.s32 %r1794, %r1793, -32; setp.gt.u32 %p252, %r1794, 2; @%p252 bra $L__BB107_129; ld.local.v4.f32 {%f8988, %f8989, %f8990, %f8991}, [%rd41+228]; add.ftz.f32 %f8995, %f8988, %f8989; add.ftz.f32 %f8996, %f8990, %f8995; mul.ftz.f32 %f8997, %f8996, 0f3EAAAAAB; abs.ftz.f32 %f8998, %f8997; ld.local.f32 %f8999, [%rd41+248]; ld.local.f32 %f9000, [%rd41+252]; ld.local.f32 %f9001, [%rd41+256]; fma.rn.ftz.f32 %f28971, %f8998, %f8999, %f28971; fma.rn.ftz.f32 %f28972, %f8998, %f9000, %f28972; fma.rn.ftz.f32 %f28973, %f8998, %f9001, %f28973; $L__BB107_129: cvt.u32.u64 %r1795, %rd40; add.s32 %r5562, %r1795, 4; add.s32 %r5561, %r5561, -4; setp.ne.s32 %p253, %r5561, 0; @%p253 bra $L__BB107_121; $L__BB107_130: setp.eq.s32 %p254, %r5564, 0; @%p254 bra $L__BB107_134; $L__BB107_131: .pragma "nounroll"; cvt.s64.s32 %rd42, %r5562; mul.wide.s32 %rd1473, %r5562, 80; add.s64 %rd1474, %rd2, %rd1473; add.s64 %rd43, %rd1474, 316; ld.local.u32 %r1796, [%rd1474+316]; add.s32 %r1797, %r1796, -32; setp.gt.u32 %p255, %r1797, 2; @%p255 bra $L__BB107_133; ld.local.f32 %f9002, [%rd43+-8]; ld.local.f32 %f9003, [%rd43+-12]; add.ftz.f32 %f9004, %f9003, %f9002; ld.local.f32 %f9005, [%rd43+-4]; add.ftz.f32 %f9006, %f9005, %f9004; mul.ftz.f32 %f9007, %f9006, 0f3EAAAAAB; abs.ftz.f32 %f9008, %f9007; ld.local.f32 %f9009, [%rd43+8]; ld.local.f32 %f9010, [%rd43+12]; ld.local.f32 %f9011, [%rd43+16]; fma.rn.ftz.f32 %f28971, %f9008, %f9009, %f28971; fma.rn.ftz.f32 %f28972, %f9008, %f9010, %f28972; fma.rn.ftz.f32 %f28973, %f9008, %f9011, %f28973; $L__BB107_133: cvt.u32.u64 %r1798, %rd42; add.s32 %r5562, %r1798, 1; add.s32 %r5564, %r5564, -1; setp.ne.s32 %p256, %r5564, 0; @%p256 bra $L__BB107_131; $L__BB107_134: setp.eq.ftz.f32 %p257, %f28972, 0f00000000; setp.eq.ftz.f32 %p258, %f28971, 0f00000000; and.pred %p259, %p258, %p257; setp.eq.ftz.f32 %p260, %f28973, 0f00000000; and.pred %p261, %p260, %p259; @%p261 bra $L__BB107_136; mul.ftz.f32 %f9012, %f28972, %f28972; fma.rn.ftz.f32 %f9013, %f28971, %f28971, %f9012; fma.rn.ftz.f32 %f9014, %f28973, %f28973, %f9013; sqrt.approx.ftz.f32 %f9015, %f9014; div.approx.ftz.f32 %f29000, %f28973, %f9015; div.approx.ftz.f32 %f28999, %f28972, %f9015; div.approx.ftz.f32 %f28998, %f28971, %f9015; $L__BB107_136: and.b32 %r1799, %r5618, -1279; st.local.u32 [%rd5+-36], %r1799; mov.u32 %r1800, 0; st.local.u32 [%rd5+172], %r1800; ld.const.u32 %r168, [kernel_params+4]; st.local.u32 [%rd5+176], %r168; and.b32 %r1801, %r45, 4194304; setp.eq.s32 %p262, %r1801, 0; mov.f32 %f9016, 0f3F800000; abs.ftz.f32 %f649, %f9016; @%p262 bra $L__BB107_378; bra.uni $L__BB107_137; $L__BB107_378: setp.ltu.ftz.f32 %p267, %f649, 0f3727C5AC; mov.u64 %rd4741, 0; mov.u64 %rd4742, %rd4741; @%p267 bra $L__BB107_383; setp.eq.s32 %p268, %r168, 0; mov.u64 %rd4741, 0; mov.u64 %rd4739, %rd4741; mov.u64 %rd4740, %rd4741; @%p268 bra $L__BB107_381; add.s64 %rd4739, %rd2, 304; add.s64 %rd4740, %rd1061, 304; mov.u64 %rd1489, 1065353216; st.local.u64 [%rd5+224], %rd1489; mov.u64 %rd1490, 4575657222473777152; st.local.u64 [%rd5+216], %rd1490; mov.u32 %r1809, 1; st.local.u32 [%rd5+172], %r1809; add.s32 %r1810, %r168, -1; st.local.u32 [%rd5+176], %r1810; $L__BB107_381: setp.eq.s64 %p269, %rd4740, 0; mov.u64 %rd4742, %rd4741; @%p269 bra $L__BB107_383; st.local.f32 [%rd4739+16], %f649; mov.u64 %rd4741, %rd4739; mov.u64 %rd4742, %rd4740; $L__BB107_383: setp.eq.s64 %p270, %rd4742, 0; @%p270 bra $L__BB107_385; st.local.f32 [%rd4741+20], %f28998; st.local.f32 [%rd4741+24], %f28999; st.local.f32 [%rd4741+28], %f29000; mov.u32 %r1811, 2; st.local.u32 [%rd4741+12], %r1811; ld.local.u32 %r1812, [%rd5+-36]; or.b32 %r1813, %r1812, 12; st.local.u32 [%rd5+-36], %r1813; bra.uni $L__BB107_385; $L__BB107_386: ld.const.u32 %r169, [kernel_params+1624]; setp.eq.s32 %p271, %r169, 0; @%p271 bra $L__BB107_520; and.b32 %r1816, %r169, 1; setp.eq.b32 %p272, %r1816, 1; mov.pred %p273, 0; xor.pred %p274, %p272, %p273; not.pred %p275, %p274; @%p275 bra $L__BB107_389; mov.u32 %r1817, 0; mov.f32 %f9017, 0f00000000; st.local.v2.f32 [%rd5+192], {%f9017, %f9017}; st.local.u32 [%rd5+200], %r1817; $L__BB107_389: and.b32 %r1818, %r169, 32; setp.eq.s32 %p276, %r1818, 0; @%p276 bra $L__BB107_391; and.b32 %r5618, %r5618, -9; st.local.u32 [%rd5+-36], %r5618; $L__BB107_391: and.b32 %r1819, %r45, 1; setp.eq.b32 %p277, %r1819, 1; xor.pred %p279, %p277, %p273; not.pred %p280, %p279; @%p280 bra $L__BB107_520; ld.local.u32 %r172, [%rd5+172]; setp.lt.s32 %p281, %r172, 1; @%p281 bra $L__BB107_520; and.b32 %r1820, %r169, 2; setp.eq.s32 %p282, %r1820, 0; and.b32 %r173, %r169, 4; and.b32 %r174, %r169, 8; and.b32 %r175, %r169, 16; add.s32 %r176, %r172, -1; and.b32 %r177, %r172, 3; @%p282 bra $L__BB107_462; bra.uni $L__BB107_394; $L__BB107_462: setp.eq.s32 %p377, %r173, 0; @%p377 bra $L__BB107_494; bra.uni $L__BB107_463; $L__BB107_494: setp.lt.u32 %p427, %r176, 3; mov.u32 %r5613, 0; @%p427 bra $L__BB107_513; sub.s32 %r5607, %r172, %r177; mov.u32 %r5613, 0; setp.eq.s32 %p429, %r174, 0; setp.eq.s32 %p431, %r175, 0; $L__BB107_496: cvt.s64.s32 %rd68, %r5613; mul.wide.s32 %rd1505, %r5613, 80; add.s64 %rd1506, %rd2, %rd1505; add.s64 %rd69, %rd1506, 316; ld.local.u32 %r266, [%rd1506+316]; add.s32 %r1915, %r266, -21; setp.gt.u32 %p428, %r1915, 9; or.pred %p430, %p428, %p429; @%p430 bra $L__BB107_498; bra.uni $L__BB107_497; $L__BB107_498: setp.ne.s32 %p432, %r266, 31; or.pred %p433, %p432, %p431; @%p433 bra $L__BB107_500; mov.u32 %r1917, 35; st.local.u32 [%rd69], %r1917; mov.u32 %r1918, 0; st.local.u32 [%rd69+4], %r1918; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_500; $L__BB107_497: mov.u32 %r1916, 0; st.local.u32 [%rd69], %r1916; st.local.u32 [%rd69+4], %r1916; $L__BB107_500: ld.local.u32 %r269, [%rd69+80]; add.s32 %r1919, %r269, -21; setp.gt.u32 %p434, %r1919, 9; or.pred %p436, %p434, %p429; @%p436 bra $L__BB107_502; bra.uni $L__BB107_501; $L__BB107_502: setp.ne.s32 %p438, %r269, 31; or.pred %p439, %p438, %p431; @%p439 bra $L__BB107_504; mov.u32 %r1921, 35; st.local.u32 [%rd69+80], %r1921; mov.u32 %r1922, 0; st.local.u32 [%rd69+84], %r1922; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_504; $L__BB107_501: mov.u32 %r1920, 0; st.local.u32 [%rd69+80], %r1920; st.local.u32 [%rd69+84], %r1920; $L__BB107_504: ld.local.u32 %r272, [%rd69+160]; add.s32 %r1923, %r272, -21; setp.gt.u32 %p440, %r1923, 9; or.pred %p442, %p440, %p429; @%p442 bra $L__BB107_506; bra.uni $L__BB107_505; $L__BB107_506: setp.ne.s32 %p444, %r272, 31; or.pred %p445, %p444, %p431; @%p445 bra $L__BB107_508; mov.u32 %r1925, 35; st.local.u32 [%rd69+160], %r1925; mov.u32 %r1926, 0; st.local.u32 [%rd69+164], %r1926; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_508; $L__BB107_505: mov.u32 %r1924, 0; st.local.u32 [%rd69+160], %r1924; st.local.u32 [%rd69+164], %r1924; $L__BB107_508: ld.local.u32 %r275, [%rd69+240]; add.s32 %r1927, %r275, -21; setp.gt.u32 %p446, %r1927, 9; or.pred %p448, %p446, %p429; @%p448 bra $L__BB107_510; bra.uni $L__BB107_509; $L__BB107_510: setp.ne.s32 %p450, %r275, 31; or.pred %p451, %p450, %p431; @%p451 bra $L__BB107_512; mov.u32 %r1929, 35; st.local.u32 [%rd69+240], %r1929; mov.u32 %r1930, 0; st.local.u32 [%rd69+244], %r1930; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_512; $L__BB107_509: mov.u32 %r1928, 0; st.local.u32 [%rd69+240], %r1928; st.local.u32 [%rd69+244], %r1928; $L__BB107_512: cvt.u32.u64 %r1931, %rd68; add.s32 %r5613, %r1931, 4; add.s32 %r5607, %r5607, -4; setp.ne.s32 %p452, %r5607, 0; @%p452 bra $L__BB107_496; $L__BB107_513: setp.eq.s32 %p453, %r177, 0; @%p453 bra $L__BB107_520; setp.eq.s32 %p455, %r174, 0; setp.eq.s32 %p457, %r175, 0; $L__BB107_515: .pragma "nounroll"; cvt.s64.s32 %rd70, %r5613; mul.wide.s32 %rd1507, %r5613, 80; add.s64 %rd1508, %rd2, %rd1507; add.s64 %rd71, %rd1508, 316; ld.local.u32 %r285, [%rd1508+316]; add.s32 %r1932, %r285, -21; setp.gt.u32 %p454, %r1932, 9; or.pred %p456, %p454, %p455; @%p456 bra $L__BB107_517; bra.uni $L__BB107_516; $L__BB107_517: setp.ne.s32 %p458, %r285, 31; or.pred %p459, %p458, %p457; @%p459 bra $L__BB107_519; mov.u32 %r1934, 35; st.local.u32 [%rd71], %r1934; mov.u32 %r1935, 0; st.local.u32 [%rd71+4], %r1935; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_519; $L__BB107_516: mov.u32 %r1933, 0; st.local.u32 [%rd71], %r1933; st.local.u32 [%rd71+4], %r1933; $L__BB107_519: cvt.u32.u64 %r1936, %rd70; add.s32 %r5613, %r1936, 1; add.s32 %r177, %r177, -1; setp.ne.s32 %p460, %r177, 0; @%p460 bra $L__BB107_515; bra.uni $L__BB107_520; $L__BB107_137: setp.ltu.ftz.f32 %p263, %f649, 0f3727C5AC; mov.u64 %rd4737, 0; mov.u64 %rd4738, %rd4737; @%p263 bra $L__BB107_142; setp.eq.s32 %p264, %r168, 0; mov.u64 %rd4737, 0; mov.u64 %rd4735, %rd4737; mov.u64 %rd4736, %rd4737; @%p264 bra $L__BB107_140; add.s64 %rd4735, %rd2, 304; add.s64 %rd4736, %rd1061, 304; mov.u64 %rd1480, 1065353216; st.local.u64 [%rd5+224], %rd1480; mov.u64 %rd1481, 4575657222473777152; st.local.u64 [%rd5+216], %rd1481; mov.u32 %r1802, 1; st.local.u32 [%rd5+172], %r1802; add.s32 %r1803, %r168, -1; st.local.u32 [%rd5+176], %r1803; $L__BB107_140: setp.eq.s64 %p265, %rd4736, 0; mov.u64 %rd4738, %rd4737; @%p265 bra $L__BB107_142; st.local.f32 [%rd4735+16], %f649; mov.u64 %rd4737, %rd4735; mov.u64 %rd4738, %rd4736; $L__BB107_142: setp.eq.s64 %p266, %rd4738, 0; @%p266 bra $L__BB107_385; st.local.f32 [%rd4737+20], %f28998; st.local.f32 [%rd4737+24], %f28999; st.local.f32 [%rd4737+28], %f29000; mov.u32 %r1804, 2139095039; st.local.u32 [%rd4737+32], %r1804; mov.u32 %r1805, 5; st.local.u32 [%rd4737+12], %r1805; mov.u32 %r1806, 4; st.local.u32 [%rd4737+36], %r1806; ld.local.u32 %r1807, [%rd5+-36]; or.b32 %r1808, %r1807, 12; st.local.u32 [%rd5+-36], %r1808; $L__BB107_385: ld.global.u32 %r1814, [%rd27]; and.b32 %r1815, %r1814, -15728641; st.global.u32 [%rd27], %r1815; $L__BB107_1312: ld.const.v2.u64 {%rd4921, %rd4920}, [kernel_params+2528]; cvta.to.global.u64 %rd2147, %rd4921; add.s64 %rd2149, %rd2147, %rd1282; ld.global.u32 %r601, [%rd2149]; cvta.to.global.u64 %rd2150, %rd4920; shl.b64 %rd2151, %rd3, 1; add.s64 %rd2152, %rd2150, %rd2151; ld.global.u16 %rs28, [%rd2152]; cvt.u32.u16 %r602, %rs28; ld.const.u64 %rd2153, [kernel_params+2456]; cvta.to.global.u64 %rd2154, %rd2153; add.s64 %rd2155, %rd2154, %rd1282; ld.global.u32 %r603, [%rd2155]; ld.const.u32 %r2543, [kernel_params+1488]; setp.eq.s32 %p1246, %r2543, 0; ld.local.u8 %rs329, [%rd5+-36]; and.b16 %rs330, %rs329, 8; setp.eq.s16 %p1247, %rs330, 0; add.u64 %rd2156, %SP, 192; add.u64 %rd2157, %SPL, 192; add.s64 %rd277, %rd2157, 8; or.pred %p1248, %p1246, %p1247; @%p1248 bra $L__BB107_3809; ld.global.u32 %r604, [%rd27]; ld.const.u64 %rd2158, [kernel_params+2464]; cvta.to.global.u64 %rd2159, %rd2158; add.s64 %rd278, %rd2159, %rd2151; ld.global.u16 %rs29, [%rd278]; add.s32 %r605, %r602, 1; ld.const.u32 %r606, [kernel_params+1596]; setp.eq.s32 %p1249, %r606, 0; brev.b32 %r2544, %r603; mul.lo.s32 %r2545, %r2544, 1025551850; xor.b32 %r607, %r2545, %r2544; @%p1249 bra $L__BB107_1316; ld.const.v2.u32 {%r2546, %r2547}, [kernel_params+1600]; mov.b32 %f1922, %r2547; setp.geu.ftz.f32 %p1250, %f1922, 0f3F800000; setp.lt.ftz.f32 %p1251, %f1922, 0f3F800000; ld.const.u32 %r2550, [kernel_params+1576]; selp.b32 %r2551, %r2550, %r601, %p1251; add.s16 %rs331, %rs28, 1; and.b16 %rs332, %rs331, 255; cvt.u32.u16 %r2552, %rs332; xor.b32 %r2553, %r2551, %r2552; mul.lo.s32 %r2554, %r2553, -512718531; shr.u32 %r2555, %r2551, 16; xor.b32 %r2556, %r2554, %r2555; bfe.u32 %r2557, %r2556, 4, 4; shr.u32 %r2558, %r2551, 8; xor.b32 %r2559, %r2556, %r2558; xor.b32 %r2560, %r2559, %r2557; mul.lo.s32 %r2561, %r2560, 153742143; shr.u32 %r2562, %r2551, 23; xor.b32 %r2563, %r2561, %r2562; bfe.u32 %r2564, %r2563, 1, 7; xor.b32 %r2565, %r2564, %r2563; shr.u32 %r2566, %r2551, 27; or.b32 %r2567, %r2566, 1; mul.lo.s32 %r2568, %r2567, %r2565; mul.lo.s32 %r2569, %r2568, -1074439621; bfe.u32 %r2570, %r2569, 2, 6; xor.b32 %r2571, %r2570, %r2569; mul.lo.s32 %r2572, %r2571, -1638916925; bfe.u32 %r2573, %r2572, 2, 6; xor.b32 %r2574, %r2573, %r2572; mul.lo.s32 %r2575, %r2574, -933190689; and.b32 %r2576, %r2575, 255; bfe.u32 %r2577, %r2575, 5, 3; xor.b32 %r2578, %r2577, %r2576; add.s32 %r2579, %r2546, -1; xor.b32 %r608, %r605, 61; xor.b32 %r2580, %r2551, %r608; mul.lo.s32 %r2581, %r2580, 9; shr.u32 %r2582, %r2581, 4; xor.b32 %r2583, %r2582, %r2581; mul.lo.s32 %r2584, %r2583, 668265261; add.s32 %r2585, %r2584, %r607; shr.u32 %r2586, %r2584, 16; or.b32 %r2587, %r2586, 1; mul.lo.s32 %r2588, %r2587, %r2585; mul.lo.s32 %r2589, %r2588, 89287766; xor.b32 %r2590, %r2589, %r2588; mul.lo.s32 %r2591, %r2590, 1403136100; xor.b32 %r2592, %r2591, %r2590; brev.b32 %r2593, %r2592; neg.s32 %r2594, %r2546; and.b32 %r2595, %r603, %r2594; and.b32 %r2596, %r2593, %r2579; or.b32 %r2597, %r2596, %r2595; mad.lo.s32 %r2598, %r2578, %r2546, %r2597; shl.b32 %r2599, %r2546, 8; rem.u32 %r2600, %r2598, %r2599; ld.const.u64 %rd2161, [kernel_params+2424]; cvta.to.global.u64 %rd2162, %rd2161; shl.b32 %r2601, %r2600, 1; mul.wide.u32 %rd2163, %r2601, 4; add.s64 %rd2164, %rd2162, %rd2163; ld.global.f32 %f29434, [%rd2164]; ld.global.f32 %f29435, [%rd2164+4]; @%p1250 bra $L__BB107_1320; xor.b32 %r2602, %r608, %r601; mul.lo.s32 %r2603, %r2602, 9; shr.u32 %r2604, %r2603, 4; xor.b32 %r2605, %r2604, %r2603; mul.lo.s32 %r2606, %r2605, 668265261; cvt.rn.f32.u32 %f13108, %r2606; mul.ftz.f32 %f13109, %f13108, 0f2F7FFFFE; xor.b32 %r2607, %r601, %r605; xor.b32 %r2608, %r2607, -905047700; mul.lo.s32 %r2609, %r2608, 9; shr.u32 %r2610, %r2609, 4; xor.b32 %r2611, %r2610, %r2609; mul.lo.s32 %r2612, %r2611, 668265261; cvt.rn.f32.u32 %f13110, %r2612; mul.ftz.f32 %f13111, %f13110, 0f2F7FFFFE; fma.rn.ftz.f32 %f13112, %f13109, %f1922, %f29434; fma.rn.ftz.f32 %f13113, %f13111, %f1922, %f29435; cvt.rmi.ftz.f32.f32 %f13114, %f13112; sub.ftz.f32 %f29434, %f13112, %f13114; cvt.rmi.ftz.f32.f32 %f13115, %f13113; sub.ftz.f32 %f29435, %f13113, %f13115; bra.uni $L__BB107_1320; $L__BB107_1316: shr.u32 %r2614, %r605, 16; xor.b32 %r2615, %r2614, %r605; mul.lo.s32 %r2616, %r2615, 569420461; shr.u32 %r2617, %r2616, 15; xor.b32 %r2618, %r2617, %r2616; mul.lo.s32 %r2619, %r2618, -749064809; shr.u32 %r2620, %r2619, 15; xor.b32 %r2621, %r601, %r2619; xor.b32 %r2622, %r2621, %r2620; xor.b32 %r609, %r2622, -419546133; xor.b32 %r2623, %r2622, 508809841; add.s32 %r2624, %r2623, %r607; shr.u32 %r2625, %r2623, 16; or.b32 %r2626, %r2625, 1; mul.lo.s32 %r2627, %r2626, %r2624; mul.lo.s32 %r2628, %r2627, 89287766; xor.b32 %r2629, %r2628, %r2627; mul.lo.s32 %r2630, %r2629, 1403136100; xor.b32 %r610, %r2630, %r2629; setp.eq.s32 %p1252, %r610, 0; mov.u32 %r5709, 0; @%p1252 bra $L__BB107_1319; mov.u32 %r5706, 0; mov.u64 %rd2166, sobol_burley_table; mov.u32 %r5707, %r610; mov.u32 %r5709, %r5706; $L__BB107_1318: clz.b32 %r2633, %r5707; add.s32 %r2634, %r2633, %r5706; mul.wide.u32 %rd2165, %r2634, 4; add.s64 %rd2167, %rd2166, %rd2165; ld.const.u32 %r2635, [%rd2167+128]; xor.b32 %r5709, %r2635, %r5709; add.s32 %r5706, %r2634, 1; shl.b32 %r2636, %r5707, %r2633; shl.b32 %r5707, %r2636, 1; setp.ne.s32 %p1253, %r5707, 0; @%p1253 bra $L__BB107_1318; $L__BB107_1319: xor.b32 %r2637, %r609, -525684874; brev.b32 %r2638, %r610; shr.u32 %r2639, %r2637, 16; or.b32 %r2640, %r2639, 1; mul.lo.s32 %r2641, %r2638, 1025551850; xor.b32 %r2642, %r2641, %r2638; add.s32 %r2643, %r2642, %r2637; mul.lo.s32 %r2644, %r2643, %r2640; mul.lo.s32 %r2645, %r2644, 89287766; xor.b32 %r2646, %r2645, %r2644; xor.b32 %r2647, %r609, -1802089138; mul.lo.s32 %r2648, %r5709, 1025551850; xor.b32 %r2649, %r2648, %r5709; add.s32 %r2650, %r2649, %r2647; shr.u32 %r2651, %r2647, 16; or.b32 %r2652, %r2651, 1; mul.lo.s32 %r2653, %r2650, %r2652; mul.lo.s32 %r2654, %r2653, 89287766; xor.b32 %r2655, %r2654, %r2653; mul.lo.s32 %r2656, %r2655, 1403136100; xor.b32 %r2657, %r2656, %r2655; brev.b32 %r2658, %r2657; cvt.rn.f32.u32 %f13116, %r2658; mul.lo.s32 %r2659, %r2646, 1403136100; xor.b32 %r2660, %r2659, %r2646; brev.b32 %r2661, %r2660; cvt.rn.f32.u32 %f13117, %r2661; mul.ftz.f32 %f29435, %f13116, 0f2F7FFFFE; mul.ftz.f32 %f29434, %f13117, 0f2F7FFFFE; $L__BB107_1320: ld.local.f32 %f13118, [%rd5+-4]; add.u64 %rd279, %SPL, 3424; ld.local.v4.f32 {%f13119, %f13120, %f13121, %f13122}, [%rd279]; cvt.u32.u16 %r2662, %rs29; add.u64 %rd2169, %SP, 2672; { // callseq 331, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f29434; .param .b32 param1; st.param.f32 [param1+0], %f29435; .param .b32 param2; st.param.f32 [param2+0], %f13118; .param .align 4 .b8 param3[12]; st.param.f32 [param3+0], %f13119; st.param.f32 [param3+4], %f13120; st.param.f32 [param3+8], %f13121; .param .b32 param4; st.param.b32 [param4+0], %r2662; .param .b32 param5; st.param.b32 [param5+0], %r604; .param .b64 param6; st.param.b64 [param6+0], %rd2169; .param .b32 retval0; call.uni (retval0), _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample, ( param0, param1, param2, param3, param4, param5, param6 ); ld.param.b32 %r2663, [retval0+0]; } // callseq 331 cvt.u16.u32 %rs333, %r2663; setp.eq.s16 %p1254, %rs333, 0; @%p1254 bra $L__BB107_3809; cvta.to.local.u64 %rd2171, %rd2169; add.s64 %rd280, %rd2171, 24; ld.local.f32 %f13127, [%rd5+-76]; ld.local.f32 %f30148, [%rd2171+24]; ld.local.v2.f32 {%f13128, %f13129}, [%rd5+-72]; mov.u32 %r5802, 0; ld.local.f32 %f30147, [%rd2171+28]; mul.ftz.f32 %f13132, %f30147, %f13128; fma.rn.ftz.f32 %f13133, %f30148, %f13127, %f13132; ld.local.f32 %f30146, [%rd2171+32]; fma.rn.ftz.f32 %f1938, %f30146, %f13129, %f13133; ld.local.u32 %r618, [%rd2171+68]; cvt.s64.s32 %rd281, %r618; setp.eq.s32 %p1255, %r618, -1; add.u64 %rd2172, %SP, 2752; add.u64 %rd2173, %SPL, 2752; add.s64 %rd282, %rd2173, 12; add.u64 %rd2174, %SP, 3376; add.u64 %rd283, %SPL, 3376; @%p1255 bra $L__BB107_3392; ld.const.u64 %rd2175, [kernel_params+2368]; cvta.to.global.u64 %rd2176, %rd2175; mul.lo.s64 %rd2177, %rd281, 192; add.s64 %rd2178, %rd2176, %rd2177; add.s64 %rd284, %rd2178, 40; ld.global.u32 %r2666, [%rd2178+40]; setp.eq.s32 %p1256, %r2666, 0; @%p1256 bra $L__BB107_3392; setp.geu.ftz.f32 %p1257, %f1938, 0f00000000; @%p1257 bra $L__BB107_1325; ld.local.u8 %rs334, [%rd5+-31]; and.b16 %rs335, %rs334, 2; setp.ne.s16 %p1258, %rs335, 0; @%p1258 bra $L__BB107_3809; $L__BB107_1325: setp.lt.ftz.f32 %p1259, %f1938, 0f00000000; @%p1259 bra $L__BB107_3392; ld.local.u8 %rs336, [%rd5+-31]; and.b16 %rs337, %rs336, 4; setp.eq.s16 %p1260, %rs337, 0; @%p1260 bra $L__BB107_3392; ld.local.u32 %r2669, [%rd5+-12]; add.u64 %rd2179, %SP, 1428; add.u64 %rd2180, %SPL, 1428; add.s64 %rd285, %rd2180, 40; st.local.u32 [%rd2180+40], %r2669; ld.local.u32 %r2670, [%rd5+-28]; st.local.u32 [%rd2180+36], %r2670; ld.local.u32 %r619, [%rd280+32]; st.local.u32 [%rd2180+48], %r619; ld.local.u32 %r620, [%rd280+36]; st.local.u32 [%rd2180+44], %r620; ld.local.v4.f32 {%f13137, %f13138, %f13139, %f13140}, [%rd279]; st.local.f32 [%rd2180], %f13137; st.local.f32 [%rd2180+4], %f13138; st.local.f32 [%rd2180+8], %f13139; mov.u32 %r5710, 0; st.local.u32 [%rd2180+24], %r5710; ld.local.f32 %f30145, [%rd280+12]; setp.eq.ftz.f32 %p1261, %f30145, 0f7F7FFFFF; @%p1261 bra $L__BB107_1329; bra.uni $L__BB107_1328; $L__BB107_1329: st.local.f32 [%rd285+-28], %f30148; st.local.f32 [%rd285+-24], %f30147; st.local.f32 [%rd285+-20], %f30146; mov.u32 %r2672, 2139095039; st.local.u32 [%rd285+-12], %r2672; mov.f32 %f1956, %f30146; mov.f32 %f1957, %f30147; mov.f32 %f1958, %f30148; bra.uni $L__BB107_1330; $L__BB107_90: ld.local.v2.f32 {%f8665, %f8666}, [%rd5+32]; ld.local.v4.f32 {%f8669, %f8670, %f8671, %f8672}, [%rd5+56]; ld.local.f32 %f8676, [%rd5+28]; mul.ftz.f32 %f8677, %f8665, %f8670; fma.rn.ftz.f32 %f8678, %f8676, %f8669, %f8677; fma.rn.ftz.f32 %f28945, %f8666, %f8671, %f8678; ld.local.v4.f32 {%f8679, %f8680, %f8681, %f8682}, [%rd5+72]; mul.ftz.f32 %f8686, %f8665, %f8680; fma.rn.ftz.f32 %f8687, %f8676, %f8679, %f8686; fma.rn.ftz.f32 %f28946, %f8666, %f8681, %f8687; ld.local.v4.f32 {%f8688, %f8689, %f8690, %f8691}, [%rd5+88]; mul.ftz.f32 %f8695, %f8665, %f8689; fma.rn.ftz.f32 %f8696, %f8676, %f8688, %f8695; fma.rn.ftz.f32 %f28947, %f8666, %f8690, %f8696; $L__BB107_92: st.local.f32 [%rd5+28], %f28945; st.local.f32 [%rd5+32], %f28946; st.local.f32 [%rd5+36], %f28947; @%p80 bra $L__BB107_94; bra.uni $L__BB107_93; $L__BB107_94: ld.const.u64 %rd1270, [kernel_params+2160]; cvta.to.global.u64 %rd1271, %rd1270; ld.local.u32 %r1678, [%rd5+-12]; mul.wide.s32 %rd1272, %r1678, 208; add.s64 %rd1273, %rd1271, %rd1272; ld.global.v4.f32 {%f8766, %f8767, %f8768, %f8769}, [%rd1273]; ld.global.v4.f32 {%f8773, %f8774, %f8775, %f8776}, [%rd1273+16]; ld.global.v4.f32 {%f8780, %f8781, %f8782, %f8783}, [%rd1273+32]; ld.local.v4.f32 {%f8787, %f8788, %f8789, %f8790}, [%rd5+40]; mul.ftz.f32 %f8794, %f8767, %f8788; fma.rn.ftz.f32 %f8795, %f8787, %f8766, %f8794; mul.ftz.f32 %f8796, %f8774, %f8788; fma.rn.ftz.f32 %f8797, %f8787, %f8773, %f8796; mul.ftz.f32 %f8798, %f8781, %f8788; fma.rn.ftz.f32 %f8799, %f8787, %f8780, %f8798; fma.rn.ftz.f32 %f8800, %f8782, %f8789, %f8799; fma.rn.ftz.f32 %f8801, %f8775, %f8789, %f8797; fma.rn.ftz.f32 %f8802, %f8768, %f8789, %f8795; st.local.v2.f32 [%rd5+40], {%f8802, %f8801}; st.local.f32 [%rd5+48], %f8800; bra.uni $L__BB107_95; $L__BB107_93: ld.local.v4.f32 {%f8729, %f8730, %f8731, %f8732}, [%rd5+40]; ld.local.v4.f32 {%f8736, %f8737, %f8738, %f8739}, [%rd5+56]; mul.ftz.f32 %f8743, %f8730, %f8737; fma.rn.ftz.f32 %f8744, %f8729, %f8736, %f8743; ld.local.v4.f32 {%f8745, %f8746, %f8747, %f8748}, [%rd5+72]; mul.ftz.f32 %f8752, %f8730, %f8746; fma.rn.ftz.f32 %f8753, %f8729, %f8745, %f8752; ld.local.v4.f32 {%f8754, %f8755, %f8756, %f8757}, [%rd5+88]; mul.ftz.f32 %f8761, %f8730, %f8755; fma.rn.ftz.f32 %f8762, %f8729, %f8754, %f8761; fma.rn.ftz.f32 %f8763, %f8731, %f8756, %f8762; fma.rn.ftz.f32 %f8764, %f8731, %f8747, %f8753; fma.rn.ftz.f32 %f8765, %f8731, %f8738, %f8744; st.local.v2.f32 [%rd5+40], {%f8765, %f8764}; st.local.f32 [%rd5+48], %f8763; bra.uni $L__BB107_95; $L__BB107_394: setp.eq.s32 %p283, %r173, 0; @%p283 bra $L__BB107_431; bra.uni $L__BB107_395; $L__BB107_431: setp.lt.u32 %p338, %r176, 3; mov.u32 %r5587, 0; @%p338 bra $L__BB107_454; sub.s32 %r5581, %r172, %r177; mov.u32 %r5587, 0; setp.ne.s32 %p361, %r174, 0; setp.eq.s32 %p364, %r175, 0; $L__BB107_433: cvt.s64.s32 %rd60, %r5587; mul.wide.s32 %rd1497, %r5587, 80; add.s64 %rd1498, %rd2, %rd1497; add.s64 %rd61, %rd1498, 316; ld.local.u32 %r210, [%rd1498+316]; add.s32 %r1857, %r210, -2; setp.lt.u32 %p339, %r1857, 7; @%p339 bra $L__BB107_437; add.s32 %r1858, %r210, -21; setp.lt.u32 %p341, %r1858, 10; and.pred %p342, %p341, %p361; @%p342 bra $L__BB107_437; setp.ne.s32 %p344, %r210, 31; or.pred %p345, %p344, %p364; @%p345 bra $L__BB107_438; mov.u32 %r1859, 35; st.local.u32 [%rd61], %r1859; mov.u32 %r1860, 0; st.local.u32 [%rd61+4], %r1860; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_438; $L__BB107_437: mov.u32 %r1861, 0; st.local.u32 [%rd61], %r1861; st.local.u32 [%rd61+4], %r1861; $L__BB107_438: ld.local.u32 %r213, [%rd61+80]; add.s32 %r1862, %r213, -2; setp.lt.u32 %p346, %r1862, 7; @%p346 bra $L__BB107_442; add.s32 %r1863, %r213, -21; setp.lt.u32 %p348, %r1863, 10; and.pred %p349, %p348, %p361; @%p349 bra $L__BB107_442; setp.ne.s32 %p351, %r213, 31; or.pred %p352, %p351, %p364; @%p352 bra $L__BB107_443; mov.u32 %r1864, 35; st.local.u32 [%rd61+80], %r1864; mov.u32 %r1865, 0; st.local.u32 [%rd61+84], %r1865; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_443; $L__BB107_442: mov.u32 %r1866, 0; st.local.u32 [%rd61+80], %r1866; st.local.u32 [%rd61+84], %r1866; $L__BB107_443: ld.local.u32 %r216, [%rd61+160]; add.s32 %r1867, %r216, -2; setp.lt.u32 %p353, %r1867, 7; @%p353 bra $L__BB107_447; add.s32 %r1868, %r216, -21; setp.lt.u32 %p355, %r1868, 10; and.pred %p356, %p355, %p361; @%p356 bra $L__BB107_447; setp.ne.s32 %p358, %r216, 31; or.pred %p359, %p358, %p364; @%p359 bra $L__BB107_448; mov.u32 %r1869, 35; st.local.u32 [%rd61+160], %r1869; mov.u32 %r1870, 0; st.local.u32 [%rd61+164], %r1870; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_448; $L__BB107_447: mov.u32 %r1871, 0; st.local.u32 [%rd61+160], %r1871; st.local.u32 [%rd61+164], %r1871; $L__BB107_448: ld.local.u32 %r219, [%rd61+240]; add.s32 %r1872, %r219, -2; setp.lt.u32 %p360, %r1872, 7; @%p360 bra $L__BB107_452; add.s32 %r1873, %r219, -21; setp.lt.u32 %p362, %r1873, 10; and.pred %p363, %p362, %p361; @%p363 bra $L__BB107_452; setp.ne.s32 %p365, %r219, 31; or.pred %p366, %p365, %p364; @%p366 bra $L__BB107_453; mov.u32 %r1874, 35; st.local.u32 [%rd61+240], %r1874; mov.u32 %r1875, 0; st.local.u32 [%rd61+244], %r1875; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_453; $L__BB107_452: mov.u32 %r1876, 0; st.local.u32 [%rd61+240], %r1876; st.local.u32 [%rd61+244], %r1876; $L__BB107_453: cvt.u32.u64 %r1877, %rd60; add.s32 %r5587, %r1877, 4; add.s32 %r5581, %r5581, -4; setp.ne.s32 %p367, %r5581, 0; @%p367 bra $L__BB107_433; $L__BB107_454: setp.eq.s32 %p368, %r177, 0; @%p368 bra $L__BB107_520; setp.ne.s32 %p370, %r174, 0; setp.eq.s32 %p373, %r175, 0; $L__BB107_456: .pragma "nounroll"; cvt.s64.s32 %rd62, %r5587; mul.wide.s32 %rd1499, %r5587, 80; add.s64 %rd1500, %rd2, %rd1499; add.s64 %rd63, %rd1500, 316; ld.local.u32 %r229, [%rd1500+316]; add.s32 %r1878, %r229, -2; setp.lt.u32 %p369, %r1878, 7; @%p369 bra $L__BB107_460; add.s32 %r1879, %r229, -21; setp.lt.u32 %p371, %r1879, 10; and.pred %p372, %p371, %p370; @%p372 bra $L__BB107_460; setp.ne.s32 %p374, %r229, 31; or.pred %p375, %p374, %p373; @%p375 bra $L__BB107_461; mov.u32 %r1880, 35; st.local.u32 [%rd63], %r1880; mov.u32 %r1881, 0; st.local.u32 [%rd63+4], %r1881; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_461; $L__BB107_460: mov.u32 %r1882, 0; st.local.u32 [%rd63], %r1882; st.local.u32 [%rd63+4], %r1882; $L__BB107_461: cvt.u32.u64 %r1883, %rd62; add.s32 %r5587, %r1883, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p376, %r177, 0; @%p376 bra $L__BB107_520; bra.uni $L__BB107_456; $L__BB107_463: setp.lt.u32 %p378, %r176, 3; mov.u32 %r5600, 0; @%p378 bra $L__BB107_486; sub.s32 %r5594, %r172, %r177; mov.u32 %r5600, 0; setp.ne.s32 %p409, %r174, 0; setp.eq.s32 %p412, %r175, 0; $L__BB107_465: cvt.s64.s32 %rd64, %r5600; mul.wide.s32 %rd1501, %r5600, 80; add.s64 %rd1502, %rd2, %rd1501; add.s64 %rd65, %rd1502, 316; ld.local.u32 %r238, [%rd1502+316]; add.s32 %r1886, %r238, -9; setp.lt.u32 %p379, %r1886, 12; setp.eq.s32 %p380, %r238, 29; or.pred %p381, %p379, %p380; @%p381 bra $L__BB107_469; add.s32 %r1887, %r238, -21; setp.lt.u32 %p383, %r1887, 10; and.pred %p384, %p383, %p409; @%p384 bra $L__BB107_469; setp.ne.s32 %p386, %r238, 31; or.pred %p387, %p386, %p412; @%p387 bra $L__BB107_470; mov.u32 %r1888, 35; st.local.u32 [%rd65], %r1888; mov.u32 %r1889, 0; st.local.u32 [%rd65+4], %r1889; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_470; $L__BB107_469: mov.u32 %r1890, 0; st.local.u32 [%rd65], %r1890; st.local.u32 [%rd65+4], %r1890; $L__BB107_470: ld.local.u32 %r241, [%rd65+80]; add.s32 %r1891, %r241, -9; setp.lt.u32 %p388, %r1891, 12; setp.eq.s32 %p389, %r241, 29; or.pred %p390, %p388, %p389; @%p390 bra $L__BB107_474; add.s32 %r1892, %r241, -21; setp.lt.u32 %p392, %r1892, 10; and.pred %p393, %p392, %p409; @%p393 bra $L__BB107_474; setp.ne.s32 %p395, %r241, 31; or.pred %p396, %p395, %p412; @%p396 bra $L__BB107_475; mov.u32 %r1893, 35; st.local.u32 [%rd65+80], %r1893; mov.u32 %r1894, 0; st.local.u32 [%rd65+84], %r1894; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_475; $L__BB107_474: mov.u32 %r1895, 0; st.local.u32 [%rd65+80], %r1895; st.local.u32 [%rd65+84], %r1895; $L__BB107_475: ld.local.u32 %r244, [%rd65+160]; add.s32 %r1896, %r244, -9; setp.lt.u32 %p397, %r1896, 12; setp.eq.s32 %p398, %r244, 29; or.pred %p399, %p397, %p398; @%p399 bra $L__BB107_479; add.s32 %r1897, %r244, -21; setp.lt.u32 %p401, %r1897, 10; and.pred %p402, %p401, %p409; @%p402 bra $L__BB107_479; setp.ne.s32 %p404, %r244, 31; or.pred %p405, %p404, %p412; @%p405 bra $L__BB107_480; mov.u32 %r1898, 35; st.local.u32 [%rd65+160], %r1898; mov.u32 %r1899, 0; st.local.u32 [%rd65+164], %r1899; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_480; $L__BB107_479: mov.u32 %r1900, 0; st.local.u32 [%rd65+160], %r1900; st.local.u32 [%rd65+164], %r1900; $L__BB107_480: ld.local.u32 %r247, [%rd65+240]; add.s32 %r1901, %r247, -9; setp.lt.u32 %p406, %r1901, 12; setp.eq.s32 %p407, %r247, 29; or.pred %p408, %p406, %p407; @%p408 bra $L__BB107_484; add.s32 %r1902, %r247, -21; setp.lt.u32 %p410, %r1902, 10; and.pred %p411, %p410, %p409; @%p411 bra $L__BB107_484; setp.ne.s32 %p413, %r247, 31; or.pred %p414, %p413, %p412; @%p414 bra $L__BB107_485; mov.u32 %r1903, 35; st.local.u32 [%rd65+240], %r1903; mov.u32 %r1904, 0; st.local.u32 [%rd65+244], %r1904; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_485; $L__BB107_484: mov.u32 %r1905, 0; st.local.u32 [%rd65+240], %r1905; st.local.u32 [%rd65+244], %r1905; $L__BB107_485: cvt.u32.u64 %r1906, %rd64; add.s32 %r5600, %r1906, 4; add.s32 %r5594, %r5594, -4; setp.ne.s32 %p415, %r5594, 0; @%p415 bra $L__BB107_465; $L__BB107_486: setp.eq.s32 %p416, %r177, 0; @%p416 bra $L__BB107_520; setp.ne.s32 %p420, %r174, 0; setp.eq.s32 %p423, %r175, 0; $L__BB107_488: .pragma "nounroll"; cvt.s64.s32 %rd66, %r5600; mul.wide.s32 %rd1503, %r5600, 80; add.s64 %rd1504, %rd2, %rd1503; add.s64 %rd67, %rd1504, 316; ld.local.u32 %r257, [%rd1504+316]; add.s32 %r1907, %r257, -9; setp.lt.u32 %p417, %r1907, 12; setp.eq.s32 %p418, %r257, 29; or.pred %p419, %p417, %p418; @%p419 bra $L__BB107_492; add.s32 %r1908, %r257, -21; setp.lt.u32 %p421, %r1908, 10; and.pred %p422, %p421, %p420; @%p422 bra $L__BB107_492; setp.ne.s32 %p424, %r257, 31; or.pred %p425, %p424, %p423; @%p425 bra $L__BB107_493; mov.u32 %r1909, 35; st.local.u32 [%rd67], %r1909; mov.u32 %r1910, 0; st.local.u32 [%rd67+4], %r1910; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_493; $L__BB107_492: mov.u32 %r1911, 0; st.local.u32 [%rd67], %r1911; st.local.u32 [%rd67+4], %r1911; $L__BB107_493: cvt.u32.u64 %r1912, %rd66; add.s32 %r5600, %r1912, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p426, %r177, 0; @%p426 bra $L__BB107_520; bra.uni $L__BB107_488; $L__BB107_395: setp.lt.u32 %p284, %r176, 3; mov.u32 %r5574, 0; @%p284 bra $L__BB107_422; sub.s32 %r5568, %r172, %r177; mov.u32 %r5574, 0; setp.ne.s32 %p319, %r174, 0; setp.eq.s32 %p322, %r175, 0; $L__BB107_397: cvt.s64.s32 %rd56, %r5574; mul.wide.s32 %rd1493, %r5574, 80; add.s64 %rd1494, %rd2, %rd1493; add.s64 %rd57, %rd1494, 316; ld.local.u32 %r182, [%rd1494+316]; add.s32 %r1823, %r182, -2; setp.lt.u32 %p285, %r1823, 7; @%p285 bra $L__BB107_402; add.s32 %r1824, %r182, -9; setp.lt.u32 %p286, %r1824, 12; setp.eq.s32 %p287, %r182, 29; or.pred %p288, %p286, %p287; @%p288 bra $L__BB107_402; add.s32 %r1825, %r182, -21; setp.lt.u32 %p290, %r1825, 10; and.pred %p291, %p290, %p319; @%p291 bra $L__BB107_402; setp.ne.s32 %p293, %r182, 31; or.pred %p294, %p293, %p322; @%p294 bra $L__BB107_403; mov.u32 %r1826, 35; st.local.u32 [%rd57], %r1826; mov.u32 %r1827, 0; st.local.u32 [%rd57+4], %r1827; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_403; $L__BB107_402: mov.u32 %r1828, 0; st.local.u32 [%rd57], %r1828; st.local.u32 [%rd57+4], %r1828; $L__BB107_403: ld.local.u32 %r185, [%rd57+80]; add.s32 %r1829, %r185, -2; setp.lt.u32 %p295, %r1829, 7; @%p295 bra $L__BB107_408; add.s32 %r1830, %r185, -9; setp.lt.u32 %p296, %r1830, 12; setp.eq.s32 %p297, %r185, 29; or.pred %p298, %p296, %p297; @%p298 bra $L__BB107_408; add.s32 %r1831, %r185, -21; setp.lt.u32 %p300, %r1831, 10; and.pred %p301, %p300, %p319; @%p301 bra $L__BB107_408; setp.ne.s32 %p303, %r185, 31; or.pred %p304, %p303, %p322; @%p304 bra $L__BB107_409; mov.u32 %r1832, 35; st.local.u32 [%rd57+80], %r1832; mov.u32 %r1833, 0; st.local.u32 [%rd57+84], %r1833; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_409; $L__BB107_408: mov.u32 %r1834, 0; st.local.u32 [%rd57+80], %r1834; st.local.u32 [%rd57+84], %r1834; $L__BB107_409: ld.local.u32 %r188, [%rd57+160]; add.s32 %r1835, %r188, -2; setp.lt.u32 %p305, %r1835, 7; @%p305 bra $L__BB107_414; add.s32 %r1836, %r188, -9; setp.lt.u32 %p306, %r1836, 12; setp.eq.s32 %p307, %r188, 29; or.pred %p308, %p306, %p307; @%p308 bra $L__BB107_414; add.s32 %r1837, %r188, -21; setp.lt.u32 %p310, %r1837, 10; and.pred %p311, %p310, %p319; @%p311 bra $L__BB107_414; setp.ne.s32 %p313, %r188, 31; or.pred %p314, %p313, %p322; @%p314 bra $L__BB107_415; mov.u32 %r1838, 35; st.local.u32 [%rd57+160], %r1838; mov.u32 %r1839, 0; st.local.u32 [%rd57+164], %r1839; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_415; $L__BB107_414: mov.u32 %r1840, 0; st.local.u32 [%rd57+160], %r1840; st.local.u32 [%rd57+164], %r1840; $L__BB107_415: ld.local.u32 %r191, [%rd57+240]; add.s32 %r1841, %r191, -2; setp.lt.u32 %p315, %r1841, 7; @%p315 bra $L__BB107_420; add.s32 %r1842, %r191, -9; setp.lt.u32 %p316, %r1842, 12; setp.eq.s32 %p317, %r191, 29; or.pred %p318, %p316, %p317; @%p318 bra $L__BB107_420; add.s32 %r1843, %r191, -21; setp.lt.u32 %p320, %r1843, 10; and.pred %p321, %p320, %p319; @%p321 bra $L__BB107_420; setp.ne.s32 %p323, %r191, 31; or.pred %p324, %p323, %p322; @%p324 bra $L__BB107_421; mov.u32 %r1844, 35; st.local.u32 [%rd57+240], %r1844; mov.u32 %r1845, 0; st.local.u32 [%rd57+244], %r1845; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_421; $L__BB107_420: mov.u32 %r1846, 0; st.local.u32 [%rd57+240], %r1846; st.local.u32 [%rd57+244], %r1846; $L__BB107_421: cvt.u32.u64 %r1847, %rd56; add.s32 %r5574, %r1847, 4; add.s32 %r5568, %r5568, -4; setp.ne.s32 %p325, %r5568, 0; @%p325 bra $L__BB107_397; $L__BB107_422: setp.eq.s32 %p326, %r177, 0; @%p326 bra $L__BB107_520; setp.ne.s32 %p331, %r174, 0; setp.eq.s32 %p334, %r175, 0; $L__BB107_424: .pragma "nounroll"; cvt.s64.s32 %rd58, %r5574; mul.wide.s32 %rd1495, %r5574, 80; add.s64 %rd1496, %rd2, %rd1495; add.s64 %rd59, %rd1496, 316; ld.local.u32 %r201, [%rd1496+316]; add.s32 %r1848, %r201, -2; setp.lt.u32 %p327, %r1848, 7; @%p327 bra $L__BB107_429; add.s32 %r1849, %r201, -9; setp.lt.u32 %p328, %r1849, 12; setp.eq.s32 %p329, %r201, 29; or.pred %p330, %p328, %p329; @%p330 bra $L__BB107_429; add.s32 %r1850, %r201, -21; setp.lt.u32 %p332, %r1850, 10; and.pred %p333, %p332, %p331; @%p333 bra $L__BB107_429; setp.ne.s32 %p335, %r201, 31; or.pred %p336, %p335, %p334; @%p336 bra $L__BB107_430; mov.u32 %r1851, 35; st.local.u32 [%rd59], %r1851; mov.u32 %r1852, 0; st.local.u32 [%rd59+4], %r1852; or.b32 %r5618, %r5618, 32; st.local.u32 [%rd5+-36], %r5618; bra.uni $L__BB107_430; $L__BB107_429: mov.u32 %r1853, 0; st.local.u32 [%rd59], %r1853; st.local.u32 [%rd59+4], %r1853; $L__BB107_430: cvt.u32.u64 %r1854, %rd58; add.s32 %r5574, %r1854, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p337, %r177, 0; @%p337 bra $L__BB107_520; bra.uni $L__BB107_424; $L__BB107_520: ld.const.v2.u64 {%rd1509, %rd1510}, [kernel_params+2464]; cvta.to.global.u64 %rd1513, %rd1509; shl.b64 %rd1514, %rd3, 1; add.s64 %rd72, %rd1513, %rd1514; ld.global.u16 %r1937, [%rd72]; cvta.to.global.u64 %rd1515, %rd1510; add.s64 %rd1516, %rd1515, %rd1514; ld.global.u16 %r1938, [%rd1516]; neg.s32 %r1939, %r1938; setp.ne.s32 %p461, %r1937, %r1939; ld.local.u32 %r291, [%rd5+172]; setp.lt.s32 %p462, %r291, 2; or.pred %p463, %p461, %p462; @%p463 bra $L__BB107_553; add.s32 %r292, %r291, -1; and.b32 %r5628, %r291, 3; setp.lt.u32 %p464, %r292, 3; mov.f32 %f29002, 0f00000000; mov.u32 %r5621, 0; @%p464 bra $L__BB107_532; sub.s32 %r5620, %r291, %r5628; mov.f32 %f29002, 0f00000000; mov.u32 %r5621, 0; $L__BB107_523: cvt.s64.s32 %rd73, %r5621; mul.wide.s32 %rd1517, %r5621, 80; add.s64 %rd1518, %rd2, %rd1517; add.s64 %rd74, %rd1518, 316; ld.local.u32 %r1942, [%rd1518+316]; setp.gt.s32 %p465, %r1942, 34; @%p465 bra $L__BB107_525; ld.local.f32 %f9021, [%rd74+4]; add.ftz.f32 %f29002, %f29002, %f9021; $L__BB107_525: ld.local.u32 %r1943, [%rd74+80]; setp.gt.s32 %p466, %r1943, 34; @%p466 bra $L__BB107_527; ld.local.f32 %f9022, [%rd74+84]; add.ftz.f32 %f29002, %f29002, %f9022; $L__BB107_527: ld.local.u32 %r1944, [%rd74+160]; setp.gt.s32 %p467, %r1944, 34; @%p467 bra $L__BB107_529; ld.local.f32 %f9023, [%rd74+164]; add.ftz.f32 %f29002, %f29002, %f9023; $L__BB107_529: ld.local.u32 %r1945, [%rd74+240]; setp.gt.s32 %p468, %r1945, 34; @%p468 bra $L__BB107_531; ld.local.f32 %f9024, [%rd74+244]; add.ftz.f32 %f29002, %f29002, %f9024; $L__BB107_531: cvt.u32.u64 %r1946, %rd73; add.s32 %r5621, %r1946, 4; add.s32 %r5620, %r5620, -4; setp.ne.s32 %p469, %r5620, 0; @%p469 bra $L__BB107_523; $L__BB107_532: setp.eq.s32 %p470, %r5628, 0; @%p470 bra $L__BB107_537; mov.u32 %r5623, %r5628; $L__BB107_534: .pragma "nounroll"; cvt.s64.s32 %rd75, %r5621; mul.wide.s32 %rd1519, %r5621, 80; add.s64 %rd1520, %rd2, %rd1519; add.s64 %rd76, %rd1520, 316; ld.local.u32 %r1947, [%rd1520+316]; setp.gt.s32 %p471, %r1947, 34; @%p471 bra $L__BB107_536; ld.local.f32 %f9025, [%rd76+4]; add.ftz.f32 %f29002, %f29002, %f9025; $L__BB107_536: cvt.u32.u64 %r1948, %rd75; add.s32 %r5621, %r1948, 1; add.s32 %r5623, %r5623, -1; setp.ne.s32 %p472, %r5623, 0; @%p472 bra $L__BB107_534; $L__BB107_537: setp.lt.s32 %p473, %r291, 1; @%p473 bra $L__BB107_553; mul.ftz.f32 %f665, %f29002, 0f3E000000; mov.u32 %r5626, 0; @%p464 bra $L__BB107_549; sub.s32 %r5625, %r291, %r5628; mov.u32 %r5626, 0; $L__BB107_540: cvt.s64.s32 %rd77, %r5626; mul.wide.s32 %rd1521, %r5626, 80; add.s64 %rd1522, %rd2, %rd1521; add.s64 %rd78, %rd1522, 316; ld.local.u32 %r1951, [%rd1522+316]; setp.gt.s32 %p475, %r1951, 34; @%p475 bra $L__BB107_542; ld.local.f32 %f9026, [%rd78+4]; max.ftz.f32 %f9027, %f9026, %f665; st.local.f32 [%rd78+4], %f9027; $L__BB107_542: ld.local.u32 %r1952, [%rd78+80]; setp.gt.s32 %p476, %r1952, 34; @%p476 bra $L__BB107_544; ld.local.f32 %f9028, [%rd78+84]; max.ftz.f32 %f9029, %f9028, %f665; st.local.f32 [%rd78+84], %f9029; $L__BB107_544: ld.local.u32 %r1953, [%rd78+160]; setp.gt.s32 %p477, %r1953, 34; @%p477 bra $L__BB107_546; ld.local.f32 %f9030, [%rd78+164]; max.ftz.f32 %f9031, %f9030, %f665; st.local.f32 [%rd78+164], %f9031; $L__BB107_546: ld.local.u32 %r1954, [%rd78+240]; setp.gt.s32 %p478, %r1954, 34; @%p478 bra $L__BB107_548; ld.local.f32 %f9032, [%rd78+244]; max.ftz.f32 %f9033, %f9032, %f665; st.local.f32 [%rd78+244], %f9033; $L__BB107_548: cvt.u32.u64 %r1955, %rd77; add.s32 %r5626, %r1955, 4; add.s32 %r5625, %r5625, -4; setp.ne.s32 %p479, %r5625, 0; @%p479 bra $L__BB107_540; $L__BB107_549: @%p470 bra $L__BB107_553; $L__BB107_550: .pragma "nounroll"; cvt.s64.s32 %rd79, %r5626; mul.wide.s32 %rd1523, %r5626, 80; add.s64 %rd1524, %rd2, %rd1523; add.s64 %rd80, %rd1524, 316; ld.local.u32 %r1956, [%rd1524+316]; setp.gt.s32 %p481, %r1956, 34; @%p481 bra $L__BB107_552; ld.local.f32 %f9034, [%rd80+4]; max.ftz.f32 %f9035, %f9034, %f665; st.local.f32 [%rd80+4], %f9035; $L__BB107_552: cvt.u32.u64 %r1957, %rd79; add.s32 %r5626, %r1957, 1; add.s32 %r5628, %r5628, -1; setp.ne.s32 %p482, %r5628, 0; @%p482 bra $L__BB107_550; $L__BB107_553: ld.const.f32 %f666, [kernel_params+1572]; setp.eq.ftz.f32 %p483, %f666, 0f7F7FFFFF; @%p483 bra $L__BB107_647; ld.const.u64 %rd1525, [kernel_params+2552]; cvta.to.global.u64 %rd1526, %rd1525; add.s64 %rd1527, %rd1526, %rd3; ld.global.u8 %rs312, [%rd1527]; and.b16 %rs313, %rs312, 1; setp.eq.b16 %p484, %rs313, 1; mov.pred %p485, 0; xor.pred %p486, %p484, %p485; @%p486 bra $L__BB107_647; ld.const.u64 %rd1528, [kernel_params+2568]; cvta.to.global.u64 %rd1529, %rd1528; add.s64 %rd1531, %rd1529, %rd1282; ld.global.f32 %f9036, [%rd1531]; mul.ftz.f32 %f667, %f666, %f9036; setp.geu.ftz.f32 %p487, %f667, 0f3F800000; @%p487 bra $L__BB107_647; mov.f32 %f9037, 0f3F800000; sub.ftz.f32 %f9038, %f9037, %f667; sqrt.approx.ftz.f32 %f9039, %f9038; mul.ftz.f32 %f668, %f9039, 0f3F000000; setp.lt.s32 %p488, %r291, 1; @%p488 bra $L__BB107_647; max.s32 %r314, %r291, 1; add.s32 %r1959, %r314, -1; and.b32 %r5633, %r314, 3; setp.lt.u32 %p489, %r1959, 3; mov.u32 %r5631, 0; @%p489 bra $L__BB107_628; sub.s32 %r5630, %r314, %r5633; mov.u32 %r5631, 0; $L__BB107_559: cvt.s64.s32 %rd81, %r5631; mul.wide.s32 %rd1532, %r5631, 80; add.s64 %rd1533, %rd2, %rd1532; add.s64 %rd82, %rd1533, 316; ld.local.u32 %r319, [%rd1533+316]; setp.gt.s32 %p490, %r319, 31; @%p490 bra $L__BB107_576; setp.gt.s32 %p491, %r319, 15; @%p491 bra $L__BB107_565; add.s32 %r1961, %r319, -10; setp.lt.u32 %p499, %r1961, 3; @%p499 bra $L__BB107_569; add.s32 %r1962, %r319, -14; setp.lt.u32 %p500, %r1962, 2; @%p500 bra $L__BB107_575; setp.eq.s32 %p501, %r319, 13; @%p501 bra $L__BB107_564; bra.uni $L__BB107_576; $L__BB107_565: setp.gt.s32 %p492, %r319, 23; @%p492 bra $L__BB107_570; setp.eq.s32 %p496, %r319, 16; @%p496 bra $L__BB107_574; setp.eq.s32 %p497, %r319, 22; @%p497 bra $L__BB107_564; setp.eq.s32 %p498, %r319, 23; @%p498 bra $L__BB107_569; bra.uni $L__BB107_576; $L__BB107_569: ld.local.v2.f32 {%f9064, %f9065}, [%rd82+20]; max.ftz.f32 %f9068, %f668, %f9065; max.ftz.f32 %f9069, %f668, %f9064; st.local.v2.f32 [%rd82+20], {%f9069, %f9068}; bra.uni $L__BB107_576; $L__BB107_570: setp.eq.s32 %p493, %r319, 24; @%p493 bra $L__BB107_575; setp.eq.s32 %p494, %r319, 27; @%p494 bra $L__BB107_575; setp.ne.s32 %p495, %r319, 29; @%p495 bra $L__BB107_576; ld.local.f32 %f9040, [%rd82+32]; max.ftz.f32 %f9041, %f668, %f9040; st.local.f32 [%rd82+32], %f9041; ld.local.f32 %f9042, [%rd82+36]; max.ftz.f32 %f9043, %f668, %f9042; st.local.f32 [%rd82+36], %f9043; ld.local.f32 %f9044, [%rd82+48]; max.ftz.f32 %f9045, %f668, %f9044; st.local.f32 [%rd82+48], %f9045; bra.uni $L__BB107_576; $L__BB107_575: ld.local.v2.f32 {%f9052, %f9053}, [%rd82+20]; max.ftz.f32 %f9056, %f668, %f9053; max.ftz.f32 %f9057, %f668, %f9052; st.local.v2.f32 [%rd82+20], {%f9057, %f9056}; bra.uni $L__BB107_576; $L__BB107_564: ld.local.v2.f32 {%f9058, %f9059}, [%rd82+20]; max.ftz.f32 %f9062, %f668, %f9059; max.ftz.f32 %f9063, %f668, %f9058; st.local.v2.f32 [%rd82+20], {%f9063, %f9062}; bra.uni $L__BB107_576; $L__BB107_574: ld.local.v2.f32 {%f9046, %f9047}, [%rd82+20]; max.ftz.f32 %f9050, %f668, %f9047; max.ftz.f32 %f9051, %f668, %f9046; st.local.v2.f32 [%rd82+20], {%f9051, %f9050}; $L__BB107_576: ld.local.u32 %r320, [%rd82+80]; setp.gt.s32 %p502, %r320, 31; @%p502 bra $L__BB107_593; setp.gt.s32 %p503, %r320, 15; @%p503 bra $L__BB107_582; add.s32 %r1963, %r320, -10; setp.lt.u32 %p511, %r1963, 3; @%p511 bra $L__BB107_586; add.s32 %r1964, %r320, -14; setp.lt.u32 %p512, %r1964, 2; @%p512 bra $L__BB107_592; setp.eq.s32 %p513, %r320, 13; @%p513 bra $L__BB107_581; bra.uni $L__BB107_593; $L__BB107_582: setp.gt.s32 %p504, %r320, 23; @%p504 bra $L__BB107_587; setp.eq.s32 %p508, %r320, 16; @%p508 bra $L__BB107_591; setp.eq.s32 %p509, %r320, 22; @%p509 bra $L__BB107_581; setp.eq.s32 %p510, %r320, 23; @%p510 bra $L__BB107_586; bra.uni $L__BB107_593; $L__BB107_586: ld.local.v2.f32 {%f9094, %f9095}, [%rd82+100]; max.ftz.f32 %f9098, %f668, %f9095; max.ftz.f32 %f9099, %f668, %f9094; st.local.v2.f32 [%rd82+100], {%f9099, %f9098}; bra.uni $L__BB107_593; $L__BB107_587: setp.eq.s32 %p505, %r320, 24; @%p505 bra $L__BB107_592; setp.eq.s32 %p506, %r320, 27; @%p506 bra $L__BB107_592; setp.ne.s32 %p507, %r320, 29; @%p507 bra $L__BB107_593; ld.local.f32 %f9070, [%rd82+112]; max.ftz.f32 %f9071, %f668, %f9070; st.local.f32 [%rd82+112], %f9071; ld.local.f32 %f9072, [%rd82+116]; max.ftz.f32 %f9073, %f668, %f9072; st.local.f32 [%rd82+116], %f9073; ld.local.f32 %f9074, [%rd82+128]; max.ftz.f32 %f9075, %f668, %f9074; st.local.f32 [%rd82+128], %f9075; bra.uni $L__BB107_593; $L__BB107_592: ld.local.v2.f32 {%f9082, %f9083}, [%rd82+100]; max.ftz.f32 %f9086, %f668, %f9083; max.ftz.f32 %f9087, %f668, %f9082; st.local.v2.f32 [%rd82+100], {%f9087, %f9086}; bra.uni $L__BB107_593; $L__BB107_581: ld.local.v2.f32 {%f9088, %f9089}, [%rd82+100]; max.ftz.f32 %f9092, %f668, %f9089; max.ftz.f32 %f9093, %f668, %f9088; st.local.v2.f32 [%rd82+100], {%f9093, %f9092}; bra.uni $L__BB107_593; $L__BB107_591: ld.local.v2.f32 {%f9076, %f9077}, [%rd82+100]; max.ftz.f32 %f9080, %f668, %f9077; max.ftz.f32 %f9081, %f668, %f9076; st.local.v2.f32 [%rd82+100], {%f9081, %f9080}; $L__BB107_593: ld.local.u32 %r321, [%rd82+160]; setp.gt.s32 %p514, %r321, 31; @%p514 bra $L__BB107_610; setp.gt.s32 %p515, %r321, 15; @%p515 bra $L__BB107_599; add.s32 %r1965, %r321, -10; setp.lt.u32 %p523, %r1965, 3; @%p523 bra $L__BB107_603; add.s32 %r1966, %r321, -14; setp.lt.u32 %p524, %r1966, 2; @%p524 bra $L__BB107_609; setp.eq.s32 %p525, %r321, 13; @%p525 bra $L__BB107_598; bra.uni $L__BB107_610; $L__BB107_599: setp.gt.s32 %p516, %r321, 23; @%p516 bra $L__BB107_604; setp.eq.s32 %p520, %r321, 16; @%p520 bra $L__BB107_608; setp.eq.s32 %p521, %r321, 22; @%p521 bra $L__BB107_598; setp.eq.s32 %p522, %r321, 23; @%p522 bra $L__BB107_603; bra.uni $L__BB107_610; $L__BB107_603: ld.local.v2.f32 {%f9124, %f9125}, [%rd82+180]; max.ftz.f32 %f9128, %f668, %f9125; max.ftz.f32 %f9129, %f668, %f9124; st.local.v2.f32 [%rd82+180], {%f9129, %f9128}; bra.uni $L__BB107_610; $L__BB107_604: setp.eq.s32 %p517, %r321, 24; @%p517 bra $L__BB107_609; setp.eq.s32 %p518, %r321, 27; @%p518 bra $L__BB107_609; setp.ne.s32 %p519, %r321, 29; @%p519 bra $L__BB107_610; ld.local.f32 %f9100, [%rd82+192]; max.ftz.f32 %f9101, %f668, %f9100; st.local.f32 [%rd82+192], %f9101; ld.local.f32 %f9102, [%rd82+196]; max.ftz.f32 %f9103, %f668, %f9102; st.local.f32 [%rd82+196], %f9103; ld.local.f32 %f9104, [%rd82+208]; max.ftz.f32 %f9105, %f668, %f9104; st.local.f32 [%rd82+208], %f9105; bra.uni $L__BB107_610; $L__BB107_609: ld.local.v2.f32 {%f9112, %f9113}, [%rd82+180]; max.ftz.f32 %f9116, %f668, %f9113; max.ftz.f32 %f9117, %f668, %f9112; st.local.v2.f32 [%rd82+180], {%f9117, %f9116}; bra.uni $L__BB107_610; $L__BB107_598: ld.local.v2.f32 {%f9118, %f9119}, [%rd82+180]; max.ftz.f32 %f9122, %f668, %f9119; max.ftz.f32 %f9123, %f668, %f9118; st.local.v2.f32 [%rd82+180], {%f9123, %f9122}; bra.uni $L__BB107_610; $L__BB107_608: ld.local.v2.f32 {%f9106, %f9107}, [%rd82+180]; max.ftz.f32 %f9110, %f668, %f9107; max.ftz.f32 %f9111, %f668, %f9106; st.local.v2.f32 [%rd82+180], {%f9111, %f9110}; $L__BB107_610: ld.local.u32 %r322, [%rd82+240]; setp.gt.s32 %p526, %r322, 31; @%p526 bra $L__BB107_627; setp.gt.s32 %p527, %r322, 15; @%p527 bra $L__BB107_616; add.s32 %r1967, %r322, -10; setp.lt.u32 %p535, %r1967, 3; @%p535 bra $L__BB107_620; add.s32 %r1968, %r322, -14; setp.lt.u32 %p536, %r1968, 2; @%p536 bra $L__BB107_626; setp.eq.s32 %p537, %r322, 13; @%p537 bra $L__BB107_615; bra.uni $L__BB107_627; $L__BB107_616: setp.gt.s32 %p528, %r322, 23; @%p528 bra $L__BB107_621; setp.eq.s32 %p532, %r322, 16; @%p532 bra $L__BB107_625; setp.eq.s32 %p533, %r322, 22; @%p533 bra $L__BB107_615; setp.eq.s32 %p534, %r322, 23; @%p534 bra $L__BB107_620; bra.uni $L__BB107_627; $L__BB107_620: ld.local.v2.f32 {%f9154, %f9155}, [%rd82+260]; max.ftz.f32 %f9158, %f668, %f9155; max.ftz.f32 %f9159, %f668, %f9154; st.local.v2.f32 [%rd82+260], {%f9159, %f9158}; bra.uni $L__BB107_627; $L__BB107_621: setp.eq.s32 %p529, %r322, 24; @%p529 bra $L__BB107_626; setp.eq.s32 %p530, %r322, 27; @%p530 bra $L__BB107_626; setp.ne.s32 %p531, %r322, 29; @%p531 bra $L__BB107_627; ld.local.f32 %f9130, [%rd82+272]; max.ftz.f32 %f9131, %f668, %f9130; st.local.f32 [%rd82+272], %f9131; ld.local.f32 %f9132, [%rd82+276]; max.ftz.f32 %f9133, %f668, %f9132; st.local.f32 [%rd82+276], %f9133; ld.local.f32 %f9134, [%rd82+288]; max.ftz.f32 %f9135, %f668, %f9134; st.local.f32 [%rd82+288], %f9135; bra.uni $L__BB107_627; $L__BB107_626: ld.local.v2.f32 {%f9142, %f9143}, [%rd82+260]; max.ftz.f32 %f9146, %f668, %f9143; max.ftz.f32 %f9147, %f668, %f9142; st.local.v2.f32 [%rd82+260], {%f9147, %f9146}; bra.uni $L__BB107_627; $L__BB107_615: ld.local.v2.f32 {%f9148, %f9149}, [%rd82+260]; max.ftz.f32 %f9152, %f668, %f9149; max.ftz.f32 %f9153, %f668, %f9148; st.local.v2.f32 [%rd82+260], {%f9153, %f9152}; bra.uni $L__BB107_627; $L__BB107_625: ld.local.v2.f32 {%f9136, %f9137}, [%rd82+260]; max.ftz.f32 %f9140, %f668, %f9137; max.ftz.f32 %f9141, %f668, %f9136; st.local.v2.f32 [%rd82+260], {%f9141, %f9140}; $L__BB107_627: cvt.u32.u64 %r1969, %rd81; add.s32 %r5631, %r1969, 4; add.s32 %r5630, %r5630, -4; setp.ne.s32 %p538, %r5630, 0; @%p538 bra $L__BB107_559; $L__BB107_628: setp.eq.s32 %p539, %r5633, 0; @%p539 bra $L__BB107_647; $L__BB107_629: .pragma "nounroll"; cvt.s64.s32 %rd83, %r5631; mul.wide.s32 %rd1534, %r5631, 80; add.s64 %rd1535, %rd2, %rd1534; add.s64 %rd84, %rd1535, 316; ld.local.u32 %r328, [%rd1535+316]; setp.gt.s32 %p540, %r328, 31; @%p540 bra $L__BB107_646; setp.gt.s32 %p541, %r328, 15; @%p541 bra $L__BB107_635; add.s32 %r1970, %r328, -10; setp.lt.u32 %p549, %r1970, 3; @%p549 bra $L__BB107_639; add.s32 %r1971, %r328, -14; setp.lt.u32 %p550, %r1971, 2; @%p550 bra $L__BB107_645; setp.eq.s32 %p551, %r328, 13; @%p551 bra $L__BB107_634; bra.uni $L__BB107_646; $L__BB107_635: setp.gt.s32 %p542, %r328, 23; @%p542 bra $L__BB107_640; setp.eq.s32 %p546, %r328, 16; @%p546 bra $L__BB107_644; setp.eq.s32 %p547, %r328, 22; @%p547 bra $L__BB107_634; setp.eq.s32 %p548, %r328, 23; @%p548 bra $L__BB107_639; bra.uni $L__BB107_646; $L__BB107_639: ld.local.v2.f32 {%f9184, %f9185}, [%rd84+20]; max.ftz.f32 %f9188, %f668, %f9185; max.ftz.f32 %f9189, %f668, %f9184; st.local.v2.f32 [%rd84+20], {%f9189, %f9188}; bra.uni $L__BB107_646; $L__BB107_640: setp.eq.s32 %p543, %r328, 24; @%p543 bra $L__BB107_645; setp.eq.s32 %p544, %r328, 27; @%p544 bra $L__BB107_645; setp.ne.s32 %p545, %r328, 29; @%p545 bra $L__BB107_646; ld.local.f32 %f9160, [%rd84+32]; max.ftz.f32 %f9161, %f668, %f9160; st.local.f32 [%rd84+32], %f9161; ld.local.f32 %f9162, [%rd84+36]; max.ftz.f32 %f9163, %f668, %f9162; st.local.f32 [%rd84+36], %f9163; ld.local.f32 %f9164, [%rd84+48]; max.ftz.f32 %f9165, %f668, %f9164; st.local.f32 [%rd84+48], %f9165; bra.uni $L__BB107_646; $L__BB107_645: ld.local.v2.f32 {%f9172, %f9173}, [%rd84+20]; max.ftz.f32 %f9176, %f668, %f9173; max.ftz.f32 %f9177, %f668, %f9172; st.local.v2.f32 [%rd84+20], {%f9177, %f9176}; bra.uni $L__BB107_646; $L__BB107_634: ld.local.v2.f32 {%f9178, %f9179}, [%rd84+20]; max.ftz.f32 %f9182, %f668, %f9179; max.ftz.f32 %f9183, %f668, %f9178; st.local.v2.f32 [%rd84+20], {%f9183, %f9182}; bra.uni $L__BB107_646; $L__BB107_644: ld.local.v2.f32 {%f9166, %f9167}, [%rd84+20]; max.ftz.f32 %f9170, %f668, %f9167; max.ftz.f32 %f9171, %f668, %f9166; st.local.v2.f32 [%rd84+20], {%f9171, %f9170}; $L__BB107_646: cvt.u32.u64 %r1972, %rd83; add.s32 %r5631, %r1972, 1; add.s32 %r5633, %r5633, -1; setp.ne.s32 %p552, %r5633, 0; @%p552 bra $L__BB107_629; $L__BB107_647: ld.global.u32 %r331, [%rd27]; and.b32 %r1973, %r5618, 32; setp.eq.s32 %p553, %r1973, 0; and.b32 %r332, %r331, 16384; @%p553 bra $L__BB107_649; bra.uni $L__BB107_648; $L__BB107_649: ld.local.u32 %r1974, [%rd5+-32]; and.b32 %r1975, %r1974, 1; setp.eq.b32 %p555, %r1975, 1; not.pred %p556, %p555; setp.eq.s32 %p557, %r332, 0; or.pred %p558, %p557, %p556; @%p558 bra $L__BB107_690; bra.uni $L__BB107_650; $L__BB107_648: setp.eq.s32 %p554, %r332, 0; @%p554 bra $L__BB107_690; $L__BB107_650: ld.local.u32 %r1976, [%rd5+-32]; and.b32 %r1977, %r1976, 1; setp.eq.b32 %p559, %r1977, 1; mov.pred %p560, 0; xor.pred %p561, %p559, %p560; not.pred %p562, %p561; @%p562 bra $L__BB107_669; bra.uni $L__BB107_651; $L__BB107_669: ld.local.u32 %r346, [%rd5+172]; setp.lt.s32 %p574, %r346, 1; mov.f32 %f29014, 0f00000000; mov.f32 %f29015, %f29014; mov.f32 %f29016, %f29014; @%p574 bra $L__BB107_685; add.s32 %r1993, %r346, -1; and.b32 %r5643, %r346, 3; setp.lt.u32 %p575, %r1993, 3; mov.f32 %f29016, 0f00000000; mov.u32 %r5641, 0; mov.f32 %f29015, %f29016; mov.f32 %f29014, %f29016; @%p575 bra $L__BB107_681; sub.s32 %r5640, %r346, %r5643; mov.f32 %f29016, 0f00000000; mov.u32 %r5641, 0; mov.f32 %f29015, %f29016; mov.f32 %f29014, %f29016; $L__BB107_672: cvt.s64.s32 %rd89, %r5641; mul.wide.s32 %rd1538, %r5641, 80; add.s64 %rd1539, %rd2, %rd1538; add.s64 %rd90, %rd1539, 316; ld.local.u32 %r1995, [%rd1539+316]; setp.ne.s32 %p576, %r1995, 35; @%p576 bra $L__BB107_674; ld.local.v4.f32 {%f9206, %f9207, %f9208, %f9209}, [%rd90+-12]; add.ftz.f32 %f29014, %f29014, %f9206; add.ftz.f32 %f29015, %f29015, %f9207; add.ftz.f32 %f29016, %f29016, %f9208; $L__BB107_674: ld.local.u32 %r1996, [%rd90+80]; setp.ne.s32 %p577, %r1996, 35; @%p577 bra $L__BB107_676; ld.local.v4.f32 {%f9213, %f9214, %f9215, %f9216}, [%rd90+68]; add.ftz.f32 %f29014, %f29014, %f9213; add.ftz.f32 %f29015, %f29015, %f9214; add.ftz.f32 %f29016, %f29016, %f9215; $L__BB107_676: ld.local.u32 %r1997, [%rd90+160]; setp.ne.s32 %p578, %r1997, 35; @%p578 bra $L__BB107_678; ld.local.v4.f32 {%f9220, %f9221, %f9222, %f9223}, [%rd90+148]; add.ftz.f32 %f29014, %f29014, %f9220; add.ftz.f32 %f29015, %f29015, %f9221; add.ftz.f32 %f29016, %f29016, %f9222; $L__BB107_678: ld.local.u32 %r1998, [%rd90+240]; setp.ne.s32 %p579, %r1998, 35; @%p579 bra $L__BB107_680; ld.local.v4.f32 {%f9227, %f9228, %f9229, %f9230}, [%rd90+228]; add.ftz.f32 %f29014, %f29014, %f9227; add.ftz.f32 %f29015, %f29015, %f9228; add.ftz.f32 %f29016, %f29016, %f9229; $L__BB107_680: cvt.u32.u64 %r1999, %rd89; add.s32 %r5641, %r1999, 4; add.s32 %r5640, %r5640, -4; setp.ne.s32 %p580, %r5640, 0; @%p580 bra $L__BB107_672; $L__BB107_681: setp.eq.s32 %p581, %r5643, 0; @%p581 bra $L__BB107_685; $L__BB107_682: .pragma "nounroll"; cvt.s64.s32 %rd91, %r5641; mul.wide.s32 %rd1540, %r5641, 80; add.s64 %rd1541, %rd2, %rd1540; add.s64 %rd92, %rd1541, 316; ld.local.u32 %r2000, [%rd1541+316]; setp.ne.s32 %p582, %r2000, 35; @%p582 bra $L__BB107_684; ld.local.f32 %f9234, [%rd92+-12]; add.ftz.f32 %f29014, %f29014, %f9234; ld.local.f32 %f9235, [%rd92+-8]; add.ftz.f32 %f29015, %f29015, %f9235; ld.local.f32 %f9236, [%rd92+-4]; add.ftz.f32 %f29016, %f29016, %f9236; $L__BB107_684: cvt.u32.u64 %r2001, %rd91; add.s32 %r5641, %r2001, 1; add.s32 %r5643, %r5643, -1; setp.ne.s32 %p583, %r5643, 0; @%p583 bra $L__BB107_682; bra.uni $L__BB107_685; $L__BB107_651: and.b32 %r1978, %r5618, 524800; setp.ne.s32 %p563, %r1978, 512; mov.f32 %f29014, 0f3F800000; mov.f32 %f29015, %f29014; mov.f32 %f29016, %f29014; @%p563 bra $L__BB107_685; ld.local.f32 %f669, [%rd5+204]; ld.local.v2.f32 {%f9193, %f9194}, [%rd5+208]; ld.local.u32 %r333, [%rd5+172]; setp.lt.s32 %p564, %r333, 1; @%p564 bra $L__BB107_668; add.s32 %r1980, %r333, -1; and.b32 %r5638, %r333, 3; setp.lt.u32 %p565, %r1980, 3; mov.u32 %r5636, 0; @%p565 bra $L__BB107_664; sub.s32 %r5635, %r333, %r5638; add.s64 %rd4743, %rd2, 316; mov.u32 %r5636, 0; $L__BB107_655: ld.local.u32 %r1982, [%rd4743]; setp.eq.s32 %p566, %r1982, 31; @%p566 bra $L__BB107_657; mov.u32 %r1983, 40; st.local.u32 [%rd4743], %r1983; $L__BB107_657: ld.local.u32 %r1984, [%rd4743+80]; setp.eq.s32 %p567, %r1984, 31; @%p567 bra $L__BB107_659; mov.u32 %r1985, 40; st.local.u32 [%rd4743+80], %r1985; $L__BB107_659: ld.local.u32 %r1986, [%rd4743+160]; setp.eq.s32 %p568, %r1986, 31; @%p568 bra $L__BB107_661; mov.u32 %r1987, 40; st.local.u32 [%rd4743+160], %r1987; $L__BB107_661: ld.local.u32 %r1988, [%rd4743+240]; setp.eq.s32 %p569, %r1988, 31; @%p569 bra $L__BB107_663; mov.u32 %r1989, 40; st.local.u32 [%rd4743+240], %r1989; $L__BB107_663: add.s32 %r5636, %r5636, 4; add.s32 %r5635, %r5635, -4; setp.ne.s32 %p570, %r5635, 0; add.s64 %rd4743, %rd4743, 320; @%p570 bra $L__BB107_655; $L__BB107_664: setp.eq.s32 %p571, %r5638, 0; @%p571 bra $L__BB107_668; $L__BB107_665: .pragma "nounroll"; mul.wide.s32 %rd1536, %r5636, 80; add.s64 %rd1537, %rd2, %rd1536; add.s64 %rd88, %rd1537, 316; ld.local.u32 %r1990, [%rd1537+316]; setp.eq.s32 %p572, %r1990, 31; @%p572 bra $L__BB107_667; mov.u32 %r1991, 40; st.local.u32 [%rd88], %r1991; $L__BB107_667: add.s32 %r5636, %r5636, 1; add.s32 %r5638, %r5638, -1; setp.ne.s32 %p573, %r5638, 0; @%p573 bra $L__BB107_665; $L__BB107_668: and.b32 %r5618, %r5618, -763; st.local.u32 [%rd5+-36], %r5618; mov.f32 %f9195, 0f3F800000; sub.ftz.f32 %f29016, %f9195, %f9194; sub.ftz.f32 %f29015, %f9195, %f9193; sub.ftz.f32 %f29014, %f9195, %f669; $L__BB107_685: ld.const.u64 %rd1542, [kernel_params+2584]; cvta.to.global.u64 %rd1543, %rd1542; mul.lo.s64 %rd1544, %rd3, 12; add.s64 %rd1545, %rd1543, %rd1544; ld.global.f32 %f9237, [%rd1545]; ld.global.f32 %f9238, [%rd1545+4]; mul.ftz.f32 %f9239, %f29015, %f9238; ld.global.f32 %f9240, [%rd1545+8]; fma.rn.ftz.f32 %f9241, %f29014, %f9237, %f9239; fma.rn.ftz.f32 %f9242, %f29016, %f9240, %f9241; mul.ftz.f32 %f720, %f9242, 0f3EAAAAAB; ld.const.u64 %rd1546, [kernel_params+2448]; cvta.to.global.u64 %rd1547, %rd1546; add.s64 %rd1549, %rd1547, %rd1282; ld.global.u32 %rd1550, [%rd1549]; ld.const.s32 %rd1551, [kernel_params+1264]; mul.lo.s64 %rd93, %rd1551, %rd1550; ld.const.u8 %rs314, [kernel_params+1260]; and.b16 %rs315, %rs314, 2; setp.eq.s16 %p584, %rs315, 0; @%p584 bra $L__BB107_687; ld.const.s32 %rd1552, [kernel_params+1268]; cvta.to.global.u64 %rd1553, %rd1059; add.s64 %rd1554, %rd1552, %rd93; shl.b64 %rd1555, %rd1554, 2; add.s64 %rd1556, %rd1553, %rd1555; add.s64 %rd1557, %rd1556, 12; atom.global.add.f32 %f9243, [%rd1557], %f720; $L__BB107_687: ld.const.u32 %r2002, [kernel_params+1620]; setp.eq.s32 %p585, %r2002, 0; shr.u32 %r2003, %r331, 29; and.b32 %r2004, %r2003, 1; setp.eq.b32 %p586, %r2004, 1; or.pred %p587, %p586, %p585; @%p587 bra $L__BB107_689; ld.const.s32 %rd1558, [kernel_params+1384]; cvta.to.global.u64 %rd1559, %rd1059; add.s64 %rd1560, %rd1558, %rd93; shl.b64 %rd1561, %rd1560, 2; add.s64 %rd1562, %rd1559, %rd1561; add.s64 %rd1563, %rd1562, 12; atom.global.add.f32 %f9244, [%rd1563], %f720; $L__BB107_689: setp.eq.ftz.f32 %p588, %f29014, 0f3F800000; setp.eq.ftz.f32 %p589, %f29015, 0f3F800000; and.pred %p590, %p588, %p589; setp.eq.ftz.f32 %p591, %f29016, 0f3F800000; and.pred %p592, %p590, %p591; mov.u16 %rs748, 0; @%p592 bra $L__BB107_4545; $L__BB107_690: and.b32 %r2005, %r5618, 2; setp.eq.s32 %p593, %r2005, 0; @%p593 bra $L__BB107_794; ld.global.u32 %r360, [%rd27]; ld.local.v2.f32 {%f9246, %f9247}, [%rd5+-64]; ld.local.v4.f32 {%f9248, %f9249, %f9250, %f9251}, [%rd5+-56]; mul.ftz.f32 %f9252, %f9247, %f9250; fma.rn.ftz.f32 %f9253, %f9246, %f9249, %f9252; fma.rn.ftz.f32 %f727, %f9248, %f9251, %f9253; ld.local.v2.f32 {%f9254, %f9255}, [%rd5+192]; ld.local.f32 %f730, [%rd5+200]; shr.u32 %r2006, %r360, 11; and.b32 %r2007, %r2006, 1; setp.eq.b32 %p594, %r2007, 1; and.b32 %r2008, %r5618, 65536; setp.eq.s32 %p595, %r2008, 0; or.pred %p596, %p594, %p595; mov.f32 %f29137, 0f3F800000; @%p596 bra $L__BB107_750; ld.local.u32 %r2009, [%rd5+-24]; and.b32 %r2010, %r2009, 1; setp.eq.b32 %p597, %r2010, 1; mov.pred %p598, 0; xor.pred %p599, %p597, %p598; not.pred %p600, %p599; @%p600 bra $L__BB107_750; ld.const.u64 %rd4717, [kernel_params+2184]; cvta.to.global.u64 %rd4716, %rd4717; ld.const.u64 %rd1564, [kernel_params+2560]; cvta.to.global.u64 %rd1565, %rd1564; add.s64 %rd1567, %rd1565, %rd1282; ld.global.f32 %f731, [%rd1567]; ld.local.f32 %f732, [%rd5]; ld.local.s32 %rd94, [%rd5+-28]; ld.local.u32 %r2011, [%rd5+-12]; cvt.s64.s32 %rd95, %r2011; mul.wide.s32 %rd1568, %r2011, 4; add.s64 %rd1569, %rd4716, %rd1568; ld.global.u32 %r361, [%rd1569]; shr.u32 %r2012, %r361, 6; and.b32 %r2013, %r2012, 1; setp.eq.b32 %p601, %r2013, 1; ld.local.f32 %f733, [%rd5+-4]; setp.ge.ftz.f32 %p602, %f733, 0f00000000; and.pred %p603, %p602, %p601; @%p603 bra $L__BB107_695; bra.uni $L__BB107_694; $L__BB107_695: ld.const.u64 %rd1581, [kernel_params+2160]; cvta.to.global.u64 %rd1582, %rd1581; mul.lo.s64 %rd1583, %rd95, 208; add.s64 %rd1584, %rd1582, %rd1583; ld.global.u32 %r362, [%rd1584+156]; ld.global.u32 %r363, [%rd1584+152]; shl.b32 %r2017, %r363, 1; cvt.rn.f32.s32 %f9257, %r2017; mul.ftz.f32 %f743, %f733, %f9257; cvt.rzi.ftz.s32.f32 %r2018, %f743; add.s32 %r2019, %r2017, -1; min.s32 %r364, %r2018, %r2019; ld.const.u64 %rd1585, [kernel_params+2312]; cvta.to.global.u64 %rd97, %rd1585; ld.global.u32 %r5647, [%rd1584+164]; mul.wide.u32 %rd1586, %r5647, 16; add.s64 %rd1587, %rd97, %rd1586; ld.global.u16 %rs612, [%rd1587+12]; ld.global.u32 %r5646, [%rd1587+8]; ld.global.u64 %rd4744, [%rd1587]; setp.eq.s64 %p604, %rd4744, 11; @%p604 bra $L__BB107_700; $L__BB107_696: setp.eq.s64 %p605, %rd4744, 0; @%p605 bra $L__BB107_698; add.s32 %r5647, %r5647, 2; bra.uni $L__BB107_699; $L__BB107_698: setp.eq.s16 %p606, %rs612, 0; mov.u32 %r5650, -1; mov.u32 %r5647, %r5646; @%p606 bra $L__BB107_701; $L__BB107_699: mul.wide.u32 %rd1588, %r5647, 16; add.s64 %rd1589, %rd97, %rd1588; ld.global.u16 %rs612, [%rd1589+12]; ld.global.u32 %r5646, [%rd1589+8]; ld.global.u64 %rd4744, [%rd1589]; setp.ne.s64 %p607, %rd4744, 11; @%p607 bra $L__BB107_696; $L__BB107_700: setp.eq.s16 %p608, %rs612, 0; selp.b32 %r5650, -1, %r5646, %p608; $L__BB107_701: ld.const.u64 %rd4763, [kernel_params+2232]; cvta.to.global.u64 %rd1590, %rd4763; shl.b64 %rd1591, %rd94, 4; add.s64 %rd1592, %rd1590, %rd1591; ld.global.v4.u32 {%r2021, %r2022, %r2023, %r2024}, [%rd1592]; setp.eq.s32 %p609, %r364, %r363; @%p609 bra $L__BB107_703; bra.uni $L__BB107_702; $L__BB107_703: ld.const.u64 %rd1598, [kernel_params+2256]; cvta.to.global.u64 %rd1599, %rd1598; mul.wide.u32 %rd1600, %r2024, 12; add.s64 %rd4745, %rd1599, %rd1600; add.s32 %r2031, %r2024, 1; mul.wide.u32 %rd1601, %r2031, 12; add.s64 %rd4748, %rd1599, %rd1601; add.s32 %r2032, %r2024, 2; mul.wide.u32 %rd1602, %r2032, 12; add.s64 %rd4751, %rd1599, %rd1602; add.s64 %rd4746, %rd4745, 4; add.s64 %rd4747, %rd4745, 8; add.s64 %rd4749, %rd4748, 4; add.s64 %rd4750, %rd4748, 8; add.s64 %rd4752, %rd4751, 4; add.s64 %rd4753, %rd4751, 8; bra.uni $L__BB107_704; $L__BB107_694: ld.const.u64 %rd4763, [kernel_params+2232]; cvta.to.global.u64 %rd1570, %rd4763; shl.b64 %rd1571, %rd94, 4; add.s64 %rd1572, %rd1570, %rd1571; ld.const.u64 %rd1573, [kernel_params+2256]; cvta.to.global.u64 %rd1574, %rd1573; ld.global.u32 %r2014, [%rd1572+12]; mul.wide.u32 %rd1575, %r2014, 12; add.s64 %rd1576, %rd1574, %rd1575; ld.global.f32 %f29066, [%rd1576]; ld.global.f32 %f29067, [%rd1576+4]; ld.global.f32 %f29068, [%rd1576+8]; add.s32 %r2015, %r2014, 1; mul.wide.u32 %rd1577, %r2015, 12; add.s64 %rd1578, %rd1574, %rd1577; ld.global.f32 %f29069, [%rd1578]; ld.global.f32 %f29070, [%rd1578+4]; ld.global.f32 %f29071, [%rd1578+8]; add.s32 %r2016, %r2014, 2; mul.wide.u32 %rd1579, %r2016, 12; add.s64 %rd1580, %rd1574, %rd1579; ld.global.f32 %f29072, [%rd1580]; ld.global.f32 %f29073, [%rd1580+4]; ld.global.f32 %f29074, [%rd1580+8]; mov.u16 %rs615, 0; bra.uni $L__BB107_708; $L__BB107_702: setp.gt.s32 %p610, %r364, %r363; selp.b32 %r2025, -1, 0, %p610; add.s32 %r2026, %r364, %r2025; mad.lo.s32 %r2027, %r2026, %r362, %r5650; ld.const.u64 %rd1593, [kernel_params+2336]; cvta.to.global.u64 %rd1594, %rd1593; add.s32 %r2028, %r2021, %r2027; mul.wide.u32 %rd1595, %r2028, 12; add.s64 %rd4745, %rd1594, %rd1595; add.s32 %r2029, %r2022, %r2027; mul.wide.u32 %rd1596, %r2029, 12; add.s64 %rd4748, %rd1594, %rd1596; add.s32 %r2030, %r2023, %r2027; mul.wide.u32 %rd1597, %r2030, 12; add.s64 %rd4751, %rd1594, %rd1597; add.s64 %rd4746, %rd4745, 4; add.s64 %rd4747, %rd4745, 8; add.s64 %rd4749, %rd4748, 4; add.s64 %rd4750, %rd4748, 8; add.s64 %rd4752, %rd4751, 4; add.s64 %rd4753, %rd4751, 8; $L__BB107_704: ld.global.f32 %f744, [%rd4753]; ld.global.f32 %f745, [%rd4752]; ld.global.f32 %f746, [%rd4751]; ld.global.f32 %f747, [%rd4750]; ld.global.f32 %f748, [%rd4749]; ld.global.f32 %f749, [%rd4748]; ld.global.f32 %f750, [%rd4747]; ld.global.f32 %f751, [%rd4746]; ld.global.f32 %f752, [%rd4745]; add.s32 %r379, %r364, 1; setp.eq.s32 %p611, %r379, %r363; @%p611 bra $L__BB107_706; bra.uni $L__BB107_705; $L__BB107_706: ld.const.u64 %rd1608, [kernel_params+2256]; cvta.to.global.u64 %rd1609, %rd1608; mul.wide.u32 %rd1610, %r2024, 12; add.s64 %rd4754, %rd1609, %rd1610; add.s32 %r2039, %r2024, 1; mul.wide.u32 %rd1611, %r2039, 12; add.s64 %rd4757, %rd1609, %rd1611; add.s32 %r2040, %r2024, 2; mul.wide.u32 %rd1612, %r2040, 12; add.s64 %rd4760, %rd1609, %rd1612; add.s64 %rd4755, %rd4754, 4; add.s64 %rd4756, %rd4754, 8; add.s64 %rd4758, %rd4757, 4; add.s64 %rd4759, %rd4757, 8; add.s64 %rd4761, %rd4760, 4; add.s64 %rd4762, %rd4760, 8; bra.uni $L__BB107_707; $L__BB107_705: setp.ge.s32 %p612, %r364, %r363; selp.b32 %r2033, -1, 0, %p612; add.s32 %r2034, %r379, %r2033; mad.lo.s32 %r2035, %r2034, %r362, %r5650; ld.const.u64 %rd1603, [kernel_params+2336]; cvta.to.global.u64 %rd1604, %rd1603; add.s32 %r2036, %r2021, %r2035; mul.wide.u32 %rd1605, %r2036, 12; add.s64 %rd4754, %rd1604, %rd1605; add.s32 %r2037, %r2022, %r2035; mul.wide.u32 %rd1606, %r2037, 12; add.s64 %rd4757, %rd1604, %rd1606; add.s32 %r2038, %r2023, %r2035; mul.wide.u32 %rd1607, %r2038, 12; add.s64 %rd4760, %rd1604, %rd1607; add.s64 %rd4755, %rd4754, 4; add.s64 %rd4756, %rd4754, 8; add.s64 %rd4758, %rd4757, 4; add.s64 %rd4759, %rd4757, 8; add.s64 %rd4761, %rd4760, 4; add.s64 %rd4762, %rd4760, 8; $L__BB107_707: cvt.rn.f32.s32 %f9258, %r364; sub.ftz.f32 %f9259, %f743, %f9258; mov.f32 %f9260, 0f3F800000; sub.ftz.f32 %f9261, %f9260, %f9259; ld.global.f32 %f9262, [%rd4754]; mul.ftz.f32 %f9263, %f9259, %f9262; ld.global.f32 %f9264, [%rd4755]; mul.ftz.f32 %f9265, %f9259, %f9264; ld.global.f32 %f9266, [%rd4756]; mul.ftz.f32 %f9267, %f9259, %f9266; fma.rn.ftz.f32 %f29066, %f9261, %f752, %f9263; fma.rn.ftz.f32 %f29067, %f9261, %f751, %f9265; fma.rn.ftz.f32 %f29068, %f9261, %f750, %f9267; ld.global.f32 %f9268, [%rd4757]; mul.ftz.f32 %f9269, %f9259, %f9268; ld.global.f32 %f9270, [%rd4758]; mul.ftz.f32 %f9271, %f9259, %f9270; ld.global.f32 %f9272, [%rd4759]; mul.ftz.f32 %f9273, %f9259, %f9272; fma.rn.ftz.f32 %f29069, %f9261, %f749, %f9269; fma.rn.ftz.f32 %f29070, %f9261, %f748, %f9271; fma.rn.ftz.f32 %f29071, %f9261, %f747, %f9273; ld.global.f32 %f9274, [%rd4760]; mul.ftz.f32 %f9275, %f9259, %f9274; ld.global.f32 %f9276, [%rd4761]; mul.ftz.f32 %f9277, %f9259, %f9276; ld.global.f32 %f9278, [%rd4762]; mul.ftz.f32 %f9279, %f9259, %f9278; fma.rn.ftz.f32 %f29072, %f9261, %f746, %f9275; fma.rn.ftz.f32 %f29073, %f9261, %f745, %f9277; fma.rn.ftz.f32 %f29074, %f9261, %f744, %f9279; mov.u16 %rs615, 1; $L__BB107_708: and.b32 %r380, %r361, 4; setp.ne.s32 %p613, %r380, 0; @%p613 bra $L__BB107_716; and.b32 %r2041, %r361, 2; setp.eq.s32 %p614, %r2041, 0; ld.const.u64 %rd1613, [kernel_params+2160]; cvta.to.global.u64 %rd1614, %rd1613; mul.lo.s64 %rd1615, %rd95, 208; add.s64 %rd157, %rd1614, %rd1615; @%p614 bra $L__BB107_714; bra.uni $L__BB107_710; $L__BB107_714: ld.global.v4.f32 {%f29054, %f29055, %f29056, %f29057}, [%rd157]; ld.global.v4.f32 {%f29058, %f29059, %f29060, %f29061}, [%rd157+16]; ld.global.v4.f32 {%f29062, %f29063, %f29064, %f29065}, [%rd157+32]; bra.uni $L__BB107_715; $L__BB107_710: ld.const.u64 %rd1616, [kernel_params+2176]; cvta.to.global.u64 %rd1617, %rd1616; ld.global.u32 %rd1618, [%rd157+168]; ld.global.u32 %r2042, [%rd157+152]; shl.b32 %r2043, %r2042, 1; cvt.rn.f32.s32 %f9280, %r2043; selp.f32 %f9281, %f733, 0f3F000000, %p602; mul.ftz.f32 %f9282, %f9281, %f9280; cvt.rzi.ftz.s32.f32 %r2044, %f9282; add.s32 %r2045, %r2043, -1; min.s32 %r2046, %r2044, %r2045; cvt.rn.f32.s32 %f9283, %r2046; sub.ftz.f32 %f771, %f9282, %f9283; cvt.s64.s32 %rd1619, %r2046; add.s64 %rd1620, %rd1619, %rd1618; shl.b64 %rd1621, %rd1620, 6; add.s64 %rd158, %rd1617, %rd1621; ld.global.v4.f32 {%f9284, %f9285, %f9286, %f9287}, [%rd158]; ld.global.v4.f32 {%f9288, %f9289, %f9290, %f9291}, [%rd158+64]; mul.ftz.f32 %f9292, %f9285, %f9289; fma.rn.ftz.f32 %f9293, %f9284, %f9288, %f9292; mul.ftz.f32 %f9294, %f9287, %f9291; fma.rn.ftz.f32 %f9295, %f9286, %f9290, %f9294; add.ftz.f32 %f780, %f9293, %f9295; setp.gt.ftz.f32 %p616, %f780, 0f3F7FDF3B; mov.f32 %f9296, 0f3F800000; sub.ftz.f32 %f781, %f9296, %f771; @%p616 bra $L__BB107_712; bra.uni $L__BB107_711; $L__BB107_712: mul.ftz.f32 %f9350, %f9284, %f781; fma.rn.ftz.f32 %f9351, %f9288, %f771, %f9350; mul.ftz.f32 %f9352, %f9285, %f781; fma.rn.ftz.f32 %f9353, %f9289, %f771, %f9352; mul.ftz.f32 %f9354, %f9286, %f781; fma.rn.ftz.f32 %f9355, %f9290, %f771, %f9354; mul.ftz.f32 %f9356, %f9287, %f781; fma.rn.ftz.f32 %f9357, %f9291, %f771, %f9356; mul.ftz.f32 %f9358, %f9353, %f9353; fma.rn.ftz.f32 %f9359, %f9351, %f9351, %f9358; mul.ftz.f32 %f9360, %f9357, %f9357; fma.rn.ftz.f32 %f9361, %f9355, %f9355, %f9360; add.ftz.f32 %f9362, %f9359, %f9361; rsqrt.approx.ftz.f32 %f9363, %f9362; mul.ftz.f32 %f29053, %f9357, %f9363; mul.ftz.f32 %f29052, %f9355, %f9363; mul.ftz.f32 %f29051, %f9353, %f9363; mul.ftz.f32 %f29050, %f9351, %f9363; bra.uni $L__BB107_713; $L__BB107_1328: ld.local.f32 %f13141, [%rd280+-24]; sub.ftz.f32 %f13142, %f13141, %f13137; ld.local.f32 %f13143, [%rd280+-20]; sub.ftz.f32 %f13144, %f13143, %f13138; ld.local.f32 %f13145, [%rd280+-16]; sub.ftz.f32 %f13146, %f13145, %f13139; mul.ftz.f32 %f13147, %f13144, %f13144; fma.rn.ftz.f32 %f13148, %f13142, %f13142, %f13147; fma.rn.ftz.f32 %f13149, %f13146, %f13146, %f13148; sqrt.approx.ftz.f32 %f13150, %f13149; st.local.f32 [%rd285+-12], %f13150; rsqrt.approx.ftz.f32 %f13151, %f13149; mul.ftz.f32 %f1958, %f13142, %f13151; mul.ftz.f32 %f1957, %f13144, %f13151; mul.ftz.f32 %f1956, %f13146, %f13151; st.local.f32 [%rd285+-28], %f1958; st.local.f32 [%rd285+-24], %f1957; st.local.f32 [%rd285+-20], %f1956; $L__BB107_1330: cvt.s64.s32 %rd286, %r619; ld.local.f32 %f1949, [%rd5+4]; st.local.f32 [%rd285+12], %f1949; st.local.u32 [%rd285+16], %r5710; ld.local.f32 %f1950, [%rd5+-4]; st.local.f32 [%rd285+-8], %f1950; ld.const.u32 %r621, [kernel_params+1108]; ld.const.u32 %r622, [kernel_params+1112]; ld.const.v2.u64 {%rd2181, %rd2182}, [kernel_params+2176]; cvta.to.global.u64 %rd287, %rd2182; add.s64 %rd2186, %rd2172, 68; add.s64 %rd2187, %rd2172, 72; add.s64 %rd2188, %rd2172, 88; cvta.to.local.u64 %rd288, %rd2188; add.s64 %rd289, %rd2172, 64; add.s64 %rd2189, %rd2172, 76; add.s64 %rd2190, %rd2172, 56; cvta.to.local.u64 %rd290, %rd2190; add.s64 %rd2191, %rd2172, 60; cvta.to.local.u64 %rd291, %rd2191; add.s64 %rd292, %rd2172, 80; add.s64 %rd2192, %rd2172, 52; cvta.to.local.u64 %rd293, %rd2192; add.s64 %rd2193, %rd2172, 84; add.s64 %rd2194, %rd2172, 36; cvta.to.local.u64 %rd294, %rd2194; add.s64 %rd295, %rd2172, 40; cvta.to.local.u64 %rd296, %rd295; add.s64 %rd297, %rd2172, 44; cvta.to.local.u64 %rd298, %rd297; add.s64 %rd299, %rd2172, 144; ld.const.u64 %rd2195, [kernel_params+2160]; cvta.to.global.u64 %rd300, %rd2195; cvta.to.global.u64 %rd301, %rd2181; add.s64 %rd302, %rd2172, 24; ld.const.v2.u64 {%rd2196, %rd2197}, [kernel_params+2400]; cvta.to.global.u64 %rd303, %rd2197; add.s64 %rd304, %rd2172, 48; cvta.to.local.u64 %rd305, %rd304; add.s64 %rd306, %rd2172, 32; cvta.to.local.u64 %rd307, %rd306; add.s64 %rd2200, %rd2172, 116; add.s64 %rd308, %rd2172, 120; add.s64 %rd2201, %rd2172, 124; add.s64 %rd309, %rd2172, 128; add.s64 %rd310, %rd2172, 176; add.s64 %rd311, %rd2172, 160; add.s64 %rd312, %rd2172, 192; ld.const.v2.u64 {%rd2202, %rd2203}, [kernel_params+2304]; cvta.to.global.u64 %rd313, %rd2202; or.b64 %rd2206, %rd2172, 8; cvta.to.local.u64 %rd314, %rd2206; or.b64 %rd2207, %rd2172, 12; cvta.to.local.u64 %rd315, %rd2207; add.s64 %rd316, %rd2172, 16; add.s64 %rd317, %rd2172, 20; cvta.to.local.u64 %rd318, %rd2200; cvta.to.local.u64 %rd319, %rd308; cvta.to.local.u64 %rd320, %rd2201; add.s64 %rd2208, %rd2172, 136; cvta.to.local.u64 %rd321, %rd2208; add.s64 %rd2209, %rd2172, 92; cvta.to.local.u64 %rd322, %rd2209; add.s64 %rd323, %rd2172, 96; cvta.to.local.u64 %rd324, %rd323; ld.const.v2.u64 {%rd2210, %rd2211}, [kernel_params+2256]; cvta.to.global.u64 %rd325, %rd2211; add.s64 %rd326, %rd2172, 224; add.s64 %rd327, %rd2172, 208; ld.const.v2.u64 {%rd2214, %rd2215}, [kernel_params+2224]; cvta.to.global.u64 %rd328, %rd2215; cvta.to.global.u64 %rd329, %rd2210; ld.const.u64 %rd2218, [kernel_params+2296]; cvta.to.global.u64 %rd330, %rd2218; cvta.to.global.u64 %rd331, %rd2203; ld.const.u64 %rd2219, [kernel_params+2272]; cvta.to.global.u64 %rd332, %rd2219; ld.const.u64 %rd2220, [kernel_params+2216]; cvta.to.global.u64 %rd333, %rd2220; add.s64 %rd2221, %rd2172, 100; cvta.to.local.u64 %rd334, %rd2221; add.s64 %rd335, %rd2172, 104; add.s64 %rd2222, %rd2172, 112; cvta.to.local.u64 %rd336, %rd2222; cvta.to.local.u64 %rd337, %rd316; cvta.to.local.u64 %rd338, %rd317; cvta.to.local.u64 %rd339, %rd2189; ld.const.v2.u64 {%rd2223, %rd2224}, [kernel_params+2336]; cvta.to.global.u64 %rd341, %rd2224; add.s64 %rd2226, %rd2172, 260; cvta.to.local.u64 %rd342, %rd2226; add.s64 %rd2227, %rd2172, 264; cvta.to.local.u64 %rd343, %rd2227; cvta.to.global.u64 %rd344, %rd2196; add.s64 %rd2228, %rd2172, 268; cvta.to.local.u64 %rd348, %rd2228; add.s64 %rd2229, %rd2172, 272; cvta.to.local.u64 %rd349, %rd2229; add.s64 %rd2230, %rd2172, 276; cvta.to.local.u64 %rd350, %rd2230; cvta.to.local.u64 %rd351, %rd2187; cvta.to.global.u64 %rd352, %rd2214; ld.const.u64 %rd2231, [kernel_params+2136]; cvta.to.global.u64 %rd353, %rd2231; cvta.to.local.u64 %rd354, %rd289; add.s32 %r2675, %r602, 3; shr.u32 %r2676, %r2675, 16; xor.b32 %r2677, %r2676, %r2675; mul.lo.s32 %r2678, %r2677, 569420461; shr.u32 %r2679, %r2678, 15; xor.b32 %r2680, %r2679, %r2678; mul.lo.s32 %r2681, %r2680, -749064809; shr.u32 %r2682, %r2681, 15; xor.b32 %r2683, %r601, %r2681; xor.b32 %r2684, %r2683, %r2682; xor.b32 %r2685, %r2684, 508809841; add.s32 %r2686, %r2685, %r607; shr.u32 %r2687, %r2685, 16; or.b32 %r2688, %r2687, 1; mul.lo.s32 %r2689, %r2688, %r2686; mul.lo.s32 %r2690, %r2689, 89287766; xor.b32 %r2691, %r2690, %r2689; mul.lo.s32 %r2692, %r2691, 1403136100; xor.b32 %r623, %r2692, %r2691; xor.b32 %r2693, %r2684, 106206365; brev.b32 %r2694, %r623; mul.lo.s32 %r2695, %r2694, 1025551850; xor.b32 %r2696, %r2695, %r2694; add.s32 %r2697, %r2696, %r2693; shr.u32 %r2698, %r2693, 16; or.b32 %r2699, %r2698, 1; mul.lo.s32 %r2700, %r2697, %r2699; mul.lo.s32 %r2701, %r2700, 89287766; xor.b32 %r2702, %r2701, %r2700; mul.lo.s32 %r2703, %r2702, 1403136100; xor.b32 %r2704, %r2703, %r2702; brev.b32 %r2705, %r2704; cvt.rn.f32.u32 %f13152, %r2705; xor.b32 %r624, %r2684, 1919448741; ld.const.v2.u32 {%r2706, %r2707}, [kernel_params+1600]; mov.b32 %f1951, %r2707; setp.lt.ftz.f32 %p1262, %f1951, 0f3F800000; ld.const.u32 %r2710, [kernel_params+1576]; selp.b32 %r2711, %r2710, %r601, %p1262; add.s16 %rs338, %rs28, 3; and.b16 %rs339, %rs338, 255; cvt.u32.u16 %r2712, %rs339; xor.b32 %r2713, %r2711, %r2712; mul.lo.s32 %r2714, %r2713, -512718531; shr.u32 %r2715, %r2711, 16; xor.b32 %r2716, %r2714, %r2715; bfe.u32 %r2717, %r2716, 4, 4; shr.u32 %r2718, %r2711, 8; xor.b32 %r2719, %r2716, %r2718; xor.b32 %r2720, %r2719, %r2717; mul.lo.s32 %r2721, %r2720, 153742143; shr.u32 %r2722, %r2711, 23; xor.b32 %r2723, %r2721, %r2722; bfe.u32 %r2724, %r2723, 1, 7; xor.b32 %r2725, %r2724, %r2723; shr.u32 %r2726, %r2711, 27; or.b32 %r2727, %r2726, 1; mul.lo.s32 %r2728, %r2727, %r2725; mul.lo.s32 %r2729, %r2728, -1074439621; bfe.u32 %r2730, %r2729, 2, 6; xor.b32 %r2731, %r2730, %r2729; mul.lo.s32 %r2732, %r2731, -1638916925; bfe.u32 %r2733, %r2732, 2, 6; xor.b32 %r2734, %r2733, %r2732; mul.lo.s32 %r2735, %r2734, -933190689; and.b32 %r2736, %r2735, 255; bfe.u32 %r2737, %r2735, 5, 3; xor.b32 %r2738, %r2737, %r2736; add.s32 %r2739, %r2706, -1; xor.b32 %r2740, %r2675, 61; xor.b32 %r2741, %r2711, %r2740; mul.lo.s32 %r2742, %r2741, 9; shr.u32 %r2743, %r2742, 4; xor.b32 %r2744, %r2743, %r2742; mul.lo.s32 %r2745, %r2744, 668265261; add.s32 %r2746, %r2745, %r607; shr.u32 %r2747, %r2745, 16; or.b32 %r2748, %r2747, 1; mul.lo.s32 %r2749, %r2748, %r2746; mul.lo.s32 %r2750, %r2749, 89287766; xor.b32 %r2751, %r2750, %r2749; mul.lo.s32 %r2752, %r2751, 1403136100; xor.b32 %r2753, %r2752, %r2751; brev.b32 %r2754, %r2753; neg.s32 %r2755, %r2706; and.b32 %r2756, %r603, %r2755; and.b32 %r2757, %r2754, %r2739; or.b32 %r2758, %r2757, %r2756; mad.lo.s32 %r625, %r2738, %r2706, %r2758; shl.b32 %r626, %r2706, 8; ld.const.u64 %rd2232, [kernel_params+2424]; cvta.to.global.u64 %rd355, %rd2232; cvta.to.local.u64 %rd356, %rd2193; cvta.to.local.u64 %rd357, %rd2186; shr.u32 %r2759, %r624, 16; or.b32 %r627, %r2759, 1; mul.ftz.f32 %f1952, %f13152, 0f2F7FFFFE; xor.b32 %r2760, %r2740, %r601; mul.lo.s32 %r2761, %r2760, 9; shr.u32 %r2762, %r2761, 4; xor.b32 %r2763, %r2762, %r2761; mul.lo.s32 %r2764, %r2763, 668265261; cvt.rn.f32.u32 %f13153, %r2764; mul.ftz.f32 %f13154, %f13153, 0f2F7FFFFE; mul.ftz.f32 %f1954, %f13154, %f1951; xor.b32 %r2765, %r601, %r2675; xor.b32 %r2766, %r2765, -905047700; mul.lo.s32 %r2767, %r2766, 9; shr.u32 %r2768, %r2767, 4; xor.b32 %r2769, %r2768, %r2767; mul.lo.s32 %r2770, %r2769, 668265261; cvt.rn.f32.u32 %f13155, %r2770; mul.ftz.f32 %f13156, %f13155, 0f2F7FFFFE; mul.ftz.f32 %f1955, %f13156, %f1951; cvt.s64.s32 %rd359, %r620; setp.nan.ftz.f32 %p1263, %f13137, %f13137; mov.u16 %rs340, 0; mov.pred %p1284, 0; cvta.to.local.u64 %rd2374, %rd335; rem.u32 %r2960, %r625, %r626; shl.b32 %r2961, %r2960, 1; mov.f32 %f1959, %f1956; mov.f32 %f1960, %f1957; mov.f32 %f1961, %f1958; mov.u32 %r5732, %r5710; $L__BB107_1331: mov.b32 %r630, %f13137; mov.u16 %rs620, %rs340; @%p1263 bra $L__BB107_1334; or.b32 %r2771, %r630, -2147483648; setp.ne.s32 %p1264, %r2771, -2147483648; add.ftz.f32 %f13157, %f13137, %f13137; setp.eq.ftz.f32 %p1265, %f13157, %f13137; and.pred %p1266, %p1265, %p1264; mov.u16 %rs620, %rs340; @%p1266 bra $L__BB107_1334; shl.b32 %r2772, %r630, 1; setp.lt.u32 %p1267, %r2772, -16777215; selp.u16 %rs620, 1, 0, %p1267; $L__BB107_1334: setp.eq.s16 %p1268, %rs620, 0; mov.u16 %rs622, %rs340; @%p1268 bra $L__BB107_1340; mov.b32 %r631, %f1961; setp.nan.ftz.f32 %p1269, %f1961, %f1961; mov.u16 %rs622, 0; mov.u16 %rs621, %rs622; @%p1269 bra $L__BB107_1338; or.b32 %r2773, %r631, -2147483648; setp.ne.s32 %p1270, %r2773, -2147483648; add.ftz.f32 %f13158, %f1961, %f1961; setp.eq.ftz.f32 %p1271, %f13158, %f1961; and.pred %p1272, %p1271, %p1270; mov.u16 %rs621, %rs622; @%p1272 bra $L__BB107_1338; shl.b32 %r2774, %r631, 1; setp.lt.u32 %p1273, %r2774, -16777215; selp.u16 %rs621, 1, 0, %p1273; $L__BB107_1338: setp.eq.s16 %p1274, %rs621, 0; @%p1274 bra $L__BB107_1340; mul.ftz.f32 %f13159, %f1960, %f1960; fma.rn.ftz.f32 %f13160, %f1961, %f1961, %f13159; fma.rn.ftz.f32 %f13161, %f1959, %f1959, %f13160; setp.neu.ftz.f32 %p1275, %f13161, 0f00000000; selp.u16 %rs622, 1, 0, %p1275; $L__BB107_1340: setp.eq.s16 %p1276, %rs622, 0; mov.u16 %rs623, %rs340; @%p1276 bra $L__BB107_1348; setp.eq.s32 %p1277, %r621, 0; @%p1277 bra $L__BB107_1345; setp.ne.s32 %p1278, %r622, 0; @%p1278 bra $L__BB107_1344; bra.uni $L__BB107_1343; $L__BB107_1344: mov.u32 %r2777, 4; add.u64 %rd2236, %SP, 1488; { // callseq 333, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2179; .param .b64 param1; st.param.b64 [param1+0], %rd2236; .param .b32 param2; st.param.b32 [param2+0], %r2777; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r2778, [retval0+0]; } // callseq 333 cvt.u16.u32 %rs623, %r2778; bra.uni $L__BB107_1348; $L__BB107_1345: setp.ne.s32 %p1279, %r622, 0; @%p1279 bra $L__BB107_1347; bra.uni $L__BB107_1346; $L__BB107_1347: mov.u32 %r2781, 4; add.u64 %rd2240, %SP, 1488; { // callseq 335, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2179; .param .b64 param1; st.param.b64 [param1+0], %rd2240; .param .b32 param2; st.param.b32 [param2+0], %r2781; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r2782, [retval0+0]; } // callseq 335 cvt.u16.u32 %rs623, %r2782; bra.uni $L__BB107_1348; $L__BB107_1343: mov.u32 %r2775, 4; add.u64 %rd2234, %SP, 1488; { // callseq 332, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2179; .param .b64 param1; st.param.b64 [param1+0], %rd2234; .param .b32 param2; st.param.b32 [param2+0], %r2775; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r2776, [retval0+0]; } // callseq 332 cvt.u16.u32 %rs623, %r2776; bra.uni $L__BB107_1348; $L__BB107_1346: mov.u32 %r2779, 4; add.u64 %rd2238, %SP, 1488; { // callseq 334, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2179; .param .b64 param1; st.param.b64 [param1+0], %rd2238; .param .b32 param2; st.param.b32 [param2+0], %r2779; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r2780, [retval0+0]; } // callseq 334 cvt.u16.u32 %rs623, %r2780; $L__BB107_1348: setp.eq.s16 %p1280, %rs623, 0; @%p1280 bra $L__BB107_1753; add.u64 %rd360, %SPL, 1488; ld.local.u32 %r632, [%rd360+16]; cvt.s64.s32 %rd4795, %r632; mul.wide.s32 %rd2242, %r632, 4; add.s64 %rd4794, %rd287, %rd2242; ld.global.u32 %r5722, [%rd4794]; and.b32 %r2783, %r5722, 512; setp.eq.s32 %p1281, %r2783, 0; @%p1281 bra $L__BB107_1751; bra.uni $L__BB107_1350; $L__BB107_1751: ld.local.u32 %r5730, [%rd360+12]; ld.local.f32 %f29669, [%rd360]; bra.uni $L__BB107_1752; $L__BB107_1350: setp.gt.s32 %p1282, %r5732, 5; mov.u32 %r5802, 0; @%p1282 bra $L__BB107_3392; ld.local.u32 %r634, [%rd360+20]; and.b32 %r2786, %r634, 1; setp.eq.b32 %p1283, %r2786, 1; xor.pred %p1285, %p1283, %p1284; not.pred %p1286, %p1285; @%p1286 bra $L__BB107_3392; ld.local.f32 %f1967, [%rd360+4]; ld.local.f32 %f1968, [%rd360+8]; ld.local.f32 %f29669, [%rd360]; st.local.f32 [%rd288], %f29669; mov.u32 %r5802, 0; mov.b32 %r2788, %f1968; mov.b32 %r2789, %f1967; st.local.v4.u32 [%rd354], {%r634, %r2789, %r2788, %r632}; st.local.u32 [%rd290], %r5722; ld.local.u32 %r5730, [%rd360+12]; cvt.s64.s32 %rd363, %r5730; st.local.u32 [%rd291], %r5730; st.local.u32 [%rd293], %r5802; mov.b32 %r2790, %f1950; mov.u32 %r5716, -1; cvta.to.local.u64 %rd2244, %rd292; st.local.v2.u32 [%rd2244], {%r5716, %r2790}; and.b32 %r636, %r5722, 2; setp.eq.s32 %p1287, %r636, 0; mul.lo.s64 %rd2245, %rd4795, 208; add.s64 %rd364, %rd300, %rd2245; @%p1287 bra $L__BB107_1359; ld.global.u32 %rd2246, [%rd364+168]; ld.global.u32 %r2792, [%rd364+152]; shl.b32 %r2793, %r2792, 1; cvt.rn.f32.s32 %f13164, %r2793; mul.ftz.f32 %f13165, %f1950, %f13164; cvt.rzi.ftz.s32.f32 %r2794, %f13165; add.s32 %r2795, %r2793, -1; min.s32 %r2796, %r2794, %r2795; cvt.rn.f32.s32 %f13166, %r2796; sub.ftz.f32 %f1970, %f13165, %f13166; cvt.s64.s32 %rd2247, %r2796; add.s64 %rd2248, %rd2247, %rd2246; shl.b64 %rd2249, %rd2248, 6; add.s64 %rd365, %rd301, %rd2249; ld.global.v4.f32 {%f13167, %f13168, %f13169, %f13170}, [%rd365]; ld.global.v4.f32 {%f13171, %f13172, %f13173, %f13174}, [%rd365+64]; mul.ftz.f32 %f13175, %f13168, %f13172; fma.rn.ftz.f32 %f13176, %f13167, %f13171, %f13175; mul.ftz.f32 %f13177, %f13170, %f13174; fma.rn.ftz.f32 %f13178, %f13169, %f13173, %f13177; add.ftz.f32 %f1979, %f13176, %f13178; setp.gt.ftz.f32 %p1288, %f1979, 0f3F7FDF3B; mov.f32 %f13179, 0f3F800000; sub.ftz.f32 %f1980, %f13179, %f1970; @%p1288 bra $L__BB107_1355; bra.uni $L__BB107_1354; $L__BB107_1355: mul.ftz.f32 %f13233, %f13167, %f1980; fma.rn.ftz.f32 %f13234, %f13171, %f1970, %f13233; mul.ftz.f32 %f13235, %f13168, %f1980; fma.rn.ftz.f32 %f13236, %f13172, %f1970, %f13235; mul.ftz.f32 %f13237, %f13169, %f1980; fma.rn.ftz.f32 %f13238, %f13173, %f1970, %f13237; mul.ftz.f32 %f13239, %f13170, %f1980; fma.rn.ftz.f32 %f13240, %f13174, %f1970, %f13239; mul.ftz.f32 %f13241, %f13236, %f13236; fma.rn.ftz.f32 %f13242, %f13234, %f13234, %f13241; mul.ftz.f32 %f13243, %f13240, %f13240; fma.rn.ftz.f32 %f13244, %f13238, %f13238, %f13243; add.ftz.f32 %f13245, %f13242, %f13244; rsqrt.approx.ftz.f32 %f13246, %f13245; mul.ftz.f32 %f29453, %f13240, %f13246; mul.ftz.f32 %f29452, %f13238, %f13246; mul.ftz.f32 %f29451, %f13236, %f13246; mul.ftz.f32 %f29450, %f13234, %f13246; bra.uni $L__BB107_1356; $L__BB107_1354: mov.f32 %f13180, 0fBF800000; max.ftz.f32 %f13181, %f1979, %f13180; min.ftz.f32 %f13183, %f13181, %f13179; abs.ftz.f32 %f13184, %f13183; sub.ftz.f32 %f13185, %f13179, %f13184; mul.ftz.f32 %f13186, %f13185, 0f3F000000; sqrt.approx.ftz.f32 %f13187, %f13186; setp.gt.ftz.f32 %p1289, %f13184, 0f3F11EB85; selp.f32 %f13188, %f13187, %f13184, %p1289; mul.ftz.f32 %f13189, %f13188, %f13188; mov.f32 %f13190, 0f3C94D2E9; mov.f32 %f13191, 0f3D53F941; fma.rn.ftz.f32 %f13192, %f13191, %f13189, %f13190; mov.f32 %f13193, 0f3D3F841F; fma.rn.ftz.f32 %f13194, %f13192, %f13189, %f13193; mov.f32 %f13195, 0f3D994929; fma.rn.ftz.f32 %f13196, %f13194, %f13189, %f13195; mov.f32 %f13197, 0f3E2AAB94; fma.rn.ftz.f32 %f13198, %f13196, %f13189, %f13197; mul.ftz.f32 %f13199, %f13189, %f13198; fma.rn.ftz.f32 %f13200, %f13199, %f13188, %f13188; add.ftz.f32 %f13201, %f13200, %f13200; mov.f32 %f13202, 0f3FC90FDB; sub.ftz.f32 %f13203, %f13202, %f13200; selp.f32 %f13204, %f13201, %f13203, %p1289; setp.lt.ftz.f32 %p1290, %f13183, 0f00000000; mov.f32 %f13205, 0f40490FDB; sub.ftz.f32 %f13206, %f13205, %f13204; selp.f32 %f13207, %f13206, %f13204, %p1290; mul.ftz.f32 %f13208, %f13167, %f1979; sub.ftz.f32 %f13209, %f13171, %f13208; mul.ftz.f32 %f13210, %f13168, %f1979; sub.ftz.f32 %f13211, %f13172, %f13210; mul.ftz.f32 %f13212, %f13169, %f1979; sub.ftz.f32 %f13213, %f13173, %f13212; mul.ftz.f32 %f13214, %f13170, %f1979; sub.ftz.f32 %f13215, %f13174, %f13214; mul.ftz.f32 %f13216, %f13211, %f13211; fma.rn.ftz.f32 %f13217, %f13209, %f13209, %f13216; mul.ftz.f32 %f13218, %f13215, %f13215; fma.rn.ftz.f32 %f13219, %f13213, %f13213, %f13218; add.ftz.f32 %f13220, %f13217, %f13219; rsqrt.approx.ftz.f32 %f13221, %f13220; mul.ftz.f32 %f13222, %f13209, %f13221; mul.ftz.f32 %f13223, %f13211, %f13221; mul.ftz.f32 %f13224, %f13213, %f13221; mul.ftz.f32 %f13225, %f13215, %f13221; mul.ftz.f32 %f13226, %f1970, %f13207; cos.approx.ftz.f32 %f13227, %f13226; sin.approx.ftz.f32 %f13228, %f13226; mul.ftz.f32 %f13229, %f13222, %f13228; mul.ftz.f32 %f13230, %f13223, %f13228; mul.ftz.f32 %f13231, %f13224, %f13228; mul.ftz.f32 %f13232, %f13225, %f13228; fma.rn.ftz.f32 %f29453, %f13170, %f13227, %f13232; fma.rn.ftz.f32 %f29452, %f13169, %f13227, %f13231; fma.rn.ftz.f32 %f29451, %f13168, %f13227, %f13230; fma.rn.ftz.f32 %f29450, %f13167, %f13227, %f13229; $L__BB107_1356: ld.global.v4.f32 {%f13247, %f13248, %f13249, %f13250}, [%rd365+16]; ld.global.v4.f32 {%f13255, %f13256, %f13257, %f13258}, [%rd365+80]; mul.ftz.f32 %f13263, %f1970, %f13255; mul.ftz.f32 %f13264, %f1970, %f13256; mul.ftz.f32 %f13265, %f1970, %f13257; mul.ftz.f32 %f13266, %f1970, %f13258; fma.rn.ftz.f32 %f1993, %f1980, %f13247, %f13263; fma.rn.ftz.f32 %f1994, %f1980, %f13248, %f13264; fma.rn.ftz.f32 %f1995, %f1980, %f13249, %f13265; fma.rn.ftz.f32 %f13267, %f1980, %f13250, %f13266; ld.global.v4.f32 {%f13268, %f13269, %f13270, %f13271}, [%rd365+32]; ld.global.v4.f32 {%f13276, %f13277, %f13278, %f13279}, [%rd365+96]; mul.ftz.f32 %f13284, %f1970, %f13276; mul.ftz.f32 %f13285, %f1970, %f13277; mul.ftz.f32 %f13286, %f1970, %f13278; mul.ftz.f32 %f13287, %f1970, %f13279; fma.rn.ftz.f32 %f13288, %f1980, %f13268, %f13284; fma.rn.ftz.f32 %f13289, %f1980, %f13269, %f13285; fma.rn.ftz.f32 %f13290, %f1980, %f13270, %f13286; fma.rn.ftz.f32 %f13291, %f1980, %f13271, %f13287; ld.global.v4.f32 {%f13292, %f13293, %f13294, %f13295}, [%rd365+48]; ld.global.v4.f32 {%f13300, %f13301, %f13302, %f13303}, [%rd365+112]; mul.ftz.f32 %f13308, %f1970, %f13300; mul.ftz.f32 %f13309, %f1970, %f13301; mul.ftz.f32 %f13310, %f1970, %f13302; mul.ftz.f32 %f13311, %f1970, %f13303; fma.rn.ftz.f32 %f13312, %f1980, %f13292, %f13308; fma.rn.ftz.f32 %f13313, %f1980, %f13293, %f13309; fma.rn.ftz.f32 %f13314, %f1980, %f13294, %f13310; fma.rn.ftz.f32 %f13315, %f1980, %f13295, %f13311; mul.ftz.f32 %f13316, %f29453, 0f3FB504F3; mul.ftz.f32 %f13317, %f29450, 0f3FB504F3; mul.ftz.f32 %f13318, %f29451, 0f3FB504F3; mul.ftz.f32 %f13319, %f29452, 0f3FB504F3; mul.ftz.f32 %f13320, %f13316, %f13317; mul.ftz.f32 %f13321, %f13316, %f13318; mul.ftz.f32 %f13322, %f13316, %f13319; mul.ftz.f32 %f13323, %f13317, %f13317; mul.ftz.f32 %f13324, %f13317, %f13318; mul.ftz.f32 %f13325, %f13317, %f13319; mul.ftz.f32 %f13326, %f13318, %f13318; mul.ftz.f32 %f13327, %f13318, %f13319; mul.ftz.f32 %f13328, %f13319, %f13319; sub.ftz.f32 %f13330, %f13179, %f13326; sub.ftz.f32 %f13331, %f13330, %f13328; sub.ftz.f32 %f13332, %f13324, %f13322; add.ftz.f32 %f13333, %f13321, %f13325; add.ftz.f32 %f13334, %f13322, %f13324; sub.ftz.f32 %f13335, %f13179, %f13323; sub.ftz.f32 %f13336, %f13335, %f13328; sub.ftz.f32 %f13337, %f13327, %f13320; sub.ftz.f32 %f13338, %f13325, %f13321; add.ftz.f32 %f13339, %f13320, %f13327; sub.ftz.f32 %f13340, %f13335, %f13326; mul.ftz.f32 %f13341, %f13332, %f13290; fma.rn.ftz.f32 %f13342, %f13331, %f13267, %f13341; fma.rn.ftz.f32 %f29459, %f13333, %f13313, %f13342; mul.ftz.f32 %f13343, %f13332, %f13291; fma.rn.ftz.f32 %f13344, %f13331, %f13288, %f13343; fma.rn.ftz.f32 %f1997, %f13333, %f13314, %f13344; mul.ftz.f32 %f13345, %f13332, %f13312; fma.rn.ftz.f32 %f13346, %f13331, %f13289, %f13345; fma.rn.ftz.f32 %f1998, %f13333, %f13315, %f13346; mul.ftz.f32 %f13347, %f13336, %f13290; fma.rn.ftz.f32 %f13348, %f13334, %f13267, %f13347; fma.rn.ftz.f32 %f1999, %f13337, %f13313, %f13348; mul.ftz.f32 %f13349, %f13336, %f13291; fma.rn.ftz.f32 %f13350, %f13334, %f13288, %f13349; fma.rn.ftz.f32 %f29458, %f13337, %f13314, %f13350; mul.ftz.f32 %f13351, %f13336, %f13312; fma.rn.ftz.f32 %f13352, %f13334, %f13289, %f13351; fma.rn.ftz.f32 %f2001, %f13337, %f13315, %f13352; mul.ftz.f32 %f13353, %f13339, %f13290; fma.rn.ftz.f32 %f13354, %f13338, %f13267, %f13353; fma.rn.ftz.f32 %f2002, %f13340, %f13313, %f13354; mul.ftz.f32 %f13355, %f13339, %f13291; fma.rn.ftz.f32 %f13356, %f13338, %f13288, %f13355; fma.rn.ftz.f32 %f2003, %f13340, %f13314, %f13356; mul.ftz.f32 %f13357, %f13339, %f13312; fma.rn.ftz.f32 %f13358, %f13338, %f13289, %f13357; fma.rn.ftz.f32 %f29457, %f13340, %f13315, %f13358; cvta.to.local.u64 %rd2250, %rd299; st.local.v4.f32 [%rd2250], {%f29459, %f1997, %f1998, %f1993}; cvta.to.local.u64 %rd2251, %rd311; st.local.v4.f32 [%rd2251], {%f1999, %f29458, %f2001, %f1994}; cvta.to.local.u64 %rd2252, %rd310; st.local.v4.f32 [%rd2252], {%f2002, %f2003, %f29457, %f1995}; mul.ftz.f32 %f13359, %f29458, %f29457; mul.ftz.f32 %f2005, %f2003, %f2001; sub.ftz.f32 %f29456, %f13359, %f2005; mul.ftz.f32 %f2007, %f2003, %f1998; mul.ftz.f32 %f13360, %f1997, %f29457; sub.ftz.f32 %f29455, %f2007, %f13360; mul.ftz.f32 %f2009, %f1997, %f2001; mul.ftz.f32 %f13361, %f29458, %f1998; sub.ftz.f32 %f29454, %f2009, %f13361; mul.ftz.f32 %f13362, %f1999, %f29455; fma.rn.ftz.f32 %f13363, %f29459, %f29456, %f13362; fma.rn.ftz.f32 %f29460, %f2002, %f29454, %f13363; setp.neu.ftz.f32 %p1291, %f29460, 0f00000000; @%p1291 bra $L__BB107_1358; add.ftz.f32 %f29459, %f29459, 0f322BCC77; add.ftz.f32 %f29457, %f29457, 0f322BCC77; add.ftz.f32 %f29458, %f29458, 0f322BCC77; mul.ftz.f32 %f13364, %f29458, %f29457; sub.ftz.f32 %f29456, %f13364, %f2005; mul.ftz.f32 %f13365, %f1997, %f29457; sub.ftz.f32 %f29455, %f2007, %f13365; mul.ftz.f32 %f13366, %f29458, %f1998; sub.ftz.f32 %f29454, %f2009, %f13366; mul.ftz.f32 %f13367, %f1999, %f29455; fma.rn.ftz.f32 %f13368, %f29459, %f29456, %f13367; fma.rn.ftz.f32 %f13369, %f2002, %f29454, %f13368; setp.eq.ftz.f32 %p1292, %f13369, 0f00000000; selp.f32 %f29460, 0f7F7FFFFF, %f13369, %p1292; $L__BB107_1358: mul.ftz.f32 %f13370, %f1999, %f29457; mul.ftz.f32 %f13371, %f2002, %f2001; sub.ftz.f32 %f13372, %f13371, %f13370; mul.ftz.f32 %f13373, %f2002, %f1998; mul.ftz.f32 %f13374, %f29457, %f29459; sub.ftz.f32 %f13375, %f13374, %f13373; mul.ftz.f32 %f13376, %f2001, %f29459; mul.ftz.f32 %f13377, %f1999, %f1998; sub.ftz.f32 %f13378, %f13377, %f13376; mul.ftz.f32 %f13379, %f2002, %f29458; mul.ftz.f32 %f13380, %f1999, %f2003; sub.ftz.f32 %f13381, %f13380, %f13379; mul.ftz.f32 %f13382, %f2003, %f29459; mul.ftz.f32 %f13383, %f2002, %f1997; sub.ftz.f32 %f13384, %f13383, %f13382; mul.ftz.f32 %f13385, %f1999, %f1997; mul.ftz.f32 %f13386, %f29458, %f29459; sub.ftz.f32 %f13387, %f13386, %f13385; div.approx.ftz.f32 %f13388, %f29455, %f29460; mul.ftz.f32 %f13389, %f1994, %f13388; div.approx.ftz.f32 %f13390, %f29456, %f29460; fma.rn.ftz.f32 %f13391, %f1993, %f13390, %f13389; div.approx.ftz.f32 %f13392, %f29454, %f29460; fma.rn.ftz.f32 %f13393, %f1995, %f13392, %f13391; div.approx.ftz.f32 %f13394, %f13375, %f29460; mul.ftz.f32 %f13395, %f1994, %f13394; div.approx.ftz.f32 %f13396, %f13372, %f29460; fma.rn.ftz.f32 %f13397, %f1993, %f13396, %f13395; div.approx.ftz.f32 %f13398, %f13378, %f29460; fma.rn.ftz.f32 %f13399, %f1995, %f13398, %f13397; div.approx.ftz.f32 %f13400, %f13384, %f29460; mul.ftz.f32 %f13401, %f1994, %f13400; div.approx.ftz.f32 %f13402, %f13381, %f29460; fma.rn.ftz.f32 %f13403, %f1993, %f13402, %f13401; div.approx.ftz.f32 %f13404, %f13387, %f29460; fma.rn.ftz.f32 %f13405, %f1995, %f13404, %f13403; neg.ftz.f32 %f13406, %f13393; cvta.to.local.u64 %rd2253, %rd312; st.local.v4.f32 [%rd2253], {%f13390, %f13388, %f13392, %f13406}; neg.ftz.f32 %f13407, %f13399; cvta.to.local.u64 %rd2254, %rd327; st.local.v4.f32 [%rd2254], {%f13396, %f13394, %f13398, %f13407}; neg.ftz.f32 %f13408, %f13405; cvta.to.local.u64 %rd2255, %rd326; st.local.v4.f32 [%rd2255], {%f13402, %f13400, %f13404, %f13408}; $L__BB107_1359: neg.ftz.f32 %f2026, %f1961; st.local.f32 [%rd294], %f2026; neg.ftz.f32 %f2027, %f1960; st.local.f32 [%rd296], %f2027; neg.ftz.f32 %f13409, %f1959; st.local.f32 [%rd298], %f13409; and.b32 %r2797, %r634, 6; setp.eq.s32 %p1293, %r2797, 0; shl.b64 %rd2256, %rd363, 4; add.s64 %rd366, %rd328, %rd2256; shl.b64 %rd2257, %rd363, 2; add.s64 %rd367, %rd333, %rd2257; @%p1293 bra $L__BB107_1397; and.b32 %r637, %r5722, 4; setp.ne.s32 %p1294, %r637, 0; mov.f32 %f29476, %f29669; mov.f32 %f29477, %f1958; mov.f32 %f29478, %f1957; mov.f32 %f29479, %f1956; mov.f32 %f29480, %f13137; mov.f32 %f29481, %f13138; mov.f32 %f29482, %f13139; @%p1294 bra $L__BB107_1366; @%p1287 bra $L__BB107_1363; bra.uni $L__BB107_1362; $L__BB107_1363: ld.global.v4.f32 {%f29461, %f29462, %f29463, %f29464}, [%rd364+48]; ld.global.v4.f32 {%f29465, %f29466, %f29467, %f29468}, [%rd364+64]; ld.global.v4.f32 {%f29469, %f29470, %f29471, %f29472}, [%rd364+80]; bra.uni $L__BB107_1364; $L__BB107_1397: and.b32 %r2824, %r634, 8; setp.eq.s32 %p1318, %r2824, 0; @%p1318 bra $L__BB107_1418; bra.uni $L__BB107_1398; $L__BB107_1418: setp.eq.s32 %p1331, %r634, 1; @%p1331 bra $L__BB107_1420; bra.uni $L__BB107_1419; $L__BB107_1420: ld.global.u32 %r2841, [%rd366+12]; mul.wide.u32 %rd2317, %r2841, 12; add.s64 %rd2318, %rd329, %rd2317; add.s32 %r2842, %r2841, 1; mul.wide.u32 %rd2319, %r2842, 12; add.s64 %rd2320, %rd329, %rd2319; add.s32 %r2843, %r2841, 2; mul.wide.u32 %rd2321, %r2843, 12; add.s64 %rd2322, %rd329, %rd2321; and.b32 %r2844, %r5722, 8; setp.eq.s32 %p1332, %r2844, 0; ld.global.f32 %f13963, [%rd2322]; ld.global.f32 %f2305, [%rd2318]; sub.ftz.f32 %f2306, %f13963, %f2305; ld.global.f32 %f13964, [%rd2322+4]; ld.global.f32 %f2307, [%rd2318+4]; sub.ftz.f32 %f2308, %f13964, %f2307; ld.global.f32 %f13965, [%rd2322+8]; ld.global.f32 %f2309, [%rd2318+8]; sub.ftz.f32 %f2310, %f13965, %f2309; ld.global.f32 %f13966, [%rd2320]; sub.ftz.f32 %f2311, %f13966, %f2305; ld.global.f32 %f13967, [%rd2320+4]; sub.ftz.f32 %f2312, %f13967, %f2307; ld.global.f32 %f13968, [%rd2320+8]; sub.ftz.f32 %f2313, %f13968, %f2309; mul.ftz.f32 %f2314, %f2313, %f2308; mul.ftz.f32 %f2315, %f2312, %f2310; mul.ftz.f32 %f2316, %f2311, %f2310; mul.ftz.f32 %f2317, %f2313, %f2306; mul.ftz.f32 %f2318, %f2312, %f2306; mul.ftz.f32 %f2319, %f2311, %f2308; @%p1332 bra $L__BB107_1422; bra.uni $L__BB107_1421; $L__BB107_1422: sub.ftz.f32 %f13976, %f2315, %f2314; sub.ftz.f32 %f13977, %f2317, %f2316; mul.ftz.f32 %f13978, %f13977, %f13977; fma.rn.ftz.f32 %f13979, %f13976, %f13976, %f13978; sub.ftz.f32 %f13980, %f2319, %f2318; fma.rn.ftz.f32 %f13981, %f13980, %f13980, %f13979; sqrt.approx.ftz.f32 %f13982, %f13981; div.approx.ftz.f32 %f29547, %f13980, %f13982; div.approx.ftz.f32 %f29546, %f13977, %f13982; div.approx.ftz.f32 %f29545, %f13976, %f13982; bra.uni $L__BB107_1423; $L__BB107_1398: add.s64 %rd2298, %rd313, %rd2257; ld.global.u32 %r2825, [%rd2298]; st.local.u32 [%rd305], %r2825; fma.rn.ftz.f32 %f2268, %f29669, %f1958, %f13137; fma.rn.ftz.f32 %f2269, %f29669, %f1957, %f13138; fma.rn.ftz.f32 %f2270, %f29669, %f1956, %f13139; st.local.v2.f32 [%rd2173], {%f2268, %f2269}; st.local.f32 [%rd314], %f2270; and.b32 %r2826, %r634, 64; setp.eq.s32 %p1319, %r2826, 0; @%p1319 bra $L__BB107_1412; bra.uni $L__BB107_1399; $L__BB107_1412: add.s64 %rd2310, %rd330, %rd2256; ld.global.v4.f32 {%f29542, %f29543, %f29544, %f13888}, [%rd2310]; bra.uni $L__BB107_1413; $L__BB107_1362: cvta.to.local.u64 %rd2258, %rd312; ld.local.v4.f32 {%f29461, %f29462, %f29463, %f29464}, [%rd2258]; cvta.to.local.u64 %rd2259, %rd327; ld.local.v4.f32 {%f29465, %f29466, %f29467, %f29468}, [%rd2259]; cvta.to.local.u64 %rd2260, %rd326; ld.local.v4.f32 {%f29469, %f29470, %f29471, %f29472}, [%rd2260]; $L__BB107_1364: mul.ftz.f32 %f13434, %f13138, %f29462; fma.rn.ftz.f32 %f13435, %f13137, %f29461, %f13434; fma.rn.ftz.f32 %f13436, %f13139, %f29463, %f13435; add.ftz.f32 %f29480, %f29464, %f13436; mul.ftz.f32 %f13437, %f13138, %f29466; fma.rn.ftz.f32 %f13438, %f13137, %f29465, %f13437; fma.rn.ftz.f32 %f13439, %f13139, %f29467, %f13438; add.ftz.f32 %f29481, %f29468, %f13439; mul.ftz.f32 %f13440, %f13138, %f29470; fma.rn.ftz.f32 %f13441, %f13137, %f29469, %f13440; fma.rn.ftz.f32 %f13442, %f13139, %f29471, %f13441; add.ftz.f32 %f29482, %f29472, %f13442; mul.ftz.f32 %f13443, %f1958, %f29669; mul.ftz.f32 %f13444, %f1957, %f29669; mul.ftz.f32 %f13445, %f13444, %f29462; fma.rn.ftz.f32 %f13446, %f13443, %f29461, %f13445; mul.ftz.f32 %f13447, %f1956, %f29669; fma.rn.ftz.f32 %f29477, %f13447, %f29463, %f13446; mul.ftz.f32 %f13448, %f13444, %f29466; fma.rn.ftz.f32 %f13449, %f13443, %f29465, %f13448; fma.rn.ftz.f32 %f29478, %f13447, %f29467, %f13449; mul.ftz.f32 %f13450, %f13444, %f29470; fma.rn.ftz.f32 %f13451, %f13443, %f29469, %f13450; fma.rn.ftz.f32 %f29479, %f13447, %f29471, %f13451; mul.ftz.f32 %f13452, %f29478, %f29478; fma.rn.ftz.f32 %f13453, %f29477, %f29477, %f13452; fma.rn.ftz.f32 %f13454, %f29479, %f29479, %f13453; sqrt.approx.ftz.f32 %f29476, %f13454; setp.neu.ftz.f32 %p1296, %f29476, 0f00000000; @%p1296 bra $L__BB107_1365; bra.uni $L__BB107_1366; $L__BB107_1365: div.approx.ftz.f32 %f29479, %f29479, %f29476; div.approx.ftz.f32 %f29478, %f29478, %f29476; div.approx.ftz.f32 %f29477, %f29477, %f29476; $L__BB107_1366: add.s64 %rd2262, %rd325, %rd2256; add.s64 %rd368, %rd2262, 4; shr.s32 %r2798, %r634, 7; ld.global.u32 %r2799, [%rd2262+4]; add.s32 %r638, %r2798, %r2799; add.s32 %r2800, %r638, -1; max.s32 %r639, %r2800, %r2799; add.s32 %r2801, %r638, 2; ld.global.u32 %r2802, [%rd2262+8]; add.s32 %r2803, %r2799, %r2802; add.s32 %r2804, %r2803, -1; min.s32 %r640, %r2801, %r2804; and.b32 %r2805, %r634, 64; setp.eq.s32 %p1297, %r2805, 0; mul.wide.s32 %rd2263, %r639, 16; add.s64 %rd369, %rd332, %rd2263; mul.wide.s32 %rd2264, %r638, 16; add.s64 %rd370, %rd332, %rd2264; mul.wide.s32 %rd2265, %r640, 16; add.s64 %rd371, %rd332, %rd2265; @%p1297 bra $L__BB107_1380; bra.uni $L__BB107_1367; $L__BB107_1380: ld.global.v4.f32 {%f29518, %f29517, %f29516, %f13539}, [%rd369]; ld.global.v4.f32 {%f29515, %f29514, %f29513, %f13543}, [%rd370]; ld.global.v4.f32 {%f29512, %f29511, %f29510, %f13547}, [%rd370+16]; ld.global.v4.f32 {%f29509, %f29508, %f29507, %f13551}, [%rd371]; bra.uni $L__BB107_1381; $L__BB107_1367: ld.global.u32 %r641, [%rd364+148]; ld.global.u32 %r642, [%rd364+152]; shl.b32 %r2806, %r642, 1; cvt.rn.f32.s32 %f13455, %r2806; mul.ftz.f32 %f2090, %f1950, %f13455; cvt.rzi.ftz.s32.f32 %r2807, %f2090; add.s32 %r2808, %r2806, -1; min.s32 %r643, %r2807, %r2808; ld.global.u32 %r5713, [%rd364+164]; mul.wide.u32 %rd2266, %r5713, 16; add.s64 %rd2267, %rd331, %rd2266; ld.global.u16 %rs624, [%rd2267+12]; ld.global.u32 %r5712, [%rd2267+8]; ld.global.u64 %rd4790, [%rd2267]; setp.eq.s64 %p1298, %rd4790, 11; @%p1298 bra $L__BB107_1372; $L__BB107_1368: setp.eq.s64 %p1299, %rd4790, 0; @%p1299 bra $L__BB107_1370; add.s32 %r5713, %r5713, 2; bra.uni $L__BB107_1371; $L__BB107_1370: setp.eq.s16 %p1300, %rs624, 0; mov.u32 %r5713, %r5712; @%p1300 bra $L__BB107_1373; $L__BB107_1371: mul.wide.u32 %rd2268, %r5713, 16; add.s64 %rd2269, %rd331, %rd2268; ld.global.u16 %rs624, [%rd2269+12]; ld.global.u32 %r5712, [%rd2269+8]; ld.global.u64 %rd4790, [%rd2269]; setp.ne.s64 %p1301, %rd4790, 11; @%p1301 bra $L__BB107_1368; $L__BB107_1372: setp.eq.s16 %p1302, %rs624, 0; selp.b32 %r5716, -1, %r5712, %p1302; $L__BB107_1373: setp.eq.s32 %p1303, %r643, %r642; @%p1303 bra $L__BB107_1375; bra.uni $L__BB107_1374; $L__BB107_1375: ld.global.v4.f32 {%f29494, %f29493, %f29492, %f13475}, [%rd369]; ld.global.v4.f32 {%f29491, %f29490, %f29489, %f13479}, [%rd370]; ld.global.v4.f32 {%f29488, %f29487, %f29486, %f13483}, [%rd370+16]; ld.global.v4.f32 {%f29485, %f29484, %f29483, %f13487}, [%rd371]; bra.uni $L__BB107_1376; $L__BB107_1374: setp.gt.s32 %p1304, %r643, %r642; selp.b32 %r2810, -1, 0, %p1304; add.s32 %r2811, %r643, %r2810; mad.lo.s32 %r2812, %r2811, %r641, %r5716; add.s32 %r2813, %r2812, %r639; mul.wide.s32 %rd2270, %r2813, 16; add.s64 %rd2271, %rd341, %rd2270; ld.global.v4.f32 {%f29494, %f29493, %f29492, %f13459}, [%rd2271]; add.s32 %r2814, %r2812, %r638; mul.wide.s32 %rd2272, %r2814, 16; add.s64 %rd2273, %rd341, %rd2272; ld.global.v4.f32 {%f29491, %f29490, %f29489, %f13463}, [%rd2273]; ld.global.v4.f32 {%f29488, %f29487, %f29486, %f13467}, [%rd2273+16]; add.s32 %r2815, %r2812, %r640; mul.wide.s32 %rd2274, %r2815, 16; add.s64 %rd2275, %rd341, %rd2274; ld.global.v4.f32 {%f29485, %f29484, %f29483, %f13471}, [%rd2275]; $L__BB107_1376: add.s32 %r654, %r643, 1; setp.eq.s32 %p1305, %r654, %r642; @%p1305 bra $L__BB107_1378; bra.uni $L__BB107_1377; $L__BB107_1378: ld.global.v4.f32 {%f29506, %f29505, %f29504, %f13507}, [%rd369]; ld.global.v4.f32 {%f29503, %f29502, %f29501, %f13511}, [%rd370]; ld.global.v4.f32 {%f29500, %f29499, %f29498, %f13515}, [%rd370+16]; ld.global.v4.f32 {%f29497, %f29496, %f29495, %f13519}, [%rd371]; bra.uni $L__BB107_1379; $L__BB107_1377: setp.ge.s32 %p1306, %r643, %r642; selp.b32 %r2816, -1, 0, %p1306; add.s32 %r2817, %r654, %r2816; mad.lo.s32 %r2818, %r2817, %r641, %r5716; add.s32 %r2819, %r2818, %r639; mul.wide.s32 %rd2276, %r2819, 16; add.s64 %rd2277, %rd341, %rd2276; ld.global.v4.f32 {%f29506, %f29505, %f29504, %f13491}, [%rd2277]; add.s32 %r2820, %r2818, %r638; mul.wide.s32 %rd2278, %r2820, 16; add.s64 %rd2279, %rd341, %rd2278; ld.global.v4.f32 {%f29503, %f29502, %f29501, %f13495}, [%rd2279]; ld.global.v4.f32 {%f29500, %f29499, %f29498, %f13499}, [%rd2279+16]; add.s32 %r2821, %r2818, %r640; mul.wide.s32 %rd2280, %r2821, 16; add.s64 %rd2281, %rd341, %rd2280; ld.global.v4.f32 {%f29497, %f29496, %f29495, %f13503}, [%rd2281]; $L__BB107_1379: cvt.rn.f32.s32 %f13520, %r643; sub.ftz.f32 %f13521, %f2090, %f13520; mov.f32 %f13522, 0f3F800000; sub.ftz.f32 %f13523, %f13522, %f13521; mul.ftz.f32 %f13524, %f13521, %f29506; fma.rn.ftz.f32 %f29518, %f13523, %f29494, %f13524; mul.ftz.f32 %f13525, %f13521, %f29505; fma.rn.ftz.f32 %f29517, %f13523, %f29493, %f13525; mul.ftz.f32 %f13526, %f13521, %f29504; fma.rn.ftz.f32 %f29516, %f13523, %f29492, %f13526; mul.ftz.f32 %f13527, %f13521, %f29503; fma.rn.ftz.f32 %f29515, %f13523, %f29491, %f13527; mul.ftz.f32 %f13528, %f13521, %f29502; fma.rn.ftz.f32 %f29514, %f13523, %f29490, %f13528; mul.ftz.f32 %f13529, %f13521, %f29501; fma.rn.ftz.f32 %f29513, %f13523, %f29489, %f13529; mul.ftz.f32 %f13530, %f13521, %f29500; fma.rn.ftz.f32 %f29512, %f13523, %f29488, %f13530; mul.ftz.f32 %f13531, %f13521, %f29499; fma.rn.ftz.f32 %f29511, %f13523, %f29487, %f13531; mul.ftz.f32 %f13532, %f13521, %f29498; fma.rn.ftz.f32 %f29510, %f13523, %f29486, %f13532; mul.ftz.f32 %f13533, %f13521, %f29497; fma.rn.ftz.f32 %f29509, %f13523, %f29485, %f13533; mul.ftz.f32 %f13534, %f13521, %f29496; fma.rn.ftz.f32 %f29508, %f13523, %f29484, %f13534; mul.ftz.f32 %f13535, %f13521, %f29495; fma.rn.ftz.f32 %f29507, %f13523, %f29483, %f13535; $L__BB107_1381: fma.rn.ftz.f32 %f29536, %f29476, %f29477, %f29480; fma.rn.ftz.f32 %f29537, %f29476, %f29478, %f29481; fma.rn.ftz.f32 %f29538, %f29476, %f29479, %f29482; mov.f32 %f13552, 0f3F800000; sub.ftz.f32 %f2202, %f13552, %f1967; mul.ftz.f32 %f2203, %f2202, %f2202; add.ftz.f32 %f13553, %f2202, %f2202; mul.ftz.f32 %f13554, %f1967, %f13553; sub.ftz.f32 %f13555, %f13554, %f2203; mul.ftz.f32 %f13556, %f1967, 0f40400000; add.ftz.f32 %f2204, %f13556, 0fC0A00000; add.ftz.f32 %f13557, %f1967, %f1967; mul.ftz.f32 %f13558, %f13557, %f2204; fma.rn.ftz.f32 %f13559, %f1967, %f13556, %f13558; add.ftz.f32 %f13560, %f13556, 0f40000000; mul.ftz.f32 %f13561, %f13553, %f13560; mul.ftz.f32 %f2205, %f2202, 0f40400000; mul.ftz.f32 %f13562, %f2202, %f2205; sub.ftz.f32 %f13563, %f13561, %f13562; mul.ftz.f32 %f13564, %f2202, 0fC0000000; mul.ftz.f32 %f2206, %f1967, %f1967; fma.rn.ftz.f32 %f13565, %f1967, %f13564, %f2206; mul.ftz.f32 %f13566, %f29515, %f13559; mul.ftz.f32 %f13567, %f29514, %f13559; mul.ftz.f32 %f13568, %f29513, %f13559; fma.rn.ftz.f32 %f13569, %f29518, %f13555, %f13566; fma.rn.ftz.f32 %f13570, %f29517, %f13555, %f13567; fma.rn.ftz.f32 %f13571, %f29516, %f13555, %f13568; fma.rn.ftz.f32 %f13572, %f29512, %f13563, %f13569; fma.rn.ftz.f32 %f13573, %f29511, %f13563, %f13570; fma.rn.ftz.f32 %f13574, %f29510, %f13563, %f13571; fma.rn.ftz.f32 %f13575, %f29509, %f13565, %f13572; fma.rn.ftz.f32 %f13576, %f29508, %f13565, %f13573; fma.rn.ftz.f32 %f13577, %f29507, %f13565, %f13574; mul.ftz.f32 %f29531, %f13575, 0f3F000000; mul.ftz.f32 %f29533, %f13576, 0f3F000000; mul.ftz.f32 %f29532, %f13577, 0f3F000000; and.b32 %r655, %r634, 4; setp.eq.s32 %p1307, %r655, 0; @%p1307 bra $L__BB107_1383; bra.uni $L__BB107_1382; $L__BB107_1383: mul.ftz.f32 %f13636, %f1967, %f2202; mul.ftz.f32 %f13637, %f2202, %f13636; fma.rn.ftz.f32 %f13638, %f2206, %f2204, 0f40000000; add.ftz.f32 %f13639, %f2205, 0fC0A00000; fma.rn.ftz.f32 %f13640, %f2203, %f13639, 0f40000000; mul.ftz.f32 %f13641, %f1967, %f13636; mul.ftz.f32 %f13642, %f29518, %f13637; mul.ftz.f32 %f13643, %f29517, %f13637; mul.ftz.f32 %f13644, %f29516, %f13637; mul.ftz.f32 %f13645, %f29515, %f13638; mul.ftz.f32 %f13646, %f29514, %f13638; mul.ftz.f32 %f13647, %f29513, %f13638; sub.ftz.f32 %f13648, %f13645, %f13642; sub.ftz.f32 %f13649, %f13646, %f13643; sub.ftz.f32 %f13650, %f13647, %f13644; fma.rn.ftz.f32 %f13651, %f29512, %f13640, %f13648; fma.rn.ftz.f32 %f13652, %f29511, %f13640, %f13649; fma.rn.ftz.f32 %f13653, %f29510, %f13640, %f13650; mul.ftz.f32 %f13654, %f29509, %f13641; mul.ftz.f32 %f13655, %f29508, %f13641; mul.ftz.f32 %f13656, %f29507, %f13641; sub.ftz.f32 %f13657, %f13651, %f13654; sub.ftz.f32 %f13658, %f13652, %f13655; sub.ftz.f32 %f13659, %f13653, %f13656; mul.ftz.f32 %f2213, %f13657, 0f3F000000; mul.ftz.f32 %f2214, %f13658, 0f3F000000; mul.ftz.f32 %f2215, %f13659, 0f3F000000; setp.eq.ftz.f32 %p1308, %f29536, %f2213; setp.eq.ftz.f32 %p1309, %f29537, %f2214; and.pred %p1310, %p1308, %p1309; setp.eq.ftz.f32 %p1311, %f29538, %f2215; and.pred %p1312, %p1311, %p1310; @%p1312 bra $L__BB107_1385; sub.ftz.f32 %f13660, %f29536, %f2213; sub.ftz.f32 %f13661, %f29537, %f2214; mul.ftz.f32 %f13662, %f13661, %f13661; fma.rn.ftz.f32 %f13663, %f13660, %f13660, %f13662; sub.ftz.f32 %f13664, %f29538, %f2215; fma.rn.ftz.f32 %f13665, %f13664, %f13664, %f13663; sqrt.approx.ftz.f32 %f13666, %f13665; div.approx.ftz.f32 %f1959, %f13664, %f13666; div.approx.ftz.f32 %f1960, %f13661, %f13666; div.approx.ftz.f32 %f1961, %f13660, %f13666; $L__BB107_1385: st.local.f32 [%rd315], %f1961; st.local.f32 [%rd337], %f1960; st.local.f32 [%rd338], %f1959; st.local.u32 [%rd351], %r5802; bra.uni $L__BB107_1386; $L__BB107_1382: mul.ftz.f32 %f13578, %f29533, %f29533; fma.rn.ftz.f32 %f13579, %f29531, %f29531, %f13578; fma.rn.ftz.f32 %f13580, %f29532, %f29532, %f13579; sqrt.approx.ftz.f32 %f13581, %f13580; div.approx.ftz.f32 %f13582, %f29531, %f13581; div.approx.ftz.f32 %f13583, %f29533, %f13581; div.approx.ftz.f32 %f13584, %f29532, %f13581; mul.ftz.f32 %f13585, %f13583, %f29479; mul.ftz.f32 %f13586, %f13584, %f29478; sub.ftz.f32 %f13587, %f13586, %f13585; mul.ftz.f32 %f13588, %f13584, %f29477; mul.ftz.f32 %f13589, %f13582, %f29479; sub.ftz.f32 %f13590, %f13589, %f13588; mul.ftz.f32 %f13591, %f13582, %f29478; mul.ftz.f32 %f13592, %f13583, %f29477; sub.ftz.f32 %f13593, %f13592, %f13591; mul.ftz.f32 %f13594, %f13590, %f13590; fma.rn.ftz.f32 %f13595, %f13587, %f13587, %f13594; fma.rn.ftz.f32 %f13596, %f13593, %f13593, %f13595; sqrt.approx.ftz.f32 %f13597, %f13596; div.approx.ftz.f32 %f13598, %f13587, %f13597; div.approx.ftz.f32 %f13599, %f13590, %f13597; div.approx.ftz.f32 %f13600, %f13593, %f13597; mul.ftz.f32 %f13601, %f1968, %f1968; sub.ftz.f32 %f13603, %f13552, %f13601; mov.f32 %f13604, 0f00000000; max.ftz.f32 %f13605, %f13603, %f13604; sqrt.approx.ftz.f32 %f13606, %f13605; mul.ftz.f32 %f13607, %f13598, %f1968; mul.ftz.f32 %f13608, %f13599, %f1968; mul.ftz.f32 %f13609, %f13600, %f1968; mul.ftz.f32 %f13610, %f13583, %f13600; mul.ftz.f32 %f13611, %f13584, %f13599; sub.ftz.f32 %f13612, %f13610, %f13611; mul.ftz.f32 %f13613, %f13584, %f13598; mul.ftz.f32 %f13614, %f13582, %f13600; sub.ftz.f32 %f13615, %f13613, %f13614; mul.ftz.f32 %f13616, %f13582, %f13599; mul.ftz.f32 %f13617, %f13583, %f13598; sub.ftz.f32 %f13618, %f13616, %f13617; mul.ftz.f32 %f13619, %f13615, %f13615; fma.rn.ftz.f32 %f13620, %f13612, %f13612, %f13619; fma.rn.ftz.f32 %f13621, %f13618, %f13618, %f13620; sqrt.approx.ftz.f32 %f13622, %f13621; div.approx.ftz.f32 %f13623, %f13612, %f13622; div.approx.ftz.f32 %f13624, %f13615, %f13622; div.approx.ftz.f32 %f13625, %f13618, %f13622; mul.ftz.f32 %f13626, %f13606, %f13623; mul.ftz.f32 %f13627, %f13606, %f13624; mul.ftz.f32 %f13628, %f13606, %f13625; sub.ftz.f32 %f13629, %f13607, %f13626; sub.ftz.f32 %f13630, %f13608, %f13627; sub.ftz.f32 %f13631, %f13609, %f13628; mul.ftz.f32 %f13632, %f13630, %f13630; fma.rn.ftz.f32 %f13633, %f13629, %f13629, %f13632; fma.rn.ftz.f32 %f13634, %f13631, %f13631, %f13633; sqrt.approx.ftz.f32 %f13635, %f13634; div.approx.ftz.f32 %f1961, %f13629, %f13635; div.approx.ftz.f32 %f1960, %f13630, %f13635; div.approx.ftz.f32 %f1959, %f13631, %f13635; st.local.f32 [%rd315], %f1961; st.local.f32 [%rd337], %f1960; st.local.f32 [%rd338], %f1959; $L__BB107_1386: st.local.f32 [%rd318], %f29531; st.local.f32 [%rd319], %f29533; st.local.f32 [%rd320], %f29532; @%p1294 bra $L__BB107_1396; @%p1287 bra $L__BB107_1389; bra.uni $L__BB107_1388; $L__BB107_1389: ld.global.v4.f32 {%f13700, %f13701, %f13702, %f13703}, [%rd364]; ld.global.v4.f32 {%f13708, %f13709, %f13710, %f13711}, [%rd364+16]; ld.global.v4.f32 {%f13716, %f13717, %f13718, %f13719}, [%rd364+32]; mul.ftz.f32 %f13724, %f29537, %f13701; fma.rn.ftz.f32 %f13725, %f29536, %f13700, %f13724; fma.rn.ftz.f32 %f13726, %f29538, %f13702, %f13725; add.ftz.f32 %f29525, %f13703, %f13726; mul.ftz.f32 %f13727, %f29537, %f13709; fma.rn.ftz.f32 %f13728, %f29536, %f13708, %f13727; fma.rn.ftz.f32 %f13729, %f29538, %f13710, %f13728; add.ftz.f32 %f29526, %f13711, %f13729; mul.ftz.f32 %f13730, %f29537, %f13717; fma.rn.ftz.f32 %f13731, %f29536, %f13716, %f13730; fma.rn.ftz.f32 %f13732, %f29538, %f13718, %f13731; add.ftz.f32 %f29538, %f13719, %f13732; bra.uni $L__BB107_1390; $L__BB107_1388: cvta.to.local.u64 %rd2282, %rd299; ld.local.v4.f32 {%f13667, %f13668, %f13669, %f13670}, [%rd2282]; mul.ftz.f32 %f13675, %f29537, %f13668; fma.rn.ftz.f32 %f13676, %f29536, %f13667, %f13675; fma.rn.ftz.f32 %f13677, %f29538, %f13669, %f13676; add.ftz.f32 %f29525, %f13670, %f13677; cvta.to.local.u64 %rd2283, %rd311; ld.local.v4.f32 {%f13678, %f13679, %f13680, %f13681}, [%rd2283]; mul.ftz.f32 %f13686, %f29537, %f13679; fma.rn.ftz.f32 %f13687, %f29536, %f13678, %f13686; fma.rn.ftz.f32 %f13688, %f29538, %f13680, %f13687; add.ftz.f32 %f29526, %f13681, %f13688; cvta.to.local.u64 %rd2284, %rd310; ld.local.v4.f32 {%f13689, %f13690, %f13691, %f13692}, [%rd2284]; mul.ftz.f32 %f13697, %f29537, %f13690; fma.rn.ftz.f32 %f13698, %f29536, %f13689, %f13697; fma.rn.ftz.f32 %f13699, %f29538, %f13691, %f13698; add.ftz.f32 %f29538, %f13692, %f13699; $L__BB107_1390: @%p1287 bra $L__BB107_1392; bra.uni $L__BB107_1391; $L__BB107_1392: ld.global.v4.f32 {%f13767, %f13768, %f13769, %f13770}, [%rd364+48]; ld.global.v4.f32 {%f13774, %f13775, %f13776, %f13777}, [%rd364+64]; ld.global.v4.f32 {%f13781, %f13782, %f13783, %f13784}, [%rd364+80]; mul.ftz.f32 %f13788, %f13774, %f1960; fma.rn.ftz.f32 %f13789, %f13767, %f1961, %f13788; fma.rn.ftz.f32 %f13790, %f13781, %f1959, %f13789; mul.ftz.f32 %f13791, %f13775, %f1960; fma.rn.ftz.f32 %f13792, %f13768, %f1961, %f13791; fma.rn.ftz.f32 %f13793, %f13782, %f1959, %f13792; mul.ftz.f32 %f13794, %f13776, %f1960; fma.rn.ftz.f32 %f13795, %f13769, %f1961, %f13794; fma.rn.ftz.f32 %f13796, %f13783, %f1959, %f13795; mul.ftz.f32 %f13797, %f13793, %f13793; fma.rn.ftz.f32 %f13798, %f13790, %f13790, %f13797; fma.rn.ftz.f32 %f13799, %f13796, %f13796, %f13798; sqrt.approx.ftz.f32 %f13800, %f13799; div.approx.ftz.f32 %f1961, %f13790, %f13800; div.approx.ftz.f32 %f1960, %f13793, %f13800; div.approx.ftz.f32 %f29528, %f13796, %f13800; bra.uni $L__BB107_1393; $L__BB107_1391: cvta.to.local.u64 %rd2285, %rd312; ld.local.v4.f32 {%f13733, %f13734, %f13735, %f13736}, [%rd2285]; cvta.to.local.u64 %rd2286, %rd327; ld.local.v4.f32 {%f13740, %f13741, %f13742, %f13743}, [%rd2286]; cvta.to.local.u64 %rd2287, %rd326; ld.local.v4.f32 {%f13747, %f13748, %f13749, %f13750}, [%rd2287]; mul.ftz.f32 %f13754, %f1960, %f13740; fma.rn.ftz.f32 %f13755, %f1961, %f13733, %f13754; fma.rn.ftz.f32 %f13756, %f1959, %f13747, %f13755; mul.ftz.f32 %f13757, %f1960, %f13741; fma.rn.ftz.f32 %f13758, %f1961, %f13734, %f13757; fma.rn.ftz.f32 %f13759, %f1959, %f13748, %f13758; mul.ftz.f32 %f13760, %f1960, %f13742; fma.rn.ftz.f32 %f13761, %f1961, %f13735, %f13760; fma.rn.ftz.f32 %f13762, %f1959, %f13749, %f13761; mul.ftz.f32 %f13763, %f13759, %f13759; fma.rn.ftz.f32 %f13764, %f13756, %f13756, %f13763; fma.rn.ftz.f32 %f13765, %f13762, %f13762, %f13764; sqrt.approx.ftz.f32 %f13766, %f13765; div.approx.ftz.f32 %f1961, %f13756, %f13766; div.approx.ftz.f32 %f1960, %f13759, %f13766; div.approx.ftz.f32 %f29528, %f13762, %f13766; $L__BB107_1393: st.local.f32 [%rd315], %f1961; st.local.f32 [%rd337], %f1960; st.local.f32 [%rd338], %f29528; @%p1287 bra $L__BB107_1395; bra.uni $L__BB107_1394; $L__BB107_1395: ld.global.v4.f32 {%f13828, %f13829, %f13830, %f13831}, [%rd364]; ld.global.v4.f32 {%f13835, %f13836, %f13837, %f13838}, [%rd364+16]; ld.global.v4.f32 {%f13842, %f13843, %f13844, %f13845}, [%rd364+32]; mul.ftz.f32 %f13849, %f13829, %f29533; fma.rn.ftz.f32 %f13850, %f29531, %f13828, %f13849; fma.rn.ftz.f32 %f2252, %f13830, %f29532, %f13850; mul.ftz.f32 %f13851, %f13836, %f29533; fma.rn.ftz.f32 %f13852, %f29531, %f13835, %f13851; fma.rn.ftz.f32 %f2253, %f13837, %f29532, %f13852; mul.ftz.f32 %f13853, %f13843, %f29533; fma.rn.ftz.f32 %f13854, %f29531, %f13842, %f13853; fma.rn.ftz.f32 %f29532, %f13844, %f29532, %f13854; st.local.f32 [%rd318], %f2252; st.local.f32 [%rd319], %f2253; st.local.f32 [%rd320], %f29532; mov.f32 %f29531, %f2252; mov.f32 %f29533, %f2253; mov.f32 %f29536, %f29525; mov.f32 %f29537, %f29526; bra.uni $L__BB107_1396; $L__BB107_1394: cvta.to.local.u64 %rd2288, %rd299; ld.local.v4.f32 {%f13801, %f13802, %f13803, %f13804}, [%rd2288]; mul.ftz.f32 %f13808, %f29533, %f13802; fma.rn.ftz.f32 %f13809, %f29531, %f13801, %f13808; fma.rn.ftz.f32 %f2249, %f29532, %f13803, %f13809; cvta.to.local.u64 %rd2289, %rd311; ld.local.v4.f32 {%f13810, %f13811, %f13812, %f13813}, [%rd2289]; mul.ftz.f32 %f13817, %f29533, %f13811; fma.rn.ftz.f32 %f13818, %f29531, %f13810, %f13817; fma.rn.ftz.f32 %f2250, %f29532, %f13812, %f13818; cvta.to.local.u64 %rd2290, %rd310; ld.local.v4.f32 {%f13819, %f13820, %f13821, %f13822}, [%rd2290]; mul.ftz.f32 %f13826, %f29533, %f13820; fma.rn.ftz.f32 %f13827, %f29531, %f13819, %f13826; fma.rn.ftz.f32 %f29532, %f29532, %f13821, %f13827; st.local.f32 [%rd318], %f2249; st.local.f32 [%rd319], %f2250; st.local.f32 [%rd320], %f29532; mov.f32 %f29531, %f2249; mov.f32 %f29533, %f2250; mov.f32 %f29536, %f29525; mov.f32 %f29537, %f29526; $L__BB107_1396: st.local.v2.f32 [%rd2173], {%f29536, %f29537}; st.local.f32 [%rd314], %f29538; selp.f32 %f29575, %f1961, %f2026, %p1307; selp.f32 %f29576, %f1960, %f2027, %p1307; selp.b64 %rd2293, %rd317, %rd297, %p1307; cvta.to.local.u64 %rd2294, %rd2293; ld.local.f32 %f29577, [%rd2294]; cvta.to.local.u64 %rd2295, %rd302; st.local.v2.f32 [%rd2295], {%f29575, %f29576}; st.local.f32 [%rd307], %f29577; mul.ftz.f32 %f13855, %f29533, %f29577; mul.ftz.f32 %f13856, %f29532, %f29576; mul.ftz.f32 %f13857, %f29532, %f29575; mul.ftz.f32 %f13858, %f29531, %f29577; mul.ftz.f32 %f13859, %f29531, %f29576; mul.ftz.f32 %f13860, %f29533, %f29575; sub.ftz.f32 %f13861, %f13859, %f13860; sub.ftz.f32 %f13862, %f13857, %f13858; sub.ftz.f32 %f13863, %f13855, %f13856; cvta.to.local.u64 %rd2296, %rd309; st.local.v2.f32 [%rd2296], {%f13863, %f13862}; st.local.f32 [%rd321], %f13861; ld.global.u32 %r2823, [%rd368+-4]; st.local.u32 [%rd305], %r2823; mov.f32 %f1959, %f1956; mov.f32 %f1960, %f1957; mov.f32 %f1961, %f1958; bra.uni $L__BB107_1447; $L__BB107_1419: { // callseq 336, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 336 ld.local.u32 %r5722, [%rd290]; bra.uni $L__BB107_1433; $L__BB107_1399: ld.global.u32 %r656, [%rd364+148]; ld.global.u32 %r657, [%rd364+152]; shl.b32 %r2827, %r657, 1; cvt.rn.f32.s32 %f13864, %r2827; mul.ftz.f32 %f2271, %f1950, %f13864; cvt.rzi.ftz.s32.f32 %r2828, %f2271; add.s32 %r2829, %r2827, -1; min.s32 %r658, %r2828, %r2829; ld.global.u32 %r5718, [%rd364+164]; mul.wide.u32 %rd2301, %r5718, 16; add.s64 %rd2302, %rd331, %rd2301; ld.global.u16 %rs626, [%rd2302+12]; ld.global.u32 %r5717, [%rd2302+8]; ld.global.u64 %rd4791, [%rd2302]; setp.eq.s64 %p1320, %rd4791, 11; @%p1320 bra $L__BB107_1404; $L__BB107_1400: setp.eq.s64 %p1321, %rd4791, 0; @%p1321 bra $L__BB107_1402; add.s32 %r5718, %r5718, 2; bra.uni $L__BB107_1403; $L__BB107_1402: setp.eq.s16 %p1322, %rs626, 0; mov.u32 %r5721, -1; mov.u32 %r5718, %r5717; @%p1322 bra $L__BB107_1405; $L__BB107_1403: mul.wide.u32 %rd2303, %r5718, 16; add.s64 %rd2304, %rd331, %rd2303; ld.global.u16 %rs626, [%rd2304+12]; ld.global.u32 %r5717, [%rd2304+8]; ld.global.u64 %rd4791, [%rd2304]; setp.ne.s64 %p1323, %rd4791, 11; @%p1323 bra $L__BB107_1400; $L__BB107_1404: setp.eq.s16 %p1324, %rs626, 0; selp.b32 %r5721, -1, %r5717, %p1324; $L__BB107_1405: setp.eq.s32 %p1325, %r658, %r657; @%p1325 bra $L__BB107_1407; bra.uni $L__BB107_1406; $L__BB107_1407: add.s64 %rd4792, %rd330, %rd2256; bra.uni $L__BB107_1408; $L__BB107_1421: sub.ftz.f32 %f13969, %f2314, %f2315; sub.ftz.f32 %f13970, %f2316, %f2317; mul.ftz.f32 %f13971, %f13970, %f13970; fma.rn.ftz.f32 %f13972, %f13969, %f13969, %f13971; sub.ftz.f32 %f13973, %f2318, %f2319; fma.rn.ftz.f32 %f13974, %f13973, %f13973, %f13972; sqrt.approx.ftz.f32 %f13975, %f13974; div.approx.ftz.f32 %f29547, %f13973, %f13975; div.approx.ftz.f32 %f29546, %f13970, %f13975; div.approx.ftz.f32 %f29545, %f13969, %f13975; $L__BB107_1423: ld.global.u32 %r671, [%rd367]; st.local.u32 [%rd305], %r671; fma.rn.ftz.f32 %f13983, %f1967, %f2311, %f2305; fma.rn.ftz.f32 %f13984, %f1967, %f2312, %f2307; fma.rn.ftz.f32 %f13985, %f1967, %f2313, %f2309; fma.rn.ftz.f32 %f29560, %f1968, %f2306, %f13983; fma.rn.ftz.f32 %f29561, %f1968, %f2308, %f13984; fma.rn.ftz.f32 %f29562, %f1968, %f2310, %f13985; and.b32 %r2845, %r5722, 4; setp.ne.s32 %p1333, %r2845, 0; @%p1333 bra $L__BB107_1428; @%p1287 bra $L__BB107_1426; bra.uni $L__BB107_1425; $L__BB107_1426: ld.global.v4.f32 {%f29548, %f29549, %f29550, %f29551}, [%rd364]; ld.global.v4.f32 {%f29552, %f29553, %f29554, %f29555}, [%rd364+16]; ld.global.v4.f32 {%f29556, %f29557, %f29558, %f29559}, [%rd364+32]; bra.uni $L__BB107_1427; $L__BB107_1406: setp.gt.s32 %p1326, %r658, %r657; selp.b32 %r2831, -1, 0, %p1326; add.s32 %r2832, %r658, %r2831; mad.lo.s32 %r2833, %r2832, %r656, %r5730; add.s32 %r2834, %r2833, %r5721; mul.wide.s32 %rd2305, %r2834, 16; add.s64 %rd4792, %rd341, %rd2305; $L__BB107_1408: ld.global.v4.f32 {%f13865, %f13866, %f13867, %f13868}, [%rd4792]; add.s32 %r669, %r658, 1; setp.eq.s32 %p1327, %r669, %r657; @%p1327 bra $L__BB107_1410; bra.uni $L__BB107_1409; $L__BB107_1410: add.s64 %rd4793, %rd330, %rd2256; bra.uni $L__BB107_1411; $L__BB107_1409: setp.ge.s32 %p1328, %r658, %r657; selp.b32 %r2835, -1, 0, %p1328; add.s32 %r2836, %r669, %r2835; mad.lo.s32 %r2837, %r2836, %r656, %r5730; add.s32 %r2838, %r2837, %r5721; mul.wide.s32 %rd2307, %r2838, 16; add.s64 %rd4793, %rd341, %rd2307; $L__BB107_1411: cvt.rn.f32.s32 %f13869, %r658; sub.ftz.f32 %f13870, %f2271, %f13869; ld.global.v4.f32 {%f13871, %f13872, %f13873, %f13874}, [%rd4793]; mov.f32 %f13879, 0f3F800000; sub.ftz.f32 %f13880, %f13879, %f13870; mul.ftz.f32 %f13881, %f13870, %f13871; mul.ftz.f32 %f13882, %f13870, %f13872; mul.ftz.f32 %f13883, %f13870, %f13873; fma.rn.ftz.f32 %f29544, %f13880, %f13867, %f13883; fma.rn.ftz.f32 %f29543, %f13880, %f13866, %f13882; fma.rn.ftz.f32 %f29542, %f13880, %f13865, %f13881; $L__BB107_1413: and.b32 %r2839, %r5722, 4; setp.ne.s32 %p1329, %r2839, 0; @%p1329 bra $L__BB107_1417; @%p1287 bra $L__BB107_1416; bra.uni $L__BB107_1415; $L__BB107_1416: ld.global.v4.f32 {%f13922, %f13923, %f13924, %f13925}, [%rd364]; ld.global.v4.f32 {%f13930, %f13931, %f13932, %f13933}, [%rd364+16]; ld.global.v4.f32 {%f13938, %f13939, %f13940, %f13941}, [%rd364+32]; mul.ftz.f32 %f13946, %f29543, %f13923; fma.rn.ftz.f32 %f13947, %f29542, %f13922, %f13946; fma.rn.ftz.f32 %f13948, %f29544, %f13924, %f13947; add.ftz.f32 %f2294, %f13925, %f13948; mul.ftz.f32 %f13949, %f29543, %f13931; fma.rn.ftz.f32 %f13950, %f29542, %f13930, %f13949; fma.rn.ftz.f32 %f13951, %f29544, %f13932, %f13950; add.ftz.f32 %f2295, %f13933, %f13951; mul.ftz.f32 %f13952, %f29543, %f13939; fma.rn.ftz.f32 %f13953, %f29542, %f13938, %f13952; fma.rn.ftz.f32 %f13954, %f29544, %f13940, %f13953; add.ftz.f32 %f29544, %f13941, %f13954; mov.f32 %f29542, %f2294; mov.f32 %f29543, %f2295; bra.uni $L__BB107_1417; $L__BB107_1415: cvta.to.local.u64 %rd2311, %rd299; ld.local.v4.f32 {%f13889, %f13890, %f13891, %f13892}, [%rd2311]; mul.ftz.f32 %f13897, %f29543, %f13890; fma.rn.ftz.f32 %f13898, %f29542, %f13889, %f13897; fma.rn.ftz.f32 %f13899, %f29544, %f13891, %f13898; add.ftz.f32 %f2291, %f13892, %f13899; cvta.to.local.u64 %rd2312, %rd311; ld.local.v4.f32 {%f13900, %f13901, %f13902, %f13903}, [%rd2312]; mul.ftz.f32 %f13908, %f29543, %f13901; fma.rn.ftz.f32 %f13909, %f29542, %f13900, %f13908; fma.rn.ftz.f32 %f13910, %f29544, %f13902, %f13909; add.ftz.f32 %f2292, %f13903, %f13910; cvta.to.local.u64 %rd2313, %rd310; ld.local.v4.f32 {%f13911, %f13912, %f13913, %f13914}, [%rd2313]; mul.ftz.f32 %f13919, %f29543, %f13912; fma.rn.ftz.f32 %f13920, %f29542, %f13911, %f13919; fma.rn.ftz.f32 %f13921, %f29544, %f13913, %f13920; add.ftz.f32 %f29544, %f13914, %f13921; mov.f32 %f29542, %f2291; mov.f32 %f29543, %f2292; $L__BB107_1417: sub.ftz.f32 %f13955, %f2268, %f29542; sub.ftz.f32 %f13956, %f2269, %f29543; mul.ftz.f32 %f13957, %f13956, %f13956; fma.rn.ftz.f32 %f13958, %f13955, %f13955, %f13957; sub.ftz.f32 %f13959, %f2270, %f29544; fma.rn.ftz.f32 %f13960, %f13959, %f13959, %f13958; sqrt.approx.ftz.f32 %f13961, %f13960; div.approx.ftz.f32 %f29575, %f13955, %f13961; div.approx.ftz.f32 %f29576, %f13956, %f13961; div.approx.ftz.f32 %f29577, %f13959, %f13961; st.local.f32 [%rd307], %f29577; st.local.f32 [%rd315], %f29575; st.local.v4.f32 [%rd337], {%f29576, %f29577, %f29575, %f29576}; mov.u32 %r2840, 0; st.local.u32 [%rd318], %r2840; st.local.u32 [%rd319], %r2840; st.local.u32 [%rd320], %r2840; cvta.to.local.u64 %rd2315, %rd309; mov.f32 %f13962, 0f00000000; st.local.v2.f32 [%rd2315], {%f13962, %f13962}; st.local.u32 [%rd321], %r2840; mov.f32 %f1959, %f1956; mov.f32 %f1960, %f1957; mov.f32 %f1961, %f1958; bra.uni $L__BB107_1447; $L__BB107_1425: cvta.to.local.u64 %rd2323, %rd299; ld.local.v4.f32 {%f29548, %f29549, %f29550, %f29551}, [%rd2323]; cvta.to.local.u64 %rd2324, %rd311; ld.local.v4.f32 {%f29552, %f29553, %f29554, %f29555}, [%rd2324]; cvta.to.local.u64 %rd2325, %rd310; ld.local.v4.f32 {%f29556, %f29557, %f29558, %f29559}, [%rd2325]; $L__BB107_1427: mul.ftz.f32 %f14010, %f29561, %f29549; fma.rn.ftz.f32 %f14011, %f29560, %f29548, %f14010; fma.rn.ftz.f32 %f14012, %f29562, %f29550, %f14011; mul.ftz.f32 %f14013, %f29561, %f29553; fma.rn.ftz.f32 %f14014, %f29560, %f29552, %f14013; fma.rn.ftz.f32 %f14015, %f29562, %f29554, %f14014; mul.ftz.f32 %f14016, %f29561, %f29557; fma.rn.ftz.f32 %f14017, %f29560, %f29556, %f14016; fma.rn.ftz.f32 %f14018, %f29562, %f29558, %f14017; add.ftz.f32 %f29560, %f29551, %f14012; add.ftz.f32 %f29561, %f29555, %f14015; add.ftz.f32 %f29562, %f29559, %f14018; $L__BB107_1428: st.local.f32 [%rd307], %f29547; st.local.v4.f32 [%rd2173], {%f29560, %f29561, %f29562, %f29545}; st.local.v4.f32 [%rd337], {%f29546, %f29547, %f29545, %f29546}; setp.gt.s32 %p1335, %r671, -1; @%p1335 bra $L__BB107_1432; ld.global.v4.u32 {%r2846, %r2847, %r2848, %r2849}, [%rd366]; mul.wide.u32 %rd2329, %r2846, 12; add.s64 %rd2330, %rd352, %rd2329; mul.wide.u32 %rd2331, %r2847, 12; add.s64 %rd2332, %rd352, %rd2331; mul.wide.u32 %rd2333, %r2848, 12; add.s64 %rd2334, %rd352, %rd2333; mov.f32 %f14019, 0f3F800000; sub.ftz.f32 %f14020, %f14019, %f1967; sub.ftz.f32 %f14021, %f14020, %f1968; ld.global.f32 %f14022, [%rd2330]; ld.global.f32 %f14023, [%rd2330+4]; ld.global.f32 %f14024, [%rd2330+8]; ld.global.f32 %f14025, [%rd2332]; mul.ftz.f32 %f14026, %f1967, %f14025; ld.global.f32 %f14027, [%rd2332+4]; mul.ftz.f32 %f14028, %f1967, %f14027; ld.global.f32 %f14029, [%rd2332+8]; mul.ftz.f32 %f14030, %f1967, %f14029; fma.rn.ftz.f32 %f14031, %f14021, %f14022, %f14026; fma.rn.ftz.f32 %f14032, %f14021, %f14023, %f14028; fma.rn.ftz.f32 %f14033, %f14021, %f14024, %f14030; ld.global.f32 %f14034, [%rd2334]; ld.global.f32 %f14035, [%rd2334+4]; ld.global.f32 %f14036, [%rd2334+8]; fma.rn.ftz.f32 %f29563, %f1968, %f14034, %f14031; fma.rn.ftz.f32 %f29564, %f1968, %f14035, %f14032; fma.rn.ftz.f32 %f29565, %f1968, %f14036, %f14033; mul.ftz.f32 %f14037, %f29564, %f29564; fma.rn.ftz.f32 %f14038, %f29563, %f29563, %f14037; fma.rn.ftz.f32 %f14039, %f29565, %f29565, %f14038; sqrt.approx.ftz.f32 %f2380, %f14039; setp.neu.ftz.f32 %p1336, %f2380, 0f00000000; @%p1336 bra $L__BB107_1430; bra.uni $L__BB107_1431; $L__BB107_1430: rcp.approx.ftz.f32 %f14040, %f2380; mul.ftz.f32 %f29565, %f29565, %f14040; mul.ftz.f32 %f29564, %f29564, %f14040; mul.ftz.f32 %f29563, %f29563, %f14040; $L__BB107_1431: setp.eq.ftz.f32 %p1337, %f29563, 0f00000000; setp.eq.ftz.f32 %p1338, %f29564, 0f00000000; and.pred %p1339, %p1337, %p1338; setp.eq.ftz.f32 %p1340, %f29565, 0f00000000; and.pred %p1341, %p1340, %p1339; selp.f32 %f14041, %f29547, %f29565, %p1341; selp.f32 %f14042, %f29546, %f29564, %p1341; selp.f32 %f14043, %f29545, %f29563, %p1341; st.local.f32 [%rd315], %f14043; st.local.f32 [%rd337], %f14042; st.local.f32 [%rd338], %f14041; $L__BB107_1432: st.local.f32 [%rd318], %f2311; st.local.f32 [%rd319], %f2312; st.local.f32 [%rd320], %f2313; cvta.to.local.u64 %rd2335, %rd309; st.local.v2.f32 [%rd2335], {%f2306, %f2308}; st.local.f32 [%rd321], %f2310; $L__BB107_1433: and.b32 %r2853, %r5722, 4; setp.eq.s32 %p1342, %r2853, 0; @%p1342 bra $L__BB107_1435; bra.uni $L__BB107_1434; $L__BB107_1435: and.b32 %r673, %r5722, 2; setp.eq.s32 %p1343, %r673, 0; @%p1343 bra $L__BB107_1437; bra.uni $L__BB107_1436; $L__BB107_1437: ld.local.u32 %r2854, [%rd339]; mul.wide.s32 %rd2341, %r2854, 208; add.s64 %rd2342, %rd300, %rd2341; ld.global.v4.f32 {%f14085, %f14086, %f14087, %f14088}, [%rd2342+48]; ld.global.v4.f32 {%f14092, %f14093, %f14094, %f14095}, [%rd2342+64]; ld.global.v4.f32 {%f14099, %f14100, %f14101, %f14102}, [%rd2342+80]; ld.local.v2.f32 {%f14106, %f14107}, [%rd337]; ld.local.f32 %f14110, [%rd315]; mul.ftz.f32 %f14111, %f14092, %f14106; fma.rn.ftz.f32 %f14112, %f14085, %f14110, %f14111; fma.rn.ftz.f32 %f14113, %f14099, %f14107, %f14112; mul.ftz.f32 %f14114, %f14093, %f14106; fma.rn.ftz.f32 %f14115, %f14086, %f14110, %f14114; fma.rn.ftz.f32 %f14116, %f14100, %f14107, %f14115; mul.ftz.f32 %f14117, %f14094, %f14106; fma.rn.ftz.f32 %f14118, %f14087, %f14110, %f14117; fma.rn.ftz.f32 %f14119, %f14101, %f14107, %f14118; mul.ftz.f32 %f14120, %f14116, %f14116; fma.rn.ftz.f32 %f14121, %f14113, %f14113, %f14120; fma.rn.ftz.f32 %f14122, %f14119, %f14119, %f14121; sqrt.approx.ftz.f32 %f14123, %f14122; div.approx.ftz.f32 %f29568, %f14113, %f14123; div.approx.ftz.f32 %f29567, %f14116, %f14123; div.approx.ftz.f32 %f29566, %f14119, %f14123; bra.uni $L__BB107_1438; $L__BB107_1434: cvta.to.local.u64 %rd2336, %rd302; ld.local.v2.f32 {%f29575, %f29576}, [%rd2336]; ld.local.f32 %f29577, [%rd307]; bra.uni $L__BB107_1447; $L__BB107_1436: ld.local.v2.f32 {%f14046, %f14047}, [%rd337]; cvta.to.local.u64 %rd2338, %rd312; ld.local.v4.f32 {%f14050, %f14051, %f14052, %f14053}, [%rd2338]; cvta.to.local.u64 %rd2339, %rd327; ld.local.v4.f32 {%f14057, %f14058, %f14059, %f14060}, [%rd2339]; cvta.to.local.u64 %rd2340, %rd326; ld.local.v4.f32 {%f14064, %f14065, %f14066, %f14067}, [%rd2340]; ld.local.f32 %f14071, [%rd315]; mul.ftz.f32 %f14072, %f14046, %f14057; fma.rn.ftz.f32 %f14073, %f14071, %f14050, %f14072; fma.rn.ftz.f32 %f14074, %f14047, %f14064, %f14073; mul.ftz.f32 %f14075, %f14046, %f14058; fma.rn.ftz.f32 %f14076, %f14071, %f14051, %f14075; fma.rn.ftz.f32 %f14077, %f14047, %f14065, %f14076; mul.ftz.f32 %f14078, %f14046, %f14059; fma.rn.ftz.f32 %f14079, %f14071, %f14052, %f14078; fma.rn.ftz.f32 %f14080, %f14047, %f14066, %f14079; mul.ftz.f32 %f14081, %f14077, %f14077; fma.rn.ftz.f32 %f14082, %f14074, %f14074, %f14081; fma.rn.ftz.f32 %f14083, %f14080, %f14080, %f14082; sqrt.approx.ftz.f32 %f14084, %f14083; div.approx.ftz.f32 %f29568, %f14074, %f14084; div.approx.ftz.f32 %f29567, %f14077, %f14084; div.approx.ftz.f32 %f29566, %f14080, %f14084; $L__BB107_1438: st.local.f32 [%rd315], %f29568; st.local.f32 [%rd337], %f29567; st.local.f32 [%rd338], %f29566; @%p1343 bra $L__BB107_1440; bra.uni $L__BB107_1439; $L__BB107_1440: ld.local.u32 %r2855, [%rd339]; mul.wide.s32 %rd2348, %r2855, 208; add.s64 %rd2349, %rd300, %rd2348; ld.global.v4.f32 {%f14163, %f14164, %f14165, %f14166}, [%rd2349+48]; ld.global.v4.f32 {%f14170, %f14171, %f14172, %f14173}, [%rd2349+64]; ld.global.v4.f32 {%f14177, %f14178, %f14179, %f14180}, [%rd2349+80]; cvta.to.local.u64 %rd2350, %rd302; ld.local.v2.f32 {%f14184, %f14185}, [%rd2350]; mul.ftz.f32 %f14188, %f14170, %f14185; fma.rn.ftz.f32 %f14189, %f14163, %f14184, %f14188; ld.local.f32 %f14190, [%rd307]; fma.rn.ftz.f32 %f14191, %f14177, %f14190, %f14189; mul.ftz.f32 %f14192, %f14171, %f14185; fma.rn.ftz.f32 %f14193, %f14164, %f14184, %f14192; fma.rn.ftz.f32 %f14194, %f14178, %f14190, %f14193; mul.ftz.f32 %f14195, %f14172, %f14185; fma.rn.ftz.f32 %f14196, %f14165, %f14184, %f14195; fma.rn.ftz.f32 %f14197, %f14179, %f14190, %f14196; mul.ftz.f32 %f14198, %f14194, %f14194; fma.rn.ftz.f32 %f14199, %f14191, %f14191, %f14198; fma.rn.ftz.f32 %f14200, %f14197, %f14197, %f14199; sqrt.approx.ftz.f32 %f14201, %f14200; div.approx.ftz.f32 %f29575, %f14191, %f14201; div.approx.ftz.f32 %f29576, %f14194, %f14201; div.approx.ftz.f32 %f29577, %f14197, %f14201; bra.uni $L__BB107_1441; $L__BB107_1439: cvta.to.local.u64 %rd2344, %rd302; ld.local.v2.f32 {%f14124, %f14125}, [%rd2344]; cvta.to.local.u64 %rd2345, %rd312; ld.local.v4.f32 {%f14128, %f14129, %f14130, %f14131}, [%rd2345]; cvta.to.local.u64 %rd2346, %rd327; ld.local.v4.f32 {%f14135, %f14136, %f14137, %f14138}, [%rd2346]; cvta.to.local.u64 %rd2347, %rd326; ld.local.v4.f32 {%f14142, %f14143, %f14144, %f14145}, [%rd2347]; mul.ftz.f32 %f14149, %f14125, %f14135; fma.rn.ftz.f32 %f14150, %f14124, %f14128, %f14149; ld.local.f32 %f14151, [%rd307]; fma.rn.ftz.f32 %f14152, %f14151, %f14142, %f14150; mul.ftz.f32 %f14153, %f14125, %f14136; fma.rn.ftz.f32 %f14154, %f14124, %f14129, %f14153; fma.rn.ftz.f32 %f14155, %f14151, %f14143, %f14154; mul.ftz.f32 %f14156, %f14125, %f14137; fma.rn.ftz.f32 %f14157, %f14124, %f14130, %f14156; fma.rn.ftz.f32 %f14158, %f14151, %f14144, %f14157; mul.ftz.f32 %f14159, %f14155, %f14155; fma.rn.ftz.f32 %f14160, %f14152, %f14152, %f14159; fma.rn.ftz.f32 %f14161, %f14158, %f14158, %f14160; sqrt.approx.ftz.f32 %f14162, %f14161; div.approx.ftz.f32 %f29575, %f14152, %f14162; div.approx.ftz.f32 %f29576, %f14155, %f14162; div.approx.ftz.f32 %f29577, %f14158, %f14162; $L__BB107_1441: cvta.to.local.u64 %rd2351, %rd302; st.local.v2.f32 [%rd2351], {%f29575, %f29576}; st.local.f32 [%rd307], %f29577; @%p1343 bra $L__BB107_1443; bra.uni $L__BB107_1442; $L__BB107_1443: ld.local.u32 %r2856, [%rd339]; mul.wide.s32 %rd2356, %r2856, 208; add.s64 %rd2357, %rd300, %rd2356; ld.global.v4.f32 {%f14234, %f14235, %f14236, %f14237}, [%rd2357]; ld.global.v4.f32 {%f14241, %f14242, %f14243, %f14244}, [%rd2357+16]; ld.global.v4.f32 {%f14248, %f14249, %f14250, %f14251}, [%rd2357+32]; ld.local.v2.f32 {%f14255, %f14256}, [%rd319]; ld.local.f32 %f14259, [%rd318]; mul.ftz.f32 %f14260, %f14235, %f14255; fma.rn.ftz.f32 %f14261, %f14259, %f14234, %f14260; fma.rn.ftz.f32 %f29574, %f14236, %f14256, %f14261; mul.ftz.f32 %f14262, %f14242, %f14255; fma.rn.ftz.f32 %f14263, %f14259, %f14241, %f14262; fma.rn.ftz.f32 %f29573, %f14243, %f14256, %f14263; mul.ftz.f32 %f14264, %f14249, %f14255; fma.rn.ftz.f32 %f14265, %f14259, %f14248, %f14264; fma.rn.ftz.f32 %f29572, %f14250, %f14256, %f14265; bra.uni $L__BB107_1444; $L__BB107_1442: ld.local.v2.f32 {%f14202, %f14203}, [%rd319]; cvta.to.local.u64 %rd2353, %rd299; ld.local.v4.f32 {%f14206, %f14207, %f14208, %f14209}, [%rd2353]; ld.local.f32 %f14213, [%rd318]; mul.ftz.f32 %f14214, %f14202, %f14207; fma.rn.ftz.f32 %f14215, %f14213, %f14206, %f14214; fma.rn.ftz.f32 %f29574, %f14203, %f14208, %f14215; cvta.to.local.u64 %rd2354, %rd311; ld.local.v4.f32 {%f14216, %f14217, %f14218, %f14219}, [%rd2354]; mul.ftz.f32 %f14223, %f14202, %f14217; fma.rn.ftz.f32 %f14224, %f14213, %f14216, %f14223; fma.rn.ftz.f32 %f29573, %f14203, %f14218, %f14224; cvta.to.local.u64 %rd2355, %rd310; ld.local.v4.f32 {%f14225, %f14226, %f14227, %f14228}, [%rd2355]; mul.ftz.f32 %f14232, %f14202, %f14226; fma.rn.ftz.f32 %f14233, %f14213, %f14225, %f14232; fma.rn.ftz.f32 %f29572, %f14203, %f14227, %f14233; $L__BB107_1444: st.local.f32 [%rd318], %f29574; st.local.f32 [%rd319], %f29573; st.local.f32 [%rd320], %f29572; @%p1343 bra $L__BB107_1446; bra.uni $L__BB107_1445; $L__BB107_1446: ld.local.u32 %r2857, [%rd339]; mul.wide.s32 %rd2363, %r2857, 208; add.s64 %rd2364, %rd300, %rd2363; ld.global.v4.f32 {%f14303, %f14304, %f14305, %f14306}, [%rd2364]; ld.global.v4.f32 {%f14310, %f14311, %f14312, %f14313}, [%rd2364+16]; ld.global.v4.f32 {%f14317, %f14318, %f14319, %f14320}, [%rd2364+32]; cvta.to.local.u64 %rd2365, %rd309; ld.local.v4.f32 {%f14324, %f14325, %f14326, %f14327}, [%rd2365]; mul.ftz.f32 %f14331, %f14304, %f14325; fma.rn.ftz.f32 %f14332, %f14324, %f14303, %f14331; mul.ftz.f32 %f14333, %f14311, %f14325; fma.rn.ftz.f32 %f14334, %f14324, %f14310, %f14333; mul.ftz.f32 %f14335, %f14318, %f14325; fma.rn.ftz.f32 %f14336, %f14324, %f14317, %f14335; fma.rn.ftz.f32 %f14337, %f14319, %f14326, %f14336; fma.rn.ftz.f32 %f14338, %f14312, %f14326, %f14334; fma.rn.ftz.f32 %f14339, %f14305, %f14326, %f14332; st.local.v2.f32 [%rd2365], {%f14339, %f14338}; st.local.f32 [%rd321], %f14337; bra.uni $L__BB107_1447; $L__BB107_1445: cvta.to.local.u64 %rd2359, %rd309; ld.local.v4.f32 {%f14266, %f14267, %f14268, %f14269}, [%rd2359]; cvta.to.local.u64 %rd2360, %rd299; ld.local.v4.f32 {%f14273, %f14274, %f14275, %f14276}, [%rd2360]; mul.ftz.f32 %f14280, %f14267, %f14274; fma.rn.ftz.f32 %f14281, %f14266, %f14273, %f14280; cvta.to.local.u64 %rd2361, %rd311; ld.local.v4.f32 {%f14282, %f14283, %f14284, %f14285}, [%rd2361]; mul.ftz.f32 %f14289, %f14267, %f14283; fma.rn.ftz.f32 %f14290, %f14266, %f14282, %f14289; cvta.to.local.u64 %rd2362, %rd310; ld.local.v4.f32 {%f14291, %f14292, %f14293, %f14294}, [%rd2362]; mul.ftz.f32 %f14298, %f14267, %f14292; fma.rn.ftz.f32 %f14299, %f14266, %f14291, %f14298; fma.rn.ftz.f32 %f14300, %f14268, %f14293, %f14299; fma.rn.ftz.f32 %f14301, %f14268, %f14284, %f14290; fma.rn.ftz.f32 %f14302, %f14268, %f14275, %f14281; st.local.v2.f32 [%rd2359], {%f14302, %f14301}; st.local.f32 [%rd321], %f14300; $L__BB107_1447: ld.local.u32 %r674, [%rd305]; and.b32 %r5723, %r674, 4194303; shl.b32 %r2858, %r674, 5; cvt.u64.u32 %rd2366, %r2858; and.b64 %rd2367, %rd2366, 134217696; add.s64 %rd2368, %rd303, %rd2367; ld.global.u32 %r676, [%rd2368+16]; st.local.u32 [%rd293], %r676; ld.local.f32 %f14340, [%rd294]; ld.local.v2.f32 {%f14341, %f14342}, [%rd296]; mul.ftz.f32 %f14345, %f29576, %f14341; fma.rn.ftz.f32 %f14346, %f29575, %f14340, %f14345; fma.rn.ftz.f32 %f14347, %f29577, %f14342, %f14346; setp.lt.ftz.f32 %p1347, %f14347, 0f00000000; @%p1347 bra $L__BB107_1449; bra.uni $L__BB107_1448; $L__BB107_1449: or.b32 %r2859, %r676, 1; st.local.u32 [%rd293], %r2859; neg.ftz.f32 %f29577, %f29577; st.local.f32 [%rd307], %f29577; ld.local.f32 %f14348, [%rd315]; neg.ftz.f32 %f14349, %f14348; ld.local.v2.f32 {%f14350, %f14351}, [%rd337]; st.local.f32 [%rd315], %f14349; neg.ftz.f32 %f14354, %f14351; neg.ftz.f32 %f14355, %f14350; neg.ftz.f32 %f29576, %f29576; neg.ftz.f32 %f29575, %f29575; st.local.v4.f32 [%rd337], {%f14355, %f14354, %f29575, %f29576}; ld.local.f32 %f14356, [%rd318]; neg.ftz.f32 %f29595, %f14356; ld.local.v2.f32 {%f14357, %f14358}, [%rd319]; neg.ftz.f32 %f29596, %f14357; neg.ftz.f32 %f14361, %f14358; st.local.f32 [%rd318], %f29595; st.local.f32 [%rd319], %f29596; st.local.f32 [%rd320], %f14361; cvta.to.local.u64 %rd2372, %rd309; ld.local.v4.f32 {%f14362, %f14363, %f14364, %f14365}, [%rd2372]; neg.ftz.f32 %f14369, %f14364; neg.ftz.f32 %f14370, %f14363; neg.ftz.f32 %f14371, %f14362; st.local.v2.f32 [%rd2372], {%f14371, %f14370}; st.local.f32 [%rd321], %f14369; bra.uni $L__BB107_1450; $L__BB107_1448: ld.local.f32 %f29595, [%rd318]; ld.local.f32 %f29596, [%rd319]; $L__BB107_1450: ld.local.f32 %f14372, [%rd288]; fma.rn.ftz.f32 %f14373, %f14372, 0f00000000, %f1949; st.local.f32 [%rd322], %f14373; st.local.u32 [%rd324], %r5802; cvta.to.local.u64 %rd2373, %rd309; ld.local.v4.f32 {%f29593, %f29594, %f14376, %f14377}, [%rd2373]; setp.neu.ftz.f32 %p1348, %f29575, %f29576; setp.neu.ftz.f32 %p1349, %f29575, %f29577; or.pred %p1350, %p1348, %p1349; sub.ftz.f32 %f14378, %f29577, %f29576; neg.ftz.f32 %f14379, %f29576; neg.ftz.f32 %f14380, %f29577; selp.f32 %f14381, %f14380, %f29577, %p1350; add.ftz.f32 %f14382, %f29575, %f14381; selp.f32 %f14383, %f29576, %f14379, %p1350; sub.ftz.f32 %f14384, %f14383, %f29575; mul.ftz.f32 %f14385, %f14382, %f14382; fma.rn.ftz.f32 %f14386, %f14378, %f14378, %f14385; fma.rn.ftz.f32 %f14387, %f14384, %f14384, %f14386; sqrt.approx.ftz.f32 %f14388, %f14387; div.approx.ftz.f32 %f14389, %f14378, %f14388; div.approx.ftz.f32 %f14390, %f14382, %f14388; div.approx.ftz.f32 %f14391, %f14384, %f14388; mul.ftz.f32 %f14392, %f29576, %f14391; mul.ftz.f32 %f14393, %f29577, %f14390; sub.ftz.f32 %f14394, %f14392, %f14393; mul.ftz.f32 %f14395, %f29577, %f14389; mul.ftz.f32 %f14396, %f29575, %f14391; sub.ftz.f32 %f14397, %f14395, %f14396; mul.ftz.f32 %f14398, %f29575, %f14390; mul.ftz.f32 %f14399, %f29576, %f14389; sub.ftz.f32 %f14400, %f14398, %f14399; mul.ftz.f32 %f29589, %f14373, %f14389; mul.ftz.f32 %f29590, %f14373, %f14390; mul.ftz.f32 %f2455, %f14373, %f14391; mul.ftz.f32 %f29591, %f14373, %f14394; mul.ftz.f32 %f29592, %f14373, %f14397; mul.ftz.f32 %f2458, %f14373, %f14400; abs.ftz.f32 %f2459, %f29575; abs.ftz.f32 %f2460, %f29576; abs.ftz.f32 %f2461, %f29577; setp.geu.ftz.f32 %p1351, %f2461, %f2459; setp.geu.ftz.f32 %p1352, %f2461, %f2460; and.pred %p1353, %p1351, %p1352; @%p1353 bra $L__BB107_1452; ld.local.f32 %f2463, [%rd320]; setp.lt.ftz.f32 %p1354, %f2460, %f2461; setp.lt.ftz.f32 %p1355, %f2460, %f2459; or.pred %p1356, %p1355, %p1354; selp.f32 %f29589, %f29590, %f29589, %p1356; selp.f32 %f29591, %f29592, %f29591, %p1356; selp.f32 %f29593, %f29594, %f29593, %p1356; selp.f32 %f29595, %f29596, %f29595, %p1356; mov.f32 %f29590, %f2455; mov.f32 %f29592, %f2458; mov.f32 %f29594, %f14376; mov.f32 %f29596, %f2463; $L__BB107_1452: mul.ftz.f32 %f14401, %f29593, %f29596; mul.ftz.f32 %f14402, %f29594, %f29595; sub.ftz.f32 %f29597, %f14402, %f14401; setp.eq.ftz.f32 %p1357, %f29597, 0f00000000; @%p1357 bra $L__BB107_1454; rcp.approx.ftz.f32 %f29597, %f29597; $L__BB107_1454: mul.ftz.f32 %f14404, %f29590, %f29593; mul.ftz.f32 %f14405, %f29589, %f29594; sub.ftz.f32 %f14406, %f14405, %f14404; mul.ftz.f32 %f14407, %f14406, %f29597; st.local.f32 [%rd334], %f14407; mul.ftz.f32 %f14408, %f29589, %f29596; mul.ftz.f32 %f14409, %f29590, %f29595; sub.ftz.f32 %f14410, %f14409, %f14408; mul.ftz.f32 %f14411, %f29592, %f29593; mul.ftz.f32 %f14412, %f29591, %f29594; sub.ftz.f32 %f14413, %f14412, %f14411; mul.ftz.f32 %f14414, %f14410, %f29597; mul.ftz.f32 %f14415, %f14413, %f29597; st.local.v2.f32 [%rd2374], {%f14415, %f14414}; mul.ftz.f32 %f14416, %f29591, %f29596; mul.ftz.f32 %f14417, %f29592, %f29595; sub.ftz.f32 %f14418, %f14417, %f14416; mul.ftz.f32 %f14419, %f14418, %f29597; st.local.f32 [%rd336], %f14419; setp.gt.s32 %p1358, %r674, -1; @%p1358 bra $L__BB107_3392; mov.u32 %r5802, 0; st.local.u32 [%rd342], %r5802; mov.u32 %r2863, 4; st.local.u32 [%rd343], %r2863; add.s32 %r677, %r5732, 1; cvt.s64.s32 %rd385, %r5732; bra.uni $L__BB107_1456; $L__BB107_1626: { // callseq 337, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 337 add.s32 %r5723, %r678, 1; $L__BB107_1456: mov.u32 %r678, %r5723; mul.wide.s32 %rd2376, %r678, 16; add.s64 %rd386, %rd344, %rd2376; ld.global.v4.u32 {%r2865, %r2866, %r2867, %r2868}, [%rd386]; mul.wide.u32 %rd2377, %r2866, 4; add.s64 %rd387, %rd2157, %rd2377; mul.wide.u32 %rd2378, %r2867, 4; add.s64 %rd388, %rd2157, %rd2378; setp.gt.s32 %p1359, %r2865, 1; @%p1359 bra $L__BB107_1464; setp.eq.s32 %p1487, %r2865, 1; mov.u32 %r5723, %r2866; @%p1487 bra $L__BB107_1456; bra.uni $L__BB107_1458; $L__BB107_1464: setp.gt.s32 %p1360, %r2865, 49; @%p1360 bra $L__BB107_1545; setp.gt.s32 %p1424, %r2865, 25; @%p1424 bra $L__BB107_1506; setp.gt.s32 %p1456, %r2865, 13; @%p1456 bra $L__BB107_1487; setp.gt.s32 %p1472, %r2865, 7; @%p1472 bra $L__BB107_1477; setp.gt.s32 %p1480, %r2865, 4; @%p1480 bra $L__BB107_1473; setp.eq.s32 %p1484, %r2865, 2; @%p1484 bra $L__BB107_1693; setp.eq.s32 %p1485, %r2865, 3; @%p1485 bra $L__BB107_1692; setp.eq.s32 %p1486, %r2865, 4; @%p1486 bra $L__BB107_1472; bra.uni $L__BB107_1458; $L__BB107_1472: { // callseq 416, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2863, %r2866, %r2867, %r2868}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 416 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1545: setp.gt.s32 %p1361, %r2865, 73; @%p1361 bra $L__BB107_1585; setp.gt.s32 %p1393, %r2865, 61; @%p1393 bra $L__BB107_1566; setp.gt.s32 %p1409, %r2865, 55; @%p1409 bra $L__BB107_1557; setp.gt.s32 %p1417, %r2865, 52; @%p1417 bra $L__BB107_1553; setp.eq.s32 %p1421, %r2865, 50; @%p1421 bra $L__BB107_1660; setp.eq.s32 %p1422, %r2865, 51; @%p1422 bra $L__BB107_1659; setp.eq.s32 %p1423, %r2865, 52; @%p1423 bra $L__BB107_1552; bra.uni $L__BB107_1458; $L__BB107_1552: add.s32 %r2905, %r678, 1; { // callseq 377, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2905; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 377 bra.uni $L__BB107_1456; $L__BB107_1506: setp.gt.s32 %p1425, %r2865, 37; @%p1425 bra $L__BB107_1526; setp.gt.s32 %p1441, %r2865, 31; @%p1441 bra $L__BB107_1517; setp.gt.s32 %p1449, %r2865, 28; @%p1449 bra $L__BB107_1513; setp.eq.s32 %p1453, %r2865, 26; @%p1453 bra $L__BB107_1676; setp.eq.s32 %p1454, %r2865, 27; @%p1454 bra $L__BB107_1675; setp.eq.s32 %p1455, %r2865, 28; @%p1455 bra $L__BB107_1512; bra.uni $L__BB107_1458; $L__BB107_1512: mov.u32 %r2920, 28; { // callseq 399, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2920; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 399 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1585: setp.gt.s32 %p1362, %r2865, 85; @%p1362 bra $L__BB107_1608; setp.gt.s32 %p1378, %r2865, 79; @%p1378 bra $L__BB107_1599; setp.gt.s32 %p1386, %r2865, 76; @%p1386 bra $L__BB107_1595; setp.eq.s32 %p1390, %r2865, 74; @%p1390 bra $L__BB107_1644; setp.eq.s32 %p1391, %r2865, 75; @%p1391 bra $L__BB107_1643; setp.eq.s32 %p1392, %r2865, 76; @%p1392 bra $L__BB107_1591; bra.uni $L__BB107_1458; $L__BB107_1591: add.s32 %r5723, %r678, 1; setp.eq.s32 %p1489, %r2868, 255; @%p1489 bra $L__BB107_1456; add.s64 %rd389, %rd387, 4; add.s64 %rd390, %rd387, 8; setp.eq.s32 %p1490, %r2867, 0; mul.wide.u32 %rd2405, %r2868, 4; add.s64 %rd391, %rd2157, %rd2405; @%p1490 bra $L__BB107_1642; setp.eq.s32 %p1491, %r2867, 1; @%p1491 bra $L__BB107_1641; bra.uni $L__BB107_1594; $L__BB107_1641: ld.local.f32 %f14427, [%rd389]; st.local.f32 [%rd391], %f14427; bra.uni $L__BB107_1456; $L__BB107_1487: setp.gt.s32 %p1457, %r2865, 19; @%p1457 bra $L__BB107_1497; setp.gt.s32 %p1465, %r2865, 16; @%p1465 bra $L__BB107_1493; setp.eq.s32 %p1469, %r2865, 14; @%p1469 bra $L__BB107_1684; setp.eq.s32 %p1470, %r2865, 15; @%p1470 bra $L__BB107_1683; setp.eq.s32 %p1471, %r2865, 16; @%p1471 bra $L__BB107_1492; bra.uni $L__BB107_1458; $L__BB107_1492: mov.u32 %r2930, 16; { // callseq 410, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2930; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 410 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1566: setp.gt.s32 %p1394, %r2865, 67; @%p1394 bra $L__BB107_1576; setp.gt.s32 %p1402, %r2865, 64; @%p1402 bra $L__BB107_1572; setp.eq.s32 %p1406, %r2865, 62; @%p1406 bra $L__BB107_1652; setp.eq.s32 %p1407, %r2865, 63; @%p1407 bra $L__BB107_1651; setp.eq.s32 %p1408, %r2865, 64; @%p1408 bra $L__BB107_1571; bra.uni $L__BB107_1458; $L__BB107_1571: mov.u32 %r2891, 64; add.s32 %r2892, %r678, 1; { // callseq 365, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2891; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2892; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 365 bra.uni $L__BB107_1456; $L__BB107_1526: setp.gt.s32 %p1426, %r2865, 43; @%p1426 bra $L__BB107_1536; setp.gt.s32 %p1434, %r2865, 40; @%p1434 bra $L__BB107_1532; setp.eq.s32 %p1438, %r2865, 38; @%p1438 bra $L__BB107_1668; setp.eq.s32 %p1439, %r2865, 39; @%p1439 bra $L__BB107_1667; setp.eq.s32 %p1440, %r2865, 40; @%p1440 bra $L__BB107_1531; bra.uni $L__BB107_1458; $L__BB107_1531: add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1608: setp.gt.s32 %p1363, %r2865, 92; @%p1363 bra $L__BB107_1618; setp.gt.s32 %p1371, %r2865, 88; @%p1371 bra $L__BB107_1614; setp.eq.s32 %p1375, %r2865, 86; @%p1375 bra $L__BB107_1633; setp.eq.s32 %p1376, %r2865, 87; @%p1376 bra $L__BB107_1632; setp.eq.s32 %p1377, %r2865, 88; @%p1377 bra $L__BB107_1613; bra.uni $L__BB107_1458; $L__BB107_1613: mov.u32 %r2874, 88; ld.local.f32 %f14420, [%rd282]; ld.local.v2.f32 {%f14421, %f14422}, [%rd282+4]; { // callseq 343, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f14420; .param .b32 param1; st.param.f32 [param1+0], %f14421; .param .b32 param2; st.param.f32 [param2+0], %f14422; .param .b64 param3; st.param.b64 [param3+0], %rd2156; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r2874; st.param.b32 [param4+4], %r2866; st.param.b32 [param4+8], %r2867; st.param.b32 [param4+12], %r2868; call.uni _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 343 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1477: setp.gt.s32 %p1473, %r2865, 10; @%p1473 bra $L__BB107_1483; setp.eq.s32 %p1477, %r2865, 8; @%p1477 bra $L__BB107_1689; setp.eq.s32 %p1478, %r2865, 9; @%p1478 bra $L__BB107_1687; setp.eq.s32 %p1479, %r2865, 10; @%p1479 bra $L__BB107_1481; bra.uni $L__BB107_1458; $L__BB107_1481: ld.local.f32 %f14452, [%rd388]; setp.ltu.ftz.f32 %p1492, %f14452, 0f3F800000; add.s32 %r5723, %r678, 1; @%p1492 bra $L__BB107_1456; add.s32 %r2941, %r2866, %r678; add.s32 %r5723, %r2941, 1; bra.uni $L__BB107_1456; $L__BB107_1557: setp.gt.s32 %p1410, %r2865, 58; @%p1410 bra $L__BB107_1562; setp.eq.s32 %p1414, %r2865, 56; @%p1414 bra $L__BB107_1656; setp.eq.s32 %p1415, %r2865, 57; @%p1415 bra $L__BB107_1655; setp.eq.s32 %p1416, %r2865, 58; @%p1416 bra $L__BB107_1561; bra.uni $L__BB107_1458; $L__BB107_1561: mov.u32 %r2900, 58; { // callseq 371, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2900; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 371 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1517: setp.gt.s32 %p1442, %r2865, 34; @%p1442 bra $L__BB107_1522; setp.eq.s32 %p1446, %r2865, 32; @%p1446 bra $L__BB107_1672; setp.eq.s32 %p1447, %r2865, 33; @%p1447 bra $L__BB107_1671; setp.eq.s32 %p1448, %r2865, 34; @%p1448 bra $L__BB107_1521; bra.uni $L__BB107_1458; $L__BB107_1521: { // callseq 394, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 394 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1599: setp.gt.s32 %p1379, %r2865, 82; @%p1379 bra $L__BB107_1604; setp.eq.s32 %p1383, %r2865, 80; @%p1383 bra $L__BB107_1637; setp.eq.s32 %p1384, %r2865, 81; @%p1384 bra $L__BB107_1636; setp.eq.s32 %p1385, %r2865, 82; @%p1385 bra $L__BB107_1603; bra.uni $L__BB107_1458; $L__BB107_1603: mov.u32 %r2878, 82; { // callseq 349, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2878; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 349 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1497: setp.gt.s32 %p1458, %r2865, 22; @%p1458 bra $L__BB107_1502; setp.eq.s32 %p1462, %r2865, 20; @%p1462 bra $L__BB107_1680; setp.eq.s32 %p1463, %r2865, 21; @%p1463 bra $L__BB107_1679; setp.eq.s32 %p1464, %r2865, 22; @%p1464 bra $L__BB107_1501; bra.uni $L__BB107_1458; $L__BB107_1501: mov.u32 %r2927, 22; add.s32 %r2928, %r678, 1; { // callseq 405, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2927; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; .param .b32 param3; st.param.b32 [param3+0], %r2928; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 405 bra.uni $L__BB107_1456; $L__BB107_1576: setp.gt.s32 %p1395, %r2865, 70; @%p1395 bra $L__BB107_1581; setp.eq.s32 %p1399, %r2865, 68; @%p1399 bra $L__BB107_1648; setp.eq.s32 %p1400, %r2865, 69; @%p1400 bra $L__BB107_1647; setp.eq.s32 %p1401, %r2865, 70; @%p1401 bra $L__BB107_1580; bra.uni $L__BB107_1458; $L__BB107_1580: mov.u32 %r2885, 70; { // callseq 359, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2885; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 359 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1536: setp.gt.s32 %p1427, %r2865, 46; @%p1427 bra $L__BB107_1541; setp.eq.s32 %p1431, %r2865, 44; @%p1431 bra $L__BB107_1664; setp.eq.s32 %p1432, %r2865, 45; @%p1432 bra $L__BB107_1663; setp.eq.s32 %p1433, %r2865, 46; @%p1433 bra $L__BB107_1540; bra.uni $L__BB107_1458; $L__BB107_1540: { // callseq 383, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 383 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1618: setp.gt.s32 %p1364, %r2865, 95; @%p1364 bra $L__BB107_1623; setp.eq.s32 %p1368, %r2865, 93; @%p1368 bra $L__BB107_1628; setp.eq.s32 %p1369, %r2865, 94; @%p1369 bra $L__BB107_1627; setp.eq.s32 %p1370, %r2865, 95; @%p1370 bra $L__BB107_1622; bra.uni $L__BB107_1458; $L__BB107_1622: { // callseq 338, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 338 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1473: setp.eq.s32 %p1481, %r2865, 5; @%p1481 bra $L__BB107_1691; setp.eq.s32 %p1482, %r2865, 6; @%p1482 bra $L__BB107_1690; setp.eq.s32 %p1483, %r2865, 7; @%p1483 bra $L__BB107_1476; bra.uni $L__BB107_1458; $L__BB107_1476: mov.u32 %r2944, 7; { // callseq 415, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2944; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 415 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1553: setp.eq.s32 %p1418, %r2865, 53; @%p1418 bra $L__BB107_1658; setp.eq.s32 %p1419, %r2865, 54; @%p1419 bra $L__BB107_1657; setp.eq.s32 %p1420, %r2865, 55; @%p1420 bra $L__BB107_1556; bra.uni $L__BB107_1458; $L__BB107_1556: ld.local.v4.f32 {%f14429, %f14430, %f14431, %f14432}, [%rd2173]; { // callseq 374, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f14429; st.param.f32 [param0+4], %f14430; st.param.f32 [param0+8], %f14431; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; .param .b32 param4; st.param.b32 [param4+0], %r2868; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 374 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1513: setp.eq.s32 %p1450, %r2865, 29; @%p1450 bra $L__BB107_1674; setp.eq.s32 %p1451, %r2865, 30; @%p1451 bra $L__BB107_1673; setp.eq.s32 %p1452, %r2865, 31; @%p1452 bra $L__BB107_1516; bra.uni $L__BB107_1458; $L__BB107_1516: mov.u32 %r2917, 31; add.s32 %r2918, %r678, 1; mov.u32 %r2919, 8; { // callseq 396, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r2919; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r2917; st.param.b32 [param3+4], %r2866; st.param.b32 [param3+8], %r2867; st.param.b32 [param3+12], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2918; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 396 bra.uni $L__BB107_1456; $L__BB107_1595: setp.eq.s32 %p1387, %r2865, 77; @%p1387 bra $L__BB107_1639; setp.eq.s32 %p1388, %r2865, 78; @%p1388 bra $L__BB107_1638; setp.eq.s32 %p1389, %r2865, 79; @%p1389 bra $L__BB107_1598; bra.uni $L__BB107_1458; $L__BB107_1598: add.s32 %r2880, %r678, 1; { // callseq 352, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2880; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 352 bra.uni $L__BB107_1456; $L__BB107_1493: setp.eq.s32 %p1466, %r2865, 17; @%p1466 bra $L__BB107_1682; setp.eq.s32 %p1467, %r2865, 18; @%p1467 bra $L__BB107_1681; setp.eq.s32 %p1468, %r2865, 19; @%p1468 bra $L__BB107_1496; bra.uni $L__BB107_1458; $L__BB107_1496: { // callseq 407, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 407 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1572: setp.eq.s32 %p1403, %r2865, 65; @%p1403 bra $L__BB107_1650; setp.eq.s32 %p1404, %r2865, 66; @%p1404 bra $L__BB107_1649; setp.eq.s32 %p1405, %r2865, 67; @%p1405 bra $L__BB107_1575; bra.uni $L__BB107_1458; $L__BB107_1575: mov.u32 %r2889, 67; { // callseq 362, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2889, %r2866, %r2867, %r2868}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 362 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1532: setp.eq.s32 %p1435, %r2865, 41; @%p1435 bra $L__BB107_1666; setp.eq.s32 %p1436, %r2865, 42; @%p1436 bra $L__BB107_1665; setp.eq.s32 %p1437, %r2865, 43; @%p1437 bra $L__BB107_1535; bra.uni $L__BB107_1458; $L__BB107_1535: add.s32 %r2909, %r678, 1; { // callseq 386, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2909; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 386 bra.uni $L__BB107_1456; $L__BB107_1614: setp.eq.s32 %p1372, %r2865, 89; @%p1372 bra $L__BB107_1631; setp.eq.s32 %p1373, %r2865, 90; @%p1373 bra $L__BB107_1630; setp.eq.s32 %p1374, %r2865, 92; @%p1374 bra $L__BB107_1617; bra.uni $L__BB107_1458; $L__BB107_1617: add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1483: setp.eq.s32 %p1474, %r2865, 11; @%p1474 bra $L__BB107_1686; setp.eq.s32 %p1475, %r2865, 12; @%p1475 bra $L__BB107_1685; setp.eq.s32 %p1476, %r2865, 13; @%p1476 bra $L__BB107_1486; bra.uni $L__BB107_1458; $L__BB107_1486: mov.u32 %r2938, 13; add.s32 %r2939, %r678, 1; mov.u32 %r2940, 8; { // callseq 411, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r2940; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r2938; st.param.b32 [param3+4], %r2866; st.param.b32 [param3+8], %r2867; st.param.b32 [param3+12], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2939; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 411 bra.uni $L__BB107_1456; $L__BB107_1562: setp.eq.s32 %p1411, %r2865, 59; @%p1411 bra $L__BB107_1654; setp.eq.s32 %p1412, %r2865, 60; @%p1412 bra $L__BB107_1653; setp.eq.s32 %p1413, %r2865, 61; @%p1413 bra $L__BB107_1565; bra.uni $L__BB107_1458; $L__BB107_1565: mov.u32 %r2896, 61; add.s32 %r2897, %r678, 1; { // callseq 368, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2896; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2897; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 368 bra.uni $L__BB107_1456; $L__BB107_1522: setp.eq.s32 %p1443, %r2865, 35; @%p1443 bra $L__BB107_1670; setp.eq.s32 %p1444, %r2865, 36; @%p1444 bra $L__BB107_1669; setp.eq.s32 %p1445, %r2865, 37; @%p1445 bra $L__BB107_1525; bra.uni $L__BB107_1458; $L__BB107_1525: mov.u32 %r2912, 37; { // callseq 391, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2912, %r2866, %r2867, %r2868}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 391 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1604: setp.eq.s32 %p1380, %r2865, 83; @%p1380 bra $L__BB107_1635; setp.eq.s32 %p1381, %r2865, 84; @%p1381 bra $L__BB107_1634; setp.eq.s32 %p1382, %r2865, 85; @%p1382 bra $L__BB107_1607; bra.uni $L__BB107_1458; $L__BB107_1607: add.s32 %r2877, %r678, 1; { // callseq 346, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2877; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 346 bra.uni $L__BB107_1456; $L__BB107_1502: setp.eq.s32 %p1459, %r2865, 23; @%p1459 bra $L__BB107_1678; setp.eq.s32 %p1460, %r2865, 24; @%p1460 bra $L__BB107_1677; setp.eq.s32 %p1461, %r2865, 25; @%p1461 bra $L__BB107_1505; bra.uni $L__BB107_1458; $L__BB107_1505: add.s32 %r2923, %r678, 1; { // callseq 402, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2923; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 402 bra.uni $L__BB107_1456; $L__BB107_1581: setp.eq.s32 %p1396, %r2865, 71; @%p1396 bra $L__BB107_1646; setp.eq.s32 %p1397, %r2865, 72; @%p1397 bra $L__BB107_1645; setp.eq.s32 %p1398, %r2865, 73; @%p1398 bra $L__BB107_1584; bra.uni $L__BB107_1458; $L__BB107_1584: add.s32 %r2883, %r678, 1; { // callseq 356, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2883; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 356 bra.uni $L__BB107_1456; $L__BB107_1541: setp.eq.s32 %p1428, %r2865, 47; @%p1428 bra $L__BB107_1662; setp.eq.s32 %p1429, %r2865, 48; @%p1429 bra $L__BB107_1661; setp.eq.s32 %p1430, %r2865, 49; @%p1430 bra $L__BB107_1544; bra.uni $L__BB107_1458; $L__BB107_1544: { // callseq 380, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 380 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1623: setp.eq.s32 %p1365, %r2865, 98; @%p1365 bra $L__BB107_1694; setp.eq.s32 %p1366, %r2865, 97; @%p1366 bra $L__BB107_1629; setp.ne.s32 %p1367, %r2865, 96; @%p1367 bra $L__BB107_1458; bra.uni $L__BB107_1626; $L__BB107_1693: mov.u32 %r2947, 2; add.s32 %r2948, %r678, 1; mov.u32 %r2949, 8; { // callseq 418, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2947; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; .param .b32 param3; st.param.b32 [param3+0], %r2949; .param .b32 param4; st.param.b32 [param4+0], %r2948; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 418 bra.uni $L__BB107_1456; $L__BB107_1692: mov.u32 %r2946, 3; { // callseq 417, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2946, %r2866, %r2867, %r2868}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 417 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1660: { // callseq 379, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 379 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1659: { // callseq 378, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 378 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1676: mov.u32 %r2922, 26; { // callseq 401, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2922; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 401 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1675: mov.u32 %r2921, 27; { // callseq 400, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2921; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 400 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1644: { // callseq 355, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 355 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1643: { // callseq 354, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 354 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1684: st.local.u32 [%rd388], %r2866; add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1683: ld.global.v4.u32 {%r2931, %r2932, %r2933, %r2934}, [%rd386+16]; add.s32 %r5723, %r678, 2; st.local.u32 [%rd387], %r2932; st.local.u32 [%rd387+4], %r2933; st.local.u32 [%rd387+8], %r2934; bra.uni $L__BB107_1456; $L__BB107_1652: mov.u32 %r2894, 62; add.s32 %r2895, %r678, 1; { // callseq 367, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2894; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2895; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 367 bra.uni $L__BB107_1456; $L__BB107_1651: mov.u32 %r2893, 63; { // callseq 366, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2893; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 366 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1668: { // callseq 390, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; .param .b32 param4; st.param.b32 [param4+0], %r2868; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 390 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1667: mov.u32 %r2911, 39; { // callseq 389, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2911, %r2866, %r2867, %r2868}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 389 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1633: add.s32 %r2876, %r678, 1; { // callseq 345, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2876; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 345 bra.uni $L__BB107_1456; $L__BB107_1632: add.s32 %r2875, %r678, 1; { // callseq 344, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2875; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 344 bra.uni $L__BB107_1456; $L__BB107_1689: mov.u32 %r2943, 8; { // callseq 414, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2943; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 414 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1687: ld.local.f32 %f14453, [%rd388]; setp.gtu.ftz.f32 %p1493, %f14453, 0f00000000; add.s32 %r5723, %r678, 1; @%p1493 bra $L__BB107_1456; add.s32 %r2942, %r2866, %r678; add.s32 %r5723, %r2942, 1; bra.uni $L__BB107_1456; $L__BB107_1656: mov.u32 %r2903, 56; { // callseq 373, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2903; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 373 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1655: mov.u32 %r2901, 57; add.s32 %r2902, %r678, 1; { // callseq 372, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2901; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2902; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 372 bra.uni $L__BB107_1456; $L__BB107_1672: mov.u32 %r2914, 32; add.s32 %r2915, %r678, 1; mov.u32 %r2916, 8; { // callseq 395, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r2916; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r2914; st.param.b32 [param3+4], %r2866; st.param.b32 [param3+8], %r2867; st.param.b32 [param3+12], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2915; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 395 bra.uni $L__BB107_1456; $L__BB107_1671: ld.local.f32 %f14436, [%rd387]; ld.local.f32 %f14437, [%rd387+4]; ld.local.f32 %f14438, [%rd387+8]; st.local.f32 [%rd315], %f14436; st.local.f32 [%rd337], %f14437; st.local.f32 [%rd338], %f14438; st.local.f32 [%rd388], %f14436; st.local.f32 [%rd388+4], %f14437; st.local.f32 [%rd388+8], %f14438; add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1637: { // callseq 351, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 351 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1636: mov.u32 %r2879, 81; { // callseq 350, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2879; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 350 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1680: ld.local.f32 %f14439, [%rd387]; ld.local.v4.f32 {%f14440, %f14441, %f14442, %f14443}, [%rd2173]; ld.local.f32 %f14447, [%rd387+4]; ld.local.f32 %f14448, [%rd387+8]; add.ftz.f32 %f14449, %f14448, %f14442; add.ftz.f32 %f14450, %f14447, %f14441; add.ftz.f32 %f14451, %f14439, %f14440; st.local.v2.f32 [%rd2173], {%f14451, %f14450}; st.local.f32 [%rd314], %f14449; add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1679: mov.u32 %r2929, 21; { // callseq 406, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2929; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 406 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1648: mov.u32 %r2888, 68; { // callseq 361, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2888; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 361 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1647: mov.u32 %r2886, 69; add.s32 %r2887, %r678, 1; { // callseq 360, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2886; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2887; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 360 bra.uni $L__BB107_1456; $L__BB107_1664: mov.u32 %r2907, 44; add.s32 %r2908, %r678, 1; { // callseq 385, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2907; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2908; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 385 bra.uni $L__BB107_1456; $L__BB107_1663: { // callseq 384, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 384 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1628: add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1627: mov.u32 %r2869, 94; add.s32 %r2870, %r678, 1; { // callseq 339, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2869; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2870; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 339 bra.uni $L__BB107_1456; $L__BB107_1691: st.local.u32 [%rd348], %r2866; st.local.u32 [%rd349], %r2867; st.local.u32 [%rd350], %r2868; add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1690: ld.local.f32 %f14454, [%rd387]; ld.local.f32 %f14455, [%rd387+4]; ld.local.f32 %f14456, [%rd387+8]; st.local.f32 [%rd348], %f14454; st.local.f32 [%rd349], %f14455; st.local.f32 [%rd350], %f14456; add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1658: { // callseq 376, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 376 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1657: add.s32 %r2904, %r678, 1; { // callseq 375, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2904; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 375 bra.uni $L__BB107_1456; $L__BB107_1674: { // callseq 398, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; .param .b32 param4; st.param.b32 [param4+0], %r2868; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 398 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1673: { // callseq 397, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; .param .b32 param4; st.param.b32 [param4+0], %r2868; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 397 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1639: setp.eq.s32 %p1488, %r2868, 255; add.s32 %r5723, %r678, 1; @%p1488 bra $L__BB107_1456; ld.local.f32 %f14425, [%rd387]; add.s32 %r2882, %r2868, %r2867; mul.wide.u32 %rd2401, %r2882, 4; add.s64 %rd2402, %rd2157, %rd2401; st.local.f32 [%rd2402], %f14425; bra.uni $L__BB107_1456; $L__BB107_1638: add.s32 %r2881, %r678, 1; { // callseq 353, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2881; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 353 bra.uni $L__BB107_1456; $L__BB107_1682: { // callseq 409, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; .param .b32 param4; st.param.b32 [param4+0], %r2868; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 409 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1681: { // callseq 408, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 408 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1650: { // callseq 364, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 364 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1649: add.s32 %r2890, %r678, 1; { // callseq 363, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2890; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 363 bra.uni $L__BB107_1456; $L__BB107_1666: add.s32 %r2910, %r678, 1; { // callseq 388, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r2910; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 388 bra.uni $L__BB107_1456; $L__BB107_1665: { // callseq 387, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 387 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1631: mov.u32 %r2873, 89; { // callseq 342, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2873; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 342 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1630: mov.u32 %r2871, 90; add.s32 %r2872, %r678, 1; { // callseq 341, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2871; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; .param .b32 param3; st.param.b32 [param3+0], %r2872; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 341 bra.uni $L__BB107_1456; $L__BB107_1686: { // callseq 413, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 413 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1685: { // callseq 412, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 412 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1654: add.s32 %r2899, %r678, 1; { // callseq 370, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2899; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 370 bra.uni $L__BB107_1456; $L__BB107_1653: add.s32 %r2898, %r678, 1; { // callseq 369, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; .param .b32 param4; st.param.b32 [param4+0], %r2898; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 369 bra.uni $L__BB107_1456; $L__BB107_1670: { // callseq 393, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 393 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1669: mov.u32 %r2913, 36; { // callseq 392, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2913; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 392 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1635: { // callseq 348, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 348 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1634: { // callseq 347, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 347 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1678: mov.u32 %r2925, 23; add.s32 %r2926, %r678, 1; { // callseq 404, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2925; st.param.b32 [param1+4], %r2866; st.param.b32 [param1+8], %r2867; st.param.b32 [param1+12], %r2868; .param .b32 param2; st.param.b32 [param2+0], %r2926; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5723, [retval0+0]; } // callseq 404 bra.uni $L__BB107_1456; $L__BB107_1677: mov.u32 %r2924, 24; { // callseq 403, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2924; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 403 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1646: mov.u32 %r2884, 71; { // callseq 358, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2884; st.param.b32 [param2+4], %r2866; st.param.b32 [param2+8], %r2867; st.param.b32 [param2+12], %r2868; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 358 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1645: { // callseq 357, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; .param .b32 param3; st.param.b32 [param3+0], %r2868; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 357 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1662: mov.u32 %r2906, 8; { // callseq 382, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1602; .param .b64 param1; st.param.b64 [param1+0], %rd2172; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .b32 param3; st.param.b32 [param3+0], %r2866; .param .b32 param4; st.param.b32 [param4+0], %r2867; .param .b32 param5; st.param.b32 [param5+0], %r2906; call.uni _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 382 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1661: { // callseq 381, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r2866; .param .b32 param3; st.param.b32 [param3+0], %r2867; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 381 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1694: { // callseq 419, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 419 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1629: { // callseq 340, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r2866; .param .b32 param2; st.param.b32 [param2+0], %r2867; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 340 add.s32 %r5723, %r678, 1; bra.uni $L__BB107_1456; $L__BB107_1642: ld.local.f32 %f14428, [%rd387]; st.local.f32 [%rd391], %f14428; bra.uni $L__BB107_1456; $L__BB107_1594: ld.local.f32 %f14426, [%rd390]; st.local.f32 [%rd391], %f14426; bra.uni $L__BB107_1456; $L__BB107_1458: ld.local.u32 %r784, [%rd342]; setp.lt.s32 %p1494, %r784, 1; @%p1494 bra $L__BB107_3392; mov.u32 %r5724, 0; $L__BB107_1460: cvt.s64.s32 %rd392, %r5724; mul.wide.s32 %rd2508, %r5724, 80; add.s64 %rd2509, %rd2173, %rd2508; add.s64 %rd393, %rd2509, 316; ld.local.u32 %r786, [%rd2509+316]; setp.gt.u32 %p1495, %r786, 27; @%p1495 bra $L__BB107_1462; mov.u32 %r2953, 1; shl.b32 %r2954, %r2953, %r786; and.b32 %r2955, %r2954, 165675008; setp.ne.s32 %p1496, %r2955, 0; @%p1496 bra $L__BB107_1695; $L__BB107_1462: cvt.u32.u64 %r2956, %rd392; add.s32 %r5724, %r2956, 1; setp.eq.s32 %p1497, %r786, 28; @%p1497 bra $L__BB107_1695; setp.lt.s32 %p1498, %r5724, %r784; mov.u32 %r5802, 0; @%p1498 bra $L__BB107_1460; bra.uni $L__BB107_3392; $L__BB107_1695: ld.local.u32 %r2958, [%rd293]; and.b32 %r2959, %r2958, 1; setp.eq.b32 %p1499, %r2959, 1; xor.pred %p1501, %p1499, %p1284; not.pred %p1502, %p1501; ld.local.f32 %f29598, [%rd393+28]; @%p1502 bra $L__BB107_1697; rcp.approx.ftz.f32 %f29598, %f29598; $L__BB107_1697: ld.local.f32 %f2482, [%rd393+20]; setp.leu.ftz.f32 %p1503, %f2482, 0f00000000; mov.f32 %f29608, 0f00000000; mov.f32 %f29609, %f29608; @%p1503 bra $L__BB107_1715; ld.local.f32 %f2483, [%rd393+24]; setp.leu.ftz.f32 %p1504, %f2483, 0f00000000; mov.f32 %f29609, %f29608; @%p1504 bra $L__BB107_1715; @%p1249 bra $L__BB107_1702; setp.geu.ftz.f32 %p1506, %f1951, 0f3F800000; mul.wide.u32 %rd2510, %r2961, 4; add.s64 %rd2511, %rd355, %rd2510; ld.global.f32 %f29601, [%rd2511]; ld.global.f32 %f29602, [%rd2511+4]; @%p1506 bra $L__BB107_1706; add.ftz.f32 %f14463, %f29601, %f1954; cvt.rmi.ftz.f32.f32 %f14464, %f14463; sub.ftz.f32 %f29601, %f14463, %f14464; add.ftz.f32 %f14465, %f29602, %f1955; cvt.rmi.ftz.f32.f32 %f14466, %f14465; sub.ftz.f32 %f29602, %f14465, %f14466; bra.uni $L__BB107_1706; $L__BB107_1702: setp.eq.s32 %p1507, %r623, 0; mov.u32 %r5728, 0; @%p1507 bra $L__BB107_1705; mov.u32 %r5725, 0; mov.u32 %r5726, %r623; mov.u32 %r5728, %r5725; $L__BB107_1704: clz.b32 %r2965, %r5726; add.s32 %r2966, %r2965, %r5725; mul.wide.u32 %rd2512, %r2966, 4; mov.u64 %rd2513, sobol_burley_table; add.s64 %rd2514, %rd2513, %rd2512; ld.const.u32 %r2967, [%rd2514+128]; xor.b32 %r5728, %r2967, %r5728; add.s32 %r5725, %r2966, 1; shl.b32 %r2968, %r5726, %r2965; shl.b32 %r5726, %r2968, 1; setp.ne.s32 %p1508, %r5726, 0; @%p1508 bra $L__BB107_1704; $L__BB107_1705: mul.lo.s32 %r2969, %r5728, 1025551850; xor.b32 %r2970, %r2969, %r5728; add.s32 %r2971, %r2970, %r624; mul.lo.s32 %r2972, %r2971, %r627; mul.lo.s32 %r2973, %r2972, 89287766; xor.b32 %r2974, %r2973, %r2972; mul.lo.s32 %r2975, %r2974, 1403136100; xor.b32 %r2976, %r2975, %r2974; brev.b32 %r2977, %r2976; cvt.rn.f32.u32 %f14467, %r2977; mul.ftz.f32 %f29602, %f14467, 0f2F7FFFFE; mov.f32 %f29601, %f1952; $L__BB107_1706: mul.ftz.f32 %f2497, %f29602, 0f40C90FDB; mul.ftz.f32 %f29606, %f2482, %f2482; setp.eq.ftz.f32 %p1509, %f2482, %f2483; @%p1509 bra $L__BB107_1710; bra.uni $L__BB107_1707; $L__BB107_1710: mul.ftz.f32 %f14528, %f2497, 0f3EA2F983; mov.b32 %r2987, %f14528; and.b32 %r2988, %r2987, -2147483648; or.b32 %r2989, %r2988, 1056964608; mov.b32 %f14529, %r2989; add.ftz.f32 %f14530, %f14528, %f14529; cvt.rzi.ftz.s32.f32 %r2990, %f14530; cvt.rn.f32.s32 %f14531, %r2990; fma.rn.ftz.f32 %f14532, %f14531, 0fC0490000, %f2497; fma.rn.ftz.f32 %f14533, %f14531, 0fBA7DA000, %f14532; fma.rn.ftz.f32 %f14534, %f14531, 0fB4222000, %f14533; fma.rn.ftz.f32 %f14535, %f14531, 0fACB4611A, %f14534; add.ftz.f32 %f14536, %f14535, 0fBFC90FDB; add.ftz.f32 %f14537, %f14536, 0f3FC90FDB; mul.ftz.f32 %f14538, %f14537, %f14537; and.b32 %r2991, %r2990, 1; setp.eq.b32 %p1517, %r2991, 1; neg.ftz.f32 %f14539, %f14537; selp.f32 %f14540, %f14539, %f14537, %p1517; fma.rn.ftz.f32 %f14541, %f14538, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f14542, %f14538, %f14541, 0f3C088777; fma.rn.ftz.f32 %f14543, %f14538, %f14542, 0fBE2AAAA6; mul.ftz.f32 %f14544, %f14540, %f14543; fma.rn.ftz.f32 %f14545, %f14538, %f14544, %f14540; fma.rn.ftz.f32 %f14546, %f14538, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f14547, %f14538, %f14546, 0fBAB60B58; fma.rn.ftz.f32 %f14548, %f14538, %f14547, 0f3D2AAAAA; fma.rn.ftz.f32 %f14549, %f14538, %f14548, 0fBF000000; fma.rn.ftz.f32 %f14550, %f14538, %f14549, 0f3F800000; neg.ftz.f32 %f14551, %f14550; selp.f32 %f14552, %f14551, %f14550, %p1517; abs.ftz.f32 %f14553, %f14545; setp.gt.ftz.f32 %p1518, %f14553, 0f3F800000; selp.f32 %f29604, 0f00000000, %f14545, %p1518; abs.ftz.f32 %f14554, %f14552; setp.gt.ftz.f32 %p1519, %f14554, 0f3F800000; selp.f32 %f29605, 0f00000000, %f14552, %p1519; bra.uni $L__BB107_1711; $L__BB107_1707: div.approx.ftz.f32 %f14468, %f2483, %f2482; add.ftz.f32 %f14469, %f2497, 0f3FC90FDB; sin.approx.ftz.f32 %f14470, %f14469; cos.approx.ftz.f32 %f14471, %f14469; div.approx.ftz.f32 %f14472, %f14470, %f14471; mul.ftz.f32 %f2499, %f14468, %f14472; abs.ftz.f32 %f2500, %f2499; setp.leu.ftz.f32 %p1510, %f2500, 0f3F800000; mov.f32 %f29603, %f2500; @%p1510 bra $L__BB107_1709; rcp.approx.ftz.f32 %f29603, %f2500; $L__BB107_1709: mul.rn.ftz.f32 %f14473, %f29603, %f29603; mov.f32 %f14474, 0fC0B59883; mov.f32 %f14475, 0fBF52C7EA; fma.rn.ftz.f32 %f14476, %f14473, %f14475, %f14474; mov.f32 %f14477, 0fC0D21907; fma.rn.ftz.f32 %f14478, %f14476, %f14473, %f14477; mul.ftz.f32 %f14479, %f14473, %f14478; mul.ftz.f32 %f14480, %f29603, %f14479; add.ftz.f32 %f14481, %f14473, 0f41355DC0; mov.f32 %f14482, 0f41E6BD60; fma.rn.ftz.f32 %f14483, %f14481, %f14473, %f14482; mov.f32 %f14484, 0f419D92C8; fma.rn.ftz.f32 %f14485, %f14483, %f14473, %f14484; rcp.approx.ftz.f32 %f14486, %f14485; fma.rn.ftz.f32 %f14487, %f14480, %f14486, %f29603; mov.f32 %f14488, 0f3FC90FDB; sub.ftz.f32 %f14489, %f14488, %f14487; setp.gt.ftz.f32 %p1511, %f2500, 0f3F800000; selp.f32 %f14490, %f14489, %f14487, %p1511; mov.b32 %r2978, %f14490; mov.b32 %r2979, %f2499; and.b32 %r2980, %r2979, -2147483648; or.b32 %r2981, %r2980, %r2978; mov.b32 %f14491, %r2981; setp.le.ftz.f32 %p1512, %f2500, 0f7F800000; selp.f32 %f14492, %f14491, %f14490, %p1512; add.ftz.f32 %f14493, %f14492, 0f40490FDB; setp.gt.ftz.f32 %p1513, %f29602, 0f3F000000; selp.f32 %f14494, %f14493, %f14492, %p1513; mul.ftz.f32 %f14495, %f14494, 0f3EA2F983; mov.b32 %r2982, %f14495; and.b32 %r2983, %r2982, -2147483648; or.b32 %r2984, %r2983, 1056964608; mov.b32 %f14496, %r2984; add.ftz.f32 %f14497, %f14495, %f14496; cvt.rzi.ftz.s32.f32 %r2985, %f14497; cvt.rn.f32.s32 %f14498, %r2985; fma.rn.ftz.f32 %f14499, %f14498, 0fC0490000, %f14494; fma.rn.ftz.f32 %f14500, %f14498, 0fBA7DA000, %f14499; fma.rn.ftz.f32 %f14501, %f14498, 0fB4222000, %f14500; fma.rn.ftz.f32 %f14502, %f14498, 0fACB4611A, %f14501; add.ftz.f32 %f14503, %f14502, 0fBFC90FDB; add.ftz.f32 %f14504, %f14503, 0f3FC90FDB; mul.ftz.f32 %f14505, %f14504, %f14504; and.b32 %r2986, %r2985, 1; setp.eq.b32 %p1514, %r2986, 1; neg.ftz.f32 %f14506, %f14504; selp.f32 %f14507, %f14506, %f14504, %p1514; fma.rn.ftz.f32 %f14508, %f14505, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f14509, %f14505, %f14508, 0f3C088777; fma.rn.ftz.f32 %f14510, %f14505, %f14509, 0fBE2AAAA6; mul.ftz.f32 %f14511, %f14507, %f14510; fma.rn.ftz.f32 %f14512, %f14505, %f14511, %f14507; fma.rn.ftz.f32 %f14513, %f14505, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f14514, %f14505, %f14513, 0fBAB60B58; fma.rn.ftz.f32 %f14515, %f14505, %f14514, 0f3D2AAAAA; fma.rn.ftz.f32 %f14516, %f14505, %f14515, 0fBF000000; fma.rn.ftz.f32 %f14517, %f14505, %f14516, 0f3F800000; neg.ftz.f32 %f14518, %f14517; selp.f32 %f14519, %f14518, %f14517, %p1514; abs.ftz.f32 %f14520, %f14512; setp.gt.ftz.f32 %p1515, %f14520, 0f3F800000; selp.f32 %f29604, 0f00000000, %f14512, %p1515; abs.ftz.f32 %f14521, %f14519; setp.gt.ftz.f32 %p1516, %f14521, 0f3F800000; selp.f32 %f29605, 0f00000000, %f14519, %p1516; mul.ftz.f32 %f14522, %f29605, %f29605; div.approx.ftz.f32 %f14523, %f14522, %f29606; mul.ftz.f32 %f14524, %f29604, %f29604; mul.ftz.f32 %f14525, %f2483, %f2483; div.approx.ftz.f32 %f14526, %f14524, %f14525; add.ftz.f32 %f14527, %f14523, %f14526; rcp.approx.ftz.f32 %f29606, %f14527; $L__BB107_1711: mov.f32 %f14555, 0f3F800000; sub.ftz.f32 %f2512, %f14555, %f29601; setp.eq.s32 %p1520, %r786, 22; @%p1520 bra $L__BB107_1713; bra.uni $L__BB107_1712; $L__BB107_1713: lg2.approx.ftz.f32 %f14556, %f2512; mul.ftz.f32 %f29607, %f14556, 0fBF317218; bra.uni $L__BB107_1714; $L__BB107_1712: div.approx.ftz.f32 %f29607, %f29601, %f2512; $L__BB107_1714: fma.rn.ftz.f32 %f14557, %f29606, %f29607, 0f3F800000; rcp.approx.ftz.f32 %f14559, %f14557; sub.ftz.f32 %f14560, %f14555, %f14559; mov.f32 %f14561, 0f00000000; max.ftz.f32 %f14562, %f14560, %f14561; sqrt.approx.ftz.f32 %f14563, %f14562; mul.ftz.f32 %f29609, %f29604, %f14563; mul.ftz.f32 %f29608, %f29605, %f14563; $L__BB107_1715: setp.ne.s32 %p1521, %r632, -1; mov.u32 %r5729, %r632; @%p1521 bra $L__BB107_1717; add.s64 %rd2516, %rd353, %rd2257; ld.global.u32 %r5729, [%rd2516]; cvt.s64.s32 %rd4795, %r5729; mul.wide.s32 %rd2517, %r5729, 4; add.s64 %rd4794, %rd287, %rd2517; $L__BB107_1717: st.local.u32 [%rd339], %r5729; st.local.u32 [%rd354], %r634; mov.u32 %r2992, 0; st.local.u32 [%rd293], %r2992; ld.global.u32 %r797, [%rd4794]; st.local.u32 [%rd290], %r797; and.b32 %r798, %r797, 2; setp.eq.s32 %p1522, %r798, 0; mul.lo.s64 %rd2518, %rd4795, 208; add.s64 %rd398, %rd300, %rd2518; @%p1522 bra $L__BB107_1724; ld.global.u32 %rd2519, [%rd398+168]; ld.global.u32 %r2993, [%rd398+152]; shl.b32 %r2994, %r2993, 1; cvt.rn.f32.s32 %f14564, %r2994; mul.ftz.f32 %f14565, %f1950, %f14564; cvt.rzi.ftz.s32.f32 %r2995, %f14565; add.s32 %r2996, %r2994, -1; min.s32 %r2997, %r2995, %r2996; cvt.rn.f32.s32 %f14566, %r2997; sub.ftz.f32 %f2520, %f14565, %f14566; cvt.s64.s32 %rd2520, %r2997; add.s64 %rd2521, %rd2520, %rd2519; shl.b64 %rd2522, %rd2521, 6; add.s64 %rd399, %rd301, %rd2522; ld.global.v4.f32 {%f14567, %f14568, %f14569, %f14570}, [%rd399]; ld.global.v4.f32 {%f14571, %f14572, %f14573, %f14574}, [%rd399+64]; mul.ftz.f32 %f14575, %f14568, %f14572; fma.rn.ftz.f32 %f14576, %f14567, %f14571, %f14575; mul.ftz.f32 %f14577, %f14570, %f14574; fma.rn.ftz.f32 %f14578, %f14569, %f14573, %f14577; add.ftz.f32 %f2529, %f14576, %f14578; setp.gt.ftz.f32 %p1523, %f2529, 0f3F7FDF3B; mov.f32 %f14579, 0f3F800000; sub.ftz.f32 %f2530, %f14579, %f2520; @%p1523 bra $L__BB107_1720; bra.uni $L__BB107_1719; $L__BB107_1720: mul.ftz.f32 %f14633, %f14567, %f2530; fma.rn.ftz.f32 %f14634, %f14571, %f2520, %f14633; mul.ftz.f32 %f14635, %f14568, %f2530; fma.rn.ftz.f32 %f14636, %f14572, %f2520, %f14635; mul.ftz.f32 %f14637, %f14569, %f2530; fma.rn.ftz.f32 %f14638, %f14573, %f2520, %f14637; mul.ftz.f32 %f14639, %f14570, %f2530; fma.rn.ftz.f32 %f14640, %f14574, %f2520, %f14639; mul.ftz.f32 %f14641, %f14636, %f14636; fma.rn.ftz.f32 %f14642, %f14634, %f14634, %f14641; mul.ftz.f32 %f14643, %f14640, %f14640; fma.rn.ftz.f32 %f14644, %f14638, %f14638, %f14643; add.ftz.f32 %f14645, %f14642, %f14644; rsqrt.approx.ftz.f32 %f14646, %f14645; mul.ftz.f32 %f29613, %f14640, %f14646; mul.ftz.f32 %f29612, %f14638, %f14646; mul.ftz.f32 %f29611, %f14636, %f14646; mul.ftz.f32 %f29610, %f14634, %f14646; bra.uni $L__BB107_1721; $L__BB107_1719: mov.f32 %f14580, 0fBF800000; max.ftz.f32 %f14581, %f2529, %f14580; min.ftz.f32 %f14583, %f14581, %f14579; abs.ftz.f32 %f14584, %f14583; sub.ftz.f32 %f14585, %f14579, %f14584; mul.ftz.f32 %f14586, %f14585, 0f3F000000; sqrt.approx.ftz.f32 %f14587, %f14586; setp.gt.ftz.f32 %p1524, %f14584, 0f3F11EB85; selp.f32 %f14588, %f14587, %f14584, %p1524; mul.ftz.f32 %f14589, %f14588, %f14588; mov.f32 %f14590, 0f3C94D2E9; mov.f32 %f14591, 0f3D53F941; fma.rn.ftz.f32 %f14592, %f14591, %f14589, %f14590; mov.f32 %f14593, 0f3D3F841F; fma.rn.ftz.f32 %f14594, %f14592, %f14589, %f14593; mov.f32 %f14595, 0f3D994929; fma.rn.ftz.f32 %f14596, %f14594, %f14589, %f14595; mov.f32 %f14597, 0f3E2AAB94; fma.rn.ftz.f32 %f14598, %f14596, %f14589, %f14597; mul.ftz.f32 %f14599, %f14589, %f14598; fma.rn.ftz.f32 %f14600, %f14599, %f14588, %f14588; add.ftz.f32 %f14601, %f14600, %f14600; mov.f32 %f14602, 0f3FC90FDB; sub.ftz.f32 %f14603, %f14602, %f14600; selp.f32 %f14604, %f14601, %f14603, %p1524; setp.lt.ftz.f32 %p1525, %f14583, 0f00000000; mov.f32 %f14605, 0f40490FDB; sub.ftz.f32 %f14606, %f14605, %f14604; selp.f32 %f14607, %f14606, %f14604, %p1525; mul.ftz.f32 %f14608, %f14567, %f2529; sub.ftz.f32 %f14609, %f14571, %f14608; mul.ftz.f32 %f14610, %f14568, %f2529; sub.ftz.f32 %f14611, %f14572, %f14610; mul.ftz.f32 %f14612, %f14569, %f2529; sub.ftz.f32 %f14613, %f14573, %f14612; mul.ftz.f32 %f14614, %f14570, %f2529; sub.ftz.f32 %f14615, %f14574, %f14614; mul.ftz.f32 %f14616, %f14611, %f14611; fma.rn.ftz.f32 %f14617, %f14609, %f14609, %f14616; mul.ftz.f32 %f14618, %f14615, %f14615; fma.rn.ftz.f32 %f14619, %f14613, %f14613, %f14618; add.ftz.f32 %f14620, %f14617, %f14619; rsqrt.approx.ftz.f32 %f14621, %f14620; mul.ftz.f32 %f14622, %f14609, %f14621; mul.ftz.f32 %f14623, %f14611, %f14621; mul.ftz.f32 %f14624, %f14613, %f14621; mul.ftz.f32 %f14625, %f14615, %f14621; mul.ftz.f32 %f14626, %f2520, %f14607; cos.approx.ftz.f32 %f14627, %f14626; sin.approx.ftz.f32 %f14628, %f14626; mul.ftz.f32 %f14629, %f14622, %f14628; mul.ftz.f32 %f14630, %f14623, %f14628; mul.ftz.f32 %f14631, %f14624, %f14628; mul.ftz.f32 %f14632, %f14625, %f14628; fma.rn.ftz.f32 %f29613, %f14570, %f14627, %f14632; fma.rn.ftz.f32 %f29612, %f14569, %f14627, %f14631; fma.rn.ftz.f32 %f29611, %f14568, %f14627, %f14630; fma.rn.ftz.f32 %f29610, %f14567, %f14627, %f14629; $L__BB107_1721: ld.global.v4.f32 {%f14647, %f14648, %f14649, %f14650}, [%rd399+16]; ld.global.v4.f32 {%f14655, %f14656, %f14657, %f14658}, [%rd399+80]; mul.ftz.f32 %f14663, %f2520, %f14655; mul.ftz.f32 %f14664, %f2520, %f14656; mul.ftz.f32 %f14665, %f2520, %f14657; mul.ftz.f32 %f14666, %f2520, %f14658; fma.rn.ftz.f32 %f2543, %f2530, %f14647, %f14663; fma.rn.ftz.f32 %f2544, %f2530, %f14648, %f14664; fma.rn.ftz.f32 %f2545, %f2530, %f14649, %f14665; fma.rn.ftz.f32 %f14667, %f2530, %f14650, %f14666; ld.global.v4.f32 {%f14668, %f14669, %f14670, %f14671}, [%rd399+32]; ld.global.v4.f32 {%f14676, %f14677, %f14678, %f14679}, [%rd399+96]; mul.ftz.f32 %f14684, %f2520, %f14676; mul.ftz.f32 %f14685, %f2520, %f14677; mul.ftz.f32 %f14686, %f2520, %f14678; mul.ftz.f32 %f14687, %f2520, %f14679; fma.rn.ftz.f32 %f14688, %f2530, %f14668, %f14684; fma.rn.ftz.f32 %f14689, %f2530, %f14669, %f14685; fma.rn.ftz.f32 %f14690, %f2530, %f14670, %f14686; fma.rn.ftz.f32 %f14691, %f2530, %f14671, %f14687; ld.global.v4.f32 {%f14692, %f14693, %f14694, %f14695}, [%rd399+48]; ld.global.v4.f32 {%f14700, %f14701, %f14702, %f14703}, [%rd399+112]; mul.ftz.f32 %f14708, %f2520, %f14700; mul.ftz.f32 %f14709, %f2520, %f14701; mul.ftz.f32 %f14710, %f2520, %f14702; mul.ftz.f32 %f14711, %f2520, %f14703; fma.rn.ftz.f32 %f14712, %f2530, %f14692, %f14708; fma.rn.ftz.f32 %f14713, %f2530, %f14693, %f14709; fma.rn.ftz.f32 %f14714, %f2530, %f14694, %f14710; fma.rn.ftz.f32 %f14715, %f2530, %f14695, %f14711; mul.ftz.f32 %f14716, %f29613, 0f3FB504F3; mul.ftz.f32 %f14717, %f29610, 0f3FB504F3; mul.ftz.f32 %f14718, %f29611, 0f3FB504F3; mul.ftz.f32 %f14719, %f29612, 0f3FB504F3; mul.ftz.f32 %f14720, %f14716, %f14717; mul.ftz.f32 %f14721, %f14716, %f14718; mul.ftz.f32 %f14722, %f14716, %f14719; mul.ftz.f32 %f14723, %f14717, %f14717; mul.ftz.f32 %f14724, %f14717, %f14718; mul.ftz.f32 %f14725, %f14717, %f14719; mul.ftz.f32 %f14726, %f14718, %f14718; mul.ftz.f32 %f14727, %f14718, %f14719; mul.ftz.f32 %f14728, %f14719, %f14719; sub.ftz.f32 %f14730, %f14579, %f14726; sub.ftz.f32 %f14731, %f14730, %f14728; sub.ftz.f32 %f14732, %f14724, %f14722; add.ftz.f32 %f14733, %f14721, %f14725; add.ftz.f32 %f14734, %f14722, %f14724; sub.ftz.f32 %f14735, %f14579, %f14723; sub.ftz.f32 %f14736, %f14735, %f14728; sub.ftz.f32 %f14737, %f14727, %f14720; sub.ftz.f32 %f14738, %f14725, %f14721; add.ftz.f32 %f14739, %f14720, %f14727; sub.ftz.f32 %f14740, %f14735, %f14726; mul.ftz.f32 %f14741, %f14732, %f14690; fma.rn.ftz.f32 %f14742, %f14731, %f14667, %f14741; fma.rn.ftz.f32 %f29619, %f14733, %f14713, %f14742; mul.ftz.f32 %f14743, %f14732, %f14691; fma.rn.ftz.f32 %f14744, %f14731, %f14688, %f14743; fma.rn.ftz.f32 %f2547, %f14733, %f14714, %f14744; mul.ftz.f32 %f14745, %f14732, %f14712; fma.rn.ftz.f32 %f14746, %f14731, %f14689, %f14745; fma.rn.ftz.f32 %f2548, %f14733, %f14715, %f14746; mul.ftz.f32 %f14747, %f14736, %f14690; fma.rn.ftz.f32 %f14748, %f14734, %f14667, %f14747; fma.rn.ftz.f32 %f2549, %f14737, %f14713, %f14748; mul.ftz.f32 %f14749, %f14736, %f14691; fma.rn.ftz.f32 %f14750, %f14734, %f14688, %f14749; fma.rn.ftz.f32 %f29618, %f14737, %f14714, %f14750; mul.ftz.f32 %f14751, %f14736, %f14712; fma.rn.ftz.f32 %f14752, %f14734, %f14689, %f14751; fma.rn.ftz.f32 %f2551, %f14737, %f14715, %f14752; mul.ftz.f32 %f14753, %f14739, %f14690; fma.rn.ftz.f32 %f14754, %f14738, %f14667, %f14753; fma.rn.ftz.f32 %f2552, %f14740, %f14713, %f14754; mul.ftz.f32 %f14755, %f14739, %f14691; fma.rn.ftz.f32 %f14756, %f14738, %f14688, %f14755; fma.rn.ftz.f32 %f2553, %f14740, %f14714, %f14756; mul.ftz.f32 %f14757, %f14739, %f14712; fma.rn.ftz.f32 %f14758, %f14738, %f14689, %f14757; fma.rn.ftz.f32 %f29617, %f14740, %f14715, %f14758; cvta.to.local.u64 %rd2523, %rd299; st.local.v4.f32 [%rd2523], {%f29619, %f2547, %f2548, %f2543}; cvta.to.local.u64 %rd2524, %rd311; st.local.v4.f32 [%rd2524], {%f2549, %f29618, %f2551, %f2544}; cvta.to.local.u64 %rd2525, %rd310; st.local.v4.f32 [%rd2525], {%f2552, %f2553, %f29617, %f2545}; mul.ftz.f32 %f14759, %f29618, %f29617; mul.ftz.f32 %f2555, %f2553, %f2551; sub.ftz.f32 %f29616, %f14759, %f2555; mul.ftz.f32 %f2557, %f2553, %f2548; mul.ftz.f32 %f14760, %f2547, %f29617; sub.ftz.f32 %f29615, %f2557, %f14760; mul.ftz.f32 %f2559, %f2547, %f2551; mul.ftz.f32 %f14761, %f29618, %f2548; sub.ftz.f32 %f29614, %f2559, %f14761; mul.ftz.f32 %f14762, %f2549, %f29615; fma.rn.ftz.f32 %f14763, %f29619, %f29616, %f14762; fma.rn.ftz.f32 %f29620, %f2552, %f29614, %f14763; setp.neu.ftz.f32 %p1526, %f29620, 0f00000000; @%p1526 bra $L__BB107_1723; add.ftz.f32 %f29619, %f29619, 0f322BCC77; add.ftz.f32 %f29617, %f29617, 0f322BCC77; add.ftz.f32 %f29618, %f29618, 0f322BCC77; mul.ftz.f32 %f14764, %f29618, %f29617; sub.ftz.f32 %f29616, %f14764, %f2555; mul.ftz.f32 %f14765, %f2547, %f29617; sub.ftz.f32 %f29615, %f2557, %f14765; mul.ftz.f32 %f14766, %f29618, %f2548; sub.ftz.f32 %f29614, %f2559, %f14766; mul.ftz.f32 %f14767, %f2549, %f29615; fma.rn.ftz.f32 %f14768, %f29619, %f29616, %f14767; fma.rn.ftz.f32 %f14769, %f2552, %f29614, %f14768; setp.eq.ftz.f32 %p1527, %f14769, 0f00000000; selp.f32 %f29620, 0f7F7FFFFF, %f14769, %p1527; $L__BB107_1723: mul.ftz.f32 %f14770, %f2549, %f29617; mul.ftz.f32 %f14771, %f2552, %f2551; sub.ftz.f32 %f14772, %f14771, %f14770; mul.ftz.f32 %f14773, %f2552, %f2548; mul.ftz.f32 %f14774, %f29617, %f29619; sub.ftz.f32 %f14775, %f14774, %f14773; mul.ftz.f32 %f14776, %f2551, %f29619; mul.ftz.f32 %f14777, %f2549, %f2548; sub.ftz.f32 %f14778, %f14777, %f14776; mul.ftz.f32 %f14779, %f2552, %f29618; mul.ftz.f32 %f14780, %f2549, %f2553; sub.ftz.f32 %f14781, %f14780, %f14779; mul.ftz.f32 %f14782, %f2553, %f29619; mul.ftz.f32 %f14783, %f2552, %f2547; sub.ftz.f32 %f14784, %f14783, %f14782; mul.ftz.f32 %f14785, %f2549, %f2547; mul.ftz.f32 %f14786, %f29618, %f29619; sub.ftz.f32 %f14787, %f14786, %f14785; div.approx.ftz.f32 %f14788, %f29615, %f29620; mul.ftz.f32 %f14789, %f2544, %f14788; div.approx.ftz.f32 %f14790, %f29616, %f29620; fma.rn.ftz.f32 %f14791, %f2543, %f14790, %f14789; div.approx.ftz.f32 %f14792, %f29614, %f29620; fma.rn.ftz.f32 %f14793, %f2545, %f14792, %f14791; div.approx.ftz.f32 %f14794, %f14775, %f29620; mul.ftz.f32 %f14795, %f2544, %f14794; div.approx.ftz.f32 %f14796, %f14772, %f29620; fma.rn.ftz.f32 %f14797, %f2543, %f14796, %f14795; div.approx.ftz.f32 %f14798, %f14778, %f29620; fma.rn.ftz.f32 %f14799, %f2545, %f14798, %f14797; div.approx.ftz.f32 %f14800, %f14784, %f29620; mul.ftz.f32 %f14801, %f2544, %f14800; div.approx.ftz.f32 %f14802, %f14781, %f29620; fma.rn.ftz.f32 %f14803, %f2543, %f14802, %f14801; div.approx.ftz.f32 %f14804, %f14787, %f29620; fma.rn.ftz.f32 %f14805, %f2545, %f14804, %f14803; neg.ftz.f32 %f14806, %f14793; cvta.to.local.u64 %rd2526, %rd312; st.local.v4.f32 [%rd2526], {%f14790, %f14788, %f14792, %f14806}; neg.ftz.f32 %f14807, %f14799; cvta.to.local.u64 %rd2527, %rd327; st.local.v4.f32 [%rd2527], {%f14796, %f14794, %f14798, %f14807}; neg.ftz.f32 %f14808, %f14805; cvta.to.local.u64 %rd2528, %rd326; st.local.v4.f32 [%rd2528], {%f14802, %f14800, %f14804, %f14808}; $L__BB107_1724: st.local.f32 [%rd356], %f1950; st.local.u32 [%rd291], %r5730; st.local.f32 [%rd288], %f29669; st.local.f32 [%rd357], %f1967; st.local.f32 [%rd351], %f1968; ld.global.u32 %r799, [%rd367]; st.local.u32 [%rd305], %r799; ld.global.v4.u32 {%r2998, %r2999, %r3000, %r3001}, [%rd366]; mul.wide.u32 %rd2529, %r3001, 12; add.s64 %rd2530, %rd329, %rd2529; add.s32 %r3006, %r3001, 1; mul.wide.u32 %rd2531, %r3006, 12; add.s64 %rd2532, %rd329, %rd2531; add.s32 %r3007, %r3001, 2; mul.wide.u32 %rd2533, %r3007, 12; add.s64 %rd2534, %rd329, %rd2533; mul.wide.u32 %rd2535, %r2998, 12; add.s64 %rd2536, %rd352, %rd2535; ld.global.f32 %f29648, [%rd2536]; ld.global.f32 %f29649, [%rd2536+4]; ld.global.f32 %f29650, [%rd2536+8]; mul.wide.u32 %rd2537, %r2999, 12; add.s64 %rd2538, %rd352, %rd2537; ld.global.f32 %f29651, [%rd2538]; ld.global.f32 %f29652, [%rd2538+4]; ld.global.f32 %f29653, [%rd2538+8]; mul.wide.u32 %rd2539, %r3000, 12; add.s64 %rd2540, %rd352, %rd2539; ld.global.f32 %f29654, [%rd2540]; ld.global.f32 %f29655, [%rd2540+4]; ld.global.f32 %f29656, [%rd2540+8]; mov.f32 %f14809, 0f3F800000; sub.ftz.f32 %f14810, %f14809, %f1967; sub.ftz.f32 %f2585, %f14810, %f1968; ld.global.f32 %f29657, [%rd2530]; ld.global.f32 %f29658, [%rd2530+4]; ld.global.f32 %f29659, [%rd2530+8]; ld.global.f32 %f29660, [%rd2532]; mul.ftz.f32 %f14811, %f29660, %f1967; ld.global.f32 %f29661, [%rd2532+4]; mul.ftz.f32 %f14812, %f29661, %f1967; ld.global.f32 %f29662, [%rd2532+8]; mul.ftz.f32 %f14813, %f29662, %f1967; fma.rn.ftz.f32 %f14814, %f29657, %f2585, %f14811; fma.rn.ftz.f32 %f14815, %f29658, %f2585, %f14812; fma.rn.ftz.f32 %f14816, %f29659, %f2585, %f14813; ld.global.f32 %f29663, [%rd2534]; ld.global.f32 %f29664, [%rd2534+4]; ld.global.f32 %f29665, [%rd2534+8]; fma.rn.ftz.f32 %f14817, %f29665, %f1968, %f14816; fma.rn.ftz.f32 %f14818, %f29664, %f1968, %f14815; fma.rn.ftz.f32 %f14819, %f29663, %f1968, %f14814; st.local.v2.f32 [%rd2173], {%f14819, %f14818}; st.local.f32 [%rd314], %f14817; and.b32 %r800, %r797, 4; setp.ne.s32 %p1528, %r800, 0; @%p1528 bra $L__BB107_1729; @%p1522 bra $L__BB107_1727; bra.uni $L__BB107_1726; $L__BB107_1727: ld.global.v4.f32 {%f29621, %f29622, %f29623, %f29624}, [%rd398]; ld.global.v4.f32 {%f29625, %f29626, %f29627, %f29628}, [%rd398+16]; ld.global.v4.f32 {%f29629, %f29630, %f29631, %f29632}, [%rd398+32]; bra.uni $L__BB107_1728; $L__BB107_1726: cvta.to.local.u64 %rd2542, %rd299; ld.local.v4.f32 {%f29621, %f29622, %f29623, %f29624}, [%rd2542]; cvta.to.local.u64 %rd2543, %rd311; ld.local.v4.f32 {%f29625, %f29626, %f29627, %f29628}, [%rd2543]; cvta.to.local.u64 %rd2544, %rd310; ld.local.v4.f32 {%f29629, %f29630, %f29631, %f29632}, [%rd2544]; $L__BB107_1728: ld.local.v4.f32 {%f14844, %f14845, %f14846, %f14847}, [%rd2173]; mul.ftz.f32 %f14851, %f29622, %f14845; fma.rn.ftz.f32 %f14852, %f29621, %f14844, %f14851; fma.rn.ftz.f32 %f14853, %f29623, %f14846, %f14852; mul.ftz.f32 %f14854, %f29626, %f14845; fma.rn.ftz.f32 %f14855, %f29625, %f14844, %f14854; fma.rn.ftz.f32 %f14856, %f29627, %f14846, %f14855; mul.ftz.f32 %f14857, %f29630, %f14845; fma.rn.ftz.f32 %f14858, %f29629, %f14844, %f14857; fma.rn.ftz.f32 %f14859, %f29631, %f14846, %f14858; add.ftz.f32 %f14860, %f29632, %f14859; add.ftz.f32 %f14861, %f29624, %f14853; add.ftz.f32 %f14862, %f29628, %f14856; st.local.v2.f32 [%rd2173], {%f14861, %f14862}; st.local.f32 [%rd314], %f14860; $L__BB107_1729: @%p1528 bra $L__BB107_1748; mul.wide.s32 %rd2547, %r5729, 208; add.s64 %rd401, %rd300, %rd2547; @%p1522 bra $L__BB107_1732; bra.uni $L__BB107_1731; $L__BB107_1732: ld.global.v4.f32 {%f14896, %f14897, %f14898, %f14899}, [%rd401]; ld.global.v4.f32 {%f14904, %f14905, %f14906, %f14907}, [%rd401+16]; ld.global.v4.f32 {%f14912, %f14913, %f14914, %f14915}, [%rd401+32]; mul.ftz.f32 %f14920, %f29658, %f14897; fma.rn.ftz.f32 %f14921, %f29657, %f14896, %f14920; fma.rn.ftz.f32 %f14922, %f29659, %f14898, %f14921; add.ftz.f32 %f29633, %f14899, %f14922; mul.ftz.f32 %f14923, %f29658, %f14905; fma.rn.ftz.f32 %f14924, %f29657, %f14904, %f14923; fma.rn.ftz.f32 %f14925, %f29659, %f14906, %f14924; add.ftz.f32 %f29634, %f14907, %f14925; mul.ftz.f32 %f14926, %f29658, %f14913; fma.rn.ftz.f32 %f14927, %f29657, %f14912, %f14926; fma.rn.ftz.f32 %f14928, %f29659, %f14914, %f14927; add.ftz.f32 %f29659, %f14915, %f14928; bra.uni $L__BB107_1733; $L__BB107_1731: cvta.to.local.u64 %rd2548, %rd299; ld.local.v4.f32 {%f14863, %f14864, %f14865, %f14866}, [%rd2548]; mul.ftz.f32 %f14871, %f29658, %f14864; fma.rn.ftz.f32 %f14872, %f29657, %f14863, %f14871; fma.rn.ftz.f32 %f14873, %f29659, %f14865, %f14872; add.ftz.f32 %f29633, %f14866, %f14873; cvta.to.local.u64 %rd2549, %rd311; ld.local.v4.f32 {%f14874, %f14875, %f14876, %f14877}, [%rd2549]; mul.ftz.f32 %f14882, %f29658, %f14875; fma.rn.ftz.f32 %f14883, %f29657, %f14874, %f14882; fma.rn.ftz.f32 %f14884, %f29659, %f14876, %f14883; add.ftz.f32 %f29634, %f14877, %f14884; cvta.to.local.u64 %rd2550, %rd310; ld.local.v4.f32 {%f14885, %f14886, %f14887, %f14888}, [%rd2550]; mul.ftz.f32 %f14893, %f29658, %f14886; fma.rn.ftz.f32 %f14894, %f29657, %f14885, %f14893; fma.rn.ftz.f32 %f14895, %f29659, %f14887, %f14894; add.ftz.f32 %f29659, %f14888, %f14895; $L__BB107_1733: @%p1522 bra $L__BB107_1735; bra.uni $L__BB107_1734; $L__BB107_1735: ld.global.v4.f32 {%f14962, %f14963, %f14964, %f14965}, [%rd401]; ld.global.v4.f32 {%f14970, %f14971, %f14972, %f14973}, [%rd401+16]; ld.global.v4.f32 {%f14978, %f14979, %f14980, %f14981}, [%rd401+32]; mul.ftz.f32 %f14986, %f29661, %f14963; fma.rn.ftz.f32 %f14987, %f29660, %f14962, %f14986; fma.rn.ftz.f32 %f14988, %f29662, %f14964, %f14987; add.ftz.f32 %f29636, %f14965, %f14988; mul.ftz.f32 %f14989, %f29661, %f14971; fma.rn.ftz.f32 %f14990, %f29660, %f14970, %f14989; fma.rn.ftz.f32 %f14991, %f29662, %f14972, %f14990; add.ftz.f32 %f29637, %f14973, %f14991; mul.ftz.f32 %f14992, %f29661, %f14979; fma.rn.ftz.f32 %f14993, %f29660, %f14978, %f14992; fma.rn.ftz.f32 %f14994, %f29662, %f14980, %f14993; add.ftz.f32 %f29662, %f14981, %f14994; bra.uni $L__BB107_1736; $L__BB107_1734: cvta.to.local.u64 %rd2551, %rd299; ld.local.v4.f32 {%f14929, %f14930, %f14931, %f14932}, [%rd2551]; mul.ftz.f32 %f14937, %f29661, %f14930; fma.rn.ftz.f32 %f14938, %f29660, %f14929, %f14937; fma.rn.ftz.f32 %f14939, %f29662, %f14931, %f14938; add.ftz.f32 %f29636, %f14932, %f14939; cvta.to.local.u64 %rd2552, %rd311; ld.local.v4.f32 {%f14940, %f14941, %f14942, %f14943}, [%rd2552]; mul.ftz.f32 %f14948, %f29661, %f14941; fma.rn.ftz.f32 %f14949, %f29660, %f14940, %f14948; fma.rn.ftz.f32 %f14950, %f29662, %f14942, %f14949; add.ftz.f32 %f29637, %f14943, %f14950; cvta.to.local.u64 %rd2553, %rd310; ld.local.v4.f32 {%f14951, %f14952, %f14953, %f14954}, [%rd2553]; mul.ftz.f32 %f14959, %f29661, %f14952; fma.rn.ftz.f32 %f14960, %f29660, %f14951, %f14959; fma.rn.ftz.f32 %f14961, %f29662, %f14953, %f14960; add.ftz.f32 %f29662, %f14954, %f14961; $L__BB107_1736: @%p1522 bra $L__BB107_1738; bra.uni $L__BB107_1737; $L__BB107_1738: ld.global.v4.f32 {%f15028, %f15029, %f15030, %f15031}, [%rd401]; ld.global.v4.f32 {%f15036, %f15037, %f15038, %f15039}, [%rd401+16]; ld.global.v4.f32 {%f15044, %f15045, %f15046, %f15047}, [%rd401+32]; mul.ftz.f32 %f15052, %f29664, %f15029; fma.rn.ftz.f32 %f15053, %f29663, %f15028, %f15052; fma.rn.ftz.f32 %f15054, %f29665, %f15030, %f15053; add.ftz.f32 %f29639, %f15031, %f15054; mul.ftz.f32 %f15055, %f29664, %f15037; fma.rn.ftz.f32 %f15056, %f29663, %f15036, %f15055; fma.rn.ftz.f32 %f15057, %f29665, %f15038, %f15056; add.ftz.f32 %f29640, %f15039, %f15057; mul.ftz.f32 %f15058, %f29664, %f15045; fma.rn.ftz.f32 %f15059, %f29663, %f15044, %f15058; fma.rn.ftz.f32 %f15060, %f29665, %f15046, %f15059; add.ftz.f32 %f29665, %f15047, %f15060; bra.uni $L__BB107_1739; $L__BB107_1737: cvta.to.local.u64 %rd2554, %rd299; ld.local.v4.f32 {%f14995, %f14996, %f14997, %f14998}, [%rd2554]; mul.ftz.f32 %f15003, %f29664, %f14996; fma.rn.ftz.f32 %f15004, %f29663, %f14995, %f15003; fma.rn.ftz.f32 %f15005, %f29665, %f14997, %f15004; add.ftz.f32 %f29639, %f14998, %f15005; cvta.to.local.u64 %rd2555, %rd311; ld.local.v4.f32 {%f15006, %f15007, %f15008, %f15009}, [%rd2555]; mul.ftz.f32 %f15014, %f29664, %f15007; fma.rn.ftz.f32 %f15015, %f29663, %f15006, %f15014; fma.rn.ftz.f32 %f15016, %f29665, %f15008, %f15015; add.ftz.f32 %f29640, %f15009, %f15016; cvta.to.local.u64 %rd2556, %rd310; ld.local.v4.f32 {%f15017, %f15018, %f15019, %f15020}, [%rd2556]; mul.ftz.f32 %f15025, %f29664, %f15018; fma.rn.ftz.f32 %f15026, %f29663, %f15017, %f15025; fma.rn.ftz.f32 %f15027, %f29665, %f15019, %f15026; add.ftz.f32 %f29665, %f15020, %f15027; $L__BB107_1739: @%p1522 bra $L__BB107_1741; bra.uni $L__BB107_1740; $L__BB107_1741: ld.global.v4.f32 {%f15095, %f15096, %f15097, %f15098}, [%rd401+48]; ld.global.v4.f32 {%f15102, %f15103, %f15104, %f15105}, [%rd401+64]; ld.global.v4.f32 {%f15109, %f15110, %f15111, %f15112}, [%rd401+80]; mul.ftz.f32 %f15116, %f29649, %f15102; fma.rn.ftz.f32 %f15117, %f29648, %f15095, %f15116; fma.rn.ftz.f32 %f15118, %f29650, %f15109, %f15117; mul.ftz.f32 %f15119, %f29649, %f15103; fma.rn.ftz.f32 %f15120, %f29648, %f15096, %f15119; fma.rn.ftz.f32 %f15121, %f29650, %f15110, %f15120; mul.ftz.f32 %f15122, %f29649, %f15104; fma.rn.ftz.f32 %f15123, %f29648, %f15097, %f15122; fma.rn.ftz.f32 %f15124, %f29650, %f15111, %f15123; mul.ftz.f32 %f15125, %f15121, %f15121; fma.rn.ftz.f32 %f15126, %f15118, %f15118, %f15125; fma.rn.ftz.f32 %f15127, %f15124, %f15124, %f15126; sqrt.approx.ftz.f32 %f15128, %f15127; div.approx.ftz.f32 %f29648, %f15118, %f15128; div.approx.ftz.f32 %f29649, %f15121, %f15128; div.approx.ftz.f32 %f29650, %f15124, %f15128; bra.uni $L__BB107_1742; $L__BB107_1740: cvta.to.local.u64 %rd2557, %rd312; ld.local.v4.f32 {%f15061, %f15062, %f15063, %f15064}, [%rd2557]; cvta.to.local.u64 %rd2558, %rd327; ld.local.v4.f32 {%f15068, %f15069, %f15070, %f15071}, [%rd2558]; cvta.to.local.u64 %rd2559, %rd326; ld.local.v4.f32 {%f15075, %f15076, %f15077, %f15078}, [%rd2559]; mul.ftz.f32 %f15082, %f29649, %f15068; fma.rn.ftz.f32 %f15083, %f29648, %f15061, %f15082; fma.rn.ftz.f32 %f15084, %f29650, %f15075, %f15083; mul.ftz.f32 %f15085, %f29649, %f15069; fma.rn.ftz.f32 %f15086, %f29648, %f15062, %f15085; fma.rn.ftz.f32 %f15087, %f29650, %f15076, %f15086; mul.ftz.f32 %f15088, %f29649, %f15070; fma.rn.ftz.f32 %f15089, %f29648, %f15063, %f15088; fma.rn.ftz.f32 %f15090, %f29650, %f15077, %f15089; mul.ftz.f32 %f15091, %f15087, %f15087; fma.rn.ftz.f32 %f15092, %f15084, %f15084, %f15091; fma.rn.ftz.f32 %f15093, %f15090, %f15090, %f15092; sqrt.approx.ftz.f32 %f15094, %f15093; div.approx.ftz.f32 %f29648, %f15084, %f15094; div.approx.ftz.f32 %f29649, %f15087, %f15094; div.approx.ftz.f32 %f29650, %f15090, %f15094; $L__BB107_1742: @%p1522 bra $L__BB107_1744; bra.uni $L__BB107_1743; $L__BB107_1744: ld.global.v4.f32 {%f15163, %f15164, %f15165, %f15166}, [%rd401+48]; ld.global.v4.f32 {%f15170, %f15171, %f15172, %f15173}, [%rd401+64]; ld.global.v4.f32 {%f15177, %f15178, %f15179, %f15180}, [%rd401+80]; mul.ftz.f32 %f15184, %f29652, %f15170; fma.rn.ftz.f32 %f15185, %f29651, %f15163, %f15184; fma.rn.ftz.f32 %f15186, %f29653, %f15177, %f15185; mul.ftz.f32 %f15187, %f29652, %f15171; fma.rn.ftz.f32 %f15188, %f29651, %f15164, %f15187; fma.rn.ftz.f32 %f15189, %f29653, %f15178, %f15188; mul.ftz.f32 %f15190, %f29652, %f15172; fma.rn.ftz.f32 %f15191, %f29651, %f15165, %f15190; fma.rn.ftz.f32 %f15192, %f29653, %f15179, %f15191; mul.ftz.f32 %f15193, %f15189, %f15189; fma.rn.ftz.f32 %f15194, %f15186, %f15186, %f15193; fma.rn.ftz.f32 %f15195, %f15192, %f15192, %f15194; sqrt.approx.ftz.f32 %f15196, %f15195; div.approx.ftz.f32 %f29651, %f15186, %f15196; div.approx.ftz.f32 %f29652, %f15189, %f15196; div.approx.ftz.f32 %f29653, %f15192, %f15196; bra.uni $L__BB107_1745; $L__BB107_1743: cvta.to.local.u64 %rd2560, %rd312; ld.local.v4.f32 {%f15129, %f15130, %f15131, %f15132}, [%rd2560]; cvta.to.local.u64 %rd2561, %rd327; ld.local.v4.f32 {%f15136, %f15137, %f15138, %f15139}, [%rd2561]; cvta.to.local.u64 %rd2562, %rd326; ld.local.v4.f32 {%f15143, %f15144, %f15145, %f15146}, [%rd2562]; mul.ftz.f32 %f15150, %f29652, %f15136; fma.rn.ftz.f32 %f15151, %f29651, %f15129, %f15150; fma.rn.ftz.f32 %f15152, %f29653, %f15143, %f15151; mul.ftz.f32 %f15153, %f29652, %f15137; fma.rn.ftz.f32 %f15154, %f29651, %f15130, %f15153; fma.rn.ftz.f32 %f15155, %f29653, %f15144, %f15154; mul.ftz.f32 %f15156, %f29652, %f15138; fma.rn.ftz.f32 %f15157, %f29651, %f15131, %f15156; fma.rn.ftz.f32 %f15158, %f29653, %f15145, %f15157; mul.ftz.f32 %f15159, %f15155, %f15155; fma.rn.ftz.f32 %f15160, %f15152, %f15152, %f15159; fma.rn.ftz.f32 %f15161, %f15158, %f15158, %f15160; sqrt.approx.ftz.f32 %f15162, %f15161; div.approx.ftz.f32 %f29651, %f15152, %f15162; div.approx.ftz.f32 %f29652, %f15155, %f15162; div.approx.ftz.f32 %f29653, %f15158, %f15162; $L__BB107_1745: @%p1522 bra $L__BB107_1747; bra.uni $L__BB107_1746; $L__BB107_1747: ld.global.v4.f32 {%f15231, %f15232, %f15233, %f15234}, [%rd401+48]; ld.global.v4.f32 {%f15238, %f15239, %f15240, %f15241}, [%rd401+64]; ld.global.v4.f32 {%f15245, %f15246, %f15247, %f15248}, [%rd401+80]; mul.ftz.f32 %f15252, %f29655, %f15238; fma.rn.ftz.f32 %f15253, %f29654, %f15231, %f15252; fma.rn.ftz.f32 %f15254, %f29656, %f15245, %f15253; mul.ftz.f32 %f15255, %f29655, %f15239; fma.rn.ftz.f32 %f15256, %f29654, %f15232, %f15255; fma.rn.ftz.f32 %f15257, %f29656, %f15246, %f15256; mul.ftz.f32 %f15258, %f29655, %f15240; fma.rn.ftz.f32 %f15259, %f29654, %f15233, %f15258; fma.rn.ftz.f32 %f15260, %f29656, %f15247, %f15259; mul.ftz.f32 %f15261, %f15257, %f15257; fma.rn.ftz.f32 %f15262, %f15254, %f15254, %f15261; fma.rn.ftz.f32 %f15263, %f15260, %f15260, %f15262; sqrt.approx.ftz.f32 %f15264, %f15263; div.approx.ftz.f32 %f29654, %f15254, %f15264; div.approx.ftz.f32 %f29655, %f15257, %f15264; div.approx.ftz.f32 %f29656, %f15260, %f15264; mov.f32 %f29657, %f29633; mov.f32 %f29658, %f29634; mov.f32 %f29660, %f29636; mov.f32 %f29661, %f29637; mov.f32 %f29663, %f29639; mov.f32 %f29664, %f29640; bra.uni $L__BB107_1748; $L__BB107_1746: cvta.to.local.u64 %rd2563, %rd312; ld.local.v4.f32 {%f15197, %f15198, %f15199, %f15200}, [%rd2563]; cvta.to.local.u64 %rd2564, %rd327; ld.local.v4.f32 {%f15204, %f15205, %f15206, %f15207}, [%rd2564]; cvta.to.local.u64 %rd2565, %rd326; ld.local.v4.f32 {%f15211, %f15212, %f15213, %f15214}, [%rd2565]; mul.ftz.f32 %f15218, %f29655, %f15204; fma.rn.ftz.f32 %f15219, %f29654, %f15197, %f15218; fma.rn.ftz.f32 %f15220, %f29656, %f15211, %f15219; mul.ftz.f32 %f15221, %f29655, %f15205; fma.rn.ftz.f32 %f15222, %f29654, %f15198, %f15221; fma.rn.ftz.f32 %f15223, %f29656, %f15212, %f15222; mul.ftz.f32 %f15224, %f29655, %f15206; fma.rn.ftz.f32 %f15225, %f29654, %f15199, %f15224; fma.rn.ftz.f32 %f15226, %f29656, %f15213, %f15225; mul.ftz.f32 %f15227, %f15223, %f15223; fma.rn.ftz.f32 %f15228, %f15220, %f15220, %f15227; fma.rn.ftz.f32 %f15229, %f15226, %f15226, %f15228; sqrt.approx.ftz.f32 %f15230, %f15229; div.approx.ftz.f32 %f29654, %f15220, %f15230; div.approx.ftz.f32 %f29655, %f15223, %f15230; div.approx.ftz.f32 %f29656, %f15226, %f15230; mov.f32 %f29657, %f29633; mov.f32 %f29658, %f29634; mov.f32 %f29660, %f29636; mov.f32 %f29661, %f29637; mov.f32 %f29663, %f29639; mov.f32 %f29664, %f29640; $L__BB107_1748: sub.ftz.f32 %f2700, %f29665, %f29659; sub.ftz.f32 %f2701, %f29661, %f29658; mul.ftz.f32 %f15265, %f2701, %f2700; sub.ftz.f32 %f2702, %f29664, %f29658; sub.ftz.f32 %f2703, %f29662, %f29659; mul.ftz.f32 %f15266, %f2703, %f2702; sub.ftz.f32 %f15267, %f15265, %f15266; sub.ftz.f32 %f2704, %f29663, %f29657; mul.ftz.f32 %f15268, %f2703, %f2704; sub.ftz.f32 %f2705, %f29660, %f29657; mul.ftz.f32 %f15269, %f2705, %f2700; sub.ftz.f32 %f15270, %f15268, %f15269; mul.ftz.f32 %f15271, %f2705, %f2702; mul.ftz.f32 %f15272, %f2701, %f2704; sub.ftz.f32 %f15273, %f15271, %f15272; mul.ftz.f32 %f15274, %f15270, %f15270; fma.rn.ftz.f32 %f15275, %f15267, %f15267, %f15274; fma.rn.ftz.f32 %f15276, %f15273, %f15273, %f15275; sqrt.approx.ftz.f32 %f15277, %f15276; div.approx.ftz.f32 %f29668, %f15273, %f15277; div.approx.ftz.f32 %f29667, %f15270, %f15277; div.approx.ftz.f32 %f29666, %f15267, %f15277; mul.lo.s64 %rd2566, %rd385, 192; add.s64 %rd2567, %rd1, %rd2566; add.s64 %rd402, %rd2567, 48; st.local.v2.f32 [%rd2567+48], {%f29666, %f29667}; st.local.f32 [%rd2567+56], %f29668; and.b32 %r3008, %r797, 8; setp.eq.s32 %p1537, %r3008, 0; @%p1537 bra $L__BB107_1750; mov.b32 %r3009, %f29666; mov.b64 %rd2568, {%r3009, %r3010}; cvt.u32.u64 %r3011, %rd2568; mov.b32 %f15278, %r3011; ld.local.f32 %f15279, [%rd402+4]; neg.ftz.f32 %f29668, %f29668; neg.ftz.f32 %f29667, %f15279; neg.ftz.f32 %f29666, %f15278; st.local.v2.f32 [%rd402], {%f29666, %f29667}; st.local.f32 [%rd402+8], %f29668; $L__BB107_1750: mul.ftz.f32 %f15280, %f29651, %f1967; fma.rn.ftz.f32 %f15281, %f29648, %f2585, %f15280; mul.ftz.f32 %f15282, %f29652, %f1967; fma.rn.ftz.f32 %f15283, %f29649, %f2585, %f15282; mul.ftz.f32 %f15284, %f29653, %f1967; fma.rn.ftz.f32 %f15285, %f29650, %f2585, %f15284; fma.rn.ftz.f32 %f15286, %f29654, %f1968, %f15281; fma.rn.ftz.f32 %f15287, %f29655, %f1968, %f15283; fma.rn.ftz.f32 %f15288, %f29656, %f1968, %f15285; mul.ftz.f32 %f15289, %f15287, %f15287; fma.rn.ftz.f32 %f15290, %f15286, %f15286, %f15289; fma.rn.ftz.f32 %f15291, %f15288, %f15288, %f15290; rsqrt.approx.ftz.f32 %f15292, %f15291; sub.ftz.f32 %f15293, %f29651, %f29648; mul.ftz.f32 %f15294, %f15293, %f15292; sub.ftz.f32 %f15295, %f29652, %f29649; mul.ftz.f32 %f15296, %f15295, %f15292; sub.ftz.f32 %f15297, %f29653, %f29650; mul.ftz.f32 %f15298, %f15297, %f15292; sub.ftz.f32 %f15299, %f29654, %f29648; mul.ftz.f32 %f15300, %f15299, %f15292; sub.ftz.f32 %f15301, %f29655, %f29649; mul.ftz.f32 %f15302, %f15301, %f15292; sub.ftz.f32 %f15303, %f29656, %f29650; mul.ftz.f32 %f15304, %f15303, %f15292; mul.ftz.f32 %f15305, %f15287, %f15292; mul.ftz.f32 %f15306, %f15296, %f15305; mul.ftz.f32 %f15307, %f15286, %f15292; fma.rn.ftz.f32 %f15308, %f15294, %f15307, %f15306; mul.ftz.f32 %f15309, %f15288, %f15292; fma.rn.ftz.f32 %f15310, %f15298, %f15309, %f15308; mul.ftz.f32 %f15311, %f15307, %f15310; mul.ftz.f32 %f15312, %f15305, %f15310; mul.ftz.f32 %f15313, %f15310, %f15309; sub.ftz.f32 %f15314, %f15294, %f15311; sub.ftz.f32 %f15315, %f15296, %f15312; sub.ftz.f32 %f15316, %f15298, %f15313; mul.ftz.f32 %f15317, %f15302, %f15305; fma.rn.ftz.f32 %f15318, %f15300, %f15307, %f15317; fma.rn.ftz.f32 %f15319, %f15304, %f15309, %f15318; mul.ftz.f32 %f15320, %f15307, %f15319; mul.ftz.f32 %f15321, %f15305, %f15319; mul.ftz.f32 %f15322, %f15319, %f15309; sub.ftz.f32 %f15323, %f15300, %f15320; sub.ftz.f32 %f15324, %f15302, %f15321; sub.ftz.f32 %f15325, %f15304, %f15322; mul.ftz.f32 %f15326, %f2701, %f2701; fma.rn.ftz.f32 %f15327, %f2705, %f2705, %f15326; fma.rn.ftz.f32 %f15328, %f2703, %f2703, %f15327; rsqrt.approx.ftz.f32 %f15329, %f15328; mul.ftz.f32 %f15330, %f2705, %f15329; mul.ftz.f32 %f15331, %f2701, %f15329; mul.ftz.f32 %f15332, %f2703, %f15329; mul.ftz.f32 %f15333, %f15314, %f15329; mul.ftz.f32 %f15334, %f15315, %f15329; mul.ftz.f32 %f15335, %f15316, %f15329; mul.ftz.f32 %f15336, %f2702, %f15331; fma.rn.ftz.f32 %f15337, %f2704, %f15330, %f15336; fma.rn.ftz.f32 %f15338, %f2700, %f15332, %f15337; mul.ftz.f32 %f15339, %f15330, %f15338; mul.ftz.f32 %f15340, %f15331, %f15338; mul.ftz.f32 %f15341, %f15332, %f15338; sub.ftz.f32 %f15342, %f2704, %f15339; sub.ftz.f32 %f15343, %f2702, %f15340; sub.ftz.f32 %f15344, %f2700, %f15341; mul.ftz.f32 %f15345, %f15333, %f15338; mul.ftz.f32 %f15346, %f15334, %f15338; mul.ftz.f32 %f15347, %f15335, %f15338; sub.ftz.f32 %f15348, %f15323, %f15345; sub.ftz.f32 %f15349, %f15324, %f15346; sub.ftz.f32 %f15350, %f15325, %f15347; mul.ftz.f32 %f15351, %f15343, %f15343; fma.rn.ftz.f32 %f15352, %f15342, %f15342, %f15351; fma.rn.ftz.f32 %f15353, %f15344, %f15344, %f15352; rsqrt.approx.ftz.f32 %f15354, %f15353; mul.ftz.f32 %f15355, %f15354, %f15342; mul.ftz.f32 %f15356, %f15354, %f15343; mul.ftz.f32 %f15357, %f15354, %f15344; mul.ftz.f32 %f15358, %f15354, %f15348; mul.ftz.f32 %f15359, %f15354, %f15349; mul.ftz.f32 %f15360, %f15354, %f15350; setp.neu.ftz.f32 %p1538, %f29666, %f29667; setp.neu.ftz.f32 %p1539, %f29666, %f29668; or.pred %p1540, %p1538, %p1539; sub.ftz.f32 %f15361, %f29668, %f29667; neg.ftz.f32 %f15362, %f29667; neg.ftz.f32 %f15363, %f29668; selp.f32 %f15364, %f15363, %f29668, %p1540; add.ftz.f32 %f15365, %f29666, %f15364; selp.f32 %f15366, %f29667, %f15362, %p1540; sub.ftz.f32 %f15367, %f15366, %f29666; mul.ftz.f32 %f15368, %f15365, %f15365; fma.rn.ftz.f32 %f15369, %f15361, %f15361, %f15368; fma.rn.ftz.f32 %f15370, %f15367, %f15367, %f15369; sqrt.approx.ftz.f32 %f15371, %f15370; div.approx.ftz.f32 %f15372, %f15367, %f15371; mul.ftz.f32 %f15373, %f29667, %f15372; div.approx.ftz.f32 %f15374, %f15365, %f15371; mul.ftz.f32 %f15375, %f29668, %f15374; div.approx.ftz.f32 %f15376, %f15361, %f15371; mul.ftz.f32 %f15377, %f29668, %f15376; mul.ftz.f32 %f15378, %f29666, %f15372; mul.ftz.f32 %f15379, %f29666, %f15374; mul.ftz.f32 %f15380, %f29667, %f15376; sub.ftz.f32 %f15381, %f15377, %f15378; sub.ftz.f32 %f15382, %f15373, %f15375; st.local.v4.f32 [%rd402+-32], {%f15374, %f15372, %f15382, %f15381}; sub.ftz.f32 %f15383, %f15379, %f15380; st.local.v4.f32 [%rd402+-16], {%f15383, %f15307, %f15305, %f15309}; mul.ftz.f32 %f15384, %f15331, %f15374; fma.rn.ftz.f32 %f15385, %f15330, %f15376, %f15384; fma.rn.ftz.f32 %f15386, %f15332, %f15372, %f15385; mul.ftz.f32 %f15387, %f15356, %f15374; fma.rn.ftz.f32 %f15388, %f15355, %f15376, %f15387; fma.rn.ftz.f32 %f15389, %f15357, %f15372, %f15388; mul.ftz.f32 %f15390, %f15358, %f15389; mul.ftz.f32 %f15391, %f15359, %f15389; mul.ftz.f32 %f15392, %f15360, %f15389; fma.rn.ftz.f32 %f15393, %f15333, %f15386, %f15390; fma.rn.ftz.f32 %f15394, %f15334, %f15386, %f15391; fma.rn.ftz.f32 %f15395, %f15335, %f15386, %f15392; st.local.f32 [%rd402+12], %f15393; st.local.f32 [%rd402+16], %f15394; st.local.f32 [%rd402+20], %f15395; mul.ftz.f32 %f15396, %f15333, %f15389; mul.ftz.f32 %f15397, %f15334, %f15389; mul.ftz.f32 %f15398, %f15335, %f15389; mul.ftz.f32 %f15399, %f15358, %f15386; mul.ftz.f32 %f15400, %f15359, %f15386; mul.ftz.f32 %f15401, %f15360, %f15386; sub.ftz.f32 %f15402, %f15401, %f15398; sub.ftz.f32 %f15403, %f15400, %f15397; sub.ftz.f32 %f15404, %f15399, %f15396; st.local.v2.f32 [%rd402+24], {%f15404, %f15403}; st.local.f32 [%rd402+32], %f15402; ld.local.v4.f32 {%f15405, %f15406, %f15407, %f15408}, [%rd2173]; st.local.v4.f32 [%rd402+-48], {%f15405, %f15406, %f15407, %f15376}; mov.f32 %f15412, 0f00000000; st.local.v4.f32 [%rd402+128], {%f15412, %f15412, %f15412, %f15412}; st.local.v4.f32 [%rd402+96], {%f15412, %f15412, %f15412, %f15412}; st.local.v4.f32 [%rd402+112], {%f14809, %f15412, %f15412, %f14809}; st.local.v2.f32 [%rd402+80], {%f15412, %f15412}; st.local.v2.f32 [%rd402+72], {%f29608, %f29609}; mul.lo.s64 %rd2571, %rd392, 80; add.s64 %rd2572, %rd2172, %rd2571; add.s64 %rd2573, %rd2572, 304; st.local.u64 [%rd402+64], %rd2573; st.local.v2.f32 [%rd402+40], {%f1967, %f1968}; mov.b32 %r3012, %f29598; st.local.v4.u32 [%rd402+48], {%r5729, %r5730, %r799, %r3012}; mov.u32 %r5732, %r677; $L__BB107_1752: st.local.u32 [%rd285], %r632; st.local.u32 [%rd285+-4], %r5730; mov.b32 %r3013, %f29669; add.s32 %r3014, %r3013, 1; mov.b32 %f15414, %r3014; setp.eq.ftz.f32 %p1541, %f29669, 0f00000000; selp.f32 %f15415, 0f00000001, %f15414, %p1541; st.local.f32 [%rd285+-16], %f15415; add.s32 %r5710, %r5710, 1; setp.lt.u32 %p1542, %r5710, 10; @%p1542 bra $L__BB107_1331; $L__BB107_1753: ld.const.u64 %rd2574, [kernel_params+2552]; cvta.to.global.u64 %rd2575, %rd2574; add.s64 %rd403, %rd2575, %rd3; ld.global.u8 %rs49, [%rd403]; and.b16 %rs347, %rs49, 254; st.global.u8 [%rd403], %rs347; setp.eq.s32 %p1543, %r5732, 0; mov.u32 %r5802, 0; @%p1543 bra $L__BB107_3392; ld.const.u64 %rd2576, [kernel_params+2496]; cvta.to.global.u64 %rd2577, %rd2576; add.s64 %rd404, %rd2577, %rd2151; ld.global.u16 %r3017, [%rd404]; add.s32 %r3018, %r5732, %r3017; ld.const.u32 %r3019, [kernel_params+1528]; setp.gt.s32 %p1544, %r3018, %r3019; @%p1544 bra $L__BB107_3392; ld.const.u64 %rd2579, [kernel_params+2480]; cvta.to.global.u64 %rd2580, %rd2579; add.s64 %rd405, %rd2580, %rd2151; ld.global.u16 %r3021, [%rd405]; add.s32 %r3022, %r3021, 1; ld.const.u32 %r3023, [kernel_params+1520]; setp.ge.s32 %p1545, %r3022, %r3023; @%p1545 bra $L__BB107_3392; ld.global.u16 %r3025, [%rd278]; add.s32 %r3026, %r5732, %r3025; ld.const.u32 %r3027, [kernel_params+1516]; setp.ge.s32 %p1546, %r3026, %r3027; @%p1546 bra $L__BB107_3392; or.b16 %rs348, %rs49, 1; st.global.u8 [%rd403], %rs348; add.u64 %rd2582, %SP, 1344; add.u64 %rd2583, %SPL, 1344; add.s64 %rd406, %rd2583, 48; mov.u32 %r3028, -1; st.local.u32 [%rd2583+48], %r3028; st.local.u32 [%rd2583+44], %r3028; ld.local.f32 %f15420, [%rd5+4]; st.local.f32 [%rd2583+52], %f15420; mov.u32 %r5733, 0; st.local.u32 [%rd2583+56], %r5733; st.local.u32 [%rd2583+24], %r5733; ld.local.f32 %f2730, [%rd5+-4]; st.local.f32 [%rd2583+32], %f2730; @%p1261 bra $L__BB107_1759; bra.uni $L__BB107_1758; $L__BB107_1759: mov.f32 %f29680, %f30148; mov.f32 %f29681, %f30147; mov.f32 %f29682, %f30146; bra.uni $L__BB107_1760; $L__BB107_711: mov.f32 %f9297, 0fBF800000; max.ftz.f32 %f9298, %f780, %f9297; min.ftz.f32 %f9300, %f9298, %f9296; abs.ftz.f32 %f9301, %f9300; sub.ftz.f32 %f9302, %f9296, %f9301; mul.ftz.f32 %f9303, %f9302, 0f3F000000; sqrt.approx.ftz.f32 %f9304, %f9303; setp.gt.ftz.f32 %p617, %f9301, 0f3F11EB85; selp.f32 %f9305, %f9304, %f9301, %p617; mul.ftz.f32 %f9306, %f9305, %f9305; mov.f32 %f9307, 0f3C94D2E9; mov.f32 %f9308, 0f3D53F941; fma.rn.ftz.f32 %f9309, %f9308, %f9306, %f9307; mov.f32 %f9310, 0f3D3F841F; fma.rn.ftz.f32 %f9311, %f9309, %f9306, %f9310; mov.f32 %f9312, 0f3D994929; fma.rn.ftz.f32 %f9313, %f9311, %f9306, %f9312; mov.f32 %f9314, 0f3E2AAB94; fma.rn.ftz.f32 %f9315, %f9313, %f9306, %f9314; mul.ftz.f32 %f9316, %f9306, %f9315; fma.rn.ftz.f32 %f9317, %f9316, %f9305, %f9305; add.ftz.f32 %f9318, %f9317, %f9317; mov.f32 %f9319, 0f3FC90FDB; sub.ftz.f32 %f9320, %f9319, %f9317; selp.f32 %f9321, %f9318, %f9320, %p617; setp.lt.ftz.f32 %p618, %f9300, 0f00000000; mov.f32 %f9322, 0f40490FDB; sub.ftz.f32 %f9323, %f9322, %f9321; selp.f32 %f9324, %f9323, %f9321, %p618; mul.ftz.f32 %f9325, %f9284, %f780; sub.ftz.f32 %f9326, %f9288, %f9325; mul.ftz.f32 %f9327, %f9285, %f780; sub.ftz.f32 %f9328, %f9289, %f9327; mul.ftz.f32 %f9329, %f9286, %f780; sub.ftz.f32 %f9330, %f9290, %f9329; mul.ftz.f32 %f9331, %f9287, %f780; sub.ftz.f32 %f9332, %f9291, %f9331; mul.ftz.f32 %f9333, %f9328, %f9328; fma.rn.ftz.f32 %f9334, %f9326, %f9326, %f9333; mul.ftz.f32 %f9335, %f9332, %f9332; fma.rn.ftz.f32 %f9336, %f9330, %f9330, %f9335; add.ftz.f32 %f9337, %f9334, %f9336; rsqrt.approx.ftz.f32 %f9338, %f9337; mul.ftz.f32 %f9339, %f9326, %f9338; mul.ftz.f32 %f9340, %f9328, %f9338; mul.ftz.f32 %f9341, %f9330, %f9338; mul.ftz.f32 %f9342, %f9332, %f9338; mul.ftz.f32 %f9343, %f771, %f9324; cos.approx.ftz.f32 %f9344, %f9343; sin.approx.ftz.f32 %f9345, %f9343; mul.ftz.f32 %f9346, %f9339, %f9345; mul.ftz.f32 %f9347, %f9340, %f9345; mul.ftz.f32 %f9348, %f9341, %f9345; mul.ftz.f32 %f9349, %f9342, %f9345; fma.rn.ftz.f32 %f29053, %f9287, %f9344, %f9349; fma.rn.ftz.f32 %f29052, %f9286, %f9344, %f9348; fma.rn.ftz.f32 %f29051, %f9285, %f9344, %f9347; fma.rn.ftz.f32 %f29050, %f9284, %f9344, %f9346; $L__BB107_713: ld.global.v4.f32 {%f9364, %f9365, %f9366, %f9367}, [%rd158+16]; ld.global.v4.f32 {%f9372, %f9373, %f9374, %f9375}, [%rd158+80]; mul.ftz.f32 %f9380, %f771, %f9372; mul.ftz.f32 %f9381, %f771, %f9373; mul.ftz.f32 %f9382, %f771, %f9374; mul.ftz.f32 %f9383, %f771, %f9375; fma.rn.ftz.f32 %f9384, %f781, %f9367, %f9383; ld.global.v4.f32 {%f9385, %f9386, %f9387, %f9388}, [%rd158+32]; ld.global.v4.f32 {%f9393, %f9394, %f9395, %f9396}, [%rd158+96]; mul.ftz.f32 %f9401, %f771, %f9393; mul.ftz.f32 %f9402, %f771, %f9394; mul.ftz.f32 %f9403, %f771, %f9395; mul.ftz.f32 %f9404, %f771, %f9396; fma.rn.ftz.f32 %f9405, %f781, %f9385, %f9401; fma.rn.ftz.f32 %f9406, %f781, %f9386, %f9402; fma.rn.ftz.f32 %f9407, %f781, %f9387, %f9403; fma.rn.ftz.f32 %f9408, %f781, %f9388, %f9404; ld.global.v4.f32 {%f9409, %f9410, %f9411, %f9412}, [%rd158+48]; ld.global.v4.f32 {%f9417, %f9418, %f9419, %f9420}, [%rd158+112]; mul.ftz.f32 %f9425, %f771, %f9417; mul.ftz.f32 %f9426, %f771, %f9418; mul.ftz.f32 %f9427, %f771, %f9419; mul.ftz.f32 %f9428, %f771, %f9420; fma.rn.ftz.f32 %f9429, %f781, %f9409, %f9425; fma.rn.ftz.f32 %f9430, %f781, %f9410, %f9426; fma.rn.ftz.f32 %f9431, %f781, %f9411, %f9427; fma.rn.ftz.f32 %f9432, %f781, %f9412, %f9428; mul.ftz.f32 %f9433, %f29053, 0f3FB504F3; mul.ftz.f32 %f9434, %f29050, 0f3FB504F3; mul.ftz.f32 %f9435, %f29051, 0f3FB504F3; mul.ftz.f32 %f9436, %f29052, 0f3FB504F3; mul.ftz.f32 %f9437, %f9433, %f9434; mul.ftz.f32 %f9438, %f9433, %f9435; mul.ftz.f32 %f9439, %f9433, %f9436; mul.ftz.f32 %f9440, %f9434, %f9434; mul.ftz.f32 %f9441, %f9434, %f9435; mul.ftz.f32 %f9442, %f9434, %f9436; mul.ftz.f32 %f9443, %f9435, %f9435; mul.ftz.f32 %f9444, %f9435, %f9436; mul.ftz.f32 %f9445, %f9436, %f9436; sub.ftz.f32 %f9447, %f9296, %f9443; sub.ftz.f32 %f9448, %f9447, %f9445; sub.ftz.f32 %f9449, %f9441, %f9439; add.ftz.f32 %f9450, %f9438, %f9442; add.ftz.f32 %f9451, %f9439, %f9441; sub.ftz.f32 %f9452, %f9296, %f9440; sub.ftz.f32 %f9453, %f9452, %f9445; sub.ftz.f32 %f9454, %f9444, %f9437; sub.ftz.f32 %f9455, %f9442, %f9438; add.ftz.f32 %f9456, %f9437, %f9444; sub.ftz.f32 %f9457, %f9452, %f9443; mul.ftz.f32 %f9458, %f9449, %f9407; fma.rn.ftz.f32 %f9459, %f9448, %f9384, %f9458; mul.ftz.f32 %f9460, %f9449, %f9408; fma.rn.ftz.f32 %f9461, %f9448, %f9405, %f9460; mul.ftz.f32 %f9462, %f9449, %f9429; fma.rn.ftz.f32 %f9463, %f9448, %f9406, %f9462; mul.ftz.f32 %f9464, %f9453, %f9407; fma.rn.ftz.f32 %f9465, %f9451, %f9384, %f9464; mul.ftz.f32 %f9466, %f9453, %f9408; fma.rn.ftz.f32 %f9467, %f9451, %f9405, %f9466; mul.ftz.f32 %f9468, %f9453, %f9429; fma.rn.ftz.f32 %f9469, %f9451, %f9406, %f9468; mul.ftz.f32 %f9470, %f9456, %f9407; fma.rn.ftz.f32 %f9471, %f9455, %f9384, %f9470; mul.ftz.f32 %f9472, %f9456, %f9408; fma.rn.ftz.f32 %f9473, %f9455, %f9405, %f9472; mul.ftz.f32 %f9474, %f9456, %f9429; fma.rn.ftz.f32 %f9475, %f9455, %f9406, %f9474; fma.rn.ftz.f32 %f29060, %f9454, %f9432, %f9469; fma.rn.ftz.f32 %f29059, %f9454, %f9431, %f9467; fma.rn.ftz.f32 %f29058, %f9454, %f9430, %f9465; fma.rn.ftz.f32 %f29064, %f9457, %f9432, %f9475; fma.rn.ftz.f32 %f29063, %f9457, %f9431, %f9473; fma.rn.ftz.f32 %f29062, %f9457, %f9430, %f9471; fma.rn.ftz.f32 %f29056, %f9450, %f9432, %f9463; fma.rn.ftz.f32 %f29055, %f9450, %f9431, %f9461; fma.rn.ftz.f32 %f29054, %f9450, %f9430, %f9459; fma.rn.ftz.f32 %f29065, %f781, %f9366, %f9382; fma.rn.ftz.f32 %f29061, %f781, %f9365, %f9381; fma.rn.ftz.f32 %f29057, %f781, %f9364, %f9380; $L__BB107_715: mul.ftz.f32 %f9488, %f29067, %f29055; fma.rn.ftz.f32 %f9489, %f29066, %f29054, %f9488; fma.rn.ftz.f32 %f9490, %f29068, %f29056, %f9489; add.ftz.f32 %f830, %f29057, %f9490; mul.ftz.f32 %f9491, %f29067, %f29059; fma.rn.ftz.f32 %f9492, %f29066, %f29058, %f9491; fma.rn.ftz.f32 %f9493, %f29068, %f29060, %f9492; add.ftz.f32 %f831, %f29061, %f9493; mul.ftz.f32 %f9494, %f29067, %f29063; fma.rn.ftz.f32 %f9495, %f29066, %f29062, %f9494; fma.rn.ftz.f32 %f9496, %f29068, %f29064, %f9495; add.ftz.f32 %f29068, %f29065, %f9496; mul.ftz.f32 %f9497, %f29070, %f29055; fma.rn.ftz.f32 %f9498, %f29069, %f29054, %f9497; fma.rn.ftz.f32 %f9499, %f29071, %f29056, %f9498; add.ftz.f32 %f833, %f29057, %f9499; mul.ftz.f32 %f9500, %f29070, %f29059; fma.rn.ftz.f32 %f9501, %f29069, %f29058, %f9500; fma.rn.ftz.f32 %f9502, %f29071, %f29060, %f9501; add.ftz.f32 %f834, %f29061, %f9502; mul.ftz.f32 %f9503, %f29070, %f29063; fma.rn.ftz.f32 %f9504, %f29069, %f29062, %f9503; fma.rn.ftz.f32 %f9505, %f29071, %f29064, %f9504; add.ftz.f32 %f29071, %f29065, %f9505; mul.ftz.f32 %f9506, %f29073, %f29055; fma.rn.ftz.f32 %f9507, %f29072, %f29054, %f9506; fma.rn.ftz.f32 %f9508, %f29074, %f29056, %f9507; add.ftz.f32 %f836, %f29057, %f9508; mul.ftz.f32 %f9509, %f29073, %f29059; fma.rn.ftz.f32 %f9510, %f29072, %f29058, %f9509; fma.rn.ftz.f32 %f9511, %f29074, %f29060, %f9510; add.ftz.f32 %f837, %f29061, %f9511; mul.ftz.f32 %f9512, %f29073, %f29063; fma.rn.ftz.f32 %f9513, %f29072, %f29062, %f9512; fma.rn.ftz.f32 %f9514, %f29074, %f29064, %f9513; add.ftz.f32 %f29074, %f29065, %f9514; mov.u16 %rs615, 1; mov.f32 %f29066, %f830; mov.f32 %f29067, %f831; mov.f32 %f29069, %f833; mov.f32 %f29070, %f834; mov.f32 %f29072, %f836; mov.f32 %f29073, %f837; $L__BB107_716: sub.ftz.f32 %f9515, %f29069, %f29066; sub.ftz.f32 %f9516, %f29070, %f29067; mul.ftz.f32 %f9517, %f9516, %f9516; fma.rn.ftz.f32 %f9518, %f9515, %f9515, %f9517; sub.ftz.f32 %f9519, %f29071, %f29068; fma.rn.ftz.f32 %f9520, %f9519, %f9519, %f9518; sub.ftz.f32 %f9521, %f29072, %f29066; sub.ftz.f32 %f9522, %f29073, %f29067; mul.ftz.f32 %f9523, %f9522, %f9522; fma.rn.ftz.f32 %f9524, %f9521, %f9521, %f9523; sub.ftz.f32 %f9525, %f29074, %f29068; fma.rn.ftz.f32 %f9526, %f9525, %f9525, %f9524; sub.ftz.f32 %f9527, %f29072, %f29069; sub.ftz.f32 %f9528, %f29073, %f29070; mul.ftz.f32 %f9529, %f9528, %f9528; fma.rn.ftz.f32 %f9530, %f9527, %f9527, %f9529; sub.ftz.f32 %f9531, %f29074, %f29071; fma.rn.ftz.f32 %f9532, %f9531, %f9531, %f9530; max.ftz.f32 %f9533, %f9526, %f9532; max.ftz.f32 %f9534, %f9520, %f9533; mul.ftz.f32 %f9535, %f9516, %f9525; mul.ftz.f32 %f9536, %f9519, %f9522; sub.ftz.f32 %f9537, %f9535, %f9536; mul.ftz.f32 %f9538, %f9519, %f9521; mul.ftz.f32 %f9539, %f9515, %f9525; sub.ftz.f32 %f9540, %f9538, %f9539; mul.ftz.f32 %f9541, %f9515, %f9522; mul.ftz.f32 %f9542, %f9516, %f9521; sub.ftz.f32 %f9543, %f9541, %f9542; ld.local.f32 %f848, [%rd5+-52]; mul.ftz.f32 %f849, %f732, %f848; ld.local.v2.f32 {%f9544, %f9545}, [%rd5+-48]; mul.ftz.f32 %f852, %f732, %f9544; mul.ftz.f32 %f853, %f732, %f9545; mul.ftz.f32 %f9546, %f9540, %f852; fma.rn.ftz.f32 %f9547, %f9537, %f849, %f9546; fma.rn.ftz.f32 %f9548, %f9543, %f853, %f9547; abs.ftz.f32 %f9549, %f9548; mul.ftz.f32 %f9550, %f9540, %f9540; fma.rn.ftz.f32 %f9551, %f9537, %f9537, %f9550; fma.rn.ftz.f32 %f854, %f9543, %f9543, %f9551; div.approx.ftz.f32 %f9552, %f9549, %f854; mul.ftz.f32 %f9553, %f9552, %f9552; setp.gt.ftz.f32 %p619, %f9534, %f9553; @%p619 bra $L__BB107_730; bra.uni $L__BB107_717; $L__BB107_730: cvta.to.local.u64 %rd1644, %rd1061; ld.local.v4.f32 {%f9819, %f9820, %f9821, %f9822}, [%rd1644]; add.ftz.f32 %f9826, %f849, %f9819; add.ftz.f32 %f9827, %f852, %f9820; add.ftz.f32 %f9828, %f853, %f9821; sub.ftz.f32 %f946, %f29066, %f9826; sub.ftz.f32 %f947, %f29067, %f9827; sub.ftz.f32 %f948, %f29068, %f9828; sub.ftz.f32 %f949, %f29069, %f9826; sub.ftz.f32 %f950, %f29070, %f9827; sub.ftz.f32 %f951, %f29071, %f9828; sub.ftz.f32 %f952, %f29072, %f9826; sub.ftz.f32 %f953, %f29073, %f9827; sub.ftz.f32 %f954, %f29074, %f9828; mul.ftz.f32 %f9829, %f947, %f951; mul.ftz.f32 %f9830, %f950, %f948; sub.ftz.f32 %f29101, %f9829, %f9830; mul.ftz.f32 %f9831, %f949, %f948; mul.ftz.f32 %f9832, %f946, %f951; sub.ftz.f32 %f29102, %f9831, %f9832; mul.ftz.f32 %f9833, %f946, %f950; mul.ftz.f32 %f9834, %f949, %f947; sub.ftz.f32 %f29103, %f9833, %f9834; mul.ftz.f32 %f9835, %f29102, %f29102; fma.rn.ftz.f32 %f9836, %f29101, %f29101, %f9835; fma.rn.ftz.f32 %f9837, %f29103, %f29103, %f9836; sqrt.approx.ftz.f32 %f958, %f9837; setp.neu.ftz.f32 %p628, %f958, 0f00000000; @%p628 bra $L__BB107_731; bra.uni $L__BB107_732; $L__BB107_731: rcp.approx.ftz.f32 %f9838, %f958; mul.ftz.f32 %f29103, %f9838, %f29103; mul.ftz.f32 %f29102, %f9838, %f29102; mul.ftz.f32 %f29101, %f9838, %f29101; $L__BB107_732: mul.ftz.f32 %f9839, %f953, %f948; mul.ftz.f32 %f9840, %f947, %f954; sub.ftz.f32 %f29104, %f9840, %f9839; mul.ftz.f32 %f9841, %f946, %f954; mul.ftz.f32 %f9842, %f952, %f948; sub.ftz.f32 %f29105, %f9842, %f9841; mul.ftz.f32 %f9843, %f952, %f947; mul.ftz.f32 %f9844, %f946, %f953; sub.ftz.f32 %f29106, %f9844, %f9843; mul.ftz.f32 %f9845, %f29105, %f29105; fma.rn.ftz.f32 %f9846, %f29104, %f29104, %f9845; fma.rn.ftz.f32 %f9847, %f29106, %f29106, %f9846; sqrt.approx.ftz.f32 %f974, %f9847; setp.neu.ftz.f32 %p629, %f974, 0f00000000; @%p629 bra $L__BB107_733; bra.uni $L__BB107_734; $L__BB107_733: rcp.approx.ftz.f32 %f9848, %f974; mul.ftz.f32 %f29106, %f29106, %f9848; mul.ftz.f32 %f29105, %f29105, %f9848; mul.ftz.f32 %f29104, %f29104, %f9848; $L__BB107_734: mul.ftz.f32 %f9849, %f953, %f951; mul.ftz.f32 %f9850, %f950, %f954; sub.ftz.f32 %f29107, %f9850, %f9849; mul.ftz.f32 %f9851, %f949, %f954; mul.ftz.f32 %f9852, %f952, %f951; sub.ftz.f32 %f29108, %f9852, %f9851; mul.ftz.f32 %f9853, %f952, %f950; mul.ftz.f32 %f9854, %f949, %f953; sub.ftz.f32 %f29109, %f9854, %f9853; mul.ftz.f32 %f9855, %f29108, %f29108; fma.rn.ftz.f32 %f9856, %f29107, %f29107, %f9855; fma.rn.ftz.f32 %f9857, %f29109, %f29109, %f9856; sqrt.approx.ftz.f32 %f990, %f9857; setp.neu.ftz.f32 %p630, %f990, 0f00000000; @%p630 bra $L__BB107_735; bra.uni $L__BB107_736; $L__BB107_735: rcp.approx.ftz.f32 %f9858, %f990; mul.ftz.f32 %f29109, %f29109, %f9858; mul.ftz.f32 %f29108, %f29108, %f9858; mul.ftz.f32 %f29107, %f29107, %f9858; $L__BB107_736: mul.ftz.f32 %f9860, %f29102, %f29105; fma.rn.ftz.f32 %f9861, %f29101, %f29104, %f9860; fma.rn.ftz.f32 %f9862, %f29103, %f29106, %f9861; abs.ftz.f32 %f9863, %f9862; setp.lt.ftz.f32 %p631, %f9863, 0f3F800000; mov.f32 %f9864, 0f3F800000; add.ftz.f32 %f9865, %f9863, 0fBF800000; add.ftz.f32 %f9866, %f9865, 0f3F800000; selp.f32 %f9867, %f9866, 0f3F800000, %p631; sub.ftz.f32 %f9868, %f9864, %f9867; sqrt.approx.ftz.f32 %f9869, %f9868; mul.ftz.f32 %f9870, %f9867, 0f3CB1485D; mov.f32 %f9871, 0f3D9FB43B; sub.ftz.f32 %f9872, %f9871, %f9870; fma.rn.ftz.f32 %f9873, %f9867, %f9872, 0fBE5A6B93; fma.rn.ftz.f32 %f9874, %f9867, %f9873, 0f3FC90FDB; mul.ftz.f32 %f9875, %f9869, %f9874; setp.lt.ftz.f32 %p632, %f9862, 0f00000000; mov.f32 %f29136, 0f00000000; mov.f32 %f9876, 0f40490FDB; sub.ftz.f32 %f9877, %f9876, %f9875; selp.f32 %f9878, %f9877, %f9875, %p632; mul.ftz.f32 %f9879, %f29102, %f29108; fma.rn.ftz.f32 %f9880, %f29101, %f29107, %f9879; fma.rn.ftz.f32 %f9881, %f29103, %f29109, %f9880; neg.ftz.f32 %f9882, %f9881; abs.ftz.f32 %f9883, %f9882; setp.lt.ftz.f32 %p633, %f9883, 0f3F800000; add.ftz.f32 %f9884, %f9883, 0fBF800000; add.ftz.f32 %f9885, %f9884, 0f3F800000; selp.f32 %f9886, %f9885, 0f3F800000, %p633; sub.ftz.f32 %f9887, %f9864, %f9886; sqrt.approx.ftz.f32 %f9888, %f9887; mul.ftz.f32 %f9889, %f9886, 0f3CB1485D; sub.ftz.f32 %f9890, %f9871, %f9889; fma.rn.ftz.f32 %f9891, %f9886, %f9890, 0fBE5A6B93; fma.rn.ftz.f32 %f9892, %f9886, %f9891, 0f3FC90FDB; mul.ftz.f32 %f9893, %f9888, %f9892; setp.gt.ftz.f32 %p634, %f9881, 0f80000000; sub.ftz.f32 %f9894, %f9876, %f9893; selp.f32 %f9895, %f9894, %f9893, %p634; mul.ftz.f32 %f9896, %f29105, %f29108; fma.rn.ftz.f32 %f9897, %f29104, %f29107, %f9896; fma.rn.ftz.f32 %f9898, %f29106, %f29109, %f9897; abs.ftz.f32 %f9899, %f9898; setp.lt.ftz.f32 %p635, %f9899, 0f3F800000; add.ftz.f32 %f9900, %f9899, 0fBF800000; add.ftz.f32 %f9901, %f9900, 0f3F800000; selp.f32 %f9902, %f9901, 0f3F800000, %p635; sub.ftz.f32 %f9903, %f9864, %f9902; sqrt.approx.ftz.f32 %f9904, %f9903; mul.ftz.f32 %f9905, %f9902, 0f3CB1485D; sub.ftz.f32 %f9906, %f9871, %f9905; fma.rn.ftz.f32 %f9907, %f9902, %f9906, 0fBE5A6B93; fma.rn.ftz.f32 %f9908, %f9902, %f9907, 0f3FC90FDB; mul.ftz.f32 %f9909, %f9904, %f9908; setp.lt.ftz.f32 %p636, %f9898, 0f00000000; sub.ftz.f32 %f9910, %f9876, %f9909; selp.f32 %f9911, %f9910, %f9909, %p636; add.ftz.f32 %f9912, %f9878, %f9895; add.ftz.f32 %f9913, %f9912, %f9911; add.ftz.f32 %f1000, %f9913, 0fC0490FDB; setp.eq.ftz.f32 %p637, %f1000, 0f00000000; @%p637 bra $L__BB107_749; setp.eq.s16 %p638, %rs615, 0; @%p638 bra $L__BB107_747; cvt.u32.u64 %r2060, %rd94; cvta.to.global.u64 %rd1645, %rd4763; mul.wide.s32 %rd1646, %r2060, 16; add.s64 %rd1647, %rd1645, %rd1646; ld.const.u64 %rd1648, [kernel_params+2256]; cvta.to.global.u64 %rd1649, %rd1648; ld.global.u32 %r2061, [%rd1647+12]; mul.wide.u32 %rd1650, %r2061, 12; add.s64 %rd1651, %rd1649, %rd1650; ld.global.f32 %f29126, [%rd1651]; ld.global.f32 %f29127, [%rd1651+4]; ld.global.f32 %f29128, [%rd1651+8]; add.s32 %r2062, %r2061, 1; mul.wide.u32 %rd1652, %r2062, 12; add.s64 %rd1653, %rd1649, %rd1652; ld.global.f32 %f29129, [%rd1653]; ld.global.f32 %f29130, [%rd1653+4]; ld.global.f32 %f29131, [%rd1653+8]; add.s32 %r2063, %r2061, 2; mul.wide.u32 %rd1654, %r2063, 12; add.s64 %rd1655, %rd1649, %rd1654; ld.global.f32 %f29132, [%rd1655]; ld.global.f32 %f29133, [%rd1655+4]; ld.global.f32 %f29134, [%rd1655+8]; @%p613 bra $L__BB107_746; and.b32 %r2064, %r361, 2; setp.eq.s32 %p640, %r2064, 0; ld.const.u64 %rd1656, [kernel_params+2160]; cvta.to.global.u64 %rd1657, %rd1656; mul.lo.s64 %rd1658, %rd95, 208; add.s64 %rd161, %rd1657, %rd1658; @%p640 bra $L__BB107_744; bra.uni $L__BB107_740; $L__BB107_744: ld.global.v4.f32 {%f29114, %f29115, %f29116, %f29117}, [%rd161]; ld.global.v4.f32 {%f29118, %f29119, %f29120, %f29121}, [%rd161+16]; ld.global.v4.f32 {%f29122, %f29123, %f29124, %f29125}, [%rd161+32]; bra.uni $L__BB107_745; $L__BB107_717: mul.ftz.f32 %f9555, %f9247, %f9544; fma.rn.ftz.f32 %f9556, %f9246, %f848, %f9555; mov.b32 %r2047, %f9248; mov.b64 %rd1622, {%r2047, %r2048}; cvt.u32.u64 %r2049, %rd1622; mov.b32 %f9557, %r2049; fma.rn.ftz.f32 %f9558, %f9557, %f9545, %f9556; abs.ftz.f32 %f855, %f9558; setp.eq.ftz.f32 %p620, %f855, 0f00000000; mov.f32 %f29075, 0f00000000; @%p620 bra $L__BB107_719; ld.const.f32 %f9559, [kernel_params+1500]; mul.ftz.f32 %f9560, %f732, %f732; mul.ftz.f32 %f9561, %f9560, %f9559; div.approx.ftz.f32 %f29075, %f9561, %f855; $L__BB107_719: setp.eq.s16 %p621, %rs615, 0; mov.f32 %f29136, %f29075; @%p621 bra $L__BB107_749; sqrt.approx.ftz.f32 %f9563, %f854; mul.ftz.f32 %f858, %f9563, 0f3F000000; setp.eq.ftz.f32 %p622, %f858, 0f00000000; mov.f32 %f29136, 0f00000000; @%p622 bra $L__BB107_749; cvt.u32.u64 %r2050, %rd94; cvta.to.global.u64 %rd1623, %rd4763; mul.wide.s32 %rd1624, %r2050, 16; add.s64 %rd1625, %rd1623, %rd1624; ld.const.u64 %rd1626, [kernel_params+2256]; cvta.to.global.u64 %rd1627, %rd1626; ld.global.u32 %r2051, [%rd1625+12]; mul.wide.u32 %rd1628, %r2051, 12; add.s64 %rd1629, %rd1627, %rd1628; ld.global.f32 %f29092, [%rd1629]; ld.global.f32 %f29093, [%rd1629+4]; ld.global.f32 %f29094, [%rd1629+8]; add.s32 %r2052, %r2051, 1; mul.wide.u32 %rd1630, %r2052, 12; add.s64 %rd1631, %rd1627, %rd1630; ld.global.f32 %f29095, [%rd1631]; ld.global.f32 %f29096, [%rd1631+4]; ld.global.f32 %f29097, [%rd1631+8]; add.s32 %r2053, %r2051, 2; mul.wide.u32 %rd1632, %r2053, 12; add.s64 %rd1633, %rd1627, %rd1632; ld.global.f32 %f29098, [%rd1633]; ld.global.f32 %f29099, [%rd1633+4]; ld.global.f32 %f29100, [%rd1633+8]; @%p613 bra $L__BB107_729; and.b32 %r2054, %r361, 2; setp.eq.s32 %p624, %r2054, 0; ld.const.u64 %rd1634, [kernel_params+2160]; cvta.to.global.u64 %rd1635, %rd1634; mul.lo.s64 %rd1636, %rd95, 208; add.s64 %rd159, %rd1635, %rd1636; @%p624 bra $L__BB107_727; bra.uni $L__BB107_723; $L__BB107_727: ld.global.v4.f32 {%f29080, %f29081, %f29082, %f29083}, [%rd159]; ld.global.v4.f32 {%f29084, %f29085, %f29086, %f29087}, [%rd159+16]; ld.global.v4.f32 {%f29088, %f29089, %f29090, %f29091}, [%rd159+32]; bra.uni $L__BB107_728; $L__BB107_747: sqrt.approx.ftz.f32 %f29135, %f854; bra.uni $L__BB107_748; $L__BB107_740: ld.const.u64 %rd1659, [kernel_params+2176]; cvta.to.global.u64 %rd1660, %rd1659; ld.global.u32 %rd1661, [%rd161+168]; ld.global.u32 %r2065, [%rd161+152]; shl.b32 %r2066, %r2065, 1; cvt.rn.f32.s32 %f9914, %r2066; mul.ftz.f32 %f9915, %f9914, 0f3F000000; cvt.rzi.ftz.s32.f32 %r2067, %f9915; add.s32 %r2068, %r2066, -1; min.s32 %r2069, %r2067, %r2068; cvt.rn.f32.s32 %f9916, %r2069; sub.ftz.f32 %f1010, %f9915, %f9916; cvt.s64.s32 %rd1662, %r2069; add.s64 %rd1663, %rd1662, %rd1661; shl.b64 %rd1664, %rd1663, 6; add.s64 %rd162, %rd1660, %rd1664; ld.global.v4.f32 {%f9917, %f9918, %f9919, %f9920}, [%rd162]; ld.global.v4.f32 {%f9921, %f9922, %f9923, %f9924}, [%rd162+64]; mul.ftz.f32 %f9925, %f9918, %f9922; fma.rn.ftz.f32 %f9926, %f9917, %f9921, %f9925; mul.ftz.f32 %f9927, %f9920, %f9924; fma.rn.ftz.f32 %f9928, %f9919, %f9923, %f9927; add.ftz.f32 %f1019, %f9926, %f9928; setp.gt.ftz.f32 %p641, %f1019, 0f3F7FDF3B; sub.ftz.f32 %f1020, %f9864, %f1010; @%p641 bra $L__BB107_742; bra.uni $L__BB107_741; $L__BB107_742: mul.ftz.f32 %f9983, %f9917, %f1020; fma.rn.ftz.f32 %f9984, %f9921, %f1010, %f9983; mul.ftz.f32 %f9985, %f9918, %f1020; fma.rn.ftz.f32 %f9986, %f9922, %f1010, %f9985; mul.ftz.f32 %f9987, %f9919, %f1020; fma.rn.ftz.f32 %f9988, %f9923, %f1010, %f9987; mul.ftz.f32 %f9989, %f9920, %f1020; fma.rn.ftz.f32 %f9990, %f9924, %f1010, %f9989; mul.ftz.f32 %f9991, %f9986, %f9986; fma.rn.ftz.f32 %f9992, %f9984, %f9984, %f9991; mul.ftz.f32 %f9993, %f9990, %f9990; fma.rn.ftz.f32 %f9994, %f9988, %f9988, %f9993; add.ftz.f32 %f9995, %f9992, %f9994; rsqrt.approx.ftz.f32 %f9996, %f9995; mul.ftz.f32 %f29113, %f9990, %f9996; mul.ftz.f32 %f29112, %f9988, %f9996; mul.ftz.f32 %f29111, %f9986, %f9996; mul.ftz.f32 %f29110, %f9984, %f9996; bra.uni $L__BB107_743; $L__BB107_723: ld.const.u64 %rd1637, [kernel_params+2176]; cvta.to.global.u64 %rd1638, %rd1637; ld.global.u32 %rd1639, [%rd159+168]; ld.global.u32 %r2055, [%rd159+152]; shl.b32 %r2056, %r2055, 1; cvt.rn.f32.s32 %f9564, %r2056; mul.ftz.f32 %f9565, %f9564, 0f3F000000; cvt.rzi.ftz.s32.f32 %r2057, %f9565; add.s32 %r2058, %r2056, -1; min.s32 %r2059, %r2057, %r2058; cvt.rn.f32.s32 %f9566, %r2059; sub.ftz.f32 %f868, %f9565, %f9566; cvt.s64.s32 %rd1640, %r2059; add.s64 %rd1641, %rd1640, %rd1639; shl.b64 %rd1642, %rd1641, 6; add.s64 %rd160, %rd1638, %rd1642; ld.global.v4.f32 {%f9567, %f9568, %f9569, %f9570}, [%rd160]; ld.global.v4.f32 {%f9571, %f9572, %f9573, %f9574}, [%rd160+64]; mul.ftz.f32 %f9575, %f9568, %f9572; fma.rn.ftz.f32 %f9576, %f9567, %f9571, %f9575; mul.ftz.f32 %f9577, %f9570, %f9574; fma.rn.ftz.f32 %f9578, %f9569, %f9573, %f9577; add.ftz.f32 %f877, %f9576, %f9578; setp.gt.ftz.f32 %p625, %f877, 0f3F7FDF3B; mov.f32 %f9579, 0f3F800000; sub.ftz.f32 %f878, %f9579, %f868; @%p625 bra $L__BB107_725; bra.uni $L__BB107_724; $L__BB107_725: mul.ftz.f32 %f9633, %f9567, %f878; fma.rn.ftz.f32 %f9634, %f9571, %f868, %f9633; mul.ftz.f32 %f9635, %f9568, %f878; fma.rn.ftz.f32 %f9636, %f9572, %f868, %f9635; mul.ftz.f32 %f9637, %f9569, %f878; fma.rn.ftz.f32 %f9638, %f9573, %f868, %f9637; mul.ftz.f32 %f9639, %f9570, %f878; fma.rn.ftz.f32 %f9640, %f9574, %f868, %f9639; mul.ftz.f32 %f9641, %f9636, %f9636; fma.rn.ftz.f32 %f9642, %f9634, %f9634, %f9641; mul.ftz.f32 %f9643, %f9640, %f9640; fma.rn.ftz.f32 %f9644, %f9638, %f9638, %f9643; add.ftz.f32 %f9645, %f9642, %f9644; rsqrt.approx.ftz.f32 %f9646, %f9645; mul.ftz.f32 %f29079, %f9640, %f9646; mul.ftz.f32 %f29078, %f9638, %f9646; mul.ftz.f32 %f29077, %f9636, %f9646; mul.ftz.f32 %f29076, %f9634, %f9646; bra.uni $L__BB107_726; $L__BB107_741: mov.f32 %f9930, 0fBF800000; max.ftz.f32 %f9931, %f1019, %f9930; min.ftz.f32 %f9933, %f9931, %f9864; abs.ftz.f32 %f9934, %f9933; sub.ftz.f32 %f9935, %f9864, %f9934; mul.ftz.f32 %f9936, %f9935, 0f3F000000; sqrt.approx.ftz.f32 %f9937, %f9936; setp.gt.ftz.f32 %p642, %f9934, 0f3F11EB85; selp.f32 %f9938, %f9937, %f9934, %p642; mul.ftz.f32 %f9939, %f9938, %f9938; mov.f32 %f9940, 0f3C94D2E9; mov.f32 %f9941, 0f3D53F941; fma.rn.ftz.f32 %f9942, %f9941, %f9939, %f9940; mov.f32 %f9943, 0f3D3F841F; fma.rn.ftz.f32 %f9944, %f9942, %f9939, %f9943; mov.f32 %f9945, 0f3D994929; fma.rn.ftz.f32 %f9946, %f9944, %f9939, %f9945; mov.f32 %f9947, 0f3E2AAB94; fma.rn.ftz.f32 %f9948, %f9946, %f9939, %f9947; mul.ftz.f32 %f9949, %f9939, %f9948; fma.rn.ftz.f32 %f9950, %f9949, %f9938, %f9938; add.ftz.f32 %f9951, %f9950, %f9950; mov.f32 %f9952, 0f3FC90FDB; sub.ftz.f32 %f9953, %f9952, %f9950; selp.f32 %f9954, %f9951, %f9953, %p642; setp.lt.ftz.f32 %p643, %f9933, 0f00000000; sub.ftz.f32 %f9956, %f9876, %f9954; selp.f32 %f9957, %f9956, %f9954, %p643; mul.ftz.f32 %f9958, %f9917, %f1019; sub.ftz.f32 %f9959, %f9921, %f9958; mul.ftz.f32 %f9960, %f9918, %f1019; sub.ftz.f32 %f9961, %f9922, %f9960; mul.ftz.f32 %f9962, %f9919, %f1019; sub.ftz.f32 %f9963, %f9923, %f9962; mul.ftz.f32 %f9964, %f9920, %f1019; sub.ftz.f32 %f9965, %f9924, %f9964; mul.ftz.f32 %f9966, %f9961, %f9961; fma.rn.ftz.f32 %f9967, %f9959, %f9959, %f9966; mul.ftz.f32 %f9968, %f9965, %f9965; fma.rn.ftz.f32 %f9969, %f9963, %f9963, %f9968; add.ftz.f32 %f9970, %f9967, %f9969; rsqrt.approx.ftz.f32 %f9971, %f9970; mul.ftz.f32 %f9972, %f9959, %f9971; mul.ftz.f32 %f9973, %f9961, %f9971; mul.ftz.f32 %f9974, %f9963, %f9971; mul.ftz.f32 %f9975, %f9965, %f9971; mul.ftz.f32 %f9976, %f1010, %f9957; cos.approx.ftz.f32 %f9977, %f9976; sin.approx.ftz.f32 %f9978, %f9976; mul.ftz.f32 %f9979, %f9972, %f9978; mul.ftz.f32 %f9980, %f9973, %f9978; mul.ftz.f32 %f9981, %f9974, %f9978; mul.ftz.f32 %f9982, %f9975, %f9978; fma.rn.ftz.f32 %f29113, %f9920, %f9977, %f9982; fma.rn.ftz.f32 %f29112, %f9919, %f9977, %f9981; fma.rn.ftz.f32 %f29111, %f9918, %f9977, %f9980; fma.rn.ftz.f32 %f29110, %f9917, %f9977, %f9979; $L__BB107_743: ld.global.v4.f32 {%f9997, %f9998, %f9999, %f10000}, [%rd162+16]; ld.global.v4.f32 {%f10005, %f10006, %f10007, %f10008}, [%rd162+80]; mul.ftz.f32 %f10013, %f1010, %f10005; mul.ftz.f32 %f10014, %f1010, %f10006; mul.ftz.f32 %f10015, %f1010, %f10007; mul.ftz.f32 %f10016, %f1010, %f10008; fma.rn.ftz.f32 %f10017, %f1020, %f10000, %f10016; ld.global.v4.f32 {%f10018, %f10019, %f10020, %f10021}, [%rd162+32]; ld.global.v4.f32 {%f10026, %f10027, %f10028, %f10029}, [%rd162+96]; mul.ftz.f32 %f10034, %f1010, %f10026; mul.ftz.f32 %f10035, %f1010, %f10027; mul.ftz.f32 %f10036, %f1010, %f10028; mul.ftz.f32 %f10037, %f1010, %f10029; fma.rn.ftz.f32 %f10038, %f1020, %f10018, %f10034; fma.rn.ftz.f32 %f10039, %f1020, %f10019, %f10035; fma.rn.ftz.f32 %f10040, %f1020, %f10020, %f10036; fma.rn.ftz.f32 %f10041, %f1020, %f10021, %f10037; ld.global.v4.f32 {%f10042, %f10043, %f10044, %f10045}, [%rd162+48]; ld.global.v4.f32 {%f10050, %f10051, %f10052, %f10053}, [%rd162+112]; mul.ftz.f32 %f10058, %f1010, %f10050; mul.ftz.f32 %f10059, %f1010, %f10051; mul.ftz.f32 %f10060, %f1010, %f10052; mul.ftz.f32 %f10061, %f1010, %f10053; fma.rn.ftz.f32 %f10062, %f1020, %f10042, %f10058; fma.rn.ftz.f32 %f10063, %f1020, %f10043, %f10059; fma.rn.ftz.f32 %f10064, %f1020, %f10044, %f10060; fma.rn.ftz.f32 %f10065, %f1020, %f10045, %f10061; mul.ftz.f32 %f10066, %f29113, 0f3FB504F3; mul.ftz.f32 %f10067, %f29110, 0f3FB504F3; mul.ftz.f32 %f10068, %f29111, 0f3FB504F3; mul.ftz.f32 %f10069, %f29112, 0f3FB504F3; mul.ftz.f32 %f10070, %f10066, %f10067; mul.ftz.f32 %f10071, %f10066, %f10068; mul.ftz.f32 %f10072, %f10066, %f10069; mul.ftz.f32 %f10073, %f10067, %f10067; mul.ftz.f32 %f10074, %f10067, %f10068; mul.ftz.f32 %f10075, %f10067, %f10069; mul.ftz.f32 %f10076, %f10068, %f10068; mul.ftz.f32 %f10077, %f10068, %f10069; mul.ftz.f32 %f10078, %f10069, %f10069; sub.ftz.f32 %f10080, %f9864, %f10076; sub.ftz.f32 %f10081, %f10080, %f10078; sub.ftz.f32 %f10082, %f10074, %f10072; add.ftz.f32 %f10083, %f10071, %f10075; add.ftz.f32 %f10084, %f10072, %f10074; sub.ftz.f32 %f10085, %f9864, %f10073; sub.ftz.f32 %f10086, %f10085, %f10078; sub.ftz.f32 %f10087, %f10077, %f10070; sub.ftz.f32 %f10088, %f10075, %f10071; add.ftz.f32 %f10089, %f10070, %f10077; sub.ftz.f32 %f10090, %f10085, %f10076; mul.ftz.f32 %f10091, %f10082, %f10040; fma.rn.ftz.f32 %f10092, %f10081, %f10017, %f10091; mul.ftz.f32 %f10093, %f10082, %f10041; fma.rn.ftz.f32 %f10094, %f10081, %f10038, %f10093; mul.ftz.f32 %f10095, %f10082, %f10062; fma.rn.ftz.f32 %f10096, %f10081, %f10039, %f10095; mul.ftz.f32 %f10097, %f10086, %f10040; fma.rn.ftz.f32 %f10098, %f10084, %f10017, %f10097; mul.ftz.f32 %f10099, %f10086, %f10041; fma.rn.ftz.f32 %f10100, %f10084, %f10038, %f10099; mul.ftz.f32 %f10101, %f10086, %f10062; fma.rn.ftz.f32 %f10102, %f10084, %f10039, %f10101; mul.ftz.f32 %f10103, %f10089, %f10040; fma.rn.ftz.f32 %f10104, %f10088, %f10017, %f10103; mul.ftz.f32 %f10105, %f10089, %f10041; fma.rn.ftz.f32 %f10106, %f10088, %f10038, %f10105; mul.ftz.f32 %f10107, %f10089, %f10062; fma.rn.ftz.f32 %f10108, %f10088, %f10039, %f10107; fma.rn.ftz.f32 %f29120, %f10087, %f10065, %f10102; fma.rn.ftz.f32 %f29119, %f10087, %f10064, %f10100; fma.rn.ftz.f32 %f29118, %f10087, %f10063, %f10098; fma.rn.ftz.f32 %f29124, %f10090, %f10065, %f10108; fma.rn.ftz.f32 %f29123, %f10090, %f10064, %f10106; fma.rn.ftz.f32 %f29122, %f10090, %f10063, %f10104; fma.rn.ftz.f32 %f29116, %f10083, %f10065, %f10096; fma.rn.ftz.f32 %f29115, %f10083, %f10064, %f10094; fma.rn.ftz.f32 %f29114, %f10083, %f10063, %f10092; fma.rn.ftz.f32 %f29125, %f1020, %f9999, %f10015; fma.rn.ftz.f32 %f29121, %f1020, %f9998, %f10014; fma.rn.ftz.f32 %f29117, %f1020, %f9997, %f10013; $L__BB107_745: mul.ftz.f32 %f10121, %f29127, %f29115; fma.rn.ftz.f32 %f10122, %f29126, %f29114, %f10121; fma.rn.ftz.f32 %f10123, %f29128, %f29116, %f10122; add.ftz.f32 %f1069, %f29117, %f10123; mul.ftz.f32 %f10124, %f29127, %f29119; fma.rn.ftz.f32 %f10125, %f29126, %f29118, %f10124; fma.rn.ftz.f32 %f10126, %f29128, %f29120, %f10125; add.ftz.f32 %f1070, %f29121, %f10126; mul.ftz.f32 %f10127, %f29127, %f29123; fma.rn.ftz.f32 %f10128, %f29126, %f29122, %f10127; fma.rn.ftz.f32 %f10129, %f29128, %f29124, %f10128; add.ftz.f32 %f29128, %f29125, %f10129; mul.ftz.f32 %f10130, %f29130, %f29115; fma.rn.ftz.f32 %f10131, %f29129, %f29114, %f10130; fma.rn.ftz.f32 %f10132, %f29131, %f29116, %f10131; add.ftz.f32 %f1072, %f29117, %f10132; mul.ftz.f32 %f10133, %f29130, %f29119; fma.rn.ftz.f32 %f10134, %f29129, %f29118, %f10133; fma.rn.ftz.f32 %f10135, %f29131, %f29120, %f10134; add.ftz.f32 %f1073, %f29121, %f10135; mul.ftz.f32 %f10136, %f29130, %f29123; fma.rn.ftz.f32 %f10137, %f29129, %f29122, %f10136; fma.rn.ftz.f32 %f10138, %f29131, %f29124, %f10137; add.ftz.f32 %f29131, %f29125, %f10138; mul.ftz.f32 %f10139, %f29133, %f29115; fma.rn.ftz.f32 %f10140, %f29132, %f29114, %f10139; fma.rn.ftz.f32 %f10141, %f29134, %f29116, %f10140; add.ftz.f32 %f1075, %f29117, %f10141; mul.ftz.f32 %f10142, %f29133, %f29119; fma.rn.ftz.f32 %f10143, %f29132, %f29118, %f10142; fma.rn.ftz.f32 %f10144, %f29134, %f29120, %f10143; add.ftz.f32 %f1076, %f29121, %f10144; mul.ftz.f32 %f10145, %f29133, %f29123; fma.rn.ftz.f32 %f10146, %f29132, %f29122, %f10145; fma.rn.ftz.f32 %f10147, %f29134, %f29124, %f10146; add.ftz.f32 %f29134, %f29125, %f10147; mov.f32 %f29126, %f1069; mov.f32 %f29127, %f1070; mov.f32 %f29129, %f1072; mov.f32 %f29130, %f1073; mov.f32 %f29132, %f1075; mov.f32 %f29133, %f1076; $L__BB107_746: sub.ftz.f32 %f10148, %f29128, %f29131; sub.ftz.f32 %f10149, %f29133, %f29130; mul.ftz.f32 %f10150, %f10148, %f10149; sub.ftz.f32 %f10151, %f29127, %f29130; sub.ftz.f32 %f10152, %f29134, %f29131; mul.ftz.f32 %f10153, %f10151, %f10152; sub.ftz.f32 %f10154, %f10150, %f10153; sub.ftz.f32 %f10155, %f29126, %f29129; mul.ftz.f32 %f10156, %f10155, %f10152; sub.ftz.f32 %f10157, %f29132, %f29129; mul.ftz.f32 %f10158, %f10148, %f10157; sub.ftz.f32 %f10159, %f10156, %f10158; mul.ftz.f32 %f10160, %f10151, %f10157; mul.ftz.f32 %f10161, %f10155, %f10149; sub.ftz.f32 %f10162, %f10160, %f10161; mul.ftz.f32 %f10163, %f10159, %f10159; fma.rn.ftz.f32 %f10164, %f10154, %f10154, %f10163; fma.rn.ftz.f32 %f10165, %f10162, %f10162, %f10164; sqrt.approx.ftz.f32 %f29135, %f10165; $L__BB107_748: ld.const.f32 %f10166, [kernel_params+1500]; mul.ftz.f32 %f10167, %f29135, 0f3F000000; mul.ftz.f32 %f10168, %f10167, %f10166; div.approx.ftz.f32 %f29136, %f10168, %f1000; bra.uni $L__BB107_749; $L__BB107_724: mov.f32 %f9580, 0fBF800000; max.ftz.f32 %f9581, %f877, %f9580; min.ftz.f32 %f9583, %f9581, %f9579; abs.ftz.f32 %f9584, %f9583; sub.ftz.f32 %f9585, %f9579, %f9584; mul.ftz.f32 %f9586, %f9585, 0f3F000000; sqrt.approx.ftz.f32 %f9587, %f9586; setp.gt.ftz.f32 %p626, %f9584, 0f3F11EB85; selp.f32 %f9588, %f9587, %f9584, %p626; mul.ftz.f32 %f9589, %f9588, %f9588; mov.f32 %f9590, 0f3C94D2E9; mov.f32 %f9591, 0f3D53F941; fma.rn.ftz.f32 %f9592, %f9591, %f9589, %f9590; mov.f32 %f9593, 0f3D3F841F; fma.rn.ftz.f32 %f9594, %f9592, %f9589, %f9593; mov.f32 %f9595, 0f3D994929; fma.rn.ftz.f32 %f9596, %f9594, %f9589, %f9595; mov.f32 %f9597, 0f3E2AAB94; fma.rn.ftz.f32 %f9598, %f9596, %f9589, %f9597; mul.ftz.f32 %f9599, %f9589, %f9598; fma.rn.ftz.f32 %f9600, %f9599, %f9588, %f9588; add.ftz.f32 %f9601, %f9600, %f9600; mov.f32 %f9602, 0f3FC90FDB; sub.ftz.f32 %f9603, %f9602, %f9600; selp.f32 %f9604, %f9601, %f9603, %p626; setp.lt.ftz.f32 %p627, %f9583, 0f00000000; mov.f32 %f9605, 0f40490FDB; sub.ftz.f32 %f9606, %f9605, %f9604; selp.f32 %f9607, %f9606, %f9604, %p627; mul.ftz.f32 %f9608, %f9567, %f877; sub.ftz.f32 %f9609, %f9571, %f9608; mul.ftz.f32 %f9610, %f9568, %f877; sub.ftz.f32 %f9611, %f9572, %f9610; mul.ftz.f32 %f9612, %f9569, %f877; sub.ftz.f32 %f9613, %f9573, %f9612; mul.ftz.f32 %f9614, %f9570, %f877; sub.ftz.f32 %f9615, %f9574, %f9614; mul.ftz.f32 %f9616, %f9611, %f9611; fma.rn.ftz.f32 %f9617, %f9609, %f9609, %f9616; mul.ftz.f32 %f9618, %f9615, %f9615; fma.rn.ftz.f32 %f9619, %f9613, %f9613, %f9618; add.ftz.f32 %f9620, %f9617, %f9619; rsqrt.approx.ftz.f32 %f9621, %f9620; mul.ftz.f32 %f9622, %f9609, %f9621; mul.ftz.f32 %f9623, %f9611, %f9621; mul.ftz.f32 %f9624, %f9613, %f9621; mul.ftz.f32 %f9625, %f9615, %f9621; mul.ftz.f32 %f9626, %f868, %f9607; cos.approx.ftz.f32 %f9627, %f9626; sin.approx.ftz.f32 %f9628, %f9626; mul.ftz.f32 %f9629, %f9622, %f9628; mul.ftz.f32 %f9630, %f9623, %f9628; mul.ftz.f32 %f9631, %f9624, %f9628; mul.ftz.f32 %f9632, %f9625, %f9628; fma.rn.ftz.f32 %f29079, %f9570, %f9627, %f9632; fma.rn.ftz.f32 %f29078, %f9569, %f9627, %f9631; fma.rn.ftz.f32 %f29077, %f9568, %f9627, %f9630; fma.rn.ftz.f32 %f29076, %f9567, %f9627, %f9629; $L__BB107_726: ld.global.v4.f32 {%f9647, %f9648, %f9649, %f9650}, [%rd160+16]; ld.global.v4.f32 {%f9655, %f9656, %f9657, %f9658}, [%rd160+80]; mul.ftz.f32 %f9663, %f868, %f9655; mul.ftz.f32 %f9664, %f868, %f9656; mul.ftz.f32 %f9665, %f868, %f9657; mul.ftz.f32 %f9666, %f868, %f9658; fma.rn.ftz.f32 %f9667, %f878, %f9650, %f9666; ld.global.v4.f32 {%f9668, %f9669, %f9670, %f9671}, [%rd160+32]; ld.global.v4.f32 {%f9676, %f9677, %f9678, %f9679}, [%rd160+96]; mul.ftz.f32 %f9684, %f868, %f9676; mul.ftz.f32 %f9685, %f868, %f9677; mul.ftz.f32 %f9686, %f868, %f9678; mul.ftz.f32 %f9687, %f868, %f9679; fma.rn.ftz.f32 %f9688, %f878, %f9668, %f9684; fma.rn.ftz.f32 %f9689, %f878, %f9669, %f9685; fma.rn.ftz.f32 %f9690, %f878, %f9670, %f9686; fma.rn.ftz.f32 %f9691, %f878, %f9671, %f9687; ld.global.v4.f32 {%f9692, %f9693, %f9694, %f9695}, [%rd160+48]; ld.global.v4.f32 {%f9700, %f9701, %f9702, %f9703}, [%rd160+112]; mul.ftz.f32 %f9708, %f868, %f9700; mul.ftz.f32 %f9709, %f868, %f9701; mul.ftz.f32 %f9710, %f868, %f9702; mul.ftz.f32 %f9711, %f868, %f9703; fma.rn.ftz.f32 %f9712, %f878, %f9692, %f9708; fma.rn.ftz.f32 %f9713, %f878, %f9693, %f9709; fma.rn.ftz.f32 %f9714, %f878, %f9694, %f9710; fma.rn.ftz.f32 %f9715, %f878, %f9695, %f9711; mul.ftz.f32 %f9716, %f29079, 0f3FB504F3; mul.ftz.f32 %f9717, %f29076, 0f3FB504F3; mul.ftz.f32 %f9718, %f29077, 0f3FB504F3; mul.ftz.f32 %f9719, %f29078, 0f3FB504F3; mul.ftz.f32 %f9720, %f9716, %f9717; mul.ftz.f32 %f9721, %f9716, %f9718; mul.ftz.f32 %f9722, %f9716, %f9719; mul.ftz.f32 %f9723, %f9717, %f9717; mul.ftz.f32 %f9724, %f9717, %f9718; mul.ftz.f32 %f9725, %f9717, %f9719; mul.ftz.f32 %f9726, %f9718, %f9718; mul.ftz.f32 %f9727, %f9718, %f9719; mul.ftz.f32 %f9728, %f9719, %f9719; sub.ftz.f32 %f9730, %f9579, %f9726; sub.ftz.f32 %f9731, %f9730, %f9728; sub.ftz.f32 %f9732, %f9724, %f9722; add.ftz.f32 %f9733, %f9721, %f9725; add.ftz.f32 %f9734, %f9722, %f9724; sub.ftz.f32 %f9735, %f9579, %f9723; sub.ftz.f32 %f9736, %f9735, %f9728; sub.ftz.f32 %f9737, %f9727, %f9720; sub.ftz.f32 %f9738, %f9725, %f9721; add.ftz.f32 %f9739, %f9720, %f9727; sub.ftz.f32 %f9740, %f9735, %f9726; mul.ftz.f32 %f9741, %f9732, %f9690; fma.rn.ftz.f32 %f9742, %f9731, %f9667, %f9741; mul.ftz.f32 %f9743, %f9732, %f9691; fma.rn.ftz.f32 %f9744, %f9731, %f9688, %f9743; mul.ftz.f32 %f9745, %f9732, %f9712; fma.rn.ftz.f32 %f9746, %f9731, %f9689, %f9745; mul.ftz.f32 %f9747, %f9736, %f9690; fma.rn.ftz.f32 %f9748, %f9734, %f9667, %f9747; mul.ftz.f32 %f9749, %f9736, %f9691; fma.rn.ftz.f32 %f9750, %f9734, %f9688, %f9749; mul.ftz.f32 %f9751, %f9736, %f9712; fma.rn.ftz.f32 %f9752, %f9734, %f9689, %f9751; mul.ftz.f32 %f9753, %f9739, %f9690; fma.rn.ftz.f32 %f9754, %f9738, %f9667, %f9753; mul.ftz.f32 %f9755, %f9739, %f9691; fma.rn.ftz.f32 %f9756, %f9738, %f9688, %f9755; mul.ftz.f32 %f9757, %f9739, %f9712; fma.rn.ftz.f32 %f9758, %f9738, %f9689, %f9757; fma.rn.ftz.f32 %f29086, %f9737, %f9715, %f9752; fma.rn.ftz.f32 %f29085, %f9737, %f9714, %f9750; fma.rn.ftz.f32 %f29084, %f9737, %f9713, %f9748; fma.rn.ftz.f32 %f29090, %f9740, %f9715, %f9758; fma.rn.ftz.f32 %f29089, %f9740, %f9714, %f9756; fma.rn.ftz.f32 %f29088, %f9740, %f9713, %f9754; fma.rn.ftz.f32 %f29082, %f9733, %f9715, %f9746; fma.rn.ftz.f32 %f29081, %f9733, %f9714, %f9744; fma.rn.ftz.f32 %f29080, %f9733, %f9713, %f9742; fma.rn.ftz.f32 %f29091, %f878, %f9649, %f9665; fma.rn.ftz.f32 %f29087, %f878, %f9648, %f9664; fma.rn.ftz.f32 %f29083, %f878, %f9647, %f9663; $L__BB107_728: mul.ftz.f32 %f9771, %f29093, %f29081; fma.rn.ftz.f32 %f9772, %f29092, %f29080, %f9771; fma.rn.ftz.f32 %f9773, %f29094, %f29082, %f9772; add.ftz.f32 %f927, %f29083, %f9773; mul.ftz.f32 %f9774, %f29093, %f29085; fma.rn.ftz.f32 %f9775, %f29092, %f29084, %f9774; fma.rn.ftz.f32 %f9776, %f29094, %f29086, %f9775; add.ftz.f32 %f928, %f29087, %f9776; mul.ftz.f32 %f9777, %f29093, %f29089; fma.rn.ftz.f32 %f9778, %f29092, %f29088, %f9777; fma.rn.ftz.f32 %f9779, %f29094, %f29090, %f9778; add.ftz.f32 %f29094, %f29091, %f9779; mul.ftz.f32 %f9780, %f29096, %f29081; fma.rn.ftz.f32 %f9781, %f29095, %f29080, %f9780; fma.rn.ftz.f32 %f9782, %f29097, %f29082, %f9781; add.ftz.f32 %f930, %f29083, %f9782; mul.ftz.f32 %f9783, %f29096, %f29085; fma.rn.ftz.f32 %f9784, %f29095, %f29084, %f9783; fma.rn.ftz.f32 %f9785, %f29097, %f29086, %f9784; add.ftz.f32 %f931, %f29087, %f9785; mul.ftz.f32 %f9786, %f29096, %f29089; fma.rn.ftz.f32 %f9787, %f29095, %f29088, %f9786; fma.rn.ftz.f32 %f9788, %f29097, %f29090, %f9787; add.ftz.f32 %f29097, %f29091, %f9788; mul.ftz.f32 %f9789, %f29099, %f29081; fma.rn.ftz.f32 %f9790, %f29098, %f29080, %f9789; fma.rn.ftz.f32 %f9791, %f29100, %f29082, %f9790; add.ftz.f32 %f933, %f29083, %f9791; mul.ftz.f32 %f9792, %f29099, %f29085; fma.rn.ftz.f32 %f9793, %f29098, %f29084, %f9792; fma.rn.ftz.f32 %f9794, %f29100, %f29086, %f9793; add.ftz.f32 %f934, %f29087, %f9794; mul.ftz.f32 %f9795, %f29099, %f29089; fma.rn.ftz.f32 %f9796, %f29098, %f29088, %f9795; fma.rn.ftz.f32 %f9797, %f29100, %f29090, %f9796; add.ftz.f32 %f29100, %f29091, %f9797; mov.f32 %f29092, %f927; mov.f32 %f29093, %f928; mov.f32 %f29095, %f930; mov.f32 %f29096, %f931; mov.f32 %f29098, %f933; mov.f32 %f29099, %f934; $L__BB107_729: sub.ftz.f32 %f9798, %f29094, %f29097; sub.ftz.f32 %f9799, %f29099, %f29096; mul.ftz.f32 %f9800, %f9798, %f9799; sub.ftz.f32 %f9801, %f29093, %f29096; sub.ftz.f32 %f9802, %f29100, %f29097; mul.ftz.f32 %f9803, %f9801, %f9802; sub.ftz.f32 %f9804, %f9800, %f9803; sub.ftz.f32 %f9805, %f29092, %f29095; mul.ftz.f32 %f9806, %f9805, %f9802; sub.ftz.f32 %f9807, %f29098, %f29095; mul.ftz.f32 %f9808, %f9798, %f9807; sub.ftz.f32 %f9809, %f9806, %f9808; mul.ftz.f32 %f9810, %f9801, %f9807; mul.ftz.f32 %f9811, %f9805, %f9799; sub.ftz.f32 %f9812, %f9810, %f9811; mul.ftz.f32 %f9813, %f9809, %f9809; fma.rn.ftz.f32 %f9814, %f9804, %f9804, %f9813; fma.rn.ftz.f32 %f9815, %f9812, %f9812, %f9814; sqrt.approx.ftz.f32 %f9816, %f9815; mul.ftz.f32 %f9817, %f9816, 0f3F000000; mul.ftz.f32 %f9818, %f29075, %f9817; div.approx.ftz.f32 %f29136, %f9818, %f858; $L__BB107_749: mul.ftz.f32 %f10169, %f731, %f731; fma.rn.ftz.f32 %f10170, %f29136, %f29136, %f10169; div.approx.ftz.f32 %f29137, %f10169, %f10170; $L__BB107_750: ld.local.u32 %r381, [%rd5+-12]; setp.eq.s32 %p644, %r381, -1; mov.u32 %r5651, -1; @%p644 bra $L__BB107_752; ld.const.u64 %rd1665, [kernel_params+2160]; cvta.to.global.u64 %rd1666, %rd1665; mul.wide.s32 %rd1667, %r381, 208; add.s64 %rd1668, %rd1666, %rd1667; ld.global.u32 %r5651, [%rd1668+192]; $L__BB107_752: abs.ftz.f32 %f10172, %f727; setp.gt.ftz.f32 %p645, %f10172, 0f00000000; mov.f32 %f29146, 0f00000000; selp.f32 %f10173, 0f3F800000, 0f00000000, %p645; ld.const.u64 %rd1669, [kernel_params+2584]; cvta.to.global.u64 %rd1670, %rd1669; mul.lo.s64 %rd1671, %rd3, 12; add.s64 %rd1672, %rd1670, %rd1671; ld.global.f32 %f10174, [%rd1672]; mul.ftz.f32 %f10175, %f9254, %f10173; mul.ftz.f32 %f10176, %f10175, %f10174; ld.global.f32 %f10177, [%rd1672+4]; mul.ftz.f32 %f10178, %f10173, %f9255; mul.ftz.f32 %f10179, %f10178, %f10177; ld.global.f32 %f10180, [%rd1672+8]; mul.ftz.f32 %f10181, %f10173, %f730; mul.ftz.f32 %f10182, %f10181, %f10180; mul.ftz.f32 %f1094, %f29137, %f10176; mul.ftz.f32 %f1095, %f29137, %f10179; mul.ftz.f32 %f1096, %f29137, %f10182; ld.global.u16 %rs15, [%rd72]; mov.b32 %r384, %f1094; setp.nan.ftz.f32 %p646, %f1094, %f1094; mov.f32 %f29144, %f29146; @%p646 bra $L__BB107_755; or.b32 %r2071, %r384, -2147483648; setp.ne.s32 %p647, %r2071, -2147483648; add.ftz.f32 %f10184, %f1094, %f1094; setp.eq.ftz.f32 %p648, %f10184, %f1094; and.pred %p649, %p648, %p647; mov.f32 %f29144, %f29146; @%p649 bra $L__BB107_755; shl.b32 %r2072, %r384, 1; setp.lt.u32 %p650, %r2072, -16777215; selp.f32 %f29144, %f1094, 0f00000000, %p650; $L__BB107_755: mov.b32 %r385, %f1095; setp.nan.ftz.f32 %p651, %f1095, %f1095; mov.f32 %f29145, %f29146; @%p651 bra $L__BB107_758; or.b32 %r2073, %r385, -2147483648; setp.ne.s32 %p652, %r2073, -2147483648; add.ftz.f32 %f10187, %f1095, %f1095; setp.eq.ftz.f32 %p653, %f10187, %f1095; and.pred %p654, %p653, %p652; mov.f32 %f29145, %f29146; @%p654 bra $L__BB107_758; shl.b32 %r2074, %r385, 1; setp.lt.u32 %p655, %r2074, -16777215; selp.f32 %f29145, %f1095, 0f00000000, %p655; $L__BB107_758: mov.b32 %r386, %f1096; setp.nan.ftz.f32 %p656, %f1096, %f1096; @%p656 bra $L__BB107_761; or.b32 %r2075, %r386, -2147483648; setp.ne.s32 %p657, %r2075, -2147483648; add.ftz.f32 %f10190, %f1096, %f1096; setp.eq.ftz.f32 %p658, %f10190, %f1096; and.pred %p659, %p658, %p657; @%p659 bra $L__BB107_761; shl.b32 %r2076, %r386, 1; setp.lt.u32 %p660, %r2076, -16777215; selp.f32 %f29146, %f1096, 0f00000000, %p660; $L__BB107_761: ld.const.f32 %f10191, [kernel_params+1580]; ld.const.f32 %f10192, [kernel_params+1584]; setp.gt.u16 %p661, %rs15, 1; selp.f32 %f1103, %f10192, %f10191, %p661; abs.ftz.f32 %f10193, %f29145; abs.ftz.f32 %f10194, %f29144; add.ftz.f32 %f10195, %f10194, %f10193; abs.ftz.f32 %f10196, %f29146; add.ftz.f32 %f1104, %f10196, %f10195; setp.leu.ftz.f32 %p662, %f1104, %f1103; @%p662 bra $L__BB107_763; div.approx.ftz.f32 %f10197, %f1103, %f1104; mul.ftz.f32 %f29144, %f29144, %f10197; mul.ftz.f32 %f29145, %f29145, %f10197; mul.ftz.f32 %f29146, %f29146, %f10197; $L__BB107_763: ld.const.v2.u64 {%rd1673, %rd1674}, [kernel_params+2448]; cvta.to.global.u64 %rd1677, %rd1673; add.s64 %rd1679, %rd1677, %rd1282; ld.global.u32 %rd1680, [%rd1679]; ld.const.s32 %rd1681, [kernel_params+1264]; mul.lo.s64 %rd163, %rd1681, %rd1680; cvta.to.global.u64 %rd1682, %rd1674; add.s64 %rd1683, %rd1682, %rd1282; ld.global.u32 %r387, [%rd1683]; ld.const.u32 %r2077, [kernel_params+1620]; setp.eq.s32 %p663, %r2077, 0; @%p663 bra $L__BB107_768; and.b32 %r2078, %r360, 536870912; setp.ne.s32 %p664, %r2078, 0; @%p664 bra $L__BB107_766; ld.const.s32 %rd1684, [kernel_params+1384]; add.s64 %rd1685, %rd163, %rd1684; cvta.to.global.u64 %rd1686, %rd1059; shl.b64 %rd1687, %rd1685, 2; add.s64 %rd1688, %rd1686, %rd1687; add.s64 %rd1689, %rd1688, 4; add.s64 %rd1690, %rd1688, 8; atom.global.add.f32 %f10198, [%rd1688], %f29144; atom.global.add.f32 %f10199, [%rd1689], %f29145; atom.global.add.f32 %f10200, [%rd1690], %f29146; $L__BB107_766: and.b32 %r2079, %r360, 1073741824; setp.eq.s32 %p665, %r2079, 0; @%p665 bra $L__BB107_768; bra.uni $L__BB107_767; $L__BB107_768: ld.const.u32 %r5652, [kernel_params+1260]; and.b32 %r2080, %r5652, 2; setp.eq.s32 %p666, %r2080, 0; @%p666 bra $L__BB107_770; ld.const.s32 %rd1698, [kernel_params+1268]; add.s64 %rd1699, %rd163, %rd1698; cvta.to.global.u64 %rd1700, %rd1059; shl.b64 %rd1701, %rd1699, 2; add.s64 %rd1702, %rd1700, %rd1701; add.s64 %rd1703, %rd1702, 4; add.s64 %rd1704, %rd1702, 8; atom.global.add.f32 %f10204, [%rd1702], %f29144; atom.global.add.f32 %f10205, [%rd1703], %f29145; atom.global.add.f32 %f10206, [%rd1704], %f29146; $L__BB107_770: ld.const.u32 %r390, [kernel_params+1400]; setp.eq.s32 %p667, %r390, -1; @%p667 bra $L__BB107_773; and.b32 %r2081, %r387, -1431655766; popc.b32 %r2082, %r2081; and.b32 %r2083, %r2082, 1; setp.eq.b32 %p668, %r2083, 1; mov.pred %p669, 0; xor.pred %p670, %p668, %p669; not.pred %p671, %p670; @%p671 bra $L__BB107_773; cvt.s64.s32 %rd1705, %r390; add.s64 %rd1706, %rd163, %rd1705; cvta.to.global.u64 %rd1707, %rd1059; shl.b64 %rd1708, %rd1706, 2; add.s64 %rd1709, %rd1707, %rd1708; add.s64 %rd1710, %rd1709, 4; add.s64 %rd1711, %rd1709, 8; add.s64 %rd1712, %rd1709, 12; add.ftz.f32 %f10207, %f29144, %f29144; atom.global.add.f32 %f10208, [%rd1709], %f10207; add.ftz.f32 %f10209, %f29145, %f29145; atom.global.add.f32 %f10210, [%rd1710], %f10209; add.ftz.f32 %f10211, %f29146, %f29146; atom.global.add.f32 %f10212, [%rd1711], %f10211; atom.global.add.f32 %f10213, [%rd1712], 0f00000000; bra.uni $L__BB107_773; $L__BB107_767: ld.const.s32 %rd1691, [kernel_params+1376]; add.s64 %rd1692, %rd163, %rd1691; cvta.to.global.u64 %rd1693, %rd1059; shl.b64 %rd1694, %rd1692, 2; add.s64 %rd1695, %rd1693, %rd1694; add.s64 %rd1696, %rd1695, 4; add.s64 %rd1697, %rd1695, 8; atom.global.add.f32 %f10201, [%rd1695], %f29144; atom.global.add.f32 %f10202, [%rd1696], %f29145; atom.global.add.f32 %f10203, [%rd1697], %f29146; ld.const.u32 %r5652, [kernel_params+1260]; $L__BB107_773: setp.eq.s32 %p672, %r5652, 0; @%p672 bra $L__BB107_794; ld.const.u32 %r5653, [kernel_params+1352]; and.b32 %r2084, %r360, 16777216; setp.eq.s32 %p673, %r2084, 0; ld.const.u32 %r393, [kernel_params+1428]; setp.eq.s32 %p674, %r393, -1; or.pred %p675, %p673, %p674; @%p675 bra $L__BB107_776; ld.const.u64 %rd1713, [kernel_params+2616]; cvta.to.global.u64 %rd1714, %rd1713; add.s64 %rd1716, %rd1714, %rd1671; ld.global.f32 %f10214, [%rd1716]; mul.ftz.f32 %f10215, %f29144, %f10214; ld.global.f32 %f10216, [%rd1716+4]; mul.ftz.f32 %f10217, %f29145, %f10216; ld.global.f32 %f10218, [%rd1716+8]; mul.ftz.f32 %f10219, %f29146, %f10218; cvta.to.global.u64 %rd1717, %rd1059; cvt.s64.s32 %rd1718, %r393; add.s64 %rd1719, %rd163, %rd1718; shl.b64 %rd1720, %rd1719, 2; add.s64 %rd1721, %rd1717, %rd1720; add.s64 %rd1722, %rd1721, 4; add.s64 %rd1723, %rd1721, 8; atom.global.add.f32 %f10220, [%rd1721], %f10215; atom.global.add.f32 %f10221, [%rd1722], %f10217; atom.global.add.f32 %f10222, [%rd1723], %f10219; $L__BB107_776: setp.eq.s32 %p676, %r5651, -1; ld.const.u32 %r394, [kernel_params+1444]; setp.eq.s32 %p677, %r394, -1; or.pred %p678, %p676, %p677; @%p678 bra $L__BB107_778; mul.lo.s32 %r2085, %r5651, 3; cvt.s64.s32 %rd1724, %r2085; add.s64 %rd1725, %rd163, %rd1724; cvt.s64.s32 %rd1726, %r394; add.s64 %rd1727, %rd1725, %rd1726; cvta.to.global.u64 %rd1728, %rd1059; shl.b64 %rd1729, %rd1727, 2; add.s64 %rd1730, %rd1728, %rd1729; add.s64 %rd1731, %rd1730, 4; add.s64 %rd1732, %rd1730, 8; atom.global.add.f32 %f10223, [%rd1730], %f29144; atom.global.add.f32 %f10224, [%rd1731], %f29145; atom.global.add.f32 %f10225, [%rd1732], %f29146; $L__BB107_778: and.b32 %r2086, %r360, 100663296; setp.eq.s32 %p679, %r2086, 0; @%p679 bra $L__BB107_792; ld.const.u8 %rs320, [kernel_params+2]; and.b16 %rs321, %rs320, 32; setp.eq.s16 %p680, %rs321, 0; mov.u32 %r5653, -1; @%p680 bra $L__BB107_792; bra.uni $L__BB107_780; $L__BB107_792: setp.eq.s32 %p691, %r5653, -1; @%p691 bra $L__BB107_794; cvt.s64.s32 %rd1754, %r5653; add.s64 %rd1755, %rd163, %rd1754; cvta.to.global.u64 %rd1756, %rd1059; shl.b64 %rd1757, %rd1755, 2; add.s64 %rd1758, %rd1756, %rd1757; add.s64 %rd1759, %rd1758, 4; add.s64 %rd1760, %rd1758, 8; atom.global.add.f32 %f10245, [%rd1758], %f29144; atom.global.add.f32 %f10246, [%rd1759], %f29145; atom.global.add.f32 %f10247, [%rd1760], %f29146; bra.uni $L__BB107_794; $L__BB107_780: and.b32 %r2088, %r360, 536870912; setp.ne.s32 %p681, %r2088, 0; @%p681 bra $L__BB107_794; and.b32 %r2089, %r360, 33554432; setp.eq.s32 %p682, %r2089, 0; @%p682 bra $L__BB107_788; bra.uni $L__BB107_782; $L__BB107_788: and.b32 %r2098, %r360, 67108864; setp.eq.s32 %p689, %r2098, 0; @%p689 bra $L__BB107_792; setp.eq.s16 %p690, %rs15, 1; @%p690 bra $L__BB107_791; bra.uni $L__BB107_790; $L__BB107_791: ld.const.u32 %r5653, [kernel_params+1348]; bra.uni $L__BB107_792; $L__BB107_794: cvt.u16.u32 %rs322, %r45; setp.lt.s16 %p692, %rs322, 0; mov.f32 %f29147, 0f00000000; @%p692 bra $L__BB107_796; ld.const.u64 %rd1761, [kernel_params+2576]; cvta.to.global.u64 %rd1762, %rd1761; add.s64 %rd1764, %rd1762, %rd1282; ld.global.f32 %f29147, [%rd1764]; $L__BB107_796: setp.eq.ftz.f32 %p693, %f29147, 0f00000000; mov.u16 %rs748, 0; @%p693 bra $L__BB107_4545; setp.eq.ftz.f32 %p694, %f29147, 0f3F800000; @%p694 bra $L__BB107_799; ld.const.u64 %rd1765, [kernel_params+2584]; cvta.to.global.u64 %rd1766, %rd1765; mul.lo.s64 %rd1767, %rd3, 12; add.s64 %rd1768, %rd1766, %rd1767; ld.global.f32 %f10249, [%rd1768]; rcp.approx.ftz.f32 %f10250, %f29147; mul.ftz.f32 %f10251, %f10250, %f10249; ld.global.f32 %f10252, [%rd1768+4]; mul.ftz.f32 %f10253, %f10250, %f10252; ld.global.f32 %f10254, [%rd1768+8]; mul.ftz.f32 %f10255, %f10250, %f10254; st.global.f32 [%rd1768], %f10251; st.global.f32 [%rd1768+4], %f10253; st.global.f32 [%rd1768+8], %f10255; $L__BB107_799: ld.global.u32 %r401, [%rd27]; and.b32 %r2099, %r401, 16384; setp.eq.s32 %p695, %r2099, 0; @%p695 bra $L__BB107_1257; ld.const.u32 %r402, [kernel_params+1256]; setp.eq.s32 %p696, %r402, 0; @%p696 bra $L__BB107_1257; ld.const.u64 %rd1769, [kernel_params+2448]; cvta.to.global.u64 %rd1770, %rd1769; add.s64 %rd1772, %rd1770, %rd1282; ld.global.u32 %rd1773, [%rd1772]; ld.const.s32 %rd1774, [kernel_params+1264]; mul.lo.s64 %rd164, %rd1774, %rd1773; and.b32 %r2100, %r401, 8192; setp.ne.s32 %p697, %r2100, 0; @%p697 bra $L__BB107_1146; ld.const.u64 %rd1775, [kernel_params+2456]; cvta.to.global.u64 %rd1776, %rd1775; add.s64 %rd1778, %rd1776, %rd1282; ld.global.u32 %r2101, [%rd1778]; setp.ne.s32 %p698, %r2101, 0; @%p698 bra $L__BB107_816; and.b32 %r2102, %r402, 1; setp.eq.b32 %p699, %r2102, 1; mov.pred %p700, 0; xor.pred %p701, %p699, %p700; not.pred %p702, %p701; @%p702 bra $L__BB107_808; cvta.to.local.u64 %rd1780, %rd1061; ld.local.v4.f32 {%f10256, %f10257, %f10258, %f10259}, [%rd1780]; ld.const.u32 %r2103, [kernel_params+16]; setp.eq.s32 %p703, %r2103, 2; @%p703 bra $L__BB107_806; bra.uni $L__BB107_805; $L__BB107_806: ld.const.v4.f32 {%f10271, %f10272, %f10273, %f10274}, [kernel_params+96]; ld.const.v4.f32 {%f10276, %f10277, %f10278, %f10279}, [kernel_params+112]; ld.const.v4.f32 {%f10281, %f10282, %f10283, %f10284}, [kernel_params+128]; sub.ftz.f32 %f10286, %f10256, %f10274; sub.ftz.f32 %f10287, %f10257, %f10279; sub.ftz.f32 %f10288, %f10258, %f10284; mul.ftz.f32 %f10289, %f10287, %f10287; fma.rn.ftz.f32 %f10290, %f10286, %f10286, %f10289; fma.rn.ftz.f32 %f10291, %f10288, %f10288, %f10290; sqrt.approx.ftz.f32 %f29148, %f10291; bra.uni $L__BB107_807; $L__BB107_782: ld.const.u64 %rd1733, [kernel_params+2600]; cvta.to.global.u64 %rd1734, %rd1733; add.s64 %rd1736, %rd1734, %rd1671; ld.global.f32 %f1111, [%rd1736]; ld.global.f32 %f1112, [%rd1736+4]; ld.global.f32 %f1113, [%rd1736+8]; ld.const.u64 %rd1737, [kernel_params+2608]; cvta.to.global.u64 %rd1738, %rd1737; add.s64 %rd1739, %rd1738, %rd1671; ld.global.f32 %f1114, [%rd1739]; ld.global.f32 %f1115, [%rd1739+4]; ld.global.f32 %f1116, [%rd1739+8]; ld.const.u32 %r2090, [kernel_params+1324]; ld.const.u32 %r2091, [kernel_params+1340]; setp.eq.s16 %p683, %rs15, 1; selp.b32 %r395, %r2091, %r2090, %p683; setp.eq.s32 %p684, %r395, -1; @%p684 bra $L__BB107_784; cvt.s64.s32 %rd1740, %r395; add.s64 %rd1741, %rd163, %rd1740; cvta.to.global.u64 %rd1742, %rd1059; shl.b64 %rd1743, %rd1741, 2; add.s64 %rd1744, %rd1742, %rd1743; add.s64 %rd1745, %rd1744, 4; add.s64 %rd1746, %rd1744, 8; mul.ftz.f32 %f10226, %f29144, %f1114; atom.global.add.f32 %f10227, [%rd1744], %f10226; mul.ftz.f32 %f10228, %f29145, %f1115; atom.global.add.f32 %f10229, [%rd1745], %f10228; mul.ftz.f32 %f10230, %f29146, %f1116; atom.global.add.f32 %f10231, [%rd1746], %f10230; $L__BB107_784: ld.const.u32 %r2092, [kernel_params+1328]; ld.const.u32 %r2093, [kernel_params+1344]; selp.b32 %r396, %r2093, %r2092, %p683; setp.eq.s32 %p686, %r396, -1; @%p686 bra $L__BB107_786; mov.f32 %f10232, 0f3F800000; sub.ftz.f32 %f10233, %f10232, %f1111; sub.ftz.f32 %f10234, %f10233, %f1114; sub.ftz.f32 %f10235, %f10232, %f1112; sub.ftz.f32 %f10236, %f10235, %f1115; sub.ftz.f32 %f10237, %f10232, %f1113; sub.ftz.f32 %f10238, %f10237, %f1116; cvt.s64.s32 %rd1747, %r396; add.s64 %rd1748, %rd163, %rd1747; cvta.to.global.u64 %rd1749, %rd1059; shl.b64 %rd1750, %rd1748, 2; add.s64 %rd1751, %rd1749, %rd1750; mul.ftz.f32 %f10239, %f29144, %f10234; mul.ftz.f32 %f10240, %f29145, %f10236; mul.ftz.f32 %f10241, %f29146, %f10238; add.s64 %rd1752, %rd1751, 4; add.s64 %rd1753, %rd1751, 8; atom.global.add.f32 %f10242, [%rd1751], %f10239; atom.global.add.f32 %f10243, [%rd1752], %f10240; atom.global.add.f32 %f10244, [%rd1753], %f10241; $L__BB107_786: ld.const.u32 %r2095, [kernel_params+1320]; ld.const.u32 %r2096, [kernel_params+1336]; selp.b32 %r397, %r2096, %r2095, %p683; setp.eq.s32 %p688, %r397, -1; @%p688 bra $L__BB107_792; mul.ftz.f32 %f29144, %f29144, %f1111; mul.ftz.f32 %f29145, %f29145, %f1112; mul.ftz.f32 %f29146, %f29146, %f1113; mov.u32 %r5653, %r397; bra.uni $L__BB107_792; $L__BB107_790: ld.const.u32 %r5653, [kernel_params+1332]; bra.uni $L__BB107_792; $L__BB107_805: ld.const.v4.f32 {%f10260, %f10261, %f10262, %f10263}, [kernel_params+720]; mul.ftz.f32 %f10268, %f10257, %f10261; fma.rn.ftz.f32 %f10269, %f10256, %f10260, %f10268; fma.rn.ftz.f32 %f10270, %f10258, %f10262, %f10269; add.ftz.f32 %f29148, %f10263, %f10270; $L__BB107_807: ld.const.s32 %rd1781, [kernel_params+1272]; add.s64 %rd1782, %rd164, %rd1781; cvta.to.global.u64 %rd1783, %rd1059; shl.b64 %rd1784, %rd1782, 2; add.s64 %rd1785, %rd1783, %rd1784; st.global.f32 [%rd1785], %f29148; $L__BB107_808: and.b32 %r2104, %r402, 32; setp.eq.s32 %p704, %r2104, 0; @%p704 bra $L__BB107_812; ld.local.u32 %r403, [%rd5+-12]; setp.eq.s32 %p705, %r403, -1; mov.f32 %f29149, 0f00000000; @%p705 bra $L__BB107_811; ld.const.u64 %rd1786, [kernel_params+2160]; cvta.to.global.u64 %rd1787, %rd1786; mul.wide.s32 %rd1788, %r403, 208; add.s64 %rd1789, %rd1787, %rd1788; ld.global.f32 %f29149, [%rd1789+100]; $L__BB107_811: ld.const.s32 %rd1790, [kernel_params+1300]; add.s64 %rd1791, %rd164, %rd1790; cvta.to.global.u64 %rd1792, %rd1059; shl.b64 %rd1793, %rd1791, 2; add.s64 %rd1794, %rd1792, %rd1793; st.global.f32 [%rd1794], %f29149; $L__BB107_812: and.b32 %r2105, %r402, 64; setp.eq.s32 %p706, %r2105, 0; @%p706 bra $L__BB107_814; ld.local.u32 %rd1795, [%rd5+-40]; shl.b64 %rd1796, %rd1795, 5; and.b64 %rd1797, %rd1796, 134217696; add.s64 %rd1798, %rd26, %rd1797; ld.global.u32 %r2106, [%rd1798+20]; cvt.rn.f32.s32 %f10293, %r2106; ld.const.s32 %rd1799, [kernel_params+1304]; add.s64 %rd1800, %rd164, %rd1799; cvta.to.global.u64 %rd1801, %rd1059; shl.b64 %rd1802, %rd1800, 2; add.s64 %rd1803, %rd1801, %rd1802; st.global.f32 [%rd1803], %f10293; $L__BB107_814: and.b32 %r2107, %r402, 2; setp.eq.s32 %p707, %r2107, 0; @%p707 bra $L__BB107_816; cvta.to.local.u64 %rd1805, %rd1061; ld.local.v4.f32 {%f10294, %f10295, %f10296, %f10297}, [%rd1805]; ld.const.s32 %rd1806, [kernel_params+1276]; add.s64 %rd1807, %rd164, %rd1806; cvta.to.global.u64 %rd1808, %rd1059; shl.b64 %rd1809, %rd1807, 2; add.s64 %rd1810, %rd1808, %rd1809; st.global.f32 [%rd1810], %f10294; st.global.f32 [%rd1810+4], %f10295; st.global.f32 [%rd1810+8], %f10296; $L__BB107_816: and.b32 %r2108, %r5618, 512; setp.eq.s32 %p708, %r2108, 0; ld.const.f32 %f1133, [kernel_params+1364]; setp.eq.ftz.f32 %p709, %f1133, 0f00000000; or.pred %p710, %p708, %p709; @%p710 bra $L__BB107_820; and.b32 %r2109, %r5618, 524288; setp.ne.s32 %p711, %r2109, 0; mov.f32 %f10303, 0f3F800000; mov.f32 %f29150, %f10303; mov.f32 %f29151, %f10303; mov.f32 %f29152, %f10303; @%p711 bra $L__BB107_819; ld.local.v2.f32 {%f29151, %f29152}, [%rd5+208]; ld.local.f32 %f29150, [%rd5+204]; $L__BB107_819: sub.ftz.f32 %f10307, %f10303, %f29150; sub.ftz.f32 %f10308, %f10303, %f29151; sub.ftz.f32 %f10309, %f10303, %f29152; cvt.ftz.sat.f32.f32 %f10310, %f10307; cvt.ftz.sat.f32.f32 %f10311, %f10308; cvt.ftz.sat.f32.f32 %f10312, %f10309; add.ftz.f32 %f10313, %f10310, %f10311; add.ftz.f32 %f10314, %f10312, %f10313; mul.ftz.f32 %f10315, %f10314, 0f3EAAAAAB; setp.ltu.ftz.f32 %p712, %f10315, %f1133; @%p712 bra $L__BB107_1146; $L__BB107_820: and.b32 %r2110, %r402, 4; setp.eq.s32 %p713, %r2110, 0; @%p713 bra $L__BB107_841; ld.local.u32 %r404, [%rd5+172]; setp.lt.s32 %p714, %r404, 1; mov.f32 %f29156, 0f00000000; mov.f32 %f29157, %f29156; mov.f32 %f29158, %f29156; @%p714 bra $L__BB107_837; add.s32 %r2112, %r404, -1; and.b32 %r5658, %r404, 3; setp.lt.u32 %p715, %r2112, 3; mov.f32 %f29158, 0f00000000; mov.u32 %r5656, 0; mov.f32 %f29157, %f29158; mov.f32 %f29156, %f29158; @%p715 bra $L__BB107_833; sub.s32 %r5655, %r404, %r5658; mov.f32 %f29158, 0f00000000; mov.u32 %r5656, 0; mov.f32 %f29157, %f29158; mov.f32 %f29156, %f29158; $L__BB107_824: cvt.s64.s32 %rd165, %r5656; mul.wide.s32 %rd1811, %r5656, 80; add.s64 %rd1812, %rd2, %rd1811; add.s64 %rd166, %rd1812, 316; ld.local.u32 %r2114, [%rd1812+316]; setp.gt.s32 %p716, %r2114, 34; @%p716 bra $L__BB107_826; ld.local.v4.f32 {%f10326, %f10327, %f10328, %f10329}, [%rd166+-12]; add.ftz.f32 %f10333, %f10326, %f10327; add.ftz.f32 %f10334, %f10328, %f10333; mul.ftz.f32 %f10335, %f10334, 0f3EAAAAAB; abs.ftz.f32 %f10336, %f10335; ld.local.f32 %f10337, [%rd166+8]; ld.local.f32 %f10338, [%rd166+12]; ld.local.f32 %f10339, [%rd166+16]; fma.rn.ftz.f32 %f29156, %f10336, %f10337, %f29156; fma.rn.ftz.f32 %f29157, %f10336, %f10338, %f29157; fma.rn.ftz.f32 %f29158, %f10336, %f10339, %f29158; $L__BB107_826: ld.local.u32 %r2115, [%rd166+80]; setp.gt.s32 %p717, %r2115, 34; @%p717 bra $L__BB107_828; ld.local.v4.f32 {%f10340, %f10341, %f10342, %f10343}, [%rd166+68]; add.ftz.f32 %f10347, %f10340, %f10341; add.ftz.f32 %f10348, %f10342, %f10347; mul.ftz.f32 %f10349, %f10348, 0f3EAAAAAB; abs.ftz.f32 %f10350, %f10349; ld.local.f32 %f10351, [%rd166+88]; ld.local.f32 %f10352, [%rd166+92]; ld.local.f32 %f10353, [%rd166+96]; fma.rn.ftz.f32 %f29156, %f10350, %f10351, %f29156; fma.rn.ftz.f32 %f29157, %f10350, %f10352, %f29157; fma.rn.ftz.f32 %f29158, %f10350, %f10353, %f29158; $L__BB107_828: ld.local.u32 %r2116, [%rd166+160]; setp.gt.s32 %p718, %r2116, 34; @%p718 bra $L__BB107_830; ld.local.v4.f32 {%f10354, %f10355, %f10356, %f10357}, [%rd166+148]; add.ftz.f32 %f10361, %f10354, %f10355; add.ftz.f32 %f10362, %f10356, %f10361; mul.ftz.f32 %f10363, %f10362, 0f3EAAAAAB; abs.ftz.f32 %f10364, %f10363; ld.local.f32 %f10365, [%rd166+168]; ld.local.f32 %f10366, [%rd166+172]; ld.local.f32 %f10367, [%rd166+176]; fma.rn.ftz.f32 %f29156, %f10364, %f10365, %f29156; fma.rn.ftz.f32 %f29157, %f10364, %f10366, %f29157; fma.rn.ftz.f32 %f29158, %f10364, %f10367, %f29158; $L__BB107_830: ld.local.u32 %r2117, [%rd166+240]; setp.gt.s32 %p719, %r2117, 34; @%p719 bra $L__BB107_832; ld.local.v4.f32 {%f10368, %f10369, %f10370, %f10371}, [%rd166+228]; add.ftz.f32 %f10375, %f10368, %f10369; add.ftz.f32 %f10376, %f10370, %f10375; mul.ftz.f32 %f10377, %f10376, 0f3EAAAAAB; abs.ftz.f32 %f10378, %f10377; ld.local.f32 %f10379, [%rd166+248]; ld.local.f32 %f10380, [%rd166+252]; ld.local.f32 %f10381, [%rd166+256]; fma.rn.ftz.f32 %f29156, %f10378, %f10379, %f29156; fma.rn.ftz.f32 %f29157, %f10378, %f10380, %f29157; fma.rn.ftz.f32 %f29158, %f10378, %f10381, %f29158; $L__BB107_832: cvt.u32.u64 %r2118, %rd165; add.s32 %r5656, %r2118, 4; add.s32 %r5655, %r5655, -4; setp.ne.s32 %p720, %r5655, 0; @%p720 bra $L__BB107_824; $L__BB107_833: setp.eq.s32 %p721, %r5658, 0; @%p721 bra $L__BB107_837; $L__BB107_834: .pragma "nounroll"; cvt.s64.s32 %rd167, %r5656; mul.wide.s32 %rd1813, %r5656, 80; add.s64 %rd1814, %rd2, %rd1813; add.s64 %rd168, %rd1814, 316; ld.local.u32 %r2119, [%rd1814+316]; setp.gt.s32 %p722, %r2119, 34; @%p722 bra $L__BB107_836; ld.local.f32 %f10382, [%rd168+-8]; ld.local.f32 %f10383, [%rd168+-12]; add.ftz.f32 %f10384, %f10383, %f10382; ld.local.f32 %f10385, [%rd168+-4]; add.ftz.f32 %f10386, %f10385, %f10384; mul.ftz.f32 %f10387, %f10386, 0f3EAAAAAB; abs.ftz.f32 %f10388, %f10387; ld.local.f32 %f10389, [%rd168+8]; ld.local.f32 %f10390, [%rd168+12]; ld.local.f32 %f10391, [%rd168+16]; fma.rn.ftz.f32 %f29156, %f10388, %f10389, %f29156; fma.rn.ftz.f32 %f29157, %f10388, %f10390, %f29157; fma.rn.ftz.f32 %f29158, %f10388, %f10391, %f29158; $L__BB107_836: cvt.u32.u64 %r2120, %rd167; add.s32 %r5656, %r2120, 1; add.s32 %r5658, %r5658, -1; setp.ne.s32 %p723, %r5658, 0; @%p723 bra $L__BB107_834; $L__BB107_837: setp.eq.ftz.f32 %p724, %f29157, 0f00000000; setp.eq.ftz.f32 %p725, %f29156, 0f00000000; and.pred %p726, %p725, %p724; setp.eq.ftz.f32 %p727, %f29158, 0f00000000; and.pred %p728, %p727, %p726; @%p728 bra $L__BB107_839; bra.uni $L__BB107_838; $L__BB107_839: ld.local.v2.f32 {%f29184, %f29185}, [%rd5+-72]; ld.local.f32 %f29183, [%rd5+-76]; bra.uni $L__BB107_840; $L__BB107_838: mul.ftz.f32 %f10392, %f29157, %f29157; fma.rn.ftz.f32 %f10393, %f29156, %f29156, %f10392; fma.rn.ftz.f32 %f10394, %f29158, %f29158, %f10393; sqrt.approx.ftz.f32 %f10395, %f10394; div.approx.ftz.f32 %f29185, %f29158, %f10395; div.approx.ftz.f32 %f29184, %f29157, %f10395; div.approx.ftz.f32 %f29183, %f29156, %f10395; $L__BB107_840: ld.const.s32 %rd1815, [kernel_params+1280]; add.s64 %rd1816, %rd164, %rd1815; cvta.to.global.u64 %rd1817, %rd1059; shl.b64 %rd1818, %rd1816, 2; add.s64 %rd1819, %rd1817, %rd1818; add.s64 %rd1820, %rd1819, 4; add.s64 %rd1821, %rd1819, 8; atom.global.add.f32 %f10398, [%rd1819], %f29183; atom.global.add.f32 %f10399, [%rd1820], %f29184; atom.global.add.f32 %f10400, [%rd1821], %f29185; $L__BB107_841: and.b32 %r2121, %r402, 8; setp.eq.s32 %p729, %r2121, 0; @%p729 bra $L__BB107_886; ld.local.u32 %r416, [%rd5+172]; setp.lt.s32 %p730, %r416, 1; mov.f32 %f29201, 0f00000000; mov.f32 %f29189, %f29201; mov.f32 %f29190, %f29201; @%p730 bra $L__BB107_883; and.b32 %r417, %r416, 1; setp.eq.s32 %p731, %r416, 1; mov.f32 %f29190, 0f00000000; mov.u32 %r5661, 0; mov.f32 %f29189, %f29190; @%p731 bra $L__BB107_870; sub.s32 %r5660, %r416, %r417; mov.f32 %f29190, 0f00000000; mov.u32 %r5661, 0; mov.f32 %f29189, %f29190; $L__BB107_845: cvt.s64.s32 %rd169, %r5661; mul.wide.s32 %rd1822, %r5661, 80; add.s64 %rd1823, %rd2, %rd1822; add.s64 %rd170, %rd1823, 316; ld.local.u32 %r421, [%rd1823+316]; setp.gt.s32 %p732, %r421, 31; @%p732 bra $L__BB107_857; ld.local.v4.f32 {%f10408, %f10409, %f10410, %f10411}, [%rd170+-12]; setp.eq.s32 %p733, %r421, 3; @%p733 bra $L__BB107_855; bra.uni $L__BB107_847; $L__BB107_855: ld.local.f32 %f10421, [%rd170+20]; mul.ftz.f32 %f10422, %f10421, %f10421; mul.ftz.f32 %f29188, %f10422, %f10422; bra.uni $L__BB107_856; $L__BB107_847: setp.eq.s32 %p734, %r421, 5; @%p734 bra $L__BB107_854; bra.uni $L__BB107_848; $L__BB107_854: ld.local.f32 %f10419, [%rd170+20]; mul.ftz.f32 %f10420, %f10419, %f10419; mul.ftz.f32 %f29188, %f10420, %f10420; bra.uni $L__BB107_856; $L__BB107_848: add.s32 %r2124, %r421, -2; setp.lt.u32 %p735, %r2124, 7; mov.f32 %f29188, 0f00000000; @%p735 bra $L__BB107_856; setp.gt.u32 %p736, %r421, 31; @%p736 bra $L__BB107_851; mov.u32 %r2126, 1; shl.b32 %r2127, %r2126, %r421; and.b32 %r2128, %r2127, -2145385984; setp.ne.s32 %p737, %r2128, 0; @%p737 bra $L__BB107_856; $L__BB107_851: add.s32 %r2129, %r421, -10; setp.lt.u32 %p738, %r2129, 7; @%p738 bra $L__BB107_853; add.s32 %r2130, %r421, -22; setp.gt.u32 %p739, %r2130, 2; setp.ne.s32 %p740, %r421, 27; mov.f32 %f29188, 0f3F800000; and.pred %p741, %p739, %p740; @%p741 bra $L__BB107_856; $L__BB107_853: ld.local.v2.f32 {%f10415, %f10416}, [%rd170+20]; mul.ftz.f32 %f29188, %f10415, %f10416; $L__BB107_856: add.ftz.f32 %f10423, %f10408, %f10409; add.ftz.f32 %f10424, %f10410, %f10423; mul.ftz.f32 %f10425, %f10424, 0f3EAAAAAB; abs.ftz.f32 %f10426, %f10425; mov.f32 %f10427, 0f00000000; max.ftz.f32 %f10428, %f29188, %f10427; sqrt.approx.ftz.f32 %f10429, %f10428; sqrt.approx.ftz.f32 %f10430, %f10429; fma.rn.ftz.f32 %f29189, %f10426, %f10430, %f29189; add.ftz.f32 %f29190, %f29190, %f10426; $L__BB107_857: ld.local.u32 %r422, [%rd170+80]; setp.gt.s32 %p742, %r422, 31; @%p742 bra $L__BB107_869; ld.local.v4.f32 {%f10431, %f10432, %f10433, %f10434}, [%rd170+68]; setp.eq.s32 %p743, %r422, 3; @%p743 bra $L__BB107_867; bra.uni $L__BB107_859; $L__BB107_867: ld.local.f32 %f10444, [%rd170+100]; mul.ftz.f32 %f10445, %f10444, %f10444; mul.ftz.f32 %f29191, %f10445, %f10445; bra.uni $L__BB107_868; $L__BB107_859: setp.eq.s32 %p744, %r422, 5; @%p744 bra $L__BB107_866; bra.uni $L__BB107_860; $L__BB107_866: ld.local.f32 %f10442, [%rd170+100]; mul.ftz.f32 %f10443, %f10442, %f10442; mul.ftz.f32 %f29191, %f10443, %f10443; bra.uni $L__BB107_868; $L__BB107_860: add.s32 %r2131, %r422, -2; setp.lt.u32 %p745, %r2131, 7; mov.f32 %f29191, 0f00000000; @%p745 bra $L__BB107_868; setp.gt.u32 %p746, %r422, 31; @%p746 bra $L__BB107_863; mov.u32 %r2133, 1; shl.b32 %r2134, %r2133, %r422; and.b32 %r2135, %r2134, -2145385984; setp.ne.s32 %p747, %r2135, 0; @%p747 bra $L__BB107_868; $L__BB107_863: add.s32 %r2136, %r422, -10; setp.lt.u32 %p748, %r2136, 7; @%p748 bra $L__BB107_865; add.s32 %r2137, %r422, -22; setp.gt.u32 %p749, %r2137, 2; setp.ne.s32 %p750, %r422, 27; mov.f32 %f29191, 0f3F800000; and.pred %p751, %p749, %p750; @%p751 bra $L__BB107_868; $L__BB107_865: ld.local.v2.f32 {%f10438, %f10439}, [%rd170+100]; mul.ftz.f32 %f29191, %f10438, %f10439; $L__BB107_868: add.ftz.f32 %f10446, %f10431, %f10432; add.ftz.f32 %f10447, %f10433, %f10446; mul.ftz.f32 %f10448, %f10447, 0f3EAAAAAB; abs.ftz.f32 %f10449, %f10448; mov.f32 %f10450, 0f00000000; max.ftz.f32 %f10451, %f29191, %f10450; sqrt.approx.ftz.f32 %f10452, %f10451; sqrt.approx.ftz.f32 %f10453, %f10452; fma.rn.ftz.f32 %f29189, %f10449, %f10453, %f29189; add.ftz.f32 %f29190, %f29190, %f10449; $L__BB107_869: cvt.u32.u64 %r2138, %rd169; add.s32 %r5661, %r2138, 2; add.s32 %r5660, %r5660, -2; setp.ne.s32 %p752, %r5660, 0; @%p752 bra $L__BB107_845; $L__BB107_870: setp.eq.s32 %p753, %r417, 0; @%p753 bra $L__BB107_883; mul.wide.s32 %rd1824, %r5661, 80; add.s64 %rd1825, %rd2, %rd1824; add.s64 %rd171, %rd1825, 316; ld.local.u32 %r426, [%rd1825+316]; setp.gt.s32 %p754, %r426, 31; @%p754 bra $L__BB107_883; ld.local.f32 %f10454, [%rd171+-8]; ld.local.f32 %f10455, [%rd171+-12]; add.ftz.f32 %f10456, %f10455, %f10454; ld.local.f32 %f10457, [%rd171+-4]; add.ftz.f32 %f1224, %f10457, %f10456; setp.eq.s32 %p755, %r426, 3; @%p755 bra $L__BB107_881; bra.uni $L__BB107_873; $L__BB107_881: ld.local.f32 %f10467, [%rd171+20]; mul.ftz.f32 %f10468, %f10467, %f10467; mul.ftz.f32 %f29198, %f10468, %f10468; bra.uni $L__BB107_882; $L__BB107_873: setp.eq.s32 %p756, %r426, 5; @%p756 bra $L__BB107_880; bra.uni $L__BB107_874; $L__BB107_880: ld.local.f32 %f10465, [%rd171+20]; mul.ftz.f32 %f10466, %f10465, %f10465; mul.ftz.f32 %f29198, %f10466, %f10466; bra.uni $L__BB107_882; $L__BB107_874: add.s32 %r2139, %r426, -2; setp.lt.u32 %p757, %r2139, 7; mov.f32 %f29198, 0f00000000; @%p757 bra $L__BB107_882; setp.gt.u32 %p758, %r426, 31; @%p758 bra $L__BB107_877; mov.u32 %r2141, 1; shl.b32 %r2142, %r2141, %r426; and.b32 %r2143, %r2142, -2145385984; setp.ne.s32 %p759, %r2143, 0; @%p759 bra $L__BB107_882; $L__BB107_877: add.s32 %r2144, %r426, -10; setp.lt.u32 %p760, %r2144, 7; @%p760 bra $L__BB107_879; add.s32 %r2145, %r426, -22; setp.gt.u32 %p761, %r2145, 2; setp.ne.s32 %p762, %r426, 27; mov.f32 %f29198, 0f3F800000; and.pred %p763, %p761, %p762; @%p763 bra $L__BB107_882; $L__BB107_879: ld.local.v2.f32 {%f10461, %f10462}, [%rd171+20]; mul.ftz.f32 %f29198, %f10461, %f10462; $L__BB107_882: mul.ftz.f32 %f10469, %f1224, 0f3EAAAAAB; abs.ftz.f32 %f10470, %f10469; mov.f32 %f10471, 0f00000000; max.ftz.f32 %f10472, %f29198, %f10471; sqrt.approx.ftz.f32 %f10473, %f10472; sqrt.approx.ftz.f32 %f10474, %f10473; fma.rn.ftz.f32 %f29189, %f10470, %f10474, %f29189; add.ftz.f32 %f29190, %f29190, %f10470; $L__BB107_883: setp.leu.ftz.f32 %p764, %f29190, 0f00000000; @%p764 bra $L__BB107_885; div.approx.ftz.f32 %f29201, %f29189, %f29190; $L__BB107_885: ld.const.s32 %rd1826, [kernel_params+1284]; add.s64 %rd1827, %rd164, %rd1826; cvta.to.global.u64 %rd1828, %rd1059; shl.b64 %rd1829, %rd1827, 2; add.s64 %rd1830, %rd1828, %rd1829; atom.global.add.f32 %f10476, [%rd1830], %f29201; $L__BB107_886: and.b32 %r2146, %r402, 16; setp.eq.s32 %p765, %r2146, 0; @%p765 bra $L__BB107_924; ld.local.u32 %r427, [%rd5+-12]; ld.local.s32 %rd172, [%rd5+-28]; setp.eq.s32 %p766, %r427, -1; mov.u32 %r448, -1; mov.u32 %r445, 0; mov.u32 %r5669, %r445; mov.u32 %r5670, %r445; @%p766 bra $L__BB107_902; ld.const.u64 %rd1831, [kernel_params+2160]; cvta.to.global.u64 %rd1832, %rd1831; mul.wide.s32 %rd1833, %r427, 208; add.s64 %rd1834, %rd1832, %rd1833; ld.global.u32 %r428, [%rd1834+164]; ld.local.u32 %r2152, [%rd5+-24]; and.b32 %r2153, %r2152, 1; setp.eq.b32 %p767, %r2153, 1; mov.pred %p768, 0; xor.pred %p769, %p767, %p768; not.pred %p770, %p769; mov.u32 %r5663, 0; @%p770 bra $L__BB107_892; cvt.u32.u64 %r2155, %rd172; setp.eq.s32 %p771, %r2155, -1; mov.u32 %r5662, -1; @%p771 bra $L__BB107_891; ld.const.u64 %rd1835, [kernel_params+2240]; cvta.to.global.u64 %rd1836, %rd1835; shl.b64 %rd1837, %rd172, 2; add.s64 %rd1838, %rd1836, %rd1837; ld.global.u32 %r5662, [%rd1838]; $L__BB107_891: setp.ne.s32 %p772, %r5662, -1; selp.u32 %r5663, 1, 0, %p772; $L__BB107_892: ld.const.u64 %rd1839, [kernel_params+2312]; cvta.to.global.u64 %rd173, %rd1839; add.s32 %r5665, %r5663, %r428; mul.wide.u32 %rd1840, %r5665, 16; add.s64 %rd1841, %rd173, %rd1840; add.s64 %rd4766, %rd1841, 14; ld.global.u16 %rs616, [%rd1841+12]; ld.global.u32 %r5664, [%rd1841+8]; ld.global.u64 %rd4764, [%rd1841]; setp.eq.s64 %p773, %rd4764, 3; add.s64 %rd4765, %rd1841, 15; @%p773 bra $L__BB107_898; $L__BB107_893: setp.eq.s64 %p774, %rd4764, 0; @%p774 bra $L__BB107_895; add.s32 %r5665, %r5665, 2; bra.uni $L__BB107_896; $L__BB107_895: setp.eq.s16 %p775, %rs616, 0; mov.u32 %r445, 0; mov.u32 %r5665, %r5664; mov.u32 %r5669, %r445; mov.u32 %r5670, %r445; @%p775 bra $L__BB107_902; $L__BB107_896: cvt.u64.u32 %rd178, %r5665; mul.wide.u32 %rd1842, %r5665, 16; add.s64 %rd1843, %rd173, %rd1842; ld.global.u16 %rs616, [%rd1843+12]; ld.global.u32 %r5664, [%rd1843+8]; ld.global.u64 %rd4764, [%rd1843]; setp.ne.s64 %p776, %rd4764, 3; @%p776 bra $L__BB107_893; shl.b64 %rd1844, %rd178, 4; add.s64 %rd1845, %rd173, %rd1844; add.s64 %rd4766, %rd1845, 14; add.s64 %rd4765, %rd1845, 15; $L__BB107_898: ld.global.u8 %rs20, [%rd4766]; ld.global.u8 %rs21, [%rd4765]; cvt.u32.u64 %r2160, %rd172; setp.ne.s32 %p777, %r2160, -1; @%p777 bra $L__BB107_901; mov.u32 %r445, 0; add.s16 %rs324, %rs616, -1; setp.lt.u16 %p778, %rs324, 2; @%p778 bra $L__BB107_901; setp.ne.s16 %p779, %rs616, 1024; mov.u32 %r5669, %r445; mov.u32 %r5670, %r445; @%p779 bra $L__BB107_902; $L__BB107_901: setp.eq.s16 %p780, %rs616, 0; selp.b32 %r448, -1, %r5664, %p780; cvt.u32.u16 %r445, %rs616; cvt.u32.u16 %r2165, %rs21; and.b32 %r5670, %r2165, 255; cvt.u32.u16 %r2166, %rs20; and.b32 %r5669, %r2166, 255; $L__BB107_902: setp.eq.s32 %p781, %r448, -1; mov.f32 %f29208, 0f00000000; mov.f32 %f29209, %f29208; mov.f32 %f29212, %f29208; @%p781 bra $L__BB107_923; ld.local.u32 %r451, [%rd5+-24]; and.b32 %r2167, %r451, 1; setp.eq.b32 %p782, %r2167, 1; mov.pred %p783, 0; xor.pred %p784, %p782, %p783; not.pred %p785, %p784; @%p785 bra $L__BB107_915; bra.uni $L__BB107_904; $L__BB107_915: and.b32 %r2188, %r451, 6; setp.eq.s32 %p792, %r2188, 0; @%p792 bra $L__BB107_920; and.b32 %r2189, %r445, 768; setp.eq.s32 %p793, %r2189, 0; @%p793 bra $L__BB107_918; cvt.u32.u64 %r2190, %rd172; ld.const.u64 %rd1872, [kernel_params+2264]; cvta.to.global.u64 %rd1873, %rd1872; mul.wide.s32 %rd1874, %r2190, 16; add.s64 %rd1875, %rd1873, %rd1874; shr.s32 %r2191, %r451, 7; ld.global.u32 %r2192, [%rd1875+4]; add.s32 %r2193, %r2191, %r2192; ld.const.u64 %rd1876, [kernel_params+2328]; cvta.to.global.u64 %rd1877, %rd1876; add.s32 %r2194, %r2193, %r448; mul.wide.s32 %rd1878, %r2194, 8; add.s64 %rd1879, %rd1877, %rd1878; ld.global.v2.f32 {%f10505, %f10506}, [%rd1879]; ld.global.v2.f32 {%f10509, %f10510}, [%rd1879+8]; ld.local.f32 %f10513, [%rd5+-20]; mov.f32 %f10514, 0f3F800000; sub.ftz.f32 %f10515, %f10514, %f10513; mul.ftz.f32 %f10516, %f10505, %f10515; mul.ftz.f32 %f10517, %f10506, %f10515; fma.rn.ftz.f32 %f29209, %f10510, %f10513, %f10517; fma.rn.ftz.f32 %f29208, %f10509, %f10513, %f10516; bra.uni $L__BB107_922; $L__BB107_904: cvt.u32.u64 %r2169, %rd172; setp.eq.s32 %p786, %r2169, -1; mov.u32 %r5672, -1; @%p786 bra $L__BB107_906; ld.const.u64 %rd1846, [kernel_params+2240]; cvta.to.global.u64 %rd1847, %rd1846; shl.b64 %rd1848, %rd172, 2; add.s64 %rd1849, %rd1847, %rd1848; ld.global.u32 %r5672, [%rd1849]; $L__BB107_906: setp.eq.s32 %p787, %r5672, -1; @%p787 bra $L__BB107_908; mov.u64 %rd1850, 0; { // callseq 328, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r445; st.param.b32 [param1+4], %r5669; st.param.b32 [param1+8], %r5670; st.param.b32 [param1+12], %r448; .param .b64 param2; st.param.b64 [param2+0], %rd1850; .param .b64 param3; st.param.b64 [param3+0], %rd1850; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f29208, [retval0+0]; ld.param.f32 %f29209, [retval0+4]; } // callseq 328 bra.uni $L__BB107_922; $L__BB107_920: and.b32 %r2199, %r451, 8; setp.eq.s32 %p796, %r2199, 0; setp.ne.s32 %p797, %r445, 8; mov.f32 %f29208, 0f00000000; or.pred %p798, %p797, %p796; mov.f32 %f29209, %f29208; @%p798 bra $L__BB107_922; cvt.u32.u64 %r2200, %rd172; ld.const.u64 %rd1884, [kernel_params+2328]; cvta.to.global.u64 %rd1885, %rd1884; add.s32 %r2201, %r2200, %r448; mul.wide.s32 %rd1886, %r2201, 8; add.s64 %rd1887, %rd1885, %rd1886; ld.global.v2.f32 {%f29208, %f29209}, [%rd1887]; bra.uni $L__BB107_922; $L__BB107_908: and.b32 %r2170, %r445, 56; setp.eq.s32 %p788, %r2170, 0; @%p788 bra $L__BB107_913; and.b32 %r2171, %r445, 24; setp.eq.s32 %p789, %r2171, 0; @%p789 bra $L__BB107_911; ld.const.u64 %rd1852, [kernel_params+2232]; cvta.to.global.u64 %rd1853, %rd1852; shl.b64 %rd1854, %rd172, 4; add.s64 %rd1855, %rd1853, %rd1854; ld.global.v4.u32 {%r2172, %r2173, %r2174, %r2175}, [%rd1855]; ld.const.u64 %rd1856, [kernel_params+2328]; cvta.to.global.u64 %rd1857, %rd1856; add.s32 %r2179, %r2172, %r448; mul.wide.u32 %rd1858, %r2179, 8; add.s64 %rd1859, %rd1857, %rd1858; ld.global.v2.f32 {%f29206, %f29207}, [%rd1859]; add.s32 %r2180, %r2173, %r448; mul.wide.u32 %rd1860, %r2180, 8; add.s64 %rd1861, %rd1857, %rd1860; ld.global.v2.f32 {%f29204, %f29205}, [%rd1861]; add.s32 %r2181, %r2174, %r448; mul.wide.u32 %rd1862, %r2181, 8; add.s64 %rd1863, %rd1857, %rd1862; ld.global.v2.f32 {%f29202, %f29203}, [%rd1863]; bra.uni $L__BB107_912; $L__BB107_918: and.b32 %r2195, %r445, 131; setp.eq.s32 %p794, %r2195, 0; mov.f32 %f29208, 0f00000000; mov.f32 %f29209, %f29208; @%p794 bra $L__BB107_922; cvt.u32.u64 %r2196, %rd172; setp.eq.s32 %p795, %r445, 128; selp.b32 %r2197, %r2196, 0, %p795; add.s32 %r2198, %r2197, %r448; ld.const.u64 %rd1880, [kernel_params+2328]; cvta.to.global.u64 %rd1881, %rd1880; mul.wide.s32 %rd1882, %r2198, 8; add.s64 %rd1883, %rd1881, %rd1882; ld.global.v2.f32 {%f29208, %f29209}, [%rd1883]; bra.uni $L__BB107_922; $L__BB107_913: and.b32 %r2184, %r445, 7; setp.eq.s32 %p790, %r2184, 0; mov.f32 %f29208, 0f00000000; mov.f32 %f29209, %f29208; @%p790 bra $L__BB107_922; setp.eq.s32 %p791, %r445, 4; selp.b32 %r2186, %r2169, 0, %p791; add.s32 %r2187, %r2186, %r448; ld.const.u64 %rd1868, [kernel_params+2328]; cvta.to.global.u64 %rd1869, %rd1868; mul.wide.s32 %rd1870, %r2187, 8; add.s64 %rd1871, %rd1869, %rd1870; ld.global.v2.f32 {%f29208, %f29209}, [%rd1871]; bra.uni $L__BB107_922; $L__BB107_911: mad.lo.s32 %r2183, %r2169, 3, %r448; ld.const.u64 %rd1864, [kernel_params+2328]; cvta.to.global.u64 %rd1865, %rd1864; mul.wide.s32 %rd1866, %r2183, 8; add.s64 %rd1867, %rd1865, %rd1866; ld.global.v2.f32 {%f29206, %f29207}, [%rd1867]; ld.global.v2.f32 {%f29204, %f29205}, [%rd1867+8]; ld.global.v2.f32 {%f29202, %f29203}, [%rd1867+16]; $L__BB107_912: ld.local.f32 %f10492, [%rd5+-20]; ld.local.f32 %f10493, [%rd5+-16]; mul.ftz.f32 %f10494, %f29202, %f10493; mul.ftz.f32 %f10495, %f29203, %f10493; fma.rn.ftz.f32 %f10496, %f29204, %f10492, %f10494; fma.rn.ftz.f32 %f10497, %f29205, %f10492, %f10495; mov.f32 %f10498, 0f3F800000; sub.ftz.f32 %f10499, %f10498, %f10492; sub.ftz.f32 %f10500, %f10499, %f10493; fma.rn.ftz.f32 %f29209, %f29207, %f10500, %f10497; fma.rn.ftz.f32 %f29208, %f29206, %f10500, %f10496; $L__BB107_922: mov.f32 %f29212, 0f3F800000; $L__BB107_923: ld.const.s32 %rd1888, [kernel_params+1296]; add.s64 %rd1889, %rd164, %rd1888; cvta.to.global.u64 %rd1890, %rd1059; shl.b64 %rd1891, %rd1889, 2; add.s64 %rd1892, %rd1890, %rd1891; add.s64 %rd1893, %rd1892, 4; add.s64 %rd1894, %rd1892, 8; atom.global.add.f32 %f10526, [%rd1892], %f29208; atom.global.add.f32 %f10527, [%rd1893], %f29209; atom.global.add.f32 %f10528, [%rd1894], %f29212; $L__BB107_924: and.b32 %r2202, %r402, 128; setp.eq.s32 %p799, %r2202, 0; @%p799 bra $L__BB107_1145; ld.local.u32 %r454, [%rd5+-24]; and.b32 %r455, %r454, 14; setp.ne.s32 %p800, %r455, 0; @%p800 bra $L__BB107_927; bra.uni $L__BB107_926; $L__BB107_927: and.b32 %r2203, %r454, 6; setp.eq.s32 %p801, %r2203, 0; @%p801 bra $L__BB107_929; ld.const.u64 %rd1897, [kernel_params+2264]; cvta.to.global.u64 %rd1898, %rd1897; ld.local.u32 %r2204, [%rd5+-28]; mul.wide.s32 %rd1899, %r2204, 16; add.s64 %rd1900, %rd1898, %rd1899; shr.s32 %r2205, %r454, 7; ld.global.u32 %r2206, [%rd1900+4]; add.s32 %r2207, %r2205, %r2206; ld.const.u64 %rd1901, [kernel_params+2272]; cvta.to.global.u64 %rd1902, %rd1901; mul.wide.s32 %rd1903, %r2207, 16; add.s64 %rd1904, %rd1902, %rd1903; ld.global.v4.f32 {%f10533, %f10534, %f10535, %f10536}, [%rd1904]; ld.global.v4.f32 {%f10540, %f10541, %f10542, %f10543}, [%rd1904+16]; ld.local.f32 %f10547, [%rd5+-20]; mov.f32 %f10548, 0f3F800000; sub.ftz.f32 %f10549, %f10548, %f10547; mul.ftz.f32 %f10550, %f10533, %f10549; mul.ftz.f32 %f10551, %f10534, %f10549; mul.ftz.f32 %f10552, %f10535, %f10549; fma.rn.ftz.f32 %f29216, %f10540, %f10547, %f10550; fma.rn.ftz.f32 %f29217, %f10541, %f10547, %f10551; fma.rn.ftz.f32 %f29218, %f10542, %f10547, %f10552; bra.uni $L__BB107_931; $L__BB107_926: cvta.to.local.u64 %rd1896, %rd1061; ld.local.v4.f32 {%f29216, %f29217, %f29218, %f10532}, [%rd1896]; bra.uni $L__BB107_935; $L__BB107_929: and.b32 %r2208, %r454, 8; setp.eq.s32 %p802, %r2208, 0; mov.f32 %f29216, 0f00000000; mov.f32 %f29217, %f29216; mov.f32 %f29218, %f29216; @%p802 bra $L__BB107_931; ld.const.u64 %rd1905, [kernel_params+2296]; cvta.to.global.u64 %rd1906, %rd1905; ld.local.u32 %r2209, [%rd5+-28]; mul.wide.s32 %rd1907, %r2209, 16; add.s64 %rd1908, %rd1906, %rd1907; ld.global.v4.f32 {%f29216, %f29217, %f29218, %f10559}, [%rd1908]; $L__BB107_931: ld.local.u32 %r456, [%rd5+-32]; and.b32 %r2210, %r456, 4; setp.ne.s32 %p803, %r2210, 0; @%p803 bra $L__BB107_935; and.b32 %r2211, %r456, 2; setp.eq.s32 %p804, %r2211, 0; @%p804 bra $L__BB107_934; bra.uni $L__BB107_933; $L__BB107_934: ld.const.u64 %rd1909, [kernel_params+2160]; cvta.to.global.u64 %rd1910, %rd1909; ld.local.u32 %r2212, [%rd5+-12]; mul.wide.s32 %rd1911, %r2212, 208; add.s64 %rd1912, %rd1910, %rd1911; ld.global.v4.f32 {%f10593, %f10594, %f10595, %f10596}, [%rd1912]; ld.global.v4.f32 {%f10601, %f10602, %f10603, %f10604}, [%rd1912+16]; ld.global.v4.f32 {%f10609, %f10610, %f10611, %f10612}, [%rd1912+32]; mul.ftz.f32 %f10617, %f29217, %f10594; fma.rn.ftz.f32 %f10618, %f29216, %f10593, %f10617; fma.rn.ftz.f32 %f10619, %f29218, %f10595, %f10618; add.ftz.f32 %f1288, %f10596, %f10619; mul.ftz.f32 %f10620, %f29217, %f10602; fma.rn.ftz.f32 %f10621, %f29216, %f10601, %f10620; fma.rn.ftz.f32 %f10622, %f29218, %f10603, %f10621; add.ftz.f32 %f1289, %f10604, %f10622; mul.ftz.f32 %f10623, %f29217, %f10610; fma.rn.ftz.f32 %f10624, %f29216, %f10609, %f10623; fma.rn.ftz.f32 %f10625, %f29218, %f10611, %f10624; add.ftz.f32 %f29218, %f10612, %f10625; mov.f32 %f29216, %f1288; mov.f32 %f29217, %f1289; bra.uni $L__BB107_935; $L__BB107_933: ld.local.v4.f32 {%f10560, %f10561, %f10562, %f10563}, [%rd5+56]; mul.ftz.f32 %f10568, %f29217, %f10561; fma.rn.ftz.f32 %f10569, %f29216, %f10560, %f10568; fma.rn.ftz.f32 %f10570, %f29218, %f10562, %f10569; add.ftz.f32 %f1285, %f10563, %f10570; ld.local.v4.f32 {%f10571, %f10572, %f10573, %f10574}, [%rd5+72]; mul.ftz.f32 %f10579, %f29217, %f10572; fma.rn.ftz.f32 %f10580, %f29216, %f10571, %f10579; fma.rn.ftz.f32 %f10581, %f29218, %f10573, %f10580; add.ftz.f32 %f1286, %f10574, %f10581; ld.local.v4.f32 {%f10582, %f10583, %f10584, %f10585}, [%rd5+88]; mul.ftz.f32 %f10590, %f29217, %f10583; fma.rn.ftz.f32 %f10591, %f29216, %f10582, %f10590; fma.rn.ftz.f32 %f10592, %f29218, %f10584, %f10591; add.ftz.f32 %f29218, %f10585, %f10592; mov.f32 %f29216, %f1285; mov.f32 %f29217, %f1286; $L__BB107_935: ld.local.u32 %r5684, [%rd5+-12]; cvt.s64.s32 %rd184, %r5684; ld.local.s32 %rd185, [%rd5+-28]; setp.eq.s32 %p805, %r5684, -1; mov.u32 %r478, -1; mov.u32 %r475, 0; mov.u32 %r5680, %r475; mov.u32 %r5681, %r475; @%p805 bra $L__BB107_950; ld.const.u64 %rd1913, [kernel_params+2160]; cvta.to.global.u64 %rd1914, %rd1913; mul.lo.s64 %rd1915, %rd184, 208; add.s64 %rd1916, %rd1914, %rd1915; ld.global.u32 %r458, [%rd1916+164]; and.b32 %r2218, %r454, 1; setp.eq.b32 %p806, %r2218, 1; mov.pred %p807, 0; xor.pred %p808, %p806, %p807; not.pred %p809, %p808; mov.u32 %r5674, 0; @%p809 bra $L__BB107_940; cvt.u32.u64 %r2220, %rd185; setp.eq.s32 %p810, %r2220, -1; mov.u32 %r5673, -1; @%p810 bra $L__BB107_939; ld.const.u64 %rd1917, [kernel_params+2240]; cvta.to.global.u64 %rd1918, %rd1917; shl.b64 %rd1919, %rd185, 2; add.s64 %rd1920, %rd1918, %rd1919; ld.global.u32 %r5673, [%rd1920]; $L__BB107_939: setp.ne.s32 %p811, %r5673, -1; selp.u32 %r5674, 1, 0, %p811; $L__BB107_940: ld.const.u64 %rd1921, [kernel_params+2312]; cvta.to.global.u64 %rd186, %rd1921; add.s32 %r5676, %r5674, %r458; mul.wide.u32 %rd1922, %r5676, 16; add.s64 %rd1923, %rd186, %rd1922; add.s64 %rd4769, %rd1923, 14; ld.global.u16 %rs618, [%rd1923+12]; ld.global.u32 %r5675, [%rd1923+8]; ld.global.u64 %rd4767, [%rd1923]; setp.eq.s64 %p812, %rd4767, 11; add.s64 %rd4768, %rd1923, 15; @%p812 bra $L__BB107_946; $L__BB107_941: setp.eq.s64 %p813, %rd4767, 0; @%p813 bra $L__BB107_943; add.s32 %r5676, %r5676, 2; bra.uni $L__BB107_944; $L__BB107_943: setp.eq.s16 %p814, %rs618, 0; mov.u32 %r475, 0; mov.u32 %r5676, %r5675; mov.u32 %r5680, %r475; mov.u32 %r5681, %r475; @%p814 bra $L__BB107_950; $L__BB107_944: cvt.u64.u32 %rd191, %r5676; mul.wide.u32 %rd1924, %r5676, 16; add.s64 %rd1925, %rd186, %rd1924; ld.global.u16 %rs618, [%rd1925+12]; ld.global.u32 %r5675, [%rd1925+8]; ld.global.u64 %rd4767, [%rd1925]; setp.ne.s64 %p815, %rd4767, 11; @%p815 bra $L__BB107_941; shl.b64 %rd1926, %rd191, 4; add.s64 %rd1927, %rd186, %rd1926; add.s64 %rd4769, %rd1927, 14; add.s64 %rd4768, %rd1927, 15; $L__BB107_946: ld.global.u8 %rs26, [%rd4769]; ld.global.u8 %rs27, [%rd4768]; cvt.u32.u64 %r2225, %rd185; setp.ne.s32 %p816, %r2225, -1; @%p816 bra $L__BB107_949; mov.u32 %r475, 0; add.s16 %rs325, %rs618, -1; setp.lt.u16 %p817, %rs325, 2; @%p817 bra $L__BB107_949; setp.ne.s16 %p818, %rs618, 1024; mov.u32 %r5680, %r475; mov.u32 %r5681, %r475; @%p818 bra $L__BB107_950; $L__BB107_949: setp.eq.s16 %p819, %rs618, 0; selp.b32 %r478, -1, %r5675, %p819; cvt.u32.u16 %r475, %rs618; cvt.u32.u16 %r2230, %rs27; and.b32 %r5681, %r2230, 255; cvt.u32.u16 %r2231, %rs26; and.b32 %r5680, %r2231, 255; $L__BB107_950: setp.eq.s32 %p820, %r478, -1; mov.f32 %f29234, %f29216; mov.f32 %f29235, %f29217; mov.f32 %f29236, %f29218; mov.f32 %f29237, %f29216; mov.f32 %f29238, %f29217; mov.f32 %f29239, %f29218; @%p820 bra $L__BB107_986; ld.const.u64 %rd1928, [kernel_params+2160]; cvta.to.global.u64 %rd1929, %rd1928; mul.lo.s64 %rd1930, %rd184, 208; add.s64 %rd1931, %rd1929, %rd1930; add.s64 %rd197, %rd1931, 148; ld.global.u32 %r481, [%rd1931+156]; @%p800 bra $L__BB107_971; bra.uni $L__BB107_952; $L__BB107_971: ld.global.u32 %r487, [%rd197]; and.b32 %r488, %r475, 768; setp.eq.s32 %p836, %r488, 0; @%p836 bra $L__BB107_973; ld.const.u64 %rd1970, [kernel_params+2264]; cvta.to.global.u64 %rd1971, %rd1970; shl.b64 %rd1972, %rd185, 4; add.s64 %rd1973, %rd1971, %rd1972; shr.s32 %r2270, %r454, 7; ld.global.u32 %r2271, [%rd1973+4]; add.s32 %r2272, %r2270, %r2271; ld.const.u64 %rd1974, [kernel_params+2344]; cvta.to.global.u64 %rd1975, %rd1974; add.s32 %r2273, %r2272, %r478; mul.wide.s32 %rd1976, %r2273, 16; add.s64 %rd1977, %rd1975, %rd1976; ld.global.v4.f32 {%f10672, %f10673, %f10674, %f10675}, [%rd1977]; ld.global.v4.f32 {%f10680, %f10681, %f10682, %f10683}, [%rd1977+16]; ld.local.f32 %f10688, [%rd5+-20]; mov.f32 %f10689, 0f3F800000; sub.ftz.f32 %f10690, %f10689, %f10688; mul.ftz.f32 %f10691, %f10672, %f10690; mul.ftz.f32 %f10692, %f10673, %f10690; mul.ftz.f32 %f10693, %f10674, %f10690; fma.rn.ftz.f32 %f29239, %f10682, %f10688, %f10693; fma.rn.ftz.f32 %f29238, %f10681, %f10688, %f10692; fma.rn.ftz.f32 %f29237, %f10680, %f10688, %f10691; bra.uni $L__BB107_975; $L__BB107_952: and.b32 %r2232, %r454, 1; setp.eq.b32 %p822, %r2232, 1; mov.pred %p823, 0; xor.pred %p824, %p822, %p823; not.pred %p825, %p824; mov.f32 %f29234, %f29216; mov.f32 %f29235, %f29217; mov.f32 %f29236, %f29218; mov.f32 %f29237, %f29216; mov.f32 %f29238, %f29217; mov.f32 %f29239, %f29218; @%p825 bra $L__BB107_986; cvt.u32.u64 %r2234, %rd185; setp.eq.s32 %p826, %r2234, -1; mov.u32 %r5683, -1; @%p826 bra $L__BB107_955; ld.const.u64 %rd1932, [kernel_params+2240]; cvta.to.global.u64 %rd1933, %rd1932; shl.b64 %rd1934, %rd185, 2; add.s64 %rd1935, %rd1933, %rd1934; ld.global.u32 %r5683, [%rd1935]; $L__BB107_955: setp.eq.s32 %p827, %r5683, -1; @%p827 bra $L__BB107_957; mov.u64 %rd1936, 0; { // callseq 329, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r5680; st.param.b32 [param1+8], %r5681; st.param.b32 [param1+12], %r478; .param .b64 param2; st.param.b64 [param2+0], %rd1936; .param .b64 param3; st.param.b64 [param3+0], %rd1936; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f29237, [retval0+0]; ld.param.f32 %f29238, [retval0+4]; ld.param.f32 %f29239, [retval0+8]; } // callseq 329 add.s32 %r2235, %r481, %r478; { // callseq 330, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1061; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r5680; st.param.b32 [param1+8], %r5681; st.param.b32 [param1+12], %r2235; .param .b64 param2; st.param.b64 [param2+0], %rd1936; .param .b64 param3; st.param.b64 [param3+0], %rd1936; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f29234, [retval0+0]; ld.param.f32 %f29235, [retval0+4]; ld.param.f32 %f29236, [retval0+8]; } // callseq 330 ld.local.u32 %r5684, [%rd5+-12]; bra.uni $L__BB107_986; $L__BB107_973: and.b32 %r2274, %r475, 131; setp.eq.s32 %p837, %r2274, 0; mov.f32 %f29237, 0f00000000; mov.f32 %f29238, %f29237; mov.f32 %f29239, %f29237; @%p837 bra $L__BB107_975; cvt.u32.u64 %r2275, %rd185; setp.eq.s32 %p838, %r475, 128; selp.b32 %r2276, %r2275, 0, %p838; add.s32 %r2277, %r2276, %r478; ld.const.u64 %rd1978, [kernel_params+2344]; cvta.to.global.u64 %rd1979, %rd1978; mul.wide.s32 %rd1980, %r2277, 16; add.s64 %rd1981, %rd1979, %rd1980; ld.global.v4.f32 {%f29237, %f29238, %f29239, %f10702}, [%rd1981]; $L__BB107_975: add.s32 %r489, %r487, %r478; @%p836 bra $L__BB107_977; cvt.u32.u64 %r2278, %rd185; ld.const.u64 %rd1982, [kernel_params+2264]; cvta.to.global.u64 %rd1983, %rd1982; mul.wide.s32 %rd1984, %r2278, 16; add.s64 %rd1985, %rd1983, %rd1984; shr.s32 %r2279, %r454, 7; ld.global.u32 %r2280, [%rd1985+4]; add.s32 %r2281, %r2279, %r2280; ld.const.u64 %rd1986, [kernel_params+2344]; cvta.to.global.u64 %rd1987, %rd1986; add.s32 %r2282, %r2281, %r489; mul.wide.s32 %rd1988, %r2282, 16; add.s64 %rd1989, %rd1987, %rd1988; ld.global.v4.f32 {%f10703, %f10704, %f10705, %f10706}, [%rd1989]; ld.global.v4.f32 {%f10711, %f10712, %f10713, %f10714}, [%rd1989+16]; ld.local.f32 %f10719, [%rd5+-20]; mov.f32 %f10720, 0f3F800000; sub.ftz.f32 %f10721, %f10720, %f10719; mul.ftz.f32 %f10722, %f10703, %f10721; mul.ftz.f32 %f10723, %f10704, %f10721; mul.ftz.f32 %f10724, %f10705, %f10721; fma.rn.ftz.f32 %f29236, %f10713, %f10719, %f10724; fma.rn.ftz.f32 %f29235, %f10712, %f10719, %f10723; fma.rn.ftz.f32 %f29234, %f10711, %f10719, %f10722; bra.uni $L__BB107_979; $L__BB107_977: and.b32 %r2283, %r475, 131; setp.eq.s32 %p840, %r2283, 0; mov.f32 %f29234, 0f00000000; mov.f32 %f29235, %f29234; mov.f32 %f29236, %f29234; @%p840 bra $L__BB107_979; cvt.u32.u64 %r2284, %rd185; setp.eq.s32 %p841, %r475, 128; selp.b32 %r2285, %r2284, 0, %p841; add.s32 %r2286, %r2285, %r489; ld.const.u64 %rd1990, [kernel_params+2344]; cvta.to.global.u64 %rd1991, %rd1990; mul.wide.s32 %rd1992, %r2286, 16; add.s64 %rd1993, %rd1991, %rd1992; ld.global.v4.f32 {%f29234, %f29235, %f29236, %f10733}, [%rd1993]; $L__BB107_979: ld.local.u32 %r490, [%rd5+-32]; and.b32 %r2287, %r490, 64; setp.ne.s32 %p842, %r2287, 0; @%p842 bra $L__BB107_986; and.b32 %r491, %r490, 2; setp.eq.s32 %p843, %r491, 0; @%p843 bra $L__BB107_982; bra.uni $L__BB107_981; $L__BB107_982: ld.global.v4.f32 {%f10767, %f10768, %f10769, %f10770}, [%rd197+-148]; ld.global.v4.f32 {%f10775, %f10776, %f10777, %f10778}, [%rd197+-132]; ld.global.v4.f32 {%f10783, %f10784, %f10785, %f10786}, [%rd197+-116]; mul.ftz.f32 %f10791, %f29238, %f10768; fma.rn.ftz.f32 %f10792, %f29237, %f10767, %f10791; fma.rn.ftz.f32 %f10793, %f29239, %f10769, %f10792; add.ftz.f32 %f29231, %f10770, %f10793; mul.ftz.f32 %f10794, %f29238, %f10776; fma.rn.ftz.f32 %f10795, %f29237, %f10775, %f10794; fma.rn.ftz.f32 %f10796, %f29239, %f10777, %f10795; add.ftz.f32 %f29232, %f10778, %f10796; mul.ftz.f32 %f10797, %f29238, %f10784; fma.rn.ftz.f32 %f10798, %f29237, %f10783, %f10797; fma.rn.ftz.f32 %f10799, %f29239, %f10785, %f10798; add.ftz.f32 %f29239, %f10786, %f10799; bra.uni $L__BB107_983; $L__BB107_981: ld.local.v4.f32 {%f10734, %f10735, %f10736, %f10737}, [%rd5+56]; mul.ftz.f32 %f10742, %f29238, %f10735; fma.rn.ftz.f32 %f10743, %f29237, %f10734, %f10742; fma.rn.ftz.f32 %f10744, %f29239, %f10736, %f10743; add.ftz.f32 %f29231, %f10737, %f10744; ld.local.v4.f32 {%f10745, %f10746, %f10747, %f10748}, [%rd5+72]; mul.ftz.f32 %f10753, %f29238, %f10746; fma.rn.ftz.f32 %f10754, %f29237, %f10745, %f10753; fma.rn.ftz.f32 %f10755, %f29239, %f10747, %f10754; add.ftz.f32 %f29232, %f10748, %f10755; ld.local.v4.f32 {%f10756, %f10757, %f10758, %f10759}, [%rd5+88]; mul.ftz.f32 %f10764, %f29238, %f10757; fma.rn.ftz.f32 %f10765, %f29237, %f10756, %f10764; fma.rn.ftz.f32 %f10766, %f29239, %f10758, %f10765; add.ftz.f32 %f29239, %f10759, %f10766; $L__BB107_983: @%p843 bra $L__BB107_985; bra.uni $L__BB107_984; $L__BB107_985: ld.global.v4.f32 {%f10833, %f10834, %f10835, %f10836}, [%rd197+-148]; ld.global.v4.f32 {%f10841, %f10842, %f10843, %f10844}, [%rd197+-132]; ld.global.v4.f32 {%f10849, %f10850, %f10851, %f10852}, [%rd197+-116]; mul.ftz.f32 %f10857, %f29235, %f10834; fma.rn.ftz.f32 %f10858, %f29234, %f10833, %f10857; fma.rn.ftz.f32 %f10859, %f29236, %f10835, %f10858; add.ftz.f32 %f1366, %f10836, %f10859; mul.ftz.f32 %f10860, %f29235, %f10842; fma.rn.ftz.f32 %f10861, %f29234, %f10841, %f10860; fma.rn.ftz.f32 %f10862, %f29236, %f10843, %f10861; add.ftz.f32 %f1367, %f10844, %f10862; mul.ftz.f32 %f10863, %f29235, %f10850; fma.rn.ftz.f32 %f10864, %f29234, %f10849, %f10863; fma.rn.ftz.f32 %f10865, %f29236, %f10851, %f10864; add.ftz.f32 %f29236, %f10852, %f10865; mov.f32 %f29234, %f1366; mov.f32 %f29235, %f1367; mov.f32 %f29237, %f29231; mov.f32 %f29238, %f29232; bra.uni $L__BB107_986; $L__BB107_984: ld.local.v4.f32 {%f10800, %f10801, %f10802, %f10803}, [%rd5+56]; mul.ftz.f32 %f10808, %f29235, %f10801; fma.rn.ftz.f32 %f10809, %f29234, %f10800, %f10808; fma.rn.ftz.f32 %f10810, %f29236, %f10802, %f10809; add.ftz.f32 %f1363, %f10803, %f10810; ld.local.v4.f32 {%f10811, %f10812, %f10813, %f10814}, [%rd5+72]; mul.ftz.f32 %f10819, %f29235, %f10812; fma.rn.ftz.f32 %f10820, %f29234, %f10811, %f10819; fma.rn.ftz.f32 %f10821, %f29236, %f10813, %f10820; add.ftz.f32 %f1364, %f10814, %f10821; ld.local.v4.f32 {%f10822, %f10823, %f10824, %f10825}, [%rd5+88]; mul.ftz.f32 %f10830, %f29235, %f10823; fma.rn.ftz.f32 %f10831, %f29234, %f10822, %f10830; fma.rn.ftz.f32 %f10832, %f29236, %f10824, %f10831; add.ftz.f32 %f29236, %f10825, %f10832; mov.f32 %f29234, %f1363; mov.f32 %f29235, %f1364; mov.f32 %f29237, %f29231; mov.f32 %f29238, %f29232; bra.uni $L__BB107_986; $L__BB107_957: and.b32 %r485, %r475, 56; setp.eq.s32 %p828, %r485, 0; @%p828 bra $L__BB107_962; and.b32 %r2236, %r475, 24; setp.eq.s32 %p829, %r2236, 0; @%p829 bra $L__BB107_960; ld.const.u64 %rd1938, [kernel_params+2232]; cvta.to.global.u64 %rd1939, %rd1938; shl.b64 %rd1940, %rd185, 4; add.s64 %rd1941, %rd1939, %rd1940; ld.global.v4.u32 {%r2237, %r2238, %r2239, %r2240}, [%rd1941]; ld.const.u64 %rd1942, [kernel_params+2336]; cvta.to.global.u64 %rd1943, %rd1942; add.s32 %r2244, %r2237, %r478; mul.wide.u32 %rd1944, %r2244, 12; add.s64 %rd4776, %rd1943, %rd1944; add.s32 %r2245, %r2238, %r478; mul.wide.u32 %rd1945, %r2245, 12; add.s64 %rd4773, %rd1943, %rd1945; add.s32 %r2246, %r2239, %r478; mul.wide.u32 %rd1946, %r2246, 12; add.s64 %rd4770, %rd1943, %rd1946; add.s64 %rd4777, %rd4776, 4; add.s64 %rd4778, %rd4776, 8; add.s64 %rd4774, %rd4773, 4; add.s64 %rd4775, %rd4773, 8; add.s64 %rd4771, %rd4770, 4; add.s64 %rd4772, %rd4770, 8; bra.uni $L__BB107_961; $L__BB107_1758: ld.local.f32 %f29682, [%rd280+-16]; ld.local.f32 %f29681, [%rd280+-20]; ld.local.f32 %f29680, [%rd280+-24]; $L__BB107_1760: cvta.to.global.u64 %rd407, %rd2223; add.s32 %r3031, %r5732, -1; cvt.s64.s32 %rd408, %r3031; add.u64 %rd409, %SPL, 144; mul.wide.s32 %rd2585, %r3031, 8; add.s64 %rd410, %rd409, %rd2585; add.s32 %r806, %r5732, -2; and.b32 %r807, %r5732, 3; sub.s32 %r808, %r5732, %r807; and.b32 %r5793, %r3031, 3; add.s32 %r810, %r5732, -3; add.s32 %r811, %r5732, -4; cvt.s64.s32 %rd411, %r811; add.s32 %r812, %r5732, -5; mov.u16 %rs349, 1; mov.f32 %f29683, 0f3DCCCCCD; mul.lo.s64 %rd2586, %rd411, 192; add.s64 %rd2587, %rd1, %rd2586; add.s64 %rd413, %rd2587, 176; setp.lt.s32 %p1589, %r5732, 1; mov.u16 %rs351, 0; setp.eq.s32 %p1593, %r621, 0; setp.ne.s32 %p1969, %r622, 0; mov.u16 %rs628, %rs349; bra.uni $L__BB107_1761; $L__BB107_962: and.b32 %r2249, %r475, 7; setp.eq.s32 %p830, %r2249, 0; mov.f32 %f29237, 0f00000000; mov.f32 %f29238, %f29237; mov.f32 %f29239, %f29237; @%p830 bra $L__BB107_964; setp.eq.s32 %p831, %r475, 4; selp.b32 %r2251, %r2234, 0, %p831; add.s32 %r2252, %r2251, %r478; ld.const.u64 %rd1950, [kernel_params+2336]; cvta.to.global.u64 %rd1951, %rd1950; mul.wide.s32 %rd1952, %r2252, 12; add.s64 %rd1953, %rd1951, %rd1952; ld.global.f32 %f29239, [%rd1953+8]; ld.global.f32 %f29238, [%rd1953+4]; ld.global.f32 %f29237, [%rd1953]; bra.uni $L__BB107_964; $L__BB107_1794: setp.lt.s32 %p1574, %r5732, 2; ld.local.v2.f32 {%f15917, %f15918}, [%rd412+92]; add.u64 %rd2598, %SPL, 96; st.local.v2.f32 [%rd2598], {%f15917, %f15918}; @%p1574 bra $L__BB107_1800; add.u64 %rd418, %SPL, 0; mov.u32 %r5738, 1; $L__BB107_1796: cvt.s64.s32 %rd419, %r5738; mul.wide.s32 %rd2600, %r5738, 192; add.s64 %rd2601, %rd1, %rd2600; add.s64 %rd420, %rd2601, 144; ld.local.v4.f32 {%f15922, %f15923, %f15924, %f15925}, [%rd2601+144]; mul.wide.s32 %rd2602, %r5738, 16; add.s64 %rd2603, %rd418, %rd2602; add.s64 %rd421, %rd2603, -16; ld.local.v4.f32 {%f15930, %f15931, %f15932, %f15933}, [%rd2603+-16]; mul.ftz.f32 %f15938, %f15923, %f15932; fma.rn.ftz.f32 %f2839, %f15922, %f15930, %f15938; mul.ftz.f32 %f15939, %f15923, %f15933; fma.rn.ftz.f32 %f2840, %f15922, %f15931, %f15939; mul.ftz.f32 %f15940, %f15925, %f15932; fma.rn.ftz.f32 %f2841, %f15924, %f15930, %f15940; mul.ftz.f32 %f15941, %f15925, %f15933; fma.rn.ftz.f32 %f2842, %f15924, %f15931, %f15941; ld.local.v4.f32 {%f15942, %f15943, %f15944, %f15945}, [%rd2601+-16]; mul.ftz.f32 %f15950, %f15944, %f2840; fma.rn.ftz.f32 %f15951, %f15942, %f2839, %f15950; mul.ftz.f32 %f15952, %f15945, %f2840; fma.rn.ftz.f32 %f15953, %f15943, %f2839, %f15952; mul.ftz.f32 %f15954, %f15944, %f2842; fma.rn.ftz.f32 %f15955, %f15942, %f2841, %f15954; mul.ftz.f32 %f15956, %f15945, %f2842; fma.rn.ftz.f32 %f15957, %f15943, %f2841, %f15956; ld.local.v4.f32 {%f15958, %f15959, %f15960, %f15961}, [%rd2601+160]; sub.ftz.f32 %f2843, %f15958, %f15951; sub.ftz.f32 %f2844, %f15959, %f15953; sub.ftz.f32 %f2845, %f15960, %f15955; sub.ftz.f32 %f2846, %f15961, %f15957; mul.ftz.f32 %f15966, %f2843, %f2846; mul.ftz.f32 %f15967, %f2844, %f2845; sub.ftz.f32 %f2847, %f15966, %f15967; abs.ftz.f32 %f15968, %f2847; setp.lt.ftz.f32 %p1575, %f15968, 0f38D1B717; mov.f32 %f29705, 0f00000000; @%p1575 bra $L__BB107_1798; rcp.approx.ftz.f32 %f15969, %f2847; mul.ftz.f32 %f15970, %f2844, %f15969; mul.ftz.f32 %f15971, %f2845, %f15969; mul.ftz.f32 %f15972, %f2843, %f15969; mul.ftz.f32 %f15973, %f2846, %f15969; neg.ftz.f32 %f15974, %f15971; neg.ftz.f32 %f15975, %f15970; st.local.v4.f32 [%rd421+16], {%f15973, %f15975, %f15974, %f15972}; mov.f32 %f29705, %f2847; $L__BB107_1798: setp.eq.ftz.f32 %p1577, %f29705, 0f00000000; mov.pred %p3920, %p3921; @%p1577 bra $L__BB107_1807; cvt.u32.u64 %r3045, %rd419; shl.b64 %rd2606, %rd419, 3; add.s64 %rd2607, %rd2598, %rd2606; ld.local.v2.f32 {%f15976, %f15977}, [%rd2607+-8]; mul.ftz.f32 %f15980, %f2840, %f15977; fma.rn.ftz.f32 %f15981, %f2839, %f15976, %f15980; mul.ftz.f32 %f15982, %f2842, %f15977; fma.rn.ftz.f32 %f15983, %f2841, %f15976, %f15982; ld.local.v2.f32 {%f15984, %f15985}, [%rd420+-16]; sub.ftz.f32 %f15988, %f15985, %f15983; sub.ftz.f32 %f15989, %f15984, %f15981; st.local.v2.f32 [%rd2607], {%f15989, %f15988}; add.s32 %r5738, %r3045, 1; setp.lt.s32 %p1578, %r5738, %r5732; @%p1578 bra $L__BB107_1796; $L__BB107_1800: add.u64 %rd422, %SPL, 0; shl.b64 %rd2609, %rd408, 4; add.s64 %rd423, %rd422, %rd2609; ld.local.v4.f32 {%f15990, %f15991, %f15992, %f15993}, [%rd423]; shl.b64 %rd2611, %rd408, 3; add.s64 %rd425, %rd2598, %rd2611; ld.local.v2.f32 {%f15998, %f15999}, [%rd425]; mul.ftz.f32 %f16002, %f15991, %f15999; fma.rn.ftz.f32 %f2849, %f15990, %f15998, %f16002; mul.ftz.f32 %f16003, %f15993, %f15999; fma.rn.ftz.f32 %f2850, %f15992, %f15998, %f16003; st.local.v2.f32 [%rd410], {%f2849, %f2850}; mov.pred %p3920, 0; @%p1574 bra $L__BB107_1807; setp.eq.s32 %p1581, %r5793, 0; mov.u32 %r5739, %r806; @%p1581 bra $L__BB107_1805; setp.eq.s32 %p1582, %r5793, 1; ld.local.v4.f32 {%f16004, %f16005, %f16006, %f16007}, [%rd423+-16]; ld.local.v4.f32 {%f16012, %f16013, %f16014, %f16015}, [%rd413+384]; mul.ftz.f32 %f16020, %f16013, %f2850; fma.rn.ftz.f32 %f16021, %f16012, %f2849, %f16020; mul.ftz.f32 %f16022, %f16015, %f2850; fma.rn.ftz.f32 %f16023, %f16014, %f2849, %f16022; ld.local.v2.f32 {%f16024, %f16025}, [%rd425+-8]; sub.ftz.f32 %f16028, %f16024, %f16021; sub.ftz.f32 %f16029, %f16025, %f16023; mul.ftz.f32 %f16030, %f16005, %f16029; fma.rn.ftz.f32 %f2851, %f16004, %f16028, %f16030; mul.ftz.f32 %f16031, %f16007, %f16029; fma.rn.ftz.f32 %f2852, %f16006, %f16028, %f16031; st.local.v2.f32 [%rd410+-8], {%f2851, %f2852}; mov.u32 %r5739, %r810; @%p1582 bra $L__BB107_1805; setp.eq.s32 %p1583, %r5793, 2; ld.local.v4.f32 {%f16032, %f16033, %f16034, %f16035}, [%rd423+-32]; ld.local.v4.f32 {%f16040, %f16041, %f16042, %f16043}, [%rd413+192]; mul.ftz.f32 %f16048, %f16041, %f2852; fma.rn.ftz.f32 %f16049, %f16040, %f2851, %f16048; mul.ftz.f32 %f16050, %f16043, %f2852; fma.rn.ftz.f32 %f16051, %f16042, %f2851, %f16050; ld.local.v2.f32 {%f16052, %f16053}, [%rd425+-16]; sub.ftz.f32 %f16056, %f16052, %f16049; sub.ftz.f32 %f16057, %f16053, %f16051; mul.ftz.f32 %f16058, %f16033, %f16057; fma.rn.ftz.f32 %f2853, %f16032, %f16056, %f16058; mul.ftz.f32 %f16059, %f16035, %f16057; fma.rn.ftz.f32 %f2854, %f16034, %f16056, %f16059; st.local.v2.f32 [%rd410+-16], {%f2853, %f2854}; mov.u32 %r5739, %r811; @%p1583 bra $L__BB107_1805; ld.local.v4.f32 {%f16060, %f16061, %f16062, %f16063}, [%rd423+-48]; ld.local.v4.f32 {%f16068, %f16069, %f16070, %f16071}, [%rd413]; mul.ftz.f32 %f16076, %f16069, %f2854; fma.rn.ftz.f32 %f16077, %f16068, %f2853, %f16076; mul.ftz.f32 %f16078, %f16071, %f2854; fma.rn.ftz.f32 %f16079, %f16070, %f2853, %f16078; ld.local.v2.f32 {%f16080, %f16081}, [%rd425+-24]; sub.ftz.f32 %f16084, %f16080, %f16077; sub.ftz.f32 %f16085, %f16081, %f16079; mul.ftz.f32 %f16086, %f16061, %f16085; mul.ftz.f32 %f16087, %f16063, %f16085; fma.rn.ftz.f32 %f16088, %f16062, %f16084, %f16087; fma.rn.ftz.f32 %f16089, %f16060, %f16084, %f16086; st.local.v2.f32 [%rd410+-24], {%f16089, %f16088}; mov.u32 %r5739, %r812; $L__BB107_1805: setp.lt.u32 %p1585, %r806, 3; @%p1585 bra $L__BB107_1807; $L__BB107_1806: mul.wide.s32 %rd2612, %r5739, 16; add.s64 %rd2613, %rd422, %rd2612; ld.local.v4.f32 {%f16090, %f16091, %f16092, %f16093}, [%rd2613]; mul.wide.s32 %rd2614, %r5739, 192; add.s64 %rd2615, %rd1, %rd2614; ld.local.v4.f32 {%f16098, %f16099, %f16100, %f16101}, [%rd2615+176]; mul.wide.s32 %rd2616, %r5739, 8; add.s64 %rd2617, %rd409, %rd2616; ld.local.v2.f32 {%f16106, %f16107}, [%rd2617+8]; mul.ftz.f32 %f16110, %f16099, %f16107; fma.rn.ftz.f32 %f16111, %f16098, %f16106, %f16110; mul.ftz.f32 %f16112, %f16101, %f16107; fma.rn.ftz.f32 %f16113, %f16100, %f16106, %f16112; add.s64 %rd2618, %rd2598, %rd2616; ld.local.v2.f32 {%f16114, %f16115}, [%rd2618]; sub.ftz.f32 %f16118, %f16114, %f16111; sub.ftz.f32 %f16119, %f16115, %f16113; mul.ftz.f32 %f16120, %f16091, %f16119; mul.ftz.f32 %f16121, %f16093, %f16119; fma.rn.ftz.f32 %f16122, %f16092, %f16118, %f16121; fma.rn.ftz.f32 %f16123, %f16090, %f16118, %f16120; st.local.v2.f32 [%rd2617], {%f16123, %f16122}; ld.local.v4.f32 {%f16124, %f16125, %f16126, %f16127}, [%rd2613+-16]; ld.local.v4.f32 {%f16132, %f16133, %f16134, %f16135}, [%rd2615+-16]; mul.ftz.f32 %f16140, %f16133, %f16122; fma.rn.ftz.f32 %f16141, %f16132, %f16123, %f16140; mul.ftz.f32 %f16142, %f16135, %f16122; fma.rn.ftz.f32 %f16143, %f16134, %f16123, %f16142; ld.local.v2.f32 {%f16144, %f16145}, [%rd2618+-8]; sub.ftz.f32 %f16148, %f16144, %f16141; sub.ftz.f32 %f16149, %f16145, %f16143; mul.ftz.f32 %f16150, %f16125, %f16149; mul.ftz.f32 %f16151, %f16127, %f16149; fma.rn.ftz.f32 %f16152, %f16126, %f16148, %f16151; fma.rn.ftz.f32 %f16153, %f16124, %f16148, %f16150; st.local.v2.f32 [%rd2617+-8], {%f16153, %f16152}; ld.local.v4.f32 {%f16154, %f16155, %f16156, %f16157}, [%rd2613+-32]; ld.local.v4.f32 {%f16162, %f16163, %f16164, %f16165}, [%rd2615+-208]; mul.ftz.f32 %f16170, %f16163, %f16152; fma.rn.ftz.f32 %f16171, %f16162, %f16153, %f16170; mul.ftz.f32 %f16172, %f16165, %f16152; fma.rn.ftz.f32 %f16173, %f16164, %f16153, %f16172; ld.local.v2.f32 {%f16174, %f16175}, [%rd2618+-16]; sub.ftz.f32 %f16178, %f16174, %f16171; sub.ftz.f32 %f16179, %f16175, %f16173; mul.ftz.f32 %f16180, %f16155, %f16179; mul.ftz.f32 %f16181, %f16157, %f16179; fma.rn.ftz.f32 %f16182, %f16156, %f16178, %f16181; fma.rn.ftz.f32 %f16183, %f16154, %f16178, %f16180; st.local.v2.f32 [%rd2617+-16], {%f16183, %f16182}; ld.local.v4.f32 {%f16184, %f16185, %f16186, %f16187}, [%rd2613+-48]; ld.local.v4.f32 {%f16192, %f16193, %f16194, %f16195}, [%rd2615+-400]; mul.ftz.f32 %f16200, %f16193, %f16182; fma.rn.ftz.f32 %f16201, %f16192, %f16183, %f16200; mul.ftz.f32 %f16202, %f16195, %f16182; fma.rn.ftz.f32 %f16203, %f16194, %f16183, %f16202; ld.local.v2.f32 {%f16204, %f16205}, [%rd2618+-24]; sub.ftz.f32 %f16208, %f16204, %f16201; sub.ftz.f32 %f16209, %f16205, %f16203; mul.ftz.f32 %f16210, %f16185, %f16209; mul.ftz.f32 %f16211, %f16187, %f16209; fma.rn.ftz.f32 %f16212, %f16186, %f16208, %f16211; fma.rn.ftz.f32 %f16213, %f16184, %f16208, %f16210; st.local.v2.f32 [%rd2617+-24], {%f16213, %f16212}; add.s32 %r825, %r5739, -4; setp.gt.s32 %p1587, %r5739, 3; mov.u32 %r5739, %r825; @%p1587 bra $L__BB107_1806; bra.uni $L__BB107_1807; $L__BB107_1987: st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3173, [%rd427+12]; st.local.u32 [%rd406+-12], %r3173; ld.local.f32 %f16357, [%rd427]; setp.eq.ftz.f32 %p1783, %f16357, 0f00000000; mov.b32 %r3174, %f16357; add.s32 %r3175, %r3174, 1; mov.b32 %f16358, %r3175; selp.f32 %f16359, 0f00000001, %f16358, %p1783; st.local.f32 [%rd406+-24], %f16359; mov.u16 %rs713, %rs349; bra.uni $L__BB107_2228; $L__BB107_2158: st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3293, [%rd427+12]; st.local.u32 [%rd406+-12], %r3293; ld.local.f32 %f16437, [%rd427]; setp.eq.ftz.f32 %p1973, %f16437, 0f00000000; mov.b32 %r3294, %f16437; add.s32 %r3295, %r3294, 1; mov.b32 %f16438, %r3295; selp.f32 %f16439, 0f00000001, %f16438, %p1973; st.local.f32 [%rd406+-24], %f16439; mov.u16 %rs713, %rs349; bra.uni $L__BB107_2228; $L__BB107_1761: add.s64 %rd412, %rd1, 36; and.b16 %rs350, %rs628, 255; setp.eq.s16 %p1548, %rs350, 0; @%p1548 bra $L__BB107_1808; mov.f32 %f29703, 0f00000000; @%p1589 bra $L__BB107_1790; ld.local.v4.f32 {%f15423, %f15424, %f15425, %f15426}, [%rd279]; mov.u32 %r5734, 0; $L__BB107_1764: cvt.s64.s32 %rd414, %r5734; mul.wide.s32 %rd2588, %r5734, 192; add.s64 %rd415, %rd1, %rd2588; add.s32 %r3033, %r5734, -1; mul.wide.s32 %rd2589, %r3033, 192; add.s64 %rd2590, %rd1, %rd2589; add.s64 %rd416, %rd2590, 4; setp.eq.s32 %p1550, %r5734, 0; @%p1550 bra $L__BB107_1766; ld.local.v4.f32 {%f29686, %f29685, %f29684, %f15430}, [%rd415]; ld.local.f32 %f15431, [%rd415+-192]; sub.ftz.f32 %f29687, %f15431, %f29686; ld.local.f32 %f15432, [%rd416]; sub.ftz.f32 %f29688, %f15432, %f29685; ld.local.f32 %f15433, [%rd416+4]; sub.ftz.f32 %f29689, %f15433, %f29684; bra.uni $L__BB107_1767; $L__BB107_1766: ld.local.v4.f32 {%f29686, %f29685, %f29684, %f15437}, [%rd415]; sub.ftz.f32 %f29687, %f15423, %f29686; sub.ftz.f32 %f29688, %f15424, %f29685; sub.ftz.f32 %f29689, %f15425, %f29684; $L__BB107_1767: mul.ftz.f32 %f15438, %f29688, %f29688; fma.rn.ftz.f32 %f15439, %f29687, %f29687, %f15438; fma.rn.ftz.f32 %f15440, %f29689, %f29689, %f15439; sqrt.approx.ftz.f32 %f2765, %f15440; setp.lt.ftz.f32 %p1552, %f2765, 0f3A83126F; mov.pred %p3921, -1; @%p1552 bra $L__BB107_2250; cvt.u32.u64 %r3034, %rd408; rcp.approx.ftz.f32 %f2766, %f2765; mul.ftz.f32 %f2767, %f29687, %f2766; mul.ftz.f32 %f2768, %f29688, %f2766; mul.ftz.f32 %f2769, %f29689, %f2766; cvt.u32.u64 %r3035, %rd414; setp.eq.s32 %p1553, %r3035, %r3034; @%p1553 bra $L__BB107_1770; bra.uni $L__BB107_1769; $L__BB107_1770: mov.f32 %f29690, %f29680; mov.f32 %f29691, %f29681; mov.f32 %f29692, %f29682; @%p1261 bra $L__BB107_1772; sub.ftz.f32 %f29690, %f29680, %f29686; sub.ftz.f32 %f29691, %f29681, %f29685; sub.ftz.f32 %f29692, %f29682, %f29684; bra.uni $L__BB107_1772; $L__BB107_1769: ld.local.v4.f32 {%f15441, %f15442, %f15443, %f15444}, [%rd415+192]; sub.ftz.f32 %f29690, %f15441, %f29686; sub.ftz.f32 %f29691, %f15442, %f29685; sub.ftz.f32 %f29692, %f15443, %f29684; $L__BB107_1772: mul.ftz.f32 %f15448, %f29691, %f29691; fma.rn.ftz.f32 %f15449, %f29690, %f29690, %f15448; fma.rn.ftz.f32 %f15450, %f29692, %f29692, %f15449; sqrt.approx.ftz.f32 %f2779, %f15450; setp.lt.ftz.f32 %p1556, %f2779, 0f3A83126F; @%p1556 bra $L__BB107_2250; rcp.approx.ftz.f32 %f2780, %f2779; mul.ftz.f32 %f2781, %f29690, %f2780; mul.ftz.f32 %f2782, %f29691, %f2780; mul.ftz.f32 %f2783, %f29692, %f2780; ld.local.f32 %f29693, [%rd415+108]; ld.local.v4.f32 {%f15451, %f15452, %f15453, %f15454}, [%rd415+48]; mul.ftz.f32 %f15458, %f2768, %f15452; fma.rn.ftz.f32 %f15459, %f2767, %f15451, %f15458; fma.rn.ftz.f32 %f15460, %f2769, %f15453, %f15459; setp.geu.ftz.f32 %p1557, %f15460, 0f00000000; @%p1557 bra $L__BB107_1775; rcp.approx.ftz.f32 %f29693, %f29693; $L__BB107_1775: setp.eq.s32 %p1558, %r3035, 0; fma.rn.ftz.f32 %f15461, %f2781, %f29693, %f2767; fma.rn.ftz.f32 %f15462, %f2782, %f29693, %f2768; fma.rn.ftz.f32 %f15463, %f2783, %f29693, %f2769; mul.ftz.f32 %f15464, %f15462, %f15462; fma.rn.ftz.f32 %f15465, %f15461, %f15461, %f15464; fma.rn.ftz.f32 %f15466, %f15463, %f15463, %f15465; rsqrt.approx.ftz.f32 %f15467, %f15466; mul.ftz.f32 %f15468, %f15461, %f15467; neg.ftz.f32 %f2787, %f15468; mul.ftz.f32 %f15469, %f15462, %f15467; neg.ftz.f32 %f2788, %f15469; mul.ftz.f32 %f15470, %f15463, %f15467; neg.ftz.f32 %f2789, %f15470; mul.ftz.f32 %f15471, %f29693, %f15467; mul.ftz.f32 %f2790, %f2780, %f15471; mul.ftz.f32 %f2791, %f2766, %f15467; ld.local.f32 %f2792, [%rd415+36]; ld.local.f32 %f2793, [%rd415+12]; ld.local.f32 %f2794, [%rd415+40]; ld.local.f32 %f2795, [%rd415+16]; mul.ftz.f32 %f15472, %f2795, %f2794; fma.rn.ftz.f32 %f15473, %f2793, %f2792, %f15472; ld.local.f32 %f2796, [%rd415+44]; ld.local.f32 %f2797, [%rd415+20]; fma.rn.ftz.f32 %f2798, %f2797, %f2796, %f15473; mul.ftz.f32 %f15474, %f2792, %f2798; mul.ftz.f32 %f15475, %f2794, %f2798; mul.ftz.f32 %f15476, %f2798, %f2796; sub.ftz.f32 %f15477, %f2793, %f15474; sub.ftz.f32 %f15478, %f2795, %f15475; sub.ftz.f32 %f15479, %f2797, %f15476; mul.ftz.f32 %f15480, %f15478, %f15478; fma.rn.ftz.f32 %f15481, %f15477, %f15477, %f15480; fma.rn.ftz.f32 %f15482, %f15479, %f15479, %f15481; rsqrt.approx.ftz.f32 %f2799, %f15482; mul.ftz.f32 %f2800, %f15477, %f2799; mul.ftz.f32 %f2801, %f15478, %f2799; mul.ftz.f32 %f2802, %f15479, %f2799; mul.ftz.f32 %f15483, %f2794, %f2802; mul.ftz.f32 %f15484, %f2796, %f2801; sub.ftz.f32 %f2803, %f15483, %f15484; mul.ftz.f32 %f15485, %f2796, %f2800; mul.ftz.f32 %f15486, %f2802, %f2792; sub.ftz.f32 %f2804, %f15485, %f15486; mul.ftz.f32 %f15487, %f2801, %f2792; mul.ftz.f32 %f15488, %f2800, %f2794; sub.ftz.f32 %f2805, %f15487, %f15488; @%p1558 bra $L__BB107_1777; ld.local.f32 %f15489, [%rd415+-180]; ld.local.f32 %f15490, [%rd415+-176]; mul.ftz.f32 %f15491, %f2768, %f15490; fma.rn.ftz.f32 %f15492, %f2767, %f15489, %f15491; ld.local.f32 %f15493, [%rd415+-172]; fma.rn.ftz.f32 %f15494, %f2769, %f15493, %f15492; mul.ftz.f32 %f15495, %f2767, %f15494; mul.ftz.f32 %f15496, %f2768, %f15494; mul.ftz.f32 %f15497, %f2769, %f15494; sub.ftz.f32 %f15498, %f15489, %f15495; sub.ftz.f32 %f15499, %f15490, %f15496; sub.ftz.f32 %f15500, %f15493, %f15497; mul.ftz.f32 %f15501, %f2791, %f15498; mul.ftz.f32 %f15502, %f2791, %f15499; mul.ftz.f32 %f15503, %f2791, %f15500; ld.local.f32 %f15504, [%rd416+20]; ld.local.f32 %f15505, [%rd416+24]; mul.ftz.f32 %f15506, %f2768, %f15505; fma.rn.ftz.f32 %f15507, %f2767, %f15504, %f15506; ld.local.f32 %f15508, [%rd416+28]; fma.rn.ftz.f32 %f15509, %f2769, %f15508, %f15507; mul.ftz.f32 %f15510, %f2767, %f15509; mul.ftz.f32 %f15511, %f2768, %f15509; mul.ftz.f32 %f15512, %f2769, %f15509; sub.ftz.f32 %f15513, %f15504, %f15510; sub.ftz.f32 %f15514, %f15505, %f15511; sub.ftz.f32 %f15515, %f15508, %f15512; mul.ftz.f32 %f15516, %f2791, %f15513; mul.ftz.f32 %f15517, %f2791, %f15514; mul.ftz.f32 %f15518, %f2791, %f15515; mul.ftz.f32 %f15519, %f15502, %f2788; fma.rn.ftz.f32 %f15520, %f15501, %f2787, %f15519; fma.rn.ftz.f32 %f15521, %f15503, %f2789, %f15520; mul.ftz.f32 %f15522, %f15521, %f2787; mul.ftz.f32 %f15523, %f15521, %f2788; mul.ftz.f32 %f15524, %f15521, %f2789; sub.ftz.f32 %f15525, %f15501, %f15522; sub.ftz.f32 %f15526, %f15502, %f15523; sub.ftz.f32 %f15527, %f15503, %f15524; mul.ftz.f32 %f15528, %f15517, %f2788; fma.rn.ftz.f32 %f15529, %f15516, %f2787, %f15528; fma.rn.ftz.f32 %f15530, %f15518, %f2789, %f15529; mul.ftz.f32 %f15531, %f15530, %f2787; mul.ftz.f32 %f15532, %f15530, %f2788; mul.ftz.f32 %f15533, %f15530, %f2789; sub.ftz.f32 %f15534, %f15516, %f15531; sub.ftz.f32 %f15535, %f15517, %f15532; sub.ftz.f32 %f15536, %f15518, %f15533; neg.ftz.f32 %f15537, %f15525; neg.ftz.f32 %f15538, %f15534; mul.ftz.f32 %f15539, %f2800, %f15537; mul.ftz.f32 %f15540, %f2801, %f15526; sub.ftz.f32 %f15541, %f15539, %f15540; mul.ftz.f32 %f15542, %f2802, %f15527; mul.ftz.f32 %f15543, %f2800, %f15538; mul.ftz.f32 %f15544, %f2801, %f15535; sub.ftz.f32 %f15545, %f15543, %f15544; mul.ftz.f32 %f15546, %f2802, %f15536; mul.ftz.f32 %f15547, %f2803, %f15537; mul.ftz.f32 %f15548, %f2804, %f15526; sub.ftz.f32 %f15549, %f15547, %f15548; mul.ftz.f32 %f15550, %f2805, %f15527; mul.ftz.f32 %f15551, %f2803, %f15538; mul.ftz.f32 %f15552, %f2804, %f15535; sub.ftz.f32 %f15553, %f15551, %f15552; mul.ftz.f32 %f15554, %f2805, %f15536; sub.ftz.f32 %f15555, %f15553, %f15554; sub.ftz.f32 %f15556, %f15549, %f15550; sub.ftz.f32 %f15557, %f15545, %f15546; sub.ftz.f32 %f15558, %f15541, %f15542; st.local.v4.f32 [%rd415+144], {%f15558, %f15557, %f15556, %f15555}; $L__BB107_1777: and.pred %p1561, %p1261, %p1553; mul.ftz.f32 %f15559, %f2768, %f2795; fma.rn.ftz.f32 %f15560, %f2767, %f2793, %f15559; fma.rn.ftz.f32 %f2806, %f2769, %f2797, %f15560; @%p1561 bra $L__BB107_1779; bra.uni $L__BB107_1778; $L__BB107_1779: mul.ftz.f32 %f15598, %f2767, %f2806; sub.ftz.f32 %f15599, %f15598, %f2793; mul.ftz.f32 %f15600, %f2768, %f2806; sub.ftz.f32 %f15601, %f15600, %f2795; mul.ftz.f32 %f15602, %f2769, %f2806; sub.ftz.f32 %f15603, %f15602, %f2797; mul.ftz.f32 %f29697, %f2791, %f15599; mul.ftz.f32 %f29698, %f2791, %f15601; mul.ftz.f32 %f29699, %f2791, %f15603; ld.local.v2.f32 {%f15604, %f15605}, [%rd415+24]; mul.ftz.f32 %f15608, %f2768, %f15605; fma.rn.ftz.f32 %f15609, %f2767, %f15604, %f15608; ld.local.f32 %f15610, [%rd415+32]; fma.rn.ftz.f32 %f15611, %f2769, %f15610, %f15609; mul.ftz.f32 %f15612, %f2767, %f15611; mul.ftz.f32 %f15613, %f2768, %f15611; mul.ftz.f32 %f15614, %f2769, %f15611; sub.ftz.f32 %f15615, %f15612, %f15604; sub.ftz.f32 %f15616, %f15613, %f15605; sub.ftz.f32 %f15617, %f15614, %f15610; mul.ftz.f32 %f29694, %f2791, %f15615; mul.ftz.f32 %f29695, %f2791, %f15616; mul.ftz.f32 %f29696, %f2791, %f15617; bra.uni $L__BB107_1780; $L__BB107_1778: add.ftz.f32 %f15561, %f2791, %f2790; mul.ftz.f32 %f15562, %f15561, %f2793; mul.ftz.f32 %f15563, %f15561, %f2795; mul.ftz.f32 %f15564, %f15561, %f2797; mul.ftz.f32 %f15565, %f2791, %f2806; mul.ftz.f32 %f15566, %f2767, %f15565; mul.ftz.f32 %f15567, %f2768, %f15565; mul.ftz.f32 %f15568, %f2769, %f15565; sub.ftz.f32 %f15569, %f15566, %f15562; sub.ftz.f32 %f15570, %f15567, %f15563; sub.ftz.f32 %f15571, %f15568, %f15564; mul.ftz.f32 %f15572, %f2782, %f2795; fma.rn.ftz.f32 %f15573, %f2781, %f2793, %f15572; fma.rn.ftz.f32 %f15574, %f2783, %f2797, %f15573; mul.ftz.f32 %f15575, %f2790, %f15574; fma.rn.ftz.f32 %f29697, %f2781, %f15575, %f15569; fma.rn.ftz.f32 %f29698, %f2782, %f15575, %f15570; fma.rn.ftz.f32 %f29699, %f2783, %f15575, %f15571; ld.local.v2.f32 {%f15576, %f15577}, [%rd415+24]; mul.ftz.f32 %f15580, %f15561, %f15576; mul.ftz.f32 %f15581, %f15561, %f15577; ld.local.f32 %f15582, [%rd415+32]; mul.ftz.f32 %f15583, %f15561, %f15582; mul.ftz.f32 %f15584, %f2768, %f15577; fma.rn.ftz.f32 %f15585, %f2767, %f15576, %f15584; fma.rn.ftz.f32 %f15586, %f2769, %f15582, %f15585; mul.ftz.f32 %f15587, %f2791, %f15586; mul.ftz.f32 %f15588, %f2767, %f15587; mul.ftz.f32 %f15589, %f2768, %f15587; mul.ftz.f32 %f15590, %f2769, %f15587; sub.ftz.f32 %f15591, %f15588, %f15580; sub.ftz.f32 %f15592, %f15589, %f15581; sub.ftz.f32 %f15593, %f15590, %f15583; mul.ftz.f32 %f15594, %f2782, %f15577; fma.rn.ftz.f32 %f15595, %f2781, %f15576, %f15594; fma.rn.ftz.f32 %f15596, %f2783, %f15582, %f15595; mul.ftz.f32 %f15597, %f2790, %f15596; fma.rn.ftz.f32 %f29694, %f2781, %f15597, %f15591; fma.rn.ftz.f32 %f29695, %f2782, %f15597, %f15592; fma.rn.ftz.f32 %f29696, %f2783, %f15597, %f15593; $L__BB107_1780: mul.ftz.f32 %f15618, %f29698, %f2788; fma.rn.ftz.f32 %f15619, %f29697, %f2787, %f15618; fma.rn.ftz.f32 %f15620, %f29699, %f2789, %f15619; mul.ftz.f32 %f15621, %f15620, %f2787; mul.ftz.f32 %f15622, %f15620, %f2788; mul.ftz.f32 %f15623, %f15620, %f2789; sub.ftz.f32 %f15624, %f29697, %f15621; sub.ftz.f32 %f15625, %f29698, %f15622; sub.ftz.f32 %f15626, %f29699, %f15623; mul.ftz.f32 %f15627, %f29695, %f2788; fma.rn.ftz.f32 %f15628, %f29694, %f2787, %f15627; fma.rn.ftz.f32 %f15629, %f29696, %f2789, %f15628; mul.ftz.f32 %f15630, %f15629, %f2787; mul.ftz.f32 %f15631, %f15629, %f2788; mul.ftz.f32 %f15632, %f15629, %f2789; sub.ftz.f32 %f15633, %f29694, %f15630; sub.ftz.f32 %f15634, %f29695, %f15631; sub.ftz.f32 %f15635, %f29696, %f15632; neg.ftz.f32 %f15636, %f15624; neg.ftz.f32 %f15637, %f15633; ld.local.f32 %f15638, [%rd415+60]; ld.local.f32 %f15639, [%rd415+64]; mul.ftz.f32 %f15640, %f2795, %f15639; fma.rn.ftz.f32 %f15641, %f2793, %f15638, %f15640; ld.local.f32 %f15642, [%rd415+68]; fma.rn.ftz.f32 %f15643, %f2797, %f15642, %f15641; mul.ftz.f32 %f15644, %f2798, %f15638; mul.ftz.f32 %f15645, %f2798, %f15639; mul.ftz.f32 %f15646, %f2798, %f15642; fma.rn.ftz.f32 %f15647, %f2792, %f15643, %f15644; fma.rn.ftz.f32 %f15648, %f2794, %f15643, %f15645; fma.rn.ftz.f32 %f15649, %f15643, %f2796, %f15646; neg.ftz.f32 %f15650, %f2799; mul.ftz.f32 %f15651, %f15647, %f15650; mul.ftz.f32 %f15652, %f15648, %f15650; mul.ftz.f32 %f15653, %f15649, %f15650; ld.local.f32 %f15654, [%rd415+72]; ld.local.f32 %f15655, [%rd415+76]; mul.ftz.f32 %f15656, %f2795, %f15655; fma.rn.ftz.f32 %f15657, %f2793, %f15654, %f15656; ld.local.f32 %f15658, [%rd415+80]; fma.rn.ftz.f32 %f15659, %f2797, %f15658, %f15657; mul.ftz.f32 %f15660, %f2798, %f15654; mul.ftz.f32 %f15661, %f2798, %f15655; mul.ftz.f32 %f15662, %f2798, %f15658; fma.rn.ftz.f32 %f15663, %f2792, %f15659, %f15660; fma.rn.ftz.f32 %f15664, %f2794, %f15659, %f15661; fma.rn.ftz.f32 %f15665, %f15659, %f2796, %f15662; mul.ftz.f32 %f15666, %f15663, %f15650; mul.ftz.f32 %f15667, %f15664, %f15650; mul.ftz.f32 %f15668, %f15665, %f15650; mul.ftz.f32 %f15669, %f2801, %f15652; fma.rn.ftz.f32 %f15670, %f2800, %f15651, %f15669; fma.rn.ftz.f32 %f15671, %f2802, %f15653, %f15670; mul.ftz.f32 %f15672, %f2800, %f15671; mul.ftz.f32 %f15673, %f2801, %f15671; mul.ftz.f32 %f15674, %f2802, %f15671; sub.ftz.f32 %f15675, %f15651, %f15672; sub.ftz.f32 %f15676, %f15652, %f15673; sub.ftz.f32 %f15677, %f15653, %f15674; mul.ftz.f32 %f15678, %f2801, %f15667; fma.rn.ftz.f32 %f15679, %f2800, %f15666, %f15678; fma.rn.ftz.f32 %f15680, %f2802, %f15668, %f15679; mul.ftz.f32 %f15681, %f2800, %f15680; mul.ftz.f32 %f15682, %f2801, %f15680; mul.ftz.f32 %f15683, %f2802, %f15680; sub.ftz.f32 %f15684, %f15666, %f15681; sub.ftz.f32 %f15685, %f15667, %f15682; sub.ftz.f32 %f15686, %f15668, %f15683; mul.ftz.f32 %f15687, %f2802, %f15639; mul.ftz.f32 %f15688, %f2801, %f15642; sub.ftz.f32 %f15689, %f15687, %f15688; mul.ftz.f32 %f15690, %f2800, %f15642; mul.ftz.f32 %f15691, %f2802, %f15638; sub.ftz.f32 %f15692, %f15690, %f15691; mul.ftz.f32 %f15693, %f2801, %f15638; mul.ftz.f32 %f15694, %f2800, %f15639; sub.ftz.f32 %f15695, %f15693, %f15694; mul.ftz.f32 %f15696, %f15677, %f2794; mul.ftz.f32 %f15697, %f15676, %f2796; sub.ftz.f32 %f15698, %f15696, %f15697; mul.ftz.f32 %f15699, %f15675, %f2796; mul.ftz.f32 %f15700, %f15677, %f2792; sub.ftz.f32 %f15701, %f15699, %f15700; mul.ftz.f32 %f15702, %f15676, %f2792; mul.ftz.f32 %f15703, %f15675, %f2794; sub.ftz.f32 %f15704, %f15702, %f15703; add.ftz.f32 %f15705, %f15689, %f15698; add.ftz.f32 %f15706, %f15692, %f15701; add.ftz.f32 %f15707, %f15695, %f15704; mul.ftz.f32 %f15708, %f2802, %f15655; mul.ftz.f32 %f15709, %f2801, %f15658; sub.ftz.f32 %f15710, %f15708, %f15709; mul.ftz.f32 %f15711, %f2800, %f15658; mul.ftz.f32 %f15712, %f2802, %f15654; sub.ftz.f32 %f15713, %f15711, %f15712; mul.ftz.f32 %f15714, %f2801, %f15654; mul.ftz.f32 %f15715, %f2800, %f15655; sub.ftz.f32 %f15716, %f15714, %f15715; mul.ftz.f32 %f15717, %f15686, %f2794; mul.ftz.f32 %f15718, %f15685, %f2796; sub.ftz.f32 %f15719, %f15717, %f15718; mul.ftz.f32 %f15720, %f15684, %f2796; mul.ftz.f32 %f15721, %f15686, %f2792; sub.ftz.f32 %f15722, %f15720, %f15721; mul.ftz.f32 %f15723, %f15685, %f2792; mul.ftz.f32 %f15724, %f15684, %f2794; sub.ftz.f32 %f15725, %f15723, %f15724; add.ftz.f32 %f15726, %f15710, %f15719; add.ftz.f32 %f15727, %f15713, %f15722; add.ftz.f32 %f15728, %f15716, %f15725; mul.ftz.f32 %f15729, %f2800, %f15636; mul.ftz.f32 %f15730, %f2801, %f15625; sub.ftz.f32 %f15731, %f15729, %f15730; mul.ftz.f32 %f15732, %f2802, %f15626; sub.ftz.f32 %f15733, %f15731, %f15732; mul.ftz.f32 %f15734, %f15676, %f2788; fma.rn.ftz.f32 %f15735, %f15675, %f2787, %f15734; fma.rn.ftz.f32 %f15736, %f15677, %f2789, %f15735; mul.ftz.f32 %f15737, %f2800, %f15637; mul.ftz.f32 %f15738, %f2801, %f15634; sub.ftz.f32 %f15739, %f15737, %f15738; mul.ftz.f32 %f15740, %f2802, %f15635; sub.ftz.f32 %f15741, %f15739, %f15740; mul.ftz.f32 %f15742, %f15685, %f2788; fma.rn.ftz.f32 %f15743, %f15684, %f2787, %f15742; fma.rn.ftz.f32 %f15744, %f15686, %f2789, %f15743; mul.ftz.f32 %f15745, %f2803, %f15636; mul.ftz.f32 %f15746, %f2804, %f15625; sub.ftz.f32 %f15747, %f15745, %f15746; mul.ftz.f32 %f15748, %f2805, %f15626; sub.ftz.f32 %f15749, %f15747, %f15748; mul.ftz.f32 %f15750, %f15706, %f2788; fma.rn.ftz.f32 %f15751, %f15705, %f2787, %f15750; fma.rn.ftz.f32 %f15752, %f15707, %f2789, %f15751; mul.ftz.f32 %f15753, %f2803, %f15637; mul.ftz.f32 %f15754, %f2804, %f15634; sub.ftz.f32 %f15755, %f15753, %f15754; mul.ftz.f32 %f15756, %f2805, %f15635; sub.ftz.f32 %f15757, %f15755, %f15756; mul.ftz.f32 %f15758, %f15727, %f2788; fma.rn.ftz.f32 %f15759, %f15726, %f2787, %f15758; fma.rn.ftz.f32 %f15760, %f15728, %f2789, %f15759; add.ftz.f32 %f15761, %f15757, %f15760; add.ftz.f32 %f15762, %f15749, %f15752; add.ftz.f32 %f15763, %f15741, %f15744; add.ftz.f32 %f15764, %f15733, %f15736; st.local.v4.f32 [%rd415+160], {%f15764, %f15763, %f15762, %f15761}; setp.ge.s32 %p1562, %r3035, %r3034; add.s32 %r5734, %r3035, 1; @%p1562 bra $L__BB107_1782; ld.local.f32 %f15765, [%rd415+204]; ld.local.f32 %f15766, [%rd415+208]; mul.ftz.f32 %f15767, %f2782, %f15766; fma.rn.ftz.f32 %f15768, %f2781, %f15765, %f15767; ld.local.f32 %f15769, [%rd415+212]; fma.rn.ftz.f32 %f15770, %f2783, %f15769, %f15768; mul.ftz.f32 %f15771, %f2781, %f15770; mul.ftz.f32 %f15772, %f2782, %f15770; mul.ftz.f32 %f15773, %f2783, %f15770; sub.ftz.f32 %f15774, %f15765, %f15771; sub.ftz.f32 %f15775, %f15766, %f15772; sub.ftz.f32 %f15776, %f15769, %f15773; mul.ftz.f32 %f15777, %f2790, %f15774; mul.ftz.f32 %f15778, %f2790, %f15775; mul.ftz.f32 %f15779, %f2790, %f15776; ld.local.f32 %f15780, [%rd415+216]; ld.local.f32 %f15781, [%rd415+220]; mul.ftz.f32 %f15782, %f2782, %f15781; fma.rn.ftz.f32 %f15783, %f2781, %f15780, %f15782; ld.local.f32 %f15784, [%rd415+224]; fma.rn.ftz.f32 %f15785, %f2783, %f15784, %f15783; mul.ftz.f32 %f15786, %f2781, %f15785; mul.ftz.f32 %f15787, %f2782, %f15785; mul.ftz.f32 %f15788, %f2783, %f15785; sub.ftz.f32 %f15789, %f15780, %f15786; sub.ftz.f32 %f15790, %f15781, %f15787; sub.ftz.f32 %f15791, %f15784, %f15788; mul.ftz.f32 %f15792, %f2790, %f15789; mul.ftz.f32 %f15793, %f2790, %f15790; mul.ftz.f32 %f15794, %f2790, %f15791; mul.ftz.f32 %f15795, %f15778, %f2788; fma.rn.ftz.f32 %f15796, %f15777, %f2787, %f15795; fma.rn.ftz.f32 %f15797, %f15779, %f2789, %f15796; mul.ftz.f32 %f15798, %f15797, %f2787; mul.ftz.f32 %f15799, %f15797, %f2788; mul.ftz.f32 %f15800, %f15797, %f2789; sub.ftz.f32 %f15801, %f15777, %f15798; sub.ftz.f32 %f15802, %f15778, %f15799; sub.ftz.f32 %f15803, %f15779, %f15800; mul.ftz.f32 %f15804, %f15793, %f2788; fma.rn.ftz.f32 %f15805, %f15792, %f2787, %f15804; fma.rn.ftz.f32 %f15806, %f15794, %f2789, %f15805; mul.ftz.f32 %f15807, %f15806, %f2787; mul.ftz.f32 %f15808, %f15806, %f2788; mul.ftz.f32 %f15809, %f15806, %f2789; sub.ftz.f32 %f15810, %f15792, %f15807; sub.ftz.f32 %f15811, %f15793, %f15808; sub.ftz.f32 %f15812, %f15794, %f15809; neg.ftz.f32 %f15813, %f15801; neg.ftz.f32 %f15814, %f15810; mul.ftz.f32 %f15815, %f2800, %f15813; mul.ftz.f32 %f15816, %f2801, %f15802; sub.ftz.f32 %f15817, %f15815, %f15816; mul.ftz.f32 %f15818, %f2802, %f15803; mul.ftz.f32 %f15819, %f2800, %f15814; mul.ftz.f32 %f15820, %f2801, %f15811; sub.ftz.f32 %f15821, %f15819, %f15820; mul.ftz.f32 %f15822, %f2802, %f15812; mul.ftz.f32 %f15823, %f2803, %f15813; mul.ftz.f32 %f15824, %f2804, %f15802; sub.ftz.f32 %f15825, %f15823, %f15824; mul.ftz.f32 %f15826, %f2805, %f15803; mul.ftz.f32 %f15827, %f2803, %f15814; mul.ftz.f32 %f15828, %f2804, %f15811; sub.ftz.f32 %f15829, %f15827, %f15828; mul.ftz.f32 %f15830, %f2805, %f15812; sub.ftz.f32 %f15831, %f15829, %f15830; sub.ftz.f32 %f15832, %f15825, %f15826; sub.ftz.f32 %f15833, %f15821, %f15822; sub.ftz.f32 %f15834, %f15817, %f15818; st.local.v4.f32 [%rd415+176], {%f15834, %f15833, %f15832, %f15831}; $L__BB107_1782: mul.ftz.f32 %f15835, %f2801, %f2788; fma.rn.ftz.f32 %f15836, %f2800, %f2787, %f15835; fma.rn.ftz.f32 %f15837, %f2802, %f2789, %f15836; mul.ftz.f32 %f15838, %f2804, %f2788; fma.rn.ftz.f32 %f15839, %f2803, %f2787, %f15838; fma.rn.ftz.f32 %f15840, %f2805, %f2789, %f15839; ld.local.v2.f32 {%f15841, %f15842}, [%rd415+120]; sub.ftz.f32 %f15845, %f15840, %f15842; sub.ftz.f32 %f15846, %f15837, %f15841; st.local.v2.f32 [%rd415+128], {%f15846, %f15845}; setp.lt.s32 %p1563, %r5734, %r5732; @%p1563 bra $L__BB107_1764; setp.lt.u32 %p1564, %r3034, 3; mov.f32 %f29703, 0f00000000; mov.u32 %r5737, 0; @%p1564 bra $L__BB107_1786; mov.f32 %f29703, 0f00000000; mov.u32 %r5737, 0; mov.u32 %r5736, %r808; $L__BB107_1785: mul.wide.s32 %rd2591, %r5737, 192; add.s64 %rd2592, %rd1, %rd2591; ld.local.v2.f32 {%f15850, %f15851}, [%rd2592+128]; mul.ftz.f32 %f15854, %f15851, %f15851; fma.rn.ftz.f32 %f15855, %f15850, %f15850, %f15854; sqrt.approx.ftz.f32 %f15856, %f15855; max.ftz.f32 %f15857, %f29703, %f15856; ld.local.v2.f32 {%f15858, %f15859}, [%rd2592+320]; mul.ftz.f32 %f15862, %f15859, %f15859; fma.rn.ftz.f32 %f15863, %f15858, %f15858, %f15862; sqrt.approx.ftz.f32 %f15864, %f15863; max.ftz.f32 %f15865, %f15857, %f15864; ld.local.v2.f32 {%f15866, %f15867}, [%rd2592+512]; mul.ftz.f32 %f15870, %f15867, %f15867; fma.rn.ftz.f32 %f15871, %f15866, %f15866, %f15870; sqrt.approx.ftz.f32 %f15872, %f15871; max.ftz.f32 %f15873, %f15865, %f15872; ld.local.v2.f32 {%f15874, %f15875}, [%rd2592+704]; mul.ftz.f32 %f15878, %f15875, %f15875; fma.rn.ftz.f32 %f15879, %f15874, %f15874, %f15878; sqrt.approx.ftz.f32 %f15880, %f15879; max.ftz.f32 %f29703, %f15873, %f15880; add.s32 %r5737, %r5737, 4; add.s32 %r5736, %r5736, -4; setp.ne.s32 %p1565, %r5736, 0; @%p1565 bra $L__BB107_1785; $L__BB107_1786: setp.eq.s32 %p1566, %r807, 0; @%p1566 bra $L__BB107_1790; setp.eq.s32 %p1567, %r807, 1; mul.wide.s32 %rd2593, %r5737, 192; add.s64 %rd2594, %rd1, %rd2593; add.s64 %rd417, %rd2594, 128; ld.local.v2.f32 {%f15881, %f15882}, [%rd2594+128]; mul.ftz.f32 %f15885, %f15882, %f15882; fma.rn.ftz.f32 %f15886, %f15881, %f15881, %f15885; sqrt.approx.ftz.f32 %f15887, %f15886; max.ftz.f32 %f29703, %f29703, %f15887; @%p1567 bra $L__BB107_1790; setp.eq.s32 %p1568, %r807, 2; ld.local.v2.f32 {%f15888, %f15889}, [%rd417+192]; mul.ftz.f32 %f15892, %f15889, %f15889; fma.rn.ftz.f32 %f15893, %f15888, %f15888, %f15892; sqrt.approx.ftz.f32 %f15894, %f15893; max.ftz.f32 %f29703, %f29703, %f15894; @%p1568 bra $L__BB107_1790; ld.local.v2.f32 {%f15895, %f15896}, [%rd417+384]; mul.ftz.f32 %f15899, %f15896, %f15896; fma.rn.ftz.f32 %f15900, %f15895, %f15895, %f15899; sqrt.approx.ftz.f32 %f15901, %f15900; max.ftz.f32 %f29703, %f29703, %f15901; $L__BB107_1790: setp.lt.ftz.f32 %p1570, %f29703, 0f3A83126F; mov.pred %p3921, %p1284; @%p1570 bra $L__BB107_2250; ld.local.v4.f32 {%f15903, %f15904, %f15905, %f15906}, [%rd412+124]; mul.ftz.f32 %f15907, %f15903, %f15906; mul.ftz.f32 %f15908, %f15904, %f15905; sub.ftz.f32 %f2837, %f15907, %f15908; abs.ftz.f32 %f15909, %f2837; setp.lt.ftz.f32 %p1571, %f15909, 0f38D1B717; mov.f32 %f29704, 0f00000000; @%p1571 bra $L__BB107_1793; rcp.approx.ftz.f32 %f15910, %f2837; mul.ftz.f32 %f15911, %f15904, %f15910; mul.ftz.f32 %f15912, %f15905, %f15910; mul.ftz.f32 %f15913, %f15903, %f15910; mul.ftz.f32 %f15914, %f15906, %f15910; neg.ftz.f32 %f15915, %f15912; neg.ftz.f32 %f15916, %f15911; add.u64 %rd2596, %SPL, 0; st.local.v4.f32 [%rd2596], {%f15914, %f15916, %f15915, %f15913}; mov.f32 %f29704, %f2837; $L__BB107_1793: setp.eq.ftz.f32 %p1573, %f29704, 0f00000000; mov.pred %p3921, -1; mov.pred %p3920, %p3921; @%p1573 bra $L__BB107_1807; bra.uni $L__BB107_1794; $L__BB107_1807: @%p3920 bra $L__BB107_2250; $L__BB107_1808: mov.u16 %rs713, %rs351; @%p1589 bra $L__BB107_2228; add.u64 %rd2619, %SP, 1404; add.u64 %rd427, %SPL, 1404; mov.u32 %r5741, 0; $L__BB107_1810: cvt.s64.s32 %rd428, %r5741; mul.wide.s32 %rd2620, %r5741, 192; add.s64 %rd429, %rd1, %rd2620; mul.wide.s32 %rd2621, %r5741, 8; add.s64 %rd2622, %rd409, %rd2621; ld.local.v2.f32 {%f16214, %f16215}, [%rd2622]; ld.local.v4.f32 {%f16218, %f16219, %f16220, %f16221}, [%rd429]; ld.local.v4.f32 {%f16222, %f16223, %f16224, %f16225}, [%rd429+16]; mul.ftz.f32 %f16226, %f16215, %f16224; mul.ftz.f32 %f16227, %f16215, %f16225; ld.local.f32 %f2871, [%rd429+32]; mul.ftz.f32 %f16228, %f16215, %f2871; fma.rn.ftz.f32 %f16229, %f16214, %f16221, %f16226; fma.rn.ftz.f32 %f16230, %f16214, %f16222, %f16227; fma.rn.ftz.f32 %f16231, %f16214, %f16223, %f16228; mul.ftz.f32 %f16232, %f29683, %f16229; mul.ftz.f32 %f16233, %f29683, %f16230; mul.ftz.f32 %f16234, %f29683, %f16231; sub.ftz.f32 %f29706, %f16218, %f16232; sub.ftz.f32 %f29707, %f16219, %f16233; sub.ftz.f32 %f29708, %f16220, %f16234; setp.ne.s32 %p1590, %r5741, 0; @%p1590 bra $L__BB107_1813; ld.local.v4.f32 {%f16235, %f16236, %f16237, %f16238}, [%rd279]; sub.ftz.f32 %f16242, %f29706, %f16235; sub.ftz.f32 %f16243, %f29707, %f16236; sub.ftz.f32 %f16244, %f29708, %f16237; ld.local.v2.f32 {%f16245, %f16246}, [%rd5+-64]; mul.ftz.f32 %f16249, %f16243, %f16246; fma.rn.ftz.f32 %f16250, %f16242, %f16245, %f16249; ld.local.f32 %f16251, [%rd5+-56]; fma.rn.ftz.f32 %f16252, %f16244, %f16251, %f16250; setp.gtu.ftz.f32 %p1591, %f16252, 0f00000000; @%p1591 bra $L__BB107_1813; ld.local.v2.f32 {%f16253, %f16254}, [%rd409]; mul.ftz.f32 %f16257, %f16254, %f16224; mul.ftz.f32 %f16258, %f16254, %f16225; mul.ftz.f32 %f16259, %f16254, %f2871; fma.rn.ftz.f32 %f16260, %f16253, %f16221, %f16257; fma.rn.ftz.f32 %f16261, %f16253, %f16222, %f16258; fma.rn.ftz.f32 %f16262, %f16253, %f16223, %f16259; fma.rn.ftz.f32 %f29706, %f29683, %f16260, %f16218; fma.rn.ftz.f32 %f29707, %f29683, %f16261, %f16219; fma.rn.ftz.f32 %f29708, %f29683, %f16262, %f16220; $L__BB107_1813: cvt.u32.u64 %r3047, %rd428; setp.eq.s32 %p1592, %r3047, 0; @%p1592 bra $L__BB107_1815; add.s32 %r3049, %r3047, -1; mul.wide.s32 %rd2625, %r3049, 192; add.s64 %rd2626, %rd1, %rd2625; ld.local.v2.u32 {%r3050, %r3051}, [%rd2626+96]; st.local.u32 [%rd406+-8], %r3050; st.local.u32 [%rd406+-12], %r3051; ld.local.v4.f32 {%f29709, %f29710, %f29711, %f16266}, [%rd429+-192]; bra.uni $L__BB107_1816; $L__BB107_1815: ld.local.u32 %r3054, [%rd5+-12]; st.local.u32 [%rd406+-8], %r3054; ld.local.u32 %r3055, [%rd5+-28]; st.local.u32 [%rd406+-12], %r3055; ld.local.v4.f32 {%f29709, %f29710, %f29711, %f16270}, [%rd279]; $L__BB107_1816: st.local.f32 [%rd406+-48], %f29709; st.local.f32 [%rd406+-44], %f29710; st.local.f32 [%rd406+-40], %f29711; sub.ftz.f32 %f16271, %f29706, %f29709; sub.ftz.f32 %f16272, %f29707, %f29710; mul.ftz.f32 %f16273, %f16272, %f16272; fma.rn.ftz.f32 %f16274, %f16271, %f16271, %f16273; sub.ftz.f32 %f16275, %f29708, %f29711; fma.rn.ftz.f32 %f16276, %f16275, %f16275, %f16274; sqrt.approx.ftz.f32 %f16277, %f16276; rsqrt.approx.ftz.f32 %f16278, %f16276; mul.ftz.f32 %f2892, %f16271, %f16278; mul.ftz.f32 %f2893, %f16272, %f16278; mul.ftz.f32 %f2894, %f16275, %f16278; st.local.f32 [%rd406+-36], %f2892; st.local.f32 [%rd406+-32], %f2893; st.local.f32 [%rd406+-28], %f2894; add.ftz.f32 %f16279, %f16277, %f16277; st.local.f32 [%rd406+-20], %f16279; mov.b32 %r827, %f29709; @%p1593 bra $L__BB107_1988; setp.nan.ftz.f32 %p1594, %f29709, %f29709; mov.u16 %rs632, 0; mov.u16 %rs629, %rs632; @%p1594 bra $L__BB107_1820; or.b32 %r3056, %r827, -2147483648; setp.ne.s32 %p1595, %r3056, -2147483648; add.ftz.f32 %f16280, %f29709, %f29709; setp.eq.ftz.f32 %p1596, %f16280, %f29709; and.pred %p1597, %p1596, %p1595; mov.u16 %rs629, %rs632; @%p1597 bra $L__BB107_1820; shl.b32 %r3057, %r827, 1; setp.lt.u32 %p1598, %r3057, -16777215; selp.u16 %rs629, 1, 0, %p1598; $L__BB107_1820: setp.eq.s16 %p1599, %rs629, 0; mov.u16 %rs631, %rs632; @%p1599 bra $L__BB107_1826; mov.b32 %r828, %f2892; setp.nan.ftz.f32 %p1600, %f2892, %f2892; mov.u16 %rs631, 0; mov.u16 %rs630, %rs631; @%p1600 bra $L__BB107_1824; or.b32 %r3058, %r828, -2147483648; setp.ne.s32 %p1601, %r3058, -2147483648; add.ftz.f32 %f16281, %f2892, %f2892; setp.eq.ftz.f32 %p1602, %f16281, %f2892; and.pred %p1603, %p1602, %p1601; mov.u16 %rs630, %rs631; @%p1603 bra $L__BB107_1824; shl.b32 %r3059, %r828, 1; setp.lt.u32 %p1604, %r3059, -16777215; selp.u16 %rs630, 1, 0, %p1604; $L__BB107_1824: setp.eq.s16 %p1605, %rs630, 0; @%p1605 bra $L__BB107_1826; mul.ftz.f32 %f16282, %f2893, %f2893; fma.rn.ftz.f32 %f16283, %f2892, %f2892, %f16282; fma.rn.ftz.f32 %f16284, %f2894, %f2894, %f16283; setp.neu.ftz.f32 %p1606, %f16284, 0f00000000; selp.u16 %rs631, 1, 0, %p1606; $L__BB107_1826: setp.eq.s16 %p1607, %rs631, 0; @%p1607 bra $L__BB107_1830; @%p1969 bra $L__BB107_1829; bra.uni $L__BB107_1828; $L__BB107_1829: mov.u32 %r3062, 4; { // callseq 421, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3062; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3063, [retval0+0]; } // callseq 421 cvt.u16.u32 %rs632, %r3063; bra.uni $L__BB107_1830; $L__BB107_1988: setp.nan.ftz.f32 %p1784, %f29709, %f29709; mov.u16 %rs669, %rs351; @%p1784 bra $L__BB107_1991; or.b32 %r3176, %r827, -2147483648; setp.ne.s32 %p1785, %r3176, -2147483648; add.ftz.f32 %f16360, %f29709, %f29709; setp.eq.ftz.f32 %p1786, %f16360, %f29709; and.pred %p1787, %p1786, %p1785; mov.u16 %rs669, %rs351; @%p1787 bra $L__BB107_1991; shl.b32 %r3177, %r827, 1; setp.lt.u32 %p1788, %r3177, -16777215; selp.u16 %rs669, 1, 0, %p1788; $L__BB107_1991: setp.eq.s16 %p1789, %rs669, 0; mov.u16 %rs671, %rs351; @%p1789 bra $L__BB107_1997; mov.b32 %r869, %f2892; setp.nan.ftz.f32 %p1790, %f2892, %f2892; mov.u16 %rs671, 0; mov.u16 %rs670, %rs671; @%p1790 bra $L__BB107_1995; or.b32 %r3178, %r869, -2147483648; setp.ne.s32 %p1791, %r3178, -2147483648; add.ftz.f32 %f16361, %f2892, %f2892; setp.eq.ftz.f32 %p1792, %f16361, %f2892; and.pred %p1793, %p1792, %p1791; mov.u16 %rs670, %rs671; @%p1793 bra $L__BB107_1995; shl.b32 %r3179, %r869, 1; setp.lt.u32 %p1794, %r3179, -16777215; selp.u16 %rs670, 1, 0, %p1794; $L__BB107_1995: setp.eq.s16 %p1795, %rs670, 0; @%p1795 bra $L__BB107_1997; mul.ftz.f32 %f16362, %f2893, %f2893; fma.rn.ftz.f32 %f16363, %f2892, %f2892, %f16362; fma.rn.ftz.f32 %f16364, %f2894, %f2894, %f16363; setp.neu.ftz.f32 %p1796, %f16364, 0f00000000; selp.u16 %rs671, 1, 0, %p1796; $L__BB107_1997: setp.eq.s16 %p1797, %rs671, 0; mov.u16 %rs672, %rs351; @%p1797 bra $L__BB107_2001; @%p1969 bra $L__BB107_2000; bra.uni $L__BB107_1999; $L__BB107_2000: mov.u32 %r3182, 4; { // callseq 441, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3182; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3183, [retval0+0]; } // callseq 441 cvt.u16.u32 %rs672, %r3183; bra.uni $L__BB107_2001; $L__BB107_1828: mov.u32 %r3060, 4; { // callseq 420, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3060; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3061, [retval0+0]; } // callseq 420 cvt.u16.u32 %rs632, %r3061; $L__BB107_1830: setp.eq.s16 %p1609, %rs632, 0; mov.u16 %rs713, %rs349; @%p1609 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1610, %r5762, -1; mov.u32 %r5742, %r5762; @%p1610 bra $L__BB107_1833; ld.local.u32 %r3064, [%rd427+12]; mul.wide.s32 %rd2631, %r3064, 4; add.s64 %rd2632, %rd353, %rd2631; ld.global.u32 %r5742, [%rd2632]; $L__BB107_1833: ld.local.u32 %r832, [%rd429+96]; setp.eq.s32 %p1611, %r5742, %r832; @%p1611 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3065, [%rd427+12]; st.local.u32 [%rd406+-12], %r3065; ld.local.f32 %f16285, [%rd427]; setp.eq.ftz.f32 %p1613, %f16285, 0f00000000; mov.b32 %r3066, %f16285; add.s32 %r3067, %r3066, 1; mov.b32 %f16286, %r3067; selp.f32 %f16287, 0f00000001, %f16286, %p1613; st.local.f32 [%rd406+-24], %f16287; mov.u16 %rs636, 0; mov.u16 %rs633, %rs636; @%p1594 bra $L__BB107_1837; or.b32 %r3068, %r827, -2147483648; setp.ne.s32 %p1614, %r3068, -2147483648; add.ftz.f32 %f16288, %f29709, %f29709; setp.eq.ftz.f32 %p1615, %f16288, %f29709; and.pred %p1616, %p1615, %p1614; mov.u16 %rs633, %rs636; @%p1616 bra $L__BB107_1837; shl.b32 %r3069, %r827, 1; setp.lt.u32 %p1617, %r3069, -16777215; selp.u16 %rs633, 1, 0, %p1617; $L__BB107_1837: setp.eq.s16 %p1618, %rs633, 0; mov.u16 %rs635, %rs636; @%p1618 bra $L__BB107_1843; mov.b32 %r833, %f2892; setp.nan.ftz.f32 %p1619, %f2892, %f2892; mov.u16 %rs635, 0; mov.u16 %rs634, %rs635; @%p1619 bra $L__BB107_1841; or.b32 %r3070, %r833, -2147483648; setp.ne.s32 %p1620, %r3070, -2147483648; add.ftz.f32 %f16289, %f2892, %f2892; setp.eq.ftz.f32 %p1621, %f16289, %f2892; and.pred %p1622, %p1621, %p1620; mov.u16 %rs634, %rs635; @%p1622 bra $L__BB107_1841; shl.b32 %r3071, %r833, 1; setp.lt.u32 %p1623, %r3071, -16777215; selp.u16 %rs634, 1, 0, %p1623; $L__BB107_1841: setp.eq.s16 %p1624, %rs634, 0; @%p1624 bra $L__BB107_1843; mul.ftz.f32 %f16290, %f2893, %f2893; fma.rn.ftz.f32 %f16291, %f2892, %f2892, %f16290; fma.rn.ftz.f32 %f16292, %f2894, %f2894, %f16291; setp.neu.ftz.f32 %p1625, %f16292, 0f00000000; selp.u16 %rs635, 1, 0, %p1625; $L__BB107_1843: setp.eq.s16 %p1626, %rs635, 0; @%p1626 bra $L__BB107_1847; @%p1969 bra $L__BB107_1846; bra.uni $L__BB107_1845; $L__BB107_1846: mov.u32 %r3074, 4; { // callseq 423, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3074; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3075, [retval0+0]; } // callseq 423 cvt.u16.u32 %rs636, %r3075; bra.uni $L__BB107_1847; $L__BB107_1999: mov.u32 %r3180, 4; { // callseq 440, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3180; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3181, [retval0+0]; } // callseq 440 cvt.u16.u32 %rs672, %r3181; $L__BB107_2001: setp.eq.s16 %p1799, %rs672, 0; mov.u16 %rs713, %rs349; @%p1799 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1800, %r5762, -1; mov.u32 %r5752, %r5762; @%p1800 bra $L__BB107_2004; ld.local.u32 %r3184, [%rd427+12]; mul.wide.s32 %rd2691, %r3184, 4; add.s64 %rd2692, %rd353, %rd2691; ld.global.u32 %r5752, [%rd2692]; $L__BB107_2004: ld.local.u32 %r873, [%rd429+96]; setp.eq.s32 %p1801, %r5752, %r873; @%p1801 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3185, [%rd427+12]; st.local.u32 [%rd406+-12], %r3185; ld.local.f32 %f16365, [%rd427]; setp.eq.ftz.f32 %p1803, %f16365, 0f00000000; mov.b32 %r3186, %f16365; add.s32 %r3187, %r3186, 1; mov.b32 %f16366, %r3187; selp.f32 %f16367, 0f00000001, %f16366, %p1803; st.local.f32 [%rd406+-24], %f16367; mov.u16 %rs673, %rs351; @%p1784 bra $L__BB107_2008; or.b32 %r3188, %r827, -2147483648; setp.ne.s32 %p1804, %r3188, -2147483648; add.ftz.f32 %f16368, %f29709, %f29709; setp.eq.ftz.f32 %p1805, %f16368, %f29709; and.pred %p1806, %p1805, %p1804; mov.u16 %rs673, %rs351; @%p1806 bra $L__BB107_2008; shl.b32 %r3189, %r827, 1; setp.lt.u32 %p1807, %r3189, -16777215; selp.u16 %rs673, 1, 0, %p1807; $L__BB107_2008: setp.eq.s16 %p1808, %rs673, 0; mov.u16 %rs675, %rs351; @%p1808 bra $L__BB107_2014; mov.b32 %r874, %f2892; setp.nan.ftz.f32 %p1809, %f2892, %f2892; mov.u16 %rs675, 0; mov.u16 %rs674, %rs675; @%p1809 bra $L__BB107_2012; or.b32 %r3190, %r874, -2147483648; setp.ne.s32 %p1810, %r3190, -2147483648; add.ftz.f32 %f16369, %f2892, %f2892; setp.eq.ftz.f32 %p1811, %f16369, %f2892; and.pred %p1812, %p1811, %p1810; mov.u16 %rs674, %rs675; @%p1812 bra $L__BB107_2012; shl.b32 %r3191, %r874, 1; setp.lt.u32 %p1813, %r3191, -16777215; selp.u16 %rs674, 1, 0, %p1813; $L__BB107_2012: setp.eq.s16 %p1814, %rs674, 0; @%p1814 bra $L__BB107_2014; mul.ftz.f32 %f16370, %f2893, %f2893; fma.rn.ftz.f32 %f16371, %f2892, %f2892, %f16370; fma.rn.ftz.f32 %f16372, %f2894, %f2894, %f16371; setp.neu.ftz.f32 %p1815, %f16372, 0f00000000; selp.u16 %rs675, 1, 0, %p1815; $L__BB107_2014: setp.eq.s16 %p1816, %rs675, 0; mov.u16 %rs676, %rs351; @%p1816 bra $L__BB107_2018; @%p1969 bra $L__BB107_2017; bra.uni $L__BB107_2016; $L__BB107_2017: mov.u32 %r3194, 4; { // callseq 443, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3194; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3195, [retval0+0]; } // callseq 443 cvt.u16.u32 %rs676, %r3195; bra.uni $L__BB107_2018; $L__BB107_1845: mov.u32 %r3072, 4; { // callseq 422, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3072; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3073, [retval0+0]; } // callseq 422 cvt.u16.u32 %rs636, %r3073; $L__BB107_1847: setp.eq.s16 %p1628, %rs636, 0; mov.u16 %rs713, %rs349; @%p1628 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1629, %r5762, -1; mov.u32 %r5743, %r5762; @%p1629 bra $L__BB107_1850; ld.local.u32 %r3076, [%rd427+12]; mul.wide.s32 %rd2637, %r3076, 4; add.s64 %rd2638, %rd353, %rd2637; ld.global.u32 %r5743, [%rd2638]; $L__BB107_1850: setp.eq.s32 %p1630, %r5743, %r832; @%p1630 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3077, [%rd427+12]; st.local.u32 [%rd406+-12], %r3077; ld.local.f32 %f16293, [%rd427]; setp.eq.ftz.f32 %p1632, %f16293, 0f00000000; mov.b32 %r3078, %f16293; add.s32 %r3079, %r3078, 1; mov.b32 %f16294, %r3079; selp.f32 %f16295, 0f00000001, %f16294, %p1632; st.local.f32 [%rd406+-24], %f16295; mov.u16 %rs640, 0; mov.u16 %rs637, %rs640; @%p1594 bra $L__BB107_1854; or.b32 %r3080, %r827, -2147483648; setp.ne.s32 %p1633, %r3080, -2147483648; add.ftz.f32 %f16296, %f29709, %f29709; setp.eq.ftz.f32 %p1634, %f16296, %f29709; and.pred %p1635, %p1634, %p1633; mov.u16 %rs637, %rs640; @%p1635 bra $L__BB107_1854; shl.b32 %r3081, %r827, 1; setp.lt.u32 %p1636, %r3081, -16777215; selp.u16 %rs637, 1, 0, %p1636; $L__BB107_1854: setp.eq.s16 %p1637, %rs637, 0; mov.u16 %rs639, %rs640; @%p1637 bra $L__BB107_1860; mov.b32 %r837, %f2892; setp.nan.ftz.f32 %p1638, %f2892, %f2892; mov.u16 %rs639, 0; mov.u16 %rs638, %rs639; @%p1638 bra $L__BB107_1858; or.b32 %r3082, %r837, -2147483648; setp.ne.s32 %p1639, %r3082, -2147483648; add.ftz.f32 %f16297, %f2892, %f2892; setp.eq.ftz.f32 %p1640, %f16297, %f2892; and.pred %p1641, %p1640, %p1639; mov.u16 %rs638, %rs639; @%p1641 bra $L__BB107_1858; shl.b32 %r3083, %r837, 1; setp.lt.u32 %p1642, %r3083, -16777215; selp.u16 %rs638, 1, 0, %p1642; $L__BB107_1858: setp.eq.s16 %p1643, %rs638, 0; @%p1643 bra $L__BB107_1860; mul.ftz.f32 %f16298, %f2893, %f2893; fma.rn.ftz.f32 %f16299, %f2892, %f2892, %f16298; fma.rn.ftz.f32 %f16300, %f2894, %f2894, %f16299; setp.neu.ftz.f32 %p1644, %f16300, 0f00000000; selp.u16 %rs639, 1, 0, %p1644; $L__BB107_1860: setp.eq.s16 %p1645, %rs639, 0; @%p1645 bra $L__BB107_1864; @%p1969 bra $L__BB107_1863; bra.uni $L__BB107_1862; $L__BB107_1863: mov.u32 %r3086, 4; { // callseq 425, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3086; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3087, [retval0+0]; } // callseq 425 cvt.u16.u32 %rs640, %r3087; bra.uni $L__BB107_1864; $L__BB107_2016: mov.u32 %r3192, 4; { // callseq 442, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3192; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3193, [retval0+0]; } // callseq 442 cvt.u16.u32 %rs676, %r3193; $L__BB107_2018: setp.eq.s16 %p1818, %rs676, 0; mov.u16 %rs713, %rs349; @%p1818 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1819, %r5762, -1; mov.u32 %r5753, %r5762; @%p1819 bra $L__BB107_2021; ld.local.u32 %r3196, [%rd427+12]; mul.wide.s32 %rd2697, %r3196, 4; add.s64 %rd2698, %rd353, %rd2697; ld.global.u32 %r5753, [%rd2698]; $L__BB107_2021: setp.eq.s32 %p1820, %r5753, %r873; @%p1820 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3197, [%rd427+12]; st.local.u32 [%rd406+-12], %r3197; ld.local.f32 %f16373, [%rd427]; setp.eq.ftz.f32 %p1822, %f16373, 0f00000000; mov.b32 %r3198, %f16373; add.s32 %r3199, %r3198, 1; mov.b32 %f16374, %r3199; selp.f32 %f16375, 0f00000001, %f16374, %p1822; st.local.f32 [%rd406+-24], %f16375; mov.u16 %rs677, %rs351; @%p1784 bra $L__BB107_2025; or.b32 %r3200, %r827, -2147483648; setp.ne.s32 %p1823, %r3200, -2147483648; add.ftz.f32 %f16376, %f29709, %f29709; setp.eq.ftz.f32 %p1824, %f16376, %f29709; and.pred %p1825, %p1824, %p1823; mov.u16 %rs677, %rs351; @%p1825 bra $L__BB107_2025; shl.b32 %r3201, %r827, 1; setp.lt.u32 %p1826, %r3201, -16777215; selp.u16 %rs677, 1, 0, %p1826; $L__BB107_2025: setp.eq.s16 %p1827, %rs677, 0; mov.u16 %rs679, %rs351; @%p1827 bra $L__BB107_2031; mov.b32 %r878, %f2892; setp.nan.ftz.f32 %p1828, %f2892, %f2892; mov.u16 %rs679, 0; mov.u16 %rs678, %rs679; @%p1828 bra $L__BB107_2029; or.b32 %r3202, %r878, -2147483648; setp.ne.s32 %p1829, %r3202, -2147483648; add.ftz.f32 %f16377, %f2892, %f2892; setp.eq.ftz.f32 %p1830, %f16377, %f2892; and.pred %p1831, %p1830, %p1829; mov.u16 %rs678, %rs679; @%p1831 bra $L__BB107_2029; shl.b32 %r3203, %r878, 1; setp.lt.u32 %p1832, %r3203, -16777215; selp.u16 %rs678, 1, 0, %p1832; $L__BB107_2029: setp.eq.s16 %p1833, %rs678, 0; @%p1833 bra $L__BB107_2031; mul.ftz.f32 %f16378, %f2893, %f2893; fma.rn.ftz.f32 %f16379, %f2892, %f2892, %f16378; fma.rn.ftz.f32 %f16380, %f2894, %f2894, %f16379; setp.neu.ftz.f32 %p1834, %f16380, 0f00000000; selp.u16 %rs679, 1, 0, %p1834; $L__BB107_2031: setp.eq.s16 %p1835, %rs679, 0; mov.u16 %rs680, %rs351; @%p1835 bra $L__BB107_2035; @%p1969 bra $L__BB107_2034; bra.uni $L__BB107_2033; $L__BB107_2034: mov.u32 %r3206, 4; { // callseq 445, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3206; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3207, [retval0+0]; } // callseq 445 cvt.u16.u32 %rs680, %r3207; bra.uni $L__BB107_2035; $L__BB107_1862: mov.u32 %r3084, 4; { // callseq 424, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3084; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3085, [retval0+0]; } // callseq 424 cvt.u16.u32 %rs640, %r3085; $L__BB107_1864: setp.eq.s16 %p1647, %rs640, 0; mov.u16 %rs713, %rs349; @%p1647 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1648, %r5762, -1; mov.u32 %r5744, %r5762; @%p1648 bra $L__BB107_1867; ld.local.u32 %r3088, [%rd427+12]; mul.wide.s32 %rd2643, %r3088, 4; add.s64 %rd2644, %rd353, %rd2643; ld.global.u32 %r5744, [%rd2644]; $L__BB107_1867: setp.eq.s32 %p1649, %r5744, %r832; @%p1649 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3089, [%rd427+12]; st.local.u32 [%rd406+-12], %r3089; ld.local.f32 %f16301, [%rd427]; setp.eq.ftz.f32 %p1651, %f16301, 0f00000000; mov.b32 %r3090, %f16301; add.s32 %r3091, %r3090, 1; mov.b32 %f16302, %r3091; selp.f32 %f16303, 0f00000001, %f16302, %p1651; st.local.f32 [%rd406+-24], %f16303; mov.u16 %rs644, 0; mov.u16 %rs641, %rs644; @%p1594 bra $L__BB107_1871; or.b32 %r3092, %r827, -2147483648; setp.ne.s32 %p1652, %r3092, -2147483648; add.ftz.f32 %f16304, %f29709, %f29709; setp.eq.ftz.f32 %p1653, %f16304, %f29709; and.pred %p1654, %p1653, %p1652; mov.u16 %rs641, %rs644; @%p1654 bra $L__BB107_1871; shl.b32 %r3093, %r827, 1; setp.lt.u32 %p1655, %r3093, -16777215; selp.u16 %rs641, 1, 0, %p1655; $L__BB107_1871: setp.eq.s16 %p1656, %rs641, 0; mov.u16 %rs643, %rs644; @%p1656 bra $L__BB107_1877; mov.b32 %r841, %f2892; setp.nan.ftz.f32 %p1657, %f2892, %f2892; mov.u16 %rs643, 0; mov.u16 %rs642, %rs643; @%p1657 bra $L__BB107_1875; or.b32 %r3094, %r841, -2147483648; setp.ne.s32 %p1658, %r3094, -2147483648; add.ftz.f32 %f16305, %f2892, %f2892; setp.eq.ftz.f32 %p1659, %f16305, %f2892; and.pred %p1660, %p1659, %p1658; mov.u16 %rs642, %rs643; @%p1660 bra $L__BB107_1875; shl.b32 %r3095, %r841, 1; setp.lt.u32 %p1661, %r3095, -16777215; selp.u16 %rs642, 1, 0, %p1661; $L__BB107_1875: setp.eq.s16 %p1662, %rs642, 0; @%p1662 bra $L__BB107_1877; mul.ftz.f32 %f16306, %f2893, %f2893; fma.rn.ftz.f32 %f16307, %f2892, %f2892, %f16306; fma.rn.ftz.f32 %f16308, %f2894, %f2894, %f16307; setp.neu.ftz.f32 %p1663, %f16308, 0f00000000; selp.u16 %rs643, 1, 0, %p1663; $L__BB107_1877: setp.eq.s16 %p1664, %rs643, 0; @%p1664 bra $L__BB107_1881; @%p1969 bra $L__BB107_1880; bra.uni $L__BB107_1879; $L__BB107_1880: mov.u32 %r3098, 4; { // callseq 427, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3098; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3099, [retval0+0]; } // callseq 427 cvt.u16.u32 %rs644, %r3099; bra.uni $L__BB107_1881; $L__BB107_2033: mov.u32 %r3204, 4; { // callseq 444, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3204; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3205, [retval0+0]; } // callseq 444 cvt.u16.u32 %rs680, %r3205; $L__BB107_2035: setp.eq.s16 %p1837, %rs680, 0; mov.u16 %rs713, %rs349; @%p1837 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1838, %r5762, -1; mov.u32 %r5754, %r5762; @%p1838 bra $L__BB107_2038; ld.local.u32 %r3208, [%rd427+12]; mul.wide.s32 %rd2703, %r3208, 4; add.s64 %rd2704, %rd353, %rd2703; ld.global.u32 %r5754, [%rd2704]; $L__BB107_2038: setp.eq.s32 %p1839, %r5754, %r873; @%p1839 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3209, [%rd427+12]; st.local.u32 [%rd406+-12], %r3209; ld.local.f32 %f16381, [%rd427]; setp.eq.ftz.f32 %p1841, %f16381, 0f00000000; mov.b32 %r3210, %f16381; add.s32 %r3211, %r3210, 1; mov.b32 %f16382, %r3211; selp.f32 %f16383, 0f00000001, %f16382, %p1841; st.local.f32 [%rd406+-24], %f16383; mov.u16 %rs684, 0; mov.u16 %rs681, %rs684; @%p1784 bra $L__BB107_2042; or.b32 %r3212, %r827, -2147483648; setp.ne.s32 %p1842, %r3212, -2147483648; add.ftz.f32 %f16384, %f29709, %f29709; setp.eq.ftz.f32 %p1843, %f16384, %f29709; and.pred %p1844, %p1843, %p1842; mov.u16 %rs681, %rs684; @%p1844 bra $L__BB107_2042; shl.b32 %r3213, %r827, 1; setp.lt.u32 %p1845, %r3213, -16777215; selp.u16 %rs681, 1, 0, %p1845; $L__BB107_2042: setp.eq.s16 %p1846, %rs681, 0; mov.u16 %rs683, %rs684; @%p1846 bra $L__BB107_2048; mov.b32 %r882, %f2892; setp.nan.ftz.f32 %p1847, %f2892, %f2892; mov.u16 %rs683, 0; mov.u16 %rs682, %rs683; @%p1847 bra $L__BB107_2046; or.b32 %r3214, %r882, -2147483648; setp.ne.s32 %p1848, %r3214, -2147483648; add.ftz.f32 %f16385, %f2892, %f2892; setp.eq.ftz.f32 %p1849, %f16385, %f2892; and.pred %p1850, %p1849, %p1848; mov.u16 %rs682, %rs683; @%p1850 bra $L__BB107_2046; shl.b32 %r3215, %r882, 1; setp.lt.u32 %p1851, %r3215, -16777215; selp.u16 %rs682, 1, 0, %p1851; $L__BB107_2046: setp.eq.s16 %p1852, %rs682, 0; @%p1852 bra $L__BB107_2048; mul.ftz.f32 %f16386, %f2893, %f2893; fma.rn.ftz.f32 %f16387, %f2892, %f2892, %f16386; fma.rn.ftz.f32 %f16388, %f2894, %f2894, %f16387; setp.neu.ftz.f32 %p1853, %f16388, 0f00000000; selp.u16 %rs683, 1, 0, %p1853; $L__BB107_2048: setp.eq.s16 %p1854, %rs683, 0; @%p1854 bra $L__BB107_2052; @%p1969 bra $L__BB107_2051; bra.uni $L__BB107_2050; $L__BB107_2051: mov.u32 %r3218, 4; { // callseq 447, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3218; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3219, [retval0+0]; } // callseq 447 cvt.u16.u32 %rs684, %r3219; bra.uni $L__BB107_2052; $L__BB107_1879: mov.u32 %r3096, 4; { // callseq 426, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3096; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3097, [retval0+0]; } // callseq 426 cvt.u16.u32 %rs644, %r3097; $L__BB107_1881: setp.eq.s16 %p1666, %rs644, 0; mov.u16 %rs713, %rs349; @%p1666 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1667, %r5762, -1; mov.u32 %r5745, %r5762; @%p1667 bra $L__BB107_1884; ld.local.u32 %r3100, [%rd427+12]; mul.wide.s32 %rd2649, %r3100, 4; add.s64 %rd2650, %rd353, %rd2649; ld.global.u32 %r5745, [%rd2650]; $L__BB107_1884: setp.eq.s32 %p1668, %r5745, %r832; @%p1668 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3101, [%rd427+12]; st.local.u32 [%rd406+-12], %r3101; ld.local.f32 %f16309, [%rd427]; setp.eq.ftz.f32 %p1670, %f16309, 0f00000000; mov.b32 %r3102, %f16309; add.s32 %r3103, %r3102, 1; mov.b32 %f16310, %r3103; selp.f32 %f16311, 0f00000001, %f16310, %p1670; st.local.f32 [%rd406+-24], %f16311; mov.u16 %rs648, 0; mov.u16 %rs645, %rs648; @%p1594 bra $L__BB107_1888; or.b32 %r3104, %r827, -2147483648; setp.ne.s32 %p1671, %r3104, -2147483648; add.ftz.f32 %f16312, %f29709, %f29709; setp.eq.ftz.f32 %p1672, %f16312, %f29709; and.pred %p1673, %p1672, %p1671; mov.u16 %rs645, %rs648; @%p1673 bra $L__BB107_1888; shl.b32 %r3105, %r827, 1; setp.lt.u32 %p1674, %r3105, -16777215; selp.u16 %rs645, 1, 0, %p1674; $L__BB107_1888: setp.eq.s16 %p1675, %rs645, 0; mov.u16 %rs647, %rs648; @%p1675 bra $L__BB107_1894; mov.b32 %r845, %f2892; setp.nan.ftz.f32 %p1676, %f2892, %f2892; mov.u16 %rs647, 0; mov.u16 %rs646, %rs647; @%p1676 bra $L__BB107_1892; or.b32 %r3106, %r845, -2147483648; setp.ne.s32 %p1677, %r3106, -2147483648; add.ftz.f32 %f16313, %f2892, %f2892; setp.eq.ftz.f32 %p1678, %f16313, %f2892; and.pred %p1679, %p1678, %p1677; mov.u16 %rs646, %rs647; @%p1679 bra $L__BB107_1892; shl.b32 %r3107, %r845, 1; setp.lt.u32 %p1680, %r3107, -16777215; selp.u16 %rs646, 1, 0, %p1680; $L__BB107_1892: setp.eq.s16 %p1681, %rs646, 0; @%p1681 bra $L__BB107_1894; mul.ftz.f32 %f16314, %f2893, %f2893; fma.rn.ftz.f32 %f16315, %f2892, %f2892, %f16314; fma.rn.ftz.f32 %f16316, %f2894, %f2894, %f16315; setp.neu.ftz.f32 %p1682, %f16316, 0f00000000; selp.u16 %rs647, 1, 0, %p1682; $L__BB107_1894: setp.eq.s16 %p1683, %rs647, 0; @%p1683 bra $L__BB107_1898; @%p1969 bra $L__BB107_1897; bra.uni $L__BB107_1896; $L__BB107_1897: mov.u32 %r3110, 4; { // callseq 429, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3110; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3111, [retval0+0]; } // callseq 429 cvt.u16.u32 %rs648, %r3111; bra.uni $L__BB107_1898; $L__BB107_2050: mov.u32 %r3216, 4; { // callseq 446, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3216; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3217, [retval0+0]; } // callseq 446 cvt.u16.u32 %rs684, %r3217; $L__BB107_2052: setp.eq.s16 %p1856, %rs684, 0; mov.u16 %rs713, %rs349; @%p1856 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1857, %r5762, -1; mov.u32 %r5755, %r5762; @%p1857 bra $L__BB107_2055; ld.local.u32 %r3220, [%rd427+12]; mul.wide.s32 %rd2709, %r3220, 4; add.s64 %rd2710, %rd353, %rd2709; ld.global.u32 %r5755, [%rd2710]; $L__BB107_2055: setp.eq.s32 %p1858, %r5755, %r873; @%p1858 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3221, [%rd427+12]; st.local.u32 [%rd406+-12], %r3221; ld.local.f32 %f16389, [%rd427]; setp.eq.ftz.f32 %p1860, %f16389, 0f00000000; mov.b32 %r3222, %f16389; add.s32 %r3223, %r3222, 1; mov.b32 %f16390, %r3223; selp.f32 %f16391, 0f00000001, %f16390, %p1860; st.local.f32 [%rd406+-24], %f16391; mov.u16 %rs688, 0; mov.u16 %rs685, %rs688; @%p1784 bra $L__BB107_2059; or.b32 %r3224, %r827, -2147483648; setp.ne.s32 %p1861, %r3224, -2147483648; add.ftz.f32 %f16392, %f29709, %f29709; setp.eq.ftz.f32 %p1862, %f16392, %f29709; and.pred %p1863, %p1862, %p1861; mov.u16 %rs685, %rs688; @%p1863 bra $L__BB107_2059; shl.b32 %r3225, %r827, 1; setp.lt.u32 %p1864, %r3225, -16777215; selp.u16 %rs685, 1, 0, %p1864; $L__BB107_2059: setp.eq.s16 %p1865, %rs685, 0; mov.u16 %rs687, %rs688; @%p1865 bra $L__BB107_2065; mov.b32 %r886, %f2892; setp.nan.ftz.f32 %p1866, %f2892, %f2892; mov.u16 %rs687, 0; mov.u16 %rs686, %rs687; @%p1866 bra $L__BB107_2063; or.b32 %r3226, %r886, -2147483648; setp.ne.s32 %p1867, %r3226, -2147483648; add.ftz.f32 %f16393, %f2892, %f2892; setp.eq.ftz.f32 %p1868, %f16393, %f2892; and.pred %p1869, %p1868, %p1867; mov.u16 %rs686, %rs687; @%p1869 bra $L__BB107_2063; shl.b32 %r3227, %r886, 1; setp.lt.u32 %p1870, %r3227, -16777215; selp.u16 %rs686, 1, 0, %p1870; $L__BB107_2063: setp.eq.s16 %p1871, %rs686, 0; @%p1871 bra $L__BB107_2065; mul.ftz.f32 %f16394, %f2893, %f2893; fma.rn.ftz.f32 %f16395, %f2892, %f2892, %f16394; fma.rn.ftz.f32 %f16396, %f2894, %f2894, %f16395; setp.neu.ftz.f32 %p1872, %f16396, 0f00000000; selp.u16 %rs687, 1, 0, %p1872; $L__BB107_2065: setp.eq.s16 %p1873, %rs687, 0; @%p1873 bra $L__BB107_2069; @%p1969 bra $L__BB107_2068; bra.uni $L__BB107_2067; $L__BB107_2068: mov.u32 %r3230, 4; { // callseq 449, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3230; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3231, [retval0+0]; } // callseq 449 cvt.u16.u32 %rs688, %r3231; bra.uni $L__BB107_2069; $L__BB107_1896: mov.u32 %r3108, 4; { // callseq 428, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3108; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3109, [retval0+0]; } // callseq 428 cvt.u16.u32 %rs648, %r3109; $L__BB107_1898: setp.eq.s16 %p1685, %rs648, 0; mov.u16 %rs713, %rs349; @%p1685 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1686, %r5762, -1; mov.u32 %r5746, %r5762; @%p1686 bra $L__BB107_1901; ld.local.u32 %r3112, [%rd427+12]; mul.wide.s32 %rd2655, %r3112, 4; add.s64 %rd2656, %rd353, %rd2655; ld.global.u32 %r5746, [%rd2656]; $L__BB107_1901: setp.eq.s32 %p1687, %r5746, %r832; @%p1687 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3113, [%rd427+12]; st.local.u32 [%rd406+-12], %r3113; ld.local.f32 %f16317, [%rd427]; setp.eq.ftz.f32 %p1689, %f16317, 0f00000000; mov.b32 %r3114, %f16317; add.s32 %r3115, %r3114, 1; mov.b32 %f16318, %r3115; selp.f32 %f16319, 0f00000001, %f16318, %p1689; st.local.f32 [%rd406+-24], %f16319; mov.u16 %rs652, 0; mov.u16 %rs649, %rs652; @%p1594 bra $L__BB107_1905; or.b32 %r3116, %r827, -2147483648; setp.ne.s32 %p1690, %r3116, -2147483648; add.ftz.f32 %f16320, %f29709, %f29709; setp.eq.ftz.f32 %p1691, %f16320, %f29709; and.pred %p1692, %p1691, %p1690; mov.u16 %rs649, %rs652; @%p1692 bra $L__BB107_1905; shl.b32 %r3117, %r827, 1; setp.lt.u32 %p1693, %r3117, -16777215; selp.u16 %rs649, 1, 0, %p1693; $L__BB107_1905: setp.eq.s16 %p1694, %rs649, 0; mov.u16 %rs651, %rs652; @%p1694 bra $L__BB107_1911; mov.b32 %r849, %f2892; setp.nan.ftz.f32 %p1695, %f2892, %f2892; mov.u16 %rs651, 0; mov.u16 %rs650, %rs651; @%p1695 bra $L__BB107_1909; or.b32 %r3118, %r849, -2147483648; setp.ne.s32 %p1696, %r3118, -2147483648; add.ftz.f32 %f16321, %f2892, %f2892; setp.eq.ftz.f32 %p1697, %f16321, %f2892; and.pred %p1698, %p1697, %p1696; mov.u16 %rs650, %rs651; @%p1698 bra $L__BB107_1909; shl.b32 %r3119, %r849, 1; setp.lt.u32 %p1699, %r3119, -16777215; selp.u16 %rs650, 1, 0, %p1699; $L__BB107_1909: setp.eq.s16 %p1700, %rs650, 0; @%p1700 bra $L__BB107_1911; mul.ftz.f32 %f16322, %f2893, %f2893; fma.rn.ftz.f32 %f16323, %f2892, %f2892, %f16322; fma.rn.ftz.f32 %f16324, %f2894, %f2894, %f16323; setp.neu.ftz.f32 %p1701, %f16324, 0f00000000; selp.u16 %rs651, 1, 0, %p1701; $L__BB107_1911: setp.eq.s16 %p1702, %rs651, 0; @%p1702 bra $L__BB107_1915; @%p1969 bra $L__BB107_1914; bra.uni $L__BB107_1913; $L__BB107_1914: mov.u32 %r3122, 4; { // callseq 431, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3122; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3123, [retval0+0]; } // callseq 431 cvt.u16.u32 %rs652, %r3123; bra.uni $L__BB107_1915; $L__BB107_2067: mov.u32 %r3228, 4; { // callseq 448, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3228; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 448 cvt.u16.u32 %rs688, %r3229; $L__BB107_2069: setp.eq.s16 %p1875, %rs688, 0; mov.u16 %rs713, %rs349; @%p1875 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1876, %r5762, -1; mov.u32 %r5756, %r5762; @%p1876 bra $L__BB107_2072; ld.local.u32 %r3232, [%rd427+12]; mul.wide.s32 %rd2715, %r3232, 4; add.s64 %rd2716, %rd353, %rd2715; ld.global.u32 %r5756, [%rd2716]; $L__BB107_2072: setp.eq.s32 %p1877, %r5756, %r873; @%p1877 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3233, [%rd427+12]; st.local.u32 [%rd406+-12], %r3233; ld.local.f32 %f16397, [%rd427]; setp.eq.ftz.f32 %p1879, %f16397, 0f00000000; mov.b32 %r3234, %f16397; add.s32 %r3235, %r3234, 1; mov.b32 %f16398, %r3235; selp.f32 %f16399, 0f00000001, %f16398, %p1879; st.local.f32 [%rd406+-24], %f16399; mov.u16 %rs692, 0; mov.u16 %rs689, %rs692; @%p1784 bra $L__BB107_2076; or.b32 %r3236, %r827, -2147483648; setp.ne.s32 %p1880, %r3236, -2147483648; add.ftz.f32 %f16400, %f29709, %f29709; setp.eq.ftz.f32 %p1881, %f16400, %f29709; and.pred %p1882, %p1881, %p1880; mov.u16 %rs689, %rs692; @%p1882 bra $L__BB107_2076; shl.b32 %r3237, %r827, 1; setp.lt.u32 %p1883, %r3237, -16777215; selp.u16 %rs689, 1, 0, %p1883; $L__BB107_2076: setp.eq.s16 %p1884, %rs689, 0; mov.u16 %rs691, %rs692; @%p1884 bra $L__BB107_2082; mov.b32 %r890, %f2892; setp.nan.ftz.f32 %p1885, %f2892, %f2892; mov.u16 %rs691, 0; mov.u16 %rs690, %rs691; @%p1885 bra $L__BB107_2080; or.b32 %r3238, %r890, -2147483648; setp.ne.s32 %p1886, %r3238, -2147483648; add.ftz.f32 %f16401, %f2892, %f2892; setp.eq.ftz.f32 %p1887, %f16401, %f2892; and.pred %p1888, %p1887, %p1886; mov.u16 %rs690, %rs691; @%p1888 bra $L__BB107_2080; shl.b32 %r3239, %r890, 1; setp.lt.u32 %p1889, %r3239, -16777215; selp.u16 %rs690, 1, 0, %p1889; $L__BB107_2080: setp.eq.s16 %p1890, %rs690, 0; @%p1890 bra $L__BB107_2082; mul.ftz.f32 %f16402, %f2893, %f2893; fma.rn.ftz.f32 %f16403, %f2892, %f2892, %f16402; fma.rn.ftz.f32 %f16404, %f2894, %f2894, %f16403; setp.neu.ftz.f32 %p1891, %f16404, 0f00000000; selp.u16 %rs691, 1, 0, %p1891; $L__BB107_2082: setp.eq.s16 %p1892, %rs691, 0; @%p1892 bra $L__BB107_2086; @%p1969 bra $L__BB107_2085; bra.uni $L__BB107_2084; $L__BB107_2085: mov.u32 %r3242, 4; { // callseq 451, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3242; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3243, [retval0+0]; } // callseq 451 cvt.u16.u32 %rs692, %r3243; bra.uni $L__BB107_2086; $L__BB107_1913: mov.u32 %r3120, 4; { // callseq 430, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3120; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3121, [retval0+0]; } // callseq 430 cvt.u16.u32 %rs652, %r3121; $L__BB107_1915: setp.eq.s16 %p1704, %rs652, 0; mov.u16 %rs713, %rs349; @%p1704 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1705, %r5762, -1; mov.u32 %r5747, %r5762; @%p1705 bra $L__BB107_1918; ld.local.u32 %r3124, [%rd427+12]; mul.wide.s32 %rd2661, %r3124, 4; add.s64 %rd2662, %rd353, %rd2661; ld.global.u32 %r5747, [%rd2662]; $L__BB107_1918: setp.eq.s32 %p1706, %r5747, %r832; @%p1706 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3125, [%rd427+12]; st.local.u32 [%rd406+-12], %r3125; ld.local.f32 %f16325, [%rd427]; setp.eq.ftz.f32 %p1708, %f16325, 0f00000000; mov.b32 %r3126, %f16325; add.s32 %r3127, %r3126, 1; mov.b32 %f16326, %r3127; selp.f32 %f16327, 0f00000001, %f16326, %p1708; st.local.f32 [%rd406+-24], %f16327; mov.u16 %rs656, 0; mov.u16 %rs653, %rs656; @%p1594 bra $L__BB107_1922; or.b32 %r3128, %r827, -2147483648; setp.ne.s32 %p1709, %r3128, -2147483648; add.ftz.f32 %f16328, %f29709, %f29709; setp.eq.ftz.f32 %p1710, %f16328, %f29709; and.pred %p1711, %p1710, %p1709; mov.u16 %rs653, %rs656; @%p1711 bra $L__BB107_1922; shl.b32 %r3129, %r827, 1; setp.lt.u32 %p1712, %r3129, -16777215; selp.u16 %rs653, 1, 0, %p1712; $L__BB107_1922: setp.eq.s16 %p1713, %rs653, 0; mov.u16 %rs655, %rs656; @%p1713 bra $L__BB107_1928; mov.b32 %r853, %f2892; setp.nan.ftz.f32 %p1714, %f2892, %f2892; mov.u16 %rs655, 0; mov.u16 %rs654, %rs655; @%p1714 bra $L__BB107_1926; or.b32 %r3130, %r853, -2147483648; setp.ne.s32 %p1715, %r3130, -2147483648; add.ftz.f32 %f16329, %f2892, %f2892; setp.eq.ftz.f32 %p1716, %f16329, %f2892; and.pred %p1717, %p1716, %p1715; mov.u16 %rs654, %rs655; @%p1717 bra $L__BB107_1926; shl.b32 %r3131, %r853, 1; setp.lt.u32 %p1718, %r3131, -16777215; selp.u16 %rs654, 1, 0, %p1718; $L__BB107_1926: setp.eq.s16 %p1719, %rs654, 0; @%p1719 bra $L__BB107_1928; mul.ftz.f32 %f16330, %f2893, %f2893; fma.rn.ftz.f32 %f16331, %f2892, %f2892, %f16330; fma.rn.ftz.f32 %f16332, %f2894, %f2894, %f16331; setp.neu.ftz.f32 %p1720, %f16332, 0f00000000; selp.u16 %rs655, 1, 0, %p1720; $L__BB107_1928: setp.eq.s16 %p1721, %rs655, 0; @%p1721 bra $L__BB107_1932; @%p1969 bra $L__BB107_1931; bra.uni $L__BB107_1930; $L__BB107_1931: mov.u32 %r3134, 4; { // callseq 433, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3134; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3135, [retval0+0]; } // callseq 433 cvt.u16.u32 %rs656, %r3135; bra.uni $L__BB107_1932; $L__BB107_2084: mov.u32 %r3240, 4; { // callseq 450, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3240; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3241, [retval0+0]; } // callseq 450 cvt.u16.u32 %rs692, %r3241; $L__BB107_2086: setp.eq.s16 %p1894, %rs692, 0; mov.u16 %rs713, %rs349; @%p1894 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1895, %r5762, -1; mov.u32 %r5757, %r5762; @%p1895 bra $L__BB107_2089; ld.local.u32 %r3244, [%rd427+12]; mul.wide.s32 %rd2721, %r3244, 4; add.s64 %rd2722, %rd353, %rd2721; ld.global.u32 %r5757, [%rd2722]; $L__BB107_2089: setp.eq.s32 %p1896, %r5757, %r873; @%p1896 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3245, [%rd427+12]; st.local.u32 [%rd406+-12], %r3245; ld.local.f32 %f16405, [%rd427]; setp.eq.ftz.f32 %p1898, %f16405, 0f00000000; mov.b32 %r3246, %f16405; add.s32 %r3247, %r3246, 1; mov.b32 %f16406, %r3247; selp.f32 %f16407, 0f00000001, %f16406, %p1898; st.local.f32 [%rd406+-24], %f16407; mov.u16 %rs696, 0; mov.u16 %rs693, %rs696; @%p1784 bra $L__BB107_2093; or.b32 %r3248, %r827, -2147483648; setp.ne.s32 %p1899, %r3248, -2147483648; add.ftz.f32 %f16408, %f29709, %f29709; setp.eq.ftz.f32 %p1900, %f16408, %f29709; and.pred %p1901, %p1900, %p1899; mov.u16 %rs693, %rs696; @%p1901 bra $L__BB107_2093; shl.b32 %r3249, %r827, 1; setp.lt.u32 %p1902, %r3249, -16777215; selp.u16 %rs693, 1, 0, %p1902; $L__BB107_2093: setp.eq.s16 %p1903, %rs693, 0; mov.u16 %rs695, %rs696; @%p1903 bra $L__BB107_2099; mov.b32 %r894, %f2892; setp.nan.ftz.f32 %p1904, %f2892, %f2892; mov.u16 %rs695, 0; mov.u16 %rs694, %rs695; @%p1904 bra $L__BB107_2097; or.b32 %r3250, %r894, -2147483648; setp.ne.s32 %p1905, %r3250, -2147483648; add.ftz.f32 %f16409, %f2892, %f2892; setp.eq.ftz.f32 %p1906, %f16409, %f2892; and.pred %p1907, %p1906, %p1905; mov.u16 %rs694, %rs695; @%p1907 bra $L__BB107_2097; shl.b32 %r3251, %r894, 1; setp.lt.u32 %p1908, %r3251, -16777215; selp.u16 %rs694, 1, 0, %p1908; $L__BB107_2097: setp.eq.s16 %p1909, %rs694, 0; @%p1909 bra $L__BB107_2099; mul.ftz.f32 %f16410, %f2893, %f2893; fma.rn.ftz.f32 %f16411, %f2892, %f2892, %f16410; fma.rn.ftz.f32 %f16412, %f2894, %f2894, %f16411; setp.neu.ftz.f32 %p1910, %f16412, 0f00000000; selp.u16 %rs695, 1, 0, %p1910; $L__BB107_2099: setp.eq.s16 %p1911, %rs695, 0; @%p1911 bra $L__BB107_2103; @%p1969 bra $L__BB107_2102; bra.uni $L__BB107_2101; $L__BB107_2102: mov.u32 %r3254, 4; { // callseq 453, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3254; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3255, [retval0+0]; } // callseq 453 cvt.u16.u32 %rs696, %r3255; bra.uni $L__BB107_2103; $L__BB107_1930: mov.u32 %r3132, 4; { // callseq 432, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3132; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3133, [retval0+0]; } // callseq 432 cvt.u16.u32 %rs656, %r3133; $L__BB107_1932: setp.eq.s16 %p1723, %rs656, 0; mov.u16 %rs713, %rs349; @%p1723 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1724, %r5762, -1; mov.u32 %r5748, %r5762; @%p1724 bra $L__BB107_1935; ld.local.u32 %r3136, [%rd427+12]; mul.wide.s32 %rd2667, %r3136, 4; add.s64 %rd2668, %rd353, %rd2667; ld.global.u32 %r5748, [%rd2668]; $L__BB107_1935: setp.eq.s32 %p1725, %r5748, %r832; @%p1725 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3137, [%rd427+12]; st.local.u32 [%rd406+-12], %r3137; ld.local.f32 %f16333, [%rd427]; setp.eq.ftz.f32 %p1727, %f16333, 0f00000000; mov.b32 %r3138, %f16333; add.s32 %r3139, %r3138, 1; mov.b32 %f16334, %r3139; selp.f32 %f16335, 0f00000001, %f16334, %p1727; st.local.f32 [%rd406+-24], %f16335; mov.u16 %rs660, 0; mov.u16 %rs657, %rs660; @%p1594 bra $L__BB107_1939; or.b32 %r3140, %r827, -2147483648; setp.ne.s32 %p1728, %r3140, -2147483648; add.ftz.f32 %f16336, %f29709, %f29709; setp.eq.ftz.f32 %p1729, %f16336, %f29709; and.pred %p1730, %p1729, %p1728; mov.u16 %rs657, %rs660; @%p1730 bra $L__BB107_1939; shl.b32 %r3141, %r827, 1; setp.lt.u32 %p1731, %r3141, -16777215; selp.u16 %rs657, 1, 0, %p1731; $L__BB107_1939: setp.eq.s16 %p1732, %rs657, 0; mov.u16 %rs659, %rs660; @%p1732 bra $L__BB107_1945; mov.b32 %r857, %f2892; setp.nan.ftz.f32 %p1733, %f2892, %f2892; mov.u16 %rs659, 0; mov.u16 %rs658, %rs659; @%p1733 bra $L__BB107_1943; or.b32 %r3142, %r857, -2147483648; setp.ne.s32 %p1734, %r3142, -2147483648; add.ftz.f32 %f16337, %f2892, %f2892; setp.eq.ftz.f32 %p1735, %f16337, %f2892; and.pred %p1736, %p1735, %p1734; mov.u16 %rs658, %rs659; @%p1736 bra $L__BB107_1943; shl.b32 %r3143, %r857, 1; setp.lt.u32 %p1737, %r3143, -16777215; selp.u16 %rs658, 1, 0, %p1737; $L__BB107_1943: setp.eq.s16 %p1738, %rs658, 0; @%p1738 bra $L__BB107_1945; mul.ftz.f32 %f16338, %f2893, %f2893; fma.rn.ftz.f32 %f16339, %f2892, %f2892, %f16338; fma.rn.ftz.f32 %f16340, %f2894, %f2894, %f16339; setp.neu.ftz.f32 %p1739, %f16340, 0f00000000; selp.u16 %rs659, 1, 0, %p1739; $L__BB107_1945: setp.eq.s16 %p1740, %rs659, 0; @%p1740 bra $L__BB107_1949; @%p1969 bra $L__BB107_1948; bra.uni $L__BB107_1947; $L__BB107_1948: mov.u32 %r3146, 4; { // callseq 435, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3146; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3147, [retval0+0]; } // callseq 435 cvt.u16.u32 %rs660, %r3147; bra.uni $L__BB107_1949; $L__BB107_2101: mov.u32 %r3252, 4; { // callseq 452, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3252; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3253, [retval0+0]; } // callseq 452 cvt.u16.u32 %rs696, %r3253; $L__BB107_2103: setp.eq.s16 %p1913, %rs696, 0; mov.u16 %rs713, %rs349; @%p1913 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1914, %r5762, -1; mov.u32 %r5758, %r5762; @%p1914 bra $L__BB107_2106; ld.local.u32 %r3256, [%rd427+12]; mul.wide.s32 %rd2727, %r3256, 4; add.s64 %rd2728, %rd353, %rd2727; ld.global.u32 %r5758, [%rd2728]; $L__BB107_2106: setp.eq.s32 %p1915, %r5758, %r873; @%p1915 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3257, [%rd427+12]; st.local.u32 [%rd406+-12], %r3257; ld.local.f32 %f16413, [%rd427]; setp.eq.ftz.f32 %p1917, %f16413, 0f00000000; mov.b32 %r3258, %f16413; add.s32 %r3259, %r3258, 1; mov.b32 %f16414, %r3259; selp.f32 %f16415, 0f00000001, %f16414, %p1917; st.local.f32 [%rd406+-24], %f16415; mov.u16 %rs700, 0; mov.u16 %rs697, %rs700; @%p1784 bra $L__BB107_2110; or.b32 %r3260, %r827, -2147483648; setp.ne.s32 %p1918, %r3260, -2147483648; add.ftz.f32 %f16416, %f29709, %f29709; setp.eq.ftz.f32 %p1919, %f16416, %f29709; and.pred %p1920, %p1919, %p1918; mov.u16 %rs697, %rs700; @%p1920 bra $L__BB107_2110; shl.b32 %r3261, %r827, 1; setp.lt.u32 %p1921, %r3261, -16777215; selp.u16 %rs697, 1, 0, %p1921; $L__BB107_2110: setp.eq.s16 %p1922, %rs697, 0; mov.u16 %rs699, %rs700; @%p1922 bra $L__BB107_2116; mov.b32 %r898, %f2892; setp.nan.ftz.f32 %p1923, %f2892, %f2892; mov.u16 %rs699, 0; mov.u16 %rs698, %rs699; @%p1923 bra $L__BB107_2114; or.b32 %r3262, %r898, -2147483648; setp.ne.s32 %p1924, %r3262, -2147483648; add.ftz.f32 %f16417, %f2892, %f2892; setp.eq.ftz.f32 %p1925, %f16417, %f2892; and.pred %p1926, %p1925, %p1924; mov.u16 %rs698, %rs699; @%p1926 bra $L__BB107_2114; shl.b32 %r3263, %r898, 1; setp.lt.u32 %p1927, %r3263, -16777215; selp.u16 %rs698, 1, 0, %p1927; $L__BB107_2114: setp.eq.s16 %p1928, %rs698, 0; @%p1928 bra $L__BB107_2116; mul.ftz.f32 %f16418, %f2893, %f2893; fma.rn.ftz.f32 %f16419, %f2892, %f2892, %f16418; fma.rn.ftz.f32 %f16420, %f2894, %f2894, %f16419; setp.neu.ftz.f32 %p1929, %f16420, 0f00000000; selp.u16 %rs699, 1, 0, %p1929; $L__BB107_2116: setp.eq.s16 %p1930, %rs699, 0; @%p1930 bra $L__BB107_2120; @%p1969 bra $L__BB107_2119; bra.uni $L__BB107_2118; $L__BB107_2119: mov.u32 %r3266, 4; { // callseq 455, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3266; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3267, [retval0+0]; } // callseq 455 cvt.u16.u32 %rs700, %r3267; bra.uni $L__BB107_2120; $L__BB107_1947: mov.u32 %r3144, 4; { // callseq 434, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3144; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3145, [retval0+0]; } // callseq 434 cvt.u16.u32 %rs660, %r3145; $L__BB107_1949: setp.eq.s16 %p1742, %rs660, 0; mov.u16 %rs713, %rs349; @%p1742 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1743, %r5762, -1; mov.u32 %r5749, %r5762; @%p1743 bra $L__BB107_1952; ld.local.u32 %r3148, [%rd427+12]; mul.wide.s32 %rd2673, %r3148, 4; add.s64 %rd2674, %rd353, %rd2673; ld.global.u32 %r5749, [%rd2674]; $L__BB107_1952: setp.eq.s32 %p1744, %r5749, %r832; @%p1744 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3149, [%rd427+12]; st.local.u32 [%rd406+-12], %r3149; ld.local.f32 %f16341, [%rd427]; setp.eq.ftz.f32 %p1746, %f16341, 0f00000000; mov.b32 %r3150, %f16341; add.s32 %r3151, %r3150, 1; mov.b32 %f16342, %r3151; selp.f32 %f16343, 0f00000001, %f16342, %p1746; st.local.f32 [%rd406+-24], %f16343; mov.u16 %rs664, 0; mov.u16 %rs661, %rs664; @%p1594 bra $L__BB107_1956; or.b32 %r3152, %r827, -2147483648; setp.ne.s32 %p1747, %r3152, -2147483648; add.ftz.f32 %f16344, %f29709, %f29709; setp.eq.ftz.f32 %p1748, %f16344, %f29709; and.pred %p1749, %p1748, %p1747; mov.u16 %rs661, %rs664; @%p1749 bra $L__BB107_1956; shl.b32 %r3153, %r827, 1; setp.lt.u32 %p1750, %r3153, -16777215; selp.u16 %rs661, 1, 0, %p1750; $L__BB107_1956: setp.eq.s16 %p1751, %rs661, 0; mov.u16 %rs663, %rs664; @%p1751 bra $L__BB107_1962; mov.b32 %r861, %f2892; setp.nan.ftz.f32 %p1752, %f2892, %f2892; mov.u16 %rs663, 0; mov.u16 %rs662, %rs663; @%p1752 bra $L__BB107_1960; or.b32 %r3154, %r861, -2147483648; setp.ne.s32 %p1753, %r3154, -2147483648; add.ftz.f32 %f16345, %f2892, %f2892; setp.eq.ftz.f32 %p1754, %f16345, %f2892; and.pred %p1755, %p1754, %p1753; mov.u16 %rs662, %rs663; @%p1755 bra $L__BB107_1960; shl.b32 %r3155, %r861, 1; setp.lt.u32 %p1756, %r3155, -16777215; selp.u16 %rs662, 1, 0, %p1756; $L__BB107_1960: setp.eq.s16 %p1757, %rs662, 0; @%p1757 bra $L__BB107_1962; mul.ftz.f32 %f16346, %f2893, %f2893; fma.rn.ftz.f32 %f16347, %f2892, %f2892, %f16346; fma.rn.ftz.f32 %f16348, %f2894, %f2894, %f16347; setp.neu.ftz.f32 %p1758, %f16348, 0f00000000; selp.u16 %rs663, 1, 0, %p1758; $L__BB107_1962: setp.eq.s16 %p1759, %rs663, 0; @%p1759 bra $L__BB107_1966; @%p1969 bra $L__BB107_1965; bra.uni $L__BB107_1964; $L__BB107_1965: mov.u32 %r3158, 4; { // callseq 437, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3158; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3159, [retval0+0]; } // callseq 437 cvt.u16.u32 %rs664, %r3159; bra.uni $L__BB107_1966; $L__BB107_2118: mov.u32 %r3264, 4; { // callseq 454, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3264; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3265, [retval0+0]; } // callseq 454 cvt.u16.u32 %rs700, %r3265; $L__BB107_2120: setp.eq.s16 %p1932, %rs700, 0; mov.u16 %rs713, %rs349; @%p1932 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1933, %r5762, -1; mov.u32 %r5759, %r5762; @%p1933 bra $L__BB107_2123; ld.local.u32 %r3268, [%rd427+12]; mul.wide.s32 %rd2733, %r3268, 4; add.s64 %rd2734, %rd353, %rd2733; ld.global.u32 %r5759, [%rd2734]; $L__BB107_2123: setp.eq.s32 %p1934, %r5759, %r873; @%p1934 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3269, [%rd427+12]; st.local.u32 [%rd406+-12], %r3269; ld.local.f32 %f16421, [%rd427]; setp.eq.ftz.f32 %p1936, %f16421, 0f00000000; mov.b32 %r3270, %f16421; add.s32 %r3271, %r3270, 1; mov.b32 %f16422, %r3271; selp.f32 %f16423, 0f00000001, %f16422, %p1936; st.local.f32 [%rd406+-24], %f16423; mov.u16 %rs704, 0; mov.u16 %rs701, %rs704; @%p1784 bra $L__BB107_2127; or.b32 %r3272, %r827, -2147483648; setp.ne.s32 %p1937, %r3272, -2147483648; add.ftz.f32 %f16424, %f29709, %f29709; setp.eq.ftz.f32 %p1938, %f16424, %f29709; and.pred %p1939, %p1938, %p1937; mov.u16 %rs701, %rs704; @%p1939 bra $L__BB107_2127; shl.b32 %r3273, %r827, 1; setp.lt.u32 %p1940, %r3273, -16777215; selp.u16 %rs701, 1, 0, %p1940; $L__BB107_2127: setp.eq.s16 %p1941, %rs701, 0; mov.u16 %rs703, %rs704; @%p1941 bra $L__BB107_2133; mov.b32 %r902, %f2892; setp.nan.ftz.f32 %p1942, %f2892, %f2892; mov.u16 %rs703, 0; mov.u16 %rs702, %rs703; @%p1942 bra $L__BB107_2131; or.b32 %r3274, %r902, -2147483648; setp.ne.s32 %p1943, %r3274, -2147483648; add.ftz.f32 %f16425, %f2892, %f2892; setp.eq.ftz.f32 %p1944, %f16425, %f2892; and.pred %p1945, %p1944, %p1943; mov.u16 %rs702, %rs703; @%p1945 bra $L__BB107_2131; shl.b32 %r3275, %r902, 1; setp.lt.u32 %p1946, %r3275, -16777215; selp.u16 %rs702, 1, 0, %p1946; $L__BB107_2131: setp.eq.s16 %p1947, %rs702, 0; @%p1947 bra $L__BB107_2133; mul.ftz.f32 %f16426, %f2893, %f2893; fma.rn.ftz.f32 %f16427, %f2892, %f2892, %f16426; fma.rn.ftz.f32 %f16428, %f2894, %f2894, %f16427; setp.neu.ftz.f32 %p1948, %f16428, 0f00000000; selp.u16 %rs703, 1, 0, %p1948; $L__BB107_2133: setp.eq.s16 %p1949, %rs703, 0; @%p1949 bra $L__BB107_2137; @%p1969 bra $L__BB107_2136; bra.uni $L__BB107_2135; $L__BB107_2136: mov.u32 %r3278, 4; { // callseq 457, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3278; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3279, [retval0+0]; } // callseq 457 cvt.u16.u32 %rs704, %r3279; bra.uni $L__BB107_2137; $L__BB107_1964: mov.u32 %r3156, 4; { // callseq 436, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3156; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3157, [retval0+0]; } // callseq 436 cvt.u16.u32 %rs664, %r3157; $L__BB107_1966: setp.eq.s16 %p1761, %rs664, 0; mov.u16 %rs713, %rs349; @%p1761 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1762, %r5762, -1; mov.u32 %r5750, %r5762; @%p1762 bra $L__BB107_1969; ld.local.u32 %r3160, [%rd427+12]; mul.wide.s32 %rd2679, %r3160, 4; add.s64 %rd2680, %rd353, %rd2679; ld.global.u32 %r5750, [%rd2680]; $L__BB107_1969: setp.eq.s32 %p1763, %r5750, %r832; @%p1763 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3161, [%rd427+12]; st.local.u32 [%rd406+-12], %r3161; ld.local.f32 %f16349, [%rd427]; setp.eq.ftz.f32 %p1765, %f16349, 0f00000000; mov.b32 %r3162, %f16349; add.s32 %r3163, %r3162, 1; mov.b32 %f16350, %r3163; selp.f32 %f16351, 0f00000001, %f16350, %p1765; st.local.f32 [%rd406+-24], %f16351; mov.u16 %rs668, 0; mov.u16 %rs665, %rs668; @%p1594 bra $L__BB107_1973; or.b32 %r3164, %r827, -2147483648; setp.ne.s32 %p1766, %r3164, -2147483648; add.ftz.f32 %f16352, %f29709, %f29709; setp.eq.ftz.f32 %p1767, %f16352, %f29709; and.pred %p1768, %p1767, %p1766; mov.u16 %rs665, %rs668; @%p1768 bra $L__BB107_1973; shl.b32 %r3165, %r827, 1; setp.lt.u32 %p1769, %r3165, -16777215; selp.u16 %rs665, 1, 0, %p1769; $L__BB107_1973: setp.eq.s16 %p1770, %rs665, 0; mov.u16 %rs667, %rs668; @%p1770 bra $L__BB107_1979; mov.b32 %r865, %f2892; setp.nan.ftz.f32 %p1771, %f2892, %f2892; mov.u16 %rs667, 0; mov.u16 %rs666, %rs667; @%p1771 bra $L__BB107_1977; or.b32 %r3166, %r865, -2147483648; setp.ne.s32 %p1772, %r3166, -2147483648; add.ftz.f32 %f16353, %f2892, %f2892; setp.eq.ftz.f32 %p1773, %f16353, %f2892; and.pred %p1774, %p1773, %p1772; mov.u16 %rs666, %rs667; @%p1774 bra $L__BB107_1977; shl.b32 %r3167, %r865, 1; setp.lt.u32 %p1775, %r3167, -16777215; selp.u16 %rs666, 1, 0, %p1775; $L__BB107_1977: setp.eq.s16 %p1776, %rs666, 0; @%p1776 bra $L__BB107_1979; mul.ftz.f32 %f16354, %f2893, %f2893; fma.rn.ftz.f32 %f16355, %f2892, %f2892, %f16354; fma.rn.ftz.f32 %f16356, %f2894, %f2894, %f16355; setp.neu.ftz.f32 %p1777, %f16356, 0f00000000; selp.u16 %rs667, 1, 0, %p1777; $L__BB107_1979: setp.eq.s16 %p1778, %rs667, 0; @%p1778 bra $L__BB107_1983; @%p1969 bra $L__BB107_1982; bra.uni $L__BB107_1981; $L__BB107_1982: mov.u32 %r3170, 4; { // callseq 439, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3170; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3171, [retval0+0]; } // callseq 439 cvt.u16.u32 %rs668, %r3171; bra.uni $L__BB107_1983; $L__BB107_2135: mov.u32 %r3276, 4; { // callseq 456, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3276; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3277, [retval0+0]; } // callseq 456 cvt.u16.u32 %rs704, %r3277; $L__BB107_2137: setp.eq.s16 %p1951, %rs704, 0; mov.u16 %rs713, %rs349; @%p1951 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1952, %r5762, -1; mov.u32 %r5760, %r5762; @%p1952 bra $L__BB107_2140; ld.local.u32 %r3280, [%rd427+12]; mul.wide.s32 %rd2739, %r3280, 4; add.s64 %rd2740, %rd353, %rd2739; ld.global.u32 %r5760, [%rd2740]; $L__BB107_2140: setp.eq.s32 %p1953, %r5760, %r873; @%p1953 bra $L__BB107_2159; st.local.u32 [%rd406+-8], %r5762; ld.local.u32 %r3281, [%rd427+12]; st.local.u32 [%rd406+-12], %r3281; ld.local.f32 %f16429, [%rd427]; setp.eq.ftz.f32 %p1955, %f16429, 0f00000000; mov.b32 %r3282, %f16429; add.s32 %r3283, %r3282, 1; mov.b32 %f16430, %r3283; selp.f32 %f16431, 0f00000001, %f16430, %p1955; st.local.f32 [%rd406+-24], %f16431; mov.u16 %rs708, 0; mov.u16 %rs705, %rs708; @%p1784 bra $L__BB107_2144; or.b32 %r3284, %r827, -2147483648; setp.ne.s32 %p1956, %r3284, -2147483648; add.ftz.f32 %f16432, %f29709, %f29709; setp.eq.ftz.f32 %p1957, %f16432, %f29709; and.pred %p1958, %p1957, %p1956; mov.u16 %rs705, %rs708; @%p1958 bra $L__BB107_2144; shl.b32 %r3285, %r827, 1; setp.lt.u32 %p1959, %r3285, -16777215; selp.u16 %rs705, 1, 0, %p1959; $L__BB107_2144: setp.eq.s16 %p1960, %rs705, 0; mov.u16 %rs707, %rs708; @%p1960 bra $L__BB107_2150; mov.b32 %r906, %f2892; setp.nan.ftz.f32 %p1961, %f2892, %f2892; mov.u16 %rs707, 0; mov.u16 %rs706, %rs707; @%p1961 bra $L__BB107_2148; or.b32 %r3286, %r906, -2147483648; setp.ne.s32 %p1962, %r3286, -2147483648; add.ftz.f32 %f16433, %f2892, %f2892; setp.eq.ftz.f32 %p1963, %f16433, %f2892; and.pred %p1964, %p1963, %p1962; mov.u16 %rs706, %rs707; @%p1964 bra $L__BB107_2148; shl.b32 %r3287, %r906, 1; setp.lt.u32 %p1965, %r3287, -16777215; selp.u16 %rs706, 1, 0, %p1965; $L__BB107_2148: setp.eq.s16 %p1966, %rs706, 0; @%p1966 bra $L__BB107_2150; mul.ftz.f32 %f16434, %f2893, %f2893; fma.rn.ftz.f32 %f16435, %f2892, %f2892, %f16434; fma.rn.ftz.f32 %f16436, %f2894, %f2894, %f16435; setp.neu.ftz.f32 %p1967, %f16436, 0f00000000; selp.u16 %rs707, 1, 0, %p1967; $L__BB107_2150: setp.eq.s16 %p1968, %rs707, 0; @%p1968 bra $L__BB107_2154; @%p1969 bra $L__BB107_2153; bra.uni $L__BB107_2152; $L__BB107_2153: mov.u32 %r3290, 4; { // callseq 459, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3290; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3291, [retval0+0]; } // callseq 459 cvt.u16.u32 %rs708, %r3291; bra.uni $L__BB107_2154; $L__BB107_1981: mov.u32 %r3168, 4; { // callseq 438, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3168; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3169, [retval0+0]; } // callseq 438 cvt.u16.u32 %rs668, %r3169; $L__BB107_1983: setp.eq.s16 %p1780, %rs668, 0; mov.u16 %rs713, %rs349; @%p1780 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1781, %r5762, -1; mov.u32 %r5751, %r5762; @%p1781 bra $L__BB107_1986; ld.local.u32 %r3172, [%rd427+12]; mul.wide.s32 %rd2685, %r3172, 4; add.s64 %rd2686, %rd353, %rd2685; ld.global.u32 %r5751, [%rd2686]; $L__BB107_1986: setp.eq.s32 %p1782, %r5751, %r832; @%p1782 bra $L__BB107_2159; bra.uni $L__BB107_1987; $L__BB107_2152: mov.u32 %r3288, 4; { // callseq 458, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2582; .param .b64 param1; st.param.b64 [param1+0], %rd2619; .param .b32 param2; st.param.b32 [param2+0], %r3288; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r3289, [retval0+0]; } // callseq 458 cvt.u16.u32 %rs708, %r3289; $L__BB107_2154: setp.eq.s16 %p1970, %rs708, 0; mov.u16 %rs713, %rs349; @%p1970 bra $L__BB107_2228; ld.local.u32 %r5762, [%rd427+16]; setp.ne.s32 %p1971, %r5762, -1; mov.u32 %r5761, %r5762; @%p1971 bra $L__BB107_2157; ld.local.u32 %r3292, [%rd427+12]; mul.wide.s32 %rd2745, %r3292, 4; add.s64 %rd2746, %rd353, %rd2745; ld.global.u32 %r5761, [%rd2746]; $L__BB107_2157: setp.eq.s32 %p1972, %r5761, %r873; @%p1972 bra $L__BB107_2159; bra.uni $L__BB107_2158; $L__BB107_2159: mul.lo.s64 %rd2749, %rd428, 192; add.s64 %rd433, %rd2157, %rd2749; st.local.v4.f32 [%rd433], {%f16218, %f16219, %f16220, %f16221}; ld.local.v2.f32 {%f16440, %f16441}, [%rd429+24]; st.local.v4.f32 [%rd433+16], {%f16222, %f16223, %f16440, %f16441}; st.local.f32 [%rd433+32], %f2871; ld.local.u64 %rd434, [%rd429+112]; ld.local.u32 %r911, [%rd429+108]; ld.local.v2.f32 {%f16444, %f16445}, [%rd429+120]; setp.ne.s32 %p1974, %r5762, -1; @%p1974 bra $L__BB107_2161; ld.local.u32 %r3296, [%rd427+12]; mul.wide.s32 %rd2750, %r3296, 4; add.s64 %rd2751, %rd353, %rd2750; ld.global.u32 %r5762, [%rd2751]; $L__BB107_2161: st.local.u32 [%rd339], %r5762; ld.local.u32 %r914, [%rd427+20]; st.local.u32 [%rd354], %r914; mov.u32 %r3297, 0; st.local.u32 [%rd293], %r3297; cvt.s64.s32 %rd435, %r5762; mul.wide.s32 %rd2752, %r5762, 4; add.s64 %rd2753, %rd287, %rd2752; ld.global.u32 %r915, [%rd2753]; st.local.u32 [%rd290], %r915; and.b32 %r916, %r915, 2; setp.eq.s32 %p1975, %r916, 0; mul.wide.s32 %rd2754, %r5762, 208; add.s64 %rd436, %rd300, %rd2754; @%p1975 bra $L__BB107_2168; ld.global.u32 %rd2755, [%rd436+168]; ld.global.u32 %r3298, [%rd436+152]; shl.b32 %r3299, %r3298, 1; cvt.rn.f32.s32 %f16446, %r3299; mul.ftz.f32 %f16447, %f2730, %f16446; cvt.rzi.ftz.s32.f32 %r3300, %f16447; add.s32 %r3301, %r3299, -1; min.s32 %r3302, %r3300, %r3301; cvt.rn.f32.s32 %f16448, %r3302; sub.ftz.f32 %f2897, %f16447, %f16448; cvt.s64.s32 %rd2756, %r3302; add.s64 %rd2757, %rd2756, %rd2755; shl.b64 %rd2758, %rd2757, 6; add.s64 %rd437, %rd301, %rd2758; ld.global.v4.f32 {%f16449, %f16450, %f16451, %f16452}, [%rd437]; ld.global.v4.f32 {%f16453, %f16454, %f16455, %f16456}, [%rd437+64]; mul.ftz.f32 %f16457, %f16450, %f16454; fma.rn.ftz.f32 %f16458, %f16449, %f16453, %f16457; mul.ftz.f32 %f16459, %f16452, %f16456; fma.rn.ftz.f32 %f16460, %f16451, %f16455, %f16459; add.ftz.f32 %f2906, %f16458, %f16460; setp.gt.ftz.f32 %p1976, %f2906, 0f3F7FDF3B; mov.f32 %f16461, 0f3F800000; sub.ftz.f32 %f2907, %f16461, %f2897; @%p1976 bra $L__BB107_2164; bra.uni $L__BB107_2163; $L__BB107_2164: mul.ftz.f32 %f16515, %f16449, %f2907; fma.rn.ftz.f32 %f16516, %f16453, %f2897, %f16515; mul.ftz.f32 %f16517, %f16450, %f2907; fma.rn.ftz.f32 %f16518, %f16454, %f2897, %f16517; mul.ftz.f32 %f16519, %f16451, %f2907; fma.rn.ftz.f32 %f16520, %f16455, %f2897, %f16519; mul.ftz.f32 %f16521, %f16452, %f2907; fma.rn.ftz.f32 %f16522, %f16456, %f2897, %f16521; mul.ftz.f32 %f16523, %f16518, %f16518; fma.rn.ftz.f32 %f16524, %f16516, %f16516, %f16523; mul.ftz.f32 %f16525, %f16522, %f16522; fma.rn.ftz.f32 %f16526, %f16520, %f16520, %f16525; add.ftz.f32 %f16527, %f16524, %f16526; rsqrt.approx.ftz.f32 %f16528, %f16527; mul.ftz.f32 %f29715, %f16522, %f16528; mul.ftz.f32 %f29714, %f16520, %f16528; mul.ftz.f32 %f29713, %f16518, %f16528; mul.ftz.f32 %f29712, %f16516, %f16528; bra.uni $L__BB107_2165; $L__BB107_2163: mov.f32 %f16462, 0fBF800000; max.ftz.f32 %f16463, %f2906, %f16462; min.ftz.f32 %f16465, %f16463, %f16461; abs.ftz.f32 %f16466, %f16465; sub.ftz.f32 %f16467, %f16461, %f16466; mul.ftz.f32 %f16468, %f16467, 0f3F000000; sqrt.approx.ftz.f32 %f16469, %f16468; setp.gt.ftz.f32 %p1977, %f16466, 0f3F11EB85; selp.f32 %f16470, %f16469, %f16466, %p1977; mul.ftz.f32 %f16471, %f16470, %f16470; mov.f32 %f16472, 0f3C94D2E9; mov.f32 %f16473, 0f3D53F941; fma.rn.ftz.f32 %f16474, %f16473, %f16471, %f16472; mov.f32 %f16475, 0f3D3F841F; fma.rn.ftz.f32 %f16476, %f16474, %f16471, %f16475; mov.f32 %f16477, 0f3D994929; fma.rn.ftz.f32 %f16478, %f16476, %f16471, %f16477; mov.f32 %f16479, 0f3E2AAB94; fma.rn.ftz.f32 %f16480, %f16478, %f16471, %f16479; mul.ftz.f32 %f16481, %f16471, %f16480; fma.rn.ftz.f32 %f16482, %f16481, %f16470, %f16470; add.ftz.f32 %f16483, %f16482, %f16482; mov.f32 %f16484, 0f3FC90FDB; sub.ftz.f32 %f16485, %f16484, %f16482; selp.f32 %f16486, %f16483, %f16485, %p1977; setp.lt.ftz.f32 %p1978, %f16465, 0f00000000; mov.f32 %f16487, 0f40490FDB; sub.ftz.f32 %f16488, %f16487, %f16486; selp.f32 %f16489, %f16488, %f16486, %p1978; mul.ftz.f32 %f16490, %f16449, %f2906; sub.ftz.f32 %f16491, %f16453, %f16490; mul.ftz.f32 %f16492, %f16450, %f2906; sub.ftz.f32 %f16493, %f16454, %f16492; mul.ftz.f32 %f16494, %f16451, %f2906; sub.ftz.f32 %f16495, %f16455, %f16494; mul.ftz.f32 %f16496, %f16452, %f2906; sub.ftz.f32 %f16497, %f16456, %f16496; mul.ftz.f32 %f16498, %f16493, %f16493; fma.rn.ftz.f32 %f16499, %f16491, %f16491, %f16498; mul.ftz.f32 %f16500, %f16497, %f16497; fma.rn.ftz.f32 %f16501, %f16495, %f16495, %f16500; add.ftz.f32 %f16502, %f16499, %f16501; rsqrt.approx.ftz.f32 %f16503, %f16502; mul.ftz.f32 %f16504, %f16491, %f16503; mul.ftz.f32 %f16505, %f16493, %f16503; mul.ftz.f32 %f16506, %f16495, %f16503; mul.ftz.f32 %f16507, %f16497, %f16503; mul.ftz.f32 %f16508, %f2897, %f16489; cos.approx.ftz.f32 %f16509, %f16508; sin.approx.ftz.f32 %f16510, %f16508; mul.ftz.f32 %f16511, %f16504, %f16510; mul.ftz.f32 %f16512, %f16505, %f16510; mul.ftz.f32 %f16513, %f16506, %f16510; mul.ftz.f32 %f16514, %f16507, %f16510; fma.rn.ftz.f32 %f29715, %f16452, %f16509, %f16514; fma.rn.ftz.f32 %f29714, %f16451, %f16509, %f16513; fma.rn.ftz.f32 %f29713, %f16450, %f16509, %f16512; fma.rn.ftz.f32 %f29712, %f16449, %f16509, %f16511; $L__BB107_2165: ld.global.v4.f32 {%f16529, %f16530, %f16531, %f16532}, [%rd437+16]; ld.global.v4.f32 {%f16537, %f16538, %f16539, %f16540}, [%rd437+80]; mul.ftz.f32 %f16545, %f2897, %f16537; mul.ftz.f32 %f16546, %f2897, %f16538; mul.ftz.f32 %f16547, %f2897, %f16539; mul.ftz.f32 %f16548, %f2897, %f16540; fma.rn.ftz.f32 %f2920, %f2907, %f16529, %f16545; fma.rn.ftz.f32 %f2921, %f2907, %f16530, %f16546; fma.rn.ftz.f32 %f2922, %f2907, %f16531, %f16547; fma.rn.ftz.f32 %f16549, %f2907, %f16532, %f16548; ld.global.v4.f32 {%f16550, %f16551, %f16552, %f16553}, [%rd437+32]; ld.global.v4.f32 {%f16558, %f16559, %f16560, %f16561}, [%rd437+96]; mul.ftz.f32 %f16566, %f2897, %f16558; mul.ftz.f32 %f16567, %f2897, %f16559; mul.ftz.f32 %f16568, %f2897, %f16560; mul.ftz.f32 %f16569, %f2897, %f16561; fma.rn.ftz.f32 %f16570, %f2907, %f16550, %f16566; fma.rn.ftz.f32 %f16571, %f2907, %f16551, %f16567; fma.rn.ftz.f32 %f16572, %f2907, %f16552, %f16568; fma.rn.ftz.f32 %f16573, %f2907, %f16553, %f16569; ld.global.v4.f32 {%f16574, %f16575, %f16576, %f16577}, [%rd437+48]; ld.global.v4.f32 {%f16582, %f16583, %f16584, %f16585}, [%rd437+112]; mul.ftz.f32 %f16590, %f2897, %f16582; mul.ftz.f32 %f16591, %f2897, %f16583; mul.ftz.f32 %f16592, %f2897, %f16584; mul.ftz.f32 %f16593, %f2897, %f16585; fma.rn.ftz.f32 %f16594, %f2907, %f16574, %f16590; fma.rn.ftz.f32 %f16595, %f2907, %f16575, %f16591; fma.rn.ftz.f32 %f16596, %f2907, %f16576, %f16592; fma.rn.ftz.f32 %f16597, %f2907, %f16577, %f16593; mul.ftz.f32 %f16598, %f29715, 0f3FB504F3; mul.ftz.f32 %f16599, %f29712, 0f3FB504F3; mul.ftz.f32 %f16600, %f29713, 0f3FB504F3; mul.ftz.f32 %f16601, %f29714, 0f3FB504F3; mul.ftz.f32 %f16602, %f16598, %f16599; mul.ftz.f32 %f16603, %f16598, %f16600; mul.ftz.f32 %f16604, %f16598, %f16601; mul.ftz.f32 %f16605, %f16599, %f16599; mul.ftz.f32 %f16606, %f16599, %f16600; mul.ftz.f32 %f16607, %f16599, %f16601; mul.ftz.f32 %f16608, %f16600, %f16600; mul.ftz.f32 %f16609, %f16600, %f16601; mul.ftz.f32 %f16610, %f16601, %f16601; sub.ftz.f32 %f16612, %f16461, %f16608; sub.ftz.f32 %f16613, %f16612, %f16610; sub.ftz.f32 %f16614, %f16606, %f16604; add.ftz.f32 %f16615, %f16603, %f16607; add.ftz.f32 %f16616, %f16604, %f16606; sub.ftz.f32 %f16617, %f16461, %f16605; sub.ftz.f32 %f16618, %f16617, %f16610; sub.ftz.f32 %f16619, %f16609, %f16602; sub.ftz.f32 %f16620, %f16607, %f16603; add.ftz.f32 %f16621, %f16602, %f16609; sub.ftz.f32 %f16622, %f16617, %f16608; mul.ftz.f32 %f16623, %f16614, %f16572; fma.rn.ftz.f32 %f16624, %f16613, %f16549, %f16623; fma.rn.ftz.f32 %f29721, %f16615, %f16595, %f16624; mul.ftz.f32 %f16625, %f16614, %f16573; fma.rn.ftz.f32 %f16626, %f16613, %f16570, %f16625; fma.rn.ftz.f32 %f2924, %f16615, %f16596, %f16626; mul.ftz.f32 %f16627, %f16614, %f16594; fma.rn.ftz.f32 %f16628, %f16613, %f16571, %f16627; fma.rn.ftz.f32 %f2925, %f16615, %f16597, %f16628; mul.ftz.f32 %f16629, %f16618, %f16572; fma.rn.ftz.f32 %f16630, %f16616, %f16549, %f16629; fma.rn.ftz.f32 %f2926, %f16619, %f16595, %f16630; mul.ftz.f32 %f16631, %f16618, %f16573; fma.rn.ftz.f32 %f16632, %f16616, %f16570, %f16631; fma.rn.ftz.f32 %f29720, %f16619, %f16596, %f16632; mul.ftz.f32 %f16633, %f16618, %f16594; fma.rn.ftz.f32 %f16634, %f16616, %f16571, %f16633; fma.rn.ftz.f32 %f2928, %f16619, %f16597, %f16634; mul.ftz.f32 %f16635, %f16621, %f16572; fma.rn.ftz.f32 %f16636, %f16620, %f16549, %f16635; fma.rn.ftz.f32 %f2929, %f16622, %f16595, %f16636; mul.ftz.f32 %f16637, %f16621, %f16573; fma.rn.ftz.f32 %f16638, %f16620, %f16570, %f16637; fma.rn.ftz.f32 %f2930, %f16622, %f16596, %f16638; mul.ftz.f32 %f16639, %f16621, %f16594; fma.rn.ftz.f32 %f16640, %f16620, %f16571, %f16639; fma.rn.ftz.f32 %f29719, %f16622, %f16597, %f16640; cvta.to.local.u64 %rd2759, %rd299; st.local.v4.f32 [%rd2759], {%f29721, %f2924, %f2925, %f2920}; cvta.to.local.u64 %rd2760, %rd311; st.local.v4.f32 [%rd2760], {%f2926, %f29720, %f2928, %f2921}; cvta.to.local.u64 %rd2761, %rd310; st.local.v4.f32 [%rd2761], {%f2929, %f2930, %f29719, %f2922}; mul.ftz.f32 %f16641, %f29720, %f29719; mul.ftz.f32 %f2932, %f2930, %f2928; sub.ftz.f32 %f29718, %f16641, %f2932; mul.ftz.f32 %f2934, %f2930, %f2925; mul.ftz.f32 %f16642, %f2924, %f29719; sub.ftz.f32 %f29717, %f2934, %f16642; mul.ftz.f32 %f2936, %f2924, %f2928; mul.ftz.f32 %f16643, %f29720, %f2925; sub.ftz.f32 %f29716, %f2936, %f16643; mul.ftz.f32 %f16644, %f2926, %f29717; fma.rn.ftz.f32 %f16645, %f29721, %f29718, %f16644; fma.rn.ftz.f32 %f29722, %f2929, %f29716, %f16645; setp.neu.ftz.f32 %p1979, %f29722, 0f00000000; @%p1979 bra $L__BB107_2167; add.ftz.f32 %f29721, %f29721, 0f322BCC77; add.ftz.f32 %f29719, %f29719, 0f322BCC77; add.ftz.f32 %f29720, %f29720, 0f322BCC77; mul.ftz.f32 %f16646, %f29720, %f29719; sub.ftz.f32 %f29718, %f16646, %f2932; mul.ftz.f32 %f16647, %f2924, %f29719; sub.ftz.f32 %f29717, %f2934, %f16647; mul.ftz.f32 %f16648, %f29720, %f2925; sub.ftz.f32 %f29716, %f2936, %f16648; mul.ftz.f32 %f16649, %f2926, %f29717; fma.rn.ftz.f32 %f16650, %f29721, %f29718, %f16649; fma.rn.ftz.f32 %f16651, %f2929, %f29716, %f16650; setp.eq.ftz.f32 %p1980, %f16651, 0f00000000; selp.f32 %f29722, 0f7F7FFFFF, %f16651, %p1980; $L__BB107_2167: mul.ftz.f32 %f16652, %f2926, %f29719; mul.ftz.f32 %f16653, %f2929, %f2928; sub.ftz.f32 %f16654, %f16653, %f16652; mul.ftz.f32 %f16655, %f2929, %f2925; mul.ftz.f32 %f16656, %f29719, %f29721; sub.ftz.f32 %f16657, %f16656, %f16655; mul.ftz.f32 %f16658, %f2928, %f29721; mul.ftz.f32 %f16659, %f2926, %f2925; sub.ftz.f32 %f16660, %f16659, %f16658; mul.ftz.f32 %f16661, %f2929, %f29720; mul.ftz.f32 %f16662, %f2926, %f2930; sub.ftz.f32 %f16663, %f16662, %f16661; mul.ftz.f32 %f16664, %f2930, %f29721; mul.ftz.f32 %f16665, %f2929, %f2924; sub.ftz.f32 %f16666, %f16665, %f16664; mul.ftz.f32 %f16667, %f2926, %f2924; mul.ftz.f32 %f16668, %f29720, %f29721; sub.ftz.f32 %f16669, %f16668, %f16667; div.approx.ftz.f32 %f16670, %f29717, %f29722; mul.ftz.f32 %f16671, %f2921, %f16670; div.approx.ftz.f32 %f16672, %f29718, %f29722; fma.rn.ftz.f32 %f16673, %f2920, %f16672, %f16671; div.approx.ftz.f32 %f16674, %f29716, %f29722; fma.rn.ftz.f32 %f16675, %f2922, %f16674, %f16673; div.approx.ftz.f32 %f16676, %f16657, %f29722; mul.ftz.f32 %f16677, %f2921, %f16676; div.approx.ftz.f32 %f16678, %f16654, %f29722; fma.rn.ftz.f32 %f16679, %f2920, %f16678, %f16677; div.approx.ftz.f32 %f16680, %f16660, %f29722; fma.rn.ftz.f32 %f16681, %f2922, %f16680, %f16679; div.approx.ftz.f32 %f16682, %f16666, %f29722; mul.ftz.f32 %f16683, %f2921, %f16682; div.approx.ftz.f32 %f16684, %f16663, %f29722; fma.rn.ftz.f32 %f16685, %f2920, %f16684, %f16683; div.approx.ftz.f32 %f16686, %f16669, %f29722; fma.rn.ftz.f32 %f16687, %f2922, %f16686, %f16685; neg.ftz.f32 %f16688, %f16675; cvta.to.local.u64 %rd2762, %rd312; st.local.v4.f32 [%rd2762], {%f16672, %f16670, %f16674, %f16688}; neg.ftz.f32 %f16689, %f16681; cvta.to.local.u64 %rd2763, %rd327; st.local.v4.f32 [%rd2763], {%f16678, %f16676, %f16680, %f16689}; neg.ftz.f32 %f16690, %f16687; cvta.to.local.u64 %rd2764, %rd326; st.local.v4.f32 [%rd2764], {%f16684, %f16682, %f16686, %f16690}; $L__BB107_2168: st.local.f32 [%rd356], %f2730; ld.local.u32 %r3303, [%rd427+12]; st.local.u32 [%rd291], %r3303; ld.local.f32 %f16691, [%rd427]; st.local.f32 [%rd288], %f16691; ld.local.f32 %f2953, [%rd427+4]; st.local.f32 [%rd357], %f2953; ld.local.f32 %f2954, [%rd427+8]; st.local.f32 [%rd351], %f2954; cvt.s64.s32 %rd438, %r3303; mul.wide.s32 %rd2765, %r3303, 4; add.s64 %rd2766, %rd333, %rd2765; ld.global.u32 %r917, [%rd2766]; st.local.u32 [%rd305], %r917; and.b32 %r3304, %r914, 1; setp.eq.b32 %p1981, %r3304, 1; xor.pred %p1983, %p1981, %p1284; not.pred %p1984, %p1983; mul.wide.s32 %rd2767, %r3303, 16; add.s64 %rd439, %rd328, %rd2767; @%p1984 bra $L__BB107_2174; bra.uni $L__BB107_2169; $L__BB107_2174: ld.global.u32 %r919, [%rd436+156]; ld.global.u32 %r920, [%rd436+152]; shl.b32 %r3316, %r920, 1; cvt.rn.f32.s32 %f16747, %r3316; mul.ftz.f32 %f16748, %f2730, %f16747; cvt.rzi.ftz.s32.f32 %r3317, %f16748; add.s32 %r3318, %r3316, -1; min.s32 %r921, %r3317, %r3318; cvt.rn.f32.s32 %f16749, %r921; sub.ftz.f32 %f3009, %f16748, %f16749; ld.global.u32 %r5770, [%rd436+164]; mul.wide.u32 %rd2786, %r5770, 16; add.s64 %rd2787, %rd331, %rd2786; ld.global.u16 %rs711, [%rd2787+12]; ld.global.u32 %r5769, [%rd2787+8]; ld.global.u64 %rd4815, [%rd2787]; setp.eq.s64 %p1987, %rd4815, 11; mov.u16 %rs709, %rs711; mov.u32 %r5764, %r5769; @%p1987 bra $L__BB107_2180; mov.u64 %rd4796, %rd4815; mov.u32 %r5764, %r5769; mov.u16 %rs709, %rs711; mov.u32 %r5765, %r5770; $L__BB107_2176: setp.eq.s64 %p1988, %rd4796, 0; @%p1988 bra $L__BB107_2178; add.s32 %r5765, %r5765, 2; bra.uni $L__BB107_2179; $L__BB107_2178: setp.eq.s16 %p1989, %rs709, 0; mov.u32 %r5765, %r5764; mov.u32 %r5768, %r3028; @%p1989 bra $L__BB107_2181; $L__BB107_2179: mul.wide.u32 %rd2788, %r5765, 16; add.s64 %rd2789, %rd331, %rd2788; ld.global.u16 %rs709, [%rd2789+12]; ld.global.u32 %r5764, [%rd2789+8]; ld.global.u64 %rd4796, [%rd2789]; setp.ne.s64 %p1990, %rd4796, 11; @%p1990 bra $L__BB107_2176; $L__BB107_2180: setp.eq.s16 %p1991, %rs709, 0; selp.b32 %r5768, -1, %r5764, %p1991; $L__BB107_2181: ld.global.v4.u32 {%r3320, %r3321, %r3322, %r3323}, [%rd439]; setp.eq.s32 %p1992, %r921, %r920; @%p1992 bra $L__BB107_2183; bra.uni $L__BB107_2182; $L__BB107_2183: mul.wide.u32 %rd2793, %r3323, 12; add.s64 %rd4797, %rd329, %rd2793; add.s32 %r3330, %r3323, 1; mul.wide.u32 %rd2794, %r3330, 12; add.s64 %rd4800, %rd329, %rd2794; add.s32 %r3331, %r3323, 2; mul.wide.u32 %rd2795, %r3331, 12; add.s64 %rd4803, %rd329, %rd2795; add.s64 %rd4798, %rd4797, 4; add.s64 %rd4799, %rd4797, 8; add.s64 %rd4801, %rd4800, 4; add.s64 %rd4802, %rd4800, 8; add.s64 %rd4804, %rd4803, 4; add.s64 %rd4805, %rd4803, 8; bra.uni $L__BB107_2184; $L__BB107_2169: ld.global.v4.u32 {%r3305, %r3306, %r3307, %r3308}, [%rd439]; mul.wide.u32 %rd2768, %r3308, 12; add.s64 %rd2769, %rd329, %rd2768; add.s32 %r3313, %r3308, 1; mul.wide.u32 %rd2770, %r3313, 12; add.s64 %rd2771, %rd329, %rd2770; add.s32 %r3314, %r3308, 2; mul.wide.u32 %rd2772, %r3314, 12; add.s64 %rd2773, %rd329, %rd2772; mul.wide.u32 %rd2774, %r3305, 12; add.s64 %rd2775, %rd352, %rd2774; ld.global.f32 %f29750, [%rd2775]; ld.global.f32 %f29751, [%rd2775+4]; ld.global.f32 %f29752, [%rd2775+8]; mul.wide.u32 %rd2776, %r3306, 12; add.s64 %rd2777, %rd352, %rd2776; ld.global.f32 %f29753, [%rd2777]; ld.global.f32 %f29754, [%rd2777+4]; ld.global.f32 %f29755, [%rd2777+8]; mul.wide.u32 %rd2778, %r3307, 12; add.s64 %rd2779, %rd352, %rd2778; ld.global.f32 %f29756, [%rd2779]; ld.global.f32 %f29757, [%rd2779+4]; ld.global.f32 %f29758, [%rd2779+8]; mov.f32 %f16692, 0f3F800000; sub.ftz.f32 %f16693, %f16692, %f2953; sub.ftz.f32 %f16694, %f16693, %f2954; ld.global.f32 %f29759, [%rd2769]; ld.global.f32 %f29760, [%rd2769+4]; ld.global.f32 %f29761, [%rd2769+8]; ld.global.f32 %f29762, [%rd2771]; mul.ftz.f32 %f16695, %f29762, %f2953; ld.global.f32 %f29763, [%rd2771+4]; mul.ftz.f32 %f16696, %f29763, %f2953; ld.global.f32 %f29764, [%rd2771+8]; mul.ftz.f32 %f16697, %f29764, %f2953; fma.rn.ftz.f32 %f16698, %f29759, %f16694, %f16695; fma.rn.ftz.f32 %f16699, %f29760, %f16694, %f16696; fma.rn.ftz.f32 %f16700, %f29761, %f16694, %f16697; ld.global.f32 %f29765, [%rd2773]; ld.global.f32 %f29766, [%rd2773+4]; ld.global.f32 %f29767, [%rd2773+8]; fma.rn.ftz.f32 %f16701, %f29767, %f2954, %f16700; fma.rn.ftz.f32 %f16702, %f29766, %f2954, %f16699; fma.rn.ftz.f32 %f16703, %f29765, %f2954, %f16698; st.local.v2.f32 [%rd2173], {%f16703, %f16702}; st.local.f32 [%rd314], %f16701; and.b32 %r5774, %r915, 4; setp.ne.s32 %p1985, %r5774, 0; @%p1985 bra $L__BB107_2205; @%p1975 bra $L__BB107_2172; bra.uni $L__BB107_2171; $L__BB107_2172: ld.global.v4.f32 {%f29723, %f29724, %f29725, %f29726}, [%rd436]; ld.global.v4.f32 {%f29727, %f29728, %f29729, %f29730}, [%rd436+16]; ld.global.v4.f32 {%f29731, %f29732, %f29733, %f29734}, [%rd436+32]; bra.uni $L__BB107_2173; $L__BB107_2182: setp.gt.s32 %p1993, %r921, %r920; selp.b32 %r3324, -1, 0, %p1993; add.s32 %r3325, %r921, %r3324; mad.lo.s32 %r3326, %r3325, %r919, %r5768; add.s32 %r3327, %r3320, %r3326; mul.wide.u32 %rd2790, %r3327, 12; add.s64 %rd4797, %rd407, %rd2790; add.s32 %r3328, %r3321, %r3326; mul.wide.u32 %rd2791, %r3328, 12; add.s64 %rd4800, %rd407, %rd2791; add.s32 %r3329, %r3322, %r3326; mul.wide.u32 %rd2792, %r3329, 12; add.s64 %rd4803, %rd407, %rd2792; add.s64 %rd4798, %rd4797, 4; add.s64 %rd4799, %rd4797, 8; add.s64 %rd4801, %rd4800, 4; add.s64 %rd4802, %rd4800, 8; add.s64 %rd4804, %rd4803, 4; add.s64 %rd4805, %rd4803, 8; $L__BB107_2184: ld.global.f32 %f3010, [%rd4805]; ld.global.f32 %f3011, [%rd4804]; ld.global.f32 %f3012, [%rd4803]; ld.global.f32 %f3013, [%rd4802]; ld.global.f32 %f3014, [%rd4801]; ld.global.f32 %f3015, [%rd4800]; ld.global.f32 %f3016, [%rd4799]; ld.global.f32 %f3017, [%rd4798]; ld.global.f32 %f3018, [%rd4797]; add.s32 %r936, %r921, 1; setp.eq.s32 %p1994, %r936, %r920; @%p1994 bra $L__BB107_2186; bra.uni $L__BB107_2185; $L__BB107_2186: mul.wide.u32 %rd2799, %r3323, 12; add.s64 %rd4806, %rd329, %rd2799; add.s32 %r3338, %r3323, 1; mul.wide.u32 %rd2800, %r3338, 12; add.s64 %rd4809, %rd329, %rd2800; add.s32 %r3339, %r3323, 2; mul.wide.u32 %rd2801, %r3339, 12; add.s64 %rd4812, %rd329, %rd2801; add.s64 %rd4807, %rd4806, 4; add.s64 %rd4808, %rd4806, 8; add.s64 %rd4810, %rd4809, 4; add.s64 %rd4811, %rd4809, 8; add.s64 %rd4813, %rd4812, 4; add.s64 %rd4814, %rd4812, 8; bra.uni $L__BB107_2187; $L__BB107_2185: setp.ge.s32 %p1995, %r921, %r920; selp.b32 %r3332, -1, 0, %p1995; add.s32 %r3333, %r936, %r3332; mad.lo.s32 %r3334, %r3333, %r919, %r5768; add.s32 %r3335, %r3320, %r3334; mul.wide.u32 %rd2796, %r3335, 12; add.s64 %rd4806, %rd407, %rd2796; add.s32 %r3336, %r3321, %r3334; mul.wide.u32 %rd2797, %r3336, 12; add.s64 %rd4809, %rd407, %rd2797; add.s32 %r3337, %r3322, %r3334; mul.wide.u32 %rd2798, %r3337, 12; add.s64 %rd4812, %rd407, %rd2798; add.s64 %rd4807, %rd4806, 4; add.s64 %rd4808, %rd4806, 8; add.s64 %rd4810, %rd4809, 4; add.s64 %rd4811, %rd4809, 8; add.s64 %rd4813, %rd4812, 4; add.s64 %rd4814, %rd4812, 8; $L__BB107_2187: mov.f32 %f16750, 0f3F800000; sub.ftz.f32 %f3019, %f16750, %f3009; ld.global.f32 %f16751, [%rd4806]; mul.ftz.f32 %f16752, %f3009, %f16751; ld.global.f32 %f16753, [%rd4807]; mul.ftz.f32 %f16754, %f3009, %f16753; ld.global.f32 %f16755, [%rd4808]; mul.ftz.f32 %f16756, %f3009, %f16755; fma.rn.ftz.f32 %f29759, %f3019, %f3018, %f16752; fma.rn.ftz.f32 %f29760, %f3019, %f3017, %f16754; fma.rn.ftz.f32 %f29761, %f3019, %f3016, %f16756; ld.global.f32 %f16757, [%rd4809]; mul.ftz.f32 %f16758, %f3009, %f16757; ld.global.f32 %f16759, [%rd4810]; mul.ftz.f32 %f16760, %f3009, %f16759; ld.global.f32 %f16761, [%rd4811]; mul.ftz.f32 %f16762, %f3009, %f16761; fma.rn.ftz.f32 %f29762, %f3019, %f3015, %f16758; fma.rn.ftz.f32 %f29763, %f3019, %f3014, %f16760; fma.rn.ftz.f32 %f29764, %f3019, %f3013, %f16762; ld.global.f32 %f16763, [%rd4812]; mul.ftz.f32 %f16764, %f3009, %f16763; ld.global.f32 %f16765, [%rd4813]; mul.ftz.f32 %f16766, %f3009, %f16765; ld.global.f32 %f16767, [%rd4814]; mul.ftz.f32 %f16768, %f3009, %f16767; fma.rn.ftz.f32 %f29765, %f3019, %f3012, %f16764; fma.rn.ftz.f32 %f29766, %f3019, %f3011, %f16766; fma.rn.ftz.f32 %f29767, %f3019, %f3010, %f16768; setp.eq.s64 %p1996, %rd4815, 12; @%p1996 bra $L__BB107_2192; $L__BB107_2188: setp.eq.s64 %p1997, %rd4815, 0; @%p1997 bra $L__BB107_2190; add.s32 %r5770, %r5770, 2; bra.uni $L__BB107_2191; $L__BB107_2190: setp.eq.s16 %p1998, %rs711, 0; mov.u32 %r5773, -1; mov.u32 %r5770, %r5769; @%p1998 bra $L__BB107_2193; $L__BB107_2191: mul.wide.u32 %rd2802, %r5770, 16; add.s64 %rd2803, %rd331, %rd2802; ld.global.u16 %rs711, [%rd2803+12]; ld.global.u32 %r5769, [%rd2803+8]; ld.global.u64 %rd4815, [%rd2803]; setp.ne.s64 %p1999, %rd4815, 12; @%p1999 bra $L__BB107_2188; $L__BB107_2192: setp.eq.s16 %p2000, %rs711, 0; selp.b32 %r5773, -1, %r5769, %p2000; $L__BB107_2193: @%p1992 bra $L__BB107_2195; bra.uni $L__BB107_2194; $L__BB107_2195: mul.wide.u32 %rd2807, %r3320, 12; add.s64 %rd4816, %rd352, %rd2807; mul.wide.u32 %rd2808, %r3321, 12; add.s64 %rd4819, %rd352, %rd2808; mul.wide.u32 %rd2809, %r3322, 12; add.s64 %rd4822, %rd352, %rd2809; add.s64 %rd4817, %rd4816, 4; add.s64 %rd4818, %rd4816, 8; add.s64 %rd4820, %rd4819, 4; add.s64 %rd4821, %rd4819, 8; add.s64 %rd4823, %rd4822, 4; add.s64 %rd4824, %rd4822, 8; bra.uni $L__BB107_2196; $L__BB107_2194: setp.gt.s32 %p2002, %r921, %r920; selp.b32 %r3341, -1, 0, %p2002; add.s32 %r3342, %r921, %r3341; mad.lo.s32 %r3343, %r3342, %r919, %r5773; add.s32 %r3344, %r3343, %r3320; mul.wide.u32 %rd2804, %r3344, 12; add.s64 %rd4816, %rd407, %rd2804; add.s32 %r3345, %r3343, %r3321; mul.wide.u32 %rd2805, %r3345, 12; add.s64 %rd4819, %rd407, %rd2805; add.s32 %r3346, %r3343, %r3322; mul.wide.u32 %rd2806, %r3346, 12; add.s64 %rd4822, %rd407, %rd2806; add.s64 %rd4817, %rd4816, 4; add.s64 %rd4818, %rd4816, 8; add.s64 %rd4820, %rd4819, 4; add.s64 %rd4821, %rd4819, 8; add.s64 %rd4823, %rd4822, 4; add.s64 %rd4824, %rd4822, 8; $L__BB107_2196: ld.global.f32 %f3029, [%rd4824]; ld.global.f32 %f3030, [%rd4823]; ld.global.f32 %f3031, [%rd4822]; ld.global.f32 %f3032, [%rd4821]; ld.global.f32 %f3033, [%rd4820]; ld.global.f32 %f3034, [%rd4819]; ld.global.f32 %f3035, [%rd4818]; ld.global.f32 %f3036, [%rd4817]; ld.global.f32 %f3037, [%rd4816]; @%p1994 bra $L__BB107_2198; bra.uni $L__BB107_2197; $L__BB107_2198: mul.wide.u32 %rd2813, %r3320, 12; add.s64 %rd4825, %rd352, %rd2813; mul.wide.u32 %rd2814, %r3321, 12; add.s64 %rd4828, %rd352, %rd2814; mul.wide.u32 %rd2815, %r3322, 12; add.s64 %rd4831, %rd352, %rd2815; add.s64 %rd4826, %rd4825, 4; add.s64 %rd4827, %rd4825, 8; add.s64 %rd4829, %rd4828, 4; add.s64 %rd4830, %rd4828, 8; add.s64 %rd4832, %rd4831, 4; add.s64 %rd4833, %rd4831, 8; bra.uni $L__BB107_2199; $L__BB107_2197: setp.ge.s32 %p2004, %r921, %r920; selp.b32 %r3347, -1, 0, %p2004; add.s32 %r3348, %r936, %r3347; mad.lo.s32 %r3349, %r3348, %r919, %r5773; add.s32 %r3350, %r3349, %r3320; mul.wide.u32 %rd2810, %r3350, 12; add.s64 %rd4825, %rd407, %rd2810; add.s32 %r3351, %r3349, %r3321; mul.wide.u32 %rd2811, %r3351, 12; add.s64 %rd4828, %rd407, %rd2811; add.s32 %r3352, %r3349, %r3322; mul.wide.u32 %rd2812, %r3352, 12; add.s64 %rd4831, %rd407, %rd2812; add.s64 %rd4826, %rd4825, 4; add.s64 %rd4827, %rd4825, 8; add.s64 %rd4829, %rd4828, 4; add.s64 %rd4830, %rd4828, 8; add.s64 %rd4832, %rd4831, 4; add.s64 %rd4833, %rd4831, 8; $L__BB107_2199: ld.global.f32 %f16769, [%rd4825]; mul.ftz.f32 %f16770, %f3009, %f16769; ld.global.f32 %f16771, [%rd4826]; mul.ftz.f32 %f16772, %f3009, %f16771; ld.global.f32 %f16773, [%rd4827]; mul.ftz.f32 %f16774, %f3009, %f16773; fma.rn.ftz.f32 %f29750, %f3019, %f3037, %f16770; fma.rn.ftz.f32 %f29751, %f3019, %f3036, %f16772; fma.rn.ftz.f32 %f29752, %f3019, %f3035, %f16774; ld.global.f32 %f16775, [%rd4828]; mul.ftz.f32 %f16776, %f3009, %f16775; ld.global.f32 %f16777, [%rd4829]; mul.ftz.f32 %f16778, %f3009, %f16777; ld.global.f32 %f16779, [%rd4830]; mul.ftz.f32 %f16780, %f3009, %f16779; fma.rn.ftz.f32 %f29753, %f3019, %f3034, %f16776; fma.rn.ftz.f32 %f29754, %f3019, %f3033, %f16778; fma.rn.ftz.f32 %f29755, %f3019, %f3032, %f16780; ld.global.f32 %f16781, [%rd4831]; mul.ftz.f32 %f16782, %f3009, %f16781; ld.global.f32 %f16783, [%rd4832]; mul.ftz.f32 %f16784, %f3009, %f16783; ld.global.f32 %f16785, [%rd4833]; mul.ftz.f32 %f16786, %f3009, %f16785; fma.rn.ftz.f32 %f29756, %f3019, %f3031, %f16782; fma.rn.ftz.f32 %f29757, %f3019, %f3030, %f16784; fma.rn.ftz.f32 %f29758, %f3019, %f3029, %f16786; sub.ftz.f32 %f16787, %f29762, %f29759; sub.ftz.f32 %f16788, %f29763, %f29760; sub.ftz.f32 %f16789, %f29764, %f29761; fma.rn.ftz.f32 %f16790, %f16787, %f2953, %f29759; fma.rn.ftz.f32 %f16791, %f16788, %f2953, %f29760; fma.rn.ftz.f32 %f16792, %f16789, %f2953, %f29761; sub.ftz.f32 %f16793, %f29765, %f29759; sub.ftz.f32 %f16794, %f29766, %f29760; sub.ftz.f32 %f16795, %f29767, %f29761; fma.rn.ftz.f32 %f29747, %f16793, %f2954, %f16790; fma.rn.ftz.f32 %f29748, %f16794, %f2954, %f16791; fma.rn.ftz.f32 %f29749, %f16795, %f2954, %f16792; and.b32 %r5774, %r915, 4; setp.ne.s32 %p2005, %r5774, 0; @%p2005 bra $L__BB107_2204; @%p1975 bra $L__BB107_2202; bra.uni $L__BB107_2201; $L__BB107_2202: ld.global.v4.f32 {%f29735, %f29736, %f29737, %f29738}, [%rd436]; ld.global.v4.f32 {%f29739, %f29740, %f29741, %f29742}, [%rd436+16]; ld.global.v4.f32 {%f29743, %f29744, %f29745, %f29746}, [%rd436+32]; bra.uni $L__BB107_2203; $L__BB107_2171: cvta.to.local.u64 %rd2781, %rd299; ld.local.v4.f32 {%f29723, %f29724, %f29725, %f29726}, [%rd2781]; cvta.to.local.u64 %rd2782, %rd311; ld.local.v4.f32 {%f29727, %f29728, %f29729, %f29730}, [%rd2782]; cvta.to.local.u64 %rd2783, %rd310; ld.local.v4.f32 {%f29731, %f29732, %f29733, %f29734}, [%rd2783]; $L__BB107_2173: ld.local.v4.f32 {%f16728, %f16729, %f16730, %f16731}, [%rd2173]; mul.ftz.f32 %f16735, %f29724, %f16729; fma.rn.ftz.f32 %f16736, %f29723, %f16728, %f16735; fma.rn.ftz.f32 %f16737, %f29725, %f16730, %f16736; mul.ftz.f32 %f16738, %f29728, %f16729; fma.rn.ftz.f32 %f16739, %f29727, %f16728, %f16738; fma.rn.ftz.f32 %f16740, %f29729, %f16730, %f16739; mul.ftz.f32 %f16741, %f29732, %f16729; fma.rn.ftz.f32 %f16742, %f29731, %f16728, %f16741; fma.rn.ftz.f32 %f16743, %f29733, %f16730, %f16742; add.ftz.f32 %f16744, %f29734, %f16743; add.ftz.f32 %f16745, %f29726, %f16737; add.ftz.f32 %f16746, %f29730, %f16740; st.local.v2.f32 [%rd2173], {%f16745, %f16746}; st.local.f32 [%rd314], %f16744; mov.u32 %r5774, %r3297; bra.uni $L__BB107_2205; $L__BB107_2201: cvta.to.local.u64 %rd2816, %rd299; ld.local.v4.f32 {%f29735, %f29736, %f29737, %f29738}, [%rd2816]; cvta.to.local.u64 %rd2817, %rd311; ld.local.v4.f32 {%f29739, %f29740, %f29741, %f29742}, [%rd2817]; cvta.to.local.u64 %rd2818, %rd310; ld.local.v4.f32 {%f29743, %f29744, %f29745, %f29746}, [%rd2818]; $L__BB107_2203: mul.ftz.f32 %f16820, %f29748, %f29736; fma.rn.ftz.f32 %f16821, %f29747, %f29735, %f16820; fma.rn.ftz.f32 %f16822, %f29749, %f29737, %f16821; mul.ftz.f32 %f16823, %f29748, %f29740; fma.rn.ftz.f32 %f16824, %f29747, %f29739, %f16823; fma.rn.ftz.f32 %f16825, %f29749, %f29741, %f16824; mul.ftz.f32 %f16826, %f29748, %f29744; fma.rn.ftz.f32 %f16827, %f29747, %f29743, %f16826; fma.rn.ftz.f32 %f16828, %f29749, %f29745, %f16827; add.ftz.f32 %f29747, %f29738, %f16822; add.ftz.f32 %f29748, %f29742, %f16825; add.ftz.f32 %f29749, %f29746, %f16828; $L__BB107_2204: st.local.v2.f32 [%rd2173], {%f29747, %f29748}; st.local.f32 [%rd314], %f29749; $L__BB107_2205: setp.ne.s32 %p2007, %r5774, 0; @%p2007 bra $L__BB107_2224; @%p1975 bra $L__BB107_2208; bra.uni $L__BB107_2207; $L__BB107_2208: ld.global.v4.f32 {%f16862, %f16863, %f16864, %f16865}, [%rd436]; ld.global.v4.f32 {%f16870, %f16871, %f16872, %f16873}, [%rd436+16]; ld.global.v4.f32 {%f16878, %f16879, %f16880, %f16881}, [%rd436+32]; mul.ftz.f32 %f16886, %f29760, %f16863; fma.rn.ftz.f32 %f16887, %f29759, %f16862, %f16886; fma.rn.ftz.f32 %f16888, %f29761, %f16864, %f16887; add.ftz.f32 %f29768, %f16865, %f16888; mul.ftz.f32 %f16889, %f29760, %f16871; fma.rn.ftz.f32 %f16890, %f29759, %f16870, %f16889; fma.rn.ftz.f32 %f16891, %f29761, %f16872, %f16890; add.ftz.f32 %f29769, %f16873, %f16891; mul.ftz.f32 %f16892, %f29760, %f16879; fma.rn.ftz.f32 %f16893, %f29759, %f16878, %f16892; fma.rn.ftz.f32 %f16894, %f29761, %f16880, %f16893; add.ftz.f32 %f29761, %f16881, %f16894; bra.uni $L__BB107_2209; $L__BB107_2207: cvta.to.local.u64 %rd2821, %rd299; ld.local.v4.f32 {%f16829, %f16830, %f16831, %f16832}, [%rd2821]; mul.ftz.f32 %f16837, %f29760, %f16830; fma.rn.ftz.f32 %f16838, %f29759, %f16829, %f16837; fma.rn.ftz.f32 %f16839, %f29761, %f16831, %f16838; add.ftz.f32 %f29768, %f16832, %f16839; cvta.to.local.u64 %rd2822, %rd311; ld.local.v4.f32 {%f16840, %f16841, %f16842, %f16843}, [%rd2822]; mul.ftz.f32 %f16848, %f29760, %f16841; fma.rn.ftz.f32 %f16849, %f29759, %f16840, %f16848; fma.rn.ftz.f32 %f16850, %f29761, %f16842, %f16849; add.ftz.f32 %f29769, %f16843, %f16850; cvta.to.local.u64 %rd2823, %rd310; ld.local.v4.f32 {%f16851, %f16852, %f16853, %f16854}, [%rd2823]; mul.ftz.f32 %f16859, %f29760, %f16852; fma.rn.ftz.f32 %f16860, %f29759, %f16851, %f16859; fma.rn.ftz.f32 %f16861, %f29761, %f16853, %f16860; add.ftz.f32 %f29761, %f16854, %f16861; $L__BB107_2209: @%p1975 bra $L__BB107_2211; bra.uni $L__BB107_2210; $L__BB107_2211: ld.global.v4.f32 {%f16928, %f16929, %f16930, %f16931}, [%rd436]; ld.global.v4.f32 {%f16936, %f16937, %f16938, %f16939}, [%rd436+16]; ld.global.v4.f32 {%f16944, %f16945, %f16946, %f16947}, [%rd436+32]; mul.ftz.f32 %f16952, %f29763, %f16929; fma.rn.ftz.f32 %f16953, %f29762, %f16928, %f16952; fma.rn.ftz.f32 %f16954, %f29764, %f16930, %f16953; add.ftz.f32 %f29771, %f16931, %f16954; mul.ftz.f32 %f16955, %f29763, %f16937; fma.rn.ftz.f32 %f16956, %f29762, %f16936, %f16955; fma.rn.ftz.f32 %f16957, %f29764, %f16938, %f16956; add.ftz.f32 %f29772, %f16939, %f16957; mul.ftz.f32 %f16958, %f29763, %f16945; fma.rn.ftz.f32 %f16959, %f29762, %f16944, %f16958; fma.rn.ftz.f32 %f16960, %f29764, %f16946, %f16959; add.ftz.f32 %f29764, %f16947, %f16960; bra.uni $L__BB107_2212; $L__BB107_2210: cvta.to.local.u64 %rd2824, %rd299; ld.local.v4.f32 {%f16895, %f16896, %f16897, %f16898}, [%rd2824]; mul.ftz.f32 %f16903, %f29763, %f16896; fma.rn.ftz.f32 %f16904, %f29762, %f16895, %f16903; fma.rn.ftz.f32 %f16905, %f29764, %f16897, %f16904; add.ftz.f32 %f29771, %f16898, %f16905; cvta.to.local.u64 %rd2825, %rd311; ld.local.v4.f32 {%f16906, %f16907, %f16908, %f16909}, [%rd2825]; mul.ftz.f32 %f16914, %f29763, %f16907; fma.rn.ftz.f32 %f16915, %f29762, %f16906, %f16914; fma.rn.ftz.f32 %f16916, %f29764, %f16908, %f16915; add.ftz.f32 %f29772, %f16909, %f16916; cvta.to.local.u64 %rd2826, %rd310; ld.local.v4.f32 {%f16917, %f16918, %f16919, %f16920}, [%rd2826]; mul.ftz.f32 %f16925, %f29763, %f16918; fma.rn.ftz.f32 %f16926, %f29762, %f16917, %f16925; fma.rn.ftz.f32 %f16927, %f29764, %f16919, %f16926; add.ftz.f32 %f29764, %f16920, %f16927; $L__BB107_2212: @%p1975 bra $L__BB107_2214; bra.uni $L__BB107_2213; $L__BB107_2214: ld.global.v4.f32 {%f16994, %f16995, %f16996, %f16997}, [%rd436]; ld.global.v4.f32 {%f17002, %f17003, %f17004, %f17005}, [%rd436+16]; ld.global.v4.f32 {%f17010, %f17011, %f17012, %f17013}, [%rd436+32]; mul.ftz.f32 %f17018, %f29766, %f16995; fma.rn.ftz.f32 %f17019, %f29765, %f16994, %f17018; fma.rn.ftz.f32 %f17020, %f29767, %f16996, %f17019; add.ftz.f32 %f29774, %f16997, %f17020; mul.ftz.f32 %f17021, %f29766, %f17003; fma.rn.ftz.f32 %f17022, %f29765, %f17002, %f17021; fma.rn.ftz.f32 %f17023, %f29767, %f17004, %f17022; add.ftz.f32 %f29775, %f17005, %f17023; mul.ftz.f32 %f17024, %f29766, %f17011; fma.rn.ftz.f32 %f17025, %f29765, %f17010, %f17024; fma.rn.ftz.f32 %f17026, %f29767, %f17012, %f17025; add.ftz.f32 %f29767, %f17013, %f17026; bra.uni $L__BB107_2215; $L__BB107_2213: cvta.to.local.u64 %rd2827, %rd299; ld.local.v4.f32 {%f16961, %f16962, %f16963, %f16964}, [%rd2827]; mul.ftz.f32 %f16969, %f29766, %f16962; fma.rn.ftz.f32 %f16970, %f29765, %f16961, %f16969; fma.rn.ftz.f32 %f16971, %f29767, %f16963, %f16970; add.ftz.f32 %f29774, %f16964, %f16971; cvta.to.local.u64 %rd2828, %rd311; ld.local.v4.f32 {%f16972, %f16973, %f16974, %f16975}, [%rd2828]; mul.ftz.f32 %f16980, %f29766, %f16973; fma.rn.ftz.f32 %f16981, %f29765, %f16972, %f16980; fma.rn.ftz.f32 %f16982, %f29767, %f16974, %f16981; add.ftz.f32 %f29775, %f16975, %f16982; cvta.to.local.u64 %rd2829, %rd310; ld.local.v4.f32 {%f16983, %f16984, %f16985, %f16986}, [%rd2829]; mul.ftz.f32 %f16991, %f29766, %f16984; fma.rn.ftz.f32 %f16992, %f29765, %f16983, %f16991; fma.rn.ftz.f32 %f16993, %f29767, %f16985, %f16992; add.ftz.f32 %f29767, %f16986, %f16993; $L__BB107_2215: @%p1975 bra $L__BB107_2217; bra.uni $L__BB107_2216; $L__BB107_2217: ld.global.v4.f32 {%f17061, %f17062, %f17063, %f17064}, [%rd436+48]; ld.global.v4.f32 {%f17068, %f17069, %f17070, %f17071}, [%rd436+64]; ld.global.v4.f32 {%f17075, %f17076, %f17077, %f17078}, [%rd436+80]; mul.ftz.f32 %f17082, %f29751, %f17068; fma.rn.ftz.f32 %f17083, %f29750, %f17061, %f17082; fma.rn.ftz.f32 %f17084, %f29752, %f17075, %f17083; mul.ftz.f32 %f17085, %f29751, %f17069; fma.rn.ftz.f32 %f17086, %f29750, %f17062, %f17085; fma.rn.ftz.f32 %f17087, %f29752, %f17076, %f17086; mul.ftz.f32 %f17088, %f29751, %f17070; fma.rn.ftz.f32 %f17089, %f29750, %f17063, %f17088; fma.rn.ftz.f32 %f17090, %f29752, %f17077, %f17089; mul.ftz.f32 %f17091, %f17087, %f17087; fma.rn.ftz.f32 %f17092, %f17084, %f17084, %f17091; fma.rn.ftz.f32 %f17093, %f17090, %f17090, %f17092; sqrt.approx.ftz.f32 %f17094, %f17093; div.approx.ftz.f32 %f29750, %f17084, %f17094; div.approx.ftz.f32 %f29751, %f17087, %f17094; div.approx.ftz.f32 %f29752, %f17090, %f17094; bra.uni $L__BB107_2218; $L__BB107_2216: cvta.to.local.u64 %rd2830, %rd312; ld.local.v4.f32 {%f17027, %f17028, %f17029, %f17030}, [%rd2830]; cvta.to.local.u64 %rd2831, %rd327; ld.local.v4.f32 {%f17034, %f17035, %f17036, %f17037}, [%rd2831]; cvta.to.local.u64 %rd2832, %rd326; ld.local.v4.f32 {%f17041, %f17042, %f17043, %f17044}, [%rd2832]; mul.ftz.f32 %f17048, %f29751, %f17034; fma.rn.ftz.f32 %f17049, %f29750, %f17027, %f17048; fma.rn.ftz.f32 %f17050, %f29752, %f17041, %f17049; mul.ftz.f32 %f17051, %f29751, %f17035; fma.rn.ftz.f32 %f17052, %f29750, %f17028, %f17051; fma.rn.ftz.f32 %f17053, %f29752, %f17042, %f17052; mul.ftz.f32 %f17054, %f29751, %f17036; fma.rn.ftz.f32 %f17055, %f29750, %f17029, %f17054; fma.rn.ftz.f32 %f17056, %f29752, %f17043, %f17055; mul.ftz.f32 %f17057, %f17053, %f17053; fma.rn.ftz.f32 %f17058, %f17050, %f17050, %f17057; fma.rn.ftz.f32 %f17059, %f17056, %f17056, %f17058; sqrt.approx.ftz.f32 %f17060, %f17059; div.approx.ftz.f32 %f29750, %f17050, %f17060; div.approx.ftz.f32 %f29751, %f17053, %f17060; div.approx.ftz.f32 %f29752, %f17056, %f17060; $L__BB107_2218: @%p1975 bra $L__BB107_2220; bra.uni $L__BB107_2219; $L__BB107_2220: ld.global.v4.f32 {%f17129, %f17130, %f17131, %f17132}, [%rd436+48]; ld.global.v4.f32 {%f17136, %f17137, %f17138, %f17139}, [%rd436+64]; ld.global.v4.f32 {%f17143, %f17144, %f17145, %f17146}, [%rd436+80]; mul.ftz.f32 %f17150, %f29754, %f17136; fma.rn.ftz.f32 %f17151, %f29753, %f17129, %f17150; fma.rn.ftz.f32 %f17152, %f29755, %f17143, %f17151; mul.ftz.f32 %f17153, %f29754, %f17137; fma.rn.ftz.f32 %f17154, %f29753, %f17130, %f17153; fma.rn.ftz.f32 %f17155, %f29755, %f17144, %f17154; mul.ftz.f32 %f17156, %f29754, %f17138; fma.rn.ftz.f32 %f17157, %f29753, %f17131, %f17156; fma.rn.ftz.f32 %f17158, %f29755, %f17145, %f17157; mul.ftz.f32 %f17159, %f17155, %f17155; fma.rn.ftz.f32 %f17160, %f17152, %f17152, %f17159; fma.rn.ftz.f32 %f17161, %f17158, %f17158, %f17160; sqrt.approx.ftz.f32 %f17162, %f17161; div.approx.ftz.f32 %f29753, %f17152, %f17162; div.approx.ftz.f32 %f29754, %f17155, %f17162; div.approx.ftz.f32 %f29755, %f17158, %f17162; bra.uni $L__BB107_2221; $L__BB107_2219: cvta.to.local.u64 %rd2833, %rd312; ld.local.v4.f32 {%f17095, %f17096, %f17097, %f17098}, [%rd2833]; cvta.to.local.u64 %rd2834, %rd327; ld.local.v4.f32 {%f17102, %f17103, %f17104, %f17105}, [%rd2834]; cvta.to.local.u64 %rd2835, %rd326; ld.local.v4.f32 {%f17109, %f17110, %f17111, %f17112}, [%rd2835]; mul.ftz.f32 %f17116, %f29754, %f17102; fma.rn.ftz.f32 %f17117, %f29753, %f17095, %f17116; fma.rn.ftz.f32 %f17118, %f29755, %f17109, %f17117; mul.ftz.f32 %f17119, %f29754, %f17103; fma.rn.ftz.f32 %f17120, %f29753, %f17096, %f17119; fma.rn.ftz.f32 %f17121, %f29755, %f17110, %f17120; mul.ftz.f32 %f17122, %f29754, %f17104; fma.rn.ftz.f32 %f17123, %f29753, %f17097, %f17122; fma.rn.ftz.f32 %f17124, %f29755, %f17111, %f17123; mul.ftz.f32 %f17125, %f17121, %f17121; fma.rn.ftz.f32 %f17126, %f17118, %f17118, %f17125; fma.rn.ftz.f32 %f17127, %f17124, %f17124, %f17126; sqrt.approx.ftz.f32 %f17128, %f17127; div.approx.ftz.f32 %f29753, %f17118, %f17128; div.approx.ftz.f32 %f29754, %f17121, %f17128; div.approx.ftz.f32 %f29755, %f17124, %f17128; $L__BB107_2221: @%p1975 bra $L__BB107_2223; bra.uni $L__BB107_2222; $L__BB107_2223: ld.global.v4.f32 {%f17197, %f17198, %f17199, %f17200}, [%rd436+48]; ld.global.v4.f32 {%f17204, %f17205, %f17206, %f17207}, [%rd436+64]; ld.global.v4.f32 {%f17211, %f17212, %f17213, %f17214}, [%rd436+80]; mul.ftz.f32 %f17218, %f29757, %f17204; fma.rn.ftz.f32 %f17219, %f29756, %f17197, %f17218; fma.rn.ftz.f32 %f17220, %f29758, %f17211, %f17219; mul.ftz.f32 %f17221, %f29757, %f17205; fma.rn.ftz.f32 %f17222, %f29756, %f17198, %f17221; fma.rn.ftz.f32 %f17223, %f29758, %f17212, %f17222; mul.ftz.f32 %f17224, %f29757, %f17206; fma.rn.ftz.f32 %f17225, %f29756, %f17199, %f17224; fma.rn.ftz.f32 %f17226, %f29758, %f17213, %f17225; mul.ftz.f32 %f17227, %f17223, %f17223; fma.rn.ftz.f32 %f17228, %f17220, %f17220, %f17227; fma.rn.ftz.f32 %f17229, %f17226, %f17226, %f17228; sqrt.approx.ftz.f32 %f17230, %f17229; div.approx.ftz.f32 %f29756, %f17220, %f17230; div.approx.ftz.f32 %f29757, %f17223, %f17230; div.approx.ftz.f32 %f29758, %f17226, %f17230; mov.f32 %f29759, %f29768; mov.f32 %f29760, %f29769; mov.f32 %f29762, %f29771; mov.f32 %f29763, %f29772; mov.f32 %f29765, %f29774; mov.f32 %f29766, %f29775; bra.uni $L__BB107_2224; $L__BB107_2222: cvta.to.local.u64 %rd2836, %rd312; ld.local.v4.f32 {%f17163, %f17164, %f17165, %f17166}, [%rd2836]; cvta.to.local.u64 %rd2837, %rd327; ld.local.v4.f32 {%f17170, %f17171, %f17172, %f17173}, [%rd2837]; cvta.to.local.u64 %rd2838, %rd326; ld.local.v4.f32 {%f17177, %f17178, %f17179, %f17180}, [%rd2838]; mul.ftz.f32 %f17184, %f29757, %f17170; fma.rn.ftz.f32 %f17185, %f29756, %f17163, %f17184; fma.rn.ftz.f32 %f17186, %f29758, %f17177, %f17185; mul.ftz.f32 %f17187, %f29757, %f17171; fma.rn.ftz.f32 %f17188, %f29756, %f17164, %f17187; fma.rn.ftz.f32 %f17189, %f29758, %f17178, %f17188; mul.ftz.f32 %f17190, %f29757, %f17172; fma.rn.ftz.f32 %f17191, %f29756, %f17165, %f17190; fma.rn.ftz.f32 %f17192, %f29758, %f17179, %f17191; mul.ftz.f32 %f17193, %f17189, %f17189; fma.rn.ftz.f32 %f17194, %f17186, %f17186, %f17193; fma.rn.ftz.f32 %f17195, %f17192, %f17192, %f17194; sqrt.approx.ftz.f32 %f17196, %f17195; div.approx.ftz.f32 %f29756, %f17186, %f17196; div.approx.ftz.f32 %f29757, %f17189, %f17196; div.approx.ftz.f32 %f29758, %f17192, %f17196; mov.f32 %f29759, %f29768; mov.f32 %f29760, %f29769; mov.f32 %f29762, %f29771; mov.f32 %f29763, %f29772; mov.f32 %f29765, %f29774; mov.f32 %f29766, %f29775; $L__BB107_2224: sub.ftz.f32 %f3182, %f29767, %f29761; sub.ftz.f32 %f3183, %f29763, %f29760; mul.ftz.f32 %f17231, %f3183, %f3182; sub.ftz.f32 %f3184, %f29766, %f29760; sub.ftz.f32 %f3185, %f29764, %f29761; mul.ftz.f32 %f17232, %f3185, %f3184; sub.ftz.f32 %f17233, %f17231, %f17232; sub.ftz.f32 %f3186, %f29765, %f29759; mul.ftz.f32 %f17234, %f3185, %f3186; sub.ftz.f32 %f3187, %f29762, %f29759; mul.ftz.f32 %f17235, %f3187, %f3182; sub.ftz.f32 %f17236, %f17234, %f17235; mul.ftz.f32 %f17237, %f3187, %f3184; mul.ftz.f32 %f17238, %f3183, %f3186; sub.ftz.f32 %f17239, %f17237, %f17238; mul.ftz.f32 %f17240, %f17236, %f17236; fma.rn.ftz.f32 %f17241, %f17233, %f17233, %f17240; fma.rn.ftz.f32 %f17242, %f17239, %f17239, %f17241; sqrt.approx.ftz.f32 %f17243, %f17242; div.approx.ftz.f32 %f29803, %f17239, %f17243; div.approx.ftz.f32 %f29802, %f17236, %f17243; div.approx.ftz.f32 %f29801, %f17233, %f17243; st.local.v2.f32 [%rd433+48], {%f29801, %f29802}; st.local.f32 [%rd433+56], %f29803; and.b32 %r3353, %r915, 8; setp.eq.s32 %p2014, %r3353, 0; @%p2014 bra $L__BB107_2226; ld.local.v4.f32 {%f17244, %f17245, %f17246, %f17247}, [%rd433+48]; neg.ftz.f32 %f29803, %f17246; neg.ftz.f32 %f29802, %f17245; neg.ftz.f32 %f29801, %f17244; st.local.v2.f32 [%rd433+48], {%f29801, %f29802}; st.local.f32 [%rd433+56], %f29803; $L__BB107_2226: mov.f32 %f17251, 0f3F800000; sub.ftz.f32 %f17252, %f17251, %f2953; sub.ftz.f32 %f17253, %f17252, %f2954; mul.ftz.f32 %f17254, %f29753, %f2953; fma.rn.ftz.f32 %f17255, %f29750, %f17253, %f17254; mul.ftz.f32 %f17256, %f29754, %f2953; fma.rn.ftz.f32 %f17257, %f29751, %f17253, %f17256; mul.ftz.f32 %f17258, %f29755, %f2953; fma.rn.ftz.f32 %f17259, %f29752, %f17253, %f17258; fma.rn.ftz.f32 %f17260, %f29756, %f2954, %f17255; fma.rn.ftz.f32 %f17261, %f29757, %f2954, %f17257; fma.rn.ftz.f32 %f17262, %f29758, %f2954, %f17259; mul.ftz.f32 %f17263, %f17261, %f17261; fma.rn.ftz.f32 %f17264, %f17260, %f17260, %f17263; fma.rn.ftz.f32 %f17265, %f17262, %f17262, %f17264; rsqrt.approx.ftz.f32 %f17266, %f17265; sub.ftz.f32 %f17267, %f29753, %f29750; mul.ftz.f32 %f17268, %f17267, %f17266; sub.ftz.f32 %f17269, %f29754, %f29751; mul.ftz.f32 %f17270, %f17269, %f17266; sub.ftz.f32 %f17271, %f29755, %f29752; mul.ftz.f32 %f17272, %f17271, %f17266; sub.ftz.f32 %f17273, %f29756, %f29750; mul.ftz.f32 %f17274, %f17273, %f17266; sub.ftz.f32 %f17275, %f29757, %f29751; mul.ftz.f32 %f17276, %f17275, %f17266; sub.ftz.f32 %f17277, %f29758, %f29752; mul.ftz.f32 %f17278, %f17277, %f17266; mul.ftz.f32 %f17279, %f17261, %f17266; mul.ftz.f32 %f17280, %f17270, %f17279; mul.ftz.f32 %f17281, %f17260, %f17266; fma.rn.ftz.f32 %f17282, %f17268, %f17281, %f17280; mul.ftz.f32 %f17283, %f17262, %f17266; fma.rn.ftz.f32 %f17284, %f17272, %f17283, %f17282; mul.ftz.f32 %f17285, %f17281, %f17284; mul.ftz.f32 %f17286, %f17279, %f17284; mul.ftz.f32 %f17287, %f17284, %f17283; sub.ftz.f32 %f17288, %f17268, %f17285; sub.ftz.f32 %f17289, %f17270, %f17286; sub.ftz.f32 %f17290, %f17272, %f17287; mul.ftz.f32 %f17291, %f17276, %f17279; fma.rn.ftz.f32 %f17292, %f17274, %f17281, %f17291; fma.rn.ftz.f32 %f17293, %f17278, %f17283, %f17292; mul.ftz.f32 %f17294, %f17281, %f17293; mul.ftz.f32 %f17295, %f17279, %f17293; mul.ftz.f32 %f17296, %f17293, %f17283; sub.ftz.f32 %f17297, %f17274, %f17294; sub.ftz.f32 %f17298, %f17276, %f17295; sub.ftz.f32 %f17299, %f17278, %f17296; mul.ftz.f32 %f17300, %f3183, %f3183; fma.rn.ftz.f32 %f17301, %f3187, %f3187, %f17300; fma.rn.ftz.f32 %f17302, %f3185, %f3185, %f17301; rsqrt.approx.ftz.f32 %f17303, %f17302; mul.ftz.f32 %f17304, %f3187, %f17303; mul.ftz.f32 %f17305, %f3183, %f17303; mul.ftz.f32 %f17306, %f3185, %f17303; mul.ftz.f32 %f17307, %f17288, %f17303; mul.ftz.f32 %f17308, %f17289, %f17303; mul.ftz.f32 %f17309, %f17290, %f17303; mul.ftz.f32 %f17310, %f3184, %f17305; fma.rn.ftz.f32 %f17311, %f3186, %f17304, %f17310; fma.rn.ftz.f32 %f17312, %f3182, %f17306, %f17311; mul.ftz.f32 %f17313, %f17304, %f17312; mul.ftz.f32 %f17314, %f17305, %f17312; mul.ftz.f32 %f17315, %f17306, %f17312; sub.ftz.f32 %f17316, %f3186, %f17313; sub.ftz.f32 %f17317, %f3184, %f17314; sub.ftz.f32 %f17318, %f3182, %f17315; mul.ftz.f32 %f17319, %f17307, %f17312; mul.ftz.f32 %f17320, %f17308, %f17312; mul.ftz.f32 %f17321, %f17309, %f17312; sub.ftz.f32 %f17322, %f17297, %f17319; sub.ftz.f32 %f17323, %f17298, %f17320; sub.ftz.f32 %f17324, %f17299, %f17321; mul.ftz.f32 %f17325, %f17317, %f17317; fma.rn.ftz.f32 %f17326, %f17316, %f17316, %f17325; fma.rn.ftz.f32 %f17327, %f17318, %f17318, %f17326; rsqrt.approx.ftz.f32 %f17328, %f17327; mul.ftz.f32 %f17329, %f17328, %f17316; mul.ftz.f32 %f17330, %f17328, %f17317; mul.ftz.f32 %f17331, %f17328, %f17318; mul.ftz.f32 %f17332, %f17328, %f17322; mul.ftz.f32 %f17333, %f17328, %f17323; mul.ftz.f32 %f17334, %f17328, %f17324; setp.neu.ftz.f32 %p2016, %f29801, %f29802; setp.neu.ftz.f32 %p2017, %f29801, %f29803; or.pred %p2018, %p2016, %p2017; sub.ftz.f32 %f17335, %f29803, %f29802; neg.ftz.f32 %f17336, %f29802; neg.ftz.f32 %f17337, %f29803; selp.f32 %f17338, %f17337, %f29803, %p2018; add.ftz.f32 %f17339, %f29801, %f17338; selp.f32 %f17340, %f29802, %f17336, %p2018; sub.ftz.f32 %f17341, %f17340, %f29801; mul.ftz.f32 %f17342, %f17339, %f17339; fma.rn.ftz.f32 %f17343, %f17335, %f17335, %f17342; fma.rn.ftz.f32 %f17344, %f17341, %f17341, %f17343; sqrt.approx.ftz.f32 %f17345, %f17344; div.approx.ftz.f32 %f17346, %f17341, %f17345; mul.ftz.f32 %f17347, %f29802, %f17346; div.approx.ftz.f32 %f17348, %f17339, %f17345; mul.ftz.f32 %f17349, %f29803, %f17348; div.approx.ftz.f32 %f17350, %f17335, %f17345; mul.ftz.f32 %f17351, %f29803, %f17350; mul.ftz.f32 %f17352, %f29801, %f17346; mul.ftz.f32 %f17353, %f29801, %f17348; mul.ftz.f32 %f17354, %f29802, %f17350; sub.ftz.f32 %f17355, %f17351, %f17352; sub.ftz.f32 %f17356, %f17347, %f17349; st.local.v4.f32 [%rd433+16], {%f17348, %f17346, %f17356, %f17355}; sub.ftz.f32 %f17357, %f17353, %f17354; st.local.v4.f32 [%rd433+32], {%f17357, %f17281, %f17279, %f17283}; mul.ftz.f32 %f17358, %f17305, %f17348; fma.rn.ftz.f32 %f17359, %f17304, %f17350, %f17358; fma.rn.ftz.f32 %f17360, %f17306, %f17346, %f17359; mul.ftz.f32 %f17361, %f17330, %f17348; fma.rn.ftz.f32 %f17362, %f17329, %f17350, %f17361; fma.rn.ftz.f32 %f17363, %f17331, %f17346, %f17362; mul.ftz.f32 %f17364, %f17332, %f17363; mul.ftz.f32 %f17365, %f17333, %f17363; mul.ftz.f32 %f17366, %f17334, %f17363; fma.rn.ftz.f32 %f17367, %f17307, %f17360, %f17364; fma.rn.ftz.f32 %f17368, %f17308, %f17360, %f17365; fma.rn.ftz.f32 %f17369, %f17309, %f17360, %f17366; st.local.f32 [%rd433+60], %f17367; st.local.f32 [%rd433+64], %f17368; st.local.f32 [%rd433+68], %f17369; mul.ftz.f32 %f17370, %f17307, %f17363; mul.ftz.f32 %f17371, %f17308, %f17363; mul.ftz.f32 %f17372, %f17309, %f17363; mul.ftz.f32 %f17373, %f17332, %f17360; mul.ftz.f32 %f17374, %f17333, %f17360; mul.ftz.f32 %f17375, %f17334, %f17360; sub.ftz.f32 %f17376, %f17375, %f17372; sub.ftz.f32 %f17377, %f17374, %f17371; sub.ftz.f32 %f17378, %f17373, %f17370; st.local.v2.f32 [%rd433+72], {%f17378, %f17377}; st.local.f32 [%rd433+80], %f17376; ld.local.v4.f32 {%f17379, %f17380, %f17381, %f17382}, [%rd2173]; st.local.v4.f32 [%rd433], {%f17379, %f17380, %f17381, %f17350}; mov.f32 %f17386, 0f00000000; st.local.v4.f32 [%rd433+176], {%f17386, %f17386, %f17386, %f17386}; st.local.v4.f32 [%rd433+144], {%f17386, %f17386, %f17386, %f17386}; st.local.v4.f32 [%rd433+160], {%f17251, %f17386, %f17386, %f17251}; st.local.v2.f32 [%rd433+128], {%f17386, %f17386}; st.local.v2.f32 [%rd433+120], {%f16444, %f16445}; st.local.u64 [%rd433+112], %rd434; st.local.v2.f32 [%rd433+88], {%f2953, %f2954}; cvt.u32.u64 %r3354, %rd438; cvt.u32.u64 %r3355, %rd435; st.local.v4.u32 [%rd433+96], {%r3355, %r3354, %r917, %r911}; sub.ftz.f32 %f17387, %f17379, %f16218; sub.ftz.f32 %f17388, %f17380, %f16219; sub.ftz.f32 %f17389, %f17381, %f16220; mul.ftz.f32 %f17390, %f17388, %f17388; fma.rn.ftz.f32 %f17391, %f17387, %f17387, %f17390; fma.rn.ftz.f32 %f17392, %f17389, %f17389, %f17391; sqrt.approx.ftz.f32 %f17393, %f17392; setp.lt.ftz.f32 %p2019, %f17393, 0f38D1B717; add.s32 %r5741, %r3047, 1; mov.pred %p3921, -1; @%p2019 bra $L__BB107_2250; setp.lt.s32 %p2020, %r5741, %r5732; mov.u16 %rs713, %rs351; @%p2020 bra $L__BB107_1810; $L__BB107_2228: setp.ne.s16 %p2021, %rs713, 0; or.pred %p2023, %p2021, %p1589; @%p2023 bra $L__BB107_2238; ld.local.f32 %f3197, [%rd280+-24]; ld.local.f32 %f3198, [%rd280+-20]; ld.local.f32 %f3199, [%rd280+-16]; mov.u32 %r5775, 0; $L__BB107_2230: cvt.s64.s32 %rd556, %r5775; setp.eq.s32 %p2024, %r5775, 0; mov.u64 %rd4834, %rd2; @%p2024 bra $L__BB107_2232; cvt.u32.u64 %r3358, %rd556; add.s32 %r3359, %r3358, -1; mul.wide.s32 %rd2844, %r3359, 192; add.s64 %rd4834, %rd2157, %rd2844; $L__BB107_2232: mul.lo.s64 %rd2845, %rd556, 192; add.s64 %rd559, %rd2157, %rd2845; ld.local.v4.f32 {%f17394, %f17395, %f17396, %f17397}, [%rd559]; ld.local.f32 %f3203, [%rd4834]; ld.local.f32 %f3204, [%rd4834+4]; ld.local.f32 %f3205, [%rd4834+8]; cvt.u32.u64 %r3360, %rd556; cvt.u32.u64 %r3361, %rd408; setp.eq.s32 %p2025, %r3360, %r3361; @%p2025 bra $L__BB107_2234; bra.uni $L__BB107_2233; $L__BB107_2234: mov.f32 %f29804, %f30148; mov.f32 %f29805, %f30147; mov.f32 %f29806, %f30146; @%p1261 bra $L__BB107_2236; sub.ftz.f32 %f29804, %f3197, %f17394; sub.ftz.f32 %f29805, %f3198, %f17395; sub.ftz.f32 %f29806, %f3199, %f17396; bra.uni $L__BB107_2236; $L__BB107_2233: ld.local.v4.f32 {%f17398, %f17399, %f17400, %f17401}, [%rd559+192]; sub.ftz.f32 %f29804, %f17398, %f17394; sub.ftz.f32 %f29805, %f17399, %f17395; sub.ftz.f32 %f29806, %f17400, %f17396; $L__BB107_2236: ld.local.f32 %f17405, [%rd559+36]; sub.ftz.f32 %f17406, %f3203, %f17394; ld.local.f32 %f17407, [%rd559+40]; sub.ftz.f32 %f17408, %f3204, %f17395; mul.ftz.f32 %f17409, %f17408, %f17407; fma.rn.ftz.f32 %f17410, %f17406, %f17405, %f17409; ld.local.f32 %f17411, [%rd559+44]; sub.ftz.f32 %f17412, %f3205, %f17396; fma.rn.ftz.f32 %f17413, %f17412, %f17411, %f17410; mul.ftz.f32 %f17414, %f29805, %f17407; fma.rn.ftz.f32 %f17415, %f29804, %f17405, %f17414; fma.rn.ftz.f32 %f17416, %f29806, %f17411, %f17415; mul.ftz.f32 %f17417, %f17413, %f17416; setp.ge.ftz.f32 %p2027, %f17417, 0f00000000; add.s32 %r5775, %r3360, 1; mov.u16 %rs713, 1; @%p2027 bra $L__BB107_2238; setp.lt.s32 %p2028, %r5775, %r5732; mov.u16 %rs713, 0; @%p2028 bra $L__BB107_2230; $L__BB107_2238: setp.eq.s16 %p2029, %rs713, 0; @%p2029 bra $L__BB107_2240; mul.ftz.f32 %f29683, %f29683, 0f3F000000; setp.lt.ftz.f32 %p2031, %f29683, 0f38D1B717; mov.u16 %rs628, 0; mov.pred %p3921, -1; @%p2031 bra $L__BB107_2250; bra.uni $L__BB107_2249; $L__BB107_2240: @%p1589 bra $L__BB107_2248; cvt.u32.u64 %r3364, %rd408; setp.lt.u32 %p2033, %r3364, 3; mov.u32 %r5778, 0; @%p2033 bra $L__BB107_2244; mov.u32 %r5778, 0; mov.u32 %r5777, %r808; $L__BB107_2243: mul.wide.s32 %rd2847, %r5778, 192; add.s64 %rd2848, %rd1, %rd2847; add.s64 %rd2849, %rd2157, %rd2847; ld.local.v4.u32 {%r3366, %r3367, %r3368, %r3369}, [%rd2849]; ld.local.v4.u32 {%r3374, %r3375, %r3376, %r3377}, [%rd2849+16]; ld.local.v4.u32 {%r3382, %r3383, %r3384, %r3385}, [%rd2849+32]; ld.local.v4.u32 {%r3390, %r3391, %r3392, %r3393}, [%rd2849+48]; ld.local.v4.u32 {%r3398, %r3399, %r3400, %r3401}, [%rd2849+64]; ld.local.v4.u32 {%r3406, %r3407, %r3408, %r3409}, [%rd2849+80]; ld.local.v4.u32 {%r3414, %r3415, %r3416, %r3417}, [%rd2849+96]; ld.local.v4.u32 {%r3422, %r3423, %r3424, %r3425}, [%rd2849+112]; ld.local.v4.u32 {%r3430, %r3431, %r3432, %r3433}, [%rd2849+128]; ld.local.v4.u32 {%r3438, %r3439, %r3440, %r3441}, [%rd2849+144]; ld.local.v4.u32 {%r3446, %r3447, %r3448, %r3449}, [%rd2849+160]; ld.local.v4.u32 {%r3454, %r3455, %r3456, %r3457}, [%rd2849+176]; st.local.v4.u32 [%rd2848], {%r3366, %r3367, %r3368, %r3369}; st.local.v4.u32 [%rd2848+16], {%r3374, %r3375, %r3376, %r3377}; st.local.v4.u32 [%rd2848+32], {%r3382, %r3383, %r3384, %r3385}; st.local.v4.u32 [%rd2848+48], {%r3390, %r3391, %r3392, %r3393}; st.local.v4.u32 [%rd2848+64], {%r3398, %r3399, %r3400, %r3401}; st.local.v4.u32 [%rd2848+80], {%r3406, %r3407, %r3408, %r3409}; st.local.v4.u32 [%rd2848+96], {%r3414, %r3415, %r3416, %r3417}; st.local.v4.u32 [%rd2848+112], {%r3422, %r3423, %r3424, %r3425}; st.local.v4.u32 [%rd2848+128], {%r3430, %r3431, %r3432, %r3433}; st.local.v4.u32 [%rd2848+144], {%r3438, %r3439, %r3440, %r3441}; st.local.v4.u32 [%rd2848+160], {%r3446, %r3447, %r3448, %r3449}; st.local.v4.u32 [%rd2848+176], {%r3454, %r3455, %r3456, %r3457}; ld.local.v4.u32 {%r3462, %r3463, %r3464, %r3465}, [%rd2849+192]; ld.local.v4.u32 {%r3470, %r3471, %r3472, %r3473}, [%rd2849+208]; ld.local.v4.u32 {%r3478, %r3479, %r3480, %r3481}, [%rd2849+224]; ld.local.v4.u32 {%r3486, %r3487, %r3488, %r3489}, [%rd2849+240]; ld.local.v4.u32 {%r3494, %r3495, %r3496, %r3497}, [%rd2849+256]; ld.local.v4.u32 {%r3502, %r3503, %r3504, %r3505}, [%rd2849+272]; ld.local.v4.u32 {%r3510, %r3511, %r3512, %r3513}, [%rd2849+288]; ld.local.v4.u32 {%r3518, %r3519, %r3520, %r3521}, [%rd2849+304]; ld.local.v4.u32 {%r3526, %r3527, %r3528, %r3529}, [%rd2849+320]; ld.local.v4.u32 {%r3534, %r3535, %r3536, %r3537}, [%rd2849+336]; ld.local.v4.u32 {%r3542, %r3543, %r3544, %r3545}, [%rd2849+352]; ld.local.v4.u32 {%r3550, %r3551, %r3552, %r3553}, [%rd2849+368]; st.local.v4.u32 [%rd2848+192], {%r3462, %r3463, %r3464, %r3465}; st.local.v4.u32 [%rd2848+208], {%r3470, %r3471, %r3472, %r3473}; st.local.v4.u32 [%rd2848+224], {%r3478, %r3479, %r3480, %r3481}; st.local.v4.u32 [%rd2848+240], {%r3486, %r3487, %r3488, %r3489}; st.local.v4.u32 [%rd2848+256], {%r3494, %r3495, %r3496, %r3497}; st.local.v4.u32 [%rd2848+272], {%r3502, %r3503, %r3504, %r3505}; st.local.v4.u32 [%rd2848+288], {%r3510, %r3511, %r3512, %r3513}; st.local.v4.u32 [%rd2848+304], {%r3518, %r3519, %r3520, %r3521}; st.local.v4.u32 [%rd2848+320], {%r3526, %r3527, %r3528, %r3529}; st.local.v4.u32 [%rd2848+336], {%r3534, %r3535, %r3536, %r3537}; st.local.v4.u32 [%rd2848+352], {%r3542, %r3543, %r3544, %r3545}; st.local.v4.u32 [%rd2848+368], {%r3550, %r3551, %r3552, %r3553}; ld.local.v4.u32 {%r3558, %r3559, %r3560, %r3561}, [%rd2849+384]; ld.local.v4.u32 {%r3566, %r3567, %r3568, %r3569}, [%rd2849+400]; ld.local.v4.u32 {%r3574, %r3575, %r3576, %r3577}, [%rd2849+416]; ld.local.v4.u32 {%r3582, %r3583, %r3584, %r3585}, [%rd2849+432]; ld.local.v4.u32 {%r3590, %r3591, %r3592, %r3593}, [%rd2849+448]; ld.local.v4.u32 {%r3598, %r3599, %r3600, %r3601}, [%rd2849+464]; ld.local.v4.u32 {%r3606, %r3607, %r3608, %r3609}, [%rd2849+480]; ld.local.v4.u32 {%r3614, %r3615, %r3616, %r3617}, [%rd2849+496]; ld.local.v4.u32 {%r3622, %r3623, %r3624, %r3625}, [%rd2849+512]; ld.local.v4.u32 {%r3630, %r3631, %r3632, %r3633}, [%rd2849+528]; ld.local.v4.u32 {%r3638, %r3639, %r3640, %r3641}, [%rd2849+544]; ld.local.v4.u32 {%r3646, %r3647, %r3648, %r3649}, [%rd2849+560]; st.local.v4.u32 [%rd2848+384], {%r3558, %r3559, %r3560, %r3561}; st.local.v4.u32 [%rd2848+400], {%r3566, %r3567, %r3568, %r3569}; st.local.v4.u32 [%rd2848+416], {%r3574, %r3575, %r3576, %r3577}; st.local.v4.u32 [%rd2848+432], {%r3582, %r3583, %r3584, %r3585}; st.local.v4.u32 [%rd2848+448], {%r3590, %r3591, %r3592, %r3593}; st.local.v4.u32 [%rd2848+464], {%r3598, %r3599, %r3600, %r3601}; st.local.v4.u32 [%rd2848+480], {%r3606, %r3607, %r3608, %r3609}; st.local.v4.u32 [%rd2848+496], {%r3614, %r3615, %r3616, %r3617}; st.local.v4.u32 [%rd2848+512], {%r3622, %r3623, %r3624, %r3625}; st.local.v4.u32 [%rd2848+528], {%r3630, %r3631, %r3632, %r3633}; st.local.v4.u32 [%rd2848+544], {%r3638, %r3639, %r3640, %r3641}; st.local.v4.u32 [%rd2848+560], {%r3646, %r3647, %r3648, %r3649}; ld.local.v4.u32 {%r3654, %r3655, %r3656, %r3657}, [%rd2849+576]; ld.local.v4.u32 {%r3662, %r3663, %r3664, %r3665}, [%rd2849+592]; ld.local.v4.u32 {%r3670, %r3671, %r3672, %r3673}, [%rd2849+608]; ld.local.v4.u32 {%r3678, %r3679, %r3680, %r3681}, [%rd2849+624]; ld.local.v4.u32 {%r3686, %r3687, %r3688, %r3689}, [%rd2849+640]; ld.local.v4.u32 {%r3694, %r3695, %r3696, %r3697}, [%rd2849+656]; ld.local.v4.u32 {%r3702, %r3703, %r3704, %r3705}, [%rd2849+672]; ld.local.v4.u32 {%r3710, %r3711, %r3712, %r3713}, [%rd2849+688]; ld.local.v4.u32 {%r3718, %r3719, %r3720, %r3721}, [%rd2849+704]; ld.local.v4.u32 {%r3726, %r3727, %r3728, %r3729}, [%rd2849+720]; ld.local.v4.u32 {%r3734, %r3735, %r3736, %r3737}, [%rd2849+736]; ld.local.v4.u32 {%r3742, %r3743, %r3744, %r3745}, [%rd2849+752]; st.local.v4.u32 [%rd2848+576], {%r3654, %r3655, %r3656, %r3657}; st.local.v4.u32 [%rd2848+592], {%r3662, %r3663, %r3664, %r3665}; st.local.v4.u32 [%rd2848+608], {%r3670, %r3671, %r3672, %r3673}; st.local.v4.u32 [%rd2848+624], {%r3678, %r3679, %r3680, %r3681}; st.local.v4.u32 [%rd2848+640], {%r3686, %r3687, %r3688, %r3689}; st.local.v4.u32 [%rd2848+656], {%r3694, %r3695, %r3696, %r3697}; st.local.v4.u32 [%rd2848+672], {%r3702, %r3703, %r3704, %r3705}; st.local.v4.u32 [%rd2848+688], {%r3710, %r3711, %r3712, %r3713}; st.local.v4.u32 [%rd2848+704], {%r3718, %r3719, %r3720, %r3721}; st.local.v4.u32 [%rd2848+720], {%r3726, %r3727, %r3728, %r3729}; st.local.v4.u32 [%rd2848+736], {%r3734, %r3735, %r3736, %r3737}; st.local.v4.u32 [%rd2848+752], {%r3742, %r3743, %r3744, %r3745}; add.s32 %r5778, %r5778, 4; add.s32 %r5777, %r5777, -4; setp.ne.s32 %p2034, %r5777, 0; @%p2034 bra $L__BB107_2243; $L__BB107_2244: setp.eq.s32 %p2035, %r807, 0; @%p2035 bra $L__BB107_2248; setp.eq.s32 %p2036, %r807, 1; mul.wide.s32 %rd2850, %r5778, 192; add.s64 %rd562, %rd1, %rd2850; add.s64 %rd561, %rd2157, %rd2850; ld.local.v4.u32 {%r3750, %r3751, %r3752, %r3753}, [%rd561]; ld.local.v4.u32 {%r3758, %r3759, %r3760, %r3761}, [%rd561+16]; ld.local.v4.u32 {%r3766, %r3767, %r3768, %r3769}, [%rd561+32]; ld.local.v4.u32 {%r3774, %r3775, %r3776, %r3777}, [%rd561+48]; ld.local.v4.u32 {%r3782, %r3783, %r3784, %r3785}, [%rd561+64]; ld.local.v4.u32 {%r3790, %r3791, %r3792, %r3793}, [%rd561+80]; ld.local.v4.u32 {%r3798, %r3799, %r3800, %r3801}, [%rd561+96]; ld.local.v4.u32 {%r3806, %r3807, %r3808, %r3809}, [%rd561+112]; ld.local.v4.u32 {%r3814, %r3815, %r3816, %r3817}, [%rd561+128]; ld.local.v4.u32 {%r3822, %r3823, %r3824, %r3825}, [%rd561+144]; ld.local.v4.u32 {%r3830, %r3831, %r3832, %r3833}, [%rd561+160]; ld.local.v4.u32 {%r3838, %r3839, %r3840, %r3841}, [%rd561+176]; st.local.v4.u32 [%rd562], {%r3750, %r3751, %r3752, %r3753}; st.local.v4.u32 [%rd562+16], {%r3758, %r3759, %r3760, %r3761}; st.local.v4.u32 [%rd562+32], {%r3766, %r3767, %r3768, %r3769}; st.local.v4.u32 [%rd562+48], {%r3774, %r3775, %r3776, %r3777}; st.local.v4.u32 [%rd562+64], {%r3782, %r3783, %r3784, %r3785}; st.local.v4.u32 [%rd562+80], {%r3790, %r3791, %r3792, %r3793}; st.local.v4.u32 [%rd562+96], {%r3798, %r3799, %r3800, %r3801}; st.local.v4.u32 [%rd562+112], {%r3806, %r3807, %r3808, %r3809}; st.local.v4.u32 [%rd562+128], {%r3814, %r3815, %r3816, %r3817}; st.local.v4.u32 [%rd562+144], {%r3822, %r3823, %r3824, %r3825}; st.local.v4.u32 [%rd562+160], {%r3830, %r3831, %r3832, %r3833}; st.local.v4.u32 [%rd562+176], {%r3838, %r3839, %r3840, %r3841}; @%p2036 bra $L__BB107_2248; setp.eq.s32 %p2037, %r807, 2; ld.local.v4.u32 {%r3846, %r3847, %r3848, %r3849}, [%rd561+192]; ld.local.v4.u32 {%r3854, %r3855, %r3856, %r3857}, [%rd561+208]; ld.local.v4.u32 {%r3862, %r3863, %r3864, %r3865}, [%rd561+224]; ld.local.v4.u32 {%r3870, %r3871, %r3872, %r3873}, [%rd561+240]; ld.local.v4.u32 {%r3878, %r3879, %r3880, %r3881}, [%rd561+256]; ld.local.v4.u32 {%r3886, %r3887, %r3888, %r3889}, [%rd561+272]; ld.local.v4.u32 {%r3894, %r3895, %r3896, %r3897}, [%rd561+288]; ld.local.v4.u32 {%r3902, %r3903, %r3904, %r3905}, [%rd561+304]; ld.local.v4.u32 {%r3910, %r3911, %r3912, %r3913}, [%rd561+320]; ld.local.v4.u32 {%r3918, %r3919, %r3920, %r3921}, [%rd561+336]; ld.local.v4.u32 {%r3926, %r3927, %r3928, %r3929}, [%rd561+352]; ld.local.v4.u32 {%r3934, %r3935, %r3936, %r3937}, [%rd561+368]; st.local.v4.u32 [%rd562+192], {%r3846, %r3847, %r3848, %r3849}; st.local.v4.u32 [%rd562+208], {%r3854, %r3855, %r3856, %r3857}; st.local.v4.u32 [%rd562+224], {%r3862, %r3863, %r3864, %r3865}; st.local.v4.u32 [%rd562+240], {%r3870, %r3871, %r3872, %r3873}; st.local.v4.u32 [%rd562+256], {%r3878, %r3879, %r3880, %r3881}; st.local.v4.u32 [%rd562+272], {%r3886, %r3887, %r3888, %r3889}; st.local.v4.u32 [%rd562+288], {%r3894, %r3895, %r3896, %r3897}; st.local.v4.u32 [%rd562+304], {%r3902, %r3903, %r3904, %r3905}; st.local.v4.u32 [%rd562+320], {%r3910, %r3911, %r3912, %r3913}; st.local.v4.u32 [%rd562+336], {%r3918, %r3919, %r3920, %r3921}; st.local.v4.u32 [%rd562+352], {%r3926, %r3927, %r3928, %r3929}; st.local.v4.u32 [%rd562+368], {%r3934, %r3935, %r3936, %r3937}; @%p2037 bra $L__BB107_2248; ld.local.v4.u32 {%r3942, %r3943, %r3944, %r3945}, [%rd561+384]; ld.local.v4.u32 {%r3950, %r3951, %r3952, %r3953}, [%rd561+400]; ld.local.v4.u32 {%r3958, %r3959, %r3960, %r3961}, [%rd561+416]; ld.local.v4.u32 {%r3966, %r3967, %r3968, %r3969}, [%rd561+432]; ld.local.v4.u32 {%r3974, %r3975, %r3976, %r3977}, [%rd561+448]; ld.local.v4.u32 {%r3982, %r3983, %r3984, %r3985}, [%rd561+464]; ld.local.v4.u32 {%r3990, %r3991, %r3992, %r3993}, [%rd561+480]; ld.local.v4.u32 {%r3998, %r3999, %r4000, %r4001}, [%rd561+496]; ld.local.v4.u32 {%r4006, %r4007, %r4008, %r4009}, [%rd561+512]; ld.local.v4.u32 {%r4014, %r4015, %r4016, %r4017}, [%rd561+528]; ld.local.v4.u32 {%r4022, %r4023, %r4024, %r4025}, [%rd561+544]; ld.local.v4.u32 {%r4030, %r4031, %r4032, %r4033}, [%rd561+560]; st.local.v4.u32 [%rd562+384], {%r3942, %r3943, %r3944, %r3945}; st.local.v4.u32 [%rd562+400], {%r3950, %r3951, %r3952, %r3953}; st.local.v4.u32 [%rd562+416], {%r3958, %r3959, %r3960, %r3961}; st.local.v4.u32 [%rd562+432], {%r3966, %r3967, %r3968, %r3969}; st.local.v4.u32 [%rd562+448], {%r3974, %r3975, %r3976, %r3977}; st.local.v4.u32 [%rd562+464], {%r3982, %r3983, %r3984, %r3985}; st.local.v4.u32 [%rd562+480], {%r3990, %r3991, %r3992, %r3993}; st.local.v4.u32 [%rd562+496], {%r3998, %r3999, %r4000, %r4001}; st.local.v4.u32 [%rd562+512], {%r4006, %r4007, %r4008, %r4009}; st.local.v4.u32 [%rd562+528], {%r4014, %r4015, %r4016, %r4017}; st.local.v4.u32 [%rd562+544], {%r4022, %r4023, %r4024, %r4025}; st.local.v4.u32 [%rd562+560], {%r4030, %r4031, %r4032, %r4033}; $L__BB107_2248: add.ftz.f32 %f17418, %f29683, %f29683; mov.f32 %f17419, 0f3F800000; min.ftz.f32 %f29683, %f17419, %f17418; mov.u16 %rs628, 1; $L__BB107_2249: add.s32 %r5733, %r5733, 1; setp.lt.u32 %p2039, %r5733, 64; mov.pred %p3921, -1; @%p2039 bra $L__BB107_1761; $L__BB107_2250: mov.u32 %r5802, 0; @%p3921 bra $L__BB107_3392; ld.local.v4.f32 {%f17430, %f30255, %f30253, %f17433}, [%rd1]; ld.local.v4.f32 {%f17435, %f17436, %f17437, %f17438}, [%rd279]; sub.ftz.f32 %f17442, %f17430, %f17435; sub.ftz.f32 %f17443, %f30255, %f17436; sub.ftz.f32 %f17444, %f30253, %f17437; mul.ftz.f32 %f17445, %f17443, %f17443; fma.rn.ftz.f32 %f17446, %f17442, %f17442, %f17445; fma.rn.ftz.f32 %f17447, %f17444, %f17444, %f17446; sqrt.approx.ftz.f32 %f30257, %f17447; rsqrt.approx.ftz.f32 %f17448, %f17447; mul.ftz.f32 %f30260, %f17442, %f17448; mul.ftz.f32 %f30259, %f17443, %f17448; mul.ftz.f32 %f30258, %f17444, %f17448; ld.local.u32 %r956, [%rd280+40]; mov.f32 %f29808, 0f00000000; mov.u32 %r5802, 0; st.local.u32 [%rd283], %r5802; st.local.u32 [%rd283+4], %r5802; st.local.u32 [%rd283+8], %r5802; st.local.u32 [%rd283+12], %r5802; st.local.u32 [%rd283+16], %r5802; st.local.u32 [%rd283+20], %r5802; st.local.u32 [%rd283+24], %r5802; st.local.u32 [%rd283+28], %r5802; st.local.u32 [%rd283+32], %r5802; ld.local.u32 %r957, [%rd5+172]; setp.lt.s32 %p2040, %r957, 1; mov.f32 %f29809, %f29808; mov.f32 %f29810, %f29808; mov.f32 %f30127, %f29808; mov.f32 %f30128, %f29808; mov.f32 %f30129, %f29808; mov.f32 %f30124, %f29808; mov.f32 %f30125, %f29808; mov.f32 %f30126, %f29808; @%p2040 bra $L__BB107_2796; and.b32 %r958, %r956, 264241152; and.b32 %r959, %r956, 134217728; and.b32 %r960, %r956, 67108864; and.b32 %r961, %r956, 33554432; ld.const.v4.f32 {%f17458, %f17459, %f17460, %f17461}, [kernel_params+1184]; mov.u32 %r5779, 0; mov.f32 %f29808, 0f00000000; setp.eq.s32 %p2042, %r958, 0; mov.u16 %rs519, 0; mov.u16 %rs534, 1; mov.pred %p2411, -1; mov.pred %p2422, 0; mov.f32 %f29809, %f29808; mov.f32 %f29810, %f29808; mov.f32 %f30124, %f29808; mov.f32 %f30125, %f29808; mov.f32 %f30126, %f29808; mov.f32 %f30127, %f29808; mov.f32 %f30128, %f29808; mov.f32 %f30129, %f29808; bra.uni $L__BB107_2253; $L__BB107_960: mad.lo.s32 %r2248, %r2234, 3, %r478; ld.const.u64 %rd1947, [kernel_params+2336]; cvta.to.global.u64 %rd1948, %rd1947; mul.wide.s32 %rd1949, %r2248, 12; add.s64 %rd4776, %rd1948, %rd1949; add.s64 %rd4777, %rd4776, 4; add.s64 %rd4778, %rd4776, 8; add.s64 %rd4773, %rd4776, 12; add.s64 %rd4774, %rd4776, 16; add.s64 %rd4775, %rd4776, 20; add.s64 %rd4770, %rd4776, 24; add.s64 %rd4771, %rd4776, 28; add.s64 %rd4772, %rd4776, 32; $L__BB107_961: ld.local.f32 %f10626, [%rd5+-20]; ld.global.f32 %f10627, [%rd4773]; ld.global.f32 %f10628, [%rd4774]; ld.global.f32 %f10629, [%rd4775]; ld.local.f32 %f10630, [%rd5+-16]; ld.global.f32 %f10631, [%rd4770]; mul.ftz.f32 %f10632, %f10631, %f10630; ld.global.f32 %f10633, [%rd4771]; mul.ftz.f32 %f10634, %f10633, %f10630; ld.global.f32 %f10635, [%rd4772]; mul.ftz.f32 %f10636, %f10635, %f10630; fma.rn.ftz.f32 %f10637, %f10627, %f10626, %f10632; fma.rn.ftz.f32 %f10638, %f10628, %f10626, %f10634; fma.rn.ftz.f32 %f10639, %f10629, %f10626, %f10636; mov.f32 %f10640, 0f3F800000; sub.ftz.f32 %f10641, %f10640, %f10626; sub.ftz.f32 %f10642, %f10641, %f10630; ld.global.f32 %f10643, [%rd4776]; ld.global.f32 %f10644, [%rd4777]; ld.global.f32 %f10645, [%rd4778]; fma.rn.ftz.f32 %f29239, %f10645, %f10642, %f10639; fma.rn.ftz.f32 %f29238, %f10644, %f10642, %f10638; fma.rn.ftz.f32 %f29237, %f10643, %f10642, %f10637; $L__BB107_964: add.s32 %r486, %r481, %r478; @%p828 bra $L__BB107_969; and.b32 %r2253, %r475, 24; setp.eq.s32 %p833, %r2253, 0; @%p833 bra $L__BB107_967; ld.const.u64 %rd1954, [kernel_params+2232]; cvta.to.global.u64 %rd1955, %rd1954; shl.b64 %rd1956, %rd185, 4; add.s64 %rd1957, %rd1955, %rd1956; ld.global.v4.u32 {%r2254, %r2255, %r2256, %r2257}, [%rd1957]; ld.const.u64 %rd1958, [kernel_params+2336]; cvta.to.global.u64 %rd1959, %rd1958; add.s32 %r2261, %r2254, %r486; mul.wide.u32 %rd1960, %r2261, 12; add.s64 %rd4785, %rd1959, %rd1960; add.s32 %r2262, %r2255, %r486; mul.wide.u32 %rd1961, %r2262, 12; add.s64 %rd4782, %rd1959, %rd1961; add.s32 %r2263, %r2256, %r486; mul.wide.u32 %rd1962, %r2263, 12; add.s64 %rd4779, %rd1959, %rd1962; add.s64 %rd4786, %rd4785, 4; add.s64 %rd4787, %rd4785, 8; add.s64 %rd4783, %rd4782, 4; add.s64 %rd4784, %rd4782, 8; add.s64 %rd4780, %rd4779, 4; add.s64 %rd4781, %rd4779, 8; bra.uni $L__BB107_968; $L__BB107_969: and.b32 %r2266, %r475, 7; setp.eq.s32 %p834, %r2266, 0; mov.f32 %f29234, 0f00000000; mov.f32 %f29235, %f29234; mov.f32 %f29236, %f29234; @%p834 bra $L__BB107_986; setp.eq.s32 %p835, %r475, 4; selp.b32 %r2268, %r2234, 0, %p835; add.s32 %r2269, %r2268, %r486; ld.const.u64 %rd1966, [kernel_params+2336]; cvta.to.global.u64 %rd1967, %rd1966; mul.wide.s32 %rd1968, %r2269, 12; add.s64 %rd1969, %rd1967, %rd1968; ld.global.f32 %f29236, [%rd1969+8]; ld.global.f32 %f29235, [%rd1969+4]; ld.global.f32 %f29234, [%rd1969]; bra.uni $L__BB107_986; $L__BB107_967: mad.lo.s32 %r2265, %r2234, 3, %r486; ld.const.u64 %rd1963, [kernel_params+2336]; cvta.to.global.u64 %rd1964, %rd1963; mul.wide.s32 %rd1965, %r2265, 12; add.s64 %rd4785, %rd1964, %rd1965; add.s64 %rd4786, %rd4785, 4; add.s64 %rd4787, %rd4785, 8; add.s64 %rd4782, %rd4785, 12; add.s64 %rd4783, %rd4785, 16; add.s64 %rd4784, %rd4785, 20; add.s64 %rd4779, %rd4785, 24; add.s64 %rd4780, %rd4785, 28; add.s64 %rd4781, %rd4785, 32; $L__BB107_968: ld.local.f32 %f10649, [%rd5+-20]; ld.global.f32 %f10650, [%rd4782]; ld.global.f32 %f10651, [%rd4783]; ld.global.f32 %f10652, [%rd4784]; ld.local.f32 %f10653, [%rd5+-16]; ld.global.f32 %f10654, [%rd4779]; mul.ftz.f32 %f10655, %f10654, %f10653; ld.global.f32 %f10656, [%rd4780]; mul.ftz.f32 %f10657, %f10656, %f10653; ld.global.f32 %f10658, [%rd4781]; mul.ftz.f32 %f10659, %f10658, %f10653; fma.rn.ftz.f32 %f10660, %f10650, %f10649, %f10655; fma.rn.ftz.f32 %f10661, %f10651, %f10649, %f10657; fma.rn.ftz.f32 %f10662, %f10652, %f10649, %f10659; mov.f32 %f10663, 0f3F800000; sub.ftz.f32 %f10664, %f10663, %f10649; sub.ftz.f32 %f10665, %f10664, %f10653; ld.global.f32 %f10666, [%rd4785]; ld.global.f32 %f10667, [%rd4786]; ld.global.f32 %f10668, [%rd4787]; fma.rn.ftz.f32 %f29236, %f10668, %f10665, %f10662; fma.rn.ftz.f32 %f29235, %f10667, %f10665, %f10661; fma.rn.ftz.f32 %f29234, %f10666, %f10665, %f10660; $L__BB107_986: ld.const.u64 %rd1994, [kernel_params+2168]; cvta.to.global.u64 %rd1995, %rd1994; shl.b32 %r2288, %r5684, 1; mul.wide.s32 %rd1996, %r2288, 48; add.s64 %rd1997, %rd1995, %rd1996; ld.global.v4.f32 {%f10866, %f10867, %f10868, %f10869}, [%rd1997]; ld.global.v4.f32 {%f10874, %f10875, %f10876, %f10877}, [%rd1997+16]; ld.global.v4.f32 {%f10882, %f10883, %f10884, %f10885}, [%rd1997+32]; mul.ftz.f32 %f10890, %f29238, %f10867; fma.rn.ftz.f32 %f10891, %f29237, %f10866, %f10890; fma.rn.ftz.f32 %f10892, %f29239, %f10868, %f10891; add.ftz.f32 %f1375, %f10869, %f10892; mul.ftz.f32 %f10893, %f29238, %f10875; fma.rn.ftz.f32 %f10894, %f29237, %f10874, %f10893; fma.rn.ftz.f32 %f10895, %f29239, %f10876, %f10894; add.ftz.f32 %f1376, %f10877, %f10895; mul.ftz.f32 %f10896, %f29238, %f10883; fma.rn.ftz.f32 %f10897, %f29237, %f10882, %f10896; fma.rn.ftz.f32 %f10898, %f29239, %f10884, %f10897; add.ftz.f32 %f1377, %f10885, %f10898; ld.global.v4.f32 {%f10899, %f10900, %f10901, %f10902}, [%rd1997+48]; ld.global.v4.f32 {%f10907, %f10908, %f10909, %f10910}, [%rd1997+64]; ld.global.v4.f32 {%f10915, %f10916, %f10917, %f10918}, [%rd1997+80]; mul.ftz.f32 %f10923, %f29235, %f10900; fma.rn.ftz.f32 %f10924, %f29234, %f10899, %f10923; fma.rn.ftz.f32 %f10925, %f29236, %f10901, %f10924; add.ftz.f32 %f1378, %f10902, %f10925; mul.ftz.f32 %f10926, %f29235, %f10908; fma.rn.ftz.f32 %f10927, %f29234, %f10907, %f10926; fma.rn.ftz.f32 %f10928, %f29236, %f10909, %f10927; add.ftz.f32 %f1379, %f10910, %f10928; mul.ftz.f32 %f10929, %f29235, %f10916; fma.rn.ftz.f32 %f10930, %f29234, %f10915, %f10929; fma.rn.ftz.f32 %f10931, %f29236, %f10917, %f10930; add.ftz.f32 %f1380, %f10918, %f10931; ld.const.u32 %r2289, [kernel_params+16]; setp.eq.s32 %p845, %r2289, 2; @%p845 bra $L__BB107_993; bra.uni $L__BB107_987; $L__BB107_993: ld.const.v4.f32 {%f11073, %f11074, %f11075, %f11076}, [kernel_params+688]; ld.const.v4.f32 {%f11081, %f11082, %f11083, %f11084}, [kernel_params+704]; ld.const.v4.f32 {%f11089, %f11090, %f11091, %f11092}, [kernel_params+720]; mul.ftz.f32 %f11097, %f29217, %f11074; fma.rn.ftz.f32 %f11098, %f29216, %f11073, %f11097; fma.rn.ftz.f32 %f11099, %f29218, %f11075, %f11098; add.ftz.f32 %f11100, %f11076, %f11099; mul.ftz.f32 %f11101, %f29217, %f11082; fma.rn.ftz.f32 %f11102, %f29216, %f11081, %f11101; fma.rn.ftz.f32 %f11103, %f29218, %f11083, %f11102; add.ftz.f32 %f11104, %f11084, %f11103; mul.ftz.f32 %f11105, %f29217, %f11090; fma.rn.ftz.f32 %f11106, %f29216, %f11089, %f11105; fma.rn.ftz.f32 %f11107, %f29218, %f11091, %f11106; add.ftz.f32 %f11108, %f11092, %f11107; mul.ftz.f32 %f11109, %f11104, %f11104; fma.rn.ftz.f32 %f11110, %f11100, %f11100, %f11109; fma.rn.ftz.f32 %f11111, %f11108, %f11108, %f11110; sqrt.approx.ftz.f32 %f11112, %f11111; div.approx.ftz.f32 %f1417, %f11100, %f11112; div.approx.ftz.f32 %f1418, %f11104, %f11112; div.approx.ftz.f32 %f1419, %f11108, %f11112; ld.const.u32 %r493, [kernel_params+20]; mov.f32 %f11113, 0fBF800000; max.ftz.f32 %f11114, %f1417, %f11113; mov.f32 %f11115, 0f3F800000; min.ftz.f32 %f11116, %f11114, %f11115; abs.ftz.f32 %f11117, %f11116; sub.ftz.f32 %f11118, %f11115, %f11117; mul.ftz.f32 %f11119, %f11118, 0f3F000000; sqrt.approx.ftz.f32 %f11120, %f11119; setp.gt.ftz.f32 %p849, %f11117, 0f3F11EB85; selp.f32 %f11121, %f11120, %f11117, %p849; mul.ftz.f32 %f11122, %f11121, %f11121; mov.f32 %f11123, 0f3C94D2E9; mov.f32 %f11124, 0f3D53F941; fma.rn.ftz.f32 %f11125, %f11124, %f11122, %f11123; mov.f32 %f11126, 0f3D3F841F; fma.rn.ftz.f32 %f11127, %f11125, %f11122, %f11126; mov.f32 %f11128, 0f3D994929; fma.rn.ftz.f32 %f11129, %f11127, %f11122, %f11128; mov.f32 %f11130, 0f3E2AAB94; fma.rn.ftz.f32 %f11131, %f11129, %f11122, %f11130; mul.ftz.f32 %f11132, %f11122, %f11131; fma.rn.ftz.f32 %f11133, %f11132, %f11121, %f11121; add.ftz.f32 %f11134, %f11133, %f11133; mov.f32 %f11135, 0f3FC90FDB; sub.ftz.f32 %f11136, %f11135, %f11133; selp.f32 %f11137, %f11134, %f11136, %p849; setp.lt.ftz.f32 %p850, %f11116, 0f00000000; mov.f32 %f11138, 0f40490FDB; sub.ftz.f32 %f11139, %f11138, %f11137; selp.f32 %f1420, %f11139, %f11137, %p850; abs.ftz.f32 %f1421, %f1418; abs.ftz.f32 %f1422, %f1419; setp.eq.ftz.f32 %p851, %f1421, 0f00000000; setp.eq.ftz.f32 %p852, %f1422, 0f00000000; and.pred %p2, %p851, %p852; mov.b32 %r494, %f1418; mov.b32 %r2290, %f1419; and.b32 %r495, %r2290, -2147483648; setp.gt.s32 %p853, %r493, 2; @%p853 bra $L__BB107_999; bra.uni $L__BB107_994; $L__BB107_999: setp.eq.s32 %p854, %r493, 3; @%p854 bra $L__BB107_1008; setp.eq.s32 %p855, %r493, 5; @%p855 bra $L__BB107_1011; setp.ne.s32 %p856, %r493, 4; @%p856 bra $L__BB107_1037; ld.const.f32 %f11140, [kernel_params+48]; ld.const.v4.f32 {%f11141, %f11142, %f11143, %f11144}, [kernel_params+64]; neg.ftz.f32 %f11149, %f1420; sub.ftz.f32 %f11150, %f11149, %f11140; div.approx.ftz.f32 %f11151, %f11150, %f11141; mul.ftz.f32 %f11152, %f11151, %f11151; mul.ftz.f32 %f11153, %f11151, %f11152; mul.ftz.f32 %f11154, %f11152, %f11152; mul.ftz.f32 %f11155, %f11142, %f11152; fma.rn.ftz.f32 %f11156, %f11141, %f11151, %f11155; mul.ftz.f32 %f11157, %f11144, %f11154; fma.rn.ftz.f32 %f11158, %f11143, %f11153, %f11157; add.ftz.f32 %f11159, %f11156, %f11158; add.ftz.f32 %f11160, %f11140, %f11159; sub.ftz.f32 %f11161, %f11149, %f11160; div.approx.ftz.f32 %f11162, %f11161, %f11141; mul.ftz.f32 %f11163, %f11162, %f11162; mul.ftz.f32 %f11164, %f11162, %f11163; mul.ftz.f32 %f11165, %f11163, %f11163; mul.ftz.f32 %f11166, %f11142, %f11163; fma.rn.ftz.f32 %f11167, %f11141, %f11162, %f11166; mul.ftz.f32 %f11168, %f11144, %f11165; fma.rn.ftz.f32 %f11169, %f11143, %f11164, %f11168; add.ftz.f32 %f11170, %f11167, %f11169; add.ftz.f32 %f11171, %f11140, %f11170; sub.ftz.f32 %f11172, %f11149, %f11171; div.approx.ftz.f32 %f11173, %f11172, %f11141; mul.ftz.f32 %f11174, %f11173, %f11173; mul.ftz.f32 %f11175, %f11173, %f11174; mul.ftz.f32 %f11176, %f11174, %f11174; mul.ftz.f32 %f11177, %f11142, %f11174; fma.rn.ftz.f32 %f11178, %f11141, %f11173, %f11177; mul.ftz.f32 %f11179, %f11144, %f11176; fma.rn.ftz.f32 %f11180, %f11143, %f11175, %f11179; add.ftz.f32 %f11181, %f11178, %f11180; add.ftz.f32 %f11182, %f11140, %f11181; sub.ftz.f32 %f11183, %f11149, %f11182; div.approx.ftz.f32 %f11184, %f11183, %f11141; mul.ftz.f32 %f11185, %f11184, %f11184; mul.ftz.f32 %f11186, %f11184, %f11185; mul.ftz.f32 %f11187, %f11185, %f11185; mul.ftz.f32 %f11188, %f11142, %f11185; fma.rn.ftz.f32 %f11189, %f11141, %f11184, %f11188; mul.ftz.f32 %f11190, %f11144, %f11187; fma.rn.ftz.f32 %f11191, %f11143, %f11186, %f11190; add.ftz.f32 %f11192, %f11189, %f11191; add.ftz.f32 %f11193, %f11140, %f11192; sub.ftz.f32 %f11194, %f11149, %f11193; div.approx.ftz.f32 %f11195, %f11194, %f11141; mul.ftz.f32 %f11196, %f11195, %f11195; mul.ftz.f32 %f11197, %f11195, %f11196; mul.ftz.f32 %f11198, %f11196, %f11196; mul.ftz.f32 %f11199, %f11142, %f11196; fma.rn.ftz.f32 %f11200, %f11141, %f11195, %f11199; mul.ftz.f32 %f11201, %f11144, %f11198; fma.rn.ftz.f32 %f11202, %f11143, %f11197, %f11201; add.ftz.f32 %f11203, %f11200, %f11202; add.ftz.f32 %f11204, %f11140, %f11203; sub.ftz.f32 %f11205, %f11149, %f11204; div.approx.ftz.f32 %f11206, %f11205, %f11141; mul.ftz.f32 %f11207, %f11206, %f11206; mul.ftz.f32 %f11208, %f11206, %f11207; mul.ftz.f32 %f11209, %f11207, %f11207; mul.ftz.f32 %f11210, %f11142, %f11207; fma.rn.ftz.f32 %f11211, %f11141, %f11206, %f11210; mul.ftz.f32 %f11212, %f11144, %f11209; fma.rn.ftz.f32 %f11213, %f11143, %f11208, %f11212; add.ftz.f32 %f11214, %f11211, %f11213; add.ftz.f32 %f11215, %f11140, %f11214; sub.ftz.f32 %f11216, %f11149, %f11215; div.approx.ftz.f32 %f11217, %f11216, %f11141; mul.ftz.f32 %f11218, %f11217, %f11217; mul.ftz.f32 %f11219, %f11217, %f11218; mul.ftz.f32 %f11220, %f11218, %f11218; mul.ftz.f32 %f11221, %f11142, %f11218; fma.rn.ftz.f32 %f11222, %f11141, %f11217, %f11221; mul.ftz.f32 %f11223, %f11144, %f11220; fma.rn.ftz.f32 %f11224, %f11143, %f11219, %f11223; add.ftz.f32 %f11225, %f11222, %f11224; add.ftz.f32 %f11226, %f11140, %f11225; sub.ftz.f32 %f11227, %f11149, %f11226; div.approx.ftz.f32 %f11228, %f11227, %f11141; mul.ftz.f32 %f11229, %f11228, %f11228; mul.ftz.f32 %f11230, %f11228, %f11229; mul.ftz.f32 %f11231, %f11229, %f11229; mul.ftz.f32 %f11232, %f11142, %f11229; fma.rn.ftz.f32 %f11233, %f11141, %f11228, %f11232; mul.ftz.f32 %f11234, %f11144, %f11231; fma.rn.ftz.f32 %f11235, %f11143, %f11230, %f11234; add.ftz.f32 %f11236, %f11233, %f11235; add.ftz.f32 %f11237, %f11140, %f11236; sub.ftz.f32 %f11238, %f11149, %f11237; div.approx.ftz.f32 %f11239, %f11238, %f11141; mul.ftz.f32 %f11240, %f11239, %f11239; mul.ftz.f32 %f11241, %f11239, %f11240; mul.ftz.f32 %f11242, %f11240, %f11240; mul.ftz.f32 %f11243, %f11142, %f11240; fma.rn.ftz.f32 %f11244, %f11141, %f11239, %f11243; mul.ftz.f32 %f11245, %f11144, %f11242; fma.rn.ftz.f32 %f11246, %f11143, %f11241, %f11245; add.ftz.f32 %f11247, %f11244, %f11246; add.ftz.f32 %f11248, %f11140, %f11247; sub.ftz.f32 %f11249, %f11149, %f11248; div.approx.ftz.f32 %f11250, %f11249, %f11141; mul.ftz.f32 %f11251, %f11250, %f11250; mul.ftz.f32 %f11252, %f11250, %f11251; mul.ftz.f32 %f11253, %f11251, %f11251; mul.ftz.f32 %f11254, %f11142, %f11251; fma.rn.ftz.f32 %f11255, %f11141, %f11250, %f11254; mul.ftz.f32 %f11256, %f11144, %f11253; fma.rn.ftz.f32 %f11257, %f11143, %f11252, %f11256; add.ftz.f32 %f11258, %f11255, %f11257; add.ftz.f32 %f11259, %f11140, %f11258; sub.ftz.f32 %f11260, %f11149, %f11259; div.approx.ftz.f32 %f11261, %f11260, %f11141; mul.ftz.f32 %f11262, %f11261, %f11261; mul.ftz.f32 %f11263, %f11261, %f11262; mul.ftz.f32 %f11264, %f11262, %f11262; mul.ftz.f32 %f11265, %f11142, %f11262; fma.rn.ftz.f32 %f11266, %f11141, %f11261, %f11265; mul.ftz.f32 %f11267, %f11144, %f11264; fma.rn.ftz.f32 %f11268, %f11143, %f11263, %f11267; add.ftz.f32 %f11269, %f11266, %f11268; add.ftz.f32 %f11270, %f11140, %f11269; sub.ftz.f32 %f11271, %f11149, %f11270; div.approx.ftz.f32 %f11272, %f11271, %f11141; mul.ftz.f32 %f11273, %f11272, %f11272; mul.ftz.f32 %f11274, %f11272, %f11273; mul.ftz.f32 %f11275, %f11273, %f11273; mul.ftz.f32 %f11276, %f11142, %f11273; fma.rn.ftz.f32 %f11277, %f11141, %f11272, %f11276; mul.ftz.f32 %f11278, %f11144, %f11275; fma.rn.ftz.f32 %f11279, %f11143, %f11274, %f11278; add.ftz.f32 %f11280, %f11277, %f11279; add.ftz.f32 %f11281, %f11140, %f11280; sub.ftz.f32 %f11282, %f11149, %f11281; div.approx.ftz.f32 %f11283, %f11282, %f11141; mul.ftz.f32 %f11284, %f11283, %f11283; mul.ftz.f32 %f11285, %f11283, %f11284; mul.ftz.f32 %f11286, %f11284, %f11284; mul.ftz.f32 %f11287, %f11142, %f11284; fma.rn.ftz.f32 %f11288, %f11141, %f11283, %f11287; mul.ftz.f32 %f11289, %f11144, %f11286; fma.rn.ftz.f32 %f11290, %f11143, %f11285, %f11289; add.ftz.f32 %f11291, %f11288, %f11290; add.ftz.f32 %f11292, %f11140, %f11291; sub.ftz.f32 %f11293, %f11149, %f11292; div.approx.ftz.f32 %f11294, %f11293, %f11141; mul.ftz.f32 %f11295, %f11294, %f11294; mul.ftz.f32 %f11296, %f11294, %f11295; mul.ftz.f32 %f11297, %f11295, %f11295; mul.ftz.f32 %f11298, %f11142, %f11295; fma.rn.ftz.f32 %f11299, %f11141, %f11294, %f11298; mul.ftz.f32 %f11300, %f11144, %f11297; fma.rn.ftz.f32 %f11301, %f11143, %f11296, %f11300; add.ftz.f32 %f11302, %f11299, %f11301; add.ftz.f32 %f11303, %f11140, %f11302; sub.ftz.f32 %f11304, %f11149, %f11303; div.approx.ftz.f32 %f11305, %f11304, %f11141; mul.ftz.f32 %f11306, %f11305, %f11305; mul.ftz.f32 %f11307, %f11305, %f11306; mul.ftz.f32 %f11308, %f11306, %f11306; mul.ftz.f32 %f11309, %f11142, %f11306; fma.rn.ftz.f32 %f11310, %f11141, %f11305, %f11309; mul.ftz.f32 %f11311, %f11144, %f11308; fma.rn.ftz.f32 %f11312, %f11143, %f11307, %f11311; add.ftz.f32 %f11313, %f11310, %f11312; add.ftz.f32 %f11314, %f11140, %f11313; sub.ftz.f32 %f11315, %f11149, %f11314; div.approx.ftz.f32 %f11316, %f11315, %f11141; mul.ftz.f32 %f11317, %f11316, %f11316; mul.ftz.f32 %f11318, %f11316, %f11317; mul.ftz.f32 %f11319, %f11317, %f11317; mul.ftz.f32 %f11320, %f11142, %f11317; fma.rn.ftz.f32 %f11321, %f11141, %f11316, %f11320; mul.ftz.f32 %f11322, %f11144, %f11319; fma.rn.ftz.f32 %f11323, %f11143, %f11318, %f11322; add.ftz.f32 %f11324, %f11321, %f11323; add.ftz.f32 %f11325, %f11140, %f11324; sub.ftz.f32 %f11326, %f11149, %f11325; div.approx.ftz.f32 %f11327, %f11326, %f11141; mul.ftz.f32 %f11328, %f11327, %f11327; mul.ftz.f32 %f11329, %f11327, %f11328; mul.ftz.f32 %f11330, %f11328, %f11328; mul.ftz.f32 %f11331, %f11142, %f11328; fma.rn.ftz.f32 %f11332, %f11141, %f11327, %f11331; mul.ftz.f32 %f11333, %f11144, %f11330; fma.rn.ftz.f32 %f11334, %f11143, %f11329, %f11333; add.ftz.f32 %f11335, %f11332, %f11334; add.ftz.f32 %f11336, %f11140, %f11335; sub.ftz.f32 %f11337, %f11149, %f11336; div.approx.ftz.f32 %f11338, %f11337, %f11141; mul.ftz.f32 %f11339, %f11338, %f11338; mul.ftz.f32 %f11340, %f11338, %f11339; mul.ftz.f32 %f11341, %f11339, %f11339; mul.ftz.f32 %f11342, %f11142, %f11339; fma.rn.ftz.f32 %f11343, %f11141, %f11338, %f11342; mul.ftz.f32 %f11344, %f11144, %f11341; fma.rn.ftz.f32 %f11345, %f11143, %f11340, %f11344; add.ftz.f32 %f11346, %f11343, %f11345; add.ftz.f32 %f11347, %f11140, %f11346; sub.ftz.f32 %f11348, %f11149, %f11347; div.approx.ftz.f32 %f11349, %f11348, %f11141; mul.ftz.f32 %f11350, %f11349, %f11349; mul.ftz.f32 %f11351, %f11349, %f11350; mul.ftz.f32 %f11352, %f11350, %f11350; mul.ftz.f32 %f11353, %f11142, %f11350; fma.rn.ftz.f32 %f11354, %f11141, %f11349, %f11353; mul.ftz.f32 %f11355, %f11144, %f11352; fma.rn.ftz.f32 %f11356, %f11143, %f11351, %f11355; add.ftz.f32 %f11357, %f11354, %f11356; add.ftz.f32 %f11358, %f11140, %f11357; sub.ftz.f32 %f11359, %f11149, %f11358; div.approx.ftz.f32 %f11360, %f11359, %f11141; mul.ftz.f32 %f11361, %f11360, %f11360; mul.ftz.f32 %f11362, %f11360, %f11361; mul.ftz.f32 %f11363, %f11361, %f11361; mul.ftz.f32 %f11364, %f11142, %f11361; fma.rn.ftz.f32 %f11365, %f11141, %f11360, %f11364; mul.ftz.f32 %f11366, %f11144, %f11363; fma.rn.ftz.f32 %f11367, %f11143, %f11362, %f11366; add.ftz.f32 %f11368, %f11365, %f11367; add.ftz.f32 %f11369, %f11140, %f11368; sub.ftz.f32 %f11370, %f11149, %f11369; div.approx.ftz.f32 %f1425, %f11370, %f11141; @%p2 bra $L__BB107_1006; bra.uni $L__BB107_1003; $L__BB107_1006: shr.s32 %r2295, %r494, 31; and.b32 %r2296, %r2295, 1078530011; or.b32 %r2297, %r2296, %r495; mov.b32 %f29246, %r2297; bra.uni $L__BB107_1007; $L__BB107_987: ld.const.v4.f32 {%f10935, %f10936, %f10937, %f10938}, [kernel_params+560]; ld.const.v4.f32 {%f10943, %f10944, %f10945, %f10946}, [kernel_params+576]; ld.const.v4.f32 {%f10959, %f10960, %f10961, %f10962}, [kernel_params+608]; mul.ftz.f32 %f10967, %f29217, %f10936; fma.rn.ftz.f32 %f10968, %f29216, %f10935, %f10967; fma.rn.ftz.f32 %f10969, %f29218, %f10937, %f10938; add.ftz.f32 %f1381, %f10968, %f10969; mul.ftz.f32 %f10970, %f29217, %f10944; fma.rn.ftz.f32 %f10971, %f29216, %f10943, %f10970; fma.rn.ftz.f32 %f10972, %f29218, %f10945, %f10946; add.ftz.f32 %f1382, %f10971, %f10972; mul.ftz.f32 %f10976, %f29217, %f10960; fma.rn.ftz.f32 %f10977, %f29216, %f10959, %f10976; fma.rn.ftz.f32 %f10978, %f29218, %f10961, %f10962; add.ftz.f32 %f1384, %f10977, %f10978; setp.eq.ftz.f32 %p846, %f1384, 0f00000000; mov.f32 %f29281, 0f00000000; mov.f32 %f29279, %f29281; mov.f32 %f29280, %f29281; @%p846 bra $L__BB107_989; div.approx.ftz.f32 %f29280, %f1382, %f1384; div.approx.ftz.f32 %f29279, %f1381, %f1384; $L__BB107_989: ld.const.v4.f32 {%f10982, %f10983, %f10984, %f10985}, [kernel_params+736]; ld.const.v4.f32 {%f10990, %f10991, %f10992, %f10993}, [kernel_params+752]; ld.const.v4.f32 {%f11006, %f11007, %f11008, %f11009}, [kernel_params+784]; mul.ftz.f32 %f11014, %f1376, %f10983; fma.rn.ftz.f32 %f11015, %f1375, %f10982, %f11014; fma.rn.ftz.f32 %f11016, %f1377, %f10984, %f10985; add.ftz.f32 %f1393, %f11015, %f11016; mul.ftz.f32 %f11017, %f1376, %f10991; fma.rn.ftz.f32 %f11018, %f1375, %f10990, %f11017; fma.rn.ftz.f32 %f11019, %f1377, %f10992, %f10993; add.ftz.f32 %f1394, %f11018, %f11019; mul.ftz.f32 %f11023, %f1376, %f11007; fma.rn.ftz.f32 %f11024, %f1375, %f11006, %f11023; fma.rn.ftz.f32 %f11025, %f1377, %f11008, %f11009; add.ftz.f32 %f1396, %f11024, %f11025; setp.eq.ftz.f32 %p847, %f1396, 0f00000000; mov.f32 %f29283, %f29281; mov.f32 %f29284, %f29281; @%p847 bra $L__BB107_991; div.approx.ftz.f32 %f29284, %f1394, %f1396; div.approx.ftz.f32 %f29283, %f1393, %f1396; $L__BB107_991: ld.const.v4.f32 {%f11029, %f11030, %f11031, %f11032}, [kernel_params+800]; ld.const.v4.f32 {%f11037, %f11038, %f11039, %f11040}, [kernel_params+816]; ld.const.v4.f32 {%f11053, %f11054, %f11055, %f11056}, [kernel_params+848]; mul.ftz.f32 %f11061, %f1379, %f11030; fma.rn.ftz.f32 %f11062, %f1378, %f11029, %f11061; fma.rn.ftz.f32 %f11063, %f1380, %f11031, %f11032; add.ftz.f32 %f1405, %f11062, %f11063; mul.ftz.f32 %f11064, %f1379, %f11038; fma.rn.ftz.f32 %f11065, %f1378, %f11037, %f11064; fma.rn.ftz.f32 %f11066, %f1380, %f11039, %f11040; add.ftz.f32 %f1406, %f11065, %f11066; mul.ftz.f32 %f11070, %f1379, %f11054; fma.rn.ftz.f32 %f11071, %f1378, %f11053, %f11070; fma.rn.ftz.f32 %f11072, %f1380, %f11055, %f11056; add.ftz.f32 %f1408, %f11071, %f11072; setp.eq.ftz.f32 %p848, %f1408, 0f00000000; mov.f32 %f29282, %f29281; @%p848 bra $L__BB107_1144; div.approx.ftz.f32 %f29282, %f1406, %f1408; div.approx.ftz.f32 %f29281, %f1405, %f1408; bra.uni $L__BB107_1144; $L__BB107_994: setp.eq.s32 %p857, %r493, 0; @%p857 bra $L__BB107_1030; setp.eq.s32 %p858, %r493, 1; @%p858 bra $L__BB107_996; bra.uni $L__BB107_1037; $L__BB107_996: ld.const.f32 %f1450, [kernel_params+24]; mul.ftz.f32 %f11470, %f1419, %f1419; fma.rn.ftz.f32 %f11471, %f1418, %f1418, %f11470; sqrt.approx.ftz.f32 %f11472, %f11471; abs.ftz.f32 %f1451, %f11472; abs.ftz.f32 %f1452, %f1417; setp.eq.ftz.f32 %p887, %f1452, 0f00000000; setp.eq.ftz.f32 %p888, %f1451, 0f00000000; and.pred %p889, %p887, %p888; mov.b32 %r498, %f1417; mov.b32 %r2312, %f11472; and.b32 %r499, %r2312, -2147483648; @%p889 bra $L__BB107_1023; bra.uni $L__BB107_997; $L__BB107_1023: shr.s32 %r2317, %r498, 31; and.b32 %r2318, %r2317, 1078530011; or.b32 %r2319, %r499, %r2318; mov.b32 %f29251, %r2319; bra.uni $L__BB107_1024; $L__BB107_1008: add.ftz.f32 %f11403, %f1418, 0fBF800000; mul.ftz.f32 %f11404, %f11403, 0fBF000000; mov.f32 %f11405, 0f00000000; max.ftz.f32 %f11406, %f11404, %f11405; sqrt.approx.ftz.f32 %f11407, %f11406; add.ftz.f32 %f1432, %f11407, %f11407; setp.leu.ftz.f32 %p866, %f1432, 0f00000000; @%p866 bra $L__BB107_1010; rcp.approx.ftz.f32 %f11408, %f1432; mul.ftz.f32 %f1417, %f1417, %f11408; mul.ftz.f32 %f1419, %f1419, %f11408; $L__BB107_1010: add.ftz.f32 %f11409, %f1417, 0f3F800000; add.ftz.f32 %f11410, %f1419, 0f3F800000; mul.ftz.f32 %f29256, %f11410, 0f3F000000; mul.ftz.f32 %f29255, %f11409, 0f3F000000; bra.uni $L__BB107_1043; $L__BB107_1011: abs.ftz.f32 %f1439, %f1417; setp.eq.ftz.f32 %p868, %f1439, 0f00000000; and.pred %p869, %p868, %p851; mov.b32 %r496, %f1417; and.b32 %r497, %r494, -2147483648; @%p869 bra $L__BB107_1015; bra.uni $L__BB107_1012; $L__BB107_1015: shr.s32 %r2302, %r496, 31; and.b32 %r2303, %r2302, 1078530011; or.b32 %r2304, %r2303, %r497; mov.b32 %f29249, %r2304; bra.uni $L__BB107_1016; $L__BB107_1030: setp.eq.ftz.f32 %p904, %f1418, 0f00000000; setp.eq.ftz.f32 %p905, %f1417, 0f00000000; and.pred %p906, %p905, %p904; setp.eq.ftz.f32 %p907, %f1419, 0f00000000; and.pred %p908, %p907, %p906; mov.f32 %f29255, 0f00000000; mov.f32 %f29256, %f29255; @%p908 bra $L__BB107_1043; ld.const.v4.f32 {%f11529, %f11530, %f11531, %f11532}, [kernel_params+32]; abs.ftz.f32 %f1468, %f1417; setp.eq.ftz.f32 %p910, %f1468, 0f00000000; and.pred %p911, %p910, %p851; mov.b32 %r500, %f1417; and.b32 %r501, %r494, -2147483648; @%p911 bra $L__BB107_1035; bra.uni $L__BB107_1032; $L__BB107_1035: shr.s32 %r2331, %r500, 31; and.b32 %r2332, %r2331, 1078530011; or.b32 %r2333, %r2332, %r501; mov.b32 %f29253, %r2333; bra.uni $L__BB107_1036; $L__BB107_1037: ld.const.f32 %f11591, [kernel_params+28]; ld.const.f32 %f1475, [kernel_params+276]; ld.const.f32 %f1476, [kernel_params+280]; add.ftz.f32 %f11592, %f11591, %f11591; mul.ftz.f32 %f11593, %f1420, 0f3F000000; sin.approx.ftz.f32 %f11594, %f11593; mul.ftz.f32 %f1477, %f11592, %f11594; @%p2 bra $L__BB107_1041; bra.uni $L__BB107_1038; $L__BB107_1041: shr.s32 %r2338, %r494, 31; and.b32 %r2339, %r2338, 1078530011; or.b32 %r2340, %r2339, %r495; mov.b32 %f29254, %r2340; bra.uni $L__BB107_1042; $L__BB107_1003: setp.eq.ftz.f32 %p859, %f1421, 0f7F800000; setp.eq.ftz.f32 %p860, %f1422, 0f7F800000; and.pred %p861, %p859, %p860; @%p861 bra $L__BB107_1005; bra.uni $L__BB107_1004; $L__BB107_1005: setp.lt.s32 %p865, %r494, 0; selp.b32 %r2293, 1075235812, 1061752795, %p865; or.b32 %r2294, %r2293, %r495; mov.b32 %f29246, %r2294; bra.uni $L__BB107_1007; $L__BB107_997: setp.eq.ftz.f32 %p890, %f1452, 0f7F800000; setp.eq.ftz.f32 %p891, %f1451, 0f7F800000; and.pred %p892, %p890, %p891; @%p892 bra $L__BB107_1022; bra.uni $L__BB107_998; $L__BB107_1022: setp.lt.s32 %p896, %r498, 0; selp.b32 %r2315, 1075235812, 1061752795, %p896; or.b32 %r2316, %r499, %r2315; mov.b32 %f29251, %r2316; bra.uni $L__BB107_1024; $L__BB107_1012: setp.eq.ftz.f32 %p870, %f1439, 0f7F800000; setp.eq.ftz.f32 %p871, %f1421, 0f7F800000; and.pred %p872, %p870, %p871; @%p872 bra $L__BB107_1014; bra.uni $L__BB107_1013; $L__BB107_1014: setp.lt.s32 %p876, %r496, 0; selp.b32 %r2300, 1075235812, 1061752795, %p876; or.b32 %r2301, %r2300, %r497; mov.b32 %f29249, %r2301; bra.uni $L__BB107_1016; $L__BB107_1038: setp.eq.ftz.f32 %p921, %f1421, 0f7F800000; setp.eq.ftz.f32 %p922, %f1422, 0f7F800000; and.pred %p923, %p921, %p922; @%p923 bra $L__BB107_1040; bra.uni $L__BB107_1039; $L__BB107_1040: setp.lt.s32 %p927, %r494, 0; selp.b32 %r2336, 1075235812, 1061752795, %p927; or.b32 %r2337, %r2336, %r495; mov.b32 %f29254, %r2337; bra.uni $L__BB107_1042; $L__BB107_1032: setp.eq.ftz.f32 %p912, %f1468, 0f7F800000; setp.eq.ftz.f32 %p913, %f1421, 0f7F800000; and.pred %p914, %p912, %p913; @%p914 bra $L__BB107_1034; bra.uni $L__BB107_1033; $L__BB107_1034: setp.lt.s32 %p918, %r500, 0; selp.b32 %r2329, 1075235812, 1061752795, %p918; or.b32 %r2330, %r2329, %r501; mov.b32 %f29253, %r2330; bra.uni $L__BB107_1036; $L__BB107_1004: mov.f32 %f28768, 0f3FC90FDB; setp.lt.s32 %p862, %r494, 0; min.ftz.f32 %f11371, %f1422, %f1421; max.ftz.f32 %f11372, %f1422, %f1421; div.full.ftz.f32 %f11373, %f11371, %f11372; mul.rn.ftz.f32 %f11374, %f11373, %f11373; mov.f32 %f11375, 0fC0B59883; mov.f32 %f11376, 0fBF52C7EA; fma.rn.ftz.f32 %f11377, %f11374, %f11376, %f11375; mov.f32 %f11378, 0fC0D21907; fma.rn.ftz.f32 %f11379, %f11377, %f11374, %f11378; mul.ftz.f32 %f11380, %f11374, %f11379; mul.ftz.f32 %f11381, %f11373, %f11380; add.ftz.f32 %f11382, %f11374, 0f41355DC0; mov.f32 %f11383, 0f41E6BD60; fma.rn.ftz.f32 %f11384, %f11382, %f11374, %f11383; mov.f32 %f11385, 0f419D92C8; fma.rn.ftz.f32 %f11386, %f11384, %f11374, %f11385; rcp.approx.ftz.f32 %f11387, %f11386; fma.rn.ftz.f32 %f11388, %f11381, %f11387, %f11373; sub.ftz.f32 %f11390, %f28768, %f11388; setp.gt.ftz.f32 %p863, %f1422, %f1421; selp.f32 %f11391, %f11390, %f11388, %p863; sub.ftz.f32 %f11393, %f11138, %f11391; selp.f32 %f11394, %f11393, %f11391, %p862; mov.b32 %r2291, %f11394; or.b32 %r2292, %r495, %r2291; mov.b32 %f11395, %r2292; add.ftz.f32 %f11396, %f1421, %f1422; setp.le.ftz.f32 %p864, %f11396, 0f7F800000; selp.f32 %f29246, %f11395, %f11396, %p864; $L__BB107_1007: ld.const.f32 %f28718, [kernel_params+280]; ld.const.f32 %f28717, [kernel_params+276]; cos.approx.ftz.f32 %f11397, %f29246; mul.ftz.f32 %f11398, %f1425, %f11397; div.approx.ftz.f32 %f11399, %f11398, %f28717; sin.approx.ftz.f32 %f11400, %f29246; mul.ftz.f32 %f11401, %f1425, %f11400; div.approx.ftz.f32 %f11402, %f11401, %f28718; add.ftz.f32 %f29256, %f11402, 0f3F000000; add.ftz.f32 %f29255, %f11399, 0f3F000000; bra.uni $L__BB107_1043; $L__BB107_998: setp.lt.s32 %p893, %r498, 0; min.ftz.f32 %f11473, %f1451, %f1452; max.ftz.f32 %f11474, %f1451, %f1452; div.full.ftz.f32 %f11475, %f11473, %f11474; mul.rn.ftz.f32 %f11476, %f11475, %f11475; mov.f32 %f11477, 0fC0B59883; mov.f32 %f11478, 0fBF52C7EA; fma.rn.ftz.f32 %f11479, %f11476, %f11478, %f11477; mov.f32 %f11480, 0fC0D21907; fma.rn.ftz.f32 %f11481, %f11479, %f11476, %f11480; mul.ftz.f32 %f11482, %f11476, %f11481; mul.ftz.f32 %f11483, %f11475, %f11482; add.ftz.f32 %f11484, %f11476, 0f41355DC0; mov.f32 %f11485, 0f41E6BD60; fma.rn.ftz.f32 %f11486, %f11484, %f11476, %f11485; mov.f32 %f11487, 0f419D92C8; fma.rn.ftz.f32 %f11488, %f11486, %f11476, %f11487; rcp.approx.ftz.f32 %f11489, %f11488; fma.rn.ftz.f32 %f11490, %f11483, %f11489, %f11475; sub.ftz.f32 %f11492, %f11135, %f11490; setp.gt.ftz.f32 %p894, %f1451, %f1452; selp.f32 %f11493, %f11492, %f11490, %p894; sub.ftz.f32 %f11495, %f11138, %f11493; selp.f32 %f11496, %f11495, %f11493, %p893; mov.b32 %r2313, %f11496; or.b32 %r2314, %r499, %r2313; mov.b32 %f11497, %r2314; add.ftz.f32 %f11498, %f1452, %f1451; setp.le.ftz.f32 %p895, %f11498, 0f7F800000; selp.f32 %f29251, %f11497, %f11498, %p895; $L__BB107_1024: div.approx.ftz.f32 %f1457, %f29251, %f1450; @%p2 bra $L__BB107_1028; bra.uni $L__BB107_1025; $L__BB107_1028: shr.s32 %r2324, %r494, 31; and.b32 %r2325, %r2324, 1078530011; or.b32 %r2326, %r2325, %r495; mov.b32 %f29252, %r2326; bra.uni $L__BB107_1029; $L__BB107_1025: setp.eq.ftz.f32 %p897, %f1421, 0f7F800000; setp.eq.ftz.f32 %p898, %f1422, 0f7F800000; and.pred %p899, %p897, %p898; @%p899 bra $L__BB107_1027; bra.uni $L__BB107_1026; $L__BB107_1027: setp.lt.s32 %p903, %r494, 0; selp.b32 %r2322, 1075235812, 1061752795, %p903; or.b32 %r2323, %r2322, %r495; mov.b32 %f29252, %r2323; bra.uni $L__BB107_1029; $L__BB107_1013: setp.lt.s32 %p873, %r496, 0; min.ftz.f32 %f11411, %f1421, %f1439; max.ftz.f32 %f11412, %f1421, %f1439; div.full.ftz.f32 %f11413, %f11411, %f11412; mul.rn.ftz.f32 %f11414, %f11413, %f11413; mov.f32 %f11415, 0fC0B59883; mov.f32 %f11416, 0fBF52C7EA; fma.rn.ftz.f32 %f11417, %f11414, %f11416, %f11415; mov.f32 %f11418, 0fC0D21907; fma.rn.ftz.f32 %f11419, %f11417, %f11414, %f11418; mul.ftz.f32 %f11420, %f11414, %f11419; mul.ftz.f32 %f11421, %f11413, %f11420; add.ftz.f32 %f11422, %f11414, 0f41355DC0; mov.f32 %f11423, 0f41E6BD60; fma.rn.ftz.f32 %f11424, %f11422, %f11414, %f11423; mov.f32 %f11425, 0f419D92C8; fma.rn.ftz.f32 %f11426, %f11424, %f11414, %f11425; rcp.approx.ftz.f32 %f11427, %f11426; fma.rn.ftz.f32 %f11428, %f11421, %f11427, %f11413; sub.ftz.f32 %f11430, %f11135, %f11428; setp.gt.ftz.f32 %p874, %f1421, %f1439; selp.f32 %f11431, %f11430, %f11428, %p874; sub.ftz.f32 %f11433, %f11138, %f11431; selp.f32 %f11434, %f11433, %f11431, %p873; mov.b32 %r2298, %f11434; or.b32 %r2299, %r497, %r2298; mov.b32 %f11435, %r2299; add.ftz.f32 %f11436, %f1439, %f1421; setp.le.ftz.f32 %p875, %f11436, 0f7F800000; selp.f32 %f29249, %f11435, %f11436, %p875; $L__BB107_1016: and.pred %p879, %p868, %p852; @%p879 bra $L__BB107_1020; bra.uni $L__BB107_1017; $L__BB107_1020: shr.s32 %r2309, %r496, 31; and.b32 %r2310, %r2309, 1078530011; or.b32 %r2311, %r495, %r2310; mov.b32 %f29250, %r2311; bra.uni $L__BB107_1021; $L__BB107_1017: setp.eq.ftz.f32 %p880, %f1439, 0f7F800000; setp.eq.ftz.f32 %p881, %f1422, 0f7F800000; and.pred %p882, %p880, %p881; @%p882 bra $L__BB107_1019; bra.uni $L__BB107_1018; $L__BB107_1019: setp.lt.s32 %p886, %r496, 0; selp.b32 %r2307, 1075235812, 1061752795, %p886; or.b32 %r2308, %r495, %r2307; mov.b32 %f29250, %r2308; bra.uni $L__BB107_1021; $L__BB107_1039: setp.lt.s32 %p924, %r494, 0; min.ftz.f32 %f11595, %f1422, %f1421; max.ftz.f32 %f11596, %f1422, %f1421; div.full.ftz.f32 %f11597, %f11595, %f11596; mul.rn.ftz.f32 %f11598, %f11597, %f11597; mov.f32 %f11599, 0fC0B59883; mov.f32 %f11600, 0fBF52C7EA; fma.rn.ftz.f32 %f11601, %f11598, %f11600, %f11599; mov.f32 %f11602, 0fC0D21907; fma.rn.ftz.f32 %f11603, %f11601, %f11598, %f11602; mul.ftz.f32 %f11604, %f11598, %f11603; mul.ftz.f32 %f11605, %f11597, %f11604; add.ftz.f32 %f11606, %f11598, 0f41355DC0; mov.f32 %f11607, 0f41E6BD60; fma.rn.ftz.f32 %f11608, %f11606, %f11598, %f11607; mov.f32 %f11609, 0f419D92C8; fma.rn.ftz.f32 %f11610, %f11608, %f11598, %f11609; rcp.approx.ftz.f32 %f11611, %f11610; fma.rn.ftz.f32 %f11612, %f11605, %f11611, %f11597; sub.ftz.f32 %f11614, %f11135, %f11612; setp.gt.ftz.f32 %p925, %f1422, %f1421; selp.f32 %f11615, %f11614, %f11612, %p925; sub.ftz.f32 %f11617, %f11138, %f11615; selp.f32 %f11618, %f11617, %f11615, %p924; mov.b32 %r2334, %f11618; or.b32 %r2335, %r495, %r2334; mov.b32 %f11619, %r2335; add.ftz.f32 %f11620, %f1421, %f1422; setp.le.ftz.f32 %p926, %f11620, 0f7F800000; selp.f32 %f29254, %f11619, %f11620, %p926; $L__BB107_1042: cos.approx.ftz.f32 %f11621, %f29254; mul.ftz.f32 %f11622, %f1477, %f11621; div.approx.ftz.f32 %f11623, %f11622, %f1475; sin.approx.ftz.f32 %f11624, %f29254; mul.ftz.f32 %f11625, %f1477, %f11624; div.approx.ftz.f32 %f11626, %f11625, %f1476; add.ftz.f32 %f29256, %f11626, 0f3F000000; add.ftz.f32 %f29255, %f11623, 0f3F000000; bra.uni $L__BB107_1043; $L__BB107_1026: setp.lt.s32 %p900, %r494, 0; min.ftz.f32 %f11499, %f1422, %f1421; max.ftz.f32 %f11500, %f1422, %f1421; div.full.ftz.f32 %f11501, %f11499, %f11500; mul.rn.ftz.f32 %f11502, %f11501, %f11501; mov.f32 %f11503, 0fC0B59883; mov.f32 %f11504, 0fBF52C7EA; fma.rn.ftz.f32 %f11505, %f11502, %f11504, %f11503; mov.f32 %f11506, 0fC0D21907; fma.rn.ftz.f32 %f11507, %f11505, %f11502, %f11506; mul.ftz.f32 %f11508, %f11502, %f11507; mul.ftz.f32 %f11509, %f11501, %f11508; add.ftz.f32 %f11510, %f11502, 0f41355DC0; mov.f32 %f11511, 0f41E6BD60; fma.rn.ftz.f32 %f11512, %f11510, %f11502, %f11511; mov.f32 %f11513, 0f419D92C8; fma.rn.ftz.f32 %f11514, %f11512, %f11502, %f11513; rcp.approx.ftz.f32 %f11515, %f11514; fma.rn.ftz.f32 %f11516, %f11509, %f11515, %f11501; sub.ftz.f32 %f11518, %f11135, %f11516; setp.gt.ftz.f32 %p901, %f1422, %f1421; selp.f32 %f11519, %f11518, %f11516, %p901; sub.ftz.f32 %f11521, %f11138, %f11519; selp.f32 %f11522, %f11521, %f11519, %p900; mov.b32 %r2320, %f11522; or.b32 %r2321, %r495, %r2320; mov.b32 %f11523, %r2321; add.ftz.f32 %f11524, %f1421, %f1422; setp.le.ftz.f32 %p902, %f11524, 0f7F800000; selp.f32 %f29252, %f11523, %f11524, %p902; $L__BB107_1029: cos.approx.ftz.f32 %f11525, %f29252; sin.approx.ftz.f32 %f11526, %f29252; fma.rn.ftz.f32 %f29256, %f1457, %f11526, 0f3F000000; fma.rn.ftz.f32 %f29255, %f1457, %f11525, 0f3F000000; bra.uni $L__BB107_1043; $L__BB107_1018: setp.lt.s32 %p883, %r496, 0; min.ftz.f32 %f11437, %f1422, %f1439; max.ftz.f32 %f11438, %f1422, %f1439; div.full.ftz.f32 %f11439, %f11437, %f11438; mul.rn.ftz.f32 %f11440, %f11439, %f11439; mov.f32 %f11441, 0fC0B59883; mov.f32 %f11442, 0fBF52C7EA; fma.rn.ftz.f32 %f11443, %f11440, %f11442, %f11441; mov.f32 %f11444, 0fC0D21907; fma.rn.ftz.f32 %f11445, %f11443, %f11440, %f11444; mul.ftz.f32 %f11446, %f11440, %f11445; mul.ftz.f32 %f11447, %f11439, %f11446; add.ftz.f32 %f11448, %f11440, 0f41355DC0; mov.f32 %f11449, 0f41E6BD60; fma.rn.ftz.f32 %f11450, %f11448, %f11440, %f11449; mov.f32 %f11451, 0f419D92C8; fma.rn.ftz.f32 %f11452, %f11450, %f11440, %f11451; rcp.approx.ftz.f32 %f11453, %f11452; fma.rn.ftz.f32 %f11454, %f11447, %f11453, %f11439; sub.ftz.f32 %f11456, %f11135, %f11454; setp.gt.ftz.f32 %p884, %f1422, %f1439; selp.f32 %f11457, %f11456, %f11454, %p884; sub.ftz.f32 %f11459, %f11138, %f11457; selp.f32 %f11460, %f11459, %f11457, %p883; mov.b32 %r2305, %f11460; or.b32 %r2306, %r495, %r2305; mov.b32 %f11461, %r2306; add.ftz.f32 %f11462, %f1439, %f1422; setp.le.ftz.f32 %p885, %f11462, 0f7F800000; selp.f32 %f29250, %f11461, %f11462, %p885; $L__BB107_1021: add.ftz.f32 %f11463, %f29250, %f29250; div.approx.ftz.f32 %f11465, %f11463, %f11138; add.ftz.f32 %f11466, %f29249, %f29249; div.approx.ftz.f32 %f11467, %f11466, %f11138; add.ftz.f32 %f11468, %f11467, 0f3F000000; sub.ftz.f32 %f29255, %f11115, %f11468; add.ftz.f32 %f29256, %f11465, 0f3F000000; bra.uni $L__BB107_1043; $L__BB107_1033: setp.lt.s32 %p915, %r500, 0; min.ftz.f32 %f11533, %f1421, %f1468; max.ftz.f32 %f11534, %f1421, %f1468; div.full.ftz.f32 %f11535, %f11533, %f11534; mul.rn.ftz.f32 %f11536, %f11535, %f11535; mov.f32 %f11537, 0fC0B59883; mov.f32 %f11538, 0fBF52C7EA; fma.rn.ftz.f32 %f11539, %f11536, %f11538, %f11537; mov.f32 %f11540, 0fC0D21907; fma.rn.ftz.f32 %f11541, %f11539, %f11536, %f11540; mul.ftz.f32 %f11542, %f11536, %f11541; mul.ftz.f32 %f11543, %f11535, %f11542; add.ftz.f32 %f11544, %f11536, 0f41355DC0; mov.f32 %f11545, 0f41E6BD60; fma.rn.ftz.f32 %f11546, %f11544, %f11536, %f11545; mov.f32 %f11547, 0f419D92C8; fma.rn.ftz.f32 %f11548, %f11546, %f11536, %f11547; rcp.approx.ftz.f32 %f11549, %f11548; fma.rn.ftz.f32 %f11550, %f11543, %f11549, %f11535; sub.ftz.f32 %f11552, %f11135, %f11550; setp.gt.ftz.f32 %p916, %f1421, %f1468; selp.f32 %f11553, %f11552, %f11550, %p916; sub.ftz.f32 %f11555, %f11138, %f11553; selp.f32 %f11556, %f11555, %f11553, %p915; mov.b32 %r2327, %f11556; or.b32 %r2328, %r501, %r2327; mov.b32 %f11557, %r2328; add.ftz.f32 %f11558, %f1468, %f1421; setp.le.ftz.f32 %p917, %f11558, 0f7F800000; selp.f32 %f29253, %f11557, %f11558, %p917; $L__BB107_1036: sub.ftz.f32 %f11559, %f29253, %f11530; mul.ftz.f32 %f11560, %f1418, %f1418; fma.rn.ftz.f32 %f11561, %f1417, %f1417, %f11560; fma.rn.ftz.f32 %f11562, %f1419, %f1419, %f11561; sqrt.approx.ftz.f32 %f11563, %f11562; div.approx.ftz.f32 %f11564, %f1419, %f11563; abs.ftz.f32 %f11565, %f11564; sub.ftz.f32 %f11567, %f11115, %f11565; mul.ftz.f32 %f11568, %f11567, 0f3F000000; sqrt.approx.ftz.f32 %f11569, %f11568; setp.gt.ftz.f32 %p919, %f11565, 0f3F11EB85; selp.f32 %f11570, %f11569, %f11565, %p919; mul.ftz.f32 %f11571, %f11570, %f11570; fma.rn.ftz.f32 %f11574, %f11124, %f11571, %f11123; fma.rn.ftz.f32 %f11576, %f11574, %f11571, %f11126; fma.rn.ftz.f32 %f11578, %f11576, %f11571, %f11128; fma.rn.ftz.f32 %f11580, %f11578, %f11571, %f11130; mul.ftz.f32 %f11581, %f11571, %f11580; fma.rn.ftz.f32 %f11582, %f11581, %f11570, %f11570; add.ftz.f32 %f11583, %f11582, %f11582; sub.ftz.f32 %f11585, %f11135, %f11582; selp.f32 %f11586, %f11583, %f11585, %p919; setp.lt.ftz.f32 %p920, %f11564, 0f00000000; sub.ftz.f32 %f11588, %f11138, %f11586; selp.f32 %f11589, %f11588, %f11586, %p920; sub.ftz.f32 %f11590, %f11589, %f11532; div.approx.ftz.f32 %f29256, %f11590, %f11531; div.approx.ftz.f32 %f29255, %f11559, %f11529; $L__BB107_1043: ld.const.u32 %r5545, [kernel_params+20]; setp.gt.s32 %p3919, %r5545, 2; mov.f32 %f28769, 0f3FC90FDB; mov.f32 %f28725, 0fBF800000; mov.f32 %f28724, 0f3E2AAB94; mov.f32 %f28723, 0f3D994929; mov.f32 %f28722, 0f3D3F841F; mov.f32 %f28721, 0f3C94D2E9; mov.f32 %f28720, 0f3D53F941; mov.f32 %f28719, 0f3F800000; ld.const.v4.f32 {%f11627, %f11628, %f11629, %f11630}, [kernel_params+864]; ld.const.v4.f32 {%f11635, %f11636, %f11637, %f11638}, [kernel_params+880]; ld.const.v4.f32 {%f11643, %f11644, %f11645, %f11646}, [kernel_params+896]; mul.ftz.f32 %f11651, %f1376, %f11628; fma.rn.ftz.f32 %f11652, %f1375, %f11627, %f11651; fma.rn.ftz.f32 %f11653, %f1377, %f11629, %f11652; add.ftz.f32 %f11654, %f11630, %f11653; mul.ftz.f32 %f11655, %f1376, %f11636; fma.rn.ftz.f32 %f11656, %f1375, %f11635, %f11655; fma.rn.ftz.f32 %f11657, %f1377, %f11637, %f11656; add.ftz.f32 %f11658, %f11638, %f11657; mul.ftz.f32 %f11659, %f1376, %f11644; fma.rn.ftz.f32 %f11660, %f1375, %f11643, %f11659; fma.rn.ftz.f32 %f11661, %f1377, %f11645, %f11660; add.ftz.f32 %f11662, %f11646, %f11661; mul.ftz.f32 %f11663, %f11658, %f11658; fma.rn.ftz.f32 %f11664, %f11654, %f11654, %f11663; fma.rn.ftz.f32 %f11665, %f11662, %f11662, %f11664; sqrt.approx.ftz.f32 %f11666, %f11665; div.approx.ftz.f32 %f1488, %f11654, %f11666; div.approx.ftz.f32 %f1489, %f11658, %f11666; div.approx.ftz.f32 %f1490, %f11662, %f11666; max.ftz.f32 %f11668, %f1488, %f28725; min.ftz.f32 %f11670, %f11668, %f28719; abs.ftz.f32 %f11671, %f11670; sub.ftz.f32 %f11672, %f28719, %f11671; mul.ftz.f32 %f11673, %f11672, 0f3F000000; sqrt.approx.ftz.f32 %f11674, %f11673; setp.gt.ftz.f32 %p928, %f11671, 0f3F11EB85; selp.f32 %f11675, %f11674, %f11671, %p928; mul.ftz.f32 %f11676, %f11675, %f11675; fma.rn.ftz.f32 %f11679, %f28720, %f11676, %f28721; fma.rn.ftz.f32 %f11681, %f11679, %f11676, %f28722; fma.rn.ftz.f32 %f11683, %f11681, %f11676, %f28723; fma.rn.ftz.f32 %f11685, %f11683, %f11676, %f28724; mul.ftz.f32 %f11686, %f11676, %f11685; fma.rn.ftz.f32 %f11687, %f11686, %f11675, %f11675; add.ftz.f32 %f11688, %f11687, %f11687; sub.ftz.f32 %f11690, %f28769, %f11687; selp.f32 %f11691, %f11688, %f11690, %p928; setp.lt.ftz.f32 %p929, %f11670, 0f00000000; sub.ftz.f32 %f11693, %f11138, %f11691; selp.f32 %f1491, %f11693, %f11691, %p929; abs.ftz.f32 %f1492, %f1489; abs.ftz.f32 %f1493, %f1490; setp.eq.ftz.f32 %p930, %f1492, 0f00000000; setp.eq.ftz.f32 %p931, %f1493, 0f00000000; and.pred %p3, %p930, %p931; mov.b32 %r502, %f1489; mov.b32 %r2341, %f1490; and.b32 %r503, %r2341, -2147483648; @%p3919 bra $L__BB107_1049; bra.uni $L__BB107_1044; $L__BB107_1049: ld.const.u32 %r5530, [kernel_params+20]; setp.eq.s32 %p933, %r5530, 3; @%p933 bra $L__BB107_1058; ld.const.u32 %r5531, [kernel_params+20]; setp.eq.s32 %p934, %r5531, 5; @%p934 bra $L__BB107_1061; ld.const.u32 %r5532, [kernel_params+20]; setp.ne.s32 %p935, %r5532, 4; @%p935 bra $L__BB107_1087; ld.const.f32 %f11694, [kernel_params+48]; ld.const.v4.f32 {%f11695, %f11696, %f11697, %f11698}, [kernel_params+64]; neg.ftz.f32 %f11703, %f1491; sub.ftz.f32 %f11704, %f11703, %f11694; div.approx.ftz.f32 %f11705, %f11704, %f11695; mul.ftz.f32 %f11706, %f11705, %f11705; mul.ftz.f32 %f11707, %f11705, %f11706; mul.ftz.f32 %f11708, %f11706, %f11706; mul.ftz.f32 %f11709, %f11696, %f11706; fma.rn.ftz.f32 %f11710, %f11695, %f11705, %f11709; mul.ftz.f32 %f11711, %f11698, %f11708; fma.rn.ftz.f32 %f11712, %f11697, %f11707, %f11711; add.ftz.f32 %f11713, %f11710, %f11712; add.ftz.f32 %f11714, %f11694, %f11713; sub.ftz.f32 %f11715, %f11703, %f11714; div.approx.ftz.f32 %f11716, %f11715, %f11695; mul.ftz.f32 %f11717, %f11716, %f11716; mul.ftz.f32 %f11718, %f11716, %f11717; mul.ftz.f32 %f11719, %f11717, %f11717; mul.ftz.f32 %f11720, %f11696, %f11717; fma.rn.ftz.f32 %f11721, %f11695, %f11716, %f11720; mul.ftz.f32 %f11722, %f11698, %f11719; fma.rn.ftz.f32 %f11723, %f11697, %f11718, %f11722; add.ftz.f32 %f11724, %f11721, %f11723; add.ftz.f32 %f11725, %f11694, %f11724; sub.ftz.f32 %f11726, %f11703, %f11725; div.approx.ftz.f32 %f11727, %f11726, %f11695; mul.ftz.f32 %f11728, %f11727, %f11727; mul.ftz.f32 %f11729, %f11727, %f11728; mul.ftz.f32 %f11730, %f11728, %f11728; mul.ftz.f32 %f11731, %f11696, %f11728; fma.rn.ftz.f32 %f11732, %f11695, %f11727, %f11731; mul.ftz.f32 %f11733, %f11698, %f11730; fma.rn.ftz.f32 %f11734, %f11697, %f11729, %f11733; add.ftz.f32 %f11735, %f11732, %f11734; add.ftz.f32 %f11736, %f11694, %f11735; sub.ftz.f32 %f11737, %f11703, %f11736; div.approx.ftz.f32 %f11738, %f11737, %f11695; mul.ftz.f32 %f11739, %f11738, %f11738; mul.ftz.f32 %f11740, %f11738, %f11739; mul.ftz.f32 %f11741, %f11739, %f11739; mul.ftz.f32 %f11742, %f11696, %f11739; fma.rn.ftz.f32 %f11743, %f11695, %f11738, %f11742; mul.ftz.f32 %f11744, %f11698, %f11741; fma.rn.ftz.f32 %f11745, %f11697, %f11740, %f11744; add.ftz.f32 %f11746, %f11743, %f11745; add.ftz.f32 %f11747, %f11694, %f11746; sub.ftz.f32 %f11748, %f11703, %f11747; div.approx.ftz.f32 %f11749, %f11748, %f11695; mul.ftz.f32 %f11750, %f11749, %f11749; mul.ftz.f32 %f11751, %f11749, %f11750; mul.ftz.f32 %f11752, %f11750, %f11750; mul.ftz.f32 %f11753, %f11696, %f11750; fma.rn.ftz.f32 %f11754, %f11695, %f11749, %f11753; mul.ftz.f32 %f11755, %f11698, %f11752; fma.rn.ftz.f32 %f11756, %f11697, %f11751, %f11755; add.ftz.f32 %f11757, %f11754, %f11756; add.ftz.f32 %f11758, %f11694, %f11757; sub.ftz.f32 %f11759, %f11703, %f11758; div.approx.ftz.f32 %f11760, %f11759, %f11695; mul.ftz.f32 %f11761, %f11760, %f11760; mul.ftz.f32 %f11762, %f11760, %f11761; mul.ftz.f32 %f11763, %f11761, %f11761; mul.ftz.f32 %f11764, %f11696, %f11761; fma.rn.ftz.f32 %f11765, %f11695, %f11760, %f11764; mul.ftz.f32 %f11766, %f11698, %f11763; fma.rn.ftz.f32 %f11767, %f11697, %f11762, %f11766; add.ftz.f32 %f11768, %f11765, %f11767; add.ftz.f32 %f11769, %f11694, %f11768; sub.ftz.f32 %f11770, %f11703, %f11769; div.approx.ftz.f32 %f11771, %f11770, %f11695; mul.ftz.f32 %f11772, %f11771, %f11771; mul.ftz.f32 %f11773, %f11771, %f11772; mul.ftz.f32 %f11774, %f11772, %f11772; mul.ftz.f32 %f11775, %f11696, %f11772; fma.rn.ftz.f32 %f11776, %f11695, %f11771, %f11775; mul.ftz.f32 %f11777, %f11698, %f11774; fma.rn.ftz.f32 %f11778, %f11697, %f11773, %f11777; add.ftz.f32 %f11779, %f11776, %f11778; add.ftz.f32 %f11780, %f11694, %f11779; sub.ftz.f32 %f11781, %f11703, %f11780; div.approx.ftz.f32 %f11782, %f11781, %f11695; mul.ftz.f32 %f11783, %f11782, %f11782; mul.ftz.f32 %f11784, %f11782, %f11783; mul.ftz.f32 %f11785, %f11783, %f11783; mul.ftz.f32 %f11786, %f11696, %f11783; fma.rn.ftz.f32 %f11787, %f11695, %f11782, %f11786; mul.ftz.f32 %f11788, %f11698, %f11785; fma.rn.ftz.f32 %f11789, %f11697, %f11784, %f11788; add.ftz.f32 %f11790, %f11787, %f11789; add.ftz.f32 %f11791, %f11694, %f11790; sub.ftz.f32 %f11792, %f11703, %f11791; div.approx.ftz.f32 %f11793, %f11792, %f11695; mul.ftz.f32 %f11794, %f11793, %f11793; mul.ftz.f32 %f11795, %f11793, %f11794; mul.ftz.f32 %f11796, %f11794, %f11794; mul.ftz.f32 %f11797, %f11696, %f11794; fma.rn.ftz.f32 %f11798, %f11695, %f11793, %f11797; mul.ftz.f32 %f11799, %f11698, %f11796; fma.rn.ftz.f32 %f11800, %f11697, %f11795, %f11799; add.ftz.f32 %f11801, %f11798, %f11800; add.ftz.f32 %f11802, %f11694, %f11801; sub.ftz.f32 %f11803, %f11703, %f11802; div.approx.ftz.f32 %f11804, %f11803, %f11695; mul.ftz.f32 %f11805, %f11804, %f11804; mul.ftz.f32 %f11806, %f11804, %f11805; mul.ftz.f32 %f11807, %f11805, %f11805; mul.ftz.f32 %f11808, %f11696, %f11805; fma.rn.ftz.f32 %f11809, %f11695, %f11804, %f11808; mul.ftz.f32 %f11810, %f11698, %f11807; fma.rn.ftz.f32 %f11811, %f11697, %f11806, %f11810; add.ftz.f32 %f11812, %f11809, %f11811; add.ftz.f32 %f11813, %f11694, %f11812; sub.ftz.f32 %f11814, %f11703, %f11813; div.approx.ftz.f32 %f11815, %f11814, %f11695; mul.ftz.f32 %f11816, %f11815, %f11815; mul.ftz.f32 %f11817, %f11815, %f11816; mul.ftz.f32 %f11818, %f11816, %f11816; mul.ftz.f32 %f11819, %f11696, %f11816; fma.rn.ftz.f32 %f11820, %f11695, %f11815, %f11819; mul.ftz.f32 %f11821, %f11698, %f11818; fma.rn.ftz.f32 %f11822, %f11697, %f11817, %f11821; add.ftz.f32 %f11823, %f11820, %f11822; add.ftz.f32 %f11824, %f11694, %f11823; sub.ftz.f32 %f11825, %f11703, %f11824; div.approx.ftz.f32 %f11826, %f11825, %f11695; mul.ftz.f32 %f11827, %f11826, %f11826; mul.ftz.f32 %f11828, %f11826, %f11827; mul.ftz.f32 %f11829, %f11827, %f11827; mul.ftz.f32 %f11830, %f11696, %f11827; fma.rn.ftz.f32 %f11831, %f11695, %f11826, %f11830; mul.ftz.f32 %f11832, %f11698, %f11829; fma.rn.ftz.f32 %f11833, %f11697, %f11828, %f11832; add.ftz.f32 %f11834, %f11831, %f11833; add.ftz.f32 %f11835, %f11694, %f11834; sub.ftz.f32 %f11836, %f11703, %f11835; div.approx.ftz.f32 %f11837, %f11836, %f11695; mul.ftz.f32 %f11838, %f11837, %f11837; mul.ftz.f32 %f11839, %f11837, %f11838; mul.ftz.f32 %f11840, %f11838, %f11838; mul.ftz.f32 %f11841, %f11696, %f11838; fma.rn.ftz.f32 %f11842, %f11695, %f11837, %f11841; mul.ftz.f32 %f11843, %f11698, %f11840; fma.rn.ftz.f32 %f11844, %f11697, %f11839, %f11843; add.ftz.f32 %f11845, %f11842, %f11844; add.ftz.f32 %f11846, %f11694, %f11845; sub.ftz.f32 %f11847, %f11703, %f11846; div.approx.ftz.f32 %f11848, %f11847, %f11695; mul.ftz.f32 %f11849, %f11848, %f11848; mul.ftz.f32 %f11850, %f11848, %f11849; mul.ftz.f32 %f11851, %f11849, %f11849; mul.ftz.f32 %f11852, %f11696, %f11849; fma.rn.ftz.f32 %f11853, %f11695, %f11848, %f11852; mul.ftz.f32 %f11854, %f11698, %f11851; fma.rn.ftz.f32 %f11855, %f11697, %f11850, %f11854; add.ftz.f32 %f11856, %f11853, %f11855; add.ftz.f32 %f11857, %f11694, %f11856; sub.ftz.f32 %f11858, %f11703, %f11857; div.approx.ftz.f32 %f11859, %f11858, %f11695; mul.ftz.f32 %f11860, %f11859, %f11859; mul.ftz.f32 %f11861, %f11859, %f11860; mul.ftz.f32 %f11862, %f11860, %f11860; mul.ftz.f32 %f11863, %f11696, %f11860; fma.rn.ftz.f32 %f11864, %f11695, %f11859, %f11863; mul.ftz.f32 %f11865, %f11698, %f11862; fma.rn.ftz.f32 %f11866, %f11697, %f11861, %f11865; add.ftz.f32 %f11867, %f11864, %f11866; add.ftz.f32 %f11868, %f11694, %f11867; sub.ftz.f32 %f11869, %f11703, %f11868; div.approx.ftz.f32 %f11870, %f11869, %f11695; mul.ftz.f32 %f11871, %f11870, %f11870; mul.ftz.f32 %f11872, %f11870, %f11871; mul.ftz.f32 %f11873, %f11871, %f11871; mul.ftz.f32 %f11874, %f11696, %f11871; fma.rn.ftz.f32 %f11875, %f11695, %f11870, %f11874; mul.ftz.f32 %f11876, %f11698, %f11873; fma.rn.ftz.f32 %f11877, %f11697, %f11872, %f11876; add.ftz.f32 %f11878, %f11875, %f11877; add.ftz.f32 %f11879, %f11694, %f11878; sub.ftz.f32 %f11880, %f11703, %f11879; div.approx.ftz.f32 %f11881, %f11880, %f11695; mul.ftz.f32 %f11882, %f11881, %f11881; mul.ftz.f32 %f11883, %f11881, %f11882; mul.ftz.f32 %f11884, %f11882, %f11882; mul.ftz.f32 %f11885, %f11696, %f11882; fma.rn.ftz.f32 %f11886, %f11695, %f11881, %f11885; mul.ftz.f32 %f11887, %f11698, %f11884; fma.rn.ftz.f32 %f11888, %f11697, %f11883, %f11887; add.ftz.f32 %f11889, %f11886, %f11888; add.ftz.f32 %f11890, %f11694, %f11889; sub.ftz.f32 %f11891, %f11703, %f11890; div.approx.ftz.f32 %f11892, %f11891, %f11695; mul.ftz.f32 %f11893, %f11892, %f11892; mul.ftz.f32 %f11894, %f11892, %f11893; mul.ftz.f32 %f11895, %f11893, %f11893; mul.ftz.f32 %f11896, %f11696, %f11893; fma.rn.ftz.f32 %f11897, %f11695, %f11892, %f11896; mul.ftz.f32 %f11898, %f11698, %f11895; fma.rn.ftz.f32 %f11899, %f11697, %f11894, %f11898; add.ftz.f32 %f11900, %f11897, %f11899; add.ftz.f32 %f11901, %f11694, %f11900; sub.ftz.f32 %f11902, %f11703, %f11901; div.approx.ftz.f32 %f11903, %f11902, %f11695; mul.ftz.f32 %f11904, %f11903, %f11903; mul.ftz.f32 %f11905, %f11903, %f11904; mul.ftz.f32 %f11906, %f11904, %f11904; mul.ftz.f32 %f11907, %f11696, %f11904; fma.rn.ftz.f32 %f11908, %f11695, %f11903, %f11907; mul.ftz.f32 %f11909, %f11698, %f11906; fma.rn.ftz.f32 %f11910, %f11697, %f11905, %f11909; add.ftz.f32 %f11911, %f11908, %f11910; add.ftz.f32 %f11912, %f11694, %f11911; sub.ftz.f32 %f11913, %f11703, %f11912; div.approx.ftz.f32 %f11914, %f11913, %f11695; mul.ftz.f32 %f11915, %f11914, %f11914; mul.ftz.f32 %f11916, %f11914, %f11915; mul.ftz.f32 %f11917, %f11915, %f11915; mul.ftz.f32 %f11918, %f11696, %f11915; fma.rn.ftz.f32 %f11919, %f11695, %f11914, %f11918; mul.ftz.f32 %f11920, %f11698, %f11917; fma.rn.ftz.f32 %f11921, %f11697, %f11916, %f11920; add.ftz.f32 %f11922, %f11919, %f11921; add.ftz.f32 %f11923, %f11694, %f11922; sub.ftz.f32 %f11924, %f11703, %f11923; div.approx.ftz.f32 %f1496, %f11924, %f11695; @%p3 bra $L__BB107_1056; bra.uni $L__BB107_1053; $L__BB107_1056: shr.s32 %r2346, %r502, 31; and.b32 %r2347, %r2346, 1078530011; or.b32 %r2348, %r2347, %r503; mov.b32 %f29257, %r2348; bra.uni $L__BB107_1057; $L__BB107_1044: ld.const.u32 %r5542, [kernel_params+20]; setp.eq.s32 %p936, %r5542, 0; @%p936 bra $L__BB107_1080; ld.const.u32 %r5543, [kernel_params+20]; setp.eq.s32 %p937, %r5543, 1; @%p937 bra $L__BB107_1046; bra.uni $L__BB107_1087; $L__BB107_1046: ld.const.f32 %f1521, [kernel_params+24]; mul.ftz.f32 %f12024, %f1490, %f1490; fma.rn.ftz.f32 %f12025, %f1489, %f1489, %f12024; sqrt.approx.ftz.f32 %f12026, %f12025; abs.ftz.f32 %f1522, %f12026; abs.ftz.f32 %f1523, %f1488; setp.eq.ftz.f32 %p966, %f1523, 0f00000000; setp.eq.ftz.f32 %p967, %f1522, 0f00000000; and.pred %p968, %p966, %p967; mov.b32 %r506, %f1488; mov.b32 %r2363, %f12026; and.b32 %r507, %r2363, -2147483648; @%p968 bra $L__BB107_1073; bra.uni $L__BB107_1047; $L__BB107_1073: shr.s32 %r2368, %r506, 31; and.b32 %r2369, %r2368, 1078530011; or.b32 %r2370, %r507, %r2369; mov.b32 %f29262, %r2370; bra.uni $L__BB107_1074; $L__BB107_1058: add.ftz.f32 %f11957, %f1489, 0fBF800000; mul.ftz.f32 %f11958, %f11957, 0fBF000000; mov.f32 %f11959, 0f00000000; max.ftz.f32 %f11960, %f11958, %f11959; sqrt.approx.ftz.f32 %f11961, %f11960; add.ftz.f32 %f1503, %f11961, %f11961; setp.leu.ftz.f32 %p945, %f1503, 0f00000000; @%p945 bra $L__BB107_1060; rcp.approx.ftz.f32 %f11962, %f1503; mul.ftz.f32 %f1488, %f1488, %f11962; mul.ftz.f32 %f1490, %f1490, %f11962; $L__BB107_1060: add.ftz.f32 %f11963, %f1488, 0f3F800000; add.ftz.f32 %f11964, %f1490, 0f3F800000; mul.ftz.f32 %f29267, %f11964, 0f3F000000; mul.ftz.f32 %f29266, %f11963, 0f3F000000; bra.uni $L__BB107_1093; $L__BB107_1061: abs.ftz.f32 %f1510, %f1488; setp.eq.ftz.f32 %p947, %f1510, 0f00000000; and.pred %p948, %p947, %p930; mov.b32 %r504, %f1488; and.b32 %r505, %r502, -2147483648; @%p948 bra $L__BB107_1065; bra.uni $L__BB107_1062; $L__BB107_1065: shr.s32 %r2353, %r504, 31; and.b32 %r2354, %r2353, 1078530011; or.b32 %r2355, %r2354, %r505; mov.b32 %f29260, %r2355; bra.uni $L__BB107_1066; $L__BB107_1080: setp.eq.ftz.f32 %p983, %f1489, 0f00000000; setp.eq.ftz.f32 %p984, %f1488, 0f00000000; and.pred %p985, %p984, %p983; setp.eq.ftz.f32 %p986, %f1490, 0f00000000; and.pred %p987, %p986, %p985; mov.f32 %f29266, 0f00000000; mov.f32 %f29267, %f29266; @%p987 bra $L__BB107_1093; ld.const.v4.f32 {%f12083, %f12084, %f12085, %f12086}, [kernel_params+32]; abs.ftz.f32 %f1539, %f1488; setp.eq.ftz.f32 %p989, %f1539, 0f00000000; and.pred %p990, %p989, %p930; mov.b32 %r508, %f1488; and.b32 %r509, %r502, -2147483648; @%p990 bra $L__BB107_1085; bra.uni $L__BB107_1082; $L__BB107_1085: shr.s32 %r2382, %r508, 31; and.b32 %r2383, %r2382, 1078530011; or.b32 %r2384, %r2383, %r509; mov.b32 %f29264, %r2384; bra.uni $L__BB107_1086; $L__BB107_1087: ld.const.f32 %f12145, [kernel_params+28]; ld.const.f32 %f1546, [kernel_params+276]; ld.const.f32 %f1547, [kernel_params+280]; add.ftz.f32 %f12146, %f12145, %f12145; mul.ftz.f32 %f12147, %f1491, 0f3F000000; sin.approx.ftz.f32 %f12148, %f12147; mul.ftz.f32 %f1548, %f12146, %f12148; @%p3 bra $L__BB107_1091; bra.uni $L__BB107_1088; $L__BB107_1091: shr.s32 %r2389, %r502, 31; and.b32 %r2390, %r2389, 1078530011; or.b32 %r2391, %r2390, %r503; mov.b32 %f29265, %r2391; bra.uni $L__BB107_1092; $L__BB107_1053: setp.eq.ftz.f32 %p938, %f1492, 0f7F800000; setp.eq.ftz.f32 %p939, %f1493, 0f7F800000; and.pred %p940, %p938, %p939; @%p940 bra $L__BB107_1055; bra.uni $L__BB107_1054; $L__BB107_1055: setp.lt.s32 %p944, %r502, 0; selp.b32 %r2344, 1075235812, 1061752795, %p944; or.b32 %r2345, %r2344, %r503; mov.b32 %f29257, %r2345; bra.uni $L__BB107_1057; $L__BB107_1047: setp.eq.ftz.f32 %p969, %f1523, 0f7F800000; setp.eq.ftz.f32 %p970, %f1522, 0f7F800000; and.pred %p971, %p969, %p970; @%p971 bra $L__BB107_1072; bra.uni $L__BB107_1048; $L__BB107_1072: setp.lt.s32 %p975, %r506, 0; selp.b32 %r2366, 1075235812, 1061752795, %p975; or.b32 %r2367, %r507, %r2366; mov.b32 %f29262, %r2367; bra.uni $L__BB107_1074; $L__BB107_1062: setp.eq.ftz.f32 %p949, %f1510, 0f7F800000; setp.eq.ftz.f32 %p950, %f1492, 0f7F800000; and.pred %p951, %p949, %p950; @%p951 bra $L__BB107_1064; bra.uni $L__BB107_1063; $L__BB107_1064: setp.lt.s32 %p955, %r504, 0; selp.b32 %r2351, 1075235812, 1061752795, %p955; or.b32 %r2352, %r2351, %r505; mov.b32 %f29260, %r2352; bra.uni $L__BB107_1066; $L__BB107_1088: setp.eq.ftz.f32 %p1000, %f1492, 0f7F800000; setp.eq.ftz.f32 %p1001, %f1493, 0f7F800000; and.pred %p1002, %p1000, %p1001; @%p1002 bra $L__BB107_1090; bra.uni $L__BB107_1089; $L__BB107_1090: setp.lt.s32 %p1006, %r502, 0; selp.b32 %r2387, 1075235812, 1061752795, %p1006; or.b32 %r2388, %r2387, %r503; mov.b32 %f29265, %r2388; bra.uni $L__BB107_1092; $L__BB107_1082: setp.eq.ftz.f32 %p991, %f1539, 0f7F800000; setp.eq.ftz.f32 %p992, %f1492, 0f7F800000; and.pred %p993, %p991, %p992; @%p993 bra $L__BB107_1084; bra.uni $L__BB107_1083; $L__BB107_1084: setp.lt.s32 %p997, %r508, 0; selp.b32 %r2380, 1075235812, 1061752795, %p997; or.b32 %r2381, %r2380, %r509; mov.b32 %f29264, %r2381; bra.uni $L__BB107_1086; $L__BB107_1054: mov.f32 %f28747, 0f3FC90FDB; setp.lt.s32 %p941, %r502, 0; min.ftz.f32 %f11925, %f1493, %f1492; max.ftz.f32 %f11926, %f1493, %f1492; div.full.ftz.f32 %f11927, %f11925, %f11926; mul.rn.ftz.f32 %f11928, %f11927, %f11927; mov.f32 %f11929, 0fC0B59883; mov.f32 %f11930, 0fBF52C7EA; fma.rn.ftz.f32 %f11931, %f11928, %f11930, %f11929; mov.f32 %f11932, 0fC0D21907; fma.rn.ftz.f32 %f11933, %f11931, %f11928, %f11932; mul.ftz.f32 %f11934, %f11928, %f11933; mul.ftz.f32 %f11935, %f11927, %f11934; add.ftz.f32 %f11936, %f11928, 0f41355DC0; mov.f32 %f11937, 0f41E6BD60; fma.rn.ftz.f32 %f11938, %f11936, %f11928, %f11937; mov.f32 %f11939, 0f419D92C8; fma.rn.ftz.f32 %f11940, %f11938, %f11928, %f11939; rcp.approx.ftz.f32 %f11941, %f11940; fma.rn.ftz.f32 %f11942, %f11935, %f11941, %f11927; sub.ftz.f32 %f11944, %f28747, %f11942; setp.gt.ftz.f32 %p942, %f1493, %f1492; selp.f32 %f11945, %f11944, %f11942, %p942; sub.ftz.f32 %f11947, %f11138, %f11945; selp.f32 %f11948, %f11947, %f11945, %p941; mov.b32 %r2342, %f11948; or.b32 %r2343, %r503, %r2342; mov.b32 %f11949, %r2343; add.ftz.f32 %f11950, %f1492, %f1493; setp.le.ftz.f32 %p943, %f11950, 0f7F800000; selp.f32 %f29257, %f11949, %f11950, %p943; $L__BB107_1057: ld.const.f32 %f28749, [kernel_params+280]; ld.const.f32 %f28748, [kernel_params+276]; cos.approx.ftz.f32 %f11951, %f29257; mul.ftz.f32 %f11952, %f1496, %f11951; div.approx.ftz.f32 %f11953, %f11952, %f28748; sin.approx.ftz.f32 %f11954, %f29257; mul.ftz.f32 %f11955, %f1496, %f11954; div.approx.ftz.f32 %f11956, %f11955, %f28749; add.ftz.f32 %f29267, %f11956, 0f3F000000; add.ftz.f32 %f29266, %f11953, 0f3F000000; bra.uni $L__BB107_1093; $L__BB107_1048: mov.f32 %f28773, 0f3FC90FDB; setp.lt.s32 %p972, %r506, 0; min.ftz.f32 %f12027, %f1522, %f1523; max.ftz.f32 %f12028, %f1522, %f1523; div.full.ftz.f32 %f12029, %f12027, %f12028; mul.rn.ftz.f32 %f12030, %f12029, %f12029; mov.f32 %f12031, 0fC0B59883; mov.f32 %f12032, 0fBF52C7EA; fma.rn.ftz.f32 %f12033, %f12030, %f12032, %f12031; mov.f32 %f12034, 0fC0D21907; fma.rn.ftz.f32 %f12035, %f12033, %f12030, %f12034; mul.ftz.f32 %f12036, %f12030, %f12035; mul.ftz.f32 %f12037, %f12029, %f12036; add.ftz.f32 %f12038, %f12030, 0f41355DC0; mov.f32 %f12039, 0f41E6BD60; fma.rn.ftz.f32 %f12040, %f12038, %f12030, %f12039; mov.f32 %f12041, 0f419D92C8; fma.rn.ftz.f32 %f12042, %f12040, %f12030, %f12041; rcp.approx.ftz.f32 %f12043, %f12042; fma.rn.ftz.f32 %f12044, %f12037, %f12043, %f12029; sub.ftz.f32 %f12046, %f28773, %f12044; setp.gt.ftz.f32 %p973, %f1522, %f1523; selp.f32 %f12047, %f12046, %f12044, %p973; sub.ftz.f32 %f12049, %f11138, %f12047; selp.f32 %f12050, %f12049, %f12047, %p972; mov.b32 %r2364, %f12050; or.b32 %r2365, %r507, %r2364; mov.b32 %f12051, %r2365; add.ftz.f32 %f12052, %f1523, %f1522; setp.le.ftz.f32 %p974, %f12052, 0f7F800000; selp.f32 %f29262, %f12051, %f12052, %p974; $L__BB107_1074: div.approx.ftz.f32 %f1528, %f29262, %f1521; @%p3 bra $L__BB107_1078; bra.uni $L__BB107_1075; $L__BB107_1078: shr.s32 %r2375, %r502, 31; and.b32 %r2376, %r2375, 1078530011; or.b32 %r2377, %r2376, %r503; mov.b32 %f29263, %r2377; bra.uni $L__BB107_1079; $L__BB107_1075: setp.eq.ftz.f32 %p976, %f1492, 0f7F800000; setp.eq.ftz.f32 %p977, %f1493, 0f7F800000; and.pred %p978, %p976, %p977; @%p978 bra $L__BB107_1077; bra.uni $L__BB107_1076; $L__BB107_1077: setp.lt.s32 %p982, %r502, 0; selp.b32 %r2373, 1075235812, 1061752795, %p982; or.b32 %r2374, %r2373, %r503; mov.b32 %f29263, %r2374; bra.uni $L__BB107_1079; $L__BB107_1063: mov.f32 %f28771, 0f3FC90FDB; setp.lt.s32 %p952, %r504, 0; min.ftz.f32 %f11965, %f1492, %f1510; max.ftz.f32 %f11966, %f1492, %f1510; div.full.ftz.f32 %f11967, %f11965, %f11966; mul.rn.ftz.f32 %f11968, %f11967, %f11967; mov.f32 %f11969, 0fC0B59883; mov.f32 %f11970, 0fBF52C7EA; fma.rn.ftz.f32 %f11971, %f11968, %f11970, %f11969; mov.f32 %f11972, 0fC0D21907; fma.rn.ftz.f32 %f11973, %f11971, %f11968, %f11972; mul.ftz.f32 %f11974, %f11968, %f11973; mul.ftz.f32 %f11975, %f11967, %f11974; add.ftz.f32 %f11976, %f11968, 0f41355DC0; mov.f32 %f11977, 0f41E6BD60; fma.rn.ftz.f32 %f11978, %f11976, %f11968, %f11977; mov.f32 %f11979, 0f419D92C8; fma.rn.ftz.f32 %f11980, %f11978, %f11968, %f11979; rcp.approx.ftz.f32 %f11981, %f11980; fma.rn.ftz.f32 %f11982, %f11975, %f11981, %f11967; sub.ftz.f32 %f11984, %f28771, %f11982; setp.gt.ftz.f32 %p953, %f1492, %f1510; selp.f32 %f11985, %f11984, %f11982, %p953; sub.ftz.f32 %f11987, %f11138, %f11985; selp.f32 %f11988, %f11987, %f11985, %p952; mov.b32 %r2349, %f11988; or.b32 %r2350, %r505, %r2349; mov.b32 %f11989, %r2350; add.ftz.f32 %f11990, %f1510, %f1492; setp.le.ftz.f32 %p954, %f11990, 0f7F800000; selp.f32 %f29260, %f11989, %f11990, %p954; $L__BB107_1066: and.pred %p958, %p947, %p931; @%p958 bra $L__BB107_1070; bra.uni $L__BB107_1067; $L__BB107_1070: shr.s32 %r2360, %r504, 31; and.b32 %r2361, %r2360, 1078530011; or.b32 %r2362, %r503, %r2361; mov.b32 %f29261, %r2362; bra.uni $L__BB107_1071; $L__BB107_1067: setp.eq.ftz.f32 %p959, %f1510, 0f7F800000; setp.eq.ftz.f32 %p960, %f1493, 0f7F800000; and.pred %p961, %p959, %p960; @%p961 bra $L__BB107_1069; bra.uni $L__BB107_1068; $L__BB107_1069: setp.lt.s32 %p965, %r504, 0; selp.b32 %r2358, 1075235812, 1061752795, %p965; or.b32 %r2359, %r503, %r2358; mov.b32 %f29261, %r2359; bra.uni $L__BB107_1071; $L__BB107_1089: mov.f32 %f28770, 0f3FC90FDB; setp.lt.s32 %p1003, %r502, 0; min.ftz.f32 %f12149, %f1493, %f1492; max.ftz.f32 %f12150, %f1493, %f1492; div.full.ftz.f32 %f12151, %f12149, %f12150; mul.rn.ftz.f32 %f12152, %f12151, %f12151; mov.f32 %f12153, 0fC0B59883; mov.f32 %f12154, 0fBF52C7EA; fma.rn.ftz.f32 %f12155, %f12152, %f12154, %f12153; mov.f32 %f12156, 0fC0D21907; fma.rn.ftz.f32 %f12157, %f12155, %f12152, %f12156; mul.ftz.f32 %f12158, %f12152, %f12157; mul.ftz.f32 %f12159, %f12151, %f12158; add.ftz.f32 %f12160, %f12152, 0f41355DC0; mov.f32 %f12161, 0f41E6BD60; fma.rn.ftz.f32 %f12162, %f12160, %f12152, %f12161; mov.f32 %f12163, 0f419D92C8; fma.rn.ftz.f32 %f12164, %f12162, %f12152, %f12163; rcp.approx.ftz.f32 %f12165, %f12164; fma.rn.ftz.f32 %f12166, %f12159, %f12165, %f12151; sub.ftz.f32 %f12168, %f28770, %f12166; setp.gt.ftz.f32 %p1004, %f1493, %f1492; selp.f32 %f12169, %f12168, %f12166, %p1004; sub.ftz.f32 %f12171, %f11138, %f12169; selp.f32 %f12172, %f12171, %f12169, %p1003; mov.b32 %r2385, %f12172; or.b32 %r2386, %r503, %r2385; mov.b32 %f12173, %r2386; add.ftz.f32 %f12174, %f1492, %f1493; setp.le.ftz.f32 %p1005, %f12174, 0f7F800000; selp.f32 %f29265, %f12173, %f12174, %p1005; $L__BB107_1092: cos.approx.ftz.f32 %f12175, %f29265; mul.ftz.f32 %f12176, %f1548, %f12175; div.approx.ftz.f32 %f12177, %f12176, %f1546; sin.approx.ftz.f32 %f12178, %f29265; mul.ftz.f32 %f12179, %f1548, %f12178; div.approx.ftz.f32 %f12180, %f12179, %f1547; add.ftz.f32 %f29267, %f12180, 0f3F000000; add.ftz.f32 %f29266, %f12177, 0f3F000000; bra.uni $L__BB107_1093; $L__BB107_1076: mov.f32 %f28774, 0f3FC90FDB; setp.lt.s32 %p979, %r502, 0; min.ftz.f32 %f12053, %f1493, %f1492; max.ftz.f32 %f12054, %f1493, %f1492; div.full.ftz.f32 %f12055, %f12053, %f12054; mul.rn.ftz.f32 %f12056, %f12055, %f12055; mov.f32 %f12057, 0fC0B59883; mov.f32 %f12058, 0fBF52C7EA; fma.rn.ftz.f32 %f12059, %f12056, %f12058, %f12057; mov.f32 %f12060, 0fC0D21907; fma.rn.ftz.f32 %f12061, %f12059, %f12056, %f12060; mul.ftz.f32 %f12062, %f12056, %f12061; mul.ftz.f32 %f12063, %f12055, %f12062; add.ftz.f32 %f12064, %f12056, 0f41355DC0; mov.f32 %f12065, 0f41E6BD60; fma.rn.ftz.f32 %f12066, %f12064, %f12056, %f12065; mov.f32 %f12067, 0f419D92C8; fma.rn.ftz.f32 %f12068, %f12066, %f12056, %f12067; rcp.approx.ftz.f32 %f12069, %f12068; fma.rn.ftz.f32 %f12070, %f12063, %f12069, %f12055; sub.ftz.f32 %f12072, %f28774, %f12070; setp.gt.ftz.f32 %p980, %f1493, %f1492; selp.f32 %f12073, %f12072, %f12070, %p980; sub.ftz.f32 %f12075, %f11138, %f12073; selp.f32 %f12076, %f12075, %f12073, %p979; mov.b32 %r2371, %f12076; or.b32 %r2372, %r503, %r2371; mov.b32 %f12077, %r2372; add.ftz.f32 %f12078, %f1492, %f1493; setp.le.ftz.f32 %p981, %f12078, 0f7F800000; selp.f32 %f29263, %f12077, %f12078, %p981; $L__BB107_1079: cos.approx.ftz.f32 %f12079, %f29263; sin.approx.ftz.f32 %f12080, %f29263; fma.rn.ftz.f32 %f29267, %f1528, %f12080, 0f3F000000; fma.rn.ftz.f32 %f29266, %f1528, %f12079, 0f3F000000; bra.uni $L__BB107_1093; $L__BB107_1068: mov.f32 %f28772, 0f3FC90FDB; setp.lt.s32 %p962, %r504, 0; min.ftz.f32 %f11991, %f1493, %f1510; max.ftz.f32 %f11992, %f1493, %f1510; div.full.ftz.f32 %f11993, %f11991, %f11992; mul.rn.ftz.f32 %f11994, %f11993, %f11993; mov.f32 %f11995, 0fC0B59883; mov.f32 %f11996, 0fBF52C7EA; fma.rn.ftz.f32 %f11997, %f11994, %f11996, %f11995; mov.f32 %f11998, 0fC0D21907; fma.rn.ftz.f32 %f11999, %f11997, %f11994, %f11998; mul.ftz.f32 %f12000, %f11994, %f11999; mul.ftz.f32 %f12001, %f11993, %f12000; add.ftz.f32 %f12002, %f11994, 0f41355DC0; mov.f32 %f12003, 0f41E6BD60; fma.rn.ftz.f32 %f12004, %f12002, %f11994, %f12003; mov.f32 %f12005, 0f419D92C8; fma.rn.ftz.f32 %f12006, %f12004, %f11994, %f12005; rcp.approx.ftz.f32 %f12007, %f12006; fma.rn.ftz.f32 %f12008, %f12001, %f12007, %f11993; sub.ftz.f32 %f12010, %f28772, %f12008; setp.gt.ftz.f32 %p963, %f1493, %f1510; selp.f32 %f12011, %f12010, %f12008, %p963; sub.ftz.f32 %f12013, %f11138, %f12011; selp.f32 %f12014, %f12013, %f12011, %p962; mov.b32 %r2356, %f12014; or.b32 %r2357, %r503, %r2356; mov.b32 %f12015, %r2357; add.ftz.f32 %f12016, %f1510, %f1493; setp.le.ftz.f32 %p964, %f12016, 0f7F800000; selp.f32 %f29261, %f12015, %f12016, %p964; $L__BB107_1071: mov.f32 %f28740, 0f3F800000; add.ftz.f32 %f12017, %f29261, %f29261; div.approx.ftz.f32 %f12019, %f12017, %f11138; add.ftz.f32 %f12020, %f29260, %f29260; div.approx.ftz.f32 %f12021, %f12020, %f11138; add.ftz.f32 %f12022, %f12021, 0f3F000000; sub.ftz.f32 %f29266, %f28740, %f12022; add.ftz.f32 %f29267, %f12019, 0f3F000000; bra.uni $L__BB107_1093; $L__BB107_1083: mov.f32 %f28775, 0f3FC90FDB; setp.lt.s32 %p994, %r508, 0; min.ftz.f32 %f12087, %f1492, %f1539; max.ftz.f32 %f12088, %f1492, %f1539; div.full.ftz.f32 %f12089, %f12087, %f12088; mul.rn.ftz.f32 %f12090, %f12089, %f12089; mov.f32 %f12091, 0fC0B59883; mov.f32 %f12092, 0fBF52C7EA; fma.rn.ftz.f32 %f12093, %f12090, %f12092, %f12091; mov.f32 %f12094, 0fC0D21907; fma.rn.ftz.f32 %f12095, %f12093, %f12090, %f12094; mul.ftz.f32 %f12096, %f12090, %f12095; mul.ftz.f32 %f12097, %f12089, %f12096; add.ftz.f32 %f12098, %f12090, 0f41355DC0; mov.f32 %f12099, 0f41E6BD60; fma.rn.ftz.f32 %f12100, %f12098, %f12090, %f12099; mov.f32 %f12101, 0f419D92C8; fma.rn.ftz.f32 %f12102, %f12100, %f12090, %f12101; rcp.approx.ftz.f32 %f12103, %f12102; fma.rn.ftz.f32 %f12104, %f12097, %f12103, %f12089; sub.ftz.f32 %f12106, %f28775, %f12104; setp.gt.ftz.f32 %p995, %f1492, %f1539; selp.f32 %f12107, %f12106, %f12104, %p995; sub.ftz.f32 %f12109, %f11138, %f12107; selp.f32 %f12110, %f12109, %f12107, %p994; mov.b32 %r2378, %f12110; or.b32 %r2379, %r509, %r2378; mov.b32 %f12111, %r2379; add.ftz.f32 %f12112, %f1539, %f1492; setp.le.ftz.f32 %p996, %f12112, 0f7F800000; selp.f32 %f29264, %f12111, %f12112, %p996; $L__BB107_1086: mov.f32 %f28776, 0f3FC90FDB; mov.f32 %f28746, 0f3E2AAB94; mov.f32 %f28745, 0f3D994929; mov.f32 %f28744, 0f3D3F841F; mov.f32 %f28743, 0f3C94D2E9; mov.f32 %f28742, 0f3D53F941; mov.f32 %f28741, 0f3F800000; sub.ftz.f32 %f12113, %f29264, %f12084; mul.ftz.f32 %f12114, %f1489, %f1489; fma.rn.ftz.f32 %f12115, %f1488, %f1488, %f12114; fma.rn.ftz.f32 %f12116, %f1490, %f1490, %f12115; sqrt.approx.ftz.f32 %f12117, %f12116; div.approx.ftz.f32 %f12118, %f1490, %f12117; abs.ftz.f32 %f12119, %f12118; sub.ftz.f32 %f12121, %f28741, %f12119; mul.ftz.f32 %f12122, %f12121, 0f3F000000; sqrt.approx.ftz.f32 %f12123, %f12122; setp.gt.ftz.f32 %p998, %f12119, 0f3F11EB85; selp.f32 %f12124, %f12123, %f12119, %p998; mul.ftz.f32 %f12125, %f12124, %f12124; fma.rn.ftz.f32 %f12128, %f28742, %f12125, %f28743; fma.rn.ftz.f32 %f12130, %f12128, %f12125, %f28744; fma.rn.ftz.f32 %f12132, %f12130, %f12125, %f28745; fma.rn.ftz.f32 %f12134, %f12132, %f12125, %f28746; mul.ftz.f32 %f12135, %f12125, %f12134; fma.rn.ftz.f32 %f12136, %f12135, %f12124, %f12124; add.ftz.f32 %f12137, %f12136, %f12136; sub.ftz.f32 %f12139, %f28776, %f12136; selp.f32 %f12140, %f12137, %f12139, %p998; setp.lt.ftz.f32 %p999, %f12118, 0f00000000; sub.ftz.f32 %f12142, %f11138, %f12140; selp.f32 %f12143, %f12142, %f12140, %p999; sub.ftz.f32 %f12144, %f12143, %f12086; div.approx.ftz.f32 %f29267, %f12144, %f12085; div.approx.ftz.f32 %f29266, %f12113, %f12083; $L__BB107_1093: ld.const.u32 %r5544, [kernel_params+20]; setp.gt.s32 %p3918, %r5544, 2; mov.f32 %f28750, 0f3FC90FDB; mov.f32 %f28732, 0fBF800000; mov.f32 %f28731, 0f3E2AAB94; mov.f32 %f28730, 0f3D994929; mov.f32 %f28729, 0f3D3F841F; mov.f32 %f28728, 0f3C94D2E9; mov.f32 %f28727, 0f3D53F941; mov.f32 %f28726, 0f3F800000; ld.const.v4.f32 {%f12181, %f12182, %f12183, %f12184}, [kernel_params+912]; ld.const.v4.f32 {%f12189, %f12190, %f12191, %f12192}, [kernel_params+928]; ld.const.v4.f32 {%f12197, %f12198, %f12199, %f12200}, [kernel_params+944]; mul.ftz.f32 %f12205, %f1379, %f12182; fma.rn.ftz.f32 %f12206, %f1378, %f12181, %f12205; fma.rn.ftz.f32 %f12207, %f1380, %f12183, %f12206; add.ftz.f32 %f12208, %f12184, %f12207; mul.ftz.f32 %f12209, %f1379, %f12190; fma.rn.ftz.f32 %f12210, %f1378, %f12189, %f12209; fma.rn.ftz.f32 %f12211, %f1380, %f12191, %f12210; add.ftz.f32 %f12212, %f12192, %f12211; mul.ftz.f32 %f12213, %f1379, %f12198; fma.rn.ftz.f32 %f12214, %f1378, %f12197, %f12213; fma.rn.ftz.f32 %f12215, %f1380, %f12199, %f12214; add.ftz.f32 %f12216, %f12200, %f12215; mul.ftz.f32 %f12217, %f12212, %f12212; fma.rn.ftz.f32 %f12218, %f12208, %f12208, %f12217; fma.rn.ftz.f32 %f12219, %f12216, %f12216, %f12218; sqrt.approx.ftz.f32 %f12220, %f12219; div.approx.ftz.f32 %f1557, %f12208, %f12220; div.approx.ftz.f32 %f1558, %f12212, %f12220; div.approx.ftz.f32 %f1559, %f12216, %f12220; max.ftz.f32 %f12222, %f1557, %f28732; min.ftz.f32 %f12224, %f12222, %f28726; abs.ftz.f32 %f12225, %f12224; sub.ftz.f32 %f12226, %f28726, %f12225; mul.ftz.f32 %f12227, %f12226, 0f3F000000; sqrt.approx.ftz.f32 %f12228, %f12227; setp.gt.ftz.f32 %p1007, %f12225, 0f3F11EB85; selp.f32 %f12229, %f12228, %f12225, %p1007; mul.ftz.f32 %f12230, %f12229, %f12229; fma.rn.ftz.f32 %f12233, %f28727, %f12230, %f28728; fma.rn.ftz.f32 %f12235, %f12233, %f12230, %f28729; fma.rn.ftz.f32 %f12237, %f12235, %f12230, %f28730; fma.rn.ftz.f32 %f12239, %f12237, %f12230, %f28731; mul.ftz.f32 %f12240, %f12230, %f12239; fma.rn.ftz.f32 %f12241, %f12240, %f12229, %f12229; add.ftz.f32 %f12242, %f12241, %f12241; sub.ftz.f32 %f12244, %f28750, %f12241; selp.f32 %f12245, %f12242, %f12244, %p1007; setp.lt.ftz.f32 %p1008, %f12224, 0f00000000; sub.ftz.f32 %f12247, %f11138, %f12245; selp.f32 %f1560, %f12247, %f12245, %p1008; abs.ftz.f32 %f1561, %f1558; abs.ftz.f32 %f1562, %f1559; setp.eq.ftz.f32 %p1009, %f1561, 0f00000000; setp.eq.ftz.f32 %p1010, %f1562, 0f00000000; and.pred %p4, %p1009, %p1010; mov.b32 %r510, %f1558; mov.b32 %r2392, %f1559; and.b32 %r511, %r2392, -2147483648; @%p3918 bra $L__BB107_1099; bra.uni $L__BB107_1094; $L__BB107_1099: ld.const.u32 %r5533, [kernel_params+20]; setp.eq.s32 %p1012, %r5533, 3; @%p1012 bra $L__BB107_1108; ld.const.u32 %r5534, [kernel_params+20]; setp.eq.s32 %p1013, %r5534, 5; @%p1013 bra $L__BB107_1111; ld.const.u32 %r5535, [kernel_params+20]; setp.ne.s32 %p1014, %r5535, 4; @%p1014 bra $L__BB107_1137; ld.const.f32 %f12248, [kernel_params+48]; ld.const.v4.f32 {%f12249, %f12250, %f12251, %f12252}, [kernel_params+64]; neg.ftz.f32 %f12257, %f1560; sub.ftz.f32 %f12258, %f12257, %f12248; div.approx.ftz.f32 %f12259, %f12258, %f12249; mul.ftz.f32 %f12260, %f12259, %f12259; mul.ftz.f32 %f12261, %f12259, %f12260; mul.ftz.f32 %f12262, %f12260, %f12260; mul.ftz.f32 %f12263, %f12250, %f12260; fma.rn.ftz.f32 %f12264, %f12249, %f12259, %f12263; mul.ftz.f32 %f12265, %f12252, %f12262; fma.rn.ftz.f32 %f12266, %f12251, %f12261, %f12265; add.ftz.f32 %f12267, %f12264, %f12266; add.ftz.f32 %f12268, %f12248, %f12267; sub.ftz.f32 %f12269, %f12257, %f12268; div.approx.ftz.f32 %f12270, %f12269, %f12249; mul.ftz.f32 %f12271, %f12270, %f12270; mul.ftz.f32 %f12272, %f12270, %f12271; mul.ftz.f32 %f12273, %f12271, %f12271; mul.ftz.f32 %f12274, %f12250, %f12271; fma.rn.ftz.f32 %f12275, %f12249, %f12270, %f12274; mul.ftz.f32 %f12276, %f12252, %f12273; fma.rn.ftz.f32 %f12277, %f12251, %f12272, %f12276; add.ftz.f32 %f12278, %f12275, %f12277; add.ftz.f32 %f12279, %f12248, %f12278; sub.ftz.f32 %f12280, %f12257, %f12279; div.approx.ftz.f32 %f12281, %f12280, %f12249; mul.ftz.f32 %f12282, %f12281, %f12281; mul.ftz.f32 %f12283, %f12281, %f12282; mul.ftz.f32 %f12284, %f12282, %f12282; mul.ftz.f32 %f12285, %f12250, %f12282; fma.rn.ftz.f32 %f12286, %f12249, %f12281, %f12285; mul.ftz.f32 %f12287, %f12252, %f12284; fma.rn.ftz.f32 %f12288, %f12251, %f12283, %f12287; add.ftz.f32 %f12289, %f12286, %f12288; add.ftz.f32 %f12290, %f12248, %f12289; sub.ftz.f32 %f12291, %f12257, %f12290; div.approx.ftz.f32 %f12292, %f12291, %f12249; mul.ftz.f32 %f12293, %f12292, %f12292; mul.ftz.f32 %f12294, %f12292, %f12293; mul.ftz.f32 %f12295, %f12293, %f12293; mul.ftz.f32 %f12296, %f12250, %f12293; fma.rn.ftz.f32 %f12297, %f12249, %f12292, %f12296; mul.ftz.f32 %f12298, %f12252, %f12295; fma.rn.ftz.f32 %f12299, %f12251, %f12294, %f12298; add.ftz.f32 %f12300, %f12297, %f12299; add.ftz.f32 %f12301, %f12248, %f12300; sub.ftz.f32 %f12302, %f12257, %f12301; div.approx.ftz.f32 %f12303, %f12302, %f12249; mul.ftz.f32 %f12304, %f12303, %f12303; mul.ftz.f32 %f12305, %f12303, %f12304; mul.ftz.f32 %f12306, %f12304, %f12304; mul.ftz.f32 %f12307, %f12250, %f12304; fma.rn.ftz.f32 %f12308, %f12249, %f12303, %f12307; mul.ftz.f32 %f12309, %f12252, %f12306; fma.rn.ftz.f32 %f12310, %f12251, %f12305, %f12309; add.ftz.f32 %f12311, %f12308, %f12310; add.ftz.f32 %f12312, %f12248, %f12311; sub.ftz.f32 %f12313, %f12257, %f12312; div.approx.ftz.f32 %f12314, %f12313, %f12249; mul.ftz.f32 %f12315, %f12314, %f12314; mul.ftz.f32 %f12316, %f12314, %f12315; mul.ftz.f32 %f12317, %f12315, %f12315; mul.ftz.f32 %f12318, %f12250, %f12315; fma.rn.ftz.f32 %f12319, %f12249, %f12314, %f12318; mul.ftz.f32 %f12320, %f12252, %f12317; fma.rn.ftz.f32 %f12321, %f12251, %f12316, %f12320; add.ftz.f32 %f12322, %f12319, %f12321; add.ftz.f32 %f12323, %f12248, %f12322; sub.ftz.f32 %f12324, %f12257, %f12323; div.approx.ftz.f32 %f12325, %f12324, %f12249; mul.ftz.f32 %f12326, %f12325, %f12325; mul.ftz.f32 %f12327, %f12325, %f12326; mul.ftz.f32 %f12328, %f12326, %f12326; mul.ftz.f32 %f12329, %f12250, %f12326; fma.rn.ftz.f32 %f12330, %f12249, %f12325, %f12329; mul.ftz.f32 %f12331, %f12252, %f12328; fma.rn.ftz.f32 %f12332, %f12251, %f12327, %f12331; add.ftz.f32 %f12333, %f12330, %f12332; add.ftz.f32 %f12334, %f12248, %f12333; sub.ftz.f32 %f12335, %f12257, %f12334; div.approx.ftz.f32 %f12336, %f12335, %f12249; mul.ftz.f32 %f12337, %f12336, %f12336; mul.ftz.f32 %f12338, %f12336, %f12337; mul.ftz.f32 %f12339, %f12337, %f12337; mul.ftz.f32 %f12340, %f12250, %f12337; fma.rn.ftz.f32 %f12341, %f12249, %f12336, %f12340; mul.ftz.f32 %f12342, %f12252, %f12339; fma.rn.ftz.f32 %f12343, %f12251, %f12338, %f12342; add.ftz.f32 %f12344, %f12341, %f12343; add.ftz.f32 %f12345, %f12248, %f12344; sub.ftz.f32 %f12346, %f12257, %f12345; div.approx.ftz.f32 %f12347, %f12346, %f12249; mul.ftz.f32 %f12348, %f12347, %f12347; mul.ftz.f32 %f12349, %f12347, %f12348; mul.ftz.f32 %f12350, %f12348, %f12348; mul.ftz.f32 %f12351, %f12250, %f12348; fma.rn.ftz.f32 %f12352, %f12249, %f12347, %f12351; mul.ftz.f32 %f12353, %f12252, %f12350; fma.rn.ftz.f32 %f12354, %f12251, %f12349, %f12353; add.ftz.f32 %f12355, %f12352, %f12354; add.ftz.f32 %f12356, %f12248, %f12355; sub.ftz.f32 %f12357, %f12257, %f12356; div.approx.ftz.f32 %f12358, %f12357, %f12249; mul.ftz.f32 %f12359, %f12358, %f12358; mul.ftz.f32 %f12360, %f12358, %f12359; mul.ftz.f32 %f12361, %f12359, %f12359; mul.ftz.f32 %f12362, %f12250, %f12359; fma.rn.ftz.f32 %f12363, %f12249, %f12358, %f12362; mul.ftz.f32 %f12364, %f12252, %f12361; fma.rn.ftz.f32 %f12365, %f12251, %f12360, %f12364; add.ftz.f32 %f12366, %f12363, %f12365; add.ftz.f32 %f12367, %f12248, %f12366; sub.ftz.f32 %f12368, %f12257, %f12367; div.approx.ftz.f32 %f12369, %f12368, %f12249; mul.ftz.f32 %f12370, %f12369, %f12369; mul.ftz.f32 %f12371, %f12369, %f12370; mul.ftz.f32 %f12372, %f12370, %f12370; mul.ftz.f32 %f12373, %f12250, %f12370; fma.rn.ftz.f32 %f12374, %f12249, %f12369, %f12373; mul.ftz.f32 %f12375, %f12252, %f12372; fma.rn.ftz.f32 %f12376, %f12251, %f12371, %f12375; add.ftz.f32 %f12377, %f12374, %f12376; add.ftz.f32 %f12378, %f12248, %f12377; sub.ftz.f32 %f12379, %f12257, %f12378; div.approx.ftz.f32 %f12380, %f12379, %f12249; mul.ftz.f32 %f12381, %f12380, %f12380; mul.ftz.f32 %f12382, %f12380, %f12381; mul.ftz.f32 %f12383, %f12381, %f12381; mul.ftz.f32 %f12384, %f12250, %f12381; fma.rn.ftz.f32 %f12385, %f12249, %f12380, %f12384; mul.ftz.f32 %f12386, %f12252, %f12383; fma.rn.ftz.f32 %f12387, %f12251, %f12382, %f12386; add.ftz.f32 %f12388, %f12385, %f12387; add.ftz.f32 %f12389, %f12248, %f12388; sub.ftz.f32 %f12390, %f12257, %f12389; div.approx.ftz.f32 %f12391, %f12390, %f12249; mul.ftz.f32 %f12392, %f12391, %f12391; mul.ftz.f32 %f12393, %f12391, %f12392; mul.ftz.f32 %f12394, %f12392, %f12392; mul.ftz.f32 %f12395, %f12250, %f12392; fma.rn.ftz.f32 %f12396, %f12249, %f12391, %f12395; mul.ftz.f32 %f12397, %f12252, %f12394; fma.rn.ftz.f32 %f12398, %f12251, %f12393, %f12397; add.ftz.f32 %f12399, %f12396, %f12398; add.ftz.f32 %f12400, %f12248, %f12399; sub.ftz.f32 %f12401, %f12257, %f12400; div.approx.ftz.f32 %f12402, %f12401, %f12249; mul.ftz.f32 %f12403, %f12402, %f12402; mul.ftz.f32 %f12404, %f12402, %f12403; mul.ftz.f32 %f12405, %f12403, %f12403; mul.ftz.f32 %f12406, %f12250, %f12403; fma.rn.ftz.f32 %f12407, %f12249, %f12402, %f12406; mul.ftz.f32 %f12408, %f12252, %f12405; fma.rn.ftz.f32 %f12409, %f12251, %f12404, %f12408; add.ftz.f32 %f12410, %f12407, %f12409; add.ftz.f32 %f12411, %f12248, %f12410; sub.ftz.f32 %f12412, %f12257, %f12411; div.approx.ftz.f32 %f12413, %f12412, %f12249; mul.ftz.f32 %f12414, %f12413, %f12413; mul.ftz.f32 %f12415, %f12413, %f12414; mul.ftz.f32 %f12416, %f12414, %f12414; mul.ftz.f32 %f12417, %f12250, %f12414; fma.rn.ftz.f32 %f12418, %f12249, %f12413, %f12417; mul.ftz.f32 %f12419, %f12252, %f12416; fma.rn.ftz.f32 %f12420, %f12251, %f12415, %f12419; add.ftz.f32 %f12421, %f12418, %f12420; add.ftz.f32 %f12422, %f12248, %f12421; sub.ftz.f32 %f12423, %f12257, %f12422; div.approx.ftz.f32 %f12424, %f12423, %f12249; mul.ftz.f32 %f12425, %f12424, %f12424; mul.ftz.f32 %f12426, %f12424, %f12425; mul.ftz.f32 %f12427, %f12425, %f12425; mul.ftz.f32 %f12428, %f12250, %f12425; fma.rn.ftz.f32 %f12429, %f12249, %f12424, %f12428; mul.ftz.f32 %f12430, %f12252, %f12427; fma.rn.ftz.f32 %f12431, %f12251, %f12426, %f12430; add.ftz.f32 %f12432, %f12429, %f12431; add.ftz.f32 %f12433, %f12248, %f12432; sub.ftz.f32 %f12434, %f12257, %f12433; div.approx.ftz.f32 %f12435, %f12434, %f12249; mul.ftz.f32 %f12436, %f12435, %f12435; mul.ftz.f32 %f12437, %f12435, %f12436; mul.ftz.f32 %f12438, %f12436, %f12436; mul.ftz.f32 %f12439, %f12250, %f12436; fma.rn.ftz.f32 %f12440, %f12249, %f12435, %f12439; mul.ftz.f32 %f12441, %f12252, %f12438; fma.rn.ftz.f32 %f12442, %f12251, %f12437, %f12441; add.ftz.f32 %f12443, %f12440, %f12442; add.ftz.f32 %f12444, %f12248, %f12443; sub.ftz.f32 %f12445, %f12257, %f12444; div.approx.ftz.f32 %f12446, %f12445, %f12249; mul.ftz.f32 %f12447, %f12446, %f12446; mul.ftz.f32 %f12448, %f12446, %f12447; mul.ftz.f32 %f12449, %f12447, %f12447; mul.ftz.f32 %f12450, %f12250, %f12447; fma.rn.ftz.f32 %f12451, %f12249, %f12446, %f12450; mul.ftz.f32 %f12452, %f12252, %f12449; fma.rn.ftz.f32 %f12453, %f12251, %f12448, %f12452; add.ftz.f32 %f12454, %f12451, %f12453; add.ftz.f32 %f12455, %f12248, %f12454; sub.ftz.f32 %f12456, %f12257, %f12455; div.approx.ftz.f32 %f12457, %f12456, %f12249; mul.ftz.f32 %f12458, %f12457, %f12457; mul.ftz.f32 %f12459, %f12457, %f12458; mul.ftz.f32 %f12460, %f12458, %f12458; mul.ftz.f32 %f12461, %f12250, %f12458; fma.rn.ftz.f32 %f12462, %f12249, %f12457, %f12461; mul.ftz.f32 %f12463, %f12252, %f12460; fma.rn.ftz.f32 %f12464, %f12251, %f12459, %f12463; add.ftz.f32 %f12465, %f12462, %f12464; add.ftz.f32 %f12466, %f12248, %f12465; sub.ftz.f32 %f12467, %f12257, %f12466; div.approx.ftz.f32 %f12468, %f12467, %f12249; mul.ftz.f32 %f12469, %f12468, %f12468; mul.ftz.f32 %f12470, %f12468, %f12469; mul.ftz.f32 %f12471, %f12469, %f12469; mul.ftz.f32 %f12472, %f12250, %f12469; fma.rn.ftz.f32 %f12473, %f12249, %f12468, %f12472; mul.ftz.f32 %f12474, %f12252, %f12471; fma.rn.ftz.f32 %f12475, %f12251, %f12470, %f12474; add.ftz.f32 %f12476, %f12473, %f12475; add.ftz.f32 %f12477, %f12248, %f12476; sub.ftz.f32 %f12478, %f12257, %f12477; div.approx.ftz.f32 %f1565, %f12478, %f12249; @%p4 bra $L__BB107_1106; bra.uni $L__BB107_1103; $L__BB107_1106: shr.s32 %r2397, %r510, 31; and.b32 %r2398, %r2397, 1078530011; or.b32 %r2399, %r2398, %r511; mov.b32 %f29268, %r2399; bra.uni $L__BB107_1107; $L__BB107_1094: ld.const.u32 %r5540, [kernel_params+20]; setp.eq.s32 %p1015, %r5540, 0; @%p1015 bra $L__BB107_1130; ld.const.u32 %r5541, [kernel_params+20]; setp.eq.s32 %p1016, %r5541, 1; @%p1016 bra $L__BB107_1096; bra.uni $L__BB107_1137; $L__BB107_1096: ld.const.f32 %f1590, [kernel_params+24]; mul.ftz.f32 %f12578, %f1559, %f1559; fma.rn.ftz.f32 %f12579, %f1558, %f1558, %f12578; sqrt.approx.ftz.f32 %f12580, %f12579; abs.ftz.f32 %f1591, %f12580; abs.ftz.f32 %f1592, %f1557; setp.eq.ftz.f32 %p1045, %f1592, 0f00000000; setp.eq.ftz.f32 %p1046, %f1591, 0f00000000; and.pred %p1047, %p1045, %p1046; mov.b32 %r514, %f1557; mov.b32 %r2414, %f12580; and.b32 %r515, %r2414, -2147483648; @%p1047 bra $L__BB107_1123; bra.uni $L__BB107_1097; $L__BB107_1123: shr.s32 %r2419, %r514, 31; and.b32 %r2420, %r2419, 1078530011; or.b32 %r2421, %r515, %r2420; mov.b32 %f29273, %r2421; bra.uni $L__BB107_1124; $L__BB107_1108: add.ftz.f32 %f12511, %f1558, 0fBF800000; mul.ftz.f32 %f12512, %f12511, 0fBF000000; mov.f32 %f12513, 0f00000000; max.ftz.f32 %f12514, %f12512, %f12513; sqrt.approx.ftz.f32 %f12515, %f12514; add.ftz.f32 %f1572, %f12515, %f12515; setp.leu.ftz.f32 %p1024, %f1572, 0f00000000; @%p1024 bra $L__BB107_1110; rcp.approx.ftz.f32 %f12516, %f1572; mul.ftz.f32 %f1557, %f1557, %f12516; mul.ftz.f32 %f1559, %f1559, %f12516; $L__BB107_1110: add.ftz.f32 %f12517, %f1557, 0f3F800000; add.ftz.f32 %f12518, %f1559, 0f3F800000; mul.ftz.f32 %f29278, %f12518, 0f3F000000; mul.ftz.f32 %f29277, %f12517, 0f3F000000; bra.uni $L__BB107_1143; $L__BB107_1111: abs.ftz.f32 %f1579, %f1557; setp.eq.ftz.f32 %p1026, %f1579, 0f00000000; and.pred %p1027, %p1026, %p1009; mov.b32 %r512, %f1557; and.b32 %r513, %r510, -2147483648; @%p1027 bra $L__BB107_1115; bra.uni $L__BB107_1112; $L__BB107_1115: shr.s32 %r2404, %r512, 31; and.b32 %r2405, %r2404, 1078530011; or.b32 %r2406, %r2405, %r513; mov.b32 %f29271, %r2406; bra.uni $L__BB107_1116; $L__BB107_1130: setp.eq.ftz.f32 %p1062, %f1558, 0f00000000; setp.eq.ftz.f32 %p1063, %f1557, 0f00000000; and.pred %p1064, %p1063, %p1062; setp.eq.ftz.f32 %p1065, %f1559, 0f00000000; and.pred %p1066, %p1065, %p1064; mov.f32 %f29277, 0f00000000; mov.f32 %f29278, %f29277; @%p1066 bra $L__BB107_1143; ld.const.v4.f32 {%f12637, %f12638, %f12639, %f12640}, [kernel_params+32]; abs.ftz.f32 %f1608, %f1557; setp.eq.ftz.f32 %p1068, %f1608, 0f00000000; and.pred %p1069, %p1068, %p1009; mov.b32 %r516, %f1557; and.b32 %r517, %r510, -2147483648; @%p1069 bra $L__BB107_1135; bra.uni $L__BB107_1132; $L__BB107_1135: shr.s32 %r2433, %r516, 31; and.b32 %r2434, %r2433, 1078530011; or.b32 %r2435, %r2434, %r517; mov.b32 %f29275, %r2435; bra.uni $L__BB107_1136; $L__BB107_1137: ld.const.f32 %f12699, [kernel_params+28]; ld.const.f32 %f1615, [kernel_params+276]; ld.const.f32 %f1616, [kernel_params+280]; add.ftz.f32 %f12700, %f12699, %f12699; mul.ftz.f32 %f12701, %f1560, 0f3F000000; sin.approx.ftz.f32 %f12702, %f12701; mul.ftz.f32 %f1617, %f12700, %f12702; @%p4 bra $L__BB107_1141; bra.uni $L__BB107_1138; $L__BB107_1141: shr.s32 %r2440, %r510, 31; and.b32 %r2441, %r2440, 1078530011; or.b32 %r2442, %r2441, %r511; mov.b32 %f29276, %r2442; bra.uni $L__BB107_1142; $L__BB107_1103: setp.eq.ftz.f32 %p1017, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1018, %f1562, 0f7F800000; and.pred %p1019, %p1017, %p1018; @%p1019 bra $L__BB107_1105; bra.uni $L__BB107_1104; $L__BB107_1105: setp.lt.s32 %p1023, %r510, 0; selp.b32 %r2395, 1075235812, 1061752795, %p1023; or.b32 %r2396, %r2395, %r511; mov.b32 %f29268, %r2396; bra.uni $L__BB107_1107; $L__BB107_1097: setp.eq.ftz.f32 %p1048, %f1592, 0f7F800000; setp.eq.ftz.f32 %p1049, %f1591, 0f7F800000; and.pred %p1050, %p1048, %p1049; @%p1050 bra $L__BB107_1122; bra.uni $L__BB107_1098; $L__BB107_1122: setp.lt.s32 %p1054, %r514, 0; selp.b32 %r2417, 1075235812, 1061752795, %p1054; or.b32 %r2418, %r515, %r2417; mov.b32 %f29273, %r2418; bra.uni $L__BB107_1124; $L__BB107_1112: setp.eq.ftz.f32 %p1028, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1029, %f1561, 0f7F800000; and.pred %p1030, %p1028, %p1029; @%p1030 bra $L__BB107_1114; bra.uni $L__BB107_1113; $L__BB107_1114: setp.lt.s32 %p1034, %r512, 0; selp.b32 %r2402, 1075235812, 1061752795, %p1034; or.b32 %r2403, %r2402, %r513; mov.b32 %f29271, %r2403; bra.uni $L__BB107_1116; $L__BB107_1138: setp.eq.ftz.f32 %p1079, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1080, %f1562, 0f7F800000; and.pred %p1081, %p1079, %p1080; @%p1081 bra $L__BB107_1140; bra.uni $L__BB107_1139; $L__BB107_1140: setp.lt.s32 %p1085, %r510, 0; selp.b32 %r2438, 1075235812, 1061752795, %p1085; or.b32 %r2439, %r2438, %r511; mov.b32 %f29276, %r2439; bra.uni $L__BB107_1142; $L__BB107_1132: setp.eq.ftz.f32 %p1070, %f1608, 0f7F800000; setp.eq.ftz.f32 %p1071, %f1561, 0f7F800000; and.pred %p1072, %p1070, %p1071; @%p1072 bra $L__BB107_1134; bra.uni $L__BB107_1133; $L__BB107_1134: setp.lt.s32 %p1076, %r516, 0; selp.b32 %r2431, 1075235812, 1061752795, %p1076; or.b32 %r2432, %r2431, %r517; mov.b32 %f29275, %r2432; bra.uni $L__BB107_1136; $L__BB107_1104: mov.f32 %f28761, 0f40490FDB; mov.f32 %f28754, 0f3FC90FDB; setp.lt.s32 %p1020, %r510, 0; min.ftz.f32 %f12479, %f1562, %f1561; max.ftz.f32 %f12480, %f1562, %f1561; div.full.ftz.f32 %f12481, %f12479, %f12480; mul.rn.ftz.f32 %f12482, %f12481, %f12481; mov.f32 %f12483, 0fC0B59883; mov.f32 %f12484, 0fBF52C7EA; fma.rn.ftz.f32 %f12485, %f12482, %f12484, %f12483; mov.f32 %f12486, 0fC0D21907; fma.rn.ftz.f32 %f12487, %f12485, %f12482, %f12486; mul.ftz.f32 %f12488, %f12482, %f12487; mul.ftz.f32 %f12489, %f12481, %f12488; add.ftz.f32 %f12490, %f12482, 0f41355DC0; mov.f32 %f12491, 0f41E6BD60; fma.rn.ftz.f32 %f12492, %f12490, %f12482, %f12491; mov.f32 %f12493, 0f419D92C8; fma.rn.ftz.f32 %f12494, %f12492, %f12482, %f12493; rcp.approx.ftz.f32 %f12495, %f12494; fma.rn.ftz.f32 %f12496, %f12489, %f12495, %f12481; sub.ftz.f32 %f12498, %f28754, %f12496; setp.gt.ftz.f32 %p1021, %f1562, %f1561; selp.f32 %f12499, %f12498, %f12496, %p1021; sub.ftz.f32 %f12501, %f28761, %f12499; selp.f32 %f12502, %f12501, %f12499, %p1020; mov.b32 %r2393, %f12502; or.b32 %r2394, %r511, %r2393; mov.b32 %f12503, %r2394; add.ftz.f32 %f12504, %f1561, %f1562; setp.le.ftz.f32 %p1022, %f12504, 0f7F800000; selp.f32 %f29268, %f12503, %f12504, %p1022; $L__BB107_1107: ld.const.f32 %f28763, [kernel_params+280]; ld.const.f32 %f28762, [kernel_params+276]; cos.approx.ftz.f32 %f12505, %f29268; mul.ftz.f32 %f12506, %f1565, %f12505; div.approx.ftz.f32 %f12507, %f12506, %f28762; sin.approx.ftz.f32 %f12508, %f29268; mul.ftz.f32 %f12509, %f1565, %f12508; div.approx.ftz.f32 %f12510, %f12509, %f28763; add.ftz.f32 %f29278, %f12510, 0f3F000000; add.ftz.f32 %f29277, %f12507, 0f3F000000; bra.uni $L__BB107_1143; $L__BB107_1098: mov.f32 %f28757, 0f3FC90FDB; setp.lt.s32 %p1051, %r514, 0; min.ftz.f32 %f12581, %f1591, %f1592; max.ftz.f32 %f12582, %f1591, %f1592; div.full.ftz.f32 %f12583, %f12581, %f12582; mul.rn.ftz.f32 %f12584, %f12583, %f12583; mov.f32 %f12585, 0fC0B59883; mov.f32 %f12586, 0fBF52C7EA; fma.rn.ftz.f32 %f12587, %f12584, %f12586, %f12585; mov.f32 %f12588, 0fC0D21907; fma.rn.ftz.f32 %f12589, %f12587, %f12584, %f12588; mul.ftz.f32 %f12590, %f12584, %f12589; mul.ftz.f32 %f12591, %f12583, %f12590; add.ftz.f32 %f12592, %f12584, 0f41355DC0; mov.f32 %f12593, 0f41E6BD60; fma.rn.ftz.f32 %f12594, %f12592, %f12584, %f12593; mov.f32 %f12595, 0f419D92C8; fma.rn.ftz.f32 %f12596, %f12594, %f12584, %f12595; rcp.approx.ftz.f32 %f12597, %f12596; fma.rn.ftz.f32 %f12598, %f12591, %f12597, %f12583; sub.ftz.f32 %f12600, %f28757, %f12598; setp.gt.ftz.f32 %p1052, %f1591, %f1592; selp.f32 %f12601, %f12600, %f12598, %p1052; sub.ftz.f32 %f12603, %f11138, %f12601; selp.f32 %f12604, %f12603, %f12601, %p1051; mov.b32 %r2415, %f12604; or.b32 %r2416, %r515, %r2415; mov.b32 %f12605, %r2416; add.ftz.f32 %f12606, %f1592, %f1591; setp.le.ftz.f32 %p1053, %f12606, 0f7F800000; selp.f32 %f29273, %f12605, %f12606, %p1053; $L__BB107_1124: div.approx.ftz.f32 %f1597, %f29273, %f1590; @%p4 bra $L__BB107_1128; bra.uni $L__BB107_1125; $L__BB107_1128: shr.s32 %r2426, %r510, 31; and.b32 %r2427, %r2426, 1078530011; or.b32 %r2428, %r2427, %r511; mov.b32 %f29274, %r2428; bra.uni $L__BB107_1129; $L__BB107_1125: setp.eq.ftz.f32 %p1055, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1056, %f1562, 0f7F800000; and.pred %p1057, %p1055, %p1056; @%p1057 bra $L__BB107_1127; bra.uni $L__BB107_1126; $L__BB107_1127: setp.lt.s32 %p1061, %r510, 0; selp.b32 %r2424, 1075235812, 1061752795, %p1061; or.b32 %r2425, %r2424, %r511; mov.b32 %f29274, %r2425; bra.uni $L__BB107_1129; $L__BB107_1113: mov.f32 %f28755, 0f3FC90FDB; setp.lt.s32 %p1031, %r512, 0; min.ftz.f32 %f12519, %f1561, %f1579; max.ftz.f32 %f12520, %f1561, %f1579; div.full.ftz.f32 %f12521, %f12519, %f12520; mul.rn.ftz.f32 %f12522, %f12521, %f12521; mov.f32 %f12523, 0fC0B59883; mov.f32 %f12524, 0fBF52C7EA; fma.rn.ftz.f32 %f12525, %f12522, %f12524, %f12523; mov.f32 %f12526, 0fC0D21907; fma.rn.ftz.f32 %f12527, %f12525, %f12522, %f12526; mul.ftz.f32 %f12528, %f12522, %f12527; mul.ftz.f32 %f12529, %f12521, %f12528; add.ftz.f32 %f12530, %f12522, 0f41355DC0; mov.f32 %f12531, 0f41E6BD60; fma.rn.ftz.f32 %f12532, %f12530, %f12522, %f12531; mov.f32 %f12533, 0f419D92C8; fma.rn.ftz.f32 %f12534, %f12532, %f12522, %f12533; rcp.approx.ftz.f32 %f12535, %f12534; fma.rn.ftz.f32 %f12536, %f12529, %f12535, %f12521; sub.ftz.f32 %f12538, %f28755, %f12536; setp.gt.ftz.f32 %p1032, %f1561, %f1579; selp.f32 %f12539, %f12538, %f12536, %p1032; sub.ftz.f32 %f12541, %f11138, %f12539; selp.f32 %f12542, %f12541, %f12539, %p1031; mov.b32 %r2400, %f12542; or.b32 %r2401, %r513, %r2400; mov.b32 %f12543, %r2401; add.ftz.f32 %f12544, %f1579, %f1561; setp.le.ftz.f32 %p1033, %f12544, 0f7F800000; selp.f32 %f29271, %f12543, %f12544, %p1033; $L__BB107_1116: and.pred %p1037, %p1026, %p1010; @%p1037 bra $L__BB107_1120; bra.uni $L__BB107_1117; $L__BB107_1120: shr.s32 %r2411, %r512, 31; and.b32 %r2412, %r2411, 1078530011; or.b32 %r2413, %r511, %r2412; mov.b32 %f29272, %r2413; bra.uni $L__BB107_1121; $L__BB107_1117: setp.eq.ftz.f32 %p1038, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1039, %f1562, 0f7F800000; and.pred %p1040, %p1038, %p1039; @%p1040 bra $L__BB107_1119; bra.uni $L__BB107_1118; $L__BB107_1119: setp.lt.s32 %p1044, %r512, 0; selp.b32 %r2409, 1075235812, 1061752795, %p1044; or.b32 %r2410, %r511, %r2409; mov.b32 %f29272, %r2410; bra.uni $L__BB107_1121; $L__BB107_1139: mov.f32 %f28751, 0f3FC90FDB; setp.lt.s32 %p1082, %r510, 0; min.ftz.f32 %f12703, %f1562, %f1561; max.ftz.f32 %f12704, %f1562, %f1561; div.full.ftz.f32 %f12705, %f12703, %f12704; mul.rn.ftz.f32 %f12706, %f12705, %f12705; mov.f32 %f12707, 0fC0B59883; mov.f32 %f12708, 0fBF52C7EA; fma.rn.ftz.f32 %f12709, %f12706, %f12708, %f12707; mov.f32 %f12710, 0fC0D21907; fma.rn.ftz.f32 %f12711, %f12709, %f12706, %f12710; mul.ftz.f32 %f12712, %f12706, %f12711; mul.ftz.f32 %f12713, %f12705, %f12712; add.ftz.f32 %f12714, %f12706, 0f41355DC0; mov.f32 %f12715, 0f41E6BD60; fma.rn.ftz.f32 %f12716, %f12714, %f12706, %f12715; mov.f32 %f12717, 0f419D92C8; fma.rn.ftz.f32 %f12718, %f12716, %f12706, %f12717; rcp.approx.ftz.f32 %f12719, %f12718; fma.rn.ftz.f32 %f12720, %f12713, %f12719, %f12705; sub.ftz.f32 %f12722, %f28751, %f12720; setp.gt.ftz.f32 %p1083, %f1562, %f1561; selp.f32 %f12723, %f12722, %f12720, %p1083; sub.ftz.f32 %f12725, %f11138, %f12723; selp.f32 %f12726, %f12725, %f12723, %p1082; mov.b32 %r2436, %f12726; or.b32 %r2437, %r511, %r2436; mov.b32 %f12727, %r2437; add.ftz.f32 %f12728, %f1561, %f1562; setp.le.ftz.f32 %p1084, %f12728, 0f7F800000; selp.f32 %f29276, %f12727, %f12728, %p1084; $L__BB107_1142: cos.approx.ftz.f32 %f12729, %f29276; mul.ftz.f32 %f12730, %f1617, %f12729; div.approx.ftz.f32 %f12731, %f12730, %f1615; sin.approx.ftz.f32 %f12732, %f29276; mul.ftz.f32 %f12733, %f1617, %f12732; div.approx.ftz.f32 %f12734, %f12733, %f1616; add.ftz.f32 %f29278, %f12734, 0f3F000000; add.ftz.f32 %f29277, %f12731, 0f3F000000; bra.uni $L__BB107_1143; $L__BB107_1126: mov.f32 %f28758, 0f3FC90FDB; setp.lt.s32 %p1058, %r510, 0; min.ftz.f32 %f12607, %f1562, %f1561; max.ftz.f32 %f12608, %f1562, %f1561; div.full.ftz.f32 %f12609, %f12607, %f12608; mul.rn.ftz.f32 %f12610, %f12609, %f12609; mov.f32 %f12611, 0fC0B59883; mov.f32 %f12612, 0fBF52C7EA; fma.rn.ftz.f32 %f12613, %f12610, %f12612, %f12611; mov.f32 %f12614, 0fC0D21907; fma.rn.ftz.f32 %f12615, %f12613, %f12610, %f12614; mul.ftz.f32 %f12616, %f12610, %f12615; mul.ftz.f32 %f12617, %f12609, %f12616; add.ftz.f32 %f12618, %f12610, 0f41355DC0; mov.f32 %f12619, 0f41E6BD60; fma.rn.ftz.f32 %f12620, %f12618, %f12610, %f12619; mov.f32 %f12621, 0f419D92C8; fma.rn.ftz.f32 %f12622, %f12620, %f12610, %f12621; rcp.approx.ftz.f32 %f12623, %f12622; fma.rn.ftz.f32 %f12624, %f12617, %f12623, %f12609; sub.ftz.f32 %f12626, %f28758, %f12624; setp.gt.ftz.f32 %p1059, %f1562, %f1561; selp.f32 %f12627, %f12626, %f12624, %p1059; sub.ftz.f32 %f12629, %f11138, %f12627; selp.f32 %f12630, %f12629, %f12627, %p1058; mov.b32 %r2422, %f12630; or.b32 %r2423, %r511, %r2422; mov.b32 %f12631, %r2423; add.ftz.f32 %f12632, %f1561, %f1562; setp.le.ftz.f32 %p1060, %f12632, 0f7F800000; selp.f32 %f29274, %f12631, %f12632, %p1060; $L__BB107_1129: cos.approx.ftz.f32 %f12633, %f29274; sin.approx.ftz.f32 %f12634, %f29274; fma.rn.ftz.f32 %f29278, %f1597, %f12634, 0f3F000000; fma.rn.ftz.f32 %f29277, %f1597, %f12633, 0f3F000000; bra.uni $L__BB107_1143; $L__BB107_1118: mov.f32 %f28756, 0f3FC90FDB; setp.lt.s32 %p1041, %r512, 0; min.ftz.f32 %f12545, %f1562, %f1579; max.ftz.f32 %f12546, %f1562, %f1579; div.full.ftz.f32 %f12547, %f12545, %f12546; mul.rn.ftz.f32 %f12548, %f12547, %f12547; mov.f32 %f12549, 0fC0B59883; mov.f32 %f12550, 0fBF52C7EA; fma.rn.ftz.f32 %f12551, %f12548, %f12550, %f12549; mov.f32 %f12552, 0fC0D21907; fma.rn.ftz.f32 %f12553, %f12551, %f12548, %f12552; mul.ftz.f32 %f12554, %f12548, %f12553; mul.ftz.f32 %f12555, %f12547, %f12554; add.ftz.f32 %f12556, %f12548, 0f41355DC0; mov.f32 %f12557, 0f41E6BD60; fma.rn.ftz.f32 %f12558, %f12556, %f12548, %f12557; mov.f32 %f12559, 0f419D92C8; fma.rn.ftz.f32 %f12560, %f12558, %f12548, %f12559; rcp.approx.ftz.f32 %f12561, %f12560; fma.rn.ftz.f32 %f12562, %f12555, %f12561, %f12547; sub.ftz.f32 %f12564, %f28756, %f12562; setp.gt.ftz.f32 %p1042, %f1562, %f1579; selp.f32 %f12565, %f12564, %f12562, %p1042; sub.ftz.f32 %f12567, %f11138, %f12565; selp.f32 %f12568, %f12567, %f12565, %p1041; mov.b32 %r2407, %f12568; or.b32 %r2408, %r511, %r2407; mov.b32 %f12569, %r2408; add.ftz.f32 %f12570, %f1579, %f1562; setp.le.ftz.f32 %p1043, %f12570, 0f7F800000; selp.f32 %f29272, %f12569, %f12570, %p1043; $L__BB107_1121: mov.f32 %f28733, 0f3F800000; add.ftz.f32 %f12571, %f29272, %f29272; div.approx.ftz.f32 %f12573, %f12571, %f11138; add.ftz.f32 %f12574, %f29271, %f29271; div.approx.ftz.f32 %f12575, %f12574, %f11138; add.ftz.f32 %f12576, %f12575, 0f3F000000; sub.ftz.f32 %f29277, %f28733, %f12576; add.ftz.f32 %f29278, %f12573, 0f3F000000; bra.uni $L__BB107_1143; $L__BB107_1133: mov.f32 %f28759, 0f3FC90FDB; setp.lt.s32 %p1073, %r516, 0; min.ftz.f32 %f12641, %f1561, %f1608; max.ftz.f32 %f12642, %f1561, %f1608; div.full.ftz.f32 %f12643, %f12641, %f12642; mul.rn.ftz.f32 %f12644, %f12643, %f12643; mov.f32 %f12645, 0fC0B59883; mov.f32 %f12646, 0fBF52C7EA; fma.rn.ftz.f32 %f12647, %f12644, %f12646, %f12645; mov.f32 %f12648, 0fC0D21907; fma.rn.ftz.f32 %f12649, %f12647, %f12644, %f12648; mul.ftz.f32 %f12650, %f12644, %f12649; mul.ftz.f32 %f12651, %f12643, %f12650; add.ftz.f32 %f12652, %f12644, 0f41355DC0; mov.f32 %f12653, 0f41E6BD60; fma.rn.ftz.f32 %f12654, %f12652, %f12644, %f12653; mov.f32 %f12655, 0f419D92C8; fma.rn.ftz.f32 %f12656, %f12654, %f12644, %f12655; rcp.approx.ftz.f32 %f12657, %f12656; fma.rn.ftz.f32 %f12658, %f12651, %f12657, %f12643; sub.ftz.f32 %f12660, %f28759, %f12658; setp.gt.ftz.f32 %p1074, %f1561, %f1608; selp.f32 %f12661, %f12660, %f12658, %p1074; sub.ftz.f32 %f12663, %f11138, %f12661; selp.f32 %f12664, %f12663, %f12661, %p1073; mov.b32 %r2429, %f12664; or.b32 %r2430, %r517, %r2429; mov.b32 %f12665, %r2430; add.ftz.f32 %f12666, %f1608, %f1561; setp.le.ftz.f32 %p1075, %f12666, 0f7F800000; selp.f32 %f29275, %f12665, %f12666, %p1075; $L__BB107_1136: mov.f32 %f28760, 0f3FC90FDB; mov.f32 %f28739, 0f3E2AAB94; mov.f32 %f28738, 0f3D994929; mov.f32 %f28737, 0f3D3F841F; mov.f32 %f28736, 0f3C94D2E9; mov.f32 %f28735, 0f3D53F941; mov.f32 %f28734, 0f3F800000; sub.ftz.f32 %f12667, %f29275, %f12638; mul.ftz.f32 %f12668, %f1558, %f1558; fma.rn.ftz.f32 %f12669, %f1557, %f1557, %f12668; fma.rn.ftz.f32 %f12670, %f1559, %f1559, %f12669; sqrt.approx.ftz.f32 %f12671, %f12670; div.approx.ftz.f32 %f12672, %f1559, %f12671; abs.ftz.f32 %f12673, %f12672; sub.ftz.f32 %f12675, %f28734, %f12673; mul.ftz.f32 %f12676, %f12675, 0f3F000000; sqrt.approx.ftz.f32 %f12677, %f12676; setp.gt.ftz.f32 %p1077, %f12673, 0f3F11EB85; selp.f32 %f12678, %f12677, %f12673, %p1077; mul.ftz.f32 %f12679, %f12678, %f12678; fma.rn.ftz.f32 %f12682, %f28735, %f12679, %f28736; fma.rn.ftz.f32 %f12684, %f12682, %f12679, %f28737; fma.rn.ftz.f32 %f12686, %f12684, %f12679, %f28738; fma.rn.ftz.f32 %f12688, %f12686, %f12679, %f28739; mul.ftz.f32 %f12689, %f12679, %f12688; fma.rn.ftz.f32 %f12690, %f12689, %f12678, %f12678; add.ftz.f32 %f12691, %f12690, %f12690; sub.ftz.f32 %f12693, %f28760, %f12690; selp.f32 %f12694, %f12691, %f12693, %p1077; setp.lt.ftz.f32 %p1078, %f12672, 0f00000000; sub.ftz.f32 %f12696, %f11138, %f12694; selp.f32 %f12697, %f12696, %f12694, %p1078; sub.ftz.f32 %f12698, %f12697, %f12640; div.approx.ftz.f32 %f29278, %f12698, %f12639; div.approx.ftz.f32 %f29277, %f12667, %f12637; $L__BB107_1143: ld.const.f32 %f28753, [kernel_params+288]; ld.const.f32 %f28752, [kernel_params+284]; mul.ftz.f32 %f29281, %f29277, %f28752; mul.ftz.f32 %f29282, %f29278, %f28753; mul.ftz.f32 %f29279, %f29255, %f28752; mul.ftz.f32 %f29280, %f29256, %f28753; mul.ftz.f32 %f29284, %f29267, %f28753; mul.ftz.f32 %f29283, %f29266, %f28752; $L__BB107_1144: ld.const.s32 %rd1998, [kernel_params+1288]; add.s64 %rd1999, %rd164, %rd1998; cvta.to.global.u64 %rd2000, %rd1059; shl.b64 %rd2001, %rd1999, 2; add.s64 %rd2002, %rd2000, %rd2001; add.s64 %rd2003, %rd2002, 4; add.s64 %rd2004, %rd2002, 8; add.s64 %rd2005, %rd2002, 12; sub.ftz.f32 %f12735, %f29283, %f29279; atom.global.add.f32 %f12736, [%rd2002], %f12735; sub.ftz.f32 %f12737, %f29284, %f29280; atom.global.add.f32 %f12738, [%rd2003], %f12737; sub.ftz.f32 %f12739, %f29279, %f29281; atom.global.add.f32 %f12740, [%rd2004], %f12739; sub.ftz.f32 %f12741, %f29280, %f29282; atom.global.add.f32 %f12742, [%rd2005], %f12741; ld.const.s32 %rd2006, [kernel_params+1292]; add.s64 %rd2007, %rd164, %rd2006; shl.b64 %rd2008, %rd2007, 2; add.s64 %rd2009, %rd2000, %rd2008; atom.global.add.f32 %f12743, [%rd2009], 0f3F800000; $L__BB107_1145: ld.global.u32 %r2443, [%rd27]; or.b32 %r2444, %r2443, 8192; st.global.u32 [%rd27], %r2444; $L__BB107_1146: ld.const.u32 %r518, [kernel_params+1388]; setp.eq.s32 %p1086, %r518, 0; @%p1086 bra $L__BB107_1184; ld.const.u64 %rd2010, [kernel_params+2584]; cvta.to.global.u64 %rd2011, %rd2010; mul.lo.s64 %rd2012, %rd3, 12; add.s64 %rd2013, %rd2011, %rd2012; ld.global.f32 %f12747, [%rd2013+4]; ld.global.f32 %f12748, [%rd2013]; add.ftz.f32 %f12749, %f12748, %f12747; ld.global.f32 %f12750, [%rd2013+8]; add.ftz.f32 %f1638, %f12750, %f12749; ld.local.u32 %r519, [%rd5+-36]; and.b32 %r2445, %r519, 524288; setp.ne.s32 %p1087, %r2445, 0; mov.f32 %f12746, 0f3F800000; mov.f32 %f29285, %f12746; mov.f32 %f29286, %f12746; mov.f32 %f29287, %f12746; @%p1087 bra $L__BB107_1150; and.b32 %r2446, %r519, 512; setp.eq.s32 %p1088, %r2446, 0; mov.f32 %f29285, 0f00000000; mov.f32 %f29286, %f29285; mov.f32 %f29287, %f29285; @%p1088 bra $L__BB107_1150; ld.local.v2.f32 {%f29286, %f29287}, [%rd5+208]; ld.local.f32 %f29285, [%rd5+204]; $L__BB107_1150: add.ftz.f32 %f12756, %f29285, %f29286; add.ftz.f32 %f12757, %f29287, %f12756; mul.ftz.f32 %f12758, %f12757, 0f3EAAAAAB; sub.ftz.f32 %f12760, %f12746, %f12758; mul.ftz.f32 %f12761, %f1638, 0f3EAAAAAB; mul.ftz.f32 %f1645, %f12761, %f12760; setp.leu.ftz.f32 %p1089, %f1645, 0f00000000; @%p1089 bra $L__BB107_1184; ld.const.s32 %rd2014, [kernel_params+1396]; add.s64 %rd2015, %rd164, %rd2014; cvta.to.global.u64 %rd2016, %rd1059; shl.b64 %rd2017, %rd2015, 2; add.s64 %rd4789, %rd2016, %rd2017; and.b32 %r2447, %r518, 1; setp.eq.b32 %p1090, %r2447, 1; mov.pred %p1091, 0; xor.pred %p1092, %p1090, %p1091; not.pred %p1093, %p1092; @%p1093 bra $L__BB107_1163; ld.local.u32 %r520, [%rd5+-12]; setp.eq.s32 %p1094, %r520, -1; mov.f32 %f1647, 0f00000000; mov.u32 %r522, 0; @%p1094 bra $L__BB107_1154; ld.const.u64 %rd2018, [kernel_params+2160]; cvta.to.global.u64 %rd2019, %rd2018; mul.wide.s32 %rd2020, %r520, 208; add.s64 %rd2021, %rd2019, %rd2020; ld.global.f32 %f1647, [%rd2021+172]; mov.b32 %r522, %f1647; $L__BB107_1154: ld.const.u32 %r523, [kernel_params+1392]; shl.b32 %r524, %r523, 1; setp.lt.s32 %p1095, %r524, 1; setp.eq.ftz.f32 %p1096, %f1645, 0f00000000; or.pred %p1097, %p1096, %p1095; @%p1097 bra $L__BB107_1162; add.s32 %r525, %r524, -1; mov.u32 %r526, 0; $L__BB107_1156: mul.wide.s32 %rd2022, %r526, 8; add.s64 %rd253, %rd4789, %rd2022; ld.global.f32 %f1648, [%rd253]; setp.eq.ftz.f32 %p1098, %f1648, 0f00000000; @%p1098 bra $L__BB107_1159; bra.uni $L__BB107_1157; $L__BB107_1159: shl.b32 %r2450, %r526, 1; mul.wide.s32 %rd2024, %r2450, 4; add.s64 %rd2025, %rd4789, %rd2024; mov.u32 %r2451, 0; atom.global.cas.b32 %r2452, [%rd2025], %r2451, %r522; mov.b32 %f12764, %r2452; setp.neu.ftz.f32 %p1102, %f12764, 0f00000000; setp.neu.ftz.f32 %p1103, %f1647, %f12764; and.pred %p1104, %p1102, %p1103; @%p1104 bra $L__BB107_1161; bra.uni $L__BB107_1160; $L__BB107_1157: setp.neu.ftz.f32 %p1099, %f1648, %f1647; setp.ne.s32 %p1100, %r526, %r525; and.pred %p1101, %p1100, %p1099; @%p1101 bra $L__BB107_1161; bra.uni $L__BB107_1158; $L__BB107_1161: add.s32 %r526, %r526, 1; setp.lt.s32 %p1105, %r526, %r524; @%p1105 bra $L__BB107_1156; bra.uni $L__BB107_1162; $L__BB107_1160: add.s64 %rd2026, %rd253, 4; atom.global.add.f32 %f12765, [%rd2026], %f1645; bra.uni $L__BB107_1162; $L__BB107_1158: add.s64 %rd2023, %rd253, 4; atom.global.add.f32 %f12763, [%rd2023], %f1645; $L__BB107_1162: mul.wide.s32 %rd2027, %r523, 4; shl.b64 %rd2028, %rd2027, 2; add.s64 %rd4789, %rd4789, %rd2028; $L__BB107_1163: and.b32 %r2453, %r518, 2; setp.eq.s32 %p1106, %r2453, 0; @%p1106 bra $L__BB107_1173; ld.const.u64 %rd4719, [kernel_params+2408]; cvta.to.global.u64 %rd4718, %rd4719; ld.local.u32 %rd2029, [%rd5+-40]; shl.b64 %rd2030, %rd2029, 5; and.b64 %rd2031, %rd2030, 134217696; add.s64 %rd2032, %rd4718, %rd2031; ld.global.f32 %f1649, [%rd2032+12]; ld.const.u32 %r528, [kernel_params+1392]; shl.b32 %r529, %r528, 1; setp.lt.s32 %p1107, %r529, 1; setp.eq.ftz.f32 %p1108, %f1645, 0f00000000; or.pred %p1109, %p1108, %p1107; @%p1109 bra $L__BB107_1172; mov.b32 %r530, %f1649; add.s32 %r531, %r529, -1; mov.u32 %r532, 0; $L__BB107_1166: mul.wide.s32 %rd2033, %r532, 8; add.s64 %rd256, %rd4789, %rd2033; ld.global.f32 %f1650, [%rd256]; setp.eq.ftz.f32 %p1110, %f1650, 0f00000000; @%p1110 bra $L__BB107_1169; bra.uni $L__BB107_1167; $L__BB107_1169: mov.u32 %r2455, 0; atom.global.cas.b32 %r2456, [%rd256], %r2455, %r530; mov.b32 %f12767, %r2456; setp.neu.ftz.f32 %p1114, %f12767, 0f00000000; setp.neu.ftz.f32 %p1115, %f1649, %f12767; and.pred %p1116, %p1114, %p1115; @%p1116 bra $L__BB107_1171; bra.uni $L__BB107_1170; $L__BB107_1167: setp.neu.ftz.f32 %p1111, %f1650, %f1649; setp.ne.s32 %p1112, %r532, %r531; and.pred %p1113, %p1112, %p1111; @%p1113 bra $L__BB107_1171; bra.uni $L__BB107_1168; $L__BB107_1171: add.s32 %r532, %r532, 1; setp.lt.s32 %p1117, %r532, %r529; @%p1117 bra $L__BB107_1166; bra.uni $L__BB107_1172; $L__BB107_1170: add.s64 %rd2035, %rd256, 4; atom.global.add.f32 %f12768, [%rd2035], %f1645; bra.uni $L__BB107_1172; $L__BB107_1168: add.s64 %rd2034, %rd256, 4; atom.global.add.f32 %f12766, [%rd2034], %f1645; $L__BB107_1172: mul.wide.s32 %rd2036, %r528, 4; shl.b64 %rd2037, %rd2036, 2; add.s64 %rd4789, %rd4789, %rd2037; $L__BB107_1173: and.b32 %r2457, %r518, 4; setp.eq.s32 %p1118, %r2457, 0; @%p1118 bra $L__BB107_1184; ld.local.u32 %r534, [%rd5+-12]; setp.eq.s32 %p1119, %r534, -1; mov.f32 %f1652, 0f00000000; mov.u32 %r536, 0; @%p1119 bra $L__BB107_1176; ld.const.u64 %rd2038, [kernel_params+2160]; cvta.to.global.u64 %rd2039, %rd2038; mul.wide.s32 %rd2040, %r534, 208; add.s64 %rd2041, %rd2039, %rd2040; ld.global.f32 %f1652, [%rd2041+176]; mov.b32 %r536, %f1652; $L__BB107_1176: ld.const.u32 %r2459, [kernel_params+1392]; shl.b32 %r537, %r2459, 1; setp.lt.s32 %p1120, %r537, 1; setp.eq.ftz.f32 %p1121, %f1645, 0f00000000; or.pred %p1122, %p1121, %p1120; @%p1122 bra $L__BB107_1184; add.s32 %r538, %r537, -1; mov.u32 %r539, 0; $L__BB107_1178: mul.wide.s32 %rd2042, %r539, 8; add.s64 %rd259, %rd4789, %rd2042; ld.global.f32 %f1653, [%rd259]; setp.eq.ftz.f32 %p1123, %f1653, 0f00000000; @%p1123 bra $L__BB107_1181; bra.uni $L__BB107_1179; $L__BB107_1181: mov.u32 %r2461, 0; atom.global.cas.b32 %r2462, [%rd259], %r2461, %r536; mov.b32 %f12771, %r2462; setp.neu.ftz.f32 %p1127, %f12771, 0f00000000; setp.neu.ftz.f32 %p1128, %f1652, %f12771; and.pred %p1129, %p1127, %p1128; @%p1129 bra $L__BB107_1183; bra.uni $L__BB107_1182; $L__BB107_1179: setp.neu.ftz.f32 %p1124, %f1653, %f1652; setp.ne.s32 %p1125, %r539, %r538; and.pred %p1126, %p1125, %p1124; @%p1126 bra $L__BB107_1183; bra.uni $L__BB107_1180; $L__BB107_1183: add.s32 %r539, %r539, 1; setp.lt.s32 %p1130, %r539, %r537; @%p1130 bra $L__BB107_1178; bra.uni $L__BB107_1184; $L__BB107_1182: add.s64 %rd2044, %rd259, 4; atom.global.add.f32 %f12772, [%rd2044], %f1645; bra.uni $L__BB107_1184; $L__BB107_1180: add.s64 %rd2043, %rd259, 4; atom.global.add.f32 %f12770, [%rd2043], %f1645; $L__BB107_1184: ld.const.u32 %r5536, [kernel_params+1256]; and.b32 %r2463, %r5536, 16384; setp.eq.s32 %p1131, %r2463, 0; @%p1131 bra $L__BB107_1207; ld.const.u64 %rd2045, [kernel_params+2584]; cvta.to.global.u64 %rd2046, %rd2045; mul.lo.s64 %rd2047, %rd3, 12; add.s64 %rd2048, %rd2046, %rd2047; ld.global.f32 %f1654, [%rd2048]; ld.global.f32 %f1655, [%rd2048+4]; ld.global.f32 %f1656, [%rd2048+8]; ld.const.u32 %r541, [kernel_params+1308]; ld.local.u32 %r542, [%rd5+172]; setp.lt.s32 %p1132, %r542, 1; mov.f32 %f29292, 0f00000000; mov.f32 %f29291, %f29292; mov.f32 %f29290, %f29292; @%p1132 bra $L__BB107_1206; add.s32 %r2465, %r542, -1; and.b32 %r5694, %r542, 3; setp.lt.u32 %p1133, %r2465, 3; mov.f32 %f29290, 0f00000000; mov.u32 %r5692, 0; mov.f32 %f29291, %f29290; mov.f32 %f29292, %f29290; @%p1133 bra $L__BB107_1201; sub.s32 %r5691, %r542, %r5694; mov.f32 %f29290, 0f00000000; mov.u32 %r5692, 0; mov.f32 %f29291, %f29290; mov.f32 %f29292, %f29290; $L__BB107_1188: cvt.s64.s32 %rd260, %r5692; mul.wide.s32 %rd2049, %r5692, 80; add.s64 %rd2050, %rd2, %rd2049; add.s64 %rd261, %rd2050, 316; ld.local.u32 %r547, [%rd2050+316]; add.s32 %r2467, %r547, -2; setp.lt.u32 %p1134, %r2467, 7; @%p1134 bra $L__BB107_1190; add.s32 %r2468, %r547, -32; setp.gt.u32 %p1135, %r2468, 2; @%p1135 bra $L__BB107_1191; $L__BB107_1190: ld.local.v4.f32 {%f12783, %f12784, %f12785, %f12786}, [%rd261+-12]; add.ftz.f32 %f29292, %f29292, %f12783; add.ftz.f32 %f29291, %f29291, %f12784; add.ftz.f32 %f29290, %f29290, %f12785; $L__BB107_1191: ld.local.u32 %r548, [%rd261+80]; add.s32 %r2469, %r548, -2; setp.lt.u32 %p1136, %r2469, 7; @%p1136 bra $L__BB107_1193; add.s32 %r2470, %r548, -32; setp.gt.u32 %p1137, %r2470, 2; @%p1137 bra $L__BB107_1194; $L__BB107_1193: ld.local.v4.f32 {%f12790, %f12791, %f12792, %f12793}, [%rd261+68]; add.ftz.f32 %f29292, %f29292, %f12790; add.ftz.f32 %f29291, %f29291, %f12791; add.ftz.f32 %f29290, %f29290, %f12792; $L__BB107_1194: ld.local.u32 %r549, [%rd261+160]; add.s32 %r2471, %r549, -2; setp.lt.u32 %p1138, %r2471, 7; @%p1138 bra $L__BB107_1196; add.s32 %r2472, %r549, -32; setp.gt.u32 %p1139, %r2472, 2; @%p1139 bra $L__BB107_1197; $L__BB107_1196: ld.local.v4.f32 {%f12797, %f12798, %f12799, %f12800}, [%rd261+148]; add.ftz.f32 %f29292, %f29292, %f12797; add.ftz.f32 %f29291, %f29291, %f12798; add.ftz.f32 %f29290, %f29290, %f12799; $L__BB107_1197: ld.local.u32 %r550, [%rd261+240]; add.s32 %r2473, %r550, -2; setp.lt.u32 %p1140, %r2473, 7; @%p1140 bra $L__BB107_1199; add.s32 %r2474, %r550, -32; setp.gt.u32 %p1141, %r2474, 2; @%p1141 bra $L__BB107_1200; $L__BB107_1199: ld.local.v4.f32 {%f12804, %f12805, %f12806, %f12807}, [%rd261+228]; add.ftz.f32 %f29292, %f29292, %f12804; add.ftz.f32 %f29291, %f29291, %f12805; add.ftz.f32 %f29290, %f29290, %f12806; $L__BB107_1200: cvt.u32.u64 %r2475, %rd260; add.s32 %r5692, %r2475, 4; add.s32 %r5691, %r5691, -4; setp.ne.s32 %p1142, %r5691, 0; @%p1142 bra $L__BB107_1188; $L__BB107_1201: setp.eq.s32 %p1143, %r5694, 0; @%p1143 bra $L__BB107_1206; $L__BB107_1202: .pragma "nounroll"; cvt.s64.s32 %rd262, %r5692; mul.wide.s32 %rd2051, %r5692, 80; add.s64 %rd2052, %rd2, %rd2051; add.s64 %rd263, %rd2052, 316; ld.local.u32 %r556, [%rd2052+316]; add.s32 %r2476, %r556, -2; setp.lt.u32 %p1144, %r2476, 7; @%p1144 bra $L__BB107_1204; add.s32 %r2477, %r556, -32; setp.gt.u32 %p1145, %r2477, 2; @%p1145 bra $L__BB107_1205; $L__BB107_1204: ld.local.f32 %f12811, [%rd263+-12]; add.ftz.f32 %f29292, %f29292, %f12811; ld.local.f32 %f12812, [%rd263+-8]; add.ftz.f32 %f29291, %f29291, %f12812; ld.local.f32 %f12813, [%rd263+-4]; add.ftz.f32 %f29290, %f29290, %f12813; $L__BB107_1205: cvt.u32.u64 %r2478, %rd262; add.s32 %r5692, %r2478, 1; add.s32 %r5694, %r5694, -1; setp.ne.s32 %p1146, %r5694, 0; @%p1146 bra $L__BB107_1202; $L__BB107_1206: cvta.to.global.u64 %rd2053, %rd1059; cvt.s64.s32 %rd2054, %r541; add.s64 %rd2055, %rd164, %rd2054; shl.b64 %rd2056, %rd2055, 2; add.s64 %rd2057, %rd2053, %rd2056; mul.ftz.f32 %f12814, %f1654, %f29292; atom.global.add.f32 %f12815, [%rd2057], %f12814; add.s64 %rd2058, %rd2057, 4; mul.ftz.f32 %f12816, %f1655, %f29291; atom.global.add.f32 %f12817, [%rd2058], %f12816; add.s64 %rd2059, %rd2057, 8; mul.ftz.f32 %f12818, %f1656, %f29290; atom.global.add.f32 %f12819, [%rd2059], %f12818; $L__BB107_1207: ld.const.u32 %r5537, [kernel_params+1256]; cvt.u16.u32 %rs326, %r5537; setp.gt.s16 %p1147, %rs326, -1; @%p1147 bra $L__BB107_1225; ld.const.u64 %rd2060, [kernel_params+2584]; cvta.to.global.u64 %rd2061, %rd2060; mul.lo.s64 %rd2062, %rd3, 12; add.s64 %rd2063, %rd2061, %rd2062; ld.global.f32 %f1702, [%rd2063]; ld.global.f32 %f1703, [%rd2063+4]; ld.global.f32 %f1704, [%rd2063+8]; ld.const.u32 %r559, [kernel_params+1312]; ld.local.u32 %r560, [%rd5+172]; setp.lt.s32 %p1148, %r560, 1; mov.f32 %f29323, 0f00000000; mov.f32 %f29324, %f29323; mov.f32 %f29325, %f29323; @%p1148 bra $L__BB107_1224; add.s32 %r2480, %r560, -1; and.b32 %r5699, %r560, 3; setp.lt.u32 %p1149, %r2480, 3; mov.f32 %f29325, 0f00000000; mov.u32 %r5697, 0; mov.f32 %f29324, %f29325; mov.f32 %f29323, %f29325; @%p1149 bra $L__BB107_1220; sub.s32 %r5696, %r560, %r5699; mov.f32 %f29325, 0f00000000; mov.u32 %r5697, 0; mov.f32 %f29324, %f29325; mov.f32 %f29323, %f29325; $L__BB107_1211: cvt.s64.s32 %rd264, %r5697; mul.wide.s32 %rd2064, %r5697, 80; add.s64 %rd2065, %rd2, %rd2064; add.s64 %rd265, %rd2065, 316; ld.local.u32 %r2482, [%rd2065+316]; add.s32 %r2483, %r2482, -9; setp.gt.u32 %p1150, %r2483, 11; setp.ne.s32 %p1151, %r2482, 29; and.pred %p1152, %p1150, %p1151; @%p1152 bra $L__BB107_1213; ld.local.v4.f32 {%f12830, %f12831, %f12832, %f12833}, [%rd265+-12]; add.ftz.f32 %f29323, %f29323, %f12830; add.ftz.f32 %f29324, %f29324, %f12831; add.ftz.f32 %f29325, %f29325, %f12832; $L__BB107_1213: ld.local.u32 %r2484, [%rd265+80]; add.s32 %r2485, %r2484, -9; setp.gt.u32 %p1153, %r2485, 11; setp.ne.s32 %p1154, %r2484, 29; and.pred %p1155, %p1153, %p1154; @%p1155 bra $L__BB107_1215; ld.local.v4.f32 {%f12837, %f12838, %f12839, %f12840}, [%rd265+68]; add.ftz.f32 %f29323, %f29323, %f12837; add.ftz.f32 %f29324, %f29324, %f12838; add.ftz.f32 %f29325, %f29325, %f12839; $L__BB107_1215: ld.local.u32 %r2486, [%rd265+160]; add.s32 %r2487, %r2486, -9; setp.gt.u32 %p1156, %r2487, 11; setp.ne.s32 %p1157, %r2486, 29; and.pred %p1158, %p1156, %p1157; @%p1158 bra $L__BB107_1217; ld.local.v4.f32 {%f12844, %f12845, %f12846, %f12847}, [%rd265+148]; add.ftz.f32 %f29323, %f29323, %f12844; add.ftz.f32 %f29324, %f29324, %f12845; add.ftz.f32 %f29325, %f29325, %f12846; $L__BB107_1217: ld.local.u32 %r2488, [%rd265+240]; add.s32 %r2489, %r2488, -9; setp.gt.u32 %p1159, %r2489, 11; setp.ne.s32 %p1160, %r2488, 29; and.pred %p1161, %p1159, %p1160; @%p1161 bra $L__BB107_1219; ld.local.v4.f32 {%f12851, %f12852, %f12853, %f12854}, [%rd265+228]; add.ftz.f32 %f29323, %f29323, %f12851; add.ftz.f32 %f29324, %f29324, %f12852; add.ftz.f32 %f29325, %f29325, %f12853; $L__BB107_1219: cvt.u32.u64 %r2490, %rd264; add.s32 %r5697, %r2490, 4; add.s32 %r5696, %r5696, -4; setp.ne.s32 %p1162, %r5696, 0; @%p1162 bra $L__BB107_1211; $L__BB107_1220: setp.eq.s32 %p1163, %r5699, 0; @%p1163 bra $L__BB107_1224; $L__BB107_1221: .pragma "nounroll"; cvt.s64.s32 %rd266, %r5697; mul.wide.s32 %rd2066, %r5697, 80; add.s64 %rd2067, %rd2, %rd2066; add.s64 %rd267, %rd2067, 316; ld.local.u32 %r2491, [%rd2067+316]; add.s32 %r2492, %r2491, -9; setp.gt.u32 %p1164, %r2492, 11; setp.ne.s32 %p1165, %r2491, 29; and.pred %p1166, %p1164, %p1165; @%p1166 bra $L__BB107_1223; ld.local.f32 %f12858, [%rd267+-12]; add.ftz.f32 %f29323, %f29323, %f12858; ld.local.f32 %f12859, [%rd267+-8]; add.ftz.f32 %f29324, %f29324, %f12859; ld.local.f32 %f12860, [%rd267+-4]; add.ftz.f32 %f29325, %f29325, %f12860; $L__BB107_1223: cvt.u32.u64 %r2493, %rd266; add.s32 %r5697, %r2493, 1; add.s32 %r5699, %r5699, -1; setp.ne.s32 %p1167, %r5699, 0; @%p1167 bra $L__BB107_1221; $L__BB107_1224: cvta.to.global.u64 %rd2068, %rd1059; cvt.s64.s32 %rd2069, %r559; add.s64 %rd2070, %rd164, %rd2069; shl.b64 %rd2071, %rd2070, 2; add.s64 %rd2072, %rd2068, %rd2071; mul.ftz.f32 %f12861, %f1702, %f29323; atom.global.add.f32 %f12862, [%rd2072], %f12861; add.s64 %rd2073, %rd2072, 4; mul.ftz.f32 %f12863, %f1703, %f29324; atom.global.add.f32 %f12864, [%rd2073], %f12863; add.s64 %rd2074, %rd2072, 8; mul.ftz.f32 %f12865, %f1704, %f29325; atom.global.add.f32 %f12866, [%rd2074], %f12865; $L__BB107_1225: ld.const.u32 %r5538, [kernel_params+1256]; and.b32 %r2494, %r5538, 65536; setp.eq.s32 %p1168, %r2494, 0; @%p1168 bra $L__BB107_1243; ld.const.u64 %rd2075, [kernel_params+2584]; cvta.to.global.u64 %rd2076, %rd2075; mul.lo.s64 %rd2077, %rd3, 12; add.s64 %rd2078, %rd2076, %rd2077; ld.global.f32 %f1750, [%rd2078]; ld.global.f32 %f1751, [%rd2078+4]; ld.global.f32 %f1752, [%rd2078+8]; ld.const.u32 %r572, [kernel_params+1316]; ld.local.u32 %r573, [%rd5+172]; setp.lt.s32 %p1169, %r573, 1; mov.f32 %f29353, 0f00000000; mov.f32 %f29354, %f29353; mov.f32 %f29355, %f29353; @%p1169 bra $L__BB107_1242; add.s32 %r2496, %r573, -1; and.b32 %r5704, %r573, 3; setp.lt.u32 %p1170, %r2496, 3; mov.f32 %f29355, 0f00000000; mov.u32 %r5702, 0; mov.f32 %f29354, %f29355; mov.f32 %f29353, %f29355; @%p1170 bra $L__BB107_1238; sub.s32 %r5701, %r573, %r5704; mov.f32 %f29355, 0f00000000; mov.u32 %r5702, 0; mov.f32 %f29354, %f29355; mov.f32 %f29353, %f29355; $L__BB107_1229: cvt.s64.s32 %rd268, %r5702; mul.wide.s32 %rd2079, %r5702, 80; add.s64 %rd2080, %rd2, %rd2079; add.s64 %rd269, %rd2080, 316; ld.local.u32 %r2498, [%rd2080+316]; add.s32 %r2499, %r2498, -21; setp.gt.u32 %p1171, %r2499, 9; @%p1171 bra $L__BB107_1231; ld.local.v4.f32 {%f12877, %f12878, %f12879, %f12880}, [%rd269+-12]; add.ftz.f32 %f29353, %f29353, %f12877; add.ftz.f32 %f29354, %f29354, %f12878; add.ftz.f32 %f29355, %f29355, %f12879; $L__BB107_1231: ld.local.u32 %r2500, [%rd269+80]; add.s32 %r2501, %r2500, -21; setp.gt.u32 %p1172, %r2501, 9; @%p1172 bra $L__BB107_1233; ld.local.v4.f32 {%f12884, %f12885, %f12886, %f12887}, [%rd269+68]; add.ftz.f32 %f29353, %f29353, %f12884; add.ftz.f32 %f29354, %f29354, %f12885; add.ftz.f32 %f29355, %f29355, %f12886; $L__BB107_1233: ld.local.u32 %r2502, [%rd269+160]; add.s32 %r2503, %r2502, -21; setp.gt.u32 %p1173, %r2503, 9; @%p1173 bra $L__BB107_1235; ld.local.v4.f32 {%f12891, %f12892, %f12893, %f12894}, [%rd269+148]; add.ftz.f32 %f29353, %f29353, %f12891; add.ftz.f32 %f29354, %f29354, %f12892; add.ftz.f32 %f29355, %f29355, %f12893; $L__BB107_1235: ld.local.u32 %r2504, [%rd269+240]; add.s32 %r2505, %r2504, -21; setp.gt.u32 %p1174, %r2505, 9; @%p1174 bra $L__BB107_1237; ld.local.v4.f32 {%f12898, %f12899, %f12900, %f12901}, [%rd269+228]; add.ftz.f32 %f29353, %f29353, %f12898; add.ftz.f32 %f29354, %f29354, %f12899; add.ftz.f32 %f29355, %f29355, %f12900; $L__BB107_1237: cvt.u32.u64 %r2506, %rd268; add.s32 %r5702, %r2506, 4; add.s32 %r5701, %r5701, -4; setp.ne.s32 %p1175, %r5701, 0; @%p1175 bra $L__BB107_1229; $L__BB107_1238: setp.eq.s32 %p1176, %r5704, 0; @%p1176 bra $L__BB107_1242; $L__BB107_1239: .pragma "nounroll"; cvt.s64.s32 %rd270, %r5702; mul.wide.s32 %rd2081, %r5702, 80; add.s64 %rd2082, %rd2, %rd2081; add.s64 %rd271, %rd2082, 316; ld.local.u32 %r2507, [%rd2082+316]; add.s32 %r2508, %r2507, -21; setp.gt.u32 %p1177, %r2508, 9; @%p1177 bra $L__BB107_1241; ld.local.f32 %f12905, [%rd271+-12]; add.ftz.f32 %f29353, %f29353, %f12905; ld.local.f32 %f12906, [%rd271+-8]; add.ftz.f32 %f29354, %f29354, %f12906; ld.local.f32 %f12907, [%rd271+-4]; add.ftz.f32 %f29355, %f29355, %f12907; $L__BB107_1241: cvt.u32.u64 %r2509, %rd270; add.s32 %r5702, %r2509, 1; add.s32 %r5704, %r5704, -1; setp.ne.s32 %p1178, %r5704, 0; @%p1178 bra $L__BB107_1239; $L__BB107_1242: cvta.to.global.u64 %rd2083, %rd1059; cvt.s64.s32 %rd2084, %r572; add.s64 %rd2085, %rd164, %rd2084; shl.b64 %rd2086, %rd2085, 2; add.s64 %rd2087, %rd2083, %rd2086; mul.ftz.f32 %f12908, %f1750, %f29353; atom.global.add.f32 %f12909, [%rd2087], %f12908; add.s64 %rd2088, %rd2087, 4; mul.ftz.f32 %f12910, %f1751, %f29354; atom.global.add.f32 %f12911, [%rd2088], %f12910; add.s64 %rd2089, %rd2087, 8; mul.ftz.f32 %f12912, %f1752, %f29355; atom.global.add.f32 %f12913, [%rd2089], %f12912; $L__BB107_1243: ld.const.u32 %r5539, [kernel_params+1256]; and.b32 %r2510, %r5539, 131072; setp.eq.s32 %p1179, %r2510, 0; @%p1179 bra $L__BB107_1257; ld.const.f32 %f1798, [kernel_params+1412]; ld.const.f32 %f1799, [kernel_params+1416]; add.u64 %rd2091, %SPL, 3424; ld.local.v4.f32 {%f12914, %f12915, %f12916, %f12917}, [%rd2091]; ld.const.v4.f32 {%f12921, %f12922, %f12923, %f12924}, [kernel_params+96]; ld.const.v4.f32 {%f12925, %f12926, %f12927, %f12928}, [kernel_params+112]; ld.const.v4.f32 {%f12929, %f12930, %f12931, %f12932}, [kernel_params+128]; ld.const.u32 %r2511, [kernel_params+16]; setp.eq.s32 %p1180, %r2511, 1; sub.ftz.f32 %f1812, %f12914, %f12924; sub.ftz.f32 %f1813, %f12915, %f12928; sub.ftz.f32 %f1814, %f12916, %f12932; @%p1180 bra $L__BB107_1246; bra.uni $L__BB107_1245; $L__BB107_1246: mul.ftz.f32 %f12936, %f12927, %f1813; fma.rn.ftz.f32 %f12937, %f12923, %f1812, %f12936; fma.rn.ftz.f32 %f12938, %f12931, %f1814, %f12937; abs.ftz.f32 %f29380, %f12938; bra.uni $L__BB107_1247; $L__BB107_1245: mul.ftz.f32 %f12933, %f1813, %f1813; fma.rn.ftz.f32 %f12934, %f1812, %f1812, %f12933; fma.rn.ftz.f32 %f12935, %f1814, %f1814, %f12934; sqrt.approx.ftz.f32 %f29380, %f12935; $L__BB107_1247: sub.ftz.f32 %f12939, %f29380, %f1798; mul.ftz.f32 %f12940, %f1799, %f12939; cvt.ftz.sat.f32.f32 %f29381, %f12940; ld.const.f32 %f1819, [kernel_params+1420]; setp.eq.ftz.f32 %p1181, %f1819, 0f3F800000; @%p1181 bra $L__BB107_1253; setp.eq.ftz.f32 %p1182, %f1819, 0f40000000; @%p1182 bra $L__BB107_1252; bra.uni $L__BB107_1249; $L__BB107_1252: mul.ftz.f32 %f29381, %f29381, %f29381; bra.uni $L__BB107_1253; $L__BB107_1249: setp.eq.ftz.f32 %p1183, %f1819, 0f3F000000; @%p1183 bra $L__BB107_1251; bra.uni $L__BB107_1250; $L__BB107_1251: sqrt.approx.ftz.f32 %f29381, %f29381; bra.uni $L__BB107_1253; $L__BB107_1250: lg2.approx.ftz.f32 %f12941, %f29381; mul.ftz.f32 %f12942, %f1819, %f12941; ex2.approx.ftz.f32 %f29381, %f12942; $L__BB107_1253: ld.const.u64 %rd2092, [kernel_params+2584]; cvta.to.global.u64 %rd2093, %rd2092; mul.lo.s64 %rd2094, %rd3, 12; add.s64 %rd2095, %rd2093, %rd2094; ld.global.f32 %f1824, [%rd2095]; ld.global.f32 %f1825, [%rd2095+4]; ld.global.f32 %f1826, [%rd2095+8]; ld.local.u32 %r585, [%rd5+-36]; and.b32 %r2512, %r585, 524288; setp.ne.s32 %p1184, %r2512, 0; mov.f32 %f12945, 0f3F800000; mov.f32 %f29382, %f12945; mov.f32 %f29383, %f12945; mov.f32 %f29384, %f12945; @%p1184 bra $L__BB107_1256; and.b32 %r2513, %r585, 512; setp.eq.s32 %p1185, %r2513, 0; mov.f32 %f29382, 0f00000000; mov.f32 %f29383, %f29382; mov.f32 %f29384, %f29382; @%p1185 bra $L__BB107_1256; ld.local.v2.f32 {%f29383, %f29384}, [%rd5+208]; ld.local.f32 %f29382, [%rd5+204]; $L__BB107_1256: sub.ftz.f32 %f12952, %f12945, %f29382; sub.ftz.f32 %f12953, %f12945, %f29383; sub.ftz.f32 %f12954, %f12945, %f29384; cvt.ftz.sat.f32.f32 %f12955, %f12952; cvt.ftz.sat.f32.f32 %f12956, %f12953; cvt.ftz.sat.f32.f32 %f12957, %f12954; mul.ftz.f32 %f12958, %f1825, %f12956; fma.rn.ftz.f32 %f12959, %f1824, %f12955, %f12958; fma.rn.ftz.f32 %f12960, %f1826, %f12957, %f12959; mul.ftz.f32 %f12961, %f12960, 0f3EAAAAAB; sub.ftz.f32 %f12962, %f12945, %f29381; mul.ftz.f32 %f12963, %f12962, %f12961; ld.const.s32 %rd2096, [kernel_params+1408]; add.s64 %rd2097, %rd164, %rd2096; cvta.to.global.u64 %rd2098, %rd1059; shl.b64 %rd2099, %rd2097, 2; add.s64 %rd2100, %rd2098, %rd2099; atom.global.add.f32 %f12964, [%rd2100], %f12963; $L__BB107_1257: ld.global.u32 %r586, [%rd27]; and.b32 %r2514, %r586, 16777216; setp.eq.s32 %p1186, %r2514, 0; @%p1186 bra $L__BB107_1312; ld.local.u8 %rs327, [%rd5+-34]; and.b16 %rs328, %rs327, 8; setp.ne.s16 %p1187, %rs328, 0; @%p1187 bra $L__BB107_1312; ld.const.u64 %rd2101, [kernel_params+2448]; cvta.to.global.u64 %rd2102, %rd2101; add.s64 %rd2104, %rd2102, %rd1282; ld.global.u32 %rd2105, [%rd2104]; ld.const.s32 %rd2106, [kernel_params+1264]; mul.lo.s64 %rd272, %rd2106, %rd2105; ld.const.u32 %r587, [kernel_params+1432]; setp.eq.s32 %p1188, %r587, -1; @%p1188 bra $L__BB107_1264; ld.const.u64 %rd2107, [kernel_params+2616]; cvta.to.global.u64 %rd2108, %rd2107; mul.lo.s64 %rd2109, %rd3, 12; add.s64 %rd2110, %rd2108, %rd2109; ld.const.u64 %rd2111, [kernel_params+2648]; cvta.to.global.u64 %rd2112, %rd2111; add.s64 %rd2114, %rd2112, %rd1282; ld.global.f32 %f12966, [%rd2114]; ld.local.f32 %f12967, [%rd5]; sub.ftz.f32 %f12968, %f12967, %f12966; ld.global.f32 %f12969, [%rd2110+4]; ld.global.f32 %f12970, [%rd2110]; add.ftz.f32 %f12971, %f12970, %f12969; ld.global.f32 %f12972, [%rd2110+8]; add.ftz.f32 %f12973, %f12972, %f12971; mul.ftz.f32 %f12974, %f12973, 0f3EAAAAAB; mul.ftz.f32 %f1833, %f12974, %f12968; mov.b32 %r588, %f1833; setp.nan.ftz.f32 %p1189, %f1833, %f1833; mov.f32 %f29385, 0f00000000; @%p1189 bra $L__BB107_1263; or.b32 %r2515, %r588, -2147483648; setp.ne.s32 %p1190, %r2515, -2147483648; add.ftz.f32 %f12976, %f1833, %f1833; setp.eq.ftz.f32 %p1191, %f12976, %f1833; and.pred %p1192, %p1191, %p1190; @%p1192 bra $L__BB107_1263; shl.b32 %r2516, %r588, 1; setp.lt.u32 %p1193, %r2516, -16777215; selp.f32 %f29385, %f1833, 0f00000000, %p1193; $L__BB107_1263: cvta.to.global.u64 %rd2115, %rd1059; cvt.s64.s32 %rd2116, %r587; add.s64 %rd2117, %rd272, %rd2116; shl.b64 %rd2118, %rd2117, 2; add.s64 %rd2119, %rd2115, %rd2118; atom.global.add.f32 %f12977, [%rd2119], %f29385; $L__BB107_1264: ld.local.u32 %r589, [%rd5+172]; setp.lt.s32 %p1194, %r589, 1; mov.f32 %f29401, 0f00000000; mov.f32 %f29402, %f29401; mov.f32 %f29403, %f29401; mov.f32 %f29404, %f29401; mov.f32 %f29405, %f29401; mov.f32 %f29406, %f29401; mov.f32 %f29407, %f29401; mov.f32 %f29408, %f29401; mov.f32 %f29409, %f29401; mov.f32 %f29410, %f29401; mov.f32 %f29411, %f29401; @%p1194 bra $L__BB107_1285; mov.f32 %f29411, 0f00000000; mov.u32 %r5705, 0; mov.f32 %f29410, %f29411; mov.f32 %f29409, %f29411; mov.f32 %f29408, %f29411; mov.f32 %f29407, %f29411; mov.f32 %f29406, %f29411; mov.f32 %f29405, %f29411; mov.f32 %f29404, %f29411; mov.f32 %f29403, %f29411; mov.f32 %f29402, %f29411; mov.f32 %f29401, %f29411; bra.uni $L__BB107_1266; $L__BB107_1270: setp.eq.s32 %p1198, %r591, 6; @%p1198 bra $L__BB107_1275; bra.uni $L__BB107_1271; $L__BB107_1275: ld.local.f32 %f13039, [%rd274+20]; mul.ftz.f32 %f29397, %f29397, %f13039; mul.ftz.f32 %f29398, %f29398, %f13039; mul.ftz.f32 %f29399, %f29399, %f13039; bra.uni $L__BB107_1276; $L__BB107_1271: setp.eq.s32 %p1199, %r591, 29; @%p1199 bra $L__BB107_1274; bra.uni $L__BB107_1272; $L__BB107_1274: ld.local.v4.f32 {%f13014, %f13015, %f13016, %f13017}, [%rd274+20]; sqrt.approx.ftz.f32 %f13022, %f13014; sqrt.approx.ftz.f32 %f13023, %f13015; sqrt.approx.ftz.f32 %f13024, %f13016; fma.rn.ftz.f32 %f13025, %f13017, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f13026, %f13017, %f13025, 0fC12BAE14; fma.rn.ftz.f32 %f13027, %f13017, %f13026, 0f40220C4A; fma.rn.ftz.f32 %f13028, %f13017, %f13027, 0fBE5C28F6; fma.rn.ftz.f32 %f13029, %f13017, %f13028, 0f40BF020C; mul.ftz.f32 %f13030, %f13022, %f13029; mul.ftz.f32 %f13031, %f13023, %f13029; mul.ftz.f32 %f13032, %f13024, %f13029; mul.ftz.f32 %f13033, %f13030, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13034, %f13033; mul.ftz.f32 %f13035, %f13031, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13036, %f13035; mul.ftz.f32 %f13037, %f13032, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13038, %f13037; mul.ftz.f32 %f29397, %f29397, %f13034; mul.ftz.f32 %f29398, %f29398, %f13036; mul.ftz.f32 %f29399, %f29399, %f13038; bra.uni $L__BB107_1276; $L__BB107_1272: setp.ne.s32 %p1200, %r591, 5; @%p1200 bra $L__BB107_1276; ld.local.u32 %r2522, [%rd274+24]; setp.eq.s32 %p1201, %r2522, 8; @%p1201 bra $L__BB107_1284; bra.uni $L__BB107_1276; $L__BB107_1266: cvt.s64.s32 %rd273, %r5705; mul.wide.s32 %rd2120, %r5705, 80; add.s64 %rd2121, %rd2, %rd2120; add.s64 %rd274, %rd2121, 316; ld.local.u32 %r591, [%rd2121+316]; setp.gt.s32 %p1195, %r591, 34; @%p1195 bra $L__BB107_1284; ld.local.v4.f32 {%f13000, %f13001, %f13002, %f13003}, [%rd274+4]; fma.rn.ftz.f32 %f29407, %f13000, %f13001, %f29407; fma.rn.ftz.f32 %f29408, %f13000, %f13002, %f29408; fma.rn.ftz.f32 %f29409, %f13000, %f13003, %f29409; add.ftz.f32 %f29410, %f29410, %f13000; ld.local.v4.f32 {%f29397, %f29398, %f29399, %f13010}, [%rd274+-12]; setp.gt.u32 %p1196, %r591, 27; @%p1196 bra $L__BB107_1270; mov.u32 %r2519, 1; shl.b32 %r2520, %r2519, %r591; and.b32 %r2521, %r2520, 134256640; setp.ne.s32 %p1197, %r2521, 0; @%p1197 bra $L__BB107_1269; bra.uni $L__BB107_1270; $L__BB107_1269: ld.local.u64 %rd2122, [%rd274+36]; ld.f32 %f13011, [%rd2122+24]; mul.ftz.f32 %f29397, %f29397, %f13011; ld.f32 %f13012, [%rd2122+28]; mul.ftz.f32 %f29398, %f29398, %f13012; ld.f32 %f13013, [%rd2122+32]; mul.ftz.f32 %f29399, %f29399, %f13013; $L__BB107_1276: mov.f32 %f29400, 0f00000000; setp.gt.u32 %p1202, %r591, 31; @%p1202 bra $L__BB107_1278; mov.u32 %r2524, 1; shl.b32 %r2525, %r2524, %r591; and.b32 %r2526, %r2525, -2145385984; setp.ne.s32 %p1203, %r2526, 0; @%p1203 bra $L__BB107_1281; $L__BB107_1278: add.s32 %r2527, %r591, -10; setp.lt.u32 %p1204, %r2527, 7; @%p1204 bra $L__BB107_1280; add.s32 %r2528, %r591, -22; setp.gt.u32 %p1205, %r2528, 2; setp.ne.s32 %p1206, %r591, 27; mov.f32 %f29400, 0f3F800000; and.pred %p1207, %p1205, %p1206; @%p1207 bra $L__BB107_1281; $L__BB107_1280: ld.local.v2.f32 {%f13042, %f13043}, [%rd274+20]; mul.ftz.f32 %f29400, %f13042, %f13043; $L__BB107_1281: setp.gt.ftz.f32 %p1208, %f29400, 0f3BB851EC; @%p1208 bra $L__BB107_1283; bra.uni $L__BB107_1282; $L__BB107_1283: add.ftz.f32 %f29404, %f29404, %f29397; add.ftz.f32 %f29405, %f29405, %f29398; add.ftz.f32 %f29406, %f29406, %f29399; add.ftz.f32 %f29411, %f29411, %f13000; bra.uni $L__BB107_1284; $L__BB107_1282: add.ftz.f32 %f29401, %f29401, %f29397; add.ftz.f32 %f29402, %f29402, %f29398; add.ftz.f32 %f29403, %f29403, %f29399; $L__BB107_1284: cvt.u32.u64 %r2529, %rd273; add.s32 %r5705, %r2529, 1; setp.lt.s32 %p1209, %r5705, %r589; @%p1209 bra $L__BB107_1266; $L__BB107_1285: mul.ftz.f32 %f13046, %f29411, 0f40800000; setp.gt.ftz.f32 %p1210, %f13046, %f29410; setp.eq.ftz.f32 %p1211, %f29410, 0f00000000; or.pred %p1212, %p1211, %p1210; @%p1212 bra $L__BB107_1287; bra.uni $L__BB107_1286; $L__BB107_1287: @%p1211 bra $L__BB107_1289; rcp.approx.ftz.f32 %f13053, %f29410; mul.ftz.f32 %f29407, %f29407, %f13053; mul.ftz.f32 %f29408, %f29408, %f13053; mul.ftz.f32 %f29409, %f29409, %f13053; $L__BB107_1289: ld.const.u32 %r593, [kernel_params+1424]; setp.eq.s32 %p1214, %r593, -1; @%p1214 bra $L__BB107_1300; ld.const.v4.f32 {%f13055, %f13056, %f13057, %f13058}, [kernel_params+688]; ld.const.v4.f32 {%f13062, %f13063, %f13064, %f13065}, [kernel_params+704]; ld.const.v4.f32 {%f13069, %f13070, %f13071, %f13072}, [kernel_params+720]; mul.ftz.f32 %f13076, %f29408, %f13056; fma.rn.ftz.f32 %f13077, %f29407, %f13055, %f13076; fma.rn.ftz.f32 %f1904, %f29409, %f13057, %f13077; mul.ftz.f32 %f13078, %f29408, %f13063; fma.rn.ftz.f32 %f13079, %f29407, %f13062, %f13078; fma.rn.ftz.f32 %f1905, %f29409, %f13064, %f13079; mul.ftz.f32 %f13080, %f29408, %f13070; fma.rn.ftz.f32 %f13081, %f29407, %f13069, %f13080; fma.rn.ftz.f32 %f1906, %f29409, %f13071, %f13081; mov.b32 %r594, %f1904; setp.nan.ftz.f32 %p1215, %f1904, %f1904; mov.f32 %f29428, 0f00000000; mov.f32 %f29426, %f29428; @%p1215 bra $L__BB107_1293; or.b32 %r2530, %r594, -2147483648; setp.ne.s32 %p1216, %r2530, -2147483648; add.ftz.f32 %f13083, %f1904, %f1904; setp.eq.ftz.f32 %p1217, %f13083, %f1904; and.pred %p1218, %p1217, %p1216; mov.f32 %f29426, %f29428; @%p1218 bra $L__BB107_1293; shl.b32 %r2531, %r594, 1; setp.lt.u32 %p1219, %r2531, -16777215; selp.f32 %f29426, %f1904, 0f00000000, %p1219; $L__BB107_1293: mov.b32 %r595, %f1905; setp.nan.ftz.f32 %p1220, %f1905, %f1905; mov.f32 %f29427, %f29428; @%p1220 bra $L__BB107_1296; or.b32 %r2532, %r595, -2147483648; setp.ne.s32 %p1221, %r2532, -2147483648; add.ftz.f32 %f13086, %f1905, %f1905; setp.eq.ftz.f32 %p1222, %f13086, %f1905; and.pred %p1223, %p1222, %p1221; mov.f32 %f29427, %f29428; @%p1223 bra $L__BB107_1296; shl.b32 %r2533, %r595, 1; setp.lt.u32 %p1224, %r2533, -16777215; selp.f32 %f29427, %f1905, 0f00000000, %p1224; $L__BB107_1296: mov.b32 %r596, %f1906; setp.nan.ftz.f32 %p1225, %f1906, %f1906; @%p1225 bra $L__BB107_1299; or.b32 %r2534, %r596, -2147483648; setp.ne.s32 %p1226, %r2534, -2147483648; add.ftz.f32 %f13089, %f1906, %f1906; setp.eq.ftz.f32 %p1227, %f13089, %f1906; and.pred %p1228, %p1227, %p1226; @%p1228 bra $L__BB107_1299; shl.b32 %r2535, %r596, 1; setp.lt.u32 %p1229, %r2535, -16777215; selp.f32 %f29428, %f1906, 0f00000000, %p1229; $L__BB107_1299: cvta.to.global.u64 %rd2127, %rd1059; cvt.s64.s32 %rd2128, %r593; add.s64 %rd2129, %rd272, %rd2128; shl.b64 %rd2130, %rd2129, 2; add.s64 %rd2131, %rd2127, %rd2130; add.s64 %rd2132, %rd2131, 4; add.s64 %rd2133, %rd2131, 8; atom.global.add.f32 %f13090, [%rd2131], %f29426; atom.global.add.f32 %f13091, [%rd2132], %f29427; atom.global.add.f32 %f13092, [%rd2133], %f29428; $L__BB107_1300: ld.const.u32 %r597, [kernel_params+1428]; setp.eq.s32 %p1230, %r597, -1; @%p1230 bra $L__BB107_1311; ld.const.u64 %rd2134, [kernel_params+2616]; cvta.to.global.u64 %rd2135, %rd2134; mul.lo.s64 %rd2136, %rd3, 12; add.s64 %rd2137, %rd2135, %rd2136; ld.global.f32 %f13094, [%rd2137]; mul.ftz.f32 %f1913, %f29404, %f13094; ld.global.f32 %f13095, [%rd2137+4]; mul.ftz.f32 %f1914, %f29405, %f13095; ld.global.f32 %f13096, [%rd2137+8]; mul.ftz.f32 %f1915, %f29406, %f13096; mov.b32 %r598, %f1913; setp.nan.ftz.f32 %p1231, %f1913, %f1913; mov.f32 %f29431, 0f00000000; mov.f32 %f29429, %f29431; @%p1231 bra $L__BB107_1304; or.b32 %r2536, %r598, -2147483648; setp.ne.s32 %p1232, %r2536, -2147483648; add.ftz.f32 %f13098, %f1913, %f1913; setp.eq.ftz.f32 %p1233, %f13098, %f1913; and.pred %p1234, %p1233, %p1232; mov.f32 %f29429, %f29431; @%p1234 bra $L__BB107_1304; shl.b32 %r2537, %r598, 1; setp.lt.u32 %p1235, %r2537, -16777215; selp.f32 %f29429, %f1913, 0f00000000, %p1235; $L__BB107_1304: mov.b32 %r599, %f1914; setp.nan.ftz.f32 %p1236, %f1914, %f1914; mov.f32 %f29430, %f29431; @%p1236 bra $L__BB107_1307; or.b32 %r2538, %r599, -2147483648; setp.ne.s32 %p1237, %r2538, -2147483648; add.ftz.f32 %f13101, %f1914, %f1914; setp.eq.ftz.f32 %p1238, %f13101, %f1914; and.pred %p1239, %p1238, %p1237; mov.f32 %f29430, %f29431; @%p1239 bra $L__BB107_1307; shl.b32 %r2539, %r599, 1; setp.lt.u32 %p1240, %r2539, -16777215; selp.f32 %f29430, %f1914, 0f00000000, %p1240; $L__BB107_1307: mov.b32 %r600, %f1915; setp.nan.ftz.f32 %p1241, %f1915, %f1915; @%p1241 bra $L__BB107_1310; or.b32 %r2540, %r600, -2147483648; setp.ne.s32 %p1242, %r2540, -2147483648; add.ftz.f32 %f13104, %f1915, %f1915; setp.eq.ftz.f32 %p1243, %f13104, %f1915; and.pred %p1244, %p1243, %p1242; @%p1244 bra $L__BB107_1310; shl.b32 %r2541, %r600, 1; setp.lt.u32 %p1245, %r2541, -16777215; selp.f32 %f29431, %f1915, 0f00000000, %p1245; $L__BB107_1310: cvta.to.global.u64 %rd2138, %rd1059; cvt.s64.s32 %rd2139, %r597; add.s64 %rd2140, %rd272, %rd2139; shl.b64 %rd2141, %rd2140, 2; add.s64 %rd2142, %rd2138, %rd2141; add.s64 %rd2143, %rd2142, 4; add.s64 %rd2144, %rd2142, 8; atom.global.add.f32 %f13105, [%rd2142], %f29429; atom.global.add.f32 %f13106, [%rd2143], %f29430; atom.global.add.f32 %f13107, [%rd2144], %f29431; $L__BB107_1311: and.b32 %r2542, %r586, -16777217; st.global.u32 [%rd27], %r2542; bra.uni $L__BB107_1312; $L__BB107_1286: ld.const.u64 %rd2123, [kernel_params+2616]; cvta.to.global.u64 %rd2124, %rd2123; mul.lo.s64 %rd2125, %rd3, 12; add.s64 %rd2126, %rd2124, %rd2125; ld.global.f32 %f13047, [%rd2126]; mul.ftz.f32 %f13048, %f29401, %f13047; ld.global.f32 %f13049, [%rd2126+4]; mul.ftz.f32 %f13050, %f29402, %f13049; ld.global.f32 %f13051, [%rd2126+8]; mul.ftz.f32 %f13052, %f29403, %f13051; st.global.f32 [%rd2126], %f13048; st.global.f32 [%rd2126+4], %f13050; st.global.f32 [%rd2126+8], %f13052; bra.uni $L__BB107_1312; $L__BB107_2652: neg.ftz.f32 %f19722, %f4171; div.approx.ftz.f32 %f19723, %f19722, %f19513; neg.ftz.f32 %f19724, %f4172; div.approx.ftz.f32 %f19725, %f19724, %f19514; fma.rn.ftz.f32 %f19726, %f19723, %f19723, %f4174; fma.rn.ftz.f32 %f19727, %f19725, %f19725, %f19726; mul.ftz.f32 %f19728, %f19727, 0f40490FDB; mul.ftz.f32 %f19729, %f19727, %f19728; mul.ftz.f32 %f19730, %f19513, %f19729; mul.ftz.f32 %f19731, %f19514, %f19730; mov.f32 %f19732, 0f33D6BF95; max.ftz.f32 %f19733, %f19731, %f19732; rcp.approx.ftz.f32 %f30001, %f19733; $L__BB107_2654: add.ftz.f32 %f19743, %f30050, 0f3F800000; mov.f32 %f30026, 0f3F800000; sub.ftz.f32 %f19744, %f30026, %f19743; add.ftz.f32 %f19745, %f19744, %f30000; rcp.approx.ftz.f32 %f19746, %f19745; mul.ftz.f32 %f19747, %f19746, 0f3E800000; div.approx.ftz.f32 %f19748, %f19747, %f29998; mul.ftz.f32 %f30104, %f19748, %f30001; fma.rn.ftz.f32 %f4179, %f4155, %f4155, 0fBF800000; abs.ftz.f32 %f4180, %f30026; fma.rn.ftz.f32 %f4181, %f4180, %f4180, %f4179; setp.leu.ftz.f32 %p2407, %f4181, 0f00000000; mov.f32 %f30002, %f30026; @%p2407 bra $L__BB107_2656; sqrt.approx.ftz.f32 %f19749, %f4181; sub.ftz.f32 %f19750, %f19749, %f4180; add.ftz.f32 %f19751, %f4180, %f19749; div.approx.ftz.f32 %f19752, %f19750, %f19751; fma.rn.ftz.f32 %f19753, %f4180, %f19751, 0fBF800000; fma.rn.ftz.f32 %f19754, %f4180, %f19750, 0f3F800000; div.approx.ftz.f32 %f19755, %f19753, %f19754; mul.ftz.f32 %f19756, %f19752, 0f3F000000; mul.ftz.f32 %f19757, %f19752, %f19756; fma.rn.ftz.f32 %f19758, %f19755, %f19755, 0f3F800000; mul.ftz.f32 %f30002, %f19757, %f19758; $L__BB107_2656: sub.ftz.f32 %f4184, %f30026, %f30002; rcp.approx.ftz.f32 %f4185, %f4184; setp.ne.s32 %p2408, %r963, 15; mov.f32 %f30047, %f30026; mov.f32 %f30048, %f30026; mov.f32 %f30049, %f30026; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2408 bra $L__BB107_2660; mul.ftz.f32 %f19762, %f29997, %f4172; fma.rn.ftz.f32 %f19763, %f29996, %f4171, %f19762; fma.rn.ftz.f32 %f19764, %f29998, %f4173, %f19763; abs.ftz.f32 %f4186, %f19764; fma.rn.ftz.f32 %f4187, %f4186, %f4186, %f4179; setp.leu.ftz.f32 %p2409, %f4187, 0f00000000; mov.f32 %f30003, %f4184; @%p2409 bra $L__BB107_2659; sqrt.approx.ftz.f32 %f19765, %f4187; sub.ftz.f32 %f19766, %f19765, %f4186; add.ftz.f32 %f19767, %f4186, %f19765; div.approx.ftz.f32 %f19768, %f19766, %f19767; fma.rn.ftz.f32 %f19769, %f4186, %f19767, 0fBF800000; fma.rn.ftz.f32 %f19770, %f4186, %f19766, 0f3F800000; div.approx.ftz.f32 %f19771, %f19769, %f19770; mul.ftz.f32 %f19772, %f19768, 0f3F000000; mul.ftz.f32 %f19773, %f19768, %f19772; fma.rn.ftz.f32 %f19774, %f19771, %f19771, 0f3F800000; mul.ftz.f32 %f19775, %f19773, %f19774; sub.ftz.f32 %f30003, %f19775, %f30002; $L__BB107_2659: mul.ftz.f32 %f19776, %f4185, %f30003; mov.f32 %f19777, 0f3F800000; sub.ftz.f32 %f19778, %f19777, %f19776; fma.rn.ftz.f32 %f30047, %f4156, %f19778, %f19776; fma.rn.ftz.f32 %f30048, %f4157, %f19778, %f19776; fma.rn.ftz.f32 %f30049, %f4158, %f19778, %f19776; mul.ftz.f32 %f4193, %f30104, %f30047; mul.ftz.f32 %f30105, %f30104, %f30048; mul.ftz.f32 %f30106, %f30104, %f30049; mov.f32 %f30104, %f4193; $L__BB107_2660: mul.ftz.f32 %f4202, %f19513, %f19513; ld.local.u32 %r5786, [%rd5+168]; selp.f32 %f4203, 0f3F800000, 0f00000000, %p2403; mov.f32 %f30024, 0f00000000; neg.ftz.f32 %f30046, %f29998; mov.u32 %r5785, 0; mov.f32 %f30025, %f30026; $L__BB107_2661: mad.lo.s32 %r5786, %r5786, 1103515245, 12345; cvt.rn.f32.u32 %f19782, %r5786; mul.ftz.f32 %f4218, %f19782, 0f2F800000; setp.gt.ftz.f32 %p2412, %f30046, 0f3F7FF972; mov.pred %p3924, %p2411; @%p2412 bra $L__BB107_2671; bra.uni $L__BB107_2662; $L__BB107_2671: @%p3924 bra $L__BB107_2704; mad.lo.s32 %r4127, %r5786, 1103515245, 12345; cvt.rn.f32.u32 %f19802, %r4127; mul.ftz.f32 %f4229, %f19802, 0f2F800000; mad.lo.s32 %r5786, %r4127, 1103515245, 12345; cvt.rn.f32.u32 %f19803, %r5786; mul.ftz.f32 %f4230, %f19803, 0f2F800000; neg.ftz.f32 %f4231, %f30044; mul.ftz.f32 %f19804, %f19513, %f4231; neg.ftz.f32 %f4232, %f30045; mul.ftz.f32 %f19805, %f19514, %f4232; mul.ftz.f32 %f19806, %f19805, %f19805; fma.rn.ftz.f32 %f19807, %f19804, %f19804, %f19806; fma.rn.ftz.f32 %f19808, %f30046, %f30046, %f19807; sqrt.approx.ftz.f32 %f19809, %f19808; div.approx.ftz.f32 %f30029, %f19804, %f19809; div.approx.ftz.f32 %f30030, %f19805, %f19809; neg.ftz.f32 %f4235, %f30046; div.approx.ftz.f32 %f4236, %f4235, %f19809; setp.gt.ftz.f32 %p2425, %f4236, 0f3F7FF972; @%p2425 bra $L__BB107_2678; abs.ftz.f32 %f19810, %f4236; setp.lt.ftz.f32 %p2426, %f19810, 0f358637BD; @%p2426 bra $L__BB107_2678; bra.uni $L__BB107_2674; $L__BB107_2678: mov.f32 %f19854, 0f3F800000; sub.ftz.f32 %f19855, %f19854, %f4230; mov.f32 %f19856, 0f33D6BF95; max.ftz.f32 %f19857, %f19855, %f19856; div.approx.ftz.f32 %f19858, %f4230, %f19857; sqrt.approx.ftz.f32 %f19859, %f19858; mul.ftz.f32 %f19860, %f4229, 0f40C90FDB; cos.approx.ftz.f32 %f19861, %f19860; sin.approx.ftz.f32 %f19862, %f19860; mul.ftz.f32 %f30028, %f19859, %f19862; mul.ftz.f32 %f30027, %f19859, %f19861; $L__BB107_2679: mul.ftz.f32 %f19863, %f30030, %f30030; fma.rn.ftz.f32 %f19864, %f30029, %f30029, %f19863; add.ftz.f32 %f19865, %f19864, 0f00000000; sqrt.approx.ftz.f32 %f4252, %f19865; setp.neu.ftz.f32 %p2434, %f4252, 0f00000000; @%p2434 bra $L__BB107_2680; bra.uni $L__BB107_2681; $L__BB107_2680: rcp.approx.ftz.f32 %f19866, %f4252; mul.ftz.f32 %f30030, %f30030, %f19866; mul.ftz.f32 %f30029, %f30029, %f19866; $L__BB107_2681: mul.ftz.f32 %f19867, %f30027, %f30029; mul.ftz.f32 %f19868, %f30028, %f30030; sub.ftz.f32 %f19869, %f19867, %f19868; mul.ftz.f32 %f19870, %f19513, %f19869; mul.ftz.f32 %f19871, %f30028, %f30029; fma.rn.ftz.f32 %f19872, %f30027, %f30030, %f19871; mul.ftz.f32 %f19873, %f19514, %f19872; neg.ftz.f32 %f19874, %f19870; neg.ftz.f32 %f19875, %f19873; mul.ftz.f32 %f19876, %f19873, %f19873; fma.rn.ftz.f32 %f19877, %f19870, %f19870, %f19876; add.ftz.f32 %f19878, %f19877, 0f3F800000; mov.f32 %f19879, 0f3F800000; sqrt.approx.ftz.f32 %f19880, %f19878; div.approx.ftz.f32 %f4262, %f19874, %f19880; div.approx.ftz.f32 %f4263, %f19875, %f19880; div.approx.ftz.f32 %f4264, %f19879, %f19880; setp.eq.s32 %p2435, %r5785, 0; @%p2435 bra $L__BB107_2692; mov.f32 %f30032, 0f00000000; @%p2412 bra $L__BB107_2689; sub.ftz.f32 %f19887, %f29993, %f30044; sub.ftz.f32 %f19888, %f29994, %f30045; mul.ftz.f32 %f19889, %f19888, %f19888; fma.rn.ftz.f32 %f19890, %f19887, %f19887, %f19889; sub.ftz.f32 %f19891, %f29995, %f30046; fma.rn.ftz.f32 %f19892, %f19891, %f19891, %f19890; sqrt.approx.ftz.f32 %f19893, %f19892; div.approx.ftz.f32 %f4265, %f19887, %f19893; div.approx.ftz.f32 %f4266, %f19888, %f19893; div.approx.ftz.f32 %f4267, %f19891, %f19893; setp.lt.ftz.f32 %p2437, %f4267, 0f00000000; @%p2437 bra $L__BB107_2689; mul.ftz.f32 %f19897, %f4266, %f4232; fma.rn.ftz.f32 %f19898, %f4265, %f4231, %f19897; fma.rn.ftz.f32 %f4268, %f4267, %f4235, %f19898; setp.lt.ftz.f32 %p2438, %f4268, 0f00000000; @%p2438 bra $L__BB107_2689; mul.ftz.f32 %f4269, %f4267, %f4267; @%p2406 bra $L__BB107_2687; bra.uni $L__BB107_2686; $L__BB107_2687: sub.ftz.f32 %f19912, %f19879, %f4269; fma.rn.ftz.f32 %f19913, %f4202, %f4269, %f19912; mul.ftz.f32 %f19914, %f19913, 0f40490FDB; mul.ftz.f32 %f19915, %f19913, %f19914; mov.f32 %f19916, 0f33D6BF95; max.ftz.f32 %f19917, %f19915, %f19916; div.approx.ftz.f32 %f30031, %f4202, %f19917; bra.uni $L__BB107_2688; $L__BB107_2662: setp.lt.ftz.f32 %p2413, %f30046, 0fBF7FF972; @%p2413 bra $L__BB107_2669; bra.uni $L__BB107_2663; $L__BB107_2669: mul.ftz.f32 %f30025, %f30025, %f4218; cvt.ftz.sat.f32.f32 %f19799, %f30025; fma.rn.ftz.f32 %f30026, %f19799, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p2423, %f30046, 0f3727C5AC; mov.f32 %f30024, 0f00000000; mov.pred %p3924, %p2422; @%p2423 bra $L__BB107_2671; lg2.approx.ftz.f32 %f19800, %f30025; mul.ftz.f32 %f19801, %f30050, %f19800; ex2.approx.ftz.f32 %f30024, %f19801; mov.pred %p3924, %p2422; bra.uni $L__BB107_2671; $L__BB107_2663: abs.ftz.f32 %f19783, %f30046; setp.ltu.ftz.f32 %p2415, %f19783, 0f38D1B717; mov.pred %p3924, %p2422; @%p2415 bra $L__BB107_2671; mov.f32 %f30023, 0f3F800000; sub.ftz.f32 %f19785, %f30023, %f30024; setp.lt.ftz.f32 %p2417, %f19785, %f4218; mov.pred %p3924, %p2411; @%p2417 bra $L__BB107_2671; setp.ge.ftz.f32 %p2418, %f30050, 0f00000000; @%p2418 bra $L__BB107_2667; mov.f32 %f19787, 0f3F800000; sub.ftz.f32 %f19788, %f19787, %f4218; lg2.approx.ftz.f32 %f19789, %f19788; mov.f32 %f19790, 0fBF800000; div.approx.ftz.f32 %f19791, %f19790, %f30050; mul.ftz.f32 %f19792, %f19791, %f19789; ex2.approx.ftz.f32 %f19793, %f19792; mul.ftz.f32 %f30023, %f30025, %f19793; $L__BB107_2667: cvt.ftz.sat.f32.f32 %f19795, %f30023; fma.rn.ftz.f32 %f30026, %f19795, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p2420, %f30046, 0f3727C5AC; mov.f32 %f30024, 0f00000000; mov.f32 %f30025, %f30023; mov.pred %p3924, %p2422; @%p2420 bra $L__BB107_2671; lg2.approx.ftz.f32 %f19796, %f30023; mul.ftz.f32 %f19797, %f30050, %f19796; ex2.approx.ftz.f32 %f30024, %f19797; mov.f32 %f30025, %f30023; mov.pred %p3924, %p2422; bra.uni $L__BB107_2671; $L__BB107_2674: add.ftz.f32 %f19813, %f4236, 0f3F800000; mul.ftz.f32 %f4237, %f19813, 0f3F000000; setp.lt.ftz.f32 %p2427, %f4237, 0f38D1B717; mov.f32 %f30027, 0f00000000; mov.f32 %f30028, %f30027; @%p2427 bra $L__BB107_2679; add.ftz.f32 %f19816, %f4230, %f4230; mul.ftz.f32 %f19817, %f19816, %f4237; div.approx.ftz.f32 %f19818, %f19817, %f4236; add.ftz.f32 %f4238, %f19818, 0fBF800000; mul.ftz.f32 %f4239, %f4238, %f4238; add.ftz.f32 %f4240, %f4239, 0fBF800000; abs.ftz.f32 %f19819, %f4240; setp.lt.ftz.f32 %p2428, %f19819, 0f33D6BF95; mov.f32 %f30028, %f30027; @%p2428 bra $L__BB107_2679; mul.ftz.f32 %f19820, %f4236, %f4236; mov.f32 %f19821, 0f3F800000; sub.ftz.f32 %f19822, %f19821, %f19820; mov.f32 %f19823, 0f00000000; max.ftz.f32 %f19824, %f19822, %f19823; sqrt.approx.ftz.f32 %f19825, %f19824; div.approx.ftz.f32 %f19826, %f19825, %f4236; rcp.approx.ftz.f32 %f19827, %f4240; mul.ftz.f32 %f19828, %f19826, %f19826; mul.ftz.f32 %f19829, %f19828, %f19827; mul.ftz.f32 %f19830, %f19827, %f19829; sub.ftz.f32 %f19831, %f4239, %f19828; mul.ftz.f32 %f19832, %f19831, %f19827; sub.ftz.f32 %f19833, %f19830, %f19832; max.ftz.f32 %f19834, %f19833, %f19823; sqrt.approx.ftz.f32 %f19835, %f19834; mul.ftz.f32 %f19836, %f19826, %f19827; add.ftz.f32 %f19837, %f19836, %f19835; rcp.approx.ftz.f32 %f19838, %f19826; setp.gt.ftz.f32 %p2429, %f19837, %f19838; setp.lt.ftz.f32 %p2430, %f4238, 0f00000000; or.pred %p2431, %p2430, %p2429; sub.ftz.f32 %f19839, %f19836, %f19835; mov.f32 %f19840, 0f3F000000; sub.ftz.f32 %f19841, %f19840, %f4229; add.ftz.f32 %f19842, %f4229, 0fBF000000; setp.ge.ftz.f32 %p2432, %f4229, 0f3F000000; selp.f32 %f19843, %f19842, %f19841, %p2432; add.ftz.f32 %f19844, %f19843, %f19843; fma.rn.ftz.f32 %f19845, %f19844, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f19846, %f19844, %f19845, 0f3EED4413; mul.ftz.f32 %f19847, %f19844, %f19846; fma.rn.ftz.f32 %f19848, %f19844, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f19849, %f19844, %f19848, 0fBF800000; fma.rn.ftz.f32 %f19850, %f19844, %f19849, 0f3F191676; div.approx.ftz.f32 %f19851, %f19847, %f19850; selp.f32 %f30027, %f19839, %f19837, %p2431; fma.rn.ftz.f32 %f19852, %f30027, %f30027, 0f3F800000; sqrt.approx.ftz.f32 %f19853, %f19852; mul.ftz.f32 %f30028, %f19851, %f19853; setp.ltu.ftz.f32 %p2433, %f4229, 0f3F000000; @%p2433 bra $L__BB107_2677; bra.uni $L__BB107_2679; $L__BB107_2677: neg.ftz.f32 %f30028, %f30028; bra.uni $L__BB107_2679; $L__BB107_2686: neg.ftz.f32 %f19899, %f4265; div.approx.ftz.f32 %f19900, %f19899, %f19513; neg.ftz.f32 %f19901, %f4266; div.approx.ftz.f32 %f19902, %f19901, %f19514; fma.rn.ftz.f32 %f19903, %f19900, %f19900, %f4269; fma.rn.ftz.f32 %f19904, %f19902, %f19902, %f19903; mul.ftz.f32 %f19905, %f19904, 0f40490FDB; mul.ftz.f32 %f19906, %f19904, %f19905; mul.ftz.f32 %f19907, %f19513, %f19906; mul.ftz.f32 %f19908, %f19514, %f19907; mov.f32 %f19909, 0f33D6BF95; max.ftz.f32 %f19910, %f19908, %f19909; rcp.approx.ftz.f32 %f30031, %f19910; $L__BB107_2688: setp.lt.ftz.f32 %p2440, %f30046, 0fBF7FF972; mul.ftz.f32 %f19918, %f30046, %f30050; selp.f32 %f19919, 0f3F800000, %f19918, %p2440; mul.ftz.f32 %f19920, %f19919, %f4268; mov.f32 %f19921, 0f33D6BF95; max.ftz.f32 %f19922, %f19920, %f19921; mov.f32 %f19923, 0f00000000; max.ftz.f32 %f19924, %f19923, %f4268; mul.ftz.f32 %f19925, %f19924, 0f3E800000; div.approx.ftz.f32 %f19926, %f19925, %f19922; mul.ftz.f32 %f30032, %f19926, %f30031; $L__BB107_2689: mov.f32 %f30033, %f4203; @%p2403 bra $L__BB107_2691; add.ftz.f32 %f19927, %f30026, 0f3F800000; mul.ftz.f32 %f19928, %f19927, 0f3F000000; cvt.ftz.sat.f32.f32 %f19929, %f19928; lg2.approx.ftz.f32 %f19930, %f19929; mul.ftz.f32 %f19931, %f30000, %f19930; ex2.approx.ftz.f32 %f30033, %f19931; $L__BB107_2691: mul.ftz.f32 %f19932, %f30047, %f30032; mul.ftz.f32 %f19933, %f30047, %f19932; fma.rn.ftz.f32 %f30104, %f19933, %f30033, %f30104; mul.ftz.f32 %f19934, %f30048, %f30032; mul.ftz.f32 %f19935, %f30048, %f19934; fma.rn.ftz.f32 %f30105, %f19935, %f30033, %f30105; mul.ftz.f32 %f19936, %f30049, %f30032; mul.ftz.f32 %f19937, %f30049, %f19936; fma.rn.ftz.f32 %f30106, %f19937, %f30033, %f30106; $L__BB107_2692: add.s32 %r5785, %r5785, 1; setp.gt.u32 %p2442, %r5785, 9; @%p2442 bra $L__BB107_2703; mul.ftz.f32 %f19938, %f4263, %f4232; fma.rn.ftz.f32 %f19939, %f4262, %f4231, %f19938; fma.rn.ftz.f32 %f4287, %f4264, %f4235, %f19939; or.pred %p2445, %p2408, %p2435; @%p2445 bra $L__BB107_2697; abs.ftz.f32 %f4288, %f4287; fma.rn.ftz.f32 %f4289, %f4288, %f4288, %f4179; setp.leu.ftz.f32 %p2446, %f4289, 0f00000000; mov.f32 %f30037, %f4184; @%p2446 bra $L__BB107_2696; sqrt.approx.ftz.f32 %f19940, %f4289; sub.ftz.f32 %f19941, %f19940, %f4288; add.ftz.f32 %f19942, %f4288, %f19940; div.approx.ftz.f32 %f19943, %f19941, %f19942; fma.rn.ftz.f32 %f19944, %f4288, %f19942, 0fBF800000; fma.rn.ftz.f32 %f19945, %f4288, %f19941, 0f3F800000; div.approx.ftz.f32 %f19946, %f19944, %f19945; mul.ftz.f32 %f19947, %f19943, 0f3F000000; mul.ftz.f32 %f19948, %f19943, %f19947; fma.rn.ftz.f32 %f19949, %f19946, %f19946, 0f3F800000; mul.ftz.f32 %f19950, %f19948, %f19949; sub.ftz.f32 %f30037, %f19950, %f30002; $L__BB107_2696: mul.ftz.f32 %f19951, %f4185, %f30037; mov.f32 %f19952, 0f3F800000; sub.ftz.f32 %f19953, %f19952, %f19951; fma.rn.ftz.f32 %f19954, %f4156, %f19953, %f19951; fma.rn.ftz.f32 %f19955, %f4157, %f19953, %f19951; fma.rn.ftz.f32 %f19956, %f4158, %f19953, %f19951; mul.ftz.f32 %f30047, %f30047, %f19954; mul.ftz.f32 %f30048, %f30048, %f19955; mul.ftz.f32 %f30049, %f30049, %f19956; $L__BB107_2697: add.ftz.f32 %f19958, %f4262, %f4262; add.ftz.f32 %f19959, %f4263, %f4263; add.ftz.f32 %f19960, %f4264, %f4264; fma.rn.ftz.f32 %f30044, %f19958, %f4287, %f30044; fma.rn.ftz.f32 %f30045, %f19959, %f4287, %f30045; fma.rn.ftz.f32 %f30046, %f19960, %f4287, %f30046; setp.gt.ftz.f32 %p2447, %f30046, 0f3F7FF972; mov.f32 %f30050, 0f00000000; @%p2447 bra $L__BB107_2700; setp.lt.ftz.f32 %p2448, %f30046, 0fBF7FF972; mov.f32 %f30050, 0fBF7FF972; @%p2448 bra $L__BB107_2700; mul.ftz.f32 %f19962, %f30046, %f30046; mov.f32 %f19963, 0f33D6BF95; max.ftz.f32 %f19964, %f19962, %f19963; rcp.approx.ftz.f32 %f19965, %f19964; mul.ftz.f32 %f19966, %f19513, %f30044; mul.ftz.f32 %f19967, %f19514, %f30045; mul.ftz.f32 %f19968, %f19967, %f19967; fma.rn.ftz.f32 %f19969, %f19966, %f19966, %f19968; fma.rn.ftz.f32 %f19970, %f19969, %f19965, 0f3F800000; sqrt.approx.ftz.f32 %f19971, %f19970; neg.ftz.f32 %f19972, %f19971; setp.le.ftz.f32 %p2449, %f30046, 0f00000000; selp.f32 %f19973, %f19972, %f19971, %p2449; add.ftz.f32 %f19974, %f19973, 0fBF800000; mul.ftz.f32 %f30050, %f19974, 0f3F000000; $L__BB107_2700: mul.ftz.f32 %f19976, %f4152, %f30047; setp.eq.s32 %p2450, %r963, 15; selp.f32 %f30047, %f30047, %f19976, %p2450; mul.ftz.f32 %f19977, %f4153, %f30048; selp.f32 %f30048, %f30048, %f19977, %p2450; mul.ftz.f32 %f19978, %f4154, %f30049; selp.f32 %f30049, %f30049, %f19978, %p2450; add.ftz.f32 %f19979, %f30026, 0f3F800000; mul.ftz.f32 %f19980, %f19979, 0f3F000000; cvt.ftz.sat.f32.f32 %f30025, %f19980; mov.f32 %f30024, %f19879; @%p2447 bra $L__BB107_2703; setp.lt.ftz.f32 %p2452, %f30046, 0f3727C5AC; mov.f32 %f30024, 0f00000000; @%p2452 bra $L__BB107_2703; lg2.approx.ftz.f32 %f19982, %f30025; mul.ftz.f32 %f19983, %f30050, %f19982; ex2.approx.ftz.f32 %f30024, %f19983; $L__BB107_2703: setp.lt.u32 %p2453, %r5785, 10; @%p2453 bra $L__BB107_2661; $L__BB107_2704: st.local.u32 [%rd5+168], %r5786; setp.eq.s16 %p2454, %rs721, 0; @%p2454 bra $L__BB107_2706; div.approx.ftz.f32 %f19984, %f29998, %f29995; abs.ftz.f32 %f19985, %f19984; mul.ftz.f32 %f30104, %f30104, %f19985; mul.ftz.f32 %f30105, %f30105, %f19985; mul.ftz.f32 %f30106, %f30106, %f19985; $L__BB107_2706: mov.f32 %f30103, %f29992; bra.uni $L__BB107_2774; $L__BB107_2253: cvt.s64.s32 %rd563, %r5779; mul.wide.s32 %rd2855, %r5779, 80; add.s64 %rd2856, %rd2, %rd2855; add.s64 %rd564, %rd2856, 316; ld.local.u32 %r963, [%rd2856+316]; setp.gt.s32 %p2041, %r963, 31; @%p2041 bra $L__BB107_2794; mov.u16 %rs716, %rs519; @%p2042 bra $L__BB107_2259; setp.ne.s32 %p2043, %r959, 0; add.s32 %r4041, %r963, -2; setp.lt.u32 %p2044, %r4041, 7; and.pred %p2045, %p2043, %p2044; mov.u16 %rs716, 1; @%p2045 bra $L__BB107_2259; setp.eq.s32 %p2046, %r960, 0; @%p2046 bra $L__BB107_2258; add.s32 %r4042, %r963, -9; setp.lt.u32 %p2047, %r4042, 12; setp.eq.s32 %p2048, %r963, 29; or.pred %p2049, %p2048, %p2047; @%p2049 bra $L__BB107_2259; $L__BB107_2258: setp.ne.s32 %p2050, %r961, 0; add.s32 %r4043, %r963, -21; setp.lt.u32 %p2051, %r4043, 10; and.pred %p2052, %p2050, %p2051; selp.u16 %rs716, 1, 0, %p2052; $L__BB107_2259: setp.ne.s16 %p2053, %rs716, 0; @%p2053 bra $L__BB107_2794; mov.f32 %f30103, 0f00000000; setp.gt.s32 %p2054, %r963, 13; @%p2054 bra $L__BB107_2291; bra.uni $L__BB107_2261; $L__BB107_2291: setp.gt.s32 %p2055, %r963, 21; @%p2055 bra $L__BB107_2311; setp.gt.s32 %p2063, %r963, 16; @%p2063 bra $L__BB107_2302; add.s32 %r4044, %r963, -14; setp.lt.u32 %p2067, %r4044, 2; @%p2067 bra $L__BB107_2625; bra.uni $L__BB107_2294; $L__BB107_2625: ld.local.v2.f32 {%f19499, %f19500}, [%rd5+-72]; ld.local.f32 %f4107, [%rd5+-52]; ld.local.v2.f32 {%f19503, %f19504}, [%rd5+-48]; ld.local.f32 %f19505, [%rd5+-76]; mul.ftz.f32 %f19506, %f30259, %f19499; fma.rn.ftz.f32 %f19507, %f30260, %f19505, %f19506; fma.rn.ftz.f32 %f19508, %f30258, %f19500, %f19507; setp.lt.ftz.f32 %p2382, %f19508, 0f00000000; mov.f32 %f30104, 0f00000000; mov.f32 %f30103, %f30104; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2382 bra $L__BB107_2774; ld.local.v2.f32 {%f19513, %f19514}, [%rd564+20]; mul.ftz.f32 %f4112, %f19513, %f19514; setp.lt.ftz.f32 %p2383, %f4112, 0f33D6BF95; mov.f32 %f30103, %f30104; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2383 bra $L__BB107_2774; ld.local.f32 %f4113, [%rd564+8]; ld.local.f32 %f4114, [%rd564+12]; mul.ftz.f32 %f19519, %f19503, %f4114; fma.rn.ftz.f32 %f19520, %f4107, %f4113, %f19519; ld.local.f32 %f4115, [%rd564+16]; fma.rn.ftz.f32 %f29998, %f19504, %f4115, %f19520; setp.le.ftz.f32 %p2384, %f29998, 0f00000000; mov.f32 %f30103, %f30104; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2384 bra $L__BB107_2774; mul.ftz.f32 %f19525, %f30259, %f4114; fma.rn.ftz.f32 %f19526, %f30260, %f4113, %f19525; fma.rn.ftz.f32 %f4117, %f30258, %f4115, %f19526; setp.le.ftz.f32 %p2385, %f4117, 0f00000000; mov.f32 %f30103, %f30104; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2385 bra $L__BB107_2774; setp.neu.ftz.f32 %p2386, %f19513, %f19514; @%p2386 bra $L__BB107_2631; bra.uni $L__BB107_2630; $L__BB107_2631: ld.local.v2.f32 {%f19544, %f19545}, [%rd564+44]; ld.local.f32 %f19548, [%rd564+52]; mul.ftz.f32 %f19549, %f4114, %f19548; mul.ftz.f32 %f19550, %f4115, %f19545; sub.ftz.f32 %f19551, %f19549, %f19550; mul.ftz.f32 %f19552, %f4115, %f19544; mul.ftz.f32 %f19553, %f4113, %f19548; sub.ftz.f32 %f19554, %f19552, %f19553; mul.ftz.f32 %f19555, %f4113, %f19545; mul.ftz.f32 %f19556, %f4114, %f19544; sub.ftz.f32 %f19557, %f19555, %f19556; mul.ftz.f32 %f19558, %f19554, %f19554; fma.rn.ftz.f32 %f19559, %f19551, %f19551, %f19558; fma.rn.ftz.f32 %f19560, %f19557, %f19557, %f19559; sqrt.approx.ftz.f32 %f19561, %f19560; div.approx.ftz.f32 %f29984, %f19551, %f19561; div.approx.ftz.f32 %f29985, %f19554, %f19561; div.approx.ftz.f32 %f29986, %f19557, %f19561; mul.ftz.f32 %f19562, %f4115, %f29985; mul.ftz.f32 %f19563, %f4114, %f29986; sub.ftz.f32 %f29987, %f19562, %f19563; mul.ftz.f32 %f19564, %f4113, %f29986; mul.ftz.f32 %f19565, %f4115, %f29984; sub.ftz.f32 %f29988, %f19564, %f19565; mul.ftz.f32 %f19566, %f4114, %f29984; mul.ftz.f32 %f19567, %f4113, %f29985; sub.ftz.f32 %f29989, %f19566, %f19567; bra.uni $L__BB107_2632; $L__BB107_2261: setp.gt.s32 %p2069, %r963, 7; @%p2069 bra $L__BB107_2280; setp.gt.s32 %p2073, %r963, 4; @%p2073 bra $L__BB107_2270; bra.uni $L__BB107_2263; $L__BB107_2270: setp.eq.s32 %p2074, %r963, 5; @%p2074 bra $L__BB107_2764; setp.eq.s32 %p2075, %r963, 6; @%p2075 bra $L__BB107_2760; setp.eq.s32 %p2076, %r963, 7; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2076 bra $L__BB107_2273; bra.uni $L__BB107_2774; $L__BB107_2273: ld.local.f32 %f20688, [%rd564+8]; ld.local.f32 %f20689, [%rd564+12]; mul.ftz.f32 %f20690, %f30259, %f20689; fma.rn.ftz.f32 %f20691, %f30260, %f20688, %f20690; ld.local.f32 %f20692, [%rd564+16]; fma.rn.ftz.f32 %f4514, %f30258, %f20692, %f20691; setp.ltu.ftz.f32 %p2515, %f4514, 0f00000000; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2515 bra $L__BB107_2279; ld.local.v2.f32 {%f20694, %f20695}, [%rd564+20]; mul.ftz.f32 %f4515, %f20694, 0f3FC90FDB; mov.f32 %f20698, 0f3FC90FDB; mul.ftz.f32 %f4516, %f20695, 0f3FC90FDB; mov.f32 %f30103, 0f00000000; max.ftz.f32 %f20700, %f4514, %f30103; mov.f32 %f20701, 0fBF800000; max.ftz.f32 %f20702, %f20700, %f20701; mov.f32 %f30091, 0f3F800000; min.ftz.f32 %f20703, %f20702, %f30091; abs.ftz.f32 %f20704, %f20703; sub.ftz.f32 %f20705, %f30091, %f20704; mul.ftz.f32 %f20706, %f20705, 0f3F000000; sqrt.approx.ftz.f32 %f20707, %f20706; setp.gt.ftz.f32 %p2516, %f20704, 0f3F11EB85; selp.f32 %f20708, %f20707, %f20704, %p2516; mul.ftz.f32 %f20709, %f20708, %f20708; mov.f32 %f20710, 0f3C94D2E9; mov.f32 %f20711, 0f3D53F941; fma.rn.ftz.f32 %f20712, %f20711, %f20709, %f20710; mov.f32 %f20713, 0f3D3F841F; fma.rn.ftz.f32 %f20714, %f20712, %f20709, %f20713; mov.f32 %f20715, 0f3D994929; fma.rn.ftz.f32 %f20716, %f20714, %f20709, %f20715; mov.f32 %f20717, 0f3E2AAB94; fma.rn.ftz.f32 %f20718, %f20716, %f20709, %f20717; mul.ftz.f32 %f20719, %f20709, %f20718; fma.rn.ftz.f32 %f20720, %f20719, %f20708, %f20708; add.ftz.f32 %f20721, %f20720, %f20720; sub.ftz.f32 %f20722, %f20698, %f20720; selp.f32 %f20723, %f20721, %f20722, %p2516; setp.lt.ftz.f32 %p2517, %f20703, 0f00000000; mov.f32 %f20724, 0f40490FDB; sub.ftz.f32 %f20725, %f20724, %f20723; selp.f32 %f4517, %f20725, %f20723, %p2517; setp.lt.ftz.f32 %p2518, %f4517, %f4515; @%p2518 bra $L__BB107_2277; add.ftz.f32 %f20727, %f4515, %f4516; setp.leu.ftz.f32 %p2519, %f20727, %f4517; setp.eq.ftz.f32 %p2520, %f4516, 0f00000000; or.pred %p2521, %p2520, %p2519; mov.f32 %f30091, %f30103; @%p2521 bra $L__BB107_2277; sub.ftz.f32 %f20728, %f4517, %f4515; div.approx.ftz.f32 %f20729, %f20728, %f4516; mov.f32 %f20730, 0f3F800000; sub.ftz.f32 %f30091, %f20730, %f20729; $L__BB107_2277: setp.leu.ftz.f32 %p2522, %f30091, 0f00000000; mov.f32 %f30104, %f30103; @%p2522 bra $L__BB107_2279; add.ftz.f32 %f20735, %f4515, %f4516; min.ftz.f32 %f20737, %f20735, %f20698; cos.approx.ftz.f32 %f20738, %f20737; mov.f32 %f20739, 0f3F800000; sub.ftz.f32 %f20740, %f20739, %f20738; mov.f32 %f20741, 0f3E22F983; div.approx.ftz.f32 %f30103, %f20741, %f20740; mul.ftz.f32 %f30104, %f30091, %f30103; $L__BB107_2279: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2311: setp.gt.s32 %p2056, %r963, 26; @%p2056 bra $L__BB107_2315; setp.eq.s32 %p2060, %r963, 22; @%p2060 bra $L__BB107_2283; setp.eq.s32 %p2061, %r963, 23; @%p2061 bra $L__BB107_2719; setp.eq.s32 %p2062, %r963, 24; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2062 bra $L__BB107_2318; bra.uni $L__BB107_2774; $L__BB107_2280: add.s32 %r4045, %r963, -10; setp.lt.u32 %p2070, %r4045, 3; @%p2070 bra $L__BB107_2719; bra.uni $L__BB107_2281; $L__BB107_2719: ld.local.v2.f32 {%f20212, %f20213}, [%rd5+-72]; ld.local.f32 %f4404, [%rd5+-52]; ld.local.v2.f32 {%f20216, %f20217}, [%rd5+-48]; ld.local.v2.f32 {%f20218, %f20219}, [%rd564+20]; ld.local.f32 %f20220, [%rd5+-76]; mul.ftz.f32 %f20221, %f30259, %f20212; fma.rn.ftz.f32 %f20222, %f30260, %f20220, %f20221; fma.rn.ftz.f32 %f4409, %f30258, %f20213, %f20222; setp.lt.ftz.f32 %p2474, %f4409, 0f00000000; setp.eq.s32 %p2475, %r963, 23; xor.pred %p2476, %p2474, %p2475; mul.ftz.f32 %f30070, %f20218, %f20219; setp.le.ftz.f32 %p2477, %f30070, 0f33D6BF95; or.pred %p2478, %p2476, %p2477; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2478 bra $L__BB107_2774; ld.local.f32 %f4411, [%rd564+8]; ld.local.f32 %f4412, [%rd564+12]; mul.ftz.f32 %f20223, %f20216, %f4412; fma.rn.ftz.f32 %f20224, %f4404, %f4411, %f20223; ld.local.f32 %f4413, [%rd564+16]; fma.rn.ftz.f32 %f4414, %f20217, %f4413, %f20224; mul.ftz.f32 %f20225, %f30259, %f4412; fma.rn.ftz.f32 %f20226, %f30260, %f4411, %f20225; fma.rn.ftz.f32 %f4415, %f30258, %f4413, %f20226; @%p2474 bra $L__BB107_2738; bra.uni $L__BB107_2721; $L__BB107_2738: setp.le.ftz.f32 %p2490, %f4414, 0f00000000; setp.ge.ftz.f32 %p2491, %f4415, 0f00000000; or.pred %p2492, %p2490, %p2491; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2492 bra $L__BB107_2774; ld.local.f32 %f20412, [%rd564+28]; fma.rn.ftz.f32 %f20413, %f30260, %f20412, %f4404; fma.rn.ftz.f32 %f20414, %f30259, %f20412, %f20216; fma.rn.ftz.f32 %f20415, %f30258, %f20412, %f20217; neg.ftz.f32 %f20416, %f20413; neg.ftz.f32 %f20417, %f20414; neg.ftz.f32 %f20418, %f20415; mul.ftz.f32 %f20419, %f20414, %f20414; fma.rn.ftz.f32 %f20420, %f20413, %f20413, %f20419; fma.rn.ftz.f32 %f20421, %f20415, %f20415, %f20420; sqrt.approx.ftz.f32 %f20422, %f20421; div.approx.ftz.f32 %f20423, %f20416, %f20422; div.approx.ftz.f32 %f20424, %f20417, %f20422; div.approx.ftz.f32 %f20425, %f20418, %f20422; mul.ftz.f32 %f20426, %f20216, %f20424; fma.rn.ftz.f32 %f20427, %f4404, %f20423, %f20426; fma.rn.ftz.f32 %f20428, %f20217, %f20425, %f20427; mul.ftz.f32 %f20429, %f30259, %f20424; fma.rn.ftz.f32 %f20430, %f30260, %f20423, %f20429; fma.rn.ftz.f32 %f20431, %f30258, %f20425, %f20430; mul.ftz.f32 %f20432, %f4412, %f20424; fma.rn.ftz.f32 %f20433, %f4411, %f20423, %f20432; fma.rn.ftz.f32 %f20434, %f4413, %f20425, %f20433; mul.ftz.f32 %f20435, %f20434, %f20434; mov.f32 %f20436, 0f3F800000; sub.ftz.f32 %f20437, %f20436, %f20435; div.approx.ftz.f32 %f20438, %f20437, %f20435; mul.ftz.f32 %f20439, %f20435, %f20435; mul.ftz.f32 %f20440, %f20439, 0f40490FDB; add.ftz.f32 %f20441, %f30070, %f20438; mul.ftz.f32 %f20442, %f20441, %f20440; mul.ftz.f32 %f20443, %f20441, %f20442; div.approx.ftz.f32 %f20444, %f30070, %f20443; mul.ftz.f32 %f20445, %f4414, %f4414; sub.ftz.f32 %f20446, %f20436, %f20445; mul.ftz.f32 %f20447, %f30070, %f20446; div.approx.ftz.f32 %f20448, %f20447, %f20445; add.ftz.f32 %f20449, %f20448, 0f3F800000; mov.f32 %f20450, 0f00000000; max.ftz.f32 %f20451, %f20449, %f20450; sqrt.approx.ftz.f32 %f20452, %f20451; add.ftz.f32 %f20453, %f20452, 0f3F800000; mov.f32 %f20454, 0f40000000; div.approx.ftz.f32 %f20455, %f20454, %f20453; mul.ftz.f32 %f20456, %f4415, %f4415; sub.ftz.f32 %f20457, %f20436, %f20456; mul.ftz.f32 %f20458, %f30070, %f20457; div.approx.ftz.f32 %f20459, %f20458, %f20456; add.ftz.f32 %f20460, %f20459, 0f3F800000; max.ftz.f32 %f20461, %f20460, %f20450; sqrt.approx.ftz.f32 %f20462, %f20461; add.ftz.f32 %f20463, %f20462, 0f3F800000; div.approx.ftz.f32 %f20464, %f20454, %f20463; mul.ftz.f32 %f20465, %f20455, %f20464; mul.ftz.f32 %f20466, %f20412, %f20412; mul.ftz.f32 %f20467, %f20466, %f20444; mul.ftz.f32 %f20468, %f4414, %f20421; div.approx.ftz.f32 %f20469, %f20467, %f20468; mul.ftz.f32 %f20470, %f20431, %f20428; abs.ftz.f32 %f20471, %f20470; mul.ftz.f32 %f20472, %f20465, %f20471; mul.ftz.f32 %f20473, %f20455, %f20471; mul.ftz.f32 %f30103, %f20469, %f20473; mul.ftz.f32 %f30104, %f20469, %f20472; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2302: setp.eq.s32 %p2064, %r963, 17; @%p2064 bra $L__BB107_2618; setp.eq.s32 %p2065, %r963, 19; @%p2065 bra $L__BB107_2612; setp.eq.s32 %p2066, %r963, 20; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2066 bra $L__BB107_2305; bra.uni $L__BB107_2774; $L__BB107_2305: ld.local.f32 %f19071, [%rd564+8]; ld.local.f32 %f19072, [%rd564+12]; mul.ftz.f32 %f19073, %f30259, %f19072; fma.rn.ftz.f32 %f19074, %f30260, %f19071, %f19073; ld.local.f32 %f19075, [%rd564+16]; fma.rn.ftz.f32 %f19076, %f30258, %f19075, %f19074; setp.lt.ftz.f32 %p2338, %f19076, 0f00000000; mov.f32 %f30103, 0f00000000; @%p2338 bra $L__BB107_2611; ld.local.f32 %f19081, [%rd5+-52]; ld.local.v2.f32 {%f19082, %f19083}, [%rd5+-48]; ld.local.f32 %f3993, [%rd564+40]; ld.local.v4.f32 {%f19086, %f19087, %f19088, %f19089}, [%rd564+20]; mul.ftz.f32 %f19093, %f19082, %f19087; fma.rn.ftz.f32 %f19094, %f19081, %f19086, %f19093; fma.rn.ftz.f32 %f19095, %f19083, %f19088, %f19094; mul.ftz.f32 %f19096, %f19086, %f19095; mul.ftz.f32 %f19097, %f19087, %f19095; mul.ftz.f32 %f19098, %f19088, %f19095; sub.ftz.f32 %f19099, %f19081, %f19096; sub.ftz.f32 %f19100, %f19082, %f19097; sub.ftz.f32 %f19101, %f19083, %f19098; mul.ftz.f32 %f19102, %f19100, %f19100; fma.rn.ftz.f32 %f19103, %f19099, %f19099, %f19102; fma.rn.ftz.f32 %f19104, %f19101, %f19101, %f19103; sqrt.approx.ftz.f32 %f19105, %f19104; div.approx.ftz.f32 %f19106, %f19099, %f19105; div.approx.ftz.f32 %f19107, %f19100, %f19105; div.approx.ftz.f32 %f19108, %f19101, %f19105; abs.ftz.f32 %f19109, %f19095; setp.lt.ftz.f32 %p2339, %f19109, 0f3F800000; mov.f32 %f19110, 0f3F800000; add.ftz.f32 %f19111, %f19109, 0fBF800000; add.ftz.f32 %f19112, %f19111, 0f3F800000; selp.f32 %f19113, %f19112, 0f3F800000, %p2339; sub.ftz.f32 %f19114, %f19110, %f19113; sqrt.approx.ftz.f32 %f19115, %f19114; mul.ftz.f32 %f19116, %f19113, 0f3CB1485D; mov.f32 %f19117, 0f3D9FB43B; sub.ftz.f32 %f19118, %f19117, %f19116; fma.rn.ftz.f32 %f19119, %f19113, %f19118, 0fBE5A6B93; fma.rn.ftz.f32 %f19120, %f19113, %f19119, 0f3FC90FDB; mov.f32 %f19121, 0f3FC90FDB; mul.ftz.f32 %f19122, %f19115, %f19120; setp.lt.ftz.f32 %p2340, %f19095, 0f00000000; mov.f32 %f19123, 0f40490FDB; sub.ftz.f32 %f19124, %f19123, %f19122; selp.f32 %f19125, %f19124, %f19122, %p2340; sub.ftz.f32 %f3995, %f19121, %f19125; mul.ftz.f32 %f19126, %f30259, %f19087; fma.rn.ftz.f32 %f19127, %f30260, %f19086, %f19126; fma.rn.ftz.f32 %f19128, %f30258, %f19088, %f19127; mul.ftz.f32 %f19129, %f19086, %f19128; mul.ftz.f32 %f19130, %f19087, %f19128; mul.ftz.f32 %f19131, %f19088, %f19128; sub.ftz.f32 %f19132, %f30260, %f19129; sub.ftz.f32 %f19133, %f30259, %f19130; sub.ftz.f32 %f19134, %f30258, %f19131; mul.ftz.f32 %f19135, %f19133, %f19133; fma.rn.ftz.f32 %f19136, %f19132, %f19132, %f19135; fma.rn.ftz.f32 %f19137, %f19134, %f19134, %f19136; sqrt.approx.ftz.f32 %f19138, %f19137; div.approx.ftz.f32 %f19139, %f19132, %f19138; div.approx.ftz.f32 %f19140, %f19133, %f19138; div.approx.ftz.f32 %f19141, %f19134, %f19138; abs.ftz.f32 %f19142, %f19128; setp.lt.ftz.f32 %p2341, %f19142, 0f3F800000; add.ftz.f32 %f19143, %f19142, 0fBF800000; add.ftz.f32 %f19144, %f19143, 0f3F800000; selp.f32 %f19145, %f19144, 0f3F800000, %p2341; sub.ftz.f32 %f19146, %f19110, %f19145; sqrt.approx.ftz.f32 %f19147, %f19146; mul.ftz.f32 %f19148, %f19145, 0f3CB1485D; sub.ftz.f32 %f19149, %f19117, %f19148; fma.rn.ftz.f32 %f19150, %f19145, %f19149, 0fBE5A6B93; fma.rn.ftz.f32 %f19151, %f19145, %f19150, 0f3FC90FDB; mul.ftz.f32 %f19152, %f19147, %f19151; setp.lt.ftz.f32 %p2342, %f19128, 0f00000000; sub.ftz.f32 %f19153, %f19123, %f19152; selp.f32 %f19154, %f19153, %f19152, %p2342; sub.ftz.f32 %f3996, %f19121, %f19154; mul.ftz.f32 %f19155, %f19107, %f19140; fma.rn.ftz.f32 %f19156, %f19106, %f19139, %f19155; fma.rn.ftz.f32 %f3997, %f19108, %f19141, %f19156; abs.ftz.f32 %f19157, %f3996; sub.ftz.f32 %f19158, %f19121, %f19157; setp.lt.ftz.f32 %p2343, %f19158, 0f3A83126F; setp.lt.ftz.f32 %p2344, %f3997, 0f00000000; or.pred %p2345, %p2344, %p2343; @%p2345 bra $L__BB107_2611; ld.local.f32 %f19160, [%rd564+36]; rcp.approx.ftz.f32 %f3998, %f19160; abs.ftz.f32 %f19162, %f3997; setp.lt.ftz.f32 %p2346, %f19162, 0f3F800000; add.ftz.f32 %f19163, %f19162, 0fBF800000; add.ftz.f32 %f19164, %f19163, 0f3F800000; selp.f32 %f19165, %f19164, 0f3F800000, %p2346; sub.ftz.f32 %f19166, %f19110, %f19165; sqrt.approx.ftz.f32 %f19167, %f19166; fma.rn.ftz.f32 %f19168, %f19165, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f19169, %f19165, %f19168, 0fBE5A6B93; fma.rn.ftz.f32 %f19170, %f19165, %f19169, 0f3FC90FDB; mul.ftz.f32 %f3999, %f19167, %f19170; mul.ftz.f32 %f19171, %f3996, 0f3EA2F983; mov.b32 %r4108, %f19171; and.b32 %r4109, %r4108, -2147483648; or.b32 %r4110, %r4109, 1056964608; mov.b32 %f19172, %r4110; add.ftz.f32 %f19173, %f19171, %f19172; cvt.rzi.ftz.s32.f32 %r977, %f19173; cvt.rn.f32.s32 %f19174, %r977; fma.rn.ftz.f32 %f19175, %f19174, 0fC0490000, %f3996; fma.rn.ftz.f32 %f19176, %f19174, 0fBA7DA000, %f19175; fma.rn.ftz.f32 %f19177, %f19174, 0fB4222000, %f19176; fma.rn.ftz.f32 %f19178, %f19174, 0fACB4611A, %f19177; add.ftz.f32 %f19179, %f19178, 0fBFC90FDB; add.ftz.f32 %f19180, %f19179, 0f3FC90FDB; mul.ftz.f32 %f19181, %f19180, %f19180; fma.rn.ftz.f32 %f19182, %f19181, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f19183, %f19181, %f19182, 0fBAB60B58; fma.rn.ftz.f32 %f19184, %f19181, %f19183, 0f3D2AAAAA; fma.rn.ftz.f32 %f19185, %f19181, %f19184, 0fBF000000; fma.rn.ftz.f32 %f4000, %f19181, %f19185, 0f3F800000; add.ftz.f32 %f19186, %f3995, 0f3FC90FDB; mul.ftz.f32 %f19187, %f19186, 0f3F000000; sub.ftz.f32 %f19188, %f19187, %f3993; rcp.approx.ftz.f32 %f4001, %f19089; mul.ftz.f32 %f4002, %f4001, %f19188; abs.ftz.f32 %f4003, %f4002; setp.eq.ftz.f32 %p2347, %f4003, 0f00000000; mov.f32 %f29972, 0f00000000; mov.f32 %f29971, %f29972; @%p2347 bra $L__BB107_2605; setp.eq.ftz.f32 %p2348, %f4003, 0f3F800000; mov.f32 %f29971, %f19110; @%p2348 bra $L__BB107_2605; setp.gt.ftz.f32 %p2349, %f4003, 0f3F800000; @%p2349 bra $L__BB107_2604; bra.uni $L__BB107_2310; $L__BB107_2604: mov.f32 %f19191, 0f3F800000; div.approx.ftz.f32 %f29971, %f19191, %f4003; bra.uni $L__BB107_2605; $L__BB107_2315: setp.eq.s32 %p2057, %r963, 30; @%p2057 bra $L__BB107_2740; setp.eq.s32 %p2058, %r963, 29; @%p2058 bra $L__BB107_2459; setp.ne.s32 %p2059, %r963, 27; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2059 bra $L__BB107_2774; $L__BB107_2318: ld.local.v2.f32 {%f17470, %f17471}, [%rd564+20]; mul.ftz.f32 %f17472, %f17470, %f17471; setp.lt.ftz.f32 %p2079, %f17472, 0f33D6BF95; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2079 bra $L__BB107_2774; ld.local.f32 %f17474, [%rd5+-52]; ld.local.v2.f32 {%f17475, %f17476}, [%rd5+-48]; ld.local.f32 %f17479, [%rd564+12]; ld.local.f32 %f17480, [%rd564+8]; setp.neu.ftz.f32 %p2080, %f17480, %f17479; ld.local.f32 %f17481, [%rd564+16]; setp.neu.ftz.f32 %p2081, %f17480, %f17481; or.pred %p2082, %p2080, %p2081; sub.ftz.f32 %f17482, %f17481, %f17479; neg.ftz.f32 %f17483, %f17479; neg.ftz.f32 %f17484, %f17481; selp.f32 %f17485, %f17484, %f17481, %p2082; add.ftz.f32 %f17486, %f17480, %f17485; selp.f32 %f17487, %f17479, %f17483, %p2082; sub.ftz.f32 %f17488, %f17487, %f17480; mul.ftz.f32 %f17489, %f17486, %f17486; fma.rn.ftz.f32 %f17490, %f17482, %f17482, %f17489; fma.rn.ftz.f32 %f17491, %f17488, %f17488, %f17490; sqrt.approx.ftz.f32 %f17492, %f17491; div.approx.ftz.f32 %f17493, %f17482, %f17492; div.approx.ftz.f32 %f17494, %f17486, %f17492; div.approx.ftz.f32 %f17495, %f17488, %f17492; mul.ftz.f32 %f17496, %f17479, %f17495; mul.ftz.f32 %f17497, %f17481, %f17494; sub.ftz.f32 %f17498, %f17496, %f17497; mul.ftz.f32 %f17499, %f17481, %f17493; mul.ftz.f32 %f17500, %f17480, %f17495; sub.ftz.f32 %f17501, %f17499, %f17500; mul.ftz.f32 %f17502, %f17480, %f17494; mul.ftz.f32 %f17503, %f17479, %f17493; sub.ftz.f32 %f17504, %f17502, %f17503; mul.ftz.f32 %f17505, %f17475, %f17494; fma.rn.ftz.f32 %f17506, %f17474, %f17493, %f17505; fma.rn.ftz.f32 %f29827, %f17476, %f17495, %f17506; mul.ftz.f32 %f17507, %f17475, %f17501; fma.rn.ftz.f32 %f17508, %f17474, %f17498, %f17507; fma.rn.ftz.f32 %f29828, %f17476, %f17504, %f17508; mul.ftz.f32 %f17509, %f17475, %f17479; fma.rn.ftz.f32 %f17510, %f17474, %f17480, %f17509; fma.rn.ftz.f32 %f29829, %f17476, %f17481, %f17510; mul.ftz.f32 %f17511, %f30259, %f17494; fma.rn.ftz.f32 %f17512, %f30260, %f17493, %f17511; fma.rn.ftz.f32 %f3241, %f30258, %f17495, %f17512; mul.ftz.f32 %f17513, %f30259, %f17501; fma.rn.ftz.f32 %f17514, %f30260, %f17498, %f17513; fma.rn.ftz.f32 %f3242, %f30258, %f17504, %f17514; mul.ftz.f32 %f17515, %f30259, %f17479; fma.rn.ftz.f32 %f17516, %f30260, %f17480, %f17515; fma.rn.ftz.f32 %f3243, %f30258, %f17481, %f17516; mov.f32 %f29817, 0f00000000; setp.geu.ftz.f32 %p7, %f3243, 0f00000000; setp.eq.s32 %p2083, %r963, 27; and.pred %p8, %p2083, %p7; mul.ftz.f32 %f3244, %f29829, %f3243; setp.gt.ftz.f32 %p2084, %f3244, 0f00000000; ld.local.f32 %f3245, [%rd564+28]; mul.ftz.f32 %f17517, %f3241, %f3245; mul.ftz.f32 %f17518, %f3242, %f3245; mul.ftz.f32 %f17519, %f3243, %f3245; selp.f32 %f17520, %f3241, %f17517, %p2084; selp.f32 %f17521, %f3242, %f17518, %p2084; selp.f32 %f17522, %f3243, %f17519, %p2084; add.ftz.f32 %f17523, %f29827, %f17520; add.ftz.f32 %f17524, %f29828, %f17521; add.ftz.f32 %f17525, %f29829, %f17522; mul.ftz.f32 %f17526, %f17524, %f17524; fma.rn.ftz.f32 %f17527, %f17523, %f17523, %f17526; fma.rn.ftz.f32 %f17528, %f17525, %f17525, %f17527; sqrt.approx.ftz.f32 %f3246, %f17528; rsqrt.approx.ftz.f32 %f17529, %f17528; mul.ftz.f32 %f17530, %f17529, %f17523; mul.ftz.f32 %f17531, %f17529, %f17524; mul.ftz.f32 %f17532, %f17529, %f17525; setp.lt.ftz.f32 %p2085, %f17532, 0f00000000; neg.ftz.f32 %f17533, %f17530; neg.ftz.f32 %f17534, %f17531; neg.ftz.f32 %f17535, %f17532; selp.f32 %f3247, %f17533, %f17530, %p2085; selp.f32 %f3248, %f17534, %f17531, %p2085; selp.f32 %f3249, %f17535, %f17532, %p2085; setp.lt.ftz.f32 %p2086, %f29829, 0f00000000; neg.ftz.f32 %f3250, %f29827; neg.ftz.f32 %f3251, %f29828; neg.ftz.f32 %f3252, %f29829; selp.f32 %f3253, %f3250, %f29827, %p2086; selp.f32 %f3254, %f3251, %f29828, %p2086; selp.f32 %f3255, %f3252, %f29829, %p2086; setp.gt.ftz.f32 %p2087, %f3255, 0f3F7FF972; @%p2087 bra $L__BB107_2322; setp.lt.ftz.f32 %p2088, %f3255, 0fBF7FF972; mov.f32 %f29817, 0fBF7FF972; @%p2088 bra $L__BB107_2322; mul.ftz.f32 %f17537, %f3255, %f3255; mov.f32 %f17538, 0f33D6BF95; max.ftz.f32 %f17539, %f17537, %f17538; rcp.approx.ftz.f32 %f17540, %f17539; mul.ftz.f32 %f17541, %f3253, %f17470; mul.ftz.f32 %f17542, %f3254, %f17470; mul.ftz.f32 %f17543, %f17542, %f17542; fma.rn.ftz.f32 %f17544, %f17541, %f17541, %f17543; fma.rn.ftz.f32 %f17545, %f17544, %f17540, 0f3F800000; sqrt.approx.ftz.f32 %f17546, %f17545; neg.ftz.f32 %f17547, %f17546; setp.le.ftz.f32 %p2089, %f3255, 0f00000000; selp.f32 %f17548, %f17547, %f17546, %p2089; add.ftz.f32 %f17549, %f17548, 0fBF800000; mul.ftz.f32 %f29817, %f17549, 0f3F000000; $L__BB107_2322: mul.ftz.f32 %f17551, %f3249, %f3249; mov.f32 %f29818, 0f3F800000; sub.ftz.f32 %f17552, %f29818, %f17551; mul.ftz.f32 %f3258, %f17470, %f17470; fma.rn.ftz.f32 %f17553, %f3258, %f17551, %f17552; mul.ftz.f32 %f17554, %f17553, 0f40490FDB; mul.ftz.f32 %f17555, %f17553, %f17554; mov.f32 %f17556, 0f33D6BF95; max.ftz.f32 %f17557, %f17555, %f17556; div.approx.ftz.f32 %f3259, %f3258, %f17557; mul.ftz.f32 %f17558, %f3254, %f3248; fma.rn.ftz.f32 %f17559, %f3253, %f3247, %f17558; fma.rn.ftz.f32 %f3260, %f3255, %f3249, %f17559; abs.ftz.f32 %f3261, %f3260; fma.rn.ftz.f32 %f3262, %f3245, %f3245, 0fBF800000; fma.rn.ftz.f32 %f3263, %f3261, %f3261, %f3262; setp.leu.ftz.f32 %p2090, %f3263, 0f00000000; @%p2090 bra $L__BB107_2324; sqrt.approx.ftz.f32 %f17560, %f3263; sub.ftz.f32 %f17561, %f17560, %f3261; add.ftz.f32 %f17562, %f3261, %f17560; div.approx.ftz.f32 %f17563, %f17561, %f17562; fma.rn.ftz.f32 %f17564, %f3261, %f17562, 0fBF800000; fma.rn.ftz.f32 %f17565, %f3261, %f17561, 0f3F800000; div.approx.ftz.f32 %f17566, %f17564, %f17565; mul.ftz.f32 %f17567, %f17563, 0f3F000000; mul.ftz.f32 %f17568, %f17563, %f17567; fma.rn.ftz.f32 %f17569, %f17566, %f17566, 0f3F800000; mul.ftz.f32 %f29818, %f17568, %f17569; $L__BB107_2324: mul.ftz.f32 %f17570, %f3243, 0f3EA2F983; abs.ftz.f32 %f3266, %f17570; add.ftz.f32 %f17571, %f29817, 0f3F800000; mul.ftz.f32 %f3267, %f3255, %f17571; @%p2084 bra $L__BB107_2328; bra.uni $L__BB107_2325; $L__BB107_2328: mul.ftz.f32 %f17629, %f3259, 0f3E800000; max.ftz.f32 %f17631, %f3267, %f17556; div.approx.ftz.f32 %f17632, %f17629, %f17631; mul.ftz.f32 %f17633, %f17470, 0fBFFE59F3; mul.ftz.f32 %f17634, %f17470, %f17633; mul.ftz.f32 %f17635, %f17634, 0f3FB8AA3B; ex2.approx.ftz.f32 %f17636, %f17635; fma.rn.ftz.f32 %f17637, %f17636, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f17638, %f17470, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f17639, %f17470, %f17638, 0f417C8B44; fma.rn.ftz.f32 %f17640, %f17470, %f17639, 0fC1665E35; fma.rn.ftz.f32 %f17641, %f17470, %f17640, 0f40CE7881; fma.rn.ftz.f32 %f17642, %f17470, %f17641, 0fBF993E81; fma.rn.ftz.f32 %f17643, %f17470, %f17642, 0f3CE3C31E; fma.rn.ftz.f32 %f17644, %f17470, %f17643, 0f3BBA5D49; sub.ftz.f32 %f17645, %f17637, %f17644; cvt.ftz.sat.f32.f32 %f17646, %f17645; mov.f32 %f17647, 0f3F800000; sub.ftz.f32 %f17648, %f17647, %f17646; mul.ftz.f32 %f17649, %f3266, %f17648; fma.rn.ftz.f32 %f17650, %f17632, %f17646, %f17649; mul.ftz.f32 %f30103, %f29818, %f17650; bra.uni $L__BB107_2329; $L__BB107_2281: setp.eq.s32 %p2071, %r963, 8; @%p2071 bra $L__BB107_2759; setp.eq.s32 %p2072, %r963, 13; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2072 bra $L__BB107_2283; bra.uni $L__BB107_2774; $L__BB107_2283: ld.local.v2.f32 {%f19990, %f19991}, [%rd5+-72]; ld.local.f32 %f4336, [%rd5+-52]; ld.local.v2.f32 {%f19994, %f19995}, [%rd5+-48]; ld.local.v2.f32 {%f19996, %f19997}, [%rd564+20]; ld.local.f32 %f19998, [%rd5+-76]; mul.ftz.f32 %f19999, %f30259, %f19990; fma.rn.ftz.f32 %f20000, %f30260, %f19998, %f19999; fma.rn.ftz.f32 %f20001, %f30258, %f19991, %f20000; setp.lt.ftz.f32 %p2455, %f20001, 0f00000000; setp.eq.s32 %p2456, %r963, 22; xor.pred %p2457, %p2455, %p2456; mul.ftz.f32 %f4341, %f19996, %f19997; setp.le.ftz.f32 %p2458, %f4341, 0f33D6BF95; or.pred %p2459, %p2457, %p2458; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2459 bra $L__BB107_2718; bra.uni $L__BB107_2284; $L__BB107_2718: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2263: setp.eq.s32 %p2077, %r963, 2; @%p2077 bra $L__BB107_2773; setp.eq.s32 %p2078, %r963, 3; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2078 bra $L__BB107_2265; bra.uni $L__BB107_2774; $L__BB107_2265: ld.local.f32 %f4575, [%rd564+8]; ld.local.f32 %f4576, [%rd564+12]; mul.ftz.f32 %f20823, %f30259, %f4576; fma.rn.ftz.f32 %f20824, %f30260, %f4575, %f20823; ld.local.f32 %f4577, [%rd564+16]; fma.rn.ftz.f32 %f4578, %f30258, %f4577, %f20824; setp.leu.ftz.f32 %p2532, %f4578, 0f00000000; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2532 bra $L__BB107_2269; ld.local.f32 %f20825, [%rd5+-52]; ld.local.v2.f32 {%f20826, %f20827}, [%rd5+-48]; mov.f32 %f20830, 0f00000000; max.ftz.f32 %f4579, %f4578, %f20830; mul.ftz.f32 %f20831, %f20826, %f4576; fma.rn.ftz.f32 %f20832, %f20825, %f4575, %f20831; fma.rn.ftz.f32 %f20833, %f20827, %f4577, %f20832; max.ftz.f32 %f4580, %f20833, %f20830; mul.ftz.f32 %f20834, %f30259, %f20826; fma.rn.ftz.f32 %f20835, %f30260, %f20825, %f20834; fma.rn.ftz.f32 %f20836, %f30258, %f20827, %f20835; mul.ftz.f32 %f20837, %f4579, %f4580; sub.ftz.f32 %f30100, %f20836, %f20837; setp.leu.ftz.f32 %p2533, %f30100, 0f00000000; @%p2533 bra $L__BB107_2268; max.ftz.f32 %f20838, %f4579, %f4580; add.ftz.f32 %f20839, %f20838, 0f00800000; div.approx.ftz.f32 %f30100, %f30100, %f20839; $L__BB107_2268: ld.local.f32 %f20841, [%rd564+28]; ld.local.f32 %f20842, [%rd564+24]; fma.rn.ftz.f32 %f20843, %f30100, %f20841, %f20842; mul.ftz.f32 %f30104, %f4579, %f20843; mov.f32 %f30103, 0f3E22F983; $L__BB107_2269: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2294: setp.eq.s32 %p2068, %r963, 16; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2068 bra $L__BB107_2295; bra.uni $L__BB107_2774; $L__BB107_2295: ld.local.v2.f32 {%f19383, %f19384}, [%rd5+-72]; ld.local.v2.f32 {%f19387, %f19388}, [%rd5+-48]; ld.local.f32 %f19389, [%rd5+-76]; mul.ftz.f32 %f19390, %f30259, %f19383; fma.rn.ftz.f32 %f19391, %f30260, %f19389, %f19390; fma.rn.ftz.f32 %f19392, %f30258, %f19384, %f19391; ld.local.f32 %f4071, [%rd564+8]; ld.local.f32 %f4072, [%rd5+-52]; ld.local.f32 %f4073, [%rd564+12]; mul.ftz.f32 %f19393, %f19387, %f4073; fma.rn.ftz.f32 %f19394, %f4072, %f4071, %f19393; ld.local.f32 %f4074, [%rd564+16]; fma.rn.ftz.f32 %f4075, %f19388, %f4074, %f19394; mul.ftz.f32 %f19395, %f30259, %f4073; fma.rn.ftz.f32 %f19396, %f30260, %f4071, %f19395; fma.rn.ftz.f32 %f4076, %f30258, %f4074, %f19396; setp.lt.ftz.f32 %p2375, %f19392, 0f00000000; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2375 bra $L__BB107_2624; ld.local.v2.f32 {%f19401, %f19402}, [%rd564+20]; max.ftz.f32 %f19403, %f19401, %f19402; setp.le.ftz.f32 %p2376, %f19403, 0f38D1B717; mov.f32 %f30104, %f30103; @%p2376 bra $L__BB107_2624; setp.leu.ftz.f32 %p2377, %f4075, 0f00000000; setp.leu.ftz.f32 %p2378, %f4076, 0f00000000; or.pred %p2379, %p2378, %p2377; mov.f32 %f30104, %f30103; @%p2379 bra $L__BB107_2624; add.ftz.f32 %f19408, %f30260, %f4072; add.ftz.f32 %f19409, %f30259, %f19387; mul.ftz.f32 %f19410, %f19409, %f19409; fma.rn.ftz.f32 %f19411, %f19408, %f19408, %f19410; add.ftz.f32 %f19412, %f30258, %f19388; fma.rn.ftz.f32 %f19413, %f19412, %f19412, %f19411; sqrt.approx.ftz.f32 %f19414, %f19413; div.approx.ftz.f32 %f4079, %f19408, %f19414; div.approx.ftz.f32 %f4080, %f19409, %f19414; div.approx.ftz.f32 %f4081, %f19412, %f19414; mul.ftz.f32 %f19415, %f19387, %f4080; fma.rn.ftz.f32 %f19416, %f4072, %f4079, %f19415; fma.rn.ftz.f32 %f19417, %f19388, %f4081, %f19416; abs.ftz.f32 %f19418, %f19417; mov.f32 %f19419, 0f358637BD; max.ftz.f32 %f4082, %f19418, %f19419; mul.ftz.f32 %f19420, %f4073, %f4080; fma.rn.ftz.f32 %f19421, %f4071, %f4079, %f19420; fma.rn.ftz.f32 %f19422, %f4074, %f4081, %f19421; max.ftz.f32 %f4083, %f19422, %f19419; max.ftz.f32 %f4084, %f4076, %f19419; mul.ftz.f32 %f19423, %f19401, %f19401; mov.f32 %f19424, 0f40000000; div.approx.ftz.f32 %f19425, %f19424, %f19423; add.ftz.f32 %f4085, %f19425, 0fC0000000; mul.ftz.f32 %f19426, %f19402, %f19402; div.approx.ftz.f32 %f19427, %f19424, %f19426; add.ftz.f32 %f4086, %f19427, 0fC0000000; setp.eq.ftz.f32 %p2380, %f4085, %f4086; @%p2380 bra $L__BB107_2622; bra.uni $L__BB107_2299; $L__BB107_2622: lg2.approx.ftz.f32 %f19481, %f4083; mul.ftz.f32 %f19482, %f4085, %f19481; ex2.approx.ftz.f32 %f19483, %f19482; add.ftz.f32 %f19484, %f4085, 0f3F800000; mov.f32 %f19485, 0f41C90FDB; div.approx.ftz.f32 %f19486, %f19484, %f19485; mul.ftz.f32 %f19487, %f4084, %f19486; mul.ftz.f32 %f29981, %f19483, %f19487; mul.ftz.f32 %f19488, %f19483, %f19486; div.approx.ftz.f32 %f30103, %f19488, %f4082; bra.uni $L__BB107_2623; $L__BB107_2764: ld.local.f32 %f4549, [%rd5+-52]; ld.local.v2.f32 {%f20783, %f20784}, [%rd5+-48]; ld.local.f32 %f4552, [%rd564+8]; ld.local.f32 %f4553, [%rd564+12]; mul.ftz.f32 %f20785, %f30259, %f4553; fma.rn.ftz.f32 %f20786, %f30260, %f4552, %f20785; ld.local.f32 %f4554, [%rd564+16]; fma.rn.ftz.f32 %f4555, %f30258, %f4554, %f20786; setp.leu.ftz.f32 %p2527, %f4555, 0f00000000; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2527 bra $L__BB107_2772; mov.f32 %f30104, 0f00000000; max.ftz.f32 %f20790, %f4555, %f30104; mul.ftz.f32 %f30103, %f20790, 0f3EA2F983; setp.le.ftz.f32 %p2528, %f4555, 0f00000000; @%p2528 bra $L__BB107_2772; mul.ftz.f32 %f20791, %f20783, %f4553; fma.rn.ftz.f32 %f20792, %f4549, %f4552, %f20791; fma.rn.ftz.f32 %f20793, %f20784, %f4554, %f20792; mov.f32 %f20794, 0f3F800000; sub.ftz.f32 %f20795, %f20794, %f20793; mov.f32 %f20796, 0f00000000; max.ftz.f32 %f20797, %f20795, %f20796; min.ftz.f32 %f20798, %f20797, %f20794; mul.ftz.f32 %f20799, %f20798, %f20798; mul.ftz.f32 %f20800, %f20799, %f20799; mul.ftz.f32 %f4557, %f20798, %f20800; sub.ftz.f32 %f20801, %f20794, %f4555; max.ftz.f32 %f20802, %f20801, %f20796; min.ftz.f32 %f20803, %f20802, %f20794; mul.ftz.f32 %f20804, %f20803, %f20803; mul.ftz.f32 %f20805, %f20804, %f20804; mul.ftz.f32 %f4558, %f20803, %f20805; ld.local.u32 %r988, [%rd564+24]; and.b32 %r4147, %r988, 3; setp.eq.s32 %p2529, %r4147, 0; fma.rn.ftz.f32 %f4559, %f4558, 0fBF000000, 0f3F800000; @%p2529 bra $L__BB107_2768; fma.rn.ftz.f32 %f20806, %f4557, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f30097, %f20806, %f4559, 0f00000000; bra.uni $L__BB107_2769; $L__BB107_2760: ld.local.f32 %f4531, [%rd564+8]; ld.local.f32 %f4532, [%rd564+12]; mul.ftz.f32 %f20746, %f30259, %f4532; fma.rn.ftz.f32 %f20747, %f30260, %f4531, %f20746; ld.local.f32 %f4533, [%rd564+16]; fma.rn.ftz.f32 %f4534, %f30258, %f4533, %f20747; setp.leu.ftz.f32 %p2523, %f4534, 0f00000000; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2523 bra $L__BB107_2763; ld.local.f32 %f20752, [%rd5+-52]; ld.local.v2.f32 {%f20753, %f20754}, [%rd5+-48]; add.ftz.f32 %f20757, %f30260, %f20752; add.ftz.f32 %f20758, %f30259, %f20753; mul.ftz.f32 %f20759, %f20758, %f20758; fma.rn.ftz.f32 %f20760, %f20757, %f20757, %f20759; add.ftz.f32 %f20761, %f30258, %f20754; fma.rn.ftz.f32 %f20762, %f20761, %f20761, %f20760; sqrt.approx.ftz.f32 %f20763, %f20762; div.approx.ftz.f32 %f4535, %f20757, %f20763; div.approx.ftz.f32 %f4536, %f20758, %f20763; div.approx.ftz.f32 %f4537, %f20761, %f20763; mul.ftz.f32 %f20764, %f20753, %f4532; fma.rn.ftz.f32 %f20765, %f20752, %f4531, %f20764; fma.rn.ftz.f32 %f20766, %f20754, %f4533, %f20765; setp.lt.ftz.f32 %p2524, %f20766, 0f00000000; setp.lt.ftz.f32 %p2525, %f4534, 0f00000000; or.pred %p2526, %p2525, %p2524; mov.f32 %f30104, %f30103; @%p2526 bra $L__BB107_2763; mov.f32 %f20767, 0f00000000; max.ftz.f32 %f20768, %f4534, %f20767; mul.ftz.f32 %f30103, %f20768, 0f3EA2F983; mul.ftz.f32 %f20769, %f30259, %f4536; fma.rn.ftz.f32 %f20770, %f30260, %f4535, %f20769; fma.rn.ftz.f32 %f20771, %f30258, %f4537, %f20770; mov.f32 %f20772, 0f3F800000; sub.ftz.f32 %f20773, %f20772, %f20771; max.ftz.f32 %f20774, %f20773, %f20767; min.ftz.f32 %f20775, %f20774, %f20772; mul.ftz.f32 %f20776, %f20775, %f20775; mul.ftz.f32 %f20777, %f20776, %f20776; mul.ftz.f32 %f20778, %f20775, %f20777; mul.ftz.f32 %f30104, %f4534, %f20778; $L__BB107_2763: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2721: setp.leu.ftz.f32 %p2480, %f4415, 0f00000000; setp.leu.ftz.f32 %p2481, %f4414, 0f00000000; or.pred %p2482, %p2481, %p2480; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; @%p2482 bra $L__BB107_2774; add.ftz.f32 %f20231, %f30260, %f4404; add.ftz.f32 %f20232, %f30259, %f20216; mul.ftz.f32 %f20233, %f20232, %f20232; fma.rn.ftz.f32 %f20234, %f20231, %f20231, %f20233; add.ftz.f32 %f20235, %f30258, %f20217; fma.rn.ftz.f32 %f20236, %f20235, %f20235, %f20234; sqrt.approx.ftz.f32 %f20237, %f20236; div.approx.ftz.f32 %f4416, %f20231, %f20237; div.approx.ftz.f32 %f4417, %f20232, %f20237; div.approx.ftz.f32 %f4418, %f20235, %f20237; mul.ftz.f32 %f20238, %f4412, %f4417; fma.rn.ftz.f32 %f20239, %f4411, %f4416, %f20238; fma.rn.ftz.f32 %f4419, %f4413, %f4418, %f20239; mul.ftz.f32 %f4420, %f4419, %f4419; setp.eq.ftz.f32 %p2483, %f20218, %f20219; @%p2483 bra $L__BB107_2724; bra.uni $L__BB107_2723; $L__BB107_2724: setp.eq.s32 %p2484, %r963, 12; @%p2484 bra $L__BB107_2726; bra.uni $L__BB107_2725; $L__BB107_2726: setp.ge.ftz.f32 %p2485, %f20218, 0f3F800000; mov.f32 %f30071, 0f3EA2F983; mov.f32 %f30070, 0f3D800000; @%p2485 bra $L__BB107_2728; mul.ftz.f32 %f20339, %f20218, %f20218; add.ftz.f32 %f20340, %f20339, 0fBF800000; mul.ftz.f32 %f20341, %f4419, %f20340; fma.rn.ftz.f32 %f20342, %f4419, %f20341, 0f3F800000; lg2.approx.ftz.f32 %f20343, %f20339; mul.ftz.f32 %f20344, %f20343, 0f3F317218; mul.ftz.f32 %f20345, %f20344, 0f40490FDB; mul.ftz.f32 %f20346, %f20342, %f20345; div.approx.ftz.f32 %f30071, %f20340, %f20346; bra.uni $L__BB107_2728; $L__BB107_2618: ld.local.v2.f32 {%f19327, %f19328}, [%rd5+-48]; ld.local.f32 %f4051, [%rd564+8]; ld.local.f32 %f4052, [%rd5+-52]; ld.local.f32 %f4053, [%rd564+12]; mul.ftz.f32 %f19329, %f19327, %f4053; fma.rn.ftz.f32 %f19330, %f4052, %f4051, %f19329; ld.local.f32 %f4054, [%rd564+16]; fma.rn.ftz.f32 %f4055, %f19328, %f4054, %f19330; mul.ftz.f32 %f19331, %f30259, %f4053; fma.rn.ftz.f32 %f19332, %f30260, %f4051, %f19331; fma.rn.ftz.f32 %f4056, %f30258, %f4054, %f19332; setp.leu.ftz.f32 %p2369, %f4055, 0f00000000; setp.leu.ftz.f32 %p2370, %f4056, 0f00000000; or.pred %p2371, %p2370, %p2369; mov.f32 %f30103, 0f00000000; mov.f32 %f30104, %f30103; @%p2371 bra $L__BB107_2621; add.ftz.f32 %f19337, %f30260, %f4052; add.ftz.f32 %f19338, %f30259, %f19327; mul.ftz.f32 %f19339, %f19338, %f19338; fma.rn.ftz.f32 %f19340, %f19337, %f19337, %f19339; add.ftz.f32 %f19341, %f30258, %f19328; fma.rn.ftz.f32 %f19342, %f19341, %f19341, %f19340; sqrt.approx.ftz.f32 %f19343, %f19342; div.approx.ftz.f32 %f19344, %f19337, %f19343; div.approx.ftz.f32 %f19345, %f19338, %f19343; div.approx.ftz.f32 %f19346, %f19341, %f19343; mul.ftz.f32 %f19347, %f4053, %f19345; fma.rn.ftz.f32 %f19348, %f4051, %f19344, %f19347; fma.rn.ftz.f32 %f4057, %f4054, %f19346, %f19348; mul.ftz.f32 %f19349, %f19327, %f19345; fma.rn.ftz.f32 %f19350, %f4052, %f19344, %f19349; fma.rn.ftz.f32 %f19351, %f19328, %f19346, %f19350; abs.ftz.f32 %f4058, %f19351; abs.ftz.f32 %f19352, %f4057; setp.geu.ftz.f32 %p2372, %f19352, 0f3F7FFF58; setp.leu.ftz.f32 %p2373, %f4058, 0f3727C5AC; or.pred %p2374, %p2373, %p2372; mov.f32 %f30104, %f30103; @%p2374 bra $L__BB107_2621; ld.local.f32 %f19354, [%rd564+24]; div.approx.ftz.f32 %f19355, %f4057, %f4058; mov.f32 %f19356, 0f3727C5AC; max.ftz.f32 %f19357, %f19355, %f19356; mul.ftz.f32 %f19358, %f4055, %f19357; abs.ftz.f32 %f19359, %f19358; add.ftz.f32 %f19360, %f19359, %f19359; mul.ftz.f32 %f19361, %f4056, %f19357; abs.ftz.f32 %f19362, %f19361; add.ftz.f32 %f19363, %f19362, %f19362; mul.ftz.f32 %f19364, %f4057, %f4057; mov.f32 %f19365, 0f3F800000; sub.ftz.f32 %f19366, %f19365, %f19364; mul.ftz.f32 %f19367, %f19366, %f19366; div.approx.ftz.f32 %f19368, %f19364, %f19366; mul.ftz.f32 %f19369, %f19354, %f19368; mul.ftz.f32 %f19370, %f19369, 0fBFB8AA3B; ex2.approx.ftz.f32 %f19371, %f19370; mul.ftz.f32 %f19372, %f19354, %f19371; mul.ftz.f32 %f19373, %f19372, 0f3EA2F983; div.approx.ftz.f32 %f19374, %f19373, %f19367; min.ftz.f32 %f19375, %f19360, %f19363; min.ftz.f32 %f19376, %f19365, %f19375; mul.ftz.f32 %f19377, %f19374, %f19376; mul.ftz.f32 %f19378, %f19377, 0f3E800000; div.approx.ftz.f32 %f30104, %f19378, %f4055; mov.f32 %f30103, 0f3E22F983; $L__BB107_2621: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2612: ld.local.v2.f32 {%f19265, %f19266}, [%rd5+-48]; ld.local.f32 %f19267, [%rd564+20]; mul.ftz.f32 %f4027, %f19267, 0f3FC90FDB; ld.local.f32 %f19268, [%rd564+24]; mul.ftz.f32 %f4028, %f19268, 0f3FC90FDB; ld.local.f32 %f4029, [%rd564+8]; ld.local.f32 %f4030, [%rd564+12]; mul.ftz.f32 %f19269, %f30259, %f4030; fma.rn.ftz.f32 %f19270, %f30260, %f4029, %f19269; ld.local.f32 %f4031, [%rd564+16]; fma.rn.ftz.f32 %f19271, %f30258, %f4031, %f19270; ld.local.f32 %f4032, [%rd5+-52]; mul.ftz.f32 %f19272, %f19265, %f4030; fma.rn.ftz.f32 %f19273, %f4032, %f4029, %f19272; fma.rn.ftz.f32 %f4033, %f19266, %f4031, %f19273; setp.leu.ftz.f32 %p2360, %f19271, 0f00000000; setp.leu.ftz.f32 %p2361, %f4033, 0f00000000; mov.f32 %f30103, 0f00000000; or.pred %p2362, %p2360, %p2361; mov.f32 %f30104, %f30103; @%p2362 bra $L__BB107_2617; add.ftz.f32 %f19275, %f4033, %f4033; mul.ftz.f32 %f19276, %f19275, %f4029; mul.ftz.f32 %f19277, %f19275, %f4030; mul.ftz.f32 %f19278, %f19275, %f4031; sub.ftz.f32 %f19279, %f19276, %f4032; sub.ftz.f32 %f19280, %f19277, %f19265; sub.ftz.f32 %f19281, %f19278, %f19266; mul.ftz.f32 %f19282, %f30259, %f19280; fma.rn.ftz.f32 %f19283, %f30260, %f19279, %f19282; fma.rn.ftz.f32 %f19284, %f30258, %f19281, %f19283; mov.f32 %f19285, 0f00000000; max.ftz.f32 %f19286, %f19284, %f19285; mov.f32 %f19287, 0fBF800000; max.ftz.f32 %f19288, %f19286, %f19287; mov.f32 %f19274, 0f3F800000; min.ftz.f32 %f19289, %f19288, %f19274; abs.ftz.f32 %f19290, %f19289; sub.ftz.f32 %f19291, %f19274, %f19290; mul.ftz.f32 %f19292, %f19291, 0f3F000000; sqrt.approx.ftz.f32 %f19293, %f19292; setp.gt.ftz.f32 %p2363, %f19290, 0f3F11EB85; selp.f32 %f19294, %f19293, %f19290, %p2363; mul.ftz.f32 %f19295, %f19294, %f19294; mov.f32 %f19296, 0f3C94D2E9; mov.f32 %f19297, 0f3D53F941; fma.rn.ftz.f32 %f19298, %f19297, %f19295, %f19296; mov.f32 %f19299, 0f3D3F841F; fma.rn.ftz.f32 %f19300, %f19298, %f19295, %f19299; mov.f32 %f19301, 0f3D994929; fma.rn.ftz.f32 %f19302, %f19300, %f19295, %f19301; mov.f32 %f19303, 0f3E2AAB94; fma.rn.ftz.f32 %f19304, %f19302, %f19295, %f19303; mul.ftz.f32 %f19305, %f19295, %f19304; fma.rn.ftz.f32 %f19306, %f19305, %f19294, %f19294; add.ftz.f32 %f19307, %f19306, %f19306; mov.f32 %f19308, 0f3FC90FDB; sub.ftz.f32 %f19309, %f19308, %f19306; selp.f32 %f19310, %f19307, %f19309, %p2363; setp.lt.ftz.f32 %p2364, %f19289, 0f00000000; mov.f32 %f19311, 0f40490FDB; sub.ftz.f32 %f19312, %f19311, %f19310; selp.f32 %f4034, %f19312, %f19310, %p2364; setp.lt.ftz.f32 %p2365, %f4034, %f4027; add.ftz.f32 %f4035, %f4027, %f4028; mov.f32 %f29974, %f19274; @%p2365 bra $L__BB107_2616; setp.leu.ftz.f32 %p2366, %f4035, %f4034; setp.eq.ftz.f32 %p2367, %f4028, 0f00000000; or.pred %p2368, %p2367, %p2366; mov.f32 %f29974, %f19285; @%p2368 bra $L__BB107_2616; sub.ftz.f32 %f19314, %f4034, %f4027; div.approx.ftz.f32 %f19315, %f19314, %f4028; mov.f32 %f19316, 0f3F800000; sub.ftz.f32 %f29974, %f19316, %f19315; $L__BB107_2616: min.ftz.f32 %f19318, %f4035, %f19308; cos.approx.ftz.f32 %f19319, %f19318; sub.ftz.f32 %f19321, %f19274, %f19319; mov.f32 %f19322, 0f3E22F983; div.approx.ftz.f32 %f30103, %f19322, %f19321; mul.ftz.f32 %f30104, %f29974, %f30103; $L__BB107_2617: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2740: ld.local.f32 %f20478, [%rd564+8]; ld.local.f32 %f20479, [%rd564+12]; mul.ftz.f32 %f20480, %f30259, %f20479; fma.rn.ftz.f32 %f20481, %f30260, %f20478, %f20480; ld.local.f32 %f20482, [%rd564+16]; fma.rn.ftz.f32 %f20483, %f30258, %f20482, %f20481; setp.ge.ftz.f32 %p2493, %f20483, 0f00000000; mov.f32 %f30103, 0f00000000; @%p2493 bra $L__BB107_2758; ld.local.f32 %f20488, [%rd5+-52]; ld.local.v2.f32 {%f20489, %f20490}, [%rd5+-48]; ld.local.v2.f32 {%f20493, %f20494}, [%rd564+36]; ld.local.v4.f32 {%f20495, %f20496, %f20497, %f20498}, [%rd564+20]; mul.ftz.f32 %f20502, %f20489, %f20496; fma.rn.ftz.f32 %f20503, %f20488, %f20495, %f20502; fma.rn.ftz.f32 %f20504, %f20490, %f20497, %f20503; mul.ftz.f32 %f20505, %f20495, %f20504; mul.ftz.f32 %f20506, %f20496, %f20504; mul.ftz.f32 %f20507, %f20497, %f20504; sub.ftz.f32 %f20508, %f20488, %f20505; sub.ftz.f32 %f20509, %f20489, %f20506; sub.ftz.f32 %f20510, %f20490, %f20507; mul.ftz.f32 %f20511, %f20509, %f20509; fma.rn.ftz.f32 %f20512, %f20508, %f20508, %f20511; fma.rn.ftz.f32 %f20513, %f20510, %f20510, %f20512; sqrt.approx.ftz.f32 %f20514, %f20513; div.approx.ftz.f32 %f20515, %f20508, %f20514; div.approx.ftz.f32 %f20516, %f20509, %f20514; div.approx.ftz.f32 %f20517, %f20510, %f20514; abs.ftz.f32 %f20518, %f20504; setp.lt.ftz.f32 %p2494, %f20518, 0f3F800000; mov.f32 %f20519, 0f3F800000; add.ftz.f32 %f20520, %f20518, 0fBF800000; add.ftz.f32 %f20521, %f20520, 0f3F800000; selp.f32 %f20522, %f20521, 0f3F800000, %p2494; sub.ftz.f32 %f20523, %f20519, %f20522; sqrt.approx.ftz.f32 %f20524, %f20523; mul.ftz.f32 %f20525, %f20522, 0f3CB1485D; mov.f32 %f20526, 0f3D9FB43B; sub.ftz.f32 %f20527, %f20526, %f20525; fma.rn.ftz.f32 %f20528, %f20522, %f20527, 0fBE5A6B93; fma.rn.ftz.f32 %f20529, %f20522, %f20528, 0f3FC90FDB; mov.f32 %f20530, 0f3FC90FDB; mul.ftz.f32 %f20531, %f20524, %f20529; setp.lt.ftz.f32 %p2495, %f20504, 0f00000000; mov.f32 %f20532, 0f40490FDB; sub.ftz.f32 %f20533, %f20532, %f20531; selp.f32 %f20534, %f20533, %f20531, %p2495; sub.ftz.f32 %f4481, %f20530, %f20534; mul.ftz.f32 %f20535, %f30259, %f20496; fma.rn.ftz.f32 %f20536, %f30260, %f20495, %f20535; fma.rn.ftz.f32 %f20537, %f30258, %f20497, %f20536; mul.ftz.f32 %f20538, %f20495, %f20537; mul.ftz.f32 %f20539, %f20496, %f20537; mul.ftz.f32 %f20540, %f20497, %f20537; sub.ftz.f32 %f20541, %f30260, %f20538; sub.ftz.f32 %f20542, %f30259, %f20539; sub.ftz.f32 %f20543, %f30258, %f20540; mul.ftz.f32 %f20544, %f20542, %f20542; fma.rn.ftz.f32 %f20545, %f20541, %f20541, %f20544; fma.rn.ftz.f32 %f20546, %f20543, %f20543, %f20545; sqrt.approx.ftz.f32 %f20547, %f20546; div.approx.ftz.f32 %f20548, %f20541, %f20547; div.approx.ftz.f32 %f20549, %f20542, %f20547; div.approx.ftz.f32 %f20550, %f20543, %f20547; abs.ftz.f32 %f20551, %f20537; setp.lt.ftz.f32 %p2496, %f20551, 0f3F800000; add.ftz.f32 %f20552, %f20551, 0fBF800000; add.ftz.f32 %f20553, %f20552, 0f3F800000; selp.f32 %f20554, %f20553, 0f3F800000, %p2496; sub.ftz.f32 %f20555, %f20519, %f20554; sqrt.approx.ftz.f32 %f20556, %f20555; mul.ftz.f32 %f20557, %f20554, 0f3CB1485D; sub.ftz.f32 %f20558, %f20526, %f20557; fma.rn.ftz.f32 %f20559, %f20554, %f20558, 0fBE5A6B93; fma.rn.ftz.f32 %f20560, %f20554, %f20559, 0f3FC90FDB; mul.ftz.f32 %f20561, %f20556, %f20560; setp.lt.ftz.f32 %p2497, %f20537, 0f00000000; sub.ftz.f32 %f20562, %f20532, %f20561; selp.f32 %f20563, %f20562, %f20561, %p2497; sub.ftz.f32 %f4482, %f20530, %f20563; mul.ftz.f32 %f20564, %f20516, %f20549; fma.rn.ftz.f32 %f20565, %f20515, %f20548, %f20564; fma.rn.ftz.f32 %f4483, %f20517, %f20550, %f20565; abs.ftz.f32 %f20566, %f4482; sub.ftz.f32 %f20567, %f20530, %f20566; setp.lt.ftz.f32 %p2498, %f20567, 0f3A83126F; @%p2498 bra $L__BB107_2758; abs.ftz.f32 %f20569, %f4483; setp.lt.ftz.f32 %p2499, %f20569, 0f3F800000; add.ftz.f32 %f20571, %f20569, 0fBF800000; add.ftz.f32 %f20572, %f20571, 0f3F800000; selp.f32 %f20573, %f20572, 0f3F800000, %p2499; sub.ftz.f32 %f20574, %f20519, %f20573; sqrt.approx.ftz.f32 %f20575, %f20574; fma.rn.ftz.f32 %f20576, %f20573, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f20577, %f20573, %f20576, 0fBE5A6B93; fma.rn.ftz.f32 %f20578, %f20573, %f20577, 0f3FC90FDB; mul.ftz.f32 %f20579, %f20575, %f20578; sub.ftz.f32 %f20581, %f20532, %f20579; setp.lt.ftz.f32 %p2500, %f4483, 0f00000000; mov.f32 %f30089, 0f00000000; selp.f32 %f4484, %f20581, %f20579, %p2500; mul.ftz.f32 %f20582, %f4482, 0f3EA2F983; mov.b32 %r4129, %f20582; and.b32 %r4130, %r4129, -2147483648; or.b32 %r4131, %r4130, 1056964608; mov.b32 %f20583, %r4131; add.ftz.f32 %f20584, %f20582, %f20583; cvt.rzi.ftz.s32.f32 %r4132, %f20584; cvt.rn.f32.s32 %f20585, %r4132; fma.rn.ftz.f32 %f20586, %f20585, 0fC0490000, %f4482; fma.rn.ftz.f32 %f20587, %f20585, 0fBA7DA000, %f20586; fma.rn.ftz.f32 %f20588, %f20585, 0fB4222000, %f20587; fma.rn.ftz.f32 %f20589, %f20585, 0fACB4611A, %f20588; add.ftz.f32 %f20590, %f20589, 0fBFC90FDB; add.ftz.f32 %f20591, %f20590, 0f3FC90FDB; mul.ftz.f32 %f20592, %f20591, %f20591; fma.rn.ftz.f32 %f20593, %f20592, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f20594, %f20592, %f20593, 0fBAB60B58; fma.rn.ftz.f32 %f20595, %f20592, %f20594, 0f3D2AAAAA; fma.rn.ftz.f32 %f20596, %f20592, %f20595, 0fBF000000; fma.rn.ftz.f32 %f20597, %f20592, %f20596, 0f3F800000; and.b32 %r4133, %r4132, 1; setp.eq.b32 %p2501, %r4133, 1; neg.ftz.f32 %f20598, %f20597; selp.f32 %f4485, %f20598, %f20597, %p2501; add.ftz.f32 %f20599, %f4481, 0f3FC90FDB; mov.f32 %f20600, 0f40000000; div.approx.ftz.f32 %f20601, %f20599, %f20600; sub.ftz.f32 %f20602, %f20601, %f20494; rcp.approx.ftz.f32 %f4486, %f20498; mul.ftz.f32 %f4487, %f4486, %f20602; abs.ftz.f32 %f4488, %f4487; setp.eq.ftz.f32 %p2502, %f4488, 0f00000000; mov.f32 %f30087, %f30089; @%p2502 bra $L__BB107_2747; setp.eq.ftz.f32 %p2503, %f4488, 0f3F800000; mov.f32 %f30087, %f20519; @%p2503 bra $L__BB107_2747; setp.gt.ftz.f32 %p2504, %f4488, 0f3F800000; @%p2504 bra $L__BB107_2746; bra.uni $L__BB107_2745; $L__BB107_2746: mov.f32 %f20605, 0f3F800000; div.approx.ftz.f32 %f30087, %f20605, %f4488; bra.uni $L__BB107_2747; $L__BB107_2459: ld.local.u64 %rd566, [%rd564+52]; ld.v4.f32 {%f18397, %f18398, %f18399, %f18400}, [%rd566]; ld.local.v2.f32 {%f29933, %f29934}, [%rd5+32]; ld.local.f32 %f29932, [%rd5+28]; mul.ftz.f32 %f18403, %f29933, %f29933; fma.rn.ftz.f32 %f18404, %f29932, %f29932, %f18403; fma.rn.ftz.f32 %f18405, %f29934, %f29934, %f18404; sqrt.approx.ftz.f32 %f3663, %f18405; setp.neu.ftz.f32 %p2217, %f3663, 0f00000000; @%p2217 bra $L__BB107_2460; bra.uni $L__BB107_2461; $L__BB107_2460: rcp.approx.ftz.f32 %f18406, %f3663; mul.ftz.f32 %f29934, %f29934, %f18406; mul.ftz.f32 %f29933, %f29933, %f18406; mul.ftz.f32 %f29932, %f29932, %f18406; $L__BB107_2461: mul.ftz.f32 %f18407, %f18399, %f29933; mul.ftz.f32 %f18408, %f18398, %f29934; sub.ftz.f32 %f29935, %f18407, %f18408; mul.ftz.f32 %f18409, %f18397, %f29934; mul.ftz.f32 %f18410, %f18399, %f29932; sub.ftz.f32 %f29936, %f18409, %f18410; mul.ftz.f32 %f18411, %f18398, %f29932; mul.ftz.f32 %f18412, %f18397, %f29933; sub.ftz.f32 %f29937, %f18411, %f18412; mul.ftz.f32 %f18413, %f29936, %f29936; fma.rn.ftz.f32 %f18414, %f29935, %f29935, %f18413; fma.rn.ftz.f32 %f18415, %f29937, %f29937, %f18414; sqrt.approx.ftz.f32 %f3679, %f18415; setp.neu.ftz.f32 %p2218, %f3679, 0f00000000; @%p2218 bra $L__BB107_2462; bra.uni $L__BB107_2463; $L__BB107_2462: rcp.approx.ftz.f32 %f18416, %f3679; mul.ftz.f32 %f29937, %f18416, %f29937; mul.ftz.f32 %f29936, %f18416, %f29936; mul.ftz.f32 %f29935, %f18416, %f29935; $L__BB107_2463: ld.local.f32 %f18417, [%rd5+-52]; ld.local.v2.f32 {%f18418, %f18419}, [%rd5+-48]; mul.ftz.f32 %f18422, %f29933, %f18418; fma.rn.ftz.f32 %f18423, %f29932, %f18417, %f18422; fma.rn.ftz.f32 %f3689, %f29934, %f18419, %f18423; mul.ftz.f32 %f18424, %f18398, %f18418; fma.rn.ftz.f32 %f18425, %f18397, %f18417, %f18424; fma.rn.ftz.f32 %f18426, %f18399, %f18419, %f18425; mul.ftz.f32 %f18427, %f29936, %f18418; fma.rn.ftz.f32 %f18428, %f29935, %f18417, %f18427; fma.rn.ftz.f32 %f18429, %f29937, %f18419, %f18428; mul.ftz.f32 %f18430, %f30259, %f29933; fma.rn.ftz.f32 %f18431, %f30260, %f29932, %f18430; fma.rn.ftz.f32 %f3690, %f30258, %f29934, %f18431; mul.ftz.f32 %f18432, %f30259, %f18398; fma.rn.ftz.f32 %f18433, %f30260, %f18397, %f18432; fma.rn.ftz.f32 %f3691, %f30258, %f18399, %f18433; mul.ftz.f32 %f18434, %f30259, %f29936; fma.rn.ftz.f32 %f18435, %f30260, %f29935, %f18434; fma.rn.ftz.f32 %f3692, %f30258, %f29937, %f18435; mul.ftz.f32 %f3693, %f3689, %f3689; mov.f32 %f18436, 0f3F800000; sub.ftz.f32 %f18437, %f18436, %f3693; mov.f32 %f18438, 0f00000000; max.ftz.f32 %f18439, %f18437, %f18438; sqrt.approx.ftz.f32 %f3694, %f18439; abs.ftz.f32 %f3695, %f18426; abs.ftz.f32 %f3696, %f18429; setp.eq.ftz.f32 %p2219, %f3695, 0f00000000; setp.eq.ftz.f32 %p2220, %f3696, 0f00000000; and.pred %p2221, %p2219, %p2220; mov.b32 %r973, %f18426; mov.b32 %r4048, %f18429; and.b32 %r974, %r4048, -2147483648; @%p2221 bra $L__BB107_2467; bra.uni $L__BB107_2464; $L__BB107_2467: shr.s32 %r4053, %r973, 31; and.b32 %r4054, %r4053, 1078530011; or.b32 %r4055, %r974, %r4054; mov.b32 %f29938, %r4055; bra.uni $L__BB107_2468; $L__BB107_2759: ld.local.f32 %f20675, [%rd564+8]; ld.local.f32 %f20676, [%rd564+12]; mul.ftz.f32 %f20677, %f30259, %f20676; fma.rn.ftz.f32 %f20678, %f30260, %f20675, %f20677; ld.local.f32 %f20679, [%rd564+16]; fma.rn.ftz.f32 %f20680, %f30258, %f20679, %f20678; neg.ftz.f32 %f20681, %f20680; mov.f32 %f20682, 0f00000000; max.ftz.f32 %f20683, %f20681, %f20682; mul.ftz.f32 %f30103, %f20683, 0f3EA2F983; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; bra.uni $L__BB107_2774; $L__BB107_2284: ld.local.f32 %f4342, [%rd564+8]; ld.local.f32 %f4343, [%rd564+12]; mul.ftz.f32 %f20002, %f19994, %f4343; fma.rn.ftz.f32 %f20003, %f4336, %f4342, %f20002; ld.local.f32 %f4344, [%rd564+16]; fma.rn.ftz.f32 %f4345, %f19995, %f4344, %f20003; mul.ftz.f32 %f20004, %f30259, %f4343; fma.rn.ftz.f32 %f20005, %f30260, %f4342, %f20004; fma.rn.ftz.f32 %f4346, %f30258, %f4344, %f20005; setp.lt.ftz.f32 %p2460, %f4346, 0f00000000; @%p2460 bra $L__BB107_2712; bra.uni $L__BB107_2285; $L__BB107_2712: setp.le.ftz.f32 %p2469, %f4345, 0f00000000; setp.ge.ftz.f32 %p2470, %f4346, 0f00000000; or.pred %p2471, %p2469, %p2470; mov.f32 %f30104, %f30103; @%p2471 bra $L__BB107_2718; ld.local.f32 %f4382, [%rd564+28]; fma.rn.ftz.f32 %f20146, %f30260, %f4382, %f4336; fma.rn.ftz.f32 %f20147, %f30259, %f4382, %f19994; fma.rn.ftz.f32 %f20148, %f30258, %f4382, %f19995; neg.ftz.f32 %f20149, %f20146; neg.ftz.f32 %f20150, %f20147; neg.ftz.f32 %f20151, %f20148; mul.ftz.f32 %f20152, %f20147, %f20147; fma.rn.ftz.f32 %f20153, %f20146, %f20146, %f20152; fma.rn.ftz.f32 %f4383, %f20148, %f20148, %f20153; sqrt.approx.ftz.f32 %f20154, %f4383; div.approx.ftz.f32 %f20155, %f20149, %f20154; div.approx.ftz.f32 %f20156, %f20150, %f20154; div.approx.ftz.f32 %f20157, %f20151, %f20154; mul.ftz.f32 %f20158, %f19994, %f20156; fma.rn.ftz.f32 %f20159, %f4336, %f20155, %f20158; fma.rn.ftz.f32 %f4384, %f19995, %f20157, %f20159; mul.ftz.f32 %f20160, %f30259, %f20156; fma.rn.ftz.f32 %f20161, %f30260, %f20155, %f20160; fma.rn.ftz.f32 %f4385, %f30258, %f20157, %f20161; mul.ftz.f32 %f20162, %f4343, %f20156; fma.rn.ftz.f32 %f20163, %f4342, %f20155, %f20162; fma.rn.ftz.f32 %f20164, %f4344, %f20157, %f20163; mov.f32 %f30067, 0f3F800000; min.ftz.f32 %f20165, %f20164, %f30067; mul.ftz.f32 %f20166, %f20165, %f20165; sub.ftz.f32 %f20167, %f30067, %f20166; div.approx.ftz.f32 %f20168, %f20167, %f20166; mul.ftz.f32 %f20169, %f20166, %f20166; neg.ftz.f32 %f20170, %f20168; div.approx.ftz.f32 %f20171, %f20170, %f4341; mul.ftz.f32 %f20172, %f20171, 0f3FB8AA3B; ex2.approx.ftz.f32 %f20173, %f20172; mul.ftz.f32 %f20174, %f4341, 0f40490FDB; mul.ftz.f32 %f20175, %f20174, %f20169; div.approx.ftz.f32 %f4386, %f20173, %f20175; mul.ftz.f32 %f20176, %f4345, %f4345; sub.ftz.f32 %f20177, %f30067, %f20176; div.approx.ftz.f32 %f20178, %f20177, %f20176; mov.f32 %f20179, 0f00000000; max.ftz.f32 %f20180, %f20178, %f20179; sqrt.approx.ftz.f32 %f20181, %f20180; mul.ftz.f32 %f4387, %f19996, %f20181; setp.lt.ftz.f32 %p2472, %f4387, 0f3F200000; mov.f32 %f30066, %f30067; @%p2472 bra $L__BB107_2715; rcp.approx.ftz.f32 %f20182, %f4387; fma.rn.ftz.f32 %f20183, %f20182, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20184, %f20182, %f20183; fma.rn.ftz.f32 %f20185, %f20182, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20186, %f20182, %f20185, 0f3F800000; div.approx.ftz.f32 %f30066, %f20184, %f20186; $L__BB107_2715: mul.ftz.f32 %f20188, %f4346, %f4346; sub.ftz.f32 %f20189, %f30067, %f20188; div.approx.ftz.f32 %f20190, %f20189, %f20188; max.ftz.f32 %f20192, %f20190, %f20179; sqrt.approx.ftz.f32 %f20193, %f20192; mul.ftz.f32 %f4390, %f19996, %f20193; setp.lt.ftz.f32 %p2473, %f4390, 0f3F200000; @%p2473 bra $L__BB107_2717; rcp.approx.ftz.f32 %f20194, %f4390; fma.rn.ftz.f32 %f20195, %f20194, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20196, %f20194, %f20195; fma.rn.ftz.f32 %f20197, %f20194, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20198, %f20194, %f20197, 0f3F800000; div.approx.ftz.f32 %f30067, %f20196, %f20198; $L__BB107_2717: mul.ftz.f32 %f20199, %f4382, %f4382; mul.ftz.f32 %f20200, %f20199, %f4386; mul.ftz.f32 %f20201, %f4345, %f4383; div.approx.ftz.f32 %f20202, %f20200, %f20201; mul.ftz.f32 %f20203, %f4385, %f4384; abs.ftz.f32 %f20204, %f20203; mul.ftz.f32 %f20205, %f30066, %f30067; mul.ftz.f32 %f20206, %f20205, %f20204; mul.ftz.f32 %f20207, %f30066, %f20204; mul.ftz.f32 %f30103, %f20202, %f20207; mul.ftz.f32 %f30104, %f20202, %f20206; bra.uni $L__BB107_2718; $L__BB107_2773: ld.local.f32 %f20844, [%rd564+8]; ld.local.f32 %f20845, [%rd564+12]; mul.ftz.f32 %f20846, %f30259, %f20845; fma.rn.ftz.f32 %f20847, %f30260, %f20844, %f20846; ld.local.f32 %f20848, [%rd564+16]; fma.rn.ftz.f32 %f20849, %f30258, %f20848, %f20847; mov.f32 %f20850, 0f00000000; max.ftz.f32 %f20851, %f20849, %f20850; mul.ftz.f32 %f30103, %f20851, 0f3EA2F983; mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; bra.uni $L__BB107_2774; $L__BB107_2325: mul.ftz.f32 %f17572, %f3242, %f3248; fma.rn.ftz.f32 %f17573, %f3241, %f3247, %f17572; fma.rn.ftz.f32 %f17574, %f3243, %f3249, %f17573; mul.ftz.f32 %f17575, %f3260, %f17574; mul.ftz.f32 %f17576, %f17575, %f3259; mul.ftz.f32 %f17577, %f3245, %f17576; mul.ftz.f32 %f17578, %f3245, %f17577; mul.ftz.f32 %f17579, %f3246, %f3267; mul.ftz.f32 %f17580, %f3246, %f17579; max.ftz.f32 %f17582, %f17580, %f17556; div.approx.ftz.f32 %f3268, %f17578, %f17582; setp.geu.ftz.f32 %p2092, %f3245, 0f3F800000; mov.f32 %f29819, %f3245; @%p2092 bra $L__BB107_2327; rcp.approx.ftz.f32 %f29819, %f3245; $L__BB107_2327: mov.f32 %f17583, 0f3F800000; max.ftz.f32 %f17584, %f29819, %f17583; mov.f32 %f17585, 0f40400000; min.ftz.f32 %f17586, %f17584, %f17585; add.ftz.f32 %f17587, %f17586, 0fBF2814F9; mul.ftz.f32 %f17588, %f17587, 0fBF7A7547; mul.ftz.f32 %f17589, %f17587, %f17588; mul.ftz.f32 %f17590, %f17589, 0f3FB8AA3B; ex2.approx.ftz.f32 %f17591, %f17590; mul.ftz.f32 %f17592, %f17586, 0fBD0A43BB; fma.rn.ftz.f32 %f17593, %f17591, 0f3D435661, %f17592; add.ftz.f32 %f17594, %f17593, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f17595, %f17470; fma.rn.ftz.f32 %f17596, %f17595, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f17597, %f17595, %f17596, 0f3D4D5C09; fma.rn.ftz.f32 %f17598, %f17595, %f17597, 0fBB86D80B; fma.rn.ftz.f32 %f17599, %f17595, %f17598, 0f3F80029F; fma.rn.ftz.f32 %f17600, %f17586, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f17601, %f17586, %f17600, 0fC2FFCFDF; fma.rn.ftz.f32 %f17602, %f17586, %f17601, 0f43927AA0; fma.rn.ftz.f32 %f17603, %f17586, %f17602, 0fC38FF917; mul.ftz.f32 %f17604, %f17586, %f17586; mov.f32 %f17605, 0f4347CD91; div.approx.ftz.f32 %f17606, %f17605, %f17604; add.ftz.f32 %f17607, %f17606, %f17603; mul.ftz.f32 %f17608, %f17586, %f17604; mov.f32 %f17609, 0f42CB5604; div.approx.ftz.f32 %f17610, %f17609, %f17608; sub.ftz.f32 %f17611, %f17607, %f17610; fma.rn.ftz.f32 %f17612, %f17595, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f17613, %f17595, %f17612, 0f41B5F319; fma.rn.ftz.f32 %f17614, %f17595, %f17613, 0fC05A14A5; fma.rn.ftz.f32 %f17615, %f17595, %f17614, 0f3DC9FFD6; fma.rn.ftz.f32 %f17616, %f17595, %f17615, 0f3BA1B61E; mul.ftz.f32 %f17617, %f17616, %f17611; fma.rn.ftz.f32 %f17618, %f17617, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f17619, %f17583, %f17594; sub.ftz.f32 %f17620, %f17583, %f17599; mul.ftz.f32 %f17621, %f17619, %f17620; fma.rn.ftz.f32 %f17622, %f17621, 0fC11520C5, %f17618; cvt.ftz.sat.f32.f32 %f17623, %f17622; abs.ftz.f32 %f17624, %f3268; sub.ftz.f32 %f17625, %f17583, %f17623; mul.ftz.f32 %f17626, %f3266, %f17625; fma.rn.ftz.f32 %f17627, %f17624, %f17623, %f17626; sub.ftz.f32 %f17628, %f17583, %f29818; mul.ftz.f32 %f30103, %f17628, %f17627; $L__BB107_2329: selp.u16 %rs243, 1, 0, %p7; ld.local.u64 %rd565, [%rd564+36]; ld.f32 %f3274, [%rd565]; ld.f32 %f3275, [%rd565+4]; ld.f32 %f3276, [%rd565+8]; setp.lt.ftz.f32 %p2093, %f3243, 0f00000000; @%p2093 bra $L__BB107_2331; bra.uni $L__BB107_2330; $L__BB107_2331: mov.f32 %f29821, %f3274; mov.f32 %f29822, %f3275; mov.f32 %f29823, %f3276; bra.uni $L__BB107_2332; $L__BB107_2464: setp.eq.ftz.f32 %p2222, %f3695, 0f7F800000; setp.eq.ftz.f32 %p2223, %f3696, 0f7F800000; and.pred %p2224, %p2222, %p2223; @%p2224 bra $L__BB107_2466; bra.uni $L__BB107_2465; $L__BB107_2466: setp.lt.s32 %p2228, %r973, 0; selp.b32 %r4051, 1075235812, 1061752795, %p2228; or.b32 %r4052, %r974, %r4051; mov.b32 %f29938, %r4052; bra.uni $L__BB107_2468; $L__BB107_2285: setp.leu.ftz.f32 %p2461, %f4345, 0f00000000; setp.leu.ftz.f32 %p2462, %f4346, 0f00000000; or.pred %p2463, %p2462, %p2461; mov.f32 %f30104, %f30103; @%p2463 bra $L__BB107_2718; add.ftz.f32 %f20010, %f30260, %f4336; add.ftz.f32 %f20011, %f30259, %f19994; mul.ftz.f32 %f20012, %f20011, %f20011; fma.rn.ftz.f32 %f20013, %f20010, %f20010, %f20012; add.ftz.f32 %f20014, %f30258, %f19995; fma.rn.ftz.f32 %f20015, %f20014, %f20014, %f20013; sqrt.approx.ftz.f32 %f20016, %f20015; div.approx.ftz.f32 %f4347, %f20010, %f20016; div.approx.ftz.f32 %f4348, %f20011, %f20016; div.approx.ftz.f32 %f4349, %f20014, %f20016; mul.ftz.f32 %f20017, %f4343, %f4348; fma.rn.ftz.f32 %f20018, %f4342, %f4347, %f20017; fma.rn.ftz.f32 %f4350, %f4344, %f4349, %f20018; mul.ftz.f32 %f4351, %f4350, %f4350; mul.ftz.f32 %f20019, %f4351, %f4351; mul.ftz.f32 %f20020, %f4341, 0f40490FDB; mul.ftz.f32 %f4352, %f20020, %f20019; mul.ftz.f32 %f4353, %f4345, %f4345; mov.f32 %f30065, 0f3F800000; sub.ftz.f32 %f4354, %f30065, %f4353; setp.eq.ftz.f32 %p2464, %f19996, %f19997; @%p2464 bra $L__BB107_2707; bra.uni $L__BB107_2287; $L__BB107_2707: mov.f32 %f30065, 0f3F800000; sub.ftz.f32 %f20111, %f30065, %f4351; div.approx.ftz.f32 %f20112, %f20111, %f4351; neg.ftz.f32 %f20113, %f20112; div.approx.ftz.f32 %f20114, %f20113, %f4341; mul.ftz.f32 %f20115, %f20114, 0f3FB8AA3B; ex2.approx.ftz.f32 %f20116, %f20115; div.approx.ftz.f32 %f30063, %f20116, %f4352; div.approx.ftz.f32 %f20117, %f4354, %f4353; mov.f32 %f20118, 0f00000000; max.ftz.f32 %f20119, %f20117, %f20118; sqrt.approx.ftz.f32 %f20120, %f20119; mul.ftz.f32 %f4370, %f19996, %f20120; setp.lt.ftz.f32 %p2467, %f4370, 0f3F200000; mov.f32 %f30064, %f30065; @%p2467 bra $L__BB107_2709; rcp.approx.ftz.f32 %f20121, %f4370; fma.rn.ftz.f32 %f20122, %f20121, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20123, %f20121, %f20122; fma.rn.ftz.f32 %f20124, %f20121, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20125, %f20121, %f20124, 0f3F800000; div.approx.ftz.f32 %f30064, %f20123, %f20125; $L__BB107_2709: mul.ftz.f32 %f20127, %f4346, %f4346; sub.ftz.f32 %f20128, %f30065, %f20127; div.approx.ftz.f32 %f20129, %f20128, %f20127; max.ftz.f32 %f20131, %f20129, %f20118; sqrt.approx.ftz.f32 %f20132, %f20131; mul.ftz.f32 %f4373, %f19996, %f20132; setp.lt.ftz.f32 %p2468, %f4373, 0f3F200000; @%p2468 bra $L__BB107_2711; rcp.approx.ftz.f32 %f20133, %f4373; fma.rn.ftz.f32 %f20134, %f20133, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20135, %f20133, %f20134; fma.rn.ftz.f32 %f20136, %f20133, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20137, %f20133, %f20136, 0f3F800000; div.approx.ftz.f32 %f30065, %f20135, %f20137; bra.uni $L__BB107_2711; $L__BB107_2330: ld.f32 %f29823, [%rd565+20]; ld.f32 %f29822, [%rd565+16]; ld.f32 %f29821, [%rd565+12]; $L__BB107_2332: setp.lt.ftz.f32 %p2094, %f3244, 0f00000000; @%p2094 bra $L__BB107_2335; bra.uni $L__BB107_2333; $L__BB107_2335: neg.ftz.f32 %f3286, %f3243; setp.leu.ftz.f32 %p2096, %f29829, %f3286; mov.f32 %f29824, %f3241; mov.f32 %f29825, %f3242; mov.f32 %f29826, %f3243; mov.u16 %rs717, %rs519; @%p2096 bra $L__BB107_2337; neg.ftz.f32 %f29827, %f3241; neg.ftz.f32 %f29828, %f3242; mov.f32 %f29824, %f3250; mov.f32 %f29825, %f3251; mov.f32 %f29826, %f3252; mov.f32 %f29829, %f3286; mov.u16 %rs717, %rs534; bra.uni $L__BB107_2337; $L__BB107_2333: setp.geu.ftz.f32 %p2095, %f3243, %f29829; mov.f32 %f29824, %f3241; mov.f32 %f29825, %f3242; mov.f32 %f29826, %f3243; mov.u16 %rs717, %rs519; @%p2095 bra $L__BB107_2337; mov.f32 %f29824, %f29827; mov.f32 %f29825, %f29828; mov.f32 %f29826, %f29829; mov.f32 %f29827, %f3241; mov.f32 %f29828, %f3242; mov.f32 %f29829, %f3243; mov.u16 %rs717, %rs534; $L__BB107_2337: setp.lt.ftz.f32 %p2097, %f29826, 0f3727C5AC; and.pred %p2098, %p2097, %p7; setp.lt.ftz.f32 %p2099, %f29829, 0f3727C5AC; or.pred %p2100, %p2098, %p2099; setp.gt.ftz.f32 %p2101, %f29826, 0fB727C5AC; and.pred %p2103, %p2093, %p2101; or.pred %p2104, %p2103, %p2100; mov.f32 %f30104, 0f00000000; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2104 bra $L__BB107_2774; neg.ftz.f32 %f29915, %f29827; neg.ftz.f32 %f29916, %f29828; neg.ftz.f32 %f29917, %f29829; setp.lt.ftz.f32 %p2105, %f29829, 0fBF7FF972; mov.f32 %f29921, 0f00000000; @%p2105 bra $L__BB107_2341; setp.gt.ftz.f32 %p2106, %f29829, 0f3F7FF972; mov.f32 %f29921, 0fBF7FF972; @%p2106 bra $L__BB107_2341; mul.ftz.f32 %f17656, %f29829, %f29829; max.ftz.f32 %f17658, %f17656, %f17556; rcp.approx.ftz.f32 %f17659, %f17658; mul.ftz.f32 %f17660, %f17470, %f29915; mul.ftz.f32 %f17661, %f17471, %f29916; mul.ftz.f32 %f17662, %f17661, %f17661; fma.rn.ftz.f32 %f17663, %f17660, %f17660, %f17662; fma.rn.ftz.f32 %f17664, %f17663, %f17659, 0f3F800000; sqrt.approx.ftz.f32 %f17665, %f17664; neg.ftz.f32 %f17666, %f17665; setp.ge.ftz.f32 %p2107, %f29829, 0f80000000; selp.f32 %f17667, %f17666, %f17665, %p2107; add.ftz.f32 %f17668, %f17667, 0fBF800000; mul.ftz.f32 %f29921, %f17668, 0f3F000000; $L__BB107_2341: mov.f32 %f29831, %f29824; mov.f32 %f29832, %f29825; mov.f32 %f29833, %f29826; @%p2093 bra $L__BB107_2342; bra.uni $L__BB107_2343; $L__BB107_2342: neg.ftz.f32 %f29833, %f29826; neg.ftz.f32 %f29832, %f29825; neg.ftz.f32 %f29831, %f29824; $L__BB107_2343: setp.gt.ftz.f32 %p2109, %f29833, 0f3F7FF972; mov.f32 %f29838, 0f00000000; mov.f32 %f29834, %f29838; @%p2109 bra $L__BB107_2346; setp.lt.ftz.f32 %p2110, %f29833, 0fBF7FF972; mov.f32 %f29834, 0fBF7FF972; @%p2110 bra $L__BB107_2346; mul.ftz.f32 %f17671, %f29833, %f29833; max.ftz.f32 %f17673, %f17671, %f17556; rcp.approx.ftz.f32 %f17674, %f17673; mul.ftz.f32 %f17675, %f17470, %f29831; mul.ftz.f32 %f17676, %f17471, %f29832; mul.ftz.f32 %f17677, %f17676, %f17676; fma.rn.ftz.f32 %f17678, %f17675, %f17675, %f17677; fma.rn.ftz.f32 %f17679, %f17678, %f17674, 0f3F800000; sqrt.approx.ftz.f32 %f17680, %f17679; neg.ftz.f32 %f17681, %f17680; setp.le.ftz.f32 %p2111, %f29833, 0f00000000; selp.f32 %f17682, %f17681, %f17680, %p2111; add.ftz.f32 %f17683, %f17682, 0fBF800000; mul.ftz.f32 %f29834, %f17683, 0f3F000000; $L__BB107_2346: add.ftz.f32 %f17687, %f29824, %f29827; add.ftz.f32 %f17688, %f29825, %f29828; mul.ftz.f32 %f17689, %f17688, %f17688; fma.rn.ftz.f32 %f17690, %f17687, %f17687, %f17689; add.ftz.f32 %f17691, %f29826, %f29829; fma.rn.ftz.f32 %f17692, %f17691, %f17691, %f17690; sqrt.approx.ftz.f32 %f17693, %f17692; div.approx.ftz.f32 %f3312, %f17687, %f17693; div.approx.ftz.f32 %f3313, %f17688, %f17693; div.approx.ftz.f32 %f3314, %f17691, %f17693; @%p2105 bra $L__BB107_2356; mul.ftz.f32 %f17694, %f29921, %f29917; setp.gt.ftz.f32 %p2114, %f29829, 0f3F7FF972; selp.f32 %f3315, 0f3F800000, %f17694, %p2114; @%p2093 bra $L__BB107_2352; bra.uni $L__BB107_2348; $L__BB107_2352: fma.rn.ftz.f32 %f17729, %f3245, %f29824, %f29827; fma.rn.ftz.f32 %f17730, %f3245, %f29825, %f29828; fma.rn.ftz.f32 %f17731, %f3245, %f29826, %f29829; mul.ftz.f32 %f17732, %f17730, %f17730; fma.rn.ftz.f32 %f17733, %f17729, %f17729, %f17732; fma.rn.ftz.f32 %f17734, %f17731, %f17731, %f17733; sqrt.approx.ftz.f32 %f17735, %f17734; div.approx.ftz.f32 %f17736, %f17729, %f17735; div.approx.ftz.f32 %f17737, %f17730, %f17735; div.approx.ftz.f32 %f17738, %f17731, %f17735; setp.lt.ftz.f32 %p2117, %f17738, 0f00000000; neg.ftz.f32 %f17739, %f17736; neg.ftz.f32 %f17740, %f17737; neg.ftz.f32 %f17741, %f17738; selp.f32 %f17742, %f17739, %f17736, %p2117; selp.f32 %f17743, %f17740, %f17737, %p2117; selp.f32 %f3322, %f17741, %f17738, %p2117; mul.ftz.f32 %f17744, %f29828, %f17743; fma.rn.ftz.f32 %f17745, %f29827, %f17742, %f17744; fma.rn.ftz.f32 %f3323, %f29829, %f3322, %f17745; mul.ftz.f32 %f17746, %f29825, %f17743; fma.rn.ftz.f32 %f17747, %f29824, %f17742, %f17746; fma.rn.ftz.f32 %f3324, %f29826, %f3322, %f17747; setp.lt.ftz.f32 %p2118, %f3323, 0f00000000; @%p2118 bra $L__BB107_2356; fma.rn.ftz.f32 %f3325, %f3245, %f3324, %f3323; abs.ftz.f32 %f3326, %f3323; fma.rn.ftz.f32 %f3327, %f3326, %f3326, %f3262; setp.leu.ftz.f32 %p2119, %f3327, 0f00000000; mov.f32 %f17748, 0f3F800000; mov.f32 %f29836, %f17748; @%p2119 bra $L__BB107_2355; sqrt.approx.ftz.f32 %f17749, %f3327; sub.ftz.f32 %f17750, %f17749, %f3326; add.ftz.f32 %f17751, %f3326, %f17749; div.approx.ftz.f32 %f17752, %f17750, %f17751; fma.rn.ftz.f32 %f17753, %f3326, %f17751, 0fBF800000; fma.rn.ftz.f32 %f17754, %f3326, %f17750, 0f3F800000; div.approx.ftz.f32 %f17755, %f17753, %f17754; mul.ftz.f32 %f17756, %f17752, 0f3F000000; mul.ftz.f32 %f17757, %f17752, %f17756; fma.rn.ftz.f32 %f17758, %f17755, %f17755, 0f3F800000; mul.ftz.f32 %f29836, %f17757, %f17758; $L__BB107_2355: mov.f32 %f17759, 0f00000000; max.ftz.f32 %f17760, %f17759, %f3323; sub.ftz.f32 %f17762, %f17748, %f29836; mul.ftz.f32 %f17763, %f17762, %f17760; neg.ftz.f32 %f17764, %f3324; max.ftz.f32 %f17765, %f17759, %f17764; mul.ftz.f32 %f17766, %f17763, %f17765; mul.ftz.f32 %f17767, %f3322, %f3322; sub.ftz.f32 %f17768, %f17748, %f17767; fma.rn.ftz.f32 %f17769, %f3258, %f17767, %f17768; mul.ftz.f32 %f17770, %f17769, 0f40490FDB; mul.ftz.f32 %f17771, %f17769, %f17770; max.ftz.f32 %f17773, %f17771, %f17556; div.approx.ftz.f32 %f17774, %f3258, %f17773; mul.ftz.f32 %f17775, %f17766, %f17774; mul.ftz.f32 %f17776, %f3315, %f3325; mul.ftz.f32 %f17777, %f3325, %f17776; div.approx.ftz.f32 %f29838, %f17775, %f17777; bra.uni $L__BB107_2356; $L__BB107_2465: setp.lt.s32 %p2225, %r973, 0; min.ftz.f32 %f18440, %f3696, %f3695; max.ftz.f32 %f18441, %f3696, %f3695; div.full.ftz.f32 %f18442, %f18440, %f18441; mul.rn.ftz.f32 %f18443, %f18442, %f18442; mov.f32 %f18444, 0fC0B59883; mov.f32 %f18445, 0fBF52C7EA; fma.rn.ftz.f32 %f18446, %f18443, %f18445, %f18444; mov.f32 %f18447, 0fC0D21907; fma.rn.ftz.f32 %f18448, %f18446, %f18443, %f18447; mul.ftz.f32 %f18449, %f18443, %f18448; mul.ftz.f32 %f18450, %f18442, %f18449; add.ftz.f32 %f18451, %f18443, 0f41355DC0; mov.f32 %f18452, 0f41E6BD60; fma.rn.ftz.f32 %f18453, %f18451, %f18443, %f18452; mov.f32 %f18454, 0f419D92C8; fma.rn.ftz.f32 %f18455, %f18453, %f18443, %f18454; rcp.approx.ftz.f32 %f18456, %f18455; fma.rn.ftz.f32 %f18457, %f18450, %f18456, %f18442; mov.f32 %f18458, 0f3FC90FDB; sub.ftz.f32 %f18459, %f18458, %f18457; setp.gt.ftz.f32 %p2226, %f3696, %f3695; selp.f32 %f18460, %f18459, %f18457, %p2226; mov.f32 %f18461, 0f40490FDB; sub.ftz.f32 %f18462, %f18461, %f18460; selp.f32 %f18463, %f18462, %f18460, %p2225; mov.b32 %r4049, %f18463; or.b32 %r4050, %r974, %r4049; mov.b32 %f18464, %r4050; add.ftz.f32 %f18465, %f3695, %f3696; setp.le.ftz.f32 %p2227, %f18465, 0f7F800000; selp.f32 %f29938, %f18464, %f18465, %p2227; $L__BB107_2468: ld.local.f32 %f3701, [%rd564+44]; ld.f32 %f3702, [%rd566+12]; mul.ftz.f32 %f18467, %f3702, %f3702; sub.ftz.f32 %f18468, %f18436, %f18467; max.ftz.f32 %f18470, %f18468, %f18438; sqrt.approx.ftz.f32 %f18471, %f18470; mul.ftz.f32 %f18472, %f3694, %f3702; mul.ftz.f32 %f18473, %f3701, %f3701; sub.ftz.f32 %f18474, %f18473, %f3693; sqrt.approx.ftz.f32 %f18475, %f18474; div.approx.ftz.f32 %f3703, %f18472, %f18475; mov.f32 %f18476, 0fBF800000; max.ftz.f32 %f18477, %f3703, %f18476; min.ftz.f32 %f18478, %f18477, %f18436; abs.ftz.f32 %f18479, %f18478; sub.ftz.f32 %f18480, %f18436, %f18479; mul.ftz.f32 %f18481, %f18480, 0f3F000000; sqrt.approx.ftz.f32 %f18482, %f18481; setp.gt.ftz.f32 %p2229, %f18479, 0f3F11EB85; selp.f32 %f18483, %f18482, %f18479, %p2229; mul.ftz.f32 %f18484, %f18483, %f18483; mov.f32 %f18485, 0f3C94D2E9; mov.f32 %f18486, 0f3D53F941; fma.rn.ftz.f32 %f18487, %f18486, %f18484, %f18485; mov.f32 %f18488, 0f3D3F841F; fma.rn.ftz.f32 %f18489, %f18487, %f18484, %f18488; mov.f32 %f18490, 0f3D994929; fma.rn.ftz.f32 %f18491, %f18489, %f18484, %f18490; mov.f32 %f18492, 0f3E2AAB94; fma.rn.ftz.f32 %f18493, %f18491, %f18484, %f18492; mul.ftz.f32 %f18494, %f18484, %f18493; fma.rn.ftz.f32 %f18495, %f18494, %f18483, %f18483; mov.f32 %f18496, 0f3FC90FDB; mov.f32 %f18497, 0fC0000000; fma.rn.ftz.f32 %f18498, %f18497, %f18495, %f18496; selp.f32 %f18499, %f18498, %f18495, %p2229; setp.le.ftz.f32 %p2230, %f18499, 0f7F800000; mov.b32 %r4056, %f18499; mov.b32 %r4057, %f18478; and.b32 %r4058, %r4057, -2147483648; or.b32 %r4059, %r4058, %r4056; mov.b32 %f18500, %r4059; selp.f32 %f3704, %f18500, %f18499, %p2230; ld.local.v4.f32 {%f18501, %f18502, %f18503, %f18504}, [%rd564+20]; mul.ftz.f32 %f18505, %f3694, %f18471; abs.ftz.f32 %f3709, %f18505; add.ftz.f32 %f18506, %f18473, 0fBF800000; fma.rn.ftz.f32 %f3710, %f3709, %f3709, %f18506; setp.leu.ftz.f32 %p2231, %f3710, 0f00000000; mov.f32 %f29939, %f18436; @%p2231 bra $L__BB107_2470; sqrt.approx.ftz.f32 %f18507, %f3710; sub.ftz.f32 %f18508, %f18507, %f3709; add.ftz.f32 %f18509, %f3709, %f18507; div.approx.ftz.f32 %f18510, %f18508, %f18509; fma.rn.ftz.f32 %f18511, %f3709, %f18509, 0fBF800000; fma.rn.ftz.f32 %f18512, %f3709, %f18508, 0f3F800000; div.approx.ftz.f32 %f18513, %f18511, %f18512; mul.ftz.f32 %f18514, %f18510, 0f3F000000; mul.ftz.f32 %f18515, %f18510, %f18514; fma.rn.ftz.f32 %f18516, %f18513, %f18513, 0f3F800000; mul.ftz.f32 %f29939, %f18515, %f18516; $L__BB107_2470: div.approx.ftz.f32 %f18518, %f3689, %f3701; mul.ftz.f32 %f18519, %f3703, %f3703; sub.ftz.f32 %f18521, %f18436, %f18519; max.ftz.f32 %f18522, %f18521, %f18438; sqrt.approx.ftz.f32 %f18523, %f18522; add.ftz.f32 %f18524, %f18523, %f18523; mul.ftz.f32 %f18525, %f18518, %f18518; sub.ftz.f32 %f18526, %f18436, %f18525; max.ftz.f32 %f18527, %f18526, %f18438; sqrt.approx.ftz.f32 %f18528, %f18527; div.approx.ftz.f32 %f18529, %f18524, %f18528; mul.ftz.f32 %f18530, %f18501, %f18529; mul.ftz.f32 %f18531, %f18530, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18532, %f18531; mul.ftz.f32 %f18533, %f18502, %f18529; mul.ftz.f32 %f18534, %f18533, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18535, %f18534; mul.ftz.f32 %f18536, %f18503, %f18529; mul.ftz.f32 %f18537, %f18536, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18538, %f18537; sub.ftz.f32 %f18539, %f18436, %f29939; mul.ftz.f32 %f18540, %f18539, %f18539; mul.ftz.f32 %f3713, %f18532, %f18540; mul.ftz.f32 %f3714, %f18535, %f18540; mul.ftz.f32 %f3715, %f18538, %f18540; mul.ftz.f32 %f18541, %f3714, %f17459; fma.rn.ftz.f32 %f18542, %f3713, %f17458, %f18541; fma.rn.ftz.f32 %f3716, %f3715, %f17460, %f18542; mul.ftz.f32 %f3717, %f18532, %f29939; mul.ftz.f32 %f3718, %f3717, %f3713; mul.ftz.f32 %f3719, %f18535, %f29939; mul.ftz.f32 %f3720, %f3719, %f3714; mul.ftz.f32 %f3721, %f18538, %f29939; mul.ftz.f32 %f3722, %f3721, %f3715; mul.ftz.f32 %f18543, %f3720, %f17459; fma.rn.ftz.f32 %f18544, %f3718, %f17458, %f18543; fma.rn.ftz.f32 %f3723, %f3722, %f17460, %f18544; sub.ftz.f32 %f3724, %f18436, %f3717; sub.ftz.f32 %f3725, %f18436, %f3719; sub.ftz.f32 %f3726, %f18436, %f3721; setp.eq.ftz.f32 %p2232, %f3724, 0f00000000; mov.f32 %f29940, %f18438; @%p2232 bra $L__BB107_2472; div.approx.ftz.f32 %f29940, %f3717, %f3724; $L__BB107_2472: setp.eq.ftz.f32 %p2233, %f3725, 0f00000000; mov.f32 %f29941, %f18438; @%p2233 bra $L__BB107_2474; div.approx.ftz.f32 %f29941, %f3719, %f3725; $L__BB107_2474: setp.eq.ftz.f32 %p2234, %f3726, 0f00000000; mov.f32 %f29942, %f18438; @%p2234 bra $L__BB107_2476; div.approx.ftz.f32 %f29942, %f3721, %f3726; $L__BB107_2476: mul.ftz.f32 %f3733, %f3718, %f29940; mul.ftz.f32 %f3734, %f3720, %f29941; mul.ftz.f32 %f18548, %f3734, %f17459; fma.rn.ftz.f32 %f18549, %f3733, %f17458, %f18548; mul.ftz.f32 %f3735, %f3722, %f29942; fma.rn.ftz.f32 %f3736, %f3735, %f17460, %f18549; add.ftz.f32 %f18550, %f29939, %f3716; add.ftz.f32 %f18551, %f18550, %f3723; add.ftz.f32 %f3737, %f3736, %f18551; setp.eq.ftz.f32 %p2235, %f3737, 0f00000000; mov.f32 %f29943, %f18438; @%p2235 bra $L__BB107_2478; mov.f32 %f18552, 0f3F800000; div.approx.ftz.f32 %f29943, %f18552, %f3737; $L__BB107_2478: mul.ftz.f32 %f18553, %f3690, %f3690; mov.f32 %f18554, 0f3F800000; sub.ftz.f32 %f18555, %f18554, %f18553; max.ftz.f32 %f18557, %f18555, %f18438; sqrt.approx.ftz.f32 %f3740, %f18557; abs.ftz.f32 %f3741, %f3691; setp.eq.ftz.f32 %p2236, %f3741, 0f00000000; abs.ftz.f32 %f3742, %f3692; setp.eq.ftz.f32 %p2237, %f3742, 0f00000000; and.pred %p2238, %p2236, %p2237; mov.b32 %r975, %f3691; mov.b32 %r4060, %f3692; and.b32 %r976, %r4060, -2147483648; @%p2238 bra $L__BB107_2482; bra.uni $L__BB107_2479; $L__BB107_2482: shr.s32 %r4065, %r975, 31; and.b32 %r4066, %r4065, 1078530011; or.b32 %r4067, %r976, %r4066; mov.b32 %f29944, %r4067; bra.uni $L__BB107_2483; $L__BB107_2479: setp.eq.ftz.f32 %p2239, %f3741, 0f7F800000; setp.eq.ftz.f32 %p2240, %f3742, 0f7F800000; and.pred %p2241, %p2239, %p2240; @%p2241 bra $L__BB107_2481; bra.uni $L__BB107_2480; $L__BB107_2481: setp.lt.s32 %p2245, %r975, 0; selp.b32 %r4063, 1075235812, 1061752795, %p2245; or.b32 %r4064, %r976, %r4063; mov.b32 %f29944, %r4064; bra.uni $L__BB107_2483; $L__BB107_2480: setp.lt.s32 %p2242, %r975, 0; min.ftz.f32 %f18558, %f3742, %f3741; max.ftz.f32 %f18559, %f3742, %f3741; div.full.ftz.f32 %f18560, %f18558, %f18559; mul.rn.ftz.f32 %f18561, %f18560, %f18560; mov.f32 %f18562, 0fC0B59883; mov.f32 %f18563, 0fBF52C7EA; fma.rn.ftz.f32 %f18564, %f18561, %f18563, %f18562; mov.f32 %f18565, 0fC0D21907; fma.rn.ftz.f32 %f18566, %f18564, %f18561, %f18565; mul.ftz.f32 %f18567, %f18561, %f18566; mul.ftz.f32 %f18568, %f18560, %f18567; add.ftz.f32 %f18569, %f18561, 0f41355DC0; mov.f32 %f18570, 0f41E6BD60; fma.rn.ftz.f32 %f18571, %f18569, %f18561, %f18570; mov.f32 %f18572, 0f419D92C8; fma.rn.ftz.f32 %f18573, %f18571, %f18561, %f18572; rcp.approx.ftz.f32 %f18574, %f18573; fma.rn.ftz.f32 %f18575, %f18568, %f18574, %f18560; sub.ftz.f32 %f18577, %f18496, %f18575; setp.gt.ftz.f32 %p2243, %f3742, %f3741; selp.f32 %f18578, %f18577, %f18575, %p2243; mov.f32 %f18579, 0f40490FDB; sub.ftz.f32 %f18580, %f18579, %f18578; selp.f32 %f18581, %f18580, %f18578, %p2242; mov.b32 %r4061, %f18581; or.b32 %r4062, %r976, %r4061; mov.b32 %f18582, %r4062; add.ftz.f32 %f18583, %f3741, %f3742; setp.le.ftz.f32 %p2244, %f18583, 0f7F800000; selp.f32 %f29944, %f18582, %f18583, %p2244; $L__BB107_2483: max.ftz.f32 %f18585, %f3702, %f18476; min.ftz.f32 %f18587, %f18585, %f18554; abs.ftz.f32 %f18588, %f18587; setp.gt.ftz.f32 %p2246, %f18588, 0f3F11EB85; sub.ftz.f32 %f18589, %f18554, %f18588; mul.ftz.f32 %f18590, %f18589, 0f3F000000; sqrt.approx.ftz.f32 %f18591, %f18590; selp.f32 %f18592, %f18591, %f18588, %p2246; mul.ftz.f32 %f18593, %f18592, %f18592; fma.rn.ftz.f32 %f18596, %f18486, %f18593, %f18485; fma.rn.ftz.f32 %f18598, %f18596, %f18593, %f18488; fma.rn.ftz.f32 %f18600, %f18598, %f18593, %f18490; fma.rn.ftz.f32 %f18602, %f18600, %f18593, %f18492; mul.ftz.f32 %f18603, %f18593, %f18602; fma.rn.ftz.f32 %f18604, %f18603, %f18592, %f18592; fma.rn.ftz.f32 %f18607, %f18497, %f18604, %f18496; selp.f32 %f18608, %f18607, %f18604, %p2246; sub.ftz.f32 %f3747, %f29944, %f29938; ld.local.v2.f32 {%f18609, %f18610}, [%rd564+36]; sin.approx.ftz.f32 %f18612, %f18610; mul.ftz.f32 %f18613, %f18612, %f18612; sub.ftz.f32 %f18614, %f18554, %f18613; mov.f32 %f29951, 0f00000000; max.ftz.f32 %f18616, %f18614, %f29951; sqrt.approx.ftz.f32 %f18617, %f18616; add.ftz.f32 %f18618, %f18612, %f18612; mul.ftz.f32 %f18619, %f18618, %f18617; mul.ftz.f32 %f18620, %f18617, %f18617; sub.ftz.f32 %f18621, %f18620, %f18613; add.ftz.f32 %f18622, %f18619, %f18619; mul.ftz.f32 %f18623, %f18622, %f18621; mul.ftz.f32 %f18624, %f18621, %f18621; mul.ftz.f32 %f18625, %f18619, %f18619; sub.ftz.f32 %f18626, %f18624, %f18625; mul.ftz.f32 %f18627, %f3690, %f18621; mul.ftz.f32 %f18628, %f3740, %f18621; mul.ftz.f32 %f18629, %f3690, %f18619; sub.ftz.f32 %f18630, %f18628, %f18629; fma.rn.ftz.f32 %f18631, %f3740, %f18619, %f18627; abs.ftz.f32 %f18632, %f18630; st.local.v2.f32 [%rd2157], {%f18631, %f18632}; mul.ftz.f32 %f18633, %f3690, %f18617; mul.ftz.f32 %f18634, %f3740, %f18612; sub.ftz.f32 %f3749, %f18633, %f18634; mul.ftz.f32 %f18635, %f3740, %f18617; fma.rn.ftz.f32 %f18636, %f3690, %f18612, %f18635; abs.ftz.f32 %f3750, %f18636; st.local.v2.f32 [%rd277], {%f3749, %f3750}; mul.ftz.f32 %f18637, %f3690, %f18626; mul.ftz.f32 %f18638, %f3740, %f18623; sub.ftz.f32 %f3751, %f18637, %f18638; mul.ftz.f32 %f18639, %f3740, %f18626; fma.rn.ftz.f32 %f18640, %f3690, %f18623, %f18639; abs.ftz.f32 %f3752, %f18640; st.local.v2.f32 [%rd277+8], {%f3751, %f3752}; setp.le.ftz.f32 %p2247, %f18608, 0f7F800000; mov.b32 %r4068, %f18608; mov.b32 %r4069, %f18587; and.b32 %r4070, %r4069, -2147483648; or.b32 %r4071, %r4070, %r4068; mov.b32 %f18641, %r4071; selp.f32 %f18642, %f18641, %f18608, %p2247; add.ftz.f32 %f3753, %f18642, %f18642; mov.f32 %f18643, 0f40490FDB; div.approx.ftz.f32 %f3754, %f18643, %f18609; mul.ftz.f32 %f3755, %f3754, 0f3FB8AA3B; ld.local.f32 %f3756, [%rd564+48]; rcp.approx.ftz.f32 %f3757, %f3756; mul.ftz.f32 %f18644, %f3694, %f18632; mul.ftz.f32 %f3758, %f18644, %f3757; mul.ftz.f32 %f18645, %f3689, %f18631; mul.ftz.f32 %f3759, %f18645, %f3757; setp.gtu.ftz.f32 %p2248, %f3756, 0f3DCCCCCD; @%p2248 bra $L__BB107_2497; bra.uni $L__BB107_2484; $L__BB107_2497: mul.ftz.f32 %f3782, %f3758, %f3758; fma.rn.ftz.f32 %f29947, %f3782, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3784, %f3782, %f3782; mov.u64 %rd2867, 64; cvt.rn.f32.u64 %f18676, %rd2867; div.approx.ftz.f32 %f18677, %f3784, %f18676; add.ftz.f32 %f3785, %f29947, %f18677; setp.eq.ftz.f32 %p2258, %f29947, %f3785; @%p2258 bra $L__BB107_2506; mul.ftz.f32 %f3786, %f3782, %f3784; mov.u64 %rd2868, 2304; cvt.rn.f32.u64 %f18678, %rd2868; div.approx.ftz.f32 %f18679, %f3786, %f18678; add.ftz.f32 %f3787, %f3785, %f18679; setp.eq.ftz.f32 %p2259, %f3785, %f3787; mov.f32 %f29947, %f3785; @%p2259 bra $L__BB107_2506; mul.ftz.f32 %f3788, %f3782, %f3786; mov.u64 %rd2869, 147456; cvt.rn.f32.u64 %f18680, %rd2869; div.approx.ftz.f32 %f18681, %f3788, %f18680; add.ftz.f32 %f3789, %f3787, %f18681; setp.eq.ftz.f32 %p2260, %f3787, %f3789; mov.f32 %f29947, %f3787; @%p2260 bra $L__BB107_2506; mul.ftz.f32 %f3790, %f3782, %f3788; mov.u64 %rd2870, 14745600; cvt.rn.f32.u64 %f18682, %rd2870; div.approx.ftz.f32 %f18683, %f3790, %f18682; add.ftz.f32 %f3791, %f3789, %f18683; setp.eq.ftz.f32 %p2261, %f3789, %f3791; mov.f32 %f29947, %f3789; @%p2261 bra $L__BB107_2506; mul.ftz.f32 %f3792, %f3782, %f3790; mov.u64 %rd2871, 2123366400; cvt.rn.f32.u64 %f18684, %rd2871; div.approx.ftz.f32 %f18685, %f3792, %f18684; add.ftz.f32 %f3793, %f3791, %f18685; setp.eq.ftz.f32 %p2262, %f3791, %f3793; mov.f32 %f29947, %f3791; @%p2262 bra $L__BB107_2506; mul.ftz.f32 %f3794, %f3782, %f3792; mov.u64 %rd2872, 416179814400; cvt.rn.f32.u64 %f18686, %rd2872; div.approx.ftz.f32 %f18687, %f3794, %f18686; add.ftz.f32 %f3795, %f3793, %f18687; setp.eq.ftz.f32 %p2263, %f3793, %f3795; mov.f32 %f29947, %f3793; @%p2263 bra $L__BB107_2506; mul.ftz.f32 %f3796, %f3782, %f3794; mov.u64 %rd2873, 106542032486400; cvt.rn.f32.u64 %f18688, %rd2873; div.approx.ftz.f32 %f18689, %f3796, %f18688; add.ftz.f32 %f3797, %f3795, %f18689; setp.eq.ftz.f32 %p2264, %f3795, %f3797; mov.f32 %f29947, %f3795; @%p2264 bra $L__BB107_2506; mul.ftz.f32 %f18690, %f3782, %f3796; mov.u64 %rd2874, 34519618525593600; cvt.rn.f32.u64 %f18691, %rd2874; div.approx.ftz.f32 %f18692, %f18690, %f18691; add.ftz.f32 %f3798, %f3797, %f18692; setp.eq.ftz.f32 %p2265, %f3797, %f3798; mov.f32 %f29947, %f3797; @%p2265 bra $L__BB107_2506; mov.f32 %f29947, %f3798; $L__BB107_2506: mul.ftz.f32 %f18693, %f3759, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18694, %f18693; mul.ftz.f32 %f3800, %f29947, %f18694; abs.ftz.f32 %f3801, %f3757; setp.ltu.ftz.f32 %p2266, %f3801, 0f3F800000; @%p2266 bra $L__BB107_2508; bra.uni $L__BB107_2507; $L__BB107_2508: mul.ftz.f32 %f18716, %f3757, %f3757; mov.f32 %f18717, 0f394FFF49; mov.f32 %f18718, 0f363D0ADA; fma.rn.ftz.f32 %f18719, %f18718, %f18716, %f18717; mov.f32 %f18720, 0f3C08889A; fma.rn.ftz.f32 %f18721, %f18719, %f18716, %f18720; mov.f32 %f18722, 0f3E2AAAAB; fma.rn.ftz.f32 %f18723, %f18721, %f18716, %f18722; mul.ftz.f32 %f18724, %f18716, %f18723; fma.rn.ftz.f32 %f29948, %f18724, %f3757, %f3757; bra.uni $L__BB107_2509; $L__BB107_2484: setp.gt.ftz.f32 %p2249, %f3758, 0f41400000; @%p2249 bra $L__BB107_2495; bra.uni $L__BB107_2485; $L__BB107_2495: mul.ftz.f32 %f18664, %f3758, 0f41000000; rcp.approx.ftz.f32 %f18665, %f18664; add.ftz.f32 %f18666, %f18665, 0fBFEB3F8E; lg2.approx.ftz.f32 %f18667, %f3758; fma.rn.ftz.f32 %f18668, %f18667, 0fBF317218, %f18666; fma.rn.ftz.f32 %f29946, %f18668, 0f3F000000, %f3758; bra.uni $L__BB107_2496; $L__BB107_2507: mov.f32 %f18695, 0f3FB8AA3B; mul.rn.ftz.f32 %f18696, %f3801, %f18695; cvt.rzi.ftz.f32.f32 %f18697, %f18696; abs.ftz.f32 %f18698, %f18697; setp.gt.ftz.f32 %p2267, %f18698, 0f42FC0000; mov.b32 %r4072, %f18697; and.b32 %r4073, %r4072, -2147483648; or.b32 %r4074, %r4073, 1123811328; mov.b32 %f18699, %r4074; selp.f32 %f18700, %f18699, %f18697, %p2267; mov.f32 %f18701, 0fBF317218; fma.rn.ftz.f32 %f18702, %f18700, %f18701, %f3801; mov.f32 %f18703, 0f3102E308; fma.rn.ftz.f32 %f18704, %f18700, %f18703, %f18702; mul.ftz.f32 %f18705, %f18704, 0f3FB8AA3B; add.ftz.f32 %f18706, %f18700, 0f4B40007D; mov.b32 %r4075, %f18706; shl.b32 %r4076, %r4075, 23; mov.b32 %f18707, %r4076; ex2.approx.ftz.f32 %f18708, %f18705; mul.ftz.f32 %f18709, %f18708, %f18707; mov.f32 %f18710, 0f3E000000; div.approx.ftz.f32 %f18711, %f18710, %f18709; neg.ftz.f32 %f18712, %f18711; mov.f32 %f18713, 0f40000000; fma.rn.ftz.f32 %f18714, %f18713, %f18709, %f18712; setp.ge.ftz.f32 %p2268, %f3801, 0f42B40000; selp.f32 %f18715, 0f7F800000, %f18714, %p2268; mov.b32 %r4077, %f18715; mov.b32 %r4078, %f3757; and.b32 %r4079, %r4078, -2147483648; or.b32 %r4080, %r4079, %r4077; mov.b32 %f29948, %r4080; $L__BB107_2509: add.ftz.f32 %f18725, %f29948, %f29948; mul.ftz.f32 %f18726, %f3756, %f18725; div.approx.ftz.f32 %f29949, %f3800, %f18726; bra.uni $L__BB107_2510; $L__BB107_2485: mul.ftz.f32 %f3760, %f3758, %f3758; fma.rn.ftz.f32 %f29945, %f3760, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3762, %f3760, %f3760; mov.u64 %rd2859, 64; cvt.rn.f32.u64 %f18646, %rd2859; div.approx.ftz.f32 %f18647, %f3762, %f18646; add.ftz.f32 %f3763, %f29945, %f18647; setp.eq.ftz.f32 %p2250, %f29945, %f3763; @%p2250 bra $L__BB107_2494; mul.ftz.f32 %f3764, %f3760, %f3762; mov.u64 %rd2860, 2304; cvt.rn.f32.u64 %f18648, %rd2860; div.approx.ftz.f32 %f18649, %f3764, %f18648; add.ftz.f32 %f3765, %f3763, %f18649; setp.eq.ftz.f32 %p2251, %f3763, %f3765; mov.f32 %f29945, %f3763; @%p2251 bra $L__BB107_2494; mul.ftz.f32 %f3766, %f3760, %f3764; mov.u64 %rd2861, 147456; cvt.rn.f32.u64 %f18650, %rd2861; div.approx.ftz.f32 %f18651, %f3766, %f18650; add.ftz.f32 %f3767, %f3765, %f18651; setp.eq.ftz.f32 %p2252, %f3765, %f3767; mov.f32 %f29945, %f3765; @%p2252 bra $L__BB107_2494; mul.ftz.f32 %f3768, %f3760, %f3766; mov.u64 %rd2862, 14745600; cvt.rn.f32.u64 %f18652, %rd2862; div.approx.ftz.f32 %f18653, %f3768, %f18652; add.ftz.f32 %f3769, %f3767, %f18653; setp.eq.ftz.f32 %p2253, %f3767, %f3769; mov.f32 %f29945, %f3767; @%p2253 bra $L__BB107_2494; mul.ftz.f32 %f3770, %f3760, %f3768; mov.u64 %rd2863, 2123366400; cvt.rn.f32.u64 %f18654, %rd2863; div.approx.ftz.f32 %f18655, %f3770, %f18654; add.ftz.f32 %f3771, %f3769, %f18655; setp.eq.ftz.f32 %p2254, %f3769, %f3771; mov.f32 %f29945, %f3769; @%p2254 bra $L__BB107_2494; mul.ftz.f32 %f3772, %f3760, %f3770; mov.u64 %rd2864, 416179814400; cvt.rn.f32.u64 %f18656, %rd2864; div.approx.ftz.f32 %f18657, %f3772, %f18656; add.ftz.f32 %f3773, %f3771, %f18657; setp.eq.ftz.f32 %p2255, %f3771, %f3773; mov.f32 %f29945, %f3771; @%p2255 bra $L__BB107_2494; mul.ftz.f32 %f3774, %f3760, %f3772; mov.u64 %rd2865, 106542032486400; cvt.rn.f32.u64 %f18658, %rd2865; div.approx.ftz.f32 %f18659, %f3774, %f18658; add.ftz.f32 %f3775, %f3773, %f18659; setp.eq.ftz.f32 %p2256, %f3773, %f3775; mov.f32 %f29945, %f3773; @%p2256 bra $L__BB107_2494; mul.ftz.f32 %f18660, %f3760, %f3774; mov.u64 %rd2866, 34519618525593600; cvt.rn.f32.u64 %f18661, %rd2866; div.approx.ftz.f32 %f18662, %f18660, %f18661; add.ftz.f32 %f3776, %f3775, %f18662; setp.eq.ftz.f32 %p2257, %f3775, %f3776; mov.f32 %f29945, %f3775; @%p2257 bra $L__BB107_2494; mov.f32 %f29945, %f3776; $L__BB107_2494: lg2.approx.ftz.f32 %f18663, %f29945; mul.ftz.f32 %f29946, %f18663, 0f3F317218; $L__BB107_2496: sub.ftz.f32 %f18669, %f29946, %f3759; sub.ftz.f32 %f18670, %f18669, %f3757; add.ftz.f32 %f18671, %f18670, 0f3F316F00; mul.ftz.f32 %f18672, %f3757, 0f3F000000; lg2.approx.ftz.f32 %f18673, %f18672; fma.rn.ftz.f32 %f18674, %f18673, 0f3F317218, %f18671; mul.ftz.f32 %f18675, %f18674, 0f3FB8AA3B; ex2.approx.ftz.f32 %f29949, %f18675; $L__BB107_2510: setp.gt.ftz.f32 %p2269, %f3754, 0f42B00000; mul.ftz.f32 %f3807, %f29943, %f3716; mul.ftz.f32 %f3808, %f29943, %f3723; mul.ftz.f32 %f3809, %f29943, %f3736; mov.f32 %f29950, %f29951; @%p2269 bra $L__BB107_2512; ex2.approx.ftz.f32 %f18728, %f3755; add.ftz.f32 %f18729, %f18728, 0f3F800000; rcp.approx.ftz.f32 %f18730, %f18729; add.ftz.f32 %f29950, %f18730, %f18730; $L__BB107_2512: sub.ftz.f32 %f3812, %f18554, %f29950; setp.eq.ftz.f32 %p2270, %f3812, 0f00000000; @%p2270 bra $L__BB107_2514; mul.ftz.f32 %f18733, %f3704, 0f00000000; sub.ftz.f32 %f18734, %f18733, %f3753; add.ftz.f32 %f18735, %f18734, 0f00000000; sub.ftz.f32 %f18736, %f3747, %f18735; add.ftz.f32 %f18737, %f18736, 0f40490FDB; mov.f32 %f18738, 0f40C90FDB; div.approx.ftz.f32 %f18739, %f18737, %f18738; cvt.rmi.ftz.f32.f32 %f18740, %f18739; mul.ftz.f32 %f18741, %f18740, 0f40C90FDB; sub.ftz.f32 %f18742, %f18737, %f18741; add.ftz.f32 %f18743, %f18742, 0fC0490FDB; abs.ftz.f32 %f18744, %f18743; neg.ftz.f32 %f18745, %f18744; div.approx.ftz.f32 %f18746, %f18745, %f18609; mul.ftz.f32 %f18747, %f18746, 0f3FB8AA3B; ex2.approx.ftz.f32 %f18748, %f18747; add.ftz.f32 %f18749, %f18748, 0f3F800000; mul.ftz.f32 %f18750, %f18749, %f18749; mul.ftz.f32 %f18751, %f18609, %f18750; div.approx.ftz.f32 %f18752, %f18748, %f18751; div.approx.ftz.f32 %f29951, %f18752, %f3812; $L__BB107_2514: mul.ftz.f32 %f18753, %f29949, %f29939; fma.rn.ftz.f32 %f3815, %f29951, %f18753, 0f00000000; mul.ftz.f32 %f18754, %f29943, %f29939; mul.ftz.f32 %f18755, %f29949, %f18754; fma.rn.ftz.f32 %f3816, %f29951, %f18755, 0f00000000; ld.local.f32 %f3817, [%rd564+32]; mul.ftz.f32 %f3818, %f3817, 0f3E800000; rcp.approx.ftz.f32 %f3819, %f3818; mul.ftz.f32 %f18756, %f3694, %f3750; mul.ftz.f32 %f3820, %f18756, %f3819; mul.ftz.f32 %f18757, %f3689, %f3749; mul.ftz.f32 %f3821, %f18757, %f3819; setp.gtu.ftz.f32 %p2271, %f3818, 0f3DCCCCCD; @%p2271 bra $L__BB107_2528; bra.uni $L__BB107_2515; $L__BB107_2528: mul.ftz.f32 %f3844, %f3820, %f3820; fma.rn.ftz.f32 %f29954, %f3844, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3846, %f3844, %f3844; mov.u64 %rd2883, 64; cvt.rn.f32.u64 %f18788, %rd2883; div.approx.ftz.f32 %f18789, %f3846, %f18788; add.ftz.f32 %f3847, %f29954, %f18789; setp.eq.ftz.f32 %p2281, %f29954, %f3847; @%p2281 bra $L__BB107_2537; mul.ftz.f32 %f3848, %f3844, %f3846; mov.u64 %rd2884, 2304; cvt.rn.f32.u64 %f18790, %rd2884; div.approx.ftz.f32 %f18791, %f3848, %f18790; add.ftz.f32 %f3849, %f3847, %f18791; setp.eq.ftz.f32 %p2282, %f3847, %f3849; mov.f32 %f29954, %f3847; @%p2282 bra $L__BB107_2537; mul.ftz.f32 %f3850, %f3844, %f3848; mov.u64 %rd2885, 147456; cvt.rn.f32.u64 %f18792, %rd2885; div.approx.ftz.f32 %f18793, %f3850, %f18792; add.ftz.f32 %f3851, %f3849, %f18793; setp.eq.ftz.f32 %p2283, %f3849, %f3851; mov.f32 %f29954, %f3849; @%p2283 bra $L__BB107_2537; mul.ftz.f32 %f3852, %f3844, %f3850; mov.u64 %rd2886, 14745600; cvt.rn.f32.u64 %f18794, %rd2886; div.approx.ftz.f32 %f18795, %f3852, %f18794; add.ftz.f32 %f3853, %f3851, %f18795; setp.eq.ftz.f32 %p2284, %f3851, %f3853; mov.f32 %f29954, %f3851; @%p2284 bra $L__BB107_2537; mul.ftz.f32 %f3854, %f3844, %f3852; mov.u64 %rd2887, 2123366400; cvt.rn.f32.u64 %f18796, %rd2887; div.approx.ftz.f32 %f18797, %f3854, %f18796; add.ftz.f32 %f3855, %f3853, %f18797; setp.eq.ftz.f32 %p2285, %f3853, %f3855; mov.f32 %f29954, %f3853; @%p2285 bra $L__BB107_2537; mul.ftz.f32 %f3856, %f3844, %f3854; mov.u64 %rd2888, 416179814400; cvt.rn.f32.u64 %f18798, %rd2888; div.approx.ftz.f32 %f18799, %f3856, %f18798; add.ftz.f32 %f3857, %f3855, %f18799; setp.eq.ftz.f32 %p2286, %f3855, %f3857; mov.f32 %f29954, %f3855; @%p2286 bra $L__BB107_2537; mul.ftz.f32 %f3858, %f3844, %f3856; mov.u64 %rd2889, 106542032486400; cvt.rn.f32.u64 %f18800, %rd2889; div.approx.ftz.f32 %f18801, %f3858, %f18800; add.ftz.f32 %f3859, %f3857, %f18801; setp.eq.ftz.f32 %p2287, %f3857, %f3859; mov.f32 %f29954, %f3857; @%p2287 bra $L__BB107_2537; mul.ftz.f32 %f18802, %f3844, %f3858; mov.u64 %rd2890, 34519618525593600; cvt.rn.f32.u64 %f18803, %rd2890; div.approx.ftz.f32 %f18804, %f18802, %f18803; add.ftz.f32 %f3860, %f3859, %f18804; setp.eq.ftz.f32 %p2288, %f3859, %f3860; mov.f32 %f29954, %f3859; @%p2288 bra $L__BB107_2537; mov.f32 %f29954, %f3860; $L__BB107_2537: mul.ftz.f32 %f18805, %f3821, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18806, %f18805; mul.ftz.f32 %f3862, %f29954, %f18806; abs.ftz.f32 %f3863, %f3819; setp.ltu.ftz.f32 %p2289, %f3863, 0f3F800000; @%p2289 bra $L__BB107_2539; bra.uni $L__BB107_2538; $L__BB107_2539: mul.ftz.f32 %f18828, %f3819, %f3819; mov.f32 %f18829, 0f394FFF49; mov.f32 %f18830, 0f363D0ADA; fma.rn.ftz.f32 %f18831, %f18830, %f18828, %f18829; mov.f32 %f18832, 0f3C08889A; fma.rn.ftz.f32 %f18833, %f18831, %f18828, %f18832; mov.f32 %f18834, 0f3E2AAAAB; fma.rn.ftz.f32 %f18835, %f18833, %f18828, %f18834; mul.ftz.f32 %f18836, %f18828, %f18835; fma.rn.ftz.f32 %f29955, %f18836, %f3819, %f3819; bra.uni $L__BB107_2540; $L__BB107_2515: setp.gt.ftz.f32 %p2272, %f3820, 0f41400000; @%p2272 bra $L__BB107_2526; bra.uni $L__BB107_2516; $L__BB107_2526: mul.ftz.f32 %f18776, %f3820, 0f41000000; rcp.approx.ftz.f32 %f18777, %f18776; add.ftz.f32 %f18778, %f18777, 0fBFEB3F8E; lg2.approx.ftz.f32 %f18779, %f3820; fma.rn.ftz.f32 %f18780, %f18779, 0fBF317218, %f18778; fma.rn.ftz.f32 %f29953, %f18780, 0f3F000000, %f3820; bra.uni $L__BB107_2527; $L__BB107_2538: mov.f32 %f18807, 0f3FB8AA3B; mul.rn.ftz.f32 %f18808, %f3863, %f18807; cvt.rzi.ftz.f32.f32 %f18809, %f18808; abs.ftz.f32 %f18810, %f18809; setp.gt.ftz.f32 %p2290, %f18810, 0f42FC0000; mov.b32 %r4081, %f18809; and.b32 %r4082, %r4081, -2147483648; or.b32 %r4083, %r4082, 1123811328; mov.b32 %f18811, %r4083; selp.f32 %f18812, %f18811, %f18809, %p2290; mov.f32 %f18813, 0fBF317218; fma.rn.ftz.f32 %f18814, %f18812, %f18813, %f3863; mov.f32 %f18815, 0f3102E308; fma.rn.ftz.f32 %f18816, %f18812, %f18815, %f18814; mul.ftz.f32 %f18817, %f18816, 0f3FB8AA3B; add.ftz.f32 %f18818, %f18812, 0f4B40007D; mov.b32 %r4084, %f18818; shl.b32 %r4085, %r4084, 23; mov.b32 %f18819, %r4085; ex2.approx.ftz.f32 %f18820, %f18817; mul.ftz.f32 %f18821, %f18820, %f18819; mov.f32 %f18822, 0f3E000000; div.approx.ftz.f32 %f18823, %f18822, %f18821; neg.ftz.f32 %f18824, %f18823; mov.f32 %f18825, 0f40000000; fma.rn.ftz.f32 %f18826, %f18825, %f18821, %f18824; setp.ge.ftz.f32 %p2291, %f3863, 0f42B40000; selp.f32 %f18827, 0f7F800000, %f18826, %p2291; mov.b32 %r4086, %f18827; mov.b32 %r4087, %f3819; and.b32 %r4088, %r4087, -2147483648; or.b32 %r4089, %r4088, %r4086; mov.b32 %f29955, %r4089; $L__BB107_2540: add.ftz.f32 %f18837, %f29955, %f29955; mul.ftz.f32 %f18838, %f3818, %f18837; div.approx.ftz.f32 %f29956, %f3862, %f18838; bra.uni $L__BB107_2541; $L__BB107_2516: mul.ftz.f32 %f3822, %f3820, %f3820; fma.rn.ftz.f32 %f29952, %f3822, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3824, %f3822, %f3822; mov.u64 %rd2875, 64; cvt.rn.f32.u64 %f18758, %rd2875; div.approx.ftz.f32 %f18759, %f3824, %f18758; add.ftz.f32 %f3825, %f29952, %f18759; setp.eq.ftz.f32 %p2273, %f29952, %f3825; @%p2273 bra $L__BB107_2525; mul.ftz.f32 %f3826, %f3822, %f3824; mov.u64 %rd2876, 2304; cvt.rn.f32.u64 %f18760, %rd2876; div.approx.ftz.f32 %f18761, %f3826, %f18760; add.ftz.f32 %f3827, %f3825, %f18761; setp.eq.ftz.f32 %p2274, %f3825, %f3827; mov.f32 %f29952, %f3825; @%p2274 bra $L__BB107_2525; mul.ftz.f32 %f3828, %f3822, %f3826; mov.u64 %rd2877, 147456; cvt.rn.f32.u64 %f18762, %rd2877; div.approx.ftz.f32 %f18763, %f3828, %f18762; add.ftz.f32 %f3829, %f3827, %f18763; setp.eq.ftz.f32 %p2275, %f3827, %f3829; mov.f32 %f29952, %f3827; @%p2275 bra $L__BB107_2525; mul.ftz.f32 %f3830, %f3822, %f3828; mov.u64 %rd2878, 14745600; cvt.rn.f32.u64 %f18764, %rd2878; div.approx.ftz.f32 %f18765, %f3830, %f18764; add.ftz.f32 %f3831, %f3829, %f18765; setp.eq.ftz.f32 %p2276, %f3829, %f3831; mov.f32 %f29952, %f3829; @%p2276 bra $L__BB107_2525; mul.ftz.f32 %f3832, %f3822, %f3830; mov.u64 %rd2879, 2123366400; cvt.rn.f32.u64 %f18766, %rd2879; div.approx.ftz.f32 %f18767, %f3832, %f18766; add.ftz.f32 %f3833, %f3831, %f18767; setp.eq.ftz.f32 %p2277, %f3831, %f3833; mov.f32 %f29952, %f3831; @%p2277 bra $L__BB107_2525; mul.ftz.f32 %f3834, %f3822, %f3832; mov.u64 %rd2880, 416179814400; cvt.rn.f32.u64 %f18768, %rd2880; div.approx.ftz.f32 %f18769, %f3834, %f18768; add.ftz.f32 %f3835, %f3833, %f18769; setp.eq.ftz.f32 %p2278, %f3833, %f3835; mov.f32 %f29952, %f3833; @%p2278 bra $L__BB107_2525; mul.ftz.f32 %f3836, %f3822, %f3834; mov.u64 %rd2881, 106542032486400; cvt.rn.f32.u64 %f18770, %rd2881; div.approx.ftz.f32 %f18771, %f3836, %f18770; add.ftz.f32 %f3837, %f3835, %f18771; setp.eq.ftz.f32 %p2279, %f3835, %f3837; mov.f32 %f29952, %f3835; @%p2279 bra $L__BB107_2525; mul.ftz.f32 %f18772, %f3822, %f3836; mov.u64 %rd2882, 34519618525593600; cvt.rn.f32.u64 %f18773, %rd2882; div.approx.ftz.f32 %f18774, %f18772, %f18773; add.ftz.f32 %f3838, %f3837, %f18774; setp.eq.ftz.f32 %p2280, %f3837, %f3838; mov.f32 %f29952, %f3837; @%p2280 bra $L__BB107_2525; mov.f32 %f29952, %f3838; $L__BB107_2525: lg2.approx.ftz.f32 %f18775, %f29952; mul.ftz.f32 %f29953, %f18775, 0f3F317218; $L__BB107_2527: sub.ftz.f32 %f18781, %f29953, %f3821; sub.ftz.f32 %f18782, %f18781, %f3819; add.ftz.f32 %f18783, %f18782, 0f3F316F00; mul.ftz.f32 %f18784, %f3819, 0f3F000000; lg2.approx.ftz.f32 %f18785, %f18784; fma.rn.ftz.f32 %f18786, %f18785, 0f3F317218, %f18783; mul.ftz.f32 %f18787, %f18786, 0f3FB8AA3B; ex2.approx.ftz.f32 %f29956, %f18787; $L__BB107_2541: mov.f32 %f29958, 0f00000000; mov.f32 %f29957, %f29958; @%p2269 bra $L__BB107_2543; ex2.approx.ftz.f32 %f18840, %f3755; add.ftz.f32 %f18841, %f18840, 0f3F800000; rcp.approx.ftz.f32 %f18842, %f18841; add.ftz.f32 %f29957, %f18842, %f18842; $L__BB107_2543: sub.ftz.f32 %f3871, %f18554, %f29957; setp.eq.ftz.f32 %p2293, %f3871, 0f00000000; @%p2293 bra $L__BB107_2545; add.ftz.f32 %f18845, %f3704, %f3704; sub.ftz.f32 %f18846, %f18845, %f3753; add.ftz.f32 %f18847, %f18846, 0f40490FDB; sub.ftz.f32 %f18848, %f3747, %f18847; add.ftz.f32 %f18849, %f18848, 0f40490FDB; mov.f32 %f18850, 0f40C90FDB; div.approx.ftz.f32 %f18851, %f18849, %f18850; cvt.rmi.ftz.f32.f32 %f18852, %f18851; mul.ftz.f32 %f18853, %f18852, 0f40C90FDB; sub.ftz.f32 %f18854, %f18849, %f18853; add.ftz.f32 %f18855, %f18854, 0fC0490FDB; abs.ftz.f32 %f18856, %f18855; neg.ftz.f32 %f18857, %f18856; div.approx.ftz.f32 %f18858, %f18857, %f18609; mul.ftz.f32 %f18859, %f18858, 0f3FB8AA3B; ex2.approx.ftz.f32 %f18860, %f18859; add.ftz.f32 %f18861, %f18860, 0f3F800000; mul.ftz.f32 %f18862, %f18861, %f18861; mul.ftz.f32 %f18863, %f18609, %f18862; div.approx.ftz.f32 %f18864, %f18860, %f18863; div.approx.ftz.f32 %f29958, %f18864, %f3871; $L__BB107_2545: mul.ftz.f32 %f18865, %f29956, %f3713; mul.ftz.f32 %f18866, %f29956, %f3714; mul.ftz.f32 %f18867, %f29956, %f3715; fma.rn.ftz.f32 %f3874, %f29958, %f18865, %f3815; fma.rn.ftz.f32 %f3875, %f29958, %f18866, %f3815; fma.rn.ftz.f32 %f3876, %f29958, %f18867, %f3815; mul.ftz.f32 %f18868, %f29956, %f3807; fma.rn.ftz.f32 %f3877, %f29958, %f18868, %f3816; mul.ftz.f32 %f3878, %f3817, 0f40800000; rcp.approx.ftz.f32 %f3879, %f3878; mul.ftz.f32 %f18869, %f3694, %f3752; mul.ftz.f32 %f3880, %f18869, %f3879; mul.ftz.f32 %f18870, %f3689, %f3751; mul.ftz.f32 %f3881, %f18870, %f3879; setp.gtu.ftz.f32 %p2294, %f3878, 0f3DCCCCCD; @%p2294 bra $L__BB107_2559; bra.uni $L__BB107_2546; $L__BB107_2559: mul.ftz.f32 %f3904, %f3880, %f3880; fma.rn.ftz.f32 %f29961, %f3904, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3906, %f3904, %f3904; mov.u64 %rd2899, 64; cvt.rn.f32.u64 %f18901, %rd2899; div.approx.ftz.f32 %f18902, %f3906, %f18901; add.ftz.f32 %f3907, %f29961, %f18902; setp.eq.ftz.f32 %p2304, %f29961, %f3907; @%p2304 bra $L__BB107_2568; mul.ftz.f32 %f3908, %f3904, %f3906; mov.u64 %rd2900, 2304; cvt.rn.f32.u64 %f18903, %rd2900; div.approx.ftz.f32 %f18904, %f3908, %f18903; add.ftz.f32 %f3909, %f3907, %f18904; setp.eq.ftz.f32 %p2305, %f3907, %f3909; mov.f32 %f29961, %f3907; @%p2305 bra $L__BB107_2568; mul.ftz.f32 %f3910, %f3904, %f3908; mov.u64 %rd2901, 147456; cvt.rn.f32.u64 %f18905, %rd2901; div.approx.ftz.f32 %f18906, %f3910, %f18905; add.ftz.f32 %f3911, %f3909, %f18906; setp.eq.ftz.f32 %p2306, %f3909, %f3911; mov.f32 %f29961, %f3909; @%p2306 bra $L__BB107_2568; mul.ftz.f32 %f3912, %f3904, %f3910; mov.u64 %rd2902, 14745600; cvt.rn.f32.u64 %f18907, %rd2902; div.approx.ftz.f32 %f18908, %f3912, %f18907; add.ftz.f32 %f3913, %f3911, %f18908; setp.eq.ftz.f32 %p2307, %f3911, %f3913; mov.f32 %f29961, %f3911; @%p2307 bra $L__BB107_2568; mul.ftz.f32 %f3914, %f3904, %f3912; mov.u64 %rd2903, 2123366400; cvt.rn.f32.u64 %f18909, %rd2903; div.approx.ftz.f32 %f18910, %f3914, %f18909; add.ftz.f32 %f3915, %f3913, %f18910; setp.eq.ftz.f32 %p2308, %f3913, %f3915; mov.f32 %f29961, %f3913; @%p2308 bra $L__BB107_2568; mul.ftz.f32 %f3916, %f3904, %f3914; mov.u64 %rd2904, 416179814400; cvt.rn.f32.u64 %f18911, %rd2904; div.approx.ftz.f32 %f18912, %f3916, %f18911; add.ftz.f32 %f3917, %f3915, %f18912; setp.eq.ftz.f32 %p2309, %f3915, %f3917; mov.f32 %f29961, %f3915; @%p2309 bra $L__BB107_2568; mul.ftz.f32 %f3918, %f3904, %f3916; mov.u64 %rd2905, 106542032486400; cvt.rn.f32.u64 %f18913, %rd2905; div.approx.ftz.f32 %f18914, %f3918, %f18913; add.ftz.f32 %f3919, %f3917, %f18914; setp.eq.ftz.f32 %p2310, %f3917, %f3919; mov.f32 %f29961, %f3917; @%p2310 bra $L__BB107_2568; mul.ftz.f32 %f18915, %f3904, %f3918; mov.u64 %rd2906, 34519618525593600; cvt.rn.f32.u64 %f18916, %rd2906; div.approx.ftz.f32 %f18917, %f18915, %f18916; add.ftz.f32 %f3920, %f3919, %f18917; setp.eq.ftz.f32 %p2311, %f3919, %f3920; mov.f32 %f29961, %f3919; @%p2311 bra $L__BB107_2568; mov.f32 %f29961, %f3920; $L__BB107_2568: mul.ftz.f32 %f18918, %f3881, 0fBFB8AA3B; ex2.approx.ftz.f32 %f18919, %f18918; mul.ftz.f32 %f3922, %f29961, %f18919; abs.ftz.f32 %f3923, %f3879; setp.ltu.ftz.f32 %p2312, %f3923, 0f3F800000; @%p2312 bra $L__BB107_2570; bra.uni $L__BB107_2569; $L__BB107_2570: mul.ftz.f32 %f18941, %f3879, %f3879; mov.f32 %f18942, 0f394FFF49; mov.f32 %f18943, 0f363D0ADA; fma.rn.ftz.f32 %f18944, %f18943, %f18941, %f18942; mov.f32 %f18945, 0f3C08889A; fma.rn.ftz.f32 %f18946, %f18944, %f18941, %f18945; mov.f32 %f18947, 0f3E2AAAAB; fma.rn.ftz.f32 %f18948, %f18946, %f18941, %f18947; mul.ftz.f32 %f18949, %f18941, %f18948; fma.rn.ftz.f32 %f29962, %f18949, %f3879, %f3879; bra.uni $L__BB107_2571; $L__BB107_2546: setp.gt.ftz.f32 %p2295, %f3880, 0f41400000; @%p2295 bra $L__BB107_2557; bra.uni $L__BB107_2547; $L__BB107_2557: mul.ftz.f32 %f18889, %f3880, 0f41000000; rcp.approx.ftz.f32 %f18890, %f18889; add.ftz.f32 %f18891, %f18890, 0fBFEB3F8E; lg2.approx.ftz.f32 %f18892, %f3880; fma.rn.ftz.f32 %f18893, %f18892, 0fBF317218, %f18891; fma.rn.ftz.f32 %f29960, %f18893, 0f3F000000, %f3880; bra.uni $L__BB107_2558; $L__BB107_2569: mov.f32 %f18920, 0f3FB8AA3B; mul.rn.ftz.f32 %f18921, %f3923, %f18920; cvt.rzi.ftz.f32.f32 %f18922, %f18921; abs.ftz.f32 %f18923, %f18922; setp.gt.ftz.f32 %p2313, %f18923, 0f42FC0000; mov.b32 %r4090, %f18922; and.b32 %r4091, %r4090, -2147483648; or.b32 %r4092, %r4091, 1123811328; mov.b32 %f18924, %r4092; selp.f32 %f18925, %f18924, %f18922, %p2313; mov.f32 %f18926, 0fBF317218; fma.rn.ftz.f32 %f18927, %f18925, %f18926, %f3923; mov.f32 %f18928, 0f3102E308; fma.rn.ftz.f32 %f18929, %f18925, %f18928, %f18927; mul.ftz.f32 %f18930, %f18929, 0f3FB8AA3B; add.ftz.f32 %f18931, %f18925, 0f4B40007D; mov.b32 %r4093, %f18931; shl.b32 %r4094, %r4093, 23; mov.b32 %f18932, %r4094; ex2.approx.ftz.f32 %f18933, %f18930; mul.ftz.f32 %f18934, %f18933, %f18932; mov.f32 %f18935, 0f3E000000; div.approx.ftz.f32 %f18936, %f18935, %f18934; neg.ftz.f32 %f18937, %f18936; mov.f32 %f18938, 0f40000000; fma.rn.ftz.f32 %f18939, %f18938, %f18934, %f18937; setp.ge.ftz.f32 %p2314, %f3923, 0f42B40000; selp.f32 %f18940, 0f7F800000, %f18939, %p2314; mov.b32 %r4095, %f18940; mov.b32 %r4096, %f3879; and.b32 %r4097, %r4096, -2147483648; or.b32 %r4098, %r4097, %r4095; mov.b32 %f29962, %r4098; $L__BB107_2571: add.ftz.f32 %f18950, %f29962, %f29962; mul.ftz.f32 %f18951, %f3878, %f18950; div.approx.ftz.f32 %f29963, %f3922, %f18951; bra.uni $L__BB107_2572; $L__BB107_2547: mul.ftz.f32 %f3882, %f3880, %f3880; fma.rn.ftz.f32 %f29959, %f3882, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3884, %f3882, %f3882; mov.u64 %rd2891, 64; cvt.rn.f32.u64 %f18871, %rd2891; div.approx.ftz.f32 %f18872, %f3884, %f18871; add.ftz.f32 %f3885, %f29959, %f18872; setp.eq.ftz.f32 %p2296, %f29959, %f3885; @%p2296 bra $L__BB107_2556; mul.ftz.f32 %f3886, %f3882, %f3884; mov.u64 %rd2892, 2304; cvt.rn.f32.u64 %f18873, %rd2892; div.approx.ftz.f32 %f18874, %f3886, %f18873; add.ftz.f32 %f3887, %f3885, %f18874; setp.eq.ftz.f32 %p2297, %f3885, %f3887; mov.f32 %f29959, %f3885; @%p2297 bra $L__BB107_2556; mul.ftz.f32 %f3888, %f3882, %f3886; mov.u64 %rd2893, 147456; cvt.rn.f32.u64 %f18875, %rd2893; div.approx.ftz.f32 %f18876, %f3888, %f18875; add.ftz.f32 %f3889, %f3887, %f18876; setp.eq.ftz.f32 %p2298, %f3887, %f3889; mov.f32 %f29959, %f3887; @%p2298 bra $L__BB107_2556; mul.ftz.f32 %f3890, %f3882, %f3888; mov.u64 %rd2894, 14745600; cvt.rn.f32.u64 %f18877, %rd2894; div.approx.ftz.f32 %f18878, %f3890, %f18877; add.ftz.f32 %f3891, %f3889, %f18878; setp.eq.ftz.f32 %p2299, %f3889, %f3891; mov.f32 %f29959, %f3889; @%p2299 bra $L__BB107_2556; mul.ftz.f32 %f3892, %f3882, %f3890; mov.u64 %rd2895, 2123366400; cvt.rn.f32.u64 %f18879, %rd2895; div.approx.ftz.f32 %f18880, %f3892, %f18879; add.ftz.f32 %f3893, %f3891, %f18880; setp.eq.ftz.f32 %p2300, %f3891, %f3893; mov.f32 %f29959, %f3891; @%p2300 bra $L__BB107_2556; mul.ftz.f32 %f3894, %f3882, %f3892; mov.u64 %rd2896, 416179814400; cvt.rn.f32.u64 %f18881, %rd2896; div.approx.ftz.f32 %f18882, %f3894, %f18881; add.ftz.f32 %f3895, %f3893, %f18882; setp.eq.ftz.f32 %p2301, %f3893, %f3895; mov.f32 %f29959, %f3893; @%p2301 bra $L__BB107_2556; mul.ftz.f32 %f3896, %f3882, %f3894; mov.u64 %rd2897, 106542032486400; cvt.rn.f32.u64 %f18883, %rd2897; div.approx.ftz.f32 %f18884, %f3896, %f18883; add.ftz.f32 %f3897, %f3895, %f18884; setp.eq.ftz.f32 %p2302, %f3895, %f3897; mov.f32 %f29959, %f3895; @%p2302 bra $L__BB107_2556; mul.ftz.f32 %f18885, %f3882, %f3896; mov.u64 %rd2898, 34519618525593600; cvt.rn.f32.u64 %f18886, %rd2898; div.approx.ftz.f32 %f18887, %f18885, %f18886; add.ftz.f32 %f3898, %f3897, %f18887; setp.eq.ftz.f32 %p2303, %f3897, %f3898; mov.f32 %f29959, %f3897; @%p2303 bra $L__BB107_2556; mov.f32 %f29959, %f3898; $L__BB107_2556: lg2.approx.ftz.f32 %f18888, %f29959; mul.ftz.f32 %f29960, %f18888, 0f3F317218; $L__BB107_2558: sub.ftz.f32 %f18894, %f29960, %f3881; sub.ftz.f32 %f18895, %f18894, %f3879; add.ftz.f32 %f18896, %f18895, 0f3F316F00; mul.ftz.f32 %f18897, %f3879, 0f3F000000; lg2.approx.ftz.f32 %f18898, %f18897; fma.rn.ftz.f32 %f18899, %f18898, 0f3F317218, %f18896; mul.ftz.f32 %f18900, %f18899, 0f3FB8AA3B; ex2.approx.ftz.f32 %f29963, %f18900; $L__BB107_2572: mul.ftz.f32 %f18953, %f3704, 0f40800000; sub.ftz.f32 %f18954, %f18953, %f3753; add.ftz.f32 %f18955, %f18954, 0f40C90FDB; sub.ftz.f32 %f3929, %f3747, %f18955; mov.f32 %f29965, 0f00000000; mov.f32 %f29964, %f29965; @%p2269 bra $L__BB107_2574; ex2.approx.ftz.f32 %f18956, %f3755; add.ftz.f32 %f18957, %f18956, 0f3F800000; rcp.approx.ftz.f32 %f18958, %f18957; add.ftz.f32 %f29964, %f18958, %f18958; $L__BB107_2574: add.ftz.f32 %f18960, %f3929, 0f40490FDB; mov.f32 %f18961, 0f40C90FDB; div.approx.ftz.f32 %f18962, %f18960, %f18961; cvt.rmi.ftz.f32.f32 %f18963, %f18962; mul.ftz.f32 %f18964, %f18963, 0f40C90FDB; sub.ftz.f32 %f18965, %f18960, %f18964; add.ftz.f32 %f18966, %f18965, 0fC0490FDB; abs.ftz.f32 %f18967, %f18966; neg.ftz.f32 %f18968, %f18967; div.approx.ftz.f32 %f18969, %f18968, %f18609; mul.ftz.f32 %f18970, %f18969, 0f3FB8AA3B; ex2.approx.ftz.f32 %f18971, %f18970; add.ftz.f32 %f18972, %f18971, 0f3F800000; mul.ftz.f32 %f18974, %f18972, %f18972; mul.ftz.f32 %f18975, %f18609, %f18974; div.approx.ftz.f32 %f3932, %f18971, %f18975; sub.ftz.f32 %f3933, %f18554, %f29964; setp.eq.ftz.f32 %p2316, %f3933, 0f00000000; @%p2316 bra $L__BB107_2576; div.approx.ftz.f32 %f29965, %f3932, %f3933; $L__BB107_2576: mul.ftz.f32 %f18976, %f29963, %f3718; mul.ftz.f32 %f18977, %f29963, %f3720; mul.ftz.f32 %f18978, %f29963, %f3722; fma.rn.ftz.f32 %f3936, %f29965, %f18976, %f3874; fma.rn.ftz.f32 %f3937, %f29965, %f18977, %f3875; fma.rn.ftz.f32 %f3938, %f29965, %f18978, %f3876; mul.ftz.f32 %f18979, %f29963, %f3808; fma.rn.ftz.f32 %f3939, %f29965, %f18979, %f3877; mul.ftz.f32 %f18980, %f3694, %f3740; mul.ftz.f32 %f3940, %f18980, %f3879; mul.ftz.f32 %f18981, %f3690, %f3689; mul.ftz.f32 %f3941, %f18981, %f3879; @%p2294 bra $L__BB107_2590; bra.uni $L__BB107_2577; $L__BB107_2590: mul.ftz.f32 %f3964, %f3940, %f3940; fma.rn.ftz.f32 %f29968, %f3964, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3966, %f3964, %f3964; mov.u64 %rd2915, 64; cvt.rn.f32.u64 %f19012, %rd2915; div.approx.ftz.f32 %f19013, %f3966, %f19012; add.ftz.f32 %f3967, %f29968, %f19013; setp.eq.ftz.f32 %p2327, %f29968, %f3967; @%p2327 bra $L__BB107_2599; mul.ftz.f32 %f3968, %f3964, %f3966; mov.u64 %rd2916, 2304; cvt.rn.f32.u64 %f19014, %rd2916; div.approx.ftz.f32 %f19015, %f3968, %f19014; add.ftz.f32 %f3969, %f3967, %f19015; setp.eq.ftz.f32 %p2328, %f3967, %f3969; mov.f32 %f29968, %f3967; @%p2328 bra $L__BB107_2599; mul.ftz.f32 %f3970, %f3964, %f3968; mov.u64 %rd2917, 147456; cvt.rn.f32.u64 %f19016, %rd2917; div.approx.ftz.f32 %f19017, %f3970, %f19016; add.ftz.f32 %f3971, %f3969, %f19017; setp.eq.ftz.f32 %p2329, %f3969, %f3971; mov.f32 %f29968, %f3969; @%p2329 bra $L__BB107_2599; mul.ftz.f32 %f3972, %f3964, %f3970; mov.u64 %rd2918, 14745600; cvt.rn.f32.u64 %f19018, %rd2918; div.approx.ftz.f32 %f19019, %f3972, %f19018; add.ftz.f32 %f3973, %f3971, %f19019; setp.eq.ftz.f32 %p2330, %f3971, %f3973; mov.f32 %f29968, %f3971; @%p2330 bra $L__BB107_2599; mul.ftz.f32 %f3974, %f3964, %f3972; mov.u64 %rd2919, 2123366400; cvt.rn.f32.u64 %f19020, %rd2919; div.approx.ftz.f32 %f19021, %f3974, %f19020; add.ftz.f32 %f3975, %f3973, %f19021; setp.eq.ftz.f32 %p2331, %f3973, %f3975; mov.f32 %f29968, %f3973; @%p2331 bra $L__BB107_2599; mul.ftz.f32 %f3976, %f3964, %f3974; mov.u64 %rd2920, 416179814400; cvt.rn.f32.u64 %f19022, %rd2920; div.approx.ftz.f32 %f19023, %f3976, %f19022; add.ftz.f32 %f3977, %f3975, %f19023; setp.eq.ftz.f32 %p2332, %f3975, %f3977; mov.f32 %f29968, %f3975; @%p2332 bra $L__BB107_2599; mul.ftz.f32 %f3978, %f3964, %f3976; mov.u64 %rd2921, 106542032486400; cvt.rn.f32.u64 %f19024, %rd2921; div.approx.ftz.f32 %f19025, %f3978, %f19024; add.ftz.f32 %f3979, %f3977, %f19025; setp.eq.ftz.f32 %p2333, %f3977, %f3979; mov.f32 %f29968, %f3977; @%p2333 bra $L__BB107_2599; mul.ftz.f32 %f19026, %f3964, %f3978; mov.u64 %rd2922, 34519618525593600; cvt.rn.f32.u64 %f19027, %rd2922; div.approx.ftz.f32 %f19028, %f19026, %f19027; add.ftz.f32 %f3980, %f3979, %f19028; setp.eq.ftz.f32 %p2334, %f3979, %f3980; mov.f32 %f29968, %f3979; @%p2334 bra $L__BB107_2599; mov.f32 %f29968, %f3980; $L__BB107_2599: mul.ftz.f32 %f19029, %f3941, 0fBFB8AA3B; ex2.approx.ftz.f32 %f19030, %f19029; mul.ftz.f32 %f3982, %f29968, %f19030; abs.ftz.f32 %f3983, %f3879; setp.ltu.ftz.f32 %p2335, %f3983, 0f3F800000; @%p2335 bra $L__BB107_2601; bra.uni $L__BB107_2600; $L__BB107_2601: mul.ftz.f32 %f19052, %f3879, %f3879; mov.f32 %f19053, 0f394FFF49; mov.f32 %f19054, 0f363D0ADA; fma.rn.ftz.f32 %f19055, %f19054, %f19052, %f19053; mov.f32 %f19056, 0f3C08889A; fma.rn.ftz.f32 %f19057, %f19055, %f19052, %f19056; mov.f32 %f19058, 0f3E2AAAAB; fma.rn.ftz.f32 %f19059, %f19057, %f19052, %f19058; mul.ftz.f32 %f19060, %f19052, %f19059; fma.rn.ftz.f32 %f29969, %f19060, %f3879, %f3879; bra.uni $L__BB107_2602; $L__BB107_2577: setp.gt.ftz.f32 %p2318, %f3940, 0f41400000; @%p2318 bra $L__BB107_2588; bra.uni $L__BB107_2578; $L__BB107_2588: mul.ftz.f32 %f19000, %f3940, 0f41000000; rcp.approx.ftz.f32 %f19001, %f19000; add.ftz.f32 %f19002, %f19001, 0fBFEB3F8E; lg2.approx.ftz.f32 %f19003, %f3940; fma.rn.ftz.f32 %f19004, %f19003, 0fBF317218, %f19002; fma.rn.ftz.f32 %f29967, %f19004, 0f3F000000, %f3940; bra.uni $L__BB107_2589; $L__BB107_2600: mov.f32 %f19031, 0f3FB8AA3B; mul.rn.ftz.f32 %f19032, %f3983, %f19031; cvt.rzi.ftz.f32.f32 %f19033, %f19032; abs.ftz.f32 %f19034, %f19033; setp.gt.ftz.f32 %p2336, %f19034, 0f42FC0000; mov.b32 %r4099, %f19033; and.b32 %r4100, %r4099, -2147483648; or.b32 %r4101, %r4100, 1123811328; mov.b32 %f19035, %r4101; selp.f32 %f19036, %f19035, %f19033, %p2336; mov.f32 %f19037, 0fBF317218; fma.rn.ftz.f32 %f19038, %f19036, %f19037, %f3983; mov.f32 %f19039, 0f3102E308; fma.rn.ftz.f32 %f19040, %f19036, %f19039, %f19038; mul.ftz.f32 %f19041, %f19040, 0f3FB8AA3B; add.ftz.f32 %f19042, %f19036, 0f4B40007D; mov.b32 %r4102, %f19042; shl.b32 %r4103, %r4102, 23; mov.b32 %f19043, %r4103; ex2.approx.ftz.f32 %f19044, %f19041; mul.ftz.f32 %f19045, %f19044, %f19043; mov.f32 %f19046, 0f3E000000; div.approx.ftz.f32 %f19047, %f19046, %f19045; neg.ftz.f32 %f19048, %f19047; mov.f32 %f19049, 0f40000000; fma.rn.ftz.f32 %f19050, %f19049, %f19045, %f19048; setp.ge.ftz.f32 %p2337, %f3983, 0f42B40000; selp.f32 %f19051, 0f7F800000, %f19050, %p2337; mov.b32 %r4104, %f19051; mov.b32 %r4105, %f3879; and.b32 %r4106, %r4105, -2147483648; or.b32 %r4107, %r4106, %r4104; mov.b32 %f29969, %r4107; $L__BB107_2602: add.ftz.f32 %f19061, %f29969, %f29969; mul.ftz.f32 %f19062, %f3878, %f19061; div.approx.ftz.f32 %f29970, %f3982, %f19062; bra.uni $L__BB107_2603; $L__BB107_2578: mul.ftz.f32 %f3942, %f3940, %f3940; fma.rn.ftz.f32 %f29966, %f3942, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3944, %f3942, %f3942; mov.u64 %rd2907, 64; cvt.rn.f32.u64 %f18982, %rd2907; div.approx.ftz.f32 %f18983, %f3944, %f18982; add.ftz.f32 %f3945, %f29966, %f18983; setp.eq.ftz.f32 %p2319, %f29966, %f3945; @%p2319 bra $L__BB107_2587; mul.ftz.f32 %f3946, %f3942, %f3944; mov.u64 %rd2908, 2304; cvt.rn.f32.u64 %f18984, %rd2908; div.approx.ftz.f32 %f18985, %f3946, %f18984; add.ftz.f32 %f3947, %f3945, %f18985; setp.eq.ftz.f32 %p2320, %f3945, %f3947; mov.f32 %f29966, %f3945; @%p2320 bra $L__BB107_2587; mul.ftz.f32 %f3948, %f3942, %f3946; mov.u64 %rd2909, 147456; cvt.rn.f32.u64 %f18986, %rd2909; div.approx.ftz.f32 %f18987, %f3948, %f18986; add.ftz.f32 %f3949, %f3947, %f18987; setp.eq.ftz.f32 %p2321, %f3947, %f3949; mov.f32 %f29966, %f3947; @%p2321 bra $L__BB107_2587; mul.ftz.f32 %f3950, %f3942, %f3948; mov.u64 %rd2910, 14745600; cvt.rn.f32.u64 %f18988, %rd2910; div.approx.ftz.f32 %f18989, %f3950, %f18988; add.ftz.f32 %f3951, %f3949, %f18989; setp.eq.ftz.f32 %p2322, %f3949, %f3951; mov.f32 %f29966, %f3949; @%p2322 bra $L__BB107_2587; mul.ftz.f32 %f3952, %f3942, %f3950; mov.u64 %rd2911, 2123366400; cvt.rn.f32.u64 %f18990, %rd2911; div.approx.ftz.f32 %f18991, %f3952, %f18990; add.ftz.f32 %f3953, %f3951, %f18991; setp.eq.ftz.f32 %p2323, %f3951, %f3953; mov.f32 %f29966, %f3951; @%p2323 bra $L__BB107_2587; mul.ftz.f32 %f3954, %f3942, %f3952; mov.u64 %rd2912, 416179814400; cvt.rn.f32.u64 %f18992, %rd2912; div.approx.ftz.f32 %f18993, %f3954, %f18992; add.ftz.f32 %f3955, %f3953, %f18993; setp.eq.ftz.f32 %p2324, %f3953, %f3955; mov.f32 %f29966, %f3953; @%p2324 bra $L__BB107_2587; mul.ftz.f32 %f3956, %f3942, %f3954; mov.u64 %rd2913, 106542032486400; cvt.rn.f32.u64 %f18994, %rd2913; div.approx.ftz.f32 %f18995, %f3956, %f18994; add.ftz.f32 %f3957, %f3955, %f18995; setp.eq.ftz.f32 %p2325, %f3955, %f3957; mov.f32 %f29966, %f3955; @%p2325 bra $L__BB107_2587; mul.ftz.f32 %f18996, %f3942, %f3956; mov.u64 %rd2914, 34519618525593600; cvt.rn.f32.u64 %f18997, %rd2914; div.approx.ftz.f32 %f18998, %f18996, %f18997; add.ftz.f32 %f3958, %f3957, %f18998; setp.eq.ftz.f32 %p2326, %f3957, %f3958; mov.f32 %f29966, %f3957; @%p2326 bra $L__BB107_2587; mov.f32 %f29966, %f3958; $L__BB107_2587: lg2.approx.ftz.f32 %f18999, %f29966; mul.ftz.f32 %f29967, %f18999, 0f3F317218; $L__BB107_2589: sub.ftz.f32 %f19005, %f29967, %f3941; sub.ftz.f32 %f19006, %f19005, %f3879; add.ftz.f32 %f19007, %f19006, 0f3F316F00; mul.ftz.f32 %f19008, %f3879, 0f3F000000; lg2.approx.ftz.f32 %f19009, %f19008; fma.rn.ftz.f32 %f19010, %f19009, 0f3F317218, %f19007; mul.ftz.f32 %f19011, %f19010, 0f3FB8AA3B; ex2.approx.ftz.f32 %f29970, %f19011; $L__BB107_2603: mul.ftz.f32 %f19063, %f29970, %f3733; mul.ftz.f32 %f19064, %f29970, %f3734; mul.ftz.f32 %f19065, %f29970, %f3735; fma.rn.ftz.f32 %f30104, %f19063, 0f3E22F983, %f3936; fma.rn.ftz.f32 %f30105, %f19064, 0f3E22F983, %f3937; fma.rn.ftz.f32 %f30106, %f19065, 0f3E22F983, %f3938; mul.ftz.f32 %f19066, %f29970, %f3809; fma.rn.ftz.f32 %f30103, %f19066, 0f3E22F983, %f3939; bra.uni $L__BB107_2774; $L__BB107_2723: ld.local.v2.f32 {%f20240, %f20241}, [%rd564+44]; ld.local.f32 %f20244, [%rd564+52]; mul.ftz.f32 %f20245, %f4412, %f20244; mul.ftz.f32 %f20246, %f4413, %f20241; sub.ftz.f32 %f20247, %f20245, %f20246; mul.ftz.f32 %f20248, %f4413, %f20240; mul.ftz.f32 %f20249, %f4411, %f20244; sub.ftz.f32 %f20250, %f20248, %f20249; mul.ftz.f32 %f20251, %f4411, %f20241; mul.ftz.f32 %f20252, %f4412, %f20240; sub.ftz.f32 %f20253, %f20251, %f20252; mul.ftz.f32 %f20254, %f20250, %f20250; fma.rn.ftz.f32 %f20255, %f20247, %f20247, %f20254; fma.rn.ftz.f32 %f20256, %f20253, %f20253, %f20255; sqrt.approx.ftz.f32 %f20257, %f20256; div.approx.ftz.f32 %f20258, %f20247, %f20257; div.approx.ftz.f32 %f20259, %f20250, %f20257; div.approx.ftz.f32 %f20260, %f20253, %f20257; mul.ftz.f32 %f20261, %f4413, %f20259; mul.ftz.f32 %f20262, %f4412, %f20260; sub.ftz.f32 %f20263, %f20261, %f20262; mul.ftz.f32 %f20264, %f4411, %f20260; mul.ftz.f32 %f20265, %f4413, %f20258; sub.ftz.f32 %f20266, %f20264, %f20265; mul.ftz.f32 %f20267, %f4412, %f20258; mul.ftz.f32 %f20268, %f4411, %f20259; sub.ftz.f32 %f20269, %f20267, %f20268; mul.ftz.f32 %f20270, %f4417, %f20266; fma.rn.ftz.f32 %f20271, %f4416, %f20263, %f20270; fma.rn.ftz.f32 %f20272, %f4418, %f20269, %f20271; mul.ftz.f32 %f20273, %f4417, %f20259; fma.rn.ftz.f32 %f20274, %f4416, %f20258, %f20273; fma.rn.ftz.f32 %f20275, %f4418, %f20260, %f20274; neg.ftz.f32 %f20276, %f20272; mul.ftz.f32 %f20277, %f20218, %f4419; div.approx.ftz.f32 %f20278, %f20276, %f20277; neg.ftz.f32 %f20279, %f20275; mul.ftz.f32 %f20280, %f20219, %f4419; div.approx.ftz.f32 %f20281, %f20279, %f20280; fma.rn.ftz.f32 %f20282, %f20278, %f20278, 0f3F800000; mov.f32 %f20283, 0f3F800000; fma.rn.ftz.f32 %f20284, %f20281, %f20281, %f20282; mul.ftz.f32 %f20285, %f20284, %f20284; mul.ftz.f32 %f20286, %f20285, 0f40490FDB; mul.ftz.f32 %f20287, %f30070, %f20286; mul.ftz.f32 %f20288, %f4420, %f4420; mul.ftz.f32 %f20289, %f20288, %f20287; rcp.approx.ftz.f32 %f30071, %f20289; mul.ftz.f32 %f20290, %f4414, %f4414; sub.ftz.f32 %f20291, %f20283, %f20290; div.approx.ftz.f32 %f20292, %f20291, %f20290; mul.ftz.f32 %f20293, %f20216, %f20266; fma.rn.ftz.f32 %f20294, %f4404, %f20263, %f20293; fma.rn.ftz.f32 %f20295, %f20217, %f20269, %f20294; mul.ftz.f32 %f20296, %f20216, %f20259; fma.rn.ftz.f32 %f20297, %f4404, %f20258, %f20296; fma.rn.ftz.f32 %f20298, %f20217, %f20260, %f20297; mul.ftz.f32 %f20299, %f20295, %f20295; mul.ftz.f32 %f20300, %f20218, %f20218; mul.ftz.f32 %f20301, %f20300, %f20299; mul.ftz.f32 %f20302, %f20298, %f20298; mul.ftz.f32 %f20303, %f20219, %f20219; fma.rn.ftz.f32 %f20304, %f20303, %f20302, %f20301; add.ftz.f32 %f20305, %f20302, %f20299; div.approx.ftz.f32 %f20306, %f20304, %f20305; fma.rn.ftz.f32 %f20307, %f20292, %f20306, 0f3F800000; mov.f32 %f20308, 0f00000000; max.ftz.f32 %f20309, %f20307, %f20308; sqrt.approx.ftz.f32 %f20310, %f20309; add.ftz.f32 %f20311, %f20310, 0f3F800000; mov.f32 %f20312, 0f40000000; div.approx.ftz.f32 %f30073, %f20312, %f20311; mul.ftz.f32 %f20313, %f4415, %f4415; sub.ftz.f32 %f20314, %f20283, %f20313; div.approx.ftz.f32 %f20315, %f20314, %f20313; mul.ftz.f32 %f20316, %f30259, %f20266; fma.rn.ftz.f32 %f20317, %f30260, %f20263, %f20316; fma.rn.ftz.f32 %f20318, %f30258, %f20269, %f20317; mul.ftz.f32 %f20319, %f30259, %f20259; fma.rn.ftz.f32 %f20320, %f30260, %f20258, %f20319; fma.rn.ftz.f32 %f20321, %f30258, %f20260, %f20320; mul.ftz.f32 %f20322, %f20318, %f20318; mul.ftz.f32 %f20323, %f20300, %f20322; mul.ftz.f32 %f20324, %f20321, %f20321; fma.rn.ftz.f32 %f20325, %f20303, %f20324, %f20323; add.ftz.f32 %f20326, %f20324, %f20322; div.approx.ftz.f32 %f20327, %f20325, %f20326; mul.ftz.f32 %f30074, %f20315, %f20327; bra.uni $L__BB107_2729; $L__BB107_2725: mul.ftz.f32 %f20328, %f4420, %f4420; mul.ftz.f32 %f20329, %f20328, 0f40490FDB; mov.f32 %f20330, 0f3F800000; sub.ftz.f32 %f20331, %f20330, %f4420; div.approx.ftz.f32 %f20332, %f20331, %f4420; add.ftz.f32 %f20333, %f30070, %f20332; mul.ftz.f32 %f20334, %f20329, %f20333; mul.ftz.f32 %f20335, %f20333, %f20334; div.approx.ftz.f32 %f30071, %f30070, %f20335; $L__BB107_2728: mul.ftz.f32 %f20347, %f4414, %f4414; mov.f32 %f20348, 0f3F800000; sub.ftz.f32 %f20349, %f20348, %f20347; mul.ftz.f32 %f20350, %f20349, %f30070; div.approx.ftz.f32 %f20351, %f20350, %f20347; add.ftz.f32 %f20352, %f20351, 0f3F800000; mov.f32 %f20353, 0f00000000; max.ftz.f32 %f20354, %f20352, %f20353; sqrt.approx.ftz.f32 %f20355, %f20354; add.ftz.f32 %f20356, %f20355, 0f3F800000; mov.f32 %f20357, 0f40000000; div.approx.ftz.f32 %f30073, %f20357, %f20356; mul.ftz.f32 %f20358, %f4415, %f4415; sub.ftz.f32 %f20359, %f20348, %f20358; mul.ftz.f32 %f20360, %f20359, %f30070; div.approx.ftz.f32 %f30074, %f20360, %f20358; $L__BB107_2729: add.ftz.f32 %f20364, %f30074, 0f3F800000; mov.f32 %f20365, 0f00000000; max.ftz.f32 %f20366, %f20364, %f20365; sqrt.approx.ftz.f32 %f20367, %f20366; add.ftz.f32 %f20368, %f20367, 0f3F800000; mov.f32 %f20369, 0f40000000; div.approx.ftz.f32 %f20370, %f20369, %f20368; mul.ftz.f32 %f4433, %f30073, %f20370; mul.ftz.f32 %f20371, %f30071, 0f3E800000; div.approx.ftz.f32 %f4434, %f20371, %f4414; add.s32 %r4128, %r963, -11; setp.gt.u32 %p2486, %r4128, 1; mov.f32 %f30080, 0f3F800000; mov.f32 %f30081, %f30080; mov.f32 %f30082, %f30080; @%p2486 bra $L__BB107_2735; ld.local.f32 %f20373, [%rd564+28]; fma.rn.ftz.f32 %f4435, %f20373, %f20373, 0fBF800000; mov.f32 %f20372, 0f3F800000; abs.ftz.f32 %f4436, %f20372; fma.rn.ftz.f32 %f4437, %f4436, %f4436, %f4435; setp.leu.ftz.f32 %p2487, %f4437, 0f00000000; mov.f32 %f30075, %f20372; @%p2487 bra $L__BB107_2732; sqrt.approx.ftz.f32 %f20374, %f4437; sub.ftz.f32 %f20375, %f20374, %f4436; add.ftz.f32 %f20376, %f4436, %f20374; div.approx.ftz.f32 %f20377, %f20375, %f20376; fma.rn.ftz.f32 %f20378, %f4436, %f20376, 0fBF800000; fma.rn.ftz.f32 %f20379, %f4436, %f20375, 0f3F800000; div.approx.ftz.f32 %f20380, %f20378, %f20379; mul.ftz.f32 %f20381, %f20377, 0f3F000000; mul.ftz.f32 %f20382, %f20377, %f20381; fma.rn.ftz.f32 %f20383, %f20380, %f20380, 0f3F800000; mul.ftz.f32 %f30075, %f20382, %f20383; $L__BB107_2732: ld.local.u64 %rd2924, [%rd564+36]; ld.f32 %f4440, [%rd2924+12]; ld.f32 %f4441, [%rd2924+16]; ld.f32 %f4442, [%rd2924+20]; sub.ftz.f32 %f4443, %f20372, %f30075; mul.ftz.f32 %f20385, %f30259, %f4417; fma.rn.ftz.f32 %f20386, %f30260, %f4416, %f20385; fma.rn.ftz.f32 %f20387, %f30258, %f4418, %f20386; abs.ftz.f32 %f4444, %f20387; fma.rn.ftz.f32 %f4445, %f4444, %f4444, %f4435; setp.leu.ftz.f32 %p2488, %f4445, 0f00000000; mov.f32 %f30076, %f4443; @%p2488 bra $L__BB107_2734; sqrt.approx.ftz.f32 %f20388, %f4445; sub.ftz.f32 %f20389, %f20388, %f4444; add.ftz.f32 %f20390, %f4444, %f20388; div.approx.ftz.f32 %f20391, %f20389, %f20390; fma.rn.ftz.f32 %f20392, %f4444, %f20390, 0fBF800000; fma.rn.ftz.f32 %f20393, %f4444, %f20389, 0f3F800000; div.approx.ftz.f32 %f20394, %f20392, %f20393; mul.ftz.f32 %f20395, %f20391, 0f3F000000; mul.ftz.f32 %f20396, %f20391, %f20395; fma.rn.ftz.f32 %f20397, %f20394, %f20394, 0f3F800000; mul.ftz.f32 %f20398, %f20396, %f20397; sub.ftz.f32 %f30076, %f20398, %f30075; $L__BB107_2734: rcp.approx.ftz.f32 %f20399, %f4443; mov.f32 %f20400, 0f3F800000; mul.ftz.f32 %f20401, %f20399, %f30076; sub.ftz.f32 %f20402, %f20400, %f20401; fma.rn.ftz.f32 %f30082, %f4442, %f20402, %f20401; fma.rn.ftz.f32 %f30081, %f4441, %f20402, %f20401; fma.rn.ftz.f32 %f30080, %f4440, %f20402, %f20401; $L__BB107_2735: setp.ne.s32 %p2489, %r963, 12; @%p2489 bra $L__BB107_2737; ld.local.u64 %rd2925, [%rd564+36]; ld.f32 %f20403, [%rd2925+36]; mul.ftz.f32 %f20404, %f20403, 0f3E800000; mul.ftz.f32 %f30080, %f30080, %f20404; mul.ftz.f32 %f30081, %f30081, %f20404; mul.ftz.f32 %f30082, %f30082, %f20404; $L__BB107_2737: mul.ftz.f32 %f20405, %f4433, %f30080; mul.ftz.f32 %f20406, %f4433, %f30081; mul.ftz.f32 %f20407, %f4433, %f30082; mul.ftz.f32 %f30106, %f4434, %f20407; mul.ftz.f32 %f30105, %f4434, %f20406; mul.ftz.f32 %f30104, %f4434, %f20405; mul.ftz.f32 %f30103, %f30073, %f4434; bra.uni $L__BB107_2774; $L__BB107_2348: setp.lt.ftz.f32 %p2115, %f3314, 0f00000000; @%p2115 bra $L__BB107_2356; mul.ftz.f32 %f17699, %f29828, %f3313; fma.rn.ftz.f32 %f17700, %f29827, %f3312, %f17699; fma.rn.ftz.f32 %f3316, %f29829, %f3314, %f17700; abs.ftz.f32 %f3317, %f3316; fma.rn.ftz.f32 %f3318, %f3317, %f3317, %f3262; setp.leu.ftz.f32 %p2116, %f3318, 0f00000000; mov.f32 %f17698, 0f3F800000; mov.f32 %f29835, %f17698; @%p2116 bra $L__BB107_2351; sqrt.approx.ftz.f32 %f17701, %f3318; sub.ftz.f32 %f17702, %f17701, %f3317; add.ftz.f32 %f17703, %f3317, %f17701; div.approx.ftz.f32 %f17704, %f17702, %f17703; fma.rn.ftz.f32 %f17705, %f3317, %f17703, 0fBF800000; fma.rn.ftz.f32 %f17706, %f3317, %f17702, 0f3F800000; div.approx.ftz.f32 %f17707, %f17705, %f17706; mul.ftz.f32 %f17708, %f17704, 0f3F000000; mul.ftz.f32 %f17709, %f17704, %f17708; fma.rn.ftz.f32 %f17710, %f17707, %f17707, 0f3F800000; mul.ftz.f32 %f29835, %f17709, %f17710; $L__BB107_2351: mov.f32 %f17711, 0f00000000; max.ftz.f32 %f17712, %f17711, %f3316; mul.ftz.f32 %f17713, %f29835, %f17712; mul.ftz.f32 %f17714, %f3314, %f3314; sub.ftz.f32 %f17716, %f17698, %f17714; fma.rn.ftz.f32 %f17717, %f3258, %f17714, %f17716; mul.ftz.f32 %f17718, %f17717, 0f40490FDB; mul.ftz.f32 %f17719, %f17717, %f17718; max.ftz.f32 %f17721, %f17719, %f17556; div.approx.ftz.f32 %f17722, %f3258, %f17721; mul.ftz.f32 %f17723, %f17713, %f17722; mul.ftz.f32 %f17724, %f17723, 0f3E800000; mul.ftz.f32 %f17725, %f3315, %f3316; div.approx.ftz.f32 %f29838, %f17724, %f17725; $L__BB107_2356: neg.ftz.f32 %f3338, %f29921; @%p2093 bra $L__BB107_2358; bra.uni $L__BB107_2357; $L__BB107_2358: add.ftz.f32 %f17779, %f29834, 0f3F800000; { // callseq 460, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f3338; .param .b32 param1; st.param.f32 [param1+0], %f17779; .param .b32 retval0; call.uni (retval0), _Z4betaff, ( param0, param1 ); ld.param.f32 %f17780, [retval0+0]; } // callseq 460 mul.ftz.f32 %f29839, %f17780, %f3338; bra.uni $L__BB107_2359; $L__BB107_2357: sub.ftz.f32 %f17778, %f29834, %f29921; div.approx.ftz.f32 %f29839, %f3338, %f17778; $L__BB107_2359: mul.ftz.f32 %f30105, %f29838, %f29839; mul.ftz.f32 %f30104, %f29838, %f29839; mul.ftz.f32 %f30106, %f29838, %f29839; mov.f32 %f29864, 0f3F800000; abs.ftz.f32 %f3345, %f29864; fma.rn.ftz.f32 %f3346, %f3345, %f3345, %f3262; setp.leu.ftz.f32 %p2121, %f3346, 0f00000000; mov.f32 %f29840, %f29864; @%p2121 bra $L__BB107_2361; sqrt.approx.ftz.f32 %f17782, %f3346; sub.ftz.f32 %f17783, %f17782, %f3345; add.ftz.f32 %f17784, %f3345, %f17782; div.approx.ftz.f32 %f17785, %f17783, %f17784; fma.rn.ftz.f32 %f17786, %f3345, %f17784, 0fBF800000; fma.rn.ftz.f32 %f17787, %f3345, %f17783, 0f3F800000; div.approx.ftz.f32 %f17788, %f17786, %f17787; mul.ftz.f32 %f17789, %f17785, 0f3F000000; mul.ftz.f32 %f17790, %f17785, %f17789; fma.rn.ftz.f32 %f17791, %f17788, %f17788, 0f3F800000; mul.ftz.f32 %f29840, %f17790, %f17791; $L__BB107_2361: setp.ne.s32 %p2122, %r963, 27; or.pred %p9, %p2093, %p2122; sub.ftz.f32 %f3352, %f29864, %f29840; rcp.approx.ftz.f32 %f3353, %f3352; mov.f32 %f29918, %f29864; mov.f32 %f29919, %f29864; mov.f32 %f29920, %f29864; @%p9 bra $L__BB107_2365; mul.ftz.f32 %f17795, %f29828, %f3313; fma.rn.ftz.f32 %f17796, %f29827, %f3312, %f17795; fma.rn.ftz.f32 %f17797, %f29829, %f3314, %f17796; abs.ftz.f32 %f3354, %f17797; fma.rn.ftz.f32 %f3355, %f3354, %f3354, %f3262; setp.leu.ftz.f32 %p2124, %f3355, 0f00000000; mov.f32 %f29841, %f3352; @%p2124 bra $L__BB107_2364; sqrt.approx.ftz.f32 %f17798, %f3355; sub.ftz.f32 %f17799, %f17798, %f3354; add.ftz.f32 %f17800, %f3354, %f17798; div.approx.ftz.f32 %f17801, %f17799, %f17800; fma.rn.ftz.f32 %f17802, %f3354, %f17800, 0fBF800000; fma.rn.ftz.f32 %f17803, %f3354, %f17799, 0f3F800000; div.approx.ftz.f32 %f17804, %f17802, %f17803; mul.ftz.f32 %f17805, %f17801, 0f3F000000; mul.ftz.f32 %f17806, %f17801, %f17805; fma.rn.ftz.f32 %f17807, %f17804, %f17804, 0f3F800000; mul.ftz.f32 %f17808, %f17806, %f17807; sub.ftz.f32 %f29841, %f17808, %f29840; $L__BB107_2364: mul.ftz.f32 %f17809, %f3353, %f29841; mov.f32 %f17810, 0f3F800000; sub.ftz.f32 %f17811, %f17810, %f17809; fma.rn.ftz.f32 %f29918, %f29821, %f17811, %f17809; fma.rn.ftz.f32 %f29919, %f29822, %f17811, %f17809; fma.rn.ftz.f32 %f29920, %f29823, %f17811, %f17809; mul.ftz.f32 %f30104, %f30104, %f29918; mul.ftz.f32 %f30105, %f30105, %f29919; mul.ftz.f32 %f30106, %f30106, %f29920; $L__BB107_2365: neg.ftz.f32 %f3370, %f29824; neg.ftz.f32 %f3371, %f29825; neg.ftz.f32 %f3372, %f29826; rcp.approx.ftz.f32 %f3373, %f3245; mul.ftz.f32 %f3374, %f3373, %f3370; mul.ftz.f32 %f3375, %f3373, %f3371; mul.ftz.f32 %f3376, %f3373, %f3372; mul.ftz.f32 %f3377, %f3245, %f29824; mul.ftz.f32 %f3378, %f3245, %f29825; mul.ftz.f32 %f3379, %f3245, %f29826; fma.rn.ftz.f32 %f3380, %f3373, %f3373, 0fBF800000; ld.local.u32 %r5782, [%rd5+168]; mov.f32 %f29862, 0f00000000; selp.f32 %f3381, %f3372, %f29826, %p2093; mov.u32 %r5781, 0; mov.u16 %rs720, %rs534; mov.f32 %f29863, %f29864; $L__BB107_2366: mad.lo.s32 %r5782, %r5782, 1103515245, 12345; cvt.rn.f32.u32 %f17815, %r5782; mul.ftz.f32 %f3395, %f17815, 0f2F800000; setp.gt.ftz.f32 %p2127, %f29917, 0f3F7FF972; mov.pred %p2126, -1; mov.pred %p3922, %p2126; @%p2127 bra $L__BB107_2376; bra.uni $L__BB107_2367; $L__BB107_2376: @%p3922 bra $L__BB107_2457; mad.lo.s32 %r4047, %r5782, 1103515245, 12345; cvt.rn.f32.u32 %f17835, %r4047; mul.ftz.f32 %f3406, %f17835, 0f2F800000; mad.lo.s32 %r5782, %r4047, 1103515245, 12345; cvt.rn.f32.u32 %f17836, %r5782; mul.ftz.f32 %f3407, %f17836, 0f2F800000; neg.ftz.f32 %f3408, %f29915; mul.ftz.f32 %f17837, %f17470, %f3408; neg.ftz.f32 %f3409, %f29916; mul.ftz.f32 %f17838, %f17471, %f3409; mul.ftz.f32 %f17839, %f17838, %f17838; fma.rn.ftz.f32 %f17840, %f17837, %f17837, %f17839; fma.rn.ftz.f32 %f17841, %f29917, %f29917, %f17840; sqrt.approx.ftz.f32 %f17842, %f17841; div.approx.ftz.f32 %f29867, %f17837, %f17842; div.approx.ftz.f32 %f29868, %f17838, %f17842; neg.ftz.f32 %f3412, %f29917; div.approx.ftz.f32 %f3413, %f3412, %f17842; setp.gt.ftz.f32 %p2140, %f3413, 0f3F7FF972; @%p2140 bra $L__BB107_2383; abs.ftz.f32 %f17843, %f3413; setp.lt.ftz.f32 %p2141, %f17843, 0f358637BD; @%p2141 bra $L__BB107_2383; bra.uni $L__BB107_2379; $L__BB107_2383: mov.f32 %f17887, 0f3F800000; sub.ftz.f32 %f17888, %f17887, %f3407; max.ftz.f32 %f17890, %f17888, %f17556; div.approx.ftz.f32 %f17891, %f3407, %f17890; sqrt.approx.ftz.f32 %f17892, %f17891; mul.ftz.f32 %f17893, %f3406, 0f40C90FDB; cos.approx.ftz.f32 %f17894, %f17893; sin.approx.ftz.f32 %f17895, %f17893; mul.ftz.f32 %f29866, %f17892, %f17895; mul.ftz.f32 %f29865, %f17892, %f17894; $L__BB107_2384: mul.ftz.f32 %f17896, %f29868, %f29868; fma.rn.ftz.f32 %f17897, %f29867, %f29867, %f17896; add.ftz.f32 %f17898, %f17897, 0f00000000; sqrt.approx.ftz.f32 %f3429, %f17898; setp.neu.ftz.f32 %p2149, %f3429, 0f00000000; @%p2149 bra $L__BB107_2385; bra.uni $L__BB107_2386; $L__BB107_2385: rcp.approx.ftz.f32 %f17899, %f3429; mul.ftz.f32 %f29868, %f29868, %f17899; mul.ftz.f32 %f29867, %f29867, %f17899; $L__BB107_2386: mul.ftz.f32 %f17900, %f29865, %f29867; mul.ftz.f32 %f17901, %f29866, %f29868; sub.ftz.f32 %f17902, %f17900, %f17901; mul.ftz.f32 %f17903, %f17470, %f17902; mul.ftz.f32 %f17904, %f29866, %f29867; fma.rn.ftz.f32 %f17905, %f29865, %f29868, %f17904; mul.ftz.f32 %f17906, %f17471, %f17905; neg.ftz.f32 %f17907, %f17903; neg.ftz.f32 %f17908, %f17906; mul.ftz.f32 %f17909, %f17906, %f17906; fma.rn.ftz.f32 %f17910, %f17903, %f17903, %f17909; add.ftz.f32 %f17911, %f17910, 0f3F800000; mov.f32 %f17912, 0f3F800000; sqrt.approx.ftz.f32 %f17913, %f17911; div.approx.ftz.f32 %f3439, %f17907, %f17913; div.approx.ftz.f32 %f3440, %f17908, %f17913; div.approx.ftz.f32 %f3441, %f17912, %f17913; setp.ne.s32 %p2150, %r5781, 0; not.pred %p2151, %p8; or.pred %p2152, %p2151, %p2150; @%p2152 bra $L__BB107_2411; and.b16 %rs527, %rs720, 255; mov.f32 %f29872, 0f00000000; setp.eq.s16 %p2153, %rs527, 0; @%p2153 bra $L__BB107_2398; @%p2127 bra $L__BB107_2408; mul.ftz.f32 %f17917, %f29917, %f29921; setp.lt.ftz.f32 %p2156, %f29917, 0fBF7FF972; selp.f32 %f3442, 0f3F800000, %f17917, %p2156; @%p2093 bra $L__BB107_2394; bra.uni $L__BB107_2390; $L__BB107_2394: sub.ftz.f32 %f17959, %f3377, %f29915; sub.ftz.f32 %f17960, %f3378, %f29916; mul.ftz.f32 %f17961, %f17960, %f17960; fma.rn.ftz.f32 %f17962, %f17959, %f17959, %f17961; sub.ftz.f32 %f17963, %f3379, %f29917; fma.rn.ftz.f32 %f17964, %f17963, %f17963, %f17962; sqrt.approx.ftz.f32 %f17965, %f17964; div.approx.ftz.f32 %f17966, %f17959, %f17965; div.approx.ftz.f32 %f17967, %f17960, %f17965; div.approx.ftz.f32 %f17968, %f17963, %f17965; setp.lt.ftz.f32 %p2159, %f17968, 0f00000000; neg.ftz.f32 %f17969, %f17966; neg.ftz.f32 %f17970, %f17967; neg.ftz.f32 %f17971, %f17968; selp.f32 %f17972, %f17969, %f17966, %p2159; selp.f32 %f17973, %f17970, %f17967, %p2159; selp.f32 %f3452, %f17971, %f17968, %p2159; mul.ftz.f32 %f17974, %f17973, %f3409; fma.rn.ftz.f32 %f17975, %f17972, %f3408, %f17974; fma.rn.ftz.f32 %f3453, %f3452, %f3412, %f17975; mul.ftz.f32 %f17976, %f29825, %f17973; fma.rn.ftz.f32 %f17977, %f29824, %f17972, %f17976; fma.rn.ftz.f32 %f3454, %f29826, %f3452, %f17977; setp.lt.ftz.f32 %p2160, %f3453, 0f00000000; @%p2160 bra $L__BB107_2408; fma.rn.ftz.f32 %f3455, %f3245, %f3454, %f3453; abs.ftz.f32 %f3456, %f3453; fma.rn.ftz.f32 %f3457, %f3456, %f3456, %f3262; setp.leu.ftz.f32 %p2161, %f3457, 0f00000000; mov.f32 %f17978, 0f3F800000; mov.f32 %f29870, %f17978; @%p2161 bra $L__BB107_2397; sqrt.approx.ftz.f32 %f17979, %f3457; sub.ftz.f32 %f17980, %f17979, %f3456; add.ftz.f32 %f17981, %f3456, %f17979; div.approx.ftz.f32 %f17982, %f17980, %f17981; fma.rn.ftz.f32 %f17983, %f3456, %f17981, 0fBF800000; fma.rn.ftz.f32 %f17984, %f3456, %f17980, 0f3F800000; div.approx.ftz.f32 %f17985, %f17983, %f17984; mul.ftz.f32 %f17986, %f17982, 0f3F000000; mul.ftz.f32 %f17987, %f17982, %f17986; fma.rn.ftz.f32 %f17988, %f17985, %f17985, 0f3F800000; mul.ftz.f32 %f29870, %f17987, %f17988; $L__BB107_2397: mov.f32 %f17989, 0f00000000; max.ftz.f32 %f17990, %f17989, %f3453; sub.ftz.f32 %f17992, %f17978, %f29870; mul.ftz.f32 %f17993, %f17992, %f17990; neg.ftz.f32 %f17994, %f3454; max.ftz.f32 %f17995, %f17989, %f17994; mul.ftz.f32 %f17996, %f17993, %f17995; mul.ftz.f32 %f17997, %f3452, %f3452; sub.ftz.f32 %f17998, %f17978, %f17997; fma.rn.ftz.f32 %f17999, %f3258, %f17997, %f17998; mul.ftz.f32 %f18000, %f17999, 0f40490FDB; mul.ftz.f32 %f18001, %f17999, %f18000; max.ftz.f32 %f18003, %f18001, %f17556; div.approx.ftz.f32 %f18004, %f3258, %f18003; mul.ftz.f32 %f18005, %f17996, %f18004; mul.ftz.f32 %f18006, %f3442, %f3455; mul.ftz.f32 %f18007, %f3455, %f18006; div.approx.ftz.f32 %f29872, %f18005, %f18007; bra.uni $L__BB107_2408; $L__BB107_2367: setp.lt.ftz.f32 %p2128, %f29917, 0fBF7FF972; @%p2128 bra $L__BB107_2374; bra.uni $L__BB107_2368; $L__BB107_2374: mul.ftz.f32 %f29863, %f29863, %f3395; cvt.ftz.sat.f32.f32 %f17832, %f29863; fma.rn.ftz.f32 %f29864, %f17832, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p2138, %f29917, 0f3727C5AC; mov.pred %p3922, 0; mov.f32 %f29862, 0f00000000; @%p2138 bra $L__BB107_2376; lg2.approx.ftz.f32 %f17833, %f29863; mul.ftz.f32 %f17834, %f29921, %f17833; ex2.approx.ftz.f32 %f29862, %f17834; bra.uni $L__BB107_2376; $L__BB107_2368: abs.ftz.f32 %f17816, %f29917; setp.ltu.ftz.f32 %p2130, %f17816, 0f38D1B717; mov.pred %p2129, 0; mov.pred %p3922, %p2129; @%p2130 bra $L__BB107_2376; mov.f32 %f29861, 0f3F800000; sub.ftz.f32 %f17818, %f29861, %f29862; setp.lt.ftz.f32 %p2132, %f17818, %f3395; mov.pred %p3922, %p2126; @%p2132 bra $L__BB107_2376; setp.ge.ftz.f32 %p2133, %f29921, 0f00000000; @%p2133 bra $L__BB107_2372; mov.f32 %f17820, 0f3F800000; sub.ftz.f32 %f17821, %f17820, %f3395; lg2.approx.ftz.f32 %f17822, %f17821; mov.f32 %f17823, 0fBF800000; div.approx.ftz.f32 %f17824, %f17823, %f29921; mul.ftz.f32 %f17825, %f17824, %f17822; ex2.approx.ftz.f32 %f17826, %f17825; mul.ftz.f32 %f29861, %f29863, %f17826; $L__BB107_2372: cvt.ftz.sat.f32.f32 %f17828, %f29861; fma.rn.ftz.f32 %f29864, %f17828, 0f40000000, 0fBF800000; setp.lt.ftz.f32 %p2135, %f29917, 0f3727C5AC; mov.f32 %f29862, 0f00000000; mov.f32 %f29863, %f29861; mov.pred %p3922, %p2129; @%p2135 bra $L__BB107_2376; lg2.approx.ftz.f32 %f17829, %f29861; mul.ftz.f32 %f17830, %f29921, %f17829; ex2.approx.ftz.f32 %f29862, %f17830; mov.f32 %f29863, %f29861; mov.pred %p3922, %p2129; bra.uni $L__BB107_2376; $L__BB107_2379: add.ftz.f32 %f17846, %f3413, 0f3F800000; mul.ftz.f32 %f3414, %f17846, 0f3F000000; setp.lt.ftz.f32 %p2142, %f3414, 0f38D1B717; mov.f32 %f29865, 0f00000000; mov.f32 %f29866, %f29865; @%p2142 bra $L__BB107_2384; add.ftz.f32 %f17849, %f3407, %f3407; mul.ftz.f32 %f17850, %f17849, %f3414; div.approx.ftz.f32 %f17851, %f17850, %f3413; add.ftz.f32 %f3415, %f17851, 0fBF800000; mul.ftz.f32 %f3416, %f3415, %f3415; add.ftz.f32 %f3417, %f3416, 0fBF800000; abs.ftz.f32 %f17852, %f3417; setp.lt.ftz.f32 %p2143, %f17852, 0f33D6BF95; mov.f32 %f29866, %f29865; @%p2143 bra $L__BB107_2384; mul.ftz.f32 %f17853, %f3413, %f3413; mov.f32 %f17854, 0f3F800000; sub.ftz.f32 %f17855, %f17854, %f17853; mov.f32 %f17856, 0f00000000; max.ftz.f32 %f17857, %f17855, %f17856; sqrt.approx.ftz.f32 %f17858, %f17857; div.approx.ftz.f32 %f17859, %f17858, %f3413; rcp.approx.ftz.f32 %f17860, %f3417; mul.ftz.f32 %f17861, %f17859, %f17859; mul.ftz.f32 %f17862, %f17861, %f17860; mul.ftz.f32 %f17863, %f17860, %f17862; sub.ftz.f32 %f17864, %f3416, %f17861; mul.ftz.f32 %f17865, %f17864, %f17860; sub.ftz.f32 %f17866, %f17863, %f17865; max.ftz.f32 %f17867, %f17866, %f17856; sqrt.approx.ftz.f32 %f17868, %f17867; mul.ftz.f32 %f17869, %f17859, %f17860; add.ftz.f32 %f17870, %f17869, %f17868; rcp.approx.ftz.f32 %f17871, %f17859; setp.gt.ftz.f32 %p2144, %f17870, %f17871; setp.lt.ftz.f32 %p2145, %f3415, 0f00000000; or.pred %p2146, %p2145, %p2144; sub.ftz.f32 %f17872, %f17869, %f17868; mov.f32 %f17873, 0f3F000000; sub.ftz.f32 %f17874, %f17873, %f3406; add.ftz.f32 %f17875, %f3406, 0fBF000000; setp.ge.ftz.f32 %p2147, %f3406, 0f3F000000; selp.f32 %f17876, %f17875, %f17874, %p2147; add.ftz.f32 %f17877, %f17876, %f17876; fma.rn.ftz.f32 %f17878, %f17877, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f17879, %f17877, %f17878, 0f3EED4413; mul.ftz.f32 %f17880, %f17877, %f17879; fma.rn.ftz.f32 %f17881, %f17877, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f17882, %f17877, %f17881, 0fBF800000; fma.rn.ftz.f32 %f17883, %f17877, %f17882, 0f3F191676; div.approx.ftz.f32 %f17884, %f17880, %f17883; selp.f32 %f29865, %f17872, %f17870, %p2146; fma.rn.ftz.f32 %f17885, %f29865, %f29865, 0f3F800000; sqrt.approx.ftz.f32 %f17886, %f17885; mul.ftz.f32 %f29866, %f17884, %f17886; setp.ltu.ftz.f32 %p2148, %f3406, 0f3F000000; @%p2148 bra $L__BB107_2382; bra.uni $L__BB107_2384; $L__BB107_2382: neg.ftz.f32 %f29866, %f29866; bra.uni $L__BB107_2384; $L__BB107_2398: @%p2127 bra $L__BB107_2408; mul.ftz.f32 %f18011, %f29917, %f29921; setp.lt.ftz.f32 %p2164, %f29917, 0fBF7FF972; selp.f32 %f3471, 0f3F800000, %f18011, %p2164; @%p2093 bra $L__BB107_2404; bra.uni $L__BB107_2400; $L__BB107_2404: sub.ftz.f32 %f18067, %f3370, %f29915; sub.ftz.f32 %f18068, %f3371, %f29916; mul.ftz.f32 %f18069, %f18068, %f18068; fma.rn.ftz.f32 %f18070, %f18067, %f18067, %f18069; sub.ftz.f32 %f18071, %f3372, %f29917; fma.rn.ftz.f32 %f18072, %f18071, %f18071, %f18070; sqrt.approx.ftz.f32 %f18073, %f18072; div.approx.ftz.f32 %f3481, %f18067, %f18073; div.approx.ftz.f32 %f3482, %f18068, %f18073; div.approx.ftz.f32 %f3483, %f18071, %f18073; setp.lt.ftz.f32 %p2168, %f3483, 0f00000000; @%p2168 bra $L__BB107_2408; mul.ftz.f32 %f18075, %f3482, %f3409; fma.rn.ftz.f32 %f18076, %f3481, %f3408, %f18075; fma.rn.ftz.f32 %f3484, %f3483, %f3412, %f18076; abs.ftz.f32 %f3485, %f3484; fma.rn.ftz.f32 %f3486, %f3485, %f3485, %f3380; setp.leu.ftz.f32 %p2169, %f3486, 0f00000000; mov.f32 %f18074, 0f3F800000; mov.f32 %f29874, %f18074; @%p2169 bra $L__BB107_2407; sqrt.approx.ftz.f32 %f18077, %f3486; sub.ftz.f32 %f18078, %f18077, %f3485; add.ftz.f32 %f18079, %f3485, %f18077; div.approx.ftz.f32 %f18080, %f18078, %f18079; fma.rn.ftz.f32 %f18081, %f3485, %f18079, 0fBF800000; fma.rn.ftz.f32 %f18082, %f3485, %f18078, 0f3F800000; div.approx.ftz.f32 %f18083, %f18081, %f18082; mul.ftz.f32 %f18084, %f18080, 0f3F000000; mul.ftz.f32 %f18085, %f18080, %f18084; fma.rn.ftz.f32 %f18086, %f18083, %f18083, 0f3F800000; mul.ftz.f32 %f29874, %f18085, %f18086; $L__BB107_2407: mov.f32 %f18087, 0f00000000; max.ftz.f32 %f18088, %f18087, %f3484; mul.ftz.f32 %f18089, %f29874, %f18088; mul.ftz.f32 %f18090, %f3483, %f3483; sub.ftz.f32 %f18092, %f18074, %f18090; fma.rn.ftz.f32 %f18093, %f3258, %f18090, %f18092; mul.ftz.f32 %f18094, %f18093, 0f40490FDB; mul.ftz.f32 %f18095, %f18093, %f18094; max.ftz.f32 %f18097, %f18095, %f17556; div.approx.ftz.f32 %f18098, %f3258, %f18097; mul.ftz.f32 %f18099, %f18089, %f18098; mul.ftz.f32 %f18100, %f18099, 0f3E800000; mul.ftz.f32 %f18101, %f3471, %f3484; div.approx.ftz.f32 %f29872, %f18100, %f18101; bra.uni $L__BB107_2408; $L__BB107_2390: sub.ftz.f32 %f17921, %f29824, %f29915; sub.ftz.f32 %f17922, %f29825, %f29916; mul.ftz.f32 %f17923, %f17922, %f17922; fma.rn.ftz.f32 %f17924, %f17921, %f17921, %f17923; sub.ftz.f32 %f17925, %f29826, %f29917; fma.rn.ftz.f32 %f17926, %f17925, %f17925, %f17924; sqrt.approx.ftz.f32 %f17927, %f17926; div.approx.ftz.f32 %f3443, %f17921, %f17927; div.approx.ftz.f32 %f3444, %f17922, %f17927; div.approx.ftz.f32 %f3445, %f17925, %f17927; setp.lt.ftz.f32 %p2157, %f3445, 0f00000000; @%p2157 bra $L__BB107_2408; mul.ftz.f32 %f17929, %f3444, %f3409; fma.rn.ftz.f32 %f17930, %f3443, %f3408, %f17929; fma.rn.ftz.f32 %f3446, %f3445, %f3412, %f17930; abs.ftz.f32 %f3447, %f3446; fma.rn.ftz.f32 %f3448, %f3447, %f3447, %f3262; setp.leu.ftz.f32 %p2158, %f3448, 0f00000000; mov.f32 %f17928, 0f3F800000; mov.f32 %f29869, %f17928; @%p2158 bra $L__BB107_2393; sqrt.approx.ftz.f32 %f17931, %f3448; sub.ftz.f32 %f17932, %f17931, %f3447; add.ftz.f32 %f17933, %f3447, %f17931; div.approx.ftz.f32 %f17934, %f17932, %f17933; fma.rn.ftz.f32 %f17935, %f3447, %f17933, 0fBF800000; fma.rn.ftz.f32 %f17936, %f3447, %f17932, 0f3F800000; div.approx.ftz.f32 %f17937, %f17935, %f17936; mul.ftz.f32 %f17938, %f17934, 0f3F000000; mul.ftz.f32 %f17939, %f17934, %f17938; fma.rn.ftz.f32 %f17940, %f17937, %f17937, 0f3F800000; mul.ftz.f32 %f29869, %f17939, %f17940; $L__BB107_2393: mov.f32 %f17941, 0f00000000; max.ftz.f32 %f17942, %f17941, %f3446; mul.ftz.f32 %f17943, %f29869, %f17942; mul.ftz.f32 %f17944, %f3445, %f3445; sub.ftz.f32 %f17946, %f17928, %f17944; fma.rn.ftz.f32 %f17947, %f3258, %f17944, %f17946; mul.ftz.f32 %f17948, %f17947, 0f40490FDB; mul.ftz.f32 %f17949, %f17947, %f17948; max.ftz.f32 %f17951, %f17949, %f17556; div.approx.ftz.f32 %f17952, %f3258, %f17951; mul.ftz.f32 %f17953, %f17943, %f17952; mul.ftz.f32 %f17954, %f17953, 0f3E800000; mul.ftz.f32 %f17955, %f3442, %f3446; div.approx.ftz.f32 %f29872, %f17954, %f17955; bra.uni $L__BB107_2408; $L__BB107_2400: sub.ftz.f32 %f18015, %f3374, %f29915; sub.ftz.f32 %f18016, %f3375, %f29916; mul.ftz.f32 %f18017, %f18016, %f18016; fma.rn.ftz.f32 %f18018, %f18015, %f18015, %f18017; sub.ftz.f32 %f18019, %f3376, %f29917; fma.rn.ftz.f32 %f18020, %f18019, %f18019, %f18018; sqrt.approx.ftz.f32 %f18021, %f18020; div.approx.ftz.f32 %f18022, %f18015, %f18021; div.approx.ftz.f32 %f18023, %f18016, %f18021; div.approx.ftz.f32 %f18024, %f18019, %f18021; setp.lt.ftz.f32 %p2165, %f18024, 0f00000000; neg.ftz.f32 %f18025, %f18022; neg.ftz.f32 %f18026, %f18023; neg.ftz.f32 %f18027, %f18024; selp.f32 %f18028, %f18025, %f18022, %p2165; selp.f32 %f18029, %f18026, %f18023, %p2165; selp.f32 %f3472, %f18027, %f18024, %p2165; mul.ftz.f32 %f18030, %f18029, %f3409; fma.rn.ftz.f32 %f18031, %f18028, %f3408, %f18030; fma.rn.ftz.f32 %f3473, %f3472, %f3412, %f18031; mul.ftz.f32 %f18032, %f18029, %f3371; fma.rn.ftz.f32 %f18033, %f18028, %f3370, %f18032; fma.rn.ftz.f32 %f3474, %f3472, %f3372, %f18033; setp.lt.ftz.f32 %p2166, %f3473, 0f00000000; @%p2166 bra $L__BB107_2408; fma.rn.ftz.f32 %f3475, %f3373, %f3474, %f3473; abs.ftz.f32 %f3476, %f3473; fma.rn.ftz.f32 %f3477, %f3476, %f3476, %f3380; setp.leu.ftz.f32 %p2167, %f3477, 0f00000000; mov.f32 %f18034, 0f3F800000; mov.f32 %f29873, %f18034; @%p2167 bra $L__BB107_2403; sqrt.approx.ftz.f32 %f18035, %f3477; sub.ftz.f32 %f18036, %f18035, %f3476; add.ftz.f32 %f18037, %f3476, %f18035; div.approx.ftz.f32 %f18038, %f18036, %f18037; fma.rn.ftz.f32 %f18039, %f3476, %f18037, 0fBF800000; fma.rn.ftz.f32 %f18040, %f3476, %f18036, 0f3F800000; div.approx.ftz.f32 %f18041, %f18039, %f18040; mul.ftz.f32 %f18042, %f18038, 0f3F000000; mul.ftz.f32 %f18043, %f18038, %f18042; fma.rn.ftz.f32 %f18044, %f18041, %f18041, 0f3F800000; mul.ftz.f32 %f29873, %f18043, %f18044; $L__BB107_2403: mov.f32 %f18045, 0f00000000; max.ftz.f32 %f18046, %f18045, %f3473; sub.ftz.f32 %f18048, %f18034, %f29873; mul.ftz.f32 %f18049, %f18048, %f18046; neg.ftz.f32 %f18050, %f3474; max.ftz.f32 %f18051, %f18045, %f18050; mul.ftz.f32 %f18052, %f18049, %f18051; mul.ftz.f32 %f18053, %f3472, %f3472; sub.ftz.f32 %f18054, %f18034, %f18053; fma.rn.ftz.f32 %f18055, %f3258, %f18053, %f18054; mul.ftz.f32 %f18056, %f18055, 0f40490FDB; mul.ftz.f32 %f18057, %f18055, %f18056; max.ftz.f32 %f18059, %f18057, %f17556; div.approx.ftz.f32 %f18060, %f3258, %f18059; mul.ftz.f32 %f18061, %f18052, %f18060; mul.ftz.f32 %f18062, %f3471, %f3475; mul.ftz.f32 %f18063, %f3475, %f18062; div.approx.ftz.f32 %f29872, %f18061, %f18063; $L__BB107_2408: setp.gt.ftz.f32 %p2170, %f3381, 0f3F7FF972; setp.lt.ftz.f32 %p2171, %f3381, 0f3727C5AC; or.pred %p2172, %p2170, %p2171; selp.f32 %f29880, 0f3F800000, 0f00000000, %p2170; @%p2172 bra $L__BB107_2410; setp.eq.s16 %p2173, %rs527, %rs243; neg.ftz.f32 %f18102, %f29864; selp.f32 %f18103, %f29864, %f18102, %p2173; add.ftz.f32 %f18104, %f18103, 0f3F800000; mul.ftz.f32 %f18105, %f18104, 0f3F000000; cvt.ftz.sat.f32.f32 %f18106, %f18105; lg2.approx.ftz.f32 %f18107, %f18106; mul.ftz.f32 %f18108, %f29834, %f18107; ex2.approx.ftz.f32 %f29880, %f18108; $L__BB107_2410: mul.ftz.f32 %f18109, %f29918, %f29872; mul.ftz.f32 %f30104, %f18109, %f29880; mul.ftz.f32 %f18110, %f29919, %f29872; mul.ftz.f32 %f30105, %f18110, %f29880; mul.ftz.f32 %f18111, %f29920, %f29872; mul.ftz.f32 %f30106, %f18111, %f29880; $L__BB107_2411: setp.eq.s32 %p2174, %r5781, 0; @%p2174 bra $L__BB107_2436; and.b16 %rs529, %rs720, 255; mov.f32 %f29887, 0f00000000; setp.eq.s16 %p2175, %rs529, 0; @%p2175 bra $L__BB107_2423; @%p2127 bra $L__BB107_2433; mul.ftz.f32 %f18115, %f29917, %f29921; setp.lt.ftz.f32 %p2178, %f29917, 0fBF7FF972; selp.f32 %f3512, 0f3F800000, %f18115, %p2178; @%p2093 bra $L__BB107_2419; bra.uni $L__BB107_2415; $L__BB107_2419: sub.ftz.f32 %f18157, %f3377, %f29915; sub.ftz.f32 %f18158, %f3378, %f29916; mul.ftz.f32 %f18159, %f18158, %f18158; fma.rn.ftz.f32 %f18160, %f18157, %f18157, %f18159; sub.ftz.f32 %f18161, %f3379, %f29917; fma.rn.ftz.f32 %f18162, %f18161, %f18161, %f18160; sqrt.approx.ftz.f32 %f18163, %f18162; div.approx.ftz.f32 %f18164, %f18157, %f18163; div.approx.ftz.f32 %f18165, %f18158, %f18163; div.approx.ftz.f32 %f18166, %f18161, %f18163; setp.lt.ftz.f32 %p2181, %f18166, 0f00000000; neg.ftz.f32 %f18167, %f18164; neg.ftz.f32 %f18168, %f18165; neg.ftz.f32 %f18169, %f18166; selp.f32 %f18170, %f18167, %f18164, %p2181; selp.f32 %f18171, %f18168, %f18165, %p2181; selp.f32 %f3522, %f18169, %f18166, %p2181; mul.ftz.f32 %f18172, %f18171, %f3409; fma.rn.ftz.f32 %f18173, %f18170, %f3408, %f18172; fma.rn.ftz.f32 %f3523, %f3522, %f3412, %f18173; mul.ftz.f32 %f18174, %f29825, %f18171; fma.rn.ftz.f32 %f18175, %f29824, %f18170, %f18174; fma.rn.ftz.f32 %f3524, %f29826, %f3522, %f18175; setp.lt.ftz.f32 %p2182, %f3523, 0f00000000; @%p2182 bra $L__BB107_2433; fma.rn.ftz.f32 %f3525, %f3245, %f3524, %f3523; abs.ftz.f32 %f3526, %f3523; fma.rn.ftz.f32 %f3527, %f3526, %f3526, %f3262; setp.leu.ftz.f32 %p2183, %f3527, 0f00000000; mov.f32 %f18176, 0f3F800000; mov.f32 %f29885, %f18176; @%p2183 bra $L__BB107_2422; sqrt.approx.ftz.f32 %f18177, %f3527; sub.ftz.f32 %f18178, %f18177, %f3526; add.ftz.f32 %f18179, %f3526, %f18177; div.approx.ftz.f32 %f18180, %f18178, %f18179; fma.rn.ftz.f32 %f18181, %f3526, %f18179, 0fBF800000; fma.rn.ftz.f32 %f18182, %f3526, %f18178, 0f3F800000; div.approx.ftz.f32 %f18183, %f18181, %f18182; mul.ftz.f32 %f18184, %f18180, 0f3F000000; mul.ftz.f32 %f18185, %f18180, %f18184; fma.rn.ftz.f32 %f18186, %f18183, %f18183, 0f3F800000; mul.ftz.f32 %f29885, %f18185, %f18186; $L__BB107_2422: mov.f32 %f18187, 0f00000000; max.ftz.f32 %f18188, %f18187, %f3523; sub.ftz.f32 %f18190, %f18176, %f29885; mul.ftz.f32 %f18191, %f18190, %f18188; neg.ftz.f32 %f18192, %f3524; max.ftz.f32 %f18193, %f18187, %f18192; mul.ftz.f32 %f18194, %f18191, %f18193; mul.ftz.f32 %f18195, %f3522, %f3522; sub.ftz.f32 %f18196, %f18176, %f18195; fma.rn.ftz.f32 %f18197, %f3258, %f18195, %f18196; mul.ftz.f32 %f18198, %f18197, 0f40490FDB; mul.ftz.f32 %f18199, %f18197, %f18198; max.ftz.f32 %f18201, %f18199, %f17556; div.approx.ftz.f32 %f18202, %f3258, %f18201; mul.ftz.f32 %f18203, %f18194, %f18202; mul.ftz.f32 %f18204, %f3512, %f3525; mul.ftz.f32 %f18205, %f3525, %f18204; div.approx.ftz.f32 %f29887, %f18203, %f18205; bra.uni $L__BB107_2433; $L__BB107_2423: @%p2127 bra $L__BB107_2433; mul.ftz.f32 %f18209, %f29917, %f29921; setp.lt.ftz.f32 %p2186, %f29917, 0fBF7FF972; selp.f32 %f3541, 0f3F800000, %f18209, %p2186; @%p2093 bra $L__BB107_2429; bra.uni $L__BB107_2425; $L__BB107_2429: sub.ftz.f32 %f18265, %f3370, %f29915; sub.ftz.f32 %f18266, %f3371, %f29916; mul.ftz.f32 %f18267, %f18266, %f18266; fma.rn.ftz.f32 %f18268, %f18265, %f18265, %f18267; sub.ftz.f32 %f18269, %f3372, %f29917; fma.rn.ftz.f32 %f18270, %f18269, %f18269, %f18268; sqrt.approx.ftz.f32 %f18271, %f18270; div.approx.ftz.f32 %f3551, %f18265, %f18271; div.approx.ftz.f32 %f3552, %f18266, %f18271; div.approx.ftz.f32 %f3553, %f18269, %f18271; setp.lt.ftz.f32 %p2190, %f3553, 0f00000000; @%p2190 bra $L__BB107_2433; mul.ftz.f32 %f18273, %f3552, %f3409; fma.rn.ftz.f32 %f18274, %f3551, %f3408, %f18273; fma.rn.ftz.f32 %f3554, %f3553, %f3412, %f18274; abs.ftz.f32 %f3555, %f3554; fma.rn.ftz.f32 %f3556, %f3555, %f3555, %f3380; setp.leu.ftz.f32 %p2191, %f3556, 0f00000000; mov.f32 %f18272, 0f3F800000; mov.f32 %f29889, %f18272; @%p2191 bra $L__BB107_2432; sqrt.approx.ftz.f32 %f18275, %f3556; sub.ftz.f32 %f18276, %f18275, %f3555; add.ftz.f32 %f18277, %f3555, %f18275; div.approx.ftz.f32 %f18278, %f18276, %f18277; fma.rn.ftz.f32 %f18279, %f3555, %f18277, 0fBF800000; fma.rn.ftz.f32 %f18280, %f3555, %f18276, 0f3F800000; div.approx.ftz.f32 %f18281, %f18279, %f18280; mul.ftz.f32 %f18282, %f18278, 0f3F000000; mul.ftz.f32 %f18283, %f18278, %f18282; fma.rn.ftz.f32 %f18284, %f18281, %f18281, 0f3F800000; mul.ftz.f32 %f29889, %f18283, %f18284; $L__BB107_2432: mov.f32 %f18285, 0f00000000; max.ftz.f32 %f18286, %f18285, %f3554; mul.ftz.f32 %f18287, %f29889, %f18286; mul.ftz.f32 %f18288, %f3553, %f3553; sub.ftz.f32 %f18290, %f18272, %f18288; fma.rn.ftz.f32 %f18291, %f3258, %f18288, %f18290; mul.ftz.f32 %f18292, %f18291, 0f40490FDB; mul.ftz.f32 %f18293, %f18291, %f18292; max.ftz.f32 %f18295, %f18293, %f17556; div.approx.ftz.f32 %f18296, %f3258, %f18295; mul.ftz.f32 %f18297, %f18287, %f18296; mul.ftz.f32 %f18298, %f18297, 0f3E800000; mul.ftz.f32 %f18299, %f3541, %f3554; div.approx.ftz.f32 %f29887, %f18298, %f18299; bra.uni $L__BB107_2433; $L__BB107_2415: sub.ftz.f32 %f18119, %f29824, %f29915; sub.ftz.f32 %f18120, %f29825, %f29916; mul.ftz.f32 %f18121, %f18120, %f18120; fma.rn.ftz.f32 %f18122, %f18119, %f18119, %f18121; sub.ftz.f32 %f18123, %f29826, %f29917; fma.rn.ftz.f32 %f18124, %f18123, %f18123, %f18122; sqrt.approx.ftz.f32 %f18125, %f18124; div.approx.ftz.f32 %f3513, %f18119, %f18125; div.approx.ftz.f32 %f3514, %f18120, %f18125; div.approx.ftz.f32 %f3515, %f18123, %f18125; setp.lt.ftz.f32 %p2179, %f3515, 0f00000000; @%p2179 bra $L__BB107_2433; mul.ftz.f32 %f18127, %f3514, %f3409; fma.rn.ftz.f32 %f18128, %f3513, %f3408, %f18127; fma.rn.ftz.f32 %f3516, %f3515, %f3412, %f18128; abs.ftz.f32 %f3517, %f3516; fma.rn.ftz.f32 %f3518, %f3517, %f3517, %f3262; setp.leu.ftz.f32 %p2180, %f3518, 0f00000000; mov.f32 %f29884, %f17912; @%p2180 bra $L__BB107_2418; sqrt.approx.ftz.f32 %f18129, %f3518; sub.ftz.f32 %f18130, %f18129, %f3517; add.ftz.f32 %f18131, %f3517, %f18129; div.approx.ftz.f32 %f18132, %f18130, %f18131; fma.rn.ftz.f32 %f18133, %f3517, %f18131, 0fBF800000; fma.rn.ftz.f32 %f18134, %f3517, %f18130, 0f3F800000; div.approx.ftz.f32 %f18135, %f18133, %f18134; mul.ftz.f32 %f18136, %f18132, 0f3F000000; mul.ftz.f32 %f18137, %f18132, %f18136; fma.rn.ftz.f32 %f18138, %f18135, %f18135, 0f3F800000; mul.ftz.f32 %f29884, %f18137, %f18138; $L__BB107_2418: mov.f32 %f18139, 0f00000000; max.ftz.f32 %f18140, %f18139, %f3516; mul.ftz.f32 %f18141, %f29884, %f18140; mul.ftz.f32 %f18142, %f3515, %f3515; sub.ftz.f32 %f18144, %f17912, %f18142; fma.rn.ftz.f32 %f18145, %f3258, %f18142, %f18144; mul.ftz.f32 %f18146, %f18145, 0f40490FDB; mul.ftz.f32 %f18147, %f18145, %f18146; max.ftz.f32 %f18149, %f18147, %f17556; div.approx.ftz.f32 %f18150, %f3258, %f18149; mul.ftz.f32 %f18151, %f18141, %f18150; mul.ftz.f32 %f18152, %f18151, 0f3E800000; mul.ftz.f32 %f18153, %f3512, %f3516; div.approx.ftz.f32 %f29887, %f18152, %f18153; bra.uni $L__BB107_2433; $L__BB107_2425: sub.ftz.f32 %f18213, %f3374, %f29915; sub.ftz.f32 %f18214, %f3375, %f29916; mul.ftz.f32 %f18215, %f18214, %f18214; fma.rn.ftz.f32 %f18216, %f18213, %f18213, %f18215; sub.ftz.f32 %f18217, %f3376, %f29917; fma.rn.ftz.f32 %f18218, %f18217, %f18217, %f18216; sqrt.approx.ftz.f32 %f18219, %f18218; div.approx.ftz.f32 %f18220, %f18213, %f18219; div.approx.ftz.f32 %f18221, %f18214, %f18219; div.approx.ftz.f32 %f18222, %f18217, %f18219; setp.lt.ftz.f32 %p2187, %f18222, 0f00000000; neg.ftz.f32 %f18223, %f18220; neg.ftz.f32 %f18224, %f18221; neg.ftz.f32 %f18225, %f18222; selp.f32 %f18226, %f18223, %f18220, %p2187; selp.f32 %f18227, %f18224, %f18221, %p2187; selp.f32 %f3542, %f18225, %f18222, %p2187; mul.ftz.f32 %f18228, %f18227, %f3409; fma.rn.ftz.f32 %f18229, %f18226, %f3408, %f18228; fma.rn.ftz.f32 %f3543, %f3542, %f3412, %f18229; mul.ftz.f32 %f18230, %f18227, %f3371; fma.rn.ftz.f32 %f18231, %f18226, %f3370, %f18230; fma.rn.ftz.f32 %f3544, %f3542, %f3372, %f18231; setp.lt.ftz.f32 %p2188, %f3543, 0f00000000; @%p2188 bra $L__BB107_2433; fma.rn.ftz.f32 %f3545, %f3373, %f3544, %f3543; abs.ftz.f32 %f3546, %f3543; fma.rn.ftz.f32 %f3547, %f3546, %f3546, %f3380; setp.leu.ftz.f32 %p2189, %f3547, 0f00000000; mov.f32 %f18232, 0f3F800000; mov.f32 %f29888, %f18232; @%p2189 bra $L__BB107_2428; sqrt.approx.ftz.f32 %f18233, %f3547; sub.ftz.f32 %f18234, %f18233, %f3546; add.ftz.f32 %f18235, %f3546, %f18233; div.approx.ftz.f32 %f18236, %f18234, %f18235; fma.rn.ftz.f32 %f18237, %f3546, %f18235, 0fBF800000; fma.rn.ftz.f32 %f18238, %f3546, %f18234, 0f3F800000; div.approx.ftz.f32 %f18239, %f18237, %f18238; mul.ftz.f32 %f18240, %f18236, 0f3F000000; mul.ftz.f32 %f18241, %f18236, %f18240; fma.rn.ftz.f32 %f18242, %f18239, %f18239, 0f3F800000; mul.ftz.f32 %f29888, %f18241, %f18242; $L__BB107_2428: mov.f32 %f18243, 0f00000000; max.ftz.f32 %f18244, %f18243, %f3543; sub.ftz.f32 %f18246, %f18232, %f29888; mul.ftz.f32 %f18247, %f18246, %f18244; neg.ftz.f32 %f18248, %f3544; max.ftz.f32 %f18249, %f18243, %f18248; mul.ftz.f32 %f18250, %f18247, %f18249; mul.ftz.f32 %f18251, %f3542, %f3542; sub.ftz.f32 %f18252, %f18232, %f18251; fma.rn.ftz.f32 %f18253, %f3258, %f18251, %f18252; mul.ftz.f32 %f18254, %f18253, 0f40490FDB; mul.ftz.f32 %f18255, %f18253, %f18254; max.ftz.f32 %f18257, %f18255, %f17556; div.approx.ftz.f32 %f18258, %f3258, %f18257; mul.ftz.f32 %f18259, %f18250, %f18258; mul.ftz.f32 %f18260, %f3541, %f3545; mul.ftz.f32 %f18261, %f3545, %f18260; div.approx.ftz.f32 %f29887, %f18259, %f18261; $L__BB107_2433: setp.gt.ftz.f32 %p2192, %f3381, 0f3F7FF972; setp.lt.ftz.f32 %p2193, %f3381, 0f3727C5AC; or.pred %p2194, %p2192, %p2193; selp.f32 %f29895, 0f3F800000, 0f00000000, %p2192; @%p2194 bra $L__BB107_2435; setp.eq.s16 %p2195, %rs529, %rs243; neg.ftz.f32 %f18300, %f29864; selp.f32 %f18301, %f29864, %f18300, %p2195; add.ftz.f32 %f18302, %f18301, 0f3F800000; mul.ftz.f32 %f18303, %f18302, 0f3F000000; cvt.ftz.sat.f32.f32 %f18304, %f18303; lg2.approx.ftz.f32 %f18305, %f18304; mul.ftz.f32 %f18306, %f29834, %f18305; ex2.approx.ftz.f32 %f29895, %f18306; $L__BB107_2435: mul.ftz.f32 %f18307, %f29918, %f29887; fma.rn.ftz.f32 %f30104, %f18307, %f29895, %f30104; mul.ftz.f32 %f18308, %f29919, %f29887; fma.rn.ftz.f32 %f30105, %f18308, %f29895, %f30105; mul.ftz.f32 %f18309, %f29920, %f29887; fma.rn.ftz.f32 %f30106, %f18309, %f29895, %f30106; $L__BB107_2436: add.s32 %r5781, %r5781, 1; setp.gt.u32 %p2196, %r5781, 9; @%p2196 bra $L__BB107_2456; mad.lo.s32 %r5782, %r5782, 1103515245, 12345; and.b16 %rs531, %rs720, 255; setp.eq.s16 %p2197, %rs531, 0; selp.f32 %f3582, %f3373, %f3245, %p2197; mul.ftz.f32 %f18311, %f3440, %f3409; fma.rn.ftz.f32 %f18312, %f3439, %f3408, %f18311; fma.rn.ftz.f32 %f3583, %f3441, %f3412, %f18312; abs.ftz.f32 %f3584, %f3583; fma.rn.ftz.f32 %f18313, %f3582, %f3582, 0fBF800000; mov.f32 %f29899, 0f3F800000; mul.ftz.f32 %f3585, %f3584, %f3584; add.ftz.f32 %f3586, %f18313, %f3585; setp.leu.ftz.f32 %p2198, %f3586, 0f00000000; @%p2198 bra $L__BB107_2439; sqrt.approx.ftz.f32 %f18314, %f3586; sub.ftz.f32 %f18315, %f18314, %f3584; add.ftz.f32 %f18316, %f3584, %f18314; div.approx.ftz.f32 %f18317, %f18315, %f18316; fma.rn.ftz.f32 %f18318, %f3584, %f18316, 0fBF800000; fma.rn.ftz.f32 %f18319, %f3584, %f18315, 0f3F800000; div.approx.ftz.f32 %f18320, %f18318, %f18319; mul.ftz.f32 %f18321, %f18317, 0f3F000000; mul.ftz.f32 %f18322, %f18317, %f18321; fma.rn.ftz.f32 %f18323, %f18320, %f18320, 0f3F800000; mul.ftz.f32 %f29899, %f18322, %f18323; $L__BB107_2439: cvt.rn.f32.u32 %f18324, %r5782; mul.ftz.f32 %f18325, %f18324, 0f2F800000; setp.gt.ftz.f32 %p2199, %f29899, %f18325; @%p2199 bra $L__BB107_2441; bra.uni $L__BB107_2440; $L__BB107_2441: add.ftz.f32 %f18351, %f3439, %f3439; add.ftz.f32 %f18352, %f3440, %f3440; add.ftz.f32 %f18353, %f3441, %f3441; fma.rn.ftz.f32 %f29917, %f18353, %f3583, %f29917; fma.rn.ftz.f32 %f29916, %f18352, %f3583, %f29916; fma.rn.ftz.f32 %f29915, %f18351, %f3583, %f29915; mov.pred %p3923, 0; bra.uni $L__BB107_2442; $L__BB107_2440: rcp.approx.ftz.f32 %f18326, %f3582; mov.f32 %f18327, 0f3F800000; mul.ftz.f32 %f18328, %f3583, %f3583; sub.ftz.f32 %f18329, %f18327, %f18328; mul.ftz.f32 %f18330, %f18329, %f18326; mul.ftz.f32 %f18331, %f18326, %f18330; sub.ftz.f32 %f18332, %f18327, %f18331; mov.f32 %f18333, 0f00000000; max.ftz.f32 %f18334, %f18332, %f18333; sqrt.approx.ftz.f32 %f18335, %f18334; mul.ftz.f32 %f18336, %f3583, %f18326; sub.ftz.f32 %f18337, %f18336, %f18335; mul.ftz.f32 %f18338, %f3439, %f18337; mul.ftz.f32 %f18339, %f3440, %f18337; mul.ftz.f32 %f18340, %f3441, %f18337; mul.ftz.f32 %f18341, %f18326, %f3408; mul.ftz.f32 %f18342, %f18326, %f3409; mul.ftz.f32 %f18343, %f18326, %f3412; sub.ftz.f32 %f18344, %f18338, %f18341; sub.ftz.f32 %f18345, %f18339, %f18342; sub.ftz.f32 %f18346, %f18340, %f18343; mul.ftz.f32 %f18347, %f18345, %f18345; fma.rn.ftz.f32 %f18348, %f18344, %f18344, %f18347; fma.rn.ftz.f32 %f18349, %f18346, %f18346, %f18348; sqrt.approx.ftz.f32 %f18350, %f18349; div.approx.ftz.f32 %f29917, %f18346, %f18350; div.approx.ftz.f32 %f29916, %f18345, %f18350; div.approx.ftz.f32 %f29915, %f18344, %f18350; mov.pred %p3923, -1; $L__BB107_2442: not.pred %p2202, %p3923; @%p2202 bra $L__BB107_2444; selp.u16 %rs720, 1, 0, %p2197; neg.ftz.f32 %f29915, %f29915; neg.ftz.f32 %f29916, %f29916; neg.ftz.f32 %f29917, %f29917; neg.ftz.f32 %f29864, %f29864; $L__BB107_2444: or.pred %p2205, %p9, %p2202; @%p2205 bra $L__BB107_2446; bra.uni $L__BB107_2445; $L__BB107_2446: or.pred %p2208, %p2151, %p2174; @%p2208 bra $L__BB107_2450; add.ftz.f32 %f3612, %f3262, %f3585; setp.leu.ftz.f32 %p2209, %f3612, 0f00000000; mov.f32 %f29907, %f3352; @%p2209 bra $L__BB107_2449; sqrt.approx.ftz.f32 %f18354, %f3612; sub.ftz.f32 %f18355, %f18354, %f3584; add.ftz.f32 %f18356, %f3584, %f18354; div.approx.ftz.f32 %f18357, %f18355, %f18356; fma.rn.ftz.f32 %f18358, %f3584, %f18356, 0fBF800000; fma.rn.ftz.f32 %f18359, %f3584, %f18355, 0f3F800000; div.approx.ftz.f32 %f18360, %f18358, %f18359; mul.ftz.f32 %f18361, %f18357, 0f3F000000; mul.ftz.f32 %f18362, %f18357, %f18361; fma.rn.ftz.f32 %f18363, %f18360, %f18360, 0f3F800000; mul.ftz.f32 %f18364, %f18362, %f18363; sub.ftz.f32 %f29907, %f18364, %f29840; $L__BB107_2449: mul.ftz.f32 %f18365, %f3353, %f29907; mov.f32 %f18366, 0f3F800000; sub.ftz.f32 %f18367, %f18366, %f18365; fma.rn.ftz.f32 %f18368, %f29821, %f18367, %f18365; fma.rn.ftz.f32 %f18369, %f29822, %f18367, %f18365; fma.rn.ftz.f32 %f18370, %f29823, %f18367, %f18365; mul.ftz.f32 %f29918, %f29918, %f18368; mul.ftz.f32 %f29919, %f29919, %f18369; mul.ftz.f32 %f29920, %f29920, %f18370; bra.uni $L__BB107_2450; $L__BB107_2445: mul.ftz.f32 %f29918, %f3274, %f29918; mul.ftz.f32 %f29919, %f3275, %f29919; mul.ftz.f32 %f29920, %f3276, %f29920; $L__BB107_2450: setp.gt.ftz.f32 %p2210, %f29917, 0f3F7FF972; mov.f32 %f29921, 0f00000000; @%p2210 bra $L__BB107_2453; setp.lt.ftz.f32 %p2211, %f29917, 0fBF7FF972; mov.f32 %f29921, 0fBF7FF972; @%p2211 bra $L__BB107_2453; mul.ftz.f32 %f18373, %f29917, %f29917; max.ftz.f32 %f18375, %f18373, %f17556; rcp.approx.ftz.f32 %f18376, %f18375; mul.ftz.f32 %f18377, %f17470, %f29915; mul.ftz.f32 %f18378, %f17471, %f29916; mul.ftz.f32 %f18379, %f18378, %f18378; fma.rn.ftz.f32 %f18380, %f18377, %f18377, %f18379; fma.rn.ftz.f32 %f18381, %f18380, %f18376, 0f3F800000; sqrt.approx.ftz.f32 %f18382, %f18381; neg.ftz.f32 %f18383, %f18382; setp.le.ftz.f32 %p2212, %f29917, 0f00000000; selp.f32 %f18384, %f18383, %f18382, %p2212; add.ftz.f32 %f18385, %f18384, 0fBF800000; mul.ftz.f32 %f29921, %f18385, 0f3F000000; $L__BB107_2453: mul.ftz.f32 %f18387, %f3274, %f29918; selp.f32 %f29918, %f18387, %f29918, %p9; mul.ftz.f32 %f18388, %f3275, %f29919; selp.f32 %f29919, %f18388, %f29919, %p9; mul.ftz.f32 %f18389, %f3276, %f29920; selp.f32 %f29920, %f18389, %f29920, %p9; add.ftz.f32 %f18390, %f29864, 0f3F800000; mov.f32 %f29862, 0f3F800000; mul.ftz.f32 %f18391, %f18390, 0f3F000000; cvt.ftz.sat.f32.f32 %f29863, %f18391; @%p2210 bra $L__BB107_2456; setp.lt.ftz.f32 %p2214, %f29917, 0f3727C5AC; mov.f32 %f29862, 0f00000000; @%p2214 bra $L__BB107_2456; lg2.approx.ftz.f32 %f18393, %f29863; mul.ftz.f32 %f18394, %f29921, %f18393; ex2.approx.ftz.f32 %f29862, %f18394; $L__BB107_2456: setp.lt.u32 %p2215, %r5781, 10; @%p2215 bra $L__BB107_2366; $L__BB107_2457: st.local.u32 [%rd5+168], %r5782; setp.eq.s16 %p2216, %rs717, 0; @%p2216 bra $L__BB107_2774; div.approx.ftz.f32 %f18395, %f29829, %f29826; abs.ftz.f32 %f18396, %f18395; mul.ftz.f32 %f30104, %f30104, %f18396; mul.ftz.f32 %f30105, %f30105, %f18396; mul.ftz.f32 %f30106, %f30106, %f18396; bra.uni $L__BB107_2774; $L__BB107_2768: and.b32 %r4148, %r988, 4; setp.eq.s32 %p2530, %r4148, 0; add.ftz.f32 %f20807, %f4559, 0f00000000; selp.f32 %f30097, 0f00000000, %f20807, %p2530; $L__BB107_2769: and.b32 %r4149, %r988, 9; setp.eq.s32 %p2531, %r4149, 0; @%p2531 bra $L__BB107_2771; mul.ftz.f32 %f20808, %f30259, %f20783; fma.rn.ftz.f32 %f20809, %f30260, %f4549, %f20808; fma.rn.ftz.f32 %f20810, %f30258, %f20784, %f20809; add.ftz.f32 %f20811, %f20810, 0f3F800000; ld.local.f32 %f20812, [%rd564+20]; mul.ftz.f32 %f20813, %f20811, %f20812; add.ftz.f32 %f20814, %f20813, 0fBF800000; mul.ftz.f32 %f20815, %f4557, %f4558; add.ftz.f32 %f20816, %f4557, %f4558; fma.rn.ftz.f32 %f20817, %f20815, %f20814, %f20816; fma.rn.ftz.f32 %f30097, %f20813, %f20817, %f30097; $L__BB107_2771: mul.ftz.f32 %f20818, %f4555, 0f3EA2F983; mul.ftz.f32 %f30104, %f20818, %f30097; $L__BB107_2772: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2287: ld.local.v2.f32 {%f20023, %f20024}, [%rd564+44]; ld.local.f32 %f20027, [%rd564+52]; mul.ftz.f32 %f20028, %f4343, %f20027; mul.ftz.f32 %f20029, %f4344, %f20024; sub.ftz.f32 %f20030, %f20028, %f20029; mul.ftz.f32 %f20031, %f4344, %f20023; mul.ftz.f32 %f20032, %f4342, %f20027; sub.ftz.f32 %f20033, %f20031, %f20032; mul.ftz.f32 %f20034, %f4342, %f20024; mul.ftz.f32 %f20035, %f4343, %f20023; sub.ftz.f32 %f20036, %f20034, %f20035; mul.ftz.f32 %f20037, %f20033, %f20033; fma.rn.ftz.f32 %f20038, %f20030, %f20030, %f20037; fma.rn.ftz.f32 %f20039, %f20036, %f20036, %f20038; sqrt.approx.ftz.f32 %f20040, %f20039; div.approx.ftz.f32 %f4355, %f20030, %f20040; div.approx.ftz.f32 %f4356, %f20033, %f20040; div.approx.ftz.f32 %f4357, %f20036, %f20040; mul.ftz.f32 %f20041, %f4344, %f4356; mul.ftz.f32 %f20042, %f4343, %f4357; sub.ftz.f32 %f4358, %f20041, %f20042; mul.ftz.f32 %f20043, %f4342, %f4357; mul.ftz.f32 %f20044, %f4344, %f4355; sub.ftz.f32 %f4359, %f20043, %f20044; mul.ftz.f32 %f20045, %f4343, %f4355; mul.ftz.f32 %f20046, %f4342, %f4356; sub.ftz.f32 %f4360, %f20045, %f20046; mul.ftz.f32 %f20047, %f4348, %f4359; fma.rn.ftz.f32 %f20048, %f4347, %f4358, %f20047; fma.rn.ftz.f32 %f20049, %f4349, %f4360, %f20048; mul.ftz.f32 %f20050, %f4348, %f4356; fma.rn.ftz.f32 %f20051, %f4347, %f4355, %f20050; fma.rn.ftz.f32 %f20052, %f4349, %f4357, %f20051; neg.ftz.f32 %f20053, %f20049; mul.ftz.f32 %f20054, %f19996, %f4350; div.approx.ftz.f32 %f20055, %f20053, %f20054; neg.ftz.f32 %f20056, %f20052; mul.ftz.f32 %f20057, %f19997, %f4350; div.approx.ftz.f32 %f20058, %f20056, %f20057; mul.ftz.f32 %f20059, %f20055, %f20055; neg.ftz.f32 %f20060, %f20059; mul.ftz.f32 %f20061, %f20058, %f20058; sub.ftz.f32 %f20062, %f20060, %f20061; mul.ftz.f32 %f20063, %f20062, 0f3FB8AA3B; ex2.approx.ftz.f32 %f20064, %f20063; div.approx.ftz.f32 %f30063, %f20064, %f4352; mul.ftz.f32 %f20065, %f19994, %f4359; fma.rn.ftz.f32 %f20066, %f4336, %f4358, %f20065; fma.rn.ftz.f32 %f20067, %f19995, %f4360, %f20066; mul.ftz.f32 %f20068, %f19994, %f4356; fma.rn.ftz.f32 %f20069, %f4336, %f4355, %f20068; fma.rn.ftz.f32 %f20070, %f19995, %f4357, %f20069; mul.ftz.f32 %f20071, %f20070, %f20070; mul.ftz.f32 %f20072, %f20067, %f20067; mul.ftz.f32 %f4362, %f19996, %f19996; mul.ftz.f32 %f20073, %f4362, %f20072; mul.ftz.f32 %f4363, %f19997, %f19997; fma.rn.ftz.f32 %f20074, %f4363, %f20071, %f20073; add.ftz.f32 %f20075, %f20071, %f20072; div.approx.ftz.f32 %f20076, %f20074, %f20075; mul.ftz.f32 %f20077, %f4354, %f20076; div.approx.ftz.f32 %f20078, %f20077, %f4353; mov.f32 %f20079, 0f00000000; max.ftz.f32 %f20080, %f20078, %f20079; sqrt.approx.ftz.f32 %f4364, %f20080; setp.lt.ftz.f32 %p2465, %f4364, 0f3F200000; mov.f32 %f30064, %f30065; @%p2465 bra $L__BB107_2289; rcp.approx.ftz.f32 %f20081, %f4364; fma.rn.ftz.f32 %f20082, %f20081, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20083, %f20081, %f20082; fma.rn.ftz.f32 %f20084, %f20081, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20085, %f20081, %f20084, 0f3F800000; div.approx.ftz.f32 %f30064, %f20083, %f20085; $L__BB107_2289: mul.ftz.f32 %f20087, %f30259, %f4359; fma.rn.ftz.f32 %f20088, %f30260, %f4358, %f20087; fma.rn.ftz.f32 %f20089, %f30258, %f4360, %f20088; mul.ftz.f32 %f20090, %f30259, %f4356; fma.rn.ftz.f32 %f20091, %f30260, %f4355, %f20090; fma.rn.ftz.f32 %f20092, %f30258, %f4357, %f20091; mul.ftz.f32 %f20093, %f20092, %f20092; mul.ftz.f32 %f20094, %f20089, %f20089; mul.ftz.f32 %f20095, %f4362, %f20094; fma.rn.ftz.f32 %f20096, %f4363, %f20093, %f20095; add.ftz.f32 %f20097, %f20093, %f20094; div.approx.ftz.f32 %f20098, %f20096, %f20097; mul.ftz.f32 %f20099, %f4346, %f4346; sub.ftz.f32 %f20100, %f30065, %f20099; mul.ftz.f32 %f20101, %f20100, %f20098; div.approx.ftz.f32 %f20102, %f20101, %f20099; max.ftz.f32 %f20104, %f20102, %f20079; sqrt.approx.ftz.f32 %f4367, %f20104; setp.lt.ftz.f32 %p2466, %f4367, 0f3F200000; @%p2466 bra $L__BB107_2711; rcp.approx.ftz.f32 %f20105, %f4367; fma.rn.ftz.f32 %f20106, %f20105, 0f400B9581, 0f40623D71; mul.ftz.f32 %f20107, %f20105, %f20106; fma.rn.ftz.f32 %f20108, %f20105, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f20109, %f20105, %f20108, 0f3F800000; div.approx.ftz.f32 %f30065, %f20107, %f20109; $L__BB107_2711: mul.ftz.f32 %f20138, %f30063, 0f3E800000; div.approx.ftz.f32 %f20139, %f20138, %f4345; mul.ftz.f32 %f20140, %f30064, %f30065; mul.ftz.f32 %f30103, %f30064, %f20139; mul.ftz.f32 %f30104, %f20140, %f20139; bra.uni $L__BB107_2718; $L__BB107_2630: setp.neu.ftz.f32 %p2387, %f4113, %f4114; setp.neu.ftz.f32 %p2388, %f4113, %f4115; or.pred %p2389, %p2387, %p2388; neg.ftz.f32 %f19527, %f4115; selp.f32 %f19528, %f19527, %f4115, %p2389; add.ftz.f32 %f19529, %f4113, %f19528; neg.ftz.f32 %f19530, %f4114; selp.f32 %f19531, %f4114, %f19530, %p2389; sub.ftz.f32 %f19532, %f19531, %f4113; sub.ftz.f32 %f19533, %f4115, %f4114; mul.ftz.f32 %f19534, %f19529, %f19529; fma.rn.ftz.f32 %f19535, %f19533, %f19533, %f19534; fma.rn.ftz.f32 %f19536, %f19532, %f19532, %f19535; sqrt.approx.ftz.f32 %f19537, %f19536; div.approx.ftz.f32 %f29987, %f19533, %f19537; div.approx.ftz.f32 %f29988, %f19529, %f19537; div.approx.ftz.f32 %f29989, %f19532, %f19537; mul.ftz.f32 %f19538, %f4114, %f29989; mul.ftz.f32 %f19539, %f4115, %f29988; sub.ftz.f32 %f29984, %f19538, %f19539; mul.ftz.f32 %f19540, %f4115, %f29987; mul.ftz.f32 %f19541, %f4113, %f29989; sub.ftz.f32 %f29985, %f19540, %f19541; mul.ftz.f32 %f19542, %f4113, %f29988; mul.ftz.f32 %f19543, %f4114, %f29987; sub.ftz.f32 %f29986, %f19542, %f19543; $L__BB107_2632: mul.ftz.f32 %f19568, %f19503, %f29988; fma.rn.ftz.f32 %f19569, %f4107, %f29987, %f19568; fma.rn.ftz.f32 %f29996, %f19504, %f29989, %f19569; mul.ftz.f32 %f19570, %f19503, %f29985; fma.rn.ftz.f32 %f19571, %f4107, %f29984, %f19570; fma.rn.ftz.f32 %f29997, %f19504, %f29986, %f19571; mul.ftz.f32 %f19572, %f30259, %f29988; fma.rn.ftz.f32 %f19573, %f30260, %f29987, %f19572; fma.rn.ftz.f32 %f4138, %f30258, %f29989, %f19573; mul.ftz.f32 %f19574, %f30259, %f29985; fma.rn.ftz.f32 %f19575, %f30260, %f29984, %f19574; fma.rn.ftz.f32 %f4139, %f30258, %f29986, %f19575; add.ftz.f32 %f19576, %f29996, %f4138; add.ftz.f32 %f19577, %f29997, %f4139; mul.ftz.f32 %f19578, %f19576, %f19576; fma.rn.ftz.f32 %f19579, %f19577, %f19577, %f19578; add.ftz.f32 %f19580, %f29998, %f4117; fma.rn.ftz.f32 %f19581, %f19580, %f19580, %f19579; sqrt.approx.ftz.f32 %f19582, %f19581; div.approx.ftz.f32 %f4140, %f19576, %f19582; div.approx.ftz.f32 %f4141, %f19577, %f19582; div.approx.ftz.f32 %f19583, %f19580, %f19582; mul.ftz.f32 %f4142, %f19583, %f19583; @%p2386 bra $L__BB107_2637; bra.uni $L__BB107_2633; $L__BB107_2637: setp.gt.ftz.f32 %p2393, %f29998, 0f3F7FF972; neg.ftz.f32 %f19630, %f4140; div.approx.ftz.f32 %f19631, %f19630, %f19513; neg.ftz.f32 %f19632, %f4141; div.approx.ftz.f32 %f19633, %f19632, %f19514; fma.rn.ftz.f32 %f19634, %f19631, %f19631, %f4142; fma.rn.ftz.f32 %f4147, %f19633, %f19633, %f19634; mov.f32 %f29991, 0f00000000; @%p2393 bra $L__BB107_2640; setp.lt.ftz.f32 %p2394, %f29998, 0fBF7FF972; mov.f32 %f29991, 0fBF7FF972; @%p2394 bra $L__BB107_2640; mul.ftz.f32 %f19636, %f29998, %f29998; mov.f32 %f19637, 0f33D6BF95; max.ftz.f32 %f19638, %f19636, %f19637; rcp.approx.ftz.f32 %f19639, %f19638; mul.ftz.f32 %f19640, %f19513, %f29996; mul.ftz.f32 %f19641, %f29997, %f19514; mul.ftz.f32 %f19642, %f19641, %f19641; fma.rn.ftz.f32 %f19643, %f19640, %f19640, %f19642; fma.rn.ftz.f32 %f19644, %f19643, %f19639, 0f3F800000; sqrt.approx.ftz.f32 %f19645, %f19644; add.ftz.f32 %f19646, %f19645, 0fBF800000; mul.ftz.f32 %f29991, %f19646, 0f3F000000; $L__BB107_2640: add.ftz.f32 %f19647, %f29991, 0f3F800000; mov.f32 %f19648, 0f3F800000; mul.ftz.f32 %f19649, %f29998, %f19647; mov.f32 %f19650, 0f33D6BF95; max.ftz.f32 %f19651, %f19649, %f19650; mul.ftz.f32 %f19652, %f4147, 0f40490FDB; mul.ftz.f32 %f19653, %f4147, %f19652; mul.ftz.f32 %f19654, %f19513, %f19653; mul.ftz.f32 %f19655, %f19514, %f19654; max.ftz.f32 %f19656, %f19655, %f19650; rcp.approx.ftz.f32 %f19657, %f19656; mul.ftz.f32 %f19658, %f19657, 0f3E800000; div.approx.ftz.f32 %f19659, %f19658, %f19651; sqrt.approx.ftz.f32 %f19660, %f4112; mul.ftz.f32 %f19661, %f19660, 0fBFFE59F3; mul.ftz.f32 %f19662, %f19660, %f19661; mul.ftz.f32 %f19663, %f19662, 0f3FB8AA3B; ex2.approx.ftz.f32 %f19664, %f19663; fma.rn.ftz.f32 %f19665, %f19664, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f19666, %f19660, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f19667, %f19660, %f19666, 0f417C8B44; fma.rn.ftz.f32 %f19668, %f19660, %f19667, 0fC1665E35; fma.rn.ftz.f32 %f19669, %f19660, %f19668, 0f40CE7881; fma.rn.ftz.f32 %f19670, %f19660, %f19669, 0fBF993E81; fma.rn.ftz.f32 %f19671, %f19660, %f19670, 0f3CE3C31E; fma.rn.ftz.f32 %f19672, %f19660, %f19671, 0f3BBA5D49; sub.ftz.f32 %f19673, %f19665, %f19672; cvt.ftz.sat.f32.f32 %f19674, %f19673; sub.ftz.f32 %f19675, %f19648, %f19674; mul.ftz.f32 %f19676, %f4117, 0f3EA2F983; mul.ftz.f32 %f19677, %f19676, %f19675; fma.rn.ftz.f32 %f29992, %f19659, %f19674, %f19677; bra.uni $L__BB107_2641; $L__BB107_2299: ld.local.v2.f32 {%f19429, %f19430}, [%rd564+44]; ld.local.f32 %f19433, [%rd564+52]; mul.ftz.f32 %f19434, %f4073, %f19433; mul.ftz.f32 %f19435, %f4074, %f19430; sub.ftz.f32 %f19436, %f19434, %f19435; mul.ftz.f32 %f19437, %f4074, %f19429; mul.ftz.f32 %f19438, %f4071, %f19433; sub.ftz.f32 %f19439, %f19437, %f19438; mul.ftz.f32 %f19440, %f4071, %f19430; mul.ftz.f32 %f19441, %f4073, %f19429; sub.ftz.f32 %f19442, %f19440, %f19441; mul.ftz.f32 %f19443, %f19439, %f19439; fma.rn.ftz.f32 %f19444, %f19436, %f19436, %f19443; fma.rn.ftz.f32 %f19445, %f19442, %f19442, %f19444; sqrt.approx.ftz.f32 %f19446, %f19445; div.approx.ftz.f32 %f19447, %f19436, %f19446; div.approx.ftz.f32 %f19448, %f19439, %f19446; div.approx.ftz.f32 %f19449, %f19442, %f19446; mul.ftz.f32 %f19450, %f4074, %f19448; mul.ftz.f32 %f19451, %f4073, %f19449; sub.ftz.f32 %f19452, %f19450, %f19451; mul.ftz.f32 %f19453, %f4071, %f19449; mul.ftz.f32 %f19454, %f4074, %f19447; sub.ftz.f32 %f19455, %f19453, %f19454; mul.ftz.f32 %f19456, %f4073, %f19447; mul.ftz.f32 %f19457, %f4071, %f19448; sub.ftz.f32 %f19458, %f19456, %f19457; mul.ftz.f32 %f19459, %f4080, %f19455; fma.rn.ftz.f32 %f19460, %f4079, %f19452, %f19459; fma.rn.ftz.f32 %f4087, %f4081, %f19458, %f19460; mul.ftz.f32 %f19461, %f4080, %f19448; fma.rn.ftz.f32 %f19462, %f4079, %f19447, %f19461; fma.rn.ftz.f32 %f4088, %f4081, %f19449, %f19462; setp.geu.ftz.f32 %p2381, %f4083, 0f3F800000; mov.f32 %f29979, 0f3F800000; @%p2381 bra $L__BB107_2301; mul.ftz.f32 %f19463, %f4085, %f4087; mul.ftz.f32 %f19464, %f4087, %f19463; mul.ftz.f32 %f19465, %f4086, %f4088; fma.rn.ftz.f32 %f19466, %f4088, %f19465, %f19464; mul.ftz.f32 %f19467, %f4083, %f4083; mov.f32 %f19468, 0f3F800000; sub.ftz.f32 %f19469, %f19468, %f19467; div.approx.ftz.f32 %f19470, %f19466, %f19469; lg2.approx.ftz.f32 %f19471, %f4083; mul.ftz.f32 %f19472, %f19470, %f19471; ex2.approx.ftz.f32 %f29979, %f19472; $L__BB107_2301: add.ftz.f32 %f19473, %f4086, 0f3F800000; add.ftz.f32 %f19474, %f4085, 0f3F800000; mul.ftz.f32 %f19475, %f19474, %f19473; sqrt.approx.ftz.f32 %f19476, %f19475; mov.f32 %f19477, 0f41C90FDB; div.approx.ftz.f32 %f19478, %f19476, %f19477; mul.ftz.f32 %f19479, %f4084, %f19478; mul.ftz.f32 %f29981, %f29979, %f19479; mul.ftz.f32 %f19480, %f29979, %f19478; div.approx.ftz.f32 %f30103, %f19480, %f4082; $L__BB107_2623: max.ftz.f32 %f19490, %f4075, %f19419; max.ftz.f32 %f19491, %f4084, %f19490; mul.ftz.f32 %f19492, %f4082, %f19491; max.ftz.f32 %f19493, %f19419, %f19492; rcp.approx.ftz.f32 %f19494, %f19493; mul.ftz.f32 %f30104, %f19494, %f29981; $L__BB107_2624: mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; bra.uni $L__BB107_2774; $L__BB107_2633: setp.gt.ftz.f32 %p2391, %f29998, 0f3F7FF972; mul.ftz.f32 %f19585, %f19513, %f19513; mov.f32 %f19586, 0f3F800000; sub.ftz.f32 %f19587, %f19586, %f4142; fma.rn.ftz.f32 %f19588, %f19585, %f4142, %f19587; mul.ftz.f32 %f19589, %f19588, 0f40490FDB; mul.ftz.f32 %f19590, %f19588, %f19589; mov.f32 %f19591, 0f33D6BF95; max.ftz.f32 %f19592, %f19590, %f19591; div.approx.ftz.f32 %f4143, %f19585, %f19592; mov.f32 %f29990, 0f00000000; @%p2391 bra $L__BB107_2636; setp.lt.ftz.f32 %p2392, %f29998, 0fBF7FF972; mov.f32 %f29990, 0fBF7FF972; @%p2392 bra $L__BB107_2636; mul.ftz.f32 %f19594, %f29998, %f29998; max.ftz.f32 %f19596, %f19594, %f19591; rcp.approx.ftz.f32 %f19597, %f19596; mul.ftz.f32 %f19598, %f19513, %f29996; mul.ftz.f32 %f19599, %f19598, %f19598; mul.ftz.f32 %f19600, %f29997, %f19513; fma.rn.ftz.f32 %f19601, %f19600, %f19600, %f19599; fma.rn.ftz.f32 %f19602, %f19601, %f19597, 0f3F800000; sqrt.approx.ftz.f32 %f19603, %f19602; add.ftz.f32 %f19604, %f19603, 0fBF800000; mul.ftz.f32 %f29990, %f19604, 0f3F000000; $L__BB107_2636: add.ftz.f32 %f19605, %f29990, 0f3F800000; mul.ftz.f32 %f19607, %f29998, %f19605; max.ftz.f32 %f19609, %f19607, %f19591; mul.ftz.f32 %f19610, %f4143, 0f3E800000; div.approx.ftz.f32 %f19611, %f19610, %f19609; mul.ftz.f32 %f19612, %f19513, 0fBFFE59F3; mul.ftz.f32 %f19613, %f19513, %f19612; mul.ftz.f32 %f19614, %f19613, 0f3FB8AA3B; ex2.approx.ftz.f32 %f19615, %f19614; fma.rn.ftz.f32 %f19616, %f19615, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f19617, %f19513, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f19618, %f19513, %f19617, 0f417C8B44; fma.rn.ftz.f32 %f19619, %f19513, %f19618, 0fC1665E35; fma.rn.ftz.f32 %f19620, %f19513, %f19619, 0f40CE7881; fma.rn.ftz.f32 %f19621, %f19513, %f19620, 0fBF993E81; fma.rn.ftz.f32 %f19622, %f19513, %f19621, 0f3CE3C31E; fma.rn.ftz.f32 %f19623, %f19513, %f19622, 0f3BBA5D49; sub.ftz.f32 %f19624, %f19616, %f19623; cvt.ftz.sat.f32.f32 %f19625, %f19624; sub.ftz.f32 %f19626, %f19586, %f19625; mul.ftz.f32 %f19627, %f4117, 0f3EA2F983; mul.ftz.f32 %f19628, %f19627, %f19626; fma.rn.ftz.f32 %f29992, %f19611, %f19625, %f19628; $L__BB107_2641: setp.le.ftz.f32 %p2395, %f29992, 0f00000000; mov.f32 %f30103, %f30104; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2395 bra $L__BB107_2774; ld.local.u64 %rd2923, [%rd564+36]; ld.f32 %f4152, [%rd2923]; ld.f32 %f4153, [%rd2923+4]; ld.f32 %f4154, [%rd2923+8]; ld.local.f32 %f4155, [%rd564+28]; ld.f32 %f4156, [%rd2923+12]; ld.f32 %f4157, [%rd2923+16]; ld.f32 %f4158, [%rd2923+20]; setp.geu.ftz.f32 %p2396, %f4117, %f29998; mov.f32 %f29993, %f4138; mov.f32 %f29994, %f4139; mov.f32 %f29995, %f4117; mov.u16 %rs721, %rs519; @%p2396 bra $L__BB107_2644; mov.f32 %f29993, %f29996; mov.f32 %f29994, %f29997; mov.f32 %f29995, %f29998; mov.f32 %f29996, %f4138; mov.f32 %f29997, %f4139; mov.f32 %f29998, %f4117; mov.u16 %rs721, %rs534; $L__BB107_2644: setp.lt.ftz.f32 %p2397, %f29995, 0f3727C5AC; setp.lt.ftz.f32 %p2398, %f29998, 0f3727C5AC; or.pred %p2399, %p2397, %p2398; mov.f32 %f30103, %f29992; mov.f32 %f30105, %f30104; mov.f32 %f30106, %f30104; @%p2399 bra $L__BB107_2774; neg.ftz.f32 %f30044, %f29996; neg.ftz.f32 %f30045, %f29997; setp.lt.ftz.f32 %p2400, %f29998, 0fBF7FF972; mov.f32 %f30000, 0f00000000; mov.f32 %f30050, %f30000; @%p2400 bra $L__BB107_2648; setp.gt.ftz.f32 %p2401, %f29998, 0f3F7FF972; mov.f32 %f30050, 0fBF7FF972; @%p2401 bra $L__BB107_2648; mul.ftz.f32 %f19687, %f29998, %f29998; mov.f32 %f19688, 0f33D6BF95; max.ftz.f32 %f19689, %f19687, %f19688; rcp.approx.ftz.f32 %f19690, %f19689; mul.ftz.f32 %f19691, %f19513, %f30044; mul.ftz.f32 %f19692, %f19514, %f30045; mul.ftz.f32 %f19693, %f19692, %f19692; fma.rn.ftz.f32 %f19694, %f19691, %f19691, %f19693; fma.rn.ftz.f32 %f19695, %f19694, %f19690, 0f3F800000; sqrt.approx.ftz.f32 %f19696, %f19695; neg.ftz.f32 %f19697, %f19696; setp.ge.ftz.f32 %p2402, %f29998, 0f80000000; selp.f32 %f19698, %f19697, %f19696, %p2402; add.ftz.f32 %f19699, %f19698, 0fBF800000; mul.ftz.f32 %f30050, %f19699, 0f3F000000; $L__BB107_2648: setp.gt.ftz.f32 %p2403, %f29995, 0f3F7FF972; @%p2403 bra $L__BB107_2651; setp.lt.ftz.f32 %p2404, %f29995, 0fBF7FF972; mov.f32 %f30000, 0fBF7FF972; @%p2404 bra $L__BB107_2651; mul.ftz.f32 %f19702, %f29995, %f29995; mov.f32 %f19703, 0f33D6BF95; max.ftz.f32 %f19704, %f19702, %f19703; rcp.approx.ftz.f32 %f19705, %f19704; mul.ftz.f32 %f19706, %f19513, %f29993; mul.ftz.f32 %f19707, %f19514, %f29994; mul.ftz.f32 %f19708, %f19707, %f19707; fma.rn.ftz.f32 %f19709, %f19706, %f19706, %f19708; fma.rn.ftz.f32 %f19710, %f19709, %f19705, 0f3F800000; sqrt.approx.ftz.f32 %f19711, %f19710; neg.ftz.f32 %f19712, %f19711; setp.le.ftz.f32 %p2405, %f29995, 0f00000000; selp.f32 %f19713, %f19712, %f19711, %p2405; add.ftz.f32 %f19714, %f19713, 0fBF800000; mul.ftz.f32 %f30000, %f19714, 0f3F000000; $L__BB107_2651: add.ftz.f32 %f19715, %f29993, %f29996; add.ftz.f32 %f19716, %f29994, %f29997; mul.ftz.f32 %f19717, %f19716, %f19716; fma.rn.ftz.f32 %f19718, %f19715, %f19715, %f19717; add.ftz.f32 %f19719, %f29995, %f29998; fma.rn.ftz.f32 %f19720, %f19719, %f19719, %f19718; sqrt.approx.ftz.f32 %f19721, %f19720; div.approx.ftz.f32 %f4171, %f19715, %f19721; div.approx.ftz.f32 %f4172, %f19716, %f19721; div.approx.ftz.f32 %f4173, %f19719, %f19721; mul.ftz.f32 %f4174, %f4173, %f4173; setp.eq.ftz.f32 %p2406, %f19513, %f19514; @%p2406 bra $L__BB107_2653; bra.uni $L__BB107_2652; $L__BB107_2653: mul.ftz.f32 %f19734, %f19513, %f19513; mov.f32 %f19735, 0f3F800000; sub.ftz.f32 %f19736, %f19735, %f4174; fma.rn.ftz.f32 %f19737, %f19734, %f4174, %f19736; mul.ftz.f32 %f19738, %f19737, 0f40490FDB; mul.ftz.f32 %f19739, %f19737, %f19738; mov.f32 %f19740, 0f33D6BF95; max.ftz.f32 %f19741, %f19739, %f19740; div.approx.ftz.f32 %f30001, %f19734, %f19741; bra.uni $L__BB107_2654; $L__BB107_2310: mov.f32 %f19190, 0f3F800000; div.approx.ftz.f32 %f29971, %f4003, %f19190; $L__BB107_2605: mul.ftz.f32 %f4007, %f3998, %f3999; neg.ftz.f32 %f19193, %f4000; and.b32 %r4111, %r977, 1; setp.eq.b32 %p2350, %r4111, 1; selp.f32 %f4008, %f19193, %f4000, %p2350; add.ftz.f32 %f19194, %f29971, 0fBF800000; add.ftz.f32 %f19195, %f19194, 0f3F800000; mul.ftz.f32 %f19196, %f19195, %f19195; fma.rn.ftz.f32 %f19197, %f19196, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f19198, %f19195, %f19197; fma.rn.ftz.f32 %f19199, %f19196, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f19200, %f19196, %f19199, 0f3F800000; div.approx.ftz.f32 %f19201, %f19198, %f19200; sub.ftz.f32 %f19203, %f19121, %f19201; setp.gt.ftz.f32 %p2351, %f4003, 0f3F800000; selp.f32 %f19204, %f19203, %f19201, %p2351; mov.b32 %r4112, %f4002; and.b32 %r4113, %r4112, -2147483648; mov.b32 %r4114, %f19204; and.b32 %r4115, %r4114, 2147483647; or.b32 %r978, %r4115, %r4113; add.ftz.f32 %f19205, %f3995, 0fBFC90FDB; mul.ftz.f32 %f19206, %f19205, 0f3F000000; sub.ftz.f32 %f19207, %f19206, %f3993; mul.ftz.f32 %f4009, %f4001, %f19207; abs.ftz.f32 %f4010, %f4009; setp.eq.ftz.f32 %p2352, %f4010, 0f00000000; @%p2352 bra $L__BB107_2610; setp.eq.ftz.f32 %p2353, %f4010, 0f3F800000; mov.f32 %f29972, 0f3F800000; @%p2353 bra $L__BB107_2610; setp.gt.ftz.f32 %p2354, %f4010, 0f3F800000; @%p2354 bra $L__BB107_2609; bra.uni $L__BB107_2608; $L__BB107_2609: mov.f32 %f19210, 0f3F800000; div.approx.ftz.f32 %f29972, %f19210, %f4010; bra.uni $L__BB107_2610; $L__BB107_2745: mov.f32 %f20604, 0f3F800000; div.approx.ftz.f32 %f30087, %f4488, %f20604; $L__BB107_2747: add.ftz.f32 %f20607, %f30087, 0fBF800000; add.ftz.f32 %f20608, %f20607, 0f3F800000; mul.ftz.f32 %f20609, %f20608, %f20608; fma.rn.ftz.f32 %f20610, %f20609, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f20611, %f20608, %f20610; fma.rn.ftz.f32 %f20612, %f20609, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f20613, %f20609, %f20612, 0f3F800000; div.approx.ftz.f32 %f20614, %f20611, %f20613; sub.ftz.f32 %f20616, %f20530, %f20614; setp.gt.ftz.f32 %p2505, %f4488, 0f3F800000; selp.f32 %f20617, %f20616, %f20614, %p2505; mov.b32 %r4134, %f4487; and.b32 %r4135, %r4134, -2147483648; mov.b32 %r4136, %f20617; and.b32 %r4137, %r4136, 2147483647; or.b32 %r986, %r4137, %r4135; add.ftz.f32 %f20618, %f4481, 0fBFC90FDB; div.approx.ftz.f32 %f20620, %f20618, %f20600; sub.ftz.f32 %f20621, %f20620, %f20494; mul.ftz.f32 %f4492, %f4486, %f20621; abs.ftz.f32 %f4493, %f4492; setp.eq.ftz.f32 %p2506, %f4493, 0f00000000; mov.f32 %f30088, %f30089; @%p2506 bra $L__BB107_2752; setp.eq.ftz.f32 %p2507, %f4493, 0f3F800000; mov.f32 %f30088, 0f3F800000; @%p2507 bra $L__BB107_2752; setp.gt.ftz.f32 %p2508, %f4493, 0f3F800000; @%p2508 bra $L__BB107_2751; bra.uni $L__BB107_2750; $L__BB107_2751: mov.f32 %f20624, 0f3F800000; div.approx.ftz.f32 %f30088, %f20624, %f4493; bra.uni $L__BB107_2752; $L__BB107_2608: mov.f32 %f19209, 0f3F800000; div.approx.ftz.f32 %f29972, %f4010, %f19209; $L__BB107_2610: add.ftz.f32 %f19211, %f29972, 0fBF800000; add.ftz.f32 %f19212, %f19211, 0f3F800000; mul.ftz.f32 %f19213, %f19212, %f19212; fma.rn.ftz.f32 %f19214, %f19213, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f19215, %f19212, %f19214; fma.rn.ftz.f32 %f19216, %f19213, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f19217, %f19213, %f19216, 0f3F800000; div.approx.ftz.f32 %f19218, %f19215, %f19217; sub.ftz.f32 %f19220, %f19121, %f19218; setp.gt.ftz.f32 %p2355, %f4010, 0f3F800000; selp.f32 %f19221, %f19220, %f19218, %p2355; mov.b32 %r4116, %f4009; and.b32 %r4117, %r4116, -2147483648; mov.b32 %r4118, %f19221; and.b32 %r4119, %r4118, 2147483647; or.b32 %r4120, %r4119, %r4117; mov.b32 %f19222, %r4120; add.ftz.f32 %f19223, %f3995, %f3996; mul.ftz.f32 %f19224, %f19223, 0f3F000000; sub.ftz.f32 %f19225, %f19224, %f3993; mul.ftz.f32 %f19226, %f4007, 0f3F000000; abs.ftz.f32 %f19227, %f4007; setp.lt.ftz.f32 %p2356, %f19227, 0f40490FDB; selp.f32 %f19228, %f19226, 0f3FC90FDB, %p2356; mul.ftz.f32 %f19229, %f19228, 0f3EA2F983; mov.b32 %r4121, %f19229; and.b32 %r4122, %r4121, -2147483648; or.b32 %r4123, %r4122, 1056964608; mov.b32 %f19230, %r4123; add.ftz.f32 %f19231, %f19229, %f19230; cvt.rzi.ftz.s32.f32 %r4124, %f19231; cvt.rn.f32.s32 %f19232, %r4124; fma.rn.ftz.f32 %f19233, %f19232, 0fC0490000, %f19228; fma.rn.ftz.f32 %f19234, %f19232, 0fBA7DA000, %f19233; fma.rn.ftz.f32 %f19235, %f19232, 0fB4222000, %f19234; fma.rn.ftz.f32 %f19236, %f19232, 0fACB4611A, %f19235; add.ftz.f32 %f19237, %f19236, 0fBFC90FDB; add.ftz.f32 %f19238, %f19237, 0f3FC90FDB; mul.ftz.f32 %f19239, %f19238, %f19238; fma.rn.ftz.f32 %f19240, %f19239, 0fB491ED89, 0f37D0078B; fma.rn.ftz.f32 %f19241, %f19239, %f19240, 0fBAB60B58; fma.rn.ftz.f32 %f19242, %f19239, %f19241, 0f3D2AAAAA; fma.rn.ftz.f32 %f19243, %f19239, %f19242, 0fBF000000; fma.rn.ftz.f32 %f19244, %f19239, %f19243, 0f3F800000; and.b32 %r4125, %r4124, 1; setp.eq.b32 %p2357, %r4125, 1; neg.ftz.f32 %f19245, %f19244; selp.f32 %f19246, %f19245, %f19244, %p2357; abs.ftz.f32 %f19247, %f19246; setp.gt.ftz.f32 %p2358, %f19247, 0f3F800000; mul.ftz.f32 %f19248, %f19246, 0f3E800000; selp.f32 %f19249, 0f00000000, %f19248, %p2358; mul.ftz.f32 %f19250, %f3998, %f19249; mul.ftz.f32 %f19251, %f19225, %f19225; fma.rn.ftz.f32 %f19252, %f19089, %f19089, %f19251; add.ftz.f32 %f19253, %f19252, %f19252; mov.b32 %f19254, %r978; sub.ftz.f32 %f19255, %f19254, %f19222; mul.ftz.f32 %f19256, %f19253, %f19255; abs.ftz.f32 %f19257, %f4008; setp.gt.ftz.f32 %p2359, %f19257, 0f3F800000; selp.f32 %f19258, 0f00000000, %f4008, %p2359; mul.ftz.f32 %f19259, %f19258, %f19256; div.approx.ftz.f32 %f19260, %f19089, %f19259; mul.ftz.f32 %f30103, %f19260, %f19250; $L__BB107_2611: mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; bra.uni $L__BB107_2774; $L__BB107_2750: mov.f32 %f20623, 0f3F800000; div.approx.ftz.f32 %f30088, %f4493, %f20623; $L__BB107_2752: add.ftz.f32 %f20626, %f30088, 0fBF800000; add.ftz.f32 %f20627, %f20626, 0f3F800000; mul.ftz.f32 %f20628, %f20627, %f20627; fma.rn.ftz.f32 %f20629, %f20628, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f20630, %f20627, %f20629; fma.rn.ftz.f32 %f20631, %f20628, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f20632, %f20628, %f20631, 0f3F800000; div.approx.ftz.f32 %f20633, %f20630, %f20632; sub.ftz.f32 %f20635, %f20530, %f20633; setp.gt.ftz.f32 %p2509, %f4493, 0f3F800000; selp.f32 %f20636, %f20635, %f20633, %p2509; mov.b32 %r4138, %f4492; and.b32 %r4139, %r4138, -2147483648; mov.b32 %r4140, %f20636; and.b32 %r4141, %r4140, 2147483647; or.b32 %r987, %r4141, %r4139; div.approx.ftz.f32 %f4497, %f20530, %f20493; abs.ftz.f32 %f4498, %f4497; setp.eq.ftz.f32 %p2510, %f4498, 0f00000000; @%p2510 bra $L__BB107_2757; setp.eq.ftz.f32 %p2511, %f4498, 0f3F800000; mov.f32 %f30089, 0f3F800000; @%p2511 bra $L__BB107_2757; setp.gt.ftz.f32 %p2512, %f4498, 0f3F800000; @%p2512 bra $L__BB107_2756; bra.uni $L__BB107_2755; $L__BB107_2756: mov.f32 %f20639, 0f3F800000; div.approx.ftz.f32 %f30089, %f20639, %f4498; bra.uni $L__BB107_2757; $L__BB107_2755: mov.f32 %f20638, 0f3F800000; div.approx.ftz.f32 %f30089, %f4498, %f20638; $L__BB107_2757: add.ftz.f32 %f20640, %f30089, 0fBF800000; add.ftz.f32 %f20641, %f20640, 0f3F800000; mul.ftz.f32 %f20642, %f20641, %f20641; fma.rn.ftz.f32 %f20643, %f20642, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f20644, %f20641, %f20643; fma.rn.ftz.f32 %f20645, %f20642, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f20646, %f20642, %f20645, 0f3F800000; div.approx.ftz.f32 %f20647, %f20644, %f20646; sub.ftz.f32 %f20649, %f20530, %f20647; setp.gt.ftz.f32 %p2513, %f4498, 0f3F800000; selp.f32 %f20650, %f20649, %f20647, %p2513; mov.b32 %r4142, %f4497; and.b32 %r4143, %r4142, -2147483648; mov.b32 %r4144, %f20650; and.b32 %r4145, %r4144, 2147483647; or.b32 %r4146, %r4145, %r4143; mov.b32 %f20651, %r4146; add.ftz.f32 %f20652, %f20651, %f20651; add.ftz.f32 %f20654, %f4481, %f4482; div.approx.ftz.f32 %f20655, %f20654, %f20600; sub.ftz.f32 %f20656, %f20655, %f20494; abs.ftz.f32 %f20657, %f4484; sub.ftz.f32 %f20659, %f20532, %f20657; mul.ftz.f32 %f20660, %f20656, %f20656; fma.rn.ftz.f32 %f20661, %f20498, %f20498, %f20660; add.ftz.f32 %f20662, %f20661, %f20661; mov.b32 %f20663, %r986; mov.b32 %f20664, %r987; sub.ftz.f32 %f20665, %f20663, %f20664; mul.ftz.f32 %f20666, %f20665, %f20662; abs.ftz.f32 %f20667, %f4485; setp.gt.ftz.f32 %p2514, %f20667, 0f3F800000; selp.f32 %f20668, 0f00000000, %f4485, %p2514; mul.ftz.f32 %f20669, %f20668, %f20666; div.approx.ftz.f32 %f20670, %f20498, %f20669; mul.ftz.f32 %f20671, %f20659, %f20659; fma.rn.ftz.f32 %f20672, %f20493, %f20493, %f20671; mul.ftz.f32 %f20673, %f20672, %f20652; div.approx.ftz.f32 %f20674, %f20493, %f20673; mul.ftz.f32 %f30103, %f20670, %f20674; $L__BB107_2758: mov.f32 %f30104, %f30103; mov.f32 %f30105, %f30103; mov.f32 %f30106, %f30103; $L__BB107_2774: add.s32 %r989, %r963, -2; setp.gt.u32 %p2534, %r989, 6; @%p2534 bra $L__BB107_2785; ld.local.f32 %f4599, [%rd564+16]; ld.local.f32 %f4600, [%rd5+-68]; ld.local.f32 %f4601, [%rd5+-76]; ld.local.f32 %f4602, [%rd564+8]; setp.neu.ftz.f32 %p2535, %f4602, %f4601; mov.u16 %rs722, %rs519; @%p2535 bra $L__BB107_2778; ld.local.f32 %f20852, [%rd564+12]; ld.local.f32 %f20853, [%rd5+-72]; setp.neu.ftz.f32 %p2536, %f20852, %f20853; mov.u16 %rs722, %rs519; @%p2536 bra $L__BB107_2778; setp.eq.ftz.f32 %p2537, %f4599, %f4600; selp.u16 %rs722, 1, 0, %p2537; $L__BB107_2778: setp.ne.s16 %p2538, %rs722, 0; @%p2538 bra $L__BB107_2785; ld.local.f32 %f20855, [%rd564+12]; mul.ftz.f32 %f20856, %f30259, %f20855; fma.rn.ftz.f32 %f20857, %f30260, %f4602, %f20856; fma.rn.ftz.f32 %f20858, %f30258, %f4599, %f20857; setp.lt.ftz.f32 %p2539, %f20858, 0f00000000; mov.f32 %f30107, 0f00000000; ld.local.f32 %f20859, [%rd5+-72]; neg.ftz.f32 %f20860, %f20859; neg.ftz.f32 %f20861, %f4601; selp.f32 %f20862, %f20861, %f4601, %p2539; selp.f32 %f20863, %f20860, %f20859, %p2539; neg.ftz.f32 %f20864, %f4600; selp.f32 %f20865, %f20864, %f4600, %p2539; mul.ftz.f32 %f20866, %f30259, %f20863; fma.rn.ftz.f32 %f20867, %f30260, %f20862, %f20866; fma.rn.ftz.f32 %f4603, %f30258, %f20865, %f20867; mul.ftz.f32 %f20868, %f20855, %f20863; fma.rn.ftz.f32 %f20869, %f4602, %f20862, %f20868; fma.rn.ftz.f32 %f20870, %f4599, %f20865, %f20869; mul.ftz.f32 %f4604, %f20858, %f20870; setp.eq.ftz.f32 %p2540, %f4604, 0f00000000; @%p2540 bra $L__BB107_2781; div.approx.ftz.f32 %f30107, %f4603, %f4604; $L__BB107_2781: setp.ge.ftz.f32 %p2541, %f30107, 0f3F800000; mov.f32 %f30108, 0f3F800000; @%p2541 bra $L__BB107_2784; setp.lt.ftz.f32 %p2542, %f30107, 0f00000000; mov.f32 %f30108, 0f00000000; @%p2542 bra $L__BB107_2784; mul.ftz.f32 %f20873, %f30107, %f30107; mul.ftz.f32 %f20874, %f30107, %f20873; sub.ftz.f32 %f20875, %f20873, %f20874; add.ftz.f32 %f30108, %f30107, %f20875; $L__BB107_2784: mul.ftz.f32 %f30104, %f30104, %f30108; mul.ftz.f32 %f30105, %f30105, %f30108; mul.ftz.f32 %f30106, %f30106, %f30108; $L__BB107_2785: ld.local.u32 %r4150, [%rd5+-12]; mul.wide.s32 %rd2926, %r4150, 208; add.s64 %rd2927, %rd300, %rd2926; ld.global.f32 %f4615, [%rd2927+180]; setp.leu.ftz.f32 %p2543, %f4615, 0f3F800000; @%p2543 bra $L__BB107_2788; ld.local.f32 %f20876, [%rd564+8]; ld.local.f32 %f20877, [%rd564+12]; mul.ftz.f32 %f20878, %f30259, %f20877; fma.rn.ftz.f32 %f20879, %f30260, %f20876, %f20878; ld.local.f32 %f20880, [%rd564+16]; fma.rn.ftz.f32 %f4616, %f30258, %f20880, %f20879; setp.ltu.ftz.f32 %p2544, %f4616, 0f00000000; @%p2544 bra $L__BB107_2788; mov.f32 %f20881, 0f3F800000; min.ftz.f32 %f20882, %f4616, %f20881; abs.ftz.f32 %f20883, %f20882; setp.lt.ftz.f32 %p2545, %f20883, 0f3F800000; add.ftz.f32 %f20884, %f20883, 0fBF800000; add.ftz.f32 %f20885, %f20884, 0f3F800000; selp.f32 %f20886, %f20885, 0f3F800000, %p2545; sub.ftz.f32 %f20887, %f20881, %f20886; sqrt.approx.ftz.f32 %f20888, %f20887; fma.rn.ftz.f32 %f20889, %f20886, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f20890, %f20886, %f20889, 0fBE5A6B93; fma.rn.ftz.f32 %f20891, %f20886, %f20890, 0f3FC90FDB; mul.ftz.f32 %f20892, %f20888, %f20891; setp.lt.ftz.f32 %p2546, %f20882, 0f00000000; mov.f32 %f20893, 0f00000000; mov.f32 %f20894, 0f40490FDB; sub.ftz.f32 %f20895, %f20894, %f20892; selp.f32 %f20896, %f20895, %f20892, %p2546; mul.ftz.f32 %f20897, %f4615, %f20896; cos.approx.ftz.f32 %f20898, %f20897; max.ftz.f32 %f20899, %f20898, %f20893; div.approx.ftz.f32 %f20900, %f20899, %f20882; mul.ftz.f32 %f30104, %f30104, %f20900; mul.ftz.f32 %f30105, %f30105, %f20900; mul.ftz.f32 %f30106, %f30106, %f20900; $L__BB107_2788: setp.eq.ftz.f32 %p2547, %f30103, 0f00000000; @%p2547 bra $L__BB107_2794; setp.lt.u32 %p2548, %r989, 7; ld.local.v4.f32 {%f20901, %f20902, %f20903, %f20904}, [%rd564+-12]; mul.ftz.f32 %f4623, %f30104, %f20901; mul.ftz.f32 %f4624, %f30105, %f20902; mul.ftz.f32 %f4625, %f30106, %f20903; @%p2548 bra $L__BB107_2792; bra.uni $L__BB107_2790; $L__BB107_2792: add.ftz.f32 %f30126, %f4623, %f30126; add.ftz.f32 %f30125, %f4624, %f30125; add.ftz.f32 %f30124, %f4625, %f30124; bra.uni $L__BB107_2793; $L__BB107_2790: add.s32 %r4151, %r963, -9; setp.gt.u32 %p2549, %r4151, 11; setp.ne.s32 %p2550, %r963, 29; and.pred %p2551, %p2550, %p2549; @%p2551 bra $L__BB107_2793; add.ftz.f32 %f30129, %f4623, %f30129; add.ftz.f32 %f30128, %f4624, %f30128; add.ftz.f32 %f30127, %f4625, %f30127; $L__BB107_2793: add.ftz.f32 %f29810, %f4623, %f29810; add.ftz.f32 %f29809, %f4624, %f29809; add.ftz.f32 %f29808, %f4625, %f29808; $L__BB107_2794: cvt.u32.u64 %r4152, %rd563; add.s32 %r5779, %r4152, 1; setp.lt.s32 %p2552, %r5779, %r957; @%p2552 bra $L__BB107_2253; st.local.f32 [%rd283+12], %f30129; st.local.f32 [%rd283+16], %f30128; st.local.f32 [%rd283+20], %f30127; st.local.f32 [%rd283], %f30126; st.local.f32 [%rd283+4], %f30125; st.local.f32 [%rd283+8], %f30124; st.local.f32 [%rd283+24], %f29810; st.local.f32 [%rd283+28], %f29809; st.local.f32 [%rd283+32], %f29808; $L__BB107_2796: ld.local.v4.f32 {%f20908, %f20909, %f20910, %f20911}, [%rd413+400]; mov.b32 %r4153, %f20908; ld.local.u32 %r991, [%rd280+52]; or.b32 %r4154, %r991, 4; setp.eq.s32 %p2553, %r4154, 4; mov.b64 %rd2928, {%r4153, %r4155}; cvt.u32.u64 %r4156, %rd2928; mov.b32 %f4662, %r4156; @%p2553 bra $L__BB107_2801; bra.uni $L__BB107_2797; $L__BB107_2801: ld.local.f32 %f20946, [%rd280+-24]; sub.ftz.f32 %f20947, %f20946, %f20908; ld.local.f32 %f20948, [%rd280+-20]; sub.ftz.f32 %f20949, %f20948, %f20909; ld.local.f32 %f20950, [%rd280+-16]; sub.ftz.f32 %f20951, %f20950, %f20910; mul.ftz.f32 %f20952, %f20949, %f20949; fma.rn.ftz.f32 %f20953, %f20947, %f20947, %f20952; fma.rn.ftz.f32 %f20954, %f20951, %f20951, %f20953; sqrt.approx.ftz.f32 %f30145, %f20954; st.local.f32 [%rd280+12], %f30145; rsqrt.approx.ftz.f32 %f20955, %f20954; mul.ftz.f32 %f30148, %f20947, %f20955; mul.ftz.f32 %f30147, %f20949, %f20955; mul.ftz.f32 %f30146, %f20951, %f20955; st.local.f32 [%rd280], %f30148; st.local.f32 [%rd280+4], %f30147; st.local.f32 [%rd280+8], %f30146; neg.ftz.f32 %f4674, %f30148; neg.ftz.f32 %f4675, %f30147; neg.ftz.f32 %f4676, %f30146; st.local.f32 [%rd280+-12], %f4674; st.local.f32 [%rd280+-8], %f4675; st.local.f32 [%rd280+-4], %f4676; mul.ftz.f32 %f20956, %f30147, %f30147; fma.rn.ftz.f32 %f20957, %f30148, %f30148, %f20956; fma.rn.ftz.f32 %f4677, %f30146, %f30146, %f20957; setp.leu.ftz.f32 %p2556, %f4677, 0f00000000; mov.f32 %f30141, 0f00000000; mov.f32 %f30142, %f30141; @%p2556 bra $L__BB107_2810; setp.eq.ftz.f32 %p2557, %f30148, 0f80000000; mov.f32 %f30141, 0f00000000; setp.eq.ftz.f32 %p2558, %f30147, 0f80000000; and.pred %p2559, %p2557, %p2558; @%p2559 bra $L__BB107_2809; abs.ftz.f32 %f4678, %f4675; setp.eq.ftz.f32 %p2560, %f4678, 0f00000000; abs.ftz.f32 %f4679, %f4674; setp.eq.ftz.f32 %p2561, %f4679, 0f00000000; and.pred %p2562, %p2560, %p2561; mov.b32 %r992, %f4675; mov.b32 %r4157, %f4674; and.b32 %r993, %r4157, -2147483648; @%p2562 bra $L__BB107_2807; bra.uni $L__BB107_2804; $L__BB107_2807: shr.s32 %r4162, %r992, 31; and.b32 %r4163, %r4162, 1078530011; or.b32 %r4164, %r4163, %r993; mov.b32 %f30139, %r4164; bra.uni $L__BB107_2808; $L__BB107_2797: setp.eq.s32 %p2554, %r991, 3; @%p2554 bra $L__BB107_2799; bra.uni $L__BB107_2798; $L__BB107_2799: ld.global.f32 %f20912, [%rd284+116]; abs.ftz.f32 %f30144, %f20912; ld.local.f32 %f20913, [%rd280+-24]; sub.ftz.f32 %f20914, %f20913, %f20908; ld.local.f32 %f20915, [%rd280+-20]; sub.ftz.f32 %f20916, %f20915, %f20909; ld.local.f32 %f20917, [%rd280+-16]; sub.ftz.f32 %f20918, %f20917, %f20910; mul.ftz.f32 %f20919, %f20916, %f20916; fma.rn.ftz.f32 %f20920, %f20914, %f20914, %f20919; fma.rn.ftz.f32 %f20921, %f20918, %f20918, %f20920; sqrt.approx.ftz.f32 %f30145, %f20921; st.local.f32 [%rd280+12], %f30145; rsqrt.approx.ftz.f32 %f20922, %f20921; mul.ftz.f32 %f30148, %f20914, %f20922; mul.ftz.f32 %f30147, %f20916, %f20922; mul.ftz.f32 %f30146, %f20918, %f20922; st.local.f32 [%rd280], %f30148; st.local.f32 [%rd280+4], %f30147; st.local.f32 [%rd280+8], %f30146; st.local.f32 [%rd280+24], %f30144; ld.global.f32 %f4669, [%rd284+132]; setp.leu.ftz.f32 %p2555, %f4669, 0f00000000; @%p2555 bra $L__BB107_2815; mul.ftz.f32 %f20923, %f30144, 0f3E800000; ld.local.f32 %f20924, [%rd280+-12]; ld.local.f32 %f20925, [%rd280+-8]; mul.ftz.f32 %f20926, %f30147, %f20925; fma.rn.ftz.f32 %f20927, %f30148, %f20924, %f20926; ld.local.f32 %f20928, [%rd280+-4]; fma.rn.ftz.f32 %f20929, %f30146, %f20928, %f20927; neg.ftz.f32 %f20930, %f20929; mul.ftz.f32 %f20931, %f20929, %f20929; mov.f32 %f20932, 0f3F800000; sub.ftz.f32 %f20933, %f20932, %f20931; mov.f32 %f20934, 0f00000000; max.ftz.f32 %f20935, %f20933, %f20934; sqrt.approx.ftz.f32 %f20936, %f20935; div.approx.ftz.f32 %f20937, %f20936, %f20930; mul.ftz.f32 %f20938, %f4669, %f20937; sub.ftz.f32 %f20939, %f20932, %f20938; ld.global.f32 %f20940, [%rd284+148]; mul.ftz.f32 %f20941, %f20940, %f20939; max.ftz.f32 %f20942, %f20941, %f20934; mul.ftz.f32 %f20943, %f20942, %f20923; st.local.f32 [%rd280+28], %f20943; bra.uni $L__BB107_2815; $L__BB107_2798: ld.local.f32 %f30144, [%rd280+24]; bra.uni $L__BB107_2815; $L__BB107_2804: setp.eq.ftz.f32 %p2563, %f4678, 0f7F800000; setp.eq.ftz.f32 %p2564, %f4679, 0f7F800000; and.pred %p2565, %p2563, %p2564; @%p2565 bra $L__BB107_2806; bra.uni $L__BB107_2805; $L__BB107_2806: setp.lt.s32 %p2569, %r992, 0; selp.b32 %r4160, 1075235812, 1061752795, %p2569; or.b32 %r4161, %r4160, %r993; mov.b32 %f30139, %r4161; bra.uni $L__BB107_2808; $L__BB107_2805: setp.lt.s32 %p2566, %r992, 0; min.ftz.f32 %f20959, %f4679, %f4678; max.ftz.f32 %f20960, %f4679, %f4678; div.full.ftz.f32 %f20961, %f20959, %f20960; mul.rn.ftz.f32 %f20962, %f20961, %f20961; mov.f32 %f20963, 0fC0B59883; mov.f32 %f20964, 0fBF52C7EA; fma.rn.ftz.f32 %f20965, %f20962, %f20964, %f20963; mov.f32 %f20966, 0fC0D21907; fma.rn.ftz.f32 %f20967, %f20965, %f20962, %f20966; mul.ftz.f32 %f20968, %f20962, %f20967; mul.ftz.f32 %f20969, %f20961, %f20968; add.ftz.f32 %f20970, %f20962, 0f41355DC0; mov.f32 %f20971, 0f41E6BD60; fma.rn.ftz.f32 %f20972, %f20970, %f20962, %f20971; mov.f32 %f20973, 0f419D92C8; fma.rn.ftz.f32 %f20974, %f20972, %f20962, %f20973; rcp.approx.ftz.f32 %f20975, %f20974; fma.rn.ftz.f32 %f20976, %f20969, %f20975, %f20961; mov.f32 %f20977, 0f3FC90FDB; sub.ftz.f32 %f20978, %f20977, %f20976; setp.gt.ftz.f32 %p2567, %f4679, %f4678; selp.f32 %f20979, %f20978, %f20976, %p2567; mov.f32 %f20980, 0f40490FDB; sub.ftz.f32 %f20981, %f20980, %f20979; selp.f32 %f20982, %f20981, %f20979, %p2566; mov.b32 %r4158, %f20982; or.b32 %r4159, %r993, %r4158; mov.b32 %f20983, %r4159; add.ftz.f32 %f20984, %f4678, %f4679; setp.le.ftz.f32 %p2568, %f20984, 0f7F800000; selp.f32 %f30139, %f20983, %f20984, %p2568; $L__BB107_2808: fma.rn.ftz.f32 %f30141, %f30139, 0fBE22F983, 0f3F000000; $L__BB107_2809: sqrt.approx.ftz.f32 %f20985, %f4677; div.approx.ftz.f32 %f20986, %f4676, %f20985; mov.f32 %f20987, 0fBF800000; max.ftz.f32 %f20988, %f20986, %f20987; mov.f32 %f20989, 0f3F800000; min.ftz.f32 %f20990, %f20988, %f20989; abs.ftz.f32 %f20991, %f20990; sub.ftz.f32 %f20992, %f20989, %f20991; mul.ftz.f32 %f20993, %f20992, 0f3F000000; sqrt.approx.ftz.f32 %f20994, %f20993; setp.gt.ftz.f32 %p2570, %f20991, 0f3F11EB85; selp.f32 %f20995, %f20994, %f20991, %p2570; mul.ftz.f32 %f20996, %f20995, %f20995; mov.f32 %f20997, 0f3C94D2E9; mov.f32 %f20998, 0f3D53F941; fma.rn.ftz.f32 %f20999, %f20998, %f20996, %f20997; mov.f32 %f21000, 0f3D3F841F; fma.rn.ftz.f32 %f21001, %f20999, %f20996, %f21000; mov.f32 %f21002, 0f3D994929; fma.rn.ftz.f32 %f21003, %f21001, %f20996, %f21002; mov.f32 %f21004, 0f3E2AAB94; fma.rn.ftz.f32 %f21005, %f21003, %f20996, %f21004; mul.ftz.f32 %f21006, %f20996, %f21005; fma.rn.ftz.f32 %f21007, %f21006, %f20995, %f20995; add.ftz.f32 %f21008, %f21007, %f21007; mov.f32 %f21009, 0f3FC90FDB; sub.ftz.f32 %f21010, %f21009, %f21007; selp.f32 %f21011, %f21008, %f21010, %p2570; setp.lt.ftz.f32 %p2571, %f20990, 0f00000000; mov.f32 %f21012, 0f40490FDB; sub.ftz.f32 %f21013, %f21012, %f21011; selp.f32 %f21014, %f21013, %f21011, %p2571; fma.rn.ftz.f32 %f30142, %f21014, 0fBEA2F983, 0f3F800000; $L__BB107_2810: st.local.f32 [%rd280+16], %f30141; st.local.f32 [%rd280+20], %f30142; ld.global.f32 %f30144, [%rd284+108]; mul.ftz.f32 %f4690, %f30144, 0f3DA2F983; st.local.f32 [%rd280+28], %f4690; st.local.f32 [%rd280+24], %f30144; setp.ne.s32 %p2572, %r991, 4; @%p2572 bra $L__BB107_2815; ld.global.v4.f32 {%f21016, %f21017, %f21018, %f21019}, [%rd284+120]; ld.global.v2.f32 {%f21023, %f21024}, [%rd284+112]; mul.ftz.f32 %f21025, %f21017, %f4675; fma.rn.ftz.f32 %f21026, %f21016, %f4674, %f21025; fma.rn.ftz.f32 %f4693, %f21018, %f4676, %f21026; setp.le.ftz.f32 %p2573, %f4693, %f21023; mov.f32 %f30143, 0f00000000; @%p2573 bra $L__BB107_2814; sub.ftz.f32 %f4694, %f4693, %f21023; setp.geu.ftz.f32 %p2574, %f4694, %f21024; setp.eq.ftz.f32 %p2575, %f21024, 0f00000000; or.pred %p2576, %p2575, %p2574; mov.f32 %f30143, %f4693; @%p2576 bra $L__BB107_2814; div.approx.ftz.f32 %f21027, %f4694, %f21024; mul.ftz.f32 %f21028, %f21027, %f21027; mul.ftz.f32 %f21029, %f21028, 0fC0000000; mul.ftz.f32 %f21030, %f21027, %f21029; fma.rn.ftz.f32 %f21031, %f21028, 0f40400000, %f21030; mul.ftz.f32 %f30143, %f4693, %f21031; $L__BB107_2814: mul.ftz.f32 %f21032, %f30143, %f4690; st.local.f32 [%rd280+28], %f21032; $L__BB107_2815: ld.const.f32 %f21036, [kernel_params+1504]; mul.ftz.f32 %f4702, %f21036, %f30144; st.local.f32 [%rd280+24], %f4702; ld.global.u16 %rs252, [%rd404]; ld.global.u16 %rs253, [%rd278]; ld.global.u16 %rs254, [%rd405]; add.s16 %rs537, %rs254, 1; st.global.u16 [%rd405], %rs537; cvt.u16.u32 %rs538, %r5732; add.s16 %rs539, %rs538, %rs252; add.s16 %rs540, %rs539, -1; st.global.u16 [%rd404], %rs540; add.s16 %rs541, %rs253, %rs538; st.global.u16 [%rd278], %rs541; ld.local.f32 %f4703, [%rd5+-4]; and.b32 %r5790, %r956, 4194303; shl.b32 %r4165, %r956, 5; cvt.u64.u32 %rd2929, %r4165; and.b64 %rd2930, %rd2929, 134217696; add.s64 %rd2931, %rd303, %rd2930; add.s64 %rd568, %rd2931, 16; ld.global.u32 %r995, [%rd2931+16]; and.b32 %r4166, %r995, 134217728; setp.eq.s32 %p2578, %r4166, 0; mov.pred %p3925, -1; mov.f32 %f30190, 0f00000000; mov.f32 %f30191, %f30190; mov.f32 %f30192, %f30190; @%p2578 bra $L__BB107_2817; ld.global.f32 %f30190, [%rd568+-16]; ld.global.f32 %f30191, [%rd568+-12]; ld.global.f32 %f30192, [%rd568+-8]; mov.pred %p3925, 0; $L__BB107_2817: @%p3925 bra $L__BB107_2821; bra.uni $L__BB107_2818; $L__BB107_2821: setp.eq.s32 %p2582, %r991, 2; ld.local.f32 %f4713, [%rd280+-24]; ld.local.f32 %f4714, [%rd280+-20]; ld.local.f32 %f4715, [%rd280+-16]; neg.ftz.f32 %f4716, %f30148; neg.ftz.f32 %f4717, %f30147; neg.ftz.f32 %f4718, %f30146; @%p2582 bra $L__BB107_2854; bra.uni $L__BB107_2822; $L__BB107_2854: add.s64 %rd2979, %rd2172, 248; cvta.to.local.u64 %rd2980, %rd2979; add.s64 %rd2981, %rd2172, 240; cvta.to.local.u64 %rd2982, %rd2981; st.local.v2.f32 [%rd2982], {%f4713, %f4714}; st.local.f32 [%rd2980], %f4715; st.local.v4.f32 [%rd2173], {%f30148, %f30147, %f30146, %f4716}; st.local.v4.f32 [%rd337], {%f4717, %f4718, %f4716, %f4717}; st.local.v4.f32 [%rd307], {%f4718, %f4716, %f4717, %f4718}; ld.const.u32 %r4200, [kernel_params+1024]; and.b32 %r5790, %r4200, 4194303; shl.b32 %r4201, %r4200, 5; cvt.u64.u32 %rd2986, %r4201; and.b64 %rd2987, %rd2986, 134217696; add.s64 %rd2988, %rd303, %rd2987; ld.global.u32 %r4202, [%rd2988+16]; st.local.u32 [%rd339], %r3028; mov.u32 %r4204, 0; st.local.v4.u32 [%rd305], {%r4200, %r4202, %r4204, %r3028}; st.local.u32 [%rd357], %r4204; st.local.u32 [%rd351], %r4204; cvta.to.local.u64 %rd2990, %rd309; mov.f32 %f21520, 0f00000000; st.local.v2.f32 [%rd2990], {%f21520, %f21520}; st.local.u32 [%rd321], %r4204; mov.b32 %r4205, %f4703; cvta.to.local.u64 %rd2991, %rd292; st.local.v4.u32 [%rd2991], {%r3028, %r4205, %r4204, %r4204}; st.local.v4.f32 [%rd324], {%f21520, %f21520, %f21520, %f21520}; add.s64 %rd2993, %rd289, 48; cvta.to.local.u64 %rd2994, %rd2993; st.local.v4.f32 [%rd2994], {%f21520, %f21520, %f21520, %f21520}; bra.uni $L__BB107_2855; $L__BB107_2818: setp.eq.s32 %p2580, %r620, -1; @%p2580 bra $L__BB107_3066; ld.local.f32 %f4710, [%rd280+-12]; ld.local.f32 %f4711, [%rd280+-8]; mul.ftz.f32 %f21037, %f4711, %f30147; fma.rn.ftz.f32 %f21038, %f4710, %f30148, %f21037; ld.local.f32 %f4712, [%rd280+-4]; fma.rn.ftz.f32 %f21039, %f4712, %f30146, %f21038; setp.leu.ftz.f32 %p2581, %f21039, 0f00000000; @%p2581 bra $L__BB107_3066; neg.ftz.f32 %f21040, %f4710; st.local.f32 [%rd280+-12], %f21040; neg.ftz.f32 %f21041, %f4711; st.local.f32 [%rd280+-8], %f21041; neg.ftz.f32 %f21042, %f4712; st.local.f32 [%rd280+-4], %f21042; bra.uni $L__BB107_3066; $L__BB107_2822: ld.local.f32 %f30182, [%rd280+-12]; ld.local.f32 %f30183, [%rd280+-8]; ld.local.f32 %f4721, [%rd280+-4]; ld.local.f32 %f4722, [%rd280+16]; ld.local.f32 %f4723, [%rd280+20]; mov.u32 %r5787, 1; st.local.v4.f32 [%rd2173], {%f4713, %f4714, %f4715, %f30182}; st.local.v4.f32 [%rd337], {%f30183, %f4721, %f30182, %f30183}; st.local.v4.f32 [%rd307], {%f4721, %f4716, %f4717, %f4718}; st.local.u32 [%rd305], %r956; setp.ne.s32 %p2583, %r620, -1; @%p2583 bra $L__BB107_2824; mov.u32 %r5787, 32; $L__BB107_2824: st.local.u32 [%rd291], %r620; mov.u32 %r4171, 0; mov.b32 %r4172, %f4723; mov.b32 %r4173, %f4722; st.local.v4.u32 [%rd354], {%r5787, %r4173, %r4172, %r619}; mov.b32 %r4174, %f4703; cvta.to.local.u64 %rd2936, %rd292; st.local.v2.u32 [%rd2936], {%r3028, %r4174}; st.local.f32 [%rd288], %f30145; st.local.u32 [%rd293], %r995; st.local.u32 [%rd290], %r4171; setp.eq.s32 %p2584, %r619, -1; @%p2584 bra $L__BB107_2849; shl.b64 %rd2937, %rd286, 2; add.s64 %rd2938, %rd287, %rd2937; ld.global.u32 %r998, [%rd2938]; st.local.u32 [%rd290], %r998; and.b32 %r999, %r998, 2; setp.eq.s32 %p2585, %r999, 0; mul.lo.s64 %rd2939, %rd286, 208; add.s64 %rd570, %rd300, %rd2939; @%p2585 bra $L__BB107_2832; ld.global.u32 %rd2940, [%rd570+168]; ld.global.u32 %r4176, [%rd570+152]; shl.b32 %r4177, %r4176, 1; cvt.rn.f32.s32 %f21043, %r4177; mul.ftz.f32 %f21044, %f4703, %f21043; cvt.rzi.ftz.s32.f32 %r4178, %f21044; add.s32 %r4179, %r4177, -1; min.s32 %r4180, %r4178, %r4179; cvt.rn.f32.s32 %f21045, %r4180; sub.ftz.f32 %f4732, %f21044, %f21045; cvt.s64.s32 %rd2941, %r4180; add.s64 %rd2942, %rd2941, %rd2940; shl.b64 %rd2943, %rd2942, 6; add.s64 %rd571, %rd301, %rd2943; ld.global.v4.f32 {%f21046, %f21047, %f21048, %f21049}, [%rd571]; ld.global.v4.f32 {%f21050, %f21051, %f21052, %f21053}, [%rd571+64]; mul.ftz.f32 %f21054, %f21047, %f21051; fma.rn.ftz.f32 %f21055, %f21046, %f21050, %f21054; mul.ftz.f32 %f21056, %f21049, %f21053; fma.rn.ftz.f32 %f21057, %f21048, %f21052, %f21056; add.ftz.f32 %f4741, %f21055, %f21057; setp.gt.ftz.f32 %p2586, %f4741, 0f3F7FDF3B; mov.f32 %f21058, 0f3F800000; sub.ftz.f32 %f4742, %f21058, %f4732; @%p2586 bra $L__BB107_2828; bra.uni $L__BB107_2827; $L__BB107_2828: mul.ftz.f32 %f21112, %f21046, %f4742; fma.rn.ftz.f32 %f21113, %f21050, %f4732, %f21112; mul.ftz.f32 %f21114, %f21047, %f4742; fma.rn.ftz.f32 %f21115, %f21051, %f4732, %f21114; mul.ftz.f32 %f21116, %f21048, %f4742; fma.rn.ftz.f32 %f21117, %f21052, %f4732, %f21116; mul.ftz.f32 %f21118, %f21049, %f4742; fma.rn.ftz.f32 %f21119, %f21053, %f4732, %f21118; mul.ftz.f32 %f21120, %f21115, %f21115; fma.rn.ftz.f32 %f21121, %f21113, %f21113, %f21120; mul.ftz.f32 %f21122, %f21119, %f21119; fma.rn.ftz.f32 %f21123, %f21117, %f21117, %f21122; add.ftz.f32 %f21124, %f21121, %f21123; rsqrt.approx.ftz.f32 %f21125, %f21124; mul.ftz.f32 %f30155, %f21119, %f21125; mul.ftz.f32 %f30154, %f21117, %f21125; mul.ftz.f32 %f30153, %f21115, %f21125; mul.ftz.f32 %f30152, %f21113, %f21125; bra.uni $L__BB107_2829; $L__BB107_2849: st.local.u32 [%rd2936], %r618; mov.u32 %r4194, 0; st.local.u32 [%rd318], %r4194; st.local.u32 [%rd319], %r4194; st.local.u32 [%rd320], %r4194; cvta.to.local.u64 %rd2974, %rd309; mov.f32 %f30172, 0f00000000; st.local.v2.f32 [%rd2974], {%f30172, %f30172}; st.local.u32 [%rd321], %r4194; mov.f32 %f30173, %f30172; mov.f32 %f30174, %f30172; mov.f32 %f30175, %f30172; mov.f32 %f30176, %f30172; mov.f32 %f30177, %f30172; mov.f32 %f30178, %f4721; mov.f32 %f30179, %f30183; mov.f32 %f30180, %f30182; bra.uni $L__BB107_2850; $L__BB107_2827: mov.f32 %f21059, 0fBF800000; max.ftz.f32 %f21060, %f4741, %f21059; min.ftz.f32 %f21062, %f21060, %f21058; abs.ftz.f32 %f21063, %f21062; sub.ftz.f32 %f21064, %f21058, %f21063; mul.ftz.f32 %f21065, %f21064, 0f3F000000; sqrt.approx.ftz.f32 %f21066, %f21065; setp.gt.ftz.f32 %p2587, %f21063, 0f3F11EB85; selp.f32 %f21067, %f21066, %f21063, %p2587; mul.ftz.f32 %f21068, %f21067, %f21067; mov.f32 %f21069, 0f3C94D2E9; mov.f32 %f21070, 0f3D53F941; fma.rn.ftz.f32 %f21071, %f21070, %f21068, %f21069; mov.f32 %f21072, 0f3D3F841F; fma.rn.ftz.f32 %f21073, %f21071, %f21068, %f21072; mov.f32 %f21074, 0f3D994929; fma.rn.ftz.f32 %f21075, %f21073, %f21068, %f21074; mov.f32 %f21076, 0f3E2AAB94; fma.rn.ftz.f32 %f21077, %f21075, %f21068, %f21076; mul.ftz.f32 %f21078, %f21068, %f21077; fma.rn.ftz.f32 %f21079, %f21078, %f21067, %f21067; add.ftz.f32 %f21080, %f21079, %f21079; mov.f32 %f21081, 0f3FC90FDB; sub.ftz.f32 %f21082, %f21081, %f21079; selp.f32 %f21083, %f21080, %f21082, %p2587; setp.lt.ftz.f32 %p2588, %f21062, 0f00000000; mov.f32 %f21084, 0f40490FDB; sub.ftz.f32 %f21085, %f21084, %f21083; selp.f32 %f21086, %f21085, %f21083, %p2588; mul.ftz.f32 %f21087, %f21046, %f4741; sub.ftz.f32 %f21088, %f21050, %f21087; mul.ftz.f32 %f21089, %f21047, %f4741; sub.ftz.f32 %f21090, %f21051, %f21089; mul.ftz.f32 %f21091, %f21048, %f4741; sub.ftz.f32 %f21092, %f21052, %f21091; mul.ftz.f32 %f21093, %f21049, %f4741; sub.ftz.f32 %f21094, %f21053, %f21093; mul.ftz.f32 %f21095, %f21090, %f21090; fma.rn.ftz.f32 %f21096, %f21088, %f21088, %f21095; mul.ftz.f32 %f21097, %f21094, %f21094; fma.rn.ftz.f32 %f21098, %f21092, %f21092, %f21097; add.ftz.f32 %f21099, %f21096, %f21098; rsqrt.approx.ftz.f32 %f21100, %f21099; mul.ftz.f32 %f21101, %f21088, %f21100; mul.ftz.f32 %f21102, %f21090, %f21100; mul.ftz.f32 %f21103, %f21092, %f21100; mul.ftz.f32 %f21104, %f21094, %f21100; mul.ftz.f32 %f21105, %f4732, %f21086; cos.approx.ftz.f32 %f21106, %f21105; sin.approx.ftz.f32 %f21107, %f21105; mul.ftz.f32 %f21108, %f21101, %f21107; mul.ftz.f32 %f21109, %f21102, %f21107; mul.ftz.f32 %f21110, %f21103, %f21107; mul.ftz.f32 %f21111, %f21104, %f21107; fma.rn.ftz.f32 %f30155, %f21049, %f21106, %f21111; fma.rn.ftz.f32 %f30154, %f21048, %f21106, %f21110; fma.rn.ftz.f32 %f30153, %f21047, %f21106, %f21109; fma.rn.ftz.f32 %f30152, %f21046, %f21106, %f21108; $L__BB107_2829: ld.global.v4.f32 {%f21126, %f21127, %f21128, %f21129}, [%rd571+16]; ld.global.v4.f32 {%f21134, %f21135, %f21136, %f21137}, [%rd571+80]; mul.ftz.f32 %f21142, %f4732, %f21134; mul.ftz.f32 %f21143, %f4732, %f21135; mul.ftz.f32 %f21144, %f4732, %f21136; mul.ftz.f32 %f21145, %f4732, %f21137; fma.rn.ftz.f32 %f4755, %f4742, %f21126, %f21142; fma.rn.ftz.f32 %f4756, %f4742, %f21127, %f21143; fma.rn.ftz.f32 %f4757, %f4742, %f21128, %f21144; fma.rn.ftz.f32 %f21146, %f4742, %f21129, %f21145; ld.global.v4.f32 {%f21147, %f21148, %f21149, %f21150}, [%rd571+32]; ld.global.v4.f32 {%f21155, %f21156, %f21157, %f21158}, [%rd571+96]; mul.ftz.f32 %f21163, %f4732, %f21155; mul.ftz.f32 %f21164, %f4732, %f21156; mul.ftz.f32 %f21165, %f4732, %f21157; mul.ftz.f32 %f21166, %f4732, %f21158; fma.rn.ftz.f32 %f21167, %f4742, %f21147, %f21163; fma.rn.ftz.f32 %f21168, %f4742, %f21148, %f21164; fma.rn.ftz.f32 %f21169, %f4742, %f21149, %f21165; fma.rn.ftz.f32 %f21170, %f4742, %f21150, %f21166; ld.global.v4.f32 {%f21171, %f21172, %f21173, %f21174}, [%rd571+48]; ld.global.v4.f32 {%f21179, %f21180, %f21181, %f21182}, [%rd571+112]; mul.ftz.f32 %f21187, %f4732, %f21179; mul.ftz.f32 %f21188, %f4732, %f21180; mul.ftz.f32 %f21189, %f4732, %f21181; mul.ftz.f32 %f21190, %f4732, %f21182; fma.rn.ftz.f32 %f21191, %f4742, %f21171, %f21187; fma.rn.ftz.f32 %f21192, %f4742, %f21172, %f21188; fma.rn.ftz.f32 %f21193, %f4742, %f21173, %f21189; fma.rn.ftz.f32 %f21194, %f4742, %f21174, %f21190; mul.ftz.f32 %f21195, %f30155, 0f3FB504F3; mul.ftz.f32 %f21196, %f30152, 0f3FB504F3; mul.ftz.f32 %f21197, %f30153, 0f3FB504F3; mul.ftz.f32 %f21198, %f30154, 0f3FB504F3; mul.ftz.f32 %f21199, %f21195, %f21196; mul.ftz.f32 %f21200, %f21195, %f21197; mul.ftz.f32 %f21201, %f21195, %f21198; mul.ftz.f32 %f21202, %f21196, %f21196; mul.ftz.f32 %f21203, %f21196, %f21197; mul.ftz.f32 %f21204, %f21196, %f21198; mul.ftz.f32 %f21205, %f21197, %f21197; mul.ftz.f32 %f21206, %f21197, %f21198; mul.ftz.f32 %f21207, %f21198, %f21198; sub.ftz.f32 %f21209, %f21058, %f21205; sub.ftz.f32 %f21210, %f21209, %f21207; sub.ftz.f32 %f21211, %f21203, %f21201; add.ftz.f32 %f21212, %f21200, %f21204; add.ftz.f32 %f21213, %f21201, %f21203; sub.ftz.f32 %f21214, %f21058, %f21202; sub.ftz.f32 %f21215, %f21214, %f21207; sub.ftz.f32 %f21216, %f21206, %f21199; sub.ftz.f32 %f21217, %f21204, %f21200; add.ftz.f32 %f21218, %f21199, %f21206; sub.ftz.f32 %f21219, %f21214, %f21205; mul.ftz.f32 %f21220, %f21211, %f21169; fma.rn.ftz.f32 %f21221, %f21210, %f21146, %f21220; fma.rn.ftz.f32 %f30161, %f21212, %f21192, %f21221; mul.ftz.f32 %f21222, %f21211, %f21170; fma.rn.ftz.f32 %f21223, %f21210, %f21167, %f21222; fma.rn.ftz.f32 %f4759, %f21212, %f21193, %f21223; mul.ftz.f32 %f21224, %f21211, %f21191; fma.rn.ftz.f32 %f21225, %f21210, %f21168, %f21224; fma.rn.ftz.f32 %f4760, %f21212, %f21194, %f21225; mul.ftz.f32 %f21226, %f21215, %f21169; fma.rn.ftz.f32 %f21227, %f21213, %f21146, %f21226; fma.rn.ftz.f32 %f4761, %f21216, %f21192, %f21227; mul.ftz.f32 %f21228, %f21215, %f21170; fma.rn.ftz.f32 %f21229, %f21213, %f21167, %f21228; fma.rn.ftz.f32 %f30160, %f21216, %f21193, %f21229; mul.ftz.f32 %f21230, %f21215, %f21191; fma.rn.ftz.f32 %f21231, %f21213, %f21168, %f21230; fma.rn.ftz.f32 %f4763, %f21216, %f21194, %f21231; mul.ftz.f32 %f21232, %f21218, %f21169; fma.rn.ftz.f32 %f21233, %f21217, %f21146, %f21232; fma.rn.ftz.f32 %f4764, %f21219, %f21192, %f21233; mul.ftz.f32 %f21234, %f21218, %f21170; fma.rn.ftz.f32 %f21235, %f21217, %f21167, %f21234; fma.rn.ftz.f32 %f4765, %f21219, %f21193, %f21235; mul.ftz.f32 %f21236, %f21218, %f21191; fma.rn.ftz.f32 %f21237, %f21217, %f21168, %f21236; fma.rn.ftz.f32 %f30159, %f21219, %f21194, %f21237; cvta.to.local.u64 %rd2944, %rd299; st.local.v4.f32 [%rd2944], {%f30161, %f4759, %f4760, %f4755}; cvta.to.local.u64 %rd2945, %rd311; st.local.v4.f32 [%rd2945], {%f4761, %f30160, %f4763, %f4756}; cvta.to.local.u64 %rd2946, %rd310; st.local.v4.f32 [%rd2946], {%f4764, %f4765, %f30159, %f4757}; mul.ftz.f32 %f21238, %f30160, %f30159; mul.ftz.f32 %f4767, %f4765, %f4763; sub.ftz.f32 %f30158, %f21238, %f4767; mul.ftz.f32 %f4769, %f4765, %f4760; mul.ftz.f32 %f21239, %f4759, %f30159; sub.ftz.f32 %f30157, %f4769, %f21239; mul.ftz.f32 %f4771, %f4759, %f4763; mul.ftz.f32 %f21240, %f30160, %f4760; sub.ftz.f32 %f30156, %f4771, %f21240; mul.ftz.f32 %f21241, %f4761, %f30157; fma.rn.ftz.f32 %f21242, %f30161, %f30158, %f21241; fma.rn.ftz.f32 %f30162, %f4764, %f30156, %f21242; setp.neu.ftz.f32 %p2589, %f30162, 0f00000000; @%p2589 bra $L__BB107_2831; add.ftz.f32 %f30161, %f30161, 0f322BCC77; add.ftz.f32 %f30159, %f30159, 0f322BCC77; add.ftz.f32 %f30160, %f30160, 0f322BCC77; mul.ftz.f32 %f21243, %f30160, %f30159; sub.ftz.f32 %f30158, %f21243, %f4767; mul.ftz.f32 %f21244, %f4759, %f30159; sub.ftz.f32 %f30157, %f4769, %f21244; mul.ftz.f32 %f21245, %f30160, %f4760; sub.ftz.f32 %f30156, %f4771, %f21245; mul.ftz.f32 %f21246, %f4761, %f30157; fma.rn.ftz.f32 %f21247, %f30161, %f30158, %f21246; fma.rn.ftz.f32 %f21248, %f4764, %f30156, %f21247; setp.eq.ftz.f32 %p2590, %f21248, 0f00000000; selp.f32 %f30162, 0f7F7FFFFF, %f21248, %p2590; $L__BB107_2831: mul.ftz.f32 %f21249, %f4761, %f30159; mul.ftz.f32 %f21250, %f4764, %f4763; sub.ftz.f32 %f21251, %f21250, %f21249; mul.ftz.f32 %f21252, %f4764, %f4760; mul.ftz.f32 %f21253, %f30159, %f30161; sub.ftz.f32 %f21254, %f21253, %f21252; mul.ftz.f32 %f21255, %f4763, %f30161; mul.ftz.f32 %f21256, %f4761, %f4760; sub.ftz.f32 %f21257, %f21256, %f21255; mul.ftz.f32 %f21258, %f4764, %f30160; mul.ftz.f32 %f21259, %f4761, %f4765; sub.ftz.f32 %f21260, %f21259, %f21258; mul.ftz.f32 %f21261, %f4765, %f30161; mul.ftz.f32 %f21262, %f4764, %f4759; sub.ftz.f32 %f21263, %f21262, %f21261; mul.ftz.f32 %f21264, %f4761, %f4759; mul.ftz.f32 %f21265, %f30160, %f30161; sub.ftz.f32 %f21266, %f21265, %f21264; div.approx.ftz.f32 %f21267, %f30157, %f30162; mul.ftz.f32 %f21268, %f4756, %f21267; div.approx.ftz.f32 %f21269, %f30158, %f30162; fma.rn.ftz.f32 %f21270, %f4755, %f21269, %f21268; div.approx.ftz.f32 %f21271, %f30156, %f30162; fma.rn.ftz.f32 %f21272, %f4757, %f21271, %f21270; div.approx.ftz.f32 %f21273, %f21254, %f30162; mul.ftz.f32 %f21274, %f4756, %f21273; div.approx.ftz.f32 %f21275, %f21251, %f30162; fma.rn.ftz.f32 %f21276, %f4755, %f21275, %f21274; div.approx.ftz.f32 %f21277, %f21257, %f30162; fma.rn.ftz.f32 %f21278, %f4757, %f21277, %f21276; div.approx.ftz.f32 %f21279, %f21263, %f30162; mul.ftz.f32 %f21280, %f4756, %f21279; div.approx.ftz.f32 %f21281, %f21260, %f30162; fma.rn.ftz.f32 %f21282, %f4755, %f21281, %f21280; div.approx.ftz.f32 %f21283, %f21266, %f30162; fma.rn.ftz.f32 %f21284, %f4757, %f21283, %f21282; neg.ftz.f32 %f21285, %f21272; cvta.to.local.u64 %rd2947, %rd312; st.local.v4.f32 [%rd2947], {%f21269, %f21267, %f21271, %f21285}; neg.ftz.f32 %f21286, %f21278; cvta.to.local.u64 %rd2948, %rd327; st.local.v4.f32 [%rd2948], {%f21275, %f21273, %f21277, %f21286}; neg.ftz.f32 %f21287, %f21284; cvta.to.local.u64 %rd2949, %rd326; st.local.v4.f32 [%rd2949], {%f21281, %f21279, %f21283, %f21287}; $L__BB107_2832: setp.eq.s32 %p2591, %r5787, 1; @%p2591 bra $L__BB107_2834; bra.uni $L__BB107_2833; $L__BB107_2834: setp.lt.s32 %p2592, %r956, 0; shl.b64 %rd2951, %rd359, 4; add.s64 %rd572, %rd328, %rd2951; @%p2592 bra $L__BB107_2836; and.b32 %r5788, %r998, 4; mov.f32 %f30178, %f4721; mov.f32 %f30179, %f30183; mov.f32 %f30180, %f30182; bra.uni $L__BB107_2842; $L__BB107_2833: st.local.u32 [%rd318], %r4171; st.local.u32 [%rd319], %r4171; st.local.u32 [%rd320], %r4171; cvta.to.local.u64 %rd2950, %rd309; mov.f32 %f30172, 0f00000000; st.local.v2.f32 [%rd2950], {%f30172, %f30172}; st.local.u32 [%rd321], %r4171; mov.f32 %f30173, %f30172; mov.f32 %f30174, %f30172; mov.f32 %f30175, %f30172; mov.f32 %f30176, %f30172; mov.f32 %f30177, %f30172; mov.f32 %f30178, %f4721; mov.f32 %f30179, %f30183; mov.f32 %f30180, %f30182; bra.uni $L__BB107_2850; $L__BB107_2836: ld.global.v4.u32 {%r4182, %r4183, %r4184, %r4185}, [%rd572]; mul.wide.u32 %rd2952, %r4182, 12; add.s64 %rd2953, %rd352, %rd2952; mul.wide.u32 %rd2954, %r4183, 12; add.s64 %rd2955, %rd352, %rd2954; mul.wide.u32 %rd2956, %r4184, 12; add.s64 %rd2957, %rd352, %rd2956; mov.f32 %f21294, 0f3F800000; sub.ftz.f32 %f21295, %f21294, %f4722; sub.ftz.f32 %f21296, %f21295, %f4723; ld.global.f32 %f21297, [%rd2953]; ld.global.f32 %f21298, [%rd2953+4]; ld.global.f32 %f21299, [%rd2953+8]; ld.global.f32 %f21300, [%rd2955]; mul.ftz.f32 %f21301, %f4722, %f21300; ld.global.f32 %f21302, [%rd2955+4]; mul.ftz.f32 %f21303, %f4722, %f21302; ld.global.f32 %f21304, [%rd2955+8]; mul.ftz.f32 %f21305, %f4722, %f21304; fma.rn.ftz.f32 %f21306, %f21296, %f21297, %f21301; fma.rn.ftz.f32 %f21307, %f21296, %f21298, %f21303; fma.rn.ftz.f32 %f21308, %f21296, %f21299, %f21305; ld.global.f32 %f21309, [%rd2957]; ld.global.f32 %f21310, [%rd2957+4]; ld.global.f32 %f21311, [%rd2957+8]; fma.rn.ftz.f32 %f30163, %f4723, %f21309, %f21306; fma.rn.ftz.f32 %f30164, %f4723, %f21310, %f21307; fma.rn.ftz.f32 %f30165, %f4723, %f21311, %f21308; mul.ftz.f32 %f21312, %f30164, %f30164; fma.rn.ftz.f32 %f21313, %f30163, %f30163, %f21312; fma.rn.ftz.f32 %f21314, %f30165, %f30165, %f21313; sqrt.approx.ftz.f32 %f4791, %f21314; setp.neu.ftz.f32 %p2593, %f4791, 0f00000000; @%p2593 bra $L__BB107_2837; bra.uni $L__BB107_2838; $L__BB107_2837: rcp.approx.ftz.f32 %f21315, %f4791; mul.ftz.f32 %f30165, %f30165, %f21315; mul.ftz.f32 %f30164, %f30164, %f21315; mul.ftz.f32 %f30163, %f30163, %f21315; $L__BB107_2838: setp.eq.ftz.f32 %p2594, %f30163, 0f00000000; setp.eq.ftz.f32 %p2595, %f30164, 0f00000000; and.pred %p2596, %p2594, %p2595; setp.eq.ftz.f32 %p2597, %f30165, 0f00000000; and.pred %p2598, %p2597, %p2596; selp.f32 %f30178, %f4721, %f30165, %p2598; selp.f32 %f30179, %f30183, %f30164, %p2598; selp.f32 %f30180, %f30182, %f30163, %p2598; st.local.f32 [%rd315], %f30180; st.local.f32 [%rd337], %f30179; st.local.f32 [%rd338], %f30178; and.b32 %r5788, %r998, 4; setp.ne.s32 %p2599, %r5788, 0; @%p2599 bra $L__BB107_2842; @%p2585 bra $L__BB107_2841; bra.uni $L__BB107_2840; $L__BB107_2841: ld.global.v4.f32 {%f21350, %f21351, %f21352, %f21353}, [%rd570+48]; ld.global.v4.f32 {%f21357, %f21358, %f21359, %f21360}, [%rd570+64]; ld.global.v4.f32 {%f21364, %f21365, %f21366, %f21367}, [%rd570+80]; mul.ftz.f32 %f21371, %f21357, %f30179; fma.rn.ftz.f32 %f21372, %f21350, %f30180, %f21371; fma.rn.ftz.f32 %f21373, %f21364, %f30178, %f21372; mul.ftz.f32 %f21374, %f21358, %f30179; fma.rn.ftz.f32 %f21375, %f21351, %f30180, %f21374; fma.rn.ftz.f32 %f21376, %f21365, %f30178, %f21375; mul.ftz.f32 %f21377, %f21359, %f30179; fma.rn.ftz.f32 %f21378, %f21352, %f30180, %f21377; fma.rn.ftz.f32 %f21379, %f21366, %f30178, %f21378; mul.ftz.f32 %f21380, %f21376, %f21376; fma.rn.ftz.f32 %f21381, %f21373, %f21373, %f21380; fma.rn.ftz.f32 %f21382, %f21379, %f21379, %f21381; sqrt.approx.ftz.f32 %f21383, %f21382; div.approx.ftz.f32 %f30180, %f21373, %f21383; div.approx.ftz.f32 %f30179, %f21376, %f21383; div.approx.ftz.f32 %f30178, %f21379, %f21383; st.local.f32 [%rd315], %f30180; st.local.f32 [%rd337], %f30179; st.local.f32 [%rd338], %f30178; mov.u32 %r5788, %r4171; bra.uni $L__BB107_2842; $L__BB107_2840: cvta.to.local.u64 %rd2958, %rd312; ld.local.v4.f32 {%f21316, %f21317, %f21318, %f21319}, [%rd2958]; cvta.to.local.u64 %rd2959, %rd327; ld.local.v4.f32 {%f21323, %f21324, %f21325, %f21326}, [%rd2959]; cvta.to.local.u64 %rd2960, %rd326; ld.local.v4.f32 {%f21330, %f21331, %f21332, %f21333}, [%rd2960]; mul.ftz.f32 %f21337, %f30179, %f21323; fma.rn.ftz.f32 %f21338, %f30180, %f21316, %f21337; fma.rn.ftz.f32 %f21339, %f30178, %f21330, %f21338; mul.ftz.f32 %f21340, %f30179, %f21324; fma.rn.ftz.f32 %f21341, %f30180, %f21317, %f21340; fma.rn.ftz.f32 %f21342, %f30178, %f21331, %f21341; mul.ftz.f32 %f21343, %f30179, %f21325; fma.rn.ftz.f32 %f21344, %f30180, %f21318, %f21343; fma.rn.ftz.f32 %f21345, %f30178, %f21332, %f21344; mul.ftz.f32 %f21346, %f21342, %f21342; fma.rn.ftz.f32 %f21347, %f21339, %f21339, %f21346; fma.rn.ftz.f32 %f21348, %f21345, %f21345, %f21347; sqrt.approx.ftz.f32 %f21349, %f21348; div.approx.ftz.f32 %f30180, %f21339, %f21349; div.approx.ftz.f32 %f30179, %f21342, %f21349; div.approx.ftz.f32 %f30178, %f21345, %f21349; st.local.f32 [%rd315], %f30180; st.local.f32 [%rd337], %f30179; st.local.f32 [%rd338], %f30178; mov.u32 %r5788, %r4171; $L__BB107_2842: ld.global.u32 %r4191, [%rd572+12]; mul.wide.u32 %rd2961, %r4191, 12; add.s64 %rd2962, %rd329, %rd2961; add.s32 %r4192, %r4191, 1; mul.wide.u32 %rd2963, %r4192, 12; add.s64 %rd2964, %rd329, %rd2963; add.s32 %r4193, %r4191, 2; mul.wide.u32 %rd2965, %r4193, 12; add.s64 %rd2966, %rd329, %rd2965; ld.global.f32 %f21384, [%rd2966]; ld.global.f32 %f21385, [%rd2966+4]; ld.global.f32 %f21386, [%rd2966+8]; ld.global.f32 %f21387, [%rd2964]; ld.global.f32 %f21388, [%rd2962]; sub.ftz.f32 %f30177, %f21387, %f21388; ld.global.f32 %f21389, [%rd2964+4]; ld.global.f32 %f21390, [%rd2962+4]; sub.ftz.f32 %f30176, %f21389, %f21390; ld.global.f32 %f21391, [%rd2964+8]; ld.global.f32 %f21392, [%rd2962+8]; sub.ftz.f32 %f30175, %f21391, %f21392; st.local.f32 [%rd318], %f30177; st.local.f32 [%rd319], %f30176; st.local.f32 [%rd320], %f30175; sub.ftz.f32 %f30174, %f21384, %f21388; sub.ftz.f32 %f30173, %f21385, %f21390; sub.ftz.f32 %f30172, %f21386, %f21392; cvta.to.local.u64 %rd573, %rd309; st.local.v2.f32 [%rd573], {%f30174, %f30173}; st.local.f32 [%rd321], %f30172; setp.ne.s32 %p2601, %r5788, 0; @%p2601 bra $L__BB107_2850; @%p2585 bra $L__BB107_2845; bra.uni $L__BB107_2844; $L__BB107_2845: ld.global.v4.f32 {%f21420, %f21421, %f21422, %f21423}, [%rd570]; ld.global.v4.f32 {%f21427, %f21428, %f21429, %f21430}, [%rd570+16]; ld.global.v4.f32 {%f21434, %f21435, %f21436, %f21437}, [%rd570+32]; mul.ftz.f32 %f21441, %f21421, %f30176; fma.rn.ftz.f32 %f21442, %f30177, %f21420, %f21441; fma.rn.ftz.f32 %f30171, %f21422, %f30175, %f21442; mul.ftz.f32 %f21443, %f21428, %f30176; fma.rn.ftz.f32 %f21444, %f30177, %f21427, %f21443; fma.rn.ftz.f32 %f30170, %f21429, %f30175, %f21444; mul.ftz.f32 %f21445, %f21435, %f30176; fma.rn.ftz.f32 %f21446, %f30177, %f21434, %f21445; fma.rn.ftz.f32 %f30175, %f21436, %f30175, %f21446; bra.uni $L__BB107_2846; $L__BB107_2844: cvta.to.local.u64 %rd2967, %rd299; ld.local.v4.f32 {%f21393, %f21394, %f21395, %f21396}, [%rd2967]; mul.ftz.f32 %f21400, %f30176, %f21394; fma.rn.ftz.f32 %f21401, %f30177, %f21393, %f21400; fma.rn.ftz.f32 %f30171, %f30175, %f21395, %f21401; cvta.to.local.u64 %rd2968, %rd311; ld.local.v4.f32 {%f21402, %f21403, %f21404, %f21405}, [%rd2968]; mul.ftz.f32 %f21409, %f30176, %f21403; fma.rn.ftz.f32 %f21410, %f30177, %f21402, %f21409; fma.rn.ftz.f32 %f30170, %f30175, %f21404, %f21410; cvta.to.local.u64 %rd2969, %rd310; ld.local.v4.f32 {%f21411, %f21412, %f21413, %f21414}, [%rd2969]; mul.ftz.f32 %f21418, %f30176, %f21412; fma.rn.ftz.f32 %f21419, %f30177, %f21411, %f21418; fma.rn.ftz.f32 %f30175, %f30175, %f21413, %f21419; $L__BB107_2846: st.local.f32 [%rd318], %f30171; st.local.f32 [%rd319], %f30170; st.local.f32 [%rd320], %f30175; @%p2585 bra $L__BB107_2848; bra.uni $L__BB107_2847; $L__BB107_2848: ld.global.v4.f32 {%f21474, %f21475, %f21476, %f21477}, [%rd570]; ld.global.v4.f32 {%f21481, %f21482, %f21483, %f21484}, [%rd570+16]; ld.global.v4.f32 {%f21488, %f21489, %f21490, %f21491}, [%rd570+32]; mul.ftz.f32 %f21495, %f21475, %f30173; fma.rn.ftz.f32 %f21496, %f30174, %f21474, %f21495; fma.rn.ftz.f32 %f4831, %f21476, %f30172, %f21496; mul.ftz.f32 %f21497, %f21482, %f30173; fma.rn.ftz.f32 %f21498, %f30174, %f21481, %f21497; fma.rn.ftz.f32 %f4832, %f21483, %f30172, %f21498; mul.ftz.f32 %f21499, %f21489, %f30173; fma.rn.ftz.f32 %f21500, %f30174, %f21488, %f21499; fma.rn.ftz.f32 %f30172, %f21490, %f30172, %f21500; st.local.v2.f32 [%rd573], {%f4831, %f4832}; st.local.f32 [%rd321], %f30172; mov.f32 %f30173, %f4832; mov.f32 %f30174, %f4831; mov.f32 %f30176, %f30170; mov.f32 %f30177, %f30171; bra.uni $L__BB107_2850; $L__BB107_2847: cvta.to.local.u64 %rd2970, %rd299; ld.local.v4.f32 {%f21447, %f21448, %f21449, %f21450}, [%rd2970]; mul.ftz.f32 %f21454, %f30173, %f21448; fma.rn.ftz.f32 %f21455, %f30174, %f21447, %f21454; fma.rn.ftz.f32 %f4828, %f30172, %f21449, %f21455; cvta.to.local.u64 %rd2971, %rd311; ld.local.v4.f32 {%f21456, %f21457, %f21458, %f21459}, [%rd2971]; mul.ftz.f32 %f21463, %f30173, %f21457; fma.rn.ftz.f32 %f21464, %f30174, %f21456, %f21463; fma.rn.ftz.f32 %f4829, %f30172, %f21458, %f21464; cvta.to.local.u64 %rd2972, %rd310; ld.local.v4.f32 {%f21465, %f21466, %f21467, %f21468}, [%rd2972]; mul.ftz.f32 %f21472, %f30173, %f21466; fma.rn.ftz.f32 %f21473, %f30174, %f21465, %f21472; fma.rn.ftz.f32 %f30172, %f30172, %f21467, %f21473; st.local.v2.f32 [%rd573], {%f4828, %f4829}; st.local.f32 [%rd321], %f30172; mov.f32 %f30173, %f4829; mov.f32 %f30174, %f4828; mov.f32 %f30176, %f30170; mov.f32 %f30177, %f30171; $L__BB107_2850: mov.b32 %r4195, %f4721; mov.b64 %rd2975, {%r4195, %r4196}; cvt.u32.u64 %r4197, %rd2975; mov.b32 %f30181, %r4197; setp.eq.s32 %p2604, %r620, -1; @%p2604 bra $L__BB107_2853; mul.ftz.f32 %f21507, %f30183, %f4717; fma.rn.ftz.f32 %f21508, %f30182, %f4716, %f21507; fma.rn.ftz.f32 %f21509, %f4721, %f4718, %f21508; setp.geu.ftz.f32 %p2605, %f21509, 0f00000000; @%p2605 bra $L__BB107_2853; or.b32 %r4198, %r995, 1; st.local.u32 [%rd293], %r4198; neg.ftz.f32 %f30181, %f4721; st.local.f32 [%rd307], %f30181; neg.ftz.f32 %f21510, %f30180; st.local.f32 [%rd315], %f21510; neg.ftz.f32 %f21511, %f30178; neg.ftz.f32 %f21512, %f30179; neg.ftz.f32 %f30183, %f30183; neg.ftz.f32 %f30182, %f30182; st.local.v4.f32 [%rd337], {%f21512, %f21511, %f30182, %f30183}; neg.ftz.f32 %f21513, %f30177; st.local.f32 [%rd318], %f21513; neg.ftz.f32 %f21514, %f30176; st.local.f32 [%rd319], %f21514; neg.ftz.f32 %f21515, %f30175; st.local.f32 [%rd320], %f21515; neg.ftz.f32 %f21516, %f30173; neg.ftz.f32 %f21517, %f30174; cvta.to.local.u64 %rd2976, %rd309; st.local.v2.f32 [%rd2976], {%f21517, %f21516}; neg.ftz.f32 %f21518, %f30172; st.local.f32 [%rd321], %f21518; $L__BB107_2853: mov.u32 %r4199, 0; st.local.u32 [%rd322], %r4199; mov.f32 %f21519, 0f00000000; st.local.v4.f32 [%rd324], {%f21519, %f21519, %f21519, %f21519}; st.local.u32 [%rd336], %r4199; st.local.f32 [%rd280+-12], %f30182; st.local.f32 [%rd280+-8], %f30183; st.local.f32 [%rd280+-4], %f30181; $L__BB107_2855: mov.u32 %r4206, 0; st.local.u32 [%rd342], %r4206; st.local.u32 [%rd343], %r4206; bra.uni $L__BB107_2856; $L__BB107_3008: { // callseq 461, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 461 $L__BB107_2856: mov.u32 %r1005, %r5790; mul.wide.s32 %rd2996, %r1005, 16; add.s64 %rd575, %rd344, %rd2996; ld.global.v4.u32 {%r4208, %r4209, %r4210, %r4211}, [%rd575]; mul.wide.u32 %rd2997, %r4209, 4; add.s64 %rd576, %rd2157, %rd2997; mul.wide.u32 %rd2998, %r4210, 4; add.s64 %rd577, %rd2157, %rd2998; add.s32 %r5790, %r1005, 1; setp.gt.s32 %p2606, %r4208, 36; @%p2606 bra $L__BB107_2905; setp.gt.s32 %p2688, %r4208, 25; @%p2688 bra $L__BB107_2901; bra.uni $L__BB107_2858; $L__BB107_2901: add.s32 %r4213, %r4208, -27; setp.lt.u32 %p2689, %r4213, 9; @%p2689 bra $L__BB107_2856; setp.eq.s32 %p2690, %r4208, 26; @%p2690 bra $L__BB107_3047; setp.eq.s32 %p2691, %r4208, 36; @%p2691 bra $L__BB107_2904; bra.uni $L__BB107_3061; $L__BB107_2904: mov.u32 %r4254, 36; { // callseq 514, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4254; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 514 bra.uni $L__BB107_2856; $L__BB107_2905: setp.gt.s32 %p2607, %r4208, 68; @%p2607 bra $L__BB107_2960; bra.uni $L__BB107_2906; $L__BB107_2960: setp.gt.s32 %p2608, %r4208, 88; @%p2608 bra $L__BB107_2996; bra.uni $L__BB107_2961; $L__BB107_2996: setp.gt.s32 %p2609, %r4208, 93; @%p2609 bra $L__BB107_3001; bra.uni $L__BB107_2997; $L__BB107_3001: setp.gt.s32 %p2610, %r4208, 95; @%p2610 bra $L__BB107_3005; bra.uni $L__BB107_3002; $L__BB107_3005: setp.eq.s32 %p2611, %r4208, 98; @%p2611 bra $L__BB107_3060; setp.eq.s32 %p2612, %r4208, 97; @%p2612 bra $L__BB107_3010; setp.ne.s32 %p2613, %r4208, 96; @%p2613 bra $L__BB107_3061; bra.uni $L__BB107_3008; $L__BB107_2906: setp.gt.s32 %p2646, %r4208, 48; @%p2646 bra $L__BB107_2925; bra.uni $L__BB107_2907; $L__BB107_2925: setp.gt.s32 %p2647, %r4208, 58; @%p2647 bra $L__BB107_2943; setp.gt.s32 %p2661, %r4208, 53; @%p2661 bra $L__BB107_2935; setp.gt.s32 %p2668, %r4208, 50; @%p2668 bra $L__BB107_2931; bra.uni $L__BB107_2928; $L__BB107_2931: setp.eq.s32 %p2669, %r4208, 51; @%p2669 bra $L__BB107_3039; setp.eq.s32 %p2670, %r4208, 52; @%p2670 bra $L__BB107_3038; setp.eq.s32 %p2671, %r4208, 53; @%p2671 bra $L__BB107_2934; bra.uni $L__BB107_3061; $L__BB107_2934: { // callseq 499, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 499 bra.uni $L__BB107_2856; $L__BB107_2858: setp.gt.s32 %p2692, %r4208, 17; @%p2692 bra $L__BB107_2892; bra.uni $L__BB107_2859; $L__BB107_2892: setp.lt.s32 %p2693, %r4208, 21; @%p2693 bra $L__BB107_2856; setp.gt.s32 %p2694, %r4208, 22; @%p2694 bra $L__BB107_2897; bra.uni $L__BB107_2894; $L__BB107_2897: setp.eq.s32 %p2695, %r4208, 23; @%p2695 bra $L__BB107_3049; setp.eq.s32 %p2696, %r4208, 24; @%p2696 bra $L__BB107_3048; setp.eq.s32 %p2697, %r4208, 25; @%p2697 bra $L__BB107_2900; bra.uni $L__BB107_3061; $L__BB107_2900: add.s32 %r4256, %r1005, 1; { // callseq 516, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4256; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 516 bra.uni $L__BB107_2856; $L__BB107_2961: setp.gt.s32 %p2619, %r4208, 78; @%p2619 bra $L__BB107_2979; setp.gt.s32 %p2633, %r4208, 73; @%p2633 bra $L__BB107_2971; setp.gt.s32 %p2640, %r4208, 70; @%p2640 bra $L__BB107_2967; bra.uni $L__BB107_2964; $L__BB107_2967: setp.eq.s32 %p2641, %r4208, 71; @%p2641 bra $L__BB107_3027; setp.eq.s32 %p2642, %r4208, 72; @%p2642 bra $L__BB107_3026; setp.eq.s32 %p2643, %r4208, 73; @%p2643 bra $L__BB107_2970; bra.uni $L__BB107_3061; $L__BB107_2970: add.s32 %r4226, %r1005, 1; { // callseq 479, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4226; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 479 bra.uni $L__BB107_2856; $L__BB107_2907: setp.gt.s32 %p2674, %r4208, 41; @%p2674 bra $L__BB107_2913; setp.gt.s32 %p2684, %r4208, 39; @%p2684 bra $L__BB107_2856; setp.eq.s32 %p2685, %r4208, 37; @%p2685 bra $L__BB107_3046; setp.eq.s32 %p2686, %r4208, 38; @%p2686 bra $L__BB107_3045; setp.eq.s32 %p2687, %r4208, 39; @%p2687 bra $L__BB107_2912; bra.uni $L__BB107_3061; $L__BB107_2912: mov.u32 %r4252, 39; { // callseq 511, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4252, %r4209, %r4210, %r4211}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 511 bra.uni $L__BB107_2856; $L__BB107_2859: setp.gt.s32 %p2700, %r4208, 8; @%p2700 bra $L__BB107_2875; bra.uni $L__BB107_2860; $L__BB107_2875: setp.gt.s32 %p2701, %r4208, 12; @%p2701 bra $L__BB107_2884; bra.uni $L__BB107_2876; $L__BB107_2884: setp.gt.s32 %p2702, %r4208, 14; @%p2702 bra $L__BB107_2888; bra.uni $L__BB107_2885; $L__BB107_2888: setp.eq.s32 %p2703, %r4208, 15; @%p2703 bra $L__BB107_3052; setp.eq.s32 %p2704, %r4208, 16; @%p2704 bra $L__BB107_3051; setp.eq.s32 %p2705, %r4208, 17; @%p2705 bra $L__BB107_2891; bra.uni $L__BB107_3061; $L__BB107_2891: { // callseq 521, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; .param .b32 param4; st.param.b32 [param4+0], %r4211; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 521 bra.uni $L__BB107_2856; $L__BB107_2943: setp.gt.s32 %p2648, %r4208, 63; @%p2648 bra $L__BB107_2952; setp.gt.s32 %p2655, %r4208, 60; @%p2655 bra $L__BB107_2948; bra.uni $L__BB107_2945; $L__BB107_2948: setp.eq.s32 %p2656, %r4208, 61; @%p2656 bra $L__BB107_3033; setp.eq.s32 %p2657, %r4208, 62; @%p2657 bra $L__BB107_3032; setp.eq.s32 %p2658, %r4208, 63; @%p2658 bra $L__BB107_2951; bra.uni $L__BB107_3061; $L__BB107_2951: mov.u32 %r4236, 63; { // callseq 489, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4236; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 489 bra.uni $L__BB107_2856; $L__BB107_2979: setp.gt.s32 %p2620, %r4208, 83; @%p2620 bra $L__BB107_2988; setp.gt.s32 %p2627, %r4208, 80; @%p2627 bra $L__BB107_2984; bra.uni $L__BB107_2981; $L__BB107_2984: setp.eq.s32 %p2628, %r4208, 81; @%p2628 bra $L__BB107_3015; setp.eq.s32 %p2629, %r4208, 82; @%p2629 bra $L__BB107_3014; setp.eq.s32 %p2630, %r4208, 83; @%p2630 bra $L__BB107_2987; bra.uni $L__BB107_3061; $L__BB107_2987: { // callseq 471, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 471 bra.uni $L__BB107_2856; $L__BB107_2860: setp.gt.s32 %p2713, %r4208, 4; @%p2713 bra $L__BB107_2868; setp.gt.s32 %p2719, %r4208, 2; @%p2719 bra $L__BB107_2865; setp.eq.s32 %p2722, %r4208, 1; mov.u32 %r5790, %r4209; @%p2722 bra $L__BB107_2856; setp.eq.s32 %p2723, %r4208, 2; @%p2723 bra $L__BB107_2864; bra.uni $L__BB107_3061; $L__BB107_2864: mov.u32 %r4280, 2; add.s32 %r4281, %r1005, 1; { // callseq 530, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r4280; st.param.b32 [param0+4], %r4209; st.param.b32 [param0+8], %r4210; st.param.b32 [param0+12], %r4211; .param .b32 param1; st.param.b32 [param1+0], %r4281; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 530 bra.uni $L__BB107_2856; $L__BB107_2913: setp.gt.s32 %p2675, %r4208, 44; @%p2675 bra $L__BB107_2918; bra.uni $L__BB107_2914; $L__BB107_2918: setp.gt.s32 %p2676, %r4208, 46; @%p2676 bra $L__BB107_2922; setp.eq.s32 %p2679, %r4208, 45; @%p2679 bra $L__BB107_3042; setp.eq.s32 %p2680, %r4208, 46; @%p2680 bra $L__BB107_2921; bra.uni $L__BB107_3061; $L__BB107_2921: { // callseq 506, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 506 bra.uni $L__BB107_2856; $L__BB107_3002: setp.eq.s32 %p2614, %r4208, 94; @%p2614 bra $L__BB107_3009; setp.eq.s32 %p2615, %r4208, 95; @%p2615 bra $L__BB107_3004; bra.uni $L__BB107_3061; $L__BB107_3004: { // callseq 462, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 462 bra.uni $L__BB107_2856; $L__BB107_2997: add.s32 %r4212, %r4208, -92; setp.lt.u32 %p2616, %r4212, 2; @%p2616 bra $L__BB107_2856; setp.eq.s32 %p2617, %r4208, 90; @%p2617 bra $L__BB107_2856; setp.eq.s32 %p2618, %r4208, 89; @%p2618 bra $L__BB107_3000; bra.uni $L__BB107_3061; $L__BB107_3000: mov.u32 %r4216, 89; { // callseq 465, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4216; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 465 bra.uni $L__BB107_2856; $L__BB107_2935: setp.gt.s32 %p2662, %r4208, 55; @%p2662 bra $L__BB107_2939; bra.uni $L__BB107_2936; $L__BB107_2939: setp.eq.s32 %p2663, %r4208, 56; @%p2663 bra $L__BB107_3036; setp.eq.s32 %p2664, %r4208, 57; @%p2664 bra $L__BB107_3035; setp.eq.s32 %p2665, %r4208, 58; @%p2665 bra $L__BB107_2942; bra.uni $L__BB107_3061; $L__BB107_2942: mov.u32 %r4243, 58; { // callseq 494, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4243; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 494 bra.uni $L__BB107_2856; $L__BB107_2971: setp.gt.s32 %p2634, %r4208, 75; @%p2634 bra $L__BB107_2975; bra.uni $L__BB107_2972; $L__BB107_2975: setp.eq.s32 %p2635, %r4208, 76; @%p2635 bra $L__BB107_3019; setp.eq.s32 %p2636, %r4208, 77; @%p2636 bra $L__BB107_3017; setp.eq.s32 %p2637, %r4208, 78; @%p2637 bra $L__BB107_2978; bra.uni $L__BB107_3061; $L__BB107_2978: add.s32 %r4224, %r1005, 1; { // callseq 476, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4224; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 476 bra.uni $L__BB107_2856; $L__BB107_2952: setp.gt.s32 %p2649, %r4208, 65; @%p2649 bra $L__BB107_2956; bra.uni $L__BB107_2953; $L__BB107_2956: setp.eq.s32 %p2650, %r4208, 66; @%p2650 bra $L__BB107_3030; setp.eq.s32 %p2651, %r4208, 67; @%p2651 bra $L__BB107_3029; setp.eq.s32 %p2652, %r4208, 68; @%p2652 bra $L__BB107_2959; bra.uni $L__BB107_3061; $L__BB107_2959: mov.u32 %r4231, 68; { // callseq 484, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4231; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 484 bra.uni $L__BB107_2856; $L__BB107_2988: setp.gt.s32 %p2621, %r4208, 85; @%p2621 bra $L__BB107_2992; bra.uni $L__BB107_2989; $L__BB107_2992: setp.eq.s32 %p2622, %r4208, 86; @%p2622 bra $L__BB107_3012; setp.eq.s32 %p2623, %r4208, 87; @%p2623 bra $L__BB107_3011; setp.eq.s32 %p2624, %r4208, 88; @%p2624 bra $L__BB107_2995; bra.uni $L__BB107_3061; $L__BB107_2995: mov.u32 %r4217, 88; ld.local.f32 %f21521, [%rd282]; ld.local.v2.f32 {%f21522, %f21523}, [%rd282+4]; { // callseq 466, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f21521; .param .b32 param1; st.param.f32 [param1+0], %f21522; .param .b32 param2; st.param.f32 [param2+0], %f21523; .param .b64 param3; st.param.b64 [param3+0], %rd2156; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r4217; st.param.b32 [param4+4], %r4209; st.param.b32 [param4+8], %r4210; st.param.b32 [param4+12], %r4211; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 466 bra.uni $L__BB107_2856; $L__BB107_2876: setp.gt.s32 %p2708, %r4208, 10; @%p2708 bra $L__BB107_2881; setp.eq.s32 %p2711, %r4208, 9; @%p2711 bra $L__BB107_3055; setp.eq.s32 %p2712, %r4208, 10; @%p2712 bra $L__BB107_2879; bra.uni $L__BB107_3061; $L__BB107_2879: ld.local.f32 %f21537, [%rd577]; setp.ltu.ftz.f32 %p2728, %f21537, 0f3F800000; @%p2728 bra $L__BB107_2856; add.s32 %r4274, %r4209, %r1005; add.s32 %r5790, %r4274, 1; bra.uni $L__BB107_2856; $L__BB107_2868: setp.gt.s32 %p2714, %r4208, 6; @%p2714 bra $L__BB107_2872; setp.eq.s32 %p2717, %r4208, 5; @%p2717 bra $L__BB107_3058; setp.eq.s32 %p2718, %r4208, 6; @%p2718 bra $L__BB107_2871; bra.uni $L__BB107_3061; $L__BB107_2871: ld.local.f32 %f21539, [%rd576]; ld.local.f32 %f21540, [%rd576+4]; ld.local.f32 %f21541, [%rd576+8]; st.local.f32 [%rd348], %f21539; st.local.f32 [%rd349], %f21540; st.local.f32 [%rd350], %f21541; bra.uni $L__BB107_2856; $L__BB107_2914: setp.eq.s32 %p2681, %r4208, 42; @%p2681 bra $L__BB107_3044; setp.eq.s32 %p2682, %r4208, 43; @%p2682 bra $L__BB107_3043; setp.eq.s32 %p2683, %r4208, 44; @%p2683 bra $L__BB107_2917; bra.uni $L__BB107_3061; $L__BB107_2917: mov.u32 %r4249, 44; add.s32 %r4250, %r1005, 1; { // callseq 508, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4249; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4250; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 508 bra.uni $L__BB107_2856; $L__BB107_3060: { // callseq 531, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 531 bra.uni $L__BB107_2856; $L__BB107_3010: { // callseq 464, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 464 bra.uni $L__BB107_2856; $L__BB107_3009: mov.u32 %r4214, 94; add.s32 %r4215, %r1005, 1; { // callseq 463, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4214; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4215; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 463 bra.uni $L__BB107_2856; $L__BB107_2894: setp.eq.s32 %p2698, %r4208, 21; @%p2698 bra $L__BB107_3050; setp.eq.s32 %p2699, %r4208, 22; @%p2699 bra $L__BB107_2896; bra.uni $L__BB107_3061; $L__BB107_2896: mov.u32 %r4260, 22; add.s32 %r4261, %r1005, 1; { // callseq 519, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4260; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4261; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 519 bra.uni $L__BB107_2856; $L__BB107_2885: setp.eq.s32 %p2706, %r4208, 13; @%p2706 bra $L__BB107_3053; setp.eq.s32 %p2707, %r4208, 14; @%p2707 bra $L__BB107_2887; bra.uni $L__BB107_3061; $L__BB107_2887: st.local.u32 [%rd577], %r4209; bra.uni $L__BB107_2856; $L__BB107_2865: setp.eq.s32 %p2720, %r4208, 3; @%p2720 bra $L__BB107_3059; setp.eq.s32 %p2721, %r4208, 4; @%p2721 bra $L__BB107_2867; bra.uni $L__BB107_3061; $L__BB107_2867: mov.u32 %r4278, 4; { // callseq 528, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4278, %r4209, %r4210, %r4211}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 528 bra.uni $L__BB107_2856; $L__BB107_2881: setp.eq.s32 %p2709, %r4208, 11; @%p2709 bra $L__BB107_3054; setp.eq.s32 %p2710, %r4208, 12; @%p2710 bra $L__BB107_2883; bra.uni $L__BB107_3061; $L__BB107_2883: { // callseq 524, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 524 bra.uni $L__BB107_2856; $L__BB107_2872: setp.eq.s32 %p2715, %r4208, 7; @%p2715 bra $L__BB107_3057; setp.eq.s32 %p2716, %r4208, 8; @%p2716 bra $L__BB107_2874; bra.uni $L__BB107_3061; $L__BB107_2874: mov.u32 %r4276, 8; { // callseq 526, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4276; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 526 bra.uni $L__BB107_2856; $L__BB107_2928: setp.eq.s32 %p2672, %r4208, 49; @%p2672 bra $L__BB107_3040; setp.eq.s32 %p2673, %r4208, 50; @%p2673 bra $L__BB107_2930; bra.uni $L__BB107_3061; $L__BB107_2930: { // callseq 502, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 502 bra.uni $L__BB107_2856; $L__BB107_2964: setp.eq.s32 %p2644, %r4208, 69; @%p2644 bra $L__BB107_3028; setp.eq.s32 %p2645, %r4208, 70; @%p2645 bra $L__BB107_2966; bra.uni $L__BB107_3061; $L__BB107_2966: mov.u32 %r4228, 70; { // callseq 482, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4228; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 482 bra.uni $L__BB107_2856; $L__BB107_2945: setp.eq.s32 %p2659, %r4208, 59; @%p2659 bra $L__BB107_3034; setp.eq.s32 %p2660, %r4208, 60; @%p2660 bra $L__BB107_2947; bra.uni $L__BB107_3061; $L__BB107_2947: add.s32 %r4241, %r1005, 1; { // callseq 492, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4241; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 492 bra.uni $L__BB107_2856; $L__BB107_2981: setp.eq.s32 %p2631, %r4208, 79; @%p2631 bra $L__BB107_3016; setp.eq.s32 %p2632, %r4208, 80; @%p2632 bra $L__BB107_2983; bra.uni $L__BB107_3061; $L__BB107_2983: { // callseq 474, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 474 bra.uni $L__BB107_2856; $L__BB107_2922: setp.eq.s32 %p2677, %r4208, 47; @%p2677 bra $L__BB107_3041; setp.eq.s32 %p2678, %r4208, 48; @%p2678 bra $L__BB107_2924; bra.uni $L__BB107_3061; $L__BB107_2924: { // callseq 504, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 504 bra.uni $L__BB107_2856; $L__BB107_2936: setp.eq.s32 %p2666, %r4208, 54; @%p2666 bra $L__BB107_3037; setp.eq.s32 %p2667, %r4208, 55; @%p2667 bra $L__BB107_2938; bra.uni $L__BB107_3061; $L__BB107_2938: ld.local.v4.f32 {%f21530, %f21531, %f21532, %f21533}, [%rd2173]; { // callseq 497, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f21530; st.param.f32 [param0+4], %f21531; st.param.f32 [param0+8], %f21532; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; .param .b32 param4; st.param.b32 [param4+0], %r4211; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 497 bra.uni $L__BB107_2856; $L__BB107_2972: setp.eq.s32 %p2638, %r4208, 74; @%p2638 bra $L__BB107_3025; setp.eq.s32 %p2639, %r4208, 75; @%p2639 bra $L__BB107_2974; bra.uni $L__BB107_3061; $L__BB107_2974: { // callseq 477, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 477 bra.uni $L__BB107_2856; $L__BB107_2953: setp.eq.s32 %p2653, %r4208, 64; @%p2653 bra $L__BB107_3031; setp.eq.s32 %p2654, %r4208, 65; @%p2654 bra $L__BB107_2955; bra.uni $L__BB107_3061; $L__BB107_2955: { // callseq 487, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 487 bra.uni $L__BB107_2856; $L__BB107_2989: setp.eq.s32 %p2625, %r4208, 84; @%p2625 bra $L__BB107_3013; setp.eq.s32 %p2626, %r4208, 85; @%p2626 bra $L__BB107_2991; bra.uni $L__BB107_3061; $L__BB107_2991: add.s32 %r4220, %r1005, 1; { // callseq 469, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4220; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 469 bra.uni $L__BB107_2856; $L__BB107_3049: mov.u32 %r4258, 23; add.s32 %r4259, %r1005, 1; { // callseq 518, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4258; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4259; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 518 bra.uni $L__BB107_2856; $L__BB107_3048: mov.u32 %r4257, 24; { // callseq 517, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4257; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 517 bra.uni $L__BB107_2856; $L__BB107_3052: ld.global.v4.u32 {%r4264, %r4265, %r4266, %r4267}, [%rd575+16]; add.s32 %r5790, %r1005, 2; st.local.u32 [%rd576], %r4265; st.local.u32 [%rd576+4], %r4266; st.local.u32 [%rd576+8], %r4267; bra.uni $L__BB107_2856; $L__BB107_3051: mov.u32 %r4263, 16; { // callseq 522, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4263; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 522 bra.uni $L__BB107_2856; $L__BB107_3055: ld.local.f32 %f21538, [%rd577]; setp.gtu.ftz.f32 %p2729, %f21538, 0f00000000; @%p2729 bra $L__BB107_2856; add.s32 %r4275, %r4209, %r1005; add.s32 %r5790, %r4275, 1; bra.uni $L__BB107_2856; $L__BB107_3058: st.local.u32 [%rd348], %r4209; st.local.u32 [%rd349], %r4210; st.local.u32 [%rd350], %r4211; bra.uni $L__BB107_2856; $L__BB107_3050: mov.u32 %r4262, 21; { // callseq 520, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4262; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 520 bra.uni $L__BB107_2856; $L__BB107_3053: mov.u32 %r4271, 13; add.s32 %r4272, %r1005, 1; mov.u32 %r4273, 524288; { // callseq 523, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r4273; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r4271; st.param.b32 [param3+4], %r4209; st.param.b32 [param3+8], %r4210; st.param.b32 [param3+12], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4272; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 523 bra.uni $L__BB107_2856; $L__BB107_3059: mov.u32 %r4279, 3; { // callseq 529, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4279, %r4209, %r4210, %r4211}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 529 bra.uni $L__BB107_2856; $L__BB107_3054: { // callseq 525, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 525 bra.uni $L__BB107_2856; $L__BB107_3057: mov.u32 %r4277, 7; { // callseq 527, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4277; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 527 bra.uni $L__BB107_2856; $L__BB107_3039: { // callseq 501, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 501 bra.uni $L__BB107_2856; $L__BB107_3038: add.s32 %r4248, %r1005, 1; { // callseq 500, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4248; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 500 bra.uni $L__BB107_2856; $L__BB107_3027: mov.u32 %r4227, 71; { // callseq 481, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4227; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 481 bra.uni $L__BB107_2856; $L__BB107_3026: { // callseq 480, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 480 bra.uni $L__BB107_2856; $L__BB107_3046: mov.u32 %r4253, 37; { // callseq 513, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4253, %r4209, %r4210, %r4211}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 513 bra.uni $L__BB107_2856; $L__BB107_3045: { // callseq 512, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; .param .b32 param4; st.param.b32 [param4+0], %r4211; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 512 bra.uni $L__BB107_2856; $L__BB107_3033: mov.u32 %r4239, 61; add.s32 %r4240, %r1005, 1; { // callseq 491, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4239; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4240; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 491 bra.uni $L__BB107_2856; $L__BB107_3032: mov.u32 %r4237, 62; add.s32 %r4238, %r1005, 1; { // callseq 490, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4237; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4238; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 490 bra.uni $L__BB107_2856; $L__BB107_3015: mov.u32 %r4222, 81; { // callseq 473, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4222; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 473 bra.uni $L__BB107_2856; $L__BB107_3014: mov.u32 %r4221, 82; { // callseq 472, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4221; st.param.b32 [param2+4], %r4209; st.param.b32 [param2+8], %r4210; st.param.b32 [param2+12], %r4211; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 472 bra.uni $L__BB107_2856; $L__BB107_3042: { // callseq 507, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 507 bra.uni $L__BB107_2856; $L__BB107_3036: mov.u32 %r4246, 56; { // callseq 496, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4246; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 496 bra.uni $L__BB107_2856; $L__BB107_3035: mov.u32 %r4244, 57; add.s32 %r4245, %r1005, 1; { // callseq 495, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4244; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4245; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 495 bra.uni $L__BB107_2856; $L__BB107_3019: setp.eq.s32 %p2725, %r4211, 255; @%p2725 bra $L__BB107_2856; add.s64 %rd578, %rd576, 4; add.s64 %rd579, %rd576, 8; setp.eq.s32 %p2726, %r4210, 0; mul.wide.u32 %rd3023, %r4211, 4; add.s64 %rd580, %rd2157, %rd3023; @%p2726 bra $L__BB107_3024; setp.eq.s32 %p2727, %r4210, 1; @%p2727 bra $L__BB107_3023; bra.uni $L__BB107_3022; $L__BB107_3023: ld.local.f32 %f21528, [%rd578]; st.local.f32 [%rd580], %f21528; bra.uni $L__BB107_2856; $L__BB107_3017: setp.eq.s32 %p2724, %r4211, 255; @%p2724 bra $L__BB107_2856; ld.local.f32 %f21526, [%rd576]; add.s32 %r4225, %r4211, %r4210; mul.wide.u32 %rd3019, %r4225, 4; add.s64 %rd3020, %rd2157, %rd3019; st.local.f32 [%rd3020], %f21526; bra.uni $L__BB107_2856; $L__BB107_3030: add.s32 %r4233, %r1005, 1; { // callseq 486, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4233; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 486 bra.uni $L__BB107_2856; $L__BB107_3029: mov.u32 %r4232, 67; { // callseq 485, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4232, %r4209, %r4210, %r4211}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 485 bra.uni $L__BB107_2856; $L__BB107_3012: add.s32 %r4219, %r1005, 1; { // callseq 468, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4219; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 468 bra.uni $L__BB107_2856; $L__BB107_3011: add.s32 %r4218, %r1005, 1; { // callseq 467, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4218; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 467 bra.uni $L__BB107_2856; $L__BB107_3044: { // callseq 510, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 510 bra.uni $L__BB107_2856; $L__BB107_3043: add.s32 %r4251, %r1005, 1; { // callseq 509, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4251; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 509 bra.uni $L__BB107_2856; $L__BB107_3040: { // callseq 503, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4209; .param .b32 param3; st.param.b32 [param3+0], %r4210; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 503 bra.uni $L__BB107_2856; $L__BB107_3028: mov.u32 %r4229, 69; add.s32 %r4230, %r1005, 1; { // callseq 483, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4229; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4230; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 483 bra.uni $L__BB107_2856; $L__BB107_3034: add.s32 %r4242, %r1005, 1; { // callseq 493, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4242; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 493 bra.uni $L__BB107_2856; $L__BB107_3016: add.s32 %r4223, %r1005, 1; { // callseq 475, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; .param .b32 param4; st.param.b32 [param4+0], %r4223; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 475 bra.uni $L__BB107_2856; $L__BB107_3041: { // callseq 505, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1602; .param .b64 param1; st.param.b64 [param1+0], %rd2172; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .b32 param3; st.param.b32 [param3+0], %r4209; .param .b32 param4; st.param.b32 [param4+0], %r4210; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 505 bra.uni $L__BB107_2856; $L__BB107_3037: add.s32 %r4247, %r1005, 1; { // callseq 498, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4247; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 498 bra.uni $L__BB107_2856; $L__BB107_3025: { // callseq 478, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; .param .b32 param3; st.param.b32 [param3+0], %r4211; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 478 bra.uni $L__BB107_2856; $L__BB107_3031: mov.u32 %r4234, 64; add.s32 %r4235, %r1005, 1; { // callseq 488, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4234; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; .param .b32 param2; st.param.b32 [param2+0], %r4235; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5790, [retval0+0]; } // callseq 488 bra.uni $L__BB107_2856; $L__BB107_3013: { // callseq 470, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4209; .param .b32 param2; st.param.b32 [param2+0], %r4210; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 470 bra.uni $L__BB107_2856; $L__BB107_3047: mov.u32 %r4255, 26; { // callseq 515, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4255; st.param.b32 [param1+4], %r4209; st.param.b32 [param1+8], %r4210; st.param.b32 [param1+12], %r4211; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 515 bra.uni $L__BB107_2856; $L__BB107_3024: ld.local.f32 %f21529, [%rd576]; st.local.f32 [%rd580], %f21529; bra.uni $L__BB107_2856; $L__BB107_3022: ld.local.f32 %f21527, [%rd579]; st.local.f32 [%rd580], %f21527; bra.uni $L__BB107_2856; $L__BB107_3061: ld.local.u32 %r4282, [%rd293]; and.b32 %r1095, %r4282, 2; @%p2582 bra $L__BB107_3064; bra.uni $L__BB107_3062; $L__BB107_3064: setp.eq.s32 %p2733, %r1095, 0; mov.f32 %f30190, 0f00000000; mov.f32 %f30191, %f30190; mov.f32 %f30192, %f30190; @%p2733 bra $L__BB107_3066; add.s64 %rd3107, %rd2172, 280; cvta.to.local.u64 %rd3108, %rd3107; ld.local.v2.f32 {%f30190, %f30191}, [%rd3108]; add.s64 %rd3109, %rd2172, 288; cvta.to.local.u64 %rd3110, %rd3109; ld.local.f32 %f30192, [%rd3110]; bra.uni $L__BB107_3066; $L__BB107_3062: setp.eq.s32 %p2731, %r1095, 0; mov.f32 %f30190, 0f00000000; mov.f32 %f30191, %f30190; mov.f32 %f30192, %f30190; @%p2731 bra $L__BB107_3066; cvta.to.local.u64 %rd3099, %rd302; ld.local.v2.f32 {%f21545, %f21546}, [%rd3099]; ld.local.v4.f32 {%f21549, %f21550, %f21551, %f21552}, [%rd307]; mul.ftz.f32 %f21557, %f21546, %f21551; fma.rn.ftz.f32 %f21558, %f21545, %f21550, %f21557; fma.rn.ftz.f32 %f21559, %f21549, %f21552, %f21558; abs.ftz.f32 %f21560, %f21559; setp.gt.ftz.f32 %p2732, %f21560, 0f00000000; selp.f32 %f21561, 0f3F800000, 0f00000000, %p2732; add.s64 %rd3102, %rd2172, 280; cvta.to.local.u64 %rd3103, %rd3102; ld.local.v2.f32 {%f21562, %f21563}, [%rd3103]; add.s64 %rd3104, %rd2172, 288; cvta.to.local.u64 %rd3105, %rd3104; ld.local.f32 %f21566, [%rd3105]; mul.ftz.f32 %f30192, %f21561, %f21566; mul.ftz.f32 %f30191, %f21561, %f21563; mul.ftz.f32 %f30190, %f21562, %f21561; $L__BB107_3066: ld.local.f32 %f21574, [%rd280+28]; mul.ftz.f32 %f21575, %f30190, %f21574; mul.ftz.f32 %f21576, %f30191, %f21574; mul.ftz.f32 %f21577, %f30192, %f21574; ld.global.f32 %f21578, [%rd284+-12]; mul.ftz.f32 %f21579, %f21575, %f21578; ld.global.f32 %f21580, [%rd284+-8]; mul.ftz.f32 %f21581, %f21576, %f21580; ld.global.f32 %f21582, [%rd284+-4]; mul.ftz.f32 %f21583, %f21577, %f21582; div.approx.ftz.f32 %f21584, %f21579, %f4702; div.approx.ftz.f32 %f21585, %f21581, %f4702; div.approx.ftz.f32 %f21586, %f21583, %f4702; mul.ftz.f32 %f30298, %f21584, %f30126; mul.ftz.f32 %f4874, %f21585, %f30125; mul.ftz.f32 %f4875, %f21586, %f30124; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f4874; st.local.f32 [%rd283+8], %f4875; mul.ftz.f32 %f4876, %f21584, %f30129; mul.ftz.f32 %f4877, %f21585, %f30128; mul.ftz.f32 %f4878, %f21586, %f30127; st.local.f32 [%rd283+12], %f4876; st.local.f32 [%rd283+16], %f4877; st.local.f32 [%rd283+20], %f4878; mul.ftz.f32 %f4879, %f21584, %f29810; mul.ftz.f32 %f4880, %f21585, %f29809; mul.ftz.f32 %f4881, %f21586, %f29808; st.local.f32 [%rd283+24], %f4879; st.local.f32 [%rd283+28], %f4880; st.local.f32 [%rd283+32], %f4881; ld.local.v4.f32 {%f21587, %f21588, %f21589, %f21590}, [%rd412+124]; mul.ftz.f32 %f21591, %f21587, %f21590; mul.ftz.f32 %f21592, %f21588, %f21589; sub.ftz.f32 %f4886, %f21591, %f21592; abs.ftz.f32 %f21593, %f4886; setp.lt.ftz.f32 %p2734, %f21593, 0f38D1B717; mov.f32 %f30197, 0f00000000; @%p2734 bra $L__BB107_3068; rcp.approx.ftz.f32 %f21594, %f4886; mul.ftz.f32 %f30202, %f21590, %f21594; mul.ftz.f32 %f21595, %f21588, %f21594; neg.ftz.f32 %f30203, %f21595; mul.ftz.f32 %f21596, %f21589, %f21594; neg.ftz.f32 %f30204, %f21596; mul.ftz.f32 %f30205, %f21587, %f21594; mov.f32 %f30197, %f4886; $L__BB107_3068: setp.eq.ftz.f32 %p2736, %f30197, 0f00000000; mov.pred %p3926, -1; @%p2736 bra $L__BB107_3090; setp.lt.s32 %p2737, %r5732, 2; @%p2737 bra $L__BB107_3075; mov.u32 %r5791, 1; mov.f32 %f21598, 0f00000000; $L__BB107_3071: add.s32 %r4284, %r5791, -1; mul.wide.s32 %rd3112, %r4284, 192; add.s64 %rd3113, %rd1, %rd3112; ld.local.v4.f32 {%f21599, %f21600, %f21601, %f21602}, [%rd3113+176]; mul.ftz.f32 %f21607, %f30203, %f21601; mul.ftz.f32 %f21608, %f30203, %f21602; mul.ftz.f32 %f21609, %f30205, %f21601; mul.ftz.f32 %f21610, %f30205, %f21602; mul.wide.s32 %rd3114, %r4284, 16; add.s64 %rd3115, %rd2157, %rd3114; fma.rn.ftz.f32 %f21611, %f30204, %f21600, %f21610; fma.rn.ftz.f32 %f21612, %f30204, %f21599, %f21609; fma.rn.ftz.f32 %f21613, %f30202, %f21600, %f21608; fma.rn.ftz.f32 %f21614, %f30202, %f21599, %f21607; st.local.v4.f32 [%rd3115], {%f21614, %f21613, %f21612, %f21611}; ld.local.v4.f32 {%f21615, %f21616, %f21617, %f21618}, [%rd3113+336]; mul.ftz.f32 %f21623, %f21616, %f21612; fma.rn.ftz.f32 %f21624, %f21615, %f21614, %f21623; mul.ftz.f32 %f21625, %f21616, %f21611; fma.rn.ftz.f32 %f21626, %f21615, %f21613, %f21625; mul.ftz.f32 %f21627, %f21618, %f21612; fma.rn.ftz.f32 %f21628, %f21617, %f21614, %f21627; mul.ftz.f32 %f21629, %f21618, %f21611; fma.rn.ftz.f32 %f21630, %f21617, %f21613, %f21629; ld.local.v4.f32 {%f21631, %f21632, %f21633, %f21634}, [%rd3113+352]; sub.ftz.f32 %f4900, %f21631, %f21624; sub.ftz.f32 %f4901, %f21632, %f21626; sub.ftz.f32 %f4902, %f21633, %f21628; sub.ftz.f32 %f4903, %f21634, %f21630; mul.ftz.f32 %f21639, %f4900, %f4903; mul.ftz.f32 %f21640, %f4901, %f4902; sub.ftz.f32 %f4904, %f21639, %f21640; abs.ftz.f32 %f21641, %f4904; setp.lt.ftz.f32 %p2738, %f21641, 0f38D1B717; mov.f32 %f30206, %f21598; @%p2738 bra $L__BB107_3073; rcp.approx.ftz.f32 %f21642, %f4904; mul.ftz.f32 %f30202, %f4903, %f21642; mul.ftz.f32 %f21643, %f4901, %f21642; neg.ftz.f32 %f30203, %f21643; mul.ftz.f32 %f21644, %f4902, %f21642; neg.ftz.f32 %f30204, %f21644; mul.ftz.f32 %f30205, %f4900, %f21642; mov.f32 %f30206, %f4904; $L__BB107_3073: add.s32 %r5791, %r5791, 1; setp.eq.ftz.f32 %p2740, %f30206, 0f00000000; @%p2740 bra $L__BB107_3090; setp.lt.s32 %p2741, %r5791, %r5732; @%p2741 bra $L__BB107_3071; $L__BB107_3075: ld.local.f32 %f4918, [%rd413+436]; ld.local.f32 %f21646, [%rd413+412]; ld.local.f32 %f4919, [%rd413+440]; ld.local.f32 %f21647, [%rd413+416]; mul.ftz.f32 %f21648, %f21647, %f4919; fma.rn.ftz.f32 %f21649, %f21646, %f4918, %f21648; ld.local.f32 %f4920, [%rd413+444]; ld.local.f32 %f21650, [%rd413+420]; fma.rn.ftz.f32 %f21651, %f21650, %f4920, %f21649; mul.ftz.f32 %f21652, %f4918, %f21651; mul.ftz.f32 %f21653, %f4919, %f21651; mul.ftz.f32 %f21654, %f21651, %f4920; sub.ftz.f32 %f21655, %f21646, %f21652; sub.ftz.f32 %f21656, %f21647, %f21653; sub.ftz.f32 %f21657, %f21650, %f21654; mul.ftz.f32 %f21658, %f21656, %f21656; fma.rn.ftz.f32 %f21659, %f21655, %f21655, %f21658; fma.rn.ftz.f32 %f21660, %f21657, %f21657, %f21659; sqrt.approx.ftz.f32 %f21661, %f21660; div.approx.ftz.f32 %f4921, %f21655, %f21661; div.approx.ftz.f32 %f4922, %f21656, %f21661; div.approx.ftz.f32 %f4923, %f21657, %f21661; mul.ftz.f32 %f21662, %f4923, %f4919; mul.ftz.f32 %f21663, %f4922, %f4920; sub.ftz.f32 %f4924, %f21662, %f21663; mul.ftz.f32 %f21664, %f4921, %f4920; mul.ftz.f32 %f21665, %f4923, %f4918; sub.ftz.f32 %f4925, %f21664, %f21665; mul.ftz.f32 %f21666, %f4922, %f4918; mul.ftz.f32 %f21667, %f4921, %f4919; sub.ftz.f32 %f4926, %f21666, %f21667; ld.local.f32 %f4927, [%rd280+-8]; ld.local.f32 %f4928, [%rd280+-12]; setp.neu.ftz.f32 %p2742, %f4928, %f4927; ld.local.f32 %f4929, [%rd280+-4]; setp.neu.ftz.f32 %p2743, %f4928, %f4929; or.pred %p2744, %p2742, %p2743; sub.ftz.f32 %f21668, %f4929, %f4927; neg.ftz.f32 %f21669, %f4927; neg.ftz.f32 %f21670, %f4929; selp.f32 %f21671, %f21670, %f4929, %p2744; add.ftz.f32 %f21672, %f4928, %f21671; selp.f32 %f21673, %f4927, %f21669, %p2744; sub.ftz.f32 %f21674, %f21673, %f4928; mul.ftz.f32 %f21675, %f21672, %f21672; fma.rn.ftz.f32 %f21676, %f21668, %f21668, %f21675; fma.rn.ftz.f32 %f21677, %f21674, %f21674, %f21676; sqrt.approx.ftz.f32 %f21678, %f21677; div.approx.ftz.f32 %f4930, %f21668, %f21678; div.approx.ftz.f32 %f4931, %f21672, %f21678; div.approx.ftz.f32 %f4932, %f21674, %f21678; setp.eq.s32 %p2745, %r5732, 1; @%p2745 bra $L__BB107_3077; bra.uni $L__BB107_3076; $L__BB107_3077: ld.local.v4.f32 {%f21682, %f21683, %f21684, %f21685}, [%rd279]; sub.ftz.f32 %f30213, %f21682, %f4662; ld.local.f32 %f30212, [%rd413+404]; sub.ftz.f32 %f30214, %f21683, %f30212; ld.local.f32 %f30211, [%rd413+408]; sub.ftz.f32 %f30215, %f21684, %f30211; bra.uni $L__BB107_3078; $L__BB107_3076: ld.local.f32 %f21679, [%rd413+208]; sub.ftz.f32 %f30213, %f21679, %f4662; ld.local.f32 %f30212, [%rd413+404]; ld.local.f32 %f21680, [%rd413+212]; sub.ftz.f32 %f30214, %f21680, %f30212; ld.local.f32 %f30211, [%rd413+408]; ld.local.f32 %f21681, [%rd413+216]; sub.ftz.f32 %f30215, %f21681, %f30211; $L__BB107_3078: mul.ftz.f32 %f21689, %f4929, %f4931; mul.ftz.f32 %f21690, %f4927, %f4932; sub.ftz.f32 %f4948, %f21690, %f21689; mul.ftz.f32 %f21691, %f4928, %f4932; mul.ftz.f32 %f21692, %f4929, %f4930; sub.ftz.f32 %f4949, %f21692, %f21691; mul.ftz.f32 %f21693, %f4927, %f4930; mul.ftz.f32 %f21694, %f4928, %f4931; sub.ftz.f32 %f4950, %f21694, %f21693; mul.ftz.f32 %f21695, %f30214, %f30214; fma.rn.ftz.f32 %f21696, %f30213, %f30213, %f21695; fma.rn.ftz.f32 %f21697, %f30215, %f30215, %f21696; rsqrt.approx.ftz.f32 %f21698, %f21697; mul.ftz.f32 %f4951, %f30213, %f21698; mul.ftz.f32 %f4952, %f30214, %f21698; mul.ftz.f32 %f4953, %f30215, %f21698; ld.local.f32 %f30216, [%rd413+508]; ld.local.f32 %f21699, [%rd413+448]; ld.local.f32 %f21700, [%rd413+452]; mul.ftz.f32 %f21701, %f4952, %f21700; fma.rn.ftz.f32 %f21702, %f4951, %f21699, %f21701; ld.local.f32 %f21703, [%rd413+456]; fma.rn.ftz.f32 %f21704, %f4953, %f21703, %f21702; setp.geu.ftz.f32 %p2746, %f21704, 0f00000000; @%p2746 bra $L__BB107_3080; rcp.approx.ftz.f32 %f30216, %f30216; $L__BB107_3080: setp.eq.ftz.f32 %p2747, %f30145, 0f7F7FFFFF; @%p2747 bra $L__BB107_3082; bra.uni $L__BB107_3081; $L__BB107_3082: fma.rn.ftz.f32 %f21790, %f30216, %f30148, %f4951; fma.rn.ftz.f32 %f21791, %f30216, %f30147, %f4952; fma.rn.ftz.f32 %f21792, %f30216, %f30146, %f4953; mul.ftz.f32 %f21793, %f21791, %f21791; fma.rn.ftz.f32 %f21794, %f21790, %f21790, %f21793; fma.rn.ftz.f32 %f21795, %f21792, %f21792, %f21794; rsqrt.approx.ftz.f32 %f21796, %f21795; mov.f32 %f21797, 0f3F800000; mul.ftz.f32 %f21798, %f21790, %f21796; neg.ftz.f32 %f21799, %f21798; mul.ftz.f32 %f21800, %f21791, %f21796; mul.ftz.f32 %f21801, %f21792, %f21796; mul.ftz.f32 %f21802, %f30216, %f21796; neg.ftz.f32 %f21803, %f21802; mul.ftz.f32 %f21804, %f30147, %f4919; fma.rn.ftz.f32 %f21805, %f30148, %f4918, %f21804; fma.rn.ftz.f32 %f21806, %f30146, %f4920, %f21805; mul.ftz.f32 %f21807, %f21806, %f21806; sub.ftz.f32 %f21808, %f21797, %f21807; mov.f32 %f21809, 0f00000000; max.ftz.f32 %f21810, %f21808, %f21809; sqrt.approx.ftz.f32 %f21811, %f21810; mul.ftz.f32 %f21812, %f4922, %f30147; fma.rn.ftz.f32 %f21813, %f4921, %f30148, %f21812; fma.rn.ftz.f32 %f21814, %f4923, %f30146, %f21813; mul.ftz.f32 %f21815, %f21814, %f21814; sub.ftz.f32 %f21816, %f21797, %f21815; max.ftz.f32 %f21817, %f21816, %f21809; sqrt.approx.ftz.f32 %f21818, %f21817; mul.ftz.f32 %f21819, %f4924, %f21818; mul.ftz.f32 %f21820, %f4925, %f21818; mul.ftz.f32 %f21821, %f4926, %f21818; fma.rn.ftz.f32 %f21822, %f4921, %f21814, %f21819; fma.rn.ftz.f32 %f21823, %f4922, %f21814, %f21820; fma.rn.ftz.f32 %f21824, %f4923, %f21814, %f21821; mul.ftz.f32 %f21825, %f21806, %f21822; mul.ftz.f32 %f21826, %f21806, %f21823; mul.ftz.f32 %f21827, %f21806, %f21824; mul.ftz.f32 %f21828, %f21811, %f4918; mul.ftz.f32 %f21829, %f21811, %f4919; mul.ftz.f32 %f21830, %f21811, %f4920; sub.ftz.f32 %f21831, %f21825, %f21828; sub.ftz.f32 %f21832, %f21826, %f21829; sub.ftz.f32 %f21833, %f21827, %f21830; mul.ftz.f32 %f21834, %f21831, %f21803; mul.ftz.f32 %f21835, %f21832, %f21803; mul.ftz.f32 %f21836, %f21833, %f21803; mul.ftz.f32 %f21837, %f21811, %f21803; mul.ftz.f32 %f21838, %f4921, %f21818; mul.ftz.f32 %f21839, %f4922, %f21818; mul.ftz.f32 %f21840, %f4923, %f21818; mul.ftz.f32 %f21841, %f4924, %f21814; mul.ftz.f32 %f21842, %f4925, %f21814; mul.ftz.f32 %f21843, %f4926, %f21814; sub.ftz.f32 %f21844, %f21841, %f21838; sub.ftz.f32 %f21845, %f21842, %f21839; sub.ftz.f32 %f21846, %f21843, %f21840; mul.ftz.f32 %f21847, %f21837, %f21844; mul.ftz.f32 %f21848, %f21837, %f21845; mul.ftz.f32 %f21849, %f21837, %f21846; mul.ftz.f32 %f21850, %f21834, %f21799; mul.ftz.f32 %f21851, %f21835, %f21800; sub.ftz.f32 %f21852, %f21850, %f21851; mul.ftz.f32 %f21853, %f21836, %f21801; sub.ftz.f32 %f21854, %f21852, %f21853; fma.rn.ftz.f32 %f21855, %f21854, %f21798, %f21834; fma.rn.ftz.f32 %f21856, %f21854, %f21800, %f21835; fma.rn.ftz.f32 %f21857, %f21854, %f21801, %f21836; mul.ftz.f32 %f21858, %f21847, %f21799; mul.ftz.f32 %f21859, %f21848, %f21800; sub.ftz.f32 %f21860, %f21858, %f21859; mul.ftz.f32 %f21861, %f21849, %f21801; sub.ftz.f32 %f21862, %f21860, %f21861; fma.rn.ftz.f32 %f21863, %f21862, %f21798, %f21847; fma.rn.ftz.f32 %f21864, %f21862, %f21800, %f21848; fma.rn.ftz.f32 %f21865, %f21862, %f21801, %f21849; mul.ftz.f32 %f21866, %f4922, %f21856; fma.rn.ftz.f32 %f21867, %f4921, %f21855, %f21866; mul.ftz.f32 %f21868, %f4922, %f21864; fma.rn.ftz.f32 %f21869, %f4921, %f21863, %f21868; mul.ftz.f32 %f21870, %f4925, %f21856; fma.rn.ftz.f32 %f21871, %f4924, %f21855, %f21870; mul.ftz.f32 %f21872, %f4925, %f21864; fma.rn.ftz.f32 %f21873, %f4924, %f21863, %f21872; fma.rn.ftz.f32 %f30221, %f4926, %f21865, %f21873; fma.rn.ftz.f32 %f30219, %f4923, %f21865, %f21869; fma.rn.ftz.f32 %f30220, %f4926, %f21857, %f21871; fma.rn.ftz.f32 %f30218, %f4923, %f21857, %f21867; abs.ftz.f32 %f21874, %f21811; mov.f32 %f21875, 0f3A83126F; max.ftz.f32 %f21876, %f21875, %f21874; rcp.approx.ftz.f32 %f30217, %f21876; bra.uni $L__BB107_3083; $L__BB107_3081: ld.local.f32 %f21706, [%rd280+-24]; sub.ftz.f32 %f21707, %f21706, %f4662; ld.local.f32 %f21708, [%rd280+-20]; sub.ftz.f32 %f21709, %f21708, %f30212; ld.local.f32 %f21710, [%rd280+-16]; sub.ftz.f32 %f21711, %f21710, %f30211; mul.ftz.f32 %f21712, %f21709, %f21709; fma.rn.ftz.f32 %f21713, %f21707, %f21707, %f21712; fma.rn.ftz.f32 %f21714, %f21711, %f21711, %f21713; rsqrt.approx.ftz.f32 %f21715, %f21714; mov.f32 %f30217, 0f3F800000; mul.ftz.f32 %f21716, %f21707, %f21715; mul.ftz.f32 %f21717, %f21709, %f21715; mul.ftz.f32 %f21718, %f21711, %f21715; fma.rn.ftz.f32 %f21719, %f30216, %f21716, %f4951; fma.rn.ftz.f32 %f21720, %f30216, %f21717, %f4952; fma.rn.ftz.f32 %f21721, %f30216, %f21718, %f4953; mul.ftz.f32 %f21722, %f21720, %f21720; fma.rn.ftz.f32 %f21723, %f21719, %f21719, %f21722; fma.rn.ftz.f32 %f21724, %f21721, %f21721, %f21723; rsqrt.approx.ftz.f32 %f21725, %f21724; mul.ftz.f32 %f21726, %f21725, %f21719; neg.ftz.f32 %f21727, %f21726; mul.ftz.f32 %f21728, %f21725, %f21720; mul.ftz.f32 %f21729, %f21725, %f21721; mul.ftz.f32 %f21730, %f30216, %f21725; mul.ftz.f32 %f21731, %f21715, %f21730; mul.ftz.f32 %f21732, %f4931, %f21717; fma.rn.ftz.f32 %f21733, %f4930, %f21716, %f21732; fma.rn.ftz.f32 %f21734, %f4932, %f21718, %f21733; mul.ftz.f32 %f21735, %f21716, %f21734; mul.ftz.f32 %f21736, %f21717, %f21734; mul.ftz.f32 %f21737, %f21718, %f21734; sub.ftz.f32 %f21738, %f4930, %f21735; sub.ftz.f32 %f21739, %f4931, %f21736; sub.ftz.f32 %f21740, %f4932, %f21737; mul.ftz.f32 %f21741, %f21731, %f21738; mul.ftz.f32 %f21742, %f21731, %f21739; mul.ftz.f32 %f21743, %f21731, %f21740; mul.ftz.f32 %f21744, %f4949, %f21717; fma.rn.ftz.f32 %f21745, %f4948, %f21716, %f21744; fma.rn.ftz.f32 %f21746, %f4950, %f21718, %f21745; mul.ftz.f32 %f21747, %f21716, %f21746; mul.ftz.f32 %f21748, %f21717, %f21746; mul.ftz.f32 %f21749, %f21718, %f21746; sub.ftz.f32 %f21750, %f4948, %f21747; sub.ftz.f32 %f21751, %f4949, %f21748; sub.ftz.f32 %f21752, %f4950, %f21749; mul.ftz.f32 %f21753, %f21731, %f21750; mul.ftz.f32 %f21754, %f21731, %f21751; mul.ftz.f32 %f21755, %f21731, %f21752; mul.ftz.f32 %f21756, %f21741, %f21727; mul.ftz.f32 %f21757, %f21742, %f21728; sub.ftz.f32 %f21758, %f21756, %f21757; mul.ftz.f32 %f21759, %f21743, %f21729; sub.ftz.f32 %f21760, %f21758, %f21759; fma.rn.ftz.f32 %f21761, %f21760, %f21726, %f21741; fma.rn.ftz.f32 %f21762, %f21760, %f21728, %f21742; fma.rn.ftz.f32 %f21763, %f21760, %f21729, %f21743; mul.ftz.f32 %f21764, %f21753, %f21727; mul.ftz.f32 %f21765, %f21754, %f21728; sub.ftz.f32 %f21766, %f21764, %f21765; mul.ftz.f32 %f21767, %f21755, %f21729; sub.ftz.f32 %f21768, %f21766, %f21767; fma.rn.ftz.f32 %f21769, %f21768, %f21726, %f21753; fma.rn.ftz.f32 %f21770, %f21768, %f21728, %f21754; fma.rn.ftz.f32 %f21771, %f21768, %f21729, %f21755; neg.ftz.f32 %f21772, %f21761; neg.ftz.f32 %f21773, %f21769; mul.ftz.f32 %f21774, %f4921, %f21772; mul.ftz.f32 %f21775, %f4922, %f21762; sub.ftz.f32 %f21776, %f21774, %f21775; mul.ftz.f32 %f21777, %f4923, %f21763; mul.ftz.f32 %f21778, %f4921, %f21773; mul.ftz.f32 %f21779, %f4922, %f21770; sub.ftz.f32 %f21780, %f21778, %f21779; mul.ftz.f32 %f21781, %f4923, %f21771; mul.ftz.f32 %f21782, %f4924, %f21772; mul.ftz.f32 %f21783, %f4925, %f21762; sub.ftz.f32 %f21784, %f21782, %f21783; mul.ftz.f32 %f21785, %f4926, %f21763; mul.ftz.f32 %f21786, %f4924, %f21773; mul.ftz.f32 %f21787, %f4925, %f21770; sub.ftz.f32 %f21788, %f21786, %f21787; mul.ftz.f32 %f21789, %f4926, %f21771; sub.ftz.f32 %f30221, %f21788, %f21789; sub.ftz.f32 %f30220, %f21784, %f21785; sub.ftz.f32 %f30219, %f21780, %f21781; sub.ftz.f32 %f30218, %f21776, %f21777; $L__BB107_3083: mul.ftz.f32 %f21877, %f30203, %f30220; fma.rn.ftz.f32 %f21878, %f30202, %f30218, %f21877; mul.ftz.f32 %f21879, %f30203, %f30221; fma.rn.ftz.f32 %f21880, %f30202, %f30219, %f21879; mul.ftz.f32 %f21881, %f30205, %f30220; fma.rn.ftz.f32 %f21882, %f30204, %f30218, %f21881; mul.ftz.f32 %f21883, %f30205, %f30221; fma.rn.ftz.f32 %f21884, %f30204, %f30219, %f21883; neg.ftz.f32 %f30240, %f21882; neg.ftz.f32 %f30238, %f21878; neg.ftz.f32 %f30241, %f21884; neg.ftz.f32 %f30239, %f21880; @%p2737 bra $L__BB107_3089; setp.eq.s32 %p2749, %r5793, 0; mov.u32 %r5794, %r806; @%p2749 bra $L__BB107_3087; mov.u32 %r5794, %r806; $L__BB107_3086: .pragma "nounroll"; mul.wide.s32 %rd3117, %r5794, 16; add.s64 %rd3118, %rd2157, %rd3117; ld.local.v4.f32 {%f21893, %f21894, %f21895, %f21896}, [%rd3118]; mul.ftz.f32 %f21901, %f30240, %f21894; fma.rn.ftz.f32 %f21902, %f30238, %f21893, %f21901; mul.ftz.f32 %f21903, %f30241, %f21894; fma.rn.ftz.f32 %f21904, %f30239, %f21893, %f21903; mul.ftz.f32 %f21905, %f30240, %f21896; fma.rn.ftz.f32 %f21906, %f30238, %f21895, %f21905; mul.ftz.f32 %f21907, %f30241, %f21896; fma.rn.ftz.f32 %f21908, %f30239, %f21895, %f21907; neg.ftz.f32 %f30241, %f21908; neg.ftz.f32 %f30240, %f21906; neg.ftz.f32 %f30239, %f21904; neg.ftz.f32 %f30238, %f21902; add.s32 %r5794, %r5794, -1; add.s32 %r5793, %r5793, -1; setp.ne.s32 %p2750, %r5793, 0; @%p2750 bra $L__BB107_3086; $L__BB107_3087: setp.lt.u32 %p2751, %r806, 3; @%p2751 bra $L__BB107_3089; $L__BB107_3088: mul.wide.s32 %rd3120, %r5794, 16; add.s64 %rd3121, %rd2157, %rd3120; ld.local.v4.f32 {%f21909, %f21910, %f21911, %f21912}, [%rd3121]; mul.ftz.f32 %f21917, %f30240, %f21910; fma.rn.ftz.f32 %f21918, %f30238, %f21909, %f21917; mul.ftz.f32 %f21919, %f30241, %f21910; fma.rn.ftz.f32 %f21920, %f30239, %f21909, %f21919; mul.ftz.f32 %f21921, %f30240, %f21912; fma.rn.ftz.f32 %f21922, %f30238, %f21911, %f21921; mul.ftz.f32 %f21923, %f30241, %f21912; fma.rn.ftz.f32 %f21924, %f30239, %f21911, %f21923; neg.ftz.f32 %f21925, %f21918; neg.ftz.f32 %f21926, %f21920; ld.local.v4.f32 {%f21927, %f21928, %f21929, %f21930}, [%rd3121+-16]; mul.ftz.f32 %f21935, %f21927, %f21925; mul.ftz.f32 %f21936, %f21928, %f21922; sub.ftz.f32 %f21937, %f21935, %f21936; mul.ftz.f32 %f21938, %f21927, %f21926; mul.ftz.f32 %f21939, %f21928, %f21924; sub.ftz.f32 %f21940, %f21938, %f21939; mul.ftz.f32 %f21941, %f21929, %f21925; mul.ftz.f32 %f21942, %f21930, %f21922; sub.ftz.f32 %f21943, %f21941, %f21942; mul.ftz.f32 %f21944, %f21929, %f21926; mul.ftz.f32 %f21945, %f21930, %f21924; sub.ftz.f32 %f21946, %f21944, %f21945; neg.ftz.f32 %f21947, %f21937; neg.ftz.f32 %f21948, %f21940; ld.local.v4.f32 {%f21949, %f21950, %f21951, %f21952}, [%rd3121+-32]; mul.ftz.f32 %f21957, %f21949, %f21947; mul.ftz.f32 %f21958, %f21950, %f21943; sub.ftz.f32 %f21959, %f21957, %f21958; mul.ftz.f32 %f21960, %f21949, %f21948; mul.ftz.f32 %f21961, %f21950, %f21946; sub.ftz.f32 %f21962, %f21960, %f21961; mul.ftz.f32 %f21963, %f21951, %f21947; mul.ftz.f32 %f21964, %f21952, %f21943; sub.ftz.f32 %f21965, %f21963, %f21964; mul.ftz.f32 %f21966, %f21951, %f21948; mul.ftz.f32 %f21967, %f21952, %f21946; sub.ftz.f32 %f21968, %f21966, %f21967; neg.ftz.f32 %f21969, %f21959; neg.ftz.f32 %f21970, %f21962; ld.local.v4.f32 {%f21971, %f21972, %f21973, %f21974}, [%rd3121+-48]; mul.ftz.f32 %f21979, %f21971, %f21969; mul.ftz.f32 %f21980, %f21972, %f21965; sub.ftz.f32 %f21981, %f21979, %f21980; mul.ftz.f32 %f21982, %f21971, %f21970; mul.ftz.f32 %f21983, %f21972, %f21968; sub.ftz.f32 %f21984, %f21982, %f21983; mul.ftz.f32 %f21985, %f21973, %f21969; mul.ftz.f32 %f21986, %f21974, %f21965; sub.ftz.f32 %f21987, %f21985, %f21986; mul.ftz.f32 %f21988, %f21973, %f21970; mul.ftz.f32 %f21989, %f21974, %f21968; sub.ftz.f32 %f21990, %f21988, %f21989; neg.ftz.f32 %f30241, %f21990; neg.ftz.f32 %f30240, %f21987; neg.ftz.f32 %f30239, %f21984; neg.ftz.f32 %f30238, %f21981; add.s32 %r1105, %r5794, -4; setp.gt.s32 %p2752, %r5794, 3; mov.u32 %r5794, %r1105; @%p2752 bra $L__BB107_3088; $L__BB107_3089: mul.ftz.f32 %f21991, %f30238, %f30241; mul.ftz.f32 %f21992, %f30239, %f30240; sub.ftz.f32 %f21993, %f21991, %f21992; abs.ftz.f32 %f21994, %f21993; mul.ftz.f32 %f30242, %f30217, %f21994; mov.pred %p3926, 0; $L__BB107_3090: @%p3926 bra $L__BB107_3392; ld.local.f32 %f21995, [%rd412]; ld.local.v2.f32 {%f21996, %f21997}, [%rd412+4]; mul.ftz.f32 %f22000, %f30259, %f21996; fma.rn.ftz.f32 %f22001, %f30260, %f21995, %f22000; fma.rn.ftz.f32 %f22002, %f30258, %f21997, %f22001; abs.ftz.f32 %f22003, %f22002; mul.ftz.f32 %f22004, %f30257, %f30257; div.approx.ftz.f32 %f22005, %f22003, %f22004; mul.ftz.f32 %f22006, %f30242, %f22005; mov.f32 %f22007, 0f40000000; min.ftz.f32 %f22008, %f22006, %f22007; mul.ftz.f32 %f30298, %f22008, %f30298; mul.ftz.f32 %f30250, %f22008, %f4874; mul.ftz.f32 %f30249, %f22008, %f4875; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f30250; st.local.f32 [%rd283+8], %f30249; mul.ftz.f32 %f30248, %f22008, %f4876; mul.ftz.f32 %f30247, %f22008, %f4877; mul.ftz.f32 %f30246, %f22008, %f4878; st.local.f32 [%rd283+12], %f30248; st.local.f32 [%rd283+16], %f30247; st.local.f32 [%rd283+20], %f30246; mul.ftz.f32 %f30245, %f22008, %f4879; mul.ftz.f32 %f30244, %f22008, %f4880; mul.ftz.f32 %f30243, %f22008, %f4881; st.local.f32 [%rd283+24], %f30245; st.local.f32 [%rd283+28], %f30244; st.local.f32 [%rd283+32], %f30243; add.u64 %rd3122, %SP, 0; add.u64 %rd3123, %SPL, 0; add.s64 %rd584, %rd3123, 48; st.local.u32 [%rd3123+48], %r619; st.local.u32 [%rd3123+44], %r620; mov.u32 %r5796, 0; st.local.u32 [%rd3123+24], %r5796; ld.local.f32 %f22009, [%rd5+4]; st.local.f32 [%rd3123+52], %f22009; st.local.u32 [%rd3123+56], %r5796; ld.local.f32 %f22010, [%rd5+-4]; st.local.f32 [%rd3123+32], %f22010; ld.local.u32 %r4288, [%rd5+-12]; st.local.u32 [%rd3123+40], %r4288; ld.local.u32 %r4289, [%rd5+-28]; st.local.u32 [%rd3123+36], %r4289; ld.local.v4.f32 {%f30256, %f30254, %f30252, %f22014}, [%rd279]; st.local.f32 [%rd3123], %f30256; st.local.f32 [%rd3123+4], %f30254; st.local.f32 [%rd3123+8], %f30252; @%p1589 bra $L__BB107_3391; cvt.u32.u16 %r4291, %rs253; add.s32 %r1106, %r4291, 1; mov.u16 %rs542, 0; setp.eq.ftz.f32 %p2935, %f30145, 0f7F7FFFFF; bra.uni $L__BB107_3093; $L__BB107_4546: ld.local.f32 %f30255, [%rd586+196]; ld.local.f32 %f30253, [%rd586+200]; mov.f32 %f30258, %f5177; mov.f32 %f30259, %f5176; mov.f32 %f30260, %f5175; $L__BB107_3093: cvt.s64.s32 %rd585, %r5796; mul.wide.s32 %rd3124, %r5796, 192; add.s64 %rd586, %rd1, %rd3124; ld.local.f32 %f22015, [%rd586]; sub.ftz.f32 %f22016, %f22015, %f30256; sub.ftz.f32 %f22017, %f30255, %f30254; mul.ftz.f32 %f22018, %f22017, %f22017; fma.rn.ftz.f32 %f22019, %f22016, %f22016, %f22018; sub.ftz.f32 %f22020, %f30253, %f30252; fma.rn.ftz.f32 %f22021, %f22020, %f22020, %f22019; sqrt.approx.ftz.f32 %f5038, %f22021; st.local.f32 [%rd584+-20], %f5038; rsqrt.approx.ftz.f32 %f22022, %f22021; mul.ftz.f32 %f5039, %f22016, %f22022; mul.ftz.f32 %f5040, %f22017, %f22022; mul.ftz.f32 %f5041, %f22020, %f22022; st.local.f32 [%rd584+-36], %f5039; st.local.f32 [%rd584+-32], %f5040; st.local.f32 [%rd584+-28], %f5041; mov.b32 %r1108, %f30256; setp.nan.ftz.f32 %p2755, %f30256, %f30256; mov.u16 %rs723, %rs542; @%p2755 bra $L__BB107_3096; or.b32 %r4292, %r1108, -2147483648; setp.ne.s32 %p2756, %r4292, -2147483648; add.ftz.f32 %f22023, %f30256, %f30256; setp.eq.ftz.f32 %p2757, %f22023, %f30256; and.pred %p2758, %p2757, %p2756; mov.u16 %rs723, %rs542; @%p2758 bra $L__BB107_3096; shl.b32 %r4293, %r1108, 1; setp.lt.u32 %p2759, %r4293, -16777215; selp.u16 %rs723, 1, 0, %p2759; $L__BB107_3096: setp.eq.s16 %p2760, %rs723, 0; mov.u16 %rs725, %rs542; @%p2760 bra $L__BB107_3102; mov.b32 %r1109, %f5039; setp.nan.ftz.f32 %p2761, %f5039, %f5039; mov.u16 %rs725, 0; mov.u16 %rs724, %rs725; @%p2761 bra $L__BB107_3100; or.b32 %r4294, %r1109, -2147483648; setp.ne.s32 %p2762, %r4294, -2147483648; add.ftz.f32 %f22024, %f5039, %f5039; setp.eq.ftz.f32 %p2763, %f22024, %f5039; and.pred %p2764, %p2763, %p2762; mov.u16 %rs724, %rs725; @%p2764 bra $L__BB107_3100; shl.b32 %r4295, %r1109, 1; setp.lt.u32 %p2765, %r4295, -16777215; selp.u16 %rs724, 1, 0, %p2765; $L__BB107_3100: setp.eq.s16 %p2766, %rs724, 0; @%p2766 bra $L__BB107_3102; mul.ftz.f32 %f22025, %f5040, %f5040; fma.rn.ftz.f32 %f22026, %f5039, %f5039, %f22025; fma.rn.ftz.f32 %f22027, %f5041, %f5041, %f22026; setp.neu.ftz.f32 %p2767, %f22027, 0f00000000; selp.u16 %rs725, 1, 0, %p2767; $L__BB107_3102: setp.eq.s16 %p2768, %rs725, 0; mov.u16 %rs726, %rs542; @%p2768 bra $L__BB107_3110; @%p1593 bra $L__BB107_3107; @%p1969 bra $L__BB107_3106; bra.uni $L__BB107_3105; $L__BB107_3106: mov.u32 %r4298, 4; { // callseq 533, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd3122; .param .b64 param1; st.param.b64 [param1+0], %rd2582; .param .b32 param2; st.param.b32 [param2+0], %r4298; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r4299, [retval0+0]; } // callseq 533 cvt.u16.u32 %rs726, %r4299; bra.uni $L__BB107_3110; $L__BB107_3107: @%p1969 bra $L__BB107_3109; bra.uni $L__BB107_3108; $L__BB107_3109: mov.u32 %r4302, 4; { // callseq 535, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd3122; .param .b64 param1; st.param.b64 [param1+0], %rd2582; .param .b32 param2; st.param.b32 [param2+0], %r4302; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r4303, [retval0+0]; } // callseq 535 cvt.u16.u32 %rs726, %r4303; bra.uni $L__BB107_3110; $L__BB107_3105: mov.u32 %r4296, 4; { // callseq 532, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd3122; .param .b64 param1; st.param.b64 [param1+0], %rd2582; .param .b32 param2; st.param.b32 [param2+0], %r4296; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r4297, [retval0+0]; } // callseq 532 cvt.u16.u32 %rs726, %r4297; bra.uni $L__BB107_3110; $L__BB107_3108: mov.u32 %r4300, 4; { // callseq 534, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd3122; .param .b64 param1; st.param.b64 [param1+0], %rd2582; .param .b32 param2; st.param.b32 [param2+0], %r4300; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r4301, [retval0+0]; } // callseq 534 cvt.u16.u32 %rs726, %r4301; $L__BB107_3110: setp.eq.s16 %p2772, %rs726, 0; @%p2772 bra $L__BB107_3115; ld.local.u32 %r5798, [%rd2583+16]; setp.ne.s32 %p2773, %r5798, -1; @%p2773 bra $L__BB107_3113; ld.local.u32 %r4304, [%rd2583+12]; mul.wide.s32 %rd3134, %r4304, 4; add.s64 %rd3135, %rd353, %rd3134; ld.global.u32 %r5798, [%rd3135]; $L__BB107_3113: ld.local.u32 %r4306, [%rd586+96]; setp.ne.s32 %p2774, %r5798, %r4306; mov.u32 %r5802, 0; @%p2774 bra $L__BB107_3392; ld.local.f32 %f22028, [%rd2583]; sub.ftz.f32 %f22029, %f5038, %f22028; abs.ftz.f32 %f22030, %f22029; setp.gt.ftz.f32 %p2775, %f22030, 0f3A83126F; @%p2775 bra $L__BB107_3392; bra.uni $L__BB107_3116; $L__BB107_3115: ld.local.u32 %r5798, [%rd586+96]; $L__BB107_3116: st.local.u32 [%rd584+-8], %r5798; ld.local.u32 %r4308, [%rd586+100]; cvt.s64.s32 %rd588, %r4308; st.local.u32 [%rd584+-12], %r4308; ld.local.v4.f32 {%f30256, %f30254, %f30252, %f22034}, [%rd586]; st.local.f32 [%rd584+-48], %f30256; st.local.f32 [%rd584+-44], %f30254; st.local.f32 [%rd584+-40], %f30252; ld.local.f32 %f5045, [%rd586+36]; ld.local.f32 %f5046, [%rd586+40]; ld.local.f32 %f5047, [%rd586+44]; ld.local.u32 %r1115, [%rd586+104]; ld.local.v2.f32 {%f22035, %f22036}, [%rd586+88]; ld.local.f32 %f5050, [%rd5+-4]; st.local.v4.f32 [%rd2173], {%f30256, %f30254, %f30252, %f5045}; st.local.v4.f32 [%rd337], {%f5046, %f5047, %f5045, %f5046}; neg.ftz.f32 %f5051, %f30260; neg.ftz.f32 %f5052, %f30259; neg.ftz.f32 %f5053, %f30258; st.local.v4.f32 [%rd307], {%f5047, %f5051, %f5052, %f5053}; setp.ne.s32 %p2776, %r4308, -1; selp.u32 %r4309, 1, 0, %p2776; mov.b32 %r4310, %f22036; mov.b32 %r4311, %f22035; st.local.v4.u32 [%rd354], {%r4309, %r4311, %r4310, %r5798}; mov.b32 %r4312, %f5050; cvta.to.local.u64 %rd3139, %rd292; st.local.v2.u32 [%rd3139], {%r3028, %r4312}; st.local.f32 [%rd288], %f30257; and.b32 %r5800, %r1115, 4194303; shl.b32 %r4314, %r1115, 5; cvt.u64.u32 %rd3140, %r4314; and.b64 %rd3141, %rd3140, 134217696; add.s64 %rd3142, %rd303, %rd3141; ld.global.u32 %r1117, [%rd3142+16]; mov.u32 %r4315, 0; st.local.v4.u32 [%rd305], {%r1115, %r1117, %r4315, %r4308}; setp.eq.s32 %p2777, %r5798, -1; @%p2777 bra $L__BB107_3141; mul.wide.s32 %rd3144, %r5798, 4; add.s64 %rd3145, %rd287, %rd3144; ld.global.u32 %r1118, [%rd3145]; st.local.u32 [%rd290], %r1118; and.b32 %r1119, %r1118, 2; setp.eq.s32 %p2778, %r1119, 0; mul.wide.s32 %rd3146, %r5798, 208; add.s64 %rd591, %rd300, %rd3146; @%p2778 bra $L__BB107_3124; ld.global.u32 %rd3147, [%rd591+168]; ld.global.u32 %r4316, [%rd591+152]; shl.b32 %r4317, %r4316, 1; cvt.rn.f32.s32 %f22037, %r4317; mul.ftz.f32 %f22038, %f5050, %f22037; cvt.rzi.ftz.s32.f32 %r4318, %f22038; add.s32 %r4319, %r4317, -1; min.s32 %r4320, %r4318, %r4319; cvt.rn.f32.s32 %f22039, %r4320; sub.ftz.f32 %f5054, %f22038, %f22039; cvt.s64.s32 %rd3148, %r4320; add.s64 %rd3149, %rd3148, %rd3147; shl.b64 %rd3150, %rd3149, 6; add.s64 %rd592, %rd301, %rd3150; ld.global.v4.f32 {%f22040, %f22041, %f22042, %f22043}, [%rd592]; ld.global.v4.f32 {%f22044, %f22045, %f22046, %f22047}, [%rd592+64]; mul.ftz.f32 %f22048, %f22041, %f22045; fma.rn.ftz.f32 %f22049, %f22040, %f22044, %f22048; mul.ftz.f32 %f22050, %f22043, %f22047; fma.rn.ftz.f32 %f22051, %f22042, %f22046, %f22050; add.ftz.f32 %f5063, %f22049, %f22051; setp.gt.ftz.f32 %p2779, %f5063, 0f3F7FDF3B; mov.f32 %f22052, 0f3F800000; sub.ftz.f32 %f5064, %f22052, %f5054; @%p2779 bra $L__BB107_3120; bra.uni $L__BB107_3119; $L__BB107_3120: mul.ftz.f32 %f22106, %f22040, %f5064; fma.rn.ftz.f32 %f22107, %f22044, %f5054, %f22106; mul.ftz.f32 %f22108, %f22041, %f5064; fma.rn.ftz.f32 %f22109, %f22045, %f5054, %f22108; mul.ftz.f32 %f22110, %f22042, %f5064; fma.rn.ftz.f32 %f22111, %f22046, %f5054, %f22110; mul.ftz.f32 %f22112, %f22043, %f5064; fma.rn.ftz.f32 %f22113, %f22047, %f5054, %f22112; mul.ftz.f32 %f22114, %f22109, %f22109; fma.rn.ftz.f32 %f22115, %f22107, %f22107, %f22114; mul.ftz.f32 %f22116, %f22113, %f22113; fma.rn.ftz.f32 %f22117, %f22111, %f22111, %f22116; add.ftz.f32 %f22118, %f22115, %f22117; rsqrt.approx.ftz.f32 %f22119, %f22118; mul.ftz.f32 %f30264, %f22113, %f22119; mul.ftz.f32 %f30263, %f22111, %f22119; mul.ftz.f32 %f30262, %f22109, %f22119; mul.ftz.f32 %f30261, %f22107, %f22119; bra.uni $L__BB107_3121; $L__BB107_3141: mov.u32 %r4335, 0; st.local.u32 [%rd318], %r4335; st.local.u32 [%rd319], %r4335; st.local.u32 [%rd320], %r4335; cvta.to.local.u64 %rd3180, %rd309; mov.f32 %f30281, 0f00000000; st.local.v2.f32 [%rd3180], {%f30281, %f30281}; st.local.u32 [%rd321], %r4335; mov.f32 %f30282, %f30281; mov.f32 %f30283, %f30281; mov.f32 %f30284, %f30281; mov.f32 %f30285, %f30281; mov.f32 %f30286, %f30281; mov.f32 %f30287, %f5047; mov.f32 %f30288, %f5046; mov.f32 %f30289, %f5045; bra.uni $L__BB107_3142; $L__BB107_3119: mov.f32 %f22053, 0fBF800000; max.ftz.f32 %f22054, %f5063, %f22053; min.ftz.f32 %f22056, %f22054, %f22052; abs.ftz.f32 %f22057, %f22056; sub.ftz.f32 %f22058, %f22052, %f22057; mul.ftz.f32 %f22059, %f22058, 0f3F000000; sqrt.approx.ftz.f32 %f22060, %f22059; setp.gt.ftz.f32 %p2780, %f22057, 0f3F11EB85; selp.f32 %f22061, %f22060, %f22057, %p2780; mul.ftz.f32 %f22062, %f22061, %f22061; mov.f32 %f22063, 0f3C94D2E9; mov.f32 %f22064, 0f3D53F941; fma.rn.ftz.f32 %f22065, %f22064, %f22062, %f22063; mov.f32 %f22066, 0f3D3F841F; fma.rn.ftz.f32 %f22067, %f22065, %f22062, %f22066; mov.f32 %f22068, 0f3D994929; fma.rn.ftz.f32 %f22069, %f22067, %f22062, %f22068; mov.f32 %f22070, 0f3E2AAB94; fma.rn.ftz.f32 %f22071, %f22069, %f22062, %f22070; mul.ftz.f32 %f22072, %f22062, %f22071; fma.rn.ftz.f32 %f22073, %f22072, %f22061, %f22061; add.ftz.f32 %f22074, %f22073, %f22073; mov.f32 %f22075, 0f3FC90FDB; sub.ftz.f32 %f22076, %f22075, %f22073; selp.f32 %f22077, %f22074, %f22076, %p2780; setp.lt.ftz.f32 %p2781, %f22056, 0f00000000; mov.f32 %f22078, 0f40490FDB; sub.ftz.f32 %f22079, %f22078, %f22077; selp.f32 %f22080, %f22079, %f22077, %p2781; mul.ftz.f32 %f22081, %f22040, %f5063; sub.ftz.f32 %f22082, %f22044, %f22081; mul.ftz.f32 %f22083, %f22041, %f5063; sub.ftz.f32 %f22084, %f22045, %f22083; mul.ftz.f32 %f22085, %f22042, %f5063; sub.ftz.f32 %f22086, %f22046, %f22085; mul.ftz.f32 %f22087, %f22043, %f5063; sub.ftz.f32 %f22088, %f22047, %f22087; mul.ftz.f32 %f22089, %f22084, %f22084; fma.rn.ftz.f32 %f22090, %f22082, %f22082, %f22089; mul.ftz.f32 %f22091, %f22088, %f22088; fma.rn.ftz.f32 %f22092, %f22086, %f22086, %f22091; add.ftz.f32 %f22093, %f22090, %f22092; rsqrt.approx.ftz.f32 %f22094, %f22093; mul.ftz.f32 %f22095, %f22082, %f22094; mul.ftz.f32 %f22096, %f22084, %f22094; mul.ftz.f32 %f22097, %f22086, %f22094; mul.ftz.f32 %f22098, %f22088, %f22094; mul.ftz.f32 %f22099, %f5054, %f22080; cos.approx.ftz.f32 %f22100, %f22099; sin.approx.ftz.f32 %f22101, %f22099; mul.ftz.f32 %f22102, %f22095, %f22101; mul.ftz.f32 %f22103, %f22096, %f22101; mul.ftz.f32 %f22104, %f22097, %f22101; mul.ftz.f32 %f22105, %f22098, %f22101; fma.rn.ftz.f32 %f30264, %f22043, %f22100, %f22105; fma.rn.ftz.f32 %f30263, %f22042, %f22100, %f22104; fma.rn.ftz.f32 %f30262, %f22041, %f22100, %f22103; fma.rn.ftz.f32 %f30261, %f22040, %f22100, %f22102; $L__BB107_3121: ld.global.v4.f32 {%f22120, %f22121, %f22122, %f22123}, [%rd592+16]; ld.global.v4.f32 {%f22128, %f22129, %f22130, %f22131}, [%rd592+80]; mul.ftz.f32 %f22136, %f5054, %f22128; mul.ftz.f32 %f22137, %f5054, %f22129; mul.ftz.f32 %f22138, %f5054, %f22130; mul.ftz.f32 %f22139, %f5054, %f22131; fma.rn.ftz.f32 %f5077, %f5064, %f22120, %f22136; fma.rn.ftz.f32 %f5078, %f5064, %f22121, %f22137; fma.rn.ftz.f32 %f5079, %f5064, %f22122, %f22138; fma.rn.ftz.f32 %f22140, %f5064, %f22123, %f22139; ld.global.v4.f32 {%f22141, %f22142, %f22143, %f22144}, [%rd592+32]; ld.global.v4.f32 {%f22149, %f22150, %f22151, %f22152}, [%rd592+96]; mul.ftz.f32 %f22157, %f5054, %f22149; mul.ftz.f32 %f22158, %f5054, %f22150; mul.ftz.f32 %f22159, %f5054, %f22151; mul.ftz.f32 %f22160, %f5054, %f22152; fma.rn.ftz.f32 %f22161, %f5064, %f22141, %f22157; fma.rn.ftz.f32 %f22162, %f5064, %f22142, %f22158; fma.rn.ftz.f32 %f22163, %f5064, %f22143, %f22159; fma.rn.ftz.f32 %f22164, %f5064, %f22144, %f22160; ld.global.v4.f32 {%f22165, %f22166, %f22167, %f22168}, [%rd592+48]; ld.global.v4.f32 {%f22173, %f22174, %f22175, %f22176}, [%rd592+112]; mul.ftz.f32 %f22181, %f5054, %f22173; mul.ftz.f32 %f22182, %f5054, %f22174; mul.ftz.f32 %f22183, %f5054, %f22175; mul.ftz.f32 %f22184, %f5054, %f22176; fma.rn.ftz.f32 %f22185, %f5064, %f22165, %f22181; fma.rn.ftz.f32 %f22186, %f5064, %f22166, %f22182; fma.rn.ftz.f32 %f22187, %f5064, %f22167, %f22183; fma.rn.ftz.f32 %f22188, %f5064, %f22168, %f22184; mul.ftz.f32 %f22189, %f30264, 0f3FB504F3; mul.ftz.f32 %f22190, %f30261, 0f3FB504F3; mul.ftz.f32 %f22191, %f30262, 0f3FB504F3; mul.ftz.f32 %f22192, %f30263, 0f3FB504F3; mul.ftz.f32 %f22193, %f22189, %f22190; mul.ftz.f32 %f22194, %f22189, %f22191; mul.ftz.f32 %f22195, %f22189, %f22192; mul.ftz.f32 %f22196, %f22190, %f22190; mul.ftz.f32 %f22197, %f22190, %f22191; mul.ftz.f32 %f22198, %f22190, %f22192; mul.ftz.f32 %f22199, %f22191, %f22191; mul.ftz.f32 %f22200, %f22191, %f22192; mul.ftz.f32 %f22201, %f22192, %f22192; sub.ftz.f32 %f22203, %f22052, %f22199; sub.ftz.f32 %f22204, %f22203, %f22201; sub.ftz.f32 %f22205, %f22197, %f22195; add.ftz.f32 %f22206, %f22194, %f22198; add.ftz.f32 %f22207, %f22195, %f22197; sub.ftz.f32 %f22208, %f22052, %f22196; sub.ftz.f32 %f22209, %f22208, %f22201; sub.ftz.f32 %f22210, %f22200, %f22193; sub.ftz.f32 %f22211, %f22198, %f22194; add.ftz.f32 %f22212, %f22193, %f22200; sub.ftz.f32 %f22213, %f22208, %f22199; mul.ftz.f32 %f22214, %f22205, %f22163; fma.rn.ftz.f32 %f22215, %f22204, %f22140, %f22214; fma.rn.ftz.f32 %f30270, %f22206, %f22186, %f22215; mul.ftz.f32 %f22216, %f22205, %f22164; fma.rn.ftz.f32 %f22217, %f22204, %f22161, %f22216; fma.rn.ftz.f32 %f5081, %f22206, %f22187, %f22217; mul.ftz.f32 %f22218, %f22205, %f22185; fma.rn.ftz.f32 %f22219, %f22204, %f22162, %f22218; fma.rn.ftz.f32 %f5082, %f22206, %f22188, %f22219; mul.ftz.f32 %f22220, %f22209, %f22163; fma.rn.ftz.f32 %f22221, %f22207, %f22140, %f22220; fma.rn.ftz.f32 %f5083, %f22210, %f22186, %f22221; mul.ftz.f32 %f22222, %f22209, %f22164; fma.rn.ftz.f32 %f22223, %f22207, %f22161, %f22222; fma.rn.ftz.f32 %f30269, %f22210, %f22187, %f22223; mul.ftz.f32 %f22224, %f22209, %f22185; fma.rn.ftz.f32 %f22225, %f22207, %f22162, %f22224; fma.rn.ftz.f32 %f5085, %f22210, %f22188, %f22225; mul.ftz.f32 %f22226, %f22212, %f22163; fma.rn.ftz.f32 %f22227, %f22211, %f22140, %f22226; fma.rn.ftz.f32 %f5086, %f22213, %f22186, %f22227; mul.ftz.f32 %f22228, %f22212, %f22164; fma.rn.ftz.f32 %f22229, %f22211, %f22161, %f22228; fma.rn.ftz.f32 %f5087, %f22213, %f22187, %f22229; mul.ftz.f32 %f22230, %f22212, %f22185; fma.rn.ftz.f32 %f22231, %f22211, %f22162, %f22230; fma.rn.ftz.f32 %f30268, %f22213, %f22188, %f22231; cvta.to.local.u64 %rd3151, %rd299; st.local.v4.f32 [%rd3151], {%f30270, %f5081, %f5082, %f5077}; cvta.to.local.u64 %rd3152, %rd311; st.local.v4.f32 [%rd3152], {%f5083, %f30269, %f5085, %f5078}; cvta.to.local.u64 %rd3153, %rd310; st.local.v4.f32 [%rd3153], {%f5086, %f5087, %f30268, %f5079}; mul.ftz.f32 %f22232, %f30269, %f30268; mul.ftz.f32 %f5089, %f5087, %f5085; sub.ftz.f32 %f30267, %f22232, %f5089; mul.ftz.f32 %f5091, %f5087, %f5082; mul.ftz.f32 %f22233, %f5081, %f30268; sub.ftz.f32 %f30266, %f5091, %f22233; mul.ftz.f32 %f5093, %f5081, %f5085; mul.ftz.f32 %f22234, %f30269, %f5082; sub.ftz.f32 %f30265, %f5093, %f22234; mul.ftz.f32 %f22235, %f5083, %f30266; fma.rn.ftz.f32 %f22236, %f30270, %f30267, %f22235; fma.rn.ftz.f32 %f30271, %f5086, %f30265, %f22236; setp.neu.ftz.f32 %p2782, %f30271, 0f00000000; @%p2782 bra $L__BB107_3123; add.ftz.f32 %f30270, %f30270, 0f322BCC77; add.ftz.f32 %f30268, %f30268, 0f322BCC77; add.ftz.f32 %f30269, %f30269, 0f322BCC77; mul.ftz.f32 %f22237, %f30269, %f30268; sub.ftz.f32 %f30267, %f22237, %f5089; mul.ftz.f32 %f22238, %f5081, %f30268; sub.ftz.f32 %f30266, %f5091, %f22238; mul.ftz.f32 %f22239, %f30269, %f5082; sub.ftz.f32 %f30265, %f5093, %f22239; mul.ftz.f32 %f22240, %f5083, %f30266; fma.rn.ftz.f32 %f22241, %f30270, %f30267, %f22240; fma.rn.ftz.f32 %f22242, %f5086, %f30265, %f22241; setp.eq.ftz.f32 %p2783, %f22242, 0f00000000; selp.f32 %f30271, 0f7F7FFFFF, %f22242, %p2783; $L__BB107_3123: mul.ftz.f32 %f22243, %f5083, %f30268; mul.ftz.f32 %f22244, %f5086, %f5085; sub.ftz.f32 %f22245, %f22244, %f22243; mul.ftz.f32 %f22246, %f5086, %f5082; mul.ftz.f32 %f22247, %f30268, %f30270; sub.ftz.f32 %f22248, %f22247, %f22246; mul.ftz.f32 %f22249, %f5085, %f30270; mul.ftz.f32 %f22250, %f5083, %f5082; sub.ftz.f32 %f22251, %f22250, %f22249; mul.ftz.f32 %f22252, %f5086, %f30269; mul.ftz.f32 %f22253, %f5083, %f5087; sub.ftz.f32 %f22254, %f22253, %f22252; mul.ftz.f32 %f22255, %f5087, %f30270; mul.ftz.f32 %f22256, %f5086, %f5081; sub.ftz.f32 %f22257, %f22256, %f22255; mul.ftz.f32 %f22258, %f5083, %f5081; mul.ftz.f32 %f22259, %f30269, %f30270; sub.ftz.f32 %f22260, %f22259, %f22258; div.approx.ftz.f32 %f22261, %f30266, %f30271; mul.ftz.f32 %f22262, %f5078, %f22261; div.approx.ftz.f32 %f22263, %f30267, %f30271; fma.rn.ftz.f32 %f22264, %f5077, %f22263, %f22262; div.approx.ftz.f32 %f22265, %f30265, %f30271; fma.rn.ftz.f32 %f22266, %f5079, %f22265, %f22264; div.approx.ftz.f32 %f22267, %f22248, %f30271; mul.ftz.f32 %f22268, %f5078, %f22267; div.approx.ftz.f32 %f22269, %f22245, %f30271; fma.rn.ftz.f32 %f22270, %f5077, %f22269, %f22268; div.approx.ftz.f32 %f22271, %f22251, %f30271; fma.rn.ftz.f32 %f22272, %f5079, %f22271, %f22270; div.approx.ftz.f32 %f22273, %f22257, %f30271; mul.ftz.f32 %f22274, %f5078, %f22273; div.approx.ftz.f32 %f22275, %f22254, %f30271; fma.rn.ftz.f32 %f22276, %f5077, %f22275, %f22274; div.approx.ftz.f32 %f22277, %f22260, %f30271; fma.rn.ftz.f32 %f22278, %f5079, %f22277, %f22276; neg.ftz.f32 %f22279, %f22266; cvta.to.local.u64 %rd3154, %rd312; st.local.v4.f32 [%rd3154], {%f22263, %f22261, %f22265, %f22279}; neg.ftz.f32 %f22280, %f22272; cvta.to.local.u64 %rd3155, %rd327; st.local.v4.f32 [%rd3155], {%f22269, %f22267, %f22271, %f22280}; neg.ftz.f32 %f22281, %f22278; cvta.to.local.u64 %rd3156, %rd326; st.local.v4.f32 [%rd3156], {%f22275, %f22273, %f22277, %f22281}; $L__BB107_3124: cvt.u32.u64 %r4321, %rd588; setp.ne.s32 %p2784, %r4321, -1; @%p2784 bra $L__BB107_3126; bra.uni $L__BB107_3125; $L__BB107_3126: setp.lt.s32 %p2785, %r1115, 0; shl.b64 %rd3158, %rd588, 4; add.s64 %rd593, %rd328, %rd3158; @%p2785 bra $L__BB107_3128; and.b32 %r5799, %r1118, 4; mov.f32 %f30287, %f5047; mov.f32 %f30288, %f5046; mov.f32 %f30289, %f5045; bra.uni $L__BB107_3134; $L__BB107_3125: st.local.u32 [%rd318], %r4315; st.local.u32 [%rd319], %r4315; st.local.u32 [%rd320], %r4315; cvta.to.local.u64 %rd3157, %rd309; mov.f32 %f30281, 0f00000000; st.local.v2.f32 [%rd3157], {%f30281, %f30281}; st.local.u32 [%rd321], %r4315; mov.f32 %f30282, %f30281; mov.f32 %f30283, %f30281; mov.f32 %f30284, %f30281; mov.f32 %f30285, %f30281; mov.f32 %f30286, %f30281; mov.f32 %f30287, %f5047; mov.f32 %f30288, %f5046; mov.f32 %f30289, %f5045; bra.uni $L__BB107_3142; $L__BB107_3128: ld.global.v4.u32 {%r4323, %r4324, %r4325, %r4326}, [%rd593]; mul.wide.u32 %rd3159, %r4323, 12; add.s64 %rd3160, %rd352, %rd3159; mul.wide.u32 %rd3161, %r4324, 12; add.s64 %rd3162, %rd352, %rd3161; mul.wide.u32 %rd3163, %r4325, 12; add.s64 %rd3164, %rd352, %rd3163; mov.f32 %f22288, 0f3F800000; sub.ftz.f32 %f22289, %f22288, %f22035; sub.ftz.f32 %f22290, %f22289, %f22036; ld.global.f32 %f22291, [%rd3160]; ld.global.f32 %f22292, [%rd3160+4]; ld.global.f32 %f22293, [%rd3160+8]; ld.global.f32 %f22294, [%rd3162]; mul.ftz.f32 %f22295, %f22035, %f22294; ld.global.f32 %f22296, [%rd3162+4]; mul.ftz.f32 %f22297, %f22035, %f22296; ld.global.f32 %f22298, [%rd3162+8]; mul.ftz.f32 %f22299, %f22035, %f22298; fma.rn.ftz.f32 %f22300, %f22290, %f22291, %f22295; fma.rn.ftz.f32 %f22301, %f22290, %f22292, %f22297; fma.rn.ftz.f32 %f22302, %f22290, %f22293, %f22299; ld.global.f32 %f22303, [%rd3164]; ld.global.f32 %f22304, [%rd3164+4]; ld.global.f32 %f22305, [%rd3164+8]; fma.rn.ftz.f32 %f30272, %f22036, %f22303, %f22300; fma.rn.ftz.f32 %f30273, %f22036, %f22304, %f22301; fma.rn.ftz.f32 %f30274, %f22036, %f22305, %f22302; mul.ftz.f32 %f22306, %f30273, %f30273; fma.rn.ftz.f32 %f22307, %f30272, %f30272, %f22306; fma.rn.ftz.f32 %f22308, %f30274, %f30274, %f22307; sqrt.approx.ftz.f32 %f5113, %f22308; setp.neu.ftz.f32 %p2786, %f5113, 0f00000000; @%p2786 bra $L__BB107_3129; bra.uni $L__BB107_3130; $L__BB107_3129: rcp.approx.ftz.f32 %f22309, %f5113; mul.ftz.f32 %f30274, %f30274, %f22309; mul.ftz.f32 %f30273, %f30273, %f22309; mul.ftz.f32 %f30272, %f30272, %f22309; $L__BB107_3130: setp.eq.ftz.f32 %p2787, %f30272, 0f00000000; setp.eq.ftz.f32 %p2788, %f30273, 0f00000000; and.pred %p2789, %p2787, %p2788; setp.eq.ftz.f32 %p2790, %f30274, 0f00000000; and.pred %p2791, %p2790, %p2789; selp.f32 %f30287, %f5047, %f30274, %p2791; selp.f32 %f30288, %f5046, %f30273, %p2791; selp.f32 %f30289, %f5045, %f30272, %p2791; st.local.f32 [%rd315], %f30289; st.local.f32 [%rd337], %f30288; st.local.f32 [%rd338], %f30287; and.b32 %r5799, %r1118, 4; setp.ne.s32 %p2792, %r5799, 0; @%p2792 bra $L__BB107_3134; @%p2778 bra $L__BB107_3133; bra.uni $L__BB107_3132; $L__BB107_3133: ld.global.v4.f32 {%f22344, %f22345, %f22346, %f22347}, [%rd591+48]; ld.global.v4.f32 {%f22351, %f22352, %f22353, %f22354}, [%rd591+64]; ld.global.v4.f32 {%f22358, %f22359, %f22360, %f22361}, [%rd591+80]; mul.ftz.f32 %f22365, %f22351, %f30288; fma.rn.ftz.f32 %f22366, %f22344, %f30289, %f22365; fma.rn.ftz.f32 %f22367, %f22358, %f30287, %f22366; mul.ftz.f32 %f22368, %f22352, %f30288; fma.rn.ftz.f32 %f22369, %f22345, %f30289, %f22368; fma.rn.ftz.f32 %f22370, %f22359, %f30287, %f22369; mul.ftz.f32 %f22371, %f22353, %f30288; fma.rn.ftz.f32 %f22372, %f22346, %f30289, %f22371; fma.rn.ftz.f32 %f22373, %f22360, %f30287, %f22372; mul.ftz.f32 %f22374, %f22370, %f22370; fma.rn.ftz.f32 %f22375, %f22367, %f22367, %f22374; fma.rn.ftz.f32 %f22376, %f22373, %f22373, %f22375; sqrt.approx.ftz.f32 %f22377, %f22376; div.approx.ftz.f32 %f30289, %f22367, %f22377; div.approx.ftz.f32 %f30288, %f22370, %f22377; div.approx.ftz.f32 %f30287, %f22373, %f22377; st.local.f32 [%rd315], %f30289; st.local.f32 [%rd337], %f30288; st.local.f32 [%rd338], %f30287; mov.u32 %r5799, %r4315; bra.uni $L__BB107_3134; $L__BB107_3132: cvta.to.local.u64 %rd3165, %rd312; ld.local.v4.f32 {%f22310, %f22311, %f22312, %f22313}, [%rd3165]; cvta.to.local.u64 %rd3166, %rd327; ld.local.v4.f32 {%f22317, %f22318, %f22319, %f22320}, [%rd3166]; cvta.to.local.u64 %rd3167, %rd326; ld.local.v4.f32 {%f22324, %f22325, %f22326, %f22327}, [%rd3167]; mul.ftz.f32 %f22331, %f30288, %f22317; fma.rn.ftz.f32 %f22332, %f30289, %f22310, %f22331; fma.rn.ftz.f32 %f22333, %f30287, %f22324, %f22332; mul.ftz.f32 %f22334, %f30288, %f22318; fma.rn.ftz.f32 %f22335, %f30289, %f22311, %f22334; fma.rn.ftz.f32 %f22336, %f30287, %f22325, %f22335; mul.ftz.f32 %f22337, %f30288, %f22319; fma.rn.ftz.f32 %f22338, %f30289, %f22312, %f22337; fma.rn.ftz.f32 %f22339, %f30287, %f22326, %f22338; mul.ftz.f32 %f22340, %f22336, %f22336; fma.rn.ftz.f32 %f22341, %f22333, %f22333, %f22340; fma.rn.ftz.f32 %f22342, %f22339, %f22339, %f22341; sqrt.approx.ftz.f32 %f22343, %f22342; div.approx.ftz.f32 %f30289, %f22333, %f22343; div.approx.ftz.f32 %f30288, %f22336, %f22343; div.approx.ftz.f32 %f30287, %f22339, %f22343; st.local.f32 [%rd315], %f30289; st.local.f32 [%rd337], %f30288; st.local.f32 [%rd338], %f30287; mov.u32 %r5799, %r4315; $L__BB107_3134: ld.global.u32 %r4332, [%rd593+12]; mul.wide.u32 %rd3168, %r4332, 12; add.s64 %rd3169, %rd329, %rd3168; add.s32 %r4333, %r4332, 1; mul.wide.u32 %rd3170, %r4333, 12; add.s64 %rd3171, %rd329, %rd3170; add.s32 %r4334, %r4332, 2; mul.wide.u32 %rd3172, %r4334, 12; add.s64 %rd3173, %rd329, %rd3172; ld.global.f32 %f22378, [%rd3173]; ld.global.f32 %f22379, [%rd3173+4]; ld.global.f32 %f22380, [%rd3173+8]; ld.global.f32 %f22381, [%rd3171]; ld.global.f32 %f22382, [%rd3169]; sub.ftz.f32 %f30286, %f22381, %f22382; ld.global.f32 %f22383, [%rd3171+4]; ld.global.f32 %f22384, [%rd3169+4]; sub.ftz.f32 %f30285, %f22383, %f22384; ld.global.f32 %f22385, [%rd3171+8]; ld.global.f32 %f22386, [%rd3169+8]; sub.ftz.f32 %f30284, %f22385, %f22386; st.local.f32 [%rd318], %f30286; st.local.f32 [%rd319], %f30285; st.local.f32 [%rd320], %f30284; sub.ftz.f32 %f30283, %f22378, %f22382; sub.ftz.f32 %f30282, %f22379, %f22384; sub.ftz.f32 %f30281, %f22380, %f22386; cvta.to.local.u64 %rd594, %rd309; st.local.v2.f32 [%rd594], {%f30283, %f30282}; st.local.f32 [%rd321], %f30281; setp.ne.s32 %p2794, %r5799, 0; @%p2794 bra $L__BB107_3142; @%p2778 bra $L__BB107_3137; bra.uni $L__BB107_3136; $L__BB107_3137: ld.global.v4.f32 {%f22414, %f22415, %f22416, %f22417}, [%rd591]; ld.global.v4.f32 {%f22421, %f22422, %f22423, %f22424}, [%rd591+16]; ld.global.v4.f32 {%f22428, %f22429, %f22430, %f22431}, [%rd591+32]; mul.ftz.f32 %f22435, %f22415, %f30285; fma.rn.ftz.f32 %f22436, %f30286, %f22414, %f22435; fma.rn.ftz.f32 %f30280, %f22416, %f30284, %f22436; mul.ftz.f32 %f22437, %f22422, %f30285; fma.rn.ftz.f32 %f22438, %f30286, %f22421, %f22437; fma.rn.ftz.f32 %f30279, %f22423, %f30284, %f22438; mul.ftz.f32 %f22439, %f22429, %f30285; fma.rn.ftz.f32 %f22440, %f30286, %f22428, %f22439; fma.rn.ftz.f32 %f30284, %f22430, %f30284, %f22440; bra.uni $L__BB107_3138; $L__BB107_3136: cvta.to.local.u64 %rd3174, %rd299; ld.local.v4.f32 {%f22387, %f22388, %f22389, %f22390}, [%rd3174]; mul.ftz.f32 %f22394, %f30285, %f22388; fma.rn.ftz.f32 %f22395, %f30286, %f22387, %f22394; fma.rn.ftz.f32 %f30280, %f30284, %f22389, %f22395; cvta.to.local.u64 %rd3175, %rd311; ld.local.v4.f32 {%f22396, %f22397, %f22398, %f22399}, [%rd3175]; mul.ftz.f32 %f22403, %f30285, %f22397; fma.rn.ftz.f32 %f22404, %f30286, %f22396, %f22403; fma.rn.ftz.f32 %f30279, %f30284, %f22398, %f22404; cvta.to.local.u64 %rd3176, %rd310; ld.local.v4.f32 {%f22405, %f22406, %f22407, %f22408}, [%rd3176]; mul.ftz.f32 %f22412, %f30285, %f22406; fma.rn.ftz.f32 %f22413, %f30286, %f22405, %f22412; fma.rn.ftz.f32 %f30284, %f30284, %f22407, %f22413; $L__BB107_3138: st.local.f32 [%rd318], %f30280; st.local.f32 [%rd319], %f30279; st.local.f32 [%rd320], %f30284; @%p2778 bra $L__BB107_3140; bra.uni $L__BB107_3139; $L__BB107_3140: ld.global.v4.f32 {%f22468, %f22469, %f22470, %f22471}, [%rd591]; ld.global.v4.f32 {%f22475, %f22476, %f22477, %f22478}, [%rd591+16]; ld.global.v4.f32 {%f22482, %f22483, %f22484, %f22485}, [%rd591+32]; mul.ftz.f32 %f22489, %f22469, %f30282; fma.rn.ftz.f32 %f22490, %f30283, %f22468, %f22489; fma.rn.ftz.f32 %f5153, %f22470, %f30281, %f22490; mul.ftz.f32 %f22491, %f22476, %f30282; fma.rn.ftz.f32 %f22492, %f30283, %f22475, %f22491; fma.rn.ftz.f32 %f5154, %f22477, %f30281, %f22492; mul.ftz.f32 %f22493, %f22483, %f30282; fma.rn.ftz.f32 %f22494, %f30283, %f22482, %f22493; fma.rn.ftz.f32 %f30281, %f22484, %f30281, %f22494; st.local.v2.f32 [%rd594], {%f5153, %f5154}; st.local.f32 [%rd321], %f30281; mov.f32 %f30282, %f5154; mov.f32 %f30283, %f5153; mov.f32 %f30285, %f30279; mov.f32 %f30286, %f30280; bra.uni $L__BB107_3142; $L__BB107_3139: cvta.to.local.u64 %rd3177, %rd299; ld.local.v4.f32 {%f22441, %f22442, %f22443, %f22444}, [%rd3177]; mul.ftz.f32 %f22448, %f30282, %f22442; fma.rn.ftz.f32 %f22449, %f30283, %f22441, %f22448; fma.rn.ftz.f32 %f5150, %f30281, %f22443, %f22449; cvta.to.local.u64 %rd3178, %rd311; ld.local.v4.f32 {%f22450, %f22451, %f22452, %f22453}, [%rd3178]; mul.ftz.f32 %f22457, %f30282, %f22451; fma.rn.ftz.f32 %f22458, %f30283, %f22450, %f22457; fma.rn.ftz.f32 %f5151, %f30281, %f22452, %f22458; cvta.to.local.u64 %rd3179, %rd310; ld.local.v4.f32 {%f22459, %f22460, %f22461, %f22462}, [%rd3179]; mul.ftz.f32 %f22466, %f30282, %f22460; fma.rn.ftz.f32 %f22467, %f30283, %f22459, %f22466; fma.rn.ftz.f32 %f30281, %f30281, %f22461, %f22467; st.local.v2.f32 [%rd594], {%f5150, %f5151}; st.local.f32 [%rd321], %f30281; mov.f32 %f30282, %f5151; mov.f32 %f30283, %f5150; mov.f32 %f30285, %f30279; mov.f32 %f30286, %f30280; $L__BB107_3142: cvt.u32.u64 %r4336, %rd588; setp.eq.s32 %p2797, %r4336, -1; @%p2797 bra $L__BB107_3145; mul.ftz.f32 %f22501, %f5046, %f5052; fma.rn.ftz.f32 %f22502, %f5045, %f5051, %f22501; fma.rn.ftz.f32 %f22503, %f5047, %f5053, %f22502; setp.geu.ftz.f32 %p2798, %f22503, 0f00000000; @%p2798 bra $L__BB107_3145; or.b32 %r4337, %r1117, 1; st.local.u32 [%rd293], %r4337; neg.ftz.f32 %f22504, %f5047; st.local.f32 [%rd307], %f22504; neg.ftz.f32 %f22505, %f30289; st.local.f32 [%rd315], %f22505; neg.ftz.f32 %f22506, %f5046; neg.ftz.f32 %f22507, %f5045; neg.ftz.f32 %f22508, %f30287; neg.ftz.f32 %f22509, %f30288; st.local.v4.f32 [%rd337], {%f22509, %f22508, %f22507, %f22506}; neg.ftz.f32 %f22510, %f30286; st.local.f32 [%rd318], %f22510; neg.ftz.f32 %f22511, %f30285; st.local.f32 [%rd319], %f22511; neg.ftz.f32 %f22512, %f30284; st.local.f32 [%rd320], %f22512; neg.ftz.f32 %f22513, %f30282; neg.ftz.f32 %f22514, %f30283; cvta.to.local.u64 %rd3181, %rd309; st.local.v2.f32 [%rd3181], {%f22514, %f22513}; neg.ftz.f32 %f22515, %f30281; st.local.f32 [%rd321], %f22515; $L__BB107_3145: cvt.u32.u64 %r4338, %rd585; mov.u32 %r4339, 0; st.local.u32 [%rd322], %r4339; mov.f32 %f22516, 0f00000000; st.local.v4.f32 [%rd324], {%f22516, %f22516, %f22516, %f22516}; st.local.u32 [%rd336], %r4339; cvt.u16.u64 %rs549, %rd585; add.s16 %rs550, %rs252, %rs549; st.global.u16 [%rd404], %rs550; add.s32 %r4340, %r1106, %r4338; st.global.u16 [%rd278], %r4340; st.local.u32 [%rd342], %r4339; mov.u32 %r4341, 4; st.local.u32 [%rd343], %r4341; bra.uni $L__BB107_3146; $L__BB107_3312: { // callseq 536, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 536 add.s32 %r5800, %r1123, 1; $L__BB107_3146: mov.u32 %r1123, %r5800; mul.wide.s32 %rd3184, %r1123, 16; add.s64 %rd596, %rd344, %rd3184; ld.global.v4.u32 {%r4343, %r4344, %r4345, %r4346}, [%rd596]; mul.wide.u32 %rd3185, %r4344, 4; add.s64 %rd597, %rd2157, %rd3185; mul.wide.u32 %rd3186, %r4345, 4; add.s64 %rd598, %rd2157, %rd3186; setp.gt.s32 %p2799, %r4343, 1; @%p2799 bra $L__BB107_3150; setp.eq.s32 %p2927, %r4343, 1; mov.u32 %r5800, %r4344; @%p2927 bra $L__BB107_3146; bra.uni $L__BB107_3148; $L__BB107_3150: setp.gt.s32 %p2800, %r4343, 49; @%p2800 bra $L__BB107_3231; setp.gt.s32 %p2864, %r4343, 25; @%p2864 bra $L__BB107_3192; setp.gt.s32 %p2896, %r4343, 13; @%p2896 bra $L__BB107_3173; setp.gt.s32 %p2912, %r4343, 7; @%p2912 bra $L__BB107_3163; setp.gt.s32 %p2920, %r4343, 4; @%p2920 bra $L__BB107_3159; setp.eq.s32 %p2924, %r4343, 2; @%p2924 bra $L__BB107_3379; setp.eq.s32 %p2925, %r4343, 3; @%p2925 bra $L__BB107_3378; setp.eq.s32 %p2926, %r4343, 4; @%p2926 bra $L__BB107_3158; bra.uni $L__BB107_3148; $L__BB107_3158: { // callseq 615, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4341, %r4344, %r4345, %r4346}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 615 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3231: setp.gt.s32 %p2801, %r4343, 73; @%p2801 bra $L__BB107_3271; setp.gt.s32 %p2833, %r4343, 61; @%p2833 bra $L__BB107_3252; setp.gt.s32 %p2849, %r4343, 55; @%p2849 bra $L__BB107_3243; setp.gt.s32 %p2857, %r4343, 52; @%p2857 bra $L__BB107_3239; setp.eq.s32 %p2861, %r4343, 50; @%p2861 bra $L__BB107_3346; setp.eq.s32 %p2862, %r4343, 51; @%p2862 bra $L__BB107_3345; setp.eq.s32 %p2863, %r4343, 52; @%p2863 bra $L__BB107_3238; bra.uni $L__BB107_3148; $L__BB107_3238: add.s32 %r4383, %r1123, 1; { // callseq 576, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4383; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 576 bra.uni $L__BB107_3146; $L__BB107_3192: setp.gt.s32 %p2865, %r4343, 37; @%p2865 bra $L__BB107_3212; setp.gt.s32 %p2881, %r4343, 31; @%p2881 bra $L__BB107_3203; setp.gt.s32 %p2889, %r4343, 28; @%p2889 bra $L__BB107_3199; setp.eq.s32 %p2893, %r4343, 26; @%p2893 bra $L__BB107_3362; setp.eq.s32 %p2894, %r4343, 27; @%p2894 bra $L__BB107_3361; setp.eq.s32 %p2895, %r4343, 28; @%p2895 bra $L__BB107_3198; bra.uni $L__BB107_3148; $L__BB107_3198: mov.u32 %r4398, 28; { // callseq 598, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4398; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 598 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3271: setp.gt.s32 %p2802, %r4343, 85; @%p2802 bra $L__BB107_3294; setp.gt.s32 %p2818, %r4343, 79; @%p2818 bra $L__BB107_3285; setp.gt.s32 %p2826, %r4343, 76; @%p2826 bra $L__BB107_3281; setp.eq.s32 %p2830, %r4343, 74; @%p2830 bra $L__BB107_3330; setp.eq.s32 %p2831, %r4343, 75; @%p2831 bra $L__BB107_3329; setp.eq.s32 %p2832, %r4343, 76; @%p2832 bra $L__BB107_3277; bra.uni $L__BB107_3148; $L__BB107_3277: add.s32 %r5800, %r1123, 1; setp.eq.s32 %p2929, %r4346, 255; @%p2929 bra $L__BB107_3146; add.s64 %rd599, %rd597, 4; add.s64 %rd600, %rd597, 8; setp.eq.s32 %p2930, %r4345, 0; mul.wide.u32 %rd3213, %r4346, 4; add.s64 %rd601, %rd2157, %rd3213; @%p2930 bra $L__BB107_3328; setp.eq.s32 %p2931, %r4345, 1; @%p2931 bra $L__BB107_3327; bra.uni $L__BB107_3280; $L__BB107_3327: ld.local.f32 %f22524, [%rd599]; st.local.f32 [%rd601], %f22524; bra.uni $L__BB107_3146; $L__BB107_3173: setp.gt.s32 %p2897, %r4343, 19; @%p2897 bra $L__BB107_3183; setp.gt.s32 %p2905, %r4343, 16; @%p2905 bra $L__BB107_3179; setp.eq.s32 %p2909, %r4343, 14; @%p2909 bra $L__BB107_3370; setp.eq.s32 %p2910, %r4343, 15; @%p2910 bra $L__BB107_3369; setp.eq.s32 %p2911, %r4343, 16; @%p2911 bra $L__BB107_3178; bra.uni $L__BB107_3148; $L__BB107_3178: mov.u32 %r4408, 16; { // callseq 609, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4408; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 609 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3252: setp.gt.s32 %p2834, %r4343, 67; @%p2834 bra $L__BB107_3262; setp.gt.s32 %p2842, %r4343, 64; @%p2842 bra $L__BB107_3258; setp.eq.s32 %p2846, %r4343, 62; @%p2846 bra $L__BB107_3338; setp.eq.s32 %p2847, %r4343, 63; @%p2847 bra $L__BB107_3337; setp.eq.s32 %p2848, %r4343, 64; @%p2848 bra $L__BB107_3257; bra.uni $L__BB107_3148; $L__BB107_3257: mov.u32 %r4369, 64; add.s32 %r4370, %r1123, 1; { // callseq 564, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4369; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4370; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 564 bra.uni $L__BB107_3146; $L__BB107_3212: setp.gt.s32 %p2866, %r4343, 43; @%p2866 bra $L__BB107_3222; setp.gt.s32 %p2874, %r4343, 40; @%p2874 bra $L__BB107_3218; setp.eq.s32 %p2878, %r4343, 38; @%p2878 bra $L__BB107_3354; setp.eq.s32 %p2879, %r4343, 39; @%p2879 bra $L__BB107_3353; setp.eq.s32 %p2880, %r4343, 40; @%p2880 bra $L__BB107_3217; bra.uni $L__BB107_3148; $L__BB107_3217: add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3294: setp.gt.s32 %p2803, %r4343, 92; @%p2803 bra $L__BB107_3304; setp.gt.s32 %p2811, %r4343, 88; @%p2811 bra $L__BB107_3300; setp.eq.s32 %p2815, %r4343, 86; @%p2815 bra $L__BB107_3319; setp.eq.s32 %p2816, %r4343, 87; @%p2816 bra $L__BB107_3318; setp.eq.s32 %p2817, %r4343, 88; @%p2817 bra $L__BB107_3299; bra.uni $L__BB107_3148; $L__BB107_3299: mov.u32 %r4352, 88; ld.local.f32 %f22517, [%rd282]; ld.local.v2.f32 {%f22518, %f22519}, [%rd282+4]; { // callseq 542, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f22517; .param .b32 param1; st.param.f32 [param1+0], %f22518; .param .b32 param2; st.param.f32 [param2+0], %f22519; .param .b64 param3; st.param.b64 [param3+0], %rd2156; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r4352; st.param.b32 [param4+4], %r4344; st.param.b32 [param4+8], %r4345; st.param.b32 [param4+12], %r4346; call.uni _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 542 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3163: setp.gt.s32 %p2913, %r4343, 10; @%p2913 bra $L__BB107_3169; setp.eq.s32 %p2917, %r4343, 8; @%p2917 bra $L__BB107_3375; setp.eq.s32 %p2918, %r4343, 9; @%p2918 bra $L__BB107_3373; setp.eq.s32 %p2919, %r4343, 10; @%p2919 bra $L__BB107_3167; bra.uni $L__BB107_3148; $L__BB107_3167: ld.local.f32 %f22549, [%rd598]; setp.ltu.ftz.f32 %p2932, %f22549, 0f3F800000; add.s32 %r5800, %r1123, 1; @%p2932 bra $L__BB107_3146; add.s32 %r4419, %r4344, %r1123; add.s32 %r5800, %r4419, 1; bra.uni $L__BB107_3146; $L__BB107_3243: setp.gt.s32 %p2850, %r4343, 58; @%p2850 bra $L__BB107_3248; setp.eq.s32 %p2854, %r4343, 56; @%p2854 bra $L__BB107_3342; setp.eq.s32 %p2855, %r4343, 57; @%p2855 bra $L__BB107_3341; setp.eq.s32 %p2856, %r4343, 58; @%p2856 bra $L__BB107_3247; bra.uni $L__BB107_3148; $L__BB107_3247: mov.u32 %r4378, 58; { // callseq 570, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4378; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 570 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3203: setp.gt.s32 %p2882, %r4343, 34; @%p2882 bra $L__BB107_3208; setp.eq.s32 %p2886, %r4343, 32; @%p2886 bra $L__BB107_3358; setp.eq.s32 %p2887, %r4343, 33; @%p2887 bra $L__BB107_3357; setp.eq.s32 %p2888, %r4343, 34; @%p2888 bra $L__BB107_3207; bra.uni $L__BB107_3148; $L__BB107_3207: { // callseq 593, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 593 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3285: setp.gt.s32 %p2819, %r4343, 82; @%p2819 bra $L__BB107_3290; setp.eq.s32 %p2823, %r4343, 80; @%p2823 bra $L__BB107_3323; setp.eq.s32 %p2824, %r4343, 81; @%p2824 bra $L__BB107_3322; setp.eq.s32 %p2825, %r4343, 82; @%p2825 bra $L__BB107_3289; bra.uni $L__BB107_3148; $L__BB107_3289: mov.u32 %r4356, 82; { // callseq 548, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4356; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 548 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3183: setp.gt.s32 %p2898, %r4343, 22; @%p2898 bra $L__BB107_3188; setp.eq.s32 %p2902, %r4343, 20; @%p2902 bra $L__BB107_3366; setp.eq.s32 %p2903, %r4343, 21; @%p2903 bra $L__BB107_3365; setp.eq.s32 %p2904, %r4343, 22; @%p2904 bra $L__BB107_3187; bra.uni $L__BB107_3148; $L__BB107_3187: mov.u32 %r4405, 22; add.s32 %r4406, %r1123, 1; { // callseq 604, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4405; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; .param .b32 param3; st.param.b32 [param3+0], %r4406; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 604 bra.uni $L__BB107_3146; $L__BB107_3262: setp.gt.s32 %p2835, %r4343, 70; @%p2835 bra $L__BB107_3267; setp.eq.s32 %p2839, %r4343, 68; @%p2839 bra $L__BB107_3334; setp.eq.s32 %p2840, %r4343, 69; @%p2840 bra $L__BB107_3333; setp.eq.s32 %p2841, %r4343, 70; @%p2841 bra $L__BB107_3266; bra.uni $L__BB107_3148; $L__BB107_3266: mov.u32 %r4363, 70; { // callseq 558, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4363; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 558 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3222: setp.gt.s32 %p2867, %r4343, 46; @%p2867 bra $L__BB107_3227; setp.eq.s32 %p2871, %r4343, 44; @%p2871 bra $L__BB107_3350; setp.eq.s32 %p2872, %r4343, 45; @%p2872 bra $L__BB107_3349; setp.eq.s32 %p2873, %r4343, 46; @%p2873 bra $L__BB107_3226; bra.uni $L__BB107_3148; $L__BB107_3226: { // callseq 582, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 582 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3304: setp.gt.s32 %p2804, %r4343, 95; @%p2804 bra $L__BB107_3309; setp.eq.s32 %p2808, %r4343, 93; @%p2808 bra $L__BB107_3314; setp.eq.s32 %p2809, %r4343, 94; @%p2809 bra $L__BB107_3313; setp.eq.s32 %p2810, %r4343, 95; @%p2810 bra $L__BB107_3308; bra.uni $L__BB107_3148; $L__BB107_3308: { // callseq 537, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 537 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3159: setp.eq.s32 %p2921, %r4343, 5; @%p2921 bra $L__BB107_3377; setp.eq.s32 %p2922, %r4343, 6; @%p2922 bra $L__BB107_3376; setp.eq.s32 %p2923, %r4343, 7; @%p2923 bra $L__BB107_3162; bra.uni $L__BB107_3148; $L__BB107_3162: mov.u32 %r4422, 7; { // callseq 614, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4422; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 614 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3239: setp.eq.s32 %p2858, %r4343, 53; @%p2858 bra $L__BB107_3344; setp.eq.s32 %p2859, %r4343, 54; @%p2859 bra $L__BB107_3343; setp.eq.s32 %p2860, %r4343, 55; @%p2860 bra $L__BB107_3242; bra.uni $L__BB107_3148; $L__BB107_3242: ld.local.v4.f32 {%f22526, %f22527, %f22528, %f22529}, [%rd2173]; { // callseq 573, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f22526; st.param.f32 [param0+4], %f22527; st.param.f32 [param0+8], %f22528; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; .param .b32 param4; st.param.b32 [param4+0], %r4346; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 573 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3199: setp.eq.s32 %p2890, %r4343, 29; @%p2890 bra $L__BB107_3360; setp.eq.s32 %p2891, %r4343, 30; @%p2891 bra $L__BB107_3359; setp.eq.s32 %p2892, %r4343, 31; @%p2892 bra $L__BB107_3202; bra.uni $L__BB107_3148; $L__BB107_3202: mov.u32 %r4395, 31; add.s32 %r4396, %r1123, 1; mov.u32 %r4397, 8; { // callseq 595, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r4397; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r4395; st.param.b32 [param3+4], %r4344; st.param.b32 [param3+8], %r4345; st.param.b32 [param3+12], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4396; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 595 bra.uni $L__BB107_3146; $L__BB107_3281: setp.eq.s32 %p2827, %r4343, 77; @%p2827 bra $L__BB107_3325; setp.eq.s32 %p2828, %r4343, 78; @%p2828 bra $L__BB107_3324; setp.eq.s32 %p2829, %r4343, 79; @%p2829 bra $L__BB107_3284; bra.uni $L__BB107_3148; $L__BB107_3284: add.s32 %r4358, %r1123, 1; { // callseq 551, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4358; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 551 bra.uni $L__BB107_3146; $L__BB107_3179: setp.eq.s32 %p2906, %r4343, 17; @%p2906 bra $L__BB107_3368; setp.eq.s32 %p2907, %r4343, 18; @%p2907 bra $L__BB107_3367; setp.eq.s32 %p2908, %r4343, 19; @%p2908 bra $L__BB107_3182; bra.uni $L__BB107_3148; $L__BB107_3182: { // callseq 606, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 606 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3258: setp.eq.s32 %p2843, %r4343, 65; @%p2843 bra $L__BB107_3336; setp.eq.s32 %p2844, %r4343, 66; @%p2844 bra $L__BB107_3335; setp.eq.s32 %p2845, %r4343, 67; @%p2845 bra $L__BB107_3261; bra.uni $L__BB107_3148; $L__BB107_3261: mov.u32 %r4367, 67; { // callseq 561, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4367, %r4344, %r4345, %r4346}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 561 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3218: setp.eq.s32 %p2875, %r4343, 41; @%p2875 bra $L__BB107_3352; setp.eq.s32 %p2876, %r4343, 42; @%p2876 bra $L__BB107_3351; setp.eq.s32 %p2877, %r4343, 43; @%p2877 bra $L__BB107_3221; bra.uni $L__BB107_3148; $L__BB107_3221: add.s32 %r4387, %r1123, 1; { // callseq 585, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4387; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 585 bra.uni $L__BB107_3146; $L__BB107_3300: setp.eq.s32 %p2812, %r4343, 89; @%p2812 bra $L__BB107_3317; setp.eq.s32 %p2813, %r4343, 90; @%p2813 bra $L__BB107_3316; setp.eq.s32 %p2814, %r4343, 92; @%p2814 bra $L__BB107_3303; bra.uni $L__BB107_3148; $L__BB107_3303: add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3169: setp.eq.s32 %p2914, %r4343, 11; @%p2914 bra $L__BB107_3372; setp.eq.s32 %p2915, %r4343, 12; @%p2915 bra $L__BB107_3371; setp.eq.s32 %p2916, %r4343, 13; @%p2916 bra $L__BB107_3172; bra.uni $L__BB107_3148; $L__BB107_3172: mov.u32 %r4416, 13; add.s32 %r4417, %r1123, 1; mov.u32 %r4418, 8; { // callseq 610, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r4418; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r4416; st.param.b32 [param3+4], %r4344; st.param.b32 [param3+8], %r4345; st.param.b32 [param3+12], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4417; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 610 bra.uni $L__BB107_3146; $L__BB107_3248: setp.eq.s32 %p2851, %r4343, 59; @%p2851 bra $L__BB107_3340; setp.eq.s32 %p2852, %r4343, 60; @%p2852 bra $L__BB107_3339; setp.eq.s32 %p2853, %r4343, 61; @%p2853 bra $L__BB107_3251; bra.uni $L__BB107_3148; $L__BB107_3251: mov.u32 %r4374, 61; add.s32 %r4375, %r1123, 1; { // callseq 567, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4374; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4375; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 567 bra.uni $L__BB107_3146; $L__BB107_3208: setp.eq.s32 %p2883, %r4343, 35; @%p2883 bra $L__BB107_3356; setp.eq.s32 %p2884, %r4343, 36; @%p2884 bra $L__BB107_3355; setp.eq.s32 %p2885, %r4343, 37; @%p2885 bra $L__BB107_3211; bra.uni $L__BB107_3148; $L__BB107_3211: mov.u32 %r4390, 37; { // callseq 590, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4390, %r4344, %r4345, %r4346}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 590 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3290: setp.eq.s32 %p2820, %r4343, 83; @%p2820 bra $L__BB107_3321; setp.eq.s32 %p2821, %r4343, 84; @%p2821 bra $L__BB107_3320; setp.eq.s32 %p2822, %r4343, 85; @%p2822 bra $L__BB107_3293; bra.uni $L__BB107_3148; $L__BB107_3293: add.s32 %r4355, %r1123, 1; { // callseq 545, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4355; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 545 bra.uni $L__BB107_3146; $L__BB107_3188: setp.eq.s32 %p2899, %r4343, 23; @%p2899 bra $L__BB107_3364; setp.eq.s32 %p2900, %r4343, 24; @%p2900 bra $L__BB107_3363; setp.eq.s32 %p2901, %r4343, 25; @%p2901 bra $L__BB107_3191; bra.uni $L__BB107_3148; $L__BB107_3191: add.s32 %r4401, %r1123, 1; { // callseq 601, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4401; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 601 bra.uni $L__BB107_3146; $L__BB107_3267: setp.eq.s32 %p2836, %r4343, 71; @%p2836 bra $L__BB107_3332; setp.eq.s32 %p2837, %r4343, 72; @%p2837 bra $L__BB107_3331; setp.eq.s32 %p2838, %r4343, 73; @%p2838 bra $L__BB107_3270; bra.uni $L__BB107_3148; $L__BB107_3270: add.s32 %r4361, %r1123, 1; { // callseq 555, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4361; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 555 bra.uni $L__BB107_3146; $L__BB107_3227: setp.eq.s32 %p2868, %r4343, 47; @%p2868 bra $L__BB107_3348; setp.eq.s32 %p2869, %r4343, 48; @%p2869 bra $L__BB107_3347; setp.eq.s32 %p2870, %r4343, 49; @%p2870 bra $L__BB107_3230; bra.uni $L__BB107_3148; $L__BB107_3230: { // callseq 579, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 579 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3309: setp.eq.s32 %p2805, %r4343, 98; @%p2805 bra $L__BB107_3380; setp.eq.s32 %p2806, %r4343, 97; @%p2806 bra $L__BB107_3315; setp.ne.s32 %p2807, %r4343, 96; @%p2807 bra $L__BB107_3148; bra.uni $L__BB107_3312; $L__BB107_3379: mov.u32 %r4425, 2; add.s32 %r4426, %r1123, 1; mov.u32 %r4427, 8; { // callseq 617, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4425; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; .param .b32 param3; st.param.b32 [param3+0], %r4427; .param .b32 param4; st.param.b32 [param4+0], %r4426; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 617 bra.uni $L__BB107_3146; $L__BB107_3378: mov.u32 %r4424, 3; { // callseq 616, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4424, %r4344, %r4345, %r4346}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 616 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3346: { // callseq 578, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 578 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3345: { // callseq 577, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 577 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3362: mov.u32 %r4400, 26; { // callseq 600, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4400; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 600 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3361: mov.u32 %r4399, 27; { // callseq 599, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4399; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 599 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3330: { // callseq 554, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 554 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3329: { // callseq 553, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 553 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3370: st.local.u32 [%rd598], %r4344; add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3369: ld.global.v4.u32 {%r4409, %r4410, %r4411, %r4412}, [%rd596+16]; add.s32 %r5800, %r1123, 2; st.local.u32 [%rd597], %r4410; st.local.u32 [%rd597+4], %r4411; st.local.u32 [%rd597+8], %r4412; bra.uni $L__BB107_3146; $L__BB107_3338: mov.u32 %r4372, 62; add.s32 %r4373, %r1123, 1; { // callseq 566, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4372; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4373; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 566 bra.uni $L__BB107_3146; $L__BB107_3337: mov.u32 %r4371, 63; { // callseq 565, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4371; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 565 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3354: { // callseq 589, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; .param .b32 param4; st.param.b32 [param4+0], %r4346; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 589 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3353: mov.u32 %r4389, 39; { // callseq 588, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4389, %r4344, %r4345, %r4346}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 588 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3319: add.s32 %r4354, %r1123, 1; { // callseq 544, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4354; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 544 bra.uni $L__BB107_3146; $L__BB107_3318: add.s32 %r4353, %r1123, 1; { // callseq 543, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4353; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 543 bra.uni $L__BB107_3146; $L__BB107_3375: mov.u32 %r4421, 8; { // callseq 613, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4421; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 613 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3373: ld.local.f32 %f22550, [%rd598]; setp.gtu.ftz.f32 %p2933, %f22550, 0f00000000; add.s32 %r5800, %r1123, 1; @%p2933 bra $L__BB107_3146; add.s32 %r4420, %r4344, %r1123; add.s32 %r5800, %r4420, 1; bra.uni $L__BB107_3146; $L__BB107_3342: mov.u32 %r4381, 56; { // callseq 572, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4381; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 572 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3341: mov.u32 %r4379, 57; add.s32 %r4380, %r1123, 1; { // callseq 571, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4379; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4380; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 571 bra.uni $L__BB107_3146; $L__BB107_3358: mov.u32 %r4392, 32; add.s32 %r4393, %r1123, 1; mov.u32 %r4394, 8; { // callseq 594, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r4394; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r4392; st.param.b32 [param3+4], %r4344; st.param.b32 [param3+8], %r4345; st.param.b32 [param3+12], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4393; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 594 bra.uni $L__BB107_3146; $L__BB107_3357: ld.local.f32 %f22533, [%rd597]; ld.local.f32 %f22534, [%rd597+4]; ld.local.f32 %f22535, [%rd597+8]; st.local.f32 [%rd315], %f22533; st.local.f32 [%rd337], %f22534; st.local.f32 [%rd338], %f22535; st.local.f32 [%rd598], %f22533; st.local.f32 [%rd598+4], %f22534; st.local.f32 [%rd598+8], %f22535; add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3323: { // callseq 550, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 550 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3322: mov.u32 %r4357, 81; { // callseq 549, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4357; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 549 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3366: ld.local.f32 %f22536, [%rd597]; ld.local.v4.f32 {%f22537, %f22538, %f22539, %f22540}, [%rd2173]; ld.local.f32 %f22544, [%rd597+4]; ld.local.f32 %f22545, [%rd597+8]; add.ftz.f32 %f22546, %f22545, %f22539; add.ftz.f32 %f22547, %f22544, %f22538; add.ftz.f32 %f22548, %f22536, %f22537; st.local.v2.f32 [%rd2173], {%f22548, %f22547}; st.local.f32 [%rd314], %f22546; add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3365: mov.u32 %r4407, 21; { // callseq 605, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4407; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 605 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3334: mov.u32 %r4366, 68; { // callseq 560, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4366; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 560 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3333: mov.u32 %r4364, 69; add.s32 %r4365, %r1123, 1; { // callseq 559, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4364; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4365; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 559 bra.uni $L__BB107_3146; $L__BB107_3350: mov.u32 %r4385, 44; add.s32 %r4386, %r1123, 1; { // callseq 584, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4385; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4386; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 584 bra.uni $L__BB107_3146; $L__BB107_3349: { // callseq 583, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 583 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3314: add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3313: mov.u32 %r4347, 94; add.s32 %r4348, %r1123, 1; { // callseq 538, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4347; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4348; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 538 bra.uni $L__BB107_3146; $L__BB107_3377: st.local.u32 [%rd348], %r4344; st.local.u32 [%rd349], %r4345; st.local.u32 [%rd350], %r4346; add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3376: ld.local.f32 %f22551, [%rd597]; ld.local.f32 %f22552, [%rd597+4]; ld.local.f32 %f22553, [%rd597+8]; st.local.f32 [%rd348], %f22551; st.local.f32 [%rd349], %f22552; st.local.f32 [%rd350], %f22553; add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3344: { // callseq 575, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 575 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3343: add.s32 %r4382, %r1123, 1; { // callseq 574, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4382; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 574 bra.uni $L__BB107_3146; $L__BB107_3360: { // callseq 597, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; .param .b32 param4; st.param.b32 [param4+0], %r4346; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 597 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3359: { // callseq 596, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; .param .b32 param4; st.param.b32 [param4+0], %r4346; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 596 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3325: setp.eq.s32 %p2928, %r4346, 255; add.s32 %r5800, %r1123, 1; @%p2928 bra $L__BB107_3146; ld.local.f32 %f22522, [%rd597]; add.s32 %r4360, %r4346, %r4345; mul.wide.u32 %rd3209, %r4360, 4; add.s64 %rd3210, %rd2157, %rd3209; st.local.f32 [%rd3210], %f22522; bra.uni $L__BB107_3146; $L__BB107_3324: add.s32 %r4359, %r1123, 1; { // callseq 552, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4359; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 552 bra.uni $L__BB107_3146; $L__BB107_3368: { // callseq 608, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; .param .b32 param4; st.param.b32 [param4+0], %r4346; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 608 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3367: { // callseq 607, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 607 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3336: { // callseq 563, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 563 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3335: add.s32 %r4368, %r1123, 1; { // callseq 562, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4368; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 562 bra.uni $L__BB107_3146; $L__BB107_3352: add.s32 %r4388, %r1123, 1; { // callseq 587, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r4388; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 587 bra.uni $L__BB107_3146; $L__BB107_3351: { // callseq 586, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 586 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3317: mov.u32 %r4351, 89; { // callseq 541, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4351; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 541 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3316: mov.u32 %r4349, 90; add.s32 %r4350, %r1123, 1; { // callseq 540, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4349; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; .param .b32 param3; st.param.b32 [param3+0], %r4350; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 540 bra.uni $L__BB107_3146; $L__BB107_3372: { // callseq 612, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 612 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3371: { // callseq 611, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 611 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3340: add.s32 %r4377, %r1123, 1; { // callseq 569, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4377; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 569 bra.uni $L__BB107_3146; $L__BB107_3339: add.s32 %r4376, %r1123, 1; { // callseq 568, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; .param .b32 param4; st.param.b32 [param4+0], %r4376; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 568 bra.uni $L__BB107_3146; $L__BB107_3356: { // callseq 592, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 592 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3355: mov.u32 %r4391, 36; { // callseq 591, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4391; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 591 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3321: { // callseq 547, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 547 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3320: { // callseq 546, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 546 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3364: mov.u32 %r4403, 23; add.s32 %r4404, %r1123, 1; { // callseq 603, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4403; st.param.b32 [param1+4], %r4344; st.param.b32 [param1+8], %r4345; st.param.b32 [param1+12], %r4346; .param .b32 param2; st.param.b32 [param2+0], %r4404; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5800, [retval0+0]; } // callseq 603 bra.uni $L__BB107_3146; $L__BB107_3363: mov.u32 %r4402, 24; { // callseq 602, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4402; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 602 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3332: mov.u32 %r4362, 71; { // callseq 557, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4362; st.param.b32 [param2+4], %r4344; st.param.b32 [param2+8], %r4345; st.param.b32 [param2+12], %r4346; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 557 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3331: { // callseq 556, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; .param .b32 param3; st.param.b32 [param3+0], %r4346; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 556 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3348: mov.u32 %r4384, 8; { // callseq 581, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1602; .param .b64 param1; st.param.b64 [param1+0], %rd2172; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .b32 param3; st.param.b32 [param3+0], %r4344; .param .b32 param4; st.param.b32 [param4+0], %r4345; .param .b32 param5; st.param.b32 [param5+0], %r4384; call.uni _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 581 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3347: { // callseq 580, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4344; .param .b32 param3; st.param.b32 [param3+0], %r4345; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 580 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3380: { // callseq 618, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 618 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3315: { // callseq 539, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4344; .param .b32 param2; st.param.b32 [param2+0], %r4345; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 539 add.s32 %r5800, %r1123, 1; bra.uni $L__BB107_3146; $L__BB107_3328: ld.local.f32 %f22525, [%rd597]; st.local.f32 [%rd601], %f22525; bra.uni $L__BB107_3146; $L__BB107_3280: ld.local.f32 %f22523, [%rd600]; st.local.f32 [%rd601], %f22523; bra.uni $L__BB107_3146; $L__BB107_3148: cvt.u32.u64 %r4429, %rd408; setp.eq.s32 %p2934, %r4338, %r4429; @%p2934 bra $L__BB107_3381; bra.uni $L__BB107_3149; $L__BB107_3381: mov.f32 %f30290, %f30148; mov.f32 %f30291, %f30147; mov.f32 %f30292, %f30146; @%p2935 bra $L__BB107_3383; ld.local.f32 %f22561, [%rd280+-24]; ld.local.f32 %f22562, [%rd280+-20]; ld.local.f32 %f22563, [%rd280+-16]; sub.ftz.f32 %f30292, %f22563, %f30252; sub.ftz.f32 %f30291, %f22562, %f30254; sub.ftz.f32 %f30290, %f22561, %f30256; bra.uni $L__BB107_3383; $L__BB107_3149: ld.local.v4.f32 {%f22554, %f22555, %f22556, %f22557}, [%rd586+192]; sub.ftz.f32 %f30292, %f22556, %f30252; sub.ftz.f32 %f30291, %f22555, %f30254; sub.ftz.f32 %f30290, %f22554, %f30256; $L__BB107_3383: mul.ftz.f32 %f22564, %f30291, %f30291; fma.rn.ftz.f32 %f22565, %f30290, %f30290, %f22564; fma.rn.ftz.f32 %f22566, %f30292, %f30292, %f22565; sqrt.approx.ftz.f32 %f30257, %f22566; rsqrt.approx.ftz.f32 %f22567, %f22566; mov.f32 %f30294, 0f3F800000; mul.ftz.f32 %f5175, %f30290, %f22567; mul.ftz.f32 %f5176, %f30291, %f22567; mul.ftz.f32 %f5177, %f30292, %f22567; ld.local.u64 %rd3314, [%rd586+112]; add.s64 %rd602, %rd3314, 20; ld.f32 %f22569, [%rd3314+20]; ld.f32 %f22570, [%rd3314+24]; mul.ftz.f32 %f22571, %f22570, %f5052; fma.rn.ftz.f32 %f22572, %f22569, %f5051, %f22571; ld.f32 %f22573, [%rd3314+28]; fma.rn.ftz.f32 %f5178, %f22573, %f5053, %f22572; mul.ftz.f32 %f22574, %f5176, %f22570; fma.rn.ftz.f32 %f22575, %f5175, %f22569, %f22574; fma.rn.ftz.f32 %f5179, %f5177, %f22573, %f22575; ld.f32 %f22576, [%rd3314+40]; mul.ftz.f32 %f22577, %f5175, %f22576; mul.ftz.f32 %f22578, %f5176, %f22576; mul.ftz.f32 %f22579, %f5177, %f22576; sub.ftz.f32 %f22580, %f22577, %f30260; sub.ftz.f32 %f22581, %f22578, %f30259; sub.ftz.f32 %f22582, %f22579, %f30258; neg.ftz.f32 %f22583, %f22580; neg.ftz.f32 %f22584, %f22581; neg.ftz.f32 %f22585, %f22582; mul.ftz.f32 %f22586, %f22581, %f22581; fma.rn.ftz.f32 %f22587, %f22580, %f22580, %f22586; fma.rn.ftz.f32 %f22588, %f22582, %f22582, %f22587; sqrt.approx.ftz.f32 %f22589, %f22588; div.approx.ftz.f32 %f22590, %f22583, %f22589; div.approx.ftz.f32 %f22591, %f22584, %f22589; div.approx.ftz.f32 %f22592, %f22585, %f22589; mul.ftz.f32 %f22593, %f22591, %f5052; fma.rn.ftz.f32 %f22594, %f22590, %f5051, %f22593; fma.rn.ftz.f32 %f5180, %f22592, %f5053, %f22594; ld.v2.f32 {%f22595, %f22596}, [%rd3314+32]; mul.ftz.f32 %f22597, %f22591, %f22570; fma.rn.ftz.f32 %f22598, %f22590, %f22569, %f22597; fma.rn.ftz.f32 %f5184, %f22592, %f22573, %f22598; ld.u32 %r4430, [%rd3314+12]; setp.eq.s32 %p2936, %r4430, 22; mul.ftz.f32 %f5185, %f5178, %f5178; sub.ftz.f32 %f5186, %f30294, %f5185; @%p2936 bra $L__BB107_3385; bra.uni $L__BB107_3384; $L__BB107_3385: div.approx.ftz.f32 %f22620, %f5186, %f5185; max.ftz.f32 %f22622, %f22620, %f22516; sqrt.approx.ftz.f32 %f22623, %f22622; mul.ftz.f32 %f5188, %f22595, %f22623; setp.lt.ftz.f32 %p2937, %f5188, 0f3F200000; mov.f32 %f30293, %f30294; @%p2937 bra $L__BB107_3387; rcp.approx.ftz.f32 %f22624, %f5188; fma.rn.ftz.f32 %f22625, %f22624, 0f400B9581, 0f40623D71; mul.ftz.f32 %f22626, %f22624, %f22625; fma.rn.ftz.f32 %f22627, %f22624, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f22628, %f22624, %f22627, 0f3F800000; div.approx.ftz.f32 %f30293, %f22626, %f22628; $L__BB107_3387: mul.ftz.f32 %f22630, %f5179, %f5179; sub.ftz.f32 %f22631, %f30294, %f22630; div.approx.ftz.f32 %f22632, %f22631, %f22630; max.ftz.f32 %f22634, %f22632, %f22516; sqrt.approx.ftz.f32 %f22635, %f22634; mul.ftz.f32 %f5191, %f22595, %f22635; setp.lt.ftz.f32 %p2938, %f5191, 0f3F200000; @%p2938 bra $L__BB107_3389; rcp.approx.ftz.f32 %f22636, %f5191; fma.rn.ftz.f32 %f22637, %f22636, 0f400B9581, 0f40623D71; mul.ftz.f32 %f22638, %f22636, %f22637; fma.rn.ftz.f32 %f22639, %f22636, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f22640, %f22636, %f22639, 0f3F800000; div.approx.ftz.f32 %f30294, %f22638, %f22640; $L__BB107_3389: mul.ftz.f32 %f30295, %f30293, %f30294; bra.uni $L__BB107_3390; $L__BB107_3384: mul.ftz.f32 %f22599, %f22595, %f22596; mul.ftz.f32 %f22600, %f5186, %f22599; div.approx.ftz.f32 %f22601, %f22600, %f5185; add.ftz.f32 %f22602, %f22601, 0f3F800000; max.ftz.f32 %f22605, %f22602, %f22516; sqrt.approx.ftz.f32 %f22606, %f22605; add.ftz.f32 %f22607, %f22606, 0f3F800000; div.approx.ftz.f32 %f22609, %f22007, %f22607; mul.ftz.f32 %f22610, %f5179, %f5179; sub.ftz.f32 %f22611, %f30294, %f22610; mul.ftz.f32 %f22612, %f22611, %f22599; div.approx.ftz.f32 %f22613, %f22612, %f22610; add.ftz.f32 %f22614, %f22613, 0f3F800000; max.ftz.f32 %f22615, %f22614, %f22516; sqrt.approx.ftz.f32 %f22616, %f22615; add.ftz.f32 %f22617, %f22616, 0f3F800000; div.approx.ftz.f32 %f22618, %f22007, %f22617; mul.ftz.f32 %f30295, %f22609, %f22618; $L__BB107_3390: ld.f32 %f22641, [%rd602+-20]; mul.ftz.f32 %f22642, %f30295, %f22641; ld.f32 %f22643, [%rd602+-16]; mul.ftz.f32 %f22644, %f30295, %f22643; ld.f32 %f22645, [%rd602+-12]; mul.ftz.f32 %f22646, %f30295, %f22645; mul.ftz.f32 %f22647, %f5184, %f5184; mul.ftz.f32 %f22648, %f5178, %f22647; div.approx.ftz.f32 %f22649, %f5180, %f22648; abs.ftz.f32 %f22650, %f22649; mul.ftz.f32 %f22651, %f22642, %f22650; mul.ftz.f32 %f22652, %f22644, %f22650; mul.ftz.f32 %f22653, %f22646, %f22650; mul.ftz.f32 %f30298, %f30298, %f22651; mul.ftz.f32 %f30250, %f22652, %f30250; mul.ftz.f32 %f30249, %f22653, %f30249; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f30250; st.local.f32 [%rd283+8], %f30249; mul.ftz.f32 %f30248, %f22651, %f30248; mul.ftz.f32 %f30247, %f22652, %f30247; mul.ftz.f32 %f30246, %f22653, %f30246; st.local.f32 [%rd283+12], %f30248; st.local.f32 [%rd283+16], %f30247; st.local.f32 [%rd283+20], %f30246; mul.ftz.f32 %f30245, %f22651, %f30245; mul.ftz.f32 %f30244, %f22652, %f30244; mul.ftz.f32 %f30243, %f22653, %f30243; st.local.f32 [%rd283+24], %f30245; st.local.f32 [%rd283+28], %f30244; st.local.f32 [%rd283+32], %f30243; add.s32 %r5796, %r4338, 1; setp.lt.s32 %p2939, %r5796, %r5732; @%p2939 bra $L__BB107_4546; $L__BB107_3391: st.global.u16 [%rd404], %rs252; st.global.u16 [%rd405], %rs254; st.global.u16 [%rd278], %rs253; mov.u32 %r5802, %r5732; $L__BB107_3392: setp.gt.s32 %p2940, %r5802, 0; @%p2940 bra $L__BB107_3711; bra.uni $L__BB107_3393; $L__BB107_3711: ld.local.v4.f32 {%f30424, %f30425, %f30426, %f23530}, [%rd2173]; add.s64 %rd3732, %rd2172, 64; cvta.to.local.u64 %rd3733, %rd3732; ld.local.u32 %r1372, [%rd3733]; and.b32 %r4697, %r1372, 1; setp.eq.b32 %p3153, %r4697, 1; mov.pred %p3154, 0; xor.pred %p3155, %p3153, %p3154; not.pred %p3156, %p3155; mov.u16 %rs736, 1; @%p3156 bra $L__BB107_3755; add.s64 %rd3735, %rd2172, 48; cvta.to.local.u64 %rd3736, %rd3735; ld.local.u32 %r4698, [%rd3736]; setp.gt.s32 %p3157, %r4698, -1; @%p3157 bra $L__BB107_3755; ld.const.u64 %rd3737, [kernel_params+2160]; cvta.to.global.u64 %rd3738, %rd3737; add.s64 %rd3740, %rd2172, 76; cvta.to.local.u64 %rd3741, %rd3740; ld.local.u32 %r4699, [%rd3741]; mul.wide.s32 %rd3742, %r4699, 208; add.s64 %rd3743, %rd3738, %rd3742; add.s64 %rd752, %rd3743, 184; ld.global.f32 %f5586, [%rd3743+184]; setp.leu.ftz.f32 %p3158, %f5586, 0f00000000; @%p3158 bra $L__BB107_3755; or.b64 %rd3745, %rd2172, 12; cvta.to.local.u64 %rd3746, %rd3745; ld.local.f32 %f23531, [%rd3746]; add.s64 %rd3747, %rd2172, 16; cvta.to.local.u64 %rd3748, %rd3747; ld.local.v2.f32 {%f23532, %f23533}, [%rd3748]; mul.ftz.f32 %f23536, %f30147, %f23532; fma.rn.ftz.f32 %f23537, %f30148, %f23531, %f23536; fma.rn.ftz.f32 %f23538, %f30146, %f23533, %f23537; setp.lt.ftz.f32 %p3159, %f23538, 0f00000000; neg.ftz.f32 %f23539, %f23538; selp.f32 %f5587, %f23539, %f23538, %p3159; add.s64 %rd753, %rd2172, 24; @%p3159 bra $L__BB107_3716; bra.uni $L__BB107_3715; $L__BB107_3716: add.s64 %rd3752, %rd753, 8; cvta.to.local.u64 %rd3753, %rd753; ld.local.v2.f32 {%f23542, %f23543}, [%rd3753]; neg.ftz.f32 %f30399, %f23542; neg.ftz.f32 %f30400, %f23543; cvta.to.local.u64 %rd3754, %rd3752; ld.local.f32 %f23546, [%rd3754]; neg.ftz.f32 %f30401, %f23546; bra.uni $L__BB107_3717; $L__BB107_3393: ld.const.u64 %rd4727, [kernel_params+2408]; cvta.to.global.u64 %rd4726, %rd4727; ld.local.f32 %f5217, [%rd5+-4]; ld.local.u32 %r1233, [%rd280+40]; and.b32 %r5807, %r1233, 4194303; shl.b32 %r4432, %r1233, 5; cvt.u64.u32 %rd3316, %r4432; and.b64 %rd3317, %rd3316, 134217696; add.s64 %rd3318, %rd4726, %rd3317; add.s64 %rd604, %rd3318, 16; ld.global.u32 %r1235, [%rd3318+16]; and.b32 %r4433, %r1235, 134217728; setp.eq.s32 %p2942, %r4433, 0; mov.pred %p3927, -1; mov.f32 %f30347, 0f00000000; mov.f32 %f30348, %f30347; mov.f32 %f30349, %f30347; @%p2942 bra $L__BB107_3395; ld.global.f32 %f30347, [%rd604+-16]; ld.global.f32 %f30348, [%rd604+-12]; ld.global.f32 %f30349, [%rd604+-8]; mov.pred %p3927, 0; $L__BB107_3395: @%p3927 bra $L__BB107_3399; bra.uni $L__BB107_3396; $L__BB107_3399: ld.local.u32 %r1236, [%rd280+52]; setp.eq.s32 %p2946, %r1236, 2; ld.local.f32 %f5227, [%rd280+-24]; ld.local.f32 %f5228, [%rd280+-20]; ld.local.f32 %f5229, [%rd280+-16]; or.b64 %rd605, %rd2172, 12; neg.ftz.f32 %f5230, %f30148; neg.ftz.f32 %f5231, %f30147; neg.ftz.f32 %f5232, %f30146; @%p2946 bra $L__BB107_3437; bra.uni $L__BB107_3400; $L__BB107_3437: ld.const.u64 %rd4730, [kernel_params+2408]; cvta.to.global.u64 %rd4729, %rd4730; add.s64 %rd3482, %rd605, 20; add.s64 %rd3483, %rd605, 4; add.s64 %rd3485, %rd2172, 248; cvta.to.local.u64 %rd3486, %rd3485; add.s64 %rd3487, %rd2172, 240; cvta.to.local.u64 %rd3488, %rd3487; st.local.v2.f32 [%rd3488], {%f5227, %f5228}; st.local.f32 [%rd3486], %f5229; st.local.v4.f32 [%rd2173], {%f30148, %f30147, %f30146, %f5230}; cvta.to.local.u64 %rd3490, %rd3483; st.local.v4.f32 [%rd3490], {%f5231, %f5232, %f5230, %f5231}; cvta.to.local.u64 %rd3491, %rd3482; st.local.v4.f32 [%rd3491], {%f5232, %f5230, %f5231, %f5232}; ld.const.u32 %r4465, [kernel_params+1024]; and.b32 %r5807, %r4465, 4194303; shl.b32 %r4466, %r4465, 5; cvt.u64.u32 %rd3492, %r4466; and.b64 %rd3493, %rd3492, 134217696; add.s64 %rd3494, %rd4729, %rd3493; add.s64 %rd3495, %rd2172, 76; cvta.to.local.u64 %rd3496, %rd3495; ld.global.u32 %r4467, [%rd3494+16]; mov.u32 %r4468, -1; st.local.u32 [%rd3496], %r4468; mov.u32 %r4469, 0; add.s64 %rd3497, %rd605, 36; cvta.to.local.u64 %rd3498, %rd3497; st.local.v4.u32 [%rd3498], {%r4465, %r4467, %r4469, %r4468}; add.s64 %rd3499, %rd2172, 68; cvta.to.local.u64 %rd3500, %rd3499; st.local.u32 [%rd3500], %r4469; add.s64 %rd3501, %rd2172, 72; cvta.to.local.u64 %rd3502, %rd3501; st.local.u32 [%rd3502], %r4469; add.s64 %rd3503, %rd2172, 136; cvta.to.local.u64 %rd3504, %rd3503; add.s64 %rd3505, %rd2172, 128; cvta.to.local.u64 %rd3506, %rd3505; mov.f32 %f23140, 0f00000000; st.local.v2.f32 [%rd3506], {%f23140, %f23140}; st.local.u32 [%rd3504], %r4469; mov.b32 %r4470, %f5217; add.s64 %rd3507, %rd2172, 80; cvta.to.local.u64 %rd3508, %rd3507; st.local.v4.u32 [%rd3508], {%r4468, %r4470, %r4469, %r4469}; add.s64 %rd3509, %rd2172, 96; cvta.to.local.u64 %rd3510, %rd3509; st.local.v4.f32 [%rd3510], {%f23140, %f23140, %f23140, %f23140}; add.s64 %rd3511, %rd2172, 112; cvta.to.local.u64 %rd3512, %rd3511; st.local.v4.f32 [%rd3512], {%f23140, %f23140, %f23140, %f23140}; bra.uni $L__BB107_3438; $L__BB107_3396: ld.local.u32 %r4434, [%rd280+36]; setp.eq.s32 %p2944, %r4434, -1; @%p2944 bra $L__BB107_3649; ld.local.f32 %f5224, [%rd280+-12]; ld.local.f32 %f5225, [%rd280+-8]; mul.ftz.f32 %f22657, %f5225, %f30147; fma.rn.ftz.f32 %f22658, %f5224, %f30148, %f22657; ld.local.f32 %f5226, [%rd280+-4]; fma.rn.ftz.f32 %f22659, %f5226, %f30146, %f22658; setp.leu.ftz.f32 %p2945, %f22659, 0f00000000; @%p2945 bra $L__BB107_3649; neg.ftz.f32 %f22660, %f5224; st.local.f32 [%rd280+-12], %f22660; neg.ftz.f32 %f22661, %f5225; st.local.f32 [%rd280+-8], %f22661; neg.ftz.f32 %f22662, %f5226; st.local.f32 [%rd280+-4], %f22662; bra.uni $L__BB107_3649; $L__BB107_3400: add.s64 %rd3320, %rd605, 20; add.s64 %rd3321, %rd605, 4; add.s64 %rd3322, %rd605, 36; ld.local.f32 %f30339, [%rd280+-12]; ld.local.f32 %f30340, [%rd280+-8]; ld.local.f32 %f5235, [%rd280+-4]; ld.local.u32 %r1237, [%rd280+32]; cvt.s64.s32 %rd606, %r1237; ld.local.u32 %r1238, [%rd280+36]; cvt.s64.s32 %rd607, %r1238; ld.local.f32 %f5236, [%rd280+16]; ld.local.f32 %f5237, [%rd280+20]; ld.local.f32 %f5238, [%rd280+12]; st.local.v4.f32 [%rd2173], {%f5227, %f5228, %f5229, %f30339}; cvta.to.local.u64 %rd608, %rd3321; st.local.v4.f32 [%rd608], {%f30340, %f5235, %f30339, %f30340}; cvta.to.local.u64 %rd3325, %rd3320; st.local.v4.f32 [%rd3325], {%f5235, %f5230, %f5231, %f5232}; cvta.to.local.u64 %rd3326, %rd3322; st.local.u32 [%rd3326], %r1233; setp.eq.s32 %p2947, %r1238, -1; add.s64 %rd3327, %rd2172, 64; cvta.to.local.u64 %rd609, %rd3327; @%p2947 bra $L__BB107_3402; mov.u32 %r5804, 1; st.local.u32 [%rd609], %r5804; bra.uni $L__BB107_3405; $L__BB107_3402: @%p1255 bra $L__BB107_3404; mov.u32 %r5804, 32; st.local.u32 [%rd609], %r5804; bra.uni $L__BB107_3405; $L__BB107_3715: add.s64 %rd3749, %rd753, 8; cvta.to.local.u64 %rd3750, %rd753; ld.local.v2.f32 {%f30399, %f30400}, [%rd3750]; cvta.to.local.u64 %rd3751, %rd3749; ld.local.f32 %f30401, [%rd3751]; $L__BB107_3717: mul.ftz.f32 %f23547, %f30147, %f30400; fma.rn.ftz.f32 %f23548, %f30148, %f30399, %f23547; fma.rn.ftz.f32 %f5597, %f30146, %f30401, %f23548; add.ftz.f32 %f23549, %f5587, %f5597; setp.lt.ftz.f32 %p3160, %f5587, %f5586; selp.f32 %f23550, 0f40000000, 0f3F800000, %p3160; mov.f32 %f23551, 0f3F800000; selp.f32 %f23552, %f23549, %f5597, %p3160; div.approx.ftz.f32 %f23553, %f23552, %f5586; sub.ftz.f32 %f23554, %f23550, %f23553; mov.f32 %f23555, 0f00000000; max.ftz.f32 %f23556, %f23554, %f23555; min.ftz.f32 %f5598, %f23556, %f23551; setp.leu.ftz.f32 %p3161, %f5598, 0f00000000; @%p3161 bra $L__BB107_3755; setp.eq.s32 %p3162, %r1372, 65; add.s64 %rd3756, %rd2172, 60; cvta.to.local.u64 %rd3757, %rd3756; ld.local.s32 %rd754, [%rd3757]; @%p3162 bra $L__BB107_3720; bra.uni $L__BB107_3719; $L__BB107_3720: add.s64 %rd3781, %rd2172, 84; cvta.to.local.u64 %rd3782, %rd3781; ld.global.u32 %r1373, [%rd752+-28]; ld.global.u32 %r1374, [%rd752+-32]; shl.b32 %r4711, %r1374, 1; cvt.rn.f32.s32 %f23557, %r4711; ld.local.f32 %f23558, [%rd3782]; mul.ftz.f32 %f23559, %f23558, %f23557; cvt.rzi.ftz.s32.f32 %r4712, %f23559; add.s32 %r4713, %r4711, -1; min.s32 %r1375, %r4712, %r4713; cvt.rn.f32.s32 %f23560, %r1375; sub.ftz.f32 %f5617, %f23559, %f23560; ld.const.u64 %rd3783, [kernel_params+2312]; cvta.to.global.u64 %rd755, %rd3783; ld.global.u32 %r5824, [%rd752+-20]; mul.wide.u32 %rd3784, %r5824, 16; add.s64 %rd3785, %rd755, %rd3784; ld.global.u16 %rs734, [%rd3785+12]; ld.global.u32 %r5823, [%rd3785+8]; ld.global.u64 %rd4892, [%rd3785]; setp.eq.s64 %p3163, %rd4892, 11; mov.u16 %rs732, %rs734; mov.u32 %r5818, %r5823; @%p3163 bra $L__BB107_3726; mov.u64 %rd4873, %rd4892; mov.u32 %r5818, %r5823; mov.u16 %rs732, %rs734; mov.u32 %r5819, %r5824; $L__BB107_3722: setp.eq.s64 %p3164, %rd4873, 0; @%p3164 bra $L__BB107_3724; add.s32 %r5819, %r5819, 2; bra.uni $L__BB107_3725; $L__BB107_3724: setp.eq.s16 %p3165, %rs732, 0; mov.u32 %r5822, -1; mov.u32 %r5819, %r5818; @%p3165 bra $L__BB107_3727; $L__BB107_3725: mul.wide.u32 %rd3786, %r5819, 16; add.s64 %rd3787, %rd755, %rd3786; ld.global.u16 %rs732, [%rd3787+12]; ld.global.u32 %r5818, [%rd3787+8]; ld.global.u64 %rd4873, [%rd3787]; setp.ne.s64 %p3166, %rd4873, 11; @%p3166 bra $L__BB107_3722; $L__BB107_3726: setp.eq.s16 %p3167, %rs732, 0; selp.b32 %r5822, -1, %r5818, %p3167; $L__BB107_3727: ld.const.u64 %rd3788, [kernel_params+2232]; cvta.to.global.u64 %rd3789, %rd3788; shl.b64 %rd3790, %rd754, 4; add.s64 %rd3791, %rd3789, %rd3790; ld.global.v4.u32 {%r4715, %r4716, %r4717, %r4718}, [%rd3791]; setp.eq.s32 %p3168, %r1375, %r1374; @%p3168 bra $L__BB107_3729; bra.uni $L__BB107_3728; $L__BB107_3729: ld.const.u64 %rd3797, [kernel_params+2256]; cvta.to.global.u64 %rd3798, %rd3797; mul.wide.u32 %rd3799, %r4718, 12; add.s64 %rd4874, %rd3798, %rd3799; add.s32 %r4725, %r4718, 1; mul.wide.u32 %rd3800, %r4725, 12; add.s64 %rd4877, %rd3798, %rd3800; add.s32 %r4726, %r4718, 2; mul.wide.u32 %rd3801, %r4726, 12; add.s64 %rd4880, %rd3798, %rd3801; add.s64 %rd4875, %rd4874, 4; add.s64 %rd4876, %rd4874, 8; add.s64 %rd4878, %rd4877, 4; add.s64 %rd4879, %rd4877, 8; add.s64 %rd4881, %rd4880, 4; add.s64 %rd4882, %rd4880, 8; bra.uni $L__BB107_3730; $L__BB107_3719: cvt.u32.u64 %r4700, %rd754; ld.const.v2.u64 {%rd3758, %rd3759}, [kernel_params+2224]; cvta.to.global.u64 %rd3762, %rd3759; mul.wide.s32 %rd3763, %r4700, 16; add.s64 %rd3764, %rd3762, %rd3763; ld.global.v4.u32 {%r4701, %r4702, %r4703, %r4704}, [%rd3764]; ld.const.u64 %rd3765, [kernel_params+2256]; cvta.to.global.u64 %rd3766, %rd3765; mul.wide.u32 %rd3767, %r4704, 12; add.s64 %rd3768, %rd3766, %rd3767; ld.global.f32 %f30411, [%rd3768]; ld.global.f32 %f30412, [%rd3768+4]; ld.global.f32 %f30413, [%rd3768+8]; add.s32 %r4709, %r4704, 1; mul.wide.u32 %rd3769, %r4709, 12; add.s64 %rd3770, %rd3766, %rd3769; ld.global.f32 %f30414, [%rd3770]; ld.global.f32 %f30415, [%rd3770+4]; ld.global.f32 %f30416, [%rd3770+8]; add.s32 %r4710, %r4704, 2; mul.wide.u32 %rd3771, %r4710, 12; add.s64 %rd3772, %rd3766, %rd3771; ld.global.f32 %f30417, [%rd3772]; ld.global.f32 %f30418, [%rd3772+4]; ld.global.f32 %f30419, [%rd3772+8]; cvta.to.global.u64 %rd3773, %rd3758; mul.wide.u32 %rd3774, %r4701, 12; add.s64 %rd3775, %rd3773, %rd3774; ld.global.f32 %f30402, [%rd3775]; ld.global.f32 %f30403, [%rd3775+4]; ld.global.f32 %f30404, [%rd3775+8]; mul.wide.u32 %rd3776, %r4702, 12; add.s64 %rd3777, %rd3773, %rd3776; ld.global.f32 %f30405, [%rd3777]; ld.global.f32 %f30406, [%rd3777+4]; ld.global.f32 %f30407, [%rd3777+8]; mul.wide.u32 %rd3778, %r4703, 12; add.s64 %rd3779, %rd3773, %rd3778; ld.global.f32 %f30408, [%rd3779]; ld.global.f32 %f30409, [%rd3779+4]; ld.global.f32 %f30410, [%rd3779+8]; bra.uni $L__BB107_3747; $L__BB107_3404: mov.u32 %r5804, 0; st.local.u32 [%rd609], %r5804; $L__BB107_3405: add.s64 %rd3329, %rd2172, 76; cvta.to.local.u64 %rd3330, %rd3329; st.local.u32 [%rd3330], %r1237; add.s64 %rd3331, %rd2172, 60; cvta.to.local.u64 %rd3332, %rd3331; st.local.u32 [%rd3332], %r1238; add.s64 %rd3333, %rd2172, 68; cvta.to.local.u64 %rd3334, %rd3333; st.local.f32 [%rd3334], %f5236; add.s64 %rd3335, %rd2172, 72; cvta.to.local.u64 %rd3336, %rd3335; st.local.f32 [%rd3336], %f5237; mov.b32 %r4438, %f5217; mov.u32 %r4439, -1; add.s64 %rd610, %rd2172, 80; cvta.to.local.u64 %rd3337, %rd610; st.local.v2.u32 [%rd3337], {%r4439, %r4438}; add.s64 %rd3338, %rd2172, 88; cvta.to.local.u64 %rd3339, %rd3338; st.local.f32 [%rd3339], %f5238; add.s64 %rd3340, %rd2172, 52; cvta.to.local.u64 %rd611, %rd3340; st.local.u32 [%rd611], %r1235; add.s64 %rd3341, %rd2172, 56; cvta.to.local.u64 %rd612, %rd3341; mov.u32 %r4440, 0; st.local.u32 [%rd612], %r4440; setp.eq.s32 %p2949, %r1237, -1; @%p2949 bra $L__BB107_3430; ld.const.u64 %rd4715, [kernel_params+2184]; cvta.to.global.u64 %rd4714, %rd4715; shl.b64 %rd3344, %rd606, 2; add.s64 %rd3345, %rd4714, %rd3344; ld.global.u32 %r1240, [%rd3345]; st.local.u32 [%rd612], %r1240; and.b32 %r1241, %r1240, 2; setp.eq.s32 %p2950, %r1241, 0; @%p2950 bra $L__BB107_3413; ld.const.u64 %rd3346, [kernel_params+2160]; cvta.to.global.u64 %rd3347, %rd3346; mul.lo.s64 %rd3348, %rd606, 208; add.s64 %rd3349, %rd3347, %rd3348; ld.const.u64 %rd3350, [kernel_params+2176]; cvta.to.global.u64 %rd3351, %rd3350; ld.global.u32 %rd3352, [%rd3349+168]; ld.global.u32 %r4441, [%rd3349+152]; shl.b32 %r4442, %r4441, 1; cvt.rn.f32.s32 %f22663, %r4442; mul.ftz.f32 %f22664, %f5217, %f22663; cvt.rzi.ftz.s32.f32 %r4443, %f22664; add.s32 %r4444, %r4442, -1; min.s32 %r4445, %r4443, %r4444; cvt.rn.f32.s32 %f22665, %r4445; sub.ftz.f32 %f5247, %f22664, %f22665; cvt.s64.s32 %rd3353, %r4445; add.s64 %rd3354, %rd3353, %rd3352; shl.b64 %rd3355, %rd3354, 6; add.s64 %rd613, %rd3351, %rd3355; ld.global.v4.f32 {%f22666, %f22667, %f22668, %f22669}, [%rd613]; ld.global.v4.f32 {%f22670, %f22671, %f22672, %f22673}, [%rd613+64]; mul.ftz.f32 %f22674, %f22667, %f22671; fma.rn.ftz.f32 %f22675, %f22666, %f22670, %f22674; mul.ftz.f32 %f22676, %f22669, %f22673; fma.rn.ftz.f32 %f22677, %f22668, %f22672, %f22676; add.ftz.f32 %f5256, %f22675, %f22677; setp.gt.ftz.f32 %p2951, %f5256, 0f3F7FDF3B; mov.f32 %f22678, 0f3F800000; sub.ftz.f32 %f5257, %f22678, %f5247; @%p2951 bra $L__BB107_3409; bra.uni $L__BB107_3408; $L__BB107_3409: mul.ftz.f32 %f22732, %f22666, %f5257; fma.rn.ftz.f32 %f22733, %f22670, %f5247, %f22732; mul.ftz.f32 %f22734, %f22667, %f5257; fma.rn.ftz.f32 %f22735, %f22671, %f5247, %f22734; mul.ftz.f32 %f22736, %f22668, %f5257; fma.rn.ftz.f32 %f22737, %f22672, %f5247, %f22736; mul.ftz.f32 %f22738, %f22669, %f5257; fma.rn.ftz.f32 %f22739, %f22673, %f5247, %f22738; mul.ftz.f32 %f22740, %f22735, %f22735; fma.rn.ftz.f32 %f22741, %f22733, %f22733, %f22740; mul.ftz.f32 %f22742, %f22739, %f22739; fma.rn.ftz.f32 %f22743, %f22737, %f22737, %f22742; add.ftz.f32 %f22744, %f22741, %f22743; rsqrt.approx.ftz.f32 %f22745, %f22744; mul.ftz.f32 %f30312, %f22739, %f22745; mul.ftz.f32 %f30311, %f22737, %f22745; mul.ftz.f32 %f30310, %f22735, %f22745; mul.ftz.f32 %f30309, %f22733, %f22745; bra.uni $L__BB107_3410; $L__BB107_3430: @%p1255 bra $L__BB107_3432; st.local.u32 [%rd3337], %r618; $L__BB107_3432: add.s64 %rd3450, %rd2172, 116; cvta.to.local.u64 %rd3451, %rd3450; mov.u32 %r4459, 0; st.local.u32 [%rd3451], %r4459; add.s64 %rd3452, %rd2172, 120; cvta.to.local.u64 %rd3453, %rd3452; st.local.u32 [%rd3453], %r4459; add.s64 %rd3454, %rd2172, 124; cvta.to.local.u64 %rd3455, %rd3454; st.local.u32 [%rd3455], %r4459; add.s64 %rd3456, %rd2172, 136; cvta.to.local.u64 %rd3457, %rd3456; add.s64 %rd3458, %rd2172, 128; cvta.to.local.u64 %rd3459, %rd3458; mov.f32 %f30329, 0f00000000; st.local.v2.f32 [%rd3459], {%f30329, %f30329}; st.local.u32 [%rd3457], %r4459; mov.f32 %f30330, %f30329; mov.f32 %f30331, %f30329; mov.f32 %f30332, %f30329; mov.f32 %f30333, %f30329; mov.f32 %f30334, %f30329; mov.f32 %f30335, %f5235; mov.f32 %f30336, %f30340; mov.f32 %f30337, %f30339; bra.uni $L__BB107_3433; $L__BB107_3408: mov.f32 %f22679, 0fBF800000; max.ftz.f32 %f22680, %f5256, %f22679; min.ftz.f32 %f22682, %f22680, %f22678; abs.ftz.f32 %f22683, %f22682; sub.ftz.f32 %f22684, %f22678, %f22683; mul.ftz.f32 %f22685, %f22684, 0f3F000000; sqrt.approx.ftz.f32 %f22686, %f22685; setp.gt.ftz.f32 %p2952, %f22683, 0f3F11EB85; selp.f32 %f22687, %f22686, %f22683, %p2952; mul.ftz.f32 %f22688, %f22687, %f22687; mov.f32 %f22689, 0f3C94D2E9; mov.f32 %f22690, 0f3D53F941; fma.rn.ftz.f32 %f22691, %f22690, %f22688, %f22689; mov.f32 %f22692, 0f3D3F841F; fma.rn.ftz.f32 %f22693, %f22691, %f22688, %f22692; mov.f32 %f22694, 0f3D994929; fma.rn.ftz.f32 %f22695, %f22693, %f22688, %f22694; mov.f32 %f22696, 0f3E2AAB94; fma.rn.ftz.f32 %f22697, %f22695, %f22688, %f22696; mul.ftz.f32 %f22698, %f22688, %f22697; fma.rn.ftz.f32 %f22699, %f22698, %f22687, %f22687; add.ftz.f32 %f22700, %f22699, %f22699; mov.f32 %f22701, 0f3FC90FDB; sub.ftz.f32 %f22702, %f22701, %f22699; selp.f32 %f22703, %f22700, %f22702, %p2952; setp.lt.ftz.f32 %p2953, %f22682, 0f00000000; mov.f32 %f22704, 0f40490FDB; sub.ftz.f32 %f22705, %f22704, %f22703; selp.f32 %f22706, %f22705, %f22703, %p2953; mul.ftz.f32 %f22707, %f22666, %f5256; sub.ftz.f32 %f22708, %f22670, %f22707; mul.ftz.f32 %f22709, %f22667, %f5256; sub.ftz.f32 %f22710, %f22671, %f22709; mul.ftz.f32 %f22711, %f22668, %f5256; sub.ftz.f32 %f22712, %f22672, %f22711; mul.ftz.f32 %f22713, %f22669, %f5256; sub.ftz.f32 %f22714, %f22673, %f22713; mul.ftz.f32 %f22715, %f22710, %f22710; fma.rn.ftz.f32 %f22716, %f22708, %f22708, %f22715; mul.ftz.f32 %f22717, %f22714, %f22714; fma.rn.ftz.f32 %f22718, %f22712, %f22712, %f22717; add.ftz.f32 %f22719, %f22716, %f22718; rsqrt.approx.ftz.f32 %f22720, %f22719; mul.ftz.f32 %f22721, %f22708, %f22720; mul.ftz.f32 %f22722, %f22710, %f22720; mul.ftz.f32 %f22723, %f22712, %f22720; mul.ftz.f32 %f22724, %f22714, %f22720; mul.ftz.f32 %f22725, %f5247, %f22706; cos.approx.ftz.f32 %f22726, %f22725; sin.approx.ftz.f32 %f22727, %f22725; mul.ftz.f32 %f22728, %f22721, %f22727; mul.ftz.f32 %f22729, %f22722, %f22727; mul.ftz.f32 %f22730, %f22723, %f22727; mul.ftz.f32 %f22731, %f22724, %f22727; fma.rn.ftz.f32 %f30312, %f22669, %f22726, %f22731; fma.rn.ftz.f32 %f30311, %f22668, %f22726, %f22730; fma.rn.ftz.f32 %f30310, %f22667, %f22726, %f22729; fma.rn.ftz.f32 %f30309, %f22666, %f22726, %f22728; $L__BB107_3410: ld.global.v4.f32 {%f22746, %f22747, %f22748, %f22749}, [%rd613+16]; ld.global.v4.f32 {%f22754, %f22755, %f22756, %f22757}, [%rd613+80]; mul.ftz.f32 %f22762, %f5247, %f22754; mul.ftz.f32 %f22763, %f5247, %f22755; mul.ftz.f32 %f22764, %f5247, %f22756; mul.ftz.f32 %f22765, %f5247, %f22757; fma.rn.ftz.f32 %f5270, %f5257, %f22746, %f22762; fma.rn.ftz.f32 %f5271, %f5257, %f22747, %f22763; fma.rn.ftz.f32 %f5272, %f5257, %f22748, %f22764; fma.rn.ftz.f32 %f22766, %f5257, %f22749, %f22765; ld.global.v4.f32 {%f22767, %f22768, %f22769, %f22770}, [%rd613+32]; ld.global.v4.f32 {%f22775, %f22776, %f22777, %f22778}, [%rd613+96]; mul.ftz.f32 %f22783, %f5247, %f22775; mul.ftz.f32 %f22784, %f5247, %f22776; mul.ftz.f32 %f22785, %f5247, %f22777; mul.ftz.f32 %f22786, %f5247, %f22778; fma.rn.ftz.f32 %f22787, %f5257, %f22767, %f22783; fma.rn.ftz.f32 %f22788, %f5257, %f22768, %f22784; fma.rn.ftz.f32 %f22789, %f5257, %f22769, %f22785; fma.rn.ftz.f32 %f22790, %f5257, %f22770, %f22786; ld.global.v4.f32 {%f22791, %f22792, %f22793, %f22794}, [%rd613+48]; ld.global.v4.f32 {%f22799, %f22800, %f22801, %f22802}, [%rd613+112]; mul.ftz.f32 %f22807, %f5247, %f22799; mul.ftz.f32 %f22808, %f5247, %f22800; mul.ftz.f32 %f22809, %f5247, %f22801; mul.ftz.f32 %f22810, %f5247, %f22802; fma.rn.ftz.f32 %f22811, %f5257, %f22791, %f22807; fma.rn.ftz.f32 %f22812, %f5257, %f22792, %f22808; fma.rn.ftz.f32 %f22813, %f5257, %f22793, %f22809; fma.rn.ftz.f32 %f22814, %f5257, %f22794, %f22810; mul.ftz.f32 %f22815, %f30312, 0f3FB504F3; mul.ftz.f32 %f22816, %f30309, 0f3FB504F3; mul.ftz.f32 %f22817, %f30310, 0f3FB504F3; mul.ftz.f32 %f22818, %f30311, 0f3FB504F3; mul.ftz.f32 %f22819, %f22815, %f22816; mul.ftz.f32 %f22820, %f22815, %f22817; mul.ftz.f32 %f22821, %f22815, %f22818; mul.ftz.f32 %f22822, %f22816, %f22816; mul.ftz.f32 %f22823, %f22816, %f22817; mul.ftz.f32 %f22824, %f22816, %f22818; mul.ftz.f32 %f22825, %f22817, %f22817; mul.ftz.f32 %f22826, %f22817, %f22818; mul.ftz.f32 %f22827, %f22818, %f22818; sub.ftz.f32 %f22829, %f22678, %f22825; sub.ftz.f32 %f22830, %f22829, %f22827; sub.ftz.f32 %f22831, %f22823, %f22821; add.ftz.f32 %f22832, %f22820, %f22824; add.ftz.f32 %f22833, %f22821, %f22823; sub.ftz.f32 %f22834, %f22678, %f22822; sub.ftz.f32 %f22835, %f22834, %f22827; sub.ftz.f32 %f22836, %f22826, %f22819; sub.ftz.f32 %f22837, %f22824, %f22820; add.ftz.f32 %f22838, %f22819, %f22826; sub.ftz.f32 %f22839, %f22834, %f22825; mul.ftz.f32 %f22840, %f22831, %f22789; fma.rn.ftz.f32 %f22841, %f22830, %f22766, %f22840; fma.rn.ftz.f32 %f30318, %f22832, %f22812, %f22841; mul.ftz.f32 %f22842, %f22831, %f22790; fma.rn.ftz.f32 %f22843, %f22830, %f22787, %f22842; fma.rn.ftz.f32 %f5274, %f22832, %f22813, %f22843; mul.ftz.f32 %f22844, %f22831, %f22811; fma.rn.ftz.f32 %f22845, %f22830, %f22788, %f22844; fma.rn.ftz.f32 %f5275, %f22832, %f22814, %f22845; mul.ftz.f32 %f22846, %f22835, %f22789; fma.rn.ftz.f32 %f22847, %f22833, %f22766, %f22846; fma.rn.ftz.f32 %f5276, %f22836, %f22812, %f22847; mul.ftz.f32 %f22848, %f22835, %f22790; fma.rn.ftz.f32 %f22849, %f22833, %f22787, %f22848; fma.rn.ftz.f32 %f30317, %f22836, %f22813, %f22849; mul.ftz.f32 %f22850, %f22835, %f22811; fma.rn.ftz.f32 %f22851, %f22833, %f22788, %f22850; fma.rn.ftz.f32 %f5278, %f22836, %f22814, %f22851; mul.ftz.f32 %f22852, %f22838, %f22789; fma.rn.ftz.f32 %f22853, %f22837, %f22766, %f22852; fma.rn.ftz.f32 %f5279, %f22839, %f22812, %f22853; mul.ftz.f32 %f22854, %f22838, %f22790; fma.rn.ftz.f32 %f22855, %f22837, %f22787, %f22854; fma.rn.ftz.f32 %f5280, %f22839, %f22813, %f22855; mul.ftz.f32 %f22856, %f22838, %f22811; fma.rn.ftz.f32 %f22857, %f22837, %f22788, %f22856; fma.rn.ftz.f32 %f30316, %f22839, %f22814, %f22857; add.s64 %rd3357, %rd2172, 144; cvta.to.local.u64 %rd3358, %rd3357; st.local.v4.f32 [%rd3358], {%f30318, %f5274, %f5275, %f5270}; add.s64 %rd3359, %rd2172, 160; cvta.to.local.u64 %rd3360, %rd3359; st.local.v4.f32 [%rd3360], {%f5276, %f30317, %f5278, %f5271}; add.s64 %rd3361, %rd2172, 176; cvta.to.local.u64 %rd3362, %rd3361; st.local.v4.f32 [%rd3362], {%f5279, %f5280, %f30316, %f5272}; mul.ftz.f32 %f22858, %f30317, %f30316; mul.ftz.f32 %f5282, %f5280, %f5278; sub.ftz.f32 %f30315, %f22858, %f5282; mul.ftz.f32 %f5284, %f5280, %f5275; mul.ftz.f32 %f22859, %f5274, %f30316; sub.ftz.f32 %f30314, %f5284, %f22859; mul.ftz.f32 %f5286, %f5274, %f5278; mul.ftz.f32 %f22860, %f30317, %f5275; sub.ftz.f32 %f30313, %f5286, %f22860; mul.ftz.f32 %f22861, %f5276, %f30314; fma.rn.ftz.f32 %f22862, %f30318, %f30315, %f22861; fma.rn.ftz.f32 %f30319, %f5279, %f30313, %f22862; setp.neu.ftz.f32 %p2954, %f30319, 0f00000000; @%p2954 bra $L__BB107_3412; add.ftz.f32 %f30318, %f30318, 0f322BCC77; add.ftz.f32 %f30316, %f30316, 0f322BCC77; add.ftz.f32 %f30317, %f30317, 0f322BCC77; mul.ftz.f32 %f22863, %f30317, %f30316; sub.ftz.f32 %f30315, %f22863, %f5282; mul.ftz.f32 %f22864, %f5274, %f30316; sub.ftz.f32 %f30314, %f5284, %f22864; mul.ftz.f32 %f22865, %f30317, %f5275; sub.ftz.f32 %f30313, %f5286, %f22865; mul.ftz.f32 %f22866, %f5276, %f30314; fma.rn.ftz.f32 %f22867, %f30318, %f30315, %f22866; fma.rn.ftz.f32 %f22868, %f5279, %f30313, %f22867; setp.eq.ftz.f32 %p2955, %f22868, 0f00000000; selp.f32 %f30319, 0f7F7FFFFF, %f22868, %p2955; $L__BB107_3412: mul.ftz.f32 %f22869, %f5276, %f30316; mul.ftz.f32 %f22870, %f5279, %f5278; sub.ftz.f32 %f22871, %f22870, %f22869; mul.ftz.f32 %f22872, %f5279, %f5275; mul.ftz.f32 %f22873, %f30316, %f30318; sub.ftz.f32 %f22874, %f22873, %f22872; mul.ftz.f32 %f22875, %f5278, %f30318; mul.ftz.f32 %f22876, %f5276, %f5275; sub.ftz.f32 %f22877, %f22876, %f22875; mul.ftz.f32 %f22878, %f5279, %f30317; mul.ftz.f32 %f22879, %f5276, %f5280; sub.ftz.f32 %f22880, %f22879, %f22878; mul.ftz.f32 %f22881, %f5280, %f30318; mul.ftz.f32 %f22882, %f5279, %f5274; sub.ftz.f32 %f22883, %f22882, %f22881; mul.ftz.f32 %f22884, %f5276, %f5274; mul.ftz.f32 %f22885, %f30317, %f30318; sub.ftz.f32 %f22886, %f22885, %f22884; div.approx.ftz.f32 %f22887, %f30314, %f30319; mul.ftz.f32 %f22888, %f5271, %f22887; div.approx.ftz.f32 %f22889, %f30315, %f30319; fma.rn.ftz.f32 %f22890, %f5270, %f22889, %f22888; div.approx.ftz.f32 %f22891, %f30313, %f30319; fma.rn.ftz.f32 %f22892, %f5272, %f22891, %f22890; div.approx.ftz.f32 %f22893, %f22874, %f30319; mul.ftz.f32 %f22894, %f5271, %f22893; div.approx.ftz.f32 %f22895, %f22871, %f30319; fma.rn.ftz.f32 %f22896, %f5270, %f22895, %f22894; div.approx.ftz.f32 %f22897, %f22877, %f30319; fma.rn.ftz.f32 %f22898, %f5272, %f22897, %f22896; div.approx.ftz.f32 %f22899, %f22883, %f30319; mul.ftz.f32 %f22900, %f5271, %f22899; div.approx.ftz.f32 %f22901, %f22880, %f30319; fma.rn.ftz.f32 %f22902, %f5270, %f22901, %f22900; div.approx.ftz.f32 %f22903, %f22886, %f30319; fma.rn.ftz.f32 %f22904, %f5272, %f22903, %f22902; neg.ftz.f32 %f22905, %f22892; add.s64 %rd3364, %rd2172, 192; cvta.to.local.u64 %rd3365, %rd3364; st.local.v4.f32 [%rd3365], {%f22889, %f22887, %f22891, %f22905}; neg.ftz.f32 %f22906, %f22898; add.s64 %rd3366, %rd2172, 208; cvta.to.local.u64 %rd3367, %rd3366; st.local.v4.f32 [%rd3367], {%f22895, %f22893, %f22897, %f22906}; neg.ftz.f32 %f22907, %f22904; add.s64 %rd3368, %rd2172, 224; cvta.to.local.u64 %rd3369, %rd3368; st.local.v4.f32 [%rd3369], {%f22901, %f22899, %f22903, %f22907}; $L__BB107_3413: setp.eq.s32 %p2956, %r5804, 1; @%p2956 bra $L__BB107_3415; bra.uni $L__BB107_3414; $L__BB107_3415: setp.lt.s32 %p2957, %r1233, 0; @%p2957 bra $L__BB107_3417; and.b32 %r5805, %r1240, 4; mov.f32 %f30335, %f5235; mov.f32 %f30336, %f30340; mov.f32 %f30337, %f30339; bra.uni $L__BB107_3423; $L__BB107_3414: add.s64 %rd3371, %rd2172, 116; cvta.to.local.u64 %rd3372, %rd3371; st.local.u32 [%rd3372], %r4440; add.s64 %rd3373, %rd2172, 120; cvta.to.local.u64 %rd3374, %rd3373; st.local.u32 [%rd3374], %r4440; add.s64 %rd3375, %rd2172, 124; cvta.to.local.u64 %rd3376, %rd3375; st.local.u32 [%rd3376], %r4440; add.s64 %rd3377, %rd2172, 136; cvta.to.local.u64 %rd3378, %rd3377; add.s64 %rd3379, %rd2172, 128; cvta.to.local.u64 %rd3380, %rd3379; mov.f32 %f30329, 0f00000000; st.local.v2.f32 [%rd3380], {%f30329, %f30329}; st.local.u32 [%rd3378], %r4440; mov.f32 %f30330, %f30329; mov.f32 %f30331, %f30329; mov.f32 %f30332, %f30329; mov.f32 %f30333, %f30329; mov.f32 %f30334, %f30329; mov.f32 %f30335, %f5235; mov.f32 %f30336, %f30340; mov.f32 %f30337, %f30339; bra.uni $L__BB107_3433; $L__BB107_3417: ld.const.v2.u64 {%rd3381, %rd3382}, [kernel_params+2224]; cvta.to.global.u64 %rd3385, %rd3382; shl.b64 %rd3386, %rd607, 4; add.s64 %rd3387, %rd3385, %rd3386; ld.global.v4.u32 {%r4447, %r4448, %r4449, %r4450}, [%rd3387]; cvta.to.global.u64 %rd3388, %rd3381; mul.wide.u32 %rd3389, %r4447, 12; add.s64 %rd3390, %rd3388, %rd3389; mul.wide.u32 %rd3391, %r4448, 12; add.s64 %rd3392, %rd3388, %rd3391; mul.wide.u32 %rd3393, %r4449, 12; add.s64 %rd3394, %rd3388, %rd3393; mov.f32 %f22914, 0f3F800000; sub.ftz.f32 %f22915, %f22914, %f5236; sub.ftz.f32 %f22916, %f22915, %f5237; ld.global.f32 %f22917, [%rd3390]; ld.global.f32 %f22918, [%rd3390+4]; ld.global.f32 %f22919, [%rd3390+8]; ld.global.f32 %f22920, [%rd3392]; mul.ftz.f32 %f22921, %f5236, %f22920; ld.global.f32 %f22922, [%rd3392+4]; mul.ftz.f32 %f22923, %f5236, %f22922; ld.global.f32 %f22924, [%rd3392+8]; mul.ftz.f32 %f22925, %f5236, %f22924; fma.rn.ftz.f32 %f22926, %f22916, %f22917, %f22921; fma.rn.ftz.f32 %f22927, %f22916, %f22918, %f22923; fma.rn.ftz.f32 %f22928, %f22916, %f22919, %f22925; ld.global.f32 %f22929, [%rd3394]; ld.global.f32 %f22930, [%rd3394+4]; ld.global.f32 %f22931, [%rd3394+8]; fma.rn.ftz.f32 %f30320, %f5237, %f22929, %f22926; fma.rn.ftz.f32 %f30321, %f5237, %f22930, %f22927; fma.rn.ftz.f32 %f30322, %f5237, %f22931, %f22928; mul.ftz.f32 %f22932, %f30321, %f30321; fma.rn.ftz.f32 %f22933, %f30320, %f30320, %f22932; fma.rn.ftz.f32 %f22934, %f30322, %f30322, %f22933; sqrt.approx.ftz.f32 %f5306, %f22934; setp.neu.ftz.f32 %p2958, %f5306, 0f00000000; @%p2958 bra $L__BB107_3418; bra.uni $L__BB107_3419; $L__BB107_3418: rcp.approx.ftz.f32 %f22935, %f5306; mul.ftz.f32 %f30322, %f30322, %f22935; mul.ftz.f32 %f30321, %f30321, %f22935; mul.ftz.f32 %f30320, %f30320, %f22935; $L__BB107_3419: add.s64 %rd3395, %rd605, 8; setp.eq.ftz.f32 %p2959, %f30320, 0f00000000; setp.eq.ftz.f32 %p2960, %f30321, 0f00000000; and.pred %p2961, %p2959, %p2960; setp.eq.ftz.f32 %p2962, %f30322, 0f00000000; and.pred %p2963, %p2962, %p2961; selp.f32 %f30335, %f5235, %f30322, %p2963; selp.f32 %f30336, %f30340, %f30321, %p2963; selp.f32 %f30337, %f30339, %f30320, %p2963; cvta.to.local.u64 %rd614, %rd605; st.local.f32 [%rd614], %f30337; st.local.f32 [%rd608], %f30336; cvta.to.local.u64 %rd616, %rd3395; st.local.f32 [%rd616], %f30335; and.b32 %r5805, %r1240, 4; setp.ne.s32 %p2964, %r5805, 0; @%p2964 bra $L__BB107_3423; @%p2950 bra $L__BB107_3422; bra.uni $L__BB107_3421; $L__BB107_3422: ld.const.u64 %rd3404, [kernel_params+2160]; cvta.to.global.u64 %rd3405, %rd3404; mul.lo.s64 %rd3406, %rd606, 208; add.s64 %rd3407, %rd3405, %rd3406; ld.global.v4.f32 {%f22970, %f22971, %f22972, %f22973}, [%rd3407+48]; ld.global.v4.f32 {%f22977, %f22978, %f22979, %f22980}, [%rd3407+64]; ld.global.v4.f32 {%f22984, %f22985, %f22986, %f22987}, [%rd3407+80]; mul.ftz.f32 %f22991, %f22977, %f30336; fma.rn.ftz.f32 %f22992, %f22970, %f30337, %f22991; fma.rn.ftz.f32 %f22993, %f22984, %f30335, %f22992; mul.ftz.f32 %f22994, %f22978, %f30336; fma.rn.ftz.f32 %f22995, %f22971, %f30337, %f22994; fma.rn.ftz.f32 %f22996, %f22985, %f30335, %f22995; mul.ftz.f32 %f22997, %f22979, %f30336; fma.rn.ftz.f32 %f22998, %f22972, %f30337, %f22997; fma.rn.ftz.f32 %f22999, %f22986, %f30335, %f22998; mul.ftz.f32 %f23000, %f22996, %f22996; fma.rn.ftz.f32 %f23001, %f22993, %f22993, %f23000; fma.rn.ftz.f32 %f23002, %f22999, %f22999, %f23001; sqrt.approx.ftz.f32 %f23003, %f23002; div.approx.ftz.f32 %f30337, %f22993, %f23003; div.approx.ftz.f32 %f30336, %f22996, %f23003; div.approx.ftz.f32 %f30335, %f22999, %f23003; st.local.f32 [%rd614], %f30337; st.local.f32 [%rd608], %f30336; st.local.f32 [%rd616], %f30335; mov.u32 %r5805, %r4440; bra.uni $L__BB107_3423; $L__BB107_3728: setp.gt.s32 %p3169, %r1375, %r1374; selp.b32 %r4719, -1, 0, %p3169; add.s32 %r4720, %r1375, %r4719; mad.lo.s32 %r4721, %r4720, %r1373, %r5822; ld.const.u64 %rd3792, [kernel_params+2336]; cvta.to.global.u64 %rd3793, %rd3792; add.s32 %r4722, %r4715, %r4721; mul.wide.u32 %rd3794, %r4722, 12; add.s64 %rd4874, %rd3793, %rd3794; add.s32 %r4723, %r4716, %r4721; mul.wide.u32 %rd3795, %r4723, 12; add.s64 %rd4877, %rd3793, %rd3795; add.s32 %r4724, %r4717, %r4721; mul.wide.u32 %rd3796, %r4724, 12; add.s64 %rd4880, %rd3793, %rd3796; add.s64 %rd4875, %rd4874, 4; add.s64 %rd4876, %rd4874, 8; add.s64 %rd4878, %rd4877, 4; add.s64 %rd4879, %rd4877, 8; add.s64 %rd4881, %rd4880, 4; add.s64 %rd4882, %rd4880, 8; $L__BB107_3730: ld.global.f32 %f5618, [%rd4882]; ld.global.f32 %f5619, [%rd4881]; ld.global.f32 %f5620, [%rd4880]; ld.global.f32 %f5621, [%rd4879]; ld.global.f32 %f5622, [%rd4878]; ld.global.f32 %f5623, [%rd4877]; ld.global.f32 %f5624, [%rd4876]; ld.global.f32 %f5625, [%rd4875]; ld.global.f32 %f5626, [%rd4874]; add.s32 %r1390, %r1375, 1; setp.eq.s32 %p3170, %r1390, %r1374; @%p3170 bra $L__BB107_3732; bra.uni $L__BB107_3731; $L__BB107_3732: ld.const.u64 %rd3807, [kernel_params+2256]; cvta.to.global.u64 %rd3808, %rd3807; mul.wide.u32 %rd3809, %r4718, 12; add.s64 %rd4883, %rd3808, %rd3809; add.s32 %r4733, %r4718, 1; mul.wide.u32 %rd3810, %r4733, 12; add.s64 %rd4886, %rd3808, %rd3810; add.s32 %r4734, %r4718, 2; mul.wide.u32 %rd3811, %r4734, 12; add.s64 %rd4889, %rd3808, %rd3811; add.s64 %rd4884, %rd4883, 4; add.s64 %rd4885, %rd4883, 8; add.s64 %rd4887, %rd4886, 4; add.s64 %rd4888, %rd4886, 8; add.s64 %rd4890, %rd4889, 4; add.s64 %rd4891, %rd4889, 8; bra.uni $L__BB107_3733; $L__BB107_3731: setp.ge.s32 %p3171, %r1375, %r1374; selp.b32 %r4727, -1, 0, %p3171; add.s32 %r4728, %r1390, %r4727; mad.lo.s32 %r4729, %r4728, %r1373, %r5822; ld.const.u64 %rd3802, [kernel_params+2336]; cvta.to.global.u64 %rd3803, %rd3802; add.s32 %r4730, %r4715, %r4729; mul.wide.u32 %rd3804, %r4730, 12; add.s64 %rd4883, %rd3803, %rd3804; add.s32 %r4731, %r4716, %r4729; mul.wide.u32 %rd3805, %r4731, 12; add.s64 %rd4886, %rd3803, %rd3805; add.s32 %r4732, %r4717, %r4729; mul.wide.u32 %rd3806, %r4732, 12; add.s64 %rd4889, %rd3803, %rd3806; add.s64 %rd4884, %rd4883, 4; add.s64 %rd4885, %rd4883, 8; add.s64 %rd4887, %rd4886, 4; add.s64 %rd4888, %rd4886, 8; add.s64 %rd4890, %rd4889, 4; add.s64 %rd4891, %rd4889, 8; $L__BB107_3733: sub.ftz.f32 %f5627, %f23551, %f5617; ld.global.f32 %f23562, [%rd4883]; mul.ftz.f32 %f23563, %f5617, %f23562; ld.global.f32 %f23564, [%rd4884]; mul.ftz.f32 %f23565, %f5617, %f23564; ld.global.f32 %f23566, [%rd4885]; mul.ftz.f32 %f23567, %f5617, %f23566; fma.rn.ftz.f32 %f30411, %f5627, %f5626, %f23563; fma.rn.ftz.f32 %f30412, %f5627, %f5625, %f23565; fma.rn.ftz.f32 %f30413, %f5627, %f5624, %f23567; ld.global.f32 %f23568, [%rd4886]; mul.ftz.f32 %f23569, %f5617, %f23568; ld.global.f32 %f23570, [%rd4887]; mul.ftz.f32 %f23571, %f5617, %f23570; ld.global.f32 %f23572, [%rd4888]; mul.ftz.f32 %f23573, %f5617, %f23572; fma.rn.ftz.f32 %f30414, %f5627, %f5623, %f23569; fma.rn.ftz.f32 %f30415, %f5627, %f5622, %f23571; fma.rn.ftz.f32 %f30416, %f5627, %f5621, %f23573; ld.global.f32 %f23574, [%rd4889]; mul.ftz.f32 %f23575, %f5617, %f23574; ld.global.f32 %f23576, [%rd4890]; mul.ftz.f32 %f23577, %f5617, %f23576; ld.global.f32 %f23578, [%rd4891]; mul.ftz.f32 %f23579, %f5617, %f23578; fma.rn.ftz.f32 %f30417, %f5627, %f5620, %f23575; fma.rn.ftz.f32 %f30418, %f5627, %f5619, %f23577; fma.rn.ftz.f32 %f30419, %f5627, %f5618, %f23579; setp.eq.s64 %p3172, %rd4892, 12; @%p3172 bra $L__BB107_3739; mov.u32 %r5827, -1; $L__BB107_3735: setp.eq.s64 %p3173, %rd4892, 0; @%p3173 bra $L__BB107_3737; add.s32 %r5824, %r5824, 2; bra.uni $L__BB107_3738; $L__BB107_3737: setp.eq.s16 %p3174, %rs734, 0; mov.u32 %r5824, %r5823; @%p3174 bra $L__BB107_3740; $L__BB107_3738: mul.wide.u32 %rd3812, %r5824, 16; add.s64 %rd3813, %rd755, %rd3812; ld.global.u16 %rs734, [%rd3813+12]; ld.global.u32 %r5823, [%rd3813+8]; ld.global.u64 %rd4892, [%rd3813]; setp.ne.s64 %p3175, %rd4892, 12; @%p3175 bra $L__BB107_3735; $L__BB107_3739: setp.eq.s16 %p3176, %rs734, 0; selp.b32 %r5827, -1, %r5823, %p3176; $L__BB107_3740: @%p3168 bra $L__BB107_3742; bra.uni $L__BB107_3741; $L__BB107_3742: ld.const.u64 %rd3819, [kernel_params+2224]; cvta.to.global.u64 %rd3820, %rd3819; mul.wide.u32 %rd3821, %r4715, 12; add.s64 %rd4893, %rd3820, %rd3821; mul.wide.u32 %rd3822, %r4716, 12; add.s64 %rd4896, %rd3820, %rd3822; mul.wide.u32 %rd3823, %r4717, 12; add.s64 %rd4899, %rd3820, %rd3823; add.s64 %rd4894, %rd4893, 4; add.s64 %rd4895, %rd4893, 8; add.s64 %rd4897, %rd4896, 4; add.s64 %rd4898, %rd4896, 8; add.s64 %rd4900, %rd4899, 4; add.s64 %rd4901, %rd4899, 8; bra.uni $L__BB107_3743; $L__BB107_3741: setp.gt.s32 %p3178, %r1375, %r1374; selp.b32 %r4736, -1, 0, %p3178; add.s32 %r4737, %r1375, %r4736; mad.lo.s32 %r4738, %r4737, %r1373, %r5827; ld.const.u64 %rd3814, [kernel_params+2336]; cvta.to.global.u64 %rd3815, %rd3814; add.s32 %r4739, %r4738, %r4715; mul.wide.u32 %rd3816, %r4739, 12; add.s64 %rd4893, %rd3815, %rd3816; add.s32 %r4740, %r4738, %r4716; mul.wide.u32 %rd3817, %r4740, 12; add.s64 %rd4896, %rd3815, %rd3817; add.s32 %r4741, %r4738, %r4717; mul.wide.u32 %rd3818, %r4741, 12; add.s64 %rd4899, %rd3815, %rd3818; add.s64 %rd4894, %rd4893, 4; add.s64 %rd4895, %rd4893, 8; add.s64 %rd4897, %rd4896, 4; add.s64 %rd4898, %rd4896, 8; add.s64 %rd4900, %rd4899, 4; add.s64 %rd4901, %rd4899, 8; $L__BB107_3743: ld.global.f32 %f5637, [%rd4901]; ld.global.f32 %f5638, [%rd4900]; ld.global.f32 %f5639, [%rd4899]; ld.global.f32 %f5640, [%rd4898]; ld.global.f32 %f5641, [%rd4897]; ld.global.f32 %f5642, [%rd4896]; ld.global.f32 %f5643, [%rd4895]; ld.global.f32 %f5644, [%rd4894]; ld.global.f32 %f5645, [%rd4893]; @%p3170 bra $L__BB107_3745; bra.uni $L__BB107_3744; $L__BB107_3745: ld.const.u64 %rd3829, [kernel_params+2224]; cvta.to.global.u64 %rd3830, %rd3829; mul.wide.u32 %rd3831, %r4715, 12; add.s64 %rd4902, %rd3830, %rd3831; mul.wide.u32 %rd3832, %r4716, 12; add.s64 %rd4905, %rd3830, %rd3832; mul.wide.u32 %rd3833, %r4717, 12; add.s64 %rd4908, %rd3830, %rd3833; add.s64 %rd4903, %rd4902, 4; add.s64 %rd4904, %rd4902, 8; add.s64 %rd4906, %rd4905, 4; add.s64 %rd4907, %rd4905, 8; add.s64 %rd4909, %rd4908, 4; add.s64 %rd4910, %rd4908, 8; bra.uni $L__BB107_3746; $L__BB107_3744: setp.ge.s32 %p3180, %r1375, %r1374; selp.b32 %r4742, -1, 0, %p3180; add.s32 %r4743, %r1390, %r4742; mad.lo.s32 %r4744, %r4743, %r1373, %r5827; ld.const.u64 %rd3824, [kernel_params+2336]; cvta.to.global.u64 %rd3825, %rd3824; add.s32 %r4745, %r4744, %r4715; mul.wide.u32 %rd3826, %r4745, 12; add.s64 %rd4902, %rd3825, %rd3826; add.s32 %r4746, %r4744, %r4716; mul.wide.u32 %rd3827, %r4746, 12; add.s64 %rd4905, %rd3825, %rd3827; add.s32 %r4747, %r4744, %r4717; mul.wide.u32 %rd3828, %r4747, 12; add.s64 %rd4908, %rd3825, %rd3828; add.s64 %rd4903, %rd4902, 4; add.s64 %rd4904, %rd4902, 8; add.s64 %rd4906, %rd4905, 4; add.s64 %rd4907, %rd4905, 8; add.s64 %rd4909, %rd4908, 4; add.s64 %rd4910, %rd4908, 8; $L__BB107_3746: ld.global.f32 %f23580, [%rd4902]; mul.ftz.f32 %f23581, %f5617, %f23580; ld.global.f32 %f23582, [%rd4903]; mul.ftz.f32 %f23583, %f5617, %f23582; ld.global.f32 %f23584, [%rd4904]; mul.ftz.f32 %f23585, %f5617, %f23584; fma.rn.ftz.f32 %f30402, %f5627, %f5645, %f23581; fma.rn.ftz.f32 %f30403, %f5627, %f5644, %f23583; fma.rn.ftz.f32 %f30404, %f5627, %f5643, %f23585; ld.global.f32 %f23586, [%rd4905]; mul.ftz.f32 %f23587, %f5617, %f23586; ld.global.f32 %f23588, [%rd4906]; mul.ftz.f32 %f23589, %f5617, %f23588; ld.global.f32 %f23590, [%rd4907]; mul.ftz.f32 %f23591, %f5617, %f23590; fma.rn.ftz.f32 %f30405, %f5627, %f5642, %f23587; fma.rn.ftz.f32 %f30406, %f5627, %f5641, %f23589; fma.rn.ftz.f32 %f30407, %f5627, %f5640, %f23591; ld.global.f32 %f23592, [%rd4908]; mul.ftz.f32 %f23593, %f5617, %f23592; ld.global.f32 %f23594, [%rd4909]; mul.ftz.f32 %f23595, %f5617, %f23594; ld.global.f32 %f23596, [%rd4910]; mul.ftz.f32 %f23597, %f5617, %f23596; fma.rn.ftz.f32 %f30408, %f5627, %f5639, %f23593; fma.rn.ftz.f32 %f30409, %f5627, %f5638, %f23595; fma.rn.ftz.f32 %f30410, %f5627, %f5637, %f23597; $L__BB107_3747: add.s64 %rd3835, %rd2172, 68; cvta.to.local.u64 %rd3836, %rd3835; ld.local.f32 %f5673, [%rd3836]; sub.ftz.f32 %f23599, %f23551, %f5673; add.s64 %rd3837, %rd2172, 72; cvta.to.local.u64 %rd3838, %rd3837; ld.local.f32 %f23600, [%rd3838]; sub.ftz.f32 %f5674, %f23599, %f23600; mul.ftz.f32 %f23601, %f30411, %f5674; mul.ftz.f32 %f23602, %f30412, %f5674; mul.ftz.f32 %f23603, %f30413, %f5674; fma.rn.ftz.f32 %f23604, %f30414, %f5673, %f23601; fma.rn.ftz.f32 %f23605, %f30415, %f5673, %f23602; fma.rn.ftz.f32 %f23606, %f30416, %f5673, %f23603; fma.rn.ftz.f32 %f5675, %f30417, %f23600, %f23604; fma.rn.ftz.f32 %f5676, %f30418, %f23600, %f23605; fma.rn.ftz.f32 %f5677, %f30419, %f23600, %f23606; mul.ftz.f32 %f23607, %f30402, %f5674; mul.ftz.f32 %f23608, %f30403, %f5674; mul.ftz.f32 %f23609, %f30404, %f5674; fma.rn.ftz.f32 %f23610, %f30405, %f5673, %f23607; fma.rn.ftz.f32 %f23611, %f30406, %f5673, %f23608; fma.rn.ftz.f32 %f23612, %f30407, %f5673, %f23609; fma.rn.ftz.f32 %f30420, %f30408, %f23600, %f23610; fma.rn.ftz.f32 %f30421, %f30409, %f23600, %f23611; fma.rn.ftz.f32 %f30422, %f30410, %f23600, %f23612; add.s64 %rd3839, %rd2172, 56; cvta.to.local.u64 %rd3840, %rd3839; ld.local.u32 %r1399, [%rd3840]; and.b32 %r4748, %r1399, 4; setp.ne.s32 %p3181, %r4748, 0; @%p3181 bra $L__BB107_3751; and.b32 %r4749, %r1399, 2; setp.eq.s32 %p3182, %r4749, 0; @%p3182 bra $L__BB107_3750; bra.uni $L__BB107_3749; $L__BB107_3750: ld.global.v4.f32 {%f23640, %f23641, %f23642, %f23643}, [%rd752+-184]; ld.global.v4.f32 {%f23647, %f23648, %f23649, %f23650}, [%rd752+-168]; ld.global.v4.f32 {%f23654, %f23655, %f23656, %f23657}, [%rd752+-152]; mul.ftz.f32 %f23661, %f30421, %f23641; fma.rn.ftz.f32 %f23662, %f30420, %f23640, %f23661; fma.rn.ftz.f32 %f5684, %f30422, %f23642, %f23662; mul.ftz.f32 %f23663, %f30421, %f23648; fma.rn.ftz.f32 %f23664, %f30420, %f23647, %f23663; fma.rn.ftz.f32 %f5685, %f30422, %f23649, %f23664; mul.ftz.f32 %f23665, %f30421, %f23655; fma.rn.ftz.f32 %f23666, %f30420, %f23654, %f23665; fma.rn.ftz.f32 %f30422, %f30422, %f23656, %f23666; mov.f32 %f30420, %f5684; mov.f32 %f30421, %f5685; bra.uni $L__BB107_3751; $L__BB107_3749: add.s64 %rd3842, %rd2172, 144; cvta.to.local.u64 %rd3843, %rd3842; ld.local.v4.f32 {%f23613, %f23614, %f23615, %f23616}, [%rd3843]; mul.ftz.f32 %f23620, %f30421, %f23614; fma.rn.ftz.f32 %f23621, %f30420, %f23613, %f23620; fma.rn.ftz.f32 %f5681, %f30422, %f23615, %f23621; add.s64 %rd3844, %rd2172, 160; cvta.to.local.u64 %rd3845, %rd3844; ld.local.v4.f32 {%f23622, %f23623, %f23624, %f23625}, [%rd3845]; mul.ftz.f32 %f23629, %f30421, %f23623; fma.rn.ftz.f32 %f23630, %f30420, %f23622, %f23629; fma.rn.ftz.f32 %f5682, %f30422, %f23624, %f23630; add.s64 %rd3846, %rd2172, 176; cvta.to.local.u64 %rd3847, %rd3846; ld.local.v4.f32 {%f23631, %f23632, %f23633, %f23634}, [%rd3847]; mul.ftz.f32 %f23638, %f30421, %f23632; fma.rn.ftz.f32 %f23639, %f30420, %f23631, %f23638; fma.rn.ftz.f32 %f30422, %f30422, %f23633, %f23639; mov.f32 %f30420, %f5681; mov.f32 %f30421, %f5682; $L__BB107_3751: sub.ftz.f32 %f5690, %f30411, %f30417; sub.ftz.f32 %f23667, %f30408, %f30402; sub.ftz.f32 %f5691, %f30412, %f30418; sub.ftz.f32 %f23668, %f30409, %f30403; mul.ftz.f32 %f23669, %f23668, %f5691; fma.rn.ftz.f32 %f23670, %f23667, %f5690, %f23669; sub.ftz.f32 %f5692, %f30413, %f30419; sub.ftz.f32 %f23671, %f30410, %f30404; fma.rn.ftz.f32 %f23672, %f23671, %f5692, %f23670; sub.ftz.f32 %f5693, %f30414, %f30417; sub.ftz.f32 %f23673, %f30408, %f30405; sub.ftz.f32 %f5694, %f30415, %f30418; sub.ftz.f32 %f23674, %f30409, %f30406; mul.ftz.f32 %f23675, %f23674, %f5694; fma.rn.ftz.f32 %f23676, %f23673, %f5693, %f23675; sub.ftz.f32 %f5695, %f30416, %f30419; sub.ftz.f32 %f23677, %f30410, %f30407; fma.rn.ftz.f32 %f23678, %f23677, %f5695, %f23676; sub.ftz.f32 %f5696, %f30414, %f30411; sub.ftz.f32 %f23679, %f30405, %f30402; sub.ftz.f32 %f5697, %f30415, %f30412; sub.ftz.f32 %f23680, %f30406, %f30403; mul.ftz.f32 %f23681, %f23680, %f5697; fma.rn.ftz.f32 %f23682, %f23679, %f5696, %f23681; sub.ftz.f32 %f5698, %f30416, %f30413; sub.ftz.f32 %f23683, %f30407, %f30404; fma.rn.ftz.f32 %f23684, %f23683, %f5698, %f23682; mul.ftz.f32 %f23685, %f23672, %f5674; add.ftz.f32 %f23686, %f5674, 0fBF800000; add.ftz.f32 %f23687, %f23672, %f23678; add.ftz.f32 %f23688, %f23684, %f23687; mul.ftz.f32 %f23689, %f5674, %f23688; mul.ftz.f32 %f23690, %f5673, %f23689; fma.rn.ftz.f32 %f23691, %f23685, %f23686, %f23690; mul.ftz.f32 %f23692, %f5673, %f23678; add.ftz.f32 %f23693, %f5673, 0fBF800000; fma.rn.ftz.f32 %f23694, %f23693, %f23692, %f23691; mul.ftz.f32 %f23695, %f30400, %f30421; fma.rn.ftz.f32 %f23696, %f30399, %f30420, %f23695; fma.rn.ftz.f32 %f23697, %f30401, %f30422, %f23696; setp.gt.ftz.f32 %p3183, %f23697, 0f00000000; sub.ftz.f32 %f5699, %f30411, %f30414; sub.ftz.f32 %f5700, %f30412, %f30415; sub.ftz.f32 %f5701, %f30413, %f30416; sub.ftz.f32 %f5702, %f30417, %f30411; sub.ftz.f32 %f5703, %f30418, %f30412; sub.ftz.f32 %f5704, %f30419, %f30413; sub.ftz.f32 %f5705, %f30417, %f30414; sub.ftz.f32 %f5706, %f30418, %f30415; sub.ftz.f32 %f5707, %f30419, %f30416; mul.ftz.f32 %f5708, %f23694, 0f3F000000; @%p3183 bra $L__BB107_3753; bra.uni $L__BB107_3752; $L__BB107_3753: mul.ftz.f32 %f23750, %f30403, %f5697; fma.rn.ftz.f32 %f23751, %f30402, %f5696, %f23750; fma.rn.ftz.f32 %f23752, %f30404, %f5698, %f23751; mul.ftz.f32 %f23753, %f30403, %f5703; fma.rn.ftz.f32 %f23754, %f30402, %f5702, %f23753; fma.rn.ftz.f32 %f23755, %f30404, %f5704, %f23754; max.ftz.f32 %f23756, %f23752, %f23755; max.ftz.f32 %f23758, %f23756, %f23555; mul.ftz.f32 %f23759, %f30406, %f5700; fma.rn.ftz.f32 %f23760, %f30405, %f5699, %f23759; fma.rn.ftz.f32 %f23761, %f30407, %f5701, %f23760; mul.ftz.f32 %f23762, %f30406, %f5706; fma.rn.ftz.f32 %f23763, %f30405, %f5705, %f23762; fma.rn.ftz.f32 %f23764, %f30407, %f5707, %f23763; max.ftz.f32 %f23765, %f23761, %f23764; max.ftz.f32 %f23766, %f23765, %f23555; mul.ftz.f32 %f23767, %f30409, %f5691; fma.rn.ftz.f32 %f23768, %f30408, %f5690, %f23767; fma.rn.ftz.f32 %f23769, %f30410, %f5692, %f23768; mul.ftz.f32 %f23770, %f30409, %f5694; fma.rn.ftz.f32 %f23771, %f30408, %f5693, %f23770; fma.rn.ftz.f32 %f23772, %f30410, %f5695, %f23771; max.ftz.f32 %f23773, %f23769, %f23772; max.ftz.f32 %f23774, %f23773, %f23555; sub.ftz.f32 %f23775, %f30411, %f5675; sub.ftz.f32 %f23776, %f30412, %f5676; mul.ftz.f32 %f23777, %f30403, %f23776; fma.rn.ftz.f32 %f23778, %f30402, %f23775, %f23777; sub.ftz.f32 %f23779, %f30413, %f5677; fma.rn.ftz.f32 %f23780, %f30404, %f23779, %f23778; add.ftz.f32 %f23781, %f23780, %f23758; max.ftz.f32 %f23782, %f23781, %f23555; sub.ftz.f32 %f23783, %f30414, %f5675; sub.ftz.f32 %f23784, %f30415, %f5676; mul.ftz.f32 %f23785, %f30406, %f23784; fma.rn.ftz.f32 %f23786, %f30405, %f23783, %f23785; sub.ftz.f32 %f23787, %f30416, %f5677; fma.rn.ftz.f32 %f23788, %f30407, %f23787, %f23786; add.ftz.f32 %f23789, %f23788, %f23766; max.ftz.f32 %f23790, %f23789, %f23555; sub.ftz.f32 %f23791, %f30417, %f5675; sub.ftz.f32 %f23792, %f30418, %f5676; mul.ftz.f32 %f23793, %f30409, %f23792; fma.rn.ftz.f32 %f23794, %f30408, %f23791, %f23793; sub.ftz.f32 %f23795, %f30419, %f5677; fma.rn.ftz.f32 %f23796, %f30410, %f23795, %f23794; add.ftz.f32 %f23797, %f23796, %f23774; max.ftz.f32 %f23798, %f23797, %f23555; min.ftz.f32 %f23799, %f23782, %f23790; min.ftz.f32 %f23800, %f23799, %f23798; max.ftz.f32 %f30423, %f23800, %f5708; bra.uni $L__BB107_3754; $L__BB107_3752: mul.ftz.f32 %f23698, %f30403, %f5700; fma.rn.ftz.f32 %f23699, %f30402, %f5699, %f23698; fma.rn.ftz.f32 %f23700, %f30404, %f5701, %f23699; mul.ftz.f32 %f23701, %f30403, %f5691; fma.rn.ftz.f32 %f23702, %f30402, %f5690, %f23701; fma.rn.ftz.f32 %f23703, %f30404, %f5692, %f23702; max.ftz.f32 %f23704, %f23700, %f23703; max.ftz.f32 %f23706, %f23704, %f23555; mul.ftz.f32 %f23707, %f30406, %f5697; fma.rn.ftz.f32 %f23708, %f30405, %f5696, %f23707; fma.rn.ftz.f32 %f23709, %f30407, %f5698, %f23708; mul.ftz.f32 %f23710, %f30406, %f5694; fma.rn.ftz.f32 %f23711, %f30405, %f5693, %f23710; fma.rn.ftz.f32 %f23712, %f30407, %f5695, %f23711; max.ftz.f32 %f23713, %f23709, %f23712; max.ftz.f32 %f23714, %f23713, %f23555; mul.ftz.f32 %f23715, %f30409, %f5703; fma.rn.ftz.f32 %f23716, %f30408, %f5702, %f23715; fma.rn.ftz.f32 %f23717, %f30410, %f5704, %f23716; mul.ftz.f32 %f23718, %f30409, %f5706; fma.rn.ftz.f32 %f23719, %f30408, %f5705, %f23718; fma.rn.ftz.f32 %f23720, %f30410, %f5707, %f23719; max.ftz.f32 %f23721, %f23717, %f23720; max.ftz.f32 %f23722, %f23721, %f23555; sub.ftz.f32 %f23723, %f5675, %f30411; sub.ftz.f32 %f23724, %f5676, %f30412; mul.ftz.f32 %f23725, %f30403, %f23724; fma.rn.ftz.f32 %f23726, %f30402, %f23723, %f23725; sub.ftz.f32 %f23727, %f5677, %f30413; fma.rn.ftz.f32 %f23728, %f30404, %f23727, %f23726; add.ftz.f32 %f23729, %f23728, %f23706; max.ftz.f32 %f23730, %f23729, %f23555; sub.ftz.f32 %f23731, %f5675, %f30414; sub.ftz.f32 %f23732, %f5676, %f30415; mul.ftz.f32 %f23733, %f30406, %f23732; fma.rn.ftz.f32 %f23734, %f30405, %f23731, %f23733; sub.ftz.f32 %f23735, %f5677, %f30416; fma.rn.ftz.f32 %f23736, %f30407, %f23735, %f23734; add.ftz.f32 %f23737, %f23736, %f23714; max.ftz.f32 %f23738, %f23737, %f23555; sub.ftz.f32 %f23739, %f5675, %f30417; sub.ftz.f32 %f23740, %f5676, %f30418; mul.ftz.f32 %f23741, %f30409, %f23740; fma.rn.ftz.f32 %f23742, %f30408, %f23739, %f23741; sub.ftz.f32 %f23743, %f5677, %f30419; fma.rn.ftz.f32 %f23744, %f30410, %f23743, %f23742; add.ftz.f32 %f23745, %f23744, %f23722; max.ftz.f32 %f23746, %f23745, %f23555; min.ftz.f32 %f23747, %f23730, %f23738; min.ftz.f32 %f23748, %f23747, %f23746; neg.ftz.f32 %f23749, %f23748; min.ftz.f32 %f30423, %f23749, %f5708; $L__BB107_3754: mul.ftz.f32 %f23801, %f30420, %f30423; mul.ftz.f32 %f23802, %f30421, %f30423; mul.ftz.f32 %f23803, %f30422, %f30423; fma.rn.ftz.f32 %f30424, %f5598, %f23801, %f30424; fma.rn.ftz.f32 %f30425, %f5598, %f23802, %f30425; fma.rn.ftz.f32 %f30426, %f5598, %f23803, %f30426; setp.gt.ftz.f32 %p3184, %f5597, 0f00000000; selp.u16 %rs736, 1, 0, %p3184; $L__BB107_3755: ld.local.u8 %rs560, [%rd280+43]; and.b16 %rs561, %rs560, 64; setp.eq.s16 %p3185, %rs561, 0; mov.f32 %f30392, 0f00000000; mov.f32 %f30393, %f30392; mov.f32 %f30394, %f30392; mov.f32 %f30395, %f30392; mov.f32 %f30396, %f30392; mov.f32 %f30397, %f30392; mov.f32 %f30398, %f30392; @%p3185 bra $L__BB107_3758; ld.local.f32 %f23812, [%rd280+12]; setp.eq.ftz.f32 %p3186, %f23812, 0f7F7FFFFF; mov.f32 %f30398, 0f7F7FFFFF; mov.f32 %f30392, %f30424; mov.f32 %f30393, %f30425; mov.f32 %f30394, %f30426; mov.f32 %f30395, %f30148; mov.f32 %f30396, %f30147; mov.f32 %f30397, %f30146; @%p3186 bra $L__BB107_3758; ld.local.f32 %f23813, [%rd280+-24]; sub.ftz.f32 %f23814, %f23813, %f30424; ld.local.f32 %f23815, [%rd280+-20]; sub.ftz.f32 %f23816, %f23815, %f30425; ld.local.f32 %f23817, [%rd280+-16]; sub.ftz.f32 %f23818, %f23817, %f30426; mul.ftz.f32 %f23819, %f23816, %f23816; fma.rn.ftz.f32 %f23820, %f23814, %f23814, %f23819; fma.rn.ftz.f32 %f23821, %f23818, %f23818, %f23820; sqrt.approx.ftz.f32 %f30398, %f23821; rsqrt.approx.ftz.f32 %f23822, %f23821; mul.ftz.f32 %f30395, %f23814, %f23822; mul.ftz.f32 %f30396, %f23816, %f23822; mul.ftz.f32 %f30397, %f23818, %f23822; mov.f32 %f30392, %f30424; mov.f32 %f30393, %f30425; mov.f32 %f30394, %f30426; $L__BB107_3758: add.s64 %rd3849, %rd2172, 92; cvta.to.local.u64 %rd3850, %rd3849; ld.local.f32 %f30443, [%rd3850]; add.s64 %rd3851, %rd2172, 84; cvta.to.local.u64 %rd3852, %rd3851; ld.local.f32 %f30442, [%rd3852]; add.s64 %rd3853, %rd2172, 76; cvta.to.local.u64 %rd3854, %rd3853; ld.local.u32 %r4750, [%rd3854]; setp.eq.s16 %p3187, %rs736, 0; selp.b32 %r5829, -1, %r4750, %p3187; add.s64 %rd3855, %rd2172, 60; cvta.to.local.u64 %rd3856, %rd3855; ld.local.u32 %r4751, [%rd3856]; selp.b32 %r5828, -1, %r4751, %p3187; ld.local.u32 %r5831, [%rd280+32]; ld.local.u32 %r5830, [%rd280+36]; $L__BB107_3759: ld.local.u32 %r1408, [%rd280+52]; ld.const.u64 %rd3857, [kernel_params+4376]; cvta.to.global.u64 %rd3858, %rd3857; atom.global.add.u32 %r4752, [%rd3858], 1; ld.const.u64 %rd3859, [kernel_params+4256]; cvta.to.global.u64 %rd3860, %rd3859; add.s64 %rd3861, %rd3860, 12; atom.global.add.u32 %r4753, [%rd3861], 1; ld.const.u64 %rd3862, [kernel_params+3416]; cvta.to.global.u64 %rd3863, %rd3862; cvt.s64.s32 %rd869, %r4752; mul.wide.s32 %rd3864, %r4752, 2; add.s64 %rd3865, %rd3863, %rd3864; mov.u16 %rs562, 3; st.global.u16 [%rd3865], %rs562; ld.const.u32 %r1409, [kernel_params]; and.b32 %r4754, %r1409, 131072; setp.eq.s32 %p3188, %r4754, 0; @%p3188 bra $L__BB107_3762; mov.u32 %r5832, 0; mov.u64 %rd3867, kernel_params; $L__BB107_3761: mul.wide.s32 %rd3866, %r5832, 16; add.s64 %rd3868, %rd3867, %rd3866; ld.const.v2.u64 {%rd3869, %rd3870}, [%rd3868+2768]; cvta.to.global.u64 %rd3873, %rd3870; add.s64 %rd3875, %rd3873, %rd1282; ld.global.u32 %r4756, [%rd3875]; cvta.to.global.u64 %rd3876, %rd3869; add.s64 %rd3877, %rd3876, %rd1282; ld.global.u32 %r4757, [%rd3877]; ld.const.v2.u64 {%rd3878, %rd3879}, [%rd3868+3744]; cvta.to.global.u64 %rd3882, %rd3878; shl.b64 %rd3883, %rd869, 2; add.s64 %rd3884, %rd3882, %rd3883; st.global.u32 [%rd3884], %r4757; cvta.to.global.u64 %rd3885, %rd3879; add.s64 %rd3886, %rd3885, %rd3883; st.global.u32 [%rd3886], %r4756; add.s32 %r5832, %r5832, 1; setp.ne.s32 %p3189, %r4756, -1; @%p3189 bra $L__BB107_3761; $L__BB107_3762: setp.geu.ftz.f32 %p3190, %f1938, 0f00000000; @%p3190 bra $L__BB107_3777; ld.local.u32 %r1412, [%rd5+-36]; and.b32 %r4758, %r1412, 262144; setp.eq.s32 %p3191, %r4758, 0; @%p3191 bra $L__BB107_3777; bra.uni $L__BB107_3764; $L__BB107_3777: setp.eq.s32 %p3208, %r5829, -1; @%p3208 bra $L__BB107_3786; ld.local.u32 %r4773, [%rd5+-24]; and.b32 %r4774, %r4773, 1; setp.eq.b32 %p3209, %r4774, 1; mov.pred %p3210, 0; xor.pred %p3211, %p3209, %p3210; not.pred %p3212, %p3211; @%p3212 bra $L__BB107_3786; ld.const.u64 %rd3950, [kernel_params+2232]; cvta.to.global.u64 %rd3951, %rd3950; ld.local.v2.u32 {%r4775, %r4776}, [%rd5+-32]; mul.wide.s32 %rd3952, %r4776, 16; add.s64 %rd3953, %rd3951, %rd3952; ld.const.u64 %rd3954, [kernel_params+2256]; cvta.to.global.u64 %rd3955, %rd3954; ld.global.u32 %r4778, [%rd3953+12]; mul.wide.u32 %rd3956, %r4778, 12; add.s64 %rd3957, %rd3955, %rd3956; ld.global.f32 %f5744, [%rd3957]; ld.global.f32 %f5745, [%rd3957+4]; ld.global.f32 %f5746, [%rd3957+8]; add.s32 %r4779, %r4778, 1; mul.wide.u32 %rd3958, %r4779, 12; add.s64 %rd3959, %rd3955, %rd3958; ld.global.f32 %f5747, [%rd3959]; ld.global.f32 %f5748, [%rd3959+4]; ld.global.f32 %f5749, [%rd3959+8]; add.s32 %r4780, %r4778, 2; mul.wide.u32 %rd3960, %r4780, 12; add.s64 %rd3961, %rd3955, %rd3960; ld.global.f32 %f5750, [%rd3961]; ld.global.f32 %f5751, [%rd3961+4]; ld.global.f32 %f5752, [%rd3961+8]; and.b32 %r4781, %r4775, 4; setp.ne.s32 %p3213, %r4781, 0; mov.f32 %f30456, %f30392; mov.f32 %f30457, %f30393; mov.f32 %f30458, %f30394; mov.f32 %f30459, %f30395; mov.f32 %f30460, %f30396; mov.f32 %f30461, %f30397; @%p3213 bra $L__BB107_3784; and.b32 %r4782, %r4775, 2; setp.eq.s32 %p3214, %r4782, 0; @%p3214 bra $L__BB107_3782; bra.uni $L__BB107_3781; $L__BB107_3782: ld.const.u64 %rd3962, [kernel_params+2160]; cvta.to.global.u64 %rd3963, %rd3962; ld.local.u32 %r4783, [%rd5+-12]; mul.wide.s32 %rd3964, %r4783, 208; add.s64 %rd3965, %rd3963, %rd3964; ld.global.v4.f32 {%f30444, %f30445, %f30446, %f30447}, [%rd3965+48]; ld.global.v4.f32 {%f30448, %f30449, %f30450, %f30451}, [%rd3965+64]; ld.global.v4.f32 {%f30452, %f30453, %f30454, %f30455}, [%rd3965+80]; bra.uni $L__BB107_3783; $L__BB107_3764: and.b32 %r4759, %r1412, 1; setp.eq.b32 %p3192, %r4759, 1; mov.pred %p3193, 0; xor.pred %p3194, %p3192, %p3193; not.pred %p3195, %p3194; ld.const.u64 %rd4911, [kernel_params+3752]; cvta.to.global.u64 %rd3887, %rd4911; shl.b64 %rd3888, %rd869, 2; add.s64 %rd3889, %rd3887, %rd3888; ld.global.u32 %r5833, [%rd3889]; @%p3195 bra $L__BB107_3771; bra.uni $L__BB107_3765; $L__BB107_3771: mov.u64 %rd3920, kernel_params; add.s64 %rd4917, %rd3920, 3744; setp.eq.s32 %p3202, %r5833, -1; mov.u32 %r5838, 0; @%p3202 bra $L__BB107_3775; ld.local.u32 %r1423, [%rd5+-12]; add.s64 %rd4917, %rd3920, 3744; ld.local.u32 %r1424, [%rd5+-40]; mov.u32 %r5838, 0; $L__BB107_3773: ld.const.u64 %rd3923, [%rd4917]; cvta.to.global.u64 %rd3924, %rd3923; add.s64 %rd3926, %rd3924, %rd3888; ld.global.u32 %r4766, [%rd3926]; setp.eq.s32 %p3203, %r4766, %r1423; setp.eq.s32 %p3204, %r5833, %r1424; and.pred %p3205, %p3203, %p3204; @%p3205 bra $L__BB107_3777; add.s32 %r5838, %r5838, 1; mul.wide.s32 %rd3927, %r5838, 16; add.s64 %rd3929, %rd3920, %rd3927; add.s64 %rd4917, %rd3929, 3744; ld.const.u64 %rd4911, [%rd3929+3752]; cvta.to.global.u64 %rd3930, %rd4911; add.s64 %rd3932, %rd3930, %rd3888; ld.global.u32 %r5833, [%rd3932]; setp.ne.s32 %p3206, %r5833, -1; @%p3206 bra $L__BB107_3773; $L__BB107_3775: ld.const.u32 %r4767, [kernel_params+12]; add.s32 %r4768, %r4767, -1; setp.ge.u32 %p3207, %r5838, %r4768; @%p3207 bra $L__BB107_3777; ld.local.u32 %r4769, [%rd5+-12]; ld.local.u32 %r4770, [%rd5+-40]; ld.const.u64 %rd3933, [%rd4917]; cvta.to.global.u64 %rd3934, %rd3933; add.s64 %rd3936, %rd3934, %rd3888; st.global.u32 [%rd3936], %r4769; cvta.to.global.u64 %rd3937, %rd4911; add.s64 %rd3938, %rd3937, %rd3888; st.global.u32 [%rd3938], %r4770; add.s32 %r4771, %r5838, 1; mul.wide.s32 %rd3939, %r4771, 16; add.s64 %rd3941, %rd3920, %rd3939; ld.const.v2.u64 {%rd3942, %rd3943}, [%rd3941+3744]; cvta.to.global.u64 %rd3946, %rd3942; add.s64 %rd3947, %rd3946, %rd3888; mov.u32 %r4772, -1; st.global.u32 [%rd3947], %r4772; cvta.to.global.u64 %rd3948, %rd3943; add.s64 %rd3949, %rd3948, %rd3888; st.global.u32 [%rd3949], %r4772; bra.uni $L__BB107_3777; $L__BB107_3765: setp.eq.s32 %p3196, %r5833, -1; @%p3196 bra $L__BB107_3777; ld.local.u32 %r1414, [%rd5+-12]; mov.u64 %rd3891, kernel_params; add.s64 %rd4912, %rd3891, 3744; ld.local.u32 %r1415, [%rd5+-40]; mov.u32 %r5834, 0; $L__BB107_3767: ld.const.u64 %rd3892, [%rd4912]; cvta.to.global.u64 %rd3893, %rd3892; add.s64 %rd3895, %rd3893, %rd3888; ld.global.u32 %r4761, [%rd3895]; setp.eq.s32 %p3197, %r4761, %r1414; setp.eq.s32 %p3198, %r5833, %r1415; and.pred %p3199, %p3197, %p3198; add.s32 %r1418, %r5834, 1; mul.wide.s32 %rd3896, %r1418, 16; add.s64 %rd3898, %rd3891, %rd3896; add.s64 %rd873, %rd3898, 3752; @%p3199 bra $L__BB107_3769; add.s64 %rd4912, %rd3898, 3744; ld.const.u64 %rd4911, [%rd873]; cvta.to.global.u64 %rd3902, %rd4911; add.s64 %rd3904, %rd3902, %rd3888; ld.global.u32 %r5833, [%rd3904]; setp.eq.s32 %p3200, %r5833, -1; mov.u32 %r5834, %r1418; @%p3200 bra $L__BB107_3777; bra.uni $L__BB107_3767; $L__BB107_3781: ld.local.v4.f32 {%f30444, %f30445, %f30446, %f30447}, [%rd5+104]; ld.local.v4.f32 {%f30448, %f30449, %f30450, %f30451}, [%rd5+120]; ld.local.v4.f32 {%f30452, %f30453, %f30454, %f30455}, [%rd5+136]; $L__BB107_3783: mul.ftz.f32 %f23847, %f30393, %f30445; fma.rn.ftz.f32 %f23848, %f30392, %f30444, %f23847; fma.rn.ftz.f32 %f23849, %f30394, %f30446, %f23848; mul.ftz.f32 %f23850, %f30393, %f30449; fma.rn.ftz.f32 %f23851, %f30392, %f30448, %f23850; fma.rn.ftz.f32 %f23852, %f30394, %f30450, %f23851; mul.ftz.f32 %f23853, %f30393, %f30453; fma.rn.ftz.f32 %f23854, %f30392, %f30452, %f23853; fma.rn.ftz.f32 %f23855, %f30394, %f30454, %f23854; add.ftz.f32 %f30456, %f30447, %f23849; add.ftz.f32 %f30457, %f30451, %f23852; add.ftz.f32 %f30458, %f30455, %f23855; mul.ftz.f32 %f23856, %f30396, %f30445; fma.rn.ftz.f32 %f23857, %f30395, %f30444, %f23856; mul.ftz.f32 %f23858, %f30396, %f30449; fma.rn.ftz.f32 %f23859, %f30395, %f30448, %f23858; mul.ftz.f32 %f23860, %f30396, %f30453; fma.rn.ftz.f32 %f23861, %f30395, %f30452, %f23860; fma.rn.ftz.f32 %f30459, %f30397, %f30446, %f23857; fma.rn.ftz.f32 %f30460, %f30397, %f30450, %f23859; fma.rn.ftz.f32 %f30461, %f30397, %f30454, %f23861; $L__BB107_3784: sub.ftz.f32 %f23862, %f5744, %f30456; sub.ftz.f32 %f23863, %f5745, %f30457; sub.ftz.f32 %f23864, %f5746, %f30458; sub.ftz.f32 %f23865, %f5747, %f30456; sub.ftz.f32 %f23866, %f5748, %f30457; sub.ftz.f32 %f23867, %f5749, %f30458; sub.ftz.f32 %f23868, %f5750, %f30456; sub.ftz.f32 %f23869, %f5751, %f30457; sub.ftz.f32 %f23870, %f5752, %f30458; sub.ftz.f32 %f23871, %f23868, %f23862; sub.ftz.f32 %f23872, %f23869, %f23863; sub.ftz.f32 %f23873, %f23870, %f23864; sub.ftz.f32 %f23874, %f23862, %f23865; sub.ftz.f32 %f23875, %f23863, %f23866; sub.ftz.f32 %f23876, %f23864, %f23867; sub.ftz.f32 %f23877, %f23865, %f23868; sub.ftz.f32 %f23878, %f23866, %f23869; sub.ftz.f32 %f23879, %f23867, %f23870; add.ftz.f32 %f23880, %f23868, %f23862; add.ftz.f32 %f23881, %f23869, %f23863; add.ftz.f32 %f23882, %f23870, %f23864; mul.ftz.f32 %f23883, %f23881, %f23873; mul.ftz.f32 %f23884, %f23882, %f23872; sub.ftz.f32 %f23885, %f23883, %f23884; mul.ftz.f32 %f23886, %f23882, %f23871; mul.ftz.f32 %f23887, %f23880, %f23873; sub.ftz.f32 %f23888, %f23886, %f23887; mul.ftz.f32 %f23889, %f23880, %f23872; mul.ftz.f32 %f23890, %f23881, %f23871; sub.ftz.f32 %f23891, %f23889, %f23890; mul.ftz.f32 %f23892, %f30460, %f23888; fma.rn.ftz.f32 %f23893, %f30459, %f23885, %f23892; fma.rn.ftz.f32 %f23894, %f30461, %f23891, %f23893; add.ftz.f32 %f23895, %f23862, %f23865; add.ftz.f32 %f23896, %f23863, %f23866; add.ftz.f32 %f23897, %f23864, %f23867; mul.ftz.f32 %f23898, %f23896, %f23876; mul.ftz.f32 %f23899, %f23897, %f23875; sub.ftz.f32 %f23900, %f23898, %f23899; mul.ftz.f32 %f23901, %f23897, %f23874; mul.ftz.f32 %f23902, %f23895, %f23876; sub.ftz.f32 %f23903, %f23901, %f23902; mul.ftz.f32 %f23904, %f23895, %f23875; mul.ftz.f32 %f23905, %f23896, %f23874; sub.ftz.f32 %f23906, %f23904, %f23905; mul.ftz.f32 %f23907, %f30460, %f23903; fma.rn.ftz.f32 %f23908, %f30459, %f23900, %f23907; fma.rn.ftz.f32 %f23909, %f30461, %f23906, %f23908; add.ftz.f32 %f23910, %f23865, %f23868; add.ftz.f32 %f23911, %f23866, %f23869; add.ftz.f32 %f23912, %f23867, %f23870; mul.ftz.f32 %f23913, %f23911, %f23879; mul.ftz.f32 %f23914, %f23912, %f23878; sub.ftz.f32 %f23915, %f23913, %f23914; mul.ftz.f32 %f23916, %f23912, %f23877; mul.ftz.f32 %f23917, %f23910, %f23879; sub.ftz.f32 %f23918, %f23916, %f23917; mul.ftz.f32 %f23919, %f23910, %f23878; mul.ftz.f32 %f23920, %f23911, %f23877; sub.ftz.f32 %f23921, %f23919, %f23920; mul.ftz.f32 %f23922, %f30460, %f23918; fma.rn.ftz.f32 %f23923, %f30459, %f23915, %f23922; fma.rn.ftz.f32 %f23924, %f30461, %f23921, %f23923; add.ftz.f32 %f23925, %f23894, %f23909; add.ftz.f32 %f23926, %f23924, %f23925; abs.ftz.f32 %f23927, %f23926; mul.ftz.f32 %f23928, %f23927, 0f34000000; min.ftz.f32 %f23929, %f23909, %f23924; min.ftz.f32 %f23930, %f23894, %f23929; max.ftz.f32 %f23931, %f23909, %f23924; max.ftz.f32 %f23932, %f23894, %f23931; setp.ge.ftz.f32 %p3215, %f23930, %f23928; mul.ftz.f32 %f23933, %f23927, 0fB4000000; setp.le.ftz.f32 %p3216, %f23932, %f23933; or.pred %p3217, %p3215, %p3216; @%p3217 bra $L__BB107_3786; ld.local.v2.f32 {%f23934, %f23935}, [%rd5+-64]; mul.ftz.f32 %f23938, %f23934, 0f43800000; cvt.rzi.ftz.s32.f32 %r4784, %f23938; mul.ftz.f32 %f23939, %f23935, 0f43800000; cvt.rzi.ftz.s32.f32 %r4785, %f23939; ld.local.f32 %f23940, [%rd5+-56]; mul.ftz.f32 %f23941, %f23940, 0f43800000; cvt.rzi.ftz.s32.f32 %r4786, %f23941; neg.s32 %r4787, %r4784; setp.lt.ftz.f32 %p3218, %f30392, 0f00000000; selp.b32 %r4788, %r4787, %r4784, %p3218; mov.b32 %r4789, %f30392; add.s32 %r4790, %r4788, %r4789; mov.b32 %f23942, %r4790; neg.s32 %r4791, %r4785; setp.lt.ftz.f32 %p3219, %f30393, 0f00000000; selp.b32 %r4792, %r4791, %r4785, %p3219; mov.b32 %r4793, %f30393; add.s32 %r4794, %r4792, %r4793; mov.b32 %f23943, %r4794; neg.s32 %r4795, %r4786; setp.lt.ftz.f32 %p3220, %f30394, 0f00000000; selp.b32 %r4796, %r4795, %r4786, %p3220; mov.b32 %r4797, %f30394; add.s32 %r4798, %r4796, %r4797; mov.b32 %f23944, %r4798; abs.ftz.f32 %f23945, %f30392; setp.lt.ftz.f32 %p3221, %f23945, 0f3D000000; fma.rn.ftz.f32 %f23946, %f23934, 0f37800000, %f30392; abs.ftz.f32 %f23947, %f30393; setp.lt.ftz.f32 %p3222, %f23947, 0f3D000000; fma.rn.ftz.f32 %f23948, %f23935, 0f37800000, %f30393; abs.ftz.f32 %f23949, %f30394; setp.lt.ftz.f32 %p3223, %f23949, 0f3D000000; fma.rn.ftz.f32 %f23950, %f23940, 0f37800000, %f30394; selp.f32 %f30394, %f23950, %f23944, %p3223; selp.f32 %f30393, %f23948, %f23943, %p3222; selp.f32 %f30392, %f23946, %f23942, %p3221; $L__BB107_3786: ld.const.u64 %rd3966, [kernel_params+3496]; cvta.to.global.u64 %rd3967, %rd3966; mul.lo.s64 %rd3968, %rd869, 12; add.s64 %rd3969, %rd3967, %rd3968; st.global.f32 [%rd3969], %f30392; st.global.f32 [%rd3969+4], %f30393; st.global.f32 [%rd3969+8], %f30394; ld.const.v2.u64 {%rd3970, %rd3971}, [kernel_params+3504]; mov.u32 %r4799, 0; cvta.to.global.u64 %rd3974, %rd3970; add.s64 %rd3975, %rd3974, %rd3968; st.global.f32 [%rd3975], %f30395; st.global.f32 [%rd3975+4], %f30396; st.global.f32 [%rd3975+8], %f30397; cvta.to.global.u64 %rd3976, %rd3971; shl.b64 %rd3977, %rd869, 2; add.s64 %rd3978, %rd3976, %rd3977; st.global.u32 [%rd3978], %r4799; ld.const.v2.u64 {%rd3979, %rd3980}, [kernel_params+3520]; cvta.to.global.u64 %rd3983, %rd3979; add.s64 %rd3984, %rd3983, %rd3977; st.global.f32 [%rd3984], %f30398; cvta.to.global.u64 %rd3985, %rd3980; add.s64 %rd3986, %rd3985, %rd3977; st.global.f32 [%rd3986], %f30442; ld.const.u64 %rd3987, [kernel_params+3536]; cvta.to.global.u64 %rd3988, %rd3987; add.s64 %rd3989, %rd3988, %rd3977; st.global.f32 [%rd3989], %f30443; ld.const.u64 %rd3990, [kernel_params+3584]; cvta.to.global.u64 %rd3991, %rd3990; add.s64 %rd3992, %rd3991, %rd3977; st.global.u32 [%rd3992], %r5829; ld.const.u64 %rd3993, [kernel_params+3576]; cvta.to.global.u64 %rd3994, %rd3993; add.s64 %rd3995, %rd3994, %rd3977; st.global.u32 [%rd3995], %r5828; ld.const.u64 %rd3996, [kernel_params+3632]; cvta.to.global.u64 %rd3997, %rd3996; add.s64 %rd3998, %rd3997, %rd3977; st.global.u32 [%rd3998], %r5831; ld.const.u64 %rd3999, [kernel_params+3624]; cvta.to.global.u64 %rd4000, %rd3999; add.s64 %rd4001, %rd4000, %rd3977; st.global.u32 [%rd4001], %r5830; setp.ne.s32 %p3224, %r5830, -1; setp.eq.s32 %p3225, %r1408, 2; or.pred %p3226, %p3224, %p3225; selp.b32 %r4800, 0, 134217728, %p3226; ld.global.u32 %r1431, [%rd27]; or.b32 %r5840, %r1431, %r4800; ld.const.u64 %rd4006, [kernel_params+2584]; cvta.to.global.u64 %rd4007, %rd4006; mul.lo.s64 %rd4008, %rd3, 12; add.s64 %rd4009, %rd4007, %rd4008; ld.local.f32 %f5816, [%rd283+24]; ld.global.f32 %f23951, [%rd4009]; mul.ftz.f32 %f5817, %f23951, %f5816; ld.local.f32 %f5818, [%rd283+28]; ld.global.f32 %f23952, [%rd4009+4]; mul.ftz.f32 %f5819, %f23952, %f5818; ld.local.f32 %f5820, [%rd283+32]; ld.global.f32 %f23953, [%rd4009+8]; mul.ftz.f32 %f5821, %f23953, %f5820; and.b32 %r4801, %r1409, 2097152; setp.eq.s32 %p3227, %r4801, 0; @%p3227 bra $L__BB107_3803; and.b32 %r4802, %r1431, 100663296; setp.eq.s32 %p3228, %r4802, 0; @%p3228 bra $L__BB107_3789; ld.const.u64 %rd4010, [kernel_params+2600]; cvta.to.global.u64 %rd4011, %rd4010; add.s64 %rd4013, %rd4011, %rd4008; ld.global.f32 %f30476, [%rd4013+8]; ld.global.f32 %f30475, [%rd4013+4]; ld.global.f32 %f30474, [%rd4013]; ld.const.u64 %rd4014, [kernel_params+2608]; cvta.to.global.u64 %rd4015, %rd4014; add.s64 %rd4016, %rd4015, %rd4008; ld.global.f32 %f30479, [%rd4016+8]; ld.global.f32 %f30478, [%rd4016+4]; ld.global.f32 %f30477, [%rd4016]; bra.uni $L__BB107_3802; $L__BB107_3789: ld.local.f32 %f5828, [%rd283+4]; ld.local.f32 %f5829, [%rd283+8]; setp.eq.ftz.f32 %p3229, %f5816, 0f00000000; mov.f32 %f30479, 0f00000000; mov.f32 %f30474, %f30479; @%p3229 bra $L__BB107_3791; div.approx.ftz.f32 %f30474, %f30298, %f5816; $L__BB107_3791: setp.eq.ftz.f32 %p3230, %f5818, 0f00000000; mov.f32 %f30475, %f30479; @%p3230 bra $L__BB107_3793; div.approx.ftz.f32 %f30475, %f5828, %f5818; $L__BB107_3793: setp.eq.ftz.f32 %p3231, %f5820, 0f00000000; mov.f32 %f30476, %f30479; @%p3231 bra $L__BB107_3795; div.approx.ftz.f32 %f30476, %f5829, %f5820; $L__BB107_3795: ld.local.f32 %f5839, [%rd283+16]; ld.local.f32 %f5840, [%rd283+20]; mov.f32 %f30477, %f30479; @%p3229 bra $L__BB107_3797; ld.local.f32 %f23958, [%rd283+12]; div.approx.ftz.f32 %f30477, %f23958, %f5816; $L__BB107_3797: mov.f32 %f30478, %f30479; @%p3230 bra $L__BB107_3799; div.approx.ftz.f32 %f30478, %f5839, %f5818; $L__BB107_3799: @%p3231 bra $L__BB107_3801; div.approx.ftz.f32 %f30479, %f5840, %f5820; $L__BB107_3801: or.b32 %r5840, %r5840, 33554432; $L__BB107_3802: ld.const.u64 %rd4019, [kernel_params+3448]; cvta.to.global.u64 %rd4020, %rd4019; add.s64 %rd4022, %rd4020, %rd3968; st.global.f32 [%rd4022], %f30474; st.global.f32 [%rd4022+4], %f30475; st.global.f32 [%rd4022+8], %f30476; ld.const.u64 %rd4023, [kernel_params+3456]; cvta.to.global.u64 %rd4024, %rd4023; add.s64 %rd4025, %rd4024, %rd3968; st.global.f32 [%rd4025], %f30477; st.global.f32 [%rd4025+4], %f30478; st.global.f32 [%rd4025+8], %f30479; $L__BB107_3803: ld.const.v2.u64 {%rd4026, %rd4027}, [kernel_params+2448]; cvta.to.global.u64 %rd4030, %rd4026; add.s64 %rd4032, %rd4030, %rd1282; ld.global.u32 %r4803, [%rd4032]; ld.const.v2.u64 {%rd4033, %rd4034}, [kernel_params+3344]; cvta.to.global.u64 %rd4037, %rd4033; add.s64 %rd4039, %rd4037, %rd3977; st.global.u32 [%rd4039], %r4803; cvta.to.global.u64 %rd4040, %rd4920; add.s64 %rd4042, %rd4040, %rd2151; ld.global.u16 %rs563, [%rd4042]; ld.const.v2.u64 {%rd4043, %rd4044}, [kernel_params+3360]; cvta.to.global.u64 %rd4047, %rd4044; shl.b64 %rd4048, %rd869, 1; add.s64 %rd4049, %rd4047, %rd4048; st.global.u16 [%rd4049], %rs563; cvta.to.global.u64 %rd4050, %rd4921; add.s64 %rd4051, %rd4050, %rd1282; ld.global.u32 %r4804, [%rd4051]; cvta.to.global.u64 %rd4052, %rd4043; add.s64 %rd4053, %rd4052, %rd3977; st.global.u32 [%rd4053], %r4804; cvta.to.global.u64 %rd4054, %rd4027; add.s64 %rd4055, %rd4054, %rd1282; ld.global.u32 %r4805, [%rd4055]; cvta.to.global.u64 %rd4056, %rd4034; add.s64 %rd4057, %rd4056, %rd3977; st.global.u32 [%rd4057], %r4805; ld.const.u64 %rd4058, [kernel_params+3424]; cvta.to.global.u64 %rd4059, %rd4058; add.s64 %rd4060, %rd4059, %rd3977; st.global.u32 [%rd4060], %r5840; ld.const.u64 %rd4061, [kernel_params+2472]; cvta.to.global.u64 %rd4062, %rd4061; add.s64 %rd4063, %rd4062, %rd2151; ld.global.u16 %rs564, [%rd4063]; ld.const.u64 %rd4064, [kernel_params+3384]; cvta.to.global.u64 %rd4065, %rd4064; add.s64 %rd4066, %rd4065, %rd4048; st.global.u16 [%rd4066], %rs564; ld.const.u64 %rd4067, [kernel_params+2488]; cvta.to.global.u64 %rd4068, %rd4067; add.s64 %rd4069, %rd4068, %rd2151; ld.global.u16 %rs565, [%rd4069]; ld.const.u64 %rd4070, [kernel_params+3400]; cvta.to.global.u64 %rd4071, %rd4070; add.s64 %rd4072, %rd4071, %rd4048; st.global.u16 [%rd4072], %rs565; ld.const.u64 %rd4073, [kernel_params+2496]; cvta.to.global.u64 %rd4074, %rd4073; add.s64 %rd4075, %rd4074, %rd2151; ld.global.u16 %rs284, [%rd4075]; @%p2940 bra $L__BB107_3805; bra.uni $L__BB107_3804; $L__BB107_3805: cvt.u16.u32 %rs568, %r5802; add.s16 %rs569, %rs568, %rs284; add.s16 %rs570, %rs569, -1; ld.const.v2.u64 {%rd4099, %rd4100}, [kernel_params+3392]; ld.const.u64 %rd4102, [kernel_params+3408]; cvta.to.global.u64 %rd4103, %rd4102; add.s64 %rd4105, %rd4103, %rd4048; st.global.u16 [%rd4105], %rs570; ld.const.v2.u64 {%rd4106, %rd4107}, [kernel_params+2464]; ld.const.u64 %rd4109, [kernel_params+2480]; cvta.to.global.u64 %rd4110, %rd4109; add.s64 %rd4112, %rd4110, %rd2151; ld.global.u16 %rs571, [%rd4112]; add.s16 %rs572, %rs571, 1; ld.const.v2.u64 {%rd4113, %rd4114}, [kernel_params+3376]; cvta.to.global.u64 %rd4116, %rd4099; add.s64 %rd4117, %rd4116, %rd4048; st.global.u16 [%rd4117], %rs572; cvta.to.global.u64 %rd4118, %rd4106; add.s64 %rd4119, %rd4118, %rd2151; ld.global.u16 %rs573, [%rd4119]; add.s16 %rs574, %rs573, %rs568; cvta.to.global.u64 %rd4120, %rd4113; add.s64 %rd4121, %rd4120, %rd4048; st.global.u16 [%rd4121], %rs574; bra.uni $L__BB107_3806; $L__BB107_3804: ld.const.v2.u64 {%rd4076, %rd4077}, [kernel_params+3392]; ld.const.u64 %rd4079, [kernel_params+3408]; cvta.to.global.u64 %rd4080, %rd4079; add.s64 %rd4082, %rd4080, %rd4048; st.global.u16 [%rd4082], %rs284; ld.const.v2.u64 {%rd4083, %rd4084}, [kernel_params+2464]; ld.const.u64 %rd4086, [kernel_params+2480]; cvta.to.global.u64 %rd4087, %rd4086; add.s64 %rd4089, %rd4087, %rd2151; ld.global.u16 %rs566, [%rd4089]; ld.const.v2.u64 {%rd4090, %rd4091}, [kernel_params+3376]; cvta.to.global.u64 %rd4093, %rd4076; add.s64 %rd4094, %rd4093, %rd4048; st.global.u16 [%rd4094], %rs566; cvta.to.global.u64 %rd4095, %rd4083; add.s64 %rd4096, %rd4095, %rd2151; ld.global.u16 %rs567, [%rd4096]; cvta.to.global.u64 %rd4097, %rd4090; add.s64 %rd4098, %rd4097, %rd4048; st.global.u16 [%rd4098], %rs567; $L__BB107_3806: ld.const.u64 %rd4122, [kernel_params+3432]; cvta.to.global.u64 %rd4123, %rd4122; add.s64 %rd4125, %rd4123, %rd3968; st.global.f32 [%rd4125], %f5817; st.global.f32 [%rd4125+4], %f5819; st.global.f32 [%rd4125+8], %f5821; and.b32 %r4806, %r1409, 4194304; setp.eq.s32 %p3236, %r4806, 0; @%p3236 bra $L__BB107_3808; ld.const.u64 %rd4126, [kernel_params+3440]; cvta.to.global.u64 %rd4127, %rd4126; add.s64 %rd4129, %rd4127, %rd3968; st.global.f32 [%rd4129], %f5817; st.global.f32 [%rd4129+4], %f5819; st.global.f32 [%rd4129+8], %f5821; $L__BB107_3808: ld.const.u32 %r4807, [kernel_params+1076]; ld.local.u32 %r4808, [%rd280+48]; selp.b32 %r4809, %r4807, %r4808, %p3225; cvt.u16.u32 %rs575, %r4809; add.s16 %rs576, %rs575, 1; ld.const.u64 %rd4130, [kernel_params+3472]; cvta.to.global.u64 %rd4131, %rd4130; add.s64 %rd4132, %rd4131, %rd869; st.global.u8 [%rd4132], %rs576; bra.uni $L__BB107_3809; $L__BB107_3769: ld.const.u64 %rd4914, [%rd873+-24]; $L__BB107_3770: add.s32 %r5834, %r5834, 1; mul.wide.s32 %rd3905, %r5834, 16; add.s64 %rd3907, %rd3891, %rd3905; ld.const.v2.u64 {%rd3908, %rd3909}, [%rd3907+3744]; cvta.to.global.u64 %rd3910, %rd3908; add.s64 %rd3912, %rd3910, %rd3888; ld.global.u32 %r4762, [%rd3912]; cvta.to.global.u64 %rd3913, %rd3909; add.s64 %rd3914, %rd3913, %rd3888; ld.global.u32 %r4763, [%rd3914]; cvta.to.global.u64 %rd3915, %rd4914; add.s64 %rd3916, %rd3915, %rd3888; st.global.u32 [%rd3916], %r4762; cvta.to.global.u64 %rd3917, %rd4911; add.s64 %rd3918, %rd3917, %rd3888; st.global.u32 [%rd3918], %r4763; setp.eq.s32 %p3201, %r4763, -1; mov.u64 %rd4911, %rd3909; mov.u64 %rd4914, %rd3908; @%p3201 bra $L__BB107_3777; bra.uni $L__BB107_3770; $L__BB107_3421: add.s64 %rd3398, %rd2172, 192; cvta.to.local.u64 %rd3399, %rd3398; ld.local.v4.f32 {%f22936, %f22937, %f22938, %f22939}, [%rd3399]; add.s64 %rd3400, %rd2172, 208; cvta.to.local.u64 %rd3401, %rd3400; ld.local.v4.f32 {%f22943, %f22944, %f22945, %f22946}, [%rd3401]; add.s64 %rd3402, %rd2172, 224; cvta.to.local.u64 %rd3403, %rd3402; ld.local.v4.f32 {%f22950, %f22951, %f22952, %f22953}, [%rd3403]; mul.ftz.f32 %f22957, %f30336, %f22943; fma.rn.ftz.f32 %f22958, %f30337, %f22936, %f22957; fma.rn.ftz.f32 %f22959, %f30335, %f22950, %f22958; mul.ftz.f32 %f22960, %f30336, %f22944; fma.rn.ftz.f32 %f22961, %f30337, %f22937, %f22960; fma.rn.ftz.f32 %f22962, %f30335, %f22951, %f22961; mul.ftz.f32 %f22963, %f30336, %f22945; fma.rn.ftz.f32 %f22964, %f30337, %f22938, %f22963; fma.rn.ftz.f32 %f22965, %f30335, %f22952, %f22964; mul.ftz.f32 %f22966, %f22962, %f22962; fma.rn.ftz.f32 %f22967, %f22959, %f22959, %f22966; fma.rn.ftz.f32 %f22968, %f22965, %f22965, %f22967; sqrt.approx.ftz.f32 %f22969, %f22968; div.approx.ftz.f32 %f30337, %f22959, %f22969; div.approx.ftz.f32 %f30336, %f22962, %f22969; div.approx.ftz.f32 %f30335, %f22965, %f22969; st.local.f32 [%rd614], %f30337; st.local.f32 [%rd608], %f30336; st.local.f32 [%rd616], %f30335; mov.u32 %r5805, %r4440; $L__BB107_3423: ld.const.u64 %rd3408, [kernel_params+2232]; cvta.to.global.u64 %rd3409, %rd3408; shl.b64 %rd3410, %rd607, 4; add.s64 %rd3411, %rd3409, %rd3410; ld.const.u64 %rd3412, [kernel_params+2256]; cvta.to.global.u64 %rd3413, %rd3412; ld.global.u32 %r4456, [%rd3411+12]; mul.wide.u32 %rd3414, %r4456, 12; add.s64 %rd3415, %rd3413, %rd3414; add.s32 %r4457, %r4456, 1; mul.wide.u32 %rd3416, %r4457, 12; add.s64 %rd3417, %rd3413, %rd3416; add.s32 %r4458, %r4456, 2; mul.wide.u32 %rd3418, %r4458, 12; add.s64 %rd3419, %rd3413, %rd3418; ld.global.f32 %f23004, [%rd3419]; ld.global.f32 %f23005, [%rd3419+4]; ld.global.f32 %f23006, [%rd3419+8]; ld.global.f32 %f23007, [%rd3417]; ld.global.f32 %f23008, [%rd3415]; sub.ftz.f32 %f30334, %f23007, %f23008; ld.global.f32 %f23009, [%rd3417+4]; ld.global.f32 %f23010, [%rd3415+4]; sub.ftz.f32 %f30333, %f23009, %f23010; ld.global.f32 %f23011, [%rd3417+8]; ld.global.f32 %f23012, [%rd3415+8]; sub.ftz.f32 %f30332, %f23011, %f23012; add.s64 %rd3421, %rd2172, 116; cvta.to.local.u64 %rd617, %rd3421; st.local.f32 [%rd617], %f30334; add.s64 %rd3422, %rd2172, 120; cvta.to.local.u64 %rd618, %rd3422; st.local.f32 [%rd618], %f30333; add.s64 %rd3423, %rd2172, 124; cvta.to.local.u64 %rd619, %rd3423; st.local.f32 [%rd619], %f30332; sub.ftz.f32 %f30331, %f23004, %f23008; sub.ftz.f32 %f30330, %f23005, %f23010; sub.ftz.f32 %f30329, %f23006, %f23012; add.s64 %rd3424, %rd2172, 128; cvta.to.local.u64 %rd620, %rd3424; st.local.v2.f32 [%rd620], {%f30331, %f30330}; add.s64 %rd3425, %rd2172, 136; cvta.to.local.u64 %rd621, %rd3425; st.local.f32 [%rd621], %f30329; setp.ne.s32 %p2966, %r5805, 0; @%p2966 bra $L__BB107_3433; @%p2950 bra $L__BB107_3426; bra.uni $L__BB107_3425; $L__BB107_3426: ld.const.u64 %rd3433, [kernel_params+2160]; cvta.to.global.u64 %rd3434, %rd3433; mul.lo.s64 %rd3435, %rd606, 208; add.s64 %rd3436, %rd3434, %rd3435; ld.global.v4.f32 {%f23040, %f23041, %f23042, %f23043}, [%rd3436]; ld.global.v4.f32 {%f23047, %f23048, %f23049, %f23050}, [%rd3436+16]; ld.global.v4.f32 {%f23054, %f23055, %f23056, %f23057}, [%rd3436+32]; mul.ftz.f32 %f23061, %f23041, %f30333; fma.rn.ftz.f32 %f23062, %f30334, %f23040, %f23061; fma.rn.ftz.f32 %f30328, %f23042, %f30332, %f23062; mul.ftz.f32 %f23063, %f23048, %f30333; fma.rn.ftz.f32 %f23064, %f30334, %f23047, %f23063; fma.rn.ftz.f32 %f30327, %f23049, %f30332, %f23064; mul.ftz.f32 %f23065, %f23055, %f30333; fma.rn.ftz.f32 %f23066, %f30334, %f23054, %f23065; fma.rn.ftz.f32 %f30332, %f23056, %f30332, %f23066; bra.uni $L__BB107_3427; $L__BB107_3425: add.s64 %rd3427, %rd2172, 144; cvta.to.local.u64 %rd3428, %rd3427; ld.local.v4.f32 {%f23013, %f23014, %f23015, %f23016}, [%rd3428]; mul.ftz.f32 %f23020, %f30333, %f23014; fma.rn.ftz.f32 %f23021, %f30334, %f23013, %f23020; fma.rn.ftz.f32 %f30328, %f30332, %f23015, %f23021; add.s64 %rd3429, %rd2172, 160; cvta.to.local.u64 %rd3430, %rd3429; ld.local.v4.f32 {%f23022, %f23023, %f23024, %f23025}, [%rd3430]; mul.ftz.f32 %f23029, %f30333, %f23023; fma.rn.ftz.f32 %f23030, %f30334, %f23022, %f23029; fma.rn.ftz.f32 %f30327, %f30332, %f23024, %f23030; add.s64 %rd3431, %rd2172, 176; cvta.to.local.u64 %rd3432, %rd3431; ld.local.v4.f32 {%f23031, %f23032, %f23033, %f23034}, [%rd3432]; mul.ftz.f32 %f23038, %f30333, %f23032; fma.rn.ftz.f32 %f23039, %f30334, %f23031, %f23038; fma.rn.ftz.f32 %f30332, %f30332, %f23033, %f23039; $L__BB107_3427: st.local.f32 [%rd617], %f30328; st.local.f32 [%rd618], %f30327; st.local.f32 [%rd619], %f30332; @%p2950 bra $L__BB107_3429; bra.uni $L__BB107_3428; $L__BB107_3429: ld.const.u64 %rd3444, [kernel_params+2160]; cvta.to.global.u64 %rd3445, %rd3444; mul.lo.s64 %rd3446, %rd606, 208; add.s64 %rd3447, %rd3445, %rd3446; ld.global.v4.f32 {%f23094, %f23095, %f23096, %f23097}, [%rd3447]; ld.global.v4.f32 {%f23101, %f23102, %f23103, %f23104}, [%rd3447+16]; ld.global.v4.f32 {%f23108, %f23109, %f23110, %f23111}, [%rd3447+32]; mul.ftz.f32 %f23115, %f23095, %f30330; fma.rn.ftz.f32 %f23116, %f30331, %f23094, %f23115; fma.rn.ftz.f32 %f5346, %f23096, %f30329, %f23116; mul.ftz.f32 %f23117, %f23102, %f30330; fma.rn.ftz.f32 %f23118, %f30331, %f23101, %f23117; fma.rn.ftz.f32 %f5347, %f23103, %f30329, %f23118; mul.ftz.f32 %f23119, %f23109, %f30330; fma.rn.ftz.f32 %f23120, %f30331, %f23108, %f23119; fma.rn.ftz.f32 %f30329, %f23110, %f30329, %f23120; st.local.v2.f32 [%rd620], {%f5346, %f5347}; st.local.f32 [%rd621], %f30329; mov.f32 %f30330, %f5347; mov.f32 %f30331, %f5346; mov.f32 %f30333, %f30327; mov.f32 %f30334, %f30328; bra.uni $L__BB107_3433; $L__BB107_3428: add.s64 %rd3438, %rd2172, 144; cvta.to.local.u64 %rd3439, %rd3438; ld.local.v4.f32 {%f23067, %f23068, %f23069, %f23070}, [%rd3439]; mul.ftz.f32 %f23074, %f30330, %f23068; fma.rn.ftz.f32 %f23075, %f30331, %f23067, %f23074; fma.rn.ftz.f32 %f5343, %f30329, %f23069, %f23075; add.s64 %rd3440, %rd2172, 160; cvta.to.local.u64 %rd3441, %rd3440; ld.local.v4.f32 {%f23076, %f23077, %f23078, %f23079}, [%rd3441]; mul.ftz.f32 %f23083, %f30330, %f23077; fma.rn.ftz.f32 %f23084, %f30331, %f23076, %f23083; fma.rn.ftz.f32 %f5344, %f30329, %f23078, %f23084; add.s64 %rd3442, %rd2172, 176; cvta.to.local.u64 %rd3443, %rd3442; ld.local.v4.f32 {%f23085, %f23086, %f23087, %f23088}, [%rd3443]; mul.ftz.f32 %f23092, %f30330, %f23086; fma.rn.ftz.f32 %f23093, %f30331, %f23085, %f23092; fma.rn.ftz.f32 %f30329, %f30329, %f23087, %f23093; st.local.v2.f32 [%rd620], {%f5343, %f5344}; st.local.f32 [%rd621], %f30329; mov.f32 %f30330, %f5344; mov.f32 %f30331, %f5343; mov.f32 %f30333, %f30327; mov.f32 %f30334, %f30328; $L__BB107_3433: mov.b32 %r4460, %f5235; mov.b64 %rd3460, {%r4460, %r4461}; cvt.u32.u64 %r4462, %rd3460; mov.b32 %f30338, %r4462; @%p2947 bra $L__BB107_3436; mul.ftz.f32 %f23127, %f30340, %f5231; fma.rn.ftz.f32 %f23128, %f30339, %f5230, %f23127; fma.rn.ftz.f32 %f23129, %f5235, %f5232, %f23128; setp.geu.ftz.f32 %p2971, %f23129, 0f00000000; @%p2971 bra $L__BB107_3436; or.b32 %r4463, %r1235, 1; st.local.u32 [%rd611], %r4463; neg.ftz.f32 %f30338, %f5235; st.local.f32 [%rd3325], %f30338; cvta.to.local.u64 %rd3463, %rd605; neg.ftz.f32 %f23130, %f30337; st.local.f32 [%rd3463], %f23130; neg.ftz.f32 %f23131, %f30335; neg.ftz.f32 %f23132, %f30336; neg.ftz.f32 %f30340, %f30340; neg.ftz.f32 %f30339, %f30339; st.local.v4.f32 [%rd608], {%f23132, %f23131, %f30339, %f30340}; add.s64 %rd3465, %rd2172, 116; add.s64 %rd3466, %rd2172, 120; add.s64 %rd3467, %rd2172, 124; cvta.to.local.u64 %rd3468, %rd3465; neg.ftz.f32 %f23133, %f30334; st.local.f32 [%rd3468], %f23133; cvta.to.local.u64 %rd3469, %rd3466; neg.ftz.f32 %f23134, %f30333; st.local.f32 [%rd3469], %f23134; cvta.to.local.u64 %rd3470, %rd3467; neg.ftz.f32 %f23135, %f30332; st.local.f32 [%rd3470], %f23135; add.s64 %rd3471, %rd2172, 136; cvta.to.local.u64 %rd3472, %rd3471; neg.ftz.f32 %f23136, %f30330; neg.ftz.f32 %f23137, %f30331; add.s64 %rd3473, %rd2172, 128; cvta.to.local.u64 %rd3474, %rd3473; st.local.v2.f32 [%rd3474], {%f23137, %f23136}; neg.ftz.f32 %f23138, %f30329; st.local.f32 [%rd3472], %f23138; $L__BB107_3436: add.s64 %rd3476, %rd2172, 92; cvta.to.local.u64 %rd3477, %rd3476; mov.u32 %r4464, 0; st.local.u32 [%rd3477], %r4464; add.s64 %rd3478, %rd2172, 96; cvta.to.local.u64 %rd3479, %rd3478; mov.f32 %f23139, 0f00000000; st.local.v4.f32 [%rd3479], {%f23139, %f23139, %f23139, %f23139}; add.s64 %rd3480, %rd2172, 112; cvta.to.local.u64 %rd3481, %rd3480; st.local.u32 [%rd3481], %r4464; st.local.f32 [%rd280+-12], %f30339; st.local.f32 [%rd280+-8], %f30340; st.local.f32 [%rd280+-4], %f30338; $L__BB107_3438: add.s64 %rd3514, %rd2172, 260; cvta.to.local.u64 %rd3515, %rd3514; mov.u32 %r4471, 0; st.local.u32 [%rd3515], %r4471; add.s64 %rd3516, %rd2172, 264; cvta.to.local.u64 %rd3517, %rd3516; st.local.u32 [%rd3517], %r4471; ld.const.u64 %rd3518, [kernel_params+2400]; cvta.to.global.u64 %rd622, %rd3518; add.s64 %rd3519, %rd2172, 268; cvta.to.local.u64 %rd626, %rd3519; add.s64 %rd3520, %rd2172, 272; cvta.to.local.u64 %rd627, %rd3520; add.s64 %rd3521, %rd2172, 276; cvta.to.local.u64 %rd628, %rd3521; bra.uni $L__BB107_3439; $L__BB107_3591: { // callseq 619, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 619 $L__BB107_3439: mov.u32 %r1247, %r5807; mul.wide.s32 %rd3523, %r1247, 16; add.s64 %rd630, %rd622, %rd3523; ld.global.v4.u32 {%r4473, %r4474, %r4475, %r4476}, [%rd630]; mul.wide.u32 %rd3524, %r4474, 4; add.s64 %rd631, %rd2157, %rd3524; mul.wide.u32 %rd3525, %r4475, 4; add.s64 %rd632, %rd2157, %rd3525; add.s32 %r5807, %r1247, 1; setp.gt.s32 %p2972, %r4473, 36; @%p2972 bra $L__BB107_3488; setp.gt.s32 %p3054, %r4473, 25; @%p3054 bra $L__BB107_3484; bra.uni $L__BB107_3441; $L__BB107_3484: add.s32 %r4478, %r4473, -27; setp.lt.u32 %p3055, %r4478, 9; @%p3055 bra $L__BB107_3439; setp.eq.s32 %p3056, %r4473, 26; @%p3056 bra $L__BB107_3630; setp.eq.s32 %p3057, %r4473, 36; @%p3057 bra $L__BB107_3487; bra.uni $L__BB107_3644; $L__BB107_3487: mov.u32 %r4519, 36; { // callseq 672, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4519; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 672 bra.uni $L__BB107_3439; $L__BB107_3488: setp.gt.s32 %p2973, %r4473, 68; @%p2973 bra $L__BB107_3543; bra.uni $L__BB107_3489; $L__BB107_3543: setp.gt.s32 %p2974, %r4473, 88; @%p2974 bra $L__BB107_3579; bra.uni $L__BB107_3544; $L__BB107_3579: setp.gt.s32 %p2975, %r4473, 93; @%p2975 bra $L__BB107_3584; bra.uni $L__BB107_3580; $L__BB107_3584: setp.gt.s32 %p2976, %r4473, 95; @%p2976 bra $L__BB107_3588; bra.uni $L__BB107_3585; $L__BB107_3588: setp.eq.s32 %p2977, %r4473, 98; @%p2977 bra $L__BB107_3643; setp.eq.s32 %p2978, %r4473, 97; @%p2978 bra $L__BB107_3593; setp.ne.s32 %p2979, %r4473, 96; @%p2979 bra $L__BB107_3644; bra.uni $L__BB107_3591; $L__BB107_3489: setp.gt.s32 %p3012, %r4473, 48; @%p3012 bra $L__BB107_3508; bra.uni $L__BB107_3490; $L__BB107_3508: setp.gt.s32 %p3013, %r4473, 58; @%p3013 bra $L__BB107_3526; setp.gt.s32 %p3027, %r4473, 53; @%p3027 bra $L__BB107_3518; setp.gt.s32 %p3034, %r4473, 50; @%p3034 bra $L__BB107_3514; bra.uni $L__BB107_3511; $L__BB107_3514: setp.eq.s32 %p3035, %r4473, 51; @%p3035 bra $L__BB107_3622; setp.eq.s32 %p3036, %r4473, 52; @%p3036 bra $L__BB107_3621; setp.eq.s32 %p3037, %r4473, 53; @%p3037 bra $L__BB107_3517; bra.uni $L__BB107_3644; $L__BB107_3517: { // callseq 657, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 657 bra.uni $L__BB107_3439; $L__BB107_3441: setp.gt.s32 %p3058, %r4473, 17; @%p3058 bra $L__BB107_3475; bra.uni $L__BB107_3442; $L__BB107_3475: setp.lt.s32 %p3059, %r4473, 21; @%p3059 bra $L__BB107_3439; setp.gt.s32 %p3060, %r4473, 22; @%p3060 bra $L__BB107_3480; bra.uni $L__BB107_3477; $L__BB107_3480: setp.eq.s32 %p3061, %r4473, 23; @%p3061 bra $L__BB107_3632; setp.eq.s32 %p3062, %r4473, 24; @%p3062 bra $L__BB107_3631; setp.eq.s32 %p3063, %r4473, 25; @%p3063 bra $L__BB107_3483; bra.uni $L__BB107_3644; $L__BB107_3483: add.s32 %r4521, %r1247, 1; { // callseq 674, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4521; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 674 bra.uni $L__BB107_3439; $L__BB107_3544: setp.gt.s32 %p2985, %r4473, 78; @%p2985 bra $L__BB107_3562; setp.gt.s32 %p2999, %r4473, 73; @%p2999 bra $L__BB107_3554; setp.gt.s32 %p3006, %r4473, 70; @%p3006 bra $L__BB107_3550; bra.uni $L__BB107_3547; $L__BB107_3550: setp.eq.s32 %p3007, %r4473, 71; @%p3007 bra $L__BB107_3610; setp.eq.s32 %p3008, %r4473, 72; @%p3008 bra $L__BB107_3609; setp.eq.s32 %p3009, %r4473, 73; @%p3009 bra $L__BB107_3553; bra.uni $L__BB107_3644; $L__BB107_3553: add.s32 %r4491, %r1247, 1; { // callseq 637, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4491; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 637 bra.uni $L__BB107_3439; $L__BB107_3490: setp.gt.s32 %p3040, %r4473, 41; @%p3040 bra $L__BB107_3496; setp.gt.s32 %p3050, %r4473, 39; @%p3050 bra $L__BB107_3439; setp.eq.s32 %p3051, %r4473, 37; @%p3051 bra $L__BB107_3629; setp.eq.s32 %p3052, %r4473, 38; @%p3052 bra $L__BB107_3628; setp.eq.s32 %p3053, %r4473, 39; @%p3053 bra $L__BB107_3495; bra.uni $L__BB107_3644; $L__BB107_3495: mov.u32 %r4517, 39; { // callseq 669, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4517, %r4474, %r4475, %r4476}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 669 bra.uni $L__BB107_3439; $L__BB107_3442: setp.gt.s32 %p3066, %r4473, 8; @%p3066 bra $L__BB107_3458; bra.uni $L__BB107_3443; $L__BB107_3458: setp.gt.s32 %p3067, %r4473, 12; @%p3067 bra $L__BB107_3467; bra.uni $L__BB107_3459; $L__BB107_3467: setp.gt.s32 %p3068, %r4473, 14; @%p3068 bra $L__BB107_3471; bra.uni $L__BB107_3468; $L__BB107_3471: setp.eq.s32 %p3069, %r4473, 15; @%p3069 bra $L__BB107_3635; setp.eq.s32 %p3070, %r4473, 16; @%p3070 bra $L__BB107_3634; setp.eq.s32 %p3071, %r4473, 17; @%p3071 bra $L__BB107_3474; bra.uni $L__BB107_3644; $L__BB107_3474: { // callseq 679, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; .param .b32 param4; st.param.b32 [param4+0], %r4476; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 679 bra.uni $L__BB107_3439; $L__BB107_3526: setp.gt.s32 %p3014, %r4473, 63; @%p3014 bra $L__BB107_3535; setp.gt.s32 %p3021, %r4473, 60; @%p3021 bra $L__BB107_3531; bra.uni $L__BB107_3528; $L__BB107_3531: setp.eq.s32 %p3022, %r4473, 61; @%p3022 bra $L__BB107_3616; setp.eq.s32 %p3023, %r4473, 62; @%p3023 bra $L__BB107_3615; setp.eq.s32 %p3024, %r4473, 63; @%p3024 bra $L__BB107_3534; bra.uni $L__BB107_3644; $L__BB107_3534: mov.u32 %r4501, 63; { // callseq 647, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4501; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 647 bra.uni $L__BB107_3439; $L__BB107_3562: setp.gt.s32 %p2986, %r4473, 83; @%p2986 bra $L__BB107_3571; setp.gt.s32 %p2993, %r4473, 80; @%p2993 bra $L__BB107_3567; bra.uni $L__BB107_3564; $L__BB107_3567: setp.eq.s32 %p2994, %r4473, 81; @%p2994 bra $L__BB107_3598; setp.eq.s32 %p2995, %r4473, 82; @%p2995 bra $L__BB107_3597; setp.eq.s32 %p2996, %r4473, 83; @%p2996 bra $L__BB107_3570; bra.uni $L__BB107_3644; $L__BB107_3570: { // callseq 629, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 629 bra.uni $L__BB107_3439; $L__BB107_3443: setp.gt.s32 %p3079, %r4473, 4; @%p3079 bra $L__BB107_3451; setp.gt.s32 %p3085, %r4473, 2; @%p3085 bra $L__BB107_3448; setp.eq.s32 %p3088, %r4473, 1; mov.u32 %r5807, %r4474; @%p3088 bra $L__BB107_3439; setp.eq.s32 %p3089, %r4473, 2; @%p3089 bra $L__BB107_3447; bra.uni $L__BB107_3644; $L__BB107_3447: mov.u32 %r4545, 2; add.s32 %r4546, %r1247, 1; { // callseq 688, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r4545; st.param.b32 [param0+4], %r4474; st.param.b32 [param0+8], %r4475; st.param.b32 [param0+12], %r4476; .param .b32 param1; st.param.b32 [param1+0], %r4546; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 688 bra.uni $L__BB107_3439; $L__BB107_3496: setp.gt.s32 %p3041, %r4473, 44; @%p3041 bra $L__BB107_3501; bra.uni $L__BB107_3497; $L__BB107_3501: setp.gt.s32 %p3042, %r4473, 46; @%p3042 bra $L__BB107_3505; setp.eq.s32 %p3045, %r4473, 45; @%p3045 bra $L__BB107_3625; setp.eq.s32 %p3046, %r4473, 46; @%p3046 bra $L__BB107_3504; bra.uni $L__BB107_3644; $L__BB107_3504: { // callseq 664, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 664 bra.uni $L__BB107_3439; $L__BB107_3585: setp.eq.s32 %p2980, %r4473, 94; @%p2980 bra $L__BB107_3592; setp.eq.s32 %p2981, %r4473, 95; @%p2981 bra $L__BB107_3587; bra.uni $L__BB107_3644; $L__BB107_3587: { // callseq 620, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 620 bra.uni $L__BB107_3439; $L__BB107_3580: add.s32 %r4477, %r4473, -92; setp.lt.u32 %p2982, %r4477, 2; @%p2982 bra $L__BB107_3439; setp.eq.s32 %p2983, %r4473, 90; @%p2983 bra $L__BB107_3439; setp.eq.s32 %p2984, %r4473, 89; @%p2984 bra $L__BB107_3583; bra.uni $L__BB107_3644; $L__BB107_3583: mov.u32 %r4481, 89; { // callseq 623, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4481; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 623 bra.uni $L__BB107_3439; $L__BB107_3518: setp.gt.s32 %p3028, %r4473, 55; @%p3028 bra $L__BB107_3522; bra.uni $L__BB107_3519; $L__BB107_3522: setp.eq.s32 %p3029, %r4473, 56; @%p3029 bra $L__BB107_3619; setp.eq.s32 %p3030, %r4473, 57; @%p3030 bra $L__BB107_3618; setp.eq.s32 %p3031, %r4473, 58; @%p3031 bra $L__BB107_3525; bra.uni $L__BB107_3644; $L__BB107_3525: mov.u32 %r4508, 58; { // callseq 652, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4508; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 652 bra.uni $L__BB107_3439; $L__BB107_3554: setp.gt.s32 %p3000, %r4473, 75; @%p3000 bra $L__BB107_3558; bra.uni $L__BB107_3555; $L__BB107_3558: setp.eq.s32 %p3001, %r4473, 76; @%p3001 bra $L__BB107_3602; setp.eq.s32 %p3002, %r4473, 77; @%p3002 bra $L__BB107_3600; setp.eq.s32 %p3003, %r4473, 78; @%p3003 bra $L__BB107_3561; bra.uni $L__BB107_3644; $L__BB107_3561: add.s32 %r4489, %r1247, 1; { // callseq 634, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4489; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 634 bra.uni $L__BB107_3439; $L__BB107_3535: setp.gt.s32 %p3015, %r4473, 65; @%p3015 bra $L__BB107_3539; bra.uni $L__BB107_3536; $L__BB107_3539: setp.eq.s32 %p3016, %r4473, 66; @%p3016 bra $L__BB107_3613; setp.eq.s32 %p3017, %r4473, 67; @%p3017 bra $L__BB107_3612; setp.eq.s32 %p3018, %r4473, 68; @%p3018 bra $L__BB107_3542; bra.uni $L__BB107_3644; $L__BB107_3542: mov.u32 %r4496, 68; { // callseq 642, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4496; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 642 bra.uni $L__BB107_3439; $L__BB107_3571: setp.gt.s32 %p2987, %r4473, 85; @%p2987 bra $L__BB107_3575; bra.uni $L__BB107_3572; $L__BB107_3575: setp.eq.s32 %p2988, %r4473, 86; @%p2988 bra $L__BB107_3595; setp.eq.s32 %p2989, %r4473, 87; @%p2989 bra $L__BB107_3594; setp.eq.s32 %p2990, %r4473, 88; @%p2990 bra $L__BB107_3578; bra.uni $L__BB107_3644; $L__BB107_3578: mov.u32 %r4482, 88; ld.local.f32 %f23141, [%rd282]; ld.local.v2.f32 {%f23142, %f23143}, [%rd282+4]; { // callseq 624, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f23141; .param .b32 param1; st.param.f32 [param1+0], %f23142; .param .b32 param2; st.param.f32 [param2+0], %f23143; .param .b64 param3; st.param.b64 [param3+0], %rd2156; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r4482; st.param.b32 [param4+4], %r4474; st.param.b32 [param4+8], %r4475; st.param.b32 [param4+12], %r4476; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 624 bra.uni $L__BB107_3439; $L__BB107_3459: setp.gt.s32 %p3074, %r4473, 10; @%p3074 bra $L__BB107_3464; setp.eq.s32 %p3077, %r4473, 9; @%p3077 bra $L__BB107_3638; setp.eq.s32 %p3078, %r4473, 10; @%p3078 bra $L__BB107_3462; bra.uni $L__BB107_3644; $L__BB107_3462: ld.local.f32 %f23157, [%rd632]; setp.ltu.ftz.f32 %p3094, %f23157, 0f3F800000; @%p3094 bra $L__BB107_3439; add.s32 %r4539, %r4474, %r1247; add.s32 %r5807, %r4539, 1; bra.uni $L__BB107_3439; $L__BB107_3451: setp.gt.s32 %p3080, %r4473, 6; @%p3080 bra $L__BB107_3455; setp.eq.s32 %p3083, %r4473, 5; @%p3083 bra $L__BB107_3641; setp.eq.s32 %p3084, %r4473, 6; @%p3084 bra $L__BB107_3454; bra.uni $L__BB107_3644; $L__BB107_3454: ld.local.f32 %f23159, [%rd631]; ld.local.f32 %f23160, [%rd631+4]; ld.local.f32 %f23161, [%rd631+8]; st.local.f32 [%rd626], %f23159; st.local.f32 [%rd627], %f23160; st.local.f32 [%rd628], %f23161; bra.uni $L__BB107_3439; $L__BB107_3497: setp.eq.s32 %p3047, %r4473, 42; @%p3047 bra $L__BB107_3627; setp.eq.s32 %p3048, %r4473, 43; @%p3048 bra $L__BB107_3626; setp.eq.s32 %p3049, %r4473, 44; @%p3049 bra $L__BB107_3500; bra.uni $L__BB107_3644; $L__BB107_3500: mov.u32 %r4514, 44; add.s32 %r4515, %r1247, 1; { // callseq 666, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4514; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4515; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 666 bra.uni $L__BB107_3439; $L__BB107_3643: { // callseq 689, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 689 bra.uni $L__BB107_3439; $L__BB107_3593: { // callseq 622, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 622 bra.uni $L__BB107_3439; $L__BB107_3592: mov.u32 %r4479, 94; add.s32 %r4480, %r1247, 1; { // callseq 621, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4479; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4480; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 621 bra.uni $L__BB107_3439; $L__BB107_3477: setp.eq.s32 %p3064, %r4473, 21; @%p3064 bra $L__BB107_3633; setp.eq.s32 %p3065, %r4473, 22; @%p3065 bra $L__BB107_3479; bra.uni $L__BB107_3644; $L__BB107_3479: mov.u32 %r4525, 22; add.s32 %r4526, %r1247, 1; { // callseq 677, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4525; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4526; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 677 bra.uni $L__BB107_3439; $L__BB107_3468: setp.eq.s32 %p3072, %r4473, 13; @%p3072 bra $L__BB107_3636; setp.eq.s32 %p3073, %r4473, 14; @%p3073 bra $L__BB107_3470; bra.uni $L__BB107_3644; $L__BB107_3470: st.local.u32 [%rd632], %r4474; bra.uni $L__BB107_3439; $L__BB107_3448: setp.eq.s32 %p3086, %r4473, 3; @%p3086 bra $L__BB107_3642; setp.eq.s32 %p3087, %r4473, 4; @%p3087 bra $L__BB107_3450; bra.uni $L__BB107_3644; $L__BB107_3450: mov.u32 %r4543, 4; { // callseq 686, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4543, %r4474, %r4475, %r4476}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 686 bra.uni $L__BB107_3439; $L__BB107_3464: setp.eq.s32 %p3075, %r4473, 11; @%p3075 bra $L__BB107_3637; setp.eq.s32 %p3076, %r4473, 12; @%p3076 bra $L__BB107_3466; bra.uni $L__BB107_3644; $L__BB107_3466: { // callseq 682, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 682 bra.uni $L__BB107_3439; $L__BB107_3455: setp.eq.s32 %p3081, %r4473, 7; @%p3081 bra $L__BB107_3640; setp.eq.s32 %p3082, %r4473, 8; @%p3082 bra $L__BB107_3457; bra.uni $L__BB107_3644; $L__BB107_3457: mov.u32 %r4541, 8; { // callseq 684, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4541; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 684 bra.uni $L__BB107_3439; $L__BB107_3511: setp.eq.s32 %p3038, %r4473, 49; @%p3038 bra $L__BB107_3623; setp.eq.s32 %p3039, %r4473, 50; @%p3039 bra $L__BB107_3513; bra.uni $L__BB107_3644; $L__BB107_3513: { // callseq 660, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 660 bra.uni $L__BB107_3439; $L__BB107_3547: setp.eq.s32 %p3010, %r4473, 69; @%p3010 bra $L__BB107_3611; setp.eq.s32 %p3011, %r4473, 70; @%p3011 bra $L__BB107_3549; bra.uni $L__BB107_3644; $L__BB107_3549: mov.u32 %r4493, 70; { // callseq 640, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4493; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 640 bra.uni $L__BB107_3439; $L__BB107_3528: setp.eq.s32 %p3025, %r4473, 59; @%p3025 bra $L__BB107_3617; setp.eq.s32 %p3026, %r4473, 60; @%p3026 bra $L__BB107_3530; bra.uni $L__BB107_3644; $L__BB107_3530: add.s32 %r4506, %r1247, 1; { // callseq 650, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4506; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 650 bra.uni $L__BB107_3439; $L__BB107_3564: setp.eq.s32 %p2997, %r4473, 79; @%p2997 bra $L__BB107_3599; setp.eq.s32 %p2998, %r4473, 80; @%p2998 bra $L__BB107_3566; bra.uni $L__BB107_3644; $L__BB107_3566: { // callseq 632, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 632 bra.uni $L__BB107_3439; $L__BB107_3505: setp.eq.s32 %p3043, %r4473, 47; @%p3043 bra $L__BB107_3624; setp.eq.s32 %p3044, %r4473, 48; @%p3044 bra $L__BB107_3507; bra.uni $L__BB107_3644; $L__BB107_3507: { // callseq 662, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 662 bra.uni $L__BB107_3439; $L__BB107_3519: setp.eq.s32 %p3032, %r4473, 54; @%p3032 bra $L__BB107_3620; setp.eq.s32 %p3033, %r4473, 55; @%p3033 bra $L__BB107_3521; bra.uni $L__BB107_3644; $L__BB107_3521: ld.local.v4.f32 {%f23150, %f23151, %f23152, %f23153}, [%rd2173]; { // callseq 655, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f23150; st.param.f32 [param0+4], %f23151; st.param.f32 [param0+8], %f23152; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; .param .b32 param4; st.param.b32 [param4+0], %r4476; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 655 bra.uni $L__BB107_3439; $L__BB107_3555: setp.eq.s32 %p3004, %r4473, 74; @%p3004 bra $L__BB107_3608; setp.eq.s32 %p3005, %r4473, 75; @%p3005 bra $L__BB107_3557; bra.uni $L__BB107_3644; $L__BB107_3557: { // callseq 635, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 635 bra.uni $L__BB107_3439; $L__BB107_3536: setp.eq.s32 %p3019, %r4473, 64; @%p3019 bra $L__BB107_3614; setp.eq.s32 %p3020, %r4473, 65; @%p3020 bra $L__BB107_3538; bra.uni $L__BB107_3644; $L__BB107_3538: { // callseq 645, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 645 bra.uni $L__BB107_3439; $L__BB107_3572: setp.eq.s32 %p2991, %r4473, 84; @%p2991 bra $L__BB107_3596; setp.eq.s32 %p2992, %r4473, 85; @%p2992 bra $L__BB107_3574; bra.uni $L__BB107_3644; $L__BB107_3574: add.s32 %r4485, %r1247, 1; { // callseq 627, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4485; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 627 bra.uni $L__BB107_3439; $L__BB107_3632: mov.u32 %r4523, 23; add.s32 %r4524, %r1247, 1; { // callseq 676, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4523; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4524; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 676 bra.uni $L__BB107_3439; $L__BB107_3631: mov.u32 %r4522, 24; { // callseq 675, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4522; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 675 bra.uni $L__BB107_3439; $L__BB107_3635: ld.global.v4.u32 {%r4529, %r4530, %r4531, %r4532}, [%rd630+16]; add.s32 %r5807, %r1247, 2; st.local.u32 [%rd631], %r4530; st.local.u32 [%rd631+4], %r4531; st.local.u32 [%rd631+8], %r4532; bra.uni $L__BB107_3439; $L__BB107_3634: mov.u32 %r4528, 16; { // callseq 680, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4528; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 680 bra.uni $L__BB107_3439; $L__BB107_3638: ld.local.f32 %f23158, [%rd632]; setp.gtu.ftz.f32 %p3095, %f23158, 0f00000000; @%p3095 bra $L__BB107_3439; add.s32 %r4540, %r4474, %r1247; add.s32 %r5807, %r4540, 1; bra.uni $L__BB107_3439; $L__BB107_3641: st.local.u32 [%rd626], %r4474; st.local.u32 [%rd627], %r4475; st.local.u32 [%rd628], %r4476; bra.uni $L__BB107_3439; $L__BB107_3633: mov.u32 %r4527, 21; { // callseq 678, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4527; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 678 bra.uni $L__BB107_3439; $L__BB107_3636: mov.u32 %r4536, 13; add.s32 %r4537, %r1247, 1; mov.u32 %r4538, 524288; { // callseq 681, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b32 param1; st.param.b32 [param1+0], %r4538; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r4536; st.param.b32 [param3+4], %r4474; st.param.b32 [param3+8], %r4475; st.param.b32 [param3+12], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4537; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 681 bra.uni $L__BB107_3439; $L__BB107_3642: mov.u32 %r4544, 3; { // callseq 687, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4544, %r4474, %r4475, %r4476}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 687 bra.uni $L__BB107_3439; $L__BB107_3637: { // callseq 683, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 683 bra.uni $L__BB107_3439; $L__BB107_3640: mov.u32 %r4542, 7; { // callseq 685, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4542; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 685 bra.uni $L__BB107_3439; $L__BB107_3622: { // callseq 659, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 659 bra.uni $L__BB107_3439; $L__BB107_3621: add.s32 %r4513, %r1247, 1; { // callseq 658, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4513; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 658 bra.uni $L__BB107_3439; $L__BB107_3610: mov.u32 %r4492, 71; { // callseq 639, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4492; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 639 bra.uni $L__BB107_3439; $L__BB107_3609: { // callseq 638, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 638 bra.uni $L__BB107_3439; $L__BB107_3629: mov.u32 %r4518, 37; { // callseq 671, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4518, %r4474, %r4475, %r4476}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 671 bra.uni $L__BB107_3439; $L__BB107_3628: { // callseq 670, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; .param .b32 param4; st.param.b32 [param4+0], %r4476; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 670 bra.uni $L__BB107_3439; $L__BB107_3616: mov.u32 %r4504, 61; add.s32 %r4505, %r1247, 1; { // callseq 649, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4504; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4505; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 649 bra.uni $L__BB107_3439; $L__BB107_3615: mov.u32 %r4502, 62; add.s32 %r4503, %r1247, 1; { // callseq 648, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4502; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4503; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 648 bra.uni $L__BB107_3439; $L__BB107_3598: mov.u32 %r4487, 81; { // callseq 631, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4487; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 631 bra.uni $L__BB107_3439; $L__BB107_3597: mov.u32 %r4486, 82; { // callseq 630, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r4486; st.param.b32 [param2+4], %r4474; st.param.b32 [param2+8], %r4475; st.param.b32 [param2+12], %r4476; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 630 bra.uni $L__BB107_3439; $L__BB107_3625: { // callseq 665, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 665 bra.uni $L__BB107_3439; $L__BB107_3619: mov.u32 %r4511, 56; { // callseq 654, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4511; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 654 bra.uni $L__BB107_3439; $L__BB107_3618: mov.u32 %r4509, 57; add.s32 %r4510, %r1247, 1; { // callseq 653, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4509; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4510; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 653 bra.uni $L__BB107_3439; $L__BB107_3602: setp.eq.s32 %p3091, %r4476, 255; @%p3091 bra $L__BB107_3439; add.s64 %rd633, %rd631, 4; add.s64 %rd634, %rd631, 8; setp.eq.s32 %p3092, %r4475, 0; mul.wide.u32 %rd3550, %r4476, 4; add.s64 %rd635, %rd2157, %rd3550; @%p3092 bra $L__BB107_3607; setp.eq.s32 %p3093, %r4475, 1; @%p3093 bra $L__BB107_3606; bra.uni $L__BB107_3605; $L__BB107_3606: ld.local.f32 %f23148, [%rd633]; st.local.f32 [%rd635], %f23148; bra.uni $L__BB107_3439; $L__BB107_3600: setp.eq.s32 %p3090, %r4476, 255; @%p3090 bra $L__BB107_3439; ld.local.f32 %f23146, [%rd631]; add.s32 %r4490, %r4476, %r4475; mul.wide.u32 %rd3546, %r4490, 4; add.s64 %rd3547, %rd2157, %rd3546; st.local.f32 [%rd3547], %f23146; bra.uni $L__BB107_3439; $L__BB107_3613: add.s32 %r4498, %r1247, 1; { // callseq 644, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4498; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 644 bra.uni $L__BB107_3439; $L__BB107_3612: mov.u32 %r4497, 67; { // callseq 643, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r4497, %r4474, %r4475, %r4476}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 643 bra.uni $L__BB107_3439; $L__BB107_3595: add.s32 %r4484, %r1247, 1; { // callseq 626, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4484; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 626 bra.uni $L__BB107_3439; $L__BB107_3594: add.s32 %r4483, %r1247, 1; { // callseq 625, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4483; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 625 bra.uni $L__BB107_3439; $L__BB107_3627: { // callseq 668, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 668 bra.uni $L__BB107_3439; $L__BB107_3626: add.s32 %r4516, %r1247, 1; { // callseq 667, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4516; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 667 bra.uni $L__BB107_3439; $L__BB107_3623: { // callseq 661, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2172; .param .b64 param1; st.param.b64 [param1+0], %rd2156; .param .b32 param2; st.param.b32 [param2+0], %r4474; .param .b32 param3; st.param.b32 [param3+0], %r4475; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 661 bra.uni $L__BB107_3439; $L__BB107_3611: mov.u32 %r4494, 69; add.s32 %r4495, %r1247, 1; { // callseq 641, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4494; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4495; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 641 bra.uni $L__BB107_3439; $L__BB107_3617: add.s32 %r4507, %r1247, 1; { // callseq 651, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4507; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 651 bra.uni $L__BB107_3439; $L__BB107_3599: add.s32 %r4488, %r1247, 1; { // callseq 633, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; .param .b32 param4; st.param.b32 [param4+0], %r4488; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 633 bra.uni $L__BB107_3439; $L__BB107_3624: { // callseq 663, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1602; .param .b64 param1; st.param.b64 [param1+0], %rd2172; .param .b64 param2; st.param.b64 [param2+0], %rd2156; .param .b32 param3; st.param.b32 [param3+0], %r4474; .param .b32 param4; st.param.b32 [param4+0], %r4475; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 663 bra.uni $L__BB107_3439; $L__BB107_3620: add.s32 %r4512, %r1247, 1; { // callseq 656, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4512; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 656 bra.uni $L__BB107_3439; $L__BB107_3608: { // callseq 636, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; .param .b32 param3; st.param.b32 [param3+0], %r4476; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 636 bra.uni $L__BB107_3439; $L__BB107_3614: mov.u32 %r4499, 64; add.s32 %r4500, %r1247, 1; { // callseq 646, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4499; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; .param .b32 param2; st.param.b32 [param2+0], %r4500; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r5807, [retval0+0]; } // callseq 646 bra.uni $L__BB107_3439; $L__BB107_3596: { // callseq 628, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .b32 param1; st.param.b32 [param1+0], %r4474; .param .b32 param2; st.param.b32 [param2+0], %r4475; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 628 bra.uni $L__BB107_3439; $L__BB107_3630: mov.u32 %r4520, 26; { // callseq 673, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2156; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r4520; st.param.b32 [param1+4], %r4474; st.param.b32 [param1+8], %r4475; st.param.b32 [param1+12], %r4476; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 673 bra.uni $L__BB107_3439; $L__BB107_3607: ld.local.f32 %f23149, [%rd631]; st.local.f32 [%rd635], %f23149; bra.uni $L__BB107_3439; $L__BB107_3605: ld.local.f32 %f23147, [%rd634]; st.local.f32 [%rd635], %f23147; bra.uni $L__BB107_3439; $L__BB107_3644: add.s64 %rd3627, %rd2172, 52; cvta.to.local.u64 %rd3628, %rd3627; ld.local.u32 %r4547, [%rd3628]; and.b32 %r1337, %r4547, 2; @%p2946 bra $L__BB107_3647; bra.uni $L__BB107_3645; $L__BB107_3647: setp.eq.s32 %p3099, %r1337, 0; mov.f32 %f30347, 0f00000000; mov.f32 %f30348, %f30347; mov.f32 %f30349, %f30347; @%p3099 bra $L__BB107_3649; add.s64 %rd3639, %rd2172, 280; cvta.to.local.u64 %rd3640, %rd3639; ld.local.v2.f32 {%f30347, %f30348}, [%rd3640]; add.s64 %rd3641, %rd2172, 288; cvta.to.local.u64 %rd3642, %rd3641; ld.local.f32 %f30349, [%rd3642]; bra.uni $L__BB107_3649; $L__BB107_3645: setp.eq.s32 %p3097, %r1337, 0; mov.f32 %f30347, 0f00000000; mov.f32 %f30348, %f30347; mov.f32 %f30349, %f30347; @%p3097 bra $L__BB107_3649; add.s64 %rd3629, %rd605, 20; add.s64 %rd3630, %rd605, 12; cvta.to.local.u64 %rd3631, %rd3630; ld.local.v2.f32 {%f23165, %f23166}, [%rd3631]; cvta.to.local.u64 %rd3632, %rd3629; ld.local.v4.f32 {%f23169, %f23170, %f23171, %f23172}, [%rd3632]; mul.ftz.f32 %f23177, %f23166, %f23171; fma.rn.ftz.f32 %f23178, %f23165, %f23170, %f23177; fma.rn.ftz.f32 %f23179, %f23169, %f23172, %f23178; abs.ftz.f32 %f23180, %f23179; setp.gt.ftz.f32 %p3098, %f23180, 0f00000000; selp.f32 %f23181, 0f3F800000, 0f00000000, %p3098; add.s64 %rd3634, %rd2172, 280; cvta.to.local.u64 %rd3635, %rd3634; ld.local.v2.f32 {%f23182, %f23183}, [%rd3635]; add.s64 %rd3636, %rd2172, 288; cvta.to.local.u64 %rd3637, %rd3636; ld.local.f32 %f23186, [%rd3637]; mul.ftz.f32 %f30349, %f23181, %f23186; mul.ftz.f32 %f30348, %f23181, %f23183; mul.ftz.f32 %f30347, %f23182, %f23181; $L__BB107_3649: ld.local.f32 %f23192, [%rd280+28]; mul.ftz.f32 %f30350, %f30347, %f23192; mul.ftz.f32 %f30351, %f30348, %f23192; mul.ftz.f32 %f30352, %f30349, %f23192; @%p1255 bra $L__BB107_3651; ld.const.u64 %rd3643, [kernel_params+2368]; cvta.to.global.u64 %rd3644, %rd3643; mul.lo.s64 %rd3645, %rd281, 192; add.s64 %rd3646, %rd3644, %rd3645; ld.global.f32 %f23193, [%rd3646+28]; mul.ftz.f32 %f30350, %f30350, %f23193; ld.global.f32 %f23194, [%rd3646+32]; mul.ftz.f32 %f30351, %f30351, %f23194; ld.global.f32 %f23195, [%rd3646+36]; mul.ftz.f32 %f30352, %f30352, %f23195; $L__BB107_3651: setp.eq.ftz.f32 %p3101, %f30351, 0f00000000; setp.eq.ftz.f32 %p3102, %f30350, 0f00000000; and.pred %p3103, %p3102, %p3101; setp.eq.ftz.f32 %p3104, %f30352, 0f00000000; and.pred %p3105, %p3104, %p3103; @%p3105 bra $L__BB107_3809; add.u64 %rd4728, %SP, 3424; { // callseq 690, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd4728; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f30148; st.param.f32 [param1+4], %f30147; st.param.f32 [param1+8], %f30146; .param .b64 param2; st.param.b64 [param2+0], %rd2174; .param .b32 param3; st.param.b32 [param3+0], %r1233; .param .b32 retval0; call.uni (retval0), _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj, ( param0, param1, param2, param3 ); ld.param.f32 %f5397, [retval0+0]; } // callseq 690 ld.local.f32 %f5398, [%rd280+24]; div.approx.ftz.f32 %f23196, %f30350, %f5398; div.approx.ftz.f32 %f23197, %f30351, %f5398; div.approx.ftz.f32 %f23198, %f30352, %f5398; ld.local.f32 %f23199, [%rd283]; mul.ftz.f32 %f30298, %f23196, %f23199; ld.local.f32 %f23200, [%rd283+4]; mul.ftz.f32 %f30357, %f23197, %f23200; ld.local.f32 %f23201, [%rd283+8]; mul.ftz.f32 %f30356, %f23198, %f23201; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f30357; st.local.f32 [%rd283+8], %f30356; ld.local.f32 %f23202, [%rd283+12]; mul.ftz.f32 %f30355, %f23196, %f23202; ld.local.f32 %f23203, [%rd283+16]; mul.ftz.f32 %f30354, %f23197, %f23203; ld.local.f32 %f23204, [%rd283+20]; mul.ftz.f32 %f30353, %f23198, %f23204; st.local.f32 [%rd283+12], %f30355; st.local.f32 [%rd283+16], %f30354; st.local.f32 [%rd283+20], %f30353; ld.local.f32 %f23205, [%rd283+24]; mul.ftz.f32 %f30361, %f23196, %f23205; ld.local.f32 %f23206, [%rd283+28]; mul.ftz.f32 %f30360, %f23197, %f23206; ld.local.f32 %f23207, [%rd283+32]; mul.ftz.f32 %f30359, %f23198, %f23207; st.local.f32 [%rd283+24], %f30361; st.local.f32 [%rd283+28], %f30360; st.local.f32 [%rd283+32], %f30359; and.b32 %r4548, %r1233, 268435456; setp.eq.s32 %p3106, %r4548, 0; @%p3106 bra $L__BB107_3654; mul.ftz.f32 %f23208, %f5398, %f5398; fma.rn.ftz.f32 %f23209, %f5397, %f5397, %f23208; div.approx.ftz.f32 %f23210, %f23208, %f23209; mul.ftz.f32 %f30298, %f23210, %f30298; mul.ftz.f32 %f30357, %f23210, %f30357; mul.ftz.f32 %f30356, %f23210, %f30356; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f30357; st.local.f32 [%rd283+8], %f30356; mul.ftz.f32 %f30355, %f23210, %f30355; mul.ftz.f32 %f30354, %f23210, %f30354; mul.ftz.f32 %f30353, %f23210, %f30353; st.local.f32 [%rd283+12], %f30355; st.local.f32 [%rd283+16], %f30354; st.local.f32 [%rd283+20], %f30353; mul.ftz.f32 %f30361, %f23210, %f30361; mul.ftz.f32 %f30360, %f23210, %f30360; mul.ftz.f32 %f30359, %f23210, %f30359; st.local.f32 [%rd283+24], %f30361; st.local.f32 [%rd283+28], %f30360; st.local.f32 [%rd283+32], %f30359; $L__BB107_3654: ld.const.f32 %f5426, [kernel_params+1508]; setp.leu.ftz.f32 %p3107, %f5426, 0f00000000; mov.f32 %f30362, 0f00000000; @%p3107 bra $L__BB107_3659; add.s32 %r1338, %r602, 2; @%p1249 bra $L__BB107_3658; ld.const.v2.u32 {%r4549, %r4550}, [kernel_params+1600]; mov.b32 %f5427, %r4550; setp.geu.ftz.f32 %p3109, %f5427, 0f3F800000; setp.lt.ftz.f32 %p3110, %f5427, 0f3F800000; ld.const.u32 %r4553, [kernel_params+1576]; selp.b32 %r4554, %r4553, %r601, %p3110; add.s16 %rs551, %rs28, 2; and.b16 %rs552, %rs551, 255; cvt.u32.u16 %r4555, %rs552; xor.b32 %r4556, %r4554, %r4555; mul.lo.s32 %r4557, %r4556, -512718531; shr.u32 %r4558, %r4554, 16; xor.b32 %r4559, %r4557, %r4558; bfe.u32 %r4560, %r4559, 4, 4; shr.u32 %r4561, %r4554, 8; xor.b32 %r4562, %r4559, %r4561; xor.b32 %r4563, %r4562, %r4560; mul.lo.s32 %r4564, %r4563, 153742143; shr.u32 %r4565, %r4554, 23; xor.b32 %r4566, %r4564, %r4565; bfe.u32 %r4567, %r4566, 1, 7; xor.b32 %r4568, %r4567, %r4566; shr.u32 %r4569, %r4554, 27; or.b32 %r4570, %r4569, 1; mul.lo.s32 %r4571, %r4570, %r4568; mul.lo.s32 %r4572, %r4571, -1074439621; bfe.u32 %r4573, %r4572, 2, 6; xor.b32 %r4574, %r4573, %r4572; mul.lo.s32 %r4575, %r4574, -1638916925; bfe.u32 %r4576, %r4575, 2, 6; xor.b32 %r4577, %r4576, %r4575; mul.lo.s32 %r4578, %r4577, -933190689; and.b32 %r4579, %r4578, 255; bfe.u32 %r4580, %r4578, 5, 3; xor.b32 %r4581, %r4580, %r4579; add.s32 %r4582, %r4549, -1; xor.b32 %r1339, %r1338, 61; xor.b32 %r4583, %r4554, %r1339; mul.lo.s32 %r4584, %r4583, 9; shr.u32 %r4585, %r4584, 4; xor.b32 %r4586, %r4585, %r4584; mul.lo.s32 %r4587, %r4586, 668265261; add.s32 %r4588, %r4587, %r607; shr.u32 %r4589, %r4587, 16; or.b32 %r4590, %r4589, 1; mul.lo.s32 %r4591, %r4590, %r4588; mul.lo.s32 %r4592, %r4591, 89287766; xor.b32 %r4593, %r4592, %r4591; mul.lo.s32 %r4594, %r4593, 1403136100; xor.b32 %r4595, %r4594, %r4593; brev.b32 %r4596, %r4595; neg.s32 %r4597, %r4549; and.b32 %r4598, %r603, %r4597; and.b32 %r4599, %r4596, %r4582; or.b32 %r4600, %r4599, %r4598; mad.lo.s32 %r4601, %r4581, %r4549, %r4600; shl.b32 %r4602, %r4549, 8; rem.u32 %r4603, %r4601, %r4602; ld.const.u64 %rd3649, [kernel_params+2424]; cvta.to.global.u64 %rd3650, %rd3649; shl.b32 %r4604, %r4603, 1; mul.wide.u32 %rd3651, %r4604, 4; add.s64 %rd3652, %rd3650, %rd3651; ld.global.f32 %f30362, [%rd3652]; @%p3109 bra $L__BB107_3659; xor.b32 %r4605, %r1339, %r601; mul.lo.s32 %r4606, %r4605, 9; shr.u32 %r4607, %r4606, 4; xor.b32 %r4608, %r4607, %r4606; mul.lo.s32 %r4609, %r4608, 668265261; cvt.rn.f32.u32 %f23212, %r4609; mul.ftz.f32 %f23213, %f23212, 0f2F7FFFFE; fma.rn.ftz.f32 %f23214, %f23213, %f5427, %f30362; cvt.rmi.ftz.f32.f32 %f23215, %f23214; sub.ftz.f32 %f30362, %f23214, %f23215; bra.uni $L__BB107_3659; $L__BB107_3658: shr.u32 %r4610, %r1338, 16; xor.b32 %r4611, %r4610, %r1338; mul.lo.s32 %r4612, %r4611, 569420461; shr.u32 %r4613, %r4612, 15; xor.b32 %r4614, %r4613, %r4612; mul.lo.s32 %r4615, %r4614, -749064809; shr.u32 %r4616, %r4615, 15; xor.b32 %r4617, %r601, %r4615; xor.b32 %r4618, %r4617, %r4616; xor.b32 %r4619, %r4618, 1493655573; add.s32 %r4620, %r4619, %r607; shr.u32 %r4621, %r4619, 16; or.b32 %r4622, %r4621, 1; mul.lo.s32 %r4623, %r4622, %r4620; mul.lo.s32 %r4624, %r4623, 89287766; xor.b32 %r4625, %r4624, %r4623; mul.lo.s32 %r4626, %r4625, 1403136100; xor.b32 %r4627, %r4626, %r4625; xor.b32 %r4628, %r4618, -2052961194; brev.b32 %r4629, %r4627; mul.lo.s32 %r4630, %r4629, 1025551850; xor.b32 %r4631, %r4630, %r4629; add.s32 %r4632, %r4631, %r4628; shr.u32 %r4633, %r4628, 16; or.b32 %r4634, %r4633, 1; mul.lo.s32 %r4635, %r4632, %r4634; mul.lo.s32 %r4636, %r4635, 89287766; xor.b32 %r4637, %r4636, %r4635; mul.lo.s32 %r4638, %r4637, 1403136100; xor.b32 %r4639, %r4638, %r4637; brev.b32 %r4640, %r4639; cvt.rn.f32.u32 %f23216, %r4640; mul.ftz.f32 %f30362, %f23216, 0f2F7FFFFE; $L__BB107_3659: setp.eq.ftz.f32 %p3111, %f30360, 0f00000000; setp.eq.ftz.f32 %p3112, %f30361, 0f00000000; and.pred %p3113, %p3112, %p3111; setp.eq.ftz.f32 %p3114, %f30359, 0f00000000; and.pred %p3115, %p3114, %p3113; @%p3115 bra $L__BB107_3809; bra.uni $L__BB107_3660; $L__BB107_3809: ld.const.u32 %r1436, [kernel_params]; and.b32 %r4810, %r1436, 25165824; setp.eq.s32 %p3238, %r4810, 0; @%p3238 bra $L__BB107_3899; and.b32 %r1437, %r1436, 16777216; setp.ne.s32 %p3239, %r1437, 0; @%p3239 bra $L__BB107_3812; ld.global.u32 %r4811, [%rd27]; and.b32 %r4812, %r4811, 1; setp.eq.b32 %p3240, %r4812, 1; mov.pred %p3241, 0; xor.pred %p3242, %p3240, %p3241; not.pred %p3243, %p3242; @%p3243 bra $L__BB107_3899; $L__BB107_3812: add.s32 %r1438, %r602, 3; ld.const.u32 %r4813, [kernel_params+1596]; setp.eq.s32 %p3244, %r4813, 0; brev.b32 %r4814, %r603; mul.lo.s32 %r4815, %r4814, 1025551850; xor.b32 %r1439, %r4815, %r4814; @%p3244 bra $L__BB107_3815; ld.const.v2.u32 {%r4816, %r4817}, [kernel_params+1600]; mov.b32 %f5856, %r4817; setp.geu.ftz.f32 %p3245, %f5856, 0f3F800000; setp.lt.ftz.f32 %p3246, %f5856, 0f3F800000; ld.const.u32 %r4820, [kernel_params+1576]; selp.b32 %r4821, %r4820, %r601, %p3246; add.s16 %rs577, %rs28, 3; and.b16 %rs578, %rs577, 255; cvt.u32.u16 %r4822, %rs578; xor.b32 %r4823, %r4821, %r4822; mul.lo.s32 %r4824, %r4823, -512718531; shr.u32 %r4825, %r4821, 16; xor.b32 %r4826, %r4824, %r4825; bfe.u32 %r4827, %r4826, 4, 4; shr.u32 %r4828, %r4821, 8; xor.b32 %r4829, %r4826, %r4828; xor.b32 %r4830, %r4829, %r4827; mul.lo.s32 %r4831, %r4830, 153742143; shr.u32 %r4832, %r4821, 23; xor.b32 %r4833, %r4831, %r4832; bfe.u32 %r4834, %r4833, 1, 7; xor.b32 %r4835, %r4834, %r4833; shr.u32 %r4836, %r4821, 27; or.b32 %r4837, %r4836, 1; mul.lo.s32 %r4838, %r4837, %r4835; mul.lo.s32 %r4839, %r4838, -1074439621; bfe.u32 %r4840, %r4839, 2, 6; xor.b32 %r4841, %r4840, %r4839; mul.lo.s32 %r4842, %r4841, -1638916925; bfe.u32 %r4843, %r4842, 2, 6; xor.b32 %r4844, %r4843, %r4842; mul.lo.s32 %r4845, %r4844, -933190689; and.b32 %r4846, %r4845, 255; bfe.u32 %r4847, %r4845, 5, 3; xor.b32 %r4848, %r4847, %r4846; add.s32 %r4849, %r4816, -1; xor.b32 %r1440, %r1438, 61; xor.b32 %r4850, %r4821, %r1440; mul.lo.s32 %r4851, %r4850, 9; shr.u32 %r4852, %r4851, 4; xor.b32 %r4853, %r4852, %r4851; mul.lo.s32 %r4854, %r4853, 668265261; add.s32 %r4855, %r4854, %r1439; shr.u32 %r4856, %r4854, 16; or.b32 %r4857, %r4856, 1; mul.lo.s32 %r4858, %r4857, %r4855; mul.lo.s32 %r4859, %r4858, 89287766; xor.b32 %r4860, %r4859, %r4858; mul.lo.s32 %r4861, %r4860, 1403136100; xor.b32 %r4862, %r4861, %r4860; brev.b32 %r4863, %r4862; neg.s32 %r4864, %r4816; and.b32 %r4865, %r603, %r4864; and.b32 %r4866, %r4863, %r4849; or.b32 %r4867, %r4866, %r4865; mad.lo.s32 %r4868, %r4848, %r4816, %r4867; shl.b32 %r4869, %r4816, 8; rem.u32 %r4870, %r4868, %r4869; ld.const.u64 %rd4137, [kernel_params+2424]; cvta.to.global.u64 %rd4138, %rd4137; shl.b32 %r4871, %r4870, 1; mul.wide.u32 %rd4139, %r4871, 4; add.s64 %rd4140, %rd4138, %rd4139; ld.global.f32 %f30482, [%rd4140]; ld.global.f32 %f30483, [%rd4140+4]; @%p3245 bra $L__BB107_3819; xor.b32 %r4872, %r1440, %r601; mul.lo.s32 %r4873, %r4872, 9; shr.u32 %r4874, %r4873, 4; xor.b32 %r4875, %r4874, %r4873; mul.lo.s32 %r4876, %r4875, 668265261; cvt.rn.f32.u32 %f23961, %r4876; mul.ftz.f32 %f23962, %f23961, 0f2F7FFFFE; xor.b32 %r4877, %r601, %r1438; xor.b32 %r4878, %r4877, -905047700; mul.lo.s32 %r4879, %r4878, 9; shr.u32 %r4880, %r4879, 4; xor.b32 %r4881, %r4880, %r4879; mul.lo.s32 %r4882, %r4881, 668265261; cvt.rn.f32.u32 %f23963, %r4882; mul.ftz.f32 %f23964, %f23963, 0f2F7FFFFE; fma.rn.ftz.f32 %f23965, %f23962, %f5856, %f30482; fma.rn.ftz.f32 %f23966, %f23964, %f5856, %f30483; cvt.rmi.ftz.f32.f32 %f23967, %f23965; sub.ftz.f32 %f30482, %f23965, %f23967; cvt.rmi.ftz.f32.f32 %f23968, %f23966; sub.ftz.f32 %f30483, %f23966, %f23968; bra.uni $L__BB107_3819; $L__BB107_3815: shr.u32 %r4884, %r1438, 16; xor.b32 %r4885, %r4884, %r1438; mul.lo.s32 %r4886, %r4885, 569420461; shr.u32 %r4887, %r4886, 15; xor.b32 %r4888, %r4887, %r4886; mul.lo.s32 %r4889, %r4888, -749064809; shr.u32 %r4890, %r4889, 15; xor.b32 %r4891, %r601, %r4889; xor.b32 %r4892, %r4891, %r4890; xor.b32 %r1441, %r4892, -419546133; xor.b32 %r4893, %r4892, 508809841; add.s32 %r4894, %r4893, %r1439; shr.u32 %r4895, %r4893, 16; or.b32 %r4896, %r4895, 1; mul.lo.s32 %r4897, %r4896, %r4894; mul.lo.s32 %r4898, %r4897, 89287766; xor.b32 %r4899, %r4898, %r4897; mul.lo.s32 %r4900, %r4899, 1403136100; xor.b32 %r1442, %r4900, %r4899; setp.eq.s32 %p3247, %r1442, 0; mov.u32 %r5844, 0; @%p3247 bra $L__BB107_3818; mov.u32 %r5841, 0; mov.u64 %rd4142, sobol_burley_table; mov.u32 %r5842, %r1442; mov.u32 %r5844, %r5841; $L__BB107_3817: clz.b32 %r4903, %r5842; add.s32 %r4904, %r4903, %r5841; mul.wide.u32 %rd4141, %r4904, 4; add.s64 %rd4143, %rd4142, %rd4141; ld.const.u32 %r4905, [%rd4143+128]; xor.b32 %r5844, %r4905, %r5844; add.s32 %r5841, %r4904, 1; shl.b32 %r4906, %r5842, %r4903; shl.b32 %r5842, %r4906, 1; setp.ne.s32 %p3248, %r5842, 0; @%p3248 bra $L__BB107_3817; $L__BB107_3818: xor.b32 %r4907, %r1441, -525684874; brev.b32 %r4908, %r1442; shr.u32 %r4909, %r4907, 16; or.b32 %r4910, %r4909, 1; mul.lo.s32 %r4911, %r4908, 1025551850; xor.b32 %r4912, %r4911, %r4908; add.s32 %r4913, %r4912, %r4907; mul.lo.s32 %r4914, %r4913, %r4910; mul.lo.s32 %r4915, %r4914, 89287766; xor.b32 %r4916, %r4915, %r4914; xor.b32 %r4917, %r1441, -1802089138; mul.lo.s32 %r4918, %r5844, 1025551850; xor.b32 %r4919, %r4918, %r5844; add.s32 %r4920, %r4919, %r4917; shr.u32 %r4921, %r4917, 16; or.b32 %r4922, %r4921, 1; mul.lo.s32 %r4923, %r4920, %r4922; mul.lo.s32 %r4924, %r4923, 89287766; xor.b32 %r4925, %r4924, %r4923; mul.lo.s32 %r4926, %r4925, 1403136100; xor.b32 %r4927, %r4926, %r4925; brev.b32 %r4928, %r4927; cvt.rn.f32.u32 %f23969, %r4928; mul.lo.s32 %r4929, %r4916, 1403136100; xor.b32 %r4930, %r4929, %r4916; brev.b32 %r4931, %r4930; cvt.rn.f32.u32 %f23970, %r4931; mul.ftz.f32 %f30483, %f23969, 0f2F7FFFFE; mul.ftz.f32 %f30482, %f23970, 0f2F7FFFFE; $L__BB107_3819: ld.const.f32 %f5869, [kernel_params+1548]; ld.local.u32 %r1450, [%rd5+172]; setp.lt.s32 %p3249, %r1450, 1; mov.f32 %f30490, 0f00000000; mov.f32 %f30491, %f30490; mov.f32 %f30492, %f30490; mov.f32 %f30493, %f30490; mov.f32 %f30494, %f30490; mov.f32 %f30495, %f30490; @%p3249 bra $L__BB107_3836; add.s32 %r4933, %r1450, -1; and.b32 %r5848, %r1450, 3; setp.lt.u32 %p3250, %r4933, 3; mov.f32 %f30495, 0f00000000; mov.u32 %r5847, 0; mov.f32 %f30494, %f30495; mov.f32 %f30493, %f30495; mov.f32 %f30492, %f30495; mov.f32 %f30491, %f30495; mov.f32 %f30490, %f30495; @%p3250 bra $L__BB107_3831; sub.s32 %r5846, %r1450, %r5848; mov.f32 %f30495, 0f00000000; mov.u32 %r5847, 0; mov.f32 %f30494, %f30495; mov.f32 %f30493, %f30495; mov.f32 %f30492, %f30495; mov.f32 %f30491, %f30495; mov.f32 %f30490, %f30495; $L__BB107_3822: cvt.s64.s32 %rd896, %r5847; mul.wide.s32 %rd4144, %r5847, 80; add.s64 %rd4145, %rd2, %rd4144; add.s64 %rd897, %rd4145, 316; ld.local.u32 %r4935, [%rd4145+316]; add.s32 %r4936, %r4935, -2; setp.gt.u32 %p3251, %r4936, 6; @%p3251 bra $L__BB107_3824; ld.local.v4.f32 {%f23990, %f23991, %f23992, %f23993}, [%rd897+-12]; fma.rn.ftz.f32 %f30493, %f5869, %f23990, %f30493; fma.rn.ftz.f32 %f30494, %f5869, %f23991, %f30494; fma.rn.ftz.f32 %f30495, %f5869, %f23992, %f30495; add.ftz.f32 %f23997, %f23990, %f23991; add.ftz.f32 %f23998, %f23992, %f23997; mul.ftz.f32 %f23999, %f23998, 0f3EAAAAAB; abs.ftz.f32 %f24000, %f23999; ld.local.f32 %f24001, [%rd897+8]; ld.local.f32 %f24002, [%rd897+12]; ld.local.f32 %f24003, [%rd897+16]; fma.rn.ftz.f32 %f30490, %f24000, %f24001, %f30490; fma.rn.ftz.f32 %f30491, %f24000, %f24002, %f30491; fma.rn.ftz.f32 %f30492, %f24000, %f24003, %f30492; $L__BB107_3824: ld.local.u32 %r4937, [%rd897+80]; add.s32 %r4938, %r4937, -2; setp.gt.u32 %p3252, %r4938, 6; @%p3252 bra $L__BB107_3826; ld.local.v4.f32 {%f24004, %f24005, %f24006, %f24007}, [%rd897+68]; fma.rn.ftz.f32 %f30493, %f5869, %f24004, %f30493; fma.rn.ftz.f32 %f30494, %f5869, %f24005, %f30494; fma.rn.ftz.f32 %f30495, %f5869, %f24006, %f30495; add.ftz.f32 %f24011, %f24004, %f24005; add.ftz.f32 %f24012, %f24006, %f24011; mul.ftz.f32 %f24013, %f24012, 0f3EAAAAAB; abs.ftz.f32 %f24014, %f24013; ld.local.f32 %f24015, [%rd897+88]; ld.local.f32 %f24016, [%rd897+92]; ld.local.f32 %f24017, [%rd897+96]; fma.rn.ftz.f32 %f30490, %f24014, %f24015, %f30490; fma.rn.ftz.f32 %f30491, %f24014, %f24016, %f30491; fma.rn.ftz.f32 %f30492, %f24014, %f24017, %f30492; $L__BB107_3826: ld.local.u32 %r4939, [%rd897+160]; add.s32 %r4940, %r4939, -2; setp.gt.u32 %p3253, %r4940, 6; @%p3253 bra $L__BB107_3828; ld.local.v4.f32 {%f24018, %f24019, %f24020, %f24021}, [%rd897+148]; fma.rn.ftz.f32 %f30493, %f5869, %f24018, %f30493; fma.rn.ftz.f32 %f30494, %f5869, %f24019, %f30494; fma.rn.ftz.f32 %f30495, %f5869, %f24020, %f30495; add.ftz.f32 %f24025, %f24018, %f24019; add.ftz.f32 %f24026, %f24020, %f24025; mul.ftz.f32 %f24027, %f24026, 0f3EAAAAAB; abs.ftz.f32 %f24028, %f24027; ld.local.f32 %f24029, [%rd897+168]; ld.local.f32 %f24030, [%rd897+172]; ld.local.f32 %f24031, [%rd897+176]; fma.rn.ftz.f32 %f30490, %f24028, %f24029, %f30490; fma.rn.ftz.f32 %f30491, %f24028, %f24030, %f30491; fma.rn.ftz.f32 %f30492, %f24028, %f24031, %f30492; $L__BB107_3828: ld.local.u32 %r4941, [%rd897+240]; add.s32 %r4942, %r4941, -2; setp.gt.u32 %p3254, %r4942, 6; @%p3254 bra $L__BB107_3830; ld.local.v4.f32 {%f24032, %f24033, %f24034, %f24035}, [%rd897+228]; fma.rn.ftz.f32 %f30493, %f5869, %f24032, %f30493; fma.rn.ftz.f32 %f30494, %f5869, %f24033, %f30494; fma.rn.ftz.f32 %f30495, %f5869, %f24034, %f30495; add.ftz.f32 %f24039, %f24032, %f24033; add.ftz.f32 %f24040, %f24034, %f24039; mul.ftz.f32 %f24041, %f24040, 0f3EAAAAAB; abs.ftz.f32 %f24042, %f24041; ld.local.f32 %f24043, [%rd897+248]; ld.local.f32 %f24044, [%rd897+252]; ld.local.f32 %f24045, [%rd897+256]; fma.rn.ftz.f32 %f30490, %f24042, %f24043, %f30490; fma.rn.ftz.f32 %f30491, %f24042, %f24044, %f30491; fma.rn.ftz.f32 %f30492, %f24042, %f24045, %f30492; $L__BB107_3830: cvt.u32.u64 %r4943, %rd896; add.s32 %r5847, %r4943, 4; add.s32 %r5846, %r5846, -4; setp.ne.s32 %p3255, %r5846, 0; @%p3255 bra $L__BB107_3822; $L__BB107_3831: setp.eq.s32 %p3256, %r5848, 0; @%p3256 bra $L__BB107_3836; mul.wide.s32 %rd4146, %r5847, 80; add.s64 %rd4147, %rd2, %rd4146; add.s64 %rd4922, %rd4147, 332; $L__BB107_3833: .pragma "nounroll"; ld.local.u32 %r4944, [%rd4922+-16]; add.s32 %r4945, %r4944, -2; setp.gt.u32 %p3257, %r4945, 6; @%p3257 bra $L__BB107_3835; ld.local.f32 %f24046, [%rd4922+-28]; ld.local.f32 %f24047, [%rd4922+-24]; ld.local.f32 %f24048, [%rd4922+-20]; fma.rn.ftz.f32 %f30493, %f5869, %f24046, %f30493; fma.rn.ftz.f32 %f30494, %f5869, %f24047, %f30494; fma.rn.ftz.f32 %f30495, %f5869, %f24048, %f30495; add.ftz.f32 %f24049, %f24046, %f24047; add.ftz.f32 %f24050, %f24048, %f24049; mul.ftz.f32 %f24051, %f24050, 0f3EAAAAAB; abs.ftz.f32 %f24052, %f24051; ld.local.f32 %f24053, [%rd4922+-8]; ld.local.f32 %f24054, [%rd4922+-4]; ld.local.f32 %f24055, [%rd4922]; fma.rn.ftz.f32 %f30490, %f24052, %f24053, %f30490; fma.rn.ftz.f32 %f30491, %f24052, %f24054, %f30491; fma.rn.ftz.f32 %f30492, %f24052, %f24055, %f30492; $L__BB107_3835: add.s64 %rd4922, %rd4922, 80; add.s32 %r5848, %r5848, -1; setp.ne.s32 %p3258, %r5848, 0; @%p3258 bra $L__BB107_3833; $L__BB107_3836: setp.eq.ftz.f32 %p3259, %f30491, 0f00000000; setp.eq.ftz.f32 %p3260, %f30490, 0f00000000; and.pred %p3261, %p3260, %p3259; setp.eq.ftz.f32 %p3262, %f30492, 0f00000000; and.pred %p3263, %p3262, %p3261; @%p3263 bra $L__BB107_3838; bra.uni $L__BB107_3837; $L__BB107_3838: ld.local.v2.f32 {%f30545, %f30546}, [%rd5+-72]; ld.local.f32 %f30544, [%rd5+-76]; bra.uni $L__BB107_3839; $L__BB107_3837: mul.ftz.f32 %f24056, %f30491, %f30491; fma.rn.ftz.f32 %f24057, %f30490, %f30490, %f24056; fma.rn.ftz.f32 %f24058, %f30492, %f30492, %f24057; sqrt.approx.ftz.f32 %f24059, %f24058; div.approx.ftz.f32 %f30546, %f30492, %f24059; div.approx.ftz.f32 %f30545, %f30491, %f24059; div.approx.ftz.f32 %f30544, %f30490, %f24059; $L__BB107_3839: mul.ftz.f32 %f24062, %f30482, 0f40C90FDB; cos.approx.ftz.f32 %f24063, %f24062; sqrt.approx.ftz.f32 %f24064, %f30483; mul.ftz.f32 %f24065, %f24064, %f24063; sin.approx.ftz.f32 %f24066, %f24062; mul.ftz.f32 %f24067, %f24064, %f24066; mul.ftz.f32 %f24068, %f24065, %f24065; mov.f32 %f24069, 0f3F800000; sub.ftz.f32 %f24070, %f24069, %f24068; mul.ftz.f32 %f24071, %f24067, %f24067; sub.ftz.f32 %f24072, %f24070, %f24071; mov.f32 %f24073, 0f00000000; max.ftz.f32 %f24074, %f24072, %f24073; sqrt.approx.ftz.f32 %f24075, %f24074; setp.neu.ftz.f32 %p3264, %f30544, %f30545; setp.neu.ftz.f32 %p3265, %f30544, %f30546; or.pred %p3266, %p3264, %p3265; sub.ftz.f32 %f24076, %f30546, %f30545; neg.ftz.f32 %f24077, %f30545; neg.ftz.f32 %f24078, %f30546; selp.f32 %f24079, %f24078, %f30546, %p3266; add.ftz.f32 %f24080, %f30544, %f24079; selp.f32 %f24081, %f30545, %f24077, %p3266; sub.ftz.f32 %f24082, %f24081, %f30544; mul.ftz.f32 %f24083, %f24080, %f24080; fma.rn.ftz.f32 %f24084, %f24076, %f24076, %f24083; fma.rn.ftz.f32 %f24085, %f24082, %f24082, %f24084; sqrt.approx.ftz.f32 %f24086, %f24085; div.approx.ftz.f32 %f24087, %f24076, %f24086; div.approx.ftz.f32 %f24088, %f24080, %f24086; div.approx.ftz.f32 %f24089, %f24082, %f24086; mul.ftz.f32 %f24090, %f30545, %f24089; mul.ftz.f32 %f24091, %f30546, %f24088; sub.ftz.f32 %f24092, %f24090, %f24091; mul.ftz.f32 %f24093, %f30546, %f24087; mul.ftz.f32 %f24094, %f30544, %f24089; sub.ftz.f32 %f24095, %f24093, %f24094; mul.ftz.f32 %f24096, %f30544, %f24088; mul.ftz.f32 %f24097, %f30545, %f24087; sub.ftz.f32 %f24098, %f24096, %f24097; mul.ftz.f32 %f24099, %f24067, %f24092; mul.ftz.f32 %f24100, %f24067, %f24095; mul.ftz.f32 %f24101, %f24067, %f24098; fma.rn.ftz.f32 %f24102, %f24065, %f24087, %f24099; fma.rn.ftz.f32 %f24103, %f24065, %f24088, %f24100; fma.rn.ftz.f32 %f24104, %f24065, %f24089, %f24101; fma.rn.ftz.f32 %f5969, %f30544, %f24075, %f24102; fma.rn.ftz.f32 %f5970, %f30545, %f24075, %f24103; fma.rn.ftz.f32 %f5971, %f30546, %f24075, %f24104; add.u64 %rd4149, %SPL, 3424; ld.local.v4.f32 {%f30572, %f30573, %f30574, %f24108}, [%rd4149]; ld.local.u32 %r1460, [%rd5+-24]; and.b32 %r1461, %r1460, 1; setp.eq.s32 %p3267, %r1461, 0; ld.local.u32 %r4946, [%rd5+-40]; setp.gt.s32 %p3268, %r4946, -1; mov.u16 %rs741, 1; or.pred %p3269, %p3267, %p3268; @%p3269 bra $L__BB107_3882; ld.const.u64 %rd4150, [kernel_params+2160]; cvta.to.global.u64 %rd4151, %rd4150; ld.local.u32 %r4947, [%rd5+-12]; mul.wide.s32 %rd4152, %r4947, 208; add.s64 %rd4153, %rd4151, %rd4152; add.s64 %rd901, %rd4153, 184; ld.global.f32 %f5975, [%rd4153+184]; setp.leu.ftz.f32 %p3270, %f5975, 0f00000000; @%p3270 bra $L__BB107_3882; ld.local.f32 %f24109, [%rd5+-76]; ld.local.v2.f32 {%f24110, %f24111}, [%rd5+-72]; mul.ftz.f32 %f24114, %f5970, %f24110; fma.rn.ftz.f32 %f24115, %f5969, %f24109, %f24114; fma.rn.ftz.f32 %f24116, %f5971, %f24111, %f24115; setp.lt.ftz.f32 %p3271, %f24116, 0f00000000; neg.ftz.f32 %f24117, %f24116; selp.f32 %f5976, %f24117, %f24116, %p3271; @%p3271 bra $L__BB107_3843; bra.uni $L__BB107_3842; $L__BB107_3843: ld.local.v2.f32 {%f24120, %f24121}, [%rd5+-64]; neg.ftz.f32 %f30547, %f24120; neg.ftz.f32 %f30548, %f24121; ld.local.f32 %f24124, [%rd5+-56]; neg.ftz.f32 %f30549, %f24124; bra.uni $L__BB107_3844; $L__BB107_3842: ld.local.v2.f32 {%f30547, %f30548}, [%rd5+-64]; ld.local.f32 %f30549, [%rd5+-56]; $L__BB107_3844: mul.ftz.f32 %f24125, %f5970, %f30548; fma.rn.ftz.f32 %f24126, %f5969, %f30547, %f24125; fma.rn.ftz.f32 %f5986, %f5971, %f30549, %f24126; add.ftz.f32 %f24127, %f5976, %f5986; setp.lt.ftz.f32 %p3272, %f5976, %f5975; selp.f32 %f24128, 0f40000000, 0f3F800000, %p3272; mov.f32 %f24129, 0f3F800000; selp.f32 %f24130, %f24127, %f5986, %p3272; div.approx.ftz.f32 %f24131, %f24130, %f5975; sub.ftz.f32 %f24132, %f24128, %f24131; mov.f32 %f24133, 0f00000000; max.ftz.f32 %f24134, %f24132, %f24133; min.ftz.f32 %f5987, %f24134, %f24129; setp.leu.ftz.f32 %p3273, %f5987, 0f00000000; @%p3273 bra $L__BB107_3882; setp.eq.s32 %p3274, %r1460, 65; ld.local.s32 %rd902, [%rd5+-28]; @%p3274 bra $L__BB107_3847; bra.uni $L__BB107_3846; $L__BB107_3847: ld.global.u32 %r1462, [%rd901+-28]; ld.global.u32 %r1463, [%rd901+-32]; shl.b32 %r4959, %r1463, 1; cvt.rn.f32.s32 %f24135, %r4959; ld.local.f32 %f24136, [%rd5+-4]; mul.ftz.f32 %f24137, %f24136, %f24135; cvt.rzi.ftz.s32.f32 %r4960, %f24137; add.s32 %r4961, %r4959, -1; min.s32 %r1464, %r4960, %r4961; cvt.rn.f32.s32 %f24138, %r1464; sub.ftz.f32 %f6006, %f24137, %f24138; ld.const.u64 %rd4176, [kernel_params+2312]; cvta.to.global.u64 %rd903, %rd4176; ld.global.u32 %r5855, [%rd901+-20]; mul.wide.u32 %rd4177, %r5855, 16; add.s64 %rd4178, %rd903, %rd4177; ld.global.u16 %rs739, [%rd4178+12]; ld.global.u32 %r5854, [%rd4178+8]; ld.global.u64 %rd4942, [%rd4178]; setp.eq.s64 %p3275, %rd4942, 11; mov.u16 %rs737, %rs739; mov.u32 %r5849, %r5854; @%p3275 bra $L__BB107_3853; mov.u64 %rd4923, %rd4942; mov.u32 %r5849, %r5854; mov.u16 %rs737, %rs739; mov.u32 %r5850, %r5855; $L__BB107_3849: setp.eq.s64 %p3276, %rd4923, 0; @%p3276 bra $L__BB107_3851; add.s32 %r5850, %r5850, 2; bra.uni $L__BB107_3852; $L__BB107_3851: setp.eq.s16 %p3277, %rs737, 0; mov.u32 %r5853, -1; mov.u32 %r5850, %r5849; @%p3277 bra $L__BB107_3854; $L__BB107_3852: mul.wide.u32 %rd4179, %r5850, 16; add.s64 %rd4180, %rd903, %rd4179; ld.global.u16 %rs737, [%rd4180+12]; ld.global.u32 %r5849, [%rd4180+8]; ld.global.u64 %rd4923, [%rd4180]; setp.ne.s64 %p3278, %rd4923, 11; @%p3278 bra $L__BB107_3849; $L__BB107_3853: setp.eq.s16 %p3279, %rs737, 0; selp.b32 %r5853, -1, %r5849, %p3279; $L__BB107_3854: ld.const.u64 %rd4181, [kernel_params+2232]; cvta.to.global.u64 %rd4182, %rd4181; shl.b64 %rd4183, %rd902, 4; add.s64 %rd4184, %rd4182, %rd4183; ld.global.v4.u32 {%r4963, %r4964, %r4965, %r4966}, [%rd4184]; setp.eq.s32 %p3280, %r1464, %r1463; @%p3280 bra $L__BB107_3856; bra.uni $L__BB107_3855; $L__BB107_3856: ld.const.u64 %rd4190, [kernel_params+2256]; cvta.to.global.u64 %rd4191, %rd4190; mul.wide.u32 %rd4192, %r4966, 12; add.s64 %rd4924, %rd4191, %rd4192; add.s32 %r4973, %r4966, 1; mul.wide.u32 %rd4193, %r4973, 12; add.s64 %rd4927, %rd4191, %rd4193; add.s32 %r4974, %r4966, 2; mul.wide.u32 %rd4194, %r4974, 12; add.s64 %rd4930, %rd4191, %rd4194; add.s64 %rd4925, %rd4924, 4; add.s64 %rd4926, %rd4924, 8; add.s64 %rd4928, %rd4927, 4; add.s64 %rd4929, %rd4927, 8; add.s64 %rd4931, %rd4930, 4; add.s64 %rd4932, %rd4930, 8; bra.uni $L__BB107_3857; $L__BB107_3846: cvt.u32.u64 %r4948, %rd902; ld.const.v2.u64 {%rd4154, %rd4155}, [kernel_params+2224]; cvta.to.global.u64 %rd4158, %rd4155; mul.wide.s32 %rd4159, %r4948, 16; add.s64 %rd4160, %rd4158, %rd4159; ld.global.v4.u32 {%r4949, %r4950, %r4951, %r4952}, [%rd4160]; ld.const.u64 %rd4161, [kernel_params+2256]; cvta.to.global.u64 %rd4162, %rd4161; mul.wide.u32 %rd4163, %r4952, 12; add.s64 %rd4164, %rd4162, %rd4163; ld.global.f32 %f30559, [%rd4164]; ld.global.f32 %f30560, [%rd4164+4]; ld.global.f32 %f30561, [%rd4164+8]; add.s32 %r4957, %r4952, 1; mul.wide.u32 %rd4165, %r4957, 12; add.s64 %rd4166, %rd4162, %rd4165; ld.global.f32 %f30562, [%rd4166]; ld.global.f32 %f30563, [%rd4166+4]; ld.global.f32 %f30564, [%rd4166+8]; add.s32 %r4958, %r4952, 2; mul.wide.u32 %rd4167, %r4958, 12; add.s64 %rd4168, %rd4162, %rd4167; ld.global.f32 %f30565, [%rd4168]; ld.global.f32 %f30566, [%rd4168+4]; ld.global.f32 %f30567, [%rd4168+8]; cvta.to.global.u64 %rd4169, %rd4154; mul.wide.u32 %rd4170, %r4949, 12; add.s64 %rd4171, %rd4169, %rd4170; ld.global.f32 %f30550, [%rd4171]; ld.global.f32 %f30551, [%rd4171+4]; ld.global.f32 %f30552, [%rd4171+8]; mul.wide.u32 %rd4172, %r4950, 12; add.s64 %rd4173, %rd4169, %rd4172; ld.global.f32 %f30553, [%rd4173]; ld.global.f32 %f30554, [%rd4173+4]; ld.global.f32 %f30555, [%rd4173+8]; mul.wide.u32 %rd4174, %r4951, 12; add.s64 %rd4175, %rd4169, %rd4174; ld.global.f32 %f30556, [%rd4175]; ld.global.f32 %f30557, [%rd4175+4]; ld.global.f32 %f30558, [%rd4175+8]; bra.uni $L__BB107_3874; $L__BB107_3660: @%p3107 bra $L__BB107_3664; abs.ftz.f32 %f23217, %f30361; abs.ftz.f32 %f23218, %f30360; max.ftz.f32 %f23219, %f23217, %f23218; abs.ftz.f32 %f23220, %f30359; max.ftz.f32 %f23221, %f23219, %f23220; mul.ftz.f32 %f5432, %f23221, %f5426; setp.geu.ftz.f32 %p3117, %f5432, 0f3F800000; @%p3117 bra $L__BB107_3664; setp.le.ftz.f32 %p3118, %f5432, %f30362; @%p3118 bra $L__BB107_3809; rcp.approx.ftz.f32 %f23222, %f5432; mul.ftz.f32 %f30298, %f23222, %f30298; mul.ftz.f32 %f23223, %f23222, %f30357; mul.ftz.f32 %f23224, %f23222, %f30356; st.local.f32 [%rd283], %f30298; st.local.f32 [%rd283+4], %f23223; st.local.f32 [%rd283+8], %f23224; mul.ftz.f32 %f23225, %f23222, %f30355; mul.ftz.f32 %f23226, %f23222, %f30354; mul.ftz.f32 %f23227, %f23222, %f30353; st.local.f32 [%rd283+12], %f23225; st.local.f32 [%rd283+16], %f23226; st.local.f32 [%rd283+20], %f23227; mul.ftz.f32 %f23228, %f23222, %f30361; mul.ftz.f32 %f23229, %f23222, %f30360; mul.ftz.f32 %f23230, %f23222, %f30359; st.local.f32 [%rd283+24], %f23228; st.local.f32 [%rd283+28], %f23229; st.local.f32 [%rd283+32], %f23230; $L__BB107_3664: ld.local.v4.f32 {%f30389, %f30390, %f30391, %f23234}, [%rd279]; ld.local.u32 %r1340, [%rd5+-24]; and.b32 %r4641, %r1340, 1; setp.eq.b32 %p3119, %r4641, 1; not.pred %p3120, %p3119; ld.local.u32 %r4642, [%rd5+-40]; setp.gt.s32 %p3121, %r4642, -1; mov.u16 %rs731, 1; or.pred %p3122, %p3120, %p3121; @%p3122 bra $L__BB107_3707; ld.const.u64 %rd3653, [kernel_params+2160]; cvta.to.global.u64 %rd3654, %rd3653; ld.local.u32 %r4643, [%rd5+-12]; mul.wide.s32 %rd3655, %r4643, 208; add.s64 %rd3656, %rd3654, %rd3655; add.s64 %rd636, %rd3656, 184; ld.global.f32 %f5438, [%rd3656+184]; setp.leu.ftz.f32 %p3123, %f5438, 0f00000000; @%p3123 bra $L__BB107_3707; ld.local.f32 %f23235, [%rd5+-76]; ld.local.v2.f32 {%f23236, %f23237}, [%rd5+-72]; mul.ftz.f32 %f23240, %f30147, %f23236; fma.rn.ftz.f32 %f23241, %f30148, %f23235, %f23240; fma.rn.ftz.f32 %f23242, %f30146, %f23237, %f23241; setp.lt.ftz.f32 %p3124, %f23242, 0f00000000; neg.ftz.f32 %f23243, %f23242; selp.f32 %f5439, %f23243, %f23242, %p3124; @%p3124 bra $L__BB107_3668; bra.uni $L__BB107_3667; $L__BB107_3668: ld.local.v2.f32 {%f23246, %f23247}, [%rd5+-64]; neg.ftz.f32 %f30364, %f23246; neg.ftz.f32 %f30365, %f23247; ld.local.f32 %f23250, [%rd5+-56]; neg.ftz.f32 %f30366, %f23250; bra.uni $L__BB107_3669; $L__BB107_3855: setp.gt.s32 %p3281, %r1464, %r1463; selp.b32 %r4967, -1, 0, %p3281; add.s32 %r4968, %r1464, %r4967; mad.lo.s32 %r4969, %r4968, %r1462, %r5853; ld.const.u64 %rd4185, [kernel_params+2336]; cvta.to.global.u64 %rd4186, %rd4185; add.s32 %r4970, %r4963, %r4969; mul.wide.u32 %rd4187, %r4970, 12; add.s64 %rd4924, %rd4186, %rd4187; add.s32 %r4971, %r4964, %r4969; mul.wide.u32 %rd4188, %r4971, 12; add.s64 %rd4927, %rd4186, %rd4188; add.s32 %r4972, %r4965, %r4969; mul.wide.u32 %rd4189, %r4972, 12; add.s64 %rd4930, %rd4186, %rd4189; add.s64 %rd4925, %rd4924, 4; add.s64 %rd4926, %rd4924, 8; add.s64 %rd4928, %rd4927, 4; add.s64 %rd4929, %rd4927, 8; add.s64 %rd4931, %rd4930, 4; add.s64 %rd4932, %rd4930, 8; $L__BB107_3857: ld.global.f32 %f6007, [%rd4932]; ld.global.f32 %f6008, [%rd4931]; ld.global.f32 %f6009, [%rd4930]; ld.global.f32 %f6010, [%rd4929]; ld.global.f32 %f6011, [%rd4928]; ld.global.f32 %f6012, [%rd4927]; ld.global.f32 %f6013, [%rd4926]; ld.global.f32 %f6014, [%rd4925]; ld.global.f32 %f6015, [%rd4924]; add.s32 %r1479, %r1464, 1; setp.eq.s32 %p3282, %r1479, %r1463; @%p3282 bra $L__BB107_3859; bra.uni $L__BB107_3858; $L__BB107_3859: ld.const.u64 %rd4200, [kernel_params+2256]; cvta.to.global.u64 %rd4201, %rd4200; mul.wide.u32 %rd4202, %r4966, 12; add.s64 %rd4933, %rd4201, %rd4202; add.s32 %r4981, %r4966, 1; mul.wide.u32 %rd4203, %r4981, 12; add.s64 %rd4936, %rd4201, %rd4203; add.s32 %r4982, %r4966, 2; mul.wide.u32 %rd4204, %r4982, 12; add.s64 %rd4939, %rd4201, %rd4204; add.s64 %rd4934, %rd4933, 4; add.s64 %rd4935, %rd4933, 8; add.s64 %rd4937, %rd4936, 4; add.s64 %rd4938, %rd4936, 8; add.s64 %rd4940, %rd4939, 4; add.s64 %rd4941, %rd4939, 8; bra.uni $L__BB107_3860; $L__BB107_3858: setp.ge.s32 %p3283, %r1464, %r1463; selp.b32 %r4975, -1, 0, %p3283; add.s32 %r4976, %r1479, %r4975; mad.lo.s32 %r4977, %r4976, %r1462, %r5853; ld.const.u64 %rd4195, [kernel_params+2336]; cvta.to.global.u64 %rd4196, %rd4195; add.s32 %r4978, %r4963, %r4977; mul.wide.u32 %rd4197, %r4978, 12; add.s64 %rd4933, %rd4196, %rd4197; add.s32 %r4979, %r4964, %r4977; mul.wide.u32 %rd4198, %r4979, 12; add.s64 %rd4936, %rd4196, %rd4198; add.s32 %r4980, %r4965, %r4977; mul.wide.u32 %rd4199, %r4980, 12; add.s64 %rd4939, %rd4196, %rd4199; add.s64 %rd4934, %rd4933, 4; add.s64 %rd4935, %rd4933, 8; add.s64 %rd4937, %rd4936, 4; add.s64 %rd4938, %rd4936, 8; add.s64 %rd4940, %rd4939, 4; add.s64 %rd4941, %rd4939, 8; $L__BB107_3860: sub.ftz.f32 %f6016, %f24129, %f6006; ld.global.f32 %f24140, [%rd4933]; mul.ftz.f32 %f24141, %f6006, %f24140; ld.global.f32 %f24142, [%rd4934]; mul.ftz.f32 %f24143, %f6006, %f24142; ld.global.f32 %f24144, [%rd4935]; mul.ftz.f32 %f24145, %f6006, %f24144; fma.rn.ftz.f32 %f30559, %f6016, %f6015, %f24141; fma.rn.ftz.f32 %f30560, %f6016, %f6014, %f24143; fma.rn.ftz.f32 %f30561, %f6016, %f6013, %f24145; ld.global.f32 %f24146, [%rd4936]; mul.ftz.f32 %f24147, %f6006, %f24146; ld.global.f32 %f24148, [%rd4937]; mul.ftz.f32 %f24149, %f6006, %f24148; ld.global.f32 %f24150, [%rd4938]; mul.ftz.f32 %f24151, %f6006, %f24150; fma.rn.ftz.f32 %f30562, %f6016, %f6012, %f24147; fma.rn.ftz.f32 %f30563, %f6016, %f6011, %f24149; fma.rn.ftz.f32 %f30564, %f6016, %f6010, %f24151; ld.global.f32 %f24152, [%rd4939]; mul.ftz.f32 %f24153, %f6006, %f24152; ld.global.f32 %f24154, [%rd4940]; mul.ftz.f32 %f24155, %f6006, %f24154; ld.global.f32 %f24156, [%rd4941]; mul.ftz.f32 %f24157, %f6006, %f24156; fma.rn.ftz.f32 %f30565, %f6016, %f6009, %f24153; fma.rn.ftz.f32 %f30566, %f6016, %f6008, %f24155; fma.rn.ftz.f32 %f30567, %f6016, %f6007, %f24157; setp.eq.s64 %p3284, %rd4942, 12; @%p3284 bra $L__BB107_3866; mov.u32 %r5858, -1; $L__BB107_3862: setp.eq.s64 %p3285, %rd4942, 0; @%p3285 bra $L__BB107_3864; add.s32 %r5855, %r5855, 2; bra.uni $L__BB107_3865; $L__BB107_3864: setp.eq.s16 %p3286, %rs739, 0; mov.u32 %r5855, %r5854; @%p3286 bra $L__BB107_3867; $L__BB107_3865: mul.wide.u32 %rd4205, %r5855, 16; add.s64 %rd4206, %rd903, %rd4205; ld.global.u16 %rs739, [%rd4206+12]; ld.global.u32 %r5854, [%rd4206+8]; ld.global.u64 %rd4942, [%rd4206]; setp.ne.s64 %p3287, %rd4942, 12; @%p3287 bra $L__BB107_3862; $L__BB107_3866: setp.eq.s16 %p3288, %rs739, 0; selp.b32 %r5858, -1, %r5854, %p3288; $L__BB107_3867: @%p3280 bra $L__BB107_3869; bra.uni $L__BB107_3868; $L__BB107_3869: ld.const.u64 %rd4212, [kernel_params+2224]; cvta.to.global.u64 %rd4213, %rd4212; mul.wide.u32 %rd4214, %r4963, 12; add.s64 %rd4943, %rd4213, %rd4214; mul.wide.u32 %rd4215, %r4964, 12; add.s64 %rd4946, %rd4213, %rd4215; mul.wide.u32 %rd4216, %r4965, 12; add.s64 %rd4949, %rd4213, %rd4216; add.s64 %rd4944, %rd4943, 4; add.s64 %rd4945, %rd4943, 8; add.s64 %rd4947, %rd4946, 4; add.s64 %rd4948, %rd4946, 8; add.s64 %rd4950, %rd4949, 4; add.s64 %rd4951, %rd4949, 8; bra.uni $L__BB107_3870; $L__BB107_3868: setp.gt.s32 %p3290, %r1464, %r1463; selp.b32 %r4984, -1, 0, %p3290; add.s32 %r4985, %r1464, %r4984; mad.lo.s32 %r4986, %r4985, %r1462, %r5858; ld.const.u64 %rd4207, [kernel_params+2336]; cvta.to.global.u64 %rd4208, %rd4207; add.s32 %r4987, %r4986, %r4963; mul.wide.u32 %rd4209, %r4987, 12; add.s64 %rd4943, %rd4208, %rd4209; add.s32 %r4988, %r4986, %r4964; mul.wide.u32 %rd4210, %r4988, 12; add.s64 %rd4946, %rd4208, %rd4210; add.s32 %r4989, %r4986, %r4965; mul.wide.u32 %rd4211, %r4989, 12; add.s64 %rd4949, %rd4208, %rd4211; add.s64 %rd4944, %rd4943, 4; add.s64 %rd4945, %rd4943, 8; add.s64 %rd4947, %rd4946, 4; add.s64 %rd4948, %rd4946, 8; add.s64 %rd4950, %rd4949, 4; add.s64 %rd4951, %rd4949, 8; $L__BB107_3870: ld.global.f32 %f6026, [%rd4951]; ld.global.f32 %f6027, [%rd4950]; ld.global.f32 %f6028, [%rd4949]; ld.global.f32 %f6029, [%rd4948]; ld.global.f32 %f6030, [%rd4947]; ld.global.f32 %f6031, [%rd4946]; ld.global.f32 %f6032, [%rd4945]; ld.global.f32 %f6033, [%rd4944]; ld.global.f32 %f6034, [%rd4943]; @%p3282 bra $L__BB107_3872; bra.uni $L__BB107_3871; $L__BB107_3872: ld.const.u64 %rd4222, [kernel_params+2224]; cvta.to.global.u64 %rd4223, %rd4222; mul.wide.u32 %rd4224, %r4963, 12; add.s64 %rd4952, %rd4223, %rd4224; mul.wide.u32 %rd4225, %r4964, 12; add.s64 %rd4955, %rd4223, %rd4225; mul.wide.u32 %rd4226, %r4965, 12; add.s64 %rd4958, %rd4223, %rd4226; add.s64 %rd4953, %rd4952, 4; add.s64 %rd4954, %rd4952, 8; add.s64 %rd4956, %rd4955, 4; add.s64 %rd4957, %rd4955, 8; add.s64 %rd4959, %rd4958, 4; add.s64 %rd4960, %rd4958, 8; bra.uni $L__BB107_3873; $L__BB107_3871: setp.ge.s32 %p3292, %r1464, %r1463; selp.b32 %r4990, -1, 0, %p3292; add.s32 %r4991, %r1479, %r4990; mad.lo.s32 %r4992, %r4991, %r1462, %r5858; ld.const.u64 %rd4217, [kernel_params+2336]; cvta.to.global.u64 %rd4218, %rd4217; add.s32 %r4993, %r4992, %r4963; mul.wide.u32 %rd4219, %r4993, 12; add.s64 %rd4952, %rd4218, %rd4219; add.s32 %r4994, %r4992, %r4964; mul.wide.u32 %rd4220, %r4994, 12; add.s64 %rd4955, %rd4218, %rd4220; add.s32 %r4995, %r4992, %r4965; mul.wide.u32 %rd4221, %r4995, 12; add.s64 %rd4958, %rd4218, %rd4221; add.s64 %rd4953, %rd4952, 4; add.s64 %rd4954, %rd4952, 8; add.s64 %rd4956, %rd4955, 4; add.s64 %rd4957, %rd4955, 8; add.s64 %rd4959, %rd4958, 4; add.s64 %rd4960, %rd4958, 8; $L__BB107_3873: ld.global.f32 %f24158, [%rd4952]; mul.ftz.f32 %f24159, %f6006, %f24158; ld.global.f32 %f24160, [%rd4953]; mul.ftz.f32 %f24161, %f6006, %f24160; ld.global.f32 %f24162, [%rd4954]; mul.ftz.f32 %f24163, %f6006, %f24162; fma.rn.ftz.f32 %f30550, %f6016, %f6034, %f24159; fma.rn.ftz.f32 %f30551, %f6016, %f6033, %f24161; fma.rn.ftz.f32 %f30552, %f6016, %f6032, %f24163; ld.global.f32 %f24164, [%rd4955]; mul.ftz.f32 %f24165, %f6006, %f24164; ld.global.f32 %f24166, [%rd4956]; mul.ftz.f32 %f24167, %f6006, %f24166; ld.global.f32 %f24168, [%rd4957]; mul.ftz.f32 %f24169, %f6006, %f24168; fma.rn.ftz.f32 %f30553, %f6016, %f6031, %f24165; fma.rn.ftz.f32 %f30554, %f6016, %f6030, %f24167; fma.rn.ftz.f32 %f30555, %f6016, %f6029, %f24169; ld.global.f32 %f24170, [%rd4958]; mul.ftz.f32 %f24171, %f6006, %f24170; ld.global.f32 %f24172, [%rd4959]; mul.ftz.f32 %f24173, %f6006, %f24172; ld.global.f32 %f24174, [%rd4960]; mul.ftz.f32 %f24175, %f6006, %f24174; fma.rn.ftz.f32 %f30556, %f6016, %f6028, %f24171; fma.rn.ftz.f32 %f30557, %f6016, %f6027, %f24173; fma.rn.ftz.f32 %f30558, %f6016, %f6026, %f24175; $L__BB107_3874: ld.local.f32 %f6062, [%rd5+-20]; sub.ftz.f32 %f24177, %f24129, %f6062; ld.local.f32 %f24178, [%rd5+-16]; sub.ftz.f32 %f6063, %f24177, %f24178; mul.ftz.f32 %f24179, %f30559, %f6063; mul.ftz.f32 %f24180, %f30560, %f6063; mul.ftz.f32 %f24181, %f30561, %f6063; fma.rn.ftz.f32 %f24182, %f30562, %f6062, %f24179; fma.rn.ftz.f32 %f24183, %f30563, %f6062, %f24180; fma.rn.ftz.f32 %f24184, %f30564, %f6062, %f24181; fma.rn.ftz.f32 %f6064, %f30565, %f24178, %f24182; fma.rn.ftz.f32 %f6065, %f30566, %f24178, %f24183; fma.rn.ftz.f32 %f6066, %f30567, %f24178, %f24184; mul.ftz.f32 %f24185, %f30550, %f6063; mul.ftz.f32 %f24186, %f30551, %f6063; mul.ftz.f32 %f24187, %f30552, %f6063; fma.rn.ftz.f32 %f24188, %f30553, %f6062, %f24185; fma.rn.ftz.f32 %f24189, %f30554, %f6062, %f24186; fma.rn.ftz.f32 %f24190, %f30555, %f6062, %f24187; fma.rn.ftz.f32 %f30568, %f30556, %f24178, %f24188; fma.rn.ftz.f32 %f30569, %f30557, %f24178, %f24189; fma.rn.ftz.f32 %f30570, %f30558, %f24178, %f24190; ld.local.u32 %r1488, [%rd5+-32]; and.b32 %r4996, %r1488, 4; setp.ne.s32 %p3293, %r4996, 0; @%p3293 bra $L__BB107_3878; and.b32 %r4997, %r1488, 2; setp.eq.s32 %p3294, %r4997, 0; @%p3294 bra $L__BB107_3877; bra.uni $L__BB107_3876; $L__BB107_3877: ld.global.v4.f32 {%f24218, %f24219, %f24220, %f24221}, [%rd901+-184]; ld.global.v4.f32 {%f24225, %f24226, %f24227, %f24228}, [%rd901+-168]; ld.global.v4.f32 {%f24232, %f24233, %f24234, %f24235}, [%rd901+-152]; mul.ftz.f32 %f24239, %f30569, %f24219; fma.rn.ftz.f32 %f24240, %f30568, %f24218, %f24239; fma.rn.ftz.f32 %f6073, %f30570, %f24220, %f24240; mul.ftz.f32 %f24241, %f30569, %f24226; fma.rn.ftz.f32 %f24242, %f30568, %f24225, %f24241; fma.rn.ftz.f32 %f6074, %f30570, %f24227, %f24242; mul.ftz.f32 %f24243, %f30569, %f24233; fma.rn.ftz.f32 %f24244, %f30568, %f24232, %f24243; fma.rn.ftz.f32 %f30570, %f30570, %f24234, %f24244; mov.f32 %f30568, %f6073; mov.f32 %f30569, %f6074; bra.uni $L__BB107_3878; $L__BB107_3876: ld.local.v4.f32 {%f24191, %f24192, %f24193, %f24194}, [%rd5+56]; mul.ftz.f32 %f24198, %f30569, %f24192; fma.rn.ftz.f32 %f24199, %f30568, %f24191, %f24198; fma.rn.ftz.f32 %f6070, %f30570, %f24193, %f24199; ld.local.v4.f32 {%f24200, %f24201, %f24202, %f24203}, [%rd5+72]; mul.ftz.f32 %f24207, %f30569, %f24201; fma.rn.ftz.f32 %f24208, %f30568, %f24200, %f24207; fma.rn.ftz.f32 %f6071, %f30570, %f24202, %f24208; ld.local.v4.f32 {%f24209, %f24210, %f24211, %f24212}, [%rd5+88]; mul.ftz.f32 %f24216, %f30569, %f24210; fma.rn.ftz.f32 %f24217, %f30568, %f24209, %f24216; fma.rn.ftz.f32 %f30570, %f30570, %f24211, %f24217; mov.f32 %f30568, %f6070; mov.f32 %f30569, %f6071; $L__BB107_3878: sub.ftz.f32 %f6079, %f30559, %f30565; sub.ftz.f32 %f24245, %f30556, %f30550; sub.ftz.f32 %f6080, %f30560, %f30566; sub.ftz.f32 %f24246, %f30557, %f30551; mul.ftz.f32 %f24247, %f24246, %f6080; fma.rn.ftz.f32 %f24248, %f24245, %f6079, %f24247; sub.ftz.f32 %f6081, %f30561, %f30567; sub.ftz.f32 %f24249, %f30558, %f30552; fma.rn.ftz.f32 %f24250, %f24249, %f6081, %f24248; sub.ftz.f32 %f6082, %f30562, %f30565; sub.ftz.f32 %f24251, %f30556, %f30553; sub.ftz.f32 %f6083, %f30563, %f30566; sub.ftz.f32 %f24252, %f30557, %f30554; mul.ftz.f32 %f24253, %f24252, %f6083; fma.rn.ftz.f32 %f24254, %f24251, %f6082, %f24253; sub.ftz.f32 %f6084, %f30564, %f30567; sub.ftz.f32 %f24255, %f30558, %f30555; fma.rn.ftz.f32 %f24256, %f24255, %f6084, %f24254; sub.ftz.f32 %f6085, %f30562, %f30559; sub.ftz.f32 %f24257, %f30553, %f30550; sub.ftz.f32 %f6086, %f30563, %f30560; sub.ftz.f32 %f24258, %f30554, %f30551; mul.ftz.f32 %f24259, %f24258, %f6086; fma.rn.ftz.f32 %f24260, %f24257, %f6085, %f24259; sub.ftz.f32 %f6087, %f30564, %f30561; sub.ftz.f32 %f24261, %f30555, %f30552; fma.rn.ftz.f32 %f24262, %f24261, %f6087, %f24260; mul.ftz.f32 %f24263, %f24250, %f6063; add.ftz.f32 %f24264, %f6063, 0fBF800000; add.ftz.f32 %f24265, %f24250, %f24256; add.ftz.f32 %f24266, %f24262, %f24265; mul.ftz.f32 %f24267, %f6063, %f24266; mul.ftz.f32 %f24268, %f6062, %f24267; fma.rn.ftz.f32 %f24269, %f24263, %f24264, %f24268; mul.ftz.f32 %f24270, %f6062, %f24256; add.ftz.f32 %f24271, %f6062, 0fBF800000; fma.rn.ftz.f32 %f24272, %f24271, %f24270, %f24269; mul.ftz.f32 %f24273, %f30548, %f30569; fma.rn.ftz.f32 %f24274, %f30547, %f30568, %f24273; fma.rn.ftz.f32 %f24275, %f30549, %f30570, %f24274; setp.gt.ftz.f32 %p3295, %f24275, 0f00000000; sub.ftz.f32 %f6088, %f30559, %f30562; sub.ftz.f32 %f6089, %f30560, %f30563; sub.ftz.f32 %f6090, %f30561, %f30564; sub.ftz.f32 %f6091, %f30565, %f30559; sub.ftz.f32 %f6092, %f30566, %f30560; sub.ftz.f32 %f6093, %f30567, %f30561; sub.ftz.f32 %f6094, %f30565, %f30562; sub.ftz.f32 %f6095, %f30566, %f30563; sub.ftz.f32 %f6096, %f30567, %f30564; mul.ftz.f32 %f6097, %f24272, 0f3F000000; @%p3295 bra $L__BB107_3880; bra.uni $L__BB107_3879; $L__BB107_3880: mul.ftz.f32 %f24328, %f30551, %f6086; fma.rn.ftz.f32 %f24329, %f30550, %f6085, %f24328; fma.rn.ftz.f32 %f24330, %f30552, %f6087, %f24329; mul.ftz.f32 %f24331, %f30551, %f6092; fma.rn.ftz.f32 %f24332, %f30550, %f6091, %f24331; fma.rn.ftz.f32 %f24333, %f30552, %f6093, %f24332; max.ftz.f32 %f24334, %f24330, %f24333; max.ftz.f32 %f24336, %f24334, %f24133; mul.ftz.f32 %f24337, %f30554, %f6089; fma.rn.ftz.f32 %f24338, %f30553, %f6088, %f24337; fma.rn.ftz.f32 %f24339, %f30555, %f6090, %f24338; mul.ftz.f32 %f24340, %f30554, %f6095; fma.rn.ftz.f32 %f24341, %f30553, %f6094, %f24340; fma.rn.ftz.f32 %f24342, %f30555, %f6096, %f24341; max.ftz.f32 %f24343, %f24339, %f24342; max.ftz.f32 %f24344, %f24343, %f24133; mul.ftz.f32 %f24345, %f30557, %f6080; fma.rn.ftz.f32 %f24346, %f30556, %f6079, %f24345; fma.rn.ftz.f32 %f24347, %f30558, %f6081, %f24346; mul.ftz.f32 %f24348, %f30557, %f6083; fma.rn.ftz.f32 %f24349, %f30556, %f6082, %f24348; fma.rn.ftz.f32 %f24350, %f30558, %f6084, %f24349; max.ftz.f32 %f24351, %f24347, %f24350; max.ftz.f32 %f24352, %f24351, %f24133; sub.ftz.f32 %f24353, %f30559, %f6064; sub.ftz.f32 %f24354, %f30560, %f6065; mul.ftz.f32 %f24355, %f30551, %f24354; fma.rn.ftz.f32 %f24356, %f30550, %f24353, %f24355; sub.ftz.f32 %f24357, %f30561, %f6066; fma.rn.ftz.f32 %f24358, %f30552, %f24357, %f24356; add.ftz.f32 %f24359, %f24358, %f24336; max.ftz.f32 %f24360, %f24359, %f24133; sub.ftz.f32 %f24361, %f30562, %f6064; sub.ftz.f32 %f24362, %f30563, %f6065; mul.ftz.f32 %f24363, %f30554, %f24362; fma.rn.ftz.f32 %f24364, %f30553, %f24361, %f24363; sub.ftz.f32 %f24365, %f30564, %f6066; fma.rn.ftz.f32 %f24366, %f30555, %f24365, %f24364; add.ftz.f32 %f24367, %f24366, %f24344; max.ftz.f32 %f24368, %f24367, %f24133; sub.ftz.f32 %f24369, %f30565, %f6064; sub.ftz.f32 %f24370, %f30566, %f6065; mul.ftz.f32 %f24371, %f30557, %f24370; fma.rn.ftz.f32 %f24372, %f30556, %f24369, %f24371; sub.ftz.f32 %f24373, %f30567, %f6066; fma.rn.ftz.f32 %f24374, %f30558, %f24373, %f24372; add.ftz.f32 %f24375, %f24374, %f24352; max.ftz.f32 %f24376, %f24375, %f24133; min.ftz.f32 %f24377, %f24360, %f24368; min.ftz.f32 %f24378, %f24377, %f24376; max.ftz.f32 %f30571, %f24378, %f6097; bra.uni $L__BB107_3881; $L__BB107_3879: mul.ftz.f32 %f24276, %f30551, %f6089; fma.rn.ftz.f32 %f24277, %f30550, %f6088, %f24276; fma.rn.ftz.f32 %f24278, %f30552, %f6090, %f24277; mul.ftz.f32 %f24279, %f30551, %f6080; fma.rn.ftz.f32 %f24280, %f30550, %f6079, %f24279; fma.rn.ftz.f32 %f24281, %f30552, %f6081, %f24280; max.ftz.f32 %f24282, %f24278, %f24281; max.ftz.f32 %f24284, %f24282, %f24133; mul.ftz.f32 %f24285, %f30554, %f6086; fma.rn.ftz.f32 %f24286, %f30553, %f6085, %f24285; fma.rn.ftz.f32 %f24287, %f30555, %f6087, %f24286; mul.ftz.f32 %f24288, %f30554, %f6083; fma.rn.ftz.f32 %f24289, %f30553, %f6082, %f24288; fma.rn.ftz.f32 %f24290, %f30555, %f6084, %f24289; max.ftz.f32 %f24291, %f24287, %f24290; max.ftz.f32 %f24292, %f24291, %f24133; mul.ftz.f32 %f24293, %f30557, %f6092; fma.rn.ftz.f32 %f24294, %f30556, %f6091, %f24293; fma.rn.ftz.f32 %f24295, %f30558, %f6093, %f24294; mul.ftz.f32 %f24296, %f30557, %f6095; fma.rn.ftz.f32 %f24297, %f30556, %f6094, %f24296; fma.rn.ftz.f32 %f24298, %f30558, %f6096, %f24297; max.ftz.f32 %f24299, %f24295, %f24298; max.ftz.f32 %f24300, %f24299, %f24133; sub.ftz.f32 %f24301, %f6064, %f30559; sub.ftz.f32 %f24302, %f6065, %f30560; mul.ftz.f32 %f24303, %f30551, %f24302; fma.rn.ftz.f32 %f24304, %f30550, %f24301, %f24303; sub.ftz.f32 %f24305, %f6066, %f30561; fma.rn.ftz.f32 %f24306, %f30552, %f24305, %f24304; add.ftz.f32 %f24307, %f24306, %f24284; max.ftz.f32 %f24308, %f24307, %f24133; sub.ftz.f32 %f24309, %f6064, %f30562; sub.ftz.f32 %f24310, %f6065, %f30563; mul.ftz.f32 %f24311, %f30554, %f24310; fma.rn.ftz.f32 %f24312, %f30553, %f24309, %f24311; sub.ftz.f32 %f24313, %f6066, %f30564; fma.rn.ftz.f32 %f24314, %f30555, %f24313, %f24312; add.ftz.f32 %f24315, %f24314, %f24292; max.ftz.f32 %f24316, %f24315, %f24133; sub.ftz.f32 %f24317, %f6064, %f30565; sub.ftz.f32 %f24318, %f6065, %f30566; mul.ftz.f32 %f24319, %f30557, %f24318; fma.rn.ftz.f32 %f24320, %f30556, %f24317, %f24319; sub.ftz.f32 %f24321, %f6066, %f30567; fma.rn.ftz.f32 %f24322, %f30558, %f24321, %f24320; add.ftz.f32 %f24323, %f24322, %f24300; max.ftz.f32 %f24324, %f24323, %f24133; min.ftz.f32 %f24325, %f24308, %f24316; min.ftz.f32 %f24326, %f24325, %f24324; neg.ftz.f32 %f24327, %f24326; min.ftz.f32 %f30571, %f24327, %f6097; $L__BB107_3881: mul.ftz.f32 %f24379, %f30568, %f30571; mul.ftz.f32 %f24380, %f30569, %f30571; mul.ftz.f32 %f24381, %f30570, %f30571; fma.rn.ftz.f32 %f30572, %f5987, %f24379, %f30572; fma.rn.ftz.f32 %f30573, %f5987, %f24380, %f30573; fma.rn.ftz.f32 %f30574, %f5987, %f24381, %f30574; setp.gt.ftz.f32 %p3296, %f5986, 0f00000000; selp.u16 %rs741, 1, 0, %p3296; $L__BB107_3882: setp.eq.s16 %p3297, %rs741, 0; @%p3297 bra $L__BB107_3891; @%p3267 bra $L__BB107_3891; ld.const.u64 %rd4227, [kernel_params+2232]; cvta.to.global.u64 %rd4228, %rd4227; ld.local.v2.u32 {%r4998, %r4999}, [%rd5+-32]; mul.wide.s32 %rd4229, %r4999, 16; add.s64 %rd4230, %rd4228, %rd4229; ld.const.u64 %rd4231, [kernel_params+2256]; cvta.to.global.u64 %rd4232, %rd4231; ld.global.u32 %r5001, [%rd4230+12]; mul.wide.u32 %rd4233, %r5001, 12; add.s64 %rd4234, %rd4232, %rd4233; ld.global.f32 %f6110, [%rd4234]; ld.global.f32 %f6111, [%rd4234+4]; ld.global.f32 %f6112, [%rd4234+8]; add.s32 %r5002, %r5001, 1; mul.wide.u32 %rd4235, %r5002, 12; add.s64 %rd4236, %rd4232, %rd4235; ld.global.f32 %f6113, [%rd4236]; ld.global.f32 %f6114, [%rd4236+4]; ld.global.f32 %f6115, [%rd4236+8]; add.s32 %r5003, %r5001, 2; mul.wide.u32 %rd4237, %r5003, 12; add.s64 %rd4238, %rd4232, %rd4237; ld.global.f32 %f6116, [%rd4238]; ld.global.f32 %f6117, [%rd4238+4]; ld.global.f32 %f6118, [%rd4238+8]; and.b32 %r5004, %r4998, 4; setp.ne.s32 %p3299, %r5004, 0; mov.f32 %f30587, %f30572; mov.f32 %f30588, %f30573; mov.f32 %f30589, %f30574; mov.f32 %f30590, %f5969; mov.f32 %f30591, %f5970; mov.f32 %f30592, %f5971; @%p3299 bra $L__BB107_3889; and.b32 %r5005, %r4998, 2; setp.eq.s32 %p3300, %r5005, 0; @%p3300 bra $L__BB107_3887; bra.uni $L__BB107_3886; $L__BB107_3887: ld.const.u64 %rd4239, [kernel_params+2160]; cvta.to.global.u64 %rd4240, %rd4239; ld.local.u32 %r5006, [%rd5+-12]; mul.wide.s32 %rd4241, %r5006, 208; add.s64 %rd4242, %rd4240, %rd4241; ld.global.v4.f32 {%f30575, %f30576, %f30577, %f30578}, [%rd4242+48]; ld.global.v4.f32 {%f30579, %f30580, %f30581, %f30582}, [%rd4242+64]; ld.global.v4.f32 {%f30583, %f30584, %f30585, %f30586}, [%rd4242+80]; bra.uni $L__BB107_3888; $L__BB107_3886: ld.local.v4.f32 {%f30575, %f30576, %f30577, %f30578}, [%rd5+104]; ld.local.v4.f32 {%f30579, %f30580, %f30581, %f30582}, [%rd5+120]; ld.local.v4.f32 {%f30583, %f30584, %f30585, %f30586}, [%rd5+136]; $L__BB107_3888: mul.ftz.f32 %f24406, %f30573, %f30576; fma.rn.ftz.f32 %f24407, %f30572, %f30575, %f24406; fma.rn.ftz.f32 %f24408, %f30574, %f30577, %f24407; mul.ftz.f32 %f24409, %f30573, %f30580; fma.rn.ftz.f32 %f24410, %f30572, %f30579, %f24409; fma.rn.ftz.f32 %f24411, %f30574, %f30581, %f24410; mul.ftz.f32 %f24412, %f30573, %f30584; fma.rn.ftz.f32 %f24413, %f30572, %f30583, %f24412; fma.rn.ftz.f32 %f24414, %f30574, %f30585, %f24413; add.ftz.f32 %f30587, %f30578, %f24408; add.ftz.f32 %f30588, %f30582, %f24411; add.ftz.f32 %f30589, %f30586, %f24414; mul.ftz.f32 %f24415, %f5970, %f30576; fma.rn.ftz.f32 %f24416, %f5969, %f30575, %f24415; mul.ftz.f32 %f24417, %f5970, %f30580; fma.rn.ftz.f32 %f24418, %f5969, %f30579, %f24417; mul.ftz.f32 %f24419, %f5970, %f30584; fma.rn.ftz.f32 %f24420, %f5969, %f30583, %f24419; fma.rn.ftz.f32 %f30590, %f5971, %f30577, %f24416; fma.rn.ftz.f32 %f30591, %f5971, %f30581, %f24418; fma.rn.ftz.f32 %f30592, %f5971, %f30585, %f24420; $L__BB107_3889: sub.ftz.f32 %f24421, %f6110, %f30587; sub.ftz.f32 %f24422, %f6111, %f30588; sub.ftz.f32 %f24423, %f6112, %f30589; sub.ftz.f32 %f24424, %f6113, %f30587; sub.ftz.f32 %f24425, %f6114, %f30588; sub.ftz.f32 %f24426, %f6115, %f30589; sub.ftz.f32 %f24427, %f6116, %f30587; sub.ftz.f32 %f24428, %f6117, %f30588; sub.ftz.f32 %f24429, %f6118, %f30589; sub.ftz.f32 %f24430, %f24427, %f24421; sub.ftz.f32 %f24431, %f24428, %f24422; sub.ftz.f32 %f24432, %f24429, %f24423; sub.ftz.f32 %f24433, %f24421, %f24424; sub.ftz.f32 %f24434, %f24422, %f24425; sub.ftz.f32 %f24435, %f24423, %f24426; sub.ftz.f32 %f24436, %f24424, %f24427; sub.ftz.f32 %f24437, %f24425, %f24428; sub.ftz.f32 %f24438, %f24426, %f24429; add.ftz.f32 %f24439, %f24427, %f24421; add.ftz.f32 %f24440, %f24428, %f24422; add.ftz.f32 %f24441, %f24429, %f24423; mul.ftz.f32 %f24442, %f24440, %f24432; mul.ftz.f32 %f24443, %f24441, %f24431; sub.ftz.f32 %f24444, %f24442, %f24443; mul.ftz.f32 %f24445, %f24441, %f24430; mul.ftz.f32 %f24446, %f24439, %f24432; sub.ftz.f32 %f24447, %f24445, %f24446; mul.ftz.f32 %f24448, %f24439, %f24431; mul.ftz.f32 %f24449, %f24440, %f24430; sub.ftz.f32 %f24450, %f24448, %f24449; mul.ftz.f32 %f24451, %f30591, %f24447; fma.rn.ftz.f32 %f24452, %f30590, %f24444, %f24451; fma.rn.ftz.f32 %f24453, %f30592, %f24450, %f24452; add.ftz.f32 %f24454, %f24421, %f24424; add.ftz.f32 %f24455, %f24422, %f24425; add.ftz.f32 %f24456, %f24423, %f24426; mul.ftz.f32 %f24457, %f24455, %f24435; mul.ftz.f32 %f24458, %f24456, %f24434; sub.ftz.f32 %f24459, %f24457, %f24458; mul.ftz.f32 %f24460, %f24456, %f24433; mul.ftz.f32 %f24461, %f24454, %f24435; sub.ftz.f32 %f24462, %f24460, %f24461; mul.ftz.f32 %f24463, %f24454, %f24434; mul.ftz.f32 %f24464, %f24455, %f24433; sub.ftz.f32 %f24465, %f24463, %f24464; mul.ftz.f32 %f24466, %f30591, %f24462; fma.rn.ftz.f32 %f24467, %f30590, %f24459, %f24466; fma.rn.ftz.f32 %f24468, %f30592, %f24465, %f24467; add.ftz.f32 %f24469, %f24424, %f24427; add.ftz.f32 %f24470, %f24425, %f24428; add.ftz.f32 %f24471, %f24426, %f24429; mul.ftz.f32 %f24472, %f24470, %f24438; mul.ftz.f32 %f24473, %f24471, %f24437; sub.ftz.f32 %f24474, %f24472, %f24473; mul.ftz.f32 %f24475, %f24471, %f24436; mul.ftz.f32 %f24476, %f24469, %f24438; sub.ftz.f32 %f24477, %f24475, %f24476; mul.ftz.f32 %f24478, %f24469, %f24437; mul.ftz.f32 %f24479, %f24470, %f24436; sub.ftz.f32 %f24480, %f24478, %f24479; mul.ftz.f32 %f24481, %f30591, %f24477; fma.rn.ftz.f32 %f24482, %f30590, %f24474, %f24481; fma.rn.ftz.f32 %f24483, %f30592, %f24480, %f24482; add.ftz.f32 %f24484, %f24453, %f24468; add.ftz.f32 %f24485, %f24483, %f24484; abs.ftz.f32 %f24486, %f24485; mul.ftz.f32 %f24487, %f24486, 0f34000000; min.ftz.f32 %f24488, %f24468, %f24483; min.ftz.f32 %f24489, %f24453, %f24488; max.ftz.f32 %f24490, %f24468, %f24483; max.ftz.f32 %f24491, %f24453, %f24490; setp.ge.ftz.f32 %p3301, %f24489, %f24487; mul.ftz.f32 %f24492, %f24486, 0fB4000000; setp.le.ftz.f32 %p3302, %f24491, %f24492; or.pred %p3303, %p3301, %p3302; @%p3303 bra $L__BB107_3891; ld.local.v2.f32 {%f24493, %f24494}, [%rd5+-64]; mul.ftz.f32 %f24497, %f24493, 0f43800000; cvt.rzi.ftz.s32.f32 %r5007, %f24497; mul.ftz.f32 %f24498, %f24494, 0f43800000; cvt.rzi.ftz.s32.f32 %r5008, %f24498; ld.local.f32 %f24499, [%rd5+-56]; mul.ftz.f32 %f24500, %f24499, 0f43800000; cvt.rzi.ftz.s32.f32 %r5009, %f24500; neg.s32 %r5010, %r5007; setp.lt.ftz.f32 %p3304, %f30572, 0f00000000; selp.b32 %r5011, %r5010, %r5007, %p3304; mov.b32 %r5012, %f30572; add.s32 %r5013, %r5011, %r5012; mov.b32 %f24501, %r5013; neg.s32 %r5014, %r5008; setp.lt.ftz.f32 %p3305, %f30573, 0f00000000; selp.b32 %r5015, %r5014, %r5008, %p3305; mov.b32 %r5016, %f30573; add.s32 %r5017, %r5015, %r5016; mov.b32 %f24502, %r5017; neg.s32 %r5018, %r5009; setp.lt.ftz.f32 %p3306, %f30574, 0f00000000; selp.b32 %r5019, %r5018, %r5009, %p3306; mov.b32 %r5020, %f30574; add.s32 %r5021, %r5019, %r5020; mov.b32 %f24503, %r5021; abs.ftz.f32 %f24504, %f30572; setp.lt.ftz.f32 %p3307, %f24504, 0f3D000000; fma.rn.ftz.f32 %f24505, %f24493, 0f37800000, %f30572; abs.ftz.f32 %f24506, %f30573; setp.lt.ftz.f32 %p3308, %f24506, 0f3D000000; fma.rn.ftz.f32 %f24507, %f24494, 0f37800000, %f30573; abs.ftz.f32 %f24508, %f30574; setp.lt.ftz.f32 %p3309, %f24508, 0f3D000000; fma.rn.ftz.f32 %f24509, %f24499, 0f37800000, %f30574; selp.f32 %f30574, %f24509, %f24503, %p3309; selp.f32 %f30573, %f24507, %f24502, %p3308; selp.f32 %f30572, %f24505, %f24501, %p3307; $L__BB107_3891: ld.const.f32 %f6182, [kernel_params+1540]; ld.local.f32 %f6183, [%rd5+-4]; ld.local.u32 %r5022, [%rd5+-12]; selp.b32 %r1490, -1, %r5022, %p3297; ld.local.u32 %r5023, [%rd5+-28]; selp.b32 %r1491, -1, %r5023, %p3297; ld.const.u64 %rd4243, [kernel_params+4376]; cvta.to.global.u64 %rd4244, %rd4243; atom.global.add.u32 %r5024, [%rd4244], 1; ld.const.u64 %rd4245, [kernel_params+4256]; cvta.to.global.u64 %rd4246, %rd4245; add.s64 %rd4247, %rd4246, 12; atom.global.add.u32 %r5025, [%rd4247], 1; ld.const.u64 %rd4248, [kernel_params+3416]; cvta.to.global.u64 %rd4249, %rd4248; cvt.s64.s32 %rd1017, %r5024; mul.wide.s32 %rd4250, %r5024, 2; add.s64 %rd4251, %rd4249, %rd4250; mov.u16 %rs582, 3; st.global.u16 [%rd4251], %rs582; and.b32 %r5026, %r1436, 131072; setp.eq.s32 %p3311, %r5026, 0; @%p3311 bra $L__BB107_3894; mov.u32 %r5859, 0; mov.u64 %rd4253, kernel_params; $L__BB107_3893: mul.wide.s32 %rd4252, %r5859, 16; add.s64 %rd4254, %rd4253, %rd4252; ld.const.v2.u64 {%rd4255, %rd4256}, [%rd4254+2768]; cvta.to.global.u64 %rd4259, %rd4256; add.s64 %rd4261, %rd4259, %rd1282; ld.global.u32 %r5028, [%rd4261]; cvta.to.global.u64 %rd4262, %rd4255; add.s64 %rd4263, %rd4262, %rd1282; ld.global.u32 %r5029, [%rd4263]; ld.const.v2.u64 {%rd4264, %rd4265}, [%rd4254+3744]; cvta.to.global.u64 %rd4268, %rd4264; shl.b64 %rd4269, %rd1017, 2; add.s64 %rd4270, %rd4268, %rd4269; st.global.u32 [%rd4270], %r5029; cvta.to.global.u64 %rd4271, %rd4265; add.s64 %rd4272, %rd4271, %rd4269; st.global.u32 [%rd4272], %r5028; add.s32 %r5859, %r5859, 1; setp.ne.s32 %p3312, %r5028, -1; @%p3312 bra $L__BB107_3893; $L__BB107_3894: ld.const.u64 %rd4273, [kernel_params+3496]; cvta.to.global.u64 %rd4274, %rd4273; mul.lo.s64 %rd4275, %rd1017, 12; add.s64 %rd4276, %rd4274, %rd4275; st.global.f32 [%rd4276], %f30572; st.global.f32 [%rd4276+4], %f30573; st.global.f32 [%rd4276+8], %f30574; ld.const.v2.u64 {%rd4277, %rd4278}, [kernel_params+3504]; mov.u32 %r5030, 0; cvta.to.global.u64 %rd4281, %rd4277; add.s64 %rd4282, %rd4281, %rd4275; st.global.f32 [%rd4282], %f5969; st.global.f32 [%rd4282+4], %f5970; st.global.f32 [%rd4282+8], %f5971; cvta.to.global.u64 %rd4283, %rd4278; shl.b64 %rd4284, %rd1017, 2; add.s64 %rd4285, %rd4283, %rd4284; st.global.u32 [%rd4285], %r5030; ld.const.v2.u64 {%rd4286, %rd4287}, [kernel_params+3520]; cvta.to.global.u64 %rd4290, %rd4286; add.s64 %rd4291, %rd4290, %rd4284; st.global.f32 [%rd4291], %f6182; cvta.to.global.u64 %rd4292, %rd4287; add.s64 %rd4293, %rd4292, %rd4284; st.global.f32 [%rd4293], %f6183; ld.const.u64 %rd4294, [kernel_params+3536]; cvta.to.global.u64 %rd4295, %rd4294; add.s64 %rd4296, %rd4295, %rd4284; st.global.u32 [%rd4296], %r5030; ld.const.u64 %rd4297, [kernel_params+3584]; cvta.to.global.u64 %rd4298, %rd4297; add.s64 %rd4299, %rd4298, %rd4284; st.global.u32 [%rd4299], %r1490; ld.const.u64 %rd4300, [kernel_params+3576]; cvta.to.global.u64 %rd4301, %rd4300; add.s64 %rd4302, %rd4301, %rd4284; st.global.u32 [%rd4302], %r1491; ld.const.u64 %rd4303, [kernel_params+3632]; cvta.to.global.u64 %rd4304, %rd4303; add.s64 %rd4305, %rd4304, %rd4284; mov.u32 %r5031, -1; st.global.u32 [%rd4305], %r5031; ld.const.u64 %rd4306, [kernel_params+3624]; cvta.to.global.u64 %rd4307, %rd4306; add.s64 %rd4308, %rd4307, %rd4284; st.global.u32 [%rd4308], %r5031; ld.const.v2.u64 {%rd4309, %rd4310}, [kernel_params+2464]; cvta.to.global.u64 %rd4313, %rd4309; add.s64 %rd4315, %rd4313, %rd2151; ld.global.u16 %rs294, [%rd4315]; cvta.to.global.u64 %rd4316, %rd4310; add.s64 %rd4317, %rd4316, %rd2151; ld.global.u16 %rs295, [%rd4317]; ld.global.u32 %r1494, [%rd27]; ld.const.u64 %rd1018, [kernel_params+2584]; ld.local.u32 %r1495, [%rd5+-36]; and.b32 %r5032, %r1495, 524288; setp.ne.s32 %p3313, %r5032, 0; mov.f32 %f30599, %f24069; mov.f32 %f30600, %f24069; mov.f32 %f30601, %f24069; @%p3313 bra $L__BB107_3897; and.b32 %r5033, %r1495, 512; setp.eq.s32 %p3314, %r5033, 0; mov.f32 %f30599, %f24073; mov.f32 %f30600, %f24073; mov.f32 %f30601, %f24073; @%p3314 bra $L__BB107_3897; ld.local.v2.f32 {%f30600, %f30601}, [%rd5+208]; ld.local.f32 %f30599, [%rd5+204]; $L__BB107_3897: sub.ftz.f32 %f24519, %f24069, %f30599; sub.ftz.f32 %f24520, %f24069, %f30600; sub.ftz.f32 %f24521, %f24069, %f30601; cvt.ftz.sat.f32.f32 %f24522, %f24519; cvt.ftz.sat.f32.f32 %f24523, %f24520; cvt.ftz.sat.f32.f32 %f24524, %f24521; cvta.to.global.u64 %rd4322, %rd1018; mul.lo.s64 %rd4323, %rd3, 12; add.s64 %rd4324, %rd4322, %rd4323; ld.global.f32 %f24525, [%rd4324]; mul.ftz.f32 %f24526, %f24522, %f24525; ld.global.f32 %f24527, [%rd4324+4]; mul.ftz.f32 %f24528, %f24523, %f24527; ld.global.f32 %f24529, [%rd4324+8]; mul.ftz.f32 %f24530, %f24524, %f24529; ld.const.v2.u64 {%rd4325, %rd4326}, [kernel_params+2448]; cvta.to.global.u64 %rd4329, %rd4325; add.s64 %rd4331, %rd4329, %rd1282; ld.global.u32 %r5034, [%rd4331]; ld.const.v2.u64 {%rd4332, %rd4333}, [kernel_params+3344]; cvta.to.global.u64 %rd4336, %rd4332; add.s64 %rd4338, %rd4336, %rd4284; st.global.u32 [%rd4338], %r5034; ld.const.v2.u64 {%rd4339, %rd4340}, [kernel_params+2528]; cvta.to.global.u64 %rd4343, %rd4340; add.s64 %rd4345, %rd4343, %rd2151; ld.global.u16 %rs583, [%rd4345]; ld.const.v2.u64 {%rd4346, %rd4347}, [kernel_params+3360]; cvta.to.global.u64 %rd4350, %rd4347; shl.b64 %rd4351, %rd1017, 1; add.s64 %rd4352, %rd4350, %rd4351; st.global.u16 [%rd4352], %rs583; cvta.to.global.u64 %rd4353, %rd4339; add.s64 %rd4354, %rd4353, %rd1282; ld.global.u32 %r5035, [%rd4354]; cvta.to.global.u64 %rd4355, %rd4346; add.s64 %rd4356, %rd4355, %rd4284; st.global.u32 [%rd4356], %r5035; cvta.to.global.u64 %rd4357, %rd4326; add.s64 %rd4358, %rd4357, %rd1282; ld.global.u32 %r5036, [%rd4358]; cvta.to.global.u64 %rd4359, %rd4333; add.s64 %rd4360, %rd4359, %rd4284; st.global.u32 [%rd4360], %r5036; ld.const.v2.u64 {%rd4361, %rd4362}, [kernel_params+3424]; cvta.to.global.u64 %rd4365, %rd4361; add.s64 %rd4366, %rd4365, %rd4284; or.b32 %r5037, %r1494, 268435456; st.global.u32 [%rd4366], %r5037; ld.const.v2.u64 {%rd4367, %rd4368}, [kernel_params+3376]; cvta.to.global.u64 %rd4371, %rd4367; add.s64 %rd4372, %rd4371, %rd4351; st.global.u16 [%rd4372], %rs294; cvta.to.global.u64 %rd4373, %rd4368; add.s64 %rd4374, %rd4373, %rd4351; st.global.u16 [%rd4374], %rs295; cvta.to.global.u64 %rd4375, %rd4362; add.s64 %rd4377, %rd4375, %rd4275; st.global.f32 [%rd4377], %f24526; st.global.f32 [%rd4377+4], %f24528; st.global.f32 [%rd4377+8], %f24530; setp.eq.s32 %p3315, %r1437, 0; @%p3315 bra $L__BB107_3899; ld.const.u64 %rd4378, [kernel_params+3440]; cvta.to.global.u64 %rd4379, %rd4378; add.s64 %rd4381, %rd4379, %rd4275; st.global.f32 [%rd4381], %f30493; st.global.f32 [%rd4381+4], %f30494; st.global.f32 [%rd4381+8], %f30495; $L__BB107_3899: ld.local.u32 %r1496, [%rd5+-36]; and.b32 %r5039, %r1496, 20; setp.eq.s32 %p3316, %r5039, 0; mov.u32 %r5886, 0; @%p3316 bra $L__BB107_4529; add.s32 %r1497, %r602, 3; ld.const.u32 %r5040, [kernel_params+1596]; setp.eq.s32 %p3317, %r5040, 0; brev.b32 %r5041, %r603; mul.lo.s32 %r5042, %r5041, 1025551850; xor.b32 %r1498, %r5042, %r5041; @%p3317 bra $L__BB107_3903; ld.const.v2.u32 {%r5043, %r5044}, [kernel_params+1600]; mov.b32 %f6190, %r5044; setp.geu.ftz.f32 %p3318, %f6190, 0f3F800000; setp.lt.ftz.f32 %p3319, %f6190, 0f3F800000; ld.const.u32 %r5047, [kernel_params+1576]; selp.b32 %r5048, %r5047, %r601, %p3319; add.s16 %rs584, %rs28, 3; and.b16 %rs585, %rs584, 255; cvt.u32.u16 %r5049, %rs585; xor.b32 %r5050, %r5048, %r5049; mul.lo.s32 %r5051, %r5050, -512718531; shr.u32 %r5052, %r5048, 16; xor.b32 %r5053, %r5051, %r5052; bfe.u32 %r5054, %r5053, 4, 4; shr.u32 %r5055, %r5048, 8; xor.b32 %r5056, %r5053, %r5055; xor.b32 %r5057, %r5056, %r5054; mul.lo.s32 %r5058, %r5057, 153742143; shr.u32 %r5059, %r5048, 23; xor.b32 %r5060, %r5058, %r5059; bfe.u32 %r5061, %r5060, 1, 7; xor.b32 %r5062, %r5061, %r5060; shr.u32 %r5063, %r5048, 27; or.b32 %r5064, %r5063, 1; mul.lo.s32 %r5065, %r5064, %r5062; mul.lo.s32 %r5066, %r5065, -1074439621; bfe.u32 %r5067, %r5066, 2, 6; xor.b32 %r5068, %r5067, %r5066; mul.lo.s32 %r5069, %r5068, -1638916925; bfe.u32 %r5070, %r5069, 2, 6; xor.b32 %r5071, %r5070, %r5069; mul.lo.s32 %r5072, %r5071, -933190689; and.b32 %r5073, %r5072, 255; bfe.u32 %r5074, %r5072, 5, 3; xor.b32 %r5075, %r5074, %r5073; add.s32 %r5076, %r5043, -1; xor.b32 %r1499, %r1497, 61; xor.b32 %r5077, %r5048, %r1499; mul.lo.s32 %r5078, %r5077, 9; shr.u32 %r5079, %r5078, 4; xor.b32 %r5080, %r5079, %r5078; mul.lo.s32 %r5081, %r5080, 668265261; add.s32 %r5082, %r5081, %r1498; shr.u32 %r5083, %r5081, 16; or.b32 %r5084, %r5083, 1; mul.lo.s32 %r5085, %r5084, %r5082; mul.lo.s32 %r5086, %r5085, 89287766; xor.b32 %r5087, %r5086, %r5085; mul.lo.s32 %r5088, %r5087, 1403136100; xor.b32 %r5089, %r5088, %r5087; brev.b32 %r5090, %r5089; neg.s32 %r5091, %r5043; and.b32 %r5092, %r603, %r5091; and.b32 %r5093, %r5090, %r5076; or.b32 %r5094, %r5093, %r5092; mad.lo.s32 %r5095, %r5075, %r5043, %r5094; shl.b32 %r5096, %r5043, 8; rem.u32 %r5097, %r5095, %r5096; ld.const.u64 %rd4382, [kernel_params+2424]; cvta.to.global.u64 %rd4383, %rd4382; shl.b32 %r5098, %r5097, 1; mul.wide.u32 %rd4384, %r5098, 4; add.s64 %rd4385, %rd4383, %rd4384; ld.global.f32 %f6226, [%rd4385]; ld.global.f32 %f6202, [%rd4385+4]; @%p3318 bra $L__BB107_3907; xor.b32 %r5099, %r1499, %r601; mul.lo.s32 %r5100, %r5099, 9; shr.u32 %r5101, %r5100, 4; xor.b32 %r5102, %r5101, %r5100; mul.lo.s32 %r5103, %r5102, 668265261; cvt.rn.f32.u32 %f24531, %r5103; mul.ftz.f32 %f24532, %f24531, 0f2F7FFFFE; xor.b32 %r5104, %r601, %r1497; xor.b32 %r5105, %r5104, -905047700; mul.lo.s32 %r5106, %r5105, 9; shr.u32 %r5107, %r5106, 4; xor.b32 %r5108, %r5107, %r5106; mul.lo.s32 %r5109, %r5108, 668265261; cvt.rn.f32.u32 %f24533, %r5109; mul.ftz.f32 %f24534, %f24533, 0f2F7FFFFE; fma.rn.ftz.f32 %f24535, %f24532, %f6190, %f6226; fma.rn.ftz.f32 %f24536, %f24534, %f6190, %f6202; cvt.rmi.ftz.f32.f32 %f24537, %f24535; sub.ftz.f32 %f6226, %f24535, %f24537; cvt.rmi.ftz.f32.f32 %f24538, %f24536; sub.ftz.f32 %f6202, %f24536, %f24538; bra.uni $L__BB107_3907; $L__BB107_3903: shr.u32 %r5111, %r1497, 16; xor.b32 %r5112, %r5111, %r1497; mul.lo.s32 %r5113, %r5112, 569420461; shr.u32 %r5114, %r5113, 15; xor.b32 %r5115, %r5114, %r5113; mul.lo.s32 %r5116, %r5115, -749064809; shr.u32 %r5117, %r5116, 15; xor.b32 %r5118, %r601, %r5116; xor.b32 %r5119, %r5118, %r5117; xor.b32 %r1500, %r5119, -419546133; xor.b32 %r5120, %r5119, 508809841; add.s32 %r5121, %r5120, %r1498; shr.u32 %r5122, %r5120, 16; or.b32 %r5123, %r5122, 1; mul.lo.s32 %r5124, %r5123, %r5121; mul.lo.s32 %r5125, %r5124, 89287766; xor.b32 %r5126, %r5125, %r5124; mul.lo.s32 %r5127, %r5126, 1403136100; xor.b32 %r1501, %r5127, %r5126; setp.eq.s32 %p3320, %r1501, 0; mov.u32 %r5863, 0; @%p3320 bra $L__BB107_3906; mov.u32 %r5860, 0; mov.u64 %rd4387, sobol_burley_table; mov.u32 %r5861, %r1501; mov.u32 %r5863, %r5860; $L__BB107_3905: clz.b32 %r5130, %r5861; add.s32 %r5131, %r5130, %r5860; mul.wide.u32 %rd4386, %r5131, 4; add.s64 %rd4388, %rd4387, %rd4386; ld.const.u32 %r5132, [%rd4388+128]; xor.b32 %r5863, %r5132, %r5863; add.s32 %r5860, %r5131, 1; shl.b32 %r5133, %r5861, %r5130; shl.b32 %r5861, %r5133, 1; setp.ne.s32 %p3321, %r5861, 0; @%p3321 bra $L__BB107_3905; $L__BB107_3906: xor.b32 %r5134, %r1500, -525684874; brev.b32 %r5135, %r1501; shr.u32 %r5136, %r5134, 16; or.b32 %r5137, %r5136, 1; mul.lo.s32 %r5138, %r5135, 1025551850; xor.b32 %r5139, %r5138, %r5135; add.s32 %r5140, %r5139, %r5134; mul.lo.s32 %r5141, %r5140, %r5137; mul.lo.s32 %r5142, %r5141, 89287766; xor.b32 %r5143, %r5142, %r5141; xor.b32 %r5144, %r1500, -1802089138; mul.lo.s32 %r5145, %r5863, 1025551850; xor.b32 %r5146, %r5145, %r5863; add.s32 %r5147, %r5146, %r5144; shr.u32 %r5148, %r5144, 16; or.b32 %r5149, %r5148, 1; mul.lo.s32 %r5150, %r5147, %r5149; mul.lo.s32 %r5151, %r5150, 89287766; xor.b32 %r5152, %r5151, %r5150; mul.lo.s32 %r5153, %r5152, 1403136100; xor.b32 %r5154, %r5153, %r5152; brev.b32 %r5155, %r5154; cvt.rn.f32.u32 %f24539, %r5155; mul.lo.s32 %r5156, %r5143, 1403136100; xor.b32 %r5157, %r5156, %r5143; brev.b32 %r5158, %r5157; cvt.rn.f32.u32 %f24540, %r5158; mul.ftz.f32 %f6202, %f24539, 0f2F7FFFFE; mul.ftz.f32 %f6226, %f24540, 0f2F7FFFFE; $L__BB107_3907: ld.local.u32 %r1509, [%rd5+172]; setp.lt.s32 %p3322, %r1509, 2; mov.u32 %r5869, 0; @%p3322 bra $L__BB107_3929; add.s32 %r5161, %r1509, -1; and.b32 %r5867, %r1509, 3; setp.lt.u32 %p3323, %r5161, 3; mov.f32 %f30607, 0f00000000; mov.u32 %r5866, 0; @%p3323 bra $L__BB107_3919; sub.s32 %r5865, %r1509, %r5867; mov.f32 %f30607, 0f00000000; mov.u32 %r5866, 0; $L__BB107_3910: cvt.s64.s32 %rd1019, %r5866; mul.wide.s32 %rd4389, %r5866, 80; add.s64 %rd4390, %rd2, %rd4389; add.s64 %rd1020, %rd4390, 316; ld.local.u32 %r5163, [%rd4390+316]; setp.gt.s32 %p3324, %r5163, 34; @%p3324 bra $L__BB107_3912; ld.local.f32 %f24544, [%rd1020+4]; add.ftz.f32 %f30607, %f30607, %f24544; $L__BB107_3912: ld.local.u32 %r5164, [%rd1020+80]; setp.gt.s32 %p3325, %r5164, 34; @%p3325 bra $L__BB107_3914; ld.local.f32 %f24545, [%rd1020+84]; add.ftz.f32 %f30607, %f30607, %f24545; $L__BB107_3914: ld.local.u32 %r5165, [%rd1020+160]; setp.gt.s32 %p3326, %r5165, 34; @%p3326 bra $L__BB107_3916; ld.local.f32 %f24546, [%rd1020+164]; add.ftz.f32 %f30607, %f30607, %f24546; $L__BB107_3916: ld.local.u32 %r5166, [%rd1020+240]; setp.gt.s32 %p3327, %r5166, 34; @%p3327 bra $L__BB107_3918; ld.local.f32 %f24547, [%rd1020+244]; add.ftz.f32 %f30607, %f30607, %f24547; $L__BB107_3918: cvt.u32.u64 %r5167, %rd1019; add.s32 %r5866, %r5167, 4; add.s32 %r5865, %r5865, -4; setp.ne.s32 %p3328, %r5865, 0; @%p3328 bra $L__BB107_3910; $L__BB107_3919: setp.eq.s32 %p3329, %r5867, 0; @%p3329 bra $L__BB107_3924; mul.wide.s32 %rd4391, %r5866, 80; add.s64 %rd4392, %rd2, %rd4391; add.s64 %rd4961, %rd4392, 320; $L__BB107_3921: .pragma "nounroll"; ld.local.u32 %r5168, [%rd4961+-4]; setp.gt.s32 %p3330, %r5168, 34; @%p3330 bra $L__BB107_3923; ld.local.f32 %f24548, [%rd4961]; add.ftz.f32 %f30607, %f30607, %f24548; $L__BB107_3923: add.s64 %rd4961, %rd4961, 80; add.s32 %r5867, %r5867, -1; setp.ne.s32 %p3331, %r5867, 0; @%p3331 bra $L__BB107_3921; $L__BB107_3924: mul.ftz.f32 %f6220, %f6226, %f30607; mov.f32 %f30616, 0f00000000; mov.u32 %r5869, 0; mov.u32 %r5868, %r5869; mov.f32 %f30617, %f30616; $L__BB107_3925: mul.wide.s32 %rd4393, %r5868, 80; add.s64 %rd4394, %rd2, %rd4393; add.s64 %rd1025, %rd4394, 316; ld.local.u32 %r5170, [%rd4394+316]; setp.gt.s32 %p3332, %r5170, 34; @%p3332 bra $L__BB107_3927; ld.local.f32 %f6222, [%rd1025+4]; add.ftz.f32 %f30617, %f30616, %f6222; setp.lt.ftz.f32 %p3333, %f6220, %f30617; @%p3333 bra $L__BB107_3928; $L__BB107_3927: add.s32 %r5868, %r5868, 1; setp.lt.s32 %p3334, %r5868, %r1509; mov.f32 %f30616, %f30617; @%p3334 bra $L__BB107_3925; bra.uni $L__BB107_3929; $L__BB107_3928: sub.ftz.f32 %f24550, %f6220, %f30616; div.approx.ftz.f32 %f6226, %f24550, %f6222; mov.u32 %r5869, %r5868; $L__BB107_3929: cvt.s64.s32 %rd1026, %r5869; mul.wide.s32 %rd4395, %r5869, 80; add.s64 %rd4396, %rd2, %rd4395; add.s64 %rd1027, %rd4396, 316; ld.local.u32 %r1522, [%rd4396+316]; add.s32 %r5172, %r1522, -32; setp.lt.u32 %p3335, %r5172, 3; @%p3335 bra $L__BB107_4507; bra.uni $L__BB107_3930; $L__BB107_4507: ld.const.u64 %rd4713, [kernel_params+2632]; cvta.to.global.u64 %rd4712, %rd4713; add.u64 %rd4583, %SPL, 3424; ld.local.v4.f32 {%f28672, %f28673, %f28674, %f28675}, [%rd4583]; mul.lo.s64 %rd4584, %rd3, 12; add.s64 %rd4585, %rd4712, %rd4584; st.global.f32 [%rd4585], %f28672; st.global.f32 [%rd4585+4], %f28673; st.global.f32 [%rd4585+8], %f28674; ld.const.v2.u64 {%rd4586, %rd4587}, [kernel_params+2640]; cvta.to.global.u64 %rd4590, %rd4586; ld.local.f32 %f28679, [%rd1027+8]; ld.local.f32 %f28680, [%rd1027+12]; ld.local.f32 %f28681, [%rd1027+16]; add.s64 %rd4591, %rd4590, %rd4584; st.global.f32 [%rd4591], %f28679; st.global.f32 [%rd4591+4], %f28680; st.global.f32 [%rd4591+8], %f28681; cvta.to.global.u64 %rd4592, %rd4587; add.s64 %rd4594, %rd4592, %rd1282; mov.u32 %r5483, 0; st.global.u32 [%rd4594], %r5483; ld.const.u64 %rd4595, [kernel_params+2656]; cvta.to.global.u64 %rd4596, %rd4595; add.s64 %rd4597, %rd4596, %rd1282; mov.u32 %r5484, 2139095039; st.global.u32 [%rd4597], %r5484; ld.local.f32 %f28682, [%rd5+4]; add.s64 %rd4603, %rd1101, %rd1282; st.global.f32 [%rd4603], %f28682; add.s64 %rd4605, %rd1103, %rd1282; st.global.u32 [%rd4605], %r5483; ld.const.u64 %rd4606, [kernel_params+2760]; cvta.to.global.u64 %rd4607, %rd4606; ld.local.v2.f32 {%f28683, %f28684}, [%rd5+-64]; ld.local.f32 %f28687, [%rd5+-56]; add.s64 %rd4608, %rd4607, %rd4584; st.global.f32 [%rd4608], %f28683; st.global.f32 [%rd4608+4], %f28684; st.global.f32 [%rd4608+8], %f28687; ld.global.u32 %r5485, [%rd27]; and.b32 %r5486, %r5485, -2; setp.eq.s32 %p3882, %r1522, 32; selp.b32 %r5487, 2097152, 1048576, %p3882; or.b32 %r1573, %r5487, %r5486; ld.local.f32 %f30953, [%rd1027+-12]; ld.local.f32 %f30954, [%rd1027+-8]; ld.local.f32 %f30955, [%rd1027+-4]; @%p3322 bra $L__BB107_4525; add.s32 %r5489, %r1509, -1; and.b32 %r5885, %r1509, 3; setp.lt.u32 %p3883, %r5489, 3; mov.f32 %f30944, 0f00000000; mov.u32 %r5884, %r5483; @%p3883 bra $L__BB107_4519; sub.s32 %r5883, %r1509, %r5885; mov.f32 %f30944, 0f00000000; mov.u32 %r5884, 0; $L__BB107_4510: cvt.s64.s32 %rd1037, %r5884; mul.wide.s32 %rd4611, %r5884, 80; add.s64 %rd4612, %rd2, %rd4611; add.s64 %rd1038, %rd4612, 316; ld.local.u32 %r5491, [%rd4612+316]; setp.gt.s32 %p3884, %r5491, 34; @%p3884 bra $L__BB107_4512; ld.local.f32 %f28691, [%rd1038+4]; add.ftz.f32 %f30944, %f30944, %f28691; $L__BB107_4512: ld.local.u32 %r5492, [%rd1038+80]; setp.gt.s32 %p3885, %r5492, 34; @%p3885 bra $L__BB107_4514; ld.local.f32 %f28692, [%rd1038+84]; add.ftz.f32 %f30944, %f30944, %f28692; $L__BB107_4514: ld.local.u32 %r5493, [%rd1038+160]; setp.gt.s32 %p3886, %r5493, 34; @%p3886 bra $L__BB107_4516; ld.local.f32 %f28693, [%rd1038+164]; add.ftz.f32 %f30944, %f30944, %f28693; $L__BB107_4516: ld.local.u32 %r5494, [%rd1038+240]; setp.gt.s32 %p3887, %r5494, 34; @%p3887 bra $L__BB107_4518; ld.local.f32 %f28694, [%rd1038+244]; add.ftz.f32 %f30944, %f30944, %f28694; $L__BB107_4518: cvt.u32.u64 %r5495, %rd1037; add.s32 %r5884, %r5495, 4; add.s32 %r5883, %r5883, -4; setp.ne.s32 %p3888, %r5883, 0; @%p3888 bra $L__BB107_4510; $L__BB107_4519: setp.eq.s32 %p3889, %r5885, 0; @%p3889 bra $L__BB107_4524; mul.wide.s32 %rd4613, %r5884, 80; add.s64 %rd4614, %rd2, %rd4613; add.s64 %rd4962, %rd4614, 320; $L__BB107_4521: .pragma "nounroll"; add.s64 %rd1041, %rd4962, -4; ld.local.u32 %r5496, [%rd4962+-4]; setp.gt.s32 %p3890, %r5496, 34; @%p3890 bra $L__BB107_4523; ld.local.f32 %f28695, [%rd1041+4]; add.ftz.f32 %f30944, %f30944, %f28695; $L__BB107_4523: add.s64 %rd4962, %rd4962, 80; add.s32 %r5885, %r5885, -1; setp.ne.s32 %p3891, %r5885, 0; @%p3891 bra $L__BB107_4521; $L__BB107_4524: ld.local.f32 %f28696, [%rd1027+4]; div.approx.ftz.f32 %f28697, %f30944, %f28696; mul.ftz.f32 %f30953, %f30953, %f28697; mul.ftz.f32 %f30954, %f30954, %f28697; mul.ftz.f32 %f30955, %f30955, %f28697; $L__BB107_4525: ld.local.f32 %f28698, [%rd1027+44]; setp.neu.ftz.f32 %p3892, %f28698, 0f7F7FFFFF; or.b32 %r5497, %r1573, 4194304; selp.b32 %r5498, %r5497, %r1573, %p3892; shl.b32 %r5499, %r1496, 23; and.b32 %r5500, %r5499, 8388608; or.b32 %r5501, %r5500, %r5498; ld.const.u64 %rd4615, [kernel_params+2584]; cvta.to.global.u64 %rd4616, %rd4615; add.s64 %rd4618, %rd4616, %rd4584; ld.global.f32 %f28699, [%rd4618]; mul.ftz.f32 %f28700, %f30953, %f28699; ld.global.f32 %f28701, [%rd4618+4]; mul.ftz.f32 %f28702, %f30954, %f28701; ld.global.f32 %f28703, [%rd4618+8]; mul.ftz.f32 %f28704, %f30955, %f28703; st.global.f32 [%rd4618], %f28700; st.global.f32 [%rd4618+4], %f28702; st.global.f32 [%rd4618+8], %f28704; st.global.u32 [%rd27], %r5501; ld.const.u64 %rd4619, [kernel_params+2536]; cvta.to.global.u64 %rd4620, %rd4619; add.s64 %rd4622, %rd4620, %rd2151; ld.global.u16 %rs605, [%rd4622]; add.s16 %rs606, %rs605, 16; st.global.u16 [%rd4622], %rs606; and.b32 %r5502, %r1436, 2097152; setp.eq.s32 %p3893, %r5502, 0; @%p3893 bra $L__BB107_4528; ld.const.u64 %rd4623, [kernel_params+2464]; cvta.to.global.u64 %rd4624, %rd4623; add.s64 %rd4626, %rd4624, %rd2151; ld.global.u16 %rs607, [%rd4626]; setp.ne.s16 %p3894, %rs607, 0; @%p3894 bra $L__BB107_4528; ld.const.u64 %rd4627, [kernel_params+2600]; cvta.to.global.u64 %rd4628, %rd4627; add.s64 %rd4630, %rd4628, %rd4584; mov.u32 %r5503, 1065353216; st.global.u32 [%rd4630], %r5503; st.global.u32 [%rd4630+4], %r5503; st.global.u32 [%rd4630+8], %r5503; ld.const.u64 %rd4631, [kernel_params+2608]; cvta.to.global.u64 %rd4632, %rd4631; add.s64 %rd4633, %rd4632, %rd4584; st.global.u32 [%rd4633], %r5483; st.global.u32 [%rd4633+4], %r5483; st.global.u32 [%rd4633+8], %r5483; $L__BB107_4528: ld.const.v2.u64 {%rd4634, %rd4635}, [kernel_params+2736]; cvta.to.global.u64 %rd4638, %rd4634; ld.local.f32 %f28705, [%rd1027+32]; ld.local.f32 %f28706, [%rd1027+36]; ld.local.f32 %f28707, [%rd1027+40]; add.s64 %rd4640, %rd4638, %rd4584; st.global.f32 [%rd4640], %f28705; st.global.f32 [%rd4640+4], %f28706; st.global.f32 [%rd4640+8], %f28707; cvta.to.global.u64 %rd4641, %rd4635; ld.local.f32 %f28708, [%rd1027+20]; ld.local.f32 %f28709, [%rd1027+24]; ld.local.f32 %f28710, [%rd1027+28]; add.s64 %rd4642, %rd4641, %rd4584; st.global.f32 [%rd4642], %f28708; st.global.f32 [%rd4642+4], %f28709; st.global.f32 [%rd4642+8], %f28710; ld.local.f32 %f28711, [%rd1027+48]; ld.const.u64 %rd4643, [kernel_params+2752]; cvta.to.global.u64 %rd4644, %rd4643; add.s64 %rd4646, %rd4644, %rd1282; st.global.f32 [%rd4646], %f28711; mov.u32 %r5886, 256; bra.uni $L__BB107_4529; $L__BB107_3930: ld.local.u8 %rs586, [%rd5+-24]; and.b16 %rs587, %rs586, 6; setp.eq.s16 %p3336, %rs587, 0; @%p3336 bra $L__BB107_3932; ld.local.f32 %f6235, [%rd1027+16]; ld.local.f32 %f6234, [%rd1027+12]; ld.local.f32 %f6233, [%rd1027+8]; bra.uni $L__BB107_3933; $L__BB107_3932: ld.local.v2.f32 {%f6233, %f6234}, [%rd5+-64]; ld.local.f32 %f6235, [%rd5+-56]; $L__BB107_3933: mov.u32 %r5877, 0; mov.f32 %f30902, 0f00000000; setp.gt.s32 %p3337, %r1522, 15; @%p3337 bra $L__BB107_3960; setp.gt.s32 %p3353, %r1522, 9; @%p3353 bra $L__BB107_3957; setp.gt.s32 %p3357, %r1522, 5; @%p3357 bra $L__BB107_3943; bra.uni $L__BB107_3936; $L__BB107_3943: setp.gt.s32 %p3358, %r1522, 7; @%p3358 bra $L__BB107_3952; setp.eq.s32 %p3361, %r1522, 6; @%p3361 bra $L__BB107_4418; setp.eq.s32 %p3362, %r1522, 7; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3362 bra $L__BB107_3946; bra.uni $L__BB107_4433; $L__BB107_3946: ld.local.f32 %f28107, [%rd1027+20]; mul.ftz.f32 %f7474, %f28107, 0f3FC90FDB; mov.f32 %f28108, 0f3FC90FDB; ld.local.f32 %f28109, [%rd1027+24]; mul.ftz.f32 %f7475, %f28109, 0f3FC90FDB; add.ftz.f32 %f7476, %f7474, %f7475; min.ftz.f32 %f7477, %f7476, %f28108; mul.ftz.f32 %f7478, %f6226, %f7477; setp.leu.ftz.f32 %p3764, %f7477, 0f00000000; mov.u32 %r5877, 6; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3764 bra $L__BB107_4433; cos.approx.ftz.f32 %f28114, %f7477; mul.ftz.f32 %f28115, %f6226, %f28114; sub.ftz.f32 %f28116, %f28114, %f28115; add.ftz.f32 %f28117, %f6226, %f28116; mul.ftz.f32 %f28118, %f28117, %f28117; mov.f32 %f30892, 0f3F800000; sub.ftz.f32 %f28120, %f30892, %f28118; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28121, %f28120, %f30902; sqrt.approx.ftz.f32 %f28122, %f28121; mul.ftz.f32 %f28123, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f28124, %f28123; mul.ftz.f32 %f28125, %f28122, %f28124; sin.approx.ftz.f32 %f28126, %f28123; mul.ftz.f32 %f28127, %f28122, %f28126; ld.local.f32 %f28128, [%rd1027+12]; ld.local.f32 %f28129, [%rd1027+8]; setp.neu.ftz.f32 %p3765, %f28129, %f28128; ld.local.f32 %f28130, [%rd1027+16]; setp.neu.ftz.f32 %p3766, %f28129, %f28130; or.pred %p3767, %p3765, %p3766; sub.ftz.f32 %f28131, %f28130, %f28128; neg.ftz.f32 %f28132, %f28128; neg.ftz.f32 %f28133, %f28130; selp.f32 %f28134, %f28133, %f28130, %p3767; add.ftz.f32 %f28135, %f28129, %f28134; selp.f32 %f28136, %f28128, %f28132, %p3767; sub.ftz.f32 %f28137, %f28136, %f28129; mul.ftz.f32 %f28138, %f28135, %f28135; fma.rn.ftz.f32 %f28139, %f28131, %f28131, %f28138; fma.rn.ftz.f32 %f28140, %f28137, %f28137, %f28139; sqrt.approx.ftz.f32 %f28141, %f28140; div.approx.ftz.f32 %f28142, %f28131, %f28141; div.approx.ftz.f32 %f28143, %f28135, %f28141; div.approx.ftz.f32 %f28144, %f28137, %f28141; mul.ftz.f32 %f28145, %f28128, %f28144; mul.ftz.f32 %f28146, %f28130, %f28143; sub.ftz.f32 %f28147, %f28145, %f28146; mul.ftz.f32 %f28148, %f28130, %f28142; mul.ftz.f32 %f28149, %f28129, %f28144; sub.ftz.f32 %f28150, %f28148, %f28149; mul.ftz.f32 %f28151, %f28129, %f28143; mul.ftz.f32 %f28152, %f28128, %f28142; sub.ftz.f32 %f28153, %f28151, %f28152; mul.ftz.f32 %f28154, %f28127, %f28147; mul.ftz.f32 %f28155, %f28127, %f28150; mul.ftz.f32 %f28156, %f28127, %f28153; fma.rn.ftz.f32 %f28157, %f28125, %f28142, %f28154; fma.rn.ftz.f32 %f28158, %f28125, %f28143, %f28155; fma.rn.ftz.f32 %f28159, %f28125, %f28144, %f28156; fma.rn.ftz.f32 %f30899, %f28129, %f28117, %f28157; fma.rn.ftz.f32 %f30900, %f28128, %f28117, %f28158; fma.rn.ftz.f32 %f30901, %f28130, %f28117, %f28159; sub.ftz.f32 %f28160, %f30892, %f28114; mov.f32 %f28161, 0f3E22F983; div.approx.ftz.f32 %f7482, %f28161, %f28160; mul.ftz.f32 %f28162, %f6234, %f30900; fma.rn.ftz.f32 %f28163, %f6233, %f30899, %f28162; fma.rn.ftz.f32 %f28164, %f6235, %f30901, %f28163; setp.leu.ftz.f32 %p3768, %f28164, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3768 bra $L__BB107_4433; setp.lt.ftz.f32 %p3769, %f7478, %f7474; @%p3769 bra $L__BB107_3951; setp.leu.ftz.f32 %p3770, %f7476, %f7478; setp.eq.ftz.f32 %p3771, %f7475, 0f00000000; mov.f32 %f30892, 0f00000000; or.pred %p3772, %p3771, %p3770; @%p3772 bra $L__BB107_3951; sub.ftz.f32 %f28167, %f7478, %f7474; div.approx.ftz.f32 %f28168, %f28167, %f7475; mov.f32 %f28169, 0f3F800000; sub.ftz.f32 %f30892, %f28169, %f28168; $L__BB107_3951: mul.ftz.f32 %f30903, %f7482, %f30892; mov.f32 %f30902, %f7482; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_3960: setp.gt.s32 %p3338, %r1522, 22; @%p3338 bra $L__BB107_3977; setp.gt.s32 %p3346, %r1522, 19; @%p3346 bra $L__BB107_3972; setp.eq.s32 %p3350, %r1522, 16; @%p3350 bra $L__BB107_4237; setp.eq.s32 %p3351, %r1522, 17; @%p3351 bra $L__BB107_4232; setp.eq.s32 %p3352, %r1522, 19; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3352 bra $L__BB107_3965; bra.uni $L__BB107_4433; $L__BB107_3965: ld.local.v2.f32 {%f26207, %f26208}, [%rd5+-48]; ld.local.f32 %f26209, [%rd1027+20]; mul.ftz.f32 %f6881, %f26209, 0f3FC90FDB; ld.local.f32 %f26210, [%rd1027+24]; mul.ftz.f32 %f6882, %f26210, 0f3FC90FDB; ld.local.f32 %f6883, [%rd1027+8]; ld.local.f32 %f6884, [%rd5+-52]; ld.local.f32 %f6885, [%rd1027+12]; mul.ftz.f32 %f26211, %f26207, %f6885; fma.rn.ftz.f32 %f26212, %f6884, %f6883, %f26211; ld.local.f32 %f6886, [%rd1027+16]; fma.rn.ftz.f32 %f6887, %f26208, %f6886, %f26212; setp.leu.ftz.f32 %p3597, %f6887, 0f00000000; mov.u32 %r5877, 10; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3597 bra $L__BB107_4433; add.ftz.f32 %f26217, %f6887, %f6887; mul.ftz.f32 %f26218, %f26217, %f6883; mul.ftz.f32 %f26219, %f26217, %f6885; mul.ftz.f32 %f26220, %f26217, %f6886; sub.ftz.f32 %f26221, %f26218, %f6884; sub.ftz.f32 %f26222, %f26219, %f26207; sub.ftz.f32 %f26223, %f26220, %f26208; add.ftz.f32 %f6888, %f6881, %f6882; mov.f32 %f26224, 0f3FC90FDB; min.ftz.f32 %f26225, %f6888, %f26224; mul.ftz.f32 %f6889, %f6226, %f26225; cos.approx.ftz.f32 %f26226, %f26225; mul.ftz.f32 %f26227, %f6226, %f26226; sub.ftz.f32 %f26228, %f26226, %f26227; add.ftz.f32 %f26229, %f6226, %f26228; mul.ftz.f32 %f26230, %f26229, %f26229; mov.f32 %f30730, 0f3F800000; sub.ftz.f32 %f26232, %f30730, %f26230; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f26233, %f26232, %f30902; sqrt.approx.ftz.f32 %f26234, %f26233; mul.ftz.f32 %f26235, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f26236, %f26235; mul.ftz.f32 %f26237, %f26234, %f26236; sin.approx.ftz.f32 %f26238, %f26235; mul.ftz.f32 %f26239, %f26234, %f26238; setp.neu.ftz.f32 %p3598, %f26221, %f26222; setp.neu.ftz.f32 %p3599, %f26221, %f26223; or.pred %p3600, %p3598, %p3599; sub.ftz.f32 %f26240, %f26223, %f26222; neg.ftz.f32 %f26241, %f26222; neg.ftz.f32 %f26242, %f26223; selp.f32 %f26243, %f26242, %f26223, %p3600; add.ftz.f32 %f26244, %f26221, %f26243; selp.f32 %f26245, %f26222, %f26241, %p3600; sub.ftz.f32 %f26246, %f26245, %f26221; mul.ftz.f32 %f26247, %f26244, %f26244; fma.rn.ftz.f32 %f26248, %f26240, %f26240, %f26247; fma.rn.ftz.f32 %f26249, %f26246, %f26246, %f26248; sqrt.approx.ftz.f32 %f26250, %f26249; div.approx.ftz.f32 %f26251, %f26240, %f26250; div.approx.ftz.f32 %f26252, %f26244, %f26250; div.approx.ftz.f32 %f26253, %f26246, %f26250; mul.ftz.f32 %f26254, %f26222, %f26253; mul.ftz.f32 %f26255, %f26223, %f26252; sub.ftz.f32 %f26256, %f26254, %f26255; mul.ftz.f32 %f26257, %f26223, %f26251; mul.ftz.f32 %f26258, %f26221, %f26253; sub.ftz.f32 %f26259, %f26257, %f26258; mul.ftz.f32 %f26260, %f26221, %f26252; mul.ftz.f32 %f26261, %f26222, %f26251; sub.ftz.f32 %f26262, %f26260, %f26261; mul.ftz.f32 %f26263, %f26239, %f26256; mul.ftz.f32 %f26264, %f26239, %f26259; mul.ftz.f32 %f26265, %f26239, %f26262; fma.rn.ftz.f32 %f26266, %f26237, %f26251, %f26263; fma.rn.ftz.f32 %f26267, %f26237, %f26252, %f26264; fma.rn.ftz.f32 %f26268, %f26237, %f26253, %f26265; fma.rn.ftz.f32 %f30899, %f26221, %f26229, %f26266; fma.rn.ftz.f32 %f30900, %f26222, %f26229, %f26267; fma.rn.ftz.f32 %f30901, %f26223, %f26229, %f26268; sub.ftz.f32 %f26269, %f30730, %f26226; mov.f32 %f26270, 0f3E22F983; div.approx.ftz.f32 %f6893, %f26270, %f26269; mul.ftz.f32 %f26271, %f6234, %f30900; fma.rn.ftz.f32 %f26272, %f6233, %f30899, %f26271; fma.rn.ftz.f32 %f26273, %f6235, %f30901, %f26272; setp.leu.ftz.f32 %p3601, %f26273, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3601 bra $L__BB107_4433; mul.ftz.f32 %f26278, %f30900, %f6885; fma.rn.ftz.f32 %f26279, %f30899, %f6883, %f26278; fma.rn.ftz.f32 %f26280, %f30901, %f6886, %f26279; setp.leu.ftz.f32 %p3602, %f26280, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3602 bra $L__BB107_4433; setp.lt.ftz.f32 %p3603, %f6889, %f6881; @%p3603 bra $L__BB107_3971; setp.leu.ftz.f32 %p3604, %f6888, %f6889; setp.eq.ftz.f32 %p3605, %f6882, 0f00000000; mov.f32 %f30730, 0f00000000; or.pred %p3606, %p3605, %p3604; @%p3606 bra $L__BB107_3971; sub.ftz.f32 %f26283, %f6889, %f6881; div.approx.ftz.f32 %f26284, %f26283, %f6882; mov.f32 %f26285, 0f3F800000; sub.ftz.f32 %f30730, %f26285, %f26284; $L__BB107_3971: mul.ftz.f32 %f30903, %f6893, %f30730; mov.f32 %f30902, %f6893; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_3957: add.s32 %r5174, %r1522, -10; setp.lt.u32 %p3354, %r5174, 3; @%p3354 bra $L__BB107_4370; add.s32 %r5175, %r1522, -14; setp.lt.u32 %p3355, %r5175, 2; @%p3355 bra $L__BB107_4267; bra.uni $L__BB107_3959; $L__BB107_4267: ld.local.v2.f32 {%f26719, %f26720}, [%rd5+-48]; ld.local.f32 %f7000, [%rd1027+8]; ld.local.f32 %f7001, [%rd5+-52]; ld.local.f32 %f7002, [%rd1027+12]; mul.ftz.f32 %f26721, %f26719, %f7002; fma.rn.ftz.f32 %f26722, %f7001, %f7000, %f26721; ld.local.f32 %f7003, [%rd1027+16]; fma.rn.ftz.f32 %f7004, %f26720, %f7003, %f26722; setp.le.ftz.f32 %p3642, %f7004, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3642 bra $L__BB107_4433; ld.local.v2.f32 {%f26723, %f26724}, [%rd1027+20]; mul.ftz.f32 %f26725, %f26723, %f26724; setp.lt.ftz.f32 %p3643, %f26725, 0f33D6BF95; @%p3643 bra $L__BB107_4320; bra.uni $L__BB107_4269; $L__BB107_4320: add.ftz.f32 %f27056, %f7004, %f7004; mul.ftz.f32 %f27057, %f7000, %f27056; mul.ftz.f32 %f27058, %f7002, %f27056; mul.ftz.f32 %f27059, %f7003, %f27056; sub.ftz.f32 %f30899, %f27057, %f7001; sub.ftz.f32 %f30900, %f27058, %f26719; sub.ftz.f32 %f30901, %f27059, %f26720; mul.ftz.f32 %f27060, %f6234, %f30900; fma.rn.ftz.f32 %f27061, %f6233, %f30899, %f27060; fma.rn.ftz.f32 %f27062, %f6235, %f30901, %f27061; setp.le.ftz.f32 %p3690, %f27062, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3690 bra $L__BB107_4433; mov.u32 %r5877, 18; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_3977: setp.gt.s32 %p3339, %r1522, 28; @%p3339 bra $L__BB107_3994; setp.eq.s32 %p3343, %r1522, 23; @%p3343 bra $L__BB107_4370; setp.eq.s32 %p3344, %r1522, 24; @%p3344 bra $L__BB107_3981; setp.eq.s32 %p3345, %r1522, 27; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3345 bra $L__BB107_3981; bra.uni $L__BB107_4433; $L__BB107_3981: ld.local.v2.f32 {%f25507, %f25508}, [%rd5+-48]; ld.local.v2.f32 {%f25509, %f25510}, [%rd1027+20]; mul.ftz.f32 %f25511, %f25509, %f25510; setp.lt.ftz.f32 %p3510, %f25511, 0f33D6BF95; ld.local.f32 %f30710, [%rd1027+8]; ld.local.f32 %f6624, [%rd5+-52]; ld.local.f32 %f30711, [%rd1027+12]; mul.ftz.f32 %f25512, %f25507, %f30711; fma.rn.ftz.f32 %f25513, %f6624, %f30710, %f25512; ld.local.f32 %f30712, [%rd1027+16]; fma.rn.ftz.f32 %f30713, %f25508, %f30712, %f25513; @%p3510 bra $L__BB107_4209; bra.uni $L__BB107_3982; $L__BB107_4209: ld.local.f32 %f6803, [%rd1027+28]; setp.gt.ftz.f32 %p3568, %f30713, 0f00000000; @%p3568 bra $L__BB107_4211; bra.uni $L__BB107_4210; $L__BB107_4211: rcp.approx.ftz.f32 %f30714, %f6803; bra.uni $L__BB107_4212; $L__BB107_4370: ld.local.v2.f32 {%f27618, %f27619}, [%rd5+-48]; ld.local.v2.f32 {%f27620, %f27621}, [%rd1027+20]; ld.local.f32 %f7325, [%rd1027+8]; ld.local.f32 %f7326, [%rd5+-52]; ld.local.f32 %f7327, [%rd1027+12]; mul.ftz.f32 %f27622, %f27618, %f7327; fma.rn.ftz.f32 %f27623, %f7326, %f7325, %f27622; ld.local.f32 %f7328, [%rd1027+16]; fma.rn.ftz.f32 %f7329, %f27619, %f7328, %f27623; setp.gt.ftz.f32 %p3725, %f7329, 0f00000000; @%p3725 bra $L__BB107_4372; bra.uni $L__BB107_4371; $L__BB107_4372: setp.eq.ftz.f32 %p3727, %f27620, %f27621; @%p3727 bra $L__BB107_4374; bra.uni $L__BB107_4373; $L__BB107_4374: setp.neu.ftz.f32 %p3728, %f7325, %f7327; setp.neu.ftz.f32 %p3729, %f7325, %f7328; or.pred %p3730, %p3728, %p3729; neg.ftz.f32 %f27651, %f7328; selp.f32 %f27652, %f27651, %f7328, %p3730; add.ftz.f32 %f27653, %f7325, %f27652; neg.ftz.f32 %f27654, %f7327; selp.f32 %f27655, %f7327, %f27654, %p3730; sub.ftz.f32 %f27656, %f27655, %f7325; sub.ftz.f32 %f27657, %f7328, %f7327; mul.ftz.f32 %f27658, %f27653, %f27653; fma.rn.ftz.f32 %f27659, %f27657, %f27657, %f27658; fma.rn.ftz.f32 %f27660, %f27656, %f27656, %f27659; sqrt.approx.ftz.f32 %f27661, %f27660; div.approx.ftz.f32 %f30855, %f27657, %f27661; div.approx.ftz.f32 %f30856, %f27653, %f27661; div.approx.ftz.f32 %f30857, %f27656, %f27661; mul.ftz.f32 %f27662, %f7327, %f30857; mul.ftz.f32 %f27663, %f7328, %f30856; sub.ftz.f32 %f30852, %f27662, %f27663; mul.ftz.f32 %f27664, %f7328, %f30855; mul.ftz.f32 %f27665, %f7325, %f30857; sub.ftz.f32 %f30853, %f27664, %f27665; mul.ftz.f32 %f27666, %f7325, %f30856; mul.ftz.f32 %f27667, %f7327, %f30855; sub.ftz.f32 %f30854, %f27666, %f27667; bra.uni $L__BB107_4375; $L__BB107_3972: setp.eq.s32 %p3347, %r1522, 20; @%p3347 bra $L__BB107_4221; setp.eq.s32 %p3348, %r1522, 21; @%p3348 bra $L__BB107_4215; setp.eq.s32 %p3349, %r1522, 22; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3349 bra $L__BB107_3975; bra.uni $L__BB107_4433; $L__BB107_3994: setp.eq.s32 %p3340, %r1522, 29; @%p3340 bra $L__BB107_4014; setp.eq.s32 %p3341, %r1522, 30; @%p3341 bra $L__BB107_3998; setp.ne.s32 %p3342, %r1522, 31; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3342 bra $L__BB107_4433; ld.local.v2.f32 {%f24562, %f24563}, [%rd5+-48]; ld.local.f32 %f24566, [%rd5+-52]; neg.ftz.f32 %f30899, %f24566; neg.ftz.f32 %f30900, %f24562; neg.ftz.f32 %f30901, %f24563; mov.u32 %r5877, 33; mov.f32 %f30902, 0f3F800000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_3936: setp.eq.s32 %p3363, %r1522, 2; @%p3363 bra $L__BB107_4431; setp.eq.s32 %p3364, %r1522, 3; @%p3364 bra $L__BB107_4427; setp.eq.s32 %p3365, %r1522, 5; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3365 bra $L__BB107_3939; bra.uni $L__BB107_4433; $L__BB107_3939: ld.local.f32 %f7508, [%rd5+-52]; ld.local.v2.f32 {%f28254, %f28255}, [%rd5+-48]; mul.ftz.f32 %f28256, %f6226, 0f40C90FDB; cos.approx.ftz.f32 %f28257, %f28256; sqrt.approx.ftz.f32 %f28258, %f6202; mul.ftz.f32 %f28259, %f28258, %f28257; sin.approx.ftz.f32 %f28260, %f28256; mul.ftz.f32 %f28261, %f28258, %f28260; mul.ftz.f32 %f28262, %f28259, %f28259; mov.f32 %f28263, 0f3F800000; sub.ftz.f32 %f28264, %f28263, %f28262; mul.ftz.f32 %f28265, %f28261, %f28261; sub.ftz.f32 %f28266, %f28264, %f28265; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28267, %f28266, %f30902; sqrt.approx.ftz.f32 %f7511, %f28267; ld.local.f32 %f7512, [%rd1027+12]; ld.local.f32 %f7513, [%rd1027+8]; setp.neu.ftz.f32 %p3780, %f7513, %f7512; ld.local.f32 %f7514, [%rd1027+16]; setp.neu.ftz.f32 %p3781, %f7513, %f7514; or.pred %p3782, %p3780, %p3781; sub.ftz.f32 %f28268, %f7514, %f7512; neg.ftz.f32 %f28269, %f7512; neg.ftz.f32 %f28270, %f7514; selp.f32 %f28271, %f28270, %f7514, %p3782; add.ftz.f32 %f28272, %f7513, %f28271; selp.f32 %f28273, %f7512, %f28269, %p3782; sub.ftz.f32 %f28274, %f28273, %f7513; mul.ftz.f32 %f28275, %f28272, %f28272; fma.rn.ftz.f32 %f28276, %f28268, %f28268, %f28275; fma.rn.ftz.f32 %f28277, %f28274, %f28274, %f28276; sqrt.approx.ftz.f32 %f28278, %f28277; div.approx.ftz.f32 %f28279, %f28268, %f28278; div.approx.ftz.f32 %f28280, %f28272, %f28278; div.approx.ftz.f32 %f28281, %f28274, %f28278; mul.ftz.f32 %f28282, %f7512, %f28281; mul.ftz.f32 %f28283, %f7514, %f28280; sub.ftz.f32 %f28284, %f28282, %f28283; mul.ftz.f32 %f28285, %f7514, %f28279; mul.ftz.f32 %f28286, %f7513, %f28281; sub.ftz.f32 %f28287, %f28285, %f28286; mul.ftz.f32 %f28288, %f7513, %f28280; mul.ftz.f32 %f28289, %f7512, %f28279; sub.ftz.f32 %f28290, %f28288, %f28289; mul.ftz.f32 %f28291, %f28261, %f28284; mul.ftz.f32 %f28292, %f28261, %f28287; mul.ftz.f32 %f28293, %f28261, %f28290; fma.rn.ftz.f32 %f28294, %f28259, %f28279, %f28291; fma.rn.ftz.f32 %f28295, %f28259, %f28280, %f28292; fma.rn.ftz.f32 %f28296, %f28259, %f28281, %f28293; fma.rn.ftz.f32 %f30899, %f7513, %f7511, %f28294; fma.rn.ftz.f32 %f30900, %f7512, %f7511, %f28295; fma.rn.ftz.f32 %f30901, %f7514, %f7511, %f28296; mul.ftz.f32 %f28297, %f6234, %f30900; fma.rn.ftz.f32 %f28298, %f6233, %f30899, %f28297; fma.rn.ftz.f32 %f28299, %f6235, %f30901, %f28298; setp.leu.ftz.f32 %p3783, %f28299, 0f00000000; mov.u32 %r5877, 6; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3783 bra $L__BB107_4433; mul.ftz.f32 %f28303, %f7512, %f30900; fma.rn.ftz.f32 %f28304, %f7513, %f30899, %f28303; fma.rn.ftz.f32 %f7518, %f7514, %f30901, %f28304; setp.le.ftz.f32 %p3784, %f7518, 0f00000000; mov.f32 %f30903, 0f00000000; @%p3784 bra $L__BB107_4426; mul.ftz.f32 %f28305, %f28254, %f7512; fma.rn.ftz.f32 %f28306, %f7508, %f7513, %f28305; fma.rn.ftz.f32 %f28307, %f28255, %f7514, %f28306; sub.ftz.f32 %f28309, %f28263, %f28307; mov.f32 %f28310, 0f00000000; max.ftz.f32 %f28311, %f28309, %f28310; min.ftz.f32 %f28312, %f28311, %f28263; mul.ftz.f32 %f28313, %f28312, %f28312; mul.ftz.f32 %f28314, %f28313, %f28313; mul.ftz.f32 %f7519, %f28312, %f28314; sub.ftz.f32 %f28315, %f28263, %f7518; max.ftz.f32 %f28316, %f28315, %f28310; min.ftz.f32 %f28317, %f28316, %f28263; mul.ftz.f32 %f28318, %f28317, %f28317; mul.ftz.f32 %f28319, %f28318, %f28318; mul.ftz.f32 %f7520, %f28317, %f28319; ld.local.u32 %r1552, [%rd1027+24]; and.b32 %r5379, %r1552, 3; setp.eq.s32 %p3785, %r5379, 0; fma.rn.ftz.f32 %f7521, %f7520, 0fBF000000, 0f3F800000; @%p3785 bra $L__BB107_4422; fma.rn.ftz.f32 %f28320, %f7519, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f30896, %f28320, %f7521, 0f00000000; bra.uni $L__BB107_4423; $L__BB107_3952: setp.eq.s32 %p3359, %r1522, 8; @%p3359 bra $L__BB107_4416; setp.eq.s32 %p3360, %r1522, 9; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3360 bra $L__BB107_3954; bra.uni $L__BB107_4433; $L__BB107_3954: ld.local.v2.f32 {%f28029, %f28030}, [%rd5+-48]; ld.local.f32 %f7461, [%rd1027+8]; ld.local.f32 %f7462, [%rd5+-52]; ld.local.f32 %f7463, [%rd1027+12]; mul.ftz.f32 %f28031, %f28029, %f7463; fma.rn.ftz.f32 %f28032, %f7462, %f7461, %f28031; ld.local.f32 %f7464, [%rd1027+16]; fma.rn.ftz.f32 %f7465, %f28030, %f7464, %f28032; setp.leu.ftz.f32 %p3758, %f7465, 0f00000000; mov.u32 %r5877, 18; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3758 bra $L__BB107_4433; add.ftz.f32 %f28037, %f7465, %f7465; mul.ftz.f32 %f28038, %f7461, %f28037; mul.ftz.f32 %f28039, %f7463, %f28037; mul.ftz.f32 %f28040, %f7464, %f28037; sub.ftz.f32 %f30899, %f28038, %f7462; sub.ftz.f32 %f30900, %f28039, %f28029; sub.ftz.f32 %f30901, %f28040, %f28030; mul.ftz.f32 %f28041, %f6234, %f30900; fma.rn.ftz.f32 %f28042, %f6233, %f30899, %f28041; fma.rn.ftz.f32 %f28043, %f6235, %f30901, %f28042; setp.leu.ftz.f32 %p3759, %f28043, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3759 bra $L__BB107_4433; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4371: setp.eq.s32 %p3726, %r1522, 23; selp.b32 %r5877, 9, 10, %p3726; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_3959: setp.eq.s32 %p3356, %r1522, 13; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3356 bra $L__BB107_3975; bra.uni $L__BB107_4433; $L__BB107_3975: ld.local.v2.f32 {%f27067, %f27068}, [%rd5+-48]; ld.local.v2.f32 {%f27069, %f27070}, [%rd1027+20]; ld.local.f32 %f7163, [%rd1027+8]; ld.local.f32 %f7164, [%rd5+-52]; ld.local.f32 %f7165, [%rd1027+12]; mul.ftz.f32 %f27071, %f27067, %f7165; fma.rn.ftz.f32 %f27072, %f7164, %f7163, %f27071; ld.local.f32 %f7166, [%rd1027+16]; fma.rn.ftz.f32 %f7167, %f27068, %f7166, %f27072; setp.gt.ftz.f32 %p3691, %f7167, 0f00000000; @%p3691 bra $L__BB107_4322; bra.uni $L__BB107_3976; $L__BB107_4322: setp.eq.ftz.f32 %p3693, %f27069, %f27070; @%p3693 bra $L__BB107_4324; bra.uni $L__BB107_4323; $L__BB107_4324: setp.neu.ftz.f32 %p3694, %f7163, %f7165; setp.neu.ftz.f32 %p3695, %f7163, %f7166; or.pred %p3696, %p3694, %p3695; neg.ftz.f32 %f27100, %f7166; selp.f32 %f27101, %f27100, %f7166, %p3696; add.ftz.f32 %f27102, %f7163, %f27101; neg.ftz.f32 %f27103, %f7165; selp.f32 %f27104, %f7165, %f27103, %p3696; sub.ftz.f32 %f27105, %f27104, %f7163; sub.ftz.f32 %f27106, %f7166, %f7165; mul.ftz.f32 %f27107, %f27102, %f27102; fma.rn.ftz.f32 %f27108, %f27106, %f27106, %f27107; fma.rn.ftz.f32 %f27109, %f27105, %f27105, %f27108; sqrt.approx.ftz.f32 %f27110, %f27109; div.approx.ftz.f32 %f30788, %f27106, %f27110; div.approx.ftz.f32 %f30789, %f27102, %f27110; div.approx.ftz.f32 %f30790, %f27105, %f27110; mul.ftz.f32 %f27111, %f7165, %f30790; mul.ftz.f32 %f27112, %f7166, %f30789; sub.ftz.f32 %f30785, %f27111, %f27112; mul.ftz.f32 %f27113, %f7166, %f30788; mul.ftz.f32 %f27114, %f7163, %f30790; sub.ftz.f32 %f30786, %f27113, %f27114; mul.ftz.f32 %f27115, %f7163, %f30789; mul.ftz.f32 %f27116, %f7165, %f30788; sub.ftz.f32 %f30787, %f27115, %f27116; bra.uni $L__BB107_4325; $L__BB107_4418: mul.ftz.f32 %f28174, %f6226, 0f40C90FDB; cos.approx.ftz.f32 %f28175, %f28174; sqrt.approx.ftz.f32 %f28176, %f6202; mul.ftz.f32 %f28177, %f28176, %f28175; sin.approx.ftz.f32 %f28178, %f28174; mul.ftz.f32 %f28179, %f28176, %f28178; mul.ftz.f32 %f28180, %f28177, %f28177; mov.f32 %f28181, 0f3F800000; sub.ftz.f32 %f28182, %f28181, %f28180; mul.ftz.f32 %f28183, %f28179, %f28179; sub.ftz.f32 %f28184, %f28182, %f28183; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28185, %f28184, %f30902; sqrt.approx.ftz.f32 %f7486, %f28185; ld.local.f32 %f7487, [%rd1027+12]; ld.local.f32 %f7488, [%rd1027+8]; setp.neu.ftz.f32 %p3773, %f7488, %f7487; ld.local.f32 %f7489, [%rd1027+16]; setp.neu.ftz.f32 %p3774, %f7488, %f7489; or.pred %p3775, %p3773, %p3774; sub.ftz.f32 %f28186, %f7489, %f7487; neg.ftz.f32 %f28187, %f7487; neg.ftz.f32 %f28188, %f7489; selp.f32 %f28189, %f28188, %f7489, %p3775; add.ftz.f32 %f28190, %f7488, %f28189; selp.f32 %f28191, %f7487, %f28187, %p3775; sub.ftz.f32 %f28192, %f28191, %f7488; mul.ftz.f32 %f28193, %f28190, %f28190; fma.rn.ftz.f32 %f28194, %f28186, %f28186, %f28193; fma.rn.ftz.f32 %f28195, %f28192, %f28192, %f28194; sqrt.approx.ftz.f32 %f28196, %f28195; div.approx.ftz.f32 %f28197, %f28186, %f28196; div.approx.ftz.f32 %f28198, %f28190, %f28196; div.approx.ftz.f32 %f28199, %f28192, %f28196; mul.ftz.f32 %f28200, %f7487, %f28199; mul.ftz.f32 %f28201, %f7489, %f28198; sub.ftz.f32 %f28202, %f28200, %f28201; mul.ftz.f32 %f28203, %f7489, %f28197; mul.ftz.f32 %f28204, %f7488, %f28199; sub.ftz.f32 %f28205, %f28203, %f28204; mul.ftz.f32 %f28206, %f7488, %f28198; mul.ftz.f32 %f28207, %f7487, %f28197; sub.ftz.f32 %f28208, %f28206, %f28207; mul.ftz.f32 %f28209, %f28179, %f28202; mul.ftz.f32 %f28210, %f28179, %f28205; mul.ftz.f32 %f28211, %f28179, %f28208; fma.rn.ftz.f32 %f28212, %f28177, %f28197, %f28209; fma.rn.ftz.f32 %f28213, %f28177, %f28198, %f28210; fma.rn.ftz.f32 %f28214, %f28177, %f28199, %f28211; fma.rn.ftz.f32 %f30899, %f7488, %f7486, %f28212; fma.rn.ftz.f32 %f30900, %f7487, %f7486, %f28213; fma.rn.ftz.f32 %f30901, %f7489, %f7486, %f28214; mul.ftz.f32 %f28215, %f6234, %f30900; fma.rn.ftz.f32 %f28216, %f6233, %f30899, %f28215; fma.rn.ftz.f32 %f28217, %f6235, %f30901, %f28216; setp.leu.ftz.f32 %p3776, %f28217, 0f00000000; mov.u32 %r5877, 6; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3776 bra $L__BB107_4433; ld.local.f32 %f28222, [%rd5+-52]; ld.local.v2.f32 {%f28223, %f28224}, [%rd5+-48]; add.ftz.f32 %f28227, %f28222, %f30899; add.ftz.f32 %f28228, %f28223, %f30900; mul.ftz.f32 %f28229, %f28228, %f28228; fma.rn.ftz.f32 %f28230, %f28227, %f28227, %f28229; add.ftz.f32 %f28231, %f28224, %f30901; fma.rn.ftz.f32 %f28232, %f28231, %f28231, %f28230; sqrt.approx.ftz.f32 %f28233, %f28232; div.approx.ftz.f32 %f7493, %f28227, %f28233; div.approx.ftz.f32 %f7494, %f28228, %f28233; div.approx.ftz.f32 %f7495, %f28231, %f28233; mul.ftz.f32 %f28234, %f7487, %f30900; fma.rn.ftz.f32 %f28235, %f7488, %f30899, %f28234; fma.rn.ftz.f32 %f7496, %f7489, %f30901, %f28235; mul.ftz.f32 %f28236, %f28223, %f7487; fma.rn.ftz.f32 %f28237, %f28222, %f7488, %f28236; fma.rn.ftz.f32 %f28238, %f28224, %f7489, %f28237; setp.lt.ftz.f32 %p3777, %f7496, 0f00000000; setp.lt.ftz.f32 %p3778, %f28238, 0f00000000; or.pred %p3779, %p3778, %p3777; mov.f32 %f30902, 0f00000000; mov.f32 %f30903, %f30902; @%p3779 bra $L__BB107_4421; mul.ftz.f32 %f28239, %f30900, %f7494; fma.rn.ftz.f32 %f28240, %f30899, %f7493, %f28239; fma.rn.ftz.f32 %f28241, %f30901, %f7495, %f28240; sub.ftz.f32 %f28243, %f28181, %f28241; mov.f32 %f28244, 0f00000000; max.ftz.f32 %f28245, %f28243, %f28244; min.ftz.f32 %f28246, %f28245, %f28181; mul.ftz.f32 %f28247, %f28246, %f28246; mul.ftz.f32 %f28248, %f28247, %f28247; mul.ftz.f32 %f28249, %f28246, %f28248; mul.ftz.f32 %f30903, %f7496, %f28249; mul.ftz.f32 %f30902, %f7486, 0f3EA2F983; $L__BB107_4421: mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_4237: ld.local.v2.f32 {%f26392, %f26393}, [%rd5+-48]; ld.local.f32 %f6911, [%rd1027+8]; ld.local.f32 %f6912, [%rd5+-52]; ld.local.f32 %f6913, [%rd1027+12]; mul.ftz.f32 %f26394, %f26392, %f6913; fma.rn.ftz.f32 %f26395, %f6912, %f6911, %f26394; ld.local.f32 %f6914, [%rd1027+16]; fma.rn.ftz.f32 %f6915, %f26393, %f6914, %f26395; setp.leu.ftz.f32 %p3616, %f6915, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3616 bra $L__BB107_4433; ld.local.v2.f32 {%f26396, %f26397}, [%rd1027+20]; mul.ftz.f32 %f26398, %f26396, %f26396; mov.f32 %f26399, 0f40000000; div.approx.ftz.f32 %f26400, %f26399, %f26398; add.ftz.f32 %f6918, %f26400, 0fC0000000; mul.ftz.f32 %f26401, %f26397, %f26397; div.approx.ftz.f32 %f26402, %f26399, %f26401; add.ftz.f32 %f6919, %f26402, 0fC0000000; setp.eq.ftz.f32 %p3617, %f6918, %f6919; @%p3617 bra $L__BB107_4240; bra.uni $L__BB107_4239; $L__BB107_4240: setp.neu.ftz.f32 %p3618, %f6911, %f6913; setp.neu.ftz.f32 %p3619, %f6911, %f6914; or.pred %p3620, %p3618, %p3619; neg.ftz.f32 %f26425, %f6914; selp.f32 %f26426, %f26425, %f6914, %p3620; add.ftz.f32 %f26427, %f6911, %f26426; neg.ftz.f32 %f26428, %f6913; selp.f32 %f26429, %f6913, %f26428, %p3620; sub.ftz.f32 %f26430, %f26429, %f6911; sub.ftz.f32 %f26431, %f6914, %f6913; mul.ftz.f32 %f26432, %f26427, %f26427; fma.rn.ftz.f32 %f26433, %f26431, %f26431, %f26432; fma.rn.ftz.f32 %f26434, %f26430, %f26430, %f26433; sqrt.approx.ftz.f32 %f26435, %f26434; div.approx.ftz.f32 %f30734, %f26431, %f26435; div.approx.ftz.f32 %f30735, %f26427, %f26435; div.approx.ftz.f32 %f30736, %f26430, %f26435; mul.ftz.f32 %f26436, %f6913, %f30736; mul.ftz.f32 %f26437, %f6914, %f30735; sub.ftz.f32 %f30731, %f26436, %f26437; mul.ftz.f32 %f26438, %f6914, %f30734; mul.ftz.f32 %f26439, %f6911, %f30736; sub.ftz.f32 %f30732, %f26438, %f26439; mul.ftz.f32 %f26440, %f6911, %f30735; mul.ftz.f32 %f26441, %f6913, %f30734; sub.ftz.f32 %f30733, %f26440, %f26441; bra.uni $L__BB107_4241; $L__BB107_4232: mul.ftz.f32 %f26290, %f6226, %f6226; mov.f32 %f26291, 0f3F800000; sub.ftz.f32 %f26292, %f26291, %f26290; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f26293, %f30902, %f26292; sqrt.approx.ftz.f32 %f26294, %f26293; mul.ftz.f32 %f26295, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f26296, %f26295; mul.ftz.f32 %f26297, %f26294, %f26296; sin.approx.ftz.f32 %f26298, %f26295; mul.ftz.f32 %f26299, %f26294, %f26298; ld.local.f32 %f6897, [%rd1027+12]; ld.local.f32 %f6898, [%rd1027+8]; setp.neu.ftz.f32 %p3607, %f6898, %f6897; ld.local.f32 %f6899, [%rd1027+16]; setp.neu.ftz.f32 %p3608, %f6898, %f6899; or.pred %p3609, %p3607, %p3608; sub.ftz.f32 %f26300, %f6899, %f6897; neg.ftz.f32 %f26301, %f6897; neg.ftz.f32 %f26302, %f6899; selp.f32 %f26303, %f26302, %f6899, %p3609; add.ftz.f32 %f26304, %f6898, %f26303; selp.f32 %f26305, %f6897, %f26301, %p3609; sub.ftz.f32 %f26306, %f26305, %f6898; mul.ftz.f32 %f26307, %f26304, %f26304; fma.rn.ftz.f32 %f26308, %f26300, %f26300, %f26307; fma.rn.ftz.f32 %f26309, %f26306, %f26306, %f26308; sqrt.approx.ftz.f32 %f26310, %f26309; div.approx.ftz.f32 %f26311, %f26300, %f26310; div.approx.ftz.f32 %f26312, %f26304, %f26310; div.approx.ftz.f32 %f26313, %f26306, %f26310; mul.ftz.f32 %f26314, %f6897, %f26313; mul.ftz.f32 %f26315, %f6899, %f26312; sub.ftz.f32 %f26316, %f26314, %f26315; mul.ftz.f32 %f26317, %f6899, %f26311; mul.ftz.f32 %f26318, %f6898, %f26313; sub.ftz.f32 %f26319, %f26317, %f26318; mul.ftz.f32 %f26320, %f6898, %f26312; mul.ftz.f32 %f26321, %f6897, %f26311; sub.ftz.f32 %f26322, %f26320, %f26321; mul.ftz.f32 %f26323, %f26299, %f26316; mul.ftz.f32 %f26324, %f26299, %f26319; mul.ftz.f32 %f26325, %f26299, %f26322; fma.rn.ftz.f32 %f26326, %f26297, %f26311, %f26323; fma.rn.ftz.f32 %f26327, %f26297, %f26312, %f26324; fma.rn.ftz.f32 %f26328, %f26297, %f26313, %f26325; fma.rn.ftz.f32 %f30899, %f6226, %f6898, %f26326; fma.rn.ftz.f32 %f30900, %f6226, %f6897, %f26327; fma.rn.ftz.f32 %f30901, %f6226, %f6899, %f26328; mul.ftz.f32 %f26329, %f6234, %f30900; fma.rn.ftz.f32 %f26330, %f6233, %f30899, %f26329; fma.rn.ftz.f32 %f26331, %f6235, %f30901, %f26330; setp.leu.ftz.f32 %p3610, %f26331, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3610 bra $L__BB107_4433; ld.local.f32 %f26332, [%rd5+-52]; ld.local.v2.f32 {%f26333, %f26334}, [%rd5+-48]; ld.local.f32 %f6903, [%rd1027+24]; add.ftz.f32 %f26337, %f26332, %f30899; add.ftz.f32 %f26338, %f26333, %f30900; mul.ftz.f32 %f26339, %f26338, %f26338; fma.rn.ftz.f32 %f26340, %f26337, %f26337, %f26339; add.ftz.f32 %f26341, %f26334, %f30901; fma.rn.ftz.f32 %f26342, %f26341, %f26341, %f26340; sqrt.approx.ftz.f32 %f26343, %f26342; div.approx.ftz.f32 %f26344, %f26337, %f26343; div.approx.ftz.f32 %f26345, %f26338, %f26343; div.approx.ftz.f32 %f26346, %f26341, %f26343; mul.ftz.f32 %f26347, %f6897, %f30900; fma.rn.ftz.f32 %f26348, %f6898, %f30899, %f26347; fma.rn.ftz.f32 %f6904, %f6899, %f30901, %f26348; mul.ftz.f32 %f26349, %f26333, %f6897; fma.rn.ftz.f32 %f26350, %f26332, %f6898, %f26349; fma.rn.ftz.f32 %f6905, %f26334, %f6899, %f26350; mul.ftz.f32 %f26351, %f6897, %f26345; fma.rn.ftz.f32 %f26352, %f6898, %f26344, %f26351; fma.rn.ftz.f32 %f6906, %f6899, %f26346, %f26352; mul.ftz.f32 %f26353, %f26333, %f26345; fma.rn.ftz.f32 %f26354, %f26332, %f26344, %f26353; fma.rn.ftz.f32 %f26355, %f26334, %f26346, %f26354; abs.ftz.f32 %f6907, %f26355; abs.ftz.f32 %f26356, %f6905; setp.leu.ftz.f32 %p3612, %f26356, 0f3727C5AC; mov.pred %p3931, 0; @%p3612 bra $L__BB107_4235; abs.ftz.f32 %f26357, %f6906; setp.lt.ftz.f32 %p3931, %f26357, 0f3F7FFF58; $L__BB107_4235: setp.leu.ftz.f32 %p3613, %f6907, 0f3727C5AC; not.pred %p3614, %p3931; or.pred %p3615, %p3613, %p3614; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3615 bra $L__BB107_4433; div.approx.ftz.f32 %f26363, %f6906, %f6907; mov.f32 %f26364, 0f3727C5AC; max.ftz.f32 %f26365, %f26363, %f26364; mul.ftz.f32 %f26366, %f6905, %f26365; abs.ftz.f32 %f26367, %f26366; add.ftz.f32 %f26368, %f26367, %f26367; mul.ftz.f32 %f26369, %f6904, %f26365; abs.ftz.f32 %f26370, %f26369; add.ftz.f32 %f26371, %f26370, %f26370; mul.ftz.f32 %f26372, %f6906, %f6906; sub.ftz.f32 %f26374, %f26291, %f26372; mul.ftz.f32 %f26375, %f26374, %f26374; div.approx.ftz.f32 %f26376, %f26372, %f26374; mul.ftz.f32 %f26377, %f6903, %f26376; mul.ftz.f32 %f26378, %f26377, 0fBFB8AA3B; ex2.approx.ftz.f32 %f26379, %f26378; mul.ftz.f32 %f26380, %f6903, %f26379; mul.ftz.f32 %f26381, %f26380, 0f3EA2F983; div.approx.ftz.f32 %f26382, %f26381, %f26375; min.ftz.f32 %f26383, %f26368, %f26371; min.ftz.f32 %f26384, %f26291, %f26383; mul.ftz.f32 %f26385, %f26382, %f26384; mul.ftz.f32 %f26386, %f26385, 0f3E800000; div.approx.ftz.f32 %f30903, %f26386, %f6905; mov.u32 %r5877, 6; mov.f32 %f30902, 0f3E22F983; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_3982: setp.neu.ftz.f32 %p3511, %f30710, %f30711; setp.neu.ftz.f32 %p3512, %f30710, %f30712; or.pred %p3513, %p3511, %p3512; neg.ftz.f32 %f25515, %f30712; selp.f32 %f25516, %f25515, %f30712, %p3513; add.ftz.f32 %f25517, %f30710, %f25516; neg.ftz.f32 %f25518, %f30711; selp.f32 %f25519, %f30711, %f25518, %p3513; sub.ftz.f32 %f25520, %f25519, %f30710; sub.ftz.f32 %f25521, %f30712, %f30711; mul.ftz.f32 %f25522, %f25517, %f25517; fma.rn.ftz.f32 %f25523, %f25521, %f25521, %f25522; fma.rn.ftz.f32 %f25524, %f25520, %f25520, %f25523; sqrt.approx.ftz.f32 %f25525, %f25524; div.approx.ftz.f32 %f6628, %f25521, %f25525; div.approx.ftz.f32 %f6629, %f25517, %f25525; div.approx.ftz.f32 %f6630, %f25520, %f25525; mul.ftz.f32 %f25526, %f30711, %f6630; mul.ftz.f32 %f25527, %f30712, %f6629; sub.ftz.f32 %f6631, %f25526, %f25527; mul.ftz.f32 %f25528, %f30712, %f6628; mul.ftz.f32 %f25529, %f30710, %f6630; sub.ftz.f32 %f6632, %f25528, %f25529; mul.ftz.f32 %f25530, %f30710, %f6629; mul.ftz.f32 %f25531, %f30711, %f6628; sub.ftz.f32 %f6633, %f25530, %f25531; mul.ftz.f32 %f25532, %f25507, %f6629; fma.rn.ftz.f32 %f25533, %f6624, %f6628, %f25532; fma.rn.ftz.f32 %f6634, %f25508, %f6630, %f25533; mul.ftz.f32 %f25534, %f25507, %f6632; fma.rn.ftz.f32 %f25535, %f6624, %f6631, %f25534; fma.rn.ftz.f32 %f6635, %f25508, %f6633, %f25535; ld.local.u64 %rd4467, [%rd1027+36]; ld.f32 %f6636, [%rd4467]; ld.f32 %f6637, [%rd4467+4]; ld.f32 %f6638, [%rd4467+8]; ld.local.f32 %f6639, [%rd1027+28]; ld.f32 %f6640, [%rd4467+12]; ld.f32 %f6641, [%rd4467+16]; ld.f32 %f6642, [%rd4467+20]; neg.ftz.f32 %f6643, %f6634; neg.ftz.f32 %f6644, %f6635; neg.ftz.f32 %f6645, %f30713; setp.lt.ftz.f32 %p3514, %f30713, 0fBF7FF972; mov.f32 %f30696, 0f00000000; @%p3514 bra $L__BB107_3985; setp.gt.ftz.f32 %p3515, %f30713, 0f3F7FF972; mov.f32 %f30696, 0fBF7FF972; @%p3515 bra $L__BB107_3985; mul.ftz.f32 %f25537, %f30713, %f30713; mov.f32 %f25538, 0f33D6BF95; max.ftz.f32 %f25539, %f25537, %f25538; rcp.approx.ftz.f32 %f25540, %f25539; mul.ftz.f32 %f25541, %f25509, %f6643; mul.ftz.f32 %f25542, %f25510, %f6644; mul.ftz.f32 %f25543, %f25542, %f25542; fma.rn.ftz.f32 %f25544, %f25541, %f25541, %f25543; fma.rn.ftz.f32 %f25545, %f25544, %f25540, 0f3F800000; sqrt.approx.ftz.f32 %f25546, %f25545; neg.ftz.f32 %f25547, %f25546; setp.ge.ftz.f32 %p3516, %f30713, 0f80000000; selp.f32 %f25548, %f25547, %f25546, %p3516; add.ftz.f32 %f25549, %f25548, 0fBF800000; mul.ftz.f32 %f30696, %f25549, 0f3F000000; $L__BB107_3985: fma.rn.ftz.f32 %f6648, %f6639, %f6639, 0fBF800000; mov.f32 %f30695, 0f3F800000; abs.ftz.f32 %f6649, %f30695; fma.rn.ftz.f32 %f6650, %f6649, %f6649, %f6648; setp.leu.ftz.f32 %p3517, %f6650, 0f00000000; mov.f32 %f30670, %f30695; @%p3517 bra $L__BB107_3987; sqrt.approx.ftz.f32 %f25551, %f6650; sub.ftz.f32 %f25552, %f25551, %f6649; add.ftz.f32 %f25553, %f6649, %f25551; div.approx.ftz.f32 %f25554, %f25552, %f25553; fma.rn.ftz.f32 %f25555, %f6649, %f25553, 0fBF800000; fma.rn.ftz.f32 %f25556, %f6649, %f25552, 0f3F800000; div.approx.ftz.f32 %f25557, %f25555, %f25556; mul.ftz.f32 %f25558, %f25554, 0f3F000000; mul.ftz.f32 %f25559, %f25554, %f25558; fma.rn.ftz.f32 %f25560, %f25557, %f25557, 0f3F800000; mul.ftz.f32 %f30670, %f25559, %f25560; $L__BB107_3987: setp.ne.s32 %p17, %r1522, 27; rcp.approx.ftz.f32 %f6653, %f6639; sub.ftz.f32 %f6654, %f30695, %f30670; rcp.approx.ftz.f32 %f6655, %f6654; ld.local.u32 %r5871, [%rd5+168]; mov.f32 %f30697, 0f00000000; mov.u16 %rs742, 1; mov.u32 %r5872, 0; mov.pred %p3518, -1; mov.pred %p3545, 0; mov.f32 %f30694, %f30695; mov.f32 %f30693, %f30695; mov.f32 %f6659, %f6645; mov.f32 %f6660, %f6644; mov.f32 %f6661, %f6643; mov.f32 %f30680, %f30695; $L__BB107_3988: mov.f32 %f6663, %f30680; mad.lo.s32 %r1534, %r5871, 1103515245, 12345; cvt.rn.f32.u32 %f25566, %r1534; mul.ftz.f32 %f6665, %f25566, 0f2F800000; setp.gt.ftz.f32 %p3519, %f6659, 0f3F7FF972; mov.pred %p3929, %p3518; @%p3519 bra $L__BB107_4166; setp.lt.ftz.f32 %p3520, %f6659, 0fBF7FF972; @%p3520 bra $L__BB107_4165; bra.uni $L__BB107_3990; $L__BB107_4165: mul.ftz.f32 %f30680, %f6663, %f6665; mov.pred %p3929, %p3545; bra.uni $L__BB107_4166; $L__BB107_3990: abs.ftz.f32 %f25567, %f6659; setp.ltu.ftz.f32 %p3522, %f25567, 0f38D1B717; mov.f32 %f30680, %f6663; mov.pred %p3929, %p3545; @%p3522 bra $L__BB107_4166; mov.f32 %f25568, 0f3F800000; sub.ftz.f32 %f25569, %f25568, %f30697; setp.lt.ftz.f32 %p3524, %f25569, %f6665; mov.f32 %f30680, %f6663; mov.pred %p3929, %p3518; @%p3524 bra $L__BB107_4166; setp.ge.ftz.f32 %p3526, %f30696, 0f00000000; mov.f32 %f30680, %f25568; mov.pred %p3929, %p3545; @%p3526 bra $L__BB107_4166; mov.f32 %f25571, 0f3F800000; sub.ftz.f32 %f25572, %f25571, %f6665; lg2.approx.ftz.f32 %f25573, %f25572; mov.f32 %f25574, 0fBF800000; div.approx.ftz.f32 %f25575, %f25574, %f30696; mul.ftz.f32 %f25576, %f25575, %f25573; ex2.approx.ftz.f32 %f25577, %f25576; mul.ftz.f32 %f30680, %f6663, %f25577; mov.pred %p3929, %p3545; $L__BB107_4166: @%p3929 bra $L__BB107_4198; mad.lo.s32 %r5260, %r1534, 1103515245, 12345; cvt.rn.f32.u32 %f25578, %r5260; mul.ftz.f32 %f6669, %f25578, 0f2F800000; mad.lo.s32 %r1535, %r5260, 1103515245, 12345; cvt.rn.f32.u32 %f25579, %r1535; mul.ftz.f32 %f6670, %f25579, 0f2F800000; neg.ftz.f32 %f6671, %f6661; mul.ftz.f32 %f25580, %f25509, %f6671; neg.ftz.f32 %f6672, %f6660; mul.ftz.f32 %f25581, %f25510, %f6672; mul.ftz.f32 %f25582, %f25581, %f25581; fma.rn.ftz.f32 %f25583, %f25580, %f25580, %f25582; fma.rn.ftz.f32 %f25584, %f6659, %f6659, %f25583; sqrt.approx.ftz.f32 %f25585, %f25584; div.approx.ftz.f32 %f30683, %f25580, %f25585; div.approx.ftz.f32 %f30684, %f25581, %f25585; neg.ftz.f32 %f6675, %f6659; div.approx.ftz.f32 %f6676, %f6675, %f25585; setp.gt.ftz.f32 %p3529, %f6676, 0f3F7FF972; @%p3529 bra $L__BB107_4173; abs.ftz.f32 %f25586, %f6676; setp.lt.ftz.f32 %p3530, %f25586, 0f358637BD; @%p3530 bra $L__BB107_4173; bra.uni $L__BB107_4169; $L__BB107_4173: mov.f32 %f25630, 0f3F800000; sub.ftz.f32 %f25631, %f25630, %f6670; mov.f32 %f25632, 0f33D6BF95; max.ftz.f32 %f25633, %f25631, %f25632; div.approx.ftz.f32 %f25634, %f6670, %f25633; sqrt.approx.ftz.f32 %f25635, %f25634; mul.ftz.f32 %f25636, %f6669, 0f40C90FDB; cos.approx.ftz.f32 %f25637, %f25636; sin.approx.ftz.f32 %f25638, %f25636; mul.ftz.f32 %f30682, %f25635, %f25638; mul.ftz.f32 %f30681, %f25635, %f25637; $L__BB107_4174: mul.ftz.f32 %f25639, %f30684, %f30684; fma.rn.ftz.f32 %f25640, %f30683, %f30683, %f25639; add.ftz.f32 %f25641, %f25640, 0f00000000; sqrt.approx.ftz.f32 %f6694, %f25641; setp.neu.ftz.f32 %p3538, %f6694, 0f00000000; @%p3538 bra $L__BB107_4175; bra.uni $L__BB107_4176; $L__BB107_4175: rcp.approx.ftz.f32 %f25642, %f6694; mul.ftz.f32 %f30684, %f30684, %f25642; mul.ftz.f32 %f30683, %f30683, %f25642; $L__BB107_4176: mul.ftz.f32 %f25644, %f30681, %f30683; mul.ftz.f32 %f25645, %f30682, %f30684; sub.ftz.f32 %f25646, %f25644, %f25645; mul.ftz.f32 %f25647, %f25509, %f25646; mul.ftz.f32 %f25648, %f30682, %f30683; fma.rn.ftz.f32 %f25649, %f30681, %f30684, %f25648; mul.ftz.f32 %f25650, %f25510, %f25649; neg.ftz.f32 %f25651, %f25647; neg.ftz.f32 %f25652, %f25650; mul.ftz.f32 %f25653, %f25650, %f25650; fma.rn.ftz.f32 %f25654, %f25647, %f25647, %f25653; add.ftz.f32 %f25655, %f25654, 0f3F800000; mov.f32 %f30685, 0f3F800000; sqrt.approx.ftz.f32 %f25656, %f25655; div.approx.ftz.f32 %f6704, %f25651, %f25656; div.approx.ftz.f32 %f6705, %f25652, %f25656; div.approx.ftz.f32 %f6706, %f30685, %f25656; setp.ne.s32 %p3539, %r5872, 0; and.pred %p3540, %p3539, %p17; mul.ftz.f32 %f25657, %f6636, %f30693; selp.f32 %f6707, %f25657, %f30693, %p3540; mul.ftz.f32 %f25658, %f6637, %f30694; selp.f32 %f6708, %f25658, %f30694, %p3540; mul.ftz.f32 %f25659, %f6638, %f30695; selp.f32 %f6709, %f25659, %f30695, %p3540; mad.lo.s32 %r5871, %r1535, 1103515245, 12345; and.b16 %rs589, %rs742, 255; setp.eq.s16 %p3541, %rs589, 0; selp.f32 %f6710, %f6653, %f6639, %p3541; mul.ftz.f32 %f25660, %f6705, %f6672; fma.rn.ftz.f32 %f25661, %f6704, %f6671, %f25660; fma.rn.ftz.f32 %f6711, %f6706, %f6675, %f25661; abs.ftz.f32 %f6712, %f6711; fma.rn.ftz.f32 %f25662, %f6710, %f6710, 0fBF800000; mul.ftz.f32 %f6713, %f6712, %f6712; add.ftz.f32 %f6714, %f25662, %f6713; setp.leu.ftz.f32 %p3542, %f6714, 0f00000000; @%p3542 bra $L__BB107_4178; sqrt.approx.ftz.f32 %f25663, %f6714; sub.ftz.f32 %f25664, %f25663, %f6712; add.ftz.f32 %f25665, %f6712, %f25663; div.approx.ftz.f32 %f25666, %f25664, %f25665; fma.rn.ftz.f32 %f25667, %f6712, %f25665, 0fBF800000; fma.rn.ftz.f32 %f25668, %f6712, %f25664, 0f3F800000; div.approx.ftz.f32 %f25669, %f25667, %f25668; mul.ftz.f32 %f25670, %f25666, 0f3F000000; mul.ftz.f32 %f25671, %f25666, %f25670; fma.rn.ftz.f32 %f25672, %f25669, %f25669, 0f3F800000; mul.ftz.f32 %f30685, %f25671, %f25672; $L__BB107_4178: cvt.rn.f32.u32 %f25673, %r5871; mul.ftz.f32 %f25674, %f25673, 0f2F800000; setp.gt.ftz.f32 %p3543, %f30685, %f25674; @%p3543 bra $L__BB107_4180; bra.uni $L__BB107_4179; $L__BB107_4180: add.ftz.f32 %f25700, %f6704, %f6704; add.ftz.f32 %f25701, %f6705, %f6705; add.ftz.f32 %f25702, %f6706, %f6706; fma.rn.ftz.f32 %f6659, %f25702, %f6711, %f6659; fma.rn.ftz.f32 %f6660, %f25701, %f6711, %f6660; fma.rn.ftz.f32 %f6661, %f25700, %f6711, %f6661; mov.pred %p3930, %p3545; bra.uni $L__BB107_4181; $L__BB107_4179: rcp.approx.ftz.f32 %f25675, %f6710; mov.f32 %f25676, 0f3F800000; mul.ftz.f32 %f25677, %f6711, %f6711; sub.ftz.f32 %f25678, %f25676, %f25677; mul.ftz.f32 %f25679, %f25678, %f25675; mul.ftz.f32 %f25680, %f25675, %f25679; sub.ftz.f32 %f25681, %f25676, %f25680; mov.f32 %f25682, 0f00000000; max.ftz.f32 %f25683, %f25681, %f25682; sqrt.approx.ftz.f32 %f25684, %f25683; mul.ftz.f32 %f25685, %f6711, %f25675; sub.ftz.f32 %f25686, %f25685, %f25684; mul.ftz.f32 %f25687, %f6704, %f25686; mul.ftz.f32 %f25688, %f6705, %f25686; mul.ftz.f32 %f25689, %f6706, %f25686; mul.ftz.f32 %f25690, %f25675, %f6671; mul.ftz.f32 %f25691, %f25675, %f6672; mul.ftz.f32 %f25692, %f25675, %f6675; sub.ftz.f32 %f25693, %f25687, %f25690; sub.ftz.f32 %f25694, %f25688, %f25691; sub.ftz.f32 %f25695, %f25689, %f25692; mul.ftz.f32 %f25696, %f25694, %f25694; fma.rn.ftz.f32 %f25697, %f25693, %f25693, %f25696; fma.rn.ftz.f32 %f25698, %f25695, %f25695, %f25697; sqrt.approx.ftz.f32 %f25699, %f25698; div.approx.ftz.f32 %f6659, %f25695, %f25699; div.approx.ftz.f32 %f6660, %f25694, %f25699; div.approx.ftz.f32 %f6661, %f25693, %f25699; mov.pred %p3930, %p3518; $L__BB107_4181: not.pred %p3546, %p3930; @%p3546 bra $L__BB107_4183; neg.ftz.f32 %f6661, %f6661; neg.ftz.f32 %f6660, %f6660; neg.ftz.f32 %f6659, %f6659; selp.u16 %rs742, 1, 0, %p3541; $L__BB107_4183: mov.f32 %f30693, %f6707; mov.f32 %f30694, %f6708; mov.f32 %f30695, %f6709; @%p17 bra $L__BB107_4190; @%p3930 bra $L__BB107_4189; bra.uni $L__BB107_4185; $L__BB107_4189: mul.ftz.f32 %f30693, %f6636, %f6707; mul.ftz.f32 %f30694, %f6637, %f6708; mul.ftz.f32 %f30695, %f6638, %f6709; bra.uni $L__BB107_4190; $L__BB107_4169: add.ftz.f32 %f25589, %f6676, 0f3F800000; mul.ftz.f32 %f6677, %f25589, 0f3F000000; setp.lt.ftz.f32 %p3531, %f6677, 0f38D1B717; mov.f32 %f30681, 0f00000000; mov.f32 %f30682, %f30681; @%p3531 bra $L__BB107_4174; add.ftz.f32 %f25592, %f6670, %f6670; mul.ftz.f32 %f25593, %f25592, %f6677; div.approx.ftz.f32 %f25594, %f25593, %f6676; add.ftz.f32 %f6678, %f25594, 0fBF800000; mul.ftz.f32 %f6679, %f6678, %f6678; add.ftz.f32 %f6680, %f6679, 0fBF800000; abs.ftz.f32 %f25595, %f6680; setp.lt.ftz.f32 %p3532, %f25595, 0f33D6BF95; mov.f32 %f30682, %f30681; @%p3532 bra $L__BB107_4174; mul.ftz.f32 %f25596, %f6676, %f6676; mov.f32 %f25597, 0f3F800000; sub.ftz.f32 %f25598, %f25597, %f25596; mov.f32 %f25599, 0f00000000; max.ftz.f32 %f25600, %f25598, %f25599; sqrt.approx.ftz.f32 %f25601, %f25600; div.approx.ftz.f32 %f25602, %f25601, %f6676; rcp.approx.ftz.f32 %f25603, %f6680; mul.ftz.f32 %f25604, %f25602, %f25602; mul.ftz.f32 %f25605, %f25604, %f25603; mul.ftz.f32 %f25606, %f25603, %f25605; sub.ftz.f32 %f25607, %f6679, %f25604; mul.ftz.f32 %f25608, %f25607, %f25603; sub.ftz.f32 %f25609, %f25606, %f25608; max.ftz.f32 %f25610, %f25609, %f25599; sqrt.approx.ftz.f32 %f25611, %f25610; mul.ftz.f32 %f25612, %f25602, %f25603; add.ftz.f32 %f25613, %f25612, %f25611; rcp.approx.ftz.f32 %f25614, %f25602; setp.gt.ftz.f32 %p3533, %f25613, %f25614; setp.lt.ftz.f32 %p3534, %f6678, 0f00000000; or.pred %p3535, %p3534, %p3533; sub.ftz.f32 %f25615, %f25612, %f25611; mov.f32 %f25616, 0f3F000000; sub.ftz.f32 %f25617, %f25616, %f6669; add.ftz.f32 %f25618, %f6669, 0fBF000000; setp.ge.ftz.f32 %p3536, %f6669, 0f3F000000; selp.f32 %f25619, %f25618, %f25617, %p3536; add.ftz.f32 %f25620, %f25619, %f25619; fma.rn.ftz.f32 %f25621, %f25620, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f25622, %f25620, %f25621, 0f3EED4413; mul.ftz.f32 %f25623, %f25620, %f25622; fma.rn.ftz.f32 %f25624, %f25620, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f25625, %f25620, %f25624, 0fBF800000; fma.rn.ftz.f32 %f25626, %f25620, %f25625, 0f3F191676; div.approx.ftz.f32 %f25627, %f25623, %f25626; selp.f32 %f30681, %f25615, %f25613, %p3535; fma.rn.ftz.f32 %f25628, %f30681, %f30681, 0f3F800000; sqrt.approx.ftz.f32 %f25629, %f25628; mul.ftz.f32 %f30682, %f25627, %f25629; setp.ltu.ftz.f32 %p3537, %f6669, 0f3F000000; @%p3537 bra $L__BB107_4172; bra.uni $L__BB107_4174; $L__BB107_4172: neg.ftz.f32 %f30682, %f30682; bra.uni $L__BB107_4174; $L__BB107_4185: add.ftz.f32 %f6735, %f6648, %f6713; setp.leu.ftz.f32 %p3549, %f6735, 0f00000000; mov.f32 %f30692, %f6654; @%p3549 bra $L__BB107_4187; sqrt.approx.ftz.f32 %f25703, %f6735; sub.ftz.f32 %f25704, %f25703, %f6712; add.ftz.f32 %f25705, %f6712, %f25703; div.approx.ftz.f32 %f25706, %f25704, %f25705; fma.rn.ftz.f32 %f25707, %f6712, %f25705, 0fBF800000; fma.rn.ftz.f32 %f25708, %f6712, %f25704, 0f3F800000; div.approx.ftz.f32 %f25709, %f25707, %f25708; mul.ftz.f32 %f25710, %f25706, 0f3F000000; mul.ftz.f32 %f25711, %f25706, %f25710; fma.rn.ftz.f32 %f25712, %f25709, %f25709, 0f3F800000; mul.ftz.f32 %f25713, %f25711, %f25712; sub.ftz.f32 %f30692, %f25713, %f30670; $L__BB107_4187: mul.ftz.f32 %f25714, %f6655, %f30692; mov.f32 %f25715, 0f3F800000; sub.ftz.f32 %f25716, %f25715, %f25714; fma.rn.ftz.f32 %f30693, %f6640, %f25716, %f25714; fma.rn.ftz.f32 %f30694, %f6641, %f25716, %f25714; fma.rn.ftz.f32 %f30695, %f6642, %f25716, %f25714; setp.eq.s32 %p3550, %r5872, 0; @%p3550 bra $L__BB107_4190; mul.ftz.f32 %f30693, %f6707, %f30693; mul.ftz.f32 %f30694, %f6708, %f30694; mul.ftz.f32 %f30695, %f6709, %f30695; $L__BB107_4190: setp.gt.ftz.f32 %p3551, %f6659, 0f3F7FF972; mov.f32 %f30696, 0f00000000; @%p3551 bra $L__BB107_4193; setp.lt.ftz.f32 %p3552, %f6659, 0fBF7FF972; mov.f32 %f30696, 0fBF7FF972; @%p3552 bra $L__BB107_4193; mul.ftz.f32 %f25719, %f6659, %f6659; mov.f32 %f25720, 0f33D6BF95; max.ftz.f32 %f25721, %f25719, %f25720; rcp.approx.ftz.f32 %f25722, %f25721; mul.ftz.f32 %f25723, %f25509, %f6661; mul.ftz.f32 %f25724, %f25510, %f6660; mul.ftz.f32 %f25725, %f25724, %f25724; fma.rn.ftz.f32 %f25726, %f25723, %f25723, %f25725; fma.rn.ftz.f32 %f25727, %f25726, %f25722, 0f3F800000; sqrt.approx.ftz.f32 %f25728, %f25727; neg.ftz.f32 %f25729, %f25728; setp.le.ftz.f32 %p3553, %f6659, 0f00000000; selp.f32 %f25730, %f25729, %f25728, %p3553; add.ftz.f32 %f25731, %f25730, 0fBF800000; mul.ftz.f32 %f30696, %f25731, 0f3F000000; $L__BB107_4193: mov.f32 %f30697, 0f3F800000; @%p3551 bra $L__BB107_4196; setp.lt.ftz.f32 %p3555, %f6659, 0f3727C5AC; mov.f32 %f30697, 0f00000000; @%p3555 bra $L__BB107_4196; lg2.approx.ftz.f32 %f25734, %f30680; mul.ftz.f32 %f25735, %f30696, %f25734; ex2.approx.ftz.f32 %f30697, %f25735; $L__BB107_4196: add.s32 %r5872, %r5872, 1; setp.lt.u32 %p3556, %r5872, 10; @%p3556 bra $L__BB107_3988; st.local.u32 [%rd5+168], %r5871; mov.f32 %f6661, 0f00000000; mov.f32 %f6659, 0f3F800000; mov.f32 %f6660, %f6661; mov.f32 %f30693, %f6661; mov.f32 %f30694, %f6661; mov.f32 %f30695, %f6661; bra.uni $L__BB107_4200; $L__BB107_4221: ld.local.v2.f32 {%f25966, %f25967}, [%rd5+-48]; ld.local.f32 %f6854, [%rd1027+36]; ld.local.f32 %f6855, [%rd1027+20]; ld.local.f32 %f25970, [%rd5+-52]; ld.local.f32 %f6856, [%rd1027+24]; mul.ftz.f32 %f25971, %f25966, %f6856; fma.rn.ftz.f32 %f25972, %f25970, %f6855, %f25971; ld.local.f32 %f6857, [%rd1027+28]; fma.rn.ftz.f32 %f25973, %f25967, %f6857, %f25972; mul.ftz.f32 %f25974, %f6855, %f25973; mul.ftz.f32 %f25975, %f6856, %f25973; mul.ftz.f32 %f25976, %f6857, %f25973; sub.ftz.f32 %f25977, %f25970, %f25974; sub.ftz.f32 %f25978, %f25966, %f25975; sub.ftz.f32 %f25979, %f25967, %f25976; mul.ftz.f32 %f25980, %f25978, %f25978; fma.rn.ftz.f32 %f25981, %f25977, %f25977, %f25980; fma.rn.ftz.f32 %f25982, %f25979, %f25979, %f25981; sqrt.approx.ftz.f32 %f25983, %f25982; div.approx.ftz.f32 %f6858, %f25977, %f25983; div.approx.ftz.f32 %f6859, %f25978, %f25983; div.approx.ftz.f32 %f6860, %f25979, %f25983; abs.ftz.f32 %f25984, %f25973; setp.lt.ftz.f32 %p3576, %f25984, 0f3F800000; mov.f32 %f25985, 0f3F800000; add.ftz.f32 %f25986, %f25984, 0fBF800000; add.ftz.f32 %f25987, %f25986, 0f3F800000; selp.f32 %f25988, %f25987, 0f3F800000, %p3576; sub.ftz.f32 %f25989, %f25985, %f25988; sqrt.approx.ftz.f32 %f25990, %f25989; fma.rn.ftz.f32 %f25991, %f25988, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f25992, %f25988, %f25991, 0fBE5A6B93; fma.rn.ftz.f32 %f25993, %f25988, %f25992, 0f3FC90FDB; mov.f32 %f25994, 0f3FC90FDB; mul.ftz.f32 %f25995, %f25990, %f25993; setp.lt.ftz.f32 %p3577, %f25973, 0f00000000; mov.f32 %f30729, 0f00000000; mov.f32 %f25996, 0f40490FDB; sub.ftz.f32 %f25997, %f25996, %f25995; selp.f32 %f25998, %f25997, %f25995, %p3577; sub.ftz.f32 %f6861, %f25994, %f25998; ld.local.f32 %f6862, [%rd1027+32]; rcp.approx.ftz.f32 %f6863, %f6862; add.ftz.f32 %f25999, %f6861, 0f3FC90FDB; mul.ftz.f32 %f26000, %f25999, 0f3F000000; ld.local.f32 %f6864, [%rd1027+40]; sub.ftz.f32 %f26001, %f26000, %f6864; mul.ftz.f32 %f6865, %f6863, %f26001; abs.ftz.f32 %f6866, %f6865; setp.eq.ftz.f32 %p3578, %f6866, 0f00000000; mov.f32 %f30728, %f30729; @%p3578 bra $L__BB107_4226; setp.eq.ftz.f32 %p3579, %f6866, 0f3F800000; mov.f32 %f30728, %f25985; @%p3579 bra $L__BB107_4226; setp.gt.ftz.f32 %p3580, %f6866, 0f3F800000; @%p3580 bra $L__BB107_4225; bra.uni $L__BB107_4224; $L__BB107_4225: mov.f32 %f26004, 0f3F800000; div.approx.ftz.f32 %f30728, %f26004, %f6866; bra.uni $L__BB107_4226; $L__BB107_4215: ld.local.v2.f32 {%f25929, %f25930}, [%rd5+-48]; ld.local.f32 %f6828, [%rd1027+28]; ld.local.f32 %f30719, [%rd1027+8]; ld.local.f32 %f6830, [%rd5+-52]; ld.local.f32 %f30720, [%rd1027+12]; mul.ftz.f32 %f25931, %f25929, %f30720; fma.rn.ftz.f32 %f25932, %f6830, %f30719, %f25931; ld.local.f32 %f30721, [%rd1027+16]; fma.rn.ftz.f32 %f30722, %f25930, %f30721, %f25932; setp.gt.ftz.f32 %p3571, %f30722, 0f00000000; @%p3571 bra $L__BB107_4217; bra.uni $L__BB107_4216; $L__BB107_4217: rcp.approx.ftz.f32 %f30723, %f6828; mov.u16 %rs744, 0; bra.uni $L__BB107_4218; $L__BB107_4014: ld.local.u64 %rd1028, [%rd1027+52]; ld.v4.f32 {%f24776, %f24777, %f24778, %f24779}, [%rd1028]; ld.local.v2.f32 {%f30626, %f30627}, [%rd5+32]; ld.local.f32 %f30625, [%rd5+28]; mul.ftz.f32 %f24782, %f30626, %f30626; fma.rn.ftz.f32 %f24783, %f30625, %f30625, %f24782; fma.rn.ftz.f32 %f24784, %f30627, %f30627, %f24783; sqrt.approx.ftz.f32 %f6275, %f24784; setp.neu.ftz.f32 %p3387, %f6275, 0f00000000; @%p3387 bra $L__BB107_4015; bra.uni $L__BB107_4016; $L__BB107_4015: rcp.approx.ftz.f32 %f24785, %f6275; mul.ftz.f32 %f30627, %f30627, %f24785; mul.ftz.f32 %f30626, %f30626, %f24785; mul.ftz.f32 %f30625, %f30625, %f24785; $L__BB107_4016: mul.ftz.f32 %f24786, %f24778, %f30626; mul.ftz.f32 %f24787, %f24777, %f30627; sub.ftz.f32 %f30628, %f24786, %f24787; mul.ftz.f32 %f24788, %f24776, %f30627; mul.ftz.f32 %f24789, %f24778, %f30625; sub.ftz.f32 %f30629, %f24788, %f24789; mul.ftz.f32 %f24790, %f24777, %f30625; mul.ftz.f32 %f24791, %f24776, %f30626; sub.ftz.f32 %f30630, %f24790, %f24791; mul.ftz.f32 %f24792, %f30629, %f30629; fma.rn.ftz.f32 %f24793, %f30628, %f30628, %f24792; fma.rn.ftz.f32 %f24794, %f30630, %f30630, %f24793; sqrt.approx.ftz.f32 %f6291, %f24794; setp.neu.ftz.f32 %p3388, %f6291, 0f00000000; @%p3388 bra $L__BB107_4017; bra.uni $L__BB107_4018; $L__BB107_4017: rcp.approx.ftz.f32 %f24795, %f6291; mul.ftz.f32 %f30630, %f24795, %f30630; mul.ftz.f32 %f30629, %f24795, %f30629; mul.ftz.f32 %f30628, %f24795, %f30628; $L__BB107_4018: ld.local.f32 %f24796, [%rd5+-52]; ld.local.v2.f32 {%f24797, %f24798}, [%rd5+-48]; mul.ftz.f32 %f24801, %f30626, %f24797; fma.rn.ftz.f32 %f24802, %f30625, %f24796, %f24801; fma.rn.ftz.f32 %f6301, %f30627, %f24798, %f24802; mul.ftz.f32 %f24803, %f24777, %f24797; fma.rn.ftz.f32 %f24804, %f24776, %f24796, %f24803; fma.rn.ftz.f32 %f24805, %f24778, %f24798, %f24804; mul.ftz.f32 %f24806, %f30629, %f24797; fma.rn.ftz.f32 %f24807, %f30628, %f24796, %f24806; fma.rn.ftz.f32 %f24808, %f30630, %f24798, %f24807; ld.local.u32 %r5201, [%rd5+168]; mad.lo.s32 %r1525, %r5201, 1103515245, 12345; mad.lo.s32 %r1526, %r1525, 1103515245, 12345; st.local.u32 [%rd5+168], %r1526; mul.ftz.f32 %f6305, %f6301, %f6301; mov.f32 %f24809, 0f3F800000; sub.ftz.f32 %f24810, %f24809, %f6305; mov.f32 %f30636, 0f00000000; max.ftz.f32 %f24812, %f24810, %f30636; sqrt.approx.ftz.f32 %f6306, %f24812; abs.ftz.f32 %f6307, %f24805; abs.ftz.f32 %f6308, %f24808; setp.eq.ftz.f32 %p3389, %f6307, 0f00000000; setp.eq.ftz.f32 %p3390, %f6308, 0f00000000; and.pred %p3391, %p3389, %p3390; mov.b32 %r1527, %f24805; mov.b32 %r5202, %f24808; and.b32 %r1528, %r5202, -2147483648; @%p3391 bra $L__BB107_4022; bra.uni $L__BB107_4019; $L__BB107_4022: shr.s32 %r5207, %r1527, 31; and.b32 %r5208, %r5207, 1078530011; or.b32 %r5209, %r1528, %r5208; mov.b32 %f30631, %r5209; bra.uni $L__BB107_4023; $L__BB107_3998: ld.local.v2.f32 {%f24568, %f24569}, [%rd5+-48]; ld.local.f32 %f6239, [%rd1027+36]; ld.local.f32 %f6240, [%rd1027+20]; ld.local.f32 %f24572, [%rd5+-52]; ld.local.f32 %f6241, [%rd1027+24]; mul.ftz.f32 %f24573, %f24568, %f6241; fma.rn.ftz.f32 %f24574, %f24572, %f6240, %f24573; ld.local.f32 %f6242, [%rd1027+28]; fma.rn.ftz.f32 %f24575, %f24569, %f6242, %f24574; mul.ftz.f32 %f24576, %f6240, %f24575; mul.ftz.f32 %f24577, %f6241, %f24575; mul.ftz.f32 %f24578, %f6242, %f24575; sub.ftz.f32 %f24579, %f24572, %f24576; sub.ftz.f32 %f24580, %f24568, %f24577; sub.ftz.f32 %f24581, %f24569, %f24578; mul.ftz.f32 %f24582, %f24580, %f24580; fma.rn.ftz.f32 %f24583, %f24579, %f24579, %f24582; fma.rn.ftz.f32 %f24584, %f24581, %f24581, %f24583; sqrt.approx.ftz.f32 %f24585, %f24584; div.approx.ftz.f32 %f6243, %f24579, %f24585; div.approx.ftz.f32 %f6244, %f24580, %f24585; div.approx.ftz.f32 %f6245, %f24581, %f24585; abs.ftz.f32 %f24586, %f24575; setp.lt.ftz.f32 %p3366, %f24586, 0f3F800000; mov.f32 %f24587, 0f3F800000; add.ftz.f32 %f24588, %f24586, 0fBF800000; add.ftz.f32 %f24589, %f24588, 0f3F800000; selp.f32 %f24590, %f24589, 0f3F800000, %p3366; sub.ftz.f32 %f24591, %f24587, %f24590; sqrt.approx.ftz.f32 %f24592, %f24591; fma.rn.ftz.f32 %f24593, %f24590, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f24594, %f24590, %f24593, 0fBE5A6B93; fma.rn.ftz.f32 %f24595, %f24590, %f24594, 0f3FC90FDB; mov.f32 %f24596, 0f3FC90FDB; mul.ftz.f32 %f24597, %f24592, %f24595; setp.lt.ftz.f32 %p3367, %f24575, 0f00000000; mov.f32 %f30624, 0f00000000; mov.f32 %f24598, 0f40490FDB; sub.ftz.f32 %f24599, %f24598, %f24597; selp.f32 %f24600, %f24599, %f24597, %p3367; sub.ftz.f32 %f6246, %f24596, %f24600; ld.local.f32 %f6247, [%rd1027+32]; rcp.approx.ftz.f32 %f6248, %f6247; add.ftz.f32 %f24601, %f6246, 0f3FC90FDB; mov.f32 %f24602, 0f40000000; div.approx.ftz.f32 %f24603, %f24601, %f24602; ld.local.f32 %f6249, [%rd1027+40]; sub.ftz.f32 %f24604, %f24603, %f6249; mul.ftz.f32 %f6250, %f6248, %f24604; abs.ftz.f32 %f6251, %f6250; setp.eq.ftz.f32 %p3368, %f6251, 0f00000000; mov.f32 %f30622, %f30624; @%p3368 bra $L__BB107_4003; setp.eq.ftz.f32 %p3369, %f6251, 0f3F800000; mov.f32 %f30622, %f24587; @%p3369 bra $L__BB107_4003; setp.gt.ftz.f32 %p3370, %f6251, 0f3F800000; @%p3370 bra $L__BB107_4002; bra.uni $L__BB107_4001; $L__BB107_4002: mov.f32 %f24607, 0f3F800000; div.approx.ftz.f32 %f30622, %f24607, %f6251; bra.uni $L__BB107_4003; $L__BB107_4431: mul.ftz.f32 %f28405, %f6226, 0f40C90FDB; cos.approx.ftz.f32 %f28406, %f28405; sqrt.approx.ftz.f32 %f28407, %f6202; mul.ftz.f32 %f28408, %f28407, %f28406; sin.approx.ftz.f32 %f28409, %f28405; mul.ftz.f32 %f28410, %f28407, %f28409; mul.ftz.f32 %f28411, %f28408, %f28408; mov.f32 %f28412, 0f3F800000; sub.ftz.f32 %f28413, %f28412, %f28411; mul.ftz.f32 %f28414, %f28410, %f28410; sub.ftz.f32 %f28415, %f28413, %f28414; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28416, %f28415, %f30902; sqrt.approx.ftz.f32 %f7549, %f28416; ld.local.f32 %f28417, [%rd1027+12]; ld.local.f32 %f28418, [%rd1027+8]; setp.neu.ftz.f32 %p3793, %f28418, %f28417; ld.local.f32 %f28419, [%rd1027+16]; setp.neu.ftz.f32 %p3794, %f28418, %f28419; or.pred %p3795, %p3793, %p3794; sub.ftz.f32 %f28420, %f28419, %f28417; neg.ftz.f32 %f28421, %f28417; neg.ftz.f32 %f28422, %f28419; selp.f32 %f28423, %f28422, %f28419, %p3795; add.ftz.f32 %f28424, %f28418, %f28423; selp.f32 %f28425, %f28417, %f28421, %p3795; sub.ftz.f32 %f28426, %f28425, %f28418; mul.ftz.f32 %f28427, %f28424, %f28424; fma.rn.ftz.f32 %f28428, %f28420, %f28420, %f28427; fma.rn.ftz.f32 %f28429, %f28426, %f28426, %f28428; sqrt.approx.ftz.f32 %f28430, %f28429; div.approx.ftz.f32 %f28431, %f28420, %f28430; div.approx.ftz.f32 %f28432, %f28424, %f28430; div.approx.ftz.f32 %f28433, %f28426, %f28430; mul.ftz.f32 %f28434, %f28417, %f28433; mul.ftz.f32 %f28435, %f28419, %f28432; sub.ftz.f32 %f28436, %f28434, %f28435; mul.ftz.f32 %f28437, %f28419, %f28431; mul.ftz.f32 %f28438, %f28418, %f28433; sub.ftz.f32 %f28439, %f28437, %f28438; mul.ftz.f32 %f28440, %f28418, %f28432; mul.ftz.f32 %f28441, %f28417, %f28431; sub.ftz.f32 %f28442, %f28440, %f28441; mul.ftz.f32 %f28443, %f28410, %f28436; mul.ftz.f32 %f28444, %f28410, %f28439; mul.ftz.f32 %f28445, %f28410, %f28442; fma.rn.ftz.f32 %f28446, %f28408, %f28431, %f28443; fma.rn.ftz.f32 %f28447, %f28408, %f28432, %f28444; fma.rn.ftz.f32 %f28448, %f28408, %f28433, %f28445; fma.rn.ftz.f32 %f30899, %f28418, %f7549, %f28446; fma.rn.ftz.f32 %f30900, %f28417, %f7549, %f28447; fma.rn.ftz.f32 %f30901, %f28419, %f7549, %f28448; mul.ftz.f32 %f28449, %f6234, %f30900; fma.rn.ftz.f32 %f28450, %f6233, %f30899, %f28449; fma.rn.ftz.f32 %f28451, %f6235, %f30901, %f28450; setp.leu.ftz.f32 %p3796, %f28451, 0f00000000; mov.u32 %r5877, 6; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3796 bra $L__BB107_4433; mul.ftz.f32 %f30902, %f7549, 0f3EA2F983; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4427: mul.ftz.f32 %f28337, %f6226, %f6226; mov.f32 %f28338, 0f3F800000; sub.ftz.f32 %f28339, %f28338, %f28337; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28340, %f30902, %f28339; sqrt.approx.ftz.f32 %f28341, %f28340; mul.ftz.f32 %f28342, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f28343, %f28342; mul.ftz.f32 %f28344, %f28341, %f28343; sin.approx.ftz.f32 %f28345, %f28342; mul.ftz.f32 %f28346, %f28341, %f28345; ld.local.f32 %f7537, [%rd1027+12]; ld.local.f32 %f7538, [%rd1027+8]; setp.neu.ftz.f32 %p3788, %f7538, %f7537; ld.local.f32 %f7539, [%rd1027+16]; setp.neu.ftz.f32 %p3789, %f7538, %f7539; or.pred %p3790, %p3788, %p3789; sub.ftz.f32 %f28347, %f7539, %f7537; neg.ftz.f32 %f28348, %f7537; neg.ftz.f32 %f28349, %f7539; selp.f32 %f28350, %f28349, %f7539, %p3790; add.ftz.f32 %f28351, %f7538, %f28350; selp.f32 %f28352, %f7537, %f28348, %p3790; sub.ftz.f32 %f28353, %f28352, %f7538; mul.ftz.f32 %f28354, %f28351, %f28351; fma.rn.ftz.f32 %f28355, %f28347, %f28347, %f28354; fma.rn.ftz.f32 %f28356, %f28353, %f28353, %f28355; sqrt.approx.ftz.f32 %f28357, %f28356; div.approx.ftz.f32 %f28358, %f28347, %f28357; div.approx.ftz.f32 %f28359, %f28351, %f28357; div.approx.ftz.f32 %f28360, %f28353, %f28357; mul.ftz.f32 %f28361, %f7537, %f28360; mul.ftz.f32 %f28362, %f7539, %f28359; sub.ftz.f32 %f28363, %f28361, %f28362; mul.ftz.f32 %f28364, %f7539, %f28358; mul.ftz.f32 %f28365, %f7538, %f28360; sub.ftz.f32 %f28366, %f28364, %f28365; mul.ftz.f32 %f28367, %f7538, %f28359; mul.ftz.f32 %f28368, %f7537, %f28358; sub.ftz.f32 %f28369, %f28367, %f28368; mul.ftz.f32 %f28370, %f28346, %f28363; mul.ftz.f32 %f28371, %f28346, %f28366; mul.ftz.f32 %f28372, %f28346, %f28369; fma.rn.ftz.f32 %f28373, %f28344, %f28358, %f28370; fma.rn.ftz.f32 %f28374, %f28344, %f28359, %f28371; fma.rn.ftz.f32 %f28375, %f28344, %f28360, %f28372; fma.rn.ftz.f32 %f30899, %f6226, %f7538, %f28373; fma.rn.ftz.f32 %f30900, %f6226, %f7537, %f28374; fma.rn.ftz.f32 %f30901, %f6226, %f7539, %f28375; mul.ftz.f32 %f28376, %f6234, %f30900; fma.rn.ftz.f32 %f28377, %f6233, %f30899, %f28376; fma.rn.ftz.f32 %f28378, %f6235, %f30901, %f28377; setp.leu.ftz.f32 %p3791, %f28378, 0f00000000; mov.u32 %r5877, 6; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3791 bra $L__BB107_4433; ld.local.f32 %f28379, [%rd5+-52]; ld.local.v2.f32 {%f28380, %f28381}, [%rd5+-48]; mul.ftz.f32 %f28384, %f30900, %f7537; fma.rn.ftz.f32 %f28385, %f30899, %f7538, %f28384; fma.rn.ftz.f32 %f28386, %f30901, %f7539, %f28385; mov.f32 %f28387, 0f00000000; max.ftz.f32 %f7543, %f28386, %f28387; mul.ftz.f32 %f28388, %f28380, %f7537; fma.rn.ftz.f32 %f28389, %f28379, %f7538, %f28388; fma.rn.ftz.f32 %f28390, %f28381, %f7539, %f28389; max.ftz.f32 %f7544, %f28390, %f28387; mul.ftz.f32 %f28391, %f28380, %f30900; fma.rn.ftz.f32 %f28392, %f28379, %f30899, %f28391; fma.rn.ftz.f32 %f28393, %f28381, %f30901, %f28392; mul.ftz.f32 %f28394, %f7543, %f7544; sub.ftz.f32 %f30898, %f28393, %f28394; setp.leu.ftz.f32 %p3792, %f30898, 0f00000000; @%p3792 bra $L__BB107_4430; max.ftz.f32 %f28395, %f7543, %f7544; add.ftz.f32 %f28396, %f28395, 0f00800000; div.approx.ftz.f32 %f30898, %f30898, %f28396; $L__BB107_4430: ld.local.f32 %f28398, [%rd1027+28]; ld.local.f32 %f28399, [%rd1027+24]; fma.rn.ftz.f32 %f28400, %f30898, %f28398, %f28399; mul.ftz.f32 %f30903, %f7543, %f28400; mov.f32 %f30902, 0f3E22F983; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_4416: ld.local.f32 %f28052, [%rd1027+8]; ld.local.f32 %f28053, [%rd1027+12]; neg.ftz.f32 %f28054, %f28053; ld.local.f32 %f28055, [%rd1027+16]; neg.ftz.f32 %f28056, %f28055; mul.ftz.f32 %f28057, %f6226, 0f40C90FDB; cos.approx.ftz.f32 %f28058, %f28057; sqrt.approx.ftz.f32 %f28059, %f6202; mul.ftz.f32 %f28060, %f28059, %f28058; sin.approx.ftz.f32 %f28061, %f28057; mul.ftz.f32 %f28062, %f28059, %f28061; mul.ftz.f32 %f28063, %f28060, %f28060; mov.f32 %f28064, 0f3F800000; sub.ftz.f32 %f28065, %f28064, %f28063; mul.ftz.f32 %f28066, %f28062, %f28062; sub.ftz.f32 %f28067, %f28065, %f28066; mov.f32 %f30902, 0f00000000; max.ftz.f32 %f28068, %f28067, %f30902; sqrt.approx.ftz.f32 %f7469, %f28068; setp.neu.ftz.f32 %p3760, %f28052, %f28053; setp.neu.ftz.f32 %p3761, %f28052, %f28055; or.pred %p3762, %p3760, %p3761; sub.ftz.f32 %f28069, %f28053, %f28055; selp.f32 %f28070, %f28055, %f28056, %p3762; selp.f32 %f28071, %f28054, %f28053, %p3762; add.ftz.f32 %f28072, %f28052, %f28071; sub.ftz.f32 %f28073, %f28070, %f28052; mul.ftz.f32 %f28074, %f28073, %f28073; fma.rn.ftz.f32 %f28075, %f28069, %f28069, %f28074; fma.rn.ftz.f32 %f28076, %f28072, %f28072, %f28075; sqrt.approx.ftz.f32 %f28077, %f28076; div.approx.ftz.f32 %f28078, %f28069, %f28077; div.approx.ftz.f32 %f28079, %f28073, %f28077; div.approx.ftz.f32 %f28080, %f28072, %f28077; mul.ftz.f32 %f28081, %f28080, %f28053; mul.ftz.f32 %f28082, %f28079, %f28055; sub.ftz.f32 %f28083, %f28082, %f28081; mul.ftz.f32 %f28084, %f28078, %f28055; mul.ftz.f32 %f28085, %f28080, %f28052; sub.ftz.f32 %f28086, %f28085, %f28084; mul.ftz.f32 %f28087, %f28079, %f28052; mul.ftz.f32 %f28088, %f28078, %f28053; sub.ftz.f32 %f28089, %f28088, %f28087; mul.ftz.f32 %f28090, %f28062, %f28083; mul.ftz.f32 %f28091, %f28062, %f28086; mul.ftz.f32 %f28092, %f28062, %f28089; fma.rn.ftz.f32 %f28093, %f28060, %f28078, %f28090; fma.rn.ftz.f32 %f28094, %f28060, %f28079, %f28091; fma.rn.ftz.f32 %f28095, %f28060, %f28080, %f28092; mul.ftz.f32 %f28096, %f7469, %f28052; mul.ftz.f32 %f28097, %f7469, %f28053; mul.ftz.f32 %f28098, %f7469, %f28055; sub.ftz.f32 %f30899, %f28093, %f28096; sub.ftz.f32 %f30900, %f28094, %f28097; sub.ftz.f32 %f30901, %f28095, %f28098; mul.ftz.f32 %f28099, %f6234, %f30900; fma.rn.ftz.f32 %f28100, %f6233, %f30899, %f28099; fma.rn.ftz.f32 %f28101, %f6235, %f30901, %f28100; setp.geu.ftz.f32 %p3763, %f28101, 0f00000000; mov.u32 %r5877, 5; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3763 bra $L__BB107_4433; mul.ftz.f32 %f30902, %f7469, 0f3EA2F983; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_3976: setp.eq.s32 %p3692, %r1522, 22; selp.b32 %r5877, 9, 10, %p3692; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4373: ld.local.f32 %f27629, [%rd1027+52]; mul.ftz.f32 %f27630, %f7327, %f27629; ld.local.f32 %f27631, [%rd1027+48]; mul.ftz.f32 %f27632, %f7328, %f27631; sub.ftz.f32 %f27633, %f27630, %f27632; ld.local.f32 %f27634, [%rd1027+44]; mul.ftz.f32 %f27635, %f7328, %f27634; mul.ftz.f32 %f27636, %f7325, %f27629; sub.ftz.f32 %f27637, %f27635, %f27636; mul.ftz.f32 %f27638, %f7325, %f27631; mul.ftz.f32 %f27639, %f7327, %f27634; sub.ftz.f32 %f27640, %f27638, %f27639; mul.ftz.f32 %f27641, %f27637, %f27637; fma.rn.ftz.f32 %f27642, %f27633, %f27633, %f27641; fma.rn.ftz.f32 %f27643, %f27640, %f27640, %f27642; sqrt.approx.ftz.f32 %f27644, %f27643; div.approx.ftz.f32 %f30852, %f27633, %f27644; div.approx.ftz.f32 %f30853, %f27637, %f27644; div.approx.ftz.f32 %f30854, %f27640, %f27644; mul.ftz.f32 %f27645, %f7328, %f30853; mul.ftz.f32 %f27646, %f7327, %f30854; sub.ftz.f32 %f30855, %f27645, %f27646; mul.ftz.f32 %f27647, %f7325, %f30854; mul.ftz.f32 %f27648, %f7328, %f30852; sub.ftz.f32 %f30856, %f27647, %f27648; mul.ftz.f32 %f27649, %f7327, %f30852; mul.ftz.f32 %f27650, %f7325, %f30853; sub.ftz.f32 %f30857, %f27649, %f27650; $L__BB107_4375: mul.ftz.f32 %f27672, %f27618, %f30856; fma.rn.ftz.f32 %f27673, %f7326, %f30855, %f27672; fma.rn.ftz.f32 %f27674, %f27619, %f30857, %f27673; mul.ftz.f32 %f27675, %f27618, %f30853; fma.rn.ftz.f32 %f27676, %f7326, %f30852, %f27675; fma.rn.ftz.f32 %f27677, %f27619, %f30854, %f27676; mul.ftz.f32 %f27678, %f27620, %f27674; mul.ftz.f32 %f27679, %f27621, %f27677; mul.ftz.f32 %f27680, %f27678, %f27678; fma.rn.ftz.f32 %f27681, %f27679, %f27679, %f27680; mul.ftz.f32 %f7348, %f7329, %f7329; add.ftz.f32 %f27682, %f7348, %f27681; sqrt.approx.ftz.f32 %f27683, %f27682; div.approx.ftz.f32 %f7349, %f27678, %f27683; div.approx.ftz.f32 %f7350, %f27679, %f27683; div.approx.ftz.f32 %f7351, %f7329, %f27683; setp.geu.ftz.f32 %p3731, %f7351, 0f3F7FFF58; mov.f32 %f30859, 0f00000000; mov.f32 %f30858, 0f3F800000; mov.f32 %f30860, %f30858; mov.f32 %f30861, %f30859; @%p3731 bra $L__BB107_4377; mul.ftz.f32 %f27684, %f7351, %f7351; mov.f32 %f27685, 0f3F800000; sub.ftz.f32 %f27686, %f27685, %f27684; mov.f32 %f27687, 0f00000000; max.ftz.f32 %f27688, %f27686, %f27687; sqrt.approx.ftz.f32 %f30859, %f27688; rsqrt.approx.ftz.f32 %f27689, %f27688; mul.ftz.f32 %f30860, %f7349, %f27689; mul.ftz.f32 %f30861, %f7350, %f27689; mov.f32 %f30858, %f7351; $L__BB107_4377: setp.ltu.ftz.f32 %p3732, %f30858, 0f3F7FFF58; @%p3732 bra $L__BB107_4379; bra.uni $L__BB107_4378; $L__BB107_4379: div.approx.ftz.f32 %f27697, %f30859, %f30858; mul.ftz.f32 %f27698, %f27697, %f27697; add.ftz.f32 %f27699, %f27698, 0f3F800000; mov.f32 %f27700, 0f00000000; max.ftz.f32 %f27701, %f27699, %f27700; sqrt.approx.ftz.f32 %f27702, %f27701; add.ftz.f32 %f27703, %f27702, 0f3F800000; mul.ftz.f32 %f27704, %f27703, 0f3F000000; mov.f32 %f27705, 0f3F000000; rcp.approx.ftz.f32 %f7364, %f27704; add.ftz.f32 %f27706, %f6226, %f6226; fma.rn.ftz.f32 %f27707, %f27706, %f27704, 0fBF800000; mul.ftz.f32 %f27708, %f27707, %f27707; add.ftz.f32 %f27709, %f27708, 0fBF800000; rcp.approx.ftz.f32 %f27710, %f27709; mul.ftz.f32 %f27711, %f27710, %f27710; mul.ftz.f32 %f27712, %f27698, %f27711; sub.ftz.f32 %f27713, %f27708, %f27698; mul.ftz.f32 %f27714, %f27713, %f27710; sub.ftz.f32 %f27715, %f27712, %f27714; max.ftz.f32 %f27716, %f27715, %f27700; sqrt.approx.ftz.f32 %f27717, %f27716; mul.ftz.f32 %f27718, %f27697, %f27710; sub.ftz.f32 %f27719, %f27718, %f27717; add.ftz.f32 %f27720, %f27717, %f27718; setp.lt.ftz.f32 %p3733, %f27707, 0f00000000; mul.ftz.f32 %f27721, %f27697, %f27720; setp.gt.ftz.f32 %p3734, %f27721, 0f3F800000; or.pred %p3735, %p3733, %p3734; selp.f32 %f30864, %f27719, %f27720, %p3735; sub.ftz.f32 %f27722, %f27705, %f6202; add.ftz.f32 %f27723, %f6202, 0fBF000000; setp.gt.ftz.f32 %p3736, %f6202, 0f3F000000; selp.f32 %f27724, %f27723, %f27722, %p3736; add.ftz.f32 %f27725, %f27724, %f27724; selp.f32 %f27726, 0f3F800000, 0fBF800000, %p3736; fma.rn.ftz.f32 %f27727, %f27725, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f27728, %f27725, %f27727, 0f3EED4413; mul.ftz.f32 %f27729, %f27725, %f27728; fma.rn.ftz.f32 %f27730, %f27725, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f27731, %f27725, %f27730, 0fBF800000; fma.rn.ftz.f32 %f27732, %f27725, %f27731, 0f3F191676; div.approx.ftz.f32 %f27733, %f27729, %f27732; mul.ftz.f32 %f27734, %f27726, %f27733; fma.rn.ftz.f32 %f27735, %f30864, %f30864, 0f3F800000; max.ftz.f32 %f27736, %f27735, %f27700; sqrt.approx.ftz.f32 %f27737, %f27736; mul.ftz.f32 %f30863, %f27734, %f27737; bra.uni $L__BB107_4380; $L__BB107_4378: mov.f32 %f7364, 0f3F800000; sub.ftz.f32 %f27691, %f7364, %f6226; div.approx.ftz.f32 %f27692, %f6226, %f27691; sqrt.approx.ftz.f32 %f27693, %f27692; mul.ftz.f32 %f27694, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f27695, %f27694; mul.ftz.f32 %f30864, %f27693, %f27695; sin.approx.ftz.f32 %f27696, %f27694; mul.ftz.f32 %f30863, %f27693, %f27696; $L__BB107_4380: mul.ftz.f32 %f27738, %f30861, %f30863; mul.ftz.f32 %f27739, %f30860, %f30864; sub.ftz.f32 %f27740, %f27739, %f27738; mul.ftz.f32 %f27741, %f30861, %f30864; fma.rn.ftz.f32 %f27742, %f30860, %f30863, %f27741; mul.ftz.f32 %f27743, %f27620, %f27740; mul.ftz.f32 %f27744, %f27621, %f27742; neg.ftz.f32 %f27745, %f27743; neg.ftz.f32 %f27746, %f27744; mul.ftz.f32 %f27747, %f27744, %f27744; fma.rn.ftz.f32 %f27748, %f27743, %f27743, %f27747; add.ftz.f32 %f27749, %f27748, 0f3F800000; mov.f32 %f27750, 0f3F800000; sqrt.approx.ftz.f32 %f27751, %f27749; div.approx.ftz.f32 %f27752, %f27745, %f27751; div.approx.ftz.f32 %f27753, %f27746, %f27751; div.approx.ftz.f32 %f7367, %f27750, %f27751; mul.ftz.f32 %f27754, %f30852, %f27753; mul.ftz.f32 %f27755, %f30853, %f27753; mul.ftz.f32 %f27756, %f30854, %f27753; fma.rn.ftz.f32 %f27757, %f30855, %f27752, %f27754; fma.rn.ftz.f32 %f27758, %f30856, %f27752, %f27755; fma.rn.ftz.f32 %f27759, %f30857, %f27752, %f27756; fma.rn.ftz.f32 %f7368, %f7325, %f7367, %f27757; fma.rn.ftz.f32 %f7369, %f7327, %f7367, %f27758; fma.rn.ftz.f32 %f7370, %f7328, %f7367, %f27759; mul.ftz.f32 %f27760, %f27618, %f7369; fma.rn.ftz.f32 %f27761, %f7326, %f7368, %f27760; fma.rn.ftz.f32 %f7371, %f27619, %f7370, %f27761; setp.eq.s32 %p3737, %r1522, 23; @%p3737 bra $L__BB107_4407; bra.uni $L__BB107_4381; $L__BB107_4407: setp.gt.ftz.f32 %p3751, %f7371, 0f00000000; ld.local.f32 %f7436, [%rd1027+28]; @%p3751 bra $L__BB107_4409; bra.uni $L__BB107_4408; $L__BB107_4409: rcp.approx.ftz.f32 %f30887, %f7436; mov.u16 %rs746, 0; mov.f32 %f30883, %f7368; mov.f32 %f30884, %f7369; mov.f32 %f30885, %f7370; mov.f32 %f30886, %f7371; bra.uni $L__BB107_4410; $L__BB107_4381: setp.leu.ftz.f32 %p3738, %f7371, 0f00000000; mov.u32 %r5877, 10; mov.f32 %f30902, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3738 bra $L__BB107_4433; add.ftz.f32 %f27771, %f7371, %f7371; mul.ftz.f32 %f27772, %f7368, %f27771; mul.ftz.f32 %f27773, %f7369, %f27771; mul.ftz.f32 %f27774, %f7370, %f27771; sub.ftz.f32 %f30899, %f27772, %f7326; sub.ftz.f32 %f30900, %f27773, %f27618; sub.ftz.f32 %f30901, %f27774, %f27619; mul.ftz.f32 %f27775, %f6234, %f30900; fma.rn.ftz.f32 %f27776, %f6233, %f30899, %f27775; fma.rn.ftz.f32 %f27777, %f6235, %f30901, %f27776; setp.leu.ftz.f32 %p3739, %f27777, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3739 bra $L__BB107_4433; mul.ftz.f32 %f30869, %f27620, %f27621; setp.gtu.ftz.f32 %p3740, %f30869, 0f33D6BF95; @%p3740 bra $L__BB107_4390; bra.uni $L__BB107_4384; $L__BB107_4390: mul.ftz.f32 %f27817, %f7327, %f30900; fma.rn.ftz.f32 %f27818, %f7325, %f30899, %f27817; fma.rn.ftz.f32 %f7392, %f7328, %f30901, %f27818; @%p3727 bra $L__BB107_4392; bra.uni $L__BB107_4391; $L__BB107_4392: setp.eq.s32 %p3745, %r1522, 12; @%p3745 bra $L__BB107_4394; bra.uni $L__BB107_4393; $L__BB107_4394: setp.ge.ftz.f32 %p3746, %f27620, 0f3F800000; mov.f32 %f30870, 0f3EA2F983; @%p3746 bra $L__BB107_4396; mul.ftz.f32 %f27869, %f27620, %f27620; add.ftz.f32 %f27870, %f27869, 0fBF800000; mul.ftz.f32 %f27871, %f7367, %f27870; fma.rn.ftz.f32 %f27872, %f7367, %f27871, 0f3F800000; lg2.approx.ftz.f32 %f27873, %f27869; mul.ftz.f32 %f27874, %f27873, 0f3F317218; mul.ftz.f32 %f27875, %f27874, 0f40490FDB; mul.ftz.f32 %f27876, %f27872, %f27875; div.approx.ftz.f32 %f30870, %f27870, %f27876; $L__BB107_4396: mov.f32 %f27878, 0f3F800000; sub.ftz.f32 %f27879, %f27878, %f7348; mul.ftz.f32 %f27880, %f27879, 0f3D800000; mov.f32 %f30869, 0f3D800000; div.approx.ftz.f32 %f27881, %f27880, %f7348; add.ftz.f32 %f27882, %f27881, 0f3F800000; mov.f32 %f27883, 0f00000000; max.ftz.f32 %f27884, %f27882, %f27883; sqrt.approx.ftz.f32 %f27885, %f27884; add.ftz.f32 %f27886, %f27885, 0f3F800000; mov.f32 %f27887, 0f40000000; div.approx.ftz.f32 %f7364, %f27887, %f27886; bra.uni $L__BB107_4397; $L__BB107_4198: st.local.u32 [%rd5+168], %r1534; and.b16 %rs591, %rs742, 255; setp.eq.s16 %p3557, %rs591, 0; @%p3557 bra $L__BB107_4199; bra.uni $L__BB107_4200; $L__BB107_4199: neg.ftz.f32 %f6659, %f6659; neg.ftz.f32 %f6660, %f6660; neg.ftz.f32 %f6661, %f6661; $L__BB107_4200: ld.local.f32 %f6775, [%rd1027+20]; mul.ftz.f32 %f6776, %f30713, %f6659; setp.gt.ftz.f32 %p3558, %f6776, 0f00000000; mov.f32 %f30707, 0f00000000; mul.ftz.f32 %f25743, %f6661, %f6639; selp.f32 %f25744, %f6661, %f25743, %p3558; mul.ftz.f32 %f25745, %f6660, %f6639; selp.f32 %f25746, %f6660, %f25745, %p3558; mul.ftz.f32 %f25747, %f6659, %f6639; selp.f32 %f25748, %f6659, %f25747, %p3558; add.ftz.f32 %f25749, %f6634, %f25744; add.ftz.f32 %f25750, %f6635, %f25746; add.ftz.f32 %f25751, %f30713, %f25748; mul.ftz.f32 %f25752, %f25750, %f25750; fma.rn.ftz.f32 %f25753, %f25749, %f25749, %f25752; fma.rn.ftz.f32 %f25754, %f25751, %f25751, %f25753; sqrt.approx.ftz.f32 %f6777, %f25754; rsqrt.approx.ftz.f32 %f25755, %f25754; mul.ftz.f32 %f25756, %f25755, %f25749; mul.ftz.f32 %f25757, %f25755, %f25750; mul.ftz.f32 %f25758, %f25755, %f25751; setp.lt.ftz.f32 %p3559, %f25758, 0f00000000; neg.ftz.f32 %f25759, %f25756; neg.ftz.f32 %f25760, %f25757; neg.ftz.f32 %f25761, %f25758; selp.f32 %f6778, %f25759, %f25756, %p3559; selp.f32 %f6779, %f25760, %f25757, %p3559; selp.f32 %f6780, %f25761, %f25758, %p3559; setp.lt.ftz.f32 %p3560, %f30713, 0f00000000; selp.f32 %f6781, %f6643, %f6634, %p3560; selp.f32 %f6782, %f6644, %f6635, %p3560; selp.f32 %f6783, %f6645, %f30713, %p3560; setp.gt.ftz.f32 %p3561, %f6783, 0f3F7FF972; @%p3561 bra $L__BB107_4203; setp.lt.ftz.f32 %p3562, %f6783, 0fBF7FF972; mov.f32 %f30707, 0fBF7FF972; @%p3562 bra $L__BB107_4203; mul.ftz.f32 %f25763, %f6783, %f6783; mov.f32 %f25764, 0f33D6BF95; max.ftz.f32 %f25765, %f25763, %f25764; rcp.approx.ftz.f32 %f25766, %f25765; mul.ftz.f32 %f25767, %f6781, %f6775; mul.ftz.f32 %f25768, %f6782, %f6775; mul.ftz.f32 %f25769, %f25768, %f25768; fma.rn.ftz.f32 %f25770, %f25767, %f25767, %f25769; fma.rn.ftz.f32 %f25771, %f25770, %f25766, 0f3F800000; sqrt.approx.ftz.f32 %f25772, %f25771; neg.ftz.f32 %f25773, %f25772; setp.le.ftz.f32 %p3563, %f6783, 0f00000000; selp.f32 %f25774, %f25773, %f25772, %p3563; add.ftz.f32 %f25775, %f25774, 0fBF800000; mul.ftz.f32 %f30707, %f25775, 0f3F000000; $L__BB107_4203: mul.ftz.f32 %f25777, %f6780, %f6780; mov.f32 %f30708, 0f3F800000; sub.ftz.f32 %f25778, %f30708, %f25777; mul.ftz.f32 %f25779, %f6775, %f6775; fma.rn.ftz.f32 %f25780, %f25779, %f25777, %f25778; mul.ftz.f32 %f25781, %f25780, 0f40490FDB; mul.ftz.f32 %f25782, %f25780, %f25781; mov.f32 %f25783, 0f33D6BF95; max.ftz.f32 %f25784, %f25782, %f25783; div.approx.ftz.f32 %f6786, %f25779, %f25784; mul.ftz.f32 %f25785, %f6782, %f6779; fma.rn.ftz.f32 %f25786, %f6781, %f6778, %f25785; fma.rn.ftz.f32 %f6787, %f6783, %f6780, %f25786; abs.ftz.f32 %f6788, %f6787; fma.rn.ftz.f32 %f6789, %f6788, %f6788, %f6648; setp.leu.ftz.f32 %p3564, %f6789, 0f00000000; @%p3564 bra $L__BB107_4205; sqrt.approx.ftz.f32 %f25787, %f6789; sub.ftz.f32 %f25788, %f25787, %f6788; add.ftz.f32 %f25789, %f6788, %f25787; div.approx.ftz.f32 %f25790, %f25788, %f25789; fma.rn.ftz.f32 %f25791, %f6788, %f25789, 0fBF800000; fma.rn.ftz.f32 %f25792, %f6788, %f25788, 0f3F800000; div.approx.ftz.f32 %f25793, %f25791, %f25792; mul.ftz.f32 %f25794, %f25790, 0f3F000000; mul.ftz.f32 %f25795, %f25790, %f25794; fma.rn.ftz.f32 %f25796, %f25793, %f25793, 0f3F800000; mul.ftz.f32 %f30708, %f25795, %f25796; $L__BB107_4205: mul.ftz.f32 %f25797, %f6659, 0f3EA2F983; abs.ftz.f32 %f6792, %f25797; add.ftz.f32 %f25798, %f30707, 0f3F800000; mul.ftz.f32 %f6793, %f6783, %f25798; @%p3558 bra $L__BB107_4207; bra.uni $L__BB107_4206; $L__BB107_4207: mul.ftz.f32 %f25858, %f6786, 0f3E800000; max.ftz.f32 %f25860, %f6793, %f25783; div.approx.ftz.f32 %f25861, %f25858, %f25860; mul.ftz.f32 %f25862, %f6775, 0fBFFE59F3; mul.ftz.f32 %f25863, %f6775, %f25862; mul.ftz.f32 %f25864, %f25863, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25865, %f25864; fma.rn.ftz.f32 %f25866, %f25865, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f25867, %f6775, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f25868, %f6775, %f25867, 0f417C8B44; fma.rn.ftz.f32 %f25869, %f6775, %f25868, 0fC1665E35; fma.rn.ftz.f32 %f25870, %f6775, %f25869, 0f40CE7881; fma.rn.ftz.f32 %f25871, %f6775, %f25870, 0fBF993E81; fma.rn.ftz.f32 %f25872, %f6775, %f25871, 0f3CE3C31E; fma.rn.ftz.f32 %f25873, %f6775, %f25872, 0f3BBA5D49; sub.ftz.f32 %f25874, %f25866, %f25873; cvt.ftz.sat.f32.f32 %f25875, %f25874; mov.f32 %f25876, 0f3F800000; sub.ftz.f32 %f25877, %f25876, %f25875; mul.ftz.f32 %f25878, %f6792, %f25877; fma.rn.ftz.f32 %f25879, %f25861, %f25875, %f25878; mul.ftz.f32 %f30902, %f30708, %f25879; bra.uni $L__BB107_4208; $L__BB107_4269: setp.neu.ftz.f32 %p3644, %f26723, %f26724; @%p3644 bra $L__BB107_4271; bra.uni $L__BB107_4270; $L__BB107_4271: ld.local.f32 %f26743, [%rd1027+52]; mul.ftz.f32 %f26744, %f7002, %f26743; ld.local.f32 %f26745, [%rd1027+48]; mul.ftz.f32 %f26746, %f7003, %f26745; sub.ftz.f32 %f26747, %f26744, %f26746; ld.local.f32 %f26748, [%rd1027+44]; mul.ftz.f32 %f26749, %f7003, %f26748; mul.ftz.f32 %f26750, %f7000, %f26743; sub.ftz.f32 %f26751, %f26749, %f26750; mul.ftz.f32 %f26752, %f7000, %f26745; mul.ftz.f32 %f26753, %f7002, %f26748; sub.ftz.f32 %f26754, %f26752, %f26753; mul.ftz.f32 %f26755, %f26751, %f26751; fma.rn.ftz.f32 %f26756, %f26747, %f26747, %f26755; fma.rn.ftz.f32 %f26757, %f26754, %f26754, %f26756; sqrt.approx.ftz.f32 %f26758, %f26757; div.approx.ftz.f32 %f30748, %f26747, %f26758; div.approx.ftz.f32 %f30749, %f26751, %f26758; div.approx.ftz.f32 %f30750, %f26754, %f26758; mul.ftz.f32 %f26759, %f7003, %f30749; mul.ftz.f32 %f26760, %f7002, %f30750; sub.ftz.f32 %f30751, %f26759, %f26760; mul.ftz.f32 %f26761, %f7000, %f30750; mul.ftz.f32 %f26762, %f7003, %f30748; sub.ftz.f32 %f30752, %f26761, %f26762; mul.ftz.f32 %f26763, %f7002, %f30748; mul.ftz.f32 %f26764, %f7000, %f30749; sub.ftz.f32 %f30753, %f26763, %f26764; bra.uni $L__BB107_4272; $L__BB107_4210: neg.ftz.f32 %f30713, %f30713; neg.ftz.f32 %f30710, %f30710; neg.ftz.f32 %f30711, %f30711; neg.ftz.f32 %f30712, %f30712; mov.f32 %f30714, %f6803; $L__BB107_4212: mul.ftz.f32 %f25890, %f30713, %f30713; mov.f32 %f30718, 0f3F800000; sub.ftz.f32 %f25891, %f30718, %f25890; mul.ftz.f32 %f25892, %f30714, %f30714; mul.ftz.f32 %f25893, %f25892, %f25891; sub.ftz.f32 %f6814, %f30718, %f25893; setp.lt.ftz.f32 %p3569, %f6814, 0f00000000; mov.f32 %f30715, 0f00000000; mov.f32 %f30716, %f30715; mov.f32 %f30717, %f30715; @%p3569 bra $L__BB107_4214; sqrt.approx.ftz.f32 %f25894, %f6814; mov.f32 %f25895, 0f33D6BF95; max.ftz.f32 %f25896, %f25894, %f25895; mul.ftz.f32 %f25897, %f30713, %f30714; sub.ftz.f32 %f25898, %f25897, %f25896; mul.ftz.f32 %f25899, %f30710, %f25898; mul.ftz.f32 %f25900, %f30711, %f25898; mul.ftz.f32 %f25901, %f30712, %f25898; mul.ftz.f32 %f25902, %f6624, %f30714; sub.ftz.f32 %f30715, %f25899, %f25902; mul.ftz.f32 %f25903, %f25507, %f30714; sub.ftz.f32 %f30716, %f25900, %f25903; mul.ftz.f32 %f25904, %f25508, %f30714; sub.ftz.f32 %f30717, %f25901, %f25904; mul.ftz.f32 %f25905, %f30711, %f30716; fma.rn.ftz.f32 %f25906, %f30710, %f30715, %f25905; fma.rn.ftz.f32 %f25907, %f30712, %f30717, %f25906; mul.ftz.f32 %f25908, %f6803, %f25907; add.ftz.f32 %f25909, %f30713, %f25908; sub.ftz.f32 %f25910, %f30713, %f25908; div.approx.ftz.f32 %f25911, %f25909, %f25910; mul.ftz.f32 %f25912, %f6803, %f30713; add.ftz.f32 %f25913, %f25912, %f25907; sub.ftz.f32 %f25914, %f25912, %f25907; div.approx.ftz.f32 %f25915, %f25913, %f25914; mul.ftz.f32 %f25916, %f25915, %f25915; fma.rn.ftz.f32 %f25917, %f25911, %f25911, %f25916; mul.ftz.f32 %f30718, %f25917, 0f3F000000; $L__BB107_4214: add.ftz.f32 %f25922, %f30713, %f30713; setp.gt.ftz.f32 %p3570, %f30718, %f6226; mul.ftz.f32 %f25923, %f30710, %f25922; sub.ftz.f32 %f25924, %f25923, %f6624; selp.f32 %f30899, %f25924, %f30715, %p3570; mul.ftz.f32 %f25925, %f30711, %f25922; sub.ftz.f32 %f25926, %f25925, %f25507; selp.f32 %f30900, %f25926, %f30716, %p3570; mul.ftz.f32 %f25927, %f30712, %f25922; sub.ftz.f32 %f25928, %f25927, %f25508; selp.f32 %f30901, %f25928, %f30717, %p3570; selp.b32 %r5877, 18, 17, %p3570; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4216: neg.ftz.f32 %f30722, %f30722; neg.ftz.f32 %f30719, %f30719; neg.ftz.f32 %f30720, %f30720; neg.ftz.f32 %f30721, %f30721; mov.u16 %rs744, 1; mov.f32 %f30723, %f6828; $L__BB107_4218: mul.ftz.f32 %f25937, %f30722, %f30722; mov.f32 %f30727, 0f3F800000; sub.ftz.f32 %f25938, %f30727, %f25937; mul.ftz.f32 %f25939, %f30723, %f30723; mul.ftz.f32 %f25940, %f25939, %f25938; sub.ftz.f32 %f6844, %f30727, %f25940; setp.lt.ftz.f32 %p3572, %f6844, 0f00000000; mov.f32 %f30899, 0f00000000; mov.f32 %f30900, %f30899; mov.f32 %f30901, %f30899; @%p3572 bra $L__BB107_4220; sqrt.approx.ftz.f32 %f25941, %f6844; mov.f32 %f25942, 0f33D6BF95; max.ftz.f32 %f25943, %f25941, %f25942; mul.ftz.f32 %f25944, %f30722, %f30723; sub.ftz.f32 %f25945, %f25944, %f25943; mul.ftz.f32 %f25946, %f30719, %f25945; mul.ftz.f32 %f25947, %f30720, %f25945; mul.ftz.f32 %f25948, %f30721, %f25945; mul.ftz.f32 %f25949, %f6830, %f30723; sub.ftz.f32 %f30899, %f25946, %f25949; mul.ftz.f32 %f25950, %f25929, %f30723; sub.ftz.f32 %f30900, %f25947, %f25950; mul.ftz.f32 %f25951, %f25930, %f30723; sub.ftz.f32 %f30901, %f25948, %f25951; mul.ftz.f32 %f25952, %f30720, %f30900; fma.rn.ftz.f32 %f25953, %f30719, %f30899, %f25952; fma.rn.ftz.f32 %f25954, %f30721, %f30901, %f25953; mul.ftz.f32 %f25955, %f6828, %f25954; add.ftz.f32 %f25956, %f30722, %f25955; sub.ftz.f32 %f25957, %f30722, %f25955; div.approx.ftz.f32 %f25958, %f25956, %f25957; mul.ftz.f32 %f25959, %f6828, %f30722; add.ftz.f32 %f25960, %f25959, %f25954; sub.ftz.f32 %f25961, %f25959, %f25954; div.approx.ftz.f32 %f25962, %f25960, %f25961; mul.ftz.f32 %f25963, %f25962, %f25962; fma.rn.ftz.f32 %f25964, %f25958, %f25958, %f25963; mul.ftz.f32 %f30727, %f25964, 0f3F000000; $L__BB107_4220: setp.neu.ftz.f32 %p3573, %f30727, 0f3F800000; setp.eq.s16 %p3574, %rs744, 0; and.pred %p3575, %p3574, %p3573; selp.f32 %f30902, 0f49742400, 0f00000000, %p3575; mov.u32 %r5877, 17; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4019: setp.eq.ftz.f32 %p3392, %f6307, 0f7F800000; setp.eq.ftz.f32 %p3393, %f6308, 0f7F800000; and.pred %p3394, %p3392, %p3393; @%p3394 bra $L__BB107_4021; bra.uni $L__BB107_4020; $L__BB107_4021: setp.lt.s32 %p3398, %r1527, 0; selp.b32 %r5205, 1075235812, 1061752795, %p3398; or.b32 %r5206, %r1528, %r5205; mov.b32 %f30631, %r5206; bra.uni $L__BB107_4023; $L__BB107_4408: neg.ftz.f32 %f30886, %f7371; neg.ftz.f32 %f30883, %f7368; neg.ftz.f32 %f30884, %f7369; neg.ftz.f32 %f30885, %f7370; mov.u16 %rs746, 1; mov.f32 %f30887, %f7436; $L__BB107_4410: mul.ftz.f32 %f27942, %f30886, %f30886; mov.f32 %f30891, 0f3F800000; sub.ftz.f32 %f27943, %f30891, %f27942; mul.ftz.f32 %f27944, %f30887, %f30887; mul.ftz.f32 %f27945, %f27944, %f27943; sub.ftz.f32 %f7447, %f30891, %f27945; setp.lt.ftz.f32 %p3752, %f7447, 0f00000000; mov.f32 %f30902, 0f00000000; mov.f32 %f30899, %f30902; mov.f32 %f30900, %f30902; mov.f32 %f30901, %f30902; @%p3752 bra $L__BB107_4412; sqrt.approx.ftz.f32 %f27946, %f7447; mov.f32 %f27947, 0f33D6BF95; max.ftz.f32 %f27948, %f27946, %f27947; mul.ftz.f32 %f27949, %f30886, %f30887; sub.ftz.f32 %f27950, %f27949, %f27948; mul.ftz.f32 %f27951, %f30883, %f27950; mul.ftz.f32 %f27952, %f30884, %f27950; mul.ftz.f32 %f27953, %f30885, %f27950; mul.ftz.f32 %f27954, %f7326, %f30887; sub.ftz.f32 %f30899, %f27951, %f27954; mul.ftz.f32 %f27955, %f27618, %f30887; sub.ftz.f32 %f30900, %f27952, %f27955; mul.ftz.f32 %f27956, %f27619, %f30887; sub.ftz.f32 %f30901, %f27953, %f27956; mul.ftz.f32 %f27957, %f30884, %f30900; fma.rn.ftz.f32 %f27958, %f30883, %f30899, %f27957; fma.rn.ftz.f32 %f27959, %f30885, %f30901, %f27958; mul.ftz.f32 %f27960, %f7436, %f27959; add.ftz.f32 %f27961, %f30886, %f27960; sub.ftz.f32 %f27962, %f30886, %f27960; div.approx.ftz.f32 %f27963, %f27961, %f27962; mul.ftz.f32 %f27964, %f7436, %f30886; add.ftz.f32 %f27965, %f27964, %f27959; sub.ftz.f32 %f27966, %f27964, %f27959; div.approx.ftz.f32 %f27967, %f27965, %f27966; mul.ftz.f32 %f27968, %f27967, %f27967; fma.rn.ftz.f32 %f27969, %f27963, %f27963, %f27968; mul.ftz.f32 %f30891, %f27969, 0f3F000000; $L__BB107_4412: setp.eq.ftz.f32 %p3753, %f30891, 0f3F800000; setp.ne.s16 %p3754, %rs746, 0; mov.u32 %r5364, 9; or.pred %p3755, %p3754, %p3753; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; mov.u32 %r5877, %r5364; @%p3755 bra $L__BB107_4433; mul.ftz.f32 %f7456, %f27620, %f27621; setp.le.ftz.f32 %p3756, %f7456, 0f33D6BF95; mov.u32 %r5877, 17; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3756 bra $L__BB107_4433; add.ftz.f32 %f27983, %f7436, 0fBF800000; abs.ftz.f32 %f27984, %f27983; setp.lt.ftz.f32 %p3757, %f27984, 0f38D1B717; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3757 bra $L__BB107_4433; mul.ftz.f32 %f27985, %f7367, %f7367; mul.ftz.f32 %f27986, %f27985, %f27985; rcp.approx.ftz.f32 %f27987, %f27985; mov.f32 %f27988, 0f3F800000; add.ftz.f32 %f27989, %f27987, 0fBF800000; mul.ftz.f32 %f27990, %f27986, 0f40490FDB; add.ftz.f32 %f27991, %f7456, %f27989; mul.ftz.f32 %f27992, %f27990, %f27991; mul.ftz.f32 %f27993, %f27991, %f27992; div.approx.ftz.f32 %f27994, %f7456, %f27993; mul.ftz.f32 %f27995, %f7327, %f30900; fma.rn.ftz.f32 %f27996, %f7325, %f30899, %f27995; fma.rn.ftz.f32 %f27997, %f7328, %f30901, %f27996; mul.ftz.f32 %f27998, %f27997, %f27997; sub.ftz.f32 %f27999, %f27988, %f27998; mul.ftz.f32 %f28000, %f7456, %f27999; div.approx.ftz.f32 %f28001, %f28000, %f27998; add.ftz.f32 %f28002, %f28001, 0f3F800000; mov.f32 %f28003, 0f00000000; max.ftz.f32 %f28004, %f28002, %f28003; sqrt.approx.ftz.f32 %f28005, %f28004; add.ftz.f32 %f28006, %f28005, 0f3F800000; mov.f32 %f28007, 0f40000000; div.approx.ftz.f32 %f28008, %f28007, %f28006; mul.ftz.f32 %f28009, %f7369, %f30900; fma.rn.ftz.f32 %f28010, %f7368, %f30899, %f28009; fma.rn.ftz.f32 %f28011, %f7370, %f30901, %f28010; fma.rn.ftz.f32 %f28012, %f7436, %f28011, %f7371; mul.ftz.f32 %f28013, %f28012, %f28012; mul.ftz.f32 %f28014, %f7364, %f27994; mul.ftz.f32 %f28015, %f7436, %f7436; mul.ftz.f32 %f28016, %f28015, %f28014; mul.ftz.f32 %f28017, %f7329, %f28013; div.approx.ftz.f32 %f28018, %f28016, %f28017; mul.ftz.f32 %f28019, %f7371, %f28011; abs.ftz.f32 %f28020, %f28019; mul.ftz.f32 %f28021, %f28008, %f28020; mul.ftz.f32 %f30903, %f28018, %f28021; abs.ftz.f32 %f28022, %f28011; mul.ftz.f32 %f28023, %f7371, %f28022; mul.ftz.f32 %f30902, %f28018, %f28023; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; mov.u32 %r5877, %r5364; bra.uni $L__BB107_4433; $L__BB107_4206: mul.ftz.f32 %f25799, %f6660, %f6779; fma.rn.ftz.f32 %f25800, %f6661, %f6778, %f25799; fma.rn.ftz.f32 %f25801, %f6659, %f6780, %f25800; mul.ftz.f32 %f25802, %f6787, %f25801; mul.ftz.f32 %f25803, %f25802, %f6786; mul.ftz.f32 %f25804, %f6639, %f25803; mul.ftz.f32 %f25805, %f6639, %f25804; mul.ftz.f32 %f25806, %f6777, %f6793; mul.ftz.f32 %f25807, %f6777, %f25806; max.ftz.f32 %f25809, %f25807, %f25783; div.approx.ftz.f32 %f25810, %f25805, %f25809; abs.ftz.f32 %f25811, %f25810; setp.lt.ftz.f32 %p3566, %f6639, 0f3F800000; mov.f32 %f25812, 0f3F800000; selp.f32 %f25813, %f6653, %f6639, %p3566; max.ftz.f32 %f25814, %f25813, %f25812; mov.f32 %f25815, 0f40400000; min.ftz.f32 %f25816, %f25814, %f25815; add.ftz.f32 %f25817, %f25816, 0fBF2814F9; mul.ftz.f32 %f25818, %f25817, 0fBF7A7547; mul.ftz.f32 %f25819, %f25817, %f25818; mul.ftz.f32 %f25820, %f25819, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25821, %f25820; mul.ftz.f32 %f25822, %f25816, 0fBD0A43BB; fma.rn.ftz.f32 %f25823, %f25821, 0f3D435661, %f25822; add.ftz.f32 %f25824, %f25823, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f25825, %f6775; fma.rn.ftz.f32 %f25826, %f25825, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f25827, %f25825, %f25826, 0f3D4D5C09; fma.rn.ftz.f32 %f25828, %f25825, %f25827, 0fBB86D80B; fma.rn.ftz.f32 %f25829, %f25825, %f25828, 0f3F80029F; fma.rn.ftz.f32 %f25830, %f25816, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f25831, %f25816, %f25830, 0fC2FFCFDF; fma.rn.ftz.f32 %f25832, %f25816, %f25831, 0f43927AA0; fma.rn.ftz.f32 %f25833, %f25816, %f25832, 0fC38FF917; mul.ftz.f32 %f25834, %f25816, %f25816; mov.f32 %f25835, 0f4347CD91; div.approx.ftz.f32 %f25836, %f25835, %f25834; add.ftz.f32 %f25837, %f25836, %f25833; mul.ftz.f32 %f25838, %f25816, %f25834; mov.f32 %f25839, 0f42CB5604; div.approx.ftz.f32 %f25840, %f25839, %f25838; sub.ftz.f32 %f25841, %f25837, %f25840; fma.rn.ftz.f32 %f25842, %f25825, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f25843, %f25825, %f25842, 0f41B5F319; fma.rn.ftz.f32 %f25844, %f25825, %f25843, 0fC05A14A5; fma.rn.ftz.f32 %f25845, %f25825, %f25844, 0f3DC9FFD6; fma.rn.ftz.f32 %f25846, %f25825, %f25845, 0f3BA1B61E; mul.ftz.f32 %f25847, %f25846, %f25841; fma.rn.ftz.f32 %f25848, %f25847, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f25849, %f25812, %f25824; sub.ftz.f32 %f25850, %f25812, %f25829; mul.ftz.f32 %f25851, %f25849, %f25850; fma.rn.ftz.f32 %f25852, %f25851, 0fC11520C5, %f25848; cvt.ftz.sat.f32.f32 %f25853, %f25852; sub.ftz.f32 %f25854, %f25812, %f25853; mul.ftz.f32 %f25855, %f6792, %f25854; fma.rn.ftz.f32 %f25856, %f25811, %f25853, %f25855; sub.ftz.f32 %f25857, %f25812, %f30708; mul.ftz.f32 %f30902, %f25857, %f25856; $L__BB107_4208: mul.ftz.f32 %f30903, %f30693, %f30902; mul.ftz.f32 %f30904, %f30694, %f30902; mul.ftz.f32 %f30905, %f30695, %f30902; mul.ftz.f32 %f25880, %f6631, %f6660; fma.rn.ftz.f32 %f25881, %f6628, %f6661, %f25880; mul.ftz.f32 %f25882, %f6632, %f6660; fma.rn.ftz.f32 %f25883, %f6629, %f6661, %f25882; mul.ftz.f32 %f25884, %f6633, %f6660; fma.rn.ftz.f32 %f25885, %f6630, %f6661, %f25884; fma.rn.ftz.f32 %f30899, %f30710, %f6659, %f25881; fma.rn.ftz.f32 %f30900, %f30711, %f6659, %f25883; fma.rn.ftz.f32 %f30901, %f30712, %f6659, %f25885; selp.b32 %r5877, 10, 9, %p3558; bra.uni $L__BB107_4433; $L__BB107_4323: ld.local.f32 %f27078, [%rd1027+52]; mul.ftz.f32 %f27079, %f7165, %f27078; ld.local.f32 %f27080, [%rd1027+48]; mul.ftz.f32 %f27081, %f7166, %f27080; sub.ftz.f32 %f27082, %f27079, %f27081; ld.local.f32 %f27083, [%rd1027+44]; mul.ftz.f32 %f27084, %f7166, %f27083; mul.ftz.f32 %f27085, %f7163, %f27078; sub.ftz.f32 %f27086, %f27084, %f27085; mul.ftz.f32 %f27087, %f7163, %f27080; mul.ftz.f32 %f27088, %f7165, %f27083; sub.ftz.f32 %f27089, %f27087, %f27088; mul.ftz.f32 %f27090, %f27086, %f27086; fma.rn.ftz.f32 %f27091, %f27082, %f27082, %f27090; fma.rn.ftz.f32 %f27092, %f27089, %f27089, %f27091; sqrt.approx.ftz.f32 %f27093, %f27092; div.approx.ftz.f32 %f30785, %f27082, %f27093; div.approx.ftz.f32 %f30786, %f27086, %f27093; div.approx.ftz.f32 %f30787, %f27089, %f27093; mul.ftz.f32 %f27094, %f7166, %f30786; mul.ftz.f32 %f27095, %f7165, %f30787; sub.ftz.f32 %f30788, %f27094, %f27095; mul.ftz.f32 %f27096, %f7163, %f30787; mul.ftz.f32 %f27097, %f7166, %f30785; sub.ftz.f32 %f30789, %f27096, %f27097; mul.ftz.f32 %f27098, %f7165, %f30785; mul.ftz.f32 %f27099, %f7163, %f30786; sub.ftz.f32 %f30790, %f27098, %f27099; $L__BB107_4325: mul.ftz.f32 %f27121, %f27067, %f30789; fma.rn.ftz.f32 %f27122, %f7164, %f30788, %f27121; fma.rn.ftz.f32 %f27123, %f27068, %f30790, %f27122; mul.ftz.f32 %f27124, %f27067, %f30786; fma.rn.ftz.f32 %f27125, %f7164, %f30785, %f27124; fma.rn.ftz.f32 %f27126, %f27068, %f30787, %f27125; mul.ftz.f32 %f27127, %f27069, %f27123; mul.ftz.f32 %f27128, %f27069, %f27126; mul.ftz.f32 %f27129, %f27127, %f27127; fma.rn.ftz.f32 %f27130, %f27128, %f27128, %f27129; fma.rn.ftz.f32 %f27131, %f7167, %f7167, %f27130; sqrt.approx.ftz.f32 %f27132, %f27131; div.approx.ftz.f32 %f7186, %f27127, %f27132; div.approx.ftz.f32 %f7187, %f27128, %f27132; div.approx.ftz.f32 %f7188, %f7167, %f27132; setp.geu.ftz.f32 %p3697, %f7188, 0f3F7FFF58; mov.f32 %f30792, 0f00000000; mov.f32 %f30791, 0f3F800000; mov.f32 %f30793, %f30791; mov.f32 %f30794, %f30792; @%p3697 bra $L__BB107_4327; mul.ftz.f32 %f27133, %f7188, %f7188; mov.f32 %f27134, 0f3F800000; sub.ftz.f32 %f27135, %f27134, %f27133; mov.f32 %f27136, 0f00000000; max.ftz.f32 %f27137, %f27135, %f27136; sqrt.approx.ftz.f32 %f30792, %f27137; rsqrt.approx.ftz.f32 %f27138, %f27137; mul.ftz.f32 %f30793, %f7186, %f27138; mul.ftz.f32 %f30794, %f7187, %f27138; mov.f32 %f30791, %f7188; $L__BB107_4327: setp.ltu.ftz.f32 %p3698, %f30791, 0f3F7FFF58; @%p3698 bra $L__BB107_4329; bra.uni $L__BB107_4328; $L__BB107_4329: div.approx.ftz.f32 %f7198, %f30792, %f30791; rcp.approx.ftz.f32 %f7199, %f7198; abs.ftz.f32 %f7200, %f7199; setp.ge.ftz.f32 %p3699, %f7200, 0f4144CCCD; mov.u32 %r5875, 1065353216; @%p3699 bra $L__BB107_4331; add.ftz.f32 %f27146, %f7200, 0fBF800000; mov.f32 %f27147, 0f3F800000; add.ftz.f32 %f27148, %f27146, 0f3F800000; fma.rn.ftz.f32 %f27149, %f27148, 0f38349F67, 0f39910039; fma.rn.ftz.f32 %f27150, %f27148, %f27149, 0f391F6607; fma.rn.ftz.f32 %f27151, %f27148, %f27150, 0f3C17E369; fma.rn.ftz.f32 %f27152, %f27148, %f27151, 0f3D2D2FE7; fma.rn.ftz.f32 %f27153, %f27148, %f27152, 0f3D906E67; fma.rn.ftz.f32 %f27154, %f27148, %f27153, 0f3F800000; mul.ftz.f32 %f27155, %f27154, %f27154; mul.ftz.f32 %f27156, %f27155, %f27155; mul.ftz.f32 %f27157, %f27156, %f27156; mul.ftz.f32 %f27158, %f27157, %f27157; rcp.approx.ftz.f32 %f27159, %f27158; sub.ftz.f32 %f27160, %f27147, %f27159; mov.b32 %r5325, %f27160; and.b32 %r5875, %r5325, 2147483647; $L__BB107_4331: mov.b32 %r5326, %f7199; and.b32 %r5327, %r5326, -2147483648; or.b32 %r5328, %r5875, %r5327; mov.b32 %f27161, %r5328; mul.ftz.f32 %f27162, %f7199, %f7199; mul.ftz.f32 %f27163, %f27162, 0fBFB8AA3B; ex2.approx.ftz.f32 %f27164, %f27163; add.ftz.f32 %f27165, %f27161, 0fBF800000; mov.f32 %f27166, 0f3F800000; mul.ftz.f32 %f27167, %f27165, 0f3F000000; mul.ftz.f32 %f27168, %f7198, %f27164; fma.rn.ftz.f32 %f7201, %f27168, 0f3E906EBB, %f27167; add.ftz.f32 %f27169, %f27161, 0f3F800000; mul.ftz.f32 %f27170, %f27161, %f27161; sub.ftz.f32 %f27171, %f27166, %f27170; mul.ftz.f32 %f7202, %f7198, 0f3F106EBB; fma.rn.ftz.f32 %f27172, %f7202, %f27171, %f27169; mul.ftz.f32 %f7203, %f6226, %f27172; fma.rn.ftz.f32 %f27173, %f7202, %f27164, %f27169; mul.ftz.f32 %f7204, %f6226, %f27173; setp.gt.ftz.f32 %p3700, %f7202, 0f00000000; @%p3700 bra $L__BB107_4333; bra.uni $L__BB107_4332; $L__BB107_4333: sub.ftz.f32 %f27174, %f7202, %f7203; add.ftz.f32 %f27175, %f27174, 0f3F800000; fma.rn.ftz.f32 %f27176, %f7202, %f27175, 0f3E800000; sqrt.approx.ftz.f32 %f27177, %f27176; mov.f32 %f27178, 0f3F000000; sub.ftz.f32 %f27179, %f27178, %f27177; div.approx.ftz.f32 %f30795, %f27179, %f7202; bra.uni $L__BB107_4334; $L__BB107_4328: lg2.approx.ftz.f32 %f27140, %f6226; mul.ftz.f32 %f27141, %f27140, 0fBF317218; sqrt.approx.ftz.f32 %f27142, %f27141; mul.ftz.f32 %f27143, %f6202, 0f40C90FDB; cos.approx.ftz.f32 %f27144, %f27143; mul.ftz.f32 %f30839, %f27142, %f27144; sin.approx.ftz.f32 %f27145, %f27143; mul.ftz.f32 %f30838, %f27142, %f27145; mov.f32 %f30837, 0f3F800000; bra.uni $L__BB107_4349; $L__BB107_4020: setp.lt.s32 %p3395, %r1527, 0; min.ftz.f32 %f24813, %f6308, %f6307; max.ftz.f32 %f24814, %f6308, %f6307; div.full.ftz.f32 %f24815, %f24813, %f24814; mul.rn.ftz.f32 %f24816, %f24815, %f24815; mov.f32 %f24817, 0fC0B59883; mov.f32 %f24818, 0fBF52C7EA; fma.rn.ftz.f32 %f24819, %f24816, %f24818, %f24817; mov.f32 %f24820, 0fC0D21907; fma.rn.ftz.f32 %f24821, %f24819, %f24816, %f24820; mul.ftz.f32 %f24822, %f24816, %f24821; mul.ftz.f32 %f24823, %f24815, %f24822; add.ftz.f32 %f24824, %f24816, 0f41355DC0; mov.f32 %f24825, 0f41E6BD60; fma.rn.ftz.f32 %f24826, %f24824, %f24816, %f24825; mov.f32 %f24827, 0f419D92C8; fma.rn.ftz.f32 %f24828, %f24826, %f24816, %f24827; rcp.approx.ftz.f32 %f24829, %f24828; fma.rn.ftz.f32 %f24830, %f24823, %f24829, %f24815; mov.f32 %f24831, 0f3FC90FDB; sub.ftz.f32 %f24832, %f24831, %f24830; setp.gt.ftz.f32 %p3396, %f6308, %f6307; selp.f32 %f24833, %f24832, %f24830, %p3396; mov.f32 %f24834, 0f40490FDB; sub.ftz.f32 %f24835, %f24834, %f24833; selp.f32 %f24836, %f24835, %f24833, %p3395; mov.b32 %r5203, %f24836; or.b32 %r5204, %r1528, %r5203; mov.b32 %f24837, %r5204; add.ftz.f32 %f24838, %f6307, %f6308; setp.le.ftz.f32 %p3397, %f24838, 0f7F800000; selp.f32 %f30631, %f24837, %f24838, %p3397; $L__BB107_4023: ld.local.f32 %f24840, [%rd1027+44]; div.approx.ftz.f32 %f24841, %f6301, %f24840; mul.ftz.f32 %f24842, %f24841, %f24841; sub.ftz.f32 %f24843, %f24809, %f24842; max.ftz.f32 %f24845, %f24843, %f30636; sqrt.approx.ftz.f32 %f24846, %f24845; ld.f32 %f24847, [%rd1028+12]; mul.ftz.f32 %f24848, %f24847, %f24847; sub.ftz.f32 %f24849, %f24809, %f24848; max.ftz.f32 %f24850, %f24849, %f30636; sqrt.approx.ftz.f32 %f24851, %f24850; mov.f32 %f24852, 0fBF800000; max.ftz.f32 %f24853, %f24847, %f24852; min.ftz.f32 %f24854, %f24853, %f24809; abs.ftz.f32 %f24855, %f24854; sub.ftz.f32 %f24856, %f24809, %f24855; mul.ftz.f32 %f24857, %f24856, 0f3F000000; sqrt.approx.ftz.f32 %f24858, %f24857; setp.gt.ftz.f32 %p3399, %f24855, 0f3F11EB85; selp.f32 %f24859, %f24858, %f24855, %p3399; mul.ftz.f32 %f24860, %f24859, %f24859; mov.f32 %f24861, 0f3C94D2E9; mov.f32 %f24862, 0f3D53F941; fma.rn.ftz.f32 %f24863, %f24862, %f24860, %f24861; mov.f32 %f24864, 0f3D3F841F; fma.rn.ftz.f32 %f24865, %f24863, %f24860, %f24864; mov.f32 %f24866, 0f3D994929; fma.rn.ftz.f32 %f24867, %f24865, %f24860, %f24866; mov.f32 %f24868, 0f3E2AAB94; fma.rn.ftz.f32 %f24869, %f24867, %f24860, %f24868; mul.ftz.f32 %f24870, %f24860, %f24869; fma.rn.ftz.f32 %f24871, %f24870, %f24859, %f24859; mov.f32 %f24872, 0f3FC90FDB; mov.f32 %f24873, 0fC0000000; fma.rn.ftz.f32 %f24874, %f24873, %f24871, %f24872; selp.f32 %f24875, %f24874, %f24871, %p3399; setp.le.ftz.f32 %p3400, %f24875, 0f7F800000; mov.b32 %r5210, %f24875; mov.b32 %r5211, %f24854; and.b32 %r5212, %r5211, -2147483648; or.b32 %r5213, %r5212, %r5210; mov.b32 %f24876, %r5213; selp.f32 %f6313, %f24876, %f24875, %p3400; mul.ftz.f32 %f24877, %f6306, %f24847; mul.ftz.f32 %f24878, %f24840, %f24840; sub.ftz.f32 %f24879, %f24878, %f6305; sqrt.approx.ftz.f32 %f24880, %f24879; div.approx.ftz.f32 %f24881, %f24877, %f24880; mul.ftz.f32 %f24882, %f24881, %f24881; sub.ftz.f32 %f24883, %f24809, %f24882; max.ftz.f32 %f24884, %f24883, %f30636; sqrt.approx.ftz.f32 %f24885, %f24884; max.ftz.f32 %f24886, %f24881, %f24852; min.ftz.f32 %f24887, %f24886, %f24809; abs.ftz.f32 %f24888, %f24887; sub.ftz.f32 %f24889, %f24809, %f24888; mul.ftz.f32 %f24890, %f24889, 0f3F000000; sqrt.approx.ftz.f32 %f24891, %f24890; setp.gt.ftz.f32 %p3401, %f24888, 0f3F11EB85; selp.f32 %f24892, %f24891, %f24888, %p3401; mul.ftz.f32 %f24893, %f24892, %f24892; fma.rn.ftz.f32 %f24894, %f24862, %f24893, %f24861; fma.rn.ftz.f32 %f24895, %f24894, %f24893, %f24864; fma.rn.ftz.f32 %f24896, %f24895, %f24893, %f24866; fma.rn.ftz.f32 %f24897, %f24896, %f24893, %f24868; mul.ftz.f32 %f24898, %f24893, %f24897; fma.rn.ftz.f32 %f24899, %f24898, %f24892, %f24892; fma.rn.ftz.f32 %f24900, %f24873, %f24899, %f24872; selp.f32 %f24901, %f24900, %f24899, %p3401; setp.le.ftz.f32 %p3402, %f24901, 0f7F800000; mov.b32 %r5214, %f24901; mov.b32 %r5215, %f24887; and.b32 %r5216, %r5215, -2147483648; or.b32 %r5217, %r5216, %r5214; mov.b32 %f24902, %r5217; selp.f32 %f6314, %f24902, %f24901, %p3402; add.ftz.f32 %f24903, %f24885, %f24885; div.approx.ftz.f32 %f24904, %f24903, %f24846; ld.local.f32 %f24905, [%rd1027+20]; mul.ftz.f32 %f6315, %f24905, %f24904; ld.local.f32 %f24906, [%rd1027+24]; mul.ftz.f32 %f6316, %f24906, %f24904; ld.local.f32 %f24907, [%rd1027+28]; mul.ftz.f32 %f6317, %f24907, %f24904; mul.ftz.f32 %f24908, %f6306, %f24851; abs.ftz.f32 %f6318, %f24908; add.ftz.f32 %f24909, %f24878, 0fBF800000; fma.rn.ftz.f32 %f6319, %f6318, %f6318, %f24909; setp.leu.ftz.f32 %p3403, %f6319, 0f00000000; mov.f32 %f30632, %f24809; @%p3403 bra $L__BB107_4025; sqrt.approx.ftz.f32 %f24910, %f6319; sub.ftz.f32 %f24911, %f24910, %f6318; add.ftz.f32 %f24912, %f6318, %f24910; div.approx.ftz.f32 %f24913, %f24911, %f24912; fma.rn.ftz.f32 %f24914, %f6318, %f24912, 0fBF800000; fma.rn.ftz.f32 %f24915, %f6318, %f24911, 0f3F800000; div.approx.ftz.f32 %f24916, %f24914, %f24915; mul.ftz.f32 %f24917, %f24913, 0f3F000000; mul.ftz.f32 %f24918, %f24913, %f24917; fma.rn.ftz.f32 %f24919, %f24916, %f24916, 0f3F800000; mul.ftz.f32 %f30632, %f24918, %f24919; $L__BB107_4025: mul.ftz.f32 %f24921, %f6315, 0fBFB8AA3B; ex2.approx.ftz.f32 %f24922, %f24921; mul.ftz.f32 %f24923, %f6316, 0fBFB8AA3B; ex2.approx.ftz.f32 %f24924, %f24923; mul.ftz.f32 %f24925, %f6317, 0fBFB8AA3B; ex2.approx.ftz.f32 %f24926, %f24925; sub.ftz.f32 %f24928, %f24809, %f30632; mul.ftz.f32 %f24929, %f24928, %f24928; mul.ftz.f32 %f6322, %f24922, %f24929; mul.ftz.f32 %f6323, %f24924, %f24929; mul.ftz.f32 %f6324, %f24926, %f24929; ld.const.v4.f32 {%f24930, %f24931, %f24932, %f24933}, [kernel_params+1184]; mul.ftz.f32 %f24934, %f6323, %f24931; fma.rn.ftz.f32 %f24935, %f6322, %f24930, %f24934; fma.rn.ftz.f32 %f6328, %f6324, %f24932, %f24935; mul.ftz.f32 %f6329, %f24922, %f30632; mul.ftz.f32 %f6330, %f6329, %f6322; mul.ftz.f32 %f6331, %f24924, %f30632; mul.ftz.f32 %f6332, %f6331, %f6323; mul.ftz.f32 %f6333, %f24926, %f30632; mul.ftz.f32 %f6334, %f6333, %f6324; mul.ftz.f32 %f24936, %f6332, %f24931; fma.rn.ftz.f32 %f24937, %f6330, %f24930, %f24936; fma.rn.ftz.f32 %f6335, %f6334, %f24932, %f24937; sub.ftz.f32 %f6336, %f24809, %f6329; sub.ftz.f32 %f6337, %f24809, %f6331; sub.ftz.f32 %f6338, %f24809, %f6333; setp.eq.ftz.f32 %p3404, %f6336, 0f00000000; mov.f32 %f30633, %f30636; @%p3404 bra $L__BB107_4027; div.approx.ftz.f32 %f30633, %f6329, %f6336; $L__BB107_4027: setp.eq.ftz.f32 %p3405, %f6337, 0f00000000; mov.f32 %f30634, %f30636; @%p3405 bra $L__BB107_4029; div.approx.ftz.f32 %f30634, %f6331, %f6337; $L__BB107_4029: setp.eq.ftz.f32 %p3406, %f6338, 0f00000000; mov.f32 %f30635, %f30636; @%p3406 bra $L__BB107_4031; div.approx.ftz.f32 %f30635, %f6333, %f6338; $L__BB107_4031: mul.ftz.f32 %f6345, %f6330, %f30633; mul.ftz.f32 %f6346, %f6332, %f30634; mul.ftz.f32 %f24941, %f6346, %f24931; fma.rn.ftz.f32 %f24942, %f6345, %f24930, %f24941; mul.ftz.f32 %f6347, %f6334, %f30635; fma.rn.ftz.f32 %f6348, %f6347, %f24932, %f24942; add.ftz.f32 %f24943, %f30632, %f6328; add.ftz.f32 %f24944, %f24943, %f6335; add.ftz.f32 %f6349, %f6348, %f24944; setp.eq.ftz.f32 %p3407, %f6349, 0f00000000; @%p3407 bra $L__BB107_4033; mov.f32 %f24945, 0f3F800000; div.approx.ftz.f32 %f30636, %f24945, %f6349; $L__BB107_4033: mul.ftz.f32 %f6352, %f30636, %f6328; mul.ftz.f32 %f6353, %f30636, %f6335; mul.ftz.f32 %f6354, %f30636, %f6348; mul.ftz.f32 %f6355, %f30636, %f30632; setp.lt.ftz.f32 %p3409, %f6226, %f6355; mov.pred %p3928, -1; mov.u32 %r5870, 0; @%p3409 bra $L__BB107_4037; sub.ftz.f32 %f6356, %f6226, %f6355; setp.lt.ftz.f32 %p3411, %f6356, %f6352; mov.u32 %r5870, 1; @%p3411 bra $L__BB107_4037; sub.ftz.f32 %f24946, %f6356, %f6352; setp.lt.ftz.f32 %p3413, %f24946, %f6353; mov.u32 %r5870, 2; @%p3413 bra $L__BB107_4037; mov.pred %p3928, 0; mov.u32 %r5870, 3; $L__BB107_4037: ld.local.f32 %f6357, [%rd1027+32]; mul.ftz.f32 %f6358, %f6357, 0f3E800000; setp.eq.s32 %p3415, %r5870, 1; selp.f32 %f24947, %f6358, %f6357, %p3415; mul.ftz.f32 %f24948, %f24947, 0f40800000; setp.gt.u32 %p3416, %r5870, 1; selp.f32 %f24949, %f24948, %f24947, %p3416; cvt.rn.f32.u32 %f24950, %r1525; mul.ftz.f32 %f24951, %f24950, 0f2F800000; mov.f32 %f24952, 0f3727C5AC; max.ftz.f32 %f24953, %f24951, %f24952; mov.f32 %f24954, 0f3F800000; sub.ftz.f32 %f24955, %f24954, %f24953; div.approx.ftz.f32 %f24957, %f24873, %f24949; mul.ftz.f32 %f24958, %f24957, 0f3FB8AA3B; ex2.approx.ftz.f32 %f24959, %f24958; fma.rn.ftz.f32 %f24960, %f24955, %f24959, %f24953; lg2.approx.ftz.f32 %f24961, %f24960; mul.ftz.f32 %f24962, %f24961, 0f3F317218; fma.rn.ftz.f32 %f24963, %f24949, %f24962, 0f3F800000; mul.ftz.f32 %f24964, %f6301, %f24963; mul.ftz.f32 %f24965, %f24963, %f24963; sub.ftz.f32 %f24966, %f24954, %f24965; mov.f32 %f30649, 0f00000000; max.ftz.f32 %f24968, %f24966, %f30649; sqrt.approx.ftz.f32 %f24969, %f24968; cvt.rn.f32.u32 %f24970, %r1526; mul.ftz.f32 %f24971, %f24970, 0f2F800000; mul.ftz.f32 %f24972, %f24971, 0f40C90FDB; cos.approx.ftz.f32 %f24973, %f24972; mul.ftz.f32 %f24974, %f24969, %f24973; mul.ftz.f32 %f24975, %f6306, %f24974; sub.ftz.f32 %f30637, %f24975, %f24964; mul.ftz.f32 %f24976, %f30637, %f30637; sub.ftz.f32 %f24977, %f24954, %f24976; max.ftz.f32 %f24978, %f24977, %f30649; sqrt.approx.ftz.f32 %f30638, %f24978; not.pred %p3417, %p3928; @%p3417 bra $L__BB107_4039; ld.local.f32 %f24979, [%rd1027+40]; neg.ftz.f32 %f24980, %f24979; sin.approx.ftz.f32 %f24981, %f24980; mul.ftz.f32 %f24982, %f24981, %f24981; sub.ftz.f32 %f24984, %f24954, %f24982; mov.f32 %f24985, 0f00000000; max.ftz.f32 %f24986, %f24984, %f24985; sqrt.approx.ftz.f32 %f24987, %f24986; add.ftz.f32 %f24988, %f24981, %f24981; mul.ftz.f32 %f24989, %f24988, %f24987; mul.ftz.f32 %f24990, %f24987, %f24987; sub.ftz.f32 %f24991, %f24990, %f24982; add.ftz.f32 %f24992, %f24989, %f24989; mul.ftz.f32 %f24993, %f24992, %f24991; mul.ftz.f32 %f24994, %f24991, %f24991; mul.ftz.f32 %f24995, %f24989, %f24989; sub.ftz.f32 %f24996, %f24994, %f24995; mul.ftz.f32 %f24997, %f30637, %f24991; mul.ftz.f32 %f24998, %f30638, %f24991; mul.ftz.f32 %f24999, %f30637, %f24989; sub.ftz.f32 %f25000, %f24998, %f24999; fma.rn.ftz.f32 %f25001, %f30638, %f24989, %f24997; abs.ftz.f32 %f25002, %f25000; st.local.v2.f32 [%rd2157], {%f25001, %f25002}; mul.ftz.f32 %f25003, %f30637, %f24987; mul.ftz.f32 %f25004, %f30638, %f24981; mul.ftz.f32 %f25005, %f30638, %f24987; fma.rn.ftz.f32 %f25006, %f30637, %f24981, %f25005; sub.ftz.f32 %f25007, %f25003, %f25004; abs.ftz.f32 %f25008, %f25006; st.local.v2.f32 [%rd277], {%f25007, %f25008}; mul.ftz.f32 %f25009, %f30637, %f24996; mul.ftz.f32 %f25010, %f30638, %f24993; mul.ftz.f32 %f25011, %f30638, %f24996; fma.rn.ftz.f32 %f25012, %f30637, %f24993, %f25011; sub.ftz.f32 %f25013, %f25009, %f25010; abs.ftz.f32 %f25014, %f25012; st.local.v2.f32 [%rd277+8], {%f25013, %f25014}; shl.b32 %r5222, %r5870, 1; mul.wide.u32 %rd4399, %r5222, 4; add.s64 %rd4400, %rd2157, %rd4399; ld.local.v2.f32 {%f30637, %f30638}, [%rd4400]; $L__BB107_4039: add.ftz.f32 %f6365, %f6313, %f6313; @%p3928 bra $L__BB107_4041; bra.uni $L__BB107_4040; $L__BB107_4041: cvt.rn.f32.s32 %f25019, %r5870; add.ftz.f32 %f25020, %f25019, %f25019; mul.ftz.f32 %f25021, %f6314, %f25020; sub.ftz.f32 %f25022, %f25021, %f6365; mov.f32 %f25023, 0f40490FDB; fma.rn.ftz.f32 %f6369, %f25019, 0f40490FDB, %f25022; ld.local.f32 %f30641, [%rd1027+36]; div.approx.ftz.f32 %f30640, %f25023, %f30641; setp.gt.ftz.f32 %p3418, %f30640, 0f42B00000; mov.f32 %f30639, 0f00000000; @%p3418 bra $L__BB107_4043; mul.ftz.f32 %f25024, %f30640, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25025, %f25024; add.ftz.f32 %f25026, %f25025, 0f3F800000; rcp.approx.ftz.f32 %f30639, %f25026; $L__BB107_4043: add.ftz.f32 %f25027, %f30639, %f30639; sub.ftz.f32 %f25029, %f24954, %f25027; fma.rn.ftz.f32 %f25030, %f6202, %f25029, %f30639; rcp.approx.ftz.f32 %f25031, %f25030; add.ftz.f32 %f25032, %f25031, 0fBF800000; lg2.approx.ftz.f32 %f25033, %f25032; mul.ftz.f32 %f25034, %f25033, 0fBF317218; mul.ftz.f32 %f25035, %f30641, %f25034; mov.f32 %f25036, 0fC0490FDB; max.ftz.f32 %f25037, %f25035, %f25036; min.ftz.f32 %f25039, %f25037, %f25023; add.ftz.f32 %f30642, %f6369, %f25039; bra.uni $L__BB107_4044; $L__BB107_4040: mul.ftz.f32 %f30642, %f6202, 0f40C90FDB; ld.local.f32 %f30641, [%rd1027+36]; mov.f32 %f25017, 0f40490FDB; div.approx.ftz.f32 %f30640, %f25017, %f30641; $L__BB107_4044: ld.local.f32 %f25040, [%rd1027+40]; sin.approx.ftz.f32 %f25041, %f25040; mul.ftz.f32 %f25042, %f25041, %f25041; sub.ftz.f32 %f25044, %f24954, %f25042; max.ftz.f32 %f25046, %f25044, %f30649; sqrt.approx.ftz.f32 %f25047, %f25046; add.ftz.f32 %f25048, %f25041, %f25041; mul.ftz.f32 %f25049, %f25048, %f25047; mul.ftz.f32 %f25050, %f25047, %f25047; sub.ftz.f32 %f25051, %f25050, %f25042; add.ftz.f32 %f25052, %f25049, %f25049; mul.ftz.f32 %f25053, %f25052, %f25051; mul.ftz.f32 %f25054, %f25051, %f25051; mul.ftz.f32 %f25055, %f25049, %f25049; sub.ftz.f32 %f25056, %f25054, %f25055; mul.ftz.f32 %f25057, %f30637, %f25051; mul.ftz.f32 %f25058, %f30638, %f25051; mul.ftz.f32 %f25059, %f30637, %f25049; sub.ftz.f32 %f25060, %f25058, %f25059; fma.rn.ftz.f32 %f25061, %f30638, %f25049, %f25057; abs.ftz.f32 %f25062, %f25060; st.local.v2.f32 [%rd2157], {%f25061, %f25062}; mul.ftz.f32 %f25063, %f30637, %f25047; mul.ftz.f32 %f25064, %f30638, %f25041; sub.ftz.f32 %f6378, %f25063, %f25064; mul.ftz.f32 %f25065, %f30638, %f25047; fma.rn.ftz.f32 %f25066, %f30637, %f25041, %f25065; abs.ftz.f32 %f6379, %f25066; st.local.v2.f32 [%rd277], {%f6378, %f6379}; mul.ftz.f32 %f25067, %f30637, %f25056; mul.ftz.f32 %f25068, %f30638, %f25053; sub.ftz.f32 %f6380, %f25067, %f25068; mul.ftz.f32 %f25069, %f30638, %f25056; fma.rn.ftz.f32 %f25070, %f30637, %f25053, %f25069; abs.ftz.f32 %f6381, %f25070; st.local.v2.f32 [%rd277+8], {%f6380, %f6381}; mul.ftz.f32 %f6382, %f30640, 0f3FB8AA3B; ld.local.f32 %f6383, [%rd1027+48]; rcp.approx.ftz.f32 %f6384, %f6383; mul.ftz.f32 %f25071, %f6306, %f25062; mul.ftz.f32 %f6385, %f25071, %f6384; mul.ftz.f32 %f25072, %f6301, %f25061; mul.ftz.f32 %f6386, %f25072, %f6384; setp.gtu.ftz.f32 %p3419, %f6383, 0f3DCCCCCD; @%p3419 bra $L__BB107_4058; bra.uni $L__BB107_4045; $L__BB107_4058: mul.ftz.f32 %f6409, %f6385, %f6385; fma.rn.ftz.f32 %f30645, %f6409, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6411, %f6409, %f6409; mov.u64 %rd4411, 64; cvt.rn.f32.u64 %f25103, %rd4411; div.approx.ftz.f32 %f25104, %f6411, %f25103; add.ftz.f32 %f6412, %f30645, %f25104; setp.eq.ftz.f32 %p3429, %f30645, %f6412; @%p3429 bra $L__BB107_4067; mul.ftz.f32 %f6413, %f6409, %f6411; mov.u64 %rd4412, 2304; cvt.rn.f32.u64 %f25105, %rd4412; div.approx.ftz.f32 %f25106, %f6413, %f25105; add.ftz.f32 %f6414, %f6412, %f25106; setp.eq.ftz.f32 %p3430, %f6412, %f6414; mov.f32 %f30645, %f6412; @%p3430 bra $L__BB107_4067; mul.ftz.f32 %f6415, %f6409, %f6413; mov.u64 %rd4413, 147456; cvt.rn.f32.u64 %f25107, %rd4413; div.approx.ftz.f32 %f25108, %f6415, %f25107; add.ftz.f32 %f6416, %f6414, %f25108; setp.eq.ftz.f32 %p3431, %f6414, %f6416; mov.f32 %f30645, %f6414; @%p3431 bra $L__BB107_4067; mul.ftz.f32 %f6417, %f6409, %f6415; mov.u64 %rd4414, 14745600; cvt.rn.f32.u64 %f25109, %rd4414; div.approx.ftz.f32 %f25110, %f6417, %f25109; add.ftz.f32 %f6418, %f6416, %f25110; setp.eq.ftz.f32 %p3432, %f6416, %f6418; mov.f32 %f30645, %f6416; @%p3432 bra $L__BB107_4067; mul.ftz.f32 %f6419, %f6409, %f6417; mov.u64 %rd4415, 2123366400; cvt.rn.f32.u64 %f25111, %rd4415; div.approx.ftz.f32 %f25112, %f6419, %f25111; add.ftz.f32 %f6420, %f6418, %f25112; setp.eq.ftz.f32 %p3433, %f6418, %f6420; mov.f32 %f30645, %f6418; @%p3433 bra $L__BB107_4067; mul.ftz.f32 %f6421, %f6409, %f6419; mov.u64 %rd4416, 416179814400; cvt.rn.f32.u64 %f25113, %rd4416; div.approx.ftz.f32 %f25114, %f6421, %f25113; add.ftz.f32 %f6422, %f6420, %f25114; setp.eq.ftz.f32 %p3434, %f6420, %f6422; mov.f32 %f30645, %f6420; @%p3434 bra $L__BB107_4067; mul.ftz.f32 %f6423, %f6409, %f6421; mov.u64 %rd4417, 106542032486400; cvt.rn.f32.u64 %f25115, %rd4417; div.approx.ftz.f32 %f25116, %f6423, %f25115; add.ftz.f32 %f6424, %f6422, %f25116; setp.eq.ftz.f32 %p3435, %f6422, %f6424; mov.f32 %f30645, %f6422; @%p3435 bra $L__BB107_4067; mul.ftz.f32 %f25117, %f6409, %f6423; mov.u64 %rd4418, 34519618525593600; cvt.rn.f32.u64 %f25118, %rd4418; div.approx.ftz.f32 %f25119, %f25117, %f25118; add.ftz.f32 %f6425, %f6424, %f25119; setp.eq.ftz.f32 %p3436, %f6424, %f6425; mov.f32 %f30645, %f6424; @%p3436 bra $L__BB107_4067; mov.f32 %f30645, %f6425; $L__BB107_4067: mul.ftz.f32 %f25120, %f6386, 0fBFB8AA3B; ex2.approx.ftz.f32 %f25121, %f25120; mul.ftz.f32 %f6427, %f30645, %f25121; abs.ftz.f32 %f6428, %f6384; setp.ltu.ftz.f32 %p3437, %f6428, 0f3F800000; @%p3437 bra $L__BB107_4069; bra.uni $L__BB107_4068; $L__BB107_4069: mul.ftz.f32 %f25143, %f6384, %f6384; mov.f32 %f25144, 0f394FFF49; mov.f32 %f25145, 0f363D0ADA; fma.rn.ftz.f32 %f25146, %f25145, %f25143, %f25144; mov.f32 %f25147, 0f3C08889A; fma.rn.ftz.f32 %f25148, %f25146, %f25143, %f25147; mov.f32 %f25149, 0f3E2AAAAB; fma.rn.ftz.f32 %f25150, %f25148, %f25143, %f25149; mul.ftz.f32 %f25151, %f25143, %f25150; fma.rn.ftz.f32 %f30646, %f25151, %f6384, %f6384; bra.uni $L__BB107_4070; $L__BB107_4045: setp.gt.ftz.f32 %p3420, %f6385, 0f41400000; @%p3420 bra $L__BB107_4056; bra.uni $L__BB107_4046; $L__BB107_4056: mul.ftz.f32 %f25091, %f6385, 0f41000000; rcp.approx.ftz.f32 %f25092, %f25091; add.ftz.f32 %f25093, %f25092, 0fBFEB3F8E; lg2.approx.ftz.f32 %f25094, %f6385; fma.rn.ftz.f32 %f25095, %f25094, 0fBF317218, %f25093; fma.rn.ftz.f32 %f30644, %f25095, 0f3F000000, %f6385; bra.uni $L__BB107_4057; $L__BB107_4332: add.ftz.f32 %f30795, %f7203, 0fBF800000; $L__BB107_4334: abs.ftz.f32 %f27180, %f30795; setp.gt.ftz.f32 %p3701, %f27180, 0f3F7FFFFF; selp.f32 %f27181, 0f3F7FFFFF, %f27180, %p3701; sub.ftz.f32 %f27183, %f27166, %f27181; add.ftz.f32 %f27184, %f27181, 0f3F800000; mul.ftz.f32 %f27185, %f27183, %f27184; mov.f32 %f27186, 0f00800000; max.ftz.f32 %f27187, %f27185, %f27186; mov.f32 %f27188, 0f7F7FFFFF; min.ftz.f32 %f27189, %f27187, %f27188; mov.b32 %r5329, %f27189; shr.u32 %r5330, %r5329, 23; add.s32 %r5331, %r5330, -127; and.b32 %r5332, %r5329, 8388607; or.b32 %r5333, %r5332, 1065353216; mov.b32 %f27190, %r5333; add.ftz.f32 %f27191, %f27190, 0fBF800000; mul.ftz.f32 %f27192, %f27191, %f27191; mul.ftz.f32 %f27193, %f27192, %f27192; fma.rn.ftz.f32 %f27194, %f27191, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f27195, %f27191, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f27196, %f27191, %f27194, 0fBE0CD4FD; fma.rn.ftz.f32 %f27197, %f27191, %f27196, 0f3E77ADBD; fma.rn.ftz.f32 %f27198, %f27191, %f27197, 0fBEB1D206; fma.rn.ftz.f32 %f27199, %f27191, %f27195, 0f3FB8AA10; mul.ftz.f32 %f27200, %f27193, %f27198; fma.rn.ftz.f32 %f27201, %f27191, %f27199, %f27200; cvt.rn.f32.s32 %f27202, %r5331; add.ftz.f32 %f27203, %f27201, %f27202; mul.ftz.f32 %f7208, %f27203, 0f3F317218; setp.gt.ftz.f32 %p3702, %f7208, 0fC0A00000; @%p3702 bra $L__BB107_4336; bra.uni $L__BB107_4335; $L__BB107_4336: mov.f32 %f27224, 0fC0200000; sub.ftz.f32 %f30797, %f27224, %f7208; mov.f32 %f30805, 0f32F16588; mov.f32 %f30804, 0f34B84B36; mov.f32 %f30803, 0fB66C7357; mov.f32 %f30802, 0fB6935AC1; mov.f32 %f30801, 0f396532DB; mov.f32 %f30800, 0fBAA45408; mov.f32 %f30799, 0fBB88E4EF; mov.f32 %f30798, 0f3E7C8F63; mov.f32 %f30796, 0f3FC02E2F; bra.uni $L__BB107_4337; $L__BB107_4068: mov.f32 %f25122, 0f3FB8AA3B; mul.rn.ftz.f32 %f25123, %f6428, %f25122; cvt.rzi.ftz.f32.f32 %f25124, %f25123; abs.ftz.f32 %f25125, %f25124; setp.gt.ftz.f32 %p3438, %f25125, 0f42FC0000; mov.b32 %r5223, %f25124; and.b32 %r5224, %r5223, -2147483648; or.b32 %r5225, %r5224, 1123811328; mov.b32 %f25126, %r5225; selp.f32 %f25127, %f25126, %f25124, %p3438; mov.f32 %f25128, 0fBF317218; fma.rn.ftz.f32 %f25129, %f25127, %f25128, %f6428; mov.f32 %f25130, 0f3102E308; fma.rn.ftz.f32 %f25131, %f25127, %f25130, %f25129; mul.ftz.f32 %f25132, %f25131, 0f3FB8AA3B; add.ftz.f32 %f25133, %f25127, 0f4B40007D; mov.b32 %r5226, %f25133; shl.b32 %r5227, %r5226, 23; mov.b32 %f25134, %r5227; ex2.approx.ftz.f32 %f25135, %f25132; mul.ftz.f32 %f25136, %f25135, %f25134; mov.f32 %f25137, 0f3E000000; div.approx.ftz.f32 %f25138, %f25137, %f25136; neg.ftz.f32 %f25139, %f25138; mov.f32 %f25140, 0f40000000; fma.rn.ftz.f32 %f25141, %f25140, %f25136, %f25139; setp.ge.ftz.f32 %p3439, %f6428, 0f42B40000; selp.f32 %f25142, 0f7F800000, %f25141, %p3439; mov.b32 %r5228, %f25142; mov.b32 %r5229, %f6384; and.b32 %r5230, %r5229, -2147483648; or.b32 %r5231, %r5230, %r5228; mov.b32 %f30646, %r5231; $L__BB107_4070: add.ftz.f32 %f25152, %f30646, %f30646; mul.ftz.f32 %f25153, %f6383, %f25152; div.approx.ftz.f32 %f30647, %f6427, %f25153; bra.uni $L__BB107_4071; $L__BB107_4046: mul.ftz.f32 %f6387, %f6385, %f6385; fma.rn.ftz.f32 %f30643, %f6387, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6389, %f6387, %f6387; mov.u64 %rd4403, 64; cvt.rn.f32.u64 %f25073, %rd4403; div.approx.ftz.f32 %f25074, %f6389, %f25073; add.ftz.f32 %f6390, %f30643, %f25074; setp.eq.ftz.f32 %p3421, %f30643, %f6390; @%p3421 bra $L__BB107_4055; mul.ftz.f32 %f6391, %f6387, %f6389; mov.u64 %rd4404, 2304; cvt.rn.f32.u64 %f25075, %rd4404; div.approx.ftz.f32 %f25076, %f6391, %f25075; add.ftz.f32 %f6392, %f6390, %f25076; setp.eq.ftz.f32 %p3422, %f6390, %f6392; mov.f32 %f30643, %f6390; @%p3422 bra $L__BB107_4055; mul.ftz.f32 %f6393, %f6387, %f6391; mov.u64 %rd4405, 147456; cvt.rn.f32.u64 %f25077, %rd4405; div.approx.ftz.f32 %f25078, %f6393, %f25077; add.ftz.f32 %f6394, %f6392, %f25078; setp.eq.ftz.f32 %p3423, %f6392, %f6394; mov.f32 %f30643, %f6392; @%p3423 bra $L__BB107_4055; mul.ftz.f32 %f6395, %f6387, %f6393; mov.u64 %rd4406, 14745600; cvt.rn.f32.u64 %f25079, %rd4406; div.approx.ftz.f32 %f25080, %f6395, %f25079; add.ftz.f32 %f6396, %f6394, %f25080; setp.eq.ftz.f32 %p3424, %f6394, %f6396; mov.f32 %f30643, %f6394; @%p3424 bra $L__BB107_4055; mul.ftz.f32 %f6397, %f6387, %f6395; mov.u64 %rd4407, 2123366400; cvt.rn.f32.u64 %f25081, %rd4407; div.approx.ftz.f32 %f25082, %f6397, %f25081; add.ftz.f32 %f6398, %f6396, %f25082; setp.eq.ftz.f32 %p3425, %f6396, %f6398; mov.f32 %f30643, %f6396; @%p3425 bra $L__BB107_4055; mul.ftz.f32 %f6399, %f6387, %f6397; mov.u64 %rd4408, 416179814400; cvt.rn.f32.u64 %f25083, %rd4408; div.approx.ftz.f32 %f25084, %f6399, %f25083; add.ftz.f32 %f6400, %f6398, %f25084; setp.eq.ftz.f32 %p3426, %f6398, %f6400; mov.f32 %f30643, %f6398; @%p3426 bra $L__BB107_4055; mul.ftz.f32 %f6401, %f6387, %f6399; mov.u64 %rd4409, 106542032486400; cvt.rn.f32.u64 %f25085, %rd4409; div.approx.ftz.f32 %f25086, %f6401, %f25085; add.ftz.f32 %f6402, %f6400, %f25086; setp.eq.ftz.f32 %p3427, %f6400, %f6402; mov.f32 %f30643, %f6400; @%p3427 bra $L__BB107_4055; mul.ftz.f32 %f25087, %f6387, %f6401; mov.u64 %rd4410, 34519618525593600; cvt.rn.f32.u64 %f25088, %rd4410; div.approx.ftz.f32 %f25089, %f25087, %f25088; add.ftz.f32 %f6403, %f6402, %f25089; setp.eq.ftz.f32 %p3428, %f6402, %f6403; mov.f32 %f30643, %f6402; @%p3428 bra $L__BB107_4055; mov.f32 %f30643, %f6403; $L__BB107_4055: lg2.approx.ftz.f32 %f25090, %f30643; mul.ftz.f32 %f30644, %f25090, 0f3F317218; $L__BB107_4057: sub.ftz.f32 %f25096, %f30644, %f6386; sub.ftz.f32 %f25097, %f25096, %f6384; add.ftz.f32 %f25098, %f25097, 0f3F316F00; mul.ftz.f32 %f25099, %f6384, 0f3F000000; lg2.approx.ftz.f32 %f25100, %f25099; fma.rn.ftz.f32 %f25101, %f25100, 0f3F317218, %f25098; mul.ftz.f32 %f25102, %f25101, 0f3FB8AA3B; ex2.approx.ftz.f32 %f30647, %f25102; $L__BB107_4071: setp.gt.ftz.f32 %p3440, %f30640, 0f42B00000; mov.f32 %f30648, %f30649; @%p3440 bra $L__BB107_4073; ex2.approx.ftz.f32 %f25155, %f6382; add.ftz.f32 %f25156, %f25155, 0f3F800000; rcp.approx.ftz.f32 %f25157, %f25156; add.ftz.f32 %f30648, %f25157, %f25157; $L__BB107_4073: sub.ftz.f32 %f6436, %f24954, %f30648; setp.eq.ftz.f32 %p3441, %f6436, 0f00000000; @%p3441 bra $L__BB107_4075; mul.ftz.f32 %f25160, %f6314, 0f00000000; sub.ftz.f32 %f25161, %f25160, %f6365; add.ftz.f32 %f25162, %f25161, 0f00000000; sub.ftz.f32 %f25163, %f30642, %f25162; add.ftz.f32 %f25164, %f25163, 0f40490FDB; mov.f32 %f25165, 0f40C90FDB; div.approx.ftz.f32 %f25166, %f25164, %f25165; cvt.rmi.ftz.f32.f32 %f25167, %f25166; mul.ftz.f32 %f25168, %f25167, 0f40C90FDB; sub.ftz.f32 %f25169, %f25164, %f25168; add.ftz.f32 %f25170, %f25169, 0fC0490FDB; abs.ftz.f32 %f25171, %f25170; neg.ftz.f32 %f25172, %f25171; div.approx.ftz.f32 %f25173, %f25172, %f30641; mul.ftz.f32 %f25174, %f25173, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25175, %f25174; add.ftz.f32 %f25176, %f25175, 0f3F800000; mul.ftz.f32 %f25177, %f25176, %f25176; mul.ftz.f32 %f25178, %f30641, %f25177; div.approx.ftz.f32 %f25179, %f25175, %f25178; div.approx.ftz.f32 %f30649, %f25179, %f6436; $L__BB107_4075: mul.ftz.f32 %f25180, %f30647, %f30632; fma.rn.ftz.f32 %f6439, %f30649, %f25180, 0f00000000; mul.ftz.f32 %f25181, %f30647, %f6355; fma.rn.ftz.f32 %f6440, %f30649, %f25181, 0f00000000; mul.ftz.f32 %f25182, %f6306, %f6379; rcp.approx.ftz.f32 %f6441, %f6358; mul.ftz.f32 %f6442, %f25182, %f6441; mul.ftz.f32 %f25183, %f6301, %f6378; mul.ftz.f32 %f6443, %f25183, %f6441; setp.gtu.ftz.f32 %p3442, %f6358, 0f3DCCCCCD; @%p3442 bra $L__BB107_4089; bra.uni $L__BB107_4076; $L__BB107_4089: mul.ftz.f32 %f6466, %f6442, %f6442; fma.rn.ftz.f32 %f30652, %f6466, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6468, %f6466, %f6466; mov.u64 %rd4427, 64; cvt.rn.f32.u64 %f25214, %rd4427; div.approx.ftz.f32 %f25215, %f6468, %f25214; add.ftz.f32 %f6469, %f30652, %f25215; setp.eq.ftz.f32 %p3452, %f30652, %f6469; @%p3452 bra $L__BB107_4098; mul.ftz.f32 %f6470, %f6466, %f6468; mov.u64 %rd4428, 2304; cvt.rn.f32.u64 %f25216, %rd4428; div.approx.ftz.f32 %f25217, %f6470, %f25216; add.ftz.f32 %f6471, %f6469, %f25217; setp.eq.ftz.f32 %p3453, %f6469, %f6471; mov.f32 %f30652, %f6469; @%p3453 bra $L__BB107_4098; mul.ftz.f32 %f6472, %f6466, %f6470; mov.u64 %rd4429, 147456; cvt.rn.f32.u64 %f25218, %rd4429; div.approx.ftz.f32 %f25219, %f6472, %f25218; add.ftz.f32 %f6473, %f6471, %f25219; setp.eq.ftz.f32 %p3454, %f6471, %f6473; mov.f32 %f30652, %f6471; @%p3454 bra $L__BB107_4098; mul.ftz.f32 %f6474, %f6466, %f6472; mov.u64 %rd4430, 14745600; cvt.rn.f32.u64 %f25220, %rd4430; div.approx.ftz.f32 %f25221, %f6474, %f25220; add.ftz.f32 %f6475, %f6473, %f25221; setp.eq.ftz.f32 %p3455, %f6473, %f6475; mov.f32 %f30652, %f6473; @%p3455 bra $L__BB107_4098; mul.ftz.f32 %f6476, %f6466, %f6474; mov.u64 %rd4431, 2123366400; cvt.rn.f32.u64 %f25222, %rd4431; div.approx.ftz.f32 %f25223, %f6476, %f25222; add.ftz.f32 %f6477, %f6475, %f25223; setp.eq.ftz.f32 %p3456, %f6475, %f6477; mov.f32 %f30652, %f6475; @%p3456 bra $L__BB107_4098; mul.ftz.f32 %f6478, %f6466, %f6476; mov.u64 %rd4432, 416179814400; cvt.rn.f32.u64 %f25224, %rd4432; div.approx.ftz.f32 %f25225, %f6478, %f25224; add.ftz.f32 %f6479, %f6477, %f25225; setp.eq.ftz.f32 %p3457, %f6477, %f6479; mov.f32 %f30652, %f6477; @%p3457 bra $L__BB107_4098; mul.ftz.f32 %f6480, %f6466, %f6478; mov.u64 %rd4433, 106542032486400; cvt.rn.f32.u64 %f25226, %rd4433; div.approx.ftz.f32 %f25227, %f6480, %f25226; add.ftz.f32 %f6481, %f6479, %f25227; setp.eq.ftz.f32 %p3458, %f6479, %f6481; mov.f32 %f30652, %f6479; @%p3458 bra $L__BB107_4098; mul.ftz.f32 %f25228, %f6466, %f6480; mov.u64 %rd4434, 34519618525593600; cvt.rn.f32.u64 %f25229, %rd4434; div.approx.ftz.f32 %f25230, %f25228, %f25229; add.ftz.f32 %f6482, %f6481, %f25230; setp.eq.ftz.f32 %p3459, %f6481, %f6482; mov.f32 %f30652, %f6481; @%p3459 bra $L__BB107_4098; mov.f32 %f30652, %f6482; $L__BB107_4098: mul.ftz.f32 %f25231, %f6443, 0fBFB8AA3B; ex2.approx.ftz.f32 %f25232, %f25231; mul.ftz.f32 %f6484, %f30652, %f25232; abs.ftz.f32 %f6485, %f6441; setp.ltu.ftz.f32 %p3460, %f6485, 0f3F800000; @%p3460 bra $L__BB107_4100; bra.uni $L__BB107_4099; $L__BB107_4100: mul.ftz.f32 %f25254, %f6441, %f6441; mov.f32 %f25255, 0f394FFF49; mov.f32 %f25256, 0f363D0ADA; fma.rn.ftz.f32 %f25257, %f25256, %f25254, %f25255; mov.f32 %f25258, 0f3C08889A; fma.rn.ftz.f32 %f25259, %f25257, %f25254, %f25258; mov.f32 %f25260, 0f3E2AAAAB; fma.rn.ftz.f32 %f25261, %f25259, %f25254, %f25260; mul.ftz.f32 %f25262, %f25254, %f25261; fma.rn.ftz.f32 %f30653, %f25262, %f6441, %f6441; bra.uni $L__BB107_4101; $L__BB107_4076: setp.gt.ftz.f32 %p3443, %f6442, 0f41400000; @%p3443 bra $L__BB107_4087; bra.uni $L__BB107_4077; $L__BB107_4087: mul.ftz.f32 %f25202, %f6442, 0f41000000; rcp.approx.ftz.f32 %f25203, %f25202; add.ftz.f32 %f25204, %f25203, 0fBFEB3F8E; lg2.approx.ftz.f32 %f25205, %f6442; fma.rn.ftz.f32 %f25206, %f25205, 0fBF317218, %f25204; fma.rn.ftz.f32 %f30651, %f25206, 0f3F000000, %f6442; bra.uni $L__BB107_4088; $L__BB107_4335: neg.ftz.f32 %f27213, %f7208; sqrt.approx.ftz.f32 %f27214, %f27213; add.ftz.f32 %f30797, %f27214, 0fC0400000; mov.f32 %f30805, 0fB951F09B; mov.f32 %f30804, 0f38D3B56B; mov.f32 %f30803, 0f3AB0DC72; mov.f32 %f30802, 0fBB70BDE7; mov.f32 %f30801, 0f3BBC127B; mov.f32 %f30800, 0fBBF9C5D7; mov.f32 %f30799, 0f3C1AA57E; mov.f32 %f30798, 0f3F8036DB; mov.f32 %f30796, 0f40354F7E; $L__BB107_4337: fma.rn.ftz.f32 %f27225, %f30797, %f30805, %f30804; fma.rn.ftz.f32 %f27226, %f30797, %f27225, %f30803; fma.rn.ftz.f32 %f27227, %f30797, %f27226, %f30802; fma.rn.ftz.f32 %f27228, %f30797, %f27227, %f30801; fma.rn.ftz.f32 %f27229, %f30797, %f27228, %f30800; fma.rn.ftz.f32 %f27230, %f30797, %f27229, %f30799; fma.rn.ftz.f32 %f27231, %f30797, %f27230, %f30798; fma.rn.ftz.f32 %f27232, %f30797, %f27231, %f30796; mul.ftz.f32 %f30839, %f30795, %f27232; mul.ftz.f32 %f27233, %f30839, %f30839; mul.ftz.f32 %f27234, %f27233, 0fBFB8AA3B; ex2.approx.ftz.f32 %f27235, %f27234; add.ftz.f32 %f27236, %f30795, 0f3F800000; fma.rn.ftz.f32 %f27237, %f7202, %f27235, %f27236; sub.ftz.f32 %f7222, %f27237, %f7204; abs.ftz.f32 %f27238, %f7222; setp.leu.ftz.f32 %p3703, %f27238, 0f358637BD; @%p3703 bra $L__BB107_4345; mul.ftz.f32 %f27239, %f7198, %f30839; sub.ftz.f32 %f27241, %f27166, %f27239; div.approx.ftz.f32 %f27242, %f7222, %f27241; sub.ftz.f32 %f7223, %f30795, %f27242; abs.ftz.f32 %f27243, %f7223; setp.gt.ftz.f32 %p3704, %f27243, 0f3F7FFFFF; selp.f32 %f27244, 0f3F7FFFFF, %f27243, %p3704; sub.ftz.f32 %f27245, %f27166, %f27244; add.ftz.f32 %f27246, %f27244, 0f3F800000; mul.ftz.f32 %f27247, %f27245, %f27246; max.ftz.f32 %f27249, %f27247, %f27186; min.ftz.f32 %f27251, %f27249, %f27188; mov.b32 %r5334, %f27251; shr.u32 %r5335, %r5334, 23; add.s32 %r5336, %r5335, -127; and.b32 %r5337, %r5334, 8388607; or.b32 %r5338, %r5337, 1065353216; mov.b32 %f27252, %r5338; add.ftz.f32 %f27253, %f27252, 0fBF800000; mul.ftz.f32 %f27254, %f27253, %f27253; mul.ftz.f32 %f27255, %f27254, %f27254; fma.rn.ftz.f32 %f27256, %f27253, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f27257, %f27253, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f27258, %f27253, %f27256, 0fBE0CD4FD; fma.rn.ftz.f32 %f27259, %f27253, %f27258, 0f3E77ADBD; fma.rn.ftz.f32 %f27260, %f27253, %f27259, 0fBEB1D206; fma.rn.ftz.f32 %f27261, %f27253, %f27257, 0f3FB8AA10; mul.ftz.f32 %f27262, %f27255, %f27260; fma.rn.ftz.f32 %f27263, %f27253, %f27261, %f27262; cvt.rn.f32.s32 %f27264, %r5336; add.ftz.f32 %f27265, %f27263, %f27264; mul.ftz.f32 %f7224, %f27265, 0f3F317218; setp.gt.ftz.f32 %p3705, %f7224, 0fC0A00000; @%p3705 bra $L__BB107_4340; bra.uni $L__BB107_4339; $L__BB107_4340: mov.f32 %f27286, 0fC0200000; sub.ftz.f32 %f30807, %f27286, %f7224; mov.f32 %f30815, 0f32F16588; mov.f32 %f30814, 0f34B84B36; mov.f32 %f30813, 0fB66C7357; mov.f32 %f30812, 0fB6935AC1; mov.f32 %f30811, 0f396532DB; mov.f32 %f30810, 0fBAA45408; mov.f32 %f30809, 0fBB88E4EF; mov.f32 %f30808, 0f3E7C8F63; mov.f32 %f30806, 0f3FC02E2F; bra.uni $L__BB107_4341; $L__BB107_4099: mov.f32 %f25233, 0f3FB8AA3B; mul.rn.ftz.f32 %f25234, %f6485, %f25233; cvt.rzi.ftz.f32.f32 %f25235, %f25234; abs.ftz.f32 %f25236, %f25235; setp.gt.ftz.f32 %p3461, %f25236, 0f42FC0000; mov.b32 %r5232, %f25235; and.b32 %r5233, %r5232, -2147483648; or.b32 %r5234, %r5233, 1123811328; mov.b32 %f25237, %r5234; selp.f32 %f25238, %f25237, %f25235, %p3461; mov.f32 %f25239, 0fBF317218; fma.rn.ftz.f32 %f25240, %f25238, %f25239, %f6485; mov.f32 %f25241, 0f3102E308; fma.rn.ftz.f32 %f25242, %f25238, %f25241, %f25240; mul.ftz.f32 %f25243, %f25242, 0f3FB8AA3B; add.ftz.f32 %f25244, %f25238, 0f4B40007D; mov.b32 %r5235, %f25244; shl.b32 %r5236, %r5235, 23; mov.b32 %f25245, %r5236; ex2.approx.ftz.f32 %f25246, %f25243; mul.ftz.f32 %f25247, %f25246, %f25245; mov.f32 %f25248, 0f3E000000; div.approx.ftz.f32 %f25249, %f25248, %f25247; neg.ftz.f32 %f25250, %f25249; mov.f32 %f25251, 0f40000000; fma.rn.ftz.f32 %f25252, %f25251, %f25247, %f25250; setp.ge.ftz.f32 %p3462, %f6485, 0f42B40000; selp.f32 %f25253, 0f7F800000, %f25252, %p3462; mov.b32 %r5237, %f25253; mov.b32 %r5238, %f6441; and.b32 %r5239, %r5238, -2147483648; or.b32 %r5240, %r5239, %r5237; mov.b32 %f30653, %r5240; $L__BB107_4101: add.ftz.f32 %f25263, %f30653, %f30653; mul.ftz.f32 %f25264, %f6358, %f25263; div.approx.ftz.f32 %f30654, %f6484, %f25264; bra.uni $L__BB107_4102; $L__BB107_4077: mul.ftz.f32 %f6444, %f6442, %f6442; fma.rn.ftz.f32 %f30650, %f6444, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6446, %f6444, %f6444; mov.u64 %rd4419, 64; cvt.rn.f32.u64 %f25184, %rd4419; div.approx.ftz.f32 %f25185, %f6446, %f25184; add.ftz.f32 %f6447, %f30650, %f25185; setp.eq.ftz.f32 %p3444, %f30650, %f6447; @%p3444 bra $L__BB107_4086; mul.ftz.f32 %f6448, %f6444, %f6446; mov.u64 %rd4420, 2304; cvt.rn.f32.u64 %f25186, %rd4420; div.approx.ftz.f32 %f25187, %f6448, %f25186; add.ftz.f32 %f6449, %f6447, %f25187; setp.eq.ftz.f32 %p3445, %f6447, %f6449; mov.f32 %f30650, %f6447; @%p3445 bra $L__BB107_4086; mul.ftz.f32 %f6450, %f6444, %f6448; mov.u64 %rd4421, 147456; cvt.rn.f32.u64 %f25188, %rd4421; div.approx.ftz.f32 %f25189, %f6450, %f25188; add.ftz.f32 %f6451, %f6449, %f25189; setp.eq.ftz.f32 %p3446, %f6449, %f6451; mov.f32 %f30650, %f6449; @%p3446 bra $L__BB107_4086; mul.ftz.f32 %f6452, %f6444, %f6450; mov.u64 %rd4422, 14745600; cvt.rn.f32.u64 %f25190, %rd4422; div.approx.ftz.f32 %f25191, %f6452, %f25190; add.ftz.f32 %f6453, %f6451, %f25191; setp.eq.ftz.f32 %p3447, %f6451, %f6453; mov.f32 %f30650, %f6451; @%p3447 bra $L__BB107_4086; mul.ftz.f32 %f6454, %f6444, %f6452; mov.u64 %rd4423, 2123366400; cvt.rn.f32.u64 %f25192, %rd4423; div.approx.ftz.f32 %f25193, %f6454, %f25192; add.ftz.f32 %f6455, %f6453, %f25193; setp.eq.ftz.f32 %p3448, %f6453, %f6455; mov.f32 %f30650, %f6453; @%p3448 bra $L__BB107_4086; mul.ftz.f32 %f6456, %f6444, %f6454; mov.u64 %rd4424, 416179814400; cvt.rn.f32.u64 %f25194, %rd4424; div.approx.ftz.f32 %f25195, %f6456, %f25194; add.ftz.f32 %f6457, %f6455, %f25195; setp.eq.ftz.f32 %p3449, %f6455, %f6457; mov.f32 %f30650, %f6455; @%p3449 bra $L__BB107_4086; mul.ftz.f32 %f6458, %f6444, %f6456; mov.u64 %rd4425, 106542032486400; cvt.rn.f32.u64 %f25196, %rd4425; div.approx.ftz.f32 %f25197, %f6458, %f25196; add.ftz.f32 %f6459, %f6457, %f25197; setp.eq.ftz.f32 %p3450, %f6457, %f6459; mov.f32 %f30650, %f6457; @%p3450 bra $L__BB107_4086; mul.ftz.f32 %f25198, %f6444, %f6458; mov.u64 %rd4426, 34519618525593600; cvt.rn.f32.u64 %f25199, %rd4426; div.approx.ftz.f32 %f25200, %f25198, %f25199; add.ftz.f32 %f6460, %f6459, %f25200; setp.eq.ftz.f32 %p3451, %f6459, %f6460; mov.f32 %f30650, %f6459; @%p3451 bra $L__BB107_4086; mov.f32 %f30650, %f6460; $L__BB107_4086: lg2.approx.ftz.f32 %f25201, %f30650; mul.ftz.f32 %f30651, %f25201, 0f3F317218; $L__BB107_4088: sub.ftz.f32 %f25207, %f30651, %f6443; sub.ftz.f32 %f25208, %f25207, %f6441; add.ftz.f32 %f25209, %f25208, 0f3F316F00; mul.ftz.f32 %f25210, %f6441, 0f3F000000; lg2.approx.ftz.f32 %f25211, %f25210; fma.rn.ftz.f32 %f25212, %f25211, 0f3F317218, %f25209; mul.ftz.f32 %f25213, %f25212, 0f3FB8AA3B; ex2.approx.ftz.f32 %f30654, %f25213; $L__BB107_4102: mov.f32 %f30656, 0f00000000; mov.f32 %f30655, %f30656; @%p3440 bra $L__BB107_4104; ex2.approx.ftz.f32 %f25266, %f6382; add.ftz.f32 %f25267, %f25266, 0f3F800000; rcp.approx.ftz.f32 %f25268, %f25267; add.ftz.f32 %f30655, %f25268, %f25268; $L__BB107_4104: sub.ftz.f32 %f6493, %f24954, %f30655; setp.eq.ftz.f32 %p3464, %f6493, 0f00000000; @%p3464 bra $L__BB107_4106; add.ftz.f32 %f25271, %f6314, %f6314; sub.ftz.f32 %f25272, %f25271, %f6365; add.ftz.f32 %f25273, %f25272, 0f40490FDB; sub.ftz.f32 %f25274, %f30642, %f25273; add.ftz.f32 %f25275, %f25274, 0f40490FDB; mov.f32 %f25276, 0f40C90FDB; div.approx.ftz.f32 %f25277, %f25275, %f25276; cvt.rmi.ftz.f32.f32 %f25278, %f25277; mul.ftz.f32 %f25279, %f25278, 0f40C90FDB; sub.ftz.f32 %f25280, %f25275, %f25279; add.ftz.f32 %f25281, %f25280, 0fC0490FDB; abs.ftz.f32 %f25282, %f25281; neg.ftz.f32 %f25283, %f25282; div.approx.ftz.f32 %f25284, %f25283, %f30641; mul.ftz.f32 %f25285, %f25284, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25286, %f25285; add.ftz.f32 %f25287, %f25286, 0f3F800000; mul.ftz.f32 %f25288, %f25287, %f25287; mul.ftz.f32 %f25289, %f30641, %f25288; div.approx.ftz.f32 %f25290, %f25286, %f25289; div.approx.ftz.f32 %f30656, %f25290, %f6493; $L__BB107_4106: mul.ftz.f32 %f25291, %f30654, %f6322; mul.ftz.f32 %f25292, %f30654, %f6323; mul.ftz.f32 %f25293, %f30654, %f6324; fma.rn.ftz.f32 %f6496, %f30656, %f25291, %f6439; fma.rn.ftz.f32 %f6497, %f30656, %f25292, %f6439; fma.rn.ftz.f32 %f6498, %f30656, %f25293, %f6439; mul.ftz.f32 %f25294, %f30654, %f6352; fma.rn.ftz.f32 %f6499, %f30656, %f25294, %f6440; mul.ftz.f32 %f6500, %f6357, 0f40800000; rcp.approx.ftz.f32 %f6501, %f6500; mul.ftz.f32 %f25295, %f6306, %f6381; mul.ftz.f32 %f6502, %f25295, %f6501; mul.ftz.f32 %f25296, %f6301, %f6380; mul.ftz.f32 %f6503, %f25296, %f6501; setp.gtu.ftz.f32 %p3465, %f6500, 0f3DCCCCCD; @%p3465 bra $L__BB107_4120; bra.uni $L__BB107_4107; $L__BB107_4120: mul.ftz.f32 %f6526, %f6502, %f6502; fma.rn.ftz.f32 %f30659, %f6526, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6528, %f6526, %f6526; mov.u64 %rd4443, 64; cvt.rn.f32.u64 %f25327, %rd4443; div.approx.ftz.f32 %f25328, %f6528, %f25327; add.ftz.f32 %f6529, %f30659, %f25328; setp.eq.ftz.f32 %p3475, %f30659, %f6529; @%p3475 bra $L__BB107_4129; mul.ftz.f32 %f6530, %f6526, %f6528; mov.u64 %rd4444, 2304; cvt.rn.f32.u64 %f25329, %rd4444; div.approx.ftz.f32 %f25330, %f6530, %f25329; add.ftz.f32 %f6531, %f6529, %f25330; setp.eq.ftz.f32 %p3476, %f6529, %f6531; mov.f32 %f30659, %f6529; @%p3476 bra $L__BB107_4129; mul.ftz.f32 %f6532, %f6526, %f6530; mov.u64 %rd4445, 147456; cvt.rn.f32.u64 %f25331, %rd4445; div.approx.ftz.f32 %f25332, %f6532, %f25331; add.ftz.f32 %f6533, %f6531, %f25332; setp.eq.ftz.f32 %p3477, %f6531, %f6533; mov.f32 %f30659, %f6531; @%p3477 bra $L__BB107_4129; mul.ftz.f32 %f6534, %f6526, %f6532; mov.u64 %rd4446, 14745600; cvt.rn.f32.u64 %f25333, %rd4446; div.approx.ftz.f32 %f25334, %f6534, %f25333; add.ftz.f32 %f6535, %f6533, %f25334; setp.eq.ftz.f32 %p3478, %f6533, %f6535; mov.f32 %f30659, %f6533; @%p3478 bra $L__BB107_4129; mul.ftz.f32 %f6536, %f6526, %f6534; mov.u64 %rd4447, 2123366400; cvt.rn.f32.u64 %f25335, %rd4447; div.approx.ftz.f32 %f25336, %f6536, %f25335; add.ftz.f32 %f6537, %f6535, %f25336; setp.eq.ftz.f32 %p3479, %f6535, %f6537; mov.f32 %f30659, %f6535; @%p3479 bra $L__BB107_4129; mul.ftz.f32 %f6538, %f6526, %f6536; mov.u64 %rd4448, 416179814400; cvt.rn.f32.u64 %f25337, %rd4448; div.approx.ftz.f32 %f25338, %f6538, %f25337; add.ftz.f32 %f6539, %f6537, %f25338; setp.eq.ftz.f32 %p3480, %f6537, %f6539; mov.f32 %f30659, %f6537; @%p3480 bra $L__BB107_4129; mul.ftz.f32 %f6540, %f6526, %f6538; mov.u64 %rd4449, 106542032486400; cvt.rn.f32.u64 %f25339, %rd4449; div.approx.ftz.f32 %f25340, %f6540, %f25339; add.ftz.f32 %f6541, %f6539, %f25340; setp.eq.ftz.f32 %p3481, %f6539, %f6541; mov.f32 %f30659, %f6539; @%p3481 bra $L__BB107_4129; mul.ftz.f32 %f25341, %f6526, %f6540; mov.u64 %rd4450, 34519618525593600; cvt.rn.f32.u64 %f25342, %rd4450; div.approx.ftz.f32 %f25343, %f25341, %f25342; add.ftz.f32 %f6542, %f6541, %f25343; setp.eq.ftz.f32 %p3482, %f6541, %f6542; mov.f32 %f30659, %f6541; @%p3482 bra $L__BB107_4129; mov.f32 %f30659, %f6542; $L__BB107_4129: mul.ftz.f32 %f25344, %f6503, 0fBFB8AA3B; ex2.approx.ftz.f32 %f25345, %f25344; mul.ftz.f32 %f6544, %f30659, %f25345; abs.ftz.f32 %f6545, %f6501; setp.ltu.ftz.f32 %p3483, %f6545, 0f3F800000; @%p3483 bra $L__BB107_4131; bra.uni $L__BB107_4130; $L__BB107_4131: mul.ftz.f32 %f25367, %f6501, %f6501; mov.f32 %f25368, 0f394FFF49; mov.f32 %f25369, 0f363D0ADA; fma.rn.ftz.f32 %f25370, %f25369, %f25367, %f25368; mov.f32 %f25371, 0f3C08889A; fma.rn.ftz.f32 %f25372, %f25370, %f25367, %f25371; mov.f32 %f25373, 0f3E2AAAAB; fma.rn.ftz.f32 %f25374, %f25372, %f25367, %f25373; mul.ftz.f32 %f25375, %f25367, %f25374; fma.rn.ftz.f32 %f30660, %f25375, %f6501, %f6501; bra.uni $L__BB107_4132; $L__BB107_4107: setp.gt.ftz.f32 %p3466, %f6502, 0f41400000; @%p3466 bra $L__BB107_4118; bra.uni $L__BB107_4108; $L__BB107_4118: mul.ftz.f32 %f25315, %f6502, 0f41000000; rcp.approx.ftz.f32 %f25316, %f25315; add.ftz.f32 %f25317, %f25316, 0fBFEB3F8E; lg2.approx.ftz.f32 %f25318, %f6502; fma.rn.ftz.f32 %f25319, %f25318, 0fBF317218, %f25317; fma.rn.ftz.f32 %f30658, %f25319, 0f3F000000, %f6502; bra.uni $L__BB107_4119; $L__BB107_4130: mov.f32 %f25346, 0f3FB8AA3B; mul.rn.ftz.f32 %f25347, %f6545, %f25346; cvt.rzi.ftz.f32.f32 %f25348, %f25347; abs.ftz.f32 %f25349, %f25348; setp.gt.ftz.f32 %p3484, %f25349, 0f42FC0000; mov.b32 %r5241, %f25348; and.b32 %r5242, %r5241, -2147483648; or.b32 %r5243, %r5242, 1123811328; mov.b32 %f25350, %r5243; selp.f32 %f25351, %f25350, %f25348, %p3484; mov.f32 %f25352, 0fBF317218; fma.rn.ftz.f32 %f25353, %f25351, %f25352, %f6545; mov.f32 %f25354, 0f3102E308; fma.rn.ftz.f32 %f25355, %f25351, %f25354, %f25353; mul.ftz.f32 %f25356, %f25355, 0f3FB8AA3B; add.ftz.f32 %f25357, %f25351, 0f4B40007D; mov.b32 %r5244, %f25357; shl.b32 %r5245, %r5244, 23; mov.b32 %f25358, %r5245; ex2.approx.ftz.f32 %f25359, %f25356; mul.ftz.f32 %f25360, %f25359, %f25358; mov.f32 %f25361, 0f3E000000; div.approx.ftz.f32 %f25362, %f25361, %f25360; neg.ftz.f32 %f25363, %f25362; mov.f32 %f25364, 0f40000000; fma.rn.ftz.f32 %f25365, %f25364, %f25360, %f25363; setp.ge.ftz.f32 %p3485, %f6545, 0f42B40000; selp.f32 %f25366, 0f7F800000, %f25365, %p3485; mov.b32 %r5246, %f25366; mov.b32 %r5247, %f6501; and.b32 %r5248, %r5247, -2147483648; or.b32 %r5249, %r5248, %r5246; mov.b32 %f30660, %r5249; $L__BB107_4132: add.ftz.f32 %f25376, %f30660, %f30660; mul.ftz.f32 %f25377, %f6500, %f25376; div.approx.ftz.f32 %f30661, %f6544, %f25377; bra.uni $L__BB107_4133; $L__BB107_4108: mul.ftz.f32 %f6504, %f6502, %f6502; fma.rn.ftz.f32 %f30657, %f6504, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6506, %f6504, %f6504; mov.u64 %rd4435, 64; cvt.rn.f32.u64 %f25297, %rd4435; div.approx.ftz.f32 %f25298, %f6506, %f25297; add.ftz.f32 %f6507, %f30657, %f25298; setp.eq.ftz.f32 %p3467, %f30657, %f6507; @%p3467 bra $L__BB107_4117; mul.ftz.f32 %f6508, %f6504, %f6506; mov.u64 %rd4436, 2304; cvt.rn.f32.u64 %f25299, %rd4436; div.approx.ftz.f32 %f25300, %f6508, %f25299; add.ftz.f32 %f6509, %f6507, %f25300; setp.eq.ftz.f32 %p3468, %f6507, %f6509; mov.f32 %f30657, %f6507; @%p3468 bra $L__BB107_4117; mul.ftz.f32 %f6510, %f6504, %f6508; mov.u64 %rd4437, 147456; cvt.rn.f32.u64 %f25301, %rd4437; div.approx.ftz.f32 %f25302, %f6510, %f25301; add.ftz.f32 %f6511, %f6509, %f25302; setp.eq.ftz.f32 %p3469, %f6509, %f6511; mov.f32 %f30657, %f6509; @%p3469 bra $L__BB107_4117; mul.ftz.f32 %f6512, %f6504, %f6510; mov.u64 %rd4438, 14745600; cvt.rn.f32.u64 %f25303, %rd4438; div.approx.ftz.f32 %f25304, %f6512, %f25303; add.ftz.f32 %f6513, %f6511, %f25304; setp.eq.ftz.f32 %p3470, %f6511, %f6513; mov.f32 %f30657, %f6511; @%p3470 bra $L__BB107_4117; mul.ftz.f32 %f6514, %f6504, %f6512; mov.u64 %rd4439, 2123366400; cvt.rn.f32.u64 %f25305, %rd4439; div.approx.ftz.f32 %f25306, %f6514, %f25305; add.ftz.f32 %f6515, %f6513, %f25306; setp.eq.ftz.f32 %p3471, %f6513, %f6515; mov.f32 %f30657, %f6513; @%p3471 bra $L__BB107_4117; mul.ftz.f32 %f6516, %f6504, %f6514; mov.u64 %rd4440, 416179814400; cvt.rn.f32.u64 %f25307, %rd4440; div.approx.ftz.f32 %f25308, %f6516, %f25307; add.ftz.f32 %f6517, %f6515, %f25308; setp.eq.ftz.f32 %p3472, %f6515, %f6517; mov.f32 %f30657, %f6515; @%p3472 bra $L__BB107_4117; mul.ftz.f32 %f6518, %f6504, %f6516; mov.u64 %rd4441, 106542032486400; cvt.rn.f32.u64 %f25309, %rd4441; div.approx.ftz.f32 %f25310, %f6518, %f25309; add.ftz.f32 %f6519, %f6517, %f25310; setp.eq.ftz.f32 %p3473, %f6517, %f6519; mov.f32 %f30657, %f6517; @%p3473 bra $L__BB107_4117; mul.ftz.f32 %f25311, %f6504, %f6518; mov.u64 %rd4442, 34519618525593600; cvt.rn.f32.u64 %f25312, %rd4442; div.approx.ftz.f32 %f25313, %f25311, %f25312; add.ftz.f32 %f6520, %f6519, %f25313; setp.eq.ftz.f32 %p3474, %f6519, %f6520; mov.f32 %f30657, %f6519; @%p3474 bra $L__BB107_4117; mov.f32 %f30657, %f6520; $L__BB107_4117: lg2.approx.ftz.f32 %f25314, %f30657; mul.ftz.f32 %f30658, %f25314, 0f3F317218; $L__BB107_4119: sub.ftz.f32 %f25320, %f30658, %f6503; sub.ftz.f32 %f25321, %f25320, %f6501; add.ftz.f32 %f25322, %f25321, 0f3F316F00; mul.ftz.f32 %f25323, %f6501, 0f3F000000; lg2.approx.ftz.f32 %f25324, %f25323; fma.rn.ftz.f32 %f25325, %f25324, 0f3F317218, %f25322; mul.ftz.f32 %f25326, %f25325, 0f3FB8AA3B; ex2.approx.ftz.f32 %f30661, %f25326; $L__BB107_4133: mul.ftz.f32 %f25379, %f6314, 0f40800000; sub.ftz.f32 %f6551, %f25379, %f6365; mov.f32 %f30663, 0f00000000; mov.f32 %f30662, %f30663; @%p3440 bra $L__BB107_4135; ex2.approx.ftz.f32 %f25380, %f6382; add.ftz.f32 %f25381, %f25380, 0f3F800000; rcp.approx.ftz.f32 %f25382, %f25381; add.ftz.f32 %f30662, %f25382, %f25382; $L__BB107_4135: add.ftz.f32 %f25384, %f6551, 0f40C90FDB; mov.f32 %f25385, 0f40C90FDB; sub.ftz.f32 %f25386, %f30642, %f25384; add.ftz.f32 %f25387, %f25386, 0f40490FDB; div.approx.ftz.f32 %f25388, %f25387, %f25385; cvt.rmi.ftz.f32.f32 %f25389, %f25388; mul.ftz.f32 %f25390, %f25389, 0f40C90FDB; sub.ftz.f32 %f25391, %f25387, %f25390; add.ftz.f32 %f25392, %f25391, 0fC0490FDB; abs.ftz.f32 %f25393, %f25392; neg.ftz.f32 %f25394, %f25393; div.approx.ftz.f32 %f25395, %f25394, %f30641; mul.ftz.f32 %f25396, %f25395, 0f3FB8AA3B; ex2.approx.ftz.f32 %f25397, %f25396; add.ftz.f32 %f25398, %f25397, 0f3F800000; mul.ftz.f32 %f25400, %f25398, %f25398; mul.ftz.f32 %f25401, %f30641, %f25400; div.approx.ftz.f32 %f6554, %f25397, %f25401; sub.ftz.f32 %f6555, %f24954, %f30662; setp.eq.ftz.f32 %p3487, %f6555, 0f00000000; @%p3487 bra $L__BB107_4137; div.approx.ftz.f32 %f30663, %f6554, %f6555; $L__BB107_4137: mul.ftz.f32 %f25402, %f30661, %f6330; mul.ftz.f32 %f25403, %f30661, %f6332; mul.ftz.f32 %f25404, %f30661, %f6334; fma.rn.ftz.f32 %f6558, %f30663, %f25402, %f6496; fma.rn.ftz.f32 %f6559, %f30663, %f25403, %f6497; fma.rn.ftz.f32 %f6560, %f30663, %f25404, %f6498; mul.ftz.f32 %f25405, %f30661, %f6353; fma.rn.ftz.f32 %f6561, %f30663, %f25405, %f6499; add.ftz.f32 %f6562, %f30631, %f30642; mul.ftz.f32 %f25406, %f6306, %f30638; mul.ftz.f32 %f6563, %f25406, %f6501; mul.ftz.f32 %f25407, %f6301, %f30637; mul.ftz.f32 %f6564, %f25407, %f6501; @%p3465 bra $L__BB107_4151; bra.uni $L__BB107_4138; $L__BB107_4151: mul.ftz.f32 %f6587, %f6563, %f6563; fma.rn.ftz.f32 %f30666, %f6587, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6589, %f6587, %f6587; mov.u64 %rd4459, 64; cvt.rn.f32.u64 %f25438, %rd4459; div.approx.ftz.f32 %f25439, %f6589, %f25438; add.ftz.f32 %f6590, %f30666, %f25439; setp.eq.ftz.f32 %p3498, %f30666, %f6590; @%p3498 bra $L__BB107_4160; mul.ftz.f32 %f6591, %f6587, %f6589; mov.u64 %rd4460, 2304; cvt.rn.f32.u64 %f25440, %rd4460; div.approx.ftz.f32 %f25441, %f6591, %f25440; add.ftz.f32 %f6592, %f6590, %f25441; setp.eq.ftz.f32 %p3499, %f6590, %f6592; mov.f32 %f30666, %f6590; @%p3499 bra $L__BB107_4160; mul.ftz.f32 %f6593, %f6587, %f6591; mov.u64 %rd4461, 147456; cvt.rn.f32.u64 %f25442, %rd4461; div.approx.ftz.f32 %f25443, %f6593, %f25442; add.ftz.f32 %f6594, %f6592, %f25443; setp.eq.ftz.f32 %p3500, %f6592, %f6594; mov.f32 %f30666, %f6592; @%p3500 bra $L__BB107_4160; mul.ftz.f32 %f6595, %f6587, %f6593; mov.u64 %rd4462, 14745600; cvt.rn.f32.u64 %f25444, %rd4462; div.approx.ftz.f32 %f25445, %f6595, %f25444; add.ftz.f32 %f6596, %f6594, %f25445; setp.eq.ftz.f32 %p3501, %f6594, %f6596; mov.f32 %f30666, %f6594; @%p3501 bra $L__BB107_4160; mul.ftz.f32 %f6597, %f6587, %f6595; mov.u64 %rd4463, 2123366400; cvt.rn.f32.u64 %f25446, %rd4463; div.approx.ftz.f32 %f25447, %f6597, %f25446; add.ftz.f32 %f6598, %f6596, %f25447; setp.eq.ftz.f32 %p3502, %f6596, %f6598; mov.f32 %f30666, %f6596; @%p3502 bra $L__BB107_4160; mul.ftz.f32 %f6599, %f6587, %f6597; mov.u64 %rd4464, 416179814400; cvt.rn.f32.u64 %f25448, %rd4464; div.approx.ftz.f32 %f25449, %f6599, %f25448; add.ftz.f32 %f6600, %f6598, %f25449; setp.eq.ftz.f32 %p3503, %f6598, %f6600; mov.f32 %f30666, %f6598; @%p3503 bra $L__BB107_4160; mul.ftz.f32 %f6601, %f6587, %f6599; mov.u64 %rd4465, 106542032486400; cvt.rn.f32.u64 %f25450, %rd4465; div.approx.ftz.f32 %f25451, %f6601, %f25450; add.ftz.f32 %f6602, %f6600, %f25451; setp.eq.ftz.f32 %p3504, %f6600, %f6602; mov.f32 %f30666, %f6600; @%p3504 bra $L__BB107_4160; mul.ftz.f32 %f25452, %f6587, %f6601; mov.u64 %rd4466, 34519618525593600; cvt.rn.f32.u64 %f25453, %rd4466; div.approx.ftz.f32 %f25454, %f25452, %f25453; add.ftz.f32 %f6603, %f6602, %f25454; setp.eq.ftz.f32 %p3505, %f6602, %f6603; mov.f32 %f30666, %f6602; @%p3505 bra $L__BB107_4160; mov.f32 %f30666, %f6603; $L__BB107_4160: mul.ftz.f32 %f25455, %f6564, 0fBFB8AA3B; ex2.approx.ftz.f32 %f25456, %f25455; mul.ftz.f32 %f6605, %f30666, %f25456; abs.ftz.f32 %f6606, %f6501; setp.ltu.ftz.f32 %p3506, %f6606, 0f3F800000; @%p3506 bra $L__BB107_4162; bra.uni $L__BB107_4161; $L__BB107_4162: mul.ftz.f32 %f25478, %f6501, %f6501; mov.f32 %f25479, 0f394FFF49; mov.f32 %f25480, 0f363D0ADA; fma.rn.ftz.f32 %f25481, %f25480, %f25478, %f25479; mov.f32 %f25482, 0f3C08889A; fma.rn.ftz.f32 %f25483, %f25481, %f25478, %f25482; mov.f32 %f25484, 0f3E2AAAAB; fma.rn.ftz.f32 %f25485, %f25483, %f25478, %f25484; mul.ftz.f32 %f25486, %f25478, %f25485; fma.rn.ftz.f32 %f30667, %f25486, %f6501, %f6501; bra.uni $L__BB107_4163; $L__BB107_4138: setp.gt.ftz.f32 %p3489, %f6563, 0f41400000; @%p3489 bra $L__BB107_4149; bra.uni $L__BB107_4139; $L__BB107_4149: mul.ftz.f32 %f25426, %f6563, 0f41000000; rcp.approx.ftz.f32 %f25427, %f25426; add.ftz.f32 %f25428, %f25427, 0fBFEB3F8E; lg2.approx.ftz.f32 %f25429, %f6563; fma.rn.ftz.f32 %f25430, %f25429, 0fBF317218, %f25428; fma.rn.ftz.f32 %f30665, %f25430, 0f3F000000, %f6563; bra.uni $L__BB107_4150; $L__BB107_4161: mov.f32 %f25457, 0f3FB8AA3B; mul.rn.ftz.f32 %f25458, %f6606, %f25457; cvt.rzi.ftz.f32.f32 %f25459, %f25458; abs.ftz.f32 %f25460, %f25459; setp.gt.ftz.f32 %p3507, %f25460, 0f42FC0000; mov.b32 %r5250, %f25459; and.b32 %r5251, %r5250, -2147483648; or.b32 %r5252, %r5251, 1123811328; mov.b32 %f25461, %r5252; selp.f32 %f25462, %f25461, %f25459, %p3507; mov.f32 %f25463, 0fBF317218; fma.rn.ftz.f32 %f25464, %f25462, %f25463, %f6606; mov.f32 %f25465, 0f3102E308; fma.rn.ftz.f32 %f25466, %f25462, %f25465, %f25464; mul.ftz.f32 %f25467, %f25466, 0f3FB8AA3B; add.ftz.f32 %f25468, %f25462, 0f4B40007D; mov.b32 %r5253, %f25468; shl.b32 %r5254, %r5253, 23; mov.b32 %f25469, %r5254; ex2.approx.ftz.f32 %f25470, %f25467; mul.ftz.f32 %f25471, %f25470, %f25469; mov.f32 %f25472, 0f3E000000; div.approx.ftz.f32 %f25473, %f25472, %f25471; neg.ftz.f32 %f25474, %f25473; mov.f32 %f25475, 0f40000000; fma.rn.ftz.f32 %f25476, %f25475, %f25471, %f25474; setp.ge.ftz.f32 %p3508, %f6606, 0f42B40000; selp.f32 %f25477, 0f7F800000, %f25476, %p3508; mov.b32 %r5255, %f25477; mov.b32 %r5256, %f6501; and.b32 %r5257, %r5256, -2147483648; or.b32 %r5258, %r5257, %r5255; mov.b32 %f30667, %r5258; $L__BB107_4163: add.ftz.f32 %f25487, %f30667, %f30667; mul.ftz.f32 %f25488, %f6500, %f25487; div.approx.ftz.f32 %f30668, %f6605, %f25488; bra.uni $L__BB107_4164; $L__BB107_4139: mul.ftz.f32 %f6565, %f6563, %f6563; fma.rn.ftz.f32 %f30664, %f6565, 0f3E800000, 0f3F800000; mul.ftz.f32 %f6567, %f6565, %f6565; mov.u64 %rd4451, 64; cvt.rn.f32.u64 %f25408, %rd4451; div.approx.ftz.f32 %f25409, %f6567, %f25408; add.ftz.f32 %f6568, %f30664, %f25409; setp.eq.ftz.f32 %p3490, %f30664, %f6568; @%p3490 bra $L__BB107_4148; mul.ftz.f32 %f6569, %f6565, %f6567; mov.u64 %rd4452, 2304; cvt.rn.f32.u64 %f25410, %rd4452; div.approx.ftz.f32 %f25411, %f6569, %f25410; add.ftz.f32 %f6570, %f6568, %f25411; setp.eq.ftz.f32 %p3491, %f6568, %f6570; mov.f32 %f30664, %f6568; @%p3491 bra $L__BB107_4148; mul.ftz.f32 %f6571, %f6565, %f6569; mov.u64 %rd4453, 147456; cvt.rn.f32.u64 %f25412, %rd4453; div.approx.ftz.f32 %f25413, %f6571, %f25412; add.ftz.f32 %f6572, %f6570, %f25413; setp.eq.ftz.f32 %p3492, %f6570, %f6572; mov.f32 %f30664, %f6570; @%p3492 bra $L__BB107_4148; mul.ftz.f32 %f6573, %f6565, %f6571; mov.u64 %rd4454, 14745600; cvt.rn.f32.u64 %f25414, %rd4454; div.approx.ftz.f32 %f25415, %f6573, %f25414; add.ftz.f32 %f6574, %f6572, %f25415; setp.eq.ftz.f32 %p3493, %f6572, %f6574; mov.f32 %f30664, %f6572; @%p3493 bra $L__BB107_4148; mul.ftz.f32 %f6575, %f6565, %f6573; mov.u64 %rd4455, 2123366400; cvt.rn.f32.u64 %f25416, %rd4455; div.approx.ftz.f32 %f25417, %f6575, %f25416; add.ftz.f32 %f6576, %f6574, %f25417; setp.eq.ftz.f32 %p3494, %f6574, %f6576; mov.f32 %f30664, %f6574; @%p3494 bra $L__BB107_4148; mul.ftz.f32 %f6577, %f6565, %f6575; mov.u64 %rd4456, 416179814400; cvt.rn.f32.u64 %f25418, %rd4456; div.approx.ftz.f32 %f25419, %f6577, %f25418; add.ftz.f32 %f6578, %f6576, %f25419; setp.eq.ftz.f32 %p3495, %f6576, %f6578; mov.f32 %f30664, %f6576; @%p3495 bra $L__BB107_4148; mul.ftz.f32 %f6579, %f6565, %f6577; mov.u64 %rd4457, 106542032486400; cvt.rn.f32.u64 %f25420, %rd4457; div.approx.ftz.f32 %f25421, %f6579, %f25420; add.ftz.f32 %f6580, %f6578, %f25421; setp.eq.ftz.f32 %p3496, %f6578, %f6580; mov.f32 %f30664, %f6578; @%p3496 bra $L__BB107_4148; mul.ftz.f32 %f25422, %f6565, %f6579; mov.u64 %rd4458, 34519618525593600; cvt.rn.f32.u64 %f25423, %rd4458; div.approx.ftz.f32 %f25424, %f25422, %f25423; add.ftz.f32 %f6581, %f6580, %f25424; setp.eq.ftz.f32 %p3497, %f6580, %f6581; mov.f32 %f30664, %f6580; @%p3497 bra $L__BB107_4148; mov.f32 %f30664, %f6581; $L__BB107_4148: lg2.approx.ftz.f32 %f25425, %f30664; mul.ftz.f32 %f30665, %f25425, 0f3F317218; $L__BB107_4150: sub.ftz.f32 %f25431, %f30665, %f6564; sub.ftz.f32 %f25432, %f25431, %f6501; add.ftz.f32 %f25433, %f25432, 0f3F316F00; mul.ftz.f32 %f25434, %f6501, 0f3F000000; lg2.approx.ftz.f32 %f25435, %f25434; fma.rn.ftz.f32 %f25436, %f25435, 0f3F317218, %f25433; mul.ftz.f32 %f25437, %f25436, 0f3FB8AA3B; ex2.approx.ftz.f32 %f30668, %f25437; $L__BB107_4164: mul.ftz.f32 %f25489, %f30668, %f6345; mul.ftz.f32 %f25490, %f30668, %f6346; mul.ftz.f32 %f25491, %f30668, %f6347; fma.rn.ftz.f32 %f30903, %f25489, 0f3E22F983, %f6558; fma.rn.ftz.f32 %f30904, %f25490, 0f3E22F983, %f6559; fma.rn.ftz.f32 %f30905, %f25491, 0f3E22F983, %f6560; mul.ftz.f32 %f25492, %f30668, %f6354; fma.rn.ftz.f32 %f30902, %f25492, 0f3E22F983, %f6561; cos.approx.ftz.f32 %f25493, %f6562; mul.ftz.f32 %f25494, %f24776, %f30638; mul.ftz.f32 %f25495, %f25494, %f25493; mul.ftz.f32 %f25496, %f24777, %f30638; mul.ftz.f32 %f25497, %f25496, %f25493; mul.ftz.f32 %f25498, %f24778, %f30638; mul.ftz.f32 %f25499, %f25498, %f25493; fma.rn.ftz.f32 %f25500, %f30625, %f30637, %f25495; fma.rn.ftz.f32 %f25501, %f30626, %f30637, %f25497; fma.rn.ftz.f32 %f25502, %f30627, %f30637, %f25499; sin.approx.ftz.f32 %f25503, %f6562; mul.ftz.f32 %f25504, %f30628, %f30638; mul.ftz.f32 %f25505, %f30629, %f30638; mul.ftz.f32 %f25506, %f30630, %f30638; fma.rn.ftz.f32 %f30899, %f25504, %f25503, %f25500; fma.rn.ftz.f32 %f30900, %f25505, %f25503, %f25501; fma.rn.ftz.f32 %f30901, %f25506, %f25503, %f25502; setp.eq.s32 %p3509, %r5870, 0; selp.b32 %r5877, 10, 9, %p3509; bra.uni $L__BB107_4433; $L__BB107_3667: ld.local.v2.f32 {%f30364, %f30365}, [%rd5+-64]; ld.local.f32 %f30366, [%rd5+-56]; $L__BB107_3669: mul.ftz.f32 %f23251, %f30147, %f30365; fma.rn.ftz.f32 %f23252, %f30148, %f30364, %f23251; fma.rn.ftz.f32 %f5449, %f30146, %f30366, %f23252; add.ftz.f32 %f23253, %f5439, %f5449; setp.lt.ftz.f32 %p3125, %f5439, %f5438; selp.f32 %f23254, 0f40000000, 0f3F800000, %p3125; mov.f32 %f23255, 0f3F800000; selp.f32 %f23256, %f23253, %f5449, %p3125; div.approx.ftz.f32 %f23257, %f23256, %f5438; sub.ftz.f32 %f23258, %f23254, %f23257; mov.f32 %f23259, 0f00000000; max.ftz.f32 %f23260, %f23258, %f23259; min.ftz.f32 %f5450, %f23260, %f23255; setp.leu.ftz.f32 %p3126, %f5450, 0f00000000; @%p3126 bra $L__BB107_3707; setp.eq.s32 %p3127, %r1340, 65; ld.local.s32 %rd637, [%rd5+-28]; @%p3127 bra $L__BB107_3672; bra.uni $L__BB107_3671; $L__BB107_3672: ld.global.u32 %r1341, [%rd636+-28]; ld.global.u32 %r1342, [%rd636+-32]; shl.b32 %r4655, %r1342, 1; cvt.rn.f32.s32 %f23261, %r4655; ld.local.f32 %f23262, [%rd5+-4]; mul.ftz.f32 %f23263, %f23262, %f23261; cvt.rzi.ftz.s32.f32 %r4656, %f23263; add.s32 %r4657, %r4655, -1; min.s32 %r1343, %r4656, %r4657; cvt.rn.f32.s32 %f23264, %r1343; sub.ftz.f32 %f5469, %f23263, %f23264; ld.const.u64 %rd3679, [kernel_params+2312]; cvta.to.global.u64 %rd638, %rd3679; ld.global.u32 %r5814, [%rd636+-20]; mul.wide.u32 %rd3680, %r5814, 16; add.s64 %rd3681, %rd638, %rd3680; ld.global.u16 %rs729, [%rd3681+12]; ld.global.u32 %r5813, [%rd3681+8]; ld.global.u64 %rd4854, [%rd3681]; setp.eq.s64 %p3128, %rd4854, 11; mov.u16 %rs727, %rs729; mov.u32 %r5808, %r5813; @%p3128 bra $L__BB107_3678; mov.u64 %rd4835, %rd4854; mov.u32 %r5808, %r5813; mov.u16 %rs727, %rs729; mov.u32 %r5809, %r5814; $L__BB107_3674: setp.eq.s64 %p3129, %rd4835, 0; @%p3129 bra $L__BB107_3676; add.s32 %r5809, %r5809, 2; bra.uni $L__BB107_3677; $L__BB107_3676: setp.eq.s16 %p3130, %rs727, 0; mov.u32 %r5812, -1; mov.u32 %r5809, %r5808; @%p3130 bra $L__BB107_3679; $L__BB107_3677: mul.wide.u32 %rd3682, %r5809, 16; add.s64 %rd3683, %rd638, %rd3682; ld.global.u16 %rs727, [%rd3683+12]; ld.global.u32 %r5808, [%rd3683+8]; ld.global.u64 %rd4835, [%rd3683]; setp.ne.s64 %p3131, %rd4835, 11; @%p3131 bra $L__BB107_3674; $L__BB107_3678: setp.eq.s16 %p3132, %rs727, 0; selp.b32 %r5812, -1, %r5808, %p3132; $L__BB107_3679: ld.const.u64 %rd3684, [kernel_params+2232]; cvta.to.global.u64 %rd3685, %rd3684; shl.b64 %rd3686, %rd637, 4; add.s64 %rd3687, %rd3685, %rd3686; ld.global.v4.u32 {%r4659, %r4660, %r4661, %r4662}, [%rd3687]; setp.eq.s32 %p3133, %r1343, %r1342; @%p3133 bra $L__BB107_3681; bra.uni $L__BB107_3680; $L__BB107_3681: ld.const.u64 %rd3693, [kernel_params+2256]; cvta.to.global.u64 %rd3694, %rd3693; mul.wide.u32 %rd3695, %r4662, 12; add.s64 %rd4836, %rd3694, %rd3695; add.s32 %r4669, %r4662, 1; mul.wide.u32 %rd3696, %r4669, 12; add.s64 %rd4839, %rd3694, %rd3696; add.s32 %r4670, %r4662, 2; mul.wide.u32 %rd3697, %r4670, 12; add.s64 %rd4842, %rd3694, %rd3697; add.s64 %rd4837, %rd4836, 4; add.s64 %rd4838, %rd4836, 8; add.s64 %rd4840, %rd4839, 4; add.s64 %rd4841, %rd4839, 8; add.s64 %rd4843, %rd4842, 4; add.s64 %rd4844, %rd4842, 8; bra.uni $L__BB107_3682; $L__BB107_4239: ld.local.f32 %f26403, [%rd1027+52]; mul.ftz.f32 %f26404, %f6913, %f26403; ld.local.f32 %f26405, [%rd1027+48]; mul.ftz.f32 %f26406, %f6914, %f26405; sub.ftz.f32 %f26407, %f26404, %f26406; ld.local.f32 %f26408, [%rd1027+44]; mul.ftz.f32 %f26409, %f6914, %f26408; mul.ftz.f32 %f26410, %f6911, %f26403; sub.ftz.f32 %f26411, %f26409, %f26410; mul.ftz.f32 %f26412, %f6911, %f26405; mul.ftz.f32 %f26413, %f6913, %f26408; sub.ftz.f32 %f26414, %f26412, %f26413; mul.ftz.f32 %f26415, %f26411, %f26411; fma.rn.ftz.f32 %f26416, %f26407, %f26407, %f26415; fma.rn.ftz.f32 %f26417, %f26414, %f26414, %f26416; sqrt.approx.ftz.f32 %f26418, %f26417; div.approx.ftz.f32 %f30731, %f26407, %f26418; div.approx.ftz.f32 %f30732, %f26411, %f26418; div.approx.ftz.f32 %f30733, %f26414, %f26418; mul.ftz.f32 %f26419, %f6914, %f30732; mul.ftz.f32 %f26420, %f6913, %f30733; sub.ftz.f32 %f30734, %f26419, %f26420; mul.ftz.f32 %f26421, %f6911, %f30733; mul.ftz.f32 %f26422, %f6914, %f30731; sub.ftz.f32 %f30735, %f26421, %f26422; mul.ftz.f32 %f26423, %f6913, %f30731; mul.ftz.f32 %f26424, %f6911, %f30732; sub.ftz.f32 %f30736, %f26423, %f26424; $L__BB107_4241: add.ftz.f32 %f6938, %f6918, 0f3F800000; @%p3617 bra $L__BB107_4257; bra.uni $L__BB107_4242; $L__BB107_4257: mul.ftz.f32 %f30742, %f6226, 0f40C90FDB; lg2.approx.ftz.f32 %f26590, %f6202; rcp.approx.ftz.f32 %f26591, %f6938; mul.ftz.f32 %f26592, %f26591, %f26590; ex2.approx.ftz.f32 %f30741, %f26592; bra.uni $L__BB107_4258; $L__BB107_4270: setp.neu.ftz.f32 %p3645, %f7000, %f7002; setp.neu.ftz.f32 %p3646, %f7000, %f7003; or.pred %p3647, %p3645, %p3646; neg.ftz.f32 %f26726, %f7003; selp.f32 %f26727, %f26726, %f7003, %p3647; add.ftz.f32 %f26728, %f7000, %f26727; neg.ftz.f32 %f26729, %f7002; selp.f32 %f26730, %f7002, %f26729, %p3647; sub.ftz.f32 %f26731, %f26730, %f7000; sub.ftz.f32 %f26732, %f7003, %f7002; mul.ftz.f32 %f26733, %f26728, %f26728; fma.rn.ftz.f32 %f26734, %f26732, %f26732, %f26733; fma.rn.ftz.f32 %f26735, %f26731, %f26731, %f26734; sqrt.approx.ftz.f32 %f26736, %f26735; div.approx.ftz.f32 %f30751, %f26732, %f26736; div.approx.ftz.f32 %f30752, %f26728, %f26736; div.approx.ftz.f32 %f30753, %f26731, %f26736; mul.ftz.f32 %f26737, %f7002, %f30753; mul.ftz.f32 %f26738, %f7003, %f30752; sub.ftz.f32 %f30748, %f26737, %f26738; mul.ftz.f32 %f26739, %f7003, %f30751; mul.ftz.f32 %f26740, %f7000, %f30753; sub.ftz.f32 %f30749, %f26739, %f26740; mul.ftz.f32 %f26741, %f7000, %f30752; mul.ftz.f32 %f26742, %f7002, %f30751; sub.ftz.f32 %f30750, %f26741, %f26742; $L__BB107_4272: mul.ftz.f32 %f26766, %f26719, %f30752; fma.rn.ftz.f32 %f26767, %f7001, %f30751, %f26766; fma.rn.ftz.f32 %f7028, %f26720, %f30753, %f26767; mul.ftz.f32 %f26768, %f26719, %f30749; fma.rn.ftz.f32 %f26769, %f7001, %f30748, %f26768; fma.rn.ftz.f32 %f7029, %f26720, %f30750, %f26769; ld.local.u64 %rd4468, [%rd1027+36]; ld.f32 %f7030, [%rd4468]; ld.f32 %f7031, [%rd4468+4]; ld.f32 %f7032, [%rd4468+8]; ld.local.f32 %f7033, [%rd1027+28]; ld.f32 %f7034, [%rd4468+12]; ld.f32 %f7035, [%rd4468+16]; ld.f32 %f7036, [%rd4468+20]; neg.ftz.f32 %f30761, %f7028; neg.ftz.f32 %f30760, %f7029; setp.lt.ftz.f32 %p3648, %f7004, 0fBF7FF972; mov.f32 %f30774, 0f00000000; @%p3648 bra $L__BB107_4275; setp.gt.ftz.f32 %p3649, %f7004, 0f3F7FF972; mov.f32 %f30774, 0fBF7FF972; @%p3649 bra $L__BB107_4275; mul.ftz.f32 %f26771, %f7004, %f7004; mov.f32 %f26772, 0f33D6BF95; max.ftz.f32 %f26773, %f26771, %f26772; rcp.approx.ftz.f32 %f26774, %f26773; mul.ftz.f32 %f26775, %f26723, %f30761; mul.ftz.f32 %f26776, %f26724, %f30760; mul.ftz.f32 %f26777, %f26776, %f26776; fma.rn.ftz.f32 %f26778, %f26775, %f26775, %f26777; fma.rn.ftz.f32 %f26779, %f26778, %f26774, 0f3F800000; sqrt.approx.ftz.f32 %f26780, %f26779; neg.ftz.f32 %f26781, %f26780; setp.ge.ftz.f32 %p3650, %f7004, 0f80000000; selp.f32 %f26782, %f26781, %f26780, %p3650; add.ftz.f32 %f26783, %f26782, 0fBF800000; mul.ftz.f32 %f30774, %f26783, 0f3F000000; $L__BB107_4275: fma.rn.ftz.f32 %f7041, %f7033, %f7033, 0fBF800000; mov.f32 %f30905, 0f3F800000; abs.ftz.f32 %f7042, %f30905; fma.rn.ftz.f32 %f7043, %f7042, %f7042, %f7041; setp.leu.ftz.f32 %p3651, %f7043, 0f00000000; mov.f32 %f30755, %f30905; @%p3651 bra $L__BB107_4277; sqrt.approx.ftz.f32 %f26785, %f7043; sub.ftz.f32 %f26786, %f26785, %f7042; add.ftz.f32 %f26787, %f7042, %f26785; div.approx.ftz.f32 %f26788, %f26786, %f26787; fma.rn.ftz.f32 %f26789, %f7042, %f26787, 0fBF800000; fma.rn.ftz.f32 %f26790, %f7042, %f26786, 0f3F800000; div.approx.ftz.f32 %f26791, %f26789, %f26790; mul.ftz.f32 %f26792, %f26788, 0f3F000000; mul.ftz.f32 %f26793, %f26788, %f26792; fma.rn.ftz.f32 %f26794, %f26791, %f26791, 0f3F800000; mul.ftz.f32 %f30755, %f26793, %f26794; $L__BB107_4277: setp.ne.s32 %p22, %r1522, 15; sub.ftz.f32 %f7046, %f30905, %f30755; rcp.approx.ftz.f32 %f7047, %f7046; ld.local.u32 %r5873, [%rd5+168]; neg.ftz.f32 %f30778, %f7004; mov.f32 %f30775, 0f00000000; mov.u32 %r5874, 0; mov.pred %p3652, -1; mov.pred %p3662, 0; mov.f32 %f30904, %f30905; mov.f32 %f30903, %f30905; mov.f32 %f30765, %f30905; $L__BB107_4278: mov.f32 %f7056, %f30765; mad.lo.s32 %r1544, %r5873, 1103515245, 12345; cvt.rn.f32.u32 %f26800, %r1544; mul.ftz.f32 %f7058, %f26800, 0f2F800000; setp.gt.ftz.f32 %p3653, %f30778, 0f3F7FF972; mov.pred %p3932, %p3652; @%p3653 bra $L__BB107_4285; setp.lt.ftz.f32 %p3654, %f30778, 0fBF7FF972; @%p3654 bra $L__BB107_4284; bra.uni $L__BB107_4280; $L__BB107_4284: mul.ftz.f32 %f30765, %f7056, %f7058; mov.pred %p3932, %p3662; bra.uni $L__BB107_4285; $L__BB107_4280: abs.ftz.f32 %f26801, %f30778; setp.ltu.ftz.f32 %p3656, %f26801, 0f38D1B717; mov.f32 %f30765, %f7056; mov.pred %p3932, %p3662; @%p3656 bra $L__BB107_4285; mov.f32 %f26802, 0f3F800000; sub.ftz.f32 %f26803, %f26802, %f30775; setp.lt.ftz.f32 %p3658, %f26803, %f7058; mov.f32 %f30765, %f7056; mov.pred %p3932, %p3652; @%p3658 bra $L__BB107_4285; setp.ge.ftz.f32 %p3660, %f30774, 0f00000000; mov.f32 %f30765, %f26802; mov.pred %p3932, %p3662; @%p3660 bra $L__BB107_4285; mov.f32 %f26805, 0f3F800000; sub.ftz.f32 %f26806, %f26805, %f7058; lg2.approx.ftz.f32 %f26807, %f26806; mov.f32 %f26808, 0fBF800000; div.approx.ftz.f32 %f26809, %f26808, %f30774; mul.ftz.f32 %f26810, %f26809, %f26807; ex2.approx.ftz.f32 %f26811, %f26810; mul.ftz.f32 %f30765, %f7056, %f26811; mov.pred %p3932, %p3662; $L__BB107_4285: @%p3932 bra $L__BB107_4308; mad.lo.s32 %r5319, %r1544, 1103515245, 12345; cvt.rn.f32.u32 %f26812, %r5319; mul.ftz.f32 %f7062, %f26812, 0f2F800000; mad.lo.s32 %r5873, %r5319, 1103515245, 12345; cvt.rn.f32.u32 %f26813, %r5873; mul.ftz.f32 %f7063, %f26813, 0f2F800000; neg.ftz.f32 %f7064, %f30761; mul.ftz.f32 %f26814, %f26723, %f7064; neg.ftz.f32 %f7065, %f30760; mul.ftz.f32 %f26815, %f26724, %f7065; mul.ftz.f32 %f26816, %f26815, %f26815; fma.rn.ftz.f32 %f26817, %f26814, %f26814, %f26816; fma.rn.ftz.f32 %f26818, %f30778, %f30778, %f26817; sqrt.approx.ftz.f32 %f26819, %f26818; div.approx.ftz.f32 %f30768, %f26814, %f26819; div.approx.ftz.f32 %f30769, %f26815, %f26819; neg.ftz.f32 %f7068, %f30778; div.approx.ftz.f32 %f7069, %f7068, %f26819; setp.gt.ftz.f32 %p3663, %f7069, 0f3F7FF972; @%p3663 bra $L__BB107_4292; abs.ftz.f32 %f26820, %f7069; setp.lt.ftz.f32 %p3664, %f26820, 0f358637BD; @%p3664 bra $L__BB107_4292; bra.uni $L__BB107_4288; $L__BB107_4292: mov.f32 %f26864, 0f3F800000; sub.ftz.f32 %f26865, %f26864, %f7063; mov.f32 %f26866, 0f33D6BF95; max.ftz.f32 %f26867, %f26865, %f26866; div.approx.ftz.f32 %f26868, %f7063, %f26867; sqrt.approx.ftz.f32 %f26869, %f26868; mul.ftz.f32 %f26870, %f7062, 0f40C90FDB; cos.approx.ftz.f32 %f26871, %f26870; sin.approx.ftz.f32 %f26872, %f26870; mul.ftz.f32 %f30767, %f26869, %f26872; mul.ftz.f32 %f30766, %f26869, %f26871; $L__BB107_4293: mul.ftz.f32 %f26873, %f30769, %f30769; fma.rn.ftz.f32 %f26874, %f30768, %f30768, %f26873; add.ftz.f32 %f26875, %f26874, 0f00000000; sqrt.approx.ftz.f32 %f7087, %f26875; setp.neu.ftz.f32 %p3672, %f7087, 0f00000000; @%p3672 bra $L__BB107_4294; bra.uni $L__BB107_4295; $L__BB107_4294: rcp.approx.ftz.f32 %f26876, %f7087; mul.ftz.f32 %f30769, %f30769, %f26876; mul.ftz.f32 %f30768, %f30768, %f26876; $L__BB107_4295: mul.ftz.f32 %f26877, %f30766, %f30768; mul.ftz.f32 %f26878, %f30767, %f30769; sub.ftz.f32 %f26879, %f26877, %f26878; mul.ftz.f32 %f26880, %f26723, %f26879; mul.ftz.f32 %f26881, %f30767, %f30768; fma.rn.ftz.f32 %f26882, %f30766, %f30769, %f26881; mul.ftz.f32 %f26883, %f26724, %f26882; neg.ftz.f32 %f26884, %f26880; neg.ftz.f32 %f26885, %f26883; mul.ftz.f32 %f26886, %f26883, %f26883; fma.rn.ftz.f32 %f26887, %f26880, %f26880, %f26886; add.ftz.f32 %f26888, %f26887, 0f3F800000; mov.f32 %f30775, 0f3F800000; sqrt.approx.ftz.f32 %f26890, %f26888; div.approx.ftz.f32 %f7097, %f26884, %f26890; div.approx.ftz.f32 %f7098, %f26885, %f26890; div.approx.ftz.f32 %f7099, %f30775, %f26890; setp.ne.s32 %p3673, %r5874, 0; and.pred %p3674, %p3673, %p22; mul.ftz.f32 %f26891, %f7030, %f30903; selp.f32 %f7100, %f26891, %f30903, %p3674; mul.ftz.f32 %f26892, %f7031, %f30904; selp.f32 %f7101, %f26892, %f30904, %p3674; mul.ftz.f32 %f26893, %f7032, %f30905; selp.f32 %f7102, %f26893, %f30905, %p3674; mul.ftz.f32 %f26894, %f7098, %f7065; fma.rn.ftz.f32 %f26895, %f7097, %f7064, %f26894; fma.rn.ftz.f32 %f7103, %f7099, %f7068, %f26895; mov.f32 %f30903, %f7100; mov.f32 %f30904, %f7101; mov.f32 %f30905, %f7102; @%p22 bra $L__BB107_4300; abs.ftz.f32 %f7104, %f7103; fma.rn.ftz.f32 %f7105, %f7104, %f7104, %f7041; setp.leu.ftz.f32 %p3676, %f7105, 0f00000000; mov.f32 %f30770, %f7046; @%p3676 bra $L__BB107_4298; sqrt.approx.ftz.f32 %f26896, %f7105; sub.ftz.f32 %f26897, %f26896, %f7104; add.ftz.f32 %f26898, %f7104, %f26896; div.approx.ftz.f32 %f26899, %f26897, %f26898; fma.rn.ftz.f32 %f26900, %f7104, %f26898, 0fBF800000; fma.rn.ftz.f32 %f26901, %f7104, %f26897, 0f3F800000; div.approx.ftz.f32 %f26902, %f26900, %f26901; mul.ftz.f32 %f26903, %f26899, 0f3F000000; mul.ftz.f32 %f26904, %f26899, %f26903; fma.rn.ftz.f32 %f26905, %f26902, %f26902, 0f3F800000; mul.ftz.f32 %f26906, %f26904, %f26905; sub.ftz.f32 %f30770, %f26906, %f30755; $L__BB107_4298: mul.ftz.f32 %f26907, %f7047, %f30770; mov.f32 %f26908, 0f3F800000; sub.ftz.f32 %f26909, %f26908, %f26907; fma.rn.ftz.f32 %f30903, %f7034, %f26909, %f26907; fma.rn.ftz.f32 %f30904, %f7035, %f26909, %f26907; fma.rn.ftz.f32 %f30905, %f7036, %f26909, %f26907; setp.eq.s32 %p3677, %r5874, 0; @%p3677 bra $L__BB107_4300; mul.ftz.f32 %f30903, %f7100, %f30903; mul.ftz.f32 %f30904, %f7101, %f30904; mul.ftz.f32 %f30905, %f7102, %f30905; $L__BB107_4300: add.ftz.f32 %f26911, %f7097, %f7097; add.ftz.f32 %f26912, %f7098, %f7098; add.ftz.f32 %f26913, %f7099, %f7099; fma.rn.ftz.f32 %f30761, %f26911, %f7103, %f30761; fma.rn.ftz.f32 %f30760, %f26912, %f7103, %f30760; fma.rn.ftz.f32 %f30778, %f26913, %f7103, %f30778; setp.gt.ftz.f32 %p3678, %f30778, 0f3F7FF972; mov.f32 %f30774, 0f00000000; @%p3678 bra $L__BB107_4303; setp.lt.ftz.f32 %p3679, %f30778, 0fBF7FF972; mov.f32 %f30774, 0fBF7FF972; @%p3679 bra $L__BB107_4303; mul.ftz.f32 %f26915, %f30778, %f30778; mov.f32 %f26916, 0f33D6BF95; max.ftz.f32 %f26917, %f26915, %f26916; rcp.approx.ftz.f32 %f26918, %f26917; mul.ftz.f32 %f26919, %f26723, %f30761; mul.ftz.f32 %f26920, %f26724, %f30760; mul.ftz.f32 %f26921, %f26920, %f26920; fma.rn.ftz.f32 %f26922, %f26919, %f26919, %f26921; fma.rn.ftz.f32 %f26923, %f26922, %f26918, 0f3F800000; sqrt.approx.ftz.f32 %f26924, %f26923; neg.ftz.f32 %f26925, %f26924; setp.le.ftz.f32 %p3680, %f30778, 0f00000000; selp.f32 %f26926, %f26925, %f26924, %p3680; add.ftz.f32 %f26927, %f26926, 0fBF800000; mul.ftz.f32 %f30774, %f26927, 0f3F000000; $L__BB107_4303: @%p3678 bra $L__BB107_4306; setp.lt.ftz.f32 %p3682, %f30778, 0f3727C5AC; mov.f32 %f30775, 0f00000000; @%p3682 bra $L__BB107_4306; lg2.approx.ftz.f32 %f26930, %f30765; mul.ftz.f32 %f26931, %f30774, %f26930; ex2.approx.ftz.f32 %f30775, %f26931; $L__BB107_4306: add.s32 %r5874, %r5874, 1; setp.lt.u32 %p3683, %r5874, 10; @%p3683 bra $L__BB107_4278; bra.uni $L__BB107_4307; $L__BB107_4288: add.ftz.f32 %f26823, %f7069, 0f3F800000; mul.ftz.f32 %f7070, %f26823, 0f3F000000; setp.lt.ftz.f32 %p3665, %f7070, 0f38D1B717; mov.f32 %f30766, 0f00000000; mov.f32 %f30767, %f30766; @%p3665 bra $L__BB107_4293; add.ftz.f32 %f26826, %f7063, %f7063; mul.ftz.f32 %f26827, %f26826, %f7070; div.approx.ftz.f32 %f26828, %f26827, %f7069; add.ftz.f32 %f7071, %f26828, 0fBF800000; mul.ftz.f32 %f7072, %f7071, %f7071; add.ftz.f32 %f7073, %f7072, 0fBF800000; abs.ftz.f32 %f26829, %f7073; setp.lt.ftz.f32 %p3666, %f26829, 0f33D6BF95; mov.f32 %f30767, %f30766; @%p3666 bra $L__BB107_4293; mul.ftz.f32 %f26830, %f7069, %f7069; mov.f32 %f26831, 0f3F800000; sub.ftz.f32 %f26832, %f26831, %f26830; mov.f32 %f26833, 0f00000000; max.ftz.f32 %f26834, %f26832, %f26833; sqrt.approx.ftz.f32 %f26835, %f26834; div.approx.ftz.f32 %f26836, %f26835, %f7069; rcp.approx.ftz.f32 %f26837, %f7073; mul.ftz.f32 %f26838, %f26836, %f26836; mul.ftz.f32 %f26839, %f26838, %f26837; mul.ftz.f32 %f26840, %f26837, %f26839; sub.ftz.f32 %f26841, %f7072, %f26838; mul.ftz.f32 %f26842, %f26841, %f26837; sub.ftz.f32 %f26843, %f26840, %f26842; max.ftz.f32 %f26844, %f26843, %f26833; sqrt.approx.ftz.f32 %f26845, %f26844; mul.ftz.f32 %f26846, %f26836, %f26837; add.ftz.f32 %f26847, %f26846, %f26845; rcp.approx.ftz.f32 %f26848, %f26836; setp.gt.ftz.f32 %p3667, %f26847, %f26848; setp.lt.ftz.f32 %p3668, %f7071, 0f00000000; or.pred %p3669, %p3668, %p3667; sub.ftz.f32 %f26849, %f26846, %f26845; mov.f32 %f26850, 0f3F000000; sub.ftz.f32 %f26851, %f26850, %f7062; add.ftz.f32 %f26852, %f7062, 0fBF000000; setp.ge.ftz.f32 %p3670, %f7062, 0f3F000000; selp.f32 %f26853, %f26852, %f26851, %p3670; add.ftz.f32 %f26854, %f26853, %f26853; fma.rn.ftz.f32 %f26855, %f26854, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f26856, %f26854, %f26855, 0f3EED4413; mul.ftz.f32 %f26857, %f26854, %f26856; fma.rn.ftz.f32 %f26858, %f26854, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f26859, %f26854, %f26858, 0fBF800000; fma.rn.ftz.f32 %f26860, %f26854, %f26859, 0f3F191676; div.approx.ftz.f32 %f26861, %f26857, %f26860; selp.f32 %f30766, %f26849, %f26847, %p3669; fma.rn.ftz.f32 %f26862, %f30766, %f30766, 0f3F800000; sqrt.approx.ftz.f32 %f26863, %f26862; mul.ftz.f32 %f30767, %f26861, %f26863; setp.ltu.ftz.f32 %p3671, %f7062, 0f3F000000; @%p3671 bra $L__BB107_4291; bra.uni $L__BB107_4293; $L__BB107_4291: neg.ftz.f32 %f30767, %f30767; bra.uni $L__BB107_4293; $L__BB107_4308: st.local.u32 [%rd5+168], %r1544; bra.uni $L__BB107_4309; $L__BB107_4307: st.local.u32 [%rd5+168], %r5873; mov.f32 %f30761, 0f00000000; mov.f32 %f30778, 0f3F800000; mov.f32 %f30760, %f30761; mov.f32 %f30903, %f30761; mov.f32 %f30904, %f30761; mov.f32 %f30905, %f30761; $L__BB107_4309: mul.ftz.f32 %f26939, %f30748, %f30760; fma.rn.ftz.f32 %f26940, %f30751, %f30761, %f26939; mul.ftz.f32 %f26941, %f30749, %f30760; fma.rn.ftz.f32 %f26942, %f30752, %f30761, %f26941; mul.ftz.f32 %f26943, %f30750, %f30760; fma.rn.ftz.f32 %f26944, %f30753, %f30761, %f26943; fma.rn.ftz.f32 %f30899, %f7000, %f30778, %f26940; fma.rn.ftz.f32 %f30900, %f7002, %f30778, %f26942; fma.rn.ftz.f32 %f30901, %f7003, %f30778, %f26944; mul.ftz.f32 %f26945, %f6234, %f30900; fma.rn.ftz.f32 %f26946, %f6233, %f30899, %f26945; fma.rn.ftz.f32 %f26947, %f6235, %f30901, %f26946; setp.le.ftz.f32 %p3684, %f26947, 0f00000000; mov.u32 %r5877, 0; mov.f32 %f30902, 0f00000000; @%p3684 bra $L__BB107_4433; ld.local.f32 %f7139, [%rd1027+20]; add.ftz.f32 %f26948, %f7028, %f30761; add.ftz.f32 %f26949, %f7029, %f30760; mul.ftz.f32 %f26950, %f26949, %f26949; fma.rn.ftz.f32 %f26951, %f26948, %f26948, %f26950; add.ftz.f32 %f26952, %f7004, %f30778; fma.rn.ftz.f32 %f26953, %f26952, %f26952, %f26951; sqrt.approx.ftz.f32 %f26954, %f26953; div.approx.ftz.f32 %f7140, %f26948, %f26954; div.approx.ftz.f32 %f7141, %f26949, %f26954; div.approx.ftz.f32 %f26955, %f26952, %f26954; mul.ftz.f32 %f7142, %f26955, %f26955; @%p3644 bra $L__BB107_4315; bra.uni $L__BB107_4311; $L__BB107_4315: setp.gt.ftz.f32 %p3688, %f7004, 0f3F7FF972; neg.ftz.f32 %f27002, %f7140; div.approx.ftz.f32 %f27003, %f27002, %f7139; neg.ftz.f32 %f27004, %f7141; div.approx.ftz.f32 %f27005, %f27004, %f26724; fma.rn.ftz.f32 %f27006, %f27003, %f27003, %f7142; fma.rn.ftz.f32 %f7147, %f27005, %f27005, %f27006; mov.f32 %f30783, 0f00000000; @%p3688 bra $L__BB107_4318; mov.f32 %f30783, 0fBF7FF972; @%p3648 bra $L__BB107_4318; mul.ftz.f32 %f27008, %f7004, %f7004; mov.f32 %f27009, 0f33D6BF95; max.ftz.f32 %f27010, %f27008, %f27009; rcp.approx.ftz.f32 %f27011, %f27010; mul.ftz.f32 %f27012, %f7028, %f7139; mul.ftz.f32 %f27013, %f7029, %f26724; mul.ftz.f32 %f27014, %f27013, %f27013; fma.rn.ftz.f32 %f27015, %f27012, %f27012, %f27014; fma.rn.ftz.f32 %f27016, %f27015, %f27011, 0f3F800000; sqrt.approx.ftz.f32 %f27017, %f27016; add.ftz.f32 %f27018, %f27017, 0fBF800000; mul.ftz.f32 %f30783, %f27018, 0f3F000000; $L__BB107_4318: add.ftz.f32 %f27019, %f30783, 0f3F800000; mov.f32 %f27020, 0f3F800000; mul.ftz.f32 %f27021, %f7004, %f27019; mov.f32 %f27022, 0f33D6BF95; max.ftz.f32 %f27023, %f27021, %f27022; mul.ftz.f32 %f27024, %f7147, 0f40490FDB; mul.ftz.f32 %f27025, %f7147, %f27024; mul.ftz.f32 %f27026, %f7139, %f27025; mul.ftz.f32 %f27027, %f26724, %f27026; max.ftz.f32 %f27028, %f27027, %f27022; rcp.approx.ftz.f32 %f27029, %f27028; mul.ftz.f32 %f27030, %f27029, 0f3E800000; div.approx.ftz.f32 %f27031, %f27030, %f27023; mul.ftz.f32 %f27032, %f7139, %f26724; sqrt.approx.ftz.f32 %f27033, %f27032; mul.ftz.f32 %f27034, %f27033, 0fBFFE59F3; mul.ftz.f32 %f27035, %f27033, %f27034; mul.ftz.f32 %f27036, %f27035, 0f3FB8AA3B; ex2.approx.ftz.f32 %f27037, %f27036; fma.rn.ftz.f32 %f27038, %f27037, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f27039, %f27033, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f27040, %f27033, %f27039, 0f417C8B44; fma.rn.ftz.f32 %f27041, %f27033, %f27040, 0fC1665E35; fma.rn.ftz.f32 %f27042, %f27033, %f27041, 0f40CE7881; fma.rn.ftz.f32 %f27043, %f27033, %f27042, 0fBF993E81; fma.rn.ftz.f32 %f27044, %f27033, %f27043, 0f3CE3C31E; fma.rn.ftz.f32 %f27045, %f27033, %f27044, 0f3BBA5D49; sub.ftz.f32 %f27046, %f27038, %f27045; cvt.ftz.sat.f32.f32 %f27047, %f27046; sub.ftz.f32 %f27048, %f27020, %f27047; mul.ftz.f32 %f27049, %f30778, 0f3EA2F983; mul.ftz.f32 %f27050, %f27049, %f27048; fma.rn.ftz.f32 %f30784, %f27031, %f27047, %f27050; bra.uni $L__BB107_4319; $L__BB107_4242: setp.lt.ftz.f32 %p3622, %f6226, 0f3E800000; add.ftz.f32 %f26442, %f6919, 0f3F800000; div.approx.ftz.f32 %f26443, %f6938, %f26442; sqrt.approx.ftz.f32 %f6939, %f26443; @%p3622 bra $L__BB107_4254; bra.uni $L__BB107_4243; $L__BB107_4254: mul.ftz.f32 %f26556, %f6226, 0f40800000; mul.ftz.f32 %f26557, %f26556, 0f3FC90FDB; sin.approx.ftz.f32 %f26558, %f26557; cos.approx.ftz.f32 %f26559, %f26557; div.approx.ftz.f32 %f26560, %f26558, %f26559; mul.ftz.f32 %f6958, %f6939, %f26560; abs.ftz.f32 %f6959, %f6958; setp.leu.ftz.f32 %p3634, %f6959, 0f3F800000; mov.f32 %f30740, %f6959; @%p3634 bra $L__BB107_4256; rcp.approx.ftz.f32 %f30740, %f6959; $L__BB107_4256: mul.rn.ftz.f32 %f26561, %f30740, %f30740; mov.f32 %f26562, 0fC0B59883; mov.f32 %f26563, 0fBF52C7EA; fma.rn.ftz.f32 %f26564, %f26561, %f26563, %f26562; mov.f32 %f26565, 0fC0D21907; fma.rn.ftz.f32 %f26566, %f26564, %f26561, %f26565; mul.ftz.f32 %f26567, %f26561, %f26566; mul.ftz.f32 %f26568, %f30740, %f26567; add.ftz.f32 %f26569, %f26561, 0f41355DC0; mov.f32 %f26570, 0f41E6BD60; fma.rn.ftz.f32 %f26571, %f26569, %f26561, %f26570; mov.f32 %f26572, 0f419D92C8; fma.rn.ftz.f32 %f26573, %f26571, %f26561, %f26572; rcp.approx.ftz.f32 %f26574, %f26573; fma.rn.ftz.f32 %f26575, %f26568, %f26574, %f30740; mov.f32 %f26576, 0f3FC90FDB; sub.ftz.f32 %f26577, %f26576, %f26575; setp.gt.ftz.f32 %p3635, %f6959, 0f3F800000; selp.f32 %f26578, %f26577, %f26575, %p3635; mov.b32 %r5311, %f26578; mov.b32 %r5312, %f6958; and.b32 %r5313, %r5312, -2147483648; or.b32 %r5314, %r5313, %r5311; mov.b32 %f26579, %r5314; setp.le.ftz.f32 %p3636, %f6959, 0f7F800000; selp.f32 %f30742, %f26579, %f26578, %p3636; cos.approx.ftz.f32 %f26580, %f30742; sin.approx.ftz.f32 %f26581, %f30742; mul.ftz.f32 %f26582, %f6918, %f26580; mul.ftz.f32 %f26583, %f6919, %f26581; mul.ftz.f32 %f26584, %f26581, %f26583; fma.rn.ftz.f32 %f26585, %f26580, %f26582, %f26584; add.ftz.f32 %f26586, %f26585, 0f3F800000; rcp.approx.ftz.f32 %f26587, %f26586; lg2.approx.ftz.f32 %f26588, %f6202; mul.ftz.f32 %f26589, %f26588, %f26587; ex2.approx.ftz.f32 %f30741, %f26589; bra.uni $L__BB107_4258; $L__BB107_4384: add.s32 %r5360, %r1522, -11; setp.gt.u32 %p3741, %r5360, 1; mov.u32 %r5877, 18; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3741 bra $L__BB107_4405; ld.local.f32 %f27783, [%rd1027+28]; fma.rn.ftz.f32 %f7376, %f27783, %f27783, 0fBF800000; abs.ftz.f32 %f7377, %f27750; fma.rn.ftz.f32 %f7378, %f7377, %f7377, %f7376; setp.leu.ftz.f32 %p3742, %f7378, 0f00000000; mov.f32 %f30865, %f27750; @%p3742 bra $L__BB107_4387; sqrt.approx.ftz.f32 %f27784, %f7378; sub.ftz.f32 %f27785, %f27784, %f7377; add.ftz.f32 %f27786, %f7377, %f27784; div.approx.ftz.f32 %f27787, %f27785, %f27786; fma.rn.ftz.f32 %f27788, %f7377, %f27786, 0fBF800000; fma.rn.ftz.f32 %f27789, %f7377, %f27785, 0f3F800000; div.approx.ftz.f32 %f27790, %f27788, %f27789; mul.ftz.f32 %f27791, %f27787, 0f3F000000; mul.ftz.f32 %f27792, %f27787, %f27791; fma.rn.ftz.f32 %f27793, %f27790, %f27790, 0f3F800000; mul.ftz.f32 %f30865, %f27792, %f27793; $L__BB107_4387: ld.local.u64 %rd4469, [%rd1027+36]; ld.f32 %f7381, [%rd4469+12]; ld.f32 %f7382, [%rd4469+16]; ld.f32 %f7383, [%rd4469+20]; sub.ftz.f32 %f7384, %f27750, %f30865; mul.ftz.f32 %f27795, %f7369, %f30900; fma.rn.ftz.f32 %f27796, %f7368, %f30899, %f27795; fma.rn.ftz.f32 %f27797, %f7370, %f30901, %f27796; abs.ftz.f32 %f7385, %f27797; fma.rn.ftz.f32 %f7386, %f7385, %f7385, %f7376; setp.leu.ftz.f32 %p3743, %f7386, 0f00000000; mov.f32 %f30866, %f7384; @%p3743 bra $L__BB107_4389; sqrt.approx.ftz.f32 %f27798, %f7386; sub.ftz.f32 %f27799, %f27798, %f7385; add.ftz.f32 %f27800, %f7385, %f27798; div.approx.ftz.f32 %f27801, %f27799, %f27800; fma.rn.ftz.f32 %f27802, %f7385, %f27800, 0fBF800000; fma.rn.ftz.f32 %f27803, %f7385, %f27799, 0f3F800000; div.approx.ftz.f32 %f27804, %f27802, %f27803; mul.ftz.f32 %f27805, %f27801, 0f3F000000; mul.ftz.f32 %f27806, %f27801, %f27805; fma.rn.ftz.f32 %f27807, %f27804, %f27804, 0f3F800000; mul.ftz.f32 %f27808, %f27806, %f27807; sub.ftz.f32 %f30866, %f27808, %f30865; $L__BB107_4389: rcp.approx.ftz.f32 %f27810, %f7384; mov.f32 %f27811, 0f3F800000; mul.ftz.f32 %f27812, %f27810, %f30866; sub.ftz.f32 %f27813, %f27811, %f27812; fma.rn.ftz.f32 %f27814, %f7383, %f27813, %f27812; fma.rn.ftz.f32 %f27815, %f7382, %f27813, %f27812; fma.rn.ftz.f32 %f27816, %f7381, %f27813, %f27812; mul.ftz.f32 %f30903, %f27816, 0f49742400; mul.ftz.f32 %f30904, %f27815, 0f49742400; mul.ftz.f32 %f30905, %f27814, 0f49742400; bra.uni $L__BB107_4405; $L__BB107_4339: neg.ftz.f32 %f27275, %f7224; sqrt.approx.ftz.f32 %f27276, %f27275; add.ftz.f32 %f30807, %f27276, 0fC0400000; mov.f32 %f30815, 0fB951F09B; mov.f32 %f30814, 0f38D3B56B; mov.f32 %f30813, 0f3AB0DC72; mov.f32 %f30812, 0fBB70BDE7; mov.f32 %f30811, 0f3BBC127B; mov.f32 %f30810, 0fBBF9C5D7; mov.f32 %f30809, 0f3C1AA57E; mov.f32 %f30808, 0f3F8036DB; mov.f32 %f30806, 0f40354F7E; $L__BB107_4341: fma.rn.ftz.f32 %f27287, %f30807, %f30815, %f30814; fma.rn.ftz.f32 %f27288, %f30807, %f27287, %f30813; fma.rn.ftz.f32 %f27289, %f30807, %f27288, %f30812; fma.rn.ftz.f32 %f27290, %f30807, %f27289, %f30811; fma.rn.ftz.f32 %f27291, %f30807, %f27290, %f30810; fma.rn.ftz.f32 %f27292, %f30807, %f27291, %f30809; fma.rn.ftz.f32 %f27293, %f30807, %f27292, %f30808; fma.rn.ftz.f32 %f27294, %f30807, %f27293, %f30806; mul.ftz.f32 %f27295, %f7223, %f27294; mul.ftz.f32 %f27296, %f27295, %f27295; mul.ftz.f32 %f27297, %f27296, 0fBFB8AA3B; ex2.approx.ftz.f32 %f27298, %f27297; add.ftz.f32 %f27299, %f7223, 0f3F800000; fma.rn.ftz.f32 %f27301, %f7202, %f27298, %f27299; sub.ftz.f32 %f27302, %f27301, %f7204; mul.ftz.f32 %f27303, %f7198, %f27295; sub.ftz.f32 %f27304, %f27166, %f27303; div.approx.ftz.f32 %f27305, %f27302, %f27304; sub.ftz.f32 %f7237, %f7223, %f27305; abs.ftz.f32 %f27306, %f7237; setp.gt.ftz.f32 %p3706, %f27306, 0f3F7FFFFF; selp.f32 %f27307, 0f3F7FFFFF, %f27306, %p3706; sub.ftz.f32 %f27308, %f27166, %f27307; add.ftz.f32 %f27309, %f27307, 0f3F800000; mul.ftz.f32 %f27310, %f27308, %f27309; max.ftz.f32 %f27312, %f27310, %f27186; min.ftz.f32 %f27314, %f27312, %f27188; mov.b32 %r5339, %f27314; shr.u32 %r5340, %r5339, 23; add.s32 %r5341, %r5340, -127; and.b32 %r5342, %r5339, 8388607; or.b32 %r5343, %r5342, 1065353216; mov.b32 %f27315, %r5343; add.ftz.f32 %f27316, %f27315, 0fBF800000; mul.ftz.f32 %f27317, %f27316, %f27316; mul.ftz.f32 %f27318, %f27317, %f27317; fma.rn.ftz.f32 %f27319, %f27316, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f27320, %f27316, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f27321, %f27316, %f27319, 0fBE0CD4FD; fma.rn.ftz.f32 %f27322, %f27316, %f27321, 0f3E77ADBD; fma.rn.ftz.f32 %f27323, %f27316, %f27322, 0fBEB1D206; fma.rn.ftz.f32 %f27324, %f27316, %f27320, 0f3FB8AA10; mul.ftz.f32 %f27325, %f27318, %f27323; fma.rn.ftz.f32 %f27326, %f27316, %f27324, %f27325; cvt.rn.f32.s32 %f27327, %r5341; add.ftz.f32 %f27328, %f27326, %f27327; mul.ftz.f32 %f7238, %f27328, 0f3F317218; setp.gt.ftz.f32 %p3707, %f7238, 0fC0A00000; @%p3707 bra $L__BB107_4343; bra.uni $L__BB107_4342; $L__BB107_4343: mov.f32 %f27349, 0fC0200000; sub.ftz.f32 %f30817, %f27349, %f7238; mov.f32 %f30825, 0f32F16588; mov.f32 %f30824, 0f34B84B36; mov.f32 %f30823, 0fB66C7357; mov.f32 %f30822, 0fB6935AC1; mov.f32 %f30821, 0f396532DB; mov.f32 %f30820, 0fBAA45408; mov.f32 %f30819, 0fBB88E4EF; mov.f32 %f30818, 0f3E7C8F63; mov.f32 %f30816, 0f3FC02E2F; bra.uni $L__BB107_4344; $L__BB107_4342: neg.ftz.f32 %f27338, %f7238; sqrt.approx.ftz.f32 %f27339, %f27338; add.ftz.f32 %f30817, %f27339, 0fC0400000; mov.f32 %f30825, 0fB951F09B; mov.f32 %f30824, 0f38D3B56B; mov.f32 %f30823, 0f3AB0DC72; mov.f32 %f30822, 0fBB70BDE7; mov.f32 %f30821, 0f3BBC127B; mov.f32 %f30820, 0fBBF9C5D7; mov.f32 %f30819, 0f3C1AA57E; mov.f32 %f30818, 0f3F8036DB; mov.f32 %f30816, 0f40354F7E; $L__BB107_4344: fma.rn.ftz.f32 %f27350, %f30817, %f30825, %f30824; fma.rn.ftz.f32 %f27351, %f30817, %f27350, %f30823; fma.rn.ftz.f32 %f27352, %f30817, %f27351, %f30822; fma.rn.ftz.f32 %f27353, %f30817, %f27352, %f30821; fma.rn.ftz.f32 %f27354, %f30817, %f27353, %f30820; fma.rn.ftz.f32 %f27355, %f30817, %f27354, %f30819; fma.rn.ftz.f32 %f27356, %f30817, %f27355, %f30818; fma.rn.ftz.f32 %f27357, %f30817, %f27356, %f30816; mul.ftz.f32 %f30839, %f7237, %f27357; $L__BB107_4345: fma.rn.ftz.f32 %f7253, %f6202, 0f40000000, 0fBF800000; abs.ftz.f32 %f27359, %f7253; setp.gt.ftz.f32 %p3708, %f27359, 0f3F7FFFFF; selp.f32 %f27360, 0f3F7FFFFF, %f27359, %p3708; sub.ftz.f32 %f27361, %f27166, %f27360; add.ftz.f32 %f27362, %f27360, 0f3F800000; mul.ftz.f32 %f27363, %f27361, %f27362; max.ftz.f32 %f27365, %f27363, %f27186; min.ftz.f32 %f27367, %f27365, %f27188; mov.b32 %r5344, %f27367; shr.u32 %r5345, %r5344, 23; add.s32 %r5346, %r5345, -127; and.b32 %r5347, %r5344, 8388607; or.b32 %r5348, %r5347, 1065353216; mov.b32 %f27368, %r5348; add.ftz.f32 %f27369, %f27368, 0fBF800000; mul.ftz.f32 %f27370, %f27369, %f27369; mul.ftz.f32 %f27371, %f27370, %f27370; fma.rn.ftz.f32 %f27372, %f27369, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f27373, %f27369, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f27374, %f27369, %f27372, 0fBE0CD4FD; fma.rn.ftz.f32 %f27375, %f27369, %f27374, 0f3E77ADBD; fma.rn.ftz.f32 %f27376, %f27369, %f27375, 0fBEB1D206; fma.rn.ftz.f32 %f27377, %f27369, %f27373, 0f3FB8AA10; mul.ftz.f32 %f27378, %f27371, %f27376; fma.rn.ftz.f32 %f27379, %f27369, %f27377, %f27378; cvt.rn.f32.s32 %f27380, %r5346; add.ftz.f32 %f27381, %f27379, %f27380; mul.ftz.f32 %f7254, %f27381, 0f3F317218; setp.gt.ftz.f32 %p3709, %f7254, 0fC0A00000; @%p3709 bra $L__BB107_4347; bra.uni $L__BB107_4346; $L__BB107_4347: mov.f32 %f27402, 0fC0200000; sub.ftz.f32 %f30828, %f27402, %f7254; mov.f32 %f30836, 0f32F16588; mov.f32 %f30835, 0f34B84B36; mov.f32 %f30834, 0fB66C7357; mov.f32 %f30833, 0fB6935AC1; mov.f32 %f30832, 0f396532DB; mov.f32 %f30831, 0fBAA45408; mov.f32 %f30830, 0fBB88E4EF; mov.f32 %f30829, 0f3E7C8F63; mov.f32 %f30827, 0f3FC02E2F; bra.uni $L__BB107_4348; $L__BB107_4346: neg.ftz.f32 %f27391, %f7254; sqrt.approx.ftz.f32 %f27392, %f27391; add.ftz.f32 %f30828, %f27392, 0fC0400000; mov.f32 %f30836, 0fB951F09B; mov.f32 %f30835, 0f38D3B56B; mov.f32 %f30834, 0f3AB0DC72; mov.f32 %f30833, 0fBB70BDE7; mov.f32 %f30832, 0f3BBC127B; mov.f32 %f30831, 0fBBF9C5D7; mov.f32 %f30830, 0f3C1AA57E; mov.f32 %f30829, 0f3F8036DB; mov.f32 %f30827, 0f40354F7E; $L__BB107_4348: fma.rn.ftz.f32 %f27403, %f30828, %f30836, %f30835; fma.rn.ftz.f32 %f27404, %f30828, %f27403, %f30834; fma.rn.ftz.f32 %f27405, %f30828, %f27404, %f30833; fma.rn.ftz.f32 %f27406, %f30828, %f27405, %f30832; fma.rn.ftz.f32 %f27407, %f30828, %f27406, %f30831; fma.rn.ftz.f32 %f27408, %f30828, %f27407, %f30830; fma.rn.ftz.f32 %f27409, %f30828, %f27408, %f30829; fma.rn.ftz.f32 %f27410, %f30828, %f27409, %f30827; mul.ftz.f32 %f30838, %f7253, %f27410; add.ftz.f32 %f27411, %f7201, 0f3F800000; rcp.approx.ftz.f32 %f30837, %f27411; $L__BB107_4349: mul.ftz.f32 %f27412, %f30794, %f30838; mul.ftz.f32 %f27413, %f30793, %f30839; sub.ftz.f32 %f27414, %f27413, %f27412; mul.ftz.f32 %f27415, %f30794, %f30839; fma.rn.ftz.f32 %f27416, %f30793, %f30838, %f27415; mul.ftz.f32 %f27417, %f27069, %f27414; mul.ftz.f32 %f27418, %f27069, %f27416; neg.ftz.f32 %f27419, %f27417; neg.ftz.f32 %f27420, %f27418; mul.ftz.f32 %f27421, %f27418, %f27418; fma.rn.ftz.f32 %f27422, %f27417, %f27417, %f27421; add.ftz.f32 %f27423, %f27422, 0f3F800000; mov.f32 %f30841, 0f3F800000; sqrt.approx.ftz.f32 %f27425, %f27423; div.approx.ftz.f32 %f27426, %f27419, %f27425; div.approx.ftz.f32 %f27427, %f27420, %f27425; div.approx.ftz.f32 %f7272, %f30841, %f27425; mul.ftz.f32 %f27428, %f30785, %f27427; mul.ftz.f32 %f27429, %f30786, %f27427; mul.ftz.f32 %f27430, %f30787, %f27427; fma.rn.ftz.f32 %f27431, %f30788, %f27426, %f27428; fma.rn.ftz.f32 %f27432, %f30789, %f27426, %f27429; fma.rn.ftz.f32 %f27433, %f30790, %f27426, %f27430; fma.rn.ftz.f32 %f7273, %f7163, %f7272, %f27431; fma.rn.ftz.f32 %f7274, %f7165, %f7272, %f27432; fma.rn.ftz.f32 %f7275, %f7166, %f7272, %f27433; mul.ftz.f32 %f27434, %f27067, %f7274; fma.rn.ftz.f32 %f27435, %f7164, %f7273, %f27434; fma.rn.ftz.f32 %f7276, %f27068, %f7275, %f27435; setp.eq.s32 %p3710, %r1522, 22; @%p3710 bra $L__BB107_4359; bra.uni $L__BB107_4350; $L__BB107_4359: setp.gt.ftz.f32 %p3717, %f7276, 0f00000000; ld.local.f32 %f7294, [%rd1027+28]; @%p3717 bra $L__BB107_4361; bra.uni $L__BB107_4360; $L__BB107_4361: rcp.approx.ftz.f32 %f30846, %f7294; mov.u16 %rs745, 0; mov.f32 %f30842, %f7273; mov.f32 %f30843, %f7274; mov.f32 %f30844, %f7275; mov.f32 %f30845, %f7276; bra.uni $L__BB107_4362; $L__BB107_4350: setp.leu.ftz.f32 %p3711, %f7276, 0f00000000; mov.u32 %r5349, 10; mov.f32 %f30902, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; mov.u32 %r5877, %r5349; @%p3711 bra $L__BB107_4433; add.ftz.f32 %f27445, %f7276, %f7276; mul.ftz.f32 %f27446, %f7273, %f27445; mul.ftz.f32 %f27447, %f7274, %f27445; mul.ftz.f32 %f27448, %f7275, %f27445; sub.ftz.f32 %f30899, %f27446, %f7164; sub.ftz.f32 %f30900, %f27447, %f27067; sub.ftz.f32 %f30901, %f27448, %f27068; mul.ftz.f32 %f27449, %f6234, %f30900; fma.rn.ftz.f32 %f27450, %f6233, %f30899, %f27449; fma.rn.ftz.f32 %f27451, %f6235, %f30901, %f27450; setp.leu.ftz.f32 %p3712, %f27451, 0f00000000; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; mov.u32 %r5877, %r5349; @%p3712 bra $L__BB107_4433; mul.ftz.f32 %f7280, %f27069, %f27070; setp.le.ftz.f32 %p3713, %f7280, 0f33D6BF95; mov.u32 %r5877, 18; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3713 bra $L__BB107_4433; mul.ftz.f32 %f7281, %f7280, 0f40490FDB; mul.ftz.f32 %f27456, %f7165, %f30900; fma.rn.ftz.f32 %f27457, %f7163, %f30899, %f27456; fma.rn.ftz.f32 %f27458, %f7166, %f30901, %f27457; mul.ftz.f32 %f7282, %f27458, %f27458; sub.ftz.f32 %f7283, %f30841, %f7282; @%p3693 bra $L__BB107_4356; bra.uni $L__BB107_4354; $L__BB107_4356: mul.ftz.f32 %f27509, %f7272, %f7272; mul.ftz.f32 %f27510, %f27509, %f27509; rcp.approx.ftz.f32 %f27511, %f27509; add.ftz.f32 %f27512, %f27511, 0fBF800000; neg.ftz.f32 %f27513, %f27512; div.approx.ftz.f32 %f27514, %f27513, %f7280; mul.ftz.f32 %f27515, %f27514, 0f3FB8AA3B; ex2.approx.ftz.f32 %f27516, %f27515; mul.ftz.f32 %f27517, %f7281, %f27510; div.approx.ftz.f32 %f30840, %f27516, %f27517; div.approx.ftz.f32 %f27518, %f7283, %f7282; mov.f32 %f27519, 0f00000000; max.ftz.f32 %f27520, %f27518, %f27519; sqrt.approx.ftz.f32 %f27521, %f27520; mul.ftz.f32 %f7288, %f27069, %f27521; setp.lt.ftz.f32 %p3716, %f7288, 0f3F200000; @%p3716 bra $L__BB107_4358; rcp.approx.ftz.f32 %f27522, %f7288; fma.rn.ftz.f32 %f27523, %f27522, 0f400B9581, 0f40623D71; mul.ftz.f32 %f27524, %f27522, %f27523; fma.rn.ftz.f32 %f27525, %f27522, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f27526, %f27522, %f27525, 0f3F800000; div.approx.ftz.f32 %f30841, %f27524, %f27526; bra.uni $L__BB107_4358; $L__BB107_4360: neg.ftz.f32 %f30845, %f7276; neg.ftz.f32 %f30842, %f7273; neg.ftz.f32 %f30843, %f7274; neg.ftz.f32 %f30844, %f7275; mov.u16 %rs745, 1; mov.f32 %f30846, %f7294; $L__BB107_4362: mul.ftz.f32 %f27534, %f30845, %f30845; mov.f32 %f30850, 0f3F800000; sub.ftz.f32 %f27535, %f30850, %f27534; mul.ftz.f32 %f27536, %f30846, %f30846; mul.ftz.f32 %f27537, %f27536, %f27535; sub.ftz.f32 %f7305, %f30850, %f27537; setp.lt.ftz.f32 %p3718, %f7305, 0f00000000; mov.f32 %f30902, 0f00000000; mov.f32 %f30899, %f30902; mov.f32 %f30900, %f30902; mov.f32 %f30901, %f30902; @%p3718 bra $L__BB107_4364; sqrt.approx.ftz.f32 %f27538, %f7305; mov.f32 %f27539, 0f33D6BF95; max.ftz.f32 %f27540, %f27538, %f27539; mul.ftz.f32 %f27541, %f30845, %f30846; sub.ftz.f32 %f27542, %f27541, %f27540; mul.ftz.f32 %f27543, %f30842, %f27542; mul.ftz.f32 %f27544, %f30843, %f27542; mul.ftz.f32 %f27545, %f30844, %f27542; mul.ftz.f32 %f27546, %f7164, %f30846; sub.ftz.f32 %f30899, %f27543, %f27546; mul.ftz.f32 %f27547, %f27067, %f30846; sub.ftz.f32 %f30900, %f27544, %f27547; mul.ftz.f32 %f27548, %f27068, %f30846; sub.ftz.f32 %f30901, %f27545, %f27548; mul.ftz.f32 %f27549, %f30843, %f30900; fma.rn.ftz.f32 %f27550, %f30842, %f30899, %f27549; fma.rn.ftz.f32 %f27551, %f30844, %f30901, %f27550; mul.ftz.f32 %f27552, %f7294, %f27551; add.ftz.f32 %f27553, %f30845, %f27552; sub.ftz.f32 %f27554, %f30845, %f27552; div.approx.ftz.f32 %f27555, %f27553, %f27554; mul.ftz.f32 %f27556, %f7294, %f30845; add.ftz.f32 %f27557, %f27556, %f27551; sub.ftz.f32 %f27558, %f27556, %f27551; div.approx.ftz.f32 %f27559, %f27557, %f27558; mul.ftz.f32 %f27560, %f27559, %f27559; fma.rn.ftz.f32 %f27561, %f27555, %f27555, %f27560; mul.ftz.f32 %f30850, %f27561, 0f3F000000; $L__BB107_4364: setp.eq.ftz.f32 %p3719, %f30850, 0f3F800000; setp.ne.s16 %p3720, %rs745, 0; mov.u32 %r5353, 9; or.pred %p3721, %p3720, %p3719; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; mov.u32 %r5877, %r5353; @%p3721 bra $L__BB107_4433; mul.ftz.f32 %f7314, %f27069, %f27070; setp.le.ftz.f32 %p3722, %f7314, 0f33D6BF95; mov.u32 %r5877, 17; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3722 bra $L__BB107_4433; add.ftz.f32 %f27575, %f7294, 0fBF800000; abs.ftz.f32 %f27576, %f27575; setp.lt.ftz.f32 %p3723, %f27576, 0f38D1B717; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3723 bra $L__BB107_4433; mul.ftz.f32 %f27578, %f7272, %f7272; mul.ftz.f32 %f27579, %f27578, %f27578; rcp.approx.ftz.f32 %f27580, %f27578; mov.f32 %f30851, 0f3F800000; add.ftz.f32 %f27581, %f27580, 0fBF800000; neg.ftz.f32 %f27582, %f27581; div.approx.ftz.f32 %f27583, %f27582, %f7314; mul.ftz.f32 %f27584, %f27583, 0f3FB8AA3B; ex2.approx.ftz.f32 %f27585, %f27584; mul.ftz.f32 %f27586, %f7314, 0f40490FDB; mul.ftz.f32 %f27587, %f27586, %f27579; div.approx.ftz.f32 %f7315, %f27585, %f27587; mul.ftz.f32 %f27588, %f7165, %f30900; fma.rn.ftz.f32 %f27589, %f7163, %f30899, %f27588; fma.rn.ftz.f32 %f27590, %f7166, %f30901, %f27589; mul.ftz.f32 %f27591, %f27590, %f27590; sub.ftz.f32 %f27592, %f30851, %f27591; div.approx.ftz.f32 %f27593, %f27592, %f27591; mov.f32 %f27594, 0f00000000; max.ftz.f32 %f27595, %f27593, %f27594; sqrt.approx.ftz.f32 %f27596, %f27595; mul.ftz.f32 %f7316, %f27069, %f27596; setp.lt.ftz.f32 %p3724, %f7316, 0f3F200000; @%p3724 bra $L__BB107_4369; rcp.approx.ftz.f32 %f27597, %f7316; fma.rn.ftz.f32 %f27598, %f27597, 0f400B9581, 0f40623D71; mul.ftz.f32 %f27599, %f27597, %f27598; fma.rn.ftz.f32 %f27600, %f27597, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f27601, %f27597, %f27600, 0f3F800000; div.approx.ftz.f32 %f30851, %f27599, %f27601; $L__BB107_4369: mul.ftz.f32 %f27602, %f7274, %f30900; fma.rn.ftz.f32 %f27603, %f7273, %f30899, %f27602; fma.rn.ftz.f32 %f27604, %f7275, %f30901, %f27603; fma.rn.ftz.f32 %f27605, %f7294, %f27604, %f7276; mul.ftz.f32 %f27606, %f27605, %f27605; mul.ftz.f32 %f27607, %f7294, %f7294; mul.ftz.f32 %f27608, %f27607, %f7315; mul.ftz.f32 %f27609, %f7167, %f27606; div.approx.ftz.f32 %f27610, %f27608, %f27609; mul.ftz.f32 %f27611, %f7276, %f27604; abs.ftz.f32 %f27612, %f27611; mul.ftz.f32 %f27613, %f30837, %f30851; mul.ftz.f32 %f27614, %f27613, %f27612; mul.ftz.f32 %f30903, %f27610, %f27614; abs.ftz.f32 %f27615, %f27604; mul.ftz.f32 %f27616, %f30837, %f7276; mul.ftz.f32 %f27617, %f27616, %f27615; mul.ftz.f32 %f30902, %f27610, %f27617; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; mov.u32 %r5877, %r5353; bra.uni $L__BB107_4433; $L__BB107_3671: cvt.u32.u64 %r4644, %rd637; ld.const.v2.u64 {%rd3657, %rd3658}, [kernel_params+2224]; cvta.to.global.u64 %rd3661, %rd3658; mul.wide.s32 %rd3662, %r4644, 16; add.s64 %rd3663, %rd3661, %rd3662; ld.global.v4.u32 {%r4645, %r4646, %r4647, %r4648}, [%rd3663]; ld.const.u64 %rd3664, [kernel_params+2256]; cvta.to.global.u64 %rd3665, %rd3664; mul.wide.u32 %rd3666, %r4648, 12; add.s64 %rd3667, %rd3665, %rd3666; ld.global.f32 %f30376, [%rd3667]; ld.global.f32 %f30377, [%rd3667+4]; ld.global.f32 %f30378, [%rd3667+8]; add.s32 %r4653, %r4648, 1; mul.wide.u32 %rd3668, %r4653, 12; add.s64 %rd3669, %rd3665, %rd3668; ld.global.f32 %f30379, [%rd3669]; ld.global.f32 %f30380, [%rd3669+4]; ld.global.f32 %f30381, [%rd3669+8]; add.s32 %r4654, %r4648, 2; mul.wide.u32 %rd3670, %r4654, 12; add.s64 %rd3671, %rd3665, %rd3670; ld.global.f32 %f30382, [%rd3671]; ld.global.f32 %f30383, [%rd3671+4]; ld.global.f32 %f30384, [%rd3671+8]; cvta.to.global.u64 %rd3672, %rd3657; mul.wide.u32 %rd3673, %r4645, 12; add.s64 %rd3674, %rd3672, %rd3673; ld.global.f32 %f30367, [%rd3674]; ld.global.f32 %f30368, [%rd3674+4]; ld.global.f32 %f30369, [%rd3674+8]; mul.wide.u32 %rd3675, %r4646, 12; add.s64 %rd3676, %rd3672, %rd3675; ld.global.f32 %f30370, [%rd3676]; ld.global.f32 %f30371, [%rd3676+4]; ld.global.f32 %f30372, [%rd3676+8]; mul.wide.u32 %rd3677, %r4647, 12; add.s64 %rd3678, %rd3672, %rd3677; ld.global.f32 %f30373, [%rd3678]; ld.global.f32 %f30374, [%rd3678+4]; ld.global.f32 %f30375, [%rd3678+8]; bra.uni $L__BB107_3699; $L__BB107_4422: and.b32 %r5380, %r1552, 4; setp.eq.s32 %p3786, %r5380, 0; add.ftz.f32 %f28321, %f7521, 0f00000000; selp.f32 %f30896, 0f00000000, %f28321, %p3786; $L__BB107_4423: and.b32 %r5381, %r1552, 9; setp.eq.s32 %p3787, %r5381, 0; @%p3787 bra $L__BB107_4425; mul.ftz.f32 %f28322, %f28254, %f30900; fma.rn.ftz.f32 %f28323, %f7508, %f30899, %f28322; fma.rn.ftz.f32 %f28324, %f28255, %f30901, %f28323; add.ftz.f32 %f28325, %f28324, 0f3F800000; ld.local.f32 %f28326, [%rd1027+20]; mul.ftz.f32 %f28327, %f28325, %f28326; add.ftz.f32 %f28328, %f28327, 0fBF800000; mul.ftz.f32 %f28329, %f7519, %f7520; add.ftz.f32 %f28330, %f7519, %f7520; fma.rn.ftz.f32 %f28331, %f28329, %f28328, %f28330; fma.rn.ftz.f32 %f30896, %f28327, %f28331, %f30896; $L__BB107_4425: mul.ftz.f32 %f28332, %f7518, 0f3EA2F983; mul.ftz.f32 %f30903, %f28332, %f30896; $L__BB107_4426: mul.ftz.f32 %f30902, %f7511, 0f3EA2F983; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; bra.uni $L__BB107_4433; $L__BB107_4311: setp.gt.ftz.f32 %p3686, %f7004, 0f3F7FF972; mul.ftz.f32 %f26957, %f7139, %f7139; mov.f32 %f26958, 0f3F800000; sub.ftz.f32 %f26959, %f26958, %f7142; fma.rn.ftz.f32 %f26960, %f26957, %f7142, %f26959; mul.ftz.f32 %f26961, %f26960, 0f40490FDB; mul.ftz.f32 %f26962, %f26960, %f26961; mov.f32 %f26963, 0f33D6BF95; max.ftz.f32 %f26964, %f26962, %f26963; div.approx.ftz.f32 %f7143, %f26957, %f26964; mov.f32 %f30782, 0f00000000; @%p3686 bra $L__BB107_4314; mov.f32 %f30782, 0fBF7FF972; @%p3648 bra $L__BB107_4314; mul.ftz.f32 %f26966, %f7004, %f7004; max.ftz.f32 %f26968, %f26966, %f26963; rcp.approx.ftz.f32 %f26969, %f26968; mul.ftz.f32 %f26970, %f7028, %f7139; mul.ftz.f32 %f26971, %f7029, %f7139; mul.ftz.f32 %f26972, %f26971, %f26971; fma.rn.ftz.f32 %f26973, %f26970, %f26970, %f26972; fma.rn.ftz.f32 %f26974, %f26973, %f26969, 0f3F800000; sqrt.approx.ftz.f32 %f26975, %f26974; add.ftz.f32 %f26976, %f26975, 0fBF800000; mul.ftz.f32 %f30782, %f26976, 0f3F000000; $L__BB107_4314: add.ftz.f32 %f26977, %f30782, 0f3F800000; mul.ftz.f32 %f26979, %f7004, %f26977; max.ftz.f32 %f26981, %f26979, %f26963; mul.ftz.f32 %f26982, %f7143, 0f3E800000; div.approx.ftz.f32 %f26983, %f26982, %f26981; mul.ftz.f32 %f26984, %f7139, 0fBFFE59F3; mul.ftz.f32 %f26985, %f7139, %f26984; mul.ftz.f32 %f26986, %f26985, 0f3FB8AA3B; ex2.approx.ftz.f32 %f26987, %f26986; fma.rn.ftz.f32 %f26988, %f26987, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f26989, %f7139, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f26990, %f7139, %f26989, 0f417C8B44; fma.rn.ftz.f32 %f26991, %f7139, %f26990, 0fC1665E35; fma.rn.ftz.f32 %f26992, %f7139, %f26991, 0f40CE7881; fma.rn.ftz.f32 %f26993, %f7139, %f26992, 0fBF993E81; fma.rn.ftz.f32 %f26994, %f7139, %f26993, 0f3CE3C31E; fma.rn.ftz.f32 %f26995, %f7139, %f26994, 0f3BBA5D49; sub.ftz.f32 %f26996, %f26988, %f26995; cvt.ftz.sat.f32.f32 %f26997, %f26996; sub.ftz.f32 %f26998, %f26958, %f26997; mul.ftz.f32 %f26999, %f30778, 0f3EA2F983; mul.ftz.f32 %f27000, %f26999, %f26998; fma.rn.ftz.f32 %f30784, %f26983, %f26997, %f27000; $L__BB107_4319: mov.f32 %f27051, 0f00000000; max.ftz.f32 %f30902, %f27051, %f30784; mul.ftz.f32 %f30903, %f30903, %f30902; mul.ftz.f32 %f30904, %f30904, %f30902; mul.ftz.f32 %f30905, %f30905, %f30902; mov.u32 %r5877, 10; bra.uni $L__BB107_4433; $L__BB107_4243: setp.lt.ftz.f32 %p3623, %f6226, 0f3F000000; @%p3623 bra $L__BB107_4251; bra.uni $L__BB107_4244; $L__BB107_4251: mov.f32 %f26518, 0f3F000000; sub.ftz.f32 %f26519, %f26518, %f6226; mul.ftz.f32 %f26520, %f26519, 0f40800000; mul.ftz.f32 %f26521, %f26520, 0f3FC90FDB; sin.approx.ftz.f32 %f26522, %f26521; cos.approx.ftz.f32 %f26523, %f26521; div.approx.ftz.f32 %f26524, %f26522, %f26523; mul.ftz.f32 %f6952, %f6939, %f26524; abs.ftz.f32 %f6953, %f6952; setp.leu.ftz.f32 %p3631, %f6953, 0f3F800000; mov.f32 %f30739, %f6953; @%p3631 bra $L__BB107_4253; rcp.approx.ftz.f32 %f30739, %f6953; $L__BB107_4253: mul.rn.ftz.f32 %f26525, %f30739, %f30739; mov.f32 %f26526, 0fC0B59883; mov.f32 %f26527, 0fBF52C7EA; fma.rn.ftz.f32 %f26528, %f26525, %f26527, %f26526; mov.f32 %f26529, 0fC0D21907; fma.rn.ftz.f32 %f26530, %f26528, %f26525, %f26529; mul.ftz.f32 %f26531, %f26525, %f26530; mul.ftz.f32 %f26532, %f30739, %f26531; add.ftz.f32 %f26533, %f26525, 0f41355DC0; mov.f32 %f26534, 0f41E6BD60; fma.rn.ftz.f32 %f26535, %f26533, %f26525, %f26534; mov.f32 %f26536, 0f419D92C8; fma.rn.ftz.f32 %f26537, %f26535, %f26525, %f26536; rcp.approx.ftz.f32 %f26538, %f26537; fma.rn.ftz.f32 %f26539, %f26532, %f26538, %f30739; mov.f32 %f26540, 0f3FC90FDB; sub.ftz.f32 %f26541, %f26540, %f26539; setp.gt.ftz.f32 %p3632, %f6953, 0f3F800000; selp.f32 %f26542, %f26541, %f26539, %p3632; mov.b32 %r5307, %f26542; mov.b32 %r5308, %f6952; and.b32 %r5309, %r5308, -2147483648; or.b32 %r5310, %r5309, %r5307; mov.b32 %f26543, %r5310; setp.le.ftz.f32 %p3633, %f6953, 0f7F800000; selp.f32 %f26544, %f26543, %f26542, %p3633; cos.approx.ftz.f32 %f26545, %f26544; sin.approx.ftz.f32 %f26546, %f26544; mul.ftz.f32 %f26547, %f6918, %f26545; mul.ftz.f32 %f26548, %f6919, %f26546; mul.ftz.f32 %f26549, %f26546, %f26548; fma.rn.ftz.f32 %f26550, %f26545, %f26547, %f26549; add.ftz.f32 %f26551, %f26550, 0f3F800000; rcp.approx.ftz.f32 %f26552, %f26551; lg2.approx.ftz.f32 %f26553, %f6202; mul.ftz.f32 %f26554, %f26553, %f26552; ex2.approx.ftz.f32 %f30741, %f26554; mov.f32 %f26555, 0f40490FDB; sub.ftz.f32 %f30742, %f26555, %f26544; bra.uni $L__BB107_4258; $L__BB107_4224: mov.f32 %f26003, 0f3F800000; div.approx.ftz.f32 %f30728, %f6866, %f26003; $L__BB107_4226: add.ftz.f32 %f26006, %f30728, 0fBF800000; add.ftz.f32 %f26007, %f26006, 0f3F800000; mul.ftz.f32 %f26008, %f26007, %f26007; fma.rn.ftz.f32 %f26009, %f26008, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f26010, %f26007, %f26009; fma.rn.ftz.f32 %f26011, %f26008, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f26012, %f26008, %f26011, 0f3F800000; div.approx.ftz.f32 %f26013, %f26010, %f26012; sub.ftz.f32 %f26015, %f25994, %f26013; setp.gt.ftz.f32 %p3581, %f6866, 0f3F800000; selp.f32 %f26016, %f26015, %f26013, %p3581; mov.b32 %r5262, %f6865; and.b32 %r5263, %r5262, -2147483648; mov.b32 %r5264, %f26016; and.b32 %r5265, %r5264, 2147483647; or.b32 %r1540, %r5265, %r5263; add.ftz.f32 %f26017, %f6861, 0fBFC90FDB; mul.ftz.f32 %f26018, %f26017, 0f3F000000; sub.ftz.f32 %f26019, %f26018, %f6864; mul.ftz.f32 %f6870, %f6863, %f26019; abs.ftz.f32 %f6871, %f6870; setp.eq.ftz.f32 %p3582, %f6871, 0f00000000; @%p3582 bra $L__BB107_4231; setp.eq.ftz.f32 %p3583, %f6871, 0f3F800000; mov.f32 %f30729, 0f3F800000; @%p3583 bra $L__BB107_4231; setp.gt.ftz.f32 %p3584, %f6871, 0f3F800000; @%p3584 bra $L__BB107_4230; bra.uni $L__BB107_4229; $L__BB107_4230: mov.f32 %f26022, 0f3F800000; div.approx.ftz.f32 %f30729, %f26022, %f6871; bra.uni $L__BB107_4231; $L__BB107_4001: mov.f32 %f24606, 0f3F800000; div.approx.ftz.f32 %f30622, %f6251, %f24606; $L__BB107_4003: add.ftz.f32 %f24609, %f30622, 0fBF800000; add.ftz.f32 %f24610, %f24609, 0f3F800000; mul.ftz.f32 %f24611, %f24610, %f24610; fma.rn.ftz.f32 %f24612, %f24611, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f24613, %f24610, %f24612; fma.rn.ftz.f32 %f24614, %f24611, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f24615, %f24611, %f24614, 0f3F800000; div.approx.ftz.f32 %f24616, %f24613, %f24615; sub.ftz.f32 %f24618, %f24596, %f24616; setp.gt.ftz.f32 %p3371, %f6251, 0f3F800000; selp.f32 %f24619, %f24618, %f24616, %p3371; mov.b32 %r5177, %f6250; and.b32 %r5178, %r5177, -2147483648; mov.b32 %r5179, %f24619; and.b32 %r5180, %r5179, 2147483647; or.b32 %r1523, %r5180, %r5178; add.ftz.f32 %f24620, %f6246, 0fBFC90FDB; div.approx.ftz.f32 %f24622, %f24620, %f24602; sub.ftz.f32 %f24623, %f24622, %f6249; mul.ftz.f32 %f6255, %f6248, %f24623; abs.ftz.f32 %f6256, %f6255; setp.eq.ftz.f32 %p3372, %f6256, 0f00000000; mov.f32 %f30623, %f30624; @%p3372 bra $L__BB107_4008; setp.eq.ftz.f32 %p3373, %f6256, 0f3F800000; mov.f32 %f30623, 0f3F800000; @%p3373 bra $L__BB107_4008; setp.gt.ftz.f32 %p3374, %f6256, 0f3F800000; @%p3374 bra $L__BB107_4007; bra.uni $L__BB107_4006; $L__BB107_4007: mov.f32 %f24626, 0f3F800000; div.approx.ftz.f32 %f30623, %f24626, %f6256; bra.uni $L__BB107_4008; $L__BB107_4229: mov.f32 %f26021, 0f3F800000; div.approx.ftz.f32 %f30729, %f6871, %f26021; $L__BB107_4231: add.ftz.f32 %f26023, %f30729, 0fBF800000; mov.f32 %f26024, 0fBF800000; mov.f32 %f26025, 0f3F800000; add.ftz.f32 %f26026, %f26023, 0f3F800000; mul.ftz.f32 %f26027, %f26026, %f26026; fma.rn.ftz.f32 %f26028, %f26027, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f26029, %f26026, %f26028; fma.rn.ftz.f32 %f26030, %f26027, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f26031, %f26027, %f26030, 0f3F800000; div.approx.ftz.f32 %f26032, %f26029, %f26031; sub.ftz.f32 %f26034, %f25994, %f26032; setp.gt.ftz.f32 %p3585, %f6871, 0f3F800000; selp.f32 %f26035, %f26034, %f26032, %p3585; mov.b32 %r5267, %f6870; and.b32 %r5268, %r5267, -2147483648; mov.b32 %r5269, %f26035; and.b32 %r5270, %r5269, 2147483647; or.b32 %r5271, %r5270, %r5268; mov.b32 %f26036, %r5271; mov.b32 %f26037, %r1540; sub.ftz.f32 %f26038, %f26037, %f26036; fma.rn.ftz.f32 %f26039, %f6226, %f26038, %f26036; sin.approx.ftz.f32 %f26040, %f26039; cos.approx.ftz.f32 %f26041, %f26039; div.approx.ftz.f32 %f26042, %f26040, %f26041; mul.ftz.f32 %f26043, %f6862, %f26042; add.ftz.f32 %f26044, %f6864, %f26043; add.ftz.f32 %f26045, %f26044, %f26044; sub.ftz.f32 %f26046, %f26045, %f6861; mul.ftz.f32 %f26047, %f26046, 0f3EA2F983; mov.b32 %r5272, %f26047; and.b32 %r5273, %r5272, -2147483648; or.b32 %r5274, %r5273, 1056964608; mov.b32 %f26048, %r5274; add.ftz.f32 %f26049, %f26047, %f26048; cvt.rzi.ftz.s32.f32 %r5275, %f26049; cvt.rn.f32.s32 %f26050, %r5275; mul.ftz.f32 %f26051, %f26050, 0f40490000; sub.ftz.f32 %f26052, %f26046, %f26051; mul.ftz.f32 %f26053, %f26050, 0f3A7DA000; sub.ftz.f32 %f26054, %f26052, %f26053; mul.ftz.f32 %f26055, %f26050, 0f34222000; sub.ftz.f32 %f26056, %f26054, %f26055; mul.ftz.f32 %f26057, %f26050, 0f2CB4611A; sub.ftz.f32 %f26058, %f26056, %f26057; add.ftz.f32 %f26059, %f26058, 0fBFC90FDB; add.ftz.f32 %f26060, %f26059, 0f3FC90FDB; mul.ftz.f32 %f26061, %f26060, %f26060; and.b32 %r5276, %r5275, 1; setp.eq.b32 %p3586, %r5276, 1; neg.ftz.f32 %f26062, %f26060; selp.f32 %f26063, %f26062, %f26060, %p3586; fma.rn.ftz.f32 %f26064, %f26061, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f26065, %f26061, %f26064, 0f3C088777; fma.rn.ftz.f32 %f26066, %f26061, %f26065, 0fBE2AAAA6; mul.ftz.f32 %f26067, %f26063, %f26066; fma.rn.ftz.f32 %f26068, %f26061, %f26067, %f26063; mul.ftz.f32 %f26069, %f26061, 0f3491ED89; mov.f32 %f26070, 0f37D0078B; sub.ftz.f32 %f26071, %f26070, %f26069; fma.rn.ftz.f32 %f26072, %f26061, %f26071, 0fBAB60B58; fma.rn.ftz.f32 %f26073, %f26061, %f26072, 0f3D2AAAAA; fma.rn.ftz.f32 %f26074, %f26061, %f26073, 0fBF000000; fma.rn.ftz.f32 %f26075, %f26061, %f26074, 0f3F800000; neg.ftz.f32 %f26076, %f26075; selp.f32 %f26077, %f26076, %f26075, %p3586; abs.ftz.f32 %f26078, %f26068; setp.gt.ftz.f32 %p3587, %f26078, 0f3F800000; selp.f32 %f26079, 0f00000000, %f26068, %p3587; abs.ftz.f32 %f26080, %f26077; setp.gt.ftz.f32 %p3588, %f26080, 0f3F800000; selp.f32 %f26081, 0f00000000, %f26077, %p3588; add.ftz.f32 %f26082, %f6202, %f6202; sub.ftz.f32 %f26083, %f26025, %f26082; max.ftz.f32 %f26084, %f26083, %f26024; min.ftz.f32 %f26085, %f26084, %f26025; abs.ftz.f32 %f26086, %f26085; sub.ftz.f32 %f26087, %f26025, %f26086; mul.ftz.f32 %f26088, %f26087, 0f3F000000; sqrt.approx.ftz.f32 %f26089, %f26088; setp.gt.ftz.f32 %p3589, %f26086, 0f3F11EB85; selp.f32 %f26090, %f26089, %f26086, %p3589; mul.ftz.f32 %f26091, %f26090, %f26090; mov.f32 %f26092, 0f3C94D2E9; mov.f32 %f26093, 0f3D53F941; fma.rn.ftz.f32 %f26094, %f26093, %f26091, %f26092; mov.f32 %f26095, 0f3D3F841F; fma.rn.ftz.f32 %f26096, %f26094, %f26091, %f26095; mov.f32 %f26097, 0f3D994929; fma.rn.ftz.f32 %f26098, %f26096, %f26091, %f26097; mov.f32 %f26099, 0f3E2AAB94; fma.rn.ftz.f32 %f26100, %f26098, %f26091, %f26099; mul.ftz.f32 %f26101, %f26091, %f26100; fma.rn.ftz.f32 %f26102, %f26101, %f26090, %f26090; mov.f32 %f26103, 0fC0000000; fma.rn.ftz.f32 %f26104, %f26103, %f26102, %f25994; selp.f32 %f26105, %f26104, %f26102, %p3589; setp.le.ftz.f32 %p3590, %f26105, 0f7F800000; mov.b32 %r5277, %f26105; mov.b32 %r5278, %f26085; and.b32 %r5279, %r5278, -2147483648; or.b32 %r5280, %r5279, %r5277; mov.b32 %f26106, %r5280; selp.f32 %f26107, %f26106, %f26105, %p3590; add.ftz.f32 %f26108, %f26107, %f26107; mul.ftz.f32 %f26109, %f6854, %f26108; mul.ftz.f32 %f26110, %f26109, 0f3F000000; mul.ftz.f32 %f26111, %f26110, 0f3EA2F983; mov.b32 %r5281, %f26111; and.b32 %r5282, %r5281, -2147483648; or.b32 %r5283, %r5282, 1056964608; mov.b32 %f26112, %r5283; add.ftz.f32 %f26113, %f26111, %f26112; cvt.rzi.ftz.s32.f32 %r5284, %f26113; cvt.rn.f32.s32 %f26114, %r5284; mul.ftz.f32 %f26115, %f26114, 0f40490000; sub.ftz.f32 %f26116, %f26110, %f26115; mul.ftz.f32 %f26117, %f26114, 0f3A7DA000; sub.ftz.f32 %f26118, %f26116, %f26117; mul.ftz.f32 %f26119, %f26114, 0f34222000; sub.ftz.f32 %f26120, %f26118, %f26119; mul.ftz.f32 %f26121, %f26114, 0f2CB4611A; sub.ftz.f32 %f26122, %f26120, %f26121; add.ftz.f32 %f26123, %f26122, 0fBFC90FDB; add.ftz.f32 %f26124, %f26123, 0f3FC90FDB; mul.ftz.f32 %f26125, %f26124, %f26124; mul.ftz.f32 %f26126, %f26125, 0f3491ED89; sub.ftz.f32 %f26127, %f26070, %f26126; fma.rn.ftz.f32 %f26128, %f26125, %f26127, 0fBAB60B58; fma.rn.ftz.f32 %f26129, %f26125, %f26128, 0f3D2AAAAA; fma.rn.ftz.f32 %f26130, %f26125, %f26129, 0fBF000000; fma.rn.ftz.f32 %f26131, %f26125, %f26130, 0f3F800000; and.b32 %r5285, %r5284, 1; setp.eq.b32 %p3591, %r5285, 1; neg.ftz.f32 %f26132, %f26131; selp.f32 %f26133, %f26132, %f26131, %p3591; abs.ftz.f32 %f26134, %f26133; setp.gt.ftz.f32 %p3592, %f26134, 0f3F800000; mul.ftz.f32 %f26135, %f26133, 0f3E800000; selp.f32 %f26136, 0f00000000, %f26135, %p3592; div.approx.ftz.f32 %f26137, %f26136, %f6854; mul.ftz.f32 %f26138, %f26043, %f26043; fma.rn.ftz.f32 %f26139, %f6862, %f6862, %f26138; add.ftz.f32 %f26140, %f26139, %f26139; mul.ftz.f32 %f26141, %f26140, %f26038; mul.ftz.f32 %f26142, %f26141, %f26081; div.approx.ftz.f32 %f26143, %f6862, %f26142; mul.ftz.f32 %f26144, %f26109, 0f3EA2F983; mov.b32 %r5286, %f26144; and.b32 %r5287, %r5286, -2147483648; or.b32 %r5288, %r5287, 1056964608; mov.b32 %f26145, %r5288; add.ftz.f32 %f26146, %f26144, %f26145; cvt.rzi.ftz.s32.f32 %r5289, %f26146; cvt.rn.f32.s32 %f26147, %r5289; mul.ftz.f32 %f26148, %f26147, 0f40490000; sub.ftz.f32 %f26149, %f26109, %f26148; mul.ftz.f32 %f26150, %f26147, 0f3A7DA000; sub.ftz.f32 %f26151, %f26149, %f26150; mul.ftz.f32 %f26152, %f26147, 0f34222000; sub.ftz.f32 %f26153, %f26151, %f26152; mul.ftz.f32 %f26154, %f26147, 0f2CB4611A; sub.ftz.f32 %f26155, %f26153, %f26154; add.ftz.f32 %f26156, %f26155, 0fBFC90FDB; add.ftz.f32 %f26157, %f26156, 0f3FC90FDB; mul.ftz.f32 %f26158, %f26157, %f26157; and.b32 %r5290, %r5289, 1; setp.eq.b32 %p3593, %r5290, 1; neg.ftz.f32 %f26159, %f26157; selp.f32 %f26160, %f26159, %f26157, %p3593; fma.rn.ftz.f32 %f26161, %f26158, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f26162, %f26158, %f26161, 0f3C088777; fma.rn.ftz.f32 %f26163, %f26158, %f26162, 0fBE2AAAA6; mul.ftz.f32 %f26164, %f26160, %f26163; fma.rn.ftz.f32 %f26165, %f26158, %f26164, %f26160; mul.ftz.f32 %f26166, %f26158, 0f3491ED89; sub.ftz.f32 %f26167, %f26070, %f26166; fma.rn.ftz.f32 %f26168, %f26158, %f26167, 0fBAB60B58; fma.rn.ftz.f32 %f26169, %f26158, %f26168, 0f3D2AAAAA; fma.rn.ftz.f32 %f26170, %f26158, %f26169, 0fBF000000; fma.rn.ftz.f32 %f26171, %f26158, %f26170, 0f3F800000; neg.ftz.f32 %f26172, %f26171; selp.f32 %f26173, %f26172, %f26171, %p3593; abs.ftz.f32 %f26174, %f26165; setp.gt.ftz.f32 %p3594, %f26174, 0f3F800000; selp.f32 %f26175, 0f00000000, %f26165, %p3594; abs.ftz.f32 %f26176, %f26173; setp.gt.ftz.f32 %p3595, %f26176, 0f3F800000; selp.f32 %f26177, 0f00000000, %f26173, %p3595; mul.ftz.f32 %f26178, %f26081, %f26177; mul.ftz.f32 %f26179, %f6858, %f26178; mul.ftz.f32 %f26180, %f6859, %f26178; mul.ftz.f32 %f26181, %f6860, %f26178; mul.ftz.f32 %f26182, %f26081, %f26175; mul.ftz.f32 %f26183, %f6857, %f6859; mul.ftz.f32 %f26184, %f6856, %f6860; sub.ftz.f32 %f26185, %f26183, %f26184; mul.ftz.f32 %f26186, %f26185, %f26182; mul.ftz.f32 %f26187, %f6855, %f6860; mul.ftz.f32 %f26188, %f6857, %f6858; sub.ftz.f32 %f26189, %f26187, %f26188; mul.ftz.f32 %f26190, %f26189, %f26182; mul.ftz.f32 %f26191, %f6856, %f6858; mul.ftz.f32 %f26192, %f6855, %f6859; sub.ftz.f32 %f26193, %f26191, %f26192; mul.ftz.f32 %f26194, %f26193, %f26182; sub.ftz.f32 %f26195, %f26179, %f26186; sub.ftz.f32 %f26196, %f26180, %f26190; sub.ftz.f32 %f26197, %f26181, %f26194; fma.rn.ftz.f32 %f30899, %f6855, %f26079, %f26195; fma.rn.ftz.f32 %f30900, %f6856, %f26079, %f26196; fma.rn.ftz.f32 %f30901, %f6857, %f26079, %f26197; mul.ftz.f32 %f26198, %f26137, %f26143; abs.ftz.f32 %f26199, %f26198; abs.ftz.f32 %f26200, %f26046; sub.ftz.f32 %f26201, %f25994, %f26200; setp.lt.ftz.f32 %p3596, %f26201, 0f3A83126F; selp.f32 %f30902, 0f00000000, %f26199, %p3596; mov.u32 %r5877, 10; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_4006: mov.f32 %f24625, 0f3F800000; div.approx.ftz.f32 %f30623, %f6256, %f24625; $L__BB107_4008: add.ftz.f32 %f24628, %f30623, 0fBF800000; add.ftz.f32 %f24629, %f24628, 0f3F800000; mul.ftz.f32 %f24630, %f24629, %f24629; fma.rn.ftz.f32 %f24631, %f24630, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f24632, %f24629, %f24631; fma.rn.ftz.f32 %f24633, %f24630, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f24634, %f24630, %f24633, 0f3F800000; div.approx.ftz.f32 %f24635, %f24632, %f24634; sub.ftz.f32 %f24637, %f24596, %f24635; setp.gt.ftz.f32 %p3375, %f6256, 0f3F800000; selp.f32 %f24638, %f24637, %f24635, %p3375; mov.b32 %r5181, %f6255; and.b32 %r5182, %r5181, -2147483648; mov.b32 %r5183, %f24638; and.b32 %r5184, %r5183, 2147483647; or.b32 %r1524, %r5184, %r5182; div.approx.ftz.f32 %f6260, %f24596, %f6239; abs.ftz.f32 %f6261, %f6260; setp.eq.ftz.f32 %p3376, %f6261, 0f00000000; @%p3376 bra $L__BB107_4013; setp.eq.ftz.f32 %p3377, %f6261, 0f3F800000; mov.f32 %f30624, 0f3F800000; @%p3377 bra $L__BB107_4013; setp.gt.ftz.f32 %p3378, %f6261, 0f3F800000; @%p3378 bra $L__BB107_4012; bra.uni $L__BB107_4011; $L__BB107_4012: mov.f32 %f24641, 0f3F800000; div.approx.ftz.f32 %f30624, %f24641, %f6261; bra.uni $L__BB107_4013; $L__BB107_4011: mov.f32 %f24640, 0f3F800000; div.approx.ftz.f32 %f30624, %f6261, %f24640; $L__BB107_4013: mov.b32 %f24642, %r1524; add.ftz.f32 %f24643, %f30624, 0fBF800000; add.ftz.f32 %f24644, %f24643, 0f3F800000; mul.ftz.f32 %f24645, %f24644, %f24644; fma.rn.ftz.f32 %f24646, %f24645, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f24647, %f24644, %f24646; fma.rn.ftz.f32 %f24648, %f24645, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f24649, %f24645, %f24648, 0f3F800000; div.approx.ftz.f32 %f24650, %f24647, %f24649; sub.ftz.f32 %f24652, %f24596, %f24650; setp.gt.ftz.f32 %p3379, %f6261, 0f3F800000; selp.f32 %f24653, %f24652, %f24650, %p3379; mov.b32 %r5186, %f6260; and.b32 %r5187, %r5186, -2147483648; mov.b32 %r5188, %f24653; and.b32 %r5189, %r5188, 2147483647; or.b32 %r5190, %r5189, %r5187; mov.b32 %f24654, %r5190; add.ftz.f32 %f24655, %f24654, %f24654; mov.b32 %f24656, %r1523; sub.ftz.f32 %f24657, %f24656, %f24642; fma.rn.ftz.f32 %f24658, %f6226, %f24657, %f24642; sin.approx.ftz.f32 %f24659, %f24658; cos.approx.ftz.f32 %f24660, %f24658; div.approx.ftz.f32 %f24661, %f24659, %f24660; mul.ftz.f32 %f24662, %f6247, %f24661; add.ftz.f32 %f24663, %f6249, %f24662; add.ftz.f32 %f24664, %f24663, %f24663; sub.ftz.f32 %f24665, %f24664, %f6246; mul.ftz.f32 %f24666, %f24665, 0f3EA2F983; mov.b32 %r5191, %f24666; and.b32 %r5192, %r5191, -2147483648; or.b32 %r5193, %r5192, 1056964608; mov.b32 %f24667, %r5193; add.ftz.f32 %f24668, %f24666, %f24667; cvt.rzi.ftz.s32.f32 %r5194, %f24668; cvt.rn.f32.s32 %f24669, %r5194; mul.ftz.f32 %f24670, %f24669, 0f40490000; sub.ftz.f32 %f24671, %f24665, %f24670; mul.ftz.f32 %f24672, %f24669, 0f3A7DA000; sub.ftz.f32 %f24673, %f24671, %f24672; mul.ftz.f32 %f24674, %f24669, 0f34222000; sub.ftz.f32 %f24675, %f24673, %f24674; mul.ftz.f32 %f24676, %f24669, 0f2CB4611A; sub.ftz.f32 %f24677, %f24675, %f24676; add.ftz.f32 %f24678, %f24677, 0fBFC90FDB; add.ftz.f32 %f24679, %f24678, 0f3FC90FDB; mul.ftz.f32 %f24680, %f24679, %f24679; and.b32 %r5195, %r5194, 1; setp.eq.b32 %p3380, %r5195, 1; neg.ftz.f32 %f24681, %f24679; selp.f32 %f24682, %f24681, %f24679, %p3380; fma.rn.ftz.f32 %f24683, %f24680, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f24684, %f24680, %f24683, 0f3C088777; fma.rn.ftz.f32 %f24685, %f24680, %f24684, 0fBE2AAAA6; mul.ftz.f32 %f24686, %f24682, %f24685; fma.rn.ftz.f32 %f24687, %f24680, %f24686, %f24682; mul.ftz.f32 %f24688, %f24680, 0f3491ED89; mov.f32 %f24689, 0f37D0078B; sub.ftz.f32 %f24690, %f24689, %f24688; fma.rn.ftz.f32 %f24691, %f24680, %f24690, 0fBAB60B58; fma.rn.ftz.f32 %f24692, %f24680, %f24691, 0f3D2AAAAA; fma.rn.ftz.f32 %f24693, %f24680, %f24692, 0fBF000000; fma.rn.ftz.f32 %f24694, %f24680, %f24693, 0f3F800000; neg.ftz.f32 %f24695, %f24694; selp.f32 %f24696, %f24695, %f24694, %p3380; abs.ftz.f32 %f24697, %f24687; setp.gt.ftz.f32 %p3381, %f24697, 0f3F800000; selp.f32 %f24698, 0f00000000, %f24687, %p3381; abs.ftz.f32 %f24699, %f24696; setp.gt.ftz.f32 %p3382, %f24699, 0f3F800000; selp.f32 %f24700, 0f00000000, %f24696, %p3382; add.ftz.f32 %f24701, %f6202, 0fBF000000; mul.ftz.f32 %f24702, %f24701, %f24655; sin.approx.ftz.f32 %f24703, %f24702; cos.approx.ftz.f32 %f24704, %f24702; div.approx.ftz.f32 %f24705, %f24703, %f24704; mul.ftz.f32 %f24706, %f6239, %f24705; add.ftz.f32 %f24707, %f24706, 0f40490FDB; mul.ftz.f32 %f24708, %f24662, %f24662; fma.rn.ftz.f32 %f24709, %f6247, %f6247, %f24708; add.ftz.f32 %f24710, %f24709, %f24709; mul.ftz.f32 %f24711, %f24657, %f24710; mul.ftz.f32 %f24712, %f24711, %f24700; div.approx.ftz.f32 %f24713, %f6247, %f24712; mul.ftz.f32 %f24714, %f24706, %f24706; fma.rn.ftz.f32 %f24715, %f6239, %f6239, %f24714; mul.ftz.f32 %f24716, %f24655, %f24715; div.approx.ftz.f32 %f24717, %f6239, %f24716; mul.ftz.f32 %f24718, %f24707, 0f3EA2F983; mov.b32 %r5196, %f24718; and.b32 %r5197, %r5196, -2147483648; or.b32 %r5198, %r5197, 1056964608; mov.b32 %f24719, %r5198; add.ftz.f32 %f24720, %f24718, %f24719; cvt.rzi.ftz.s32.f32 %r5199, %f24720; cvt.rn.f32.s32 %f24721, %r5199; mul.ftz.f32 %f24722, %f24721, 0f40490000; sub.ftz.f32 %f24723, %f24707, %f24722; mul.ftz.f32 %f24724, %f24721, 0f3A7DA000; sub.ftz.f32 %f24725, %f24723, %f24724; mul.ftz.f32 %f24726, %f24721, 0f34222000; sub.ftz.f32 %f24727, %f24725, %f24726; mul.ftz.f32 %f24728, %f24721, 0f2CB4611A; sub.ftz.f32 %f24729, %f24727, %f24728; add.ftz.f32 %f24730, %f24729, 0fBFC90FDB; add.ftz.f32 %f24731, %f24730, 0f3FC90FDB; mul.ftz.f32 %f24732, %f24731, %f24731; and.b32 %r5200, %r5199, 1; setp.eq.b32 %p3383, %r5200, 1; neg.ftz.f32 %f24733, %f24731; selp.f32 %f24734, %f24733, %f24731, %p3383; fma.rn.ftz.f32 %f24735, %f24732, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f24736, %f24732, %f24735, 0f3C088777; fma.rn.ftz.f32 %f24737, %f24732, %f24736, 0fBE2AAAA6; mul.ftz.f32 %f24738, %f24734, %f24737; fma.rn.ftz.f32 %f24739, %f24732, %f24738, %f24734; mul.ftz.f32 %f24740, %f24732, 0f3491ED89; sub.ftz.f32 %f24741, %f24689, %f24740; fma.rn.ftz.f32 %f24742, %f24732, %f24741, 0fBAB60B58; fma.rn.ftz.f32 %f24743, %f24732, %f24742, 0f3D2AAAAA; fma.rn.ftz.f32 %f24744, %f24732, %f24743, 0fBF000000; fma.rn.ftz.f32 %f24745, %f24732, %f24744, 0f3F800000; neg.ftz.f32 %f24746, %f24745; selp.f32 %f24747, %f24746, %f24745, %p3383; abs.ftz.f32 %f24748, %f24739; setp.gt.ftz.f32 %p3384, %f24748, 0f3F800000; selp.f32 %f24749, 0f00000000, %f24739, %p3384; abs.ftz.f32 %f24750, %f24747; setp.gt.ftz.f32 %p3385, %f24750, 0f3F800000; selp.f32 %f24751, 0f00000000, %f24747, %p3385; mul.ftz.f32 %f24752, %f24700, %f24751; mul.ftz.f32 %f24753, %f6243, %f24752; mul.ftz.f32 %f24754, %f6244, %f24752; mul.ftz.f32 %f24755, %f6245, %f24752; mul.ftz.f32 %f24756, %f24700, %f24749; mul.ftz.f32 %f24757, %f6241, %f6245; mul.ftz.f32 %f24758, %f6242, %f6244; sub.ftz.f32 %f24759, %f24758, %f24757; mul.ftz.f32 %f24760, %f24759, %f24756; mul.ftz.f32 %f24761, %f6242, %f6243; mul.ftz.f32 %f24762, %f6240, %f6245; sub.ftz.f32 %f24763, %f24762, %f24761; mul.ftz.f32 %f24764, %f24763, %f24756; mul.ftz.f32 %f24765, %f6240, %f6244; mul.ftz.f32 %f24766, %f6241, %f6243; sub.ftz.f32 %f24767, %f24766, %f24765; mul.ftz.f32 %f24768, %f24767, %f24756; sub.ftz.f32 %f24769, %f24753, %f24760; sub.ftz.f32 %f24770, %f24754, %f24764; sub.ftz.f32 %f24771, %f24755, %f24768; fma.rn.ftz.f32 %f30899, %f6240, %f24698, %f24769; fma.rn.ftz.f32 %f30900, %f6241, %f24698, %f24770; fma.rn.ftz.f32 %f30901, %f6242, %f24698, %f24771; mul.ftz.f32 %f24772, %f24713, %f24717; abs.ftz.f32 %f24773, %f24772; abs.ftz.f32 %f24774, %f24665; sub.ftz.f32 %f24775, %f24596, %f24774; setp.lt.ftz.f32 %p3386, %f24775, 0f3A83126F; selp.f32 %f30902, 0f00000000, %f24773, %p3386; mov.u32 %r5877, 9; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; bra.uni $L__BB107_4433; $L__BB107_3680: setp.gt.s32 %p3134, %r1343, %r1342; selp.b32 %r4663, -1, 0, %p3134; add.s32 %r4664, %r1343, %r4663; mad.lo.s32 %r4665, %r4664, %r1341, %r5812; ld.const.u64 %rd3688, [kernel_params+2336]; cvta.to.global.u64 %rd3689, %rd3688; add.s32 %r4666, %r4659, %r4665; mul.wide.u32 %rd3690, %r4666, 12; add.s64 %rd4836, %rd3689, %rd3690; add.s32 %r4667, %r4660, %r4665; mul.wide.u32 %rd3691, %r4667, 12; add.s64 %rd4839, %rd3689, %rd3691; add.s32 %r4668, %r4661, %r4665; mul.wide.u32 %rd3692, %r4668, 12; add.s64 %rd4842, %rd3689, %rd3692; add.s64 %rd4837, %rd4836, 4; add.s64 %rd4838, %rd4836, 8; add.s64 %rd4840, %rd4839, 4; add.s64 %rd4841, %rd4839, 8; add.s64 %rd4843, %rd4842, 4; add.s64 %rd4844, %rd4842, 8; $L__BB107_3682: ld.global.f32 %f5470, [%rd4844]; ld.global.f32 %f5471, [%rd4843]; ld.global.f32 %f5472, [%rd4842]; ld.global.f32 %f5473, [%rd4841]; ld.global.f32 %f5474, [%rd4840]; ld.global.f32 %f5475, [%rd4839]; ld.global.f32 %f5476, [%rd4838]; ld.global.f32 %f5477, [%rd4837]; ld.global.f32 %f5478, [%rd4836]; add.s32 %r1358, %r1343, 1; setp.eq.s32 %p3135, %r1358, %r1342; @%p3135 bra $L__BB107_3684; bra.uni $L__BB107_3683; $L__BB107_3684: ld.const.u64 %rd3703, [kernel_params+2256]; cvta.to.global.u64 %rd3704, %rd3703; mul.wide.u32 %rd3705, %r4662, 12; add.s64 %rd4845, %rd3704, %rd3705; add.s32 %r4677, %r4662, 1; mul.wide.u32 %rd3706, %r4677, 12; add.s64 %rd4848, %rd3704, %rd3706; add.s32 %r4678, %r4662, 2; mul.wide.u32 %rd3707, %r4678, 12; add.s64 %rd4851, %rd3704, %rd3707; add.s64 %rd4846, %rd4845, 4; add.s64 %rd4847, %rd4845, 8; add.s64 %rd4849, %rd4848, 4; add.s64 %rd4850, %rd4848, 8; add.s64 %rd4852, %rd4851, 4; add.s64 %rd4853, %rd4851, 8; bra.uni $L__BB107_3685; $L__BB107_3683: setp.ge.s32 %p3136, %r1343, %r1342; selp.b32 %r4671, -1, 0, %p3136; add.s32 %r4672, %r1358, %r4671; mad.lo.s32 %r4673, %r4672, %r1341, %r5812; ld.const.u64 %rd3698, [kernel_params+2336]; cvta.to.global.u64 %rd3699, %rd3698; add.s32 %r4674, %r4659, %r4673; mul.wide.u32 %rd3700, %r4674, 12; add.s64 %rd4845, %rd3699, %rd3700; add.s32 %r4675, %r4660, %r4673; mul.wide.u32 %rd3701, %r4675, 12; add.s64 %rd4848, %rd3699, %rd3701; add.s32 %r4676, %r4661, %r4673; mul.wide.u32 %rd3702, %r4676, 12; add.s64 %rd4851, %rd3699, %rd3702; add.s64 %rd4846, %rd4845, 4; add.s64 %rd4847, %rd4845, 8; add.s64 %rd4849, %rd4848, 4; add.s64 %rd4850, %rd4848, 8; add.s64 %rd4852, %rd4851, 4; add.s64 %rd4853, %rd4851, 8; $L__BB107_3685: sub.ftz.f32 %f5479, %f23255, %f5469; ld.global.f32 %f23266, [%rd4845]; mul.ftz.f32 %f23267, %f5469, %f23266; ld.global.f32 %f23268, [%rd4846]; mul.ftz.f32 %f23269, %f5469, %f23268; ld.global.f32 %f23270, [%rd4847]; mul.ftz.f32 %f23271, %f5469, %f23270; fma.rn.ftz.f32 %f30376, %f5479, %f5478, %f23267; fma.rn.ftz.f32 %f30377, %f5479, %f5477, %f23269; fma.rn.ftz.f32 %f30378, %f5479, %f5476, %f23271; ld.global.f32 %f23272, [%rd4848]; mul.ftz.f32 %f23273, %f5469, %f23272; ld.global.f32 %f23274, [%rd4849]; mul.ftz.f32 %f23275, %f5469, %f23274; ld.global.f32 %f23276, [%rd4850]; mul.ftz.f32 %f23277, %f5469, %f23276; fma.rn.ftz.f32 %f30379, %f5479, %f5475, %f23273; fma.rn.ftz.f32 %f30380, %f5479, %f5474, %f23275; fma.rn.ftz.f32 %f30381, %f5479, %f5473, %f23277; ld.global.f32 %f23278, [%rd4851]; mul.ftz.f32 %f23279, %f5469, %f23278; ld.global.f32 %f23280, [%rd4852]; mul.ftz.f32 %f23281, %f5469, %f23280; ld.global.f32 %f23282, [%rd4853]; mul.ftz.f32 %f23283, %f5469, %f23282; fma.rn.ftz.f32 %f30382, %f5479, %f5472, %f23279; fma.rn.ftz.f32 %f30383, %f5479, %f5471, %f23281; fma.rn.ftz.f32 %f30384, %f5479, %f5470, %f23283; setp.eq.s64 %p3137, %rd4854, 12; @%p3137 bra $L__BB107_3691; mov.u32 %r5817, -1; $L__BB107_3687: setp.eq.s64 %p3138, %rd4854, 0; @%p3138 bra $L__BB107_3689; add.s32 %r5814, %r5814, 2; bra.uni $L__BB107_3690; $L__BB107_3689: setp.eq.s16 %p3139, %rs729, 0; mov.u32 %r5814, %r5813; @%p3139 bra $L__BB107_3692; $L__BB107_3690: mul.wide.u32 %rd3708, %r5814, 16; add.s64 %rd3709, %rd638, %rd3708; ld.global.u16 %rs729, [%rd3709+12]; ld.global.u32 %r5813, [%rd3709+8]; ld.global.u64 %rd4854, [%rd3709]; setp.ne.s64 %p3140, %rd4854, 12; @%p3140 bra $L__BB107_3687; $L__BB107_3691: setp.eq.s16 %p3141, %rs729, 0; selp.b32 %r5817, -1, %r5813, %p3141; $L__BB107_3692: @%p3133 bra $L__BB107_3694; bra.uni $L__BB107_3693; $L__BB107_3694: ld.const.u64 %rd3715, [kernel_params+2224]; cvta.to.global.u64 %rd3716, %rd3715; mul.wide.u32 %rd3717, %r4659, 12; add.s64 %rd4855, %rd3716, %rd3717; mul.wide.u32 %rd3718, %r4660, 12; add.s64 %rd4858, %rd3716, %rd3718; mul.wide.u32 %rd3719, %r4661, 12; add.s64 %rd4861, %rd3716, %rd3719; add.s64 %rd4856, %rd4855, 4; add.s64 %rd4857, %rd4855, 8; add.s64 %rd4859, %rd4858, 4; add.s64 %rd4860, %rd4858, 8; add.s64 %rd4862, %rd4861, 4; add.s64 %rd4863, %rd4861, 8; bra.uni $L__BB107_3695; $L__BB107_3693: setp.gt.s32 %p3143, %r1343, %r1342; selp.b32 %r4680, -1, 0, %p3143; add.s32 %r4681, %r1343, %r4680; mad.lo.s32 %r4682, %r4681, %r1341, %r5817; ld.const.u64 %rd3710, [kernel_params+2336]; cvta.to.global.u64 %rd3711, %rd3710; add.s32 %r4683, %r4682, %r4659; mul.wide.u32 %rd3712, %r4683, 12; add.s64 %rd4855, %rd3711, %rd3712; add.s32 %r4684, %r4682, %r4660; mul.wide.u32 %rd3713, %r4684, 12; add.s64 %rd4858, %rd3711, %rd3713; add.s32 %r4685, %r4682, %r4661; mul.wide.u32 %rd3714, %r4685, 12; add.s64 %rd4861, %rd3711, %rd3714; add.s64 %rd4856, %rd4855, 4; add.s64 %rd4857, %rd4855, 8; add.s64 %rd4859, %rd4858, 4; add.s64 %rd4860, %rd4858, 8; add.s64 %rd4862, %rd4861, 4; add.s64 %rd4863, %rd4861, 8; $L__BB107_3695: ld.global.f32 %f5489, [%rd4863]; ld.global.f32 %f5490, [%rd4862]; ld.global.f32 %f5491, [%rd4861]; ld.global.f32 %f5492, [%rd4860]; ld.global.f32 %f5493, [%rd4859]; ld.global.f32 %f5494, [%rd4858]; ld.global.f32 %f5495, [%rd4857]; ld.global.f32 %f5496, [%rd4856]; ld.global.f32 %f5497, [%rd4855]; @%p3135 bra $L__BB107_3697; bra.uni $L__BB107_3696; $L__BB107_3697: ld.const.u64 %rd3725, [kernel_params+2224]; cvta.to.global.u64 %rd3726, %rd3725; mul.wide.u32 %rd3727, %r4659, 12; add.s64 %rd4864, %rd3726, %rd3727; mul.wide.u32 %rd3728, %r4660, 12; add.s64 %rd4867, %rd3726, %rd3728; mul.wide.u32 %rd3729, %r4661, 12; add.s64 %rd4870, %rd3726, %rd3729; add.s64 %rd4865, %rd4864, 4; add.s64 %rd4866, %rd4864, 8; add.s64 %rd4868, %rd4867, 4; add.s64 %rd4869, %rd4867, 8; add.s64 %rd4871, %rd4870, 4; add.s64 %rd4872, %rd4870, 8; bra.uni $L__BB107_3698; $L__BB107_3696: setp.ge.s32 %p3145, %r1343, %r1342; selp.b32 %r4686, -1, 0, %p3145; add.s32 %r4687, %r1358, %r4686; mad.lo.s32 %r4688, %r4687, %r1341, %r5817; ld.const.u64 %rd3720, [kernel_params+2336]; cvta.to.global.u64 %rd3721, %rd3720; add.s32 %r4689, %r4688, %r4659; mul.wide.u32 %rd3722, %r4689, 12; add.s64 %rd4864, %rd3721, %rd3722; add.s32 %r4690, %r4688, %r4660; mul.wide.u32 %rd3723, %r4690, 12; add.s64 %rd4867, %rd3721, %rd3723; add.s32 %r4691, %r4688, %r4661; mul.wide.u32 %rd3724, %r4691, 12; add.s64 %rd4870, %rd3721, %rd3724; add.s64 %rd4865, %rd4864, 4; add.s64 %rd4866, %rd4864, 8; add.s64 %rd4868, %rd4867, 4; add.s64 %rd4869, %rd4867, 8; add.s64 %rd4871, %rd4870, 4; add.s64 %rd4872, %rd4870, 8; $L__BB107_3698: ld.global.f32 %f23284, [%rd4864]; mul.ftz.f32 %f23285, %f5469, %f23284; ld.global.f32 %f23286, [%rd4865]; mul.ftz.f32 %f23287, %f5469, %f23286; ld.global.f32 %f23288, [%rd4866]; mul.ftz.f32 %f23289, %f5469, %f23288; fma.rn.ftz.f32 %f30367, %f5479, %f5497, %f23285; fma.rn.ftz.f32 %f30368, %f5479, %f5496, %f23287; fma.rn.ftz.f32 %f30369, %f5479, %f5495, %f23289; ld.global.f32 %f23290, [%rd4867]; mul.ftz.f32 %f23291, %f5469, %f23290; ld.global.f32 %f23292, [%rd4868]; mul.ftz.f32 %f23293, %f5469, %f23292; ld.global.f32 %f23294, [%rd4869]; mul.ftz.f32 %f23295, %f5469, %f23294; fma.rn.ftz.f32 %f30370, %f5479, %f5494, %f23291; fma.rn.ftz.f32 %f30371, %f5479, %f5493, %f23293; fma.rn.ftz.f32 %f30372, %f5479, %f5492, %f23295; ld.global.f32 %f23296, [%rd4870]; mul.ftz.f32 %f23297, %f5469, %f23296; ld.global.f32 %f23298, [%rd4871]; mul.ftz.f32 %f23299, %f5469, %f23298; ld.global.f32 %f23300, [%rd4872]; mul.ftz.f32 %f23301, %f5469, %f23300; fma.rn.ftz.f32 %f30373, %f5479, %f5491, %f23297; fma.rn.ftz.f32 %f30374, %f5479, %f5490, %f23299; fma.rn.ftz.f32 %f30375, %f5479, %f5489, %f23301; $L__BB107_3699: ld.local.f32 %f5525, [%rd5+-20]; sub.ftz.f32 %f23303, %f23255, %f5525; ld.local.f32 %f23304, [%rd5+-16]; sub.ftz.f32 %f5526, %f23303, %f23304; mul.ftz.f32 %f23305, %f30376, %f5526; mul.ftz.f32 %f23306, %f30377, %f5526; mul.ftz.f32 %f23307, %f30378, %f5526; fma.rn.ftz.f32 %f23308, %f30379, %f5525, %f23305; fma.rn.ftz.f32 %f23309, %f30380, %f5525, %f23306; fma.rn.ftz.f32 %f23310, %f30381, %f5525, %f23307; fma.rn.ftz.f32 %f5527, %f30382, %f23304, %f23308; fma.rn.ftz.f32 %f5528, %f30383, %f23304, %f23309; fma.rn.ftz.f32 %f5529, %f30384, %f23304, %f23310; mul.ftz.f32 %f23311, %f30367, %f5526; mul.ftz.f32 %f23312, %f30368, %f5526; mul.ftz.f32 %f23313, %f30369, %f5526; fma.rn.ftz.f32 %f23314, %f30370, %f5525, %f23311; fma.rn.ftz.f32 %f23315, %f30371, %f5525, %f23312; fma.rn.ftz.f32 %f23316, %f30372, %f5525, %f23313; fma.rn.ftz.f32 %f30385, %f30373, %f23304, %f23314; fma.rn.ftz.f32 %f30386, %f30374, %f23304, %f23315; fma.rn.ftz.f32 %f30387, %f30375, %f23304, %f23316; ld.local.u32 %r1367, [%rd5+-32]; and.b32 %r4692, %r1367, 4; setp.ne.s32 %p3146, %r4692, 0; @%p3146 bra $L__BB107_3703; and.b32 %r4693, %r1367, 2; setp.eq.s32 %p3147, %r4693, 0; @%p3147 bra $L__BB107_3702; bra.uni $L__BB107_3701; $L__BB107_3702: ld.global.v4.f32 {%f23344, %f23345, %f23346, %f23347}, [%rd636+-184]; ld.global.v4.f32 {%f23351, %f23352, %f23353, %f23354}, [%rd636+-168]; ld.global.v4.f32 {%f23358, %f23359, %f23360, %f23361}, [%rd636+-152]; mul.ftz.f32 %f23365, %f30386, %f23345; fma.rn.ftz.f32 %f23366, %f30385, %f23344, %f23365; fma.rn.ftz.f32 %f5536, %f30387, %f23346, %f23366; mul.ftz.f32 %f23367, %f30386, %f23352; fma.rn.ftz.f32 %f23368, %f30385, %f23351, %f23367; fma.rn.ftz.f32 %f5537, %f30387, %f23353, %f23368; mul.ftz.f32 %f23369, %f30386, %f23359; fma.rn.ftz.f32 %f23370, %f30385, %f23358, %f23369; fma.rn.ftz.f32 %f30387, %f30387, %f23360, %f23370; mov.f32 %f30385, %f5536; mov.f32 %f30386, %f5537; bra.uni $L__BB107_3703; $L__BB107_3701: ld.local.v4.f32 {%f23317, %f23318, %f23319, %f23320}, [%rd5+56]; mul.ftz.f32 %f23324, %f30386, %f23318; fma.rn.ftz.f32 %f23325, %f30385, %f23317, %f23324; fma.rn.ftz.f32 %f5533, %f30387, %f23319, %f23325; ld.local.v4.f32 {%f23326, %f23327, %f23328, %f23329}, [%rd5+72]; mul.ftz.f32 %f23333, %f30386, %f23327; fma.rn.ftz.f32 %f23334, %f30385, %f23326, %f23333; fma.rn.ftz.f32 %f5534, %f30387, %f23328, %f23334; ld.local.v4.f32 {%f23335, %f23336, %f23337, %f23338}, [%rd5+88]; mul.ftz.f32 %f23342, %f30386, %f23336; fma.rn.ftz.f32 %f23343, %f30385, %f23335, %f23342; fma.rn.ftz.f32 %f30387, %f30387, %f23337, %f23343; mov.f32 %f30385, %f5533; mov.f32 %f30386, %f5534; $L__BB107_3703: sub.ftz.f32 %f5542, %f30376, %f30382; sub.ftz.f32 %f23371, %f30373, %f30367; sub.ftz.f32 %f5543, %f30377, %f30383; sub.ftz.f32 %f23372, %f30374, %f30368; mul.ftz.f32 %f23373, %f23372, %f5543; fma.rn.ftz.f32 %f23374, %f23371, %f5542, %f23373; sub.ftz.f32 %f5544, %f30378, %f30384; sub.ftz.f32 %f23375, %f30375, %f30369; fma.rn.ftz.f32 %f23376, %f23375, %f5544, %f23374; sub.ftz.f32 %f5545, %f30379, %f30382; sub.ftz.f32 %f23377, %f30373, %f30370; sub.ftz.f32 %f5546, %f30380, %f30383; sub.ftz.f32 %f23378, %f30374, %f30371; mul.ftz.f32 %f23379, %f23378, %f5546; fma.rn.ftz.f32 %f23380, %f23377, %f5545, %f23379; sub.ftz.f32 %f5547, %f30381, %f30384; sub.ftz.f32 %f23381, %f30375, %f30372; fma.rn.ftz.f32 %f23382, %f23381, %f5547, %f23380; sub.ftz.f32 %f5548, %f30379, %f30376; sub.ftz.f32 %f23383, %f30370, %f30367; sub.ftz.f32 %f5549, %f30380, %f30377; sub.ftz.f32 %f23384, %f30371, %f30368; mul.ftz.f32 %f23385, %f23384, %f5549; fma.rn.ftz.f32 %f23386, %f23383, %f5548, %f23385; sub.ftz.f32 %f5550, %f30381, %f30378; sub.ftz.f32 %f23387, %f30372, %f30369; fma.rn.ftz.f32 %f23388, %f23387, %f5550, %f23386; mul.ftz.f32 %f23389, %f23376, %f5526; add.ftz.f32 %f23390, %f5526, 0fBF800000; add.ftz.f32 %f23391, %f23376, %f23382; add.ftz.f32 %f23392, %f23388, %f23391; mul.ftz.f32 %f23393, %f5526, %f23392; mul.ftz.f32 %f23394, %f5525, %f23393; fma.rn.ftz.f32 %f23395, %f23389, %f23390, %f23394; mul.ftz.f32 %f23396, %f5525, %f23382; add.ftz.f32 %f23397, %f5525, 0fBF800000; fma.rn.ftz.f32 %f23398, %f23397, %f23396, %f23395; mul.ftz.f32 %f23399, %f30365, %f30386; fma.rn.ftz.f32 %f23400, %f30364, %f30385, %f23399; fma.rn.ftz.f32 %f23401, %f30366, %f30387, %f23400; setp.gt.ftz.f32 %p3148, %f23401, 0f00000000; sub.ftz.f32 %f5551, %f30376, %f30379; sub.ftz.f32 %f5552, %f30377, %f30380; sub.ftz.f32 %f5553, %f30378, %f30381; sub.ftz.f32 %f5554, %f30382, %f30376; sub.ftz.f32 %f5555, %f30383, %f30377; sub.ftz.f32 %f5556, %f30384, %f30378; sub.ftz.f32 %f5557, %f30382, %f30379; sub.ftz.f32 %f5558, %f30383, %f30380; sub.ftz.f32 %f5559, %f30384, %f30381; mul.ftz.f32 %f5560, %f23398, 0f3F000000; @%p3148 bra $L__BB107_3705; bra.uni $L__BB107_3704; $L__BB107_3705: mul.ftz.f32 %f23454, %f30368, %f5549; fma.rn.ftz.f32 %f23455, %f30367, %f5548, %f23454; fma.rn.ftz.f32 %f23456, %f30369, %f5550, %f23455; mul.ftz.f32 %f23457, %f30368, %f5555; fma.rn.ftz.f32 %f23458, %f30367, %f5554, %f23457; fma.rn.ftz.f32 %f23459, %f30369, %f5556, %f23458; max.ftz.f32 %f23460, %f23456, %f23459; max.ftz.f32 %f23462, %f23460, %f23259; mul.ftz.f32 %f23463, %f30371, %f5552; fma.rn.ftz.f32 %f23464, %f30370, %f5551, %f23463; fma.rn.ftz.f32 %f23465, %f30372, %f5553, %f23464; mul.ftz.f32 %f23466, %f30371, %f5558; fma.rn.ftz.f32 %f23467, %f30370, %f5557, %f23466; fma.rn.ftz.f32 %f23468, %f30372, %f5559, %f23467; max.ftz.f32 %f23469, %f23465, %f23468; max.ftz.f32 %f23470, %f23469, %f23259; mul.ftz.f32 %f23471, %f30374, %f5543; fma.rn.ftz.f32 %f23472, %f30373, %f5542, %f23471; fma.rn.ftz.f32 %f23473, %f30375, %f5544, %f23472; mul.ftz.f32 %f23474, %f30374, %f5546; fma.rn.ftz.f32 %f23475, %f30373, %f5545, %f23474; fma.rn.ftz.f32 %f23476, %f30375, %f5547, %f23475; max.ftz.f32 %f23477, %f23473, %f23476; max.ftz.f32 %f23478, %f23477, %f23259; sub.ftz.f32 %f23479, %f30376, %f5527; sub.ftz.f32 %f23480, %f30377, %f5528; mul.ftz.f32 %f23481, %f30368, %f23480; fma.rn.ftz.f32 %f23482, %f30367, %f23479, %f23481; sub.ftz.f32 %f23483, %f30378, %f5529; fma.rn.ftz.f32 %f23484, %f30369, %f23483, %f23482; add.ftz.f32 %f23485, %f23484, %f23462; max.ftz.f32 %f23486, %f23485, %f23259; sub.ftz.f32 %f23487, %f30379, %f5527; sub.ftz.f32 %f23488, %f30380, %f5528; mul.ftz.f32 %f23489, %f30371, %f23488; fma.rn.ftz.f32 %f23490, %f30370, %f23487, %f23489; sub.ftz.f32 %f23491, %f30381, %f5529; fma.rn.ftz.f32 %f23492, %f30372, %f23491, %f23490; add.ftz.f32 %f23493, %f23492, %f23470; max.ftz.f32 %f23494, %f23493, %f23259; sub.ftz.f32 %f23495, %f30382, %f5527; sub.ftz.f32 %f23496, %f30383, %f5528; mul.ftz.f32 %f23497, %f30374, %f23496; fma.rn.ftz.f32 %f23498, %f30373, %f23495, %f23497; sub.ftz.f32 %f23499, %f30384, %f5529; fma.rn.ftz.f32 %f23500, %f30375, %f23499, %f23498; add.ftz.f32 %f23501, %f23500, %f23478; max.ftz.f32 %f23502, %f23501, %f23259; min.ftz.f32 %f23503, %f23486, %f23494; min.ftz.f32 %f23504, %f23503, %f23502; max.ftz.f32 %f30388, %f23504, %f5560; bra.uni $L__BB107_3706; $L__BB107_3704: mul.ftz.f32 %f23402, %f30368, %f5552; fma.rn.ftz.f32 %f23403, %f30367, %f5551, %f23402; fma.rn.ftz.f32 %f23404, %f30369, %f5553, %f23403; mul.ftz.f32 %f23405, %f30368, %f5543; fma.rn.ftz.f32 %f23406, %f30367, %f5542, %f23405; fma.rn.ftz.f32 %f23407, %f30369, %f5544, %f23406; max.ftz.f32 %f23408, %f23404, %f23407; max.ftz.f32 %f23410, %f23408, %f23259; mul.ftz.f32 %f23411, %f30371, %f5549; fma.rn.ftz.f32 %f23412, %f30370, %f5548, %f23411; fma.rn.ftz.f32 %f23413, %f30372, %f5550, %f23412; mul.ftz.f32 %f23414, %f30371, %f5546; fma.rn.ftz.f32 %f23415, %f30370, %f5545, %f23414; fma.rn.ftz.f32 %f23416, %f30372, %f5547, %f23415; max.ftz.f32 %f23417, %f23413, %f23416; max.ftz.f32 %f23418, %f23417, %f23259; mul.ftz.f32 %f23419, %f30374, %f5555; fma.rn.ftz.f32 %f23420, %f30373, %f5554, %f23419; fma.rn.ftz.f32 %f23421, %f30375, %f5556, %f23420; mul.ftz.f32 %f23422, %f30374, %f5558; fma.rn.ftz.f32 %f23423, %f30373, %f5557, %f23422; fma.rn.ftz.f32 %f23424, %f30375, %f5559, %f23423; max.ftz.f32 %f23425, %f23421, %f23424; max.ftz.f32 %f23426, %f23425, %f23259; sub.ftz.f32 %f23427, %f5527, %f30376; sub.ftz.f32 %f23428, %f5528, %f30377; mul.ftz.f32 %f23429, %f30368, %f23428; fma.rn.ftz.f32 %f23430, %f30367, %f23427, %f23429; sub.ftz.f32 %f23431, %f5529, %f30378; fma.rn.ftz.f32 %f23432, %f30369, %f23431, %f23430; add.ftz.f32 %f23433, %f23432, %f23410; max.ftz.f32 %f23434, %f23433, %f23259; sub.ftz.f32 %f23435, %f5527, %f30379; sub.ftz.f32 %f23436, %f5528, %f30380; mul.ftz.f32 %f23437, %f30371, %f23436; fma.rn.ftz.f32 %f23438, %f30370, %f23435, %f23437; sub.ftz.f32 %f23439, %f5529, %f30381; fma.rn.ftz.f32 %f23440, %f30372, %f23439, %f23438; add.ftz.f32 %f23441, %f23440, %f23418; max.ftz.f32 %f23442, %f23441, %f23259; sub.ftz.f32 %f23443, %f5527, %f30382; sub.ftz.f32 %f23444, %f5528, %f30383; mul.ftz.f32 %f23445, %f30374, %f23444; fma.rn.ftz.f32 %f23446, %f30373, %f23443, %f23445; sub.ftz.f32 %f23447, %f5529, %f30384; fma.rn.ftz.f32 %f23448, %f30375, %f23447, %f23446; add.ftz.f32 %f23449, %f23448, %f23426; max.ftz.f32 %f23450, %f23449, %f23259; min.ftz.f32 %f23451, %f23434, %f23442; min.ftz.f32 %f23452, %f23451, %f23450; neg.ftz.f32 %f23453, %f23452; min.ftz.f32 %f30388, %f23453, %f5560; $L__BB107_3706: mul.ftz.f32 %f23505, %f30385, %f30388; mul.ftz.f32 %f23506, %f30386, %f30388; mul.ftz.f32 %f23507, %f30387, %f30388; fma.rn.ftz.f32 %f30389, %f5450, %f23505, %f30389; fma.rn.ftz.f32 %f30390, %f5450, %f23506, %f30390; fma.rn.ftz.f32 %f30391, %f5450, %f23507, %f30391; setp.gt.ftz.f32 %p3149, %f5449, 0f00000000; selp.u16 %rs731, 1, 0, %p3149; $L__BB107_3707: and.b32 %r4694, %r1233, 1073741824; setp.eq.s32 %p3150, %r4694, 0; mov.f32 %f30392, 0f00000000; mov.f32 %f30393, %f30392; mov.f32 %f30394, %f30392; mov.f32 %f30395, %f30392; mov.f32 %f30396, %f30392; mov.f32 %f30397, %f30392; mov.f32 %f30398, %f30392; @%p3150 bra $L__BB107_3710; ld.local.f32 %f23516, [%rd280+12]; setp.eq.ftz.f32 %p3151, %f23516, 0f7F7FFFFF; mov.f32 %f30398, 0f7F7FFFFF; mov.f32 %f30392, %f30389; mov.f32 %f30393, %f30390; mov.f32 %f30394, %f30391; mov.f32 %f30395, %f30148; mov.f32 %f30396, %f30147; mov.f32 %f30397, %f30146; @%p3151 bra $L__BB107_3710; ld.local.f32 %f23517, [%rd280+-24]; sub.ftz.f32 %f23518, %f23517, %f30389; ld.local.f32 %f23519, [%rd280+-20]; sub.ftz.f32 %f23520, %f23519, %f30390; ld.local.f32 %f23521, [%rd280+-16]; sub.ftz.f32 %f23522, %f23521, %f30391; mul.ftz.f32 %f23523, %f23520, %f23520; fma.rn.ftz.f32 %f23524, %f23518, %f23518, %f23523; fma.rn.ftz.f32 %f23525, %f23522, %f23522, %f23524; sqrt.approx.ftz.f32 %f30398, %f23525; rsqrt.approx.ftz.f32 %f23526, %f23525; mul.ftz.f32 %f30395, %f23518, %f23526; mul.ftz.f32 %f30396, %f23520, %f23526; mul.ftz.f32 %f30397, %f23522, %f23526; mov.f32 %f30392, %f30389; mov.f32 %f30393, %f30390; mov.f32 %f30394, %f30391; $L__BB107_3710: ld.local.f32 %f30443, [%rd5+4]; ld.local.f32 %f30442, [%rd5+-4]; ld.local.u32 %r4695, [%rd5+-12]; setp.eq.s16 %p3152, %rs731, 0; selp.b32 %r5829, -1, %r4695, %p3152; ld.local.u32 %r4696, [%rd5+-28]; selp.b32 %r5828, -1, %r4696, %p3152; ld.local.u32 %r5831, [%rd280+32]; ld.local.u32 %r5830, [%rd280+36]; bra.uni $L__BB107_3759; $L__BB107_4391: mul.ftz.f32 %f27819, %f30856, %f7369; fma.rn.ftz.f32 %f27820, %f30855, %f7368, %f27819; fma.rn.ftz.f32 %f27821, %f30857, %f7370, %f27820; mul.ftz.f32 %f27822, %f30853, %f7369; fma.rn.ftz.f32 %f27823, %f30852, %f7368, %f27822; fma.rn.ftz.f32 %f27824, %f30854, %f7370, %f27823; mul.ftz.f32 %f27825, %f7327, %f7369; fma.rn.ftz.f32 %f27826, %f7325, %f7368, %f27825; fma.rn.ftz.f32 %f27827, %f7328, %f7370, %f27826; neg.ftz.f32 %f27828, %f27821; mul.ftz.f32 %f27829, %f27620, %f27827; div.approx.ftz.f32 %f27830, %f27828, %f27829; neg.ftz.f32 %f27831, %f27824; mul.ftz.f32 %f27832, %f27621, %f27827; div.approx.ftz.f32 %f27833, %f27831, %f27832; fma.rn.ftz.f32 %f27834, %f27830, %f27830, 0f3F800000; mov.f32 %f27835, 0f3F800000; fma.rn.ftz.f32 %f27836, %f27833, %f27833, %f27834; mul.ftz.f32 %f27837, %f27827, %f27827; mul.ftz.f32 %f27838, %f27837, %f27837; mul.ftz.f32 %f27839, %f27836, %f27836; mul.ftz.f32 %f27840, %f27839, 0f40490FDB; mul.ftz.f32 %f27841, %f30869, %f27840; mul.ftz.f32 %f27842, %f27838, %f27841; rcp.approx.ftz.f32 %f30870, %f27842; mul.ftz.f32 %f27843, %f7392, %f7392; sub.ftz.f32 %f27844, %f27835, %f27843; div.approx.ftz.f32 %f27845, %f27844, %f27843; mul.ftz.f32 %f27846, %f30856, %f30900; fma.rn.ftz.f32 %f27847, %f30855, %f30899, %f27846; fma.rn.ftz.f32 %f27848, %f30857, %f30901, %f27847; mul.ftz.f32 %f27849, %f30853, %f30900; fma.rn.ftz.f32 %f27850, %f30852, %f30899, %f27849; fma.rn.ftz.f32 %f27851, %f30854, %f30901, %f27850; mul.ftz.f32 %f27852, %f27848, %f27848; mul.ftz.f32 %f27853, %f27620, %f27620; mul.ftz.f32 %f27854, %f27851, %f27851; mul.ftz.f32 %f27855, %f27621, %f27621; mul.ftz.f32 %f27856, %f27855, %f27854; fma.rn.ftz.f32 %f27857, %f27853, %f27852, %f27856; add.ftz.f32 %f27858, %f27852, %f27854; div.approx.ftz.f32 %f27859, %f27857, %f27858; mul.ftz.f32 %f30873, %f27845, %f27859; bra.uni $L__BB107_4398; $L__BB107_4244: setp.lt.ftz.f32 %p3624, %f6226, 0f3F400000; @%p3624 bra $L__BB107_4248; bra.uni $L__BB107_4245; $L__BB107_4248: add.ftz.f32 %f26482, %f6226, 0fBF000000; mul.ftz.f32 %f26483, %f26482, 0f40800000; mul.ftz.f32 %f26484, %f26483, 0f3FC90FDB; sin.approx.ftz.f32 %f26485, %f26484; cos.approx.ftz.f32 %f26486, %f26484; div.approx.ftz.f32 %f26487, %f26485, %f26486; mul.ftz.f32 %f6946, %f6939, %f26487; abs.ftz.f32 %f6947, %f6946; setp.leu.ftz.f32 %p3628, %f6947, 0f3F800000; mov.f32 %f30738, %f6947; @%p3628 bra $L__BB107_4250; rcp.approx.ftz.f32 %f30738, %f6947; $L__BB107_4250: mul.rn.ftz.f32 %f26488, %f30738, %f30738; mov.f32 %f26489, 0fC0B59883; mov.f32 %f26490, 0fBF52C7EA; fma.rn.ftz.f32 %f26491, %f26488, %f26490, %f26489; mov.f32 %f26492, 0fC0D21907; fma.rn.ftz.f32 %f26493, %f26491, %f26488, %f26492; mul.ftz.f32 %f26494, %f26488, %f26493; mul.ftz.f32 %f26495, %f30738, %f26494; add.ftz.f32 %f26496, %f26488, 0f41355DC0; mov.f32 %f26497, 0f41E6BD60; fma.rn.ftz.f32 %f26498, %f26496, %f26488, %f26497; mov.f32 %f26499, 0f419D92C8; fma.rn.ftz.f32 %f26500, %f26498, %f26488, %f26499; rcp.approx.ftz.f32 %f26501, %f26500; fma.rn.ftz.f32 %f26502, %f26495, %f26501, %f30738; mov.f32 %f26503, 0f3FC90FDB; sub.ftz.f32 %f26504, %f26503, %f26502; setp.gt.ftz.f32 %p3629, %f6947, 0f3F800000; selp.f32 %f26505, %f26504, %f26502, %p3629; mov.b32 %r5303, %f26505; mov.b32 %r5304, %f6946; and.b32 %r5305, %r5304, -2147483648; or.b32 %r5306, %r5305, %r5303; mov.b32 %f26506, %r5306; setp.le.ftz.f32 %p3630, %f6947, 0f7F800000; selp.f32 %f26507, %f26506, %f26505, %p3630; cos.approx.ftz.f32 %f26508, %f26507; sin.approx.ftz.f32 %f26509, %f26507; mul.ftz.f32 %f26510, %f6918, %f26508; mul.ftz.f32 %f26511, %f6919, %f26509; mul.ftz.f32 %f26512, %f26509, %f26511; fma.rn.ftz.f32 %f26513, %f26508, %f26510, %f26512; add.ftz.f32 %f26514, %f26513, 0f3F800000; rcp.approx.ftz.f32 %f26515, %f26514; lg2.approx.ftz.f32 %f26516, %f6202; mul.ftz.f32 %f26517, %f26516, %f26515; ex2.approx.ftz.f32 %f30741, %f26517; add.ftz.f32 %f30742, %f26507, 0f40490FDB; bra.uni $L__BB107_4258; $L__BB107_4393: mul.ftz.f32 %f27860, %f7367, %f7367; rcp.approx.ftz.f32 %f27861, %f27860; add.ftz.f32 %f27862, %f27861, 0fBF800000; mul.ftz.f32 %f27863, %f27860, %f27860; mul.ftz.f32 %f27864, %f27863, 0f40490FDB; add.ftz.f32 %f27865, %f30869, %f27862; mul.ftz.f32 %f27866, %f27864, %f27865; mul.ftz.f32 %f27867, %f27865, %f27866; div.approx.ftz.f32 %f30870, %f30869, %f27867; $L__BB107_4397: mul.ftz.f32 %f27888, %f7392, %f7392; mov.f32 %f27889, 0f3F800000; sub.ftz.f32 %f27890, %f27889, %f27888; mul.ftz.f32 %f27891, %f27890, %f30869; div.approx.ftz.f32 %f30873, %f27891, %f27888; $L__BB107_4398: mul.ftz.f32 %f27895, %f7364, %f30870; mul.ftz.f32 %f27896, %f27895, 0f3E800000; div.approx.ftz.f32 %f30902, %f27896, %f7329; add.s32 %r5362, %r1522, -11; setp.gt.u32 %p3747, %r5362, 1; mov.f32 %f30876, 0f3F800000; mov.f32 %f30877, %f30876; mov.f32 %f30878, %f30876; @%p3747 bra $L__BB107_4404; ld.local.f32 %f27898, [%rd1027+28]; fma.rn.ftz.f32 %f7407, %f27898, %f27898, 0fBF800000; mov.f32 %f27897, 0f3F800000; abs.ftz.f32 %f7408, %f27897; fma.rn.ftz.f32 %f7409, %f7408, %f7408, %f7407; setp.leu.ftz.f32 %p3748, %f7409, 0f00000000; mov.f32 %f30874, %f27897; @%p3748 bra $L__BB107_4401; sqrt.approx.ftz.f32 %f27899, %f7409; sub.ftz.f32 %f27900, %f27899, %f7408; add.ftz.f32 %f27901, %f7408, %f27899; div.approx.ftz.f32 %f27902, %f27900, %f27901; fma.rn.ftz.f32 %f27903, %f7408, %f27901, 0fBF800000; fma.rn.ftz.f32 %f27904, %f7408, %f27900, 0f3F800000; div.approx.ftz.f32 %f27905, %f27903, %f27904; mul.ftz.f32 %f27906, %f27902, 0f3F000000; mul.ftz.f32 %f27907, %f27902, %f27906; fma.rn.ftz.f32 %f27908, %f27905, %f27905, 0f3F800000; mul.ftz.f32 %f30874, %f27907, %f27908; $L__BB107_4401: ld.local.u64 %rd4470, [%rd1027+36]; ld.f32 %f7412, [%rd4470+12]; ld.f32 %f7413, [%rd4470+16]; ld.f32 %f7414, [%rd4470+20]; sub.ftz.f32 %f7415, %f27897, %f30874; mul.ftz.f32 %f27910, %f7369, %f30900; fma.rn.ftz.f32 %f27911, %f7368, %f30899, %f27910; fma.rn.ftz.f32 %f27912, %f7370, %f30901, %f27911; abs.ftz.f32 %f7416, %f27912; fma.rn.ftz.f32 %f7417, %f7416, %f7416, %f7407; setp.leu.ftz.f32 %p3749, %f7417, 0f00000000; mov.f32 %f30875, %f7415; @%p3749 bra $L__BB107_4403; sqrt.approx.ftz.f32 %f27913, %f7417; sub.ftz.f32 %f27914, %f27913, %f7416; add.ftz.f32 %f27915, %f7416, %f27913; div.approx.ftz.f32 %f27916, %f27914, %f27915; fma.rn.ftz.f32 %f27917, %f7416, %f27915, 0fBF800000; fma.rn.ftz.f32 %f27918, %f7416, %f27914, 0f3F800000; div.approx.ftz.f32 %f27919, %f27917, %f27918; mul.ftz.f32 %f27920, %f27916, 0f3F000000; mul.ftz.f32 %f27921, %f27916, %f27920; fma.rn.ftz.f32 %f27922, %f27919, %f27919, 0f3F800000; mul.ftz.f32 %f27923, %f27921, %f27922; sub.ftz.f32 %f30875, %f27923, %f30874; $L__BB107_4403: rcp.approx.ftz.f32 %f27924, %f7415; mov.f32 %f27925, 0f3F800000; mul.ftz.f32 %f27926, %f27924, %f30875; sub.ftz.f32 %f27927, %f27925, %f27926; fma.rn.ftz.f32 %f30878, %f7414, %f27927, %f27926; fma.rn.ftz.f32 %f30877, %f7413, %f27927, %f27926; fma.rn.ftz.f32 %f30876, %f7412, %f27927, %f27926; $L__BB107_4404: add.ftz.f32 %f27928, %f30873, 0f3F800000; mov.f32 %f27929, 0f00000000; max.ftz.f32 %f27930, %f27928, %f27929; sqrt.approx.ftz.f32 %f27931, %f27930; add.ftz.f32 %f27932, %f27931, 0f3F800000; mov.f32 %f27933, 0f40000000; div.approx.ftz.f32 %f27934, %f27933, %f27932; mul.ftz.f32 %f27935, %f27934, %f30902; mul.ftz.f32 %f30903, %f27935, %f30876; mul.ftz.f32 %f30904, %f27935, %f30877; mul.ftz.f32 %f30905, %f27935, %f30878; mov.u32 %r5877, 10; $L__BB107_4405: setp.ne.s32 %p3750, %r1522, 12; @%p3750 bra $L__BB107_4433; ld.local.u64 %rd4471, [%rd1027+36]; ld.f32 %f27936, [%rd4471+36]; mul.ftz.f32 %f27937, %f27936, 0f3E800000; mul.ftz.f32 %f30903, %f30903, %f27937; mul.ftz.f32 %f30904, %f30904, %f27937; mul.ftz.f32 %f30905, %f30905, %f27937; bra.uni $L__BB107_4433; $L__BB107_4354: mul.ftz.f32 %f27461, %f30789, %f7274; fma.rn.ftz.f32 %f27462, %f30788, %f7273, %f27461; fma.rn.ftz.f32 %f27463, %f30790, %f7275, %f27462; mul.ftz.f32 %f27464, %f30786, %f7274; fma.rn.ftz.f32 %f27465, %f30785, %f7273, %f27464; fma.rn.ftz.f32 %f27466, %f30787, %f7275, %f27465; mul.ftz.f32 %f27467, %f7165, %f7274; fma.rn.ftz.f32 %f27468, %f7163, %f7273, %f27467; fma.rn.ftz.f32 %f27469, %f7166, %f7275, %f27468; neg.ftz.f32 %f27470, %f27463; mul.ftz.f32 %f27471, %f27069, %f27469; div.approx.ftz.f32 %f27472, %f27470, %f27471; neg.ftz.f32 %f27473, %f27466; mul.ftz.f32 %f27474, %f27070, %f27469; div.approx.ftz.f32 %f27475, %f27473, %f27474; mul.ftz.f32 %f27476, %f27469, %f27469; mul.ftz.f32 %f27477, %f27476, %f27476; mul.ftz.f32 %f27478, %f27472, %f27472; neg.ftz.f32 %f27479, %f27478; mul.ftz.f32 %f27480, %f27475, %f27475; sub.ftz.f32 %f27481, %f27479, %f27480; mul.ftz.f32 %f27482, %f27481, 0f3FB8AA3B; ex2.approx.ftz.f32 %f27483, %f27482; mul.ftz.f32 %f27484, %f7281, %f27477; div.approx.ftz.f32 %f30840, %f27483, %f27484; mul.ftz.f32 %f27485, %f30789, %f30900; fma.rn.ftz.f32 %f27486, %f30788, %f30899, %f27485; fma.rn.ftz.f32 %f27487, %f30790, %f30901, %f27486; mul.ftz.f32 %f27488, %f30786, %f30900; fma.rn.ftz.f32 %f27489, %f30785, %f30899, %f27488; fma.rn.ftz.f32 %f27490, %f30787, %f30901, %f27489; mul.ftz.f32 %f27491, %f27490, %f27490; mul.ftz.f32 %f27492, %f27487, %f27487; mul.ftz.f32 %f27493, %f27069, %f27069; mul.ftz.f32 %f27494, %f27070, %f27070; mul.ftz.f32 %f27495, %f27494, %f27491; fma.rn.ftz.f32 %f27496, %f27493, %f27492, %f27495; add.ftz.f32 %f27497, %f27492, %f27491; div.approx.ftz.f32 %f27498, %f27496, %f27497; mul.ftz.f32 %f27499, %f7283, %f27498; div.approx.ftz.f32 %f27500, %f27499, %f7282; mov.f32 %f27501, 0f00000000; max.ftz.f32 %f27502, %f27500, %f27501; sqrt.approx.ftz.f32 %f7285, %f27502; setp.lt.ftz.f32 %p3715, %f7285, 0f3F200000; @%p3715 bra $L__BB107_4358; rcp.approx.ftz.f32 %f27503, %f7285; fma.rn.ftz.f32 %f27504, %f27503, 0f400B9581, 0f40623D71; mul.ftz.f32 %f27505, %f27503, %f27504; fma.rn.ftz.f32 %f27506, %f27503, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f27507, %f27503, %f27506, 0f3F800000; div.approx.ftz.f32 %f30841, %f27505, %f27507; $L__BB107_4358: mul.ftz.f32 %f27527, %f30840, 0f3E800000; div.approx.ftz.f32 %f27528, %f27527, %f7167; mul.ftz.f32 %f27529, %f30837, %f30841; mul.ftz.f32 %f30903, %f27529, %f27528; mul.ftz.f32 %f30902, %f30837, %f27528; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; mov.u32 %r5877, %r5349; bra.uni $L__BB107_4433; $L__BB107_4245: mov.f32 %f26444, 0f3F800000; sub.ftz.f32 %f26445, %f26444, %f6226; mul.ftz.f32 %f26446, %f26445, 0f40800000; mul.ftz.f32 %f26447, %f26446, 0f3FC90FDB; sin.approx.ftz.f32 %f26448, %f26447; cos.approx.ftz.f32 %f26449, %f26447; div.approx.ftz.f32 %f26450, %f26448, %f26449; mul.ftz.f32 %f6940, %f6939, %f26450; abs.ftz.f32 %f6941, %f6940; setp.leu.ftz.f32 %p3625, %f6941, 0f3F800000; mov.f32 %f30737, %f6941; @%p3625 bra $L__BB107_4247; rcp.approx.ftz.f32 %f30737, %f6941; $L__BB107_4247: mul.rn.ftz.f32 %f26451, %f30737, %f30737; mov.f32 %f26452, 0fC0B59883; mov.f32 %f26453, 0fBF52C7EA; fma.rn.ftz.f32 %f26454, %f26451, %f26453, %f26452; mov.f32 %f26455, 0fC0D21907; fma.rn.ftz.f32 %f26456, %f26454, %f26451, %f26455; mul.ftz.f32 %f26457, %f26451, %f26456; mul.ftz.f32 %f26458, %f30737, %f26457; add.ftz.f32 %f26459, %f26451, 0f41355DC0; mov.f32 %f26460, 0f41E6BD60; fma.rn.ftz.f32 %f26461, %f26459, %f26451, %f26460; mov.f32 %f26462, 0f419D92C8; fma.rn.ftz.f32 %f26463, %f26461, %f26451, %f26462; rcp.approx.ftz.f32 %f26464, %f26463; fma.rn.ftz.f32 %f26465, %f26458, %f26464, %f30737; mov.f32 %f26466, 0f3FC90FDB; sub.ftz.f32 %f26467, %f26466, %f26465; setp.gt.ftz.f32 %p3626, %f6941, 0f3F800000; selp.f32 %f26468, %f26467, %f26465, %p3626; mov.b32 %r5299, %f26468; mov.b32 %r5300, %f6940; and.b32 %r5301, %r5300, -2147483648; or.b32 %r5302, %r5301, %r5299; mov.b32 %f26469, %r5302; setp.le.ftz.f32 %p3627, %f6941, 0f7F800000; selp.f32 %f26470, %f26469, %f26468, %p3627; cos.approx.ftz.f32 %f26471, %f26470; sin.approx.ftz.f32 %f26472, %f26470; mul.ftz.f32 %f26473, %f6918, %f26471; mul.ftz.f32 %f26474, %f6919, %f26472; mul.ftz.f32 %f26475, %f26472, %f26474; fma.rn.ftz.f32 %f26476, %f26471, %f26473, %f26475; add.ftz.f32 %f26477, %f26476, 0f3F800000; rcp.approx.ftz.f32 %f26478, %f26477; lg2.approx.ftz.f32 %f26479, %f6202; mul.ftz.f32 %f26480, %f26479, %f26478; ex2.approx.ftz.f32 %f30741, %f26480; mov.f32 %f26481, 0f40C90FDB; sub.ftz.f32 %f30742, %f26481, %f26470; $L__BB107_4258: mul.ftz.f32 %f26597, %f30741, %f30741; mov.f32 %f30743, 0f3F800000; sub.ftz.f32 %f26599, %f30743, %f26597; mov.f32 %f30746, 0f00000000; max.ftz.f32 %f26601, %f30746, %f26599; sqrt.approx.ftz.f32 %f26602, %f26601; cos.approx.ftz.f32 %f26603, %f30742; mul.ftz.f32 %f26604, %f26602, %f26603; sin.approx.ftz.f32 %f26605, %f30742; mul.ftz.f32 %f26606, %f26602, %f26605; mul.ftz.f32 %f26607, %f30731, %f26606; mul.ftz.f32 %f26608, %f30732, %f26606; mul.ftz.f32 %f26609, %f30733, %f26606; fma.rn.ftz.f32 %f26610, %f30734, %f26604, %f26607; fma.rn.ftz.f32 %f26611, %f30735, %f26604, %f26608; fma.rn.ftz.f32 %f26612, %f30736, %f26604, %f26609; fma.rn.ftz.f32 %f26613, %f6911, %f30741, %f26610; fma.rn.ftz.f32 %f26614, %f6913, %f30741, %f26611; fma.rn.ftz.f32 %f26615, %f6914, %f30741, %f26612; mul.ftz.f32 %f26616, %f26392, %f26614; fma.rn.ftz.f32 %f26617, %f6912, %f26613, %f26616; fma.rn.ftz.f32 %f26618, %f26393, %f26615, %f26617; setp.lt.ftz.f32 %p3637, %f26618, 0f00000000; neg.ftz.f32 %f26619, %f26613; neg.ftz.f32 %f26620, %f26614; neg.ftz.f32 %f26621, %f26615; selp.f32 %f26622, %f26619, %f26613, %p3637; selp.f32 %f26623, %f26620, %f26614, %p3637; selp.f32 %f26624, %f26621, %f26615, %p3637; add.ftz.f32 %f26625, %f26618, %f26618; mul.ftz.f32 %f26626, %f26625, %f26622; mul.ftz.f32 %f26627, %f26625, %f26623; mul.ftz.f32 %f26628, %f26625, %f26624; sub.ftz.f32 %f30899, %f26626, %f6912; sub.ftz.f32 %f30900, %f26627, %f26392; sub.ftz.f32 %f30901, %f26628, %f26393; max.ftz.f32 %f26629, %f26396, %f26397; setp.le.ftz.f32 %p3638, %f26629, 0f38D1B717; mov.u32 %r5877, 18; mov.f32 %f30902, 0f49742400; mov.f32 %f30903, %f30902; mov.f32 %f30904, %f30902; mov.f32 %f30905, %f30902; @%p3638 bra $L__BB107_4433; mul.ftz.f32 %f26634, %f6913, %f30900; fma.rn.ftz.f32 %f26635, %f6911, %f30899, %f26634; fma.rn.ftz.f32 %f6971, %f6914, %f30901, %f26635; setp.leu.ftz.f32 %p3639, %f6971, 0f00000000; mov.f32 %f30903, %f30746; @%p3639 bra $L__BB107_4266; add.ftz.f32 %f26636, %f6912, %f30899; add.ftz.f32 %f26637, %f26392, %f30900; mul.ftz.f32 %f26638, %f26637, %f26637; fma.rn.ftz.f32 %f26639, %f26636, %f26636, %f26638; add.ftz.f32 %f26640, %f26393, %f30901; fma.rn.ftz.f32 %f26641, %f26640, %f26640, %f26639; sqrt.approx.ftz.f32 %f26642, %f26641; div.approx.ftz.f32 %f6972, %f26636, %f26642; div.approx.ftz.f32 %f6973, %f26637, %f26642; div.approx.ftz.f32 %f6974, %f26640, %f26642; mul.ftz.f32 %f26643, %f26392, %f6973; fma.rn.ftz.f32 %f26644, %f6912, %f6972, %f26643; fma.rn.ftz.f32 %f26645, %f26393, %f6974, %f26644; abs.ftz.f32 %f26646, %f26645; mov.f32 %f26647, 0f358637BD; max.ftz.f32 %f6975, %f26646, %f26647; mul.ftz.f32 %f26648, %f6913, %f6973; fma.rn.ftz.f32 %f26649, %f6911, %f6972, %f26648; fma.rn.ftz.f32 %f26650, %f6914, %f6974, %f26649; max.ftz.f32 %f6976, %f26650, %f26647; max.ftz.f32 %f6977, %f6971, %f26647; @%p3617 bra $L__BB107_4264; bra.uni $L__BB107_4261; $L__BB107_4264: lg2.approx.ftz.f32 %f26701, %f6976; mul.ftz.f32 %f26702, %f6918, %f26701; ex2.approx.ftz.f32 %f26703, %f26702; mov.f32 %f26704, 0f41C90FDB; div.approx.ftz.f32 %f26705, %f6938, %f26704; mul.ftz.f32 %f26706, %f6977, %f26705; mul.ftz.f32 %f30745, %f26703, %f26706; mul.ftz.f32 %f26707, %f26703, %f26705; div.approx.ftz.f32 %f30746, %f26707, %f6975; bra.uni $L__BB107_4265; $L__BB107_4261: ld.local.f32 %f26652, [%rd1027+52]; mul.ftz.f32 %f26653, %f6913, %f26652; ld.local.f32 %f26654, [%rd1027+48]; mul.ftz.f32 %f26655, %f6914, %f26654; sub.ftz.f32 %f26656, %f26653, %f26655; ld.local.f32 %f26657, [%rd1027+44]; mul.ftz.f32 %f26658, %f6914, %f26657; mul.ftz.f32 %f26659, %f6911, %f26652; sub.ftz.f32 %f26660, %f26658, %f26659; mul.ftz.f32 %f26661, %f6911, %f26654; mul.ftz.f32 %f26662, %f6913, %f26657; sub.ftz.f32 %f26663, %f26661, %f26662; mul.ftz.f32 %f26664, %f26660, %f26660; fma.rn.ftz.f32 %f26665, %f26656, %f26656, %f26664; fma.rn.ftz.f32 %f26666, %f26663, %f26663, %f26665; sqrt.approx.ftz.f32 %f26667, %f26666; div.approx.ftz.f32 %f26668, %f26656, %f26667; div.approx.ftz.f32 %f26669, %f26660, %f26667; div.approx.ftz.f32 %f26670, %f26663, %f26667; mul.ftz.f32 %f26671, %f6914, %f26669; mul.ftz.f32 %f26672, %f6913, %f26670; sub.ftz.f32 %f26673, %f26671, %f26672; mul.ftz.f32 %f26674, %f6911, %f26670; mul.ftz.f32 %f26675, %f6914, %f26668; sub.ftz.f32 %f26676, %f26674, %f26675; mul.ftz.f32 %f26677, %f6913, %f26668; mul.ftz.f32 %f26678, %f6911, %f26669; sub.ftz.f32 %f26679, %f26677, %f26678; mul.ftz.f32 %f26680, %f6973, %f26676; fma.rn.ftz.f32 %f26681, %f6972, %f26673, %f26680; fma.rn.ftz.f32 %f6978, %f6974, %f26679, %f26681; mul.ftz.f32 %f26682, %f6973, %f26669; fma.rn.ftz.f32 %f26683, %f6972, %f26668, %f26682; fma.rn.ftz.f32 %f6979, %f6974, %f26670, %f26683; setp.geu.ftz.f32 %p3641, %f6976, 0f3F800000; @%p3641 bra $L__BB107_4263; mul.ftz.f32 %f26684, %f6918, %f6978; mul.ftz.f32 %f26685, %f6978, %f26684; mul.ftz.f32 %f26686, %f6919, %f6979; fma.rn.ftz.f32 %f26687, %f6979, %f26686, %f26685; mul.ftz.f32 %f26688, %f6976, %f6976; mov.f32 %f26689, 0f3F800000; sub.ftz.f32 %f26690, %f26689, %f26688; div.approx.ftz.f32 %f26691, %f26687, %f26690; lg2.approx.ftz.f32 %f26692, %f6976; mul.ftz.f32 %f26693, %f26691, %f26692; ex2.approx.ftz.f32 %f30743, %f26693; $L__BB107_4263: add.ftz.f32 %f26694, %f6919, 0f3F800000; mul.ftz.f32 %f26695, %f6938, %f26694; sqrt.approx.ftz.f32 %f26696, %f26695; mov.f32 %f26697, 0f41C90FDB; div.approx.ftz.f32 %f26698, %f26696, %f26697; mul.ftz.f32 %f26699, %f6977, %f26698; mul.ftz.f32 %f30745, %f30743, %f26699; mul.ftz.f32 %f26700, %f30743, %f26698; div.approx.ftz.f32 %f30746, %f26700, %f6975; $L__BB107_4265: max.ftz.f32 %f26709, %f6915, %f26647; max.ftz.f32 %f26710, %f6977, %f26709; mul.ftz.f32 %f26711, %f6975, %f26710; max.ftz.f32 %f26712, %f26647, %f26711; rcp.approx.ftz.f32 %f26713, %f26712; mul.ftz.f32 %f30903, %f26713, %f30745; $L__BB107_4266: mov.u32 %r5877, 10; mov.f32 %f30902, %f30746; mov.f32 %f30904, %f30903; mov.f32 %f30905, %f30903; $L__BB107_4433: and.b32 %r5387, %r5877, 1; setp.eq.b32 %p3797, %r5387, 1; mov.pred %p3798, 0; xor.pred %p3799, %p3797, %p3798; not.pred %p3800, %p3799; @%p3800 bra $L__BB107_4441; bra.uni $L__BB107_4434; $L__BB107_4441: ld.const.u64 %rd4472, [kernel_params+2160]; cvta.to.global.u64 %rd4473, %rd4472; ld.local.u32 %r5397, [%rd5+-12]; mul.wide.s32 %rd4474, %r5397, 208; add.s64 %rd4475, %rd4473, %rd4474; ld.global.f32 %f7564, [%rd4475+180]; setp.leu.ftz.f32 %p3811, %f7564, 0f3F800000; @%p3811 bra $L__BB107_4443; ld.local.f32 %f28458, [%rd1027+8]; ld.local.f32 %f28459, [%rd1027+12]; mul.ftz.f32 %f28460, %f30900, %f28459; fma.rn.ftz.f32 %f28461, %f30899, %f28458, %f28460; ld.local.f32 %f28462, [%rd1027+16]; fma.rn.ftz.f32 %f28463, %f30901, %f28462, %f28461; mov.f32 %f28464, 0f3F800000; min.ftz.f32 %f28465, %f28463, %f28464; abs.ftz.f32 %f28466, %f28465; setp.lt.ftz.f32 %p3812, %f28466, 0f3F800000; add.ftz.f32 %f28467, %f28466, 0fBF800000; add.ftz.f32 %f28468, %f28467, 0f3F800000; selp.f32 %f28469, %f28468, 0f3F800000, %p3812; sub.ftz.f32 %f28470, %f28464, %f28469; sqrt.approx.ftz.f32 %f28471, %f28470; fma.rn.ftz.f32 %f28472, %f28469, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f28473, %f28469, %f28472, 0fBE5A6B93; fma.rn.ftz.f32 %f28474, %f28469, %f28473, 0f3FC90FDB; mul.ftz.f32 %f28475, %f28471, %f28474; setp.lt.ftz.f32 %p3813, %f28465, 0f00000000; mov.f32 %f28476, 0f00000000; mov.f32 %f28477, 0f40490FDB; sub.ftz.f32 %f28478, %f28477, %f28475; selp.f32 %f28479, %f28478, %f28475, %p3813; mul.ftz.f32 %f28480, %f7564, %f28479; cos.approx.ftz.f32 %f28481, %f28480; max.ftz.f32 %f28482, %f28481, %f28476; div.approx.ftz.f32 %f28483, %f28482, %f28465; mul.ftz.f32 %f30903, %f30903, %f28483; mul.ftz.f32 %f30904, %f30904, %f28483; mul.ftz.f32 %f30905, %f30905, %f28483; $L__BB107_4443: and.b32 %r5398, %r5877, 4; setp.eq.s32 %p3814, %r5398, 0; @%p3814 bra $L__BB107_4454; ld.local.f32 %f7571, [%rd1027+16]; ld.local.f32 %f7572, [%rd5+-68]; ld.local.f32 %f7573, [%rd5+-76]; ld.local.f32 %f7574, [%rd1027+8]; setp.neu.ftz.f32 %p3815, %f7574, %f7573; mov.u16 %rs747, 0; @%p3815 bra $L__BB107_4447; ld.local.f32 %f28484, [%rd1027+12]; ld.local.f32 %f28485, [%rd5+-72]; setp.neu.ftz.f32 %p3816, %f28484, %f28485; @%p3816 bra $L__BB107_4447; setp.eq.ftz.f32 %p3817, %f7571, %f7572; selp.u16 %rs747, 1, 0, %p3817; $L__BB107_4447: setp.ne.s16 %p3818, %rs747, 0; @%p3818 bra $L__BB107_4454; ld.local.f32 %f28487, [%rd1027+12]; mul.ftz.f32 %f28488, %f30900, %f28487; fma.rn.ftz.f32 %f28489, %f30899, %f7574, %f28488; fma.rn.ftz.f32 %f28490, %f30901, %f7571, %f28489; setp.lt.ftz.f32 %p3819, %f28490, 0f00000000; mov.f32 %f30910, 0f00000000; ld.local.f32 %f28491, [%rd5+-72]; neg.ftz.f32 %f28492, %f28491; neg.ftz.f32 %f28493, %f7573; selp.f32 %f28494, %f28493, %f7573, %p3819; selp.f32 %f28495, %f28492, %f28491, %p3819; neg.ftz.f32 %f28496, %f7572; selp.f32 %f28497, %f28496, %f7572, %p3819; mul.ftz.f32 %f28498, %f30900, %f28495; fma.rn.ftz.f32 %f28499, %f30899, %f28494, %f28498; fma.rn.ftz.f32 %f7575, %f30901, %f28497, %f28499; mul.ftz.f32 %f28500, %f28487, %f28495; fma.rn.ftz.f32 %f28501, %f7574, %f28494, %f28500; fma.rn.ftz.f32 %f28502, %f7571, %f28497, %f28501; mul.ftz.f32 %f7576, %f28490, %f28502; setp.eq.ftz.f32 %p3820, %f7576, 0f00000000; @%p3820 bra $L__BB107_4450; div.approx.ftz.f32 %f30910, %f7575, %f7576; $L__BB107_4450: setp.ge.ftz.f32 %p3821, %f30910, 0f3F800000; mov.f32 %f30911, 0f3F800000; @%p3821 bra $L__BB107_4453; setp.lt.ftz.f32 %p3822, %f30910, 0f00000000; mov.f32 %f30911, 0f00000000; @%p3822 bra $L__BB107_4453; mul.ftz.f32 %f28505, %f30910, %f30910; mul.ftz.f32 %f28506, %f30910, %f28505; sub.ftz.f32 %f28507, %f28505, %f28506; add.ftz.f32 %f30911, %f30910, %f28507; $L__BB107_4453: mul.ftz.f32 %f30903, %f30903, %f30911; mul.ftz.f32 %f30904, %f30904, %f30911; mul.ftz.f32 %f30905, %f30905, %f30911; bra.uni $L__BB107_4454; $L__BB107_4434: ld.const.f32 %f7561, [kernel_params+1040]; setp.ltu.ftz.f32 %p3801, %f7561, 0f00000000; shr.u32 %r5388, %r5877, 2; and.b32 %r5389, %r5388, 1; setp.eq.b32 %p3802, %r5389, 1; or.pred %p3803, %p3802, %p3801; @%p3803 bra $L__BB107_4454; mov.f32 %f30906, 0f00000000; setp.gt.u32 %p3804, %r1522, 31; @%p3804 bra $L__BB107_4437; mov.u32 %r5391, 1; shl.b32 %r5392, %r5391, %r1522; and.b32 %r5393, %r5392, -2145385984; setp.ne.s32 %p3805, %r5393, 0; @%p3805 bra $L__BB107_4440; bra.uni $L__BB107_4437; $L__BB107_4440: or.b32 %r5396, %r5877, 128; setp.le.ftz.f32 %p3810, %f30906, %f7561; selp.b32 %r5877, %r5396, %r5877, %p3810; $L__BB107_4454: add.s32 %r1556, %r1522, -2; setp.neu.ftz.f32 %p3823, %f30902, 0f00000000; @%p3823 bra $L__BB107_4460; bra.uni $L__BB107_4455; $L__BB107_4460: setp.lt.u32 %p3828, %r1556, 7; ld.local.f32 %f28508, [%rd1027+-12]; mul.ftz.f32 %f7587, %f30903, %f28508; ld.local.f32 %f28509, [%rd1027+-8]; mul.ftz.f32 %f7588, %f30904, %f28509; ld.local.f32 %f28510, [%rd1027+-4]; mul.ftz.f32 %f7589, %f30905, %f28510; mov.u32 %r5404, 0; st.local.u32 [%rd2157], %r5404; st.local.u32 [%rd2157+4], %r5404; st.local.u32 [%rd2157+8], %r5404; st.local.u32 [%rd2157+12], %r5404; st.local.u32 [%rd2157+16], %r5404; st.local.u32 [%rd2157+20], %r5404; @%p3828 bra $L__BB107_4463; bra.uni $L__BB107_4461; $L__BB107_4463: st.local.f32 [%rd2157], %f7587; st.local.f32 [%rd2157+4], %f7588; st.local.f32 [%rd2157+8], %f7589; bra.uni $L__BB107_4464; $L__BB107_4455: setp.lt.u32 %p3824, %r1556, 7; mov.u32 %r5399, 0; st.local.u32 [%rd2157], %r5399; st.local.u32 [%rd2157+4], %r5399; st.local.u32 [%rd2157+8], %r5399; st.local.u32 [%rd2157+12], %r5399; st.local.u32 [%rd2157+16], %r5399; st.local.u32 [%rd2157+20], %r5399; @%p3824 bra $L__BB107_4458; bra.uni $L__BB107_4456; $L__BB107_4458: st.local.u32 [%rd2157], %r5399; st.local.u32 [%rd2157+4], %r5399; st.local.u32 [%rd2157+8], %r5399; bra.uni $L__BB107_4459; $L__BB107_4461: add.s32 %r5405, %r1522, -9; setp.gt.u32 %p3829, %r5405, 11; setp.ne.s32 %p3830, %r1522, 29; and.pred %p3831, %p3830, %p3829; @%p3831 bra $L__BB107_4464; st.local.f32 [%rd2157+12], %f7587; st.local.f32 [%rd2157+16], %f7588; st.local.f32 [%rd2157+20], %f7589; $L__BB107_4464: st.local.f32 [%rd2157+24], %f7587; st.local.f32 [%rd2157+28], %f7588; st.local.f32 [%rd2157+32], %f7589; @%p3322 bra $L__BB107_4466; add.u64 %rd4723, %SP, 3424; ld.local.f32 %f28511, [%rd1027+4]; mul.ftz.f32 %f28512, %f30902, %f28511; mul.lo.s64 %rd4477, %rd1026, 80; add.s64 %rd4479, %rd4723, %rd4477; add.s64 %rd4480, %rd4479, 304; { // callseq 691, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd4723; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f30899; st.param.f32 [param1+4], %f30900; st.param.f32 [param1+8], %f30901; .param .b64 param2; st.param.b64 [param2+0], %rd4480; .param .b64 param3; st.param.b64 [param3+0], %rd2156; .param .b32 param4; st.param.f32 [param4+0], %f28512; .param .b32 param5; st.param.f32 [param5+0], %f28511; .param .b32 retval0; call.uni (retval0), _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj, ( param0, param1, param2, param3, param4, param5 ); ld.param.f32 %f30902, [retval0+0]; } // callseq 691 bra.uni $L__BB107_4466; $L__BB107_4456: add.s32 %r5400, %r1522, -9; setp.gt.u32 %p3825, %r5400, 11; setp.ne.s32 %p3826, %r1522, 29; and.pred %p3827, %p3826, %p3825; @%p3827 bra $L__BB107_4459; st.local.u32 [%rd2157+12], %r5399; st.local.u32 [%rd2157+16], %r5399; st.local.u32 [%rd2157+20], %r5399; $L__BB107_4459: st.local.u32 [%rd2157+24], %r5399; st.local.u32 [%rd2157+28], %r5399; st.local.u32 [%rd2157+32], %r5399; $L__BB107_4466: setp.eq.ftz.f32 %p3833, %f30902, 0f00000000; mov.u32 %r5886, 0; @%p3833 bra $L__BB107_4529; ld.local.f32 %f7592, [%rd2157+24]; setp.eq.ftz.f32 %p3834, %f7592, 0f00000000; ld.local.f32 %f7593, [%rd2157+28]; setp.eq.ftz.f32 %p3835, %f7593, 0f00000000; and.pred %p3836, %p3834, %p3835; ld.local.f32 %f7594, [%rd2157+32]; setp.eq.ftz.f32 %p3837, %f7594, 0f00000000; and.pred %p3838, %p3837, %p3836; @%p3838 bra $L__BB107_4529; and.b32 %r1557, %r5877, 32; setp.eq.s32 %p3839, %r1557, 0; @%p3839 bra $L__BB107_4470; bra.uni $L__BB107_4469; $L__BB107_4470: mul.ftz.f32 %f28516, %f30900, %f30900; fma.rn.ftz.f32 %f28517, %f30899, %f30899, %f28516; fma.rn.ftz.f32 %f28518, %f30901, %f30901, %f28517; sqrt.approx.ftz.f32 %f28519, %f28518; div.approx.ftz.f32 %f7595, %f30899, %f28519; div.approx.ftz.f32 %f7596, %f30900, %f28519; div.approx.ftz.f32 %f7597, %f30901, %f28519; add.u64 %rd4487, %SPL, 3424; ld.local.v4.f32 {%f30934, %f30935, %f30936, %f28523}, [%rd4487]; ld.local.u32 %r5410, [%rd5+-24]; and.b32 %r5411, %r5410, 1; setp.eq.b32 %p3841, %r5411, 1; xor.pred %p3843, %p3841, %p3798; not.pred %p3844, %p3843; @%p3844 bra $L__BB107_4478; ld.const.u64 %rd4488, [kernel_params+2232]; cvta.to.global.u64 %rd4489, %rd4488; ld.local.v2.u32 {%r5412, %r5413}, [%rd5+-32]; mul.wide.s32 %rd4490, %r5413, 16; add.s64 %rd4491, %rd4489, %rd4490; ld.const.u64 %rd4492, [kernel_params+2256]; cvta.to.global.u64 %rd4493, %rd4492; ld.global.u32 %r5415, [%rd4491+12]; mul.wide.u32 %rd4494, %r5415, 12; add.s64 %rd4495, %rd4493, %rd4494; ld.global.f32 %f7604, [%rd4495]; ld.global.f32 %f7605, [%rd4495+4]; ld.global.f32 %f7606, [%rd4495+8]; add.s32 %r5416, %r5415, 1; mul.wide.u32 %rd4496, %r5416, 12; add.s64 %rd4497, %rd4493, %rd4496; ld.global.f32 %f7607, [%rd4497]; ld.global.f32 %f7608, [%rd4497+4]; ld.global.f32 %f7609, [%rd4497+8]; add.s32 %r5417, %r5415, 2; mul.wide.u32 %rd4498, %r5417, 12; add.s64 %rd4499, %rd4493, %rd4498; ld.global.f32 %f7610, [%rd4499]; ld.global.f32 %f7611, [%rd4499+4]; ld.global.f32 %f7612, [%rd4499+8]; and.b32 %r5418, %r5412, 4; setp.ne.s32 %p3845, %r5418, 0; mov.f32 %f30928, %f30934; mov.f32 %f30929, %f30935; mov.f32 %f30930, %f30936; mov.f32 %f30931, %f7595; mov.f32 %f30932, %f7596; mov.f32 %f30933, %f7597; @%p3845 bra $L__BB107_4476; and.b32 %r5419, %r5412, 2; setp.eq.s32 %p3846, %r5419, 0; @%p3846 bra $L__BB107_4474; bra.uni $L__BB107_4473; $L__BB107_4474: ld.const.u64 %rd4500, [kernel_params+2160]; cvta.to.global.u64 %rd4501, %rd4500; ld.local.u32 %r5420, [%rd5+-12]; mul.wide.s32 %rd4502, %r5420, 208; add.s64 %rd4503, %rd4501, %rd4502; ld.global.v4.f32 {%f30916, %f30917, %f30918, %f30919}, [%rd4503+48]; ld.global.v4.f32 {%f30920, %f30921, %f30922, %f30923}, [%rd4503+64]; ld.global.v4.f32 {%f30924, %f30925, %f30926, %f30927}, [%rd4503+80]; bra.uni $L__BB107_4475; $L__BB107_4469: ld.local.f32 %f28513, [%rd5]; setp.eq.ftz.f32 %p3840, %f28513, 0f00000000; mov.b32 %r5408, %f28513; add.s32 %r5409, %r5408, 1; mov.b32 %f28514, %r5409; selp.f32 %f28515, 0f00000001, %f28514, %p3840; ld.const.u64 %rd4482, [kernel_params+2648]; cvta.to.global.u64 %rd4483, %rd4482; add.s64 %rd4485, %rd4483, %rd1282; st.global.f32 [%rd4485], %f28515; bra.uni $L__BB107_4479; $L__BB107_4437: add.s32 %r5394, %r1522, -10; setp.lt.u32 %p3806, %r5394, 7; @%p3806 bra $L__BB107_4439; add.s32 %r5395, %r1522, -22; setp.gt.u32 %p3807, %r5395, 2; setp.ne.s32 %p3808, %r1522, 27; mov.f32 %f30906, 0f3F800000; and.pred %p3809, %p3807, %p3808; @%p3809 bra $L__BB107_4440; $L__BB107_4439: ld.local.v2.f32 {%f28454, %f28455}, [%rd1027+20]; mul.ftz.f32 %f30906, %f28454, %f28455; bra.uni $L__BB107_4440; $L__BB107_4473: ld.local.v4.f32 {%f30916, %f30917, %f30918, %f30919}, [%rd5+104]; ld.local.v4.f32 {%f30920, %f30921, %f30922, %f30923}, [%rd5+120]; ld.local.v4.f32 {%f30924, %f30925, %f30926, %f30927}, [%rd5+136]; $L__BB107_4475: mul.ftz.f32 %f28548, %f30935, %f30917; fma.rn.ftz.f32 %f28549, %f30934, %f30916, %f28548; fma.rn.ftz.f32 %f28550, %f30936, %f30918, %f28549; mul.ftz.f32 %f28551, %f30935, %f30921; fma.rn.ftz.f32 %f28552, %f30934, %f30920, %f28551; fma.rn.ftz.f32 %f28553, %f30936, %f30922, %f28552; mul.ftz.f32 %f28554, %f30935, %f30925; fma.rn.ftz.f32 %f28555, %f30934, %f30924, %f28554; fma.rn.ftz.f32 %f28556, %f30936, %f30926, %f28555; add.ftz.f32 %f30928, %f30919, %f28550; add.ftz.f32 %f30929, %f30923, %f28553; add.ftz.f32 %f30930, %f30927, %f28556; mul.ftz.f32 %f28557, %f7596, %f30917; fma.rn.ftz.f32 %f28558, %f7595, %f30916, %f28557; mul.ftz.f32 %f28559, %f7596, %f30921; fma.rn.ftz.f32 %f28560, %f7595, %f30920, %f28559; mul.ftz.f32 %f28561, %f7596, %f30925; fma.rn.ftz.f32 %f28562, %f7595, %f30924, %f28561; fma.rn.ftz.f32 %f30931, %f7597, %f30918, %f28558; fma.rn.ftz.f32 %f30932, %f7597, %f30922, %f28560; fma.rn.ftz.f32 %f30933, %f7597, %f30926, %f28562; $L__BB107_4476: sub.ftz.f32 %f28563, %f7604, %f30928; sub.ftz.f32 %f28564, %f7605, %f30929; sub.ftz.f32 %f28565, %f7606, %f30930; sub.ftz.f32 %f28566, %f7607, %f30928; sub.ftz.f32 %f28567, %f7608, %f30929; sub.ftz.f32 %f28568, %f7609, %f30930; sub.ftz.f32 %f28569, %f7610, %f30928; sub.ftz.f32 %f28570, %f7611, %f30929; sub.ftz.f32 %f28571, %f7612, %f30930; sub.ftz.f32 %f28572, %f28569, %f28563; sub.ftz.f32 %f28573, %f28570, %f28564; sub.ftz.f32 %f28574, %f28571, %f28565; sub.ftz.f32 %f28575, %f28563, %f28566; sub.ftz.f32 %f28576, %f28564, %f28567; sub.ftz.f32 %f28577, %f28565, %f28568; sub.ftz.f32 %f28578, %f28566, %f28569; sub.ftz.f32 %f28579, %f28567, %f28570; sub.ftz.f32 %f28580, %f28568, %f28571; add.ftz.f32 %f28581, %f28569, %f28563; add.ftz.f32 %f28582, %f28570, %f28564; add.ftz.f32 %f28583, %f28571, %f28565; mul.ftz.f32 %f28584, %f28582, %f28574; mul.ftz.f32 %f28585, %f28583, %f28573; sub.ftz.f32 %f28586, %f28584, %f28585; mul.ftz.f32 %f28587, %f28583, %f28572; mul.ftz.f32 %f28588, %f28581, %f28574; sub.ftz.f32 %f28589, %f28587, %f28588; mul.ftz.f32 %f28590, %f28581, %f28573; mul.ftz.f32 %f28591, %f28582, %f28572; sub.ftz.f32 %f28592, %f28590, %f28591; mul.ftz.f32 %f28593, %f30932, %f28589; fma.rn.ftz.f32 %f28594, %f30931, %f28586, %f28593; fma.rn.ftz.f32 %f28595, %f30933, %f28592, %f28594; add.ftz.f32 %f28596, %f28563, %f28566; add.ftz.f32 %f28597, %f28564, %f28567; add.ftz.f32 %f28598, %f28565, %f28568; mul.ftz.f32 %f28599, %f28597, %f28577; mul.ftz.f32 %f28600, %f28598, %f28576; sub.ftz.f32 %f28601, %f28599, %f28600; mul.ftz.f32 %f28602, %f28598, %f28575; mul.ftz.f32 %f28603, %f28596, %f28577; sub.ftz.f32 %f28604, %f28602, %f28603; mul.ftz.f32 %f28605, %f28596, %f28576; mul.ftz.f32 %f28606, %f28597, %f28575; sub.ftz.f32 %f28607, %f28605, %f28606; mul.ftz.f32 %f28608, %f30932, %f28604; fma.rn.ftz.f32 %f28609, %f30931, %f28601, %f28608; fma.rn.ftz.f32 %f28610, %f30933, %f28607, %f28609; add.ftz.f32 %f28611, %f28566, %f28569; add.ftz.f32 %f28612, %f28567, %f28570; add.ftz.f32 %f28613, %f28568, %f28571; mul.ftz.f32 %f28614, %f28612, %f28580; mul.ftz.f32 %f28615, %f28613, %f28579; sub.ftz.f32 %f28616, %f28614, %f28615; mul.ftz.f32 %f28617, %f28613, %f28578; mul.ftz.f32 %f28618, %f28611, %f28580; sub.ftz.f32 %f28619, %f28617, %f28618; mul.ftz.f32 %f28620, %f28611, %f28579; mul.ftz.f32 %f28621, %f28612, %f28578; sub.ftz.f32 %f28622, %f28620, %f28621; mul.ftz.f32 %f28623, %f30932, %f28619; fma.rn.ftz.f32 %f28624, %f30931, %f28616, %f28623; fma.rn.ftz.f32 %f28625, %f30933, %f28622, %f28624; add.ftz.f32 %f28626, %f28595, %f28610; add.ftz.f32 %f28627, %f28625, %f28626; abs.ftz.f32 %f28628, %f28627; mul.ftz.f32 %f28629, %f28628, 0f34000000; min.ftz.f32 %f28630, %f28610, %f28625; min.ftz.f32 %f28631, %f28595, %f28630; max.ftz.f32 %f28632, %f28610, %f28625; max.ftz.f32 %f28633, %f28595, %f28632; setp.ge.ftz.f32 %p3847, %f28631, %f28629; mul.ftz.f32 %f28634, %f28628, 0fB4000000; setp.le.ftz.f32 %p3848, %f28633, %f28634; or.pred %p3849, %p3847, %p3848; @%p3849 bra $L__BB107_4478; ld.local.v2.f32 {%f28635, %f28636}, [%rd5+-64]; mul.ftz.f32 %f28639, %f28635, 0f43800000; cvt.rzi.ftz.s32.f32 %r5421, %f28639; mul.ftz.f32 %f28640, %f28636, 0f43800000; cvt.rzi.ftz.s32.f32 %r5422, %f28640; ld.local.f32 %f28641, [%rd5+-56]; mul.ftz.f32 %f28642, %f28641, 0f43800000; cvt.rzi.ftz.s32.f32 %r5423, %f28642; neg.s32 %r5424, %r5421; setp.lt.ftz.f32 %p3850, %f30934, 0f00000000; selp.b32 %r5425, %r5424, %r5421, %p3850; mov.b32 %r5426, %f30934; add.s32 %r5427, %r5425, %r5426; mov.b32 %f28643, %r5427; neg.s32 %r5428, %r5422; setp.lt.ftz.f32 %p3851, %f30935, 0f00000000; selp.b32 %r5429, %r5428, %r5422, %p3851; mov.b32 %r5430, %f30935; add.s32 %r5431, %r5429, %r5430; mov.b32 %f28644, %r5431; neg.s32 %r5432, %r5423; setp.lt.ftz.f32 %p3852, %f30936, 0f00000000; selp.b32 %r5433, %r5432, %r5423, %p3852; mov.b32 %r5434, %f30936; add.s32 %r5435, %r5433, %r5434; mov.b32 %f28645, %r5435; abs.ftz.f32 %f28646, %f30934; setp.lt.ftz.f32 %p3853, %f28646, 0f3D000000; fma.rn.ftz.f32 %f28647, %f28635, 0f37800000, %f30934; abs.ftz.f32 %f28648, %f30935; setp.lt.ftz.f32 %p3854, %f28648, 0f3D000000; fma.rn.ftz.f32 %f28649, %f28636, 0f37800000, %f30935; abs.ftz.f32 %f28650, %f30936; setp.lt.ftz.f32 %p3855, %f28650, 0f3D000000; fma.rn.ftz.f32 %f28651, %f28641, 0f37800000, %f30936; selp.f32 %f30936, %f28651, %f28645, %p3855; selp.f32 %f30935, %f28649, %f28644, %p3854; selp.f32 %f30934, %f28647, %f28643, %p3853; $L__BB107_4478: ld.const.u64 %rd4711, [kernel_params+2632]; cvta.to.global.u64 %rd4710, %rd4711; mul.lo.s64 %rd4505, %rd3, 12; add.s64 %rd4506, %rd4710, %rd4505; st.global.f32 [%rd4506], %f30934; st.global.f32 [%rd4506+4], %f30935; st.global.f32 [%rd4506+8], %f30936; ld.const.v2.u64 {%rd4507, %rd4508}, [kernel_params+2640]; cvta.to.global.u64 %rd4511, %rd4507; add.s64 %rd4512, %rd4511, %rd4505; st.global.f32 [%rd4512], %f7595; st.global.f32 [%rd4512+4], %f7596; st.global.f32 [%rd4512+8], %f7597; cvta.to.global.u64 %rd4513, %rd4508; add.s64 %rd4515, %rd4513, %rd1282; mov.u32 %r5436, 0; st.global.u32 [%rd4515], %r5436; ld.const.v2.u64 {%rd4516, %rd4517}, [kernel_params+2656]; cvta.to.global.u64 %rd4519, %rd4516; add.s64 %rd4520, %rd4519, %rd1282; mov.u32 %r5437, 2139095039; st.global.u32 [%rd4520], %r5437; ld.local.f32 %f28652, [%rd5+4]; ld.const.u64 %rd4521, [kernel_params+2672]; cvta.to.global.u64 %rd4522, %rd4521; add.s64 %rd4523, %rd4522, %rd1282; st.global.f32 [%rd4523], %f28652; $L__BB107_4479: div.approx.ftz.f32 %f28653, %f7592, %f30902; ld.const.u64 %rd4524, [kernel_params+2584]; cvta.to.global.u64 %rd4525, %rd4524; mul.lo.s64 %rd4526, %rd3, 12; add.s64 %rd4527, %rd4525, %rd4526; ld.global.f32 %f28654, [%rd4527]; mul.ftz.f32 %f28655, %f28653, %f28654; ld.global.f32 %f28656, [%rd4527+4]; div.approx.ftz.f32 %f28657, %f7593, %f30902; mul.ftz.f32 %f28658, %f28657, %f28656; ld.global.f32 %f28659, [%rd4527+8]; div.approx.ftz.f32 %f28660, %f7594, %f30902; mul.ftz.f32 %f28661, %f28660, %f28659; st.global.f32 [%rd4527], %f28655; st.global.f32 [%rd4527+4], %f28658; st.global.f32 [%rd4527+8], %f28661; and.b32 %r5438, %r1436, 2097152; setp.eq.s32 %p3856, %r5438, 0; @%p3856 bra $L__BB107_4494; ld.const.u64 %rd4528, [kernel_params+2464]; cvta.to.global.u64 %rd4529, %rd4528; add.s64 %rd4531, %rd4529, %rd2151; ld.global.u16 %rs600, [%rd4531]; setp.ne.s16 %p3857, %rs600, 0; @%p3857 bra $L__BB107_4494; ld.const.u64 %rd1032, [kernel_params+2600]; ld.local.f32 %f7670, [%rd2157+4]; ld.local.f32 %f7671, [%rd2157+8]; mov.f32 %f30942, 0f00000000; mov.f32 %f30937, %f30942; @%p3834 bra $L__BB107_4483; ld.local.f32 %f28663, [%rd2157]; div.approx.ftz.f32 %f30937, %f28663, %f7592; $L__BB107_4483: mov.f32 %f30938, %f30942; @%p3835 bra $L__BB107_4485; div.approx.ftz.f32 %f30938, %f7670, %f7593; $L__BB107_4485: mov.f32 %f30939, %f30942; @%p3837 bra $L__BB107_4487; div.approx.ftz.f32 %f30939, %f7671, %f7594; $L__BB107_4487: cvta.to.global.u64 %rd4532, %rd1032; add.s64 %rd4534, %rd4532, %rd4526; st.global.f32 [%rd4534], %f30937; st.global.f32 [%rd4534+4], %f30938; st.global.f32 [%rd4534+8], %f30939; ld.const.u64 %rd1033, [kernel_params+2608]; ld.local.f32 %f7678, [%rd2157+16]; ld.local.f32 %f7679, [%rd2157+20]; mov.f32 %f30940, %f30942; @%p3834 bra $L__BB107_4489; ld.local.f32 %f28667, [%rd2157+12]; div.approx.ftz.f32 %f30940, %f28667, %f7592; $L__BB107_4489: mov.f32 %f30941, %f30942; @%p3835 bra $L__BB107_4491; div.approx.ftz.f32 %f30941, %f7678, %f7593; $L__BB107_4491: @%p3837 bra $L__BB107_4493; div.approx.ftz.f32 %f30942, %f7679, %f7594; $L__BB107_4493: cvta.to.global.u64 %rd4535, %rd1033; add.s64 %rd4537, %rd4535, %rd4526; st.global.f32 [%rd4537], %f30940; st.global.f32 [%rd4537+4], %f30941; st.global.f32 [%rd4537+8], %f30942; $L__BB107_4494: setp.ne.s32 %p3864, %r1557, 0; @%p3864 bra $L__BB107_4496; ld.const.v2.u64 {%rd4538, %rd4539}, [kernel_params+2560]; cvta.to.global.u64 %rd4542, %rd4538; add.s64 %rd4544, %rd4542, %rd1282; st.global.f32 [%rd4544], %f30902; cvta.to.global.u64 %rd4545, %rd4539; add.s64 %rd4546, %rd4545, %rd1282; ld.global.f32 %f28670, [%rd4546]; min.ftz.f32 %f28671, %f30902, %f28670; st.global.f32 [%rd4546], %f28671; $L__BB107_4496: ld.global.u32 %r1559, [%rd27]; @%p3839 bra $L__BB107_4498; bra.uni $L__BB107_4497; $L__BB107_4498: ld.const.u64 %rd4557, [kernel_params+2464]; cvta.to.global.u64 %rd4558, %rd4557; add.s64 %rd1035, %rd4558, %rd2151; ld.global.u16 %r5447, [%rd1035]; add.s32 %r1560, %r5447, 1; ld.const.u32 %r5448, [kernel_params+1516]; setp.lt.u32 %p3868, %r1560, %r5448; or.b32 %r5449, %r1559, 131072; selp.b32 %r1561, %r1559, %r5449, %p3868; and.b32 %r5450, %r5877, 64; setp.eq.s32 %p3869, %r5450, 0; @%p3869 bra $L__BB107_4500; bra.uni $L__BB107_4499; $L__BB107_4500: and.b32 %r5459, %r5877, 2; setp.eq.s32 %p3872, %r5459, 0; and.b32 %r5879, %r5877, 4; @%p3872 bra $L__BB107_4504; bra.uni $L__BB107_4501; $L__BB107_4504: and.b32 %r5468, %r1561, -4096; or.b32 %r5469, %r5468, 4; and.b32 %r5470, %r5877, 128; setp.eq.s32 %p3876, %r5470, 0; and.b32 %r5471, %r5469, -20476; selp.b32 %r5472, %r5471, %r5469, %p3876; ld.const.u64 %rd4572, [kernel_params+2496]; cvta.to.global.u64 %rd4573, %rd4572; add.s64 %rd4575, %rd4573, %rd2151; ld.global.u16 %r5473, [%rd4575]; add.s32 %r5474, %r5473, 1; st.global.u16 [%rd4575], %r5474; or.b32 %r5475, %r5472, 131072; ld.const.u32 %r5476, [kernel_params+1528]; setp.lt.s32 %p3877, %r5474, %r5476; selp.b32 %r5880, %r5472, %r5475, %p3877; bra.uni $L__BB107_4505; $L__BB107_4497: ld.const.u64 %rd4550, [kernel_params+2472]; cvta.to.global.u64 %rd4551, %rd4550; add.s64 %rd4553, %rd4551, %rd2151; ld.global.u16 %r5439, [%rd4553]; add.s32 %r5440, %r5439, 1; ld.const.u32 %r5441, [kernel_params+1556]; setp.lt.u32 %p3866, %r5440, %r5441; selp.b32 %r5442, 64, 32832, %p3866; or.b32 %r5443, %r5442, %r1559; ld.const.u32 %r5444, [kernel_params+1560]; setp.eq.s32 %p3867, %r5444, 0; or.b32 %r5445, %r5443, 2048; selp.b32 %r5446, %r5445, %r5443, %p3867; st.global.u32 [%rd27], %r5446; st.global.u16 [%rd4553], %r5440; ld.const.u64 %rd4554, [kernel_params+2536]; cvta.to.global.u64 %rd4555, %rd4554; add.s64 %rd4556, %rd4555, %rd2151; ld.global.u16 %rs601, [%rd4556]; add.s16 %rs602, %rs601, 16; st.global.u16 [%rd4556], %rs602; mov.u32 %r5886, %r5877; bra.uni $L__BB107_4529; $L__BB107_4499: and.b32 %r5451, %r1561, -20480; and.b32 %r5452, %r1561, 100663296; setp.eq.s32 %p3870, %r5452, 0; selp.b32 %r5453, 67108992, 128, %p3870; or.b32 %r5454, %r5453, %r5451; ld.const.u64 %rd4560, [kernel_params+2504]; cvta.to.global.u64 %rd4561, %rd4560; add.s64 %rd4563, %rd4561, %rd2151; ld.global.u16 %r5455, [%rd4563]; add.s32 %r5456, %r5455, 1; st.global.u16 [%rd4563], %r5456; ld.const.u32 %r5457, [kernel_params+1532]; setp.lt.s32 %p3871, %r5456, %r5457; or.b32 %r5458, %r5454, 131072; selp.b32 %r5881, %r5454, %r5458, %p3871; bra.uni $L__BB107_4506; $L__BB107_4501: and.b32 %r5460, %r1561, -20480; or.b32 %r1564, %r5460, 2; or.b32 %r1565, %r5460, 131074; setp.eq.s32 %p3873, %r5879, 0; @%p3873 bra $L__BB107_4503; bra.uni $L__BB107_4502; $L__BB107_4503: ld.const.u64 %rd4568, [kernel_params+2488]; cvta.to.global.u64 %rd4569, %rd4568; add.s64 %rd4571, %rd4569, %rd2151; ld.global.u16 %r5465, [%rd4571]; add.s32 %r5466, %r5465, 1; st.global.u16 [%rd4571], %r5466; ld.const.u32 %r5467, [kernel_params+1524]; setp.lt.s32 %p3875, %r5466, %r5467; selp.b32 %r5880, %r1564, %r1565, %p3875; mov.u32 %r5879, 0; bra.uni $L__BB107_4505; $L__BB107_4502: ld.const.u64 %rd4564, [kernel_params+2480]; cvta.to.global.u64 %rd4565, %rd4564; add.s64 %rd4567, %rd4565, %rd2151; ld.global.u16 %r5461, [%rd4567]; add.s32 %r5462, %r5461, 1; st.global.u16 [%rd4567], %r5462; ld.const.u32 %r5463, [kernel_params+1520]; setp.lt.s32 %p3874, %r5462, %r5463; selp.b32 %r5880, %r1564, %r1565, %p3874; $L__BB107_4505: and.b32 %r5477, %r5877, 8; setp.eq.s32 %p3878, %r5477, 0; selp.b32 %r5478, 2096, 16, %p3878; setp.eq.s32 %p3879, %r5879, 0; selp.b32 %r5479, %r5478, 4104, %p3879; or.b32 %r5480, %r5479, %r5880; and.b32 %r5481, %r5880, 100679680; setp.eq.s32 %p3880, %r5481, 0; or.b32 %r5482, %r5480, 33554432; selp.b32 %r5881, %r5482, %r5480, %p3880; $L__BB107_4506: st.global.u32 [%rd27], %r5881; add.s32 %r5523, %r5447, 1; st.global.u16 [%rd1035], %r5523; ld.const.u64 %rd4576, [kernel_params+2536]; cvta.to.global.u64 %rd4577, %rd4576; add.s64 %rd4579, %rd4577, %rd2151; ld.global.u16 %rs603, [%rd4579]; add.s16 %rs604, %rs603, 16; st.global.u16 [%rd4579], %rs604; mov.u32 %r5886, %r5877; $L__BB107_4529: and.b32 %r5506, %r5886, 1; setp.eq.b32 %p3895, %r5506, 1; mov.pred %p3896, 0; xor.pred %p3897, %p3895, %p3896; not.pred %p3898, %p3897; @%p3898 bra $L__BB107_4544; ld.local.u32 %r1585, [%rd5+-36]; and.b32 %r5507, %r1585, 262144; setp.eq.s32 %p3899, %r5507, 0; @%p3899 bra $L__BB107_4544; bra.uni $L__BB107_4531; $L__BB107_4544: setp.ne.s32 %p3916, %r5886, 0; selp.u16 %rs748, 1, 0, %p3916; $L__BB107_4545: cvt.u32.u16 %r5522, %rs748; st.param.b32 [func_retval0+0], %r5522; ret; $L__BB107_4531: and.b32 %r5508, %r1585, 1; setp.eq.b32 %p3900, %r5508, 1; xor.pred %p3902, %p3900, %p3896; not.pred %p3903, %p3902; ld.const.u64 %rd4963, [kernel_params+2776]; cvta.to.global.u64 %rd4647, %rd4963; add.s64 %rd4649, %rd4647, %rd1282; ld.global.u32 %r5888, [%rd4649]; @%p3903 bra $L__BB107_4538; bra.uni $L__BB107_4532; $L__BB107_4538: mov.u64 %rd4680, kernel_params; add.s64 %rd4969, %rd4680, 2768; setp.eq.s32 %p3910, %r5888, -1; mov.u32 %r5893, 0; @%p3910 bra $L__BB107_4542; ld.local.u32 %r1595, [%rd5+-12]; add.s64 %rd4969, %rd4680, 2768; ld.local.u32 %r1596, [%rd5+-40]; mov.u32 %r5893, 0; $L__BB107_4540: ld.const.u64 %rd4683, [%rd4969]; cvta.to.global.u64 %rd4684, %rd4683; add.s64 %rd4686, %rd4684, %rd1282; ld.global.u32 %r5515, [%rd4686]; setp.eq.s32 %p3911, %r5515, %r1595; setp.eq.s32 %p3912, %r5888, %r1596; and.pred %p3913, %p3911, %p3912; @%p3913 bra $L__BB107_4544; add.s32 %r5893, %r5893, 1; mul.wide.s32 %rd4687, %r5893, 16; add.s64 %rd4689, %rd4680, %rd4687; add.s64 %rd4969, %rd4689, 2768; ld.const.u64 %rd4963, [%rd4689+2776]; cvta.to.global.u64 %rd4690, %rd4963; add.s64 %rd4692, %rd4690, %rd1282; ld.global.u32 %r5888, [%rd4692]; setp.ne.s32 %p3914, %r5888, -1; @%p3914 bra $L__BB107_4540; $L__BB107_4542: ld.const.u32 %r5516, [kernel_params+12]; add.s32 %r5517, %r5516, -1; setp.ge.u32 %p3915, %r5893, %r5517; @%p3915 bra $L__BB107_4544; ld.local.u32 %r5518, [%rd5+-12]; ld.local.u32 %r5519, [%rd5+-40]; ld.const.u64 %rd4693, [%rd4969]; cvta.to.global.u64 %rd4694, %rd4693; add.s64 %rd4696, %rd4694, %rd1282; st.global.u32 [%rd4696], %r5518; cvta.to.global.u64 %rd4697, %rd4963; add.s64 %rd4698, %rd4697, %rd1282; st.global.u32 [%rd4698], %r5519; add.s32 %r5520, %r5893, 1; mul.wide.s32 %rd4699, %r5520, 16; add.s64 %rd4701, %rd4680, %rd4699; ld.const.v2.u64 {%rd4702, %rd4703}, [%rd4701+2768]; cvta.to.global.u64 %rd4706, %rd4702; add.s64 %rd4707, %rd4706, %rd1282; mov.u32 %r5521, -1; st.global.u32 [%rd4707], %r5521; cvta.to.global.u64 %rd4708, %rd4703; add.s64 %rd4709, %rd4708, %rd1282; st.global.u32 [%rd4709], %r5521; bra.uni $L__BB107_4544; $L__BB107_4532: setp.eq.s32 %p3904, %r5888, -1; @%p3904 bra $L__BB107_4544; ld.local.u32 %r1587, [%rd5+-12]; mov.u64 %rd4651, kernel_params; add.s64 %rd4964, %rd4651, 2768; ld.local.u32 %r1588, [%rd5+-40]; mov.u32 %r5889, 0; $L__BB107_4534: ld.const.u64 %rd4652, [%rd4964]; cvta.to.global.u64 %rd4653, %rd4652; add.s64 %rd4655, %rd4653, %rd1282; ld.global.u32 %r5510, [%rd4655]; setp.eq.s32 %p3905, %r5510, %r1587; setp.eq.s32 %p3906, %r5888, %r1588; and.pred %p3907, %p3905, %p3906; add.s32 %r1591, %r5889, 1; mul.wide.s32 %rd4656, %r1591, 16; add.s64 %rd4658, %rd4651, %rd4656; add.s64 %rd1046, %rd4658, 2776; @%p3907 bra $L__BB107_4536; add.s64 %rd4964, %rd4658, 2768; ld.const.u64 %rd4963, [%rd1046]; cvta.to.global.u64 %rd4662, %rd4963; add.s64 %rd4664, %rd4662, %rd1282; ld.global.u32 %r5888, [%rd4664]; setp.eq.s32 %p3908, %r5888, -1; mov.u32 %r5889, %r1591; @%p3908 bra $L__BB107_4544; bra.uni $L__BB107_4534; $L__BB107_4536: ld.const.u64 %rd4966, [%rd1046+-24]; $L__BB107_4537: add.s32 %r5889, %r5889, 1; mul.wide.s32 %rd4665, %r5889, 16; add.s64 %rd4667, %rd4651, %rd4665; ld.const.v2.u64 {%rd4668, %rd4669}, [%rd4667+2768]; cvta.to.global.u64 %rd4670, %rd4668; add.s64 %rd4672, %rd4670, %rd1282; ld.global.u32 %r5511, [%rd4672]; cvta.to.global.u64 %rd4673, %rd4669; add.s64 %rd4674, %rd4673, %rd1282; ld.global.u32 %r5512, [%rd4674]; cvta.to.global.u64 %rd4675, %rd4966; add.s64 %rd4676, %rd4675, %rd1282; st.global.u32 [%rd4676], %r5511; cvta.to.global.u64 %rd4677, %rd4963; add.s64 %rd4678, %rd4677, %rd1282; st.global.u32 [%rd4678], %r5512; setp.eq.s32 %p3909, %r5512, -1; mov.u64 %rd4963, %rd4669; mov.u64 %rd4966, %rd4668; @%p3909 bra $L__BB107_4544; bra.uni $L__BB107_4537; } .func (.param .b32 func_retval0) _Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf( .param .b32 _Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf_param_0, .param .b64 _Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf_param_1 ) { .local .align 16 .b8 __local_depot108[7152]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<2190>; .reg .b16 %rs<162>; .reg .f32 %f<16797>; .reg .b32 %r<3302>; .reg .b64 %rd<3076>; mov.u64 %SPL, __local_depot108; cvta.local.u64 %SP, %SPL; ld.param.u32 %r953, [_Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf_param_0]; ld.param.u64 %rd618, [_Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf_param_1]; add.u64 %rd619, %SP, 1728; add.u64 %rd1, %SPL, 1728; ld.const.v2.u64 {%rd620, %rd621}, [kernel_params+2704]; mov.u32 %r954, 0; cvta.to.global.u64 %rd624, %rd621; cvt.s64.s32 %rd2, %r953; mul.wide.s32 %rd625, %r953, 4; add.s64 %rd626, %rd624, %rd625; ld.global.u32 %r955, [%rd626]; cvt.s64.s32 %rd3, %r955; ld.const.v2.u64 {%rd627, %rd628}, [kernel_params+2720]; cvta.to.global.u64 %rd631, %rd627; add.s64 %rd632, %rd631, %rd625; cvta.to.global.u64 %rd633, %rd628; add.s64 %rd634, %rd633, %rd625; ld.global.u32 %r1, [%rd634]; ld.const.v2.u64 {%rd635, %rd636}, [kernel_params+2688]; cvta.to.global.u64 %rd639, %rd636; add.s64 %rd640, %rd639, %rd625; ld.global.f32 %f1, [%rd640]; cvta.to.global.u64 %rd641, %rd620; add.s64 %rd642, %rd641, %rd625; ld.global.f32 %f2, [%rd642]; cvta.to.global.u64 %rd643, %rd635; add.s64 %rd644, %rd643, %rd625; ld.global.f32 %f15602, [%rd644]; ld.const.u64 %rd645, [kernel_params+2632]; cvta.to.global.u64 %rd646, %rd645; mul.wide.s32 %rd647, %r953, 12; add.s64 %rd4, %rd646, %rd647; ld.global.f32 %f15598, [%rd4]; ld.global.f32 %f15597, [%rd4+4]; ld.global.f32 %f15596, [%rd4+8]; ld.const.u64 %rd648, [kernel_params+2640]; cvta.to.global.u64 %rd649, %rd648; add.s64 %rd650, %rd649, %rd647; ld.global.f32 %f7, [%rd650]; ld.global.f32 %f8, [%rd650+4]; ld.global.f32 %f9, [%rd650+8]; ld.const.u64 %rd651, [kernel_params+2664]; cvta.to.global.u64 %rd652, %rd651; add.s64 %rd653, %rd652, %rd625; ld.global.f32 %f10, [%rd653]; ld.const.v2.u64 {%rd654, %rd655}, [kernel_params+2672]; cvta.to.global.u64 %rd656, %rd654; add.s64 %rd657, %rd656, %rd625; ld.global.f32 %f11, [%rd657]; cvta.to.global.u64 %rd658, %rd655; add.s64 %rd659, %rd658, %rd625; ld.global.f32 %f12, [%rd659]; add.s64 %rd7, %rd1, 88; ld.global.u32 %r956, [%rd632]; st.local.f32 [%rd1+88], %f15602; mov.b32 %r957, %f2; mov.b32 %r958, %f1; st.local.v4.u32 [%rd1+64], {%r1, %r958, %r957, %r956}; ld.const.u64 %rd660, [kernel_params+2184]; cvta.to.global.u64 %rd8, %rd660; cvt.s64.s32 %rd9, %r956; mul.wide.s32 %rd661, %r956, 4; add.s64 %rd662, %rd8, %rd661; ld.global.u32 %r3072, [%rd662]; st.local.u32 [%rd1+56], %r3072; st.local.u32 [%rd1+60], %r955; st.local.u32 [%rd1+52], %r954; mov.b32 %r959, %f10; mov.u32 %r3066, -1; st.local.v2.u32 [%rd1+80], {%r3066, %r959}; and.b32 %r3, %r3072, 2; setp.eq.s32 %p15, %r3, 0; @%p15 bra $L__BB108_7; ld.const.u64 %rd663, [kernel_params+2160]; cvta.to.global.u64 %rd664, %rd663; mul.lo.s64 %rd665, %rd9, 208; add.s64 %rd666, %rd664, %rd665; ld.const.u64 %rd667, [kernel_params+2176]; cvta.to.global.u64 %rd668, %rd667; ld.global.u32 %rd669, [%rd666+168]; ld.global.u32 %r961, [%rd666+152]; shl.b32 %r962, %r961, 1; cvt.rn.f32.s32 %f4315, %r962; mul.ftz.f32 %f4316, %f10, %f4315; cvt.rzi.ftz.s32.f32 %r963, %f4316; add.s32 %r964, %r962, -1; min.s32 %r965, %r963, %r964; cvt.rn.f32.s32 %f4317, %r965; sub.ftz.f32 %f13, %f4316, %f4317; cvt.s64.s32 %rd670, %r965; add.s64 %rd671, %rd670, %rd669; shl.b64 %rd672, %rd671, 6; add.s64 %rd10, %rd668, %rd672; ld.global.v4.f32 {%f4318, %f4319, %f4320, %f4321}, [%rd10]; ld.global.v4.f32 {%f4322, %f4323, %f4324, %f4325}, [%rd10+64]; mul.ftz.f32 %f4326, %f4319, %f4323; fma.rn.ftz.f32 %f4327, %f4318, %f4322, %f4326; mul.ftz.f32 %f4328, %f4321, %f4325; fma.rn.ftz.f32 %f4329, %f4320, %f4324, %f4328; add.ftz.f32 %f22, %f4327, %f4329; setp.gt.ftz.f32 %p16, %f22, 0f3F7FDF3B; @%p16 bra $L__BB108_3; bra.uni $L__BB108_2; $L__BB108_3: mov.f32 %f15515, 0f3F800000; sub.ftz.f32 %f15514, %f15515, %f13; mul.ftz.f32 %f4384, %f4318, %f15514; fma.rn.ftz.f32 %f4385, %f4322, %f13, %f4384; mul.ftz.f32 %f4386, %f4319, %f15514; fma.rn.ftz.f32 %f4387, %f4323, %f13, %f4386; mul.ftz.f32 %f4388, %f4320, %f15514; fma.rn.ftz.f32 %f4389, %f4324, %f13, %f4388; mul.ftz.f32 %f4390, %f4321, %f15514; fma.rn.ftz.f32 %f4391, %f4325, %f13, %f4390; mul.ftz.f32 %f4392, %f4387, %f4387; fma.rn.ftz.f32 %f4393, %f4385, %f4385, %f4392; mul.ftz.f32 %f4394, %f4391, %f4391; fma.rn.ftz.f32 %f4395, %f4389, %f4389, %f4394; add.ftz.f32 %f4396, %f4393, %f4395; rsqrt.approx.ftz.f32 %f4397, %f4396; mul.ftz.f32 %f15528, %f4391, %f4397; mul.ftz.f32 %f15527, %f4389, %f4397; mul.ftz.f32 %f15526, %f4387, %f4397; mul.ftz.f32 %f15525, %f4385, %f4397; bra.uni $L__BB108_4; $L__BB108_2: mov.f32 %f15511, 0f3F800000; mov.f32 %f4331, 0fBF800000; max.ftz.f32 %f4332, %f22, %f4331; min.ftz.f32 %f4334, %f4332, %f15511; abs.ftz.f32 %f4335, %f4334; sub.ftz.f32 %f4336, %f15511, %f4335; mul.ftz.f32 %f4337, %f4336, 0f3F000000; sqrt.approx.ftz.f32 %f4338, %f4337; setp.gt.ftz.f32 %p17, %f4335, 0f3F11EB85; selp.f32 %f4339, %f4338, %f4335, %p17; mul.ftz.f32 %f4340, %f4339, %f4339; mov.f32 %f4341, 0f3C94D2E9; mov.f32 %f4342, 0f3D53F941; fma.rn.ftz.f32 %f4343, %f4342, %f4340, %f4341; mov.f32 %f4344, 0f3D3F841F; fma.rn.ftz.f32 %f4345, %f4343, %f4340, %f4344; mov.f32 %f4346, 0f3D994929; fma.rn.ftz.f32 %f4347, %f4345, %f4340, %f4346; mov.f32 %f4348, 0f3E2AAB94; fma.rn.ftz.f32 %f4349, %f4347, %f4340, %f4348; mul.ftz.f32 %f4350, %f4340, %f4349; fma.rn.ftz.f32 %f4351, %f4350, %f4339, %f4339; add.ftz.f32 %f4352, %f4351, %f4351; mov.f32 %f4353, 0f3FC90FDB; sub.ftz.f32 %f4354, %f4353, %f4351; selp.f32 %f4355, %f4352, %f4354, %p17; setp.lt.ftz.f32 %p18, %f4334, 0f00000000; mov.f32 %f4356, 0f40490FDB; sub.ftz.f32 %f4357, %f4356, %f4355; selp.f32 %f4358, %f4357, %f4355, %p18; mul.ftz.f32 %f4359, %f4318, %f22; sub.ftz.f32 %f4360, %f4322, %f4359; mul.ftz.f32 %f4361, %f4319, %f22; sub.ftz.f32 %f4362, %f4323, %f4361; mul.ftz.f32 %f4363, %f4320, %f22; sub.ftz.f32 %f4364, %f4324, %f4363; mul.ftz.f32 %f4365, %f4321, %f22; sub.ftz.f32 %f4366, %f4325, %f4365; mul.ftz.f32 %f4367, %f4362, %f4362; fma.rn.ftz.f32 %f4368, %f4360, %f4360, %f4367; mul.ftz.f32 %f4369, %f4366, %f4366; fma.rn.ftz.f32 %f4370, %f4364, %f4364, %f4369; add.ftz.f32 %f4371, %f4368, %f4370; rsqrt.approx.ftz.f32 %f4372, %f4371; mul.ftz.f32 %f4373, %f4360, %f4372; mul.ftz.f32 %f4374, %f4362, %f4372; mul.ftz.f32 %f4375, %f4364, %f4372; mul.ftz.f32 %f4376, %f4366, %f4372; mul.ftz.f32 %f4377, %f13, %f4358; cos.approx.ftz.f32 %f4378, %f4377; sin.approx.ftz.f32 %f4379, %f4377; mul.ftz.f32 %f4380, %f4373, %f4379; mul.ftz.f32 %f4381, %f4374, %f4379; mul.ftz.f32 %f4382, %f4375, %f4379; mul.ftz.f32 %f4383, %f4376, %f4379; fma.rn.ftz.f32 %f15528, %f4321, %f4378, %f4383; fma.rn.ftz.f32 %f15527, %f4320, %f4378, %f4382; fma.rn.ftz.f32 %f15526, %f4319, %f4378, %f4381; fma.rn.ftz.f32 %f15525, %f4318, %f4378, %f4380; $L__BB108_4: mov.f32 %f15513, 0f3F800000; sub.ftz.f32 %f15512, %f15513, %f13; ld.global.v4.f32 {%f4398, %f4399, %f4400, %f4401}, [%rd10+16]; ld.global.v4.f32 {%f4406, %f4407, %f4408, %f4409}, [%rd10+80]; mul.ftz.f32 %f4414, %f13, %f4406; mul.ftz.f32 %f4415, %f13, %f4407; mul.ftz.f32 %f4416, %f13, %f4408; mul.ftz.f32 %f4417, %f13, %f4409; fma.rn.ftz.f32 %f144, %f15512, %f4398, %f4414; fma.rn.ftz.f32 %f148, %f15512, %f4399, %f4415; fma.rn.ftz.f32 %f152, %f15512, %f4400, %f4416; fma.rn.ftz.f32 %f4418, %f15512, %f4401, %f4417; ld.global.v4.f32 {%f4419, %f4420, %f4421, %f4422}, [%rd10+32]; ld.global.v4.f32 {%f4427, %f4428, %f4429, %f4430}, [%rd10+96]; mul.ftz.f32 %f4435, %f13, %f4427; mul.ftz.f32 %f4436, %f13, %f4428; mul.ftz.f32 %f4437, %f13, %f4429; mul.ftz.f32 %f4438, %f13, %f4430; fma.rn.ftz.f32 %f4439, %f15512, %f4419, %f4435; fma.rn.ftz.f32 %f4440, %f15512, %f4420, %f4436; fma.rn.ftz.f32 %f4441, %f15512, %f4421, %f4437; fma.rn.ftz.f32 %f4442, %f15512, %f4422, %f4438; ld.global.v4.f32 {%f4443, %f4444, %f4445, %f4446}, [%rd10+48]; ld.global.v4.f32 {%f4451, %f4452, %f4453, %f4454}, [%rd10+112]; mul.ftz.f32 %f4459, %f13, %f4451; mul.ftz.f32 %f4460, %f13, %f4452; mul.ftz.f32 %f4461, %f13, %f4453; mul.ftz.f32 %f4462, %f13, %f4454; fma.rn.ftz.f32 %f4463, %f15512, %f4443, %f4459; fma.rn.ftz.f32 %f4464, %f15512, %f4444, %f4460; fma.rn.ftz.f32 %f4465, %f15512, %f4445, %f4461; fma.rn.ftz.f32 %f4466, %f15512, %f4446, %f4462; mul.ftz.f32 %f4467, %f15528, 0f3FB504F3; mul.ftz.f32 %f4468, %f15525, 0f3FB504F3; mul.ftz.f32 %f4469, %f15526, 0f3FB504F3; mul.ftz.f32 %f4470, %f15527, 0f3FB504F3; mul.ftz.f32 %f4471, %f4467, %f4468; mul.ftz.f32 %f4472, %f4467, %f4469; mul.ftz.f32 %f4473, %f4467, %f4470; mul.ftz.f32 %f4474, %f4468, %f4468; mul.ftz.f32 %f4475, %f4468, %f4469; mul.ftz.f32 %f4476, %f4468, %f4470; mul.ftz.f32 %f4477, %f4469, %f4469; mul.ftz.f32 %f4478, %f4469, %f4470; mul.ftz.f32 %f4479, %f4470, %f4470; sub.ftz.f32 %f4481, %f15513, %f4477; sub.ftz.f32 %f4482, %f4481, %f4479; sub.ftz.f32 %f4483, %f4475, %f4473; add.ftz.f32 %f4484, %f4472, %f4476; add.ftz.f32 %f4485, %f4473, %f4475; sub.ftz.f32 %f4486, %f15513, %f4474; sub.ftz.f32 %f4487, %f4486, %f4479; sub.ftz.f32 %f4488, %f4478, %f4471; sub.ftz.f32 %f4489, %f4476, %f4472; add.ftz.f32 %f4490, %f4471, %f4478; sub.ftz.f32 %f4491, %f4486, %f4477; mul.ftz.f32 %f4492, %f4483, %f4441; fma.rn.ftz.f32 %f4493, %f4482, %f4418, %f4492; fma.rn.ftz.f32 %f141, %f4484, %f4464, %f4493; mul.ftz.f32 %f4494, %f4483, %f4442; fma.rn.ftz.f32 %f4495, %f4482, %f4439, %f4494; fma.rn.ftz.f32 %f142, %f4484, %f4465, %f4495; mul.ftz.f32 %f4496, %f4483, %f4463; fma.rn.ftz.f32 %f4497, %f4482, %f4440, %f4496; fma.rn.ftz.f32 %f143, %f4484, %f4466, %f4497; mul.ftz.f32 %f4498, %f4487, %f4441; fma.rn.ftz.f32 %f4499, %f4485, %f4418, %f4498; fma.rn.ftz.f32 %f145, %f4488, %f4464, %f4499; mul.ftz.f32 %f4500, %f4487, %f4442; fma.rn.ftz.f32 %f4501, %f4485, %f4439, %f4500; fma.rn.ftz.f32 %f146, %f4488, %f4465, %f4501; mul.ftz.f32 %f4502, %f4487, %f4463; fma.rn.ftz.f32 %f4503, %f4485, %f4440, %f4502; fma.rn.ftz.f32 %f147, %f4488, %f4466, %f4503; mul.ftz.f32 %f4504, %f4490, %f4441; fma.rn.ftz.f32 %f4505, %f4489, %f4418, %f4504; fma.rn.ftz.f32 %f149, %f4491, %f4464, %f4505; mul.ftz.f32 %f4506, %f4490, %f4442; fma.rn.ftz.f32 %f4507, %f4489, %f4439, %f4506; fma.rn.ftz.f32 %f150, %f4491, %f4465, %f4507; mul.ftz.f32 %f4508, %f4490, %f4463; fma.rn.ftz.f32 %f4509, %f4489, %f4440, %f4508; fma.rn.ftz.f32 %f151, %f4491, %f4466, %f4509; st.local.v4.f32 [%rd7+56], {%f141, %f142, %f143, %f144}; st.local.v4.f32 [%rd7+72], {%f145, %f146, %f147, %f148}; st.local.v4.f32 [%rd7+88], {%f149, %f150, %f151, %f152}; mul.ftz.f32 %f4510, %f146, %f151; mul.ftz.f32 %f60, %f150, %f147; sub.ftz.f32 %f15531, %f4510, %f60; mul.ftz.f32 %f62, %f150, %f143; mul.ftz.f32 %f4511, %f142, %f151; sub.ftz.f32 %f15530, %f62, %f4511; mul.ftz.f32 %f64, %f142, %f147; mul.ftz.f32 %f4512, %f146, %f143; sub.ftz.f32 %f15529, %f64, %f4512; mul.ftz.f32 %f4513, %f145, %f15530; fma.rn.ftz.f32 %f4514, %f141, %f15531, %f4513; fma.rn.ftz.f32 %f15535, %f149, %f15529, %f4514; setp.neu.ftz.f32 %p19, %f15535, 0f00000000; mov.f32 %f15532, %f141; mov.f32 %f15533, %f146; mov.f32 %f15534, %f151; @%p19 bra $L__BB108_6; add.ftz.f32 %f15532, %f141, 0f322BCC77; add.ftz.f32 %f15534, %f151, 0f322BCC77; add.ftz.f32 %f15533, %f146, 0f322BCC77; mul.ftz.f32 %f4515, %f15533, %f15534; sub.ftz.f32 %f15531, %f4515, %f60; mul.ftz.f32 %f4516, %f142, %f15534; sub.ftz.f32 %f15530, %f62, %f4516; mul.ftz.f32 %f4517, %f15533, %f143; sub.ftz.f32 %f15529, %f64, %f4517; mul.ftz.f32 %f4518, %f145, %f15530; fma.rn.ftz.f32 %f4519, %f15532, %f15531, %f4518; fma.rn.ftz.f32 %f4520, %f149, %f15529, %f4519; setp.eq.ftz.f32 %p20, %f4520, 0f00000000; selp.f32 %f15535, 0f7F7FFFFF, %f4520, %p20; $L__BB108_6: mul.ftz.f32 %f4521, %f145, %f15534; mul.ftz.f32 %f4522, %f149, %f147; sub.ftz.f32 %f4523, %f4522, %f4521; mul.ftz.f32 %f4524, %f149, %f143; mul.ftz.f32 %f4525, %f15532, %f15534; sub.ftz.f32 %f4526, %f4525, %f4524; mul.ftz.f32 %f4527, %f147, %f15532; mul.ftz.f32 %f4528, %f145, %f143; sub.ftz.f32 %f4529, %f4528, %f4527; mul.ftz.f32 %f4530, %f149, %f15533; mul.ftz.f32 %f4531, %f145, %f150; sub.ftz.f32 %f4532, %f4531, %f4530; mul.ftz.f32 %f4533, %f150, %f15532; mul.ftz.f32 %f4534, %f149, %f142; sub.ftz.f32 %f4535, %f4534, %f4533; mul.ftz.f32 %f4536, %f145, %f142; mul.ftz.f32 %f4537, %f15532, %f15533; sub.ftz.f32 %f4538, %f4537, %f4536; div.approx.ftz.f32 %f15570, %f15530, %f15535; mul.ftz.f32 %f4539, %f148, %f15570; div.approx.ftz.f32 %f15569, %f15531, %f15535; fma.rn.ftz.f32 %f4540, %f144, %f15569, %f4539; div.approx.ftz.f32 %f15571, %f15529, %f15535; fma.rn.ftz.f32 %f4541, %f152, %f15571, %f4540; div.approx.ftz.f32 %f15574, %f4526, %f15535; mul.ftz.f32 %f4542, %f148, %f15574; div.approx.ftz.f32 %f15573, %f4523, %f15535; fma.rn.ftz.f32 %f4543, %f144, %f15573, %f4542; div.approx.ftz.f32 %f15575, %f4529, %f15535; fma.rn.ftz.f32 %f4544, %f152, %f15575, %f4543; div.approx.ftz.f32 %f15578, %f4535, %f15535; mul.ftz.f32 %f4545, %f148, %f15578; div.approx.ftz.f32 %f15577, %f4532, %f15535; fma.rn.ftz.f32 %f4546, %f144, %f15577, %f4545; div.approx.ftz.f32 %f15579, %f4538, %f15535; fma.rn.ftz.f32 %f4547, %f152, %f15579, %f4546; neg.ftz.f32 %f15584, %f4541; neg.ftz.f32 %f15592, %f4547; neg.ftz.f32 %f15588, %f4544; st.local.v4.f32 [%rd7+104], {%f15569, %f15570, %f15571, %f15584}; st.local.v4.f32 [%rd7+120], {%f15573, %f15574, %f15575, %f15588}; st.local.v4.f32 [%rd7+136], {%f15577, %f15578, %f15579, %f15592}; $L__BB108_7: neg.ftz.f32 %f165, %f7; st.local.f32 [%rd7+-52], %f165; neg.ftz.f32 %f166, %f8; st.local.f32 [%rd7+-48], %f166; neg.ftz.f32 %f167, %f9; st.local.f32 [%rd7+-44], %f167; and.b32 %r966, %r1, 6; setp.eq.s32 %p21, %r966, 0; @%p21 bra $L__BB108_46; and.b32 %r4, %r3072, 4; setp.ne.s32 %p22, %r4, 0; mov.f32 %f15599, %f9; mov.f32 %f15600, %f8; mov.f32 %f15601, %f7; @%p22 bra $L__BB108_14; and.b32 %r3042, %r3072, 2; setp.ne.s32 %p23, %r3042, 0; mov.f32 %f15581, %f15569; mov.f32 %f15582, %f15570; mov.f32 %f15583, %f15571; mov.f32 %f15585, %f15573; mov.f32 %f15586, %f15574; mov.f32 %f15587, %f15575; mov.f32 %f15589, %f15577; mov.f32 %f15590, %f15578; mov.f32 %f15591, %f15579; @%p23 bra $L__BB108_11; ld.const.u64 %rd673, [kernel_params+2160]; cvta.to.global.u64 %rd674, %rd673; mul.lo.s64 %rd675, %rd9, 208; add.s64 %rd676, %rd674, %rd675; ld.global.v4.f32 {%f15581, %f15582, %f15583, %f15584}, [%rd676+48]; ld.global.v4.f32 {%f15585, %f15586, %f15587, %f15588}, [%rd676+64]; ld.global.v4.f32 {%f15589, %f15590, %f15591, %f15592}, [%rd676+80]; $L__BB108_11: mul.ftz.f32 %f4560, %f15597, %f15582; fma.rn.ftz.f32 %f4561, %f15598, %f15581, %f4560; fma.rn.ftz.f32 %f4562, %f15596, %f15583, %f4561; add.ftz.f32 %f192, %f15584, %f4562; mul.ftz.f32 %f4563, %f15597, %f15586; fma.rn.ftz.f32 %f4564, %f15598, %f15585, %f4563; fma.rn.ftz.f32 %f4565, %f15596, %f15587, %f4564; add.ftz.f32 %f193, %f15588, %f4565; mul.ftz.f32 %f4566, %f15597, %f15590; fma.rn.ftz.f32 %f4567, %f15598, %f15589, %f4566; fma.rn.ftz.f32 %f4568, %f15596, %f15591, %f4567; add.ftz.f32 %f15596, %f15592, %f4568; mul.ftz.f32 %f4569, %f15602, %f7; mul.ftz.f32 %f4570, %f15602, %f8; mul.ftz.f32 %f4571, %f4570, %f15582; fma.rn.ftz.f32 %f4572, %f4569, %f15581, %f4571; mul.ftz.f32 %f4573, %f15602, %f9; fma.rn.ftz.f32 %f15601, %f4573, %f15583, %f4572; mul.ftz.f32 %f4574, %f4570, %f15586; fma.rn.ftz.f32 %f4575, %f4569, %f15585, %f4574; fma.rn.ftz.f32 %f15600, %f4573, %f15587, %f4575; mul.ftz.f32 %f4576, %f4570, %f15590; fma.rn.ftz.f32 %f4577, %f4569, %f15589, %f4576; fma.rn.ftz.f32 %f15599, %f4573, %f15591, %f4577; mul.ftz.f32 %f4578, %f15600, %f15600; fma.rn.ftz.f32 %f4579, %f15601, %f15601, %f4578; fma.rn.ftz.f32 %f4580, %f15599, %f15599, %f4579; sqrt.approx.ftz.f32 %f15602, %f4580; setp.neu.ftz.f32 %p24, %f15602, 0f00000000; @%p24 bra $L__BB108_12; bra.uni $L__BB108_13; $L__BB108_12: div.approx.ftz.f32 %f15599, %f15599, %f15602; div.approx.ftz.f32 %f15600, %f15600, %f15602; div.approx.ftz.f32 %f15601, %f15601, %f15602; $L__BB108_13: mov.f32 %f15597, %f193; mov.f32 %f15598, %f192; $L__BB108_14: ld.const.u64 %rd677, [kernel_params+2264]; cvta.to.global.u64 %rd678, %rd677; shl.b64 %rd679, %rd3, 4; add.s64 %rd680, %rd678, %rd679; add.s64 %rd11, %rd680, 4; shr.s32 %r967, %r1, 7; ld.global.u32 %r968, [%rd680+4]; add.s32 %r5, %r967, %r968; add.s32 %r969, %r5, -1; max.s32 %r6, %r969, %r968; add.s32 %r970, %r5, 2; ld.global.u32 %r971, [%rd680+8]; add.s32 %r972, %r968, %r971; add.s32 %r973, %r972, -1; min.s32 %r7, %r970, %r973; and.b32 %r974, %r1, 64; setp.eq.s32 %p25, %r974, 0; @%p25 bra $L__BB108_28; bra.uni $L__BB108_15; $L__BB108_28: ld.const.u64 %rd722, [kernel_params+2272]; cvta.to.global.u64 %rd723, %rd722; mul.wide.s32 %rd724, %r6, 16; add.s64 %rd725, %rd723, %rd724; ld.global.v4.f32 {%f15638, %f15637, %f15636, %f4665}, [%rd725]; mul.wide.s32 %rd726, %r5, 16; add.s64 %rd727, %rd723, %rd726; ld.global.v4.f32 {%f15635, %f15634, %f15633, %f4669}, [%rd727]; ld.global.v4.f32 {%f15632, %f15631, %f15630, %f4673}, [%rd727+16]; mul.wide.s32 %rd728, %r7, 16; add.s64 %rd729, %rd723, %rd728; ld.global.v4.f32 {%f15629, %f15628, %f15627, %f4677}, [%rd729]; bra.uni $L__BB108_29; $L__BB108_46: and.b32 %r993, %r1, 8; setp.eq.s32 %p46, %r993, 0; @%p46 bra $L__BB108_67; bra.uni $L__BB108_47; $L__BB108_67: setp.eq.s32 %p59, %r1, 1; @%p59 bra $L__BB108_69; bra.uni $L__BB108_68; $L__BB108_69: cvt.u32.u64 %r1012, %rd3; ld.const.u64 %rd783, [kernel_params+2232]; cvta.to.global.u64 %rd784, %rd783; mul.wide.s32 %rd785, %r1012, 16; add.s64 %rd786, %rd784, %rd785; add.s64 %rd27, %rd786, 12; ld.const.u64 %rd787, [kernel_params+2256]; cvta.to.global.u64 %rd788, %rd787; ld.global.u32 %r1013, [%rd786+12]; mul.wide.u32 %rd789, %r1013, 12; add.s64 %rd790, %rd788, %rd789; add.s32 %r1014, %r1013, 1; mul.wide.u32 %rd791, %r1014, 12; add.s64 %rd792, %rd788, %rd791; add.s32 %r1015, %r1013, 2; mul.wide.u32 %rd793, %r1015, 12; add.s64 %rd794, %rd788, %rd793; and.b32 %r1016, %r3072, 8; setp.eq.s32 %p60, %r1016, 0; ld.global.f32 %f5011, [%rd794]; ld.global.f32 %f425, [%rd790]; sub.ftz.f32 %f426, %f5011, %f425; ld.global.f32 %f5012, [%rd794+4]; ld.global.f32 %f427, [%rd790+4]; sub.ftz.f32 %f428, %f5012, %f427; ld.global.f32 %f5013, [%rd794+8]; ld.global.f32 %f429, [%rd790+8]; sub.ftz.f32 %f430, %f5013, %f429; ld.global.f32 %f5014, [%rd792]; sub.ftz.f32 %f431, %f5014, %f425; ld.global.f32 %f5015, [%rd792+4]; sub.ftz.f32 %f432, %f5015, %f427; ld.global.f32 %f5016, [%rd792+8]; sub.ftz.f32 %f433, %f5016, %f429; mul.ftz.f32 %f434, %f433, %f428; mul.ftz.f32 %f435, %f432, %f430; mul.ftz.f32 %f436, %f431, %f430; mul.ftz.f32 %f437, %f433, %f426; mul.ftz.f32 %f438, %f432, %f426; mul.ftz.f32 %f439, %f431, %f428; @%p60 bra $L__BB108_71; bra.uni $L__BB108_70; $L__BB108_71: sub.ftz.f32 %f5024, %f435, %f434; sub.ftz.f32 %f5025, %f437, %f436; mul.ftz.f32 %f5026, %f5025, %f5025; fma.rn.ftz.f32 %f5027, %f5024, %f5024, %f5026; sub.ftz.f32 %f5028, %f439, %f438; fma.rn.ftz.f32 %f5029, %f5028, %f5028, %f5027; sqrt.approx.ftz.f32 %f5030, %f5029; div.approx.ftz.f32 %f15668, %f5028, %f5030; div.approx.ftz.f32 %f15667, %f5025, %f5030; div.approx.ftz.f32 %f15666, %f5024, %f5030; bra.uni $L__BB108_72; $L__BB108_15: ld.const.u64 %rd681, [kernel_params+2160]; cvta.to.global.u64 %rd682, %rd681; mul.lo.s64 %rd683, %rd9, 208; add.s64 %rd684, %rd682, %rd683; ld.global.u32 %r8, [%rd684+148]; ld.global.u32 %r9, [%rd684+152]; shl.b32 %r975, %r9, 1; cvt.rn.f32.s32 %f4581, %r975; mul.ftz.f32 %f218, %f10, %f4581; cvt.rzi.ftz.s32.f32 %r976, %f218; add.s32 %r977, %r975, -1; min.s32 %r10, %r976, %r977; ld.const.u64 %rd685, [kernel_params+2312]; cvta.to.global.u64 %rd12, %rd685; ld.global.u32 %r3063, [%rd684+164]; mul.wide.u32 %rd686, %r3063, 16; add.s64 %rd687, %rd12, %rd686; ld.global.u32 %r3062, [%rd687+8]; ld.global.u16 %rs133, [%rd687+12]; ld.global.u64 %rd2923, [%rd687]; setp.eq.s64 %p26, %rd2923, 11; @%p26 bra $L__BB108_20; $L__BB108_16: setp.eq.s64 %p27, %rd2923, 0; @%p27 bra $L__BB108_18; add.s32 %r3063, %r3063, 2; bra.uni $L__BB108_19; $L__BB108_18: setp.eq.s16 %p28, %rs133, 0; mov.u32 %r3063, %r3062; @%p28 bra $L__BB108_21; $L__BB108_19: mul.wide.u32 %rd688, %r3063, 16; add.s64 %rd689, %rd12, %rd688; ld.global.u32 %r3062, [%rd689+8]; ld.global.u16 %rs133, [%rd689+12]; ld.global.u64 %rd2923, [%rd689]; setp.ne.s64 %p29, %rd2923, 11; @%p29 bra $L__BB108_16; $L__BB108_20: setp.eq.s16 %p30, %rs133, 0; selp.b32 %r3066, -1, %r3062, %p30; $L__BB108_21: setp.eq.s32 %p31, %r10, %r9; @%p31 bra $L__BB108_23; bra.uni $L__BB108_22; $L__BB108_23: ld.const.u64 %rd698, [kernel_params+2272]; cvta.to.global.u64 %rd699, %rd698; mul.wide.s32 %rd700, %r6, 16; add.s64 %rd701, %rd699, %rd700; ld.global.v4.f32 {%f15614, %f15613, %f15612, %f4601}, [%rd701]; mul.wide.s32 %rd702, %r5, 16; add.s64 %rd703, %rd699, %rd702; ld.global.v4.f32 {%f15611, %f15610, %f15609, %f4605}, [%rd703]; ld.global.v4.f32 {%f15608, %f15607, %f15606, %f4609}, [%rd703+16]; mul.wide.s32 %rd704, %r7, 16; add.s64 %rd705, %rd699, %rd704; ld.global.v4.f32 {%f15605, %f15604, %f15603, %f4613}, [%rd705]; bra.uni $L__BB108_24; $L__BB108_47: ld.const.u64 %rd744, [kernel_params+2304]; cvta.to.global.u64 %rd745, %rd744; shl.b64 %rd746, %rd3, 2; add.s64 %rd747, %rd745, %rd746; ld.global.u32 %r994, [%rd747]; st.local.u32 [%rd7+-40], %r994; fma.rn.ftz.f32 %f4925, %f15602, %f9, %f15596; fma.rn.ftz.f32 %f4926, %f15602, %f8, %f15597; fma.rn.ftz.f32 %f4927, %f15602, %f7, %f15598; st.local.v2.f32 [%rd1], {%f4927, %f4926}; st.local.f32 [%rd7+-80], %f4925; and.b32 %r995, %r1, 64; setp.eq.s32 %p47, %r995, 0; @%p47 bra $L__BB108_61; bra.uni $L__BB108_48; $L__BB108_61: ld.const.u64 %rd771, [kernel_params+2296]; cvta.to.global.u64 %rd772, %rd771; shl.b64 %rd773, %rd3, 4; add.s64 %rd774, %rd772, %rd773; ld.global.v4.f32 {%f15663, %f15664, %f15665, %f4952}, [%rd774]; bra.uni $L__BB108_62; $L__BB108_22: setp.gt.s32 %p32, %r10, %r9; selp.b32 %r979, -1, 0, %p32; add.s32 %r980, %r10, %r979; mad.lo.s32 %r981, %r980, %r8, %r3066; ld.const.u64 %rd690, [kernel_params+2344]; cvta.to.global.u64 %rd691, %rd690; add.s32 %r982, %r981, %r6; mul.wide.s32 %rd692, %r982, 16; add.s64 %rd693, %rd691, %rd692; ld.global.v4.f32 {%f15614, %f15613, %f15612, %f4585}, [%rd693]; add.s32 %r983, %r981, %r5; mul.wide.s32 %rd694, %r983, 16; add.s64 %rd695, %rd691, %rd694; ld.global.v4.f32 {%f15611, %f15610, %f15609, %f4589}, [%rd695]; ld.global.v4.f32 {%f15608, %f15607, %f15606, %f4593}, [%rd695+16]; add.s32 %r984, %r981, %r7; mul.wide.s32 %rd696, %r984, 16; add.s64 %rd697, %rd691, %rd696; ld.global.v4.f32 {%f15605, %f15604, %f15603, %f4597}, [%rd697]; $L__BB108_24: add.s32 %r21, %r10, 1; setp.eq.s32 %p33, %r21, %r9; @%p33 bra $L__BB108_26; bra.uni $L__BB108_25; $L__BB108_26: ld.const.u64 %rd714, [kernel_params+2272]; cvta.to.global.u64 %rd715, %rd714; mul.wide.s32 %rd716, %r6, 16; add.s64 %rd717, %rd715, %rd716; ld.global.v4.f32 {%f15626, %f15625, %f15624, %f4633}, [%rd717]; mul.wide.s32 %rd718, %r5, 16; add.s64 %rd719, %rd715, %rd718; ld.global.v4.f32 {%f15623, %f15622, %f15621, %f4637}, [%rd719]; ld.global.v4.f32 {%f15620, %f15619, %f15618, %f4641}, [%rd719+16]; mul.wide.s32 %rd720, %r7, 16; add.s64 %rd721, %rd715, %rd720; ld.global.v4.f32 {%f15617, %f15616, %f15615, %f4645}, [%rd721]; bra.uni $L__BB108_27; $L__BB108_25: setp.ge.s32 %p34, %r10, %r9; selp.b32 %r985, -1, 0, %p34; add.s32 %r986, %r21, %r985; mad.lo.s32 %r987, %r986, %r8, %r3066; ld.const.u64 %rd706, [kernel_params+2344]; cvta.to.global.u64 %rd707, %rd706; add.s32 %r988, %r987, %r6; mul.wide.s32 %rd708, %r988, 16; add.s64 %rd709, %rd707, %rd708; ld.global.v4.f32 {%f15626, %f15625, %f15624, %f4617}, [%rd709]; add.s32 %r989, %r987, %r5; mul.wide.s32 %rd710, %r989, 16; add.s64 %rd711, %rd707, %rd710; ld.global.v4.f32 {%f15623, %f15622, %f15621, %f4621}, [%rd711]; ld.global.v4.f32 {%f15620, %f15619, %f15618, %f4625}, [%rd711+16]; add.s32 %r990, %r987, %r7; mul.wide.s32 %rd712, %r990, 16; add.s64 %rd713, %rd707, %rd712; ld.global.v4.f32 {%f15617, %f15616, %f15615, %f4629}, [%rd713]; $L__BB108_27: shl.b32 %r3044, %r9, 1; cvt.rn.f32.s32 %f15463, %r3044; mul.ftz.f32 %f15462, %f10, %f15463; cvt.rn.f32.s32 %f4646, %r10; sub.ftz.f32 %f4647, %f15462, %f4646; mov.f32 %f4648, 0f3F800000; sub.ftz.f32 %f4649, %f4648, %f4647; mul.ftz.f32 %f4650, %f4647, %f15626; fma.rn.ftz.f32 %f15638, %f4649, %f15614, %f4650; mul.ftz.f32 %f4651, %f4647, %f15625; fma.rn.ftz.f32 %f15637, %f4649, %f15613, %f4651; mul.ftz.f32 %f4652, %f4647, %f15624; fma.rn.ftz.f32 %f15636, %f4649, %f15612, %f4652; mul.ftz.f32 %f4653, %f4647, %f15623; fma.rn.ftz.f32 %f15635, %f4649, %f15611, %f4653; mul.ftz.f32 %f4654, %f4647, %f15622; fma.rn.ftz.f32 %f15634, %f4649, %f15610, %f4654; mul.ftz.f32 %f4655, %f4647, %f15621; fma.rn.ftz.f32 %f15633, %f4649, %f15609, %f4655; mul.ftz.f32 %f4656, %f4647, %f15620; fma.rn.ftz.f32 %f15632, %f4649, %f15608, %f4656; mul.ftz.f32 %f4657, %f4647, %f15619; fma.rn.ftz.f32 %f15631, %f4649, %f15607, %f4657; mul.ftz.f32 %f4658, %f4647, %f15618; fma.rn.ftz.f32 %f15630, %f4649, %f15606, %f4658; mul.ftz.f32 %f4659, %f4647, %f15617; fma.rn.ftz.f32 %f15629, %f4649, %f15605, %f4659; mul.ftz.f32 %f4660, %f4647, %f15616; fma.rn.ftz.f32 %f15628, %f4649, %f15604, %f4660; mul.ftz.f32 %f4661, %f4647, %f15615; fma.rn.ftz.f32 %f15627, %f4649, %f15603, %f4661; $L__BB108_29: fma.rn.ftz.f32 %f15659, %f15601, %f15602, %f15598; fma.rn.ftz.f32 %f15658, %f15600, %f15602, %f15597; fma.rn.ftz.f32 %f15657, %f15599, %f15602, %f15596; mov.f32 %f4678, 0f3F800000; sub.ftz.f32 %f330, %f4678, %f1; mul.ftz.f32 %f331, %f330, %f330; add.ftz.f32 %f4679, %f330, %f330; mul.ftz.f32 %f4680, %f1, %f4679; sub.ftz.f32 %f4681, %f4680, %f331; mul.ftz.f32 %f4682, %f1, 0f40400000; add.ftz.f32 %f332, %f4682, 0fC0A00000; add.ftz.f32 %f4683, %f1, %f1; mul.ftz.f32 %f4684, %f4683, %f332; fma.rn.ftz.f32 %f4685, %f1, %f4682, %f4684; add.ftz.f32 %f4686, %f4682, 0f40000000; mul.ftz.f32 %f4687, %f4679, %f4686; mul.ftz.f32 %f333, %f330, 0f40400000; mul.ftz.f32 %f4688, %f330, %f333; sub.ftz.f32 %f4689, %f4687, %f4688; mul.ftz.f32 %f4690, %f330, 0fC0000000; mul.ftz.f32 %f334, %f1, %f1; fma.rn.ftz.f32 %f4691, %f1, %f4690, %f334; mul.ftz.f32 %f4692, %f15635, %f4685; mul.ftz.f32 %f4693, %f15634, %f4685; mul.ftz.f32 %f4694, %f15633, %f4685; fma.rn.ftz.f32 %f4695, %f15638, %f4681, %f4692; fma.rn.ftz.f32 %f4696, %f15637, %f4681, %f4693; fma.rn.ftz.f32 %f4697, %f15636, %f4681, %f4694; fma.rn.ftz.f32 %f4698, %f15632, %f4689, %f4695; fma.rn.ftz.f32 %f4699, %f15631, %f4689, %f4696; fma.rn.ftz.f32 %f4700, %f15630, %f4689, %f4697; fma.rn.ftz.f32 %f4701, %f15629, %f4691, %f4698; fma.rn.ftz.f32 %f4702, %f15628, %f4691, %f4699; fma.rn.ftz.f32 %f4703, %f15627, %f4691, %f4700; mul.ftz.f32 %f15651, %f4701, 0f3F000000; mul.ftz.f32 %f15653, %f4702, 0f3F000000; mul.ftz.f32 %f15652, %f4703, 0f3F000000; and.b32 %r22, %r1, 4; setp.eq.s32 %p35, %r22, 0; @%p35 bra $L__BB108_31; bra.uni $L__BB108_30; $L__BB108_31: mul.ftz.f32 %f4762, %f1, %f330; mul.ftz.f32 %f4763, %f330, %f4762; fma.rn.ftz.f32 %f4764, %f334, %f332, 0f40000000; add.ftz.f32 %f4765, %f333, 0fC0A00000; fma.rn.ftz.f32 %f4766, %f331, %f4765, 0f40000000; mul.ftz.f32 %f4767, %f1, %f4762; mul.ftz.f32 %f4768, %f15638, %f4763; mul.ftz.f32 %f4769, %f15637, %f4763; mul.ftz.f32 %f4770, %f15636, %f4763; mul.ftz.f32 %f4771, %f15635, %f4764; mul.ftz.f32 %f4772, %f15634, %f4764; mul.ftz.f32 %f4773, %f15633, %f4764; sub.ftz.f32 %f4774, %f4771, %f4768; sub.ftz.f32 %f4775, %f4772, %f4769; sub.ftz.f32 %f4776, %f4773, %f4770; fma.rn.ftz.f32 %f4777, %f15632, %f4766, %f4774; fma.rn.ftz.f32 %f4778, %f15631, %f4766, %f4775; fma.rn.ftz.f32 %f4779, %f15630, %f4766, %f4776; mul.ftz.f32 %f4780, %f15629, %f4767; mul.ftz.f32 %f4781, %f15628, %f4767; mul.ftz.f32 %f4782, %f15627, %f4767; sub.ftz.f32 %f4783, %f4777, %f4780; sub.ftz.f32 %f4784, %f4778, %f4781; sub.ftz.f32 %f4785, %f4779, %f4782; mul.ftz.f32 %f341, %f4783, 0f3F000000; mul.ftz.f32 %f342, %f4784, 0f3F000000; mul.ftz.f32 %f343, %f4785, 0f3F000000; setp.eq.ftz.f32 %p36, %f15659, %f341; setp.eq.ftz.f32 %p37, %f15658, %f342; and.pred %p38, %p36, %p37; setp.eq.ftz.f32 %p39, %f15657, %f343; and.pred %p40, %p39, %p38; @%p40 bra $L__BB108_33; bra.uni $L__BB108_32; $L__BB108_33: mov.f32 %f15656, %f7; mov.f32 %f15655, %f8; mov.f32 %f15654, %f9; bra.uni $L__BB108_34; $L__BB108_30: mul.ftz.f32 %f4704, %f15653, %f15653; fma.rn.ftz.f32 %f4705, %f15651, %f15651, %f4704; fma.rn.ftz.f32 %f4706, %f15652, %f15652, %f4705; sqrt.approx.ftz.f32 %f4707, %f4706; div.approx.ftz.f32 %f4708, %f15651, %f4707; div.approx.ftz.f32 %f4709, %f15653, %f4707; div.approx.ftz.f32 %f4710, %f15652, %f4707; mul.ftz.f32 %f4711, %f4709, %f15599; mul.ftz.f32 %f4712, %f4710, %f15600; sub.ftz.f32 %f4713, %f4712, %f4711; mul.ftz.f32 %f4714, %f4710, %f15601; mul.ftz.f32 %f4715, %f4708, %f15599; sub.ftz.f32 %f4716, %f4715, %f4714; mul.ftz.f32 %f4717, %f4708, %f15600; mul.ftz.f32 %f4718, %f4709, %f15601; sub.ftz.f32 %f4719, %f4718, %f4717; mul.ftz.f32 %f4720, %f4716, %f4716; fma.rn.ftz.f32 %f4721, %f4713, %f4713, %f4720; fma.rn.ftz.f32 %f4722, %f4719, %f4719, %f4721; sqrt.approx.ftz.f32 %f4723, %f4722; div.approx.ftz.f32 %f4724, %f4713, %f4723; div.approx.ftz.f32 %f4725, %f4716, %f4723; div.approx.ftz.f32 %f4726, %f4719, %f4723; mul.ftz.f32 %f4727, %f2, %f2; sub.ftz.f32 %f4729, %f4678, %f4727; mov.f32 %f4730, 0f00000000; max.ftz.f32 %f4731, %f4729, %f4730; sqrt.approx.ftz.f32 %f4732, %f4731; mul.ftz.f32 %f4733, %f4724, %f2; mul.ftz.f32 %f4734, %f4725, %f2; mul.ftz.f32 %f4735, %f4726, %f2; mul.ftz.f32 %f4736, %f4709, %f4726; mul.ftz.f32 %f4737, %f4710, %f4725; sub.ftz.f32 %f4738, %f4736, %f4737; mul.ftz.f32 %f4739, %f4710, %f4724; mul.ftz.f32 %f4740, %f4708, %f4726; sub.ftz.f32 %f4741, %f4739, %f4740; mul.ftz.f32 %f4742, %f4708, %f4725; mul.ftz.f32 %f4743, %f4709, %f4724; sub.ftz.f32 %f4744, %f4742, %f4743; mul.ftz.f32 %f4745, %f4741, %f4741; fma.rn.ftz.f32 %f4746, %f4738, %f4738, %f4745; fma.rn.ftz.f32 %f4747, %f4744, %f4744, %f4746; sqrt.approx.ftz.f32 %f4748, %f4747; div.approx.ftz.f32 %f4749, %f4738, %f4748; div.approx.ftz.f32 %f4750, %f4741, %f4748; div.approx.ftz.f32 %f4751, %f4744, %f4748; mul.ftz.f32 %f4752, %f4732, %f4749; mul.ftz.f32 %f4753, %f4732, %f4750; mul.ftz.f32 %f4754, %f4732, %f4751; sub.ftz.f32 %f4755, %f4733, %f4752; sub.ftz.f32 %f4756, %f4734, %f4753; sub.ftz.f32 %f4757, %f4735, %f4754; mul.ftz.f32 %f4758, %f4756, %f4756; fma.rn.ftz.f32 %f4759, %f4755, %f4755, %f4758; fma.rn.ftz.f32 %f4760, %f4757, %f4757, %f4759; sqrt.approx.ftz.f32 %f4761, %f4760; div.approx.ftz.f32 %f15656, %f4755, %f4761; div.approx.ftz.f32 %f15655, %f4756, %f4761; div.approx.ftz.f32 %f15654, %f4757, %f4761; st.local.f32 [%rd7+-76], %f15656; st.local.f32 [%rd7+-72], %f15655; st.local.f32 [%rd7+-68], %f15654; bra.uni $L__BB108_35; $L__BB108_32: sub.ftz.f32 %f4786, %f15659, %f341; sub.ftz.f32 %f4787, %f15658, %f342; mul.ftz.f32 %f4788, %f4787, %f4787; fma.rn.ftz.f32 %f4789, %f4786, %f4786, %f4788; sub.ftz.f32 %f4790, %f15657, %f343; fma.rn.ftz.f32 %f4791, %f4790, %f4790, %f4789; sqrt.approx.ftz.f32 %f4792, %f4791; div.approx.ftz.f32 %f15654, %f4790, %f4792; div.approx.ftz.f32 %f15655, %f4787, %f4792; div.approx.ftz.f32 %f15656, %f4786, %f4792; $L__BB108_34: mov.u32 %r3041, 0; st.local.f32 [%rd7+-76], %f15656; st.local.f32 [%rd7+-72], %f15655; st.local.f32 [%rd7+-68], %f15654; st.local.u32 [%rd7+-16], %r3041; $L__BB108_35: and.b32 %r3045, %r3072, 4; setp.ne.s32 %p2181, %r3045, 0; st.local.f32 [%rd7+28], %f15651; st.local.f32 [%rd7+32], %f15653; st.local.f32 [%rd7+36], %f15652; @%p2181 bra $L__BB108_45; @%p15 bra $L__BB108_38; bra.uni $L__BB108_37; $L__BB108_38: ld.const.u64 %rd730, [kernel_params+2160]; cvta.to.global.u64 %rd731, %rd730; mul.lo.s64 %rd732, %rd9, 208; add.s64 %rd733, %rd731, %rd732; ld.global.v4.f32 {%f4802, %f4803, %f4804, %f4805}, [%rd733]; ld.global.v4.f32 {%f4810, %f4811, %f4812, %f4813}, [%rd733+16]; ld.global.v4.f32 {%f4818, %f4819, %f4820, %f4821}, [%rd733+32]; mul.ftz.f32 %f4826, %f15658, %f4803; fma.rn.ftz.f32 %f4827, %f15659, %f4802, %f4826; fma.rn.ftz.f32 %f4828, %f15657, %f4804, %f4827; add.ftz.f32 %f15647, %f4805, %f4828; mul.ftz.f32 %f4829, %f15658, %f4811; fma.rn.ftz.f32 %f4830, %f15659, %f4810, %f4829; fma.rn.ftz.f32 %f4831, %f15657, %f4812, %f4830; add.ftz.f32 %f15646, %f4813, %f4831; mul.ftz.f32 %f4832, %f15658, %f4819; fma.rn.ftz.f32 %f4833, %f15659, %f4818, %f4832; fma.rn.ftz.f32 %f4834, %f15657, %f4820, %f4833; add.ftz.f32 %f15657, %f4821, %f4834; bra.uni $L__BB108_39; $L__BB108_37: mul.ftz.f32 %f4793, %f15658, %f142; fma.rn.ftz.f32 %f4794, %f15659, %f141, %f4793; fma.rn.ftz.f32 %f4795, %f15657, %f143, %f4794; add.ftz.f32 %f15647, %f144, %f4795; mul.ftz.f32 %f4796, %f15658, %f146; fma.rn.ftz.f32 %f4797, %f15659, %f145, %f4796; fma.rn.ftz.f32 %f4798, %f15657, %f147, %f4797; add.ftz.f32 %f15646, %f148, %f4798; mul.ftz.f32 %f4799, %f15658, %f150; fma.rn.ftz.f32 %f4800, %f15659, %f149, %f4799; fma.rn.ftz.f32 %f4801, %f15657, %f151, %f4800; add.ftz.f32 %f15657, %f152, %f4801; $L__BB108_39: @%p15 bra $L__BB108_41; bra.uni $L__BB108_40; $L__BB108_41: ld.const.u64 %rd734, [kernel_params+2160]; cvta.to.global.u64 %rd735, %rd734; mul.lo.s64 %rd736, %rd9, 208; add.s64 %rd737, %rd735, %rd736; ld.global.v4.f32 {%f4848, %f4849, %f4850, %f4851}, [%rd737+48]; ld.global.v4.f32 {%f4855, %f4856, %f4857, %f4858}, [%rd737+64]; ld.global.v4.f32 {%f4862, %f4863, %f4864, %f4865}, [%rd737+80]; mul.ftz.f32 %f4869, %f4855, %f15655; fma.rn.ftz.f32 %f4870, %f4848, %f15656, %f4869; fma.rn.ftz.f32 %f4871, %f4862, %f15654, %f4870; mul.ftz.f32 %f4872, %f4856, %f15655; fma.rn.ftz.f32 %f4873, %f4849, %f15656, %f4872; fma.rn.ftz.f32 %f4874, %f4863, %f15654, %f4873; mul.ftz.f32 %f4875, %f4857, %f15655; fma.rn.ftz.f32 %f4876, %f4850, %f15656, %f4875; fma.rn.ftz.f32 %f4877, %f4864, %f15654, %f4876; mul.ftz.f32 %f4878, %f4874, %f4874; fma.rn.ftz.f32 %f4879, %f4871, %f4871, %f4878; fma.rn.ftz.f32 %f4880, %f4877, %f4877, %f4879; sqrt.approx.ftz.f32 %f4881, %f4880; div.approx.ftz.f32 %f15656, %f4871, %f4881; div.approx.ftz.f32 %f15655, %f4874, %f4881; div.approx.ftz.f32 %f15654, %f4877, %f4881; bra.uni $L__BB108_42; $L__BB108_40: mul.ftz.f32 %f4835, %f15655, %f15573; fma.rn.ftz.f32 %f4836, %f15656, %f15569, %f4835; fma.rn.ftz.f32 %f4837, %f15654, %f15577, %f4836; mul.ftz.f32 %f4838, %f15655, %f15574; fma.rn.ftz.f32 %f4839, %f15656, %f15570, %f4838; fma.rn.ftz.f32 %f4840, %f15654, %f15578, %f4839; mul.ftz.f32 %f4841, %f15655, %f15575; fma.rn.ftz.f32 %f4842, %f15656, %f15571, %f4841; fma.rn.ftz.f32 %f4843, %f15654, %f15579, %f4842; mul.ftz.f32 %f4844, %f4840, %f4840; fma.rn.ftz.f32 %f4845, %f4837, %f4837, %f4844; fma.rn.ftz.f32 %f4846, %f4843, %f4843, %f4845; sqrt.approx.ftz.f32 %f4847, %f4846; div.approx.ftz.f32 %f15656, %f4837, %f4847; div.approx.ftz.f32 %f15655, %f4840, %f4847; div.approx.ftz.f32 %f15654, %f4843, %f4847; $L__BB108_42: st.local.f32 [%rd7+-76], %f15656; st.local.f32 [%rd7+-72], %f15655; st.local.f32 [%rd7+-68], %f15654; @%p15 bra $L__BB108_44; bra.uni $L__BB108_43; $L__BB108_44: ld.const.u64 %rd738, [kernel_params+2160]; cvta.to.global.u64 %rd739, %rd738; mul.lo.s64 %rd740, %rd9, 208; add.s64 %rd741, %rd739, %rd740; ld.global.v4.f32 {%f4888, %f4889, %f4890, %f4891}, [%rd741]; ld.global.v4.f32 {%f4895, %f4896, %f4897, %f4898}, [%rd741+16]; ld.global.v4.f32 {%f4902, %f4903, %f4904, %f4905}, [%rd741+32]; mul.ftz.f32 %f4909, %f4889, %f15653; fma.rn.ftz.f32 %f4910, %f15651, %f4888, %f4909; fma.rn.ftz.f32 %f380, %f4890, %f15652, %f4910; mul.ftz.f32 %f4911, %f4896, %f15653; fma.rn.ftz.f32 %f4912, %f15651, %f4895, %f4911; fma.rn.ftz.f32 %f381, %f4897, %f15652, %f4912; mul.ftz.f32 %f4913, %f4903, %f15653; fma.rn.ftz.f32 %f4914, %f15651, %f4902, %f4913; fma.rn.ftz.f32 %f15652, %f4904, %f15652, %f4914; st.local.f32 [%rd7+28], %f380; st.local.f32 [%rd7+32], %f381; st.local.f32 [%rd7+36], %f15652; mov.f32 %f15651, %f380; mov.f32 %f15653, %f381; mov.f32 %f15658, %f15646; mov.f32 %f15659, %f15647; bra.uni $L__BB108_45; $L__BB108_43: mul.ftz.f32 %f4882, %f15653, %f142; fma.rn.ftz.f32 %f4883, %f15651, %f141, %f4882; fma.rn.ftz.f32 %f377, %f15652, %f143, %f4883; mul.ftz.f32 %f4884, %f15653, %f146; fma.rn.ftz.f32 %f4885, %f15651, %f145, %f4884; fma.rn.ftz.f32 %f378, %f15652, %f147, %f4885; mul.ftz.f32 %f4886, %f15653, %f150; fma.rn.ftz.f32 %f4887, %f15651, %f149, %f4886; fma.rn.ftz.f32 %f15652, %f15652, %f151, %f4887; st.local.f32 [%rd7+28], %f377; st.local.f32 [%rd7+32], %f378; st.local.f32 [%rd7+36], %f15652; mov.f32 %f15651, %f377; mov.f32 %f15653, %f378; mov.f32 %f15658, %f15646; mov.f32 %f15659, %f15647; $L__BB108_45: neg.ftz.f32 %f15461, %f7; neg.ftz.f32 %f15460, %f8; neg.ftz.f32 %f15459, %f9; st.local.v2.f32 [%rd1], {%f15659, %f15658}; st.local.f32 [%rd7+-80], %f15657; selp.f32 %f4915, %f15654, %f15459, %p35; selp.f32 %f15697, %f15655, %f15460, %p35; selp.f32 %f15696, %f15656, %f15461, %p35; st.local.v2.f32 [%rd7+-64], {%f15696, %f15697}; st.local.f32 [%rd7+-56], %f4915; mul.ftz.f32 %f4916, %f15653, %f4915; mul.ftz.f32 %f4917, %f15652, %f15697; mul.ftz.f32 %f4918, %f15652, %f15696; mul.ftz.f32 %f4919, %f15651, %f4915; mul.ftz.f32 %f4920, %f15651, %f15697; mul.ftz.f32 %f4921, %f15653, %f15696; sub.ftz.f32 %f4922, %f4920, %f4921; sub.ftz.f32 %f4923, %f4918, %f4919; sub.ftz.f32 %f4924, %f4916, %f4917; st.local.v2.f32 [%rd7+40], {%f4924, %f4923}; st.local.f32 [%rd7+48], %f4922; ld.global.u32 %r992, [%rd11+-4]; st.local.u32 [%rd7+-40], %r992; bra.uni $L__BB108_95; $L__BB108_68: { // callseq 692, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 692 ld.local.u32 %r3072, [%rd7+-32]; bra.uni $L__BB108_81; $L__BB108_48: ld.const.u64 %rd750, [kernel_params+2160]; cvta.to.global.u64 %rd751, %rd750; mul.lo.s64 %rd752, %rd9, 208; add.s64 %rd753, %rd751, %rd752; ld.global.u32 %r23, [%rd753+148]; ld.global.u32 %r24, [%rd753+152]; shl.b32 %r996, %r24, 1; cvt.rn.f32.s32 %f4928, %r996; mul.ftz.f32 %f394, %f10, %f4928; cvt.rzi.ftz.s32.f32 %r997, %f394; add.s32 %r998, %r996, -1; min.s32 %r25, %r997, %r998; ld.const.u64 %rd754, [kernel_params+2312]; cvta.to.global.u64 %rd16, %rd754; ld.global.u32 %r3068, [%rd753+164]; mul.wide.u32 %rd755, %r3068, 16; add.s64 %rd756, %rd16, %rd755; ld.global.u16 %rs135, [%rd756+12]; ld.global.u32 %r3067, [%rd756+8]; ld.global.u64 %rd2924, [%rd756]; setp.eq.s64 %p48, %rd2924, 11; @%p48 bra $L__BB108_53; $L__BB108_49: setp.eq.s64 %p49, %rd2924, 0; @%p49 bra $L__BB108_51; add.s32 %r3068, %r3068, 2; bra.uni $L__BB108_52; $L__BB108_51: setp.eq.s16 %p50, %rs135, 0; mov.u32 %r3071, -1; mov.u32 %r3068, %r3067; @%p50 bra $L__BB108_54; $L__BB108_52: mul.wide.u32 %rd757, %r3068, 16; add.s64 %rd758, %rd16, %rd757; ld.global.u16 %rs135, [%rd758+12]; ld.global.u32 %r3067, [%rd758+8]; ld.global.u64 %rd2924, [%rd758]; setp.ne.s64 %p51, %rd2924, 11; @%p51 bra $L__BB108_49; $L__BB108_53: setp.eq.s16 %p52, %rs135, 0; selp.b32 %r3071, -1, %r3067, %p52; $L__BB108_54: setp.eq.s32 %p53, %r25, %r24; @%p53 bra $L__BB108_56; bra.uni $L__BB108_55; $L__BB108_56: ld.const.u64 %rd762, [kernel_params+2296]; cvta.to.global.u64 %rd763, %rd762; shl.b64 %rd764, %rd3, 4; add.s64 %rd2925, %rd763, %rd764; bra.uni $L__BB108_57; $L__BB108_70: sub.ftz.f32 %f5017, %f434, %f435; sub.ftz.f32 %f5018, %f436, %f437; mul.ftz.f32 %f5019, %f5018, %f5018; fma.rn.ftz.f32 %f5020, %f5017, %f5017, %f5019; sub.ftz.f32 %f5021, %f438, %f439; fma.rn.ftz.f32 %f5022, %f5021, %f5021, %f5020; sqrt.approx.ftz.f32 %f5023, %f5022; div.approx.ftz.f32 %f15668, %f5021, %f5023; div.approx.ftz.f32 %f15667, %f5018, %f5023; div.approx.ftz.f32 %f15666, %f5017, %f5023; $L__BB108_72: ld.const.u64 %rd795, [kernel_params+2216]; cvta.to.global.u64 %rd796, %rd795; shl.b64 %rd797, %rd3, 2; add.s64 %rd798, %rd796, %rd797; ld.global.u32 %r38, [%rd798]; st.local.u32 [%rd7+-40], %r38; fma.rn.ftz.f32 %f5031, %f1, %f431, %f425; fma.rn.ftz.f32 %f5032, %f1, %f432, %f427; fma.rn.ftz.f32 %f5033, %f1, %f433, %f429; fma.rn.ftz.f32 %f15681, %f2, %f426, %f5031; fma.rn.ftz.f32 %f15682, %f2, %f428, %f5032; fma.rn.ftz.f32 %f15683, %f2, %f430, %f5033; and.b32 %r1017, %r3072, 4; setp.ne.s32 %p61, %r1017, 0; @%p61 bra $L__BB108_76; and.b32 %r3043, %r3072, 2; setp.ne.s32 %p62, %r3043, 0; @%p62 bra $L__BB108_75; ld.const.u64 %rd799, [kernel_params+2160]; cvta.to.global.u64 %rd800, %rd799; mul.lo.s64 %rd801, %rd9, 208; add.s64 %rd802, %rd800, %rd801; ld.global.v4.f32 {%f141, %f142, %f143, %f144}, [%rd802]; ld.global.v4.f32 {%f145, %f146, %f147, %f148}, [%rd802+16]; ld.global.v4.f32 {%f149, %f150, %f151, %f152}, [%rd802+32]; $L__BB108_75: mul.ftz.f32 %f5046, %f15682, %f142; fma.rn.ftz.f32 %f5047, %f15681, %f141, %f5046; fma.rn.ftz.f32 %f5048, %f15683, %f143, %f5047; mul.ftz.f32 %f5049, %f15682, %f146; fma.rn.ftz.f32 %f5050, %f15681, %f145, %f5049; fma.rn.ftz.f32 %f5051, %f15683, %f147, %f5050; mul.ftz.f32 %f5052, %f15682, %f150; fma.rn.ftz.f32 %f5053, %f15681, %f149, %f5052; fma.rn.ftz.f32 %f5054, %f15683, %f151, %f5053; add.ftz.f32 %f15681, %f144, %f5048; add.ftz.f32 %f15682, %f148, %f5051; add.ftz.f32 %f15683, %f152, %f5054; $L__BB108_76: st.local.f32 [%rd7+-56], %f15668; st.local.v4.f32 [%rd1], {%f15681, %f15682, %f15683, %f15666}; st.local.v4.f32 [%rd7+-72], {%f15667, %f15668, %f15666, %f15667}; setp.gt.s32 %p63, %r38, -1; @%p63 bra $L__BB108_80; ld.global.v4.u32 {%r1018, %r1019, %r1020, %r1021}, [%rd27+-12]; ld.const.u64 %rd805, [kernel_params+2224]; cvta.to.global.u64 %rd806, %rd805; mul.wide.u32 %rd807, %r1018, 12; add.s64 %rd808, %rd806, %rd807; mul.wide.u32 %rd809, %r1019, 12; add.s64 %rd810, %rd806, %rd809; mul.wide.u32 %rd811, %r1020, 12; add.s64 %rd812, %rd806, %rd811; mov.f32 %f5055, 0f3F800000; sub.ftz.f32 %f5056, %f5055, %f1; sub.ftz.f32 %f5057, %f5056, %f2; ld.global.f32 %f5058, [%rd808]; ld.global.f32 %f5059, [%rd808+4]; ld.global.f32 %f5060, [%rd808+8]; ld.global.f32 %f5061, [%rd810]; mul.ftz.f32 %f5062, %f1, %f5061; ld.global.f32 %f5063, [%rd810+4]; mul.ftz.f32 %f5064, %f1, %f5063; ld.global.f32 %f5065, [%rd810+8]; mul.ftz.f32 %f5066, %f1, %f5065; fma.rn.ftz.f32 %f5067, %f5057, %f5058, %f5062; fma.rn.ftz.f32 %f5068, %f5057, %f5059, %f5064; fma.rn.ftz.f32 %f5069, %f5057, %f5060, %f5066; ld.global.f32 %f5070, [%rd812]; ld.global.f32 %f5071, [%rd812+4]; ld.global.f32 %f5072, [%rd812+8]; fma.rn.ftz.f32 %f15684, %f2, %f5070, %f5067; fma.rn.ftz.f32 %f15685, %f2, %f5071, %f5068; fma.rn.ftz.f32 %f15686, %f2, %f5072, %f5069; mul.ftz.f32 %f5073, %f15685, %f15685; fma.rn.ftz.f32 %f5074, %f15684, %f15684, %f5073; fma.rn.ftz.f32 %f5075, %f15686, %f15686, %f5074; sqrt.approx.ftz.f32 %f488, %f5075; setp.neu.ftz.f32 %p64, %f488, 0f00000000; @%p64 bra $L__BB108_78; bra.uni $L__BB108_79; $L__BB108_78: rcp.approx.ftz.f32 %f5076, %f488; mul.ftz.f32 %f15686, %f15686, %f5076; mul.ftz.f32 %f15685, %f15685, %f5076; mul.ftz.f32 %f15684, %f15684, %f5076; $L__BB108_79: setp.eq.ftz.f32 %p65, %f15684, 0f00000000; setp.eq.ftz.f32 %p66, %f15685, 0f00000000; and.pred %p67, %p65, %p66; setp.eq.ftz.f32 %p68, %f15686, 0f00000000; and.pred %p69, %p68, %p67; selp.f32 %f5077, %f15668, %f15686, %p69; selp.f32 %f5078, %f15667, %f15685, %p69; selp.f32 %f5079, %f15666, %f15684, %p69; st.local.f32 [%rd7+-76], %f5079; st.local.f32 [%rd7+-72], %f5078; st.local.f32 [%rd7+-68], %f5077; $L__BB108_80: st.local.f32 [%rd7+28], %f431; st.local.f32 [%rd7+32], %f432; st.local.f32 [%rd7+36], %f433; st.local.v2.f32 [%rd7+40], {%f426, %f428}; st.local.f32 [%rd7+48], %f430; $L__BB108_81: and.b32 %r1025, %r3072, 4; setp.eq.s32 %p70, %r1025, 0; @%p70 bra $L__BB108_83; bra.uni $L__BB108_82; $L__BB108_83: and.b32 %r40, %r3072, 2; setp.eq.s32 %p71, %r40, 0; @%p71 bra $L__BB108_85; bra.uni $L__BB108_84; $L__BB108_85: ld.const.u64 %rd813, [kernel_params+2160]; cvta.to.global.u64 %rd814, %rd813; ld.local.u32 %r1026, [%rd7+-12]; mul.wide.s32 %rd815, %r1026, 208; add.s64 %rd816, %rd814, %rd815; ld.global.v4.f32 {%f5121, %f5122, %f5123, %f5124}, [%rd816+48]; ld.global.v4.f32 {%f5128, %f5129, %f5130, %f5131}, [%rd816+64]; ld.global.v4.f32 {%f5135, %f5136, %f5137, %f5138}, [%rd816+80]; ld.local.v2.f32 {%f5142, %f5143}, [%rd7+-72]; ld.local.f32 %f5146, [%rd7+-76]; mul.ftz.f32 %f5147, %f5128, %f5142; fma.rn.ftz.f32 %f5148, %f5121, %f5146, %f5147; fma.rn.ftz.f32 %f5149, %f5135, %f5143, %f5148; mul.ftz.f32 %f5150, %f5129, %f5142; fma.rn.ftz.f32 %f5151, %f5122, %f5146, %f5150; fma.rn.ftz.f32 %f5152, %f5136, %f5143, %f5151; mul.ftz.f32 %f5153, %f5130, %f5142; fma.rn.ftz.f32 %f5154, %f5123, %f5146, %f5153; fma.rn.ftz.f32 %f5155, %f5137, %f5143, %f5154; mul.ftz.f32 %f5156, %f5152, %f5152; fma.rn.ftz.f32 %f5157, %f5149, %f5149, %f5156; fma.rn.ftz.f32 %f5158, %f5155, %f5155, %f5157; sqrt.approx.ftz.f32 %f5159, %f5158; div.approx.ftz.f32 %f15687, %f5149, %f5159; div.approx.ftz.f32 %f15688, %f5152, %f5159; div.approx.ftz.f32 %f15689, %f5155, %f5159; bra.uni $L__BB108_86; $L__BB108_82: ld.local.v2.f32 {%f15696, %f15697}, [%rd7+-64]; bra.uni $L__BB108_95; $L__BB108_55: cvt.u32.u64 %r1000, %rd3; setp.gt.s32 %p54, %r25, %r24; selp.b32 %r1001, -1, 0, %p54; add.s32 %r1002, %r25, %r1001; ld.const.u64 %rd759, [kernel_params+2344]; cvta.to.global.u64 %rd760, %rd759; mad.lo.s32 %r1003, %r1002, %r23, %r1000; add.s32 %r1004, %r1003, %r3071; mul.wide.s32 %rd761, %r1004, 16; add.s64 %rd2925, %rd760, %rd761; $L__BB108_57: ld.global.v4.f32 {%f4929, %f4930, %f4931, %f4932}, [%rd2925]; add.s32 %r36, %r25, 1; setp.eq.s32 %p55, %r36, %r24; @%p55 bra $L__BB108_59; bra.uni $L__BB108_58; $L__BB108_59: ld.const.u64 %rd768, [kernel_params+2296]; cvta.to.global.u64 %rd769, %rd768; shl.b64 %rd770, %rd3, 4; add.s64 %rd2926, %rd769, %rd770; bra.uni $L__BB108_60; $L__BB108_84: ld.local.v2.f32 {%f5082, %f5083}, [%rd7+-72]; ld.local.v4.f32 {%f5086, %f5087, %f5088, %f5089}, [%rd7+104]; ld.local.v4.f32 {%f5093, %f5094, %f5095, %f5096}, [%rd7+120]; ld.local.v4.f32 {%f5100, %f5101, %f5102, %f5103}, [%rd7+136]; ld.local.f32 %f5107, [%rd7+-76]; mul.ftz.f32 %f5108, %f5082, %f5093; fma.rn.ftz.f32 %f5109, %f5107, %f5086, %f5108; fma.rn.ftz.f32 %f5110, %f5083, %f5100, %f5109; mul.ftz.f32 %f5111, %f5082, %f5094; fma.rn.ftz.f32 %f5112, %f5107, %f5087, %f5111; fma.rn.ftz.f32 %f5113, %f5083, %f5101, %f5112; mul.ftz.f32 %f5114, %f5082, %f5095; fma.rn.ftz.f32 %f5115, %f5107, %f5088, %f5114; fma.rn.ftz.f32 %f5116, %f5083, %f5102, %f5115; mul.ftz.f32 %f5117, %f5113, %f5113; fma.rn.ftz.f32 %f5118, %f5110, %f5110, %f5117; fma.rn.ftz.f32 %f5119, %f5116, %f5116, %f5118; sqrt.approx.ftz.f32 %f5120, %f5119; div.approx.ftz.f32 %f15687, %f5110, %f5120; div.approx.ftz.f32 %f15688, %f5113, %f5120; div.approx.ftz.f32 %f15689, %f5116, %f5120; $L__BB108_86: st.local.f32 [%rd7+-76], %f15687; st.local.f32 [%rd7+-72], %f15688; st.local.f32 [%rd7+-68], %f15689; @%p71 bra $L__BB108_88; bra.uni $L__BB108_87; $L__BB108_88: ld.const.u64 %rd817, [kernel_params+2160]; cvta.to.global.u64 %rd818, %rd817; ld.local.u32 %r1027, [%rd7+-12]; mul.wide.s32 %rd819, %r1027, 208; add.s64 %rd820, %rd818, %rd819; ld.global.v4.f32 {%f5199, %f5200, %f5201, %f5202}, [%rd820+48]; ld.global.v4.f32 {%f5206, %f5207, %f5208, %f5209}, [%rd820+64]; ld.global.v4.f32 {%f5213, %f5214, %f5215, %f5216}, [%rd820+80]; ld.local.v2.f32 {%f5220, %f5221}, [%rd7+-64]; mul.ftz.f32 %f5224, %f5206, %f5221; fma.rn.ftz.f32 %f5225, %f5199, %f5220, %f5224; ld.local.f32 %f5226, [%rd7+-56]; fma.rn.ftz.f32 %f5227, %f5213, %f5226, %f5225; mul.ftz.f32 %f5228, %f5207, %f5221; fma.rn.ftz.f32 %f5229, %f5200, %f5220, %f5228; fma.rn.ftz.f32 %f5230, %f5214, %f5226, %f5229; mul.ftz.f32 %f5231, %f5208, %f5221; fma.rn.ftz.f32 %f5232, %f5201, %f5220, %f5231; fma.rn.ftz.f32 %f5233, %f5215, %f5226, %f5232; mul.ftz.f32 %f5234, %f5230, %f5230; fma.rn.ftz.f32 %f5235, %f5227, %f5227, %f5234; fma.rn.ftz.f32 %f5236, %f5233, %f5233, %f5235; sqrt.approx.ftz.f32 %f5237, %f5236; div.approx.ftz.f32 %f15696, %f5227, %f5237; div.approx.ftz.f32 %f15697, %f5230, %f5237; div.approx.ftz.f32 %f15692, %f5233, %f5237; bra.uni $L__BB108_89; $L__BB108_58: cvt.u32.u64 %r1005, %rd3; setp.ge.s32 %p56, %r25, %r24; selp.b32 %r1006, -1, 0, %p56; add.s32 %r1007, %r36, %r1006; ld.const.u64 %rd765, [kernel_params+2344]; cvta.to.global.u64 %rd766, %rd765; mad.lo.s32 %r1008, %r1007, %r23, %r1005; add.s32 %r1009, %r1008, %r3071; mul.wide.s32 %rd767, %r1009, 16; add.s64 %rd2926, %rd766, %rd767; $L__BB108_60: cvt.rn.f32.s32 %f4933, %r25; sub.ftz.f32 %f4934, %f394, %f4933; ld.global.v4.f32 {%f4935, %f4936, %f4937, %f4938}, [%rd2926]; mov.f32 %f4943, 0f3F800000; sub.ftz.f32 %f4944, %f4943, %f4934; mul.ftz.f32 %f4945, %f4934, %f4935; mul.ftz.f32 %f4946, %f4934, %f4936; mul.ftz.f32 %f4947, %f4934, %f4937; fma.rn.ftz.f32 %f15665, %f4944, %f4931, %f4947; fma.rn.ftz.f32 %f15664, %f4944, %f4930, %f4946; fma.rn.ftz.f32 %f15663, %f4944, %f4929, %f4945; $L__BB108_62: and.b32 %r1010, %r3072, 4; setp.ne.s32 %p57, %r1010, 0; @%p57 bra $L__BB108_66; @%p15 bra $L__BB108_65; bra.uni $L__BB108_64; $L__BB108_65: ld.const.u64 %rd775, [kernel_params+2160]; cvta.to.global.u64 %rd776, %rd775; mul.lo.s64 %rd777, %rd9, 208; add.s64 %rd778, %rd776, %rd777; ld.global.v4.f32 {%f4962, %f4963, %f4964, %f4965}, [%rd778]; ld.global.v4.f32 {%f4970, %f4971, %f4972, %f4973}, [%rd778+16]; ld.global.v4.f32 {%f4978, %f4979, %f4980, %f4981}, [%rd778+32]; mul.ftz.f32 %f4986, %f15664, %f4963; fma.rn.ftz.f32 %f4987, %f15663, %f4962, %f4986; fma.rn.ftz.f32 %f4988, %f15665, %f4964, %f4987; add.ftz.f32 %f417, %f4965, %f4988; mul.ftz.f32 %f4989, %f15664, %f4971; fma.rn.ftz.f32 %f4990, %f15663, %f4970, %f4989; fma.rn.ftz.f32 %f4991, %f15665, %f4972, %f4990; add.ftz.f32 %f418, %f4973, %f4991; mul.ftz.f32 %f4992, %f15664, %f4979; fma.rn.ftz.f32 %f4993, %f15663, %f4978, %f4992; fma.rn.ftz.f32 %f4994, %f15665, %f4980, %f4993; add.ftz.f32 %f15665, %f4981, %f4994; mov.f32 %f15663, %f417; mov.f32 %f15664, %f418; bra.uni $L__BB108_66; $L__BB108_87: ld.local.v2.f32 {%f5160, %f5161}, [%rd7+-64]; ld.local.v4.f32 {%f5164, %f5165, %f5166, %f5167}, [%rd7+104]; ld.local.v4.f32 {%f5171, %f5172, %f5173, %f5174}, [%rd7+120]; ld.local.v4.f32 {%f5178, %f5179, %f5180, %f5181}, [%rd7+136]; mul.ftz.f32 %f5185, %f5161, %f5171; fma.rn.ftz.f32 %f5186, %f5160, %f5164, %f5185; ld.local.f32 %f5187, [%rd7+-56]; fma.rn.ftz.f32 %f5188, %f5187, %f5178, %f5186; mul.ftz.f32 %f5189, %f5161, %f5172; fma.rn.ftz.f32 %f5190, %f5160, %f5165, %f5189; fma.rn.ftz.f32 %f5191, %f5187, %f5179, %f5190; mul.ftz.f32 %f5192, %f5161, %f5173; fma.rn.ftz.f32 %f5193, %f5160, %f5166, %f5192; fma.rn.ftz.f32 %f5194, %f5187, %f5180, %f5193; mul.ftz.f32 %f5195, %f5191, %f5191; fma.rn.ftz.f32 %f5196, %f5188, %f5188, %f5195; fma.rn.ftz.f32 %f5197, %f5194, %f5194, %f5196; sqrt.approx.ftz.f32 %f5198, %f5197; div.approx.ftz.f32 %f15696, %f5188, %f5198; div.approx.ftz.f32 %f15697, %f5191, %f5198; div.approx.ftz.f32 %f15692, %f5194, %f5198; $L__BB108_89: st.local.v2.f32 [%rd7+-64], {%f15696, %f15697}; st.local.f32 [%rd7+-56], %f15692; @%p71 bra $L__BB108_91; bra.uni $L__BB108_90; $L__BB108_91: ld.const.u64 %rd821, [kernel_params+2160]; cvta.to.global.u64 %rd822, %rd821; ld.local.u32 %r1028, [%rd7+-12]; mul.wide.s32 %rd823, %r1028, 208; add.s64 %rd824, %rd822, %rd823; ld.global.v4.f32 {%f5270, %f5271, %f5272, %f5273}, [%rd824]; ld.global.v4.f32 {%f5277, %f5278, %f5279, %f5280}, [%rd824+16]; ld.global.v4.f32 {%f5284, %f5285, %f5286, %f5287}, [%rd824+32]; ld.local.v2.f32 {%f5291, %f5292}, [%rd7+32]; ld.local.f32 %f5295, [%rd7+28]; mul.ftz.f32 %f5296, %f5271, %f5291; fma.rn.ftz.f32 %f5297, %f5295, %f5270, %f5296; fma.rn.ftz.f32 %f15693, %f5272, %f5292, %f5297; mul.ftz.f32 %f5298, %f5278, %f5291; fma.rn.ftz.f32 %f5299, %f5295, %f5277, %f5298; fma.rn.ftz.f32 %f15694, %f5279, %f5292, %f5299; mul.ftz.f32 %f5300, %f5285, %f5291; fma.rn.ftz.f32 %f5301, %f5295, %f5284, %f5300; fma.rn.ftz.f32 %f15695, %f5286, %f5292, %f5301; bra.uni $L__BB108_92; $L__BB108_64: mul.ftz.f32 %f4953, %f15664, %f142; fma.rn.ftz.f32 %f4954, %f15663, %f141, %f4953; fma.rn.ftz.f32 %f4955, %f15665, %f143, %f4954; add.ftz.f32 %f414, %f144, %f4955; mul.ftz.f32 %f4956, %f15664, %f146; fma.rn.ftz.f32 %f4957, %f15663, %f145, %f4956; fma.rn.ftz.f32 %f4958, %f15665, %f147, %f4957; add.ftz.f32 %f415, %f148, %f4958; mul.ftz.f32 %f4959, %f15664, %f150; fma.rn.ftz.f32 %f4960, %f15663, %f149, %f4959; fma.rn.ftz.f32 %f4961, %f15665, %f151, %f4960; add.ftz.f32 %f15665, %f152, %f4961; mov.f32 %f15663, %f414; mov.f32 %f15664, %f415; $L__BB108_66: ld.local.v4.f32 {%f4995, %f4996, %f4997, %f4998}, [%rd1]; sub.ftz.f32 %f5002, %f4995, %f15663; sub.ftz.f32 %f5003, %f4996, %f15664; sub.ftz.f32 %f5004, %f4997, %f15665; mul.ftz.f32 %f5005, %f5003, %f5003; fma.rn.ftz.f32 %f5006, %f5002, %f5002, %f5005; fma.rn.ftz.f32 %f5007, %f5004, %f5004, %f5006; sqrt.approx.ftz.f32 %f5008, %f5007; div.approx.ftz.f32 %f5009, %f5004, %f5008; st.local.f32 [%rd7+-56], %f5009; div.approx.ftz.f32 %f15696, %f5002, %f5008; st.local.f32 [%rd7+-76], %f15696; div.approx.ftz.f32 %f15697, %f5003, %f5008; st.local.v4.f32 [%rd7+-72], {%f15697, %f5009, %f15696, %f15697}; mov.u32 %r1011, 0; st.local.u32 [%rd7+28], %r1011; mov.u64 %rd781, 0; st.local.u64 [%rd7+32], %rd781; mov.f32 %f5010, 0f00000000; st.local.v2.f32 [%rd7+40], {%f5010, %f5010}; st.local.u32 [%rd7+48], %r1011; $L__BB108_95: ld.const.u64 %rd829, [kernel_params+2408]; cvta.to.global.u64 %rd28, %rd829; ld.local.u32 %rd830, [%rd7+-40]; cvt.u32.u64 %r1030, %rd830; and.b32 %r3074, %r1030, 4194303; shl.b64 %rd831, %rd830, 5; and.b64 %rd832, %rd831, 134217696; add.s64 %rd833, %rd28, %rd832; ld.global.u32 %r3134, [%rd833+16]; st.local.u32 [%rd7+-36], %r3134; ld.local.v4.f32 {%f5376, %f5377, %f5378, %f5379}, [%rd7+-56]; mul.ftz.f32 %f5380, %f15697, %f5378; fma.rn.ftz.f32 %f5381, %f15696, %f5377, %f5380; fma.rn.ftz.f32 %f5382, %f5376, %f5379, %f5381; setp.lt.ftz.f32 %p75, %f5382, 0f00000000; @%p75 bra $L__BB108_97; bra.uni $L__BB108_96; $L__BB108_97: or.b32 %r3134, %r3134, 1; st.local.u32 [%rd7+-36], %r3134; neg.ftz.f32 %f15700, %f5376; neg.ftz.f32 %f15697, %f15697; neg.ftz.f32 %f15696, %f15696; st.local.f32 [%rd7+-56], %f15700; ld.local.f32 %f5387, [%rd7+-76]; neg.ftz.f32 %f5388, %f5387; ld.local.v2.f32 {%f5389, %f5390}, [%rd7+-72]; st.local.f32 [%rd7+-76], %f5388; neg.ftz.f32 %f5393, %f5390; neg.ftz.f32 %f5394, %f5389; st.local.v4.f32 [%rd7+-72], {%f5394, %f5393, %f15696, %f15697}; ld.local.f32 %f5395, [%rd7+28]; neg.ftz.f32 %f15712, %f5395; ld.local.v2.f32 {%f5396, %f5397}, [%rd7+32]; neg.ftz.f32 %f15713, %f5396; neg.ftz.f32 %f5400, %f5397; st.local.f32 [%rd7+28], %f15712; st.local.f32 [%rd7+32], %f15713; st.local.f32 [%rd7+36], %f5400; ld.local.v4.f32 {%f5401, %f5402, %f5403, %f5404}, [%rd7+40]; neg.ftz.f32 %f15703, %f5403; neg.ftz.f32 %f15711, %f5402; neg.ftz.f32 %f15710, %f5401; st.local.v2.f32 [%rd7+40], {%f15710, %f15711}; st.local.f32 [%rd7+48], %f15703; bra.uni $L__BB108_98; $L__BB108_96: ld.local.f32 %f15712, [%rd7+28]; ld.local.f32 %f15713, [%rd7+32]; ld.local.v4.f32 {%f15710, %f15711, %f15703, %f5386}, [%rd7+40]; mov.b32 %r1031, %f5376; mov.b64 %rd834, {%r1031, %r1032}; cvt.u32.u64 %r1033, %rd834; mov.b32 %f15700, %r1033; $L__BB108_98: ld.local.f32 %f564, [%rd7]; fma.rn.ftz.f32 %f5408, %f12, %f564, %f11; st.local.f32 [%rd7+4], %f5408; st.local.f32 [%rd7+8], %f12; setp.neu.ftz.f32 %p76, %f15696, %f15697; setp.neu.ftz.f32 %p77, %f15696, %f15700; or.pred %p78, %p76, %p77; sub.ftz.f32 %f5409, %f15700, %f15697; neg.ftz.f32 %f5410, %f15697; neg.ftz.f32 %f5411, %f15700; selp.f32 %f5412, %f5411, %f15700, %p78; add.ftz.f32 %f5413, %f15696, %f5412; selp.f32 %f5414, %f15697, %f5410, %p78; sub.ftz.f32 %f5415, %f5414, %f15696; mul.ftz.f32 %f5416, %f5413, %f5413; fma.rn.ftz.f32 %f5417, %f5409, %f5409, %f5416; fma.rn.ftz.f32 %f5418, %f5415, %f5415, %f5417; sqrt.approx.ftz.f32 %f5419, %f5418; div.approx.ftz.f32 %f5420, %f5409, %f5419; div.approx.ftz.f32 %f5421, %f5413, %f5419; div.approx.ftz.f32 %f5422, %f5415, %f5419; mul.ftz.f32 %f5423, %f15697, %f5422; mul.ftz.f32 %f5424, %f15700, %f5421; sub.ftz.f32 %f5425, %f5423, %f5424; mul.ftz.f32 %f5426, %f15700, %f5420; mul.ftz.f32 %f5427, %f15696, %f5422; sub.ftz.f32 %f5428, %f5426, %f5427; mul.ftz.f32 %f5429, %f15696, %f5421; mul.ftz.f32 %f5430, %f15697, %f5420; sub.ftz.f32 %f5431, %f5429, %f5430; mul.ftz.f32 %f15706, %f5408, %f5420; mul.ftz.f32 %f15707, %f5408, %f5421; mul.ftz.f32 %f567, %f5408, %f5422; mul.ftz.f32 %f15708, %f5408, %f5425; mul.ftz.f32 %f15709, %f5408, %f5428; mul.ftz.f32 %f570, %f5408, %f5431; abs.ftz.f32 %f573, %f15696; abs.ftz.f32 %f574, %f15697; abs.ftz.f32 %f575, %f15700; setp.geu.ftz.f32 %p79, %f575, %f573; setp.geu.ftz.f32 %p80, %f575, %f574; and.pred %p81, %p79, %p80; @%p81 bra $L__BB108_100; ld.local.f32 %f576, [%rd7+36]; setp.lt.ftz.f32 %p82, %f574, %f575; setp.lt.ftz.f32 %p83, %f574, %f573; or.pred %p84, %p83, %p82; selp.f32 %f15706, %f15707, %f15706, %p84; selp.f32 %f15708, %f15709, %f15708, %p84; selp.f32 %f15710, %f15711, %f15710, %p84; selp.f32 %f15712, %f15713, %f15712, %p84; mov.f32 %f15707, %f567; mov.f32 %f15709, %f570; mov.f32 %f15711, %f15703; mov.f32 %f15713, %f576; $L__BB108_100: mul.ftz.f32 %f5432, %f15710, %f15713; mul.ftz.f32 %f5433, %f15711, %f15712; sub.ftz.f32 %f15714, %f5433, %f5432; setp.eq.ftz.f32 %p85, %f15714, 0f00000000; @%p85 bra $L__BB108_102; rcp.approx.ftz.f32 %f15714, %f15714; $L__BB108_102: mul.ftz.f32 %f5434, %f15707, %f15710; mul.ftz.f32 %f5435, %f15706, %f15711; sub.ftz.f32 %f5436, %f5435, %f5434; mul.ftz.f32 %f5437, %f5436, %f15714; st.local.f32 [%rd7+12], %f5437; mul.ftz.f32 %f5438, %f15706, %f15713; mul.ftz.f32 %f5439, %f15707, %f15712; sub.ftz.f32 %f5440, %f5439, %f5438; mul.ftz.f32 %f5441, %f15709, %f15710; mul.ftz.f32 %f5442, %f15708, %f15711; sub.ftz.f32 %f5443, %f5442, %f5441; mul.ftz.f32 %f5444, %f5440, %f15714; mul.ftz.f32 %f5445, %f5443, %f15714; st.local.v2.f32 [%rd7+16], {%f5445, %f5444}; mul.ftz.f32 %f5446, %f15708, %f15713; mul.ftz.f32 %f5447, %f15709, %f15712; sub.ftz.f32 %f5448, %f5447, %f5446; mul.ftz.f32 %f5449, %f5448, %f15714; st.local.f32 [%rd7+24], %f5449; ld.const.u64 %rd835, [kernel_params+2544]; cvta.to.global.u64 %rd836, %rd835; shl.b64 %rd837, %rd2, 2; add.s64 %rd29, %rd836, %rd837; ld.global.u32 %r45, [%rd29]; and.b32 %r1034, %r3134, 524288; setp.eq.s32 %p86, %r1034, 0; @%p86 bra $L__BB108_110; bra.uni $L__BB108_103; $L__BB108_110: and.b32 %r46, %r45, 15728640; setp.ne.s32 %p92, %r46, 0; and.b32 %r1039, %r3134, 2097152; setp.eq.s32 %p93, %r1039, 0; and.pred %p94, %p93, %p92; @%p94 bra $L__BB108_116; and.b32 %r1040, %r45, 1016576; setp.eq.s32 %p95, %r1040, 0; mov.u32 %r1041, 0; ld.const.u32 %r1042, [kernel_params+4]; selp.b32 %r1043, %r1042, 0, %p95; st.local.u32 [%rd7+172], %r1041; st.local.u32 [%rd7+176], %r1043; ld.const.u64 %rd858, [kernel_params+2400]; cvta.to.global.u64 %rd30, %rd858; and.b32 %r1044, %r45, 24576; setp.ne.s32 %p96, %r1044, 16384; setp.eq.s64 %p97, %rd618, 0; or.pred %p1, %p97, %p96; ld.const.u64 %rd859, [kernel_params+2448]; cvta.to.global.u64 %rd860, %rd859; add.s64 %rd31, %rd860, %rd837; ld.const.s32 %rd32, [kernel_params+1264]; ld.const.s32 %rd33, [kernel_params+1436]; ld.const.s32 %rd34, [kernel_params+1440]; add.u64 %rd862, %SP, 704; add.u64 %rd35, %SPL, 704; bra.uni $L__BB108_112; $L__BB108_103: cvt.u16.u32 %rs60, %r45; setp.lt.s16 %p87, %rs60, 0; mov.f32 %f15715, 0f00000000; @%p87 bra $L__BB108_105; ld.const.u64 %rd838, [kernel_params+2576]; cvta.to.global.u64 %rd839, %rd838; add.s64 %rd841, %rd839, %rd837; ld.global.f32 %f15715, [%rd841]; $L__BB108_105: setp.eq.ftz.f32 %p88, %f15715, 0f00000000; mov.u16 %rs161, 0; @%p88 bra $L__BB108_2421; setp.eq.ftz.f32 %p89, %f15715, 0f3F800000; @%p89 bra $L__BB108_108; ld.const.u64 %rd842, [kernel_params+2584]; cvta.to.global.u64 %rd843, %rd842; mul.lo.s64 %rd844, %rd2, 12; add.s64 %rd845, %rd843, %rd844; ld.global.f32 %f5451, [%rd845]; rcp.approx.ftz.f32 %f5452, %f15715; mul.ftz.f32 %f5453, %f5452, %f5451; ld.global.f32 %f5454, [%rd845+4]; mul.ftz.f32 %f5455, %f5452, %f5454; ld.global.f32 %f5456, [%rd845+8]; mul.ftz.f32 %f5457, %f5452, %f5456; st.global.f32 [%rd845], %f5453; st.global.f32 [%rd845+4], %f5455; st.global.f32 [%rd845+8], %f5457; $L__BB108_108: mov.u32 %r3294, 0; ld.const.u64 %rd846, [kernel_params+2512]; cvta.to.global.u64 %rd847, %rd846; shl.b64 %rd848, %rd2, 1; add.s64 %rd849, %rd847, %rd848; ld.global.u16 %rs62, [%rd849]; add.s16 %rs63, %rs62, 1; st.global.u16 [%rd849], %rs63; setp.gt.u16 %p90, %rs62, 1023; @%p90 bra $L__BB108_2405; ld.const.u64 %rd850, [kernel_params+2536]; cvta.to.global.u64 %rd851, %rd850; add.s64 %rd853, %rd851, %rd848; ld.global.u16 %rs64, [%rd853]; add.s16 %rs65, %rs64, 16; st.global.u16 [%rd853], %rs65; mov.b32 %r1037, %f564; add.s32 %r1038, %r1037, 1; mov.b32 %f5458, %r1038; setp.eq.ftz.f32 %p91, %f564, 0f00000000; selp.f32 %f5459, 0f00000001, %f5458, %p91; ld.const.u64 %rd854, [kernel_params+2648]; cvta.to.global.u64 %rd855, %rd854; add.s64 %rd857, %rd855, %rd837; st.global.f32 [%rd857], %f5459; mov.u32 %r3294, 33; bra.uni $L__BB108_2405; $L__BB108_309: { // callseq 693, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 693 add.s32 %r3074, %r47, 1; $L__BB108_112: mov.u32 %r47, %r3074; mul.wide.s32 %rd863, %r47, 16; add.s64 %rd36, %rd30, %rd863; ld.global.v4.u32 {%r1046, %r1047, %r1048, %r1049}, [%rd36]; mul.wide.u32 %rd864, %r1047, 4; add.s64 %rd37, %rd35, %rd864; mul.wide.u32 %rd865, %r1048, 4; add.s64 %rd38, %rd35, %rd865; setp.gt.s32 %p98, %r1046, 1; @%p98 bra $L__BB108_144; setp.eq.s32 %p228, %r1046, 1; mov.u32 %r3074, %r1047; @%p228 bra $L__BB108_112; bra.uni $L__BB108_114; $L__BB108_144: setp.gt.s32 %p99, %r1046, 49; @%p99 bra $L__BB108_225; bra.uni $L__BB108_145; $L__BB108_225: setp.gt.s32 %p100, %r1046, 73; @%p100 bra $L__BB108_265; bra.uni $L__BB108_226; $L__BB108_265: setp.gt.s32 %p101, %r1046, 85; @%p101 bra $L__BB108_288; bra.uni $L__BB108_266; $L__BB108_288: setp.gt.s32 %p102, %r1046, 91; @%p102 bra $L__BB108_298; bra.uni $L__BB108_289; $L__BB108_298: setp.gt.s32 %p103, %r1046, 94; @%p103 bra $L__BB108_303; bra.uni $L__BB108_299; $L__BB108_303: setp.gt.s32 %p104, %r1046, 96; @%p104 bra $L__BB108_307; setp.eq.s32 %p107, %r1046, 95; @%p107 bra $L__BB108_310; setp.eq.s32 %p108, %r1046, 96; @%p108 bra $L__BB108_306; bra.uni $L__BB108_114; $L__BB108_306: { // callseq 694, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 694 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_145: setp.gt.s32 %p165, %r1046, 25; @%p165 bra $L__BB108_186; setp.gt.s32 %p197, %r1046, 13; @%p197 bra $L__BB108_167; setp.gt.s32 %p213, %r1046, 7; @%p213 bra $L__BB108_157; setp.gt.s32 %p221, %r1046, 4; @%p221 bra $L__BB108_153; setp.eq.s32 %p225, %r1046, 2; @%p225 bra $L__BB108_376; setp.eq.s32 %p226, %r1046, 3; @%p226 bra $L__BB108_375; setp.eq.s32 %p227, %r1046, 4; @%p227 bra $L__BB108_152; bra.uni $L__BB108_114; $L__BB108_152: mov.u32 %r1123, 4; { // callseq 772, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1123, %r1047, %r1048, %r1049}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 772 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_226: setp.gt.s32 %p134, %r1046, 61; @%p134 bra $L__BB108_246; setp.gt.s32 %p150, %r1046, 55; @%p150 bra $L__BB108_237; setp.gt.s32 %p158, %r1046, 52; @%p158 bra $L__BB108_233; setp.eq.s32 %p162, %r1046, 50; @%p162 bra $L__BB108_343; setp.eq.s32 %p163, %r1046, 51; @%p163 bra $L__BB108_342; setp.eq.s32 %p164, %r1046, 52; @%p164 bra $L__BB108_232; bra.uni $L__BB108_114; $L__BB108_232: add.s32 %r1086, %r47, 1; { // callseq 733, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 733 bra.uni $L__BB108_112; $L__BB108_186: setp.gt.s32 %p166, %r1046, 37; @%p166 bra $L__BB108_206; setp.gt.s32 %p182, %r1046, 31; @%p182 bra $L__BB108_197; setp.gt.s32 %p190, %r1046, 28; @%p190 bra $L__BB108_193; setp.eq.s32 %p194, %r1046, 26; @%p194 bra $L__BB108_359; setp.eq.s32 %p195, %r1046, 27; @%p195 bra $L__BB108_358; setp.eq.s32 %p196, %r1046, 28; @%p196 bra $L__BB108_192; bra.uni $L__BB108_114; $L__BB108_192: mov.u32 %r1099, 28; { // callseq 755, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1099; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 755 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_266: setp.gt.s32 %p119, %r1046, 79; @%p119 bra $L__BB108_279; setp.gt.s32 %p127, %r1046, 76; @%p127 bra $L__BB108_275; setp.eq.s32 %p131, %r1046, 74; @%p131 bra $L__BB108_327; setp.eq.s32 %p132, %r1046, 75; @%p132 bra $L__BB108_326; setp.eq.s32 %p133, %r1046, 76; @%p133 bra $L__BB108_271; bra.uni $L__BB108_114; $L__BB108_271: add.s32 %r3074, %r47, 1; setp.eq.s32 %p230, %r1049, 255; @%p230 bra $L__BB108_112; add.s64 %rd39, %rd37, 4; add.s64 %rd40, %rd37, 8; setp.eq.s32 %p231, %r1048, 0; mul.wide.u32 %rd911, %r1049, 4; add.s64 %rd41, %rd35, %rd911; @%p231 bra $L__BB108_325; setp.eq.s32 %p232, %r1048, 1; @%p232 bra $L__BB108_324; bra.uni $L__BB108_274; $L__BB108_324: ld.local.f32 %f5476, [%rd39]; st.local.f32 [%rd41], %f5476; bra.uni $L__BB108_112; $L__BB108_167: setp.gt.s32 %p198, %r1046, 19; @%p198 bra $L__BB108_177; setp.gt.s32 %p206, %r1046, 16; @%p206 bra $L__BB108_173; setp.eq.s32 %p210, %r1046, 14; @%p210 bra $L__BB108_367; setp.eq.s32 %p211, %r1046, 15; @%p211 bra $L__BB108_366; setp.eq.s32 %p212, %r1046, 16; @%p212 bra $L__BB108_172; bra.uni $L__BB108_114; $L__BB108_172: mov.u32 %r1109, 16; { // callseq 766, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1109; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 766 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_246: setp.gt.s32 %p135, %r1046, 67; @%p135 bra $L__BB108_256; setp.gt.s32 %p143, %r1046, 64; @%p143 bra $L__BB108_252; setp.eq.s32 %p147, %r1046, 62; @%p147 bra $L__BB108_335; setp.eq.s32 %p148, %r1046, 63; @%p148 bra $L__BB108_334; setp.eq.s32 %p149, %r1046, 64; @%p149 bra $L__BB108_251; bra.uni $L__BB108_114; $L__BB108_251: mov.u32 %r1072, 64; add.s32 %r1073, %r47, 1; { // callseq 721, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1072; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1073; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 721 bra.uni $L__BB108_112; $L__BB108_206: setp.gt.s32 %p167, %r1046, 43; @%p167 bra $L__BB108_216; setp.gt.s32 %p175, %r1046, 40; @%p175 bra $L__BB108_212; setp.eq.s32 %p179, %r1046, 38; @%p179 bra $L__BB108_351; setp.eq.s32 %p180, %r1046, 39; @%p180 bra $L__BB108_350; setp.eq.s32 %p181, %r1046, 40; @%p181 bra $L__BB108_211; bra.uni $L__BB108_114; $L__BB108_211: add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_289: setp.gt.s32 %p112, %r1046, 88; @%p112 bra $L__BB108_294; setp.eq.s32 %p116, %r1046, 86; @%p116 bra $L__BB108_316; setp.eq.s32 %p117, %r1046, 87; @%p117 bra $L__BB108_315; setp.eq.s32 %p118, %r1046, 88; @%p118 bra $L__BB108_293; bra.uni $L__BB108_114; $L__BB108_293: mov.u32 %r1055, 88; ld.local.f32 %f5469, [%rd7+-76]; ld.local.v2.f32 {%f5470, %f5471}, [%rd7+-72]; { // callseq 699, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f5469; .param .b32 param1; st.param.f32 [param1+0], %f5470; .param .b32 param2; st.param.f32 [param2+0], %f5471; .param .b64 param3; st.param.b64 [param3+0], %rd862; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r1055; st.param.b32 [param4+4], %r1047; st.param.b32 [param4+8], %r1048; st.param.b32 [param4+12], %r1049; call.uni _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 699 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_157: setp.gt.s32 %p214, %r1046, 10; @%p214 bra $L__BB108_163; setp.eq.s32 %p218, %r1046, 8; @%p218 bra $L__BB108_372; setp.eq.s32 %p219, %r1046, 9; @%p219 bra $L__BB108_370; setp.eq.s32 %p220, %r1046, 10; @%p220 bra $L__BB108_161; bra.uni $L__BB108_114; $L__BB108_161: ld.local.f32 %f5502, [%rd38]; setp.ltu.ftz.f32 %p233, %f5502, 0f3F800000; add.s32 %r3074, %r47, 1; @%p233 bra $L__BB108_112; add.s32 %r1119, %r1047, %r47; add.s32 %r3074, %r1119, 1; bra.uni $L__BB108_112; $L__BB108_237: setp.gt.s32 %p151, %r1046, 58; @%p151 bra $L__BB108_242; setp.eq.s32 %p155, %r1046, 56; @%p155 bra $L__BB108_339; setp.eq.s32 %p156, %r1046, 57; @%p156 bra $L__BB108_338; setp.eq.s32 %p157, %r1046, 58; @%p157 bra $L__BB108_241; bra.uni $L__BB108_114; $L__BB108_241: mov.u32 %r1081, 58; { // callseq 727, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1081; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 727 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_197: setp.gt.s32 %p183, %r1046, 34; @%p183 bra $L__BB108_202; setp.eq.s32 %p187, %r1046, 32; @%p187 bra $L__BB108_355; setp.eq.s32 %p188, %r1046, 33; @%p188 bra $L__BB108_354; setp.eq.s32 %p189, %r1046, 34; @%p189 bra $L__BB108_201; bra.uni $L__BB108_114; $L__BB108_201: { // callseq 750, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 750 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_279: setp.gt.s32 %p120, %r1046, 82; @%p120 bra $L__BB108_284; setp.eq.s32 %p124, %r1046, 80; @%p124 bra $L__BB108_320; setp.eq.s32 %p125, %r1046, 81; @%p125 bra $L__BB108_319; setp.eq.s32 %p126, %r1046, 82; @%p126 bra $L__BB108_283; bra.uni $L__BB108_114; $L__BB108_283: mov.u32 %r1059, 82; { // callseq 705, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1059; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 705 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_177: setp.gt.s32 %p199, %r1046, 22; @%p199 bra $L__BB108_182; setp.eq.s32 %p203, %r1046, 20; @%p203 bra $L__BB108_363; setp.eq.s32 %p204, %r1046, 21; @%p204 bra $L__BB108_362; setp.eq.s32 %p205, %r1046, 22; @%p205 bra $L__BB108_181; bra.uni $L__BB108_114; $L__BB108_181: mov.u32 %r1106, 22; add.s32 %r1107, %r47, 1; { // callseq 761, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1106; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r1107; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 761 bra.uni $L__BB108_112; $L__BB108_256: setp.gt.s32 %p136, %r1046, 70; @%p136 bra $L__BB108_261; setp.eq.s32 %p140, %r1046, 68; @%p140 bra $L__BB108_331; setp.eq.s32 %p141, %r1046, 69; @%p141 bra $L__BB108_330; setp.eq.s32 %p142, %r1046, 70; @%p142 bra $L__BB108_260; bra.uni $L__BB108_114; $L__BB108_260: mov.u32 %r1066, 70; { // callseq 715, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1066; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 715 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_216: setp.gt.s32 %p168, %r1046, 46; @%p168 bra $L__BB108_221; setp.eq.s32 %p172, %r1046, 44; @%p172 bra $L__BB108_347; setp.eq.s32 %p173, %r1046, 45; @%p173 bra $L__BB108_346; setp.eq.s32 %p174, %r1046, 46; @%p174 bra $L__BB108_220; bra.uni $L__BB108_114; $L__BB108_220: { // callseq 739, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 739 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_299: setp.eq.s32 %p109, %r1046, 92; @%p109 bra $L__BB108_312; setp.eq.s32 %p110, %r1046, 93; @%p110 bra $L__BB108_311; setp.eq.s32 %p111, %r1046, 94; @%p111 bra $L__BB108_302; bra.uni $L__BB108_114; $L__BB108_302: mov.u32 %r1050, 94; add.s32 %r1051, %r47, 1; { // callseq 696, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1050; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1051; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 696 bra.uni $L__BB108_112; $L__BB108_153: setp.eq.s32 %p222, %r1046, 5; @%p222 bra $L__BB108_374; setp.eq.s32 %p223, %r1046, 6; @%p223 bra $L__BB108_373; setp.eq.s32 %p224, %r1046, 7; @%p224 bra $L__BB108_156; bra.uni $L__BB108_114; $L__BB108_156: mov.u32 %r1122, 7; { // callseq 771, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1122; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 771 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_233: setp.eq.s32 %p159, %r1046, 53; @%p159 bra $L__BB108_341; setp.eq.s32 %p160, %r1046, 54; @%p160 bra $L__BB108_340; setp.eq.s32 %p161, %r1046, 55; @%p161 bra $L__BB108_236; bra.uni $L__BB108_114; $L__BB108_236: ld.local.v4.f32 {%f5478, %f5479, %f5480, %f5481}, [%rd1]; { // callseq 730, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f5478; st.param.f32 [param0+4], %f5479; st.param.f32 [param0+8], %f5480; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 730 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_193: setp.eq.s32 %p191, %r1046, 29; @%p191 bra $L__BB108_357; setp.eq.s32 %p192, %r1046, 30; @%p192 bra $L__BB108_356; setp.eq.s32 %p193, %r1046, 31; @%p193 bra $L__BB108_196; bra.uni $L__BB108_114; $L__BB108_196: mov.u32 %r1097, 31; add.s32 %r1098, %r47, 1; { // callseq 752, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1097; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1098; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 752 bra.uni $L__BB108_112; $L__BB108_275: setp.eq.s32 %p128, %r1046, 77; @%p128 bra $L__BB108_322; setp.eq.s32 %p129, %r1046, 78; @%p129 bra $L__BB108_321; setp.eq.s32 %p130, %r1046, 79; @%p130 bra $L__BB108_278; bra.uni $L__BB108_114; $L__BB108_278: add.s32 %r1061, %r47, 1; { // callseq 708, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1061; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 708 bra.uni $L__BB108_112; $L__BB108_173: setp.eq.s32 %p207, %r1046, 17; @%p207 bra $L__BB108_365; setp.eq.s32 %p208, %r1046, 18; @%p208 bra $L__BB108_364; setp.eq.s32 %p209, %r1046, 19; @%p209 bra $L__BB108_176; bra.uni $L__BB108_114; $L__BB108_176: { // callseq 763, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 763 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_252: setp.eq.s32 %p144, %r1046, 65; @%p144 bra $L__BB108_333; setp.eq.s32 %p145, %r1046, 66; @%p145 bra $L__BB108_332; setp.eq.s32 %p146, %r1046, 67; @%p146 bra $L__BB108_255; bra.uni $L__BB108_114; $L__BB108_255: mov.u32 %r1070, 67; { // callseq 718, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1070, %r1047, %r1048, %r1049}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 718 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_212: setp.eq.s32 %p176, %r1046, 41; @%p176 bra $L__BB108_349; setp.eq.s32 %p177, %r1046, 42; @%p177 bra $L__BB108_348; setp.eq.s32 %p178, %r1046, 43; @%p178 bra $L__BB108_215; bra.uni $L__BB108_114; $L__BB108_215: add.s32 %r1090, %r47, 1; { // callseq 742, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1090; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 742 bra.uni $L__BB108_112; $L__BB108_294: setp.eq.s32 %p113, %r1046, 89; @%p113 bra $L__BB108_314; setp.eq.s32 %p114, %r1046, 90; @%p114 bra $L__BB108_313; setp.eq.s32 %p115, %r1046, 91; @%p115 bra $L__BB108_297; bra.uni $L__BB108_114; $L__BB108_297: add.s32 %r3074, %r47, 1; @%p1 bra $L__BB108_114; bra.uni $L__BB108_112; $L__BB108_163: setp.eq.s32 %p215, %r1046, 11; @%p215 bra $L__BB108_369; setp.eq.s32 %p216, %r1046, 12; @%p216 bra $L__BB108_368; setp.eq.s32 %p217, %r1046, 13; @%p217 bra $L__BB108_166; bra.uni $L__BB108_114; $L__BB108_166: mov.u32 %r1117, 13; add.s32 %r1118, %r47, 1; { // callseq 767, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1117; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1118; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 767 bra.uni $L__BB108_112; $L__BB108_242: setp.eq.s32 %p152, %r1046, 59; @%p152 bra $L__BB108_337; setp.eq.s32 %p153, %r1046, 60; @%p153 bra $L__BB108_336; setp.eq.s32 %p154, %r1046, 61; @%p154 bra $L__BB108_245; bra.uni $L__BB108_114; $L__BB108_245: mov.u32 %r1077, 61; add.s32 %r1078, %r47, 1; { // callseq 724, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1077; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1078; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 724 bra.uni $L__BB108_112; $L__BB108_202: setp.eq.s32 %p184, %r1046, 35; @%p184 bra $L__BB108_353; setp.eq.s32 %p185, %r1046, 36; @%p185 bra $L__BB108_352; setp.eq.s32 %p186, %r1046, 37; @%p186 bra $L__BB108_205; bra.uni $L__BB108_114; $L__BB108_205: mov.u32 %r1093, 37; { // callseq 747, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1093, %r1047, %r1048, %r1049}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 747 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_284: setp.eq.s32 %p121, %r1046, 83; @%p121 bra $L__BB108_318; setp.eq.s32 %p122, %r1046, 84; @%p122 bra $L__BB108_317; setp.eq.s32 %p123, %r1046, 85; @%p123 bra $L__BB108_287; bra.uni $L__BB108_114; $L__BB108_287: add.s32 %r1058, %r47, 1; { // callseq 702, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1058; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 702 bra.uni $L__BB108_112; $L__BB108_182: setp.eq.s32 %p200, %r1046, 23; @%p200 bra $L__BB108_361; setp.eq.s32 %p201, %r1046, 24; @%p201 bra $L__BB108_360; setp.eq.s32 %p202, %r1046, 25; @%p202 bra $L__BB108_185; bra.uni $L__BB108_114; $L__BB108_185: add.s32 %r1102, %r47, 1; { // callseq 758, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1102; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 758 bra.uni $L__BB108_112; $L__BB108_261: setp.eq.s32 %p137, %r1046, 71; @%p137 bra $L__BB108_329; setp.eq.s32 %p138, %r1046, 72; @%p138 bra $L__BB108_328; setp.eq.s32 %p139, %r1046, 73; @%p139 bra $L__BB108_264; bra.uni $L__BB108_114; $L__BB108_264: add.s32 %r1064, %r47, 1; { // callseq 712, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1064; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 712 bra.uni $L__BB108_112; $L__BB108_221: setp.eq.s32 %p169, %r1046, 47; @%p169 bra $L__BB108_345; setp.eq.s32 %p170, %r1046, 48; @%p170 bra $L__BB108_344; setp.eq.s32 %p171, %r1046, 49; @%p171 bra $L__BB108_224; bra.uni $L__BB108_114; $L__BB108_224: { // callseq 736, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 736 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_307: setp.eq.s32 %p105, %r1046, 98; @%p105 bra $L__BB108_377; setp.ne.s32 %p106, %r1046, 97; @%p106 bra $L__BB108_114; bra.uni $L__BB108_309; $L__BB108_310: { // callseq 695, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 695 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_376: mov.u32 %r1125, 2; add.s32 %r1126, %r47, 1; { // callseq 774, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1125; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r45; .param .b32 param4; st.param.b32 [param4+0], %r1126; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 774 bra.uni $L__BB108_112; $L__BB108_375: mov.u32 %r1124, 3; { // callseq 773, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1124, %r1047, %r1048, %r1049}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 773 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_343: { // callseq 735, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 735 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_342: { // callseq 734, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 734 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_359: mov.u32 %r1101, 26; { // callseq 757, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1101; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 757 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_358: mov.u32 %r1100, 27; { // callseq 756, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1100; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 756 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_327: { // callseq 711, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 711 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_326: { // callseq 710, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 710 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_367: st.local.u32 [%rd38], %r1047; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_366: ld.global.v4.u32 {%r1110, %r1111, %r1112, %r1113}, [%rd36+16]; add.s32 %r3074, %r47, 2; st.local.u32 [%rd37], %r1111; st.local.u32 [%rd37+4], %r1112; st.local.u32 [%rd37+8], %r1113; bra.uni $L__BB108_112; $L__BB108_335: mov.u32 %r1075, 62; add.s32 %r1076, %r47, 1; { // callseq 723, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1075; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1076; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 723 bra.uni $L__BB108_112; $L__BB108_334: mov.u32 %r1074, 63; { // callseq 722, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1074; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 722 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_351: { // callseq 746, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 746 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_350: mov.u32 %r1092, 39; { // callseq 745, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1092, %r1047, %r1048, %r1049}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 745 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_316: add.s32 %r1057, %r47, 1; { // callseq 701, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1057; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 701 bra.uni $L__BB108_112; $L__BB108_315: add.s32 %r1056, %r47, 1; { // callseq 700, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1056; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 700 bra.uni $L__BB108_112; $L__BB108_372: mov.u32 %r1121, 8; { // callseq 770, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1121; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 770 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_370: ld.local.f32 %f5503, [%rd38]; setp.gtu.ftz.f32 %p234, %f5503, 0f00000000; add.s32 %r3074, %r47, 1; @%p234 bra $L__BB108_112; add.s32 %r1120, %r1047, %r47; add.s32 %r3074, %r1120, 1; bra.uni $L__BB108_112; $L__BB108_339: mov.u32 %r1084, 56; { // callseq 729, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1084; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 729 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_338: mov.u32 %r1082, 57; add.s32 %r1083, %r47, 1; { // callseq 728, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1082; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1083; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 728 bra.uni $L__BB108_112; $L__BB108_355: mov.u32 %r1095, 32; add.s32 %r1096, %r47, 1; { // callseq 751, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b32 param1; st.param.b32 [param1+0], %r45; .param .b64 param2; st.param.b64 [param2+0], %rd862; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1095; st.param.b32 [param3+4], %r1047; st.param.b32 [param3+8], %r1048; st.param.b32 [param3+12], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1096; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 751 bra.uni $L__BB108_112; $L__BB108_354: ld.local.f32 %f5486, [%rd37]; ld.local.f32 %f5487, [%rd37+4]; ld.local.f32 %f5488, [%rd37+8]; st.local.f32 [%rd7+-76], %f5486; st.local.f32 [%rd7+-72], %f5487; st.local.f32 [%rd7+-68], %f5488; st.local.f32 [%rd38], %f5486; st.local.f32 [%rd38+4], %f5487; st.local.f32 [%rd38+8], %f5488; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_320: { // callseq 707, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 707 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_319: mov.u32 %r1060, 81; { // callseq 706, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1060; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 706 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_363: ld.local.f32 %f5489, [%rd37]; ld.local.v4.f32 {%f5490, %f5491, %f5492, %f5493}, [%rd1]; ld.local.f32 %f5497, [%rd37+4]; ld.local.f32 %f5498, [%rd37+8]; add.ftz.f32 %f5499, %f5498, %f5492; add.ftz.f32 %f5500, %f5497, %f5491; add.ftz.f32 %f5501, %f5489, %f5490; st.local.v2.f32 [%rd1], {%f5501, %f5500}; st.local.f32 [%rd7+-80], %f5499; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_362: mov.u32 %r1108, 21; { // callseq 762, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1108; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 762 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_331: mov.u32 %r1069, 68; { // callseq 717, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1069; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 717 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_330: mov.u32 %r1067, 69; add.s32 %r1068, %r47, 1; { // callseq 716, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1067; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1068; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 716 bra.uni $L__BB108_112; $L__BB108_347: mov.u32 %r1088, 44; add.s32 %r1089, %r47, 1; { // callseq 741, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1088; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1089; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 741 bra.uni $L__BB108_112; $L__BB108_346: { // callseq 740, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 740 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_312: ld.local.f32 %f5462, [%rd37]; ld.local.f32 %f5463, [%rd37+4]; ld.local.f32 %f5464, [%rd37+8]; ld.global.u32 %rd878, [%rd31]; mul.lo.s64 %rd879, %rd32, %rd878; cvt.s64.s32 %rd880, %r1048; add.s64 %rd881, %rd33, %rd880; add.s64 %rd882, %rd881, %rd879; cvta.to.global.u64 %rd883, %rd618; shl.b64 %rd884, %rd882, 2; add.s64 %rd885, %rd883, %rd884; add.s64 %rd886, %rd885, 4; add.s64 %rd887, %rd885, 8; add.s64 %rd888, %rd885, 12; atom.global.add.f32 %f5465, [%rd885], %f5462; atom.global.add.f32 %f5466, [%rd886], %f5463; atom.global.add.f32 %f5467, [%rd887], %f5464; atom.global.add.f32 %f5468, [%rd888], 0f3F800000; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_311: ld.local.f32 %f5460, [%rd37]; ld.global.u32 %rd870, [%rd31]; mul.lo.s64 %rd871, %rd32, %rd870; cvt.s64.s32 %rd872, %r1048; add.s64 %rd873, %rd34, %rd872; add.s64 %rd874, %rd873, %rd871; cvta.to.global.u64 %rd875, %rd618; shl.b64 %rd876, %rd874, 2; add.s64 %rd877, %rd875, %rd876; atom.global.add.f32 %f5461, [%rd877], %f5460; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_374: st.local.u32 [%rd7+180], %r1047; st.local.u32 [%rd7+184], %r1048; st.local.u32 [%rd7+188], %r1049; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_373: ld.local.f32 %f5504, [%rd37]; ld.local.f32 %f5505, [%rd37+4]; ld.local.f32 %f5506, [%rd37+8]; st.local.f32 [%rd7+180], %f5504; st.local.f32 [%rd7+184], %f5505; st.local.f32 [%rd7+188], %f5506; add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_341: { // callseq 732, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 732 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_340: add.s32 %r1085, %r47, 1; { // callseq 731, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1085; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 731 bra.uni $L__BB108_112; $L__BB108_357: { // callseq 754, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 754 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_356: { // callseq 753, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 753 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_322: setp.eq.s32 %p229, %r1049, 255; add.s32 %r3074, %r47, 1; @%p229 bra $L__BB108_112; ld.local.f32 %f5474, [%rd37]; add.s32 %r1063, %r1049, %r1048; mul.wide.u32 %rd907, %r1063, 4; add.s64 %rd908, %rd35, %rd907; st.local.f32 [%rd908], %f5474; bra.uni $L__BB108_112; $L__BB108_321: add.s32 %r1062, %r47, 1; { // callseq 709, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1062; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 709 bra.uni $L__BB108_112; $L__BB108_365: { // callseq 765, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; .param .b32 param4; st.param.b32 [param4+0], %r1049; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 765 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_364: { // callseq 764, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 764 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_333: { // callseq 720, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 720 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_332: add.s32 %r1071, %r47, 1; { // callseq 719, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1071; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 719 bra.uni $L__BB108_112; $L__BB108_349: add.s32 %r1091, %r47, 1; { // callseq 744, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1091; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 744 bra.uni $L__BB108_112; $L__BB108_348: { // callseq 743, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 743 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_314: mov.u32 %r1054, 89; { // callseq 698, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1054; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 698 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_313: mov.u32 %r1052, 90; add.s32 %r1053, %r47, 1; { // callseq 697, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1052; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; .param .b32 param3; st.param.b32 [param3+0], %r1053; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 697 bra.uni $L__BB108_112; $L__BB108_369: { // callseq 769, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 769 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_368: { // callseq 768, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 768 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_337: add.s32 %r1080, %r47, 1; { // callseq 726, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1080; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 726 bra.uni $L__BB108_112; $L__BB108_336: add.s32 %r1079, %r47, 1; { // callseq 725, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; .param .b32 param4; st.param.b32 [param4+0], %r1079; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 725 bra.uni $L__BB108_112; $L__BB108_353: { // callseq 749, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 749 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_352: mov.u32 %r1094, 36; { // callseq 748, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1094; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 748 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_318: { // callseq 704, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 704 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_317: { // callseq 703, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 703 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_361: mov.u32 %r1104, 23; add.s32 %r1105, %r47, 1; { // callseq 760, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1104; st.param.b32 [param1+4], %r1047; st.param.b32 [param1+8], %r1048; st.param.b32 [param1+12], %r1049; .param .b32 param2; st.param.b32 [param2+0], %r1105; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3074, [retval0+0]; } // callseq 760 bra.uni $L__BB108_112; $L__BB108_360: mov.u32 %r1103, 24; { // callseq 759, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1103; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 759 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_329: mov.u32 %r1065, 71; { // callseq 714, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1065; st.param.b32 [param2+4], %r1047; st.param.b32 [param2+8], %r1048; st.param.b32 [param2+12], %r1049; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 714 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_328: { // callseq 713, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; .param .b32 param3; st.param.b32 [param3+0], %r1049; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 713 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_345: ld.local.u32 %r1087, [%rd7+-36]; ld.local.f32 %f5485, [%rd7]; { // callseq 738, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b32 param1; st.param.b32 [param1+0], %r1087; .param .b32 param2; st.param.f32 [param2+0], %f5485; .param .b64 param3; st.param.b64 [param3+0], %rd862; .param .b32 param4; st.param.b32 [param4+0], %r1047; .param .b32 param5; st.param.b32 [param5+0], %r1048; .param .b32 param6; st.param.b32 [param6+0], %r45; call.uni _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5, param6 ); } // callseq 738 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_344: { // callseq 737, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .b64 param1; st.param.b64 [param1+0], %rd862; .param .b32 param2; st.param.b32 [param2+0], %r1047; .param .b32 param3; st.param.b32 [param3+0], %r1048; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 737 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_377: { // callseq 775, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd862; .param .b32 param1; st.param.b32 [param1+0], %r1047; .param .b32 param2; st.param.b32 [param2+0], %r1048; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 775 add.s32 %r3074, %r47, 1; bra.uni $L__BB108_112; $L__BB108_325: ld.local.f32 %f5477, [%rd37]; st.local.f32 [%rd41], %f5477; bra.uni $L__BB108_112; $L__BB108_274: ld.local.f32 %f5475, [%rd40]; st.local.f32 [%rd41], %f5475; bra.uni $L__BB108_112; $L__BB108_114: ld.local.u32 %r3134, [%rd7+-36]; and.b32 %r1127, %r3134, 1024; setp.eq.s32 %p235, %r1127, 0; @%p235 bra $L__BB108_116; ld.const.v2.u64 {%rd1013, %rd1014}, [kernel_params+2528]; cvta.to.global.u64 %rd1017, %rd1013; add.s64 %rd1019, %rd1017, %rd837; cvta.to.global.u64 %rd1020, %rd1014; shl.b64 %rd1021, %rd2, 1; add.s64 %rd1022, %rd1020, %rd1021; ld.global.u16 %r1128, [%rd1022]; ld.const.u64 %rd1023, [kernel_params+2456]; cvta.to.global.u64 %rd1024, %rd1023; add.s64 %rd1025, %rd1024, %rd837; ld.global.u32 %r1129, [%rd1019]; add.s32 %r1130, %r1129, %r1128; ld.global.u32 %r1131, [%rd1025]; mad.lo.s32 %r1132, %r1131, -1262732973, %r1130; mad.lo.s32 %r1133, %r1132, 1103515245, 12345; st.local.u32 [%rd7+168], %r1133; $L__BB108_116: setp.eq.s32 %p236, %r46, 0; @%p236 bra $L__BB108_386; ld.local.v2.f32 {%f15747, %f15748}, [%rd7+-72]; ld.local.f32 %f15746, [%rd7+-76]; and.b32 %r1134, %r3134, 2097152; setp.eq.s32 %p237, %r1134, 0; @%p237 bra $L__BB108_136; ld.local.u32 %r156, [%rd7+172]; setp.lt.s32 %p238, %r156, 1; mov.f32 %f15719, 0f00000000; mov.f32 %f15720, %f15719; mov.f32 %f15721, %f15719; @%p238 bra $L__BB108_134; add.s32 %r1136, %r156, -1; and.b32 %r3080, %r156, 3; setp.lt.u32 %p239, %r1136, 3; mov.f32 %f15721, 0f00000000; mov.u32 %r3078, 0; mov.f32 %f15720, %f15721; mov.f32 %f15719, %f15721; @%p239 bra $L__BB108_130; sub.s32 %r3077, %r156, %r3080; mov.f32 %f15721, 0f00000000; mov.u32 %r3078, 0; mov.f32 %f15720, %f15721; mov.f32 %f15719, %f15721; $L__BB108_121: cvt.s64.s32 %rd42, %r3078; mul.wide.s32 %rd1026, %r3078, 80; add.s64 %rd1027, %rd1, %rd1026; add.s64 %rd43, %rd1027, 316; ld.local.u32 %r1138, [%rd1027+316]; add.s32 %r1139, %r1138, -32; setp.gt.u32 %p240, %r1139, 2; @%p240 bra $L__BB108_123; ld.local.v4.f32 {%f5519, %f5520, %f5521, %f5522}, [%rd43+-12]; add.ftz.f32 %f5526, %f5519, %f5520; add.ftz.f32 %f5527, %f5521, %f5526; mul.ftz.f32 %f5528, %f5527, 0f3EAAAAAB; abs.ftz.f32 %f5529, %f5528; ld.local.f32 %f5530, [%rd43+8]; ld.local.f32 %f5531, [%rd43+12]; ld.local.f32 %f5532, [%rd43+16]; fma.rn.ftz.f32 %f15719, %f5529, %f5530, %f15719; fma.rn.ftz.f32 %f15720, %f5529, %f5531, %f15720; fma.rn.ftz.f32 %f15721, %f5529, %f5532, %f15721; $L__BB108_123: ld.local.u32 %r1140, [%rd43+80]; add.s32 %r1141, %r1140, -32; setp.gt.u32 %p241, %r1141, 2; @%p241 bra $L__BB108_125; ld.local.v4.f32 {%f5533, %f5534, %f5535, %f5536}, [%rd43+68]; add.ftz.f32 %f5540, %f5533, %f5534; add.ftz.f32 %f5541, %f5535, %f5540; mul.ftz.f32 %f5542, %f5541, 0f3EAAAAAB; abs.ftz.f32 %f5543, %f5542; ld.local.f32 %f5544, [%rd43+88]; ld.local.f32 %f5545, [%rd43+92]; ld.local.f32 %f5546, [%rd43+96]; fma.rn.ftz.f32 %f15719, %f5543, %f5544, %f15719; fma.rn.ftz.f32 %f15720, %f5543, %f5545, %f15720; fma.rn.ftz.f32 %f15721, %f5543, %f5546, %f15721; $L__BB108_125: ld.local.u32 %r1142, [%rd43+160]; add.s32 %r1143, %r1142, -32; setp.gt.u32 %p242, %r1143, 2; @%p242 bra $L__BB108_127; ld.local.v4.f32 {%f5547, %f5548, %f5549, %f5550}, [%rd43+148]; add.ftz.f32 %f5554, %f5547, %f5548; add.ftz.f32 %f5555, %f5549, %f5554; mul.ftz.f32 %f5556, %f5555, 0f3EAAAAAB; abs.ftz.f32 %f5557, %f5556; ld.local.f32 %f5558, [%rd43+168]; ld.local.f32 %f5559, [%rd43+172]; ld.local.f32 %f5560, [%rd43+176]; fma.rn.ftz.f32 %f15719, %f5557, %f5558, %f15719; fma.rn.ftz.f32 %f15720, %f5557, %f5559, %f15720; fma.rn.ftz.f32 %f15721, %f5557, %f5560, %f15721; $L__BB108_127: ld.local.u32 %r1144, [%rd43+240]; add.s32 %r1145, %r1144, -32; setp.gt.u32 %p243, %r1145, 2; @%p243 bra $L__BB108_129; ld.local.v4.f32 {%f5561, %f5562, %f5563, %f5564}, [%rd43+228]; add.ftz.f32 %f5568, %f5561, %f5562; add.ftz.f32 %f5569, %f5563, %f5568; mul.ftz.f32 %f5570, %f5569, 0f3EAAAAAB; abs.ftz.f32 %f5571, %f5570; ld.local.f32 %f5572, [%rd43+248]; ld.local.f32 %f5573, [%rd43+252]; ld.local.f32 %f5574, [%rd43+256]; fma.rn.ftz.f32 %f15719, %f5571, %f5572, %f15719; fma.rn.ftz.f32 %f15720, %f5571, %f5573, %f15720; fma.rn.ftz.f32 %f15721, %f5571, %f5574, %f15721; $L__BB108_129: cvt.u32.u64 %r1146, %rd42; add.s32 %r3078, %r1146, 4; add.s32 %r3077, %r3077, -4; setp.ne.s32 %p244, %r3077, 0; @%p244 bra $L__BB108_121; $L__BB108_130: setp.eq.s32 %p245, %r3080, 0; @%p245 bra $L__BB108_134; $L__BB108_131: .pragma "nounroll"; cvt.s64.s32 %rd44, %r3078; mul.wide.s32 %rd1028, %r3078, 80; add.s64 %rd1029, %rd1, %rd1028; add.s64 %rd45, %rd1029, 316; ld.local.u32 %r1147, [%rd1029+316]; add.s32 %r1148, %r1147, -32; setp.gt.u32 %p246, %r1148, 2; @%p246 bra $L__BB108_133; ld.local.f32 %f5575, [%rd45+-8]; ld.local.f32 %f5576, [%rd45+-12]; add.ftz.f32 %f5577, %f5576, %f5575; ld.local.f32 %f5578, [%rd45+-4]; add.ftz.f32 %f5579, %f5578, %f5577; mul.ftz.f32 %f5580, %f5579, 0f3EAAAAAB; abs.ftz.f32 %f5581, %f5580; ld.local.f32 %f5582, [%rd45+8]; ld.local.f32 %f5583, [%rd45+12]; ld.local.f32 %f5584, [%rd45+16]; fma.rn.ftz.f32 %f15719, %f5581, %f5582, %f15719; fma.rn.ftz.f32 %f15720, %f5581, %f5583, %f15720; fma.rn.ftz.f32 %f15721, %f5581, %f5584, %f15721; $L__BB108_133: cvt.u32.u64 %r1149, %rd44; add.s32 %r3078, %r1149, 1; add.s32 %r3080, %r3080, -1; setp.ne.s32 %p247, %r3080, 0; @%p247 bra $L__BB108_131; $L__BB108_134: setp.eq.ftz.f32 %p248, %f15720, 0f00000000; setp.eq.ftz.f32 %p249, %f15719, 0f00000000; and.pred %p250, %p249, %p248; setp.eq.ftz.f32 %p251, %f15721, 0f00000000; and.pred %p252, %p251, %p250; @%p252 bra $L__BB108_136; mul.ftz.f32 %f5585, %f15720, %f15720; fma.rn.ftz.f32 %f5586, %f15719, %f15719, %f5585; fma.rn.ftz.f32 %f5587, %f15721, %f15721, %f5586; sqrt.approx.ftz.f32 %f5588, %f5587; div.approx.ftz.f32 %f15748, %f15721, %f5588; div.approx.ftz.f32 %f15747, %f15720, %f5588; div.approx.ftz.f32 %f15746, %f15719, %f5588; $L__BB108_136: and.b32 %r1150, %r3134, -1279; st.local.u32 [%rd7+-36], %r1150; mov.u32 %r1151, 0; st.local.u32 [%rd7+172], %r1151; ld.const.u32 %r168, [kernel_params+4]; st.local.u32 [%rd7+176], %r168; and.b32 %r1152, %r45, 4194304; setp.eq.s32 %p253, %r1152, 0; mov.f32 %f5589, 0f3F800000; abs.ftz.f32 %f649, %f5589; @%p253 bra $L__BB108_378; bra.uni $L__BB108_137; $L__BB108_378: setp.ltu.ftz.f32 %p258, %f649, 0f3727C5AC; mov.u64 %rd2933, 0; mov.u64 %rd2934, %rd2933; @%p258 bra $L__BB108_383; setp.eq.s32 %p259, %r168, 0; mov.u64 %rd2933, 0; mov.u64 %rd2931, %rd2933; mov.u64 %rd2932, %rd2933; @%p259 bra $L__BB108_381; add.s64 %rd2931, %rd1, 304; add.s64 %rd2932, %rd619, 304; mov.u64 %rd1044, 1065353216; st.local.u64 [%rd7+224], %rd1044; mov.u64 %rd1045, 4575657222473777152; st.local.u64 [%rd7+216], %rd1045; mov.u32 %r1160, 1; st.local.u32 [%rd7+172], %r1160; add.s32 %r1161, %r168, -1; st.local.u32 [%rd7+176], %r1161; $L__BB108_381: setp.eq.s64 %p260, %rd2932, 0; mov.u64 %rd2934, %rd2933; @%p260 bra $L__BB108_383; st.local.f32 [%rd2931+16], %f649; mov.u64 %rd2933, %rd2931; mov.u64 %rd2934, %rd2932; $L__BB108_383: setp.eq.s64 %p261, %rd2934, 0; @%p261 bra $L__BB108_385; st.local.f32 [%rd2933+20], %f15746; st.local.f32 [%rd2933+24], %f15747; st.local.f32 [%rd2933+28], %f15748; mov.u32 %r1162, 2; st.local.u32 [%rd2933+12], %r1162; ld.local.u32 %r1163, [%rd7+-36]; or.b32 %r1164, %r1163, 12; st.local.u32 [%rd7+-36], %r1164; bra.uni $L__BB108_385; $L__BB108_386: ld.const.u32 %r169, [kernel_params+1624]; setp.eq.s32 %p262, %r169, 0; @%p262 bra $L__BB108_520; and.b32 %r1167, %r169, 1; setp.eq.b32 %p263, %r1167, 1; mov.pred %p264, 0; xor.pred %p265, %p263, %p264; not.pred %p266, %p265; @%p266 bra $L__BB108_389; mov.u32 %r1168, 0; mov.f32 %f5590, 0f00000000; st.local.v2.f32 [%rd7+192], {%f5590, %f5590}; st.local.u32 [%rd7+200], %r1168; $L__BB108_389: and.b32 %r1169, %r169, 32; setp.eq.s32 %p267, %r1169, 0; @%p267 bra $L__BB108_391; and.b32 %r3134, %r3134, -9; st.local.u32 [%rd7+-36], %r3134; $L__BB108_391: and.b32 %r1170, %r45, 1; setp.eq.b32 %p268, %r1170, 1; xor.pred %p270, %p268, %p264; not.pred %p271, %p270; @%p271 bra $L__BB108_520; ld.local.u32 %r172, [%rd7+172]; setp.lt.s32 %p272, %r172, 1; @%p272 bra $L__BB108_520; and.b32 %r1171, %r169, 2; setp.eq.s32 %p273, %r1171, 0; and.b32 %r173, %r169, 4; and.b32 %r174, %r169, 8; and.b32 %r175, %r169, 16; add.s32 %r176, %r172, -1; and.b32 %r177, %r172, 3; @%p273 bra $L__BB108_462; bra.uni $L__BB108_394; $L__BB108_462: setp.eq.s32 %p368, %r173, 0; @%p368 bra $L__BB108_494; bra.uni $L__BB108_463; $L__BB108_494: setp.lt.u32 %p418, %r176, 3; mov.u32 %r3129, 0; @%p418 bra $L__BB108_513; sub.s32 %r3123, %r172, %r177; mov.u32 %r3129, 0; setp.eq.s32 %p420, %r174, 0; setp.eq.s32 %p422, %r175, 0; $L__BB108_496: cvt.s64.s32 %rd70, %r3129; mul.wide.s32 %rd1060, %r3129, 80; add.s64 %rd1061, %rd1, %rd1060; add.s64 %rd71, %rd1061, 316; ld.local.u32 %r266, [%rd1061+316]; add.s32 %r1266, %r266, -21; setp.gt.u32 %p419, %r1266, 9; or.pred %p421, %p419, %p420; @%p421 bra $L__BB108_498; bra.uni $L__BB108_497; $L__BB108_498: setp.ne.s32 %p423, %r266, 31; or.pred %p424, %p423, %p422; @%p424 bra $L__BB108_500; mov.u32 %r1268, 35; st.local.u32 [%rd71], %r1268; mov.u32 %r1269, 0; st.local.u32 [%rd71+4], %r1269; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_500; $L__BB108_497: mov.u32 %r1267, 0; st.local.u32 [%rd71], %r1267; st.local.u32 [%rd71+4], %r1267; $L__BB108_500: ld.local.u32 %r269, [%rd71+80]; add.s32 %r1270, %r269, -21; setp.gt.u32 %p425, %r1270, 9; or.pred %p427, %p425, %p420; @%p427 bra $L__BB108_502; bra.uni $L__BB108_501; $L__BB108_502: setp.ne.s32 %p429, %r269, 31; or.pred %p430, %p429, %p422; @%p430 bra $L__BB108_504; mov.u32 %r1272, 35; st.local.u32 [%rd71+80], %r1272; mov.u32 %r1273, 0; st.local.u32 [%rd71+84], %r1273; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_504; $L__BB108_501: mov.u32 %r1271, 0; st.local.u32 [%rd71+80], %r1271; st.local.u32 [%rd71+84], %r1271; $L__BB108_504: ld.local.u32 %r272, [%rd71+160]; add.s32 %r1274, %r272, -21; setp.gt.u32 %p431, %r1274, 9; or.pred %p433, %p431, %p420; @%p433 bra $L__BB108_506; bra.uni $L__BB108_505; $L__BB108_506: setp.ne.s32 %p435, %r272, 31; or.pred %p436, %p435, %p422; @%p436 bra $L__BB108_508; mov.u32 %r1276, 35; st.local.u32 [%rd71+160], %r1276; mov.u32 %r1277, 0; st.local.u32 [%rd71+164], %r1277; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_508; $L__BB108_505: mov.u32 %r1275, 0; st.local.u32 [%rd71+160], %r1275; st.local.u32 [%rd71+164], %r1275; $L__BB108_508: ld.local.u32 %r275, [%rd71+240]; add.s32 %r1278, %r275, -21; setp.gt.u32 %p437, %r1278, 9; or.pred %p439, %p437, %p420; @%p439 bra $L__BB108_510; bra.uni $L__BB108_509; $L__BB108_510: setp.ne.s32 %p441, %r275, 31; or.pred %p442, %p441, %p422; @%p442 bra $L__BB108_512; mov.u32 %r1280, 35; st.local.u32 [%rd71+240], %r1280; mov.u32 %r1281, 0; st.local.u32 [%rd71+244], %r1281; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_512; $L__BB108_509: mov.u32 %r1279, 0; st.local.u32 [%rd71+240], %r1279; st.local.u32 [%rd71+244], %r1279; $L__BB108_512: cvt.u32.u64 %r1282, %rd70; add.s32 %r3129, %r1282, 4; add.s32 %r3123, %r3123, -4; setp.ne.s32 %p443, %r3123, 0; @%p443 bra $L__BB108_496; $L__BB108_513: setp.eq.s32 %p444, %r177, 0; @%p444 bra $L__BB108_520; setp.eq.s32 %p446, %r174, 0; setp.eq.s32 %p448, %r175, 0; $L__BB108_515: .pragma "nounroll"; cvt.s64.s32 %rd72, %r3129; mul.wide.s32 %rd1062, %r3129, 80; add.s64 %rd1063, %rd1, %rd1062; add.s64 %rd73, %rd1063, 316; ld.local.u32 %r285, [%rd1063+316]; add.s32 %r1283, %r285, -21; setp.gt.u32 %p445, %r1283, 9; or.pred %p447, %p445, %p446; @%p447 bra $L__BB108_517; bra.uni $L__BB108_516; $L__BB108_517: setp.ne.s32 %p449, %r285, 31; or.pred %p450, %p449, %p448; @%p450 bra $L__BB108_519; mov.u32 %r1285, 35; st.local.u32 [%rd73], %r1285; mov.u32 %r1286, 0; st.local.u32 [%rd73+4], %r1286; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_519; $L__BB108_516: mov.u32 %r1284, 0; st.local.u32 [%rd73], %r1284; st.local.u32 [%rd73+4], %r1284; $L__BB108_519: cvt.u32.u64 %r1287, %rd72; add.s32 %r3129, %r1287, 1; add.s32 %r177, %r177, -1; setp.ne.s32 %p451, %r177, 0; @%p451 bra $L__BB108_515; bra.uni $L__BB108_520; $L__BB108_137: setp.ltu.ftz.f32 %p254, %f649, 0f3727C5AC; mov.u64 %rd2929, 0; mov.u64 %rd2930, %rd2929; @%p254 bra $L__BB108_142; setp.eq.s32 %p255, %r168, 0; mov.u64 %rd2929, 0; mov.u64 %rd2927, %rd2929; mov.u64 %rd2928, %rd2929; @%p255 bra $L__BB108_140; add.s64 %rd2927, %rd1, 304; add.s64 %rd2928, %rd619, 304; mov.u64 %rd1035, 1065353216; st.local.u64 [%rd7+224], %rd1035; mov.u64 %rd1036, 4575657222473777152; st.local.u64 [%rd7+216], %rd1036; mov.u32 %r1153, 1; st.local.u32 [%rd7+172], %r1153; add.s32 %r1154, %r168, -1; st.local.u32 [%rd7+176], %r1154; $L__BB108_140: setp.eq.s64 %p256, %rd2928, 0; mov.u64 %rd2930, %rd2929; @%p256 bra $L__BB108_142; st.local.f32 [%rd2927+16], %f649; mov.u64 %rd2929, %rd2927; mov.u64 %rd2930, %rd2928; $L__BB108_142: setp.eq.s64 %p257, %rd2930, 0; @%p257 bra $L__BB108_385; st.local.f32 [%rd2929+20], %f15746; st.local.f32 [%rd2929+24], %f15747; st.local.f32 [%rd2929+28], %f15748; mov.u32 %r1155, 2139095039; st.local.u32 [%rd2929+32], %r1155; mov.u32 %r1156, 5; st.local.u32 [%rd2929+12], %r1156; mov.u32 %r1157, 4; st.local.u32 [%rd2929+36], %r1157; ld.local.u32 %r1158, [%rd7+-36]; or.b32 %r1159, %r1158, 12; st.local.u32 [%rd7+-36], %r1159; $L__BB108_385: ld.global.u32 %r1165, [%rd29]; and.b32 %r1166, %r1165, -15728641; st.global.u32 [%rd29], %r1166; $L__BB108_1312: ld.const.v2.u64 {%rd1700, %rd1701}, [kernel_params+2528]; cvta.to.global.u64 %rd1704, %rd1700; add.s64 %rd277, %rd1704, %rd837; ld.global.u32 %r601, [%rd277]; cvta.to.global.u64 %rd1706, %rd1701; shl.b64 %rd1707, %rd2, 1; add.s64 %rd278, %rd1706, %rd1707; ld.global.u16 %rs28, [%rd278]; cvt.u32.u16 %r602, %rs28; ld.const.u64 %rd1708, [kernel_params+2456]; cvta.to.global.u64 %rd1709, %rd1708; add.s64 %rd279, %rd1709, %rd837; ld.global.u32 %r603, [%rd279]; ld.const.u32 %r1894, [kernel_params+1488]; setp.eq.s32 %p1237, %r1894, 0; ld.local.u8 %rs83, [%rd7+-36]; and.b16 %rs84, %rs83, 8; setp.eq.s16 %p1238, %rs84, 0; or.pred %p1239, %p1237, %p1238; @%p1239 bra $L__BB108_1685; ld.global.u32 %r604, [%rd29]; ld.const.u64 %rd1710, [kernel_params+2464]; cvta.to.global.u64 %rd1711, %rd1710; add.s64 %rd280, %rd1711, %rd1707; ld.global.u16 %rs29, [%rd280]; add.s32 %r605, %r602, 1; ld.const.u32 %r606, [kernel_params+1596]; setp.eq.s32 %p1240, %r606, 0; brev.b32 %r1895, %r603; mul.lo.s32 %r1896, %r1895, 1025551850; xor.b32 %r607, %r1896, %r1895; @%p1240 bra $L__BB108_1316; ld.const.v2.u32 {%r1897, %r1898}, [kernel_params+1600]; mov.b32 %f1922, %r1898; setp.geu.ftz.f32 %p1241, %f1922, 0f3F800000; setp.lt.ftz.f32 %p1242, %f1922, 0f3F800000; ld.const.u32 %r1901, [kernel_params+1576]; selp.b32 %r1902, %r1901, %r601, %p1242; add.s16 %rs85, %rs28, 1; and.b16 %rs86, %rs85, 255; cvt.u32.u16 %r1903, %rs86; xor.b32 %r1904, %r1902, %r1903; mul.lo.s32 %r1905, %r1904, -512718531; shr.u32 %r1906, %r1902, 16; xor.b32 %r1907, %r1905, %r1906; bfe.u32 %r1908, %r1907, 4, 4; shr.u32 %r1909, %r1902, 8; xor.b32 %r1910, %r1907, %r1909; xor.b32 %r1911, %r1910, %r1908; mul.lo.s32 %r1912, %r1911, 153742143; shr.u32 %r1913, %r1902, 23; xor.b32 %r1914, %r1912, %r1913; bfe.u32 %r1915, %r1914, 1, 7; xor.b32 %r1916, %r1915, %r1914; shr.u32 %r1917, %r1902, 27; or.b32 %r1918, %r1917, 1; mul.lo.s32 %r1919, %r1918, %r1916; mul.lo.s32 %r1920, %r1919, -1074439621; bfe.u32 %r1921, %r1920, 2, 6; xor.b32 %r1922, %r1921, %r1920; mul.lo.s32 %r1923, %r1922, -1638916925; bfe.u32 %r1924, %r1923, 2, 6; xor.b32 %r1925, %r1924, %r1923; mul.lo.s32 %r1926, %r1925, -933190689; and.b32 %r1927, %r1926, 255; bfe.u32 %r1928, %r1926, 5, 3; xor.b32 %r1929, %r1928, %r1927; add.s32 %r1930, %r1897, -1; xor.b32 %r608, %r605, 61; xor.b32 %r1931, %r1902, %r608; mul.lo.s32 %r1932, %r1931, 9; shr.u32 %r1933, %r1932, 4; xor.b32 %r1934, %r1933, %r1932; mul.lo.s32 %r1935, %r1934, 668265261; add.s32 %r1936, %r1935, %r607; shr.u32 %r1937, %r1935, 16; or.b32 %r1938, %r1937, 1; mul.lo.s32 %r1939, %r1938, %r1936; mul.lo.s32 %r1940, %r1939, 89287766; xor.b32 %r1941, %r1940, %r1939; mul.lo.s32 %r1942, %r1941, 1403136100; xor.b32 %r1943, %r1942, %r1941; brev.b32 %r1944, %r1943; neg.s32 %r1945, %r1897; and.b32 %r1946, %r603, %r1945; and.b32 %r1947, %r1944, %r1930; or.b32 %r1948, %r1947, %r1946; mad.lo.s32 %r1949, %r1929, %r1897, %r1948; shl.b32 %r1950, %r1897, 8; rem.u32 %r1951, %r1949, %r1950; ld.const.u64 %rd1713, [kernel_params+2424]; cvta.to.global.u64 %rd1714, %rd1713; shl.b32 %r1952, %r1951, 1; mul.wide.u32 %rd1715, %r1952, 4; add.s64 %rd1716, %rd1714, %rd1715; ld.global.f32 %f16182, [%rd1716]; ld.global.f32 %f16183, [%rd1716+4]; @%p1241 bra $L__BB108_1320; xor.b32 %r1953, %r608, %r601; mul.lo.s32 %r1954, %r1953, 9; shr.u32 %r1955, %r1954, 4; xor.b32 %r1956, %r1955, %r1954; mul.lo.s32 %r1957, %r1956, 668265261; cvt.rn.f32.u32 %f9681, %r1957; mul.ftz.f32 %f9682, %f9681, 0f2F7FFFFE; xor.b32 %r1958, %r601, %r605; xor.b32 %r1959, %r1958, -905047700; mul.lo.s32 %r1960, %r1959, 9; shr.u32 %r1961, %r1960, 4; xor.b32 %r1962, %r1961, %r1960; mul.lo.s32 %r1963, %r1962, 668265261; cvt.rn.f32.u32 %f9683, %r1963; mul.ftz.f32 %f9684, %f9683, 0f2F7FFFFE; fma.rn.ftz.f32 %f9685, %f9682, %f1922, %f16182; fma.rn.ftz.f32 %f9686, %f9684, %f1922, %f16183; cvt.rmi.ftz.f32.f32 %f9687, %f9685; sub.ftz.f32 %f16182, %f9685, %f9687; cvt.rmi.ftz.f32.f32 %f9688, %f9686; sub.ftz.f32 %f16183, %f9686, %f9688; bra.uni $L__BB108_1320; $L__BB108_1316: shr.u32 %r1965, %r605, 16; xor.b32 %r1966, %r1965, %r605; mul.lo.s32 %r1967, %r1966, 569420461; shr.u32 %r1968, %r1967, 15; xor.b32 %r1969, %r1968, %r1967; mul.lo.s32 %r1970, %r1969, -749064809; shr.u32 %r1971, %r1970, 15; xor.b32 %r1972, %r601, %r1970; xor.b32 %r1973, %r1972, %r1971; xor.b32 %r609, %r1973, -419546133; xor.b32 %r1974, %r1973, 508809841; add.s32 %r1975, %r1974, %r607; shr.u32 %r1976, %r1974, 16; or.b32 %r1977, %r1976, 1; mul.lo.s32 %r1978, %r1977, %r1975; mul.lo.s32 %r1979, %r1978, 89287766; xor.b32 %r1980, %r1979, %r1978; mul.lo.s32 %r1981, %r1980, 1403136100; xor.b32 %r610, %r1981, %r1980; setp.eq.s32 %p1243, %r610, 0; mov.u32 %r3225, 0; @%p1243 bra $L__BB108_1319; mov.u32 %r3222, 0; mov.u64 %rd1718, sobol_burley_table; mov.u32 %r3223, %r610; mov.u32 %r3225, %r3222; $L__BB108_1318: clz.b32 %r1984, %r3223; add.s32 %r1985, %r1984, %r3222; mul.wide.u32 %rd1717, %r1985, 4; add.s64 %rd1719, %rd1718, %rd1717; ld.const.u32 %r1986, [%rd1719+128]; xor.b32 %r3225, %r1986, %r3225; add.s32 %r3222, %r1985, 1; shl.b32 %r1987, %r3223, %r1984; shl.b32 %r3223, %r1987, 1; setp.ne.s32 %p1244, %r3223, 0; @%p1244 bra $L__BB108_1318; $L__BB108_1319: xor.b32 %r1988, %r609, -525684874; brev.b32 %r1989, %r610; shr.u32 %r1990, %r1988, 16; or.b32 %r1991, %r1990, 1; mul.lo.s32 %r1992, %r1989, 1025551850; xor.b32 %r1993, %r1992, %r1989; add.s32 %r1994, %r1993, %r1988; mul.lo.s32 %r1995, %r1994, %r1991; mul.lo.s32 %r1996, %r1995, 89287766; xor.b32 %r1997, %r1996, %r1995; xor.b32 %r1998, %r609, -1802089138; mul.lo.s32 %r1999, %r3225, 1025551850; xor.b32 %r2000, %r1999, %r3225; add.s32 %r2001, %r2000, %r1998; shr.u32 %r2002, %r1998, 16; or.b32 %r2003, %r2002, 1; mul.lo.s32 %r2004, %r2001, %r2003; mul.lo.s32 %r2005, %r2004, 89287766; xor.b32 %r2006, %r2005, %r2004; mul.lo.s32 %r2007, %r2006, 1403136100; xor.b32 %r2008, %r2007, %r2006; brev.b32 %r2009, %r2008; cvt.rn.f32.u32 %f9689, %r2009; mul.lo.s32 %r2010, %r1997, 1403136100; xor.b32 %r2011, %r2010, %r1997; brev.b32 %r2012, %r2011; cvt.rn.f32.u32 %f9690, %r2012; mul.ftz.f32 %f16183, %f9689, 0f2F7FFFFE; mul.ftz.f32 %f16182, %f9690, 0f2F7FFFFE; $L__BB108_1320: ld.local.f32 %f9691, [%rd7+-4]; ld.local.v4.f32 {%f9692, %f9693, %f9694, %f9695}, [%rd1]; cvt.u32.u16 %r2013, %rs29; add.u64 %rd1721, %SP, 0; { // callseq 779, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f16182; .param .b32 param1; st.param.f32 [param1+0], %f16183; .param .b32 param2; st.param.f32 [param2+0], %f9691; .param .align 4 .b8 param3[12]; st.param.f32 [param3+0], %f9692; st.param.f32 [param3+4], %f9693; st.param.f32 [param3+8], %f9694; .param .b32 param4; st.param.b32 [param4+0], %r2013; .param .b32 param5; st.param.b32 [param5+0], %r604; .param .b64 param6; st.param.b64 [param6+0], %rd1721; .param .b32 retval0; call.uni (retval0), _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample, ( param0, param1, param2, param3, param4, param5, param6 ); ld.param.b32 %r2014, [retval0+0]; } // callseq 779 cvt.u16.u32 %rs87, %r2014; setp.eq.s16 %p1245, %rs87, 0; @%p1245 bra $L__BB108_1685; ld.const.u64 %rd2918, [kernel_params+2408]; cvta.to.global.u64 %rd2917, %rd2918; cvta.to.local.u64 %rd1723, %rd1721; add.s64 %rd282, %rd1723, 24; ld.local.f32 %f9702, [%rd7+-76]; ld.local.f32 %f1935, [%rd1723+24]; ld.local.v2.f32 {%f9703, %f9704}, [%rd7+-72]; ld.local.f32 %f1936, [%rd1723+28]; mul.ftz.f32 %f9707, %f1936, %f9703; fma.rn.ftz.f32 %f9708, %f1935, %f9702, %f9707; ld.local.f32 %f1937, [%rd1723+32]; fma.rn.ftz.f32 %f1938, %f1937, %f9704, %f9708; ld.local.f32 %f1939, [%rd7+-4]; ld.local.u32 %r618, [%rd1723+64]; and.b32 %r3229, %r618, 4194303; shl.b32 %r2015, %r618, 5; cvt.u64.u32 %rd1724, %r2015; and.b64 %rd1725, %rd1724, 134217696; add.s64 %rd1726, %rd2917, %rd1725; add.s64 %rd283, %rd1726, 16; ld.global.u32 %r620, [%rd1726+16]; and.b32 %r2016, %r620, 134217728; setp.eq.s32 %p1247, %r2016, 0; mov.pred %p2184, -1; mov.f32 %f16225, 0f00000000; mov.f32 %f16226, %f16225; mov.f32 %f16227, %f16225; @%p1247 bra $L__BB108_1323; ld.global.f32 %f16225, [%rd283+-16]; ld.global.f32 %f16226, [%rd283+-12]; ld.global.f32 %f16227, [%rd283+-8]; mov.pred %p2184, 0; $L__BB108_1323: @%p2184 bra $L__BB108_1327; bra.uni $L__BB108_1324; $L__BB108_1327: ld.local.u32 %r621, [%rd282+52]; setp.eq.s32 %p1251, %r621, 2; ld.local.f32 %f1949, [%rd282+-24]; ld.local.f32 %f1950, [%rd282+-20]; ld.local.f32 %f1951, [%rd282+-16]; add.u64 %rd1727, %SP, 80; or.b64 %rd284, %rd1727, 12; neg.ftz.f32 %f1952, %f1935; neg.ftz.f32 %f1953, %f1936; neg.ftz.f32 %f1954, %f1937; @%p1251 bra $L__BB108_1365; bra.uni $L__BB108_1328; $L__BB108_1365: ld.const.u64 %rd2921, [kernel_params+2408]; cvta.to.global.u64 %rd2920, %rd2921; add.s64 %rd1886, %rd284, 20; add.s64 %rd1887, %rd284, 4; add.s64 %rd1889, %rd1727, 248; cvta.to.local.u64 %rd1890, %rd1889; add.s64 %rd1891, %rd1727, 240; cvta.to.local.u64 %rd1892, %rd1891; st.local.v2.f32 [%rd1892], {%f1949, %f1950}; st.local.f32 [%rd1890], %f1951; cvta.to.local.u64 %rd1893, %rd1727; st.local.v4.f32 [%rd1893], {%f1935, %f1936, %f1937, %f1952}; cvta.to.local.u64 %rd1894, %rd1887; st.local.v4.f32 [%rd1894], {%f1953, %f1954, %f1952, %f1953}; cvta.to.local.u64 %rd1895, %rd1886; st.local.v4.f32 [%rd1895], {%f1954, %f1952, %f1953, %f1954}; ld.const.u32 %r2048, [kernel_params+1024]; and.b32 %r3229, %r2048, 4194303; shl.b32 %r2049, %r2048, 5; cvt.u64.u32 %rd1896, %r2049; and.b64 %rd1897, %rd1896, 134217696; add.s64 %rd1898, %rd2920, %rd1897; add.s64 %rd1899, %rd1727, 76; cvta.to.local.u64 %rd1900, %rd1899; ld.global.u32 %r2050, [%rd1898+16]; mov.u32 %r2051, -1; st.local.u32 [%rd1900], %r2051; mov.u32 %r2052, 0; add.s64 %rd1901, %rd284, 36; cvta.to.local.u64 %rd1902, %rd1901; st.local.v4.u32 [%rd1902], {%r2048, %r2050, %r2052, %r2051}; add.s64 %rd1903, %rd1727, 68; cvta.to.local.u64 %rd1904, %rd1903; st.local.u32 [%rd1904], %r2052; add.s64 %rd1905, %rd1727, 72; cvta.to.local.u64 %rd1906, %rd1905; st.local.u32 [%rd1906], %r2052; add.s64 %rd1907, %rd1727, 136; cvta.to.local.u64 %rd1908, %rd1907; add.s64 %rd1909, %rd1727, 128; cvta.to.local.u64 %rd1910, %rd1909; mov.f32 %f10192, 0f00000000; st.local.v2.f32 [%rd1910], {%f10192, %f10192}; st.local.u32 [%rd1908], %r2052; mov.b32 %r2053, %f1939; add.s64 %rd1911, %rd1727, 80; cvta.to.local.u64 %rd1912, %rd1911; st.local.v4.u32 [%rd1912], {%r2051, %r2053, %r2052, %r2052}; add.s64 %rd1913, %rd1727, 96; cvta.to.local.u64 %rd1914, %rd1913; st.local.v4.f32 [%rd1914], {%f10192, %f10192, %f10192, %f10192}; add.s64 %rd1915, %rd1727, 112; cvta.to.local.u64 %rd1916, %rd1915; st.local.v4.f32 [%rd1916], {%f10192, %f10192, %f10192, %f10192}; bra.uni $L__BB108_1366; $L__BB108_1324: ld.local.u32 %r2017, [%rd282+36]; setp.eq.s32 %p1249, %r2017, -1; @%p1249 bra $L__BB108_1577; ld.local.f32 %f1946, [%rd282+-12]; ld.local.f32 %f1947, [%rd282+-8]; mul.ftz.f32 %f9709, %f1947, %f1936; fma.rn.ftz.f32 %f9710, %f1946, %f1935, %f9709; ld.local.f32 %f1948, [%rd282+-4]; fma.rn.ftz.f32 %f9711, %f1948, %f1937, %f9710; setp.leu.ftz.f32 %p1250, %f9711, 0f00000000; @%p1250 bra $L__BB108_1577; neg.ftz.f32 %f9712, %f1946; st.local.f32 [%rd282+-12], %f9712; neg.ftz.f32 %f9713, %f1947; st.local.f32 [%rd282+-8], %f9713; neg.ftz.f32 %f9714, %f1948; st.local.f32 [%rd282+-4], %f9714; bra.uni $L__BB108_1577; $L__BB108_90: ld.local.v2.f32 {%f5238, %f5239}, [%rd7+32]; ld.local.v4.f32 {%f5242, %f5243, %f5244, %f5245}, [%rd7+56]; ld.local.f32 %f5249, [%rd7+28]; mul.ftz.f32 %f5250, %f5238, %f5243; fma.rn.ftz.f32 %f5251, %f5249, %f5242, %f5250; fma.rn.ftz.f32 %f15693, %f5239, %f5244, %f5251; ld.local.v4.f32 {%f5252, %f5253, %f5254, %f5255}, [%rd7+72]; mul.ftz.f32 %f5259, %f5238, %f5253; fma.rn.ftz.f32 %f5260, %f5249, %f5252, %f5259; fma.rn.ftz.f32 %f15694, %f5239, %f5254, %f5260; ld.local.v4.f32 {%f5261, %f5262, %f5263, %f5264}, [%rd7+88]; mul.ftz.f32 %f5268, %f5238, %f5262; fma.rn.ftz.f32 %f5269, %f5249, %f5261, %f5268; fma.rn.ftz.f32 %f15695, %f5239, %f5263, %f5269; $L__BB108_92: st.local.f32 [%rd7+28], %f15693; st.local.f32 [%rd7+32], %f15694; st.local.f32 [%rd7+36], %f15695; @%p71 bra $L__BB108_94; bra.uni $L__BB108_93; $L__BB108_94: ld.const.u64 %rd825, [kernel_params+2160]; cvta.to.global.u64 %rd826, %rd825; ld.local.u32 %r1029, [%rd7+-12]; mul.wide.s32 %rd827, %r1029, 208; add.s64 %rd828, %rd826, %rd827; ld.global.v4.f32 {%f5339, %f5340, %f5341, %f5342}, [%rd828]; ld.global.v4.f32 {%f5346, %f5347, %f5348, %f5349}, [%rd828+16]; ld.global.v4.f32 {%f5353, %f5354, %f5355, %f5356}, [%rd828+32]; ld.local.v4.f32 {%f5360, %f5361, %f5362, %f5363}, [%rd7+40]; mul.ftz.f32 %f5367, %f5340, %f5361; fma.rn.ftz.f32 %f5368, %f5360, %f5339, %f5367; mul.ftz.f32 %f5369, %f5347, %f5361; fma.rn.ftz.f32 %f5370, %f5360, %f5346, %f5369; mul.ftz.f32 %f5371, %f5354, %f5361; fma.rn.ftz.f32 %f5372, %f5360, %f5353, %f5371; fma.rn.ftz.f32 %f5373, %f5355, %f5362, %f5372; fma.rn.ftz.f32 %f5374, %f5348, %f5362, %f5370; fma.rn.ftz.f32 %f5375, %f5341, %f5362, %f5368; st.local.v2.f32 [%rd7+40], {%f5375, %f5374}; st.local.f32 [%rd7+48], %f5373; bra.uni $L__BB108_95; $L__BB108_93: ld.local.v4.f32 {%f5302, %f5303, %f5304, %f5305}, [%rd7+40]; ld.local.v4.f32 {%f5309, %f5310, %f5311, %f5312}, [%rd7+56]; mul.ftz.f32 %f5316, %f5303, %f5310; fma.rn.ftz.f32 %f5317, %f5302, %f5309, %f5316; ld.local.v4.f32 {%f5318, %f5319, %f5320, %f5321}, [%rd7+72]; mul.ftz.f32 %f5325, %f5303, %f5319; fma.rn.ftz.f32 %f5326, %f5302, %f5318, %f5325; ld.local.v4.f32 {%f5327, %f5328, %f5329, %f5330}, [%rd7+88]; mul.ftz.f32 %f5334, %f5303, %f5328; fma.rn.ftz.f32 %f5335, %f5302, %f5327, %f5334; fma.rn.ftz.f32 %f5336, %f5304, %f5329, %f5335; fma.rn.ftz.f32 %f5337, %f5304, %f5320, %f5326; fma.rn.ftz.f32 %f5338, %f5304, %f5311, %f5317; st.local.v2.f32 [%rd7+40], {%f5338, %f5337}; st.local.f32 [%rd7+48], %f5336; bra.uni $L__BB108_95; $L__BB108_1328: add.s64 %rd1728, %rd284, 20; add.s64 %rd1729, %rd284, 4; add.s64 %rd1730, %rd284, 36; ld.local.f32 %f16217, [%rd282+-12]; ld.local.f32 %f16218, [%rd282+-8]; ld.local.f32 %f1957, [%rd282+-4]; ld.local.u32 %r622, [%rd282+32]; cvt.s64.s32 %rd285, %r622; ld.local.u32 %r623, [%rd282+36]; cvt.s64.s32 %rd286, %r623; ld.local.f32 %f1958, [%rd282+16]; ld.local.f32 %f1959, [%rd282+20]; ld.local.f32 %f1960, [%rd282+12]; ld.local.u32 %r624, [%rd282+44]; cvta.to.local.u64 %rd1732, %rd1727; st.local.v4.f32 [%rd1732], {%f1949, %f1950, %f1951, %f16217}; cvta.to.local.u64 %rd287, %rd1729; st.local.v4.f32 [%rd287], {%f16218, %f1957, %f16217, %f16218}; cvta.to.local.u64 %rd1733, %rd1728; st.local.v4.f32 [%rd1733], {%f1957, %f1952, %f1953, %f1954}; cvta.to.local.u64 %rd1734, %rd1730; st.local.u32 [%rd1734], %r618; setp.eq.s32 %p1252, %r623, -1; add.s64 %rd288, %rd1727, 64; @%p1252 bra $L__BB108_1330; cvta.to.local.u64 %rd1735, %rd288; mov.u32 %r3226, 1; st.local.u32 [%rd1735], %r3226; bra.uni $L__BB108_1333; $L__BB108_394: setp.eq.s32 %p274, %r173, 0; @%p274 bra $L__BB108_431; bra.uni $L__BB108_395; $L__BB108_431: setp.lt.u32 %p329, %r176, 3; mov.u32 %r3103, 0; @%p329 bra $L__BB108_454; sub.s32 %r3097, %r172, %r177; mov.u32 %r3103, 0; setp.ne.s32 %p352, %r174, 0; setp.eq.s32 %p355, %r175, 0; $L__BB108_433: cvt.s64.s32 %rd62, %r3103; mul.wide.s32 %rd1052, %r3103, 80; add.s64 %rd1053, %rd1, %rd1052; add.s64 %rd63, %rd1053, 316; ld.local.u32 %r210, [%rd1053+316]; add.s32 %r1208, %r210, -2; setp.lt.u32 %p330, %r1208, 7; @%p330 bra $L__BB108_437; add.s32 %r1209, %r210, -21; setp.lt.u32 %p332, %r1209, 10; and.pred %p333, %p332, %p352; @%p333 bra $L__BB108_437; setp.ne.s32 %p335, %r210, 31; or.pred %p336, %p335, %p355; @%p336 bra $L__BB108_438; mov.u32 %r1210, 35; st.local.u32 [%rd63], %r1210; mov.u32 %r1211, 0; st.local.u32 [%rd63+4], %r1211; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_438; $L__BB108_437: mov.u32 %r1212, 0; st.local.u32 [%rd63], %r1212; st.local.u32 [%rd63+4], %r1212; $L__BB108_438: ld.local.u32 %r213, [%rd63+80]; add.s32 %r1213, %r213, -2; setp.lt.u32 %p337, %r1213, 7; @%p337 bra $L__BB108_442; add.s32 %r1214, %r213, -21; setp.lt.u32 %p339, %r1214, 10; and.pred %p340, %p339, %p352; @%p340 bra $L__BB108_442; setp.ne.s32 %p342, %r213, 31; or.pred %p343, %p342, %p355; @%p343 bra $L__BB108_443; mov.u32 %r1215, 35; st.local.u32 [%rd63+80], %r1215; mov.u32 %r1216, 0; st.local.u32 [%rd63+84], %r1216; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_443; $L__BB108_442: mov.u32 %r1217, 0; st.local.u32 [%rd63+80], %r1217; st.local.u32 [%rd63+84], %r1217; $L__BB108_443: ld.local.u32 %r216, [%rd63+160]; add.s32 %r1218, %r216, -2; setp.lt.u32 %p344, %r1218, 7; @%p344 bra $L__BB108_447; add.s32 %r1219, %r216, -21; setp.lt.u32 %p346, %r1219, 10; and.pred %p347, %p346, %p352; @%p347 bra $L__BB108_447; setp.ne.s32 %p349, %r216, 31; or.pred %p350, %p349, %p355; @%p350 bra $L__BB108_448; mov.u32 %r1220, 35; st.local.u32 [%rd63+160], %r1220; mov.u32 %r1221, 0; st.local.u32 [%rd63+164], %r1221; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_448; $L__BB108_447: mov.u32 %r1222, 0; st.local.u32 [%rd63+160], %r1222; st.local.u32 [%rd63+164], %r1222; $L__BB108_448: ld.local.u32 %r219, [%rd63+240]; add.s32 %r1223, %r219, -2; setp.lt.u32 %p351, %r1223, 7; @%p351 bra $L__BB108_452; add.s32 %r1224, %r219, -21; setp.lt.u32 %p353, %r1224, 10; and.pred %p354, %p353, %p352; @%p354 bra $L__BB108_452; setp.ne.s32 %p356, %r219, 31; or.pred %p357, %p356, %p355; @%p357 bra $L__BB108_453; mov.u32 %r1225, 35; st.local.u32 [%rd63+240], %r1225; mov.u32 %r1226, 0; st.local.u32 [%rd63+244], %r1226; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_453; $L__BB108_452: mov.u32 %r1227, 0; st.local.u32 [%rd63+240], %r1227; st.local.u32 [%rd63+244], %r1227; $L__BB108_453: cvt.u32.u64 %r1228, %rd62; add.s32 %r3103, %r1228, 4; add.s32 %r3097, %r3097, -4; setp.ne.s32 %p358, %r3097, 0; @%p358 bra $L__BB108_433; $L__BB108_454: setp.eq.s32 %p359, %r177, 0; @%p359 bra $L__BB108_520; setp.ne.s32 %p361, %r174, 0; setp.eq.s32 %p364, %r175, 0; $L__BB108_456: .pragma "nounroll"; cvt.s64.s32 %rd64, %r3103; mul.wide.s32 %rd1054, %r3103, 80; add.s64 %rd1055, %rd1, %rd1054; add.s64 %rd65, %rd1055, 316; ld.local.u32 %r229, [%rd1055+316]; add.s32 %r1229, %r229, -2; setp.lt.u32 %p360, %r1229, 7; @%p360 bra $L__BB108_460; add.s32 %r1230, %r229, -21; setp.lt.u32 %p362, %r1230, 10; and.pred %p363, %p362, %p361; @%p363 bra $L__BB108_460; setp.ne.s32 %p365, %r229, 31; or.pred %p366, %p365, %p364; @%p366 bra $L__BB108_461; mov.u32 %r1231, 35; st.local.u32 [%rd65], %r1231; mov.u32 %r1232, 0; st.local.u32 [%rd65+4], %r1232; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_461; $L__BB108_460: mov.u32 %r1233, 0; st.local.u32 [%rd65], %r1233; st.local.u32 [%rd65+4], %r1233; $L__BB108_461: cvt.u32.u64 %r1234, %rd64; add.s32 %r3103, %r1234, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p367, %r177, 0; @%p367 bra $L__BB108_520; bra.uni $L__BB108_456; $L__BB108_1330: cvta.to.local.u64 %rd289, %rd288; setp.eq.s32 %p1253, %r624, -1; @%p1253 bra $L__BB108_1332; mov.u32 %r3226, 32; st.local.u32 [%rd289], %r3226; bra.uni $L__BB108_1333; $L__BB108_463: setp.lt.u32 %p369, %r176, 3; mov.u32 %r3116, 0; @%p369 bra $L__BB108_486; sub.s32 %r3110, %r172, %r177; mov.u32 %r3116, 0; setp.ne.s32 %p400, %r174, 0; setp.eq.s32 %p403, %r175, 0; $L__BB108_465: cvt.s64.s32 %rd66, %r3116; mul.wide.s32 %rd1056, %r3116, 80; add.s64 %rd1057, %rd1, %rd1056; add.s64 %rd67, %rd1057, 316; ld.local.u32 %r238, [%rd1057+316]; add.s32 %r1237, %r238, -9; setp.lt.u32 %p370, %r1237, 12; setp.eq.s32 %p371, %r238, 29; or.pred %p372, %p370, %p371; @%p372 bra $L__BB108_469; add.s32 %r1238, %r238, -21; setp.lt.u32 %p374, %r1238, 10; and.pred %p375, %p374, %p400; @%p375 bra $L__BB108_469; setp.ne.s32 %p377, %r238, 31; or.pred %p378, %p377, %p403; @%p378 bra $L__BB108_470; mov.u32 %r1239, 35; st.local.u32 [%rd67], %r1239; mov.u32 %r1240, 0; st.local.u32 [%rd67+4], %r1240; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_470; $L__BB108_469: mov.u32 %r1241, 0; st.local.u32 [%rd67], %r1241; st.local.u32 [%rd67+4], %r1241; $L__BB108_470: ld.local.u32 %r241, [%rd67+80]; add.s32 %r1242, %r241, -9; setp.lt.u32 %p379, %r1242, 12; setp.eq.s32 %p380, %r241, 29; or.pred %p381, %p379, %p380; @%p381 bra $L__BB108_474; add.s32 %r1243, %r241, -21; setp.lt.u32 %p383, %r1243, 10; and.pred %p384, %p383, %p400; @%p384 bra $L__BB108_474; setp.ne.s32 %p386, %r241, 31; or.pred %p387, %p386, %p403; @%p387 bra $L__BB108_475; mov.u32 %r1244, 35; st.local.u32 [%rd67+80], %r1244; mov.u32 %r1245, 0; st.local.u32 [%rd67+84], %r1245; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_475; $L__BB108_474: mov.u32 %r1246, 0; st.local.u32 [%rd67+80], %r1246; st.local.u32 [%rd67+84], %r1246; $L__BB108_475: ld.local.u32 %r244, [%rd67+160]; add.s32 %r1247, %r244, -9; setp.lt.u32 %p388, %r1247, 12; setp.eq.s32 %p389, %r244, 29; or.pred %p390, %p388, %p389; @%p390 bra $L__BB108_479; add.s32 %r1248, %r244, -21; setp.lt.u32 %p392, %r1248, 10; and.pred %p393, %p392, %p400; @%p393 bra $L__BB108_479; setp.ne.s32 %p395, %r244, 31; or.pred %p396, %p395, %p403; @%p396 bra $L__BB108_480; mov.u32 %r1249, 35; st.local.u32 [%rd67+160], %r1249; mov.u32 %r1250, 0; st.local.u32 [%rd67+164], %r1250; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_480; $L__BB108_479: mov.u32 %r1251, 0; st.local.u32 [%rd67+160], %r1251; st.local.u32 [%rd67+164], %r1251; $L__BB108_480: ld.local.u32 %r247, [%rd67+240]; add.s32 %r1252, %r247, -9; setp.lt.u32 %p397, %r1252, 12; setp.eq.s32 %p398, %r247, 29; or.pred %p399, %p397, %p398; @%p399 bra $L__BB108_484; add.s32 %r1253, %r247, -21; setp.lt.u32 %p401, %r1253, 10; and.pred %p402, %p401, %p400; @%p402 bra $L__BB108_484; setp.ne.s32 %p404, %r247, 31; or.pred %p405, %p404, %p403; @%p405 bra $L__BB108_485; mov.u32 %r1254, 35; st.local.u32 [%rd67+240], %r1254; mov.u32 %r1255, 0; st.local.u32 [%rd67+244], %r1255; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_485; $L__BB108_484: mov.u32 %r1256, 0; st.local.u32 [%rd67+240], %r1256; st.local.u32 [%rd67+244], %r1256; $L__BB108_485: cvt.u32.u64 %r1257, %rd66; add.s32 %r3116, %r1257, 4; add.s32 %r3110, %r3110, -4; setp.ne.s32 %p406, %r3110, 0; @%p406 bra $L__BB108_465; $L__BB108_486: setp.eq.s32 %p407, %r177, 0; @%p407 bra $L__BB108_520; setp.ne.s32 %p411, %r174, 0; setp.eq.s32 %p414, %r175, 0; $L__BB108_488: .pragma "nounroll"; cvt.s64.s32 %rd68, %r3116; mul.wide.s32 %rd1058, %r3116, 80; add.s64 %rd1059, %rd1, %rd1058; add.s64 %rd69, %rd1059, 316; ld.local.u32 %r257, [%rd1059+316]; add.s32 %r1258, %r257, -9; setp.lt.u32 %p408, %r1258, 12; setp.eq.s32 %p409, %r257, 29; or.pred %p410, %p408, %p409; @%p410 bra $L__BB108_492; add.s32 %r1259, %r257, -21; setp.lt.u32 %p412, %r1259, 10; and.pred %p413, %p412, %p411; @%p413 bra $L__BB108_492; setp.ne.s32 %p415, %r257, 31; or.pred %p416, %p415, %p414; @%p416 bra $L__BB108_493; mov.u32 %r1260, 35; st.local.u32 [%rd69], %r1260; mov.u32 %r1261, 0; st.local.u32 [%rd69+4], %r1261; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_493; $L__BB108_492: mov.u32 %r1262, 0; st.local.u32 [%rd69], %r1262; st.local.u32 [%rd69+4], %r1262; $L__BB108_493: cvt.u32.u64 %r1263, %rd68; add.s32 %r3116, %r1263, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p417, %r177, 0; @%p417 bra $L__BB108_520; bra.uni $L__BB108_488; $L__BB108_395: setp.lt.u32 %p275, %r176, 3; mov.u32 %r3090, 0; @%p275 bra $L__BB108_422; sub.s32 %r3084, %r172, %r177; mov.u32 %r3090, 0; setp.ne.s32 %p310, %r174, 0; setp.eq.s32 %p313, %r175, 0; $L__BB108_397: cvt.s64.s32 %rd58, %r3090; mul.wide.s32 %rd1048, %r3090, 80; add.s64 %rd1049, %rd1, %rd1048; add.s64 %rd59, %rd1049, 316; ld.local.u32 %r182, [%rd1049+316]; add.s32 %r1174, %r182, -2; setp.lt.u32 %p276, %r1174, 7; @%p276 bra $L__BB108_402; add.s32 %r1175, %r182, -9; setp.lt.u32 %p277, %r1175, 12; setp.eq.s32 %p278, %r182, 29; or.pred %p279, %p277, %p278; @%p279 bra $L__BB108_402; add.s32 %r1176, %r182, -21; setp.lt.u32 %p281, %r1176, 10; and.pred %p282, %p281, %p310; @%p282 bra $L__BB108_402; setp.ne.s32 %p284, %r182, 31; or.pred %p285, %p284, %p313; @%p285 bra $L__BB108_403; mov.u32 %r1177, 35; st.local.u32 [%rd59], %r1177; mov.u32 %r1178, 0; st.local.u32 [%rd59+4], %r1178; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_403; $L__BB108_402: mov.u32 %r1179, 0; st.local.u32 [%rd59], %r1179; st.local.u32 [%rd59+4], %r1179; $L__BB108_403: ld.local.u32 %r185, [%rd59+80]; add.s32 %r1180, %r185, -2; setp.lt.u32 %p286, %r1180, 7; @%p286 bra $L__BB108_408; add.s32 %r1181, %r185, -9; setp.lt.u32 %p287, %r1181, 12; setp.eq.s32 %p288, %r185, 29; or.pred %p289, %p287, %p288; @%p289 bra $L__BB108_408; add.s32 %r1182, %r185, -21; setp.lt.u32 %p291, %r1182, 10; and.pred %p292, %p291, %p310; @%p292 bra $L__BB108_408; setp.ne.s32 %p294, %r185, 31; or.pred %p295, %p294, %p313; @%p295 bra $L__BB108_409; mov.u32 %r1183, 35; st.local.u32 [%rd59+80], %r1183; mov.u32 %r1184, 0; st.local.u32 [%rd59+84], %r1184; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_409; $L__BB108_408: mov.u32 %r1185, 0; st.local.u32 [%rd59+80], %r1185; st.local.u32 [%rd59+84], %r1185; $L__BB108_409: ld.local.u32 %r188, [%rd59+160]; add.s32 %r1186, %r188, -2; setp.lt.u32 %p296, %r1186, 7; @%p296 bra $L__BB108_414; add.s32 %r1187, %r188, -9; setp.lt.u32 %p297, %r1187, 12; setp.eq.s32 %p298, %r188, 29; or.pred %p299, %p297, %p298; @%p299 bra $L__BB108_414; add.s32 %r1188, %r188, -21; setp.lt.u32 %p301, %r1188, 10; and.pred %p302, %p301, %p310; @%p302 bra $L__BB108_414; setp.ne.s32 %p304, %r188, 31; or.pred %p305, %p304, %p313; @%p305 bra $L__BB108_415; mov.u32 %r1189, 35; st.local.u32 [%rd59+160], %r1189; mov.u32 %r1190, 0; st.local.u32 [%rd59+164], %r1190; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_415; $L__BB108_414: mov.u32 %r1191, 0; st.local.u32 [%rd59+160], %r1191; st.local.u32 [%rd59+164], %r1191; $L__BB108_415: ld.local.u32 %r191, [%rd59+240]; add.s32 %r1192, %r191, -2; setp.lt.u32 %p306, %r1192, 7; @%p306 bra $L__BB108_420; add.s32 %r1193, %r191, -9; setp.lt.u32 %p307, %r1193, 12; setp.eq.s32 %p308, %r191, 29; or.pred %p309, %p307, %p308; @%p309 bra $L__BB108_420; add.s32 %r1194, %r191, -21; setp.lt.u32 %p311, %r1194, 10; and.pred %p312, %p311, %p310; @%p312 bra $L__BB108_420; setp.ne.s32 %p314, %r191, 31; or.pred %p315, %p314, %p313; @%p315 bra $L__BB108_421; mov.u32 %r1195, 35; st.local.u32 [%rd59+240], %r1195; mov.u32 %r1196, 0; st.local.u32 [%rd59+244], %r1196; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_421; $L__BB108_420: mov.u32 %r1197, 0; st.local.u32 [%rd59+240], %r1197; st.local.u32 [%rd59+244], %r1197; $L__BB108_421: cvt.u32.u64 %r1198, %rd58; add.s32 %r3090, %r1198, 4; add.s32 %r3084, %r3084, -4; setp.ne.s32 %p316, %r3084, 0; @%p316 bra $L__BB108_397; $L__BB108_422: setp.eq.s32 %p317, %r177, 0; @%p317 bra $L__BB108_520; setp.ne.s32 %p322, %r174, 0; setp.eq.s32 %p325, %r175, 0; $L__BB108_424: .pragma "nounroll"; cvt.s64.s32 %rd60, %r3090; mul.wide.s32 %rd1050, %r3090, 80; add.s64 %rd1051, %rd1, %rd1050; add.s64 %rd61, %rd1051, 316; ld.local.u32 %r201, [%rd1051+316]; add.s32 %r1199, %r201, -2; setp.lt.u32 %p318, %r1199, 7; @%p318 bra $L__BB108_429; add.s32 %r1200, %r201, -9; setp.lt.u32 %p319, %r1200, 12; setp.eq.s32 %p320, %r201, 29; or.pred %p321, %p319, %p320; @%p321 bra $L__BB108_429; add.s32 %r1201, %r201, -21; setp.lt.u32 %p323, %r1201, 10; and.pred %p324, %p323, %p322; @%p324 bra $L__BB108_429; setp.ne.s32 %p326, %r201, 31; or.pred %p327, %p326, %p325; @%p327 bra $L__BB108_430; mov.u32 %r1202, 35; st.local.u32 [%rd61], %r1202; mov.u32 %r1203, 0; st.local.u32 [%rd61+4], %r1203; or.b32 %r3134, %r3134, 32; st.local.u32 [%rd7+-36], %r3134; bra.uni $L__BB108_430; $L__BB108_429: mov.u32 %r1204, 0; st.local.u32 [%rd61], %r1204; st.local.u32 [%rd61+4], %r1204; $L__BB108_430: cvt.u32.u64 %r1205, %rd60; add.s32 %r3090, %r1205, 1; add.s32 %r177, %r177, -1; setp.eq.s32 %p328, %r177, 0; @%p328 bra $L__BB108_520; bra.uni $L__BB108_424; $L__BB108_520: ld.const.v2.u64 {%rd1064, %rd1065}, [kernel_params+2464]; cvta.to.global.u64 %rd1068, %rd1064; shl.b64 %rd1069, %rd2, 1; add.s64 %rd74, %rd1068, %rd1069; ld.global.u16 %r1288, [%rd74]; cvta.to.global.u64 %rd1070, %rd1065; add.s64 %rd1071, %rd1070, %rd1069; ld.global.u16 %r1289, [%rd1071]; neg.s32 %r1290, %r1289; setp.ne.s32 %p452, %r1288, %r1290; ld.local.u32 %r291, [%rd7+172]; setp.lt.s32 %p453, %r291, 2; or.pred %p454, %p452, %p453; @%p454 bra $L__BB108_553; add.s32 %r292, %r291, -1; and.b32 %r3144, %r291, 3; setp.lt.u32 %p455, %r292, 3; mov.f32 %f15750, 0f00000000; mov.u32 %r3137, 0; @%p455 bra $L__BB108_532; sub.s32 %r3136, %r291, %r3144; mov.f32 %f15750, 0f00000000; mov.u32 %r3137, 0; $L__BB108_523: cvt.s64.s32 %rd75, %r3137; mul.wide.s32 %rd1072, %r3137, 80; add.s64 %rd1073, %rd1, %rd1072; add.s64 %rd76, %rd1073, 316; ld.local.u32 %r1293, [%rd1073+316]; setp.gt.s32 %p456, %r1293, 34; @%p456 bra $L__BB108_525; ld.local.f32 %f5594, [%rd76+4]; add.ftz.f32 %f15750, %f15750, %f5594; $L__BB108_525: ld.local.u32 %r1294, [%rd76+80]; setp.gt.s32 %p457, %r1294, 34; @%p457 bra $L__BB108_527; ld.local.f32 %f5595, [%rd76+84]; add.ftz.f32 %f15750, %f15750, %f5595; $L__BB108_527: ld.local.u32 %r1295, [%rd76+160]; setp.gt.s32 %p458, %r1295, 34; @%p458 bra $L__BB108_529; ld.local.f32 %f5596, [%rd76+164]; add.ftz.f32 %f15750, %f15750, %f5596; $L__BB108_529: ld.local.u32 %r1296, [%rd76+240]; setp.gt.s32 %p459, %r1296, 34; @%p459 bra $L__BB108_531; ld.local.f32 %f5597, [%rd76+244]; add.ftz.f32 %f15750, %f15750, %f5597; $L__BB108_531: cvt.u32.u64 %r1297, %rd75; add.s32 %r3137, %r1297, 4; add.s32 %r3136, %r3136, -4; setp.ne.s32 %p460, %r3136, 0; @%p460 bra $L__BB108_523; $L__BB108_532: setp.eq.s32 %p461, %r3144, 0; @%p461 bra $L__BB108_537; mov.u32 %r3139, %r3144; $L__BB108_534: .pragma "nounroll"; cvt.s64.s32 %rd77, %r3137; mul.wide.s32 %rd1074, %r3137, 80; add.s64 %rd1075, %rd1, %rd1074; add.s64 %rd78, %rd1075, 316; ld.local.u32 %r1298, [%rd1075+316]; setp.gt.s32 %p462, %r1298, 34; @%p462 bra $L__BB108_536; ld.local.f32 %f5598, [%rd78+4]; add.ftz.f32 %f15750, %f15750, %f5598; $L__BB108_536: cvt.u32.u64 %r1299, %rd77; add.s32 %r3137, %r1299, 1; add.s32 %r3139, %r3139, -1; setp.ne.s32 %p463, %r3139, 0; @%p463 bra $L__BB108_534; $L__BB108_537: setp.lt.s32 %p464, %r291, 1; @%p464 bra $L__BB108_553; mul.ftz.f32 %f665, %f15750, 0f3E000000; mov.u32 %r3142, 0; @%p455 bra $L__BB108_549; sub.s32 %r3141, %r291, %r3144; mov.u32 %r3142, 0; $L__BB108_540: cvt.s64.s32 %rd79, %r3142; mul.wide.s32 %rd1076, %r3142, 80; add.s64 %rd1077, %rd1, %rd1076; add.s64 %rd80, %rd1077, 316; ld.local.u32 %r1302, [%rd1077+316]; setp.gt.s32 %p466, %r1302, 34; @%p466 bra $L__BB108_542; ld.local.f32 %f5599, [%rd80+4]; max.ftz.f32 %f5600, %f5599, %f665; st.local.f32 [%rd80+4], %f5600; $L__BB108_542: ld.local.u32 %r1303, [%rd80+80]; setp.gt.s32 %p467, %r1303, 34; @%p467 bra $L__BB108_544; ld.local.f32 %f5601, [%rd80+84]; max.ftz.f32 %f5602, %f5601, %f665; st.local.f32 [%rd80+84], %f5602; $L__BB108_544: ld.local.u32 %r1304, [%rd80+160]; setp.gt.s32 %p468, %r1304, 34; @%p468 bra $L__BB108_546; ld.local.f32 %f5603, [%rd80+164]; max.ftz.f32 %f5604, %f5603, %f665; st.local.f32 [%rd80+164], %f5604; $L__BB108_546: ld.local.u32 %r1305, [%rd80+240]; setp.gt.s32 %p469, %r1305, 34; @%p469 bra $L__BB108_548; ld.local.f32 %f5605, [%rd80+244]; max.ftz.f32 %f5606, %f5605, %f665; st.local.f32 [%rd80+244], %f5606; $L__BB108_548: cvt.u32.u64 %r1306, %rd79; add.s32 %r3142, %r1306, 4; add.s32 %r3141, %r3141, -4; setp.ne.s32 %p470, %r3141, 0; @%p470 bra $L__BB108_540; $L__BB108_549: @%p461 bra $L__BB108_553; $L__BB108_550: .pragma "nounroll"; cvt.s64.s32 %rd81, %r3142; mul.wide.s32 %rd1078, %r3142, 80; add.s64 %rd1079, %rd1, %rd1078; add.s64 %rd82, %rd1079, 316; ld.local.u32 %r1307, [%rd1079+316]; setp.gt.s32 %p472, %r1307, 34; @%p472 bra $L__BB108_552; ld.local.f32 %f5607, [%rd82+4]; max.ftz.f32 %f5608, %f5607, %f665; st.local.f32 [%rd82+4], %f5608; $L__BB108_552: cvt.u32.u64 %r1308, %rd81; add.s32 %r3142, %r1308, 1; add.s32 %r3144, %r3144, -1; setp.ne.s32 %p473, %r3144, 0; @%p473 bra $L__BB108_550; $L__BB108_553: ld.const.f32 %f666, [kernel_params+1572]; setp.eq.ftz.f32 %p474, %f666, 0f7F7FFFFF; @%p474 bra $L__BB108_647; ld.const.u64 %rd1080, [kernel_params+2552]; cvta.to.global.u64 %rd1081, %rd1080; add.s64 %rd1082, %rd1081, %rd2; ld.global.u8 %rs66, [%rd1082]; and.b16 %rs67, %rs66, 1; setp.eq.b16 %p475, %rs67, 1; mov.pred %p476, 0; xor.pred %p477, %p475, %p476; @%p477 bra $L__BB108_647; ld.const.u64 %rd1083, [kernel_params+2568]; cvta.to.global.u64 %rd1084, %rd1083; add.s64 %rd1086, %rd1084, %rd837; ld.global.f32 %f5609, [%rd1086]; mul.ftz.f32 %f667, %f666, %f5609; setp.geu.ftz.f32 %p478, %f667, 0f3F800000; @%p478 bra $L__BB108_647; mov.f32 %f5610, 0f3F800000; sub.ftz.f32 %f5611, %f5610, %f667; sqrt.approx.ftz.f32 %f5612, %f5611; mul.ftz.f32 %f668, %f5612, 0f3F000000; setp.lt.s32 %p479, %r291, 1; @%p479 bra $L__BB108_647; max.s32 %r314, %r291, 1; add.s32 %r1310, %r314, -1; and.b32 %r3149, %r314, 3; setp.lt.u32 %p480, %r1310, 3; mov.u32 %r3147, 0; @%p480 bra $L__BB108_628; sub.s32 %r3146, %r314, %r3149; mov.u32 %r3147, 0; $L__BB108_559: cvt.s64.s32 %rd83, %r3147; mul.wide.s32 %rd1087, %r3147, 80; add.s64 %rd1088, %rd1, %rd1087; add.s64 %rd84, %rd1088, 316; ld.local.u32 %r319, [%rd1088+316]; setp.gt.s32 %p481, %r319, 31; @%p481 bra $L__BB108_576; setp.gt.s32 %p482, %r319, 15; @%p482 bra $L__BB108_565; add.s32 %r1312, %r319, -10; setp.lt.u32 %p490, %r1312, 3; @%p490 bra $L__BB108_569; add.s32 %r1313, %r319, -14; setp.lt.u32 %p491, %r1313, 2; @%p491 bra $L__BB108_575; setp.eq.s32 %p492, %r319, 13; @%p492 bra $L__BB108_564; bra.uni $L__BB108_576; $L__BB108_565: setp.gt.s32 %p483, %r319, 23; @%p483 bra $L__BB108_570; setp.eq.s32 %p487, %r319, 16; @%p487 bra $L__BB108_574; setp.eq.s32 %p488, %r319, 22; @%p488 bra $L__BB108_564; setp.eq.s32 %p489, %r319, 23; @%p489 bra $L__BB108_569; bra.uni $L__BB108_576; $L__BB108_569: ld.local.v2.f32 {%f5637, %f5638}, [%rd84+20]; max.ftz.f32 %f5641, %f668, %f5638; max.ftz.f32 %f5642, %f668, %f5637; st.local.v2.f32 [%rd84+20], {%f5642, %f5641}; bra.uni $L__BB108_576; $L__BB108_570: setp.eq.s32 %p484, %r319, 24; @%p484 bra $L__BB108_575; setp.eq.s32 %p485, %r319, 27; @%p485 bra $L__BB108_575; setp.ne.s32 %p486, %r319, 29; @%p486 bra $L__BB108_576; ld.local.f32 %f5613, [%rd84+32]; max.ftz.f32 %f5614, %f668, %f5613; st.local.f32 [%rd84+32], %f5614; ld.local.f32 %f5615, [%rd84+36]; max.ftz.f32 %f5616, %f668, %f5615; st.local.f32 [%rd84+36], %f5616; ld.local.f32 %f5617, [%rd84+48]; max.ftz.f32 %f5618, %f668, %f5617; st.local.f32 [%rd84+48], %f5618; bra.uni $L__BB108_576; $L__BB108_575: ld.local.v2.f32 {%f5625, %f5626}, [%rd84+20]; max.ftz.f32 %f5629, %f668, %f5626; max.ftz.f32 %f5630, %f668, %f5625; st.local.v2.f32 [%rd84+20], {%f5630, %f5629}; bra.uni $L__BB108_576; $L__BB108_564: ld.local.v2.f32 {%f5631, %f5632}, [%rd84+20]; max.ftz.f32 %f5635, %f668, %f5632; max.ftz.f32 %f5636, %f668, %f5631; st.local.v2.f32 [%rd84+20], {%f5636, %f5635}; bra.uni $L__BB108_576; $L__BB108_574: ld.local.v2.f32 {%f5619, %f5620}, [%rd84+20]; max.ftz.f32 %f5623, %f668, %f5620; max.ftz.f32 %f5624, %f668, %f5619; st.local.v2.f32 [%rd84+20], {%f5624, %f5623}; $L__BB108_576: ld.local.u32 %r320, [%rd84+80]; setp.gt.s32 %p493, %r320, 31; @%p493 bra $L__BB108_593; setp.gt.s32 %p494, %r320, 15; @%p494 bra $L__BB108_582; add.s32 %r1314, %r320, -10; setp.lt.u32 %p502, %r1314, 3; @%p502 bra $L__BB108_586; add.s32 %r1315, %r320, -14; setp.lt.u32 %p503, %r1315, 2; @%p503 bra $L__BB108_592; setp.eq.s32 %p504, %r320, 13; @%p504 bra $L__BB108_581; bra.uni $L__BB108_593; $L__BB108_582: setp.gt.s32 %p495, %r320, 23; @%p495 bra $L__BB108_587; setp.eq.s32 %p499, %r320, 16; @%p499 bra $L__BB108_591; setp.eq.s32 %p500, %r320, 22; @%p500 bra $L__BB108_581; setp.eq.s32 %p501, %r320, 23; @%p501 bra $L__BB108_586; bra.uni $L__BB108_593; $L__BB108_586: ld.local.v2.f32 {%f5667, %f5668}, [%rd84+100]; max.ftz.f32 %f5671, %f668, %f5668; max.ftz.f32 %f5672, %f668, %f5667; st.local.v2.f32 [%rd84+100], {%f5672, %f5671}; bra.uni $L__BB108_593; $L__BB108_587: setp.eq.s32 %p496, %r320, 24; @%p496 bra $L__BB108_592; setp.eq.s32 %p497, %r320, 27; @%p497 bra $L__BB108_592; setp.ne.s32 %p498, %r320, 29; @%p498 bra $L__BB108_593; ld.local.f32 %f5643, [%rd84+112]; max.ftz.f32 %f5644, %f668, %f5643; st.local.f32 [%rd84+112], %f5644; ld.local.f32 %f5645, [%rd84+116]; max.ftz.f32 %f5646, %f668, %f5645; st.local.f32 [%rd84+116], %f5646; ld.local.f32 %f5647, [%rd84+128]; max.ftz.f32 %f5648, %f668, %f5647; st.local.f32 [%rd84+128], %f5648; bra.uni $L__BB108_593; $L__BB108_592: ld.local.v2.f32 {%f5655, %f5656}, [%rd84+100]; max.ftz.f32 %f5659, %f668, %f5656; max.ftz.f32 %f5660, %f668, %f5655; st.local.v2.f32 [%rd84+100], {%f5660, %f5659}; bra.uni $L__BB108_593; $L__BB108_581: ld.local.v2.f32 {%f5661, %f5662}, [%rd84+100]; max.ftz.f32 %f5665, %f668, %f5662; max.ftz.f32 %f5666, %f668, %f5661; st.local.v2.f32 [%rd84+100], {%f5666, %f5665}; bra.uni $L__BB108_593; $L__BB108_591: ld.local.v2.f32 {%f5649, %f5650}, [%rd84+100]; max.ftz.f32 %f5653, %f668, %f5650; max.ftz.f32 %f5654, %f668, %f5649; st.local.v2.f32 [%rd84+100], {%f5654, %f5653}; $L__BB108_593: ld.local.u32 %r321, [%rd84+160]; setp.gt.s32 %p505, %r321, 31; @%p505 bra $L__BB108_610; setp.gt.s32 %p506, %r321, 15; @%p506 bra $L__BB108_599; add.s32 %r1316, %r321, -10; setp.lt.u32 %p514, %r1316, 3; @%p514 bra $L__BB108_603; add.s32 %r1317, %r321, -14; setp.lt.u32 %p515, %r1317, 2; @%p515 bra $L__BB108_609; setp.eq.s32 %p516, %r321, 13; @%p516 bra $L__BB108_598; bra.uni $L__BB108_610; $L__BB108_599: setp.gt.s32 %p507, %r321, 23; @%p507 bra $L__BB108_604; setp.eq.s32 %p511, %r321, 16; @%p511 bra $L__BB108_608; setp.eq.s32 %p512, %r321, 22; @%p512 bra $L__BB108_598; setp.eq.s32 %p513, %r321, 23; @%p513 bra $L__BB108_603; bra.uni $L__BB108_610; $L__BB108_603: ld.local.v2.f32 {%f5697, %f5698}, [%rd84+180]; max.ftz.f32 %f5701, %f668, %f5698; max.ftz.f32 %f5702, %f668, %f5697; st.local.v2.f32 [%rd84+180], {%f5702, %f5701}; bra.uni $L__BB108_610; $L__BB108_604: setp.eq.s32 %p508, %r321, 24; @%p508 bra $L__BB108_609; setp.eq.s32 %p509, %r321, 27; @%p509 bra $L__BB108_609; setp.ne.s32 %p510, %r321, 29; @%p510 bra $L__BB108_610; ld.local.f32 %f5673, [%rd84+192]; max.ftz.f32 %f5674, %f668, %f5673; st.local.f32 [%rd84+192], %f5674; ld.local.f32 %f5675, [%rd84+196]; max.ftz.f32 %f5676, %f668, %f5675; st.local.f32 [%rd84+196], %f5676; ld.local.f32 %f5677, [%rd84+208]; max.ftz.f32 %f5678, %f668, %f5677; st.local.f32 [%rd84+208], %f5678; bra.uni $L__BB108_610; $L__BB108_609: ld.local.v2.f32 {%f5685, %f5686}, [%rd84+180]; max.ftz.f32 %f5689, %f668, %f5686; max.ftz.f32 %f5690, %f668, %f5685; st.local.v2.f32 [%rd84+180], {%f5690, %f5689}; bra.uni $L__BB108_610; $L__BB108_598: ld.local.v2.f32 {%f5691, %f5692}, [%rd84+180]; max.ftz.f32 %f5695, %f668, %f5692; max.ftz.f32 %f5696, %f668, %f5691; st.local.v2.f32 [%rd84+180], {%f5696, %f5695}; bra.uni $L__BB108_610; $L__BB108_608: ld.local.v2.f32 {%f5679, %f5680}, [%rd84+180]; max.ftz.f32 %f5683, %f668, %f5680; max.ftz.f32 %f5684, %f668, %f5679; st.local.v2.f32 [%rd84+180], {%f5684, %f5683}; $L__BB108_610: ld.local.u32 %r322, [%rd84+240]; setp.gt.s32 %p517, %r322, 31; @%p517 bra $L__BB108_627; setp.gt.s32 %p518, %r322, 15; @%p518 bra $L__BB108_616; add.s32 %r1318, %r322, -10; setp.lt.u32 %p526, %r1318, 3; @%p526 bra $L__BB108_620; add.s32 %r1319, %r322, -14; setp.lt.u32 %p527, %r1319, 2; @%p527 bra $L__BB108_626; setp.eq.s32 %p528, %r322, 13; @%p528 bra $L__BB108_615; bra.uni $L__BB108_627; $L__BB108_616: setp.gt.s32 %p519, %r322, 23; @%p519 bra $L__BB108_621; setp.eq.s32 %p523, %r322, 16; @%p523 bra $L__BB108_625; setp.eq.s32 %p524, %r322, 22; @%p524 bra $L__BB108_615; setp.eq.s32 %p525, %r322, 23; @%p525 bra $L__BB108_620; bra.uni $L__BB108_627; $L__BB108_620: ld.local.v2.f32 {%f5727, %f5728}, [%rd84+260]; max.ftz.f32 %f5731, %f668, %f5728; max.ftz.f32 %f5732, %f668, %f5727; st.local.v2.f32 [%rd84+260], {%f5732, %f5731}; bra.uni $L__BB108_627; $L__BB108_621: setp.eq.s32 %p520, %r322, 24; @%p520 bra $L__BB108_626; setp.eq.s32 %p521, %r322, 27; @%p521 bra $L__BB108_626; setp.ne.s32 %p522, %r322, 29; @%p522 bra $L__BB108_627; ld.local.f32 %f5703, [%rd84+272]; max.ftz.f32 %f5704, %f668, %f5703; st.local.f32 [%rd84+272], %f5704; ld.local.f32 %f5705, [%rd84+276]; max.ftz.f32 %f5706, %f668, %f5705; st.local.f32 [%rd84+276], %f5706; ld.local.f32 %f5707, [%rd84+288]; max.ftz.f32 %f5708, %f668, %f5707; st.local.f32 [%rd84+288], %f5708; bra.uni $L__BB108_627; $L__BB108_626: ld.local.v2.f32 {%f5715, %f5716}, [%rd84+260]; max.ftz.f32 %f5719, %f668, %f5716; max.ftz.f32 %f5720, %f668, %f5715; st.local.v2.f32 [%rd84+260], {%f5720, %f5719}; bra.uni $L__BB108_627; $L__BB108_615: ld.local.v2.f32 {%f5721, %f5722}, [%rd84+260]; max.ftz.f32 %f5725, %f668, %f5722; max.ftz.f32 %f5726, %f668, %f5721; st.local.v2.f32 [%rd84+260], {%f5726, %f5725}; bra.uni $L__BB108_627; $L__BB108_625: ld.local.v2.f32 {%f5709, %f5710}, [%rd84+260]; max.ftz.f32 %f5713, %f668, %f5710; max.ftz.f32 %f5714, %f668, %f5709; st.local.v2.f32 [%rd84+260], {%f5714, %f5713}; $L__BB108_627: cvt.u32.u64 %r1320, %rd83; add.s32 %r3147, %r1320, 4; add.s32 %r3146, %r3146, -4; setp.ne.s32 %p529, %r3146, 0; @%p529 bra $L__BB108_559; $L__BB108_628: setp.eq.s32 %p530, %r3149, 0; @%p530 bra $L__BB108_647; $L__BB108_629: .pragma "nounroll"; cvt.s64.s32 %rd85, %r3147; mul.wide.s32 %rd1089, %r3147, 80; add.s64 %rd1090, %rd1, %rd1089; add.s64 %rd86, %rd1090, 316; ld.local.u32 %r328, [%rd1090+316]; setp.gt.s32 %p531, %r328, 31; @%p531 bra $L__BB108_646; setp.gt.s32 %p532, %r328, 15; @%p532 bra $L__BB108_635; add.s32 %r1321, %r328, -10; setp.lt.u32 %p540, %r1321, 3; @%p540 bra $L__BB108_639; add.s32 %r1322, %r328, -14; setp.lt.u32 %p541, %r1322, 2; @%p541 bra $L__BB108_645; setp.eq.s32 %p542, %r328, 13; @%p542 bra $L__BB108_634; bra.uni $L__BB108_646; $L__BB108_635: setp.gt.s32 %p533, %r328, 23; @%p533 bra $L__BB108_640; setp.eq.s32 %p537, %r328, 16; @%p537 bra $L__BB108_644; setp.eq.s32 %p538, %r328, 22; @%p538 bra $L__BB108_634; setp.eq.s32 %p539, %r328, 23; @%p539 bra $L__BB108_639; bra.uni $L__BB108_646; $L__BB108_639: ld.local.v2.f32 {%f5757, %f5758}, [%rd86+20]; max.ftz.f32 %f5761, %f668, %f5758; max.ftz.f32 %f5762, %f668, %f5757; st.local.v2.f32 [%rd86+20], {%f5762, %f5761}; bra.uni $L__BB108_646; $L__BB108_640: setp.eq.s32 %p534, %r328, 24; @%p534 bra $L__BB108_645; setp.eq.s32 %p535, %r328, 27; @%p535 bra $L__BB108_645; setp.ne.s32 %p536, %r328, 29; @%p536 bra $L__BB108_646; ld.local.f32 %f5733, [%rd86+32]; max.ftz.f32 %f5734, %f668, %f5733; st.local.f32 [%rd86+32], %f5734; ld.local.f32 %f5735, [%rd86+36]; max.ftz.f32 %f5736, %f668, %f5735; st.local.f32 [%rd86+36], %f5736; ld.local.f32 %f5737, [%rd86+48]; max.ftz.f32 %f5738, %f668, %f5737; st.local.f32 [%rd86+48], %f5738; bra.uni $L__BB108_646; $L__BB108_645: ld.local.v2.f32 {%f5745, %f5746}, [%rd86+20]; max.ftz.f32 %f5749, %f668, %f5746; max.ftz.f32 %f5750, %f668, %f5745; st.local.v2.f32 [%rd86+20], {%f5750, %f5749}; bra.uni $L__BB108_646; $L__BB108_634: ld.local.v2.f32 {%f5751, %f5752}, [%rd86+20]; max.ftz.f32 %f5755, %f668, %f5752; max.ftz.f32 %f5756, %f668, %f5751; st.local.v2.f32 [%rd86+20], {%f5756, %f5755}; bra.uni $L__BB108_646; $L__BB108_644: ld.local.v2.f32 {%f5739, %f5740}, [%rd86+20]; max.ftz.f32 %f5743, %f668, %f5740; max.ftz.f32 %f5744, %f668, %f5739; st.local.v2.f32 [%rd86+20], {%f5744, %f5743}; $L__BB108_646: cvt.u32.u64 %r1323, %rd85; add.s32 %r3147, %r1323, 1; add.s32 %r3149, %r3149, -1; setp.ne.s32 %p543, %r3149, 0; @%p543 bra $L__BB108_629; $L__BB108_647: ld.global.u32 %r331, [%rd29]; and.b32 %r1324, %r3134, 32; setp.eq.s32 %p544, %r1324, 0; and.b32 %r332, %r331, 16384; @%p544 bra $L__BB108_649; bra.uni $L__BB108_648; $L__BB108_649: ld.local.u32 %r1325, [%rd7+-32]; and.b32 %r1326, %r1325, 1; setp.eq.b32 %p546, %r1326, 1; not.pred %p547, %p546; setp.eq.s32 %p548, %r332, 0; or.pred %p549, %p548, %p547; @%p549 bra $L__BB108_690; bra.uni $L__BB108_650; $L__BB108_648: setp.eq.s32 %p545, %r332, 0; @%p545 bra $L__BB108_690; $L__BB108_650: ld.local.u32 %r1327, [%rd7+-32]; and.b32 %r1328, %r1327, 1; setp.eq.b32 %p550, %r1328, 1; mov.pred %p551, 0; xor.pred %p552, %p550, %p551; not.pred %p553, %p552; @%p553 bra $L__BB108_669; bra.uni $L__BB108_651; $L__BB108_669: ld.local.u32 %r346, [%rd7+172]; setp.lt.s32 %p565, %r346, 1; mov.f32 %f15762, 0f00000000; mov.f32 %f15763, %f15762; mov.f32 %f15764, %f15762; @%p565 bra $L__BB108_685; add.s32 %r1344, %r346, -1; and.b32 %r3159, %r346, 3; setp.lt.u32 %p566, %r1344, 3; mov.f32 %f15764, 0f00000000; mov.u32 %r3157, 0; mov.f32 %f15763, %f15764; mov.f32 %f15762, %f15764; @%p566 bra $L__BB108_681; sub.s32 %r3156, %r346, %r3159; mov.f32 %f15764, 0f00000000; mov.u32 %r3157, 0; mov.f32 %f15763, %f15764; mov.f32 %f15762, %f15764; $L__BB108_672: cvt.s64.s32 %rd91, %r3157; mul.wide.s32 %rd1093, %r3157, 80; add.s64 %rd1094, %rd1, %rd1093; add.s64 %rd92, %rd1094, 316; ld.local.u32 %r1346, [%rd1094+316]; setp.ne.s32 %p567, %r1346, 35; @%p567 bra $L__BB108_674; ld.local.v4.f32 {%f5779, %f5780, %f5781, %f5782}, [%rd92+-12]; add.ftz.f32 %f15762, %f15762, %f5779; add.ftz.f32 %f15763, %f15763, %f5780; add.ftz.f32 %f15764, %f15764, %f5781; $L__BB108_674: ld.local.u32 %r1347, [%rd92+80]; setp.ne.s32 %p568, %r1347, 35; @%p568 bra $L__BB108_676; ld.local.v4.f32 {%f5786, %f5787, %f5788, %f5789}, [%rd92+68]; add.ftz.f32 %f15762, %f15762, %f5786; add.ftz.f32 %f15763, %f15763, %f5787; add.ftz.f32 %f15764, %f15764, %f5788; $L__BB108_676: ld.local.u32 %r1348, [%rd92+160]; setp.ne.s32 %p569, %r1348, 35; @%p569 bra $L__BB108_678; ld.local.v4.f32 {%f5793, %f5794, %f5795, %f5796}, [%rd92+148]; add.ftz.f32 %f15762, %f15762, %f5793; add.ftz.f32 %f15763, %f15763, %f5794; add.ftz.f32 %f15764, %f15764, %f5795; $L__BB108_678: ld.local.u32 %r1349, [%rd92+240]; setp.ne.s32 %p570, %r1349, 35; @%p570 bra $L__BB108_680; ld.local.v4.f32 {%f5800, %f5801, %f5802, %f5803}, [%rd92+228]; add.ftz.f32 %f15762, %f15762, %f5800; add.ftz.f32 %f15763, %f15763, %f5801; add.ftz.f32 %f15764, %f15764, %f5802; $L__BB108_680: cvt.u32.u64 %r1350, %rd91; add.s32 %r3157, %r1350, 4; add.s32 %r3156, %r3156, -4; setp.ne.s32 %p571, %r3156, 0; @%p571 bra $L__BB108_672; $L__BB108_681: setp.eq.s32 %p572, %r3159, 0; @%p572 bra $L__BB108_685; $L__BB108_682: .pragma "nounroll"; cvt.s64.s32 %rd93, %r3157; mul.wide.s32 %rd1095, %r3157, 80; add.s64 %rd1096, %rd1, %rd1095; add.s64 %rd94, %rd1096, 316; ld.local.u32 %r1351, [%rd1096+316]; setp.ne.s32 %p573, %r1351, 35; @%p573 bra $L__BB108_684; ld.local.f32 %f5807, [%rd94+-12]; add.ftz.f32 %f15762, %f15762, %f5807; ld.local.f32 %f5808, [%rd94+-8]; add.ftz.f32 %f15763, %f15763, %f5808; ld.local.f32 %f5809, [%rd94+-4]; add.ftz.f32 %f15764, %f15764, %f5809; $L__BB108_684: cvt.u32.u64 %r1352, %rd93; add.s32 %r3157, %r1352, 1; add.s32 %r3159, %r3159, -1; setp.ne.s32 %p574, %r3159, 0; @%p574 bra $L__BB108_682; bra.uni $L__BB108_685; $L__BB108_651: and.b32 %r1329, %r3134, 524800; setp.ne.s32 %p554, %r1329, 512; mov.f32 %f15762, 0f3F800000; mov.f32 %f15763, %f15762; mov.f32 %f15764, %f15762; @%p554 bra $L__BB108_685; ld.local.f32 %f669, [%rd7+204]; ld.local.v2.f32 {%f5766, %f5767}, [%rd7+208]; ld.local.u32 %r333, [%rd7+172]; setp.lt.s32 %p555, %r333, 1; @%p555 bra $L__BB108_668; add.s32 %r1331, %r333, -1; and.b32 %r3154, %r333, 3; setp.lt.u32 %p556, %r1331, 3; mov.u32 %r3152, 0; @%p556 bra $L__BB108_664; sub.s32 %r3151, %r333, %r3154; add.s64 %rd2935, %rd1, 316; mov.u32 %r3152, 0; $L__BB108_655: ld.local.u32 %r1333, [%rd2935]; setp.eq.s32 %p557, %r1333, 31; @%p557 bra $L__BB108_657; mov.u32 %r1334, 40; st.local.u32 [%rd2935], %r1334; $L__BB108_657: ld.local.u32 %r1335, [%rd2935+80]; setp.eq.s32 %p558, %r1335, 31; @%p558 bra $L__BB108_659; mov.u32 %r1336, 40; st.local.u32 [%rd2935+80], %r1336; $L__BB108_659: ld.local.u32 %r1337, [%rd2935+160]; setp.eq.s32 %p559, %r1337, 31; @%p559 bra $L__BB108_661; mov.u32 %r1338, 40; st.local.u32 [%rd2935+160], %r1338; $L__BB108_661: ld.local.u32 %r1339, [%rd2935+240]; setp.eq.s32 %p560, %r1339, 31; @%p560 bra $L__BB108_663; mov.u32 %r1340, 40; st.local.u32 [%rd2935+240], %r1340; $L__BB108_663: add.s32 %r3152, %r3152, 4; add.s32 %r3151, %r3151, -4; setp.ne.s32 %p561, %r3151, 0; add.s64 %rd2935, %rd2935, 320; @%p561 bra $L__BB108_655; $L__BB108_664: setp.eq.s32 %p562, %r3154, 0; @%p562 bra $L__BB108_668; $L__BB108_665: .pragma "nounroll"; mul.wide.s32 %rd1091, %r3152, 80; add.s64 %rd1092, %rd1, %rd1091; add.s64 %rd90, %rd1092, 316; ld.local.u32 %r1341, [%rd1092+316]; setp.eq.s32 %p563, %r1341, 31; @%p563 bra $L__BB108_667; mov.u32 %r1342, 40; st.local.u32 [%rd90], %r1342; $L__BB108_667: add.s32 %r3152, %r3152, 1; add.s32 %r3154, %r3154, -1; setp.ne.s32 %p564, %r3154, 0; @%p564 bra $L__BB108_665; $L__BB108_668: and.b32 %r3134, %r3134, -763; st.local.u32 [%rd7+-36], %r3134; mov.f32 %f5768, 0f3F800000; sub.ftz.f32 %f15764, %f5768, %f5767; sub.ftz.f32 %f15763, %f5768, %f5766; sub.ftz.f32 %f15762, %f5768, %f669; $L__BB108_685: ld.const.u64 %rd1097, [kernel_params+2584]; cvta.to.global.u64 %rd1098, %rd1097; mul.lo.s64 %rd1099, %rd2, 12; add.s64 %rd1100, %rd1098, %rd1099; ld.global.f32 %f5810, [%rd1100]; ld.global.f32 %f5811, [%rd1100+4]; mul.ftz.f32 %f5812, %f15763, %f5811; ld.global.f32 %f5813, [%rd1100+8]; fma.rn.ftz.f32 %f5814, %f15762, %f5810, %f5812; fma.rn.ftz.f32 %f5815, %f15764, %f5813, %f5814; mul.ftz.f32 %f720, %f5815, 0f3EAAAAAB; ld.const.u64 %rd1101, [kernel_params+2448]; cvta.to.global.u64 %rd1102, %rd1101; add.s64 %rd1104, %rd1102, %rd837; ld.global.u32 %rd1105, [%rd1104]; ld.const.s32 %rd1106, [kernel_params+1264]; mul.lo.s64 %rd95, %rd1106, %rd1105; ld.const.u8 %rs68, [kernel_params+1260]; and.b16 %rs69, %rs68, 2; setp.eq.s16 %p575, %rs69, 0; @%p575 bra $L__BB108_687; ld.const.s32 %rd1107, [kernel_params+1268]; cvta.to.global.u64 %rd1108, %rd618; add.s64 %rd1109, %rd1107, %rd95; shl.b64 %rd1110, %rd1109, 2; add.s64 %rd1111, %rd1108, %rd1110; add.s64 %rd1112, %rd1111, 12; atom.global.add.f32 %f5816, [%rd1112], %f720; $L__BB108_687: ld.const.u32 %r1353, [kernel_params+1620]; setp.eq.s32 %p576, %r1353, 0; shr.u32 %r1354, %r331, 29; and.b32 %r1355, %r1354, 1; setp.eq.b32 %p577, %r1355, 1; or.pred %p578, %p577, %p576; @%p578 bra $L__BB108_689; ld.const.s32 %rd1113, [kernel_params+1384]; cvta.to.global.u64 %rd1114, %rd618; add.s64 %rd1115, %rd1113, %rd95; shl.b64 %rd1116, %rd1115, 2; add.s64 %rd1117, %rd1114, %rd1116; add.s64 %rd1118, %rd1117, 12; atom.global.add.f32 %f5817, [%rd1118], %f720; $L__BB108_689: setp.eq.ftz.f32 %p579, %f15762, 0f3F800000; setp.eq.ftz.f32 %p580, %f15763, 0f3F800000; and.pred %p581, %p579, %p580; setp.eq.ftz.f32 %p582, %f15764, 0f3F800000; and.pred %p583, %p581, %p582; mov.u16 %rs161, 0; @%p583 bra $L__BB108_2421; $L__BB108_690: and.b32 %r1356, %r3134, 2; setp.eq.s32 %p584, %r1356, 0; @%p584 bra $L__BB108_794; ld.global.u32 %r360, [%rd29]; ld.local.v2.f32 {%f5819, %f5820}, [%rd7+-64]; ld.local.v4.f32 {%f5821, %f5822, %f5823, %f5824}, [%rd7+-56]; mul.ftz.f32 %f5825, %f5820, %f5823; fma.rn.ftz.f32 %f5826, %f5819, %f5822, %f5825; fma.rn.ftz.f32 %f727, %f5821, %f5824, %f5826; ld.local.v2.f32 {%f5827, %f5828}, [%rd7+192]; ld.local.f32 %f730, [%rd7+200]; shr.u32 %r1357, %r360, 11; and.b32 %r1358, %r1357, 1; setp.eq.b32 %p585, %r1358, 1; and.b32 %r1359, %r3134, 65536; setp.eq.s32 %p586, %r1359, 0; or.pred %p587, %p585, %p586; mov.f32 %f15885, 0f3F800000; @%p587 bra $L__BB108_750; ld.local.u32 %r1360, [%rd7+-24]; and.b32 %r1361, %r1360, 1; setp.eq.b32 %p588, %r1361, 1; mov.pred %p589, 0; xor.pred %p590, %p588, %p589; not.pred %p591, %p590; @%p591 bra $L__BB108_750; ld.const.u64 %rd2914, [kernel_params+2184]; cvta.to.global.u64 %rd2913, %rd2914; ld.const.u64 %rd1119, [kernel_params+2560]; cvta.to.global.u64 %rd1120, %rd1119; add.s64 %rd1122, %rd1120, %rd837; ld.global.f32 %f731, [%rd1122]; ld.local.f32 %f732, [%rd7]; ld.local.s32 %rd96, [%rd7+-28]; ld.local.u32 %r1362, [%rd7+-12]; cvt.s64.s32 %rd97, %r1362; mul.wide.s32 %rd1123, %r1362, 4; add.s64 %rd1124, %rd2913, %rd1123; ld.global.u32 %r361, [%rd1124]; shr.u32 %r1363, %r361, 6; and.b32 %r1364, %r1363, 1; setp.eq.b32 %p592, %r1364, 1; ld.local.f32 %f733, [%rd7+-4]; setp.ge.ftz.f32 %p593, %f733, 0f00000000; and.pred %p594, %p593, %p592; @%p594 bra $L__BB108_695; bra.uni $L__BB108_694; $L__BB108_695: ld.const.u64 %rd1136, [kernel_params+2160]; cvta.to.global.u64 %rd1137, %rd1136; mul.lo.s64 %rd1138, %rd97, 208; add.s64 %rd1139, %rd1137, %rd1138; ld.global.u32 %r362, [%rd1139+156]; ld.global.u32 %r363, [%rd1139+152]; shl.b32 %r1368, %r363, 1; cvt.rn.f32.s32 %f5830, %r1368; mul.ftz.f32 %f743, %f733, %f5830; cvt.rzi.ftz.s32.f32 %r1369, %f743; add.s32 %r1370, %r1368, -1; min.s32 %r364, %r1369, %r1370; ld.const.u64 %rd1140, [kernel_params+2312]; cvta.to.global.u64 %rd99, %rd1140; ld.global.u32 %r3163, [%rd1139+164]; mul.wide.u32 %rd1141, %r3163, 16; add.s64 %rd1142, %rd99, %rd1141; ld.global.u16 %rs137, [%rd1142+12]; ld.global.u32 %r3162, [%rd1142+8]; ld.global.u64 %rd2936, [%rd1142]; setp.eq.s64 %p595, %rd2936, 11; @%p595 bra $L__BB108_700; $L__BB108_696: setp.eq.s64 %p596, %rd2936, 0; @%p596 bra $L__BB108_698; add.s32 %r3163, %r3163, 2; bra.uni $L__BB108_699; $L__BB108_698: setp.eq.s16 %p597, %rs137, 0; mov.u32 %r3166, -1; mov.u32 %r3163, %r3162; @%p597 bra $L__BB108_701; $L__BB108_699: mul.wide.u32 %rd1143, %r3163, 16; add.s64 %rd1144, %rd99, %rd1143; ld.global.u16 %rs137, [%rd1144+12]; ld.global.u32 %r3162, [%rd1144+8]; ld.global.u64 %rd2936, [%rd1144]; setp.ne.s64 %p598, %rd2936, 11; @%p598 bra $L__BB108_696; $L__BB108_700: setp.eq.s16 %p599, %rs137, 0; selp.b32 %r3166, -1, %r3162, %p599; $L__BB108_701: ld.const.u64 %rd2955, [kernel_params+2232]; cvta.to.global.u64 %rd1145, %rd2955; shl.b64 %rd1146, %rd96, 4; add.s64 %rd1147, %rd1145, %rd1146; ld.global.v4.u32 {%r1372, %r1373, %r1374, %r1375}, [%rd1147]; setp.eq.s32 %p600, %r364, %r363; @%p600 bra $L__BB108_703; bra.uni $L__BB108_702; $L__BB108_703: ld.const.u64 %rd1153, [kernel_params+2256]; cvta.to.global.u64 %rd1154, %rd1153; mul.wide.u32 %rd1155, %r1375, 12; add.s64 %rd2937, %rd1154, %rd1155; add.s32 %r1382, %r1375, 1; mul.wide.u32 %rd1156, %r1382, 12; add.s64 %rd2940, %rd1154, %rd1156; add.s32 %r1383, %r1375, 2; mul.wide.u32 %rd1157, %r1383, 12; add.s64 %rd2943, %rd1154, %rd1157; add.s64 %rd2938, %rd2937, 4; add.s64 %rd2939, %rd2937, 8; add.s64 %rd2941, %rd2940, 4; add.s64 %rd2942, %rd2940, 8; add.s64 %rd2944, %rd2943, 4; add.s64 %rd2945, %rd2943, 8; bra.uni $L__BB108_704; $L__BB108_694: ld.const.u64 %rd2955, [kernel_params+2232]; cvta.to.global.u64 %rd1125, %rd2955; shl.b64 %rd1126, %rd96, 4; add.s64 %rd1127, %rd1125, %rd1126; ld.const.u64 %rd1128, [kernel_params+2256]; cvta.to.global.u64 %rd1129, %rd1128; ld.global.u32 %r1365, [%rd1127+12]; mul.wide.u32 %rd1130, %r1365, 12; add.s64 %rd1131, %rd1129, %rd1130; ld.global.f32 %f15814, [%rd1131]; ld.global.f32 %f15815, [%rd1131+4]; ld.global.f32 %f15816, [%rd1131+8]; add.s32 %r1366, %r1365, 1; mul.wide.u32 %rd1132, %r1366, 12; add.s64 %rd1133, %rd1129, %rd1132; ld.global.f32 %f15817, [%rd1133]; ld.global.f32 %f15818, [%rd1133+4]; ld.global.f32 %f15819, [%rd1133+8]; add.s32 %r1367, %r1365, 2; mul.wide.u32 %rd1134, %r1367, 12; add.s64 %rd1135, %rd1129, %rd1134; ld.global.f32 %f15820, [%rd1135]; ld.global.f32 %f15821, [%rd1135+4]; ld.global.f32 %f15822, [%rd1135+8]; mov.u16 %rs140, 0; bra.uni $L__BB108_708; $L__BB108_702: setp.gt.s32 %p601, %r364, %r363; selp.b32 %r1376, -1, 0, %p601; add.s32 %r1377, %r364, %r1376; mad.lo.s32 %r1378, %r1377, %r362, %r3166; ld.const.u64 %rd1148, [kernel_params+2336]; cvta.to.global.u64 %rd1149, %rd1148; add.s32 %r1379, %r1372, %r1378; mul.wide.u32 %rd1150, %r1379, 12; add.s64 %rd2937, %rd1149, %rd1150; add.s32 %r1380, %r1373, %r1378; mul.wide.u32 %rd1151, %r1380, 12; add.s64 %rd2940, %rd1149, %rd1151; add.s32 %r1381, %r1374, %r1378; mul.wide.u32 %rd1152, %r1381, 12; add.s64 %rd2943, %rd1149, %rd1152; add.s64 %rd2938, %rd2937, 4; add.s64 %rd2939, %rd2937, 8; add.s64 %rd2941, %rd2940, 4; add.s64 %rd2942, %rd2940, 8; add.s64 %rd2944, %rd2943, 4; add.s64 %rd2945, %rd2943, 8; $L__BB108_704: ld.global.f32 %f744, [%rd2945]; ld.global.f32 %f745, [%rd2944]; ld.global.f32 %f746, [%rd2943]; ld.global.f32 %f747, [%rd2942]; ld.global.f32 %f748, [%rd2941]; ld.global.f32 %f749, [%rd2940]; ld.global.f32 %f750, [%rd2939]; ld.global.f32 %f751, [%rd2938]; ld.global.f32 %f752, [%rd2937]; add.s32 %r379, %r364, 1; setp.eq.s32 %p602, %r379, %r363; @%p602 bra $L__BB108_706; bra.uni $L__BB108_705; $L__BB108_706: ld.const.u64 %rd1163, [kernel_params+2256]; cvta.to.global.u64 %rd1164, %rd1163; mul.wide.u32 %rd1165, %r1375, 12; add.s64 %rd2946, %rd1164, %rd1165; add.s32 %r1390, %r1375, 1; mul.wide.u32 %rd1166, %r1390, 12; add.s64 %rd2949, %rd1164, %rd1166; add.s32 %r1391, %r1375, 2; mul.wide.u32 %rd1167, %r1391, 12; add.s64 %rd2952, %rd1164, %rd1167; add.s64 %rd2947, %rd2946, 4; add.s64 %rd2948, %rd2946, 8; add.s64 %rd2950, %rd2949, 4; add.s64 %rd2951, %rd2949, 8; add.s64 %rd2953, %rd2952, 4; add.s64 %rd2954, %rd2952, 8; bra.uni $L__BB108_707; $L__BB108_705: setp.ge.s32 %p603, %r364, %r363; selp.b32 %r1384, -1, 0, %p603; add.s32 %r1385, %r379, %r1384; mad.lo.s32 %r1386, %r1385, %r362, %r3166; ld.const.u64 %rd1158, [kernel_params+2336]; cvta.to.global.u64 %rd1159, %rd1158; add.s32 %r1387, %r1372, %r1386; mul.wide.u32 %rd1160, %r1387, 12; add.s64 %rd2946, %rd1159, %rd1160; add.s32 %r1388, %r1373, %r1386; mul.wide.u32 %rd1161, %r1388, 12; add.s64 %rd2949, %rd1159, %rd1161; add.s32 %r1389, %r1374, %r1386; mul.wide.u32 %rd1162, %r1389, 12; add.s64 %rd2952, %rd1159, %rd1162; add.s64 %rd2947, %rd2946, 4; add.s64 %rd2948, %rd2946, 8; add.s64 %rd2950, %rd2949, 4; add.s64 %rd2951, %rd2949, 8; add.s64 %rd2953, %rd2952, 4; add.s64 %rd2954, %rd2952, 8; $L__BB108_707: cvt.rn.f32.s32 %f5831, %r364; sub.ftz.f32 %f5832, %f743, %f5831; mov.f32 %f5833, 0f3F800000; sub.ftz.f32 %f5834, %f5833, %f5832; ld.global.f32 %f5835, [%rd2946]; mul.ftz.f32 %f5836, %f5832, %f5835; ld.global.f32 %f5837, [%rd2947]; mul.ftz.f32 %f5838, %f5832, %f5837; ld.global.f32 %f5839, [%rd2948]; mul.ftz.f32 %f5840, %f5832, %f5839; fma.rn.ftz.f32 %f15814, %f5834, %f752, %f5836; fma.rn.ftz.f32 %f15815, %f5834, %f751, %f5838; fma.rn.ftz.f32 %f15816, %f5834, %f750, %f5840; ld.global.f32 %f5841, [%rd2949]; mul.ftz.f32 %f5842, %f5832, %f5841; ld.global.f32 %f5843, [%rd2950]; mul.ftz.f32 %f5844, %f5832, %f5843; ld.global.f32 %f5845, [%rd2951]; mul.ftz.f32 %f5846, %f5832, %f5845; fma.rn.ftz.f32 %f15817, %f5834, %f749, %f5842; fma.rn.ftz.f32 %f15818, %f5834, %f748, %f5844; fma.rn.ftz.f32 %f15819, %f5834, %f747, %f5846; ld.global.f32 %f5847, [%rd2952]; mul.ftz.f32 %f5848, %f5832, %f5847; ld.global.f32 %f5849, [%rd2953]; mul.ftz.f32 %f5850, %f5832, %f5849; ld.global.f32 %f5851, [%rd2954]; mul.ftz.f32 %f5852, %f5832, %f5851; fma.rn.ftz.f32 %f15820, %f5834, %f746, %f5848; fma.rn.ftz.f32 %f15821, %f5834, %f745, %f5850; fma.rn.ftz.f32 %f15822, %f5834, %f744, %f5852; mov.u16 %rs140, 1; $L__BB108_708: and.b32 %r380, %r361, 4; setp.ne.s32 %p604, %r380, 0; @%p604 bra $L__BB108_716; and.b32 %r1392, %r361, 2; setp.eq.s32 %p605, %r1392, 0; ld.const.u64 %rd1168, [kernel_params+2160]; cvta.to.global.u64 %rd1169, %rd1168; mul.lo.s64 %rd1170, %rd97, 208; add.s64 %rd159, %rd1169, %rd1170; @%p605 bra $L__BB108_714; bra.uni $L__BB108_710; $L__BB108_714: ld.global.v4.f32 {%f15802, %f15803, %f15804, %f15805}, [%rd159]; ld.global.v4.f32 {%f15806, %f15807, %f15808, %f15809}, [%rd159+16]; ld.global.v4.f32 {%f15810, %f15811, %f15812, %f15813}, [%rd159+32]; bra.uni $L__BB108_715; $L__BB108_710: ld.const.u64 %rd1171, [kernel_params+2176]; cvta.to.global.u64 %rd1172, %rd1171; ld.global.u32 %rd1173, [%rd159+168]; ld.global.u32 %r1393, [%rd159+152]; shl.b32 %r1394, %r1393, 1; cvt.rn.f32.s32 %f5853, %r1394; selp.f32 %f5854, %f733, 0f3F000000, %p593; mul.ftz.f32 %f5855, %f5854, %f5853; cvt.rzi.ftz.s32.f32 %r1395, %f5855; add.s32 %r1396, %r1394, -1; min.s32 %r1397, %r1395, %r1396; cvt.rn.f32.s32 %f5856, %r1397; sub.ftz.f32 %f771, %f5855, %f5856; cvt.s64.s32 %rd1174, %r1397; add.s64 %rd1175, %rd1174, %rd1173; shl.b64 %rd1176, %rd1175, 6; add.s64 %rd160, %rd1172, %rd1176; ld.global.v4.f32 {%f5857, %f5858, %f5859, %f5860}, [%rd160]; ld.global.v4.f32 {%f5861, %f5862, %f5863, %f5864}, [%rd160+64]; mul.ftz.f32 %f5865, %f5858, %f5862; fma.rn.ftz.f32 %f5866, %f5857, %f5861, %f5865; mul.ftz.f32 %f5867, %f5860, %f5864; fma.rn.ftz.f32 %f5868, %f5859, %f5863, %f5867; add.ftz.f32 %f780, %f5866, %f5868; setp.gt.ftz.f32 %p607, %f780, 0f3F7FDF3B; mov.f32 %f5869, 0f3F800000; sub.ftz.f32 %f781, %f5869, %f771; @%p607 bra $L__BB108_712; bra.uni $L__BB108_711; $L__BB108_712: mul.ftz.f32 %f5923, %f5857, %f781; fma.rn.ftz.f32 %f5924, %f5861, %f771, %f5923; mul.ftz.f32 %f5925, %f5858, %f781; fma.rn.ftz.f32 %f5926, %f5862, %f771, %f5925; mul.ftz.f32 %f5927, %f5859, %f781; fma.rn.ftz.f32 %f5928, %f5863, %f771, %f5927; mul.ftz.f32 %f5929, %f5860, %f781; fma.rn.ftz.f32 %f5930, %f5864, %f771, %f5929; mul.ftz.f32 %f5931, %f5926, %f5926; fma.rn.ftz.f32 %f5932, %f5924, %f5924, %f5931; mul.ftz.f32 %f5933, %f5930, %f5930; fma.rn.ftz.f32 %f5934, %f5928, %f5928, %f5933; add.ftz.f32 %f5935, %f5932, %f5934; rsqrt.approx.ftz.f32 %f5936, %f5935; mul.ftz.f32 %f15801, %f5930, %f5936; mul.ftz.f32 %f15800, %f5928, %f5936; mul.ftz.f32 %f15799, %f5926, %f5936; mul.ftz.f32 %f15798, %f5924, %f5936; bra.uni $L__BB108_713; $L__BB108_1332: mov.u32 %r3226, 0; st.local.u32 [%rd289], %r3226; $L__BB108_1333: add.s64 %rd1737, %rd1727, 76; cvta.to.local.u64 %rd1738, %rd1737; st.local.u32 [%rd1738], %r622; add.s64 %rd1739, %rd1727, 60; cvta.to.local.u64 %rd1740, %rd1739; st.local.u32 [%rd1740], %r623; add.s64 %rd1741, %rd1727, 68; cvta.to.local.u64 %rd1742, %rd1741; st.local.f32 [%rd1742], %f1958; add.s64 %rd1743, %rd1727, 72; cvta.to.local.u64 %rd1744, %rd1743; st.local.f32 [%rd1744], %f1959; mov.b32 %r2021, %f1939; mov.u32 %r2022, -1; add.s64 %rd290, %rd1727, 80; cvta.to.local.u64 %rd1745, %rd290; st.local.v2.u32 [%rd1745], {%r2022, %r2021}; add.s64 %rd1746, %rd1727, 88; cvta.to.local.u64 %rd1747, %rd1746; st.local.f32 [%rd1747], %f1960; add.s64 %rd1748, %rd1727, 52; cvta.to.local.u64 %rd291, %rd1748; st.local.u32 [%rd291], %r620; add.s64 %rd1749, %rd1727, 56; cvta.to.local.u64 %rd292, %rd1749; mov.u32 %r2023, 0; st.local.u32 [%rd292], %r2023; setp.eq.s32 %p1254, %r622, -1; @%p1254 bra $L__BB108_1358; ld.const.u64 %rd2912, [kernel_params+2184]; cvta.to.global.u64 %rd2911, %rd2912; shl.b64 %rd1750, %rd285, 2; add.s64 %rd1751, %rd2911, %rd1750; ld.global.u32 %r626, [%rd1751]; st.local.u32 [%rd292], %r626; and.b32 %r627, %r626, 2; setp.eq.s32 %p1255, %r627, 0; @%p1255 bra $L__BB108_1341; ld.const.u64 %rd1752, [kernel_params+2160]; cvta.to.global.u64 %rd1753, %rd1752; mul.lo.s64 %rd1754, %rd285, 208; add.s64 %rd1755, %rd1753, %rd1754; ld.const.u64 %rd1756, [kernel_params+2176]; cvta.to.global.u64 %rd1757, %rd1756; ld.global.u32 %rd1758, [%rd1755+168]; ld.global.u32 %r2024, [%rd1755+152]; shl.b32 %r2025, %r2024, 1; cvt.rn.f32.s32 %f9715, %r2025; mul.ftz.f32 %f9716, %f1939, %f9715; cvt.rzi.ftz.s32.f32 %r2026, %f9716; add.s32 %r2027, %r2025, -1; min.s32 %r2028, %r2026, %r2027; cvt.rn.f32.s32 %f9717, %r2028; sub.ftz.f32 %f1969, %f9716, %f9717; cvt.s64.s32 %rd1759, %r2028; add.s64 %rd1760, %rd1759, %rd1758; shl.b64 %rd1761, %rd1760, 6; add.s64 %rd293, %rd1757, %rd1761; ld.global.v4.f32 {%f9718, %f9719, %f9720, %f9721}, [%rd293]; ld.global.v4.f32 {%f9722, %f9723, %f9724, %f9725}, [%rd293+64]; mul.ftz.f32 %f9726, %f9719, %f9723; fma.rn.ftz.f32 %f9727, %f9718, %f9722, %f9726; mul.ftz.f32 %f9728, %f9721, %f9725; fma.rn.ftz.f32 %f9729, %f9720, %f9724, %f9728; add.ftz.f32 %f1978, %f9727, %f9729; setp.gt.ftz.f32 %p1256, %f1978, 0f3F7FDF3B; mov.f32 %f9730, 0f3F800000; sub.ftz.f32 %f1979, %f9730, %f1969; @%p1256 bra $L__BB108_1337; bra.uni $L__BB108_1336; $L__BB108_1337: mul.ftz.f32 %f9784, %f9718, %f1979; fma.rn.ftz.f32 %f9785, %f9722, %f1969, %f9784; mul.ftz.f32 %f9786, %f9719, %f1979; fma.rn.ftz.f32 %f9787, %f9723, %f1969, %f9786; mul.ftz.f32 %f9788, %f9720, %f1979; fma.rn.ftz.f32 %f9789, %f9724, %f1969, %f9788; mul.ftz.f32 %f9790, %f9721, %f1979; fma.rn.ftz.f32 %f9791, %f9725, %f1969, %f9790; mul.ftz.f32 %f9792, %f9787, %f9787; fma.rn.ftz.f32 %f9793, %f9785, %f9785, %f9792; mul.ftz.f32 %f9794, %f9791, %f9791; fma.rn.ftz.f32 %f9795, %f9789, %f9789, %f9794; add.ftz.f32 %f9796, %f9793, %f9795; rsqrt.approx.ftz.f32 %f9797, %f9796; mul.ftz.f32 %f16190, %f9791, %f9797; mul.ftz.f32 %f16189, %f9789, %f9797; mul.ftz.f32 %f16188, %f9787, %f9797; mul.ftz.f32 %f16187, %f9785, %f9797; bra.uni $L__BB108_1338; $L__BB108_1358: setp.eq.s32 %p1274, %r624, -1; @%p1274 bra $L__BB108_1360; st.local.u32 [%rd1745], %r624; $L__BB108_1360: add.s64 %rd1854, %rd1727, 116; cvta.to.local.u64 %rd1855, %rd1854; mov.u32 %r2042, 0; st.local.u32 [%rd1855], %r2042; add.s64 %rd1856, %rd1727, 120; cvta.to.local.u64 %rd1857, %rd1856; st.local.u32 [%rd1857], %r2042; add.s64 %rd1858, %rd1727, 124; cvta.to.local.u64 %rd1859, %rd1858; st.local.u32 [%rd1859], %r2042; add.s64 %rd1860, %rd1727, 136; cvta.to.local.u64 %rd1861, %rd1860; add.s64 %rd1862, %rd1727, 128; cvta.to.local.u64 %rd1863, %rd1862; mov.f32 %f16207, 0f00000000; st.local.v2.f32 [%rd1863], {%f16207, %f16207}; st.local.u32 [%rd1861], %r2042; mov.f32 %f16208, %f16207; mov.f32 %f16209, %f16207; mov.f32 %f16210, %f16207; mov.f32 %f16211, %f16207; mov.f32 %f16212, %f16207; mov.f32 %f16213, %f1957; mov.f32 %f16214, %f16218; mov.f32 %f16215, %f16217; bra.uni $L__BB108_1361; $L__BB108_1336: mov.f32 %f9731, 0fBF800000; max.ftz.f32 %f9732, %f1978, %f9731; min.ftz.f32 %f9734, %f9732, %f9730; abs.ftz.f32 %f9735, %f9734; sub.ftz.f32 %f9736, %f9730, %f9735; mul.ftz.f32 %f9737, %f9736, 0f3F000000; sqrt.approx.ftz.f32 %f9738, %f9737; setp.gt.ftz.f32 %p1257, %f9735, 0f3F11EB85; selp.f32 %f9739, %f9738, %f9735, %p1257; mul.ftz.f32 %f9740, %f9739, %f9739; mov.f32 %f9741, 0f3C94D2E9; mov.f32 %f9742, 0f3D53F941; fma.rn.ftz.f32 %f9743, %f9742, %f9740, %f9741; mov.f32 %f9744, 0f3D3F841F; fma.rn.ftz.f32 %f9745, %f9743, %f9740, %f9744; mov.f32 %f9746, 0f3D994929; fma.rn.ftz.f32 %f9747, %f9745, %f9740, %f9746; mov.f32 %f9748, 0f3E2AAB94; fma.rn.ftz.f32 %f9749, %f9747, %f9740, %f9748; mul.ftz.f32 %f9750, %f9740, %f9749; fma.rn.ftz.f32 %f9751, %f9750, %f9739, %f9739; add.ftz.f32 %f9752, %f9751, %f9751; mov.f32 %f9753, 0f3FC90FDB; sub.ftz.f32 %f9754, %f9753, %f9751; selp.f32 %f9755, %f9752, %f9754, %p1257; setp.lt.ftz.f32 %p1258, %f9734, 0f00000000; mov.f32 %f9756, 0f40490FDB; sub.ftz.f32 %f9757, %f9756, %f9755; selp.f32 %f9758, %f9757, %f9755, %p1258; mul.ftz.f32 %f9759, %f9718, %f1978; sub.ftz.f32 %f9760, %f9722, %f9759; mul.ftz.f32 %f9761, %f9719, %f1978; sub.ftz.f32 %f9762, %f9723, %f9761; mul.ftz.f32 %f9763, %f9720, %f1978; sub.ftz.f32 %f9764, %f9724, %f9763; mul.ftz.f32 %f9765, %f9721, %f1978; sub.ftz.f32 %f9766, %f9725, %f9765; mul.ftz.f32 %f9767, %f9762, %f9762; fma.rn.ftz.f32 %f9768, %f9760, %f9760, %f9767; mul.ftz.f32 %f9769, %f9766, %f9766; fma.rn.ftz.f32 %f9770, %f9764, %f9764, %f9769; add.ftz.f32 %f9771, %f9768, %f9770; rsqrt.approx.ftz.f32 %f9772, %f9771; mul.ftz.f32 %f9773, %f9760, %f9772; mul.ftz.f32 %f9774, %f9762, %f9772; mul.ftz.f32 %f9775, %f9764, %f9772; mul.ftz.f32 %f9776, %f9766, %f9772; mul.ftz.f32 %f9777, %f1969, %f9758; cos.approx.ftz.f32 %f9778, %f9777; sin.approx.ftz.f32 %f9779, %f9777; mul.ftz.f32 %f9780, %f9773, %f9779; mul.ftz.f32 %f9781, %f9774, %f9779; mul.ftz.f32 %f9782, %f9775, %f9779; mul.ftz.f32 %f9783, %f9776, %f9779; fma.rn.ftz.f32 %f16190, %f9721, %f9778, %f9783; fma.rn.ftz.f32 %f16189, %f9720, %f9778, %f9782; fma.rn.ftz.f32 %f16188, %f9719, %f9778, %f9781; fma.rn.ftz.f32 %f16187, %f9718, %f9778, %f9780; $L__BB108_1338: ld.global.v4.f32 {%f9798, %f9799, %f9800, %f9801}, [%rd293+16]; ld.global.v4.f32 {%f9806, %f9807, %f9808, %f9809}, [%rd293+80]; mul.ftz.f32 %f9814, %f1969, %f9806; mul.ftz.f32 %f9815, %f1969, %f9807; mul.ftz.f32 %f9816, %f1969, %f9808; mul.ftz.f32 %f9817, %f1969, %f9809; fma.rn.ftz.f32 %f1992, %f1979, %f9798, %f9814; fma.rn.ftz.f32 %f1993, %f1979, %f9799, %f9815; fma.rn.ftz.f32 %f1994, %f1979, %f9800, %f9816; fma.rn.ftz.f32 %f9818, %f1979, %f9801, %f9817; ld.global.v4.f32 {%f9819, %f9820, %f9821, %f9822}, [%rd293+32]; ld.global.v4.f32 {%f9827, %f9828, %f9829, %f9830}, [%rd293+96]; mul.ftz.f32 %f9835, %f1969, %f9827; mul.ftz.f32 %f9836, %f1969, %f9828; mul.ftz.f32 %f9837, %f1969, %f9829; mul.ftz.f32 %f9838, %f1969, %f9830; fma.rn.ftz.f32 %f9839, %f1979, %f9819, %f9835; fma.rn.ftz.f32 %f9840, %f1979, %f9820, %f9836; fma.rn.ftz.f32 %f9841, %f1979, %f9821, %f9837; fma.rn.ftz.f32 %f9842, %f1979, %f9822, %f9838; ld.global.v4.f32 {%f9843, %f9844, %f9845, %f9846}, [%rd293+48]; ld.global.v4.f32 {%f9851, %f9852, %f9853, %f9854}, [%rd293+112]; mul.ftz.f32 %f9859, %f1969, %f9851; mul.ftz.f32 %f9860, %f1969, %f9852; mul.ftz.f32 %f9861, %f1969, %f9853; mul.ftz.f32 %f9862, %f1969, %f9854; fma.rn.ftz.f32 %f9863, %f1979, %f9843, %f9859; fma.rn.ftz.f32 %f9864, %f1979, %f9844, %f9860; fma.rn.ftz.f32 %f9865, %f1979, %f9845, %f9861; fma.rn.ftz.f32 %f9866, %f1979, %f9846, %f9862; mul.ftz.f32 %f9867, %f16190, 0f3FB504F3; mul.ftz.f32 %f9868, %f16187, 0f3FB504F3; mul.ftz.f32 %f9869, %f16188, 0f3FB504F3; mul.ftz.f32 %f9870, %f16189, 0f3FB504F3; mul.ftz.f32 %f9871, %f9867, %f9868; mul.ftz.f32 %f9872, %f9867, %f9869; mul.ftz.f32 %f9873, %f9867, %f9870; mul.ftz.f32 %f9874, %f9868, %f9868; mul.ftz.f32 %f9875, %f9868, %f9869; mul.ftz.f32 %f9876, %f9868, %f9870; mul.ftz.f32 %f9877, %f9869, %f9869; mul.ftz.f32 %f9878, %f9869, %f9870; mul.ftz.f32 %f9879, %f9870, %f9870; sub.ftz.f32 %f9881, %f9730, %f9877; sub.ftz.f32 %f9882, %f9881, %f9879; sub.ftz.f32 %f9883, %f9875, %f9873; add.ftz.f32 %f9884, %f9872, %f9876; add.ftz.f32 %f9885, %f9873, %f9875; sub.ftz.f32 %f9886, %f9730, %f9874; sub.ftz.f32 %f9887, %f9886, %f9879; sub.ftz.f32 %f9888, %f9878, %f9871; sub.ftz.f32 %f9889, %f9876, %f9872; add.ftz.f32 %f9890, %f9871, %f9878; sub.ftz.f32 %f9891, %f9886, %f9877; mul.ftz.f32 %f9892, %f9883, %f9841; fma.rn.ftz.f32 %f9893, %f9882, %f9818, %f9892; fma.rn.ftz.f32 %f16196, %f9884, %f9864, %f9893; mul.ftz.f32 %f9894, %f9883, %f9842; fma.rn.ftz.f32 %f9895, %f9882, %f9839, %f9894; fma.rn.ftz.f32 %f1996, %f9884, %f9865, %f9895; mul.ftz.f32 %f9896, %f9883, %f9863; fma.rn.ftz.f32 %f9897, %f9882, %f9840, %f9896; fma.rn.ftz.f32 %f1997, %f9884, %f9866, %f9897; mul.ftz.f32 %f9898, %f9887, %f9841; fma.rn.ftz.f32 %f9899, %f9885, %f9818, %f9898; fma.rn.ftz.f32 %f1998, %f9888, %f9864, %f9899; mul.ftz.f32 %f9900, %f9887, %f9842; fma.rn.ftz.f32 %f9901, %f9885, %f9839, %f9900; fma.rn.ftz.f32 %f16195, %f9888, %f9865, %f9901; mul.ftz.f32 %f9902, %f9887, %f9863; fma.rn.ftz.f32 %f9903, %f9885, %f9840, %f9902; fma.rn.ftz.f32 %f2000, %f9888, %f9866, %f9903; mul.ftz.f32 %f9904, %f9890, %f9841; fma.rn.ftz.f32 %f9905, %f9889, %f9818, %f9904; fma.rn.ftz.f32 %f2001, %f9891, %f9864, %f9905; mul.ftz.f32 %f9906, %f9890, %f9842; fma.rn.ftz.f32 %f9907, %f9889, %f9839, %f9906; fma.rn.ftz.f32 %f2002, %f9891, %f9865, %f9907; mul.ftz.f32 %f9908, %f9890, %f9863; fma.rn.ftz.f32 %f9909, %f9889, %f9840, %f9908; fma.rn.ftz.f32 %f16194, %f9891, %f9866, %f9909; add.s64 %rd1763, %rd1727, 144; cvta.to.local.u64 %rd1764, %rd1763; st.local.v4.f32 [%rd1764], {%f16196, %f1996, %f1997, %f1992}; add.s64 %rd1765, %rd1727, 160; cvta.to.local.u64 %rd1766, %rd1765; st.local.v4.f32 [%rd1766], {%f1998, %f16195, %f2000, %f1993}; add.s64 %rd1767, %rd1727, 176; cvta.to.local.u64 %rd1768, %rd1767; st.local.v4.f32 [%rd1768], {%f2001, %f2002, %f16194, %f1994}; mul.ftz.f32 %f9910, %f16195, %f16194; mul.ftz.f32 %f2004, %f2002, %f2000; sub.ftz.f32 %f16193, %f9910, %f2004; mul.ftz.f32 %f2006, %f2002, %f1997; mul.ftz.f32 %f9911, %f1996, %f16194; sub.ftz.f32 %f16192, %f2006, %f9911; mul.ftz.f32 %f2008, %f1996, %f2000; mul.ftz.f32 %f9912, %f16195, %f1997; sub.ftz.f32 %f16191, %f2008, %f9912; mul.ftz.f32 %f9913, %f1998, %f16192; fma.rn.ftz.f32 %f9914, %f16196, %f16193, %f9913; fma.rn.ftz.f32 %f16197, %f2001, %f16191, %f9914; setp.neu.ftz.f32 %p1259, %f16197, 0f00000000; @%p1259 bra $L__BB108_1340; add.ftz.f32 %f16196, %f16196, 0f322BCC77; add.ftz.f32 %f16194, %f16194, 0f322BCC77; add.ftz.f32 %f16195, %f16195, 0f322BCC77; mul.ftz.f32 %f9915, %f16195, %f16194; sub.ftz.f32 %f16193, %f9915, %f2004; mul.ftz.f32 %f9916, %f1996, %f16194; sub.ftz.f32 %f16192, %f2006, %f9916; mul.ftz.f32 %f9917, %f16195, %f1997; sub.ftz.f32 %f16191, %f2008, %f9917; mul.ftz.f32 %f9918, %f1998, %f16192; fma.rn.ftz.f32 %f9919, %f16196, %f16193, %f9918; fma.rn.ftz.f32 %f9920, %f2001, %f16191, %f9919; setp.eq.ftz.f32 %p1260, %f9920, 0f00000000; selp.f32 %f16197, 0f7F7FFFFF, %f9920, %p1260; $L__BB108_1340: mul.ftz.f32 %f9921, %f1998, %f16194; mul.ftz.f32 %f9922, %f2001, %f2000; sub.ftz.f32 %f9923, %f9922, %f9921; mul.ftz.f32 %f9924, %f2001, %f1997; mul.ftz.f32 %f9925, %f16194, %f16196; sub.ftz.f32 %f9926, %f9925, %f9924; mul.ftz.f32 %f9927, %f2000, %f16196; mul.ftz.f32 %f9928, %f1998, %f1997; sub.ftz.f32 %f9929, %f9928, %f9927; mul.ftz.f32 %f9930, %f2001, %f16195; mul.ftz.f32 %f9931, %f1998, %f2002; sub.ftz.f32 %f9932, %f9931, %f9930; mul.ftz.f32 %f9933, %f2002, %f16196; mul.ftz.f32 %f9934, %f2001, %f1996; sub.ftz.f32 %f9935, %f9934, %f9933; mul.ftz.f32 %f9936, %f1998, %f1996; mul.ftz.f32 %f9937, %f16195, %f16196; sub.ftz.f32 %f9938, %f9937, %f9936; div.approx.ftz.f32 %f9939, %f16192, %f16197; mul.ftz.f32 %f9940, %f1993, %f9939; div.approx.ftz.f32 %f9941, %f16193, %f16197; fma.rn.ftz.f32 %f9942, %f1992, %f9941, %f9940; div.approx.ftz.f32 %f9943, %f16191, %f16197; fma.rn.ftz.f32 %f9944, %f1994, %f9943, %f9942; div.approx.ftz.f32 %f9945, %f9926, %f16197; mul.ftz.f32 %f9946, %f1993, %f9945; div.approx.ftz.f32 %f9947, %f9923, %f16197; fma.rn.ftz.f32 %f9948, %f1992, %f9947, %f9946; div.approx.ftz.f32 %f9949, %f9929, %f16197; fma.rn.ftz.f32 %f9950, %f1994, %f9949, %f9948; div.approx.ftz.f32 %f9951, %f9935, %f16197; mul.ftz.f32 %f9952, %f1993, %f9951; div.approx.ftz.f32 %f9953, %f9932, %f16197; fma.rn.ftz.f32 %f9954, %f1992, %f9953, %f9952; div.approx.ftz.f32 %f9955, %f9938, %f16197; fma.rn.ftz.f32 %f9956, %f1994, %f9955, %f9954; neg.ftz.f32 %f9957, %f9944; add.s64 %rd1770, %rd1727, 192; cvta.to.local.u64 %rd1771, %rd1770; st.local.v4.f32 [%rd1771], {%f9941, %f9939, %f9943, %f9957}; neg.ftz.f32 %f9958, %f9950; add.s64 %rd1772, %rd1727, 208; cvta.to.local.u64 %rd1773, %rd1772; st.local.v4.f32 [%rd1773], {%f9947, %f9945, %f9949, %f9958}; neg.ftz.f32 %f9959, %f9956; add.s64 %rd1774, %rd1727, 224; cvta.to.local.u64 %rd1775, %rd1774; st.local.v4.f32 [%rd1775], {%f9953, %f9951, %f9955, %f9959}; $L__BB108_1341: setp.eq.s32 %p1261, %r3226, 1; @%p1261 bra $L__BB108_1343; bra.uni $L__BB108_1342; $L__BB108_1343: setp.lt.s32 %p1262, %r618, 0; @%p1262 bra $L__BB108_1345; ld.const.u64 %rd2982, [kernel_params+2232]; and.b32 %r3227, %r626, 4; mov.f32 %f16213, %f1957; mov.f32 %f16214, %f16218; mov.f32 %f16215, %f16217; bra.uni $L__BB108_1351; $L__BB108_1342: add.s64 %rd1777, %rd1727, 116; cvta.to.local.u64 %rd1778, %rd1777; st.local.u32 [%rd1778], %r2023; add.s64 %rd1779, %rd1727, 120; cvta.to.local.u64 %rd1780, %rd1779; st.local.u32 [%rd1780], %r2023; add.s64 %rd1781, %rd1727, 124; cvta.to.local.u64 %rd1782, %rd1781; st.local.u32 [%rd1782], %r2023; add.s64 %rd1783, %rd1727, 136; cvta.to.local.u64 %rd1784, %rd1783; add.s64 %rd1785, %rd1727, 128; cvta.to.local.u64 %rd1786, %rd1785; mov.f32 %f16207, 0f00000000; st.local.v2.f32 [%rd1786], {%f16207, %f16207}; st.local.u32 [%rd1784], %r2023; mov.f32 %f16208, %f16207; mov.f32 %f16209, %f16207; mov.f32 %f16210, %f16207; mov.f32 %f16211, %f16207; mov.f32 %f16212, %f16207; mov.f32 %f16213, %f1957; mov.f32 %f16214, %f16218; mov.f32 %f16215, %f16217; bra.uni $L__BB108_1361; $L__BB108_1345: ld.const.v2.u64 {%rd1787, %rd2982}, [kernel_params+2224]; cvta.to.global.u64 %rd1790, %rd2982; shl.b64 %rd1791, %rd286, 4; add.s64 %rd1792, %rd1790, %rd1791; ld.global.v4.u32 {%r2030, %r2031, %r2032, %r2033}, [%rd1792]; cvta.to.global.u64 %rd1793, %rd1787; mul.wide.u32 %rd1794, %r2030, 12; add.s64 %rd1795, %rd1793, %rd1794; mul.wide.u32 %rd1796, %r2031, 12; add.s64 %rd1797, %rd1793, %rd1796; mul.wide.u32 %rd1798, %r2032, 12; add.s64 %rd1799, %rd1793, %rd1798; mov.f32 %f9966, 0f3F800000; sub.ftz.f32 %f9967, %f9966, %f1958; sub.ftz.f32 %f9968, %f9967, %f1959; ld.global.f32 %f9969, [%rd1795]; ld.global.f32 %f9970, [%rd1795+4]; ld.global.f32 %f9971, [%rd1795+8]; ld.global.f32 %f9972, [%rd1797]; mul.ftz.f32 %f9973, %f1958, %f9972; ld.global.f32 %f9974, [%rd1797+4]; mul.ftz.f32 %f9975, %f1958, %f9974; ld.global.f32 %f9976, [%rd1797+8]; mul.ftz.f32 %f9977, %f1958, %f9976; fma.rn.ftz.f32 %f9978, %f9968, %f9969, %f9973; fma.rn.ftz.f32 %f9979, %f9968, %f9970, %f9975; fma.rn.ftz.f32 %f9980, %f9968, %f9971, %f9977; ld.global.f32 %f9981, [%rd1799]; ld.global.f32 %f9982, [%rd1799+4]; ld.global.f32 %f9983, [%rd1799+8]; fma.rn.ftz.f32 %f16198, %f1959, %f9981, %f9978; fma.rn.ftz.f32 %f16199, %f1959, %f9982, %f9979; fma.rn.ftz.f32 %f16200, %f1959, %f9983, %f9980; mul.ftz.f32 %f9984, %f16199, %f16199; fma.rn.ftz.f32 %f9985, %f16198, %f16198, %f9984; fma.rn.ftz.f32 %f9986, %f16200, %f16200, %f9985; sqrt.approx.ftz.f32 %f2028, %f9986; setp.neu.ftz.f32 %p1263, %f2028, 0f00000000; @%p1263 bra $L__BB108_1346; bra.uni $L__BB108_1347; $L__BB108_1346: rcp.approx.ftz.f32 %f9987, %f2028; mul.ftz.f32 %f16200, %f16200, %f9987; mul.ftz.f32 %f16199, %f16199, %f9987; mul.ftz.f32 %f16198, %f16198, %f9987; $L__BB108_1347: add.s64 %rd1800, %rd284, 8; setp.eq.ftz.f32 %p1264, %f16198, 0f00000000; setp.eq.ftz.f32 %p1265, %f16199, 0f00000000; and.pred %p1266, %p1264, %p1265; setp.eq.ftz.f32 %p1267, %f16200, 0f00000000; and.pred %p1268, %p1267, %p1266; selp.f32 %f16213, %f1957, %f16200, %p1268; selp.f32 %f16214, %f16218, %f16199, %p1268; selp.f32 %f16215, %f16217, %f16198, %p1268; cvta.to.local.u64 %rd297, %rd284; st.local.f32 [%rd297], %f16215; st.local.f32 [%rd287], %f16214; cvta.to.local.u64 %rd299, %rd1800; st.local.f32 [%rd299], %f16213; and.b32 %r3227, %r626, 4; setp.ne.s32 %p1269, %r3227, 0; @%p1269 bra $L__BB108_1351; @%p1255 bra $L__BB108_1350; bra.uni $L__BB108_1349; $L__BB108_1350: ld.const.u64 %rd1809, [kernel_params+2160]; cvta.to.global.u64 %rd1810, %rd1809; mul.lo.s64 %rd1811, %rd285, 208; add.s64 %rd1812, %rd1810, %rd1811; ld.global.v4.f32 {%f10022, %f10023, %f10024, %f10025}, [%rd1812+48]; ld.global.v4.f32 {%f10029, %f10030, %f10031, %f10032}, [%rd1812+64]; ld.global.v4.f32 {%f10036, %f10037, %f10038, %f10039}, [%rd1812+80]; mul.ftz.f32 %f10043, %f10029, %f16214; fma.rn.ftz.f32 %f10044, %f10022, %f16215, %f10043; fma.rn.ftz.f32 %f10045, %f10036, %f16213, %f10044; mul.ftz.f32 %f10046, %f10030, %f16214; fma.rn.ftz.f32 %f10047, %f10023, %f16215, %f10046; fma.rn.ftz.f32 %f10048, %f10037, %f16213, %f10047; mul.ftz.f32 %f10049, %f10031, %f16214; fma.rn.ftz.f32 %f10050, %f10024, %f16215, %f10049; fma.rn.ftz.f32 %f10051, %f10038, %f16213, %f10050; mul.ftz.f32 %f10052, %f10048, %f10048; fma.rn.ftz.f32 %f10053, %f10045, %f10045, %f10052; fma.rn.ftz.f32 %f10054, %f10051, %f10051, %f10053; sqrt.approx.ftz.f32 %f10055, %f10054; div.approx.ftz.f32 %f16215, %f10045, %f10055; div.approx.ftz.f32 %f16214, %f10048, %f10055; div.approx.ftz.f32 %f16213, %f10051, %f10055; st.local.f32 [%rd297], %f16215; st.local.f32 [%rd287], %f16214; st.local.f32 [%rd299], %f16213; mov.u32 %r3227, %r2023; bra.uni $L__BB108_1351; $L__BB108_711: mov.f32 %f5870, 0fBF800000; max.ftz.f32 %f5871, %f780, %f5870; min.ftz.f32 %f5873, %f5871, %f5869; abs.ftz.f32 %f5874, %f5873; sub.ftz.f32 %f5875, %f5869, %f5874; mul.ftz.f32 %f5876, %f5875, 0f3F000000; sqrt.approx.ftz.f32 %f5877, %f5876; setp.gt.ftz.f32 %p608, %f5874, 0f3F11EB85; selp.f32 %f5878, %f5877, %f5874, %p608; mul.ftz.f32 %f5879, %f5878, %f5878; mov.f32 %f5880, 0f3C94D2E9; mov.f32 %f5881, 0f3D53F941; fma.rn.ftz.f32 %f5882, %f5881, %f5879, %f5880; mov.f32 %f5883, 0f3D3F841F; fma.rn.ftz.f32 %f5884, %f5882, %f5879, %f5883; mov.f32 %f5885, 0f3D994929; fma.rn.ftz.f32 %f5886, %f5884, %f5879, %f5885; mov.f32 %f5887, 0f3E2AAB94; fma.rn.ftz.f32 %f5888, %f5886, %f5879, %f5887; mul.ftz.f32 %f5889, %f5879, %f5888; fma.rn.ftz.f32 %f5890, %f5889, %f5878, %f5878; add.ftz.f32 %f5891, %f5890, %f5890; mov.f32 %f5892, 0f3FC90FDB; sub.ftz.f32 %f5893, %f5892, %f5890; selp.f32 %f5894, %f5891, %f5893, %p608; setp.lt.ftz.f32 %p609, %f5873, 0f00000000; mov.f32 %f5895, 0f40490FDB; sub.ftz.f32 %f5896, %f5895, %f5894; selp.f32 %f5897, %f5896, %f5894, %p609; mul.ftz.f32 %f5898, %f5857, %f780; sub.ftz.f32 %f5899, %f5861, %f5898; mul.ftz.f32 %f5900, %f5858, %f780; sub.ftz.f32 %f5901, %f5862, %f5900; mul.ftz.f32 %f5902, %f5859, %f780; sub.ftz.f32 %f5903, %f5863, %f5902; mul.ftz.f32 %f5904, %f5860, %f780; sub.ftz.f32 %f5905, %f5864, %f5904; mul.ftz.f32 %f5906, %f5901, %f5901; fma.rn.ftz.f32 %f5907, %f5899, %f5899, %f5906; mul.ftz.f32 %f5908, %f5905, %f5905; fma.rn.ftz.f32 %f5909, %f5903, %f5903, %f5908; add.ftz.f32 %f5910, %f5907, %f5909; rsqrt.approx.ftz.f32 %f5911, %f5910; mul.ftz.f32 %f5912, %f5899, %f5911; mul.ftz.f32 %f5913, %f5901, %f5911; mul.ftz.f32 %f5914, %f5903, %f5911; mul.ftz.f32 %f5915, %f5905, %f5911; mul.ftz.f32 %f5916, %f771, %f5897; cos.approx.ftz.f32 %f5917, %f5916; sin.approx.ftz.f32 %f5918, %f5916; mul.ftz.f32 %f5919, %f5912, %f5918; mul.ftz.f32 %f5920, %f5913, %f5918; mul.ftz.f32 %f5921, %f5914, %f5918; mul.ftz.f32 %f5922, %f5915, %f5918; fma.rn.ftz.f32 %f15801, %f5860, %f5917, %f5922; fma.rn.ftz.f32 %f15800, %f5859, %f5917, %f5921; fma.rn.ftz.f32 %f15799, %f5858, %f5917, %f5920; fma.rn.ftz.f32 %f15798, %f5857, %f5917, %f5919; $L__BB108_713: ld.global.v4.f32 {%f5937, %f5938, %f5939, %f5940}, [%rd160+16]; ld.global.v4.f32 {%f5945, %f5946, %f5947, %f5948}, [%rd160+80]; mul.ftz.f32 %f5953, %f771, %f5945; mul.ftz.f32 %f5954, %f771, %f5946; mul.ftz.f32 %f5955, %f771, %f5947; mul.ftz.f32 %f5956, %f771, %f5948; fma.rn.ftz.f32 %f5957, %f781, %f5940, %f5956; ld.global.v4.f32 {%f5958, %f5959, %f5960, %f5961}, [%rd160+32]; ld.global.v4.f32 {%f5966, %f5967, %f5968, %f5969}, [%rd160+96]; mul.ftz.f32 %f5974, %f771, %f5966; mul.ftz.f32 %f5975, %f771, %f5967; mul.ftz.f32 %f5976, %f771, %f5968; mul.ftz.f32 %f5977, %f771, %f5969; fma.rn.ftz.f32 %f5978, %f781, %f5958, %f5974; fma.rn.ftz.f32 %f5979, %f781, %f5959, %f5975; fma.rn.ftz.f32 %f5980, %f781, %f5960, %f5976; fma.rn.ftz.f32 %f5981, %f781, %f5961, %f5977; ld.global.v4.f32 {%f5982, %f5983, %f5984, %f5985}, [%rd160+48]; ld.global.v4.f32 {%f5990, %f5991, %f5992, %f5993}, [%rd160+112]; mul.ftz.f32 %f5998, %f771, %f5990; mul.ftz.f32 %f5999, %f771, %f5991; mul.ftz.f32 %f6000, %f771, %f5992; mul.ftz.f32 %f6001, %f771, %f5993; fma.rn.ftz.f32 %f6002, %f781, %f5982, %f5998; fma.rn.ftz.f32 %f6003, %f781, %f5983, %f5999; fma.rn.ftz.f32 %f6004, %f781, %f5984, %f6000; fma.rn.ftz.f32 %f6005, %f781, %f5985, %f6001; mul.ftz.f32 %f6006, %f15801, 0f3FB504F3; mul.ftz.f32 %f6007, %f15798, 0f3FB504F3; mul.ftz.f32 %f6008, %f15799, 0f3FB504F3; mul.ftz.f32 %f6009, %f15800, 0f3FB504F3; mul.ftz.f32 %f6010, %f6006, %f6007; mul.ftz.f32 %f6011, %f6006, %f6008; mul.ftz.f32 %f6012, %f6006, %f6009; mul.ftz.f32 %f6013, %f6007, %f6007; mul.ftz.f32 %f6014, %f6007, %f6008; mul.ftz.f32 %f6015, %f6007, %f6009; mul.ftz.f32 %f6016, %f6008, %f6008; mul.ftz.f32 %f6017, %f6008, %f6009; mul.ftz.f32 %f6018, %f6009, %f6009; sub.ftz.f32 %f6020, %f5869, %f6016; sub.ftz.f32 %f6021, %f6020, %f6018; sub.ftz.f32 %f6022, %f6014, %f6012; add.ftz.f32 %f6023, %f6011, %f6015; add.ftz.f32 %f6024, %f6012, %f6014; sub.ftz.f32 %f6025, %f5869, %f6013; sub.ftz.f32 %f6026, %f6025, %f6018; sub.ftz.f32 %f6027, %f6017, %f6010; sub.ftz.f32 %f6028, %f6015, %f6011; add.ftz.f32 %f6029, %f6010, %f6017; sub.ftz.f32 %f6030, %f6025, %f6016; mul.ftz.f32 %f6031, %f6022, %f5980; fma.rn.ftz.f32 %f6032, %f6021, %f5957, %f6031; mul.ftz.f32 %f6033, %f6022, %f5981; fma.rn.ftz.f32 %f6034, %f6021, %f5978, %f6033; mul.ftz.f32 %f6035, %f6022, %f6002; fma.rn.ftz.f32 %f6036, %f6021, %f5979, %f6035; mul.ftz.f32 %f6037, %f6026, %f5980; fma.rn.ftz.f32 %f6038, %f6024, %f5957, %f6037; mul.ftz.f32 %f6039, %f6026, %f5981; fma.rn.ftz.f32 %f6040, %f6024, %f5978, %f6039; mul.ftz.f32 %f6041, %f6026, %f6002; fma.rn.ftz.f32 %f6042, %f6024, %f5979, %f6041; mul.ftz.f32 %f6043, %f6029, %f5980; fma.rn.ftz.f32 %f6044, %f6028, %f5957, %f6043; mul.ftz.f32 %f6045, %f6029, %f5981; fma.rn.ftz.f32 %f6046, %f6028, %f5978, %f6045; mul.ftz.f32 %f6047, %f6029, %f6002; fma.rn.ftz.f32 %f6048, %f6028, %f5979, %f6047; fma.rn.ftz.f32 %f15808, %f6027, %f6005, %f6042; fma.rn.ftz.f32 %f15807, %f6027, %f6004, %f6040; fma.rn.ftz.f32 %f15806, %f6027, %f6003, %f6038; fma.rn.ftz.f32 %f15812, %f6030, %f6005, %f6048; fma.rn.ftz.f32 %f15811, %f6030, %f6004, %f6046; fma.rn.ftz.f32 %f15810, %f6030, %f6003, %f6044; fma.rn.ftz.f32 %f15804, %f6023, %f6005, %f6036; fma.rn.ftz.f32 %f15803, %f6023, %f6004, %f6034; fma.rn.ftz.f32 %f15802, %f6023, %f6003, %f6032; fma.rn.ftz.f32 %f15813, %f781, %f5939, %f5955; fma.rn.ftz.f32 %f15809, %f781, %f5938, %f5954; fma.rn.ftz.f32 %f15805, %f781, %f5937, %f5953; $L__BB108_715: mul.ftz.f32 %f6061, %f15815, %f15803; fma.rn.ftz.f32 %f6062, %f15814, %f15802, %f6061; fma.rn.ftz.f32 %f6063, %f15816, %f15804, %f6062; add.ftz.f32 %f830, %f15805, %f6063; mul.ftz.f32 %f6064, %f15815, %f15807; fma.rn.ftz.f32 %f6065, %f15814, %f15806, %f6064; fma.rn.ftz.f32 %f6066, %f15816, %f15808, %f6065; add.ftz.f32 %f831, %f15809, %f6066; mul.ftz.f32 %f6067, %f15815, %f15811; fma.rn.ftz.f32 %f6068, %f15814, %f15810, %f6067; fma.rn.ftz.f32 %f6069, %f15816, %f15812, %f6068; add.ftz.f32 %f15816, %f15813, %f6069; mul.ftz.f32 %f6070, %f15818, %f15803; fma.rn.ftz.f32 %f6071, %f15817, %f15802, %f6070; fma.rn.ftz.f32 %f6072, %f15819, %f15804, %f6071; add.ftz.f32 %f833, %f15805, %f6072; mul.ftz.f32 %f6073, %f15818, %f15807; fma.rn.ftz.f32 %f6074, %f15817, %f15806, %f6073; fma.rn.ftz.f32 %f6075, %f15819, %f15808, %f6074; add.ftz.f32 %f834, %f15809, %f6075; mul.ftz.f32 %f6076, %f15818, %f15811; fma.rn.ftz.f32 %f6077, %f15817, %f15810, %f6076; fma.rn.ftz.f32 %f6078, %f15819, %f15812, %f6077; add.ftz.f32 %f15819, %f15813, %f6078; mul.ftz.f32 %f6079, %f15821, %f15803; fma.rn.ftz.f32 %f6080, %f15820, %f15802, %f6079; fma.rn.ftz.f32 %f6081, %f15822, %f15804, %f6080; add.ftz.f32 %f836, %f15805, %f6081; mul.ftz.f32 %f6082, %f15821, %f15807; fma.rn.ftz.f32 %f6083, %f15820, %f15806, %f6082; fma.rn.ftz.f32 %f6084, %f15822, %f15808, %f6083; add.ftz.f32 %f837, %f15809, %f6084; mul.ftz.f32 %f6085, %f15821, %f15811; fma.rn.ftz.f32 %f6086, %f15820, %f15810, %f6085; fma.rn.ftz.f32 %f6087, %f15822, %f15812, %f6086; add.ftz.f32 %f15822, %f15813, %f6087; mov.u16 %rs140, 1; mov.f32 %f15814, %f830; mov.f32 %f15815, %f831; mov.f32 %f15817, %f833; mov.f32 %f15818, %f834; mov.f32 %f15820, %f836; mov.f32 %f15821, %f837; $L__BB108_716: sub.ftz.f32 %f6088, %f15817, %f15814; sub.ftz.f32 %f6089, %f15818, %f15815; mul.ftz.f32 %f6090, %f6089, %f6089; fma.rn.ftz.f32 %f6091, %f6088, %f6088, %f6090; sub.ftz.f32 %f6092, %f15819, %f15816; fma.rn.ftz.f32 %f6093, %f6092, %f6092, %f6091; sub.ftz.f32 %f6094, %f15820, %f15814; sub.ftz.f32 %f6095, %f15821, %f15815; mul.ftz.f32 %f6096, %f6095, %f6095; fma.rn.ftz.f32 %f6097, %f6094, %f6094, %f6096; sub.ftz.f32 %f6098, %f15822, %f15816; fma.rn.ftz.f32 %f6099, %f6098, %f6098, %f6097; sub.ftz.f32 %f6100, %f15820, %f15817; sub.ftz.f32 %f6101, %f15821, %f15818; mul.ftz.f32 %f6102, %f6101, %f6101; fma.rn.ftz.f32 %f6103, %f6100, %f6100, %f6102; sub.ftz.f32 %f6104, %f15822, %f15819; fma.rn.ftz.f32 %f6105, %f6104, %f6104, %f6103; max.ftz.f32 %f6106, %f6099, %f6105; max.ftz.f32 %f6107, %f6093, %f6106; mul.ftz.f32 %f6108, %f6089, %f6098; mul.ftz.f32 %f6109, %f6092, %f6095; sub.ftz.f32 %f6110, %f6108, %f6109; mul.ftz.f32 %f6111, %f6092, %f6094; mul.ftz.f32 %f6112, %f6088, %f6098; sub.ftz.f32 %f6113, %f6111, %f6112; mul.ftz.f32 %f6114, %f6088, %f6095; mul.ftz.f32 %f6115, %f6089, %f6094; sub.ftz.f32 %f6116, %f6114, %f6115; ld.local.f32 %f848, [%rd7+-52]; mul.ftz.f32 %f849, %f732, %f848; ld.local.v2.f32 {%f6117, %f6118}, [%rd7+-48]; mul.ftz.f32 %f852, %f732, %f6117; mul.ftz.f32 %f853, %f732, %f6118; mul.ftz.f32 %f6119, %f6113, %f852; fma.rn.ftz.f32 %f6120, %f6110, %f849, %f6119; fma.rn.ftz.f32 %f6121, %f6116, %f853, %f6120; abs.ftz.f32 %f6122, %f6121; mul.ftz.f32 %f6123, %f6113, %f6113; fma.rn.ftz.f32 %f6124, %f6110, %f6110, %f6123; fma.rn.ftz.f32 %f854, %f6116, %f6116, %f6124; div.approx.ftz.f32 %f6125, %f6122, %f854; mul.ftz.f32 %f6126, %f6125, %f6125; setp.gt.ftz.f32 %p610, %f6107, %f6126; @%p610 bra $L__BB108_730; bra.uni $L__BB108_717; $L__BB108_730: ld.local.v4.f32 {%f6392, %f6393, %f6394, %f6395}, [%rd1]; add.ftz.f32 %f6399, %f849, %f6392; add.ftz.f32 %f6400, %f852, %f6393; add.ftz.f32 %f6401, %f853, %f6394; sub.ftz.f32 %f946, %f15814, %f6399; sub.ftz.f32 %f947, %f15815, %f6400; sub.ftz.f32 %f948, %f15816, %f6401; sub.ftz.f32 %f949, %f15817, %f6399; sub.ftz.f32 %f950, %f15818, %f6400; sub.ftz.f32 %f951, %f15819, %f6401; sub.ftz.f32 %f952, %f15820, %f6399; sub.ftz.f32 %f953, %f15821, %f6400; sub.ftz.f32 %f954, %f15822, %f6401; mul.ftz.f32 %f6402, %f947, %f951; mul.ftz.f32 %f6403, %f950, %f948; sub.ftz.f32 %f15849, %f6402, %f6403; mul.ftz.f32 %f6404, %f949, %f948; mul.ftz.f32 %f6405, %f946, %f951; sub.ftz.f32 %f15850, %f6404, %f6405; mul.ftz.f32 %f6406, %f946, %f950; mul.ftz.f32 %f6407, %f949, %f947; sub.ftz.f32 %f15851, %f6406, %f6407; mul.ftz.f32 %f6408, %f15850, %f15850; fma.rn.ftz.f32 %f6409, %f15849, %f15849, %f6408; fma.rn.ftz.f32 %f6410, %f15851, %f15851, %f6409; sqrt.approx.ftz.f32 %f958, %f6410; setp.neu.ftz.f32 %p619, %f958, 0f00000000; @%p619 bra $L__BB108_731; bra.uni $L__BB108_732; $L__BB108_731: rcp.approx.ftz.f32 %f6411, %f958; mul.ftz.f32 %f15851, %f6411, %f15851; mul.ftz.f32 %f15850, %f6411, %f15850; mul.ftz.f32 %f15849, %f6411, %f15849; $L__BB108_732: mul.ftz.f32 %f6412, %f953, %f948; mul.ftz.f32 %f6413, %f947, %f954; sub.ftz.f32 %f15852, %f6413, %f6412; mul.ftz.f32 %f6414, %f946, %f954; mul.ftz.f32 %f6415, %f952, %f948; sub.ftz.f32 %f15853, %f6415, %f6414; mul.ftz.f32 %f6416, %f952, %f947; mul.ftz.f32 %f6417, %f946, %f953; sub.ftz.f32 %f15854, %f6417, %f6416; mul.ftz.f32 %f6418, %f15853, %f15853; fma.rn.ftz.f32 %f6419, %f15852, %f15852, %f6418; fma.rn.ftz.f32 %f6420, %f15854, %f15854, %f6419; sqrt.approx.ftz.f32 %f974, %f6420; setp.neu.ftz.f32 %p620, %f974, 0f00000000; @%p620 bra $L__BB108_733; bra.uni $L__BB108_734; $L__BB108_733: rcp.approx.ftz.f32 %f6421, %f974; mul.ftz.f32 %f15854, %f15854, %f6421; mul.ftz.f32 %f15853, %f15853, %f6421; mul.ftz.f32 %f15852, %f15852, %f6421; $L__BB108_734: mul.ftz.f32 %f6422, %f953, %f951; mul.ftz.f32 %f6423, %f950, %f954; sub.ftz.f32 %f15855, %f6423, %f6422; mul.ftz.f32 %f6424, %f949, %f954; mul.ftz.f32 %f6425, %f952, %f951; sub.ftz.f32 %f15856, %f6425, %f6424; mul.ftz.f32 %f6426, %f952, %f950; mul.ftz.f32 %f6427, %f949, %f953; sub.ftz.f32 %f15857, %f6427, %f6426; mul.ftz.f32 %f6428, %f15856, %f15856; fma.rn.ftz.f32 %f6429, %f15855, %f15855, %f6428; fma.rn.ftz.f32 %f6430, %f15857, %f15857, %f6429; sqrt.approx.ftz.f32 %f990, %f6430; setp.neu.ftz.f32 %p621, %f990, 0f00000000; @%p621 bra $L__BB108_735; bra.uni $L__BB108_736; $L__BB108_735: rcp.approx.ftz.f32 %f6431, %f990; mul.ftz.f32 %f15857, %f15857, %f6431; mul.ftz.f32 %f15856, %f15856, %f6431; mul.ftz.f32 %f15855, %f15855, %f6431; $L__BB108_736: mul.ftz.f32 %f6433, %f15850, %f15853; fma.rn.ftz.f32 %f6434, %f15849, %f15852, %f6433; fma.rn.ftz.f32 %f6435, %f15851, %f15854, %f6434; abs.ftz.f32 %f6436, %f6435; setp.lt.ftz.f32 %p622, %f6436, 0f3F800000; mov.f32 %f6437, 0f3F800000; add.ftz.f32 %f6438, %f6436, 0fBF800000; add.ftz.f32 %f6439, %f6438, 0f3F800000; selp.f32 %f6440, %f6439, 0f3F800000, %p622; sub.ftz.f32 %f6441, %f6437, %f6440; sqrt.approx.ftz.f32 %f6442, %f6441; mul.ftz.f32 %f6443, %f6440, 0f3CB1485D; mov.f32 %f6444, 0f3D9FB43B; sub.ftz.f32 %f6445, %f6444, %f6443; fma.rn.ftz.f32 %f6446, %f6440, %f6445, 0fBE5A6B93; fma.rn.ftz.f32 %f6447, %f6440, %f6446, 0f3FC90FDB; mul.ftz.f32 %f6448, %f6442, %f6447; setp.lt.ftz.f32 %p623, %f6435, 0f00000000; mov.f32 %f15884, 0f00000000; mov.f32 %f6449, 0f40490FDB; sub.ftz.f32 %f6450, %f6449, %f6448; selp.f32 %f6451, %f6450, %f6448, %p623; mul.ftz.f32 %f6452, %f15850, %f15856; fma.rn.ftz.f32 %f6453, %f15849, %f15855, %f6452; fma.rn.ftz.f32 %f6454, %f15851, %f15857, %f6453; neg.ftz.f32 %f6455, %f6454; abs.ftz.f32 %f6456, %f6455; setp.lt.ftz.f32 %p624, %f6456, 0f3F800000; add.ftz.f32 %f6457, %f6456, 0fBF800000; add.ftz.f32 %f6458, %f6457, 0f3F800000; selp.f32 %f6459, %f6458, 0f3F800000, %p624; sub.ftz.f32 %f6460, %f6437, %f6459; sqrt.approx.ftz.f32 %f6461, %f6460; mul.ftz.f32 %f6462, %f6459, 0f3CB1485D; sub.ftz.f32 %f6463, %f6444, %f6462; fma.rn.ftz.f32 %f6464, %f6459, %f6463, 0fBE5A6B93; fma.rn.ftz.f32 %f6465, %f6459, %f6464, 0f3FC90FDB; mul.ftz.f32 %f6466, %f6461, %f6465; setp.gt.ftz.f32 %p625, %f6454, 0f80000000; sub.ftz.f32 %f6467, %f6449, %f6466; selp.f32 %f6468, %f6467, %f6466, %p625; mul.ftz.f32 %f6469, %f15853, %f15856; fma.rn.ftz.f32 %f6470, %f15852, %f15855, %f6469; fma.rn.ftz.f32 %f6471, %f15854, %f15857, %f6470; abs.ftz.f32 %f6472, %f6471; setp.lt.ftz.f32 %p626, %f6472, 0f3F800000; add.ftz.f32 %f6473, %f6472, 0fBF800000; add.ftz.f32 %f6474, %f6473, 0f3F800000; selp.f32 %f6475, %f6474, 0f3F800000, %p626; sub.ftz.f32 %f6476, %f6437, %f6475; sqrt.approx.ftz.f32 %f6477, %f6476; mul.ftz.f32 %f6478, %f6475, 0f3CB1485D; sub.ftz.f32 %f6479, %f6444, %f6478; fma.rn.ftz.f32 %f6480, %f6475, %f6479, 0fBE5A6B93; fma.rn.ftz.f32 %f6481, %f6475, %f6480, 0f3FC90FDB; mul.ftz.f32 %f6482, %f6477, %f6481; setp.lt.ftz.f32 %p627, %f6471, 0f00000000; sub.ftz.f32 %f6483, %f6449, %f6482; selp.f32 %f6484, %f6483, %f6482, %p627; add.ftz.f32 %f6485, %f6451, %f6468; add.ftz.f32 %f6486, %f6485, %f6484; add.ftz.f32 %f1000, %f6486, 0fC0490FDB; setp.eq.ftz.f32 %p628, %f1000, 0f00000000; @%p628 bra $L__BB108_749; setp.eq.s16 %p629, %rs140, 0; @%p629 bra $L__BB108_747; cvt.u32.u64 %r1411, %rd96; cvta.to.global.u64 %rd1200, %rd2955; mul.wide.s32 %rd1201, %r1411, 16; add.s64 %rd1202, %rd1200, %rd1201; ld.const.u64 %rd1203, [kernel_params+2256]; cvta.to.global.u64 %rd1204, %rd1203; ld.global.u32 %r1412, [%rd1202+12]; mul.wide.u32 %rd1205, %r1412, 12; add.s64 %rd1206, %rd1204, %rd1205; ld.global.f32 %f15874, [%rd1206]; ld.global.f32 %f15875, [%rd1206+4]; ld.global.f32 %f15876, [%rd1206+8]; add.s32 %r1413, %r1412, 1; mul.wide.u32 %rd1207, %r1413, 12; add.s64 %rd1208, %rd1204, %rd1207; ld.global.f32 %f15877, [%rd1208]; ld.global.f32 %f15878, [%rd1208+4]; ld.global.f32 %f15879, [%rd1208+8]; add.s32 %r1414, %r1412, 2; mul.wide.u32 %rd1209, %r1414, 12; add.s64 %rd1210, %rd1204, %rd1209; ld.global.f32 %f15880, [%rd1210]; ld.global.f32 %f15881, [%rd1210+4]; ld.global.f32 %f15882, [%rd1210+8]; @%p604 bra $L__BB108_746; and.b32 %r1415, %r361, 2; setp.eq.s32 %p631, %r1415, 0; ld.const.u64 %rd1211, [kernel_params+2160]; cvta.to.global.u64 %rd1212, %rd1211; mul.lo.s64 %rd1213, %rd97, 208; add.s64 %rd163, %rd1212, %rd1213; @%p631 bra $L__BB108_744; bra.uni $L__BB108_740; $L__BB108_744: ld.global.v4.f32 {%f15862, %f15863, %f15864, %f15865}, [%rd163]; ld.global.v4.f32 {%f15866, %f15867, %f15868, %f15869}, [%rd163+16]; ld.global.v4.f32 {%f15870, %f15871, %f15872, %f15873}, [%rd163+32]; bra.uni $L__BB108_745; $L__BB108_717: mul.ftz.f32 %f6128, %f5820, %f6117; fma.rn.ftz.f32 %f6129, %f5819, %f848, %f6128; mov.b32 %r1398, %f5821; mov.b64 %rd1177, {%r1398, %r1399}; cvt.u32.u64 %r1400, %rd1177; mov.b32 %f6130, %r1400; fma.rn.ftz.f32 %f6131, %f6130, %f6118, %f6129; abs.ftz.f32 %f855, %f6131; setp.eq.ftz.f32 %p611, %f855, 0f00000000; mov.f32 %f15823, 0f00000000; @%p611 bra $L__BB108_719; ld.const.f32 %f6132, [kernel_params+1500]; mul.ftz.f32 %f6133, %f732, %f732; mul.ftz.f32 %f6134, %f6133, %f6132; div.approx.ftz.f32 %f15823, %f6134, %f855; $L__BB108_719: setp.eq.s16 %p612, %rs140, 0; mov.f32 %f15884, %f15823; @%p612 bra $L__BB108_749; sqrt.approx.ftz.f32 %f6136, %f854; mul.ftz.f32 %f858, %f6136, 0f3F000000; setp.eq.ftz.f32 %p613, %f858, 0f00000000; mov.f32 %f15884, 0f00000000; @%p613 bra $L__BB108_749; cvt.u32.u64 %r1401, %rd96; cvta.to.global.u64 %rd1178, %rd2955; mul.wide.s32 %rd1179, %r1401, 16; add.s64 %rd1180, %rd1178, %rd1179; ld.const.u64 %rd1181, [kernel_params+2256]; cvta.to.global.u64 %rd1182, %rd1181; ld.global.u32 %r1402, [%rd1180+12]; mul.wide.u32 %rd1183, %r1402, 12; add.s64 %rd1184, %rd1182, %rd1183; ld.global.f32 %f15840, [%rd1184]; ld.global.f32 %f15841, [%rd1184+4]; ld.global.f32 %f15842, [%rd1184+8]; add.s32 %r1403, %r1402, 1; mul.wide.u32 %rd1185, %r1403, 12; add.s64 %rd1186, %rd1182, %rd1185; ld.global.f32 %f15843, [%rd1186]; ld.global.f32 %f15844, [%rd1186+4]; ld.global.f32 %f15845, [%rd1186+8]; add.s32 %r1404, %r1402, 2; mul.wide.u32 %rd1187, %r1404, 12; add.s64 %rd1188, %rd1182, %rd1187; ld.global.f32 %f15846, [%rd1188]; ld.global.f32 %f15847, [%rd1188+4]; ld.global.f32 %f15848, [%rd1188+8]; @%p604 bra $L__BB108_729; and.b32 %r1405, %r361, 2; setp.eq.s32 %p615, %r1405, 0; ld.const.u64 %rd1189, [kernel_params+2160]; cvta.to.global.u64 %rd1190, %rd1189; mul.lo.s64 %rd1191, %rd97, 208; add.s64 %rd161, %rd1190, %rd1191; @%p615 bra $L__BB108_727; bra.uni $L__BB108_723; $L__BB108_727: ld.global.v4.f32 {%f15828, %f15829, %f15830, %f15831}, [%rd161]; ld.global.v4.f32 {%f15832, %f15833, %f15834, %f15835}, [%rd161+16]; ld.global.v4.f32 {%f15836, %f15837, %f15838, %f15839}, [%rd161+32]; bra.uni $L__BB108_728; $L__BB108_747: sqrt.approx.ftz.f32 %f15883, %f854; bra.uni $L__BB108_748; $L__BB108_1349: add.s64 %rd1803, %rd1727, 192; cvta.to.local.u64 %rd1804, %rd1803; ld.local.v4.f32 {%f9988, %f9989, %f9990, %f9991}, [%rd1804]; add.s64 %rd1805, %rd1727, 208; cvta.to.local.u64 %rd1806, %rd1805; ld.local.v4.f32 {%f9995, %f9996, %f9997, %f9998}, [%rd1806]; add.s64 %rd1807, %rd1727, 224; cvta.to.local.u64 %rd1808, %rd1807; ld.local.v4.f32 {%f10002, %f10003, %f10004, %f10005}, [%rd1808]; mul.ftz.f32 %f10009, %f16214, %f9995; fma.rn.ftz.f32 %f10010, %f16215, %f9988, %f10009; fma.rn.ftz.f32 %f10011, %f16213, %f10002, %f10010; mul.ftz.f32 %f10012, %f16214, %f9996; fma.rn.ftz.f32 %f10013, %f16215, %f9989, %f10012; fma.rn.ftz.f32 %f10014, %f16213, %f10003, %f10013; mul.ftz.f32 %f10015, %f16214, %f9997; fma.rn.ftz.f32 %f10016, %f16215, %f9990, %f10015; fma.rn.ftz.f32 %f10017, %f16213, %f10004, %f10016; mul.ftz.f32 %f10018, %f10014, %f10014; fma.rn.ftz.f32 %f10019, %f10011, %f10011, %f10018; fma.rn.ftz.f32 %f10020, %f10017, %f10017, %f10019; sqrt.approx.ftz.f32 %f10021, %f10020; div.approx.ftz.f32 %f16215, %f10011, %f10021; div.approx.ftz.f32 %f16214, %f10014, %f10021; div.approx.ftz.f32 %f16213, %f10017, %f10021; st.local.f32 [%rd297], %f16215; st.local.f32 [%rd287], %f16214; st.local.f32 [%rd299], %f16213; mov.u32 %r3227, %r2023; $L__BB108_1351: cvta.to.global.u64 %rd1813, %rd2982; shl.b64 %rd1814, %rd286, 4; add.s64 %rd1815, %rd1813, %rd1814; ld.const.u64 %rd1816, [kernel_params+2256]; cvta.to.global.u64 %rd1817, %rd1816; ld.global.u32 %r2039, [%rd1815+12]; mul.wide.u32 %rd1818, %r2039, 12; add.s64 %rd1819, %rd1817, %rd1818; add.s32 %r2040, %r2039, 1; mul.wide.u32 %rd1820, %r2040, 12; add.s64 %rd1821, %rd1817, %rd1820; add.s32 %r2041, %r2039, 2; mul.wide.u32 %rd1822, %r2041, 12; add.s64 %rd1823, %rd1817, %rd1822; ld.global.f32 %f10056, [%rd1823]; ld.global.f32 %f10057, [%rd1823+4]; ld.global.f32 %f10058, [%rd1823+8]; ld.global.f32 %f10059, [%rd1821]; ld.global.f32 %f10060, [%rd1819]; sub.ftz.f32 %f16212, %f10059, %f10060; ld.global.f32 %f10061, [%rd1821+4]; ld.global.f32 %f10062, [%rd1819+4]; sub.ftz.f32 %f16211, %f10061, %f10062; ld.global.f32 %f10063, [%rd1821+8]; ld.global.f32 %f10064, [%rd1819+8]; sub.ftz.f32 %f16210, %f10063, %f10064; add.s64 %rd1825, %rd1727, 116; cvta.to.local.u64 %rd303, %rd1825; st.local.f32 [%rd303], %f16212; add.s64 %rd1826, %rd1727, 120; cvta.to.local.u64 %rd304, %rd1826; st.local.f32 [%rd304], %f16211; add.s64 %rd1827, %rd1727, 124; cvta.to.local.u64 %rd305, %rd1827; st.local.f32 [%rd305], %f16210; sub.ftz.f32 %f16209, %f10056, %f10060; sub.ftz.f32 %f16208, %f10057, %f10062; sub.ftz.f32 %f16207, %f10058, %f10064; add.s64 %rd1828, %rd1727, 128; cvta.to.local.u64 %rd306, %rd1828; st.local.v2.f32 [%rd306], {%f16209, %f16208}; add.s64 %rd1829, %rd1727, 136; cvta.to.local.u64 %rd307, %rd1829; st.local.f32 [%rd307], %f16207; setp.ne.s32 %p1271, %r3227, 0; @%p1271 bra $L__BB108_1361; @%p1255 bra $L__BB108_1354; bra.uni $L__BB108_1353; $L__BB108_1354: ld.const.u64 %rd1837, [kernel_params+2160]; cvta.to.global.u64 %rd1838, %rd1837; mul.lo.s64 %rd1839, %rd285, 208; add.s64 %rd1840, %rd1838, %rd1839; ld.global.v4.f32 {%f10092, %f10093, %f10094, %f10095}, [%rd1840]; ld.global.v4.f32 {%f10099, %f10100, %f10101, %f10102}, [%rd1840+16]; ld.global.v4.f32 {%f10106, %f10107, %f10108, %f10109}, [%rd1840+32]; mul.ftz.f32 %f10113, %f10093, %f16211; fma.rn.ftz.f32 %f10114, %f16212, %f10092, %f10113; fma.rn.ftz.f32 %f16206, %f10094, %f16210, %f10114; mul.ftz.f32 %f10115, %f10100, %f16211; fma.rn.ftz.f32 %f10116, %f16212, %f10099, %f10115; fma.rn.ftz.f32 %f16205, %f10101, %f16210, %f10116; mul.ftz.f32 %f10117, %f10107, %f16211; fma.rn.ftz.f32 %f10118, %f16212, %f10106, %f10117; fma.rn.ftz.f32 %f16210, %f10108, %f16210, %f10118; bra.uni $L__BB108_1355; $L__BB108_1353: add.s64 %rd1831, %rd1727, 144; cvta.to.local.u64 %rd1832, %rd1831; ld.local.v4.f32 {%f10065, %f10066, %f10067, %f10068}, [%rd1832]; mul.ftz.f32 %f10072, %f16211, %f10066; fma.rn.ftz.f32 %f10073, %f16212, %f10065, %f10072; fma.rn.ftz.f32 %f16206, %f16210, %f10067, %f10073; add.s64 %rd1833, %rd1727, 160; cvta.to.local.u64 %rd1834, %rd1833; ld.local.v4.f32 {%f10074, %f10075, %f10076, %f10077}, [%rd1834]; mul.ftz.f32 %f10081, %f16211, %f10075; fma.rn.ftz.f32 %f10082, %f16212, %f10074, %f10081; fma.rn.ftz.f32 %f16205, %f16210, %f10076, %f10082; add.s64 %rd1835, %rd1727, 176; cvta.to.local.u64 %rd1836, %rd1835; ld.local.v4.f32 {%f10083, %f10084, %f10085, %f10086}, [%rd1836]; mul.ftz.f32 %f10090, %f16211, %f10084; fma.rn.ftz.f32 %f10091, %f16212, %f10083, %f10090; fma.rn.ftz.f32 %f16210, %f16210, %f10085, %f10091; $L__BB108_1355: st.local.f32 [%rd303], %f16206; st.local.f32 [%rd304], %f16205; st.local.f32 [%rd305], %f16210; @%p1255 bra $L__BB108_1357; bra.uni $L__BB108_1356; $L__BB108_1357: ld.const.u64 %rd1848, [kernel_params+2160]; cvta.to.global.u64 %rd1849, %rd1848; mul.lo.s64 %rd1850, %rd285, 208; add.s64 %rd1851, %rd1849, %rd1850; ld.global.v4.f32 {%f10146, %f10147, %f10148, %f10149}, [%rd1851]; ld.global.v4.f32 {%f10153, %f10154, %f10155, %f10156}, [%rd1851+16]; ld.global.v4.f32 {%f10160, %f10161, %f10162, %f10163}, [%rd1851+32]; mul.ftz.f32 %f10167, %f10147, %f16208; fma.rn.ftz.f32 %f10168, %f16209, %f10146, %f10167; fma.rn.ftz.f32 %f2068, %f10148, %f16207, %f10168; mul.ftz.f32 %f10169, %f10154, %f16208; fma.rn.ftz.f32 %f10170, %f16209, %f10153, %f10169; fma.rn.ftz.f32 %f2069, %f10155, %f16207, %f10170; mul.ftz.f32 %f10171, %f10161, %f16208; fma.rn.ftz.f32 %f10172, %f16209, %f10160, %f10171; fma.rn.ftz.f32 %f16207, %f10162, %f16207, %f10172; st.local.v2.f32 [%rd306], {%f2068, %f2069}; st.local.f32 [%rd307], %f16207; mov.f32 %f16208, %f2069; mov.f32 %f16209, %f2068; mov.f32 %f16211, %f16205; mov.f32 %f16212, %f16206; bra.uni $L__BB108_1361; $L__BB108_1356: add.s64 %rd1842, %rd1727, 144; cvta.to.local.u64 %rd1843, %rd1842; ld.local.v4.f32 {%f10119, %f10120, %f10121, %f10122}, [%rd1843]; mul.ftz.f32 %f10126, %f16208, %f10120; fma.rn.ftz.f32 %f10127, %f16209, %f10119, %f10126; fma.rn.ftz.f32 %f2065, %f16207, %f10121, %f10127; add.s64 %rd1844, %rd1727, 160; cvta.to.local.u64 %rd1845, %rd1844; ld.local.v4.f32 {%f10128, %f10129, %f10130, %f10131}, [%rd1845]; mul.ftz.f32 %f10135, %f16208, %f10129; fma.rn.ftz.f32 %f10136, %f16209, %f10128, %f10135; fma.rn.ftz.f32 %f2066, %f16207, %f10130, %f10136; add.s64 %rd1846, %rd1727, 176; cvta.to.local.u64 %rd1847, %rd1846; ld.local.v4.f32 {%f10137, %f10138, %f10139, %f10140}, [%rd1847]; mul.ftz.f32 %f10144, %f16208, %f10138; fma.rn.ftz.f32 %f10145, %f16209, %f10137, %f10144; fma.rn.ftz.f32 %f16207, %f16207, %f10139, %f10145; st.local.v2.f32 [%rd306], {%f2065, %f2066}; st.local.f32 [%rd307], %f16207; mov.f32 %f16208, %f2066; mov.f32 %f16209, %f2065; mov.f32 %f16211, %f16205; mov.f32 %f16212, %f16206; $L__BB108_1361: mov.b32 %r2043, %f1957; mov.b64 %rd1864, {%r2043, %r2044}; cvt.u32.u64 %r2045, %rd1864; mov.b32 %f16216, %r2045; @%p1252 bra $L__BB108_1364; mul.ftz.f32 %f10179, %f16218, %f1953; fma.rn.ftz.f32 %f10180, %f16217, %f1952, %f10179; fma.rn.ftz.f32 %f10181, %f1957, %f1954, %f10180; setp.geu.ftz.f32 %p1276, %f10181, 0f00000000; @%p1276 bra $L__BB108_1364; or.b32 %r2046, %r620, 1; st.local.u32 [%rd291], %r2046; neg.ftz.f32 %f16216, %f1957; st.local.f32 [%rd1733], %f16216; cvta.to.local.u64 %rd1867, %rd284; neg.ftz.f32 %f10182, %f16215; st.local.f32 [%rd1867], %f10182; neg.ftz.f32 %f10183, %f16213; neg.ftz.f32 %f10184, %f16214; neg.ftz.f32 %f16218, %f16218; neg.ftz.f32 %f16217, %f16217; st.local.v4.f32 [%rd287], {%f10184, %f10183, %f16217, %f16218}; add.s64 %rd1869, %rd1727, 116; add.s64 %rd1870, %rd1727, 120; add.s64 %rd1871, %rd1727, 124; cvta.to.local.u64 %rd1872, %rd1869; neg.ftz.f32 %f10185, %f16212; st.local.f32 [%rd1872], %f10185; cvta.to.local.u64 %rd1873, %rd1870; neg.ftz.f32 %f10186, %f16211; st.local.f32 [%rd1873], %f10186; cvta.to.local.u64 %rd1874, %rd1871; neg.ftz.f32 %f10187, %f16210; st.local.f32 [%rd1874], %f10187; add.s64 %rd1875, %rd1727, 136; cvta.to.local.u64 %rd1876, %rd1875; neg.ftz.f32 %f10188, %f16208; neg.ftz.f32 %f10189, %f16209; add.s64 %rd1877, %rd1727, 128; cvta.to.local.u64 %rd1878, %rd1877; st.local.v2.f32 [%rd1878], {%f10189, %f10188}; neg.ftz.f32 %f10190, %f16207; st.local.f32 [%rd1876], %f10190; $L__BB108_1364: add.s64 %rd1880, %rd1727, 92; cvta.to.local.u64 %rd1881, %rd1880; mov.u32 %r2047, 0; st.local.u32 [%rd1881], %r2047; add.s64 %rd1882, %rd1727, 96; cvta.to.local.u64 %rd1883, %rd1882; mov.f32 %f10191, 0f00000000; st.local.v4.f32 [%rd1883], {%f10191, %f10191, %f10191, %f10191}; add.s64 %rd1884, %rd1727, 112; cvta.to.local.u64 %rd1885, %rd1884; st.local.u32 [%rd1885], %r2047; st.local.f32 [%rd282+-12], %f16217; st.local.f32 [%rd282+-8], %f16218; st.local.f32 [%rd282+-4], %f16216; $L__BB108_1366: add.s64 %rd1918, %rd1727, 260; cvta.to.local.u64 %rd1919, %rd1918; mov.u32 %r2054, 0; st.local.u32 [%rd1919], %r2054; add.s64 %rd1920, %rd1727, 264; cvta.to.local.u64 %rd1921, %rd1920; st.local.u32 [%rd1921], %r2054; ld.const.u64 %rd1922, [kernel_params+2400]; cvta.to.global.u64 %rd308, %rd1922; add.s64 %rd1923, %rd1727, 268; cvta.to.local.u64 %rd312, %rd1923; add.s64 %rd1924, %rd1727, 272; cvta.to.local.u64 %rd313, %rd1924; add.s64 %rd1925, %rd1727, 276; cvta.to.local.u64 %rd314, %rd1925; add.u64 %rd1926, %SP, 704; add.u64 %rd315, %SPL, 704; bra.uni $L__BB108_1367; $L__BB108_1519: { // callseq 780, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 780 $L__BB108_1367: mov.u32 %r633, %r3229; mul.wide.s32 %rd1927, %r633, 16; add.s64 %rd316, %rd308, %rd1927; ld.global.v4.u32 {%r2056, %r2057, %r2058, %r2059}, [%rd316]; mul.wide.u32 %rd1928, %r2057, 4; add.s64 %rd317, %rd315, %rd1928; mul.wide.u32 %rd1929, %r2058, 4; add.s64 %rd318, %rd315, %rd1929; add.s32 %r3229, %r633, 1; setp.gt.s32 %p1277, %r2056, 36; @%p1277 bra $L__BB108_1416; setp.gt.s32 %p1359, %r2056, 25; @%p1359 bra $L__BB108_1412; bra.uni $L__BB108_1369; $L__BB108_1412: add.s32 %r2061, %r2056, -27; setp.lt.u32 %p1360, %r2061, 9; @%p1360 bra $L__BB108_1367; setp.eq.s32 %p1361, %r2056, 26; @%p1361 bra $L__BB108_1558; setp.eq.s32 %p1362, %r2056, 36; @%p1362 bra $L__BB108_1415; bra.uni $L__BB108_1572; $L__BB108_1415: mov.u32 %r2102, 36; { // callseq 833, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2102; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 833 bra.uni $L__BB108_1367; $L__BB108_1416: setp.gt.s32 %p1278, %r2056, 68; @%p1278 bra $L__BB108_1471; bra.uni $L__BB108_1417; $L__BB108_1471: setp.gt.s32 %p1279, %r2056, 88; @%p1279 bra $L__BB108_1507; bra.uni $L__BB108_1472; $L__BB108_1507: setp.gt.s32 %p1280, %r2056, 93; @%p1280 bra $L__BB108_1512; bra.uni $L__BB108_1508; $L__BB108_1512: setp.gt.s32 %p1281, %r2056, 95; @%p1281 bra $L__BB108_1516; bra.uni $L__BB108_1513; $L__BB108_1516: setp.eq.s32 %p1282, %r2056, 98; @%p1282 bra $L__BB108_1571; setp.eq.s32 %p1283, %r2056, 97; @%p1283 bra $L__BB108_1521; setp.ne.s32 %p1284, %r2056, 96; @%p1284 bra $L__BB108_1572; bra.uni $L__BB108_1519; $L__BB108_1417: setp.gt.s32 %p1317, %r2056, 48; @%p1317 bra $L__BB108_1436; bra.uni $L__BB108_1418; $L__BB108_1436: setp.gt.s32 %p1318, %r2056, 58; @%p1318 bra $L__BB108_1454; setp.gt.s32 %p1332, %r2056, 53; @%p1332 bra $L__BB108_1446; setp.gt.s32 %p1339, %r2056, 50; @%p1339 bra $L__BB108_1442; bra.uni $L__BB108_1439; $L__BB108_1442: setp.eq.s32 %p1340, %r2056, 51; @%p1340 bra $L__BB108_1550; setp.eq.s32 %p1341, %r2056, 52; @%p1341 bra $L__BB108_1549; setp.eq.s32 %p1342, %r2056, 53; @%p1342 bra $L__BB108_1445; bra.uni $L__BB108_1572; $L__BB108_1445: { // callseq 818, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 818 bra.uni $L__BB108_1367; $L__BB108_1369: setp.gt.s32 %p1363, %r2056, 17; @%p1363 bra $L__BB108_1403; bra.uni $L__BB108_1370; $L__BB108_1403: setp.lt.s32 %p1364, %r2056, 21; @%p1364 bra $L__BB108_1367; setp.gt.s32 %p1365, %r2056, 22; @%p1365 bra $L__BB108_1408; bra.uni $L__BB108_1405; $L__BB108_1408: setp.eq.s32 %p1366, %r2056, 23; @%p1366 bra $L__BB108_1560; setp.eq.s32 %p1367, %r2056, 24; @%p1367 bra $L__BB108_1559; setp.eq.s32 %p1368, %r2056, 25; @%p1368 bra $L__BB108_1411; bra.uni $L__BB108_1572; $L__BB108_1411: add.s32 %r2104, %r633, 1; { // callseq 835, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2104; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 835 bra.uni $L__BB108_1367; $L__BB108_1472: setp.gt.s32 %p1290, %r2056, 78; @%p1290 bra $L__BB108_1490; setp.gt.s32 %p1304, %r2056, 73; @%p1304 bra $L__BB108_1482; setp.gt.s32 %p1311, %r2056, 70; @%p1311 bra $L__BB108_1478; bra.uni $L__BB108_1475; $L__BB108_1478: setp.eq.s32 %p1312, %r2056, 71; @%p1312 bra $L__BB108_1538; setp.eq.s32 %p1313, %r2056, 72; @%p1313 bra $L__BB108_1537; setp.eq.s32 %p1314, %r2056, 73; @%p1314 bra $L__BB108_1481; bra.uni $L__BB108_1572; $L__BB108_1481: add.s32 %r2074, %r633, 1; { // callseq 798, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2074; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 798 bra.uni $L__BB108_1367; $L__BB108_1418: setp.gt.s32 %p1345, %r2056, 41; @%p1345 bra $L__BB108_1424; setp.gt.s32 %p1355, %r2056, 39; @%p1355 bra $L__BB108_1367; setp.eq.s32 %p1356, %r2056, 37; @%p1356 bra $L__BB108_1557; setp.eq.s32 %p1357, %r2056, 38; @%p1357 bra $L__BB108_1556; setp.eq.s32 %p1358, %r2056, 39; @%p1358 bra $L__BB108_1423; bra.uni $L__BB108_1572; $L__BB108_1423: mov.u32 %r2100, 39; { // callseq 830, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2100, %r2057, %r2058, %r2059}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 830 bra.uni $L__BB108_1367; $L__BB108_1370: setp.gt.s32 %p1371, %r2056, 8; @%p1371 bra $L__BB108_1386; bra.uni $L__BB108_1371; $L__BB108_1386: setp.gt.s32 %p1372, %r2056, 12; @%p1372 bra $L__BB108_1395; bra.uni $L__BB108_1387; $L__BB108_1395: setp.gt.s32 %p1373, %r2056, 14; @%p1373 bra $L__BB108_1399; bra.uni $L__BB108_1396; $L__BB108_1399: setp.eq.s32 %p1374, %r2056, 15; @%p1374 bra $L__BB108_1563; setp.eq.s32 %p1375, %r2056, 16; @%p1375 bra $L__BB108_1562; setp.eq.s32 %p1376, %r2056, 17; @%p1376 bra $L__BB108_1402; bra.uni $L__BB108_1572; $L__BB108_1402: { // callseq 840, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 840 bra.uni $L__BB108_1367; $L__BB108_1454: setp.gt.s32 %p1319, %r2056, 63; @%p1319 bra $L__BB108_1463; setp.gt.s32 %p1326, %r2056, 60; @%p1326 bra $L__BB108_1459; bra.uni $L__BB108_1456; $L__BB108_1459: setp.eq.s32 %p1327, %r2056, 61; @%p1327 bra $L__BB108_1544; setp.eq.s32 %p1328, %r2056, 62; @%p1328 bra $L__BB108_1543; setp.eq.s32 %p1329, %r2056, 63; @%p1329 bra $L__BB108_1462; bra.uni $L__BB108_1572; $L__BB108_1462: mov.u32 %r2084, 63; { // callseq 808, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2084; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 808 bra.uni $L__BB108_1367; $L__BB108_1490: setp.gt.s32 %p1291, %r2056, 83; @%p1291 bra $L__BB108_1499; setp.gt.s32 %p1298, %r2056, 80; @%p1298 bra $L__BB108_1495; bra.uni $L__BB108_1492; $L__BB108_1495: setp.eq.s32 %p1299, %r2056, 81; @%p1299 bra $L__BB108_1526; setp.eq.s32 %p1300, %r2056, 82; @%p1300 bra $L__BB108_1525; setp.eq.s32 %p1301, %r2056, 83; @%p1301 bra $L__BB108_1498; bra.uni $L__BB108_1572; $L__BB108_1498: { // callseq 790, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 790 bra.uni $L__BB108_1367; $L__BB108_1371: setp.gt.s32 %p1384, %r2056, 4; @%p1384 bra $L__BB108_1379; setp.gt.s32 %p1390, %r2056, 2; @%p1390 bra $L__BB108_1376; setp.eq.s32 %p1393, %r2056, 1; mov.u32 %r3229, %r2057; @%p1393 bra $L__BB108_1367; setp.eq.s32 %p1394, %r2056, 2; @%p1394 bra $L__BB108_1375; bra.uni $L__BB108_1572; $L__BB108_1375: mov.u32 %r2128, 2; add.s32 %r2129, %r633, 1; { // callseq 849, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r2128; st.param.b32 [param0+4], %r2057; st.param.b32 [param0+8], %r2058; st.param.b32 [param0+12], %r2059; .param .b32 param1; st.param.b32 [param1+0], %r2129; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 849 bra.uni $L__BB108_1367; $L__BB108_1424: setp.gt.s32 %p1346, %r2056, 44; @%p1346 bra $L__BB108_1429; bra.uni $L__BB108_1425; $L__BB108_1429: setp.gt.s32 %p1347, %r2056, 46; @%p1347 bra $L__BB108_1433; setp.eq.s32 %p1350, %r2056, 45; @%p1350 bra $L__BB108_1553; setp.eq.s32 %p1351, %r2056, 46; @%p1351 bra $L__BB108_1432; bra.uni $L__BB108_1572; $L__BB108_1432: { // callseq 825, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 825 bra.uni $L__BB108_1367; $L__BB108_1513: setp.eq.s32 %p1285, %r2056, 94; @%p1285 bra $L__BB108_1520; setp.eq.s32 %p1286, %r2056, 95; @%p1286 bra $L__BB108_1515; bra.uni $L__BB108_1572; $L__BB108_1515: { // callseq 781, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 781 bra.uni $L__BB108_1367; $L__BB108_1508: add.s32 %r2060, %r2056, -92; setp.lt.u32 %p1287, %r2060, 2; @%p1287 bra $L__BB108_1367; setp.eq.s32 %p1288, %r2056, 90; @%p1288 bra $L__BB108_1367; setp.eq.s32 %p1289, %r2056, 89; @%p1289 bra $L__BB108_1511; bra.uni $L__BB108_1572; $L__BB108_1511: mov.u32 %r2064, 89; { // callseq 784, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2064; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 784 bra.uni $L__BB108_1367; $L__BB108_1446: setp.gt.s32 %p1333, %r2056, 55; @%p1333 bra $L__BB108_1450; bra.uni $L__BB108_1447; $L__BB108_1450: setp.eq.s32 %p1334, %r2056, 56; @%p1334 bra $L__BB108_1547; setp.eq.s32 %p1335, %r2056, 57; @%p1335 bra $L__BB108_1546; setp.eq.s32 %p1336, %r2056, 58; @%p1336 bra $L__BB108_1453; bra.uni $L__BB108_1572; $L__BB108_1453: mov.u32 %r2091, 58; { // callseq 813, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2091; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 813 bra.uni $L__BB108_1367; $L__BB108_1482: setp.gt.s32 %p1305, %r2056, 75; @%p1305 bra $L__BB108_1486; bra.uni $L__BB108_1483; $L__BB108_1486: setp.eq.s32 %p1306, %r2056, 76; @%p1306 bra $L__BB108_1530; setp.eq.s32 %p1307, %r2056, 77; @%p1307 bra $L__BB108_1528; setp.eq.s32 %p1308, %r2056, 78; @%p1308 bra $L__BB108_1489; bra.uni $L__BB108_1572; $L__BB108_1489: add.s32 %r2072, %r633, 1; { // callseq 795, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2072; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 795 bra.uni $L__BB108_1367; $L__BB108_1463: setp.gt.s32 %p1320, %r2056, 65; @%p1320 bra $L__BB108_1467; bra.uni $L__BB108_1464; $L__BB108_1467: setp.eq.s32 %p1321, %r2056, 66; @%p1321 bra $L__BB108_1541; setp.eq.s32 %p1322, %r2056, 67; @%p1322 bra $L__BB108_1540; setp.eq.s32 %p1323, %r2056, 68; @%p1323 bra $L__BB108_1470; bra.uni $L__BB108_1572; $L__BB108_1470: mov.u32 %r2079, 68; { // callseq 803, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2079; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 803 bra.uni $L__BB108_1367; $L__BB108_1499: setp.gt.s32 %p1292, %r2056, 85; @%p1292 bra $L__BB108_1503; bra.uni $L__BB108_1500; $L__BB108_1503: setp.eq.s32 %p1293, %r2056, 86; @%p1293 bra $L__BB108_1523; setp.eq.s32 %p1294, %r2056, 87; @%p1294 bra $L__BB108_1522; setp.eq.s32 %p1295, %r2056, 88; @%p1295 bra $L__BB108_1506; bra.uni $L__BB108_1572; $L__BB108_1506: mov.u32 %r2065, 88; cvta.to.local.u64 %rd1936, %rd1727; ld.local.f32 %f10193, [%rd1936+12]; ld.local.v2.f32 {%f10194, %f10195}, [%rd1936+16]; { // callseq 785, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f10193; .param .b32 param1; st.param.f32 [param1+0], %f10194; .param .b32 param2; st.param.f32 [param2+0], %f10195; .param .b64 param3; st.param.b64 [param3+0], %rd1926; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r2065; st.param.b32 [param4+4], %r2057; st.param.b32 [param4+8], %r2058; st.param.b32 [param4+12], %r2059; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 785 bra.uni $L__BB108_1367; $L__BB108_1387: setp.gt.s32 %p1379, %r2056, 10; @%p1379 bra $L__BB108_1392; setp.eq.s32 %p1382, %r2056, 9; @%p1382 bra $L__BB108_1566; setp.eq.s32 %p1383, %r2056, 10; @%p1383 bra $L__BB108_1390; bra.uni $L__BB108_1572; $L__BB108_1390: ld.local.f32 %f10209, [%rd318]; setp.ltu.ftz.f32 %p1399, %f10209, 0f3F800000; @%p1399 bra $L__BB108_1367; add.s32 %r2122, %r2057, %r633; add.s32 %r3229, %r2122, 1; bra.uni $L__BB108_1367; $L__BB108_1379: setp.gt.s32 %p1385, %r2056, 6; @%p1385 bra $L__BB108_1383; setp.eq.s32 %p1388, %r2056, 5; @%p1388 bra $L__BB108_1569; setp.eq.s32 %p1389, %r2056, 6; @%p1389 bra $L__BB108_1382; bra.uni $L__BB108_1572; $L__BB108_1382: ld.local.f32 %f10211, [%rd317]; ld.local.f32 %f10212, [%rd317+4]; ld.local.f32 %f10213, [%rd317+8]; st.local.f32 [%rd312], %f10211; st.local.f32 [%rd313], %f10212; st.local.f32 [%rd314], %f10213; bra.uni $L__BB108_1367; $L__BB108_1425: setp.eq.s32 %p1352, %r2056, 42; @%p1352 bra $L__BB108_1555; setp.eq.s32 %p1353, %r2056, 43; @%p1353 bra $L__BB108_1554; setp.eq.s32 %p1354, %r2056, 44; @%p1354 bra $L__BB108_1428; bra.uni $L__BB108_1572; $L__BB108_1428: mov.u32 %r2097, 44; add.s32 %r2098, %r633, 1; { // callseq 827, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2097; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2098; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 827 bra.uni $L__BB108_1367; $L__BB108_1571: { // callseq 850, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 850 bra.uni $L__BB108_1367; $L__BB108_1521: { // callseq 783, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 783 bra.uni $L__BB108_1367; $L__BB108_1520: mov.u32 %r2062, 94; add.s32 %r2063, %r633, 1; { // callseq 782, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2062; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2063; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 782 bra.uni $L__BB108_1367; $L__BB108_1405: setp.eq.s32 %p1369, %r2056, 21; @%p1369 bra $L__BB108_1561; setp.eq.s32 %p1370, %r2056, 22; @%p1370 bra $L__BB108_1407; bra.uni $L__BB108_1572; $L__BB108_1407: mov.u32 %r2108, 22; add.s32 %r2109, %r633, 1; { // callseq 838, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2108; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2109; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 838 bra.uni $L__BB108_1367; $L__BB108_1396: setp.eq.s32 %p1377, %r2056, 13; @%p1377 bra $L__BB108_1564; setp.eq.s32 %p1378, %r2056, 14; @%p1378 bra $L__BB108_1398; bra.uni $L__BB108_1572; $L__BB108_1398: st.local.u32 [%rd318], %r2057; bra.uni $L__BB108_1367; $L__BB108_1376: setp.eq.s32 %p1391, %r2056, 3; @%p1391 bra $L__BB108_1570; setp.eq.s32 %p1392, %r2056, 4; @%p1392 bra $L__BB108_1378; bra.uni $L__BB108_1572; $L__BB108_1378: mov.u32 %r2126, 4; { // callseq 847, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2126, %r2057, %r2058, %r2059}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 847 bra.uni $L__BB108_1367; $L__BB108_1392: setp.eq.s32 %p1380, %r2056, 11; @%p1380 bra $L__BB108_1565; setp.eq.s32 %p1381, %r2056, 12; @%p1381 bra $L__BB108_1394; bra.uni $L__BB108_1572; $L__BB108_1394: { // callseq 843, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 843 bra.uni $L__BB108_1367; $L__BB108_1383: setp.eq.s32 %p1386, %r2056, 7; @%p1386 bra $L__BB108_1568; setp.eq.s32 %p1387, %r2056, 8; @%p1387 bra $L__BB108_1385; bra.uni $L__BB108_1572; $L__BB108_1385: mov.u32 %r2124, 8; { // callseq 845, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2124; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 845 bra.uni $L__BB108_1367; $L__BB108_1439: setp.eq.s32 %p1343, %r2056, 49; @%p1343 bra $L__BB108_1551; setp.eq.s32 %p1344, %r2056, 50; @%p1344 bra $L__BB108_1441; bra.uni $L__BB108_1572; $L__BB108_1441: { // callseq 821, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 821 bra.uni $L__BB108_1367; $L__BB108_1475: setp.eq.s32 %p1315, %r2056, 69; @%p1315 bra $L__BB108_1539; setp.eq.s32 %p1316, %r2056, 70; @%p1316 bra $L__BB108_1477; bra.uni $L__BB108_1572; $L__BB108_1477: mov.u32 %r2076, 70; { // callseq 801, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2076; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 801 bra.uni $L__BB108_1367; $L__BB108_1456: setp.eq.s32 %p1330, %r2056, 59; @%p1330 bra $L__BB108_1545; setp.eq.s32 %p1331, %r2056, 60; @%p1331 bra $L__BB108_1458; bra.uni $L__BB108_1572; $L__BB108_1458: add.s32 %r2089, %r633, 1; { // callseq 811, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2089; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 811 bra.uni $L__BB108_1367; $L__BB108_1492: setp.eq.s32 %p1302, %r2056, 79; @%p1302 bra $L__BB108_1527; setp.eq.s32 %p1303, %r2056, 80; @%p1303 bra $L__BB108_1494; bra.uni $L__BB108_1572; $L__BB108_1494: { // callseq 793, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 793 bra.uni $L__BB108_1367; $L__BB108_1433: setp.eq.s32 %p1348, %r2056, 47; @%p1348 bra $L__BB108_1552; setp.eq.s32 %p1349, %r2056, 48; @%p1349 bra $L__BB108_1435; bra.uni $L__BB108_1572; $L__BB108_1435: { // callseq 823, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 823 bra.uni $L__BB108_1367; $L__BB108_1447: setp.eq.s32 %p1337, %r2056, 54; @%p1337 bra $L__BB108_1548; setp.eq.s32 %p1338, %r2056, 55; @%p1338 bra $L__BB108_1449; bra.uni $L__BB108_1572; $L__BB108_1449: cvta.to.local.u64 %rd1981, %rd1727; ld.local.v4.f32 {%f10202, %f10203, %f10204, %f10205}, [%rd1981]; { // callseq 816, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f10202; st.param.f32 [param0+4], %f10203; st.param.f32 [param0+8], %f10204; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 816 bra.uni $L__BB108_1367; $L__BB108_1483: setp.eq.s32 %p1309, %r2056, 74; @%p1309 bra $L__BB108_1536; setp.eq.s32 %p1310, %r2056, 75; @%p1310 bra $L__BB108_1485; bra.uni $L__BB108_1572; $L__BB108_1485: { // callseq 796, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 796 bra.uni $L__BB108_1367; $L__BB108_1464: setp.eq.s32 %p1324, %r2056, 64; @%p1324 bra $L__BB108_1542; setp.eq.s32 %p1325, %r2056, 65; @%p1325 bra $L__BB108_1466; bra.uni $L__BB108_1572; $L__BB108_1466: { // callseq 806, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 806 bra.uni $L__BB108_1367; $L__BB108_1500: setp.eq.s32 %p1296, %r2056, 84; @%p1296 bra $L__BB108_1524; setp.eq.s32 %p1297, %r2056, 85; @%p1297 bra $L__BB108_1502; bra.uni $L__BB108_1572; $L__BB108_1502: add.s32 %r2068, %r633, 1; { // callseq 788, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2068; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 788 bra.uni $L__BB108_1367; $L__BB108_1560: mov.u32 %r2106, 23; add.s32 %r2107, %r633, 1; { // callseq 837, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2106; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2107; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 837 bra.uni $L__BB108_1367; $L__BB108_1559: mov.u32 %r2105, 24; { // callseq 836, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2105; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 836 bra.uni $L__BB108_1367; $L__BB108_1563: ld.global.v4.u32 {%r2112, %r2113, %r2114, %r2115}, [%rd316+16]; add.s32 %r3229, %r633, 2; st.local.u32 [%rd317], %r2113; st.local.u32 [%rd317+4], %r2114; st.local.u32 [%rd317+8], %r2115; bra.uni $L__BB108_1367; $L__BB108_1562: mov.u32 %r2111, 16; { // callseq 841, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2111; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 841 bra.uni $L__BB108_1367; $L__BB108_1566: ld.local.f32 %f10210, [%rd318]; setp.gtu.ftz.f32 %p1400, %f10210, 0f00000000; @%p1400 bra $L__BB108_1367; add.s32 %r2123, %r2057, %r633; add.s32 %r3229, %r2123, 1; bra.uni $L__BB108_1367; $L__BB108_1569: st.local.u32 [%rd312], %r2057; st.local.u32 [%rd313], %r2058; st.local.u32 [%rd314], %r2059; bra.uni $L__BB108_1367; $L__BB108_1561: mov.u32 %r2110, 21; { // callseq 839, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2110; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 839 bra.uni $L__BB108_1367; $L__BB108_1564: mov.u32 %r2119, 13; add.s32 %r2120, %r633, 1; mov.u32 %r2121, 524288; { // callseq 842, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b32 param1; st.param.b32 [param1+0], %r2121; .param .b64 param2; st.param.b64 [param2+0], %rd1926; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r2119; st.param.b32 [param3+4], %r2057; st.param.b32 [param3+8], %r2058; st.param.b32 [param3+12], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2120; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 842 bra.uni $L__BB108_1367; $L__BB108_1570: mov.u32 %r2127, 3; { // callseq 848, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2127, %r2057, %r2058, %r2059}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 848 bra.uni $L__BB108_1367; $L__BB108_1565: { // callseq 844, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 844 bra.uni $L__BB108_1367; $L__BB108_1568: mov.u32 %r2125, 7; { // callseq 846, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2125; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 846 bra.uni $L__BB108_1367; $L__BB108_1550: { // callseq 820, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 820 bra.uni $L__BB108_1367; $L__BB108_1549: add.s32 %r2096, %r633, 1; { // callseq 819, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2096; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 819 bra.uni $L__BB108_1367; $L__BB108_1538: mov.u32 %r2075, 71; { // callseq 800, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2075; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 800 bra.uni $L__BB108_1367; $L__BB108_1537: { // callseq 799, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 799 bra.uni $L__BB108_1367; $L__BB108_1557: mov.u32 %r2101, 37; { // callseq 832, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2101, %r2057, %r2058, %r2059}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 832 bra.uni $L__BB108_1367; $L__BB108_1556: { // callseq 831, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; .param .b32 param4; st.param.b32 [param4+0], %r2059; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 831 bra.uni $L__BB108_1367; $L__BB108_1544: mov.u32 %r2087, 61; add.s32 %r2088, %r633, 1; { // callseq 810, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2087; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2088; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 810 bra.uni $L__BB108_1367; $L__BB108_1543: mov.u32 %r2085, 62; add.s32 %r2086, %r633, 1; { // callseq 809, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2085; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2086; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 809 bra.uni $L__BB108_1367; $L__BB108_1526: mov.u32 %r2070, 81; { // callseq 792, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2070; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 792 bra.uni $L__BB108_1367; $L__BB108_1525: mov.u32 %r2069, 82; { // callseq 791, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r2069; st.param.b32 [param2+4], %r2057; st.param.b32 [param2+8], %r2058; st.param.b32 [param2+12], %r2059; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 791 bra.uni $L__BB108_1367; $L__BB108_1553: { // callseq 826, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 826 bra.uni $L__BB108_1367; $L__BB108_1547: mov.u32 %r2094, 56; { // callseq 815, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2094; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 815 bra.uni $L__BB108_1367; $L__BB108_1546: mov.u32 %r2092, 57; add.s32 %r2093, %r633, 1; { // callseq 814, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2092; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2093; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 814 bra.uni $L__BB108_1367; $L__BB108_1530: setp.eq.s32 %p1396, %r2059, 255; @%p1396 bra $L__BB108_1367; add.s64 %rd319, %rd317, 4; add.s64 %rd320, %rd317, 8; setp.eq.s32 %p1397, %r2058, 0; mul.wide.u32 %rd1956, %r2059, 4; add.s64 %rd321, %rd315, %rd1956; @%p1397 bra $L__BB108_1535; setp.eq.s32 %p1398, %r2058, 1; @%p1398 bra $L__BB108_1534; bra.uni $L__BB108_1533; $L__BB108_1534: ld.local.f32 %f10200, [%rd319]; st.local.f32 [%rd321], %f10200; bra.uni $L__BB108_1367; $L__BB108_1528: setp.eq.s32 %p1395, %r2059, 255; @%p1395 bra $L__BB108_1367; ld.local.f32 %f10198, [%rd317]; add.s32 %r2073, %r2059, %r2058; mul.wide.u32 %rd1952, %r2073, 4; add.s64 %rd1953, %rd315, %rd1952; st.local.f32 [%rd1953], %f10198; bra.uni $L__BB108_1367; $L__BB108_1541: add.s32 %r2081, %r633, 1; { // callseq 805, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2081; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 805 bra.uni $L__BB108_1367; $L__BB108_1540: mov.u32 %r2080, 67; { // callseq 804, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r2080, %r2057, %r2058, %r2059}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 804 bra.uni $L__BB108_1367; $L__BB108_1523: add.s32 %r2067, %r633, 1; { // callseq 787, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2067; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 787 bra.uni $L__BB108_1367; $L__BB108_1522: add.s32 %r2066, %r633, 1; { // callseq 786, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2066; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 786 bra.uni $L__BB108_1367; $L__BB108_1555: { // callseq 829, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 829 bra.uni $L__BB108_1367; $L__BB108_1554: add.s32 %r2099, %r633, 1; { // callseq 828, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2099; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 828 bra.uni $L__BB108_1367; $L__BB108_1551: { // callseq 822, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1727; .param .b64 param1; st.param.b64 [param1+0], %rd1926; .param .b32 param2; st.param.b32 [param2+0], %r2057; .param .b32 param3; st.param.b32 [param3+0], %r2058; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 822 bra.uni $L__BB108_1367; $L__BB108_1539: mov.u32 %r2077, 69; add.s32 %r2078, %r633, 1; { // callseq 802, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2077; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2078; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 802 bra.uni $L__BB108_1367; $L__BB108_1545: add.s32 %r2090, %r633, 1; { // callseq 812, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2090; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 812 bra.uni $L__BB108_1367; $L__BB108_1527: add.s32 %r2071, %r633, 1; { // callseq 794, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; .param .b32 param4; st.param.b32 [param4+0], %r2071; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 794 bra.uni $L__BB108_1367; $L__BB108_1552: { // callseq 824, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r953; .param .b64 param1; st.param.b64 [param1+0], %rd1727; .param .b64 param2; st.param.b64 [param2+0], %rd1926; .param .b32 param3; st.param.b32 [param3+0], %r2057; .param .b32 param4; st.param.b32 [param4+0], %r2058; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 824 bra.uni $L__BB108_1367; $L__BB108_1548: add.s32 %r2095, %r633, 1; { // callseq 817, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2095; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 817 bra.uni $L__BB108_1367; $L__BB108_1536: { // callseq 797, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; .param .b32 param3; st.param.b32 [param3+0], %r2059; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 797 bra.uni $L__BB108_1367; $L__BB108_1542: mov.u32 %r2082, 64; add.s32 %r2083, %r633, 1; { // callseq 807, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2082; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; .param .b32 param2; st.param.b32 [param2+0], %r2083; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r3229, [retval0+0]; } // callseq 807 bra.uni $L__BB108_1367; $L__BB108_1524: { // callseq 789, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .b32 param1; st.param.b32 [param1+0], %r2057; .param .b32 param2; st.param.b32 [param2+0], %r2058; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 789 bra.uni $L__BB108_1367; $L__BB108_1558: mov.u32 %r2103, 26; { // callseq 834, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd1926; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r2103; st.param.b32 [param1+4], %r2057; st.param.b32 [param1+8], %r2058; st.param.b32 [param1+12], %r2059; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 834 bra.uni $L__BB108_1367; $L__BB108_1535: ld.local.f32 %f10201, [%rd317]; st.local.f32 [%rd321], %f10201; bra.uni $L__BB108_1367; $L__BB108_1533: ld.local.f32 %f10199, [%rd320]; st.local.f32 [%rd321], %f10199; bra.uni $L__BB108_1367; $L__BB108_1572: add.s64 %rd2033, %rd1727, 52; cvta.to.local.u64 %rd2034, %rd2033; ld.local.u32 %r2130, [%rd2034]; and.b32 %r723, %r2130, 2; @%p1251 bra $L__BB108_1575; bra.uni $L__BB108_1573; $L__BB108_1575: setp.eq.s32 %p1404, %r723, 0; mov.f32 %f16225, 0f00000000; mov.f32 %f16226, %f16225; mov.f32 %f16227, %f16225; @%p1404 bra $L__BB108_1577; add.s64 %rd2045, %rd1727, 280; cvta.to.local.u64 %rd2046, %rd2045; ld.local.v2.f32 {%f16225, %f16226}, [%rd2046]; add.s64 %rd2047, %rd1727, 288; cvta.to.local.u64 %rd2048, %rd2047; ld.local.f32 %f16227, [%rd2048]; bra.uni $L__BB108_1577; $L__BB108_1573: setp.eq.s32 %p1402, %r723, 0; mov.f32 %f16225, 0f00000000; mov.f32 %f16226, %f16225; mov.f32 %f16227, %f16225; @%p1402 bra $L__BB108_1577; add.s64 %rd2035, %rd284, 20; add.s64 %rd2036, %rd284, 12; cvta.to.local.u64 %rd2037, %rd2036; ld.local.v2.f32 {%f10217, %f10218}, [%rd2037]; cvta.to.local.u64 %rd2038, %rd2035; ld.local.v4.f32 {%f10221, %f10222, %f10223, %f10224}, [%rd2038]; mul.ftz.f32 %f10229, %f10218, %f10223; fma.rn.ftz.f32 %f10230, %f10217, %f10222, %f10229; fma.rn.ftz.f32 %f10231, %f10221, %f10224, %f10230; abs.ftz.f32 %f10232, %f10231; setp.gt.ftz.f32 %p1403, %f10232, 0f00000000; selp.f32 %f10233, 0f3F800000, 0f00000000, %p1403; add.s64 %rd2040, %rd1727, 280; cvta.to.local.u64 %rd2041, %rd2040; ld.local.v2.f32 {%f10234, %f10235}, [%rd2041]; add.s64 %rd2042, %rd1727, 288; cvta.to.local.u64 %rd2043, %rd2042; ld.local.f32 %f10238, [%rd2043]; mul.ftz.f32 %f16227, %f10233, %f10238; mul.ftz.f32 %f16226, %f10233, %f10235; mul.ftz.f32 %f16225, %f10234, %f10233; $L__BB108_1577: ld.local.f32 %f10244, [%rd282+28]; mul.ftz.f32 %f16228, %f16225, %f10244; mul.ftz.f32 %f16229, %f16226, %f10244; mul.ftz.f32 %f16230, %f16227, %f10244; ld.local.u32 %r724, [%rd282+44]; setp.eq.s32 %p1405, %r724, -1; @%p1405 bra $L__BB108_1579; ld.const.u64 %rd2049, [kernel_params+2368]; cvta.to.global.u64 %rd2050, %rd2049; mul.wide.s32 %rd2051, %r724, 192; add.s64 %rd2052, %rd2050, %rd2051; ld.global.f32 %f10245, [%rd2052+28]; mul.ftz.f32 %f16228, %f16228, %f10245; ld.global.f32 %f10246, [%rd2052+32]; mul.ftz.f32 %f16229, %f16229, %f10246; ld.global.f32 %f10247, [%rd2052+36]; mul.ftz.f32 %f16230, %f16230, %f10247; $L__BB108_1579: setp.eq.ftz.f32 %p1406, %f16229, 0f00000000; setp.eq.ftz.f32 %p1407, %f16228, 0f00000000; and.pred %p1408, %p1407, %p1406; setp.eq.ftz.f32 %p1409, %f16230, 0f00000000; and.pred %p1410, %p1409, %p1408; @%p1410 bra $L__BB108_1685; add.u64 %rd2919, %SP, 1728; add.u64 %rd2054, %SP, 704; { // callseq 851, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2919; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f1935; st.param.f32 [param1+4], %f1936; st.param.f32 [param1+8], %f1937; .param .b64 param2; st.param.b64 [param2+0], %rd2054; .param .b32 param3; st.param.b32 [param3+0], %r618; .param .b32 retval0; call.uni (retval0), _Z24surface_shader_bsdf_evalPK16KernelGlobalsGPUiP10ShaderData6float3P8BsdfEvalj, ( param0, param1, param2, param3 ); ld.param.f32 %f2119, [retval0+0]; } // callseq 851 ld.local.f32 %f2120, [%rd282+24]; div.approx.ftz.f32 %f10248, %f16228, %f2120; div.approx.ftz.f32 %f10249, %f16229, %f2120; div.approx.ftz.f32 %f10250, %f16230, %f2120; add.u64 %rd322, %SPL, 704; ld.local.f32 %f10251, [%rd322]; mul.ftz.f32 %f16244, %f10248, %f10251; ld.local.f32 %f10252, [%rd322+4]; mul.ftz.f32 %f16246, %f10249, %f10252; ld.local.f32 %f10253, [%rd322+8]; mul.ftz.f32 %f16245, %f10250, %f10253; st.local.f32 [%rd322], %f16244; st.local.f32 [%rd322+4], %f16246; st.local.f32 [%rd322+8], %f16245; ld.local.f32 %f10254, [%rd322+12]; mul.ftz.f32 %f16241, %f10248, %f10254; ld.local.f32 %f10255, [%rd322+16]; mul.ftz.f32 %f16243, %f10249, %f10255; ld.local.f32 %f10256, [%rd322+20]; mul.ftz.f32 %f16242, %f10250, %f10256; st.local.f32 [%rd322+12], %f16241; st.local.f32 [%rd322+16], %f16243; st.local.f32 [%rd322+20], %f16242; ld.local.f32 %f10257, [%rd322+24]; mul.ftz.f32 %f16249, %f10248, %f10257; ld.local.f32 %f10258, [%rd322+28]; mul.ftz.f32 %f16248, %f10249, %f10258; ld.local.f32 %f10259, [%rd322+32]; mul.ftz.f32 %f16247, %f10250, %f10259; st.local.f32 [%rd322+24], %f16249; st.local.f32 [%rd322+28], %f16248; st.local.f32 [%rd322+32], %f16247; and.b32 %r2131, %r618, 268435456; setp.eq.s32 %p1411, %r2131, 0; @%p1411 bra $L__BB108_1582; mul.ftz.f32 %f10260, %f2120, %f2120; fma.rn.ftz.f32 %f10261, %f2119, %f2119, %f10260; div.approx.ftz.f32 %f10262, %f10260, %f10261; mul.ftz.f32 %f16244, %f10262, %f16244; mul.ftz.f32 %f16246, %f10262, %f16246; mul.ftz.f32 %f16245, %f10262, %f16245; st.local.f32 [%rd322], %f16244; st.local.f32 [%rd322+4], %f16246; st.local.f32 [%rd322+8], %f16245; mul.ftz.f32 %f16241, %f10262, %f16241; mul.ftz.f32 %f16243, %f10262, %f16243; mul.ftz.f32 %f16242, %f10262, %f16242; st.local.f32 [%rd322+12], %f16241; st.local.f32 [%rd322+16], %f16243; st.local.f32 [%rd322+20], %f16242; mul.ftz.f32 %f16249, %f10262, %f16249; mul.ftz.f32 %f16248, %f10262, %f16248; mul.ftz.f32 %f16247, %f10262, %f16247; st.local.f32 [%rd322+24], %f16249; st.local.f32 [%rd322+28], %f16248; st.local.f32 [%rd322+32], %f16247; $L__BB108_1582: ld.const.f32 %f2148, [kernel_params+1508]; setp.leu.ftz.f32 %p1412, %f2148, 0f00000000; mov.f32 %f16240, 0f00000000; @%p1412 bra $L__BB108_1587; add.s32 %r725, %r602, 2; @%p1240 bra $L__BB108_1586; ld.const.v2.u32 {%r2132, %r2133}, [kernel_params+1600]; mov.b32 %f2149, %r2133; setp.geu.ftz.f32 %p1414, %f2149, 0f3F800000; setp.lt.ftz.f32 %p1415, %f2149, 0f3F800000; ld.const.u32 %r2136, [kernel_params+1576]; selp.b32 %r2137, %r2136, %r601, %p1415; add.s16 %rs88, %rs28, 2; and.b16 %rs89, %rs88, 255; cvt.u32.u16 %r2138, %rs89; xor.b32 %r2139, %r2137, %r2138; mul.lo.s32 %r2140, %r2139, -512718531; shr.u32 %r2141, %r2137, 16; xor.b32 %r2142, %r2140, %r2141; bfe.u32 %r2143, %r2142, 4, 4; shr.u32 %r2144, %r2137, 8; xor.b32 %r2145, %r2142, %r2144; xor.b32 %r2146, %r2145, %r2143; mul.lo.s32 %r2147, %r2146, 153742143; shr.u32 %r2148, %r2137, 23; xor.b32 %r2149, %r2147, %r2148; bfe.u32 %r2150, %r2149, 1, 7; xor.b32 %r2151, %r2150, %r2149; shr.u32 %r2152, %r2137, 27; or.b32 %r2153, %r2152, 1; mul.lo.s32 %r2154, %r2153, %r2151; mul.lo.s32 %r2155, %r2154, -1074439621; bfe.u32 %r2156, %r2155, 2, 6; xor.b32 %r2157, %r2156, %r2155; mul.lo.s32 %r2158, %r2157, -1638916925; bfe.u32 %r2159, %r2158, 2, 6; xor.b32 %r2160, %r2159, %r2158; mul.lo.s32 %r2161, %r2160, -933190689; and.b32 %r2162, %r2161, 255; bfe.u32 %r2163, %r2161, 5, 3; xor.b32 %r2164, %r2163, %r2162; add.s32 %r2165, %r2132, -1; xor.b32 %r726, %r725, 61; xor.b32 %r2166, %r2137, %r726; mul.lo.s32 %r2167, %r2166, 9; shr.u32 %r2168, %r2167, 4; xor.b32 %r2169, %r2168, %r2167; mul.lo.s32 %r2170, %r2169, 668265261; add.s32 %r2171, %r2170, %r607; shr.u32 %r2172, %r2170, 16; or.b32 %r2173, %r2172, 1; mul.lo.s32 %r2174, %r2173, %r2171; mul.lo.s32 %r2175, %r2174, 89287766; xor.b32 %r2176, %r2175, %r2174; mul.lo.s32 %r2177, %r2176, 1403136100; xor.b32 %r2178, %r2177, %r2176; brev.b32 %r2179, %r2178; neg.s32 %r2180, %r2132; and.b32 %r2181, %r603, %r2180; and.b32 %r2182, %r2179, %r2165; or.b32 %r2183, %r2182, %r2181; mad.lo.s32 %r2184, %r2164, %r2132, %r2183; shl.b32 %r2185, %r2132, 8; rem.u32 %r2186, %r2184, %r2185; ld.const.u64 %rd2055, [kernel_params+2424]; cvta.to.global.u64 %rd2056, %rd2055; shl.b32 %r2187, %r2186, 1; mul.wide.u32 %rd2057, %r2187, 4; add.s64 %rd2058, %rd2056, %rd2057; ld.global.f32 %f16240, [%rd2058]; @%p1414 bra $L__BB108_1587; xor.b32 %r2188, %r726, %r601; mul.lo.s32 %r2189, %r2188, 9; shr.u32 %r2190, %r2189, 4; xor.b32 %r2191, %r2190, %r2189; mul.lo.s32 %r2192, %r2191, 668265261; cvt.rn.f32.u32 %f10264, %r2192; mul.ftz.f32 %f10265, %f10264, 0f2F7FFFFE; fma.rn.ftz.f32 %f10266, %f10265, %f2149, %f16240; cvt.rmi.ftz.f32.f32 %f10267, %f10266; sub.ftz.f32 %f16240, %f10266, %f10267; bra.uni $L__BB108_1587; $L__BB108_1586: shr.u32 %r2193, %r725, 16; xor.b32 %r2194, %r2193, %r725; mul.lo.s32 %r2195, %r2194, 569420461; shr.u32 %r2196, %r2195, 15; xor.b32 %r2197, %r2196, %r2195; mul.lo.s32 %r2198, %r2197, -749064809; shr.u32 %r2199, %r2198, 15; xor.b32 %r2200, %r601, %r2198; xor.b32 %r2201, %r2200, %r2199; xor.b32 %r2202, %r2201, 1493655573; add.s32 %r2203, %r2202, %r607; shr.u32 %r2204, %r2202, 16; or.b32 %r2205, %r2204, 1; mul.lo.s32 %r2206, %r2205, %r2203; mul.lo.s32 %r2207, %r2206, 89287766; xor.b32 %r2208, %r2207, %r2206; mul.lo.s32 %r2209, %r2208, 1403136100; xor.b32 %r2210, %r2209, %r2208; xor.b32 %r2211, %r2201, -2052961194; brev.b32 %r2212, %r2210; mul.lo.s32 %r2213, %r2212, 1025551850; xor.b32 %r2214, %r2213, %r2212; add.s32 %r2215, %r2214, %r2211; shr.u32 %r2216, %r2211, 16; or.b32 %r2217, %r2216, 1; mul.lo.s32 %r2218, %r2215, %r2217; mul.lo.s32 %r2219, %r2218, 89287766; xor.b32 %r2220, %r2219, %r2218; mul.lo.s32 %r2221, %r2220, 1403136100; xor.b32 %r2222, %r2221, %r2220; brev.b32 %r2223, %r2222; cvt.rn.f32.u32 %f10268, %r2223; mul.ftz.f32 %f16240, %f10268, 0f2F7FFFFE; $L__BB108_1587: setp.eq.ftz.f32 %p1416, %f16248, 0f00000000; setp.eq.ftz.f32 %p1417, %f16249, 0f00000000; and.pred %p1418, %p1417, %p1416; setp.eq.ftz.f32 %p1419, %f16247, 0f00000000; and.pred %p1420, %p1419, %p1418; @%p1420 bra $L__BB108_1685; @%p1412 bra $L__BB108_1592; abs.ftz.f32 %f10269, %f16249; abs.ftz.f32 %f10270, %f16248; max.ftz.f32 %f10271, %f10269, %f10270; abs.ftz.f32 %f10272, %f16247; max.ftz.f32 %f10273, %f10271, %f10272; mul.ftz.f32 %f2154, %f10273, %f2148; setp.geu.ftz.f32 %p1422, %f2154, 0f3F800000; @%p1422 bra $L__BB108_1592; setp.le.ftz.f32 %p1423, %f2154, %f16240; @%p1423 bra $L__BB108_1685; rcp.approx.ftz.f32 %f10274, %f2154; mul.ftz.f32 %f16244, %f10274, %f16244; mul.ftz.f32 %f16246, %f10274, %f16246; mul.ftz.f32 %f16245, %f10274, %f16245; st.local.f32 [%rd322], %f16244; st.local.f32 [%rd322+4], %f16246; st.local.f32 [%rd322+8], %f16245; mul.ftz.f32 %f16241, %f10274, %f16241; mul.ftz.f32 %f16243, %f10274, %f16243; mul.ftz.f32 %f16242, %f10274, %f16242; st.local.f32 [%rd322+12], %f16241; st.local.f32 [%rd322+16], %f16243; st.local.f32 [%rd322+20], %f16242; mul.ftz.f32 %f16249, %f10274, %f16249; mul.ftz.f32 %f16248, %f10274, %f16248; mul.ftz.f32 %f16247, %f10274, %f16247; st.local.f32 [%rd322+24], %f16249; st.local.f32 [%rd322+28], %f16248; st.local.f32 [%rd322+32], %f16247; $L__BB108_1592: ld.local.v4.f32 {%f16275, %f16276, %f16277, %f10278}, [%rd1]; ld.local.u32 %r727, [%rd7+-24]; and.b32 %r728, %r727, 1; setp.eq.s32 %p1424, %r728, 0; ld.local.u32 %r729, [%rd7+-40]; setp.gt.s32 %p1425, %r729, -1; mov.u16 %rs149, 1; or.pred %p1426, %p1424, %p1425; @%p1426 bra $L__BB108_1635; ld.const.u64 %rd2059, [kernel_params+2160]; cvta.to.global.u64 %rd2060, %rd2059; ld.local.u32 %r2224, [%rd7+-12]; mul.wide.s32 %rd2061, %r2224, 208; add.s64 %rd2062, %rd2060, %rd2061; add.s64 %rd323, %rd2062, 184; ld.global.f32 %f2176, [%rd2062+184]; setp.leu.ftz.f32 %p1427, %f2176, 0f00000000; @%p1427 bra $L__BB108_1635; ld.local.f32 %f10279, [%rd7+-76]; ld.local.v2.f32 {%f10280, %f10281}, [%rd7+-72]; mul.ftz.f32 %f10284, %f1936, %f10280; fma.rn.ftz.f32 %f10285, %f1935, %f10279, %f10284; fma.rn.ftz.f32 %f10286, %f1937, %f10281, %f10285; setp.lt.ftz.f32 %p1428, %f10286, 0f00000000; neg.ftz.f32 %f10287, %f10286; selp.f32 %f2177, %f10287, %f10286, %p1428; @%p1428 bra $L__BB108_1596; bra.uni $L__BB108_1595; $L__BB108_1596: ld.local.v2.f32 {%f10290, %f10291}, [%rd7+-64]; neg.ftz.f32 %f16250, %f10290; neg.ftz.f32 %f16251, %f10291; ld.local.f32 %f10294, [%rd7+-56]; neg.ftz.f32 %f16252, %f10294; bra.uni $L__BB108_1597; $L__BB108_1595: ld.local.v2.f32 {%f16250, %f16251}, [%rd7+-64]; ld.local.f32 %f16252, [%rd7+-56]; $L__BB108_1597: mul.ftz.f32 %f10295, %f1936, %f16251; fma.rn.ftz.f32 %f10296, %f1935, %f16250, %f10295; fma.rn.ftz.f32 %f2187, %f1937, %f16252, %f10296; add.ftz.f32 %f10297, %f2177, %f2187; setp.lt.ftz.f32 %p1429, %f2177, %f2176; selp.f32 %f10298, 0f40000000, 0f3F800000, %p1429; mov.f32 %f10299, 0f3F800000; selp.f32 %f10300, %f10297, %f2187, %p1429; div.approx.ftz.f32 %f10301, %f10300, %f2176; sub.ftz.f32 %f10302, %f10298, %f10301; mov.f32 %f10303, 0f00000000; max.ftz.f32 %f10304, %f10302, %f10303; min.ftz.f32 %f2188, %f10304, %f10299; setp.leu.ftz.f32 %p1430, %f2188, 0f00000000; @%p1430 bra $L__BB108_1635; setp.eq.s32 %p1431, %r727, 65; ld.local.s32 %rd324, [%rd7+-28]; @%p1431 bra $L__BB108_1600; bra.uni $L__BB108_1599; $L__BB108_1600: ld.global.u32 %r730, [%rd323+-28]; ld.global.u32 %r731, [%rd323+-32]; shl.b32 %r2236, %r731, 1; cvt.rn.f32.s32 %f10305, %r2236; ld.local.f32 %f10306, [%rd7+-4]; mul.ftz.f32 %f10307, %f10306, %f10305; cvt.rzi.ftz.s32.f32 %r2237, %f10307; add.s32 %r2238, %r2236, -1; min.s32 %r732, %r2237, %r2238; cvt.rn.f32.s32 %f10308, %r732; sub.ftz.f32 %f2207, %f10307, %f10308; ld.const.u64 %rd2085, [kernel_params+2312]; cvta.to.global.u64 %rd325, %rd2085; ld.global.u32 %r3236, [%rd323+-20]; mul.wide.u32 %rd2086, %r3236, 16; add.s64 %rd2087, %rd325, %rd2086; ld.global.u16 %rs147, [%rd2087+12]; ld.global.u32 %r3235, [%rd2087+8]; ld.global.u64 %rd3002, [%rd2087]; setp.eq.s64 %p1432, %rd3002, 11; mov.u16 %rs145, %rs147; mov.u32 %r3230, %r3235; @%p1432 bra $L__BB108_1606; mov.u64 %rd2983, %rd3002; mov.u32 %r3230, %r3235; mov.u16 %rs145, %rs147; mov.u32 %r3231, %r3236; $L__BB108_1602: setp.eq.s64 %p1433, %rd2983, 0; @%p1433 bra $L__BB108_1604; add.s32 %r3231, %r3231, 2; bra.uni $L__BB108_1605; $L__BB108_1604: setp.eq.s16 %p1434, %rs145, 0; mov.u32 %r3234, -1; mov.u32 %r3231, %r3230; @%p1434 bra $L__BB108_1607; $L__BB108_1605: mul.wide.u32 %rd2088, %r3231, 16; add.s64 %rd2089, %rd325, %rd2088; ld.global.u16 %rs145, [%rd2089+12]; ld.global.u32 %r3230, [%rd2089+8]; ld.global.u64 %rd2983, [%rd2089]; setp.ne.s64 %p1435, %rd2983, 11; @%p1435 bra $L__BB108_1602; $L__BB108_1606: setp.eq.s16 %p1436, %rs145, 0; selp.b32 %r3234, -1, %r3230, %p1436; $L__BB108_1607: ld.const.u64 %rd2090, [kernel_params+2232]; cvta.to.global.u64 %rd2091, %rd2090; shl.b64 %rd2092, %rd324, 4; add.s64 %rd2093, %rd2091, %rd2092; ld.global.v4.u32 {%r2240, %r2241, %r2242, %r2243}, [%rd2093]; setp.eq.s32 %p1437, %r732, %r731; @%p1437 bra $L__BB108_1609; bra.uni $L__BB108_1608; $L__BB108_1609: ld.const.u64 %rd2099, [kernel_params+2256]; cvta.to.global.u64 %rd2100, %rd2099; mul.wide.u32 %rd2101, %r2243, 12; add.s64 %rd2984, %rd2100, %rd2101; add.s32 %r2250, %r2243, 1; mul.wide.u32 %rd2102, %r2250, 12; add.s64 %rd2987, %rd2100, %rd2102; add.s32 %r2251, %r2243, 2; mul.wide.u32 %rd2103, %r2251, 12; add.s64 %rd2990, %rd2100, %rd2103; add.s64 %rd2985, %rd2984, 4; add.s64 %rd2986, %rd2984, 8; add.s64 %rd2988, %rd2987, 4; add.s64 %rd2989, %rd2987, 8; add.s64 %rd2991, %rd2990, 4; add.s64 %rd2992, %rd2990, 8; bra.uni $L__BB108_1610; $L__BB108_1599: cvt.u32.u64 %r2225, %rd324; ld.const.v2.u64 {%rd2063, %rd2064}, [kernel_params+2224]; cvta.to.global.u64 %rd2067, %rd2064; mul.wide.s32 %rd2068, %r2225, 16; add.s64 %rd2069, %rd2067, %rd2068; ld.global.v4.u32 {%r2226, %r2227, %r2228, %r2229}, [%rd2069]; ld.const.u64 %rd2070, [kernel_params+2256]; cvta.to.global.u64 %rd2071, %rd2070; mul.wide.u32 %rd2072, %r2229, 12; add.s64 %rd2073, %rd2071, %rd2072; ld.global.f32 %f16262, [%rd2073]; ld.global.f32 %f16263, [%rd2073+4]; ld.global.f32 %f16264, [%rd2073+8]; add.s32 %r2234, %r2229, 1; mul.wide.u32 %rd2074, %r2234, 12; add.s64 %rd2075, %rd2071, %rd2074; ld.global.f32 %f16265, [%rd2075]; ld.global.f32 %f16266, [%rd2075+4]; ld.global.f32 %f16267, [%rd2075+8]; add.s32 %r2235, %r2229, 2; mul.wide.u32 %rd2076, %r2235, 12; add.s64 %rd2077, %rd2071, %rd2076; ld.global.f32 %f16268, [%rd2077]; ld.global.f32 %f16269, [%rd2077+4]; ld.global.f32 %f16270, [%rd2077+8]; cvta.to.global.u64 %rd2078, %rd2063; mul.wide.u32 %rd2079, %r2226, 12; add.s64 %rd2080, %rd2078, %rd2079; ld.global.f32 %f16253, [%rd2080]; ld.global.f32 %f16254, [%rd2080+4]; ld.global.f32 %f16255, [%rd2080+8]; mul.wide.u32 %rd2081, %r2227, 12; add.s64 %rd2082, %rd2078, %rd2081; ld.global.f32 %f16256, [%rd2082]; ld.global.f32 %f16257, [%rd2082+4]; ld.global.f32 %f16258, [%rd2082+8]; mul.wide.u32 %rd2083, %r2228, 12; add.s64 %rd2084, %rd2078, %rd2083; ld.global.f32 %f16259, [%rd2084]; ld.global.f32 %f16260, [%rd2084+4]; ld.global.f32 %f16261, [%rd2084+8]; bra.uni $L__BB108_1627; $L__BB108_1608: setp.gt.s32 %p1438, %r732, %r731; selp.b32 %r2244, -1, 0, %p1438; add.s32 %r2245, %r732, %r2244; mad.lo.s32 %r2246, %r2245, %r730, %r3234; ld.const.u64 %rd2094, [kernel_params+2336]; cvta.to.global.u64 %rd2095, %rd2094; add.s32 %r2247, %r2240, %r2246; mul.wide.u32 %rd2096, %r2247, 12; add.s64 %rd2984, %rd2095, %rd2096; add.s32 %r2248, %r2241, %r2246; mul.wide.u32 %rd2097, %r2248, 12; add.s64 %rd2987, %rd2095, %rd2097; add.s32 %r2249, %r2242, %r2246; mul.wide.u32 %rd2098, %r2249, 12; add.s64 %rd2990, %rd2095, %rd2098; add.s64 %rd2985, %rd2984, 4; add.s64 %rd2986, %rd2984, 8; add.s64 %rd2988, %rd2987, 4; add.s64 %rd2989, %rd2987, 8; add.s64 %rd2991, %rd2990, 4; add.s64 %rd2992, %rd2990, 8; $L__BB108_1610: ld.global.f32 %f2208, [%rd2992]; ld.global.f32 %f2209, [%rd2991]; ld.global.f32 %f2210, [%rd2990]; ld.global.f32 %f2211, [%rd2989]; ld.global.f32 %f2212, [%rd2988]; ld.global.f32 %f2213, [%rd2987]; ld.global.f32 %f2214, [%rd2986]; ld.global.f32 %f2215, [%rd2985]; ld.global.f32 %f2216, [%rd2984]; add.s32 %r747, %r732, 1; setp.eq.s32 %p1439, %r747, %r731; @%p1439 bra $L__BB108_1612; bra.uni $L__BB108_1611; $L__BB108_1612: ld.const.u64 %rd2109, [kernel_params+2256]; cvta.to.global.u64 %rd2110, %rd2109; mul.wide.u32 %rd2111, %r2243, 12; add.s64 %rd2993, %rd2110, %rd2111; add.s32 %r2258, %r2243, 1; mul.wide.u32 %rd2112, %r2258, 12; add.s64 %rd2996, %rd2110, %rd2112; add.s32 %r2259, %r2243, 2; mul.wide.u32 %rd2113, %r2259, 12; add.s64 %rd2999, %rd2110, %rd2113; add.s64 %rd2994, %rd2993, 4; add.s64 %rd2995, %rd2993, 8; add.s64 %rd2997, %rd2996, 4; add.s64 %rd2998, %rd2996, 8; add.s64 %rd3000, %rd2999, 4; add.s64 %rd3001, %rd2999, 8; bra.uni $L__BB108_1613; $L__BB108_1611: setp.ge.s32 %p1440, %r732, %r731; selp.b32 %r2252, -1, 0, %p1440; add.s32 %r2253, %r747, %r2252; mad.lo.s32 %r2254, %r2253, %r730, %r3234; ld.const.u64 %rd2104, [kernel_params+2336]; cvta.to.global.u64 %rd2105, %rd2104; add.s32 %r2255, %r2240, %r2254; mul.wide.u32 %rd2106, %r2255, 12; add.s64 %rd2993, %rd2105, %rd2106; add.s32 %r2256, %r2241, %r2254; mul.wide.u32 %rd2107, %r2256, 12; add.s64 %rd2996, %rd2105, %rd2107; add.s32 %r2257, %r2242, %r2254; mul.wide.u32 %rd2108, %r2257, 12; add.s64 %rd2999, %rd2105, %rd2108; add.s64 %rd2994, %rd2993, 4; add.s64 %rd2995, %rd2993, 8; add.s64 %rd2997, %rd2996, 4; add.s64 %rd2998, %rd2996, 8; add.s64 %rd3000, %rd2999, 4; add.s64 %rd3001, %rd2999, 8; $L__BB108_1613: sub.ftz.f32 %f2217, %f10299, %f2207; ld.global.f32 %f10310, [%rd2993]; mul.ftz.f32 %f10311, %f2207, %f10310; ld.global.f32 %f10312, [%rd2994]; mul.ftz.f32 %f10313, %f2207, %f10312; ld.global.f32 %f10314, [%rd2995]; mul.ftz.f32 %f10315, %f2207, %f10314; fma.rn.ftz.f32 %f16262, %f2217, %f2216, %f10311; fma.rn.ftz.f32 %f16263, %f2217, %f2215, %f10313; fma.rn.ftz.f32 %f16264, %f2217, %f2214, %f10315; ld.global.f32 %f10316, [%rd2996]; mul.ftz.f32 %f10317, %f2207, %f10316; ld.global.f32 %f10318, [%rd2997]; mul.ftz.f32 %f10319, %f2207, %f10318; ld.global.f32 %f10320, [%rd2998]; mul.ftz.f32 %f10321, %f2207, %f10320; fma.rn.ftz.f32 %f16265, %f2217, %f2213, %f10317; fma.rn.ftz.f32 %f16266, %f2217, %f2212, %f10319; fma.rn.ftz.f32 %f16267, %f2217, %f2211, %f10321; ld.global.f32 %f10322, [%rd2999]; mul.ftz.f32 %f10323, %f2207, %f10322; ld.global.f32 %f10324, [%rd3000]; mul.ftz.f32 %f10325, %f2207, %f10324; ld.global.f32 %f10326, [%rd3001]; mul.ftz.f32 %f10327, %f2207, %f10326; fma.rn.ftz.f32 %f16268, %f2217, %f2210, %f10323; fma.rn.ftz.f32 %f16269, %f2217, %f2209, %f10325; fma.rn.ftz.f32 %f16270, %f2217, %f2208, %f10327; setp.eq.s64 %p1441, %rd3002, 12; @%p1441 bra $L__BB108_1619; mov.u32 %r3239, -1; $L__BB108_1615: setp.eq.s64 %p1442, %rd3002, 0; @%p1442 bra $L__BB108_1617; add.s32 %r3236, %r3236, 2; bra.uni $L__BB108_1618; $L__BB108_1617: setp.eq.s16 %p1443, %rs147, 0; mov.u32 %r3236, %r3235; @%p1443 bra $L__BB108_1620; $L__BB108_1618: mul.wide.u32 %rd2114, %r3236, 16; add.s64 %rd2115, %rd325, %rd2114; ld.global.u16 %rs147, [%rd2115+12]; ld.global.u32 %r3235, [%rd2115+8]; ld.global.u64 %rd3002, [%rd2115]; setp.ne.s64 %p1444, %rd3002, 12; @%p1444 bra $L__BB108_1615; $L__BB108_1619: setp.eq.s16 %p1445, %rs147, 0; selp.b32 %r3239, -1, %r3235, %p1445; $L__BB108_1620: @%p1437 bra $L__BB108_1622; bra.uni $L__BB108_1621; $L__BB108_1622: ld.const.u64 %rd2121, [kernel_params+2224]; cvta.to.global.u64 %rd2122, %rd2121; mul.wide.u32 %rd2123, %r2240, 12; add.s64 %rd3003, %rd2122, %rd2123; mul.wide.u32 %rd2124, %r2241, 12; add.s64 %rd3006, %rd2122, %rd2124; mul.wide.u32 %rd2125, %r2242, 12; add.s64 %rd3009, %rd2122, %rd2125; add.s64 %rd3004, %rd3003, 4; add.s64 %rd3005, %rd3003, 8; add.s64 %rd3007, %rd3006, 4; add.s64 %rd3008, %rd3006, 8; add.s64 %rd3010, %rd3009, 4; add.s64 %rd3011, %rd3009, 8; bra.uni $L__BB108_1623; $L__BB108_1621: setp.gt.s32 %p1447, %r732, %r731; selp.b32 %r2261, -1, 0, %p1447; add.s32 %r2262, %r732, %r2261; mad.lo.s32 %r2263, %r2262, %r730, %r3239; ld.const.u64 %rd2116, [kernel_params+2336]; cvta.to.global.u64 %rd2117, %rd2116; add.s32 %r2264, %r2263, %r2240; mul.wide.u32 %rd2118, %r2264, 12; add.s64 %rd3003, %rd2117, %rd2118; add.s32 %r2265, %r2263, %r2241; mul.wide.u32 %rd2119, %r2265, 12; add.s64 %rd3006, %rd2117, %rd2119; add.s32 %r2266, %r2263, %r2242; mul.wide.u32 %rd2120, %r2266, 12; add.s64 %rd3009, %rd2117, %rd2120; add.s64 %rd3004, %rd3003, 4; add.s64 %rd3005, %rd3003, 8; add.s64 %rd3007, %rd3006, 4; add.s64 %rd3008, %rd3006, 8; add.s64 %rd3010, %rd3009, 4; add.s64 %rd3011, %rd3009, 8; $L__BB108_1623: ld.global.f32 %f2227, [%rd3011]; ld.global.f32 %f2228, [%rd3010]; ld.global.f32 %f2229, [%rd3009]; ld.global.f32 %f2230, [%rd3008]; ld.global.f32 %f2231, [%rd3007]; ld.global.f32 %f2232, [%rd3006]; ld.global.f32 %f2233, [%rd3005]; ld.global.f32 %f2234, [%rd3004]; ld.global.f32 %f2235, [%rd3003]; @%p1439 bra $L__BB108_1625; bra.uni $L__BB108_1624; $L__BB108_1625: ld.const.u64 %rd2131, [kernel_params+2224]; cvta.to.global.u64 %rd2132, %rd2131; mul.wide.u32 %rd2133, %r2240, 12; add.s64 %rd3012, %rd2132, %rd2133; mul.wide.u32 %rd2134, %r2241, 12; add.s64 %rd3015, %rd2132, %rd2134; mul.wide.u32 %rd2135, %r2242, 12; add.s64 %rd3018, %rd2132, %rd2135; add.s64 %rd3013, %rd3012, 4; add.s64 %rd3014, %rd3012, 8; add.s64 %rd3016, %rd3015, 4; add.s64 %rd3017, %rd3015, 8; add.s64 %rd3019, %rd3018, 4; add.s64 %rd3020, %rd3018, 8; bra.uni $L__BB108_1626; $L__BB108_1624: setp.ge.s32 %p1449, %r732, %r731; selp.b32 %r2267, -1, 0, %p1449; add.s32 %r2268, %r747, %r2267; mad.lo.s32 %r2269, %r2268, %r730, %r3239; ld.const.u64 %rd2126, [kernel_params+2336]; cvta.to.global.u64 %rd2127, %rd2126; add.s32 %r2270, %r2269, %r2240; mul.wide.u32 %rd2128, %r2270, 12; add.s64 %rd3012, %rd2127, %rd2128; add.s32 %r2271, %r2269, %r2241; mul.wide.u32 %rd2129, %r2271, 12; add.s64 %rd3015, %rd2127, %rd2129; add.s32 %r2272, %r2269, %r2242; mul.wide.u32 %rd2130, %r2272, 12; add.s64 %rd3018, %rd2127, %rd2130; add.s64 %rd3013, %rd3012, 4; add.s64 %rd3014, %rd3012, 8; add.s64 %rd3016, %rd3015, 4; add.s64 %rd3017, %rd3015, 8; add.s64 %rd3019, %rd3018, 4; add.s64 %rd3020, %rd3018, 8; $L__BB108_1626: ld.global.f32 %f10328, [%rd3012]; mul.ftz.f32 %f10329, %f2207, %f10328; ld.global.f32 %f10330, [%rd3013]; mul.ftz.f32 %f10331, %f2207, %f10330; ld.global.f32 %f10332, [%rd3014]; mul.ftz.f32 %f10333, %f2207, %f10332; fma.rn.ftz.f32 %f16253, %f2217, %f2235, %f10329; fma.rn.ftz.f32 %f16254, %f2217, %f2234, %f10331; fma.rn.ftz.f32 %f16255, %f2217, %f2233, %f10333; ld.global.f32 %f10334, [%rd3015]; mul.ftz.f32 %f10335, %f2207, %f10334; ld.global.f32 %f10336, [%rd3016]; mul.ftz.f32 %f10337, %f2207, %f10336; ld.global.f32 %f10338, [%rd3017]; mul.ftz.f32 %f10339, %f2207, %f10338; fma.rn.ftz.f32 %f16256, %f2217, %f2232, %f10335; fma.rn.ftz.f32 %f16257, %f2217, %f2231, %f10337; fma.rn.ftz.f32 %f16258, %f2217, %f2230, %f10339; ld.global.f32 %f10340, [%rd3018]; mul.ftz.f32 %f10341, %f2207, %f10340; ld.global.f32 %f10342, [%rd3019]; mul.ftz.f32 %f10343, %f2207, %f10342; ld.global.f32 %f10344, [%rd3020]; mul.ftz.f32 %f10345, %f2207, %f10344; fma.rn.ftz.f32 %f16259, %f2217, %f2229, %f10341; fma.rn.ftz.f32 %f16260, %f2217, %f2228, %f10343; fma.rn.ftz.f32 %f16261, %f2217, %f2227, %f10345; $L__BB108_1627: ld.local.f32 %f2263, [%rd7+-20]; sub.ftz.f32 %f10347, %f10299, %f2263; ld.local.f32 %f10348, [%rd7+-16]; sub.ftz.f32 %f2264, %f10347, %f10348; mul.ftz.f32 %f10349, %f16262, %f2264; mul.ftz.f32 %f10350, %f16263, %f2264; mul.ftz.f32 %f10351, %f16264, %f2264; fma.rn.ftz.f32 %f10352, %f16265, %f2263, %f10349; fma.rn.ftz.f32 %f10353, %f16266, %f2263, %f10350; fma.rn.ftz.f32 %f10354, %f16267, %f2263, %f10351; fma.rn.ftz.f32 %f2265, %f16268, %f10348, %f10352; fma.rn.ftz.f32 %f2266, %f16269, %f10348, %f10353; fma.rn.ftz.f32 %f2267, %f16270, %f10348, %f10354; mul.ftz.f32 %f10355, %f16253, %f2264; mul.ftz.f32 %f10356, %f16254, %f2264; mul.ftz.f32 %f10357, %f16255, %f2264; fma.rn.ftz.f32 %f10358, %f16256, %f2263, %f10355; fma.rn.ftz.f32 %f10359, %f16257, %f2263, %f10356; fma.rn.ftz.f32 %f10360, %f16258, %f2263, %f10357; fma.rn.ftz.f32 %f16271, %f16259, %f10348, %f10358; fma.rn.ftz.f32 %f16272, %f16260, %f10348, %f10359; fma.rn.ftz.f32 %f16273, %f16261, %f10348, %f10360; ld.local.u32 %r756, [%rd7+-32]; and.b32 %r2273, %r756, 4; setp.ne.s32 %p1450, %r2273, 0; @%p1450 bra $L__BB108_1631; and.b32 %r2274, %r756, 2; setp.eq.s32 %p1451, %r2274, 0; @%p1451 bra $L__BB108_1630; bra.uni $L__BB108_1629; $L__BB108_1630: ld.global.v4.f32 {%f10388, %f10389, %f10390, %f10391}, [%rd323+-184]; ld.global.v4.f32 {%f10395, %f10396, %f10397, %f10398}, [%rd323+-168]; ld.global.v4.f32 {%f10402, %f10403, %f10404, %f10405}, [%rd323+-152]; mul.ftz.f32 %f10409, %f16272, %f10389; fma.rn.ftz.f32 %f10410, %f16271, %f10388, %f10409; fma.rn.ftz.f32 %f2274, %f16273, %f10390, %f10410; mul.ftz.f32 %f10411, %f16272, %f10396; fma.rn.ftz.f32 %f10412, %f16271, %f10395, %f10411; fma.rn.ftz.f32 %f2275, %f16273, %f10397, %f10412; mul.ftz.f32 %f10413, %f16272, %f10403; fma.rn.ftz.f32 %f10414, %f16271, %f10402, %f10413; fma.rn.ftz.f32 %f16273, %f16273, %f10404, %f10414; mov.f32 %f16271, %f2274; mov.f32 %f16272, %f2275; bra.uni $L__BB108_1631; $L__BB108_1629: ld.local.v4.f32 {%f10361, %f10362, %f10363, %f10364}, [%rd7+56]; mul.ftz.f32 %f10368, %f16272, %f10362; fma.rn.ftz.f32 %f10369, %f16271, %f10361, %f10368; fma.rn.ftz.f32 %f2271, %f16273, %f10363, %f10369; ld.local.v4.f32 {%f10370, %f10371, %f10372, %f10373}, [%rd7+72]; mul.ftz.f32 %f10377, %f16272, %f10371; fma.rn.ftz.f32 %f10378, %f16271, %f10370, %f10377; fma.rn.ftz.f32 %f2272, %f16273, %f10372, %f10378; ld.local.v4.f32 {%f10379, %f10380, %f10381, %f10382}, [%rd7+88]; mul.ftz.f32 %f10386, %f16272, %f10380; fma.rn.ftz.f32 %f10387, %f16271, %f10379, %f10386; fma.rn.ftz.f32 %f16273, %f16273, %f10381, %f10387; mov.f32 %f16271, %f2271; mov.f32 %f16272, %f2272; $L__BB108_1631: sub.ftz.f32 %f2280, %f16262, %f16268; sub.ftz.f32 %f10415, %f16259, %f16253; sub.ftz.f32 %f2281, %f16263, %f16269; sub.ftz.f32 %f10416, %f16260, %f16254; mul.ftz.f32 %f10417, %f10416, %f2281; fma.rn.ftz.f32 %f10418, %f10415, %f2280, %f10417; sub.ftz.f32 %f2282, %f16264, %f16270; sub.ftz.f32 %f10419, %f16261, %f16255; fma.rn.ftz.f32 %f10420, %f10419, %f2282, %f10418; sub.ftz.f32 %f2283, %f16265, %f16268; sub.ftz.f32 %f10421, %f16259, %f16256; sub.ftz.f32 %f2284, %f16266, %f16269; sub.ftz.f32 %f10422, %f16260, %f16257; mul.ftz.f32 %f10423, %f10422, %f2284; fma.rn.ftz.f32 %f10424, %f10421, %f2283, %f10423; sub.ftz.f32 %f2285, %f16267, %f16270; sub.ftz.f32 %f10425, %f16261, %f16258; fma.rn.ftz.f32 %f10426, %f10425, %f2285, %f10424; sub.ftz.f32 %f2286, %f16265, %f16262; sub.ftz.f32 %f10427, %f16256, %f16253; sub.ftz.f32 %f2287, %f16266, %f16263; sub.ftz.f32 %f10428, %f16257, %f16254; mul.ftz.f32 %f10429, %f10428, %f2287; fma.rn.ftz.f32 %f10430, %f10427, %f2286, %f10429; sub.ftz.f32 %f2288, %f16267, %f16264; sub.ftz.f32 %f10431, %f16258, %f16255; fma.rn.ftz.f32 %f10432, %f10431, %f2288, %f10430; mul.ftz.f32 %f10433, %f10420, %f2264; add.ftz.f32 %f10434, %f2264, 0fBF800000; add.ftz.f32 %f10435, %f10420, %f10426; add.ftz.f32 %f10436, %f10432, %f10435; mul.ftz.f32 %f10437, %f2264, %f10436; mul.ftz.f32 %f10438, %f2263, %f10437; fma.rn.ftz.f32 %f10439, %f10433, %f10434, %f10438; mul.ftz.f32 %f10440, %f2263, %f10426; add.ftz.f32 %f10441, %f2263, 0fBF800000; fma.rn.ftz.f32 %f10442, %f10441, %f10440, %f10439; mul.ftz.f32 %f10443, %f16251, %f16272; fma.rn.ftz.f32 %f10444, %f16250, %f16271, %f10443; fma.rn.ftz.f32 %f10445, %f16252, %f16273, %f10444; setp.gt.ftz.f32 %p1452, %f10445, 0f00000000; sub.ftz.f32 %f2289, %f16262, %f16265; sub.ftz.f32 %f2290, %f16263, %f16266; sub.ftz.f32 %f2291, %f16264, %f16267; sub.ftz.f32 %f2292, %f16268, %f16262; sub.ftz.f32 %f2293, %f16269, %f16263; sub.ftz.f32 %f2294, %f16270, %f16264; sub.ftz.f32 %f2295, %f16268, %f16265; sub.ftz.f32 %f2296, %f16269, %f16266; sub.ftz.f32 %f2297, %f16270, %f16267; mul.ftz.f32 %f2298, %f10442, 0f3F000000; @%p1452 bra $L__BB108_1633; bra.uni $L__BB108_1632; $L__BB108_1633: mul.ftz.f32 %f10498, %f16254, %f2287; fma.rn.ftz.f32 %f10499, %f16253, %f2286, %f10498; fma.rn.ftz.f32 %f10500, %f16255, %f2288, %f10499; mul.ftz.f32 %f10501, %f16254, %f2293; fma.rn.ftz.f32 %f10502, %f16253, %f2292, %f10501; fma.rn.ftz.f32 %f10503, %f16255, %f2294, %f10502; max.ftz.f32 %f10504, %f10500, %f10503; max.ftz.f32 %f10506, %f10504, %f10303; mul.ftz.f32 %f10507, %f16257, %f2290; fma.rn.ftz.f32 %f10508, %f16256, %f2289, %f10507; fma.rn.ftz.f32 %f10509, %f16258, %f2291, %f10508; mul.ftz.f32 %f10510, %f16257, %f2296; fma.rn.ftz.f32 %f10511, %f16256, %f2295, %f10510; fma.rn.ftz.f32 %f10512, %f16258, %f2297, %f10511; max.ftz.f32 %f10513, %f10509, %f10512; max.ftz.f32 %f10514, %f10513, %f10303; mul.ftz.f32 %f10515, %f16260, %f2281; fma.rn.ftz.f32 %f10516, %f16259, %f2280, %f10515; fma.rn.ftz.f32 %f10517, %f16261, %f2282, %f10516; mul.ftz.f32 %f10518, %f16260, %f2284; fma.rn.ftz.f32 %f10519, %f16259, %f2283, %f10518; fma.rn.ftz.f32 %f10520, %f16261, %f2285, %f10519; max.ftz.f32 %f10521, %f10517, %f10520; max.ftz.f32 %f10522, %f10521, %f10303; sub.ftz.f32 %f10523, %f16262, %f2265; sub.ftz.f32 %f10524, %f16263, %f2266; mul.ftz.f32 %f10525, %f16254, %f10524; fma.rn.ftz.f32 %f10526, %f16253, %f10523, %f10525; sub.ftz.f32 %f10527, %f16264, %f2267; fma.rn.ftz.f32 %f10528, %f16255, %f10527, %f10526; add.ftz.f32 %f10529, %f10528, %f10506; max.ftz.f32 %f10530, %f10529, %f10303; sub.ftz.f32 %f10531, %f16265, %f2265; sub.ftz.f32 %f10532, %f16266, %f2266; mul.ftz.f32 %f10533, %f16257, %f10532; fma.rn.ftz.f32 %f10534, %f16256, %f10531, %f10533; sub.ftz.f32 %f10535, %f16267, %f2267; fma.rn.ftz.f32 %f10536, %f16258, %f10535, %f10534; add.ftz.f32 %f10537, %f10536, %f10514; max.ftz.f32 %f10538, %f10537, %f10303; sub.ftz.f32 %f10539, %f16268, %f2265; sub.ftz.f32 %f10540, %f16269, %f2266; mul.ftz.f32 %f10541, %f16260, %f10540; fma.rn.ftz.f32 %f10542, %f16259, %f10539, %f10541; sub.ftz.f32 %f10543, %f16270, %f2267; fma.rn.ftz.f32 %f10544, %f16261, %f10543, %f10542; add.ftz.f32 %f10545, %f10544, %f10522; max.ftz.f32 %f10546, %f10545, %f10303; min.ftz.f32 %f10547, %f10530, %f10538; min.ftz.f32 %f10548, %f10547, %f10546; max.ftz.f32 %f16274, %f10548, %f2298; bra.uni $L__BB108_1634; $L__BB108_1632: mul.ftz.f32 %f10446, %f16254, %f2290; fma.rn.ftz.f32 %f10447, %f16253, %f2289, %f10446; fma.rn.ftz.f32 %f10448, %f16255, %f2291, %f10447; mul.ftz.f32 %f10449, %f16254, %f2281; fma.rn.ftz.f32 %f10450, %f16253, %f2280, %f10449; fma.rn.ftz.f32 %f10451, %f16255, %f2282, %f10450; max.ftz.f32 %f10452, %f10448, %f10451; max.ftz.f32 %f10454, %f10452, %f10303; mul.ftz.f32 %f10455, %f16257, %f2287; fma.rn.ftz.f32 %f10456, %f16256, %f2286, %f10455; fma.rn.ftz.f32 %f10457, %f16258, %f2288, %f10456; mul.ftz.f32 %f10458, %f16257, %f2284; fma.rn.ftz.f32 %f10459, %f16256, %f2283, %f10458; fma.rn.ftz.f32 %f10460, %f16258, %f2285, %f10459; max.ftz.f32 %f10461, %f10457, %f10460; max.ftz.f32 %f10462, %f10461, %f10303; mul.ftz.f32 %f10463, %f16260, %f2293; fma.rn.ftz.f32 %f10464, %f16259, %f2292, %f10463; fma.rn.ftz.f32 %f10465, %f16261, %f2294, %f10464; mul.ftz.f32 %f10466, %f16260, %f2296; fma.rn.ftz.f32 %f10467, %f16259, %f2295, %f10466; fma.rn.ftz.f32 %f10468, %f16261, %f2297, %f10467; max.ftz.f32 %f10469, %f10465, %f10468; max.ftz.f32 %f10470, %f10469, %f10303; sub.ftz.f32 %f10471, %f2265, %f16262; sub.ftz.f32 %f10472, %f2266, %f16263; mul.ftz.f32 %f10473, %f16254, %f10472; fma.rn.ftz.f32 %f10474, %f16253, %f10471, %f10473; sub.ftz.f32 %f10475, %f2267, %f16264; fma.rn.ftz.f32 %f10476, %f16255, %f10475, %f10474; add.ftz.f32 %f10477, %f10476, %f10454; max.ftz.f32 %f10478, %f10477, %f10303; sub.ftz.f32 %f10479, %f2265, %f16265; sub.ftz.f32 %f10480, %f2266, %f16266; mul.ftz.f32 %f10481, %f16257, %f10480; fma.rn.ftz.f32 %f10482, %f16256, %f10479, %f10481; sub.ftz.f32 %f10483, %f2267, %f16267; fma.rn.ftz.f32 %f10484, %f16258, %f10483, %f10482; add.ftz.f32 %f10485, %f10484, %f10462; max.ftz.f32 %f10486, %f10485, %f10303; sub.ftz.f32 %f10487, %f2265, %f16268; sub.ftz.f32 %f10488, %f2266, %f16269; mul.ftz.f32 %f10489, %f16260, %f10488; fma.rn.ftz.f32 %f10490, %f16259, %f10487, %f10489; sub.ftz.f32 %f10491, %f2267, %f16270; fma.rn.ftz.f32 %f10492, %f16261, %f10491, %f10490; add.ftz.f32 %f10493, %f10492, %f10470; max.ftz.f32 %f10494, %f10493, %f10303; min.ftz.f32 %f10495, %f10478, %f10486; min.ftz.f32 %f10496, %f10495, %f10494; neg.ftz.f32 %f10497, %f10496; min.ftz.f32 %f16274, %f10497, %f2298; $L__BB108_1634: mul.ftz.f32 %f10549, %f16271, %f16274; mul.ftz.f32 %f10550, %f16272, %f16274; mul.ftz.f32 %f10551, %f16273, %f16274; fma.rn.ftz.f32 %f16275, %f2188, %f10549, %f16275; fma.rn.ftz.f32 %f16276, %f2188, %f10550, %f16276; fma.rn.ftz.f32 %f16277, %f2188, %f10551, %f16277; setp.gt.ftz.f32 %p1453, %f2187, 0f00000000; selp.u16 %rs149, 1, 0, %p1453; $L__BB108_1635: and.b32 %r2275, %r618, 1073741824; setp.eq.s32 %p1454, %r2275, 0; mov.f32 %f16278, 0f00000000; mov.f32 %f16279, %f16278; mov.f32 %f16280, %f16278; mov.f32 %f16281, %f16278; mov.f32 %f16282, %f16278; mov.f32 %f16283, %f16278; mov.f32 %f16284, %f16278; @%p1454 bra $L__BB108_1638; ld.local.f32 %f10560, [%rd282+12]; setp.eq.ftz.f32 %p1455, %f10560, 0f7F7FFFFF; mov.f32 %f16284, 0f7F7FFFFF; mov.f32 %f16278, %f16275; mov.f32 %f16279, %f16276; mov.f32 %f16280, %f16277; mov.f32 %f16281, %f1935; mov.f32 %f16282, %f1936; mov.f32 %f16283, %f1937; @%p1455 bra $L__BB108_1638; ld.local.f32 %f10561, [%rd282+-24]; sub.ftz.f32 %f10562, %f10561, %f16275; ld.local.f32 %f10563, [%rd282+-20]; sub.ftz.f32 %f10564, %f10563, %f16276; ld.local.f32 %f10565, [%rd282+-16]; sub.ftz.f32 %f10566, %f10565, %f16277; mul.ftz.f32 %f10567, %f10564, %f10564; fma.rn.ftz.f32 %f10568, %f10562, %f10562, %f10567; fma.rn.ftz.f32 %f10569, %f10566, %f10566, %f10568; sqrt.approx.ftz.f32 %f16284, %f10569; rsqrt.approx.ftz.f32 %f10570, %f10569; mul.ftz.f32 %f16281, %f10562, %f10570; mul.ftz.f32 %f16282, %f10564, %f10570; mul.ftz.f32 %f16283, %f10566, %f10570; mov.f32 %f16278, %f16275; mov.f32 %f16279, %f16276; mov.f32 %f16280, %f16277; $L__BB108_1638: ld.local.f32 %f2319, [%rd7+4]; ld.local.f32 %f2320, [%rd7+-4]; ld.local.u32 %r757, [%rd7+-12]; setp.eq.s16 %p6, %rs149, 0; selp.b32 %r758, -1, %r757, %p6; ld.local.u32 %r759, [%rd7+-28]; ld.local.u32 %r760, [%rd282+32]; ld.local.u32 %r761, [%rd282+36]; ld.local.u32 %r762, [%rd282+52]; ld.const.u64 %rd2136, [kernel_params+4376]; cvta.to.global.u64 %rd2137, %rd2136; atom.global.add.u32 %r2276, [%rd2137], 1; ld.const.u64 %rd2138, [kernel_params+4256]; cvta.to.global.u64 %rd2139, %rd2138; add.s64 %rd2140, %rd2139, 12; atom.global.add.u32 %r2277, [%rd2140], 1; ld.const.u64 %rd2141, [kernel_params+3416]; cvta.to.global.u64 %rd2142, %rd2141; cvt.s64.s32 %rd439, %r2276; mul.wide.s32 %rd2143, %r2276, 2; add.s64 %rd2144, %rd2142, %rd2143; mov.u16 %rs93, 3; st.global.u16 [%rd2144], %rs93; ld.const.u32 %r763, [kernel_params]; and.b32 %r2278, %r763, 131072; setp.eq.s32 %p1456, %r2278, 0; @%p1456 bra $L__BB108_1641; mov.u32 %r3240, 0; mov.u64 %rd2146, kernel_params; $L__BB108_1640: mul.wide.s32 %rd2145, %r3240, 16; add.s64 %rd2147, %rd2146, %rd2145; ld.const.v2.u64 {%rd2148, %rd2149}, [%rd2147+2768]; cvta.to.global.u64 %rd2152, %rd2149; add.s64 %rd2154, %rd2152, %rd837; ld.global.u32 %r2280, [%rd2154]; cvta.to.global.u64 %rd2155, %rd2148; add.s64 %rd2156, %rd2155, %rd837; ld.global.u32 %r2281, [%rd2156]; ld.const.v2.u64 {%rd2157, %rd2158}, [%rd2147+3744]; cvta.to.global.u64 %rd2161, %rd2157; shl.b64 %rd2162, %rd439, 2; add.s64 %rd2163, %rd2161, %rd2162; st.global.u32 [%rd2163], %r2281; cvta.to.global.u64 %rd2164, %rd2158; add.s64 %rd2165, %rd2164, %rd2162; st.global.u32 [%rd2165], %r2280; add.s32 %r3240, %r3240, 1; setp.ne.s32 %p1457, %r2280, -1; @%p1457 bra $L__BB108_1640; $L__BB108_1641: setp.geu.ftz.f32 %p1458, %f1938, 0f00000000; @%p1458 bra $L__BB108_1656; ld.local.u32 %r766, [%rd7+-36]; and.b32 %r2282, %r766, 262144; setp.eq.s32 %p1459, %r2282, 0; @%p1459 bra $L__BB108_1656; bra.uni $L__BB108_1643; $L__BB108_1656: setp.eq.s32 %p1476, %r758, -1; selp.b32 %r780, -1, %r759, %p6; @%p1476 bra $L__BB108_1665; @%p1424 bra $L__BB108_1665; ld.const.u64 %rd2229, [kernel_params+2232]; cvta.to.global.u64 %rd2230, %rd2229; mul.wide.s32 %rd2231, %r759, 16; add.s64 %rd2232, %rd2230, %rd2231; ld.const.u64 %rd2233, [kernel_params+2256]; cvta.to.global.u64 %rd2234, %rd2233; ld.global.u32 %r2295, [%rd2232+12]; mul.wide.u32 %rd2235, %r2295, 12; add.s64 %rd2236, %rd2234, %rd2235; ld.global.f32 %f2324, [%rd2236]; ld.global.f32 %f2325, [%rd2236+4]; ld.global.f32 %f2326, [%rd2236+8]; add.s32 %r2296, %r2295, 1; mul.wide.u32 %rd2237, %r2296, 12; add.s64 %rd2238, %rd2234, %rd2237; ld.global.f32 %f2327, [%rd2238]; ld.global.f32 %f2328, [%rd2238+4]; ld.global.f32 %f2329, [%rd2238+8]; add.s32 %r2297, %r2295, 2; mul.wide.u32 %rd2239, %r2297, 12; add.s64 %rd2240, %rd2234, %rd2239; ld.global.f32 %f2330, [%rd2240]; ld.global.f32 %f2331, [%rd2240+4]; ld.global.f32 %f2332, [%rd2240+8]; ld.local.u32 %r781, [%rd7+-32]; and.b32 %r2298, %r781, 4; setp.ne.s32 %p1478, %r2298, 0; mov.f32 %f16297, %f16278; mov.f32 %f16298, %f16279; mov.f32 %f16299, %f16280; mov.f32 %f16300, %f16281; mov.f32 %f16301, %f16282; mov.f32 %f16302, %f16283; @%p1478 bra $L__BB108_1663; and.b32 %r2299, %r781, 2; setp.eq.s32 %p1479, %r2299, 0; @%p1479 bra $L__BB108_1661; bra.uni $L__BB108_1660; $L__BB108_1661: ld.const.u64 %rd2241, [kernel_params+2160]; cvta.to.global.u64 %rd2242, %rd2241; mul.wide.s32 %rd2243, %r757, 208; add.s64 %rd2244, %rd2242, %rd2243; ld.global.v4.f32 {%f16285, %f16286, %f16287, %f16288}, [%rd2244+48]; ld.global.v4.f32 {%f16289, %f16290, %f16291, %f16292}, [%rd2244+64]; ld.global.v4.f32 {%f16293, %f16294, %f16295, %f16296}, [%rd2244+80]; bra.uni $L__BB108_1662; $L__BB108_1643: and.b32 %r2283, %r766, 1; setp.eq.b32 %p1460, %r2283, 1; mov.pred %p1461, 0; xor.pred %p1462, %p1460, %p1461; not.pred %p1463, %p1462; ld.const.u64 %rd3021, [kernel_params+3752]; cvta.to.global.u64 %rd2166, %rd3021; shl.b64 %rd2167, %rd439, 2; add.s64 %rd2168, %rd2166, %rd2167; ld.global.u32 %r3241, [%rd2168]; @%p1463 bra $L__BB108_1650; bra.uni $L__BB108_1644; $L__BB108_1650: mov.u64 %rd2199, kernel_params; add.s64 %rd3027, %rd2199, 3744; setp.eq.s32 %p1470, %r3241, -1; mov.u32 %r3246, 0; @%p1470 bra $L__BB108_1654; add.s64 %rd3027, %rd2199, 3744; mov.u32 %r3246, 0; $L__BB108_1652: ld.const.u64 %rd2202, [%rd3027]; cvta.to.global.u64 %rd2203, %rd2202; add.s64 %rd2205, %rd2203, %rd2167; ld.global.u32 %r2290, [%rd2205]; setp.eq.s32 %p1471, %r2290, %r757; setp.eq.s32 %p1472, %r3241, %r729; and.pred %p1473, %p1471, %p1472; @%p1473 bra $L__BB108_1656; add.s32 %r3246, %r3246, 1; mul.wide.s32 %rd2206, %r3246, 16; add.s64 %rd2208, %rd2199, %rd2206; add.s64 %rd3027, %rd2208, 3744; ld.const.u64 %rd3021, [%rd2208+3752]; cvta.to.global.u64 %rd2209, %rd3021; add.s64 %rd2211, %rd2209, %rd2167; ld.global.u32 %r3241, [%rd2211]; setp.ne.s32 %p1474, %r3241, -1; @%p1474 bra $L__BB108_1652; $L__BB108_1654: ld.const.u32 %r2291, [kernel_params+12]; add.s32 %r2292, %r2291, -1; setp.ge.u32 %p1475, %r3246, %r2292; @%p1475 bra $L__BB108_1656; ld.const.u64 %rd2212, [%rd3027]; cvta.to.global.u64 %rd2213, %rd2212; add.s64 %rd2215, %rd2213, %rd2167; st.global.u32 [%rd2215], %r757; cvta.to.global.u64 %rd2216, %rd3021; add.s64 %rd2217, %rd2216, %rd2167; st.global.u32 [%rd2217], %r729; add.s32 %r2293, %r3246, 1; mul.wide.s32 %rd2218, %r2293, 16; add.s64 %rd2220, %rd2199, %rd2218; ld.const.v2.u64 {%rd2221, %rd2222}, [%rd2220+3744]; cvta.to.global.u64 %rd2225, %rd2221; add.s64 %rd2226, %rd2225, %rd2167; mov.u32 %r2294, -1; st.global.u32 [%rd2226], %r2294; cvta.to.global.u64 %rd2227, %rd2222; add.s64 %rd2228, %rd2227, %rd2167; st.global.u32 [%rd2228], %r2294; bra.uni $L__BB108_1656; $L__BB108_1644: setp.eq.s32 %p1464, %r3241, -1; @%p1464 bra $L__BB108_1656; mov.u64 %rd2170, kernel_params; add.s64 %rd3022, %rd2170, 3744; mov.u32 %r3242, 0; $L__BB108_1646: ld.const.u64 %rd2171, [%rd3022]; cvta.to.global.u64 %rd2172, %rd2171; add.s64 %rd2174, %rd2172, %rd2167; ld.global.u32 %r2285, [%rd2174]; setp.eq.s32 %p1465, %r2285, %r757; setp.eq.s32 %p1466, %r3241, %r729; and.pred %p1467, %p1465, %p1466; add.s32 %r770, %r3242, 1; mul.wide.s32 %rd2175, %r770, 16; add.s64 %rd2177, %rd2170, %rd2175; add.s64 %rd443, %rd2177, 3752; @%p1467 bra $L__BB108_1648; add.s64 %rd3022, %rd2177, 3744; ld.const.u64 %rd3021, [%rd443]; cvta.to.global.u64 %rd2181, %rd3021; add.s64 %rd2183, %rd2181, %rd2167; ld.global.u32 %r3241, [%rd2183]; setp.eq.s32 %p1468, %r3241, -1; mov.u32 %r3242, %r770; @%p1468 bra $L__BB108_1656; bra.uni $L__BB108_1646; $L__BB108_1660: ld.local.v4.f32 {%f16285, %f16286, %f16287, %f16288}, [%rd7+104]; ld.local.v4.f32 {%f16289, %f16290, %f16291, %f16292}, [%rd7+120]; ld.local.v4.f32 {%f16293, %f16294, %f16295, %f16296}, [%rd7+136]; $L__BB108_1662: mul.ftz.f32 %f10595, %f16279, %f16286; fma.rn.ftz.f32 %f10596, %f16278, %f16285, %f10595; fma.rn.ftz.f32 %f10597, %f16280, %f16287, %f10596; mul.ftz.f32 %f10598, %f16279, %f16290; fma.rn.ftz.f32 %f10599, %f16278, %f16289, %f10598; fma.rn.ftz.f32 %f10600, %f16280, %f16291, %f10599; mul.ftz.f32 %f10601, %f16279, %f16294; fma.rn.ftz.f32 %f10602, %f16278, %f16293, %f10601; fma.rn.ftz.f32 %f10603, %f16280, %f16295, %f10602; add.ftz.f32 %f16297, %f16288, %f10597; add.ftz.f32 %f16298, %f16292, %f10600; add.ftz.f32 %f16299, %f16296, %f10603; mul.ftz.f32 %f10604, %f16282, %f16286; fma.rn.ftz.f32 %f10605, %f16281, %f16285, %f10604; mul.ftz.f32 %f10606, %f16282, %f16290; fma.rn.ftz.f32 %f10607, %f16281, %f16289, %f10606; mul.ftz.f32 %f10608, %f16282, %f16294; fma.rn.ftz.f32 %f10609, %f16281, %f16293, %f10608; fma.rn.ftz.f32 %f16300, %f16283, %f16287, %f10605; fma.rn.ftz.f32 %f16301, %f16283, %f16291, %f10607; fma.rn.ftz.f32 %f16302, %f16283, %f16295, %f10609; $L__BB108_1663: sub.ftz.f32 %f10610, %f2324, %f16297; sub.ftz.f32 %f10611, %f2325, %f16298; sub.ftz.f32 %f10612, %f2326, %f16299; sub.ftz.f32 %f10613, %f2327, %f16297; sub.ftz.f32 %f10614, %f2328, %f16298; sub.ftz.f32 %f10615, %f2329, %f16299; sub.ftz.f32 %f10616, %f2330, %f16297; sub.ftz.f32 %f10617, %f2331, %f16298; sub.ftz.f32 %f10618, %f2332, %f16299; sub.ftz.f32 %f10619, %f10616, %f10610; sub.ftz.f32 %f10620, %f10617, %f10611; sub.ftz.f32 %f10621, %f10618, %f10612; sub.ftz.f32 %f10622, %f10610, %f10613; sub.ftz.f32 %f10623, %f10611, %f10614; sub.ftz.f32 %f10624, %f10612, %f10615; sub.ftz.f32 %f10625, %f10613, %f10616; sub.ftz.f32 %f10626, %f10614, %f10617; sub.ftz.f32 %f10627, %f10615, %f10618; add.ftz.f32 %f10628, %f10616, %f10610; add.ftz.f32 %f10629, %f10617, %f10611; add.ftz.f32 %f10630, %f10618, %f10612; mul.ftz.f32 %f10631, %f10629, %f10621; mul.ftz.f32 %f10632, %f10630, %f10620; sub.ftz.f32 %f10633, %f10631, %f10632; mul.ftz.f32 %f10634, %f10630, %f10619; mul.ftz.f32 %f10635, %f10628, %f10621; sub.ftz.f32 %f10636, %f10634, %f10635; mul.ftz.f32 %f10637, %f10628, %f10620; mul.ftz.f32 %f10638, %f10629, %f10619; sub.ftz.f32 %f10639, %f10637, %f10638; mul.ftz.f32 %f10640, %f16301, %f10636; fma.rn.ftz.f32 %f10641, %f16300, %f10633, %f10640; fma.rn.ftz.f32 %f10642, %f16302, %f10639, %f10641; add.ftz.f32 %f10643, %f10610, %f10613; add.ftz.f32 %f10644, %f10611, %f10614; add.ftz.f32 %f10645, %f10612, %f10615; mul.ftz.f32 %f10646, %f10644, %f10624; mul.ftz.f32 %f10647, %f10645, %f10623; sub.ftz.f32 %f10648, %f10646, %f10647; mul.ftz.f32 %f10649, %f10645, %f10622; mul.ftz.f32 %f10650, %f10643, %f10624; sub.ftz.f32 %f10651, %f10649, %f10650; mul.ftz.f32 %f10652, %f10643, %f10623; mul.ftz.f32 %f10653, %f10644, %f10622; sub.ftz.f32 %f10654, %f10652, %f10653; mul.ftz.f32 %f10655, %f16301, %f10651; fma.rn.ftz.f32 %f10656, %f16300, %f10648, %f10655; fma.rn.ftz.f32 %f10657, %f16302, %f10654, %f10656; add.ftz.f32 %f10658, %f10613, %f10616; add.ftz.f32 %f10659, %f10614, %f10617; add.ftz.f32 %f10660, %f10615, %f10618; mul.ftz.f32 %f10661, %f10659, %f10627; mul.ftz.f32 %f10662, %f10660, %f10626; sub.ftz.f32 %f10663, %f10661, %f10662; mul.ftz.f32 %f10664, %f10660, %f10625; mul.ftz.f32 %f10665, %f10658, %f10627; sub.ftz.f32 %f10666, %f10664, %f10665; mul.ftz.f32 %f10667, %f10658, %f10626; mul.ftz.f32 %f10668, %f10659, %f10625; sub.ftz.f32 %f10669, %f10667, %f10668; mul.ftz.f32 %f10670, %f16301, %f10666; fma.rn.ftz.f32 %f10671, %f16300, %f10663, %f10670; fma.rn.ftz.f32 %f10672, %f16302, %f10669, %f10671; add.ftz.f32 %f10673, %f10642, %f10657; add.ftz.f32 %f10674, %f10672, %f10673; abs.ftz.f32 %f10675, %f10674; mul.ftz.f32 %f10676, %f10675, 0f34000000; min.ftz.f32 %f10677, %f10657, %f10672; min.ftz.f32 %f10678, %f10642, %f10677; max.ftz.f32 %f10679, %f10657, %f10672; max.ftz.f32 %f10680, %f10642, %f10679; setp.ge.ftz.f32 %p1480, %f10678, %f10676; mul.ftz.f32 %f10681, %f10675, 0fB4000000; setp.le.ftz.f32 %p1481, %f10680, %f10681; or.pred %p1482, %p1480, %p1481; @%p1482 bra $L__BB108_1665; ld.local.v2.f32 {%f10682, %f10683}, [%rd7+-64]; mul.ftz.f32 %f10686, %f10682, 0f43800000; cvt.rzi.ftz.s32.f32 %r2300, %f10686; mul.ftz.f32 %f10687, %f10683, 0f43800000; cvt.rzi.ftz.s32.f32 %r2301, %f10687; ld.local.f32 %f10688, [%rd7+-56]; mul.ftz.f32 %f10689, %f10688, 0f43800000; cvt.rzi.ftz.s32.f32 %r2302, %f10689; neg.s32 %r2303, %r2300; setp.lt.ftz.f32 %p1483, %f16278, 0f00000000; selp.b32 %r2304, %r2303, %r2300, %p1483; mov.b32 %r2305, %f16278; add.s32 %r2306, %r2304, %r2305; mov.b32 %f10690, %r2306; neg.s32 %r2307, %r2301; setp.lt.ftz.f32 %p1484, %f16279, 0f00000000; selp.b32 %r2308, %r2307, %r2301, %p1484; mov.b32 %r2309, %f16279; add.s32 %r2310, %r2308, %r2309; mov.b32 %f10691, %r2310; neg.s32 %r2311, %r2302; setp.lt.ftz.f32 %p1485, %f16280, 0f00000000; selp.b32 %r2312, %r2311, %r2302, %p1485; mov.b32 %r2313, %f16280; add.s32 %r2314, %r2312, %r2313; mov.b32 %f10692, %r2314; abs.ftz.f32 %f10693, %f16278; setp.lt.ftz.f32 %p1486, %f10693, 0f3D000000; fma.rn.ftz.f32 %f10694, %f10682, 0f37800000, %f16278; abs.ftz.f32 %f10695, %f16279; setp.lt.ftz.f32 %p1487, %f10695, 0f3D000000; fma.rn.ftz.f32 %f10696, %f10683, 0f37800000, %f16279; abs.ftz.f32 %f10697, %f16280; setp.lt.ftz.f32 %p1488, %f10697, 0f3D000000; fma.rn.ftz.f32 %f10698, %f10688, 0f37800000, %f16280; selp.f32 %f16280, %f10698, %f10692, %p1488; selp.f32 %f16279, %f10696, %f10691, %p1487; selp.f32 %f16278, %f10694, %f10690, %p1486; $L__BB108_1665: ld.const.u64 %rd2245, [kernel_params+3496]; cvta.to.global.u64 %rd2246, %rd2245; mul.lo.s64 %rd2247, %rd439, 12; add.s64 %rd2248, %rd2246, %rd2247; st.global.f32 [%rd2248], %f16278; st.global.f32 [%rd2248+4], %f16279; st.global.f32 [%rd2248+8], %f16280; ld.const.v2.u64 {%rd2249, %rd2250}, [kernel_params+3504]; mov.u32 %r2315, 0; cvta.to.global.u64 %rd2253, %rd2249; add.s64 %rd2254, %rd2253, %rd2247; st.global.f32 [%rd2254], %f16281; st.global.f32 [%rd2254+4], %f16282; st.global.f32 [%rd2254+8], %f16283; cvta.to.global.u64 %rd2255, %rd2250; shl.b64 %rd2256, %rd439, 2; add.s64 %rd2257, %rd2255, %rd2256; st.global.u32 [%rd2257], %r2315; ld.const.v2.u64 {%rd2258, %rd2259}, [kernel_params+3520]; cvta.to.global.u64 %rd2262, %rd2258; add.s64 %rd2263, %rd2262, %rd2256; st.global.f32 [%rd2263], %f16284; cvta.to.global.u64 %rd2264, %rd2259; add.s64 %rd2265, %rd2264, %rd2256; st.global.f32 [%rd2265], %f2320; ld.const.u64 %rd2266, [kernel_params+3536]; cvta.to.global.u64 %rd2267, %rd2266; add.s64 %rd2268, %rd2267, %rd2256; st.global.f32 [%rd2268], %f2319; ld.const.u64 %rd2269, [kernel_params+3584]; cvta.to.global.u64 %rd2270, %rd2269; add.s64 %rd2271, %rd2270, %rd2256; st.global.u32 [%rd2271], %r758; ld.const.u64 %rd2272, [kernel_params+3576]; cvta.to.global.u64 %rd2273, %rd2272; add.s64 %rd2274, %rd2273, %rd2256; st.global.u32 [%rd2274], %r780; ld.const.u64 %rd2275, [kernel_params+3632]; cvta.to.global.u64 %rd2276, %rd2275; add.s64 %rd2277, %rd2276, %rd2256; st.global.u32 [%rd2277], %r760; ld.const.u64 %rd2278, [kernel_params+3624]; cvta.to.global.u64 %rd2279, %rd2278; add.s64 %rd2280, %rd2279, %rd2256; st.global.u32 [%rd2280], %r761; setp.ne.s32 %p1489, %r761, -1; setp.eq.s32 %p1490, %r762, 2; or.pred %p1491, %p1489, %p1490; selp.b32 %r2316, 0, 134217728, %p1491; ld.global.u32 %r782, [%rd29]; or.b32 %r3248, %r782, %r2316; ld.const.u64 %rd2281, [kernel_params+2584]; cvta.to.global.u64 %rd2282, %rd2281; mul.lo.s64 %rd2283, %rd2, 12; add.s64 %rd2284, %rd2282, %rd2283; ld.global.f32 %f10699, [%rd2284]; mul.ftz.f32 %f2396, %f10699, %f16249; ld.global.f32 %f10700, [%rd2284+4]; mul.ftz.f32 %f2397, %f10700, %f16248; ld.global.f32 %f10701, [%rd2284+8]; mul.ftz.f32 %f2398, %f10701, %f16247; and.b32 %r2317, %r763, 2097152; setp.eq.s32 %p1492, %r2317, 0; @%p1492 bra $L__BB108_1682; and.b32 %r2318, %r782, 100663296; setp.eq.s32 %p1493, %r2318, 0; @%p1493 bra $L__BB108_1668; ld.const.u64 %rd2285, [kernel_params+2600]; cvta.to.global.u64 %rd2286, %rd2285; add.s64 %rd2288, %rd2286, %rd2283; ld.global.f32 %f16317, [%rd2288+8]; ld.global.f32 %f16316, [%rd2288+4]; ld.global.f32 %f16315, [%rd2288]; ld.const.u64 %rd2289, [kernel_params+2608]; cvta.to.global.u64 %rd2290, %rd2289; add.s64 %rd2291, %rd2290, %rd2283; ld.global.f32 %f16320, [%rd2291+8]; ld.global.f32 %f16319, [%rd2291+4]; ld.global.f32 %f16318, [%rd2291]; bra.uni $L__BB108_1681; $L__BB108_1668: setp.eq.ftz.f32 %p1494, %f16249, 0f00000000; mov.f32 %f16320, 0f00000000; mov.f32 %f16315, %f16320; @%p1494 bra $L__BB108_1670; div.approx.ftz.f32 %f16315, %f16244, %f16249; $L__BB108_1670: setp.eq.ftz.f32 %p1495, %f16248, 0f00000000; mov.f32 %f16316, %f16320; @%p1495 bra $L__BB108_1672; div.approx.ftz.f32 %f16316, %f16246, %f16248; $L__BB108_1672: setp.eq.ftz.f32 %p1496, %f16247, 0f00000000; mov.f32 %f16317, %f16320; @%p1496 bra $L__BB108_1674; div.approx.ftz.f32 %f16317, %f16245, %f16247; $L__BB108_1674: mov.f32 %f16318, %f16320; @%p1494 bra $L__BB108_1676; div.approx.ftz.f32 %f16318, %f16241, %f16249; $L__BB108_1676: mov.f32 %f16319, %f16320; @%p1495 bra $L__BB108_1678; div.approx.ftz.f32 %f16319, %f16243, %f16248; $L__BB108_1678: @%p1496 bra $L__BB108_1680; div.approx.ftz.f32 %f16320, %f16242, %f16247; $L__BB108_1680: or.b32 %r3248, %r3248, 33554432; $L__BB108_1681: ld.const.u64 %rd2292, [kernel_params+3448]; cvta.to.global.u64 %rd2293, %rd2292; add.s64 %rd2295, %rd2293, %rd2247; st.global.f32 [%rd2295], %f16315; st.global.f32 [%rd2295+4], %f16316; st.global.f32 [%rd2295+8], %f16317; ld.const.u64 %rd2296, [kernel_params+3456]; cvta.to.global.u64 %rd2297, %rd2296; add.s64 %rd2298, %rd2297, %rd2247; st.global.f32 [%rd2298], %f16318; st.global.f32 [%rd2298+4], %f16319; st.global.f32 [%rd2298+8], %f16320; $L__BB108_1682: ld.const.u64 %rd2299, [kernel_params+2448]; cvta.to.global.u64 %rd2300, %rd2299; add.s64 %rd2302, %rd2300, %rd837; ld.global.u32 %r2319, [%rd2302]; ld.const.v2.u64 {%rd2303, %rd2304}, [kernel_params+3344]; cvta.to.global.u64 %rd2307, %rd2303; add.s64 %rd2309, %rd2307, %rd2256; st.global.u32 [%rd2309], %r2319; ld.global.u16 %rs94, [%rd278]; ld.const.v2.u64 {%rd2310, %rd2311}, [kernel_params+3360]; cvta.to.global.u64 %rd2314, %rd2311; shl.b64 %rd2315, %rd439, 1; add.s64 %rd2316, %rd2314, %rd2315; st.global.u16 [%rd2316], %rs94; ld.global.u32 %r2320, [%rd277]; cvta.to.global.u64 %rd2317, %rd2310; add.s64 %rd2318, %rd2317, %rd2256; st.global.u32 [%rd2318], %r2320; ld.global.u32 %r2321, [%rd279]; cvta.to.global.u64 %rd2319, %rd2304; add.s64 %rd2320, %rd2319, %rd2256; st.global.u32 [%rd2320], %r2321; ld.const.v2.u64 {%rd2321, %rd2322}, [kernel_params+3424]; cvta.to.global.u64 %rd2325, %rd2321; add.s64 %rd2326, %rd2325, %rd2256; st.global.u32 [%rd2326], %r3248; ld.const.u64 %rd2327, [kernel_params+2472]; cvta.to.global.u64 %rd2328, %rd2327; add.s64 %rd2330, %rd2328, %rd1707; ld.global.u16 %rs95, [%rd2330]; ld.const.u64 %rd2331, [kernel_params+3384]; cvta.to.global.u64 %rd2332, %rd2331; add.s64 %rd2333, %rd2332, %rd2315; st.global.u16 [%rd2333], %rs95; ld.const.v2.u64 {%rd2334, %rd2335}, [kernel_params+2480]; cvta.to.global.u64 %rd2338, %rd2335; add.s64 %rd2339, %rd2338, %rd1707; ld.global.u16 %rs96, [%rd2339]; ld.const.v2.u64 {%rd2340, %rd2341}, [kernel_params+3392]; cvta.to.global.u64 %rd2344, %rd2341; add.s64 %rd2345, %rd2344, %rd2315; st.global.u16 [%rd2345], %rs96; ld.const.u64 %rd2346, [kernel_params+2496]; cvta.to.global.u64 %rd2347, %rd2346; add.s64 %rd2348, %rd2347, %rd1707; ld.global.u16 %rs97, [%rd2348]; ld.const.u64 %rd2349, [kernel_params+3408]; cvta.to.global.u64 %rd2350, %rd2349; add.s64 %rd2351, %rd2350, %rd2315; st.global.u16 [%rd2351], %rs97; cvta.to.global.u64 %rd2352, %rd2334; add.s64 %rd2353, %rd2352, %rd1707; ld.global.u16 %rs98, [%rd2353]; ld.const.v2.u64 {%rd2354, %rd2355}, [kernel_params+3376]; cvta.to.global.u64 %rd2357, %rd2340; add.s64 %rd2358, %rd2357, %rd2315; st.global.u16 [%rd2358], %rs98; ld.global.u16 %rs99, [%rd280]; cvta.to.global.u64 %rd2359, %rd2354; add.s64 %rd2360, %rd2359, %rd2315; st.global.u16 [%rd2360], %rs99; cvta.to.global.u64 %rd2361, %rd2322; add.s64 %rd2363, %rd2361, %rd2247; st.global.f32 [%rd2363], %f2396; st.global.f32 [%rd2363+4], %f2397; st.global.f32 [%rd2363+8], %f2398; and.b32 %r2322, %r763, 4194304; setp.eq.s32 %p1500, %r2322, 0; @%p1500 bra $L__BB108_1684; ld.const.u64 %rd2364, [kernel_params+3440]; cvta.to.global.u64 %rd2365, %rd2364; add.s64 %rd2367, %rd2365, %rd2247; st.global.f32 [%rd2367], %f2396; st.global.f32 [%rd2367+4], %f2397; st.global.f32 [%rd2367+8], %f2398; $L__BB108_1684: ld.const.u32 %r2323, [kernel_params+1076]; ld.local.u32 %r2324, [%rd282+48]; selp.b32 %r2325, %r2323, %r2324, %p1490; cvt.u16.u32 %rs100, %r2325; add.s16 %rs101, %rs100, 1; ld.const.u64 %rd2368, [kernel_params+3472]; cvta.to.global.u64 %rd2369, %rd2368; add.s64 %rd2370, %rd2369, %rd439; st.global.u8 [%rd2370], %rs101; $L__BB108_1685: ld.const.u32 %r787, [kernel_params]; and.b32 %r2326, %r787, 25165824; setp.eq.s32 %p1502, %r2326, 0; @%p1502 bra $L__BB108_1775; and.b32 %r788, %r787, 16777216; setp.ne.s32 %p1503, %r788, 0; @%p1503 bra $L__BB108_1688; ld.global.u32 %r2327, [%rd29]; and.b32 %r2328, %r2327, 1; setp.eq.b32 %p1504, %r2328, 1; mov.pred %p1505, 0; xor.pred %p1506, %p1504, %p1505; not.pred %p1507, %p1506; @%p1507 bra $L__BB108_1775; $L__BB108_1688: add.s32 %r789, %r602, 3; ld.const.u32 %r2329, [kernel_params+1596]; setp.eq.s32 %p1508, %r2329, 0; brev.b32 %r2330, %r603; mul.lo.s32 %r2331, %r2330, 1025551850; xor.b32 %r790, %r2331, %r2330; @%p1508 bra $L__BB108_1691; ld.const.v2.u32 {%r2332, %r2333}, [kernel_params+1600]; mov.b32 %f2429, %r2333; setp.geu.ftz.f32 %p1509, %f2429, 0f3F800000; setp.lt.ftz.f32 %p1510, %f2429, 0f3F800000; ld.const.u32 %r2336, [kernel_params+1576]; selp.b32 %r2337, %r2336, %r601, %p1510; add.s16 %rs102, %rs28, 3; and.b16 %rs103, %rs102, 255; cvt.u32.u16 %r2338, %rs103; xor.b32 %r2339, %r2337, %r2338; mul.lo.s32 %r2340, %r2339, -512718531; shr.u32 %r2341, %r2337, 16; xor.b32 %r2342, %r2340, %r2341; bfe.u32 %r2343, %r2342, 4, 4; shr.u32 %r2344, %r2337, 8; xor.b32 %r2345, %r2342, %r2344; xor.b32 %r2346, %r2345, %r2343; mul.lo.s32 %r2347, %r2346, 153742143; shr.u32 %r2348, %r2337, 23; xor.b32 %r2349, %r2347, %r2348; bfe.u32 %r2350, %r2349, 1, 7; xor.b32 %r2351, %r2350, %r2349; shr.u32 %r2352, %r2337, 27; or.b32 %r2353, %r2352, 1; mul.lo.s32 %r2354, %r2353, %r2351; mul.lo.s32 %r2355, %r2354, -1074439621; bfe.u32 %r2356, %r2355, 2, 6; xor.b32 %r2357, %r2356, %r2355; mul.lo.s32 %r2358, %r2357, -1638916925; bfe.u32 %r2359, %r2358, 2, 6; xor.b32 %r2360, %r2359, %r2358; mul.lo.s32 %r2361, %r2360, -933190689; and.b32 %r2362, %r2361, 255; bfe.u32 %r2363, %r2361, 5, 3; xor.b32 %r2364, %r2363, %r2362; add.s32 %r2365, %r2332, -1; xor.b32 %r791, %r789, 61; xor.b32 %r2366, %r2337, %r791; mul.lo.s32 %r2367, %r2366, 9; shr.u32 %r2368, %r2367, 4; xor.b32 %r2369, %r2368, %r2367; mul.lo.s32 %r2370, %r2369, 668265261; add.s32 %r2371, %r2370, %r790; shr.u32 %r2372, %r2370, 16; or.b32 %r2373, %r2372, 1; mul.lo.s32 %r2374, %r2373, %r2371; mul.lo.s32 %r2375, %r2374, 89287766; xor.b32 %r2376, %r2375, %r2374; mul.lo.s32 %r2377, %r2376, 1403136100; xor.b32 %r2378, %r2377, %r2376; brev.b32 %r2379, %r2378; neg.s32 %r2380, %r2332; and.b32 %r2381, %r603, %r2380; and.b32 %r2382, %r2379, %r2365; or.b32 %r2383, %r2382, %r2381; mad.lo.s32 %r2384, %r2364, %r2332, %r2383; shl.b32 %r2385, %r2332, 8; rem.u32 %r2386, %r2384, %r2385; ld.const.u64 %rd2371, [kernel_params+2424]; cvta.to.global.u64 %rd2372, %rd2371; shl.b32 %r2387, %r2386, 1; mul.wide.u32 %rd2373, %r2387, 4; add.s64 %rd2374, %rd2372, %rd2373; ld.global.f32 %f16323, [%rd2374]; ld.global.f32 %f16324, [%rd2374+4]; @%p1509 bra $L__BB108_1695; xor.b32 %r2388, %r791, %r601; mul.lo.s32 %r2389, %r2388, 9; shr.u32 %r2390, %r2389, 4; xor.b32 %r2391, %r2390, %r2389; mul.lo.s32 %r2392, %r2391, 668265261; cvt.rn.f32.u32 %f10708, %r2392; mul.ftz.f32 %f10709, %f10708, 0f2F7FFFFE; xor.b32 %r2393, %r601, %r789; xor.b32 %r2394, %r2393, -905047700; mul.lo.s32 %r2395, %r2394, 9; shr.u32 %r2396, %r2395, 4; xor.b32 %r2397, %r2396, %r2395; mul.lo.s32 %r2398, %r2397, 668265261; cvt.rn.f32.u32 %f10710, %r2398; mul.ftz.f32 %f10711, %f10710, 0f2F7FFFFE; fma.rn.ftz.f32 %f10712, %f10709, %f2429, %f16323; fma.rn.ftz.f32 %f10713, %f10711, %f2429, %f16324; cvt.rmi.ftz.f32.f32 %f10714, %f10712; sub.ftz.f32 %f16323, %f10712, %f10714; cvt.rmi.ftz.f32.f32 %f10715, %f10713; sub.ftz.f32 %f16324, %f10713, %f10715; bra.uni $L__BB108_1695; $L__BB108_1691: shr.u32 %r2400, %r789, 16; xor.b32 %r2401, %r2400, %r789; mul.lo.s32 %r2402, %r2401, 569420461; shr.u32 %r2403, %r2402, 15; xor.b32 %r2404, %r2403, %r2402; mul.lo.s32 %r2405, %r2404, -749064809; shr.u32 %r2406, %r2405, 15; xor.b32 %r2407, %r601, %r2405; xor.b32 %r2408, %r2407, %r2406; xor.b32 %r792, %r2408, -419546133; xor.b32 %r2409, %r2408, 508809841; add.s32 %r2410, %r2409, %r790; shr.u32 %r2411, %r2409, 16; or.b32 %r2412, %r2411, 1; mul.lo.s32 %r2413, %r2412, %r2410; mul.lo.s32 %r2414, %r2413, 89287766; xor.b32 %r2415, %r2414, %r2413; mul.lo.s32 %r2416, %r2415, 1403136100; xor.b32 %r793, %r2416, %r2415; setp.eq.s32 %p1511, %r793, 0; mov.u32 %r3252, 0; @%p1511 bra $L__BB108_1694; mov.u32 %r3249, 0; mov.u64 %rd2376, sobol_burley_table; mov.u32 %r3250, %r793; mov.u32 %r3252, %r3249; $L__BB108_1693: clz.b32 %r2419, %r3250; add.s32 %r2420, %r2419, %r3249; mul.wide.u32 %rd2375, %r2420, 4; add.s64 %rd2377, %rd2376, %rd2375; ld.const.u32 %r2421, [%rd2377+128]; xor.b32 %r3252, %r2421, %r3252; add.s32 %r3249, %r2420, 1; shl.b32 %r2422, %r3250, %r2419; shl.b32 %r3250, %r2422, 1; setp.ne.s32 %p1512, %r3250, 0; @%p1512 bra $L__BB108_1693; $L__BB108_1694: xor.b32 %r2423, %r792, -525684874; brev.b32 %r2424, %r793; shr.u32 %r2425, %r2423, 16; or.b32 %r2426, %r2425, 1; mul.lo.s32 %r2427, %r2424, 1025551850; xor.b32 %r2428, %r2427, %r2424; add.s32 %r2429, %r2428, %r2423; mul.lo.s32 %r2430, %r2429, %r2426; mul.lo.s32 %r2431, %r2430, 89287766; xor.b32 %r2432, %r2431, %r2430; xor.b32 %r2433, %r792, -1802089138; mul.lo.s32 %r2434, %r3252, 1025551850; xor.b32 %r2435, %r2434, %r3252; add.s32 %r2436, %r2435, %r2433; shr.u32 %r2437, %r2433, 16; or.b32 %r2438, %r2437, 1; mul.lo.s32 %r2439, %r2436, %r2438; mul.lo.s32 %r2440, %r2439, 89287766; xor.b32 %r2441, %r2440, %r2439; mul.lo.s32 %r2442, %r2441, 1403136100; xor.b32 %r2443, %r2442, %r2441; brev.b32 %r2444, %r2443; cvt.rn.f32.u32 %f10716, %r2444; mul.lo.s32 %r2445, %r2432, 1403136100; xor.b32 %r2446, %r2445, %r2432; brev.b32 %r2447, %r2446; cvt.rn.f32.u32 %f10717, %r2447; mul.ftz.f32 %f16324, %f10716, 0f2F7FFFFE; mul.ftz.f32 %f16323, %f10717, 0f2F7FFFFE; $L__BB108_1695: ld.const.f32 %f2442, [kernel_params+1548]; ld.local.u32 %r801, [%rd7+172]; setp.lt.s32 %p1513, %r801, 1; mov.f32 %f16331, 0f00000000; mov.f32 %f16332, %f16331; mov.f32 %f16333, %f16331; mov.f32 %f16334, %f16331; mov.f32 %f16335, %f16331; mov.f32 %f16336, %f16331; @%p1513 bra $L__BB108_1712; add.s32 %r2449, %r801, -1; and.b32 %r3256, %r801, 3; setp.lt.u32 %p1514, %r2449, 3; mov.f32 %f16336, 0f00000000; mov.u32 %r3255, 0; mov.f32 %f16335, %f16336; mov.f32 %f16334, %f16336; mov.f32 %f16333, %f16336; mov.f32 %f16332, %f16336; mov.f32 %f16331, %f16336; @%p1514 bra $L__BB108_1707; sub.s32 %r3254, %r801, %r3256; mov.f32 %f16336, 0f00000000; mov.u32 %r3255, 0; mov.f32 %f16335, %f16336; mov.f32 %f16334, %f16336; mov.f32 %f16333, %f16336; mov.f32 %f16332, %f16336; mov.f32 %f16331, %f16336; $L__BB108_1698: cvt.s64.s32 %rd456, %r3255; mul.wide.s32 %rd2378, %r3255, 80; add.s64 %rd2379, %rd1, %rd2378; add.s64 %rd457, %rd2379, 316; ld.local.u32 %r2451, [%rd2379+316]; add.s32 %r2452, %r2451, -2; setp.gt.u32 %p1515, %r2452, 6; @%p1515 bra $L__BB108_1700; ld.local.v4.f32 {%f10737, %f10738, %f10739, %f10740}, [%rd457+-12]; fma.rn.ftz.f32 %f16334, %f2442, %f10737, %f16334; fma.rn.ftz.f32 %f16335, %f2442, %f10738, %f16335; fma.rn.ftz.f32 %f16336, %f2442, %f10739, %f16336; add.ftz.f32 %f10744, %f10737, %f10738; add.ftz.f32 %f10745, %f10739, %f10744; mul.ftz.f32 %f10746, %f10745, 0f3EAAAAAB; abs.ftz.f32 %f10747, %f10746; ld.local.f32 %f10748, [%rd457+8]; ld.local.f32 %f10749, [%rd457+12]; ld.local.f32 %f10750, [%rd457+16]; fma.rn.ftz.f32 %f16331, %f10747, %f10748, %f16331; fma.rn.ftz.f32 %f16332, %f10747, %f10749, %f16332; fma.rn.ftz.f32 %f16333, %f10747, %f10750, %f16333; $L__BB108_1700: ld.local.u32 %r2453, [%rd457+80]; add.s32 %r2454, %r2453, -2; setp.gt.u32 %p1516, %r2454, 6; @%p1516 bra $L__BB108_1702; ld.local.v4.f32 {%f10751, %f10752, %f10753, %f10754}, [%rd457+68]; fma.rn.ftz.f32 %f16334, %f2442, %f10751, %f16334; fma.rn.ftz.f32 %f16335, %f2442, %f10752, %f16335; fma.rn.ftz.f32 %f16336, %f2442, %f10753, %f16336; add.ftz.f32 %f10758, %f10751, %f10752; add.ftz.f32 %f10759, %f10753, %f10758; mul.ftz.f32 %f10760, %f10759, 0f3EAAAAAB; abs.ftz.f32 %f10761, %f10760; ld.local.f32 %f10762, [%rd457+88]; ld.local.f32 %f10763, [%rd457+92]; ld.local.f32 %f10764, [%rd457+96]; fma.rn.ftz.f32 %f16331, %f10761, %f10762, %f16331; fma.rn.ftz.f32 %f16332, %f10761, %f10763, %f16332; fma.rn.ftz.f32 %f16333, %f10761, %f10764, %f16333; $L__BB108_1702: ld.local.u32 %r2455, [%rd457+160]; add.s32 %r2456, %r2455, -2; setp.gt.u32 %p1517, %r2456, 6; @%p1517 bra $L__BB108_1704; ld.local.v4.f32 {%f10765, %f10766, %f10767, %f10768}, [%rd457+148]; fma.rn.ftz.f32 %f16334, %f2442, %f10765, %f16334; fma.rn.ftz.f32 %f16335, %f2442, %f10766, %f16335; fma.rn.ftz.f32 %f16336, %f2442, %f10767, %f16336; add.ftz.f32 %f10772, %f10765, %f10766; add.ftz.f32 %f10773, %f10767, %f10772; mul.ftz.f32 %f10774, %f10773, 0f3EAAAAAB; abs.ftz.f32 %f10775, %f10774; ld.local.f32 %f10776, [%rd457+168]; ld.local.f32 %f10777, [%rd457+172]; ld.local.f32 %f10778, [%rd457+176]; fma.rn.ftz.f32 %f16331, %f10775, %f10776, %f16331; fma.rn.ftz.f32 %f16332, %f10775, %f10777, %f16332; fma.rn.ftz.f32 %f16333, %f10775, %f10778, %f16333; $L__BB108_1704: ld.local.u32 %r2457, [%rd457+240]; add.s32 %r2458, %r2457, -2; setp.gt.u32 %p1518, %r2458, 6; @%p1518 bra $L__BB108_1706; ld.local.v4.f32 {%f10779, %f10780, %f10781, %f10782}, [%rd457+228]; fma.rn.ftz.f32 %f16334, %f2442, %f10779, %f16334; fma.rn.ftz.f32 %f16335, %f2442, %f10780, %f16335; fma.rn.ftz.f32 %f16336, %f2442, %f10781, %f16336; add.ftz.f32 %f10786, %f10779, %f10780; add.ftz.f32 %f10787, %f10781, %f10786; mul.ftz.f32 %f10788, %f10787, 0f3EAAAAAB; abs.ftz.f32 %f10789, %f10788; ld.local.f32 %f10790, [%rd457+248]; ld.local.f32 %f10791, [%rd457+252]; ld.local.f32 %f10792, [%rd457+256]; fma.rn.ftz.f32 %f16331, %f10789, %f10790, %f16331; fma.rn.ftz.f32 %f16332, %f10789, %f10791, %f16332; fma.rn.ftz.f32 %f16333, %f10789, %f10792, %f16333; $L__BB108_1706: cvt.u32.u64 %r2459, %rd456; add.s32 %r3255, %r2459, 4; add.s32 %r3254, %r3254, -4; setp.ne.s32 %p1519, %r3254, 0; @%p1519 bra $L__BB108_1698; $L__BB108_1707: setp.eq.s32 %p1520, %r3256, 0; @%p1520 bra $L__BB108_1712; mul.wide.s32 %rd2380, %r3255, 80; add.s64 %rd2381, %rd1, %rd2380; add.s64 %rd3028, %rd2381, 332; $L__BB108_1709: .pragma "nounroll"; ld.local.u32 %r2460, [%rd3028+-16]; add.s32 %r2461, %r2460, -2; setp.gt.u32 %p1521, %r2461, 6; @%p1521 bra $L__BB108_1711; ld.local.f32 %f10793, [%rd3028+-28]; ld.local.f32 %f10794, [%rd3028+-24]; ld.local.f32 %f10795, [%rd3028+-20]; fma.rn.ftz.f32 %f16334, %f2442, %f10793, %f16334; fma.rn.ftz.f32 %f16335, %f2442, %f10794, %f16335; fma.rn.ftz.f32 %f16336, %f2442, %f10795, %f16336; add.ftz.f32 %f10796, %f10793, %f10794; add.ftz.f32 %f10797, %f10795, %f10796; mul.ftz.f32 %f10798, %f10797, 0f3EAAAAAB; abs.ftz.f32 %f10799, %f10798; ld.local.f32 %f10800, [%rd3028+-8]; ld.local.f32 %f10801, [%rd3028+-4]; ld.local.f32 %f10802, [%rd3028]; fma.rn.ftz.f32 %f16331, %f10799, %f10800, %f16331; fma.rn.ftz.f32 %f16332, %f10799, %f10801, %f16332; fma.rn.ftz.f32 %f16333, %f10799, %f10802, %f16333; $L__BB108_1711: add.s64 %rd3028, %rd3028, 80; add.s32 %r3256, %r3256, -1; setp.ne.s32 %p1522, %r3256, 0; @%p1522 bra $L__BB108_1709; $L__BB108_1712: setp.eq.ftz.f32 %p1523, %f16332, 0f00000000; setp.eq.ftz.f32 %p1524, %f16331, 0f00000000; and.pred %p1525, %p1524, %p1523; setp.eq.ftz.f32 %p1526, %f16333, 0f00000000; and.pred %p1527, %p1526, %p1525; @%p1527 bra $L__BB108_1714; bra.uni $L__BB108_1713; $L__BB108_1714: ld.local.v2.f32 {%f16386, %f16387}, [%rd7+-72]; ld.local.f32 %f16385, [%rd7+-76]; bra.uni $L__BB108_1715; $L__BB108_1713: mul.ftz.f32 %f10803, %f16332, %f16332; fma.rn.ftz.f32 %f10804, %f16331, %f16331, %f10803; fma.rn.ftz.f32 %f10805, %f16333, %f16333, %f10804; sqrt.approx.ftz.f32 %f10806, %f10805; div.approx.ftz.f32 %f16387, %f16333, %f10806; div.approx.ftz.f32 %f16386, %f16332, %f10806; div.approx.ftz.f32 %f16385, %f16331, %f10806; $L__BB108_1715: mul.ftz.f32 %f10809, %f16323, 0f40C90FDB; cos.approx.ftz.f32 %f10810, %f10809; sqrt.approx.ftz.f32 %f10811, %f16324; mul.ftz.f32 %f10812, %f10811, %f10810; sin.approx.ftz.f32 %f10813, %f10809; mul.ftz.f32 %f10814, %f10811, %f10813; mul.ftz.f32 %f10815, %f10812, %f10812; mov.f32 %f10816, 0f3F800000; sub.ftz.f32 %f10817, %f10816, %f10815; mul.ftz.f32 %f10818, %f10814, %f10814; sub.ftz.f32 %f10819, %f10817, %f10818; mov.f32 %f10820, 0f00000000; max.ftz.f32 %f10821, %f10819, %f10820; sqrt.approx.ftz.f32 %f10822, %f10821; setp.neu.ftz.f32 %p1528, %f16385, %f16386; setp.neu.ftz.f32 %p1529, %f16385, %f16387; or.pred %p1530, %p1528, %p1529; sub.ftz.f32 %f10823, %f16387, %f16386; neg.ftz.f32 %f10824, %f16386; neg.ftz.f32 %f10825, %f16387; selp.f32 %f10826, %f10825, %f16387, %p1530; add.ftz.f32 %f10827, %f16385, %f10826; selp.f32 %f10828, %f16386, %f10824, %p1530; sub.ftz.f32 %f10829, %f10828, %f16385; mul.ftz.f32 %f10830, %f10827, %f10827; fma.rn.ftz.f32 %f10831, %f10823, %f10823, %f10830; fma.rn.ftz.f32 %f10832, %f10829, %f10829, %f10831; sqrt.approx.ftz.f32 %f10833, %f10832; div.approx.ftz.f32 %f10834, %f10823, %f10833; div.approx.ftz.f32 %f10835, %f10827, %f10833; div.approx.ftz.f32 %f10836, %f10829, %f10833; mul.ftz.f32 %f10837, %f16386, %f10836; mul.ftz.f32 %f10838, %f16387, %f10835; sub.ftz.f32 %f10839, %f10837, %f10838; mul.ftz.f32 %f10840, %f16387, %f10834; mul.ftz.f32 %f10841, %f16385, %f10836; sub.ftz.f32 %f10842, %f10840, %f10841; mul.ftz.f32 %f10843, %f16385, %f10835; mul.ftz.f32 %f10844, %f16386, %f10834; sub.ftz.f32 %f10845, %f10843, %f10844; mul.ftz.f32 %f10846, %f10814, %f10839; mul.ftz.f32 %f10847, %f10814, %f10842; mul.ftz.f32 %f10848, %f10814, %f10845; fma.rn.ftz.f32 %f10849, %f10812, %f10834, %f10846; fma.rn.ftz.f32 %f10850, %f10812, %f10835, %f10847; fma.rn.ftz.f32 %f10851, %f10812, %f10836, %f10848; fma.rn.ftz.f32 %f2542, %f16385, %f10822, %f10849; fma.rn.ftz.f32 %f2543, %f16386, %f10822, %f10850; fma.rn.ftz.f32 %f2544, %f16387, %f10822, %f10851; ld.local.v4.f32 {%f16413, %f16414, %f16415, %f10855}, [%rd1]; ld.local.u32 %r811, [%rd7+-24]; and.b32 %r812, %r811, 1; setp.eq.s32 %p1531, %r812, 0; ld.local.u32 %r2462, [%rd7+-40]; setp.gt.s32 %p1532, %r2462, -1; mov.u16 %rs154, 1; or.pred %p1533, %p1531, %p1532; @%p1533 bra $L__BB108_1758; ld.const.u64 %rd2384, [kernel_params+2160]; cvta.to.global.u64 %rd2385, %rd2384; ld.local.u32 %r2463, [%rd7+-12]; mul.wide.s32 %rd2386, %r2463, 208; add.s64 %rd2387, %rd2385, %rd2386; add.s64 %rd461, %rd2387, 184; ld.global.f32 %f2548, [%rd2387+184]; setp.leu.ftz.f32 %p1534, %f2548, 0f00000000; @%p1534 bra $L__BB108_1758; ld.local.f32 %f10856, [%rd7+-76]; ld.local.v2.f32 {%f10857, %f10858}, [%rd7+-72]; mul.ftz.f32 %f10861, %f2543, %f10857; fma.rn.ftz.f32 %f10862, %f2542, %f10856, %f10861; fma.rn.ftz.f32 %f10863, %f2544, %f10858, %f10862; setp.lt.ftz.f32 %p1535, %f10863, 0f00000000; neg.ftz.f32 %f10864, %f10863; selp.f32 %f2549, %f10864, %f10863, %p1535; @%p1535 bra $L__BB108_1719; bra.uni $L__BB108_1718; $L__BB108_1719: ld.local.v2.f32 {%f10867, %f10868}, [%rd7+-64]; neg.ftz.f32 %f16388, %f10867; neg.ftz.f32 %f16389, %f10868; ld.local.f32 %f10871, [%rd7+-56]; neg.ftz.f32 %f16390, %f10871; bra.uni $L__BB108_1720; $L__BB108_1718: ld.local.v2.f32 {%f16388, %f16389}, [%rd7+-64]; ld.local.f32 %f16390, [%rd7+-56]; $L__BB108_1720: mul.ftz.f32 %f10872, %f2543, %f16389; fma.rn.ftz.f32 %f10873, %f2542, %f16388, %f10872; fma.rn.ftz.f32 %f2559, %f2544, %f16390, %f10873; add.ftz.f32 %f10874, %f2549, %f2559; setp.lt.ftz.f32 %p1536, %f2549, %f2548; selp.f32 %f10875, 0f40000000, 0f3F800000, %p1536; mov.f32 %f10876, 0f3F800000; selp.f32 %f10877, %f10874, %f2559, %p1536; div.approx.ftz.f32 %f10878, %f10877, %f2548; sub.ftz.f32 %f10879, %f10875, %f10878; mov.f32 %f10880, 0f00000000; max.ftz.f32 %f10881, %f10879, %f10880; min.ftz.f32 %f2560, %f10881, %f10876; setp.leu.ftz.f32 %p1537, %f2560, 0f00000000; @%p1537 bra $L__BB108_1758; setp.eq.s32 %p1538, %r811, 65; ld.local.s32 %rd462, [%rd7+-28]; @%p1538 bra $L__BB108_1723; bra.uni $L__BB108_1722; $L__BB108_1723: ld.global.u32 %r813, [%rd461+-28]; ld.global.u32 %r814, [%rd461+-32]; shl.b32 %r2475, %r814, 1; cvt.rn.f32.s32 %f10882, %r2475; ld.local.f32 %f10883, [%rd7+-4]; mul.ftz.f32 %f10884, %f10883, %f10882; cvt.rzi.ftz.s32.f32 %r2476, %f10884; add.s32 %r2477, %r2475, -1; min.s32 %r815, %r2476, %r2477; cvt.rn.f32.s32 %f10885, %r815; sub.ftz.f32 %f2579, %f10884, %f10885; ld.const.u64 %rd2410, [kernel_params+2312]; cvta.to.global.u64 %rd463, %rd2410; ld.global.u32 %r3263, [%rd461+-20]; mul.wide.u32 %rd2411, %r3263, 16; add.s64 %rd2412, %rd463, %rd2411; ld.global.u16 %rs152, [%rd2412+12]; ld.global.u32 %r3262, [%rd2412+8]; ld.global.u64 %rd3048, [%rd2412]; setp.eq.s64 %p1539, %rd3048, 11; mov.u16 %rs150, %rs152; mov.u32 %r3257, %r3262; @%p1539 bra $L__BB108_1729; mov.u64 %rd3029, %rd3048; mov.u32 %r3257, %r3262; mov.u16 %rs150, %rs152; mov.u32 %r3258, %r3263; $L__BB108_1725: setp.eq.s64 %p1540, %rd3029, 0; @%p1540 bra $L__BB108_1727; add.s32 %r3258, %r3258, 2; bra.uni $L__BB108_1728; $L__BB108_1727: setp.eq.s16 %p1541, %rs150, 0; mov.u32 %r3261, -1; mov.u32 %r3258, %r3257; @%p1541 bra $L__BB108_1730; $L__BB108_1728: mul.wide.u32 %rd2413, %r3258, 16; add.s64 %rd2414, %rd463, %rd2413; ld.global.u16 %rs150, [%rd2414+12]; ld.global.u32 %r3257, [%rd2414+8]; ld.global.u64 %rd3029, [%rd2414]; setp.ne.s64 %p1542, %rd3029, 11; @%p1542 bra $L__BB108_1725; $L__BB108_1729: setp.eq.s16 %p1543, %rs150, 0; selp.b32 %r3261, -1, %r3257, %p1543; $L__BB108_1730: ld.const.u64 %rd2415, [kernel_params+2232]; cvta.to.global.u64 %rd2416, %rd2415; shl.b64 %rd2417, %rd462, 4; add.s64 %rd2418, %rd2416, %rd2417; ld.global.v4.u32 {%r2479, %r2480, %r2481, %r2482}, [%rd2418]; setp.eq.s32 %p1544, %r815, %r814; @%p1544 bra $L__BB108_1732; bra.uni $L__BB108_1731; $L__BB108_1732: ld.const.u64 %rd2424, [kernel_params+2256]; cvta.to.global.u64 %rd2425, %rd2424; mul.wide.u32 %rd2426, %r2482, 12; add.s64 %rd3030, %rd2425, %rd2426; add.s32 %r2489, %r2482, 1; mul.wide.u32 %rd2427, %r2489, 12; add.s64 %rd3033, %rd2425, %rd2427; add.s32 %r2490, %r2482, 2; mul.wide.u32 %rd2428, %r2490, 12; add.s64 %rd3036, %rd2425, %rd2428; add.s64 %rd3031, %rd3030, 4; add.s64 %rd3032, %rd3030, 8; add.s64 %rd3034, %rd3033, 4; add.s64 %rd3035, %rd3033, 8; add.s64 %rd3037, %rd3036, 4; add.s64 %rd3038, %rd3036, 8; bra.uni $L__BB108_1733; $L__BB108_1722: cvt.u32.u64 %r2464, %rd462; ld.const.v2.u64 {%rd2388, %rd2389}, [kernel_params+2224]; cvta.to.global.u64 %rd2392, %rd2389; mul.wide.s32 %rd2393, %r2464, 16; add.s64 %rd2394, %rd2392, %rd2393; ld.global.v4.u32 {%r2465, %r2466, %r2467, %r2468}, [%rd2394]; ld.const.u64 %rd2395, [kernel_params+2256]; cvta.to.global.u64 %rd2396, %rd2395; mul.wide.u32 %rd2397, %r2468, 12; add.s64 %rd2398, %rd2396, %rd2397; ld.global.f32 %f16400, [%rd2398]; ld.global.f32 %f16401, [%rd2398+4]; ld.global.f32 %f16402, [%rd2398+8]; add.s32 %r2473, %r2468, 1; mul.wide.u32 %rd2399, %r2473, 12; add.s64 %rd2400, %rd2396, %rd2399; ld.global.f32 %f16403, [%rd2400]; ld.global.f32 %f16404, [%rd2400+4]; ld.global.f32 %f16405, [%rd2400+8]; add.s32 %r2474, %r2468, 2; mul.wide.u32 %rd2401, %r2474, 12; add.s64 %rd2402, %rd2396, %rd2401; ld.global.f32 %f16406, [%rd2402]; ld.global.f32 %f16407, [%rd2402+4]; ld.global.f32 %f16408, [%rd2402+8]; cvta.to.global.u64 %rd2403, %rd2388; mul.wide.u32 %rd2404, %r2465, 12; add.s64 %rd2405, %rd2403, %rd2404; ld.global.f32 %f16391, [%rd2405]; ld.global.f32 %f16392, [%rd2405+4]; ld.global.f32 %f16393, [%rd2405+8]; mul.wide.u32 %rd2406, %r2466, 12; add.s64 %rd2407, %rd2403, %rd2406; ld.global.f32 %f16394, [%rd2407]; ld.global.f32 %f16395, [%rd2407+4]; ld.global.f32 %f16396, [%rd2407+8]; mul.wide.u32 %rd2408, %r2467, 12; add.s64 %rd2409, %rd2403, %rd2408; ld.global.f32 %f16397, [%rd2409]; ld.global.f32 %f16398, [%rd2409+4]; ld.global.f32 %f16399, [%rd2409+8]; bra.uni $L__BB108_1750; $L__BB108_1731: setp.gt.s32 %p1545, %r815, %r814; selp.b32 %r2483, -1, 0, %p1545; add.s32 %r2484, %r815, %r2483; mad.lo.s32 %r2485, %r2484, %r813, %r3261; ld.const.u64 %rd2419, [kernel_params+2336]; cvta.to.global.u64 %rd2420, %rd2419; add.s32 %r2486, %r2479, %r2485; mul.wide.u32 %rd2421, %r2486, 12; add.s64 %rd3030, %rd2420, %rd2421; add.s32 %r2487, %r2480, %r2485; mul.wide.u32 %rd2422, %r2487, 12; add.s64 %rd3033, %rd2420, %rd2422; add.s32 %r2488, %r2481, %r2485; mul.wide.u32 %rd2423, %r2488, 12; add.s64 %rd3036, %rd2420, %rd2423; add.s64 %rd3031, %rd3030, 4; add.s64 %rd3032, %rd3030, 8; add.s64 %rd3034, %rd3033, 4; add.s64 %rd3035, %rd3033, 8; add.s64 %rd3037, %rd3036, 4; add.s64 %rd3038, %rd3036, 8; $L__BB108_1733: ld.global.f32 %f2580, [%rd3038]; ld.global.f32 %f2581, [%rd3037]; ld.global.f32 %f2582, [%rd3036]; ld.global.f32 %f2583, [%rd3035]; ld.global.f32 %f2584, [%rd3034]; ld.global.f32 %f2585, [%rd3033]; ld.global.f32 %f2586, [%rd3032]; ld.global.f32 %f2587, [%rd3031]; ld.global.f32 %f2588, [%rd3030]; add.s32 %r830, %r815, 1; setp.eq.s32 %p1546, %r830, %r814; @%p1546 bra $L__BB108_1735; bra.uni $L__BB108_1734; $L__BB108_1735: ld.const.u64 %rd2434, [kernel_params+2256]; cvta.to.global.u64 %rd2435, %rd2434; mul.wide.u32 %rd2436, %r2482, 12; add.s64 %rd3039, %rd2435, %rd2436; add.s32 %r2497, %r2482, 1; mul.wide.u32 %rd2437, %r2497, 12; add.s64 %rd3042, %rd2435, %rd2437; add.s32 %r2498, %r2482, 2; mul.wide.u32 %rd2438, %r2498, 12; add.s64 %rd3045, %rd2435, %rd2438; add.s64 %rd3040, %rd3039, 4; add.s64 %rd3041, %rd3039, 8; add.s64 %rd3043, %rd3042, 4; add.s64 %rd3044, %rd3042, 8; add.s64 %rd3046, %rd3045, 4; add.s64 %rd3047, %rd3045, 8; bra.uni $L__BB108_1736; $L__BB108_1734: setp.ge.s32 %p1547, %r815, %r814; selp.b32 %r2491, -1, 0, %p1547; add.s32 %r2492, %r830, %r2491; mad.lo.s32 %r2493, %r2492, %r813, %r3261; ld.const.u64 %rd2429, [kernel_params+2336]; cvta.to.global.u64 %rd2430, %rd2429; add.s32 %r2494, %r2479, %r2493; mul.wide.u32 %rd2431, %r2494, 12; add.s64 %rd3039, %rd2430, %rd2431; add.s32 %r2495, %r2480, %r2493; mul.wide.u32 %rd2432, %r2495, 12; add.s64 %rd3042, %rd2430, %rd2432; add.s32 %r2496, %r2481, %r2493; mul.wide.u32 %rd2433, %r2496, 12; add.s64 %rd3045, %rd2430, %rd2433; add.s64 %rd3040, %rd3039, 4; add.s64 %rd3041, %rd3039, 8; add.s64 %rd3043, %rd3042, 4; add.s64 %rd3044, %rd3042, 8; add.s64 %rd3046, %rd3045, 4; add.s64 %rd3047, %rd3045, 8; $L__BB108_1736: sub.ftz.f32 %f2589, %f10876, %f2579; ld.global.f32 %f10887, [%rd3039]; mul.ftz.f32 %f10888, %f2579, %f10887; ld.global.f32 %f10889, [%rd3040]; mul.ftz.f32 %f10890, %f2579, %f10889; ld.global.f32 %f10891, [%rd3041]; mul.ftz.f32 %f10892, %f2579, %f10891; fma.rn.ftz.f32 %f16400, %f2589, %f2588, %f10888; fma.rn.ftz.f32 %f16401, %f2589, %f2587, %f10890; fma.rn.ftz.f32 %f16402, %f2589, %f2586, %f10892; ld.global.f32 %f10893, [%rd3042]; mul.ftz.f32 %f10894, %f2579, %f10893; ld.global.f32 %f10895, [%rd3043]; mul.ftz.f32 %f10896, %f2579, %f10895; ld.global.f32 %f10897, [%rd3044]; mul.ftz.f32 %f10898, %f2579, %f10897; fma.rn.ftz.f32 %f16403, %f2589, %f2585, %f10894; fma.rn.ftz.f32 %f16404, %f2589, %f2584, %f10896; fma.rn.ftz.f32 %f16405, %f2589, %f2583, %f10898; ld.global.f32 %f10899, [%rd3045]; mul.ftz.f32 %f10900, %f2579, %f10899; ld.global.f32 %f10901, [%rd3046]; mul.ftz.f32 %f10902, %f2579, %f10901; ld.global.f32 %f10903, [%rd3047]; mul.ftz.f32 %f10904, %f2579, %f10903; fma.rn.ftz.f32 %f16406, %f2589, %f2582, %f10900; fma.rn.ftz.f32 %f16407, %f2589, %f2581, %f10902; fma.rn.ftz.f32 %f16408, %f2589, %f2580, %f10904; setp.eq.s64 %p1548, %rd3048, 12; @%p1548 bra $L__BB108_1742; mov.u32 %r3266, -1; $L__BB108_1738: setp.eq.s64 %p1549, %rd3048, 0; @%p1549 bra $L__BB108_1740; add.s32 %r3263, %r3263, 2; bra.uni $L__BB108_1741; $L__BB108_1740: setp.eq.s16 %p1550, %rs152, 0; mov.u32 %r3263, %r3262; @%p1550 bra $L__BB108_1743; $L__BB108_1741: mul.wide.u32 %rd2439, %r3263, 16; add.s64 %rd2440, %rd463, %rd2439; ld.global.u16 %rs152, [%rd2440+12]; ld.global.u32 %r3262, [%rd2440+8]; ld.global.u64 %rd3048, [%rd2440]; setp.ne.s64 %p1551, %rd3048, 12; @%p1551 bra $L__BB108_1738; $L__BB108_1742: setp.eq.s16 %p1552, %rs152, 0; selp.b32 %r3266, -1, %r3262, %p1552; $L__BB108_1743: @%p1544 bra $L__BB108_1745; bra.uni $L__BB108_1744; $L__BB108_1745: ld.const.u64 %rd2446, [kernel_params+2224]; cvta.to.global.u64 %rd2447, %rd2446; mul.wide.u32 %rd2448, %r2479, 12; add.s64 %rd3049, %rd2447, %rd2448; mul.wide.u32 %rd2449, %r2480, 12; add.s64 %rd3052, %rd2447, %rd2449; mul.wide.u32 %rd2450, %r2481, 12; add.s64 %rd3055, %rd2447, %rd2450; add.s64 %rd3050, %rd3049, 4; add.s64 %rd3051, %rd3049, 8; add.s64 %rd3053, %rd3052, 4; add.s64 %rd3054, %rd3052, 8; add.s64 %rd3056, %rd3055, 4; add.s64 %rd3057, %rd3055, 8; bra.uni $L__BB108_1746; $L__BB108_1744: setp.gt.s32 %p1554, %r815, %r814; selp.b32 %r2500, -1, 0, %p1554; add.s32 %r2501, %r815, %r2500; mad.lo.s32 %r2502, %r2501, %r813, %r3266; ld.const.u64 %rd2441, [kernel_params+2336]; cvta.to.global.u64 %rd2442, %rd2441; add.s32 %r2503, %r2502, %r2479; mul.wide.u32 %rd2443, %r2503, 12; add.s64 %rd3049, %rd2442, %rd2443; add.s32 %r2504, %r2502, %r2480; mul.wide.u32 %rd2444, %r2504, 12; add.s64 %rd3052, %rd2442, %rd2444; add.s32 %r2505, %r2502, %r2481; mul.wide.u32 %rd2445, %r2505, 12; add.s64 %rd3055, %rd2442, %rd2445; add.s64 %rd3050, %rd3049, 4; add.s64 %rd3051, %rd3049, 8; add.s64 %rd3053, %rd3052, 4; add.s64 %rd3054, %rd3052, 8; add.s64 %rd3056, %rd3055, 4; add.s64 %rd3057, %rd3055, 8; $L__BB108_1746: ld.global.f32 %f2599, [%rd3057]; ld.global.f32 %f2600, [%rd3056]; ld.global.f32 %f2601, [%rd3055]; ld.global.f32 %f2602, [%rd3054]; ld.global.f32 %f2603, [%rd3053]; ld.global.f32 %f2604, [%rd3052]; ld.global.f32 %f2605, [%rd3051]; ld.global.f32 %f2606, [%rd3050]; ld.global.f32 %f2607, [%rd3049]; @%p1546 bra $L__BB108_1748; bra.uni $L__BB108_1747; $L__BB108_1748: ld.const.u64 %rd2456, [kernel_params+2224]; cvta.to.global.u64 %rd2457, %rd2456; mul.wide.u32 %rd2458, %r2479, 12; add.s64 %rd3058, %rd2457, %rd2458; mul.wide.u32 %rd2459, %r2480, 12; add.s64 %rd3061, %rd2457, %rd2459; mul.wide.u32 %rd2460, %r2481, 12; add.s64 %rd3064, %rd2457, %rd2460; add.s64 %rd3059, %rd3058, 4; add.s64 %rd3060, %rd3058, 8; add.s64 %rd3062, %rd3061, 4; add.s64 %rd3063, %rd3061, 8; add.s64 %rd3065, %rd3064, 4; add.s64 %rd3066, %rd3064, 8; bra.uni $L__BB108_1749; $L__BB108_1747: setp.ge.s32 %p1556, %r815, %r814; selp.b32 %r2506, -1, 0, %p1556; add.s32 %r2507, %r830, %r2506; mad.lo.s32 %r2508, %r2507, %r813, %r3266; ld.const.u64 %rd2451, [kernel_params+2336]; cvta.to.global.u64 %rd2452, %rd2451; add.s32 %r2509, %r2508, %r2479; mul.wide.u32 %rd2453, %r2509, 12; add.s64 %rd3058, %rd2452, %rd2453; add.s32 %r2510, %r2508, %r2480; mul.wide.u32 %rd2454, %r2510, 12; add.s64 %rd3061, %rd2452, %rd2454; add.s32 %r2511, %r2508, %r2481; mul.wide.u32 %rd2455, %r2511, 12; add.s64 %rd3064, %rd2452, %rd2455; add.s64 %rd3059, %rd3058, 4; add.s64 %rd3060, %rd3058, 8; add.s64 %rd3062, %rd3061, 4; add.s64 %rd3063, %rd3061, 8; add.s64 %rd3065, %rd3064, 4; add.s64 %rd3066, %rd3064, 8; $L__BB108_1749: ld.global.f32 %f10905, [%rd3058]; mul.ftz.f32 %f10906, %f2579, %f10905; ld.global.f32 %f10907, [%rd3059]; mul.ftz.f32 %f10908, %f2579, %f10907; ld.global.f32 %f10909, [%rd3060]; mul.ftz.f32 %f10910, %f2579, %f10909; fma.rn.ftz.f32 %f16391, %f2589, %f2607, %f10906; fma.rn.ftz.f32 %f16392, %f2589, %f2606, %f10908; fma.rn.ftz.f32 %f16393, %f2589, %f2605, %f10910; ld.global.f32 %f10911, [%rd3061]; mul.ftz.f32 %f10912, %f2579, %f10911; ld.global.f32 %f10913, [%rd3062]; mul.ftz.f32 %f10914, %f2579, %f10913; ld.global.f32 %f10915, [%rd3063]; mul.ftz.f32 %f10916, %f2579, %f10915; fma.rn.ftz.f32 %f16394, %f2589, %f2604, %f10912; fma.rn.ftz.f32 %f16395, %f2589, %f2603, %f10914; fma.rn.ftz.f32 %f16396, %f2589, %f2602, %f10916; ld.global.f32 %f10917, [%rd3064]; mul.ftz.f32 %f10918, %f2579, %f10917; ld.global.f32 %f10919, [%rd3065]; mul.ftz.f32 %f10920, %f2579, %f10919; ld.global.f32 %f10921, [%rd3066]; mul.ftz.f32 %f10922, %f2579, %f10921; fma.rn.ftz.f32 %f16397, %f2589, %f2601, %f10918; fma.rn.ftz.f32 %f16398, %f2589, %f2600, %f10920; fma.rn.ftz.f32 %f16399, %f2589, %f2599, %f10922; $L__BB108_1750: ld.local.f32 %f2635, [%rd7+-20]; sub.ftz.f32 %f10924, %f10876, %f2635; ld.local.f32 %f10925, [%rd7+-16]; sub.ftz.f32 %f2636, %f10924, %f10925; mul.ftz.f32 %f10926, %f16400, %f2636; mul.ftz.f32 %f10927, %f16401, %f2636; mul.ftz.f32 %f10928, %f16402, %f2636; fma.rn.ftz.f32 %f10929, %f16403, %f2635, %f10926; fma.rn.ftz.f32 %f10930, %f16404, %f2635, %f10927; fma.rn.ftz.f32 %f10931, %f16405, %f2635, %f10928; fma.rn.ftz.f32 %f2637, %f16406, %f10925, %f10929; fma.rn.ftz.f32 %f2638, %f16407, %f10925, %f10930; fma.rn.ftz.f32 %f2639, %f16408, %f10925, %f10931; mul.ftz.f32 %f10932, %f16391, %f2636; mul.ftz.f32 %f10933, %f16392, %f2636; mul.ftz.f32 %f10934, %f16393, %f2636; fma.rn.ftz.f32 %f10935, %f16394, %f2635, %f10932; fma.rn.ftz.f32 %f10936, %f16395, %f2635, %f10933; fma.rn.ftz.f32 %f10937, %f16396, %f2635, %f10934; fma.rn.ftz.f32 %f16409, %f16397, %f10925, %f10935; fma.rn.ftz.f32 %f16410, %f16398, %f10925, %f10936; fma.rn.ftz.f32 %f16411, %f16399, %f10925, %f10937; ld.local.u32 %r839, [%rd7+-32]; and.b32 %r2512, %r839, 4; setp.ne.s32 %p1557, %r2512, 0; @%p1557 bra $L__BB108_1754; and.b32 %r2513, %r839, 2; setp.eq.s32 %p1558, %r2513, 0; @%p1558 bra $L__BB108_1753; bra.uni $L__BB108_1752; $L__BB108_1753: ld.global.v4.f32 {%f10965, %f10966, %f10967, %f10968}, [%rd461+-184]; ld.global.v4.f32 {%f10972, %f10973, %f10974, %f10975}, [%rd461+-168]; ld.global.v4.f32 {%f10979, %f10980, %f10981, %f10982}, [%rd461+-152]; mul.ftz.f32 %f10986, %f16410, %f10966; fma.rn.ftz.f32 %f10987, %f16409, %f10965, %f10986; fma.rn.ftz.f32 %f2646, %f16411, %f10967, %f10987; mul.ftz.f32 %f10988, %f16410, %f10973; fma.rn.ftz.f32 %f10989, %f16409, %f10972, %f10988; fma.rn.ftz.f32 %f2647, %f16411, %f10974, %f10989; mul.ftz.f32 %f10990, %f16410, %f10980; fma.rn.ftz.f32 %f10991, %f16409, %f10979, %f10990; fma.rn.ftz.f32 %f16411, %f16411, %f10981, %f10991; mov.f32 %f16409, %f2646; mov.f32 %f16410, %f2647; bra.uni $L__BB108_1754; $L__BB108_1752: ld.local.v4.f32 {%f10938, %f10939, %f10940, %f10941}, [%rd7+56]; mul.ftz.f32 %f10945, %f16410, %f10939; fma.rn.ftz.f32 %f10946, %f16409, %f10938, %f10945; fma.rn.ftz.f32 %f2643, %f16411, %f10940, %f10946; ld.local.v4.f32 {%f10947, %f10948, %f10949, %f10950}, [%rd7+72]; mul.ftz.f32 %f10954, %f16410, %f10948; fma.rn.ftz.f32 %f10955, %f16409, %f10947, %f10954; fma.rn.ftz.f32 %f2644, %f16411, %f10949, %f10955; ld.local.v4.f32 {%f10956, %f10957, %f10958, %f10959}, [%rd7+88]; mul.ftz.f32 %f10963, %f16410, %f10957; fma.rn.ftz.f32 %f10964, %f16409, %f10956, %f10963; fma.rn.ftz.f32 %f16411, %f16411, %f10958, %f10964; mov.f32 %f16409, %f2643; mov.f32 %f16410, %f2644; $L__BB108_1754: sub.ftz.f32 %f2652, %f16400, %f16406; sub.ftz.f32 %f10992, %f16397, %f16391; sub.ftz.f32 %f2653, %f16401, %f16407; sub.ftz.f32 %f10993, %f16398, %f16392; mul.ftz.f32 %f10994, %f10993, %f2653; fma.rn.ftz.f32 %f10995, %f10992, %f2652, %f10994; sub.ftz.f32 %f2654, %f16402, %f16408; sub.ftz.f32 %f10996, %f16399, %f16393; fma.rn.ftz.f32 %f10997, %f10996, %f2654, %f10995; sub.ftz.f32 %f2655, %f16403, %f16406; sub.ftz.f32 %f10998, %f16397, %f16394; sub.ftz.f32 %f2656, %f16404, %f16407; sub.ftz.f32 %f10999, %f16398, %f16395; mul.ftz.f32 %f11000, %f10999, %f2656; fma.rn.ftz.f32 %f11001, %f10998, %f2655, %f11000; sub.ftz.f32 %f2657, %f16405, %f16408; sub.ftz.f32 %f11002, %f16399, %f16396; fma.rn.ftz.f32 %f11003, %f11002, %f2657, %f11001; sub.ftz.f32 %f2658, %f16403, %f16400; sub.ftz.f32 %f11004, %f16394, %f16391; sub.ftz.f32 %f2659, %f16404, %f16401; sub.ftz.f32 %f11005, %f16395, %f16392; mul.ftz.f32 %f11006, %f11005, %f2659; fma.rn.ftz.f32 %f11007, %f11004, %f2658, %f11006; sub.ftz.f32 %f2660, %f16405, %f16402; sub.ftz.f32 %f11008, %f16396, %f16393; fma.rn.ftz.f32 %f11009, %f11008, %f2660, %f11007; mul.ftz.f32 %f11010, %f10997, %f2636; add.ftz.f32 %f11011, %f2636, 0fBF800000; add.ftz.f32 %f11012, %f10997, %f11003; add.ftz.f32 %f11013, %f11009, %f11012; mul.ftz.f32 %f11014, %f2636, %f11013; mul.ftz.f32 %f11015, %f2635, %f11014; fma.rn.ftz.f32 %f11016, %f11010, %f11011, %f11015; mul.ftz.f32 %f11017, %f2635, %f11003; add.ftz.f32 %f11018, %f2635, 0fBF800000; fma.rn.ftz.f32 %f11019, %f11018, %f11017, %f11016; mul.ftz.f32 %f11020, %f16389, %f16410; fma.rn.ftz.f32 %f11021, %f16388, %f16409, %f11020; fma.rn.ftz.f32 %f11022, %f16390, %f16411, %f11021; setp.gt.ftz.f32 %p1559, %f11022, 0f00000000; sub.ftz.f32 %f2661, %f16400, %f16403; sub.ftz.f32 %f2662, %f16401, %f16404; sub.ftz.f32 %f2663, %f16402, %f16405; sub.ftz.f32 %f2664, %f16406, %f16400; sub.ftz.f32 %f2665, %f16407, %f16401; sub.ftz.f32 %f2666, %f16408, %f16402; sub.ftz.f32 %f2667, %f16406, %f16403; sub.ftz.f32 %f2668, %f16407, %f16404; sub.ftz.f32 %f2669, %f16408, %f16405; mul.ftz.f32 %f2670, %f11019, 0f3F000000; @%p1559 bra $L__BB108_1756; bra.uni $L__BB108_1755; $L__BB108_1756: mul.ftz.f32 %f11075, %f16392, %f2659; fma.rn.ftz.f32 %f11076, %f16391, %f2658, %f11075; fma.rn.ftz.f32 %f11077, %f16393, %f2660, %f11076; mul.ftz.f32 %f11078, %f16392, %f2665; fma.rn.ftz.f32 %f11079, %f16391, %f2664, %f11078; fma.rn.ftz.f32 %f11080, %f16393, %f2666, %f11079; max.ftz.f32 %f11081, %f11077, %f11080; max.ftz.f32 %f11083, %f11081, %f10880; mul.ftz.f32 %f11084, %f16395, %f2662; fma.rn.ftz.f32 %f11085, %f16394, %f2661, %f11084; fma.rn.ftz.f32 %f11086, %f16396, %f2663, %f11085; mul.ftz.f32 %f11087, %f16395, %f2668; fma.rn.ftz.f32 %f11088, %f16394, %f2667, %f11087; fma.rn.ftz.f32 %f11089, %f16396, %f2669, %f11088; max.ftz.f32 %f11090, %f11086, %f11089; max.ftz.f32 %f11091, %f11090, %f10880; mul.ftz.f32 %f11092, %f16398, %f2653; fma.rn.ftz.f32 %f11093, %f16397, %f2652, %f11092; fma.rn.ftz.f32 %f11094, %f16399, %f2654, %f11093; mul.ftz.f32 %f11095, %f16398, %f2656; fma.rn.ftz.f32 %f11096, %f16397, %f2655, %f11095; fma.rn.ftz.f32 %f11097, %f16399, %f2657, %f11096; max.ftz.f32 %f11098, %f11094, %f11097; max.ftz.f32 %f11099, %f11098, %f10880; sub.ftz.f32 %f11100, %f16400, %f2637; sub.ftz.f32 %f11101, %f16401, %f2638; mul.ftz.f32 %f11102, %f16392, %f11101; fma.rn.ftz.f32 %f11103, %f16391, %f11100, %f11102; sub.ftz.f32 %f11104, %f16402, %f2639; fma.rn.ftz.f32 %f11105, %f16393, %f11104, %f11103; add.ftz.f32 %f11106, %f11105, %f11083; max.ftz.f32 %f11107, %f11106, %f10880; sub.ftz.f32 %f11108, %f16403, %f2637; sub.ftz.f32 %f11109, %f16404, %f2638; mul.ftz.f32 %f11110, %f16395, %f11109; fma.rn.ftz.f32 %f11111, %f16394, %f11108, %f11110; sub.ftz.f32 %f11112, %f16405, %f2639; fma.rn.ftz.f32 %f11113, %f16396, %f11112, %f11111; add.ftz.f32 %f11114, %f11113, %f11091; max.ftz.f32 %f11115, %f11114, %f10880; sub.ftz.f32 %f11116, %f16406, %f2637; sub.ftz.f32 %f11117, %f16407, %f2638; mul.ftz.f32 %f11118, %f16398, %f11117; fma.rn.ftz.f32 %f11119, %f16397, %f11116, %f11118; sub.ftz.f32 %f11120, %f16408, %f2639; fma.rn.ftz.f32 %f11121, %f16399, %f11120, %f11119; add.ftz.f32 %f11122, %f11121, %f11099; max.ftz.f32 %f11123, %f11122, %f10880; min.ftz.f32 %f11124, %f11107, %f11115; min.ftz.f32 %f11125, %f11124, %f11123; max.ftz.f32 %f16412, %f11125, %f2670; bra.uni $L__BB108_1757; $L__BB108_1755: mul.ftz.f32 %f11023, %f16392, %f2662; fma.rn.ftz.f32 %f11024, %f16391, %f2661, %f11023; fma.rn.ftz.f32 %f11025, %f16393, %f2663, %f11024; mul.ftz.f32 %f11026, %f16392, %f2653; fma.rn.ftz.f32 %f11027, %f16391, %f2652, %f11026; fma.rn.ftz.f32 %f11028, %f16393, %f2654, %f11027; max.ftz.f32 %f11029, %f11025, %f11028; max.ftz.f32 %f11031, %f11029, %f10880; mul.ftz.f32 %f11032, %f16395, %f2659; fma.rn.ftz.f32 %f11033, %f16394, %f2658, %f11032; fma.rn.ftz.f32 %f11034, %f16396, %f2660, %f11033; mul.ftz.f32 %f11035, %f16395, %f2656; fma.rn.ftz.f32 %f11036, %f16394, %f2655, %f11035; fma.rn.ftz.f32 %f11037, %f16396, %f2657, %f11036; max.ftz.f32 %f11038, %f11034, %f11037; max.ftz.f32 %f11039, %f11038, %f10880; mul.ftz.f32 %f11040, %f16398, %f2665; fma.rn.ftz.f32 %f11041, %f16397, %f2664, %f11040; fma.rn.ftz.f32 %f11042, %f16399, %f2666, %f11041; mul.ftz.f32 %f11043, %f16398, %f2668; fma.rn.ftz.f32 %f11044, %f16397, %f2667, %f11043; fma.rn.ftz.f32 %f11045, %f16399, %f2669, %f11044; max.ftz.f32 %f11046, %f11042, %f11045; max.ftz.f32 %f11047, %f11046, %f10880; sub.ftz.f32 %f11048, %f2637, %f16400; sub.ftz.f32 %f11049, %f2638, %f16401; mul.ftz.f32 %f11050, %f16392, %f11049; fma.rn.ftz.f32 %f11051, %f16391, %f11048, %f11050; sub.ftz.f32 %f11052, %f2639, %f16402; fma.rn.ftz.f32 %f11053, %f16393, %f11052, %f11051; add.ftz.f32 %f11054, %f11053, %f11031; max.ftz.f32 %f11055, %f11054, %f10880; sub.ftz.f32 %f11056, %f2637, %f16403; sub.ftz.f32 %f11057, %f2638, %f16404; mul.ftz.f32 %f11058, %f16395, %f11057; fma.rn.ftz.f32 %f11059, %f16394, %f11056, %f11058; sub.ftz.f32 %f11060, %f2639, %f16405; fma.rn.ftz.f32 %f11061, %f16396, %f11060, %f11059; add.ftz.f32 %f11062, %f11061, %f11039; max.ftz.f32 %f11063, %f11062, %f10880; sub.ftz.f32 %f11064, %f2637, %f16406; sub.ftz.f32 %f11065, %f2638, %f16407; mul.ftz.f32 %f11066, %f16398, %f11065; fma.rn.ftz.f32 %f11067, %f16397, %f11064, %f11066; sub.ftz.f32 %f11068, %f2639, %f16408; fma.rn.ftz.f32 %f11069, %f16399, %f11068, %f11067; add.ftz.f32 %f11070, %f11069, %f11047; max.ftz.f32 %f11071, %f11070, %f10880; min.ftz.f32 %f11072, %f11055, %f11063; min.ftz.f32 %f11073, %f11072, %f11071; neg.ftz.f32 %f11074, %f11073; min.ftz.f32 %f16412, %f11074, %f2670; $L__BB108_1757: mul.ftz.f32 %f11126, %f16409, %f16412; mul.ftz.f32 %f11127, %f16410, %f16412; mul.ftz.f32 %f11128, %f16411, %f16412; fma.rn.ftz.f32 %f16413, %f2560, %f11126, %f16413; fma.rn.ftz.f32 %f16414, %f2560, %f11127, %f16414; fma.rn.ftz.f32 %f16415, %f2560, %f11128, %f16415; setp.gt.ftz.f32 %p1560, %f2559, 0f00000000; selp.u16 %rs154, 1, 0, %p1560; $L__BB108_1758: setp.eq.s16 %p1561, %rs154, 0; @%p1561 bra $L__BB108_1767; @%p1531 bra $L__BB108_1767; ld.const.u64 %rd2461, [kernel_params+2232]; cvta.to.global.u64 %rd2462, %rd2461; ld.local.v2.u32 {%r2514, %r2515}, [%rd7+-32]; mul.wide.s32 %rd2463, %r2515, 16; add.s64 %rd2464, %rd2462, %rd2463; ld.const.u64 %rd2465, [kernel_params+2256]; cvta.to.global.u64 %rd2466, %rd2465; ld.global.u32 %r2517, [%rd2464+12]; mul.wide.u32 %rd2467, %r2517, 12; add.s64 %rd2468, %rd2466, %rd2467; ld.global.f32 %f2683, [%rd2468]; ld.global.f32 %f2684, [%rd2468+4]; ld.global.f32 %f2685, [%rd2468+8]; add.s32 %r2518, %r2517, 1; mul.wide.u32 %rd2469, %r2518, 12; add.s64 %rd2470, %rd2466, %rd2469; ld.global.f32 %f2686, [%rd2470]; ld.global.f32 %f2687, [%rd2470+4]; ld.global.f32 %f2688, [%rd2470+8]; add.s32 %r2519, %r2517, 2; mul.wide.u32 %rd2471, %r2519, 12; add.s64 %rd2472, %rd2466, %rd2471; ld.global.f32 %f2689, [%rd2472]; ld.global.f32 %f2690, [%rd2472+4]; ld.global.f32 %f2691, [%rd2472+8]; and.b32 %r2520, %r2514, 4; setp.ne.s32 %p1563, %r2520, 0; mov.f32 %f16428, %f16413; mov.f32 %f16429, %f16414; mov.f32 %f16430, %f16415; mov.f32 %f16431, %f2542; mov.f32 %f16432, %f2543; mov.f32 %f16433, %f2544; @%p1563 bra $L__BB108_1765; and.b32 %r2521, %r2514, 2; setp.eq.s32 %p1564, %r2521, 0; @%p1564 bra $L__BB108_1763; bra.uni $L__BB108_1762; $L__BB108_1763: ld.const.u64 %rd2473, [kernel_params+2160]; cvta.to.global.u64 %rd2474, %rd2473; ld.local.u32 %r2522, [%rd7+-12]; mul.wide.s32 %rd2475, %r2522, 208; add.s64 %rd2476, %rd2474, %rd2475; ld.global.v4.f32 {%f16416, %f16417, %f16418, %f16419}, [%rd2476+48]; ld.global.v4.f32 {%f16420, %f16421, %f16422, %f16423}, [%rd2476+64]; ld.global.v4.f32 {%f16424, %f16425, %f16426, %f16427}, [%rd2476+80]; bra.uni $L__BB108_1764; $L__BB108_1762: ld.local.v4.f32 {%f16416, %f16417, %f16418, %f16419}, [%rd7+104]; ld.local.v4.f32 {%f16420, %f16421, %f16422, %f16423}, [%rd7+120]; ld.local.v4.f32 {%f16424, %f16425, %f16426, %f16427}, [%rd7+136]; $L__BB108_1764: mul.ftz.f32 %f11153, %f16414, %f16417; fma.rn.ftz.f32 %f11154, %f16413, %f16416, %f11153; fma.rn.ftz.f32 %f11155, %f16415, %f16418, %f11154; mul.ftz.f32 %f11156, %f16414, %f16421; fma.rn.ftz.f32 %f11157, %f16413, %f16420, %f11156; fma.rn.ftz.f32 %f11158, %f16415, %f16422, %f11157; mul.ftz.f32 %f11159, %f16414, %f16425; fma.rn.ftz.f32 %f11160, %f16413, %f16424, %f11159; fma.rn.ftz.f32 %f11161, %f16415, %f16426, %f11160; add.ftz.f32 %f16428, %f16419, %f11155; add.ftz.f32 %f16429, %f16423, %f11158; add.ftz.f32 %f16430, %f16427, %f11161; mul.ftz.f32 %f11162, %f2543, %f16417; fma.rn.ftz.f32 %f11163, %f2542, %f16416, %f11162; mul.ftz.f32 %f11164, %f2543, %f16421; fma.rn.ftz.f32 %f11165, %f2542, %f16420, %f11164; mul.ftz.f32 %f11166, %f2543, %f16425; fma.rn.ftz.f32 %f11167, %f2542, %f16424, %f11166; fma.rn.ftz.f32 %f16431, %f2544, %f16418, %f11163; fma.rn.ftz.f32 %f16432, %f2544, %f16422, %f11165; fma.rn.ftz.f32 %f16433, %f2544, %f16426, %f11167; $L__BB108_1765: sub.ftz.f32 %f11168, %f2683, %f16428; sub.ftz.f32 %f11169, %f2684, %f16429; sub.ftz.f32 %f11170, %f2685, %f16430; sub.ftz.f32 %f11171, %f2686, %f16428; sub.ftz.f32 %f11172, %f2687, %f16429; sub.ftz.f32 %f11173, %f2688, %f16430; sub.ftz.f32 %f11174, %f2689, %f16428; sub.ftz.f32 %f11175, %f2690, %f16429; sub.ftz.f32 %f11176, %f2691, %f16430; sub.ftz.f32 %f11177, %f11174, %f11168; sub.ftz.f32 %f11178, %f11175, %f11169; sub.ftz.f32 %f11179, %f11176, %f11170; sub.ftz.f32 %f11180, %f11168, %f11171; sub.ftz.f32 %f11181, %f11169, %f11172; sub.ftz.f32 %f11182, %f11170, %f11173; sub.ftz.f32 %f11183, %f11171, %f11174; sub.ftz.f32 %f11184, %f11172, %f11175; sub.ftz.f32 %f11185, %f11173, %f11176; add.ftz.f32 %f11186, %f11174, %f11168; add.ftz.f32 %f11187, %f11175, %f11169; add.ftz.f32 %f11188, %f11176, %f11170; mul.ftz.f32 %f11189, %f11187, %f11179; mul.ftz.f32 %f11190, %f11188, %f11178; sub.ftz.f32 %f11191, %f11189, %f11190; mul.ftz.f32 %f11192, %f11188, %f11177; mul.ftz.f32 %f11193, %f11186, %f11179; sub.ftz.f32 %f11194, %f11192, %f11193; mul.ftz.f32 %f11195, %f11186, %f11178; mul.ftz.f32 %f11196, %f11187, %f11177; sub.ftz.f32 %f11197, %f11195, %f11196; mul.ftz.f32 %f11198, %f16432, %f11194; fma.rn.ftz.f32 %f11199, %f16431, %f11191, %f11198; fma.rn.ftz.f32 %f11200, %f16433, %f11197, %f11199; add.ftz.f32 %f11201, %f11168, %f11171; add.ftz.f32 %f11202, %f11169, %f11172; add.ftz.f32 %f11203, %f11170, %f11173; mul.ftz.f32 %f11204, %f11202, %f11182; mul.ftz.f32 %f11205, %f11203, %f11181; sub.ftz.f32 %f11206, %f11204, %f11205; mul.ftz.f32 %f11207, %f11203, %f11180; mul.ftz.f32 %f11208, %f11201, %f11182; sub.ftz.f32 %f11209, %f11207, %f11208; mul.ftz.f32 %f11210, %f11201, %f11181; mul.ftz.f32 %f11211, %f11202, %f11180; sub.ftz.f32 %f11212, %f11210, %f11211; mul.ftz.f32 %f11213, %f16432, %f11209; fma.rn.ftz.f32 %f11214, %f16431, %f11206, %f11213; fma.rn.ftz.f32 %f11215, %f16433, %f11212, %f11214; add.ftz.f32 %f11216, %f11171, %f11174; add.ftz.f32 %f11217, %f11172, %f11175; add.ftz.f32 %f11218, %f11173, %f11176; mul.ftz.f32 %f11219, %f11217, %f11185; mul.ftz.f32 %f11220, %f11218, %f11184; sub.ftz.f32 %f11221, %f11219, %f11220; mul.ftz.f32 %f11222, %f11218, %f11183; mul.ftz.f32 %f11223, %f11216, %f11185; sub.ftz.f32 %f11224, %f11222, %f11223; mul.ftz.f32 %f11225, %f11216, %f11184; mul.ftz.f32 %f11226, %f11217, %f11183; sub.ftz.f32 %f11227, %f11225, %f11226; mul.ftz.f32 %f11228, %f16432, %f11224; fma.rn.ftz.f32 %f11229, %f16431, %f11221, %f11228; fma.rn.ftz.f32 %f11230, %f16433, %f11227, %f11229; add.ftz.f32 %f11231, %f11200, %f11215; add.ftz.f32 %f11232, %f11230, %f11231; abs.ftz.f32 %f11233, %f11232; mul.ftz.f32 %f11234, %f11233, 0f34000000; min.ftz.f32 %f11235, %f11215, %f11230; min.ftz.f32 %f11236, %f11200, %f11235; max.ftz.f32 %f11237, %f11215, %f11230; max.ftz.f32 %f11238, %f11200, %f11237; setp.ge.ftz.f32 %p1565, %f11236, %f11234; mul.ftz.f32 %f11239, %f11233, 0fB4000000; setp.le.ftz.f32 %p1566, %f11238, %f11239; or.pred %p1567, %p1565, %p1566; @%p1567 bra $L__BB108_1767; ld.local.v2.f32 {%f11240, %f11241}, [%rd7+-64]; mul.ftz.f32 %f11244, %f11240, 0f43800000; cvt.rzi.ftz.s32.f32 %r2523, %f11244; mul.ftz.f32 %f11245, %f11241, 0f43800000; cvt.rzi.ftz.s32.f32 %r2524, %f11245; ld.local.f32 %f11246, [%rd7+-56]; mul.ftz.f32 %f11247, %f11246, 0f43800000; cvt.rzi.ftz.s32.f32 %r2525, %f11247; neg.s32 %r2526, %r2523; setp.lt.ftz.f32 %p1568, %f16413, 0f00000000; selp.b32 %r2527, %r2526, %r2523, %p1568; mov.b32 %r2528, %f16413; add.s32 %r2529, %r2527, %r2528; mov.b32 %f11248, %r2529; neg.s32 %r2530, %r2524; setp.lt.ftz.f32 %p1569, %f16414, 0f00000000; selp.b32 %r2531, %r2530, %r2524, %p1569; mov.b32 %r2532, %f16414; add.s32 %r2533, %r2531, %r2532; mov.b32 %f11249, %r2533; neg.s32 %r2534, %r2525; setp.lt.ftz.f32 %p1570, %f16415, 0f00000000; selp.b32 %r2535, %r2534, %r2525, %p1570; mov.b32 %r2536, %f16415; add.s32 %r2537, %r2535, %r2536; mov.b32 %f11250, %r2537; abs.ftz.f32 %f11251, %f16413; setp.lt.ftz.f32 %p1571, %f11251, 0f3D000000; fma.rn.ftz.f32 %f11252, %f11240, 0f37800000, %f16413; abs.ftz.f32 %f11253, %f16414; setp.lt.ftz.f32 %p1572, %f11253, 0f3D000000; fma.rn.ftz.f32 %f11254, %f11241, 0f37800000, %f16414; abs.ftz.f32 %f11255, %f16415; setp.lt.ftz.f32 %p1573, %f11255, 0f3D000000; fma.rn.ftz.f32 %f11256, %f11246, 0f37800000, %f16415; selp.f32 %f16415, %f11256, %f11250, %p1573; selp.f32 %f16414, %f11254, %f11249, %p1572; selp.f32 %f16413, %f11252, %f11248, %p1571; $L__BB108_1767: ld.const.f32 %f2755, [kernel_params+1540]; ld.local.f32 %f2756, [%rd7+-4]; ld.local.u32 %r2538, [%rd7+-12]; selp.b32 %r841, -1, %r2538, %p1561; ld.local.u32 %r2539, [%rd7+-28]; selp.b32 %r842, -1, %r2539, %p1561; ld.const.u64 %rd2477, [kernel_params+4376]; cvta.to.global.u64 %rd2478, %rd2477; atom.global.add.u32 %r2540, [%rd2478], 1; ld.const.u64 %rd2479, [kernel_params+4256]; cvta.to.global.u64 %rd2480, %rd2479; add.s64 %rd2481, %rd2480, 12; atom.global.add.u32 %r2541, [%rd2481], 1; ld.const.u64 %rd2482, [kernel_params+3416]; cvta.to.global.u64 %rd2483, %rd2482; cvt.s64.s32 %rd577, %r2540; mul.wide.s32 %rd2484, %r2540, 2; add.s64 %rd2485, %rd2483, %rd2484; mov.u16 %rs107, 3; st.global.u16 [%rd2485], %rs107; and.b32 %r2542, %r787, 131072; setp.eq.s32 %p1575, %r2542, 0; @%p1575 bra $L__BB108_1770; mov.u32 %r3267, 0; mov.u64 %rd2487, kernel_params; $L__BB108_1769: mul.wide.s32 %rd2486, %r3267, 16; add.s64 %rd2488, %rd2487, %rd2486; ld.const.v2.u64 {%rd2489, %rd2490}, [%rd2488+2768]; cvta.to.global.u64 %rd2493, %rd2490; add.s64 %rd2495, %rd2493, %rd837; ld.global.u32 %r2544, [%rd2495]; cvta.to.global.u64 %rd2496, %rd2489; add.s64 %rd2497, %rd2496, %rd837; ld.global.u32 %r2545, [%rd2497]; ld.const.v2.u64 {%rd2498, %rd2499}, [%rd2488+3744]; cvta.to.global.u64 %rd2502, %rd2498; shl.b64 %rd2503, %rd577, 2; add.s64 %rd2504, %rd2502, %rd2503; st.global.u32 [%rd2504], %r2545; cvta.to.global.u64 %rd2505, %rd2499; add.s64 %rd2506, %rd2505, %rd2503; st.global.u32 [%rd2506], %r2544; add.s32 %r3267, %r3267, 1; setp.ne.s32 %p1576, %r2544, -1; @%p1576 bra $L__BB108_1769; $L__BB108_1770: ld.const.u64 %rd2507, [kernel_params+3496]; cvta.to.global.u64 %rd2508, %rd2507; mul.lo.s64 %rd2509, %rd577, 12; add.s64 %rd2510, %rd2508, %rd2509; st.global.f32 [%rd2510], %f16413; st.global.f32 [%rd2510+4], %f16414; st.global.f32 [%rd2510+8], %f16415; ld.const.v2.u64 {%rd2511, %rd2512}, [kernel_params+3504]; mov.u32 %r2546, 0; cvta.to.global.u64 %rd2515, %rd2511; add.s64 %rd2516, %rd2515, %rd2509; st.global.f32 [%rd2516], %f2542; st.global.f32 [%rd2516+4], %f2543; st.global.f32 [%rd2516+8], %f2544; cvta.to.global.u64 %rd2517, %rd2512; shl.b64 %rd2518, %rd577, 2; add.s64 %rd2519, %rd2517, %rd2518; st.global.u32 [%rd2519], %r2546; ld.const.v2.u64 {%rd2520, %rd2521}, [kernel_params+3520]; cvta.to.global.u64 %rd2524, %rd2520; add.s64 %rd2525, %rd2524, %rd2518; st.global.f32 [%rd2525], %f2755; cvta.to.global.u64 %rd2526, %rd2521; add.s64 %rd2527, %rd2526, %rd2518; st.global.f32 [%rd2527], %f2756; ld.const.u64 %rd2528, [kernel_params+3536]; cvta.to.global.u64 %rd2529, %rd2528; add.s64 %rd2530, %rd2529, %rd2518; st.global.u32 [%rd2530], %r2546; ld.const.u64 %rd2531, [kernel_params+3584]; cvta.to.global.u64 %rd2532, %rd2531; add.s64 %rd2533, %rd2532, %rd2518; st.global.u32 [%rd2533], %r841; ld.const.u64 %rd2534, [kernel_params+3576]; cvta.to.global.u64 %rd2535, %rd2534; add.s64 %rd2536, %rd2535, %rd2518; st.global.u32 [%rd2536], %r842; ld.const.u64 %rd2537, [kernel_params+3632]; cvta.to.global.u64 %rd2538, %rd2537; add.s64 %rd2539, %rd2538, %rd2518; mov.u32 %r2547, -1; st.global.u32 [%rd2539], %r2547; ld.const.u64 %rd2540, [kernel_params+3624]; cvta.to.global.u64 %rd2541, %rd2540; add.s64 %rd2542, %rd2541, %rd2518; st.global.u32 [%rd2542], %r2547; ld.const.v2.u64 {%rd2543, %rd2544}, [kernel_params+2464]; cvta.to.global.u64 %rd2547, %rd2543; add.s64 %rd2549, %rd2547, %rd1707; ld.global.u16 %rs48, [%rd2549]; cvta.to.global.u64 %rd2550, %rd2544; add.s64 %rd2551, %rd2550, %rd1707; ld.global.u16 %rs49, [%rd2551]; ld.global.u32 %r845, [%rd29]; ld.const.u64 %rd578, [kernel_params+2584]; ld.local.u32 %r846, [%rd7+-36]; and.b32 %r2548, %r846, 524288; setp.ne.s32 %p1577, %r2548, 0; mov.f32 %f16440, %f10816; mov.f32 %f16441, %f10816; mov.f32 %f16442, %f10816; @%p1577 bra $L__BB108_1773; and.b32 %r2549, %r846, 512; setp.eq.s32 %p1578, %r2549, 0; mov.f32 %f16440, %f10820; mov.f32 %f16441, %f10820; mov.f32 %f16442, %f10820; @%p1578 bra $L__BB108_1773; ld.local.v2.f32 {%f16441, %f16442}, [%rd7+208]; ld.local.f32 %f16440, [%rd7+204]; $L__BB108_1773: sub.ftz.f32 %f11266, %f10816, %f16440; sub.ftz.f32 %f11267, %f10816, %f16441; sub.ftz.f32 %f11268, %f10816, %f16442; cvt.ftz.sat.f32.f32 %f11269, %f11266; cvt.ftz.sat.f32.f32 %f11270, %f11267; cvt.ftz.sat.f32.f32 %f11271, %f11268; cvta.to.global.u64 %rd2552, %rd578; mul.lo.s64 %rd2553, %rd2, 12; add.s64 %rd2554, %rd2552, %rd2553; ld.global.f32 %f11272, [%rd2554]; mul.ftz.f32 %f11273, %f11269, %f11272; ld.global.f32 %f11274, [%rd2554+4]; mul.ftz.f32 %f11275, %f11270, %f11274; ld.global.f32 %f11276, [%rd2554+8]; mul.ftz.f32 %f11277, %f11271, %f11276; ld.const.u64 %rd2555, [kernel_params+2448]; cvta.to.global.u64 %rd2556, %rd2555; add.s64 %rd2558, %rd2556, %rd837; ld.global.u32 %r2550, [%rd2558]; ld.const.v2.u64 {%rd2559, %rd2560}, [kernel_params+3344]; cvta.to.global.u64 %rd2563, %rd2559; add.s64 %rd2565, %rd2563, %rd2518; st.global.u32 [%rd2565], %r2550; ld.global.u16 %rs108, [%rd278]; ld.const.v2.u64 {%rd2566, %rd2567}, [kernel_params+3360]; cvta.to.global.u64 %rd2570, %rd2567; shl.b64 %rd2571, %rd577, 1; add.s64 %rd2572, %rd2570, %rd2571; st.global.u16 [%rd2572], %rs108; ld.global.u32 %r2551, [%rd277]; cvta.to.global.u64 %rd2573, %rd2566; add.s64 %rd2574, %rd2573, %rd2518; st.global.u32 [%rd2574], %r2551; ld.global.u32 %r2552, [%rd279]; cvta.to.global.u64 %rd2575, %rd2560; add.s64 %rd2576, %rd2575, %rd2518; st.global.u32 [%rd2576], %r2552; ld.const.v2.u64 {%rd2577, %rd2578}, [kernel_params+3424]; cvta.to.global.u64 %rd2581, %rd2577; add.s64 %rd2582, %rd2581, %rd2518; or.b32 %r2553, %r845, 268435456; st.global.u32 [%rd2582], %r2553; ld.const.v2.u64 {%rd2583, %rd2584}, [kernel_params+3376]; cvta.to.global.u64 %rd2587, %rd2583; add.s64 %rd2588, %rd2587, %rd2571; st.global.u16 [%rd2588], %rs48; cvta.to.global.u64 %rd2589, %rd2584; add.s64 %rd2590, %rd2589, %rd2571; st.global.u16 [%rd2590], %rs49; cvta.to.global.u64 %rd2591, %rd2578; add.s64 %rd2593, %rd2591, %rd2509; st.global.f32 [%rd2593], %f11273; st.global.f32 [%rd2593+4], %f11275; st.global.f32 [%rd2593+8], %f11277; setp.eq.s32 %p1579, %r788, 0; @%p1579 bra $L__BB108_1775; ld.const.u64 %rd2594, [kernel_params+3440]; cvta.to.global.u64 %rd2595, %rd2594; add.s64 %rd2597, %rd2595, %rd2509; st.global.f32 [%rd2597], %f16334; st.global.f32 [%rd2597+4], %f16335; st.global.f32 [%rd2597+8], %f16336; $L__BB108_1775: ld.local.u32 %r847, [%rd7+-36]; and.b32 %r2555, %r847, 20; setp.eq.s32 %p1580, %r2555, 0; mov.u32 %r3294, 0; @%p1580 bra $L__BB108_2405; add.s32 %r848, %r602, 3; ld.const.u32 %r2556, [kernel_params+1596]; setp.eq.s32 %p1581, %r2556, 0; brev.b32 %r2557, %r603; mul.lo.s32 %r2558, %r2557, 1025551850; xor.b32 %r849, %r2558, %r2557; @%p1581 bra $L__BB108_1779; ld.const.v2.u32 {%r2559, %r2560}, [kernel_params+1600]; mov.b32 %f2763, %r2560; setp.geu.ftz.f32 %p1582, %f2763, 0f3F800000; setp.lt.ftz.f32 %p1583, %f2763, 0f3F800000; ld.const.u32 %r2563, [kernel_params+1576]; selp.b32 %r2564, %r2563, %r601, %p1583; add.s16 %rs109, %rs28, 3; and.b16 %rs110, %rs109, 255; cvt.u32.u16 %r2565, %rs110; xor.b32 %r2566, %r2564, %r2565; mul.lo.s32 %r2567, %r2566, -512718531; shr.u32 %r2568, %r2564, 16; xor.b32 %r2569, %r2567, %r2568; bfe.u32 %r2570, %r2569, 4, 4; shr.u32 %r2571, %r2564, 8; xor.b32 %r2572, %r2569, %r2571; xor.b32 %r2573, %r2572, %r2570; mul.lo.s32 %r2574, %r2573, 153742143; shr.u32 %r2575, %r2564, 23; xor.b32 %r2576, %r2574, %r2575; bfe.u32 %r2577, %r2576, 1, 7; xor.b32 %r2578, %r2577, %r2576; shr.u32 %r2579, %r2564, 27; or.b32 %r2580, %r2579, 1; mul.lo.s32 %r2581, %r2580, %r2578; mul.lo.s32 %r2582, %r2581, -1074439621; bfe.u32 %r2583, %r2582, 2, 6; xor.b32 %r2584, %r2583, %r2582; mul.lo.s32 %r2585, %r2584, -1638916925; bfe.u32 %r2586, %r2585, 2, 6; xor.b32 %r2587, %r2586, %r2585; mul.lo.s32 %r2588, %r2587, -933190689; and.b32 %r2589, %r2588, 255; bfe.u32 %r2590, %r2588, 5, 3; xor.b32 %r2591, %r2590, %r2589; add.s32 %r2592, %r2559, -1; xor.b32 %r850, %r848, 61; xor.b32 %r2593, %r2564, %r850; mul.lo.s32 %r2594, %r2593, 9; shr.u32 %r2595, %r2594, 4; xor.b32 %r2596, %r2595, %r2594; mul.lo.s32 %r2597, %r2596, 668265261; add.s32 %r2598, %r2597, %r849; shr.u32 %r2599, %r2597, 16; or.b32 %r2600, %r2599, 1; mul.lo.s32 %r2601, %r2600, %r2598; mul.lo.s32 %r2602, %r2601, 89287766; xor.b32 %r2603, %r2602, %r2601; mul.lo.s32 %r2604, %r2603, 1403136100; xor.b32 %r2605, %r2604, %r2603; brev.b32 %r2606, %r2605; neg.s32 %r2607, %r2559; and.b32 %r2608, %r603, %r2607; and.b32 %r2609, %r2606, %r2592; or.b32 %r2610, %r2609, %r2608; mad.lo.s32 %r2611, %r2591, %r2559, %r2610; shl.b32 %r2612, %r2559, 8; rem.u32 %r2613, %r2611, %r2612; ld.const.u64 %rd2598, [kernel_params+2424]; cvta.to.global.u64 %rd2599, %rd2598; shl.b32 %r2614, %r2613, 1; mul.wide.u32 %rd2600, %r2614, 4; add.s64 %rd2601, %rd2599, %rd2600; ld.global.f32 %f2799, [%rd2601]; ld.global.f32 %f2775, [%rd2601+4]; @%p1582 bra $L__BB108_1783; xor.b32 %r2615, %r850, %r601; mul.lo.s32 %r2616, %r2615, 9; shr.u32 %r2617, %r2616, 4; xor.b32 %r2618, %r2617, %r2616; mul.lo.s32 %r2619, %r2618, 668265261; cvt.rn.f32.u32 %f11278, %r2619; mul.ftz.f32 %f11279, %f11278, 0f2F7FFFFE; xor.b32 %r2620, %r601, %r848; xor.b32 %r2621, %r2620, -905047700; mul.lo.s32 %r2622, %r2621, 9; shr.u32 %r2623, %r2622, 4; xor.b32 %r2624, %r2623, %r2622; mul.lo.s32 %r2625, %r2624, 668265261; cvt.rn.f32.u32 %f11280, %r2625; mul.ftz.f32 %f11281, %f11280, 0f2F7FFFFE; fma.rn.ftz.f32 %f11282, %f11279, %f2763, %f2799; fma.rn.ftz.f32 %f11283, %f11281, %f2763, %f2775; cvt.rmi.ftz.f32.f32 %f11284, %f11282; sub.ftz.f32 %f2799, %f11282, %f11284; cvt.rmi.ftz.f32.f32 %f11285, %f11283; sub.ftz.f32 %f2775, %f11283, %f11285; bra.uni $L__BB108_1783; $L__BB108_1779: shr.u32 %r2627, %r848, 16; xor.b32 %r2628, %r2627, %r848; mul.lo.s32 %r2629, %r2628, 569420461; shr.u32 %r2630, %r2629, 15; xor.b32 %r2631, %r2630, %r2629; mul.lo.s32 %r2632, %r2631, -749064809; shr.u32 %r2633, %r2632, 15; xor.b32 %r2634, %r601, %r2632; xor.b32 %r2635, %r2634, %r2633; xor.b32 %r851, %r2635, -419546133; xor.b32 %r2636, %r2635, 508809841; add.s32 %r2637, %r2636, %r849; shr.u32 %r2638, %r2636, 16; or.b32 %r2639, %r2638, 1; mul.lo.s32 %r2640, %r2639, %r2637; mul.lo.s32 %r2641, %r2640, 89287766; xor.b32 %r2642, %r2641, %r2640; mul.lo.s32 %r2643, %r2642, 1403136100; xor.b32 %r852, %r2643, %r2642; setp.eq.s32 %p1584, %r852, 0; mov.u32 %r3271, 0; @%p1584 bra $L__BB108_1782; mov.u32 %r3268, 0; mov.u64 %rd2603, sobol_burley_table; mov.u32 %r3269, %r852; mov.u32 %r3271, %r3268; $L__BB108_1781: clz.b32 %r2646, %r3269; add.s32 %r2647, %r2646, %r3268; mul.wide.u32 %rd2602, %r2647, 4; add.s64 %rd2604, %rd2603, %rd2602; ld.const.u32 %r2648, [%rd2604+128]; xor.b32 %r3271, %r2648, %r3271; add.s32 %r3268, %r2647, 1; shl.b32 %r2649, %r3269, %r2646; shl.b32 %r3269, %r2649, 1; setp.ne.s32 %p1585, %r3269, 0; @%p1585 bra $L__BB108_1781; $L__BB108_1782: xor.b32 %r2650, %r851, -525684874; brev.b32 %r2651, %r852; shr.u32 %r2652, %r2650, 16; or.b32 %r2653, %r2652, 1; mul.lo.s32 %r2654, %r2651, 1025551850; xor.b32 %r2655, %r2654, %r2651; add.s32 %r2656, %r2655, %r2650; mul.lo.s32 %r2657, %r2656, %r2653; mul.lo.s32 %r2658, %r2657, 89287766; xor.b32 %r2659, %r2658, %r2657; xor.b32 %r2660, %r851, -1802089138; mul.lo.s32 %r2661, %r3271, 1025551850; xor.b32 %r2662, %r2661, %r3271; add.s32 %r2663, %r2662, %r2660; shr.u32 %r2664, %r2660, 16; or.b32 %r2665, %r2664, 1; mul.lo.s32 %r2666, %r2663, %r2665; mul.lo.s32 %r2667, %r2666, 89287766; xor.b32 %r2668, %r2667, %r2666; mul.lo.s32 %r2669, %r2668, 1403136100; xor.b32 %r2670, %r2669, %r2668; brev.b32 %r2671, %r2670; cvt.rn.f32.u32 %f11286, %r2671; mul.lo.s32 %r2672, %r2659, 1403136100; xor.b32 %r2673, %r2672, %r2659; brev.b32 %r2674, %r2673; cvt.rn.f32.u32 %f11287, %r2674; mul.ftz.f32 %f2775, %f11286, 0f2F7FFFFE; mul.ftz.f32 %f2799, %f11287, 0f2F7FFFFE; $L__BB108_1783: ld.local.u32 %r860, [%rd7+172]; setp.lt.s32 %p1586, %r860, 2; mov.u32 %r3277, 0; @%p1586 bra $L__BB108_1805; add.s32 %r2677, %r860, -1; and.b32 %r3275, %r860, 3; setp.lt.u32 %p1587, %r2677, 3; mov.f32 %f16448, 0f00000000; mov.u32 %r3274, 0; @%p1587 bra $L__BB108_1795; sub.s32 %r3273, %r860, %r3275; mov.f32 %f16448, 0f00000000; mov.u32 %r3274, 0; $L__BB108_1786: cvt.s64.s32 %rd579, %r3274; mul.wide.s32 %rd2605, %r3274, 80; add.s64 %rd2606, %rd1, %rd2605; add.s64 %rd580, %rd2606, 316; ld.local.u32 %r2679, [%rd2606+316]; setp.gt.s32 %p1588, %r2679, 34; @%p1588 bra $L__BB108_1788; ld.local.f32 %f11291, [%rd580+4]; add.ftz.f32 %f16448, %f16448, %f11291; $L__BB108_1788: ld.local.u32 %r2680, [%rd580+80]; setp.gt.s32 %p1589, %r2680, 34; @%p1589 bra $L__BB108_1790; ld.local.f32 %f11292, [%rd580+84]; add.ftz.f32 %f16448, %f16448, %f11292; $L__BB108_1790: ld.local.u32 %r2681, [%rd580+160]; setp.gt.s32 %p1590, %r2681, 34; @%p1590 bra $L__BB108_1792; ld.local.f32 %f11293, [%rd580+164]; add.ftz.f32 %f16448, %f16448, %f11293; $L__BB108_1792: ld.local.u32 %r2682, [%rd580+240]; setp.gt.s32 %p1591, %r2682, 34; @%p1591 bra $L__BB108_1794; ld.local.f32 %f11294, [%rd580+244]; add.ftz.f32 %f16448, %f16448, %f11294; $L__BB108_1794: cvt.u32.u64 %r2683, %rd579; add.s32 %r3274, %r2683, 4; add.s32 %r3273, %r3273, -4; setp.ne.s32 %p1592, %r3273, 0; @%p1592 bra $L__BB108_1786; $L__BB108_1795: setp.eq.s32 %p1593, %r3275, 0; @%p1593 bra $L__BB108_1800; mul.wide.s32 %rd2607, %r3274, 80; add.s64 %rd2608, %rd1, %rd2607; add.s64 %rd3067, %rd2608, 320; $L__BB108_1797: .pragma "nounroll"; ld.local.u32 %r2684, [%rd3067+-4]; setp.gt.s32 %p1594, %r2684, 34; @%p1594 bra $L__BB108_1799; ld.local.f32 %f11295, [%rd3067]; add.ftz.f32 %f16448, %f16448, %f11295; $L__BB108_1799: add.s64 %rd3067, %rd3067, 80; add.s32 %r3275, %r3275, -1; setp.ne.s32 %p1595, %r3275, 0; @%p1595 bra $L__BB108_1797; $L__BB108_1800: mul.ftz.f32 %f2793, %f2799, %f16448; mov.f32 %f16457, 0f00000000; mov.u32 %r3277, 0; mov.u32 %r3276, %r3277; mov.f32 %f16458, %f16457; $L__BB108_1801: mul.wide.s32 %rd2609, %r3276, 80; add.s64 %rd2610, %rd1, %rd2609; add.s64 %rd585, %rd2610, 316; ld.local.u32 %r2686, [%rd2610+316]; setp.gt.s32 %p1596, %r2686, 34; @%p1596 bra $L__BB108_1803; ld.local.f32 %f2795, [%rd585+4]; add.ftz.f32 %f16458, %f16457, %f2795; setp.lt.ftz.f32 %p1597, %f2793, %f16458; @%p1597 bra $L__BB108_1804; $L__BB108_1803: add.s32 %r3276, %r3276, 1; setp.lt.s32 %p1598, %r3276, %r860; mov.f32 %f16457, %f16458; @%p1598 bra $L__BB108_1801; bra.uni $L__BB108_1805; $L__BB108_1804: sub.ftz.f32 %f11297, %f2793, %f16457; div.approx.ftz.f32 %f2799, %f11297, %f2795; mov.u32 %r3277, %r3276; $L__BB108_1805: cvt.s64.s32 %rd586, %r3277; mul.wide.s32 %rd2611, %r3277, 80; add.s64 %rd2612, %rd1, %rd2611; add.s64 %rd587, %rd2612, 316; ld.local.u32 %r873, [%rd2612+316]; add.s32 %r2688, %r873, -32; setp.lt.u32 %p1599, %r2688, 3; @%p1599 bra $L__BB108_2383; bra.uni $L__BB108_1806; $L__BB108_2383: ld.const.u64 %rd2910, [kernel_params+2632]; mul.wide.s32 %rd2909, %r953, 12; cvta.to.global.u64 %rd2908, %rd2910; add.s64 %rd2907, %rd2908, %rd2909; ld.local.v4.f32 {%f15419, %f15420, %f15421, %f15422}, [%rd1]; st.global.f32 [%rd2907], %f15419; st.global.f32 [%rd2907+4], %f15420; st.global.f32 [%rd2907+8], %f15421; ld.const.v2.u64 {%rd2790, %rd2791}, [kernel_params+2640]; cvta.to.global.u64 %rd2794, %rd2790; ld.local.f32 %f15426, [%rd587+8]; ld.local.f32 %f15427, [%rd587+12]; ld.local.f32 %f15428, [%rd587+16]; mul.lo.s64 %rd2795, %rd2, 12; add.s64 %rd2796, %rd2794, %rd2795; st.global.f32 [%rd2796], %f15426; st.global.f32 [%rd2796+4], %f15427; st.global.f32 [%rd2796+8], %f15428; cvta.to.global.u64 %rd2797, %rd2791; add.s64 %rd2799, %rd2797, %rd837; mov.u32 %r2999, 0; st.global.u32 [%rd2799], %r2999; ld.const.u64 %rd2800, [kernel_params+2656]; cvta.to.global.u64 %rd2801, %rd2800; add.s64 %rd2802, %rd2801, %rd837; mov.u32 %r3000, 2139095039; st.global.u32 [%rd2802], %r3000; ld.local.f32 %f15429, [%rd7+4]; add.s64 %rd2804, %rd656, %rd837; st.global.f32 [%rd2804], %f15429; add.s64 %rd2806, %rd658, %rd837; st.global.u32 [%rd2806], %r2999; ld.const.u64 %rd2807, [kernel_params+2760]; cvta.to.global.u64 %rd2808, %rd2807; ld.local.v2.f32 {%f15430, %f15431}, [%rd7+-64]; ld.local.f32 %f15434, [%rd7+-56]; add.s64 %rd2809, %rd2808, %rd2795; st.global.f32 [%rd2809], %f15430; st.global.f32 [%rd2809+4], %f15431; st.global.f32 [%rd2809+8], %f15434; ld.global.u32 %r3001, [%rd29]; and.b32 %r3002, %r3001, -2; setp.eq.s32 %p2146, %r873, 32; selp.b32 %r3003, 2097152, 1048576, %p2146; or.b32 %r924, %r3003, %r3002; ld.local.f32 %f16794, [%rd587+-12]; ld.local.f32 %f16795, [%rd587+-8]; ld.local.f32 %f16796, [%rd587+-4]; @%p1586 bra $L__BB108_2401; add.s32 %r3005, %r860, -1; and.b32 %r3293, %r860, 3; setp.lt.u32 %p2147, %r3005, 3; mov.f32 %f16785, 0f00000000; mov.u32 %r3292, %r2999; @%p2147 bra $L__BB108_2395; sub.s32 %r3291, %r860, %r3293; mov.f32 %f16785, 0f00000000; mov.u32 %r3292, 0; $L__BB108_2386: cvt.s64.s32 %rd596, %r3292; mul.wide.s32 %rd2810, %r3292, 80; add.s64 %rd2811, %rd1, %rd2810; add.s64 %rd597, %rd2811, 316; ld.local.u32 %r3007, [%rd2811+316]; setp.gt.s32 %p2148, %r3007, 34; @%p2148 bra $L__BB108_2388; ld.local.f32 %f15438, [%rd597+4]; add.ftz.f32 %f16785, %f16785, %f15438; $L__BB108_2388: ld.local.u32 %r3008, [%rd597+80]; setp.gt.s32 %p2149, %r3008, 34; @%p2149 bra $L__BB108_2390; ld.local.f32 %f15439, [%rd597+84]; add.ftz.f32 %f16785, %f16785, %f15439; $L__BB108_2390: ld.local.u32 %r3009, [%rd597+160]; setp.gt.s32 %p2150, %r3009, 34; @%p2150 bra $L__BB108_2392; ld.local.f32 %f15440, [%rd597+164]; add.ftz.f32 %f16785, %f16785, %f15440; $L__BB108_2392: ld.local.u32 %r3010, [%rd597+240]; setp.gt.s32 %p2151, %r3010, 34; @%p2151 bra $L__BB108_2394; ld.local.f32 %f15441, [%rd597+244]; add.ftz.f32 %f16785, %f16785, %f15441; $L__BB108_2394: cvt.u32.u64 %r3011, %rd596; add.s32 %r3292, %r3011, 4; add.s32 %r3291, %r3291, -4; setp.ne.s32 %p2152, %r3291, 0; @%p2152 bra $L__BB108_2386; $L__BB108_2395: setp.eq.s32 %p2153, %r3293, 0; @%p2153 bra $L__BB108_2400; mul.wide.s32 %rd2812, %r3292, 80; add.s64 %rd2813, %rd1, %rd2812; add.s64 %rd3068, %rd2813, 320; $L__BB108_2397: .pragma "nounroll"; add.s64 %rd600, %rd3068, -4; ld.local.u32 %r3012, [%rd3068+-4]; setp.gt.s32 %p2154, %r3012, 34; @%p2154 bra $L__BB108_2399; ld.local.f32 %f15442, [%rd600+4]; add.ftz.f32 %f16785, %f16785, %f15442; $L__BB108_2399: add.s64 %rd3068, %rd3068, 80; add.s32 %r3293, %r3293, -1; setp.ne.s32 %p2155, %r3293, 0; @%p2155 bra $L__BB108_2397; $L__BB108_2400: ld.local.f32 %f15443, [%rd587+4]; div.approx.ftz.f32 %f15444, %f16785, %f15443; mul.ftz.f32 %f16794, %f16794, %f15444; mul.ftz.f32 %f16795, %f16795, %f15444; mul.ftz.f32 %f16796, %f16796, %f15444; $L__BB108_2401: ld.local.f32 %f15445, [%rd587+44]; setp.neu.ftz.f32 %p2156, %f15445, 0f7F7FFFFF; or.b32 %r3013, %r924, 4194304; selp.b32 %r3014, %r3013, %r924, %p2156; shl.b32 %r3015, %r847, 23; and.b32 %r3016, %r3015, 8388608; or.b32 %r3017, %r3016, %r3014; ld.const.u64 %rd2814, [kernel_params+2584]; cvta.to.global.u64 %rd2815, %rd2814; add.s64 %rd2817, %rd2815, %rd2795; ld.global.f32 %f15446, [%rd2817]; mul.ftz.f32 %f15447, %f16794, %f15446; ld.global.f32 %f15448, [%rd2817+4]; mul.ftz.f32 %f15449, %f16795, %f15448; ld.global.f32 %f15450, [%rd2817+8]; mul.ftz.f32 %f15451, %f16796, %f15450; st.global.f32 [%rd2817], %f15447; st.global.f32 [%rd2817+4], %f15449; st.global.f32 [%rd2817+8], %f15451; st.global.u32 [%rd29], %r3017; ld.global.u16 %rs130, [%rd278]; add.s16 %rs131, %rs130, 16; st.global.u16 [%rd278], %rs131; and.b32 %r3018, %r787, 2097152; setp.eq.s32 %p2157, %r3018, 0; @%p2157 bra $L__BB108_2404; ld.const.u64 %rd2818, [kernel_params+2464]; cvta.to.global.u64 %rd2819, %rd2818; add.s64 %rd2821, %rd2819, %rd1707; ld.global.u16 %rs132, [%rd2821]; setp.ne.s16 %p2158, %rs132, 0; @%p2158 bra $L__BB108_2404; ld.const.u64 %rd2822, [kernel_params+2600]; cvta.to.global.u64 %rd2823, %rd2822; add.s64 %rd2825, %rd2823, %rd2795; mov.u32 %r3019, 1065353216; st.global.u32 [%rd2825], %r3019; st.global.u32 [%rd2825+4], %r3019; st.global.u32 [%rd2825+8], %r3019; ld.const.u64 %rd2826, [kernel_params+2608]; cvta.to.global.u64 %rd2827, %rd2826; add.s64 %rd2828, %rd2827, %rd2795; st.global.u32 [%rd2828], %r2999; st.global.u32 [%rd2828+4], %r2999; st.global.u32 [%rd2828+8], %r2999; $L__BB108_2404: ld.const.v2.u64 {%rd2829, %rd2830}, [kernel_params+2736]; cvta.to.global.u64 %rd2833, %rd2829; ld.local.f32 %f15452, [%rd587+32]; ld.local.f32 %f15453, [%rd587+36]; ld.local.f32 %f15454, [%rd587+40]; add.s64 %rd2835, %rd2833, %rd2795; st.global.f32 [%rd2835], %f15452; st.global.f32 [%rd2835+4], %f15453; st.global.f32 [%rd2835+8], %f15454; cvta.to.global.u64 %rd2836, %rd2830; ld.local.f32 %f15455, [%rd587+20]; ld.local.f32 %f15456, [%rd587+24]; ld.local.f32 %f15457, [%rd587+28]; add.s64 %rd2837, %rd2836, %rd2795; st.global.f32 [%rd2837], %f15455; st.global.f32 [%rd2837+4], %f15456; st.global.f32 [%rd2837+8], %f15457; ld.local.f32 %f15458, [%rd587+48]; ld.const.u64 %rd2838, [kernel_params+2752]; cvta.to.global.u64 %rd2839, %rd2838; add.s64 %rd2841, %rd2839, %rd837; st.global.f32 [%rd2841], %f15458; mov.u32 %r3294, 256; bra.uni $L__BB108_2405; $L__BB108_1806: ld.local.u8 %rs111, [%rd7+-24]; and.b16 %rs112, %rs111, 6; setp.eq.s16 %p1600, %rs112, 0; @%p1600 bra $L__BB108_1808; ld.local.f32 %f2808, [%rd587+16]; ld.local.f32 %f2807, [%rd587+12]; ld.local.f32 %f2806, [%rd587+8]; bra.uni $L__BB108_1809; $L__BB108_1808: ld.local.v2.f32 {%f2806, %f2807}, [%rd7+-64]; ld.local.f32 %f2808, [%rd7+-56]; $L__BB108_1809: mov.u32 %r3285, 0; mov.f32 %f16743, 0f00000000; setp.gt.s32 %p1601, %r873, 15; @%p1601 bra $L__BB108_1836; setp.gt.s32 %p1617, %r873, 9; @%p1617 bra $L__BB108_1833; setp.gt.s32 %p1621, %r873, 5; @%p1621 bra $L__BB108_1819; bra.uni $L__BB108_1812; $L__BB108_1819: setp.gt.s32 %p1622, %r873, 7; @%p1622 bra $L__BB108_1828; setp.eq.s32 %p1625, %r873, 6; @%p1625 bra $L__BB108_2294; setp.eq.s32 %p1626, %r873, 7; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1626 bra $L__BB108_1822; bra.uni $L__BB108_2309; $L__BB108_1822: ld.local.f32 %f14854, [%rd587+20]; mul.ftz.f32 %f4047, %f14854, 0f3FC90FDB; mov.f32 %f14855, 0f3FC90FDB; ld.local.f32 %f14856, [%rd587+24]; mul.ftz.f32 %f4048, %f14856, 0f3FC90FDB; add.ftz.f32 %f4049, %f4047, %f4048; min.ftz.f32 %f4050, %f4049, %f14855; mul.ftz.f32 %f4051, %f2799, %f4050; setp.leu.ftz.f32 %p2028, %f4050, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2028 bra $L__BB108_2309; cos.approx.ftz.f32 %f14861, %f4050; mul.ftz.f32 %f14862, %f2799, %f14861; sub.ftz.f32 %f14863, %f14861, %f14862; add.ftz.f32 %f14864, %f2799, %f14863; mul.ftz.f32 %f14865, %f14864, %f14864; mov.f32 %f16733, 0f3F800000; sub.ftz.f32 %f14867, %f16733, %f14865; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f14868, %f14867, %f16743; sqrt.approx.ftz.f32 %f14869, %f14868; mul.ftz.f32 %f14870, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f14871, %f14870; mul.ftz.f32 %f14872, %f14869, %f14871; sin.approx.ftz.f32 %f14873, %f14870; mul.ftz.f32 %f14874, %f14869, %f14873; ld.local.f32 %f14875, [%rd587+12]; ld.local.f32 %f14876, [%rd587+8]; setp.neu.ftz.f32 %p2029, %f14876, %f14875; ld.local.f32 %f14877, [%rd587+16]; setp.neu.ftz.f32 %p2030, %f14876, %f14877; or.pred %p2031, %p2029, %p2030; sub.ftz.f32 %f14878, %f14877, %f14875; neg.ftz.f32 %f14879, %f14875; neg.ftz.f32 %f14880, %f14877; selp.f32 %f14881, %f14880, %f14877, %p2031; add.ftz.f32 %f14882, %f14876, %f14881; selp.f32 %f14883, %f14875, %f14879, %p2031; sub.ftz.f32 %f14884, %f14883, %f14876; mul.ftz.f32 %f14885, %f14882, %f14882; fma.rn.ftz.f32 %f14886, %f14878, %f14878, %f14885; fma.rn.ftz.f32 %f14887, %f14884, %f14884, %f14886; sqrt.approx.ftz.f32 %f14888, %f14887; div.approx.ftz.f32 %f14889, %f14878, %f14888; div.approx.ftz.f32 %f14890, %f14882, %f14888; div.approx.ftz.f32 %f14891, %f14884, %f14888; mul.ftz.f32 %f14892, %f14875, %f14891; mul.ftz.f32 %f14893, %f14877, %f14890; sub.ftz.f32 %f14894, %f14892, %f14893; mul.ftz.f32 %f14895, %f14877, %f14889; mul.ftz.f32 %f14896, %f14876, %f14891; sub.ftz.f32 %f14897, %f14895, %f14896; mul.ftz.f32 %f14898, %f14876, %f14890; mul.ftz.f32 %f14899, %f14875, %f14889; sub.ftz.f32 %f14900, %f14898, %f14899; mul.ftz.f32 %f14901, %f14874, %f14894; mul.ftz.f32 %f14902, %f14874, %f14897; mul.ftz.f32 %f14903, %f14874, %f14900; fma.rn.ftz.f32 %f14904, %f14872, %f14889, %f14901; fma.rn.ftz.f32 %f14905, %f14872, %f14890, %f14902; fma.rn.ftz.f32 %f14906, %f14872, %f14891, %f14903; fma.rn.ftz.f32 %f16740, %f14876, %f14864, %f14904; fma.rn.ftz.f32 %f16741, %f14875, %f14864, %f14905; fma.rn.ftz.f32 %f16742, %f14877, %f14864, %f14906; sub.ftz.f32 %f14907, %f16733, %f14861; mov.f32 %f14908, 0f3E22F983; div.approx.ftz.f32 %f4055, %f14908, %f14907; mul.ftz.f32 %f14909, %f2807, %f16741; fma.rn.ftz.f32 %f14910, %f2806, %f16740, %f14909; fma.rn.ftz.f32 %f14911, %f2808, %f16742, %f14910; setp.leu.ftz.f32 %p2032, %f14911, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2032 bra $L__BB108_2309; setp.lt.ftz.f32 %p2033, %f4051, %f4047; @%p2033 bra $L__BB108_1827; setp.leu.ftz.f32 %p2034, %f4049, %f4051; setp.eq.ftz.f32 %p2035, %f4048, 0f00000000; mov.f32 %f16733, 0f00000000; or.pred %p2036, %p2035, %p2034; @%p2036 bra $L__BB108_1827; sub.ftz.f32 %f14914, %f4051, %f4047; div.approx.ftz.f32 %f14915, %f14914, %f4048; mov.f32 %f14916, 0f3F800000; sub.ftz.f32 %f16733, %f14916, %f14915; $L__BB108_1827: mul.ftz.f32 %f16744, %f4055, %f16733; mov.f32 %f16743, %f4055; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_1836: setp.gt.s32 %p1602, %r873, 22; @%p1602 bra $L__BB108_1853; setp.gt.s32 %p1610, %r873, 19; @%p1610 bra $L__BB108_1848; setp.eq.s32 %p1614, %r873, 16; @%p1614 bra $L__BB108_2113; setp.eq.s32 %p1615, %r873, 17; @%p1615 bra $L__BB108_2108; setp.eq.s32 %p1616, %r873, 19; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1616 bra $L__BB108_1841; bra.uni $L__BB108_2309; $L__BB108_1841: ld.local.v2.f32 {%f12954, %f12955}, [%rd7+-48]; ld.local.f32 %f12956, [%rd587+20]; mul.ftz.f32 %f3454, %f12956, 0f3FC90FDB; ld.local.f32 %f12957, [%rd587+24]; mul.ftz.f32 %f3455, %f12957, 0f3FC90FDB; ld.local.f32 %f3456, [%rd587+8]; ld.local.f32 %f3457, [%rd7+-52]; ld.local.f32 %f3458, [%rd587+12]; mul.ftz.f32 %f12958, %f12954, %f3458; fma.rn.ftz.f32 %f12959, %f3457, %f3456, %f12958; ld.local.f32 %f3459, [%rd587+16]; fma.rn.ftz.f32 %f3460, %f12955, %f3459, %f12959; setp.leu.ftz.f32 %p1861, %f3460, 0f00000000; mov.u32 %r3285, 10; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1861 bra $L__BB108_2309; add.ftz.f32 %f12964, %f3460, %f3460; mul.ftz.f32 %f12965, %f12964, %f3456; mul.ftz.f32 %f12966, %f12964, %f3458; mul.ftz.f32 %f12967, %f12964, %f3459; sub.ftz.f32 %f12968, %f12965, %f3457; sub.ftz.f32 %f12969, %f12966, %f12954; sub.ftz.f32 %f12970, %f12967, %f12955; add.ftz.f32 %f3461, %f3454, %f3455; mov.f32 %f12971, 0f3FC90FDB; min.ftz.f32 %f12972, %f3461, %f12971; mul.ftz.f32 %f3462, %f2799, %f12972; cos.approx.ftz.f32 %f12973, %f12972; mul.ftz.f32 %f12974, %f2799, %f12973; sub.ftz.f32 %f12975, %f12973, %f12974; add.ftz.f32 %f12976, %f2799, %f12975; mul.ftz.f32 %f12977, %f12976, %f12976; mov.f32 %f16571, 0f3F800000; sub.ftz.f32 %f12979, %f16571, %f12977; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f12980, %f12979, %f16743; sqrt.approx.ftz.f32 %f12981, %f12980; mul.ftz.f32 %f12982, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f12983, %f12982; mul.ftz.f32 %f12984, %f12981, %f12983; sin.approx.ftz.f32 %f12985, %f12982; mul.ftz.f32 %f12986, %f12981, %f12985; setp.neu.ftz.f32 %p1862, %f12968, %f12969; setp.neu.ftz.f32 %p1863, %f12968, %f12970; or.pred %p1864, %p1862, %p1863; sub.ftz.f32 %f12987, %f12970, %f12969; neg.ftz.f32 %f12988, %f12969; neg.ftz.f32 %f12989, %f12970; selp.f32 %f12990, %f12989, %f12970, %p1864; add.ftz.f32 %f12991, %f12968, %f12990; selp.f32 %f12992, %f12969, %f12988, %p1864; sub.ftz.f32 %f12993, %f12992, %f12968; mul.ftz.f32 %f12994, %f12991, %f12991; fma.rn.ftz.f32 %f12995, %f12987, %f12987, %f12994; fma.rn.ftz.f32 %f12996, %f12993, %f12993, %f12995; sqrt.approx.ftz.f32 %f12997, %f12996; div.approx.ftz.f32 %f12998, %f12987, %f12997; div.approx.ftz.f32 %f12999, %f12991, %f12997; div.approx.ftz.f32 %f13000, %f12993, %f12997; mul.ftz.f32 %f13001, %f12969, %f13000; mul.ftz.f32 %f13002, %f12970, %f12999; sub.ftz.f32 %f13003, %f13001, %f13002; mul.ftz.f32 %f13004, %f12970, %f12998; mul.ftz.f32 %f13005, %f12968, %f13000; sub.ftz.f32 %f13006, %f13004, %f13005; mul.ftz.f32 %f13007, %f12968, %f12999; mul.ftz.f32 %f13008, %f12969, %f12998; sub.ftz.f32 %f13009, %f13007, %f13008; mul.ftz.f32 %f13010, %f12986, %f13003; mul.ftz.f32 %f13011, %f12986, %f13006; mul.ftz.f32 %f13012, %f12986, %f13009; fma.rn.ftz.f32 %f13013, %f12984, %f12998, %f13010; fma.rn.ftz.f32 %f13014, %f12984, %f12999, %f13011; fma.rn.ftz.f32 %f13015, %f12984, %f13000, %f13012; fma.rn.ftz.f32 %f16740, %f12968, %f12976, %f13013; fma.rn.ftz.f32 %f16741, %f12969, %f12976, %f13014; fma.rn.ftz.f32 %f16742, %f12970, %f12976, %f13015; sub.ftz.f32 %f13016, %f16571, %f12973; mov.f32 %f13017, 0f3E22F983; div.approx.ftz.f32 %f3466, %f13017, %f13016; mul.ftz.f32 %f13018, %f2807, %f16741; fma.rn.ftz.f32 %f13019, %f2806, %f16740, %f13018; fma.rn.ftz.f32 %f13020, %f2808, %f16742, %f13019; setp.leu.ftz.f32 %p1865, %f13020, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1865 bra $L__BB108_2309; mul.ftz.f32 %f13025, %f16741, %f3458; fma.rn.ftz.f32 %f13026, %f16740, %f3456, %f13025; fma.rn.ftz.f32 %f13027, %f16742, %f3459, %f13026; setp.leu.ftz.f32 %p1866, %f13027, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1866 bra $L__BB108_2309; setp.lt.ftz.f32 %p1867, %f3462, %f3454; @%p1867 bra $L__BB108_1847; setp.leu.ftz.f32 %p1868, %f3461, %f3462; setp.eq.ftz.f32 %p1869, %f3455, 0f00000000; mov.f32 %f16571, 0f00000000; or.pred %p1870, %p1869, %p1868; @%p1870 bra $L__BB108_1847; sub.ftz.f32 %f13030, %f3462, %f3454; div.approx.ftz.f32 %f13031, %f13030, %f3455; mov.f32 %f13032, 0f3F800000; sub.ftz.f32 %f16571, %f13032, %f13031; $L__BB108_1847: mul.ftz.f32 %f16744, %f3466, %f16571; mov.f32 %f16743, %f3466; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_1833: add.s32 %r2690, %r873, -10; setp.lt.u32 %p1618, %r2690, 3; @%p1618 bra $L__BB108_2246; add.s32 %r2691, %r873, -14; setp.lt.u32 %p1619, %r2691, 2; @%p1619 bra $L__BB108_2143; bra.uni $L__BB108_1835; $L__BB108_2143: ld.local.v2.f32 {%f13466, %f13467}, [%rd7+-48]; ld.local.f32 %f3573, [%rd587+8]; ld.local.f32 %f3574, [%rd7+-52]; ld.local.f32 %f3575, [%rd587+12]; mul.ftz.f32 %f13468, %f13466, %f3575; fma.rn.ftz.f32 %f13469, %f3574, %f3573, %f13468; ld.local.f32 %f3576, [%rd587+16]; fma.rn.ftz.f32 %f3577, %f13467, %f3576, %f13469; setp.le.ftz.f32 %p1906, %f3577, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1906 bra $L__BB108_2309; ld.local.v2.f32 {%f13470, %f13471}, [%rd587+20]; mul.ftz.f32 %f13472, %f13470, %f13471; setp.lt.ftz.f32 %p1907, %f13472, 0f33D6BF95; @%p1907 bra $L__BB108_2196; bra.uni $L__BB108_2145; $L__BB108_2196: add.ftz.f32 %f13803, %f3577, %f3577; mul.ftz.f32 %f13804, %f3573, %f13803; mul.ftz.f32 %f13805, %f3575, %f13803; mul.ftz.f32 %f13806, %f3576, %f13803; sub.ftz.f32 %f16740, %f13804, %f3574; sub.ftz.f32 %f16741, %f13805, %f13466; sub.ftz.f32 %f16742, %f13806, %f13467; mul.ftz.f32 %f13807, %f2807, %f16741; fma.rn.ftz.f32 %f13808, %f2806, %f16740, %f13807; fma.rn.ftz.f32 %f13809, %f2808, %f16742, %f13808; setp.le.ftz.f32 %p1954, %f13809, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1954 bra $L__BB108_2309; mov.u32 %r3285, 18; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1853: setp.gt.s32 %p1603, %r873, 28; @%p1603 bra $L__BB108_1870; setp.eq.s32 %p1607, %r873, 23; @%p1607 bra $L__BB108_2246; setp.eq.s32 %p1608, %r873, 24; @%p1608 bra $L__BB108_1857; setp.eq.s32 %p1609, %r873, 27; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1609 bra $L__BB108_1857; bra.uni $L__BB108_2309; $L__BB108_1857: ld.local.v2.f32 {%f12254, %f12255}, [%rd7+-48]; ld.local.v2.f32 {%f12256, %f12257}, [%rd587+20]; mul.ftz.f32 %f12258, %f12256, %f12257; setp.lt.ftz.f32 %p1774, %f12258, 0f33D6BF95; ld.local.f32 %f16551, [%rd587+8]; ld.local.f32 %f3197, [%rd7+-52]; ld.local.f32 %f16552, [%rd587+12]; mul.ftz.f32 %f12259, %f12254, %f16552; fma.rn.ftz.f32 %f12260, %f3197, %f16551, %f12259; ld.local.f32 %f16553, [%rd587+16]; fma.rn.ftz.f32 %f16554, %f12255, %f16553, %f12260; @%p1774 bra $L__BB108_2085; bra.uni $L__BB108_1858; $L__BB108_2085: ld.local.f32 %f3376, [%rd587+28]; setp.gt.ftz.f32 %p1832, %f16554, 0f00000000; @%p1832 bra $L__BB108_2087; bra.uni $L__BB108_2086; $L__BB108_2087: rcp.approx.ftz.f32 %f16555, %f3376; bra.uni $L__BB108_2088; $L__BB108_2246: ld.local.v2.f32 {%f14365, %f14366}, [%rd7+-48]; ld.local.v2.f32 {%f14367, %f14368}, [%rd587+20]; ld.local.f32 %f3898, [%rd587+8]; ld.local.f32 %f3899, [%rd7+-52]; ld.local.f32 %f3900, [%rd587+12]; mul.ftz.f32 %f14369, %f14365, %f3900; fma.rn.ftz.f32 %f14370, %f3899, %f3898, %f14369; ld.local.f32 %f3901, [%rd587+16]; fma.rn.ftz.f32 %f3902, %f14366, %f3901, %f14370; setp.gt.ftz.f32 %p1989, %f3902, 0f00000000; @%p1989 bra $L__BB108_2248; bra.uni $L__BB108_2247; $L__BB108_2248: setp.eq.ftz.f32 %p1991, %f14367, %f14368; @%p1991 bra $L__BB108_2250; bra.uni $L__BB108_2249; $L__BB108_2250: setp.neu.ftz.f32 %p1992, %f3898, %f3900; setp.neu.ftz.f32 %p1993, %f3898, %f3901; or.pred %p1994, %p1992, %p1993; neg.ftz.f32 %f14398, %f3901; selp.f32 %f14399, %f14398, %f3901, %p1994; add.ftz.f32 %f14400, %f3898, %f14399; neg.ftz.f32 %f14401, %f3900; selp.f32 %f14402, %f3900, %f14401, %p1994; sub.ftz.f32 %f14403, %f14402, %f3898; sub.ftz.f32 %f14404, %f3901, %f3900; mul.ftz.f32 %f14405, %f14400, %f14400; fma.rn.ftz.f32 %f14406, %f14404, %f14404, %f14405; fma.rn.ftz.f32 %f14407, %f14403, %f14403, %f14406; sqrt.approx.ftz.f32 %f14408, %f14407; div.approx.ftz.f32 %f16696, %f14404, %f14408; div.approx.ftz.f32 %f16697, %f14400, %f14408; div.approx.ftz.f32 %f16698, %f14403, %f14408; mul.ftz.f32 %f14409, %f3900, %f16698; mul.ftz.f32 %f14410, %f3901, %f16697; sub.ftz.f32 %f16693, %f14409, %f14410; mul.ftz.f32 %f14411, %f3901, %f16696; mul.ftz.f32 %f14412, %f3898, %f16698; sub.ftz.f32 %f16694, %f14411, %f14412; mul.ftz.f32 %f14413, %f3898, %f16697; mul.ftz.f32 %f14414, %f3900, %f16696; sub.ftz.f32 %f16695, %f14413, %f14414; bra.uni $L__BB108_2251; $L__BB108_1812: setp.eq.s32 %p1627, %r873, 2; @%p1627 bra $L__BB108_2307; setp.eq.s32 %p1628, %r873, 3; @%p1628 bra $L__BB108_2303; setp.eq.s32 %p1629, %r873, 5; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1629 bra $L__BB108_1815; bra.uni $L__BB108_2309; $L__BB108_1815: ld.local.f32 %f4081, [%rd7+-52]; ld.local.v2.f32 {%f15001, %f15002}, [%rd7+-48]; mul.ftz.f32 %f15003, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f15004, %f15003; sqrt.approx.ftz.f32 %f15005, %f2775; mul.ftz.f32 %f15006, %f15005, %f15004; sin.approx.ftz.f32 %f15007, %f15003; mul.ftz.f32 %f15008, %f15005, %f15007; mul.ftz.f32 %f15009, %f15006, %f15006; mov.f32 %f15010, 0f3F800000; sub.ftz.f32 %f15011, %f15010, %f15009; mul.ftz.f32 %f15012, %f15008, %f15008; sub.ftz.f32 %f15013, %f15011, %f15012; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f15014, %f15013, %f16743; sqrt.approx.ftz.f32 %f4084, %f15014; ld.local.f32 %f4085, [%rd587+12]; ld.local.f32 %f4086, [%rd587+8]; setp.neu.ftz.f32 %p2044, %f4086, %f4085; ld.local.f32 %f4087, [%rd587+16]; setp.neu.ftz.f32 %p2045, %f4086, %f4087; or.pred %p2046, %p2044, %p2045; sub.ftz.f32 %f15015, %f4087, %f4085; neg.ftz.f32 %f15016, %f4085; neg.ftz.f32 %f15017, %f4087; selp.f32 %f15018, %f15017, %f4087, %p2046; add.ftz.f32 %f15019, %f4086, %f15018; selp.f32 %f15020, %f4085, %f15016, %p2046; sub.ftz.f32 %f15021, %f15020, %f4086; mul.ftz.f32 %f15022, %f15019, %f15019; fma.rn.ftz.f32 %f15023, %f15015, %f15015, %f15022; fma.rn.ftz.f32 %f15024, %f15021, %f15021, %f15023; sqrt.approx.ftz.f32 %f15025, %f15024; div.approx.ftz.f32 %f15026, %f15015, %f15025; div.approx.ftz.f32 %f15027, %f15019, %f15025; div.approx.ftz.f32 %f15028, %f15021, %f15025; mul.ftz.f32 %f15029, %f4085, %f15028; mul.ftz.f32 %f15030, %f4087, %f15027; sub.ftz.f32 %f15031, %f15029, %f15030; mul.ftz.f32 %f15032, %f4087, %f15026; mul.ftz.f32 %f15033, %f4086, %f15028; sub.ftz.f32 %f15034, %f15032, %f15033; mul.ftz.f32 %f15035, %f4086, %f15027; mul.ftz.f32 %f15036, %f4085, %f15026; sub.ftz.f32 %f15037, %f15035, %f15036; mul.ftz.f32 %f15038, %f15008, %f15031; mul.ftz.f32 %f15039, %f15008, %f15034; mul.ftz.f32 %f15040, %f15008, %f15037; fma.rn.ftz.f32 %f15041, %f15006, %f15026, %f15038; fma.rn.ftz.f32 %f15042, %f15006, %f15027, %f15039; fma.rn.ftz.f32 %f15043, %f15006, %f15028, %f15040; fma.rn.ftz.f32 %f16740, %f4086, %f4084, %f15041; fma.rn.ftz.f32 %f16741, %f4085, %f4084, %f15042; fma.rn.ftz.f32 %f16742, %f4087, %f4084, %f15043; mul.ftz.f32 %f15044, %f2807, %f16741; fma.rn.ftz.f32 %f15045, %f2806, %f16740, %f15044; fma.rn.ftz.f32 %f15046, %f2808, %f16742, %f15045; setp.leu.ftz.f32 %p2047, %f15046, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2047 bra $L__BB108_2309; mul.ftz.f32 %f15050, %f4085, %f16741; fma.rn.ftz.f32 %f15051, %f4086, %f16740, %f15050; fma.rn.ftz.f32 %f4091, %f4087, %f16742, %f15051; setp.le.ftz.f32 %p2048, %f4091, 0f00000000; mov.f32 %f16744, 0f00000000; @%p2048 bra $L__BB108_2302; mul.ftz.f32 %f15052, %f15001, %f4085; fma.rn.ftz.f32 %f15053, %f4081, %f4086, %f15052; fma.rn.ftz.f32 %f15054, %f15002, %f4087, %f15053; sub.ftz.f32 %f15056, %f15010, %f15054; mov.f32 %f15057, 0f00000000; max.ftz.f32 %f15058, %f15056, %f15057; min.ftz.f32 %f15059, %f15058, %f15010; mul.ftz.f32 %f15060, %f15059, %f15059; mul.ftz.f32 %f15061, %f15060, %f15060; mul.ftz.f32 %f4092, %f15059, %f15061; sub.ftz.f32 %f15062, %f15010, %f4091; max.ftz.f32 %f15063, %f15062, %f15057; min.ftz.f32 %f15064, %f15063, %f15010; mul.ftz.f32 %f15065, %f15064, %f15064; mul.ftz.f32 %f15066, %f15065, %f15065; mul.ftz.f32 %f4093, %f15064, %f15066; ld.local.u32 %r903, [%rd587+24]; and.b32 %r2895, %r903, 3; setp.eq.s32 %p2049, %r2895, 0; fma.rn.ftz.f32 %f4094, %f4093, 0fBF000000, 0f3F800000; @%p2049 bra $L__BB108_2298; fma.rn.ftz.f32 %f15067, %f4092, 0fBF000000, 0f3F800000; fma.rn.ftz.f32 %f16737, %f15067, %f4094, 0f00000000; bra.uni $L__BB108_2299; $L__BB108_1848: setp.eq.s32 %p1611, %r873, 20; @%p1611 bra $L__BB108_2097; setp.eq.s32 %p1612, %r873, 21; @%p1612 bra $L__BB108_2091; setp.eq.s32 %p1613, %r873, 22; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1613 bra $L__BB108_1851; bra.uni $L__BB108_2309; $L__BB108_1870: setp.eq.s32 %p1604, %r873, 29; @%p1604 bra $L__BB108_1890; setp.eq.s32 %p1605, %r873, 30; @%p1605 bra $L__BB108_1874; setp.ne.s32 %p1606, %r873, 31; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1606 bra $L__BB108_2309; ld.local.v2.f32 {%f11309, %f11310}, [%rd7+-48]; ld.local.f32 %f11313, [%rd7+-52]; neg.ftz.f32 %f16740, %f11313; neg.ftz.f32 %f16741, %f11309; neg.ftz.f32 %f16742, %f11310; mov.u32 %r3285, 33; mov.f32 %f16743, 0f3F800000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1828: setp.eq.s32 %p1623, %r873, 8; @%p1623 bra $L__BB108_2292; setp.eq.s32 %p1624, %r873, 9; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1624 bra $L__BB108_1830; bra.uni $L__BB108_2309; $L__BB108_1830: ld.local.v2.f32 {%f14776, %f14777}, [%rd7+-48]; ld.local.f32 %f4034, [%rd587+8]; ld.local.f32 %f4035, [%rd7+-52]; ld.local.f32 %f4036, [%rd587+12]; mul.ftz.f32 %f14778, %f14776, %f4036; fma.rn.ftz.f32 %f14779, %f4035, %f4034, %f14778; ld.local.f32 %f4037, [%rd587+16]; fma.rn.ftz.f32 %f4038, %f14777, %f4037, %f14779; setp.leu.ftz.f32 %p2022, %f4038, 0f00000000; mov.u32 %r3285, 18; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2022 bra $L__BB108_2309; add.ftz.f32 %f14784, %f4038, %f4038; mul.ftz.f32 %f14785, %f4034, %f14784; mul.ftz.f32 %f14786, %f4036, %f14784; mul.ftz.f32 %f14787, %f4037, %f14784; sub.ftz.f32 %f16740, %f14785, %f4035; sub.ftz.f32 %f16741, %f14786, %f14776; sub.ftz.f32 %f16742, %f14787, %f14777; mul.ftz.f32 %f14788, %f2807, %f16741; fma.rn.ftz.f32 %f14789, %f2806, %f16740, %f14788; fma.rn.ftz.f32 %f14790, %f2808, %f16742, %f14789; setp.leu.ftz.f32 %p2023, %f14790, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2023 bra $L__BB108_2309; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_2247: setp.eq.s32 %p1990, %r873, 23; selp.b32 %r3285, 9, 10, %p1990; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1835: setp.eq.s32 %p1620, %r873, 13; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1620 bra $L__BB108_1851; bra.uni $L__BB108_2309; $L__BB108_1851: ld.local.v2.f32 {%f13814, %f13815}, [%rd7+-48]; ld.local.v2.f32 {%f13816, %f13817}, [%rd587+20]; ld.local.f32 %f3736, [%rd587+8]; ld.local.f32 %f3737, [%rd7+-52]; ld.local.f32 %f3738, [%rd587+12]; mul.ftz.f32 %f13818, %f13814, %f3738; fma.rn.ftz.f32 %f13819, %f3737, %f3736, %f13818; ld.local.f32 %f3739, [%rd587+16]; fma.rn.ftz.f32 %f3740, %f13815, %f3739, %f13819; setp.gt.ftz.f32 %p1955, %f3740, 0f00000000; @%p1955 bra $L__BB108_2198; bra.uni $L__BB108_1852; $L__BB108_2198: setp.eq.ftz.f32 %p1957, %f13816, %f13817; @%p1957 bra $L__BB108_2200; bra.uni $L__BB108_2199; $L__BB108_2200: setp.neu.ftz.f32 %p1958, %f3736, %f3738; setp.neu.ftz.f32 %p1959, %f3736, %f3739; or.pred %p1960, %p1958, %p1959; neg.ftz.f32 %f13847, %f3739; selp.f32 %f13848, %f13847, %f3739, %p1960; add.ftz.f32 %f13849, %f3736, %f13848; neg.ftz.f32 %f13850, %f3738; selp.f32 %f13851, %f3738, %f13850, %p1960; sub.ftz.f32 %f13852, %f13851, %f3736; sub.ftz.f32 %f13853, %f3739, %f3738; mul.ftz.f32 %f13854, %f13849, %f13849; fma.rn.ftz.f32 %f13855, %f13853, %f13853, %f13854; fma.rn.ftz.f32 %f13856, %f13852, %f13852, %f13855; sqrt.approx.ftz.f32 %f13857, %f13856; div.approx.ftz.f32 %f16629, %f13853, %f13857; div.approx.ftz.f32 %f16630, %f13849, %f13857; div.approx.ftz.f32 %f16631, %f13852, %f13857; mul.ftz.f32 %f13858, %f3738, %f16631; mul.ftz.f32 %f13859, %f3739, %f16630; sub.ftz.f32 %f16626, %f13858, %f13859; mul.ftz.f32 %f13860, %f3739, %f16629; mul.ftz.f32 %f13861, %f3736, %f16631; sub.ftz.f32 %f16627, %f13860, %f13861; mul.ftz.f32 %f13862, %f3736, %f16630; mul.ftz.f32 %f13863, %f3738, %f16629; sub.ftz.f32 %f16628, %f13862, %f13863; bra.uni $L__BB108_2201; $L__BB108_2294: mul.ftz.f32 %f14921, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f14922, %f14921; sqrt.approx.ftz.f32 %f14923, %f2775; mul.ftz.f32 %f14924, %f14923, %f14922; sin.approx.ftz.f32 %f14925, %f14921; mul.ftz.f32 %f14926, %f14923, %f14925; mul.ftz.f32 %f14927, %f14924, %f14924; mov.f32 %f14928, 0f3F800000; sub.ftz.f32 %f14929, %f14928, %f14927; mul.ftz.f32 %f14930, %f14926, %f14926; sub.ftz.f32 %f14931, %f14929, %f14930; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f14932, %f14931, %f16743; sqrt.approx.ftz.f32 %f4059, %f14932; ld.local.f32 %f4060, [%rd587+12]; ld.local.f32 %f4061, [%rd587+8]; setp.neu.ftz.f32 %p2037, %f4061, %f4060; ld.local.f32 %f4062, [%rd587+16]; setp.neu.ftz.f32 %p2038, %f4061, %f4062; or.pred %p2039, %p2037, %p2038; sub.ftz.f32 %f14933, %f4062, %f4060; neg.ftz.f32 %f14934, %f4060; neg.ftz.f32 %f14935, %f4062; selp.f32 %f14936, %f14935, %f4062, %p2039; add.ftz.f32 %f14937, %f4061, %f14936; selp.f32 %f14938, %f4060, %f14934, %p2039; sub.ftz.f32 %f14939, %f14938, %f4061; mul.ftz.f32 %f14940, %f14937, %f14937; fma.rn.ftz.f32 %f14941, %f14933, %f14933, %f14940; fma.rn.ftz.f32 %f14942, %f14939, %f14939, %f14941; sqrt.approx.ftz.f32 %f14943, %f14942; div.approx.ftz.f32 %f14944, %f14933, %f14943; div.approx.ftz.f32 %f14945, %f14937, %f14943; div.approx.ftz.f32 %f14946, %f14939, %f14943; mul.ftz.f32 %f14947, %f4060, %f14946; mul.ftz.f32 %f14948, %f4062, %f14945; sub.ftz.f32 %f14949, %f14947, %f14948; mul.ftz.f32 %f14950, %f4062, %f14944; mul.ftz.f32 %f14951, %f4061, %f14946; sub.ftz.f32 %f14952, %f14950, %f14951; mul.ftz.f32 %f14953, %f4061, %f14945; mul.ftz.f32 %f14954, %f4060, %f14944; sub.ftz.f32 %f14955, %f14953, %f14954; mul.ftz.f32 %f14956, %f14926, %f14949; mul.ftz.f32 %f14957, %f14926, %f14952; mul.ftz.f32 %f14958, %f14926, %f14955; fma.rn.ftz.f32 %f14959, %f14924, %f14944, %f14956; fma.rn.ftz.f32 %f14960, %f14924, %f14945, %f14957; fma.rn.ftz.f32 %f14961, %f14924, %f14946, %f14958; fma.rn.ftz.f32 %f16740, %f4061, %f4059, %f14959; fma.rn.ftz.f32 %f16741, %f4060, %f4059, %f14960; fma.rn.ftz.f32 %f16742, %f4062, %f4059, %f14961; mul.ftz.f32 %f14962, %f2807, %f16741; fma.rn.ftz.f32 %f14963, %f2806, %f16740, %f14962; fma.rn.ftz.f32 %f14964, %f2808, %f16742, %f14963; setp.leu.ftz.f32 %p2040, %f14964, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2040 bra $L__BB108_2309; ld.local.f32 %f14969, [%rd7+-52]; ld.local.v2.f32 {%f14970, %f14971}, [%rd7+-48]; add.ftz.f32 %f14974, %f14969, %f16740; add.ftz.f32 %f14975, %f14970, %f16741; mul.ftz.f32 %f14976, %f14975, %f14975; fma.rn.ftz.f32 %f14977, %f14974, %f14974, %f14976; add.ftz.f32 %f14978, %f14971, %f16742; fma.rn.ftz.f32 %f14979, %f14978, %f14978, %f14977; sqrt.approx.ftz.f32 %f14980, %f14979; div.approx.ftz.f32 %f4066, %f14974, %f14980; div.approx.ftz.f32 %f4067, %f14975, %f14980; div.approx.ftz.f32 %f4068, %f14978, %f14980; mul.ftz.f32 %f14981, %f4060, %f16741; fma.rn.ftz.f32 %f14982, %f4061, %f16740, %f14981; fma.rn.ftz.f32 %f4069, %f4062, %f16742, %f14982; mul.ftz.f32 %f14983, %f14970, %f4060; fma.rn.ftz.f32 %f14984, %f14969, %f4061, %f14983; fma.rn.ftz.f32 %f14985, %f14971, %f4062, %f14984; setp.lt.ftz.f32 %p2041, %f4069, 0f00000000; setp.lt.ftz.f32 %p2042, %f14985, 0f00000000; or.pred %p2043, %p2042, %p2041; mov.f32 %f16743, 0f00000000; mov.f32 %f16744, %f16743; @%p2043 bra $L__BB108_2297; mul.ftz.f32 %f14986, %f16741, %f4067; fma.rn.ftz.f32 %f14987, %f16740, %f4066, %f14986; fma.rn.ftz.f32 %f14988, %f16742, %f4068, %f14987; sub.ftz.f32 %f14990, %f14928, %f14988; mov.f32 %f14991, 0f00000000; max.ftz.f32 %f14992, %f14990, %f14991; min.ftz.f32 %f14993, %f14992, %f14928; mul.ftz.f32 %f14994, %f14993, %f14993; mul.ftz.f32 %f14995, %f14994, %f14994; mul.ftz.f32 %f14996, %f14993, %f14995; mul.ftz.f32 %f16744, %f4069, %f14996; mul.ftz.f32 %f16743, %f4059, 0f3EA2F983; $L__BB108_2297: mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_2113: ld.local.v2.f32 {%f13139, %f13140}, [%rd7+-48]; ld.local.f32 %f3484, [%rd587+8]; ld.local.f32 %f3485, [%rd7+-52]; ld.local.f32 %f3486, [%rd587+12]; mul.ftz.f32 %f13141, %f13139, %f3486; fma.rn.ftz.f32 %f13142, %f3485, %f3484, %f13141; ld.local.f32 %f3487, [%rd587+16]; fma.rn.ftz.f32 %f3488, %f13140, %f3487, %f13142; setp.leu.ftz.f32 %p1880, %f3488, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1880 bra $L__BB108_2309; ld.local.v2.f32 {%f13143, %f13144}, [%rd587+20]; mul.ftz.f32 %f13145, %f13143, %f13143; mov.f32 %f13146, 0f40000000; div.approx.ftz.f32 %f13147, %f13146, %f13145; add.ftz.f32 %f3491, %f13147, 0fC0000000; mul.ftz.f32 %f13148, %f13144, %f13144; div.approx.ftz.f32 %f13149, %f13146, %f13148; add.ftz.f32 %f3492, %f13149, 0fC0000000; setp.eq.ftz.f32 %p1881, %f3491, %f3492; @%p1881 bra $L__BB108_2116; bra.uni $L__BB108_2115; $L__BB108_2116: setp.neu.ftz.f32 %p1882, %f3484, %f3486; setp.neu.ftz.f32 %p1883, %f3484, %f3487; or.pred %p1884, %p1882, %p1883; neg.ftz.f32 %f13172, %f3487; selp.f32 %f13173, %f13172, %f3487, %p1884; add.ftz.f32 %f13174, %f3484, %f13173; neg.ftz.f32 %f13175, %f3486; selp.f32 %f13176, %f3486, %f13175, %p1884; sub.ftz.f32 %f13177, %f13176, %f3484; sub.ftz.f32 %f13178, %f3487, %f3486; mul.ftz.f32 %f13179, %f13174, %f13174; fma.rn.ftz.f32 %f13180, %f13178, %f13178, %f13179; fma.rn.ftz.f32 %f13181, %f13177, %f13177, %f13180; sqrt.approx.ftz.f32 %f13182, %f13181; div.approx.ftz.f32 %f16575, %f13178, %f13182; div.approx.ftz.f32 %f16576, %f13174, %f13182; div.approx.ftz.f32 %f16577, %f13177, %f13182; mul.ftz.f32 %f13183, %f3486, %f16577; mul.ftz.f32 %f13184, %f3487, %f16576; sub.ftz.f32 %f16572, %f13183, %f13184; mul.ftz.f32 %f13185, %f3487, %f16575; mul.ftz.f32 %f13186, %f3484, %f16577; sub.ftz.f32 %f16573, %f13185, %f13186; mul.ftz.f32 %f13187, %f3484, %f16576; mul.ftz.f32 %f13188, %f3486, %f16575; sub.ftz.f32 %f16574, %f13187, %f13188; bra.uni $L__BB108_2117; $L__BB108_2108: mul.ftz.f32 %f13037, %f2799, %f2799; mov.f32 %f13038, 0f3F800000; sub.ftz.f32 %f13039, %f13038, %f13037; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f13040, %f16743, %f13039; sqrt.approx.ftz.f32 %f13041, %f13040; mul.ftz.f32 %f13042, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f13043, %f13042; mul.ftz.f32 %f13044, %f13041, %f13043; sin.approx.ftz.f32 %f13045, %f13042; mul.ftz.f32 %f13046, %f13041, %f13045; ld.local.f32 %f3470, [%rd587+12]; ld.local.f32 %f3471, [%rd587+8]; setp.neu.ftz.f32 %p1871, %f3471, %f3470; ld.local.f32 %f3472, [%rd587+16]; setp.neu.ftz.f32 %p1872, %f3471, %f3472; or.pred %p1873, %p1871, %p1872; sub.ftz.f32 %f13047, %f3472, %f3470; neg.ftz.f32 %f13048, %f3470; neg.ftz.f32 %f13049, %f3472; selp.f32 %f13050, %f13049, %f3472, %p1873; add.ftz.f32 %f13051, %f3471, %f13050; selp.f32 %f13052, %f3470, %f13048, %p1873; sub.ftz.f32 %f13053, %f13052, %f3471; mul.ftz.f32 %f13054, %f13051, %f13051; fma.rn.ftz.f32 %f13055, %f13047, %f13047, %f13054; fma.rn.ftz.f32 %f13056, %f13053, %f13053, %f13055; sqrt.approx.ftz.f32 %f13057, %f13056; div.approx.ftz.f32 %f13058, %f13047, %f13057; div.approx.ftz.f32 %f13059, %f13051, %f13057; div.approx.ftz.f32 %f13060, %f13053, %f13057; mul.ftz.f32 %f13061, %f3470, %f13060; mul.ftz.f32 %f13062, %f3472, %f13059; sub.ftz.f32 %f13063, %f13061, %f13062; mul.ftz.f32 %f13064, %f3472, %f13058; mul.ftz.f32 %f13065, %f3471, %f13060; sub.ftz.f32 %f13066, %f13064, %f13065; mul.ftz.f32 %f13067, %f3471, %f13059; mul.ftz.f32 %f13068, %f3470, %f13058; sub.ftz.f32 %f13069, %f13067, %f13068; mul.ftz.f32 %f13070, %f13046, %f13063; mul.ftz.f32 %f13071, %f13046, %f13066; mul.ftz.f32 %f13072, %f13046, %f13069; fma.rn.ftz.f32 %f13073, %f13044, %f13058, %f13070; fma.rn.ftz.f32 %f13074, %f13044, %f13059, %f13071; fma.rn.ftz.f32 %f13075, %f13044, %f13060, %f13072; fma.rn.ftz.f32 %f16740, %f2799, %f3471, %f13073; fma.rn.ftz.f32 %f16741, %f2799, %f3470, %f13074; fma.rn.ftz.f32 %f16742, %f2799, %f3472, %f13075; mul.ftz.f32 %f13076, %f2807, %f16741; fma.rn.ftz.f32 %f13077, %f2806, %f16740, %f13076; fma.rn.ftz.f32 %f13078, %f2808, %f16742, %f13077; setp.leu.ftz.f32 %p1874, %f13078, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1874 bra $L__BB108_2309; ld.local.f32 %f13079, [%rd7+-52]; ld.local.v2.f32 {%f13080, %f13081}, [%rd7+-48]; ld.local.f32 %f3476, [%rd587+24]; add.ftz.f32 %f13084, %f13079, %f16740; add.ftz.f32 %f13085, %f13080, %f16741; mul.ftz.f32 %f13086, %f13085, %f13085; fma.rn.ftz.f32 %f13087, %f13084, %f13084, %f13086; add.ftz.f32 %f13088, %f13081, %f16742; fma.rn.ftz.f32 %f13089, %f13088, %f13088, %f13087; sqrt.approx.ftz.f32 %f13090, %f13089; div.approx.ftz.f32 %f13091, %f13084, %f13090; div.approx.ftz.f32 %f13092, %f13085, %f13090; div.approx.ftz.f32 %f13093, %f13088, %f13090; mul.ftz.f32 %f13094, %f3470, %f16741; fma.rn.ftz.f32 %f13095, %f3471, %f16740, %f13094; fma.rn.ftz.f32 %f3477, %f3472, %f16742, %f13095; mul.ftz.f32 %f13096, %f13080, %f3470; fma.rn.ftz.f32 %f13097, %f13079, %f3471, %f13096; fma.rn.ftz.f32 %f3478, %f13081, %f3472, %f13097; mul.ftz.f32 %f13098, %f3470, %f13092; fma.rn.ftz.f32 %f13099, %f3471, %f13091, %f13098; fma.rn.ftz.f32 %f3479, %f3472, %f13093, %f13099; mul.ftz.f32 %f13100, %f13080, %f13092; fma.rn.ftz.f32 %f13101, %f13079, %f13091, %f13100; fma.rn.ftz.f32 %f13102, %f13081, %f13093, %f13101; abs.ftz.f32 %f3480, %f13102; abs.ftz.f32 %f13103, %f3478; setp.leu.ftz.f32 %p1876, %f13103, 0f3727C5AC; mov.pred %p2188, 0; @%p1876 bra $L__BB108_2111; abs.ftz.f32 %f13104, %f3479; setp.lt.ftz.f32 %p2188, %f13104, 0f3F7FFF58; $L__BB108_2111: setp.leu.ftz.f32 %p1877, %f3480, 0f3727C5AC; not.pred %p1878, %p2188; or.pred %p1879, %p1877, %p1878; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1879 bra $L__BB108_2309; div.approx.ftz.f32 %f13110, %f3479, %f3480; mov.f32 %f13111, 0f3727C5AC; max.ftz.f32 %f13112, %f13110, %f13111; mul.ftz.f32 %f13113, %f3478, %f13112; abs.ftz.f32 %f13114, %f13113; add.ftz.f32 %f13115, %f13114, %f13114; mul.ftz.f32 %f13116, %f3477, %f13112; abs.ftz.f32 %f13117, %f13116; add.ftz.f32 %f13118, %f13117, %f13117; mul.ftz.f32 %f13119, %f3479, %f3479; sub.ftz.f32 %f13121, %f13038, %f13119; mul.ftz.f32 %f13122, %f13121, %f13121; div.approx.ftz.f32 %f13123, %f13119, %f13121; mul.ftz.f32 %f13124, %f3476, %f13123; mul.ftz.f32 %f13125, %f13124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13126, %f13125; mul.ftz.f32 %f13127, %f3476, %f13126; mul.ftz.f32 %f13128, %f13127, 0f3EA2F983; div.approx.ftz.f32 %f13129, %f13128, %f13122; min.ftz.f32 %f13130, %f13115, %f13118; min.ftz.f32 %f13131, %f13038, %f13130; mul.ftz.f32 %f13132, %f13129, %f13131; mul.ftz.f32 %f13133, %f13132, 0f3E800000; div.approx.ftz.f32 %f16744, %f13133, %f3478; mov.u32 %r3285, 6; mov.f32 %f16743, 0f3E22F983; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_1858: setp.neu.ftz.f32 %p1775, %f16551, %f16552; setp.neu.ftz.f32 %p1776, %f16551, %f16553; or.pred %p1777, %p1775, %p1776; neg.ftz.f32 %f12262, %f16553; selp.f32 %f12263, %f12262, %f16553, %p1777; add.ftz.f32 %f12264, %f16551, %f12263; neg.ftz.f32 %f12265, %f16552; selp.f32 %f12266, %f16552, %f12265, %p1777; sub.ftz.f32 %f12267, %f12266, %f16551; sub.ftz.f32 %f12268, %f16553, %f16552; mul.ftz.f32 %f12269, %f12264, %f12264; fma.rn.ftz.f32 %f12270, %f12268, %f12268, %f12269; fma.rn.ftz.f32 %f12271, %f12267, %f12267, %f12270; sqrt.approx.ftz.f32 %f12272, %f12271; div.approx.ftz.f32 %f3201, %f12268, %f12272; div.approx.ftz.f32 %f3202, %f12264, %f12272; div.approx.ftz.f32 %f3203, %f12267, %f12272; mul.ftz.f32 %f12273, %f16552, %f3203; mul.ftz.f32 %f12274, %f16553, %f3202; sub.ftz.f32 %f3204, %f12273, %f12274; mul.ftz.f32 %f12275, %f16553, %f3201; mul.ftz.f32 %f12276, %f16551, %f3203; sub.ftz.f32 %f3205, %f12275, %f12276; mul.ftz.f32 %f12277, %f16551, %f3202; mul.ftz.f32 %f12278, %f16552, %f3201; sub.ftz.f32 %f3206, %f12277, %f12278; mul.ftz.f32 %f12279, %f12254, %f3202; fma.rn.ftz.f32 %f12280, %f3197, %f3201, %f12279; fma.rn.ftz.f32 %f3207, %f12255, %f3203, %f12280; mul.ftz.f32 %f12281, %f12254, %f3205; fma.rn.ftz.f32 %f12282, %f3197, %f3204, %f12281; fma.rn.ftz.f32 %f3208, %f12255, %f3206, %f12282; ld.local.u64 %rd2685, [%rd587+36]; ld.f32 %f3209, [%rd2685]; ld.f32 %f3210, [%rd2685+4]; ld.f32 %f3211, [%rd2685+8]; ld.local.f32 %f3212, [%rd587+28]; ld.f32 %f3213, [%rd2685+12]; ld.f32 %f3214, [%rd2685+16]; ld.f32 %f3215, [%rd2685+20]; neg.ftz.f32 %f3216, %f3207; neg.ftz.f32 %f3217, %f3208; neg.ftz.f32 %f3218, %f16554; setp.lt.ftz.f32 %p1778, %f16554, 0fBF7FF972; mov.f32 %f16537, 0f00000000; @%p1778 bra $L__BB108_1861; setp.gt.ftz.f32 %p1779, %f16554, 0f3F7FF972; mov.f32 %f16537, 0fBF7FF972; @%p1779 bra $L__BB108_1861; mul.ftz.f32 %f12284, %f16554, %f16554; mov.f32 %f12285, 0f33D6BF95; max.ftz.f32 %f12286, %f12284, %f12285; rcp.approx.ftz.f32 %f12287, %f12286; mul.ftz.f32 %f12288, %f12256, %f3216; mul.ftz.f32 %f12289, %f12257, %f3217; mul.ftz.f32 %f12290, %f12289, %f12289; fma.rn.ftz.f32 %f12291, %f12288, %f12288, %f12290; fma.rn.ftz.f32 %f12292, %f12291, %f12287, 0f3F800000; sqrt.approx.ftz.f32 %f12293, %f12292; neg.ftz.f32 %f12294, %f12293; setp.ge.ftz.f32 %p1780, %f16554, 0f80000000; selp.f32 %f12295, %f12294, %f12293, %p1780; add.ftz.f32 %f12296, %f12295, 0fBF800000; mul.ftz.f32 %f16537, %f12296, 0f3F000000; $L__BB108_1861: fma.rn.ftz.f32 %f3221, %f3212, %f3212, 0fBF800000; mov.f32 %f16536, 0f3F800000; abs.ftz.f32 %f3222, %f16536; fma.rn.ftz.f32 %f3223, %f3222, %f3222, %f3221; setp.leu.ftz.f32 %p1781, %f3223, 0f00000000; mov.f32 %f16511, %f16536; @%p1781 bra $L__BB108_1863; sqrt.approx.ftz.f32 %f12298, %f3223; sub.ftz.f32 %f12299, %f12298, %f3222; add.ftz.f32 %f12300, %f3222, %f12298; div.approx.ftz.f32 %f12301, %f12299, %f12300; fma.rn.ftz.f32 %f12302, %f3222, %f12300, 0fBF800000; fma.rn.ftz.f32 %f12303, %f3222, %f12299, 0f3F800000; div.approx.ftz.f32 %f12304, %f12302, %f12303; mul.ftz.f32 %f12305, %f12301, 0f3F000000; mul.ftz.f32 %f12306, %f12301, %f12305; fma.rn.ftz.f32 %f12307, %f12304, %f12304, 0f3F800000; mul.ftz.f32 %f16511, %f12306, %f12307; $L__BB108_1863: setp.ne.s32 %p8, %r873, 27; rcp.approx.ftz.f32 %f3226, %f3212; sub.ftz.f32 %f3227, %f16536, %f16511; rcp.approx.ftz.f32 %f3228, %f3227; ld.local.u32 %r3279, [%rd7+168]; mov.f32 %f16538, 0f00000000; mov.u16 %rs155, 1; mov.u32 %r3280, 0; mov.pred %p1782, -1; mov.pred %p1809, 0; mov.f32 %f16535, %f16536; mov.f32 %f16534, %f16536; mov.f32 %f3232, %f3218; mov.f32 %f3233, %f3217; mov.f32 %f3234, %f3216; mov.f32 %f16521, %f16536; $L__BB108_1864: mov.f32 %f3236, %f16521; mad.lo.s32 %r885, %r3279, 1103515245, 12345; cvt.rn.f32.u32 %f12313, %r885; mul.ftz.f32 %f3238, %f12313, 0f2F800000; setp.gt.ftz.f32 %p1783, %f3232, 0f3F7FF972; mov.pred %p2186, %p1782; @%p1783 bra $L__BB108_2042; setp.lt.ftz.f32 %p1784, %f3232, 0fBF7FF972; @%p1784 bra $L__BB108_2041; bra.uni $L__BB108_1866; $L__BB108_2041: mul.ftz.f32 %f16521, %f3236, %f3238; mov.pred %p2186, %p1809; bra.uni $L__BB108_2042; $L__BB108_1866: abs.ftz.f32 %f12314, %f3232; setp.ltu.ftz.f32 %p1786, %f12314, 0f38D1B717; mov.f32 %f16521, %f3236; mov.pred %p2186, %p1809; @%p1786 bra $L__BB108_2042; mov.f32 %f12315, 0f3F800000; sub.ftz.f32 %f12316, %f12315, %f16538; setp.lt.ftz.f32 %p1788, %f12316, %f3238; mov.f32 %f16521, %f3236; mov.pred %p2186, %p1782; @%p1788 bra $L__BB108_2042; setp.ge.ftz.f32 %p1790, %f16537, 0f00000000; mov.f32 %f16521, %f12315; mov.pred %p2186, %p1809; @%p1790 bra $L__BB108_2042; mov.f32 %f12318, 0f3F800000; sub.ftz.f32 %f12319, %f12318, %f3238; lg2.approx.ftz.f32 %f12320, %f12319; mov.f32 %f12321, 0fBF800000; div.approx.ftz.f32 %f12322, %f12321, %f16537; mul.ftz.f32 %f12323, %f12322, %f12320; ex2.approx.ftz.f32 %f12324, %f12323; mul.ftz.f32 %f16521, %f3236, %f12324; mov.pred %p2186, %p1809; $L__BB108_2042: @%p2186 bra $L__BB108_2074; mad.lo.s32 %r2776, %r885, 1103515245, 12345; cvt.rn.f32.u32 %f12325, %r2776; mul.ftz.f32 %f3242, %f12325, 0f2F800000; mad.lo.s32 %r886, %r2776, 1103515245, 12345; cvt.rn.f32.u32 %f12326, %r886; mul.ftz.f32 %f3243, %f12326, 0f2F800000; neg.ftz.f32 %f3244, %f3234; mul.ftz.f32 %f12327, %f12256, %f3244; neg.ftz.f32 %f3245, %f3233; mul.ftz.f32 %f12328, %f12257, %f3245; mul.ftz.f32 %f12329, %f12328, %f12328; fma.rn.ftz.f32 %f12330, %f12327, %f12327, %f12329; fma.rn.ftz.f32 %f12331, %f3232, %f3232, %f12330; sqrt.approx.ftz.f32 %f12332, %f12331; div.approx.ftz.f32 %f16524, %f12327, %f12332; div.approx.ftz.f32 %f16525, %f12328, %f12332; neg.ftz.f32 %f3248, %f3232; div.approx.ftz.f32 %f3249, %f3248, %f12332; setp.gt.ftz.f32 %p1793, %f3249, 0f3F7FF972; @%p1793 bra $L__BB108_2049; abs.ftz.f32 %f12333, %f3249; setp.lt.ftz.f32 %p1794, %f12333, 0f358637BD; @%p1794 bra $L__BB108_2049; bra.uni $L__BB108_2045; $L__BB108_2049: mov.f32 %f12377, 0f3F800000; sub.ftz.f32 %f12378, %f12377, %f3243; mov.f32 %f12379, 0f33D6BF95; max.ftz.f32 %f12380, %f12378, %f12379; div.approx.ftz.f32 %f12381, %f3243, %f12380; sqrt.approx.ftz.f32 %f12382, %f12381; mul.ftz.f32 %f12383, %f3242, 0f40C90FDB; cos.approx.ftz.f32 %f12384, %f12383; sin.approx.ftz.f32 %f12385, %f12383; mul.ftz.f32 %f16523, %f12382, %f12385; mul.ftz.f32 %f16522, %f12382, %f12384; $L__BB108_2050: mul.ftz.f32 %f12386, %f16525, %f16525; fma.rn.ftz.f32 %f12387, %f16524, %f16524, %f12386; add.ftz.f32 %f12388, %f12387, 0f00000000; sqrt.approx.ftz.f32 %f3267, %f12388; setp.neu.ftz.f32 %p1802, %f3267, 0f00000000; @%p1802 bra $L__BB108_2051; bra.uni $L__BB108_2052; $L__BB108_2051: rcp.approx.ftz.f32 %f12389, %f3267; mul.ftz.f32 %f16525, %f16525, %f12389; mul.ftz.f32 %f16524, %f16524, %f12389; $L__BB108_2052: mul.ftz.f32 %f12391, %f16522, %f16524; mul.ftz.f32 %f12392, %f16523, %f16525; sub.ftz.f32 %f12393, %f12391, %f12392; mul.ftz.f32 %f12394, %f12256, %f12393; mul.ftz.f32 %f12395, %f16523, %f16524; fma.rn.ftz.f32 %f12396, %f16522, %f16525, %f12395; mul.ftz.f32 %f12397, %f12257, %f12396; neg.ftz.f32 %f12398, %f12394; neg.ftz.f32 %f12399, %f12397; mul.ftz.f32 %f12400, %f12397, %f12397; fma.rn.ftz.f32 %f12401, %f12394, %f12394, %f12400; add.ftz.f32 %f12402, %f12401, 0f3F800000; mov.f32 %f16526, 0f3F800000; sqrt.approx.ftz.f32 %f12403, %f12402; div.approx.ftz.f32 %f3277, %f12398, %f12403; div.approx.ftz.f32 %f3278, %f12399, %f12403; div.approx.ftz.f32 %f3279, %f16526, %f12403; setp.ne.s32 %p1803, %r3280, 0; and.pred %p1804, %p1803, %p8; mul.ftz.f32 %f12404, %f3209, %f16534; selp.f32 %f3280, %f12404, %f16534, %p1804; mul.ftz.f32 %f12405, %f3210, %f16535; selp.f32 %f3281, %f12405, %f16535, %p1804; mul.ftz.f32 %f12406, %f3211, %f16536; selp.f32 %f3282, %f12406, %f16536, %p1804; mad.lo.s32 %r3279, %r886, 1103515245, 12345; and.b16 %rs114, %rs155, 255; setp.eq.s16 %p1805, %rs114, 0; selp.f32 %f3283, %f3226, %f3212, %p1805; mul.ftz.f32 %f12407, %f3278, %f3245; fma.rn.ftz.f32 %f12408, %f3277, %f3244, %f12407; fma.rn.ftz.f32 %f3284, %f3279, %f3248, %f12408; abs.ftz.f32 %f3285, %f3284; fma.rn.ftz.f32 %f12409, %f3283, %f3283, 0fBF800000; mul.ftz.f32 %f3286, %f3285, %f3285; add.ftz.f32 %f3287, %f12409, %f3286; setp.leu.ftz.f32 %p1806, %f3287, 0f00000000; @%p1806 bra $L__BB108_2054; sqrt.approx.ftz.f32 %f12410, %f3287; sub.ftz.f32 %f12411, %f12410, %f3285; add.ftz.f32 %f12412, %f3285, %f12410; div.approx.ftz.f32 %f12413, %f12411, %f12412; fma.rn.ftz.f32 %f12414, %f3285, %f12412, 0fBF800000; fma.rn.ftz.f32 %f12415, %f3285, %f12411, 0f3F800000; div.approx.ftz.f32 %f12416, %f12414, %f12415; mul.ftz.f32 %f12417, %f12413, 0f3F000000; mul.ftz.f32 %f12418, %f12413, %f12417; fma.rn.ftz.f32 %f12419, %f12416, %f12416, 0f3F800000; mul.ftz.f32 %f16526, %f12418, %f12419; $L__BB108_2054: cvt.rn.f32.u32 %f12420, %r3279; mul.ftz.f32 %f12421, %f12420, 0f2F800000; setp.gt.ftz.f32 %p1807, %f16526, %f12421; @%p1807 bra $L__BB108_2056; bra.uni $L__BB108_2055; $L__BB108_2056: add.ftz.f32 %f12447, %f3277, %f3277; add.ftz.f32 %f12448, %f3278, %f3278; add.ftz.f32 %f12449, %f3279, %f3279; fma.rn.ftz.f32 %f3232, %f12449, %f3284, %f3232; fma.rn.ftz.f32 %f3233, %f12448, %f3284, %f3233; fma.rn.ftz.f32 %f3234, %f12447, %f3284, %f3234; mov.pred %p2187, %p1809; bra.uni $L__BB108_2057; $L__BB108_2055: rcp.approx.ftz.f32 %f12422, %f3283; mov.f32 %f12423, 0f3F800000; mul.ftz.f32 %f12424, %f3284, %f3284; sub.ftz.f32 %f12425, %f12423, %f12424; mul.ftz.f32 %f12426, %f12425, %f12422; mul.ftz.f32 %f12427, %f12422, %f12426; sub.ftz.f32 %f12428, %f12423, %f12427; mov.f32 %f12429, 0f00000000; max.ftz.f32 %f12430, %f12428, %f12429; sqrt.approx.ftz.f32 %f12431, %f12430; mul.ftz.f32 %f12432, %f3284, %f12422; sub.ftz.f32 %f12433, %f12432, %f12431; mul.ftz.f32 %f12434, %f3277, %f12433; mul.ftz.f32 %f12435, %f3278, %f12433; mul.ftz.f32 %f12436, %f3279, %f12433; mul.ftz.f32 %f12437, %f12422, %f3244; mul.ftz.f32 %f12438, %f12422, %f3245; mul.ftz.f32 %f12439, %f12422, %f3248; sub.ftz.f32 %f12440, %f12434, %f12437; sub.ftz.f32 %f12441, %f12435, %f12438; sub.ftz.f32 %f12442, %f12436, %f12439; mul.ftz.f32 %f12443, %f12441, %f12441; fma.rn.ftz.f32 %f12444, %f12440, %f12440, %f12443; fma.rn.ftz.f32 %f12445, %f12442, %f12442, %f12444; sqrt.approx.ftz.f32 %f12446, %f12445; div.approx.ftz.f32 %f3232, %f12442, %f12446; div.approx.ftz.f32 %f3233, %f12441, %f12446; div.approx.ftz.f32 %f3234, %f12440, %f12446; mov.pred %p2187, %p1782; $L__BB108_2057: not.pred %p1810, %p2187; @%p1810 bra $L__BB108_2059; neg.ftz.f32 %f3234, %f3234; neg.ftz.f32 %f3233, %f3233; neg.ftz.f32 %f3232, %f3232; selp.u16 %rs155, 1, 0, %p1805; $L__BB108_2059: mov.f32 %f16534, %f3280; mov.f32 %f16535, %f3281; mov.f32 %f16536, %f3282; @%p8 bra $L__BB108_2066; @%p2187 bra $L__BB108_2065; bra.uni $L__BB108_2061; $L__BB108_2065: mul.ftz.f32 %f16534, %f3209, %f3280; mul.ftz.f32 %f16535, %f3210, %f3281; mul.ftz.f32 %f16536, %f3211, %f3282; bra.uni $L__BB108_2066; $L__BB108_2045: add.ftz.f32 %f12336, %f3249, 0f3F800000; mul.ftz.f32 %f3250, %f12336, 0f3F000000; setp.lt.ftz.f32 %p1795, %f3250, 0f38D1B717; mov.f32 %f16522, 0f00000000; mov.f32 %f16523, %f16522; @%p1795 bra $L__BB108_2050; add.ftz.f32 %f12339, %f3243, %f3243; mul.ftz.f32 %f12340, %f12339, %f3250; div.approx.ftz.f32 %f12341, %f12340, %f3249; add.ftz.f32 %f3251, %f12341, 0fBF800000; mul.ftz.f32 %f3252, %f3251, %f3251; add.ftz.f32 %f3253, %f3252, 0fBF800000; abs.ftz.f32 %f12342, %f3253; setp.lt.ftz.f32 %p1796, %f12342, 0f33D6BF95; mov.f32 %f16523, %f16522; @%p1796 bra $L__BB108_2050; mul.ftz.f32 %f12343, %f3249, %f3249; mov.f32 %f12344, 0f3F800000; sub.ftz.f32 %f12345, %f12344, %f12343; mov.f32 %f12346, 0f00000000; max.ftz.f32 %f12347, %f12345, %f12346; sqrt.approx.ftz.f32 %f12348, %f12347; div.approx.ftz.f32 %f12349, %f12348, %f3249; rcp.approx.ftz.f32 %f12350, %f3253; mul.ftz.f32 %f12351, %f12349, %f12349; mul.ftz.f32 %f12352, %f12351, %f12350; mul.ftz.f32 %f12353, %f12350, %f12352; sub.ftz.f32 %f12354, %f3252, %f12351; mul.ftz.f32 %f12355, %f12354, %f12350; sub.ftz.f32 %f12356, %f12353, %f12355; max.ftz.f32 %f12357, %f12356, %f12346; sqrt.approx.ftz.f32 %f12358, %f12357; mul.ftz.f32 %f12359, %f12349, %f12350; add.ftz.f32 %f12360, %f12359, %f12358; rcp.approx.ftz.f32 %f12361, %f12349; setp.gt.ftz.f32 %p1797, %f12360, %f12361; setp.lt.ftz.f32 %p1798, %f3251, 0f00000000; or.pred %p1799, %p1798, %p1797; sub.ftz.f32 %f12362, %f12359, %f12358; mov.f32 %f12363, 0f3F000000; sub.ftz.f32 %f12364, %f12363, %f3242; add.ftz.f32 %f12365, %f3242, 0fBF000000; setp.ge.ftz.f32 %p1800, %f3242, 0f3F000000; selp.f32 %f12366, %f12365, %f12364, %p1800; add.ftz.f32 %f12367, %f12366, %f12366; fma.rn.ftz.f32 %f12368, %f12367, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f12369, %f12367, %f12368, 0f3EED4413; mul.ftz.f32 %f12370, %f12367, %f12369; fma.rn.ftz.f32 %f12371, %f12367, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f12372, %f12367, %f12371, 0fBF800000; fma.rn.ftz.f32 %f12373, %f12367, %f12372, 0f3F191676; div.approx.ftz.f32 %f12374, %f12370, %f12373; selp.f32 %f16522, %f12362, %f12360, %p1799; fma.rn.ftz.f32 %f12375, %f16522, %f16522, 0f3F800000; sqrt.approx.ftz.f32 %f12376, %f12375; mul.ftz.f32 %f16523, %f12374, %f12376; setp.ltu.ftz.f32 %p1801, %f3242, 0f3F000000; @%p1801 bra $L__BB108_2048; bra.uni $L__BB108_2050; $L__BB108_2048: neg.ftz.f32 %f16523, %f16523; bra.uni $L__BB108_2050; $L__BB108_2061: add.ftz.f32 %f3308, %f3221, %f3286; setp.leu.ftz.f32 %p1813, %f3308, 0f00000000; mov.f32 %f16533, %f3227; @%p1813 bra $L__BB108_2063; sqrt.approx.ftz.f32 %f12450, %f3308; sub.ftz.f32 %f12451, %f12450, %f3285; add.ftz.f32 %f12452, %f3285, %f12450; div.approx.ftz.f32 %f12453, %f12451, %f12452; fma.rn.ftz.f32 %f12454, %f3285, %f12452, 0fBF800000; fma.rn.ftz.f32 %f12455, %f3285, %f12451, 0f3F800000; div.approx.ftz.f32 %f12456, %f12454, %f12455; mul.ftz.f32 %f12457, %f12453, 0f3F000000; mul.ftz.f32 %f12458, %f12453, %f12457; fma.rn.ftz.f32 %f12459, %f12456, %f12456, 0f3F800000; mul.ftz.f32 %f12460, %f12458, %f12459; sub.ftz.f32 %f16533, %f12460, %f16511; $L__BB108_2063: mul.ftz.f32 %f12461, %f3228, %f16533; mov.f32 %f12462, 0f3F800000; sub.ftz.f32 %f12463, %f12462, %f12461; fma.rn.ftz.f32 %f16534, %f3213, %f12463, %f12461; fma.rn.ftz.f32 %f16535, %f3214, %f12463, %f12461; fma.rn.ftz.f32 %f16536, %f3215, %f12463, %f12461; setp.eq.s32 %p1814, %r3280, 0; @%p1814 bra $L__BB108_2066; mul.ftz.f32 %f16534, %f3280, %f16534; mul.ftz.f32 %f16535, %f3281, %f16535; mul.ftz.f32 %f16536, %f3282, %f16536; $L__BB108_2066: setp.gt.ftz.f32 %p1815, %f3232, 0f3F7FF972; mov.f32 %f16537, 0f00000000; @%p1815 bra $L__BB108_2069; setp.lt.ftz.f32 %p1816, %f3232, 0fBF7FF972; mov.f32 %f16537, 0fBF7FF972; @%p1816 bra $L__BB108_2069; mul.ftz.f32 %f12466, %f3232, %f3232; mov.f32 %f12467, 0f33D6BF95; max.ftz.f32 %f12468, %f12466, %f12467; rcp.approx.ftz.f32 %f12469, %f12468; mul.ftz.f32 %f12470, %f12256, %f3234; mul.ftz.f32 %f12471, %f12257, %f3233; mul.ftz.f32 %f12472, %f12471, %f12471; fma.rn.ftz.f32 %f12473, %f12470, %f12470, %f12472; fma.rn.ftz.f32 %f12474, %f12473, %f12469, 0f3F800000; sqrt.approx.ftz.f32 %f12475, %f12474; neg.ftz.f32 %f12476, %f12475; setp.le.ftz.f32 %p1817, %f3232, 0f00000000; selp.f32 %f12477, %f12476, %f12475, %p1817; add.ftz.f32 %f12478, %f12477, 0fBF800000; mul.ftz.f32 %f16537, %f12478, 0f3F000000; $L__BB108_2069: mov.f32 %f16538, 0f3F800000; @%p1815 bra $L__BB108_2072; setp.lt.ftz.f32 %p1819, %f3232, 0f3727C5AC; mov.f32 %f16538, 0f00000000; @%p1819 bra $L__BB108_2072; lg2.approx.ftz.f32 %f12481, %f16521; mul.ftz.f32 %f12482, %f16537, %f12481; ex2.approx.ftz.f32 %f16538, %f12482; $L__BB108_2072: add.s32 %r3280, %r3280, 1; setp.lt.u32 %p1820, %r3280, 10; @%p1820 bra $L__BB108_1864; st.local.u32 [%rd7+168], %r3279; mov.f32 %f3234, 0f00000000; mov.f32 %f3232, 0f3F800000; mov.f32 %f3233, %f3234; mov.f32 %f16534, %f3234; mov.f32 %f16535, %f3234; mov.f32 %f16536, %f3234; bra.uni $L__BB108_2076; $L__BB108_2307: mul.ftz.f32 %f15152, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f15153, %f15152; sqrt.approx.ftz.f32 %f15154, %f2775; mul.ftz.f32 %f15155, %f15154, %f15153; sin.approx.ftz.f32 %f15156, %f15152; mul.ftz.f32 %f15157, %f15154, %f15156; mul.ftz.f32 %f15158, %f15155, %f15155; mov.f32 %f15159, 0f3F800000; sub.ftz.f32 %f15160, %f15159, %f15158; mul.ftz.f32 %f15161, %f15157, %f15157; sub.ftz.f32 %f15162, %f15160, %f15161; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f15163, %f15162, %f16743; sqrt.approx.ftz.f32 %f4122, %f15163; ld.local.f32 %f15164, [%rd587+12]; ld.local.f32 %f15165, [%rd587+8]; setp.neu.ftz.f32 %p2057, %f15165, %f15164; ld.local.f32 %f15166, [%rd587+16]; setp.neu.ftz.f32 %p2058, %f15165, %f15166; or.pred %p2059, %p2057, %p2058; sub.ftz.f32 %f15167, %f15166, %f15164; neg.ftz.f32 %f15168, %f15164; neg.ftz.f32 %f15169, %f15166; selp.f32 %f15170, %f15169, %f15166, %p2059; add.ftz.f32 %f15171, %f15165, %f15170; selp.f32 %f15172, %f15164, %f15168, %p2059; sub.ftz.f32 %f15173, %f15172, %f15165; mul.ftz.f32 %f15174, %f15171, %f15171; fma.rn.ftz.f32 %f15175, %f15167, %f15167, %f15174; fma.rn.ftz.f32 %f15176, %f15173, %f15173, %f15175; sqrt.approx.ftz.f32 %f15177, %f15176; div.approx.ftz.f32 %f15178, %f15167, %f15177; div.approx.ftz.f32 %f15179, %f15171, %f15177; div.approx.ftz.f32 %f15180, %f15173, %f15177; mul.ftz.f32 %f15181, %f15164, %f15180; mul.ftz.f32 %f15182, %f15166, %f15179; sub.ftz.f32 %f15183, %f15181, %f15182; mul.ftz.f32 %f15184, %f15166, %f15178; mul.ftz.f32 %f15185, %f15165, %f15180; sub.ftz.f32 %f15186, %f15184, %f15185; mul.ftz.f32 %f15187, %f15165, %f15179; mul.ftz.f32 %f15188, %f15164, %f15178; sub.ftz.f32 %f15189, %f15187, %f15188; mul.ftz.f32 %f15190, %f15157, %f15183; mul.ftz.f32 %f15191, %f15157, %f15186; mul.ftz.f32 %f15192, %f15157, %f15189; fma.rn.ftz.f32 %f15193, %f15155, %f15178, %f15190; fma.rn.ftz.f32 %f15194, %f15155, %f15179, %f15191; fma.rn.ftz.f32 %f15195, %f15155, %f15180, %f15192; fma.rn.ftz.f32 %f16740, %f15165, %f4122, %f15193; fma.rn.ftz.f32 %f16741, %f15164, %f4122, %f15194; fma.rn.ftz.f32 %f16742, %f15166, %f4122, %f15195; mul.ftz.f32 %f15196, %f2807, %f16741; fma.rn.ftz.f32 %f15197, %f2806, %f16740, %f15196; fma.rn.ftz.f32 %f15198, %f2808, %f16742, %f15197; setp.leu.ftz.f32 %p2060, %f15198, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2060 bra $L__BB108_2309; mul.ftz.f32 %f16743, %f4122, 0f3EA2F983; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_2303: mul.ftz.f32 %f15084, %f2799, %f2799; mov.f32 %f15085, 0f3F800000; sub.ftz.f32 %f15086, %f15085, %f15084; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f15087, %f16743, %f15086; sqrt.approx.ftz.f32 %f15088, %f15087; mul.ftz.f32 %f15089, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f15090, %f15089; mul.ftz.f32 %f15091, %f15088, %f15090; sin.approx.ftz.f32 %f15092, %f15089; mul.ftz.f32 %f15093, %f15088, %f15092; ld.local.f32 %f4110, [%rd587+12]; ld.local.f32 %f4111, [%rd587+8]; setp.neu.ftz.f32 %p2052, %f4111, %f4110; ld.local.f32 %f4112, [%rd587+16]; setp.neu.ftz.f32 %p2053, %f4111, %f4112; or.pred %p2054, %p2052, %p2053; sub.ftz.f32 %f15094, %f4112, %f4110; neg.ftz.f32 %f15095, %f4110; neg.ftz.f32 %f15096, %f4112; selp.f32 %f15097, %f15096, %f4112, %p2054; add.ftz.f32 %f15098, %f4111, %f15097; selp.f32 %f15099, %f4110, %f15095, %p2054; sub.ftz.f32 %f15100, %f15099, %f4111; mul.ftz.f32 %f15101, %f15098, %f15098; fma.rn.ftz.f32 %f15102, %f15094, %f15094, %f15101; fma.rn.ftz.f32 %f15103, %f15100, %f15100, %f15102; sqrt.approx.ftz.f32 %f15104, %f15103; div.approx.ftz.f32 %f15105, %f15094, %f15104; div.approx.ftz.f32 %f15106, %f15098, %f15104; div.approx.ftz.f32 %f15107, %f15100, %f15104; mul.ftz.f32 %f15108, %f4110, %f15107; mul.ftz.f32 %f15109, %f4112, %f15106; sub.ftz.f32 %f15110, %f15108, %f15109; mul.ftz.f32 %f15111, %f4112, %f15105; mul.ftz.f32 %f15112, %f4111, %f15107; sub.ftz.f32 %f15113, %f15111, %f15112; mul.ftz.f32 %f15114, %f4111, %f15106; mul.ftz.f32 %f15115, %f4110, %f15105; sub.ftz.f32 %f15116, %f15114, %f15115; mul.ftz.f32 %f15117, %f15093, %f15110; mul.ftz.f32 %f15118, %f15093, %f15113; mul.ftz.f32 %f15119, %f15093, %f15116; fma.rn.ftz.f32 %f15120, %f15091, %f15105, %f15117; fma.rn.ftz.f32 %f15121, %f15091, %f15106, %f15118; fma.rn.ftz.f32 %f15122, %f15091, %f15107, %f15119; fma.rn.ftz.f32 %f16740, %f2799, %f4111, %f15120; fma.rn.ftz.f32 %f16741, %f2799, %f4110, %f15121; fma.rn.ftz.f32 %f16742, %f2799, %f4112, %f15122; mul.ftz.f32 %f15123, %f2807, %f16741; fma.rn.ftz.f32 %f15124, %f2806, %f16740, %f15123; fma.rn.ftz.f32 %f15125, %f2808, %f16742, %f15124; setp.leu.ftz.f32 %p2055, %f15125, 0f00000000; mov.u32 %r3285, 6; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2055 bra $L__BB108_2309; ld.local.f32 %f15126, [%rd7+-52]; ld.local.v2.f32 {%f15127, %f15128}, [%rd7+-48]; mul.ftz.f32 %f15131, %f16741, %f4110; fma.rn.ftz.f32 %f15132, %f16740, %f4111, %f15131; fma.rn.ftz.f32 %f15133, %f16742, %f4112, %f15132; mov.f32 %f15134, 0f00000000; max.ftz.f32 %f4116, %f15133, %f15134; mul.ftz.f32 %f15135, %f15127, %f4110; fma.rn.ftz.f32 %f15136, %f15126, %f4111, %f15135; fma.rn.ftz.f32 %f15137, %f15128, %f4112, %f15136; max.ftz.f32 %f4117, %f15137, %f15134; mul.ftz.f32 %f15138, %f15127, %f16741; fma.rn.ftz.f32 %f15139, %f15126, %f16740, %f15138; fma.rn.ftz.f32 %f15140, %f15128, %f16742, %f15139; mul.ftz.f32 %f15141, %f4116, %f4117; sub.ftz.f32 %f16739, %f15140, %f15141; setp.leu.ftz.f32 %p2056, %f16739, 0f00000000; @%p2056 bra $L__BB108_2306; max.ftz.f32 %f15142, %f4116, %f4117; add.ftz.f32 %f15143, %f15142, 0f00800000; div.approx.ftz.f32 %f16739, %f16739, %f15143; $L__BB108_2306: ld.local.f32 %f15145, [%rd587+28]; ld.local.f32 %f15146, [%rd587+24]; fma.rn.ftz.f32 %f15147, %f16739, %f15145, %f15146; mul.ftz.f32 %f16744, %f4116, %f15147; mov.f32 %f16743, 0f3E22F983; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_2097: ld.local.v2.f32 {%f12713, %f12714}, [%rd7+-48]; ld.local.f32 %f3427, [%rd587+36]; ld.local.f32 %f3428, [%rd587+20]; ld.local.f32 %f12717, [%rd7+-52]; ld.local.f32 %f3429, [%rd587+24]; mul.ftz.f32 %f12718, %f12713, %f3429; fma.rn.ftz.f32 %f12719, %f12717, %f3428, %f12718; ld.local.f32 %f3430, [%rd587+28]; fma.rn.ftz.f32 %f12720, %f12714, %f3430, %f12719; mul.ftz.f32 %f12721, %f3428, %f12720; mul.ftz.f32 %f12722, %f3429, %f12720; mul.ftz.f32 %f12723, %f3430, %f12720; sub.ftz.f32 %f12724, %f12717, %f12721; sub.ftz.f32 %f12725, %f12713, %f12722; sub.ftz.f32 %f12726, %f12714, %f12723; mul.ftz.f32 %f12727, %f12725, %f12725; fma.rn.ftz.f32 %f12728, %f12724, %f12724, %f12727; fma.rn.ftz.f32 %f12729, %f12726, %f12726, %f12728; sqrt.approx.ftz.f32 %f12730, %f12729; div.approx.ftz.f32 %f3431, %f12724, %f12730; div.approx.ftz.f32 %f3432, %f12725, %f12730; div.approx.ftz.f32 %f3433, %f12726, %f12730; abs.ftz.f32 %f12731, %f12720; setp.lt.ftz.f32 %p1840, %f12731, 0f3F800000; mov.f32 %f12732, 0f3F800000; add.ftz.f32 %f12733, %f12731, 0fBF800000; add.ftz.f32 %f12734, %f12733, 0f3F800000; selp.f32 %f12735, %f12734, 0f3F800000, %p1840; sub.ftz.f32 %f12736, %f12732, %f12735; sqrt.approx.ftz.f32 %f12737, %f12736; fma.rn.ftz.f32 %f12738, %f12735, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f12739, %f12735, %f12738, 0fBE5A6B93; fma.rn.ftz.f32 %f12740, %f12735, %f12739, 0f3FC90FDB; mov.f32 %f12741, 0f3FC90FDB; mul.ftz.f32 %f12742, %f12737, %f12740; setp.lt.ftz.f32 %p1841, %f12720, 0f00000000; mov.f32 %f16570, 0f00000000; mov.f32 %f12743, 0f40490FDB; sub.ftz.f32 %f12744, %f12743, %f12742; selp.f32 %f12745, %f12744, %f12742, %p1841; sub.ftz.f32 %f3434, %f12741, %f12745; ld.local.f32 %f3435, [%rd587+32]; rcp.approx.ftz.f32 %f3436, %f3435; add.ftz.f32 %f12746, %f3434, 0f3FC90FDB; mul.ftz.f32 %f12747, %f12746, 0f3F000000; ld.local.f32 %f3437, [%rd587+40]; sub.ftz.f32 %f12748, %f12747, %f3437; mul.ftz.f32 %f3438, %f3436, %f12748; abs.ftz.f32 %f3439, %f3438; setp.eq.ftz.f32 %p1842, %f3439, 0f00000000; mov.f32 %f16569, %f16570; @%p1842 bra $L__BB108_2102; setp.eq.ftz.f32 %p1843, %f3439, 0f3F800000; mov.f32 %f16569, %f12732; @%p1843 bra $L__BB108_2102; setp.gt.ftz.f32 %p1844, %f3439, 0f3F800000; @%p1844 bra $L__BB108_2101; bra.uni $L__BB108_2100; $L__BB108_2101: mov.f32 %f12751, 0f3F800000; div.approx.ftz.f32 %f16569, %f12751, %f3439; bra.uni $L__BB108_2102; $L__BB108_2091: ld.local.v2.f32 {%f12676, %f12677}, [%rd7+-48]; ld.local.f32 %f3401, [%rd587+28]; ld.local.f32 %f16560, [%rd587+8]; ld.local.f32 %f3403, [%rd7+-52]; ld.local.f32 %f16561, [%rd587+12]; mul.ftz.f32 %f12678, %f12676, %f16561; fma.rn.ftz.f32 %f12679, %f3403, %f16560, %f12678; ld.local.f32 %f16562, [%rd587+16]; fma.rn.ftz.f32 %f16563, %f12677, %f16562, %f12679; setp.gt.ftz.f32 %p1835, %f16563, 0f00000000; @%p1835 bra $L__BB108_2093; bra.uni $L__BB108_2092; $L__BB108_2093: rcp.approx.ftz.f32 %f16564, %f3401; mov.u16 %rs157, 0; bra.uni $L__BB108_2094; $L__BB108_1890: ld.local.u64 %rd588, [%rd587+52]; ld.v4.f32 {%f11523, %f11524, %f11525, %f11526}, [%rd588]; ld.local.v2.f32 {%f16467, %f16468}, [%rd7+32]; ld.local.f32 %f16466, [%rd7+28]; mul.ftz.f32 %f11529, %f16467, %f16467; fma.rn.ftz.f32 %f11530, %f16466, %f16466, %f11529; fma.rn.ftz.f32 %f11531, %f16468, %f16468, %f11530; sqrt.approx.ftz.f32 %f2848, %f11531; setp.neu.ftz.f32 %p1651, %f2848, 0f00000000; @%p1651 bra $L__BB108_1891; bra.uni $L__BB108_1892; $L__BB108_1891: rcp.approx.ftz.f32 %f11532, %f2848; mul.ftz.f32 %f16468, %f16468, %f11532; mul.ftz.f32 %f16467, %f16467, %f11532; mul.ftz.f32 %f16466, %f16466, %f11532; $L__BB108_1892: mul.ftz.f32 %f11533, %f11525, %f16467; mul.ftz.f32 %f11534, %f11524, %f16468; sub.ftz.f32 %f16469, %f11533, %f11534; mul.ftz.f32 %f11535, %f11523, %f16468; mul.ftz.f32 %f11536, %f11525, %f16466; sub.ftz.f32 %f16470, %f11535, %f11536; mul.ftz.f32 %f11537, %f11524, %f16466; mul.ftz.f32 %f11538, %f11523, %f16467; sub.ftz.f32 %f16471, %f11537, %f11538; mul.ftz.f32 %f11539, %f16470, %f16470; fma.rn.ftz.f32 %f11540, %f16469, %f16469, %f11539; fma.rn.ftz.f32 %f11541, %f16471, %f16471, %f11540; sqrt.approx.ftz.f32 %f2864, %f11541; setp.neu.ftz.f32 %p1652, %f2864, 0f00000000; @%p1652 bra $L__BB108_1893; bra.uni $L__BB108_1894; $L__BB108_1893: rcp.approx.ftz.f32 %f11542, %f2864; mul.ftz.f32 %f16471, %f11542, %f16471; mul.ftz.f32 %f16470, %f11542, %f16470; mul.ftz.f32 %f16469, %f11542, %f16469; $L__BB108_1894: ld.local.f32 %f11543, [%rd7+-52]; ld.local.v2.f32 {%f11544, %f11545}, [%rd7+-48]; mul.ftz.f32 %f11548, %f16467, %f11544; fma.rn.ftz.f32 %f11549, %f16466, %f11543, %f11548; fma.rn.ftz.f32 %f2874, %f16468, %f11545, %f11549; mul.ftz.f32 %f11550, %f11524, %f11544; fma.rn.ftz.f32 %f11551, %f11523, %f11543, %f11550; fma.rn.ftz.f32 %f11552, %f11525, %f11545, %f11551; mul.ftz.f32 %f11553, %f16470, %f11544; fma.rn.ftz.f32 %f11554, %f16469, %f11543, %f11553; fma.rn.ftz.f32 %f11555, %f16471, %f11545, %f11554; ld.local.u32 %r2717, [%rd7+168]; mad.lo.s32 %r876, %r2717, 1103515245, 12345; mad.lo.s32 %r877, %r876, 1103515245, 12345; st.local.u32 [%rd7+168], %r877; mul.ftz.f32 %f2878, %f2874, %f2874; mov.f32 %f11556, 0f3F800000; sub.ftz.f32 %f11557, %f11556, %f2878; mov.f32 %f16477, 0f00000000; max.ftz.f32 %f11559, %f11557, %f16477; sqrt.approx.ftz.f32 %f2879, %f11559; abs.ftz.f32 %f2880, %f11552; abs.ftz.f32 %f2881, %f11555; setp.eq.ftz.f32 %p1653, %f2880, 0f00000000; setp.eq.ftz.f32 %p1654, %f2881, 0f00000000; and.pred %p1655, %p1653, %p1654; mov.b32 %r878, %f11552; mov.b32 %r2718, %f11555; and.b32 %r879, %r2718, -2147483648; @%p1655 bra $L__BB108_1898; bra.uni $L__BB108_1895; $L__BB108_1898: shr.s32 %r2723, %r878, 31; and.b32 %r2724, %r2723, 1078530011; or.b32 %r2725, %r879, %r2724; mov.b32 %f16472, %r2725; bra.uni $L__BB108_1899; $L__BB108_1874: ld.local.v2.f32 {%f11315, %f11316}, [%rd7+-48]; ld.local.f32 %f2812, [%rd587+36]; ld.local.f32 %f2813, [%rd587+20]; ld.local.f32 %f11319, [%rd7+-52]; ld.local.f32 %f2814, [%rd587+24]; mul.ftz.f32 %f11320, %f11315, %f2814; fma.rn.ftz.f32 %f11321, %f11319, %f2813, %f11320; ld.local.f32 %f2815, [%rd587+28]; fma.rn.ftz.f32 %f11322, %f11316, %f2815, %f11321; mul.ftz.f32 %f11323, %f2813, %f11322; mul.ftz.f32 %f11324, %f2814, %f11322; mul.ftz.f32 %f11325, %f2815, %f11322; sub.ftz.f32 %f11326, %f11319, %f11323; sub.ftz.f32 %f11327, %f11315, %f11324; sub.ftz.f32 %f11328, %f11316, %f11325; mul.ftz.f32 %f11329, %f11327, %f11327; fma.rn.ftz.f32 %f11330, %f11326, %f11326, %f11329; fma.rn.ftz.f32 %f11331, %f11328, %f11328, %f11330; sqrt.approx.ftz.f32 %f11332, %f11331; div.approx.ftz.f32 %f2816, %f11326, %f11332; div.approx.ftz.f32 %f2817, %f11327, %f11332; div.approx.ftz.f32 %f2818, %f11328, %f11332; abs.ftz.f32 %f11333, %f11322; setp.lt.ftz.f32 %p1630, %f11333, 0f3F800000; mov.f32 %f11334, 0f3F800000; add.ftz.f32 %f11335, %f11333, 0fBF800000; add.ftz.f32 %f11336, %f11335, 0f3F800000; selp.f32 %f11337, %f11336, 0f3F800000, %p1630; sub.ftz.f32 %f11338, %f11334, %f11337; sqrt.approx.ftz.f32 %f11339, %f11338; fma.rn.ftz.f32 %f11340, %f11337, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f11341, %f11337, %f11340, 0fBE5A6B93; fma.rn.ftz.f32 %f11342, %f11337, %f11341, 0f3FC90FDB; mov.f32 %f11343, 0f3FC90FDB; mul.ftz.f32 %f11344, %f11339, %f11342; setp.lt.ftz.f32 %p1631, %f11322, 0f00000000; mov.f32 %f16465, 0f00000000; mov.f32 %f11345, 0f40490FDB; sub.ftz.f32 %f11346, %f11345, %f11344; selp.f32 %f11347, %f11346, %f11344, %p1631; sub.ftz.f32 %f2819, %f11343, %f11347; ld.local.f32 %f2820, [%rd587+32]; rcp.approx.ftz.f32 %f2821, %f2820; add.ftz.f32 %f11348, %f2819, 0f3FC90FDB; mov.f32 %f11349, 0f40000000; div.approx.ftz.f32 %f11350, %f11348, %f11349; ld.local.f32 %f2822, [%rd587+40]; sub.ftz.f32 %f11351, %f11350, %f2822; mul.ftz.f32 %f2823, %f2821, %f11351; abs.ftz.f32 %f2824, %f2823; setp.eq.ftz.f32 %p1632, %f2824, 0f00000000; mov.f32 %f16463, %f16465; @%p1632 bra $L__BB108_1879; setp.eq.ftz.f32 %p1633, %f2824, 0f3F800000; mov.f32 %f16463, %f11334; @%p1633 bra $L__BB108_1879; setp.gt.ftz.f32 %p1634, %f2824, 0f3F800000; @%p1634 bra $L__BB108_1878; bra.uni $L__BB108_1877; $L__BB108_1878: mov.f32 %f11354, 0f3F800000; div.approx.ftz.f32 %f16463, %f11354, %f2824; bra.uni $L__BB108_1879; $L__BB108_2292: ld.local.f32 %f14799, [%rd587+8]; ld.local.f32 %f14800, [%rd587+12]; neg.ftz.f32 %f14801, %f14800; ld.local.f32 %f14802, [%rd587+16]; neg.ftz.f32 %f14803, %f14802; mul.ftz.f32 %f14804, %f2799, 0f40C90FDB; cos.approx.ftz.f32 %f14805, %f14804; sqrt.approx.ftz.f32 %f14806, %f2775; mul.ftz.f32 %f14807, %f14806, %f14805; sin.approx.ftz.f32 %f14808, %f14804; mul.ftz.f32 %f14809, %f14806, %f14808; mul.ftz.f32 %f14810, %f14807, %f14807; mov.f32 %f14811, 0f3F800000; sub.ftz.f32 %f14812, %f14811, %f14810; mul.ftz.f32 %f14813, %f14809, %f14809; sub.ftz.f32 %f14814, %f14812, %f14813; mov.f32 %f16743, 0f00000000; max.ftz.f32 %f14815, %f14814, %f16743; sqrt.approx.ftz.f32 %f4042, %f14815; setp.neu.ftz.f32 %p2024, %f14799, %f14800; setp.neu.ftz.f32 %p2025, %f14799, %f14802; or.pred %p2026, %p2024, %p2025; sub.ftz.f32 %f14816, %f14800, %f14802; selp.f32 %f14817, %f14802, %f14803, %p2026; selp.f32 %f14818, %f14801, %f14800, %p2026; add.ftz.f32 %f14819, %f14799, %f14818; sub.ftz.f32 %f14820, %f14817, %f14799; mul.ftz.f32 %f14821, %f14820, %f14820; fma.rn.ftz.f32 %f14822, %f14816, %f14816, %f14821; fma.rn.ftz.f32 %f14823, %f14819, %f14819, %f14822; sqrt.approx.ftz.f32 %f14824, %f14823; div.approx.ftz.f32 %f14825, %f14816, %f14824; div.approx.ftz.f32 %f14826, %f14820, %f14824; div.approx.ftz.f32 %f14827, %f14819, %f14824; mul.ftz.f32 %f14828, %f14827, %f14800; mul.ftz.f32 %f14829, %f14826, %f14802; sub.ftz.f32 %f14830, %f14829, %f14828; mul.ftz.f32 %f14831, %f14825, %f14802; mul.ftz.f32 %f14832, %f14827, %f14799; sub.ftz.f32 %f14833, %f14832, %f14831; mul.ftz.f32 %f14834, %f14826, %f14799; mul.ftz.f32 %f14835, %f14825, %f14800; sub.ftz.f32 %f14836, %f14835, %f14834; mul.ftz.f32 %f14837, %f14809, %f14830; mul.ftz.f32 %f14838, %f14809, %f14833; mul.ftz.f32 %f14839, %f14809, %f14836; fma.rn.ftz.f32 %f14840, %f14807, %f14825, %f14837; fma.rn.ftz.f32 %f14841, %f14807, %f14826, %f14838; fma.rn.ftz.f32 %f14842, %f14807, %f14827, %f14839; mul.ftz.f32 %f14843, %f4042, %f14799; mul.ftz.f32 %f14844, %f4042, %f14800; mul.ftz.f32 %f14845, %f4042, %f14802; sub.ftz.f32 %f16740, %f14840, %f14843; sub.ftz.f32 %f16741, %f14841, %f14844; sub.ftz.f32 %f16742, %f14842, %f14845; mul.ftz.f32 %f14846, %f2807, %f16741; fma.rn.ftz.f32 %f14847, %f2806, %f16740, %f14846; fma.rn.ftz.f32 %f14848, %f2808, %f16742, %f14847; setp.geu.ftz.f32 %p2027, %f14848, 0f00000000; mov.u32 %r3285, 5; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2027 bra $L__BB108_2309; mul.ftz.f32 %f16743, %f4042, 0f3EA2F983; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1852: setp.eq.s32 %p1956, %r873, 22; selp.b32 %r3285, 9, 10, %p1956; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_2249: ld.local.f32 %f14376, [%rd587+52]; mul.ftz.f32 %f14377, %f3900, %f14376; ld.local.f32 %f14378, [%rd587+48]; mul.ftz.f32 %f14379, %f3901, %f14378; sub.ftz.f32 %f14380, %f14377, %f14379; ld.local.f32 %f14381, [%rd587+44]; mul.ftz.f32 %f14382, %f3901, %f14381; mul.ftz.f32 %f14383, %f3898, %f14376; sub.ftz.f32 %f14384, %f14382, %f14383; mul.ftz.f32 %f14385, %f3898, %f14378; mul.ftz.f32 %f14386, %f3900, %f14381; sub.ftz.f32 %f14387, %f14385, %f14386; mul.ftz.f32 %f14388, %f14384, %f14384; fma.rn.ftz.f32 %f14389, %f14380, %f14380, %f14388; fma.rn.ftz.f32 %f14390, %f14387, %f14387, %f14389; sqrt.approx.ftz.f32 %f14391, %f14390; div.approx.ftz.f32 %f16693, %f14380, %f14391; div.approx.ftz.f32 %f16694, %f14384, %f14391; div.approx.ftz.f32 %f16695, %f14387, %f14391; mul.ftz.f32 %f14392, %f3901, %f16694; mul.ftz.f32 %f14393, %f3900, %f16695; sub.ftz.f32 %f16696, %f14392, %f14393; mul.ftz.f32 %f14394, %f3898, %f16695; mul.ftz.f32 %f14395, %f3901, %f16693; sub.ftz.f32 %f16697, %f14394, %f14395; mul.ftz.f32 %f14396, %f3900, %f16693; mul.ftz.f32 %f14397, %f3898, %f16694; sub.ftz.f32 %f16698, %f14396, %f14397; $L__BB108_2251: mul.ftz.f32 %f14419, %f14365, %f16697; fma.rn.ftz.f32 %f14420, %f3899, %f16696, %f14419; fma.rn.ftz.f32 %f14421, %f14366, %f16698, %f14420; mul.ftz.f32 %f14422, %f14365, %f16694; fma.rn.ftz.f32 %f14423, %f3899, %f16693, %f14422; fma.rn.ftz.f32 %f14424, %f14366, %f16695, %f14423; mul.ftz.f32 %f14425, %f14367, %f14421; mul.ftz.f32 %f14426, %f14368, %f14424; mul.ftz.f32 %f14427, %f14425, %f14425; fma.rn.ftz.f32 %f14428, %f14426, %f14426, %f14427; mul.ftz.f32 %f3921, %f3902, %f3902; add.ftz.f32 %f14429, %f3921, %f14428; sqrt.approx.ftz.f32 %f14430, %f14429; div.approx.ftz.f32 %f3922, %f14425, %f14430; div.approx.ftz.f32 %f3923, %f14426, %f14430; div.approx.ftz.f32 %f3924, %f3902, %f14430; setp.geu.ftz.f32 %p1995, %f3924, 0f3F7FFF58; mov.f32 %f16700, 0f00000000; mov.f32 %f16699, 0f3F800000; mov.f32 %f16701, %f16699; mov.f32 %f16702, %f16700; @%p1995 bra $L__BB108_2253; mul.ftz.f32 %f14431, %f3924, %f3924; mov.f32 %f14432, 0f3F800000; sub.ftz.f32 %f14433, %f14432, %f14431; mov.f32 %f14434, 0f00000000; max.ftz.f32 %f14435, %f14433, %f14434; sqrt.approx.ftz.f32 %f16700, %f14435; rsqrt.approx.ftz.f32 %f14436, %f14435; mul.ftz.f32 %f16701, %f3922, %f14436; mul.ftz.f32 %f16702, %f3923, %f14436; mov.f32 %f16699, %f3924; $L__BB108_2253: setp.ltu.ftz.f32 %p1996, %f16699, 0f3F7FFF58; @%p1996 bra $L__BB108_2255; bra.uni $L__BB108_2254; $L__BB108_2255: div.approx.ftz.f32 %f14444, %f16700, %f16699; mul.ftz.f32 %f14445, %f14444, %f14444; add.ftz.f32 %f14446, %f14445, 0f3F800000; mov.f32 %f14447, 0f00000000; max.ftz.f32 %f14448, %f14446, %f14447; sqrt.approx.ftz.f32 %f14449, %f14448; add.ftz.f32 %f14450, %f14449, 0f3F800000; mul.ftz.f32 %f14451, %f14450, 0f3F000000; mov.f32 %f14452, 0f3F000000; rcp.approx.ftz.f32 %f3937, %f14451; add.ftz.f32 %f14453, %f2799, %f2799; fma.rn.ftz.f32 %f14454, %f14453, %f14451, 0fBF800000; mul.ftz.f32 %f14455, %f14454, %f14454; add.ftz.f32 %f14456, %f14455, 0fBF800000; rcp.approx.ftz.f32 %f14457, %f14456; mul.ftz.f32 %f14458, %f14457, %f14457; mul.ftz.f32 %f14459, %f14445, %f14458; sub.ftz.f32 %f14460, %f14455, %f14445; mul.ftz.f32 %f14461, %f14460, %f14457; sub.ftz.f32 %f14462, %f14459, %f14461; max.ftz.f32 %f14463, %f14462, %f14447; sqrt.approx.ftz.f32 %f14464, %f14463; mul.ftz.f32 %f14465, %f14444, %f14457; sub.ftz.f32 %f14466, %f14465, %f14464; add.ftz.f32 %f14467, %f14464, %f14465; setp.lt.ftz.f32 %p1997, %f14454, 0f00000000; mul.ftz.f32 %f14468, %f14444, %f14467; setp.gt.ftz.f32 %p1998, %f14468, 0f3F800000; or.pred %p1999, %p1997, %p1998; selp.f32 %f16705, %f14466, %f14467, %p1999; sub.ftz.f32 %f14469, %f14452, %f2775; add.ftz.f32 %f14470, %f2775, 0fBF000000; setp.gt.ftz.f32 %p2000, %f2775, 0f3F000000; selp.f32 %f14471, %f14470, %f14469, %p2000; add.ftz.f32 %f14472, %f14471, %f14471; selp.f32 %f14473, 0f3F800000, 0fBF800000, %p2000; fma.rn.ftz.f32 %f14474, %f14472, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f14475, %f14472, %f14474, 0f3EED4413; mul.ftz.f32 %f14476, %f14472, %f14475; fma.rn.ftz.f32 %f14477, %f14472, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f14478, %f14472, %f14477, 0fBF800000; fma.rn.ftz.f32 %f14479, %f14472, %f14478, 0f3F191676; div.approx.ftz.f32 %f14480, %f14476, %f14479; mul.ftz.f32 %f14481, %f14473, %f14480; fma.rn.ftz.f32 %f14482, %f16705, %f16705, 0f3F800000; max.ftz.f32 %f14483, %f14482, %f14447; sqrt.approx.ftz.f32 %f14484, %f14483; mul.ftz.f32 %f16704, %f14481, %f14484; bra.uni $L__BB108_2256; $L__BB108_2254: mov.f32 %f3937, 0f3F800000; sub.ftz.f32 %f14438, %f3937, %f2799; div.approx.ftz.f32 %f14439, %f2799, %f14438; sqrt.approx.ftz.f32 %f14440, %f14439; mul.ftz.f32 %f14441, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f14442, %f14441; mul.ftz.f32 %f16705, %f14440, %f14442; sin.approx.ftz.f32 %f14443, %f14441; mul.ftz.f32 %f16704, %f14440, %f14443; $L__BB108_2256: mul.ftz.f32 %f14485, %f16702, %f16704; mul.ftz.f32 %f14486, %f16701, %f16705; sub.ftz.f32 %f14487, %f14486, %f14485; mul.ftz.f32 %f14488, %f16702, %f16705; fma.rn.ftz.f32 %f14489, %f16701, %f16704, %f14488; mul.ftz.f32 %f14490, %f14367, %f14487; mul.ftz.f32 %f14491, %f14368, %f14489; neg.ftz.f32 %f14492, %f14490; neg.ftz.f32 %f14493, %f14491; mul.ftz.f32 %f14494, %f14491, %f14491; fma.rn.ftz.f32 %f14495, %f14490, %f14490, %f14494; add.ftz.f32 %f14496, %f14495, 0f3F800000; mov.f32 %f14497, 0f3F800000; sqrt.approx.ftz.f32 %f14498, %f14496; div.approx.ftz.f32 %f14499, %f14492, %f14498; div.approx.ftz.f32 %f14500, %f14493, %f14498; div.approx.ftz.f32 %f3940, %f14497, %f14498; mul.ftz.f32 %f14501, %f16693, %f14500; mul.ftz.f32 %f14502, %f16694, %f14500; mul.ftz.f32 %f14503, %f16695, %f14500; fma.rn.ftz.f32 %f14504, %f16696, %f14499, %f14501; fma.rn.ftz.f32 %f14505, %f16697, %f14499, %f14502; fma.rn.ftz.f32 %f14506, %f16698, %f14499, %f14503; fma.rn.ftz.f32 %f3941, %f3898, %f3940, %f14504; fma.rn.ftz.f32 %f3942, %f3900, %f3940, %f14505; fma.rn.ftz.f32 %f3943, %f3901, %f3940, %f14506; mul.ftz.f32 %f14507, %f14365, %f3942; fma.rn.ftz.f32 %f14508, %f3899, %f3941, %f14507; fma.rn.ftz.f32 %f3944, %f14366, %f3943, %f14508; setp.eq.s32 %p2001, %r873, 23; @%p2001 bra $L__BB108_2283; bra.uni $L__BB108_2257; $L__BB108_2283: setp.gt.ftz.f32 %p2015, %f3944, 0f00000000; ld.local.f32 %f4009, [%rd587+28]; @%p2015 bra $L__BB108_2285; bra.uni $L__BB108_2284; $L__BB108_2285: rcp.approx.ftz.f32 %f16728, %f4009; mov.u16 %rs159, 0; mov.f32 %f16724, %f3941; mov.f32 %f16725, %f3942; mov.f32 %f16726, %f3943; mov.f32 %f16727, %f3944; bra.uni $L__BB108_2286; $L__BB108_2257: setp.leu.ftz.f32 %p2002, %f3944, 0f00000000; mov.u32 %r3285, 10; mov.f32 %f16743, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2002 bra $L__BB108_2309; add.ftz.f32 %f14518, %f3944, %f3944; mul.ftz.f32 %f14519, %f3941, %f14518; mul.ftz.f32 %f14520, %f3942, %f14518; mul.ftz.f32 %f14521, %f3943, %f14518; sub.ftz.f32 %f16740, %f14519, %f3899; sub.ftz.f32 %f16741, %f14520, %f14365; sub.ftz.f32 %f16742, %f14521, %f14366; mul.ftz.f32 %f14522, %f2807, %f16741; fma.rn.ftz.f32 %f14523, %f2806, %f16740, %f14522; fma.rn.ftz.f32 %f14524, %f2808, %f16742, %f14523; setp.leu.ftz.f32 %p2003, %f14524, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2003 bra $L__BB108_2309; mul.ftz.f32 %f16710, %f14367, %f14368; setp.gtu.ftz.f32 %p2004, %f16710, 0f33D6BF95; @%p2004 bra $L__BB108_2266; bra.uni $L__BB108_2260; $L__BB108_2266: mul.ftz.f32 %f14564, %f3900, %f16741; fma.rn.ftz.f32 %f14565, %f3898, %f16740, %f14564; fma.rn.ftz.f32 %f3965, %f3901, %f16742, %f14565; @%p1991 bra $L__BB108_2268; bra.uni $L__BB108_2267; $L__BB108_2268: setp.eq.s32 %p2009, %r873, 12; @%p2009 bra $L__BB108_2270; bra.uni $L__BB108_2269; $L__BB108_2270: setp.ge.ftz.f32 %p2010, %f14367, 0f3F800000; mov.f32 %f16711, 0f3EA2F983; @%p2010 bra $L__BB108_2272; mul.ftz.f32 %f14616, %f14367, %f14367; add.ftz.f32 %f14617, %f14616, 0fBF800000; mul.ftz.f32 %f14618, %f3940, %f14617; fma.rn.ftz.f32 %f14619, %f3940, %f14618, 0f3F800000; lg2.approx.ftz.f32 %f14620, %f14616; mul.ftz.f32 %f14621, %f14620, 0f3F317218; mul.ftz.f32 %f14622, %f14621, 0f40490FDB; mul.ftz.f32 %f14623, %f14619, %f14622; div.approx.ftz.f32 %f16711, %f14617, %f14623; $L__BB108_2272: mov.f32 %f14625, 0f3F800000; sub.ftz.f32 %f14626, %f14625, %f3921; mul.ftz.f32 %f14627, %f14626, 0f3D800000; mov.f32 %f16710, 0f3D800000; div.approx.ftz.f32 %f14628, %f14627, %f3921; add.ftz.f32 %f14629, %f14628, 0f3F800000; mov.f32 %f14630, 0f00000000; max.ftz.f32 %f14631, %f14629, %f14630; sqrt.approx.ftz.f32 %f14632, %f14631; add.ftz.f32 %f14633, %f14632, 0f3F800000; mov.f32 %f14634, 0f40000000; div.approx.ftz.f32 %f3937, %f14634, %f14633; bra.uni $L__BB108_2273; $L__BB108_2074: st.local.u32 [%rd7+168], %r885; and.b16 %rs116, %rs155, 255; setp.eq.s16 %p1821, %rs116, 0; @%p1821 bra $L__BB108_2075; bra.uni $L__BB108_2076; $L__BB108_2075: neg.ftz.f32 %f3232, %f3232; neg.ftz.f32 %f3233, %f3233; neg.ftz.f32 %f3234, %f3234; $L__BB108_2076: ld.local.f32 %f3348, [%rd587+20]; mul.ftz.f32 %f3349, %f16554, %f3232; setp.gt.ftz.f32 %p1822, %f3349, 0f00000000; mov.f32 %f16548, 0f00000000; mul.ftz.f32 %f12490, %f3234, %f3212; selp.f32 %f12491, %f3234, %f12490, %p1822; mul.ftz.f32 %f12492, %f3233, %f3212; selp.f32 %f12493, %f3233, %f12492, %p1822; mul.ftz.f32 %f12494, %f3232, %f3212; selp.f32 %f12495, %f3232, %f12494, %p1822; add.ftz.f32 %f12496, %f3207, %f12491; add.ftz.f32 %f12497, %f3208, %f12493; add.ftz.f32 %f12498, %f16554, %f12495; mul.ftz.f32 %f12499, %f12497, %f12497; fma.rn.ftz.f32 %f12500, %f12496, %f12496, %f12499; fma.rn.ftz.f32 %f12501, %f12498, %f12498, %f12500; sqrt.approx.ftz.f32 %f3350, %f12501; rsqrt.approx.ftz.f32 %f12502, %f12501; mul.ftz.f32 %f12503, %f12502, %f12496; mul.ftz.f32 %f12504, %f12502, %f12497; mul.ftz.f32 %f12505, %f12502, %f12498; setp.lt.ftz.f32 %p1823, %f12505, 0f00000000; neg.ftz.f32 %f12506, %f12503; neg.ftz.f32 %f12507, %f12504; neg.ftz.f32 %f12508, %f12505; selp.f32 %f3351, %f12506, %f12503, %p1823; selp.f32 %f3352, %f12507, %f12504, %p1823; selp.f32 %f3353, %f12508, %f12505, %p1823; setp.lt.ftz.f32 %p1824, %f16554, 0f00000000; selp.f32 %f3354, %f3216, %f3207, %p1824; selp.f32 %f3355, %f3217, %f3208, %p1824; selp.f32 %f3356, %f3218, %f16554, %p1824; setp.gt.ftz.f32 %p1825, %f3356, 0f3F7FF972; @%p1825 bra $L__BB108_2079; setp.lt.ftz.f32 %p1826, %f3356, 0fBF7FF972; mov.f32 %f16548, 0fBF7FF972; @%p1826 bra $L__BB108_2079; mul.ftz.f32 %f12510, %f3356, %f3356; mov.f32 %f12511, 0f33D6BF95; max.ftz.f32 %f12512, %f12510, %f12511; rcp.approx.ftz.f32 %f12513, %f12512; mul.ftz.f32 %f12514, %f3354, %f3348; mul.ftz.f32 %f12515, %f3355, %f3348; mul.ftz.f32 %f12516, %f12515, %f12515; fma.rn.ftz.f32 %f12517, %f12514, %f12514, %f12516; fma.rn.ftz.f32 %f12518, %f12517, %f12513, 0f3F800000; sqrt.approx.ftz.f32 %f12519, %f12518; neg.ftz.f32 %f12520, %f12519; setp.le.ftz.f32 %p1827, %f3356, 0f00000000; selp.f32 %f12521, %f12520, %f12519, %p1827; add.ftz.f32 %f12522, %f12521, 0fBF800000; mul.ftz.f32 %f16548, %f12522, 0f3F000000; $L__BB108_2079: mul.ftz.f32 %f12524, %f3353, %f3353; mov.f32 %f16549, 0f3F800000; sub.ftz.f32 %f12525, %f16549, %f12524; mul.ftz.f32 %f12526, %f3348, %f3348; fma.rn.ftz.f32 %f12527, %f12526, %f12524, %f12525; mul.ftz.f32 %f12528, %f12527, 0f40490FDB; mul.ftz.f32 %f12529, %f12527, %f12528; mov.f32 %f12530, 0f33D6BF95; max.ftz.f32 %f12531, %f12529, %f12530; div.approx.ftz.f32 %f3359, %f12526, %f12531; mul.ftz.f32 %f12532, %f3355, %f3352; fma.rn.ftz.f32 %f12533, %f3354, %f3351, %f12532; fma.rn.ftz.f32 %f3360, %f3356, %f3353, %f12533; abs.ftz.f32 %f3361, %f3360; fma.rn.ftz.f32 %f3362, %f3361, %f3361, %f3221; setp.leu.ftz.f32 %p1828, %f3362, 0f00000000; @%p1828 bra $L__BB108_2081; sqrt.approx.ftz.f32 %f12534, %f3362; sub.ftz.f32 %f12535, %f12534, %f3361; add.ftz.f32 %f12536, %f3361, %f12534; div.approx.ftz.f32 %f12537, %f12535, %f12536; fma.rn.ftz.f32 %f12538, %f3361, %f12536, 0fBF800000; fma.rn.ftz.f32 %f12539, %f3361, %f12535, 0f3F800000; div.approx.ftz.f32 %f12540, %f12538, %f12539; mul.ftz.f32 %f12541, %f12537, 0f3F000000; mul.ftz.f32 %f12542, %f12537, %f12541; fma.rn.ftz.f32 %f12543, %f12540, %f12540, 0f3F800000; mul.ftz.f32 %f16549, %f12542, %f12543; $L__BB108_2081: mul.ftz.f32 %f12544, %f3232, 0f3EA2F983; abs.ftz.f32 %f3365, %f12544; add.ftz.f32 %f12545, %f16548, 0f3F800000; mul.ftz.f32 %f3366, %f3356, %f12545; @%p1822 bra $L__BB108_2083; bra.uni $L__BB108_2082; $L__BB108_2083: mul.ftz.f32 %f12605, %f3359, 0f3E800000; max.ftz.f32 %f12607, %f3366, %f12530; div.approx.ftz.f32 %f12608, %f12605, %f12607; mul.ftz.f32 %f12609, %f3348, 0fBFFE59F3; mul.ftz.f32 %f12610, %f3348, %f12609; mul.ftz.f32 %f12611, %f12610, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12612, %f12611; fma.rn.ftz.f32 %f12613, %f12612, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f12614, %f3348, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f12615, %f3348, %f12614, 0f417C8B44; fma.rn.ftz.f32 %f12616, %f3348, %f12615, 0fC1665E35; fma.rn.ftz.f32 %f12617, %f3348, %f12616, 0f40CE7881; fma.rn.ftz.f32 %f12618, %f3348, %f12617, 0fBF993E81; fma.rn.ftz.f32 %f12619, %f3348, %f12618, 0f3CE3C31E; fma.rn.ftz.f32 %f12620, %f3348, %f12619, 0f3BBA5D49; sub.ftz.f32 %f12621, %f12613, %f12620; cvt.ftz.sat.f32.f32 %f12622, %f12621; mov.f32 %f12623, 0f3F800000; sub.ftz.f32 %f12624, %f12623, %f12622; mul.ftz.f32 %f12625, %f3365, %f12624; fma.rn.ftz.f32 %f12626, %f12608, %f12622, %f12625; mul.ftz.f32 %f16743, %f16549, %f12626; bra.uni $L__BB108_2084; $L__BB108_2145: setp.neu.ftz.f32 %p1908, %f13470, %f13471; @%p1908 bra $L__BB108_2147; bra.uni $L__BB108_2146; $L__BB108_2147: ld.local.f32 %f13490, [%rd587+52]; mul.ftz.f32 %f13491, %f3575, %f13490; ld.local.f32 %f13492, [%rd587+48]; mul.ftz.f32 %f13493, %f3576, %f13492; sub.ftz.f32 %f13494, %f13491, %f13493; ld.local.f32 %f13495, [%rd587+44]; mul.ftz.f32 %f13496, %f3576, %f13495; mul.ftz.f32 %f13497, %f3573, %f13490; sub.ftz.f32 %f13498, %f13496, %f13497; mul.ftz.f32 %f13499, %f3573, %f13492; mul.ftz.f32 %f13500, %f3575, %f13495; sub.ftz.f32 %f13501, %f13499, %f13500; mul.ftz.f32 %f13502, %f13498, %f13498; fma.rn.ftz.f32 %f13503, %f13494, %f13494, %f13502; fma.rn.ftz.f32 %f13504, %f13501, %f13501, %f13503; sqrt.approx.ftz.f32 %f13505, %f13504; div.approx.ftz.f32 %f16589, %f13494, %f13505; div.approx.ftz.f32 %f16590, %f13498, %f13505; div.approx.ftz.f32 %f16591, %f13501, %f13505; mul.ftz.f32 %f13506, %f3576, %f16590; mul.ftz.f32 %f13507, %f3575, %f16591; sub.ftz.f32 %f16592, %f13506, %f13507; mul.ftz.f32 %f13508, %f3573, %f16591; mul.ftz.f32 %f13509, %f3576, %f16589; sub.ftz.f32 %f16593, %f13508, %f13509; mul.ftz.f32 %f13510, %f3575, %f16589; mul.ftz.f32 %f13511, %f3573, %f16590; sub.ftz.f32 %f16594, %f13510, %f13511; bra.uni $L__BB108_2148; $L__BB108_2086: neg.ftz.f32 %f16554, %f16554; neg.ftz.f32 %f16551, %f16551; neg.ftz.f32 %f16552, %f16552; neg.ftz.f32 %f16553, %f16553; mov.f32 %f16555, %f3376; $L__BB108_2088: mul.ftz.f32 %f12637, %f16554, %f16554; mov.f32 %f16559, 0f3F800000; sub.ftz.f32 %f12638, %f16559, %f12637; mul.ftz.f32 %f12639, %f16555, %f16555; mul.ftz.f32 %f12640, %f12639, %f12638; sub.ftz.f32 %f3387, %f16559, %f12640; setp.lt.ftz.f32 %p1833, %f3387, 0f00000000; mov.f32 %f16556, 0f00000000; mov.f32 %f16557, %f16556; mov.f32 %f16558, %f16556; @%p1833 bra $L__BB108_2090; sqrt.approx.ftz.f32 %f12641, %f3387; mov.f32 %f12642, 0f33D6BF95; max.ftz.f32 %f12643, %f12641, %f12642; mul.ftz.f32 %f12644, %f16554, %f16555; sub.ftz.f32 %f12645, %f12644, %f12643; mul.ftz.f32 %f12646, %f16551, %f12645; mul.ftz.f32 %f12647, %f16552, %f12645; mul.ftz.f32 %f12648, %f16553, %f12645; mul.ftz.f32 %f12649, %f3197, %f16555; sub.ftz.f32 %f16556, %f12646, %f12649; mul.ftz.f32 %f12650, %f12254, %f16555; sub.ftz.f32 %f16557, %f12647, %f12650; mul.ftz.f32 %f12651, %f12255, %f16555; sub.ftz.f32 %f16558, %f12648, %f12651; mul.ftz.f32 %f12652, %f16552, %f16557; fma.rn.ftz.f32 %f12653, %f16551, %f16556, %f12652; fma.rn.ftz.f32 %f12654, %f16553, %f16558, %f12653; mul.ftz.f32 %f12655, %f3376, %f12654; add.ftz.f32 %f12656, %f16554, %f12655; sub.ftz.f32 %f12657, %f16554, %f12655; div.approx.ftz.f32 %f12658, %f12656, %f12657; mul.ftz.f32 %f12659, %f3376, %f16554; add.ftz.f32 %f12660, %f12659, %f12654; sub.ftz.f32 %f12661, %f12659, %f12654; div.approx.ftz.f32 %f12662, %f12660, %f12661; mul.ftz.f32 %f12663, %f12662, %f12662; fma.rn.ftz.f32 %f12664, %f12658, %f12658, %f12663; mul.ftz.f32 %f16559, %f12664, 0f3F000000; $L__BB108_2090: add.ftz.f32 %f12669, %f16554, %f16554; setp.gt.ftz.f32 %p1834, %f16559, %f2799; mul.ftz.f32 %f12670, %f16551, %f12669; sub.ftz.f32 %f12671, %f12670, %f3197; selp.f32 %f16740, %f12671, %f16556, %p1834; mul.ftz.f32 %f12672, %f16552, %f12669; sub.ftz.f32 %f12673, %f12672, %f12254; selp.f32 %f16741, %f12673, %f16557, %p1834; mul.ftz.f32 %f12674, %f16553, %f12669; sub.ftz.f32 %f12675, %f12674, %f12255; selp.f32 %f16742, %f12675, %f16558, %p1834; selp.b32 %r3285, 18, 17, %p1834; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_2092: neg.ftz.f32 %f16563, %f16563; neg.ftz.f32 %f16560, %f16560; neg.ftz.f32 %f16561, %f16561; neg.ftz.f32 %f16562, %f16562; mov.u16 %rs157, 1; mov.f32 %f16564, %f3401; $L__BB108_2094: mul.ftz.f32 %f12684, %f16563, %f16563; mov.f32 %f16568, 0f3F800000; sub.ftz.f32 %f12685, %f16568, %f12684; mul.ftz.f32 %f12686, %f16564, %f16564; mul.ftz.f32 %f12687, %f12686, %f12685; sub.ftz.f32 %f3417, %f16568, %f12687; setp.lt.ftz.f32 %p1836, %f3417, 0f00000000; mov.f32 %f16740, 0f00000000; mov.f32 %f16741, %f16740; mov.f32 %f16742, %f16740; @%p1836 bra $L__BB108_2096; sqrt.approx.ftz.f32 %f12688, %f3417; mov.f32 %f12689, 0f33D6BF95; max.ftz.f32 %f12690, %f12688, %f12689; mul.ftz.f32 %f12691, %f16563, %f16564; sub.ftz.f32 %f12692, %f12691, %f12690; mul.ftz.f32 %f12693, %f16560, %f12692; mul.ftz.f32 %f12694, %f16561, %f12692; mul.ftz.f32 %f12695, %f16562, %f12692; mul.ftz.f32 %f12696, %f3403, %f16564; sub.ftz.f32 %f16740, %f12693, %f12696; mul.ftz.f32 %f12697, %f12676, %f16564; sub.ftz.f32 %f16741, %f12694, %f12697; mul.ftz.f32 %f12698, %f12677, %f16564; sub.ftz.f32 %f16742, %f12695, %f12698; mul.ftz.f32 %f12699, %f16561, %f16741; fma.rn.ftz.f32 %f12700, %f16560, %f16740, %f12699; fma.rn.ftz.f32 %f12701, %f16562, %f16742, %f12700; mul.ftz.f32 %f12702, %f3401, %f12701; add.ftz.f32 %f12703, %f16563, %f12702; sub.ftz.f32 %f12704, %f16563, %f12702; div.approx.ftz.f32 %f12705, %f12703, %f12704; mul.ftz.f32 %f12706, %f3401, %f16563; add.ftz.f32 %f12707, %f12706, %f12701; sub.ftz.f32 %f12708, %f12706, %f12701; div.approx.ftz.f32 %f12709, %f12707, %f12708; mul.ftz.f32 %f12710, %f12709, %f12709; fma.rn.ftz.f32 %f12711, %f12705, %f12705, %f12710; mul.ftz.f32 %f16568, %f12711, 0f3F000000; $L__BB108_2096: setp.neu.ftz.f32 %p1837, %f16568, 0f3F800000; setp.eq.s16 %p1838, %rs157, 0; and.pred %p1839, %p1838, %p1837; selp.f32 %f16743, 0f49742400, 0f00000000, %p1839; mov.u32 %r3285, 17; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1895: setp.eq.ftz.f32 %p1656, %f2880, 0f7F800000; setp.eq.ftz.f32 %p1657, %f2881, 0f7F800000; and.pred %p1658, %p1656, %p1657; @%p1658 bra $L__BB108_1897; bra.uni $L__BB108_1896; $L__BB108_1897: setp.lt.s32 %p1662, %r878, 0; selp.b32 %r2721, 1075235812, 1061752795, %p1662; or.b32 %r2722, %r879, %r2721; mov.b32 %f16472, %r2722; bra.uni $L__BB108_1899; $L__BB108_2284: neg.ftz.f32 %f16727, %f3944; neg.ftz.f32 %f16724, %f3941; neg.ftz.f32 %f16725, %f3942; neg.ftz.f32 %f16726, %f3943; mov.u16 %rs159, 1; mov.f32 %f16728, %f4009; $L__BB108_2286: mul.ftz.f32 %f14689, %f16727, %f16727; mov.f32 %f16732, 0f3F800000; sub.ftz.f32 %f14690, %f16732, %f14689; mul.ftz.f32 %f14691, %f16728, %f16728; mul.ftz.f32 %f14692, %f14691, %f14690; sub.ftz.f32 %f4020, %f16732, %f14692; setp.lt.ftz.f32 %p2016, %f4020, 0f00000000; mov.f32 %f16743, 0f00000000; mov.f32 %f16740, %f16743; mov.f32 %f16741, %f16743; mov.f32 %f16742, %f16743; @%p2016 bra $L__BB108_2288; sqrt.approx.ftz.f32 %f14693, %f4020; mov.f32 %f14694, 0f33D6BF95; max.ftz.f32 %f14695, %f14693, %f14694; mul.ftz.f32 %f14696, %f16727, %f16728; sub.ftz.f32 %f14697, %f14696, %f14695; mul.ftz.f32 %f14698, %f16724, %f14697; mul.ftz.f32 %f14699, %f16725, %f14697; mul.ftz.f32 %f14700, %f16726, %f14697; mul.ftz.f32 %f14701, %f3899, %f16728; sub.ftz.f32 %f16740, %f14698, %f14701; mul.ftz.f32 %f14702, %f14365, %f16728; sub.ftz.f32 %f16741, %f14699, %f14702; mul.ftz.f32 %f14703, %f14366, %f16728; sub.ftz.f32 %f16742, %f14700, %f14703; mul.ftz.f32 %f14704, %f16725, %f16741; fma.rn.ftz.f32 %f14705, %f16724, %f16740, %f14704; fma.rn.ftz.f32 %f14706, %f16726, %f16742, %f14705; mul.ftz.f32 %f14707, %f4009, %f14706; add.ftz.f32 %f14708, %f16727, %f14707; sub.ftz.f32 %f14709, %f16727, %f14707; div.approx.ftz.f32 %f14710, %f14708, %f14709; mul.ftz.f32 %f14711, %f4009, %f16727; add.ftz.f32 %f14712, %f14711, %f14706; sub.ftz.f32 %f14713, %f14711, %f14706; div.approx.ftz.f32 %f14714, %f14712, %f14713; mul.ftz.f32 %f14715, %f14714, %f14714; fma.rn.ftz.f32 %f14716, %f14710, %f14710, %f14715; mul.ftz.f32 %f16732, %f14716, 0f3F000000; $L__BB108_2288: setp.eq.ftz.f32 %p2017, %f16732, 0f3F800000; setp.ne.s16 %p2018, %rs159, 0; mov.u32 %r2880, 9; or.pred %p2019, %p2018, %p2017; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; mov.u32 %r3285, %r2880; @%p2019 bra $L__BB108_2309; mul.ftz.f32 %f4029, %f14367, %f14368; setp.le.ftz.f32 %p2020, %f4029, 0f33D6BF95; mov.u32 %r3285, 17; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2020 bra $L__BB108_2309; add.ftz.f32 %f14730, %f4009, 0fBF800000; abs.ftz.f32 %f14731, %f14730; setp.lt.ftz.f32 %p2021, %f14731, 0f38D1B717; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2021 bra $L__BB108_2309; mul.ftz.f32 %f14732, %f3940, %f3940; mul.ftz.f32 %f14733, %f14732, %f14732; rcp.approx.ftz.f32 %f14734, %f14732; mov.f32 %f14735, 0f3F800000; add.ftz.f32 %f14736, %f14734, 0fBF800000; mul.ftz.f32 %f14737, %f14733, 0f40490FDB; add.ftz.f32 %f14738, %f4029, %f14736; mul.ftz.f32 %f14739, %f14737, %f14738; mul.ftz.f32 %f14740, %f14738, %f14739; div.approx.ftz.f32 %f14741, %f4029, %f14740; mul.ftz.f32 %f14742, %f3900, %f16741; fma.rn.ftz.f32 %f14743, %f3898, %f16740, %f14742; fma.rn.ftz.f32 %f14744, %f3901, %f16742, %f14743; mul.ftz.f32 %f14745, %f14744, %f14744; sub.ftz.f32 %f14746, %f14735, %f14745; mul.ftz.f32 %f14747, %f4029, %f14746; div.approx.ftz.f32 %f14748, %f14747, %f14745; add.ftz.f32 %f14749, %f14748, 0f3F800000; mov.f32 %f14750, 0f00000000; max.ftz.f32 %f14751, %f14749, %f14750; sqrt.approx.ftz.f32 %f14752, %f14751; add.ftz.f32 %f14753, %f14752, 0f3F800000; mov.f32 %f14754, 0f40000000; div.approx.ftz.f32 %f14755, %f14754, %f14753; mul.ftz.f32 %f14756, %f3942, %f16741; fma.rn.ftz.f32 %f14757, %f3941, %f16740, %f14756; fma.rn.ftz.f32 %f14758, %f3943, %f16742, %f14757; fma.rn.ftz.f32 %f14759, %f4009, %f14758, %f3944; mul.ftz.f32 %f14760, %f14759, %f14759; mul.ftz.f32 %f14761, %f3937, %f14741; mul.ftz.f32 %f14762, %f4009, %f4009; mul.ftz.f32 %f14763, %f14762, %f14761; mul.ftz.f32 %f14764, %f3902, %f14760; div.approx.ftz.f32 %f14765, %f14763, %f14764; mul.ftz.f32 %f14766, %f3944, %f14758; abs.ftz.f32 %f14767, %f14766; mul.ftz.f32 %f14768, %f14755, %f14767; mul.ftz.f32 %f16744, %f14765, %f14768; abs.ftz.f32 %f14769, %f14758; mul.ftz.f32 %f14770, %f3944, %f14769; mul.ftz.f32 %f16743, %f14765, %f14770; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; mov.u32 %r3285, %r2880; bra.uni $L__BB108_2309; $L__BB108_2082: mul.ftz.f32 %f12546, %f3233, %f3352; fma.rn.ftz.f32 %f12547, %f3234, %f3351, %f12546; fma.rn.ftz.f32 %f12548, %f3232, %f3353, %f12547; mul.ftz.f32 %f12549, %f3360, %f12548; mul.ftz.f32 %f12550, %f12549, %f3359; mul.ftz.f32 %f12551, %f3212, %f12550; mul.ftz.f32 %f12552, %f3212, %f12551; mul.ftz.f32 %f12553, %f3350, %f3366; mul.ftz.f32 %f12554, %f3350, %f12553; max.ftz.f32 %f12556, %f12554, %f12530; div.approx.ftz.f32 %f12557, %f12552, %f12556; abs.ftz.f32 %f12558, %f12557; setp.lt.ftz.f32 %p1830, %f3212, 0f3F800000; mov.f32 %f12559, 0f3F800000; selp.f32 %f12560, %f3226, %f3212, %p1830; max.ftz.f32 %f12561, %f12560, %f12559; mov.f32 %f12562, 0f40400000; min.ftz.f32 %f12563, %f12561, %f12562; add.ftz.f32 %f12564, %f12563, 0fBF2814F9; mul.ftz.f32 %f12565, %f12564, 0fBF7A7547; mul.ftz.f32 %f12566, %f12564, %f12565; mul.ftz.f32 %f12567, %f12566, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12568, %f12567; mul.ftz.f32 %f12569, %f12563, 0fBD0A43BB; fma.rn.ftz.f32 %f12570, %f12568, 0f3D435661, %f12569; add.ftz.f32 %f12571, %f12570, 0f3F7E465A; cvt.ftz.sat.f32.f32 %f12572, %f3348; fma.rn.ftz.f32 %f12573, %f12572, 0f3DEF98FA, 0fBE8A6DCE; fma.rn.ftz.f32 %f12574, %f12572, %f12573, 0f3D4D5C09; fma.rn.ftz.f32 %f12575, %f12572, %f12574, 0fBB86D80B; fma.rn.ftz.f32 %f12576, %f12572, %f12575, 0f3F80029F; fma.rn.ftz.f32 %f12577, %f12563, 0fC0059210, 0f41D2A36E; fma.rn.ftz.f32 %f12578, %f12563, %f12577, 0fC2FFCFDF; fma.rn.ftz.f32 %f12579, %f12563, %f12578, 0f43927AA0; fma.rn.ftz.f32 %f12580, %f12563, %f12579, 0fC38FF917; mul.ftz.f32 %f12581, %f12563, %f12563; mov.f32 %f12582, 0f4347CD91; div.approx.ftz.f32 %f12583, %f12582, %f12581; add.ftz.f32 %f12584, %f12583, %f12580; mul.ftz.f32 %f12585, %f12563, %f12581; mov.f32 %f12586, 0f42CB5604; div.approx.ftz.f32 %f12587, %f12586, %f12585; sub.ftz.f32 %f12588, %f12584, %f12587; fma.rn.ftz.f32 %f12589, %f12572, 0f40ABEB85, 0fC1C77213; fma.rn.ftz.f32 %f12590, %f12572, %f12589, 0f41B5F319; fma.rn.ftz.f32 %f12591, %f12572, %f12590, 0fC05A14A5; fma.rn.ftz.f32 %f12592, %f12572, %f12591, 0f3DC9FFD6; fma.rn.ftz.f32 %f12593, %f12572, %f12592, 0f3BA1B61E; mul.ftz.f32 %f12594, %f12593, %f12588; fma.rn.ftz.f32 %f12595, %f12594, 0f3AFAB390, 0f3F800000; sub.ftz.f32 %f12596, %f12559, %f12571; sub.ftz.f32 %f12597, %f12559, %f12576; mul.ftz.f32 %f12598, %f12596, %f12597; fma.rn.ftz.f32 %f12599, %f12598, 0fC11520C5, %f12595; cvt.ftz.sat.f32.f32 %f12600, %f12599; sub.ftz.f32 %f12601, %f12559, %f12600; mul.ftz.f32 %f12602, %f3365, %f12601; fma.rn.ftz.f32 %f12603, %f12558, %f12600, %f12602; sub.ftz.f32 %f12604, %f12559, %f16549; mul.ftz.f32 %f16743, %f12604, %f12603; $L__BB108_2084: mul.ftz.f32 %f16744, %f16534, %f16743; mul.ftz.f32 %f16745, %f16535, %f16743; mul.ftz.f32 %f16746, %f16536, %f16743; mul.ftz.f32 %f12627, %f3204, %f3233; fma.rn.ftz.f32 %f12628, %f3201, %f3234, %f12627; mul.ftz.f32 %f12629, %f3205, %f3233; fma.rn.ftz.f32 %f12630, %f3202, %f3234, %f12629; mul.ftz.f32 %f12631, %f3206, %f3233; fma.rn.ftz.f32 %f12632, %f3203, %f3234, %f12631; fma.rn.ftz.f32 %f16740, %f16551, %f3232, %f12628; fma.rn.ftz.f32 %f16741, %f16552, %f3232, %f12630; fma.rn.ftz.f32 %f16742, %f16553, %f3232, %f12632; selp.b32 %r3285, 10, 9, %p1822; bra.uni $L__BB108_2309; $L__BB108_2199: ld.local.f32 %f13825, [%rd587+52]; mul.ftz.f32 %f13826, %f3738, %f13825; ld.local.f32 %f13827, [%rd587+48]; mul.ftz.f32 %f13828, %f3739, %f13827; sub.ftz.f32 %f13829, %f13826, %f13828; ld.local.f32 %f13830, [%rd587+44]; mul.ftz.f32 %f13831, %f3739, %f13830; mul.ftz.f32 %f13832, %f3736, %f13825; sub.ftz.f32 %f13833, %f13831, %f13832; mul.ftz.f32 %f13834, %f3736, %f13827; mul.ftz.f32 %f13835, %f3738, %f13830; sub.ftz.f32 %f13836, %f13834, %f13835; mul.ftz.f32 %f13837, %f13833, %f13833; fma.rn.ftz.f32 %f13838, %f13829, %f13829, %f13837; fma.rn.ftz.f32 %f13839, %f13836, %f13836, %f13838; sqrt.approx.ftz.f32 %f13840, %f13839; div.approx.ftz.f32 %f16626, %f13829, %f13840; div.approx.ftz.f32 %f16627, %f13833, %f13840; div.approx.ftz.f32 %f16628, %f13836, %f13840; mul.ftz.f32 %f13841, %f3739, %f16627; mul.ftz.f32 %f13842, %f3738, %f16628; sub.ftz.f32 %f16629, %f13841, %f13842; mul.ftz.f32 %f13843, %f3736, %f16628; mul.ftz.f32 %f13844, %f3739, %f16626; sub.ftz.f32 %f16630, %f13843, %f13844; mul.ftz.f32 %f13845, %f3738, %f16626; mul.ftz.f32 %f13846, %f3736, %f16627; sub.ftz.f32 %f16631, %f13845, %f13846; $L__BB108_2201: mul.ftz.f32 %f13868, %f13814, %f16630; fma.rn.ftz.f32 %f13869, %f3737, %f16629, %f13868; fma.rn.ftz.f32 %f13870, %f13815, %f16631, %f13869; mul.ftz.f32 %f13871, %f13814, %f16627; fma.rn.ftz.f32 %f13872, %f3737, %f16626, %f13871; fma.rn.ftz.f32 %f13873, %f13815, %f16628, %f13872; mul.ftz.f32 %f13874, %f13816, %f13870; mul.ftz.f32 %f13875, %f13816, %f13873; mul.ftz.f32 %f13876, %f13874, %f13874; fma.rn.ftz.f32 %f13877, %f13875, %f13875, %f13876; fma.rn.ftz.f32 %f13878, %f3740, %f3740, %f13877; sqrt.approx.ftz.f32 %f13879, %f13878; div.approx.ftz.f32 %f3759, %f13874, %f13879; div.approx.ftz.f32 %f3760, %f13875, %f13879; div.approx.ftz.f32 %f3761, %f3740, %f13879; setp.geu.ftz.f32 %p1961, %f3761, 0f3F7FFF58; mov.f32 %f16633, 0f00000000; mov.f32 %f16632, 0f3F800000; mov.f32 %f16634, %f16632; mov.f32 %f16635, %f16633; @%p1961 bra $L__BB108_2203; mul.ftz.f32 %f13880, %f3761, %f3761; mov.f32 %f13881, 0f3F800000; sub.ftz.f32 %f13882, %f13881, %f13880; mov.f32 %f13883, 0f00000000; max.ftz.f32 %f13884, %f13882, %f13883; sqrt.approx.ftz.f32 %f16633, %f13884; rsqrt.approx.ftz.f32 %f13885, %f13884; mul.ftz.f32 %f16634, %f3759, %f13885; mul.ftz.f32 %f16635, %f3760, %f13885; mov.f32 %f16632, %f3761; $L__BB108_2203: setp.ltu.ftz.f32 %p1962, %f16632, 0f3F7FFF58; @%p1962 bra $L__BB108_2205; bra.uni $L__BB108_2204; $L__BB108_2205: div.approx.ftz.f32 %f3771, %f16633, %f16632; rcp.approx.ftz.f32 %f3772, %f3771; abs.ftz.f32 %f3773, %f3772; setp.ge.ftz.f32 %p1963, %f3773, 0f4144CCCD; mov.u32 %r3283, 1065353216; @%p1963 bra $L__BB108_2207; add.ftz.f32 %f13893, %f3773, 0fBF800000; mov.f32 %f13894, 0f3F800000; add.ftz.f32 %f13895, %f13893, 0f3F800000; fma.rn.ftz.f32 %f13896, %f13895, 0f38349F67, 0f39910039; fma.rn.ftz.f32 %f13897, %f13895, %f13896, 0f391F6607; fma.rn.ftz.f32 %f13898, %f13895, %f13897, 0f3C17E369; fma.rn.ftz.f32 %f13899, %f13895, %f13898, 0f3D2D2FE7; fma.rn.ftz.f32 %f13900, %f13895, %f13899, 0f3D906E67; fma.rn.ftz.f32 %f13901, %f13895, %f13900, 0f3F800000; mul.ftz.f32 %f13902, %f13901, %f13901; mul.ftz.f32 %f13903, %f13902, %f13902; mul.ftz.f32 %f13904, %f13903, %f13903; mul.ftz.f32 %f13905, %f13904, %f13904; rcp.approx.ftz.f32 %f13906, %f13905; sub.ftz.f32 %f13907, %f13894, %f13906; mov.b32 %r2841, %f13907; and.b32 %r3283, %r2841, 2147483647; $L__BB108_2207: mov.b32 %r2842, %f3772; and.b32 %r2843, %r2842, -2147483648; or.b32 %r2844, %r3283, %r2843; mov.b32 %f13908, %r2844; mul.ftz.f32 %f13909, %f3772, %f3772; mul.ftz.f32 %f13910, %f13909, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13911, %f13910; add.ftz.f32 %f13912, %f13908, 0fBF800000; mov.f32 %f13913, 0f3F800000; mul.ftz.f32 %f13914, %f13912, 0f3F000000; mul.ftz.f32 %f13915, %f3771, %f13911; fma.rn.ftz.f32 %f3774, %f13915, 0f3E906EBB, %f13914; add.ftz.f32 %f13916, %f13908, 0f3F800000; mul.ftz.f32 %f13917, %f13908, %f13908; sub.ftz.f32 %f13918, %f13913, %f13917; mul.ftz.f32 %f3775, %f3771, 0f3F106EBB; fma.rn.ftz.f32 %f13919, %f3775, %f13918, %f13916; mul.ftz.f32 %f3776, %f2799, %f13919; fma.rn.ftz.f32 %f13920, %f3775, %f13911, %f13916; mul.ftz.f32 %f3777, %f2799, %f13920; setp.gt.ftz.f32 %p1964, %f3775, 0f00000000; @%p1964 bra $L__BB108_2209; bra.uni $L__BB108_2208; $L__BB108_2209: sub.ftz.f32 %f13921, %f3775, %f3776; add.ftz.f32 %f13922, %f13921, 0f3F800000; fma.rn.ftz.f32 %f13923, %f3775, %f13922, 0f3E800000; sqrt.approx.ftz.f32 %f13924, %f13923; mov.f32 %f13925, 0f3F000000; sub.ftz.f32 %f13926, %f13925, %f13924; div.approx.ftz.f32 %f16636, %f13926, %f3775; bra.uni $L__BB108_2210; $L__BB108_2204: lg2.approx.ftz.f32 %f13887, %f2799; mul.ftz.f32 %f13888, %f13887, 0fBF317218; sqrt.approx.ftz.f32 %f13889, %f13888; mul.ftz.f32 %f13890, %f2775, 0f40C90FDB; cos.approx.ftz.f32 %f13891, %f13890; mul.ftz.f32 %f16680, %f13889, %f13891; sin.approx.ftz.f32 %f13892, %f13890; mul.ftz.f32 %f16679, %f13889, %f13892; mov.f32 %f16678, 0f3F800000; bra.uni $L__BB108_2225; $L__BB108_1896: setp.lt.s32 %p1659, %r878, 0; min.ftz.f32 %f11560, %f2881, %f2880; max.ftz.f32 %f11561, %f2881, %f2880; div.full.ftz.f32 %f11562, %f11560, %f11561; mul.rn.ftz.f32 %f11563, %f11562, %f11562; mov.f32 %f11564, 0fC0B59883; mov.f32 %f11565, 0fBF52C7EA; fma.rn.ftz.f32 %f11566, %f11563, %f11565, %f11564; mov.f32 %f11567, 0fC0D21907; fma.rn.ftz.f32 %f11568, %f11566, %f11563, %f11567; mul.ftz.f32 %f11569, %f11563, %f11568; mul.ftz.f32 %f11570, %f11562, %f11569; add.ftz.f32 %f11571, %f11563, 0f41355DC0; mov.f32 %f11572, 0f41E6BD60; fma.rn.ftz.f32 %f11573, %f11571, %f11563, %f11572; mov.f32 %f11574, 0f419D92C8; fma.rn.ftz.f32 %f11575, %f11573, %f11563, %f11574; rcp.approx.ftz.f32 %f11576, %f11575; fma.rn.ftz.f32 %f11577, %f11570, %f11576, %f11562; mov.f32 %f11578, 0f3FC90FDB; sub.ftz.f32 %f11579, %f11578, %f11577; setp.gt.ftz.f32 %p1660, %f2881, %f2880; selp.f32 %f11580, %f11579, %f11577, %p1660; mov.f32 %f11581, 0f40490FDB; sub.ftz.f32 %f11582, %f11581, %f11580; selp.f32 %f11583, %f11582, %f11580, %p1659; mov.b32 %r2719, %f11583; or.b32 %r2720, %r879, %r2719; mov.b32 %f11584, %r2720; add.ftz.f32 %f11585, %f2880, %f2881; setp.le.ftz.f32 %p1661, %f11585, 0f7F800000; selp.f32 %f16472, %f11584, %f11585, %p1661; $L__BB108_1899: ld.local.f32 %f11587, [%rd587+44]; div.approx.ftz.f32 %f11588, %f2874, %f11587; mul.ftz.f32 %f11589, %f11588, %f11588; sub.ftz.f32 %f11590, %f11556, %f11589; max.ftz.f32 %f11592, %f11590, %f16477; sqrt.approx.ftz.f32 %f11593, %f11592; ld.f32 %f11594, [%rd588+12]; mul.ftz.f32 %f11595, %f11594, %f11594; sub.ftz.f32 %f11596, %f11556, %f11595; max.ftz.f32 %f11597, %f11596, %f16477; sqrt.approx.ftz.f32 %f11598, %f11597; mov.f32 %f11599, 0fBF800000; max.ftz.f32 %f11600, %f11594, %f11599; min.ftz.f32 %f11601, %f11600, %f11556; abs.ftz.f32 %f11602, %f11601; sub.ftz.f32 %f11603, %f11556, %f11602; mul.ftz.f32 %f11604, %f11603, 0f3F000000; sqrt.approx.ftz.f32 %f11605, %f11604; setp.gt.ftz.f32 %p1663, %f11602, 0f3F11EB85; selp.f32 %f11606, %f11605, %f11602, %p1663; mul.ftz.f32 %f11607, %f11606, %f11606; mov.f32 %f11608, 0f3C94D2E9; mov.f32 %f11609, 0f3D53F941; fma.rn.ftz.f32 %f11610, %f11609, %f11607, %f11608; mov.f32 %f11611, 0f3D3F841F; fma.rn.ftz.f32 %f11612, %f11610, %f11607, %f11611; mov.f32 %f11613, 0f3D994929; fma.rn.ftz.f32 %f11614, %f11612, %f11607, %f11613; mov.f32 %f11615, 0f3E2AAB94; fma.rn.ftz.f32 %f11616, %f11614, %f11607, %f11615; mul.ftz.f32 %f11617, %f11607, %f11616; fma.rn.ftz.f32 %f11618, %f11617, %f11606, %f11606; mov.f32 %f11619, 0f3FC90FDB; mov.f32 %f11620, 0fC0000000; fma.rn.ftz.f32 %f11621, %f11620, %f11618, %f11619; selp.f32 %f11622, %f11621, %f11618, %p1663; setp.le.ftz.f32 %p1664, %f11622, 0f7F800000; mov.b32 %r2726, %f11622; mov.b32 %r2727, %f11601; and.b32 %r2728, %r2727, -2147483648; or.b32 %r2729, %r2728, %r2726; mov.b32 %f11623, %r2729; selp.f32 %f2886, %f11623, %f11622, %p1664; mul.ftz.f32 %f11624, %f2879, %f11594; mul.ftz.f32 %f11625, %f11587, %f11587; sub.ftz.f32 %f11626, %f11625, %f2878; sqrt.approx.ftz.f32 %f11627, %f11626; div.approx.ftz.f32 %f11628, %f11624, %f11627; mul.ftz.f32 %f11629, %f11628, %f11628; sub.ftz.f32 %f11630, %f11556, %f11629; max.ftz.f32 %f11631, %f11630, %f16477; sqrt.approx.ftz.f32 %f11632, %f11631; max.ftz.f32 %f11633, %f11628, %f11599; min.ftz.f32 %f11634, %f11633, %f11556; abs.ftz.f32 %f11635, %f11634; sub.ftz.f32 %f11636, %f11556, %f11635; mul.ftz.f32 %f11637, %f11636, 0f3F000000; sqrt.approx.ftz.f32 %f11638, %f11637; setp.gt.ftz.f32 %p1665, %f11635, 0f3F11EB85; selp.f32 %f11639, %f11638, %f11635, %p1665; mul.ftz.f32 %f11640, %f11639, %f11639; fma.rn.ftz.f32 %f11641, %f11609, %f11640, %f11608; fma.rn.ftz.f32 %f11642, %f11641, %f11640, %f11611; fma.rn.ftz.f32 %f11643, %f11642, %f11640, %f11613; fma.rn.ftz.f32 %f11644, %f11643, %f11640, %f11615; mul.ftz.f32 %f11645, %f11640, %f11644; fma.rn.ftz.f32 %f11646, %f11645, %f11639, %f11639; fma.rn.ftz.f32 %f11647, %f11620, %f11646, %f11619; selp.f32 %f11648, %f11647, %f11646, %p1665; setp.le.ftz.f32 %p1666, %f11648, 0f7F800000; mov.b32 %r2730, %f11648; mov.b32 %r2731, %f11634; and.b32 %r2732, %r2731, -2147483648; or.b32 %r2733, %r2732, %r2730; mov.b32 %f11649, %r2733; selp.f32 %f2887, %f11649, %f11648, %p1666; add.ftz.f32 %f11650, %f11632, %f11632; div.approx.ftz.f32 %f11651, %f11650, %f11593; ld.local.f32 %f11652, [%rd587+20]; mul.ftz.f32 %f2888, %f11652, %f11651; ld.local.f32 %f11653, [%rd587+24]; mul.ftz.f32 %f2889, %f11653, %f11651; ld.local.f32 %f11654, [%rd587+28]; mul.ftz.f32 %f2890, %f11654, %f11651; mul.ftz.f32 %f11655, %f2879, %f11598; abs.ftz.f32 %f2891, %f11655; add.ftz.f32 %f11656, %f11625, 0fBF800000; fma.rn.ftz.f32 %f2892, %f2891, %f2891, %f11656; setp.leu.ftz.f32 %p1667, %f2892, 0f00000000; mov.f32 %f16473, %f11556; @%p1667 bra $L__BB108_1901; sqrt.approx.ftz.f32 %f11657, %f2892; sub.ftz.f32 %f11658, %f11657, %f2891; add.ftz.f32 %f11659, %f2891, %f11657; div.approx.ftz.f32 %f11660, %f11658, %f11659; fma.rn.ftz.f32 %f11661, %f2891, %f11659, 0fBF800000; fma.rn.ftz.f32 %f11662, %f2891, %f11658, 0f3F800000; div.approx.ftz.f32 %f11663, %f11661, %f11662; mul.ftz.f32 %f11664, %f11660, 0f3F000000; mul.ftz.f32 %f11665, %f11660, %f11664; fma.rn.ftz.f32 %f11666, %f11663, %f11663, 0f3F800000; mul.ftz.f32 %f16473, %f11665, %f11666; $L__BB108_1901: mul.ftz.f32 %f11668, %f2888, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11669, %f11668; mul.ftz.f32 %f11670, %f2889, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11671, %f11670; mul.ftz.f32 %f11672, %f2890, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11673, %f11672; sub.ftz.f32 %f11675, %f11556, %f16473; mul.ftz.f32 %f11676, %f11675, %f11675; mul.ftz.f32 %f2895, %f11669, %f11676; mul.ftz.f32 %f2896, %f11671, %f11676; mul.ftz.f32 %f2897, %f11673, %f11676; ld.const.v4.f32 {%f11677, %f11678, %f11679, %f11680}, [kernel_params+1184]; mul.ftz.f32 %f11681, %f2896, %f11678; fma.rn.ftz.f32 %f11682, %f2895, %f11677, %f11681; fma.rn.ftz.f32 %f2901, %f2897, %f11679, %f11682; mul.ftz.f32 %f2902, %f11669, %f16473; mul.ftz.f32 %f2903, %f2902, %f2895; mul.ftz.f32 %f2904, %f11671, %f16473; mul.ftz.f32 %f2905, %f2904, %f2896; mul.ftz.f32 %f2906, %f11673, %f16473; mul.ftz.f32 %f2907, %f2906, %f2897; mul.ftz.f32 %f11683, %f2905, %f11678; fma.rn.ftz.f32 %f11684, %f2903, %f11677, %f11683; fma.rn.ftz.f32 %f2908, %f2907, %f11679, %f11684; sub.ftz.f32 %f2909, %f11556, %f2902; sub.ftz.f32 %f2910, %f11556, %f2904; sub.ftz.f32 %f2911, %f11556, %f2906; setp.eq.ftz.f32 %p1668, %f2909, 0f00000000; mov.f32 %f16474, %f16477; @%p1668 bra $L__BB108_1903; div.approx.ftz.f32 %f16474, %f2902, %f2909; $L__BB108_1903: setp.eq.ftz.f32 %p1669, %f2910, 0f00000000; mov.f32 %f16475, %f16477; @%p1669 bra $L__BB108_1905; div.approx.ftz.f32 %f16475, %f2904, %f2910; $L__BB108_1905: setp.eq.ftz.f32 %p1670, %f2911, 0f00000000; mov.f32 %f16476, %f16477; @%p1670 bra $L__BB108_1907; div.approx.ftz.f32 %f16476, %f2906, %f2911; $L__BB108_1907: mul.ftz.f32 %f2918, %f2903, %f16474; mul.ftz.f32 %f2919, %f2905, %f16475; mul.ftz.f32 %f11688, %f2919, %f11678; fma.rn.ftz.f32 %f11689, %f2918, %f11677, %f11688; mul.ftz.f32 %f2920, %f2907, %f16476; fma.rn.ftz.f32 %f2921, %f2920, %f11679, %f11689; add.ftz.f32 %f11690, %f16473, %f2901; add.ftz.f32 %f11691, %f11690, %f2908; add.ftz.f32 %f2922, %f2921, %f11691; setp.eq.ftz.f32 %p1671, %f2922, 0f00000000; @%p1671 bra $L__BB108_1909; mov.f32 %f11692, 0f3F800000; div.approx.ftz.f32 %f16477, %f11692, %f2922; $L__BB108_1909: mul.ftz.f32 %f2925, %f16477, %f2901; mul.ftz.f32 %f2926, %f16477, %f2908; mul.ftz.f32 %f2927, %f16477, %f2921; mul.ftz.f32 %f2928, %f16477, %f16473; setp.lt.ftz.f32 %p1673, %f2799, %f2928; mov.pred %p2185, -1; mov.u32 %r3278, 0; @%p1673 bra $L__BB108_1913; sub.ftz.f32 %f2929, %f2799, %f2928; setp.lt.ftz.f32 %p1675, %f2929, %f2925; mov.u32 %r3278, 1; @%p1675 bra $L__BB108_1913; sub.ftz.f32 %f11693, %f2929, %f2925; setp.lt.ftz.f32 %p1677, %f11693, %f2926; mov.u32 %r3278, 2; @%p1677 bra $L__BB108_1913; mov.pred %p2185, 0; mov.u32 %r3278, 3; $L__BB108_1913: ld.local.f32 %f2930, [%rd587+32]; mul.ftz.f32 %f2931, %f2930, 0f3E800000; setp.eq.s32 %p1679, %r3278, 1; selp.f32 %f11694, %f2931, %f2930, %p1679; mul.ftz.f32 %f11695, %f11694, 0f40800000; setp.gt.u32 %p1680, %r3278, 1; selp.f32 %f11696, %f11695, %f11694, %p1680; cvt.rn.f32.u32 %f11697, %r876; mul.ftz.f32 %f11698, %f11697, 0f2F800000; mov.f32 %f11699, 0f3727C5AC; max.ftz.f32 %f11700, %f11698, %f11699; mov.f32 %f11701, 0f3F800000; sub.ftz.f32 %f11702, %f11701, %f11700; div.approx.ftz.f32 %f11704, %f11620, %f11696; mul.ftz.f32 %f11705, %f11704, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11706, %f11705; fma.rn.ftz.f32 %f11707, %f11702, %f11706, %f11700; lg2.approx.ftz.f32 %f11708, %f11707; mul.ftz.f32 %f11709, %f11708, 0f3F317218; fma.rn.ftz.f32 %f11710, %f11696, %f11709, 0f3F800000; mul.ftz.f32 %f11711, %f2874, %f11710; mul.ftz.f32 %f11712, %f11710, %f11710; sub.ftz.f32 %f11713, %f11701, %f11712; mov.f32 %f16490, 0f00000000; max.ftz.f32 %f11715, %f11713, %f16490; sqrt.approx.ftz.f32 %f11716, %f11715; cvt.rn.f32.u32 %f11717, %r877; mul.ftz.f32 %f11718, %f11717, 0f2F800000; mul.ftz.f32 %f11719, %f11718, 0f40C90FDB; cos.approx.ftz.f32 %f11720, %f11719; mul.ftz.f32 %f11721, %f11716, %f11720; mul.ftz.f32 %f11722, %f2879, %f11721; sub.ftz.f32 %f16478, %f11722, %f11711; mul.ftz.f32 %f11723, %f16478, %f16478; sub.ftz.f32 %f11724, %f11701, %f11723; max.ftz.f32 %f11725, %f11724, %f16490; sqrt.approx.ftz.f32 %f16479, %f11725; add.u64 %rd2614, %SPL, 704; add.s64 %rd589, %rd2614, 8; not.pred %p1681, %p2185; @%p1681 bra $L__BB108_1915; ld.local.f32 %f11726, [%rd587+40]; neg.ftz.f32 %f11727, %f11726; sin.approx.ftz.f32 %f11728, %f11727; mul.ftz.f32 %f11729, %f11728, %f11728; sub.ftz.f32 %f11731, %f11701, %f11729; mov.f32 %f11732, 0f00000000; max.ftz.f32 %f11733, %f11731, %f11732; sqrt.approx.ftz.f32 %f11734, %f11733; add.ftz.f32 %f11735, %f11728, %f11728; mul.ftz.f32 %f11736, %f11735, %f11734; mul.ftz.f32 %f11737, %f11734, %f11734; sub.ftz.f32 %f11738, %f11737, %f11729; add.ftz.f32 %f11739, %f11736, %f11736; mul.ftz.f32 %f11740, %f11739, %f11738; mul.ftz.f32 %f11741, %f11738, %f11738; mul.ftz.f32 %f11742, %f11736, %f11736; sub.ftz.f32 %f11743, %f11741, %f11742; mul.ftz.f32 %f11744, %f16478, %f11738; mul.ftz.f32 %f11745, %f16479, %f11738; mul.ftz.f32 %f11746, %f16478, %f11736; sub.ftz.f32 %f11747, %f11745, %f11746; fma.rn.ftz.f32 %f11748, %f16479, %f11736, %f11744; abs.ftz.f32 %f11749, %f11747; st.local.v2.f32 [%rd2614], {%f11748, %f11749}; mul.ftz.f32 %f11750, %f16478, %f11734; mul.ftz.f32 %f11751, %f16479, %f11728; mul.ftz.f32 %f11752, %f16479, %f11734; fma.rn.ftz.f32 %f11753, %f16478, %f11728, %f11752; sub.ftz.f32 %f11754, %f11750, %f11751; abs.ftz.f32 %f11755, %f11753; st.local.v2.f32 [%rd589], {%f11754, %f11755}; mul.ftz.f32 %f11756, %f16478, %f11743; mul.ftz.f32 %f11757, %f16479, %f11740; mul.ftz.f32 %f11758, %f16479, %f11743; fma.rn.ftz.f32 %f11759, %f16478, %f11740, %f11758; sub.ftz.f32 %f11760, %f11756, %f11757; abs.ftz.f32 %f11761, %f11759; st.local.v2.f32 [%rd589+8], {%f11760, %f11761}; shl.b32 %r2738, %r3278, 1; mul.wide.u32 %rd2617, %r2738, 4; add.s64 %rd2618, %rd2614, %rd2617; ld.local.v2.f32 {%f16478, %f16479}, [%rd2618]; $L__BB108_1915: add.ftz.f32 %f2938, %f2886, %f2886; @%p2185 bra $L__BB108_1917; bra.uni $L__BB108_1916; $L__BB108_1917: cvt.rn.f32.s32 %f11766, %r3278; add.ftz.f32 %f11767, %f11766, %f11766; mul.ftz.f32 %f11768, %f2887, %f11767; sub.ftz.f32 %f11769, %f11768, %f2938; mov.f32 %f11770, 0f40490FDB; fma.rn.ftz.f32 %f2942, %f11766, 0f40490FDB, %f11769; ld.local.f32 %f16482, [%rd587+36]; div.approx.ftz.f32 %f16481, %f11770, %f16482; setp.gt.ftz.f32 %p1682, %f16481, 0f42B00000; mov.f32 %f16480, 0f00000000; @%p1682 bra $L__BB108_1919; mul.ftz.f32 %f11771, %f16481, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11772, %f11771; add.ftz.f32 %f11773, %f11772, 0f3F800000; rcp.approx.ftz.f32 %f16480, %f11773; $L__BB108_1919: add.ftz.f32 %f11774, %f16480, %f16480; sub.ftz.f32 %f11776, %f11701, %f11774; fma.rn.ftz.f32 %f11777, %f2775, %f11776, %f16480; rcp.approx.ftz.f32 %f11778, %f11777; add.ftz.f32 %f11779, %f11778, 0fBF800000; lg2.approx.ftz.f32 %f11780, %f11779; mul.ftz.f32 %f11781, %f11780, 0fBF317218; mul.ftz.f32 %f11782, %f16482, %f11781; mov.f32 %f11783, 0fC0490FDB; max.ftz.f32 %f11784, %f11782, %f11783; min.ftz.f32 %f11786, %f11784, %f11770; add.ftz.f32 %f16483, %f2942, %f11786; bra.uni $L__BB108_1920; $L__BB108_1916: mul.ftz.f32 %f16483, %f2775, 0f40C90FDB; ld.local.f32 %f16482, [%rd587+36]; mov.f32 %f11764, 0f40490FDB; div.approx.ftz.f32 %f16481, %f11764, %f16482; $L__BB108_1920: ld.local.f32 %f11787, [%rd587+40]; sin.approx.ftz.f32 %f11788, %f11787; mul.ftz.f32 %f11789, %f11788, %f11788; sub.ftz.f32 %f11791, %f11701, %f11789; max.ftz.f32 %f11793, %f11791, %f16490; sqrt.approx.ftz.f32 %f11794, %f11793; add.ftz.f32 %f11795, %f11788, %f11788; mul.ftz.f32 %f11796, %f11795, %f11794; mul.ftz.f32 %f11797, %f11794, %f11794; sub.ftz.f32 %f11798, %f11797, %f11789; add.ftz.f32 %f11799, %f11796, %f11796; mul.ftz.f32 %f11800, %f11799, %f11798; mul.ftz.f32 %f11801, %f11798, %f11798; mul.ftz.f32 %f11802, %f11796, %f11796; sub.ftz.f32 %f11803, %f11801, %f11802; mul.ftz.f32 %f11804, %f16478, %f11798; mul.ftz.f32 %f11805, %f16479, %f11798; mul.ftz.f32 %f11806, %f16478, %f11796; sub.ftz.f32 %f11807, %f11805, %f11806; fma.rn.ftz.f32 %f11808, %f16479, %f11796, %f11804; abs.ftz.f32 %f11809, %f11807; st.local.v2.f32 [%rd2614], {%f11808, %f11809}; mul.ftz.f32 %f11810, %f16478, %f11794; mul.ftz.f32 %f11811, %f16479, %f11788; sub.ftz.f32 %f2951, %f11810, %f11811; mul.ftz.f32 %f11812, %f16479, %f11794; fma.rn.ftz.f32 %f11813, %f16478, %f11788, %f11812; abs.ftz.f32 %f2952, %f11813; st.local.v2.f32 [%rd589], {%f2951, %f2952}; mul.ftz.f32 %f11814, %f16478, %f11803; mul.ftz.f32 %f11815, %f16479, %f11800; sub.ftz.f32 %f2953, %f11814, %f11815; mul.ftz.f32 %f11816, %f16479, %f11803; fma.rn.ftz.f32 %f11817, %f16478, %f11800, %f11816; abs.ftz.f32 %f2954, %f11817; st.local.v2.f32 [%rd589+8], {%f2953, %f2954}; mul.ftz.f32 %f2955, %f16481, 0f3FB8AA3B; ld.local.f32 %f2956, [%rd587+48]; rcp.approx.ftz.f32 %f2957, %f2956; mul.ftz.f32 %f11818, %f2879, %f11809; mul.ftz.f32 %f2958, %f11818, %f2957; mul.ftz.f32 %f11819, %f2874, %f11808; mul.ftz.f32 %f2959, %f11819, %f2957; setp.gtu.ftz.f32 %p1683, %f2956, 0f3DCCCCCD; @%p1683 bra $L__BB108_1934; bra.uni $L__BB108_1921; $L__BB108_1934: mul.ftz.f32 %f2982, %f2958, %f2958; fma.rn.ftz.f32 %f16486, %f2982, 0f3E800000, 0f3F800000; mul.ftz.f32 %f2984, %f2982, %f2982; mov.u64 %rd2629, 64; cvt.rn.f32.u64 %f11850, %rd2629; div.approx.ftz.f32 %f11851, %f2984, %f11850; add.ftz.f32 %f2985, %f16486, %f11851; setp.eq.ftz.f32 %p1693, %f16486, %f2985; @%p1693 bra $L__BB108_1943; mul.ftz.f32 %f2986, %f2982, %f2984; mov.u64 %rd2630, 2304; cvt.rn.f32.u64 %f11852, %rd2630; div.approx.ftz.f32 %f11853, %f2986, %f11852; add.ftz.f32 %f2987, %f2985, %f11853; setp.eq.ftz.f32 %p1694, %f2985, %f2987; mov.f32 %f16486, %f2985; @%p1694 bra $L__BB108_1943; mul.ftz.f32 %f2988, %f2982, %f2986; mov.u64 %rd2631, 147456; cvt.rn.f32.u64 %f11854, %rd2631; div.approx.ftz.f32 %f11855, %f2988, %f11854; add.ftz.f32 %f2989, %f2987, %f11855; setp.eq.ftz.f32 %p1695, %f2987, %f2989; mov.f32 %f16486, %f2987; @%p1695 bra $L__BB108_1943; mul.ftz.f32 %f2990, %f2982, %f2988; mov.u64 %rd2632, 14745600; cvt.rn.f32.u64 %f11856, %rd2632; div.approx.ftz.f32 %f11857, %f2990, %f11856; add.ftz.f32 %f2991, %f2989, %f11857; setp.eq.ftz.f32 %p1696, %f2989, %f2991; mov.f32 %f16486, %f2989; @%p1696 bra $L__BB108_1943; mul.ftz.f32 %f2992, %f2982, %f2990; mov.u64 %rd2633, 2123366400; cvt.rn.f32.u64 %f11858, %rd2633; div.approx.ftz.f32 %f11859, %f2992, %f11858; add.ftz.f32 %f2993, %f2991, %f11859; setp.eq.ftz.f32 %p1697, %f2991, %f2993; mov.f32 %f16486, %f2991; @%p1697 bra $L__BB108_1943; mul.ftz.f32 %f2994, %f2982, %f2992; mov.u64 %rd2634, 416179814400; cvt.rn.f32.u64 %f11860, %rd2634; div.approx.ftz.f32 %f11861, %f2994, %f11860; add.ftz.f32 %f2995, %f2993, %f11861; setp.eq.ftz.f32 %p1698, %f2993, %f2995; mov.f32 %f16486, %f2993; @%p1698 bra $L__BB108_1943; mul.ftz.f32 %f2996, %f2982, %f2994; mov.u64 %rd2635, 106542032486400; cvt.rn.f32.u64 %f11862, %rd2635; div.approx.ftz.f32 %f11863, %f2996, %f11862; add.ftz.f32 %f2997, %f2995, %f11863; setp.eq.ftz.f32 %p1699, %f2995, %f2997; mov.f32 %f16486, %f2995; @%p1699 bra $L__BB108_1943; mul.ftz.f32 %f11864, %f2982, %f2996; mov.u64 %rd2636, 34519618525593600; cvt.rn.f32.u64 %f11865, %rd2636; div.approx.ftz.f32 %f11866, %f11864, %f11865; add.ftz.f32 %f2998, %f2997, %f11866; setp.eq.ftz.f32 %p1700, %f2997, %f2998; mov.f32 %f16486, %f2997; @%p1700 bra $L__BB108_1943; mov.f32 %f16486, %f2998; $L__BB108_1943: mul.ftz.f32 %f11867, %f2959, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11868, %f11867; mul.ftz.f32 %f3000, %f16486, %f11868; abs.ftz.f32 %f3001, %f2957; setp.ltu.ftz.f32 %p1701, %f3001, 0f3F800000; @%p1701 bra $L__BB108_1945; bra.uni $L__BB108_1944; $L__BB108_1945: mul.ftz.f32 %f11890, %f2957, %f2957; mov.f32 %f11891, 0f394FFF49; mov.f32 %f11892, 0f363D0ADA; fma.rn.ftz.f32 %f11893, %f11892, %f11890, %f11891; mov.f32 %f11894, 0f3C08889A; fma.rn.ftz.f32 %f11895, %f11893, %f11890, %f11894; mov.f32 %f11896, 0f3E2AAAAB; fma.rn.ftz.f32 %f11897, %f11895, %f11890, %f11896; mul.ftz.f32 %f11898, %f11890, %f11897; fma.rn.ftz.f32 %f16487, %f11898, %f2957, %f2957; bra.uni $L__BB108_1946; $L__BB108_1921: setp.gt.ftz.f32 %p1684, %f2958, 0f41400000; @%p1684 bra $L__BB108_1932; bra.uni $L__BB108_1922; $L__BB108_1932: mul.ftz.f32 %f11838, %f2958, 0f41000000; rcp.approx.ftz.f32 %f11839, %f11838; add.ftz.f32 %f11840, %f11839, 0fBFEB3F8E; lg2.approx.ftz.f32 %f11841, %f2958; fma.rn.ftz.f32 %f11842, %f11841, 0fBF317218, %f11840; fma.rn.ftz.f32 %f16485, %f11842, 0f3F000000, %f2958; bra.uni $L__BB108_1933; $L__BB108_2208: add.ftz.f32 %f16636, %f3776, 0fBF800000; $L__BB108_2210: abs.ftz.f32 %f13927, %f16636; setp.gt.ftz.f32 %p1965, %f13927, 0f3F7FFFFF; selp.f32 %f13928, 0f3F7FFFFF, %f13927, %p1965; sub.ftz.f32 %f13930, %f13913, %f13928; add.ftz.f32 %f13931, %f13928, 0f3F800000; mul.ftz.f32 %f13932, %f13930, %f13931; mov.f32 %f13933, 0f00800000; max.ftz.f32 %f13934, %f13932, %f13933; mov.f32 %f13935, 0f7F7FFFFF; min.ftz.f32 %f13936, %f13934, %f13935; mov.b32 %r2845, %f13936; shr.u32 %r2846, %r2845, 23; add.s32 %r2847, %r2846, -127; and.b32 %r2848, %r2845, 8388607; or.b32 %r2849, %r2848, 1065353216; mov.b32 %f13937, %r2849; add.ftz.f32 %f13938, %f13937, 0fBF800000; mul.ftz.f32 %f13939, %f13938, %f13938; mul.ftz.f32 %f13940, %f13939, %f13939; fma.rn.ftz.f32 %f13941, %f13938, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f13942, %f13938, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f13943, %f13938, %f13941, 0fBE0CD4FD; fma.rn.ftz.f32 %f13944, %f13938, %f13943, 0f3E77ADBD; fma.rn.ftz.f32 %f13945, %f13938, %f13944, 0fBEB1D206; fma.rn.ftz.f32 %f13946, %f13938, %f13942, 0f3FB8AA10; mul.ftz.f32 %f13947, %f13940, %f13945; fma.rn.ftz.f32 %f13948, %f13938, %f13946, %f13947; cvt.rn.f32.s32 %f13949, %r2847; add.ftz.f32 %f13950, %f13948, %f13949; mul.ftz.f32 %f3781, %f13950, 0f3F317218; setp.gt.ftz.f32 %p1966, %f3781, 0fC0A00000; @%p1966 bra $L__BB108_2212; bra.uni $L__BB108_2211; $L__BB108_2212: mov.f32 %f13971, 0fC0200000; sub.ftz.f32 %f16638, %f13971, %f3781; mov.f32 %f16646, 0f32F16588; mov.f32 %f16645, 0f34B84B36; mov.f32 %f16644, 0fB66C7357; mov.f32 %f16643, 0fB6935AC1; mov.f32 %f16642, 0f396532DB; mov.f32 %f16641, 0fBAA45408; mov.f32 %f16640, 0fBB88E4EF; mov.f32 %f16639, 0f3E7C8F63; mov.f32 %f16637, 0f3FC02E2F; bra.uni $L__BB108_2213; $L__BB108_1944: mov.f32 %f11869, 0f3FB8AA3B; mul.rn.ftz.f32 %f11870, %f3001, %f11869; cvt.rzi.ftz.f32.f32 %f11871, %f11870; abs.ftz.f32 %f11872, %f11871; setp.gt.ftz.f32 %p1702, %f11872, 0f42FC0000; mov.b32 %r2739, %f11871; and.b32 %r2740, %r2739, -2147483648; or.b32 %r2741, %r2740, 1123811328; mov.b32 %f11873, %r2741; selp.f32 %f11874, %f11873, %f11871, %p1702; mov.f32 %f11875, 0fBF317218; fma.rn.ftz.f32 %f11876, %f11874, %f11875, %f3001; mov.f32 %f11877, 0f3102E308; fma.rn.ftz.f32 %f11878, %f11874, %f11877, %f11876; mul.ftz.f32 %f11879, %f11878, 0f3FB8AA3B; add.ftz.f32 %f11880, %f11874, 0f4B40007D; mov.b32 %r2742, %f11880; shl.b32 %r2743, %r2742, 23; mov.b32 %f11881, %r2743; ex2.approx.ftz.f32 %f11882, %f11879; mul.ftz.f32 %f11883, %f11882, %f11881; mov.f32 %f11884, 0f3E000000; div.approx.ftz.f32 %f11885, %f11884, %f11883; neg.ftz.f32 %f11886, %f11885; mov.f32 %f11887, 0f40000000; fma.rn.ftz.f32 %f11888, %f11887, %f11883, %f11886; setp.ge.ftz.f32 %p1703, %f3001, 0f42B40000; selp.f32 %f11889, 0f7F800000, %f11888, %p1703; mov.b32 %r2744, %f11889; mov.b32 %r2745, %f2957; and.b32 %r2746, %r2745, -2147483648; or.b32 %r2747, %r2746, %r2744; mov.b32 %f16487, %r2747; $L__BB108_1946: add.ftz.f32 %f11899, %f16487, %f16487; mul.ftz.f32 %f11900, %f2956, %f11899; div.approx.ftz.f32 %f16488, %f3000, %f11900; bra.uni $L__BB108_1947; $L__BB108_1922: mul.ftz.f32 %f2960, %f2958, %f2958; fma.rn.ftz.f32 %f16484, %f2960, 0f3E800000, 0f3F800000; mul.ftz.f32 %f2962, %f2960, %f2960; mov.u64 %rd2621, 64; cvt.rn.f32.u64 %f11820, %rd2621; div.approx.ftz.f32 %f11821, %f2962, %f11820; add.ftz.f32 %f2963, %f16484, %f11821; setp.eq.ftz.f32 %p1685, %f16484, %f2963; @%p1685 bra $L__BB108_1931; mul.ftz.f32 %f2964, %f2960, %f2962; mov.u64 %rd2622, 2304; cvt.rn.f32.u64 %f11822, %rd2622; div.approx.ftz.f32 %f11823, %f2964, %f11822; add.ftz.f32 %f2965, %f2963, %f11823; setp.eq.ftz.f32 %p1686, %f2963, %f2965; mov.f32 %f16484, %f2963; @%p1686 bra $L__BB108_1931; mul.ftz.f32 %f2966, %f2960, %f2964; mov.u64 %rd2623, 147456; cvt.rn.f32.u64 %f11824, %rd2623; div.approx.ftz.f32 %f11825, %f2966, %f11824; add.ftz.f32 %f2967, %f2965, %f11825; setp.eq.ftz.f32 %p1687, %f2965, %f2967; mov.f32 %f16484, %f2965; @%p1687 bra $L__BB108_1931; mul.ftz.f32 %f2968, %f2960, %f2966; mov.u64 %rd2624, 14745600; cvt.rn.f32.u64 %f11826, %rd2624; div.approx.ftz.f32 %f11827, %f2968, %f11826; add.ftz.f32 %f2969, %f2967, %f11827; setp.eq.ftz.f32 %p1688, %f2967, %f2969; mov.f32 %f16484, %f2967; @%p1688 bra $L__BB108_1931; mul.ftz.f32 %f2970, %f2960, %f2968; mov.u64 %rd2625, 2123366400; cvt.rn.f32.u64 %f11828, %rd2625; div.approx.ftz.f32 %f11829, %f2970, %f11828; add.ftz.f32 %f2971, %f2969, %f11829; setp.eq.ftz.f32 %p1689, %f2969, %f2971; mov.f32 %f16484, %f2969; @%p1689 bra $L__BB108_1931; mul.ftz.f32 %f2972, %f2960, %f2970; mov.u64 %rd2626, 416179814400; cvt.rn.f32.u64 %f11830, %rd2626; div.approx.ftz.f32 %f11831, %f2972, %f11830; add.ftz.f32 %f2973, %f2971, %f11831; setp.eq.ftz.f32 %p1690, %f2971, %f2973; mov.f32 %f16484, %f2971; @%p1690 bra $L__BB108_1931; mul.ftz.f32 %f2974, %f2960, %f2972; mov.u64 %rd2627, 106542032486400; cvt.rn.f32.u64 %f11832, %rd2627; div.approx.ftz.f32 %f11833, %f2974, %f11832; add.ftz.f32 %f2975, %f2973, %f11833; setp.eq.ftz.f32 %p1691, %f2973, %f2975; mov.f32 %f16484, %f2973; @%p1691 bra $L__BB108_1931; mul.ftz.f32 %f11834, %f2960, %f2974; mov.u64 %rd2628, 34519618525593600; cvt.rn.f32.u64 %f11835, %rd2628; div.approx.ftz.f32 %f11836, %f11834, %f11835; add.ftz.f32 %f2976, %f2975, %f11836; setp.eq.ftz.f32 %p1692, %f2975, %f2976; mov.f32 %f16484, %f2975; @%p1692 bra $L__BB108_1931; mov.f32 %f16484, %f2976; $L__BB108_1931: lg2.approx.ftz.f32 %f11837, %f16484; mul.ftz.f32 %f16485, %f11837, 0f3F317218; $L__BB108_1933: sub.ftz.f32 %f11843, %f16485, %f2959; sub.ftz.f32 %f11844, %f11843, %f2957; add.ftz.f32 %f11845, %f11844, 0f3F316F00; mul.ftz.f32 %f11846, %f2957, 0f3F000000; lg2.approx.ftz.f32 %f11847, %f11846; fma.rn.ftz.f32 %f11848, %f11847, 0f3F317218, %f11845; mul.ftz.f32 %f11849, %f11848, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16488, %f11849; $L__BB108_1947: setp.gt.ftz.f32 %p1704, %f16481, 0f42B00000; mov.f32 %f16489, %f16490; @%p1704 bra $L__BB108_1949; ex2.approx.ftz.f32 %f11902, %f2955; add.ftz.f32 %f11903, %f11902, 0f3F800000; rcp.approx.ftz.f32 %f11904, %f11903; add.ftz.f32 %f16489, %f11904, %f11904; $L__BB108_1949: sub.ftz.f32 %f3009, %f11701, %f16489; setp.eq.ftz.f32 %p1705, %f3009, 0f00000000; @%p1705 bra $L__BB108_1951; mul.ftz.f32 %f11907, %f2887, 0f00000000; sub.ftz.f32 %f11908, %f11907, %f2938; add.ftz.f32 %f11909, %f11908, 0f00000000; sub.ftz.f32 %f11910, %f16483, %f11909; add.ftz.f32 %f11911, %f11910, 0f40490FDB; mov.f32 %f11912, 0f40C90FDB; div.approx.ftz.f32 %f11913, %f11911, %f11912; cvt.rmi.ftz.f32.f32 %f11914, %f11913; mul.ftz.f32 %f11915, %f11914, 0f40C90FDB; sub.ftz.f32 %f11916, %f11911, %f11915; add.ftz.f32 %f11917, %f11916, 0fC0490FDB; abs.ftz.f32 %f11918, %f11917; neg.ftz.f32 %f11919, %f11918; div.approx.ftz.f32 %f11920, %f11919, %f16482; mul.ftz.f32 %f11921, %f11920, 0f3FB8AA3B; ex2.approx.ftz.f32 %f11922, %f11921; add.ftz.f32 %f11923, %f11922, 0f3F800000; mul.ftz.f32 %f11924, %f11923, %f11923; mul.ftz.f32 %f11925, %f16482, %f11924; div.approx.ftz.f32 %f11926, %f11922, %f11925; div.approx.ftz.f32 %f16490, %f11926, %f3009; $L__BB108_1951: mul.ftz.f32 %f11927, %f16488, %f16473; fma.rn.ftz.f32 %f3012, %f16490, %f11927, 0f00000000; mul.ftz.f32 %f11928, %f16488, %f2928; fma.rn.ftz.f32 %f3013, %f16490, %f11928, 0f00000000; mul.ftz.f32 %f11929, %f2879, %f2952; rcp.approx.ftz.f32 %f3014, %f2931; mul.ftz.f32 %f3015, %f11929, %f3014; mul.ftz.f32 %f11930, %f2874, %f2951; mul.ftz.f32 %f3016, %f11930, %f3014; setp.gtu.ftz.f32 %p1706, %f2931, 0f3DCCCCCD; @%p1706 bra $L__BB108_1965; bra.uni $L__BB108_1952; $L__BB108_1965: mul.ftz.f32 %f3039, %f3015, %f3015; fma.rn.ftz.f32 %f16493, %f3039, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3041, %f3039, %f3039; mov.u64 %rd2645, 64; cvt.rn.f32.u64 %f11961, %rd2645; div.approx.ftz.f32 %f11962, %f3041, %f11961; add.ftz.f32 %f3042, %f16493, %f11962; setp.eq.ftz.f32 %p1716, %f16493, %f3042; @%p1716 bra $L__BB108_1974; mul.ftz.f32 %f3043, %f3039, %f3041; mov.u64 %rd2646, 2304; cvt.rn.f32.u64 %f11963, %rd2646; div.approx.ftz.f32 %f11964, %f3043, %f11963; add.ftz.f32 %f3044, %f3042, %f11964; setp.eq.ftz.f32 %p1717, %f3042, %f3044; mov.f32 %f16493, %f3042; @%p1717 bra $L__BB108_1974; mul.ftz.f32 %f3045, %f3039, %f3043; mov.u64 %rd2647, 147456; cvt.rn.f32.u64 %f11965, %rd2647; div.approx.ftz.f32 %f11966, %f3045, %f11965; add.ftz.f32 %f3046, %f3044, %f11966; setp.eq.ftz.f32 %p1718, %f3044, %f3046; mov.f32 %f16493, %f3044; @%p1718 bra $L__BB108_1974; mul.ftz.f32 %f3047, %f3039, %f3045; mov.u64 %rd2648, 14745600; cvt.rn.f32.u64 %f11967, %rd2648; div.approx.ftz.f32 %f11968, %f3047, %f11967; add.ftz.f32 %f3048, %f3046, %f11968; setp.eq.ftz.f32 %p1719, %f3046, %f3048; mov.f32 %f16493, %f3046; @%p1719 bra $L__BB108_1974; mul.ftz.f32 %f3049, %f3039, %f3047; mov.u64 %rd2649, 2123366400; cvt.rn.f32.u64 %f11969, %rd2649; div.approx.ftz.f32 %f11970, %f3049, %f11969; add.ftz.f32 %f3050, %f3048, %f11970; setp.eq.ftz.f32 %p1720, %f3048, %f3050; mov.f32 %f16493, %f3048; @%p1720 bra $L__BB108_1974; mul.ftz.f32 %f3051, %f3039, %f3049; mov.u64 %rd2650, 416179814400; cvt.rn.f32.u64 %f11971, %rd2650; div.approx.ftz.f32 %f11972, %f3051, %f11971; add.ftz.f32 %f3052, %f3050, %f11972; setp.eq.ftz.f32 %p1721, %f3050, %f3052; mov.f32 %f16493, %f3050; @%p1721 bra $L__BB108_1974; mul.ftz.f32 %f3053, %f3039, %f3051; mov.u64 %rd2651, 106542032486400; cvt.rn.f32.u64 %f11973, %rd2651; div.approx.ftz.f32 %f11974, %f3053, %f11973; add.ftz.f32 %f3054, %f3052, %f11974; setp.eq.ftz.f32 %p1722, %f3052, %f3054; mov.f32 %f16493, %f3052; @%p1722 bra $L__BB108_1974; mul.ftz.f32 %f11975, %f3039, %f3053; mov.u64 %rd2652, 34519618525593600; cvt.rn.f32.u64 %f11976, %rd2652; div.approx.ftz.f32 %f11977, %f11975, %f11976; add.ftz.f32 %f3055, %f3054, %f11977; setp.eq.ftz.f32 %p1723, %f3054, %f3055; mov.f32 %f16493, %f3054; @%p1723 bra $L__BB108_1974; mov.f32 %f16493, %f3055; $L__BB108_1974: mul.ftz.f32 %f11978, %f3016, 0fBFB8AA3B; ex2.approx.ftz.f32 %f11979, %f11978; mul.ftz.f32 %f3057, %f16493, %f11979; abs.ftz.f32 %f3058, %f3014; setp.ltu.ftz.f32 %p1724, %f3058, 0f3F800000; @%p1724 bra $L__BB108_1976; bra.uni $L__BB108_1975; $L__BB108_1976: mul.ftz.f32 %f12001, %f3014, %f3014; mov.f32 %f12002, 0f394FFF49; mov.f32 %f12003, 0f363D0ADA; fma.rn.ftz.f32 %f12004, %f12003, %f12001, %f12002; mov.f32 %f12005, 0f3C08889A; fma.rn.ftz.f32 %f12006, %f12004, %f12001, %f12005; mov.f32 %f12007, 0f3E2AAAAB; fma.rn.ftz.f32 %f12008, %f12006, %f12001, %f12007; mul.ftz.f32 %f12009, %f12001, %f12008; fma.rn.ftz.f32 %f16494, %f12009, %f3014, %f3014; bra.uni $L__BB108_1977; $L__BB108_1952: setp.gt.ftz.f32 %p1707, %f3015, 0f41400000; @%p1707 bra $L__BB108_1963; bra.uni $L__BB108_1953; $L__BB108_1963: mul.ftz.f32 %f11949, %f3015, 0f41000000; rcp.approx.ftz.f32 %f11950, %f11949; add.ftz.f32 %f11951, %f11950, 0fBFEB3F8E; lg2.approx.ftz.f32 %f11952, %f3015; fma.rn.ftz.f32 %f11953, %f11952, 0fBF317218, %f11951; fma.rn.ftz.f32 %f16492, %f11953, 0f3F000000, %f3015; bra.uni $L__BB108_1964; $L__BB108_2211: neg.ftz.f32 %f13960, %f3781; sqrt.approx.ftz.f32 %f13961, %f13960; add.ftz.f32 %f16638, %f13961, 0fC0400000; mov.f32 %f16646, 0fB951F09B; mov.f32 %f16645, 0f38D3B56B; mov.f32 %f16644, 0f3AB0DC72; mov.f32 %f16643, 0fBB70BDE7; mov.f32 %f16642, 0f3BBC127B; mov.f32 %f16641, 0fBBF9C5D7; mov.f32 %f16640, 0f3C1AA57E; mov.f32 %f16639, 0f3F8036DB; mov.f32 %f16637, 0f40354F7E; $L__BB108_2213: fma.rn.ftz.f32 %f13972, %f16638, %f16646, %f16645; fma.rn.ftz.f32 %f13973, %f16638, %f13972, %f16644; fma.rn.ftz.f32 %f13974, %f16638, %f13973, %f16643; fma.rn.ftz.f32 %f13975, %f16638, %f13974, %f16642; fma.rn.ftz.f32 %f13976, %f16638, %f13975, %f16641; fma.rn.ftz.f32 %f13977, %f16638, %f13976, %f16640; fma.rn.ftz.f32 %f13978, %f16638, %f13977, %f16639; fma.rn.ftz.f32 %f13979, %f16638, %f13978, %f16637; mul.ftz.f32 %f16680, %f16636, %f13979; mul.ftz.f32 %f13980, %f16680, %f16680; mul.ftz.f32 %f13981, %f13980, 0fBFB8AA3B; ex2.approx.ftz.f32 %f13982, %f13981; add.ftz.f32 %f13983, %f16636, 0f3F800000; fma.rn.ftz.f32 %f13984, %f3775, %f13982, %f13983; sub.ftz.f32 %f3795, %f13984, %f3777; abs.ftz.f32 %f13985, %f3795; setp.leu.ftz.f32 %p1967, %f13985, 0f358637BD; @%p1967 bra $L__BB108_2221; mul.ftz.f32 %f13986, %f3771, %f16680; sub.ftz.f32 %f13988, %f13913, %f13986; div.approx.ftz.f32 %f13989, %f3795, %f13988; sub.ftz.f32 %f3796, %f16636, %f13989; abs.ftz.f32 %f13990, %f3796; setp.gt.ftz.f32 %p1968, %f13990, 0f3F7FFFFF; selp.f32 %f13991, 0f3F7FFFFF, %f13990, %p1968; sub.ftz.f32 %f13992, %f13913, %f13991; add.ftz.f32 %f13993, %f13991, 0f3F800000; mul.ftz.f32 %f13994, %f13992, %f13993; max.ftz.f32 %f13996, %f13994, %f13933; min.ftz.f32 %f13998, %f13996, %f13935; mov.b32 %r2850, %f13998; shr.u32 %r2851, %r2850, 23; add.s32 %r2852, %r2851, -127; and.b32 %r2853, %r2850, 8388607; or.b32 %r2854, %r2853, 1065353216; mov.b32 %f13999, %r2854; add.ftz.f32 %f14000, %f13999, 0fBF800000; mul.ftz.f32 %f14001, %f14000, %f14000; mul.ftz.f32 %f14002, %f14001, %f14001; fma.rn.ftz.f32 %f14003, %f14000, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f14004, %f14000, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14005, %f14000, %f14003, 0fBE0CD4FD; fma.rn.ftz.f32 %f14006, %f14000, %f14005, 0f3E77ADBD; fma.rn.ftz.f32 %f14007, %f14000, %f14006, 0fBEB1D206; fma.rn.ftz.f32 %f14008, %f14000, %f14004, 0f3FB8AA10; mul.ftz.f32 %f14009, %f14002, %f14007; fma.rn.ftz.f32 %f14010, %f14000, %f14008, %f14009; cvt.rn.f32.s32 %f14011, %r2852; add.ftz.f32 %f14012, %f14010, %f14011; mul.ftz.f32 %f3797, %f14012, 0f3F317218; setp.gt.ftz.f32 %p1969, %f3797, 0fC0A00000; @%p1969 bra $L__BB108_2216; bra.uni $L__BB108_2215; $L__BB108_2216: mov.f32 %f14033, 0fC0200000; sub.ftz.f32 %f16648, %f14033, %f3797; mov.f32 %f16656, 0f32F16588; mov.f32 %f16655, 0f34B84B36; mov.f32 %f16654, 0fB66C7357; mov.f32 %f16653, 0fB6935AC1; mov.f32 %f16652, 0f396532DB; mov.f32 %f16651, 0fBAA45408; mov.f32 %f16650, 0fBB88E4EF; mov.f32 %f16649, 0f3E7C8F63; mov.f32 %f16647, 0f3FC02E2F; bra.uni $L__BB108_2217; $L__BB108_1975: mov.f32 %f11980, 0f3FB8AA3B; mul.rn.ftz.f32 %f11981, %f3058, %f11980; cvt.rzi.ftz.f32.f32 %f11982, %f11981; abs.ftz.f32 %f11983, %f11982; setp.gt.ftz.f32 %p1725, %f11983, 0f42FC0000; mov.b32 %r2748, %f11982; and.b32 %r2749, %r2748, -2147483648; or.b32 %r2750, %r2749, 1123811328; mov.b32 %f11984, %r2750; selp.f32 %f11985, %f11984, %f11982, %p1725; mov.f32 %f11986, 0fBF317218; fma.rn.ftz.f32 %f11987, %f11985, %f11986, %f3058; mov.f32 %f11988, 0f3102E308; fma.rn.ftz.f32 %f11989, %f11985, %f11988, %f11987; mul.ftz.f32 %f11990, %f11989, 0f3FB8AA3B; add.ftz.f32 %f11991, %f11985, 0f4B40007D; mov.b32 %r2751, %f11991; shl.b32 %r2752, %r2751, 23; mov.b32 %f11992, %r2752; ex2.approx.ftz.f32 %f11993, %f11990; mul.ftz.f32 %f11994, %f11993, %f11992; mov.f32 %f11995, 0f3E000000; div.approx.ftz.f32 %f11996, %f11995, %f11994; neg.ftz.f32 %f11997, %f11996; mov.f32 %f11998, 0f40000000; fma.rn.ftz.f32 %f11999, %f11998, %f11994, %f11997; setp.ge.ftz.f32 %p1726, %f3058, 0f42B40000; selp.f32 %f12000, 0f7F800000, %f11999, %p1726; mov.b32 %r2753, %f12000; mov.b32 %r2754, %f3014; and.b32 %r2755, %r2754, -2147483648; or.b32 %r2756, %r2755, %r2753; mov.b32 %f16494, %r2756; $L__BB108_1977: add.ftz.f32 %f12010, %f16494, %f16494; mul.ftz.f32 %f12011, %f2931, %f12010; div.approx.ftz.f32 %f16495, %f3057, %f12011; bra.uni $L__BB108_1978; $L__BB108_1953: mul.ftz.f32 %f3017, %f3015, %f3015; fma.rn.ftz.f32 %f16491, %f3017, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3019, %f3017, %f3017; mov.u64 %rd2637, 64; cvt.rn.f32.u64 %f11931, %rd2637; div.approx.ftz.f32 %f11932, %f3019, %f11931; add.ftz.f32 %f3020, %f16491, %f11932; setp.eq.ftz.f32 %p1708, %f16491, %f3020; @%p1708 bra $L__BB108_1962; mul.ftz.f32 %f3021, %f3017, %f3019; mov.u64 %rd2638, 2304; cvt.rn.f32.u64 %f11933, %rd2638; div.approx.ftz.f32 %f11934, %f3021, %f11933; add.ftz.f32 %f3022, %f3020, %f11934; setp.eq.ftz.f32 %p1709, %f3020, %f3022; mov.f32 %f16491, %f3020; @%p1709 bra $L__BB108_1962; mul.ftz.f32 %f3023, %f3017, %f3021; mov.u64 %rd2639, 147456; cvt.rn.f32.u64 %f11935, %rd2639; div.approx.ftz.f32 %f11936, %f3023, %f11935; add.ftz.f32 %f3024, %f3022, %f11936; setp.eq.ftz.f32 %p1710, %f3022, %f3024; mov.f32 %f16491, %f3022; @%p1710 bra $L__BB108_1962; mul.ftz.f32 %f3025, %f3017, %f3023; mov.u64 %rd2640, 14745600; cvt.rn.f32.u64 %f11937, %rd2640; div.approx.ftz.f32 %f11938, %f3025, %f11937; add.ftz.f32 %f3026, %f3024, %f11938; setp.eq.ftz.f32 %p1711, %f3024, %f3026; mov.f32 %f16491, %f3024; @%p1711 bra $L__BB108_1962; mul.ftz.f32 %f3027, %f3017, %f3025; mov.u64 %rd2641, 2123366400; cvt.rn.f32.u64 %f11939, %rd2641; div.approx.ftz.f32 %f11940, %f3027, %f11939; add.ftz.f32 %f3028, %f3026, %f11940; setp.eq.ftz.f32 %p1712, %f3026, %f3028; mov.f32 %f16491, %f3026; @%p1712 bra $L__BB108_1962; mul.ftz.f32 %f3029, %f3017, %f3027; mov.u64 %rd2642, 416179814400; cvt.rn.f32.u64 %f11941, %rd2642; div.approx.ftz.f32 %f11942, %f3029, %f11941; add.ftz.f32 %f3030, %f3028, %f11942; setp.eq.ftz.f32 %p1713, %f3028, %f3030; mov.f32 %f16491, %f3028; @%p1713 bra $L__BB108_1962; mul.ftz.f32 %f3031, %f3017, %f3029; mov.u64 %rd2643, 106542032486400; cvt.rn.f32.u64 %f11943, %rd2643; div.approx.ftz.f32 %f11944, %f3031, %f11943; add.ftz.f32 %f3032, %f3030, %f11944; setp.eq.ftz.f32 %p1714, %f3030, %f3032; mov.f32 %f16491, %f3030; @%p1714 bra $L__BB108_1962; mul.ftz.f32 %f11945, %f3017, %f3031; mov.u64 %rd2644, 34519618525593600; cvt.rn.f32.u64 %f11946, %rd2644; div.approx.ftz.f32 %f11947, %f11945, %f11946; add.ftz.f32 %f3033, %f3032, %f11947; setp.eq.ftz.f32 %p1715, %f3032, %f3033; mov.f32 %f16491, %f3032; @%p1715 bra $L__BB108_1962; mov.f32 %f16491, %f3033; $L__BB108_1962: lg2.approx.ftz.f32 %f11948, %f16491; mul.ftz.f32 %f16492, %f11948, 0f3F317218; $L__BB108_1964: sub.ftz.f32 %f11954, %f16492, %f3016; sub.ftz.f32 %f11955, %f11954, %f3014; add.ftz.f32 %f11956, %f11955, 0f3F316F00; mul.ftz.f32 %f11957, %f3014, 0f3F000000; lg2.approx.ftz.f32 %f11958, %f11957; fma.rn.ftz.f32 %f11959, %f11958, 0f3F317218, %f11956; mul.ftz.f32 %f11960, %f11959, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16495, %f11960; $L__BB108_1978: mov.f32 %f16497, 0f00000000; mov.f32 %f16496, %f16497; @%p1704 bra $L__BB108_1980; ex2.approx.ftz.f32 %f12013, %f2955; add.ftz.f32 %f12014, %f12013, 0f3F800000; rcp.approx.ftz.f32 %f12015, %f12014; add.ftz.f32 %f16496, %f12015, %f12015; $L__BB108_1980: sub.ftz.f32 %f3066, %f11701, %f16496; setp.eq.ftz.f32 %p1728, %f3066, 0f00000000; @%p1728 bra $L__BB108_1982; add.ftz.f32 %f12018, %f2887, %f2887; sub.ftz.f32 %f12019, %f12018, %f2938; add.ftz.f32 %f12020, %f12019, 0f40490FDB; sub.ftz.f32 %f12021, %f16483, %f12020; add.ftz.f32 %f12022, %f12021, 0f40490FDB; mov.f32 %f12023, 0f40C90FDB; div.approx.ftz.f32 %f12024, %f12022, %f12023; cvt.rmi.ftz.f32.f32 %f12025, %f12024; mul.ftz.f32 %f12026, %f12025, 0f40C90FDB; sub.ftz.f32 %f12027, %f12022, %f12026; add.ftz.f32 %f12028, %f12027, 0fC0490FDB; abs.ftz.f32 %f12029, %f12028; neg.ftz.f32 %f12030, %f12029; div.approx.ftz.f32 %f12031, %f12030, %f16482; mul.ftz.f32 %f12032, %f12031, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12033, %f12032; add.ftz.f32 %f12034, %f12033, 0f3F800000; mul.ftz.f32 %f12035, %f12034, %f12034; mul.ftz.f32 %f12036, %f16482, %f12035; div.approx.ftz.f32 %f12037, %f12033, %f12036; div.approx.ftz.f32 %f16497, %f12037, %f3066; $L__BB108_1982: mul.ftz.f32 %f12038, %f16495, %f2895; mul.ftz.f32 %f12039, %f16495, %f2896; mul.ftz.f32 %f12040, %f16495, %f2897; fma.rn.ftz.f32 %f3069, %f16497, %f12038, %f3012; fma.rn.ftz.f32 %f3070, %f16497, %f12039, %f3012; fma.rn.ftz.f32 %f3071, %f16497, %f12040, %f3012; mul.ftz.f32 %f12041, %f16495, %f2925; fma.rn.ftz.f32 %f3072, %f16497, %f12041, %f3013; mul.ftz.f32 %f3073, %f2930, 0f40800000; rcp.approx.ftz.f32 %f3074, %f3073; mul.ftz.f32 %f12042, %f2879, %f2954; mul.ftz.f32 %f3075, %f12042, %f3074; mul.ftz.f32 %f12043, %f2874, %f2953; mul.ftz.f32 %f3076, %f12043, %f3074; setp.gtu.ftz.f32 %p1729, %f3073, 0f3DCCCCCD; @%p1729 bra $L__BB108_1996; bra.uni $L__BB108_1983; $L__BB108_1996: mul.ftz.f32 %f3099, %f3075, %f3075; fma.rn.ftz.f32 %f16500, %f3099, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3101, %f3099, %f3099; mov.u64 %rd2661, 64; cvt.rn.f32.u64 %f12074, %rd2661; div.approx.ftz.f32 %f12075, %f3101, %f12074; add.ftz.f32 %f3102, %f16500, %f12075; setp.eq.ftz.f32 %p1739, %f16500, %f3102; @%p1739 bra $L__BB108_2005; mul.ftz.f32 %f3103, %f3099, %f3101; mov.u64 %rd2662, 2304; cvt.rn.f32.u64 %f12076, %rd2662; div.approx.ftz.f32 %f12077, %f3103, %f12076; add.ftz.f32 %f3104, %f3102, %f12077; setp.eq.ftz.f32 %p1740, %f3102, %f3104; mov.f32 %f16500, %f3102; @%p1740 bra $L__BB108_2005; mul.ftz.f32 %f3105, %f3099, %f3103; mov.u64 %rd2663, 147456; cvt.rn.f32.u64 %f12078, %rd2663; div.approx.ftz.f32 %f12079, %f3105, %f12078; add.ftz.f32 %f3106, %f3104, %f12079; setp.eq.ftz.f32 %p1741, %f3104, %f3106; mov.f32 %f16500, %f3104; @%p1741 bra $L__BB108_2005; mul.ftz.f32 %f3107, %f3099, %f3105; mov.u64 %rd2664, 14745600; cvt.rn.f32.u64 %f12080, %rd2664; div.approx.ftz.f32 %f12081, %f3107, %f12080; add.ftz.f32 %f3108, %f3106, %f12081; setp.eq.ftz.f32 %p1742, %f3106, %f3108; mov.f32 %f16500, %f3106; @%p1742 bra $L__BB108_2005; mul.ftz.f32 %f3109, %f3099, %f3107; mov.u64 %rd2665, 2123366400; cvt.rn.f32.u64 %f12082, %rd2665; div.approx.ftz.f32 %f12083, %f3109, %f12082; add.ftz.f32 %f3110, %f3108, %f12083; setp.eq.ftz.f32 %p1743, %f3108, %f3110; mov.f32 %f16500, %f3108; @%p1743 bra $L__BB108_2005; mul.ftz.f32 %f3111, %f3099, %f3109; mov.u64 %rd2666, 416179814400; cvt.rn.f32.u64 %f12084, %rd2666; div.approx.ftz.f32 %f12085, %f3111, %f12084; add.ftz.f32 %f3112, %f3110, %f12085; setp.eq.ftz.f32 %p1744, %f3110, %f3112; mov.f32 %f16500, %f3110; @%p1744 bra $L__BB108_2005; mul.ftz.f32 %f3113, %f3099, %f3111; mov.u64 %rd2667, 106542032486400; cvt.rn.f32.u64 %f12086, %rd2667; div.approx.ftz.f32 %f12087, %f3113, %f12086; add.ftz.f32 %f3114, %f3112, %f12087; setp.eq.ftz.f32 %p1745, %f3112, %f3114; mov.f32 %f16500, %f3112; @%p1745 bra $L__BB108_2005; mul.ftz.f32 %f12088, %f3099, %f3113; mov.u64 %rd2668, 34519618525593600; cvt.rn.f32.u64 %f12089, %rd2668; div.approx.ftz.f32 %f12090, %f12088, %f12089; add.ftz.f32 %f3115, %f3114, %f12090; setp.eq.ftz.f32 %p1746, %f3114, %f3115; mov.f32 %f16500, %f3114; @%p1746 bra $L__BB108_2005; mov.f32 %f16500, %f3115; $L__BB108_2005: mul.ftz.f32 %f12091, %f3076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f12092, %f12091; mul.ftz.f32 %f3117, %f16500, %f12092; abs.ftz.f32 %f3118, %f3074; setp.ltu.ftz.f32 %p1747, %f3118, 0f3F800000; @%p1747 bra $L__BB108_2007; bra.uni $L__BB108_2006; $L__BB108_2007: mul.ftz.f32 %f12114, %f3074, %f3074; mov.f32 %f12115, 0f394FFF49; mov.f32 %f12116, 0f363D0ADA; fma.rn.ftz.f32 %f12117, %f12116, %f12114, %f12115; mov.f32 %f12118, 0f3C08889A; fma.rn.ftz.f32 %f12119, %f12117, %f12114, %f12118; mov.f32 %f12120, 0f3E2AAAAB; fma.rn.ftz.f32 %f12121, %f12119, %f12114, %f12120; mul.ftz.f32 %f12122, %f12114, %f12121; fma.rn.ftz.f32 %f16501, %f12122, %f3074, %f3074; bra.uni $L__BB108_2008; $L__BB108_1983: setp.gt.ftz.f32 %p1730, %f3075, 0f41400000; @%p1730 bra $L__BB108_1994; bra.uni $L__BB108_1984; $L__BB108_1994: mul.ftz.f32 %f12062, %f3075, 0f41000000; rcp.approx.ftz.f32 %f12063, %f12062; add.ftz.f32 %f12064, %f12063, 0fBFEB3F8E; lg2.approx.ftz.f32 %f12065, %f3075; fma.rn.ftz.f32 %f12066, %f12065, 0fBF317218, %f12064; fma.rn.ftz.f32 %f16499, %f12066, 0f3F000000, %f3075; bra.uni $L__BB108_1995; $L__BB108_2006: mov.f32 %f12093, 0f3FB8AA3B; mul.rn.ftz.f32 %f12094, %f3118, %f12093; cvt.rzi.ftz.f32.f32 %f12095, %f12094; abs.ftz.f32 %f12096, %f12095; setp.gt.ftz.f32 %p1748, %f12096, 0f42FC0000; mov.b32 %r2757, %f12095; and.b32 %r2758, %r2757, -2147483648; or.b32 %r2759, %r2758, 1123811328; mov.b32 %f12097, %r2759; selp.f32 %f12098, %f12097, %f12095, %p1748; mov.f32 %f12099, 0fBF317218; fma.rn.ftz.f32 %f12100, %f12098, %f12099, %f3118; mov.f32 %f12101, 0f3102E308; fma.rn.ftz.f32 %f12102, %f12098, %f12101, %f12100; mul.ftz.f32 %f12103, %f12102, 0f3FB8AA3B; add.ftz.f32 %f12104, %f12098, 0f4B40007D; mov.b32 %r2760, %f12104; shl.b32 %r2761, %r2760, 23; mov.b32 %f12105, %r2761; ex2.approx.ftz.f32 %f12106, %f12103; mul.ftz.f32 %f12107, %f12106, %f12105; mov.f32 %f12108, 0f3E000000; div.approx.ftz.f32 %f12109, %f12108, %f12107; neg.ftz.f32 %f12110, %f12109; mov.f32 %f12111, 0f40000000; fma.rn.ftz.f32 %f12112, %f12111, %f12107, %f12110; setp.ge.ftz.f32 %p1749, %f3118, 0f42B40000; selp.f32 %f12113, 0f7F800000, %f12112, %p1749; mov.b32 %r2762, %f12113; mov.b32 %r2763, %f3074; and.b32 %r2764, %r2763, -2147483648; or.b32 %r2765, %r2764, %r2762; mov.b32 %f16501, %r2765; $L__BB108_2008: add.ftz.f32 %f12123, %f16501, %f16501; mul.ftz.f32 %f12124, %f3073, %f12123; div.approx.ftz.f32 %f16502, %f3117, %f12124; bra.uni $L__BB108_2009; $L__BB108_1984: mul.ftz.f32 %f3077, %f3075, %f3075; fma.rn.ftz.f32 %f16498, %f3077, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3079, %f3077, %f3077; mov.u64 %rd2653, 64; cvt.rn.f32.u64 %f12044, %rd2653; div.approx.ftz.f32 %f12045, %f3079, %f12044; add.ftz.f32 %f3080, %f16498, %f12045; setp.eq.ftz.f32 %p1731, %f16498, %f3080; @%p1731 bra $L__BB108_1993; mul.ftz.f32 %f3081, %f3077, %f3079; mov.u64 %rd2654, 2304; cvt.rn.f32.u64 %f12046, %rd2654; div.approx.ftz.f32 %f12047, %f3081, %f12046; add.ftz.f32 %f3082, %f3080, %f12047; setp.eq.ftz.f32 %p1732, %f3080, %f3082; mov.f32 %f16498, %f3080; @%p1732 bra $L__BB108_1993; mul.ftz.f32 %f3083, %f3077, %f3081; mov.u64 %rd2655, 147456; cvt.rn.f32.u64 %f12048, %rd2655; div.approx.ftz.f32 %f12049, %f3083, %f12048; add.ftz.f32 %f3084, %f3082, %f12049; setp.eq.ftz.f32 %p1733, %f3082, %f3084; mov.f32 %f16498, %f3082; @%p1733 bra $L__BB108_1993; mul.ftz.f32 %f3085, %f3077, %f3083; mov.u64 %rd2656, 14745600; cvt.rn.f32.u64 %f12050, %rd2656; div.approx.ftz.f32 %f12051, %f3085, %f12050; add.ftz.f32 %f3086, %f3084, %f12051; setp.eq.ftz.f32 %p1734, %f3084, %f3086; mov.f32 %f16498, %f3084; @%p1734 bra $L__BB108_1993; mul.ftz.f32 %f3087, %f3077, %f3085; mov.u64 %rd2657, 2123366400; cvt.rn.f32.u64 %f12052, %rd2657; div.approx.ftz.f32 %f12053, %f3087, %f12052; add.ftz.f32 %f3088, %f3086, %f12053; setp.eq.ftz.f32 %p1735, %f3086, %f3088; mov.f32 %f16498, %f3086; @%p1735 bra $L__BB108_1993; mul.ftz.f32 %f3089, %f3077, %f3087; mov.u64 %rd2658, 416179814400; cvt.rn.f32.u64 %f12054, %rd2658; div.approx.ftz.f32 %f12055, %f3089, %f12054; add.ftz.f32 %f3090, %f3088, %f12055; setp.eq.ftz.f32 %p1736, %f3088, %f3090; mov.f32 %f16498, %f3088; @%p1736 bra $L__BB108_1993; mul.ftz.f32 %f3091, %f3077, %f3089; mov.u64 %rd2659, 106542032486400; cvt.rn.f32.u64 %f12056, %rd2659; div.approx.ftz.f32 %f12057, %f3091, %f12056; add.ftz.f32 %f3092, %f3090, %f12057; setp.eq.ftz.f32 %p1737, %f3090, %f3092; mov.f32 %f16498, %f3090; @%p1737 bra $L__BB108_1993; mul.ftz.f32 %f12058, %f3077, %f3091; mov.u64 %rd2660, 34519618525593600; cvt.rn.f32.u64 %f12059, %rd2660; div.approx.ftz.f32 %f12060, %f12058, %f12059; add.ftz.f32 %f3093, %f3092, %f12060; setp.eq.ftz.f32 %p1738, %f3092, %f3093; mov.f32 %f16498, %f3092; @%p1738 bra $L__BB108_1993; mov.f32 %f16498, %f3093; $L__BB108_1993: lg2.approx.ftz.f32 %f12061, %f16498; mul.ftz.f32 %f16499, %f12061, 0f3F317218; $L__BB108_1995: sub.ftz.f32 %f12067, %f16499, %f3076; sub.ftz.f32 %f12068, %f12067, %f3074; add.ftz.f32 %f12069, %f12068, 0f3F316F00; mul.ftz.f32 %f12070, %f3074, 0f3F000000; lg2.approx.ftz.f32 %f12071, %f12070; fma.rn.ftz.f32 %f12072, %f12071, 0f3F317218, %f12069; mul.ftz.f32 %f12073, %f12072, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16502, %f12073; $L__BB108_2009: mul.ftz.f32 %f12126, %f2887, 0f40800000; sub.ftz.f32 %f3124, %f12126, %f2938; mov.f32 %f16504, 0f00000000; mov.f32 %f16503, %f16504; @%p1704 bra $L__BB108_2011; ex2.approx.ftz.f32 %f12127, %f2955; add.ftz.f32 %f12128, %f12127, 0f3F800000; rcp.approx.ftz.f32 %f12129, %f12128; add.ftz.f32 %f16503, %f12129, %f12129; $L__BB108_2011: add.ftz.f32 %f12131, %f3124, 0f40C90FDB; mov.f32 %f12132, 0f40C90FDB; sub.ftz.f32 %f12133, %f16483, %f12131; add.ftz.f32 %f12134, %f12133, 0f40490FDB; div.approx.ftz.f32 %f12135, %f12134, %f12132; cvt.rmi.ftz.f32.f32 %f12136, %f12135; mul.ftz.f32 %f12137, %f12136, 0f40C90FDB; sub.ftz.f32 %f12138, %f12134, %f12137; add.ftz.f32 %f12139, %f12138, 0fC0490FDB; abs.ftz.f32 %f12140, %f12139; neg.ftz.f32 %f12141, %f12140; div.approx.ftz.f32 %f12142, %f12141, %f16482; mul.ftz.f32 %f12143, %f12142, 0f3FB8AA3B; ex2.approx.ftz.f32 %f12144, %f12143; add.ftz.f32 %f12145, %f12144, 0f3F800000; mul.ftz.f32 %f12147, %f12145, %f12145; mul.ftz.f32 %f12148, %f16482, %f12147; div.approx.ftz.f32 %f3127, %f12144, %f12148; sub.ftz.f32 %f3128, %f11701, %f16503; setp.eq.ftz.f32 %p1751, %f3128, 0f00000000; @%p1751 bra $L__BB108_2013; div.approx.ftz.f32 %f16504, %f3127, %f3128; $L__BB108_2013: mul.ftz.f32 %f12149, %f16502, %f2903; mul.ftz.f32 %f12150, %f16502, %f2905; mul.ftz.f32 %f12151, %f16502, %f2907; fma.rn.ftz.f32 %f3131, %f16504, %f12149, %f3069; fma.rn.ftz.f32 %f3132, %f16504, %f12150, %f3070; fma.rn.ftz.f32 %f3133, %f16504, %f12151, %f3071; mul.ftz.f32 %f12152, %f16502, %f2926; fma.rn.ftz.f32 %f3134, %f16504, %f12152, %f3072; add.ftz.f32 %f3135, %f16472, %f16483; mul.ftz.f32 %f12153, %f2879, %f16479; mul.ftz.f32 %f3136, %f12153, %f3074; mul.ftz.f32 %f12154, %f2874, %f16478; mul.ftz.f32 %f3137, %f12154, %f3074; @%p1729 bra $L__BB108_2027; bra.uni $L__BB108_2014; $L__BB108_2027: mul.ftz.f32 %f3160, %f3136, %f3136; fma.rn.ftz.f32 %f16507, %f3160, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3162, %f3160, %f3160; mov.u64 %rd2677, 64; cvt.rn.f32.u64 %f12185, %rd2677; div.approx.ftz.f32 %f12186, %f3162, %f12185; add.ftz.f32 %f3163, %f16507, %f12186; setp.eq.ftz.f32 %p1762, %f16507, %f3163; @%p1762 bra $L__BB108_2036; mul.ftz.f32 %f3164, %f3160, %f3162; mov.u64 %rd2678, 2304; cvt.rn.f32.u64 %f12187, %rd2678; div.approx.ftz.f32 %f12188, %f3164, %f12187; add.ftz.f32 %f3165, %f3163, %f12188; setp.eq.ftz.f32 %p1763, %f3163, %f3165; mov.f32 %f16507, %f3163; @%p1763 bra $L__BB108_2036; mul.ftz.f32 %f3166, %f3160, %f3164; mov.u64 %rd2679, 147456; cvt.rn.f32.u64 %f12189, %rd2679; div.approx.ftz.f32 %f12190, %f3166, %f12189; add.ftz.f32 %f3167, %f3165, %f12190; setp.eq.ftz.f32 %p1764, %f3165, %f3167; mov.f32 %f16507, %f3165; @%p1764 bra $L__BB108_2036; mul.ftz.f32 %f3168, %f3160, %f3166; mov.u64 %rd2680, 14745600; cvt.rn.f32.u64 %f12191, %rd2680; div.approx.ftz.f32 %f12192, %f3168, %f12191; add.ftz.f32 %f3169, %f3167, %f12192; setp.eq.ftz.f32 %p1765, %f3167, %f3169; mov.f32 %f16507, %f3167; @%p1765 bra $L__BB108_2036; mul.ftz.f32 %f3170, %f3160, %f3168; mov.u64 %rd2681, 2123366400; cvt.rn.f32.u64 %f12193, %rd2681; div.approx.ftz.f32 %f12194, %f3170, %f12193; add.ftz.f32 %f3171, %f3169, %f12194; setp.eq.ftz.f32 %p1766, %f3169, %f3171; mov.f32 %f16507, %f3169; @%p1766 bra $L__BB108_2036; mul.ftz.f32 %f3172, %f3160, %f3170; mov.u64 %rd2682, 416179814400; cvt.rn.f32.u64 %f12195, %rd2682; div.approx.ftz.f32 %f12196, %f3172, %f12195; add.ftz.f32 %f3173, %f3171, %f12196; setp.eq.ftz.f32 %p1767, %f3171, %f3173; mov.f32 %f16507, %f3171; @%p1767 bra $L__BB108_2036; mul.ftz.f32 %f3174, %f3160, %f3172; mov.u64 %rd2683, 106542032486400; cvt.rn.f32.u64 %f12197, %rd2683; div.approx.ftz.f32 %f12198, %f3174, %f12197; add.ftz.f32 %f3175, %f3173, %f12198; setp.eq.ftz.f32 %p1768, %f3173, %f3175; mov.f32 %f16507, %f3173; @%p1768 bra $L__BB108_2036; mul.ftz.f32 %f12199, %f3160, %f3174; mov.u64 %rd2684, 34519618525593600; cvt.rn.f32.u64 %f12200, %rd2684; div.approx.ftz.f32 %f12201, %f12199, %f12200; add.ftz.f32 %f3176, %f3175, %f12201; setp.eq.ftz.f32 %p1769, %f3175, %f3176; mov.f32 %f16507, %f3175; @%p1769 bra $L__BB108_2036; mov.f32 %f16507, %f3176; $L__BB108_2036: mul.ftz.f32 %f12202, %f3137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f12203, %f12202; mul.ftz.f32 %f3178, %f16507, %f12203; abs.ftz.f32 %f3179, %f3074; setp.ltu.ftz.f32 %p1770, %f3179, 0f3F800000; @%p1770 bra $L__BB108_2038; bra.uni $L__BB108_2037; $L__BB108_2038: mul.ftz.f32 %f12225, %f3074, %f3074; mov.f32 %f12226, 0f394FFF49; mov.f32 %f12227, 0f363D0ADA; fma.rn.ftz.f32 %f12228, %f12227, %f12225, %f12226; mov.f32 %f12229, 0f3C08889A; fma.rn.ftz.f32 %f12230, %f12228, %f12225, %f12229; mov.f32 %f12231, 0f3E2AAAAB; fma.rn.ftz.f32 %f12232, %f12230, %f12225, %f12231; mul.ftz.f32 %f12233, %f12225, %f12232; fma.rn.ftz.f32 %f16508, %f12233, %f3074, %f3074; bra.uni $L__BB108_2039; $L__BB108_2014: setp.gt.ftz.f32 %p1753, %f3136, 0f41400000; @%p1753 bra $L__BB108_2025; bra.uni $L__BB108_2015; $L__BB108_2025: mul.ftz.f32 %f12173, %f3136, 0f41000000; rcp.approx.ftz.f32 %f12174, %f12173; add.ftz.f32 %f12175, %f12174, 0fBFEB3F8E; lg2.approx.ftz.f32 %f12176, %f3136; fma.rn.ftz.f32 %f12177, %f12176, 0fBF317218, %f12175; fma.rn.ftz.f32 %f16506, %f12177, 0f3F000000, %f3136; bra.uni $L__BB108_2026; $L__BB108_2037: mov.f32 %f12204, 0f3FB8AA3B; mul.rn.ftz.f32 %f12205, %f3179, %f12204; cvt.rzi.ftz.f32.f32 %f12206, %f12205; abs.ftz.f32 %f12207, %f12206; setp.gt.ftz.f32 %p1771, %f12207, 0f42FC0000; mov.b32 %r2766, %f12206; and.b32 %r2767, %r2766, -2147483648; or.b32 %r2768, %r2767, 1123811328; mov.b32 %f12208, %r2768; selp.f32 %f12209, %f12208, %f12206, %p1771; mov.f32 %f12210, 0fBF317218; fma.rn.ftz.f32 %f12211, %f12209, %f12210, %f3179; mov.f32 %f12212, 0f3102E308; fma.rn.ftz.f32 %f12213, %f12209, %f12212, %f12211; mul.ftz.f32 %f12214, %f12213, 0f3FB8AA3B; add.ftz.f32 %f12215, %f12209, 0f4B40007D; mov.b32 %r2769, %f12215; shl.b32 %r2770, %r2769, 23; mov.b32 %f12216, %r2770; ex2.approx.ftz.f32 %f12217, %f12214; mul.ftz.f32 %f12218, %f12217, %f12216; mov.f32 %f12219, 0f3E000000; div.approx.ftz.f32 %f12220, %f12219, %f12218; neg.ftz.f32 %f12221, %f12220; mov.f32 %f12222, 0f40000000; fma.rn.ftz.f32 %f12223, %f12222, %f12218, %f12221; setp.ge.ftz.f32 %p1772, %f3179, 0f42B40000; selp.f32 %f12224, 0f7F800000, %f12223, %p1772; mov.b32 %r2771, %f12224; mov.b32 %r2772, %f3074; and.b32 %r2773, %r2772, -2147483648; or.b32 %r2774, %r2773, %r2771; mov.b32 %f16508, %r2774; $L__BB108_2039: add.ftz.f32 %f12234, %f16508, %f16508; mul.ftz.f32 %f12235, %f3073, %f12234; div.approx.ftz.f32 %f16509, %f3178, %f12235; bra.uni $L__BB108_2040; $L__BB108_2015: mul.ftz.f32 %f3138, %f3136, %f3136; fma.rn.ftz.f32 %f16505, %f3138, 0f3E800000, 0f3F800000; mul.ftz.f32 %f3140, %f3138, %f3138; mov.u64 %rd2669, 64; cvt.rn.f32.u64 %f12155, %rd2669; div.approx.ftz.f32 %f12156, %f3140, %f12155; add.ftz.f32 %f3141, %f16505, %f12156; setp.eq.ftz.f32 %p1754, %f16505, %f3141; @%p1754 bra $L__BB108_2024; mul.ftz.f32 %f3142, %f3138, %f3140; mov.u64 %rd2670, 2304; cvt.rn.f32.u64 %f12157, %rd2670; div.approx.ftz.f32 %f12158, %f3142, %f12157; add.ftz.f32 %f3143, %f3141, %f12158; setp.eq.ftz.f32 %p1755, %f3141, %f3143; mov.f32 %f16505, %f3141; @%p1755 bra $L__BB108_2024; mul.ftz.f32 %f3144, %f3138, %f3142; mov.u64 %rd2671, 147456; cvt.rn.f32.u64 %f12159, %rd2671; div.approx.ftz.f32 %f12160, %f3144, %f12159; add.ftz.f32 %f3145, %f3143, %f12160; setp.eq.ftz.f32 %p1756, %f3143, %f3145; mov.f32 %f16505, %f3143; @%p1756 bra $L__BB108_2024; mul.ftz.f32 %f3146, %f3138, %f3144; mov.u64 %rd2672, 14745600; cvt.rn.f32.u64 %f12161, %rd2672; div.approx.ftz.f32 %f12162, %f3146, %f12161; add.ftz.f32 %f3147, %f3145, %f12162; setp.eq.ftz.f32 %p1757, %f3145, %f3147; mov.f32 %f16505, %f3145; @%p1757 bra $L__BB108_2024; mul.ftz.f32 %f3148, %f3138, %f3146; mov.u64 %rd2673, 2123366400; cvt.rn.f32.u64 %f12163, %rd2673; div.approx.ftz.f32 %f12164, %f3148, %f12163; add.ftz.f32 %f3149, %f3147, %f12164; setp.eq.ftz.f32 %p1758, %f3147, %f3149; mov.f32 %f16505, %f3147; @%p1758 bra $L__BB108_2024; mul.ftz.f32 %f3150, %f3138, %f3148; mov.u64 %rd2674, 416179814400; cvt.rn.f32.u64 %f12165, %rd2674; div.approx.ftz.f32 %f12166, %f3150, %f12165; add.ftz.f32 %f3151, %f3149, %f12166; setp.eq.ftz.f32 %p1759, %f3149, %f3151; mov.f32 %f16505, %f3149; @%p1759 bra $L__BB108_2024; mul.ftz.f32 %f3152, %f3138, %f3150; mov.u64 %rd2675, 106542032486400; cvt.rn.f32.u64 %f12167, %rd2675; div.approx.ftz.f32 %f12168, %f3152, %f12167; add.ftz.f32 %f3153, %f3151, %f12168; setp.eq.ftz.f32 %p1760, %f3151, %f3153; mov.f32 %f16505, %f3151; @%p1760 bra $L__BB108_2024; mul.ftz.f32 %f12169, %f3138, %f3152; mov.u64 %rd2676, 34519618525593600; cvt.rn.f32.u64 %f12170, %rd2676; div.approx.ftz.f32 %f12171, %f12169, %f12170; add.ftz.f32 %f3154, %f3153, %f12171; setp.eq.ftz.f32 %p1761, %f3153, %f3154; mov.f32 %f16505, %f3153; @%p1761 bra $L__BB108_2024; mov.f32 %f16505, %f3154; $L__BB108_2024: lg2.approx.ftz.f32 %f12172, %f16505; mul.ftz.f32 %f16506, %f12172, 0f3F317218; $L__BB108_2026: sub.ftz.f32 %f12178, %f16506, %f3137; sub.ftz.f32 %f12179, %f12178, %f3074; add.ftz.f32 %f12180, %f12179, 0f3F316F00; mul.ftz.f32 %f12181, %f3074, 0f3F000000; lg2.approx.ftz.f32 %f12182, %f12181; fma.rn.ftz.f32 %f12183, %f12182, 0f3F317218, %f12180; mul.ftz.f32 %f12184, %f12183, 0f3FB8AA3B; ex2.approx.ftz.f32 %f16509, %f12184; $L__BB108_2040: mul.ftz.f32 %f12236, %f16509, %f2918; mul.ftz.f32 %f12237, %f16509, %f2919; mul.ftz.f32 %f12238, %f16509, %f2920; fma.rn.ftz.f32 %f16744, %f12236, 0f3E22F983, %f3131; fma.rn.ftz.f32 %f16745, %f12237, 0f3E22F983, %f3132; fma.rn.ftz.f32 %f16746, %f12238, 0f3E22F983, %f3133; mul.ftz.f32 %f12239, %f16509, %f2927; fma.rn.ftz.f32 %f16743, %f12239, 0f3E22F983, %f3134; cos.approx.ftz.f32 %f12240, %f3135; mul.ftz.f32 %f12241, %f11523, %f16479; mul.ftz.f32 %f12242, %f12241, %f12240; mul.ftz.f32 %f12243, %f11524, %f16479; mul.ftz.f32 %f12244, %f12243, %f12240; mul.ftz.f32 %f12245, %f11525, %f16479; mul.ftz.f32 %f12246, %f12245, %f12240; fma.rn.ftz.f32 %f12247, %f16466, %f16478, %f12242; fma.rn.ftz.f32 %f12248, %f16467, %f16478, %f12244; fma.rn.ftz.f32 %f12249, %f16468, %f16478, %f12246; sin.approx.ftz.f32 %f12250, %f3135; mul.ftz.f32 %f12251, %f16469, %f16479; mul.ftz.f32 %f12252, %f16470, %f16479; mul.ftz.f32 %f12253, %f16471, %f16479; fma.rn.ftz.f32 %f16740, %f12251, %f12250, %f12247; fma.rn.ftz.f32 %f16741, %f12252, %f12250, %f12248; fma.rn.ftz.f32 %f16742, %f12253, %f12250, %f12249; setp.eq.s32 %p1773, %r3278, 0; selp.b32 %r3285, 10, 9, %p1773; bra.uni $L__BB108_2309; $L__BB108_1648: ld.const.u64 %rd3024, [%rd443+-24]; $L__BB108_1649: add.s32 %r3242, %r3242, 1; mul.wide.s32 %rd2184, %r3242, 16; add.s64 %rd2186, %rd2170, %rd2184; ld.const.v2.u64 {%rd2187, %rd2188}, [%rd2186+3744]; cvta.to.global.u64 %rd2189, %rd2187; add.s64 %rd2191, %rd2189, %rd2167; ld.global.u32 %r2286, [%rd2191]; cvta.to.global.u64 %rd2192, %rd2188; add.s64 %rd2193, %rd2192, %rd2167; ld.global.u32 %r2287, [%rd2193]; cvta.to.global.u64 %rd2194, %rd3024; add.s64 %rd2195, %rd2194, %rd2167; st.global.u32 [%rd2195], %r2286; cvta.to.global.u64 %rd2196, %rd3021; add.s64 %rd2197, %rd2196, %rd2167; st.global.u32 [%rd2197], %r2287; setp.eq.s32 %p1469, %r2287, -1; mov.u64 %rd3021, %rd2188; mov.u64 %rd3024, %rd2187; @%p1469 bra $L__BB108_1656; bra.uni $L__BB108_1649; $L__BB108_2115: ld.local.f32 %f13150, [%rd587+52]; mul.ftz.f32 %f13151, %f3486, %f13150; ld.local.f32 %f13152, [%rd587+48]; mul.ftz.f32 %f13153, %f3487, %f13152; sub.ftz.f32 %f13154, %f13151, %f13153; ld.local.f32 %f13155, [%rd587+44]; mul.ftz.f32 %f13156, %f3487, %f13155; mul.ftz.f32 %f13157, %f3484, %f13150; sub.ftz.f32 %f13158, %f13156, %f13157; mul.ftz.f32 %f13159, %f3484, %f13152; mul.ftz.f32 %f13160, %f3486, %f13155; sub.ftz.f32 %f13161, %f13159, %f13160; mul.ftz.f32 %f13162, %f13158, %f13158; fma.rn.ftz.f32 %f13163, %f13154, %f13154, %f13162; fma.rn.ftz.f32 %f13164, %f13161, %f13161, %f13163; sqrt.approx.ftz.f32 %f13165, %f13164; div.approx.ftz.f32 %f16572, %f13154, %f13165; div.approx.ftz.f32 %f16573, %f13158, %f13165; div.approx.ftz.f32 %f16574, %f13161, %f13165; mul.ftz.f32 %f13166, %f3487, %f16573; mul.ftz.f32 %f13167, %f3486, %f16574; sub.ftz.f32 %f16575, %f13166, %f13167; mul.ftz.f32 %f13168, %f3484, %f16574; mul.ftz.f32 %f13169, %f3487, %f16572; sub.ftz.f32 %f16576, %f13168, %f13169; mul.ftz.f32 %f13170, %f3486, %f16572; mul.ftz.f32 %f13171, %f3484, %f16573; sub.ftz.f32 %f16577, %f13170, %f13171; $L__BB108_2117: add.ftz.f32 %f3511, %f3491, 0f3F800000; @%p1881 bra $L__BB108_2133; bra.uni $L__BB108_2118; $L__BB108_2133: mul.ftz.f32 %f16583, %f2799, 0f40C90FDB; lg2.approx.ftz.f32 %f13337, %f2775; rcp.approx.ftz.f32 %f13338, %f3511; mul.ftz.f32 %f13339, %f13338, %f13337; ex2.approx.ftz.f32 %f16582, %f13339; bra.uni $L__BB108_2134; $L__BB108_2146: setp.neu.ftz.f32 %p1909, %f3573, %f3575; setp.neu.ftz.f32 %p1910, %f3573, %f3576; or.pred %p1911, %p1909, %p1910; neg.ftz.f32 %f13473, %f3576; selp.f32 %f13474, %f13473, %f3576, %p1911; add.ftz.f32 %f13475, %f3573, %f13474; neg.ftz.f32 %f13476, %f3575; selp.f32 %f13477, %f3575, %f13476, %p1911; sub.ftz.f32 %f13478, %f13477, %f3573; sub.ftz.f32 %f13479, %f3576, %f3575; mul.ftz.f32 %f13480, %f13475, %f13475; fma.rn.ftz.f32 %f13481, %f13479, %f13479, %f13480; fma.rn.ftz.f32 %f13482, %f13478, %f13478, %f13481; sqrt.approx.ftz.f32 %f13483, %f13482; div.approx.ftz.f32 %f16592, %f13479, %f13483; div.approx.ftz.f32 %f16593, %f13475, %f13483; div.approx.ftz.f32 %f16594, %f13478, %f13483; mul.ftz.f32 %f13484, %f3575, %f16594; mul.ftz.f32 %f13485, %f3576, %f16593; sub.ftz.f32 %f16589, %f13484, %f13485; mul.ftz.f32 %f13486, %f3576, %f16592; mul.ftz.f32 %f13487, %f3573, %f16594; sub.ftz.f32 %f16590, %f13486, %f13487; mul.ftz.f32 %f13488, %f3573, %f16593; mul.ftz.f32 %f13489, %f3575, %f16592; sub.ftz.f32 %f16591, %f13488, %f13489; $L__BB108_2148: mul.ftz.f32 %f13513, %f13466, %f16593; fma.rn.ftz.f32 %f13514, %f3574, %f16592, %f13513; fma.rn.ftz.f32 %f3601, %f13467, %f16594, %f13514; mul.ftz.f32 %f13515, %f13466, %f16590; fma.rn.ftz.f32 %f13516, %f3574, %f16589, %f13515; fma.rn.ftz.f32 %f3602, %f13467, %f16591, %f13516; ld.local.u64 %rd2686, [%rd587+36]; ld.f32 %f3603, [%rd2686]; ld.f32 %f3604, [%rd2686+4]; ld.f32 %f3605, [%rd2686+8]; ld.local.f32 %f3606, [%rd587+28]; ld.f32 %f3607, [%rd2686+12]; ld.f32 %f3608, [%rd2686+16]; ld.f32 %f3609, [%rd2686+20]; neg.ftz.f32 %f16602, %f3601; neg.ftz.f32 %f16601, %f3602; setp.lt.ftz.f32 %p1912, %f3577, 0fBF7FF972; mov.f32 %f16615, 0f00000000; @%p1912 bra $L__BB108_2151; setp.gt.ftz.f32 %p1913, %f3577, 0f3F7FF972; mov.f32 %f16615, 0fBF7FF972; @%p1913 bra $L__BB108_2151; mul.ftz.f32 %f13518, %f3577, %f3577; mov.f32 %f13519, 0f33D6BF95; max.ftz.f32 %f13520, %f13518, %f13519; rcp.approx.ftz.f32 %f13521, %f13520; mul.ftz.f32 %f13522, %f13470, %f16602; mul.ftz.f32 %f13523, %f13471, %f16601; mul.ftz.f32 %f13524, %f13523, %f13523; fma.rn.ftz.f32 %f13525, %f13522, %f13522, %f13524; fma.rn.ftz.f32 %f13526, %f13525, %f13521, 0f3F800000; sqrt.approx.ftz.f32 %f13527, %f13526; neg.ftz.f32 %f13528, %f13527; setp.ge.ftz.f32 %p1914, %f3577, 0f80000000; selp.f32 %f13529, %f13528, %f13527, %p1914; add.ftz.f32 %f13530, %f13529, 0fBF800000; mul.ftz.f32 %f16615, %f13530, 0f3F000000; $L__BB108_2151: fma.rn.ftz.f32 %f3614, %f3606, %f3606, 0fBF800000; mov.f32 %f16746, 0f3F800000; abs.ftz.f32 %f3615, %f16746; fma.rn.ftz.f32 %f3616, %f3615, %f3615, %f3614; setp.leu.ftz.f32 %p1915, %f3616, 0f00000000; mov.f32 %f16596, %f16746; @%p1915 bra $L__BB108_2153; sqrt.approx.ftz.f32 %f13532, %f3616; sub.ftz.f32 %f13533, %f13532, %f3615; add.ftz.f32 %f13534, %f3615, %f13532; div.approx.ftz.f32 %f13535, %f13533, %f13534; fma.rn.ftz.f32 %f13536, %f3615, %f13534, 0fBF800000; fma.rn.ftz.f32 %f13537, %f3615, %f13533, 0f3F800000; div.approx.ftz.f32 %f13538, %f13536, %f13537; mul.ftz.f32 %f13539, %f13535, 0f3F000000; mul.ftz.f32 %f13540, %f13535, %f13539; fma.rn.ftz.f32 %f13541, %f13538, %f13538, 0f3F800000; mul.ftz.f32 %f16596, %f13540, %f13541; $L__BB108_2153: setp.ne.s32 %p13, %r873, 15; sub.ftz.f32 %f3619, %f16746, %f16596; rcp.approx.ftz.f32 %f3620, %f3619; ld.local.u32 %r3281, [%rd7+168]; neg.ftz.f32 %f16619, %f3577; mov.f32 %f16616, 0f00000000; mov.u32 %r3282, 0; mov.pred %p1916, -1; mov.pred %p1926, 0; mov.f32 %f16745, %f16746; mov.f32 %f16744, %f16746; mov.f32 %f16606, %f16746; $L__BB108_2154: mov.f32 %f3629, %f16606; mad.lo.s32 %r895, %r3281, 1103515245, 12345; cvt.rn.f32.u32 %f13547, %r895; mul.ftz.f32 %f3631, %f13547, 0f2F800000; setp.gt.ftz.f32 %p1917, %f16619, 0f3F7FF972; mov.pred %p2189, %p1916; @%p1917 bra $L__BB108_2161; setp.lt.ftz.f32 %p1918, %f16619, 0fBF7FF972; @%p1918 bra $L__BB108_2160; bra.uni $L__BB108_2156; $L__BB108_2160: mul.ftz.f32 %f16606, %f3629, %f3631; mov.pred %p2189, %p1926; bra.uni $L__BB108_2161; $L__BB108_2156: abs.ftz.f32 %f13548, %f16619; setp.ltu.ftz.f32 %p1920, %f13548, 0f38D1B717; mov.f32 %f16606, %f3629; mov.pred %p2189, %p1926; @%p1920 bra $L__BB108_2161; mov.f32 %f13549, 0f3F800000; sub.ftz.f32 %f13550, %f13549, %f16616; setp.lt.ftz.f32 %p1922, %f13550, %f3631; mov.f32 %f16606, %f3629; mov.pred %p2189, %p1916; @%p1922 bra $L__BB108_2161; setp.ge.ftz.f32 %p1924, %f16615, 0f00000000; mov.f32 %f16606, %f13549; mov.pred %p2189, %p1926; @%p1924 bra $L__BB108_2161; mov.f32 %f13552, 0f3F800000; sub.ftz.f32 %f13553, %f13552, %f3631; lg2.approx.ftz.f32 %f13554, %f13553; mov.f32 %f13555, 0fBF800000; div.approx.ftz.f32 %f13556, %f13555, %f16615; mul.ftz.f32 %f13557, %f13556, %f13554; ex2.approx.ftz.f32 %f13558, %f13557; mul.ftz.f32 %f16606, %f3629, %f13558; mov.pred %p2189, %p1926; $L__BB108_2161: @%p2189 bra $L__BB108_2184; mad.lo.s32 %r2835, %r895, 1103515245, 12345; cvt.rn.f32.u32 %f13559, %r2835; mul.ftz.f32 %f3635, %f13559, 0f2F800000; mad.lo.s32 %r3281, %r2835, 1103515245, 12345; cvt.rn.f32.u32 %f13560, %r3281; mul.ftz.f32 %f3636, %f13560, 0f2F800000; neg.ftz.f32 %f3637, %f16602; mul.ftz.f32 %f13561, %f13470, %f3637; neg.ftz.f32 %f3638, %f16601; mul.ftz.f32 %f13562, %f13471, %f3638; mul.ftz.f32 %f13563, %f13562, %f13562; fma.rn.ftz.f32 %f13564, %f13561, %f13561, %f13563; fma.rn.ftz.f32 %f13565, %f16619, %f16619, %f13564; sqrt.approx.ftz.f32 %f13566, %f13565; div.approx.ftz.f32 %f16609, %f13561, %f13566; div.approx.ftz.f32 %f16610, %f13562, %f13566; neg.ftz.f32 %f3641, %f16619; div.approx.ftz.f32 %f3642, %f3641, %f13566; setp.gt.ftz.f32 %p1927, %f3642, 0f3F7FF972; @%p1927 bra $L__BB108_2168; abs.ftz.f32 %f13567, %f3642; setp.lt.ftz.f32 %p1928, %f13567, 0f358637BD; @%p1928 bra $L__BB108_2168; bra.uni $L__BB108_2164; $L__BB108_2168: mov.f32 %f13611, 0f3F800000; sub.ftz.f32 %f13612, %f13611, %f3636; mov.f32 %f13613, 0f33D6BF95; max.ftz.f32 %f13614, %f13612, %f13613; div.approx.ftz.f32 %f13615, %f3636, %f13614; sqrt.approx.ftz.f32 %f13616, %f13615; mul.ftz.f32 %f13617, %f3635, 0f40C90FDB; cos.approx.ftz.f32 %f13618, %f13617; sin.approx.ftz.f32 %f13619, %f13617; mul.ftz.f32 %f16608, %f13616, %f13619; mul.ftz.f32 %f16607, %f13616, %f13618; $L__BB108_2169: mul.ftz.f32 %f13620, %f16610, %f16610; fma.rn.ftz.f32 %f13621, %f16609, %f16609, %f13620; add.ftz.f32 %f13622, %f13621, 0f00000000; sqrt.approx.ftz.f32 %f3660, %f13622; setp.neu.ftz.f32 %p1936, %f3660, 0f00000000; @%p1936 bra $L__BB108_2170; bra.uni $L__BB108_2171; $L__BB108_2170: rcp.approx.ftz.f32 %f13623, %f3660; mul.ftz.f32 %f16610, %f16610, %f13623; mul.ftz.f32 %f16609, %f16609, %f13623; $L__BB108_2171: mul.ftz.f32 %f13624, %f16607, %f16609; mul.ftz.f32 %f13625, %f16608, %f16610; sub.ftz.f32 %f13626, %f13624, %f13625; mul.ftz.f32 %f13627, %f13470, %f13626; mul.ftz.f32 %f13628, %f16608, %f16609; fma.rn.ftz.f32 %f13629, %f16607, %f16610, %f13628; mul.ftz.f32 %f13630, %f13471, %f13629; neg.ftz.f32 %f13631, %f13627; neg.ftz.f32 %f13632, %f13630; mul.ftz.f32 %f13633, %f13630, %f13630; fma.rn.ftz.f32 %f13634, %f13627, %f13627, %f13633; add.ftz.f32 %f13635, %f13634, 0f3F800000; mov.f32 %f16616, 0f3F800000; sqrt.approx.ftz.f32 %f13637, %f13635; div.approx.ftz.f32 %f3670, %f13631, %f13637; div.approx.ftz.f32 %f3671, %f13632, %f13637; div.approx.ftz.f32 %f3672, %f16616, %f13637; setp.ne.s32 %p1937, %r3282, 0; and.pred %p1938, %p1937, %p13; mul.ftz.f32 %f13638, %f3603, %f16744; selp.f32 %f3673, %f13638, %f16744, %p1938; mul.ftz.f32 %f13639, %f3604, %f16745; selp.f32 %f3674, %f13639, %f16745, %p1938; mul.ftz.f32 %f13640, %f3605, %f16746; selp.f32 %f3675, %f13640, %f16746, %p1938; mul.ftz.f32 %f13641, %f3671, %f3638; fma.rn.ftz.f32 %f13642, %f3670, %f3637, %f13641; fma.rn.ftz.f32 %f3676, %f3672, %f3641, %f13642; mov.f32 %f16744, %f3673; mov.f32 %f16745, %f3674; mov.f32 %f16746, %f3675; @%p13 bra $L__BB108_2176; abs.ftz.f32 %f3677, %f3676; fma.rn.ftz.f32 %f3678, %f3677, %f3677, %f3614; setp.leu.ftz.f32 %p1940, %f3678, 0f00000000; mov.f32 %f16611, %f3619; @%p1940 bra $L__BB108_2174; sqrt.approx.ftz.f32 %f13643, %f3678; sub.ftz.f32 %f13644, %f13643, %f3677; add.ftz.f32 %f13645, %f3677, %f13643; div.approx.ftz.f32 %f13646, %f13644, %f13645; fma.rn.ftz.f32 %f13647, %f3677, %f13645, 0fBF800000; fma.rn.ftz.f32 %f13648, %f3677, %f13644, 0f3F800000; div.approx.ftz.f32 %f13649, %f13647, %f13648; mul.ftz.f32 %f13650, %f13646, 0f3F000000; mul.ftz.f32 %f13651, %f13646, %f13650; fma.rn.ftz.f32 %f13652, %f13649, %f13649, 0f3F800000; mul.ftz.f32 %f13653, %f13651, %f13652; sub.ftz.f32 %f16611, %f13653, %f16596; $L__BB108_2174: mul.ftz.f32 %f13654, %f3620, %f16611; mov.f32 %f13655, 0f3F800000; sub.ftz.f32 %f13656, %f13655, %f13654; fma.rn.ftz.f32 %f16744, %f3607, %f13656, %f13654; fma.rn.ftz.f32 %f16745, %f3608, %f13656, %f13654; fma.rn.ftz.f32 %f16746, %f3609, %f13656, %f13654; setp.eq.s32 %p1941, %r3282, 0; @%p1941 bra $L__BB108_2176; mul.ftz.f32 %f16744, %f3673, %f16744; mul.ftz.f32 %f16745, %f3674, %f16745; mul.ftz.f32 %f16746, %f3675, %f16746; $L__BB108_2176: add.ftz.f32 %f13658, %f3670, %f3670; add.ftz.f32 %f13659, %f3671, %f3671; add.ftz.f32 %f13660, %f3672, %f3672; fma.rn.ftz.f32 %f16602, %f13658, %f3676, %f16602; fma.rn.ftz.f32 %f16601, %f13659, %f3676, %f16601; fma.rn.ftz.f32 %f16619, %f13660, %f3676, %f16619; setp.gt.ftz.f32 %p1942, %f16619, 0f3F7FF972; mov.f32 %f16615, 0f00000000; @%p1942 bra $L__BB108_2179; setp.lt.ftz.f32 %p1943, %f16619, 0fBF7FF972; mov.f32 %f16615, 0fBF7FF972; @%p1943 bra $L__BB108_2179; mul.ftz.f32 %f13662, %f16619, %f16619; mov.f32 %f13663, 0f33D6BF95; max.ftz.f32 %f13664, %f13662, %f13663; rcp.approx.ftz.f32 %f13665, %f13664; mul.ftz.f32 %f13666, %f13470, %f16602; mul.ftz.f32 %f13667, %f13471, %f16601; mul.ftz.f32 %f13668, %f13667, %f13667; fma.rn.ftz.f32 %f13669, %f13666, %f13666, %f13668; fma.rn.ftz.f32 %f13670, %f13669, %f13665, 0f3F800000; sqrt.approx.ftz.f32 %f13671, %f13670; neg.ftz.f32 %f13672, %f13671; setp.le.ftz.f32 %p1944, %f16619, 0f00000000; selp.f32 %f13673, %f13672, %f13671, %p1944; add.ftz.f32 %f13674, %f13673, 0fBF800000; mul.ftz.f32 %f16615, %f13674, 0f3F000000; $L__BB108_2179: @%p1942 bra $L__BB108_2182; setp.lt.ftz.f32 %p1946, %f16619, 0f3727C5AC; mov.f32 %f16616, 0f00000000; @%p1946 bra $L__BB108_2182; lg2.approx.ftz.f32 %f13677, %f16606; mul.ftz.f32 %f13678, %f16615, %f13677; ex2.approx.ftz.f32 %f16616, %f13678; $L__BB108_2182: add.s32 %r3282, %r3282, 1; setp.lt.u32 %p1947, %r3282, 10; @%p1947 bra $L__BB108_2154; bra.uni $L__BB108_2183; $L__BB108_2164: add.ftz.f32 %f13570, %f3642, 0f3F800000; mul.ftz.f32 %f3643, %f13570, 0f3F000000; setp.lt.ftz.f32 %p1929, %f3643, 0f38D1B717; mov.f32 %f16607, 0f00000000; mov.f32 %f16608, %f16607; @%p1929 bra $L__BB108_2169; add.ftz.f32 %f13573, %f3636, %f3636; mul.ftz.f32 %f13574, %f13573, %f3643; div.approx.ftz.f32 %f13575, %f13574, %f3642; add.ftz.f32 %f3644, %f13575, 0fBF800000; mul.ftz.f32 %f3645, %f3644, %f3644; add.ftz.f32 %f3646, %f3645, 0fBF800000; abs.ftz.f32 %f13576, %f3646; setp.lt.ftz.f32 %p1930, %f13576, 0f33D6BF95; mov.f32 %f16608, %f16607; @%p1930 bra $L__BB108_2169; mul.ftz.f32 %f13577, %f3642, %f3642; mov.f32 %f13578, 0f3F800000; sub.ftz.f32 %f13579, %f13578, %f13577; mov.f32 %f13580, 0f00000000; max.ftz.f32 %f13581, %f13579, %f13580; sqrt.approx.ftz.f32 %f13582, %f13581; div.approx.ftz.f32 %f13583, %f13582, %f3642; rcp.approx.ftz.f32 %f13584, %f3646; mul.ftz.f32 %f13585, %f13583, %f13583; mul.ftz.f32 %f13586, %f13585, %f13584; mul.ftz.f32 %f13587, %f13584, %f13586; sub.ftz.f32 %f13588, %f3645, %f13585; mul.ftz.f32 %f13589, %f13588, %f13584; sub.ftz.f32 %f13590, %f13587, %f13589; max.ftz.f32 %f13591, %f13590, %f13580; sqrt.approx.ftz.f32 %f13592, %f13591; mul.ftz.f32 %f13593, %f13583, %f13584; add.ftz.f32 %f13594, %f13593, %f13592; rcp.approx.ftz.f32 %f13595, %f13583; setp.gt.ftz.f32 %p1931, %f13594, %f13595; setp.lt.ftz.f32 %p1932, %f3644, 0f00000000; or.pred %p1933, %p1932, %p1931; sub.ftz.f32 %f13596, %f13593, %f13592; mov.f32 %f13597, 0f3F000000; sub.ftz.f32 %f13598, %f13597, %f3635; add.ftz.f32 %f13599, %f3635, 0fBF000000; setp.ge.ftz.f32 %p1934, %f3635, 0f3F000000; selp.f32 %f13600, %f13599, %f13598, %p1934; add.ftz.f32 %f13601, %f13600, %f13600; fma.rn.ftz.f32 %f13602, %f13601, 0f3E8C3611, 0fBF3BD31C; fma.rn.ftz.f32 %f13603, %f13601, %f13602, 0f3EED4413; mul.ftz.f32 %f13604, %f13601, %f13603; fma.rn.ftz.f32 %f13605, %f13601, 0f3DBE9D0F, 0f3E9E6C4C; fma.rn.ftz.f32 %f13606, %f13601, %f13605, 0fBF800000; fma.rn.ftz.f32 %f13607, %f13601, %f13606, 0f3F191676; div.approx.ftz.f32 %f13608, %f13604, %f13607; selp.f32 %f16607, %f13596, %f13594, %p1933; fma.rn.ftz.f32 %f13609, %f16607, %f16607, 0f3F800000; sqrt.approx.ftz.f32 %f13610, %f13609; mul.ftz.f32 %f16608, %f13608, %f13610; setp.ltu.ftz.f32 %p1935, %f3635, 0f3F000000; @%p1935 bra $L__BB108_2167; bra.uni $L__BB108_2169; $L__BB108_2167: neg.ftz.f32 %f16608, %f16608; bra.uni $L__BB108_2169; $L__BB108_2184: st.local.u32 [%rd7+168], %r895; bra.uni $L__BB108_2185; $L__BB108_2183: st.local.u32 [%rd7+168], %r3281; mov.f32 %f16602, 0f00000000; mov.f32 %f16619, 0f3F800000; mov.f32 %f16601, %f16602; mov.f32 %f16744, %f16602; mov.f32 %f16745, %f16602; mov.f32 %f16746, %f16602; $L__BB108_2185: mul.ftz.f32 %f13686, %f16589, %f16601; fma.rn.ftz.f32 %f13687, %f16592, %f16602, %f13686; mul.ftz.f32 %f13688, %f16590, %f16601; fma.rn.ftz.f32 %f13689, %f16593, %f16602, %f13688; mul.ftz.f32 %f13690, %f16591, %f16601; fma.rn.ftz.f32 %f13691, %f16594, %f16602, %f13690; fma.rn.ftz.f32 %f16740, %f3573, %f16619, %f13687; fma.rn.ftz.f32 %f16741, %f3575, %f16619, %f13689; fma.rn.ftz.f32 %f16742, %f3576, %f16619, %f13691; mul.ftz.f32 %f13692, %f2807, %f16741; fma.rn.ftz.f32 %f13693, %f2806, %f16740, %f13692; fma.rn.ftz.f32 %f13694, %f2808, %f16742, %f13693; setp.le.ftz.f32 %p1948, %f13694, 0f00000000; mov.u32 %r3285, 0; mov.f32 %f16743, 0f00000000; @%p1948 bra $L__BB108_2309; ld.local.f32 %f3712, [%rd587+20]; add.ftz.f32 %f13695, %f3601, %f16602; add.ftz.f32 %f13696, %f3602, %f16601; mul.ftz.f32 %f13697, %f13696, %f13696; fma.rn.ftz.f32 %f13698, %f13695, %f13695, %f13697; add.ftz.f32 %f13699, %f3577, %f16619; fma.rn.ftz.f32 %f13700, %f13699, %f13699, %f13698; sqrt.approx.ftz.f32 %f13701, %f13700; div.approx.ftz.f32 %f3713, %f13695, %f13701; div.approx.ftz.f32 %f3714, %f13696, %f13701; div.approx.ftz.f32 %f13702, %f13699, %f13701; mul.ftz.f32 %f3715, %f13702, %f13702; @%p1908 bra $L__BB108_2191; bra.uni $L__BB108_2187; $L__BB108_2191: setp.gt.ftz.f32 %p1952, %f3577, 0f3F7FF972; neg.ftz.f32 %f13749, %f3713; div.approx.ftz.f32 %f13750, %f13749, %f3712; neg.ftz.f32 %f13751, %f3714; div.approx.ftz.f32 %f13752, %f13751, %f13471; fma.rn.ftz.f32 %f13753, %f13750, %f13750, %f3715; fma.rn.ftz.f32 %f3720, %f13752, %f13752, %f13753; mov.f32 %f16624, 0f00000000; @%p1952 bra $L__BB108_2194; mov.f32 %f16624, 0fBF7FF972; @%p1912 bra $L__BB108_2194; mul.ftz.f32 %f13755, %f3577, %f3577; mov.f32 %f13756, 0f33D6BF95; max.ftz.f32 %f13757, %f13755, %f13756; rcp.approx.ftz.f32 %f13758, %f13757; mul.ftz.f32 %f13759, %f3601, %f3712; mul.ftz.f32 %f13760, %f3602, %f13471; mul.ftz.f32 %f13761, %f13760, %f13760; fma.rn.ftz.f32 %f13762, %f13759, %f13759, %f13761; fma.rn.ftz.f32 %f13763, %f13762, %f13758, 0f3F800000; sqrt.approx.ftz.f32 %f13764, %f13763; add.ftz.f32 %f13765, %f13764, 0fBF800000; mul.ftz.f32 %f16624, %f13765, 0f3F000000; $L__BB108_2194: add.ftz.f32 %f13766, %f16624, 0f3F800000; mov.f32 %f13767, 0f3F800000; mul.ftz.f32 %f13768, %f3577, %f13766; mov.f32 %f13769, 0f33D6BF95; max.ftz.f32 %f13770, %f13768, %f13769; mul.ftz.f32 %f13771, %f3720, 0f40490FDB; mul.ftz.f32 %f13772, %f3720, %f13771; mul.ftz.f32 %f13773, %f3712, %f13772; mul.ftz.f32 %f13774, %f13471, %f13773; max.ftz.f32 %f13775, %f13774, %f13769; rcp.approx.ftz.f32 %f13776, %f13775; mul.ftz.f32 %f13777, %f13776, 0f3E800000; div.approx.ftz.f32 %f13778, %f13777, %f13770; mul.ftz.f32 %f13779, %f3712, %f13471; sqrt.approx.ftz.f32 %f13780, %f13779; mul.ftz.f32 %f13781, %f13780, 0fBFFE59F3; mul.ftz.f32 %f13782, %f13780, %f13781; mul.ftz.f32 %f13783, %f13782, 0f3FB8AA3B; ex2.approx.ftz.f32 %f13784, %f13783; fma.rn.ftz.f32 %f13785, %f13784, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f13786, %f13780, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f13787, %f13780, %f13786, 0f417C8B44; fma.rn.ftz.f32 %f13788, %f13780, %f13787, 0fC1665E35; fma.rn.ftz.f32 %f13789, %f13780, %f13788, 0f40CE7881; fma.rn.ftz.f32 %f13790, %f13780, %f13789, 0fBF993E81; fma.rn.ftz.f32 %f13791, %f13780, %f13790, 0f3CE3C31E; fma.rn.ftz.f32 %f13792, %f13780, %f13791, 0f3BBA5D49; sub.ftz.f32 %f13793, %f13785, %f13792; cvt.ftz.sat.f32.f32 %f13794, %f13793; sub.ftz.f32 %f13795, %f13767, %f13794; mul.ftz.f32 %f13796, %f16619, 0f3EA2F983; mul.ftz.f32 %f13797, %f13796, %f13795; fma.rn.ftz.f32 %f16625, %f13778, %f13794, %f13797; bra.uni $L__BB108_2195; $L__BB108_2118: setp.lt.ftz.f32 %p1886, %f2799, 0f3E800000; add.ftz.f32 %f13189, %f3492, 0f3F800000; div.approx.ftz.f32 %f13190, %f3511, %f13189; sqrt.approx.ftz.f32 %f3512, %f13190; @%p1886 bra $L__BB108_2130; bra.uni $L__BB108_2119; $L__BB108_2130: mul.ftz.f32 %f13303, %f2799, 0f40800000; mul.ftz.f32 %f13304, %f13303, 0f3FC90FDB; sin.approx.ftz.f32 %f13305, %f13304; cos.approx.ftz.f32 %f13306, %f13304; div.approx.ftz.f32 %f13307, %f13305, %f13306; mul.ftz.f32 %f3531, %f3512, %f13307; abs.ftz.f32 %f3532, %f3531; setp.leu.ftz.f32 %p1898, %f3532, 0f3F800000; mov.f32 %f16581, %f3532; @%p1898 bra $L__BB108_2132; rcp.approx.ftz.f32 %f16581, %f3532; $L__BB108_2132: mul.rn.ftz.f32 %f13308, %f16581, %f16581; mov.f32 %f13309, 0fC0B59883; mov.f32 %f13310, 0fBF52C7EA; fma.rn.ftz.f32 %f13311, %f13308, %f13310, %f13309; mov.f32 %f13312, 0fC0D21907; fma.rn.ftz.f32 %f13313, %f13311, %f13308, %f13312; mul.ftz.f32 %f13314, %f13308, %f13313; mul.ftz.f32 %f13315, %f16581, %f13314; add.ftz.f32 %f13316, %f13308, 0f41355DC0; mov.f32 %f13317, 0f41E6BD60; fma.rn.ftz.f32 %f13318, %f13316, %f13308, %f13317; mov.f32 %f13319, 0f419D92C8; fma.rn.ftz.f32 %f13320, %f13318, %f13308, %f13319; rcp.approx.ftz.f32 %f13321, %f13320; fma.rn.ftz.f32 %f13322, %f13315, %f13321, %f16581; mov.f32 %f13323, 0f3FC90FDB; sub.ftz.f32 %f13324, %f13323, %f13322; setp.gt.ftz.f32 %p1899, %f3532, 0f3F800000; selp.f32 %f13325, %f13324, %f13322, %p1899; mov.b32 %r2827, %f13325; mov.b32 %r2828, %f3531; and.b32 %r2829, %r2828, -2147483648; or.b32 %r2830, %r2829, %r2827; mov.b32 %f13326, %r2830; setp.le.ftz.f32 %p1900, %f3532, 0f7F800000; selp.f32 %f16583, %f13326, %f13325, %p1900; cos.approx.ftz.f32 %f13327, %f16583; sin.approx.ftz.f32 %f13328, %f16583; mul.ftz.f32 %f13329, %f3491, %f13327; mul.ftz.f32 %f13330, %f3492, %f13328; mul.ftz.f32 %f13331, %f13328, %f13330; fma.rn.ftz.f32 %f13332, %f13327, %f13329, %f13331; add.ftz.f32 %f13333, %f13332, 0f3F800000; rcp.approx.ftz.f32 %f13334, %f13333; lg2.approx.ftz.f32 %f13335, %f2775; mul.ftz.f32 %f13336, %f13335, %f13334; ex2.approx.ftz.f32 %f16582, %f13336; bra.uni $L__BB108_2134; $L__BB108_2260: add.s32 %r2876, %r873, -11; setp.gt.u32 %p2005, %r2876, 1; mov.u32 %r3285, 18; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p2005 bra $L__BB108_2281; ld.local.f32 %f14530, [%rd587+28]; fma.rn.ftz.f32 %f3949, %f14530, %f14530, 0fBF800000; abs.ftz.f32 %f3950, %f14497; fma.rn.ftz.f32 %f3951, %f3950, %f3950, %f3949; setp.leu.ftz.f32 %p2006, %f3951, 0f00000000; mov.f32 %f16706, %f14497; @%p2006 bra $L__BB108_2263; sqrt.approx.ftz.f32 %f14531, %f3951; sub.ftz.f32 %f14532, %f14531, %f3950; add.ftz.f32 %f14533, %f3950, %f14531; div.approx.ftz.f32 %f14534, %f14532, %f14533; fma.rn.ftz.f32 %f14535, %f3950, %f14533, 0fBF800000; fma.rn.ftz.f32 %f14536, %f3950, %f14532, 0f3F800000; div.approx.ftz.f32 %f14537, %f14535, %f14536; mul.ftz.f32 %f14538, %f14534, 0f3F000000; mul.ftz.f32 %f14539, %f14534, %f14538; fma.rn.ftz.f32 %f14540, %f14537, %f14537, 0f3F800000; mul.ftz.f32 %f16706, %f14539, %f14540; $L__BB108_2263: ld.local.u64 %rd2687, [%rd587+36]; ld.f32 %f3954, [%rd2687+12]; ld.f32 %f3955, [%rd2687+16]; ld.f32 %f3956, [%rd2687+20]; sub.ftz.f32 %f3957, %f14497, %f16706; mul.ftz.f32 %f14542, %f3942, %f16741; fma.rn.ftz.f32 %f14543, %f3941, %f16740, %f14542; fma.rn.ftz.f32 %f14544, %f3943, %f16742, %f14543; abs.ftz.f32 %f3958, %f14544; fma.rn.ftz.f32 %f3959, %f3958, %f3958, %f3949; setp.leu.ftz.f32 %p2007, %f3959, 0f00000000; mov.f32 %f16707, %f3957; @%p2007 bra $L__BB108_2265; sqrt.approx.ftz.f32 %f14545, %f3959; sub.ftz.f32 %f14546, %f14545, %f3958; add.ftz.f32 %f14547, %f3958, %f14545; div.approx.ftz.f32 %f14548, %f14546, %f14547; fma.rn.ftz.f32 %f14549, %f3958, %f14547, 0fBF800000; fma.rn.ftz.f32 %f14550, %f3958, %f14546, 0f3F800000; div.approx.ftz.f32 %f14551, %f14549, %f14550; mul.ftz.f32 %f14552, %f14548, 0f3F000000; mul.ftz.f32 %f14553, %f14548, %f14552; fma.rn.ftz.f32 %f14554, %f14551, %f14551, 0f3F800000; mul.ftz.f32 %f14555, %f14553, %f14554; sub.ftz.f32 %f16707, %f14555, %f16706; $L__BB108_2265: rcp.approx.ftz.f32 %f14557, %f3957; mov.f32 %f14558, 0f3F800000; mul.ftz.f32 %f14559, %f14557, %f16707; sub.ftz.f32 %f14560, %f14558, %f14559; fma.rn.ftz.f32 %f14561, %f3956, %f14560, %f14559; fma.rn.ftz.f32 %f14562, %f3955, %f14560, %f14559; fma.rn.ftz.f32 %f14563, %f3954, %f14560, %f14559; mul.ftz.f32 %f16744, %f14563, 0f49742400; mul.ftz.f32 %f16745, %f14562, 0f49742400; mul.ftz.f32 %f16746, %f14561, 0f49742400; bra.uni $L__BB108_2281; $L__BB108_2215: neg.ftz.f32 %f14022, %f3797; sqrt.approx.ftz.f32 %f14023, %f14022; add.ftz.f32 %f16648, %f14023, 0fC0400000; mov.f32 %f16656, 0fB951F09B; mov.f32 %f16655, 0f38D3B56B; mov.f32 %f16654, 0f3AB0DC72; mov.f32 %f16653, 0fBB70BDE7; mov.f32 %f16652, 0f3BBC127B; mov.f32 %f16651, 0fBBF9C5D7; mov.f32 %f16650, 0f3C1AA57E; mov.f32 %f16649, 0f3F8036DB; mov.f32 %f16647, 0f40354F7E; $L__BB108_2217: fma.rn.ftz.f32 %f14034, %f16648, %f16656, %f16655; fma.rn.ftz.f32 %f14035, %f16648, %f14034, %f16654; fma.rn.ftz.f32 %f14036, %f16648, %f14035, %f16653; fma.rn.ftz.f32 %f14037, %f16648, %f14036, %f16652; fma.rn.ftz.f32 %f14038, %f16648, %f14037, %f16651; fma.rn.ftz.f32 %f14039, %f16648, %f14038, %f16650; fma.rn.ftz.f32 %f14040, %f16648, %f14039, %f16649; fma.rn.ftz.f32 %f14041, %f16648, %f14040, %f16647; mul.ftz.f32 %f14042, %f3796, %f14041; mul.ftz.f32 %f14043, %f14042, %f14042; mul.ftz.f32 %f14044, %f14043, 0fBFB8AA3B; ex2.approx.ftz.f32 %f14045, %f14044; add.ftz.f32 %f14046, %f3796, 0f3F800000; fma.rn.ftz.f32 %f14048, %f3775, %f14045, %f14046; sub.ftz.f32 %f14049, %f14048, %f3777; mul.ftz.f32 %f14050, %f3771, %f14042; sub.ftz.f32 %f14051, %f13913, %f14050; div.approx.ftz.f32 %f14052, %f14049, %f14051; sub.ftz.f32 %f3810, %f3796, %f14052; abs.ftz.f32 %f14053, %f3810; setp.gt.ftz.f32 %p1970, %f14053, 0f3F7FFFFF; selp.f32 %f14054, 0f3F7FFFFF, %f14053, %p1970; sub.ftz.f32 %f14055, %f13913, %f14054; add.ftz.f32 %f14056, %f14054, 0f3F800000; mul.ftz.f32 %f14057, %f14055, %f14056; max.ftz.f32 %f14059, %f14057, %f13933; min.ftz.f32 %f14061, %f14059, %f13935; mov.b32 %r2855, %f14061; shr.u32 %r2856, %r2855, 23; add.s32 %r2857, %r2856, -127; and.b32 %r2858, %r2855, 8388607; or.b32 %r2859, %r2858, 1065353216; mov.b32 %f14062, %r2859; add.ftz.f32 %f14063, %f14062, 0fBF800000; mul.ftz.f32 %f14064, %f14063, %f14063; mul.ftz.f32 %f14065, %f14064, %f14064; fma.rn.ftz.f32 %f14066, %f14063, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f14067, %f14063, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14068, %f14063, %f14066, 0fBE0CD4FD; fma.rn.ftz.f32 %f14069, %f14063, %f14068, 0f3E77ADBD; fma.rn.ftz.f32 %f14070, %f14063, %f14069, 0fBEB1D206; fma.rn.ftz.f32 %f14071, %f14063, %f14067, 0f3FB8AA10; mul.ftz.f32 %f14072, %f14065, %f14070; fma.rn.ftz.f32 %f14073, %f14063, %f14071, %f14072; cvt.rn.f32.s32 %f14074, %r2857; add.ftz.f32 %f14075, %f14073, %f14074; mul.ftz.f32 %f3811, %f14075, 0f3F317218; setp.gt.ftz.f32 %p1971, %f3811, 0fC0A00000; @%p1971 bra $L__BB108_2219; bra.uni $L__BB108_2218; $L__BB108_2219: mov.f32 %f14096, 0fC0200000; sub.ftz.f32 %f16658, %f14096, %f3811; mov.f32 %f16666, 0f32F16588; mov.f32 %f16665, 0f34B84B36; mov.f32 %f16664, 0fB66C7357; mov.f32 %f16663, 0fB6935AC1; mov.f32 %f16662, 0f396532DB; mov.f32 %f16661, 0fBAA45408; mov.f32 %f16660, 0fBB88E4EF; mov.f32 %f16659, 0f3E7C8F63; mov.f32 %f16657, 0f3FC02E2F; bra.uni $L__BB108_2220; $L__BB108_2218: neg.ftz.f32 %f14085, %f3811; sqrt.approx.ftz.f32 %f14086, %f14085; add.ftz.f32 %f16658, %f14086, 0fC0400000; mov.f32 %f16666, 0fB951F09B; mov.f32 %f16665, 0f38D3B56B; mov.f32 %f16664, 0f3AB0DC72; mov.f32 %f16663, 0fBB70BDE7; mov.f32 %f16662, 0f3BBC127B; mov.f32 %f16661, 0fBBF9C5D7; mov.f32 %f16660, 0f3C1AA57E; mov.f32 %f16659, 0f3F8036DB; mov.f32 %f16657, 0f40354F7E; $L__BB108_2220: fma.rn.ftz.f32 %f14097, %f16658, %f16666, %f16665; fma.rn.ftz.f32 %f14098, %f16658, %f14097, %f16664; fma.rn.ftz.f32 %f14099, %f16658, %f14098, %f16663; fma.rn.ftz.f32 %f14100, %f16658, %f14099, %f16662; fma.rn.ftz.f32 %f14101, %f16658, %f14100, %f16661; fma.rn.ftz.f32 %f14102, %f16658, %f14101, %f16660; fma.rn.ftz.f32 %f14103, %f16658, %f14102, %f16659; fma.rn.ftz.f32 %f14104, %f16658, %f14103, %f16657; mul.ftz.f32 %f16680, %f3810, %f14104; $L__BB108_2221: fma.rn.ftz.f32 %f3826, %f2775, 0f40000000, 0fBF800000; abs.ftz.f32 %f14106, %f3826; setp.gt.ftz.f32 %p1972, %f14106, 0f3F7FFFFF; selp.f32 %f14107, 0f3F7FFFFF, %f14106, %p1972; sub.ftz.f32 %f14108, %f13913, %f14107; add.ftz.f32 %f14109, %f14107, 0f3F800000; mul.ftz.f32 %f14110, %f14108, %f14109; max.ftz.f32 %f14112, %f14110, %f13933; min.ftz.f32 %f14114, %f14112, %f13935; mov.b32 %r2860, %f14114; shr.u32 %r2861, %r2860, 23; add.s32 %r2862, %r2861, -127; and.b32 %r2863, %r2860, 8388607; or.b32 %r2864, %r2863, 1065353216; mov.b32 %f14115, %r2864; add.ftz.f32 %f14116, %f14115, 0fBF800000; mul.ftz.f32 %f14117, %f14116, %f14116; mul.ftz.f32 %f14118, %f14117, %f14117; fma.rn.ftz.f32 %f14119, %f14116, 0fBC188B0D, 0f3D5541C9; fma.rn.ftz.f32 %f14120, %f14116, 0f3EF5162D, 0fBF389E54; fma.rn.ftz.f32 %f14121, %f14116, %f14119, 0fBE0CD4FD; fma.rn.ftz.f32 %f14122, %f14116, %f14121, 0f3E77ADBD; fma.rn.ftz.f32 %f14123, %f14116, %f14122, 0fBEB1D206; fma.rn.ftz.f32 %f14124, %f14116, %f14120, 0f3FB8AA10; mul.ftz.f32 %f14125, %f14118, %f14123; fma.rn.ftz.f32 %f14126, %f14116, %f14124, %f14125; cvt.rn.f32.s32 %f14127, %r2862; add.ftz.f32 %f14128, %f14126, %f14127; mul.ftz.f32 %f3827, %f14128, 0f3F317218; setp.gt.ftz.f32 %p1973, %f3827, 0fC0A00000; @%p1973 bra $L__BB108_2223; bra.uni $L__BB108_2222; $L__BB108_2223: mov.f32 %f14149, 0fC0200000; sub.ftz.f32 %f16669, %f14149, %f3827; mov.f32 %f16677, 0f32F16588; mov.f32 %f16676, 0f34B84B36; mov.f32 %f16675, 0fB66C7357; mov.f32 %f16674, 0fB6935AC1; mov.f32 %f16673, 0f396532DB; mov.f32 %f16672, 0fBAA45408; mov.f32 %f16671, 0fBB88E4EF; mov.f32 %f16670, 0f3E7C8F63; mov.f32 %f16668, 0f3FC02E2F; bra.uni $L__BB108_2224; $L__BB108_2222: neg.ftz.f32 %f14138, %f3827; sqrt.approx.ftz.f32 %f14139, %f14138; add.ftz.f32 %f16669, %f14139, 0fC0400000; mov.f32 %f16677, 0fB951F09B; mov.f32 %f16676, 0f38D3B56B; mov.f32 %f16675, 0f3AB0DC72; mov.f32 %f16674, 0fBB70BDE7; mov.f32 %f16673, 0f3BBC127B; mov.f32 %f16672, 0fBBF9C5D7; mov.f32 %f16671, 0f3C1AA57E; mov.f32 %f16670, 0f3F8036DB; mov.f32 %f16668, 0f40354F7E; $L__BB108_2224: fma.rn.ftz.f32 %f14150, %f16669, %f16677, %f16676; fma.rn.ftz.f32 %f14151, %f16669, %f14150, %f16675; fma.rn.ftz.f32 %f14152, %f16669, %f14151, %f16674; fma.rn.ftz.f32 %f14153, %f16669, %f14152, %f16673; fma.rn.ftz.f32 %f14154, %f16669, %f14153, %f16672; fma.rn.ftz.f32 %f14155, %f16669, %f14154, %f16671; fma.rn.ftz.f32 %f14156, %f16669, %f14155, %f16670; fma.rn.ftz.f32 %f14157, %f16669, %f14156, %f16668; mul.ftz.f32 %f16679, %f3826, %f14157; add.ftz.f32 %f14158, %f3774, 0f3F800000; rcp.approx.ftz.f32 %f16678, %f14158; $L__BB108_2225: mul.ftz.f32 %f14159, %f16635, %f16679; mul.ftz.f32 %f14160, %f16634, %f16680; sub.ftz.f32 %f14161, %f14160, %f14159; mul.ftz.f32 %f14162, %f16635, %f16680; fma.rn.ftz.f32 %f14163, %f16634, %f16679, %f14162; mul.ftz.f32 %f14164, %f13816, %f14161; mul.ftz.f32 %f14165, %f13816, %f14163; neg.ftz.f32 %f14166, %f14164; neg.ftz.f32 %f14167, %f14165; mul.ftz.f32 %f14168, %f14165, %f14165; fma.rn.ftz.f32 %f14169, %f14164, %f14164, %f14168; add.ftz.f32 %f14170, %f14169, 0f3F800000; mov.f32 %f16682, 0f3F800000; sqrt.approx.ftz.f32 %f14172, %f14170; div.approx.ftz.f32 %f14173, %f14166, %f14172; div.approx.ftz.f32 %f14174, %f14167, %f14172; div.approx.ftz.f32 %f3845, %f16682, %f14172; mul.ftz.f32 %f14175, %f16626, %f14174; mul.ftz.f32 %f14176, %f16627, %f14174; mul.ftz.f32 %f14177, %f16628, %f14174; fma.rn.ftz.f32 %f14178, %f16629, %f14173, %f14175; fma.rn.ftz.f32 %f14179, %f16630, %f14173, %f14176; fma.rn.ftz.f32 %f14180, %f16631, %f14173, %f14177; fma.rn.ftz.f32 %f3846, %f3736, %f3845, %f14178; fma.rn.ftz.f32 %f3847, %f3738, %f3845, %f14179; fma.rn.ftz.f32 %f3848, %f3739, %f3845, %f14180; mul.ftz.f32 %f14181, %f13814, %f3847; fma.rn.ftz.f32 %f14182, %f3737, %f3846, %f14181; fma.rn.ftz.f32 %f3849, %f13815, %f3848, %f14182; setp.eq.s32 %p1974, %r873, 22; @%p1974 bra $L__BB108_2235; bra.uni $L__BB108_2226; $L__BB108_2235: setp.gt.ftz.f32 %p1981, %f3849, 0f00000000; ld.local.f32 %f3867, [%rd587+28]; @%p1981 bra $L__BB108_2237; bra.uni $L__BB108_2236; $L__BB108_2237: rcp.approx.ftz.f32 %f16687, %f3867; mov.u16 %rs158, 0; mov.f32 %f16683, %f3846; mov.f32 %f16684, %f3847; mov.f32 %f16685, %f3848; mov.f32 %f16686, %f3849; bra.uni $L__BB108_2238; $L__BB108_2226: setp.leu.ftz.f32 %p1975, %f3849, 0f00000000; mov.u32 %r2865, 10; mov.f32 %f16743, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; mov.u32 %r3285, %r2865; @%p1975 bra $L__BB108_2309; add.ftz.f32 %f14192, %f3849, %f3849; mul.ftz.f32 %f14193, %f3846, %f14192; mul.ftz.f32 %f14194, %f3847, %f14192; mul.ftz.f32 %f14195, %f3848, %f14192; sub.ftz.f32 %f16740, %f14193, %f3737; sub.ftz.f32 %f16741, %f14194, %f13814; sub.ftz.f32 %f16742, %f14195, %f13815; mul.ftz.f32 %f14196, %f2807, %f16741; fma.rn.ftz.f32 %f14197, %f2806, %f16740, %f14196; fma.rn.ftz.f32 %f14198, %f2808, %f16742, %f14197; setp.leu.ftz.f32 %p1976, %f14198, 0f00000000; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; mov.u32 %r3285, %r2865; @%p1976 bra $L__BB108_2309; mul.ftz.f32 %f3853, %f13816, %f13817; setp.le.ftz.f32 %p1977, %f3853, 0f33D6BF95; mov.u32 %r3285, 18; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1977 bra $L__BB108_2309; mul.ftz.f32 %f3854, %f3853, 0f40490FDB; mul.ftz.f32 %f14203, %f3738, %f16741; fma.rn.ftz.f32 %f14204, %f3736, %f16740, %f14203; fma.rn.ftz.f32 %f14205, %f3739, %f16742, %f14204; mul.ftz.f32 %f3855, %f14205, %f14205; sub.ftz.f32 %f3856, %f16682, %f3855; @%p1957 bra $L__BB108_2232; bra.uni $L__BB108_2230; $L__BB108_2232: mul.ftz.f32 %f14256, %f3845, %f3845; mul.ftz.f32 %f14257, %f14256, %f14256; rcp.approx.ftz.f32 %f14258, %f14256; add.ftz.f32 %f14259, %f14258, 0fBF800000; neg.ftz.f32 %f14260, %f14259; div.approx.ftz.f32 %f14261, %f14260, %f3853; mul.ftz.f32 %f14262, %f14261, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14263, %f14262; mul.ftz.f32 %f14264, %f3854, %f14257; div.approx.ftz.f32 %f16681, %f14263, %f14264; div.approx.ftz.f32 %f14265, %f3856, %f3855; mov.f32 %f14266, 0f00000000; max.ftz.f32 %f14267, %f14265, %f14266; sqrt.approx.ftz.f32 %f14268, %f14267; mul.ftz.f32 %f3861, %f13816, %f14268; setp.lt.ftz.f32 %p1980, %f3861, 0f3F200000; @%p1980 bra $L__BB108_2234; rcp.approx.ftz.f32 %f14269, %f3861; fma.rn.ftz.f32 %f14270, %f14269, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14271, %f14269, %f14270; fma.rn.ftz.f32 %f14272, %f14269, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14273, %f14269, %f14272, 0f3F800000; div.approx.ftz.f32 %f16682, %f14271, %f14273; bra.uni $L__BB108_2234; $L__BB108_2236: neg.ftz.f32 %f16686, %f3849; neg.ftz.f32 %f16683, %f3846; neg.ftz.f32 %f16684, %f3847; neg.ftz.f32 %f16685, %f3848; mov.u16 %rs158, 1; mov.f32 %f16687, %f3867; $L__BB108_2238: mul.ftz.f32 %f14281, %f16686, %f16686; mov.f32 %f16691, 0f3F800000; sub.ftz.f32 %f14282, %f16691, %f14281; mul.ftz.f32 %f14283, %f16687, %f16687; mul.ftz.f32 %f14284, %f14283, %f14282; sub.ftz.f32 %f3878, %f16691, %f14284; setp.lt.ftz.f32 %p1982, %f3878, 0f00000000; mov.f32 %f16743, 0f00000000; mov.f32 %f16740, %f16743; mov.f32 %f16741, %f16743; mov.f32 %f16742, %f16743; @%p1982 bra $L__BB108_2240; sqrt.approx.ftz.f32 %f14285, %f3878; mov.f32 %f14286, 0f33D6BF95; max.ftz.f32 %f14287, %f14285, %f14286; mul.ftz.f32 %f14288, %f16686, %f16687; sub.ftz.f32 %f14289, %f14288, %f14287; mul.ftz.f32 %f14290, %f16683, %f14289; mul.ftz.f32 %f14291, %f16684, %f14289; mul.ftz.f32 %f14292, %f16685, %f14289; mul.ftz.f32 %f14293, %f3737, %f16687; sub.ftz.f32 %f16740, %f14290, %f14293; mul.ftz.f32 %f14294, %f13814, %f16687; sub.ftz.f32 %f16741, %f14291, %f14294; mul.ftz.f32 %f14295, %f13815, %f16687; sub.ftz.f32 %f16742, %f14292, %f14295; mul.ftz.f32 %f14296, %f16684, %f16741; fma.rn.ftz.f32 %f14297, %f16683, %f16740, %f14296; fma.rn.ftz.f32 %f14298, %f16685, %f16742, %f14297; mul.ftz.f32 %f14299, %f3867, %f14298; add.ftz.f32 %f14300, %f16686, %f14299; sub.ftz.f32 %f14301, %f16686, %f14299; div.approx.ftz.f32 %f14302, %f14300, %f14301; mul.ftz.f32 %f14303, %f3867, %f16686; add.ftz.f32 %f14304, %f14303, %f14298; sub.ftz.f32 %f14305, %f14303, %f14298; div.approx.ftz.f32 %f14306, %f14304, %f14305; mul.ftz.f32 %f14307, %f14306, %f14306; fma.rn.ftz.f32 %f14308, %f14302, %f14302, %f14307; mul.ftz.f32 %f16691, %f14308, 0f3F000000; $L__BB108_2240: setp.eq.ftz.f32 %p1983, %f16691, 0f3F800000; setp.ne.s16 %p1984, %rs158, 0; mov.u32 %r2869, 9; or.pred %p1985, %p1984, %p1983; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; mov.u32 %r3285, %r2869; @%p1985 bra $L__BB108_2309; mul.ftz.f32 %f3887, %f13816, %f13817; setp.le.ftz.f32 %p1986, %f3887, 0f33D6BF95; mov.u32 %r3285, 17; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1986 bra $L__BB108_2309; add.ftz.f32 %f14322, %f3867, 0fBF800000; abs.ftz.f32 %f14323, %f14322; setp.lt.ftz.f32 %p1987, %f14323, 0f38D1B717; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1987 bra $L__BB108_2309; mul.ftz.f32 %f14325, %f3845, %f3845; mul.ftz.f32 %f14326, %f14325, %f14325; rcp.approx.ftz.f32 %f14327, %f14325; mov.f32 %f16692, 0f3F800000; add.ftz.f32 %f14328, %f14327, 0fBF800000; neg.ftz.f32 %f14329, %f14328; div.approx.ftz.f32 %f14330, %f14329, %f3887; mul.ftz.f32 %f14331, %f14330, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14332, %f14331; mul.ftz.f32 %f14333, %f3887, 0f40490FDB; mul.ftz.f32 %f14334, %f14333, %f14326; div.approx.ftz.f32 %f3888, %f14332, %f14334; mul.ftz.f32 %f14335, %f3738, %f16741; fma.rn.ftz.f32 %f14336, %f3736, %f16740, %f14335; fma.rn.ftz.f32 %f14337, %f3739, %f16742, %f14336; mul.ftz.f32 %f14338, %f14337, %f14337; sub.ftz.f32 %f14339, %f16692, %f14338; div.approx.ftz.f32 %f14340, %f14339, %f14338; mov.f32 %f14341, 0f00000000; max.ftz.f32 %f14342, %f14340, %f14341; sqrt.approx.ftz.f32 %f14343, %f14342; mul.ftz.f32 %f3889, %f13816, %f14343; setp.lt.ftz.f32 %p1988, %f3889, 0f3F200000; @%p1988 bra $L__BB108_2245; rcp.approx.ftz.f32 %f14344, %f3889; fma.rn.ftz.f32 %f14345, %f14344, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14346, %f14344, %f14345; fma.rn.ftz.f32 %f14347, %f14344, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14348, %f14344, %f14347, 0f3F800000; div.approx.ftz.f32 %f16692, %f14346, %f14348; $L__BB108_2245: mul.ftz.f32 %f14349, %f3847, %f16741; fma.rn.ftz.f32 %f14350, %f3846, %f16740, %f14349; fma.rn.ftz.f32 %f14351, %f3848, %f16742, %f14350; fma.rn.ftz.f32 %f14352, %f3867, %f14351, %f3849; mul.ftz.f32 %f14353, %f14352, %f14352; mul.ftz.f32 %f14354, %f3867, %f3867; mul.ftz.f32 %f14355, %f14354, %f3888; mul.ftz.f32 %f14356, %f3740, %f14353; div.approx.ftz.f32 %f14357, %f14355, %f14356; mul.ftz.f32 %f14358, %f3849, %f14351; abs.ftz.f32 %f14359, %f14358; mul.ftz.f32 %f14360, %f16678, %f16692; mul.ftz.f32 %f14361, %f14360, %f14359; mul.ftz.f32 %f16744, %f14357, %f14361; abs.ftz.f32 %f14362, %f14351; mul.ftz.f32 %f14363, %f16678, %f3849; mul.ftz.f32 %f14364, %f14363, %f14362; mul.ftz.f32 %f16743, %f14357, %f14364; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; mov.u32 %r3285, %r2869; bra.uni $L__BB108_2309; $L__BB108_740: ld.const.u64 %rd1214, [kernel_params+2176]; cvta.to.global.u64 %rd1215, %rd1214; ld.global.u32 %rd1216, [%rd163+168]; ld.global.u32 %r1416, [%rd163+152]; shl.b32 %r1417, %r1416, 1; cvt.rn.f32.s32 %f6487, %r1417; mul.ftz.f32 %f6488, %f6487, 0f3F000000; cvt.rzi.ftz.s32.f32 %r1418, %f6488; add.s32 %r1419, %r1417, -1; min.s32 %r1420, %r1418, %r1419; cvt.rn.f32.s32 %f6489, %r1420; sub.ftz.f32 %f1010, %f6488, %f6489; cvt.s64.s32 %rd1217, %r1420; add.s64 %rd1218, %rd1217, %rd1216; shl.b64 %rd1219, %rd1218, 6; add.s64 %rd164, %rd1215, %rd1219; ld.global.v4.f32 {%f6490, %f6491, %f6492, %f6493}, [%rd164]; ld.global.v4.f32 {%f6494, %f6495, %f6496, %f6497}, [%rd164+64]; mul.ftz.f32 %f6498, %f6491, %f6495; fma.rn.ftz.f32 %f6499, %f6490, %f6494, %f6498; mul.ftz.f32 %f6500, %f6493, %f6497; fma.rn.ftz.f32 %f6501, %f6492, %f6496, %f6500; add.ftz.f32 %f1019, %f6499, %f6501; setp.gt.ftz.f32 %p632, %f1019, 0f3F7FDF3B; sub.ftz.f32 %f1020, %f6437, %f1010; @%p632 bra $L__BB108_742; bra.uni $L__BB108_741; $L__BB108_742: mul.ftz.f32 %f6556, %f6490, %f1020; fma.rn.ftz.f32 %f6557, %f6494, %f1010, %f6556; mul.ftz.f32 %f6558, %f6491, %f1020; fma.rn.ftz.f32 %f6559, %f6495, %f1010, %f6558; mul.ftz.f32 %f6560, %f6492, %f1020; fma.rn.ftz.f32 %f6561, %f6496, %f1010, %f6560; mul.ftz.f32 %f6562, %f6493, %f1020; fma.rn.ftz.f32 %f6563, %f6497, %f1010, %f6562; mul.ftz.f32 %f6564, %f6559, %f6559; fma.rn.ftz.f32 %f6565, %f6557, %f6557, %f6564; mul.ftz.f32 %f6566, %f6563, %f6563; fma.rn.ftz.f32 %f6567, %f6561, %f6561, %f6566; add.ftz.f32 %f6568, %f6565, %f6567; rsqrt.approx.ftz.f32 %f6569, %f6568; mul.ftz.f32 %f15861, %f6563, %f6569; mul.ftz.f32 %f15860, %f6561, %f6569; mul.ftz.f32 %f15859, %f6559, %f6569; mul.ftz.f32 %f15858, %f6557, %f6569; bra.uni $L__BB108_743; $L__BB108_723: ld.const.u64 %rd1192, [kernel_params+2176]; cvta.to.global.u64 %rd1193, %rd1192; ld.global.u32 %rd1194, [%rd161+168]; ld.global.u32 %r1406, [%rd161+152]; shl.b32 %r1407, %r1406, 1; cvt.rn.f32.s32 %f6137, %r1407; mul.ftz.f32 %f6138, %f6137, 0f3F000000; cvt.rzi.ftz.s32.f32 %r1408, %f6138; add.s32 %r1409, %r1407, -1; min.s32 %r1410, %r1408, %r1409; cvt.rn.f32.s32 %f6139, %r1410; sub.ftz.f32 %f868, %f6138, %f6139; cvt.s64.s32 %rd1195, %r1410; add.s64 %rd1196, %rd1195, %rd1194; shl.b64 %rd1197, %rd1196, 6; add.s64 %rd162, %rd1193, %rd1197; ld.global.v4.f32 {%f6140, %f6141, %f6142, %f6143}, [%rd162]; ld.global.v4.f32 {%f6144, %f6145, %f6146, %f6147}, [%rd162+64]; mul.ftz.f32 %f6148, %f6141, %f6145; fma.rn.ftz.f32 %f6149, %f6140, %f6144, %f6148; mul.ftz.f32 %f6150, %f6143, %f6147; fma.rn.ftz.f32 %f6151, %f6142, %f6146, %f6150; add.ftz.f32 %f877, %f6149, %f6151; setp.gt.ftz.f32 %p616, %f877, 0f3F7FDF3B; mov.f32 %f6152, 0f3F800000; sub.ftz.f32 %f878, %f6152, %f868; @%p616 bra $L__BB108_725; bra.uni $L__BB108_724; $L__BB108_725: mul.ftz.f32 %f6206, %f6140, %f878; fma.rn.ftz.f32 %f6207, %f6144, %f868, %f6206; mul.ftz.f32 %f6208, %f6141, %f878; fma.rn.ftz.f32 %f6209, %f6145, %f868, %f6208; mul.ftz.f32 %f6210, %f6142, %f878; fma.rn.ftz.f32 %f6211, %f6146, %f868, %f6210; mul.ftz.f32 %f6212, %f6143, %f878; fma.rn.ftz.f32 %f6213, %f6147, %f868, %f6212; mul.ftz.f32 %f6214, %f6209, %f6209; fma.rn.ftz.f32 %f6215, %f6207, %f6207, %f6214; mul.ftz.f32 %f6216, %f6213, %f6213; fma.rn.ftz.f32 %f6217, %f6211, %f6211, %f6216; add.ftz.f32 %f6218, %f6215, %f6217; rsqrt.approx.ftz.f32 %f6219, %f6218; mul.ftz.f32 %f15827, %f6213, %f6219; mul.ftz.f32 %f15826, %f6211, %f6219; mul.ftz.f32 %f15825, %f6209, %f6219; mul.ftz.f32 %f15824, %f6207, %f6219; bra.uni $L__BB108_726; $L__BB108_2298: and.b32 %r2896, %r903, 4; setp.eq.s32 %p2050, %r2896, 0; add.ftz.f32 %f15068, %f4094, 0f00000000; selp.f32 %f16737, 0f00000000, %f15068, %p2050; $L__BB108_2299: and.b32 %r2897, %r903, 9; setp.eq.s32 %p2051, %r2897, 0; @%p2051 bra $L__BB108_2301; mul.ftz.f32 %f15069, %f15001, %f16741; fma.rn.ftz.f32 %f15070, %f4081, %f16740, %f15069; fma.rn.ftz.f32 %f15071, %f15002, %f16742, %f15070; add.ftz.f32 %f15072, %f15071, 0f3F800000; ld.local.f32 %f15073, [%rd587+20]; mul.ftz.f32 %f15074, %f15072, %f15073; add.ftz.f32 %f15075, %f15074, 0fBF800000; mul.ftz.f32 %f15076, %f4092, %f4093; add.ftz.f32 %f15077, %f4092, %f4093; fma.rn.ftz.f32 %f15078, %f15076, %f15075, %f15077; fma.rn.ftz.f32 %f16737, %f15074, %f15078, %f16737; $L__BB108_2301: mul.ftz.f32 %f15079, %f4091, 0f3EA2F983; mul.ftz.f32 %f16744, %f15079, %f16737; $L__BB108_2302: mul.ftz.f32 %f16743, %f4084, 0f3EA2F983; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_2187: setp.gt.ftz.f32 %p1950, %f3577, 0f3F7FF972; mul.ftz.f32 %f13704, %f3712, %f3712; mov.f32 %f13705, 0f3F800000; sub.ftz.f32 %f13706, %f13705, %f3715; fma.rn.ftz.f32 %f13707, %f13704, %f3715, %f13706; mul.ftz.f32 %f13708, %f13707, 0f40490FDB; mul.ftz.f32 %f13709, %f13707, %f13708; mov.f32 %f13710, 0f33D6BF95; max.ftz.f32 %f13711, %f13709, %f13710; div.approx.ftz.f32 %f3716, %f13704, %f13711; mov.f32 %f16623, 0f00000000; @%p1950 bra $L__BB108_2190; mov.f32 %f16623, 0fBF7FF972; @%p1912 bra $L__BB108_2190; mul.ftz.f32 %f13713, %f3577, %f3577; max.ftz.f32 %f13715, %f13713, %f13710; rcp.approx.ftz.f32 %f13716, %f13715; mul.ftz.f32 %f13717, %f3601, %f3712; mul.ftz.f32 %f13718, %f3602, %f3712; mul.ftz.f32 %f13719, %f13718, %f13718; fma.rn.ftz.f32 %f13720, %f13717, %f13717, %f13719; fma.rn.ftz.f32 %f13721, %f13720, %f13716, 0f3F800000; sqrt.approx.ftz.f32 %f13722, %f13721; add.ftz.f32 %f13723, %f13722, 0fBF800000; mul.ftz.f32 %f16623, %f13723, 0f3F000000; $L__BB108_2190: add.ftz.f32 %f13724, %f16623, 0f3F800000; mul.ftz.f32 %f13726, %f3577, %f13724; max.ftz.f32 %f13728, %f13726, %f13710; mul.ftz.f32 %f13729, %f3716, 0f3E800000; div.approx.ftz.f32 %f13730, %f13729, %f13728; mul.ftz.f32 %f13731, %f3712, 0fBFFE59F3; mul.ftz.f32 %f13732, %f3712, %f13731; mul.ftz.f32 %f13733, %f13732, 0f3FB8AA3B; ex2.approx.ftz.f32 %f13734, %f13733; fma.rn.ftz.f32 %f13735, %f13734, 0f3F4E7675, 0f3E4C51DB; fma.rn.ftz.f32 %f13736, %f3712, 0f3FE23A7E, 0fC10705C6; fma.rn.ftz.f32 %f13737, %f3712, %f13736, 0f417C8B44; fma.rn.ftz.f32 %f13738, %f3712, %f13737, 0fC1665E35; fma.rn.ftz.f32 %f13739, %f3712, %f13738, 0f40CE7881; fma.rn.ftz.f32 %f13740, %f3712, %f13739, 0fBF993E81; fma.rn.ftz.f32 %f13741, %f3712, %f13740, 0f3CE3C31E; fma.rn.ftz.f32 %f13742, %f3712, %f13741, 0f3BBA5D49; sub.ftz.f32 %f13743, %f13735, %f13742; cvt.ftz.sat.f32.f32 %f13744, %f13743; sub.ftz.f32 %f13745, %f13705, %f13744; mul.ftz.f32 %f13746, %f16619, 0f3EA2F983; mul.ftz.f32 %f13747, %f13746, %f13745; fma.rn.ftz.f32 %f16625, %f13730, %f13744, %f13747; $L__BB108_2195: mov.f32 %f13798, 0f00000000; max.ftz.f32 %f16743, %f13798, %f16625; mul.ftz.f32 %f16744, %f16744, %f16743; mul.ftz.f32 %f16745, %f16745, %f16743; mul.ftz.f32 %f16746, %f16746, %f16743; mov.u32 %r3285, 10; bra.uni $L__BB108_2309; $L__BB108_2119: setp.lt.ftz.f32 %p1887, %f2799, 0f3F000000; @%p1887 bra $L__BB108_2127; bra.uni $L__BB108_2120; $L__BB108_2127: mov.f32 %f13265, 0f3F000000; sub.ftz.f32 %f13266, %f13265, %f2799; mul.ftz.f32 %f13267, %f13266, 0f40800000; mul.ftz.f32 %f13268, %f13267, 0f3FC90FDB; sin.approx.ftz.f32 %f13269, %f13268; cos.approx.ftz.f32 %f13270, %f13268; div.approx.ftz.f32 %f13271, %f13269, %f13270; mul.ftz.f32 %f3525, %f3512, %f13271; abs.ftz.f32 %f3526, %f3525; setp.leu.ftz.f32 %p1895, %f3526, 0f3F800000; mov.f32 %f16580, %f3526; @%p1895 bra $L__BB108_2129; rcp.approx.ftz.f32 %f16580, %f3526; $L__BB108_2129: mul.rn.ftz.f32 %f13272, %f16580, %f16580; mov.f32 %f13273, 0fC0B59883; mov.f32 %f13274, 0fBF52C7EA; fma.rn.ftz.f32 %f13275, %f13272, %f13274, %f13273; mov.f32 %f13276, 0fC0D21907; fma.rn.ftz.f32 %f13277, %f13275, %f13272, %f13276; mul.ftz.f32 %f13278, %f13272, %f13277; mul.ftz.f32 %f13279, %f16580, %f13278; add.ftz.f32 %f13280, %f13272, 0f41355DC0; mov.f32 %f13281, 0f41E6BD60; fma.rn.ftz.f32 %f13282, %f13280, %f13272, %f13281; mov.f32 %f13283, 0f419D92C8; fma.rn.ftz.f32 %f13284, %f13282, %f13272, %f13283; rcp.approx.ftz.f32 %f13285, %f13284; fma.rn.ftz.f32 %f13286, %f13279, %f13285, %f16580; mov.f32 %f13287, 0f3FC90FDB; sub.ftz.f32 %f13288, %f13287, %f13286; setp.gt.ftz.f32 %p1896, %f3526, 0f3F800000; selp.f32 %f13289, %f13288, %f13286, %p1896; mov.b32 %r2823, %f13289; mov.b32 %r2824, %f3525; and.b32 %r2825, %r2824, -2147483648; or.b32 %r2826, %r2825, %r2823; mov.b32 %f13290, %r2826; setp.le.ftz.f32 %p1897, %f3526, 0f7F800000; selp.f32 %f13291, %f13290, %f13289, %p1897; cos.approx.ftz.f32 %f13292, %f13291; sin.approx.ftz.f32 %f13293, %f13291; mul.ftz.f32 %f13294, %f3491, %f13292; mul.ftz.f32 %f13295, %f3492, %f13293; mul.ftz.f32 %f13296, %f13293, %f13295; fma.rn.ftz.f32 %f13297, %f13292, %f13294, %f13296; add.ftz.f32 %f13298, %f13297, 0f3F800000; rcp.approx.ftz.f32 %f13299, %f13298; lg2.approx.ftz.f32 %f13300, %f2775; mul.ftz.f32 %f13301, %f13300, %f13299; ex2.approx.ftz.f32 %f16582, %f13301; mov.f32 %f13302, 0f40490FDB; sub.ftz.f32 %f16583, %f13302, %f13291; bra.uni $L__BB108_2134; $L__BB108_2100: mov.f32 %f12750, 0f3F800000; div.approx.ftz.f32 %f16569, %f3439, %f12750; $L__BB108_2102: add.ftz.f32 %f12753, %f16569, 0fBF800000; add.ftz.f32 %f12754, %f12753, 0f3F800000; mul.ftz.f32 %f12755, %f12754, %f12754; fma.rn.ftz.f32 %f12756, %f12755, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f12757, %f12754, %f12756; fma.rn.ftz.f32 %f12758, %f12755, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f12759, %f12755, %f12758, 0f3F800000; div.approx.ftz.f32 %f12760, %f12757, %f12759; sub.ftz.f32 %f12762, %f12741, %f12760; setp.gt.ftz.f32 %p1845, %f3439, 0f3F800000; selp.f32 %f12763, %f12762, %f12760, %p1845; mov.b32 %r2778, %f3438; and.b32 %r2779, %r2778, -2147483648; mov.b32 %r2780, %f12763; and.b32 %r2781, %r2780, 2147483647; or.b32 %r891, %r2781, %r2779; add.ftz.f32 %f12764, %f3434, 0fBFC90FDB; mul.ftz.f32 %f12765, %f12764, 0f3F000000; sub.ftz.f32 %f12766, %f12765, %f3437; mul.ftz.f32 %f3443, %f3436, %f12766; abs.ftz.f32 %f3444, %f3443; setp.eq.ftz.f32 %p1846, %f3444, 0f00000000; @%p1846 bra $L__BB108_2107; setp.eq.ftz.f32 %p1847, %f3444, 0f3F800000; mov.f32 %f16570, 0f3F800000; @%p1847 bra $L__BB108_2107; setp.gt.ftz.f32 %p1848, %f3444, 0f3F800000; @%p1848 bra $L__BB108_2106; bra.uni $L__BB108_2105; $L__BB108_2106: mov.f32 %f12769, 0f3F800000; div.approx.ftz.f32 %f16570, %f12769, %f3444; bra.uni $L__BB108_2107; $L__BB108_1877: mov.f32 %f11353, 0f3F800000; div.approx.ftz.f32 %f16463, %f2824, %f11353; $L__BB108_1879: add.ftz.f32 %f11356, %f16463, 0fBF800000; add.ftz.f32 %f11357, %f11356, 0f3F800000; mul.ftz.f32 %f11358, %f11357, %f11357; fma.rn.ftz.f32 %f11359, %f11358, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11360, %f11357, %f11359; fma.rn.ftz.f32 %f11361, %f11358, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11362, %f11358, %f11361, 0f3F800000; div.approx.ftz.f32 %f11363, %f11360, %f11362; sub.ftz.f32 %f11365, %f11343, %f11363; setp.gt.ftz.f32 %p1635, %f2824, 0f3F800000; selp.f32 %f11366, %f11365, %f11363, %p1635; mov.b32 %r2693, %f2823; and.b32 %r2694, %r2693, -2147483648; mov.b32 %r2695, %f11366; and.b32 %r2696, %r2695, 2147483647; or.b32 %r874, %r2696, %r2694; add.ftz.f32 %f11367, %f2819, 0fBFC90FDB; div.approx.ftz.f32 %f11369, %f11367, %f11349; sub.ftz.f32 %f11370, %f11369, %f2822; mul.ftz.f32 %f2828, %f2821, %f11370; abs.ftz.f32 %f2829, %f2828; setp.eq.ftz.f32 %p1636, %f2829, 0f00000000; mov.f32 %f16464, %f16465; @%p1636 bra $L__BB108_1884; setp.eq.ftz.f32 %p1637, %f2829, 0f3F800000; mov.f32 %f16464, 0f3F800000; @%p1637 bra $L__BB108_1884; setp.gt.ftz.f32 %p1638, %f2829, 0f3F800000; @%p1638 bra $L__BB108_1883; bra.uni $L__BB108_1882; $L__BB108_1883: mov.f32 %f11373, 0f3F800000; div.approx.ftz.f32 %f16464, %f11373, %f2829; bra.uni $L__BB108_1884; $L__BB108_2105: mov.f32 %f12768, 0f3F800000; div.approx.ftz.f32 %f16570, %f3444, %f12768; $L__BB108_2107: add.ftz.f32 %f12770, %f16570, 0fBF800000; mov.f32 %f12771, 0fBF800000; mov.f32 %f12772, 0f3F800000; add.ftz.f32 %f12773, %f12770, 0f3F800000; mul.ftz.f32 %f12774, %f12773, %f12773; fma.rn.ftz.f32 %f12775, %f12774, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f12776, %f12773, %f12775; fma.rn.ftz.f32 %f12777, %f12774, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f12778, %f12774, %f12777, 0f3F800000; div.approx.ftz.f32 %f12779, %f12776, %f12778; sub.ftz.f32 %f12781, %f12741, %f12779; setp.gt.ftz.f32 %p1849, %f3444, 0f3F800000; selp.f32 %f12782, %f12781, %f12779, %p1849; mov.b32 %r2783, %f3443; and.b32 %r2784, %r2783, -2147483648; mov.b32 %r2785, %f12782; and.b32 %r2786, %r2785, 2147483647; or.b32 %r2787, %r2786, %r2784; mov.b32 %f12783, %r2787; mov.b32 %f12784, %r891; sub.ftz.f32 %f12785, %f12784, %f12783; fma.rn.ftz.f32 %f12786, %f2799, %f12785, %f12783; sin.approx.ftz.f32 %f12787, %f12786; cos.approx.ftz.f32 %f12788, %f12786; div.approx.ftz.f32 %f12789, %f12787, %f12788; mul.ftz.f32 %f12790, %f3435, %f12789; add.ftz.f32 %f12791, %f3437, %f12790; add.ftz.f32 %f12792, %f12791, %f12791; sub.ftz.f32 %f12793, %f12792, %f3434; mul.ftz.f32 %f12794, %f12793, 0f3EA2F983; mov.b32 %r2788, %f12794; and.b32 %r2789, %r2788, -2147483648; or.b32 %r2790, %r2789, 1056964608; mov.b32 %f12795, %r2790; add.ftz.f32 %f12796, %f12794, %f12795; cvt.rzi.ftz.s32.f32 %r2791, %f12796; cvt.rn.f32.s32 %f12797, %r2791; mul.ftz.f32 %f12798, %f12797, 0f40490000; sub.ftz.f32 %f12799, %f12793, %f12798; mul.ftz.f32 %f12800, %f12797, 0f3A7DA000; sub.ftz.f32 %f12801, %f12799, %f12800; mul.ftz.f32 %f12802, %f12797, 0f34222000; sub.ftz.f32 %f12803, %f12801, %f12802; mul.ftz.f32 %f12804, %f12797, 0f2CB4611A; sub.ftz.f32 %f12805, %f12803, %f12804; add.ftz.f32 %f12806, %f12805, 0fBFC90FDB; add.ftz.f32 %f12807, %f12806, 0f3FC90FDB; mul.ftz.f32 %f12808, %f12807, %f12807; and.b32 %r2792, %r2791, 1; setp.eq.b32 %p1850, %r2792, 1; neg.ftz.f32 %f12809, %f12807; selp.f32 %f12810, %f12809, %f12807, %p1850; fma.rn.ftz.f32 %f12811, %f12808, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f12812, %f12808, %f12811, 0f3C088777; fma.rn.ftz.f32 %f12813, %f12808, %f12812, 0fBE2AAAA6; mul.ftz.f32 %f12814, %f12810, %f12813; fma.rn.ftz.f32 %f12815, %f12808, %f12814, %f12810; mul.ftz.f32 %f12816, %f12808, 0f3491ED89; mov.f32 %f12817, 0f37D0078B; sub.ftz.f32 %f12818, %f12817, %f12816; fma.rn.ftz.f32 %f12819, %f12808, %f12818, 0fBAB60B58; fma.rn.ftz.f32 %f12820, %f12808, %f12819, 0f3D2AAAAA; fma.rn.ftz.f32 %f12821, %f12808, %f12820, 0fBF000000; fma.rn.ftz.f32 %f12822, %f12808, %f12821, 0f3F800000; neg.ftz.f32 %f12823, %f12822; selp.f32 %f12824, %f12823, %f12822, %p1850; abs.ftz.f32 %f12825, %f12815; setp.gt.ftz.f32 %p1851, %f12825, 0f3F800000; selp.f32 %f12826, 0f00000000, %f12815, %p1851; abs.ftz.f32 %f12827, %f12824; setp.gt.ftz.f32 %p1852, %f12827, 0f3F800000; selp.f32 %f12828, 0f00000000, %f12824, %p1852; add.ftz.f32 %f12829, %f2775, %f2775; sub.ftz.f32 %f12830, %f12772, %f12829; max.ftz.f32 %f12831, %f12830, %f12771; min.ftz.f32 %f12832, %f12831, %f12772; abs.ftz.f32 %f12833, %f12832; sub.ftz.f32 %f12834, %f12772, %f12833; mul.ftz.f32 %f12835, %f12834, 0f3F000000; sqrt.approx.ftz.f32 %f12836, %f12835; setp.gt.ftz.f32 %p1853, %f12833, 0f3F11EB85; selp.f32 %f12837, %f12836, %f12833, %p1853; mul.ftz.f32 %f12838, %f12837, %f12837; mov.f32 %f12839, 0f3C94D2E9; mov.f32 %f12840, 0f3D53F941; fma.rn.ftz.f32 %f12841, %f12840, %f12838, %f12839; mov.f32 %f12842, 0f3D3F841F; fma.rn.ftz.f32 %f12843, %f12841, %f12838, %f12842; mov.f32 %f12844, 0f3D994929; fma.rn.ftz.f32 %f12845, %f12843, %f12838, %f12844; mov.f32 %f12846, 0f3E2AAB94; fma.rn.ftz.f32 %f12847, %f12845, %f12838, %f12846; mul.ftz.f32 %f12848, %f12838, %f12847; fma.rn.ftz.f32 %f12849, %f12848, %f12837, %f12837; mov.f32 %f12850, 0fC0000000; fma.rn.ftz.f32 %f12851, %f12850, %f12849, %f12741; selp.f32 %f12852, %f12851, %f12849, %p1853; setp.le.ftz.f32 %p1854, %f12852, 0f7F800000; mov.b32 %r2793, %f12852; mov.b32 %r2794, %f12832; and.b32 %r2795, %r2794, -2147483648; or.b32 %r2796, %r2795, %r2793; mov.b32 %f12853, %r2796; selp.f32 %f12854, %f12853, %f12852, %p1854; add.ftz.f32 %f12855, %f12854, %f12854; mul.ftz.f32 %f12856, %f3427, %f12855; mul.ftz.f32 %f12857, %f12856, 0f3F000000; mul.ftz.f32 %f12858, %f12857, 0f3EA2F983; mov.b32 %r2797, %f12858; and.b32 %r2798, %r2797, -2147483648; or.b32 %r2799, %r2798, 1056964608; mov.b32 %f12859, %r2799; add.ftz.f32 %f12860, %f12858, %f12859; cvt.rzi.ftz.s32.f32 %r2800, %f12860; cvt.rn.f32.s32 %f12861, %r2800; mul.ftz.f32 %f12862, %f12861, 0f40490000; sub.ftz.f32 %f12863, %f12857, %f12862; mul.ftz.f32 %f12864, %f12861, 0f3A7DA000; sub.ftz.f32 %f12865, %f12863, %f12864; mul.ftz.f32 %f12866, %f12861, 0f34222000; sub.ftz.f32 %f12867, %f12865, %f12866; mul.ftz.f32 %f12868, %f12861, 0f2CB4611A; sub.ftz.f32 %f12869, %f12867, %f12868; add.ftz.f32 %f12870, %f12869, 0fBFC90FDB; add.ftz.f32 %f12871, %f12870, 0f3FC90FDB; mul.ftz.f32 %f12872, %f12871, %f12871; mul.ftz.f32 %f12873, %f12872, 0f3491ED89; sub.ftz.f32 %f12874, %f12817, %f12873; fma.rn.ftz.f32 %f12875, %f12872, %f12874, 0fBAB60B58; fma.rn.ftz.f32 %f12876, %f12872, %f12875, 0f3D2AAAAA; fma.rn.ftz.f32 %f12877, %f12872, %f12876, 0fBF000000; fma.rn.ftz.f32 %f12878, %f12872, %f12877, 0f3F800000; and.b32 %r2801, %r2800, 1; setp.eq.b32 %p1855, %r2801, 1; neg.ftz.f32 %f12879, %f12878; selp.f32 %f12880, %f12879, %f12878, %p1855; abs.ftz.f32 %f12881, %f12880; setp.gt.ftz.f32 %p1856, %f12881, 0f3F800000; mul.ftz.f32 %f12882, %f12880, 0f3E800000; selp.f32 %f12883, 0f00000000, %f12882, %p1856; div.approx.ftz.f32 %f12884, %f12883, %f3427; mul.ftz.f32 %f12885, %f12790, %f12790; fma.rn.ftz.f32 %f12886, %f3435, %f3435, %f12885; add.ftz.f32 %f12887, %f12886, %f12886; mul.ftz.f32 %f12888, %f12887, %f12785; mul.ftz.f32 %f12889, %f12888, %f12828; div.approx.ftz.f32 %f12890, %f3435, %f12889; mul.ftz.f32 %f12891, %f12856, 0f3EA2F983; mov.b32 %r2802, %f12891; and.b32 %r2803, %r2802, -2147483648; or.b32 %r2804, %r2803, 1056964608; mov.b32 %f12892, %r2804; add.ftz.f32 %f12893, %f12891, %f12892; cvt.rzi.ftz.s32.f32 %r2805, %f12893; cvt.rn.f32.s32 %f12894, %r2805; mul.ftz.f32 %f12895, %f12894, 0f40490000; sub.ftz.f32 %f12896, %f12856, %f12895; mul.ftz.f32 %f12897, %f12894, 0f3A7DA000; sub.ftz.f32 %f12898, %f12896, %f12897; mul.ftz.f32 %f12899, %f12894, 0f34222000; sub.ftz.f32 %f12900, %f12898, %f12899; mul.ftz.f32 %f12901, %f12894, 0f2CB4611A; sub.ftz.f32 %f12902, %f12900, %f12901; add.ftz.f32 %f12903, %f12902, 0fBFC90FDB; add.ftz.f32 %f12904, %f12903, 0f3FC90FDB; mul.ftz.f32 %f12905, %f12904, %f12904; and.b32 %r2806, %r2805, 1; setp.eq.b32 %p1857, %r2806, 1; neg.ftz.f32 %f12906, %f12904; selp.f32 %f12907, %f12906, %f12904, %p1857; fma.rn.ftz.f32 %f12908, %f12905, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f12909, %f12905, %f12908, 0f3C088777; fma.rn.ftz.f32 %f12910, %f12905, %f12909, 0fBE2AAAA6; mul.ftz.f32 %f12911, %f12907, %f12910; fma.rn.ftz.f32 %f12912, %f12905, %f12911, %f12907; mul.ftz.f32 %f12913, %f12905, 0f3491ED89; sub.ftz.f32 %f12914, %f12817, %f12913; fma.rn.ftz.f32 %f12915, %f12905, %f12914, 0fBAB60B58; fma.rn.ftz.f32 %f12916, %f12905, %f12915, 0f3D2AAAAA; fma.rn.ftz.f32 %f12917, %f12905, %f12916, 0fBF000000; fma.rn.ftz.f32 %f12918, %f12905, %f12917, 0f3F800000; neg.ftz.f32 %f12919, %f12918; selp.f32 %f12920, %f12919, %f12918, %p1857; abs.ftz.f32 %f12921, %f12912; setp.gt.ftz.f32 %p1858, %f12921, 0f3F800000; selp.f32 %f12922, 0f00000000, %f12912, %p1858; abs.ftz.f32 %f12923, %f12920; setp.gt.ftz.f32 %p1859, %f12923, 0f3F800000; selp.f32 %f12924, 0f00000000, %f12920, %p1859; mul.ftz.f32 %f12925, %f12828, %f12924; mul.ftz.f32 %f12926, %f3431, %f12925; mul.ftz.f32 %f12927, %f3432, %f12925; mul.ftz.f32 %f12928, %f3433, %f12925; mul.ftz.f32 %f12929, %f12828, %f12922; mul.ftz.f32 %f12930, %f3430, %f3432; mul.ftz.f32 %f12931, %f3429, %f3433; sub.ftz.f32 %f12932, %f12930, %f12931; mul.ftz.f32 %f12933, %f12932, %f12929; mul.ftz.f32 %f12934, %f3428, %f3433; mul.ftz.f32 %f12935, %f3430, %f3431; sub.ftz.f32 %f12936, %f12934, %f12935; mul.ftz.f32 %f12937, %f12936, %f12929; mul.ftz.f32 %f12938, %f3429, %f3431; mul.ftz.f32 %f12939, %f3428, %f3432; sub.ftz.f32 %f12940, %f12938, %f12939; mul.ftz.f32 %f12941, %f12940, %f12929; sub.ftz.f32 %f12942, %f12926, %f12933; sub.ftz.f32 %f12943, %f12927, %f12937; sub.ftz.f32 %f12944, %f12928, %f12941; fma.rn.ftz.f32 %f16740, %f3428, %f12826, %f12942; fma.rn.ftz.f32 %f16741, %f3429, %f12826, %f12943; fma.rn.ftz.f32 %f16742, %f3430, %f12826, %f12944; mul.ftz.f32 %f12945, %f12884, %f12890; abs.ftz.f32 %f12946, %f12945; abs.ftz.f32 %f12947, %f12793; sub.ftz.f32 %f12948, %f12741, %f12947; setp.lt.ftz.f32 %p1860, %f12948, 0f3A83126F; selp.f32 %f16743, 0f00000000, %f12946, %p1860; mov.u32 %r3285, 10; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; bra.uni $L__BB108_2309; $L__BB108_1882: mov.f32 %f11372, 0f3F800000; div.approx.ftz.f32 %f16464, %f2829, %f11372; $L__BB108_1884: add.ftz.f32 %f11375, %f16464, 0fBF800000; add.ftz.f32 %f11376, %f11375, 0f3F800000; mul.ftz.f32 %f11377, %f11376, %f11376; fma.rn.ftz.f32 %f11378, %f11377, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11379, %f11376, %f11378; fma.rn.ftz.f32 %f11380, %f11377, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11381, %f11377, %f11380, 0f3F800000; div.approx.ftz.f32 %f11382, %f11379, %f11381; sub.ftz.f32 %f11384, %f11343, %f11382; setp.gt.ftz.f32 %p1639, %f2829, 0f3F800000; selp.f32 %f11385, %f11384, %f11382, %p1639; mov.b32 %r2697, %f2828; and.b32 %r2698, %r2697, -2147483648; mov.b32 %r2699, %f11385; and.b32 %r2700, %r2699, 2147483647; or.b32 %r875, %r2700, %r2698; div.approx.ftz.f32 %f2833, %f11343, %f2812; abs.ftz.f32 %f2834, %f2833; setp.eq.ftz.f32 %p1640, %f2834, 0f00000000; @%p1640 bra $L__BB108_1889; setp.eq.ftz.f32 %p1641, %f2834, 0f3F800000; mov.f32 %f16465, 0f3F800000; @%p1641 bra $L__BB108_1889; setp.gt.ftz.f32 %p1642, %f2834, 0f3F800000; @%p1642 bra $L__BB108_1888; bra.uni $L__BB108_1887; $L__BB108_1888: mov.f32 %f11388, 0f3F800000; div.approx.ftz.f32 %f16465, %f11388, %f2834; bra.uni $L__BB108_1889; $L__BB108_1887: mov.f32 %f11387, 0f3F800000; div.approx.ftz.f32 %f16465, %f2834, %f11387; $L__BB108_1889: mov.b32 %f11389, %r875; add.ftz.f32 %f11390, %f16465, 0fBF800000; add.ftz.f32 %f11391, %f11390, 0f3F800000; mul.ftz.f32 %f11392, %f11391, %f11391; fma.rn.ftz.f32 %f11393, %f11392, 0f3EDCF805, 0f3F800000; mul.ftz.f32 %f11394, %f11391, %f11393; fma.rn.ftz.f32 %f11395, %f11392, 0f3D6EE04D, 0f3F43B24E; fma.rn.ftz.f32 %f11396, %f11392, %f11395, 0f3F800000; div.approx.ftz.f32 %f11397, %f11394, %f11396; sub.ftz.f32 %f11399, %f11343, %f11397; setp.gt.ftz.f32 %p1643, %f2834, 0f3F800000; selp.f32 %f11400, %f11399, %f11397, %p1643; mov.b32 %r2702, %f2833; and.b32 %r2703, %r2702, -2147483648; mov.b32 %r2704, %f11400; and.b32 %r2705, %r2704, 2147483647; or.b32 %r2706, %r2705, %r2703; mov.b32 %f11401, %r2706; add.ftz.f32 %f11402, %f11401, %f11401; mov.b32 %f11403, %r874; sub.ftz.f32 %f11404, %f11403, %f11389; fma.rn.ftz.f32 %f11405, %f2799, %f11404, %f11389; sin.approx.ftz.f32 %f11406, %f11405; cos.approx.ftz.f32 %f11407, %f11405; div.approx.ftz.f32 %f11408, %f11406, %f11407; mul.ftz.f32 %f11409, %f2820, %f11408; add.ftz.f32 %f11410, %f2822, %f11409; add.ftz.f32 %f11411, %f11410, %f11410; sub.ftz.f32 %f11412, %f11411, %f2819; mul.ftz.f32 %f11413, %f11412, 0f3EA2F983; mov.b32 %r2707, %f11413; and.b32 %r2708, %r2707, -2147483648; or.b32 %r2709, %r2708, 1056964608; mov.b32 %f11414, %r2709; add.ftz.f32 %f11415, %f11413, %f11414; cvt.rzi.ftz.s32.f32 %r2710, %f11415; cvt.rn.f32.s32 %f11416, %r2710; mul.ftz.f32 %f11417, %f11416, 0f40490000; sub.ftz.f32 %f11418, %f11412, %f11417; mul.ftz.f32 %f11419, %f11416, 0f3A7DA000; sub.ftz.f32 %f11420, %f11418, %f11419; mul.ftz.f32 %f11421, %f11416, 0f34222000; sub.ftz.f32 %f11422, %f11420, %f11421; mul.ftz.f32 %f11423, %f11416, 0f2CB4611A; sub.ftz.f32 %f11424, %f11422, %f11423; add.ftz.f32 %f11425, %f11424, 0fBFC90FDB; add.ftz.f32 %f11426, %f11425, 0f3FC90FDB; mul.ftz.f32 %f11427, %f11426, %f11426; and.b32 %r2711, %r2710, 1; setp.eq.b32 %p1644, %r2711, 1; neg.ftz.f32 %f11428, %f11426; selp.f32 %f11429, %f11428, %f11426, %p1644; fma.rn.ftz.f32 %f11430, %f11427, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f11431, %f11427, %f11430, 0f3C088777; fma.rn.ftz.f32 %f11432, %f11427, %f11431, 0fBE2AAAA6; mul.ftz.f32 %f11433, %f11429, %f11432; fma.rn.ftz.f32 %f11434, %f11427, %f11433, %f11429; mul.ftz.f32 %f11435, %f11427, 0f3491ED89; mov.f32 %f11436, 0f37D0078B; sub.ftz.f32 %f11437, %f11436, %f11435; fma.rn.ftz.f32 %f11438, %f11427, %f11437, 0fBAB60B58; fma.rn.ftz.f32 %f11439, %f11427, %f11438, 0f3D2AAAAA; fma.rn.ftz.f32 %f11440, %f11427, %f11439, 0fBF000000; fma.rn.ftz.f32 %f11441, %f11427, %f11440, 0f3F800000; neg.ftz.f32 %f11442, %f11441; selp.f32 %f11443, %f11442, %f11441, %p1644; abs.ftz.f32 %f11444, %f11434; setp.gt.ftz.f32 %p1645, %f11444, 0f3F800000; selp.f32 %f11445, 0f00000000, %f11434, %p1645; abs.ftz.f32 %f11446, %f11443; setp.gt.ftz.f32 %p1646, %f11446, 0f3F800000; selp.f32 %f11447, 0f00000000, %f11443, %p1646; add.ftz.f32 %f11448, %f2775, 0fBF000000; mul.ftz.f32 %f11449, %f11448, %f11402; sin.approx.ftz.f32 %f11450, %f11449; cos.approx.ftz.f32 %f11451, %f11449; div.approx.ftz.f32 %f11452, %f11450, %f11451; mul.ftz.f32 %f11453, %f2812, %f11452; add.ftz.f32 %f11454, %f11453, 0f40490FDB; mul.ftz.f32 %f11455, %f11409, %f11409; fma.rn.ftz.f32 %f11456, %f2820, %f2820, %f11455; add.ftz.f32 %f11457, %f11456, %f11456; mul.ftz.f32 %f11458, %f11404, %f11457; mul.ftz.f32 %f11459, %f11458, %f11447; div.approx.ftz.f32 %f11460, %f2820, %f11459; mul.ftz.f32 %f11461, %f11453, %f11453; fma.rn.ftz.f32 %f11462, %f2812, %f2812, %f11461; mul.ftz.f32 %f11463, %f11402, %f11462; div.approx.ftz.f32 %f11464, %f2812, %f11463; mul.ftz.f32 %f11465, %f11454, 0f3EA2F983; mov.b32 %r2712, %f11465; and.b32 %r2713, %r2712, -2147483648; or.b32 %r2714, %r2713, 1056964608; mov.b32 %f11466, %r2714; add.ftz.f32 %f11467, %f11465, %f11466; cvt.rzi.ftz.s32.f32 %r2715, %f11467; cvt.rn.f32.s32 %f11468, %r2715; mul.ftz.f32 %f11469, %f11468, 0f40490000; sub.ftz.f32 %f11470, %f11454, %f11469; mul.ftz.f32 %f11471, %f11468, 0f3A7DA000; sub.ftz.f32 %f11472, %f11470, %f11471; mul.ftz.f32 %f11473, %f11468, 0f34222000; sub.ftz.f32 %f11474, %f11472, %f11473; mul.ftz.f32 %f11475, %f11468, 0f2CB4611A; sub.ftz.f32 %f11476, %f11474, %f11475; add.ftz.f32 %f11477, %f11476, 0fBFC90FDB; add.ftz.f32 %f11478, %f11477, 0f3FC90FDB; mul.ftz.f32 %f11479, %f11478, %f11478; and.b32 %r2716, %r2715, 1; setp.eq.b32 %p1647, %r2716, 1; neg.ftz.f32 %f11480, %f11478; selp.f32 %f11481, %f11480, %f11478, %p1647; fma.rn.ftz.f32 %f11482, %f11479, 0f362F0A87, 0fB94FBAEB; fma.rn.ftz.f32 %f11483, %f11479, %f11482, 0f3C088777; fma.rn.ftz.f32 %f11484, %f11479, %f11483, 0fBE2AAAA6; mul.ftz.f32 %f11485, %f11481, %f11484; fma.rn.ftz.f32 %f11486, %f11479, %f11485, %f11481; mul.ftz.f32 %f11487, %f11479, 0f3491ED89; sub.ftz.f32 %f11488, %f11436, %f11487; fma.rn.ftz.f32 %f11489, %f11479, %f11488, 0fBAB60B58; fma.rn.ftz.f32 %f11490, %f11479, %f11489, 0f3D2AAAAA; fma.rn.ftz.f32 %f11491, %f11479, %f11490, 0fBF000000; fma.rn.ftz.f32 %f11492, %f11479, %f11491, 0f3F800000; neg.ftz.f32 %f11493, %f11492; selp.f32 %f11494, %f11493, %f11492, %p1647; abs.ftz.f32 %f11495, %f11486; setp.gt.ftz.f32 %p1648, %f11495, 0f3F800000; selp.f32 %f11496, 0f00000000, %f11486, %p1648; abs.ftz.f32 %f11497, %f11494; setp.gt.ftz.f32 %p1649, %f11497, 0f3F800000; selp.f32 %f11498, 0f00000000, %f11494, %p1649; mul.ftz.f32 %f11499, %f11447, %f11498; mul.ftz.f32 %f11500, %f2816, %f11499; mul.ftz.f32 %f11501, %f2817, %f11499; mul.ftz.f32 %f11502, %f2818, %f11499; mul.ftz.f32 %f11503, %f11447, %f11496; mul.ftz.f32 %f11504, %f2814, %f2818; mul.ftz.f32 %f11505, %f2815, %f2817; sub.ftz.f32 %f11506, %f11505, %f11504; mul.ftz.f32 %f11507, %f11506, %f11503; mul.ftz.f32 %f11508, %f2815, %f2816; mul.ftz.f32 %f11509, %f2813, %f2818; sub.ftz.f32 %f11510, %f11509, %f11508; mul.ftz.f32 %f11511, %f11510, %f11503; mul.ftz.f32 %f11512, %f2813, %f2817; mul.ftz.f32 %f11513, %f2814, %f2816; sub.ftz.f32 %f11514, %f11513, %f11512; mul.ftz.f32 %f11515, %f11514, %f11503; sub.ftz.f32 %f11516, %f11500, %f11507; sub.ftz.f32 %f11517, %f11501, %f11511; sub.ftz.f32 %f11518, %f11502, %f11515; fma.rn.ftz.f32 %f16740, %f2813, %f11445, %f11516; fma.rn.ftz.f32 %f16741, %f2814, %f11445, %f11517; fma.rn.ftz.f32 %f16742, %f2815, %f11445, %f11518; mul.ftz.f32 %f11519, %f11460, %f11464; abs.ftz.f32 %f11520, %f11519; abs.ftz.f32 %f11521, %f11412; sub.ftz.f32 %f11522, %f11343, %f11521; setp.lt.ftz.f32 %p1650, %f11522, 0f3A83126F; selp.f32 %f16743, 0f00000000, %f11520, %p1650; mov.u32 %r3285, 9; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; $L__BB108_2309: and.b32 %r2903, %r3285, 1; setp.eq.b32 %p2061, %r2903, 1; mov.pred %p2062, 0; xor.pred %p2063, %p2061, %p2062; not.pred %p2064, %p2063; @%p2064 bra $L__BB108_2317; bra.uni $L__BB108_2310; $L__BB108_2317: ld.const.u64 %rd2690, [kernel_params+2160]; cvta.to.global.u64 %rd2691, %rd2690; ld.local.u32 %r2913, [%rd7+-12]; mul.wide.s32 %rd2692, %r2913, 208; add.s64 %rd2693, %rd2691, %rd2692; ld.global.f32 %f4137, [%rd2693+180]; setp.leu.ftz.f32 %p2075, %f4137, 0f3F800000; @%p2075 bra $L__BB108_2319; ld.local.f32 %f15205, [%rd587+8]; ld.local.f32 %f15206, [%rd587+12]; mul.ftz.f32 %f15207, %f16741, %f15206; fma.rn.ftz.f32 %f15208, %f16740, %f15205, %f15207; ld.local.f32 %f15209, [%rd587+16]; fma.rn.ftz.f32 %f15210, %f16742, %f15209, %f15208; mov.f32 %f15211, 0f3F800000; min.ftz.f32 %f15212, %f15210, %f15211; abs.ftz.f32 %f15213, %f15212; setp.lt.ftz.f32 %p2076, %f15213, 0f3F800000; add.ftz.f32 %f15214, %f15213, 0fBF800000; add.ftz.f32 %f15215, %f15214, 0f3F800000; selp.f32 %f15216, %f15215, 0f3F800000, %p2076; sub.ftz.f32 %f15217, %f15211, %f15216; sqrt.approx.ftz.f32 %f15218, %f15217; fma.rn.ftz.f32 %f15219, %f15216, 0fBCB1485D, 0f3D9FB43B; fma.rn.ftz.f32 %f15220, %f15216, %f15219, 0fBE5A6B93; fma.rn.ftz.f32 %f15221, %f15216, %f15220, 0f3FC90FDB; mul.ftz.f32 %f15222, %f15218, %f15221; setp.lt.ftz.f32 %p2077, %f15212, 0f00000000; mov.f32 %f15223, 0f00000000; mov.f32 %f15224, 0f40490FDB; sub.ftz.f32 %f15225, %f15224, %f15222; selp.f32 %f15226, %f15225, %f15222, %p2077; mul.ftz.f32 %f15227, %f4137, %f15226; cos.approx.ftz.f32 %f15228, %f15227; max.ftz.f32 %f15229, %f15228, %f15223; div.approx.ftz.f32 %f15230, %f15229, %f15212; mul.ftz.f32 %f16744, %f16744, %f15230; mul.ftz.f32 %f16745, %f16745, %f15230; mul.ftz.f32 %f16746, %f16746, %f15230; $L__BB108_2319: and.b32 %r2914, %r3285, 4; setp.eq.s32 %p2078, %r2914, 0; @%p2078 bra $L__BB108_2330; ld.local.f32 %f4144, [%rd587+16]; ld.local.f32 %f4145, [%rd7+-68]; ld.local.f32 %f4146, [%rd7+-76]; ld.local.f32 %f4147, [%rd587+8]; setp.neu.ftz.f32 %p2079, %f4147, %f4146; mov.u16 %rs160, 0; @%p2079 bra $L__BB108_2323; ld.local.f32 %f15231, [%rd587+12]; ld.local.f32 %f15232, [%rd7+-72]; setp.neu.ftz.f32 %p2080, %f15231, %f15232; @%p2080 bra $L__BB108_2323; setp.eq.ftz.f32 %p2081, %f4144, %f4145; selp.u16 %rs160, 1, 0, %p2081; $L__BB108_2323: setp.ne.s16 %p2082, %rs160, 0; @%p2082 bra $L__BB108_2330; ld.local.f32 %f15234, [%rd587+12]; mul.ftz.f32 %f15235, %f16741, %f15234; fma.rn.ftz.f32 %f15236, %f16740, %f4147, %f15235; fma.rn.ftz.f32 %f15237, %f16742, %f4144, %f15236; setp.lt.ftz.f32 %p2083, %f15237, 0f00000000; mov.f32 %f16751, 0f00000000; ld.local.f32 %f15238, [%rd7+-72]; neg.ftz.f32 %f15239, %f15238; neg.ftz.f32 %f15240, %f4146; selp.f32 %f15241, %f15240, %f4146, %p2083; selp.f32 %f15242, %f15239, %f15238, %p2083; neg.ftz.f32 %f15243, %f4145; selp.f32 %f15244, %f15243, %f4145, %p2083; mul.ftz.f32 %f15245, %f16741, %f15242; fma.rn.ftz.f32 %f15246, %f16740, %f15241, %f15245; fma.rn.ftz.f32 %f4148, %f16742, %f15244, %f15246; mul.ftz.f32 %f15247, %f15234, %f15242; fma.rn.ftz.f32 %f15248, %f4147, %f15241, %f15247; fma.rn.ftz.f32 %f15249, %f4144, %f15244, %f15248; mul.ftz.f32 %f4149, %f15237, %f15249; setp.eq.ftz.f32 %p2084, %f4149, 0f00000000; @%p2084 bra $L__BB108_2326; div.approx.ftz.f32 %f16751, %f4148, %f4149; $L__BB108_2326: setp.ge.ftz.f32 %p2085, %f16751, 0f3F800000; mov.f32 %f16752, 0f3F800000; @%p2085 bra $L__BB108_2329; setp.lt.ftz.f32 %p2086, %f16751, 0f00000000; mov.f32 %f16752, 0f00000000; @%p2086 bra $L__BB108_2329; mul.ftz.f32 %f15252, %f16751, %f16751; mul.ftz.f32 %f15253, %f16751, %f15252; sub.ftz.f32 %f15254, %f15252, %f15253; add.ftz.f32 %f16752, %f16751, %f15254; $L__BB108_2329: mul.ftz.f32 %f16744, %f16744, %f16752; mul.ftz.f32 %f16745, %f16745, %f16752; mul.ftz.f32 %f16746, %f16746, %f16752; bra.uni $L__BB108_2330; $L__BB108_2310: ld.const.f32 %f4134, [kernel_params+1040]; setp.ltu.ftz.f32 %p2065, %f4134, 0f00000000; shr.u32 %r2904, %r3285, 2; and.b32 %r2905, %r2904, 1; setp.eq.b32 %p2066, %r2905, 1; or.pred %p2067, %p2066, %p2065; @%p2067 bra $L__BB108_2330; mov.f32 %f16747, 0f00000000; setp.gt.u32 %p2068, %r873, 31; @%p2068 bra $L__BB108_2313; mov.u32 %r2907, 1; shl.b32 %r2908, %r2907, %r873; and.b32 %r2909, %r2908, -2145385984; setp.ne.s32 %p2069, %r2909, 0; @%p2069 bra $L__BB108_2316; $L__BB108_2313: add.s32 %r2910, %r873, -10; setp.lt.u32 %p2070, %r2910, 7; @%p2070 bra $L__BB108_2315; add.s32 %r2911, %r873, -22; setp.gt.u32 %p2071, %r2911, 2; setp.ne.s32 %p2072, %r873, 27; mov.f32 %f16747, 0f3F800000; and.pred %p2073, %p2071, %p2072; @%p2073 bra $L__BB108_2316; $L__BB108_2315: ld.local.v2.f32 {%f15201, %f15202}, [%rd587+20]; mul.ftz.f32 %f16747, %f15201, %f15202; $L__BB108_2316: or.b32 %r2912, %r3285, 128; setp.le.ftz.f32 %p2074, %f16747, %f4134; selp.b32 %r3285, %r2912, %r3285, %p2074; $L__BB108_2330: add.s32 %r907, %r873, -2; add.u64 %rd2694, %SP, 704; add.u64 %rd591, %SPL, 704; setp.neu.ftz.f32 %p2087, %f16743, 0f00000000; @%p2087 bra $L__BB108_2336; bra.uni $L__BB108_2331; $L__BB108_2336: setp.lt.u32 %p2092, %r907, 7; ld.local.f32 %f15255, [%rd587+-12]; mul.ftz.f32 %f4160, %f16744, %f15255; ld.local.f32 %f15256, [%rd587+-8]; mul.ftz.f32 %f4161, %f16745, %f15256; ld.local.f32 %f15257, [%rd587+-4]; mul.ftz.f32 %f4162, %f16746, %f15257; mov.u32 %r2920, 0; st.local.u32 [%rd591], %r2920; st.local.u32 [%rd591+4], %r2920; st.local.u32 [%rd591+8], %r2920; st.local.u32 [%rd591+12], %r2920; st.local.u32 [%rd591+16], %r2920; st.local.u32 [%rd591+20], %r2920; @%p2092 bra $L__BB108_2339; bra.uni $L__BB108_2337; $L__BB108_2339: st.local.f32 [%rd591], %f4160; st.local.f32 [%rd591+4], %f4161; st.local.f32 [%rd591+8], %f4162; bra.uni $L__BB108_2340; $L__BB108_2331: setp.lt.u32 %p2088, %r907, 7; mov.u32 %r2915, 0; st.local.u32 [%rd591], %r2915; st.local.u32 [%rd591+4], %r2915; st.local.u32 [%rd591+8], %r2915; st.local.u32 [%rd591+12], %r2915; st.local.u32 [%rd591+16], %r2915; st.local.u32 [%rd591+20], %r2915; @%p2088 bra $L__BB108_2334; bra.uni $L__BB108_2332; $L__BB108_2334: st.local.u32 [%rd591], %r2915; st.local.u32 [%rd591+4], %r2915; st.local.u32 [%rd591+8], %r2915; bra.uni $L__BB108_2335; $L__BB108_2337: add.s32 %r2921, %r873, -9; setp.gt.u32 %p2093, %r2921, 11; setp.ne.s32 %p2094, %r873, 29; and.pred %p2095, %p2094, %p2093; @%p2095 bra $L__BB108_2340; st.local.f32 [%rd591+12], %f4160; st.local.f32 [%rd591+16], %f4161; st.local.f32 [%rd591+20], %f4162; $L__BB108_2340: st.local.f32 [%rd591+24], %f4160; st.local.f32 [%rd591+28], %f4161; st.local.f32 [%rd591+32], %f4162; @%p1586 bra $L__BB108_2342; add.u64 %rd2922, %SP, 1728; ld.local.f32 %f15258, [%rd587+4]; mul.ftz.f32 %f15259, %f16743, %f15258; mul.lo.s64 %rd2695, %rd586, 80; add.s64 %rd2697, %rd2922, %rd2695; add.s64 %rd2698, %rd2697, 304; { // callseq 852, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd2922; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f16740; st.param.f32 [param1+4], %f16741; st.param.f32 [param1+8], %f16742; .param .b64 param2; st.param.b64 [param2+0], %rd2698; .param .b64 param3; st.param.b64 [param3+0], %rd2694; .param .b32 param4; st.param.f32 [param4+0], %f15259; .param .b32 param5; st.param.f32 [param5+0], %f15258; .param .b32 retval0; call.uni (retval0), _Z29_surface_shader_bsdf_eval_misPK16KernelGlobalsGPUP10ShaderData6float3PK13ShaderClosureP8BsdfEvalffj, ( param0, param1, param2, param3, param4, param5 ); ld.param.f32 %f16743, [retval0+0]; } // callseq 852 bra.uni $L__BB108_2342; $L__BB108_2332: add.s32 %r2916, %r873, -9; setp.gt.u32 %p2089, %r2916, 11; setp.ne.s32 %p2090, %r873, 29; and.pred %p2091, %p2090, %p2089; @%p2091 bra $L__BB108_2335; st.local.u32 [%rd591+12], %r2915; st.local.u32 [%rd591+16], %r2915; st.local.u32 [%rd591+20], %r2915; $L__BB108_2335: st.local.u32 [%rd591+24], %r2915; st.local.u32 [%rd591+28], %r2915; st.local.u32 [%rd591+32], %r2915; $L__BB108_2342: setp.eq.ftz.f32 %p2097, %f16743, 0f00000000; mov.u32 %r3294, 0; @%p2097 bra $L__BB108_2405; ld.local.f32 %f4165, [%rd591+24]; setp.eq.ftz.f32 %p2098, %f4165, 0f00000000; ld.local.f32 %f4166, [%rd591+28]; setp.eq.ftz.f32 %p2099, %f4166, 0f00000000; and.pred %p2100, %p2098, %p2099; ld.local.f32 %f4167, [%rd591+32]; setp.eq.ftz.f32 %p2101, %f4167, 0f00000000; and.pred %p2102, %p2101, %p2100; @%p2102 bra $L__BB108_2405; and.b32 %r908, %r3285, 32; setp.eq.s32 %p2103, %r908, 0; @%p2103 bra $L__BB108_2346; bra.uni $L__BB108_2345; $L__BB108_2346: mul.ftz.f32 %f15263, %f16741, %f16741; fma.rn.ftz.f32 %f15264, %f16740, %f16740, %f15263; fma.rn.ftz.f32 %f15265, %f16742, %f16742, %f15264; sqrt.approx.ftz.f32 %f15266, %f15265; div.approx.ftz.f32 %f4168, %f16740, %f15266; div.approx.ftz.f32 %f4169, %f16741, %f15266; div.approx.ftz.f32 %f4170, %f16742, %f15266; ld.local.v4.f32 {%f16775, %f16776, %f16777, %f15270}, [%rd1]; ld.local.u32 %r2926, [%rd7+-24]; and.b32 %r2927, %r2926, 1; setp.eq.b32 %p2105, %r2927, 1; xor.pred %p2107, %p2105, %p2062; not.pred %p2108, %p2107; @%p2108 bra $L__BB108_2354; ld.const.u64 %rd2706, [kernel_params+2232]; cvta.to.global.u64 %rd2707, %rd2706; ld.local.v2.u32 {%r2928, %r2929}, [%rd7+-32]; mul.wide.s32 %rd2708, %r2929, 16; add.s64 %rd2709, %rd2707, %rd2708; ld.const.u64 %rd2710, [kernel_params+2256]; cvta.to.global.u64 %rd2711, %rd2710; ld.global.u32 %r2931, [%rd2709+12]; mul.wide.u32 %rd2712, %r2931, 12; add.s64 %rd2713, %rd2711, %rd2712; ld.global.f32 %f4177, [%rd2713]; ld.global.f32 %f4178, [%rd2713+4]; ld.global.f32 %f4179, [%rd2713+8]; add.s32 %r2932, %r2931, 1; mul.wide.u32 %rd2714, %r2932, 12; add.s64 %rd2715, %rd2711, %rd2714; ld.global.f32 %f4180, [%rd2715]; ld.global.f32 %f4181, [%rd2715+4]; ld.global.f32 %f4182, [%rd2715+8]; add.s32 %r2933, %r2931, 2; mul.wide.u32 %rd2716, %r2933, 12; add.s64 %rd2717, %rd2711, %rd2716; ld.global.f32 %f4183, [%rd2717]; ld.global.f32 %f4184, [%rd2717+4]; ld.global.f32 %f4185, [%rd2717+8]; and.b32 %r2934, %r2928, 4; setp.ne.s32 %p2109, %r2934, 0; mov.f32 %f16769, %f16775; mov.f32 %f16770, %f16776; mov.f32 %f16771, %f16777; mov.f32 %f16772, %f4168; mov.f32 %f16773, %f4169; mov.f32 %f16774, %f4170; @%p2109 bra $L__BB108_2352; and.b32 %r2935, %r2928, 2; setp.eq.s32 %p2110, %r2935, 0; @%p2110 bra $L__BB108_2350; bra.uni $L__BB108_2349; $L__BB108_2350: ld.const.u64 %rd2718, [kernel_params+2160]; cvta.to.global.u64 %rd2719, %rd2718; ld.local.u32 %r2936, [%rd7+-12]; mul.wide.s32 %rd2720, %r2936, 208; add.s64 %rd2721, %rd2719, %rd2720; ld.global.v4.f32 {%f16757, %f16758, %f16759, %f16760}, [%rd2721+48]; ld.global.v4.f32 {%f16761, %f16762, %f16763, %f16764}, [%rd2721+64]; ld.global.v4.f32 {%f16765, %f16766, %f16767, %f16768}, [%rd2721+80]; bra.uni $L__BB108_2351; $L__BB108_2345: ld.local.f32 %f15260, [%rd7]; setp.eq.ftz.f32 %p2104, %f15260, 0f00000000; mov.b32 %r2924, %f15260; add.s32 %r2925, %r2924, 1; mov.b32 %f15261, %r2925; selp.f32 %f15262, 0f00000001, %f15261, %p2104; ld.const.u64 %rd2700, [kernel_params+2648]; cvta.to.global.u64 %rd2701, %rd2700; add.s64 %rd2703, %rd2701, %rd837; st.global.f32 [%rd2703], %f15262; bra.uni $L__BB108_2355; $L__BB108_2349: ld.local.v4.f32 {%f16757, %f16758, %f16759, %f16760}, [%rd7+104]; ld.local.v4.f32 {%f16761, %f16762, %f16763, %f16764}, [%rd7+120]; ld.local.v4.f32 {%f16765, %f16766, %f16767, %f16768}, [%rd7+136]; $L__BB108_2351: mul.ftz.f32 %f15295, %f16776, %f16758; fma.rn.ftz.f32 %f15296, %f16775, %f16757, %f15295; fma.rn.ftz.f32 %f15297, %f16777, %f16759, %f15296; mul.ftz.f32 %f15298, %f16776, %f16762; fma.rn.ftz.f32 %f15299, %f16775, %f16761, %f15298; fma.rn.ftz.f32 %f15300, %f16777, %f16763, %f15299; mul.ftz.f32 %f15301, %f16776, %f16766; fma.rn.ftz.f32 %f15302, %f16775, %f16765, %f15301; fma.rn.ftz.f32 %f15303, %f16777, %f16767, %f15302; add.ftz.f32 %f16769, %f16760, %f15297; add.ftz.f32 %f16770, %f16764, %f15300; add.ftz.f32 %f16771, %f16768, %f15303; mul.ftz.f32 %f15304, %f4169, %f16758; fma.rn.ftz.f32 %f15305, %f4168, %f16757, %f15304; mul.ftz.f32 %f15306, %f4169, %f16762; fma.rn.ftz.f32 %f15307, %f4168, %f16761, %f15306; mul.ftz.f32 %f15308, %f4169, %f16766; fma.rn.ftz.f32 %f15309, %f4168, %f16765, %f15308; fma.rn.ftz.f32 %f16772, %f4170, %f16759, %f15305; fma.rn.ftz.f32 %f16773, %f4170, %f16763, %f15307; fma.rn.ftz.f32 %f16774, %f4170, %f16767, %f15309; $L__BB108_2352: sub.ftz.f32 %f15310, %f4177, %f16769; sub.ftz.f32 %f15311, %f4178, %f16770; sub.ftz.f32 %f15312, %f4179, %f16771; sub.ftz.f32 %f15313, %f4180, %f16769; sub.ftz.f32 %f15314, %f4181, %f16770; sub.ftz.f32 %f15315, %f4182, %f16771; sub.ftz.f32 %f15316, %f4183, %f16769; sub.ftz.f32 %f15317, %f4184, %f16770; sub.ftz.f32 %f15318, %f4185, %f16771; sub.ftz.f32 %f15319, %f15316, %f15310; sub.ftz.f32 %f15320, %f15317, %f15311; sub.ftz.f32 %f15321, %f15318, %f15312; sub.ftz.f32 %f15322, %f15310, %f15313; sub.ftz.f32 %f15323, %f15311, %f15314; sub.ftz.f32 %f15324, %f15312, %f15315; sub.ftz.f32 %f15325, %f15313, %f15316; sub.ftz.f32 %f15326, %f15314, %f15317; sub.ftz.f32 %f15327, %f15315, %f15318; add.ftz.f32 %f15328, %f15316, %f15310; add.ftz.f32 %f15329, %f15317, %f15311; add.ftz.f32 %f15330, %f15318, %f15312; mul.ftz.f32 %f15331, %f15329, %f15321; mul.ftz.f32 %f15332, %f15330, %f15320; sub.ftz.f32 %f15333, %f15331, %f15332; mul.ftz.f32 %f15334, %f15330, %f15319; mul.ftz.f32 %f15335, %f15328, %f15321; sub.ftz.f32 %f15336, %f15334, %f15335; mul.ftz.f32 %f15337, %f15328, %f15320; mul.ftz.f32 %f15338, %f15329, %f15319; sub.ftz.f32 %f15339, %f15337, %f15338; mul.ftz.f32 %f15340, %f16773, %f15336; fma.rn.ftz.f32 %f15341, %f16772, %f15333, %f15340; fma.rn.ftz.f32 %f15342, %f16774, %f15339, %f15341; add.ftz.f32 %f15343, %f15310, %f15313; add.ftz.f32 %f15344, %f15311, %f15314; add.ftz.f32 %f15345, %f15312, %f15315; mul.ftz.f32 %f15346, %f15344, %f15324; mul.ftz.f32 %f15347, %f15345, %f15323; sub.ftz.f32 %f15348, %f15346, %f15347; mul.ftz.f32 %f15349, %f15345, %f15322; mul.ftz.f32 %f15350, %f15343, %f15324; sub.ftz.f32 %f15351, %f15349, %f15350; mul.ftz.f32 %f15352, %f15343, %f15323; mul.ftz.f32 %f15353, %f15344, %f15322; sub.ftz.f32 %f15354, %f15352, %f15353; mul.ftz.f32 %f15355, %f16773, %f15351; fma.rn.ftz.f32 %f15356, %f16772, %f15348, %f15355; fma.rn.ftz.f32 %f15357, %f16774, %f15354, %f15356; add.ftz.f32 %f15358, %f15313, %f15316; add.ftz.f32 %f15359, %f15314, %f15317; add.ftz.f32 %f15360, %f15315, %f15318; mul.ftz.f32 %f15361, %f15359, %f15327; mul.ftz.f32 %f15362, %f15360, %f15326; sub.ftz.f32 %f15363, %f15361, %f15362; mul.ftz.f32 %f15364, %f15360, %f15325; mul.ftz.f32 %f15365, %f15358, %f15327; sub.ftz.f32 %f15366, %f15364, %f15365; mul.ftz.f32 %f15367, %f15358, %f15326; mul.ftz.f32 %f15368, %f15359, %f15325; sub.ftz.f32 %f15369, %f15367, %f15368; mul.ftz.f32 %f15370, %f16773, %f15366; fma.rn.ftz.f32 %f15371, %f16772, %f15363, %f15370; fma.rn.ftz.f32 %f15372, %f16774, %f15369, %f15371; add.ftz.f32 %f15373, %f15342, %f15357; add.ftz.f32 %f15374, %f15372, %f15373; abs.ftz.f32 %f15375, %f15374; mul.ftz.f32 %f15376, %f15375, 0f34000000; min.ftz.f32 %f15377, %f15357, %f15372; min.ftz.f32 %f15378, %f15342, %f15377; max.ftz.f32 %f15379, %f15357, %f15372; max.ftz.f32 %f15380, %f15342, %f15379; setp.ge.ftz.f32 %p2111, %f15378, %f15376; mul.ftz.f32 %f15381, %f15375, 0fB4000000; setp.le.ftz.f32 %p2112, %f15380, %f15381; or.pred %p2113, %p2111, %p2112; @%p2113 bra $L__BB108_2354; ld.local.v2.f32 {%f15382, %f15383}, [%rd7+-64]; mul.ftz.f32 %f15386, %f15382, 0f43800000; cvt.rzi.ftz.s32.f32 %r2937, %f15386; mul.ftz.f32 %f15387, %f15383, 0f43800000; cvt.rzi.ftz.s32.f32 %r2938, %f15387; ld.local.f32 %f15388, [%rd7+-56]; mul.ftz.f32 %f15389, %f15388, 0f43800000; cvt.rzi.ftz.s32.f32 %r2939, %f15389; neg.s32 %r2940, %r2937; setp.lt.ftz.f32 %p2114, %f16775, 0f00000000; selp.b32 %r2941, %r2940, %r2937, %p2114; mov.b32 %r2942, %f16775; add.s32 %r2943, %r2941, %r2942; mov.b32 %f15390, %r2943; neg.s32 %r2944, %r2938; setp.lt.ftz.f32 %p2115, %f16776, 0f00000000; selp.b32 %r2945, %r2944, %r2938, %p2115; mov.b32 %r2946, %f16776; add.s32 %r2947, %r2945, %r2946; mov.b32 %f15391, %r2947; neg.s32 %r2948, %r2939; setp.lt.ftz.f32 %p2116, %f16777, 0f00000000; selp.b32 %r2949, %r2948, %r2939, %p2116; mov.b32 %r2950, %f16777; add.s32 %r2951, %r2949, %r2950; mov.b32 %f15392, %r2951; abs.ftz.f32 %f15393, %f16775; setp.lt.ftz.f32 %p2117, %f15393, 0f3D000000; fma.rn.ftz.f32 %f15394, %f15382, 0f37800000, %f16775; abs.ftz.f32 %f15395, %f16776; setp.lt.ftz.f32 %p2118, %f15395, 0f3D000000; fma.rn.ftz.f32 %f15396, %f15383, 0f37800000, %f16776; abs.ftz.f32 %f15397, %f16777; setp.lt.ftz.f32 %p2119, %f15397, 0f3D000000; fma.rn.ftz.f32 %f15398, %f15388, 0f37800000, %f16777; selp.f32 %f16777, %f15398, %f15392, %p2119; selp.f32 %f16776, %f15396, %f15391, %p2118; selp.f32 %f16775, %f15394, %f15390, %p2117; $L__BB108_2354: ld.const.u64 %rd2906, [kernel_params+2632]; cvta.to.global.u64 %rd2905, %rd2906; mul.lo.s64 %rd2723, %rd2, 12; add.s64 %rd2724, %rd2905, %rd2723; st.global.f32 [%rd2724], %f16775; st.global.f32 [%rd2724+4], %f16776; st.global.f32 [%rd2724+8], %f16777; ld.const.v2.u64 {%rd2725, %rd2726}, [kernel_params+2640]; cvta.to.global.u64 %rd2729, %rd2725; add.s64 %rd2730, %rd2729, %rd2723; st.global.f32 [%rd2730], %f4168; st.global.f32 [%rd2730+4], %f4169; st.global.f32 [%rd2730+8], %f4170; cvta.to.global.u64 %rd2731, %rd2726; add.s64 %rd2733, %rd2731, %rd837; mov.u32 %r2952, 0; st.global.u32 [%rd2733], %r2952; ld.const.v2.u64 {%rd2734, %rd2735}, [kernel_params+2656]; cvta.to.global.u64 %rd2737, %rd2734; add.s64 %rd2738, %rd2737, %rd837; mov.u32 %r2953, 2139095039; st.global.u32 [%rd2738], %r2953; ld.local.f32 %f15399, [%rd7+4]; ld.const.u64 %rd2739, [kernel_params+2672]; cvta.to.global.u64 %rd2740, %rd2739; add.s64 %rd2741, %rd2740, %rd837; st.global.f32 [%rd2741], %f15399; $L__BB108_2355: div.approx.ftz.f32 %f15400, %f4165, %f16743; ld.const.u64 %rd2742, [kernel_params+2584]; cvta.to.global.u64 %rd2743, %rd2742; mul.lo.s64 %rd2744, %rd2, 12; add.s64 %rd2745, %rd2743, %rd2744; ld.global.f32 %f15401, [%rd2745]; mul.ftz.f32 %f15402, %f15400, %f15401; ld.global.f32 %f15403, [%rd2745+4]; div.approx.ftz.f32 %f15404, %f4166, %f16743; mul.ftz.f32 %f15405, %f15404, %f15403; ld.global.f32 %f15406, [%rd2745+8]; div.approx.ftz.f32 %f15407, %f4167, %f16743; mul.ftz.f32 %f15408, %f15407, %f15406; st.global.f32 [%rd2745], %f15402; st.global.f32 [%rd2745+4], %f15405; st.global.f32 [%rd2745+8], %f15408; and.b32 %r2954, %r787, 2097152; setp.eq.s32 %p2120, %r2954, 0; @%p2120 bra $L__BB108_2370; ld.const.u64 %rd2746, [kernel_params+2464]; cvta.to.global.u64 %rd2747, %rd2746; add.s64 %rd2749, %rd2747, %rd1707; ld.global.u16 %rs125, [%rd2749]; setp.ne.s16 %p2121, %rs125, 0; @%p2121 bra $L__BB108_2370; ld.const.u64 %rd593, [kernel_params+2600]; ld.local.f32 %f4243, [%rd591+4]; ld.local.f32 %f4244, [%rd591+8]; mov.f32 %f16783, 0f00000000; mov.f32 %f16778, %f16783; @%p2098 bra $L__BB108_2359; ld.local.f32 %f15410, [%rd591]; div.approx.ftz.f32 %f16778, %f15410, %f4165; $L__BB108_2359: mov.f32 %f16779, %f16783; @%p2099 bra $L__BB108_2361; div.approx.ftz.f32 %f16779, %f4243, %f4166; $L__BB108_2361: mov.f32 %f16780, %f16783; @%p2101 bra $L__BB108_2363; div.approx.ftz.f32 %f16780, %f4244, %f4167; $L__BB108_2363: cvta.to.global.u64 %rd2750, %rd593; add.s64 %rd2752, %rd2750, %rd2744; st.global.f32 [%rd2752], %f16778; st.global.f32 [%rd2752+4], %f16779; st.global.f32 [%rd2752+8], %f16780; ld.const.u64 %rd594, [kernel_params+2608]; ld.local.f32 %f4251, [%rd591+16]; ld.local.f32 %f4252, [%rd591+20]; mov.f32 %f16781, %f16783; @%p2098 bra $L__BB108_2365; ld.local.f32 %f15414, [%rd591+12]; div.approx.ftz.f32 %f16781, %f15414, %f4165; $L__BB108_2365: mov.f32 %f16782, %f16783; @%p2099 bra $L__BB108_2367; div.approx.ftz.f32 %f16782, %f4251, %f4166; $L__BB108_2367: @%p2101 bra $L__BB108_2369; div.approx.ftz.f32 %f16783, %f4252, %f4167; $L__BB108_2369: cvta.to.global.u64 %rd2753, %rd594; add.s64 %rd2755, %rd2753, %rd2744; st.global.f32 [%rd2755], %f16781; st.global.f32 [%rd2755+4], %f16782; st.global.f32 [%rd2755+8], %f16783; $L__BB108_2370: setp.ne.s32 %p2128, %r908, 0; @%p2128 bra $L__BB108_2372; ld.const.v2.u64 {%rd2756, %rd2757}, [kernel_params+2560]; cvta.to.global.u64 %rd2760, %rd2756; add.s64 %rd2762, %rd2760, %rd837; st.global.f32 [%rd2762], %f16743; cvta.to.global.u64 %rd2763, %rd2757; add.s64 %rd2764, %rd2763, %rd837; ld.global.f32 %f15417, [%rd2764]; min.ftz.f32 %f15418, %f16743, %f15417; st.global.f32 [%rd2764], %f15418; $L__BB108_2372: ld.global.u32 %r910, [%rd29]; @%p2103 bra $L__BB108_2374; bra.uni $L__BB108_2373; $L__BB108_2374: ld.const.u64 %rd2769, [kernel_params+2464]; cvta.to.global.u64 %rd2770, %rd2769; add.s64 %rd595, %rd2770, %rd1707; ld.global.u16 %r2963, [%rd595]; add.s32 %r911, %r2963, 1; ld.const.u32 %r2964, [kernel_params+1516]; setp.lt.u32 %p2132, %r911, %r2964; or.b32 %r2965, %r910, 131072; selp.b32 %r912, %r910, %r2965, %p2132; and.b32 %r2966, %r3285, 64; setp.eq.s32 %p2133, %r2966, 0; @%p2133 bra $L__BB108_2376; bra.uni $L__BB108_2375; $L__BB108_2376: and.b32 %r2975, %r3285, 2; setp.eq.s32 %p2136, %r2975, 0; and.b32 %r3287, %r3285, 4; @%p2136 bra $L__BB108_2380; bra.uni $L__BB108_2377; $L__BB108_2380: and.b32 %r2984, %r912, -4096; or.b32 %r2985, %r2984, 4; and.b32 %r2986, %r3285, 128; setp.eq.s32 %p2140, %r2986, 0; and.b32 %r2987, %r2985, -20476; selp.b32 %r2988, %r2987, %r2985, %p2140; ld.const.u64 %rd2784, [kernel_params+2496]; cvta.to.global.u64 %rd2785, %rd2784; add.s64 %rd2787, %rd2785, %rd1707; ld.global.u16 %r2989, [%rd2787]; add.s32 %r2990, %r2989, 1; st.global.u16 [%rd2787], %r2990; or.b32 %r2991, %r2988, 131072; ld.const.u32 %r2992, [kernel_params+1528]; setp.lt.s32 %p2141, %r2990, %r2992; selp.b32 %r3288, %r2988, %r2991, %p2141; bra.uni $L__BB108_2381; $L__BB108_2373: ld.const.u64 %rd2765, [kernel_params+2472]; cvta.to.global.u64 %rd2766, %rd2765; add.s64 %rd2768, %rd2766, %rd1707; ld.global.u16 %r2955, [%rd2768]; add.s32 %r2956, %r2955, 1; ld.const.u32 %r2957, [kernel_params+1556]; setp.lt.u32 %p2130, %r2956, %r2957; selp.b32 %r2958, 64, 32832, %p2130; or.b32 %r2959, %r2958, %r910; ld.const.u32 %r2960, [kernel_params+1560]; setp.eq.s32 %p2131, %r2960, 0; or.b32 %r2961, %r2959, 2048; selp.b32 %r2962, %r2961, %r2959, %p2131; st.global.u32 [%rd29], %r2962; st.global.u16 [%rd2768], %r2956; ld.global.u16 %rs126, [%rd278]; add.s16 %rs127, %rs126, 16; st.global.u16 [%rd278], %rs127; mov.u32 %r3294, %r3285; bra.uni $L__BB108_2405; $L__BB108_2375: and.b32 %r2967, %r912, -20480; and.b32 %r2968, %r912, 100663296; setp.eq.s32 %p2134, %r2968, 0; selp.b32 %r2969, 67108992, 128, %p2134; or.b32 %r2970, %r2969, %r2967; ld.const.u64 %rd2772, [kernel_params+2504]; cvta.to.global.u64 %rd2773, %rd2772; add.s64 %rd2775, %rd2773, %rd1707; ld.global.u16 %r2971, [%rd2775]; add.s32 %r2972, %r2971, 1; st.global.u16 [%rd2775], %r2972; ld.const.u32 %r2973, [kernel_params+1532]; setp.lt.s32 %p2135, %r2972, %r2973; or.b32 %r2974, %r2970, 131072; selp.b32 %r3289, %r2970, %r2974, %p2135; bra.uni $L__BB108_2382; $L__BB108_2377: and.b32 %r2976, %r912, -20480; or.b32 %r915, %r2976, 2; or.b32 %r916, %r2976, 131074; setp.eq.s32 %p2137, %r3287, 0; @%p2137 bra $L__BB108_2379; bra.uni $L__BB108_2378; $L__BB108_2379: ld.const.u64 %rd2780, [kernel_params+2488]; cvta.to.global.u64 %rd2781, %rd2780; add.s64 %rd2783, %rd2781, %rd1707; ld.global.u16 %r2981, [%rd2783]; add.s32 %r2982, %r2981, 1; st.global.u16 [%rd2783], %r2982; ld.const.u32 %r2983, [kernel_params+1524]; setp.lt.s32 %p2139, %r2982, %r2983; selp.b32 %r3288, %r915, %r916, %p2139; mov.u32 %r3287, 0; bra.uni $L__BB108_2381; $L__BB108_2378: ld.const.u64 %rd2776, [kernel_params+2480]; cvta.to.global.u64 %rd2777, %rd2776; add.s64 %rd2779, %rd2777, %rd1707; ld.global.u16 %r2977, [%rd2779]; add.s32 %r2978, %r2977, 1; st.global.u16 [%rd2779], %r2978; ld.const.u32 %r2979, [kernel_params+1520]; setp.lt.s32 %p2138, %r2978, %r2979; selp.b32 %r3288, %r915, %r916, %p2138; $L__BB108_2381: and.b32 %r2993, %r3285, 8; setp.eq.s32 %p2142, %r2993, 0; selp.b32 %r2994, 2096, 16, %p2142; setp.eq.s32 %p2143, %r3287, 0; selp.b32 %r2995, %r2994, 4104, %p2143; or.b32 %r2996, %r2995, %r3288; and.b32 %r2997, %r3288, 100679680; setp.eq.s32 %p2144, %r2997, 0; or.b32 %r2998, %r2996, 33554432; selp.b32 %r3289, %r2998, %r2996, %p2144; $L__BB108_2382: st.global.u32 [%rd29], %r3289; add.s32 %r3039, %r2963, 1; st.global.u16 [%rd595], %r3039; ld.global.u16 %rs128, [%rd278]; add.s16 %rs129, %rs128, 16; st.global.u16 [%rd278], %rs129; mov.u32 %r3294, %r3285; $L__BB108_2405: and.b32 %r3022, %r3294, 1; setp.eq.b32 %p2159, %r3022, 1; mov.pred %p2160, 0; xor.pred %p2161, %p2159, %p2160; not.pred %p2162, %p2161; @%p2162 bra $L__BB108_2420; ld.local.u32 %r936, [%rd7+-36]; and.b32 %r3023, %r936, 262144; setp.eq.s32 %p2163, %r3023, 0; @%p2163 bra $L__BB108_2420; bra.uni $L__BB108_2407; $L__BB108_2420: setp.ne.s32 %p2180, %r3294, 0; selp.u16 %rs161, 1, 0, %p2180; $L__BB108_2421: cvt.u32.u16 %r3038, %rs161; st.param.b32 [func_retval0+0], %r3038; ret; $L__BB108_2407: and.b32 %r3024, %r936, 1; setp.eq.b32 %p2164, %r3024, 1; xor.pred %p2166, %p2164, %p2160; not.pred %p2167, %p2166; ld.const.u64 %rd3069, [kernel_params+2776]; cvta.to.global.u64 %rd2842, %rd3069; add.s64 %rd2844, %rd2842, %rd837; ld.global.u32 %r3296, [%rd2844]; @%p2167 bra $L__BB108_2414; bra.uni $L__BB108_2408; $L__BB108_2414: mov.u64 %rd2875, kernel_params; add.s64 %rd3075, %rd2875, 2768; setp.eq.s32 %p2174, %r3296, -1; mov.u32 %r3301, 0; @%p2174 bra $L__BB108_2418; ld.local.u32 %r946, [%rd7+-12]; add.s64 %rd3075, %rd2875, 2768; ld.local.u32 %r947, [%rd7+-40]; mov.u32 %r3301, 0; $L__BB108_2416: ld.const.u64 %rd2878, [%rd3075]; cvta.to.global.u64 %rd2879, %rd2878; add.s64 %rd2881, %rd2879, %rd837; ld.global.u32 %r3031, [%rd2881]; setp.eq.s32 %p2175, %r3031, %r946; setp.eq.s32 %p2176, %r3296, %r947; and.pred %p2177, %p2175, %p2176; @%p2177 bra $L__BB108_2420; add.s32 %r3301, %r3301, 1; mul.wide.s32 %rd2882, %r3301, 16; add.s64 %rd2884, %rd2875, %rd2882; add.s64 %rd3075, %rd2884, 2768; ld.const.u64 %rd3069, [%rd2884+2776]; cvta.to.global.u64 %rd2885, %rd3069; add.s64 %rd2887, %rd2885, %rd837; ld.global.u32 %r3296, [%rd2887]; setp.ne.s32 %p2178, %r3296, -1; @%p2178 bra $L__BB108_2416; $L__BB108_2418: ld.const.u32 %r3032, [kernel_params+12]; add.s32 %r3033, %r3032, -1; setp.ge.u32 %p2179, %r3301, %r3033; @%p2179 bra $L__BB108_2420; ld.local.u32 %r3034, [%rd7+-12]; ld.local.u32 %r3035, [%rd7+-40]; ld.const.u64 %rd2888, [%rd3075]; cvta.to.global.u64 %rd2889, %rd2888; add.s64 %rd2891, %rd2889, %rd837; st.global.u32 [%rd2891], %r3034; cvta.to.global.u64 %rd2892, %rd3069; add.s64 %rd2893, %rd2892, %rd837; st.global.u32 [%rd2893], %r3035; add.s32 %r3036, %r3301, 1; mul.wide.s32 %rd2894, %r3036, 16; add.s64 %rd2896, %rd2875, %rd2894; ld.const.v2.u64 {%rd2897, %rd2898}, [%rd2896+2768]; cvta.to.global.u64 %rd2901, %rd2897; add.s64 %rd2902, %rd2901, %rd837; mov.u32 %r3037, -1; st.global.u32 [%rd2902], %r3037; cvta.to.global.u64 %rd2903, %rd2898; add.s64 %rd2904, %rd2903, %rd837; st.global.u32 [%rd2904], %r3037; bra.uni $L__BB108_2420; $L__BB108_2408: setp.eq.s32 %p2168, %r3296, -1; @%p2168 bra $L__BB108_2420; ld.local.u32 %r938, [%rd7+-12]; mov.u64 %rd2846, kernel_params; add.s64 %rd3070, %rd2846, 2768; ld.local.u32 %r939, [%rd7+-40]; mov.u32 %r3297, 0; $L__BB108_2410: ld.const.u64 %rd2847, [%rd3070]; cvta.to.global.u64 %rd2848, %rd2847; add.s64 %rd2850, %rd2848, %rd837; ld.global.u32 %r3026, [%rd2850]; setp.eq.s32 %p2169, %r3026, %r938; setp.eq.s32 %p2170, %r3296, %r939; and.pred %p2171, %p2169, %p2170; add.s32 %r942, %r3297, 1; mul.wide.s32 %rd2851, %r942, 16; add.s64 %rd2853, %rd2846, %rd2851; add.s64 %rd605, %rd2853, 2776; @%p2171 bra $L__BB108_2412; add.s64 %rd3070, %rd2853, 2768; ld.const.u64 %rd3069, [%rd605]; cvta.to.global.u64 %rd2857, %rd3069; add.s64 %rd2859, %rd2857, %rd837; ld.global.u32 %r3296, [%rd2859]; setp.eq.s32 %p2172, %r3296, -1; mov.u32 %r3297, %r942; @%p2172 bra $L__BB108_2420; bra.uni $L__BB108_2410; $L__BB108_2412: ld.const.u64 %rd3072, [%rd605+-24]; $L__BB108_2413: add.s32 %r3297, %r3297, 1; mul.wide.s32 %rd2860, %r3297, 16; add.s64 %rd2862, %rd2846, %rd2860; ld.const.v2.u64 {%rd2863, %rd2864}, [%rd2862+2768]; cvta.to.global.u64 %rd2865, %rd2863; add.s64 %rd2867, %rd2865, %rd837; ld.global.u32 %r3027, [%rd2867]; cvta.to.global.u64 %rd2868, %rd2864; add.s64 %rd2869, %rd2868, %rd837; ld.global.u32 %r3028, [%rd2869]; cvta.to.global.u64 %rd2870, %rd3072; add.s64 %rd2871, %rd2870, %rd837; st.global.u32 [%rd2871], %r3027; cvta.to.global.u64 %rd2872, %rd3069; add.s64 %rd2873, %rd2872, %rd837; st.global.u32 [%rd2873], %r3028; setp.eq.s32 %p2173, %r3028, -1; mov.u64 %rd3069, %rd2864; mov.u64 %rd3072, %rd2863; @%p2173 bra $L__BB108_2420; bra.uni $L__BB108_2413; $L__BB108_741: mov.f32 %f6503, 0fBF800000; max.ftz.f32 %f6504, %f1019, %f6503; min.ftz.f32 %f6506, %f6504, %f6437; abs.ftz.f32 %f6507, %f6506; sub.ftz.f32 %f6508, %f6437, %f6507; mul.ftz.f32 %f6509, %f6508, 0f3F000000; sqrt.approx.ftz.f32 %f6510, %f6509; setp.gt.ftz.f32 %p633, %f6507, 0f3F11EB85; selp.f32 %f6511, %f6510, %f6507, %p633; mul.ftz.f32 %f6512, %f6511, %f6511; mov.f32 %f6513, 0f3C94D2E9; mov.f32 %f6514, 0f3D53F941; fma.rn.ftz.f32 %f6515, %f6514, %f6512, %f6513; mov.f32 %f6516, 0f3D3F841F; fma.rn.ftz.f32 %f6517, %f6515, %f6512, %f6516; mov.f32 %f6518, 0f3D994929; fma.rn.ftz.f32 %f6519, %f6517, %f6512, %f6518; mov.f32 %f6520, 0f3E2AAB94; fma.rn.ftz.f32 %f6521, %f6519, %f6512, %f6520; mul.ftz.f32 %f6522, %f6512, %f6521; fma.rn.ftz.f32 %f6523, %f6522, %f6511, %f6511; add.ftz.f32 %f6524, %f6523, %f6523; mov.f32 %f6525, 0f3FC90FDB; sub.ftz.f32 %f6526, %f6525, %f6523; selp.f32 %f6527, %f6524, %f6526, %p633; setp.lt.ftz.f32 %p634, %f6506, 0f00000000; sub.ftz.f32 %f6529, %f6449, %f6527; selp.f32 %f6530, %f6529, %f6527, %p634; mul.ftz.f32 %f6531, %f6490, %f1019; sub.ftz.f32 %f6532, %f6494, %f6531; mul.ftz.f32 %f6533, %f6491, %f1019; sub.ftz.f32 %f6534, %f6495, %f6533; mul.ftz.f32 %f6535, %f6492, %f1019; sub.ftz.f32 %f6536, %f6496, %f6535; mul.ftz.f32 %f6537, %f6493, %f1019; sub.ftz.f32 %f6538, %f6497, %f6537; mul.ftz.f32 %f6539, %f6534, %f6534; fma.rn.ftz.f32 %f6540, %f6532, %f6532, %f6539; mul.ftz.f32 %f6541, %f6538, %f6538; fma.rn.ftz.f32 %f6542, %f6536, %f6536, %f6541; add.ftz.f32 %f6543, %f6540, %f6542; rsqrt.approx.ftz.f32 %f6544, %f6543; mul.ftz.f32 %f6545, %f6532, %f6544; mul.ftz.f32 %f6546, %f6534, %f6544; mul.ftz.f32 %f6547, %f6536, %f6544; mul.ftz.f32 %f6548, %f6538, %f6544; mul.ftz.f32 %f6549, %f1010, %f6530; cos.approx.ftz.f32 %f6550, %f6549; sin.approx.ftz.f32 %f6551, %f6549; mul.ftz.f32 %f6552, %f6545, %f6551; mul.ftz.f32 %f6553, %f6546, %f6551; mul.ftz.f32 %f6554, %f6547, %f6551; mul.ftz.f32 %f6555, %f6548, %f6551; fma.rn.ftz.f32 %f15861, %f6493, %f6550, %f6555; fma.rn.ftz.f32 %f15860, %f6492, %f6550, %f6554; fma.rn.ftz.f32 %f15859, %f6491, %f6550, %f6553; fma.rn.ftz.f32 %f15858, %f6490, %f6550, %f6552; $L__BB108_743: ld.global.v4.f32 {%f6570, %f6571, %f6572, %f6573}, [%rd164+16]; ld.global.v4.f32 {%f6578, %f6579, %f6580, %f6581}, [%rd164+80]; mul.ftz.f32 %f6586, %f1010, %f6578; mul.ftz.f32 %f6587, %f1010, %f6579; mul.ftz.f32 %f6588, %f1010, %f6580; mul.ftz.f32 %f6589, %f1010, %f6581; fma.rn.ftz.f32 %f6590, %f1020, %f6573, %f6589; ld.global.v4.f32 {%f6591, %f6592, %f6593, %f6594}, [%rd164+32]; ld.global.v4.f32 {%f6599, %f6600, %f6601, %f6602}, [%rd164+96]; mul.ftz.f32 %f6607, %f1010, %f6599; mul.ftz.f32 %f6608, %f1010, %f6600; mul.ftz.f32 %f6609, %f1010, %f6601; mul.ftz.f32 %f6610, %f1010, %f6602; fma.rn.ftz.f32 %f6611, %f1020, %f6591, %f6607; fma.rn.ftz.f32 %f6612, %f1020, %f6592, %f6608; fma.rn.ftz.f32 %f6613, %f1020, %f6593, %f6609; fma.rn.ftz.f32 %f6614, %f1020, %f6594, %f6610; ld.global.v4.f32 {%f6615, %f6616, %f6617, %f6618}, [%rd164+48]; ld.global.v4.f32 {%f6623, %f6624, %f6625, %f6626}, [%rd164+112]; mul.ftz.f32 %f6631, %f1010, %f6623; mul.ftz.f32 %f6632, %f1010, %f6624; mul.ftz.f32 %f6633, %f1010, %f6625; mul.ftz.f32 %f6634, %f1010, %f6626; fma.rn.ftz.f32 %f6635, %f1020, %f6615, %f6631; fma.rn.ftz.f32 %f6636, %f1020, %f6616, %f6632; fma.rn.ftz.f32 %f6637, %f1020, %f6617, %f6633; fma.rn.ftz.f32 %f6638, %f1020, %f6618, %f6634; mul.ftz.f32 %f6639, %f15861, 0f3FB504F3; mul.ftz.f32 %f6640, %f15858, 0f3FB504F3; mul.ftz.f32 %f6641, %f15859, 0f3FB504F3; mul.ftz.f32 %f6642, %f15860, 0f3FB504F3; mul.ftz.f32 %f6643, %f6639, %f6640; mul.ftz.f32 %f6644, %f6639, %f6641; mul.ftz.f32 %f6645, %f6639, %f6642; mul.ftz.f32 %f6646, %f6640, %f6640; mul.ftz.f32 %f6647, %f6640, %f6641; mul.ftz.f32 %f6648, %f6640, %f6642; mul.ftz.f32 %f6649, %f6641, %f6641; mul.ftz.f32 %f6650, %f6641, %f6642; mul.ftz.f32 %f6651, %f6642, %f6642; sub.ftz.f32 %f6653, %f6437, %f6649; sub.ftz.f32 %f6654, %f6653, %f6651; sub.ftz.f32 %f6655, %f6647, %f6645; add.ftz.f32 %f6656, %f6644, %f6648; add.ftz.f32 %f6657, %f6645, %f6647; sub.ftz.f32 %f6658, %f6437, %f6646; sub.ftz.f32 %f6659, %f6658, %f6651; sub.ftz.f32 %f6660, %f6650, %f6643; sub.ftz.f32 %f6661, %f6648, %f6644; add.ftz.f32 %f6662, %f6643, %f6650; sub.ftz.f32 %f6663, %f6658, %f6649; mul.ftz.f32 %f6664, %f6655, %f6613; fma.rn.ftz.f32 %f6665, %f6654, %f6590, %f6664; mul.ftz.f32 %f6666, %f6655, %f6614; fma.rn.ftz.f32 %f6667, %f6654, %f6611, %f6666; mul.ftz.f32 %f6668, %f6655, %f6635; fma.rn.ftz.f32 %f6669, %f6654, %f6612, %f6668; mul.ftz.f32 %f6670, %f6659, %f6613; fma.rn.ftz.f32 %f6671, %f6657, %f6590, %f6670; mul.ftz.f32 %f6672, %f6659, %f6614; fma.rn.ftz.f32 %f6673, %f6657, %f6611, %f6672; mul.ftz.f32 %f6674, %f6659, %f6635; fma.rn.ftz.f32 %f6675, %f6657, %f6612, %f6674; mul.ftz.f32 %f6676, %f6662, %f6613; fma.rn.ftz.f32 %f6677, %f6661, %f6590, %f6676; mul.ftz.f32 %f6678, %f6662, %f6614; fma.rn.ftz.f32 %f6679, %f6661, %f6611, %f6678; mul.ftz.f32 %f6680, %f6662, %f6635; fma.rn.ftz.f32 %f6681, %f6661, %f6612, %f6680; fma.rn.ftz.f32 %f15868, %f6660, %f6638, %f6675; fma.rn.ftz.f32 %f15867, %f6660, %f6637, %f6673; fma.rn.ftz.f32 %f15866, %f6660, %f6636, %f6671; fma.rn.ftz.f32 %f15872, %f6663, %f6638, %f6681; fma.rn.ftz.f32 %f15871, %f6663, %f6637, %f6679; fma.rn.ftz.f32 %f15870, %f6663, %f6636, %f6677; fma.rn.ftz.f32 %f15864, %f6656, %f6638, %f6669; fma.rn.ftz.f32 %f15863, %f6656, %f6637, %f6667; fma.rn.ftz.f32 %f15862, %f6656, %f6636, %f6665; fma.rn.ftz.f32 %f15873, %f1020, %f6572, %f6588; fma.rn.ftz.f32 %f15869, %f1020, %f6571, %f6587; fma.rn.ftz.f32 %f15865, %f1020, %f6570, %f6586; $L__BB108_745: mul.ftz.f32 %f6694, %f15875, %f15863; fma.rn.ftz.f32 %f6695, %f15874, %f15862, %f6694; fma.rn.ftz.f32 %f6696, %f15876, %f15864, %f6695; add.ftz.f32 %f1069, %f15865, %f6696; mul.ftz.f32 %f6697, %f15875, %f15867; fma.rn.ftz.f32 %f6698, %f15874, %f15866, %f6697; fma.rn.ftz.f32 %f6699, %f15876, %f15868, %f6698; add.ftz.f32 %f1070, %f15869, %f6699; mul.ftz.f32 %f6700, %f15875, %f15871; fma.rn.ftz.f32 %f6701, %f15874, %f15870, %f6700; fma.rn.ftz.f32 %f6702, %f15876, %f15872, %f6701; add.ftz.f32 %f15876, %f15873, %f6702; mul.ftz.f32 %f6703, %f15878, %f15863; fma.rn.ftz.f32 %f6704, %f15877, %f15862, %f6703; fma.rn.ftz.f32 %f6705, %f15879, %f15864, %f6704; add.ftz.f32 %f1072, %f15865, %f6705; mul.ftz.f32 %f6706, %f15878, %f15867; fma.rn.ftz.f32 %f6707, %f15877, %f15866, %f6706; fma.rn.ftz.f32 %f6708, %f15879, %f15868, %f6707; add.ftz.f32 %f1073, %f15869, %f6708; mul.ftz.f32 %f6709, %f15878, %f15871; fma.rn.ftz.f32 %f6710, %f15877, %f15870, %f6709; fma.rn.ftz.f32 %f6711, %f15879, %f15872, %f6710; add.ftz.f32 %f15879, %f15873, %f6711; mul.ftz.f32 %f6712, %f15881, %f15863; fma.rn.ftz.f32 %f6713, %f15880, %f15862, %f6712; fma.rn.ftz.f32 %f6714, %f15882, %f15864, %f6713; add.ftz.f32 %f1075, %f15865, %f6714; mul.ftz.f32 %f6715, %f15881, %f15867; fma.rn.ftz.f32 %f6716, %f15880, %f15866, %f6715; fma.rn.ftz.f32 %f6717, %f15882, %f15868, %f6716; add.ftz.f32 %f1076, %f15869, %f6717; mul.ftz.f32 %f6718, %f15881, %f15871; fma.rn.ftz.f32 %f6719, %f15880, %f15870, %f6718; fma.rn.ftz.f32 %f6720, %f15882, %f15872, %f6719; add.ftz.f32 %f15882, %f15873, %f6720; mov.f32 %f15874, %f1069; mov.f32 %f15875, %f1070; mov.f32 %f15877, %f1072; mov.f32 %f15878, %f1073; mov.f32 %f15880, %f1075; mov.f32 %f15881, %f1076; $L__BB108_746: sub.ftz.f32 %f6721, %f15876, %f15879; sub.ftz.f32 %f6722, %f15881, %f15878; mul.ftz.f32 %f6723, %f6721, %f6722; sub.ftz.f32 %f6724, %f15875, %f15878; sub.ftz.f32 %f6725, %f15882, %f15879; mul.ftz.f32 %f6726, %f6724, %f6725; sub.ftz.f32 %f6727, %f6723, %f6726; sub.ftz.f32 %f6728, %f15874, %f15877; mul.ftz.f32 %f6729, %f6728, %f6725; sub.ftz.f32 %f6730, %f15880, %f15877; mul.ftz.f32 %f6731, %f6721, %f6730; sub.ftz.f32 %f6732, %f6729, %f6731; mul.ftz.f32 %f6733, %f6724, %f6730; mul.ftz.f32 %f6734, %f6728, %f6722; sub.ftz.f32 %f6735, %f6733, %f6734; mul.ftz.f32 %f6736, %f6732, %f6732; fma.rn.ftz.f32 %f6737, %f6727, %f6727, %f6736; fma.rn.ftz.f32 %f6738, %f6735, %f6735, %f6737; sqrt.approx.ftz.f32 %f15883, %f6738; $L__BB108_748: ld.const.f32 %f6739, [kernel_params+1500]; mul.ftz.f32 %f6740, %f15883, 0f3F000000; mul.ftz.f32 %f6741, %f6740, %f6739; div.approx.ftz.f32 %f15884, %f6741, %f1000; bra.uni $L__BB108_749; $L__BB108_724: mov.f32 %f6153, 0fBF800000; max.ftz.f32 %f6154, %f877, %f6153; min.ftz.f32 %f6156, %f6154, %f6152; abs.ftz.f32 %f6157, %f6156; sub.ftz.f32 %f6158, %f6152, %f6157; mul.ftz.f32 %f6159, %f6158, 0f3F000000; sqrt.approx.ftz.f32 %f6160, %f6159; setp.gt.ftz.f32 %p617, %f6157, 0f3F11EB85; selp.f32 %f6161, %f6160, %f6157, %p617; mul.ftz.f32 %f6162, %f6161, %f6161; mov.f32 %f6163, 0f3C94D2E9; mov.f32 %f6164, 0f3D53F941; fma.rn.ftz.f32 %f6165, %f6164, %f6162, %f6163; mov.f32 %f6166, 0f3D3F841F; fma.rn.ftz.f32 %f6167, %f6165, %f6162, %f6166; mov.f32 %f6168, 0f3D994929; fma.rn.ftz.f32 %f6169, %f6167, %f6162, %f6168; mov.f32 %f6170, 0f3E2AAB94; fma.rn.ftz.f32 %f6171, %f6169, %f6162, %f6170; mul.ftz.f32 %f6172, %f6162, %f6171; fma.rn.ftz.f32 %f6173, %f6172, %f6161, %f6161; add.ftz.f32 %f6174, %f6173, %f6173; mov.f32 %f6175, 0f3FC90FDB; sub.ftz.f32 %f6176, %f6175, %f6173; selp.f32 %f6177, %f6174, %f6176, %p617; setp.lt.ftz.f32 %p618, %f6156, 0f00000000; mov.f32 %f6178, 0f40490FDB; sub.ftz.f32 %f6179, %f6178, %f6177; selp.f32 %f6180, %f6179, %f6177, %p618; mul.ftz.f32 %f6181, %f6140, %f877; sub.ftz.f32 %f6182, %f6144, %f6181; mul.ftz.f32 %f6183, %f6141, %f877; sub.ftz.f32 %f6184, %f6145, %f6183; mul.ftz.f32 %f6185, %f6142, %f877; sub.ftz.f32 %f6186, %f6146, %f6185; mul.ftz.f32 %f6187, %f6143, %f877; sub.ftz.f32 %f6188, %f6147, %f6187; mul.ftz.f32 %f6189, %f6184, %f6184; fma.rn.ftz.f32 %f6190, %f6182, %f6182, %f6189; mul.ftz.f32 %f6191, %f6188, %f6188; fma.rn.ftz.f32 %f6192, %f6186, %f6186, %f6191; add.ftz.f32 %f6193, %f6190, %f6192; rsqrt.approx.ftz.f32 %f6194, %f6193; mul.ftz.f32 %f6195, %f6182, %f6194; mul.ftz.f32 %f6196, %f6184, %f6194; mul.ftz.f32 %f6197, %f6186, %f6194; mul.ftz.f32 %f6198, %f6188, %f6194; mul.ftz.f32 %f6199, %f868, %f6180; cos.approx.ftz.f32 %f6200, %f6199; sin.approx.ftz.f32 %f6201, %f6199; mul.ftz.f32 %f6202, %f6195, %f6201; mul.ftz.f32 %f6203, %f6196, %f6201; mul.ftz.f32 %f6204, %f6197, %f6201; mul.ftz.f32 %f6205, %f6198, %f6201; fma.rn.ftz.f32 %f15827, %f6143, %f6200, %f6205; fma.rn.ftz.f32 %f15826, %f6142, %f6200, %f6204; fma.rn.ftz.f32 %f15825, %f6141, %f6200, %f6203; fma.rn.ftz.f32 %f15824, %f6140, %f6200, %f6202; $L__BB108_726: ld.global.v4.f32 {%f6220, %f6221, %f6222, %f6223}, [%rd162+16]; ld.global.v4.f32 {%f6228, %f6229, %f6230, %f6231}, [%rd162+80]; mul.ftz.f32 %f6236, %f868, %f6228; mul.ftz.f32 %f6237, %f868, %f6229; mul.ftz.f32 %f6238, %f868, %f6230; mul.ftz.f32 %f6239, %f868, %f6231; fma.rn.ftz.f32 %f6240, %f878, %f6223, %f6239; ld.global.v4.f32 {%f6241, %f6242, %f6243, %f6244}, [%rd162+32]; ld.global.v4.f32 {%f6249, %f6250, %f6251, %f6252}, [%rd162+96]; mul.ftz.f32 %f6257, %f868, %f6249; mul.ftz.f32 %f6258, %f868, %f6250; mul.ftz.f32 %f6259, %f868, %f6251; mul.ftz.f32 %f6260, %f868, %f6252; fma.rn.ftz.f32 %f6261, %f878, %f6241, %f6257; fma.rn.ftz.f32 %f6262, %f878, %f6242, %f6258; fma.rn.ftz.f32 %f6263, %f878, %f6243, %f6259; fma.rn.ftz.f32 %f6264, %f878, %f6244, %f6260; ld.global.v4.f32 {%f6265, %f6266, %f6267, %f6268}, [%rd162+48]; ld.global.v4.f32 {%f6273, %f6274, %f6275, %f6276}, [%rd162+112]; mul.ftz.f32 %f6281, %f868, %f6273; mul.ftz.f32 %f6282, %f868, %f6274; mul.ftz.f32 %f6283, %f868, %f6275; mul.ftz.f32 %f6284, %f868, %f6276; fma.rn.ftz.f32 %f6285, %f878, %f6265, %f6281; fma.rn.ftz.f32 %f6286, %f878, %f6266, %f6282; fma.rn.ftz.f32 %f6287, %f878, %f6267, %f6283; fma.rn.ftz.f32 %f6288, %f878, %f6268, %f6284; mul.ftz.f32 %f6289, %f15827, 0f3FB504F3; mul.ftz.f32 %f6290, %f15824, 0f3FB504F3; mul.ftz.f32 %f6291, %f15825, 0f3FB504F3; mul.ftz.f32 %f6292, %f15826, 0f3FB504F3; mul.ftz.f32 %f6293, %f6289, %f6290; mul.ftz.f32 %f6294, %f6289, %f6291; mul.ftz.f32 %f6295, %f6289, %f6292; mul.ftz.f32 %f6296, %f6290, %f6290; mul.ftz.f32 %f6297, %f6290, %f6291; mul.ftz.f32 %f6298, %f6290, %f6292; mul.ftz.f32 %f6299, %f6291, %f6291; mul.ftz.f32 %f6300, %f6291, %f6292; mul.ftz.f32 %f6301, %f6292, %f6292; sub.ftz.f32 %f6303, %f6152, %f6299; sub.ftz.f32 %f6304, %f6303, %f6301; sub.ftz.f32 %f6305, %f6297, %f6295; add.ftz.f32 %f6306, %f6294, %f6298; add.ftz.f32 %f6307, %f6295, %f6297; sub.ftz.f32 %f6308, %f6152, %f6296; sub.ftz.f32 %f6309, %f6308, %f6301; sub.ftz.f32 %f6310, %f6300, %f6293; sub.ftz.f32 %f6311, %f6298, %f6294; add.ftz.f32 %f6312, %f6293, %f6300; sub.ftz.f32 %f6313, %f6308, %f6299; mul.ftz.f32 %f6314, %f6305, %f6263; fma.rn.ftz.f32 %f6315, %f6304, %f6240, %f6314; mul.ftz.f32 %f6316, %f6305, %f6264; fma.rn.ftz.f32 %f6317, %f6304, %f6261, %f6316; mul.ftz.f32 %f6318, %f6305, %f6285; fma.rn.ftz.f32 %f6319, %f6304, %f6262, %f6318; mul.ftz.f32 %f6320, %f6309, %f6263; fma.rn.ftz.f32 %f6321, %f6307, %f6240, %f6320; mul.ftz.f32 %f6322, %f6309, %f6264; fma.rn.ftz.f32 %f6323, %f6307, %f6261, %f6322; mul.ftz.f32 %f6324, %f6309, %f6285; fma.rn.ftz.f32 %f6325, %f6307, %f6262, %f6324; mul.ftz.f32 %f6326, %f6312, %f6263; fma.rn.ftz.f32 %f6327, %f6311, %f6240, %f6326; mul.ftz.f32 %f6328, %f6312, %f6264; fma.rn.ftz.f32 %f6329, %f6311, %f6261, %f6328; mul.ftz.f32 %f6330, %f6312, %f6285; fma.rn.ftz.f32 %f6331, %f6311, %f6262, %f6330; fma.rn.ftz.f32 %f15834, %f6310, %f6288, %f6325; fma.rn.ftz.f32 %f15833, %f6310, %f6287, %f6323; fma.rn.ftz.f32 %f15832, %f6310, %f6286, %f6321; fma.rn.ftz.f32 %f15838, %f6313, %f6288, %f6331; fma.rn.ftz.f32 %f15837, %f6313, %f6287, %f6329; fma.rn.ftz.f32 %f15836, %f6313, %f6286, %f6327; fma.rn.ftz.f32 %f15830, %f6306, %f6288, %f6319; fma.rn.ftz.f32 %f15829, %f6306, %f6287, %f6317; fma.rn.ftz.f32 %f15828, %f6306, %f6286, %f6315; fma.rn.ftz.f32 %f15839, %f878, %f6222, %f6238; fma.rn.ftz.f32 %f15835, %f878, %f6221, %f6237; fma.rn.ftz.f32 %f15831, %f878, %f6220, %f6236; $L__BB108_728: mul.ftz.f32 %f6344, %f15841, %f15829; fma.rn.ftz.f32 %f6345, %f15840, %f15828, %f6344; fma.rn.ftz.f32 %f6346, %f15842, %f15830, %f6345; add.ftz.f32 %f927, %f15831, %f6346; mul.ftz.f32 %f6347, %f15841, %f15833; fma.rn.ftz.f32 %f6348, %f15840, %f15832, %f6347; fma.rn.ftz.f32 %f6349, %f15842, %f15834, %f6348; add.ftz.f32 %f928, %f15835, %f6349; mul.ftz.f32 %f6350, %f15841, %f15837; fma.rn.ftz.f32 %f6351, %f15840, %f15836, %f6350; fma.rn.ftz.f32 %f6352, %f15842, %f15838, %f6351; add.ftz.f32 %f15842, %f15839, %f6352; mul.ftz.f32 %f6353, %f15844, %f15829; fma.rn.ftz.f32 %f6354, %f15843, %f15828, %f6353; fma.rn.ftz.f32 %f6355, %f15845, %f15830, %f6354; add.ftz.f32 %f930, %f15831, %f6355; mul.ftz.f32 %f6356, %f15844, %f15833; fma.rn.ftz.f32 %f6357, %f15843, %f15832, %f6356; fma.rn.ftz.f32 %f6358, %f15845, %f15834, %f6357; add.ftz.f32 %f931, %f15835, %f6358; mul.ftz.f32 %f6359, %f15844, %f15837; fma.rn.ftz.f32 %f6360, %f15843, %f15836, %f6359; fma.rn.ftz.f32 %f6361, %f15845, %f15838, %f6360; add.ftz.f32 %f15845, %f15839, %f6361; mul.ftz.f32 %f6362, %f15847, %f15829; fma.rn.ftz.f32 %f6363, %f15846, %f15828, %f6362; fma.rn.ftz.f32 %f6364, %f15848, %f15830, %f6363; add.ftz.f32 %f933, %f15831, %f6364; mul.ftz.f32 %f6365, %f15847, %f15833; fma.rn.ftz.f32 %f6366, %f15846, %f15832, %f6365; fma.rn.ftz.f32 %f6367, %f15848, %f15834, %f6366; add.ftz.f32 %f934, %f15835, %f6367; mul.ftz.f32 %f6368, %f15847, %f15837; fma.rn.ftz.f32 %f6369, %f15846, %f15836, %f6368; fma.rn.ftz.f32 %f6370, %f15848, %f15838, %f6369; add.ftz.f32 %f15848, %f15839, %f6370; mov.f32 %f15840, %f927; mov.f32 %f15841, %f928; mov.f32 %f15843, %f930; mov.f32 %f15844, %f931; mov.f32 %f15846, %f933; mov.f32 %f15847, %f934; $L__BB108_729: sub.ftz.f32 %f6371, %f15842, %f15845; sub.ftz.f32 %f6372, %f15847, %f15844; mul.ftz.f32 %f6373, %f6371, %f6372; sub.ftz.f32 %f6374, %f15841, %f15844; sub.ftz.f32 %f6375, %f15848, %f15845; mul.ftz.f32 %f6376, %f6374, %f6375; sub.ftz.f32 %f6377, %f6373, %f6376; sub.ftz.f32 %f6378, %f15840, %f15843; mul.ftz.f32 %f6379, %f6378, %f6375; sub.ftz.f32 %f6380, %f15846, %f15843; mul.ftz.f32 %f6381, %f6371, %f6380; sub.ftz.f32 %f6382, %f6379, %f6381; mul.ftz.f32 %f6383, %f6374, %f6380; mul.ftz.f32 %f6384, %f6378, %f6372; sub.ftz.f32 %f6385, %f6383, %f6384; mul.ftz.f32 %f6386, %f6382, %f6382; fma.rn.ftz.f32 %f6387, %f6377, %f6377, %f6386; fma.rn.ftz.f32 %f6388, %f6385, %f6385, %f6387; sqrt.approx.ftz.f32 %f6389, %f6388; mul.ftz.f32 %f6390, %f6389, 0f3F000000; mul.ftz.f32 %f6391, %f15823, %f6390; div.approx.ftz.f32 %f15884, %f6391, %f858; $L__BB108_749: mul.ftz.f32 %f6742, %f731, %f731; fma.rn.ftz.f32 %f6743, %f15884, %f15884, %f6742; div.approx.ftz.f32 %f15885, %f6742, %f6743; $L__BB108_750: ld.local.u32 %r381, [%rd7+-12]; setp.eq.s32 %p635, %r381, -1; mov.u32 %r3167, -1; @%p635 bra $L__BB108_752; ld.const.u64 %rd1220, [kernel_params+2160]; cvta.to.global.u64 %rd1221, %rd1220; mul.wide.s32 %rd1222, %r381, 208; add.s64 %rd1223, %rd1221, %rd1222; ld.global.u32 %r3167, [%rd1223+192]; $L__BB108_752: abs.ftz.f32 %f6745, %f727; setp.gt.ftz.f32 %p636, %f6745, 0f00000000; mov.f32 %f15894, 0f00000000; selp.f32 %f6746, 0f3F800000, 0f00000000, %p636; ld.const.u64 %rd1224, [kernel_params+2584]; cvta.to.global.u64 %rd1225, %rd1224; mul.lo.s64 %rd1226, %rd2, 12; add.s64 %rd1227, %rd1225, %rd1226; ld.global.f32 %f6747, [%rd1227]; mul.ftz.f32 %f6748, %f5827, %f6746; mul.ftz.f32 %f6749, %f6748, %f6747; ld.global.f32 %f6750, [%rd1227+4]; mul.ftz.f32 %f6751, %f6746, %f5828; mul.ftz.f32 %f6752, %f6751, %f6750; ld.global.f32 %f6753, [%rd1227+8]; mul.ftz.f32 %f6754, %f6746, %f730; mul.ftz.f32 %f6755, %f6754, %f6753; mul.ftz.f32 %f1094, %f15885, %f6749; mul.ftz.f32 %f1095, %f15885, %f6752; mul.ftz.f32 %f1096, %f15885, %f6755; ld.global.u16 %rs15, [%rd74]; mov.b32 %r384, %f1094; setp.nan.ftz.f32 %p637, %f1094, %f1094; mov.f32 %f15892, %f15894; @%p637 bra $L__BB108_755; or.b32 %r1422, %r384, -2147483648; setp.ne.s32 %p638, %r1422, -2147483648; add.ftz.f32 %f6757, %f1094, %f1094; setp.eq.ftz.f32 %p639, %f6757, %f1094; and.pred %p640, %p639, %p638; mov.f32 %f15892, %f15894; @%p640 bra $L__BB108_755; shl.b32 %r1423, %r384, 1; setp.lt.u32 %p641, %r1423, -16777215; selp.f32 %f15892, %f1094, 0f00000000, %p641; $L__BB108_755: mov.b32 %r385, %f1095; setp.nan.ftz.f32 %p642, %f1095, %f1095; mov.f32 %f15893, %f15894; @%p642 bra $L__BB108_758; or.b32 %r1424, %r385, -2147483648; setp.ne.s32 %p643, %r1424, -2147483648; add.ftz.f32 %f6760, %f1095, %f1095; setp.eq.ftz.f32 %p644, %f6760, %f1095; and.pred %p645, %p644, %p643; mov.f32 %f15893, %f15894; @%p645 bra $L__BB108_758; shl.b32 %r1425, %r385, 1; setp.lt.u32 %p646, %r1425, -16777215; selp.f32 %f15893, %f1095, 0f00000000, %p646; $L__BB108_758: mov.b32 %r386, %f1096; setp.nan.ftz.f32 %p647, %f1096, %f1096; @%p647 bra $L__BB108_761; or.b32 %r1426, %r386, -2147483648; setp.ne.s32 %p648, %r1426, -2147483648; add.ftz.f32 %f6763, %f1096, %f1096; setp.eq.ftz.f32 %p649, %f6763, %f1096; and.pred %p650, %p649, %p648; @%p650 bra $L__BB108_761; shl.b32 %r1427, %r386, 1; setp.lt.u32 %p651, %r1427, -16777215; selp.f32 %f15894, %f1096, 0f00000000, %p651; $L__BB108_761: ld.const.f32 %f6764, [kernel_params+1580]; ld.const.f32 %f6765, [kernel_params+1584]; setp.gt.u16 %p652, %rs15, 1; selp.f32 %f1103, %f6765, %f6764, %p652; abs.ftz.f32 %f6766, %f15893; abs.ftz.f32 %f6767, %f15892; add.ftz.f32 %f6768, %f6767, %f6766; abs.ftz.f32 %f6769, %f15894; add.ftz.f32 %f1104, %f6769, %f6768; setp.leu.ftz.f32 %p653, %f1104, %f1103; @%p653 bra $L__BB108_763; div.approx.ftz.f32 %f6770, %f1103, %f1104; mul.ftz.f32 %f15892, %f15892, %f6770; mul.ftz.f32 %f15893, %f15893, %f6770; mul.ftz.f32 %f15894, %f15894, %f6770; $L__BB108_763: ld.const.v2.u64 {%rd1228, %rd1229}, [kernel_params+2448]; cvta.to.global.u64 %rd1232, %rd1228; add.s64 %rd1234, %rd1232, %rd837; ld.global.u32 %rd1235, [%rd1234]; ld.const.s32 %rd1236, [kernel_params+1264]; mul.lo.s64 %rd165, %rd1236, %rd1235; cvta.to.global.u64 %rd1237, %rd1229; add.s64 %rd1238, %rd1237, %rd837; ld.global.u32 %r387, [%rd1238]; ld.const.u32 %r1428, [kernel_params+1620]; setp.eq.s32 %p654, %r1428, 0; @%p654 bra $L__BB108_768; and.b32 %r1429, %r360, 536870912; setp.ne.s32 %p655, %r1429, 0; @%p655 bra $L__BB108_766; ld.const.s32 %rd1239, [kernel_params+1384]; add.s64 %rd1240, %rd165, %rd1239; cvta.to.global.u64 %rd1241, %rd618; shl.b64 %rd1242, %rd1240, 2; add.s64 %rd1243, %rd1241, %rd1242; add.s64 %rd1244, %rd1243, 4; add.s64 %rd1245, %rd1243, 8; atom.global.add.f32 %f6771, [%rd1243], %f15892; atom.global.add.f32 %f6772, [%rd1244], %f15893; atom.global.add.f32 %f6773, [%rd1245], %f15894; $L__BB108_766: and.b32 %r1430, %r360, 1073741824; setp.eq.s32 %p656, %r1430, 0; @%p656 bra $L__BB108_768; bra.uni $L__BB108_767; $L__BB108_768: ld.const.u32 %r3168, [kernel_params+1260]; and.b32 %r1431, %r3168, 2; setp.eq.s32 %p657, %r1431, 0; @%p657 bra $L__BB108_770; ld.const.s32 %rd1253, [kernel_params+1268]; add.s64 %rd1254, %rd165, %rd1253; cvta.to.global.u64 %rd1255, %rd618; shl.b64 %rd1256, %rd1254, 2; add.s64 %rd1257, %rd1255, %rd1256; add.s64 %rd1258, %rd1257, 4; add.s64 %rd1259, %rd1257, 8; atom.global.add.f32 %f6777, [%rd1257], %f15892; atom.global.add.f32 %f6778, [%rd1258], %f15893; atom.global.add.f32 %f6779, [%rd1259], %f15894; $L__BB108_770: ld.const.u32 %r390, [kernel_params+1400]; setp.eq.s32 %p658, %r390, -1; @%p658 bra $L__BB108_773; and.b32 %r1432, %r387, -1431655766; popc.b32 %r1433, %r1432; and.b32 %r1434, %r1433, 1; setp.eq.b32 %p659, %r1434, 1; mov.pred %p660, 0; xor.pred %p661, %p659, %p660; not.pred %p662, %p661; @%p662 bra $L__BB108_773; cvt.s64.s32 %rd1260, %r390; add.s64 %rd1261, %rd165, %rd1260; cvta.to.global.u64 %rd1262, %rd618; shl.b64 %rd1263, %rd1261, 2; add.s64 %rd1264, %rd1262, %rd1263; add.s64 %rd1265, %rd1264, 4; add.s64 %rd1266, %rd1264, 8; add.s64 %rd1267, %rd1264, 12; add.ftz.f32 %f6780, %f15892, %f15892; atom.global.add.f32 %f6781, [%rd1264], %f6780; add.ftz.f32 %f6782, %f15893, %f15893; atom.global.add.f32 %f6783, [%rd1265], %f6782; add.ftz.f32 %f6784, %f15894, %f15894; atom.global.add.f32 %f6785, [%rd1266], %f6784; atom.global.add.f32 %f6786, [%rd1267], 0f00000000; bra.uni $L__BB108_773; $L__BB108_767: ld.const.s32 %rd1246, [kernel_params+1376]; add.s64 %rd1247, %rd165, %rd1246; cvta.to.global.u64 %rd1248, %rd618; shl.b64 %rd1249, %rd1247, 2; add.s64 %rd1250, %rd1248, %rd1249; add.s64 %rd1251, %rd1250, 4; add.s64 %rd1252, %rd1250, 8; atom.global.add.f32 %f6774, [%rd1250], %f15892; atom.global.add.f32 %f6775, [%rd1251], %f15893; atom.global.add.f32 %f6776, [%rd1252], %f15894; ld.const.u32 %r3168, [kernel_params+1260]; $L__BB108_773: setp.eq.s32 %p663, %r3168, 0; @%p663 bra $L__BB108_794; ld.const.u32 %r3169, [kernel_params+1352]; and.b32 %r1435, %r360, 16777216; setp.eq.s32 %p664, %r1435, 0; ld.const.u32 %r393, [kernel_params+1428]; setp.eq.s32 %p665, %r393, -1; or.pred %p666, %p664, %p665; @%p666 bra $L__BB108_776; ld.const.u64 %rd1268, [kernel_params+2616]; cvta.to.global.u64 %rd1269, %rd1268; add.s64 %rd1271, %rd1269, %rd1226; ld.global.f32 %f6787, [%rd1271]; mul.ftz.f32 %f6788, %f15892, %f6787; ld.global.f32 %f6789, [%rd1271+4]; mul.ftz.f32 %f6790, %f15893, %f6789; ld.global.f32 %f6791, [%rd1271+8]; mul.ftz.f32 %f6792, %f15894, %f6791; cvta.to.global.u64 %rd1272, %rd618; cvt.s64.s32 %rd1273, %r393; add.s64 %rd1274, %rd165, %rd1273; shl.b64 %rd1275, %rd1274, 2; add.s64 %rd1276, %rd1272, %rd1275; add.s64 %rd1277, %rd1276, 4; add.s64 %rd1278, %rd1276, 8; atom.global.add.f32 %f6793, [%rd1276], %f6788; atom.global.add.f32 %f6794, [%rd1277], %f6790; atom.global.add.f32 %f6795, [%rd1278], %f6792; $L__BB108_776: setp.eq.s32 %p667, %r3167, -1; ld.const.u32 %r394, [kernel_params+1444]; setp.eq.s32 %p668, %r394, -1; or.pred %p669, %p667, %p668; @%p669 bra $L__BB108_778; mul.lo.s32 %r1436, %r3167, 3; cvt.s64.s32 %rd1279, %r1436; add.s64 %rd1280, %rd165, %rd1279; cvt.s64.s32 %rd1281, %r394; add.s64 %rd1282, %rd1280, %rd1281; cvta.to.global.u64 %rd1283, %rd618; shl.b64 %rd1284, %rd1282, 2; add.s64 %rd1285, %rd1283, %rd1284; add.s64 %rd1286, %rd1285, 4; add.s64 %rd1287, %rd1285, 8; atom.global.add.f32 %f6796, [%rd1285], %f15892; atom.global.add.f32 %f6797, [%rd1286], %f15893; atom.global.add.f32 %f6798, [%rd1287], %f15894; $L__BB108_778: and.b32 %r1437, %r360, 100663296; setp.eq.s32 %p670, %r1437, 0; @%p670 bra $L__BB108_792; ld.const.u8 %rs74, [kernel_params+2]; and.b16 %rs75, %rs74, 32; setp.eq.s16 %p671, %rs75, 0; mov.u32 %r3169, -1; @%p671 bra $L__BB108_792; bra.uni $L__BB108_780; $L__BB108_792: setp.eq.s32 %p682, %r3169, -1; @%p682 bra $L__BB108_794; cvt.s64.s32 %rd1309, %r3169; add.s64 %rd1310, %rd165, %rd1309; cvta.to.global.u64 %rd1311, %rd618; shl.b64 %rd1312, %rd1310, 2; add.s64 %rd1313, %rd1311, %rd1312; add.s64 %rd1314, %rd1313, 4; add.s64 %rd1315, %rd1313, 8; atom.global.add.f32 %f6818, [%rd1313], %f15892; atom.global.add.f32 %f6819, [%rd1314], %f15893; atom.global.add.f32 %f6820, [%rd1315], %f15894; bra.uni $L__BB108_794; $L__BB108_780: and.b32 %r1439, %r360, 536870912; setp.ne.s32 %p672, %r1439, 0; @%p672 bra $L__BB108_794; and.b32 %r1440, %r360, 33554432; setp.eq.s32 %p673, %r1440, 0; @%p673 bra $L__BB108_788; bra.uni $L__BB108_782; $L__BB108_788: and.b32 %r1449, %r360, 67108864; setp.eq.s32 %p680, %r1449, 0; @%p680 bra $L__BB108_792; setp.eq.s16 %p681, %rs15, 1; @%p681 bra $L__BB108_791; bra.uni $L__BB108_790; $L__BB108_791: ld.const.u32 %r3169, [kernel_params+1348]; bra.uni $L__BB108_792; $L__BB108_794: cvt.u16.u32 %rs76, %r45; setp.lt.s16 %p683, %rs76, 0; mov.f32 %f15895, 0f00000000; @%p683 bra $L__BB108_796; ld.const.u64 %rd1316, [kernel_params+2576]; cvta.to.global.u64 %rd1317, %rd1316; add.s64 %rd1319, %rd1317, %rd837; ld.global.f32 %f15895, [%rd1319]; $L__BB108_796: setp.eq.ftz.f32 %p684, %f15895, 0f00000000; mov.u16 %rs161, 0; @%p684 bra $L__BB108_2421; setp.eq.ftz.f32 %p685, %f15895, 0f3F800000; @%p685 bra $L__BB108_799; ld.const.u64 %rd1320, [kernel_params+2584]; cvta.to.global.u64 %rd1321, %rd1320; mul.lo.s64 %rd1322, %rd2, 12; add.s64 %rd1323, %rd1321, %rd1322; ld.global.f32 %f6822, [%rd1323]; rcp.approx.ftz.f32 %f6823, %f15895; mul.ftz.f32 %f6824, %f6823, %f6822; ld.global.f32 %f6825, [%rd1323+4]; mul.ftz.f32 %f6826, %f6823, %f6825; ld.global.f32 %f6827, [%rd1323+8]; mul.ftz.f32 %f6828, %f6823, %f6827; st.global.f32 [%rd1323], %f6824; st.global.f32 [%rd1323+4], %f6826; st.global.f32 [%rd1323+8], %f6828; $L__BB108_799: ld.global.u32 %r401, [%rd29]; and.b32 %r1450, %r401, 16384; setp.eq.s32 %p686, %r1450, 0; @%p686 bra $L__BB108_1257; ld.const.u32 %r402, [kernel_params+1256]; setp.eq.s32 %p687, %r402, 0; @%p687 bra $L__BB108_1257; ld.const.u64 %rd1324, [kernel_params+2448]; cvta.to.global.u64 %rd1325, %rd1324; add.s64 %rd1327, %rd1325, %rd837; ld.global.u32 %rd1328, [%rd1327]; ld.const.s32 %rd1329, [kernel_params+1264]; mul.lo.s64 %rd166, %rd1329, %rd1328; and.b32 %r1451, %r401, 8192; setp.ne.s32 %p688, %r1451, 0; @%p688 bra $L__BB108_1146; ld.const.u64 %rd1330, [kernel_params+2456]; cvta.to.global.u64 %rd1331, %rd1330; add.s64 %rd1333, %rd1331, %rd837; ld.global.u32 %r1452, [%rd1333]; setp.ne.s32 %p689, %r1452, 0; @%p689 bra $L__BB108_816; and.b32 %r1453, %r402, 1; setp.eq.b32 %p690, %r1453, 1; mov.pred %p691, 0; xor.pred %p692, %p690, %p691; not.pred %p693, %p692; @%p693 bra $L__BB108_808; ld.local.v4.f32 {%f6829, %f6830, %f6831, %f6832}, [%rd1]; ld.const.u32 %r1454, [kernel_params+16]; setp.eq.s32 %p694, %r1454, 2; @%p694 bra $L__BB108_806; bra.uni $L__BB108_805; $L__BB108_806: ld.const.v4.f32 {%f6844, %f6845, %f6846, %f6847}, [kernel_params+96]; ld.const.v4.f32 {%f6849, %f6850, %f6851, %f6852}, [kernel_params+112]; ld.const.v4.f32 {%f6854, %f6855, %f6856, %f6857}, [kernel_params+128]; sub.ftz.f32 %f6859, %f6829, %f6847; sub.ftz.f32 %f6860, %f6830, %f6852; sub.ftz.f32 %f6861, %f6831, %f6857; mul.ftz.f32 %f6862, %f6860, %f6860; fma.rn.ftz.f32 %f6863, %f6859, %f6859, %f6862; fma.rn.ftz.f32 %f6864, %f6861, %f6861, %f6863; sqrt.approx.ftz.f32 %f15896, %f6864; bra.uni $L__BB108_807; $L__BB108_782: ld.const.u64 %rd1288, [kernel_params+2600]; cvta.to.global.u64 %rd1289, %rd1288; add.s64 %rd1291, %rd1289, %rd1226; ld.global.f32 %f1111, [%rd1291]; ld.global.f32 %f1112, [%rd1291+4]; ld.global.f32 %f1113, [%rd1291+8]; ld.const.u64 %rd1292, [kernel_params+2608]; cvta.to.global.u64 %rd1293, %rd1292; add.s64 %rd1294, %rd1293, %rd1226; ld.global.f32 %f1114, [%rd1294]; ld.global.f32 %f1115, [%rd1294+4]; ld.global.f32 %f1116, [%rd1294+8]; ld.const.u32 %r1441, [kernel_params+1324]; ld.const.u32 %r1442, [kernel_params+1340]; setp.eq.s16 %p674, %rs15, 1; selp.b32 %r395, %r1442, %r1441, %p674; setp.eq.s32 %p675, %r395, -1; @%p675 bra $L__BB108_784; cvt.s64.s32 %rd1295, %r395; add.s64 %rd1296, %rd165, %rd1295; cvta.to.global.u64 %rd1297, %rd618; shl.b64 %rd1298, %rd1296, 2; add.s64 %rd1299, %rd1297, %rd1298; add.s64 %rd1300, %rd1299, 4; add.s64 %rd1301, %rd1299, 8; mul.ftz.f32 %f6799, %f15892, %f1114; atom.global.add.f32 %f6800, [%rd1299], %f6799; mul.ftz.f32 %f6801, %f15893, %f1115; atom.global.add.f32 %f6802, [%rd1300], %f6801; mul.ftz.f32 %f6803, %f15894, %f1116; atom.global.add.f32 %f6804, [%rd1301], %f6803; $L__BB108_784: ld.const.u32 %r1443, [kernel_params+1328]; ld.const.u32 %r1444, [kernel_params+1344]; selp.b32 %r396, %r1444, %r1443, %p674; setp.eq.s32 %p677, %r396, -1; @%p677 bra $L__BB108_786; mov.f32 %f6805, 0f3F800000; sub.ftz.f32 %f6806, %f6805, %f1111; sub.ftz.f32 %f6807, %f6806, %f1114; sub.ftz.f32 %f6808, %f6805, %f1112; sub.ftz.f32 %f6809, %f6808, %f1115; sub.ftz.f32 %f6810, %f6805, %f1113; sub.ftz.f32 %f6811, %f6810, %f1116; cvt.s64.s32 %rd1302, %r396; add.s64 %rd1303, %rd165, %rd1302; cvta.to.global.u64 %rd1304, %rd618; shl.b64 %rd1305, %rd1303, 2; add.s64 %rd1306, %rd1304, %rd1305; mul.ftz.f32 %f6812, %f15892, %f6807; mul.ftz.f32 %f6813, %f15893, %f6809; mul.ftz.f32 %f6814, %f15894, %f6811; add.s64 %rd1307, %rd1306, 4; add.s64 %rd1308, %rd1306, 8; atom.global.add.f32 %f6815, [%rd1306], %f6812; atom.global.add.f32 %f6816, [%rd1307], %f6813; atom.global.add.f32 %f6817, [%rd1308], %f6814; $L__BB108_786: ld.const.u32 %r1446, [kernel_params+1320]; ld.const.u32 %r1447, [kernel_params+1336]; selp.b32 %r397, %r1447, %r1446, %p674; setp.eq.s32 %p679, %r397, -1; @%p679 bra $L__BB108_792; mul.ftz.f32 %f15892, %f15892, %f1111; mul.ftz.f32 %f15893, %f15893, %f1112; mul.ftz.f32 %f15894, %f15894, %f1113; mov.u32 %r3169, %r397; bra.uni $L__BB108_792; $L__BB108_790: ld.const.u32 %r3169, [kernel_params+1332]; bra.uni $L__BB108_792; $L__BB108_805: ld.const.v4.f32 {%f6833, %f6834, %f6835, %f6836}, [kernel_params+720]; mul.ftz.f32 %f6841, %f6830, %f6834; fma.rn.ftz.f32 %f6842, %f6829, %f6833, %f6841; fma.rn.ftz.f32 %f6843, %f6831, %f6835, %f6842; add.ftz.f32 %f15896, %f6836, %f6843; $L__BB108_807: ld.const.s32 %rd1336, [kernel_params+1272]; add.s64 %rd1337, %rd166, %rd1336; cvta.to.global.u64 %rd1338, %rd618; shl.b64 %rd1339, %rd1337, 2; add.s64 %rd1340, %rd1338, %rd1339; st.global.f32 [%rd1340], %f15896; $L__BB108_808: and.b32 %r1455, %r402, 32; setp.eq.s32 %p695, %r1455, 0; @%p695 bra $L__BB108_812; ld.local.u32 %r403, [%rd7+-12]; setp.eq.s32 %p696, %r403, -1; mov.f32 %f15897, 0f00000000; @%p696 bra $L__BB108_811; ld.const.u64 %rd1341, [kernel_params+2160]; cvta.to.global.u64 %rd1342, %rd1341; mul.wide.s32 %rd1343, %r403, 208; add.s64 %rd1344, %rd1342, %rd1343; ld.global.f32 %f15897, [%rd1344+100]; $L__BB108_811: ld.const.s32 %rd1345, [kernel_params+1300]; add.s64 %rd1346, %rd166, %rd1345; cvta.to.global.u64 %rd1347, %rd618; shl.b64 %rd1348, %rd1346, 2; add.s64 %rd1349, %rd1347, %rd1348; st.global.f32 [%rd1349], %f15897; $L__BB108_812: and.b32 %r1456, %r402, 64; setp.eq.s32 %p697, %r1456, 0; @%p697 bra $L__BB108_814; ld.local.u32 %rd1350, [%rd7+-40]; shl.b64 %rd1351, %rd1350, 5; and.b64 %rd1352, %rd1351, 134217696; add.s64 %rd1353, %rd28, %rd1352; ld.global.u32 %r1457, [%rd1353+20]; cvt.rn.f32.s32 %f6866, %r1457; ld.const.s32 %rd1354, [kernel_params+1304]; add.s64 %rd1355, %rd166, %rd1354; cvta.to.global.u64 %rd1356, %rd618; shl.b64 %rd1357, %rd1355, 2; add.s64 %rd1358, %rd1356, %rd1357; st.global.f32 [%rd1358], %f6866; $L__BB108_814: and.b32 %r1458, %r402, 2; setp.eq.s32 %p698, %r1458, 0; @%p698 bra $L__BB108_816; ld.local.v4.f32 {%f6867, %f6868, %f6869, %f6870}, [%rd1]; ld.const.s32 %rd1361, [kernel_params+1276]; add.s64 %rd1362, %rd166, %rd1361; cvta.to.global.u64 %rd1363, %rd618; shl.b64 %rd1364, %rd1362, 2; add.s64 %rd1365, %rd1363, %rd1364; st.global.f32 [%rd1365], %f6867; st.global.f32 [%rd1365+4], %f6868; st.global.f32 [%rd1365+8], %f6869; $L__BB108_816: and.b32 %r1459, %r3134, 512; setp.eq.s32 %p699, %r1459, 0; ld.const.f32 %f1133, [kernel_params+1364]; setp.eq.ftz.f32 %p700, %f1133, 0f00000000; or.pred %p701, %p699, %p700; @%p701 bra $L__BB108_820; and.b32 %r1460, %r3134, 524288; setp.ne.s32 %p702, %r1460, 0; mov.f32 %f6876, 0f3F800000; mov.f32 %f15898, %f6876; mov.f32 %f15899, %f6876; mov.f32 %f15900, %f6876; @%p702 bra $L__BB108_819; ld.local.v2.f32 {%f15899, %f15900}, [%rd7+208]; ld.local.f32 %f15898, [%rd7+204]; $L__BB108_819: sub.ftz.f32 %f6880, %f6876, %f15898; sub.ftz.f32 %f6881, %f6876, %f15899; sub.ftz.f32 %f6882, %f6876, %f15900; cvt.ftz.sat.f32.f32 %f6883, %f6880; cvt.ftz.sat.f32.f32 %f6884, %f6881; cvt.ftz.sat.f32.f32 %f6885, %f6882; add.ftz.f32 %f6886, %f6883, %f6884; add.ftz.f32 %f6887, %f6885, %f6886; mul.ftz.f32 %f6888, %f6887, 0f3EAAAAAB; setp.ltu.ftz.f32 %p703, %f6888, %f1133; @%p703 bra $L__BB108_1146; $L__BB108_820: and.b32 %r1461, %r402, 4; setp.eq.s32 %p704, %r1461, 0; @%p704 bra $L__BB108_841; ld.local.u32 %r404, [%rd7+172]; setp.lt.s32 %p705, %r404, 1; mov.f32 %f15904, 0f00000000; mov.f32 %f15905, %f15904; mov.f32 %f15906, %f15904; @%p705 bra $L__BB108_837; add.s32 %r1463, %r404, -1; and.b32 %r3174, %r404, 3; setp.lt.u32 %p706, %r1463, 3; mov.f32 %f15906, 0f00000000; mov.u32 %r3172, 0; mov.f32 %f15905, %f15906; mov.f32 %f15904, %f15906; @%p706 bra $L__BB108_833; sub.s32 %r3171, %r404, %r3174; mov.f32 %f15906, 0f00000000; mov.u32 %r3172, 0; mov.f32 %f15905, %f15906; mov.f32 %f15904, %f15906; $L__BB108_824: cvt.s64.s32 %rd167, %r3172; mul.wide.s32 %rd1366, %r3172, 80; add.s64 %rd1367, %rd1, %rd1366; add.s64 %rd168, %rd1367, 316; ld.local.u32 %r1465, [%rd1367+316]; setp.gt.s32 %p707, %r1465, 34; @%p707 bra $L__BB108_826; ld.local.v4.f32 {%f6899, %f6900, %f6901, %f6902}, [%rd168+-12]; add.ftz.f32 %f6906, %f6899, %f6900; add.ftz.f32 %f6907, %f6901, %f6906; mul.ftz.f32 %f6908, %f6907, 0f3EAAAAAB; abs.ftz.f32 %f6909, %f6908; ld.local.f32 %f6910, [%rd168+8]; ld.local.f32 %f6911, [%rd168+12]; ld.local.f32 %f6912, [%rd168+16]; fma.rn.ftz.f32 %f15904, %f6909, %f6910, %f15904; fma.rn.ftz.f32 %f15905, %f6909, %f6911, %f15905; fma.rn.ftz.f32 %f15906, %f6909, %f6912, %f15906; $L__BB108_826: ld.local.u32 %r1466, [%rd168+80]; setp.gt.s32 %p708, %r1466, 34; @%p708 bra $L__BB108_828; ld.local.v4.f32 {%f6913, %f6914, %f6915, %f6916}, [%rd168+68]; add.ftz.f32 %f6920, %f6913, %f6914; add.ftz.f32 %f6921, %f6915, %f6920; mul.ftz.f32 %f6922, %f6921, 0f3EAAAAAB; abs.ftz.f32 %f6923, %f6922; ld.local.f32 %f6924, [%rd168+88]; ld.local.f32 %f6925, [%rd168+92]; ld.local.f32 %f6926, [%rd168+96]; fma.rn.ftz.f32 %f15904, %f6923, %f6924, %f15904; fma.rn.ftz.f32 %f15905, %f6923, %f6925, %f15905; fma.rn.ftz.f32 %f15906, %f6923, %f6926, %f15906; $L__BB108_828: ld.local.u32 %r1467, [%rd168+160]; setp.gt.s32 %p709, %r1467, 34; @%p709 bra $L__BB108_830; ld.local.v4.f32 {%f6927, %f6928, %f6929, %f6930}, [%rd168+148]; add.ftz.f32 %f6934, %f6927, %f6928; add.ftz.f32 %f6935, %f6929, %f6934; mul.ftz.f32 %f6936, %f6935, 0f3EAAAAAB; abs.ftz.f32 %f6937, %f6936; ld.local.f32 %f6938, [%rd168+168]; ld.local.f32 %f6939, [%rd168+172]; ld.local.f32 %f6940, [%rd168+176]; fma.rn.ftz.f32 %f15904, %f6937, %f6938, %f15904; fma.rn.ftz.f32 %f15905, %f6937, %f6939, %f15905; fma.rn.ftz.f32 %f15906, %f6937, %f6940, %f15906; $L__BB108_830: ld.local.u32 %r1468, [%rd168+240]; setp.gt.s32 %p710, %r1468, 34; @%p710 bra $L__BB108_832; ld.local.v4.f32 {%f6941, %f6942, %f6943, %f6944}, [%rd168+228]; add.ftz.f32 %f6948, %f6941, %f6942; add.ftz.f32 %f6949, %f6943, %f6948; mul.ftz.f32 %f6950, %f6949, 0f3EAAAAAB; abs.ftz.f32 %f6951, %f6950; ld.local.f32 %f6952, [%rd168+248]; ld.local.f32 %f6953, [%rd168+252]; ld.local.f32 %f6954, [%rd168+256]; fma.rn.ftz.f32 %f15904, %f6951, %f6952, %f15904; fma.rn.ftz.f32 %f15905, %f6951, %f6953, %f15905; fma.rn.ftz.f32 %f15906, %f6951, %f6954, %f15906; $L__BB108_832: cvt.u32.u64 %r1469, %rd167; add.s32 %r3172, %r1469, 4; add.s32 %r3171, %r3171, -4; setp.ne.s32 %p711, %r3171, 0; @%p711 bra $L__BB108_824; $L__BB108_833: setp.eq.s32 %p712, %r3174, 0; @%p712 bra $L__BB108_837; $L__BB108_834: .pragma "nounroll"; cvt.s64.s32 %rd169, %r3172; mul.wide.s32 %rd1368, %r3172, 80; add.s64 %rd1369, %rd1, %rd1368; add.s64 %rd170, %rd1369, 316; ld.local.u32 %r1470, [%rd1369+316]; setp.gt.s32 %p713, %r1470, 34; @%p713 bra $L__BB108_836; ld.local.f32 %f6955, [%rd170+-8]; ld.local.f32 %f6956, [%rd170+-12]; add.ftz.f32 %f6957, %f6956, %f6955; ld.local.f32 %f6958, [%rd170+-4]; add.ftz.f32 %f6959, %f6958, %f6957; mul.ftz.f32 %f6960, %f6959, 0f3EAAAAAB; abs.ftz.f32 %f6961, %f6960; ld.local.f32 %f6962, [%rd170+8]; ld.local.f32 %f6963, [%rd170+12]; ld.local.f32 %f6964, [%rd170+16]; fma.rn.ftz.f32 %f15904, %f6961, %f6962, %f15904; fma.rn.ftz.f32 %f15905, %f6961, %f6963, %f15905; fma.rn.ftz.f32 %f15906, %f6961, %f6964, %f15906; $L__BB108_836: cvt.u32.u64 %r1471, %rd169; add.s32 %r3172, %r1471, 1; add.s32 %r3174, %r3174, -1; setp.ne.s32 %p714, %r3174, 0; @%p714 bra $L__BB108_834; $L__BB108_837: setp.eq.ftz.f32 %p715, %f15905, 0f00000000; setp.eq.ftz.f32 %p716, %f15904, 0f00000000; and.pred %p717, %p716, %p715; setp.eq.ftz.f32 %p718, %f15906, 0f00000000; and.pred %p719, %p718, %p717; @%p719 bra $L__BB108_839; bra.uni $L__BB108_838; $L__BB108_839: ld.local.v2.f32 {%f15932, %f15933}, [%rd7+-72]; ld.local.f32 %f15931, [%rd7+-76]; bra.uni $L__BB108_840; $L__BB108_838: mul.ftz.f32 %f6965, %f15905, %f15905; fma.rn.ftz.f32 %f6966, %f15904, %f15904, %f6965; fma.rn.ftz.f32 %f6967, %f15906, %f15906, %f6966; sqrt.approx.ftz.f32 %f6968, %f6967; div.approx.ftz.f32 %f15933, %f15906, %f6968; div.approx.ftz.f32 %f15932, %f15905, %f6968; div.approx.ftz.f32 %f15931, %f15904, %f6968; $L__BB108_840: ld.const.s32 %rd1370, [kernel_params+1280]; add.s64 %rd1371, %rd166, %rd1370; cvta.to.global.u64 %rd1372, %rd618; shl.b64 %rd1373, %rd1371, 2; add.s64 %rd1374, %rd1372, %rd1373; add.s64 %rd1375, %rd1374, 4; add.s64 %rd1376, %rd1374, 8; atom.global.add.f32 %f6971, [%rd1374], %f15931; atom.global.add.f32 %f6972, [%rd1375], %f15932; atom.global.add.f32 %f6973, [%rd1376], %f15933; $L__BB108_841: and.b32 %r1472, %r402, 8; setp.eq.s32 %p720, %r1472, 0; @%p720 bra $L__BB108_886; ld.local.u32 %r416, [%rd7+172]; setp.lt.s32 %p721, %r416, 1; mov.f32 %f15949, 0f00000000; mov.f32 %f15937, %f15949; mov.f32 %f15938, %f15949; @%p721 bra $L__BB108_883; and.b32 %r417, %r416, 1; setp.eq.s32 %p722, %r416, 1; mov.f32 %f15938, 0f00000000; mov.u32 %r3177, 0; mov.f32 %f15937, %f15938; @%p722 bra $L__BB108_870; sub.s32 %r3176, %r416, %r417; mov.f32 %f15938, 0f00000000; mov.u32 %r3177, 0; mov.f32 %f15937, %f15938; $L__BB108_845: cvt.s64.s32 %rd171, %r3177; mul.wide.s32 %rd1377, %r3177, 80; add.s64 %rd1378, %rd1, %rd1377; add.s64 %rd172, %rd1378, 316; ld.local.u32 %r421, [%rd1378+316]; setp.gt.s32 %p723, %r421, 31; @%p723 bra $L__BB108_857; ld.local.v4.f32 {%f6981, %f6982, %f6983, %f6984}, [%rd172+-12]; setp.eq.s32 %p724, %r421, 3; @%p724 bra $L__BB108_855; bra.uni $L__BB108_847; $L__BB108_855: ld.local.f32 %f6994, [%rd172+20]; mul.ftz.f32 %f6995, %f6994, %f6994; mul.ftz.f32 %f15936, %f6995, %f6995; bra.uni $L__BB108_856; $L__BB108_847: setp.eq.s32 %p725, %r421, 5; @%p725 bra $L__BB108_854; bra.uni $L__BB108_848; $L__BB108_854: ld.local.f32 %f6992, [%rd172+20]; mul.ftz.f32 %f6993, %f6992, %f6992; mul.ftz.f32 %f15936, %f6993, %f6993; bra.uni $L__BB108_856; $L__BB108_848: add.s32 %r1475, %r421, -2; setp.lt.u32 %p726, %r1475, 7; mov.f32 %f15936, 0f00000000; @%p726 bra $L__BB108_856; setp.gt.u32 %p727, %r421, 31; @%p727 bra $L__BB108_851; mov.u32 %r1477, 1; shl.b32 %r1478, %r1477, %r421; and.b32 %r1479, %r1478, -2145385984; setp.ne.s32 %p728, %r1479, 0; @%p728 bra $L__BB108_856; $L__BB108_851: add.s32 %r1480, %r421, -10; setp.lt.u32 %p729, %r1480, 7; @%p729 bra $L__BB108_853; add.s32 %r1481, %r421, -22; setp.gt.u32 %p730, %r1481, 2; setp.ne.s32 %p731, %r421, 27; mov.f32 %f15936, 0f3F800000; and.pred %p732, %p730, %p731; @%p732 bra $L__BB108_856; $L__BB108_853: ld.local.v2.f32 {%f6988, %f6989}, [%rd172+20]; mul.ftz.f32 %f15936, %f6988, %f6989; $L__BB108_856: add.ftz.f32 %f6996, %f6981, %f6982; add.ftz.f32 %f6997, %f6983, %f6996; mul.ftz.f32 %f6998, %f6997, 0f3EAAAAAB; abs.ftz.f32 %f6999, %f6998; mov.f32 %f7000, 0f00000000; max.ftz.f32 %f7001, %f15936, %f7000; sqrt.approx.ftz.f32 %f7002, %f7001; sqrt.approx.ftz.f32 %f7003, %f7002; fma.rn.ftz.f32 %f15937, %f6999, %f7003, %f15937; add.ftz.f32 %f15938, %f15938, %f6999; $L__BB108_857: ld.local.u32 %r422, [%rd172+80]; setp.gt.s32 %p733, %r422, 31; @%p733 bra $L__BB108_869; ld.local.v4.f32 {%f7004, %f7005, %f7006, %f7007}, [%rd172+68]; setp.eq.s32 %p734, %r422, 3; @%p734 bra $L__BB108_867; bra.uni $L__BB108_859; $L__BB108_867: ld.local.f32 %f7017, [%rd172+100]; mul.ftz.f32 %f7018, %f7017, %f7017; mul.ftz.f32 %f15939, %f7018, %f7018; bra.uni $L__BB108_868; $L__BB108_859: setp.eq.s32 %p735, %r422, 5; @%p735 bra $L__BB108_866; bra.uni $L__BB108_860; $L__BB108_866: ld.local.f32 %f7015, [%rd172+100]; mul.ftz.f32 %f7016, %f7015, %f7015; mul.ftz.f32 %f15939, %f7016, %f7016; bra.uni $L__BB108_868; $L__BB108_860: add.s32 %r1482, %r422, -2; setp.lt.u32 %p736, %r1482, 7; mov.f32 %f15939, 0f00000000; @%p736 bra $L__BB108_868; setp.gt.u32 %p737, %r422, 31; @%p737 bra $L__BB108_863; mov.u32 %r1484, 1; shl.b32 %r1485, %r1484, %r422; and.b32 %r1486, %r1485, -2145385984; setp.ne.s32 %p738, %r1486, 0; @%p738 bra $L__BB108_868; $L__BB108_863: add.s32 %r1487, %r422, -10; setp.lt.u32 %p739, %r1487, 7; @%p739 bra $L__BB108_865; add.s32 %r1488, %r422, -22; setp.gt.u32 %p740, %r1488, 2; setp.ne.s32 %p741, %r422, 27; mov.f32 %f15939, 0f3F800000; and.pred %p742, %p740, %p741; @%p742 bra $L__BB108_868; $L__BB108_865: ld.local.v2.f32 {%f7011, %f7012}, [%rd172+100]; mul.ftz.f32 %f15939, %f7011, %f7012; $L__BB108_868: add.ftz.f32 %f7019, %f7004, %f7005; add.ftz.f32 %f7020, %f7006, %f7019; mul.ftz.f32 %f7021, %f7020, 0f3EAAAAAB; abs.ftz.f32 %f7022, %f7021; mov.f32 %f7023, 0f00000000; max.ftz.f32 %f7024, %f15939, %f7023; sqrt.approx.ftz.f32 %f7025, %f7024; sqrt.approx.ftz.f32 %f7026, %f7025; fma.rn.ftz.f32 %f15937, %f7022, %f7026, %f15937; add.ftz.f32 %f15938, %f15938, %f7022; $L__BB108_869: cvt.u32.u64 %r1489, %rd171; add.s32 %r3177, %r1489, 2; add.s32 %r3176, %r3176, -2; setp.ne.s32 %p743, %r3176, 0; @%p743 bra $L__BB108_845; $L__BB108_870: setp.eq.s32 %p744, %r417, 0; @%p744 bra $L__BB108_883; mul.wide.s32 %rd1379, %r3177, 80; add.s64 %rd1380, %rd1, %rd1379; add.s64 %rd173, %rd1380, 316; ld.local.u32 %r426, [%rd1380+316]; setp.gt.s32 %p745, %r426, 31; @%p745 bra $L__BB108_883; ld.local.f32 %f7027, [%rd173+-8]; ld.local.f32 %f7028, [%rd173+-12]; add.ftz.f32 %f7029, %f7028, %f7027; ld.local.f32 %f7030, [%rd173+-4]; add.ftz.f32 %f1224, %f7030, %f7029; setp.eq.s32 %p746, %r426, 3; @%p746 bra $L__BB108_881; bra.uni $L__BB108_873; $L__BB108_881: ld.local.f32 %f7040, [%rd173+20]; mul.ftz.f32 %f7041, %f7040, %f7040; mul.ftz.f32 %f15946, %f7041, %f7041; bra.uni $L__BB108_882; $L__BB108_873: setp.eq.s32 %p747, %r426, 5; @%p747 bra $L__BB108_880; bra.uni $L__BB108_874; $L__BB108_880: ld.local.f32 %f7038, [%rd173+20]; mul.ftz.f32 %f7039, %f7038, %f7038; mul.ftz.f32 %f15946, %f7039, %f7039; bra.uni $L__BB108_882; $L__BB108_2267: mul.ftz.f32 %f14566, %f16697, %f3942; fma.rn.ftz.f32 %f14567, %f16696, %f3941, %f14566; fma.rn.ftz.f32 %f14568, %f16698, %f3943, %f14567; mul.ftz.f32 %f14569, %f16694, %f3942; fma.rn.ftz.f32 %f14570, %f16693, %f3941, %f14569; fma.rn.ftz.f32 %f14571, %f16695, %f3943, %f14570; mul.ftz.f32 %f14572, %f3900, %f3942; fma.rn.ftz.f32 %f14573, %f3898, %f3941, %f14572; fma.rn.ftz.f32 %f14574, %f3901, %f3943, %f14573; neg.ftz.f32 %f14575, %f14568; mul.ftz.f32 %f14576, %f14367, %f14574; div.approx.ftz.f32 %f14577, %f14575, %f14576; neg.ftz.f32 %f14578, %f14571; mul.ftz.f32 %f14579, %f14368, %f14574; div.approx.ftz.f32 %f14580, %f14578, %f14579; fma.rn.ftz.f32 %f14581, %f14577, %f14577, 0f3F800000; mov.f32 %f14582, 0f3F800000; fma.rn.ftz.f32 %f14583, %f14580, %f14580, %f14581; mul.ftz.f32 %f14584, %f14574, %f14574; mul.ftz.f32 %f14585, %f14584, %f14584; mul.ftz.f32 %f14586, %f14583, %f14583; mul.ftz.f32 %f14587, %f14586, 0f40490FDB; mul.ftz.f32 %f14588, %f16710, %f14587; mul.ftz.f32 %f14589, %f14585, %f14588; rcp.approx.ftz.f32 %f16711, %f14589; mul.ftz.f32 %f14590, %f3965, %f3965; sub.ftz.f32 %f14591, %f14582, %f14590; div.approx.ftz.f32 %f14592, %f14591, %f14590; mul.ftz.f32 %f14593, %f16697, %f16741; fma.rn.ftz.f32 %f14594, %f16696, %f16740, %f14593; fma.rn.ftz.f32 %f14595, %f16698, %f16742, %f14594; mul.ftz.f32 %f14596, %f16694, %f16741; fma.rn.ftz.f32 %f14597, %f16693, %f16740, %f14596; fma.rn.ftz.f32 %f14598, %f16695, %f16742, %f14597; mul.ftz.f32 %f14599, %f14595, %f14595; mul.ftz.f32 %f14600, %f14367, %f14367; mul.ftz.f32 %f14601, %f14598, %f14598; mul.ftz.f32 %f14602, %f14368, %f14368; mul.ftz.f32 %f14603, %f14602, %f14601; fma.rn.ftz.f32 %f14604, %f14600, %f14599, %f14603; add.ftz.f32 %f14605, %f14599, %f14601; div.approx.ftz.f32 %f14606, %f14604, %f14605; mul.ftz.f32 %f16714, %f14592, %f14606; bra.uni $L__BB108_2274; $L__BB108_2120: setp.lt.ftz.f32 %p1888, %f2799, 0f3F400000; @%p1888 bra $L__BB108_2124; bra.uni $L__BB108_2121; $L__BB108_2124: add.ftz.f32 %f13229, %f2799, 0fBF000000; mul.ftz.f32 %f13230, %f13229, 0f40800000; mul.ftz.f32 %f13231, %f13230, 0f3FC90FDB; sin.approx.ftz.f32 %f13232, %f13231; cos.approx.ftz.f32 %f13233, %f13231; div.approx.ftz.f32 %f13234, %f13232, %f13233; mul.ftz.f32 %f3519, %f3512, %f13234; abs.ftz.f32 %f3520, %f3519; setp.leu.ftz.f32 %p1892, %f3520, 0f3F800000; mov.f32 %f16579, %f3520; @%p1892 bra $L__BB108_2126; rcp.approx.ftz.f32 %f16579, %f3520; $L__BB108_2126: mul.rn.ftz.f32 %f13235, %f16579, %f16579; mov.f32 %f13236, 0fC0B59883; mov.f32 %f13237, 0fBF52C7EA; fma.rn.ftz.f32 %f13238, %f13235, %f13237, %f13236; mov.f32 %f13239, 0fC0D21907; fma.rn.ftz.f32 %f13240, %f13238, %f13235, %f13239; mul.ftz.f32 %f13241, %f13235, %f13240; mul.ftz.f32 %f13242, %f16579, %f13241; add.ftz.f32 %f13243, %f13235, 0f41355DC0; mov.f32 %f13244, 0f41E6BD60; fma.rn.ftz.f32 %f13245, %f13243, %f13235, %f13244; mov.f32 %f13246, 0f419D92C8; fma.rn.ftz.f32 %f13247, %f13245, %f13235, %f13246; rcp.approx.ftz.f32 %f13248, %f13247; fma.rn.ftz.f32 %f13249, %f13242, %f13248, %f16579; mov.f32 %f13250, 0f3FC90FDB; sub.ftz.f32 %f13251, %f13250, %f13249; setp.gt.ftz.f32 %p1893, %f3520, 0f3F800000; selp.f32 %f13252, %f13251, %f13249, %p1893; mov.b32 %r2819, %f13252; mov.b32 %r2820, %f3519; and.b32 %r2821, %r2820, -2147483648; or.b32 %r2822, %r2821, %r2819; mov.b32 %f13253, %r2822; setp.le.ftz.f32 %p1894, %f3520, 0f7F800000; selp.f32 %f13254, %f13253, %f13252, %p1894; cos.approx.ftz.f32 %f13255, %f13254; sin.approx.ftz.f32 %f13256, %f13254; mul.ftz.f32 %f13257, %f3491, %f13255; mul.ftz.f32 %f13258, %f3492, %f13256; mul.ftz.f32 %f13259, %f13256, %f13258; fma.rn.ftz.f32 %f13260, %f13255, %f13257, %f13259; add.ftz.f32 %f13261, %f13260, 0f3F800000; rcp.approx.ftz.f32 %f13262, %f13261; lg2.approx.ftz.f32 %f13263, %f2775; mul.ftz.f32 %f13264, %f13263, %f13262; ex2.approx.ftz.f32 %f16582, %f13264; add.ftz.f32 %f16583, %f13254, 0f40490FDB; bra.uni $L__BB108_2134; $L__BB108_2269: mul.ftz.f32 %f14607, %f3940, %f3940; rcp.approx.ftz.f32 %f14608, %f14607; add.ftz.f32 %f14609, %f14608, 0fBF800000; mul.ftz.f32 %f14610, %f14607, %f14607; mul.ftz.f32 %f14611, %f14610, 0f40490FDB; add.ftz.f32 %f14612, %f16710, %f14609; mul.ftz.f32 %f14613, %f14611, %f14612; mul.ftz.f32 %f14614, %f14612, %f14613; div.approx.ftz.f32 %f16711, %f16710, %f14614; $L__BB108_2273: mul.ftz.f32 %f14635, %f3965, %f3965; mov.f32 %f14636, 0f3F800000; sub.ftz.f32 %f14637, %f14636, %f14635; mul.ftz.f32 %f14638, %f14637, %f16710; div.approx.ftz.f32 %f16714, %f14638, %f14635; $L__BB108_2274: mul.ftz.f32 %f14642, %f3937, %f16711; mul.ftz.f32 %f14643, %f14642, 0f3E800000; div.approx.ftz.f32 %f16743, %f14643, %f3902; add.s32 %r2878, %r873, -11; setp.gt.u32 %p2011, %r2878, 1; mov.f32 %f16717, 0f3F800000; mov.f32 %f16718, %f16717; mov.f32 %f16719, %f16717; @%p2011 bra $L__BB108_2280; ld.local.f32 %f14645, [%rd587+28]; fma.rn.ftz.f32 %f3980, %f14645, %f14645, 0fBF800000; mov.f32 %f14644, 0f3F800000; abs.ftz.f32 %f3981, %f14644; fma.rn.ftz.f32 %f3982, %f3981, %f3981, %f3980; setp.leu.ftz.f32 %p2012, %f3982, 0f00000000; mov.f32 %f16715, %f14644; @%p2012 bra $L__BB108_2277; sqrt.approx.ftz.f32 %f14646, %f3982; sub.ftz.f32 %f14647, %f14646, %f3981; add.ftz.f32 %f14648, %f3981, %f14646; div.approx.ftz.f32 %f14649, %f14647, %f14648; fma.rn.ftz.f32 %f14650, %f3981, %f14648, 0fBF800000; fma.rn.ftz.f32 %f14651, %f3981, %f14647, 0f3F800000; div.approx.ftz.f32 %f14652, %f14650, %f14651; mul.ftz.f32 %f14653, %f14649, 0f3F000000; mul.ftz.f32 %f14654, %f14649, %f14653; fma.rn.ftz.f32 %f14655, %f14652, %f14652, 0f3F800000; mul.ftz.f32 %f16715, %f14654, %f14655; $L__BB108_2277: ld.local.u64 %rd2688, [%rd587+36]; ld.f32 %f3985, [%rd2688+12]; ld.f32 %f3986, [%rd2688+16]; ld.f32 %f3987, [%rd2688+20]; sub.ftz.f32 %f3988, %f14644, %f16715; mul.ftz.f32 %f14657, %f3942, %f16741; fma.rn.ftz.f32 %f14658, %f3941, %f16740, %f14657; fma.rn.ftz.f32 %f14659, %f3943, %f16742, %f14658; abs.ftz.f32 %f3989, %f14659; fma.rn.ftz.f32 %f3990, %f3989, %f3989, %f3980; setp.leu.ftz.f32 %p2013, %f3990, 0f00000000; mov.f32 %f16716, %f3988; @%p2013 bra $L__BB108_2279; sqrt.approx.ftz.f32 %f14660, %f3990; sub.ftz.f32 %f14661, %f14660, %f3989; add.ftz.f32 %f14662, %f3989, %f14660; div.approx.ftz.f32 %f14663, %f14661, %f14662; fma.rn.ftz.f32 %f14664, %f3989, %f14662, 0fBF800000; fma.rn.ftz.f32 %f14665, %f3989, %f14661, 0f3F800000; div.approx.ftz.f32 %f14666, %f14664, %f14665; mul.ftz.f32 %f14667, %f14663, 0f3F000000; mul.ftz.f32 %f14668, %f14663, %f14667; fma.rn.ftz.f32 %f14669, %f14666, %f14666, 0f3F800000; mul.ftz.f32 %f14670, %f14668, %f14669; sub.ftz.f32 %f16716, %f14670, %f16715; $L__BB108_2279: rcp.approx.ftz.f32 %f14671, %f3988; mov.f32 %f14672, 0f3F800000; mul.ftz.f32 %f14673, %f14671, %f16716; sub.ftz.f32 %f14674, %f14672, %f14673; fma.rn.ftz.f32 %f16719, %f3987, %f14674, %f14673; fma.rn.ftz.f32 %f16718, %f3986, %f14674, %f14673; fma.rn.ftz.f32 %f16717, %f3985, %f14674, %f14673; $L__BB108_2280: add.ftz.f32 %f14675, %f16714, 0f3F800000; mov.f32 %f14676, 0f00000000; max.ftz.f32 %f14677, %f14675, %f14676; sqrt.approx.ftz.f32 %f14678, %f14677; add.ftz.f32 %f14679, %f14678, 0f3F800000; mov.f32 %f14680, 0f40000000; div.approx.ftz.f32 %f14681, %f14680, %f14679; mul.ftz.f32 %f14682, %f14681, %f16743; mul.ftz.f32 %f16744, %f14682, %f16717; mul.ftz.f32 %f16745, %f14682, %f16718; mul.ftz.f32 %f16746, %f14682, %f16719; mov.u32 %r3285, 10; $L__BB108_2281: setp.ne.s32 %p2014, %r873, 12; @%p2014 bra $L__BB108_2309; ld.local.u64 %rd2689, [%rd587+36]; ld.f32 %f14683, [%rd2689+36]; mul.ftz.f32 %f14684, %f14683, 0f3E800000; mul.ftz.f32 %f16744, %f16744, %f14684; mul.ftz.f32 %f16745, %f16745, %f14684; mul.ftz.f32 %f16746, %f16746, %f14684; bra.uni $L__BB108_2309; $L__BB108_874: add.s32 %r1490, %r426, -2; setp.lt.u32 %p748, %r1490, 7; mov.f32 %f15946, 0f00000000; @%p748 bra $L__BB108_882; setp.gt.u32 %p749, %r426, 31; @%p749 bra $L__BB108_877; mov.u32 %r1492, 1; shl.b32 %r1493, %r1492, %r426; and.b32 %r1494, %r1493, -2145385984; setp.ne.s32 %p750, %r1494, 0; @%p750 bra $L__BB108_882; bra.uni $L__BB108_877; $L__BB108_882: mul.ftz.f32 %f7042, %f1224, 0f3EAAAAAB; abs.ftz.f32 %f7043, %f7042; mov.f32 %f7044, 0f00000000; max.ftz.f32 %f7045, %f15946, %f7044; sqrt.approx.ftz.f32 %f7046, %f7045; sqrt.approx.ftz.f32 %f7047, %f7046; fma.rn.ftz.f32 %f15937, %f7043, %f7047, %f15937; add.ftz.f32 %f15938, %f15938, %f7043; $L__BB108_883: setp.leu.ftz.f32 %p755, %f15938, 0f00000000; @%p755 bra $L__BB108_885; div.approx.ftz.f32 %f15949, %f15937, %f15938; $L__BB108_885: ld.const.s32 %rd1381, [kernel_params+1284]; add.s64 %rd1382, %rd166, %rd1381; cvta.to.global.u64 %rd1383, %rd618; shl.b64 %rd1384, %rd1382, 2; add.s64 %rd1385, %rd1383, %rd1384; atom.global.add.f32 %f7049, [%rd1385], %f15949; $L__BB108_886: and.b32 %r1497, %r402, 16; setp.eq.s32 %p756, %r1497, 0; @%p756 bra $L__BB108_924; ld.local.u32 %r427, [%rd7+-12]; ld.local.s32 %rd174, [%rd7+-28]; setp.eq.s32 %p757, %r427, -1; mov.u32 %r448, -1; mov.u32 %r445, 0; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p757 bra $L__BB108_902; ld.const.u64 %rd1386, [kernel_params+2160]; cvta.to.global.u64 %rd1387, %rd1386; mul.wide.s32 %rd1388, %r427, 208; add.s64 %rd1389, %rd1387, %rd1388; ld.global.u32 %r428, [%rd1389+164]; ld.local.u32 %r1503, [%rd7+-24]; and.b32 %r1504, %r1503, 1; setp.eq.b32 %p758, %r1504, 1; mov.pred %p759, 0; xor.pred %p760, %p758, %p759; not.pred %p761, %p760; mov.u32 %r3179, 0; @%p761 bra $L__BB108_892; cvt.u32.u64 %r1506, %rd174; setp.eq.s32 %p762, %r1506, -1; mov.u32 %r3178, -1; @%p762 bra $L__BB108_891; ld.const.u64 %rd1390, [kernel_params+2240]; cvta.to.global.u64 %rd1391, %rd1390; shl.b64 %rd1392, %rd174, 2; add.s64 %rd1393, %rd1391, %rd1392; ld.global.u32 %r3178, [%rd1393]; $L__BB108_891: setp.ne.s32 %p763, %r3178, -1; selp.u32 %r3179, 1, 0, %p763; $L__BB108_892: ld.const.u64 %rd1394, [kernel_params+2312]; cvta.to.global.u64 %rd175, %rd1394; add.s32 %r3181, %r3179, %r428; mul.wide.u32 %rd1395, %r3181, 16; add.s64 %rd1396, %rd175, %rd1395; add.s64 %rd2958, %rd1396, 14; ld.global.u16 %rs141, [%rd1396+12]; ld.global.u32 %r3180, [%rd1396+8]; ld.global.u64 %rd2956, [%rd1396]; setp.eq.s64 %p764, %rd2956, 3; add.s64 %rd2957, %rd1396, 15; @%p764 bra $L__BB108_898; $L__BB108_893: setp.eq.s64 %p765, %rd2956, 0; @%p765 bra $L__BB108_895; add.s32 %r3181, %r3181, 2; bra.uni $L__BB108_896; $L__BB108_895: setp.eq.s16 %p766, %rs141, 0; mov.u32 %r445, 0; mov.u32 %r3181, %r3180; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p766 bra $L__BB108_902; $L__BB108_896: cvt.u64.u32 %rd180, %r3181; mul.wide.u32 %rd1397, %r3181, 16; add.s64 %rd1398, %rd175, %rd1397; ld.global.u16 %rs141, [%rd1398+12]; ld.global.u32 %r3180, [%rd1398+8]; ld.global.u64 %rd2956, [%rd1398]; setp.ne.s64 %p767, %rd2956, 3; @%p767 bra $L__BB108_893; shl.b64 %rd1399, %rd180, 4; add.s64 %rd1400, %rd175, %rd1399; add.s64 %rd2958, %rd1400, 14; add.s64 %rd2957, %rd1400, 15; $L__BB108_898: ld.global.u8 %rs20, [%rd2958]; ld.global.u8 %rs21, [%rd2957]; cvt.u32.u64 %r1511, %rd174; setp.ne.s32 %p768, %r1511, -1; @%p768 bra $L__BB108_901; mov.u32 %r445, 0; add.s16 %rs78, %rs141, -1; setp.lt.u16 %p769, %rs78, 2; @%p769 bra $L__BB108_901; setp.ne.s16 %p770, %rs141, 1024; mov.u32 %r3185, %r445; mov.u32 %r3186, %r445; @%p770 bra $L__BB108_902; $L__BB108_901: setp.eq.s16 %p771, %rs141, 0; selp.b32 %r448, -1, %r3180, %p771; cvt.u32.u16 %r445, %rs141; cvt.u32.u16 %r1516, %rs21; and.b32 %r3186, %r1516, 255; cvt.u32.u16 %r1517, %rs20; and.b32 %r3185, %r1517, 255; $L__BB108_902: setp.eq.s32 %p772, %r448, -1; mov.f32 %f15956, 0f00000000; mov.f32 %f15957, %f15956; mov.f32 %f15960, %f15956; @%p772 bra $L__BB108_923; ld.local.u32 %r451, [%rd7+-24]; and.b32 %r1518, %r451, 1; setp.eq.b32 %p773, %r1518, 1; mov.pred %p774, 0; xor.pred %p775, %p773, %p774; not.pred %p776, %p775; @%p776 bra $L__BB108_915; bra.uni $L__BB108_904; $L__BB108_915: and.b32 %r1539, %r451, 6; setp.eq.s32 %p783, %r1539, 0; @%p783 bra $L__BB108_920; and.b32 %r1540, %r445, 768; setp.eq.s32 %p784, %r1540, 0; @%p784 bra $L__BB108_918; cvt.u32.u64 %r1541, %rd174; ld.const.u64 %rd1427, [kernel_params+2264]; cvta.to.global.u64 %rd1428, %rd1427; mul.wide.s32 %rd1429, %r1541, 16; add.s64 %rd1430, %rd1428, %rd1429; shr.s32 %r1542, %r451, 7; ld.global.u32 %r1543, [%rd1430+4]; add.s32 %r1544, %r1542, %r1543; ld.const.u64 %rd1431, [kernel_params+2328]; cvta.to.global.u64 %rd1432, %rd1431; add.s32 %r1545, %r1544, %r448; mul.wide.s32 %rd1433, %r1545, 8; add.s64 %rd1434, %rd1432, %rd1433; ld.global.v2.f32 {%f7078, %f7079}, [%rd1434]; ld.global.v2.f32 {%f7082, %f7083}, [%rd1434+8]; ld.local.f32 %f7086, [%rd7+-20]; mov.f32 %f7087, 0f3F800000; sub.ftz.f32 %f7088, %f7087, %f7086; mul.ftz.f32 %f7089, %f7078, %f7088; mul.ftz.f32 %f7090, %f7079, %f7088; fma.rn.ftz.f32 %f15957, %f7083, %f7086, %f7090; fma.rn.ftz.f32 %f15956, %f7082, %f7086, %f7089; bra.uni $L__BB108_922; $L__BB108_904: cvt.u32.u64 %r1520, %rd174; setp.eq.s32 %p777, %r1520, -1; mov.u32 %r3188, -1; @%p777 bra $L__BB108_906; ld.const.u64 %rd1401, [kernel_params+2240]; cvta.to.global.u64 %rd1402, %rd1401; shl.b64 %rd1403, %rd174, 2; add.s64 %rd1404, %rd1402, %rd1403; ld.global.u32 %r3188, [%rd1404]; $L__BB108_906: setp.eq.s32 %p778, %r3188, -1; @%p778 bra $L__BB108_908; mov.u64 %rd1405, 0; { // callseq 776, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r445; st.param.b32 [param1+4], %r3185; st.param.b32 [param1+8], %r3186; st.param.b32 [param1+12], %r448; .param .b64 param2; st.param.b64 [param2+0], %rd1405; .param .b64 param3; st.param.b64 [param3+0], %rd1405; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15956, [retval0+0]; ld.param.f32 %f15957, [retval0+4]; } // callseq 776 bra.uni $L__BB108_922; $L__BB108_920: and.b32 %r1550, %r451, 8; setp.eq.s32 %p787, %r1550, 0; setp.ne.s32 %p788, %r445, 8; mov.f32 %f15956, 0f00000000; or.pred %p789, %p788, %p787; mov.f32 %f15957, %f15956; @%p789 bra $L__BB108_922; cvt.u32.u64 %r1551, %rd174; ld.const.u64 %rd1439, [kernel_params+2328]; cvta.to.global.u64 %rd1440, %rd1439; add.s32 %r1552, %r1551, %r448; mul.wide.s32 %rd1441, %r1552, 8; add.s64 %rd1442, %rd1440, %rd1441; ld.global.v2.f32 {%f15956, %f15957}, [%rd1442]; bra.uni $L__BB108_922; $L__BB108_908: and.b32 %r1521, %r445, 56; setp.eq.s32 %p779, %r1521, 0; @%p779 bra $L__BB108_913; and.b32 %r1522, %r445, 24; setp.eq.s32 %p780, %r1522, 0; @%p780 bra $L__BB108_911; ld.const.u64 %rd1407, [kernel_params+2232]; cvta.to.global.u64 %rd1408, %rd1407; shl.b64 %rd1409, %rd174, 4; add.s64 %rd1410, %rd1408, %rd1409; ld.global.v4.u32 {%r1523, %r1524, %r1525, %r1526}, [%rd1410]; ld.const.u64 %rd1411, [kernel_params+2328]; cvta.to.global.u64 %rd1412, %rd1411; add.s32 %r1530, %r1523, %r448; mul.wide.u32 %rd1413, %r1530, 8; add.s64 %rd1414, %rd1412, %rd1413; ld.global.v2.f32 {%f15954, %f15955}, [%rd1414]; add.s32 %r1531, %r1524, %r448; mul.wide.u32 %rd1415, %r1531, 8; add.s64 %rd1416, %rd1412, %rd1415; ld.global.v2.f32 {%f15952, %f15953}, [%rd1416]; add.s32 %r1532, %r1525, %r448; mul.wide.u32 %rd1417, %r1532, 8; add.s64 %rd1418, %rd1412, %rd1417; ld.global.v2.f32 {%f15950, %f15951}, [%rd1418]; bra.uni $L__BB108_912; $L__BB108_918: and.b32 %r1546, %r445, 131; setp.eq.s32 %p785, %r1546, 0; mov.f32 %f15956, 0f00000000; mov.f32 %f15957, %f15956; @%p785 bra $L__BB108_922; cvt.u32.u64 %r1547, %rd174; setp.eq.s32 %p786, %r445, 128; selp.b32 %r1548, %r1547, 0, %p786; add.s32 %r1549, %r1548, %r448; ld.const.u64 %rd1435, [kernel_params+2328]; cvta.to.global.u64 %rd1436, %rd1435; mul.wide.s32 %rd1437, %r1549, 8; add.s64 %rd1438, %rd1436, %rd1437; ld.global.v2.f32 {%f15956, %f15957}, [%rd1438]; bra.uni $L__BB108_922; $L__BB108_913: and.b32 %r1535, %r445, 7; setp.eq.s32 %p781, %r1535, 0; mov.f32 %f15956, 0f00000000; mov.f32 %f15957, %f15956; @%p781 bra $L__BB108_922; setp.eq.s32 %p782, %r445, 4; selp.b32 %r1537, %r1520, 0, %p782; add.s32 %r1538, %r1537, %r448; ld.const.u64 %rd1423, [kernel_params+2328]; cvta.to.global.u64 %rd1424, %rd1423; mul.wide.s32 %rd1425, %r1538, 8; add.s64 %rd1426, %rd1424, %rd1425; ld.global.v2.f32 {%f15956, %f15957}, [%rd1426]; bra.uni $L__BB108_922; $L__BB108_2230: mul.ftz.f32 %f14208, %f16630, %f3847; fma.rn.ftz.f32 %f14209, %f16629, %f3846, %f14208; fma.rn.ftz.f32 %f14210, %f16631, %f3848, %f14209; mul.ftz.f32 %f14211, %f16627, %f3847; fma.rn.ftz.f32 %f14212, %f16626, %f3846, %f14211; fma.rn.ftz.f32 %f14213, %f16628, %f3848, %f14212; mul.ftz.f32 %f14214, %f3738, %f3847; fma.rn.ftz.f32 %f14215, %f3736, %f3846, %f14214; fma.rn.ftz.f32 %f14216, %f3739, %f3848, %f14215; neg.ftz.f32 %f14217, %f14210; mul.ftz.f32 %f14218, %f13816, %f14216; div.approx.ftz.f32 %f14219, %f14217, %f14218; neg.ftz.f32 %f14220, %f14213; mul.ftz.f32 %f14221, %f13817, %f14216; div.approx.ftz.f32 %f14222, %f14220, %f14221; mul.ftz.f32 %f14223, %f14216, %f14216; mul.ftz.f32 %f14224, %f14223, %f14223; mul.ftz.f32 %f14225, %f14219, %f14219; neg.ftz.f32 %f14226, %f14225; mul.ftz.f32 %f14227, %f14222, %f14222; sub.ftz.f32 %f14228, %f14226, %f14227; mul.ftz.f32 %f14229, %f14228, 0f3FB8AA3B; ex2.approx.ftz.f32 %f14230, %f14229; mul.ftz.f32 %f14231, %f3854, %f14224; div.approx.ftz.f32 %f16681, %f14230, %f14231; mul.ftz.f32 %f14232, %f16630, %f16741; fma.rn.ftz.f32 %f14233, %f16629, %f16740, %f14232; fma.rn.ftz.f32 %f14234, %f16631, %f16742, %f14233; mul.ftz.f32 %f14235, %f16627, %f16741; fma.rn.ftz.f32 %f14236, %f16626, %f16740, %f14235; fma.rn.ftz.f32 %f14237, %f16628, %f16742, %f14236; mul.ftz.f32 %f14238, %f14237, %f14237; mul.ftz.f32 %f14239, %f14234, %f14234; mul.ftz.f32 %f14240, %f13816, %f13816; mul.ftz.f32 %f14241, %f13817, %f13817; mul.ftz.f32 %f14242, %f14241, %f14238; fma.rn.ftz.f32 %f14243, %f14240, %f14239, %f14242; add.ftz.f32 %f14244, %f14239, %f14238; div.approx.ftz.f32 %f14245, %f14243, %f14244; mul.ftz.f32 %f14246, %f3856, %f14245; div.approx.ftz.f32 %f14247, %f14246, %f3855; mov.f32 %f14248, 0f00000000; max.ftz.f32 %f14249, %f14247, %f14248; sqrt.approx.ftz.f32 %f3858, %f14249; setp.lt.ftz.f32 %p1979, %f3858, 0f3F200000; @%p1979 bra $L__BB108_2234; rcp.approx.ftz.f32 %f14250, %f3858; fma.rn.ftz.f32 %f14251, %f14250, 0f400B9581, 0f40623D71; mul.ftz.f32 %f14252, %f14250, %f14251; fma.rn.ftz.f32 %f14253, %f14250, 0f4024ED91, 0f4011A9FC; fma.rn.ftz.f32 %f14254, %f14250, %f14253, 0f3F800000; div.approx.ftz.f32 %f16682, %f14252, %f14254; $L__BB108_2234: mul.ftz.f32 %f14274, %f16681, 0f3E800000; div.approx.ftz.f32 %f14275, %f14274, %f3740; mul.ftz.f32 %f14276, %f16678, %f16682; mul.ftz.f32 %f16744, %f14276, %f14275; mul.ftz.f32 %f16743, %f16678, %f14275; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; mov.u32 %r3285, %r2865; bra.uni $L__BB108_2309; $L__BB108_2121: mov.f32 %f13191, 0f3F800000; sub.ftz.f32 %f13192, %f13191, %f2799; mul.ftz.f32 %f13193, %f13192, 0f40800000; mul.ftz.f32 %f13194, %f13193, 0f3FC90FDB; sin.approx.ftz.f32 %f13195, %f13194; cos.approx.ftz.f32 %f13196, %f13194; div.approx.ftz.f32 %f13197, %f13195, %f13196; mul.ftz.f32 %f3513, %f3512, %f13197; abs.ftz.f32 %f3514, %f3513; setp.leu.ftz.f32 %p1889, %f3514, 0f3F800000; mov.f32 %f16578, %f3514; @%p1889 bra $L__BB108_2123; rcp.approx.ftz.f32 %f16578, %f3514; $L__BB108_2123: mul.rn.ftz.f32 %f13198, %f16578, %f16578; mov.f32 %f13199, 0fC0B59883; mov.f32 %f13200, 0fBF52C7EA; fma.rn.ftz.f32 %f13201, %f13198, %f13200, %f13199; mov.f32 %f13202, 0fC0D21907; fma.rn.ftz.f32 %f13203, %f13201, %f13198, %f13202; mul.ftz.f32 %f13204, %f13198, %f13203; mul.ftz.f32 %f13205, %f16578, %f13204; add.ftz.f32 %f13206, %f13198, 0f41355DC0; mov.f32 %f13207, 0f41E6BD60; fma.rn.ftz.f32 %f13208, %f13206, %f13198, %f13207; mov.f32 %f13209, 0f419D92C8; fma.rn.ftz.f32 %f13210, %f13208, %f13198, %f13209; rcp.approx.ftz.f32 %f13211, %f13210; fma.rn.ftz.f32 %f13212, %f13205, %f13211, %f16578; mov.f32 %f13213, 0f3FC90FDB; sub.ftz.f32 %f13214, %f13213, %f13212; setp.gt.ftz.f32 %p1890, %f3514, 0f3F800000; selp.f32 %f13215, %f13214, %f13212, %p1890; mov.b32 %r2815, %f13215; mov.b32 %r2816, %f3513; and.b32 %r2817, %r2816, -2147483648; or.b32 %r2818, %r2817, %r2815; mov.b32 %f13216, %r2818; setp.le.ftz.f32 %p1891, %f3514, 0f7F800000; selp.f32 %f13217, %f13216, %f13215, %p1891; cos.approx.ftz.f32 %f13218, %f13217; sin.approx.ftz.f32 %f13219, %f13217; mul.ftz.f32 %f13220, %f3491, %f13218; mul.ftz.f32 %f13221, %f3492, %f13219; mul.ftz.f32 %f13222, %f13219, %f13221; fma.rn.ftz.f32 %f13223, %f13218, %f13220, %f13222; add.ftz.f32 %f13224, %f13223, 0f3F800000; rcp.approx.ftz.f32 %f13225, %f13224; lg2.approx.ftz.f32 %f13226, %f2775; mul.ftz.f32 %f13227, %f13226, %f13225; ex2.approx.ftz.f32 %f16582, %f13227; mov.f32 %f13228, 0f40C90FDB; sub.ftz.f32 %f16583, %f13228, %f13217; $L__BB108_2134: mul.ftz.f32 %f13344, %f16582, %f16582; mov.f32 %f16584, 0f3F800000; sub.ftz.f32 %f13346, %f16584, %f13344; mov.f32 %f16587, 0f00000000; max.ftz.f32 %f13348, %f16587, %f13346; sqrt.approx.ftz.f32 %f13349, %f13348; cos.approx.ftz.f32 %f13350, %f16583; mul.ftz.f32 %f13351, %f13349, %f13350; sin.approx.ftz.f32 %f13352, %f16583; mul.ftz.f32 %f13353, %f13349, %f13352; mul.ftz.f32 %f13354, %f16572, %f13353; mul.ftz.f32 %f13355, %f16573, %f13353; mul.ftz.f32 %f13356, %f16574, %f13353; fma.rn.ftz.f32 %f13357, %f16575, %f13351, %f13354; fma.rn.ftz.f32 %f13358, %f16576, %f13351, %f13355; fma.rn.ftz.f32 %f13359, %f16577, %f13351, %f13356; fma.rn.ftz.f32 %f13360, %f3484, %f16582, %f13357; fma.rn.ftz.f32 %f13361, %f3486, %f16582, %f13358; fma.rn.ftz.f32 %f13362, %f3487, %f16582, %f13359; mul.ftz.f32 %f13363, %f13139, %f13361; fma.rn.ftz.f32 %f13364, %f3485, %f13360, %f13363; fma.rn.ftz.f32 %f13365, %f13140, %f13362, %f13364; setp.lt.ftz.f32 %p1901, %f13365, 0f00000000; neg.ftz.f32 %f13366, %f13360; neg.ftz.f32 %f13367, %f13361; neg.ftz.f32 %f13368, %f13362; selp.f32 %f13369, %f13366, %f13360, %p1901; selp.f32 %f13370, %f13367, %f13361, %p1901; selp.f32 %f13371, %f13368, %f13362, %p1901; add.ftz.f32 %f13372, %f13365, %f13365; mul.ftz.f32 %f13373, %f13372, %f13369; mul.ftz.f32 %f13374, %f13372, %f13370; mul.ftz.f32 %f13375, %f13372, %f13371; sub.ftz.f32 %f16740, %f13373, %f3485; sub.ftz.f32 %f16741, %f13374, %f13139; sub.ftz.f32 %f16742, %f13375, %f13140; max.ftz.f32 %f13376, %f13143, %f13144; setp.le.ftz.f32 %p1902, %f13376, 0f38D1B717; mov.u32 %r3285, 18; mov.f32 %f16743, 0f49742400; mov.f32 %f16744, %f16743; mov.f32 %f16745, %f16743; mov.f32 %f16746, %f16743; @%p1902 bra $L__BB108_2309; mul.ftz.f32 %f13381, %f3486, %f16741; fma.rn.ftz.f32 %f13382, %f3484, %f16740, %f13381; fma.rn.ftz.f32 %f3544, %f3487, %f16742, %f13382; setp.leu.ftz.f32 %p1903, %f3544, 0f00000000; mov.f32 %f16744, %f16587; @%p1903 bra $L__BB108_2142; add.ftz.f32 %f13383, %f3485, %f16740; add.ftz.f32 %f13384, %f13139, %f16741; mul.ftz.f32 %f13385, %f13384, %f13384; fma.rn.ftz.f32 %f13386, %f13383, %f13383, %f13385; add.ftz.f32 %f13387, %f13140, %f16742; fma.rn.ftz.f32 %f13388, %f13387, %f13387, %f13386; sqrt.approx.ftz.f32 %f13389, %f13388; div.approx.ftz.f32 %f3545, %f13383, %f13389; div.approx.ftz.f32 %f3546, %f13384, %f13389; div.approx.ftz.f32 %f3547, %f13387, %f13389; mul.ftz.f32 %f13390, %f13139, %f3546; fma.rn.ftz.f32 %f13391, %f3485, %f3545, %f13390; fma.rn.ftz.f32 %f13392, %f13140, %f3547, %f13391; abs.ftz.f32 %f13393, %f13392; mov.f32 %f13394, 0f358637BD; max.ftz.f32 %f3548, %f13393, %f13394; mul.ftz.f32 %f13395, %f3486, %f3546; fma.rn.ftz.f32 %f13396, %f3484, %f3545, %f13395; fma.rn.ftz.f32 %f13397, %f3487, %f3547, %f13396; max.ftz.f32 %f3549, %f13397, %f13394; max.ftz.f32 %f3550, %f3544, %f13394; @%p1881 bra $L__BB108_2140; bra.uni $L__BB108_2137; $L__BB108_2140: lg2.approx.ftz.f32 %f13448, %f3549; mul.ftz.f32 %f13449, %f3491, %f13448; ex2.approx.ftz.f32 %f13450, %f13449; mov.f32 %f13451, 0f41C90FDB; div.approx.ftz.f32 %f13452, %f3511, %f13451; mul.ftz.f32 %f13453, %f3550, %f13452; mul.ftz.f32 %f16586, %f13450, %f13453; mul.ftz.f32 %f13454, %f13450, %f13452; div.approx.ftz.f32 %f16587, %f13454, %f3548; bra.uni $L__BB108_2141; $L__BB108_2137: ld.local.f32 %f13399, [%rd587+52]; mul.ftz.f32 %f13400, %f3486, %f13399; ld.local.f32 %f13401, [%rd587+48]; mul.ftz.f32 %f13402, %f3487, %f13401; sub.ftz.f32 %f13403, %f13400, %f13402; ld.local.f32 %f13404, [%rd587+44]; mul.ftz.f32 %f13405, %f3487, %f13404; mul.ftz.f32 %f13406, %f3484, %f13399; sub.ftz.f32 %f13407, %f13405, %f13406; mul.ftz.f32 %f13408, %f3484, %f13401; mul.ftz.f32 %f13409, %f3486, %f13404; sub.ftz.f32 %f13410, %f13408, %f13409; mul.ftz.f32 %f13411, %f13407, %f13407; fma.rn.ftz.f32 %f13412, %f13403, %f13403, %f13411; fma.rn.ftz.f32 %f13413, %f13410, %f13410, %f13412; sqrt.approx.ftz.f32 %f13414, %f13413; div.approx.ftz.f32 %f13415, %f13403, %f13414; div.approx.ftz.f32 %f13416, %f13407, %f13414; div.approx.ftz.f32 %f13417, %f13410, %f13414; mul.ftz.f32 %f13418, %f3487, %f13416; mul.ftz.f32 %f13419, %f3486, %f13417; sub.ftz.f32 %f13420, %f13418, %f13419; mul.ftz.f32 %f13421, %f3484, %f13417; mul.ftz.f32 %f13422, %f3487, %f13415; sub.ftz.f32 %f13423, %f13421, %f13422; mul.ftz.f32 %f13424, %f3486, %f13415; mul.ftz.f32 %f13425, %f3484, %f13416; sub.ftz.f32 %f13426, %f13424, %f13425; mul.ftz.f32 %f13427, %f3546, %f13423; fma.rn.ftz.f32 %f13428, %f3545, %f13420, %f13427; fma.rn.ftz.f32 %f3551, %f3547, %f13426, %f13428; mul.ftz.f32 %f13429, %f3546, %f13416; fma.rn.ftz.f32 %f13430, %f3545, %f13415, %f13429; fma.rn.ftz.f32 %f3552, %f3547, %f13417, %f13430; setp.geu.ftz.f32 %p1905, %f3549, 0f3F800000; @%p1905 bra $L__BB108_2139; mul.ftz.f32 %f13431, %f3491, %f3551; mul.ftz.f32 %f13432, %f3551, %f13431; mul.ftz.f32 %f13433, %f3492, %f3552; fma.rn.ftz.f32 %f13434, %f3552, %f13433, %f13432; mul.ftz.f32 %f13435, %f3549, %f3549; mov.f32 %f13436, 0f3F800000; sub.ftz.f32 %f13437, %f13436, %f13435; div.approx.ftz.f32 %f13438, %f13434, %f13437; lg2.approx.ftz.f32 %f13439, %f3549; mul.ftz.f32 %f13440, %f13438, %f13439; ex2.approx.ftz.f32 %f16584, %f13440; $L__BB108_2139: add.ftz.f32 %f13441, %f3492, 0f3F800000; mul.ftz.f32 %f13442, %f3511, %f13441; sqrt.approx.ftz.f32 %f13443, %f13442; mov.f32 %f13444, 0f41C90FDB; div.approx.ftz.f32 %f13445, %f13443, %f13444; mul.ftz.f32 %f13446, %f3550, %f13445; mul.ftz.f32 %f16586, %f16584, %f13446; mul.ftz.f32 %f13447, %f16584, %f13445; div.approx.ftz.f32 %f16587, %f13447, %f3548; $L__BB108_2141: max.ftz.f32 %f13456, %f3488, %f13394; max.ftz.f32 %f13457, %f3550, %f13456; mul.ftz.f32 %f13458, %f3548, %f13457; max.ftz.f32 %f13459, %f13394, %f13458; rcp.approx.ftz.f32 %f13460, %f13459; mul.ftz.f32 %f16744, %f13460, %f16586; $L__BB108_2142: mov.u32 %r3285, 10; mov.f32 %f16743, %f16587; mov.f32 %f16745, %f16744; mov.f32 %f16746, %f16744; bra.uni $L__BB108_2309; $L__BB108_911: mad.lo.s32 %r1534, %r1520, 3, %r448; ld.const.u64 %rd1419, [kernel_params+2328]; cvta.to.global.u64 %rd1420, %rd1419; mul.wide.s32 %rd1421, %r1534, 8; add.s64 %rd1422, %rd1420, %rd1421; ld.global.v2.f32 {%f15954, %f15955}, [%rd1422]; ld.global.v2.f32 {%f15952, %f15953}, [%rd1422+8]; ld.global.v2.f32 {%f15950, %f15951}, [%rd1422+16]; $L__BB108_912: ld.local.f32 %f7065, [%rd7+-20]; ld.local.f32 %f7066, [%rd7+-16]; mul.ftz.f32 %f7067, %f15950, %f7066; mul.ftz.f32 %f7068, %f15951, %f7066; fma.rn.ftz.f32 %f7069, %f15952, %f7065, %f7067; fma.rn.ftz.f32 %f7070, %f15953, %f7065, %f7068; mov.f32 %f7071, 0f3F800000; sub.ftz.f32 %f7072, %f7071, %f7065; sub.ftz.f32 %f7073, %f7072, %f7066; fma.rn.ftz.f32 %f15957, %f15955, %f7073, %f7070; fma.rn.ftz.f32 %f15956, %f15954, %f7073, %f7069; $L__BB108_922: mov.f32 %f15960, 0f3F800000; $L__BB108_923: ld.const.s32 %rd1443, [kernel_params+1296]; add.s64 %rd1444, %rd166, %rd1443; cvta.to.global.u64 %rd1445, %rd618; shl.b64 %rd1446, %rd1444, 2; add.s64 %rd1447, %rd1445, %rd1446; add.s64 %rd1448, %rd1447, 4; add.s64 %rd1449, %rd1447, 8; atom.global.add.f32 %f7099, [%rd1447], %f15956; atom.global.add.f32 %f7100, [%rd1448], %f15957; atom.global.add.f32 %f7101, [%rd1449], %f15960; $L__BB108_924: and.b32 %r1553, %r402, 128; setp.eq.s32 %p790, %r1553, 0; @%p790 bra $L__BB108_1145; ld.local.u32 %r454, [%rd7+-24]; and.b32 %r455, %r454, 14; setp.ne.s32 %p791, %r455, 0; @%p791 bra $L__BB108_927; bra.uni $L__BB108_926; $L__BB108_927: and.b32 %r1554, %r454, 6; setp.eq.s32 %p792, %r1554, 0; @%p792 bra $L__BB108_929; ld.const.u64 %rd1452, [kernel_params+2264]; cvta.to.global.u64 %rd1453, %rd1452; ld.local.u32 %r1555, [%rd7+-28]; mul.wide.s32 %rd1454, %r1555, 16; add.s64 %rd1455, %rd1453, %rd1454; shr.s32 %r1556, %r454, 7; ld.global.u32 %r1557, [%rd1455+4]; add.s32 %r1558, %r1556, %r1557; ld.const.u64 %rd1456, [kernel_params+2272]; cvta.to.global.u64 %rd1457, %rd1456; mul.wide.s32 %rd1458, %r1558, 16; add.s64 %rd1459, %rd1457, %rd1458; ld.global.v4.f32 {%f7106, %f7107, %f7108, %f7109}, [%rd1459]; ld.global.v4.f32 {%f7113, %f7114, %f7115, %f7116}, [%rd1459+16]; ld.local.f32 %f7120, [%rd7+-20]; mov.f32 %f7121, 0f3F800000; sub.ftz.f32 %f7122, %f7121, %f7120; mul.ftz.f32 %f7123, %f7106, %f7122; mul.ftz.f32 %f7124, %f7107, %f7122; mul.ftz.f32 %f7125, %f7108, %f7122; fma.rn.ftz.f32 %f15964, %f7113, %f7120, %f7123; fma.rn.ftz.f32 %f15965, %f7114, %f7120, %f7124; fma.rn.ftz.f32 %f15966, %f7115, %f7120, %f7125; bra.uni $L__BB108_931; $L__BB108_926: ld.local.v4.f32 {%f15964, %f15965, %f15966, %f7105}, [%rd1]; bra.uni $L__BB108_935; $L__BB108_929: and.b32 %r1559, %r454, 8; setp.eq.s32 %p793, %r1559, 0; mov.f32 %f15964, 0f00000000; mov.f32 %f15965, %f15964; mov.f32 %f15966, %f15964; @%p793 bra $L__BB108_931; ld.const.u64 %rd1460, [kernel_params+2296]; cvta.to.global.u64 %rd1461, %rd1460; ld.local.u32 %r1560, [%rd7+-28]; mul.wide.s32 %rd1462, %r1560, 16; add.s64 %rd1463, %rd1461, %rd1462; ld.global.v4.f32 {%f15964, %f15965, %f15966, %f7132}, [%rd1463]; $L__BB108_931: ld.local.u32 %r456, [%rd7+-32]; and.b32 %r1561, %r456, 4; setp.ne.s32 %p794, %r1561, 0; @%p794 bra $L__BB108_935; and.b32 %r1562, %r456, 2; setp.eq.s32 %p795, %r1562, 0; @%p795 bra $L__BB108_934; bra.uni $L__BB108_933; $L__BB108_934: ld.const.u64 %rd1464, [kernel_params+2160]; cvta.to.global.u64 %rd1465, %rd1464; ld.local.u32 %r1563, [%rd7+-12]; mul.wide.s32 %rd1466, %r1563, 208; add.s64 %rd1467, %rd1465, %rd1466; ld.global.v4.f32 {%f7166, %f7167, %f7168, %f7169}, [%rd1467]; ld.global.v4.f32 {%f7174, %f7175, %f7176, %f7177}, [%rd1467+16]; ld.global.v4.f32 {%f7182, %f7183, %f7184, %f7185}, [%rd1467+32]; mul.ftz.f32 %f7190, %f15965, %f7167; fma.rn.ftz.f32 %f7191, %f15964, %f7166, %f7190; fma.rn.ftz.f32 %f7192, %f15966, %f7168, %f7191; add.ftz.f32 %f1288, %f7169, %f7192; mul.ftz.f32 %f7193, %f15965, %f7175; fma.rn.ftz.f32 %f7194, %f15964, %f7174, %f7193; fma.rn.ftz.f32 %f7195, %f15966, %f7176, %f7194; add.ftz.f32 %f1289, %f7177, %f7195; mul.ftz.f32 %f7196, %f15965, %f7183; fma.rn.ftz.f32 %f7197, %f15964, %f7182, %f7196; fma.rn.ftz.f32 %f7198, %f15966, %f7184, %f7197; add.ftz.f32 %f15966, %f7185, %f7198; mov.f32 %f15964, %f1288; mov.f32 %f15965, %f1289; bra.uni $L__BB108_935; $L__BB108_933: ld.local.v4.f32 {%f7133, %f7134, %f7135, %f7136}, [%rd7+56]; mul.ftz.f32 %f7141, %f15965, %f7134; fma.rn.ftz.f32 %f7142, %f15964, %f7133, %f7141; fma.rn.ftz.f32 %f7143, %f15966, %f7135, %f7142; add.ftz.f32 %f1285, %f7136, %f7143; ld.local.v4.f32 {%f7144, %f7145, %f7146, %f7147}, [%rd7+72]; mul.ftz.f32 %f7152, %f15965, %f7145; fma.rn.ftz.f32 %f7153, %f15964, %f7144, %f7152; fma.rn.ftz.f32 %f7154, %f15966, %f7146, %f7153; add.ftz.f32 %f1286, %f7147, %f7154; ld.local.v4.f32 {%f7155, %f7156, %f7157, %f7158}, [%rd7+88]; mul.ftz.f32 %f7163, %f15965, %f7156; fma.rn.ftz.f32 %f7164, %f15964, %f7155, %f7163; fma.rn.ftz.f32 %f7165, %f15966, %f7157, %f7164; add.ftz.f32 %f15966, %f7158, %f7165; mov.f32 %f15964, %f1285; mov.f32 %f15965, %f1286; $L__BB108_935: ld.local.u32 %r3200, [%rd7+-12]; cvt.s64.s32 %rd186, %r3200; ld.local.s32 %rd187, [%rd7+-28]; setp.eq.s32 %p796, %r3200, -1; mov.u32 %r478, -1; mov.u32 %r475, 0; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p796 bra $L__BB108_950; ld.const.u64 %rd1468, [kernel_params+2160]; cvta.to.global.u64 %rd1469, %rd1468; mul.lo.s64 %rd1470, %rd186, 208; add.s64 %rd1471, %rd1469, %rd1470; ld.global.u32 %r458, [%rd1471+164]; and.b32 %r1569, %r454, 1; setp.eq.b32 %p797, %r1569, 1; mov.pred %p798, 0; xor.pred %p799, %p797, %p798; not.pred %p800, %p799; mov.u32 %r3190, 0; @%p800 bra $L__BB108_940; cvt.u32.u64 %r1571, %rd187; setp.eq.s32 %p801, %r1571, -1; mov.u32 %r3189, -1; @%p801 bra $L__BB108_939; ld.const.u64 %rd1472, [kernel_params+2240]; cvta.to.global.u64 %rd1473, %rd1472; shl.b64 %rd1474, %rd187, 2; add.s64 %rd1475, %rd1473, %rd1474; ld.global.u32 %r3189, [%rd1475]; $L__BB108_939: setp.ne.s32 %p802, %r3189, -1; selp.u32 %r3190, 1, 0, %p802; $L__BB108_940: ld.const.u64 %rd1476, [kernel_params+2312]; cvta.to.global.u64 %rd188, %rd1476; add.s32 %r3192, %r3190, %r458; mul.wide.u32 %rd1477, %r3192, 16; add.s64 %rd1478, %rd188, %rd1477; add.s64 %rd2961, %rd1478, 14; ld.global.u16 %rs143, [%rd1478+12]; ld.global.u32 %r3191, [%rd1478+8]; ld.global.u64 %rd2959, [%rd1478]; setp.eq.s64 %p803, %rd2959, 11; add.s64 %rd2960, %rd1478, 15; @%p803 bra $L__BB108_946; $L__BB108_941: setp.eq.s64 %p804, %rd2959, 0; @%p804 bra $L__BB108_943; add.s32 %r3192, %r3192, 2; bra.uni $L__BB108_944; $L__BB108_943: setp.eq.s16 %p805, %rs143, 0; mov.u32 %r475, 0; mov.u32 %r3192, %r3191; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p805 bra $L__BB108_950; $L__BB108_944: cvt.u64.u32 %rd193, %r3192; mul.wide.u32 %rd1479, %r3192, 16; add.s64 %rd1480, %rd188, %rd1479; ld.global.u16 %rs143, [%rd1480+12]; ld.global.u32 %r3191, [%rd1480+8]; ld.global.u64 %rd2959, [%rd1480]; setp.ne.s64 %p806, %rd2959, 11; @%p806 bra $L__BB108_941; shl.b64 %rd1481, %rd193, 4; add.s64 %rd1482, %rd188, %rd1481; add.s64 %rd2961, %rd1482, 14; add.s64 %rd2960, %rd1482, 15; $L__BB108_946: ld.global.u8 %rs26, [%rd2961]; ld.global.u8 %rs27, [%rd2960]; cvt.u32.u64 %r1576, %rd187; setp.ne.s32 %p807, %r1576, -1; @%p807 bra $L__BB108_949; mov.u32 %r475, 0; add.s16 %rs79, %rs143, -1; setp.lt.u16 %p808, %rs79, 2; @%p808 bra $L__BB108_949; setp.ne.s16 %p809, %rs143, 1024; mov.u32 %r3196, %r475; mov.u32 %r3197, %r475; @%p809 bra $L__BB108_950; $L__BB108_949: setp.eq.s16 %p810, %rs143, 0; selp.b32 %r478, -1, %r3191, %p810; cvt.u32.u16 %r475, %rs143; cvt.u32.u16 %r1581, %rs27; and.b32 %r3197, %r1581, 255; cvt.u32.u16 %r1582, %rs26; and.b32 %r3196, %r1582, 255; $L__BB108_950: setp.eq.s32 %p811, %r478, -1; mov.f32 %f15982, %f15964; mov.f32 %f15983, %f15965; mov.f32 %f15984, %f15966; mov.f32 %f15985, %f15964; mov.f32 %f15986, %f15965; mov.f32 %f15987, %f15966; @%p811 bra $L__BB108_986; ld.const.u64 %rd1483, [kernel_params+2160]; cvta.to.global.u64 %rd1484, %rd1483; mul.lo.s64 %rd1485, %rd186, 208; add.s64 %rd1486, %rd1484, %rd1485; add.s64 %rd199, %rd1486, 148; ld.global.u32 %r481, [%rd1486+156]; @%p791 bra $L__BB108_971; bra.uni $L__BB108_952; $L__BB108_971: ld.global.u32 %r487, [%rd199]; and.b32 %r488, %r475, 768; setp.eq.s32 %p827, %r488, 0; @%p827 bra $L__BB108_973; ld.const.u64 %rd1525, [kernel_params+2264]; cvta.to.global.u64 %rd1526, %rd1525; shl.b64 %rd1527, %rd187, 4; add.s64 %rd1528, %rd1526, %rd1527; shr.s32 %r1621, %r454, 7; ld.global.u32 %r1622, [%rd1528+4]; add.s32 %r1623, %r1621, %r1622; ld.const.u64 %rd1529, [kernel_params+2344]; cvta.to.global.u64 %rd1530, %rd1529; add.s32 %r1624, %r1623, %r478; mul.wide.s32 %rd1531, %r1624, 16; add.s64 %rd1532, %rd1530, %rd1531; ld.global.v4.f32 {%f7245, %f7246, %f7247, %f7248}, [%rd1532]; ld.global.v4.f32 {%f7253, %f7254, %f7255, %f7256}, [%rd1532+16]; ld.local.f32 %f7261, [%rd7+-20]; mov.f32 %f7262, 0f3F800000; sub.ftz.f32 %f7263, %f7262, %f7261; mul.ftz.f32 %f7264, %f7245, %f7263; mul.ftz.f32 %f7265, %f7246, %f7263; mul.ftz.f32 %f7266, %f7247, %f7263; fma.rn.ftz.f32 %f15987, %f7255, %f7261, %f7266; fma.rn.ftz.f32 %f15986, %f7254, %f7261, %f7265; fma.rn.ftz.f32 %f15985, %f7253, %f7261, %f7264; bra.uni $L__BB108_975; $L__BB108_952: and.b32 %r1583, %r454, 1; setp.eq.b32 %p813, %r1583, 1; mov.pred %p814, 0; xor.pred %p815, %p813, %p814; not.pred %p816, %p815; mov.f32 %f15982, %f15964; mov.f32 %f15983, %f15965; mov.f32 %f15984, %f15966; mov.f32 %f15985, %f15964; mov.f32 %f15986, %f15965; mov.f32 %f15987, %f15966; @%p816 bra $L__BB108_986; cvt.u32.u64 %r1585, %rd187; setp.eq.s32 %p817, %r1585, -1; mov.u32 %r3199, -1; @%p817 bra $L__BB108_955; ld.const.u64 %rd1487, [kernel_params+2240]; cvta.to.global.u64 %rd1488, %rd1487; shl.b64 %rd1489, %rd187, 2; add.s64 %rd1490, %rd1488, %rd1489; ld.global.u32 %r3199, [%rd1490]; $L__BB108_955: setp.eq.s32 %p818, %r3199, -1; @%p818 bra $L__BB108_957; mov.u64 %rd1491, 0; { // callseq 777, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r3196; st.param.b32 [param1+8], %r3197; st.param.b32 [param1+12], %r478; .param .b64 param2; st.param.b64 [param2+0], %rd1491; .param .b64 param3; st.param.b64 [param3+0], %rd1491; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15985, [retval0+0]; ld.param.f32 %f15986, [retval0+4]; ld.param.f32 %f15987, [retval0+8]; } // callseq 777 add.s32 %r1586, %r481, %r478; { // callseq 778, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd619; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r475; st.param.b32 [param1+4], %r3196; st.param.b32 [param1+8], %r3197; st.param.b32 [param1+12], %r1586; .param .b64 param2; st.param.b64 [param2+0], %rd1491; .param .b64 param3; st.param.b64 [param3+0], %rd1491; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f15982, [retval0+0]; ld.param.f32 %f15983, [retval0+4]; ld.param.f32 %f15984, [retval0+8]; } // callseq 778 ld.local.u32 %r3200, [%rd7+-12]; bra.uni $L__BB108_986; $L__BB108_973: and.b32 %r1625, %r475, 131; setp.eq.s32 %p828, %r1625, 0; mov.f32 %f15985, 0f00000000; mov.f32 %f15986, %f15985; mov.f32 %f15987, %f15985; @%p828 bra $L__BB108_975; cvt.u32.u64 %r1626, %rd187; setp.eq.s32 %p829, %r475, 128; selp.b32 %r1627, %r1626, 0, %p829; add.s32 %r1628, %r1627, %r478; ld.const.u64 %rd1533, [kernel_params+2344]; cvta.to.global.u64 %rd1534, %rd1533; mul.wide.s32 %rd1535, %r1628, 16; add.s64 %rd1536, %rd1534, %rd1535; ld.global.v4.f32 {%f15985, %f15986, %f15987, %f7275}, [%rd1536]; $L__BB108_975: add.s32 %r489, %r487, %r478; @%p827 bra $L__BB108_977; cvt.u32.u64 %r1629, %rd187; ld.const.u64 %rd1537, [kernel_params+2264]; cvta.to.global.u64 %rd1538, %rd1537; mul.wide.s32 %rd1539, %r1629, 16; add.s64 %rd1540, %rd1538, %rd1539; shr.s32 %r1630, %r454, 7; ld.global.u32 %r1631, [%rd1540+4]; add.s32 %r1632, %r1630, %r1631; ld.const.u64 %rd1541, [kernel_params+2344]; cvta.to.global.u64 %rd1542, %rd1541; add.s32 %r1633, %r1632, %r489; mul.wide.s32 %rd1543, %r1633, 16; add.s64 %rd1544, %rd1542, %rd1543; ld.global.v4.f32 {%f7276, %f7277, %f7278, %f7279}, [%rd1544]; ld.global.v4.f32 {%f7284, %f7285, %f7286, %f7287}, [%rd1544+16]; ld.local.f32 %f7292, [%rd7+-20]; mov.f32 %f7293, 0f3F800000; sub.ftz.f32 %f7294, %f7293, %f7292; mul.ftz.f32 %f7295, %f7276, %f7294; mul.ftz.f32 %f7296, %f7277, %f7294; mul.ftz.f32 %f7297, %f7278, %f7294; fma.rn.ftz.f32 %f15984, %f7286, %f7292, %f7297; fma.rn.ftz.f32 %f15983, %f7285, %f7292, %f7296; fma.rn.ftz.f32 %f15982, %f7284, %f7292, %f7295; bra.uni $L__BB108_979; $L__BB108_977: and.b32 %r1634, %r475, 131; setp.eq.s32 %p831, %r1634, 0; mov.f32 %f15982, 0f00000000; mov.f32 %f15983, %f15982; mov.f32 %f15984, %f15982; @%p831 bra $L__BB108_979; cvt.u32.u64 %r1635, %rd187; setp.eq.s32 %p832, %r475, 128; selp.b32 %r1636, %r1635, 0, %p832; add.s32 %r1637, %r1636, %r489; ld.const.u64 %rd1545, [kernel_params+2344]; cvta.to.global.u64 %rd1546, %rd1545; mul.wide.s32 %rd1547, %r1637, 16; add.s64 %rd1548, %rd1546, %rd1547; ld.global.v4.f32 {%f15982, %f15983, %f15984, %f7306}, [%rd1548]; $L__BB108_979: ld.local.u32 %r490, [%rd7+-32]; and.b32 %r1638, %r490, 64; setp.ne.s32 %p833, %r1638, 0; @%p833 bra $L__BB108_986; and.b32 %r491, %r490, 2; setp.eq.s32 %p834, %r491, 0; @%p834 bra $L__BB108_982; bra.uni $L__BB108_981; $L__BB108_982: ld.global.v4.f32 {%f7340, %f7341, %f7342, %f7343}, [%rd199+-148]; ld.global.v4.f32 {%f7348, %f7349, %f7350, %f7351}, [%rd199+-132]; ld.global.v4.f32 {%f7356, %f7357, %f7358, %f7359}, [%rd199+-116]; mul.ftz.f32 %f7364, %f15986, %f7341; fma.rn.ftz.f32 %f7365, %f15985, %f7340, %f7364; fma.rn.ftz.f32 %f7366, %f15987, %f7342, %f7365; add.ftz.f32 %f15979, %f7343, %f7366; mul.ftz.f32 %f7367, %f15986, %f7349; fma.rn.ftz.f32 %f7368, %f15985, %f7348, %f7367; fma.rn.ftz.f32 %f7369, %f15987, %f7350, %f7368; add.ftz.f32 %f15980, %f7351, %f7369; mul.ftz.f32 %f7370, %f15986, %f7357; fma.rn.ftz.f32 %f7371, %f15985, %f7356, %f7370; fma.rn.ftz.f32 %f7372, %f15987, %f7358, %f7371; add.ftz.f32 %f15987, %f7359, %f7372; bra.uni $L__BB108_983; $L__BB108_981: ld.local.v4.f32 {%f7307, %f7308, %f7309, %f7310}, [%rd7+56]; mul.ftz.f32 %f7315, %f15986, %f7308; fma.rn.ftz.f32 %f7316, %f15985, %f7307, %f7315; fma.rn.ftz.f32 %f7317, %f15987, %f7309, %f7316; add.ftz.f32 %f15979, %f7310, %f7317; ld.local.v4.f32 {%f7318, %f7319, %f7320, %f7321}, [%rd7+72]; mul.ftz.f32 %f7326, %f15986, %f7319; fma.rn.ftz.f32 %f7327, %f15985, %f7318, %f7326; fma.rn.ftz.f32 %f7328, %f15987, %f7320, %f7327; add.ftz.f32 %f15980, %f7321, %f7328; ld.local.v4.f32 {%f7329, %f7330, %f7331, %f7332}, [%rd7+88]; mul.ftz.f32 %f7337, %f15986, %f7330; fma.rn.ftz.f32 %f7338, %f15985, %f7329, %f7337; fma.rn.ftz.f32 %f7339, %f15987, %f7331, %f7338; add.ftz.f32 %f15987, %f7332, %f7339; $L__BB108_983: @%p834 bra $L__BB108_985; bra.uni $L__BB108_984; $L__BB108_985: ld.global.v4.f32 {%f7406, %f7407, %f7408, %f7409}, [%rd199+-148]; ld.global.v4.f32 {%f7414, %f7415, %f7416, %f7417}, [%rd199+-132]; ld.global.v4.f32 {%f7422, %f7423, %f7424, %f7425}, [%rd199+-116]; mul.ftz.f32 %f7430, %f15983, %f7407; fma.rn.ftz.f32 %f7431, %f15982, %f7406, %f7430; fma.rn.ftz.f32 %f7432, %f15984, %f7408, %f7431; add.ftz.f32 %f1366, %f7409, %f7432; mul.ftz.f32 %f7433, %f15983, %f7415; fma.rn.ftz.f32 %f7434, %f15982, %f7414, %f7433; fma.rn.ftz.f32 %f7435, %f15984, %f7416, %f7434; add.ftz.f32 %f1367, %f7417, %f7435; mul.ftz.f32 %f7436, %f15983, %f7423; fma.rn.ftz.f32 %f7437, %f15982, %f7422, %f7436; fma.rn.ftz.f32 %f7438, %f15984, %f7424, %f7437; add.ftz.f32 %f15984, %f7425, %f7438; mov.f32 %f15982, %f1366; mov.f32 %f15983, %f1367; mov.f32 %f15985, %f15979; mov.f32 %f15986, %f15980; bra.uni $L__BB108_986; $L__BB108_984: ld.local.v4.f32 {%f7373, %f7374, %f7375, %f7376}, [%rd7+56]; mul.ftz.f32 %f7381, %f15983, %f7374; fma.rn.ftz.f32 %f7382, %f15982, %f7373, %f7381; fma.rn.ftz.f32 %f7383, %f15984, %f7375, %f7382; add.ftz.f32 %f1363, %f7376, %f7383; ld.local.v4.f32 {%f7384, %f7385, %f7386, %f7387}, [%rd7+72]; mul.ftz.f32 %f7392, %f15983, %f7385; fma.rn.ftz.f32 %f7393, %f15982, %f7384, %f7392; fma.rn.ftz.f32 %f7394, %f15984, %f7386, %f7393; add.ftz.f32 %f1364, %f7387, %f7394; ld.local.v4.f32 {%f7395, %f7396, %f7397, %f7398}, [%rd7+88]; mul.ftz.f32 %f7403, %f15983, %f7396; fma.rn.ftz.f32 %f7404, %f15982, %f7395, %f7403; fma.rn.ftz.f32 %f7405, %f15984, %f7397, %f7404; add.ftz.f32 %f15984, %f7398, %f7405; mov.f32 %f15982, %f1363; mov.f32 %f15983, %f1364; mov.f32 %f15985, %f15979; mov.f32 %f15986, %f15980; bra.uni $L__BB108_986; $L__BB108_957: and.b32 %r485, %r475, 56; setp.eq.s32 %p819, %r485, 0; @%p819 bra $L__BB108_962; and.b32 %r1587, %r475, 24; setp.eq.s32 %p820, %r1587, 0; @%p820 bra $L__BB108_960; ld.const.u64 %rd1493, [kernel_params+2232]; cvta.to.global.u64 %rd1494, %rd1493; shl.b64 %rd1495, %rd187, 4; add.s64 %rd1496, %rd1494, %rd1495; ld.global.v4.u32 {%r1588, %r1589, %r1590, %r1591}, [%rd1496]; ld.const.u64 %rd1497, [kernel_params+2336]; cvta.to.global.u64 %rd1498, %rd1497; add.s32 %r1595, %r1588, %r478; mul.wide.u32 %rd1499, %r1595, 12; add.s64 %rd2968, %rd1498, %rd1499; add.s32 %r1596, %r1589, %r478; mul.wide.u32 %rd1500, %r1596, 12; add.s64 %rd2965, %rd1498, %rd1500; add.s32 %r1597, %r1590, %r478; mul.wide.u32 %rd1501, %r1597, 12; add.s64 %rd2962, %rd1498, %rd1501; add.s64 %rd2969, %rd2968, 4; add.s64 %rd2970, %rd2968, 8; add.s64 %rd2966, %rd2965, 4; add.s64 %rd2967, %rd2965, 8; add.s64 %rd2963, %rd2962, 4; add.s64 %rd2964, %rd2962, 8; bra.uni $L__BB108_961; $L__BB108_962: and.b32 %r1600, %r475, 7; setp.eq.s32 %p821, %r1600, 0; mov.f32 %f15985, 0f00000000; mov.f32 %f15986, %f15985; mov.f32 %f15987, %f15985; @%p821 bra $L__BB108_964; setp.eq.s32 %p822, %r475, 4; selp.b32 %r1602, %r1585, 0, %p822; add.s32 %r1603, %r1602, %r478; ld.const.u64 %rd1505, [kernel_params+2336]; cvta.to.global.u64 %rd1506, %rd1505; mul.wide.s32 %rd1507, %r1603, 12; add.s64 %rd1508, %rd1506, %rd1507; ld.global.f32 %f15987, [%rd1508+8]; ld.global.f32 %f15986, [%rd1508+4]; ld.global.f32 %f15985, [%rd1508]; bra.uni $L__BB108_964; $L__BB108_960: mad.lo.s32 %r1599, %r1585, 3, %r478; ld.const.u64 %rd1502, [kernel_params+2336]; cvta.to.global.u64 %rd1503, %rd1502; mul.wide.s32 %rd1504, %r1599, 12; add.s64 %rd2968, %rd1503, %rd1504; add.s64 %rd2969, %rd2968, 4; add.s64 %rd2970, %rd2968, 8; add.s64 %rd2965, %rd2968, 12; add.s64 %rd2966, %rd2968, 16; add.s64 %rd2967, %rd2968, 20; add.s64 %rd2962, %rd2968, 24; add.s64 %rd2963, %rd2968, 28; add.s64 %rd2964, %rd2968, 32; $L__BB108_961: ld.local.f32 %f7199, [%rd7+-20]; ld.global.f32 %f7200, [%rd2965]; ld.global.f32 %f7201, [%rd2966]; ld.global.f32 %f7202, [%rd2967]; ld.local.f32 %f7203, [%rd7+-16]; ld.global.f32 %f7204, [%rd2962]; mul.ftz.f32 %f7205, %f7204, %f7203; ld.global.f32 %f7206, [%rd2963]; mul.ftz.f32 %f7207, %f7206, %f7203; ld.global.f32 %f7208, [%rd2964]; mul.ftz.f32 %f7209, %f7208, %f7203; fma.rn.ftz.f32 %f7210, %f7200, %f7199, %f7205; fma.rn.ftz.f32 %f7211, %f7201, %f7199, %f7207; fma.rn.ftz.f32 %f7212, %f7202, %f7199, %f7209; mov.f32 %f7213, 0f3F800000; sub.ftz.f32 %f7214, %f7213, %f7199; sub.ftz.f32 %f7215, %f7214, %f7203; ld.global.f32 %f7216, [%rd2968]; ld.global.f32 %f7217, [%rd2969]; ld.global.f32 %f7218, [%rd2970]; fma.rn.ftz.f32 %f15987, %f7218, %f7215, %f7212; fma.rn.ftz.f32 %f15986, %f7217, %f7215, %f7211; fma.rn.ftz.f32 %f15985, %f7216, %f7215, %f7210; $L__BB108_964: add.s32 %r486, %r481, %r478; @%p819 bra $L__BB108_969; and.b32 %r1604, %r475, 24; setp.eq.s32 %p824, %r1604, 0; @%p824 bra $L__BB108_967; ld.const.u64 %rd1509, [kernel_params+2232]; cvta.to.global.u64 %rd1510, %rd1509; shl.b64 %rd1511, %rd187, 4; add.s64 %rd1512, %rd1510, %rd1511; ld.global.v4.u32 {%r1605, %r1606, %r1607, %r1608}, [%rd1512]; ld.const.u64 %rd1513, [kernel_params+2336]; cvta.to.global.u64 %rd1514, %rd1513; add.s32 %r1612, %r1605, %r486; mul.wide.u32 %rd1515, %r1612, 12; add.s64 %rd2977, %rd1514, %rd1515; add.s32 %r1613, %r1606, %r486; mul.wide.u32 %rd1516, %r1613, 12; add.s64 %rd2974, %rd1514, %rd1516; add.s32 %r1614, %r1607, %r486; mul.wide.u32 %rd1517, %r1614, 12; add.s64 %rd2971, %rd1514, %rd1517; add.s64 %rd2978, %rd2977, 4; add.s64 %rd2979, %rd2977, 8; add.s64 %rd2975, %rd2974, 4; add.s64 %rd2976, %rd2974, 8; add.s64 %rd2972, %rd2971, 4; add.s64 %rd2973, %rd2971, 8; bra.uni $L__BB108_968; $L__BB108_969: and.b32 %r1617, %r475, 7; setp.eq.s32 %p825, %r1617, 0; mov.f32 %f15982, 0f00000000; mov.f32 %f15983, %f15982; mov.f32 %f15984, %f15982; @%p825 bra $L__BB108_986; setp.eq.s32 %p826, %r475, 4; selp.b32 %r1619, %r1585, 0, %p826; add.s32 %r1620, %r1619, %r486; ld.const.u64 %rd1521, [kernel_params+2336]; cvta.to.global.u64 %rd1522, %rd1521; mul.wide.s32 %rd1523, %r1620, 12; add.s64 %rd1524, %rd1522, %rd1523; ld.global.f32 %f15984, [%rd1524+8]; ld.global.f32 %f15983, [%rd1524+4]; ld.global.f32 %f15982, [%rd1524]; bra.uni $L__BB108_986; $L__BB108_967: mad.lo.s32 %r1616, %r1585, 3, %r486; ld.const.u64 %rd1518, [kernel_params+2336]; cvta.to.global.u64 %rd1519, %rd1518; mul.wide.s32 %rd1520, %r1616, 12; add.s64 %rd2977, %rd1519, %rd1520; add.s64 %rd2978, %rd2977, 4; add.s64 %rd2979, %rd2977, 8; add.s64 %rd2974, %rd2977, 12; add.s64 %rd2975, %rd2977, 16; add.s64 %rd2976, %rd2977, 20; add.s64 %rd2971, %rd2977, 24; add.s64 %rd2972, %rd2977, 28; add.s64 %rd2973, %rd2977, 32; $L__BB108_968: ld.local.f32 %f7222, [%rd7+-20]; ld.global.f32 %f7223, [%rd2974]; ld.global.f32 %f7224, [%rd2975]; ld.global.f32 %f7225, [%rd2976]; ld.local.f32 %f7226, [%rd7+-16]; ld.global.f32 %f7227, [%rd2971]; mul.ftz.f32 %f7228, %f7227, %f7226; ld.global.f32 %f7229, [%rd2972]; mul.ftz.f32 %f7230, %f7229, %f7226; ld.global.f32 %f7231, [%rd2973]; mul.ftz.f32 %f7232, %f7231, %f7226; fma.rn.ftz.f32 %f7233, %f7223, %f7222, %f7228; fma.rn.ftz.f32 %f7234, %f7224, %f7222, %f7230; fma.rn.ftz.f32 %f7235, %f7225, %f7222, %f7232; mov.f32 %f7236, 0f3F800000; sub.ftz.f32 %f7237, %f7236, %f7222; sub.ftz.f32 %f7238, %f7237, %f7226; ld.global.f32 %f7239, [%rd2977]; ld.global.f32 %f7240, [%rd2978]; ld.global.f32 %f7241, [%rd2979]; fma.rn.ftz.f32 %f15984, %f7241, %f7238, %f7235; fma.rn.ftz.f32 %f15983, %f7240, %f7238, %f7234; fma.rn.ftz.f32 %f15982, %f7239, %f7238, %f7233; $L__BB108_986: ld.const.u64 %rd1549, [kernel_params+2168]; cvta.to.global.u64 %rd1550, %rd1549; shl.b32 %r1639, %r3200, 1; mul.wide.s32 %rd1551, %r1639, 48; add.s64 %rd1552, %rd1550, %rd1551; ld.global.v4.f32 {%f7439, %f7440, %f7441, %f7442}, [%rd1552]; ld.global.v4.f32 {%f7447, %f7448, %f7449, %f7450}, [%rd1552+16]; ld.global.v4.f32 {%f7455, %f7456, %f7457, %f7458}, [%rd1552+32]; mul.ftz.f32 %f7463, %f15986, %f7440; fma.rn.ftz.f32 %f7464, %f15985, %f7439, %f7463; fma.rn.ftz.f32 %f7465, %f15987, %f7441, %f7464; add.ftz.f32 %f1375, %f7442, %f7465; mul.ftz.f32 %f7466, %f15986, %f7448; fma.rn.ftz.f32 %f7467, %f15985, %f7447, %f7466; fma.rn.ftz.f32 %f7468, %f15987, %f7449, %f7467; add.ftz.f32 %f1376, %f7450, %f7468; mul.ftz.f32 %f7469, %f15986, %f7456; fma.rn.ftz.f32 %f7470, %f15985, %f7455, %f7469; fma.rn.ftz.f32 %f7471, %f15987, %f7457, %f7470; add.ftz.f32 %f1377, %f7458, %f7471; ld.global.v4.f32 {%f7472, %f7473, %f7474, %f7475}, [%rd1552+48]; ld.global.v4.f32 {%f7480, %f7481, %f7482, %f7483}, [%rd1552+64]; ld.global.v4.f32 {%f7488, %f7489, %f7490, %f7491}, [%rd1552+80]; mul.ftz.f32 %f7496, %f15983, %f7473; fma.rn.ftz.f32 %f7497, %f15982, %f7472, %f7496; fma.rn.ftz.f32 %f7498, %f15984, %f7474, %f7497; add.ftz.f32 %f1378, %f7475, %f7498; mul.ftz.f32 %f7499, %f15983, %f7481; fma.rn.ftz.f32 %f7500, %f15982, %f7480, %f7499; fma.rn.ftz.f32 %f7501, %f15984, %f7482, %f7500; add.ftz.f32 %f1379, %f7483, %f7501; mul.ftz.f32 %f7502, %f15983, %f7489; fma.rn.ftz.f32 %f7503, %f15982, %f7488, %f7502; fma.rn.ftz.f32 %f7504, %f15984, %f7490, %f7503; add.ftz.f32 %f1380, %f7491, %f7504; ld.const.u32 %r1640, [kernel_params+16]; setp.eq.s32 %p836, %r1640, 2; @%p836 bra $L__BB108_993; bra.uni $L__BB108_987; $L__BB108_993: ld.const.v4.f32 {%f7646, %f7647, %f7648, %f7649}, [kernel_params+688]; ld.const.v4.f32 {%f7654, %f7655, %f7656, %f7657}, [kernel_params+704]; ld.const.v4.f32 {%f7662, %f7663, %f7664, %f7665}, [kernel_params+720]; mul.ftz.f32 %f7670, %f15965, %f7647; fma.rn.ftz.f32 %f7671, %f15964, %f7646, %f7670; fma.rn.ftz.f32 %f7672, %f15966, %f7648, %f7671; add.ftz.f32 %f7673, %f7649, %f7672; mul.ftz.f32 %f7674, %f15965, %f7655; fma.rn.ftz.f32 %f7675, %f15964, %f7654, %f7674; fma.rn.ftz.f32 %f7676, %f15966, %f7656, %f7675; add.ftz.f32 %f7677, %f7657, %f7676; mul.ftz.f32 %f7678, %f15965, %f7663; fma.rn.ftz.f32 %f7679, %f15964, %f7662, %f7678; fma.rn.ftz.f32 %f7680, %f15966, %f7664, %f7679; add.ftz.f32 %f7681, %f7665, %f7680; mul.ftz.f32 %f7682, %f7677, %f7677; fma.rn.ftz.f32 %f7683, %f7673, %f7673, %f7682; fma.rn.ftz.f32 %f7684, %f7681, %f7681, %f7683; sqrt.approx.ftz.f32 %f7685, %f7684; div.approx.ftz.f32 %f1417, %f7673, %f7685; div.approx.ftz.f32 %f1418, %f7677, %f7685; div.approx.ftz.f32 %f1419, %f7681, %f7685; ld.const.u32 %r493, [kernel_params+20]; mov.f32 %f7686, 0fBF800000; max.ftz.f32 %f7687, %f1417, %f7686; mov.f32 %f7688, 0f3F800000; min.ftz.f32 %f7689, %f7687, %f7688; abs.ftz.f32 %f7690, %f7689; sub.ftz.f32 %f7691, %f7688, %f7690; mul.ftz.f32 %f7692, %f7691, 0f3F000000; sqrt.approx.ftz.f32 %f7693, %f7692; setp.gt.ftz.f32 %p840, %f7690, 0f3F11EB85; selp.f32 %f7694, %f7693, %f7690, %p840; mul.ftz.f32 %f7695, %f7694, %f7694; mov.f32 %f7696, 0f3C94D2E9; mov.f32 %f7697, 0f3D53F941; fma.rn.ftz.f32 %f7698, %f7697, %f7695, %f7696; mov.f32 %f7699, 0f3D3F841F; fma.rn.ftz.f32 %f7700, %f7698, %f7695, %f7699; mov.f32 %f7701, 0f3D994929; fma.rn.ftz.f32 %f7702, %f7700, %f7695, %f7701; mov.f32 %f7703, 0f3E2AAB94; fma.rn.ftz.f32 %f7704, %f7702, %f7695, %f7703; mul.ftz.f32 %f7705, %f7695, %f7704; fma.rn.ftz.f32 %f7706, %f7705, %f7694, %f7694; add.ftz.f32 %f7707, %f7706, %f7706; mov.f32 %f7708, 0f3FC90FDB; sub.ftz.f32 %f7709, %f7708, %f7706; selp.f32 %f7710, %f7707, %f7709, %p840; setp.lt.ftz.f32 %p841, %f7689, 0f00000000; mov.f32 %f7711, 0f40490FDB; sub.ftz.f32 %f7712, %f7711, %f7710; selp.f32 %f1420, %f7712, %f7710, %p841; abs.ftz.f32 %f1421, %f1418; abs.ftz.f32 %f1422, %f1419; setp.eq.ftz.f32 %p842, %f1421, 0f00000000; setp.eq.ftz.f32 %p843, %f1422, 0f00000000; and.pred %p2, %p842, %p843; mov.b32 %r494, %f1418; mov.b32 %r1641, %f1419; and.b32 %r495, %r1641, -2147483648; setp.gt.s32 %p844, %r493, 2; @%p844 bra $L__BB108_999; bra.uni $L__BB108_994; $L__BB108_999: setp.eq.s32 %p845, %r493, 3; @%p845 bra $L__BB108_1008; setp.eq.s32 %p846, %r493, 5; @%p846 bra $L__BB108_1011; setp.ne.s32 %p847, %r493, 4; @%p847 bra $L__BB108_1037; ld.const.f32 %f7713, [kernel_params+48]; ld.const.v4.f32 {%f7714, %f7715, %f7716, %f7717}, [kernel_params+64]; neg.ftz.f32 %f7722, %f1420; sub.ftz.f32 %f7723, %f7722, %f7713; div.approx.ftz.f32 %f7724, %f7723, %f7714; mul.ftz.f32 %f7725, %f7724, %f7724; mul.ftz.f32 %f7726, %f7724, %f7725; mul.ftz.f32 %f7727, %f7725, %f7725; mul.ftz.f32 %f7728, %f7715, %f7725; fma.rn.ftz.f32 %f7729, %f7714, %f7724, %f7728; mul.ftz.f32 %f7730, %f7717, %f7727; fma.rn.ftz.f32 %f7731, %f7716, %f7726, %f7730; add.ftz.f32 %f7732, %f7729, %f7731; add.ftz.f32 %f7733, %f7713, %f7732; sub.ftz.f32 %f7734, %f7722, %f7733; div.approx.ftz.f32 %f7735, %f7734, %f7714; mul.ftz.f32 %f7736, %f7735, %f7735; mul.ftz.f32 %f7737, %f7735, %f7736; mul.ftz.f32 %f7738, %f7736, %f7736; mul.ftz.f32 %f7739, %f7715, %f7736; fma.rn.ftz.f32 %f7740, %f7714, %f7735, %f7739; mul.ftz.f32 %f7741, %f7717, %f7738; fma.rn.ftz.f32 %f7742, %f7716, %f7737, %f7741; add.ftz.f32 %f7743, %f7740, %f7742; add.ftz.f32 %f7744, %f7713, %f7743; sub.ftz.f32 %f7745, %f7722, %f7744; div.approx.ftz.f32 %f7746, %f7745, %f7714; mul.ftz.f32 %f7747, %f7746, %f7746; mul.ftz.f32 %f7748, %f7746, %f7747; mul.ftz.f32 %f7749, %f7747, %f7747; mul.ftz.f32 %f7750, %f7715, %f7747; fma.rn.ftz.f32 %f7751, %f7714, %f7746, %f7750; mul.ftz.f32 %f7752, %f7717, %f7749; fma.rn.ftz.f32 %f7753, %f7716, %f7748, %f7752; add.ftz.f32 %f7754, %f7751, %f7753; add.ftz.f32 %f7755, %f7713, %f7754; sub.ftz.f32 %f7756, %f7722, %f7755; div.approx.ftz.f32 %f7757, %f7756, %f7714; mul.ftz.f32 %f7758, %f7757, %f7757; mul.ftz.f32 %f7759, %f7757, %f7758; mul.ftz.f32 %f7760, %f7758, %f7758; mul.ftz.f32 %f7761, %f7715, %f7758; fma.rn.ftz.f32 %f7762, %f7714, %f7757, %f7761; mul.ftz.f32 %f7763, %f7717, %f7760; fma.rn.ftz.f32 %f7764, %f7716, %f7759, %f7763; add.ftz.f32 %f7765, %f7762, %f7764; add.ftz.f32 %f7766, %f7713, %f7765; sub.ftz.f32 %f7767, %f7722, %f7766; div.approx.ftz.f32 %f7768, %f7767, %f7714; mul.ftz.f32 %f7769, %f7768, %f7768; mul.ftz.f32 %f7770, %f7768, %f7769; mul.ftz.f32 %f7771, %f7769, %f7769; mul.ftz.f32 %f7772, %f7715, %f7769; fma.rn.ftz.f32 %f7773, %f7714, %f7768, %f7772; mul.ftz.f32 %f7774, %f7717, %f7771; fma.rn.ftz.f32 %f7775, %f7716, %f7770, %f7774; add.ftz.f32 %f7776, %f7773, %f7775; add.ftz.f32 %f7777, %f7713, %f7776; sub.ftz.f32 %f7778, %f7722, %f7777; div.approx.ftz.f32 %f7779, %f7778, %f7714; mul.ftz.f32 %f7780, %f7779, %f7779; mul.ftz.f32 %f7781, %f7779, %f7780; mul.ftz.f32 %f7782, %f7780, %f7780; mul.ftz.f32 %f7783, %f7715, %f7780; fma.rn.ftz.f32 %f7784, %f7714, %f7779, %f7783; mul.ftz.f32 %f7785, %f7717, %f7782; fma.rn.ftz.f32 %f7786, %f7716, %f7781, %f7785; add.ftz.f32 %f7787, %f7784, %f7786; add.ftz.f32 %f7788, %f7713, %f7787; sub.ftz.f32 %f7789, %f7722, %f7788; div.approx.ftz.f32 %f7790, %f7789, %f7714; mul.ftz.f32 %f7791, %f7790, %f7790; mul.ftz.f32 %f7792, %f7790, %f7791; mul.ftz.f32 %f7793, %f7791, %f7791; mul.ftz.f32 %f7794, %f7715, %f7791; fma.rn.ftz.f32 %f7795, %f7714, %f7790, %f7794; mul.ftz.f32 %f7796, %f7717, %f7793; fma.rn.ftz.f32 %f7797, %f7716, %f7792, %f7796; add.ftz.f32 %f7798, %f7795, %f7797; add.ftz.f32 %f7799, %f7713, %f7798; sub.ftz.f32 %f7800, %f7722, %f7799; div.approx.ftz.f32 %f7801, %f7800, %f7714; mul.ftz.f32 %f7802, %f7801, %f7801; mul.ftz.f32 %f7803, %f7801, %f7802; mul.ftz.f32 %f7804, %f7802, %f7802; mul.ftz.f32 %f7805, %f7715, %f7802; fma.rn.ftz.f32 %f7806, %f7714, %f7801, %f7805; mul.ftz.f32 %f7807, %f7717, %f7804; fma.rn.ftz.f32 %f7808, %f7716, %f7803, %f7807; add.ftz.f32 %f7809, %f7806, %f7808; add.ftz.f32 %f7810, %f7713, %f7809; sub.ftz.f32 %f7811, %f7722, %f7810; div.approx.ftz.f32 %f7812, %f7811, %f7714; mul.ftz.f32 %f7813, %f7812, %f7812; mul.ftz.f32 %f7814, %f7812, %f7813; mul.ftz.f32 %f7815, %f7813, %f7813; mul.ftz.f32 %f7816, %f7715, %f7813; fma.rn.ftz.f32 %f7817, %f7714, %f7812, %f7816; mul.ftz.f32 %f7818, %f7717, %f7815; fma.rn.ftz.f32 %f7819, %f7716, %f7814, %f7818; add.ftz.f32 %f7820, %f7817, %f7819; add.ftz.f32 %f7821, %f7713, %f7820; sub.ftz.f32 %f7822, %f7722, %f7821; div.approx.ftz.f32 %f7823, %f7822, %f7714; mul.ftz.f32 %f7824, %f7823, %f7823; mul.ftz.f32 %f7825, %f7823, %f7824; mul.ftz.f32 %f7826, %f7824, %f7824; mul.ftz.f32 %f7827, %f7715, %f7824; fma.rn.ftz.f32 %f7828, %f7714, %f7823, %f7827; mul.ftz.f32 %f7829, %f7717, %f7826; fma.rn.ftz.f32 %f7830, %f7716, %f7825, %f7829; add.ftz.f32 %f7831, %f7828, %f7830; add.ftz.f32 %f7832, %f7713, %f7831; sub.ftz.f32 %f7833, %f7722, %f7832; div.approx.ftz.f32 %f7834, %f7833, %f7714; mul.ftz.f32 %f7835, %f7834, %f7834; mul.ftz.f32 %f7836, %f7834, %f7835; mul.ftz.f32 %f7837, %f7835, %f7835; mul.ftz.f32 %f7838, %f7715, %f7835; fma.rn.ftz.f32 %f7839, %f7714, %f7834, %f7838; mul.ftz.f32 %f7840, %f7717, %f7837; fma.rn.ftz.f32 %f7841, %f7716, %f7836, %f7840; add.ftz.f32 %f7842, %f7839, %f7841; add.ftz.f32 %f7843, %f7713, %f7842; sub.ftz.f32 %f7844, %f7722, %f7843; div.approx.ftz.f32 %f7845, %f7844, %f7714; mul.ftz.f32 %f7846, %f7845, %f7845; mul.ftz.f32 %f7847, %f7845, %f7846; mul.ftz.f32 %f7848, %f7846, %f7846; mul.ftz.f32 %f7849, %f7715, %f7846; fma.rn.ftz.f32 %f7850, %f7714, %f7845, %f7849; mul.ftz.f32 %f7851, %f7717, %f7848; fma.rn.ftz.f32 %f7852, %f7716, %f7847, %f7851; add.ftz.f32 %f7853, %f7850, %f7852; add.ftz.f32 %f7854, %f7713, %f7853; sub.ftz.f32 %f7855, %f7722, %f7854; div.approx.ftz.f32 %f7856, %f7855, %f7714; mul.ftz.f32 %f7857, %f7856, %f7856; mul.ftz.f32 %f7858, %f7856, %f7857; mul.ftz.f32 %f7859, %f7857, %f7857; mul.ftz.f32 %f7860, %f7715, %f7857; fma.rn.ftz.f32 %f7861, %f7714, %f7856, %f7860; mul.ftz.f32 %f7862, %f7717, %f7859; fma.rn.ftz.f32 %f7863, %f7716, %f7858, %f7862; add.ftz.f32 %f7864, %f7861, %f7863; add.ftz.f32 %f7865, %f7713, %f7864; sub.ftz.f32 %f7866, %f7722, %f7865; div.approx.ftz.f32 %f7867, %f7866, %f7714; mul.ftz.f32 %f7868, %f7867, %f7867; mul.ftz.f32 %f7869, %f7867, %f7868; mul.ftz.f32 %f7870, %f7868, %f7868; mul.ftz.f32 %f7871, %f7715, %f7868; fma.rn.ftz.f32 %f7872, %f7714, %f7867, %f7871; mul.ftz.f32 %f7873, %f7717, %f7870; fma.rn.ftz.f32 %f7874, %f7716, %f7869, %f7873; add.ftz.f32 %f7875, %f7872, %f7874; add.ftz.f32 %f7876, %f7713, %f7875; sub.ftz.f32 %f7877, %f7722, %f7876; div.approx.ftz.f32 %f7878, %f7877, %f7714; mul.ftz.f32 %f7879, %f7878, %f7878; mul.ftz.f32 %f7880, %f7878, %f7879; mul.ftz.f32 %f7881, %f7879, %f7879; mul.ftz.f32 %f7882, %f7715, %f7879; fma.rn.ftz.f32 %f7883, %f7714, %f7878, %f7882; mul.ftz.f32 %f7884, %f7717, %f7881; fma.rn.ftz.f32 %f7885, %f7716, %f7880, %f7884; add.ftz.f32 %f7886, %f7883, %f7885; add.ftz.f32 %f7887, %f7713, %f7886; sub.ftz.f32 %f7888, %f7722, %f7887; div.approx.ftz.f32 %f7889, %f7888, %f7714; mul.ftz.f32 %f7890, %f7889, %f7889; mul.ftz.f32 %f7891, %f7889, %f7890; mul.ftz.f32 %f7892, %f7890, %f7890; mul.ftz.f32 %f7893, %f7715, %f7890; fma.rn.ftz.f32 %f7894, %f7714, %f7889, %f7893; mul.ftz.f32 %f7895, %f7717, %f7892; fma.rn.ftz.f32 %f7896, %f7716, %f7891, %f7895; add.ftz.f32 %f7897, %f7894, %f7896; add.ftz.f32 %f7898, %f7713, %f7897; sub.ftz.f32 %f7899, %f7722, %f7898; div.approx.ftz.f32 %f7900, %f7899, %f7714; mul.ftz.f32 %f7901, %f7900, %f7900; mul.ftz.f32 %f7902, %f7900, %f7901; mul.ftz.f32 %f7903, %f7901, %f7901; mul.ftz.f32 %f7904, %f7715, %f7901; fma.rn.ftz.f32 %f7905, %f7714, %f7900, %f7904; mul.ftz.f32 %f7906, %f7717, %f7903; fma.rn.ftz.f32 %f7907, %f7716, %f7902, %f7906; add.ftz.f32 %f7908, %f7905, %f7907; add.ftz.f32 %f7909, %f7713, %f7908; sub.ftz.f32 %f7910, %f7722, %f7909; div.approx.ftz.f32 %f7911, %f7910, %f7714; mul.ftz.f32 %f7912, %f7911, %f7911; mul.ftz.f32 %f7913, %f7911, %f7912; mul.ftz.f32 %f7914, %f7912, %f7912; mul.ftz.f32 %f7915, %f7715, %f7912; fma.rn.ftz.f32 %f7916, %f7714, %f7911, %f7915; mul.ftz.f32 %f7917, %f7717, %f7914; fma.rn.ftz.f32 %f7918, %f7716, %f7913, %f7917; add.ftz.f32 %f7919, %f7916, %f7918; add.ftz.f32 %f7920, %f7713, %f7919; sub.ftz.f32 %f7921, %f7722, %f7920; div.approx.ftz.f32 %f7922, %f7921, %f7714; mul.ftz.f32 %f7923, %f7922, %f7922; mul.ftz.f32 %f7924, %f7922, %f7923; mul.ftz.f32 %f7925, %f7923, %f7923; mul.ftz.f32 %f7926, %f7715, %f7923; fma.rn.ftz.f32 %f7927, %f7714, %f7922, %f7926; mul.ftz.f32 %f7928, %f7717, %f7925; fma.rn.ftz.f32 %f7929, %f7716, %f7924, %f7928; add.ftz.f32 %f7930, %f7927, %f7929; add.ftz.f32 %f7931, %f7713, %f7930; sub.ftz.f32 %f7932, %f7722, %f7931; div.approx.ftz.f32 %f7933, %f7932, %f7714; mul.ftz.f32 %f7934, %f7933, %f7933; mul.ftz.f32 %f7935, %f7933, %f7934; mul.ftz.f32 %f7936, %f7934, %f7934; mul.ftz.f32 %f7937, %f7715, %f7934; fma.rn.ftz.f32 %f7938, %f7714, %f7933, %f7937; mul.ftz.f32 %f7939, %f7717, %f7936; fma.rn.ftz.f32 %f7940, %f7716, %f7935, %f7939; add.ftz.f32 %f7941, %f7938, %f7940; add.ftz.f32 %f7942, %f7713, %f7941; sub.ftz.f32 %f7943, %f7722, %f7942; div.approx.ftz.f32 %f1425, %f7943, %f7714; @%p2 bra $L__BB108_1006; bra.uni $L__BB108_1003; $L__BB108_1006: shr.s32 %r1646, %r494, 31; and.b32 %r1647, %r1646, 1078530011; or.b32 %r1648, %r1647, %r495; mov.b32 %f15994, %r1648; bra.uni $L__BB108_1007; $L__BB108_987: ld.const.v4.f32 {%f7508, %f7509, %f7510, %f7511}, [kernel_params+560]; ld.const.v4.f32 {%f7516, %f7517, %f7518, %f7519}, [kernel_params+576]; ld.const.v4.f32 {%f7532, %f7533, %f7534, %f7535}, [kernel_params+608]; mul.ftz.f32 %f7540, %f15965, %f7509; fma.rn.ftz.f32 %f7541, %f15964, %f7508, %f7540; fma.rn.ftz.f32 %f7542, %f15966, %f7510, %f7511; add.ftz.f32 %f1381, %f7541, %f7542; mul.ftz.f32 %f7543, %f15965, %f7517; fma.rn.ftz.f32 %f7544, %f15964, %f7516, %f7543; fma.rn.ftz.f32 %f7545, %f15966, %f7518, %f7519; add.ftz.f32 %f1382, %f7544, %f7545; mul.ftz.f32 %f7549, %f15965, %f7533; fma.rn.ftz.f32 %f7550, %f15964, %f7532, %f7549; fma.rn.ftz.f32 %f7551, %f15966, %f7534, %f7535; add.ftz.f32 %f1384, %f7550, %f7551; setp.eq.ftz.f32 %p837, %f1384, 0f00000000; mov.f32 %f16029, 0f00000000; mov.f32 %f16027, %f16029; mov.f32 %f16028, %f16029; @%p837 bra $L__BB108_989; div.approx.ftz.f32 %f16028, %f1382, %f1384; div.approx.ftz.f32 %f16027, %f1381, %f1384; $L__BB108_989: ld.const.v4.f32 {%f7555, %f7556, %f7557, %f7558}, [kernel_params+736]; ld.const.v4.f32 {%f7563, %f7564, %f7565, %f7566}, [kernel_params+752]; ld.const.v4.f32 {%f7579, %f7580, %f7581, %f7582}, [kernel_params+784]; mul.ftz.f32 %f7587, %f1376, %f7556; fma.rn.ftz.f32 %f7588, %f1375, %f7555, %f7587; fma.rn.ftz.f32 %f7589, %f1377, %f7557, %f7558; add.ftz.f32 %f1393, %f7588, %f7589; mul.ftz.f32 %f7590, %f1376, %f7564; fma.rn.ftz.f32 %f7591, %f1375, %f7563, %f7590; fma.rn.ftz.f32 %f7592, %f1377, %f7565, %f7566; add.ftz.f32 %f1394, %f7591, %f7592; mul.ftz.f32 %f7596, %f1376, %f7580; fma.rn.ftz.f32 %f7597, %f1375, %f7579, %f7596; fma.rn.ftz.f32 %f7598, %f1377, %f7581, %f7582; add.ftz.f32 %f1396, %f7597, %f7598; setp.eq.ftz.f32 %p838, %f1396, 0f00000000; mov.f32 %f16031, %f16029; mov.f32 %f16032, %f16029; @%p838 bra $L__BB108_991; div.approx.ftz.f32 %f16032, %f1394, %f1396; div.approx.ftz.f32 %f16031, %f1393, %f1396; $L__BB108_991: ld.const.v4.f32 {%f7602, %f7603, %f7604, %f7605}, [kernel_params+800]; ld.const.v4.f32 {%f7610, %f7611, %f7612, %f7613}, [kernel_params+816]; ld.const.v4.f32 {%f7626, %f7627, %f7628, %f7629}, [kernel_params+848]; mul.ftz.f32 %f7634, %f1379, %f7603; fma.rn.ftz.f32 %f7635, %f1378, %f7602, %f7634; fma.rn.ftz.f32 %f7636, %f1380, %f7604, %f7605; add.ftz.f32 %f1405, %f7635, %f7636; mul.ftz.f32 %f7637, %f1379, %f7611; fma.rn.ftz.f32 %f7638, %f1378, %f7610, %f7637; fma.rn.ftz.f32 %f7639, %f1380, %f7612, %f7613; add.ftz.f32 %f1406, %f7638, %f7639; mul.ftz.f32 %f7643, %f1379, %f7627; fma.rn.ftz.f32 %f7644, %f1378, %f7626, %f7643; fma.rn.ftz.f32 %f7645, %f1380, %f7628, %f7629; add.ftz.f32 %f1408, %f7644, %f7645; setp.eq.ftz.f32 %p839, %f1408, 0f00000000; mov.f32 %f16030, %f16029; @%p839 bra $L__BB108_1144; div.approx.ftz.f32 %f16030, %f1406, %f1408; div.approx.ftz.f32 %f16029, %f1405, %f1408; bra.uni $L__BB108_1144; $L__BB108_994: setp.eq.s32 %p848, %r493, 0; @%p848 bra $L__BB108_1030; setp.eq.s32 %p849, %r493, 1; @%p849 bra $L__BB108_996; bra.uni $L__BB108_1037; $L__BB108_996: ld.const.f32 %f1450, [kernel_params+24]; mul.ftz.f32 %f8043, %f1419, %f1419; fma.rn.ftz.f32 %f8044, %f1418, %f1418, %f8043; sqrt.approx.ftz.f32 %f8045, %f8044; abs.ftz.f32 %f1451, %f8045; abs.ftz.f32 %f1452, %f1417; setp.eq.ftz.f32 %p878, %f1452, 0f00000000; setp.eq.ftz.f32 %p879, %f1451, 0f00000000; and.pred %p880, %p878, %p879; mov.b32 %r498, %f1417; mov.b32 %r1663, %f8045; and.b32 %r499, %r1663, -2147483648; @%p880 bra $L__BB108_1023; bra.uni $L__BB108_997; $L__BB108_1023: shr.s32 %r1668, %r498, 31; and.b32 %r1669, %r1668, 1078530011; or.b32 %r1670, %r499, %r1669; mov.b32 %f15999, %r1670; bra.uni $L__BB108_1024; $L__BB108_1008: add.ftz.f32 %f7976, %f1418, 0fBF800000; mul.ftz.f32 %f7977, %f7976, 0fBF000000; mov.f32 %f7978, 0f00000000; max.ftz.f32 %f7979, %f7977, %f7978; sqrt.approx.ftz.f32 %f7980, %f7979; add.ftz.f32 %f1432, %f7980, %f7980; setp.leu.ftz.f32 %p857, %f1432, 0f00000000; @%p857 bra $L__BB108_1010; rcp.approx.ftz.f32 %f7981, %f1432; mul.ftz.f32 %f1417, %f1417, %f7981; mul.ftz.f32 %f1419, %f1419, %f7981; $L__BB108_1010: add.ftz.f32 %f7982, %f1417, 0f3F800000; add.ftz.f32 %f7983, %f1419, 0f3F800000; mul.ftz.f32 %f16004, %f7983, 0f3F000000; mul.ftz.f32 %f16003, %f7982, 0f3F000000; bra.uni $L__BB108_1043; $L__BB108_1011: abs.ftz.f32 %f1439, %f1417; setp.eq.ftz.f32 %p859, %f1439, 0f00000000; and.pred %p860, %p859, %p842; mov.b32 %r496, %f1417; and.b32 %r497, %r494, -2147483648; @%p860 bra $L__BB108_1015; bra.uni $L__BB108_1012; $L__BB108_1015: shr.s32 %r1653, %r496, 31; and.b32 %r1654, %r1653, 1078530011; or.b32 %r1655, %r1654, %r497; mov.b32 %f15997, %r1655; bra.uni $L__BB108_1016; $L__BB108_1030: setp.eq.ftz.f32 %p895, %f1418, 0f00000000; setp.eq.ftz.f32 %p896, %f1417, 0f00000000; and.pred %p897, %p896, %p895; setp.eq.ftz.f32 %p898, %f1419, 0f00000000; and.pred %p899, %p898, %p897; mov.f32 %f16003, 0f00000000; mov.f32 %f16004, %f16003; @%p899 bra $L__BB108_1043; ld.const.v4.f32 {%f8102, %f8103, %f8104, %f8105}, [kernel_params+32]; abs.ftz.f32 %f1468, %f1417; setp.eq.ftz.f32 %p901, %f1468, 0f00000000; and.pred %p902, %p901, %p842; mov.b32 %r500, %f1417; and.b32 %r501, %r494, -2147483648; @%p902 bra $L__BB108_1035; bra.uni $L__BB108_1032; $L__BB108_1035: shr.s32 %r1682, %r500, 31; and.b32 %r1683, %r1682, 1078530011; or.b32 %r1684, %r1683, %r501; mov.b32 %f16001, %r1684; bra.uni $L__BB108_1036; $L__BB108_1037: ld.const.f32 %f8164, [kernel_params+28]; ld.const.f32 %f1475, [kernel_params+276]; ld.const.f32 %f1476, [kernel_params+280]; add.ftz.f32 %f8165, %f8164, %f8164; mul.ftz.f32 %f8166, %f1420, 0f3F000000; sin.approx.ftz.f32 %f8167, %f8166; mul.ftz.f32 %f1477, %f8165, %f8167; @%p2 bra $L__BB108_1041; bra.uni $L__BB108_1038; $L__BB108_1041: shr.s32 %r1689, %r494, 31; and.b32 %r1690, %r1689, 1078530011; or.b32 %r1691, %r1690, %r495; mov.b32 %f16002, %r1691; bra.uni $L__BB108_1042; $L__BB108_1003: setp.eq.ftz.f32 %p850, %f1421, 0f7F800000; setp.eq.ftz.f32 %p851, %f1422, 0f7F800000; and.pred %p852, %p850, %p851; @%p852 bra $L__BB108_1005; bra.uni $L__BB108_1004; $L__BB108_1005: setp.lt.s32 %p856, %r494, 0; selp.b32 %r1644, 1075235812, 1061752795, %p856; or.b32 %r1645, %r1644, %r495; mov.b32 %f15994, %r1645; bra.uni $L__BB108_1007; $L__BB108_997: setp.eq.ftz.f32 %p881, %f1452, 0f7F800000; setp.eq.ftz.f32 %p882, %f1451, 0f7F800000; and.pred %p883, %p881, %p882; @%p883 bra $L__BB108_1022; bra.uni $L__BB108_998; $L__BB108_1022: setp.lt.s32 %p887, %r498, 0; selp.b32 %r1666, 1075235812, 1061752795, %p887; or.b32 %r1667, %r499, %r1666; mov.b32 %f15999, %r1667; bra.uni $L__BB108_1024; $L__BB108_1012: setp.eq.ftz.f32 %p861, %f1439, 0f7F800000; setp.eq.ftz.f32 %p862, %f1421, 0f7F800000; and.pred %p863, %p861, %p862; @%p863 bra $L__BB108_1014; bra.uni $L__BB108_1013; $L__BB108_1014: setp.lt.s32 %p867, %r496, 0; selp.b32 %r1651, 1075235812, 1061752795, %p867; or.b32 %r1652, %r1651, %r497; mov.b32 %f15997, %r1652; bra.uni $L__BB108_1016; $L__BB108_1038: setp.eq.ftz.f32 %p912, %f1421, 0f7F800000; setp.eq.ftz.f32 %p913, %f1422, 0f7F800000; and.pred %p914, %p912, %p913; @%p914 bra $L__BB108_1040; bra.uni $L__BB108_1039; $L__BB108_1040: setp.lt.s32 %p918, %r494, 0; selp.b32 %r1687, 1075235812, 1061752795, %p918; or.b32 %r1688, %r1687, %r495; mov.b32 %f16002, %r1688; bra.uni $L__BB108_1042; $L__BB108_1032: setp.eq.ftz.f32 %p903, %f1468, 0f7F800000; setp.eq.ftz.f32 %p904, %f1421, 0f7F800000; and.pred %p905, %p903, %p904; @%p905 bra $L__BB108_1034; bra.uni $L__BB108_1033; $L__BB108_1034: setp.lt.s32 %p909, %r500, 0; selp.b32 %r1680, 1075235812, 1061752795, %p909; or.b32 %r1681, %r1680, %r501; mov.b32 %f16001, %r1681; bra.uni $L__BB108_1036; $L__BB108_1004: mov.f32 %f15516, 0f3FC90FDB; setp.lt.s32 %p853, %r494, 0; min.ftz.f32 %f7944, %f1422, %f1421; max.ftz.f32 %f7945, %f1422, %f1421; div.full.ftz.f32 %f7946, %f7944, %f7945; mul.rn.ftz.f32 %f7947, %f7946, %f7946; mov.f32 %f7948, 0fC0B59883; mov.f32 %f7949, 0fBF52C7EA; fma.rn.ftz.f32 %f7950, %f7947, %f7949, %f7948; mov.f32 %f7951, 0fC0D21907; fma.rn.ftz.f32 %f7952, %f7950, %f7947, %f7951; mul.ftz.f32 %f7953, %f7947, %f7952; mul.ftz.f32 %f7954, %f7946, %f7953; add.ftz.f32 %f7955, %f7947, 0f41355DC0; mov.f32 %f7956, 0f41E6BD60; fma.rn.ftz.f32 %f7957, %f7955, %f7947, %f7956; mov.f32 %f7958, 0f419D92C8; fma.rn.ftz.f32 %f7959, %f7957, %f7947, %f7958; rcp.approx.ftz.f32 %f7960, %f7959; fma.rn.ftz.f32 %f7961, %f7954, %f7960, %f7946; sub.ftz.f32 %f7963, %f15516, %f7961; setp.gt.ftz.f32 %p854, %f1422, %f1421; selp.f32 %f7964, %f7963, %f7961, %p854; sub.ftz.f32 %f7966, %f7711, %f7964; selp.f32 %f7967, %f7966, %f7964, %p853; mov.b32 %r1642, %f7967; or.b32 %r1643, %r495, %r1642; mov.b32 %f7968, %r1643; add.ftz.f32 %f7969, %f1421, %f1422; setp.le.ftz.f32 %p855, %f7969, 0f7F800000; selp.f32 %f15994, %f7968, %f7969, %p855; $L__BB108_1007: ld.const.f32 %f15465, [kernel_params+280]; ld.const.f32 %f15464, [kernel_params+276]; cos.approx.ftz.f32 %f7970, %f15994; mul.ftz.f32 %f7971, %f1425, %f7970; div.approx.ftz.f32 %f7972, %f7971, %f15464; sin.approx.ftz.f32 %f7973, %f15994; mul.ftz.f32 %f7974, %f1425, %f7973; div.approx.ftz.f32 %f7975, %f7974, %f15465; add.ftz.f32 %f16004, %f7975, 0f3F000000; add.ftz.f32 %f16003, %f7972, 0f3F000000; bra.uni $L__BB108_1043; $L__BB108_998: setp.lt.s32 %p884, %r498, 0; min.ftz.f32 %f8046, %f1451, %f1452; max.ftz.f32 %f8047, %f1451, %f1452; div.full.ftz.f32 %f8048, %f8046, %f8047; mul.rn.ftz.f32 %f8049, %f8048, %f8048; mov.f32 %f8050, 0fC0B59883; mov.f32 %f8051, 0fBF52C7EA; fma.rn.ftz.f32 %f8052, %f8049, %f8051, %f8050; mov.f32 %f8053, 0fC0D21907; fma.rn.ftz.f32 %f8054, %f8052, %f8049, %f8053; mul.ftz.f32 %f8055, %f8049, %f8054; mul.ftz.f32 %f8056, %f8048, %f8055; add.ftz.f32 %f8057, %f8049, 0f41355DC0; mov.f32 %f8058, 0f41E6BD60; fma.rn.ftz.f32 %f8059, %f8057, %f8049, %f8058; mov.f32 %f8060, 0f419D92C8; fma.rn.ftz.f32 %f8061, %f8059, %f8049, %f8060; rcp.approx.ftz.f32 %f8062, %f8061; fma.rn.ftz.f32 %f8063, %f8056, %f8062, %f8048; sub.ftz.f32 %f8065, %f7708, %f8063; setp.gt.ftz.f32 %p885, %f1451, %f1452; selp.f32 %f8066, %f8065, %f8063, %p885; sub.ftz.f32 %f8068, %f7711, %f8066; selp.f32 %f8069, %f8068, %f8066, %p884; mov.b32 %r1664, %f8069; or.b32 %r1665, %r499, %r1664; mov.b32 %f8070, %r1665; add.ftz.f32 %f8071, %f1452, %f1451; setp.le.ftz.f32 %p886, %f8071, 0f7F800000; selp.f32 %f15999, %f8070, %f8071, %p886; $L__BB108_1024: div.approx.ftz.f32 %f1457, %f15999, %f1450; @%p2 bra $L__BB108_1028; bra.uni $L__BB108_1025; $L__BB108_1028: shr.s32 %r1675, %r494, 31; and.b32 %r1676, %r1675, 1078530011; or.b32 %r1677, %r1676, %r495; mov.b32 %f16000, %r1677; bra.uni $L__BB108_1029; $L__BB108_1025: setp.eq.ftz.f32 %p888, %f1421, 0f7F800000; setp.eq.ftz.f32 %p889, %f1422, 0f7F800000; and.pred %p890, %p888, %p889; @%p890 bra $L__BB108_1027; bra.uni $L__BB108_1026; $L__BB108_1027: setp.lt.s32 %p894, %r494, 0; selp.b32 %r1673, 1075235812, 1061752795, %p894; or.b32 %r1674, %r1673, %r495; mov.b32 %f16000, %r1674; bra.uni $L__BB108_1029; $L__BB108_1013: setp.lt.s32 %p864, %r496, 0; min.ftz.f32 %f7984, %f1421, %f1439; max.ftz.f32 %f7985, %f1421, %f1439; div.full.ftz.f32 %f7986, %f7984, %f7985; mul.rn.ftz.f32 %f7987, %f7986, %f7986; mov.f32 %f7988, 0fC0B59883; mov.f32 %f7989, 0fBF52C7EA; fma.rn.ftz.f32 %f7990, %f7987, %f7989, %f7988; mov.f32 %f7991, 0fC0D21907; fma.rn.ftz.f32 %f7992, %f7990, %f7987, %f7991; mul.ftz.f32 %f7993, %f7987, %f7992; mul.ftz.f32 %f7994, %f7986, %f7993; add.ftz.f32 %f7995, %f7987, 0f41355DC0; mov.f32 %f7996, 0f41E6BD60; fma.rn.ftz.f32 %f7997, %f7995, %f7987, %f7996; mov.f32 %f7998, 0f419D92C8; fma.rn.ftz.f32 %f7999, %f7997, %f7987, %f7998; rcp.approx.ftz.f32 %f8000, %f7999; fma.rn.ftz.f32 %f8001, %f7994, %f8000, %f7986; sub.ftz.f32 %f8003, %f7708, %f8001; setp.gt.ftz.f32 %p865, %f1421, %f1439; selp.f32 %f8004, %f8003, %f8001, %p865; sub.ftz.f32 %f8006, %f7711, %f8004; selp.f32 %f8007, %f8006, %f8004, %p864; mov.b32 %r1649, %f8007; or.b32 %r1650, %r497, %r1649; mov.b32 %f8008, %r1650; add.ftz.f32 %f8009, %f1439, %f1421; setp.le.ftz.f32 %p866, %f8009, 0f7F800000; selp.f32 %f15997, %f8008, %f8009, %p866; $L__BB108_1016: and.pred %p870, %p859, %p843; @%p870 bra $L__BB108_1020; bra.uni $L__BB108_1017; $L__BB108_1020: shr.s32 %r1660, %r496, 31; and.b32 %r1661, %r1660, 1078530011; or.b32 %r1662, %r495, %r1661; mov.b32 %f15998, %r1662; bra.uni $L__BB108_1021; $L__BB108_1017: setp.eq.ftz.f32 %p871, %f1439, 0f7F800000; setp.eq.ftz.f32 %p872, %f1422, 0f7F800000; and.pred %p873, %p871, %p872; @%p873 bra $L__BB108_1019; bra.uni $L__BB108_1018; $L__BB108_1019: setp.lt.s32 %p877, %r496, 0; selp.b32 %r1658, 1075235812, 1061752795, %p877; or.b32 %r1659, %r495, %r1658; mov.b32 %f15998, %r1659; bra.uni $L__BB108_1021; $L__BB108_1039: setp.lt.s32 %p915, %r494, 0; min.ftz.f32 %f8168, %f1422, %f1421; max.ftz.f32 %f8169, %f1422, %f1421; div.full.ftz.f32 %f8170, %f8168, %f8169; mul.rn.ftz.f32 %f8171, %f8170, %f8170; mov.f32 %f8172, 0fC0B59883; mov.f32 %f8173, 0fBF52C7EA; fma.rn.ftz.f32 %f8174, %f8171, %f8173, %f8172; mov.f32 %f8175, 0fC0D21907; fma.rn.ftz.f32 %f8176, %f8174, %f8171, %f8175; mul.ftz.f32 %f8177, %f8171, %f8176; mul.ftz.f32 %f8178, %f8170, %f8177; add.ftz.f32 %f8179, %f8171, 0f41355DC0; mov.f32 %f8180, 0f41E6BD60; fma.rn.ftz.f32 %f8181, %f8179, %f8171, %f8180; mov.f32 %f8182, 0f419D92C8; fma.rn.ftz.f32 %f8183, %f8181, %f8171, %f8182; rcp.approx.ftz.f32 %f8184, %f8183; fma.rn.ftz.f32 %f8185, %f8178, %f8184, %f8170; sub.ftz.f32 %f8187, %f7708, %f8185; setp.gt.ftz.f32 %p916, %f1422, %f1421; selp.f32 %f8188, %f8187, %f8185, %p916; sub.ftz.f32 %f8190, %f7711, %f8188; selp.f32 %f8191, %f8190, %f8188, %p915; mov.b32 %r1685, %f8191; or.b32 %r1686, %r495, %r1685; mov.b32 %f8192, %r1686; add.ftz.f32 %f8193, %f1421, %f1422; setp.le.ftz.f32 %p917, %f8193, 0f7F800000; selp.f32 %f16002, %f8192, %f8193, %p917; $L__BB108_1042: cos.approx.ftz.f32 %f8194, %f16002; mul.ftz.f32 %f8195, %f1477, %f8194; div.approx.ftz.f32 %f8196, %f8195, %f1475; sin.approx.ftz.f32 %f8197, %f16002; mul.ftz.f32 %f8198, %f1477, %f8197; div.approx.ftz.f32 %f8199, %f8198, %f1476; add.ftz.f32 %f16004, %f8199, 0f3F000000; add.ftz.f32 %f16003, %f8196, 0f3F000000; bra.uni $L__BB108_1043; $L__BB108_1026: setp.lt.s32 %p891, %r494, 0; min.ftz.f32 %f8072, %f1422, %f1421; max.ftz.f32 %f8073, %f1422, %f1421; div.full.ftz.f32 %f8074, %f8072, %f8073; mul.rn.ftz.f32 %f8075, %f8074, %f8074; mov.f32 %f8076, 0fC0B59883; mov.f32 %f8077, 0fBF52C7EA; fma.rn.ftz.f32 %f8078, %f8075, %f8077, %f8076; mov.f32 %f8079, 0fC0D21907; fma.rn.ftz.f32 %f8080, %f8078, %f8075, %f8079; mul.ftz.f32 %f8081, %f8075, %f8080; mul.ftz.f32 %f8082, %f8074, %f8081; add.ftz.f32 %f8083, %f8075, 0f41355DC0; mov.f32 %f8084, 0f41E6BD60; fma.rn.ftz.f32 %f8085, %f8083, %f8075, %f8084; mov.f32 %f8086, 0f419D92C8; fma.rn.ftz.f32 %f8087, %f8085, %f8075, %f8086; rcp.approx.ftz.f32 %f8088, %f8087; fma.rn.ftz.f32 %f8089, %f8082, %f8088, %f8074; sub.ftz.f32 %f8091, %f7708, %f8089; setp.gt.ftz.f32 %p892, %f1422, %f1421; selp.f32 %f8092, %f8091, %f8089, %p892; sub.ftz.f32 %f8094, %f7711, %f8092; selp.f32 %f8095, %f8094, %f8092, %p891; mov.b32 %r1671, %f8095; or.b32 %r1672, %r495, %r1671; mov.b32 %f8096, %r1672; add.ftz.f32 %f8097, %f1421, %f1422; setp.le.ftz.f32 %p893, %f8097, 0f7F800000; selp.f32 %f16000, %f8096, %f8097, %p893; $L__BB108_1029: cos.approx.ftz.f32 %f8098, %f16000; sin.approx.ftz.f32 %f8099, %f16000; fma.rn.ftz.f32 %f16004, %f1457, %f8099, 0f3F000000; fma.rn.ftz.f32 %f16003, %f1457, %f8098, 0f3F000000; bra.uni $L__BB108_1043; $L__BB108_1018: setp.lt.s32 %p874, %r496, 0; min.ftz.f32 %f8010, %f1422, %f1439; max.ftz.f32 %f8011, %f1422, %f1439; div.full.ftz.f32 %f8012, %f8010, %f8011; mul.rn.ftz.f32 %f8013, %f8012, %f8012; mov.f32 %f8014, 0fC0B59883; mov.f32 %f8015, 0fBF52C7EA; fma.rn.ftz.f32 %f8016, %f8013, %f8015, %f8014; mov.f32 %f8017, 0fC0D21907; fma.rn.ftz.f32 %f8018, %f8016, %f8013, %f8017; mul.ftz.f32 %f8019, %f8013, %f8018; mul.ftz.f32 %f8020, %f8012, %f8019; add.ftz.f32 %f8021, %f8013, 0f41355DC0; mov.f32 %f8022, 0f41E6BD60; fma.rn.ftz.f32 %f8023, %f8021, %f8013, %f8022; mov.f32 %f8024, 0f419D92C8; fma.rn.ftz.f32 %f8025, %f8023, %f8013, %f8024; rcp.approx.ftz.f32 %f8026, %f8025; fma.rn.ftz.f32 %f8027, %f8020, %f8026, %f8012; sub.ftz.f32 %f8029, %f7708, %f8027; setp.gt.ftz.f32 %p875, %f1422, %f1439; selp.f32 %f8030, %f8029, %f8027, %p875; sub.ftz.f32 %f8032, %f7711, %f8030; selp.f32 %f8033, %f8032, %f8030, %p874; mov.b32 %r1656, %f8033; or.b32 %r1657, %r495, %r1656; mov.b32 %f8034, %r1657; add.ftz.f32 %f8035, %f1439, %f1422; setp.le.ftz.f32 %p876, %f8035, 0f7F800000; selp.f32 %f15998, %f8034, %f8035, %p876; $L__BB108_1021: add.ftz.f32 %f8036, %f15998, %f15998; div.approx.ftz.f32 %f8038, %f8036, %f7711; add.ftz.f32 %f8039, %f15997, %f15997; div.approx.ftz.f32 %f8040, %f8039, %f7711; add.ftz.f32 %f8041, %f8040, 0f3F000000; sub.ftz.f32 %f16003, %f7688, %f8041; add.ftz.f32 %f16004, %f8038, 0f3F000000; bra.uni $L__BB108_1043; $L__BB108_1033: setp.lt.s32 %p906, %r500, 0; min.ftz.f32 %f8106, %f1421, %f1468; max.ftz.f32 %f8107, %f1421, %f1468; div.full.ftz.f32 %f8108, %f8106, %f8107; mul.rn.ftz.f32 %f8109, %f8108, %f8108; mov.f32 %f8110, 0fC0B59883; mov.f32 %f8111, 0fBF52C7EA; fma.rn.ftz.f32 %f8112, %f8109, %f8111, %f8110; mov.f32 %f8113, 0fC0D21907; fma.rn.ftz.f32 %f8114, %f8112, %f8109, %f8113; mul.ftz.f32 %f8115, %f8109, %f8114; mul.ftz.f32 %f8116, %f8108, %f8115; add.ftz.f32 %f8117, %f8109, 0f41355DC0; mov.f32 %f8118, 0f41E6BD60; fma.rn.ftz.f32 %f8119, %f8117, %f8109, %f8118; mov.f32 %f8120, 0f419D92C8; fma.rn.ftz.f32 %f8121, %f8119, %f8109, %f8120; rcp.approx.ftz.f32 %f8122, %f8121; fma.rn.ftz.f32 %f8123, %f8116, %f8122, %f8108; sub.ftz.f32 %f8125, %f7708, %f8123; setp.gt.ftz.f32 %p907, %f1421, %f1468; selp.f32 %f8126, %f8125, %f8123, %p907; sub.ftz.f32 %f8128, %f7711, %f8126; selp.f32 %f8129, %f8128, %f8126, %p906; mov.b32 %r1678, %f8129; or.b32 %r1679, %r501, %r1678; mov.b32 %f8130, %r1679; add.ftz.f32 %f8131, %f1468, %f1421; setp.le.ftz.f32 %p908, %f8131, 0f7F800000; selp.f32 %f16001, %f8130, %f8131, %p908; $L__BB108_1036: sub.ftz.f32 %f8132, %f16001, %f8103; mul.ftz.f32 %f8133, %f1418, %f1418; fma.rn.ftz.f32 %f8134, %f1417, %f1417, %f8133; fma.rn.ftz.f32 %f8135, %f1419, %f1419, %f8134; sqrt.approx.ftz.f32 %f8136, %f8135; div.approx.ftz.f32 %f8137, %f1419, %f8136; abs.ftz.f32 %f8138, %f8137; sub.ftz.f32 %f8140, %f7688, %f8138; mul.ftz.f32 %f8141, %f8140, 0f3F000000; sqrt.approx.ftz.f32 %f8142, %f8141; setp.gt.ftz.f32 %p910, %f8138, 0f3F11EB85; selp.f32 %f8143, %f8142, %f8138, %p910; mul.ftz.f32 %f8144, %f8143, %f8143; fma.rn.ftz.f32 %f8147, %f7697, %f8144, %f7696; fma.rn.ftz.f32 %f8149, %f8147, %f8144, %f7699; fma.rn.ftz.f32 %f8151, %f8149, %f8144, %f7701; fma.rn.ftz.f32 %f8153, %f8151, %f8144, %f7703; mul.ftz.f32 %f8154, %f8144, %f8153; fma.rn.ftz.f32 %f8155, %f8154, %f8143, %f8143; add.ftz.f32 %f8156, %f8155, %f8155; sub.ftz.f32 %f8158, %f7708, %f8155; selp.f32 %f8159, %f8156, %f8158, %p910; setp.lt.ftz.f32 %p911, %f8137, 0f00000000; sub.ftz.f32 %f8161, %f7711, %f8159; selp.f32 %f8162, %f8161, %f8159, %p911; sub.ftz.f32 %f8163, %f8162, %f8105; div.approx.ftz.f32 %f16004, %f8163, %f8104; div.approx.ftz.f32 %f16003, %f8132, %f8102; $L__BB108_1043: ld.const.u32 %r3061, [kernel_params+20]; setp.gt.s32 %p2183, %r3061, 2; mov.f32 %f15517, 0f3FC90FDB; mov.f32 %f15472, 0fBF800000; mov.f32 %f15471, 0f3E2AAB94; mov.f32 %f15470, 0f3D994929; mov.f32 %f15469, 0f3D3F841F; mov.f32 %f15468, 0f3C94D2E9; mov.f32 %f15467, 0f3D53F941; mov.f32 %f15466, 0f3F800000; ld.const.v4.f32 {%f8200, %f8201, %f8202, %f8203}, [kernel_params+864]; ld.const.v4.f32 {%f8208, %f8209, %f8210, %f8211}, [kernel_params+880]; ld.const.v4.f32 {%f8216, %f8217, %f8218, %f8219}, [kernel_params+896]; mul.ftz.f32 %f8224, %f1376, %f8201; fma.rn.ftz.f32 %f8225, %f1375, %f8200, %f8224; fma.rn.ftz.f32 %f8226, %f1377, %f8202, %f8225; add.ftz.f32 %f8227, %f8203, %f8226; mul.ftz.f32 %f8228, %f1376, %f8209; fma.rn.ftz.f32 %f8229, %f1375, %f8208, %f8228; fma.rn.ftz.f32 %f8230, %f1377, %f8210, %f8229; add.ftz.f32 %f8231, %f8211, %f8230; mul.ftz.f32 %f8232, %f1376, %f8217; fma.rn.ftz.f32 %f8233, %f1375, %f8216, %f8232; fma.rn.ftz.f32 %f8234, %f1377, %f8218, %f8233; add.ftz.f32 %f8235, %f8219, %f8234; mul.ftz.f32 %f8236, %f8231, %f8231; fma.rn.ftz.f32 %f8237, %f8227, %f8227, %f8236; fma.rn.ftz.f32 %f8238, %f8235, %f8235, %f8237; sqrt.approx.ftz.f32 %f8239, %f8238; div.approx.ftz.f32 %f1488, %f8227, %f8239; div.approx.ftz.f32 %f1489, %f8231, %f8239; div.approx.ftz.f32 %f1490, %f8235, %f8239; max.ftz.f32 %f8241, %f1488, %f15472; min.ftz.f32 %f8243, %f8241, %f15466; abs.ftz.f32 %f8244, %f8243; sub.ftz.f32 %f8245, %f15466, %f8244; mul.ftz.f32 %f8246, %f8245, 0f3F000000; sqrt.approx.ftz.f32 %f8247, %f8246; setp.gt.ftz.f32 %p919, %f8244, 0f3F11EB85; selp.f32 %f8248, %f8247, %f8244, %p919; mul.ftz.f32 %f8249, %f8248, %f8248; fma.rn.ftz.f32 %f8252, %f15467, %f8249, %f15468; fma.rn.ftz.f32 %f8254, %f8252, %f8249, %f15469; fma.rn.ftz.f32 %f8256, %f8254, %f8249, %f15470; fma.rn.ftz.f32 %f8258, %f8256, %f8249, %f15471; mul.ftz.f32 %f8259, %f8249, %f8258; fma.rn.ftz.f32 %f8260, %f8259, %f8248, %f8248; add.ftz.f32 %f8261, %f8260, %f8260; sub.ftz.f32 %f8263, %f15517, %f8260; selp.f32 %f8264, %f8261, %f8263, %p919; setp.lt.ftz.f32 %p920, %f8243, 0f00000000; sub.ftz.f32 %f8266, %f7711, %f8264; selp.f32 %f1491, %f8266, %f8264, %p920; abs.ftz.f32 %f1492, %f1489; abs.ftz.f32 %f1493, %f1490; setp.eq.ftz.f32 %p921, %f1492, 0f00000000; setp.eq.ftz.f32 %p922, %f1493, 0f00000000; and.pred %p3, %p921, %p922; mov.b32 %r502, %f1489; mov.b32 %r1692, %f1490; and.b32 %r503, %r1692, -2147483648; @%p2183 bra $L__BB108_1049; bra.uni $L__BB108_1044; $L__BB108_1049: ld.const.u32 %r3046, [kernel_params+20]; setp.eq.s32 %p924, %r3046, 3; @%p924 bra $L__BB108_1058; ld.const.u32 %r3047, [kernel_params+20]; setp.eq.s32 %p925, %r3047, 5; @%p925 bra $L__BB108_1061; ld.const.u32 %r3048, [kernel_params+20]; setp.ne.s32 %p926, %r3048, 4; @%p926 bra $L__BB108_1087; ld.const.f32 %f8267, [kernel_params+48]; ld.const.v4.f32 {%f8268, %f8269, %f8270, %f8271}, [kernel_params+64]; neg.ftz.f32 %f8276, %f1491; sub.ftz.f32 %f8277, %f8276, %f8267; div.approx.ftz.f32 %f8278, %f8277, %f8268; mul.ftz.f32 %f8279, %f8278, %f8278; mul.ftz.f32 %f8280, %f8278, %f8279; mul.ftz.f32 %f8281, %f8279, %f8279; mul.ftz.f32 %f8282, %f8269, %f8279; fma.rn.ftz.f32 %f8283, %f8268, %f8278, %f8282; mul.ftz.f32 %f8284, %f8271, %f8281; fma.rn.ftz.f32 %f8285, %f8270, %f8280, %f8284; add.ftz.f32 %f8286, %f8283, %f8285; add.ftz.f32 %f8287, %f8267, %f8286; sub.ftz.f32 %f8288, %f8276, %f8287; div.approx.ftz.f32 %f8289, %f8288, %f8268; mul.ftz.f32 %f8290, %f8289, %f8289; mul.ftz.f32 %f8291, %f8289, %f8290; mul.ftz.f32 %f8292, %f8290, %f8290; mul.ftz.f32 %f8293, %f8269, %f8290; fma.rn.ftz.f32 %f8294, %f8268, %f8289, %f8293; mul.ftz.f32 %f8295, %f8271, %f8292; fma.rn.ftz.f32 %f8296, %f8270, %f8291, %f8295; add.ftz.f32 %f8297, %f8294, %f8296; add.ftz.f32 %f8298, %f8267, %f8297; sub.ftz.f32 %f8299, %f8276, %f8298; div.approx.ftz.f32 %f8300, %f8299, %f8268; mul.ftz.f32 %f8301, %f8300, %f8300; mul.ftz.f32 %f8302, %f8300, %f8301; mul.ftz.f32 %f8303, %f8301, %f8301; mul.ftz.f32 %f8304, %f8269, %f8301; fma.rn.ftz.f32 %f8305, %f8268, %f8300, %f8304; mul.ftz.f32 %f8306, %f8271, %f8303; fma.rn.ftz.f32 %f8307, %f8270, %f8302, %f8306; add.ftz.f32 %f8308, %f8305, %f8307; add.ftz.f32 %f8309, %f8267, %f8308; sub.ftz.f32 %f8310, %f8276, %f8309; div.approx.ftz.f32 %f8311, %f8310, %f8268; mul.ftz.f32 %f8312, %f8311, %f8311; mul.ftz.f32 %f8313, %f8311, %f8312; mul.ftz.f32 %f8314, %f8312, %f8312; mul.ftz.f32 %f8315, %f8269, %f8312; fma.rn.ftz.f32 %f8316, %f8268, %f8311, %f8315; mul.ftz.f32 %f8317, %f8271, %f8314; fma.rn.ftz.f32 %f8318, %f8270, %f8313, %f8317; add.ftz.f32 %f8319, %f8316, %f8318; add.ftz.f32 %f8320, %f8267, %f8319; sub.ftz.f32 %f8321, %f8276, %f8320; div.approx.ftz.f32 %f8322, %f8321, %f8268; mul.ftz.f32 %f8323, %f8322, %f8322; mul.ftz.f32 %f8324, %f8322, %f8323; mul.ftz.f32 %f8325, %f8323, %f8323; mul.ftz.f32 %f8326, %f8269, %f8323; fma.rn.ftz.f32 %f8327, %f8268, %f8322, %f8326; mul.ftz.f32 %f8328, %f8271, %f8325; fma.rn.ftz.f32 %f8329, %f8270, %f8324, %f8328; add.ftz.f32 %f8330, %f8327, %f8329; add.ftz.f32 %f8331, %f8267, %f8330; sub.ftz.f32 %f8332, %f8276, %f8331; div.approx.ftz.f32 %f8333, %f8332, %f8268; mul.ftz.f32 %f8334, %f8333, %f8333; mul.ftz.f32 %f8335, %f8333, %f8334; mul.ftz.f32 %f8336, %f8334, %f8334; mul.ftz.f32 %f8337, %f8269, %f8334; fma.rn.ftz.f32 %f8338, %f8268, %f8333, %f8337; mul.ftz.f32 %f8339, %f8271, %f8336; fma.rn.ftz.f32 %f8340, %f8270, %f8335, %f8339; add.ftz.f32 %f8341, %f8338, %f8340; add.ftz.f32 %f8342, %f8267, %f8341; sub.ftz.f32 %f8343, %f8276, %f8342; div.approx.ftz.f32 %f8344, %f8343, %f8268; mul.ftz.f32 %f8345, %f8344, %f8344; mul.ftz.f32 %f8346, %f8344, %f8345; mul.ftz.f32 %f8347, %f8345, %f8345; mul.ftz.f32 %f8348, %f8269, %f8345; fma.rn.ftz.f32 %f8349, %f8268, %f8344, %f8348; mul.ftz.f32 %f8350, %f8271, %f8347; fma.rn.ftz.f32 %f8351, %f8270, %f8346, %f8350; add.ftz.f32 %f8352, %f8349, %f8351; add.ftz.f32 %f8353, %f8267, %f8352; sub.ftz.f32 %f8354, %f8276, %f8353; div.approx.ftz.f32 %f8355, %f8354, %f8268; mul.ftz.f32 %f8356, %f8355, %f8355; mul.ftz.f32 %f8357, %f8355, %f8356; mul.ftz.f32 %f8358, %f8356, %f8356; mul.ftz.f32 %f8359, %f8269, %f8356; fma.rn.ftz.f32 %f8360, %f8268, %f8355, %f8359; mul.ftz.f32 %f8361, %f8271, %f8358; fma.rn.ftz.f32 %f8362, %f8270, %f8357, %f8361; add.ftz.f32 %f8363, %f8360, %f8362; add.ftz.f32 %f8364, %f8267, %f8363; sub.ftz.f32 %f8365, %f8276, %f8364; div.approx.ftz.f32 %f8366, %f8365, %f8268; mul.ftz.f32 %f8367, %f8366, %f8366; mul.ftz.f32 %f8368, %f8366, %f8367; mul.ftz.f32 %f8369, %f8367, %f8367; mul.ftz.f32 %f8370, %f8269, %f8367; fma.rn.ftz.f32 %f8371, %f8268, %f8366, %f8370; mul.ftz.f32 %f8372, %f8271, %f8369; fma.rn.ftz.f32 %f8373, %f8270, %f8368, %f8372; add.ftz.f32 %f8374, %f8371, %f8373; add.ftz.f32 %f8375, %f8267, %f8374; sub.ftz.f32 %f8376, %f8276, %f8375; div.approx.ftz.f32 %f8377, %f8376, %f8268; mul.ftz.f32 %f8378, %f8377, %f8377; mul.ftz.f32 %f8379, %f8377, %f8378; mul.ftz.f32 %f8380, %f8378, %f8378; mul.ftz.f32 %f8381, %f8269, %f8378; fma.rn.ftz.f32 %f8382, %f8268, %f8377, %f8381; mul.ftz.f32 %f8383, %f8271, %f8380; fma.rn.ftz.f32 %f8384, %f8270, %f8379, %f8383; add.ftz.f32 %f8385, %f8382, %f8384; add.ftz.f32 %f8386, %f8267, %f8385; sub.ftz.f32 %f8387, %f8276, %f8386; div.approx.ftz.f32 %f8388, %f8387, %f8268; mul.ftz.f32 %f8389, %f8388, %f8388; mul.ftz.f32 %f8390, %f8388, %f8389; mul.ftz.f32 %f8391, %f8389, %f8389; mul.ftz.f32 %f8392, %f8269, %f8389; fma.rn.ftz.f32 %f8393, %f8268, %f8388, %f8392; mul.ftz.f32 %f8394, %f8271, %f8391; fma.rn.ftz.f32 %f8395, %f8270, %f8390, %f8394; add.ftz.f32 %f8396, %f8393, %f8395; add.ftz.f32 %f8397, %f8267, %f8396; sub.ftz.f32 %f8398, %f8276, %f8397; div.approx.ftz.f32 %f8399, %f8398, %f8268; mul.ftz.f32 %f8400, %f8399, %f8399; mul.ftz.f32 %f8401, %f8399, %f8400; mul.ftz.f32 %f8402, %f8400, %f8400; mul.ftz.f32 %f8403, %f8269, %f8400; fma.rn.ftz.f32 %f8404, %f8268, %f8399, %f8403; mul.ftz.f32 %f8405, %f8271, %f8402; fma.rn.ftz.f32 %f8406, %f8270, %f8401, %f8405; add.ftz.f32 %f8407, %f8404, %f8406; add.ftz.f32 %f8408, %f8267, %f8407; sub.ftz.f32 %f8409, %f8276, %f8408; div.approx.ftz.f32 %f8410, %f8409, %f8268; mul.ftz.f32 %f8411, %f8410, %f8410; mul.ftz.f32 %f8412, %f8410, %f8411; mul.ftz.f32 %f8413, %f8411, %f8411; mul.ftz.f32 %f8414, %f8269, %f8411; fma.rn.ftz.f32 %f8415, %f8268, %f8410, %f8414; mul.ftz.f32 %f8416, %f8271, %f8413; fma.rn.ftz.f32 %f8417, %f8270, %f8412, %f8416; add.ftz.f32 %f8418, %f8415, %f8417; add.ftz.f32 %f8419, %f8267, %f8418; sub.ftz.f32 %f8420, %f8276, %f8419; div.approx.ftz.f32 %f8421, %f8420, %f8268; mul.ftz.f32 %f8422, %f8421, %f8421; mul.ftz.f32 %f8423, %f8421, %f8422; mul.ftz.f32 %f8424, %f8422, %f8422; mul.ftz.f32 %f8425, %f8269, %f8422; fma.rn.ftz.f32 %f8426, %f8268, %f8421, %f8425; mul.ftz.f32 %f8427, %f8271, %f8424; fma.rn.ftz.f32 %f8428, %f8270, %f8423, %f8427; add.ftz.f32 %f8429, %f8426, %f8428; add.ftz.f32 %f8430, %f8267, %f8429; sub.ftz.f32 %f8431, %f8276, %f8430; div.approx.ftz.f32 %f8432, %f8431, %f8268; mul.ftz.f32 %f8433, %f8432, %f8432; mul.ftz.f32 %f8434, %f8432, %f8433; mul.ftz.f32 %f8435, %f8433, %f8433; mul.ftz.f32 %f8436, %f8269, %f8433; fma.rn.ftz.f32 %f8437, %f8268, %f8432, %f8436; mul.ftz.f32 %f8438, %f8271, %f8435; fma.rn.ftz.f32 %f8439, %f8270, %f8434, %f8438; add.ftz.f32 %f8440, %f8437, %f8439; add.ftz.f32 %f8441, %f8267, %f8440; sub.ftz.f32 %f8442, %f8276, %f8441; div.approx.ftz.f32 %f8443, %f8442, %f8268; mul.ftz.f32 %f8444, %f8443, %f8443; mul.ftz.f32 %f8445, %f8443, %f8444; mul.ftz.f32 %f8446, %f8444, %f8444; mul.ftz.f32 %f8447, %f8269, %f8444; fma.rn.ftz.f32 %f8448, %f8268, %f8443, %f8447; mul.ftz.f32 %f8449, %f8271, %f8446; fma.rn.ftz.f32 %f8450, %f8270, %f8445, %f8449; add.ftz.f32 %f8451, %f8448, %f8450; add.ftz.f32 %f8452, %f8267, %f8451; sub.ftz.f32 %f8453, %f8276, %f8452; div.approx.ftz.f32 %f8454, %f8453, %f8268; mul.ftz.f32 %f8455, %f8454, %f8454; mul.ftz.f32 %f8456, %f8454, %f8455; mul.ftz.f32 %f8457, %f8455, %f8455; mul.ftz.f32 %f8458, %f8269, %f8455; fma.rn.ftz.f32 %f8459, %f8268, %f8454, %f8458; mul.ftz.f32 %f8460, %f8271, %f8457; fma.rn.ftz.f32 %f8461, %f8270, %f8456, %f8460; add.ftz.f32 %f8462, %f8459, %f8461; add.ftz.f32 %f8463, %f8267, %f8462; sub.ftz.f32 %f8464, %f8276, %f8463; div.approx.ftz.f32 %f8465, %f8464, %f8268; mul.ftz.f32 %f8466, %f8465, %f8465; mul.ftz.f32 %f8467, %f8465, %f8466; mul.ftz.f32 %f8468, %f8466, %f8466; mul.ftz.f32 %f8469, %f8269, %f8466; fma.rn.ftz.f32 %f8470, %f8268, %f8465, %f8469; mul.ftz.f32 %f8471, %f8271, %f8468; fma.rn.ftz.f32 %f8472, %f8270, %f8467, %f8471; add.ftz.f32 %f8473, %f8470, %f8472; add.ftz.f32 %f8474, %f8267, %f8473; sub.ftz.f32 %f8475, %f8276, %f8474; div.approx.ftz.f32 %f8476, %f8475, %f8268; mul.ftz.f32 %f8477, %f8476, %f8476; mul.ftz.f32 %f8478, %f8476, %f8477; mul.ftz.f32 %f8479, %f8477, %f8477; mul.ftz.f32 %f8480, %f8269, %f8477; fma.rn.ftz.f32 %f8481, %f8268, %f8476, %f8480; mul.ftz.f32 %f8482, %f8271, %f8479; fma.rn.ftz.f32 %f8483, %f8270, %f8478, %f8482; add.ftz.f32 %f8484, %f8481, %f8483; add.ftz.f32 %f8485, %f8267, %f8484; sub.ftz.f32 %f8486, %f8276, %f8485; div.approx.ftz.f32 %f8487, %f8486, %f8268; mul.ftz.f32 %f8488, %f8487, %f8487; mul.ftz.f32 %f8489, %f8487, %f8488; mul.ftz.f32 %f8490, %f8488, %f8488; mul.ftz.f32 %f8491, %f8269, %f8488; fma.rn.ftz.f32 %f8492, %f8268, %f8487, %f8491; mul.ftz.f32 %f8493, %f8271, %f8490; fma.rn.ftz.f32 %f8494, %f8270, %f8489, %f8493; add.ftz.f32 %f8495, %f8492, %f8494; add.ftz.f32 %f8496, %f8267, %f8495; sub.ftz.f32 %f8497, %f8276, %f8496; div.approx.ftz.f32 %f1496, %f8497, %f8268; @%p3 bra $L__BB108_1056; bra.uni $L__BB108_1053; $L__BB108_1056: shr.s32 %r1697, %r502, 31; and.b32 %r1698, %r1697, 1078530011; or.b32 %r1699, %r1698, %r503; mov.b32 %f16005, %r1699; bra.uni $L__BB108_1057; $L__BB108_1044: ld.const.u32 %r3058, [kernel_params+20]; setp.eq.s32 %p927, %r3058, 0; @%p927 bra $L__BB108_1080; ld.const.u32 %r3059, [kernel_params+20]; setp.eq.s32 %p928, %r3059, 1; @%p928 bra $L__BB108_1046; bra.uni $L__BB108_1087; $L__BB108_1046: ld.const.f32 %f1521, [kernel_params+24]; mul.ftz.f32 %f8597, %f1490, %f1490; fma.rn.ftz.f32 %f8598, %f1489, %f1489, %f8597; sqrt.approx.ftz.f32 %f8599, %f8598; abs.ftz.f32 %f1522, %f8599; abs.ftz.f32 %f1523, %f1488; setp.eq.ftz.f32 %p957, %f1523, 0f00000000; setp.eq.ftz.f32 %p958, %f1522, 0f00000000; and.pred %p959, %p957, %p958; mov.b32 %r506, %f1488; mov.b32 %r1714, %f8599; and.b32 %r507, %r1714, -2147483648; @%p959 bra $L__BB108_1073; bra.uni $L__BB108_1047; $L__BB108_1073: shr.s32 %r1719, %r506, 31; and.b32 %r1720, %r1719, 1078530011; or.b32 %r1721, %r507, %r1720; mov.b32 %f16010, %r1721; bra.uni $L__BB108_1074; $L__BB108_1058: add.ftz.f32 %f8530, %f1489, 0fBF800000; mul.ftz.f32 %f8531, %f8530, 0fBF000000; mov.f32 %f8532, 0f00000000; max.ftz.f32 %f8533, %f8531, %f8532; sqrt.approx.ftz.f32 %f8534, %f8533; add.ftz.f32 %f1503, %f8534, %f8534; setp.leu.ftz.f32 %p936, %f1503, 0f00000000; @%p936 bra $L__BB108_1060; rcp.approx.ftz.f32 %f8535, %f1503; mul.ftz.f32 %f1488, %f1488, %f8535; mul.ftz.f32 %f1490, %f1490, %f8535; $L__BB108_1060: add.ftz.f32 %f8536, %f1488, 0f3F800000; add.ftz.f32 %f8537, %f1490, 0f3F800000; mul.ftz.f32 %f16015, %f8537, 0f3F000000; mul.ftz.f32 %f16014, %f8536, 0f3F000000; bra.uni $L__BB108_1093; $L__BB108_1061: abs.ftz.f32 %f1510, %f1488; setp.eq.ftz.f32 %p938, %f1510, 0f00000000; and.pred %p939, %p938, %p921; mov.b32 %r504, %f1488; and.b32 %r505, %r502, -2147483648; @%p939 bra $L__BB108_1065; bra.uni $L__BB108_1062; $L__BB108_1065: shr.s32 %r1704, %r504, 31; and.b32 %r1705, %r1704, 1078530011; or.b32 %r1706, %r1705, %r505; mov.b32 %f16008, %r1706; bra.uni $L__BB108_1066; $L__BB108_1080: setp.eq.ftz.f32 %p974, %f1489, 0f00000000; setp.eq.ftz.f32 %p975, %f1488, 0f00000000; and.pred %p976, %p975, %p974; setp.eq.ftz.f32 %p977, %f1490, 0f00000000; and.pred %p978, %p977, %p976; mov.f32 %f16014, 0f00000000; mov.f32 %f16015, %f16014; @%p978 bra $L__BB108_1093; ld.const.v4.f32 {%f8656, %f8657, %f8658, %f8659}, [kernel_params+32]; abs.ftz.f32 %f1539, %f1488; setp.eq.ftz.f32 %p980, %f1539, 0f00000000; and.pred %p981, %p980, %p921; mov.b32 %r508, %f1488; and.b32 %r509, %r502, -2147483648; @%p981 bra $L__BB108_1085; bra.uni $L__BB108_1082; $L__BB108_1085: shr.s32 %r1733, %r508, 31; and.b32 %r1734, %r1733, 1078530011; or.b32 %r1735, %r1734, %r509; mov.b32 %f16012, %r1735; bra.uni $L__BB108_1086; $L__BB108_1087: ld.const.f32 %f8718, [kernel_params+28]; ld.const.f32 %f1546, [kernel_params+276]; ld.const.f32 %f1547, [kernel_params+280]; add.ftz.f32 %f8719, %f8718, %f8718; mul.ftz.f32 %f8720, %f1491, 0f3F000000; sin.approx.ftz.f32 %f8721, %f8720; mul.ftz.f32 %f1548, %f8719, %f8721; @%p3 bra $L__BB108_1091; bra.uni $L__BB108_1088; $L__BB108_1091: shr.s32 %r1740, %r502, 31; and.b32 %r1741, %r1740, 1078530011; or.b32 %r1742, %r1741, %r503; mov.b32 %f16013, %r1742; bra.uni $L__BB108_1092; $L__BB108_1053: setp.eq.ftz.f32 %p929, %f1492, 0f7F800000; setp.eq.ftz.f32 %p930, %f1493, 0f7F800000; and.pred %p931, %p929, %p930; @%p931 bra $L__BB108_1055; bra.uni $L__BB108_1054; $L__BB108_1055: setp.lt.s32 %p935, %r502, 0; selp.b32 %r1695, 1075235812, 1061752795, %p935; or.b32 %r1696, %r1695, %r503; mov.b32 %f16005, %r1696; bra.uni $L__BB108_1057; $L__BB108_1047: setp.eq.ftz.f32 %p960, %f1523, 0f7F800000; setp.eq.ftz.f32 %p961, %f1522, 0f7F800000; and.pred %p962, %p960, %p961; @%p962 bra $L__BB108_1072; bra.uni $L__BB108_1048; $L__BB108_1072: setp.lt.s32 %p966, %r506, 0; selp.b32 %r1717, 1075235812, 1061752795, %p966; or.b32 %r1718, %r507, %r1717; mov.b32 %f16010, %r1718; bra.uni $L__BB108_1074; $L__BB108_1062: setp.eq.ftz.f32 %p940, %f1510, 0f7F800000; setp.eq.ftz.f32 %p941, %f1492, 0f7F800000; and.pred %p942, %p940, %p941; @%p942 bra $L__BB108_1064; bra.uni $L__BB108_1063; $L__BB108_1064: setp.lt.s32 %p946, %r504, 0; selp.b32 %r1702, 1075235812, 1061752795, %p946; or.b32 %r1703, %r1702, %r505; mov.b32 %f16008, %r1703; bra.uni $L__BB108_1066; $L__BB108_1088: setp.eq.ftz.f32 %p991, %f1492, 0f7F800000; setp.eq.ftz.f32 %p992, %f1493, 0f7F800000; and.pred %p993, %p991, %p992; @%p993 bra $L__BB108_1090; bra.uni $L__BB108_1089; $L__BB108_1090: setp.lt.s32 %p997, %r502, 0; selp.b32 %r1738, 1075235812, 1061752795, %p997; or.b32 %r1739, %r1738, %r503; mov.b32 %f16013, %r1739; bra.uni $L__BB108_1092; $L__BB108_1082: setp.eq.ftz.f32 %p982, %f1539, 0f7F800000; setp.eq.ftz.f32 %p983, %f1492, 0f7F800000; and.pred %p984, %p982, %p983; @%p984 bra $L__BB108_1084; bra.uni $L__BB108_1083; $L__BB108_1084: setp.lt.s32 %p988, %r508, 0; selp.b32 %r1731, 1075235812, 1061752795, %p988; or.b32 %r1732, %r1731, %r509; mov.b32 %f16012, %r1732; bra.uni $L__BB108_1086; $L__BB108_1054: mov.f32 %f15494, 0f3FC90FDB; setp.lt.s32 %p932, %r502, 0; min.ftz.f32 %f8498, %f1493, %f1492; max.ftz.f32 %f8499, %f1493, %f1492; div.full.ftz.f32 %f8500, %f8498, %f8499; mul.rn.ftz.f32 %f8501, %f8500, %f8500; mov.f32 %f8502, 0fC0B59883; mov.f32 %f8503, 0fBF52C7EA; fma.rn.ftz.f32 %f8504, %f8501, %f8503, %f8502; mov.f32 %f8505, 0fC0D21907; fma.rn.ftz.f32 %f8506, %f8504, %f8501, %f8505; mul.ftz.f32 %f8507, %f8501, %f8506; mul.ftz.f32 %f8508, %f8500, %f8507; add.ftz.f32 %f8509, %f8501, 0f41355DC0; mov.f32 %f8510, 0f41E6BD60; fma.rn.ftz.f32 %f8511, %f8509, %f8501, %f8510; mov.f32 %f8512, 0f419D92C8; fma.rn.ftz.f32 %f8513, %f8511, %f8501, %f8512; rcp.approx.ftz.f32 %f8514, %f8513; fma.rn.ftz.f32 %f8515, %f8508, %f8514, %f8500; sub.ftz.f32 %f8517, %f15494, %f8515; setp.gt.ftz.f32 %p933, %f1493, %f1492; selp.f32 %f8518, %f8517, %f8515, %p933; sub.ftz.f32 %f8520, %f7711, %f8518; selp.f32 %f8521, %f8520, %f8518, %p932; mov.b32 %r1693, %f8521; or.b32 %r1694, %r503, %r1693; mov.b32 %f8522, %r1694; add.ftz.f32 %f8523, %f1492, %f1493; setp.le.ftz.f32 %p934, %f8523, 0f7F800000; selp.f32 %f16005, %f8522, %f8523, %p934; $L__BB108_1057: ld.const.f32 %f15496, [kernel_params+280]; ld.const.f32 %f15495, [kernel_params+276]; cos.approx.ftz.f32 %f8524, %f16005; mul.ftz.f32 %f8525, %f1496, %f8524; div.approx.ftz.f32 %f8526, %f8525, %f15495; sin.approx.ftz.f32 %f8527, %f16005; mul.ftz.f32 %f8528, %f1496, %f8527; div.approx.ftz.f32 %f8529, %f8528, %f15496; add.ftz.f32 %f16015, %f8529, 0f3F000000; add.ftz.f32 %f16014, %f8526, 0f3F000000; bra.uni $L__BB108_1093; $L__BB108_1048: mov.f32 %f15521, 0f3FC90FDB; setp.lt.s32 %p963, %r506, 0; min.ftz.f32 %f8600, %f1522, %f1523; max.ftz.f32 %f8601, %f1522, %f1523; div.full.ftz.f32 %f8602, %f8600, %f8601; mul.rn.ftz.f32 %f8603, %f8602, %f8602; mov.f32 %f8604, 0fC0B59883; mov.f32 %f8605, 0fBF52C7EA; fma.rn.ftz.f32 %f8606, %f8603, %f8605, %f8604; mov.f32 %f8607, 0fC0D21907; fma.rn.ftz.f32 %f8608, %f8606, %f8603, %f8607; mul.ftz.f32 %f8609, %f8603, %f8608; mul.ftz.f32 %f8610, %f8602, %f8609; add.ftz.f32 %f8611, %f8603, 0f41355DC0; mov.f32 %f8612, 0f41E6BD60; fma.rn.ftz.f32 %f8613, %f8611, %f8603, %f8612; mov.f32 %f8614, 0f419D92C8; fma.rn.ftz.f32 %f8615, %f8613, %f8603, %f8614; rcp.approx.ftz.f32 %f8616, %f8615; fma.rn.ftz.f32 %f8617, %f8610, %f8616, %f8602; sub.ftz.f32 %f8619, %f15521, %f8617; setp.gt.ftz.f32 %p964, %f1522, %f1523; selp.f32 %f8620, %f8619, %f8617, %p964; sub.ftz.f32 %f8622, %f7711, %f8620; selp.f32 %f8623, %f8622, %f8620, %p963; mov.b32 %r1715, %f8623; or.b32 %r1716, %r507, %r1715; mov.b32 %f8624, %r1716; add.ftz.f32 %f8625, %f1523, %f1522; setp.le.ftz.f32 %p965, %f8625, 0f7F800000; selp.f32 %f16010, %f8624, %f8625, %p965; $L__BB108_1074: div.approx.ftz.f32 %f1528, %f16010, %f1521; @%p3 bra $L__BB108_1078; bra.uni $L__BB108_1075; $L__BB108_1078: shr.s32 %r1726, %r502, 31; and.b32 %r1727, %r1726, 1078530011; or.b32 %r1728, %r1727, %r503; mov.b32 %f16011, %r1728; bra.uni $L__BB108_1079; $L__BB108_1075: setp.eq.ftz.f32 %p967, %f1492, 0f7F800000; setp.eq.ftz.f32 %p968, %f1493, 0f7F800000; and.pred %p969, %p967, %p968; @%p969 bra $L__BB108_1077; bra.uni $L__BB108_1076; $L__BB108_1077: setp.lt.s32 %p973, %r502, 0; selp.b32 %r1724, 1075235812, 1061752795, %p973; or.b32 %r1725, %r1724, %r503; mov.b32 %f16011, %r1725; bra.uni $L__BB108_1079; $L__BB108_1063: mov.f32 %f15519, 0f3FC90FDB; setp.lt.s32 %p943, %r504, 0; min.ftz.f32 %f8538, %f1492, %f1510; max.ftz.f32 %f8539, %f1492, %f1510; div.full.ftz.f32 %f8540, %f8538, %f8539; mul.rn.ftz.f32 %f8541, %f8540, %f8540; mov.f32 %f8542, 0fC0B59883; mov.f32 %f8543, 0fBF52C7EA; fma.rn.ftz.f32 %f8544, %f8541, %f8543, %f8542; mov.f32 %f8545, 0fC0D21907; fma.rn.ftz.f32 %f8546, %f8544, %f8541, %f8545; mul.ftz.f32 %f8547, %f8541, %f8546; mul.ftz.f32 %f8548, %f8540, %f8547; add.ftz.f32 %f8549, %f8541, 0f41355DC0; mov.f32 %f8550, 0f41E6BD60; fma.rn.ftz.f32 %f8551, %f8549, %f8541, %f8550; mov.f32 %f8552, 0f419D92C8; fma.rn.ftz.f32 %f8553, %f8551, %f8541, %f8552; rcp.approx.ftz.f32 %f8554, %f8553; fma.rn.ftz.f32 %f8555, %f8548, %f8554, %f8540; sub.ftz.f32 %f8557, %f15519, %f8555; setp.gt.ftz.f32 %p944, %f1492, %f1510; selp.f32 %f8558, %f8557, %f8555, %p944; sub.ftz.f32 %f8560, %f7711, %f8558; selp.f32 %f8561, %f8560, %f8558, %p943; mov.b32 %r1700, %f8561; or.b32 %r1701, %r505, %r1700; mov.b32 %f8562, %r1701; add.ftz.f32 %f8563, %f1510, %f1492; setp.le.ftz.f32 %p945, %f8563, 0f7F800000; selp.f32 %f16008, %f8562, %f8563, %p945; $L__BB108_1066: and.pred %p949, %p938, %p922; @%p949 bra $L__BB108_1070; bra.uni $L__BB108_1067; $L__BB108_1070: shr.s32 %r1711, %r504, 31; and.b32 %r1712, %r1711, 1078530011; or.b32 %r1713, %r503, %r1712; mov.b32 %f16009, %r1713; bra.uni $L__BB108_1071; $L__BB108_1067: setp.eq.ftz.f32 %p950, %f1510, 0f7F800000; setp.eq.ftz.f32 %p951, %f1493, 0f7F800000; and.pred %p952, %p950, %p951; @%p952 bra $L__BB108_1069; bra.uni $L__BB108_1068; $L__BB108_1069: setp.lt.s32 %p956, %r504, 0; selp.b32 %r1709, 1075235812, 1061752795, %p956; or.b32 %r1710, %r503, %r1709; mov.b32 %f16009, %r1710; bra.uni $L__BB108_1071; $L__BB108_1089: mov.f32 %f15518, 0f3FC90FDB; setp.lt.s32 %p994, %r502, 0; min.ftz.f32 %f8722, %f1493, %f1492; max.ftz.f32 %f8723, %f1493, %f1492; div.full.ftz.f32 %f8724, %f8722, %f8723; mul.rn.ftz.f32 %f8725, %f8724, %f8724; mov.f32 %f8726, 0fC0B59883; mov.f32 %f8727, 0fBF52C7EA; fma.rn.ftz.f32 %f8728, %f8725, %f8727, %f8726; mov.f32 %f8729, 0fC0D21907; fma.rn.ftz.f32 %f8730, %f8728, %f8725, %f8729; mul.ftz.f32 %f8731, %f8725, %f8730; mul.ftz.f32 %f8732, %f8724, %f8731; add.ftz.f32 %f8733, %f8725, 0f41355DC0; mov.f32 %f8734, 0f41E6BD60; fma.rn.ftz.f32 %f8735, %f8733, %f8725, %f8734; mov.f32 %f8736, 0f419D92C8; fma.rn.ftz.f32 %f8737, %f8735, %f8725, %f8736; rcp.approx.ftz.f32 %f8738, %f8737; fma.rn.ftz.f32 %f8739, %f8732, %f8738, %f8724; sub.ftz.f32 %f8741, %f15518, %f8739; setp.gt.ftz.f32 %p995, %f1493, %f1492; selp.f32 %f8742, %f8741, %f8739, %p995; sub.ftz.f32 %f8744, %f7711, %f8742; selp.f32 %f8745, %f8744, %f8742, %p994; mov.b32 %r1736, %f8745; or.b32 %r1737, %r503, %r1736; mov.b32 %f8746, %r1737; add.ftz.f32 %f8747, %f1492, %f1493; setp.le.ftz.f32 %p996, %f8747, 0f7F800000; selp.f32 %f16013, %f8746, %f8747, %p996; $L__BB108_1092: cos.approx.ftz.f32 %f8748, %f16013; mul.ftz.f32 %f8749, %f1548, %f8748; div.approx.ftz.f32 %f8750, %f8749, %f1546; sin.approx.ftz.f32 %f8751, %f16013; mul.ftz.f32 %f8752, %f1548, %f8751; div.approx.ftz.f32 %f8753, %f8752, %f1547; add.ftz.f32 %f16015, %f8753, 0f3F000000; add.ftz.f32 %f16014, %f8750, 0f3F000000; bra.uni $L__BB108_1093; $L__BB108_1076: mov.f32 %f15522, 0f3FC90FDB; setp.lt.s32 %p970, %r502, 0; min.ftz.f32 %f8626, %f1493, %f1492; max.ftz.f32 %f8627, %f1493, %f1492; div.full.ftz.f32 %f8628, %f8626, %f8627; mul.rn.ftz.f32 %f8629, %f8628, %f8628; mov.f32 %f8630, 0fC0B59883; mov.f32 %f8631, 0fBF52C7EA; fma.rn.ftz.f32 %f8632, %f8629, %f8631, %f8630; mov.f32 %f8633, 0fC0D21907; fma.rn.ftz.f32 %f8634, %f8632, %f8629, %f8633; mul.ftz.f32 %f8635, %f8629, %f8634; mul.ftz.f32 %f8636, %f8628, %f8635; add.ftz.f32 %f8637, %f8629, 0f41355DC0; mov.f32 %f8638, 0f41E6BD60; fma.rn.ftz.f32 %f8639, %f8637, %f8629, %f8638; mov.f32 %f8640, 0f419D92C8; fma.rn.ftz.f32 %f8641, %f8639, %f8629, %f8640; rcp.approx.ftz.f32 %f8642, %f8641; fma.rn.ftz.f32 %f8643, %f8636, %f8642, %f8628; sub.ftz.f32 %f8645, %f15522, %f8643; setp.gt.ftz.f32 %p971, %f1493, %f1492; selp.f32 %f8646, %f8645, %f8643, %p971; sub.ftz.f32 %f8648, %f7711, %f8646; selp.f32 %f8649, %f8648, %f8646, %p970; mov.b32 %r1722, %f8649; or.b32 %r1723, %r503, %r1722; mov.b32 %f8650, %r1723; add.ftz.f32 %f8651, %f1492, %f1493; setp.le.ftz.f32 %p972, %f8651, 0f7F800000; selp.f32 %f16011, %f8650, %f8651, %p972; $L__BB108_1079: cos.approx.ftz.f32 %f8652, %f16011; sin.approx.ftz.f32 %f8653, %f16011; fma.rn.ftz.f32 %f16015, %f1528, %f8653, 0f3F000000; fma.rn.ftz.f32 %f16014, %f1528, %f8652, 0f3F000000; bra.uni $L__BB108_1093; $L__BB108_1068: mov.f32 %f15520, 0f3FC90FDB; setp.lt.s32 %p953, %r504, 0; min.ftz.f32 %f8564, %f1493, %f1510; max.ftz.f32 %f8565, %f1493, %f1510; div.full.ftz.f32 %f8566, %f8564, %f8565; mul.rn.ftz.f32 %f8567, %f8566, %f8566; mov.f32 %f8568, 0fC0B59883; mov.f32 %f8569, 0fBF52C7EA; fma.rn.ftz.f32 %f8570, %f8567, %f8569, %f8568; mov.f32 %f8571, 0fC0D21907; fma.rn.ftz.f32 %f8572, %f8570, %f8567, %f8571; mul.ftz.f32 %f8573, %f8567, %f8572; mul.ftz.f32 %f8574, %f8566, %f8573; add.ftz.f32 %f8575, %f8567, 0f41355DC0; mov.f32 %f8576, 0f41E6BD60; fma.rn.ftz.f32 %f8577, %f8575, %f8567, %f8576; mov.f32 %f8578, 0f419D92C8; fma.rn.ftz.f32 %f8579, %f8577, %f8567, %f8578; rcp.approx.ftz.f32 %f8580, %f8579; fma.rn.ftz.f32 %f8581, %f8574, %f8580, %f8566; sub.ftz.f32 %f8583, %f15520, %f8581; setp.gt.ftz.f32 %p954, %f1493, %f1510; selp.f32 %f8584, %f8583, %f8581, %p954; sub.ftz.f32 %f8586, %f7711, %f8584; selp.f32 %f8587, %f8586, %f8584, %p953; mov.b32 %r1707, %f8587; or.b32 %r1708, %r503, %r1707; mov.b32 %f8588, %r1708; add.ftz.f32 %f8589, %f1510, %f1493; setp.le.ftz.f32 %p955, %f8589, 0f7F800000; selp.f32 %f16009, %f8588, %f8589, %p955; $L__BB108_1071: mov.f32 %f15487, 0f3F800000; add.ftz.f32 %f8590, %f16009, %f16009; div.approx.ftz.f32 %f8592, %f8590, %f7711; add.ftz.f32 %f8593, %f16008, %f16008; div.approx.ftz.f32 %f8594, %f8593, %f7711; add.ftz.f32 %f8595, %f8594, 0f3F000000; sub.ftz.f32 %f16014, %f15487, %f8595; add.ftz.f32 %f16015, %f8592, 0f3F000000; bra.uni $L__BB108_1093; $L__BB108_1083: mov.f32 %f15523, 0f3FC90FDB; setp.lt.s32 %p985, %r508, 0; min.ftz.f32 %f8660, %f1492, %f1539; max.ftz.f32 %f8661, %f1492, %f1539; div.full.ftz.f32 %f8662, %f8660, %f8661; mul.rn.ftz.f32 %f8663, %f8662, %f8662; mov.f32 %f8664, 0fC0B59883; mov.f32 %f8665, 0fBF52C7EA; fma.rn.ftz.f32 %f8666, %f8663, %f8665, %f8664; mov.f32 %f8667, 0fC0D21907; fma.rn.ftz.f32 %f8668, %f8666, %f8663, %f8667; mul.ftz.f32 %f8669, %f8663, %f8668; mul.ftz.f32 %f8670, %f8662, %f8669; add.ftz.f32 %f8671, %f8663, 0f41355DC0; mov.f32 %f8672, 0f41E6BD60; fma.rn.ftz.f32 %f8673, %f8671, %f8663, %f8672; mov.f32 %f8674, 0f419D92C8; fma.rn.ftz.f32 %f8675, %f8673, %f8663, %f8674; rcp.approx.ftz.f32 %f8676, %f8675; fma.rn.ftz.f32 %f8677, %f8670, %f8676, %f8662; sub.ftz.f32 %f8679, %f15523, %f8677; setp.gt.ftz.f32 %p986, %f1492, %f1539; selp.f32 %f8680, %f8679, %f8677, %p986; sub.ftz.f32 %f8682, %f7711, %f8680; selp.f32 %f8683, %f8682, %f8680, %p985; mov.b32 %r1729, %f8683; or.b32 %r1730, %r509, %r1729; mov.b32 %f8684, %r1730; add.ftz.f32 %f8685, %f1539, %f1492; setp.le.ftz.f32 %p987, %f8685, 0f7F800000; selp.f32 %f16012, %f8684, %f8685, %p987; $L__BB108_1086: mov.f32 %f15524, 0f3FC90FDB; mov.f32 %f15493, 0f3E2AAB94; mov.f32 %f15492, 0f3D994929; mov.f32 %f15491, 0f3D3F841F; mov.f32 %f15490, 0f3C94D2E9; mov.f32 %f15489, 0f3D53F941; mov.f32 %f15488, 0f3F800000; sub.ftz.f32 %f8686, %f16012, %f8657; mul.ftz.f32 %f8687, %f1489, %f1489; fma.rn.ftz.f32 %f8688, %f1488, %f1488, %f8687; fma.rn.ftz.f32 %f8689, %f1490, %f1490, %f8688; sqrt.approx.ftz.f32 %f8690, %f8689; div.approx.ftz.f32 %f8691, %f1490, %f8690; abs.ftz.f32 %f8692, %f8691; sub.ftz.f32 %f8694, %f15488, %f8692; mul.ftz.f32 %f8695, %f8694, 0f3F000000; sqrt.approx.ftz.f32 %f8696, %f8695; setp.gt.ftz.f32 %p989, %f8692, 0f3F11EB85; selp.f32 %f8697, %f8696, %f8692, %p989; mul.ftz.f32 %f8698, %f8697, %f8697; fma.rn.ftz.f32 %f8701, %f15489, %f8698, %f15490; fma.rn.ftz.f32 %f8703, %f8701, %f8698, %f15491; fma.rn.ftz.f32 %f8705, %f8703, %f8698, %f15492; fma.rn.ftz.f32 %f8707, %f8705, %f8698, %f15493; mul.ftz.f32 %f8708, %f8698, %f8707; fma.rn.ftz.f32 %f8709, %f8708, %f8697, %f8697; add.ftz.f32 %f8710, %f8709, %f8709; sub.ftz.f32 %f8712, %f15524, %f8709; selp.f32 %f8713, %f8710, %f8712, %p989; setp.lt.ftz.f32 %p990, %f8691, 0f00000000; sub.ftz.f32 %f8715, %f7711, %f8713; selp.f32 %f8716, %f8715, %f8713, %p990; sub.ftz.f32 %f8717, %f8716, %f8659; div.approx.ftz.f32 %f16015, %f8717, %f8658; div.approx.ftz.f32 %f16014, %f8686, %f8656; $L__BB108_1093: ld.const.u32 %r3060, [kernel_params+20]; setp.gt.s32 %p2182, %r3060, 2; mov.f32 %f15497, 0f3FC90FDB; mov.f32 %f15479, 0fBF800000; mov.f32 %f15478, 0f3E2AAB94; mov.f32 %f15477, 0f3D994929; mov.f32 %f15476, 0f3D3F841F; mov.f32 %f15475, 0f3C94D2E9; mov.f32 %f15474, 0f3D53F941; mov.f32 %f15473, 0f3F800000; ld.const.v4.f32 {%f8754, %f8755, %f8756, %f8757}, [kernel_params+912]; ld.const.v4.f32 {%f8762, %f8763, %f8764, %f8765}, [kernel_params+928]; ld.const.v4.f32 {%f8770, %f8771, %f8772, %f8773}, [kernel_params+944]; mul.ftz.f32 %f8778, %f1379, %f8755; fma.rn.ftz.f32 %f8779, %f1378, %f8754, %f8778; fma.rn.ftz.f32 %f8780, %f1380, %f8756, %f8779; add.ftz.f32 %f8781, %f8757, %f8780; mul.ftz.f32 %f8782, %f1379, %f8763; fma.rn.ftz.f32 %f8783, %f1378, %f8762, %f8782; fma.rn.ftz.f32 %f8784, %f1380, %f8764, %f8783; add.ftz.f32 %f8785, %f8765, %f8784; mul.ftz.f32 %f8786, %f1379, %f8771; fma.rn.ftz.f32 %f8787, %f1378, %f8770, %f8786; fma.rn.ftz.f32 %f8788, %f1380, %f8772, %f8787; add.ftz.f32 %f8789, %f8773, %f8788; mul.ftz.f32 %f8790, %f8785, %f8785; fma.rn.ftz.f32 %f8791, %f8781, %f8781, %f8790; fma.rn.ftz.f32 %f8792, %f8789, %f8789, %f8791; sqrt.approx.ftz.f32 %f8793, %f8792; div.approx.ftz.f32 %f1557, %f8781, %f8793; div.approx.ftz.f32 %f1558, %f8785, %f8793; div.approx.ftz.f32 %f1559, %f8789, %f8793; max.ftz.f32 %f8795, %f1557, %f15479; min.ftz.f32 %f8797, %f8795, %f15473; abs.ftz.f32 %f8798, %f8797; sub.ftz.f32 %f8799, %f15473, %f8798; mul.ftz.f32 %f8800, %f8799, 0f3F000000; sqrt.approx.ftz.f32 %f8801, %f8800; setp.gt.ftz.f32 %p998, %f8798, 0f3F11EB85; selp.f32 %f8802, %f8801, %f8798, %p998; mul.ftz.f32 %f8803, %f8802, %f8802; fma.rn.ftz.f32 %f8806, %f15474, %f8803, %f15475; fma.rn.ftz.f32 %f8808, %f8806, %f8803, %f15476; fma.rn.ftz.f32 %f8810, %f8808, %f8803, %f15477; fma.rn.ftz.f32 %f8812, %f8810, %f8803, %f15478; mul.ftz.f32 %f8813, %f8803, %f8812; fma.rn.ftz.f32 %f8814, %f8813, %f8802, %f8802; add.ftz.f32 %f8815, %f8814, %f8814; sub.ftz.f32 %f8817, %f15497, %f8814; selp.f32 %f8818, %f8815, %f8817, %p998; setp.lt.ftz.f32 %p999, %f8797, 0f00000000; sub.ftz.f32 %f8820, %f7711, %f8818; selp.f32 %f1560, %f8820, %f8818, %p999; abs.ftz.f32 %f1561, %f1558; abs.ftz.f32 %f1562, %f1559; setp.eq.ftz.f32 %p1000, %f1561, 0f00000000; setp.eq.ftz.f32 %p1001, %f1562, 0f00000000; and.pred %p4, %p1000, %p1001; mov.b32 %r510, %f1558; mov.b32 %r1743, %f1559; and.b32 %r511, %r1743, -2147483648; @%p2182 bra $L__BB108_1099; bra.uni $L__BB108_1094; $L__BB108_1099: ld.const.u32 %r3049, [kernel_params+20]; setp.eq.s32 %p1003, %r3049, 3; @%p1003 bra $L__BB108_1108; ld.const.u32 %r3050, [kernel_params+20]; setp.eq.s32 %p1004, %r3050, 5; @%p1004 bra $L__BB108_1111; ld.const.u32 %r3051, [kernel_params+20]; setp.ne.s32 %p1005, %r3051, 4; @%p1005 bra $L__BB108_1137; ld.const.f32 %f8821, [kernel_params+48]; ld.const.v4.f32 {%f8822, %f8823, %f8824, %f8825}, [kernel_params+64]; neg.ftz.f32 %f8830, %f1560; sub.ftz.f32 %f8831, %f8830, %f8821; div.approx.ftz.f32 %f8832, %f8831, %f8822; mul.ftz.f32 %f8833, %f8832, %f8832; mul.ftz.f32 %f8834, %f8832, %f8833; mul.ftz.f32 %f8835, %f8833, %f8833; mul.ftz.f32 %f8836, %f8823, %f8833; fma.rn.ftz.f32 %f8837, %f8822, %f8832, %f8836; mul.ftz.f32 %f8838, %f8825, %f8835; fma.rn.ftz.f32 %f8839, %f8824, %f8834, %f8838; add.ftz.f32 %f8840, %f8837, %f8839; add.ftz.f32 %f8841, %f8821, %f8840; sub.ftz.f32 %f8842, %f8830, %f8841; div.approx.ftz.f32 %f8843, %f8842, %f8822; mul.ftz.f32 %f8844, %f8843, %f8843; mul.ftz.f32 %f8845, %f8843, %f8844; mul.ftz.f32 %f8846, %f8844, %f8844; mul.ftz.f32 %f8847, %f8823, %f8844; fma.rn.ftz.f32 %f8848, %f8822, %f8843, %f8847; mul.ftz.f32 %f8849, %f8825, %f8846; fma.rn.ftz.f32 %f8850, %f8824, %f8845, %f8849; add.ftz.f32 %f8851, %f8848, %f8850; add.ftz.f32 %f8852, %f8821, %f8851; sub.ftz.f32 %f8853, %f8830, %f8852; div.approx.ftz.f32 %f8854, %f8853, %f8822; mul.ftz.f32 %f8855, %f8854, %f8854; mul.ftz.f32 %f8856, %f8854, %f8855; mul.ftz.f32 %f8857, %f8855, %f8855; mul.ftz.f32 %f8858, %f8823, %f8855; fma.rn.ftz.f32 %f8859, %f8822, %f8854, %f8858; mul.ftz.f32 %f8860, %f8825, %f8857; fma.rn.ftz.f32 %f8861, %f8824, %f8856, %f8860; add.ftz.f32 %f8862, %f8859, %f8861; add.ftz.f32 %f8863, %f8821, %f8862; sub.ftz.f32 %f8864, %f8830, %f8863; div.approx.ftz.f32 %f8865, %f8864, %f8822; mul.ftz.f32 %f8866, %f8865, %f8865; mul.ftz.f32 %f8867, %f8865, %f8866; mul.ftz.f32 %f8868, %f8866, %f8866; mul.ftz.f32 %f8869, %f8823, %f8866; fma.rn.ftz.f32 %f8870, %f8822, %f8865, %f8869; mul.ftz.f32 %f8871, %f8825, %f8868; fma.rn.ftz.f32 %f8872, %f8824, %f8867, %f8871; add.ftz.f32 %f8873, %f8870, %f8872; add.ftz.f32 %f8874, %f8821, %f8873; sub.ftz.f32 %f8875, %f8830, %f8874; div.approx.ftz.f32 %f8876, %f8875, %f8822; mul.ftz.f32 %f8877, %f8876, %f8876; mul.ftz.f32 %f8878, %f8876, %f8877; mul.ftz.f32 %f8879, %f8877, %f8877; mul.ftz.f32 %f8880, %f8823, %f8877; fma.rn.ftz.f32 %f8881, %f8822, %f8876, %f8880; mul.ftz.f32 %f8882, %f8825, %f8879; fma.rn.ftz.f32 %f8883, %f8824, %f8878, %f8882; add.ftz.f32 %f8884, %f8881, %f8883; add.ftz.f32 %f8885, %f8821, %f8884; sub.ftz.f32 %f8886, %f8830, %f8885; div.approx.ftz.f32 %f8887, %f8886, %f8822; mul.ftz.f32 %f8888, %f8887, %f8887; mul.ftz.f32 %f8889, %f8887, %f8888; mul.ftz.f32 %f8890, %f8888, %f8888; mul.ftz.f32 %f8891, %f8823, %f8888; fma.rn.ftz.f32 %f8892, %f8822, %f8887, %f8891; mul.ftz.f32 %f8893, %f8825, %f8890; fma.rn.ftz.f32 %f8894, %f8824, %f8889, %f8893; add.ftz.f32 %f8895, %f8892, %f8894; add.ftz.f32 %f8896, %f8821, %f8895; sub.ftz.f32 %f8897, %f8830, %f8896; div.approx.ftz.f32 %f8898, %f8897, %f8822; mul.ftz.f32 %f8899, %f8898, %f8898; mul.ftz.f32 %f8900, %f8898, %f8899; mul.ftz.f32 %f8901, %f8899, %f8899; mul.ftz.f32 %f8902, %f8823, %f8899; fma.rn.ftz.f32 %f8903, %f8822, %f8898, %f8902; mul.ftz.f32 %f8904, %f8825, %f8901; fma.rn.ftz.f32 %f8905, %f8824, %f8900, %f8904; add.ftz.f32 %f8906, %f8903, %f8905; add.ftz.f32 %f8907, %f8821, %f8906; sub.ftz.f32 %f8908, %f8830, %f8907; div.approx.ftz.f32 %f8909, %f8908, %f8822; mul.ftz.f32 %f8910, %f8909, %f8909; mul.ftz.f32 %f8911, %f8909, %f8910; mul.ftz.f32 %f8912, %f8910, %f8910; mul.ftz.f32 %f8913, %f8823, %f8910; fma.rn.ftz.f32 %f8914, %f8822, %f8909, %f8913; mul.ftz.f32 %f8915, %f8825, %f8912; fma.rn.ftz.f32 %f8916, %f8824, %f8911, %f8915; add.ftz.f32 %f8917, %f8914, %f8916; add.ftz.f32 %f8918, %f8821, %f8917; sub.ftz.f32 %f8919, %f8830, %f8918; div.approx.ftz.f32 %f8920, %f8919, %f8822; mul.ftz.f32 %f8921, %f8920, %f8920; mul.ftz.f32 %f8922, %f8920, %f8921; mul.ftz.f32 %f8923, %f8921, %f8921; mul.ftz.f32 %f8924, %f8823, %f8921; fma.rn.ftz.f32 %f8925, %f8822, %f8920, %f8924; mul.ftz.f32 %f8926, %f8825, %f8923; fma.rn.ftz.f32 %f8927, %f8824, %f8922, %f8926; add.ftz.f32 %f8928, %f8925, %f8927; add.ftz.f32 %f8929, %f8821, %f8928; sub.ftz.f32 %f8930, %f8830, %f8929; div.approx.ftz.f32 %f8931, %f8930, %f8822; mul.ftz.f32 %f8932, %f8931, %f8931; mul.ftz.f32 %f8933, %f8931, %f8932; mul.ftz.f32 %f8934, %f8932, %f8932; mul.ftz.f32 %f8935, %f8823, %f8932; fma.rn.ftz.f32 %f8936, %f8822, %f8931, %f8935; mul.ftz.f32 %f8937, %f8825, %f8934; fma.rn.ftz.f32 %f8938, %f8824, %f8933, %f8937; add.ftz.f32 %f8939, %f8936, %f8938; add.ftz.f32 %f8940, %f8821, %f8939; sub.ftz.f32 %f8941, %f8830, %f8940; div.approx.ftz.f32 %f8942, %f8941, %f8822; mul.ftz.f32 %f8943, %f8942, %f8942; mul.ftz.f32 %f8944, %f8942, %f8943; mul.ftz.f32 %f8945, %f8943, %f8943; mul.ftz.f32 %f8946, %f8823, %f8943; fma.rn.ftz.f32 %f8947, %f8822, %f8942, %f8946; mul.ftz.f32 %f8948, %f8825, %f8945; fma.rn.ftz.f32 %f8949, %f8824, %f8944, %f8948; add.ftz.f32 %f8950, %f8947, %f8949; add.ftz.f32 %f8951, %f8821, %f8950; sub.ftz.f32 %f8952, %f8830, %f8951; div.approx.ftz.f32 %f8953, %f8952, %f8822; mul.ftz.f32 %f8954, %f8953, %f8953; mul.ftz.f32 %f8955, %f8953, %f8954; mul.ftz.f32 %f8956, %f8954, %f8954; mul.ftz.f32 %f8957, %f8823, %f8954; fma.rn.ftz.f32 %f8958, %f8822, %f8953, %f8957; mul.ftz.f32 %f8959, %f8825, %f8956; fma.rn.ftz.f32 %f8960, %f8824, %f8955, %f8959; add.ftz.f32 %f8961, %f8958, %f8960; add.ftz.f32 %f8962, %f8821, %f8961; sub.ftz.f32 %f8963, %f8830, %f8962; div.approx.ftz.f32 %f8964, %f8963, %f8822; mul.ftz.f32 %f8965, %f8964, %f8964; mul.ftz.f32 %f8966, %f8964, %f8965; mul.ftz.f32 %f8967, %f8965, %f8965; mul.ftz.f32 %f8968, %f8823, %f8965; fma.rn.ftz.f32 %f8969, %f8822, %f8964, %f8968; mul.ftz.f32 %f8970, %f8825, %f8967; fma.rn.ftz.f32 %f8971, %f8824, %f8966, %f8970; add.ftz.f32 %f8972, %f8969, %f8971; add.ftz.f32 %f8973, %f8821, %f8972; sub.ftz.f32 %f8974, %f8830, %f8973; div.approx.ftz.f32 %f8975, %f8974, %f8822; mul.ftz.f32 %f8976, %f8975, %f8975; mul.ftz.f32 %f8977, %f8975, %f8976; mul.ftz.f32 %f8978, %f8976, %f8976; mul.ftz.f32 %f8979, %f8823, %f8976; fma.rn.ftz.f32 %f8980, %f8822, %f8975, %f8979; mul.ftz.f32 %f8981, %f8825, %f8978; fma.rn.ftz.f32 %f8982, %f8824, %f8977, %f8981; add.ftz.f32 %f8983, %f8980, %f8982; add.ftz.f32 %f8984, %f8821, %f8983; sub.ftz.f32 %f8985, %f8830, %f8984; div.approx.ftz.f32 %f8986, %f8985, %f8822; mul.ftz.f32 %f8987, %f8986, %f8986; mul.ftz.f32 %f8988, %f8986, %f8987; mul.ftz.f32 %f8989, %f8987, %f8987; mul.ftz.f32 %f8990, %f8823, %f8987; fma.rn.ftz.f32 %f8991, %f8822, %f8986, %f8990; mul.ftz.f32 %f8992, %f8825, %f8989; fma.rn.ftz.f32 %f8993, %f8824, %f8988, %f8992; add.ftz.f32 %f8994, %f8991, %f8993; add.ftz.f32 %f8995, %f8821, %f8994; sub.ftz.f32 %f8996, %f8830, %f8995; div.approx.ftz.f32 %f8997, %f8996, %f8822; mul.ftz.f32 %f8998, %f8997, %f8997; mul.ftz.f32 %f8999, %f8997, %f8998; mul.ftz.f32 %f9000, %f8998, %f8998; mul.ftz.f32 %f9001, %f8823, %f8998; fma.rn.ftz.f32 %f9002, %f8822, %f8997, %f9001; mul.ftz.f32 %f9003, %f8825, %f9000; fma.rn.ftz.f32 %f9004, %f8824, %f8999, %f9003; add.ftz.f32 %f9005, %f9002, %f9004; add.ftz.f32 %f9006, %f8821, %f9005; sub.ftz.f32 %f9007, %f8830, %f9006; div.approx.ftz.f32 %f9008, %f9007, %f8822; mul.ftz.f32 %f9009, %f9008, %f9008; mul.ftz.f32 %f9010, %f9008, %f9009; mul.ftz.f32 %f9011, %f9009, %f9009; mul.ftz.f32 %f9012, %f8823, %f9009; fma.rn.ftz.f32 %f9013, %f8822, %f9008, %f9012; mul.ftz.f32 %f9014, %f8825, %f9011; fma.rn.ftz.f32 %f9015, %f8824, %f9010, %f9014; add.ftz.f32 %f9016, %f9013, %f9015; add.ftz.f32 %f9017, %f8821, %f9016; sub.ftz.f32 %f9018, %f8830, %f9017; div.approx.ftz.f32 %f9019, %f9018, %f8822; mul.ftz.f32 %f9020, %f9019, %f9019; mul.ftz.f32 %f9021, %f9019, %f9020; mul.ftz.f32 %f9022, %f9020, %f9020; mul.ftz.f32 %f9023, %f8823, %f9020; fma.rn.ftz.f32 %f9024, %f8822, %f9019, %f9023; mul.ftz.f32 %f9025, %f8825, %f9022; fma.rn.ftz.f32 %f9026, %f8824, %f9021, %f9025; add.ftz.f32 %f9027, %f9024, %f9026; add.ftz.f32 %f9028, %f8821, %f9027; sub.ftz.f32 %f9029, %f8830, %f9028; div.approx.ftz.f32 %f9030, %f9029, %f8822; mul.ftz.f32 %f9031, %f9030, %f9030; mul.ftz.f32 %f9032, %f9030, %f9031; mul.ftz.f32 %f9033, %f9031, %f9031; mul.ftz.f32 %f9034, %f8823, %f9031; fma.rn.ftz.f32 %f9035, %f8822, %f9030, %f9034; mul.ftz.f32 %f9036, %f8825, %f9033; fma.rn.ftz.f32 %f9037, %f8824, %f9032, %f9036; add.ftz.f32 %f9038, %f9035, %f9037; add.ftz.f32 %f9039, %f8821, %f9038; sub.ftz.f32 %f9040, %f8830, %f9039; div.approx.ftz.f32 %f9041, %f9040, %f8822; mul.ftz.f32 %f9042, %f9041, %f9041; mul.ftz.f32 %f9043, %f9041, %f9042; mul.ftz.f32 %f9044, %f9042, %f9042; mul.ftz.f32 %f9045, %f8823, %f9042; fma.rn.ftz.f32 %f9046, %f8822, %f9041, %f9045; mul.ftz.f32 %f9047, %f8825, %f9044; fma.rn.ftz.f32 %f9048, %f8824, %f9043, %f9047; add.ftz.f32 %f9049, %f9046, %f9048; add.ftz.f32 %f9050, %f8821, %f9049; sub.ftz.f32 %f9051, %f8830, %f9050; div.approx.ftz.f32 %f1565, %f9051, %f8822; @%p4 bra $L__BB108_1106; bra.uni $L__BB108_1103; $L__BB108_1106: shr.s32 %r1748, %r510, 31; and.b32 %r1749, %r1748, 1078530011; or.b32 %r1750, %r1749, %r511; mov.b32 %f16016, %r1750; bra.uni $L__BB108_1107; $L__BB108_1094: ld.const.u32 %r3056, [kernel_params+20]; setp.eq.s32 %p1006, %r3056, 0; @%p1006 bra $L__BB108_1130; ld.const.u32 %r3057, [kernel_params+20]; setp.eq.s32 %p1007, %r3057, 1; @%p1007 bra $L__BB108_1096; bra.uni $L__BB108_1137; $L__BB108_1096: ld.const.f32 %f1590, [kernel_params+24]; mul.ftz.f32 %f9151, %f1559, %f1559; fma.rn.ftz.f32 %f9152, %f1558, %f1558, %f9151; sqrt.approx.ftz.f32 %f9153, %f9152; abs.ftz.f32 %f1591, %f9153; abs.ftz.f32 %f1592, %f1557; setp.eq.ftz.f32 %p1036, %f1592, 0f00000000; setp.eq.ftz.f32 %p1037, %f1591, 0f00000000; and.pred %p1038, %p1036, %p1037; mov.b32 %r514, %f1557; mov.b32 %r1765, %f9153; and.b32 %r515, %r1765, -2147483648; @%p1038 bra $L__BB108_1123; bra.uni $L__BB108_1097; $L__BB108_1123: shr.s32 %r1770, %r514, 31; and.b32 %r1771, %r1770, 1078530011; or.b32 %r1772, %r515, %r1771; mov.b32 %f16021, %r1772; bra.uni $L__BB108_1124; $L__BB108_1108: add.ftz.f32 %f9084, %f1558, 0fBF800000; mul.ftz.f32 %f9085, %f9084, 0fBF000000; mov.f32 %f9086, 0f00000000; max.ftz.f32 %f9087, %f9085, %f9086; sqrt.approx.ftz.f32 %f9088, %f9087; add.ftz.f32 %f1572, %f9088, %f9088; setp.leu.ftz.f32 %p1015, %f1572, 0f00000000; @%p1015 bra $L__BB108_1110; rcp.approx.ftz.f32 %f9089, %f1572; mul.ftz.f32 %f1557, %f1557, %f9089; mul.ftz.f32 %f1559, %f1559, %f9089; $L__BB108_1110: add.ftz.f32 %f9090, %f1557, 0f3F800000; add.ftz.f32 %f9091, %f1559, 0f3F800000; mul.ftz.f32 %f16026, %f9091, 0f3F000000; mul.ftz.f32 %f16025, %f9090, 0f3F000000; bra.uni $L__BB108_1143; $L__BB108_1111: abs.ftz.f32 %f1579, %f1557; setp.eq.ftz.f32 %p1017, %f1579, 0f00000000; and.pred %p1018, %p1017, %p1000; mov.b32 %r512, %f1557; and.b32 %r513, %r510, -2147483648; @%p1018 bra $L__BB108_1115; bra.uni $L__BB108_1112; $L__BB108_1115: shr.s32 %r1755, %r512, 31; and.b32 %r1756, %r1755, 1078530011; or.b32 %r1757, %r1756, %r513; mov.b32 %f16019, %r1757; bra.uni $L__BB108_1116; $L__BB108_1130: setp.eq.ftz.f32 %p1053, %f1558, 0f00000000; setp.eq.ftz.f32 %p1054, %f1557, 0f00000000; and.pred %p1055, %p1054, %p1053; setp.eq.ftz.f32 %p1056, %f1559, 0f00000000; and.pred %p1057, %p1056, %p1055; mov.f32 %f16025, 0f00000000; mov.f32 %f16026, %f16025; @%p1057 bra $L__BB108_1143; ld.const.v4.f32 {%f9210, %f9211, %f9212, %f9213}, [kernel_params+32]; abs.ftz.f32 %f1608, %f1557; setp.eq.ftz.f32 %p1059, %f1608, 0f00000000; and.pred %p1060, %p1059, %p1000; mov.b32 %r516, %f1557; and.b32 %r517, %r510, -2147483648; @%p1060 bra $L__BB108_1135; bra.uni $L__BB108_1132; $L__BB108_1135: shr.s32 %r1784, %r516, 31; and.b32 %r1785, %r1784, 1078530011; or.b32 %r1786, %r1785, %r517; mov.b32 %f16023, %r1786; bra.uni $L__BB108_1136; $L__BB108_1137: ld.const.f32 %f9272, [kernel_params+28]; ld.const.f32 %f1615, [kernel_params+276]; ld.const.f32 %f1616, [kernel_params+280]; add.ftz.f32 %f9273, %f9272, %f9272; mul.ftz.f32 %f9274, %f1560, 0f3F000000; sin.approx.ftz.f32 %f9275, %f9274; mul.ftz.f32 %f1617, %f9273, %f9275; @%p4 bra $L__BB108_1141; bra.uni $L__BB108_1138; $L__BB108_1141: shr.s32 %r1791, %r510, 31; and.b32 %r1792, %r1791, 1078530011; or.b32 %r1793, %r1792, %r511; mov.b32 %f16024, %r1793; bra.uni $L__BB108_1142; $L__BB108_1103: setp.eq.ftz.f32 %p1008, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1009, %f1562, 0f7F800000; and.pred %p1010, %p1008, %p1009; @%p1010 bra $L__BB108_1105; bra.uni $L__BB108_1104; $L__BB108_1105: setp.lt.s32 %p1014, %r510, 0; selp.b32 %r1746, 1075235812, 1061752795, %p1014; or.b32 %r1747, %r1746, %r511; mov.b32 %f16016, %r1747; bra.uni $L__BB108_1107; $L__BB108_1097: setp.eq.ftz.f32 %p1039, %f1592, 0f7F800000; setp.eq.ftz.f32 %p1040, %f1591, 0f7F800000; and.pred %p1041, %p1039, %p1040; @%p1041 bra $L__BB108_1122; bra.uni $L__BB108_1098; $L__BB108_1122: setp.lt.s32 %p1045, %r514, 0; selp.b32 %r1768, 1075235812, 1061752795, %p1045; or.b32 %r1769, %r515, %r1768; mov.b32 %f16021, %r1769; bra.uni $L__BB108_1124; $L__BB108_1112: setp.eq.ftz.f32 %p1019, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1020, %f1561, 0f7F800000; and.pred %p1021, %p1019, %p1020; @%p1021 bra $L__BB108_1114; bra.uni $L__BB108_1113; $L__BB108_1114: setp.lt.s32 %p1025, %r512, 0; selp.b32 %r1753, 1075235812, 1061752795, %p1025; or.b32 %r1754, %r1753, %r513; mov.b32 %f16019, %r1754; bra.uni $L__BB108_1116; $L__BB108_1138: setp.eq.ftz.f32 %p1070, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1071, %f1562, 0f7F800000; and.pred %p1072, %p1070, %p1071; @%p1072 bra $L__BB108_1140; bra.uni $L__BB108_1139; $L__BB108_1140: setp.lt.s32 %p1076, %r510, 0; selp.b32 %r1789, 1075235812, 1061752795, %p1076; or.b32 %r1790, %r1789, %r511; mov.b32 %f16024, %r1790; bra.uni $L__BB108_1142; $L__BB108_1132: setp.eq.ftz.f32 %p1061, %f1608, 0f7F800000; setp.eq.ftz.f32 %p1062, %f1561, 0f7F800000; and.pred %p1063, %p1061, %p1062; @%p1063 bra $L__BB108_1134; bra.uni $L__BB108_1133; $L__BB108_1134: setp.lt.s32 %p1067, %r516, 0; selp.b32 %r1782, 1075235812, 1061752795, %p1067; or.b32 %r1783, %r1782, %r517; mov.b32 %f16023, %r1783; bra.uni $L__BB108_1136; $L__BB108_1104: mov.f32 %f15508, 0f40490FDB; mov.f32 %f15501, 0f3FC90FDB; setp.lt.s32 %p1011, %r510, 0; min.ftz.f32 %f9052, %f1562, %f1561; max.ftz.f32 %f9053, %f1562, %f1561; div.full.ftz.f32 %f9054, %f9052, %f9053; mul.rn.ftz.f32 %f9055, %f9054, %f9054; mov.f32 %f9056, 0fC0B59883; mov.f32 %f9057, 0fBF52C7EA; fma.rn.ftz.f32 %f9058, %f9055, %f9057, %f9056; mov.f32 %f9059, 0fC0D21907; fma.rn.ftz.f32 %f9060, %f9058, %f9055, %f9059; mul.ftz.f32 %f9061, %f9055, %f9060; mul.ftz.f32 %f9062, %f9054, %f9061; add.ftz.f32 %f9063, %f9055, 0f41355DC0; mov.f32 %f9064, 0f41E6BD60; fma.rn.ftz.f32 %f9065, %f9063, %f9055, %f9064; mov.f32 %f9066, 0f419D92C8; fma.rn.ftz.f32 %f9067, %f9065, %f9055, %f9066; rcp.approx.ftz.f32 %f9068, %f9067; fma.rn.ftz.f32 %f9069, %f9062, %f9068, %f9054; sub.ftz.f32 %f9071, %f15501, %f9069; setp.gt.ftz.f32 %p1012, %f1562, %f1561; selp.f32 %f9072, %f9071, %f9069, %p1012; sub.ftz.f32 %f9074, %f15508, %f9072; selp.f32 %f9075, %f9074, %f9072, %p1011; mov.b32 %r1744, %f9075; or.b32 %r1745, %r511, %r1744; mov.b32 %f9076, %r1745; add.ftz.f32 %f9077, %f1561, %f1562; setp.le.ftz.f32 %p1013, %f9077, 0f7F800000; selp.f32 %f16016, %f9076, %f9077, %p1013; $L__BB108_1107: ld.const.f32 %f15510, [kernel_params+280]; ld.const.f32 %f15509, [kernel_params+276]; cos.approx.ftz.f32 %f9078, %f16016; mul.ftz.f32 %f9079, %f1565, %f9078; div.approx.ftz.f32 %f9080, %f9079, %f15509; sin.approx.ftz.f32 %f9081, %f16016; mul.ftz.f32 %f9082, %f1565, %f9081; div.approx.ftz.f32 %f9083, %f9082, %f15510; add.ftz.f32 %f16026, %f9083, 0f3F000000; add.ftz.f32 %f16025, %f9080, 0f3F000000; bra.uni $L__BB108_1143; $L__BB108_1098: mov.f32 %f15504, 0f3FC90FDB; setp.lt.s32 %p1042, %r514, 0; min.ftz.f32 %f9154, %f1591, %f1592; max.ftz.f32 %f9155, %f1591, %f1592; div.full.ftz.f32 %f9156, %f9154, %f9155; mul.rn.ftz.f32 %f9157, %f9156, %f9156; mov.f32 %f9158, 0fC0B59883; mov.f32 %f9159, 0fBF52C7EA; fma.rn.ftz.f32 %f9160, %f9157, %f9159, %f9158; mov.f32 %f9161, 0fC0D21907; fma.rn.ftz.f32 %f9162, %f9160, %f9157, %f9161; mul.ftz.f32 %f9163, %f9157, %f9162; mul.ftz.f32 %f9164, %f9156, %f9163; add.ftz.f32 %f9165, %f9157, 0f41355DC0; mov.f32 %f9166, 0f41E6BD60; fma.rn.ftz.f32 %f9167, %f9165, %f9157, %f9166; mov.f32 %f9168, 0f419D92C8; fma.rn.ftz.f32 %f9169, %f9167, %f9157, %f9168; rcp.approx.ftz.f32 %f9170, %f9169; fma.rn.ftz.f32 %f9171, %f9164, %f9170, %f9156; sub.ftz.f32 %f9173, %f15504, %f9171; setp.gt.ftz.f32 %p1043, %f1591, %f1592; selp.f32 %f9174, %f9173, %f9171, %p1043; sub.ftz.f32 %f9176, %f7711, %f9174; selp.f32 %f9177, %f9176, %f9174, %p1042; mov.b32 %r1766, %f9177; or.b32 %r1767, %r515, %r1766; mov.b32 %f9178, %r1767; add.ftz.f32 %f9179, %f1592, %f1591; setp.le.ftz.f32 %p1044, %f9179, 0f7F800000; selp.f32 %f16021, %f9178, %f9179, %p1044; $L__BB108_1124: div.approx.ftz.f32 %f1597, %f16021, %f1590; @%p4 bra $L__BB108_1128; bra.uni $L__BB108_1125; $L__BB108_1128: shr.s32 %r1777, %r510, 31; and.b32 %r1778, %r1777, 1078530011; or.b32 %r1779, %r1778, %r511; mov.b32 %f16022, %r1779; bra.uni $L__BB108_1129; $L__BB108_1125: setp.eq.ftz.f32 %p1046, %f1561, 0f7F800000; setp.eq.ftz.f32 %p1047, %f1562, 0f7F800000; and.pred %p1048, %p1046, %p1047; @%p1048 bra $L__BB108_1127; bra.uni $L__BB108_1126; $L__BB108_1127: setp.lt.s32 %p1052, %r510, 0; selp.b32 %r1775, 1075235812, 1061752795, %p1052; or.b32 %r1776, %r1775, %r511; mov.b32 %f16022, %r1776; bra.uni $L__BB108_1129; $L__BB108_1113: mov.f32 %f15502, 0f3FC90FDB; setp.lt.s32 %p1022, %r512, 0; min.ftz.f32 %f9092, %f1561, %f1579; max.ftz.f32 %f9093, %f1561, %f1579; div.full.ftz.f32 %f9094, %f9092, %f9093; mul.rn.ftz.f32 %f9095, %f9094, %f9094; mov.f32 %f9096, 0fC0B59883; mov.f32 %f9097, 0fBF52C7EA; fma.rn.ftz.f32 %f9098, %f9095, %f9097, %f9096; mov.f32 %f9099, 0fC0D21907; fma.rn.ftz.f32 %f9100, %f9098, %f9095, %f9099; mul.ftz.f32 %f9101, %f9095, %f9100; mul.ftz.f32 %f9102, %f9094, %f9101; add.ftz.f32 %f9103, %f9095, 0f41355DC0; mov.f32 %f9104, 0f41E6BD60; fma.rn.ftz.f32 %f9105, %f9103, %f9095, %f9104; mov.f32 %f9106, 0f419D92C8; fma.rn.ftz.f32 %f9107, %f9105, %f9095, %f9106; rcp.approx.ftz.f32 %f9108, %f9107; fma.rn.ftz.f32 %f9109, %f9102, %f9108, %f9094; sub.ftz.f32 %f9111, %f15502, %f9109; setp.gt.ftz.f32 %p1023, %f1561, %f1579; selp.f32 %f9112, %f9111, %f9109, %p1023; sub.ftz.f32 %f9114, %f7711, %f9112; selp.f32 %f9115, %f9114, %f9112, %p1022; mov.b32 %r1751, %f9115; or.b32 %r1752, %r513, %r1751; mov.b32 %f9116, %r1752; add.ftz.f32 %f9117, %f1579, %f1561; setp.le.ftz.f32 %p1024, %f9117, 0f7F800000; selp.f32 %f16019, %f9116, %f9117, %p1024; $L__BB108_1116: and.pred %p1028, %p1017, %p1001; @%p1028 bra $L__BB108_1120; bra.uni $L__BB108_1117; $L__BB108_1120: shr.s32 %r1762, %r512, 31; and.b32 %r1763, %r1762, 1078530011; or.b32 %r1764, %r511, %r1763; mov.b32 %f16020, %r1764; bra.uni $L__BB108_1121; $L__BB108_1117: setp.eq.ftz.f32 %p1029, %f1579, 0f7F800000; setp.eq.ftz.f32 %p1030, %f1562, 0f7F800000; and.pred %p1031, %p1029, %p1030; @%p1031 bra $L__BB108_1119; bra.uni $L__BB108_1118; $L__BB108_1119: setp.lt.s32 %p1035, %r512, 0; selp.b32 %r1760, 1075235812, 1061752795, %p1035; or.b32 %r1761, %r511, %r1760; mov.b32 %f16020, %r1761; bra.uni $L__BB108_1121; $L__BB108_1139: mov.f32 %f15498, 0f3FC90FDB; setp.lt.s32 %p1073, %r510, 0; min.ftz.f32 %f9276, %f1562, %f1561; max.ftz.f32 %f9277, %f1562, %f1561; div.full.ftz.f32 %f9278, %f9276, %f9277; mul.rn.ftz.f32 %f9279, %f9278, %f9278; mov.f32 %f9280, 0fC0B59883; mov.f32 %f9281, 0fBF52C7EA; fma.rn.ftz.f32 %f9282, %f9279, %f9281, %f9280; mov.f32 %f9283, 0fC0D21907; fma.rn.ftz.f32 %f9284, %f9282, %f9279, %f9283; mul.ftz.f32 %f9285, %f9279, %f9284; mul.ftz.f32 %f9286, %f9278, %f9285; add.ftz.f32 %f9287, %f9279, 0f41355DC0; mov.f32 %f9288, 0f41E6BD60; fma.rn.ftz.f32 %f9289, %f9287, %f9279, %f9288; mov.f32 %f9290, 0f419D92C8; fma.rn.ftz.f32 %f9291, %f9289, %f9279, %f9290; rcp.approx.ftz.f32 %f9292, %f9291; fma.rn.ftz.f32 %f9293, %f9286, %f9292, %f9278; sub.ftz.f32 %f9295, %f15498, %f9293; setp.gt.ftz.f32 %p1074, %f1562, %f1561; selp.f32 %f9296, %f9295, %f9293, %p1074; sub.ftz.f32 %f9298, %f7711, %f9296; selp.f32 %f9299, %f9298, %f9296, %p1073; mov.b32 %r1787, %f9299; or.b32 %r1788, %r511, %r1787; mov.b32 %f9300, %r1788; add.ftz.f32 %f9301, %f1561, %f1562; setp.le.ftz.f32 %p1075, %f9301, 0f7F800000; selp.f32 %f16024, %f9300, %f9301, %p1075; $L__BB108_1142: cos.approx.ftz.f32 %f9302, %f16024; mul.ftz.f32 %f9303, %f1617, %f9302; div.approx.ftz.f32 %f9304, %f9303, %f1615; sin.approx.ftz.f32 %f9305, %f16024; mul.ftz.f32 %f9306, %f1617, %f9305; div.approx.ftz.f32 %f9307, %f9306, %f1616; add.ftz.f32 %f16026, %f9307, 0f3F000000; add.ftz.f32 %f16025, %f9304, 0f3F000000; bra.uni $L__BB108_1143; $L__BB108_1126: mov.f32 %f15505, 0f3FC90FDB; setp.lt.s32 %p1049, %r510, 0; min.ftz.f32 %f9180, %f1562, %f1561; max.ftz.f32 %f9181, %f1562, %f1561; div.full.ftz.f32 %f9182, %f9180, %f9181; mul.rn.ftz.f32 %f9183, %f9182, %f9182; mov.f32 %f9184, 0fC0B59883; mov.f32 %f9185, 0fBF52C7EA; fma.rn.ftz.f32 %f9186, %f9183, %f9185, %f9184; mov.f32 %f9187, 0fC0D21907; fma.rn.ftz.f32 %f9188, %f9186, %f9183, %f9187; mul.ftz.f32 %f9189, %f9183, %f9188; mul.ftz.f32 %f9190, %f9182, %f9189; add.ftz.f32 %f9191, %f9183, 0f41355DC0; mov.f32 %f9192, 0f41E6BD60; fma.rn.ftz.f32 %f9193, %f9191, %f9183, %f9192; mov.f32 %f9194, 0f419D92C8; fma.rn.ftz.f32 %f9195, %f9193, %f9183, %f9194; rcp.approx.ftz.f32 %f9196, %f9195; fma.rn.ftz.f32 %f9197, %f9190, %f9196, %f9182; sub.ftz.f32 %f9199, %f15505, %f9197; setp.gt.ftz.f32 %p1050, %f1562, %f1561; selp.f32 %f9200, %f9199, %f9197, %p1050; sub.ftz.f32 %f9202, %f7711, %f9200; selp.f32 %f9203, %f9202, %f9200, %p1049; mov.b32 %r1773, %f9203; or.b32 %r1774, %r511, %r1773; mov.b32 %f9204, %r1774; add.ftz.f32 %f9205, %f1561, %f1562; setp.le.ftz.f32 %p1051, %f9205, 0f7F800000; selp.f32 %f16022, %f9204, %f9205, %p1051; $L__BB108_1129: cos.approx.ftz.f32 %f9206, %f16022; sin.approx.ftz.f32 %f9207, %f16022; fma.rn.ftz.f32 %f16026, %f1597, %f9207, 0f3F000000; fma.rn.ftz.f32 %f16025, %f1597, %f9206, 0f3F000000; bra.uni $L__BB108_1143; $L__BB108_1118: mov.f32 %f15503, 0f3FC90FDB; setp.lt.s32 %p1032, %r512, 0; min.ftz.f32 %f9118, %f1562, %f1579; max.ftz.f32 %f9119, %f1562, %f1579; div.full.ftz.f32 %f9120, %f9118, %f9119; mul.rn.ftz.f32 %f9121, %f9120, %f9120; mov.f32 %f9122, 0fC0B59883; mov.f32 %f9123, 0fBF52C7EA; fma.rn.ftz.f32 %f9124, %f9121, %f9123, %f9122; mov.f32 %f9125, 0fC0D21907; fma.rn.ftz.f32 %f9126, %f9124, %f9121, %f9125; mul.ftz.f32 %f9127, %f9121, %f9126; mul.ftz.f32 %f9128, %f9120, %f9127; add.ftz.f32 %f9129, %f9121, 0f41355DC0; mov.f32 %f9130, 0f41E6BD60; fma.rn.ftz.f32 %f9131, %f9129, %f9121, %f9130; mov.f32 %f9132, 0f419D92C8; fma.rn.ftz.f32 %f9133, %f9131, %f9121, %f9132; rcp.approx.ftz.f32 %f9134, %f9133; fma.rn.ftz.f32 %f9135, %f9128, %f9134, %f9120; sub.ftz.f32 %f9137, %f15503, %f9135; setp.gt.ftz.f32 %p1033, %f1562, %f1579; selp.f32 %f9138, %f9137, %f9135, %p1033; sub.ftz.f32 %f9140, %f7711, %f9138; selp.f32 %f9141, %f9140, %f9138, %p1032; mov.b32 %r1758, %f9141; or.b32 %r1759, %r511, %r1758; mov.b32 %f9142, %r1759; add.ftz.f32 %f9143, %f1579, %f1562; setp.le.ftz.f32 %p1034, %f9143, 0f7F800000; selp.f32 %f16020, %f9142, %f9143, %p1034; $L__BB108_1121: mov.f32 %f15480, 0f3F800000; add.ftz.f32 %f9144, %f16020, %f16020; div.approx.ftz.f32 %f9146, %f9144, %f7711; add.ftz.f32 %f9147, %f16019, %f16019; div.approx.ftz.f32 %f9148, %f9147, %f7711; add.ftz.f32 %f9149, %f9148, 0f3F000000; sub.ftz.f32 %f16025, %f15480, %f9149; add.ftz.f32 %f16026, %f9146, 0f3F000000; bra.uni $L__BB108_1143; $L__BB108_1133: mov.f32 %f15506, 0f3FC90FDB; setp.lt.s32 %p1064, %r516, 0; min.ftz.f32 %f9214, %f1561, %f1608; max.ftz.f32 %f9215, %f1561, %f1608; div.full.ftz.f32 %f9216, %f9214, %f9215; mul.rn.ftz.f32 %f9217, %f9216, %f9216; mov.f32 %f9218, 0fC0B59883; mov.f32 %f9219, 0fBF52C7EA; fma.rn.ftz.f32 %f9220, %f9217, %f9219, %f9218; mov.f32 %f9221, 0fC0D21907; fma.rn.ftz.f32 %f9222, %f9220, %f9217, %f9221; mul.ftz.f32 %f9223, %f9217, %f9222; mul.ftz.f32 %f9224, %f9216, %f9223; add.ftz.f32 %f9225, %f9217, 0f41355DC0; mov.f32 %f9226, 0f41E6BD60; fma.rn.ftz.f32 %f9227, %f9225, %f9217, %f9226; mov.f32 %f9228, 0f419D92C8; fma.rn.ftz.f32 %f9229, %f9227, %f9217, %f9228; rcp.approx.ftz.f32 %f9230, %f9229; fma.rn.ftz.f32 %f9231, %f9224, %f9230, %f9216; sub.ftz.f32 %f9233, %f15506, %f9231; setp.gt.ftz.f32 %p1065, %f1561, %f1608; selp.f32 %f9234, %f9233, %f9231, %p1065; sub.ftz.f32 %f9236, %f7711, %f9234; selp.f32 %f9237, %f9236, %f9234, %p1064; mov.b32 %r1780, %f9237; or.b32 %r1781, %r517, %r1780; mov.b32 %f9238, %r1781; add.ftz.f32 %f9239, %f1608, %f1561; setp.le.ftz.f32 %p1066, %f9239, 0f7F800000; selp.f32 %f16023, %f9238, %f9239, %p1066; $L__BB108_1136: mov.f32 %f15507, 0f3FC90FDB; mov.f32 %f15486, 0f3E2AAB94; mov.f32 %f15485, 0f3D994929; mov.f32 %f15484, 0f3D3F841F; mov.f32 %f15483, 0f3C94D2E9; mov.f32 %f15482, 0f3D53F941; mov.f32 %f15481, 0f3F800000; sub.ftz.f32 %f9240, %f16023, %f9211; mul.ftz.f32 %f9241, %f1558, %f1558; fma.rn.ftz.f32 %f9242, %f1557, %f1557, %f9241; fma.rn.ftz.f32 %f9243, %f1559, %f1559, %f9242; sqrt.approx.ftz.f32 %f9244, %f9243; div.approx.ftz.f32 %f9245, %f1559, %f9244; abs.ftz.f32 %f9246, %f9245; sub.ftz.f32 %f9248, %f15481, %f9246; mul.ftz.f32 %f9249, %f9248, 0f3F000000; sqrt.approx.ftz.f32 %f9250, %f9249; setp.gt.ftz.f32 %p1068, %f9246, 0f3F11EB85; selp.f32 %f9251, %f9250, %f9246, %p1068; mul.ftz.f32 %f9252, %f9251, %f9251; fma.rn.ftz.f32 %f9255, %f15482, %f9252, %f15483; fma.rn.ftz.f32 %f9257, %f9255, %f9252, %f15484; fma.rn.ftz.f32 %f9259, %f9257, %f9252, %f15485; fma.rn.ftz.f32 %f9261, %f9259, %f9252, %f15486; mul.ftz.f32 %f9262, %f9252, %f9261; fma.rn.ftz.f32 %f9263, %f9262, %f9251, %f9251; add.ftz.f32 %f9264, %f9263, %f9263; sub.ftz.f32 %f9266, %f15507, %f9263; selp.f32 %f9267, %f9264, %f9266, %p1068; setp.lt.ftz.f32 %p1069, %f9245, 0f00000000; sub.ftz.f32 %f9269, %f7711, %f9267; selp.f32 %f9270, %f9269, %f9267, %p1069; sub.ftz.f32 %f9271, %f9270, %f9213; div.approx.ftz.f32 %f16026, %f9271, %f9212; div.approx.ftz.f32 %f16025, %f9240, %f9210; $L__BB108_1143: ld.const.f32 %f15500, [kernel_params+288]; ld.const.f32 %f15499, [kernel_params+284]; mul.ftz.f32 %f16029, %f16025, %f15499; mul.ftz.f32 %f16030, %f16026, %f15500; mul.ftz.f32 %f16027, %f16003, %f15499; mul.ftz.f32 %f16028, %f16004, %f15500; mul.ftz.f32 %f16032, %f16015, %f15500; mul.ftz.f32 %f16031, %f16014, %f15499; $L__BB108_1144: ld.const.s32 %rd1553, [kernel_params+1288]; add.s64 %rd1554, %rd166, %rd1553; cvta.to.global.u64 %rd1555, %rd618; shl.b64 %rd1556, %rd1554, 2; add.s64 %rd1557, %rd1555, %rd1556; add.s64 %rd1558, %rd1557, 4; add.s64 %rd1559, %rd1557, 8; add.s64 %rd1560, %rd1557, 12; sub.ftz.f32 %f9308, %f16031, %f16027; atom.global.add.f32 %f9309, [%rd1557], %f9308; sub.ftz.f32 %f9310, %f16032, %f16028; atom.global.add.f32 %f9311, [%rd1558], %f9310; sub.ftz.f32 %f9312, %f16027, %f16029; atom.global.add.f32 %f9313, [%rd1559], %f9312; sub.ftz.f32 %f9314, %f16028, %f16030; atom.global.add.f32 %f9315, [%rd1560], %f9314; ld.const.s32 %rd1561, [kernel_params+1292]; add.s64 %rd1562, %rd166, %rd1561; shl.b64 %rd1563, %rd1562, 2; add.s64 %rd1564, %rd1555, %rd1563; atom.global.add.f32 %f9316, [%rd1564], 0f3F800000; $L__BB108_1145: ld.global.u32 %r1794, [%rd29]; or.b32 %r1795, %r1794, 8192; st.global.u32 [%rd29], %r1795; $L__BB108_1146: ld.const.u32 %r518, [kernel_params+1388]; setp.eq.s32 %p1077, %r518, 0; @%p1077 bra $L__BB108_1184; ld.const.u64 %rd1565, [kernel_params+2584]; cvta.to.global.u64 %rd1566, %rd1565; mul.lo.s64 %rd1567, %rd2, 12; add.s64 %rd1568, %rd1566, %rd1567; ld.global.f32 %f9320, [%rd1568+4]; ld.global.f32 %f9321, [%rd1568]; add.ftz.f32 %f9322, %f9321, %f9320; ld.global.f32 %f9323, [%rd1568+8]; add.ftz.f32 %f1638, %f9323, %f9322; ld.local.u32 %r519, [%rd7+-36]; and.b32 %r1796, %r519, 524288; setp.ne.s32 %p1078, %r1796, 0; mov.f32 %f9319, 0f3F800000; mov.f32 %f16033, %f9319; mov.f32 %f16034, %f9319; mov.f32 %f16035, %f9319; @%p1078 bra $L__BB108_1150; and.b32 %r1797, %r519, 512; setp.eq.s32 %p1079, %r1797, 0; mov.f32 %f16033, 0f00000000; mov.f32 %f16034, %f16033; mov.f32 %f16035, %f16033; @%p1079 bra $L__BB108_1150; ld.local.v2.f32 {%f16034, %f16035}, [%rd7+208]; ld.local.f32 %f16033, [%rd7+204]; $L__BB108_1150: add.ftz.f32 %f9329, %f16033, %f16034; add.ftz.f32 %f9330, %f16035, %f9329; mul.ftz.f32 %f9331, %f9330, 0f3EAAAAAB; sub.ftz.f32 %f9333, %f9319, %f9331; mul.ftz.f32 %f9334, %f1638, 0f3EAAAAAB; mul.ftz.f32 %f1645, %f9334, %f9333; setp.leu.ftz.f32 %p1080, %f1645, 0f00000000; @%p1080 bra $L__BB108_1184; ld.const.s32 %rd1569, [kernel_params+1396]; add.s64 %rd1570, %rd166, %rd1569; cvta.to.global.u64 %rd1571, %rd618; shl.b64 %rd1572, %rd1570, 2; add.s64 %rd2981, %rd1571, %rd1572; and.b32 %r1798, %r518, 1; setp.eq.b32 %p1081, %r1798, 1; mov.pred %p1082, 0; xor.pred %p1083, %p1081, %p1082; not.pred %p1084, %p1083; @%p1084 bra $L__BB108_1163; ld.local.u32 %r520, [%rd7+-12]; setp.eq.s32 %p1085, %r520, -1; mov.f32 %f1647, 0f00000000; mov.u32 %r522, 0; @%p1085 bra $L__BB108_1154; ld.const.u64 %rd1573, [kernel_params+2160]; cvta.to.global.u64 %rd1574, %rd1573; mul.wide.s32 %rd1575, %r520, 208; add.s64 %rd1576, %rd1574, %rd1575; ld.global.f32 %f1647, [%rd1576+172]; mov.b32 %r522, %f1647; $L__BB108_1154: ld.const.u32 %r523, [kernel_params+1392]; shl.b32 %r524, %r523, 1; setp.lt.s32 %p1086, %r524, 1; setp.eq.ftz.f32 %p1087, %f1645, 0f00000000; or.pred %p1088, %p1087, %p1086; @%p1088 bra $L__BB108_1162; add.s32 %r525, %r524, -1; mov.u32 %r526, 0; $L__BB108_1156: mul.wide.s32 %rd1577, %r526, 8; add.s64 %rd255, %rd2981, %rd1577; ld.global.f32 %f1648, [%rd255]; setp.eq.ftz.f32 %p1089, %f1648, 0f00000000; @%p1089 bra $L__BB108_1159; bra.uni $L__BB108_1157; $L__BB108_1159: shl.b32 %r1801, %r526, 1; mul.wide.s32 %rd1579, %r1801, 4; add.s64 %rd1580, %rd2981, %rd1579; mov.u32 %r1802, 0; atom.global.cas.b32 %r1803, [%rd1580], %r1802, %r522; mov.b32 %f9337, %r1803; setp.neu.ftz.f32 %p1093, %f9337, 0f00000000; setp.neu.ftz.f32 %p1094, %f1647, %f9337; and.pred %p1095, %p1093, %p1094; @%p1095 bra $L__BB108_1161; bra.uni $L__BB108_1160; $L__BB108_1157: setp.neu.ftz.f32 %p1090, %f1648, %f1647; setp.ne.s32 %p1091, %r526, %r525; and.pred %p1092, %p1091, %p1090; @%p1092 bra $L__BB108_1161; bra.uni $L__BB108_1158; $L__BB108_1161: add.s32 %r526, %r526, 1; setp.lt.s32 %p1096, %r526, %r524; @%p1096 bra $L__BB108_1156; bra.uni $L__BB108_1162; $L__BB108_1160: add.s64 %rd1581, %rd255, 4; atom.global.add.f32 %f9338, [%rd1581], %f1645; bra.uni $L__BB108_1162; $L__BB108_1158: add.s64 %rd1578, %rd255, 4; atom.global.add.f32 %f9336, [%rd1578], %f1645; $L__BB108_1162: mul.wide.s32 %rd1582, %r523, 4; shl.b64 %rd1583, %rd1582, 2; add.s64 %rd2981, %rd2981, %rd1583; $L__BB108_1163: and.b32 %r1804, %r518, 2; setp.eq.s32 %p1097, %r1804, 0; @%p1097 bra $L__BB108_1173; ld.const.u64 %rd2916, [kernel_params+2408]; cvta.to.global.u64 %rd2915, %rd2916; ld.local.u32 %rd1584, [%rd7+-40]; shl.b64 %rd1585, %rd1584, 5; and.b64 %rd1586, %rd1585, 134217696; add.s64 %rd1587, %rd2915, %rd1586; ld.global.f32 %f1649, [%rd1587+12]; ld.const.u32 %r528, [kernel_params+1392]; shl.b32 %r529, %r528, 1; setp.lt.s32 %p1098, %r529, 1; setp.eq.ftz.f32 %p1099, %f1645, 0f00000000; or.pred %p1100, %p1099, %p1098; @%p1100 bra $L__BB108_1172; mov.b32 %r530, %f1649; add.s32 %r531, %r529, -1; mov.u32 %r532, 0; $L__BB108_1166: mul.wide.s32 %rd1588, %r532, 8; add.s64 %rd258, %rd2981, %rd1588; ld.global.f32 %f1650, [%rd258]; setp.eq.ftz.f32 %p1101, %f1650, 0f00000000; @%p1101 bra $L__BB108_1169; bra.uni $L__BB108_1167; $L__BB108_1169: mov.u32 %r1806, 0; atom.global.cas.b32 %r1807, [%rd258], %r1806, %r530; mov.b32 %f9340, %r1807; setp.neu.ftz.f32 %p1105, %f9340, 0f00000000; setp.neu.ftz.f32 %p1106, %f1649, %f9340; and.pred %p1107, %p1105, %p1106; @%p1107 bra $L__BB108_1171; bra.uni $L__BB108_1170; $L__BB108_1167: setp.neu.ftz.f32 %p1102, %f1650, %f1649; setp.ne.s32 %p1103, %r532, %r531; and.pred %p1104, %p1103, %p1102; @%p1104 bra $L__BB108_1171; bra.uni $L__BB108_1168; $L__BB108_1171: add.s32 %r532, %r532, 1; setp.lt.s32 %p1108, %r532, %r529; @%p1108 bra $L__BB108_1166; bra.uni $L__BB108_1172; $L__BB108_1170: add.s64 %rd1590, %rd258, 4; atom.global.add.f32 %f9341, [%rd1590], %f1645; bra.uni $L__BB108_1172; $L__BB108_1168: add.s64 %rd1589, %rd258, 4; atom.global.add.f32 %f9339, [%rd1589], %f1645; $L__BB108_1172: mul.wide.s32 %rd1591, %r528, 4; shl.b64 %rd1592, %rd1591, 2; add.s64 %rd2981, %rd2981, %rd1592; $L__BB108_1173: and.b32 %r1808, %r518, 4; setp.eq.s32 %p1109, %r1808, 0; @%p1109 bra $L__BB108_1184; ld.local.u32 %r534, [%rd7+-12]; setp.eq.s32 %p1110, %r534, -1; mov.f32 %f1652, 0f00000000; mov.u32 %r536, 0; @%p1110 bra $L__BB108_1176; ld.const.u64 %rd1593, [kernel_params+2160]; cvta.to.global.u64 %rd1594, %rd1593; mul.wide.s32 %rd1595, %r534, 208; add.s64 %rd1596, %rd1594, %rd1595; ld.global.f32 %f1652, [%rd1596+176]; mov.b32 %r536, %f1652; $L__BB108_1176: ld.const.u32 %r1810, [kernel_params+1392]; shl.b32 %r537, %r1810, 1; setp.lt.s32 %p1111, %r537, 1; setp.eq.ftz.f32 %p1112, %f1645, 0f00000000; or.pred %p1113, %p1112, %p1111; @%p1113 bra $L__BB108_1184; add.s32 %r538, %r537, -1; mov.u32 %r539, 0; $L__BB108_1178: mul.wide.s32 %rd1597, %r539, 8; add.s64 %rd261, %rd2981, %rd1597; ld.global.f32 %f1653, [%rd261]; setp.eq.ftz.f32 %p1114, %f1653, 0f00000000; @%p1114 bra $L__BB108_1181; bra.uni $L__BB108_1179; $L__BB108_1181: mov.u32 %r1812, 0; atom.global.cas.b32 %r1813, [%rd261], %r1812, %r536; mov.b32 %f9344, %r1813; setp.neu.ftz.f32 %p1118, %f9344, 0f00000000; setp.neu.ftz.f32 %p1119, %f1652, %f9344; and.pred %p1120, %p1118, %p1119; @%p1120 bra $L__BB108_1183; bra.uni $L__BB108_1182; $L__BB108_1179: setp.neu.ftz.f32 %p1115, %f1653, %f1652; setp.ne.s32 %p1116, %r539, %r538; and.pred %p1117, %p1116, %p1115; @%p1117 bra $L__BB108_1183; bra.uni $L__BB108_1180; $L__BB108_1183: add.s32 %r539, %r539, 1; setp.lt.s32 %p1121, %r539, %r537; @%p1121 bra $L__BB108_1178; bra.uni $L__BB108_1184; $L__BB108_1182: add.s64 %rd1599, %rd261, 4; atom.global.add.f32 %f9345, [%rd1599], %f1645; bra.uni $L__BB108_1184; $L__BB108_1180: add.s64 %rd1598, %rd261, 4; atom.global.add.f32 %f9343, [%rd1598], %f1645; $L__BB108_1184: ld.const.u32 %r3052, [kernel_params+1256]; and.b32 %r1814, %r3052, 16384; setp.eq.s32 %p1122, %r1814, 0; @%p1122 bra $L__BB108_1207; ld.const.u64 %rd1600, [kernel_params+2584]; cvta.to.global.u64 %rd1601, %rd1600; mul.lo.s64 %rd1602, %rd2, 12; add.s64 %rd1603, %rd1601, %rd1602; ld.global.f32 %f1654, [%rd1603]; ld.global.f32 %f1655, [%rd1603+4]; ld.global.f32 %f1656, [%rd1603+8]; ld.const.u32 %r541, [kernel_params+1308]; ld.local.u32 %r542, [%rd7+172]; setp.lt.s32 %p1123, %r542, 1; mov.f32 %f16040, 0f00000000; mov.f32 %f16039, %f16040; mov.f32 %f16038, %f16040; @%p1123 bra $L__BB108_1206; add.s32 %r1816, %r542, -1; and.b32 %r3210, %r542, 3; setp.lt.u32 %p1124, %r1816, 3; mov.f32 %f16038, 0f00000000; mov.u32 %r3208, 0; mov.f32 %f16039, %f16038; mov.f32 %f16040, %f16038; @%p1124 bra $L__BB108_1201; sub.s32 %r3207, %r542, %r3210; mov.f32 %f16038, 0f00000000; mov.u32 %r3208, 0; mov.f32 %f16039, %f16038; mov.f32 %f16040, %f16038; $L__BB108_1188: cvt.s64.s32 %rd262, %r3208; mul.wide.s32 %rd1604, %r3208, 80; add.s64 %rd1605, %rd1, %rd1604; add.s64 %rd263, %rd1605, 316; ld.local.u32 %r547, [%rd1605+316]; add.s32 %r1818, %r547, -2; setp.lt.u32 %p1125, %r1818, 7; @%p1125 bra $L__BB108_1190; add.s32 %r1819, %r547, -32; setp.gt.u32 %p1126, %r1819, 2; @%p1126 bra $L__BB108_1191; $L__BB108_1190: ld.local.v4.f32 {%f9356, %f9357, %f9358, %f9359}, [%rd263+-12]; add.ftz.f32 %f16040, %f16040, %f9356; add.ftz.f32 %f16039, %f16039, %f9357; add.ftz.f32 %f16038, %f16038, %f9358; $L__BB108_1191: ld.local.u32 %r548, [%rd263+80]; add.s32 %r1820, %r548, -2; setp.lt.u32 %p1127, %r1820, 7; @%p1127 bra $L__BB108_1193; add.s32 %r1821, %r548, -32; setp.gt.u32 %p1128, %r1821, 2; @%p1128 bra $L__BB108_1194; $L__BB108_1193: ld.local.v4.f32 {%f9363, %f9364, %f9365, %f9366}, [%rd263+68]; add.ftz.f32 %f16040, %f16040, %f9363; add.ftz.f32 %f16039, %f16039, %f9364; add.ftz.f32 %f16038, %f16038, %f9365; $L__BB108_1194: ld.local.u32 %r549, [%rd263+160]; add.s32 %r1822, %r549, -2; setp.lt.u32 %p1129, %r1822, 7; @%p1129 bra $L__BB108_1196; add.s32 %r1823, %r549, -32; setp.gt.u32 %p1130, %r1823, 2; @%p1130 bra $L__BB108_1197; $L__BB108_1196: ld.local.v4.f32 {%f9370, %f9371, %f9372, %f9373}, [%rd263+148]; add.ftz.f32 %f16040, %f16040, %f9370; add.ftz.f32 %f16039, %f16039, %f9371; add.ftz.f32 %f16038, %f16038, %f9372; $L__BB108_1197: ld.local.u32 %r550, [%rd263+240]; add.s32 %r1824, %r550, -2; setp.lt.u32 %p1131, %r1824, 7; @%p1131 bra $L__BB108_1199; add.s32 %r1825, %r550, -32; setp.gt.u32 %p1132, %r1825, 2; @%p1132 bra $L__BB108_1200; $L__BB108_1199: ld.local.v4.f32 {%f9377, %f9378, %f9379, %f9380}, [%rd263+228]; add.ftz.f32 %f16040, %f16040, %f9377; add.ftz.f32 %f16039, %f16039, %f9378; add.ftz.f32 %f16038, %f16038, %f9379; $L__BB108_1200: cvt.u32.u64 %r1826, %rd262; add.s32 %r3208, %r1826, 4; add.s32 %r3207, %r3207, -4; setp.ne.s32 %p1133, %r3207, 0; @%p1133 bra $L__BB108_1188; $L__BB108_1201: setp.eq.s32 %p1134, %r3210, 0; @%p1134 bra $L__BB108_1206; $L__BB108_1202: .pragma "nounroll"; cvt.s64.s32 %rd264, %r3208; mul.wide.s32 %rd1606, %r3208, 80; add.s64 %rd1607, %rd1, %rd1606; add.s64 %rd265, %rd1607, 316; ld.local.u32 %r556, [%rd1607+316]; add.s32 %r1827, %r556, -2; setp.lt.u32 %p1135, %r1827, 7; @%p1135 bra $L__BB108_1204; add.s32 %r1828, %r556, -32; setp.gt.u32 %p1136, %r1828, 2; @%p1136 bra $L__BB108_1205; $L__BB108_1204: ld.local.f32 %f9384, [%rd265+-12]; add.ftz.f32 %f16040, %f16040, %f9384; ld.local.f32 %f9385, [%rd265+-8]; add.ftz.f32 %f16039, %f16039, %f9385; ld.local.f32 %f9386, [%rd265+-4]; add.ftz.f32 %f16038, %f16038, %f9386; $L__BB108_1205: cvt.u32.u64 %r1829, %rd264; add.s32 %r3208, %r1829, 1; add.s32 %r3210, %r3210, -1; setp.ne.s32 %p1137, %r3210, 0; @%p1137 bra $L__BB108_1202; $L__BB108_1206: cvta.to.global.u64 %rd1608, %rd618; cvt.s64.s32 %rd1609, %r541; add.s64 %rd1610, %rd166, %rd1609; shl.b64 %rd1611, %rd1610, 2; add.s64 %rd1612, %rd1608, %rd1611; mul.ftz.f32 %f9387, %f1654, %f16040; atom.global.add.f32 %f9388, [%rd1612], %f9387; add.s64 %rd1613, %rd1612, 4; mul.ftz.f32 %f9389, %f1655, %f16039; atom.global.add.f32 %f9390, [%rd1613], %f9389; add.s64 %rd1614, %rd1612, 8; mul.ftz.f32 %f9391, %f1656, %f16038; atom.global.add.f32 %f9392, [%rd1614], %f9391; $L__BB108_1207: ld.const.u32 %r3053, [kernel_params+1256]; cvt.u16.u32 %rs80, %r3053; setp.gt.s16 %p1138, %rs80, -1; @%p1138 bra $L__BB108_1225; ld.const.u64 %rd1615, [kernel_params+2584]; cvta.to.global.u64 %rd1616, %rd1615; mul.lo.s64 %rd1617, %rd2, 12; add.s64 %rd1618, %rd1616, %rd1617; ld.global.f32 %f1702, [%rd1618]; ld.global.f32 %f1703, [%rd1618+4]; ld.global.f32 %f1704, [%rd1618+8]; ld.const.u32 %r559, [kernel_params+1312]; ld.local.u32 %r560, [%rd7+172]; setp.lt.s32 %p1139, %r560, 1; mov.f32 %f16071, 0f00000000; mov.f32 %f16072, %f16071; mov.f32 %f16073, %f16071; @%p1139 bra $L__BB108_1224; add.s32 %r1831, %r560, -1; and.b32 %r3215, %r560, 3; setp.lt.u32 %p1140, %r1831, 3; mov.f32 %f16073, 0f00000000; mov.u32 %r3213, 0; mov.f32 %f16072, %f16073; mov.f32 %f16071, %f16073; @%p1140 bra $L__BB108_1220; sub.s32 %r3212, %r560, %r3215; mov.f32 %f16073, 0f00000000; mov.u32 %r3213, 0; mov.f32 %f16072, %f16073; mov.f32 %f16071, %f16073; $L__BB108_1211: cvt.s64.s32 %rd266, %r3213; mul.wide.s32 %rd1619, %r3213, 80; add.s64 %rd1620, %rd1, %rd1619; add.s64 %rd267, %rd1620, 316; ld.local.u32 %r1833, [%rd1620+316]; add.s32 %r1834, %r1833, -9; setp.gt.u32 %p1141, %r1834, 11; setp.ne.s32 %p1142, %r1833, 29; and.pred %p1143, %p1141, %p1142; @%p1143 bra $L__BB108_1213; ld.local.v4.f32 {%f9403, %f9404, %f9405, %f9406}, [%rd267+-12]; add.ftz.f32 %f16071, %f16071, %f9403; add.ftz.f32 %f16072, %f16072, %f9404; add.ftz.f32 %f16073, %f16073, %f9405; $L__BB108_1213: ld.local.u32 %r1835, [%rd267+80]; add.s32 %r1836, %r1835, -9; setp.gt.u32 %p1144, %r1836, 11; setp.ne.s32 %p1145, %r1835, 29; and.pred %p1146, %p1144, %p1145; @%p1146 bra $L__BB108_1215; ld.local.v4.f32 {%f9410, %f9411, %f9412, %f9413}, [%rd267+68]; add.ftz.f32 %f16071, %f16071, %f9410; add.ftz.f32 %f16072, %f16072, %f9411; add.ftz.f32 %f16073, %f16073, %f9412; $L__BB108_1215: ld.local.u32 %r1837, [%rd267+160]; add.s32 %r1838, %r1837, -9; setp.gt.u32 %p1147, %r1838, 11; setp.ne.s32 %p1148, %r1837, 29; and.pred %p1149, %p1147, %p1148; @%p1149 bra $L__BB108_1217; ld.local.v4.f32 {%f9417, %f9418, %f9419, %f9420}, [%rd267+148]; add.ftz.f32 %f16071, %f16071, %f9417; add.ftz.f32 %f16072, %f16072, %f9418; add.ftz.f32 %f16073, %f16073, %f9419; $L__BB108_1217: ld.local.u32 %r1839, [%rd267+240]; add.s32 %r1840, %r1839, -9; setp.gt.u32 %p1150, %r1840, 11; setp.ne.s32 %p1151, %r1839, 29; and.pred %p1152, %p1150, %p1151; @%p1152 bra $L__BB108_1219; ld.local.v4.f32 {%f9424, %f9425, %f9426, %f9427}, [%rd267+228]; add.ftz.f32 %f16071, %f16071, %f9424; add.ftz.f32 %f16072, %f16072, %f9425; add.ftz.f32 %f16073, %f16073, %f9426; $L__BB108_1219: cvt.u32.u64 %r1841, %rd266; add.s32 %r3213, %r1841, 4; add.s32 %r3212, %r3212, -4; setp.ne.s32 %p1153, %r3212, 0; @%p1153 bra $L__BB108_1211; $L__BB108_1220: setp.eq.s32 %p1154, %r3215, 0; @%p1154 bra $L__BB108_1224; $L__BB108_1221: .pragma "nounroll"; cvt.s64.s32 %rd268, %r3213; mul.wide.s32 %rd1621, %r3213, 80; add.s64 %rd1622, %rd1, %rd1621; add.s64 %rd269, %rd1622, 316; ld.local.u32 %r1842, [%rd1622+316]; add.s32 %r1843, %r1842, -9; setp.gt.u32 %p1155, %r1843, 11; setp.ne.s32 %p1156, %r1842, 29; and.pred %p1157, %p1155, %p1156; @%p1157 bra $L__BB108_1223; ld.local.f32 %f9431, [%rd269+-12]; add.ftz.f32 %f16071, %f16071, %f9431; ld.local.f32 %f9432, [%rd269+-8]; add.ftz.f32 %f16072, %f16072, %f9432; ld.local.f32 %f9433, [%rd269+-4]; add.ftz.f32 %f16073, %f16073, %f9433; $L__BB108_1223: cvt.u32.u64 %r1844, %rd268; add.s32 %r3213, %r1844, 1; add.s32 %r3215, %r3215, -1; setp.ne.s32 %p1158, %r3215, 0; @%p1158 bra $L__BB108_1221; $L__BB108_1224: cvta.to.global.u64 %rd1623, %rd618; cvt.s64.s32 %rd1624, %r559; add.s64 %rd1625, %rd166, %rd1624; shl.b64 %rd1626, %rd1625, 2; add.s64 %rd1627, %rd1623, %rd1626; mul.ftz.f32 %f9434, %f1702, %f16071; atom.global.add.f32 %f9435, [%rd1627], %f9434; add.s64 %rd1628, %rd1627, 4; mul.ftz.f32 %f9436, %f1703, %f16072; atom.global.add.f32 %f9437, [%rd1628], %f9436; add.s64 %rd1629, %rd1627, 8; mul.ftz.f32 %f9438, %f1704, %f16073; atom.global.add.f32 %f9439, [%rd1629], %f9438; $L__BB108_1225: ld.const.u32 %r3054, [kernel_params+1256]; and.b32 %r1845, %r3054, 65536; setp.eq.s32 %p1159, %r1845, 0; @%p1159 bra $L__BB108_1243; ld.const.u64 %rd1630, [kernel_params+2584]; cvta.to.global.u64 %rd1631, %rd1630; mul.lo.s64 %rd1632, %rd2, 12; add.s64 %rd1633, %rd1631, %rd1632; ld.global.f32 %f1750, [%rd1633]; ld.global.f32 %f1751, [%rd1633+4]; ld.global.f32 %f1752, [%rd1633+8]; ld.const.u32 %r572, [kernel_params+1316]; ld.local.u32 %r573, [%rd7+172]; setp.lt.s32 %p1160, %r573, 1; mov.f32 %f16101, 0f00000000; mov.f32 %f16102, %f16101; mov.f32 %f16103, %f16101; @%p1160 bra $L__BB108_1242; add.s32 %r1847, %r573, -1; and.b32 %r3220, %r573, 3; setp.lt.u32 %p1161, %r1847, 3; mov.f32 %f16103, 0f00000000; mov.u32 %r3218, 0; mov.f32 %f16102, %f16103; mov.f32 %f16101, %f16103; @%p1161 bra $L__BB108_1238; sub.s32 %r3217, %r573, %r3220; mov.f32 %f16103, 0f00000000; mov.u32 %r3218, 0; mov.f32 %f16102, %f16103; mov.f32 %f16101, %f16103; $L__BB108_1229: cvt.s64.s32 %rd270, %r3218; mul.wide.s32 %rd1634, %r3218, 80; add.s64 %rd1635, %rd1, %rd1634; add.s64 %rd271, %rd1635, 316; ld.local.u32 %r1849, [%rd1635+316]; add.s32 %r1850, %r1849, -21; setp.gt.u32 %p1162, %r1850, 9; @%p1162 bra $L__BB108_1231; ld.local.v4.f32 {%f9450, %f9451, %f9452, %f9453}, [%rd271+-12]; add.ftz.f32 %f16101, %f16101, %f9450; add.ftz.f32 %f16102, %f16102, %f9451; add.ftz.f32 %f16103, %f16103, %f9452; $L__BB108_1231: ld.local.u32 %r1851, [%rd271+80]; add.s32 %r1852, %r1851, -21; setp.gt.u32 %p1163, %r1852, 9; @%p1163 bra $L__BB108_1233; ld.local.v4.f32 {%f9457, %f9458, %f9459, %f9460}, [%rd271+68]; add.ftz.f32 %f16101, %f16101, %f9457; add.ftz.f32 %f16102, %f16102, %f9458; add.ftz.f32 %f16103, %f16103, %f9459; $L__BB108_1233: ld.local.u32 %r1853, [%rd271+160]; add.s32 %r1854, %r1853, -21; setp.gt.u32 %p1164, %r1854, 9; @%p1164 bra $L__BB108_1235; ld.local.v4.f32 {%f9464, %f9465, %f9466, %f9467}, [%rd271+148]; add.ftz.f32 %f16101, %f16101, %f9464; add.ftz.f32 %f16102, %f16102, %f9465; add.ftz.f32 %f16103, %f16103, %f9466; $L__BB108_1235: ld.local.u32 %r1855, [%rd271+240]; add.s32 %r1856, %r1855, -21; setp.gt.u32 %p1165, %r1856, 9; @%p1165 bra $L__BB108_1237; ld.local.v4.f32 {%f9471, %f9472, %f9473, %f9474}, [%rd271+228]; add.ftz.f32 %f16101, %f16101, %f9471; add.ftz.f32 %f16102, %f16102, %f9472; add.ftz.f32 %f16103, %f16103, %f9473; $L__BB108_1237: cvt.u32.u64 %r1857, %rd270; add.s32 %r3218, %r1857, 4; add.s32 %r3217, %r3217, -4; setp.ne.s32 %p1166, %r3217, 0; @%p1166 bra $L__BB108_1229; $L__BB108_1238: setp.eq.s32 %p1167, %r3220, 0; @%p1167 bra $L__BB108_1242; $L__BB108_1239: .pragma "nounroll"; cvt.s64.s32 %rd272, %r3218; mul.wide.s32 %rd1636, %r3218, 80; add.s64 %rd1637, %rd1, %rd1636; add.s64 %rd273, %rd1637, 316; ld.local.u32 %r1858, [%rd1637+316]; add.s32 %r1859, %r1858, -21; setp.gt.u32 %p1168, %r1859, 9; @%p1168 bra $L__BB108_1241; ld.local.f32 %f9478, [%rd273+-12]; add.ftz.f32 %f16101, %f16101, %f9478; ld.local.f32 %f9479, [%rd273+-8]; add.ftz.f32 %f16102, %f16102, %f9479; ld.local.f32 %f9480, [%rd273+-4]; add.ftz.f32 %f16103, %f16103, %f9480; $L__BB108_1241: cvt.u32.u64 %r1860, %rd272; add.s32 %r3218, %r1860, 1; add.s32 %r3220, %r3220, -1; setp.ne.s32 %p1169, %r3220, 0; @%p1169 bra $L__BB108_1239; $L__BB108_1242: cvta.to.global.u64 %rd1638, %rd618; cvt.s64.s32 %rd1639, %r572; add.s64 %rd1640, %rd166, %rd1639; shl.b64 %rd1641, %rd1640, 2; add.s64 %rd1642, %rd1638, %rd1641; mul.ftz.f32 %f9481, %f1750, %f16101; atom.global.add.f32 %f9482, [%rd1642], %f9481; add.s64 %rd1643, %rd1642, 4; mul.ftz.f32 %f9483, %f1751, %f16102; atom.global.add.f32 %f9484, [%rd1643], %f9483; add.s64 %rd1644, %rd1642, 8; mul.ftz.f32 %f9485, %f1752, %f16103; atom.global.add.f32 %f9486, [%rd1644], %f9485; $L__BB108_1243: ld.const.u32 %r3055, [kernel_params+1256]; and.b32 %r1861, %r3055, 131072; setp.eq.s32 %p1170, %r1861, 0; @%p1170 bra $L__BB108_1257; ld.const.f32 %f1798, [kernel_params+1412]; ld.const.f32 %f1799, [kernel_params+1416]; ld.local.v4.f32 {%f9487, %f9488, %f9489, %f9490}, [%rd1]; ld.const.v4.f32 {%f9494, %f9495, %f9496, %f9497}, [kernel_params+96]; ld.const.v4.f32 {%f9498, %f9499, %f9500, %f9501}, [kernel_params+112]; ld.const.v4.f32 {%f9502, %f9503, %f9504, %f9505}, [kernel_params+128]; ld.const.u32 %r1862, [kernel_params+16]; setp.eq.s32 %p1171, %r1862, 1; sub.ftz.f32 %f1812, %f9487, %f9497; sub.ftz.f32 %f1813, %f9488, %f9501; sub.ftz.f32 %f1814, %f9489, %f9505; @%p1171 bra $L__BB108_1246; bra.uni $L__BB108_1245; $L__BB108_1246: mul.ftz.f32 %f9509, %f9500, %f1813; fma.rn.ftz.f32 %f9510, %f9496, %f1812, %f9509; fma.rn.ftz.f32 %f9511, %f9504, %f1814, %f9510; abs.ftz.f32 %f16128, %f9511; bra.uni $L__BB108_1247; $L__BB108_1245: mul.ftz.f32 %f9506, %f1813, %f1813; fma.rn.ftz.f32 %f9507, %f1812, %f1812, %f9506; fma.rn.ftz.f32 %f9508, %f1814, %f1814, %f9507; sqrt.approx.ftz.f32 %f16128, %f9508; $L__BB108_1247: sub.ftz.f32 %f9512, %f16128, %f1798; mul.ftz.f32 %f9513, %f1799, %f9512; cvt.ftz.sat.f32.f32 %f16129, %f9513; ld.const.f32 %f1819, [kernel_params+1420]; setp.eq.ftz.f32 %p1172, %f1819, 0f3F800000; @%p1172 bra $L__BB108_1253; setp.eq.ftz.f32 %p1173, %f1819, 0f40000000; @%p1173 bra $L__BB108_1252; bra.uni $L__BB108_1249; $L__BB108_1252: mul.ftz.f32 %f16129, %f16129, %f16129; bra.uni $L__BB108_1253; $L__BB108_1249: setp.eq.ftz.f32 %p1174, %f1819, 0f3F000000; @%p1174 bra $L__BB108_1251; bra.uni $L__BB108_1250; $L__BB108_1251: sqrt.approx.ftz.f32 %f16129, %f16129; bra.uni $L__BB108_1253; $L__BB108_1250: lg2.approx.ftz.f32 %f9514, %f16129; mul.ftz.f32 %f9515, %f1819, %f9514; ex2.approx.ftz.f32 %f16129, %f9515; $L__BB108_1253: ld.const.u64 %rd1647, [kernel_params+2584]; cvta.to.global.u64 %rd1648, %rd1647; mul.lo.s64 %rd1649, %rd2, 12; add.s64 %rd1650, %rd1648, %rd1649; ld.global.f32 %f1824, [%rd1650]; ld.global.f32 %f1825, [%rd1650+4]; ld.global.f32 %f1826, [%rd1650+8]; ld.local.u32 %r585, [%rd7+-36]; and.b32 %r1863, %r585, 524288; setp.ne.s32 %p1175, %r1863, 0; mov.f32 %f9518, 0f3F800000; mov.f32 %f16130, %f9518; mov.f32 %f16131, %f9518; mov.f32 %f16132, %f9518; @%p1175 bra $L__BB108_1256; and.b32 %r1864, %r585, 512; setp.eq.s32 %p1176, %r1864, 0; mov.f32 %f16130, 0f00000000; mov.f32 %f16131, %f16130; mov.f32 %f16132, %f16130; @%p1176 bra $L__BB108_1256; ld.local.v2.f32 {%f16131, %f16132}, [%rd7+208]; ld.local.f32 %f16130, [%rd7+204]; $L__BB108_1256: sub.ftz.f32 %f9525, %f9518, %f16130; sub.ftz.f32 %f9526, %f9518, %f16131; sub.ftz.f32 %f9527, %f9518, %f16132; cvt.ftz.sat.f32.f32 %f9528, %f9525; cvt.ftz.sat.f32.f32 %f9529, %f9526; cvt.ftz.sat.f32.f32 %f9530, %f9527; mul.ftz.f32 %f9531, %f1825, %f9529; fma.rn.ftz.f32 %f9532, %f1824, %f9528, %f9531; fma.rn.ftz.f32 %f9533, %f1826, %f9530, %f9532; mul.ftz.f32 %f9534, %f9533, 0f3EAAAAAB; sub.ftz.f32 %f9535, %f9518, %f16129; mul.ftz.f32 %f9536, %f9535, %f9534; ld.const.s32 %rd1651, [kernel_params+1408]; add.s64 %rd1652, %rd166, %rd1651; cvta.to.global.u64 %rd1653, %rd618; shl.b64 %rd1654, %rd1652, 2; add.s64 %rd1655, %rd1653, %rd1654; atom.global.add.f32 %f9537, [%rd1655], %f9536; $L__BB108_1257: ld.global.u32 %r586, [%rd29]; and.b32 %r1865, %r586, 16777216; setp.eq.s32 %p1177, %r1865, 0; @%p1177 bra $L__BB108_1312; ld.local.u8 %rs81, [%rd7+-34]; and.b16 %rs82, %rs81, 8; setp.ne.s16 %p1178, %rs82, 0; @%p1178 bra $L__BB108_1312; ld.const.u64 %rd1656, [kernel_params+2448]; cvta.to.global.u64 %rd1657, %rd1656; add.s64 %rd1659, %rd1657, %rd837; ld.global.u32 %rd1660, [%rd1659]; ld.const.s32 %rd1661, [kernel_params+1264]; mul.lo.s64 %rd274, %rd1661, %rd1660; ld.const.u32 %r587, [kernel_params+1432]; setp.eq.s32 %p1179, %r587, -1; @%p1179 bra $L__BB108_1264; ld.const.u64 %rd1662, [kernel_params+2616]; cvta.to.global.u64 %rd1663, %rd1662; mul.lo.s64 %rd1664, %rd2, 12; add.s64 %rd1665, %rd1663, %rd1664; ld.const.u64 %rd1666, [kernel_params+2648]; cvta.to.global.u64 %rd1667, %rd1666; add.s64 %rd1669, %rd1667, %rd837; ld.global.f32 %f9539, [%rd1669]; ld.local.f32 %f9540, [%rd7]; sub.ftz.f32 %f9541, %f9540, %f9539; ld.global.f32 %f9542, [%rd1665+4]; ld.global.f32 %f9543, [%rd1665]; add.ftz.f32 %f9544, %f9543, %f9542; ld.global.f32 %f9545, [%rd1665+8]; add.ftz.f32 %f9546, %f9545, %f9544; mul.ftz.f32 %f9547, %f9546, 0f3EAAAAAB; mul.ftz.f32 %f1833, %f9547, %f9541; mov.b32 %r588, %f1833; setp.nan.ftz.f32 %p1180, %f1833, %f1833; mov.f32 %f16133, 0f00000000; @%p1180 bra $L__BB108_1263; or.b32 %r1866, %r588, -2147483648; setp.ne.s32 %p1181, %r1866, -2147483648; add.ftz.f32 %f9549, %f1833, %f1833; setp.eq.ftz.f32 %p1182, %f9549, %f1833; and.pred %p1183, %p1182, %p1181; @%p1183 bra $L__BB108_1263; shl.b32 %r1867, %r588, 1; setp.lt.u32 %p1184, %r1867, -16777215; selp.f32 %f16133, %f1833, 0f00000000, %p1184; $L__BB108_1263: cvta.to.global.u64 %rd1670, %rd618; cvt.s64.s32 %rd1671, %r587; add.s64 %rd1672, %rd274, %rd1671; shl.b64 %rd1673, %rd1672, 2; add.s64 %rd1674, %rd1670, %rd1673; atom.global.add.f32 %f9550, [%rd1674], %f16133; $L__BB108_1264: ld.local.u32 %r589, [%rd7+172]; setp.lt.s32 %p1185, %r589, 1; mov.f32 %f16149, 0f00000000; mov.f32 %f16150, %f16149; mov.f32 %f16151, %f16149; mov.f32 %f16152, %f16149; mov.f32 %f16153, %f16149; mov.f32 %f16154, %f16149; mov.f32 %f16155, %f16149; mov.f32 %f16156, %f16149; mov.f32 %f16157, %f16149; mov.f32 %f16158, %f16149; mov.f32 %f16159, %f16149; @%p1185 bra $L__BB108_1285; mov.f32 %f16159, 0f00000000; mov.u32 %r3221, 0; mov.f32 %f16158, %f16159; mov.f32 %f16157, %f16159; mov.f32 %f16156, %f16159; mov.f32 %f16155, %f16159; mov.f32 %f16154, %f16159; mov.f32 %f16153, %f16159; mov.f32 %f16152, %f16159; mov.f32 %f16151, %f16159; mov.f32 %f16150, %f16159; mov.f32 %f16149, %f16159; bra.uni $L__BB108_1266; $L__BB108_1278: add.s32 %r1878, %r591, -10; setp.lt.u32 %p1195, %r1878, 7; @%p1195 bra $L__BB108_1280; add.s32 %r1879, %r591, -22; setp.gt.u32 %p1196, %r1879, 2; setp.ne.s32 %p1197, %r591, 27; mov.f32 %f16148, 0f3F800000; and.pred %p1198, %p1196, %p1197; @%p1198 bra $L__BB108_1281; $L__BB108_1280: ld.local.v2.f32 {%f9615, %f9616}, [%rd276+20]; mul.ftz.f32 %f16148, %f9615, %f9616; bra.uni $L__BB108_1281; $L__BB108_1270: setp.eq.s32 %p1189, %r591, 6; @%p1189 bra $L__BB108_1275; bra.uni $L__BB108_1271; $L__BB108_1275: ld.local.f32 %f9612, [%rd276+20]; mul.ftz.f32 %f16145, %f16145, %f9612; mul.ftz.f32 %f16146, %f16146, %f9612; mul.ftz.f32 %f16147, %f16147, %f9612; bra.uni $L__BB108_1276; $L__BB108_1271: setp.eq.s32 %p1190, %r591, 29; @%p1190 bra $L__BB108_1274; bra.uni $L__BB108_1272; $L__BB108_1274: ld.local.v4.f32 {%f9587, %f9588, %f9589, %f9590}, [%rd276+20]; sqrt.approx.ftz.f32 %f9595, %f9587; sqrt.approx.ftz.f32 %f9596, %f9588; sqrt.approx.ftz.f32 %f9597, %f9589; fma.rn.ftz.f32 %f9598, %f9590, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f9599, %f9590, %f9598, 0fC12BAE14; fma.rn.ftz.f32 %f9600, %f9590, %f9599, 0f40220C4A; fma.rn.ftz.f32 %f9601, %f9590, %f9600, 0fBE5C28F6; fma.rn.ftz.f32 %f9602, %f9590, %f9601, 0f40BF020C; mul.ftz.f32 %f9603, %f9595, %f9602; mul.ftz.f32 %f9604, %f9596, %f9602; mul.ftz.f32 %f9605, %f9597, %f9602; mul.ftz.f32 %f9606, %f9603, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9607, %f9606; mul.ftz.f32 %f9608, %f9604, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9609, %f9608; mul.ftz.f32 %f9610, %f9605, 0fBFB8AA3B; ex2.approx.ftz.f32 %f9611, %f9610; mul.ftz.f32 %f16145, %f16145, %f9607; mul.ftz.f32 %f16146, %f16146, %f9609; mul.ftz.f32 %f16147, %f16147, %f9611; bra.uni $L__BB108_1276; $L__BB108_1272: setp.ne.s32 %p1191, %r591, 5; @%p1191 bra $L__BB108_1276; ld.local.u32 %r1873, [%rd276+24]; setp.eq.s32 %p1192, %r1873, 8; @%p1192 bra $L__BB108_1284; bra.uni $L__BB108_1276; $L__BB108_1266: cvt.s64.s32 %rd275, %r3221; mul.wide.s32 %rd1675, %r3221, 80; add.s64 %rd1676, %rd1, %rd1675; add.s64 %rd276, %rd1676, 316; ld.local.u32 %r591, [%rd1676+316]; setp.gt.s32 %p1186, %r591, 34; @%p1186 bra $L__BB108_1284; ld.local.v4.f32 {%f9573, %f9574, %f9575, %f9576}, [%rd276+4]; fma.rn.ftz.f32 %f16155, %f9573, %f9574, %f16155; fma.rn.ftz.f32 %f16156, %f9573, %f9575, %f16156; fma.rn.ftz.f32 %f16157, %f9573, %f9576, %f16157; add.ftz.f32 %f16158, %f16158, %f9573; ld.local.v4.f32 {%f16145, %f16146, %f16147, %f9583}, [%rd276+-12]; setp.gt.u32 %p1187, %r591, 27; @%p1187 bra $L__BB108_1270; mov.u32 %r1870, 1; shl.b32 %r1871, %r1870, %r591; and.b32 %r1872, %r1871, 134256640; setp.ne.s32 %p1188, %r1872, 0; @%p1188 bra $L__BB108_1269; bra.uni $L__BB108_1270; $L__BB108_1269: ld.local.u64 %rd1677, [%rd276+36]; ld.f32 %f9584, [%rd1677+24]; mul.ftz.f32 %f16145, %f16145, %f9584; ld.f32 %f9585, [%rd1677+28]; mul.ftz.f32 %f16146, %f16146, %f9585; ld.f32 %f9586, [%rd1677+32]; mul.ftz.f32 %f16147, %f16147, %f9586; $L__BB108_1276: mov.f32 %f16148, 0f00000000; setp.gt.u32 %p1193, %r591, 31; @%p1193 bra $L__BB108_1278; mov.u32 %r1875, 1; shl.b32 %r1876, %r1875, %r591; and.b32 %r1877, %r1876, -2145385984; setp.ne.s32 %p1194, %r1877, 0; @%p1194 bra $L__BB108_1281; bra.uni $L__BB108_1278; $L__BB108_1281: setp.gt.ftz.f32 %p1199, %f16148, 0f3BB851EC; @%p1199 bra $L__BB108_1283; bra.uni $L__BB108_1282; $L__BB108_1283: add.ftz.f32 %f16152, %f16152, %f16145; add.ftz.f32 %f16153, %f16153, %f16146; add.ftz.f32 %f16154, %f16154, %f16147; add.ftz.f32 %f16159, %f16159, %f9573; bra.uni $L__BB108_1284; $L__BB108_1282: add.ftz.f32 %f16149, %f16149, %f16145; add.ftz.f32 %f16150, %f16150, %f16146; add.ftz.f32 %f16151, %f16151, %f16147; $L__BB108_1284: cvt.u32.u64 %r1880, %rd275; add.s32 %r3221, %r1880, 1; setp.lt.s32 %p1200, %r3221, %r589; @%p1200 bra $L__BB108_1266; $L__BB108_1285: mul.ftz.f32 %f9619, %f16159, 0f40800000; setp.gt.ftz.f32 %p1201, %f9619, %f16158; setp.eq.ftz.f32 %p1202, %f16158, 0f00000000; or.pred %p1203, %p1202, %p1201; @%p1203 bra $L__BB108_1287; bra.uni $L__BB108_1286; $L__BB108_1287: @%p1202 bra $L__BB108_1289; rcp.approx.ftz.f32 %f9626, %f16158; mul.ftz.f32 %f16155, %f16155, %f9626; mul.ftz.f32 %f16156, %f16156, %f9626; mul.ftz.f32 %f16157, %f16157, %f9626; $L__BB108_1289: ld.const.u32 %r593, [kernel_params+1424]; setp.eq.s32 %p1205, %r593, -1; @%p1205 bra $L__BB108_1300; ld.const.v4.f32 {%f9628, %f9629, %f9630, %f9631}, [kernel_params+688]; ld.const.v4.f32 {%f9635, %f9636, %f9637, %f9638}, [kernel_params+704]; ld.const.v4.f32 {%f9642, %f9643, %f9644, %f9645}, [kernel_params+720]; mul.ftz.f32 %f9649, %f16156, %f9629; fma.rn.ftz.f32 %f9650, %f16155, %f9628, %f9649; fma.rn.ftz.f32 %f1904, %f16157, %f9630, %f9650; mul.ftz.f32 %f9651, %f16156, %f9636; fma.rn.ftz.f32 %f9652, %f16155, %f9635, %f9651; fma.rn.ftz.f32 %f1905, %f16157, %f9637, %f9652; mul.ftz.f32 %f9653, %f16156, %f9643; fma.rn.ftz.f32 %f9654, %f16155, %f9642, %f9653; fma.rn.ftz.f32 %f1906, %f16157, %f9644, %f9654; mov.b32 %r594, %f1904; setp.nan.ftz.f32 %p1206, %f1904, %f1904; mov.f32 %f16176, 0f00000000; mov.f32 %f16174, %f16176; @%p1206 bra $L__BB108_1293; or.b32 %r1881, %r594, -2147483648; setp.ne.s32 %p1207, %r1881, -2147483648; add.ftz.f32 %f9656, %f1904, %f1904; setp.eq.ftz.f32 %p1208, %f9656, %f1904; and.pred %p1209, %p1208, %p1207; mov.f32 %f16174, %f16176; @%p1209 bra $L__BB108_1293; shl.b32 %r1882, %r594, 1; setp.lt.u32 %p1210, %r1882, -16777215; selp.f32 %f16174, %f1904, 0f00000000, %p1210; $L__BB108_1293: mov.b32 %r595, %f1905; setp.nan.ftz.f32 %p1211, %f1905, %f1905; mov.f32 %f16175, %f16176; @%p1211 bra $L__BB108_1296; or.b32 %r1883, %r595, -2147483648; setp.ne.s32 %p1212, %r1883, -2147483648; add.ftz.f32 %f9659, %f1905, %f1905; setp.eq.ftz.f32 %p1213, %f9659, %f1905; and.pred %p1214, %p1213, %p1212; mov.f32 %f16175, %f16176; @%p1214 bra $L__BB108_1296; shl.b32 %r1884, %r595, 1; setp.lt.u32 %p1215, %r1884, -16777215; selp.f32 %f16175, %f1905, 0f00000000, %p1215; $L__BB108_1296: mov.b32 %r596, %f1906; setp.nan.ftz.f32 %p1216, %f1906, %f1906; @%p1216 bra $L__BB108_1299; or.b32 %r1885, %r596, -2147483648; setp.ne.s32 %p1217, %r1885, -2147483648; add.ftz.f32 %f9662, %f1906, %f1906; setp.eq.ftz.f32 %p1218, %f9662, %f1906; and.pred %p1219, %p1218, %p1217; @%p1219 bra $L__BB108_1299; shl.b32 %r1886, %r596, 1; setp.lt.u32 %p1220, %r1886, -16777215; selp.f32 %f16176, %f1906, 0f00000000, %p1220; $L__BB108_1299: cvta.to.global.u64 %rd1682, %rd618; cvt.s64.s32 %rd1683, %r593; add.s64 %rd1684, %rd274, %rd1683; shl.b64 %rd1685, %rd1684, 2; add.s64 %rd1686, %rd1682, %rd1685; add.s64 %rd1687, %rd1686, 4; add.s64 %rd1688, %rd1686, 8; atom.global.add.f32 %f9663, [%rd1686], %f16174; atom.global.add.f32 %f9664, [%rd1687], %f16175; atom.global.add.f32 %f9665, [%rd1688], %f16176; $L__BB108_1300: ld.const.u32 %r597, [kernel_params+1428]; setp.eq.s32 %p1221, %r597, -1; @%p1221 bra $L__BB108_1311; ld.const.u64 %rd1689, [kernel_params+2616]; cvta.to.global.u64 %rd1690, %rd1689; mul.lo.s64 %rd1691, %rd2, 12; add.s64 %rd1692, %rd1690, %rd1691; ld.global.f32 %f9667, [%rd1692]; mul.ftz.f32 %f1913, %f16152, %f9667; ld.global.f32 %f9668, [%rd1692+4]; mul.ftz.f32 %f1914, %f16153, %f9668; ld.global.f32 %f9669, [%rd1692+8]; mul.ftz.f32 %f1915, %f16154, %f9669; mov.b32 %r598, %f1913; setp.nan.ftz.f32 %p1222, %f1913, %f1913; mov.f32 %f16179, 0f00000000; mov.f32 %f16177, %f16179; @%p1222 bra $L__BB108_1304; or.b32 %r1887, %r598, -2147483648; setp.ne.s32 %p1223, %r1887, -2147483648; add.ftz.f32 %f9671, %f1913, %f1913; setp.eq.ftz.f32 %p1224, %f9671, %f1913; and.pred %p1225, %p1224, %p1223; mov.f32 %f16177, %f16179; @%p1225 bra $L__BB108_1304; shl.b32 %r1888, %r598, 1; setp.lt.u32 %p1226, %r1888, -16777215; selp.f32 %f16177, %f1913, 0f00000000, %p1226; $L__BB108_1304: mov.b32 %r599, %f1914; setp.nan.ftz.f32 %p1227, %f1914, %f1914; mov.f32 %f16178, %f16179; @%p1227 bra $L__BB108_1307; or.b32 %r1889, %r599, -2147483648; setp.ne.s32 %p1228, %r1889, -2147483648; add.ftz.f32 %f9674, %f1914, %f1914; setp.eq.ftz.f32 %p1229, %f9674, %f1914; and.pred %p1230, %p1229, %p1228; mov.f32 %f16178, %f16179; @%p1230 bra $L__BB108_1307; shl.b32 %r1890, %r599, 1; setp.lt.u32 %p1231, %r1890, -16777215; selp.f32 %f16178, %f1914, 0f00000000, %p1231; $L__BB108_1307: mov.b32 %r600, %f1915; setp.nan.ftz.f32 %p1232, %f1915, %f1915; @%p1232 bra $L__BB108_1310; or.b32 %r1891, %r600, -2147483648; setp.ne.s32 %p1233, %r1891, -2147483648; add.ftz.f32 %f9677, %f1915, %f1915; setp.eq.ftz.f32 %p1234, %f9677, %f1915; and.pred %p1235, %p1234, %p1233; @%p1235 bra $L__BB108_1310; shl.b32 %r1892, %r600, 1; setp.lt.u32 %p1236, %r1892, -16777215; selp.f32 %f16179, %f1915, 0f00000000, %p1236; $L__BB108_1310: cvta.to.global.u64 %rd1693, %rd618; cvt.s64.s32 %rd1694, %r597; add.s64 %rd1695, %rd274, %rd1694; shl.b64 %rd1696, %rd1695, 2; add.s64 %rd1697, %rd1693, %rd1696; add.s64 %rd1698, %rd1697, 4; add.s64 %rd1699, %rd1697, 8; atom.global.add.f32 %f9678, [%rd1697], %f16177; atom.global.add.f32 %f9679, [%rd1698], %f16178; atom.global.add.f32 %f9680, [%rd1699], %f16179; $L__BB108_1311: and.b32 %r1893, %r586, -16777217; st.global.u32 [%rd29], %r1893; bra.uni $L__BB108_1312; $L__BB108_1286: ld.const.u64 %rd1678, [kernel_params+2616]; cvta.to.global.u64 %rd1679, %rd1678; mul.lo.s64 %rd1680, %rd2, 12; add.s64 %rd1681, %rd1679, %rd1680; ld.global.f32 %f9620, [%rd1681]; mul.ftz.f32 %f9621, %f16149, %f9620; ld.global.f32 %f9622, [%rd1681+4]; mul.ftz.f32 %f9623, %f16150, %f9622; ld.global.f32 %f9624, [%rd1681+8]; mul.ftz.f32 %f9625, %f16151, %f9624; st.global.f32 [%rd1681], %f9621; st.global.f32 [%rd1681+4], %f9623; st.global.f32 [%rd1681+8], %f9625; bra.uni $L__BB108_1312; $L__BB108_877: add.s32 %r1495, %r426, -10; setp.lt.u32 %p751, %r1495, 7; @%p751 bra $L__BB108_879; add.s32 %r1496, %r426, -22; setp.gt.u32 %p752, %r1496, 2; setp.ne.s32 %p753, %r426, 27; mov.f32 %f15946, 0f3F800000; and.pred %p754, %p752, %p753; @%p754 bra $L__BB108_882; $L__BB108_879: ld.local.v2.f32 {%f7034, %f7035}, [%rd173+20]; mul.ftz.f32 %f15946, %f7034, %f7035; bra.uni $L__BB108_882; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) { .reg .pred %p<2>; .reg .b32 %r<5>; ld.param.u32 %r1, [_Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u8 %r2, [_Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0+4]; setp.eq.s32 %p1, %r2, 39; selp.b32 %r3, 5, 1, %p1; add.s32 %r4, %r3, %r1; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<535>; .reg .b32 %r<242>; .reg .b64 %rd<290>; ld.param.u64 %rd63, [_Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd64, [_Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd65, %rd63; add.s64 %rd1, %rd65, 76; ld.local.u32 %r3, [%rd65+76]; cvt.s64.s32 %rd2, %r3; setp.eq.s32 %p1, %r3, -1; mov.u32 %r29, 0; mov.u32 %r30, %r29; mov.u32 %r236, %r64; mov.u32 %r32, %r29; @%p1 bra $L__BB110_16; cvt.u64.u32 %rd3, %r62; ld.local.s32 %rd4, [%rd1+-16]; ld.const.u64 %rd66, [kernel_params+2160]; cvta.to.global.u64 %rd67, %rd66; mul.lo.s64 %rd68, %rd2, 208; add.s64 %rd69, %rd67, %rd68; ld.global.u32 %r4, [%rd69+164]; ld.local.u32 %r69, [%rd1+-12]; and.b32 %r70, %r69, 1; setp.eq.b32 %p2, %r70, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r225, 0; @%p5 bra $L__BB110_5; cvt.u32.u64 %r72, %rd4; setp.eq.s32 %p6, %r72, -1; mov.u32 %r224, -1; @%p6 bra $L__BB110_4; ld.const.u64 %rd70, [kernel_params+2240]; cvta.to.global.u64 %rd71, %rd70; shl.b64 %rd72, %rd4, 2; add.s64 %rd73, %rd71, %rd72; ld.global.u32 %r224, [%rd73]; $L__BB110_4: setp.ne.s32 %p7, %r224, -1; selp.u32 %r225, 1, 0, %p7; $L__BB110_5: ld.const.u64 %rd74, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd74; add.s32 %r227, %r225, %r4; mul.wide.u32 %rd75, %r227, 16; add.s64 %rd76, %rd5, %rd75; add.s64 %rd276, %rd76, 14; add.s64 %rd277, %rd76, 15; ld.global.u32 %r226, [%rd76+8]; ld.global.u16 %rs37, [%rd76+12]; ld.global.u64 %rd275, [%rd76]; setp.eq.s64 %p8, %rd275, %rd3; @%p8 bra $L__BB110_11; $L__BB110_6: setp.eq.s64 %p9, %rd275, 0; @%p9 bra $L__BB110_8; add.s32 %r227, %r227, 2; bra.uni $L__BB110_9; $L__BB110_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r233, -1; mov.u32 %r230, 0; mov.u32 %r227, %r226; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p10 bra $L__BB110_15; $L__BB110_9: cvt.u64.u32 %rd10, %r227; mul.wide.u32 %rd77, %r227, 16; add.s64 %rd78, %rd5, %rd77; ld.global.u32 %r226, [%rd78+8]; ld.global.u16 %rs37, [%rd78+12]; ld.global.u64 %rd275, [%rd78]; setp.ne.s64 %p11, %rd275, %rd3; @%p11 bra $L__BB110_6; shl.b64 %rd79, %rd10, 4; add.s64 %rd80, %rd5, %rd79; add.s64 %rd276, %rd80, 14; add.s64 %rd277, %rd80, 15; $L__BB110_11: ld.global.u8 %rs5, [%rd277]; ld.global.u8 %rs6, [%rd276]; cvt.u32.u64 %r77, %rd4; setp.ne.s32 %p12, %r77, -1; @%p12 bra $L__BB110_14; mov.u32 %r233, -1; mov.u32 %r230, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB110_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p14 bra $L__BB110_15; $L__BB110_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r233, -1, %r226, %p15; cvt.u32.u16 %r230, %rs37; cvt.u32.u16 %r82, %rs5; and.b32 %r232, %r82, 255; cvt.u32.u16 %r83, %rs6; and.b32 %r231, %r83, 255; $L__BB110_15: setp.eq.s32 %p16, %r233, -1; selp.b32 %r29, 0, %r233, %p16; selp.b32 %r30, 0, %r232, %p16; selp.b32 %r236, %r64, %r231, %p16; selp.b32 %r32, 0, %r230, %p16; $L__BB110_16: setp.eq.s32 %p17, %r32, 0; setp.eq.s32 %p18, %r62, 7; and.pred %p19, %p18, %p17; cvta.to.local.u64 %rd81, %rd64; mul.wide.u32 %rd82, %r63, 4; add.s64 %rd16, %rd81, %rd82; @%p19 bra $L__BB110_148; bra.uni $L__BB110_17; $L__BB110_148: cvt.u32.u64 %r222, %rd2; setp.eq.s32 %p103, %r222, -1; ld.local.v4.f32 {%f534, %f533, %f532, %f426}, [%rd65]; @%p103 bra $L__BB110_152; ld.local.u8 %rs35, [%rd65+56]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p104, %rs36, 0; @%p104 bra $L__BB110_151; bra.uni $L__BB110_150; $L__BB110_151: ld.const.u64 %rd271, [kernel_params+2160]; cvta.to.global.u64 %rd272, %rd271; mul.wide.s32 %rd273, %r3, 208; add.s64 %rd274, %rd272, %rd273; ld.global.v4.f32 {%f460, %f461, %f462, %f463}, [%rd274+48]; ld.global.v4.f32 {%f468, %f469, %f470, %f471}, [%rd274+64]; ld.global.v4.f32 {%f476, %f477, %f478, %f479}, [%rd274+80]; mul.ftz.f32 %f484, %f533, %f461; fma.rn.ftz.f32 %f485, %f534, %f460, %f484; fma.rn.ftz.f32 %f486, %f532, %f462, %f485; add.ftz.f32 %f173, %f463, %f486; mul.ftz.f32 %f487, %f533, %f469; fma.rn.ftz.f32 %f488, %f534, %f468, %f487; fma.rn.ftz.f32 %f489, %f532, %f470, %f488; add.ftz.f32 %f174, %f471, %f489; mul.ftz.f32 %f490, %f533, %f477; fma.rn.ftz.f32 %f491, %f534, %f476, %f490; fma.rn.ftz.f32 %f492, %f532, %f478, %f491; add.ftz.f32 %f532, %f479, %f492; mov.f32 %f533, %f174; mov.f32 %f534, %f173; bra.uni $L__BB110_152; $L__BB110_17: setp.eq.s32 %p20, %r236, 0; @%p20 bra $L__BB110_123; setp.eq.s32 %p21, %r236, 1; @%p21 bra $L__BB110_98; bra.uni $L__BB110_19; $L__BB110_98: mov.u32 %r48, 1; ld.local.u32 %r51, [%rd65+64]; and.b32 %r153, %r51, 1; setp.eq.b32 %p65, %r153, 1; mov.pred %p66, 0; xor.pred %p67, %p65, %p66; not.pred %p68, %p67; @%p68 bra $L__BB110_109; bra.uni $L__BB110_99; $L__BB110_109: and.b32 %r173, %r51, 6; setp.eq.s32 %p75, %r173, 0; @%p75 bra $L__BB110_116; and.b32 %r174, %r32, 768; setp.eq.s32 %p76, %r174, 0; @%p76 bra $L__BB110_112; ld.const.u64 %rd205, [kernel_params+2264]; cvta.to.global.u64 %rd206, %rd205; ld.local.u32 %r175, [%rd65+60]; mul.wide.s32 %rd208, %r175, 16; add.s64 %rd209, %rd206, %rd208; shr.s32 %r176, %r51, 7; ld.global.u32 %r177, [%rd209+4]; add.s32 %r178, %r176, %r177; ld.const.u64 %rd210, [kernel_params+2328]; cvta.to.global.u64 %rd211, %rd210; add.s32 %r179, %r178, %r29; mul.wide.s32 %rd212, %r179, 8; add.s64 %rd213, %rd211, %rd212; ld.global.v2.f32 {%f383, %f384}, [%rd213]; ld.global.v2.f32 {%f387, %f388}, [%rd213+8]; ld.local.f32 %f391, [%rd65+68]; mov.f32 %f392, 0f3F800000; sub.ftz.f32 %f393, %f392, %f391; mul.ftz.f32 %f394, %f383, %f393; mul.ftz.f32 %f395, %f384, %f393; fma.rn.ftz.f32 %f530, %f388, %f391, %f395; fma.rn.ftz.f32 %f529, %f387, %f391, %f394; bra.uni $L__BB110_118; $L__BB110_123: ld.local.u32 %r58, [%rd65+64]; and.b32 %r187, %r58, 1; setp.eq.b32 %p84, %r187, 1; mov.pred %p85, 0; xor.pred %p86, %p84, %p85; not.pred %p87, %p86; @%p87 bra $L__BB110_134; bra.uni $L__BB110_124; $L__BB110_134: and.b32 %r209, %r58, 6; setp.eq.s32 %p94, %r209, 0; @%p94 bra $L__BB110_141; and.b32 %r210, %r32, 768; setp.eq.s32 %p95, %r210, 0; @%p95 bra $L__BB110_137; ld.const.u64 %rd249, [kernel_params+2264]; cvta.to.global.u64 %rd250, %rd249; ld.local.u32 %r211, [%rd65+60]; mul.wide.s32 %rd252, %r211, 16; add.s64 %rd253, %rd250, %rd252; shr.s32 %r212, %r58, 7; ld.global.u32 %r213, [%rd253+4]; add.s32 %r214, %r212, %r213; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r215, %r214, %r29; mul.wide.s32 %rd256, %r215, 4; add.s64 %rd257, %rd255, %rd256; ld.local.f32 %f415, [%rd65+68]; mov.f32 %f416, 0f3F800000; sub.ftz.f32 %f417, %f416, %f415; ld.global.f32 %f418, [%rd257]; mul.ftz.f32 %f419, %f418, %f417; ld.global.f32 %f420, [%rd257+4]; fma.rn.ftz.f32 %f531, %f420, %f415, %f419; bra.uni $L__BB110_143; $L__BB110_150: ld.local.v4.f32 {%f427, %f428, %f429, %f430}, [%rd65+192]; mul.ftz.f32 %f435, %f533, %f428; fma.rn.ftz.f32 %f436, %f534, %f427, %f435; fma.rn.ftz.f32 %f437, %f532, %f429, %f436; add.ftz.f32 %f170, %f430, %f437; ld.local.v4.f32 {%f438, %f439, %f440, %f441}, [%rd65+208]; mul.ftz.f32 %f446, %f533, %f439; fma.rn.ftz.f32 %f447, %f534, %f438, %f446; fma.rn.ftz.f32 %f448, %f532, %f440, %f447; add.ftz.f32 %f171, %f441, %f448; ld.local.v4.f32 {%f449, %f450, %f451, %f452}, [%rd65+224]; mul.ftz.f32 %f457, %f533, %f450; fma.rn.ftz.f32 %f458, %f534, %f449, %f457; fma.rn.ftz.f32 %f459, %f532, %f451, %f458; add.ftz.f32 %f532, %f452, %f459; mov.f32 %f533, %f171; mov.f32 %f534, %f170; $L__BB110_152: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB110_156; bra.uni $L__BB110_153; $L__BB110_156: add.ftz.f32 %f493, %f533, %f534; add.ftz.f32 %f494, %f532, %f493; mul.ftz.f32 %f495, %f494, 0f3EAAAAAB; st.local.f32 [%rd16], %f495; bra.uni $L__BB110_157; $L__BB110_153: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB110_155; mov.u32 %r223, 1065353216; st.local.u32 [%rd16], %r223; bra.uni $L__BB110_157; $L__BB110_19: add.s32 %r87, %r236, -3; setp.lt.u32 %p22, %r87, 2; ld.local.u32 %r41, [%rd65+64]; and.b32 %r42, %r41, 1; @%p22 bra $L__BB110_45; bra.uni $L__BB110_20; $L__BB110_45: setp.eq.s32 %p39, %r42, 0; @%p39 bra $L__BB110_84; bra.uni $L__BB110_46; $L__BB110_84: and.b32 %r141, %r41, 6; setp.eq.s32 %p56, %r141, 0; @%p56 bra $L__BB110_91; and.b32 %r142, %r32, 768; setp.eq.s32 %p57, %r142, 0; @%p57 bra $L__BB110_87; ld.const.u64 %rd158, [kernel_params+2264]; cvta.to.global.u64 %rd159, %rd158; ld.local.u32 %r143, [%rd65+60]; mul.wide.s32 %rd161, %r143, 16; add.s64 %rd162, %rd159, %rd161; shr.s32 %r144, %r41, 7; ld.global.u32 %r145, [%rd162+4]; add.s32 %r146, %r144, %r145; ld.const.u64 %rd163, [kernel_params+2344]; cvta.to.global.u64 %rd164, %rd163; add.s32 %r147, %r146, %r29; mul.wide.s32 %rd165, %r147, 16; add.s64 %rd166, %rd164, %rd165; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd166]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd166+16]; ld.local.f32 %f332, [%rd65+68]; mov.f32 %f333, 0f3F800000; sub.ftz.f32 %f334, %f333, %f332; mul.ftz.f32 %f335, %f316, %f334; mul.ftz.f32 %f336, %f317, %f334; mul.ftz.f32 %f337, %f318, %f334; mul.ftz.f32 %f338, %f319, %f334; fma.rn.ftz.f32 %f522, %f327, %f332, %f338; fma.rn.ftz.f32 %f521, %f326, %f332, %f337; fma.rn.ftz.f32 %f520, %f325, %f332, %f336; fma.rn.ftz.f32 %f519, %f324, %f332, %f335; bra.uni $L__BB110_93; $L__BB110_124: ld.local.s32 %rd53, [%rd65+60]; setp.eq.s64 %p88, %rd53, -1; @%p88 bra $L__BB110_127; ld.const.u64 %rd226, [kernel_params+2240]; cvta.to.global.u64 %rd227, %rd226; shl.b64 %rd228, %rd53, 2; add.s64 %rd229, %rd227, %rd228; ld.global.u32 %r188, [%rd229]; setp.eq.s32 %p89, %r188, -1; @%p89 bra $L__BB110_127; mov.u32 %r189, 0; mov.u64 %rd230, 0; { // callseq 856, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r189; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd230; .param .b64 param3; st.param.b64 [param3+0], %rd230; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f531, [retval0+0]; } // callseq 856 bra.uni $L__BB110_143; $L__BB110_155: st.local.f32 [%rd16], %f534; st.local.f32 [%rd16+4], %f533; st.local.f32 [%rd16+8], %f532; bra.uni $L__BB110_157; $L__BB110_99: ld.local.s32 %rd51, [%rd65+60]; setp.eq.s64 %p69, %rd51, -1; @%p69 bra $L__BB110_102; ld.const.u64 %rd179, [kernel_params+2240]; cvta.to.global.u64 %rd180, %rd179; shl.b64 %rd181, %rd51, 2; add.s64 %rd182, %rd180, %rd181; ld.global.u32 %r154, [%rd182]; setp.eq.s32 %p70, %r154, -1; @%p70 bra $L__BB110_102; mov.u64 %rd183, 0; { // callseq 855, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r48; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd183; .param .b64 param3; st.param.b64 [param3+0], %rd183; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f529, [retval0+0]; ld.param.f32 %f530, [retval0+4]; } // callseq 855 bra.uni $L__BB110_118; $L__BB110_20: setp.eq.s32 %p23, %r42, 0; @%p23 bra $L__BB110_31; bra.uni $L__BB110_21; $L__BB110_31: and.b32 %r109, %r41, 6; setp.eq.s32 %p30, %r109, 0; @%p30 bra $L__BB110_38; and.b32 %r110, %r32, 768; setp.eq.s32 %p31, %r110, 0; @%p31 bra $L__BB110_34; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r111, [%rd65+60]; mul.wide.s32 %rd112, %r111, 16; add.s64 %rd113, %rd110, %rd112; shr.s32 %r112, %r41, 7; ld.global.u32 %r113, [%rd113+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd114, [kernel_params+2336]; cvta.to.global.u64 %rd115, %rd114; add.s32 %r115, %r114, %r29; mul.wide.s32 %rd116, %r115, 12; add.s64 %rd117, %rd115, %rd116; ld.local.f32 %f202, [%rd65+68]; mov.f32 %f203, 0f3F800000; sub.ftz.f32 %f204, %f203, %f202; ld.global.f32 %f205, [%rd117]; mul.ftz.f32 %f206, %f205, %f204; ld.global.f32 %f207, [%rd117+4]; mul.ftz.f32 %f208, %f207, %f204; ld.global.f32 %f209, [%rd117+8]; mul.ftz.f32 %f210, %f209, %f204; ld.global.f32 %f211, [%rd117+12]; ld.global.f32 %f212, [%rd117+16]; ld.global.f32 %f213, [%rd117+20]; fma.rn.ftz.f32 %f498, %f213, %f202, %f210; fma.rn.ftz.f32 %f497, %f212, %f202, %f208; fma.rn.ftz.f32 %f496, %f211, %f202, %f206; bra.uni $L__BB110_40; $L__BB110_127: and.b32 %r190, %r32, 56; setp.eq.s32 %p90, %r190, 0; @%p90 bra $L__BB110_132; and.b32 %r191, %r32, 24; setp.eq.s32 %p91, %r191, 0; @%p91 bra $L__BB110_130; ld.const.u64 %rd231, [kernel_params+2232]; cvta.to.global.u64 %rd232, %rd231; shl.b64 %rd233, %rd53, 4; add.s64 %rd234, %rd232, %rd233; ld.global.v4.u32 {%r192, %r193, %r194, %r195}, [%rd234]; ld.const.u64 %rd235, [kernel_params+2320]; cvta.to.global.u64 %rd236, %rd235; add.s32 %r199, %r192, %r29; mul.wide.u32 %rd237, %r199, 4; add.s64 %rd287, %rd236, %rd237; add.s32 %r200, %r193, %r29; mul.wide.u32 %rd238, %r200, 4; add.s64 %rd288, %rd236, %rd238; add.s32 %r201, %r194, %r29; mul.wide.u32 %rd239, %r201, 4; add.s64 %rd289, %rd236, %rd239; bra.uni $L__BB110_131; $L__BB110_102: and.b32 %r155, %r32, 56; setp.eq.s32 %p71, %r155, 0; @%p71 bra $L__BB110_107; and.b32 %r156, %r32, 24; setp.eq.s32 %p72, %r156, 0; @%p72 bra $L__BB110_105; ld.const.u64 %rd184, [kernel_params+2232]; cvta.to.global.u64 %rd185, %rd184; shl.b64 %rd186, %rd51, 4; add.s64 %rd187, %rd185, %rd186; ld.global.v4.u32 {%r157, %r158, %r159, %r160}, [%rd187]; ld.const.u64 %rd188, [kernel_params+2328]; cvta.to.global.u64 %rd189, %rd188; add.s32 %r164, %r157, %r29; mul.wide.u32 %rd190, %r164, 8; add.s64 %rd191, %rd189, %rd190; ld.global.v2.f32 {%f527, %f528}, [%rd191]; add.s32 %r165, %r158, %r29; mul.wide.u32 %rd192, %r165, 8; add.s64 %rd193, %rd189, %rd192; ld.global.v2.f32 {%f525, %f526}, [%rd193]; add.s32 %r166, %r159, %r29; mul.wide.u32 %rd194, %r166, 8; add.s64 %rd195, %rd189, %rd194; ld.global.v2.f32 {%f523, %f524}, [%rd195]; bra.uni $L__BB110_106; $L__BB110_46: ld.local.s32 %rd47, [%rd65+60]; setp.eq.s64 %p40, %rd47, -1; @%p40 bra $L__BB110_49; ld.const.u64 %rd129, [kernel_params+2240]; cvta.to.global.u64 %rd130, %rd129; shl.b64 %rd131, %rd47, 2; add.s64 %rd132, %rd130, %rd131; ld.global.u32 %r122, [%rd132]; setp.eq.s32 %p41, %r122, -1; @%p41 bra $L__BB110_49; mov.u64 %rd133, 0; { // callseq 854, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd133; .param .b64 param3; st.param.b64 [param3+0], %rd133; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f519, [retval0+0]; ld.param.f32 %f520, [retval0+4]; ld.param.f32 %f521, [retval0+8]; ld.param.f32 %f522, [retval0+12]; } // callseq 854 bra.uni $L__BB110_93; $L__BB110_21: ld.local.s32 %rd18, [%rd65+60]; setp.eq.s64 %p24, %rd18, -1; @%p24 bra $L__BB110_24; ld.const.u64 %rd85, [kernel_params+2240]; cvta.to.global.u64 %rd86, %rd85; shl.b64 %rd87, %rd18, 2; add.s64 %rd88, %rd86, %rd87; ld.global.u32 %r88, [%rd88]; setp.eq.s32 %p25, %r88, -1; @%p25 bra $L__BB110_24; mov.u64 %rd89, 0; { // callseq 853, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd89; .param .b64 param3; st.param.b64 [param3+0], %rd89; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f496, [retval0+0]; ld.param.f32 %f497, [retval0+4]; ld.param.f32 %f498, [retval0+8]; } // callseq 853 bra.uni $L__BB110_40; $L__BB110_141: and.b32 %r218, %r58, 8; setp.eq.s32 %p98, %r218, 0; setp.ne.s32 %p99, %r32, 8; mov.f32 %f531, 0f00000000; or.pred %p100, %p99, %p98; @%p100 bra $L__BB110_143; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; ld.local.u32 %r219, [%rd65+60]; add.s32 %r220, %r219, %r29; mul.wide.s32 %rd266, %r220, 4; add.s64 %rd267, %rd264, %rd266; ld.global.f32 %f531, [%rd267]; bra.uni $L__BB110_143; $L__BB110_116: and.b32 %r182, %r51, 8; setp.eq.s32 %p79, %r182, 0; setp.ne.s32 %p80, %r32, 8; mov.f32 %f529, 0f00000000; or.pred %p81, %p80, %p79; mov.f32 %f530, %f529; @%p81 bra $L__BB110_118; ld.const.u64 %rd219, [kernel_params+2328]; cvta.to.global.u64 %rd220, %rd219; ld.local.u32 %r183, [%rd65+60]; add.s32 %r184, %r183, %r29; mul.wide.s32 %rd222, %r184, 8; add.s64 %rd223, %rd220, %rd222; ld.global.v2.f32 {%f529, %f530}, [%rd223]; bra.uni $L__BB110_118; $L__BB110_49: and.b32 %r123, %r32, 120; setp.eq.s32 %p42, %r123, 0; @%p42 bra $L__BB110_82; and.b32 %r124, %r32, 24; setp.eq.s32 %p43, %r124, 0; @%p43 bra $L__BB110_52; ld.const.u64 %rd134, [kernel_params+2232]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd47, 4; add.s64 %rd137, %rd135, %rd136; ld.global.v4.u32 {%r125, %r126, %r127, %r128}, [%rd137]; ld.const.u64 %rd138, [kernel_params+2344]; cvta.to.global.u64 %rd139, %rd138; add.s32 %r132, %r125, %r29; mul.wide.u32 %rd140, %r132, 16; add.s64 %rd141, %rd139, %rd140; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd141]; add.s32 %r133, %r126, %r29; mul.wide.u32 %rd142, %r133, 16; add.s64 %rd143, %rd139, %rd142; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd143]; add.s32 %r134, %r127, %r29; mul.wide.u32 %rd144, %r134, 16; add.s64 %rd145, %rd139, %rd144; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd145]; bra.uni $L__BB110_81; $L__BB110_24: and.b32 %r89, %r32, 56; setp.eq.s32 %p26, %r89, 0; @%p26 bra $L__BB110_29; and.b32 %r90, %r32, 24; setp.eq.s32 %p27, %r90, 0; @%p27 bra $L__BB110_27; ld.const.u64 %rd90, [kernel_params+2232]; cvta.to.global.u64 %rd91, %rd90; shl.b64 %rd92, %rd18, 4; add.s64 %rd93, %rd91, %rd92; ld.global.v4.u32 {%r91, %r92, %r93, %r94}, [%rd93]; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; add.s32 %r98, %r91, %r29; mul.wide.u32 %rd96, %r98, 12; add.s64 %rd284, %rd95, %rd96; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r99, %r92, %r29; mul.wide.u32 %rd97, %r99, 12; add.s64 %rd281, %rd95, %rd97; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r100, %r93, %r29; mul.wide.u32 %rd98, %r100, 12; add.s64 %rd278, %rd95, %rd98; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; bra.uni $L__BB110_28; $L__BB110_137: and.b32 %r216, %r32, 131; setp.eq.s32 %p96, %r216, 0; mov.f32 %f531, 0f00000000; @%p96 bra $L__BB110_143; setp.ne.s32 %p97, %r32, 128; @%p97 bra $L__BB110_140; ld.local.u32 %r217, [%rd65+60]; add.s32 %r29, %r217, %r29; $L__BB110_140: ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r29, 4; add.s64 %rd262, %rd260, %rd261; ld.global.f32 %f531, [%rd262]; bra.uni $L__BB110_143; $L__BB110_132: and.b32 %r205, %r32, 7; setp.eq.s32 %p92, %r205, 0; mov.f32 %f531, 0f00000000; @%p92 bra $L__BB110_143; cvt.u32.u64 %r206, %rd53; setp.eq.s32 %p93, %r32, 4; selp.b32 %r207, %r206, 0, %p93; add.s32 %r208, %r29, %r207; ld.const.u64 %rd245, [kernel_params+2320]; cvta.to.global.u64 %rd246, %rd245; mul.wide.s32 %rd247, %r208, 4; add.s64 %rd248, %rd246, %rd247; ld.global.f32 %f531, [%rd248]; bra.uni $L__BB110_143; $L__BB110_112: and.b32 %r180, %r32, 131; setp.eq.s32 %p77, %r180, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p77 bra $L__BB110_118; setp.ne.s32 %p78, %r32, 128; @%p78 bra $L__BB110_115; ld.local.u32 %r181, [%rd65+60]; add.s32 %r29, %r181, %r29; $L__BB110_115: ld.const.u64 %rd215, [kernel_params+2328]; cvta.to.global.u64 %rd216, %rd215; mul.wide.s32 %rd217, %r29, 8; add.s64 %rd218, %rd216, %rd217; ld.global.v2.f32 {%f529, %f530}, [%rd218]; bra.uni $L__BB110_118; $L__BB110_107: and.b32 %r169, %r32, 7; setp.eq.s32 %p73, %r169, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p73 bra $L__BB110_118; cvt.u32.u64 %r170, %rd51; setp.eq.s32 %p74, %r32, 4; selp.b32 %r171, %r170, 0, %p74; add.s32 %r172, %r29, %r171; ld.const.u64 %rd201, [kernel_params+2328]; cvta.to.global.u64 %rd202, %rd201; mul.wide.s32 %rd203, %r172, 8; add.s64 %rd204, %rd202, %rd203; ld.global.v2.f32 {%f529, %f530}, [%rd204]; bra.uni $L__BB110_118; $L__BB110_91: and.b32 %r150, %r41, 8; setp.eq.s32 %p60, %r150, 0; setp.ne.s32 %p61, %r32, 8; mov.f32 %f519, 0f00000000; or.pred %p62, %p61, %p60; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p62 bra $L__BB110_93; ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; ld.local.u32 %r151, [%rd65+60]; add.s32 %r152, %r151, %r29; mul.wide.s32 %rd175, %r152, 16; add.s64 %rd176, %rd173, %rd175; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd176]; bra.uni $L__BB110_93; $L__BB110_38: and.b32 %r118, %r41, 8; setp.eq.s32 %p34, %r118, 0; setp.ne.s32 %p35, %r32, 8; mov.f32 %f496, 0f00000000; or.pred %p36, %p35, %p34; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p36 bra $L__BB110_40; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r119, [%rd65+60]; add.s32 %r120, %r119, %r29; mul.wide.s32 %rd126, %r120, 12; add.s64 %rd127, %rd124, %rd126; ld.global.f32 %f498, [%rd127+8]; ld.global.f32 %f497, [%rd127+4]; ld.global.f32 %f496, [%rd127]; bra.uni $L__BB110_40; $L__BB110_130: cvt.u32.u64 %r202, %rd53; mad.lo.s32 %r203, %r202, 3, %r29; ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r203, 4; add.s64 %rd287, %rd241, %rd242; add.s64 %rd288, %rd287, 4; add.s32 %r204, %r203, 2; mul.wide.s32 %rd243, %r204, 4; add.s64 %rd289, %rd241, %rd243; $L__BB110_131: ld.local.f32 %f404, [%rd65+68]; ld.global.f32 %f405, [%rd288]; ld.local.f32 %f406, [%rd65+72]; ld.global.f32 %f407, [%rd289]; mul.ftz.f32 %f408, %f407, %f406; fma.rn.ftz.f32 %f409, %f405, %f404, %f408; mov.f32 %f410, 0f3F800000; sub.ftz.f32 %f411, %f410, %f404; sub.ftz.f32 %f412, %f411, %f406; ld.global.f32 %f413, [%rd287]; fma.rn.ftz.f32 %f531, %f413, %f412, %f409; $L__BB110_143: setp.eq.s32 %p101, %r64, 1; @%p101 bra $L__BB110_147; bra.uni $L__BB110_144; $L__BB110_147: st.local.f32 [%rd16], %f531; bra.uni $L__BB110_157; $L__BB110_144: setp.eq.s32 %p102, %r64, 0; @%p102 bra $L__BB110_146; mov.u32 %r221, 1065353216; st.local.u32 [%rd16], %r221; bra.uni $L__BB110_157; $L__BB110_146: st.local.f32 [%rd16], %f531; st.local.f32 [%rd16+4], %f531; st.local.f32 [%rd16+8], %f531; bra.uni $L__BB110_157; $L__BB110_105: cvt.u32.u64 %r167, %rd51; mad.lo.s32 %r168, %r167, 3, %r29; ld.const.u64 %rd196, [kernel_params+2328]; cvta.to.global.u64 %rd197, %rd196; mul.wide.s32 %rd198, %r168, 8; add.s64 %rd199, %rd197, %rd198; ld.global.v2.f32 {%f527, %f528}, [%rd199]; ld.global.v2.f32 {%f525, %f526}, [%rd199+8]; ld.global.v2.f32 {%f523, %f524}, [%rd199+16]; $L__BB110_106: ld.local.f32 %f370, [%rd65+68]; ld.local.f32 %f371, [%rd65+72]; mul.ftz.f32 %f372, %f523, %f371; mul.ftz.f32 %f373, %f524, %f371; fma.rn.ftz.f32 %f374, %f525, %f370, %f372; fma.rn.ftz.f32 %f375, %f526, %f370, %f373; mov.f32 %f376, 0f3F800000; sub.ftz.f32 %f377, %f376, %f370; sub.ftz.f32 %f378, %f377, %f371; fma.rn.ftz.f32 %f530, %f528, %f378, %f375; fma.rn.ftz.f32 %f529, %f527, %f378, %f374; $L__BB110_118: setp.eq.s32 %p82, %r64, 1; @%p82 bra $L__BB110_122; bra.uni $L__BB110_119; $L__BB110_122: st.local.f32 [%rd16], %f529; bra.uni $L__BB110_157; $L__BB110_119: setp.eq.s32 %p83, %r64, 0; @%p83 bra $L__BB110_121; mov.u32 %r185, 1065353216; st.local.u32 [%rd16], %r185; bra.uni $L__BB110_157; $L__BB110_121: st.local.f32 [%rd16], %f529; st.local.f32 [%rd16+4], %f530; mov.u32 %r186, 0; st.local.u32 [%rd16+8], %r186; bra.uni $L__BB110_157; $L__BB110_87: and.b32 %r148, %r32, 131; setp.eq.s32 %p58, %r148, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p58 bra $L__BB110_93; setp.ne.s32 %p59, %r32, 128; @%p59 bra $L__BB110_90; ld.local.u32 %r149, [%rd65+60]; add.s32 %r29, %r149, %r29; $L__BB110_90: ld.const.u64 %rd168, [kernel_params+2344]; cvta.to.global.u64 %rd169, %rd168; mul.wide.s32 %rd170, %r29, 16; add.s64 %rd171, %rd169, %rd170; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd171]; bra.uni $L__BB110_93; $L__BB110_34: and.b32 %r116, %r32, 131; setp.eq.s32 %p32, %r116, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p32 bra $L__BB110_40; setp.ne.s32 %p33, %r32, 128; @%p33 bra $L__BB110_37; ld.local.u32 %r117, [%rd65+60]; add.s32 %r29, %r117, %r29; $L__BB110_37: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r29, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f498, [%rd122+8]; ld.global.f32 %f497, [%rd122+4]; ld.global.f32 %f496, [%rd122]; bra.uni $L__BB110_40; $L__BB110_82: and.b32 %r137, %r32, 7; setp.eq.s32 %p54, %r137, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p54 bra $L__BB110_93; cvt.u32.u64 %r138, %rd47; setp.eq.s32 %p55, %r32, 4; selp.b32 %r139, %r138, 0, %p55; add.s32 %r140, %r29, %r139; ld.const.u64 %rd154, [kernel_params+2344]; cvta.to.global.u64 %rd155, %rd154; mul.wide.s32 %rd156, %r140, 16; add.s64 %rd157, %rd155, %rd156; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd157]; bra.uni $L__BB110_93; $L__BB110_29: and.b32 %r105, %r32, 7; setp.eq.s32 %p28, %r105, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p28 bra $L__BB110_40; cvt.u32.u64 %r106, %rd18; setp.eq.s32 %p29, %r32, 4; selp.b32 %r107, %r106, 0, %p29; add.s32 %r108, %r29, %r107; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r108, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f498, [%rd108+8]; ld.global.f32 %f497, [%rd108+4]; ld.global.f32 %f496, [%rd108]; bra.uni $L__BB110_40; $L__BB110_52: cvt.u32.u64 %r135, %rd47; mad.lo.s32 %r136, %r135, 3, %r29; cvt.s64.s32 %rd48, %r136; setp.eq.s32 %p44, %r32, 32; @%p44 bra $L__BB110_80; bra.uni $L__BB110_53; $L__BB110_80: ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd48, 4; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd152]; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd152+16]; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd152+32]; bra.uni $L__BB110_81; $L__BB110_27: cvt.u32.u64 %r101, %rd18; mad.lo.s32 %r102, %r101, 3, %r29; ld.const.u64 %rd99, [kernel_params+2336]; cvta.to.global.u64 %rd100, %rd99; mul.wide.s32 %rd101, %r102, 12; add.s64 %rd284, %rd100, %rd101; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r103, %r102, 1; mul.wide.s32 %rd102, %r103, 12; add.s64 %rd281, %rd100, %rd102; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r104, %r102, 2; mul.wide.s32 %rd103, %r104, 12; add.s64 %rd278, %rd100, %rd103; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; $L__BB110_28: ld.local.f32 %f179, [%rd65+68]; ld.global.f32 %f180, [%rd281]; ld.global.f32 %f181, [%rd282]; ld.global.f32 %f182, [%rd283]; ld.local.f32 %f183, [%rd65+72]; ld.global.f32 %f184, [%rd278]; mul.ftz.f32 %f185, %f184, %f183; ld.global.f32 %f186, [%rd279]; mul.ftz.f32 %f187, %f186, %f183; ld.global.f32 %f188, [%rd280]; mul.ftz.f32 %f189, %f188, %f183; fma.rn.ftz.f32 %f190, %f180, %f179, %f185; fma.rn.ftz.f32 %f191, %f181, %f179, %f187; fma.rn.ftz.f32 %f192, %f182, %f179, %f189; mov.f32 %f193, 0f3F800000; sub.ftz.f32 %f194, %f193, %f179; sub.ftz.f32 %f195, %f194, %f183; ld.global.f32 %f196, [%rd284]; ld.global.f32 %f197, [%rd285]; ld.global.f32 %f198, [%rd286]; fma.rn.ftz.f32 %f498, %f198, %f195, %f192; fma.rn.ftz.f32 %f497, %f197, %f195, %f191; fma.rn.ftz.f32 %f496, %f196, %f195, %f190; $L__BB110_40: setp.eq.s32 %p37, %r64, 1; @%p37 bra $L__BB110_44; bra.uni $L__BB110_41; $L__BB110_44: add.ftz.f32 %f220, %f496, %f497; add.ftz.f32 %f221, %f498, %f220; mul.ftz.f32 %f222, %f221, 0f3EAAAAAB; st.local.f32 [%rd16], %f222; bra.uni $L__BB110_157; $L__BB110_41: setp.eq.s32 %p38, %r64, 0; @%p38 bra $L__BB110_43; mov.u32 %r121, 1065353216; st.local.u32 [%rd16], %r121; bra.uni $L__BB110_157; $L__BB110_43: st.local.f32 [%rd16], %f496; st.local.f32 [%rd16+4], %f497; st.local.f32 [%rd16+8], %f498; bra.uni $L__BB110_157; $L__BB110_53: ld.const.u64 %rd146, [kernel_params+2352]; cvta.to.global.u64 %rd147, %rd146; shl.b64 %rd148, %rd48, 2; add.s64 %rd49, %rd147, %rd148; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd49]; cvt.rn.f32.u16 %f235, %rs20; mul.ftz.f32 %f38, %f235, 0f3B808081; cvt.rn.f32.u16 %f236, %rs21; mul.ftz.f32 %f39, %f236, 0f3B808081; cvt.rn.f32.u16 %f237, %rs22; mul.ftz.f32 %f40, %f237, 0f3B808081; setp.lt.ftz.f32 %p45, %f38, 0f3D25AEE6; @%p45 bra $L__BB110_55; bra.uni $L__BB110_54; $L__BB110_55: mul.ftz.f32 %f515, %f38, 0f3D9E8391; bra.uni $L__BB110_56; $L__BB110_54: add.ftz.f32 %f238, %f38, 0f3D6147AE; mul.ftz.f32 %f239, %f238, 0f3F72A76F; lg2.approx.ftz.f32 %f240, %f239; mul.ftz.f32 %f241, %f240, 0f4019999A; ex2.approx.ftz.f32 %f515, %f241; $L__BB110_56: setp.lt.ftz.f32 %p46, %f39, 0f3D25AEE6; @%p46 bra $L__BB110_58; bra.uni $L__BB110_57; $L__BB110_58: mul.ftz.f32 %f516, %f39, 0f3D9E8391; bra.uni $L__BB110_59; $L__BB110_57: add.ftz.f32 %f242, %f39, 0f3D6147AE; mul.ftz.f32 %f243, %f242, 0f3F72A76F; lg2.approx.ftz.f32 %f244, %f243; mul.ftz.f32 %f245, %f244, 0f4019999A; ex2.approx.ftz.f32 %f516, %f245; $L__BB110_59: setp.lt.ftz.f32 %p47, %f40, 0f3D25AEE6; @%p47 bra $L__BB110_61; bra.uni $L__BB110_60; $L__BB110_61: mul.ftz.f32 %f517, %f40, 0f3D9E8391; bra.uni $L__BB110_62; $L__BB110_60: add.ftz.f32 %f246, %f40, 0f3D6147AE; mul.ftz.f32 %f247, %f246, 0f3F72A76F; lg2.approx.ftz.f32 %f248, %f247; mul.ftz.f32 %f249, %f248, 0f4019999A; ex2.approx.ftz.f32 %f517, %f249; $L__BB110_62: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd49+4]; cvt.rn.f32.u16 %f250, %rs24; mul.ftz.f32 %f50, %f250, 0f3B808081; cvt.rn.f32.u16 %f251, %rs25; mul.ftz.f32 %f51, %f251, 0f3B808081; cvt.rn.f32.u16 %f252, %rs26; mul.ftz.f32 %f52, %f252, 0f3B808081; setp.lt.ftz.f32 %p48, %f50, 0f3D25AEE6; @%p48 bra $L__BB110_64; bra.uni $L__BB110_63; $L__BB110_64: mul.ftz.f32 %f511, %f50, 0f3D9E8391; bra.uni $L__BB110_65; $L__BB110_63: add.ftz.f32 %f253, %f50, 0f3D6147AE; mul.ftz.f32 %f254, %f253, 0f3F72A76F; lg2.approx.ftz.f32 %f255, %f254; mul.ftz.f32 %f256, %f255, 0f4019999A; ex2.approx.ftz.f32 %f511, %f256; $L__BB110_65: setp.lt.ftz.f32 %p49, %f51, 0f3D25AEE6; @%p49 bra $L__BB110_67; bra.uni $L__BB110_66; $L__BB110_67: mul.ftz.f32 %f512, %f51, 0f3D9E8391; bra.uni $L__BB110_68; $L__BB110_66: add.ftz.f32 %f257, %f51, 0f3D6147AE; mul.ftz.f32 %f258, %f257, 0f3F72A76F; lg2.approx.ftz.f32 %f259, %f258; mul.ftz.f32 %f260, %f259, 0f4019999A; ex2.approx.ftz.f32 %f512, %f260; $L__BB110_68: setp.lt.ftz.f32 %p50, %f52, 0f3D25AEE6; @%p50 bra $L__BB110_70; bra.uni $L__BB110_69; $L__BB110_70: mul.ftz.f32 %f513, %f52, 0f3D9E8391; bra.uni $L__BB110_71; $L__BB110_69: add.ftz.f32 %f261, %f52, 0f3D6147AE; mul.ftz.f32 %f262, %f261, 0f3F72A76F; lg2.approx.ftz.f32 %f263, %f262; mul.ftz.f32 %f264, %f263, 0f4019999A; ex2.approx.ftz.f32 %f513, %f264; $L__BB110_71: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd49+8]; cvt.rn.f32.u16 %f265, %rs28; mul.ftz.f32 %f62, %f265, 0f3B808081; cvt.rn.f32.u16 %f266, %rs29; mul.ftz.f32 %f63, %f266, 0f3B808081; cvt.rn.f32.u16 %f267, %rs30; mul.ftz.f32 %f64, %f267, 0f3B808081; setp.lt.ftz.f32 %p51, %f62, 0f3D25AEE6; @%p51 bra $L__BB110_73; bra.uni $L__BB110_72; $L__BB110_73: mul.ftz.f32 %f507, %f62, 0f3D9E8391; bra.uni $L__BB110_74; $L__BB110_72: add.ftz.f32 %f268, %f62, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f507, %f271; $L__BB110_74: setp.lt.ftz.f32 %p52, %f63, 0f3D25AEE6; @%p52 bra $L__BB110_76; bra.uni $L__BB110_75; $L__BB110_76: mul.ftz.f32 %f508, %f63, 0f3D9E8391; bra.uni $L__BB110_77; $L__BB110_75: add.ftz.f32 %f272, %f63, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f508, %f275; $L__BB110_77: cvt.rn.f32.u16 %f276, %rs23; mul.ftz.f32 %f518, %f276, 0f3B808081; cvt.rn.f32.u16 %f277, %rs27; mul.ftz.f32 %f514, %f277, 0f3B808081; cvt.rn.f32.u16 %f278, %rs31; mul.ftz.f32 %f510, %f278, 0f3B808081; setp.lt.ftz.f32 %p53, %f64, 0f3D25AEE6; @%p53 bra $L__BB110_79; bra.uni $L__BB110_78; $L__BB110_79: mul.ftz.f32 %f509, %f64, 0f3D9E8391; bra.uni $L__BB110_81; $L__BB110_78: add.ftz.f32 %f279, %f64, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f509, %f282; $L__BB110_81: ld.local.f32 %f295, [%rd65+68]; ld.local.f32 %f296, [%rd65+72]; mul.ftz.f32 %f297, %f507, %f296; mul.ftz.f32 %f298, %f508, %f296; mul.ftz.f32 %f299, %f509, %f296; mul.ftz.f32 %f300, %f510, %f296; fma.rn.ftz.f32 %f301, %f511, %f295, %f297; fma.rn.ftz.f32 %f302, %f512, %f295, %f298; fma.rn.ftz.f32 %f303, %f513, %f295, %f299; fma.rn.ftz.f32 %f304, %f514, %f295, %f300; mov.f32 %f305, 0f3F800000; sub.ftz.f32 %f306, %f305, %f295; sub.ftz.f32 %f307, %f306, %f296; fma.rn.ftz.f32 %f522, %f518, %f307, %f304; fma.rn.ftz.f32 %f521, %f517, %f307, %f303; fma.rn.ftz.f32 %f520, %f516, %f307, %f302; fma.rn.ftz.f32 %f519, %f515, %f307, %f301; $L__BB110_93: setp.eq.s32 %p63, %r64, 1; @%p63 bra $L__BB110_97; bra.uni $L__BB110_94; $L__BB110_97: add.ftz.f32 %f355, %f519, %f520; add.ftz.f32 %f356, %f521, %f355; mul.ftz.f32 %f357, %f356, 0f3EAAAAAB; st.local.f32 [%rd16], %f357; bra.uni $L__BB110_157; $L__BB110_94: setp.eq.s32 %p64, %r64, 0; @%p64 bra $L__BB110_96; st.local.f32 [%rd16], %f522; bra.uni $L__BB110_157; $L__BB110_96: st.local.f32 [%rd16], %f519; st.local.f32 [%rd16+4], %f520; st.local.f32 [%rd16+8], %f521; $L__BB110_157: ret; } .func _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r1, [_Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; cvta.to.local.u64 %rd2, %rd1; add.s32 %r2, %r1, 1; ld.param.u8 %r3, [_Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+7]; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r4, 0; st.local.u32 [%rd4], %r4; st.local.u32 [%rd4+4], %r4; st.local.u32 [%rd4+8], %r4; st.param.b32 [func_retval0+0], %r2; ret; } .func _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func (.param .b32 func_retval0) _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b32 _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0 ) { .reg .b32 %r<3>; ld.param.u32 %r1, [_Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; add.s32 %r2, %r1, 2; st.param.b32 [func_retval0+0], %r2; ret; } .func _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4 ) { .reg .pred %p<11>; .reg .b16 %rs<6>; .reg .f32 %f<12>; .reg .b32 %r<6>; .reg .b64 %rd<28>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u64 %rd3, [_Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd2, [_Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; cvta.to.local.u64 %rd4, %rd3; add.s64 %rd1, %rd4, 52; mov.f32 %f11, 0f00000000; setp.gt.s32 %p1, %r2, 10; @%p1 bra $L__BB115_5; bra.uni $L__BB115_1; $L__BB115_5: setp.gt.s32 %p2, %r2, 12; @%p2 bra $L__BB115_9; setp.eq.s32 %p5, %r2, 11; @%p5 bra $L__BB115_14; setp.eq.s32 %p6, %r2, 12; @%p6 bra $L__BB115_8; bra.uni $L__BB115_16; $L__BB115_8: ld.const.u64 %rd13, [kernel_params+2488]; cvta.to.global.u64 %rd14, %rd13; mul.wide.s32 %rd15, %r1, 2; add.s64 %rd16, %rd14, %rd15; ld.global.u16 %rs3, [%rd16]; cvt.rn.f32.u16 %f11, %rs3; bra.uni $L__BB115_16; $L__BB115_1: setp.eq.s32 %p7, %r2, 8; @%p7 bra $L__BB115_12; setp.eq.s32 %p8, %r2, 9; @%p8 bra $L__BB115_13; setp.eq.s32 %p9, %r2, 10; @%p9 bra $L__BB115_4; bra.uni $L__BB115_16; $L__BB115_4: ld.const.u64 %rd5, [kernel_params+2464]; cvta.to.global.u64 %rd6, %rd5; mul.wide.s32 %rd7, %r1, 2; add.s64 %rd8, %rd6, %rd7; ld.global.u16 %rs1, [%rd8]; cvt.rn.f32.u16 %f10, %rs1; add.ftz.f32 %f11, %f10, 0f3F800000; bra.uni $L__BB115_16; $L__BB115_9: setp.eq.s32 %p3, %r2, 13; @%p3 bra $L__BB115_15; setp.eq.s32 %p4, %r2, 14; @%p4 bra $L__BB115_11; bra.uni $L__BB115_16; $L__BB115_11: ld.const.u64 %rd21, [kernel_params+2496]; cvta.to.global.u64 %rd22, %rd21; mul.wide.s32 %rd23, %r1, 2; add.s64 %rd24, %rd22, %rd23; ld.global.u16 %rs5, [%rd24]; cvt.rn.f32.u16 %f11, %rs5; bra.uni $L__BB115_16; $L__BB115_14: ld.const.u64 %rd9, [kernel_params+2480]; cvta.to.global.u64 %rd10, %rd9; mul.wide.s32 %rd11, %r1, 2; add.s64 %rd12, %rd10, %rd11; ld.global.u16 %rs2, [%rd12]; cvt.rn.f32.u16 %f11, %rs2; bra.uni $L__BB115_16; $L__BB115_12: ld.local.u32 %r4, [%rd1]; and.b32 %r5, %r4, 1; setp.eq.b32 %p10, %r5, 1; selp.f32 %f11, 0f3F800000, 0f00000000, %p10; bra.uni $L__BB115_16; $L__BB115_15: ld.const.u64 %rd17, [kernel_params+2472]; cvta.to.global.u64 %rd18, %rd17; mul.wide.s32 %rd19, %r1, 2; add.s64 %rd20, %rd18, %rd19; ld.global.u16 %rs4, [%rd20]; cvt.rn.f32.u16 %f11, %rs4; bra.uni $L__BB115_16; $L__BB115_13: ld.local.f32 %f11, [%rd1+36]; $L__BB115_16: cvta.to.local.u64 %rd25, %rd2; mul.wide.u32 %rd26, %r3, 4; add.s64 %rd27, %rd25, %rd26; st.local.f32 [%rd27], %f11; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB116_2; bra.uni $L__BB116_1; $L__BB116_2: mov.b32 %f2764, %r395; bra.uni $L__BB116_3; $L__BB116_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB116_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB116_5; bra.uni $L__BB116_4; $L__BB116_5: mov.b32 %f9, %r396; bra.uni $L__BB116_6; $L__BB116_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB116_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB116_8; bra.uni $L__BB116_7; $L__BB116_8: mov.b32 %f2766, %r397; bra.uni $L__BB116_9; $L__BB116_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB116_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB116_11; bra.uni $L__BB116_10; $L__BB116_11: mov.b32 %f15, %r398; bra.uni $L__BB116_12; $L__BB116_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB116_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB116_14; bra.uni $L__BB116_13; $L__BB116_14: mov.b32 %f2768, %r399; bra.uni $L__BB116_15; $L__BB116_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB116_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB116_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB116_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB116_18; bra.uni $L__BB116_385; $L__BB116_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB116_259; bra.uni $L__BB116_19; $L__BB116_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB116_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB116_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB116_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB116_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB116_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB116_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB116_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB116_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB116_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB116_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB116_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB116_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB116_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB116_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB116_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB116_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB116_362; $L__BB116_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB116_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB116_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB116_40; bra.uni $L__BB116_29; $L__BB116_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB116_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB116_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB116_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB116_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB116_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB116_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB116_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB116_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB116_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB116_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB116_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB116_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB116_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB116_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB116_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB116_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB116_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB116_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB116_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB116_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB116_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB116_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB116_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB116_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB116_136; $L__BB116_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB116_371; bra.uni $L__BB116_366; $L__BB116_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB116_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB116_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB116_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB116_383; $L__BB116_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB116_149; bra.uni $L__BB116_139; $L__BB116_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB116_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB116_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB116_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB116_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB116_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB116_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB116_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB116_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB116_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB116_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB116_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB116_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB116_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB116_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB116_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB116_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB116_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB116_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB116_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB116_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB116_250; $L__BB116_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB116_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB116_253; bra.uni $L__BB116_362; $L__BB116_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB116_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB116_304; @%p171 bra $L__BB116_303; bra.uni $L__BB116_256; $L__BB116_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB116_305; $L__BB116_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB116_305; $L__BB116_256: @%p172 bra $L__BB116_302; bra.uni $L__BB116_257; $L__BB116_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB116_305; $L__BB116_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB116_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB116_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB116_308; bra.uni $L__BB116_306; $L__BB116_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB116_309; $L__BB116_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB116_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB116_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB116_316; @%p171 bra $L__BB116_315; bra.uni $L__BB116_311; $L__BB116_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB116_317; $L__BB116_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB116_317; $L__BB116_311: @%p172 bra $L__BB116_314; bra.uni $L__BB116_312; $L__BB116_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB116_317; $L__BB116_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB116_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB116_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB116_320; bra.uni $L__BB116_318; $L__BB116_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB116_321; $L__BB116_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB116_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB116_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB116_328; @%p171 bra $L__BB116_327; bra.uni $L__BB116_323; $L__BB116_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB116_329; $L__BB116_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB116_329; $L__BB116_323: @%p172 bra $L__BB116_326; bra.uni $L__BB116_324; $L__BB116_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB116_329; $L__BB116_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB116_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB116_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB116_332; bra.uni $L__BB116_330; $L__BB116_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB116_333; $L__BB116_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB116_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB116_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB116_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_362; $L__BB116_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB116_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB116_31; bra.uni $L__BB116_136; $L__BB116_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB116_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB116_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB116_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB116_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB116_107; @%p49 bra $L__BB116_106; bra.uni $L__BB116_37; $L__BB116_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB116_108; $L__BB116_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB116_108; $L__BB116_37: @%p50 bra $L__BB116_105; bra.uni $L__BB116_38; $L__BB116_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB116_108; $L__BB116_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB116_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB116_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB116_111; bra.uni $L__BB116_109; $L__BB116_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB116_112; $L__BB116_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB116_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB116_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB116_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB116_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB116_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB116_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_136; $L__BB116_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB116_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB116_368; bra.uni $L__BB116_383; $L__BB116_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB116_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB116_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB116_383; $L__BB116_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB116_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB116_141; bra.uni $L__BB116_250; $L__BB116_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB116_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB116_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB116_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB116_206; @%p101 bra $L__BB116_205; bra.uni $L__BB116_146; $L__BB116_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB116_207; $L__BB116_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB116_207; $L__BB116_146: @%p102 bra $L__BB116_204; bra.uni $L__BB116_147; $L__BB116_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB116_207; $L__BB116_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB116_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB116_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB116_210; bra.uni $L__BB116_208; $L__BB116_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB116_211; $L__BB116_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB116_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB116_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB116_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB116_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB116_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_250; $L__BB116_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB116_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB116_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB116_295; @%p151 bra $L__BB116_294; bra.uni $L__BB116_290; $L__BB116_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB116_296; $L__BB116_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB116_296; $L__BB116_290: @%p152 bra $L__BB116_293; bra.uni $L__BB116_291; $L__BB116_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB116_296; $L__BB116_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB116_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB116_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB116_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB116_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB116_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB116_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB116_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_362; $L__BB116_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB116_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB116_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB116_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB116_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB116_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB116_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB116_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB116_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB116_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB116_279; bra.uni $L__BB116_362; $L__BB116_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB116_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB116_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB116_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB116_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB116_96; @%p39 bra $L__BB116_95; bra.uni $L__BB116_91; $L__BB116_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB116_97; $L__BB116_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB116_97; $L__BB116_91: @%p40 bra $L__BB116_94; bra.uni $L__BB116_92; $L__BB116_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB116_97; $L__BB116_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB116_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB116_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB116_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB116_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB116_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB116_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB116_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB116_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB116_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_136; $L__BB116_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB116_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB116_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB116_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB116_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB116_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB116_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB116_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB116_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB116_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB116_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB116_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB116_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB116_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB116_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB116_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB116_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB116_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB116_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB116_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB116_76; bra.uni $L__BB116_136; $L__BB116_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB116_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB116_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB116_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB116_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB116_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_383; $L__BB116_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB116_383; $L__BB116_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB116_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB116_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB116_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB116_196; @%p92 bra $L__BB116_195; bra.uni $L__BB116_191; $L__BB116_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB116_197; $L__BB116_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB116_197; $L__BB116_191: @%p93 bra $L__BB116_194; bra.uni $L__BB116_192; $L__BB116_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB116_197; $L__BB116_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB116_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB116_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB116_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB116_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB116_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB116_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB116_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB116_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB116_250; $L__BB116_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB116_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB116_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB116_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB116_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB116_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB116_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB116_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB116_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB116_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB116_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB116_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB116_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB116_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB116_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB116_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB116_179; bra.uni $L__BB116_250; $L__BB116_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB116_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB116_337; @%p188 bra $L__BB116_336; bra.uni $L__BB116_23; $L__BB116_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB116_338; $L__BB116_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB116_338; $L__BB116_23: @%p189 bra $L__BB116_335; bra.uni $L__BB116_24; $L__BB116_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB116_338; $L__BB116_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB116_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB116_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB116_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB116_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB116_347; @%p188 bra $L__BB116_346; bra.uni $L__BB116_342; $L__BB116_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB116_348; $L__BB116_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB116_348; $L__BB116_342: @%p189 bra $L__BB116_345; bra.uni $L__BB116_343; $L__BB116_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB116_348; $L__BB116_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB116_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB116_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB116_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB116_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB116_357; @%p188 bra $L__BB116_356; bra.uni $L__BB116_352; $L__BB116_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB116_358; $L__BB116_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB116_358; $L__BB116_352: @%p189 bra $L__BB116_355; bra.uni $L__BB116_353; $L__BB116_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB116_358; $L__BB116_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB116_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB116_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB116_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB116_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB116_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB116_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB116_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB116_364: mov.f32 %f3392, %f3387; bra.uni $L__BB116_385; $L__BB116_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB116_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB116_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB116_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB116_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB116_128; @%p59 bra $L__BB116_127; bra.uni $L__BB116_123; $L__BB116_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB116_129; $L__BB116_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB116_129; $L__BB116_123: @%p60 bra $L__BB116_126; bra.uni $L__BB116_124; $L__BB116_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB116_129; $L__BB116_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB116_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB116_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB116_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB116_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB116_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB116_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB116_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB116_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB116_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB116_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB116_385; $L__BB116_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB116_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB116_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB116_385; $L__BB116_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB116_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB116_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB116_224; @%p120 bra $L__BB116_223; bra.uni $L__BB116_219; $L__BB116_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB116_225; $L__BB116_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB116_225; $L__BB116_219: @%p121 bra $L__BB116_222; bra.uni $L__BB116_220; $L__BB116_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB116_225; $L__BB116_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB116_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB116_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB116_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB116_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB116_234; @%p120 bra $L__BB116_233; bra.uni $L__BB116_229; $L__BB116_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB116_235; $L__BB116_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB116_235; $L__BB116_229: @%p121 bra $L__BB116_232; bra.uni $L__BB116_230; $L__BB116_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB116_235; $L__BB116_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB116_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB116_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB116_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB116_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB116_244; @%p120 bra $L__BB116_243; bra.uni $L__BB116_239; $L__BB116_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB116_245; $L__BB116_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB116_245; $L__BB116_239: @%p121 bra $L__BB116_242; bra.uni $L__BB116_240; $L__BB116_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB116_245; $L__BB116_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB116_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB116_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB116_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB116_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB116_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB116_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB116_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB116_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB116_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB116_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB116_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB116_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB116_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB116_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB116_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB116_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB116_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB116_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB116_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB118_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB118_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB118_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB118_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) { .reg .pred %p<2>; .reg .b32 %r<5>; ld.param.u32 %r1, [_Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u8 %r2, [_Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0+4]; setp.eq.s32 %p1, %r2, 39; selp.b32 %r3, 5, 1, %p1; add.s32 %r4, %r3, %r1; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<535>; .reg .b32 %r<242>; .reg .b64 %rd<290>; ld.param.u64 %rd63, [_Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd64, [_Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd65, %rd63; add.s64 %rd1, %rd65, 76; ld.local.u32 %r3, [%rd65+76]; cvt.s64.s32 %rd2, %r3; setp.eq.s32 %p1, %r3, -1; mov.u32 %r29, 0; mov.u32 %r30, %r29; mov.u32 %r236, %r64; mov.u32 %r32, %r29; @%p1 bra $L__BB120_16; cvt.u64.u32 %rd3, %r62; ld.local.s32 %rd4, [%rd1+-16]; ld.const.u64 %rd66, [kernel_params+2160]; cvta.to.global.u64 %rd67, %rd66; mul.lo.s64 %rd68, %rd2, 208; add.s64 %rd69, %rd67, %rd68; ld.global.u32 %r4, [%rd69+164]; ld.local.u32 %r69, [%rd1+-12]; and.b32 %r70, %r69, 1; setp.eq.b32 %p2, %r70, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r225, 0; @%p5 bra $L__BB120_5; cvt.u32.u64 %r72, %rd4; setp.eq.s32 %p6, %r72, -1; mov.u32 %r224, -1; @%p6 bra $L__BB120_4; ld.const.u64 %rd70, [kernel_params+2240]; cvta.to.global.u64 %rd71, %rd70; shl.b64 %rd72, %rd4, 2; add.s64 %rd73, %rd71, %rd72; ld.global.u32 %r224, [%rd73]; $L__BB120_4: setp.ne.s32 %p7, %r224, -1; selp.u32 %r225, 1, 0, %p7; $L__BB120_5: ld.const.u64 %rd74, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd74; add.s32 %r227, %r225, %r4; mul.wide.u32 %rd75, %r227, 16; add.s64 %rd76, %rd5, %rd75; add.s64 %rd276, %rd76, 14; add.s64 %rd277, %rd76, 15; ld.global.u32 %r226, [%rd76+8]; ld.global.u16 %rs37, [%rd76+12]; ld.global.u64 %rd275, [%rd76]; setp.eq.s64 %p8, %rd275, %rd3; @%p8 bra $L__BB120_11; $L__BB120_6: setp.eq.s64 %p9, %rd275, 0; @%p9 bra $L__BB120_8; add.s32 %r227, %r227, 2; bra.uni $L__BB120_9; $L__BB120_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r233, -1; mov.u32 %r230, 0; mov.u32 %r227, %r226; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p10 bra $L__BB120_15; $L__BB120_9: cvt.u64.u32 %rd10, %r227; mul.wide.u32 %rd77, %r227, 16; add.s64 %rd78, %rd5, %rd77; ld.global.u32 %r226, [%rd78+8]; ld.global.u16 %rs37, [%rd78+12]; ld.global.u64 %rd275, [%rd78]; setp.ne.s64 %p11, %rd275, %rd3; @%p11 bra $L__BB120_6; shl.b64 %rd79, %rd10, 4; add.s64 %rd80, %rd5, %rd79; add.s64 %rd276, %rd80, 14; add.s64 %rd277, %rd80, 15; $L__BB120_11: ld.global.u8 %rs5, [%rd277]; ld.global.u8 %rs6, [%rd276]; cvt.u32.u64 %r77, %rd4; setp.ne.s32 %p12, %r77, -1; @%p12 bra $L__BB120_14; mov.u32 %r233, -1; mov.u32 %r230, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB120_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p14 bra $L__BB120_15; $L__BB120_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r233, -1, %r226, %p15; cvt.u32.u16 %r230, %rs37; cvt.u32.u16 %r82, %rs5; and.b32 %r232, %r82, 255; cvt.u32.u16 %r83, %rs6; and.b32 %r231, %r83, 255; $L__BB120_15: setp.eq.s32 %p16, %r233, -1; selp.b32 %r29, 0, %r233, %p16; selp.b32 %r30, 0, %r232, %p16; selp.b32 %r236, %r64, %r231, %p16; selp.b32 %r32, 0, %r230, %p16; $L__BB120_16: setp.eq.s32 %p17, %r32, 0; setp.eq.s32 %p18, %r62, 7; and.pred %p19, %p18, %p17; cvta.to.local.u64 %rd81, %rd64; mul.wide.u32 %rd82, %r63, 4; add.s64 %rd16, %rd81, %rd82; @%p19 bra $L__BB120_148; bra.uni $L__BB120_17; $L__BB120_148: cvt.u32.u64 %r222, %rd2; setp.eq.s32 %p103, %r222, -1; ld.local.v4.f32 {%f534, %f533, %f532, %f426}, [%rd65]; @%p103 bra $L__BB120_152; ld.local.u8 %rs35, [%rd65+56]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p104, %rs36, 0; @%p104 bra $L__BB120_151; bra.uni $L__BB120_150; $L__BB120_151: ld.const.u64 %rd271, [kernel_params+2160]; cvta.to.global.u64 %rd272, %rd271; mul.wide.s32 %rd273, %r3, 208; add.s64 %rd274, %rd272, %rd273; ld.global.v4.f32 {%f460, %f461, %f462, %f463}, [%rd274+48]; ld.global.v4.f32 {%f468, %f469, %f470, %f471}, [%rd274+64]; ld.global.v4.f32 {%f476, %f477, %f478, %f479}, [%rd274+80]; mul.ftz.f32 %f484, %f533, %f461; fma.rn.ftz.f32 %f485, %f534, %f460, %f484; fma.rn.ftz.f32 %f486, %f532, %f462, %f485; add.ftz.f32 %f173, %f463, %f486; mul.ftz.f32 %f487, %f533, %f469; fma.rn.ftz.f32 %f488, %f534, %f468, %f487; fma.rn.ftz.f32 %f489, %f532, %f470, %f488; add.ftz.f32 %f174, %f471, %f489; mul.ftz.f32 %f490, %f533, %f477; fma.rn.ftz.f32 %f491, %f534, %f476, %f490; fma.rn.ftz.f32 %f492, %f532, %f478, %f491; add.ftz.f32 %f532, %f479, %f492; mov.f32 %f533, %f174; mov.f32 %f534, %f173; bra.uni $L__BB120_152; $L__BB120_17: setp.eq.s32 %p20, %r236, 0; @%p20 bra $L__BB120_123; setp.eq.s32 %p21, %r236, 1; @%p21 bra $L__BB120_98; bra.uni $L__BB120_19; $L__BB120_98: mov.u32 %r48, 1; ld.local.u32 %r51, [%rd65+64]; and.b32 %r153, %r51, 1; setp.eq.b32 %p65, %r153, 1; mov.pred %p66, 0; xor.pred %p67, %p65, %p66; not.pred %p68, %p67; @%p68 bra $L__BB120_109; bra.uni $L__BB120_99; $L__BB120_109: and.b32 %r173, %r51, 6; setp.eq.s32 %p75, %r173, 0; @%p75 bra $L__BB120_116; and.b32 %r174, %r32, 768; setp.eq.s32 %p76, %r174, 0; @%p76 bra $L__BB120_112; ld.const.u64 %rd205, [kernel_params+2264]; cvta.to.global.u64 %rd206, %rd205; ld.local.u32 %r175, [%rd65+60]; mul.wide.s32 %rd208, %r175, 16; add.s64 %rd209, %rd206, %rd208; shr.s32 %r176, %r51, 7; ld.global.u32 %r177, [%rd209+4]; add.s32 %r178, %r176, %r177; ld.const.u64 %rd210, [kernel_params+2328]; cvta.to.global.u64 %rd211, %rd210; add.s32 %r179, %r178, %r29; mul.wide.s32 %rd212, %r179, 8; add.s64 %rd213, %rd211, %rd212; ld.global.v2.f32 {%f383, %f384}, [%rd213]; ld.global.v2.f32 {%f387, %f388}, [%rd213+8]; ld.local.f32 %f391, [%rd65+68]; mov.f32 %f392, 0f3F800000; sub.ftz.f32 %f393, %f392, %f391; mul.ftz.f32 %f394, %f383, %f393; mul.ftz.f32 %f395, %f384, %f393; fma.rn.ftz.f32 %f530, %f388, %f391, %f395; fma.rn.ftz.f32 %f529, %f387, %f391, %f394; bra.uni $L__BB120_118; $L__BB120_123: ld.local.u32 %r58, [%rd65+64]; and.b32 %r187, %r58, 1; setp.eq.b32 %p84, %r187, 1; mov.pred %p85, 0; xor.pred %p86, %p84, %p85; not.pred %p87, %p86; @%p87 bra $L__BB120_134; bra.uni $L__BB120_124; $L__BB120_134: and.b32 %r209, %r58, 6; setp.eq.s32 %p94, %r209, 0; @%p94 bra $L__BB120_141; and.b32 %r210, %r32, 768; setp.eq.s32 %p95, %r210, 0; @%p95 bra $L__BB120_137; ld.const.u64 %rd249, [kernel_params+2264]; cvta.to.global.u64 %rd250, %rd249; ld.local.u32 %r211, [%rd65+60]; mul.wide.s32 %rd252, %r211, 16; add.s64 %rd253, %rd250, %rd252; shr.s32 %r212, %r58, 7; ld.global.u32 %r213, [%rd253+4]; add.s32 %r214, %r212, %r213; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r215, %r214, %r29; mul.wide.s32 %rd256, %r215, 4; add.s64 %rd257, %rd255, %rd256; ld.local.f32 %f415, [%rd65+68]; mov.f32 %f416, 0f3F800000; sub.ftz.f32 %f417, %f416, %f415; ld.global.f32 %f418, [%rd257]; mul.ftz.f32 %f419, %f418, %f417; ld.global.f32 %f420, [%rd257+4]; fma.rn.ftz.f32 %f531, %f420, %f415, %f419; bra.uni $L__BB120_143; $L__BB120_150: ld.local.v4.f32 {%f427, %f428, %f429, %f430}, [%rd65+192]; mul.ftz.f32 %f435, %f533, %f428; fma.rn.ftz.f32 %f436, %f534, %f427, %f435; fma.rn.ftz.f32 %f437, %f532, %f429, %f436; add.ftz.f32 %f170, %f430, %f437; ld.local.v4.f32 {%f438, %f439, %f440, %f441}, [%rd65+208]; mul.ftz.f32 %f446, %f533, %f439; fma.rn.ftz.f32 %f447, %f534, %f438, %f446; fma.rn.ftz.f32 %f448, %f532, %f440, %f447; add.ftz.f32 %f171, %f441, %f448; ld.local.v4.f32 {%f449, %f450, %f451, %f452}, [%rd65+224]; mul.ftz.f32 %f457, %f533, %f450; fma.rn.ftz.f32 %f458, %f534, %f449, %f457; fma.rn.ftz.f32 %f459, %f532, %f451, %f458; add.ftz.f32 %f532, %f452, %f459; mov.f32 %f533, %f171; mov.f32 %f534, %f170; $L__BB120_152: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB120_156; bra.uni $L__BB120_153; $L__BB120_156: add.ftz.f32 %f493, %f533, %f534; add.ftz.f32 %f494, %f532, %f493; mul.ftz.f32 %f495, %f494, 0f3EAAAAAB; st.local.f32 [%rd16], %f495; bra.uni $L__BB120_157; $L__BB120_153: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB120_155; mov.u32 %r223, 1065353216; st.local.u32 [%rd16], %r223; bra.uni $L__BB120_157; $L__BB120_19: add.s32 %r87, %r236, -3; setp.lt.u32 %p22, %r87, 2; ld.local.u32 %r41, [%rd65+64]; and.b32 %r42, %r41, 1; @%p22 bra $L__BB120_45; bra.uni $L__BB120_20; $L__BB120_45: setp.eq.s32 %p39, %r42, 0; @%p39 bra $L__BB120_84; bra.uni $L__BB120_46; $L__BB120_84: and.b32 %r141, %r41, 6; setp.eq.s32 %p56, %r141, 0; @%p56 bra $L__BB120_91; and.b32 %r142, %r32, 768; setp.eq.s32 %p57, %r142, 0; @%p57 bra $L__BB120_87; ld.const.u64 %rd158, [kernel_params+2264]; cvta.to.global.u64 %rd159, %rd158; ld.local.u32 %r143, [%rd65+60]; mul.wide.s32 %rd161, %r143, 16; add.s64 %rd162, %rd159, %rd161; shr.s32 %r144, %r41, 7; ld.global.u32 %r145, [%rd162+4]; add.s32 %r146, %r144, %r145; ld.const.u64 %rd163, [kernel_params+2344]; cvta.to.global.u64 %rd164, %rd163; add.s32 %r147, %r146, %r29; mul.wide.s32 %rd165, %r147, 16; add.s64 %rd166, %rd164, %rd165; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd166]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd166+16]; ld.local.f32 %f332, [%rd65+68]; mov.f32 %f333, 0f3F800000; sub.ftz.f32 %f334, %f333, %f332; mul.ftz.f32 %f335, %f316, %f334; mul.ftz.f32 %f336, %f317, %f334; mul.ftz.f32 %f337, %f318, %f334; mul.ftz.f32 %f338, %f319, %f334; fma.rn.ftz.f32 %f522, %f327, %f332, %f338; fma.rn.ftz.f32 %f521, %f326, %f332, %f337; fma.rn.ftz.f32 %f520, %f325, %f332, %f336; fma.rn.ftz.f32 %f519, %f324, %f332, %f335; bra.uni $L__BB120_93; $L__BB120_124: ld.local.s32 %rd53, [%rd65+60]; setp.eq.s64 %p88, %rd53, -1; @%p88 bra $L__BB120_127; ld.const.u64 %rd226, [kernel_params+2240]; cvta.to.global.u64 %rd227, %rd226; shl.b64 %rd228, %rd53, 2; add.s64 %rd229, %rd227, %rd228; ld.global.u32 %r188, [%rd229]; setp.eq.s32 %p89, %r188, -1; @%p89 bra $L__BB120_127; mov.u32 %r189, 0; mov.u64 %rd230, 0; { // callseq 860, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r189; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd230; .param .b64 param3; st.param.b64 [param3+0], %rd230; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f531, [retval0+0]; } // callseq 860 bra.uni $L__BB120_143; $L__BB120_155: st.local.f32 [%rd16], %f534; st.local.f32 [%rd16+4], %f533; st.local.f32 [%rd16+8], %f532; bra.uni $L__BB120_157; $L__BB120_99: ld.local.s32 %rd51, [%rd65+60]; setp.eq.s64 %p69, %rd51, -1; @%p69 bra $L__BB120_102; ld.const.u64 %rd179, [kernel_params+2240]; cvta.to.global.u64 %rd180, %rd179; shl.b64 %rd181, %rd51, 2; add.s64 %rd182, %rd180, %rd181; ld.global.u32 %r154, [%rd182]; setp.eq.s32 %p70, %r154, -1; @%p70 bra $L__BB120_102; mov.u64 %rd183, 0; { // callseq 859, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r48; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd183; .param .b64 param3; st.param.b64 [param3+0], %rd183; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f529, [retval0+0]; ld.param.f32 %f530, [retval0+4]; } // callseq 859 bra.uni $L__BB120_118; $L__BB120_20: setp.eq.s32 %p23, %r42, 0; @%p23 bra $L__BB120_31; bra.uni $L__BB120_21; $L__BB120_31: and.b32 %r109, %r41, 6; setp.eq.s32 %p30, %r109, 0; @%p30 bra $L__BB120_38; and.b32 %r110, %r32, 768; setp.eq.s32 %p31, %r110, 0; @%p31 bra $L__BB120_34; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r111, [%rd65+60]; mul.wide.s32 %rd112, %r111, 16; add.s64 %rd113, %rd110, %rd112; shr.s32 %r112, %r41, 7; ld.global.u32 %r113, [%rd113+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd114, [kernel_params+2336]; cvta.to.global.u64 %rd115, %rd114; add.s32 %r115, %r114, %r29; mul.wide.s32 %rd116, %r115, 12; add.s64 %rd117, %rd115, %rd116; ld.local.f32 %f202, [%rd65+68]; mov.f32 %f203, 0f3F800000; sub.ftz.f32 %f204, %f203, %f202; ld.global.f32 %f205, [%rd117]; mul.ftz.f32 %f206, %f205, %f204; ld.global.f32 %f207, [%rd117+4]; mul.ftz.f32 %f208, %f207, %f204; ld.global.f32 %f209, [%rd117+8]; mul.ftz.f32 %f210, %f209, %f204; ld.global.f32 %f211, [%rd117+12]; ld.global.f32 %f212, [%rd117+16]; ld.global.f32 %f213, [%rd117+20]; fma.rn.ftz.f32 %f498, %f213, %f202, %f210; fma.rn.ftz.f32 %f497, %f212, %f202, %f208; fma.rn.ftz.f32 %f496, %f211, %f202, %f206; bra.uni $L__BB120_40; $L__BB120_127: and.b32 %r190, %r32, 56; setp.eq.s32 %p90, %r190, 0; @%p90 bra $L__BB120_132; and.b32 %r191, %r32, 24; setp.eq.s32 %p91, %r191, 0; @%p91 bra $L__BB120_130; ld.const.u64 %rd231, [kernel_params+2232]; cvta.to.global.u64 %rd232, %rd231; shl.b64 %rd233, %rd53, 4; add.s64 %rd234, %rd232, %rd233; ld.global.v4.u32 {%r192, %r193, %r194, %r195}, [%rd234]; ld.const.u64 %rd235, [kernel_params+2320]; cvta.to.global.u64 %rd236, %rd235; add.s32 %r199, %r192, %r29; mul.wide.u32 %rd237, %r199, 4; add.s64 %rd287, %rd236, %rd237; add.s32 %r200, %r193, %r29; mul.wide.u32 %rd238, %r200, 4; add.s64 %rd288, %rd236, %rd238; add.s32 %r201, %r194, %r29; mul.wide.u32 %rd239, %r201, 4; add.s64 %rd289, %rd236, %rd239; bra.uni $L__BB120_131; $L__BB120_102: and.b32 %r155, %r32, 56; setp.eq.s32 %p71, %r155, 0; @%p71 bra $L__BB120_107; and.b32 %r156, %r32, 24; setp.eq.s32 %p72, %r156, 0; @%p72 bra $L__BB120_105; ld.const.u64 %rd184, [kernel_params+2232]; cvta.to.global.u64 %rd185, %rd184; shl.b64 %rd186, %rd51, 4; add.s64 %rd187, %rd185, %rd186; ld.global.v4.u32 {%r157, %r158, %r159, %r160}, [%rd187]; ld.const.u64 %rd188, [kernel_params+2328]; cvta.to.global.u64 %rd189, %rd188; add.s32 %r164, %r157, %r29; mul.wide.u32 %rd190, %r164, 8; add.s64 %rd191, %rd189, %rd190; ld.global.v2.f32 {%f527, %f528}, [%rd191]; add.s32 %r165, %r158, %r29; mul.wide.u32 %rd192, %r165, 8; add.s64 %rd193, %rd189, %rd192; ld.global.v2.f32 {%f525, %f526}, [%rd193]; add.s32 %r166, %r159, %r29; mul.wide.u32 %rd194, %r166, 8; add.s64 %rd195, %rd189, %rd194; ld.global.v2.f32 {%f523, %f524}, [%rd195]; bra.uni $L__BB120_106; $L__BB120_46: ld.local.s32 %rd47, [%rd65+60]; setp.eq.s64 %p40, %rd47, -1; @%p40 bra $L__BB120_49; ld.const.u64 %rd129, [kernel_params+2240]; cvta.to.global.u64 %rd130, %rd129; shl.b64 %rd131, %rd47, 2; add.s64 %rd132, %rd130, %rd131; ld.global.u32 %r122, [%rd132]; setp.eq.s32 %p41, %r122, -1; @%p41 bra $L__BB120_49; mov.u64 %rd133, 0; { // callseq 858, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd133; .param .b64 param3; st.param.b64 [param3+0], %rd133; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f519, [retval0+0]; ld.param.f32 %f520, [retval0+4]; ld.param.f32 %f521, [retval0+8]; ld.param.f32 %f522, [retval0+12]; } // callseq 858 bra.uni $L__BB120_93; $L__BB120_21: ld.local.s32 %rd18, [%rd65+60]; setp.eq.s64 %p24, %rd18, -1; @%p24 bra $L__BB120_24; ld.const.u64 %rd85, [kernel_params+2240]; cvta.to.global.u64 %rd86, %rd85; shl.b64 %rd87, %rd18, 2; add.s64 %rd88, %rd86, %rd87; ld.global.u32 %r88, [%rd88]; setp.eq.s32 %p25, %r88, -1; @%p25 bra $L__BB120_24; mov.u64 %rd89, 0; { // callseq 857, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd89; .param .b64 param3; st.param.b64 [param3+0], %rd89; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f496, [retval0+0]; ld.param.f32 %f497, [retval0+4]; ld.param.f32 %f498, [retval0+8]; } // callseq 857 bra.uni $L__BB120_40; $L__BB120_141: and.b32 %r218, %r58, 8; setp.eq.s32 %p98, %r218, 0; setp.ne.s32 %p99, %r32, 8; mov.f32 %f531, 0f00000000; or.pred %p100, %p99, %p98; @%p100 bra $L__BB120_143; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; ld.local.u32 %r219, [%rd65+60]; add.s32 %r220, %r219, %r29; mul.wide.s32 %rd266, %r220, 4; add.s64 %rd267, %rd264, %rd266; ld.global.f32 %f531, [%rd267]; bra.uni $L__BB120_143; $L__BB120_116: and.b32 %r182, %r51, 8; setp.eq.s32 %p79, %r182, 0; setp.ne.s32 %p80, %r32, 8; mov.f32 %f529, 0f00000000; or.pred %p81, %p80, %p79; mov.f32 %f530, %f529; @%p81 bra $L__BB120_118; ld.const.u64 %rd219, [kernel_params+2328]; cvta.to.global.u64 %rd220, %rd219; ld.local.u32 %r183, [%rd65+60]; add.s32 %r184, %r183, %r29; mul.wide.s32 %rd222, %r184, 8; add.s64 %rd223, %rd220, %rd222; ld.global.v2.f32 {%f529, %f530}, [%rd223]; bra.uni $L__BB120_118; $L__BB120_49: and.b32 %r123, %r32, 120; setp.eq.s32 %p42, %r123, 0; @%p42 bra $L__BB120_82; and.b32 %r124, %r32, 24; setp.eq.s32 %p43, %r124, 0; @%p43 bra $L__BB120_52; ld.const.u64 %rd134, [kernel_params+2232]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd47, 4; add.s64 %rd137, %rd135, %rd136; ld.global.v4.u32 {%r125, %r126, %r127, %r128}, [%rd137]; ld.const.u64 %rd138, [kernel_params+2344]; cvta.to.global.u64 %rd139, %rd138; add.s32 %r132, %r125, %r29; mul.wide.u32 %rd140, %r132, 16; add.s64 %rd141, %rd139, %rd140; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd141]; add.s32 %r133, %r126, %r29; mul.wide.u32 %rd142, %r133, 16; add.s64 %rd143, %rd139, %rd142; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd143]; add.s32 %r134, %r127, %r29; mul.wide.u32 %rd144, %r134, 16; add.s64 %rd145, %rd139, %rd144; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd145]; bra.uni $L__BB120_81; $L__BB120_24: and.b32 %r89, %r32, 56; setp.eq.s32 %p26, %r89, 0; @%p26 bra $L__BB120_29; and.b32 %r90, %r32, 24; setp.eq.s32 %p27, %r90, 0; @%p27 bra $L__BB120_27; ld.const.u64 %rd90, [kernel_params+2232]; cvta.to.global.u64 %rd91, %rd90; shl.b64 %rd92, %rd18, 4; add.s64 %rd93, %rd91, %rd92; ld.global.v4.u32 {%r91, %r92, %r93, %r94}, [%rd93]; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; add.s32 %r98, %r91, %r29; mul.wide.u32 %rd96, %r98, 12; add.s64 %rd284, %rd95, %rd96; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r99, %r92, %r29; mul.wide.u32 %rd97, %r99, 12; add.s64 %rd281, %rd95, %rd97; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r100, %r93, %r29; mul.wide.u32 %rd98, %r100, 12; add.s64 %rd278, %rd95, %rd98; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; bra.uni $L__BB120_28; $L__BB120_137: and.b32 %r216, %r32, 131; setp.eq.s32 %p96, %r216, 0; mov.f32 %f531, 0f00000000; @%p96 bra $L__BB120_143; setp.ne.s32 %p97, %r32, 128; @%p97 bra $L__BB120_140; ld.local.u32 %r217, [%rd65+60]; add.s32 %r29, %r217, %r29; $L__BB120_140: ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r29, 4; add.s64 %rd262, %rd260, %rd261; ld.global.f32 %f531, [%rd262]; bra.uni $L__BB120_143; $L__BB120_132: and.b32 %r205, %r32, 7; setp.eq.s32 %p92, %r205, 0; mov.f32 %f531, 0f00000000; @%p92 bra $L__BB120_143; cvt.u32.u64 %r206, %rd53; setp.eq.s32 %p93, %r32, 4; selp.b32 %r207, %r206, 0, %p93; add.s32 %r208, %r29, %r207; ld.const.u64 %rd245, [kernel_params+2320]; cvta.to.global.u64 %rd246, %rd245; mul.wide.s32 %rd247, %r208, 4; add.s64 %rd248, %rd246, %rd247; ld.global.f32 %f531, [%rd248]; bra.uni $L__BB120_143; $L__BB120_112: and.b32 %r180, %r32, 131; setp.eq.s32 %p77, %r180, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p77 bra $L__BB120_118; setp.ne.s32 %p78, %r32, 128; @%p78 bra $L__BB120_115; ld.local.u32 %r181, [%rd65+60]; add.s32 %r29, %r181, %r29; $L__BB120_115: ld.const.u64 %rd215, [kernel_params+2328]; cvta.to.global.u64 %rd216, %rd215; mul.wide.s32 %rd217, %r29, 8; add.s64 %rd218, %rd216, %rd217; ld.global.v2.f32 {%f529, %f530}, [%rd218]; bra.uni $L__BB120_118; $L__BB120_107: and.b32 %r169, %r32, 7; setp.eq.s32 %p73, %r169, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p73 bra $L__BB120_118; cvt.u32.u64 %r170, %rd51; setp.eq.s32 %p74, %r32, 4; selp.b32 %r171, %r170, 0, %p74; add.s32 %r172, %r29, %r171; ld.const.u64 %rd201, [kernel_params+2328]; cvta.to.global.u64 %rd202, %rd201; mul.wide.s32 %rd203, %r172, 8; add.s64 %rd204, %rd202, %rd203; ld.global.v2.f32 {%f529, %f530}, [%rd204]; bra.uni $L__BB120_118; $L__BB120_91: and.b32 %r150, %r41, 8; setp.eq.s32 %p60, %r150, 0; setp.ne.s32 %p61, %r32, 8; mov.f32 %f519, 0f00000000; or.pred %p62, %p61, %p60; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p62 bra $L__BB120_93; ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; ld.local.u32 %r151, [%rd65+60]; add.s32 %r152, %r151, %r29; mul.wide.s32 %rd175, %r152, 16; add.s64 %rd176, %rd173, %rd175; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd176]; bra.uni $L__BB120_93; $L__BB120_38: and.b32 %r118, %r41, 8; setp.eq.s32 %p34, %r118, 0; setp.ne.s32 %p35, %r32, 8; mov.f32 %f496, 0f00000000; or.pred %p36, %p35, %p34; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p36 bra $L__BB120_40; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r119, [%rd65+60]; add.s32 %r120, %r119, %r29; mul.wide.s32 %rd126, %r120, 12; add.s64 %rd127, %rd124, %rd126; ld.global.f32 %f498, [%rd127+8]; ld.global.f32 %f497, [%rd127+4]; ld.global.f32 %f496, [%rd127]; bra.uni $L__BB120_40; $L__BB120_130: cvt.u32.u64 %r202, %rd53; mad.lo.s32 %r203, %r202, 3, %r29; ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r203, 4; add.s64 %rd287, %rd241, %rd242; add.s64 %rd288, %rd287, 4; add.s32 %r204, %r203, 2; mul.wide.s32 %rd243, %r204, 4; add.s64 %rd289, %rd241, %rd243; $L__BB120_131: ld.local.f32 %f404, [%rd65+68]; ld.global.f32 %f405, [%rd288]; ld.local.f32 %f406, [%rd65+72]; ld.global.f32 %f407, [%rd289]; mul.ftz.f32 %f408, %f407, %f406; fma.rn.ftz.f32 %f409, %f405, %f404, %f408; mov.f32 %f410, 0f3F800000; sub.ftz.f32 %f411, %f410, %f404; sub.ftz.f32 %f412, %f411, %f406; ld.global.f32 %f413, [%rd287]; fma.rn.ftz.f32 %f531, %f413, %f412, %f409; $L__BB120_143: setp.eq.s32 %p101, %r64, 1; @%p101 bra $L__BB120_147; bra.uni $L__BB120_144; $L__BB120_147: st.local.f32 [%rd16], %f531; bra.uni $L__BB120_157; $L__BB120_144: setp.eq.s32 %p102, %r64, 0; @%p102 bra $L__BB120_146; mov.u32 %r221, 1065353216; st.local.u32 [%rd16], %r221; bra.uni $L__BB120_157; $L__BB120_146: st.local.f32 [%rd16], %f531; st.local.f32 [%rd16+4], %f531; st.local.f32 [%rd16+8], %f531; bra.uni $L__BB120_157; $L__BB120_105: cvt.u32.u64 %r167, %rd51; mad.lo.s32 %r168, %r167, 3, %r29; ld.const.u64 %rd196, [kernel_params+2328]; cvta.to.global.u64 %rd197, %rd196; mul.wide.s32 %rd198, %r168, 8; add.s64 %rd199, %rd197, %rd198; ld.global.v2.f32 {%f527, %f528}, [%rd199]; ld.global.v2.f32 {%f525, %f526}, [%rd199+8]; ld.global.v2.f32 {%f523, %f524}, [%rd199+16]; $L__BB120_106: ld.local.f32 %f370, [%rd65+68]; ld.local.f32 %f371, [%rd65+72]; mul.ftz.f32 %f372, %f523, %f371; mul.ftz.f32 %f373, %f524, %f371; fma.rn.ftz.f32 %f374, %f525, %f370, %f372; fma.rn.ftz.f32 %f375, %f526, %f370, %f373; mov.f32 %f376, 0f3F800000; sub.ftz.f32 %f377, %f376, %f370; sub.ftz.f32 %f378, %f377, %f371; fma.rn.ftz.f32 %f530, %f528, %f378, %f375; fma.rn.ftz.f32 %f529, %f527, %f378, %f374; $L__BB120_118: setp.eq.s32 %p82, %r64, 1; @%p82 bra $L__BB120_122; bra.uni $L__BB120_119; $L__BB120_122: st.local.f32 [%rd16], %f529; bra.uni $L__BB120_157; $L__BB120_119: setp.eq.s32 %p83, %r64, 0; @%p83 bra $L__BB120_121; mov.u32 %r185, 1065353216; st.local.u32 [%rd16], %r185; bra.uni $L__BB120_157; $L__BB120_121: st.local.f32 [%rd16], %f529; st.local.f32 [%rd16+4], %f530; mov.u32 %r186, 0; st.local.u32 [%rd16+8], %r186; bra.uni $L__BB120_157; $L__BB120_87: and.b32 %r148, %r32, 131; setp.eq.s32 %p58, %r148, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p58 bra $L__BB120_93; setp.ne.s32 %p59, %r32, 128; @%p59 bra $L__BB120_90; ld.local.u32 %r149, [%rd65+60]; add.s32 %r29, %r149, %r29; $L__BB120_90: ld.const.u64 %rd168, [kernel_params+2344]; cvta.to.global.u64 %rd169, %rd168; mul.wide.s32 %rd170, %r29, 16; add.s64 %rd171, %rd169, %rd170; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd171]; bra.uni $L__BB120_93; $L__BB120_34: and.b32 %r116, %r32, 131; setp.eq.s32 %p32, %r116, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p32 bra $L__BB120_40; setp.ne.s32 %p33, %r32, 128; @%p33 bra $L__BB120_37; ld.local.u32 %r117, [%rd65+60]; add.s32 %r29, %r117, %r29; $L__BB120_37: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r29, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f498, [%rd122+8]; ld.global.f32 %f497, [%rd122+4]; ld.global.f32 %f496, [%rd122]; bra.uni $L__BB120_40; $L__BB120_82: and.b32 %r137, %r32, 7; setp.eq.s32 %p54, %r137, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p54 bra $L__BB120_93; cvt.u32.u64 %r138, %rd47; setp.eq.s32 %p55, %r32, 4; selp.b32 %r139, %r138, 0, %p55; add.s32 %r140, %r29, %r139; ld.const.u64 %rd154, [kernel_params+2344]; cvta.to.global.u64 %rd155, %rd154; mul.wide.s32 %rd156, %r140, 16; add.s64 %rd157, %rd155, %rd156; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd157]; bra.uni $L__BB120_93; $L__BB120_29: and.b32 %r105, %r32, 7; setp.eq.s32 %p28, %r105, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p28 bra $L__BB120_40; cvt.u32.u64 %r106, %rd18; setp.eq.s32 %p29, %r32, 4; selp.b32 %r107, %r106, 0, %p29; add.s32 %r108, %r29, %r107; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r108, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f498, [%rd108+8]; ld.global.f32 %f497, [%rd108+4]; ld.global.f32 %f496, [%rd108]; bra.uni $L__BB120_40; $L__BB120_52: cvt.u32.u64 %r135, %rd47; mad.lo.s32 %r136, %r135, 3, %r29; cvt.s64.s32 %rd48, %r136; setp.eq.s32 %p44, %r32, 32; @%p44 bra $L__BB120_80; bra.uni $L__BB120_53; $L__BB120_80: ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd48, 4; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd152]; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd152+16]; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd152+32]; bra.uni $L__BB120_81; $L__BB120_27: cvt.u32.u64 %r101, %rd18; mad.lo.s32 %r102, %r101, 3, %r29; ld.const.u64 %rd99, [kernel_params+2336]; cvta.to.global.u64 %rd100, %rd99; mul.wide.s32 %rd101, %r102, 12; add.s64 %rd284, %rd100, %rd101; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r103, %r102, 1; mul.wide.s32 %rd102, %r103, 12; add.s64 %rd281, %rd100, %rd102; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r104, %r102, 2; mul.wide.s32 %rd103, %r104, 12; add.s64 %rd278, %rd100, %rd103; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; $L__BB120_28: ld.local.f32 %f179, [%rd65+68]; ld.global.f32 %f180, [%rd281]; ld.global.f32 %f181, [%rd282]; ld.global.f32 %f182, [%rd283]; ld.local.f32 %f183, [%rd65+72]; ld.global.f32 %f184, [%rd278]; mul.ftz.f32 %f185, %f184, %f183; ld.global.f32 %f186, [%rd279]; mul.ftz.f32 %f187, %f186, %f183; ld.global.f32 %f188, [%rd280]; mul.ftz.f32 %f189, %f188, %f183; fma.rn.ftz.f32 %f190, %f180, %f179, %f185; fma.rn.ftz.f32 %f191, %f181, %f179, %f187; fma.rn.ftz.f32 %f192, %f182, %f179, %f189; mov.f32 %f193, 0f3F800000; sub.ftz.f32 %f194, %f193, %f179; sub.ftz.f32 %f195, %f194, %f183; ld.global.f32 %f196, [%rd284]; ld.global.f32 %f197, [%rd285]; ld.global.f32 %f198, [%rd286]; fma.rn.ftz.f32 %f498, %f198, %f195, %f192; fma.rn.ftz.f32 %f497, %f197, %f195, %f191; fma.rn.ftz.f32 %f496, %f196, %f195, %f190; $L__BB120_40: setp.eq.s32 %p37, %r64, 1; @%p37 bra $L__BB120_44; bra.uni $L__BB120_41; $L__BB120_44: add.ftz.f32 %f220, %f496, %f497; add.ftz.f32 %f221, %f498, %f220; mul.ftz.f32 %f222, %f221, 0f3EAAAAAB; st.local.f32 [%rd16], %f222; bra.uni $L__BB120_157; $L__BB120_41: setp.eq.s32 %p38, %r64, 0; @%p38 bra $L__BB120_43; mov.u32 %r121, 1065353216; st.local.u32 [%rd16], %r121; bra.uni $L__BB120_157; $L__BB120_43: st.local.f32 [%rd16], %f496; st.local.f32 [%rd16+4], %f497; st.local.f32 [%rd16+8], %f498; bra.uni $L__BB120_157; $L__BB120_53: ld.const.u64 %rd146, [kernel_params+2352]; cvta.to.global.u64 %rd147, %rd146; shl.b64 %rd148, %rd48, 2; add.s64 %rd49, %rd147, %rd148; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd49]; cvt.rn.f32.u16 %f235, %rs20; mul.ftz.f32 %f38, %f235, 0f3B808081; cvt.rn.f32.u16 %f236, %rs21; mul.ftz.f32 %f39, %f236, 0f3B808081; cvt.rn.f32.u16 %f237, %rs22; mul.ftz.f32 %f40, %f237, 0f3B808081; setp.lt.ftz.f32 %p45, %f38, 0f3D25AEE6; @%p45 bra $L__BB120_55; bra.uni $L__BB120_54; $L__BB120_55: mul.ftz.f32 %f515, %f38, 0f3D9E8391; bra.uni $L__BB120_56; $L__BB120_54: add.ftz.f32 %f238, %f38, 0f3D6147AE; mul.ftz.f32 %f239, %f238, 0f3F72A76F; lg2.approx.ftz.f32 %f240, %f239; mul.ftz.f32 %f241, %f240, 0f4019999A; ex2.approx.ftz.f32 %f515, %f241; $L__BB120_56: setp.lt.ftz.f32 %p46, %f39, 0f3D25AEE6; @%p46 bra $L__BB120_58; bra.uni $L__BB120_57; $L__BB120_58: mul.ftz.f32 %f516, %f39, 0f3D9E8391; bra.uni $L__BB120_59; $L__BB120_57: add.ftz.f32 %f242, %f39, 0f3D6147AE; mul.ftz.f32 %f243, %f242, 0f3F72A76F; lg2.approx.ftz.f32 %f244, %f243; mul.ftz.f32 %f245, %f244, 0f4019999A; ex2.approx.ftz.f32 %f516, %f245; $L__BB120_59: setp.lt.ftz.f32 %p47, %f40, 0f3D25AEE6; @%p47 bra $L__BB120_61; bra.uni $L__BB120_60; $L__BB120_61: mul.ftz.f32 %f517, %f40, 0f3D9E8391; bra.uni $L__BB120_62; $L__BB120_60: add.ftz.f32 %f246, %f40, 0f3D6147AE; mul.ftz.f32 %f247, %f246, 0f3F72A76F; lg2.approx.ftz.f32 %f248, %f247; mul.ftz.f32 %f249, %f248, 0f4019999A; ex2.approx.ftz.f32 %f517, %f249; $L__BB120_62: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd49+4]; cvt.rn.f32.u16 %f250, %rs24; mul.ftz.f32 %f50, %f250, 0f3B808081; cvt.rn.f32.u16 %f251, %rs25; mul.ftz.f32 %f51, %f251, 0f3B808081; cvt.rn.f32.u16 %f252, %rs26; mul.ftz.f32 %f52, %f252, 0f3B808081; setp.lt.ftz.f32 %p48, %f50, 0f3D25AEE6; @%p48 bra $L__BB120_64; bra.uni $L__BB120_63; $L__BB120_64: mul.ftz.f32 %f511, %f50, 0f3D9E8391; bra.uni $L__BB120_65; $L__BB120_63: add.ftz.f32 %f253, %f50, 0f3D6147AE; mul.ftz.f32 %f254, %f253, 0f3F72A76F; lg2.approx.ftz.f32 %f255, %f254; mul.ftz.f32 %f256, %f255, 0f4019999A; ex2.approx.ftz.f32 %f511, %f256; $L__BB120_65: setp.lt.ftz.f32 %p49, %f51, 0f3D25AEE6; @%p49 bra $L__BB120_67; bra.uni $L__BB120_66; $L__BB120_67: mul.ftz.f32 %f512, %f51, 0f3D9E8391; bra.uni $L__BB120_68; $L__BB120_66: add.ftz.f32 %f257, %f51, 0f3D6147AE; mul.ftz.f32 %f258, %f257, 0f3F72A76F; lg2.approx.ftz.f32 %f259, %f258; mul.ftz.f32 %f260, %f259, 0f4019999A; ex2.approx.ftz.f32 %f512, %f260; $L__BB120_68: setp.lt.ftz.f32 %p50, %f52, 0f3D25AEE6; @%p50 bra $L__BB120_70; bra.uni $L__BB120_69; $L__BB120_70: mul.ftz.f32 %f513, %f52, 0f3D9E8391; bra.uni $L__BB120_71; $L__BB120_69: add.ftz.f32 %f261, %f52, 0f3D6147AE; mul.ftz.f32 %f262, %f261, 0f3F72A76F; lg2.approx.ftz.f32 %f263, %f262; mul.ftz.f32 %f264, %f263, 0f4019999A; ex2.approx.ftz.f32 %f513, %f264; $L__BB120_71: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd49+8]; cvt.rn.f32.u16 %f265, %rs28; mul.ftz.f32 %f62, %f265, 0f3B808081; cvt.rn.f32.u16 %f266, %rs29; mul.ftz.f32 %f63, %f266, 0f3B808081; cvt.rn.f32.u16 %f267, %rs30; mul.ftz.f32 %f64, %f267, 0f3B808081; setp.lt.ftz.f32 %p51, %f62, 0f3D25AEE6; @%p51 bra $L__BB120_73; bra.uni $L__BB120_72; $L__BB120_73: mul.ftz.f32 %f507, %f62, 0f3D9E8391; bra.uni $L__BB120_74; $L__BB120_72: add.ftz.f32 %f268, %f62, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f507, %f271; $L__BB120_74: setp.lt.ftz.f32 %p52, %f63, 0f3D25AEE6; @%p52 bra $L__BB120_76; bra.uni $L__BB120_75; $L__BB120_76: mul.ftz.f32 %f508, %f63, 0f3D9E8391; bra.uni $L__BB120_77; $L__BB120_75: add.ftz.f32 %f272, %f63, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f508, %f275; $L__BB120_77: cvt.rn.f32.u16 %f276, %rs23; mul.ftz.f32 %f518, %f276, 0f3B808081; cvt.rn.f32.u16 %f277, %rs27; mul.ftz.f32 %f514, %f277, 0f3B808081; cvt.rn.f32.u16 %f278, %rs31; mul.ftz.f32 %f510, %f278, 0f3B808081; setp.lt.ftz.f32 %p53, %f64, 0f3D25AEE6; @%p53 bra $L__BB120_79; bra.uni $L__BB120_78; $L__BB120_79: mul.ftz.f32 %f509, %f64, 0f3D9E8391; bra.uni $L__BB120_81; $L__BB120_78: add.ftz.f32 %f279, %f64, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f509, %f282; $L__BB120_81: ld.local.f32 %f295, [%rd65+68]; ld.local.f32 %f296, [%rd65+72]; mul.ftz.f32 %f297, %f507, %f296; mul.ftz.f32 %f298, %f508, %f296; mul.ftz.f32 %f299, %f509, %f296; mul.ftz.f32 %f300, %f510, %f296; fma.rn.ftz.f32 %f301, %f511, %f295, %f297; fma.rn.ftz.f32 %f302, %f512, %f295, %f298; fma.rn.ftz.f32 %f303, %f513, %f295, %f299; fma.rn.ftz.f32 %f304, %f514, %f295, %f300; mov.f32 %f305, 0f3F800000; sub.ftz.f32 %f306, %f305, %f295; sub.ftz.f32 %f307, %f306, %f296; fma.rn.ftz.f32 %f522, %f518, %f307, %f304; fma.rn.ftz.f32 %f521, %f517, %f307, %f303; fma.rn.ftz.f32 %f520, %f516, %f307, %f302; fma.rn.ftz.f32 %f519, %f515, %f307, %f301; $L__BB120_93: setp.eq.s32 %p63, %r64, 1; @%p63 bra $L__BB120_97; bra.uni $L__BB120_94; $L__BB120_97: add.ftz.f32 %f355, %f519, %f520; add.ftz.f32 %f356, %f521, %f355; mul.ftz.f32 %f357, %f356, 0f3EAAAAAB; st.local.f32 [%rd16], %f357; bra.uni $L__BB120_157; $L__BB120_94: setp.eq.s32 %p64, %r64, 0; @%p64 bra $L__BB120_96; st.local.f32 [%rd16], %f522; bra.uni $L__BB120_157; $L__BB120_96: st.local.f32 [%rd16], %f519; st.local.f32 [%rd16+4], %f520; st.local.f32 [%rd16+8], %f521; $L__BB120_157: ret; } .func _Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r1, [_Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; cvta.to.local.u64 %rd2, %rd1; add.s32 %r2, %r1, 1; ld.param.u8 %r3, [_Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+7]; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r4, 0; st.local.u32 [%rd4], %r4; st.local.u32 [%rd4+4], %r4; st.local.u32 [%rd4+8], %r4; st.param.b32 [func_retval0+0], %r2; ret; } .func _Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<58>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u64 %rd8, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd7, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; cvta.to.local.u64 %rd9, %rd8; add.s64 %rd1, %rd9, 88; mov.f32 %f20, 0f00000000; setp.gt.s32 %p1, %r2, 6; @%p1 bra $L__BB124_13; bra.uni $L__BB124_1; $L__BB124_13: setp.gt.s32 %p2, %r2, 10; @%p2 bra $L__BB124_23; setp.gt.s32 %p8, %r2, 8; @%p8 bra $L__BB124_18; setp.eq.s32 %p11, %r2, 7; @%p11 bra $L__BB124_41; setp.eq.s32 %p12, %r2, 8; @%p12 bra $L__BB124_17; bra.uni $L__BB124_46; $L__BB124_17: ld.local.u32 %r11, [%rd1+-36]; and.b32 %r12, %r11, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB124_46; $L__BB124_1: setp.gt.s32 %p13, %r2, 2; @%p13 bra $L__BB124_6; bra.uni $L__BB124_2; $L__BB124_6: setp.gt.s32 %p14, %r2, 4; @%p14 bra $L__BB124_10; setp.eq.s32 %p17, %r2, 3; @%p17 bra $L__BB124_43; setp.eq.s32 %p18, %r2, 4; @%p18 bra $L__BB124_9; bra.uni $L__BB124_46; $L__BB124_9: and.b32 %r16, %r4, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB124_46; $L__BB124_23: setp.gt.s32 %p3, %r2, 12; @%p3 bra $L__BB124_29; setp.eq.s32 %p6, %r2, 11; @%p6 bra $L__BB124_37; setp.eq.s32 %p7, %r2, 12; @%p7 bra $L__BB124_26; bra.uni $L__BB124_46; $L__BB124_26: and.b32 %r7, %r4, 768; mov.u64 %rd27, kernel_params; add.s64 %rd55, %rd27, 3400; setp.ne.s32 %p24, %r7, 0; @%p24 bra $L__BB124_28; add.s64 %rd55, %rd27, 2488; $L__BB124_28: ld.const.u64 %rd30, [%rd55]; cvta.to.global.u64 %rd31, %rd30; mul.wide.s32 %rd32, %r1, 2; add.s64 %rd33, %rd31, %rd32; ld.global.u16 %rs3, [%rd33]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB124_46; $L__BB124_2: setp.eq.s32 %p19, %r2, 0; @%p19 bra $L__BB124_45; setp.eq.s32 %p20, %r2, 1; @%p20 bra $L__BB124_44; setp.eq.s32 %p21, %r2, 2; @%p21 bra $L__BB124_5; bra.uni $L__BB124_46; $L__BB124_5: and.b32 %r18, %r4, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB124_46; $L__BB124_18: setp.eq.s32 %p9, %r2, 9; @%p9 bra $L__BB124_40; setp.eq.s32 %p10, %r2, 10; @%p10 bra $L__BB124_20; bra.uni $L__BB124_46; $L__BB124_20: and.b32 %r9, %r4, 768; mov.u64 %rd43, kernel_params; add.s64 %rd57, %rd43, 3376; setp.ne.s32 %p26, %r9, 0; @%p26 bra $L__BB124_22; add.s64 %rd57, %rd43, 2464; $L__BB124_22: ld.const.u64 %rd46, [%rd57]; cvta.to.global.u64 %rd47, %rd46; mul.wide.s32 %rd48, %r1, 2; add.s64 %rd49, %rd47, %rd48; ld.global.u16 %rs5, [%rd49]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r10, %r4, 525056; setp.eq.s32 %p27, %r10, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB124_46; $L__BB124_10: setp.eq.s32 %p15, %r2, 5; @%p15 bra $L__BB124_42; setp.eq.s32 %p16, %r2, 6; @%p16 bra $L__BB124_12; bra.uni $L__BB124_46; $L__BB124_12: and.b32 %r14, %r4, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB124_46; $L__BB124_29: setp.eq.s32 %p4, %r2, 13; @%p4 bra $L__BB124_34; setp.ne.s32 %p5, %r2, 14; @%p5 bra $L__BB124_46; and.b32 %r5, %r4, 768; mov.u64 %rd11, kernel_params; add.s64 %rd53, %rd11, 3408; setp.ne.s32 %p22, %r5, 0; @%p22 bra $L__BB124_33; add.s64 %rd53, %rd11, 2496; $L__BB124_33: ld.const.u64 %rd14, [%rd53]; cvta.to.global.u64 %rd15, %rd14; mul.wide.s32 %rd16, %r1, 2; add.s64 %rd17, %rd15, %rd16; ld.global.u16 %rs1, [%rd17]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB124_46; $L__BB124_41: and.b32 %r13, %r4, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB124_46; $L__BB124_43: and.b32 %r17, %r4, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB124_46; $L__BB124_37: and.b32 %r8, %r4, 768; mov.u64 %rd35, kernel_params; add.s64 %rd56, %rd35, 3392; setp.ne.s32 %p25, %r8, 0; @%p25 bra $L__BB124_39; add.s64 %rd56, %rd35, 2480; $L__BB124_39: ld.const.u64 %rd38, [%rd56]; cvta.to.global.u64 %rd39, %rd38; mul.wide.s32 %rd40, %r1, 2; add.s64 %rd41, %rd39, %rd40; ld.global.u16 %rs4, [%rd41]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB124_46; $L__BB124_45: and.b32 %r20, %r4, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB124_46; $L__BB124_44: and.b32 %r19, %r4, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB124_46; $L__BB124_40: ld.local.f32 %f20, [%rd1]; bra.uni $L__BB124_46; $L__BB124_42: and.b32 %r15, %r4, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB124_46; $L__BB124_34: and.b32 %r6, %r4, 768; mov.u64 %rd19, kernel_params; add.s64 %rd54, %rd19, 3384; setp.ne.s32 %p23, %r6, 0; @%p23 bra $L__BB124_36; add.s64 %rd54, %rd19, 2472; $L__BB124_36: ld.const.u64 %rd22, [%rd54]; cvta.to.global.u64 %rd23, %rd22; mul.wide.s32 %rd24, %r1, 2; add.s64 %rd25, %rd23, %rd24; ld.global.u16 %rs2, [%rd25]; cvt.rn.f32.u16 %f20, %rs2; $L__BB124_46: cvta.to.local.u64 %rd50, %rd7; mul.wide.u32 %rd51, %r3, 4; add.s64 %rd52, %rd50, %rd51; st.local.f32 [%rd52], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB125_2; bra.uni $L__BB125_1; $L__BB125_2: mov.b32 %f2764, %r395; bra.uni $L__BB125_3; $L__BB125_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB125_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB125_5; bra.uni $L__BB125_4; $L__BB125_5: mov.b32 %f9, %r396; bra.uni $L__BB125_6; $L__BB125_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB125_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB125_8; bra.uni $L__BB125_7; $L__BB125_8: mov.b32 %f2766, %r397; bra.uni $L__BB125_9; $L__BB125_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB125_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB125_11; bra.uni $L__BB125_10; $L__BB125_11: mov.b32 %f15, %r398; bra.uni $L__BB125_12; $L__BB125_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB125_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB125_14; bra.uni $L__BB125_13; $L__BB125_14: mov.b32 %f2768, %r399; bra.uni $L__BB125_15; $L__BB125_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB125_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB125_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB125_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB125_18; bra.uni $L__BB125_385; $L__BB125_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB125_259; bra.uni $L__BB125_19; $L__BB125_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB125_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB125_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB125_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB125_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB125_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB125_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB125_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB125_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB125_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB125_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB125_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB125_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB125_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB125_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB125_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB125_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB125_362; $L__BB125_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB125_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB125_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB125_40; bra.uni $L__BB125_29; $L__BB125_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB125_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB125_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB125_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB125_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB125_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB125_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB125_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB125_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB125_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB125_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB125_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB125_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB125_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB125_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB125_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB125_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB125_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB125_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB125_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB125_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB125_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB125_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB125_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB125_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB125_136; $L__BB125_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB125_371; bra.uni $L__BB125_366; $L__BB125_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB125_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB125_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB125_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB125_383; $L__BB125_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB125_149; bra.uni $L__BB125_139; $L__BB125_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB125_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB125_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB125_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB125_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB125_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB125_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB125_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB125_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB125_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB125_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB125_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB125_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB125_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB125_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB125_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB125_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB125_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB125_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB125_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB125_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB125_250; $L__BB125_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB125_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB125_253; bra.uni $L__BB125_362; $L__BB125_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB125_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB125_304; @%p171 bra $L__BB125_303; bra.uni $L__BB125_256; $L__BB125_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB125_305; $L__BB125_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB125_305; $L__BB125_256: @%p172 bra $L__BB125_302; bra.uni $L__BB125_257; $L__BB125_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB125_305; $L__BB125_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB125_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB125_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB125_308; bra.uni $L__BB125_306; $L__BB125_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB125_309; $L__BB125_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB125_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB125_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB125_316; @%p171 bra $L__BB125_315; bra.uni $L__BB125_311; $L__BB125_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB125_317; $L__BB125_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB125_317; $L__BB125_311: @%p172 bra $L__BB125_314; bra.uni $L__BB125_312; $L__BB125_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB125_317; $L__BB125_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB125_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB125_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB125_320; bra.uni $L__BB125_318; $L__BB125_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB125_321; $L__BB125_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB125_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB125_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB125_328; @%p171 bra $L__BB125_327; bra.uni $L__BB125_323; $L__BB125_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB125_329; $L__BB125_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB125_329; $L__BB125_323: @%p172 bra $L__BB125_326; bra.uni $L__BB125_324; $L__BB125_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB125_329; $L__BB125_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB125_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB125_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB125_332; bra.uni $L__BB125_330; $L__BB125_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB125_333; $L__BB125_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB125_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB125_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB125_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_362; $L__BB125_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB125_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB125_31; bra.uni $L__BB125_136; $L__BB125_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB125_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB125_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB125_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB125_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB125_107; @%p49 bra $L__BB125_106; bra.uni $L__BB125_37; $L__BB125_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB125_108; $L__BB125_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB125_108; $L__BB125_37: @%p50 bra $L__BB125_105; bra.uni $L__BB125_38; $L__BB125_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB125_108; $L__BB125_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB125_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB125_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB125_111; bra.uni $L__BB125_109; $L__BB125_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB125_112; $L__BB125_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB125_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB125_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB125_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB125_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB125_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB125_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_136; $L__BB125_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB125_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB125_368; bra.uni $L__BB125_383; $L__BB125_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB125_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB125_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB125_383; $L__BB125_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB125_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB125_141; bra.uni $L__BB125_250; $L__BB125_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB125_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB125_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB125_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB125_206; @%p101 bra $L__BB125_205; bra.uni $L__BB125_146; $L__BB125_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB125_207; $L__BB125_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB125_207; $L__BB125_146: @%p102 bra $L__BB125_204; bra.uni $L__BB125_147; $L__BB125_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB125_207; $L__BB125_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB125_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB125_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB125_210; bra.uni $L__BB125_208; $L__BB125_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB125_211; $L__BB125_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB125_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB125_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB125_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB125_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB125_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_250; $L__BB125_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB125_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB125_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB125_295; @%p151 bra $L__BB125_294; bra.uni $L__BB125_290; $L__BB125_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB125_296; $L__BB125_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB125_296; $L__BB125_290: @%p152 bra $L__BB125_293; bra.uni $L__BB125_291; $L__BB125_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB125_296; $L__BB125_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB125_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB125_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB125_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB125_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB125_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB125_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB125_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_362; $L__BB125_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB125_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB125_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB125_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB125_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB125_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB125_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB125_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB125_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB125_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB125_279; bra.uni $L__BB125_362; $L__BB125_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB125_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB125_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB125_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB125_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB125_96; @%p39 bra $L__BB125_95; bra.uni $L__BB125_91; $L__BB125_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB125_97; $L__BB125_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB125_97; $L__BB125_91: @%p40 bra $L__BB125_94; bra.uni $L__BB125_92; $L__BB125_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB125_97; $L__BB125_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB125_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB125_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB125_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB125_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB125_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB125_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB125_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB125_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB125_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_136; $L__BB125_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB125_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB125_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB125_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB125_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB125_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB125_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB125_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB125_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB125_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB125_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB125_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB125_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB125_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB125_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB125_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB125_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB125_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB125_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB125_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB125_76; bra.uni $L__BB125_136; $L__BB125_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB125_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB125_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB125_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB125_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB125_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_383; $L__BB125_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB125_383; $L__BB125_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB125_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB125_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB125_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB125_196; @%p92 bra $L__BB125_195; bra.uni $L__BB125_191; $L__BB125_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB125_197; $L__BB125_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB125_197; $L__BB125_191: @%p93 bra $L__BB125_194; bra.uni $L__BB125_192; $L__BB125_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB125_197; $L__BB125_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB125_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB125_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB125_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB125_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB125_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB125_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB125_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB125_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB125_250; $L__BB125_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB125_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB125_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB125_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB125_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB125_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB125_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB125_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB125_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB125_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB125_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB125_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB125_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB125_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB125_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB125_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB125_179; bra.uni $L__BB125_250; $L__BB125_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB125_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB125_337; @%p188 bra $L__BB125_336; bra.uni $L__BB125_23; $L__BB125_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB125_338; $L__BB125_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB125_338; $L__BB125_23: @%p189 bra $L__BB125_335; bra.uni $L__BB125_24; $L__BB125_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB125_338; $L__BB125_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB125_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB125_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB125_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB125_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB125_347; @%p188 bra $L__BB125_346; bra.uni $L__BB125_342; $L__BB125_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB125_348; $L__BB125_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB125_348; $L__BB125_342: @%p189 bra $L__BB125_345; bra.uni $L__BB125_343; $L__BB125_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB125_348; $L__BB125_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB125_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB125_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB125_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB125_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB125_357; @%p188 bra $L__BB125_356; bra.uni $L__BB125_352; $L__BB125_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB125_358; $L__BB125_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB125_358; $L__BB125_352: @%p189 bra $L__BB125_355; bra.uni $L__BB125_353; $L__BB125_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB125_358; $L__BB125_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB125_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB125_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB125_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB125_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB125_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB125_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB125_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB125_364: mov.f32 %f3392, %f3387; bra.uni $L__BB125_385; $L__BB125_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB125_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB125_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB125_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB125_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB125_128; @%p59 bra $L__BB125_127; bra.uni $L__BB125_123; $L__BB125_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB125_129; $L__BB125_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB125_129; $L__BB125_123: @%p60 bra $L__BB125_126; bra.uni $L__BB125_124; $L__BB125_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB125_129; $L__BB125_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB125_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB125_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB125_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB125_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB125_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB125_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB125_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB125_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB125_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB125_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB125_385; $L__BB125_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB125_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB125_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB125_385; $L__BB125_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB125_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB125_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB125_224; @%p120 bra $L__BB125_223; bra.uni $L__BB125_219; $L__BB125_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB125_225; $L__BB125_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB125_225; $L__BB125_219: @%p121 bra $L__BB125_222; bra.uni $L__BB125_220; $L__BB125_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB125_225; $L__BB125_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB125_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB125_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB125_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB125_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB125_234; @%p120 bra $L__BB125_233; bra.uni $L__BB125_229; $L__BB125_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB125_235; $L__BB125_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB125_235; $L__BB125_229: @%p121 bra $L__BB125_232; bra.uni $L__BB125_230; $L__BB125_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB125_235; $L__BB125_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB125_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB125_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB125_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB125_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB125_244; @%p120 bra $L__BB125_243; bra.uni $L__BB125_239; $L__BB125_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB125_245; $L__BB125_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB125_245; $L__BB125_239: @%p121 bra $L__BB125_242; bra.uni $L__BB125_240; $L__BB125_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB125_245; $L__BB125_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB125_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB125_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB125_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB125_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB125_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB125_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB125_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB125_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB125_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB125_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB125_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB125_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB125_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB125_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB125_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB125_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB125_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB125_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB125_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB127_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB127_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB127_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB127_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) { .reg .pred %p<2>; .reg .b32 %r<5>; ld.param.u32 %r1, [_Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u8 %r2, [_Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0+4]; setp.eq.s32 %p1, %r2, 39; selp.b32 %r3, 5, 1, %p1; add.s32 %r4, %r3, %r1; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB129_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB129_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB129_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB129_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB129_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB129_11; $L__BB129_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB129_8; add.s32 %r237, %r237, 2; bra.uni $L__BB129_9; $L__BB129_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB129_15; $L__BB129_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB129_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB129_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB129_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB129_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB129_15; $L__BB129_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB129_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB129_17; $L__BB129_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB129_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB129_159; bra.uni $L__BB129_18; $L__BB129_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB129_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB129_185; $L__BB129_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB129_150; bra.uni $L__BB129_19; $L__BB129_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB129_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB129_153; bra.uni $L__BB129_152; $L__BB129_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB129_154; $L__BB129_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB129_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB129_100; bra.uni $L__BB129_21; $L__BB129_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB129_111; bra.uni $L__BB129_101; $L__BB129_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB129_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB129_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB129_120; $L__BB129_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB129_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB129_164; bra.uni $L__BB129_163; $L__BB129_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB129_165; $L__BB129_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB129_136; bra.uni $L__BB129_126; $L__BB129_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB129_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB129_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB129_145; $L__BB129_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB129_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB129_158; bra.uni $L__BB129_155; $L__BB129_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB129_192; $L__BB129_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB129_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB129_192; $L__BB129_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB129_47; bra.uni $L__BB129_22; $L__BB129_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB129_86; bra.uni $L__BB129_48; $L__BB129_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB129_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB129_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB129_95; $L__BB129_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB129_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB129_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB129_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB129_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB129_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB129_184; bra.uni $L__BB129_170; $L__BB129_184: { // callseq 870, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 870 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB129_185; $L__BB129_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB129_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB129_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 864, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 864 bra.uni $L__BB129_145; $L__BB129_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB129_192; $L__BB129_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB129_183; bra.uni $L__BB129_171; $L__BB129_183: { // callseq 869, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 869 mov.f32 %f731, 0f3F800000; bra.uni $L__BB129_185; $L__BB129_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB129_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB129_104; mov.u64 %rd173, 0; { // callseq 863, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 863 bra.uni $L__BB129_120; $L__BB129_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB129_33; bra.uni $L__BB129_23; $L__BB129_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB129_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB129_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB129_42; $L__BB129_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB129_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB129_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB129_133; $L__BB129_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB129_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB129_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB129_108; $L__BB129_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB129_182; bra.uni $L__BB129_172; $L__BB129_182: { // callseq 868, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 868 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB129_185; $L__BB129_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB129_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB129_51; mov.u64 %rd129, 0; { // callseq 862, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 862 bra.uni $L__BB129_95; $L__BB129_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB129_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB129_26; mov.u64 %rd90, 0; { // callseq 861, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 861 bra.uni $L__BB129_42; $L__BB129_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB129_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB129_145; $L__BB129_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB129_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB129_120; $L__BB129_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB129_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB129_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB129_83; $L__BB129_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB129_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB129_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB129_30; $L__BB129_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB129_181; bra.uni $L__BB129_173; $L__BB129_181: { // callseq 867, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 867 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB129_185; $L__BB129_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB129_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB129_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB129_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB129_145; $L__BB129_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB129_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB129_145; $L__BB129_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB129_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB129_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB129_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB129_120; $L__BB129_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB129_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB129_120; $L__BB129_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB129_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB129_95; $L__BB129_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB129_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB129_42; $L__BB129_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB129_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB129_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB129_149; bra.uni $L__BB129_146; $L__BB129_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB129_192; $L__BB129_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB129_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB129_192; $L__BB129_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB129_192; $L__BB129_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB129_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB129_175; bra.uni $L__BB129_178; $L__BB129_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB129_177; bra.uni $L__BB129_176; $L__BB129_177: { // callseq 865, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 865 bra.uni $L__BB129_185; $L__BB129_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB129_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB129_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB129_124; bra.uni $L__BB129_121; $L__BB129_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB129_192; $L__BB129_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB129_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB129_192; $L__BB129_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB129_192; $L__BB129_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB129_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB129_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB129_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB129_95; $L__BB129_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB129_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB129_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB129_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB129_42; $L__BB129_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB129_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB129_95; $L__BB129_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB129_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB129_42; $L__BB129_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB129_82; bra.uni $L__BB129_55; $L__BB129_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB129_83; $L__BB129_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB129_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB129_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB129_46; bra.uni $L__BB129_43; $L__BB129_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB129_192; $L__BB129_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB129_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB129_192; $L__BB129_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB129_192; $L__BB129_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB129_185; $L__BB129_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB129_57; bra.uni $L__BB129_56; $L__BB129_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB129_58; $L__BB129_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB129_180; bra.uni $L__BB129_179; $L__BB129_180: { // callseq 866, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 866 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB129_185; $L__BB129_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB129_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB129_60; bra.uni $L__BB129_59; $L__BB129_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB129_61; $L__BB129_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB129_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB129_63; bra.uni $L__BB129_62; $L__BB129_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB129_64; $L__BB129_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB129_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB129_66; bra.uni $L__BB129_65; $L__BB129_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB129_67; $L__BB129_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB129_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB129_69; bra.uni $L__BB129_68; $L__BB129_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB129_70; $L__BB129_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB129_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB129_72; bra.uni $L__BB129_71; $L__BB129_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB129_73; $L__BB129_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB129_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB129_75; bra.uni $L__BB129_74; $L__BB129_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB129_76; $L__BB129_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB129_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB129_78; bra.uni $L__BB129_77; $L__BB129_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB129_79; $L__BB129_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB129_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB129_81; bra.uni $L__BB129_80; $L__BB129_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB129_83; $L__BB129_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB129_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB129_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB129_99; bra.uni $L__BB129_96; $L__BB129_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB129_192; $L__BB129_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB129_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB129_192; $L__BB129_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB129_192; $L__BB129_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB129_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB129_191; bra.uni $L__BB129_186; $L__BB129_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB129_192; $L__BB129_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB129_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB129_192; $L__BB129_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB129_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB129_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB129_192: ret; } .func _Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r1, [_Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; cvta.to.local.u64 %rd2, %rd1; add.s32 %r2, %r1, 1; ld.param.u8 %r3, [_Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+7]; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r4, 0; st.local.u32 [%rd4], %r4; st.local.u32 [%rd4+4], %r4; st.local.u32 [%rd4+8], %r4; st.param.b32 [func_retval0+0], %r2; ret; } .func _Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<13>; .reg .f32 %f<48>; .reg .b32 %r<24>; .reg .b64 %rd<31>; ld.param.u64 %rd10, [_Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd11, [_Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r10, [_Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r9, [_Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r8, [_Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd1, %rd10; cvta.to.local.u64 %rd2, %rd11; bfe.u32 %r2, %r8, 8, 8; bfe.u32 %r3, %r8, 16, 8; shr.u32 %r4, %r8, 24; setp.eq.s32 %p1, %r4, 255; mov.f32 %f45, 0f3F800000; @%p1 bra $L__BB133_2; mul.wide.u32 %rd12, %r4, 4; add.s64 %rd13, %rd2, %rd12; ld.local.f32 %f45, [%rd13]; $L__BB133_2: setp.eq.ftz.f32 %p2, %f45, 0f00000000; @%p2 bra $L__BB133_20; setp.eq.s32 %p3, %r2, 255; @%p3 bra $L__BB133_5; bra.uni $L__BB133_4; $L__BB133_5: mov.b32 %f46, %r9; bra.uni $L__BB133_6; $L__BB133_4: mul.wide.u32 %rd14, %r2, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f46, [%rd15]; $L__BB133_6: and.b32 %r11, %r8, 255; mov.f32 %f17, 0f00000000; max.ftz.f32 %f18, %f46, %f17; mul.ftz.f32 %f19, %f45, %f18; add.s64 %rd3, %rd1, 268; ld.local.f32 %f20, [%rd1+268]; mov.f32 %f21, 0f3F800000; sub.ftz.f32 %f22, %f21, %f20; ld.local.f32 %f23, [%rd1+272]; sub.ftz.f32 %f24, %f21, %f23; ld.local.f32 %f25, [%rd1+276]; sub.ftz.f32 %f26, %f21, %f25; setp.eq.s32 %p4, %r11, 37; selp.f32 %f27, %f26, %f25, %p4; selp.f32 %f28, %f24, %f23, %p4; selp.f32 %f29, %f22, %f20, %p4; mul.ftz.f32 %f6, %f19, %f29; mul.ftz.f32 %f7, %f19, %f28; mul.ftz.f32 %f8, %f19, %f27; setp.ne.s32 %p5, %r11, 38; @%p5 bra $L__BB133_17; max.ftz.f32 %f9, %f6, %f17; max.ftz.f32 %f10, %f7, %f17; add.ftz.f32 %f31, %f9, %f10; max.ftz.f32 %f11, %f8, %f17; add.ftz.f32 %f32, %f11, %f31; mul.ftz.f32 %f33, %f32, 0f3EAAAAAB; abs.ftz.f32 %f12, %f33; setp.ltu.ftz.f32 %p6, %f12, 0f3727C5AC; mov.u64 %rd29, 0; mov.u64 %rd30, %rd29; @%p6 bra $L__BB133_12; ld.local.u32 %r5, [%rd3+-4]; mov.u64 %rd29, 0; setp.eq.s32 %p7, %r5, 0; mov.u64 %rd27, %rd29; mov.u64 %rd28, %rd29; @%p7 bra $L__BB133_10; ld.local.u32 %r12, [%rd3+-8]; mul.wide.s32 %rd20, %r12, 80; add.s64 %rd21, %rd1, %rd20; add.s64 %rd27, %rd21, 304; add.s64 %rd22, %rd10, %rd20; add.s64 %rd28, %rd22, 304; mov.b32 %r13, %f11; mov.b32 %r14, %f10; mov.b32 %r15, %f9; mov.u32 %r16, 0; st.local.v4.u32 [%rd21+304], {%r15, %r14, %r13, %r16}; add.s32 %r17, %r12, 1; st.local.u32 [%rd3+-8], %r17; add.s32 %r18, %r5, -1; st.local.u32 [%rd3+-4], %r18; $L__BB133_10: setp.eq.s64 %p8, %rd28, 0; mov.u64 %rd30, %rd29; @%p8 bra $L__BB133_12; st.local.f32 [%rd27+16], %f12; mov.u64 %rd29, %rd27; mov.u64 %rd30, %rd28; $L__BB133_12: setp.eq.s64 %p9, %rd30, 0; @%p9 bra $L__BB133_17; setp.eq.s32 %p10, %r3, 255; @%p10 bra $L__BB133_15; bra.uni $L__BB133_14; $L__BB133_15: mov.b32 %f47, %r10; bra.uni $L__BB133_16; $L__BB133_14: mul.wide.u32 %rd25, %r3, 4; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f47, [%rd26]; $L__BB133_16: mov.u32 %r19, 38; st.local.u32 [%rd29+12], %r19; setp.lt.ftz.f32 %p11, %f47, 0f00000000; selp.f32 %f34, 0fBF800000, 0f3F800000, %p11; abs.ftz.f32 %f35, %f47; mov.f32 %f36, 0f3F7FBE77; min.ftz.f32 %f37, %f35, %f36; mul.ftz.f32 %f38, %f34, %f37; st.local.f32 [%rd29+32], %f38; ld.local.u32 %r20, [%rd3+-216]; or.b32 %r21, %r20, 128; st.local.u32 [%rd3+-216], %r21; $L__BB133_17: ld.local.u32 %r6, [%rd3+-216]; and.b32 %r22, %r6, 64; setp.eq.s32 %p12, %r22, 0; @%p12 bra $L__BB133_19; bra.uni $L__BB133_18; $L__BB133_19: or.b32 %r23, %r6, 64; st.local.u32 [%rd3+-216], %r23; st.local.f32 [%rd3+24], %f6; st.local.f32 [%rd3+28], %f7; st.local.f32 [%rd3+32], %f8; bra.uni $L__BB133_20; $L__BB133_18: ld.local.f32 %f39, [%rd3+24]; add.ftz.f32 %f40, %f6, %f39; ld.local.f32 %f41, [%rd3+28]; add.ftz.f32 %f42, %f7, %f41; ld.local.f32 %f43, [%rd3+32]; add.ftz.f32 %f44, %f8, %f43; st.local.f32 [%rd3+24], %f40; st.local.f32 [%rd3+28], %f42; st.local.f32 [%rd3+32], %f44; $L__BB133_20: ret; } .func (.param .b32 func_retval0) _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) { .reg .pred %p<126>; .reg .b16 %rs<34>; .reg .f32 %f<777>; .reg .b32 %r<259>; .reg .b64 %rd<207>; ld.param.u64 %rd64, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd65, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r113, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r114, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; ld.param.u32 %r112, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r111, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r110, [_Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd64; cvta.to.local.u64 %rd2, %rd65; ld.const.u64 %rd66, [kernel_params+2400]; cvta.to.global.u64 %rd67, %rd66; mul.wide.s32 %rd68, %r114, 16; add.s64 %rd69, %rd67, %rd68; ld.global.v4.u32 {%r115, %r116, %r117, %r118}, [%rd69]; ld.global.v4.u32 {%r119, %r120, %r121, %r122}, [%rd69+16]; and.b32 %r10, %r110, 255; bfe.u32 %r11, %r110, 8, 8; shr.u32 %r12, %r110, 24; setp.eq.s32 %p1, %r12, 255; mov.f32 %f728, 0f3F800000; @%p1 bra $L__BB134_2; mul.wide.u32 %rd70, %r12, 4; add.s64 %rd71, %rd2, %rd70; ld.local.f32 %f728, [%rd71]; $L__BB134_2: setp.eq.ftz.f32 %p2, %f728, 0f00000000; @%p2 bra $L__BB134_180; setp.eq.s32 %p3, %r10, 255; @%p3 bra $L__BB134_5; bra.uni $L__BB134_4; $L__BB134_5: mov.b32 %f729, %r115; bra.uni $L__BB134_6; $L__BB134_4: mul.wide.u32 %rd72, %r10, 4; add.s64 %rd73, %rd2, %rd72; ld.local.f32 %f729, [%rd73]; $L__BB134_6: mov.f32 %f255, 0f00000000; max.ftz.f32 %f256, %f729, %f255; mul.ftz.f32 %f740, %f728, %f256; setp.leu.ftz.f32 %p4, %f740, 0f3727C5AC; @%p4 bra $L__BB134_51; cvt.u64.u32 %rd4, %r119; ld.local.s32 %rd6, [%rd1+76]; ld.local.s32 %rd7, [%rd1+60]; setp.eq.s64 %p5, %rd6, -1; mov.u32 %r238, -1; mov.u32 %r237, 0; @%p5 bra $L__BB134_21; ld.const.u64 %rd74, [kernel_params+2160]; cvta.to.global.u64 %rd75, %rd74; mul.lo.s64 %rd76, %rd6, 208; add.s64 %rd77, %rd75, %rd76; ld.global.u32 %r13, [%rd77+164]; ld.local.u32 %r128, [%rd1+64]; and.b32 %r129, %r128, 1; setp.eq.b32 %p6, %r129, 1; mov.pred %p7, 0; xor.pred %p8, %p6, %p7; not.pred %p9, %p8; mov.u32 %r232, 0; @%p9 bra $L__BB134_12; cvt.u32.u64 %r131, %rd7; setp.eq.s32 %p10, %r131, -1; mov.u32 %r231, -1; @%p10 bra $L__BB134_11; ld.const.u64 %rd78, [kernel_params+2240]; cvta.to.global.u64 %rd79, %rd78; shl.b64 %rd80, %rd7, 2; add.s64 %rd81, %rd79, %rd80; ld.global.u32 %r231, [%rd81]; $L__BB134_11: setp.ne.s32 %p11, %r231, -1; selp.u32 %r232, 1, 0, %p11; $L__BB134_12: ld.const.u64 %rd82, [kernel_params+2312]; cvta.to.global.u64 %rd8, %rd82; add.s32 %r234, %r232, %r13; mul.wide.u32 %rd83, %r234, 16; add.s64 %rd84, %rd8, %rd83; ld.global.u16 %rs28, [%rd84+12]; ld.global.u32 %r233, [%rd84+8]; ld.global.u64 %rd199, [%rd84]; setp.eq.s64 %p12, %rd199, %rd4; @%p12 bra $L__BB134_17; $L__BB134_13: setp.eq.s64 %p13, %rd199, 0; @%p13 bra $L__BB134_15; add.s32 %r234, %r234, 2; bra.uni $L__BB134_16; $L__BB134_15: setp.eq.s16 %p14, %rs28, 0; mov.u32 %r237, 0; mov.u32 %r234, %r233; @%p14 bra $L__BB134_21; $L__BB134_16: mul.wide.u32 %rd85, %r234, 16; add.s64 %rd86, %rd8, %rd85; ld.global.u16 %rs28, [%rd86+12]; ld.global.u32 %r233, [%rd86+8]; ld.global.u64 %rd199, [%rd86]; setp.ne.s64 %p15, %rd199, %rd4; @%p15 bra $L__BB134_13; $L__BB134_17: cvt.u32.u64 %r136, %rd7; setp.ne.s32 %p16, %r136, -1; @%p16 bra $L__BB134_20; mov.u32 %r237, 0; add.s16 %rs19, %rs28, -1; setp.lt.u16 %p17, %rs19, 2; @%p17 bra $L__BB134_20; setp.ne.s16 %p18, %rs28, 1024; @%p18 bra $L__BB134_21; $L__BB134_20: setp.eq.s16 %p19, %rs28, 0; selp.b32 %r238, -1, %r233, %p19; cvt.u32.u16 %r237, %rs28; $L__BB134_21: setp.eq.s32 %p20, %r238, -1; @%p20 bra $L__BB134_51; ld.local.u32 %r143, [%rd1+64]; setp.ne.s32 %p21, %r143, 16; mov.f32 %f257, 0f00000000; mov.f32 %f739, %f257; @%p21 bra $L__BB134_50; and.b32 %r144, %r237, 3; setp.eq.s32 %p22, %r144, 0; @%p22 bra $L__BB134_25; ld.const.u64 %rd89, [kernel_params+2344]; cvta.to.global.u64 %rd90, %rd89; mul.wide.s32 %rd91, %r238, 16; add.s64 %rd92, %rd90, %rd91; ld.global.v4.f32 {%f736, %f737, %f738, %f261}, [%rd92]; bra.uni $L__BB134_49; $L__BB134_25: setp.ne.s32 %p23, %r237, 1024; mov.f32 %f736, 0f00000000; mov.f32 %f737, %f736; mov.f32 %f738, %f736; @%p23 bra $L__BB134_49; ld.local.f32 %f11, [%rd1]; ld.local.f32 %f12, [%rd1+4]; ld.local.f32 %f13, [%rd1+8]; ld.local.u8 %rs20, [%rd1+56]; and.b16 %rs21, %rs20, 2; setp.eq.s16 %p24, %rs21, 0; @%p24 bra $L__BB134_28; bra.uni $L__BB134_27; $L__BB134_28: cvt.u32.u64 %r145, %rd6; ld.const.u64 %rd93, [kernel_params+2160]; cvta.to.global.u64 %rd94, %rd93; mul.wide.s32 %rd95, %r145, 208; add.s64 %rd96, %rd94, %rd95; ld.global.v4.f32 {%f299, %f300, %f301, %f302}, [%rd96+48]; ld.global.v4.f32 {%f307, %f308, %f309, %f310}, [%rd96+64]; ld.global.v4.f32 {%f315, %f316, %f317, %f318}, [%rd96+80]; mul.ftz.f32 %f323, %f12, %f300; fma.rn.ftz.f32 %f324, %f11, %f299, %f323; fma.rn.ftz.f32 %f325, %f13, %f301, %f324; add.ftz.f32 %f26, %f302, %f325; mul.ftz.f32 %f326, %f12, %f308; fma.rn.ftz.f32 %f327, %f11, %f307, %f326; fma.rn.ftz.f32 %f328, %f13, %f309, %f327; add.ftz.f32 %f27, %f310, %f328; mul.ftz.f32 %f329, %f12, %f316; fma.rn.ftz.f32 %f330, %f11, %f315, %f329; fma.rn.ftz.f32 %f331, %f13, %f317, %f330; add.ftz.f32 %f28, %f318, %f331; bra.uni $L__BB134_29; $L__BB134_27: ld.local.v4.f32 {%f266, %f267, %f268, %f269}, [%rd1+192]; mul.ftz.f32 %f274, %f12, %f267; fma.rn.ftz.f32 %f275, %f11, %f266, %f274; fma.rn.ftz.f32 %f276, %f13, %f268, %f275; add.ftz.f32 %f26, %f269, %f276; ld.local.v4.f32 {%f277, %f278, %f279, %f280}, [%rd1+208]; mul.ftz.f32 %f285, %f12, %f278; fma.rn.ftz.f32 %f286, %f11, %f277, %f285; fma.rn.ftz.f32 %f287, %f13, %f279, %f286; add.ftz.f32 %f27, %f280, %f287; ld.local.v4.f32 {%f288, %f289, %f290, %f291}, [%rd1+224]; mul.ftz.f32 %f296, %f12, %f289; fma.rn.ftz.f32 %f297, %f11, %f288, %f296; fma.rn.ftz.f32 %f298, %f13, %f290, %f297; add.ftz.f32 %f28, %f291, %f298; $L__BB134_29: ld.local.u32 %r36, [%rd1+52]; ld.const.u64 %rd97, [kernel_params+2432]; cvta.to.global.u64 %rd98, %rd97; mul.wide.s32 %rd99, %r238, 96; add.s64 %rd19, %rd97, %rd99; add.s64 %rd100, %rd98, %rd99; add.s64 %rd20, %rd100, 32; ld.global.u32 %r146, [%rd100+32]; setp.eq.s32 %p25, %r146, 0; @%p25 bra $L__BB134_31; ld.global.v4.f32 {%f332, %f333, %f334, %f335}, [%rd20+16]; mul.ftz.f32 %f340, %f27, %f333; fma.rn.ftz.f32 %f341, %f26, %f332, %f340; fma.rn.ftz.f32 %f342, %f28, %f334, %f341; add.ftz.f32 %f23, %f335, %f342; ld.global.v4.f32 {%f343, %f344, %f345, %f346}, [%rd20+32]; mul.ftz.f32 %f351, %f27, %f344; fma.rn.ftz.f32 %f352, %f26, %f343, %f351; fma.rn.ftz.f32 %f353, %f28, %f345, %f352; add.ftz.f32 %f24, %f346, %f353; ld.global.v4.f32 {%f354, %f355, %f356, %f357}, [%rd20+48]; mul.ftz.f32 %f362, %f27, %f355; fma.rn.ftz.f32 %f363, %f26, %f354, %f362; fma.rn.ftz.f32 %f364, %f28, %f356, %f363; add.ftz.f32 %f28, %f357, %f364; mov.f32 %f26, %f23; mov.f32 %f27, %f24; $L__BB134_31: and.b32 %r147, %r36, 16777216; setp.ne.s32 %p26, %r147, 0; setp.eq.s32 %p27, %r147, 0; selp.b32 %r39, -1, 2, %p27; @%p26 bra $L__BB134_33; ld.global.u32 %r39, [%rd20+-20]; $L__BB134_33: ld.global.u32 %r40, [%rd20+-24]; setp.eq.s32 %p28, %r40, 8; @%p28 bra $L__BB134_48; bra.uni $L__BB134_34; $L__BB134_48: { // callseq 876, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .b32 param4; st.param.b32 [param4+0], %r39; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f736, [retval0+0]; } // callseq 876 mov.f32 %f737, %f736; mov.f32 %f738, %f736; bra.uni $L__BB134_49; $L__BB134_34: setp.eq.s32 %p29, %r40, 9; @%p29 bra $L__BB134_47; bra.uni $L__BB134_35; $L__BB134_47: { // callseq 875, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .b32 param4; st.param.b32 [param4+0], %r39; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f736, [retval0+0]; ld.param.f32 %f737, [retval0+4]; ld.param.f32 %f738, [retval0+8]; } // callseq 875 bra.uni $L__BB134_49; $L__BB134_35: setp.eq.s32 %p30, %r40, 10; @%p30 bra $L__BB134_46; bra.uni $L__BB134_36; $L__BB134_46: { // callseq 874, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .b32 param4; st.param.b32 [param4+0], %r39; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f736, [retval0+0]; } // callseq 874 mov.f32 %f737, %f736; mov.f32 %f738, %f736; bra.uni $L__BB134_49; $L__BB134_36: setp.eq.s32 %p31, %r40, 11; @%p31 bra $L__BB134_45; bra.uni $L__BB134_37; $L__BB134_45: { // callseq 873, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .b32 param4; st.param.b32 [param4+0], %r39; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f736, [retval0+0]; } // callseq 873 mov.f32 %f737, %f736; mov.f32 %f738, %f736; bra.uni $L__BB134_49; $L__BB134_37: or.b32 %r41, %r39, 1; setp.gt.u32 %p32, %r40, 6; @%p32 bra $L__BB134_42; mov.u32 %r149, 1; shl.b32 %r150, %r149, %r40; and.b32 %r151, %r150, 71; setp.ne.s32 %p33, %r151, 0; @%p33 bra $L__BB134_39; bra.uni $L__BB134_42; $L__BB134_39: setp.eq.s32 %p34, %r41, 3; @%p34 bra $L__BB134_41; bra.uni $L__BB134_40; $L__BB134_41: { // callseq 871, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f736, [retval0+0]; ld.param.f32 %f737, [retval0+4]; ld.param.f32 %f738, [retval0+8]; ld.param.f32 %f36, [retval0+12]; } // callseq 871 bra.uni $L__BB134_49; $L__BB134_40: ld.global.u64 %rd101, [%rd20+-32]; tex.3d.v4.f32.f32 {%f736, %f737, %f738, %f32}, [%rd101, {%f26, %f27, %f28, %f28}]; $L__BB134_49: add.ftz.f32 %f368, %f736, %f737; add.ftz.f32 %f369, %f738, %f368; mul.ftz.f32 %f739, %f369, 0f3EAAAAAB; $L__BB134_50: mul.ftz.f32 %f370, %f740, %f739; max.ftz.f32 %f740, %f370, %f257; $L__BB134_51: setp.leu.ftz.f32 %p36, %f740, 0f3727C5AC; @%p36 bra $L__BB134_111; ld.local.f32 %f756, [%rd1+268]; ld.local.f32 %f755, [%rd1+272]; ld.local.f32 %f754, [%rd1+276]; cvt.u64.u32 %rd22, %r120; ld.local.s32 %rd23, [%rd1+76]; ld.local.s32 %rd24, [%rd1+60]; setp.eq.s64 %p37, %rd23, -1; mov.u32 %r247, -1; mov.u32 %r246, 0; @%p37 bra $L__BB134_66; ld.const.u64 %rd103, [kernel_params+2160]; cvta.to.global.u64 %rd104, %rd103; mul.lo.s64 %rd105, %rd23, 208; add.s64 %rd106, %rd104, %rd105; ld.global.u32 %r42, [%rd106+164]; ld.local.u32 %r157, [%rd1+64]; and.b32 %r158, %r157, 1; setp.eq.b32 %p38, %r158, 1; mov.pred %p39, 0; xor.pred %p40, %p38, %p39; not.pred %p41, %p40; mov.u32 %r241, 0; @%p41 bra $L__BB134_57; cvt.u32.u64 %r160, %rd24; setp.eq.s32 %p42, %r160, -1; mov.u32 %r240, -1; @%p42 bra $L__BB134_56; ld.const.u64 %rd107, [kernel_params+2240]; cvta.to.global.u64 %rd108, %rd107; shl.b64 %rd109, %rd24, 2; add.s64 %rd110, %rd108, %rd109; ld.global.u32 %r240, [%rd110]; $L__BB134_56: setp.ne.s32 %p43, %r240, -1; selp.u32 %r241, 1, 0, %p43; $L__BB134_57: ld.const.u64 %rd111, [kernel_params+2312]; cvta.to.global.u64 %rd25, %rd111; add.s32 %r243, %r241, %r42; mul.wide.u32 %rd112, %r243, 16; add.s64 %rd113, %rd25, %rd112; ld.global.u16 %rs30, [%rd113+12]; ld.global.u32 %r242, [%rd113+8]; ld.global.u64 %rd200, [%rd113]; setp.eq.s64 %p44, %rd200, %rd22; @%p44 bra $L__BB134_62; $L__BB134_58: setp.eq.s64 %p45, %rd200, 0; @%p45 bra $L__BB134_60; add.s32 %r243, %r243, 2; bra.uni $L__BB134_61; $L__BB134_60: setp.eq.s16 %p46, %rs30, 0; mov.u32 %r246, 0; mov.u32 %r243, %r242; @%p46 bra $L__BB134_66; $L__BB134_61: mul.wide.u32 %rd114, %r243, 16; add.s64 %rd115, %rd25, %rd114; ld.global.u16 %rs30, [%rd115+12]; ld.global.u32 %r242, [%rd115+8]; ld.global.u64 %rd200, [%rd115]; setp.ne.s64 %p47, %rd200, %rd22; @%p47 bra $L__BB134_58; $L__BB134_62: cvt.u32.u64 %r165, %rd24; setp.ne.s32 %p48, %r165, -1; @%p48 bra $L__BB134_65; mov.u32 %r246, 0; add.s16 %rs22, %rs30, -1; setp.lt.u16 %p49, %rs22, 2; @%p49 bra $L__BB134_65; setp.ne.s16 %p50, %rs30, 1024; @%p50 bra $L__BB134_66; $L__BB134_65: setp.eq.s16 %p51, %rs30, 0; selp.b32 %r247, -1, %r242, %p51; cvt.u32.u16 %r246, %rs30; $L__BB134_66: setp.eq.s32 %p52, %r247, -1; @%p52 bra $L__BB134_97; ld.local.u32 %r172, [%rd1+64]; setp.ne.s32 %p53, %r172, 16; mov.f32 %f747, 0f00000000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; @%p53 bra $L__BB134_96; and.b32 %r173, %r246, 3; setp.eq.s32 %p54, %r173, 0; @%p54 bra $L__BB134_70; ld.const.u64 %rd118, [kernel_params+2344]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r247, 16; add.s64 %rd121, %rd119, %rd120; ld.global.v4.f32 {%f747, %f748, %f749, %f750}, [%rd121]; bra.uni $L__BB134_94; $L__BB134_70: setp.ne.s32 %p55, %r246, 1024; mov.f32 %f747, 0f00000000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; mov.f32 %f750, %f747; @%p55 bra $L__BB134_94; ld.local.f32 %f75, [%rd1]; ld.local.f32 %f76, [%rd1+4]; ld.local.f32 %f77, [%rd1+8]; ld.local.u8 %rs23, [%rd1+56]; and.b16 %rs24, %rs23, 2; setp.eq.s16 %p56, %rs24, 0; @%p56 bra $L__BB134_73; bra.uni $L__BB134_72; $L__BB134_73: cvt.u32.u64 %r174, %rd23; ld.const.u64 %rd122, [kernel_params+2160]; cvta.to.global.u64 %rd123, %rd122; mul.wide.s32 %rd124, %r174, 208; add.s64 %rd125, %rd123, %rd124; ld.global.v4.f32 {%f416, %f417, %f418, %f419}, [%rd125+48]; ld.global.v4.f32 {%f424, %f425, %f426, %f427}, [%rd125+64]; ld.global.v4.f32 {%f432, %f433, %f434, %f435}, [%rd125+80]; mul.ftz.f32 %f440, %f76, %f417; fma.rn.ftz.f32 %f441, %f75, %f416, %f440; fma.rn.ftz.f32 %f442, %f77, %f418, %f441; add.ftz.f32 %f90, %f419, %f442; mul.ftz.f32 %f443, %f76, %f425; fma.rn.ftz.f32 %f444, %f75, %f424, %f443; fma.rn.ftz.f32 %f445, %f77, %f426, %f444; add.ftz.f32 %f91, %f427, %f445; mul.ftz.f32 %f446, %f76, %f433; fma.rn.ftz.f32 %f447, %f75, %f432, %f446; fma.rn.ftz.f32 %f448, %f77, %f434, %f447; add.ftz.f32 %f92, %f435, %f448; bra.uni $L__BB134_74; $L__BB134_72: ld.local.v4.f32 {%f383, %f384, %f385, %f386}, [%rd1+192]; mul.ftz.f32 %f391, %f76, %f384; fma.rn.ftz.f32 %f392, %f75, %f383, %f391; fma.rn.ftz.f32 %f393, %f77, %f385, %f392; add.ftz.f32 %f90, %f386, %f393; ld.local.v4.f32 {%f394, %f395, %f396, %f397}, [%rd1+208]; mul.ftz.f32 %f402, %f76, %f395; fma.rn.ftz.f32 %f403, %f75, %f394, %f402; fma.rn.ftz.f32 %f404, %f77, %f396, %f403; add.ftz.f32 %f91, %f397, %f404; ld.local.v4.f32 {%f405, %f406, %f407, %f408}, [%rd1+224]; mul.ftz.f32 %f413, %f76, %f406; fma.rn.ftz.f32 %f414, %f75, %f405, %f413; fma.rn.ftz.f32 %f415, %f77, %f407, %f414; add.ftz.f32 %f92, %f408, %f415; $L__BB134_74: ld.local.u32 %r65, [%rd1+52]; ld.const.u64 %rd126, [kernel_params+2432]; cvta.to.global.u64 %rd127, %rd126; mul.wide.s32 %rd128, %r247, 96; add.s64 %rd36, %rd126, %rd128; add.s64 %rd129, %rd127, %rd128; add.s64 %rd37, %rd129, 32; ld.global.u32 %r175, [%rd129+32]; setp.eq.s32 %p57, %r175, 0; @%p57 bra $L__BB134_76; ld.global.v4.f32 {%f449, %f450, %f451, %f452}, [%rd37+16]; mul.ftz.f32 %f457, %f91, %f450; fma.rn.ftz.f32 %f458, %f90, %f449, %f457; fma.rn.ftz.f32 %f459, %f92, %f451, %f458; add.ftz.f32 %f87, %f452, %f459; ld.global.v4.f32 {%f460, %f461, %f462, %f463}, [%rd37+32]; mul.ftz.f32 %f468, %f91, %f461; fma.rn.ftz.f32 %f469, %f90, %f460, %f468; fma.rn.ftz.f32 %f470, %f92, %f462, %f469; add.ftz.f32 %f88, %f463, %f470; ld.global.v4.f32 {%f471, %f472, %f473, %f474}, [%rd37+48]; mul.ftz.f32 %f479, %f91, %f472; fma.rn.ftz.f32 %f480, %f90, %f471, %f479; fma.rn.ftz.f32 %f481, %f92, %f473, %f480; add.ftz.f32 %f92, %f474, %f481; mov.f32 %f90, %f87; mov.f32 %f91, %f88; $L__BB134_76: and.b32 %r176, %r65, 16777216; setp.ne.s32 %p58, %r176, 0; setp.eq.s32 %p59, %r176, 0; selp.b32 %r68, -1, 2, %p59; @%p58 bra $L__BB134_78; ld.global.u32 %r68, [%rd37+-20]; $L__BB134_78: ld.global.u32 %r69, [%rd37+-24]; setp.eq.s32 %p60, %r69, 8; @%p60 bra $L__BB134_93; bra.uni $L__BB134_79; $L__BB134_93: { // callseq 882, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .b32 param4; st.param.b32 [param4+0], %r68; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f747, [retval0+0]; } // callseq 882 mov.f32 %f750, 0f3F800000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; bra.uni $L__BB134_94; $L__BB134_79: setp.eq.s32 %p61, %r69, 9; @%p61 bra $L__BB134_92; bra.uni $L__BB134_80; $L__BB134_92: { // callseq 881, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .b32 param4; st.param.b32 [param4+0], %r68; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f747, [retval0+0]; ld.param.f32 %f748, [retval0+4]; ld.param.f32 %f749, [retval0+8]; } // callseq 881 mov.f32 %f750, 0f3F800000; bra.uni $L__BB134_94; $L__BB134_80: setp.eq.s32 %p62, %r69, 10; @%p62 bra $L__BB134_91; bra.uni $L__BB134_81; $L__BB134_91: { // callseq 880, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .b32 param4; st.param.b32 [param4+0], %r68; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f747, [retval0+0]; } // callseq 880 mov.f32 %f750, 0f3F800000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; bra.uni $L__BB134_94; $L__BB134_81: setp.eq.s32 %p63, %r69, 11; @%p63 bra $L__BB134_90; bra.uni $L__BB134_82; $L__BB134_90: { // callseq 879, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .b32 param4; st.param.b32 [param4+0], %r68; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f747, [retval0+0]; } // callseq 879 mov.f32 %f750, 0f3F800000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; bra.uni $L__BB134_94; $L__BB134_82: or.b32 %r70, %r68, 1; setp.gt.u32 %p64, %r69, 6; @%p64 bra $L__BB134_87; mov.u32 %r178, 1; shl.b32 %r179, %r178, %r69; and.b32 %r180, %r179, 71; setp.ne.s32 %p65, %r180, 0; @%p65 bra $L__BB134_84; bra.uni $L__BB134_87; $L__BB134_84: setp.eq.s32 %p66, %r70, 3; @%p66 bra $L__BB134_86; bra.uni $L__BB134_85; $L__BB134_86: { // callseq 877, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f747, [retval0+0]; ld.param.f32 %f748, [retval0+4]; ld.param.f32 %f749, [retval0+8]; ld.param.f32 %f750, [retval0+12]; } // callseq 877 bra.uni $L__BB134_94; $L__BB134_85: ld.global.u64 %rd130, [%rd37+-32]; tex.3d.v4.f32.f32 {%f747, %f748, %f749, %f750}, [%rd130, {%f90, %f91, %f92, %f92}]; $L__BB134_94: setp.leu.ftz.f32 %p68, %f750, 0f358637BD; setp.eq.ftz.f32 %p69, %f750, 0f3F800000; or.pred %p70, %p68, %p69; @%p70 bra $L__BB134_96; div.approx.ftz.f32 %f749, %f749, %f750; div.approx.ftz.f32 %f748, %f748, %f750; div.approx.ftz.f32 %f747, %f747, %f750; $L__BB134_96: mul.ftz.f32 %f756, %f756, %f747; mul.ftz.f32 %f755, %f755, %f748; mul.ftz.f32 %f754, %f754, %f749; $L__BB134_97: mul.ftz.f32 %f485, %f740, %f756; mov.f32 %f486, 0f00000000; max.ftz.f32 %f147, %f485, %f486; mul.ftz.f32 %f487, %f740, %f755; max.ftz.f32 %f148, %f487, %f486; mul.ftz.f32 %f488, %f740, %f754; max.ftz.f32 %f149, %f488, %f486; add.ftz.f32 %f489, %f147, %f148; add.ftz.f32 %f490, %f149, %f489; mul.ftz.f32 %f491, %f490, 0f3EAAAAAB; abs.ftz.f32 %f150, %f491; setp.ltu.ftz.f32 %p71, %f150, 0f3727C5AC; mov.u64 %rd203, 0; mov.u64 %rd204, %rd203; @%p71 bra $L__BB134_102; ld.local.u32 %r71, [%rd1+264]; mov.u64 %rd203, 0; setp.eq.s32 %p72, %r71, 0; mov.u64 %rd201, %rd203; mov.u64 %rd202, %rd203; @%p72 bra $L__BB134_100; ld.local.u32 %r181, [%rd1+260]; mul.wide.s32 %rd136, %r181, 80; add.s64 %rd137, %rd1, %rd136; add.s64 %rd201, %rd137, 304; add.s64 %rd138, %rd64, %rd136; add.s64 %rd202, %rd138, 304; mov.b32 %r182, %f149; mov.b32 %r183, %f148; mov.b32 %r184, %f147; mov.u32 %r185, 0; st.local.v4.u32 [%rd137+304], {%r184, %r183, %r182, %r185}; add.s32 %r186, %r181, 1; st.local.u32 [%rd1+260], %r186; add.s32 %r187, %r71, -1; st.local.u32 [%rd1+264], %r187; $L__BB134_100: setp.eq.s64 %p73, %rd202, 0; mov.u64 %rd204, %rd203; @%p73 bra $L__BB134_102; st.local.f32 [%rd201+16], %f150; mov.u64 %rd203, %rd201; mov.u64 %rd204, %rd202; $L__BB134_102: setp.eq.s64 %p74, %rd204, 0; @%p74 bra $L__BB134_107; setp.eq.s32 %p75, %r11, 255; @%p75 bra $L__BB134_105; bra.uni $L__BB134_104; $L__BB134_105: mov.b32 %f757, %r116; bra.uni $L__BB134_106; $L__BB134_107: ld.local.u32 %r249, [%rd1+52]; bra.uni $L__BB134_108; $L__BB134_104: mul.wide.u32 %rd141, %r11, 4; add.s64 %rd142, %rd2, %rd141; ld.local.f32 %f757, [%rd142]; $L__BB134_106: mov.u32 %r188, 38; st.local.u32 [%rd203+12], %r188; setp.lt.ftz.f32 %p76, %f757, 0f00000000; selp.f32 %f492, 0fBF800000, 0f3F800000, %p76; abs.ftz.f32 %f493, %f757; mov.f32 %f494, 0f3F7FBE77; min.ftz.f32 %f495, %f493, %f494; mul.ftz.f32 %f496, %f492, %f495; st.local.f32 [%rd203+32], %f496; ld.local.u32 %r189, [%rd1+52]; or.b32 %r249, %r189, 128; st.local.u32 [%rd1+52], %r249; $L__BB134_108: shr.u32 %r190, %r110, 14; cvt.u64.u32 %rd143, %r190; and.b64 %rd144, %rd143, 1020; add.s64 %rd145, %rd2, %rd144; ld.local.f32 %f497, [%rd145]; sqrt.approx.ftz.f32 %f498, %f497; ld.local.f32 %f499, [%rd145+4]; sqrt.approx.ftz.f32 %f500, %f499; ld.local.f32 %f501, [%rd145+8]; sqrt.approx.ftz.f32 %f502, %f501; max.ftz.f32 %f504, %f498, %f486; max.ftz.f32 %f505, %f500, %f486; max.ftz.f32 %f506, %f502, %f486; mov.f32 %f507, 0f3F800000; sub.ftz.f32 %f508, %f507, %f756; max.ftz.f32 %f509, %f508, %f486; sub.ftz.f32 %f510, %f507, %f755; max.ftz.f32 %f511, %f510, %f486; sub.ftz.f32 %f512, %f507, %f754; max.ftz.f32 %f513, %f512, %f486; sub.ftz.f32 %f514, %f507, %f504; sub.ftz.f32 %f515, %f507, %f505; sub.ftz.f32 %f516, %f507, %f506; max.ftz.f32 %f517, %f514, %f486; max.ftz.f32 %f518, %f515, %f486; max.ftz.f32 %f519, %f516, %f486; fma.rn.ftz.f32 %f520, %f509, %f517, %f756; fma.rn.ftz.f32 %f521, %f511, %f518, %f755; fma.rn.ftz.f32 %f522, %f513, %f519, %f754; mul.ftz.f32 %f154, %f740, %f520; mul.ftz.f32 %f155, %f740, %f521; mul.ftz.f32 %f156, %f740, %f522; and.b32 %r191, %r249, 64; setp.eq.s32 %p77, %r191, 0; @%p77 bra $L__BB134_110; bra.uni $L__BB134_109; $L__BB134_110: or.b32 %r192, %r249, 64; st.local.u32 [%rd1+52], %r192; st.local.f32 [%rd1+292], %f154; st.local.f32 [%rd1+296], %f155; st.local.f32 [%rd1+300], %f156; bra.uni $L__BB134_111; $L__BB134_109: ld.local.f32 %f523, [%rd1+292]; add.ftz.f32 %f524, %f154, %f523; ld.local.f32 %f525, [%rd1+296]; add.ftz.f32 %f526, %f155, %f525; ld.local.f32 %f527, [%rd1+300]; add.ftz.f32 %f528, %f156, %f527; st.local.f32 [%rd1+292], %f524; st.local.f32 [%rd1+296], %f526; st.local.f32 [%rd1+300], %f528; $L__BB134_111: and.b32 %r193, %r113, 768; setp.ne.s32 %p78, %r193, 0; @%p78 bra $L__BB134_180; and.b32 %r76, %r111, 255; bfe.u32 %r77, %r111, 16, 8; setp.eq.s32 %p79, %r76, 255; @%p79 bra $L__BB134_114; bra.uni $L__BB134_113; $L__BB134_114: mov.b32 %f758, %r117; bra.uni $L__BB134_115; $L__BB134_113: mul.wide.u32 %rd146, %r76, 4; add.s64 %rd147, %rd2, %rd146; ld.local.f32 %f758, [%rd147]; $L__BB134_115: setp.eq.s32 %p80, %r77, 255; @%p80 bra $L__BB134_117; bra.uni $L__BB134_116; $L__BB134_117: mov.b32 %f759, %r118; bra.uni $L__BB134_118; $L__BB134_116: mul.wide.u32 %rd148, %r77, 4; add.s64 %rd149, %rd2, %rd148; ld.local.f32 %f759, [%rd149]; $L__BB134_118: setp.leu.ftz.f32 %p81, %f758, 0f3727C5AC; @%p81 bra $L__BB134_122; shr.u32 %r194, %r111, 6; cvt.u64.u32 %rd150, %r194; and.b64 %rd151, %rd150, 1020; add.s64 %rd152, %rd2, %rd151; ld.local.f32 %f529, [%rd152]; mul.ftz.f32 %f163, %f758, %f529; ld.local.f32 %f530, [%rd152+4]; mul.ftz.f32 %f164, %f758, %f530; ld.local.f32 %f531, [%rd152+8]; mul.ftz.f32 %f165, %f758, %f531; ld.local.u32 %r78, [%rd1+52]; and.b32 %r195, %r78, 2; setp.eq.s32 %p82, %r195, 0; @%p82 bra $L__BB134_121; bra.uni $L__BB134_120; $L__BB134_121: or.b32 %r196, %r78, 2; st.local.u32 [%rd1+52], %r196; st.local.v2.f32 [%rd1+280], {%f163, %f164}; st.local.f32 [%rd1+288], %f165; bra.uni $L__BB134_122; $L__BB134_120: ld.local.f32 %f532, [%rd1+280]; add.ftz.f32 %f533, %f163, %f532; ld.local.f32 %f534, [%rd1+284]; add.ftz.f32 %f535, %f164, %f534; ld.local.f32 %f536, [%rd1+288]; add.ftz.f32 %f537, %f165, %f536; st.local.f32 [%rd1+280], %f533; st.local.f32 [%rd1+284], %f535; st.local.f32 [%rd1+288], %f537; $L__BB134_122: setp.leu.ftz.f32 %p83, %f759, 0f3727C5AC; @%p83 bra $L__BB134_180; shr.u32 %r201, %r111, 24; mul.wide.u32 %rd153, %r201, 4; add.s64 %rd154, %rd2, %rd153; ld.local.f32 %f770, [%rd154]; cvt.u64.u32 %rd45, %r121; ld.local.s32 %rd47, [%rd1+76]; ld.local.s32 %rd48, [%rd1+60]; setp.eq.s64 %p84, %rd47, -1; mov.u32 %r257, -1; mov.u32 %r256, 0; @%p84 bra $L__BB134_137; ld.const.u64 %rd155, [kernel_params+2160]; cvta.to.global.u64 %rd156, %rd155; mul.lo.s64 %rd157, %rd47, 208; add.s64 %rd158, %rd156, %rd157; ld.global.u32 %r79, [%rd158+164]; ld.local.u32 %r203, [%rd1+64]; and.b32 %r204, %r203, 1; setp.eq.b32 %p85, %r204, 1; mov.pred %p86, 0; xor.pred %p87, %p85, %p86; not.pred %p88, %p87; mov.u32 %r251, 0; @%p88 bra $L__BB134_128; cvt.u32.u64 %r206, %rd48; setp.eq.s32 %p89, %r206, -1; mov.u32 %r250, -1; @%p89 bra $L__BB134_127; ld.const.u64 %rd159, [kernel_params+2240]; cvta.to.global.u64 %rd160, %rd159; shl.b64 %rd161, %rd48, 2; add.s64 %rd162, %rd160, %rd161; ld.global.u32 %r250, [%rd162]; $L__BB134_127: setp.ne.s32 %p90, %r250, -1; selp.u32 %r251, 1, 0, %p90; $L__BB134_128: ld.const.u64 %rd163, [kernel_params+2312]; cvta.to.global.u64 %rd49, %rd163; add.s32 %r253, %r251, %r79; mul.wide.u32 %rd164, %r253, 16; add.s64 %rd165, %rd49, %rd164; ld.global.u16 %rs32, [%rd165+12]; ld.global.u32 %r252, [%rd165+8]; ld.global.u64 %rd205, [%rd165]; setp.eq.s64 %p91, %rd205, %rd45; @%p91 bra $L__BB134_133; $L__BB134_129: setp.eq.s64 %p92, %rd205, 0; @%p92 bra $L__BB134_131; add.s32 %r253, %r253, 2; bra.uni $L__BB134_132; $L__BB134_131: setp.eq.s16 %p93, %rs32, 0; mov.u32 %r256, 0; mov.u32 %r253, %r252; @%p93 bra $L__BB134_137; $L__BB134_132: mul.wide.u32 %rd166, %r253, 16; add.s64 %rd167, %rd49, %rd166; ld.global.u16 %rs32, [%rd167+12]; ld.global.u32 %r252, [%rd167+8]; ld.global.u64 %rd205, [%rd167]; setp.ne.s64 %p94, %rd205, %rd45; @%p94 bra $L__BB134_129; $L__BB134_133: cvt.u32.u64 %r211, %rd48; setp.ne.s32 %p95, %r211, -1; @%p95 bra $L__BB134_136; mov.u32 %r256, 0; add.s16 %rs25, %rs32, -1; setp.lt.u16 %p96, %rs25, 2; @%p96 bra $L__BB134_136; setp.ne.s16 %p97, %rs32, 1024; @%p97 bra $L__BB134_137; $L__BB134_136: setp.eq.s16 %p98, %rs32, 0; selp.b32 %r257, -1, %r252, %p98; cvt.u32.u16 %r256, %rs32; $L__BB134_137: setp.eq.s32 %p99, %r257, -1; @%p99 bra $L__BB134_167; ld.local.u32 %r218, [%rd1+64]; setp.ne.s32 %p100, %r218, 16; mov.f32 %f769, %f255; @%p100 bra $L__BB134_166; and.b32 %r219, %r256, 3; setp.eq.s32 %p101, %r219, 0; @%p101 bra $L__BB134_141; ld.const.u64 %rd170, [kernel_params+2344]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r257, 16; add.s64 %rd173, %rd171, %rd172; ld.global.v4.f32 {%f766, %f767, %f768, %f542}, [%rd173]; bra.uni $L__BB134_165; $L__BB134_141: setp.ne.s32 %p102, %r256, 1024; mov.f32 %f766, 0f00000000; mov.f32 %f767, %f766; mov.f32 %f768, %f766; @%p102 bra $L__BB134_165; ld.local.f32 %f171, [%rd1]; ld.local.f32 %f172, [%rd1+4]; ld.local.f32 %f173, [%rd1+8]; ld.local.u8 %rs26, [%rd1+56]; and.b16 %rs27, %rs26, 2; setp.eq.s16 %p103, %rs27, 0; @%p103 bra $L__BB134_144; bra.uni $L__BB134_143; $L__BB134_144: cvt.u32.u64 %r220, %rd47; ld.const.u64 %rd174, [kernel_params+2160]; cvta.to.global.u64 %rd175, %rd174; mul.wide.s32 %rd176, %r220, 208; add.s64 %rd177, %rd175, %rd176; ld.global.v4.f32 {%f580, %f581, %f582, %f583}, [%rd177+48]; ld.global.v4.f32 {%f588, %f589, %f590, %f591}, [%rd177+64]; ld.global.v4.f32 {%f596, %f597, %f598, %f599}, [%rd177+80]; mul.ftz.f32 %f604, %f172, %f581; fma.rn.ftz.f32 %f605, %f171, %f580, %f604; fma.rn.ftz.f32 %f606, %f173, %f582, %f605; add.ftz.f32 %f186, %f583, %f606; mul.ftz.f32 %f607, %f172, %f589; fma.rn.ftz.f32 %f608, %f171, %f588, %f607; fma.rn.ftz.f32 %f609, %f173, %f590, %f608; add.ftz.f32 %f187, %f591, %f609; mul.ftz.f32 %f610, %f172, %f597; fma.rn.ftz.f32 %f611, %f171, %f596, %f610; fma.rn.ftz.f32 %f612, %f173, %f598, %f611; add.ftz.f32 %f188, %f599, %f612; bra.uni $L__BB134_145; $L__BB134_143: ld.local.v4.f32 {%f547, %f548, %f549, %f550}, [%rd1+192]; mul.ftz.f32 %f555, %f172, %f548; fma.rn.ftz.f32 %f556, %f171, %f547, %f555; fma.rn.ftz.f32 %f557, %f173, %f549, %f556; add.ftz.f32 %f186, %f550, %f557; ld.local.v4.f32 {%f558, %f559, %f560, %f561}, [%rd1+208]; mul.ftz.f32 %f566, %f172, %f559; fma.rn.ftz.f32 %f567, %f171, %f558, %f566; fma.rn.ftz.f32 %f568, %f173, %f560, %f567; add.ftz.f32 %f187, %f561, %f568; ld.local.v4.f32 {%f569, %f570, %f571, %f572}, [%rd1+224]; mul.ftz.f32 %f577, %f172, %f570; fma.rn.ftz.f32 %f578, %f171, %f569, %f577; fma.rn.ftz.f32 %f579, %f173, %f571, %f578; add.ftz.f32 %f188, %f572, %f579; $L__BB134_145: ld.local.u32 %r102, [%rd1+52]; ld.const.u64 %rd178, [kernel_params+2432]; cvta.to.global.u64 %rd179, %rd178; mul.wide.s32 %rd180, %r257, 96; add.s64 %rd60, %rd178, %rd180; add.s64 %rd181, %rd179, %rd180; add.s64 %rd61, %rd181, 32; ld.global.u32 %r221, [%rd181+32]; setp.eq.s32 %p104, %r221, 0; @%p104 bra $L__BB134_147; ld.global.v4.f32 {%f613, %f614, %f615, %f616}, [%rd61+16]; mul.ftz.f32 %f621, %f187, %f614; fma.rn.ftz.f32 %f622, %f186, %f613, %f621; fma.rn.ftz.f32 %f623, %f188, %f615, %f622; add.ftz.f32 %f183, %f616, %f623; ld.global.v4.f32 {%f624, %f625, %f626, %f627}, [%rd61+32]; mul.ftz.f32 %f632, %f187, %f625; fma.rn.ftz.f32 %f633, %f186, %f624, %f632; fma.rn.ftz.f32 %f634, %f188, %f626, %f633; add.ftz.f32 %f184, %f627, %f634; ld.global.v4.f32 {%f635, %f636, %f637, %f638}, [%rd61+48]; mul.ftz.f32 %f643, %f187, %f636; fma.rn.ftz.f32 %f644, %f186, %f635, %f643; fma.rn.ftz.f32 %f645, %f188, %f637, %f644; add.ftz.f32 %f188, %f638, %f645; mov.f32 %f186, %f183; mov.f32 %f187, %f184; $L__BB134_147: and.b32 %r222, %r102, 16777216; setp.ne.s32 %p105, %r222, 0; setp.eq.s32 %p106, %r222, 0; selp.b32 %r105, -1, 2, %p106; @%p105 bra $L__BB134_149; ld.global.u32 %r105, [%rd61+-20]; $L__BB134_149: ld.global.u32 %r106, [%rd61+-24]; setp.eq.s32 %p107, %r106, 8; @%p107 bra $L__BB134_164; bra.uni $L__BB134_150; $L__BB134_164: { // callseq 888, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f766, [retval0+0]; } // callseq 888 mov.f32 %f767, %f766; mov.f32 %f768, %f766; bra.uni $L__BB134_165; $L__BB134_150: setp.eq.s32 %p108, %r106, 9; @%p108 bra $L__BB134_163; bra.uni $L__BB134_151; $L__BB134_163: { // callseq 887, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f766, [retval0+0]; ld.param.f32 %f767, [retval0+4]; ld.param.f32 %f768, [retval0+8]; } // callseq 887 bra.uni $L__BB134_165; $L__BB134_151: setp.eq.s32 %p109, %r106, 10; @%p109 bra $L__BB134_162; bra.uni $L__BB134_152; $L__BB134_162: { // callseq 886, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f766, [retval0+0]; } // callseq 886 mov.f32 %f767, %f766; mov.f32 %f768, %f766; bra.uni $L__BB134_165; $L__BB134_152: setp.eq.s32 %p110, %r106, 11; @%p110 bra $L__BB134_161; bra.uni $L__BB134_153; $L__BB134_161: { // callseq 885, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f766, [retval0+0]; } // callseq 885 mov.f32 %f767, %f766; mov.f32 %f768, %f766; bra.uni $L__BB134_165; $L__BB134_153: or.b32 %r107, %r105, 1; setp.gt.u32 %p111, %r106, 6; @%p111 bra $L__BB134_158; mov.u32 %r224, 1; shl.b32 %r225, %r224, %r106; and.b32 %r226, %r225, 71; setp.ne.s32 %p112, %r226, 0; @%p112 bra $L__BB134_155; bra.uni $L__BB134_158; $L__BB134_155: setp.eq.s32 %p113, %r107, 3; @%p113 bra $L__BB134_157; bra.uni $L__BB134_156; $L__BB134_157: { // callseq 883, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f766, [retval0+0]; ld.param.f32 %f767, [retval0+4]; ld.param.f32 %f768, [retval0+8]; ld.param.f32 %f196, [retval0+12]; } // callseq 883 bra.uni $L__BB134_165; $L__BB134_42: setp.eq.s32 %p35, %r41, 3; @%p35 bra $L__BB134_44; bra.uni $L__BB134_43; $L__BB134_44: { // callseq 872, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd19; .param .b32 param1; st.param.f32 [param1+0], %f26; .param .b32 param2; st.param.f32 [param2+0], %f27; .param .b32 param3; st.param.f32 [param3+0], %f28; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f736, [retval0+0]; } // callseq 872 mov.f32 %f737, %f736; mov.f32 %f738, %f736; bra.uni $L__BB134_49; $L__BB134_87: setp.eq.s32 %p67, %r70, 3; @%p67 bra $L__BB134_89; bra.uni $L__BB134_88; $L__BB134_89: { // callseq 878, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd36; .param .b32 param1; st.param.f32 [param1+0], %f90; .param .b32 param2; st.param.f32 [param2+0], %f91; .param .b32 param3; st.param.f32 [param3+0], %f92; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f747, [retval0+0]; } // callseq 878 mov.f32 %f750, 0f3F800000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; bra.uni $L__BB134_94; $L__BB134_156: ld.global.u64 %rd182, [%rd61+-32]; tex.3d.v4.f32.f32 {%f766, %f767, %f768, %f192}, [%rd182, {%f186, %f187, %f188, %f188}]; $L__BB134_165: add.ftz.f32 %f649, %f766, %f767; add.ftz.f32 %f650, %f768, %f649; mul.ftz.f32 %f769, %f650, 0f3EAAAAAB; $L__BB134_166: max.ftz.f32 %f652, %f769, %f255; mul.ftz.f32 %f770, %f770, %f652; $L__BB134_167: mov.f32 %f653, 0f00000000; max.ftz.f32 %f228, %f770, %f653; mul.ftz.f32 %f654, %f228, %f228; fma.rn.ftz.f32 %f655, %f654, %f654, 0fBF800000; fma.rn.ftz.f32 %f656, %f759, %f655, 0f3F800000; mul.ftz.f32 %f229, %f656, 0f28A292FC; setp.leu.ftz.f32 %p115, %f229, 0f3727C5AC; @%p115 bra $L__BB134_180; mul.wide.u32 %rd184, %r112, 4; add.s64 %rd185, %rd2, %rd184; ld.local.f32 %f660, [%rd185]; mul.ftz.f32 %f230, %f229, %f660; ld.local.f32 %f661, [%rd185+4]; mul.ftz.f32 %f231, %f229, %f661; ld.local.f32 %f662, [%rd185+8]; mul.ftz.f32 %f232, %f229, %f662; setp.ge.ftz.f32 %p116, %f228, 0f463B8000; mov.f32 %f773, 0f3FC87CC5; mov.f32 %f772, 0f3F7E9814; mov.f32 %f771, 0f3F53881A; @%p116 bra $L__BB134_175; setp.lt.ftz.f32 %p117, %f228, 0f44480000; mov.f32 %f773, 0fBDA87489; mov.f32 %f772, 0fBE501210; mov.f32 %f771, 0f40AD39B5; @%p117 bra $L__BB134_175; setp.ge.ftz.f32 %p118, %f228, 0f45C6E800; mov.u64 %rd206, 6; @%p118 bra $L__BB134_174; setp.ge.ftz.f32 %p119, %f228, 0f454F3000; mov.u64 %rd206, 5; @%p119 bra $L__BB134_174; setp.ge.ftz.f32 %p120, %f228, 0f44EDC000; mov.u64 %rd206, 4; @%p120 bra $L__BB134_174; setp.ge.ftz.f32 %p121, %f228, 0f44B52000; setp.ge.ftz.f32 %p122, %f228, 0f44714000; selp.u64 %rd189, 1, 0, %p122; setp.ge.ftz.f32 %p123, %f228, 0f4491E000; selp.b64 %rd190, 2, %rd189, %p123; selp.b64 %rd206, 3, %rd190, %p121; $L__BB134_174: mul.lo.s64 %rd191, %rd206, 12; mov.u64 %rd192, blackbody_table_r; add.s64 %rd193, %rd192, %rd191; ld.const.f32 %f666, [%rd193]; rcp.approx.ftz.f32 %f667, %f228; ld.const.f32 %f668, [%rd193+4]; mul.ftz.f32 %f669, %f228, %f668; fma.rn.ftz.f32 %f670, %f667, %f666, %f669; ld.const.f32 %f671, [%rd193+8]; mov.u64 %rd194, blackbody_table_g; add.s64 %rd195, %rd194, %rd191; ld.const.f32 %f672, [%rd195]; ld.const.f32 %f673, [%rd195+4]; mul.ftz.f32 %f674, %f228, %f673; fma.rn.ftz.f32 %f675, %f667, %f672, %f674; ld.const.f32 %f676, [%rd195+8]; shl.b64 %rd196, %rd206, 4; mov.u64 %rd197, blackbody_table_b; add.s64 %rd198, %rd197, %rd196; ld.const.f32 %f677, [%rd198]; ld.const.f32 %f678, [%rd198+4]; fma.rn.ftz.f32 %f679, %f228, %f677, %f678; ld.const.f32 %f680, [%rd198+8]; fma.rn.ftz.f32 %f681, %f228, %f679, %f680; ld.const.f32 %f682, [%rd198+12]; fma.rn.ftz.f32 %f773, %f228, %f681, %f682; add.ftz.f32 %f772, %f676, %f675; add.ftz.f32 %f771, %f671, %f670; $L__BB134_175: ld.const.u32 %r227, [kernel_params+1248]; setp.eq.s32 %p124, %r227, 0; @%p124 bra $L__BB134_176; bra.uni $L__BB134_177; $L__BB134_176: ld.const.v4.f32 {%f683, %f684, %f685, %f686}, [kernel_params+1200]; mul.ftz.f32 %f690, %f772, %f684; fma.rn.ftz.f32 %f691, %f771, %f683, %f690; ld.const.v4.f32 {%f692, %f693, %f694, %f695}, [kernel_params+1216]; mul.ftz.f32 %f699, %f772, %f693; fma.rn.ftz.f32 %f700, %f771, %f692, %f699; ld.const.v4.f32 {%f701, %f702, %f703, %f704}, [kernel_params+1232]; mul.ftz.f32 %f708, %f772, %f702; fma.rn.ftz.f32 %f709, %f771, %f701, %f708; fma.rn.ftz.f32 %f247, %f773, %f703, %f709; fma.rn.ftz.f32 %f772, %f773, %f694, %f700; fma.rn.ftz.f32 %f771, %f773, %f685, %f691; mov.f32 %f773, %f247; $L__BB134_177: mul.ftz.f32 %f251, %f230, %f771; mul.ftz.f32 %f252, %f231, %f772; mul.ftz.f32 %f253, %f232, %f773; ld.local.u32 %r108, [%rd1+52]; and.b32 %r228, %r108, 2; setp.eq.s32 %p125, %r228, 0; @%p125 bra $L__BB134_179; bra.uni $L__BB134_178; $L__BB134_179: or.b32 %r229, %r108, 2; st.local.u32 [%rd1+52], %r229; st.local.v2.f32 [%rd1+280], {%f251, %f252}; st.local.f32 [%rd1+288], %f253; bra.uni $L__BB134_180; $L__BB134_178: ld.local.f32 %f710, [%rd1+280]; add.ftz.f32 %f711, %f251, %f710; ld.local.f32 %f712, [%rd1+284]; add.ftz.f32 %f713, %f252, %f712; ld.local.f32 %f714, [%rd1+288]; add.ftz.f32 %f715, %f253, %f714; st.local.f32 [%rd1+280], %f711; st.local.f32 [%rd1+284], %f713; st.local.f32 [%rd1+288], %f715; $L__BB134_180: add.s32 %r230, %r114, 2; st.param.b32 [func_retval0+0], %r230; ret; $L__BB134_43: ld.global.u64 %rd102, [%rd20+-32]; tex.3d.v4.f32.f32 {%f736, %f365, %f366, %f367}, [%rd102, {%f26, %f27, %f28, %f28}]; mov.f32 %f737, %f736; mov.f32 %f738, %f736; bra.uni $L__BB134_49; $L__BB134_88: ld.global.u64 %rd131, [%rd37+-32]; tex.3d.v4.f32.f32 {%f747, %f482, %f483, %f484}, [%rd131, {%f90, %f91, %f92, %f92}]; mov.f32 %f750, 0f3F800000; mov.f32 %f748, %f747; mov.f32 %f749, %f747; bra.uni $L__BB134_94; $L__BB134_158: setp.eq.s32 %p114, %r107, 3; @%p114 bra $L__BB134_160; bra.uni $L__BB134_159; $L__BB134_160: { // callseq 884, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd60; .param .b32 param1; st.param.f32 [param1+0], %f186; .param .b32 param2; st.param.f32 [param2+0], %f187; .param .b32 param3; st.param.f32 [param3+0], %f188; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f766, [retval0+0]; } // callseq 884 mov.f32 %f767, %f766; mov.f32 %f768, %f766; bra.uni $L__BB134_165; $L__BB134_159: ld.global.u64 %rd183, [%rd61+-32]; tex.3d.v4.f32.f32 {%f766, %f646, %f647, %f648}, [%rd183, {%f186, %f187, %f188, %f188}]; mov.f32 %f767, %f766; mov.f32 %f768, %f766; bra.uni $L__BB134_165; } .func _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<58>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u64 %rd8, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd7, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; cvta.to.local.u64 %rd9, %rd8; add.s64 %rd1, %rd9, 88; mov.f32 %f20, 0f00000000; setp.gt.s32 %p1, %r2, 6; @%p1 bra $L__BB135_13; bra.uni $L__BB135_1; $L__BB135_13: setp.gt.s32 %p2, %r2, 10; @%p2 bra $L__BB135_23; setp.gt.s32 %p8, %r2, 8; @%p8 bra $L__BB135_18; setp.eq.s32 %p11, %r2, 7; @%p11 bra $L__BB135_41; setp.eq.s32 %p12, %r2, 8; @%p12 bra $L__BB135_17; bra.uni $L__BB135_46; $L__BB135_17: ld.local.u32 %r11, [%rd1+-36]; and.b32 %r12, %r11, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB135_46; $L__BB135_1: setp.gt.s32 %p13, %r2, 2; @%p13 bra $L__BB135_6; bra.uni $L__BB135_2; $L__BB135_6: setp.gt.s32 %p14, %r2, 4; @%p14 bra $L__BB135_10; setp.eq.s32 %p17, %r2, 3; @%p17 bra $L__BB135_43; setp.eq.s32 %p18, %r2, 4; @%p18 bra $L__BB135_9; bra.uni $L__BB135_46; $L__BB135_9: and.b32 %r16, %r4, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB135_46; $L__BB135_23: setp.gt.s32 %p3, %r2, 12; @%p3 bra $L__BB135_29; setp.eq.s32 %p6, %r2, 11; @%p6 bra $L__BB135_37; setp.eq.s32 %p7, %r2, 12; @%p7 bra $L__BB135_26; bra.uni $L__BB135_46; $L__BB135_26: and.b32 %r7, %r4, 768; mov.u64 %rd27, kernel_params; add.s64 %rd55, %rd27, 3400; setp.ne.s32 %p24, %r7, 0; @%p24 bra $L__BB135_28; add.s64 %rd55, %rd27, 2488; $L__BB135_28: ld.const.u64 %rd30, [%rd55]; cvta.to.global.u64 %rd31, %rd30; mul.wide.s32 %rd32, %r1, 2; add.s64 %rd33, %rd31, %rd32; ld.global.u16 %rs3, [%rd33]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB135_46; $L__BB135_2: setp.eq.s32 %p19, %r2, 0; @%p19 bra $L__BB135_45; setp.eq.s32 %p20, %r2, 1; @%p20 bra $L__BB135_44; setp.eq.s32 %p21, %r2, 2; @%p21 bra $L__BB135_5; bra.uni $L__BB135_46; $L__BB135_5: and.b32 %r18, %r4, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB135_46; $L__BB135_18: setp.eq.s32 %p9, %r2, 9; @%p9 bra $L__BB135_40; setp.eq.s32 %p10, %r2, 10; @%p10 bra $L__BB135_20; bra.uni $L__BB135_46; $L__BB135_20: and.b32 %r9, %r4, 768; mov.u64 %rd43, kernel_params; add.s64 %rd57, %rd43, 3376; setp.ne.s32 %p26, %r9, 0; @%p26 bra $L__BB135_22; add.s64 %rd57, %rd43, 2464; $L__BB135_22: ld.const.u64 %rd46, [%rd57]; cvta.to.global.u64 %rd47, %rd46; mul.wide.s32 %rd48, %r1, 2; add.s64 %rd49, %rd47, %rd48; ld.global.u16 %rs5, [%rd49]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r10, %r4, 525056; setp.eq.s32 %p27, %r10, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB135_46; $L__BB135_10: setp.eq.s32 %p15, %r2, 5; @%p15 bra $L__BB135_42; setp.eq.s32 %p16, %r2, 6; @%p16 bra $L__BB135_12; bra.uni $L__BB135_46; $L__BB135_12: and.b32 %r14, %r4, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB135_46; $L__BB135_29: setp.eq.s32 %p4, %r2, 13; @%p4 bra $L__BB135_34; setp.ne.s32 %p5, %r2, 14; @%p5 bra $L__BB135_46; and.b32 %r5, %r4, 768; mov.u64 %rd11, kernel_params; add.s64 %rd53, %rd11, 3408; setp.ne.s32 %p22, %r5, 0; @%p22 bra $L__BB135_33; add.s64 %rd53, %rd11, 2496; $L__BB135_33: ld.const.u64 %rd14, [%rd53]; cvta.to.global.u64 %rd15, %rd14; mul.wide.s32 %rd16, %r1, 2; add.s64 %rd17, %rd15, %rd16; ld.global.u16 %rs1, [%rd17]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB135_46; $L__BB135_41: and.b32 %r13, %r4, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB135_46; $L__BB135_43: and.b32 %r17, %r4, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB135_46; $L__BB135_37: and.b32 %r8, %r4, 768; mov.u64 %rd35, kernel_params; add.s64 %rd56, %rd35, 3392; setp.ne.s32 %p25, %r8, 0; @%p25 bra $L__BB135_39; add.s64 %rd56, %rd35, 2480; $L__BB135_39: ld.const.u64 %rd38, [%rd56]; cvta.to.global.u64 %rd39, %rd38; mul.wide.s32 %rd40, %r1, 2; add.s64 %rd41, %rd39, %rd40; ld.global.u16 %rs4, [%rd41]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB135_46; $L__BB135_45: and.b32 %r20, %r4, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB135_46; $L__BB135_44: and.b32 %r19, %r4, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB135_46; $L__BB135_40: ld.local.f32 %f20, [%rd1]; bra.uni $L__BB135_46; $L__BB135_42: and.b32 %r15, %r4, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB135_46; $L__BB135_34: and.b32 %r6, %r4, 768; mov.u64 %rd19, kernel_params; add.s64 %rd54, %rd19, 3384; setp.ne.s32 %p23, %r6, 0; @%p23 bra $L__BB135_36; add.s64 %rd54, %rd19, 2472; $L__BB135_36: ld.const.u64 %rd22, [%rd54]; cvta.to.global.u64 %rd23, %rd22; mul.wide.s32 %rd24, %r1, 2; add.s64 %rd25, %rd23, %rd24; ld.global.u16 %rs2, [%rd25]; cvt.rn.f32.u16 %f20, %rs2; $L__BB135_46: cvta.to.local.u64 %rd50, %rd7; mul.wide.u32 %rd51, %r3, 4; add.s64 %rd52, %rd50, %rd51; st.local.f32 [%rd52], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB136_2; bra.uni $L__BB136_1; $L__BB136_2: mov.b32 %f2764, %r395; bra.uni $L__BB136_3; $L__BB136_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB136_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB136_5; bra.uni $L__BB136_4; $L__BB136_5: mov.b32 %f9, %r396; bra.uni $L__BB136_6; $L__BB136_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB136_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB136_8; bra.uni $L__BB136_7; $L__BB136_8: mov.b32 %f2766, %r397; bra.uni $L__BB136_9; $L__BB136_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB136_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB136_11; bra.uni $L__BB136_10; $L__BB136_11: mov.b32 %f15, %r398; bra.uni $L__BB136_12; $L__BB136_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB136_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB136_14; bra.uni $L__BB136_13; $L__BB136_14: mov.b32 %f2768, %r399; bra.uni $L__BB136_15; $L__BB136_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB136_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB136_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB136_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB136_18; bra.uni $L__BB136_385; $L__BB136_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB136_259; bra.uni $L__BB136_19; $L__BB136_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB136_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB136_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB136_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB136_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB136_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB136_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB136_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB136_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB136_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB136_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB136_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB136_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB136_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB136_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB136_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB136_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB136_362; $L__BB136_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB136_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB136_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB136_40; bra.uni $L__BB136_29; $L__BB136_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB136_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB136_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB136_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB136_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB136_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB136_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB136_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB136_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB136_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB136_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB136_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB136_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB136_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB136_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB136_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB136_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB136_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB136_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB136_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB136_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB136_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB136_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB136_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB136_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB136_136; $L__BB136_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB136_371; bra.uni $L__BB136_366; $L__BB136_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB136_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB136_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB136_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB136_383; $L__BB136_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB136_149; bra.uni $L__BB136_139; $L__BB136_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB136_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB136_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB136_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB136_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB136_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB136_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB136_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB136_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB136_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB136_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB136_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB136_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB136_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB136_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB136_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB136_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB136_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB136_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB136_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB136_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB136_250; $L__BB136_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB136_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB136_253; bra.uni $L__BB136_362; $L__BB136_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB136_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB136_304; @%p171 bra $L__BB136_303; bra.uni $L__BB136_256; $L__BB136_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB136_305; $L__BB136_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB136_305; $L__BB136_256: @%p172 bra $L__BB136_302; bra.uni $L__BB136_257; $L__BB136_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB136_305; $L__BB136_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB136_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB136_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB136_308; bra.uni $L__BB136_306; $L__BB136_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB136_309; $L__BB136_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB136_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB136_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB136_316; @%p171 bra $L__BB136_315; bra.uni $L__BB136_311; $L__BB136_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB136_317; $L__BB136_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB136_317; $L__BB136_311: @%p172 bra $L__BB136_314; bra.uni $L__BB136_312; $L__BB136_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB136_317; $L__BB136_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB136_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB136_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB136_320; bra.uni $L__BB136_318; $L__BB136_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB136_321; $L__BB136_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB136_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB136_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB136_328; @%p171 bra $L__BB136_327; bra.uni $L__BB136_323; $L__BB136_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB136_329; $L__BB136_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB136_329; $L__BB136_323: @%p172 bra $L__BB136_326; bra.uni $L__BB136_324; $L__BB136_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB136_329; $L__BB136_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB136_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB136_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB136_332; bra.uni $L__BB136_330; $L__BB136_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB136_333; $L__BB136_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB136_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB136_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB136_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_362; $L__BB136_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB136_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB136_31; bra.uni $L__BB136_136; $L__BB136_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB136_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB136_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB136_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB136_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB136_107; @%p49 bra $L__BB136_106; bra.uni $L__BB136_37; $L__BB136_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB136_108; $L__BB136_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB136_108; $L__BB136_37: @%p50 bra $L__BB136_105; bra.uni $L__BB136_38; $L__BB136_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB136_108; $L__BB136_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB136_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB136_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB136_111; bra.uni $L__BB136_109; $L__BB136_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB136_112; $L__BB136_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB136_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB136_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB136_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB136_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB136_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB136_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_136; $L__BB136_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB136_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB136_368; bra.uni $L__BB136_383; $L__BB136_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB136_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB136_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB136_383; $L__BB136_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB136_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB136_141; bra.uni $L__BB136_250; $L__BB136_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB136_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB136_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB136_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB136_206; @%p101 bra $L__BB136_205; bra.uni $L__BB136_146; $L__BB136_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB136_207; $L__BB136_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB136_207; $L__BB136_146: @%p102 bra $L__BB136_204; bra.uni $L__BB136_147; $L__BB136_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB136_207; $L__BB136_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB136_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB136_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB136_210; bra.uni $L__BB136_208; $L__BB136_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB136_211; $L__BB136_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB136_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB136_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB136_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB136_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB136_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_250; $L__BB136_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB136_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB136_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB136_295; @%p151 bra $L__BB136_294; bra.uni $L__BB136_290; $L__BB136_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB136_296; $L__BB136_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB136_296; $L__BB136_290: @%p152 bra $L__BB136_293; bra.uni $L__BB136_291; $L__BB136_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB136_296; $L__BB136_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB136_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB136_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB136_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB136_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB136_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB136_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB136_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_362; $L__BB136_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB136_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB136_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB136_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB136_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB136_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB136_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB136_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB136_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB136_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB136_279; bra.uni $L__BB136_362; $L__BB136_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB136_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB136_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB136_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB136_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB136_96; @%p39 bra $L__BB136_95; bra.uni $L__BB136_91; $L__BB136_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB136_97; $L__BB136_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB136_97; $L__BB136_91: @%p40 bra $L__BB136_94; bra.uni $L__BB136_92; $L__BB136_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB136_97; $L__BB136_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB136_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB136_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB136_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB136_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB136_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB136_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB136_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB136_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB136_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_136; $L__BB136_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB136_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB136_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB136_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB136_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB136_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB136_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB136_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB136_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB136_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB136_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB136_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB136_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB136_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB136_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB136_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB136_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB136_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB136_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB136_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB136_76; bra.uni $L__BB136_136; $L__BB136_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB136_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB136_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB136_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB136_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB136_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_383; $L__BB136_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB136_383; $L__BB136_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB136_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB136_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB136_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB136_196; @%p92 bra $L__BB136_195; bra.uni $L__BB136_191; $L__BB136_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB136_197; $L__BB136_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB136_197; $L__BB136_191: @%p93 bra $L__BB136_194; bra.uni $L__BB136_192; $L__BB136_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB136_197; $L__BB136_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB136_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB136_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB136_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB136_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB136_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB136_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB136_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB136_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB136_250; $L__BB136_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB136_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB136_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB136_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB136_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB136_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB136_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB136_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB136_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB136_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB136_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB136_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB136_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB136_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB136_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB136_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB136_179; bra.uni $L__BB136_250; $L__BB136_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB136_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB136_337; @%p188 bra $L__BB136_336; bra.uni $L__BB136_23; $L__BB136_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB136_338; $L__BB136_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB136_338; $L__BB136_23: @%p189 bra $L__BB136_335; bra.uni $L__BB136_24; $L__BB136_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB136_338; $L__BB136_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB136_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB136_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB136_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB136_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB136_347; @%p188 bra $L__BB136_346; bra.uni $L__BB136_342; $L__BB136_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB136_348; $L__BB136_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB136_348; $L__BB136_342: @%p189 bra $L__BB136_345; bra.uni $L__BB136_343; $L__BB136_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB136_348; $L__BB136_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB136_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB136_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB136_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB136_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB136_357; @%p188 bra $L__BB136_356; bra.uni $L__BB136_352; $L__BB136_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB136_358; $L__BB136_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB136_358; $L__BB136_352: @%p189 bra $L__BB136_355; bra.uni $L__BB136_353; $L__BB136_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB136_358; $L__BB136_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB136_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB136_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB136_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB136_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB136_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB136_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB136_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB136_364: mov.f32 %f3392, %f3387; bra.uni $L__BB136_385; $L__BB136_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB136_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB136_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB136_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB136_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB136_128; @%p59 bra $L__BB136_127; bra.uni $L__BB136_123; $L__BB136_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB136_129; $L__BB136_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB136_129; $L__BB136_123: @%p60 bra $L__BB136_126; bra.uni $L__BB136_124; $L__BB136_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB136_129; $L__BB136_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB136_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB136_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB136_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB136_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB136_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB136_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB136_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB136_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB136_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB136_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB136_385; $L__BB136_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB136_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB136_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB136_385; $L__BB136_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB136_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB136_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB136_224; @%p120 bra $L__BB136_223; bra.uni $L__BB136_219; $L__BB136_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB136_225; $L__BB136_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB136_225; $L__BB136_219: @%p121 bra $L__BB136_222; bra.uni $L__BB136_220; $L__BB136_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB136_225; $L__BB136_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB136_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB136_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB136_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB136_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB136_234; @%p120 bra $L__BB136_233; bra.uni $L__BB136_229; $L__BB136_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB136_235; $L__BB136_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB136_235; $L__BB136_229: @%p121 bra $L__BB136_232; bra.uni $L__BB136_230; $L__BB136_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB136_235; $L__BB136_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB136_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB136_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB136_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB136_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB136_244; @%p120 bra $L__BB136_243; bra.uni $L__BB136_239; $L__BB136_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB136_245; $L__BB136_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB136_245; $L__BB136_239: @%p121 bra $L__BB136_242; bra.uni $L__BB136_240; $L__BB136_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB136_245; $L__BB136_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB136_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB136_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB136_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB136_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB136_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB136_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB136_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB136_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB136_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB136_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB136_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB136_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB136_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB136_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB136_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB136_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB136_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB136_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB136_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB138_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB138_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB138_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB138_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) { .reg .pred %p<463>; .reg .b16 %rs<22>; .reg .f32 %f<3060>; .reg .b32 %r<664>; .reg .b64 %rd<808>; ld.param.u64 %rd230, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd231, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r141, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r142, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; ld.param.u32 %r140, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r139, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r138, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd230; cvta.to.local.u64 %rd2, %rd231; and.b32 %r2, %r138, 255; bfe.u32 %r3, %r138, 8, 8; bfe.u32 %r4, %r138, 16, 8; shr.u32 %r5, %r138, 24; setp.eq.s32 %p7, %r5, 255; mov.f32 %f2, 0f3F800000; @%p7 bra $L__BB139_2; mul.wide.u32 %rd232, %r5, 4; add.s64 %rd233, %rd2, %rd232; ld.local.f32 %f2, [%rd233]; $L__BB139_2: ld.const.u64 %rd234, [kernel_params+2400]; cvta.to.global.u64 %rd235, %rd234; mul.wide.s32 %rd236, %r142, 16; add.s64 %rd3, %rd235, %rd236; ld.global.v4.u32 {%r143, %r144, %r145, %r146}, [%rd3]; add.s32 %r663, %r142, 1; setp.eq.ftz.f32 %p8, %f2, 0f00000000; @%p8 bra $L__BB139_516; bra.uni $L__BB139_3; $L__BB139_516: setp.eq.s32 %p461, %r2, 39; add.s32 %r636, %r142, 5; selp.b32 %r663, %r636, %r663, %p461; $L__BB139_517: st.param.b32 [func_retval0+0], %r663; ret; $L__BB139_3: setp.eq.s32 %p9, %r143, 255; add.s64 %rd4, %rd1, 12; @%p9 bra $L__BB139_5; bra.uni $L__BB139_4; $L__BB139_5: ld.local.f32 %f55, [%rd4]; ld.local.f32 %f56, [%rd4+4]; ld.local.f32 %f57, [%rd4+8]; bra.uni $L__BB139_6; $L__BB139_4: mul.wide.u32 %rd237, %r143, 4; add.s64 %rd238, %rd2, %rd237; ld.local.f32 %f55, [%rd238]; ld.local.f32 %f56, [%rd238+4]; ld.local.f32 %f57, [%rd238+8]; $L__BB139_6: ld.local.u32 %r11, [%rd4+52]; and.b32 %r12, %r11, 6; setp.ne.s32 %p10, %r12, 0; @%p10 bra $L__BB139_15; ld.local.v2.f32 {%f550, %f551}, [%rd4+12]; ld.local.v2.f32 {%f552, %f553}, [%rd4+20]; ld.local.v2.f32 {%f554, %f555}, [%rd4+28]; mul.ftz.f32 %f556, %f56, %f554; fma.rn.ftz.f32 %f557, %f55, %f553, %f556; fma.rn.ftz.f32 %f558, %f57, %f555, %f557; add.ftz.f32 %f559, %f558, %f558; mul.ftz.f32 %f560, %f55, %f559; mul.ftz.f32 %f561, %f56, %f559; mul.ftz.f32 %f562, %f57, %f559; sub.ftz.f32 %f563, %f560, %f553; sub.ftz.f32 %f564, %f561, %f554; sub.ftz.f32 %f565, %f562, %f555; mul.ftz.f32 %f566, %f551, %f554; fma.rn.ftz.f32 %f567, %f550, %f553, %f566; fma.rn.ftz.f32 %f18, %f552, %f555, %f567; mul.ftz.f32 %f568, %f18, 0f3F666666; mov.f32 %f569, 0f3C23D70A; min.ftz.f32 %f19, %f568, %f569; mul.ftz.f32 %f570, %f551, %f564; fma.rn.ftz.f32 %f571, %f550, %f563, %f570; fma.rn.ftz.f32 %f572, %f552, %f565, %f571; setp.ltu.ftz.f32 %p11, %f572, %f19; @%p11 bra $L__BB139_8; bra.uni $L__BB139_15; $L__BB139_8: mul.ftz.f32 %f573, %f56, %f551; fma.rn.ftz.f32 %f574, %f55, %f550, %f573; fma.rn.ftz.f32 %f575, %f57, %f552, %f574; mul.ftz.f32 %f576, %f550, %f575; mul.ftz.f32 %f577, %f551, %f575; mul.ftz.f32 %f578, %f552, %f575; sub.ftz.f32 %f579, %f55, %f576; sub.ftz.f32 %f580, %f56, %f577; sub.ftz.f32 %f581, %f57, %f578; mul.ftz.f32 %f582, %f580, %f580; fma.rn.ftz.f32 %f583, %f579, %f579, %f582; fma.rn.ftz.f32 %f584, %f581, %f581, %f583; sqrt.approx.ftz.f32 %f585, %f584; div.approx.ftz.f32 %f23, %f579, %f585; div.approx.ftz.f32 %f24, %f580, %f585; div.approx.ftz.f32 %f25, %f581, %f585; mul.ftz.f32 %f586, %f554, %f24; fma.rn.ftz.f32 %f587, %f553, %f23, %f586; fma.rn.ftz.f32 %f26, %f555, %f25, %f587; mul.ftz.f32 %f588, %f26, %f26; fma.rn.ftz.f32 %f589, %f18, %f18, %f588; mul.ftz.f32 %f590, %f19, %f19; sub.ftz.f32 %f591, %f589, %f590; mul.ftz.f32 %f592, %f588, %f591; mov.f32 %f593, 0f00000000; max.ftz.f32 %f594, %f592, %f593; sqrt.approx.ftz.f32 %f595, %f594; fma.rn.ftz.f32 %f596, %f19, %f18, %f589; mov.f32 %f597, 0f3F000000; div.approx.ftz.f32 %f598, %f597, %f589; add.ftz.f32 %f599, %f595, %f596; mul.ftz.f32 %f27, %f598, %f599; sub.ftz.f32 %f600, %f596, %f595; mul.ftz.f32 %f28, %f598, %f600; setp.gt.ftz.f32 %p12, %f27, 0f3727C5AC; setp.le.ftz.f32 %p13, %f27, 0f3F800054; and.pred %p1, %p12, %p13; setp.gt.ftz.f32 %p14, %f28, 0f3727C5AC; setp.le.ftz.f32 %p15, %f28, 0f3F800054; and.pred %p2, %p14, %p15; and.pred %p16, %p1, %p2; @%p16 bra $L__BB139_11; bra.uni $L__BB139_9; $L__BB139_11: mov.f32 %f607, 0f3F800000; sub.ftz.f32 %f608, %f607, %f27; max.ftz.f32 %f610, %f608, %f593; sqrt.approx.ftz.f32 %f34, %f610; max.ftz.f32 %f611, %f27, %f593; sqrt.approx.ftz.f32 %f35, %f611; sub.ftz.f32 %f612, %f607, %f28; max.ftz.f32 %f613, %f612, %f593; sqrt.approx.ftz.f32 %f36, %f613; max.ftz.f32 %f614, %f28, %f593; sqrt.approx.ftz.f32 %f37, %f614; mul.ftz.f32 %f615, %f18, %f35; fma.rn.ftz.f32 %f616, %f26, %f34, %f615; add.ftz.f32 %f617, %f616, %f616; mul.ftz.f32 %f618, %f35, %f617; sub.ftz.f32 %f38, %f618, %f18; mul.ftz.f32 %f619, %f18, %f37; fma.rn.ftz.f32 %f620, %f26, %f36, %f619; add.ftz.f32 %f621, %f620, %f620; mul.ftz.f32 %f622, %f37, %f621; sub.ftz.f32 %f39, %f622, %f18; setp.ge.ftz.f32 %p18, %f38, 0f3727C5AC; setp.ge.ftz.f32 %p19, %f39, 0f3727C5AC; and.pred %p20, %p18, %p19; @%p20 bra $L__BB139_13; bra.uni $L__BB139_12; $L__BB139_13: setp.lt.ftz.f32 %p22, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p22; selp.f32 %f2968, %f35, %f37, %p22; bra.uni $L__BB139_14; $L__BB139_9: or.pred %p17, %p1, %p2; mov.f32 %f55, %f550; mov.f32 %f56, %f551; mov.f32 %f57, %f552; @%p17 bra $L__BB139_10; bra.uni $L__BB139_15; $L__BB139_10: selp.f32 %f601, %f27, %f28, %p1; mov.f32 %f602, 0f3F800000; sub.ftz.f32 %f603, %f602, %f601; max.ftz.f32 %f605, %f603, %f593; sqrt.approx.ftz.f32 %f2967, %f605; max.ftz.f32 %f606, %f601, %f593; sqrt.approx.ftz.f32 %f2968, %f606; bra.uni $L__BB139_14; $L__BB139_12: setp.gt.ftz.f32 %p21, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p21; selp.f32 %f2968, %f35, %f37, %p21; $L__BB139_14: mul.ftz.f32 %f623, %f550, %f2968; mul.ftz.f32 %f624, %f551, %f2968; mul.ftz.f32 %f625, %f552, %f2968; fma.rn.ftz.f32 %f57, %f25, %f2967, %f625; fma.rn.ftz.f32 %f56, %f24, %f2967, %f624; fma.rn.ftz.f32 %f55, %f23, %f2967, %f623; $L__BB139_15: setp.eq.s32 %p23, %r3, 255; @%p23 bra $L__BB139_17; bra.uni $L__BB139_16; $L__BB139_17: mov.b32 %f60, %r139; bra.uni $L__BB139_18; $L__BB139_16: mul.wide.u32 %rd239, %r3, 4; add.s64 %rd240, %rd2, %rd239; ld.local.f32 %f60, [%rd240]; $L__BB139_18: setp.eq.s32 %p24, %r4, 255; @%p24 bra $L__BB139_20; bra.uni $L__BB139_19; $L__BB139_20: mov.b32 %f63, %r140; bra.uni $L__BB139_21; $L__BB139_19: mul.wide.u32 %rd241, %r4, 4; add.s64 %rd242, %rd2, %rd241; ld.local.f32 %f63, [%rd242]; $L__BB139_21: cvt.u16.u32 %rs8, %r138; and.b16 %rs7, %rs8, 255; mul.wide.u32 %rd243, %r144, 4; add.s64 %rd5, %rd2, %rd243; mul.wide.u32 %rd244, %r145, 4; add.s64 %rd6, %rd2, %rd244; mul.wide.u32 %rd245, %r146, 4; add.s64 %rd7, %rd2, %rd245; setp.gt.s16 %p25, %rs7, 23; @%p25 bra $L__BB139_63; setp.gt.s16 %p36, %rs7, 15; @%p36 bra $L__BB139_48; setp.gt.s16 %p43, %rs7, 8; @%p43 bra $L__BB139_34; bra.uni $L__BB139_24; $L__BB139_34: add.s16 %rs12, %rs7, -9; setp.lt.u16 %p44, %rs12, 2; @%p44 bra $L__BB139_36; add.s16 %rs13, %rs7, -13; setp.lt.u16 %p45, %rs13, 2; @%p45 bra $L__BB139_36; bra.uni $L__BB139_517; $L__BB139_63: setp.gt.s16 %p26, %rs7, 30; @%p26 bra $L__BB139_88; setp.gt.s16 %p30, %rs7, 27; @%p30 bra $L__BB139_77; add.s16 %rs10, %rs7, -25; setp.lt.u16 %p34, %rs10, 2; @%p34 bra $L__BB139_427; setp.eq.s16 %p35, %rs7, 24; @%p35 bra $L__BB139_67; bra.uni $L__BB139_517; $L__BB139_67: ld.const.u32 %r496, [kernel_params+1564]; setp.ne.s32 %p385, %r496, 0; @%p385 bra $L__BB139_69; ld.const.u32 %r497, [kernel_params+1568]; setp.eq.s32 %p386, %r497, 0; shr.u32 %r498, %r141, 3; and.b32 %r499, %r498, 1; setp.eq.b32 %p387, %r499, 1; and.pred %p388, %p387, %p386; @%p388 bra $L__BB139_517; $L__BB139_69: ld.local.f32 %f2726, [%rd4+256]; mov.u64 %rd774, 0; mul.ftz.f32 %f2727, %f2, %f2726; ld.local.f32 %f2728, [%rd4+260]; mul.ftz.f32 %f2729, %f2, %f2728; ld.local.f32 %f2730, [%rd4+264]; mul.ftz.f32 %f2731, %f2, %f2730; mov.f32 %f2732, 0f00000000; max.ftz.f32 %f493, %f2727, %f2732; max.ftz.f32 %f494, %f2729, %f2732; max.ftz.f32 %f495, %f2731, %f2732; add.ftz.f32 %f2733, %f493, %f494; add.ftz.f32 %f2734, %f495, %f2733; mul.ftz.f32 %f2735, %f2734, 0f3EAAAAAB; abs.ftz.f32 %f496, %f2735; setp.ltu.ftz.f32 %p389, %f496, 0f3727C5AC; mov.u64 %rd775, %rd774; @%p389 bra $L__BB139_74; ld.local.u32 %r125, [%rd4+252]; mov.u64 %rd774, 0; setp.eq.s32 %p390, %r125, 0; mov.u64 %rd772, %rd774; mov.u64 %rd773, %rd774; @%p390 bra $L__BB139_72; ld.local.u32 %r500, [%rd4+248]; mul.wide.s32 %rd596, %r500, 80; add.s64 %rd597, %rd1, %rd596; add.s64 %rd772, %rd597, 304; add.s64 %rd598, %rd230, %rd596; add.s64 %rd773, %rd598, 304; mov.b32 %r501, %f495; mov.b32 %r502, %f494; mov.b32 %r503, %f493; mov.u32 %r504, 0; st.local.v4.u32 [%rd597+304], {%r503, %r502, %r501, %r504}; add.s32 %r505, %r500, 1; st.local.u32 [%rd4+248], %r505; add.s32 %r506, %r125, -1; st.local.u32 [%rd4+252], %r506; $L__BB139_72: setp.eq.s64 %p391, %rd773, 0; mov.u64 %rd775, %rd774; @%p391 bra $L__BB139_74; st.local.f32 [%rd772+16], %f496; mov.u64 %rd774, %rd772; mov.u64 %rd775, %rd773; $L__BB139_74: setp.eq.s64 %p392, %rd775, 0; @%p392 bra $L__BB139_517; ld.local.u32 %r126, [%rd4+252]; setp.lt.s32 %p393, %r126, 1; @%p393 bra $L__BB139_457; add.s32 %r507, %r126, -1; st.local.u32 [%rd4+252], %r507; ld.local.s32 %rd601, [%rd4+248]; cvt.s64.s32 %rd602, %r507; add.s64 %rd603, %rd601, %rd602; mul.lo.s64 %rd604, %rd603, 80; add.s64 %rd605, %rd1, %rd604; add.s64 %rd776, %rd605, 304; add.s64 %rd606, %rd230, %rd604; add.s64 %rd777, %rd606, 304; bra.uni $L__BB139_458; $L__BB139_48: setp.gt.s16 %p37, %rs7, 19; @%p37 bra $L__BB139_61; bra.uni $L__BB139_49; $L__BB139_61: add.s16 %rs11, %rs7, -21; setp.lt.u16 %p38, %rs11, 3; @%p38 bra $L__BB139_462; bra.uni $L__BB139_62; $L__BB139_462: ld.const.u32 %r518, [kernel_params+1568]; setp.eq.s32 %p399, %r518, 0; shr.u32 %r519, %r141, 3; and.b32 %r520, %r519, 1; setp.eq.b32 %p400, %r520, 1; and.pred %p401, %p400, %p399; @%p401 bra $L__BB139_517; ld.local.f32 %f2756, [%rd4+256]; mov.u64 %rd780, 0; mul.ftz.f32 %f2757, %f2, %f2756; ld.local.f32 %f2758, [%rd4+260]; mul.ftz.f32 %f2759, %f2, %f2758; ld.local.f32 %f2760, [%rd4+264]; mul.ftz.f32 %f2761, %f2, %f2760; mov.f32 %f2762, 0f00000000; max.ftz.f32 %f500, %f2757, %f2762; max.ftz.f32 %f501, %f2759, %f2762; max.ftz.f32 %f502, %f2761, %f2762; add.ftz.f32 %f2763, %f500, %f501; add.ftz.f32 %f2764, %f502, %f2763; mul.ftz.f32 %f2765, %f2764, 0f3EAAAAAB; abs.ftz.f32 %f503, %f2765; setp.ltu.ftz.f32 %p402, %f503, 0f3727C5AC; mov.u64 %rd781, %rd780; @%p402 bra $L__BB139_468; ld.local.u32 %r127, [%rd4+252]; mov.u64 %rd780, 0; setp.eq.s32 %p403, %r127, 0; mov.u64 %rd778, %rd780; mov.u64 %rd779, %rd780; @%p403 bra $L__BB139_466; ld.local.u32 %r521, [%rd4+248]; mul.wide.s32 %rd616, %r521, 80; add.s64 %rd617, %rd1, %rd616; add.s64 %rd778, %rd617, 304; add.s64 %rd618, %rd230, %rd616; add.s64 %rd779, %rd618, 304; mov.b32 %r522, %f502; mov.b32 %r523, %f501; mov.b32 %r524, %f500; mov.u32 %r525, 0; st.local.v4.u32 [%rd617+304], {%r524, %r523, %r522, %r525}; add.s32 %r526, %r521, 1; st.local.u32 [%rd4+248], %r526; add.s32 %r527, %r127, -1; st.local.u32 [%rd4+252], %r527; $L__BB139_466: setp.eq.s64 %p404, %rd779, 0; mov.u64 %rd781, %rd780; @%p404 bra $L__BB139_468; st.local.f32 [%rd778+16], %f503; mov.u64 %rd780, %rd778; mov.u64 %rd781, %rd779; $L__BB139_468: setp.eq.s64 %p405, %rd781, 0; @%p405 bra $L__BB139_517; add.s64 %rd185, %rd780, 20; mov.u64 %rd621, 0; st.local.f32 [%rd780+20], %f55; st.local.f32 [%rd780+24], %f56; st.local.f32 [%rd780+28], %f57; mov.u32 %r528, 0; st.local.v2.f32 [%rd780+56], {%f2762, %f2762}; st.local.u32 [%rd780+64], %r528; st.local.u64 [%rd780+48], %rd621; mov.f32 %f2767, 0f3727C5AC; max.ftz.f32 %f3057, %f63, %f2767; ld.local.u32 %r529, [%rd4+40]; and.b32 %r530, %r529, 1; setp.eq.b32 %p406, %r530, 1; mov.pred %p407, 0; xor.pred %p408, %p406, %p407; not.pred %p409, %p408; @%p409 bra $L__BB139_471; rcp.approx.ftz.f32 %f3057, %f3057; $L__BB139_471: setp.eq.s32 %p410, %r2, 21; @%p410 bra $L__BB139_475; bra.uni $L__BB139_472; $L__BB139_475: st.local.v2.f32 [%rd185+12], {%f2762, %f2762}; st.local.f32 [%rd185+20], %f3057; mov.u32 %r537, 21; st.local.u32 [%rd185+-8], %r537; ld.local.u32 %r538, [%rd4+40]; or.b32 %r539, %r538, 4; st.local.u32 [%rd4+40], %r539; bra.uni $L__BB139_517; $L__BB139_88: add.s16 %rs9, %rs7, -32; setp.lt.u16 %p27, %rs9, 3; @%p27 bra $L__BB139_290; bra.uni $L__BB139_89; $L__BB139_290: ld.local.v2.u32 {%r330, %r331}, [%rd4+252]; mov.u64 %rd746, 0; mov.b32 %f1892, %r331; mul.ftz.f32 %f340, %f2, %f1892; ld.local.f32 %f1893, [%rd4+260]; mul.ftz.f32 %f341, %f2, %f1893; ld.local.f32 %f1894, [%rd4+264]; mul.ftz.f32 %f342, %f2, %f1894; setp.eq.s32 %p240, %r330, 0; mov.u64 %rd744, %rd746; mov.u64 %rd745, %rd746; @%p240 bra $L__BB139_292; ld.local.u32 %r333, [%rd4+248]; mul.wide.s32 %rd459, %r333, 80; add.s64 %rd460, %rd1, %rd459; add.s64 %rd744, %rd460, 304; add.s64 %rd461, %rd230, %rd459; add.s64 %rd745, %rd461, 304; mov.b32 %r334, %f342; mov.b32 %r335, %f341; mov.b32 %r336, %f340; mov.u32 %r337, 0; st.local.v4.u32 [%rd460+304], {%r336, %r335, %r334, %r337}; add.s32 %r338, %r333, 1; st.local.u32 [%rd4+248], %r338; add.s32 %r339, %r330, -1; st.local.u32 [%rd4+252], %r339; $L__BB139_292: setp.eq.s64 %p241, %rd745, 0; @%p241 bra $L__BB139_294; add.ftz.f32 %f1895, %f340, %f341; add.ftz.f32 %f1896, %f342, %f1895; mul.ftz.f32 %f1897, %f1896, 0f3EAAAAAB; abs.ftz.f32 %f1898, %f1897; st.local.f32 [%rd744+16], %f1898; setp.ge.ftz.f32 %p242, %f1898, 0f3727C5AC; selp.b64 %rd746, %rd745, 0, %p242; $L__BB139_294: setp.eq.s64 %p243, %rd746, 0; @%p243 bra $L__BB139_517; and.b32 %r341, %r141, 4096; setp.eq.s32 %p244, %r341, 0; selp.f32 %f1900, %f60, 0f00000000, %p244; mov.f32 %f3029, 0f00000000; ld.local.f32 %f1901, [%rd6]; mul.ftz.f32 %f1902, %f1900, %f1901; ld.local.f32 %f1903, [%rd6+4]; mul.ftz.f32 %f343, %f1900, %f1903; ld.local.f32 %f1904, [%rd6+8]; mul.ftz.f32 %f344, %f1900, %f1904; add.s64 %rd111, %rd746, 32; st.f32 [%rd746+32], %f1902; st.f32 [%rd746+36], %f343; st.f32 [%rd746+40], %f344; ld.local.f32 %f1905, [%rd4+256]; ld.local.f32 %f1906, [%rd4+260]; ld.local.f32 %f1907, [%rd4+264]; st.f32 [%rd746+44], %f1905; st.f32 [%rd746+48], %f1906; st.f32 [%rd746+52], %f1907; st.f32 [%rd746+20], %f55; st.f32 [%rd746+24], %f56; st.f32 [%rd746+28], %f57; mov.u32 %r342, 2139095039; st.u32 [%rd746+56], %r342; ld.local.f32 %f1908, [%rd7]; max.ftz.f32 %f1909, %f1908, %f3029; mov.f32 %f1910, 0f3F666666; min.ftz.f32 %f1911, %f1909, %f1910; st.f32 [%rd746+60], %f1911; setp.geu.ftz.f32 %p245, %f1902, 0f322BCC77; mov.u32 %r646, 3; mov.f32 %f3027, %f3029; @%p245 bra $L__BB139_297; ld.f32 %f3027, [%rd111+-32]; mov.u32 %r344, 0; st.u32 [%rd111+-32], %r344; st.u32 [%rd111], %r344; mov.u32 %r646, 2; $L__BB139_297: setp.geu.ftz.f32 %p246, %f343, 0f322BCC77; mov.f32 %f3028, %f3029; @%p246 bra $L__BB139_299; ld.f32 %f3028, [%rd111+-28]; mov.u32 %r345, 0; st.u32 [%rd111+-28], %r345; st.u32 [%rd111+4], %r345; add.s32 %r646, %r646, -1; $L__BB139_299: setp.geu.ftz.f32 %p247, %f344, 0f322BCC77; @%p247 bra $L__BB139_301; ld.f32 %f3029, [%rd111+-24]; mov.u32 %r346, 0; st.u32 [%rd111+-24], %r346; st.u32 [%rd111+8], %r346; add.s32 %r646, %r646, -1; $L__BB139_301: setp.gt.s32 %p248, %r646, 2; mov.u32 %r648, 0; @%p248 bra $L__BB139_309; mov.f32 %f1914, 0f00000000; max.ftz.f32 %f351, %f3027, %f1914; max.ftz.f32 %f352, %f3028, %f1914; add.ftz.f32 %f1915, %f351, %f352; max.ftz.f32 %f353, %f3029, %f1914; add.ftz.f32 %f1916, %f353, %f1915; mul.ftz.f32 %f1917, %f1916, 0f3EAAAAAB; abs.ftz.f32 %f354, %f1917; setp.ltu.ftz.f32 %p249, %f354, 0f3727C5AC; mov.u64 %rd749, 0; mov.u64 %rd750, %rd749; @%p249 bra $L__BB139_307; ld.local.u32 %r63, [%rd4+252]; mov.u64 %rd749, 0; setp.eq.s32 %p250, %r63, 0; mov.u64 %rd747, %rd749; mov.u64 %rd748, %rd749; @%p250 bra $L__BB139_305; ld.local.u32 %r348, [%rd4+248]; mul.wide.s32 %rd467, %r348, 80; add.s64 %rd468, %rd1, %rd467; add.s64 %rd747, %rd468, 304; add.s64 %rd469, %rd230, %rd467; add.s64 %rd748, %rd469, 304; mov.b32 %r349, %f353; mov.b32 %r350, %f352; mov.b32 %r351, %f351; mov.u32 %r352, 0; st.local.v4.u32 [%rd468+304], {%r351, %r350, %r349, %r352}; add.s32 %r353, %r348, 1; st.local.u32 [%rd4+248], %r353; add.s32 %r354, %r63, -1; st.local.u32 [%rd4+252], %r354; $L__BB139_305: setp.eq.s64 %p251, %rd748, 0; mov.u64 %rd750, %rd749; @%p251 bra $L__BB139_307; st.local.f32 [%rd747+16], %f354; mov.u64 %rd749, %rd747; mov.u64 %rd750, %rd748; $L__BB139_307: setp.eq.s64 %p252, %rd750, 0; @%p252 bra $L__BB139_309; ld.f32 %f1918, [%rd111+-12]; ld.f32 %f1919, [%rd111+-8]; ld.f32 %f1920, [%rd111+-4]; st.local.f32 [%rd749+20], %f1918; st.local.f32 [%rd749+24], %f1919; st.local.f32 [%rd749+28], %f1920; mov.u32 %r357, 2; st.local.u32 [%rd749+12], %r357; mov.u32 %r648, 12; $L__BB139_309: st.u32 [%rd111+-20], %r2; setp.gt.s32 %p253, %r646, 0; @%p253 bra $L__BB139_311; bra.uni $L__BB139_310; $L__BB139_311: ld.f32 %f1921, [%rd111+-28]; ld.f32 %f1922, [%rd111+-32]; add.ftz.f32 %f1923, %f1922, %f1921; ld.f32 %f1924, [%rd111+-24]; add.ftz.f32 %f1925, %f1924, %f1923; mul.ftz.f32 %f1926, %f1925, 0f3EAAAAAB; abs.ftz.f32 %f1927, %f1926; cvt.rn.f32.s32 %f1928, %r646; mul.ftz.f32 %f1929, %f1927, %f1928; st.f32 [%rd111+-16], %f1929; or.b32 %r359, %r2, 2; setp.eq.s32 %p254, %r359, 34; @%p254 bra $L__BB139_322; bra.uni $L__BB139_312; $L__BB139_322: ld.f32 %f2483, [%rd111]; mul.ftz.f32 %f3033, %f2483, 0f3DA2F983; ld.f32 %f2484, [%rd111+4]; mul.ftz.f32 %f3034, %f2484, 0f3DA2F983; ld.f32 %f2485, [%rd111+8]; mul.ftz.f32 %f3035, %f2485, 0f3DA2F983; bra.uni $L__BB139_323; $L__BB139_24: setp.eq.s16 %p46, %rs7, 2; @%p46 bra $L__BB139_507; setp.eq.s16 %p47, %rs7, 7; @%p47 bra $L__BB139_53; setp.eq.s16 %p48, %rs7, 8; @%p48 bra $L__BB139_27; bra.uni $L__BB139_517; $L__BB139_27: ld.local.f32 %f2922, [%rd4+256]; mov.u64 %rd798, 0; mul.ftz.f32 %f2923, %f2, %f2922; ld.local.f32 %f2924, [%rd4+260]; mul.ftz.f32 %f2925, %f2, %f2924; ld.local.f32 %f2926, [%rd4+264]; mul.ftz.f32 %f2927, %f2, %f2926; mov.f32 %f2928, 0f00000000; max.ftz.f32 %f535, %f2923, %f2928; max.ftz.f32 %f536, %f2925, %f2928; max.ftz.f32 %f537, %f2927, %f2928; add.ftz.f32 %f2929, %f535, %f536; add.ftz.f32 %f2930, %f537, %f2929; mul.ftz.f32 %f2931, %f2930, 0f3EAAAAAB; abs.ftz.f32 %f538, %f2931; setp.ltu.ftz.f32 %p447, %f538, 0f3727C5AC; mov.u64 %rd799, %rd798; @%p447 bra $L__BB139_32; ld.local.u32 %r132, [%rd4+252]; mov.u64 %rd798, 0; setp.eq.s32 %p448, %r132, 0; mov.u64 %rd796, %rd798; mov.u64 %rd797, %rd798; @%p448 bra $L__BB139_30; ld.local.u32 %r600, [%rd4+248]; mul.wide.s32 %rd668, %r600, 80; add.s64 %rd669, %rd1, %rd668; add.s64 %rd796, %rd669, 304; add.s64 %rd670, %rd230, %rd668; add.s64 %rd797, %rd670, 304; mov.b32 %r601, %f537; mov.b32 %r602, %f536; mov.b32 %r603, %f535; mov.u32 %r604, 0; st.local.v4.u32 [%rd669+304], {%r603, %r602, %r601, %r604}; add.s32 %r605, %r600, 1; st.local.u32 [%rd4+248], %r605; add.s32 %r606, %r132, -1; st.local.u32 [%rd4+252], %r606; $L__BB139_30: setp.eq.s64 %p449, %rd797, 0; mov.u64 %rd799, %rd798; @%p449 bra $L__BB139_32; st.local.f32 [%rd796+16], %f538; mov.u64 %rd798, %rd796; mov.u64 %rd799, %rd797; $L__BB139_32: setp.eq.s64 %p450, %rd799, 0; @%p450 bra $L__BB139_517; st.local.f32 [%rd798+20], %f55; st.local.f32 [%rd798+24], %f56; st.local.f32 [%rd798+28], %f57; mov.u32 %r607, 8; st.local.u32 [%rd798+12], %r607; ld.local.u32 %r608, [%rd4+40]; or.b32 %r609, %r608, 12; st.local.u32 [%rd4+40], %r609; bra.uni $L__BB139_517; $L__BB139_77: setp.eq.s16 %p31, %rs7, 28; @%p31 bra $L__BB139_427; setp.eq.s16 %p32, %rs7, 29; @%p32 bra $L__BB139_341; setp.eq.s16 %p33, %rs7, 30; @%p33 bra $L__BB139_80; bra.uni $L__BB139_517; $L__BB139_49: setp.eq.s16 %p40, %rs7, 16; @%p40 bra $L__BB139_36; setp.eq.s16 %p41, %rs7, 17; @%p41 bra $L__BB139_482; setp.eq.s16 %p42, %rs7, 19; @%p42 bra $L__BB139_52; bra.uni $L__BB139_517; $L__BB139_52: ld.const.u32 %r554, [kernel_params+1564]; setp.eq.s32 %p419, %r554, 0; shr.u32 %r555, %r141, 3; and.b32 %r556, %r555, 1; setp.eq.b32 %p420, %r556, 1; and.pred %p421, %p420, %p419; @%p421 bra $L__BB139_517; $L__BB139_53: ld.local.f32 %f2932, [%rd4+256]; mov.u64 %rd802, 0; mul.ftz.f32 %f2933, %f2, %f2932; ld.local.f32 %f2934, [%rd4+260]; mul.ftz.f32 %f2935, %f2, %f2934; ld.local.f32 %f2936, [%rd4+264]; mul.ftz.f32 %f2937, %f2, %f2936; mov.f32 %f2938, 0f00000000; max.ftz.f32 %f539, %f2933, %f2938; max.ftz.f32 %f540, %f2935, %f2938; max.ftz.f32 %f541, %f2937, %f2938; add.ftz.f32 %f2939, %f539, %f540; add.ftz.f32 %f2940, %f541, %f2939; mul.ftz.f32 %f2941, %f2940, 0f3EAAAAAB; abs.ftz.f32 %f542, %f2941; setp.ltu.ftz.f32 %p451, %f542, 0f3727C5AC; mov.u64 %rd803, %rd802; @%p451 bra $L__BB139_58; ld.local.u32 %r133, [%rd4+252]; mov.u64 %rd802, 0; setp.eq.s32 %p452, %r133, 0; mov.u64 %rd800, %rd802; mov.u64 %rd801, %rd802; @%p452 bra $L__BB139_56; ld.local.u32 %r610, [%rd4+248]; mul.wide.s32 %rd677, %r610, 80; add.s64 %rd678, %rd1, %rd677; add.s64 %rd800, %rd678, 304; add.s64 %rd679, %rd230, %rd677; add.s64 %rd801, %rd679, 304; mov.b32 %r611, %f541; mov.b32 %r612, %f540; mov.b32 %r613, %f539; mov.u32 %r614, 0; st.local.v4.u32 [%rd678+304], {%r613, %r612, %r611, %r614}; add.s32 %r615, %r610, 1; st.local.u32 [%rd4+248], %r615; add.s32 %r616, %r133, -1; st.local.u32 [%rd4+252], %r616; $L__BB139_56: setp.eq.s64 %p453, %rd801, 0; mov.u64 %rd803, %rd802; @%p453 bra $L__BB139_58; st.local.f32 [%rd800+16], %f542; mov.u64 %rd802, %rd800; mov.u64 %rd803, %rd801; $L__BB139_58: setp.eq.s64 %p454, %rd803, 0; @%p454 bra $L__BB139_517; add.s64 %rd222, %rd802, 20; st.local.f32 [%rd802+20], %f55; st.local.f32 [%rd802+24], %f56; st.local.f32 [%rd802+28], %f57; st.local.f32 [%rd802+32], %f60; st.local.f32 [%rd802+36], %f63; cvt.ftz.sat.f32.f32 %f543, %f60; cvt.ftz.sat.f32.f32 %f544, %f63; setp.eq.s32 %p455, %r2, 7; @%p455 bra $L__BB139_506; bra.uni $L__BB139_60; $L__BB139_506: mov.u32 %r620, 7; st.local.u32 [%rd222+-8], %r620; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r621, [%rd4+40]; or.b32 %r622, %r621, 12; st.local.u32 [%rd4+40], %r622; bra.uni $L__BB139_517; $L__BB139_36: ld.const.u32 %r567, [kernel_params+1564]; setp.eq.s32 %p426, %r567, 0; shr.u32 %r568, %r141, 3; and.b32 %r569, %r568, 1; setp.eq.b32 %p427, %r569, 1; and.pred %p428, %p427, %p426; @%p428 bra $L__BB139_517; ld.local.f32 %f2831, [%rd4+256]; mov.u64 %rd792, 0; mul.ftz.f32 %f2832, %f2, %f2831; ld.local.f32 %f2833, [%rd4+260]; mul.ftz.f32 %f2834, %f2, %f2833; ld.local.f32 %f2835, [%rd4+264]; mul.ftz.f32 %f2836, %f2, %f2835; mov.f32 %f2837, 0f00000000; max.ftz.f32 %f519, %f2832, %f2837; max.ftz.f32 %f520, %f2834, %f2837; max.ftz.f32 %f521, %f2836, %f2837; add.ftz.f32 %f2838, %f519, %f520; add.ftz.f32 %f2839, %f521, %f2838; mul.ftz.f32 %f2840, %f2839, 0f3EAAAAAB; abs.ftz.f32 %f522, %f2840; setp.ltu.ftz.f32 %p429, %f522, 0f3727C5AC; mov.u64 %rd793, %rd792; @%p429 bra $L__BB139_42; ld.local.u32 %r130, [%rd4+252]; mov.u64 %rd792, 0; setp.eq.s32 %p430, %r130, 0; mov.u64 %rd790, %rd792; mov.u64 %rd791, %rd792; @%p430 bra $L__BB139_40; ld.local.u32 %r570, [%rd4+248]; mul.wide.s32 %rd645, %r570, 80; add.s64 %rd646, %rd1, %rd645; add.s64 %rd790, %rd646, 304; add.s64 %rd647, %rd230, %rd645; add.s64 %rd791, %rd647, 304; mov.b32 %r571, %f521; mov.b32 %r572, %f520; mov.b32 %r573, %f519; mov.u32 %r574, 0; st.local.v4.u32 [%rd646+304], {%r573, %r572, %r571, %r574}; add.s32 %r575, %r570, 1; st.local.u32 [%rd4+248], %r575; add.s32 %r576, %r130, -1; st.local.u32 [%rd4+252], %r576; $L__BB139_40: setp.eq.s64 %p431, %rd791, 0; mov.u64 %rd793, %rd792; @%p431 bra $L__BB139_42; st.local.f32 [%rd790+16], %f522; mov.u64 %rd792, %rd790; mov.u64 %rd793, %rd791; $L__BB139_42: setp.eq.s64 %p432, %rd793, 0; @%p432 bra $L__BB139_517; mul.ftz.f32 %f3058, %f60, %f60; add.s64 %rd205, %rd792, 20; mov.u64 %rd794, 0; st.local.f32 [%rd792+20], %f55; st.local.f32 [%rd792+24], %f56; st.local.f32 [%rd792+28], %f57; mov.u32 %r577, 1065353216; st.local.u32 [%rd792+40], %r577; st.local.u64 [%rd792+48], %rd794; setp.eq.s32 %p433, %r144, 255; @%p433 bra $L__BB139_490; bra.uni $L__BB139_44; $L__BB139_490: mov.u32 %r578, 0; st.local.v2.f32 [%rd205+36], {%f2837, %f2837}; st.local.u32 [%rd205+44], %r578; st.local.v2.f32 [%rd205+12], {%f3058, %f3058}; mov.f32 %f3059, %f3058; bra.uni $L__BB139_491; $L__BB139_427: ld.const.u32 %r121, [kernel_params+1564]; setp.ne.s32 %p358, %r121, 0; @%p358 bra $L__BB139_429; ld.const.u32 %r454, [kernel_params+1568]; setp.eq.s32 %p359, %r454, 0; shr.u32 %r455, %r141, 3; and.b32 %r456, %r455, 1; setp.eq.b32 %p360, %r456, 1; and.pred %p361, %p360, %p359; @%p361 bra $L__BB139_517; $L__BB139_429: ld.local.f32 %f2682, [%rd4+256]; mul.ftz.f32 %f470, %f2, %f2682; ld.local.f32 %f2683, [%rd4+260]; mul.ftz.f32 %f471, %f2, %f2683; ld.local.f32 %f2684, [%rd4+264]; mul.ftz.f32 %f472, %f2, %f2684; mov.f32 %f2685, 0f3727C5AC; max.ftz.f32 %f3054, %f63, %f2685; ld.local.u32 %r457, [%rd4+40]; and.b32 %r458, %r457, 1; setp.eq.b32 %p362, %r458, 1; mov.pred %p363, 0; xor.pred %p364, %p362, %p363; not.pred %p365, %p364; @%p365 bra $L__BB139_431; rcp.approx.ftz.f32 %f3054, %f3054; $L__BB139_431: ld.local.f32 %f2687, [%rd4+24]; ld.local.f32 %f2688, [%rd4+28]; mul.ftz.f32 %f2689, %f56, %f2688; fma.rn.ftz.f32 %f2690, %f55, %f2687, %f2689; ld.local.f32 %f2691, [%rd4+32]; fma.rn.ftz.f32 %f2692, %f57, %f2691, %f2690; abs.ftz.f32 %f476, %f2692; fma.rn.ftz.f32 %f2693, %f3054, %f3054, 0fBF800000; mov.f32 %f3055, 0f3F800000; fma.rn.ftz.f32 %f477, %f476, %f476, %f2693; setp.leu.ftz.f32 %p366, %f477, 0f00000000; @%p366 bra $L__BB139_433; sqrt.approx.ftz.f32 %f2694, %f477; sub.ftz.f32 %f2695, %f2694, %f476; add.ftz.f32 %f2696, %f476, %f2694; div.approx.ftz.f32 %f2697, %f2695, %f2696; fma.rn.ftz.f32 %f2698, %f476, %f2696, 0fBF800000; fma.rn.ftz.f32 %f2699, %f476, %f2695, 0f3F800000; div.approx.ftz.f32 %f2700, %f2698, %f2699; mul.ftz.f32 %f2701, %f2697, 0f3F000000; mul.ftz.f32 %f2702, %f2697, %f2701; fma.rn.ftz.f32 %f2703, %f2700, %f2700, 0f3F800000; mul.ftz.f32 %f3055, %f2702, %f2703; $L__BB139_433: mul.ftz.f32 %f480, %f60, %f60; and.b32 %r122, %r141, 8; shr.u32 %r459, %r141, 3; and.b32 %r460, %r459, 1; setp.eq.b32 %p367, %r460, 1; setp.eq.s32 %p368, %r121, 0; and.pred %p369, %p367, %p368; @%p369 bra $L__BB139_445; mul.ftz.f32 %f2704, %f470, %f3055; mov.f32 %f2705, 0f00000000; max.ftz.f32 %f481, %f2704, %f2705; mul.ftz.f32 %f2706, %f471, %f3055; max.ftz.f32 %f482, %f2706, %f2705; mul.ftz.f32 %f2707, %f472, %f3055; max.ftz.f32 %f483, %f2707, %f2705; add.ftz.f32 %f2708, %f481, %f482; add.ftz.f32 %f2709, %f483, %f2708; mul.ftz.f32 %f2710, %f2709, 0f3EAAAAAB; abs.ftz.f32 %f484, %f2710; setp.ltu.ftz.f32 %p370, %f484, 0f3727C5AC; mov.u64 %rd766, 0; mov.u64 %rd767, %rd766; @%p370 bra $L__BB139_439; ld.local.u32 %r123, [%rd4+252]; mov.u64 %rd766, 0; setp.eq.s32 %p371, %r123, 0; mov.u64 %rd764, %rd766; mov.u64 %rd765, %rd766; @%p371 bra $L__BB139_437; ld.local.u32 %r461, [%rd4+248]; mul.wide.s32 %rd574, %r461, 80; add.s64 %rd575, %rd1, %rd574; add.s64 %rd764, %rd575, 304; add.s64 %rd576, %rd230, %rd574; add.s64 %rd765, %rd576, 304; mov.b32 %r462, %f483; mov.b32 %r463, %f482; mov.b32 %r464, %f481; mov.u32 %r465, 0; st.local.v4.u32 [%rd575+304], {%r464, %r463, %r462, %r465}; add.s32 %r466, %r461, 1; st.local.u32 [%rd4+248], %r466; add.s32 %r467, %r123, -1; st.local.u32 [%rd4+252], %r467; $L__BB139_437: setp.eq.s64 %p372, %rd765, 0; mov.u64 %rd767, %rd766; @%p372 bra $L__BB139_439; st.local.f32 [%rd764+16], %f484; mov.u64 %rd766, %rd764; mov.u64 %rd767, %rd765; $L__BB139_439: setp.eq.s64 %p373, %rd767, 0; @%p373 bra $L__BB139_445; add.s64 %rd160, %rd766, 20; mov.u64 %rd579, 0; st.local.f32 [%rd766+20], %f55; st.local.f32 [%rd766+24], %f56; st.local.f32 [%rd766+28], %f57; mov.u32 %r468, 0; st.local.v2.f32 [%rd766+56], {%f2705, %f2705}; st.local.u32 [%rd766+64], %r468; st.local.u64 [%rd766+48], %rd579; setp.eq.s32 %p374, %r2, 28; @%p374 bra $L__BB139_444; bra.uni $L__BB139_441; $L__BB139_444: st.local.v2.f32 [%rd160+12], {%f2705, %f2705}; st.local.f32 [%rd160+20], %f3054; mov.u32 %r475, 9; st.local.u32 [%rd160+-8], %r475; ld.local.u32 %r476, [%rd4+40]; or.b32 %r477, %r476, 4; st.local.u32 [%rd4+40], %r477; bra.uni $L__BB139_445; $L__BB139_89: setp.eq.s16 %p28, %rs7, 31; @%p28 bra $L__BB139_325; setp.ne.s16 %p29, %rs7, 39; @%p29 bra $L__BB139_517; ld.global.v4.u32 {%r147, %r148, %r149, %r150}, [%rd3+16]; ld.local.f32 %f2977, [%rd5]; ld.local.f32 %f2978, [%rd5+4]; ld.local.f32 %f2979, [%rd5+8]; shr.u32 %r152, %r147, 24; shl.b32 %r153, %r145, 2; cvt.u64.u32 %rd246, %r153; and.b64 %rd247, %rd246, 1020; add.s64 %rd248, %rd2, %rd247; ld.local.f32 %f70, [%rd248]; shr.u32 %r154, %r145, 6; cvt.u64.u32 %rd249, %r154; and.b64 %rd250, %rd249, 1020; add.s64 %rd251, %rd2, %rd250; ld.local.f32 %f71, [%rd251]; shr.u32 %r155, %r145, 14; cvt.u64.u32 %rd252, %r155; and.b64 %rd253, %rd252, 1020; add.s64 %rd254, %rd2, %rd253; ld.local.f32 %f72, [%rd254]; shr.u32 %r156, %r145, 24; mul.wide.u32 %rd255, %r156, 4; add.s64 %rd256, %rd2, %rd255; ld.local.f32 %f73, [%rd256]; shl.b32 %r157, %r146, 2; cvt.u64.u32 %rd257, %r157; and.b64 %rd258, %rd257, 1020; add.s64 %rd259, %rd2, %rd258; ld.local.f32 %f74, [%rd259]; shr.u32 %r158, %r146, 6; cvt.u64.u32 %rd260, %r158; and.b64 %rd261, %rd260, 1020; add.s64 %rd262, %rd2, %rd261; ld.local.f32 %f75, [%rd262]; shr.u32 %r159, %r146, 14; cvt.u64.u32 %rd263, %r159; and.b64 %rd264, %rd263, 1020; add.s64 %rd265, %rd2, %rd264; ld.local.f32 %f76, [%rd265]; shr.u32 %r160, %r146, 24; mul.wide.u32 %rd266, %r160, 4; add.s64 %rd267, %rd2, %rd266; ld.local.f32 %f77, [%rd267]; shr.u32 %r161, %r147, 6; cvt.u64.u32 %rd268, %r161; and.b64 %rd269, %rd268, 1020; add.s64 %rd270, %rd2, %rd269; ld.local.f32 %f78, [%rd270]; shr.u32 %r162, %r147, 14; cvt.u64.u32 %rd271, %r162; and.b64 %rd272, %rd271, 1020; add.s64 %rd273, %rd2, %rd272; mul.wide.u32 %rd274, %r152, 4; add.s64 %rd275, %rd2, %rd274; ld.local.f32 %f79, [%rd275]; shl.b32 %r163, %r147, 2; cvt.u64.u32 %rd276, %r163; and.b64 %rd277, %rd276, 1020; add.s64 %rd278, %rd2, %rd277; ld.local.f32 %f626, [%rd278]; mov.f32 %f627, 0f3727C5AC; max.ftz.f32 %f2980, %f626, %f627; ld.local.f32 %f81, [%rd273]; setp.eq.ftz.f32 %p49, %f81, 0f00000000; @%p49 bra $L__BB139_93; mul.ftz.f32 %f628, %f81, 0f40C90FDB; cos.approx.ftz.f32 %f629, %f628; sin.approx.ftz.f32 %f630, %f628; mov.f32 %f631, 0f3F800000; sub.ftz.f32 %f632, %f631, %f629; mul.ftz.f32 %f633, %f55, %f632; fma.rn.ftz.f32 %f634, %f55, %f633, %f629; mul.ftz.f32 %f635, %f56, %f633; mul.ftz.f32 %f636, %f57, %f630; sub.ftz.f32 %f637, %f635, %f636; mul.ftz.f32 %f638, %f2978, %f637; fma.rn.ftz.f32 %f639, %f2977, %f634, %f638; mul.ftz.f32 %f640, %f57, %f633; mul.ftz.f32 %f641, %f56, %f630; add.ftz.f32 %f642, %f641, %f640; add.ftz.f32 %f643, %f636, %f635; mul.ftz.f32 %f644, %f56, %f632; fma.rn.ftz.f32 %f645, %f56, %f644, %f629; mul.ftz.f32 %f646, %f2978, %f645; fma.rn.ftz.f32 %f647, %f2977, %f643, %f646; mul.ftz.f32 %f648, %f57, %f644; mul.ftz.f32 %f649, %f55, %f630; sub.ftz.f32 %f650, %f648, %f649; sub.ftz.f32 %f651, %f640, %f641; add.ftz.f32 %f652, %f649, %f648; mul.ftz.f32 %f653, %f2978, %f652; fma.rn.ftz.f32 %f654, %f2977, %f651, %f653; mul.ftz.f32 %f655, %f57, %f632; fma.rn.ftz.f32 %f656, %f57, %f655, %f629; fma.rn.ftz.f32 %f2978, %f2979, %f650, %f647; fma.rn.ftz.f32 %f84, %f2979, %f656, %f654; fma.rn.ftz.f32 %f2977, %f2979, %f642, %f639; mov.f32 %f2979, %f84; $L__BB139_93: ld.local.u32 %r164, [%rd4+40]; and.b32 %r165, %r164, 1; setp.eq.b32 %p50, %r165, 1; mov.pred %p51, 0; xor.pred %p52, %p50, %p51; not.pred %p53, %p52; @%p53 bra $L__BB139_95; rcp.approx.ftz.f32 %f2980, %f2980; $L__BB139_95: ld.local.f32 %f90, [%rd4+24]; ld.local.f32 %f91, [%rd4+28]; mul.ftz.f32 %f658, %f56, %f91; fma.rn.ftz.f32 %f659, %f55, %f90, %f658; ld.local.f32 %f92, [%rd4+32]; fma.rn.ftz.f32 %f660, %f57, %f92, %f659; abs.ftz.f32 %f93, %f660; fma.rn.ftz.f32 %f661, %f2980, %f2980, 0fBF800000; mov.f32 %f657, 0f3F800000; fma.rn.ftz.f32 %f94, %f93, %f93, %f661; setp.leu.ftz.f32 %p54, %f94, 0f00000000; mov.f32 %f2981, %f657; @%p54 bra $L__BB139_97; sqrt.approx.ftz.f32 %f662, %f94; sub.ftz.f32 %f663, %f662, %f93; add.ftz.f32 %f664, %f93, %f662; div.approx.ftz.f32 %f665, %f663, %f664; fma.rn.ftz.f32 %f666, %f93, %f664, 0fBF800000; fma.rn.ftz.f32 %f667, %f93, %f663, 0f3F800000; div.approx.ftz.f32 %f668, %f666, %f667; mul.ftz.f32 %f669, %f665, 0f3F000000; mul.ftz.f32 %f670, %f665, %f669; fma.rn.ftz.f32 %f671, %f668, %f668, 0f3F800000; mul.ftz.f32 %f2981, %f670, %f671; $L__BB139_97: cvt.ftz.sat.f32.f32 %f672, %f60; sub.ftz.f32 %f674, %f657, %f672; cvt.ftz.sat.f32.f32 %f675, %f78; sub.ftz.f32 %f676, %f657, %f675; mul.ftz.f32 %f97, %f674, %f676; mul.ftz.f32 %f98, %f675, %f674; sub.ftz.f32 %f99, %f657, %f98; ld.global.v4.u32 {%r166, %r167, %r168, %r169}, [%rd3+32]; setp.eq.s32 %p55, %r166, 255; @%p55 bra $L__BB139_99; bra.uni $L__BB139_98; $L__BB139_99: mov.b32 %f2982, %r167; mov.b32 %f2983, %r168; mov.b32 %f2984, %r169; bra.uni $L__BB139_100; $L__BB139_62: setp.eq.s16 %p39, %rs7, 20; @%p39 bra $L__BB139_80; bra.uni $L__BB139_517; $L__BB139_80: ld.local.f32 %f2770, [%rd4+256]; mov.u64 %rd784, 0; mul.ftz.f32 %f2771, %f2, %f2770; ld.local.f32 %f2772, [%rd4+260]; mul.ftz.f32 %f2773, %f2, %f2772; ld.local.f32 %f2774, [%rd4+264]; mul.ftz.f32 %f2775, %f2, %f2774; mov.f32 %f2776, 0f00000000; max.ftz.f32 %f509, %f2771, %f2776; max.ftz.f32 %f510, %f2773, %f2776; max.ftz.f32 %f511, %f2775, %f2776; add.ftz.f32 %f2777, %f509, %f510; add.ftz.f32 %f2778, %f511, %f2777; mul.ftz.f32 %f2779, %f2778, 0f3EAAAAAB; abs.ftz.f32 %f512, %f2779; setp.ltu.ftz.f32 %p412, %f512, 0f3727C5AC; mov.u64 %rd785, %rd784; @%p412 bra $L__BB139_85; ld.local.u32 %r128, [%rd4+252]; mov.u64 %rd784, 0; setp.eq.s32 %p413, %r128, 0; mov.u64 %rd782, %rd784; mov.u64 %rd783, %rd784; @%p413 bra $L__BB139_83; ld.local.u32 %r540, [%rd4+248]; mul.wide.s32 %rd627, %r540, 80; add.s64 %rd628, %rd1, %rd627; add.s64 %rd782, %rd628, 304; add.s64 %rd629, %rd230, %rd627; add.s64 %rd783, %rd629, 304; mov.b32 %r541, %f511; mov.b32 %r542, %f510; mov.b32 %r543, %f509; mov.u32 %r544, 0; st.local.v4.u32 [%rd628+304], {%r543, %r542, %r541, %r544}; add.s32 %r545, %r540, 1; st.local.u32 [%rd4+248], %r545; add.s32 %r546, %r128, -1; st.local.u32 [%rd4+252], %r546; $L__BB139_83: setp.eq.s64 %p414, %rd783, 0; mov.u64 %rd785, %rd784; @%p414 bra $L__BB139_85; st.local.f32 [%rd782+16], %f512; mov.u64 %rd784, %rd782; mov.u64 %rd785, %rd783; $L__BB139_85: setp.eq.s64 %p415, %rd785, 0; @%p415 bra $L__BB139_517; add.s64 %rd192, %rd784, 20; st.local.f32 [%rd784+20], %f55; st.local.f32 [%rd784+24], %f56; st.local.f32 [%rd784+28], %f57; st.local.f32 [%rd784+44], %f60; st.local.f32 [%rd784+48], %f63; ld.local.f32 %f2780, [%rd6]; neg.ftz.f32 %f2781, %f2780; st.local.f32 [%rd784+52], %f2781; setp.eq.s32 %p416, %r144, 255; @%p416 bra $L__BB139_476; bra.uni $L__BB139_87; $L__BB139_476: ld.local.u8 %rs18, [%rd4+52]; and.b16 %rs19, %rs18, 6; setp.eq.s16 %p417, %rs19, 0; @%p417 bra $L__BB139_478; ld.local.f32 %f2792, [%rd4+104]; ld.local.f32 %f2793, [%rd4+108]; mul.ftz.f32 %f2794, %f2793, %f2793; fma.rn.ftz.f32 %f2795, %f2792, %f2792, %f2794; ld.local.f32 %f2796, [%rd4+112]; fma.rn.ftz.f32 %f2797, %f2796, %f2796, %f2795; sqrt.approx.ftz.f32 %f2798, %f2797; div.approx.ftz.f32 %f2799, %f2792, %f2798; div.approx.ftz.f32 %f2800, %f2793, %f2798; div.approx.ftz.f32 %f2801, %f2796, %f2798; st.local.f32 [%rd192+12], %f2799; st.local.f32 [%rd192+16], %f2800; st.local.f32 [%rd192+20], %f2801; bra.uni $L__BB139_479; $L__BB139_507: ld.local.f32 %f2942, [%rd4+256]; mov.u64 %rd806, 0; mul.ftz.f32 %f2943, %f2, %f2942; ld.local.f32 %f2944, [%rd4+260]; mul.ftz.f32 %f2945, %f2, %f2944; ld.local.f32 %f2946, [%rd4+264]; mul.ftz.f32 %f2947, %f2, %f2946; mov.f32 %f2948, 0f00000000; max.ftz.f32 %f545, %f2943, %f2948; max.ftz.f32 %f546, %f2945, %f2948; max.ftz.f32 %f547, %f2947, %f2948; add.ftz.f32 %f2949, %f545, %f546; add.ftz.f32 %f2950, %f547, %f2949; mul.ftz.f32 %f2951, %f2950, 0f3EAAAAAB; abs.ftz.f32 %f548, %f2951; setp.ltu.ftz.f32 %p456, %f548, 0f3727C5AC; mov.u64 %rd807, %rd806; @%p456 bra $L__BB139_512; ld.local.u32 %r134, [%rd4+252]; mov.u64 %rd806, 0; setp.eq.s32 %p457, %r134, 0; mov.u64 %rd804, %rd806; mov.u64 %rd805, %rd806; @%p457 bra $L__BB139_510; ld.local.u32 %r623, [%rd4+248]; mul.wide.s32 %rd686, %r623, 80; add.s64 %rd687, %rd1, %rd686; add.s64 %rd804, %rd687, 304; add.s64 %rd688, %rd230, %rd686; add.s64 %rd805, %rd688, 304; mov.b32 %r624, %f547; mov.b32 %r625, %f546; mov.b32 %r626, %f545; mov.u32 %r627, 0; st.local.v4.u32 [%rd687+304], {%r626, %r625, %r624, %r627}; add.s32 %r628, %r623, 1; st.local.u32 [%rd4+248], %r628; add.s32 %r629, %r134, -1; st.local.u32 [%rd4+252], %r629; $L__BB139_510: setp.eq.s64 %p458, %rd805, 0; mov.u64 %rd807, %rd806; @%p458 bra $L__BB139_512; st.local.f32 [%rd804+16], %f548; mov.u64 %rd806, %rd804; mov.u64 %rd807, %rd805; $L__BB139_512: setp.eq.s64 %p459, %rd807, 0; @%p459 bra $L__BB139_517; add.s64 %rd229, %rd806, 20; st.local.f32 [%rd806+20], %f55; st.local.f32 [%rd806+24], %f56; st.local.f32 [%rd806+28], %f57; setp.eq.ftz.f32 %p460, %f60, 0f00000000; @%p460 bra $L__BB139_515; bra.uni $L__BB139_514; $L__BB139_515: mov.u32 %r633, 2; st.local.u32 [%rd229+-8], %r633; ld.local.u32 %r634, [%rd4+40]; or.b32 %r635, %r634, 12; st.local.u32 [%rd4+40], %r635; bra.uni $L__BB139_517; $L__BB139_341: ld.global.v4.u32 {%r375, %r376, %r377, %r378}, [%rd3+16]; ld.global.v4.u32 {%r379, %r380, %r381, %r382}, [%rd3+32]; ld.global.u32 %rd121, [%rd3+52]; ld.local.f32 %f378, [%rd4+256]; ld.local.f32 %f379, [%rd4+260]; ld.local.f32 %f380, [%rd4+264]; bfe.u32 %r82, %r144, 8, 8; and.b32 %r83, %r144, 255; setp.eq.s32 %p304, %r83, 255; @%p304 bra $L__BB139_343; bra.uni $L__BB139_342; $L__BB139_343: mov.b32 %f3036, %r145; bra.uni $L__BB139_344; $L__BB139_482: ld.local.f32 %f2816, [%rd4+256]; mov.u64 %rd788, 0; mul.ftz.f32 %f2817, %f2, %f2816; ld.local.f32 %f2818, [%rd4+260]; mul.ftz.f32 %f2819, %f2, %f2818; ld.local.f32 %f2820, [%rd4+264]; mul.ftz.f32 %f2821, %f2, %f2820; mov.f32 %f2822, 0f00000000; max.ftz.f32 %f515, %f2817, %f2822; max.ftz.f32 %f516, %f2819, %f2822; max.ftz.f32 %f517, %f2821, %f2822; add.ftz.f32 %f2823, %f515, %f516; add.ftz.f32 %f2824, %f517, %f2823; mul.ftz.f32 %f2825, %f2824, 0f3EAAAAAB; abs.ftz.f32 %f518, %f2825; setp.ltu.ftz.f32 %p422, %f518, 0f3727C5AC; mov.u64 %rd789, %rd788; @%p422 bra $L__BB139_487; ld.local.u32 %r129, [%rd4+252]; mov.u64 %rd788, 0; setp.eq.s32 %p423, %r129, 0; mov.u64 %rd786, %rd788; mov.u64 %rd787, %rd788; @%p423 bra $L__BB139_485; ld.local.u32 %r557, [%rd4+248]; mul.wide.s32 %rd636, %r557, 80; add.s64 %rd637, %rd1, %rd636; add.s64 %rd786, %rd637, 304; add.s64 %rd638, %rd230, %rd636; add.s64 %rd787, %rd638, 304; mov.b32 %r558, %f517; mov.b32 %r559, %f516; mov.b32 %r560, %f515; mov.u32 %r561, 0; st.local.v4.u32 [%rd637+304], {%r560, %r559, %r558, %r561}; add.s32 %r562, %r557, 1; st.local.u32 [%rd4+248], %r562; add.s32 %r563, %r129, -1; st.local.u32 [%rd4+252], %r563; $L__BB139_485: setp.eq.s64 %p424, %rd787, 0; mov.u64 %rd789, %rd788; @%p424 bra $L__BB139_487; st.local.f32 [%rd786+16], %f518; mov.u64 %rd788, %rd786; mov.u64 %rd789, %rd787; $L__BB139_487: setp.eq.s64 %p425, %rd789, 0; @%p425 bra $L__BB139_517; st.local.f32 [%rd788+20], %f55; st.local.f32 [%rd788+24], %f56; st.local.f32 [%rd788+28], %f57; cvt.ftz.sat.f32.f32 %f2826, %f60; st.local.f32 [%rd788+32], %f2826; mov.f32 %f2827, 0f3C23D70A; max.ftz.f32 %f2828, %f2826, %f2827; mul.ftz.f32 %f2829, %f2828, %f2828; rcp.approx.ftz.f32 %f2830, %f2829; st.local.f32 [%rd788+36], %f2830; mov.u32 %r564, 17; st.local.u32 [%rd788+12], %r564; ld.local.u32 %r565, [%rd4+40]; or.b32 %r566, %r565, 12; st.local.u32 [%rd4+40], %r566; bra.uni $L__BB139_517; $L__BB139_325: ld.local.f32 %f2486, [%rd4+256]; mul.ftz.f32 %f374, %f2, %f2486; ld.local.f32 %f2487, [%rd4+260]; mul.ftz.f32 %f375, %f2, %f2487; ld.local.f32 %f2488, [%rd4+264]; mul.ftz.f32 %f376, %f2, %f2488; add.ftz.f32 %f2489, %f374, %f375; add.ftz.f32 %f2490, %f376, %f2489; mul.ftz.f32 %f2491, %f2490, 0f3EAAAAAB; abs.ftz.f32 %f377, %f2491; setp.ltu.ftz.f32 %p294, %f377, 0f3727C5AC; @%p294 bra $L__BB139_517; ld.local.u32 %r67, [%rd4+40]; and.b32 %r362, %r67, 512; setp.eq.s32 %p295, %r362, 0; @%p295 bra $L__BB139_332; bra.uni $L__BB139_327; $L__BB139_332: or.b32 %r365, %r67, 516; st.local.u32 [%rd4+40], %r365; st.local.f32 [%rd4+280], %f374; st.local.f32 [%rd4+284], %f375; st.local.f32 [%rd4+288], %f376; and.b32 %r71, %r141, 491520; setp.eq.s32 %p299, %r71, 0; @%p299 bra $L__BB139_334; mov.u32 %r651, 1; st.local.u32 [%rd4+252], %r651; bra.uni $L__BB139_335; $L__BB139_44: ld.local.f32 %f524, [%rd5]; ld.local.f32 %f525, [%rd5+8]; ld.local.f32 %f526, [%rd5+4]; st.local.v2.f32 [%rd205+36], {%f524, %f526}; st.local.f32 [%rd205+44], %f525; ld.local.f32 %f527, [%rd6]; setp.eq.ftz.f32 %p434, %f527, 0f00000000; @%p434 bra $L__BB139_46; mul.ftz.f32 %f2841, %f527, 0f40C90FDB; cos.approx.ftz.f32 %f2842, %f2841; sin.approx.ftz.f32 %f2843, %f2841; mov.f32 %f2844, 0f3F800000; sub.ftz.f32 %f2845, %f2844, %f2842; mul.ftz.f32 %f2846, %f55, %f2845; fma.rn.ftz.f32 %f2847, %f55, %f2846, %f2842; mul.ftz.f32 %f2848, %f56, %f2846; mul.ftz.f32 %f2849, %f57, %f2843; sub.ftz.f32 %f2850, %f2848, %f2849; mul.ftz.f32 %f2851, %f526, %f2850; fma.rn.ftz.f32 %f2852, %f524, %f2847, %f2851; mul.ftz.f32 %f2853, %f57, %f2846; mul.ftz.f32 %f2854, %f56, %f2843; add.ftz.f32 %f2855, %f2854, %f2853; add.ftz.f32 %f2856, %f2849, %f2848; mul.ftz.f32 %f2857, %f56, %f2845; fma.rn.ftz.f32 %f2858, %f56, %f2857, %f2842; mul.ftz.f32 %f2859, %f526, %f2858; fma.rn.ftz.f32 %f2860, %f524, %f2856, %f2859; mul.ftz.f32 %f2861, %f57, %f2857; mul.ftz.f32 %f2862, %f55, %f2843; sub.ftz.f32 %f2863, %f2861, %f2862; sub.ftz.f32 %f2864, %f2853, %f2854; add.ftz.f32 %f2865, %f2862, %f2861; mul.ftz.f32 %f2866, %f526, %f2865; fma.rn.ftz.f32 %f2867, %f524, %f2864, %f2866; mul.ftz.f32 %f2868, %f57, %f2845; fma.rn.ftz.f32 %f2869, %f57, %f2868, %f2842; fma.rn.ftz.f32 %f2870, %f525, %f2869, %f2867; fma.rn.ftz.f32 %f2871, %f525, %f2863, %f2860; fma.rn.ftz.f32 %f2872, %f525, %f2855, %f2852; st.local.v2.f32 [%rd205+36], {%f2872, %f2871}; st.local.f32 [%rd205+44], %f2870; $L__BB139_46: mov.f32 %f2873, 0fBF7D70A4; max.ftz.f32 %f2874, %f63, %f2873; mov.f32 %f2875, 0f3F7D70A4; min.ftz.f32 %f528, %f2874, %f2875; setp.lt.ftz.f32 %p435, %f528, 0f00000000; @%p435 bra $L__BB139_489; bra.uni $L__BB139_47; $L__BB139_489: add.ftz.f32 %f2878, %f528, 0f3F800000; div.approx.ftz.f32 %f3059, %f3058, %f2878; mul.ftz.f32 %f3058, %f3058, %f2878; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; bra.uni $L__BB139_491; $L__BB139_310: mov.u32 %r358, 0; st.u32 [%rd111+-16], %r358; bra.uni $L__BB139_324; $L__BB139_87: ld.local.f32 %f2782, [%rd5]; ld.local.f32 %f2783, [%rd5+4]; mul.ftz.f32 %f2784, %f2783, %f2783; fma.rn.ftz.f32 %f2785, %f2782, %f2782, %f2784; ld.local.f32 %f2786, [%rd5+8]; fma.rn.ftz.f32 %f2787, %f2786, %f2786, %f2785; sqrt.approx.ftz.f32 %f2788, %f2787; div.approx.ftz.f32 %f2789, %f2782, %f2788; div.approx.ftz.f32 %f2790, %f2783, %f2788; div.approx.ftz.f32 %f2791, %f2786, %f2788; st.local.f32 [%rd192+12], %f2789; st.local.f32 [%rd192+16], %f2790; st.local.f32 [%rd192+20], %f2791; bra.uni $L__BB139_479; $L__BB139_312: mov.f32 %f1931, 0f3F8147AE; max.ftz.f32 %f1932, %f63, %f1931; mov.f32 %f1933, 0f40733333; min.ftz.f32 %f1934, %f1932, %f1933; rcp.approx.ftz.f32 %f1935, %f1934; mov.f32 %f1936, 0f3F800000; fma.rn.ftz.f32 %f1937, %f1935, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f1938, %f1935, %f1937, 0f3F2B020C; fma.rn.ftz.f32 %f1939, %f1934, 0f3D8240B8, %f1938; add.ftz.f32 %f1940, %f1939, 0f3F800000; mul.ftz.f32 %f1941, %f1940, 0f3FAAAAAB; sub.ftz.f32 %f1942, %f1936, %f1939; div.approx.ftz.f32 %f355, %f1941, %f1942; ld.f32 %f356, [%rd111+12]; setp.lt.ftz.f32 %p255, %f356, 0f38D1B717; mov.f32 %f3032, 0f00000000; mov.f32 %f3030, %f3032; @%p255 bra $L__BB139_315; setp.ge.ftz.f32 %p256, %f356, 0f3F7EB852; mov.f32 %f3030, 0f3F7FFFEF; @%p256 bra $L__BB139_315; sub.ftz.f32 %f1945, %f1936, 0f3F000000; mul.ftz.f32 %f1946, %f1945, 0f40400000; sqrt.approx.ftz.f32 %f1947, %f1946; mul.ftz.f32 %f1948, %f355, %f1947; mul.ftz.f32 %f1949, %f1948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1950, %f1949; add.ftz.f32 %f1951, %f1950, 0f3F800000; mul.ftz.f32 %f1952, %f1951, 0f3E800000; mul.ftz.f32 %f1953, %f1947, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1954, %f1953; mul.ftz.f32 %f1955, %f1954, %f1952; setp.lt.ftz.f32 %p257, %f1955, %f356; selp.f32 %f1956, 0f3F800000, 0f3F000000, %p257; selp.f32 %f1957, 0f3F000000, 0f00000000, %p257; selp.f32 %f1958, 0f3F400000, 0f3E800000, %p257; sub.ftz.f32 %f1959, %f1936, %f1958; mul.ftz.f32 %f1960, %f1959, 0f40400000; sqrt.approx.ftz.f32 %f1961, %f1960; mul.ftz.f32 %f1962, %f1958, 0f3F000000; mul.ftz.f32 %f1963, %f355, %f1961; mul.ftz.f32 %f1964, %f1963, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1965, 0f3F800000; mul.ftz.f32 %f1967, %f1966, %f1962; mul.ftz.f32 %f1968, %f1961, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1969, %f1968; mul.ftz.f32 %f1970, %f1969, %f1967; setp.lt.ftz.f32 %p258, %f1970, %f356; selp.f32 %f1971, %f1956, %f1958, %p258; selp.f32 %f1972, %f1958, %f1957, %p258; add.ftz.f32 %f1973, %f1971, %f1972; mul.ftz.f32 %f1974, %f1973, 0f3F000000; sub.ftz.f32 %f1975, %f1936, %f1974; mul.ftz.f32 %f1976, %f1975, 0f40400000; sqrt.approx.ftz.f32 %f1977, %f1976; mul.ftz.f32 %f1978, %f1974, 0f3F000000; mul.ftz.f32 %f1979, %f355, %f1977; mul.ftz.f32 %f1980, %f1979, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1981, %f1980; add.ftz.f32 %f1982, %f1981, 0f3F800000; mul.ftz.f32 %f1983, %f1982, %f1978; mul.ftz.f32 %f1984, %f1977, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f1985, %f1983; setp.lt.ftz.f32 %p259, %f1986, %f356; selp.f32 %f1987, %f1971, %f1974, %p259; selp.f32 %f1988, %f1974, %f1972, %p259; add.ftz.f32 %f1989, %f1987, %f1988; mul.ftz.f32 %f1990, %f1989, 0f3F000000; sub.ftz.f32 %f1991, %f1936, %f1990; mul.ftz.f32 %f1992, %f1991, 0f40400000; sqrt.approx.ftz.f32 %f1993, %f1992; mul.ftz.f32 %f1994, %f1990, 0f3F000000; mul.ftz.f32 %f1995, %f355, %f1993; mul.ftz.f32 %f1996, %f1995, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1997, %f1996; add.ftz.f32 %f1998, %f1997, 0f3F800000; mul.ftz.f32 %f1999, %f1998, %f1994; mul.ftz.f32 %f2000, %f1993, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2001, %f2000; mul.ftz.f32 %f2002, %f2001, %f1999; setp.lt.ftz.f32 %p260, %f2002, %f356; selp.f32 %f2003, %f1987, %f1990, %p260; selp.f32 %f2004, %f1990, %f1988, %p260; add.ftz.f32 %f2005, %f2003, %f2004; mul.ftz.f32 %f2006, %f2005, 0f3F000000; sub.ftz.f32 %f2007, %f1936, %f2006; mul.ftz.f32 %f2008, %f2007, 0f40400000; sqrt.approx.ftz.f32 %f2009, %f2008; mul.ftz.f32 %f2010, %f2006, 0f3F000000; mul.ftz.f32 %f2011, %f355, %f2009; mul.ftz.f32 %f2012, %f2011, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2013, %f2012; add.ftz.f32 %f2014, %f2013, 0f3F800000; mul.ftz.f32 %f2015, %f2014, %f2010; mul.ftz.f32 %f2016, %f2009, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2017, %f2016; mul.ftz.f32 %f2018, %f2017, %f2015; setp.lt.ftz.f32 %p261, %f2018, %f356; selp.f32 %f2019, %f2003, %f2006, %p261; selp.f32 %f2020, %f2006, %f2004, %p261; add.ftz.f32 %f2021, %f2019, %f2020; mul.ftz.f32 %f2022, %f2021, 0f3F000000; sub.ftz.f32 %f2023, %f1936, %f2022; mul.ftz.f32 %f2024, %f2023, 0f40400000; sqrt.approx.ftz.f32 %f2025, %f2024; mul.ftz.f32 %f2026, %f2022, 0f3F000000; mul.ftz.f32 %f2027, %f355, %f2025; mul.ftz.f32 %f2028, %f2027, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2029, %f2028; add.ftz.f32 %f2030, %f2029, 0f3F800000; mul.ftz.f32 %f2031, %f2030, %f2026; mul.ftz.f32 %f2032, %f2025, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2033, %f2032; mul.ftz.f32 %f2034, %f2033, %f2031; setp.lt.ftz.f32 %p262, %f2034, %f356; selp.f32 %f2035, %f2019, %f2022, %p262; selp.f32 %f2036, %f2022, %f2020, %p262; add.ftz.f32 %f2037, %f2035, %f2036; mul.ftz.f32 %f2038, %f2037, 0f3F000000; sub.ftz.f32 %f2039, %f1936, %f2038; mul.ftz.f32 %f2040, %f2039, 0f40400000; sqrt.approx.ftz.f32 %f2041, %f2040; mul.ftz.f32 %f2042, %f2038, 0f3F000000; mul.ftz.f32 %f2043, %f355, %f2041; mul.ftz.f32 %f2044, %f2043, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2045, 0f3F800000; mul.ftz.f32 %f2047, %f2046, %f2042; mul.ftz.f32 %f2048, %f2041, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2049, %f2048; mul.ftz.f32 %f2050, %f2049, %f2047; setp.lt.ftz.f32 %p263, %f2050, %f356; selp.f32 %f2051, %f2035, %f2038, %p263; selp.f32 %f2052, %f2038, %f2036, %p263; add.ftz.f32 %f2053, %f2051, %f2052; mul.ftz.f32 %f2054, %f2053, 0f3F000000; sub.ftz.f32 %f2055, %f1936, %f2054; mul.ftz.f32 %f2056, %f2055, 0f40400000; sqrt.approx.ftz.f32 %f2057, %f2056; mul.ftz.f32 %f2058, %f2054, 0f3F000000; mul.ftz.f32 %f2059, %f355, %f2057; mul.ftz.f32 %f2060, %f2059, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2061, %f2060; add.ftz.f32 %f2062, %f2061, 0f3F800000; mul.ftz.f32 %f2063, %f2062, %f2058; mul.ftz.f32 %f2064, %f2057, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2065, %f2064; mul.ftz.f32 %f2066, %f2065, %f2063; setp.lt.ftz.f32 %p264, %f2066, %f356; selp.f32 %f2067, %f2051, %f2054, %p264; selp.f32 %f2068, %f2054, %f2052, %p264; add.ftz.f32 %f2069, %f2067, %f2068; mul.ftz.f32 %f2070, %f2069, 0f3F000000; sub.ftz.f32 %f2071, %f1936, %f2070; mul.ftz.f32 %f2072, %f2071, 0f40400000; sqrt.approx.ftz.f32 %f2073, %f2072; mul.ftz.f32 %f2074, %f2070, 0f3F000000; mul.ftz.f32 %f2075, %f355, %f2073; mul.ftz.f32 %f2076, %f2075, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2077, %f2076; add.ftz.f32 %f2078, %f2077, 0f3F800000; mul.ftz.f32 %f2079, %f2078, %f2074; mul.ftz.f32 %f2080, %f2073, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2081, %f2080; mul.ftz.f32 %f2082, %f2081, %f2079; setp.lt.ftz.f32 %p265, %f2082, %f356; selp.f32 %f2083, %f2067, %f2070, %p265; selp.f32 %f2084, %f2070, %f2068, %p265; add.ftz.f32 %f2085, %f2083, %f2084; mul.ftz.f32 %f2086, %f2085, 0f3F000000; sub.ftz.f32 %f2087, %f1936, %f2086; mul.ftz.f32 %f2088, %f2087, 0f40400000; sqrt.approx.ftz.f32 %f2089, %f2088; mul.ftz.f32 %f2090, %f2086, 0f3F000000; mul.ftz.f32 %f2091, %f355, %f2089; mul.ftz.f32 %f2092, %f2091, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2093, %f2092; add.ftz.f32 %f2094, %f2093, 0f3F800000; mul.ftz.f32 %f2095, %f2094, %f2090; mul.ftz.f32 %f2096, %f2089, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2097, %f2096; mul.ftz.f32 %f2098, %f2097, %f2095; setp.lt.ftz.f32 %p266, %f2098, %f356; selp.f32 %f2099, %f2083, %f2086, %p266; selp.f32 %f2100, %f2086, %f2084, %p266; add.ftz.f32 %f2101, %f2099, %f2100; mul.ftz.f32 %f2102, %f2101, 0f3F000000; sub.ftz.f32 %f2103, %f1936, %f2102; mul.ftz.f32 %f2104, %f2103, 0f40400000; sqrt.approx.ftz.f32 %f2105, %f2104; mul.ftz.f32 %f2106, %f2102, 0f3F000000; mul.ftz.f32 %f2107, %f355, %f2105; mul.ftz.f32 %f2108, %f2107, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2109, %f2108; add.ftz.f32 %f2110, %f2109, 0f3F800000; mul.ftz.f32 %f2111, %f2110, %f2106; mul.ftz.f32 %f2112, %f2105, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2113, %f2112; mul.ftz.f32 %f2114, %f2113, %f2111; setp.lt.ftz.f32 %p267, %f2114, %f356; selp.f32 %f2115, %f2099, %f2102, %p267; selp.f32 %f2116, %f2102, %f2100, %p267; add.ftz.f32 %f2117, %f2115, %f2116; mul.ftz.f32 %f3030, %f2117, 0f3F000000; $L__BB139_315: ld.f32 %f359, [%rd111+16]; setp.lt.ftz.f32 %p268, %f359, 0f38D1B717; mov.f32 %f3031, %f3032; @%p268 bra $L__BB139_318; setp.ge.ftz.f32 %p269, %f359, 0f3F7EB852; mov.f32 %f3031, 0f3F7FFFEF; @%p269 bra $L__BB139_318; sub.ftz.f32 %f2121, %f1936, 0f3F000000; mul.ftz.f32 %f2122, %f2121, 0f40400000; sqrt.approx.ftz.f32 %f2123, %f2122; mul.ftz.f32 %f2124, %f355, %f2123; mul.ftz.f32 %f2125, %f2124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2126, %f2125; add.ftz.f32 %f2127, %f2126, 0f3F800000; mul.ftz.f32 %f2128, %f2127, 0f3E800000; mul.ftz.f32 %f2129, %f2123, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2130, %f2129; mul.ftz.f32 %f2131, %f2130, %f2128; setp.lt.ftz.f32 %p270, %f2131, %f359; selp.f32 %f2132, 0f3F800000, 0f3F000000, %p270; selp.f32 %f2133, 0f3F000000, 0f00000000, %p270; selp.f32 %f2134, 0f3F400000, 0f3E800000, %p270; sub.ftz.f32 %f2135, %f1936, %f2134; mul.ftz.f32 %f2136, %f2135, 0f40400000; sqrt.approx.ftz.f32 %f2137, %f2136; mul.ftz.f32 %f2138, %f2134, 0f3F000000; mul.ftz.f32 %f2139, %f355, %f2137; mul.ftz.f32 %f2140, %f2139, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2141, %f2140; add.ftz.f32 %f2142, %f2141, 0f3F800000; mul.ftz.f32 %f2143, %f2142, %f2138; mul.ftz.f32 %f2144, %f2137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2145, %f2144; mul.ftz.f32 %f2146, %f2145, %f2143; setp.lt.ftz.f32 %p271, %f2146, %f359; selp.f32 %f2147, %f2132, %f2134, %p271; selp.f32 %f2148, %f2134, %f2133, %p271; add.ftz.f32 %f2149, %f2147, %f2148; mul.ftz.f32 %f2150, %f2149, 0f3F000000; sub.ftz.f32 %f2151, %f1936, %f2150; mul.ftz.f32 %f2152, %f2151, 0f40400000; sqrt.approx.ftz.f32 %f2153, %f2152; mul.ftz.f32 %f2154, %f2150, 0f3F000000; mul.ftz.f32 %f2155, %f355, %f2153; mul.ftz.f32 %f2156, %f2155, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2157, %f2156; add.ftz.f32 %f2158, %f2157, 0f3F800000; mul.ftz.f32 %f2159, %f2158, %f2154; mul.ftz.f32 %f2160, %f2153, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2161, %f2160; mul.ftz.f32 %f2162, %f2161, %f2159; setp.lt.ftz.f32 %p272, %f2162, %f359; selp.f32 %f2163, %f2147, %f2150, %p272; selp.f32 %f2164, %f2150, %f2148, %p272; add.ftz.f32 %f2165, %f2163, %f2164; mul.ftz.f32 %f2166, %f2165, 0f3F000000; sub.ftz.f32 %f2167, %f1936, %f2166; mul.ftz.f32 %f2168, %f2167, 0f40400000; sqrt.approx.ftz.f32 %f2169, %f2168; mul.ftz.f32 %f2170, %f2166, 0f3F000000; mul.ftz.f32 %f2171, %f355, %f2169; mul.ftz.f32 %f2172, %f2171, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2173, %f2172; add.ftz.f32 %f2174, %f2173, 0f3F800000; mul.ftz.f32 %f2175, %f2174, %f2170; mul.ftz.f32 %f2176, %f2169, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2177, %f2176; mul.ftz.f32 %f2178, %f2177, %f2175; setp.lt.ftz.f32 %p273, %f2178, %f359; selp.f32 %f2179, %f2163, %f2166, %p273; selp.f32 %f2180, %f2166, %f2164, %p273; add.ftz.f32 %f2181, %f2179, %f2180; mul.ftz.f32 %f2182, %f2181, 0f3F000000; sub.ftz.f32 %f2183, %f1936, %f2182; mul.ftz.f32 %f2184, %f2183, 0f40400000; sqrt.approx.ftz.f32 %f2185, %f2184; mul.ftz.f32 %f2186, %f2182, 0f3F000000; mul.ftz.f32 %f2187, %f355, %f2185; mul.ftz.f32 %f2188, %f2187, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2189, %f2188; add.ftz.f32 %f2190, %f2189, 0f3F800000; mul.ftz.f32 %f2191, %f2190, %f2186; mul.ftz.f32 %f2192, %f2185, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2193, %f2192; mul.ftz.f32 %f2194, %f2193, %f2191; setp.lt.ftz.f32 %p274, %f2194, %f359; selp.f32 %f2195, %f2179, %f2182, %p274; selp.f32 %f2196, %f2182, %f2180, %p274; add.ftz.f32 %f2197, %f2195, %f2196; mul.ftz.f32 %f2198, %f2197, 0f3F000000; sub.ftz.f32 %f2199, %f1936, %f2198; mul.ftz.f32 %f2200, %f2199, 0f40400000; sqrt.approx.ftz.f32 %f2201, %f2200; mul.ftz.f32 %f2202, %f2198, 0f3F000000; mul.ftz.f32 %f2203, %f355, %f2201; mul.ftz.f32 %f2204, %f2203, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2205, %f2204; add.ftz.f32 %f2206, %f2205, 0f3F800000; mul.ftz.f32 %f2207, %f2206, %f2202; mul.ftz.f32 %f2208, %f2201, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2209, %f2208; mul.ftz.f32 %f2210, %f2209, %f2207; setp.lt.ftz.f32 %p275, %f2210, %f359; selp.f32 %f2211, %f2195, %f2198, %p275; selp.f32 %f2212, %f2198, %f2196, %p275; add.ftz.f32 %f2213, %f2211, %f2212; mul.ftz.f32 %f2214, %f2213, 0f3F000000; sub.ftz.f32 %f2215, %f1936, %f2214; mul.ftz.f32 %f2216, %f2215, 0f40400000; sqrt.approx.ftz.f32 %f2217, %f2216; mul.ftz.f32 %f2218, %f2214, 0f3F000000; mul.ftz.f32 %f2219, %f355, %f2217; mul.ftz.f32 %f2220, %f2219, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2221, %f2220; add.ftz.f32 %f2222, %f2221, 0f3F800000; mul.ftz.f32 %f2223, %f2222, %f2218; mul.ftz.f32 %f2224, %f2217, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2225, %f2224; mul.ftz.f32 %f2226, %f2225, %f2223; setp.lt.ftz.f32 %p276, %f2226, %f359; selp.f32 %f2227, %f2211, %f2214, %p276; selp.f32 %f2228, %f2214, %f2212, %p276; add.ftz.f32 %f2229, %f2227, %f2228; mul.ftz.f32 %f2230, %f2229, 0f3F000000; sub.ftz.f32 %f2231, %f1936, %f2230; mul.ftz.f32 %f2232, %f2231, 0f40400000; sqrt.approx.ftz.f32 %f2233, %f2232; mul.ftz.f32 %f2234, %f2230, 0f3F000000; mul.ftz.f32 %f2235, %f355, %f2233; mul.ftz.f32 %f2236, %f2235, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2237, %f2236; add.ftz.f32 %f2238, %f2237, 0f3F800000; mul.ftz.f32 %f2239, %f2238, %f2234; mul.ftz.f32 %f2240, %f2233, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2241, %f2240; mul.ftz.f32 %f2242, %f2241, %f2239; setp.lt.ftz.f32 %p277, %f2242, %f359; selp.f32 %f2243, %f2227, %f2230, %p277; selp.f32 %f2244, %f2230, %f2228, %p277; add.ftz.f32 %f2245, %f2243, %f2244; mul.ftz.f32 %f2246, %f2245, 0f3F000000; sub.ftz.f32 %f2247, %f1936, %f2246; mul.ftz.f32 %f2248, %f2247, 0f40400000; sqrt.approx.ftz.f32 %f2249, %f2248; mul.ftz.f32 %f2250, %f2246, 0f3F000000; mul.ftz.f32 %f2251, %f355, %f2249; mul.ftz.f32 %f2252, %f2251, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2253, %f2252; add.ftz.f32 %f2254, %f2253, 0f3F800000; mul.ftz.f32 %f2255, %f2254, %f2250; mul.ftz.f32 %f2256, %f2249, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2257, %f2256; mul.ftz.f32 %f2258, %f2257, %f2255; setp.lt.ftz.f32 %p278, %f2258, %f359; selp.f32 %f2259, %f2243, %f2246, %p278; selp.f32 %f2260, %f2246, %f2244, %p278; add.ftz.f32 %f2261, %f2259, %f2260; mul.ftz.f32 %f2262, %f2261, 0f3F000000; sub.ftz.f32 %f2263, %f1936, %f2262; mul.ftz.f32 %f2264, %f2263, 0f40400000; sqrt.approx.ftz.f32 %f2265, %f2264; mul.ftz.f32 %f2266, %f2262, 0f3F000000; mul.ftz.f32 %f2267, %f355, %f2265; mul.ftz.f32 %f2268, %f2267, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2269, %f2268; add.ftz.f32 %f2270, %f2269, 0f3F800000; mul.ftz.f32 %f2271, %f2270, %f2266; mul.ftz.f32 %f2272, %f2265, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2273, %f2272; mul.ftz.f32 %f2274, %f2273, %f2271; setp.lt.ftz.f32 %p279, %f2274, %f359; selp.f32 %f2275, %f2259, %f2262, %p279; selp.f32 %f2276, %f2262, %f2260, %p279; add.ftz.f32 %f2277, %f2275, %f2276; mul.ftz.f32 %f2278, %f2277, 0f3F000000; sub.ftz.f32 %f2279, %f1936, %f2278; mul.ftz.f32 %f2280, %f2279, 0f40400000; sqrt.approx.ftz.f32 %f2281, %f2280; mul.ftz.f32 %f2282, %f2278, 0f3F000000; mul.ftz.f32 %f2283, %f355, %f2281; mul.ftz.f32 %f2284, %f2283, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2285, %f2284; add.ftz.f32 %f2286, %f2285, 0f3F800000; mul.ftz.f32 %f2287, %f2286, %f2282; mul.ftz.f32 %f2288, %f2281, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f2289, %f2287; setp.lt.ftz.f32 %p280, %f2290, %f359; selp.f32 %f2291, %f2275, %f2278, %p280; selp.f32 %f2292, %f2278, %f2276, %p280; add.ftz.f32 %f2293, %f2291, %f2292; mul.ftz.f32 %f3031, %f2293, 0f3F000000; $L__BB139_318: ld.f32 %f362, [%rd111+20]; setp.lt.ftz.f32 %p281, %f362, 0f38D1B717; @%p281 bra $L__BB139_321; setp.ge.ftz.f32 %p282, %f362, 0f3F7EB852; mov.f32 %f3032, 0f3F7FFFEF; @%p282 bra $L__BB139_321; sub.ftz.f32 %f2297, %f1936, 0f3F000000; mul.ftz.f32 %f2298, %f2297, 0f40400000; sqrt.approx.ftz.f32 %f2299, %f2298; mul.ftz.f32 %f2300, %f355, %f2299; mul.ftz.f32 %f2301, %f2300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2302, %f2301; add.ftz.f32 %f2303, %f2302, 0f3F800000; mul.ftz.f32 %f2304, %f2303, 0f3E800000; mul.ftz.f32 %f2305, %f2299, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2306, %f2305; mul.ftz.f32 %f2307, %f2306, %f2304; setp.lt.ftz.f32 %p283, %f2307, %f362; selp.f32 %f2308, 0f3F800000, 0f3F000000, %p283; selp.f32 %f2309, 0f3F000000, 0f00000000, %p283; selp.f32 %f2310, 0f3F400000, 0f3E800000, %p283; sub.ftz.f32 %f2311, %f1936, %f2310; mul.ftz.f32 %f2312, %f2311, 0f40400000; sqrt.approx.ftz.f32 %f2313, %f2312; mul.ftz.f32 %f2314, %f2310, 0f3F000000; mul.ftz.f32 %f2315, %f355, %f2313; mul.ftz.f32 %f2316, %f2315, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2317, %f2316; add.ftz.f32 %f2318, %f2317, 0f3F800000; mul.ftz.f32 %f2319, %f2318, %f2314; mul.ftz.f32 %f2320, %f2313, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2321, %f2320; mul.ftz.f32 %f2322, %f2321, %f2319; setp.lt.ftz.f32 %p284, %f2322, %f362; selp.f32 %f2323, %f2308, %f2310, %p284; selp.f32 %f2324, %f2310, %f2309, %p284; add.ftz.f32 %f2325, %f2323, %f2324; mul.ftz.f32 %f2326, %f2325, 0f3F000000; sub.ftz.f32 %f2327, %f1936, %f2326; mul.ftz.f32 %f2328, %f2327, 0f40400000; sqrt.approx.ftz.f32 %f2329, %f2328; mul.ftz.f32 %f2330, %f2326, 0f3F000000; mul.ftz.f32 %f2331, %f355, %f2329; mul.ftz.f32 %f2332, %f2331, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2333, %f2332; add.ftz.f32 %f2334, %f2333, 0f3F800000; mul.ftz.f32 %f2335, %f2334, %f2330; mul.ftz.f32 %f2336, %f2329, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2337, %f2336; mul.ftz.f32 %f2338, %f2337, %f2335; setp.lt.ftz.f32 %p285, %f2338, %f362; selp.f32 %f2339, %f2323, %f2326, %p285; selp.f32 %f2340, %f2326, %f2324, %p285; add.ftz.f32 %f2341, %f2339, %f2340; mul.ftz.f32 %f2342, %f2341, 0f3F000000; sub.ftz.f32 %f2343, %f1936, %f2342; mul.ftz.f32 %f2344, %f2343, 0f40400000; sqrt.approx.ftz.f32 %f2345, %f2344; mul.ftz.f32 %f2346, %f2342, 0f3F000000; mul.ftz.f32 %f2347, %f355, %f2345; mul.ftz.f32 %f2348, %f2347, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2349, %f2348; add.ftz.f32 %f2350, %f2349, 0f3F800000; mul.ftz.f32 %f2351, %f2350, %f2346; mul.ftz.f32 %f2352, %f2345, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2353, %f2352; mul.ftz.f32 %f2354, %f2353, %f2351; setp.lt.ftz.f32 %p286, %f2354, %f362; selp.f32 %f2355, %f2339, %f2342, %p286; selp.f32 %f2356, %f2342, %f2340, %p286; add.ftz.f32 %f2357, %f2355, %f2356; mul.ftz.f32 %f2358, %f2357, 0f3F000000; sub.ftz.f32 %f2359, %f1936, %f2358; mul.ftz.f32 %f2360, %f2359, 0f40400000; sqrt.approx.ftz.f32 %f2361, %f2360; mul.ftz.f32 %f2362, %f2358, 0f3F000000; mul.ftz.f32 %f2363, %f355, %f2361; mul.ftz.f32 %f2364, %f2363, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2365, %f2364; add.ftz.f32 %f2366, %f2365, 0f3F800000; mul.ftz.f32 %f2367, %f2366, %f2362; mul.ftz.f32 %f2368, %f2361, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2369, %f2368; mul.ftz.f32 %f2370, %f2369, %f2367; setp.lt.ftz.f32 %p287, %f2370, %f362; selp.f32 %f2371, %f2355, %f2358, %p287; selp.f32 %f2372, %f2358, %f2356, %p287; add.ftz.f32 %f2373, %f2371, %f2372; mul.ftz.f32 %f2374, %f2373, 0f3F000000; sub.ftz.f32 %f2375, %f1936, %f2374; mul.ftz.f32 %f2376, %f2375, 0f40400000; sqrt.approx.ftz.f32 %f2377, %f2376; mul.ftz.f32 %f2378, %f2374, 0f3F000000; mul.ftz.f32 %f2379, %f355, %f2377; mul.ftz.f32 %f2380, %f2379, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2381, %f2380; add.ftz.f32 %f2382, %f2381, 0f3F800000; mul.ftz.f32 %f2383, %f2382, %f2378; mul.ftz.f32 %f2384, %f2377, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2385, %f2384; mul.ftz.f32 %f2386, %f2385, %f2383; setp.lt.ftz.f32 %p288, %f2386, %f362; selp.f32 %f2387, %f2371, %f2374, %p288; selp.f32 %f2388, %f2374, %f2372, %p288; add.ftz.f32 %f2389, %f2387, %f2388; mul.ftz.f32 %f2390, %f2389, 0f3F000000; sub.ftz.f32 %f2391, %f1936, %f2390; mul.ftz.f32 %f2392, %f2391, 0f40400000; sqrt.approx.ftz.f32 %f2393, %f2392; mul.ftz.f32 %f2394, %f2390, 0f3F000000; mul.ftz.f32 %f2395, %f355, %f2393; mul.ftz.f32 %f2396, %f2395, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2397, %f2396; add.ftz.f32 %f2398, %f2397, 0f3F800000; mul.ftz.f32 %f2399, %f2398, %f2394; mul.ftz.f32 %f2400, %f2393, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f2401, %f2399; setp.lt.ftz.f32 %p289, %f2402, %f362; selp.f32 %f2403, %f2387, %f2390, %p289; selp.f32 %f2404, %f2390, %f2388, %p289; add.ftz.f32 %f2405, %f2403, %f2404; mul.ftz.f32 %f2406, %f2405, 0f3F000000; sub.ftz.f32 %f2407, %f1936, %f2406; mul.ftz.f32 %f2408, %f2407, 0f40400000; sqrt.approx.ftz.f32 %f2409, %f2408; mul.ftz.f32 %f2410, %f2406, 0f3F000000; mul.ftz.f32 %f2411, %f355, %f2409; mul.ftz.f32 %f2412, %f2411, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2413, %f2412; add.ftz.f32 %f2414, %f2413, 0f3F800000; mul.ftz.f32 %f2415, %f2414, %f2410; mul.ftz.f32 %f2416, %f2409, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2417, %f2416; mul.ftz.f32 %f2418, %f2417, %f2415; setp.lt.ftz.f32 %p290, %f2418, %f362; selp.f32 %f2419, %f2403, %f2406, %p290; selp.f32 %f2420, %f2406, %f2404, %p290; add.ftz.f32 %f2421, %f2419, %f2420; mul.ftz.f32 %f2422, %f2421, 0f3F000000; sub.ftz.f32 %f2423, %f1936, %f2422; mul.ftz.f32 %f2424, %f2423, 0f40400000; sqrt.approx.ftz.f32 %f2425, %f2424; mul.ftz.f32 %f2426, %f2422, 0f3F000000; mul.ftz.f32 %f2427, %f355, %f2425; mul.ftz.f32 %f2428, %f2427, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2429, %f2428; add.ftz.f32 %f2430, %f2429, 0f3F800000; mul.ftz.f32 %f2431, %f2430, %f2426; mul.ftz.f32 %f2432, %f2425, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2433, %f2432; mul.ftz.f32 %f2434, %f2433, %f2431; setp.lt.ftz.f32 %p291, %f2434, %f362; selp.f32 %f2435, %f2419, %f2422, %p291; selp.f32 %f2436, %f2422, %f2420, %p291; add.ftz.f32 %f2437, %f2435, %f2436; mul.ftz.f32 %f2438, %f2437, 0f3F000000; sub.ftz.f32 %f2439, %f1936, %f2438; mul.ftz.f32 %f2440, %f2439, 0f40400000; sqrt.approx.ftz.f32 %f2441, %f2440; mul.ftz.f32 %f2442, %f2438, 0f3F000000; mul.ftz.f32 %f2443, %f355, %f2441; mul.ftz.f32 %f2444, %f2443, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2445, %f2444; add.ftz.f32 %f2446, %f2445, 0f3F800000; mul.ftz.f32 %f2447, %f2446, %f2442; mul.ftz.f32 %f2448, %f2441, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2449, %f2448; mul.ftz.f32 %f2450, %f2449, %f2447; setp.lt.ftz.f32 %p292, %f2450, %f362; selp.f32 %f2451, %f2435, %f2438, %p292; selp.f32 %f2452, %f2438, %f2436, %p292; add.ftz.f32 %f2453, %f2451, %f2452; mul.ftz.f32 %f2454, %f2453, 0f3F000000; sub.ftz.f32 %f2455, %f1936, %f2454; mul.ftz.f32 %f2456, %f2455, 0f40400000; sqrt.approx.ftz.f32 %f2457, %f2456; mul.ftz.f32 %f2458, %f2454, 0f3F000000; mul.ftz.f32 %f2459, %f355, %f2457; mul.ftz.f32 %f2460, %f2459, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2461, %f2460; add.ftz.f32 %f2462, %f2461, 0f3F800000; mul.ftz.f32 %f2463, %f2462, %f2458; mul.ftz.f32 %f2464, %f2457, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2465, %f2464; mul.ftz.f32 %f2466, %f2465, %f2463; setp.lt.ftz.f32 %p293, %f2466, %f362; selp.f32 %f2467, %f2451, %f2454, %p293; selp.f32 %f2468, %f2454, %f2452, %p293; add.ftz.f32 %f2469, %f2467, %f2468; mul.ftz.f32 %f3032, %f2469, 0f3F000000; $L__BB139_321: sub.ftz.f32 %f2471, %f1936, %f3030; mul.ftz.f32 %f2472, %f2471, 0f40400000; sub.ftz.f32 %f2473, %f1936, %f3031; mul.ftz.f32 %f2474, %f2473, 0f40400000; sub.ftz.f32 %f2475, %f1936, %f3032; mul.ftz.f32 %f2476, %f2475, 0f40400000; sqrt.approx.ftz.f32 %f2477, %f2472; sqrt.approx.ftz.f32 %f2478, %f2474; sqrt.approx.ftz.f32 %f2479, %f2476; ld.f32 %f2480, [%rd111]; mul.ftz.f32 %f3033, %f2477, %f2480; ld.f32 %f2481, [%rd111+4]; mul.ftz.f32 %f3034, %f2478, %f2481; ld.f32 %f2482, [%rd111+8]; mul.ftz.f32 %f3035, %f2479, %f2482; $L__BB139_323: st.f32 [%rd111], %f3033; st.f32 [%rd111+4], %f3034; st.f32 [%rd111+8], %f3035; or.b32 %r648, %r648, 16; $L__BB139_324: ld.local.u32 %r360, [%rd4+40]; or.b32 %r361, %r360, %r648; st.local.u32 [%rd4+40], %r361; bra.uni $L__BB139_517; $L__BB139_98: mul.wide.u32 %rd279, %r166, 4; add.s64 %rd280, %rd2, %rd279; ld.local.f32 %f2982, [%rd280]; ld.local.f32 %f2983, [%rd280+4]; ld.local.f32 %f2984, [%rd280+8]; $L__BB139_100: ld.global.v4.u32 {%r170, %r171, %r172, %r173}, [%rd3+48]; setp.eq.s32 %p56, %r170, 255; @%p56 bra $L__BB139_102; bra.uni $L__BB139_101; $L__BB139_102: ld.local.f32 %f2992, [%rd4+8]; ld.local.f32 %f2991, [%rd4+4]; ld.local.f32 %f2990, [%rd4]; bra.uni $L__BB139_103; $L__BB139_342: mul.wide.u32 %rd477, %r83, 4; add.s64 %rd478, %rd2, %rd477; ld.local.f32 %f3036, [%rd478]; $L__BB139_344: setp.eq.s32 %p305, %r82, 255; mul.ftz.f32 %f384, %f2, %f380; mul.ftz.f32 %f385, %f2, %f379; mul.ftz.f32 %f386, %f2, %f378; @%p305 bra $L__BB139_346; bra.uni $L__BB139_345; $L__BB139_346: mov.b32 %f3037, %r146; bra.uni $L__BB139_347; $L__BB139_101: mul.wide.u32 %rd281, %r170, 4; add.s64 %rd282, %rd2, %rd281; ld.local.f32 %f2992, [%rd282+8]; ld.local.f32 %f2991, [%rd282+4]; ld.local.f32 %f2990, [%rd282]; $L__BB139_103: @%p10 bra $L__BB139_112; ld.local.v2.f32 {%f677, %f678}, [%rd4+12]; mul.ftz.f32 %f679, %f2991, %f91; fma.rn.ftz.f32 %f680, %f2990, %f90, %f679; fma.rn.ftz.f32 %f681, %f2992, %f92, %f680; add.ftz.f32 %f682, %f681, %f681; mul.ftz.f32 %f683, %f2990, %f682; mul.ftz.f32 %f684, %f2991, %f682; mul.ftz.f32 %f685, %f2992, %f682; sub.ftz.f32 %f686, %f683, %f90; sub.ftz.f32 %f687, %f684, %f91; sub.ftz.f32 %f688, %f685, %f92; mul.ftz.f32 %f689, %f678, %f91; fma.rn.ftz.f32 %f690, %f677, %f90, %f689; ld.local.f32 %f123, [%rd4+20]; fma.rn.ftz.f32 %f124, %f123, %f92, %f690; mul.ftz.f32 %f691, %f124, 0f3F666666; mov.f32 %f692, 0f3C23D70A; min.ftz.f32 %f125, %f691, %f692; mul.ftz.f32 %f693, %f678, %f687; fma.rn.ftz.f32 %f694, %f677, %f686, %f693; fma.rn.ftz.f32 %f695, %f123, %f688, %f694; setp.ltu.ftz.f32 %p58, %f695, %f125; @%p58 bra $L__BB139_105; bra.uni $L__BB139_112; $L__BB139_105: mul.ftz.f32 %f696, %f2991, %f678; fma.rn.ftz.f32 %f697, %f2990, %f677, %f696; fma.rn.ftz.f32 %f698, %f2992, %f123, %f697; mul.ftz.f32 %f699, %f677, %f698; mul.ftz.f32 %f700, %f678, %f698; mul.ftz.f32 %f701, %f123, %f698; sub.ftz.f32 %f702, %f2990, %f699; sub.ftz.f32 %f703, %f2991, %f700; sub.ftz.f32 %f704, %f2992, %f701; mul.ftz.f32 %f705, %f703, %f703; fma.rn.ftz.f32 %f706, %f702, %f702, %f705; fma.rn.ftz.f32 %f707, %f704, %f704, %f706; sqrt.approx.ftz.f32 %f708, %f707; div.approx.ftz.f32 %f129, %f702, %f708; div.approx.ftz.f32 %f130, %f703, %f708; div.approx.ftz.f32 %f131, %f704, %f708; mul.ftz.f32 %f709, %f91, %f130; fma.rn.ftz.f32 %f710, %f90, %f129, %f709; fma.rn.ftz.f32 %f132, %f92, %f131, %f710; mul.ftz.f32 %f711, %f132, %f132; fma.rn.ftz.f32 %f712, %f124, %f124, %f711; mul.ftz.f32 %f713, %f125, %f125; sub.ftz.f32 %f714, %f712, %f713; mul.ftz.f32 %f715, %f711, %f714; mov.f32 %f716, 0f00000000; max.ftz.f32 %f717, %f715, %f716; sqrt.approx.ftz.f32 %f718, %f717; fma.rn.ftz.f32 %f719, %f125, %f124, %f712; mov.f32 %f720, 0f3F000000; div.approx.ftz.f32 %f721, %f720, %f712; add.ftz.f32 %f722, %f718, %f719; mul.ftz.f32 %f133, %f721, %f722; sub.ftz.f32 %f723, %f719, %f718; mul.ftz.f32 %f134, %f721, %f723; setp.gt.ftz.f32 %p59, %f133, 0f3727C5AC; setp.le.ftz.f32 %p60, %f133, 0f3F800054; and.pred %p3, %p59, %p60; setp.gt.ftz.f32 %p61, %f134, 0f3727C5AC; setp.le.ftz.f32 %p62, %f134, 0f3F800054; and.pred %p4, %p61, %p62; and.pred %p63, %p3, %p4; @%p63 bra $L__BB139_108; bra.uni $L__BB139_106; $L__BB139_108: mov.f32 %f730, 0f3F800000; sub.ftz.f32 %f731, %f730, %f133; max.ftz.f32 %f733, %f731, %f716; sqrt.approx.ftz.f32 %f140, %f733; max.ftz.f32 %f734, %f133, %f716; sqrt.approx.ftz.f32 %f141, %f734; sub.ftz.f32 %f735, %f730, %f134; max.ftz.f32 %f736, %f735, %f716; sqrt.approx.ftz.f32 %f142, %f736; max.ftz.f32 %f737, %f134, %f716; sqrt.approx.ftz.f32 %f143, %f737; mul.ftz.f32 %f738, %f124, %f141; fma.rn.ftz.f32 %f739, %f132, %f140, %f738; add.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f141, %f740; sub.ftz.f32 %f144, %f741, %f124; mul.ftz.f32 %f742, %f124, %f143; fma.rn.ftz.f32 %f743, %f132, %f142, %f742; add.ftz.f32 %f744, %f743, %f743; mul.ftz.f32 %f745, %f143, %f744; sub.ftz.f32 %f145, %f745, %f124; setp.ge.ftz.f32 %p65, %f144, 0f3727C5AC; setp.ge.ftz.f32 %p66, %f145, 0f3727C5AC; and.pred %p67, %p65, %p66; @%p67 bra $L__BB139_110; bra.uni $L__BB139_109; $L__BB139_110: setp.lt.ftz.f32 %p69, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p69; selp.f32 %f2989, %f141, %f143, %p69; bra.uni $L__BB139_111; $L__BB139_345: mul.wide.u32 %rd479, %r82, 4; add.s64 %rd480, %rd2, %rd479; ld.local.f32 %f3037, [%rd480]; $L__BB139_347: and.b32 %r86, %r375, 255; bfe.u32 %r87, %r375, 8, 8; bfe.u32 %r88, %r375, 16, 8; bfe.u32 %r89, %r379, 8, 8; bfe.u32 %r90, %r379, 16, 8; shr.u32 %r91, %r379, 24; ld.local.u32 %r92, [%rd4+64]; ld.local.s32 %rd122, [%rd4+48]; setp.eq.s32 %p306, %r92, -1; mov.u32 %r113, -1; mov.u32 %r110, 0; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p306 bra $L__BB139_362; ld.const.u64 %rd481, [kernel_params+2160]; cvta.to.global.u64 %rd482, %rd481; mul.wide.s32 %rd483, %r92, 208; add.s64 %rd484, %rd482, %rd483; ld.global.u32 %r93, [%rd484+164]; and.b32 %r388, %r11, 1; setp.eq.b32 %p307, %r388, 1; mov.pred %p308, 0; xor.pred %p309, %p307, %p308; not.pred %p310, %p309; mov.u32 %r653, 0; @%p310 bra $L__BB139_352; cvt.u32.u64 %r390, %rd122; setp.eq.s32 %p311, %r390, -1; mov.u32 %r652, -1; @%p311 bra $L__BB139_351; ld.const.u64 %rd485, [kernel_params+2240]; cvta.to.global.u64 %rd486, %rd485; shl.b64 %rd487, %rd122, 2; add.s64 %rd488, %rd486, %rd487; ld.global.u32 %r652, [%rd488]; $L__BB139_351: setp.ne.s32 %p312, %r652, -1; selp.u32 %r653, 1, 0, %p312; $L__BB139_352: ld.const.u64 %rd489, [kernel_params+2312]; cvta.to.global.u64 %rd123, %rd489; add.s32 %r655, %r653, %r93; mul.wide.u32 %rd490, %r655, 16; add.s64 %rd491, %rd123, %rd490; add.s64 %rd754, %rd491, 14; ld.global.u16 %rs20, [%rd491+12]; ld.global.u32 %r654, [%rd491+8]; ld.global.u64 %rd752, [%rd491]; setp.eq.s64 %p313, %rd752, %rd121; add.s64 %rd753, %rd491, 15; @%p313 bra $L__BB139_358; $L__BB139_353: setp.eq.s64 %p314, %rd752, 0; @%p314 bra $L__BB139_355; add.s32 %r655, %r655, 2; bra.uni $L__BB139_356; $L__BB139_355: setp.eq.s16 %p315, %rs20, 0; mov.u32 %r110, 0; mov.u32 %r655, %r654; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p315 bra $L__BB139_362; $L__BB139_356: cvt.u64.u32 %rd128, %r655; mul.wide.u32 %rd492, %r655, 16; add.s64 %rd493, %rd123, %rd492; ld.global.u16 %rs20, [%rd493+12]; ld.global.u32 %r654, [%rd493+8]; ld.global.u64 %rd752, [%rd493]; setp.ne.s64 %p316, %rd752, %rd121; @%p316 bra $L__BB139_353; shl.b64 %rd494, %rd128, 4; add.s64 %rd495, %rd123, %rd494; add.s64 %rd754, %rd495, 14; add.s64 %rd753, %rd495, 15; $L__BB139_358: ld.global.u8 %rs5, [%rd754]; ld.global.u8 %rs6, [%rd753]; cvt.u32.u64 %r395, %rd122; setp.ne.s32 %p317, %r395, -1; @%p317 bra $L__BB139_361; mov.u32 %r110, 0; add.s16 %rs14, %rs20, -1; setp.lt.u16 %p318, %rs14, 2; @%p318 bra $L__BB139_361; setp.ne.s16 %p319, %rs20, 1024; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p319 bra $L__BB139_362; $L__BB139_361: setp.eq.s16 %p320, %rs20, 0; selp.b32 %r113, -1, %r654, %p320; cvt.u32.u16 %r110, %rs20; cvt.u32.u16 %r400, %rs6; and.b32 %r660, %r400, 255; cvt.u32.u16 %r401, %rs5; and.b32 %r659, %r401, 255; $L__BB139_362: setp.eq.s32 %p321, %r113, -1; @%p321 bra $L__BB139_382; and.b32 %r402, %r11, 1; setp.eq.b32 %p322, %r402, 1; mov.pred %p323, 0; xor.pred %p324, %p322, %p323; not.pred %p325, %p324; @%p325 bra $L__BB139_375; bra.uni $L__BB139_364; $L__BB139_375: setp.eq.s32 %p332, %r12, 0; @%p332 bra $L__BB139_380; and.b32 %r423, %r110, 768; setp.eq.s32 %p333, %r423, 0; @%p333 bra $L__BB139_378; ld.const.u64 %rd517, [kernel_params+2264]; cvta.to.global.u64 %rd518, %rd517; shl.b64 %rd519, %rd122, 4; add.s64 %rd520, %rd518, %rd519; shr.s32 %r424, %r11, 7; ld.global.u32 %r425, [%rd520+4]; add.s32 %r426, %r424, %r425; ld.const.u64 %rd521, [kernel_params+2320]; cvta.to.global.u64 %rd522, %rd521; add.s32 %r427, %r426, %r113; mul.wide.s32 %rd523, %r427, 4; add.s64 %rd524, %rd522, %rd523; ld.local.f32 %f2520, [%rd4+56]; mov.f32 %f2521, 0f3F800000; sub.ftz.f32 %f2522, %f2521, %f2520; ld.global.f32 %f2523, [%rd524]; mul.ftz.f32 %f2524, %f2523, %f2522; ld.global.f32 %f2525, [%rd524+4]; fma.rn.ftz.f32 %f3038, %f2525, %f2520, %f2524; bra.uni $L__BB139_385; $L__BB139_472: mul.ftz.f32 %f2768, %f60, %f60; st.local.v2.f32 [%rd185+12], {%f2768, %f2768}; st.local.f32 [%rd185+20], %f3057; cvt.ftz.sat.f32.f32 %f507, %f2768; setp.eq.s32 %p411, %r2, 22; @%p411 bra $L__BB139_474; bra.uni $L__BB139_473; $L__BB139_474: st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r534, 22; st.local.u32 [%rd185+-8], %r534; ld.local.u32 %r535, [%rd4+40]; or.b32 %r536, %r535, 12; st.local.u32 [%rd4+40], %r536; bra.uni $L__BB139_517; $L__BB139_60: mov.u32 %r617, 19; st.local.u32 [%rd222+-8], %r617; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r618, [%rd4+40]; or.b32 %r619, %r618, 12; st.local.u32 [%rd4+40], %r619; bra.uni $L__BB139_517; $L__BB139_47: mov.f32 %f2876, 0f3F800000; sub.ftz.f32 %f2877, %f2876, %f528; mul.ftz.f32 %f3059, %f3058, %f2877; div.approx.ftz.f32 %f3058, %f3058, %f2877; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; $L__BB139_491: setp.eq.s32 %p436, %r2, 9; @%p436 bra $L__BB139_505; bra.uni $L__BB139_492; $L__BB139_505: mov.u32 %r597, 9; st.local.u32 [%rd205+-8], %r597; ld.local.u32 %r598, [%rd4+40]; or.b32 %r599, %r598, 4; st.local.u32 [%rd4+40], %r599; bra.uni $L__BB139_517; $L__BB139_492: setp.eq.s32 %p437, %r2, 13; @%p437 bra $L__BB139_504; bra.uni $L__BB139_493; $L__BB139_504: cvt.ftz.sat.f32.f32 %f2920, %f3058; cvt.ftz.sat.f32.f32 %f2921, %f3059; st.local.v2.f32 [%rd205+12], {%f2921, %f2920}; mov.u32 %r594, 13; st.local.u32 [%rd205+-8], %r594; ld.local.u32 %r595, [%rd4+40]; or.b32 %r596, %r595, 12; st.local.u32 [%rd4+40], %r596; bra.uni $L__BB139_517; $L__BB139_493: setp.eq.s32 %p438, %r2, 10; @%p438 bra $L__BB139_503; bra.uni $L__BB139_494; $L__BB139_503: mov.u64 %rd663, 0; st.local.u64 [%rd205+28], %rd663; cvt.ftz.sat.f32.f32 %f2918, %f3058; cvt.ftz.sat.f32.f32 %f2919, %f3059; st.local.v2.f32 [%rd205+12], {%f2919, %f2918}; mov.u32 %r591, 10; st.local.u32 [%rd205+-8], %r591; ld.local.u32 %r592, [%rd4+40]; or.b32 %r593, %r592, 12; st.local.u32 [%rd4+40], %r593; bra.uni $L__BB139_517; $L__BB139_441: setp.eq.s32 %p375, %r2, 25; st.local.v2.f32 [%rd160+12], {%f480, %f480}; st.local.f32 [%rd160+20], %f3054; cvt.ftz.sat.f32.f32 %f485, %f480; @%p375 bra $L__BB139_443; bra.uni $L__BB139_442; $L__BB139_443: st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r472, 13; st.local.u32 [%rd160+-8], %r472; ld.local.u32 %r473, [%rd4+40]; or.b32 %r474, %r473, 12; st.local.u32 [%rd4+40], %r474; bra.uni $L__BB139_445; $L__BB139_382: setp.eq.s32 %p339, %r89, 255; @%p339 bra $L__BB139_384; bra.uni $L__BB139_383; $L__BB139_384: mov.b32 %f3038, %r380; bra.uni $L__BB139_385; $L__BB139_364: cvt.u32.u64 %r404, %rd122; setp.eq.s32 %p326, %r404, -1; mov.u32 %r662, -1; @%p326 bra $L__BB139_366; ld.const.u64 %rd496, [kernel_params+2240]; cvta.to.global.u64 %rd497, %rd496; shl.b64 %rd498, %rd122, 2; add.s64 %rd499, %rd497, %rd498; ld.global.u32 %r662, [%rd499]; $L__BB139_366: setp.eq.s32 %p327, %r662, -1; @%p327 bra $L__BB139_368; mov.u64 %rd500, 0; { // callseq 889, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd230; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r110; st.param.b32 [param1+4], %r659; st.param.b32 [param1+8], %r660; st.param.b32 [param1+12], %r113; .param .b64 param2; st.param.b64 [param2+0], %rd500; .param .b64 param3; st.param.b64 [param3+0], %rd500; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f3038, [retval0+0]; } // callseq 889 bra.uni $L__BB139_385; $L__BB139_327: ld.local.f32 %f2492, [%rd4+280]; add.ftz.f32 %f2493, %f374, %f2492; ld.local.f32 %f2494, [%rd4+284]; add.ftz.f32 %f2495, %f375, %f2494; ld.local.f32 %f2496, [%rd4+288]; add.ftz.f32 %f2497, %f376, %f2496; st.local.f32 [%rd4+280], %f2493; st.local.f32 [%rd4+284], %f2495; st.local.f32 [%rd4+288], %f2497; ld.local.u32 %r68, [%rd4+248]; setp.lt.s32 %p296, %r68, 1; @%p296 bra $L__BB139_517; mov.u32 %r650, 0; $L__BB139_329: mul.wide.s32 %rd472, %r650, 80; add.s64 %rd473, %rd1, %rd472; add.s64 %rd118, %rd473, 316; ld.local.u32 %r364, [%rd473+316]; setp.eq.s32 %p297, %r364, 31; add.s32 %r650, %r650, 1; @%p297 bra $L__BB139_331; setp.lt.s32 %p298, %r650, %r68; @%p298 bra $L__BB139_329; bra.uni $L__BB139_517; $L__BB139_478: ld.local.f32 %f2802, [%rd4+116]; ld.local.f32 %f2803, [%rd4+120]; mul.ftz.f32 %f2804, %f2803, %f2803; fma.rn.ftz.f32 %f2805, %f2802, %f2802, %f2804; ld.local.f32 %f2806, [%rd4+124]; fma.rn.ftz.f32 %f2807, %f2806, %f2806, %f2805; sqrt.approx.ftz.f32 %f2808, %f2807; div.approx.ftz.f32 %f2809, %f2802, %f2808; div.approx.ftz.f32 %f2810, %f2803, %f2808; div.approx.ftz.f32 %f2811, %f2806, %f2808; st.local.f32 [%rd192+12], %f2809; st.local.f32 [%rd192+16], %f2810; st.local.f32 [%rd192+20], %f2811; mov.u32 %r547, 0; st.local.u32 [%rd192+32], %r547; $L__BB139_479: setp.eq.s32 %p418, %r2, 20; mov.f32 %f2812, 0f3A83126F; max.ftz.f32 %f2813, %f60, %f2812; mov.f32 %f2814, 0f3F800000; min.ftz.f32 %f513, %f2813, %f2814; max.ftz.f32 %f2815, %f63, %f2812; min.ftz.f32 %f514, %f2815, %f2814; @%p418 bra $L__BB139_481; bra.uni $L__BB139_480; $L__BB139_481: mov.u32 %r551, 20; st.local.u32 [%rd192+-8], %r551; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r552, [%rd4+40]; or.b32 %r553, %r552, 12; st.local.u32 [%rd4+40], %r553; bra.uni $L__BB139_517; $L__BB139_480: mov.u32 %r548, 30; st.local.u32 [%rd192+-8], %r548; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r549, [%rd4+40]; or.b32 %r550, %r549, 12; st.local.u32 [%rd4+40], %r550; bra.uni $L__BB139_517; $L__BB139_514: st.local.f32 [%rd229+12], %f60; mov.u32 %r630, 3; st.local.u32 [%rd229+-8], %r630; cvt.ftz.sat.f32.f32 %f2952, %f60; fma.rn.ftz.f32 %f2953, %f2952, 0f3F67750B, 0f40490FDB; rcp.approx.ftz.f32 %f2954, %f2953; st.local.f32 [%rd229+16], %f2954; mul.ftz.f32 %f2955, %f2952, %f2954; st.local.f32 [%rd229+20], %f2955; ld.local.u32 %r631, [%rd4+40]; or.b32 %r632, %r631, 12; st.local.u32 [%rd4+40], %r632; bra.uni $L__BB139_517; $L__BB139_473: st.local.u64 [%rd185+28], %rd621; st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r531, 23; st.local.u32 [%rd185+-8], %r531; ld.local.u32 %r532, [%rd4+40]; or.b32 %r533, %r532, 12; st.local.u32 [%rd4+40], %r533; bra.uni $L__BB139_517; $L__BB139_494: setp.eq.s32 %p439, %r2, 14; @%p439 bra $L__BB139_496; bra.uni $L__BB139_495; $L__BB139_496: ld.local.u32 %r131, [%rd4+252]; setp.lt.s32 %p440, %r131, 1; @%p440 bra $L__BB139_498; add.s32 %r582, %r131, -1; st.local.u32 [%rd4+252], %r582; ld.local.s32 %rd651, [%rd4+248]; cvt.s64.s32 %rd652, %r582; add.s64 %rd653, %rd651, %rd652; mul.lo.s64 %rd654, %rd653, 80; add.s64 %rd655, %rd1, %rd654; add.s64 %rd794, %rd655, 304; add.s64 %rd656, %rd230, %rd654; add.s64 %rd795, %rd656, 304; bra.uni $L__BB139_499; $L__BB139_457: ld.local.u32 %r508, [%rd4+248]; mov.u64 %rd776, 0; add.s32 %r509, %r508, -1; st.local.u32 [%rd4+248], %r509; add.s32 %r510, %r126, 1; st.local.u32 [%rd4+252], %r510; mov.u64 %rd777, %rd776; $L__BB139_458: setp.eq.s64 %p394, %rd777, 0; @%p394 bra $L__BB139_517; add.s64 %rd178, %rd774, 20; st.local.f32 [%rd774+20], %f55; st.local.f32 [%rd774+24], %f56; st.local.f32 [%rd774+28], %f57; st.local.u64 [%rd774+48], %rd777; mov.u32 %r511, 0; st.local.v2.f32 [%rd774+56], {%f2732, %f2732}; st.local.u32 [%rd774+64], %r511; mul.ftz.f32 %f2737, %f60, %f60; st.local.v2.f32 [%rd774+32], {%f2737, %f2737}; mov.f32 %f2738, 0f3727C5AC; max.ftz.f32 %f3056, %f63, %f2738; ld.local.u32 %r512, [%rd4+40]; and.b32 %r513, %r512, 1; setp.eq.b32 %p395, %r513, 1; mov.pred %p396, 0; xor.pred %p397, %p395, %p396; not.pred %p398, %p397; @%p398 bra $L__BB139_461; rcp.approx.ftz.f32 %f3056, %f3056; $L__BB139_461: st.local.f32 [%rd178+20], %f3056; ld.local.f32 %f2739, [%rd6]; ld.local.f32 %f2740, [%rd6+4]; ld.local.f32 %f2741, [%rd6+8]; st.local.f32 [%rd776], %f2739; st.local.f32 [%rd776+4], %f2740; st.local.f32 [%rd776+8], %f2741; ld.local.u64 %rd609, [%rd178+28]; st.u32 [%rd609+12], %r511; st.u32 [%rd609+16], %r511; st.u32 [%rd609+20], %r511; ld.local.u64 %rd610, [%rd178+28]; st.u32 [%rd610+36], %r511; ld.local.f32 %f2743, [%rd178+12]; mov.f32 %f2744, 0f38D1B717; max.ftz.f32 %f2745, %f2743, %f2744; mov.f32 %f2746, 0f3F800000; min.ftz.f32 %f2747, %f2745, %f2746; st.local.v2.f32 [%rd178+12], {%f2747, %f2747}; ld.local.f32 %f2748, [%rd178+20]; max.ftz.f32 %f2749, %f2732, %f2748; st.local.f32 [%rd178+20], %f2749; ld.local.u64 %rd611, [%rd178+28]; ld.f32 %f2750, [%rd611]; cvt.ftz.sat.f32.f32 %f2751, %f2750; ld.f32 %f2752, [%rd611+4]; cvt.ftz.sat.f32.f32 %f2753, %f2752; ld.f32 %f2754, [%rd611+8]; cvt.ftz.sat.f32.f32 %f2755, %f2754; st.f32 [%rd611], %f2751; st.f32 [%rd611+4], %f2753; st.f32 [%rd611+8], %f2755; mov.u32 %r515, 24; st.local.u32 [%rd178+-8], %r515; ld.local.u32 %r516, [%rd4+40]; or.b32 %r517, %r516, 1036; st.local.u32 [%rd4+40], %r517; bra.uni $L__BB139_517; $L__BB139_442: st.local.u64 [%rd160+28], %rd579; st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r469, 10; st.local.u32 [%rd160+-8], %r469; ld.local.u32 %r470, [%rd4+40]; or.b32 %r471, %r470, 12; st.local.u32 [%rd4+40], %r471; $L__BB139_445: setp.ne.s32 %p376, %r122, 0; ld.const.u32 %r478, [kernel_params+1568]; setp.eq.s32 %p377, %r478, 0; and.pred %p378, %p376, %p377; @%p378 bra $L__BB139_517; mov.f32 %f2713, 0f3F800000; sub.ftz.f32 %f2714, %f2713, %f3055; mov.f32 %f2715, 0f38D1B717; max.ftz.f32 %f2716, %f2715, %f2714; mul.ftz.f32 %f2717, %f470, %f2716; mul.ftz.f32 %f2718, %f471, %f2716; mul.ftz.f32 %f2719, %f472, %f2716; mov.f32 %f2720, 0f00000000; max.ftz.f32 %f487, %f2717, %f2720; max.ftz.f32 %f488, %f2718, %f2720; max.ftz.f32 %f489, %f2719, %f2720; add.ftz.f32 %f2721, %f487, %f488; add.ftz.f32 %f2722, %f489, %f2721; mul.ftz.f32 %f2723, %f2722, 0f3EAAAAAB; abs.ftz.f32 %f490, %f2723; setp.ltu.ftz.f32 %p379, %f490, 0f3727C5AC; mov.u64 %rd770, 0; mov.u64 %rd771, %rd770; @%p379 bra $L__BB139_451; ld.local.u32 %r124, [%rd4+252]; mov.u64 %rd770, 0; setp.eq.s32 %p380, %r124, 0; mov.u64 %rd768, %rd770; mov.u64 %rd769, %rd770; @%p380 bra $L__BB139_449; ld.local.u32 %r479, [%rd4+248]; mul.wide.s32 %rd585, %r479, 80; add.s64 %rd586, %rd1, %rd585; add.s64 %rd768, %rd586, 304; add.s64 %rd587, %rd230, %rd585; add.s64 %rd769, %rd587, 304; mov.b32 %r480, %f489; mov.b32 %r481, %f488; mov.b32 %r482, %f487; mov.u32 %r483, 0; st.local.v4.u32 [%rd586+304], {%r482, %r481, %r480, %r483}; add.s32 %r484, %r479, 1; st.local.u32 [%rd4+248], %r484; add.s32 %r485, %r124, -1; st.local.u32 [%rd4+252], %r485; $L__BB139_449: setp.eq.s64 %p381, %rd769, 0; mov.u64 %rd771, %rd770; @%p381 bra $L__BB139_451; st.local.f32 [%rd768+16], %f490; mov.u64 %rd770, %rd768; mov.u64 %rd771, %rd769; $L__BB139_451: setp.eq.s64 %p382, %rd771, 0; @%p382 bra $L__BB139_517; add.s64 %rd167, %rd770, 20; mov.u64 %rd590, 0; st.local.f32 [%rd770+20], %f55; st.local.f32 [%rd770+24], %f56; st.local.f32 [%rd770+28], %f57; mov.u32 %r486, 0; st.local.v2.f32 [%rd770+56], {%f2720, %f2720}; st.local.u32 [%rd770+64], %r486; st.local.u64 [%rd770+48], %rd590; setp.eq.s32 %p383, %r2, 28; @%p383 bra $L__BB139_456; bra.uni $L__BB139_453; $L__BB139_456: st.local.v2.f32 [%rd167+12], {%f2720, %f2720}; st.local.f32 [%rd167+20], %f3054; mov.u32 %r493, 21; st.local.u32 [%rd167+-8], %r493; ld.local.u32 %r494, [%rd4+40]; or.b32 %r495, %r494, 4; st.local.u32 [%rd4+40], %r495; bra.uni $L__BB139_517; $L__BB139_453: setp.eq.s32 %p384, %r2, 25; st.local.v2.f32 [%rd167+12], {%f480, %f480}; st.local.f32 [%rd167+20], %f3054; cvt.ftz.sat.f32.f32 %f491, %f480; @%p384 bra $L__BB139_455; bra.uni $L__BB139_454; $L__BB139_455: st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r490, 22; st.local.u32 [%rd167+-8], %r490; ld.local.u32 %r491, [%rd4+40]; or.b32 %r492, %r491, 12; st.local.u32 [%rd4+40], %r492; bra.uni $L__BB139_517; $L__BB139_454: st.local.u64 [%rd167+28], %rd590; st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r487, 23; st.local.u32 [%rd167+-8], %r487; ld.local.u32 %r488, [%rd4+40]; or.b32 %r489, %r488, 12; st.local.u32 [%rd4+40], %r489; bra.uni $L__BB139_517; $L__BB139_383: mul.wide.u32 %rd533, %r89, 4; add.s64 %rd534, %rd2, %rd533; ld.local.f32 %f3038, [%rd534]; bra.uni $L__BB139_385; $L__BB139_380: and.b32 %r432, %r11, 8; setp.eq.s32 %p336, %r432, 0; setp.ne.s32 %p337, %r110, 8; mov.f32 %f3038, 0f00000000; or.pred %p338, %p337, %p336; @%p338 bra $L__BB139_385; cvt.u32.u64 %r433, %rd122; ld.const.u64 %rd529, [kernel_params+2320]; cvta.to.global.u64 %rd530, %rd529; add.s32 %r434, %r433, %r113; mul.wide.s32 %rd531, %r434, 4; add.s64 %rd532, %rd530, %rd531; ld.global.f32 %f3038, [%rd532]; bra.uni $L__BB139_385; $L__BB139_368: and.b32 %r405, %r110, 56; setp.eq.s32 %p328, %r405, 0; @%p328 bra $L__BB139_373; and.b32 %r406, %r110, 24; setp.eq.s32 %p329, %r406, 0; @%p329 bra $L__BB139_371; ld.const.u64 %rd501, [kernel_params+2232]; cvta.to.global.u64 %rd502, %rd501; shl.b64 %rd503, %rd122, 4; add.s64 %rd504, %rd502, %rd503; ld.global.v4.u32 {%r407, %r408, %r409, %r410}, [%rd504]; ld.const.u64 %rd505, [kernel_params+2320]; cvta.to.global.u64 %rd506, %rd505; add.s32 %r414, %r407, %r113; mul.wide.u32 %rd507, %r414, 4; add.s64 %rd755, %rd506, %rd507; add.s32 %r415, %r408, %r113; mul.wide.u32 %rd508, %r415, 4; add.s64 %rd756, %rd506, %rd508; add.s32 %r416, %r409, %r113; mul.wide.u32 %rd509, %r416, 4; add.s64 %rd757, %rd506, %rd509; bra.uni $L__BB139_372; $L__BB139_495: mov.f32 %f2880, 0f38D1B717; max.ftz.f32 %f2881, %f3059, %f2880; mov.f32 %f2882, 0f3F800000; max.ftz.f32 %f2883, %f3058, %f2880; min.ftz.f32 %f2884, %f2883, %f2882; min.ftz.f32 %f2885, %f2881, %f2882; st.local.v2.f32 [%rd205+12], {%f2885, %f2884}; mov.u32 %r579, 16; st.local.u32 [%rd205+-8], %r579; ld.local.u32 %r580, [%rd4+40]; or.b32 %r581, %r580, 12; st.local.u32 [%rd4+40], %r581; bra.uni $L__BB139_517; $L__BB139_334: ld.local.u32 %r651, [%rd4+252]; $L__BB139_335: setp.eq.s32 %p301, %r651, 0; mov.pred %p462, -1; mov.u64 %rd751, 0; @%p301 bra $L__BB139_337; ld.local.u32 %r367, [%rd4+248]; mul.wide.s32 %rd475, %r367, 80; add.s64 %rd476, %rd1, %rd475; add.s64 %rd751, %rd476, 304; mov.b32 %r368, %f376; mov.b32 %r369, %f375; mov.b32 %r370, %f374; mov.u32 %r371, 31; st.local.v4.u32 [%rd476+304], {%r370, %r369, %r368, %r371}; add.s32 %r372, %r367, 1; st.local.u32 [%rd4+248], %r372; add.s32 %r373, %r651, -1; st.local.u32 [%rd4+252], %r373; mov.pred %p462, 0; $L__BB139_337: @%p462 bra $L__BB139_339; bra.uni $L__BB139_338; $L__BB139_339: @%p299 bra $L__BB139_517; mov.u32 %r374, 0; st.local.u32 [%rd4+252], %r374; bra.uni $L__BB139_517; $L__BB139_338: st.local.f32 [%rd751+16], %f377; ld.local.f32 %f2506, [%rd4]; ld.local.f32 %f2507, [%rd4+4]; ld.local.f32 %f2508, [%rd4+8]; st.local.f32 [%rd751+20], %f2506; st.local.f32 [%rd751+24], %f2507; st.local.f32 [%rd751+28], %f2508; bra.uni $L__BB139_517; $L__BB139_106: or.pred %p64, %p3, %p4; mov.f32 %f2990, %f677; mov.f32 %f2991, %f678; mov.f32 %f2992, %f123; @%p64 bra $L__BB139_107; bra.uni $L__BB139_112; $L__BB139_107: selp.f32 %f724, %f133, %f134, %p3; mov.f32 %f725, 0f3F800000; sub.ftz.f32 %f726, %f725, %f724; max.ftz.f32 %f728, %f726, %f716; sqrt.approx.ftz.f32 %f2988, %f728; max.ftz.f32 %f729, %f724, %f716; sqrt.approx.ftz.f32 %f2989, %f729; bra.uni $L__BB139_111; $L__BB139_331: ld.local.f32 %f2498, [%rd118+-12]; add.ftz.f32 %f2499, %f374, %f2498; ld.local.f32 %f2500, [%rd118+-8]; add.ftz.f32 %f2501, %f375, %f2500; ld.local.f32 %f2502, [%rd118+-4]; add.ftz.f32 %f2503, %f376, %f2502; st.local.f32 [%rd118+-12], %f2499; st.local.f32 [%rd118+-8], %f2501; st.local.f32 [%rd118+-4], %f2503; ld.local.f32 %f2504, [%rd118+4]; add.ftz.f32 %f2505, %f377, %f2504; st.local.f32 [%rd118+4], %f2505; bra.uni $L__BB139_517; $L__BB139_378: and.b32 %r428, %r110, 131; setp.eq.s32 %p334, %r428, 0; mov.f32 %f3038, 0f00000000; @%p334 bra $L__BB139_385; cvt.u32.u64 %r429, %rd122; setp.eq.s32 %p335, %r110, 128; selp.b32 %r430, %r429, 0, %p335; add.s32 %r431, %r113, %r430; ld.const.u64 %rd525, [kernel_params+2320]; cvta.to.global.u64 %rd526, %rd525; mul.wide.s32 %rd527, %r431, 4; add.s64 %rd528, %rd526, %rd527; ld.global.f32 %f3038, [%rd528]; bra.uni $L__BB139_385; $L__BB139_109: setp.gt.ftz.f32 %p68, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p68; selp.f32 %f2989, %f141, %f143, %p68; $L__BB139_111: mul.ftz.f32 %f746, %f677, %f2989; mul.ftz.f32 %f747, %f678, %f2989; mul.ftz.f32 %f748, %f123, %f2989; fma.rn.ftz.f32 %f2992, %f131, %f2988, %f748; fma.rn.ftz.f32 %f2991, %f130, %f2988, %f747; fma.rn.ftz.f32 %f2990, %f129, %f2988, %f746; $L__BB139_112: setp.eq.s32 %p70, %r171, 255; mov.f32 %f2993, 0f3F800000; mov.f32 %f2994, %f2993; mov.f32 %f2995, %f2993; @%p70 bra $L__BB139_114; mul.wide.u32 %rd283, %r171, 4; add.s64 %rd284, %rd2, %rd283; ld.local.f32 %f2993, [%rd284]; ld.local.f32 %f2994, [%rd284+4]; ld.local.f32 %f2995, [%rd284+8]; $L__BB139_114: setp.eq.s32 %p71, %r172, 255; mov.f32 %f2996, 0f3FB33333; @%p71 bra $L__BB139_116; mul.wide.u32 %rd285, %r172, 4; add.s64 %rd286, %rd2, %rd285; ld.local.f32 %f2996, [%rd286]; $L__BB139_116: setp.eq.s32 %p72, %r173, 255; mov.f32 %f2997, 0f00000000; @%p72 bra $L__BB139_118; mul.wide.u32 %rd287, %r173, 4; add.s64 %rd288, %rd2, %rd287; ld.local.f32 %f2997, [%rd288]; $L__BB139_118: ld.global.v4.u32 {%r174, %r175, %r176, %r177}, [%rd3+64]; setp.eq.s32 %p73, %r174, 255; @%p73 bra $L__BB139_120; bra.uni $L__BB139_119; $L__BB139_120: mov.b32 %f2998, %r175; mov.b32 %f2999, %r176; mov.b32 %f3000, %r177; bra.uni $L__BB139_121; $L__BB139_119: mul.wide.u32 %rd289, %r174, 4; add.s64 %rd290, %rd2, %rd289; ld.local.f32 %f2998, [%rd290]; ld.local.f32 %f2999, [%rd290+4]; ld.local.f32 %f3000, [%rd290+8]; $L__BB139_121: ld.local.f32 %f754, [%rd4+256]; mul.ftz.f32 %f177, %f2, %f754; ld.local.f32 %f755, [%rd4+260]; mul.ftz.f32 %f178, %f2, %f755; ld.local.f32 %f756, [%rd4+264]; mul.ftz.f32 %f179, %f2, %f756; mov.f32 %f757, 0f3F800000; sub.ftz.f32 %f758, %f757, %f63; mul.ftz.f32 %f759, %f63, %f2998; fma.rn.ftz.f32 %f180, %f758, %f2982, %f759; mul.ftz.f32 %f760, %f63, %f2999; fma.rn.ftz.f32 %f181, %f758, %f2983, %f760; mul.ftz.f32 %f761, %f63, %f3000; fma.rn.ftz.f32 %f182, %f758, %f2984, %f761; mul.ftz.f32 %f762, %f180, %f177; mul.ftz.f32 %f763, %f181, %f178; mul.ftz.f32 %f764, %f182, %f179; mul.ftz.f32 %f183, %f97, %f762; mul.ftz.f32 %f184, %f97, %f763; mul.ftz.f32 %f185, %f97, %f764; and.b32 %r178, %r141, 4096; setp.eq.s32 %p74, %r178, 0; selp.f32 %f186, %f2982, %f180, %p74; selp.f32 %f187, %f2983, %f181, %p74; selp.f32 %f188, %f2984, %f182, %p74; selp.f32 %f189, %f63, 0f00000000, %p74; add.ftz.f32 %f765, %f180, %f181; add.ftz.f32 %f766, %f182, %f765; mul.ftz.f32 %f767, %f766, 0f3EAAAAAB; abs.ftz.f32 %f768, %f767; setp.leu.ftz.f32 %p75, %f768, 0f3727C5AC; @%p75 bra $L__BB139_182; setp.le.ftz.f32 %p76, %f189, 0f3727C5AC; setp.gt.ftz.f32 %p77, %f97, 0f3727C5AC; and.pred %p78, %p76, %p77; @%p78 bra $L__BB139_175; bra.uni $L__BB139_123; $L__BB139_175: mul.ftz.f32 %f1391, %f186, %f177; mul.ftz.f32 %f1392, %f97, %f1391; mul.ftz.f32 %f1393, %f187, %f178; mul.ftz.f32 %f1394, %f97, %f1393; mul.ftz.f32 %f1395, %f188, %f179; mul.ftz.f32 %f1396, %f97, %f1395; mov.f32 %f1397, 0f00000000; max.ftz.f32 %f223, %f1392, %f1397; max.ftz.f32 %f224, %f1394, %f1397; max.ftz.f32 %f225, %f1396, %f1397; add.ftz.f32 %f1398, %f223, %f224; add.ftz.f32 %f1399, %f225, %f1398; mul.ftz.f32 %f1400, %f1399, 0f3EAAAAAB; abs.ftz.f32 %f226, %f1400; setp.ltu.ftz.f32 %p143, %f226, 0f3727C5AC; mov.u64 %rd708, 0; mov.u64 %rd709, %rd708; @%p143 bra $L__BB139_180; ld.local.u32 %r42, [%rd4+252]; mov.u64 %rd708, 0; setp.eq.s32 %p144, %r42, 0; mov.u64 %rd706, %rd708; mov.u64 %rd707, %rd708; @%p144 bra $L__BB139_178; ld.local.u32 %r226, [%rd4+248]; mul.wide.s32 %rd328, %r226, 80; add.s64 %rd329, %rd1, %rd328; add.s64 %rd706, %rd329, 304; add.s64 %rd330, %rd230, %rd328; add.s64 %rd707, %rd330, 304; mov.b32 %r227, %f225; mov.b32 %r228, %f224; mov.b32 %r229, %f223; mov.u32 %r230, 0; st.local.v4.u32 [%rd329+304], {%r229, %r228, %r227, %r230}; add.s32 %r231, %r226, 1; st.local.u32 [%rd4+248], %r231; add.s32 %r232, %r42, -1; st.local.u32 [%rd4+252], %r232; $L__BB139_178: setp.eq.s64 %p145, %rd707, 0; mov.u64 %rd709, %rd708; @%p145 bra $L__BB139_180; st.local.f32 [%rd706+16], %f226; mov.u64 %rd708, %rd706; mov.u64 %rd709, %rd707; $L__BB139_180: setp.eq.s64 %p146, %rd709, 0; @%p146 bra $L__BB139_182; st.local.f32 [%rd708+20], %f55; st.local.f32 [%rd708+24], %f56; st.local.f32 [%rd708+28], %f57; st.local.f32 [%rd708+32], %f71; mov.u32 %r233, 5; st.local.u32 [%rd708+12], %r233; mov.u32 %r234, 1; st.local.u32 [%rd708+36], %r234; ld.local.u32 %r235, [%rd4+40]; or.b32 %r236, %r235, 12; st.local.u32 [%rd4+40], %r236; bra.uni $L__BB139_182; $L__BB139_123: setp.leu.ftz.f32 %p79, %f189, 0f3727C5AC; @%p79 bra $L__BB139_182; ld.local.u32 %r29, [%rd4+252]; mov.u64 %rd693, 0; setp.eq.s32 %p80, %r29, 0; mov.u64 %rd691, %rd693; mov.u64 %rd692, %rd693; @%p80 bra $L__BB139_126; ld.local.u32 %r179, [%rd4+248]; mul.wide.s32 %rd293, %r179, 80; add.s64 %rd294, %rd1, %rd293; add.s64 %rd691, %rd294, 304; add.s64 %rd295, %rd230, %rd293; add.s64 %rd692, %rd295, 304; mov.b32 %r180, %f185; mov.b32 %r181, %f184; mov.b32 %r182, %f183; mov.u32 %r183, 0; st.local.v4.u32 [%rd294+304], {%r182, %r181, %r180, %r183}; add.s32 %r184, %r179, 1; st.local.u32 [%rd4+248], %r184; add.s32 %r185, %r29, -1; st.local.u32 [%rd4+252], %r185; $L__BB139_126: setp.eq.s64 %p81, %rd692, 0; @%p81 bra $L__BB139_128; add.ftz.f32 %f769, %f183, %f184; add.ftz.f32 %f770, %f185, %f769; mul.ftz.f32 %f771, %f770, 0f3EAAAAAB; abs.ftz.f32 %f772, %f771; st.local.f32 [%rd691+16], %f772; setp.ge.ftz.f32 %p82, %f772, 0f3727C5AC; selp.b64 %rd693, %rd692, 0, %p82; $L__BB139_128: setp.eq.s64 %p83, %rd693, 0; @%p83 bra $L__BB139_182; mul.ftz.f32 %f773, %f189, %f2993; add.s64 %rd14, %rd693, 32; st.f32 [%rd693+32], %f773; mul.ftz.f32 %f774, %f189, %f2994; st.f32 [%rd693+36], %f774; mul.ftz.f32 %f775, %f189, %f2995; st.f32 [%rd693+40], %f775; st.f32 [%rd693+44], %f180; st.f32 [%rd693+48], %f181; st.f32 [%rd693+52], %f182; st.f32 [%rd693+20], %f55; st.f32 [%rd693+24], %f56; st.f32 [%rd693+28], %f57; st.f32 [%rd693+56], %f71; mov.f32 %f3003, 0f00000000; max.ftz.f32 %f777, %f2997, %f3003; mov.f32 %f778, 0f3F666666; min.ftz.f32 %f779, %f777, %f778; st.f32 [%rd693+60], %f779; setp.eq.ftz.f32 %p84, %f71, 0f7F7FFFFF; mov.u32 %r643, 0; @%p84 bra $L__BB139_137; ld.f32 %f780, [%rd14+-32]; mov.u64 %rd696, 0; mov.f32 %f781, 0f00000000; max.ftz.f32 %f190, %f780, %f781; ld.f32 %f782, [%rd14+-28]; max.ftz.f32 %f191, %f782, %f781; ld.f32 %f783, [%rd14+-24]; max.ftz.f32 %f192, %f783, %f781; add.ftz.f32 %f784, %f190, %f191; add.ftz.f32 %f785, %f192, %f784; mul.ftz.f32 %f786, %f785, 0f3EAAAAAB; abs.ftz.f32 %f193, %f786; setp.ltu.ftz.f32 %p85, %f193, 0f3727C5AC; mov.u64 %rd697, %rd696; @%p85 bra $L__BB139_135; ld.local.u32 %r30, [%rd4+252]; mov.u64 %rd696, 0; setp.eq.s32 %p86, %r30, 0; mov.u64 %rd694, %rd696; mov.u64 %rd695, %rd696; @%p86 bra $L__BB139_133; ld.local.u32 %r187, [%rd4+248]; mul.wide.s32 %rd301, %r187, 80; add.s64 %rd302, %rd1, %rd301; add.s64 %rd694, %rd302, 304; add.s64 %rd303, %rd230, %rd301; add.s64 %rd695, %rd303, 304; mov.b32 %r188, %f192; mov.b32 %r189, %f191; mov.b32 %r190, %f190; mov.u32 %r191, 0; st.local.v4.u32 [%rd302+304], {%r190, %r189, %r188, %r191}; add.s32 %r192, %r187, 1; st.local.u32 [%rd4+248], %r192; add.s32 %r193, %r30, -1; st.local.u32 [%rd4+252], %r193; $L__BB139_133: setp.eq.s64 %p87, %rd695, 0; mov.u64 %rd697, %rd696; @%p87 bra $L__BB139_135; st.local.f32 [%rd694+16], %f193; mov.u64 %rd696, %rd694; mov.u64 %rd697, %rd695; $L__BB139_135: setp.eq.s64 %p88, %rd697, 0; @%p88 bra $L__BB139_137; ld.f32 %f787, [%rd14+-12]; ld.f32 %f788, [%rd14+-8]; ld.f32 %f789, [%rd14+-4]; st.local.f32 [%rd696+20], %f787; st.local.f32 [%rd696+24], %f788; st.local.f32 [%rd696+28], %f789; ld.f32 %f790, [%rd14+24]; st.local.f32 [%rd696+32], %f790; mov.u32 %r196, 5; st.local.u32 [%rd696+12], %r196; mov.u32 %r197, 8; st.local.u32 [%rd696+36], %r197; ld.local.f32 %f791, [%rd4+24]; ld.local.f32 %f792, [%rd4+28]; mul.ftz.f32 %f793, %f792, %f788; fma.rn.ftz.f32 %f794, %f791, %f787, %f793; ld.local.f32 %f795, [%rd4+32]; fma.rn.ftz.f32 %f796, %f795, %f789, %f794; mov.f32 %f797, 0f3F800000; sub.ftz.f32 %f798, %f797, %f796; max.ftz.f32 %f800, %f798, %f781; min.ftz.f32 %f801, %f800, %f797; mul.ftz.f32 %f802, %f801, %f801; mul.ftz.f32 %f803, %f802, %f802; mul.ftz.f32 %f804, %f801, %f803; mul.ftz.f32 %f805, %f790, %f804; ld.local.f32 %f806, [%rd696+16]; mul.ftz.f32 %f807, %f806, %f805; st.local.f32 [%rd696+16], %f807; mov.u32 %r643, 12; $L__BB139_137: ld.f32 %f809, [%rd14]; setp.geu.ftz.f32 %p89, %f809, 0f322BCC77; mov.u32 %r641, 3; mov.f32 %f3001, %f3003; @%p89 bra $L__BB139_139; ld.f32 %f3001, [%rd14+-32]; mov.u32 %r200, 0; st.u32 [%rd14+-32], %r200; st.u32 [%rd14], %r200; mov.u32 %r641, 2; $L__BB139_139: ld.f32 %f811, [%rd14+4]; setp.geu.ftz.f32 %p90, %f811, 0f322BCC77; mov.f32 %f3002, %f3003; @%p90 bra $L__BB139_141; ld.f32 %f3002, [%rd14+-28]; mov.u32 %r201, 0; st.u32 [%rd14+-28], %r201; st.u32 [%rd14+4], %r201; add.s32 %r641, %r641, -1; $L__BB139_141: ld.f32 %f813, [%rd14+8]; setp.geu.ftz.f32 %p91, %f813, 0f322BCC77; @%p91 bra $L__BB139_143; ld.f32 %f3003, [%rd14+-24]; mov.u32 %r202, 0; st.u32 [%rd14+-24], %r202; st.u32 [%rd14+8], %r202; add.s32 %r641, %r641, -1; $L__BB139_143: setp.gt.s32 %p92, %r641, 2; @%p92 bra $L__BB139_159; ld.f32 %f814, [%rd14+24]; setp.neu.ftz.f32 %p93, %f814, 0f7F7FFFFF; mov.f32 %f815, 0f00000000; max.ftz.f32 %f200, %f3002, %f815; max.ftz.f32 %f201, %f3001, %f815; add.ftz.f32 %f816, %f201, %f200; max.ftz.f32 %f202, %f3003, %f815; add.ftz.f32 %f817, %f202, %f816; mul.ftz.f32 %f818, %f817, 0f3EAAAAAB; abs.ftz.f32 %f203, %f818; @%p93 bra $L__BB139_152; bra.uni $L__BB139_145; $L__BB139_152: setp.ltu.ftz.f32 %p98, %f203, 0f3727C5AC; mov.u64 %rd704, 0; mov.u64 %rd705, %rd704; @%p98 bra $L__BB139_157; ld.local.u32 %r38, [%rd4+252]; mov.u64 %rd704, 0; setp.eq.s32 %p99, %r38, 0; mov.u64 %rd702, %rd704; mov.u64 %rd703, %rd704; @%p99 bra $L__BB139_155; ld.local.u32 %r212, [%rd4+248]; mul.wide.s32 %rd319, %r212, 80; add.s64 %rd320, %rd1, %rd319; add.s64 %rd702, %rd320, 304; add.s64 %rd321, %rd230, %rd319; add.s64 %rd703, %rd321, 304; mov.b32 %r213, %f202; mov.b32 %r214, %f200; mov.b32 %r215, %f201; mov.u32 %r216, 0; st.local.v4.u32 [%rd320+304], {%r215, %r214, %r213, %r216}; add.s32 %r217, %r212, 1; st.local.u32 [%rd4+248], %r217; add.s32 %r218, %r38, -1; st.local.u32 [%rd4+252], %r218; $L__BB139_155: setp.eq.s64 %p100, %rd703, 0; mov.u64 %rd705, %rd704; @%p100 bra $L__BB139_157; st.local.f32 [%rd702+16], %f203; mov.u64 %rd704, %rd702; mov.u64 %rd705, %rd703; $L__BB139_157: setp.eq.s64 %p101, %rd705, 0; @%p101 bra $L__BB139_159; ld.f32 %f822, [%rd14+-12]; ld.f32 %f823, [%rd14+-8]; ld.f32 %f824, [%rd14+-4]; st.local.f32 [%rd704+20], %f822; st.local.f32 [%rd704+24], %f823; st.local.f32 [%rd704+28], %f824; ld.f32 %f825, [%rd14+24]; st.local.f32 [%rd704+32], %f825; mov.u32 %r220, 5; st.local.u32 [%rd704+12], %r220; mov.u32 %r221, 2; st.local.u32 [%rd704+36], %r221; mov.u32 %r643, 12; bra.uni $L__BB139_159; $L__BB139_373: and.b32 %r419, %r110, 7; setp.eq.s32 %p330, %r419, 0; mov.f32 %f3038, 0f00000000; @%p330 bra $L__BB139_385; setp.eq.s32 %p331, %r110, 4; selp.b32 %r421, %r404, 0, %p331; add.s32 %r422, %r113, %r421; ld.const.u64 %rd513, [kernel_params+2320]; cvta.to.global.u64 %rd514, %rd513; mul.wide.s32 %rd515, %r422, 4; add.s64 %rd516, %rd514, %rd515; ld.global.f32 %f3038, [%rd516]; bra.uni $L__BB139_385; $L__BB139_498: ld.local.u32 %r583, [%rd4+248]; add.s32 %r584, %r583, -1; st.local.u32 [%rd4+248], %r584; add.s32 %r585, %r131, 1; st.local.u32 [%rd4+252], %r585; mov.u64 %rd795, %rd794; $L__BB139_499: st.local.u64 [%rd205+28], %rd795; setp.eq.s64 %p441, %rd795, 0; @%p441 bra $L__BB139_517; ld.local.f32 %f2886, [%rd7]; ld.local.f32 %f2887, [%rd7+4]; ld.local.f32 %f2888, [%rd7+8]; st.local.f32 [%rd794], %f2886; st.local.f32 [%rd794+4], %f2887; st.local.f32 [%rd794+8], %f2888; ld.local.u64 %rd659, [%rd205+28]; mov.u32 %r586, 0; st.u32 [%rd659+12], %r586; st.u32 [%rd659+16], %r586; st.u32 [%rd659+20], %r586; ld.local.u64 %rd660, [%rd205+28]; st.u32 [%rd660+36], %r586; ld.local.v2.f32 {%f2889, %f2890}, [%rd205+36]; setp.neu.ftz.f32 %p442, %f2889, 0f00000000; setp.neu.ftz.f32 %p443, %f2890, 0f00000000; or.pred %p444, %p442, %p443; ld.local.f32 %f2893, [%rd205+44]; setp.neu.ftz.f32 %p445, %f2893, 0f00000000; or.pred %p446, %p445, %p444; @%p446 bra $L__BB139_502; mov.f32 %f2895, 0f3F800000; st.local.v2.f32 [%rd205+36], {%f2895, %f2837}; st.local.u32 [%rd205+44], %r586; $L__BB139_502: mov.u32 %r588, 14; st.local.u32 [%rd205+-8], %r588; ld.local.v2.f32 {%f2896, %f2897}, [%rd205+12]; mov.f32 %f2900, 0f38D1B717; max.ftz.f32 %f2901, %f2896, %f2900; mov.f32 %f2902, 0f3F800000; max.ftz.f32 %f2903, %f2897, %f2900; min.ftz.f32 %f2904, %f2903, %f2902; min.ftz.f32 %f2905, %f2901, %f2902; st.local.v2.f32 [%rd205+12], {%f2905, %f2904}; ld.local.u64 %rd661, [%rd205+28]; ld.f32 %f2906, [%rd661]; cvt.ftz.sat.f32.f32 %f2907, %f2906; ld.f32 %f2908, [%rd661+4]; cvt.ftz.sat.f32.f32 %f2909, %f2908; ld.f32 %f2910, [%rd661+8]; cvt.ftz.sat.f32.f32 %f2911, %f2910; st.f32 [%rd661], %f2907; st.f32 [%rd661+4], %f2909; st.f32 [%rd661+8], %f2911; ld.local.u64 %rd662, [%rd205+28]; ld.f32 %f2912, [%rd662+12]; cvt.ftz.sat.f32.f32 %f2913, %f2912; ld.f32 %f2914, [%rd662+16]; cvt.ftz.sat.f32.f32 %f2915, %f2914; ld.f32 %f2916, [%rd662+20]; cvt.ftz.sat.f32.f32 %f2917, %f2916; st.f32 [%rd662+12], %f2913; st.f32 [%rd662+16], %f2915; st.f32 [%rd662+20], %f2917; ld.local.u32 %r589, [%rd4+40]; or.b32 %r590, %r589, 1036; st.local.u32 [%rd4+40], %r590; bra.uni $L__BB139_517; $L__BB139_371: mad.lo.s32 %r418, %r404, 3, %r113; ld.const.u64 %rd510, [kernel_params+2320]; cvta.to.global.u64 %rd511, %rd510; mul.wide.s32 %rd512, %r418, 4; add.s64 %rd755, %rd511, %rd512; add.s64 %rd756, %rd755, 4; add.s64 %rd757, %rd755, 8; $L__BB139_372: ld.local.f32 %f2509, [%rd4+56]; ld.global.f32 %f2510, [%rd756]; ld.local.f32 %f2511, [%rd4+60]; ld.global.f32 %f2512, [%rd757]; mul.ftz.f32 %f2513, %f2512, %f2511; fma.rn.ftz.f32 %f2514, %f2510, %f2509, %f2513; mov.f32 %f2515, 0f3F800000; sub.ftz.f32 %f2516, %f2515, %f2509; sub.ftz.f32 %f2517, %f2516, %f2511; ld.global.f32 %f2518, [%rd755]; fma.rn.ftz.f32 %f3038, %f2518, %f2517, %f2514; $L__BB139_385: mov.f32 %f2528, 0f00000000; max.ftz.f32 %f399, %f385, %f2528; max.ftz.f32 %f400, %f386, %f2528; add.ftz.f32 %f2529, %f400, %f399; max.ftz.f32 %f401, %f384, %f2528; add.ftz.f32 %f2530, %f401, %f2529; mul.ftz.f32 %f2531, %f2530, 0f3EAAAAAB; abs.ftz.f32 %f402, %f2531; setp.ltu.ftz.f32 %p340, %f402, 0f3727C5AC; mov.u64 %rd760, 0; mov.u64 %rd761, %rd760; @%p340 bra $L__BB139_390; ld.local.u32 %r118, [%rd4+252]; mov.u64 %rd760, 0; setp.eq.s32 %p341, %r118, 0; mov.u64 %rd758, %rd760; mov.u64 %rd759, %rd760; @%p341 bra $L__BB139_388; ld.local.u32 %r435, [%rd4+248]; mul.wide.s32 %rd539, %r435, 80; add.s64 %rd540, %rd1, %rd539; add.s64 %rd758, %rd540, 304; add.s64 %rd541, %rd230, %rd539; add.s64 %rd759, %rd541, 304; mov.b32 %r436, %f401; mov.b32 %r437, %f399; mov.b32 %r438, %f400; mov.u32 %r439, 0; st.local.v4.u32 [%rd540+304], {%r438, %r437, %r436, %r439}; add.s32 %r440, %r435, 1; st.local.u32 [%rd4+248], %r440; add.s32 %r441, %r118, -1; st.local.u32 [%rd4+252], %r441; $L__BB139_388: setp.eq.s64 %p342, %rd759, 0; mov.u64 %rd761, %rd760; @%p342 bra $L__BB139_390; st.local.f32 [%rd758+16], %f402; mov.u64 %rd760, %rd758; mov.u64 %rd761, %rd759; $L__BB139_390: add.s32 %r663, %r142, 4; setp.eq.s64 %p343, %rd761, 0; @%p343 bra $L__BB139_517; ld.local.u32 %r120, [%rd4+252]; setp.lt.s32 %p344, %r120, 1; @%p344 bra $L__BB139_393; add.s32 %r442, %r120, -1; st.local.u32 [%rd4+252], %r442; ld.local.s32 %rd544, [%rd4+248]; cvt.s64.s32 %rd545, %r442; add.s64 %rd546, %rd544, %rd545; mul.lo.s64 %rd547, %rd546, 80; add.s64 %rd548, %rd230, %rd547; add.s64 %rd763, %rd548, 304; add.s64 %rd549, %rd1, %rd547; add.s64 %rd762, %rd549, 304; bra.uni $L__BB139_394; $L__BB139_393: ld.local.u32 %r443, [%rd4+248]; mov.u64 %rd762, 0; add.s32 %r444, %r443, -1; st.local.u32 [%rd4+248], %r444; add.s32 %r445, %r120, 1; st.local.u32 [%rd4+252], %r445; mov.u64 %rd763, %rd762; $L__BB139_394: setp.eq.s64 %p345, %rd763, 0; @%p345 bra $L__BB139_517; setp.eq.s32 %p346, %r91, 255; @%p346 bra $L__BB139_397; bra.uni $L__BB139_396; $L__BB139_397: mov.b32 %f3039, %r382; bra.uni $L__BB139_398; $L__BB139_396: mul.wide.u32 %rd552, %r91, 4; add.s64 %rd553, %rd2, %rd552; ld.local.f32 %f3039, [%rd553]; $L__BB139_398: add.ftz.f32 %f2532, %f3038, 0fBF000000; add.ftz.f32 %f406, %f2532, %f2532; fma.rn.ftz.f32 %f2533, %f406, %f3039, 0f3F800000; mul.ftz.f32 %f407, %f60, %f2533; mul.ftz.f32 %f408, %f63, %f2533; setp.eq.s32 %p347, %r86, 255; @%p347 bra $L__BB139_400; bra.uni $L__BB139_399; $L__BB139_400: mov.b32 %f3040, %r376; bra.uni $L__BB139_401; $L__BB139_399: mul.wide.u32 %rd554, %r86, 4; add.s64 %rd555, %rd2, %rd554; ld.local.f32 %f3040, [%rd555]; $L__BB139_401: max.ftz.f32 %f2535, %f3040, %f2528; mov.f32 %f2536, 0f3F800000; min.ftz.f32 %f2537, %f2535, %f2536; sub.ftz.f32 %f412, %f2536, %f2537; add.s64 %rd153, %rd760, 20; st.local.f32 [%rd760+20], %f55; st.local.f32 [%rd760+24], %f56; st.local.f32 [%rd760+28], %f57; st.local.f32 [%rd760+44], %f407; st.local.v2.f32 [%rd760+48], {%f408, %f3036}; st.local.v2.f32 [%rd760+56], {%f3037, %f412}; st.local.u64 [%rd760+64], %rd763; shr.u32 %r446, %r144, 24; cvt.u16.u32 %rs15, %r446; setp.eq.s16 %p348, %rs15, 0; @%p348 bra $L__BB139_415; setp.eq.s16 %p349, %rs15, 1; @%p349 bra $L__BB139_405; setp.ne.s16 %p350, %rs15, 2; @%p350 bra $L__BB139_416; shr.u32 %r447, %r375, 24; mul.wide.u32 %rd556, %r447, 4; add.s64 %rd557, %rd2, %rd556; ld.local.f32 %f2538, [%rd557+8]; ld.local.f32 %f2539, [%rd557+4]; ld.local.f32 %f2540, [%rd557]; st.local.v2.f32 [%rd153+12], {%f2540, %f2539}; st.local.f32 [%rd153+20], %f2538; bra.uni $L__BB139_417; $L__BB139_415: shr.u32 %r449, %r144, 14; cvt.u64.u32 %rd567, %r449; and.b64 %rd568, %rd567, 1020; add.s64 %rd569, %rd2, %rd568; ld.local.f32 %f2581, [%rd569]; lg2.approx.ftz.f32 %f2582, %f2581; mul.ftz.f32 %f2583, %f2582, 0f3F317218; ld.local.f32 %f2584, [%rd569+4]; lg2.approx.ftz.f32 %f2585, %f2584; mul.ftz.f32 %f2586, %f2585, 0f3F317218; ld.local.f32 %f2587, [%rd569+8]; lg2.approx.ftz.f32 %f2588, %f2587; mul.ftz.f32 %f2589, %f2588, 0f3F317218; fma.rn.ftz.f32 %f2590, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2591, %f408, %f2590, 0fC12BAE14; fma.rn.ftz.f32 %f2592, %f408, %f2591, 0f40220C4A; fma.rn.ftz.f32 %f2593, %f408, %f2592, 0fBE5C28F6; fma.rn.ftz.f32 %f2594, %f408, %f2593, 0f40BF020C; div.approx.ftz.f32 %f2595, %f2583, %f2594; div.approx.ftz.f32 %f2596, %f2586, %f2594; div.approx.ftz.f32 %f2597, %f2589, %f2594; mul.ftz.f32 %f2598, %f2597, %f2597; mul.ftz.f32 %f2599, %f2596, %f2596; mul.ftz.f32 %f2600, %f2595, %f2595; st.local.v2.f32 [%rd153+12], {%f2600, %f2599}; st.local.f32 [%rd153+20], %f2598; bra.uni $L__BB139_417; $L__BB139_405: setp.eq.s32 %p351, %r87, 255; @%p351 bra $L__BB139_407; bra.uni $L__BB139_406; $L__BB139_407: mov.b32 %f3041, %r377; bra.uni $L__BB139_408; $L__BB139_416: mov.f32 %f2601, 0f3F172384; mov.f32 %f2602, 0f3E8D729D; st.local.v2.f32 [%rd153+12], {%f2602, %f2601}; mov.u32 %r450, 1069964113; st.local.u32 [%rd153+20], %r450; bra.uni $L__BB139_417; $L__BB139_406: mul.wide.u32 %rd558, %r87, 4; add.s64 %rd559, %rd2, %rd558; ld.local.f32 %f3041, [%rd559]; $L__BB139_408: setp.eq.s32 %p352, %r88, 255; @%p352 bra $L__BB139_410; bra.uni $L__BB139_409; $L__BB139_410: mov.b32 %f3042, %r378; bra.uni $L__BB139_411; $L__BB139_409: mul.wide.u32 %rd560, %r88, 4; add.s64 %rd561, %rd2, %rd560; ld.local.f32 %f3042, [%rd561]; $L__BB139_411: setp.eq.s32 %p353, %r90, 255; @%p353 bra $L__BB139_413; bra.uni $L__BB139_412; $L__BB139_413: mov.b32 %f3043, %r381; bra.uni $L__BB139_414; $L__BB139_412: mul.wide.u32 %rd562, %r90, 4; add.s64 %rd563, %rd2, %rd562; ld.local.f32 %f3043, [%rd563]; $L__BB139_414: max.ftz.f32 %f2542, %f3043, %f2528; min.ftz.f32 %f2544, %f2542, %f2536; fma.rn.ftz.f32 %f2545, %f406, %f2544, 0f3F800000; mul.ftz.f32 %f2546, %f3041, %f2545; sub.ftz.f32 %f2547, %f2536, %f2546; mov.f32 %f2548, 0f38D1B717; max.ftz.f32 %f2549, %f2547, %f2548; lg2.approx.ftz.f32 %f2550, %f2549; mul.ftz.f32 %f2551, %f2550, 0fBF317218; sub.ftz.f32 %f2552, %f2536, %f3042; mul.ftz.f32 %f2553, %f2552, %f2551; mul.ftz.f32 %f2554, %f3042, %f2551; mul.ftz.f32 %f2555, %f2554, 0f3EAF9DB2; mul.ftz.f32 %f2556, %f2554, 0f3F3BA5E3; mul.ftz.f32 %f2557, %f2554, 0f3FF645A2; fma.rn.ftz.f32 %f2558, %f2553, 0f3F018937, %f2555; fma.rn.ftz.f32 %f2559, %f2553, 0f3F574BC7, %f2556; fma.rn.ftz.f32 %f2560, %f2553, 0f3FD39581, %f2557; shl.b32 %r448, %r379, 2; cvt.u64.u32 %rd564, %r448; and.b64 %rd565, %rd564, 1020; add.s64 %rd566, %rd2, %rd565; ld.local.f32 %f2561, [%rd566]; lg2.approx.ftz.f32 %f2562, %f2561; mul.ftz.f32 %f2563, %f2562, 0f3F317218; ld.local.f32 %f2564, [%rd566+4]; lg2.approx.ftz.f32 %f2565, %f2564; mul.ftz.f32 %f2566, %f2565, 0f3F317218; ld.local.f32 %f2567, [%rd566+8]; lg2.approx.ftz.f32 %f2568, %f2567; mul.ftz.f32 %f2569, %f2568, 0f3F317218; fma.rn.ftz.f32 %f2570, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2571, %f408, %f2570, 0fC12BAE14; fma.rn.ftz.f32 %f2572, %f408, %f2571, 0f40220C4A; fma.rn.ftz.f32 %f2573, %f408, %f2572, 0fBE5C28F6; fma.rn.ftz.f32 %f2574, %f408, %f2573, 0f40BF020C; div.approx.ftz.f32 %f2575, %f2563, %f2574; div.approx.ftz.f32 %f2576, %f2566, %f2574; div.approx.ftz.f32 %f2577, %f2569, %f2574; fma.rn.ftz.f32 %f2578, %f2577, %f2577, %f2560; fma.rn.ftz.f32 %f2579, %f2576, %f2576, %f2559; fma.rn.ftz.f32 %f2580, %f2575, %f2575, %f2558; st.local.v2.f32 [%rd153+12], {%f2580, %f2579}; st.local.f32 [%rd153+20], %f2578; $L__BB139_417: mov.u32 %r451, 29; st.local.u32 [%rd153+-8], %r451; mov.f32 %f2603, 0f3A83126F; max.ftz.f32 %f2604, %f407, %f2603; min.ftz.f32 %f2606, %f2604, %f2536; max.ftz.f32 %f2607, %f408, %f2603; min.ftz.f32 %f2608, %f2607, %f2536; mul.ftz.f32 %f2609, %f2606, %f412; max.ftz.f32 %f2610, %f2609, %f2603; min.ftz.f32 %f2611, %f2610, %f2536; mul.ftz.f32 %f2612, %f2606, %f2606; mul.ftz.f32 %f2613, %f2612, 0f3F4FDF3B; fma.rn.ftz.f32 %f2614, %f2606, 0f3F39DB23, %f2613; mul.ftz.f32 %f2615, %f2612, %f2612; mul.ftz.f32 %f2616, %f2606, %f2615; mul.ftz.f32 %f2617, %f2616, %f2616; mul.ftz.f32 %f2618, %f2617, %f2617; fma.rn.ftz.f32 %f2619, %f2618, 0f406CCCCD, %f2614; mul.ftz.f32 %f2620, %f2619, %f2619; st.local.f32 [%rd153+24], %f2620; mul.ftz.f32 %f2621, %f2608, %f2608; mul.ftz.f32 %f2622, %f2621, 0f3F98D4FE; fma.rn.ftz.f32 %f2623, %f2608, 0f3E87AE14, %f2622; mul.ftz.f32 %f2624, %f2621, %f2621; mul.ftz.f32 %f2625, %f2608, %f2624; mul.ftz.f32 %f2626, %f2625, %f2625; mul.ftz.f32 %f2627, %f2608, %f2626; mul.ftz.f32 %f2628, %f2627, %f2627; fma.rn.ftz.f32 %f2629, %f2628, 0f40ABE76D, %f2623; mul.ftz.f32 %f2630, %f2629, 0f3F206C99; st.local.f32 [%rd153+28], %f2630; mul.ftz.f32 %f2631, %f2611, %f2611; mul.ftz.f32 %f2632, %f2631, 0f3F4FDF3B; fma.rn.ftz.f32 %f2633, %f2611, 0f3F39DB23, %f2632; mul.ftz.f32 %f2634, %f2631, %f2631; mul.ftz.f32 %f2635, %f2611, %f2634; mul.ftz.f32 %f2636, %f2635, %f2635; mul.ftz.f32 %f2637, %f2636, %f2636; fma.rn.ftz.f32 %f2638, %f2637, 0f406CCCCD, %f2633; mul.ftz.f32 %f2639, %f2638, %f2638; st.local.f32 [%rd153+40], %f2639; ld.local.f32 %f3044, [%rd4+104]; ld.local.f32 %f3045, [%rd4+108]; mul.ftz.f32 %f2640, %f3045, %f3045; fma.rn.ftz.f32 %f2641, %f3044, %f3044, %f2640; ld.local.f32 %f3046, [%rd4+112]; fma.rn.ftz.f32 %f2642, %f3046, %f3046, %f2641; sqrt.approx.ftz.f32 %f425, %f2642; setp.neu.ftz.f32 %p354, %f425, 0f00000000; @%p354 bra $L__BB139_418; bra.uni $L__BB139_419; $L__BB139_418: rcp.approx.ftz.f32 %f2643, %f425; mul.ftz.f32 %f3046, %f3046, %f2643; mul.ftz.f32 %f3045, %f3045, %f2643; mul.ftz.f32 %f3044, %f3044, %f2643; $L__BB139_419: ld.local.f32 %f2644, [%rd4+32]; mul.ftz.f32 %f2645, %f3045, %f2644; ld.local.f32 %f2646, [%rd4+28]; mul.ftz.f32 %f2647, %f3046, %f2646; sub.ftz.f32 %f3047, %f2645, %f2647; ld.local.f32 %f2648, [%rd4+24]; mul.ftz.f32 %f2649, %f3046, %f2648; mul.ftz.f32 %f2650, %f3044, %f2644; sub.ftz.f32 %f3048, %f2649, %f2650; mul.ftz.f32 %f2651, %f3044, %f2646; mul.ftz.f32 %f2652, %f3045, %f2648; sub.ftz.f32 %f3049, %f2651, %f2652; mul.ftz.f32 %f2653, %f3048, %f3048; fma.rn.ftz.f32 %f2654, %f3047, %f3047, %f2653; fma.rn.ftz.f32 %f2655, %f3049, %f3049, %f2654; sqrt.approx.ftz.f32 %f441, %f2655; setp.neu.ftz.f32 %p355, %f441, 0f00000000; @%p355 bra $L__BB139_420; bra.uni $L__BB139_421; $L__BB139_420: rcp.approx.ftz.f32 %f2656, %f441; mul.ftz.f32 %f3049, %f3049, %f2656; mul.ftz.f32 %f3048, %f3048, %f2656; mul.ftz.f32 %f3047, %f3047, %f2656; $L__BB139_421: mul.ftz.f32 %f2657, %f3045, %f3049; mul.ftz.f32 %f2658, %f3046, %f3048; sub.ftz.f32 %f3050, %f2657, %f2658; mul.ftz.f32 %f2659, %f3046, %f3047; mul.ftz.f32 %f2660, %f3044, %f3049; sub.ftz.f32 %f3051, %f2659, %f2660; mul.ftz.f32 %f2661, %f3044, %f3048; mul.ftz.f32 %f2662, %f3045, %f3047; sub.ftz.f32 %f3052, %f2661, %f2662; mul.ftz.f32 %f2663, %f3051, %f3051; fma.rn.ftz.f32 %f2664, %f3050, %f3050, %f2663; fma.rn.ftz.f32 %f2665, %f3052, %f3052, %f2664; sqrt.approx.ftz.f32 %f457, %f2665; setp.neu.ftz.f32 %p356, %f457, 0f00000000; @%p356 bra $L__BB139_422; bra.uni $L__BB139_423; $L__BB139_422: rcp.approx.ftz.f32 %f2666, %f457; mul.ftz.f32 %f3052, %f2666, %f3052; mul.ftz.f32 %f3051, %f2666, %f3051; mul.ftz.f32 %f3050, %f2666, %f3050; $L__BB139_423: ld.local.u8 %rs16, [%rd4+52]; and.b16 %rs17, %rs16, 4; setp.eq.s16 %p357, %rs17, 0; @%p357 bra $L__BB139_425; bra.uni $L__BB139_424; $L__BB139_425: ld.local.f32 %f2668, [%rd4+16]; mul.ftz.f32 %f2669, %f3046, %f2668; ld.local.f32 %f2670, [%rd4+20]; mul.ftz.f32 %f2671, %f3045, %f2670; sub.ftz.f32 %f2672, %f2669, %f2671; mul.ftz.f32 %f2673, %f3044, %f2670; ld.local.f32 %f2674, [%rd4+12]; mul.ftz.f32 %f2675, %f3046, %f2674; sub.ftz.f32 %f2676, %f2673, %f2675; mul.ftz.f32 %f2677, %f3045, %f2674; mul.ftz.f32 %f2678, %f3044, %f2668; sub.ftz.f32 %f2679, %f2677, %f2678; mul.ftz.f32 %f2680, %f3051, %f2676; fma.rn.ftz.f32 %f2681, %f3050, %f2672, %f2680; fma.rn.ftz.f32 %f3053, %f3052, %f2679, %f2681; bra.uni $L__BB139_426; $L__BB139_424: ld.local.f32 %f2667, [%rd4+60]; neg.ftz.f32 %f3053, %f2667; $L__BB139_426: st.local.v4.f32 [%rd762], {%f3047, %f3048, %f3049, %f3053}; ld.local.u32 %r452, [%rd4+40]; or.b32 %r453, %r452, 1036; st.local.u32 [%rd4+40], %r453; bra.uni $L__BB139_517; $L__BB139_145: setp.ltu.ftz.f32 %p94, %f203, 0f3727C5AC; mov.u64 %rd700, 0; mov.u64 %rd701, %rd700; @%p94 bra $L__BB139_150; ld.local.u32 %r37, [%rd4+252]; mov.u64 %rd700, 0; setp.eq.s32 %p95, %r37, 0; mov.u64 %rd698, %rd700; mov.u64 %rd699, %rd700; @%p95 bra $L__BB139_148; ld.local.u32 %r203, [%rd4+248]; mul.wide.s32 %rd310, %r203, 80; add.s64 %rd311, %rd1, %rd310; add.s64 %rd698, %rd311, 304; add.s64 %rd312, %rd230, %rd310; add.s64 %rd699, %rd312, 304; mov.b32 %r204, %f202; mov.b32 %r205, %f200; mov.b32 %r206, %f201; mov.u32 %r207, 0; st.local.v4.u32 [%rd311+304], {%r206, %r205, %r204, %r207}; add.s32 %r208, %r203, 1; st.local.u32 [%rd4+248], %r208; add.s32 %r209, %r37, -1; st.local.u32 [%rd4+252], %r209; $L__BB139_148: setp.eq.s64 %p96, %rd699, 0; mov.u64 %rd701, %rd700; @%p96 bra $L__BB139_150; st.local.f32 [%rd698+16], %f203; mov.u64 %rd700, %rd698; mov.u64 %rd701, %rd699; $L__BB139_150: setp.eq.s64 %p97, %rd701, 0; @%p97 bra $L__BB139_159; ld.f32 %f819, [%rd14+-12]; ld.f32 %f820, [%rd14+-8]; ld.f32 %f821, [%rd14+-4]; st.local.f32 [%rd700+20], %f819; st.local.f32 [%rd700+24], %f820; st.local.f32 [%rd700+28], %f821; mov.u32 %r211, 2; st.local.u32 [%rd700+12], %r211; mov.u32 %r643, 12; $L__BB139_159: st.u32 [%rd14+-20], %r149; setp.gt.s32 %p102, %r641, 0; @%p102 bra $L__BB139_161; bra.uni $L__BB139_160; $L__BB139_161: ld.f32 %f826, [%rd14+-28]; ld.f32 %f827, [%rd14+-32]; add.ftz.f32 %f828, %f827, %f826; ld.f32 %f829, [%rd14+-24]; add.ftz.f32 %f830, %f829, %f828; mul.ftz.f32 %f831, %f830, 0f3EAAAAAB; abs.ftz.f32 %f832, %f831; cvt.rn.f32.s32 %f833, %r641; mul.ftz.f32 %f834, %f832, %f833; st.f32 [%rd14+-16], %f834; or.b32 %r223, %r149, 2; setp.eq.s32 %p103, %r223, 34; @%p103 bra $L__BB139_172; bra.uni $L__BB139_162; $L__BB139_172: ld.f32 %f1388, [%rd14]; mul.ftz.f32 %f3007, %f1388, 0f3DA2F983; ld.f32 %f1389, [%rd14+4]; mul.ftz.f32 %f3008, %f1389, 0f3DA2F983; ld.f32 %f1390, [%rd14+8]; mul.ftz.f32 %f3009, %f1390, 0f3DA2F983; bra.uni $L__BB139_173; $L__BB139_160: mov.u32 %r222, 0; st.u32 [%rd14+-16], %r222; bra.uni $L__BB139_174; $L__BB139_162: mov.f32 %f836, 0f3F8147AE; max.ftz.f32 %f837, %f2996, %f836; mov.f32 %f838, 0f40733333; min.ftz.f32 %f839, %f837, %f838; rcp.approx.ftz.f32 %f840, %f839; mov.f32 %f841, 0f3F800000; fma.rn.ftz.f32 %f842, %f840, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f843, %f840, %f842, 0f3F2B020C; fma.rn.ftz.f32 %f844, %f839, 0f3D8240B8, %f843; add.ftz.f32 %f845, %f844, 0f3F800000; mul.ftz.f32 %f846, %f845, 0f3FAAAAAB; sub.ftz.f32 %f847, %f841, %f844; div.approx.ftz.f32 %f204, %f846, %f847; ld.f32 %f205, [%rd14+12]; setp.lt.ftz.f32 %p104, %f205, 0f38D1B717; mov.f32 %f3004, 0f00000000; @%p104 bra $L__BB139_165; setp.ge.ftz.f32 %p105, %f205, 0f3F7EB852; mov.f32 %f3004, 0f3F7FFFEF; @%p105 bra $L__BB139_165; sub.ftz.f32 %f850, %f841, 0f3F000000; mul.ftz.f32 %f851, %f850, 0f40400000; sqrt.approx.ftz.f32 %f852, %f851; mul.ftz.f32 %f853, %f204, %f852; mul.ftz.f32 %f854, %f853, 0fBFB8AA3B; ex2.approx.ftz.f32 %f855, %f854; add.ftz.f32 %f856, %f855, 0f3F800000; mul.ftz.f32 %f857, %f856, 0f3E800000; mul.ftz.f32 %f858, %f852, 0fBFB8AA3B; ex2.approx.ftz.f32 %f859, %f858; mul.ftz.f32 %f860, %f859, %f857; setp.lt.ftz.f32 %p106, %f860, %f205; selp.f32 %f861, 0f3F800000, 0f3F000000, %p106; selp.f32 %f862, 0f3F000000, 0f00000000, %p106; selp.f32 %f863, 0f3F400000, 0f3E800000, %p106; sub.ftz.f32 %f864, %f841, %f863; mul.ftz.f32 %f865, %f864, 0f40400000; sqrt.approx.ftz.f32 %f866, %f865; mul.ftz.f32 %f867, %f863, 0f3F000000; mul.ftz.f32 %f868, %f204, %f866; mul.ftz.f32 %f869, %f868, 0fBFB8AA3B; ex2.approx.ftz.f32 %f870, %f869; add.ftz.f32 %f871, %f870, 0f3F800000; mul.ftz.f32 %f872, %f871, %f867; mul.ftz.f32 %f873, %f866, 0fBFB8AA3B; ex2.approx.ftz.f32 %f874, %f873; mul.ftz.f32 %f875, %f874, %f872; setp.lt.ftz.f32 %p107, %f875, %f205; selp.f32 %f876, %f861, %f863, %p107; selp.f32 %f877, %f863, %f862, %p107; add.ftz.f32 %f878, %f876, %f877; mul.ftz.f32 %f879, %f878, 0f3F000000; sub.ftz.f32 %f880, %f841, %f879; mul.ftz.f32 %f881, %f880, 0f40400000; sqrt.approx.ftz.f32 %f882, %f881; mul.ftz.f32 %f883, %f879, 0f3F000000; mul.ftz.f32 %f884, %f204, %f882; mul.ftz.f32 %f885, %f884, 0fBFB8AA3B; ex2.approx.ftz.f32 %f886, %f885; add.ftz.f32 %f887, %f886, 0f3F800000; mul.ftz.f32 %f888, %f887, %f883; mul.ftz.f32 %f889, %f882, 0fBFB8AA3B; ex2.approx.ftz.f32 %f890, %f889; mul.ftz.f32 %f891, %f890, %f888; setp.lt.ftz.f32 %p108, %f891, %f205; selp.f32 %f892, %f876, %f879, %p108; selp.f32 %f893, %f879, %f877, %p108; add.ftz.f32 %f894, %f892, %f893; mul.ftz.f32 %f895, %f894, 0f3F000000; sub.ftz.f32 %f896, %f841, %f895; mul.ftz.f32 %f897, %f896, 0f40400000; sqrt.approx.ftz.f32 %f898, %f897; mul.ftz.f32 %f899, %f895, 0f3F000000; mul.ftz.f32 %f900, %f204, %f898; mul.ftz.f32 %f901, %f900, 0fBFB8AA3B; ex2.approx.ftz.f32 %f902, %f901; add.ftz.f32 %f903, %f902, 0f3F800000; mul.ftz.f32 %f904, %f903, %f899; mul.ftz.f32 %f905, %f898, 0fBFB8AA3B; ex2.approx.ftz.f32 %f906, %f905; mul.ftz.f32 %f907, %f906, %f904; setp.lt.ftz.f32 %p109, %f907, %f205; selp.f32 %f908, %f892, %f895, %p109; selp.f32 %f909, %f895, %f893, %p109; add.ftz.f32 %f910, %f908, %f909; mul.ftz.f32 %f911, %f910, 0f3F000000; sub.ftz.f32 %f912, %f841, %f911; mul.ftz.f32 %f913, %f912, 0f40400000; sqrt.approx.ftz.f32 %f914, %f913; mul.ftz.f32 %f915, %f911, 0f3F000000; mul.ftz.f32 %f916, %f204, %f914; mul.ftz.f32 %f917, %f916, 0fBFB8AA3B; ex2.approx.ftz.f32 %f918, %f917; add.ftz.f32 %f919, %f918, 0f3F800000; mul.ftz.f32 %f920, %f919, %f915; mul.ftz.f32 %f921, %f914, 0fBFB8AA3B; ex2.approx.ftz.f32 %f922, %f921; mul.ftz.f32 %f923, %f922, %f920; setp.lt.ftz.f32 %p110, %f923, %f205; selp.f32 %f924, %f908, %f911, %p110; selp.f32 %f925, %f911, %f909, %p110; add.ftz.f32 %f926, %f924, %f925; mul.ftz.f32 %f927, %f926, 0f3F000000; sub.ftz.f32 %f928, %f841, %f927; mul.ftz.f32 %f929, %f928, 0f40400000; sqrt.approx.ftz.f32 %f930, %f929; mul.ftz.f32 %f931, %f927, 0f3F000000; mul.ftz.f32 %f932, %f204, %f930; mul.ftz.f32 %f933, %f932, 0fBFB8AA3B; ex2.approx.ftz.f32 %f934, %f933; add.ftz.f32 %f935, %f934, 0f3F800000; mul.ftz.f32 %f936, %f935, %f931; mul.ftz.f32 %f937, %f930, 0fBFB8AA3B; ex2.approx.ftz.f32 %f938, %f937; mul.ftz.f32 %f939, %f938, %f936; setp.lt.ftz.f32 %p111, %f939, %f205; selp.f32 %f940, %f924, %f927, %p111; selp.f32 %f941, %f927, %f925, %p111; add.ftz.f32 %f942, %f940, %f941; mul.ftz.f32 %f943, %f942, 0f3F000000; sub.ftz.f32 %f944, %f841, %f943; mul.ftz.f32 %f945, %f944, 0f40400000; sqrt.approx.ftz.f32 %f946, %f945; mul.ftz.f32 %f947, %f943, 0f3F000000; mul.ftz.f32 %f948, %f204, %f946; mul.ftz.f32 %f949, %f948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f950, %f949; add.ftz.f32 %f951, %f950, 0f3F800000; mul.ftz.f32 %f952, %f951, %f947; mul.ftz.f32 %f953, %f946, 0fBFB8AA3B; ex2.approx.ftz.f32 %f954, %f953; mul.ftz.f32 %f955, %f954, %f952; setp.lt.ftz.f32 %p112, %f955, %f205; selp.f32 %f956, %f940, %f943, %p112; selp.f32 %f957, %f943, %f941, %p112; add.ftz.f32 %f958, %f956, %f957; mul.ftz.f32 %f959, %f958, 0f3F000000; sub.ftz.f32 %f960, %f841, %f959; mul.ftz.f32 %f961, %f960, 0f40400000; sqrt.approx.ftz.f32 %f962, %f961; mul.ftz.f32 %f963, %f959, 0f3F000000; mul.ftz.f32 %f964, %f204, %f962; mul.ftz.f32 %f965, %f964, 0fBFB8AA3B; ex2.approx.ftz.f32 %f966, %f965; add.ftz.f32 %f967, %f966, 0f3F800000; mul.ftz.f32 %f968, %f967, %f963; mul.ftz.f32 %f969, %f962, 0fBFB8AA3B; ex2.approx.ftz.f32 %f970, %f969; mul.ftz.f32 %f971, %f970, %f968; setp.lt.ftz.f32 %p113, %f971, %f205; selp.f32 %f972, %f956, %f959, %p113; selp.f32 %f973, %f959, %f957, %p113; add.ftz.f32 %f974, %f972, %f973; mul.ftz.f32 %f975, %f974, 0f3F000000; sub.ftz.f32 %f976, %f841, %f975; mul.ftz.f32 %f977, %f976, 0f40400000; sqrt.approx.ftz.f32 %f978, %f977; mul.ftz.f32 %f979, %f975, 0f3F000000; mul.ftz.f32 %f980, %f204, %f978; mul.ftz.f32 %f981, %f980, 0fBFB8AA3B; ex2.approx.ftz.f32 %f982, %f981; add.ftz.f32 %f983, %f982, 0f3F800000; mul.ftz.f32 %f984, %f983, %f979; mul.ftz.f32 %f985, %f978, 0fBFB8AA3B; ex2.approx.ftz.f32 %f986, %f985; mul.ftz.f32 %f987, %f986, %f984; setp.lt.ftz.f32 %p114, %f987, %f205; selp.f32 %f988, %f972, %f975, %p114; selp.f32 %f989, %f975, %f973, %p114; add.ftz.f32 %f990, %f988, %f989; mul.ftz.f32 %f991, %f990, 0f3F000000; sub.ftz.f32 %f992, %f841, %f991; mul.ftz.f32 %f993, %f992, 0f40400000; sqrt.approx.ftz.f32 %f994, %f993; mul.ftz.f32 %f995, %f991, 0f3F000000; mul.ftz.f32 %f996, %f204, %f994; mul.ftz.f32 %f997, %f996, 0fBFB8AA3B; ex2.approx.ftz.f32 %f998, %f997; add.ftz.f32 %f999, %f998, 0f3F800000; mul.ftz.f32 %f1000, %f999, %f995; mul.ftz.f32 %f1001, %f994, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1002, %f1001; mul.ftz.f32 %f1003, %f1002, %f1000; setp.lt.ftz.f32 %p115, %f1003, %f205; selp.f32 %f1004, %f988, %f991, %p115; selp.f32 %f1005, %f991, %f989, %p115; add.ftz.f32 %f1006, %f1004, %f1005; mul.ftz.f32 %f1007, %f1006, 0f3F000000; sub.ftz.f32 %f1008, %f841, %f1007; mul.ftz.f32 %f1009, %f1008, 0f40400000; sqrt.approx.ftz.f32 %f1010, %f1009; mul.ftz.f32 %f1011, %f1007, 0f3F000000; mul.ftz.f32 %f1012, %f204, %f1010; mul.ftz.f32 %f1013, %f1012, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1014, %f1013; add.ftz.f32 %f1015, %f1014, 0f3F800000; mul.ftz.f32 %f1016, %f1015, %f1011; mul.ftz.f32 %f1017, %f1010, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1018, %f1017; mul.ftz.f32 %f1019, %f1018, %f1016; setp.lt.ftz.f32 %p116, %f1019, %f205; selp.f32 %f1020, %f1004, %f1007, %p116; selp.f32 %f1021, %f1007, %f1005, %p116; add.ftz.f32 %f1022, %f1020, %f1021; mul.ftz.f32 %f3004, %f1022, 0f3F000000; $L__BB139_165: mov.f32 %f3005, 0f00000000; ld.f32 %f208, [%rd14+16]; setp.lt.ftz.f32 %p117, %f208, 0f38D1B717; @%p117 bra $L__BB139_168; setp.ge.ftz.f32 %p118, %f208, 0f3F7EB852; mov.f32 %f3005, 0f3F7FFFEF; @%p118 bra $L__BB139_168; mov.f32 %f2962, 0f3F800000; sub.ftz.f32 %f1026, %f2962, 0f3F000000; mul.ftz.f32 %f1027, %f1026, 0f40400000; sqrt.approx.ftz.f32 %f1028, %f1027; mul.ftz.f32 %f1029, %f204, %f1028; mul.ftz.f32 %f1030, %f1029, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1031, %f1030; add.ftz.f32 %f1032, %f1031, 0f3F800000; mul.ftz.f32 %f1033, %f1032, 0f3E800000; mul.ftz.f32 %f1034, %f1028, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1035, %f1034; mul.ftz.f32 %f1036, %f1035, %f1033; setp.lt.ftz.f32 %p119, %f1036, %f208; selp.f32 %f1037, 0f3F800000, 0f3F000000, %p119; selp.f32 %f1038, 0f3F000000, 0f00000000, %p119; selp.f32 %f1039, 0f3F400000, 0f3E800000, %p119; sub.ftz.f32 %f1040, %f2962, %f1039; mul.ftz.f32 %f1041, %f1040, 0f40400000; sqrt.approx.ftz.f32 %f1042, %f1041; mul.ftz.f32 %f1043, %f1039, 0f3F000000; mul.ftz.f32 %f1044, %f204, %f1042; mul.ftz.f32 %f1045, %f1044, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1046, %f1045; add.ftz.f32 %f1047, %f1046, 0f3F800000; mul.ftz.f32 %f1048, %f1047, %f1043; mul.ftz.f32 %f1049, %f1042, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1050, %f1049; mul.ftz.f32 %f1051, %f1050, %f1048; setp.lt.ftz.f32 %p120, %f1051, %f208; selp.f32 %f1052, %f1037, %f1039, %p120; selp.f32 %f1053, %f1039, %f1038, %p120; add.ftz.f32 %f1054, %f1052, %f1053; mul.ftz.f32 %f1055, %f1054, 0f3F000000; sub.ftz.f32 %f1056, %f2962, %f1055; mul.ftz.f32 %f1057, %f1056, 0f40400000; sqrt.approx.ftz.f32 %f1058, %f1057; mul.ftz.f32 %f1059, %f1055, 0f3F000000; mul.ftz.f32 %f1060, %f204, %f1058; mul.ftz.f32 %f1061, %f1060, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1062, %f1061; add.ftz.f32 %f1063, %f1062, 0f3F800000; mul.ftz.f32 %f1064, %f1063, %f1059; mul.ftz.f32 %f1065, %f1058, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1066, %f1065; mul.ftz.f32 %f1067, %f1066, %f1064; setp.lt.ftz.f32 %p121, %f1067, %f208; selp.f32 %f1068, %f1052, %f1055, %p121; selp.f32 %f1069, %f1055, %f1053, %p121; add.ftz.f32 %f1070, %f1068, %f1069; mul.ftz.f32 %f1071, %f1070, 0f3F000000; sub.ftz.f32 %f1072, %f2962, %f1071; mul.ftz.f32 %f1073, %f1072, 0f40400000; sqrt.approx.ftz.f32 %f1074, %f1073; mul.ftz.f32 %f1075, %f1071, 0f3F000000; mul.ftz.f32 %f1076, %f204, %f1074; mul.ftz.f32 %f1077, %f1076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1078, %f1077; add.ftz.f32 %f1079, %f1078, 0f3F800000; mul.ftz.f32 %f1080, %f1079, %f1075; mul.ftz.f32 %f1081, %f1074, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1082, %f1081; mul.ftz.f32 %f1083, %f1082, %f1080; setp.lt.ftz.f32 %p122, %f1083, %f208; selp.f32 %f1084, %f1068, %f1071, %p122; selp.f32 %f1085, %f1071, %f1069, %p122; add.ftz.f32 %f1086, %f1084, %f1085; mul.ftz.f32 %f1087, %f1086, 0f3F000000; sub.ftz.f32 %f1088, %f2962, %f1087; mul.ftz.f32 %f1089, %f1088, 0f40400000; sqrt.approx.ftz.f32 %f1090, %f1089; mul.ftz.f32 %f1091, %f1087, 0f3F000000; mul.ftz.f32 %f1092, %f204, %f1090; mul.ftz.f32 %f1093, %f1092, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1094, %f1093; add.ftz.f32 %f1095, %f1094, 0f3F800000; mul.ftz.f32 %f1096, %f1095, %f1091; mul.ftz.f32 %f1097, %f1090, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1098, %f1097; mul.ftz.f32 %f1099, %f1098, %f1096; setp.lt.ftz.f32 %p123, %f1099, %f208; selp.f32 %f1100, %f1084, %f1087, %p123; selp.f32 %f1101, %f1087, %f1085, %p123; add.ftz.f32 %f1102, %f1100, %f1101; mul.ftz.f32 %f1103, %f1102, 0f3F000000; sub.ftz.f32 %f1104, %f2962, %f1103; mul.ftz.f32 %f1105, %f1104, 0f40400000; sqrt.approx.ftz.f32 %f1106, %f1105; mul.ftz.f32 %f1107, %f1103, 0f3F000000; mul.ftz.f32 %f1108, %f204, %f1106; mul.ftz.f32 %f1109, %f1108, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1110, %f1109; add.ftz.f32 %f1111, %f1110, 0f3F800000; mul.ftz.f32 %f1112, %f1111, %f1107; mul.ftz.f32 %f1113, %f1106, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1114, %f1113; mul.ftz.f32 %f1115, %f1114, %f1112; setp.lt.ftz.f32 %p124, %f1115, %f208; selp.f32 %f1116, %f1100, %f1103, %p124; selp.f32 %f1117, %f1103, %f1101, %p124; add.ftz.f32 %f1118, %f1116, %f1117; mul.ftz.f32 %f1119, %f1118, 0f3F000000; sub.ftz.f32 %f1120, %f2962, %f1119; mul.ftz.f32 %f1121, %f1120, 0f40400000; sqrt.approx.ftz.f32 %f1122, %f1121; mul.ftz.f32 %f1123, %f1119, 0f3F000000; mul.ftz.f32 %f1124, %f204, %f1122; mul.ftz.f32 %f1125, %f1124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1126, %f1125; add.ftz.f32 %f1127, %f1126, 0f3F800000; mul.ftz.f32 %f1128, %f1127, %f1123; mul.ftz.f32 %f1129, %f1122, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1130, %f1129; mul.ftz.f32 %f1131, %f1130, %f1128; setp.lt.ftz.f32 %p125, %f1131, %f208; selp.f32 %f1132, %f1116, %f1119, %p125; selp.f32 %f1133, %f1119, %f1117, %p125; add.ftz.f32 %f1134, %f1132, %f1133; mul.ftz.f32 %f1135, %f1134, 0f3F000000; sub.ftz.f32 %f1136, %f2962, %f1135; mul.ftz.f32 %f1137, %f1136, 0f40400000; sqrt.approx.ftz.f32 %f1138, %f1137; mul.ftz.f32 %f1139, %f1135, 0f3F000000; mul.ftz.f32 %f1140, %f204, %f1138; mul.ftz.f32 %f1141, %f1140, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1142, %f1141; add.ftz.f32 %f1143, %f1142, 0f3F800000; mul.ftz.f32 %f1144, %f1143, %f1139; mul.ftz.f32 %f1145, %f1138, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1146, %f1145; mul.ftz.f32 %f1147, %f1146, %f1144; setp.lt.ftz.f32 %p126, %f1147, %f208; selp.f32 %f1148, %f1132, %f1135, %p126; selp.f32 %f1149, %f1135, %f1133, %p126; add.ftz.f32 %f1150, %f1148, %f1149; mul.ftz.f32 %f1151, %f1150, 0f3F000000; sub.ftz.f32 %f1152, %f2962, %f1151; mul.ftz.f32 %f1153, %f1152, 0f40400000; sqrt.approx.ftz.f32 %f1154, %f1153; mul.ftz.f32 %f1155, %f1151, 0f3F000000; mul.ftz.f32 %f1156, %f204, %f1154; mul.ftz.f32 %f1157, %f1156, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1158, %f1157; add.ftz.f32 %f1159, %f1158, 0f3F800000; mul.ftz.f32 %f1160, %f1159, %f1155; mul.ftz.f32 %f1161, %f1154, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1162, %f1161; mul.ftz.f32 %f1163, %f1162, %f1160; setp.lt.ftz.f32 %p127, %f1163, %f208; selp.f32 %f1164, %f1148, %f1151, %p127; selp.f32 %f1165, %f1151, %f1149, %p127; add.ftz.f32 %f1166, %f1164, %f1165; mul.ftz.f32 %f1167, %f1166, 0f3F000000; sub.ftz.f32 %f1168, %f2962, %f1167; mul.ftz.f32 %f1169, %f1168, 0f40400000; sqrt.approx.ftz.f32 %f1170, %f1169; mul.ftz.f32 %f1171, %f1167, 0f3F000000; mul.ftz.f32 %f1172, %f204, %f1170; mul.ftz.f32 %f1173, %f1172, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1174, %f1173; add.ftz.f32 %f1175, %f1174, 0f3F800000; mul.ftz.f32 %f1176, %f1175, %f1171; mul.ftz.f32 %f1177, %f1170, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1178, %f1177; mul.ftz.f32 %f1179, %f1178, %f1176; setp.lt.ftz.f32 %p128, %f1179, %f208; selp.f32 %f1180, %f1164, %f1167, %p128; selp.f32 %f1181, %f1167, %f1165, %p128; add.ftz.f32 %f1182, %f1180, %f1181; mul.ftz.f32 %f1183, %f1182, 0f3F000000; sub.ftz.f32 %f1184, %f2962, %f1183; mul.ftz.f32 %f1185, %f1184, 0f40400000; sqrt.approx.ftz.f32 %f1186, %f1185; mul.ftz.f32 %f1187, %f1183, 0f3F000000; mul.ftz.f32 %f1188, %f204, %f1186; mul.ftz.f32 %f1189, %f1188, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1190, %f1189; add.ftz.f32 %f1191, %f1190, 0f3F800000; mul.ftz.f32 %f1192, %f1191, %f1187; mul.ftz.f32 %f1193, %f1186, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1194, %f1193; mul.ftz.f32 %f1195, %f1194, %f1192; setp.lt.ftz.f32 %p129, %f1195, %f208; selp.f32 %f1196, %f1180, %f1183, %p129; selp.f32 %f1197, %f1183, %f1181, %p129; add.ftz.f32 %f1198, %f1196, %f1197; mul.ftz.f32 %f3005, %f1198, 0f3F000000; $L__BB139_168: mov.f32 %f3006, 0f00000000; ld.f32 %f211, [%rd14+20]; setp.lt.ftz.f32 %p130, %f211, 0f38D1B717; @%p130 bra $L__BB139_171; setp.ge.ftz.f32 %p131, %f211, 0f3F7EB852; mov.f32 %f3006, 0f3F7FFFEF; @%p131 bra $L__BB139_171; mov.f32 %f2957, 0f3F800000; sub.ftz.f32 %f1202, %f2957, 0f3F000000; mul.ftz.f32 %f1203, %f1202, 0f40400000; sqrt.approx.ftz.f32 %f1204, %f1203; mul.ftz.f32 %f1205, %f204, %f1204; mul.ftz.f32 %f1206, %f1205, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1207, %f1206; add.ftz.f32 %f1208, %f1207, 0f3F800000; mul.ftz.f32 %f1209, %f1208, 0f3E800000; mul.ftz.f32 %f1210, %f1204, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1211, %f1210; mul.ftz.f32 %f1212, %f1211, %f1209; setp.lt.ftz.f32 %p132, %f1212, %f211; selp.f32 %f1213, 0f3F800000, 0f3F000000, %p132; selp.f32 %f1214, 0f3F000000, 0f00000000, %p132; selp.f32 %f1215, 0f3F400000, 0f3E800000, %p132; sub.ftz.f32 %f1216, %f2957, %f1215; mul.ftz.f32 %f1217, %f1216, 0f40400000; sqrt.approx.ftz.f32 %f1218, %f1217; mul.ftz.f32 %f1219, %f1215, 0f3F000000; mul.ftz.f32 %f1220, %f204, %f1218; mul.ftz.f32 %f1221, %f1220, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1222, %f1221; add.ftz.f32 %f1223, %f1222, 0f3F800000; mul.ftz.f32 %f1224, %f1223, %f1219; mul.ftz.f32 %f1225, %f1218, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1226, %f1225; mul.ftz.f32 %f1227, %f1226, %f1224; setp.lt.ftz.f32 %p133, %f1227, %f211; selp.f32 %f1228, %f1213, %f1215, %p133; selp.f32 %f1229, %f1215, %f1214, %p133; add.ftz.f32 %f1230, %f1228, %f1229; mul.ftz.f32 %f1231, %f1230, 0f3F000000; sub.ftz.f32 %f1232, %f2957, %f1231; mul.ftz.f32 %f1233, %f1232, 0f40400000; sqrt.approx.ftz.f32 %f1234, %f1233; mul.ftz.f32 %f1235, %f1231, 0f3F000000; mul.ftz.f32 %f1236, %f204, %f1234; mul.ftz.f32 %f1237, %f1236, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1238, %f1237; add.ftz.f32 %f1239, %f1238, 0f3F800000; mul.ftz.f32 %f1240, %f1239, %f1235; mul.ftz.f32 %f1241, %f1234, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1242, %f1241; mul.ftz.f32 %f1243, %f1242, %f1240; setp.lt.ftz.f32 %p134, %f1243, %f211; selp.f32 %f1244, %f1228, %f1231, %p134; selp.f32 %f1245, %f1231, %f1229, %p134; add.ftz.f32 %f1246, %f1244, %f1245; mul.ftz.f32 %f1247, %f1246, 0f3F000000; sub.ftz.f32 %f1248, %f2957, %f1247; mul.ftz.f32 %f1249, %f1248, 0f40400000; sqrt.approx.ftz.f32 %f1250, %f1249; mul.ftz.f32 %f1251, %f1247, 0f3F000000; mul.ftz.f32 %f1252, %f204, %f1250; mul.ftz.f32 %f1253, %f1252, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1254, %f1253; add.ftz.f32 %f1255, %f1254, 0f3F800000; mul.ftz.f32 %f1256, %f1255, %f1251; mul.ftz.f32 %f1257, %f1250, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1258, %f1257; mul.ftz.f32 %f1259, %f1258, %f1256; setp.lt.ftz.f32 %p135, %f1259, %f211; selp.f32 %f1260, %f1244, %f1247, %p135; selp.f32 %f1261, %f1247, %f1245, %p135; add.ftz.f32 %f1262, %f1260, %f1261; mul.ftz.f32 %f1263, %f1262, 0f3F000000; sub.ftz.f32 %f1264, %f2957, %f1263; mul.ftz.f32 %f1265, %f1264, 0f40400000; sqrt.approx.ftz.f32 %f1266, %f1265; mul.ftz.f32 %f1267, %f1263, 0f3F000000; mul.ftz.f32 %f1268, %f204, %f1266; mul.ftz.f32 %f1269, %f1268, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1270, %f1269; add.ftz.f32 %f1271, %f1270, 0f3F800000; mul.ftz.f32 %f1272, %f1271, %f1267; mul.ftz.f32 %f1273, %f1266, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1274, %f1273; mul.ftz.f32 %f1275, %f1274, %f1272; setp.lt.ftz.f32 %p136, %f1275, %f211; selp.f32 %f1276, %f1260, %f1263, %p136; selp.f32 %f1277, %f1263, %f1261, %p136; add.ftz.f32 %f1278, %f1276, %f1277; mul.ftz.f32 %f1279, %f1278, 0f3F000000; sub.ftz.f32 %f1280, %f2957, %f1279; mul.ftz.f32 %f1281, %f1280, 0f40400000; sqrt.approx.ftz.f32 %f1282, %f1281; mul.ftz.f32 %f1283, %f1279, 0f3F000000; mul.ftz.f32 %f1284, %f204, %f1282; mul.ftz.f32 %f1285, %f1284, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1286, %f1285; add.ftz.f32 %f1287, %f1286, 0f3F800000; mul.ftz.f32 %f1288, %f1287, %f1283; mul.ftz.f32 %f1289, %f1282, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1290, %f1289; mul.ftz.f32 %f1291, %f1290, %f1288; setp.lt.ftz.f32 %p137, %f1291, %f211; selp.f32 %f1292, %f1276, %f1279, %p137; selp.f32 %f1293, %f1279, %f1277, %p137; add.ftz.f32 %f1294, %f1292, %f1293; mul.ftz.f32 %f1295, %f1294, 0f3F000000; sub.ftz.f32 %f1296, %f2957, %f1295; mul.ftz.f32 %f1297, %f1296, 0f40400000; sqrt.approx.ftz.f32 %f1298, %f1297; mul.ftz.f32 %f1299, %f1295, 0f3F000000; mul.ftz.f32 %f1300, %f204, %f1298; mul.ftz.f32 %f1301, %f1300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1302, %f1301; add.ftz.f32 %f1303, %f1302, 0f3F800000; mul.ftz.f32 %f1304, %f1303, %f1299; mul.ftz.f32 %f1305, %f1298, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1306, %f1305; mul.ftz.f32 %f1307, %f1306, %f1304; setp.lt.ftz.f32 %p138, %f1307, %f211; selp.f32 %f1308, %f1292, %f1295, %p138; selp.f32 %f1309, %f1295, %f1293, %p138; add.ftz.f32 %f1310, %f1308, %f1309; mul.ftz.f32 %f1311, %f1310, 0f3F000000; sub.ftz.f32 %f1312, %f2957, %f1311; mul.ftz.f32 %f1313, %f1312, 0f40400000; sqrt.approx.ftz.f32 %f1314, %f1313; mul.ftz.f32 %f1315, %f1311, 0f3F000000; mul.ftz.f32 %f1316, %f204, %f1314; mul.ftz.f32 %f1317, %f1316, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1318, %f1317; add.ftz.f32 %f1319, %f1318, 0f3F800000; mul.ftz.f32 %f1320, %f1319, %f1315; mul.ftz.f32 %f1321, %f1314, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1322, %f1321; mul.ftz.f32 %f1323, %f1322, %f1320; setp.lt.ftz.f32 %p139, %f1323, %f211; selp.f32 %f1324, %f1308, %f1311, %p139; selp.f32 %f1325, %f1311, %f1309, %p139; add.ftz.f32 %f1326, %f1324, %f1325; mul.ftz.f32 %f1327, %f1326, 0f3F000000; sub.ftz.f32 %f1328, %f2957, %f1327; mul.ftz.f32 %f1329, %f1328, 0f40400000; sqrt.approx.ftz.f32 %f1330, %f1329; mul.ftz.f32 %f1331, %f1327, 0f3F000000; mul.ftz.f32 %f1332, %f204, %f1330; mul.ftz.f32 %f1333, %f1332, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1334, %f1333; add.ftz.f32 %f1335, %f1334, 0f3F800000; mul.ftz.f32 %f1336, %f1335, %f1331; mul.ftz.f32 %f1337, %f1330, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1338, %f1337; mul.ftz.f32 %f1339, %f1338, %f1336; setp.lt.ftz.f32 %p140, %f1339, %f211; selp.f32 %f1340, %f1324, %f1327, %p140; selp.f32 %f1341, %f1327, %f1325, %p140; add.ftz.f32 %f1342, %f1340, %f1341; mul.ftz.f32 %f1343, %f1342, 0f3F000000; sub.ftz.f32 %f1344, %f2957, %f1343; mul.ftz.f32 %f1345, %f1344, 0f40400000; sqrt.approx.ftz.f32 %f1346, %f1345; mul.ftz.f32 %f1347, %f1343, 0f3F000000; mul.ftz.f32 %f1348, %f204, %f1346; mul.ftz.f32 %f1349, %f1348, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1350, %f1349; add.ftz.f32 %f1351, %f1350, 0f3F800000; mul.ftz.f32 %f1352, %f1351, %f1347; mul.ftz.f32 %f1353, %f1346, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1354, %f1353; mul.ftz.f32 %f1355, %f1354, %f1352; setp.lt.ftz.f32 %p141, %f1355, %f211; selp.f32 %f1356, %f1340, %f1343, %p141; selp.f32 %f1357, %f1343, %f1341, %p141; add.ftz.f32 %f1358, %f1356, %f1357; mul.ftz.f32 %f1359, %f1358, 0f3F000000; sub.ftz.f32 %f1360, %f2957, %f1359; mul.ftz.f32 %f1361, %f1360, 0f40400000; sqrt.approx.ftz.f32 %f1362, %f1361; mul.ftz.f32 %f1363, %f1359, 0f3F000000; mul.ftz.f32 %f1364, %f204, %f1362; mul.ftz.f32 %f1365, %f1364, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1366, %f1365; add.ftz.f32 %f1367, %f1366, 0f3F800000; mul.ftz.f32 %f1368, %f1367, %f1363; mul.ftz.f32 %f1369, %f1362, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f1370, %f1368; setp.lt.ftz.f32 %p142, %f1371, %f211; selp.f32 %f1372, %f1356, %f1359, %p142; selp.f32 %f1373, %f1359, %f1357, %p142; add.ftz.f32 %f1374, %f1372, %f1373; mul.ftz.f32 %f3006, %f1374, 0f3F000000; $L__BB139_171: mov.f32 %f2958, 0f3F800000; sub.ftz.f32 %f1376, %f2958, %f3004; mul.ftz.f32 %f1377, %f1376, 0f40400000; sub.ftz.f32 %f1378, %f2958, %f3005; mul.ftz.f32 %f1379, %f1378, 0f40400000; sub.ftz.f32 %f1380, %f2958, %f3006; mul.ftz.f32 %f1381, %f1380, 0f40400000; sqrt.approx.ftz.f32 %f1382, %f1377; sqrt.approx.ftz.f32 %f1383, %f1379; sqrt.approx.ftz.f32 %f1384, %f1381; ld.f32 %f1385, [%rd14]; mul.ftz.f32 %f3007, %f1382, %f1385; ld.f32 %f1386, [%rd14+4]; mul.ftz.f32 %f3008, %f1383, %f1386; ld.f32 %f1387, [%rd14+8]; mul.ftz.f32 %f3009, %f1384, %f1387; $L__BB139_173: st.f32 [%rd14], %f3007; st.f32 [%rd14+4], %f3008; st.f32 [%rd14+8], %f3009; or.b32 %r643, %r643, 16; $L__BB139_174: ld.local.u32 %r224, [%rd4+40]; or.b32 %r225, %r224, %r643; st.local.u32 [%rd4+40], %r225; $L__BB139_182: setp.leu.ftz.f32 %p147, %f97, 0f3727C5AC; setp.leu.ftz.f32 %p148, %f74, 0f3727C5AC; or.pred %p149, %p148, %p147; @%p149 bra $L__BB139_194; ld.const.v4.f32 {%f1404, %f1405, %f1406, %f1407}, [kernel_params+1184]; mul.ftz.f32 %f1411, %f187, %f1405; fma.rn.ftz.f32 %f1412, %f186, %f1404, %f1411; fma.rn.ftz.f32 %f227, %f188, %f1406, %f1412; setp.leu.ftz.f32 %p150, %f227, 0f00000000; mov.f32 %f1403, 0f3F800000; mov.f32 %f3010, %f1403; mov.f32 %f3011, %f1403; mov.f32 %f3012, %f1403; @%p150 bra $L__BB139_185; div.approx.ftz.f32 %f3012, %f186, %f227; div.approx.ftz.f32 %f3011, %f187, %f227; div.approx.ftz.f32 %f3010, %f188, %f227; $L__BB139_185: sub.ftz.f32 %f1414, %f1403, %f75; fma.rn.ftz.f32 %f1415, %f75, %f3012, %f1414; fma.rn.ftz.f32 %f1416, %f75, %f3011, %f1414; fma.rn.ftz.f32 %f1417, %f75, %f3010, %f1414; mul.ftz.f32 %f1418, %f74, %f177; mul.ftz.f32 %f1419, %f1418, %f1415; mul.ftz.f32 %f1420, %f74, %f178; mul.ftz.f32 %f1421, %f1420, %f1416; mul.ftz.f32 %f1422, %f74, %f179; mul.ftz.f32 %f1423, %f1422, %f1417; mul.ftz.f32 %f1424, %f97, %f1419; mul.ftz.f32 %f1425, %f97, %f1421; mul.ftz.f32 %f1426, %f97, %f1423; mov.f32 %f3013, 0f00000000; max.ftz.f32 %f234, %f1424, %f3013; max.ftz.f32 %f235, %f1425, %f3013; max.ftz.f32 %f236, %f1426, %f3013; add.ftz.f32 %f1428, %f234, %f235; add.ftz.f32 %f1429, %f236, %f1428; mul.ftz.f32 %f1430, %f1429, 0f3EAAAAAB; abs.ftz.f32 %f237, %f1430; setp.ltu.ftz.f32 %p151, %f237, 0f3727C5AC; mov.u64 %rd712, 0; mov.u64 %rd713, %rd712; @%p151 bra $L__BB139_190; ld.local.u32 %r43, [%rd4+252]; mov.u64 %rd712, 0; setp.eq.s32 %p152, %r43, 0; mov.u64 %rd710, %rd712; mov.u64 %rd711, %rd712; @%p152 bra $L__BB139_188; ld.local.u32 %r237, [%rd4+248]; mul.wide.s32 %rd337, %r237, 80; add.s64 %rd338, %rd1, %rd337; add.s64 %rd710, %rd338, 304; add.s64 %rd339, %rd230, %rd337; add.s64 %rd711, %rd339, 304; mov.b32 %r238, %f236; mov.b32 %r239, %f235; mov.b32 %r240, %f234; mov.u32 %r241, 0; st.local.v4.u32 [%rd338+304], {%r240, %r239, %r238, %r241}; add.s32 %r242, %r237, 1; st.local.u32 [%rd4+248], %r242; add.s32 %r243, %r43, -1; st.local.u32 [%rd4+252], %r243; $L__BB139_188: setp.eq.s64 %p153, %rd711, 0; mov.u64 %rd713, %rd712; @%p153 bra $L__BB139_190; st.local.f32 [%rd710+16], %f237; mov.u64 %rd712, %rd710; mov.u64 %rd713, %rd711; $L__BB139_190: setp.eq.s64 %p154, %rd713, 0; @%p154 bra $L__BB139_194; add.s64 %rd45, %rd712, 20; st.local.f32 [%rd712+20], %f55; st.local.f32 [%rd712+24], %f56; st.local.f32 [%rd712+28], %f57; mov.u32 %r244, 6; st.local.u32 [%rd712+12], %r244; ld.local.f32 %f1432, [%rd4+24]; ld.local.f32 %f1433, [%rd4+28]; mul.ftz.f32 %f1434, %f56, %f1433; fma.rn.ftz.f32 %f1435, %f55, %f1432, %f1434; ld.local.f32 %f1436, [%rd4+32]; fma.rn.ftz.f32 %f238, %f57, %f1436, %f1435; setp.lt.ftz.f32 %p155, %f238, 0f00000000; @%p155 bra $L__BB139_193; mov.f32 %f1437, 0f3F800000; sub.ftz.f32 %f1438, %f1437, %f238; mov.f32 %f1439, 0f00000000; max.ftz.f32 %f1440, %f1438, %f1439; min.ftz.f32 %f1441, %f1440, %f1437; mul.ftz.f32 %f1442, %f1441, %f1441; mul.ftz.f32 %f1443, %f1442, %f1442; mul.ftz.f32 %f1444, %f1441, %f1443; mul.ftz.f32 %f3013, %f238, %f1444; $L__BB139_193: st.local.f32 [%rd45+12], %f3013; ld.local.f32 %f1445, [%rd45+-4]; mul.ftz.f32 %f1446, %f3013, %f1445; st.local.f32 [%rd45+-4], %f1446; ld.local.u32 %r245, [%rd4+40]; or.b32 %r246, %r245, 12; st.local.u32 [%rd4+40], %r246; $L__BB139_194: ld.param.u32 %r637, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.const.u32 %r44, [kernel_params+1564]; setp.ne.s32 %p156, %r44, 0; and.b32 %r45, %r637, 8; setp.eq.s32 %p157, %r45, 0; or.pred %p5, %p157, %p156; not.pred %p158, %p5; setp.leu.ftz.f32 %p159, %f99, 0f3727C5AC; or.pred %p160, %p159, %p158; @%p160 bra $L__BB139_225; setp.leu.ftz.f32 %p161, %f70, 0f3727C5AC; setp.leu.ftz.f32 %p162, %f60, 0f3727C5AC; and.pred %p163, %p162, %p161; @%p163 bra $L__BB139_225; mul.ftz.f32 %f1447, %f99, %f177; mov.f32 %f1448, 0f00000000; max.ftz.f32 %f241, %f1447, %f1448; mul.ftz.f32 %f1449, %f99, %f178; max.ftz.f32 %f242, %f1449, %f1448; mul.ftz.f32 %f1450, %f99, %f179; max.ftz.f32 %f243, %f1450, %f1448; add.ftz.f32 %f1451, %f241, %f242; add.ftz.f32 %f1452, %f243, %f1451; mul.ftz.f32 %f1453, %f1452, 0f3EAAAAAB; abs.ftz.f32 %f244, %f1453; setp.ltu.ftz.f32 %p164, %f244, 0f3727C5AC; mov.u64 %rd720, 0; mov.u64 %rd716, %rd720; mov.u64 %rd717, %rd720; @%p164 bra $L__BB139_201; ld.local.u32 %r46, [%rd4+252]; mov.u64 %rd716, 0; setp.eq.s32 %p165, %r46, 0; mov.u64 %rd714, %rd716; mov.u64 %rd715, %rd716; @%p165 bra $L__BB139_199; ld.local.u32 %r247, [%rd4+248]; mul.wide.s32 %rd346, %r247, 80; add.s64 %rd347, %rd1, %rd346; add.s64 %rd714, %rd347, 304; add.s64 %rd348, %rd230, %rd346; add.s64 %rd715, %rd348, 304; mov.b32 %r248, %f243; mov.b32 %r249, %f242; mov.b32 %r250, %f241; mov.u32 %r251, 0; st.local.v4.u32 [%rd347+304], {%r250, %r249, %r248, %r251}; add.s32 %r252, %r247, 1; st.local.u32 [%rd4+248], %r252; add.s32 %r253, %r46, -1; st.local.u32 [%rd4+252], %r253; $L__BB139_199: setp.eq.s64 %p166, %rd715, 0; mov.u64 %rd717, %rd716; @%p166 bra $L__BB139_201; st.local.f32 [%rd714+16], %f244; mov.u64 %rd716, %rd714; mov.u64 %rd717, %rd715; $L__BB139_201: setp.eq.s64 %p167, %rd717, 0; mov.u64 %rd721, %rd720; @%p167 bra $L__BB139_205; ld.local.u32 %r47, [%rd4+252]; setp.lt.s32 %p168, %r47, 1; @%p168 bra $L__BB139_204; add.s32 %r254, %r47, -1; st.local.u32 [%rd4+252], %r254; ld.local.s32 %rd353, [%rd4+248]; cvt.s64.s32 %rd354, %r254; add.s64 %rd355, %rd353, %rd354; mul.lo.s64 %rd356, %rd355, 80; add.s64 %rd357, %rd1, %rd356; add.s64 %rd720, %rd357, 304; add.s64 %rd358, %rd230, %rd356; add.s64 %rd721, %rd358, 304; bra.uni $L__BB139_205; $L__BB139_204: ld.local.u32 %r255, [%rd4+248]; mov.u64 %rd720, 0; add.s32 %r256, %r255, -1; st.local.u32 [%rd4+248], %r256; add.s32 %r257, %r47, 1; st.local.u32 [%rd4+252], %r257; mov.u64 %rd721, %rd720; $L__BB139_205: setp.eq.s64 %p169, %rd721, 0; or.pred %p171, %p167, %p169; @%p171 bra $L__BB139_225; add.s64 %rd60, %rd716, 20; st.local.f32 [%rd716+20], %f55; st.local.f32 [%rd716+24], %f56; st.local.f32 [%rd716+28], %f57; mul.ftz.f32 %f245, %f70, 0f3DA3D70A; max.ftz.f32 %f1458, %f245, %f1448; sqrt.approx.ftz.f32 %f1459, %f1458; mov.f32 %f1456, 0f3F800000; sub.ftz.f32 %f1460, %f1456, %f1459; mov.f32 %f1461, 0f40000000; div.approx.ftz.f32 %f1462, %f1461, %f1460; add.ftz.f32 %f1463, %f1462, 0fBF800000; st.local.f32 [%rd716+40], %f1463; st.local.v2.f32 [%rd716+56], {%f2977, %f2978}; st.local.f32 [%rd716+64], %f2979; st.local.u64 [%rd716+48], %rd721; fma.rn.ftz.f32 %f1464, %f73, 0fBF666666, 0f3F800000; max.ftz.f32 %f1465, %f1464, %f1448; sqrt.approx.ftz.f32 %f1466, %f1465; mul.ftz.f32 %f1467, %f71, %f71; mul.ftz.f32 %f1468, %f1467, %f1466; div.approx.ftz.f32 %f1469, %f1467, %f1466; st.local.v2.f32 [%rd716+32], {%f1469, %f1468}; mul.ftz.f32 %f1470, %f187, 0f3F19999A; fma.rn.ftz.f32 %f1471, %f186, 0f3E99999A, %f1470; fma.rn.ftz.f32 %f246, %f188, 0f3DCCCCCD, %f1471; setp.leu.ftz.f32 %p172, %f246, 0f00000000; mov.f32 %f3014, %f1456; mov.f32 %f3015, %f1456; mov.f32 %f3016, %f1456; @%p172 bra $L__BB139_208; div.approx.ftz.f32 %f3016, %f186, %f246; div.approx.ftz.f32 %f3015, %f187, %f246; div.approx.ftz.f32 %f3014, %f188, %f246; $L__BB139_208: sub.ftz.f32 %f1473, %f1456, %f72; fma.rn.ftz.f32 %f1474, %f72, %f3016, %f1473; fma.rn.ftz.f32 %f1475, %f72, %f3015, %f1473; fma.rn.ftz.f32 %f1476, %f72, %f3014, %f1473; mul.ftz.f32 %f1477, %f245, %f1474; mul.ftz.f32 %f1478, %f245, %f1475; mul.ftz.f32 %f1479, %f245, %f1476; sub.ftz.f32 %f1480, %f1456, %f60; mul.ftz.f32 %f1481, %f1480, %f1477; mul.ftz.f32 %f1482, %f1480, %f1478; mul.ftz.f32 %f1483, %f1480, %f1479; fma.rn.ftz.f32 %f1484, %f60, %f186, %f1481; fma.rn.ftz.f32 %f1485, %f60, %f187, %f1482; fma.rn.ftz.f32 %f1486, %f60, %f188, %f1483; st.local.f32 [%rd720+12], %f1484; st.local.f32 [%rd720+16], %f1485; st.local.f32 [%rd720+20], %f1486; ld.local.u64 %rd361, [%rd60+28]; st.f32 [%rd361], %f186; st.f32 [%rd361+4], %f187; st.f32 [%rd361+8], %f188; mov.u32 %r258, 0; ld.local.u64 %rd362, [%rd60+28]; st.u32 [%rd362+36], %r258; setp.le.ftz.f32 %p173, %f71, 0f3D99999A; setp.eq.s32 %p174, %r148, 26; or.pred %p175, %p174, %p173; @%p175 bra $L__BB139_218; bra.uni $L__BB139_209; $L__BB139_218: ld.local.u64 %rd367, [%rd60+28]; ld.f32 %f1573, [%rd367+12]; cvt.ftz.sat.f32.f32 %f1574, %f1573; ld.f32 %f1575, [%rd367+16]; cvt.ftz.sat.f32.f32 %f1576, %f1575; ld.f32 %f1577, [%rd367+20]; cvt.ftz.sat.f32.f32 %f1578, %f1577; st.f32 [%rd367+12], %f1574; st.f32 [%rd367+16], %f1576; st.f32 [%rd367+20], %f1578; ld.local.v2.f32 {%f1579, %f1580}, [%rd60+12]; cvt.ftz.sat.f32.f32 %f1583, %f1580; cvt.ftz.sat.f32.f32 %f1584, %f1579; st.local.v2.f32 [%rd60+12], {%f1584, %f1583}; mov.u32 %r264, 11; st.local.u32 [%rd60+-8], %r264; ld.local.f32 %f1585, [%rd60+20]; fma.rn.ftz.f32 %f266, %f1585, %f1585, 0fBF800000; mov.f32 %f1572, 0f3F800000; abs.ftz.f32 %f267, %f1572; fma.rn.ftz.f32 %f268, %f267, %f267, %f266; setp.leu.ftz.f32 %p184, %f268, 0f00000000; mov.f32 %f3019, %f1572; @%p184 bra $L__BB139_220; sqrt.approx.ftz.f32 %f1586, %f268; sub.ftz.f32 %f1587, %f1586, %f267; add.ftz.f32 %f1588, %f267, %f1586; div.approx.ftz.f32 %f1589, %f1587, %f1588; fma.rn.ftz.f32 %f1590, %f267, %f1588, 0fBF800000; fma.rn.ftz.f32 %f1591, %f267, %f1587, 0f3F800000; div.approx.ftz.f32 %f1592, %f1590, %f1591; mul.ftz.f32 %f1593, %f1589, 0f3F000000; mul.ftz.f32 %f1594, %f1589, %f1593; fma.rn.ftz.f32 %f1595, %f1592, %f1592, 0f3F800000; mul.ftz.f32 %f3019, %f1594, %f1595; $L__BB139_220: ld.local.u64 %rd368, [%rd60+28]; add.s64 %rd62, %rd368, 12; ld.f32 %f271, [%rd368+12]; ld.f32 %f272, [%rd368+16]; ld.f32 %f273, [%rd368+20]; sub.ftz.f32 %f274, %f1572, %f3019; ld.local.f32 %f1597, [%rd60]; ld.local.f32 %f1598, [%rd4+24]; ld.local.f32 %f1599, [%rd60+4]; ld.local.f32 %f1600, [%rd4+28]; mul.ftz.f32 %f1601, %f1600, %f1599; fma.rn.ftz.f32 %f1602, %f1598, %f1597, %f1601; ld.local.f32 %f1603, [%rd60+8]; ld.local.f32 %f1604, [%rd4+32]; fma.rn.ftz.f32 %f1605, %f1604, %f1603, %f1602; abs.ftz.f32 %f275, %f1605; fma.rn.ftz.f32 %f276, %f275, %f275, %f266; setp.leu.ftz.f32 %p185, %f276, 0f00000000; mov.f32 %f3020, %f274; @%p185 bra $L__BB139_222; sqrt.approx.ftz.f32 %f1606, %f276; sub.ftz.f32 %f1607, %f1606, %f275; add.ftz.f32 %f1608, %f275, %f1606; div.approx.ftz.f32 %f1609, %f1607, %f1608; fma.rn.ftz.f32 %f1610, %f275, %f1608, 0fBF800000; fma.rn.ftz.f32 %f1611, %f275, %f1607, 0f3F800000; div.approx.ftz.f32 %f1612, %f1610, %f1611; mul.ftz.f32 %f1613, %f1609, 0f3F000000; mul.ftz.f32 %f1614, %f1609, %f1613; fma.rn.ftz.f32 %f1615, %f1612, %f1612, 0f3F800000; mul.ftz.f32 %f1616, %f1614, %f1615; sub.ftz.f32 %f3020, %f1616, %f3019; $L__BB139_222: rcp.approx.ftz.f32 %f1617, %f274; mov.f32 %f1618, 0f3F800000; mul.ftz.f32 %f1619, %f1617, %f3020; sub.ftz.f32 %f1620, %f1618, %f1619; fma.rn.ftz.f32 %f1621, %f271, %f1620, %f1619; fma.rn.ftz.f32 %f1622, %f272, %f1620, %f1619; fma.rn.ftz.f32 %f1623, %f273, %f1620, %f1619; st.f32 [%rd62+12], %f1621; st.f32 [%rd62+16], %f1622; st.f32 [%rd62+20], %f1623; ld.local.u32 %r265, [%rd60+-8]; setp.ne.s32 %p186, %r265, 12; @%p186 bra $L__BB139_224; ld.local.u64 %rd369, [%rd60+28]; ld.f32 %f1624, [%rd369+36]; mul.ftz.f32 %f1625, %f1624, 0f3E800000; ld.f32 %f1626, [%rd369+24]; mul.ftz.f32 %f1627, %f1626, %f1625; ld.f32 %f1628, [%rd369+28]; mul.ftz.f32 %f1629, %f1625, %f1628; ld.f32 %f1630, [%rd369+32]; mul.ftz.f32 %f1631, %f1625, %f1630; st.f32 [%rd369+24], %f1627; st.f32 [%rd369+28], %f1629; st.f32 [%rd369+32], %f1631; $L__BB139_224: ld.local.u64 %rd370, [%rd60+28]; ld.f32 %f1632, [%rd370+28]; ld.f32 %f1633, [%rd370+24]; add.ftz.f32 %f1634, %f1633, %f1632; ld.f32 %f1635, [%rd370+32]; add.ftz.f32 %f1636, %f1635, %f1634; mul.ftz.f32 %f1637, %f1636, 0f3EAAAAAB; ld.local.f32 %f1638, [%rd60+-4]; mul.ftz.f32 %f1639, %f1638, %f1637; st.local.f32 [%rd60+-4], %f1639; ld.local.u32 %r266, [%rd4+40]; or.b32 %r267, %r266, 12; st.local.u32 [%rd4+40], %r267; bra.uni $L__BB139_225; $L__BB139_209: ld.local.v2.f32 {%f1487, %f1488}, [%rd60+36]; setp.neu.ftz.f32 %p176, %f1487, 0f00000000; setp.neu.ftz.f32 %p177, %f1488, 0f00000000; or.pred %p178, %p176, %p177; ld.local.f32 %f1491, [%rd60+44]; setp.neu.ftz.f32 %p179, %f1491, 0f00000000; or.pred %p180, %p179, %p178; @%p180 bra $L__BB139_211; mov.f32 %f1493, 0f3F800000; st.local.v2.f32 [%rd60+36], {%f1493, %f1448}; st.local.u32 [%rd60+44], %r258; $L__BB139_211: mov.u32 %r260, 15; st.local.u32 [%rd60+-8], %r260; ld.local.f32 %f1495, [%rd60+20]; fma.rn.ftz.f32 %f253, %f1495, %f1495, 0fBF800000; mov.f32 %f1494, 0f3F800000; abs.ftz.f32 %f254, %f1494; fma.rn.ftz.f32 %f255, %f254, %f254, %f253; setp.leu.ftz.f32 %p181, %f255, 0f00000000; mov.f32 %f3017, %f1494; @%p181 bra $L__BB139_213; sqrt.approx.ftz.f32 %f1496, %f255; sub.ftz.f32 %f1497, %f1496, %f254; add.ftz.f32 %f1498, %f254, %f1496; div.approx.ftz.f32 %f1499, %f1497, %f1498; fma.rn.ftz.f32 %f1500, %f254, %f1498, 0fBF800000; fma.rn.ftz.f32 %f1501, %f254, %f1497, 0f3F800000; div.approx.ftz.f32 %f1502, %f1500, %f1501; mul.ftz.f32 %f1503, %f1499, 0f3F000000; mul.ftz.f32 %f1504, %f1499, %f1503; fma.rn.ftz.f32 %f1505, %f1502, %f1502, 0f3F800000; mul.ftz.f32 %f3017, %f1504, %f1505; $L__BB139_213: ld.local.u64 %rd363, [%rd60+28]; add.s64 %rd61, %rd363, 12; ld.f32 %f258, [%rd363+12]; ld.f32 %f259, [%rd363+16]; ld.f32 %f260, [%rd363+20]; sub.ftz.f32 %f261, %f1494, %f3017; ld.local.f32 %f1507, [%rd60]; ld.local.f32 %f1508, [%rd4+24]; ld.local.f32 %f1509, [%rd60+4]; ld.local.f32 %f1510, [%rd4+28]; mul.ftz.f32 %f1511, %f1510, %f1509; fma.rn.ftz.f32 %f1512, %f1508, %f1507, %f1511; ld.local.f32 %f1513, [%rd60+8]; ld.local.f32 %f1514, [%rd4+32]; fma.rn.ftz.f32 %f1515, %f1514, %f1513, %f1512; abs.ftz.f32 %f262, %f1515; fma.rn.ftz.f32 %f263, %f262, %f262, %f253; setp.leu.ftz.f32 %p182, %f263, 0f00000000; mov.f32 %f3018, %f261; @%p182 bra $L__BB139_215; sqrt.approx.ftz.f32 %f1516, %f263; sub.ftz.f32 %f1517, %f1516, %f262; add.ftz.f32 %f1518, %f262, %f1516; div.approx.ftz.f32 %f1519, %f1517, %f1518; fma.rn.ftz.f32 %f1520, %f262, %f1518, 0fBF800000; fma.rn.ftz.f32 %f1521, %f262, %f1517, 0f3F800000; div.approx.ftz.f32 %f1522, %f1520, %f1521; mul.ftz.f32 %f1523, %f1519, 0f3F000000; mul.ftz.f32 %f1524, %f1519, %f1523; fma.rn.ftz.f32 %f1525, %f1522, %f1522, 0f3F800000; mul.ftz.f32 %f1526, %f1524, %f1525; sub.ftz.f32 %f3018, %f1526, %f3017; $L__BB139_215: rcp.approx.ftz.f32 %f1527, %f261; mov.f32 %f1528, 0f3F800000; mul.ftz.f32 %f1529, %f1527, %f3018; sub.ftz.f32 %f1530, %f1528, %f1529; fma.rn.ftz.f32 %f1531, %f258, %f1530, %f1529; fma.rn.ftz.f32 %f1532, %f259, %f1530, %f1529; fma.rn.ftz.f32 %f1533, %f260, %f1530, %f1529; st.f32 [%rd61+12], %f1531; st.f32 [%rd61+16], %f1532; st.f32 [%rd61+20], %f1533; ld.local.u32 %r261, [%rd60+-8]; setp.ne.s32 %p183, %r261, 12; @%p183 bra $L__BB139_217; ld.local.u64 %rd364, [%rd60+28]; ld.f32 %f1534, [%rd364+36]; mul.ftz.f32 %f1535, %f1534, 0f3E800000; ld.f32 %f1536, [%rd364+24]; mul.ftz.f32 %f1537, %f1536, %f1535; ld.f32 %f1538, [%rd364+28]; mul.ftz.f32 %f1539, %f1535, %f1538; ld.f32 %f1540, [%rd364+32]; mul.ftz.f32 %f1541, %f1535, %f1540; st.f32 [%rd364+24], %f1537; st.f32 [%rd364+28], %f1539; st.f32 [%rd364+32], %f1541; $L__BB139_217: ld.local.u64 %rd365, [%rd60+28]; ld.f32 %f1542, [%rd365+28]; ld.f32 %f1543, [%rd365+24]; add.ftz.f32 %f1544, %f1543, %f1542; ld.f32 %f1545, [%rd365+32]; add.ftz.f32 %f1546, %f1545, %f1544; mul.ftz.f32 %f1547, %f1546, 0f3EAAAAAB; ld.local.f32 %f1548, [%rd60+-4]; mul.ftz.f32 %f1549, %f1548, %f1547; st.local.f32 [%rd60+-4], %f1549; ld.local.v2.f32 {%f1550, %f1551}, [%rd60+12]; mov.f32 %f1554, 0f38D1B717; max.ftz.f32 %f1555, %f1550, %f1554; max.ftz.f32 %f1557, %f1551, %f1554; min.ftz.f32 %f1558, %f1557, %f1528; min.ftz.f32 %f1559, %f1555, %f1528; st.local.v2.f32 [%rd60+12], {%f1559, %f1558}; ld.f32 %f1560, [%rd365]; cvt.ftz.sat.f32.f32 %f1561, %f1560; ld.f32 %f1562, [%rd365+4]; cvt.ftz.sat.f32.f32 %f1563, %f1562; ld.f32 %f1564, [%rd365+8]; cvt.ftz.sat.f32.f32 %f1565, %f1564; st.f32 [%rd365], %f1561; st.f32 [%rd365+4], %f1563; st.f32 [%rd365+8], %f1565; ld.local.u64 %rd366, [%rd60+28]; ld.f32 %f1566, [%rd366+12]; cvt.ftz.sat.f32.f32 %f1567, %f1566; ld.f32 %f1568, [%rd366+16]; cvt.ftz.sat.f32.f32 %f1569, %f1568; ld.f32 %f1570, [%rd366+20]; cvt.ftz.sat.f32.f32 %f1571, %f1570; st.f32 [%rd366+12], %f1567; st.f32 [%rd366+16], %f1569; st.f32 [%rd366+20], %f1571; ld.local.u32 %r262, [%rd4+40]; or.b32 %r263, %r262, 1036; st.local.u32 [%rd4+40], %r263; $L__BB139_225: ld.const.u32 %r48, [kernel_params+1568]; or.b32 %r268, %r48, %r44; setp.eq.s32 %p187, %r268, 0; @%p187 bra $L__BB139_227; setp.gt.ftz.f32 %p188, %f98, 0f3727C5AC; @%p188 bra $L__BB139_228; bra.uni $L__BB139_272; $L__BB139_227: setp.leu.ftz.f32 %p189, %f98, 0f3727C5AC; setp.ne.s32 %p190, %r45, 0; or.pred %p191, %p190, %p189; @%p191 bra $L__BB139_272; $L__BB139_228: mul.ftz.f32 %f279, %f98, %f177; mul.ftz.f32 %f280, %f98, %f178; mul.ftz.f32 %f281, %f98, %f179; mov.f32 %f1640, 0f3F800000; sub.ftz.f32 %f1641, %f1640, %f72; fma.rn.ftz.f32 %f282, %f72, %f186, %f1641; fma.rn.ftz.f32 %f283, %f72, %f187, %f1641; fma.rn.ftz.f32 %f284, %f72, %f188, %f1641; setp.eq.s32 %p192, %r148, 26; setp.le.ftz.f32 %p193, %f71, 0f3D4CCCCD; or.pred %p194, %p192, %p193; @%p194 bra $L__BB139_246; bra.uni $L__BB139_229; $L__BB139_246: @%p158 bra $L__BB139_264; mul.ftz.f32 %f1722, %f2981, %f279; mov.f32 %f1723, 0f00000000; max.ftz.f32 %f302, %f1722, %f1723; mul.ftz.f32 %f1724, %f2981, %f280; max.ftz.f32 %f303, %f1724, %f1723; mul.ftz.f32 %f1725, %f2981, %f281; max.ftz.f32 %f304, %f1725, %f1723; add.ftz.f32 %f1726, %f302, %f303; add.ftz.f32 %f1727, %f304, %f1726; mul.ftz.f32 %f1728, %f1727, 0f3EAAAAAB; abs.ftz.f32 %f305, %f1728; setp.ltu.ftz.f32 %p207, %f305, 0f3727C5AC; mov.u64 %rd732, 0; mov.u64 %rd730, %rd732; mov.u64 %rd731, %rd732; @%p207 bra $L__BB139_252; ld.local.u32 %r51, [%rd4+252]; mov.u64 %rd730, 0; setp.eq.s32 %p208, %r51, 0; mov.u64 %rd728, %rd730; mov.u64 %rd729, %rd730; @%p208 bra $L__BB139_250; ld.local.u32 %r285, [%rd4+248]; mul.wide.s32 %rd401, %r285, 80; add.s64 %rd402, %rd1, %rd401; add.s64 %rd728, %rd402, 304; add.s64 %rd403, %rd230, %rd401; add.s64 %rd729, %rd403, 304; mov.b32 %r286, %f304; mov.b32 %r287, %f303; mov.b32 %r288, %f302; mov.u32 %r289, 0; st.local.v4.u32 [%rd402+304], {%r288, %r287, %r286, %r289}; add.s32 %r290, %r285, 1; st.local.u32 [%rd4+248], %r290; add.s32 %r291, %r51, -1; st.local.u32 [%rd4+252], %r291; $L__BB139_250: setp.eq.s64 %p209, %rd729, 0; mov.u64 %rd731, %rd730; @%p209 bra $L__BB139_252; st.local.f32 [%rd728+16], %f305; mov.u64 %rd730, %rd728; mov.u64 %rd731, %rd729; $L__BB139_252: setp.eq.s64 %p210, %rd731, 0; mov.u64 %rd733, %rd732; @%p210 bra $L__BB139_256; ld.local.u32 %r52, [%rd4+252]; setp.lt.s32 %p211, %r52, 1; @%p211 bra $L__BB139_255; add.s32 %r292, %r52, -1; st.local.u32 [%rd4+252], %r292; ld.local.s32 %rd408, [%rd4+248]; cvt.s64.s32 %rd409, %r292; add.s64 %rd410, %rd408, %rd409; mul.lo.s64 %rd411, %rd410, 80; add.s64 %rd412, %rd1, %rd411; add.s64 %rd732, %rd412, 304; add.s64 %rd413, %rd230, %rd411; add.s64 %rd733, %rd413, 304; bra.uni $L__BB139_256; $L__BB139_229: mov.f32 %f1642, 0f00000000; max.ftz.f32 %f285, %f279, %f1642; max.ftz.f32 %f286, %f280, %f1642; add.ftz.f32 %f1643, %f285, %f286; max.ftz.f32 %f287, %f281, %f1642; add.ftz.f32 %f1644, %f287, %f1643; mul.ftz.f32 %f1645, %f1644, 0f3EAAAAAB; abs.ftz.f32 %f288, %f1645; setp.ltu.ftz.f32 %p195, %f288, 0f3727C5AC; mov.u64 %rd726, 0; mov.u64 %rd724, %rd726; mov.u64 %rd725, %rd726; @%p195 bra $L__BB139_234; ld.local.u32 %r49, [%rd4+252]; mov.u64 %rd724, 0; setp.eq.s32 %p196, %r49, 0; mov.u64 %rd722, %rd724; mov.u64 %rd723, %rd724; @%p196 bra $L__BB139_232; ld.local.u32 %r269, [%rd4+248]; mul.wide.s32 %rd375, %r269, 80; add.s64 %rd376, %rd1, %rd375; add.s64 %rd722, %rd376, 304; add.s64 %rd377, %rd230, %rd375; add.s64 %rd723, %rd377, 304; mov.b32 %r270, %f287; mov.b32 %r271, %f286; mov.b32 %r272, %f285; mov.u32 %r273, 0; st.local.v4.u32 [%rd376+304], {%r272, %r271, %r270, %r273}; add.s32 %r274, %r269, 1; st.local.u32 [%rd4+248], %r274; add.s32 %r275, %r49, -1; st.local.u32 [%rd4+252], %r275; $L__BB139_232: setp.eq.s64 %p197, %rd723, 0; mov.u64 %rd725, %rd724; @%p197 bra $L__BB139_234; st.local.f32 [%rd722+16], %f288; mov.u64 %rd724, %rd722; mov.u64 %rd725, %rd723; $L__BB139_234: setp.eq.s64 %p198, %rd725, 0; mov.u64 %rd727, %rd726; @%p198 bra $L__BB139_238; ld.local.u32 %r50, [%rd4+252]; setp.lt.s32 %p199, %r50, 1; @%p199 bra $L__BB139_237; add.s32 %r276, %r50, -1; st.local.u32 [%rd4+252], %r276; ld.local.s32 %rd382, [%rd4+248]; cvt.s64.s32 %rd383, %r276; add.s64 %rd384, %rd382, %rd383; mul.lo.s64 %rd385, %rd384, 80; add.s64 %rd386, %rd1, %rd385; add.s64 %rd726, %rd386, 304; add.s64 %rd387, %rd230, %rd385; add.s64 %rd727, %rd387, 304; bra.uni $L__BB139_238; $L__BB139_237: ld.local.u32 %r277, [%rd4+248]; add.s32 %r278, %r277, -1; st.local.u32 [%rd4+248], %r278; add.s32 %r279, %r50, 1; st.local.u32 [%rd4+252], %r279; mov.u64 %rd727, %rd726; $L__BB139_238: setp.eq.s64 %p200, %rd727, 0; or.pred %p202, %p198, %p200; @%p202 bra $L__BB139_272; add.s64 %rd73, %rd724, 20; st.local.f32 [%rd724+20], %f55; st.local.f32 [%rd724+24], %f56; st.local.f32 [%rd724+28], %f57; st.local.u64 [%rd724+48], %rd727; mov.u32 %r280, 0; st.local.v2.f32 [%rd724+56], {%f1642, %f1642}; st.local.u32 [%rd724+64], %r280; mul.ftz.f32 %f1648, %f71, %f71; st.local.f32 [%rd724+32], %f1648; st.local.f32 [%rd724+40], %f2980; st.local.f32 [%rd726], %f186; st.local.f32 [%rd726+4], %f187; st.local.f32 [%rd726+8], %f188; ld.local.u64 %rd390, [%rd724+48]; st.f32 [%rd390+12], %f282; st.f32 [%rd390+16], %f283; st.f32 [%rd390+20], %f284; ld.local.u64 %rd391, [%rd724+48]; st.u32 [%rd391+36], %r280; ld.local.f32 %f1649, [%rd724+32]; mov.f32 %f1650, 0f38D1B717; max.ftz.f32 %f1651, %f1649, %f1650; min.ftz.f32 %f1652, %f1651, %f1640; st.local.f32 [%rd724+32], %f1652; st.local.f32 [%rd724+36], %f1652; ld.local.f32 %f1653, [%rd724+40]; max.ftz.f32 %f1654, %f1642, %f1653; st.local.f32 [%rd724+40], %f1654; ld.local.u64 %rd392, [%rd724+48]; ld.f32 %f1655, [%rd392]; cvt.ftz.sat.f32.f32 %f1656, %f1655; ld.f32 %f1657, [%rd392+4]; cvt.ftz.sat.f32.f32 %f1658, %f1657; ld.f32 %f1659, [%rd392+8]; cvt.ftz.sat.f32.f32 %f1660, %f1659; st.f32 [%rd392], %f1656; st.f32 [%rd392+4], %f1658; st.f32 [%rd392+8], %f1660; ld.local.u64 %rd393, [%rd724+48]; ld.f32 %f1661, [%rd393+12]; cvt.ftz.sat.f32.f32 %f1662, %f1661; ld.f32 %f1663, [%rd393+16]; cvt.ftz.sat.f32.f32 %f1664, %f1663; ld.f32 %f1665, [%rd393+20]; cvt.ftz.sat.f32.f32 %f1666, %f1665; st.f32 [%rd393+12], %f1662; st.f32 [%rd393+16], %f1664; st.f32 [%rd393+20], %f1666; mov.u32 %r281, 27; st.local.u32 [%rd724+12], %r281; ld.local.f32 %f1667, [%rd724+40]; fma.rn.ftz.f32 %f289, %f1667, %f1667, 0fBF800000; abs.ftz.f32 %f290, %f1640; fma.rn.ftz.f32 %f291, %f290, %f290, %f289; setp.leu.ftz.f32 %p203, %f291, 0f00000000; mov.f32 %f3021, %f1640; @%p203 bra $L__BB139_241; sqrt.approx.ftz.f32 %f1668, %f291; sub.ftz.f32 %f1669, %f1668, %f290; add.ftz.f32 %f1670, %f290, %f1668; div.approx.ftz.f32 %f1671, %f1669, %f1670; fma.rn.ftz.f32 %f1672, %f290, %f1670, 0fBF800000; fma.rn.ftz.f32 %f1673, %f290, %f1669, 0f3F800000; div.approx.ftz.f32 %f1674, %f1672, %f1673; mul.ftz.f32 %f1675, %f1671, 0f3F000000; mul.ftz.f32 %f1676, %f1671, %f1675; fma.rn.ftz.f32 %f1677, %f1674, %f1674, 0f3F800000; mul.ftz.f32 %f3021, %f1676, %f1677; $L__BB139_241: ld.local.u64 %rd394, [%rd73+28]; add.s64 %rd74, %rd394, 12; ld.f32 %f294, [%rd394+12]; ld.f32 %f295, [%rd394+16]; ld.f32 %f296, [%rd394+20]; sub.ftz.f32 %f297, %f1640, %f3021; ld.local.f32 %f1679, [%rd73]; ld.local.f32 %f1680, [%rd4+24]; ld.local.f32 %f1681, [%rd73+4]; ld.local.f32 %f1682, [%rd4+28]; mul.ftz.f32 %f1683, %f1682, %f1681; fma.rn.ftz.f32 %f1684, %f1680, %f1679, %f1683; ld.local.f32 %f1685, [%rd73+8]; ld.local.f32 %f1686, [%rd4+32]; fma.rn.ftz.f32 %f1687, %f1686, %f1685, %f1684; abs.ftz.f32 %f298, %f1687; fma.rn.ftz.f32 %f299, %f298, %f298, %f289; setp.leu.ftz.f32 %p204, %f299, 0f00000000; mov.f32 %f3022, %f297; @%p204 bra $L__BB139_243; sqrt.approx.ftz.f32 %f1688, %f299; sub.ftz.f32 %f1689, %f1688, %f298; add.ftz.f32 %f1690, %f298, %f1688; div.approx.ftz.f32 %f1691, %f1689, %f1690; fma.rn.ftz.f32 %f1692, %f298, %f1690, 0fBF800000; fma.rn.ftz.f32 %f1693, %f298, %f1689, 0f3F800000; div.approx.ftz.f32 %f1694, %f1692, %f1693; mul.ftz.f32 %f1695, %f1691, 0f3F000000; mul.ftz.f32 %f1696, %f1691, %f1695; fma.rn.ftz.f32 %f1697, %f1694, %f1694, 0f3F800000; mul.ftz.f32 %f1698, %f1696, %f1697; sub.ftz.f32 %f3022, %f1698, %f3021; $L__BB139_243: rcp.approx.ftz.f32 %f1699, %f297; mov.f32 %f1700, 0f3F800000; mul.ftz.f32 %f1701, %f1699, %f3022; sub.ftz.f32 %f1702, %f1700, %f1701; fma.rn.ftz.f32 %f1703, %f294, %f1702, %f1701; fma.rn.ftz.f32 %f1704, %f295, %f1702, %f1701; fma.rn.ftz.f32 %f1705, %f296, %f1702, %f1701; st.f32 [%rd74+12], %f1703; st.f32 [%rd74+16], %f1704; st.f32 [%rd74+20], %f1705; ld.local.u32 %r282, [%rd73+-8]; setp.ne.s32 %p205, %r282, 12; @%p205 bra $L__BB139_245; ld.local.u64 %rd395, [%rd73+28]; ld.f32 %f1706, [%rd395+36]; mul.ftz.f32 %f1707, %f1706, 0f3E800000; ld.f32 %f1708, [%rd395+24]; mul.ftz.f32 %f1709, %f1708, %f1707; ld.f32 %f1710, [%rd395+28]; mul.ftz.f32 %f1711, %f1707, %f1710; ld.f32 %f1712, [%rd395+32]; mul.ftz.f32 %f1713, %f1707, %f1712; st.f32 [%rd395+24], %f1709; st.f32 [%rd395+28], %f1711; st.f32 [%rd395+32], %f1713; $L__BB139_245: ld.local.u64 %rd396, [%rd73+28]; ld.f32 %f1714, [%rd396+28]; ld.f32 %f1715, [%rd396+24]; add.ftz.f32 %f1716, %f1715, %f1714; ld.f32 %f1717, [%rd396+32]; add.ftz.f32 %f1718, %f1717, %f1716; mul.ftz.f32 %f1719, %f1718, 0f3EAAAAAB; ld.local.f32 %f1720, [%rd73+-4]; mul.ftz.f32 %f1721, %f1720, %f1719; st.local.f32 [%rd73+-4], %f1721; ld.local.u32 %r283, [%rd4+40]; or.b32 %r284, %r283, 1036; st.local.u32 [%rd4+40], %r284; bra.uni $L__BB139_272; $L__BB139_255: ld.local.u32 %r293, [%rd4+248]; add.s32 %r294, %r293, -1; st.local.u32 [%rd4+248], %r294; add.s32 %r295, %r52, 1; st.local.u32 [%rd4+252], %r295; mov.u64 %rd733, %rd732; $L__BB139_256: setp.eq.s64 %p212, %rd733, 0; or.pred %p214, %p210, %p212; @%p214 bra $L__BB139_264; add.s64 %rd85, %rd730, 20; st.local.f32 [%rd730+20], %f55; st.local.f32 [%rd730+24], %f56; st.local.f32 [%rd730+28], %f57; mov.u32 %r296, 0; st.local.v2.f32 [%rd730+56], {%f1723, %f1723}; st.local.u32 [%rd730+64], %r296; st.local.u64 [%rd730+48], %rd733; mul.ftz.f32 %f1731, %f71, %f71; st.local.v2.f32 [%rd730+32], {%f1731, %f1731}; st.local.f32 [%rd730+40], %f2980; st.local.f32 [%rd732], %f186; st.local.f32 [%rd732+4], %f187; st.local.f32 [%rd732+8], %f188; ld.local.u64 %rd416, [%rd730+48]; st.f32 [%rd416+12], %f282; st.f32 [%rd416+16], %f283; st.f32 [%rd416+20], %f284; ld.local.u64 %rd417, [%rd730+48]; st.u32 [%rd417+36], %r296; ld.local.u64 %rd418, [%rd730+48]; ld.f32 %f1732, [%rd418+12]; cvt.ftz.sat.f32.f32 %f1733, %f1732; ld.f32 %f1734, [%rd418+16]; cvt.ftz.sat.f32.f32 %f1735, %f1734; ld.f32 %f1736, [%rd418+20]; cvt.ftz.sat.f32.f32 %f1737, %f1736; st.f32 [%rd418+12], %f1733; st.f32 [%rd418+16], %f1735; st.f32 [%rd418+20], %f1737; ld.local.v2.f32 {%f1738, %f1739}, [%rd730+32]; cvt.ftz.sat.f32.f32 %f1742, %f1739; cvt.ftz.sat.f32.f32 %f1743, %f1738; st.local.v2.f32 [%rd730+32], {%f1743, %f1742}; mov.u32 %r297, 11; st.local.u32 [%rd730+12], %r297; ld.local.f32 %f1744, [%rd730+40]; fma.rn.ftz.f32 %f306, %f1744, %f1744, 0fBF800000; mov.f32 %f1729, 0f3F800000; abs.ftz.f32 %f307, %f1729; fma.rn.ftz.f32 %f308, %f307, %f307, %f306; setp.leu.ftz.f32 %p215, %f308, 0f00000000; mov.f32 %f3023, %f1729; @%p215 bra $L__BB139_259; sqrt.approx.ftz.f32 %f1745, %f308; sub.ftz.f32 %f1746, %f1745, %f307; add.ftz.f32 %f1747, %f307, %f1745; div.approx.ftz.f32 %f1748, %f1746, %f1747; fma.rn.ftz.f32 %f1749, %f307, %f1747, 0fBF800000; fma.rn.ftz.f32 %f1750, %f307, %f1746, 0f3F800000; div.approx.ftz.f32 %f1751, %f1749, %f1750; mul.ftz.f32 %f1752, %f1748, 0f3F000000; mul.ftz.f32 %f1753, %f1748, %f1752; fma.rn.ftz.f32 %f1754, %f1751, %f1751, 0f3F800000; mul.ftz.f32 %f3023, %f1753, %f1754; $L__BB139_259: ld.local.u64 %rd419, [%rd85+28]; add.s64 %rd86, %rd419, 12; ld.f32 %f311, [%rd419+12]; ld.f32 %f312, [%rd419+16]; ld.f32 %f313, [%rd419+20]; sub.ftz.f32 %f314, %f1729, %f3023; ld.local.f32 %f1756, [%rd85]; ld.local.f32 %f1757, [%rd4+24]; ld.local.f32 %f1758, [%rd85+4]; ld.local.f32 %f1759, [%rd4+28]; mul.ftz.f32 %f1760, %f1759, %f1758; fma.rn.ftz.f32 %f1761, %f1757, %f1756, %f1760; ld.local.f32 %f1762, [%rd85+8]; ld.local.f32 %f1763, [%rd4+32]; fma.rn.ftz.f32 %f1764, %f1763, %f1762, %f1761; abs.ftz.f32 %f315, %f1764; fma.rn.ftz.f32 %f316, %f315, %f315, %f306; setp.leu.ftz.f32 %p216, %f316, 0f00000000; mov.f32 %f3024, %f314; @%p216 bra $L__BB139_261; sqrt.approx.ftz.f32 %f1765, %f316; sub.ftz.f32 %f1766, %f1765, %f315; add.ftz.f32 %f1767, %f315, %f1765; div.approx.ftz.f32 %f1768, %f1766, %f1767; fma.rn.ftz.f32 %f1769, %f315, %f1767, 0fBF800000; fma.rn.ftz.f32 %f1770, %f315, %f1766, 0f3F800000; div.approx.ftz.f32 %f1771, %f1769, %f1770; mul.ftz.f32 %f1772, %f1768, 0f3F000000; mul.ftz.f32 %f1773, %f1768, %f1772; fma.rn.ftz.f32 %f1774, %f1771, %f1771, 0f3F800000; mul.ftz.f32 %f1775, %f1773, %f1774; sub.ftz.f32 %f3024, %f1775, %f3023; $L__BB139_261: rcp.approx.ftz.f32 %f1776, %f314; mov.f32 %f1777, 0f3F800000; mul.ftz.f32 %f1778, %f1776, %f3024; sub.ftz.f32 %f1779, %f1777, %f1778; fma.rn.ftz.f32 %f1780, %f311, %f1779, %f1778; fma.rn.ftz.f32 %f1781, %f312, %f1779, %f1778; fma.rn.ftz.f32 %f1782, %f313, %f1779, %f1778; st.f32 [%rd86+12], %f1780; st.f32 [%rd86+16], %f1781; st.f32 [%rd86+20], %f1782; ld.local.u32 %r298, [%rd85+-8]; setp.ne.s32 %p217, %r298, 12; @%p217 bra $L__BB139_263; ld.local.u64 %rd420, [%rd85+28]; ld.f32 %f1783, [%rd420+36]; mul.ftz.f32 %f1784, %f1783, 0f3E800000; ld.f32 %f1785, [%rd420+24]; mul.ftz.f32 %f1786, %f1785, %f1784; ld.f32 %f1787, [%rd420+28]; mul.ftz.f32 %f1788, %f1784, %f1787; ld.f32 %f1789, [%rd420+32]; mul.ftz.f32 %f1790, %f1784, %f1789; st.f32 [%rd420+24], %f1786; st.f32 [%rd420+28], %f1788; st.f32 [%rd420+32], %f1790; $L__BB139_263: ld.local.u64 %rd421, [%rd85+28]; ld.f32 %f1791, [%rd421+28]; ld.f32 %f1792, [%rd421+24]; add.ftz.f32 %f1793, %f1792, %f1791; ld.f32 %f1794, [%rd421+32]; add.ftz.f32 %f1795, %f1794, %f1793; mul.ftz.f32 %f1796, %f1795, 0f3EAAAAAB; ld.local.f32 %f1797, [%rd85+-4]; mul.ftz.f32 %f1798, %f1797, %f1796; st.local.f32 [%rd85+-4], %f1798; ld.local.u32 %r299, [%rd4+40]; or.b32 %r300, %r299, 12; st.local.u32 [%rd4+40], %r300; $L__BB139_264: setp.ne.s32 %p218, %r45, 0; setp.eq.s32 %p219, %r48, 0; and.pred %p220, %p218, %p219; @%p220 bra $L__BB139_272; mov.f32 %f1799, 0f3F800000; sub.ftz.f32 %f1800, %f1799, %f2981; mov.f32 %f1801, 0f38D1B717; max.ftz.f32 %f1802, %f1801, %f1800; mul.ftz.f32 %f1803, %f186, %f279; mul.ftz.f32 %f1804, %f1803, %f1802; mul.ftz.f32 %f1805, %f187, %f280; mul.ftz.f32 %f1806, %f1805, %f1802; mul.ftz.f32 %f1807, %f188, %f281; mul.ftz.f32 %f1808, %f1807, %f1802; mov.f32 %f1809, 0f00000000; max.ftz.f32 %f319, %f1804, %f1809; max.ftz.f32 %f320, %f1806, %f1809; max.ftz.f32 %f321, %f1808, %f1809; add.ftz.f32 %f1810, %f319, %f320; add.ftz.f32 %f1811, %f321, %f1810; mul.ftz.f32 %f1812, %f1811, 0f3EAAAAAB; abs.ftz.f32 %f322, %f1812; setp.ltu.ftz.f32 %p221, %f322, 0f3727C5AC; mov.u64 %rd736, 0; mov.u64 %rd737, %rd736; @%p221 bra $L__BB139_270; ld.local.u32 %r53, [%rd4+252]; mov.u64 %rd736, 0; setp.eq.s32 %p222, %r53, 0; mov.u64 %rd734, %rd736; mov.u64 %rd735, %rd736; @%p222 bra $L__BB139_268; ld.local.u32 %r301, [%rd4+248]; mul.wide.s32 %rd426, %r301, 80; add.s64 %rd427, %rd1, %rd426; add.s64 %rd734, %rd427, 304; add.s64 %rd428, %rd230, %rd426; add.s64 %rd735, %rd428, 304; mov.b32 %r302, %f321; mov.b32 %r303, %f320; mov.b32 %r304, %f319; mov.u32 %r305, 0; st.local.v4.u32 [%rd427+304], {%r304, %r303, %r302, %r305}; add.s32 %r306, %r301, 1; st.local.u32 [%rd4+248], %r306; add.s32 %r307, %r53, -1; st.local.u32 [%rd4+252], %r307; $L__BB139_268: setp.eq.s64 %p223, %rd735, 0; mov.u64 %rd737, %rd736; @%p223 bra $L__BB139_270; st.local.f32 [%rd734+16], %f322; mov.u64 %rd736, %rd734; mov.u64 %rd737, %rd735; $L__BB139_270: setp.eq.s64 %p224, %rd737, 0; @%p224 bra $L__BB139_272; mov.u64 %rd431, 0; st.local.f32 [%rd736+20], %f55; st.local.f32 [%rd736+24], %f56; st.local.f32 [%rd736+28], %f57; mov.u32 %r308, 0; st.local.v2.f32 [%rd736+56], {%f1809, %f1809}; st.local.u32 [%rd736+64], %r308; sub.ftz.f32 %f1815, %f1799, %f79; sub.ftz.f32 %f1816, %f1799, %f71; mul.ftz.f32 %f1817, %f1816, %f1815; sub.ftz.f32 %f1818, %f1799, %f1817; selp.f32 %f1819, %f1818, %f71, %p192; mul.ftz.f32 %f1820, %f1819, %f1819; st.local.f32 [%rd736+40], %f2980; st.local.u64 [%rd736+48], %rd431; cvt.ftz.sat.f32.f32 %f1821, %f1820; st.local.v2.f32 [%rd736+32], {%f1821, %f1821}; mov.u32 %r309, 23; st.local.u32 [%rd736+12], %r309; ld.local.u32 %r310, [%rd4+40]; or.b32 %r311, %r310, 12; st.local.u32 [%rd4+40], %r311; $L__BB139_272: ld.param.u32 %r638, [_Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; add.s32 %r663, %r638, 5; setp.leu.ftz.f32 %p226, %f76, 0f3727C5AC; or.pred %p228, %p226, %p158; @%p228 bra $L__BB139_517; mov.f32 %f1822, 0f00000000; max.ftz.f32 %f323, %f177, %f1822; max.ftz.f32 %f324, %f178, %f1822; add.ftz.f32 %f1823, %f323, %f324; max.ftz.f32 %f325, %f179, %f1822; add.ftz.f32 %f1824, %f325, %f1823; mul.ftz.f32 %f1825, %f1824, 0f3EAAAAAB; abs.ftz.f32 %f326, %f1825; setp.ltu.ftz.f32 %p229, %f326, 0f3727C5AC; mov.u64 %rd742, 0; mov.u64 %rd740, %rd742; mov.u64 %rd741, %rd742; @%p229 bra $L__BB139_278; ld.local.u32 %r55, [%rd4+252]; mov.u64 %rd740, 0; setp.eq.s32 %p230, %r55, 0; mov.u64 %rd738, %rd740; mov.u64 %rd739, %rd740; @%p230 bra $L__BB139_276; ld.local.u32 %r312, [%rd4+248]; mul.wide.s32 %rd436, %r312, 80; add.s64 %rd437, %rd1, %rd436; add.s64 %rd738, %rd437, 304; add.s64 %rd438, %rd230, %rd436; add.s64 %rd739, %rd438, 304; mov.b32 %r313, %f325; mov.b32 %r314, %f324; mov.b32 %r315, %f323; mov.u32 %r316, 0; st.local.v4.u32 [%rd437+304], {%r315, %r314, %r313, %r316}; add.s32 %r317, %r312, 1; st.local.u32 [%rd4+248], %r317; add.s32 %r318, %r55, -1; st.local.u32 [%rd4+252], %r318; $L__BB139_276: setp.eq.s64 %p231, %rd739, 0; mov.u64 %rd741, %rd740; @%p231 bra $L__BB139_278; st.local.f32 [%rd738+16], %f326; mov.u64 %rd740, %rd738; mov.u64 %rd741, %rd739; $L__BB139_278: setp.eq.s64 %p232, %rd741, 0; mov.u64 %rd743, %rd742; @%p232 bra $L__BB139_282; ld.local.u32 %r56, [%rd4+252]; setp.lt.s32 %p233, %r56, 1; @%p233 bra $L__BB139_281; add.s32 %r319, %r56, -1; st.local.u32 [%rd4+252], %r319; ld.local.s32 %rd443, [%rd4+248]; cvt.s64.s32 %rd444, %r319; add.s64 %rd445, %rd443, %rd444; mul.lo.s64 %rd446, %rd445, 80; add.s64 %rd447, %rd1, %rd446; add.s64 %rd742, %rd447, 304; add.s64 %rd448, %rd230, %rd446; add.s64 %rd743, %rd448, 304; bra.uni $L__BB139_282; $L__BB139_281: ld.local.u32 %r320, [%rd4+248]; add.s32 %r321, %r320, -1; st.local.u32 [%rd4+248], %r321; add.s32 %r322, %r56, 1; st.local.u32 [%rd4+252], %r322; mov.u64 %rd743, %rd742; $L__BB139_282: setp.eq.s64 %p234, %rd743, 0; or.pred %p236, %p232, %p234; @%p236 bra $L__BB139_517; mov.f32 %f3025, 0f3F800000; add.s64 %rd103, %rd740, 20; st.local.f32 [%rd740+20], %f2990; st.local.f32 [%rd740+24], %f2991; st.local.f32 [%rd740+28], %f2992; mov.u32 %r323, 0; st.local.v2.f32 [%rd740+56], {%f1822, %f1822}; st.local.u32 [%rd740+64], %r323; mov.u32 %r324, 1069547520; st.local.u32 [%rd740+40], %r324; st.local.u64 [%rd740+48], %rd743; mul.ftz.f32 %f1828, %f77, %f77; st.local.v2.f32 [%rd740+32], {%f1828, %f1828}; st.local.u32 [%rd742], %r323; st.local.u32 [%rd742+4], %r323; st.local.u32 [%rd742+8], %r323; ld.local.u64 %rd451, [%rd740+48]; mov.u32 %r325, 1025758986; st.u32 [%rd451+12], %r325; st.u32 [%rd451+16], %r325; st.u32 [%rd451+20], %r325; ld.local.u64 %rd452, [%rd740+48]; st.f32 [%rd452+36], %f76; ld.local.u64 %rd453, [%rd740+48]; ld.f32 %f1829, [%rd453+12]; cvt.ftz.sat.f32.f32 %f1830, %f1829; ld.f32 %f1831, [%rd453+16]; cvt.ftz.sat.f32.f32 %f1832, %f1831; ld.f32 %f1833, [%rd453+20]; cvt.ftz.sat.f32.f32 %f1834, %f1833; st.f32 [%rd453+12], %f1830; st.f32 [%rd453+16], %f1832; st.f32 [%rd453+20], %f1834; ld.local.f32 %f1835, [%rd740+32]; cvt.ftz.sat.f32.f32 %f1836, %f1835; st.local.v2.f32 [%rd740+32], {%f1836, %f1836}; mov.u32 %r326, 12; st.local.u32 [%rd740+12], %r326; ld.local.f32 %f1837, [%rd740+40]; fma.rn.ftz.f32 %f327, %f1837, %f1837, 0fBF800000; abs.ftz.f32 %f328, %f3025; fma.rn.ftz.f32 %f329, %f328, %f328, %f327; setp.leu.ftz.f32 %p237, %f329, 0f00000000; @%p237 bra $L__BB139_285; sqrt.approx.ftz.f32 %f1838, %f329; sub.ftz.f32 %f1839, %f1838, %f328; add.ftz.f32 %f1840, %f328, %f1838; div.approx.ftz.f32 %f1841, %f1839, %f1840; fma.rn.ftz.f32 %f1842, %f328, %f1840, 0fBF800000; fma.rn.ftz.f32 %f1843, %f328, %f1839, 0f3F800000; div.approx.ftz.f32 %f1844, %f1842, %f1843; mul.ftz.f32 %f1845, %f1841, 0f3F000000; mul.ftz.f32 %f1846, %f1841, %f1845; fma.rn.ftz.f32 %f1847, %f1844, %f1844, 0f3F800000; mul.ftz.f32 %f3025, %f1846, %f1847; $L__BB139_285: mov.f32 %f2960, 0f3F800000; ld.local.u64 %rd454, [%rd103+28]; add.s64 %rd104, %rd454, 12; ld.f32 %f332, [%rd454+12]; ld.f32 %f333, [%rd454+16]; ld.f32 %f334, [%rd454+20]; sub.ftz.f32 %f335, %f2960, %f3025; ld.local.f32 %f1849, [%rd103]; ld.local.f32 %f1850, [%rd4+24]; ld.local.f32 %f1851, [%rd103+4]; ld.local.f32 %f1852, [%rd4+28]; mul.ftz.f32 %f1853, %f1852, %f1851; fma.rn.ftz.f32 %f1854, %f1850, %f1849, %f1853; ld.local.f32 %f1855, [%rd103+8]; ld.local.f32 %f1856, [%rd4+32]; fma.rn.ftz.f32 %f1857, %f1856, %f1855, %f1854; abs.ftz.f32 %f336, %f1857; fma.rn.ftz.f32 %f337, %f336, %f336, %f327; setp.leu.ftz.f32 %p238, %f337, 0f00000000; mov.f32 %f3026, %f335; @%p238 bra $L__BB139_287; sqrt.approx.ftz.f32 %f1858, %f337; sub.ftz.f32 %f1859, %f1858, %f336; add.ftz.f32 %f1860, %f336, %f1858; div.approx.ftz.f32 %f1861, %f1859, %f1860; fma.rn.ftz.f32 %f1862, %f336, %f1860, 0fBF800000; fma.rn.ftz.f32 %f1863, %f336, %f1859, 0f3F800000; div.approx.ftz.f32 %f1864, %f1862, %f1863; mul.ftz.f32 %f1865, %f1861, 0f3F000000; mul.ftz.f32 %f1866, %f1861, %f1865; fma.rn.ftz.f32 %f1867, %f1864, %f1864, 0f3F800000; mul.ftz.f32 %f1868, %f1866, %f1867; sub.ftz.f32 %f3026, %f1868, %f3025; $L__BB139_287: rcp.approx.ftz.f32 %f1869, %f335; mov.f32 %f1870, 0f3F800000; mul.ftz.f32 %f1871, %f1869, %f3026; sub.ftz.f32 %f1872, %f1870, %f1871; fma.rn.ftz.f32 %f1873, %f332, %f1872, %f1871; fma.rn.ftz.f32 %f1874, %f333, %f1872, %f1871; fma.rn.ftz.f32 %f1875, %f334, %f1872, %f1871; st.f32 [%rd104+12], %f1873; st.f32 [%rd104+16], %f1874; st.f32 [%rd104+20], %f1875; ld.local.u32 %r327, [%rd103+-8]; setp.ne.s32 %p239, %r327, 12; @%p239 bra $L__BB139_289; ld.local.u64 %rd455, [%rd103+28]; ld.f32 %f1876, [%rd455+36]; mul.ftz.f32 %f1877, %f1876, 0f3E800000; ld.f32 %f1878, [%rd455+24]; mul.ftz.f32 %f1879, %f1878, %f1877; ld.f32 %f1880, [%rd455+28]; mul.ftz.f32 %f1881, %f1877, %f1880; ld.f32 %f1882, [%rd455+32]; mul.ftz.f32 %f1883, %f1877, %f1882; st.f32 [%rd455+24], %f1879; st.f32 [%rd455+28], %f1881; st.f32 [%rd455+32], %f1883; $L__BB139_289: ld.local.u64 %rd456, [%rd103+28]; ld.f32 %f1884, [%rd456+28]; ld.f32 %f1885, [%rd456+24]; add.ftz.f32 %f1886, %f1885, %f1884; ld.f32 %f1887, [%rd456+32]; add.ftz.f32 %f1888, %f1887, %f1886; mul.ftz.f32 %f1889, %f1888, 0f3EAAAAAB; ld.local.f32 %f1890, [%rd103+-4]; mul.ftz.f32 %f1891, %f1890, %f1889; st.local.f32 [%rd103+-4], %f1891; ld.local.u32 %r328, [%rd4+40]; or.b32 %r329, %r328, 12; st.local.u32 [%rd4+40], %r329; bra.uni $L__BB139_517; } .func _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB140_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB140_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB140_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB140_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB140_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB140_11; $L__BB140_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB140_8; add.s32 %r237, %r237, 2; bra.uni $L__BB140_9; $L__BB140_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB140_15; $L__BB140_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB140_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB140_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB140_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB140_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB140_15; $L__BB140_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB140_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB140_17; $L__BB140_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB140_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB140_159; bra.uni $L__BB140_18; $L__BB140_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB140_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB140_185; $L__BB140_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB140_150; bra.uni $L__BB140_19; $L__BB140_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB140_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB140_153; bra.uni $L__BB140_152; $L__BB140_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB140_154; $L__BB140_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB140_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB140_100; bra.uni $L__BB140_21; $L__BB140_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB140_111; bra.uni $L__BB140_101; $L__BB140_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB140_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB140_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB140_120; $L__BB140_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB140_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB140_164; bra.uni $L__BB140_163; $L__BB140_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB140_165; $L__BB140_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB140_136; bra.uni $L__BB140_126; $L__BB140_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB140_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB140_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB140_145; $L__BB140_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB140_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB140_158; bra.uni $L__BB140_155; $L__BB140_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB140_192; $L__BB140_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB140_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB140_192; $L__BB140_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB140_47; bra.uni $L__BB140_22; $L__BB140_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB140_86; bra.uni $L__BB140_48; $L__BB140_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB140_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB140_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB140_95; $L__BB140_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB140_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB140_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB140_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB140_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB140_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB140_184; bra.uni $L__BB140_170; $L__BB140_184: { // callseq 899, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 899 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB140_185; $L__BB140_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB140_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB140_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 893, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 893 bra.uni $L__BB140_145; $L__BB140_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB140_192; $L__BB140_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB140_183; bra.uni $L__BB140_171; $L__BB140_183: { // callseq 898, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 898 mov.f32 %f731, 0f3F800000; bra.uni $L__BB140_185; $L__BB140_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB140_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB140_104; mov.u64 %rd173, 0; { // callseq 892, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 892 bra.uni $L__BB140_120; $L__BB140_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB140_33; bra.uni $L__BB140_23; $L__BB140_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB140_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB140_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB140_42; $L__BB140_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB140_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB140_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB140_133; $L__BB140_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB140_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB140_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB140_108; $L__BB140_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB140_182; bra.uni $L__BB140_172; $L__BB140_182: { // callseq 897, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 897 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB140_185; $L__BB140_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB140_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB140_51; mov.u64 %rd129, 0; { // callseq 891, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 891 bra.uni $L__BB140_95; $L__BB140_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB140_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB140_26; mov.u64 %rd90, 0; { // callseq 890, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 890 bra.uni $L__BB140_42; $L__BB140_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB140_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB140_145; $L__BB140_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB140_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB140_120; $L__BB140_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB140_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB140_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB140_83; $L__BB140_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB140_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB140_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB140_30; $L__BB140_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB140_181; bra.uni $L__BB140_173; $L__BB140_181: { // callseq 896, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 896 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB140_185; $L__BB140_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB140_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB140_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB140_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB140_145; $L__BB140_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB140_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB140_145; $L__BB140_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB140_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB140_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB140_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB140_120; $L__BB140_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB140_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB140_120; $L__BB140_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB140_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB140_95; $L__BB140_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB140_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB140_42; $L__BB140_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB140_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB140_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB140_149; bra.uni $L__BB140_146; $L__BB140_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB140_192; $L__BB140_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB140_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB140_192; $L__BB140_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB140_192; $L__BB140_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB140_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB140_175; bra.uni $L__BB140_178; $L__BB140_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB140_177; bra.uni $L__BB140_176; $L__BB140_177: { // callseq 894, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 894 bra.uni $L__BB140_185; $L__BB140_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB140_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB140_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB140_124; bra.uni $L__BB140_121; $L__BB140_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB140_192; $L__BB140_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB140_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB140_192; $L__BB140_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB140_192; $L__BB140_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB140_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB140_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB140_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB140_95; $L__BB140_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB140_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB140_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB140_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB140_42; $L__BB140_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB140_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB140_95; $L__BB140_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB140_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB140_42; $L__BB140_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB140_82; bra.uni $L__BB140_55; $L__BB140_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB140_83; $L__BB140_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB140_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB140_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB140_46; bra.uni $L__BB140_43; $L__BB140_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB140_192; $L__BB140_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB140_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB140_192; $L__BB140_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB140_192; $L__BB140_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB140_185; $L__BB140_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB140_57; bra.uni $L__BB140_56; $L__BB140_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB140_58; $L__BB140_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB140_180; bra.uni $L__BB140_179; $L__BB140_180: { // callseq 895, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 895 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB140_185; $L__BB140_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB140_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB140_60; bra.uni $L__BB140_59; $L__BB140_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB140_61; $L__BB140_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB140_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB140_63; bra.uni $L__BB140_62; $L__BB140_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB140_64; $L__BB140_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB140_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB140_66; bra.uni $L__BB140_65; $L__BB140_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB140_67; $L__BB140_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB140_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB140_69; bra.uni $L__BB140_68; $L__BB140_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB140_70; $L__BB140_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB140_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB140_72; bra.uni $L__BB140_71; $L__BB140_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB140_73; $L__BB140_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB140_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB140_75; bra.uni $L__BB140_74; $L__BB140_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB140_76; $L__BB140_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB140_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB140_78; bra.uni $L__BB140_77; $L__BB140_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB140_79; $L__BB140_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB140_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB140_81; bra.uni $L__BB140_80; $L__BB140_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB140_83; $L__BB140_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB140_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB140_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB140_99; bra.uni $L__BB140_96; $L__BB140_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB140_192; $L__BB140_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB140_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB140_192; $L__BB140_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB140_192; $L__BB140_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB140_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB140_191; bra.uni $L__BB140_186; $L__BB140_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB140_192; $L__BB140_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB140_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB140_192; $L__BB140_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB140_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB140_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB140_192: ret; } .func _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB141_2; bra.uni $L__BB141_1; $L__BB141_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB141_3; $L__BB141_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB141_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB141_5; bra.uni $L__BB141_4; $L__BB141_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB141_9; bra.uni $L__BB141_6; $L__BB141_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB141_12; bra.uni $L__BB141_10; $L__BB141_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB141_13; $L__BB141_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB141_16; $L__BB141_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB141_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB141_13; $L__BB141_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB141_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB141_15; bra.uni $L__BB141_14; $L__BB141_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB141_16; $L__BB141_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB141_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB141_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB141_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB141_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB142_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB142_5; bra.uni $L__BB142_2; $L__BB142_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB142_8; bra.uni $L__BB142_6; $L__BB142_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB142_9; $L__BB142_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB142_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB142_9; $L__BB142_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB142_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB142_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB142_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB142_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB142_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB142_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB142_20; $L__BB142_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB142_17; add.s32 %r172, %r172, 2; bra.uni $L__BB142_18; $L__BB142_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB142_24; $L__BB142_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB142_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB142_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB142_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB142_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB142_24; $L__BB142_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB142_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB142_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB142_37; bra.uni $L__BB142_26; $L__BB142_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB142_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB142_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB142_45; $L__BB142_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB142_45; $L__BB142_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB142_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB142_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB142_30; mov.u64 %rd108, 0; { // callseq 900, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 900 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB142_45; $L__BB142_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB142_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB142_9; $L__BB142_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB142_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB142_45; $L__BB142_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB142_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB142_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB142_34; $L__BB142_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB142_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB142_45; $L__BB142_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB142_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB142_45; $L__BB142_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB142_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB142_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB142_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB142_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB142_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB142_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB142_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB142_56; $L__BB142_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB142_53; add.s32 %r187, %r187, 2; bra.uni $L__BB142_54; $L__BB142_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB142_60; $L__BB142_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB142_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB142_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB142_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB142_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB142_60; $L__BB142_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB142_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB142_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB142_72; bra.uni $L__BB142_62; $L__BB142_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB142_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB142_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB142_79; $L__BB142_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB142_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB142_65; mov.u64 %rd160, 0; { // callseq 901, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 901 bra.uni $L__BB142_79; $L__BB142_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB142_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB142_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB142_69; $L__BB142_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB142_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB142_79; $L__BB142_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB142_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB142_79; $L__BB142_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB142_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB142_79; $L__BB142_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB142_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB142_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB142_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB142_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB142_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB142_84; bra.uni $L__BB142_83; $L__BB142_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB142_85; $L__BB142_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB142_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB143_2; bra.uni $L__BB143_1; $L__BB143_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB143_3; $L__BB143_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB143_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB143_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB143_8; bra.uni $L__BB143_5; $L__BB143_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB143_11; bra.uni $L__BB143_9; $L__BB143_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB143_12; $L__BB143_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB143_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB143_12; $L__BB143_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB143_12: @%p8 bra $L__BB143_14; bra.uni $L__BB143_13; $L__BB143_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB143_15; $L__BB143_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB143_15: @%p8 bra $L__BB143_17; bra.uni $L__BB143_16; $L__BB143_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB143_18; $L__BB143_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB143_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB143_19; bra.uni $L__BB143_20; $L__BB143_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB143_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB143_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB143_22: @%p7 bra $L__BB143_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB143_25; bra.uni $L__BB143_24; $L__BB143_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB143_26; $L__BB143_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB143_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB143_27; bra.uni $L__BB143_34; $L__BB143_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB143_30; bra.uni $L__BB143_28; $L__BB143_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB143_32; bra.uni $L__BB143_31; $L__BB143_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB143_33; $L__BB143_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB143_29; bra.uni $L__BB143_34; $L__BB143_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB143_33; $L__BB143_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB143_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB143_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB143_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB143_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB143_12; } .func _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b64 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<58>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u64 %rd8, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd7, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; cvta.to.local.u64 %rd9, %rd8; add.s64 %rd1, %rd9, 88; mov.f32 %f20, 0f00000000; setp.gt.s32 %p1, %r2, 6; @%p1 bra $L__BB144_13; bra.uni $L__BB144_1; $L__BB144_13: setp.gt.s32 %p2, %r2, 10; @%p2 bra $L__BB144_23; setp.gt.s32 %p8, %r2, 8; @%p8 bra $L__BB144_18; setp.eq.s32 %p11, %r2, 7; @%p11 bra $L__BB144_41; setp.eq.s32 %p12, %r2, 8; @%p12 bra $L__BB144_17; bra.uni $L__BB144_46; $L__BB144_17: ld.local.u32 %r11, [%rd1+-36]; and.b32 %r12, %r11, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB144_46; $L__BB144_1: setp.gt.s32 %p13, %r2, 2; @%p13 bra $L__BB144_6; bra.uni $L__BB144_2; $L__BB144_6: setp.gt.s32 %p14, %r2, 4; @%p14 bra $L__BB144_10; setp.eq.s32 %p17, %r2, 3; @%p17 bra $L__BB144_43; setp.eq.s32 %p18, %r2, 4; @%p18 bra $L__BB144_9; bra.uni $L__BB144_46; $L__BB144_9: and.b32 %r16, %r4, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB144_46; $L__BB144_23: setp.gt.s32 %p3, %r2, 12; @%p3 bra $L__BB144_29; setp.eq.s32 %p6, %r2, 11; @%p6 bra $L__BB144_37; setp.eq.s32 %p7, %r2, 12; @%p7 bra $L__BB144_26; bra.uni $L__BB144_46; $L__BB144_26: and.b32 %r7, %r4, 768; mov.u64 %rd27, kernel_params; add.s64 %rd55, %rd27, 3400; setp.ne.s32 %p24, %r7, 0; @%p24 bra $L__BB144_28; add.s64 %rd55, %rd27, 2488; $L__BB144_28: ld.const.u64 %rd30, [%rd55]; cvta.to.global.u64 %rd31, %rd30; mul.wide.s32 %rd32, %r1, 2; add.s64 %rd33, %rd31, %rd32; ld.global.u16 %rs3, [%rd33]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB144_46; $L__BB144_2: setp.eq.s32 %p19, %r2, 0; @%p19 bra $L__BB144_45; setp.eq.s32 %p20, %r2, 1; @%p20 bra $L__BB144_44; setp.eq.s32 %p21, %r2, 2; @%p21 bra $L__BB144_5; bra.uni $L__BB144_46; $L__BB144_5: and.b32 %r18, %r4, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB144_46; $L__BB144_18: setp.eq.s32 %p9, %r2, 9; @%p9 bra $L__BB144_40; setp.eq.s32 %p10, %r2, 10; @%p10 bra $L__BB144_20; bra.uni $L__BB144_46; $L__BB144_20: and.b32 %r9, %r4, 768; mov.u64 %rd43, kernel_params; add.s64 %rd57, %rd43, 3376; setp.ne.s32 %p26, %r9, 0; @%p26 bra $L__BB144_22; add.s64 %rd57, %rd43, 2464; $L__BB144_22: ld.const.u64 %rd46, [%rd57]; cvta.to.global.u64 %rd47, %rd46; mul.wide.s32 %rd48, %r1, 2; add.s64 %rd49, %rd47, %rd48; ld.global.u16 %rs5, [%rd49]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r10, %r4, 525056; setp.eq.s32 %p27, %r10, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB144_46; $L__BB144_10: setp.eq.s32 %p15, %r2, 5; @%p15 bra $L__BB144_42; setp.eq.s32 %p16, %r2, 6; @%p16 bra $L__BB144_12; bra.uni $L__BB144_46; $L__BB144_12: and.b32 %r14, %r4, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB144_46; $L__BB144_29: setp.eq.s32 %p4, %r2, 13; @%p4 bra $L__BB144_34; setp.ne.s32 %p5, %r2, 14; @%p5 bra $L__BB144_46; and.b32 %r5, %r4, 768; mov.u64 %rd11, kernel_params; add.s64 %rd53, %rd11, 3408; setp.ne.s32 %p22, %r5, 0; @%p22 bra $L__BB144_33; add.s64 %rd53, %rd11, 2496; $L__BB144_33: ld.const.u64 %rd14, [%rd53]; cvta.to.global.u64 %rd15, %rd14; mul.wide.s32 %rd16, %r1, 2; add.s64 %rd17, %rd15, %rd16; ld.global.u16 %rs1, [%rd17]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB144_46; $L__BB144_41: and.b32 %r13, %r4, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB144_46; $L__BB144_43: and.b32 %r17, %r4, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB144_46; $L__BB144_37: and.b32 %r8, %r4, 768; mov.u64 %rd35, kernel_params; add.s64 %rd56, %rd35, 3392; setp.ne.s32 %p25, %r8, 0; @%p25 bra $L__BB144_39; add.s64 %rd56, %rd35, 2480; $L__BB144_39: ld.const.u64 %rd38, [%rd56]; cvta.to.global.u64 %rd39, %rd38; mul.wide.s32 %rd40, %r1, 2; add.s64 %rd41, %rd39, %rd40; ld.global.u16 %rs4, [%rd41]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB144_46; $L__BB144_45: and.b32 %r20, %r4, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB144_46; $L__BB144_44: and.b32 %r19, %r4, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB144_46; $L__BB144_40: ld.local.f32 %f20, [%rd1]; bra.uni $L__BB144_46; $L__BB144_42: and.b32 %r15, %r4, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB144_46; $L__BB144_34: and.b32 %r6, %r4, 768; mov.u64 %rd19, kernel_params; add.s64 %rd54, %rd19, 3384; setp.ne.s32 %p23, %r6, 0; @%p23 bra $L__BB144_36; add.s64 %rd54, %rd19, 2472; $L__BB144_36: ld.const.u64 %rd22, [%rd54]; cvta.to.global.u64 %rd23, %rd22; mul.wide.s32 %rd24, %r1, 2; add.s64 %rd25, %rd23, %rd24; ld.global.u16 %rs2, [%rd25]; cvt.rn.f32.u16 %f20, %rs2; $L__BB144_46: cvta.to.local.u64 %rd50, %rd7; mul.wide.u32 %rd51, %r3, 4; add.s64 %rd52, %rd50, %rd51; st.local.f32 [%rd52], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB145_2; bra.uni $L__BB145_1; $L__BB145_2: mov.b32 %f2764, %r395; bra.uni $L__BB145_3; $L__BB145_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB145_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB145_5; bra.uni $L__BB145_4; $L__BB145_5: mov.b32 %f9, %r396; bra.uni $L__BB145_6; $L__BB145_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB145_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB145_8; bra.uni $L__BB145_7; $L__BB145_8: mov.b32 %f2766, %r397; bra.uni $L__BB145_9; $L__BB145_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB145_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB145_11; bra.uni $L__BB145_10; $L__BB145_11: mov.b32 %f15, %r398; bra.uni $L__BB145_12; $L__BB145_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB145_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB145_14; bra.uni $L__BB145_13; $L__BB145_14: mov.b32 %f2768, %r399; bra.uni $L__BB145_15; $L__BB145_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB145_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB145_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB145_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB145_18; bra.uni $L__BB145_385; $L__BB145_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB145_259; bra.uni $L__BB145_19; $L__BB145_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB145_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB145_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB145_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB145_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB145_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB145_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB145_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB145_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB145_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB145_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB145_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB145_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB145_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB145_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB145_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB145_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB145_362; $L__BB145_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB145_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB145_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB145_40; bra.uni $L__BB145_29; $L__BB145_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB145_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB145_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB145_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB145_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB145_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB145_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB145_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB145_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB145_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB145_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB145_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB145_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB145_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB145_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB145_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB145_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB145_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB145_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB145_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB145_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB145_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB145_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB145_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB145_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB145_136; $L__BB145_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB145_371; bra.uni $L__BB145_366; $L__BB145_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB145_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB145_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB145_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB145_383; $L__BB145_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB145_149; bra.uni $L__BB145_139; $L__BB145_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB145_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB145_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB145_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB145_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB145_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB145_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB145_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB145_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB145_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB145_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB145_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB145_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB145_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB145_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB145_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB145_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB145_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB145_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB145_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB145_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB145_250; $L__BB145_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB145_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB145_253; bra.uni $L__BB145_362; $L__BB145_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB145_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB145_304; @%p171 bra $L__BB145_303; bra.uni $L__BB145_256; $L__BB145_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB145_305; $L__BB145_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB145_305; $L__BB145_256: @%p172 bra $L__BB145_302; bra.uni $L__BB145_257; $L__BB145_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB145_305; $L__BB145_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB145_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB145_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB145_308; bra.uni $L__BB145_306; $L__BB145_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB145_309; $L__BB145_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB145_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB145_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB145_316; @%p171 bra $L__BB145_315; bra.uni $L__BB145_311; $L__BB145_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB145_317; $L__BB145_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB145_317; $L__BB145_311: @%p172 bra $L__BB145_314; bra.uni $L__BB145_312; $L__BB145_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB145_317; $L__BB145_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB145_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB145_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB145_320; bra.uni $L__BB145_318; $L__BB145_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB145_321; $L__BB145_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB145_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB145_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB145_328; @%p171 bra $L__BB145_327; bra.uni $L__BB145_323; $L__BB145_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB145_329; $L__BB145_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB145_329; $L__BB145_323: @%p172 bra $L__BB145_326; bra.uni $L__BB145_324; $L__BB145_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB145_329; $L__BB145_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB145_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB145_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB145_332; bra.uni $L__BB145_330; $L__BB145_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB145_333; $L__BB145_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB145_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB145_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB145_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_362; $L__BB145_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB145_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB145_31; bra.uni $L__BB145_136; $L__BB145_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB145_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB145_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB145_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB145_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB145_107; @%p49 bra $L__BB145_106; bra.uni $L__BB145_37; $L__BB145_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB145_108; $L__BB145_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB145_108; $L__BB145_37: @%p50 bra $L__BB145_105; bra.uni $L__BB145_38; $L__BB145_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB145_108; $L__BB145_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB145_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB145_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB145_111; bra.uni $L__BB145_109; $L__BB145_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB145_112; $L__BB145_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB145_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB145_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB145_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB145_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB145_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB145_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_136; $L__BB145_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB145_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB145_368; bra.uni $L__BB145_383; $L__BB145_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB145_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB145_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB145_383; $L__BB145_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB145_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB145_141; bra.uni $L__BB145_250; $L__BB145_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB145_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB145_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB145_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB145_206; @%p101 bra $L__BB145_205; bra.uni $L__BB145_146; $L__BB145_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB145_207; $L__BB145_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB145_207; $L__BB145_146: @%p102 bra $L__BB145_204; bra.uni $L__BB145_147; $L__BB145_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB145_207; $L__BB145_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB145_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB145_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB145_210; bra.uni $L__BB145_208; $L__BB145_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB145_211; $L__BB145_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB145_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB145_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB145_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB145_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB145_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_250; $L__BB145_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB145_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB145_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB145_295; @%p151 bra $L__BB145_294; bra.uni $L__BB145_290; $L__BB145_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB145_296; $L__BB145_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB145_296; $L__BB145_290: @%p152 bra $L__BB145_293; bra.uni $L__BB145_291; $L__BB145_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB145_296; $L__BB145_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB145_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB145_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB145_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB145_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB145_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB145_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB145_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_362; $L__BB145_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB145_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB145_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB145_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB145_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB145_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB145_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB145_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB145_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB145_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB145_279; bra.uni $L__BB145_362; $L__BB145_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB145_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB145_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB145_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB145_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB145_96; @%p39 bra $L__BB145_95; bra.uni $L__BB145_91; $L__BB145_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB145_97; $L__BB145_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB145_97; $L__BB145_91: @%p40 bra $L__BB145_94; bra.uni $L__BB145_92; $L__BB145_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB145_97; $L__BB145_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB145_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB145_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB145_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB145_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB145_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB145_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB145_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB145_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB145_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_136; $L__BB145_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB145_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB145_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB145_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB145_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB145_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB145_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB145_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB145_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB145_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB145_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB145_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB145_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB145_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB145_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB145_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB145_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB145_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB145_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB145_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB145_76; bra.uni $L__BB145_136; $L__BB145_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB145_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB145_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB145_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB145_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB145_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_383; $L__BB145_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB145_383; $L__BB145_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB145_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB145_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB145_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB145_196; @%p92 bra $L__BB145_195; bra.uni $L__BB145_191; $L__BB145_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB145_197; $L__BB145_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB145_197; $L__BB145_191: @%p93 bra $L__BB145_194; bra.uni $L__BB145_192; $L__BB145_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB145_197; $L__BB145_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB145_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB145_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB145_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB145_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB145_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB145_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB145_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB145_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB145_250; $L__BB145_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB145_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB145_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB145_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB145_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB145_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB145_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB145_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB145_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB145_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB145_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB145_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB145_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB145_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB145_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB145_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB145_179; bra.uni $L__BB145_250; $L__BB145_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB145_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB145_337; @%p188 bra $L__BB145_336; bra.uni $L__BB145_23; $L__BB145_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB145_338; $L__BB145_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB145_338; $L__BB145_23: @%p189 bra $L__BB145_335; bra.uni $L__BB145_24; $L__BB145_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB145_338; $L__BB145_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB145_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB145_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB145_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB145_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB145_347; @%p188 bra $L__BB145_346; bra.uni $L__BB145_342; $L__BB145_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB145_348; $L__BB145_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB145_348; $L__BB145_342: @%p189 bra $L__BB145_345; bra.uni $L__BB145_343; $L__BB145_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB145_348; $L__BB145_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB145_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB145_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB145_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB145_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB145_357; @%p188 bra $L__BB145_356; bra.uni $L__BB145_352; $L__BB145_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB145_358; $L__BB145_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB145_358; $L__BB145_352: @%p189 bra $L__BB145_355; bra.uni $L__BB145_353; $L__BB145_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB145_358; $L__BB145_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB145_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB145_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB145_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB145_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB145_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB145_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB145_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB145_364: mov.f32 %f3392, %f3387; bra.uni $L__BB145_385; $L__BB145_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB145_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB145_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB145_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB145_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB145_128; @%p59 bra $L__BB145_127; bra.uni $L__BB145_123; $L__BB145_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB145_129; $L__BB145_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB145_129; $L__BB145_123: @%p60 bra $L__BB145_126; bra.uni $L__BB145_124; $L__BB145_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB145_129; $L__BB145_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB145_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB145_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB145_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB145_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB145_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB145_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB145_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB145_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB145_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB145_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB145_385; $L__BB145_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB145_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB145_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB145_385; $L__BB145_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB145_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB145_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB145_224; @%p120 bra $L__BB145_223; bra.uni $L__BB145_219; $L__BB145_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB145_225; $L__BB145_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB145_225; $L__BB145_219: @%p121 bra $L__BB145_222; bra.uni $L__BB145_220; $L__BB145_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB145_225; $L__BB145_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB145_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB145_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB145_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB145_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB145_234; @%p120 bra $L__BB145_233; bra.uni $L__BB145_229; $L__BB145_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB145_235; $L__BB145_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB145_235; $L__BB145_229: @%p121 bra $L__BB145_232; bra.uni $L__BB145_230; $L__BB145_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB145_235; $L__BB145_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB145_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB145_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB145_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB145_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB145_244; @%p120 bra $L__BB145_243; bra.uni $L__BB145_239; $L__BB145_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB145_245; $L__BB145_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB145_245; $L__BB145_239: @%p121 bra $L__BB145_242; bra.uni $L__BB145_240; $L__BB145_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB145_245; $L__BB145_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB145_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB145_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB145_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB145_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB145_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB145_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB145_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB145_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB145_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB145_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB145_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB145_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB145_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB145_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB145_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB145_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB145_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB145_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB145_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB147_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB147_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB147_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB147_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) { .reg .pred %p<463>; .reg .b16 %rs<22>; .reg .f32 %f<3060>; .reg .b32 %r<664>; .reg .b64 %rd<808>; ld.param.u64 %rd230, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd231, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r141, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r142, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; ld.param.u32 %r140, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r139, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r138, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd230; cvta.to.local.u64 %rd2, %rd231; and.b32 %r2, %r138, 255; bfe.u32 %r3, %r138, 8, 8; bfe.u32 %r4, %r138, 16, 8; shr.u32 %r5, %r138, 24; setp.eq.s32 %p7, %r5, 255; mov.f32 %f2, 0f3F800000; @%p7 bra $L__BB148_2; mul.wide.u32 %rd232, %r5, 4; add.s64 %rd233, %rd2, %rd232; ld.local.f32 %f2, [%rd233]; $L__BB148_2: ld.const.u64 %rd234, [kernel_params+2400]; cvta.to.global.u64 %rd235, %rd234; mul.wide.s32 %rd236, %r142, 16; add.s64 %rd3, %rd235, %rd236; ld.global.v4.u32 {%r143, %r144, %r145, %r146}, [%rd3]; add.s32 %r663, %r142, 1; setp.eq.ftz.f32 %p8, %f2, 0f00000000; @%p8 bra $L__BB148_516; bra.uni $L__BB148_3; $L__BB148_516: setp.eq.s32 %p461, %r2, 39; add.s32 %r636, %r142, 5; selp.b32 %r663, %r636, %r663, %p461; $L__BB148_517: st.param.b32 [func_retval0+0], %r663; ret; $L__BB148_3: setp.eq.s32 %p9, %r143, 255; add.s64 %rd4, %rd1, 12; @%p9 bra $L__BB148_5; bra.uni $L__BB148_4; $L__BB148_5: ld.local.f32 %f55, [%rd4]; ld.local.f32 %f56, [%rd4+4]; ld.local.f32 %f57, [%rd4+8]; bra.uni $L__BB148_6; $L__BB148_4: mul.wide.u32 %rd237, %r143, 4; add.s64 %rd238, %rd2, %rd237; ld.local.f32 %f55, [%rd238]; ld.local.f32 %f56, [%rd238+4]; ld.local.f32 %f57, [%rd238+8]; $L__BB148_6: ld.local.u32 %r11, [%rd4+52]; and.b32 %r12, %r11, 6; setp.ne.s32 %p10, %r12, 0; @%p10 bra $L__BB148_15; ld.local.v2.f32 {%f550, %f551}, [%rd4+12]; ld.local.v2.f32 {%f552, %f553}, [%rd4+20]; ld.local.v2.f32 {%f554, %f555}, [%rd4+28]; mul.ftz.f32 %f556, %f56, %f554; fma.rn.ftz.f32 %f557, %f55, %f553, %f556; fma.rn.ftz.f32 %f558, %f57, %f555, %f557; add.ftz.f32 %f559, %f558, %f558; mul.ftz.f32 %f560, %f55, %f559; mul.ftz.f32 %f561, %f56, %f559; mul.ftz.f32 %f562, %f57, %f559; sub.ftz.f32 %f563, %f560, %f553; sub.ftz.f32 %f564, %f561, %f554; sub.ftz.f32 %f565, %f562, %f555; mul.ftz.f32 %f566, %f551, %f554; fma.rn.ftz.f32 %f567, %f550, %f553, %f566; fma.rn.ftz.f32 %f18, %f552, %f555, %f567; mul.ftz.f32 %f568, %f18, 0f3F666666; mov.f32 %f569, 0f3C23D70A; min.ftz.f32 %f19, %f568, %f569; mul.ftz.f32 %f570, %f551, %f564; fma.rn.ftz.f32 %f571, %f550, %f563, %f570; fma.rn.ftz.f32 %f572, %f552, %f565, %f571; setp.ltu.ftz.f32 %p11, %f572, %f19; @%p11 bra $L__BB148_8; bra.uni $L__BB148_15; $L__BB148_8: mul.ftz.f32 %f573, %f56, %f551; fma.rn.ftz.f32 %f574, %f55, %f550, %f573; fma.rn.ftz.f32 %f575, %f57, %f552, %f574; mul.ftz.f32 %f576, %f550, %f575; mul.ftz.f32 %f577, %f551, %f575; mul.ftz.f32 %f578, %f552, %f575; sub.ftz.f32 %f579, %f55, %f576; sub.ftz.f32 %f580, %f56, %f577; sub.ftz.f32 %f581, %f57, %f578; mul.ftz.f32 %f582, %f580, %f580; fma.rn.ftz.f32 %f583, %f579, %f579, %f582; fma.rn.ftz.f32 %f584, %f581, %f581, %f583; sqrt.approx.ftz.f32 %f585, %f584; div.approx.ftz.f32 %f23, %f579, %f585; div.approx.ftz.f32 %f24, %f580, %f585; div.approx.ftz.f32 %f25, %f581, %f585; mul.ftz.f32 %f586, %f554, %f24; fma.rn.ftz.f32 %f587, %f553, %f23, %f586; fma.rn.ftz.f32 %f26, %f555, %f25, %f587; mul.ftz.f32 %f588, %f26, %f26; fma.rn.ftz.f32 %f589, %f18, %f18, %f588; mul.ftz.f32 %f590, %f19, %f19; sub.ftz.f32 %f591, %f589, %f590; mul.ftz.f32 %f592, %f588, %f591; mov.f32 %f593, 0f00000000; max.ftz.f32 %f594, %f592, %f593; sqrt.approx.ftz.f32 %f595, %f594; fma.rn.ftz.f32 %f596, %f19, %f18, %f589; mov.f32 %f597, 0f3F000000; div.approx.ftz.f32 %f598, %f597, %f589; add.ftz.f32 %f599, %f595, %f596; mul.ftz.f32 %f27, %f598, %f599; sub.ftz.f32 %f600, %f596, %f595; mul.ftz.f32 %f28, %f598, %f600; setp.gt.ftz.f32 %p12, %f27, 0f3727C5AC; setp.le.ftz.f32 %p13, %f27, 0f3F800054; and.pred %p1, %p12, %p13; setp.gt.ftz.f32 %p14, %f28, 0f3727C5AC; setp.le.ftz.f32 %p15, %f28, 0f3F800054; and.pred %p2, %p14, %p15; and.pred %p16, %p1, %p2; @%p16 bra $L__BB148_11; bra.uni $L__BB148_9; $L__BB148_11: mov.f32 %f607, 0f3F800000; sub.ftz.f32 %f608, %f607, %f27; max.ftz.f32 %f610, %f608, %f593; sqrt.approx.ftz.f32 %f34, %f610; max.ftz.f32 %f611, %f27, %f593; sqrt.approx.ftz.f32 %f35, %f611; sub.ftz.f32 %f612, %f607, %f28; max.ftz.f32 %f613, %f612, %f593; sqrt.approx.ftz.f32 %f36, %f613; max.ftz.f32 %f614, %f28, %f593; sqrt.approx.ftz.f32 %f37, %f614; mul.ftz.f32 %f615, %f18, %f35; fma.rn.ftz.f32 %f616, %f26, %f34, %f615; add.ftz.f32 %f617, %f616, %f616; mul.ftz.f32 %f618, %f35, %f617; sub.ftz.f32 %f38, %f618, %f18; mul.ftz.f32 %f619, %f18, %f37; fma.rn.ftz.f32 %f620, %f26, %f36, %f619; add.ftz.f32 %f621, %f620, %f620; mul.ftz.f32 %f622, %f37, %f621; sub.ftz.f32 %f39, %f622, %f18; setp.ge.ftz.f32 %p18, %f38, 0f3727C5AC; setp.ge.ftz.f32 %p19, %f39, 0f3727C5AC; and.pred %p20, %p18, %p19; @%p20 bra $L__BB148_13; bra.uni $L__BB148_12; $L__BB148_13: setp.lt.ftz.f32 %p22, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p22; selp.f32 %f2968, %f35, %f37, %p22; bra.uni $L__BB148_14; $L__BB148_9: or.pred %p17, %p1, %p2; mov.f32 %f55, %f550; mov.f32 %f56, %f551; mov.f32 %f57, %f552; @%p17 bra $L__BB148_10; bra.uni $L__BB148_15; $L__BB148_10: selp.f32 %f601, %f27, %f28, %p1; mov.f32 %f602, 0f3F800000; sub.ftz.f32 %f603, %f602, %f601; max.ftz.f32 %f605, %f603, %f593; sqrt.approx.ftz.f32 %f2967, %f605; max.ftz.f32 %f606, %f601, %f593; sqrt.approx.ftz.f32 %f2968, %f606; bra.uni $L__BB148_14; $L__BB148_12: setp.gt.ftz.f32 %p21, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p21; selp.f32 %f2968, %f35, %f37, %p21; $L__BB148_14: mul.ftz.f32 %f623, %f550, %f2968; mul.ftz.f32 %f624, %f551, %f2968; mul.ftz.f32 %f625, %f552, %f2968; fma.rn.ftz.f32 %f57, %f25, %f2967, %f625; fma.rn.ftz.f32 %f56, %f24, %f2967, %f624; fma.rn.ftz.f32 %f55, %f23, %f2967, %f623; $L__BB148_15: setp.eq.s32 %p23, %r3, 255; @%p23 bra $L__BB148_17; bra.uni $L__BB148_16; $L__BB148_17: mov.b32 %f60, %r139; bra.uni $L__BB148_18; $L__BB148_16: mul.wide.u32 %rd239, %r3, 4; add.s64 %rd240, %rd2, %rd239; ld.local.f32 %f60, [%rd240]; $L__BB148_18: setp.eq.s32 %p24, %r4, 255; @%p24 bra $L__BB148_20; bra.uni $L__BB148_19; $L__BB148_20: mov.b32 %f63, %r140; bra.uni $L__BB148_21; $L__BB148_19: mul.wide.u32 %rd241, %r4, 4; add.s64 %rd242, %rd2, %rd241; ld.local.f32 %f63, [%rd242]; $L__BB148_21: cvt.u16.u32 %rs8, %r138; and.b16 %rs7, %rs8, 255; mul.wide.u32 %rd243, %r144, 4; add.s64 %rd5, %rd2, %rd243; mul.wide.u32 %rd244, %r145, 4; add.s64 %rd6, %rd2, %rd244; mul.wide.u32 %rd245, %r146, 4; add.s64 %rd7, %rd2, %rd245; setp.gt.s16 %p25, %rs7, 23; @%p25 bra $L__BB148_63; setp.gt.s16 %p36, %rs7, 15; @%p36 bra $L__BB148_48; setp.gt.s16 %p43, %rs7, 8; @%p43 bra $L__BB148_34; bra.uni $L__BB148_24; $L__BB148_34: add.s16 %rs12, %rs7, -9; setp.lt.u16 %p44, %rs12, 2; @%p44 bra $L__BB148_36; add.s16 %rs13, %rs7, -13; setp.lt.u16 %p45, %rs13, 2; @%p45 bra $L__BB148_36; bra.uni $L__BB148_517; $L__BB148_63: setp.gt.s16 %p26, %rs7, 30; @%p26 bra $L__BB148_88; setp.gt.s16 %p30, %rs7, 27; @%p30 bra $L__BB148_77; add.s16 %rs10, %rs7, -25; setp.lt.u16 %p34, %rs10, 2; @%p34 bra $L__BB148_427; setp.eq.s16 %p35, %rs7, 24; @%p35 bra $L__BB148_67; bra.uni $L__BB148_517; $L__BB148_67: ld.const.u32 %r496, [kernel_params+1564]; setp.ne.s32 %p385, %r496, 0; @%p385 bra $L__BB148_69; ld.const.u32 %r497, [kernel_params+1568]; setp.eq.s32 %p386, %r497, 0; shr.u32 %r498, %r141, 3; and.b32 %r499, %r498, 1; setp.eq.b32 %p387, %r499, 1; and.pred %p388, %p387, %p386; @%p388 bra $L__BB148_517; $L__BB148_69: ld.local.f32 %f2726, [%rd4+256]; mov.u64 %rd774, 0; mul.ftz.f32 %f2727, %f2, %f2726; ld.local.f32 %f2728, [%rd4+260]; mul.ftz.f32 %f2729, %f2, %f2728; ld.local.f32 %f2730, [%rd4+264]; mul.ftz.f32 %f2731, %f2, %f2730; mov.f32 %f2732, 0f00000000; max.ftz.f32 %f493, %f2727, %f2732; max.ftz.f32 %f494, %f2729, %f2732; max.ftz.f32 %f495, %f2731, %f2732; add.ftz.f32 %f2733, %f493, %f494; add.ftz.f32 %f2734, %f495, %f2733; mul.ftz.f32 %f2735, %f2734, 0f3EAAAAAB; abs.ftz.f32 %f496, %f2735; setp.ltu.ftz.f32 %p389, %f496, 0f3727C5AC; mov.u64 %rd775, %rd774; @%p389 bra $L__BB148_74; ld.local.u32 %r125, [%rd4+252]; mov.u64 %rd774, 0; setp.eq.s32 %p390, %r125, 0; mov.u64 %rd772, %rd774; mov.u64 %rd773, %rd774; @%p390 bra $L__BB148_72; ld.local.u32 %r500, [%rd4+248]; mul.wide.s32 %rd596, %r500, 80; add.s64 %rd597, %rd1, %rd596; add.s64 %rd772, %rd597, 304; add.s64 %rd598, %rd230, %rd596; add.s64 %rd773, %rd598, 304; mov.b32 %r501, %f495; mov.b32 %r502, %f494; mov.b32 %r503, %f493; mov.u32 %r504, 0; st.local.v4.u32 [%rd597+304], {%r503, %r502, %r501, %r504}; add.s32 %r505, %r500, 1; st.local.u32 [%rd4+248], %r505; add.s32 %r506, %r125, -1; st.local.u32 [%rd4+252], %r506; $L__BB148_72: setp.eq.s64 %p391, %rd773, 0; mov.u64 %rd775, %rd774; @%p391 bra $L__BB148_74; st.local.f32 [%rd772+16], %f496; mov.u64 %rd774, %rd772; mov.u64 %rd775, %rd773; $L__BB148_74: setp.eq.s64 %p392, %rd775, 0; @%p392 bra $L__BB148_517; ld.local.u32 %r126, [%rd4+252]; setp.lt.s32 %p393, %r126, 1; @%p393 bra $L__BB148_457; add.s32 %r507, %r126, -1; st.local.u32 [%rd4+252], %r507; ld.local.s32 %rd601, [%rd4+248]; cvt.s64.s32 %rd602, %r507; add.s64 %rd603, %rd601, %rd602; mul.lo.s64 %rd604, %rd603, 80; add.s64 %rd605, %rd1, %rd604; add.s64 %rd776, %rd605, 304; add.s64 %rd606, %rd230, %rd604; add.s64 %rd777, %rd606, 304; bra.uni $L__BB148_458; $L__BB148_48: setp.gt.s16 %p37, %rs7, 19; @%p37 bra $L__BB148_61; bra.uni $L__BB148_49; $L__BB148_61: add.s16 %rs11, %rs7, -21; setp.lt.u16 %p38, %rs11, 3; @%p38 bra $L__BB148_462; bra.uni $L__BB148_62; $L__BB148_462: ld.const.u32 %r518, [kernel_params+1568]; setp.eq.s32 %p399, %r518, 0; shr.u32 %r519, %r141, 3; and.b32 %r520, %r519, 1; setp.eq.b32 %p400, %r520, 1; and.pred %p401, %p400, %p399; @%p401 bra $L__BB148_517; ld.local.f32 %f2756, [%rd4+256]; mov.u64 %rd780, 0; mul.ftz.f32 %f2757, %f2, %f2756; ld.local.f32 %f2758, [%rd4+260]; mul.ftz.f32 %f2759, %f2, %f2758; ld.local.f32 %f2760, [%rd4+264]; mul.ftz.f32 %f2761, %f2, %f2760; mov.f32 %f2762, 0f00000000; max.ftz.f32 %f500, %f2757, %f2762; max.ftz.f32 %f501, %f2759, %f2762; max.ftz.f32 %f502, %f2761, %f2762; add.ftz.f32 %f2763, %f500, %f501; add.ftz.f32 %f2764, %f502, %f2763; mul.ftz.f32 %f2765, %f2764, 0f3EAAAAAB; abs.ftz.f32 %f503, %f2765; setp.ltu.ftz.f32 %p402, %f503, 0f3727C5AC; mov.u64 %rd781, %rd780; @%p402 bra $L__BB148_468; ld.local.u32 %r127, [%rd4+252]; mov.u64 %rd780, 0; setp.eq.s32 %p403, %r127, 0; mov.u64 %rd778, %rd780; mov.u64 %rd779, %rd780; @%p403 bra $L__BB148_466; ld.local.u32 %r521, [%rd4+248]; mul.wide.s32 %rd616, %r521, 80; add.s64 %rd617, %rd1, %rd616; add.s64 %rd778, %rd617, 304; add.s64 %rd618, %rd230, %rd616; add.s64 %rd779, %rd618, 304; mov.b32 %r522, %f502; mov.b32 %r523, %f501; mov.b32 %r524, %f500; mov.u32 %r525, 0; st.local.v4.u32 [%rd617+304], {%r524, %r523, %r522, %r525}; add.s32 %r526, %r521, 1; st.local.u32 [%rd4+248], %r526; add.s32 %r527, %r127, -1; st.local.u32 [%rd4+252], %r527; $L__BB148_466: setp.eq.s64 %p404, %rd779, 0; mov.u64 %rd781, %rd780; @%p404 bra $L__BB148_468; st.local.f32 [%rd778+16], %f503; mov.u64 %rd780, %rd778; mov.u64 %rd781, %rd779; $L__BB148_468: setp.eq.s64 %p405, %rd781, 0; @%p405 bra $L__BB148_517; add.s64 %rd185, %rd780, 20; mov.u64 %rd621, 0; st.local.f32 [%rd780+20], %f55; st.local.f32 [%rd780+24], %f56; st.local.f32 [%rd780+28], %f57; mov.u32 %r528, 0; st.local.v2.f32 [%rd780+56], {%f2762, %f2762}; st.local.u32 [%rd780+64], %r528; st.local.u64 [%rd780+48], %rd621; mov.f32 %f2767, 0f3727C5AC; max.ftz.f32 %f3057, %f63, %f2767; ld.local.u32 %r529, [%rd4+40]; and.b32 %r530, %r529, 1; setp.eq.b32 %p406, %r530, 1; mov.pred %p407, 0; xor.pred %p408, %p406, %p407; not.pred %p409, %p408; @%p409 bra $L__BB148_471; rcp.approx.ftz.f32 %f3057, %f3057; $L__BB148_471: setp.eq.s32 %p410, %r2, 21; @%p410 bra $L__BB148_475; bra.uni $L__BB148_472; $L__BB148_475: st.local.v2.f32 [%rd185+12], {%f2762, %f2762}; st.local.f32 [%rd185+20], %f3057; mov.u32 %r537, 21; st.local.u32 [%rd185+-8], %r537; ld.local.u32 %r538, [%rd4+40]; or.b32 %r539, %r538, 4; st.local.u32 [%rd4+40], %r539; bra.uni $L__BB148_517; $L__BB148_88: add.s16 %rs9, %rs7, -32; setp.lt.u16 %p27, %rs9, 3; @%p27 bra $L__BB148_290; bra.uni $L__BB148_89; $L__BB148_290: ld.local.v2.u32 {%r330, %r331}, [%rd4+252]; mov.u64 %rd746, 0; mov.b32 %f1892, %r331; mul.ftz.f32 %f340, %f2, %f1892; ld.local.f32 %f1893, [%rd4+260]; mul.ftz.f32 %f341, %f2, %f1893; ld.local.f32 %f1894, [%rd4+264]; mul.ftz.f32 %f342, %f2, %f1894; setp.eq.s32 %p240, %r330, 0; mov.u64 %rd744, %rd746; mov.u64 %rd745, %rd746; @%p240 bra $L__BB148_292; ld.local.u32 %r333, [%rd4+248]; mul.wide.s32 %rd459, %r333, 80; add.s64 %rd460, %rd1, %rd459; add.s64 %rd744, %rd460, 304; add.s64 %rd461, %rd230, %rd459; add.s64 %rd745, %rd461, 304; mov.b32 %r334, %f342; mov.b32 %r335, %f341; mov.b32 %r336, %f340; mov.u32 %r337, 0; st.local.v4.u32 [%rd460+304], {%r336, %r335, %r334, %r337}; add.s32 %r338, %r333, 1; st.local.u32 [%rd4+248], %r338; add.s32 %r339, %r330, -1; st.local.u32 [%rd4+252], %r339; $L__BB148_292: setp.eq.s64 %p241, %rd745, 0; @%p241 bra $L__BB148_294; add.ftz.f32 %f1895, %f340, %f341; add.ftz.f32 %f1896, %f342, %f1895; mul.ftz.f32 %f1897, %f1896, 0f3EAAAAAB; abs.ftz.f32 %f1898, %f1897; st.local.f32 [%rd744+16], %f1898; setp.ge.ftz.f32 %p242, %f1898, 0f3727C5AC; selp.b64 %rd746, %rd745, 0, %p242; $L__BB148_294: setp.eq.s64 %p243, %rd746, 0; @%p243 bra $L__BB148_517; and.b32 %r341, %r141, 4096; setp.eq.s32 %p244, %r341, 0; selp.f32 %f1900, %f60, 0f00000000, %p244; mov.f32 %f3029, 0f00000000; ld.local.f32 %f1901, [%rd6]; mul.ftz.f32 %f1902, %f1900, %f1901; ld.local.f32 %f1903, [%rd6+4]; mul.ftz.f32 %f343, %f1900, %f1903; ld.local.f32 %f1904, [%rd6+8]; mul.ftz.f32 %f344, %f1900, %f1904; add.s64 %rd111, %rd746, 32; st.f32 [%rd746+32], %f1902; st.f32 [%rd746+36], %f343; st.f32 [%rd746+40], %f344; ld.local.f32 %f1905, [%rd4+256]; ld.local.f32 %f1906, [%rd4+260]; ld.local.f32 %f1907, [%rd4+264]; st.f32 [%rd746+44], %f1905; st.f32 [%rd746+48], %f1906; st.f32 [%rd746+52], %f1907; st.f32 [%rd746+20], %f55; st.f32 [%rd746+24], %f56; st.f32 [%rd746+28], %f57; mov.u32 %r342, 2139095039; st.u32 [%rd746+56], %r342; ld.local.f32 %f1908, [%rd7]; max.ftz.f32 %f1909, %f1908, %f3029; mov.f32 %f1910, 0f3F666666; min.ftz.f32 %f1911, %f1909, %f1910; st.f32 [%rd746+60], %f1911; setp.geu.ftz.f32 %p245, %f1902, 0f322BCC77; mov.u32 %r646, 3; mov.f32 %f3027, %f3029; @%p245 bra $L__BB148_297; ld.f32 %f3027, [%rd111+-32]; mov.u32 %r344, 0; st.u32 [%rd111+-32], %r344; st.u32 [%rd111], %r344; mov.u32 %r646, 2; $L__BB148_297: setp.geu.ftz.f32 %p246, %f343, 0f322BCC77; mov.f32 %f3028, %f3029; @%p246 bra $L__BB148_299; ld.f32 %f3028, [%rd111+-28]; mov.u32 %r345, 0; st.u32 [%rd111+-28], %r345; st.u32 [%rd111+4], %r345; add.s32 %r646, %r646, -1; $L__BB148_299: setp.geu.ftz.f32 %p247, %f344, 0f322BCC77; @%p247 bra $L__BB148_301; ld.f32 %f3029, [%rd111+-24]; mov.u32 %r346, 0; st.u32 [%rd111+-24], %r346; st.u32 [%rd111+8], %r346; add.s32 %r646, %r646, -1; $L__BB148_301: setp.gt.s32 %p248, %r646, 2; mov.u32 %r648, 0; @%p248 bra $L__BB148_309; mov.f32 %f1914, 0f00000000; max.ftz.f32 %f351, %f3027, %f1914; max.ftz.f32 %f352, %f3028, %f1914; add.ftz.f32 %f1915, %f351, %f352; max.ftz.f32 %f353, %f3029, %f1914; add.ftz.f32 %f1916, %f353, %f1915; mul.ftz.f32 %f1917, %f1916, 0f3EAAAAAB; abs.ftz.f32 %f354, %f1917; setp.ltu.ftz.f32 %p249, %f354, 0f3727C5AC; mov.u64 %rd749, 0; mov.u64 %rd750, %rd749; @%p249 bra $L__BB148_307; ld.local.u32 %r63, [%rd4+252]; mov.u64 %rd749, 0; setp.eq.s32 %p250, %r63, 0; mov.u64 %rd747, %rd749; mov.u64 %rd748, %rd749; @%p250 bra $L__BB148_305; ld.local.u32 %r348, [%rd4+248]; mul.wide.s32 %rd467, %r348, 80; add.s64 %rd468, %rd1, %rd467; add.s64 %rd747, %rd468, 304; add.s64 %rd469, %rd230, %rd467; add.s64 %rd748, %rd469, 304; mov.b32 %r349, %f353; mov.b32 %r350, %f352; mov.b32 %r351, %f351; mov.u32 %r352, 0; st.local.v4.u32 [%rd468+304], {%r351, %r350, %r349, %r352}; add.s32 %r353, %r348, 1; st.local.u32 [%rd4+248], %r353; add.s32 %r354, %r63, -1; st.local.u32 [%rd4+252], %r354; $L__BB148_305: setp.eq.s64 %p251, %rd748, 0; mov.u64 %rd750, %rd749; @%p251 bra $L__BB148_307; st.local.f32 [%rd747+16], %f354; mov.u64 %rd749, %rd747; mov.u64 %rd750, %rd748; $L__BB148_307: setp.eq.s64 %p252, %rd750, 0; @%p252 bra $L__BB148_309; ld.f32 %f1918, [%rd111+-12]; ld.f32 %f1919, [%rd111+-8]; ld.f32 %f1920, [%rd111+-4]; st.local.f32 [%rd749+20], %f1918; st.local.f32 [%rd749+24], %f1919; st.local.f32 [%rd749+28], %f1920; mov.u32 %r357, 2; st.local.u32 [%rd749+12], %r357; mov.u32 %r648, 12; $L__BB148_309: st.u32 [%rd111+-20], %r2; setp.gt.s32 %p253, %r646, 0; @%p253 bra $L__BB148_311; bra.uni $L__BB148_310; $L__BB148_311: ld.f32 %f1921, [%rd111+-28]; ld.f32 %f1922, [%rd111+-32]; add.ftz.f32 %f1923, %f1922, %f1921; ld.f32 %f1924, [%rd111+-24]; add.ftz.f32 %f1925, %f1924, %f1923; mul.ftz.f32 %f1926, %f1925, 0f3EAAAAAB; abs.ftz.f32 %f1927, %f1926; cvt.rn.f32.s32 %f1928, %r646; mul.ftz.f32 %f1929, %f1927, %f1928; st.f32 [%rd111+-16], %f1929; or.b32 %r359, %r2, 2; setp.eq.s32 %p254, %r359, 34; @%p254 bra $L__BB148_322; bra.uni $L__BB148_312; $L__BB148_322: ld.f32 %f2483, [%rd111]; mul.ftz.f32 %f3033, %f2483, 0f3DA2F983; ld.f32 %f2484, [%rd111+4]; mul.ftz.f32 %f3034, %f2484, 0f3DA2F983; ld.f32 %f2485, [%rd111+8]; mul.ftz.f32 %f3035, %f2485, 0f3DA2F983; bra.uni $L__BB148_323; $L__BB148_24: setp.eq.s16 %p46, %rs7, 2; @%p46 bra $L__BB148_507; setp.eq.s16 %p47, %rs7, 7; @%p47 bra $L__BB148_53; setp.eq.s16 %p48, %rs7, 8; @%p48 bra $L__BB148_27; bra.uni $L__BB148_517; $L__BB148_27: ld.local.f32 %f2922, [%rd4+256]; mov.u64 %rd798, 0; mul.ftz.f32 %f2923, %f2, %f2922; ld.local.f32 %f2924, [%rd4+260]; mul.ftz.f32 %f2925, %f2, %f2924; ld.local.f32 %f2926, [%rd4+264]; mul.ftz.f32 %f2927, %f2, %f2926; mov.f32 %f2928, 0f00000000; max.ftz.f32 %f535, %f2923, %f2928; max.ftz.f32 %f536, %f2925, %f2928; max.ftz.f32 %f537, %f2927, %f2928; add.ftz.f32 %f2929, %f535, %f536; add.ftz.f32 %f2930, %f537, %f2929; mul.ftz.f32 %f2931, %f2930, 0f3EAAAAAB; abs.ftz.f32 %f538, %f2931; setp.ltu.ftz.f32 %p447, %f538, 0f3727C5AC; mov.u64 %rd799, %rd798; @%p447 bra $L__BB148_32; ld.local.u32 %r132, [%rd4+252]; mov.u64 %rd798, 0; setp.eq.s32 %p448, %r132, 0; mov.u64 %rd796, %rd798; mov.u64 %rd797, %rd798; @%p448 bra $L__BB148_30; ld.local.u32 %r600, [%rd4+248]; mul.wide.s32 %rd668, %r600, 80; add.s64 %rd669, %rd1, %rd668; add.s64 %rd796, %rd669, 304; add.s64 %rd670, %rd230, %rd668; add.s64 %rd797, %rd670, 304; mov.b32 %r601, %f537; mov.b32 %r602, %f536; mov.b32 %r603, %f535; mov.u32 %r604, 0; st.local.v4.u32 [%rd669+304], {%r603, %r602, %r601, %r604}; add.s32 %r605, %r600, 1; st.local.u32 [%rd4+248], %r605; add.s32 %r606, %r132, -1; st.local.u32 [%rd4+252], %r606; $L__BB148_30: setp.eq.s64 %p449, %rd797, 0; mov.u64 %rd799, %rd798; @%p449 bra $L__BB148_32; st.local.f32 [%rd796+16], %f538; mov.u64 %rd798, %rd796; mov.u64 %rd799, %rd797; $L__BB148_32: setp.eq.s64 %p450, %rd799, 0; @%p450 bra $L__BB148_517; st.local.f32 [%rd798+20], %f55; st.local.f32 [%rd798+24], %f56; st.local.f32 [%rd798+28], %f57; mov.u32 %r607, 8; st.local.u32 [%rd798+12], %r607; ld.local.u32 %r608, [%rd4+40]; or.b32 %r609, %r608, 12; st.local.u32 [%rd4+40], %r609; bra.uni $L__BB148_517; $L__BB148_77: setp.eq.s16 %p31, %rs7, 28; @%p31 bra $L__BB148_427; setp.eq.s16 %p32, %rs7, 29; @%p32 bra $L__BB148_341; setp.eq.s16 %p33, %rs7, 30; @%p33 bra $L__BB148_80; bra.uni $L__BB148_517; $L__BB148_49: setp.eq.s16 %p40, %rs7, 16; @%p40 bra $L__BB148_36; setp.eq.s16 %p41, %rs7, 17; @%p41 bra $L__BB148_482; setp.eq.s16 %p42, %rs7, 19; @%p42 bra $L__BB148_52; bra.uni $L__BB148_517; $L__BB148_52: ld.const.u32 %r554, [kernel_params+1564]; setp.eq.s32 %p419, %r554, 0; shr.u32 %r555, %r141, 3; and.b32 %r556, %r555, 1; setp.eq.b32 %p420, %r556, 1; and.pred %p421, %p420, %p419; @%p421 bra $L__BB148_517; $L__BB148_53: ld.local.f32 %f2932, [%rd4+256]; mov.u64 %rd802, 0; mul.ftz.f32 %f2933, %f2, %f2932; ld.local.f32 %f2934, [%rd4+260]; mul.ftz.f32 %f2935, %f2, %f2934; ld.local.f32 %f2936, [%rd4+264]; mul.ftz.f32 %f2937, %f2, %f2936; mov.f32 %f2938, 0f00000000; max.ftz.f32 %f539, %f2933, %f2938; max.ftz.f32 %f540, %f2935, %f2938; max.ftz.f32 %f541, %f2937, %f2938; add.ftz.f32 %f2939, %f539, %f540; add.ftz.f32 %f2940, %f541, %f2939; mul.ftz.f32 %f2941, %f2940, 0f3EAAAAAB; abs.ftz.f32 %f542, %f2941; setp.ltu.ftz.f32 %p451, %f542, 0f3727C5AC; mov.u64 %rd803, %rd802; @%p451 bra $L__BB148_58; ld.local.u32 %r133, [%rd4+252]; mov.u64 %rd802, 0; setp.eq.s32 %p452, %r133, 0; mov.u64 %rd800, %rd802; mov.u64 %rd801, %rd802; @%p452 bra $L__BB148_56; ld.local.u32 %r610, [%rd4+248]; mul.wide.s32 %rd677, %r610, 80; add.s64 %rd678, %rd1, %rd677; add.s64 %rd800, %rd678, 304; add.s64 %rd679, %rd230, %rd677; add.s64 %rd801, %rd679, 304; mov.b32 %r611, %f541; mov.b32 %r612, %f540; mov.b32 %r613, %f539; mov.u32 %r614, 0; st.local.v4.u32 [%rd678+304], {%r613, %r612, %r611, %r614}; add.s32 %r615, %r610, 1; st.local.u32 [%rd4+248], %r615; add.s32 %r616, %r133, -1; st.local.u32 [%rd4+252], %r616; $L__BB148_56: setp.eq.s64 %p453, %rd801, 0; mov.u64 %rd803, %rd802; @%p453 bra $L__BB148_58; st.local.f32 [%rd800+16], %f542; mov.u64 %rd802, %rd800; mov.u64 %rd803, %rd801; $L__BB148_58: setp.eq.s64 %p454, %rd803, 0; @%p454 bra $L__BB148_517; add.s64 %rd222, %rd802, 20; st.local.f32 [%rd802+20], %f55; st.local.f32 [%rd802+24], %f56; st.local.f32 [%rd802+28], %f57; st.local.f32 [%rd802+32], %f60; st.local.f32 [%rd802+36], %f63; cvt.ftz.sat.f32.f32 %f543, %f60; cvt.ftz.sat.f32.f32 %f544, %f63; setp.eq.s32 %p455, %r2, 7; @%p455 bra $L__BB148_506; bra.uni $L__BB148_60; $L__BB148_506: mov.u32 %r620, 7; st.local.u32 [%rd222+-8], %r620; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r621, [%rd4+40]; or.b32 %r622, %r621, 12; st.local.u32 [%rd4+40], %r622; bra.uni $L__BB148_517; $L__BB148_36: ld.const.u32 %r567, [kernel_params+1564]; setp.eq.s32 %p426, %r567, 0; shr.u32 %r568, %r141, 3; and.b32 %r569, %r568, 1; setp.eq.b32 %p427, %r569, 1; and.pred %p428, %p427, %p426; @%p428 bra $L__BB148_517; ld.local.f32 %f2831, [%rd4+256]; mov.u64 %rd792, 0; mul.ftz.f32 %f2832, %f2, %f2831; ld.local.f32 %f2833, [%rd4+260]; mul.ftz.f32 %f2834, %f2, %f2833; ld.local.f32 %f2835, [%rd4+264]; mul.ftz.f32 %f2836, %f2, %f2835; mov.f32 %f2837, 0f00000000; max.ftz.f32 %f519, %f2832, %f2837; max.ftz.f32 %f520, %f2834, %f2837; max.ftz.f32 %f521, %f2836, %f2837; add.ftz.f32 %f2838, %f519, %f520; add.ftz.f32 %f2839, %f521, %f2838; mul.ftz.f32 %f2840, %f2839, 0f3EAAAAAB; abs.ftz.f32 %f522, %f2840; setp.ltu.ftz.f32 %p429, %f522, 0f3727C5AC; mov.u64 %rd793, %rd792; @%p429 bra $L__BB148_42; ld.local.u32 %r130, [%rd4+252]; mov.u64 %rd792, 0; setp.eq.s32 %p430, %r130, 0; mov.u64 %rd790, %rd792; mov.u64 %rd791, %rd792; @%p430 bra $L__BB148_40; ld.local.u32 %r570, [%rd4+248]; mul.wide.s32 %rd645, %r570, 80; add.s64 %rd646, %rd1, %rd645; add.s64 %rd790, %rd646, 304; add.s64 %rd647, %rd230, %rd645; add.s64 %rd791, %rd647, 304; mov.b32 %r571, %f521; mov.b32 %r572, %f520; mov.b32 %r573, %f519; mov.u32 %r574, 0; st.local.v4.u32 [%rd646+304], {%r573, %r572, %r571, %r574}; add.s32 %r575, %r570, 1; st.local.u32 [%rd4+248], %r575; add.s32 %r576, %r130, -1; st.local.u32 [%rd4+252], %r576; $L__BB148_40: setp.eq.s64 %p431, %rd791, 0; mov.u64 %rd793, %rd792; @%p431 bra $L__BB148_42; st.local.f32 [%rd790+16], %f522; mov.u64 %rd792, %rd790; mov.u64 %rd793, %rd791; $L__BB148_42: setp.eq.s64 %p432, %rd793, 0; @%p432 bra $L__BB148_517; mul.ftz.f32 %f3058, %f60, %f60; add.s64 %rd205, %rd792, 20; mov.u64 %rd794, 0; st.local.f32 [%rd792+20], %f55; st.local.f32 [%rd792+24], %f56; st.local.f32 [%rd792+28], %f57; mov.u32 %r577, 1065353216; st.local.u32 [%rd792+40], %r577; st.local.u64 [%rd792+48], %rd794; setp.eq.s32 %p433, %r144, 255; @%p433 bra $L__BB148_490; bra.uni $L__BB148_44; $L__BB148_490: mov.u32 %r578, 0; st.local.v2.f32 [%rd205+36], {%f2837, %f2837}; st.local.u32 [%rd205+44], %r578; st.local.v2.f32 [%rd205+12], {%f3058, %f3058}; mov.f32 %f3059, %f3058; bra.uni $L__BB148_491; $L__BB148_427: ld.const.u32 %r121, [kernel_params+1564]; setp.ne.s32 %p358, %r121, 0; @%p358 bra $L__BB148_429; ld.const.u32 %r454, [kernel_params+1568]; setp.eq.s32 %p359, %r454, 0; shr.u32 %r455, %r141, 3; and.b32 %r456, %r455, 1; setp.eq.b32 %p360, %r456, 1; and.pred %p361, %p360, %p359; @%p361 bra $L__BB148_517; $L__BB148_429: ld.local.f32 %f2682, [%rd4+256]; mul.ftz.f32 %f470, %f2, %f2682; ld.local.f32 %f2683, [%rd4+260]; mul.ftz.f32 %f471, %f2, %f2683; ld.local.f32 %f2684, [%rd4+264]; mul.ftz.f32 %f472, %f2, %f2684; mov.f32 %f2685, 0f3727C5AC; max.ftz.f32 %f3054, %f63, %f2685; ld.local.u32 %r457, [%rd4+40]; and.b32 %r458, %r457, 1; setp.eq.b32 %p362, %r458, 1; mov.pred %p363, 0; xor.pred %p364, %p362, %p363; not.pred %p365, %p364; @%p365 bra $L__BB148_431; rcp.approx.ftz.f32 %f3054, %f3054; $L__BB148_431: ld.local.f32 %f2687, [%rd4+24]; ld.local.f32 %f2688, [%rd4+28]; mul.ftz.f32 %f2689, %f56, %f2688; fma.rn.ftz.f32 %f2690, %f55, %f2687, %f2689; ld.local.f32 %f2691, [%rd4+32]; fma.rn.ftz.f32 %f2692, %f57, %f2691, %f2690; abs.ftz.f32 %f476, %f2692; fma.rn.ftz.f32 %f2693, %f3054, %f3054, 0fBF800000; mov.f32 %f3055, 0f3F800000; fma.rn.ftz.f32 %f477, %f476, %f476, %f2693; setp.leu.ftz.f32 %p366, %f477, 0f00000000; @%p366 bra $L__BB148_433; sqrt.approx.ftz.f32 %f2694, %f477; sub.ftz.f32 %f2695, %f2694, %f476; add.ftz.f32 %f2696, %f476, %f2694; div.approx.ftz.f32 %f2697, %f2695, %f2696; fma.rn.ftz.f32 %f2698, %f476, %f2696, 0fBF800000; fma.rn.ftz.f32 %f2699, %f476, %f2695, 0f3F800000; div.approx.ftz.f32 %f2700, %f2698, %f2699; mul.ftz.f32 %f2701, %f2697, 0f3F000000; mul.ftz.f32 %f2702, %f2697, %f2701; fma.rn.ftz.f32 %f2703, %f2700, %f2700, 0f3F800000; mul.ftz.f32 %f3055, %f2702, %f2703; $L__BB148_433: mul.ftz.f32 %f480, %f60, %f60; and.b32 %r122, %r141, 8; shr.u32 %r459, %r141, 3; and.b32 %r460, %r459, 1; setp.eq.b32 %p367, %r460, 1; setp.eq.s32 %p368, %r121, 0; and.pred %p369, %p367, %p368; @%p369 bra $L__BB148_445; mul.ftz.f32 %f2704, %f470, %f3055; mov.f32 %f2705, 0f00000000; max.ftz.f32 %f481, %f2704, %f2705; mul.ftz.f32 %f2706, %f471, %f3055; max.ftz.f32 %f482, %f2706, %f2705; mul.ftz.f32 %f2707, %f472, %f3055; max.ftz.f32 %f483, %f2707, %f2705; add.ftz.f32 %f2708, %f481, %f482; add.ftz.f32 %f2709, %f483, %f2708; mul.ftz.f32 %f2710, %f2709, 0f3EAAAAAB; abs.ftz.f32 %f484, %f2710; setp.ltu.ftz.f32 %p370, %f484, 0f3727C5AC; mov.u64 %rd766, 0; mov.u64 %rd767, %rd766; @%p370 bra $L__BB148_439; ld.local.u32 %r123, [%rd4+252]; mov.u64 %rd766, 0; setp.eq.s32 %p371, %r123, 0; mov.u64 %rd764, %rd766; mov.u64 %rd765, %rd766; @%p371 bra $L__BB148_437; ld.local.u32 %r461, [%rd4+248]; mul.wide.s32 %rd574, %r461, 80; add.s64 %rd575, %rd1, %rd574; add.s64 %rd764, %rd575, 304; add.s64 %rd576, %rd230, %rd574; add.s64 %rd765, %rd576, 304; mov.b32 %r462, %f483; mov.b32 %r463, %f482; mov.b32 %r464, %f481; mov.u32 %r465, 0; st.local.v4.u32 [%rd575+304], {%r464, %r463, %r462, %r465}; add.s32 %r466, %r461, 1; st.local.u32 [%rd4+248], %r466; add.s32 %r467, %r123, -1; st.local.u32 [%rd4+252], %r467; $L__BB148_437: setp.eq.s64 %p372, %rd765, 0; mov.u64 %rd767, %rd766; @%p372 bra $L__BB148_439; st.local.f32 [%rd764+16], %f484; mov.u64 %rd766, %rd764; mov.u64 %rd767, %rd765; $L__BB148_439: setp.eq.s64 %p373, %rd767, 0; @%p373 bra $L__BB148_445; add.s64 %rd160, %rd766, 20; mov.u64 %rd579, 0; st.local.f32 [%rd766+20], %f55; st.local.f32 [%rd766+24], %f56; st.local.f32 [%rd766+28], %f57; mov.u32 %r468, 0; st.local.v2.f32 [%rd766+56], {%f2705, %f2705}; st.local.u32 [%rd766+64], %r468; st.local.u64 [%rd766+48], %rd579; setp.eq.s32 %p374, %r2, 28; @%p374 bra $L__BB148_444; bra.uni $L__BB148_441; $L__BB148_444: st.local.v2.f32 [%rd160+12], {%f2705, %f2705}; st.local.f32 [%rd160+20], %f3054; mov.u32 %r475, 9; st.local.u32 [%rd160+-8], %r475; ld.local.u32 %r476, [%rd4+40]; or.b32 %r477, %r476, 4; st.local.u32 [%rd4+40], %r477; bra.uni $L__BB148_445; $L__BB148_89: setp.eq.s16 %p28, %rs7, 31; @%p28 bra $L__BB148_325; setp.ne.s16 %p29, %rs7, 39; @%p29 bra $L__BB148_517; ld.global.v4.u32 {%r147, %r148, %r149, %r150}, [%rd3+16]; ld.local.f32 %f2977, [%rd5]; ld.local.f32 %f2978, [%rd5+4]; ld.local.f32 %f2979, [%rd5+8]; shr.u32 %r152, %r147, 24; shl.b32 %r153, %r145, 2; cvt.u64.u32 %rd246, %r153; and.b64 %rd247, %rd246, 1020; add.s64 %rd248, %rd2, %rd247; ld.local.f32 %f70, [%rd248]; shr.u32 %r154, %r145, 6; cvt.u64.u32 %rd249, %r154; and.b64 %rd250, %rd249, 1020; add.s64 %rd251, %rd2, %rd250; ld.local.f32 %f71, [%rd251]; shr.u32 %r155, %r145, 14; cvt.u64.u32 %rd252, %r155; and.b64 %rd253, %rd252, 1020; add.s64 %rd254, %rd2, %rd253; ld.local.f32 %f72, [%rd254]; shr.u32 %r156, %r145, 24; mul.wide.u32 %rd255, %r156, 4; add.s64 %rd256, %rd2, %rd255; ld.local.f32 %f73, [%rd256]; shl.b32 %r157, %r146, 2; cvt.u64.u32 %rd257, %r157; and.b64 %rd258, %rd257, 1020; add.s64 %rd259, %rd2, %rd258; ld.local.f32 %f74, [%rd259]; shr.u32 %r158, %r146, 6; cvt.u64.u32 %rd260, %r158; and.b64 %rd261, %rd260, 1020; add.s64 %rd262, %rd2, %rd261; ld.local.f32 %f75, [%rd262]; shr.u32 %r159, %r146, 14; cvt.u64.u32 %rd263, %r159; and.b64 %rd264, %rd263, 1020; add.s64 %rd265, %rd2, %rd264; ld.local.f32 %f76, [%rd265]; shr.u32 %r160, %r146, 24; mul.wide.u32 %rd266, %r160, 4; add.s64 %rd267, %rd2, %rd266; ld.local.f32 %f77, [%rd267]; shr.u32 %r161, %r147, 6; cvt.u64.u32 %rd268, %r161; and.b64 %rd269, %rd268, 1020; add.s64 %rd270, %rd2, %rd269; ld.local.f32 %f78, [%rd270]; shr.u32 %r162, %r147, 14; cvt.u64.u32 %rd271, %r162; and.b64 %rd272, %rd271, 1020; add.s64 %rd273, %rd2, %rd272; mul.wide.u32 %rd274, %r152, 4; add.s64 %rd275, %rd2, %rd274; ld.local.f32 %f79, [%rd275]; shl.b32 %r163, %r147, 2; cvt.u64.u32 %rd276, %r163; and.b64 %rd277, %rd276, 1020; add.s64 %rd278, %rd2, %rd277; ld.local.f32 %f626, [%rd278]; mov.f32 %f627, 0f3727C5AC; max.ftz.f32 %f2980, %f626, %f627; ld.local.f32 %f81, [%rd273]; setp.eq.ftz.f32 %p49, %f81, 0f00000000; @%p49 bra $L__BB148_93; mul.ftz.f32 %f628, %f81, 0f40C90FDB; cos.approx.ftz.f32 %f629, %f628; sin.approx.ftz.f32 %f630, %f628; mov.f32 %f631, 0f3F800000; sub.ftz.f32 %f632, %f631, %f629; mul.ftz.f32 %f633, %f55, %f632; fma.rn.ftz.f32 %f634, %f55, %f633, %f629; mul.ftz.f32 %f635, %f56, %f633; mul.ftz.f32 %f636, %f57, %f630; sub.ftz.f32 %f637, %f635, %f636; mul.ftz.f32 %f638, %f2978, %f637; fma.rn.ftz.f32 %f639, %f2977, %f634, %f638; mul.ftz.f32 %f640, %f57, %f633; mul.ftz.f32 %f641, %f56, %f630; add.ftz.f32 %f642, %f641, %f640; add.ftz.f32 %f643, %f636, %f635; mul.ftz.f32 %f644, %f56, %f632; fma.rn.ftz.f32 %f645, %f56, %f644, %f629; mul.ftz.f32 %f646, %f2978, %f645; fma.rn.ftz.f32 %f647, %f2977, %f643, %f646; mul.ftz.f32 %f648, %f57, %f644; mul.ftz.f32 %f649, %f55, %f630; sub.ftz.f32 %f650, %f648, %f649; sub.ftz.f32 %f651, %f640, %f641; add.ftz.f32 %f652, %f649, %f648; mul.ftz.f32 %f653, %f2978, %f652; fma.rn.ftz.f32 %f654, %f2977, %f651, %f653; mul.ftz.f32 %f655, %f57, %f632; fma.rn.ftz.f32 %f656, %f57, %f655, %f629; fma.rn.ftz.f32 %f2978, %f2979, %f650, %f647; fma.rn.ftz.f32 %f84, %f2979, %f656, %f654; fma.rn.ftz.f32 %f2977, %f2979, %f642, %f639; mov.f32 %f2979, %f84; $L__BB148_93: ld.local.u32 %r164, [%rd4+40]; and.b32 %r165, %r164, 1; setp.eq.b32 %p50, %r165, 1; mov.pred %p51, 0; xor.pred %p52, %p50, %p51; not.pred %p53, %p52; @%p53 bra $L__BB148_95; rcp.approx.ftz.f32 %f2980, %f2980; $L__BB148_95: ld.local.f32 %f90, [%rd4+24]; ld.local.f32 %f91, [%rd4+28]; mul.ftz.f32 %f658, %f56, %f91; fma.rn.ftz.f32 %f659, %f55, %f90, %f658; ld.local.f32 %f92, [%rd4+32]; fma.rn.ftz.f32 %f660, %f57, %f92, %f659; abs.ftz.f32 %f93, %f660; fma.rn.ftz.f32 %f661, %f2980, %f2980, 0fBF800000; mov.f32 %f657, 0f3F800000; fma.rn.ftz.f32 %f94, %f93, %f93, %f661; setp.leu.ftz.f32 %p54, %f94, 0f00000000; mov.f32 %f2981, %f657; @%p54 bra $L__BB148_97; sqrt.approx.ftz.f32 %f662, %f94; sub.ftz.f32 %f663, %f662, %f93; add.ftz.f32 %f664, %f93, %f662; div.approx.ftz.f32 %f665, %f663, %f664; fma.rn.ftz.f32 %f666, %f93, %f664, 0fBF800000; fma.rn.ftz.f32 %f667, %f93, %f663, 0f3F800000; div.approx.ftz.f32 %f668, %f666, %f667; mul.ftz.f32 %f669, %f665, 0f3F000000; mul.ftz.f32 %f670, %f665, %f669; fma.rn.ftz.f32 %f671, %f668, %f668, 0f3F800000; mul.ftz.f32 %f2981, %f670, %f671; $L__BB148_97: cvt.ftz.sat.f32.f32 %f672, %f60; sub.ftz.f32 %f674, %f657, %f672; cvt.ftz.sat.f32.f32 %f675, %f78; sub.ftz.f32 %f676, %f657, %f675; mul.ftz.f32 %f97, %f674, %f676; mul.ftz.f32 %f98, %f675, %f674; sub.ftz.f32 %f99, %f657, %f98; ld.global.v4.u32 {%r166, %r167, %r168, %r169}, [%rd3+32]; setp.eq.s32 %p55, %r166, 255; @%p55 bra $L__BB148_99; bra.uni $L__BB148_98; $L__BB148_99: mov.b32 %f2982, %r167; mov.b32 %f2983, %r168; mov.b32 %f2984, %r169; bra.uni $L__BB148_100; $L__BB148_62: setp.eq.s16 %p39, %rs7, 20; @%p39 bra $L__BB148_80; bra.uni $L__BB148_517; $L__BB148_80: ld.local.f32 %f2770, [%rd4+256]; mov.u64 %rd784, 0; mul.ftz.f32 %f2771, %f2, %f2770; ld.local.f32 %f2772, [%rd4+260]; mul.ftz.f32 %f2773, %f2, %f2772; ld.local.f32 %f2774, [%rd4+264]; mul.ftz.f32 %f2775, %f2, %f2774; mov.f32 %f2776, 0f00000000; max.ftz.f32 %f509, %f2771, %f2776; max.ftz.f32 %f510, %f2773, %f2776; max.ftz.f32 %f511, %f2775, %f2776; add.ftz.f32 %f2777, %f509, %f510; add.ftz.f32 %f2778, %f511, %f2777; mul.ftz.f32 %f2779, %f2778, 0f3EAAAAAB; abs.ftz.f32 %f512, %f2779; setp.ltu.ftz.f32 %p412, %f512, 0f3727C5AC; mov.u64 %rd785, %rd784; @%p412 bra $L__BB148_85; ld.local.u32 %r128, [%rd4+252]; mov.u64 %rd784, 0; setp.eq.s32 %p413, %r128, 0; mov.u64 %rd782, %rd784; mov.u64 %rd783, %rd784; @%p413 bra $L__BB148_83; ld.local.u32 %r540, [%rd4+248]; mul.wide.s32 %rd627, %r540, 80; add.s64 %rd628, %rd1, %rd627; add.s64 %rd782, %rd628, 304; add.s64 %rd629, %rd230, %rd627; add.s64 %rd783, %rd629, 304; mov.b32 %r541, %f511; mov.b32 %r542, %f510; mov.b32 %r543, %f509; mov.u32 %r544, 0; st.local.v4.u32 [%rd628+304], {%r543, %r542, %r541, %r544}; add.s32 %r545, %r540, 1; st.local.u32 [%rd4+248], %r545; add.s32 %r546, %r128, -1; st.local.u32 [%rd4+252], %r546; $L__BB148_83: setp.eq.s64 %p414, %rd783, 0; mov.u64 %rd785, %rd784; @%p414 bra $L__BB148_85; st.local.f32 [%rd782+16], %f512; mov.u64 %rd784, %rd782; mov.u64 %rd785, %rd783; $L__BB148_85: setp.eq.s64 %p415, %rd785, 0; @%p415 bra $L__BB148_517; add.s64 %rd192, %rd784, 20; st.local.f32 [%rd784+20], %f55; st.local.f32 [%rd784+24], %f56; st.local.f32 [%rd784+28], %f57; st.local.f32 [%rd784+44], %f60; st.local.f32 [%rd784+48], %f63; ld.local.f32 %f2780, [%rd6]; neg.ftz.f32 %f2781, %f2780; st.local.f32 [%rd784+52], %f2781; setp.eq.s32 %p416, %r144, 255; @%p416 bra $L__BB148_476; bra.uni $L__BB148_87; $L__BB148_476: ld.local.u8 %rs18, [%rd4+52]; and.b16 %rs19, %rs18, 6; setp.eq.s16 %p417, %rs19, 0; @%p417 bra $L__BB148_478; ld.local.f32 %f2792, [%rd4+104]; ld.local.f32 %f2793, [%rd4+108]; mul.ftz.f32 %f2794, %f2793, %f2793; fma.rn.ftz.f32 %f2795, %f2792, %f2792, %f2794; ld.local.f32 %f2796, [%rd4+112]; fma.rn.ftz.f32 %f2797, %f2796, %f2796, %f2795; sqrt.approx.ftz.f32 %f2798, %f2797; div.approx.ftz.f32 %f2799, %f2792, %f2798; div.approx.ftz.f32 %f2800, %f2793, %f2798; div.approx.ftz.f32 %f2801, %f2796, %f2798; st.local.f32 [%rd192+12], %f2799; st.local.f32 [%rd192+16], %f2800; st.local.f32 [%rd192+20], %f2801; bra.uni $L__BB148_479; $L__BB148_507: ld.local.f32 %f2942, [%rd4+256]; mov.u64 %rd806, 0; mul.ftz.f32 %f2943, %f2, %f2942; ld.local.f32 %f2944, [%rd4+260]; mul.ftz.f32 %f2945, %f2, %f2944; ld.local.f32 %f2946, [%rd4+264]; mul.ftz.f32 %f2947, %f2, %f2946; mov.f32 %f2948, 0f00000000; max.ftz.f32 %f545, %f2943, %f2948; max.ftz.f32 %f546, %f2945, %f2948; max.ftz.f32 %f547, %f2947, %f2948; add.ftz.f32 %f2949, %f545, %f546; add.ftz.f32 %f2950, %f547, %f2949; mul.ftz.f32 %f2951, %f2950, 0f3EAAAAAB; abs.ftz.f32 %f548, %f2951; setp.ltu.ftz.f32 %p456, %f548, 0f3727C5AC; mov.u64 %rd807, %rd806; @%p456 bra $L__BB148_512; ld.local.u32 %r134, [%rd4+252]; mov.u64 %rd806, 0; setp.eq.s32 %p457, %r134, 0; mov.u64 %rd804, %rd806; mov.u64 %rd805, %rd806; @%p457 bra $L__BB148_510; ld.local.u32 %r623, [%rd4+248]; mul.wide.s32 %rd686, %r623, 80; add.s64 %rd687, %rd1, %rd686; add.s64 %rd804, %rd687, 304; add.s64 %rd688, %rd230, %rd686; add.s64 %rd805, %rd688, 304; mov.b32 %r624, %f547; mov.b32 %r625, %f546; mov.b32 %r626, %f545; mov.u32 %r627, 0; st.local.v4.u32 [%rd687+304], {%r626, %r625, %r624, %r627}; add.s32 %r628, %r623, 1; st.local.u32 [%rd4+248], %r628; add.s32 %r629, %r134, -1; st.local.u32 [%rd4+252], %r629; $L__BB148_510: setp.eq.s64 %p458, %rd805, 0; mov.u64 %rd807, %rd806; @%p458 bra $L__BB148_512; st.local.f32 [%rd804+16], %f548; mov.u64 %rd806, %rd804; mov.u64 %rd807, %rd805; $L__BB148_512: setp.eq.s64 %p459, %rd807, 0; @%p459 bra $L__BB148_517; add.s64 %rd229, %rd806, 20; st.local.f32 [%rd806+20], %f55; st.local.f32 [%rd806+24], %f56; st.local.f32 [%rd806+28], %f57; setp.eq.ftz.f32 %p460, %f60, 0f00000000; @%p460 bra $L__BB148_515; bra.uni $L__BB148_514; $L__BB148_515: mov.u32 %r633, 2; st.local.u32 [%rd229+-8], %r633; ld.local.u32 %r634, [%rd4+40]; or.b32 %r635, %r634, 12; st.local.u32 [%rd4+40], %r635; bra.uni $L__BB148_517; $L__BB148_341: ld.global.v4.u32 {%r375, %r376, %r377, %r378}, [%rd3+16]; ld.global.v4.u32 {%r379, %r380, %r381, %r382}, [%rd3+32]; ld.global.u32 %rd121, [%rd3+52]; ld.local.f32 %f378, [%rd4+256]; ld.local.f32 %f379, [%rd4+260]; ld.local.f32 %f380, [%rd4+264]; bfe.u32 %r82, %r144, 8, 8; and.b32 %r83, %r144, 255; setp.eq.s32 %p304, %r83, 255; @%p304 bra $L__BB148_343; bra.uni $L__BB148_342; $L__BB148_343: mov.b32 %f3036, %r145; bra.uni $L__BB148_344; $L__BB148_482: ld.local.f32 %f2816, [%rd4+256]; mov.u64 %rd788, 0; mul.ftz.f32 %f2817, %f2, %f2816; ld.local.f32 %f2818, [%rd4+260]; mul.ftz.f32 %f2819, %f2, %f2818; ld.local.f32 %f2820, [%rd4+264]; mul.ftz.f32 %f2821, %f2, %f2820; mov.f32 %f2822, 0f00000000; max.ftz.f32 %f515, %f2817, %f2822; max.ftz.f32 %f516, %f2819, %f2822; max.ftz.f32 %f517, %f2821, %f2822; add.ftz.f32 %f2823, %f515, %f516; add.ftz.f32 %f2824, %f517, %f2823; mul.ftz.f32 %f2825, %f2824, 0f3EAAAAAB; abs.ftz.f32 %f518, %f2825; setp.ltu.ftz.f32 %p422, %f518, 0f3727C5AC; mov.u64 %rd789, %rd788; @%p422 bra $L__BB148_487; ld.local.u32 %r129, [%rd4+252]; mov.u64 %rd788, 0; setp.eq.s32 %p423, %r129, 0; mov.u64 %rd786, %rd788; mov.u64 %rd787, %rd788; @%p423 bra $L__BB148_485; ld.local.u32 %r557, [%rd4+248]; mul.wide.s32 %rd636, %r557, 80; add.s64 %rd637, %rd1, %rd636; add.s64 %rd786, %rd637, 304; add.s64 %rd638, %rd230, %rd636; add.s64 %rd787, %rd638, 304; mov.b32 %r558, %f517; mov.b32 %r559, %f516; mov.b32 %r560, %f515; mov.u32 %r561, 0; st.local.v4.u32 [%rd637+304], {%r560, %r559, %r558, %r561}; add.s32 %r562, %r557, 1; st.local.u32 [%rd4+248], %r562; add.s32 %r563, %r129, -1; st.local.u32 [%rd4+252], %r563; $L__BB148_485: setp.eq.s64 %p424, %rd787, 0; mov.u64 %rd789, %rd788; @%p424 bra $L__BB148_487; st.local.f32 [%rd786+16], %f518; mov.u64 %rd788, %rd786; mov.u64 %rd789, %rd787; $L__BB148_487: setp.eq.s64 %p425, %rd789, 0; @%p425 bra $L__BB148_517; st.local.f32 [%rd788+20], %f55; st.local.f32 [%rd788+24], %f56; st.local.f32 [%rd788+28], %f57; cvt.ftz.sat.f32.f32 %f2826, %f60; st.local.f32 [%rd788+32], %f2826; mov.f32 %f2827, 0f3C23D70A; max.ftz.f32 %f2828, %f2826, %f2827; mul.ftz.f32 %f2829, %f2828, %f2828; rcp.approx.ftz.f32 %f2830, %f2829; st.local.f32 [%rd788+36], %f2830; mov.u32 %r564, 17; st.local.u32 [%rd788+12], %r564; ld.local.u32 %r565, [%rd4+40]; or.b32 %r566, %r565, 12; st.local.u32 [%rd4+40], %r566; bra.uni $L__BB148_517; $L__BB148_325: ld.local.f32 %f2486, [%rd4+256]; mul.ftz.f32 %f374, %f2, %f2486; ld.local.f32 %f2487, [%rd4+260]; mul.ftz.f32 %f375, %f2, %f2487; ld.local.f32 %f2488, [%rd4+264]; mul.ftz.f32 %f376, %f2, %f2488; add.ftz.f32 %f2489, %f374, %f375; add.ftz.f32 %f2490, %f376, %f2489; mul.ftz.f32 %f2491, %f2490, 0f3EAAAAAB; abs.ftz.f32 %f377, %f2491; setp.ltu.ftz.f32 %p294, %f377, 0f3727C5AC; @%p294 bra $L__BB148_517; ld.local.u32 %r67, [%rd4+40]; and.b32 %r362, %r67, 512; setp.eq.s32 %p295, %r362, 0; @%p295 bra $L__BB148_332; bra.uni $L__BB148_327; $L__BB148_332: or.b32 %r365, %r67, 516; st.local.u32 [%rd4+40], %r365; st.local.f32 [%rd4+280], %f374; st.local.f32 [%rd4+284], %f375; st.local.f32 [%rd4+288], %f376; and.b32 %r71, %r141, 491520; setp.eq.s32 %p299, %r71, 0; @%p299 bra $L__BB148_334; mov.u32 %r651, 1; st.local.u32 [%rd4+252], %r651; bra.uni $L__BB148_335; $L__BB148_44: ld.local.f32 %f524, [%rd5]; ld.local.f32 %f525, [%rd5+8]; ld.local.f32 %f526, [%rd5+4]; st.local.v2.f32 [%rd205+36], {%f524, %f526}; st.local.f32 [%rd205+44], %f525; ld.local.f32 %f527, [%rd6]; setp.eq.ftz.f32 %p434, %f527, 0f00000000; @%p434 bra $L__BB148_46; mul.ftz.f32 %f2841, %f527, 0f40C90FDB; cos.approx.ftz.f32 %f2842, %f2841; sin.approx.ftz.f32 %f2843, %f2841; mov.f32 %f2844, 0f3F800000; sub.ftz.f32 %f2845, %f2844, %f2842; mul.ftz.f32 %f2846, %f55, %f2845; fma.rn.ftz.f32 %f2847, %f55, %f2846, %f2842; mul.ftz.f32 %f2848, %f56, %f2846; mul.ftz.f32 %f2849, %f57, %f2843; sub.ftz.f32 %f2850, %f2848, %f2849; mul.ftz.f32 %f2851, %f526, %f2850; fma.rn.ftz.f32 %f2852, %f524, %f2847, %f2851; mul.ftz.f32 %f2853, %f57, %f2846; mul.ftz.f32 %f2854, %f56, %f2843; add.ftz.f32 %f2855, %f2854, %f2853; add.ftz.f32 %f2856, %f2849, %f2848; mul.ftz.f32 %f2857, %f56, %f2845; fma.rn.ftz.f32 %f2858, %f56, %f2857, %f2842; mul.ftz.f32 %f2859, %f526, %f2858; fma.rn.ftz.f32 %f2860, %f524, %f2856, %f2859; mul.ftz.f32 %f2861, %f57, %f2857; mul.ftz.f32 %f2862, %f55, %f2843; sub.ftz.f32 %f2863, %f2861, %f2862; sub.ftz.f32 %f2864, %f2853, %f2854; add.ftz.f32 %f2865, %f2862, %f2861; mul.ftz.f32 %f2866, %f526, %f2865; fma.rn.ftz.f32 %f2867, %f524, %f2864, %f2866; mul.ftz.f32 %f2868, %f57, %f2845; fma.rn.ftz.f32 %f2869, %f57, %f2868, %f2842; fma.rn.ftz.f32 %f2870, %f525, %f2869, %f2867; fma.rn.ftz.f32 %f2871, %f525, %f2863, %f2860; fma.rn.ftz.f32 %f2872, %f525, %f2855, %f2852; st.local.v2.f32 [%rd205+36], {%f2872, %f2871}; st.local.f32 [%rd205+44], %f2870; $L__BB148_46: mov.f32 %f2873, 0fBF7D70A4; max.ftz.f32 %f2874, %f63, %f2873; mov.f32 %f2875, 0f3F7D70A4; min.ftz.f32 %f528, %f2874, %f2875; setp.lt.ftz.f32 %p435, %f528, 0f00000000; @%p435 bra $L__BB148_489; bra.uni $L__BB148_47; $L__BB148_489: add.ftz.f32 %f2878, %f528, 0f3F800000; div.approx.ftz.f32 %f3059, %f3058, %f2878; mul.ftz.f32 %f3058, %f3058, %f2878; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; bra.uni $L__BB148_491; $L__BB148_310: mov.u32 %r358, 0; st.u32 [%rd111+-16], %r358; bra.uni $L__BB148_324; $L__BB148_87: ld.local.f32 %f2782, [%rd5]; ld.local.f32 %f2783, [%rd5+4]; mul.ftz.f32 %f2784, %f2783, %f2783; fma.rn.ftz.f32 %f2785, %f2782, %f2782, %f2784; ld.local.f32 %f2786, [%rd5+8]; fma.rn.ftz.f32 %f2787, %f2786, %f2786, %f2785; sqrt.approx.ftz.f32 %f2788, %f2787; div.approx.ftz.f32 %f2789, %f2782, %f2788; div.approx.ftz.f32 %f2790, %f2783, %f2788; div.approx.ftz.f32 %f2791, %f2786, %f2788; st.local.f32 [%rd192+12], %f2789; st.local.f32 [%rd192+16], %f2790; st.local.f32 [%rd192+20], %f2791; bra.uni $L__BB148_479; $L__BB148_312: mov.f32 %f1931, 0f3F8147AE; max.ftz.f32 %f1932, %f63, %f1931; mov.f32 %f1933, 0f40733333; min.ftz.f32 %f1934, %f1932, %f1933; rcp.approx.ftz.f32 %f1935, %f1934; mov.f32 %f1936, 0f3F800000; fma.rn.ftz.f32 %f1937, %f1935, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f1938, %f1935, %f1937, 0f3F2B020C; fma.rn.ftz.f32 %f1939, %f1934, 0f3D8240B8, %f1938; add.ftz.f32 %f1940, %f1939, 0f3F800000; mul.ftz.f32 %f1941, %f1940, 0f3FAAAAAB; sub.ftz.f32 %f1942, %f1936, %f1939; div.approx.ftz.f32 %f355, %f1941, %f1942; ld.f32 %f356, [%rd111+12]; setp.lt.ftz.f32 %p255, %f356, 0f38D1B717; mov.f32 %f3032, 0f00000000; mov.f32 %f3030, %f3032; @%p255 bra $L__BB148_315; setp.ge.ftz.f32 %p256, %f356, 0f3F7EB852; mov.f32 %f3030, 0f3F7FFFEF; @%p256 bra $L__BB148_315; sub.ftz.f32 %f1945, %f1936, 0f3F000000; mul.ftz.f32 %f1946, %f1945, 0f40400000; sqrt.approx.ftz.f32 %f1947, %f1946; mul.ftz.f32 %f1948, %f355, %f1947; mul.ftz.f32 %f1949, %f1948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1950, %f1949; add.ftz.f32 %f1951, %f1950, 0f3F800000; mul.ftz.f32 %f1952, %f1951, 0f3E800000; mul.ftz.f32 %f1953, %f1947, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1954, %f1953; mul.ftz.f32 %f1955, %f1954, %f1952; setp.lt.ftz.f32 %p257, %f1955, %f356; selp.f32 %f1956, 0f3F800000, 0f3F000000, %p257; selp.f32 %f1957, 0f3F000000, 0f00000000, %p257; selp.f32 %f1958, 0f3F400000, 0f3E800000, %p257; sub.ftz.f32 %f1959, %f1936, %f1958; mul.ftz.f32 %f1960, %f1959, 0f40400000; sqrt.approx.ftz.f32 %f1961, %f1960; mul.ftz.f32 %f1962, %f1958, 0f3F000000; mul.ftz.f32 %f1963, %f355, %f1961; mul.ftz.f32 %f1964, %f1963, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1965, 0f3F800000; mul.ftz.f32 %f1967, %f1966, %f1962; mul.ftz.f32 %f1968, %f1961, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1969, %f1968; mul.ftz.f32 %f1970, %f1969, %f1967; setp.lt.ftz.f32 %p258, %f1970, %f356; selp.f32 %f1971, %f1956, %f1958, %p258; selp.f32 %f1972, %f1958, %f1957, %p258; add.ftz.f32 %f1973, %f1971, %f1972; mul.ftz.f32 %f1974, %f1973, 0f3F000000; sub.ftz.f32 %f1975, %f1936, %f1974; mul.ftz.f32 %f1976, %f1975, 0f40400000; sqrt.approx.ftz.f32 %f1977, %f1976; mul.ftz.f32 %f1978, %f1974, 0f3F000000; mul.ftz.f32 %f1979, %f355, %f1977; mul.ftz.f32 %f1980, %f1979, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1981, %f1980; add.ftz.f32 %f1982, %f1981, 0f3F800000; mul.ftz.f32 %f1983, %f1982, %f1978; mul.ftz.f32 %f1984, %f1977, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f1985, %f1983; setp.lt.ftz.f32 %p259, %f1986, %f356; selp.f32 %f1987, %f1971, %f1974, %p259; selp.f32 %f1988, %f1974, %f1972, %p259; add.ftz.f32 %f1989, %f1987, %f1988; mul.ftz.f32 %f1990, %f1989, 0f3F000000; sub.ftz.f32 %f1991, %f1936, %f1990; mul.ftz.f32 %f1992, %f1991, 0f40400000; sqrt.approx.ftz.f32 %f1993, %f1992; mul.ftz.f32 %f1994, %f1990, 0f3F000000; mul.ftz.f32 %f1995, %f355, %f1993; mul.ftz.f32 %f1996, %f1995, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1997, %f1996; add.ftz.f32 %f1998, %f1997, 0f3F800000; mul.ftz.f32 %f1999, %f1998, %f1994; mul.ftz.f32 %f2000, %f1993, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2001, %f2000; mul.ftz.f32 %f2002, %f2001, %f1999; setp.lt.ftz.f32 %p260, %f2002, %f356; selp.f32 %f2003, %f1987, %f1990, %p260; selp.f32 %f2004, %f1990, %f1988, %p260; add.ftz.f32 %f2005, %f2003, %f2004; mul.ftz.f32 %f2006, %f2005, 0f3F000000; sub.ftz.f32 %f2007, %f1936, %f2006; mul.ftz.f32 %f2008, %f2007, 0f40400000; sqrt.approx.ftz.f32 %f2009, %f2008; mul.ftz.f32 %f2010, %f2006, 0f3F000000; mul.ftz.f32 %f2011, %f355, %f2009; mul.ftz.f32 %f2012, %f2011, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2013, %f2012; add.ftz.f32 %f2014, %f2013, 0f3F800000; mul.ftz.f32 %f2015, %f2014, %f2010; mul.ftz.f32 %f2016, %f2009, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2017, %f2016; mul.ftz.f32 %f2018, %f2017, %f2015; setp.lt.ftz.f32 %p261, %f2018, %f356; selp.f32 %f2019, %f2003, %f2006, %p261; selp.f32 %f2020, %f2006, %f2004, %p261; add.ftz.f32 %f2021, %f2019, %f2020; mul.ftz.f32 %f2022, %f2021, 0f3F000000; sub.ftz.f32 %f2023, %f1936, %f2022; mul.ftz.f32 %f2024, %f2023, 0f40400000; sqrt.approx.ftz.f32 %f2025, %f2024; mul.ftz.f32 %f2026, %f2022, 0f3F000000; mul.ftz.f32 %f2027, %f355, %f2025; mul.ftz.f32 %f2028, %f2027, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2029, %f2028; add.ftz.f32 %f2030, %f2029, 0f3F800000; mul.ftz.f32 %f2031, %f2030, %f2026; mul.ftz.f32 %f2032, %f2025, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2033, %f2032; mul.ftz.f32 %f2034, %f2033, %f2031; setp.lt.ftz.f32 %p262, %f2034, %f356; selp.f32 %f2035, %f2019, %f2022, %p262; selp.f32 %f2036, %f2022, %f2020, %p262; add.ftz.f32 %f2037, %f2035, %f2036; mul.ftz.f32 %f2038, %f2037, 0f3F000000; sub.ftz.f32 %f2039, %f1936, %f2038; mul.ftz.f32 %f2040, %f2039, 0f40400000; sqrt.approx.ftz.f32 %f2041, %f2040; mul.ftz.f32 %f2042, %f2038, 0f3F000000; mul.ftz.f32 %f2043, %f355, %f2041; mul.ftz.f32 %f2044, %f2043, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2045, 0f3F800000; mul.ftz.f32 %f2047, %f2046, %f2042; mul.ftz.f32 %f2048, %f2041, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2049, %f2048; mul.ftz.f32 %f2050, %f2049, %f2047; setp.lt.ftz.f32 %p263, %f2050, %f356; selp.f32 %f2051, %f2035, %f2038, %p263; selp.f32 %f2052, %f2038, %f2036, %p263; add.ftz.f32 %f2053, %f2051, %f2052; mul.ftz.f32 %f2054, %f2053, 0f3F000000; sub.ftz.f32 %f2055, %f1936, %f2054; mul.ftz.f32 %f2056, %f2055, 0f40400000; sqrt.approx.ftz.f32 %f2057, %f2056; mul.ftz.f32 %f2058, %f2054, 0f3F000000; mul.ftz.f32 %f2059, %f355, %f2057; mul.ftz.f32 %f2060, %f2059, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2061, %f2060; add.ftz.f32 %f2062, %f2061, 0f3F800000; mul.ftz.f32 %f2063, %f2062, %f2058; mul.ftz.f32 %f2064, %f2057, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2065, %f2064; mul.ftz.f32 %f2066, %f2065, %f2063; setp.lt.ftz.f32 %p264, %f2066, %f356; selp.f32 %f2067, %f2051, %f2054, %p264; selp.f32 %f2068, %f2054, %f2052, %p264; add.ftz.f32 %f2069, %f2067, %f2068; mul.ftz.f32 %f2070, %f2069, 0f3F000000; sub.ftz.f32 %f2071, %f1936, %f2070; mul.ftz.f32 %f2072, %f2071, 0f40400000; sqrt.approx.ftz.f32 %f2073, %f2072; mul.ftz.f32 %f2074, %f2070, 0f3F000000; mul.ftz.f32 %f2075, %f355, %f2073; mul.ftz.f32 %f2076, %f2075, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2077, %f2076; add.ftz.f32 %f2078, %f2077, 0f3F800000; mul.ftz.f32 %f2079, %f2078, %f2074; mul.ftz.f32 %f2080, %f2073, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2081, %f2080; mul.ftz.f32 %f2082, %f2081, %f2079; setp.lt.ftz.f32 %p265, %f2082, %f356; selp.f32 %f2083, %f2067, %f2070, %p265; selp.f32 %f2084, %f2070, %f2068, %p265; add.ftz.f32 %f2085, %f2083, %f2084; mul.ftz.f32 %f2086, %f2085, 0f3F000000; sub.ftz.f32 %f2087, %f1936, %f2086; mul.ftz.f32 %f2088, %f2087, 0f40400000; sqrt.approx.ftz.f32 %f2089, %f2088; mul.ftz.f32 %f2090, %f2086, 0f3F000000; mul.ftz.f32 %f2091, %f355, %f2089; mul.ftz.f32 %f2092, %f2091, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2093, %f2092; add.ftz.f32 %f2094, %f2093, 0f3F800000; mul.ftz.f32 %f2095, %f2094, %f2090; mul.ftz.f32 %f2096, %f2089, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2097, %f2096; mul.ftz.f32 %f2098, %f2097, %f2095; setp.lt.ftz.f32 %p266, %f2098, %f356; selp.f32 %f2099, %f2083, %f2086, %p266; selp.f32 %f2100, %f2086, %f2084, %p266; add.ftz.f32 %f2101, %f2099, %f2100; mul.ftz.f32 %f2102, %f2101, 0f3F000000; sub.ftz.f32 %f2103, %f1936, %f2102; mul.ftz.f32 %f2104, %f2103, 0f40400000; sqrt.approx.ftz.f32 %f2105, %f2104; mul.ftz.f32 %f2106, %f2102, 0f3F000000; mul.ftz.f32 %f2107, %f355, %f2105; mul.ftz.f32 %f2108, %f2107, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2109, %f2108; add.ftz.f32 %f2110, %f2109, 0f3F800000; mul.ftz.f32 %f2111, %f2110, %f2106; mul.ftz.f32 %f2112, %f2105, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2113, %f2112; mul.ftz.f32 %f2114, %f2113, %f2111; setp.lt.ftz.f32 %p267, %f2114, %f356; selp.f32 %f2115, %f2099, %f2102, %p267; selp.f32 %f2116, %f2102, %f2100, %p267; add.ftz.f32 %f2117, %f2115, %f2116; mul.ftz.f32 %f3030, %f2117, 0f3F000000; $L__BB148_315: ld.f32 %f359, [%rd111+16]; setp.lt.ftz.f32 %p268, %f359, 0f38D1B717; mov.f32 %f3031, %f3032; @%p268 bra $L__BB148_318; setp.ge.ftz.f32 %p269, %f359, 0f3F7EB852; mov.f32 %f3031, 0f3F7FFFEF; @%p269 bra $L__BB148_318; sub.ftz.f32 %f2121, %f1936, 0f3F000000; mul.ftz.f32 %f2122, %f2121, 0f40400000; sqrt.approx.ftz.f32 %f2123, %f2122; mul.ftz.f32 %f2124, %f355, %f2123; mul.ftz.f32 %f2125, %f2124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2126, %f2125; add.ftz.f32 %f2127, %f2126, 0f3F800000; mul.ftz.f32 %f2128, %f2127, 0f3E800000; mul.ftz.f32 %f2129, %f2123, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2130, %f2129; mul.ftz.f32 %f2131, %f2130, %f2128; setp.lt.ftz.f32 %p270, %f2131, %f359; selp.f32 %f2132, 0f3F800000, 0f3F000000, %p270; selp.f32 %f2133, 0f3F000000, 0f00000000, %p270; selp.f32 %f2134, 0f3F400000, 0f3E800000, %p270; sub.ftz.f32 %f2135, %f1936, %f2134; mul.ftz.f32 %f2136, %f2135, 0f40400000; sqrt.approx.ftz.f32 %f2137, %f2136; mul.ftz.f32 %f2138, %f2134, 0f3F000000; mul.ftz.f32 %f2139, %f355, %f2137; mul.ftz.f32 %f2140, %f2139, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2141, %f2140; add.ftz.f32 %f2142, %f2141, 0f3F800000; mul.ftz.f32 %f2143, %f2142, %f2138; mul.ftz.f32 %f2144, %f2137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2145, %f2144; mul.ftz.f32 %f2146, %f2145, %f2143; setp.lt.ftz.f32 %p271, %f2146, %f359; selp.f32 %f2147, %f2132, %f2134, %p271; selp.f32 %f2148, %f2134, %f2133, %p271; add.ftz.f32 %f2149, %f2147, %f2148; mul.ftz.f32 %f2150, %f2149, 0f3F000000; sub.ftz.f32 %f2151, %f1936, %f2150; mul.ftz.f32 %f2152, %f2151, 0f40400000; sqrt.approx.ftz.f32 %f2153, %f2152; mul.ftz.f32 %f2154, %f2150, 0f3F000000; mul.ftz.f32 %f2155, %f355, %f2153; mul.ftz.f32 %f2156, %f2155, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2157, %f2156; add.ftz.f32 %f2158, %f2157, 0f3F800000; mul.ftz.f32 %f2159, %f2158, %f2154; mul.ftz.f32 %f2160, %f2153, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2161, %f2160; mul.ftz.f32 %f2162, %f2161, %f2159; setp.lt.ftz.f32 %p272, %f2162, %f359; selp.f32 %f2163, %f2147, %f2150, %p272; selp.f32 %f2164, %f2150, %f2148, %p272; add.ftz.f32 %f2165, %f2163, %f2164; mul.ftz.f32 %f2166, %f2165, 0f3F000000; sub.ftz.f32 %f2167, %f1936, %f2166; mul.ftz.f32 %f2168, %f2167, 0f40400000; sqrt.approx.ftz.f32 %f2169, %f2168; mul.ftz.f32 %f2170, %f2166, 0f3F000000; mul.ftz.f32 %f2171, %f355, %f2169; mul.ftz.f32 %f2172, %f2171, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2173, %f2172; add.ftz.f32 %f2174, %f2173, 0f3F800000; mul.ftz.f32 %f2175, %f2174, %f2170; mul.ftz.f32 %f2176, %f2169, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2177, %f2176; mul.ftz.f32 %f2178, %f2177, %f2175; setp.lt.ftz.f32 %p273, %f2178, %f359; selp.f32 %f2179, %f2163, %f2166, %p273; selp.f32 %f2180, %f2166, %f2164, %p273; add.ftz.f32 %f2181, %f2179, %f2180; mul.ftz.f32 %f2182, %f2181, 0f3F000000; sub.ftz.f32 %f2183, %f1936, %f2182; mul.ftz.f32 %f2184, %f2183, 0f40400000; sqrt.approx.ftz.f32 %f2185, %f2184; mul.ftz.f32 %f2186, %f2182, 0f3F000000; mul.ftz.f32 %f2187, %f355, %f2185; mul.ftz.f32 %f2188, %f2187, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2189, %f2188; add.ftz.f32 %f2190, %f2189, 0f3F800000; mul.ftz.f32 %f2191, %f2190, %f2186; mul.ftz.f32 %f2192, %f2185, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2193, %f2192; mul.ftz.f32 %f2194, %f2193, %f2191; setp.lt.ftz.f32 %p274, %f2194, %f359; selp.f32 %f2195, %f2179, %f2182, %p274; selp.f32 %f2196, %f2182, %f2180, %p274; add.ftz.f32 %f2197, %f2195, %f2196; mul.ftz.f32 %f2198, %f2197, 0f3F000000; sub.ftz.f32 %f2199, %f1936, %f2198; mul.ftz.f32 %f2200, %f2199, 0f40400000; sqrt.approx.ftz.f32 %f2201, %f2200; mul.ftz.f32 %f2202, %f2198, 0f3F000000; mul.ftz.f32 %f2203, %f355, %f2201; mul.ftz.f32 %f2204, %f2203, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2205, %f2204; add.ftz.f32 %f2206, %f2205, 0f3F800000; mul.ftz.f32 %f2207, %f2206, %f2202; mul.ftz.f32 %f2208, %f2201, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2209, %f2208; mul.ftz.f32 %f2210, %f2209, %f2207; setp.lt.ftz.f32 %p275, %f2210, %f359; selp.f32 %f2211, %f2195, %f2198, %p275; selp.f32 %f2212, %f2198, %f2196, %p275; add.ftz.f32 %f2213, %f2211, %f2212; mul.ftz.f32 %f2214, %f2213, 0f3F000000; sub.ftz.f32 %f2215, %f1936, %f2214; mul.ftz.f32 %f2216, %f2215, 0f40400000; sqrt.approx.ftz.f32 %f2217, %f2216; mul.ftz.f32 %f2218, %f2214, 0f3F000000; mul.ftz.f32 %f2219, %f355, %f2217; mul.ftz.f32 %f2220, %f2219, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2221, %f2220; add.ftz.f32 %f2222, %f2221, 0f3F800000; mul.ftz.f32 %f2223, %f2222, %f2218; mul.ftz.f32 %f2224, %f2217, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2225, %f2224; mul.ftz.f32 %f2226, %f2225, %f2223; setp.lt.ftz.f32 %p276, %f2226, %f359; selp.f32 %f2227, %f2211, %f2214, %p276; selp.f32 %f2228, %f2214, %f2212, %p276; add.ftz.f32 %f2229, %f2227, %f2228; mul.ftz.f32 %f2230, %f2229, 0f3F000000; sub.ftz.f32 %f2231, %f1936, %f2230; mul.ftz.f32 %f2232, %f2231, 0f40400000; sqrt.approx.ftz.f32 %f2233, %f2232; mul.ftz.f32 %f2234, %f2230, 0f3F000000; mul.ftz.f32 %f2235, %f355, %f2233; mul.ftz.f32 %f2236, %f2235, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2237, %f2236; add.ftz.f32 %f2238, %f2237, 0f3F800000; mul.ftz.f32 %f2239, %f2238, %f2234; mul.ftz.f32 %f2240, %f2233, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2241, %f2240; mul.ftz.f32 %f2242, %f2241, %f2239; setp.lt.ftz.f32 %p277, %f2242, %f359; selp.f32 %f2243, %f2227, %f2230, %p277; selp.f32 %f2244, %f2230, %f2228, %p277; add.ftz.f32 %f2245, %f2243, %f2244; mul.ftz.f32 %f2246, %f2245, 0f3F000000; sub.ftz.f32 %f2247, %f1936, %f2246; mul.ftz.f32 %f2248, %f2247, 0f40400000; sqrt.approx.ftz.f32 %f2249, %f2248; mul.ftz.f32 %f2250, %f2246, 0f3F000000; mul.ftz.f32 %f2251, %f355, %f2249; mul.ftz.f32 %f2252, %f2251, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2253, %f2252; add.ftz.f32 %f2254, %f2253, 0f3F800000; mul.ftz.f32 %f2255, %f2254, %f2250; mul.ftz.f32 %f2256, %f2249, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2257, %f2256; mul.ftz.f32 %f2258, %f2257, %f2255; setp.lt.ftz.f32 %p278, %f2258, %f359; selp.f32 %f2259, %f2243, %f2246, %p278; selp.f32 %f2260, %f2246, %f2244, %p278; add.ftz.f32 %f2261, %f2259, %f2260; mul.ftz.f32 %f2262, %f2261, 0f3F000000; sub.ftz.f32 %f2263, %f1936, %f2262; mul.ftz.f32 %f2264, %f2263, 0f40400000; sqrt.approx.ftz.f32 %f2265, %f2264; mul.ftz.f32 %f2266, %f2262, 0f3F000000; mul.ftz.f32 %f2267, %f355, %f2265; mul.ftz.f32 %f2268, %f2267, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2269, %f2268; add.ftz.f32 %f2270, %f2269, 0f3F800000; mul.ftz.f32 %f2271, %f2270, %f2266; mul.ftz.f32 %f2272, %f2265, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2273, %f2272; mul.ftz.f32 %f2274, %f2273, %f2271; setp.lt.ftz.f32 %p279, %f2274, %f359; selp.f32 %f2275, %f2259, %f2262, %p279; selp.f32 %f2276, %f2262, %f2260, %p279; add.ftz.f32 %f2277, %f2275, %f2276; mul.ftz.f32 %f2278, %f2277, 0f3F000000; sub.ftz.f32 %f2279, %f1936, %f2278; mul.ftz.f32 %f2280, %f2279, 0f40400000; sqrt.approx.ftz.f32 %f2281, %f2280; mul.ftz.f32 %f2282, %f2278, 0f3F000000; mul.ftz.f32 %f2283, %f355, %f2281; mul.ftz.f32 %f2284, %f2283, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2285, %f2284; add.ftz.f32 %f2286, %f2285, 0f3F800000; mul.ftz.f32 %f2287, %f2286, %f2282; mul.ftz.f32 %f2288, %f2281, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f2289, %f2287; setp.lt.ftz.f32 %p280, %f2290, %f359; selp.f32 %f2291, %f2275, %f2278, %p280; selp.f32 %f2292, %f2278, %f2276, %p280; add.ftz.f32 %f2293, %f2291, %f2292; mul.ftz.f32 %f3031, %f2293, 0f3F000000; $L__BB148_318: ld.f32 %f362, [%rd111+20]; setp.lt.ftz.f32 %p281, %f362, 0f38D1B717; @%p281 bra $L__BB148_321; setp.ge.ftz.f32 %p282, %f362, 0f3F7EB852; mov.f32 %f3032, 0f3F7FFFEF; @%p282 bra $L__BB148_321; sub.ftz.f32 %f2297, %f1936, 0f3F000000; mul.ftz.f32 %f2298, %f2297, 0f40400000; sqrt.approx.ftz.f32 %f2299, %f2298; mul.ftz.f32 %f2300, %f355, %f2299; mul.ftz.f32 %f2301, %f2300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2302, %f2301; add.ftz.f32 %f2303, %f2302, 0f3F800000; mul.ftz.f32 %f2304, %f2303, 0f3E800000; mul.ftz.f32 %f2305, %f2299, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2306, %f2305; mul.ftz.f32 %f2307, %f2306, %f2304; setp.lt.ftz.f32 %p283, %f2307, %f362; selp.f32 %f2308, 0f3F800000, 0f3F000000, %p283; selp.f32 %f2309, 0f3F000000, 0f00000000, %p283; selp.f32 %f2310, 0f3F400000, 0f3E800000, %p283; sub.ftz.f32 %f2311, %f1936, %f2310; mul.ftz.f32 %f2312, %f2311, 0f40400000; sqrt.approx.ftz.f32 %f2313, %f2312; mul.ftz.f32 %f2314, %f2310, 0f3F000000; mul.ftz.f32 %f2315, %f355, %f2313; mul.ftz.f32 %f2316, %f2315, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2317, %f2316; add.ftz.f32 %f2318, %f2317, 0f3F800000; mul.ftz.f32 %f2319, %f2318, %f2314; mul.ftz.f32 %f2320, %f2313, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2321, %f2320; mul.ftz.f32 %f2322, %f2321, %f2319; setp.lt.ftz.f32 %p284, %f2322, %f362; selp.f32 %f2323, %f2308, %f2310, %p284; selp.f32 %f2324, %f2310, %f2309, %p284; add.ftz.f32 %f2325, %f2323, %f2324; mul.ftz.f32 %f2326, %f2325, 0f3F000000; sub.ftz.f32 %f2327, %f1936, %f2326; mul.ftz.f32 %f2328, %f2327, 0f40400000; sqrt.approx.ftz.f32 %f2329, %f2328; mul.ftz.f32 %f2330, %f2326, 0f3F000000; mul.ftz.f32 %f2331, %f355, %f2329; mul.ftz.f32 %f2332, %f2331, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2333, %f2332; add.ftz.f32 %f2334, %f2333, 0f3F800000; mul.ftz.f32 %f2335, %f2334, %f2330; mul.ftz.f32 %f2336, %f2329, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2337, %f2336; mul.ftz.f32 %f2338, %f2337, %f2335; setp.lt.ftz.f32 %p285, %f2338, %f362; selp.f32 %f2339, %f2323, %f2326, %p285; selp.f32 %f2340, %f2326, %f2324, %p285; add.ftz.f32 %f2341, %f2339, %f2340; mul.ftz.f32 %f2342, %f2341, 0f3F000000; sub.ftz.f32 %f2343, %f1936, %f2342; mul.ftz.f32 %f2344, %f2343, 0f40400000; sqrt.approx.ftz.f32 %f2345, %f2344; mul.ftz.f32 %f2346, %f2342, 0f3F000000; mul.ftz.f32 %f2347, %f355, %f2345; mul.ftz.f32 %f2348, %f2347, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2349, %f2348; add.ftz.f32 %f2350, %f2349, 0f3F800000; mul.ftz.f32 %f2351, %f2350, %f2346; mul.ftz.f32 %f2352, %f2345, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2353, %f2352; mul.ftz.f32 %f2354, %f2353, %f2351; setp.lt.ftz.f32 %p286, %f2354, %f362; selp.f32 %f2355, %f2339, %f2342, %p286; selp.f32 %f2356, %f2342, %f2340, %p286; add.ftz.f32 %f2357, %f2355, %f2356; mul.ftz.f32 %f2358, %f2357, 0f3F000000; sub.ftz.f32 %f2359, %f1936, %f2358; mul.ftz.f32 %f2360, %f2359, 0f40400000; sqrt.approx.ftz.f32 %f2361, %f2360; mul.ftz.f32 %f2362, %f2358, 0f3F000000; mul.ftz.f32 %f2363, %f355, %f2361; mul.ftz.f32 %f2364, %f2363, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2365, %f2364; add.ftz.f32 %f2366, %f2365, 0f3F800000; mul.ftz.f32 %f2367, %f2366, %f2362; mul.ftz.f32 %f2368, %f2361, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2369, %f2368; mul.ftz.f32 %f2370, %f2369, %f2367; setp.lt.ftz.f32 %p287, %f2370, %f362; selp.f32 %f2371, %f2355, %f2358, %p287; selp.f32 %f2372, %f2358, %f2356, %p287; add.ftz.f32 %f2373, %f2371, %f2372; mul.ftz.f32 %f2374, %f2373, 0f3F000000; sub.ftz.f32 %f2375, %f1936, %f2374; mul.ftz.f32 %f2376, %f2375, 0f40400000; sqrt.approx.ftz.f32 %f2377, %f2376; mul.ftz.f32 %f2378, %f2374, 0f3F000000; mul.ftz.f32 %f2379, %f355, %f2377; mul.ftz.f32 %f2380, %f2379, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2381, %f2380; add.ftz.f32 %f2382, %f2381, 0f3F800000; mul.ftz.f32 %f2383, %f2382, %f2378; mul.ftz.f32 %f2384, %f2377, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2385, %f2384; mul.ftz.f32 %f2386, %f2385, %f2383; setp.lt.ftz.f32 %p288, %f2386, %f362; selp.f32 %f2387, %f2371, %f2374, %p288; selp.f32 %f2388, %f2374, %f2372, %p288; add.ftz.f32 %f2389, %f2387, %f2388; mul.ftz.f32 %f2390, %f2389, 0f3F000000; sub.ftz.f32 %f2391, %f1936, %f2390; mul.ftz.f32 %f2392, %f2391, 0f40400000; sqrt.approx.ftz.f32 %f2393, %f2392; mul.ftz.f32 %f2394, %f2390, 0f3F000000; mul.ftz.f32 %f2395, %f355, %f2393; mul.ftz.f32 %f2396, %f2395, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2397, %f2396; add.ftz.f32 %f2398, %f2397, 0f3F800000; mul.ftz.f32 %f2399, %f2398, %f2394; mul.ftz.f32 %f2400, %f2393, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f2401, %f2399; setp.lt.ftz.f32 %p289, %f2402, %f362; selp.f32 %f2403, %f2387, %f2390, %p289; selp.f32 %f2404, %f2390, %f2388, %p289; add.ftz.f32 %f2405, %f2403, %f2404; mul.ftz.f32 %f2406, %f2405, 0f3F000000; sub.ftz.f32 %f2407, %f1936, %f2406; mul.ftz.f32 %f2408, %f2407, 0f40400000; sqrt.approx.ftz.f32 %f2409, %f2408; mul.ftz.f32 %f2410, %f2406, 0f3F000000; mul.ftz.f32 %f2411, %f355, %f2409; mul.ftz.f32 %f2412, %f2411, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2413, %f2412; add.ftz.f32 %f2414, %f2413, 0f3F800000; mul.ftz.f32 %f2415, %f2414, %f2410; mul.ftz.f32 %f2416, %f2409, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2417, %f2416; mul.ftz.f32 %f2418, %f2417, %f2415; setp.lt.ftz.f32 %p290, %f2418, %f362; selp.f32 %f2419, %f2403, %f2406, %p290; selp.f32 %f2420, %f2406, %f2404, %p290; add.ftz.f32 %f2421, %f2419, %f2420; mul.ftz.f32 %f2422, %f2421, 0f3F000000; sub.ftz.f32 %f2423, %f1936, %f2422; mul.ftz.f32 %f2424, %f2423, 0f40400000; sqrt.approx.ftz.f32 %f2425, %f2424; mul.ftz.f32 %f2426, %f2422, 0f3F000000; mul.ftz.f32 %f2427, %f355, %f2425; mul.ftz.f32 %f2428, %f2427, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2429, %f2428; add.ftz.f32 %f2430, %f2429, 0f3F800000; mul.ftz.f32 %f2431, %f2430, %f2426; mul.ftz.f32 %f2432, %f2425, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2433, %f2432; mul.ftz.f32 %f2434, %f2433, %f2431; setp.lt.ftz.f32 %p291, %f2434, %f362; selp.f32 %f2435, %f2419, %f2422, %p291; selp.f32 %f2436, %f2422, %f2420, %p291; add.ftz.f32 %f2437, %f2435, %f2436; mul.ftz.f32 %f2438, %f2437, 0f3F000000; sub.ftz.f32 %f2439, %f1936, %f2438; mul.ftz.f32 %f2440, %f2439, 0f40400000; sqrt.approx.ftz.f32 %f2441, %f2440; mul.ftz.f32 %f2442, %f2438, 0f3F000000; mul.ftz.f32 %f2443, %f355, %f2441; mul.ftz.f32 %f2444, %f2443, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2445, %f2444; add.ftz.f32 %f2446, %f2445, 0f3F800000; mul.ftz.f32 %f2447, %f2446, %f2442; mul.ftz.f32 %f2448, %f2441, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2449, %f2448; mul.ftz.f32 %f2450, %f2449, %f2447; setp.lt.ftz.f32 %p292, %f2450, %f362; selp.f32 %f2451, %f2435, %f2438, %p292; selp.f32 %f2452, %f2438, %f2436, %p292; add.ftz.f32 %f2453, %f2451, %f2452; mul.ftz.f32 %f2454, %f2453, 0f3F000000; sub.ftz.f32 %f2455, %f1936, %f2454; mul.ftz.f32 %f2456, %f2455, 0f40400000; sqrt.approx.ftz.f32 %f2457, %f2456; mul.ftz.f32 %f2458, %f2454, 0f3F000000; mul.ftz.f32 %f2459, %f355, %f2457; mul.ftz.f32 %f2460, %f2459, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2461, %f2460; add.ftz.f32 %f2462, %f2461, 0f3F800000; mul.ftz.f32 %f2463, %f2462, %f2458; mul.ftz.f32 %f2464, %f2457, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2465, %f2464; mul.ftz.f32 %f2466, %f2465, %f2463; setp.lt.ftz.f32 %p293, %f2466, %f362; selp.f32 %f2467, %f2451, %f2454, %p293; selp.f32 %f2468, %f2454, %f2452, %p293; add.ftz.f32 %f2469, %f2467, %f2468; mul.ftz.f32 %f3032, %f2469, 0f3F000000; $L__BB148_321: sub.ftz.f32 %f2471, %f1936, %f3030; mul.ftz.f32 %f2472, %f2471, 0f40400000; sub.ftz.f32 %f2473, %f1936, %f3031; mul.ftz.f32 %f2474, %f2473, 0f40400000; sub.ftz.f32 %f2475, %f1936, %f3032; mul.ftz.f32 %f2476, %f2475, 0f40400000; sqrt.approx.ftz.f32 %f2477, %f2472; sqrt.approx.ftz.f32 %f2478, %f2474; sqrt.approx.ftz.f32 %f2479, %f2476; ld.f32 %f2480, [%rd111]; mul.ftz.f32 %f3033, %f2477, %f2480; ld.f32 %f2481, [%rd111+4]; mul.ftz.f32 %f3034, %f2478, %f2481; ld.f32 %f2482, [%rd111+8]; mul.ftz.f32 %f3035, %f2479, %f2482; $L__BB148_323: st.f32 [%rd111], %f3033; st.f32 [%rd111+4], %f3034; st.f32 [%rd111+8], %f3035; or.b32 %r648, %r648, 16; $L__BB148_324: ld.local.u32 %r360, [%rd4+40]; or.b32 %r361, %r360, %r648; st.local.u32 [%rd4+40], %r361; bra.uni $L__BB148_517; $L__BB148_98: mul.wide.u32 %rd279, %r166, 4; add.s64 %rd280, %rd2, %rd279; ld.local.f32 %f2982, [%rd280]; ld.local.f32 %f2983, [%rd280+4]; ld.local.f32 %f2984, [%rd280+8]; $L__BB148_100: ld.global.v4.u32 {%r170, %r171, %r172, %r173}, [%rd3+48]; setp.eq.s32 %p56, %r170, 255; @%p56 bra $L__BB148_102; bra.uni $L__BB148_101; $L__BB148_102: ld.local.f32 %f2992, [%rd4+8]; ld.local.f32 %f2991, [%rd4+4]; ld.local.f32 %f2990, [%rd4]; bra.uni $L__BB148_103; $L__BB148_342: mul.wide.u32 %rd477, %r83, 4; add.s64 %rd478, %rd2, %rd477; ld.local.f32 %f3036, [%rd478]; $L__BB148_344: setp.eq.s32 %p305, %r82, 255; mul.ftz.f32 %f384, %f2, %f380; mul.ftz.f32 %f385, %f2, %f379; mul.ftz.f32 %f386, %f2, %f378; @%p305 bra $L__BB148_346; bra.uni $L__BB148_345; $L__BB148_346: mov.b32 %f3037, %r146; bra.uni $L__BB148_347; $L__BB148_101: mul.wide.u32 %rd281, %r170, 4; add.s64 %rd282, %rd2, %rd281; ld.local.f32 %f2992, [%rd282+8]; ld.local.f32 %f2991, [%rd282+4]; ld.local.f32 %f2990, [%rd282]; $L__BB148_103: @%p10 bra $L__BB148_112; ld.local.v2.f32 {%f677, %f678}, [%rd4+12]; mul.ftz.f32 %f679, %f2991, %f91; fma.rn.ftz.f32 %f680, %f2990, %f90, %f679; fma.rn.ftz.f32 %f681, %f2992, %f92, %f680; add.ftz.f32 %f682, %f681, %f681; mul.ftz.f32 %f683, %f2990, %f682; mul.ftz.f32 %f684, %f2991, %f682; mul.ftz.f32 %f685, %f2992, %f682; sub.ftz.f32 %f686, %f683, %f90; sub.ftz.f32 %f687, %f684, %f91; sub.ftz.f32 %f688, %f685, %f92; mul.ftz.f32 %f689, %f678, %f91; fma.rn.ftz.f32 %f690, %f677, %f90, %f689; ld.local.f32 %f123, [%rd4+20]; fma.rn.ftz.f32 %f124, %f123, %f92, %f690; mul.ftz.f32 %f691, %f124, 0f3F666666; mov.f32 %f692, 0f3C23D70A; min.ftz.f32 %f125, %f691, %f692; mul.ftz.f32 %f693, %f678, %f687; fma.rn.ftz.f32 %f694, %f677, %f686, %f693; fma.rn.ftz.f32 %f695, %f123, %f688, %f694; setp.ltu.ftz.f32 %p58, %f695, %f125; @%p58 bra $L__BB148_105; bra.uni $L__BB148_112; $L__BB148_105: mul.ftz.f32 %f696, %f2991, %f678; fma.rn.ftz.f32 %f697, %f2990, %f677, %f696; fma.rn.ftz.f32 %f698, %f2992, %f123, %f697; mul.ftz.f32 %f699, %f677, %f698; mul.ftz.f32 %f700, %f678, %f698; mul.ftz.f32 %f701, %f123, %f698; sub.ftz.f32 %f702, %f2990, %f699; sub.ftz.f32 %f703, %f2991, %f700; sub.ftz.f32 %f704, %f2992, %f701; mul.ftz.f32 %f705, %f703, %f703; fma.rn.ftz.f32 %f706, %f702, %f702, %f705; fma.rn.ftz.f32 %f707, %f704, %f704, %f706; sqrt.approx.ftz.f32 %f708, %f707; div.approx.ftz.f32 %f129, %f702, %f708; div.approx.ftz.f32 %f130, %f703, %f708; div.approx.ftz.f32 %f131, %f704, %f708; mul.ftz.f32 %f709, %f91, %f130; fma.rn.ftz.f32 %f710, %f90, %f129, %f709; fma.rn.ftz.f32 %f132, %f92, %f131, %f710; mul.ftz.f32 %f711, %f132, %f132; fma.rn.ftz.f32 %f712, %f124, %f124, %f711; mul.ftz.f32 %f713, %f125, %f125; sub.ftz.f32 %f714, %f712, %f713; mul.ftz.f32 %f715, %f711, %f714; mov.f32 %f716, 0f00000000; max.ftz.f32 %f717, %f715, %f716; sqrt.approx.ftz.f32 %f718, %f717; fma.rn.ftz.f32 %f719, %f125, %f124, %f712; mov.f32 %f720, 0f3F000000; div.approx.ftz.f32 %f721, %f720, %f712; add.ftz.f32 %f722, %f718, %f719; mul.ftz.f32 %f133, %f721, %f722; sub.ftz.f32 %f723, %f719, %f718; mul.ftz.f32 %f134, %f721, %f723; setp.gt.ftz.f32 %p59, %f133, 0f3727C5AC; setp.le.ftz.f32 %p60, %f133, 0f3F800054; and.pred %p3, %p59, %p60; setp.gt.ftz.f32 %p61, %f134, 0f3727C5AC; setp.le.ftz.f32 %p62, %f134, 0f3F800054; and.pred %p4, %p61, %p62; and.pred %p63, %p3, %p4; @%p63 bra $L__BB148_108; bra.uni $L__BB148_106; $L__BB148_108: mov.f32 %f730, 0f3F800000; sub.ftz.f32 %f731, %f730, %f133; max.ftz.f32 %f733, %f731, %f716; sqrt.approx.ftz.f32 %f140, %f733; max.ftz.f32 %f734, %f133, %f716; sqrt.approx.ftz.f32 %f141, %f734; sub.ftz.f32 %f735, %f730, %f134; max.ftz.f32 %f736, %f735, %f716; sqrt.approx.ftz.f32 %f142, %f736; max.ftz.f32 %f737, %f134, %f716; sqrt.approx.ftz.f32 %f143, %f737; mul.ftz.f32 %f738, %f124, %f141; fma.rn.ftz.f32 %f739, %f132, %f140, %f738; add.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f141, %f740; sub.ftz.f32 %f144, %f741, %f124; mul.ftz.f32 %f742, %f124, %f143; fma.rn.ftz.f32 %f743, %f132, %f142, %f742; add.ftz.f32 %f744, %f743, %f743; mul.ftz.f32 %f745, %f143, %f744; sub.ftz.f32 %f145, %f745, %f124; setp.ge.ftz.f32 %p65, %f144, 0f3727C5AC; setp.ge.ftz.f32 %p66, %f145, 0f3727C5AC; and.pred %p67, %p65, %p66; @%p67 bra $L__BB148_110; bra.uni $L__BB148_109; $L__BB148_110: setp.lt.ftz.f32 %p69, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p69; selp.f32 %f2989, %f141, %f143, %p69; bra.uni $L__BB148_111; $L__BB148_345: mul.wide.u32 %rd479, %r82, 4; add.s64 %rd480, %rd2, %rd479; ld.local.f32 %f3037, [%rd480]; $L__BB148_347: and.b32 %r86, %r375, 255; bfe.u32 %r87, %r375, 8, 8; bfe.u32 %r88, %r375, 16, 8; bfe.u32 %r89, %r379, 8, 8; bfe.u32 %r90, %r379, 16, 8; shr.u32 %r91, %r379, 24; ld.local.u32 %r92, [%rd4+64]; ld.local.s32 %rd122, [%rd4+48]; setp.eq.s32 %p306, %r92, -1; mov.u32 %r113, -1; mov.u32 %r110, 0; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p306 bra $L__BB148_362; ld.const.u64 %rd481, [kernel_params+2160]; cvta.to.global.u64 %rd482, %rd481; mul.wide.s32 %rd483, %r92, 208; add.s64 %rd484, %rd482, %rd483; ld.global.u32 %r93, [%rd484+164]; and.b32 %r388, %r11, 1; setp.eq.b32 %p307, %r388, 1; mov.pred %p308, 0; xor.pred %p309, %p307, %p308; not.pred %p310, %p309; mov.u32 %r653, 0; @%p310 bra $L__BB148_352; cvt.u32.u64 %r390, %rd122; setp.eq.s32 %p311, %r390, -1; mov.u32 %r652, -1; @%p311 bra $L__BB148_351; ld.const.u64 %rd485, [kernel_params+2240]; cvta.to.global.u64 %rd486, %rd485; shl.b64 %rd487, %rd122, 2; add.s64 %rd488, %rd486, %rd487; ld.global.u32 %r652, [%rd488]; $L__BB148_351: setp.ne.s32 %p312, %r652, -1; selp.u32 %r653, 1, 0, %p312; $L__BB148_352: ld.const.u64 %rd489, [kernel_params+2312]; cvta.to.global.u64 %rd123, %rd489; add.s32 %r655, %r653, %r93; mul.wide.u32 %rd490, %r655, 16; add.s64 %rd491, %rd123, %rd490; add.s64 %rd754, %rd491, 14; ld.global.u16 %rs20, [%rd491+12]; ld.global.u32 %r654, [%rd491+8]; ld.global.u64 %rd752, [%rd491]; setp.eq.s64 %p313, %rd752, %rd121; add.s64 %rd753, %rd491, 15; @%p313 bra $L__BB148_358; $L__BB148_353: setp.eq.s64 %p314, %rd752, 0; @%p314 bra $L__BB148_355; add.s32 %r655, %r655, 2; bra.uni $L__BB148_356; $L__BB148_355: setp.eq.s16 %p315, %rs20, 0; mov.u32 %r110, 0; mov.u32 %r655, %r654; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p315 bra $L__BB148_362; $L__BB148_356: cvt.u64.u32 %rd128, %r655; mul.wide.u32 %rd492, %r655, 16; add.s64 %rd493, %rd123, %rd492; ld.global.u16 %rs20, [%rd493+12]; ld.global.u32 %r654, [%rd493+8]; ld.global.u64 %rd752, [%rd493]; setp.ne.s64 %p316, %rd752, %rd121; @%p316 bra $L__BB148_353; shl.b64 %rd494, %rd128, 4; add.s64 %rd495, %rd123, %rd494; add.s64 %rd754, %rd495, 14; add.s64 %rd753, %rd495, 15; $L__BB148_358: ld.global.u8 %rs5, [%rd754]; ld.global.u8 %rs6, [%rd753]; cvt.u32.u64 %r395, %rd122; setp.ne.s32 %p317, %r395, -1; @%p317 bra $L__BB148_361; mov.u32 %r110, 0; add.s16 %rs14, %rs20, -1; setp.lt.u16 %p318, %rs14, 2; @%p318 bra $L__BB148_361; setp.ne.s16 %p319, %rs20, 1024; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p319 bra $L__BB148_362; $L__BB148_361: setp.eq.s16 %p320, %rs20, 0; selp.b32 %r113, -1, %r654, %p320; cvt.u32.u16 %r110, %rs20; cvt.u32.u16 %r400, %rs6; and.b32 %r660, %r400, 255; cvt.u32.u16 %r401, %rs5; and.b32 %r659, %r401, 255; $L__BB148_362: setp.eq.s32 %p321, %r113, -1; @%p321 bra $L__BB148_382; and.b32 %r402, %r11, 1; setp.eq.b32 %p322, %r402, 1; mov.pred %p323, 0; xor.pred %p324, %p322, %p323; not.pred %p325, %p324; @%p325 bra $L__BB148_375; bra.uni $L__BB148_364; $L__BB148_375: setp.eq.s32 %p332, %r12, 0; @%p332 bra $L__BB148_380; and.b32 %r423, %r110, 768; setp.eq.s32 %p333, %r423, 0; @%p333 bra $L__BB148_378; ld.const.u64 %rd517, [kernel_params+2264]; cvta.to.global.u64 %rd518, %rd517; shl.b64 %rd519, %rd122, 4; add.s64 %rd520, %rd518, %rd519; shr.s32 %r424, %r11, 7; ld.global.u32 %r425, [%rd520+4]; add.s32 %r426, %r424, %r425; ld.const.u64 %rd521, [kernel_params+2320]; cvta.to.global.u64 %rd522, %rd521; add.s32 %r427, %r426, %r113; mul.wide.s32 %rd523, %r427, 4; add.s64 %rd524, %rd522, %rd523; ld.local.f32 %f2520, [%rd4+56]; mov.f32 %f2521, 0f3F800000; sub.ftz.f32 %f2522, %f2521, %f2520; ld.global.f32 %f2523, [%rd524]; mul.ftz.f32 %f2524, %f2523, %f2522; ld.global.f32 %f2525, [%rd524+4]; fma.rn.ftz.f32 %f3038, %f2525, %f2520, %f2524; bra.uni $L__BB148_385; $L__BB148_472: mul.ftz.f32 %f2768, %f60, %f60; st.local.v2.f32 [%rd185+12], {%f2768, %f2768}; st.local.f32 [%rd185+20], %f3057; cvt.ftz.sat.f32.f32 %f507, %f2768; setp.eq.s32 %p411, %r2, 22; @%p411 bra $L__BB148_474; bra.uni $L__BB148_473; $L__BB148_474: st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r534, 22; st.local.u32 [%rd185+-8], %r534; ld.local.u32 %r535, [%rd4+40]; or.b32 %r536, %r535, 12; st.local.u32 [%rd4+40], %r536; bra.uni $L__BB148_517; $L__BB148_60: mov.u32 %r617, 19; st.local.u32 [%rd222+-8], %r617; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r618, [%rd4+40]; or.b32 %r619, %r618, 12; st.local.u32 [%rd4+40], %r619; bra.uni $L__BB148_517; $L__BB148_47: mov.f32 %f2876, 0f3F800000; sub.ftz.f32 %f2877, %f2876, %f528; mul.ftz.f32 %f3059, %f3058, %f2877; div.approx.ftz.f32 %f3058, %f3058, %f2877; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; $L__BB148_491: setp.eq.s32 %p436, %r2, 9; @%p436 bra $L__BB148_505; bra.uni $L__BB148_492; $L__BB148_505: mov.u32 %r597, 9; st.local.u32 [%rd205+-8], %r597; ld.local.u32 %r598, [%rd4+40]; or.b32 %r599, %r598, 4; st.local.u32 [%rd4+40], %r599; bra.uni $L__BB148_517; $L__BB148_492: setp.eq.s32 %p437, %r2, 13; @%p437 bra $L__BB148_504; bra.uni $L__BB148_493; $L__BB148_504: cvt.ftz.sat.f32.f32 %f2920, %f3058; cvt.ftz.sat.f32.f32 %f2921, %f3059; st.local.v2.f32 [%rd205+12], {%f2921, %f2920}; mov.u32 %r594, 13; st.local.u32 [%rd205+-8], %r594; ld.local.u32 %r595, [%rd4+40]; or.b32 %r596, %r595, 12; st.local.u32 [%rd4+40], %r596; bra.uni $L__BB148_517; $L__BB148_493: setp.eq.s32 %p438, %r2, 10; @%p438 bra $L__BB148_503; bra.uni $L__BB148_494; $L__BB148_503: mov.u64 %rd663, 0; st.local.u64 [%rd205+28], %rd663; cvt.ftz.sat.f32.f32 %f2918, %f3058; cvt.ftz.sat.f32.f32 %f2919, %f3059; st.local.v2.f32 [%rd205+12], {%f2919, %f2918}; mov.u32 %r591, 10; st.local.u32 [%rd205+-8], %r591; ld.local.u32 %r592, [%rd4+40]; or.b32 %r593, %r592, 12; st.local.u32 [%rd4+40], %r593; bra.uni $L__BB148_517; $L__BB148_441: setp.eq.s32 %p375, %r2, 25; st.local.v2.f32 [%rd160+12], {%f480, %f480}; st.local.f32 [%rd160+20], %f3054; cvt.ftz.sat.f32.f32 %f485, %f480; @%p375 bra $L__BB148_443; bra.uni $L__BB148_442; $L__BB148_443: st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r472, 13; st.local.u32 [%rd160+-8], %r472; ld.local.u32 %r473, [%rd4+40]; or.b32 %r474, %r473, 12; st.local.u32 [%rd4+40], %r474; bra.uni $L__BB148_445; $L__BB148_382: setp.eq.s32 %p339, %r89, 255; @%p339 bra $L__BB148_384; bra.uni $L__BB148_383; $L__BB148_384: mov.b32 %f3038, %r380; bra.uni $L__BB148_385; $L__BB148_364: cvt.u32.u64 %r404, %rd122; setp.eq.s32 %p326, %r404, -1; mov.u32 %r662, -1; @%p326 bra $L__BB148_366; ld.const.u64 %rd496, [kernel_params+2240]; cvta.to.global.u64 %rd497, %rd496; shl.b64 %rd498, %rd122, 2; add.s64 %rd499, %rd497, %rd498; ld.global.u32 %r662, [%rd499]; $L__BB148_366: setp.eq.s32 %p327, %r662, -1; @%p327 bra $L__BB148_368; mov.u64 %rd500, 0; { // callseq 902, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd230; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r110; st.param.b32 [param1+4], %r659; st.param.b32 [param1+8], %r660; st.param.b32 [param1+12], %r113; .param .b64 param2; st.param.b64 [param2+0], %rd500; .param .b64 param3; st.param.b64 [param3+0], %rd500; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f3038, [retval0+0]; } // callseq 902 bra.uni $L__BB148_385; $L__BB148_327: ld.local.f32 %f2492, [%rd4+280]; add.ftz.f32 %f2493, %f374, %f2492; ld.local.f32 %f2494, [%rd4+284]; add.ftz.f32 %f2495, %f375, %f2494; ld.local.f32 %f2496, [%rd4+288]; add.ftz.f32 %f2497, %f376, %f2496; st.local.f32 [%rd4+280], %f2493; st.local.f32 [%rd4+284], %f2495; st.local.f32 [%rd4+288], %f2497; ld.local.u32 %r68, [%rd4+248]; setp.lt.s32 %p296, %r68, 1; @%p296 bra $L__BB148_517; mov.u32 %r650, 0; $L__BB148_329: mul.wide.s32 %rd472, %r650, 80; add.s64 %rd473, %rd1, %rd472; add.s64 %rd118, %rd473, 316; ld.local.u32 %r364, [%rd473+316]; setp.eq.s32 %p297, %r364, 31; add.s32 %r650, %r650, 1; @%p297 bra $L__BB148_331; setp.lt.s32 %p298, %r650, %r68; @%p298 bra $L__BB148_329; bra.uni $L__BB148_517; $L__BB148_478: ld.local.f32 %f2802, [%rd4+116]; ld.local.f32 %f2803, [%rd4+120]; mul.ftz.f32 %f2804, %f2803, %f2803; fma.rn.ftz.f32 %f2805, %f2802, %f2802, %f2804; ld.local.f32 %f2806, [%rd4+124]; fma.rn.ftz.f32 %f2807, %f2806, %f2806, %f2805; sqrt.approx.ftz.f32 %f2808, %f2807; div.approx.ftz.f32 %f2809, %f2802, %f2808; div.approx.ftz.f32 %f2810, %f2803, %f2808; div.approx.ftz.f32 %f2811, %f2806, %f2808; st.local.f32 [%rd192+12], %f2809; st.local.f32 [%rd192+16], %f2810; st.local.f32 [%rd192+20], %f2811; mov.u32 %r547, 0; st.local.u32 [%rd192+32], %r547; $L__BB148_479: setp.eq.s32 %p418, %r2, 20; mov.f32 %f2812, 0f3A83126F; max.ftz.f32 %f2813, %f60, %f2812; mov.f32 %f2814, 0f3F800000; min.ftz.f32 %f513, %f2813, %f2814; max.ftz.f32 %f2815, %f63, %f2812; min.ftz.f32 %f514, %f2815, %f2814; @%p418 bra $L__BB148_481; bra.uni $L__BB148_480; $L__BB148_481: mov.u32 %r551, 20; st.local.u32 [%rd192+-8], %r551; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r552, [%rd4+40]; or.b32 %r553, %r552, 12; st.local.u32 [%rd4+40], %r553; bra.uni $L__BB148_517; $L__BB148_480: mov.u32 %r548, 30; st.local.u32 [%rd192+-8], %r548; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r549, [%rd4+40]; or.b32 %r550, %r549, 12; st.local.u32 [%rd4+40], %r550; bra.uni $L__BB148_517; $L__BB148_514: st.local.f32 [%rd229+12], %f60; mov.u32 %r630, 3; st.local.u32 [%rd229+-8], %r630; cvt.ftz.sat.f32.f32 %f2952, %f60; fma.rn.ftz.f32 %f2953, %f2952, 0f3F67750B, 0f40490FDB; rcp.approx.ftz.f32 %f2954, %f2953; st.local.f32 [%rd229+16], %f2954; mul.ftz.f32 %f2955, %f2952, %f2954; st.local.f32 [%rd229+20], %f2955; ld.local.u32 %r631, [%rd4+40]; or.b32 %r632, %r631, 12; st.local.u32 [%rd4+40], %r632; bra.uni $L__BB148_517; $L__BB148_473: st.local.u64 [%rd185+28], %rd621; st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r531, 23; st.local.u32 [%rd185+-8], %r531; ld.local.u32 %r532, [%rd4+40]; or.b32 %r533, %r532, 12; st.local.u32 [%rd4+40], %r533; bra.uni $L__BB148_517; $L__BB148_494: setp.eq.s32 %p439, %r2, 14; @%p439 bra $L__BB148_496; bra.uni $L__BB148_495; $L__BB148_496: ld.local.u32 %r131, [%rd4+252]; setp.lt.s32 %p440, %r131, 1; @%p440 bra $L__BB148_498; add.s32 %r582, %r131, -1; st.local.u32 [%rd4+252], %r582; ld.local.s32 %rd651, [%rd4+248]; cvt.s64.s32 %rd652, %r582; add.s64 %rd653, %rd651, %rd652; mul.lo.s64 %rd654, %rd653, 80; add.s64 %rd655, %rd1, %rd654; add.s64 %rd794, %rd655, 304; add.s64 %rd656, %rd230, %rd654; add.s64 %rd795, %rd656, 304; bra.uni $L__BB148_499; $L__BB148_457: ld.local.u32 %r508, [%rd4+248]; mov.u64 %rd776, 0; add.s32 %r509, %r508, -1; st.local.u32 [%rd4+248], %r509; add.s32 %r510, %r126, 1; st.local.u32 [%rd4+252], %r510; mov.u64 %rd777, %rd776; $L__BB148_458: setp.eq.s64 %p394, %rd777, 0; @%p394 bra $L__BB148_517; add.s64 %rd178, %rd774, 20; st.local.f32 [%rd774+20], %f55; st.local.f32 [%rd774+24], %f56; st.local.f32 [%rd774+28], %f57; st.local.u64 [%rd774+48], %rd777; mov.u32 %r511, 0; st.local.v2.f32 [%rd774+56], {%f2732, %f2732}; st.local.u32 [%rd774+64], %r511; mul.ftz.f32 %f2737, %f60, %f60; st.local.v2.f32 [%rd774+32], {%f2737, %f2737}; mov.f32 %f2738, 0f3727C5AC; max.ftz.f32 %f3056, %f63, %f2738; ld.local.u32 %r512, [%rd4+40]; and.b32 %r513, %r512, 1; setp.eq.b32 %p395, %r513, 1; mov.pred %p396, 0; xor.pred %p397, %p395, %p396; not.pred %p398, %p397; @%p398 bra $L__BB148_461; rcp.approx.ftz.f32 %f3056, %f3056; $L__BB148_461: st.local.f32 [%rd178+20], %f3056; ld.local.f32 %f2739, [%rd6]; ld.local.f32 %f2740, [%rd6+4]; ld.local.f32 %f2741, [%rd6+8]; st.local.f32 [%rd776], %f2739; st.local.f32 [%rd776+4], %f2740; st.local.f32 [%rd776+8], %f2741; ld.local.u64 %rd609, [%rd178+28]; st.u32 [%rd609+12], %r511; st.u32 [%rd609+16], %r511; st.u32 [%rd609+20], %r511; ld.local.u64 %rd610, [%rd178+28]; st.u32 [%rd610+36], %r511; ld.local.f32 %f2743, [%rd178+12]; mov.f32 %f2744, 0f38D1B717; max.ftz.f32 %f2745, %f2743, %f2744; mov.f32 %f2746, 0f3F800000; min.ftz.f32 %f2747, %f2745, %f2746; st.local.v2.f32 [%rd178+12], {%f2747, %f2747}; ld.local.f32 %f2748, [%rd178+20]; max.ftz.f32 %f2749, %f2732, %f2748; st.local.f32 [%rd178+20], %f2749; ld.local.u64 %rd611, [%rd178+28]; ld.f32 %f2750, [%rd611]; cvt.ftz.sat.f32.f32 %f2751, %f2750; ld.f32 %f2752, [%rd611+4]; cvt.ftz.sat.f32.f32 %f2753, %f2752; ld.f32 %f2754, [%rd611+8]; cvt.ftz.sat.f32.f32 %f2755, %f2754; st.f32 [%rd611], %f2751; st.f32 [%rd611+4], %f2753; st.f32 [%rd611+8], %f2755; mov.u32 %r515, 24; st.local.u32 [%rd178+-8], %r515; ld.local.u32 %r516, [%rd4+40]; or.b32 %r517, %r516, 1036; st.local.u32 [%rd4+40], %r517; bra.uni $L__BB148_517; $L__BB148_442: st.local.u64 [%rd160+28], %rd579; st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r469, 10; st.local.u32 [%rd160+-8], %r469; ld.local.u32 %r470, [%rd4+40]; or.b32 %r471, %r470, 12; st.local.u32 [%rd4+40], %r471; $L__BB148_445: setp.ne.s32 %p376, %r122, 0; ld.const.u32 %r478, [kernel_params+1568]; setp.eq.s32 %p377, %r478, 0; and.pred %p378, %p376, %p377; @%p378 bra $L__BB148_517; mov.f32 %f2713, 0f3F800000; sub.ftz.f32 %f2714, %f2713, %f3055; mov.f32 %f2715, 0f38D1B717; max.ftz.f32 %f2716, %f2715, %f2714; mul.ftz.f32 %f2717, %f470, %f2716; mul.ftz.f32 %f2718, %f471, %f2716; mul.ftz.f32 %f2719, %f472, %f2716; mov.f32 %f2720, 0f00000000; max.ftz.f32 %f487, %f2717, %f2720; max.ftz.f32 %f488, %f2718, %f2720; max.ftz.f32 %f489, %f2719, %f2720; add.ftz.f32 %f2721, %f487, %f488; add.ftz.f32 %f2722, %f489, %f2721; mul.ftz.f32 %f2723, %f2722, 0f3EAAAAAB; abs.ftz.f32 %f490, %f2723; setp.ltu.ftz.f32 %p379, %f490, 0f3727C5AC; mov.u64 %rd770, 0; mov.u64 %rd771, %rd770; @%p379 bra $L__BB148_451; ld.local.u32 %r124, [%rd4+252]; mov.u64 %rd770, 0; setp.eq.s32 %p380, %r124, 0; mov.u64 %rd768, %rd770; mov.u64 %rd769, %rd770; @%p380 bra $L__BB148_449; ld.local.u32 %r479, [%rd4+248]; mul.wide.s32 %rd585, %r479, 80; add.s64 %rd586, %rd1, %rd585; add.s64 %rd768, %rd586, 304; add.s64 %rd587, %rd230, %rd585; add.s64 %rd769, %rd587, 304; mov.b32 %r480, %f489; mov.b32 %r481, %f488; mov.b32 %r482, %f487; mov.u32 %r483, 0; st.local.v4.u32 [%rd586+304], {%r482, %r481, %r480, %r483}; add.s32 %r484, %r479, 1; st.local.u32 [%rd4+248], %r484; add.s32 %r485, %r124, -1; st.local.u32 [%rd4+252], %r485; $L__BB148_449: setp.eq.s64 %p381, %rd769, 0; mov.u64 %rd771, %rd770; @%p381 bra $L__BB148_451; st.local.f32 [%rd768+16], %f490; mov.u64 %rd770, %rd768; mov.u64 %rd771, %rd769; $L__BB148_451: setp.eq.s64 %p382, %rd771, 0; @%p382 bra $L__BB148_517; add.s64 %rd167, %rd770, 20; mov.u64 %rd590, 0; st.local.f32 [%rd770+20], %f55; st.local.f32 [%rd770+24], %f56; st.local.f32 [%rd770+28], %f57; mov.u32 %r486, 0; st.local.v2.f32 [%rd770+56], {%f2720, %f2720}; st.local.u32 [%rd770+64], %r486; st.local.u64 [%rd770+48], %rd590; setp.eq.s32 %p383, %r2, 28; @%p383 bra $L__BB148_456; bra.uni $L__BB148_453; $L__BB148_456: st.local.v2.f32 [%rd167+12], {%f2720, %f2720}; st.local.f32 [%rd167+20], %f3054; mov.u32 %r493, 21; st.local.u32 [%rd167+-8], %r493; ld.local.u32 %r494, [%rd4+40]; or.b32 %r495, %r494, 4; st.local.u32 [%rd4+40], %r495; bra.uni $L__BB148_517; $L__BB148_453: setp.eq.s32 %p384, %r2, 25; st.local.v2.f32 [%rd167+12], {%f480, %f480}; st.local.f32 [%rd167+20], %f3054; cvt.ftz.sat.f32.f32 %f491, %f480; @%p384 bra $L__BB148_455; bra.uni $L__BB148_454; $L__BB148_455: st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r490, 22; st.local.u32 [%rd167+-8], %r490; ld.local.u32 %r491, [%rd4+40]; or.b32 %r492, %r491, 12; st.local.u32 [%rd4+40], %r492; bra.uni $L__BB148_517; $L__BB148_454: st.local.u64 [%rd167+28], %rd590; st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r487, 23; st.local.u32 [%rd167+-8], %r487; ld.local.u32 %r488, [%rd4+40]; or.b32 %r489, %r488, 12; st.local.u32 [%rd4+40], %r489; bra.uni $L__BB148_517; $L__BB148_383: mul.wide.u32 %rd533, %r89, 4; add.s64 %rd534, %rd2, %rd533; ld.local.f32 %f3038, [%rd534]; bra.uni $L__BB148_385; $L__BB148_380: and.b32 %r432, %r11, 8; setp.eq.s32 %p336, %r432, 0; setp.ne.s32 %p337, %r110, 8; mov.f32 %f3038, 0f00000000; or.pred %p338, %p337, %p336; @%p338 bra $L__BB148_385; cvt.u32.u64 %r433, %rd122; ld.const.u64 %rd529, [kernel_params+2320]; cvta.to.global.u64 %rd530, %rd529; add.s32 %r434, %r433, %r113; mul.wide.s32 %rd531, %r434, 4; add.s64 %rd532, %rd530, %rd531; ld.global.f32 %f3038, [%rd532]; bra.uni $L__BB148_385; $L__BB148_368: and.b32 %r405, %r110, 56; setp.eq.s32 %p328, %r405, 0; @%p328 bra $L__BB148_373; and.b32 %r406, %r110, 24; setp.eq.s32 %p329, %r406, 0; @%p329 bra $L__BB148_371; ld.const.u64 %rd501, [kernel_params+2232]; cvta.to.global.u64 %rd502, %rd501; shl.b64 %rd503, %rd122, 4; add.s64 %rd504, %rd502, %rd503; ld.global.v4.u32 {%r407, %r408, %r409, %r410}, [%rd504]; ld.const.u64 %rd505, [kernel_params+2320]; cvta.to.global.u64 %rd506, %rd505; add.s32 %r414, %r407, %r113; mul.wide.u32 %rd507, %r414, 4; add.s64 %rd755, %rd506, %rd507; add.s32 %r415, %r408, %r113; mul.wide.u32 %rd508, %r415, 4; add.s64 %rd756, %rd506, %rd508; add.s32 %r416, %r409, %r113; mul.wide.u32 %rd509, %r416, 4; add.s64 %rd757, %rd506, %rd509; bra.uni $L__BB148_372; $L__BB148_495: mov.f32 %f2880, 0f38D1B717; max.ftz.f32 %f2881, %f3059, %f2880; mov.f32 %f2882, 0f3F800000; max.ftz.f32 %f2883, %f3058, %f2880; min.ftz.f32 %f2884, %f2883, %f2882; min.ftz.f32 %f2885, %f2881, %f2882; st.local.v2.f32 [%rd205+12], {%f2885, %f2884}; mov.u32 %r579, 16; st.local.u32 [%rd205+-8], %r579; ld.local.u32 %r580, [%rd4+40]; or.b32 %r581, %r580, 12; st.local.u32 [%rd4+40], %r581; bra.uni $L__BB148_517; $L__BB148_334: ld.local.u32 %r651, [%rd4+252]; $L__BB148_335: setp.eq.s32 %p301, %r651, 0; mov.pred %p462, -1; mov.u64 %rd751, 0; @%p301 bra $L__BB148_337; ld.local.u32 %r367, [%rd4+248]; mul.wide.s32 %rd475, %r367, 80; add.s64 %rd476, %rd1, %rd475; add.s64 %rd751, %rd476, 304; mov.b32 %r368, %f376; mov.b32 %r369, %f375; mov.b32 %r370, %f374; mov.u32 %r371, 31; st.local.v4.u32 [%rd476+304], {%r370, %r369, %r368, %r371}; add.s32 %r372, %r367, 1; st.local.u32 [%rd4+248], %r372; add.s32 %r373, %r651, -1; st.local.u32 [%rd4+252], %r373; mov.pred %p462, 0; $L__BB148_337: @%p462 bra $L__BB148_339; bra.uni $L__BB148_338; $L__BB148_339: @%p299 bra $L__BB148_517; mov.u32 %r374, 0; st.local.u32 [%rd4+252], %r374; bra.uni $L__BB148_517; $L__BB148_338: st.local.f32 [%rd751+16], %f377; ld.local.f32 %f2506, [%rd4]; ld.local.f32 %f2507, [%rd4+4]; ld.local.f32 %f2508, [%rd4+8]; st.local.f32 [%rd751+20], %f2506; st.local.f32 [%rd751+24], %f2507; st.local.f32 [%rd751+28], %f2508; bra.uni $L__BB148_517; $L__BB148_106: or.pred %p64, %p3, %p4; mov.f32 %f2990, %f677; mov.f32 %f2991, %f678; mov.f32 %f2992, %f123; @%p64 bra $L__BB148_107; bra.uni $L__BB148_112; $L__BB148_107: selp.f32 %f724, %f133, %f134, %p3; mov.f32 %f725, 0f3F800000; sub.ftz.f32 %f726, %f725, %f724; max.ftz.f32 %f728, %f726, %f716; sqrt.approx.ftz.f32 %f2988, %f728; max.ftz.f32 %f729, %f724, %f716; sqrt.approx.ftz.f32 %f2989, %f729; bra.uni $L__BB148_111; $L__BB148_331: ld.local.f32 %f2498, [%rd118+-12]; add.ftz.f32 %f2499, %f374, %f2498; ld.local.f32 %f2500, [%rd118+-8]; add.ftz.f32 %f2501, %f375, %f2500; ld.local.f32 %f2502, [%rd118+-4]; add.ftz.f32 %f2503, %f376, %f2502; st.local.f32 [%rd118+-12], %f2499; st.local.f32 [%rd118+-8], %f2501; st.local.f32 [%rd118+-4], %f2503; ld.local.f32 %f2504, [%rd118+4]; add.ftz.f32 %f2505, %f377, %f2504; st.local.f32 [%rd118+4], %f2505; bra.uni $L__BB148_517; $L__BB148_378: and.b32 %r428, %r110, 131; setp.eq.s32 %p334, %r428, 0; mov.f32 %f3038, 0f00000000; @%p334 bra $L__BB148_385; cvt.u32.u64 %r429, %rd122; setp.eq.s32 %p335, %r110, 128; selp.b32 %r430, %r429, 0, %p335; add.s32 %r431, %r113, %r430; ld.const.u64 %rd525, [kernel_params+2320]; cvta.to.global.u64 %rd526, %rd525; mul.wide.s32 %rd527, %r431, 4; add.s64 %rd528, %rd526, %rd527; ld.global.f32 %f3038, [%rd528]; bra.uni $L__BB148_385; $L__BB148_109: setp.gt.ftz.f32 %p68, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p68; selp.f32 %f2989, %f141, %f143, %p68; $L__BB148_111: mul.ftz.f32 %f746, %f677, %f2989; mul.ftz.f32 %f747, %f678, %f2989; mul.ftz.f32 %f748, %f123, %f2989; fma.rn.ftz.f32 %f2992, %f131, %f2988, %f748; fma.rn.ftz.f32 %f2991, %f130, %f2988, %f747; fma.rn.ftz.f32 %f2990, %f129, %f2988, %f746; $L__BB148_112: setp.eq.s32 %p70, %r171, 255; mov.f32 %f2993, 0f3F800000; mov.f32 %f2994, %f2993; mov.f32 %f2995, %f2993; @%p70 bra $L__BB148_114; mul.wide.u32 %rd283, %r171, 4; add.s64 %rd284, %rd2, %rd283; ld.local.f32 %f2993, [%rd284]; ld.local.f32 %f2994, [%rd284+4]; ld.local.f32 %f2995, [%rd284+8]; $L__BB148_114: setp.eq.s32 %p71, %r172, 255; mov.f32 %f2996, 0f3FB33333; @%p71 bra $L__BB148_116; mul.wide.u32 %rd285, %r172, 4; add.s64 %rd286, %rd2, %rd285; ld.local.f32 %f2996, [%rd286]; $L__BB148_116: setp.eq.s32 %p72, %r173, 255; mov.f32 %f2997, 0f00000000; @%p72 bra $L__BB148_118; mul.wide.u32 %rd287, %r173, 4; add.s64 %rd288, %rd2, %rd287; ld.local.f32 %f2997, [%rd288]; $L__BB148_118: ld.global.v4.u32 {%r174, %r175, %r176, %r177}, [%rd3+64]; setp.eq.s32 %p73, %r174, 255; @%p73 bra $L__BB148_120; bra.uni $L__BB148_119; $L__BB148_120: mov.b32 %f2998, %r175; mov.b32 %f2999, %r176; mov.b32 %f3000, %r177; bra.uni $L__BB148_121; $L__BB148_119: mul.wide.u32 %rd289, %r174, 4; add.s64 %rd290, %rd2, %rd289; ld.local.f32 %f2998, [%rd290]; ld.local.f32 %f2999, [%rd290+4]; ld.local.f32 %f3000, [%rd290+8]; $L__BB148_121: ld.local.f32 %f754, [%rd4+256]; mul.ftz.f32 %f177, %f2, %f754; ld.local.f32 %f755, [%rd4+260]; mul.ftz.f32 %f178, %f2, %f755; ld.local.f32 %f756, [%rd4+264]; mul.ftz.f32 %f179, %f2, %f756; mov.f32 %f757, 0f3F800000; sub.ftz.f32 %f758, %f757, %f63; mul.ftz.f32 %f759, %f63, %f2998; fma.rn.ftz.f32 %f180, %f758, %f2982, %f759; mul.ftz.f32 %f760, %f63, %f2999; fma.rn.ftz.f32 %f181, %f758, %f2983, %f760; mul.ftz.f32 %f761, %f63, %f3000; fma.rn.ftz.f32 %f182, %f758, %f2984, %f761; mul.ftz.f32 %f762, %f180, %f177; mul.ftz.f32 %f763, %f181, %f178; mul.ftz.f32 %f764, %f182, %f179; mul.ftz.f32 %f183, %f97, %f762; mul.ftz.f32 %f184, %f97, %f763; mul.ftz.f32 %f185, %f97, %f764; and.b32 %r178, %r141, 4096; setp.eq.s32 %p74, %r178, 0; selp.f32 %f186, %f2982, %f180, %p74; selp.f32 %f187, %f2983, %f181, %p74; selp.f32 %f188, %f2984, %f182, %p74; selp.f32 %f189, %f63, 0f00000000, %p74; add.ftz.f32 %f765, %f180, %f181; add.ftz.f32 %f766, %f182, %f765; mul.ftz.f32 %f767, %f766, 0f3EAAAAAB; abs.ftz.f32 %f768, %f767; setp.leu.ftz.f32 %p75, %f768, 0f3727C5AC; @%p75 bra $L__BB148_182; setp.le.ftz.f32 %p76, %f189, 0f3727C5AC; setp.gt.ftz.f32 %p77, %f97, 0f3727C5AC; and.pred %p78, %p76, %p77; @%p78 bra $L__BB148_175; bra.uni $L__BB148_123; $L__BB148_175: mul.ftz.f32 %f1391, %f186, %f177; mul.ftz.f32 %f1392, %f97, %f1391; mul.ftz.f32 %f1393, %f187, %f178; mul.ftz.f32 %f1394, %f97, %f1393; mul.ftz.f32 %f1395, %f188, %f179; mul.ftz.f32 %f1396, %f97, %f1395; mov.f32 %f1397, 0f00000000; max.ftz.f32 %f223, %f1392, %f1397; max.ftz.f32 %f224, %f1394, %f1397; max.ftz.f32 %f225, %f1396, %f1397; add.ftz.f32 %f1398, %f223, %f224; add.ftz.f32 %f1399, %f225, %f1398; mul.ftz.f32 %f1400, %f1399, 0f3EAAAAAB; abs.ftz.f32 %f226, %f1400; setp.ltu.ftz.f32 %p143, %f226, 0f3727C5AC; mov.u64 %rd708, 0; mov.u64 %rd709, %rd708; @%p143 bra $L__BB148_180; ld.local.u32 %r42, [%rd4+252]; mov.u64 %rd708, 0; setp.eq.s32 %p144, %r42, 0; mov.u64 %rd706, %rd708; mov.u64 %rd707, %rd708; @%p144 bra $L__BB148_178; ld.local.u32 %r226, [%rd4+248]; mul.wide.s32 %rd328, %r226, 80; add.s64 %rd329, %rd1, %rd328; add.s64 %rd706, %rd329, 304; add.s64 %rd330, %rd230, %rd328; add.s64 %rd707, %rd330, 304; mov.b32 %r227, %f225; mov.b32 %r228, %f224; mov.b32 %r229, %f223; mov.u32 %r230, 0; st.local.v4.u32 [%rd329+304], {%r229, %r228, %r227, %r230}; add.s32 %r231, %r226, 1; st.local.u32 [%rd4+248], %r231; add.s32 %r232, %r42, -1; st.local.u32 [%rd4+252], %r232; $L__BB148_178: setp.eq.s64 %p145, %rd707, 0; mov.u64 %rd709, %rd708; @%p145 bra $L__BB148_180; st.local.f32 [%rd706+16], %f226; mov.u64 %rd708, %rd706; mov.u64 %rd709, %rd707; $L__BB148_180: setp.eq.s64 %p146, %rd709, 0; @%p146 bra $L__BB148_182; st.local.f32 [%rd708+20], %f55; st.local.f32 [%rd708+24], %f56; st.local.f32 [%rd708+28], %f57; st.local.f32 [%rd708+32], %f71; mov.u32 %r233, 5; st.local.u32 [%rd708+12], %r233; mov.u32 %r234, 1; st.local.u32 [%rd708+36], %r234; ld.local.u32 %r235, [%rd4+40]; or.b32 %r236, %r235, 12; st.local.u32 [%rd4+40], %r236; bra.uni $L__BB148_182; $L__BB148_123: setp.leu.ftz.f32 %p79, %f189, 0f3727C5AC; @%p79 bra $L__BB148_182; ld.local.u32 %r29, [%rd4+252]; mov.u64 %rd693, 0; setp.eq.s32 %p80, %r29, 0; mov.u64 %rd691, %rd693; mov.u64 %rd692, %rd693; @%p80 bra $L__BB148_126; ld.local.u32 %r179, [%rd4+248]; mul.wide.s32 %rd293, %r179, 80; add.s64 %rd294, %rd1, %rd293; add.s64 %rd691, %rd294, 304; add.s64 %rd295, %rd230, %rd293; add.s64 %rd692, %rd295, 304; mov.b32 %r180, %f185; mov.b32 %r181, %f184; mov.b32 %r182, %f183; mov.u32 %r183, 0; st.local.v4.u32 [%rd294+304], {%r182, %r181, %r180, %r183}; add.s32 %r184, %r179, 1; st.local.u32 [%rd4+248], %r184; add.s32 %r185, %r29, -1; st.local.u32 [%rd4+252], %r185; $L__BB148_126: setp.eq.s64 %p81, %rd692, 0; @%p81 bra $L__BB148_128; add.ftz.f32 %f769, %f183, %f184; add.ftz.f32 %f770, %f185, %f769; mul.ftz.f32 %f771, %f770, 0f3EAAAAAB; abs.ftz.f32 %f772, %f771; st.local.f32 [%rd691+16], %f772; setp.ge.ftz.f32 %p82, %f772, 0f3727C5AC; selp.b64 %rd693, %rd692, 0, %p82; $L__BB148_128: setp.eq.s64 %p83, %rd693, 0; @%p83 bra $L__BB148_182; mul.ftz.f32 %f773, %f189, %f2993; add.s64 %rd14, %rd693, 32; st.f32 [%rd693+32], %f773; mul.ftz.f32 %f774, %f189, %f2994; st.f32 [%rd693+36], %f774; mul.ftz.f32 %f775, %f189, %f2995; st.f32 [%rd693+40], %f775; st.f32 [%rd693+44], %f180; st.f32 [%rd693+48], %f181; st.f32 [%rd693+52], %f182; st.f32 [%rd693+20], %f55; st.f32 [%rd693+24], %f56; st.f32 [%rd693+28], %f57; st.f32 [%rd693+56], %f71; mov.f32 %f3003, 0f00000000; max.ftz.f32 %f777, %f2997, %f3003; mov.f32 %f778, 0f3F666666; min.ftz.f32 %f779, %f777, %f778; st.f32 [%rd693+60], %f779; setp.eq.ftz.f32 %p84, %f71, 0f7F7FFFFF; mov.u32 %r643, 0; @%p84 bra $L__BB148_137; ld.f32 %f780, [%rd14+-32]; mov.u64 %rd696, 0; mov.f32 %f781, 0f00000000; max.ftz.f32 %f190, %f780, %f781; ld.f32 %f782, [%rd14+-28]; max.ftz.f32 %f191, %f782, %f781; ld.f32 %f783, [%rd14+-24]; max.ftz.f32 %f192, %f783, %f781; add.ftz.f32 %f784, %f190, %f191; add.ftz.f32 %f785, %f192, %f784; mul.ftz.f32 %f786, %f785, 0f3EAAAAAB; abs.ftz.f32 %f193, %f786; setp.ltu.ftz.f32 %p85, %f193, 0f3727C5AC; mov.u64 %rd697, %rd696; @%p85 bra $L__BB148_135; ld.local.u32 %r30, [%rd4+252]; mov.u64 %rd696, 0; setp.eq.s32 %p86, %r30, 0; mov.u64 %rd694, %rd696; mov.u64 %rd695, %rd696; @%p86 bra $L__BB148_133; ld.local.u32 %r187, [%rd4+248]; mul.wide.s32 %rd301, %r187, 80; add.s64 %rd302, %rd1, %rd301; add.s64 %rd694, %rd302, 304; add.s64 %rd303, %rd230, %rd301; add.s64 %rd695, %rd303, 304; mov.b32 %r188, %f192; mov.b32 %r189, %f191; mov.b32 %r190, %f190; mov.u32 %r191, 0; st.local.v4.u32 [%rd302+304], {%r190, %r189, %r188, %r191}; add.s32 %r192, %r187, 1; st.local.u32 [%rd4+248], %r192; add.s32 %r193, %r30, -1; st.local.u32 [%rd4+252], %r193; $L__BB148_133: setp.eq.s64 %p87, %rd695, 0; mov.u64 %rd697, %rd696; @%p87 bra $L__BB148_135; st.local.f32 [%rd694+16], %f193; mov.u64 %rd696, %rd694; mov.u64 %rd697, %rd695; $L__BB148_135: setp.eq.s64 %p88, %rd697, 0; @%p88 bra $L__BB148_137; ld.f32 %f787, [%rd14+-12]; ld.f32 %f788, [%rd14+-8]; ld.f32 %f789, [%rd14+-4]; st.local.f32 [%rd696+20], %f787; st.local.f32 [%rd696+24], %f788; st.local.f32 [%rd696+28], %f789; ld.f32 %f790, [%rd14+24]; st.local.f32 [%rd696+32], %f790; mov.u32 %r196, 5; st.local.u32 [%rd696+12], %r196; mov.u32 %r197, 8; st.local.u32 [%rd696+36], %r197; ld.local.f32 %f791, [%rd4+24]; ld.local.f32 %f792, [%rd4+28]; mul.ftz.f32 %f793, %f792, %f788; fma.rn.ftz.f32 %f794, %f791, %f787, %f793; ld.local.f32 %f795, [%rd4+32]; fma.rn.ftz.f32 %f796, %f795, %f789, %f794; mov.f32 %f797, 0f3F800000; sub.ftz.f32 %f798, %f797, %f796; max.ftz.f32 %f800, %f798, %f781; min.ftz.f32 %f801, %f800, %f797; mul.ftz.f32 %f802, %f801, %f801; mul.ftz.f32 %f803, %f802, %f802; mul.ftz.f32 %f804, %f801, %f803; mul.ftz.f32 %f805, %f790, %f804; ld.local.f32 %f806, [%rd696+16]; mul.ftz.f32 %f807, %f806, %f805; st.local.f32 [%rd696+16], %f807; mov.u32 %r643, 12; $L__BB148_137: ld.f32 %f809, [%rd14]; setp.geu.ftz.f32 %p89, %f809, 0f322BCC77; mov.u32 %r641, 3; mov.f32 %f3001, %f3003; @%p89 bra $L__BB148_139; ld.f32 %f3001, [%rd14+-32]; mov.u32 %r200, 0; st.u32 [%rd14+-32], %r200; st.u32 [%rd14], %r200; mov.u32 %r641, 2; $L__BB148_139: ld.f32 %f811, [%rd14+4]; setp.geu.ftz.f32 %p90, %f811, 0f322BCC77; mov.f32 %f3002, %f3003; @%p90 bra $L__BB148_141; ld.f32 %f3002, [%rd14+-28]; mov.u32 %r201, 0; st.u32 [%rd14+-28], %r201; st.u32 [%rd14+4], %r201; add.s32 %r641, %r641, -1; $L__BB148_141: ld.f32 %f813, [%rd14+8]; setp.geu.ftz.f32 %p91, %f813, 0f322BCC77; @%p91 bra $L__BB148_143; ld.f32 %f3003, [%rd14+-24]; mov.u32 %r202, 0; st.u32 [%rd14+-24], %r202; st.u32 [%rd14+8], %r202; add.s32 %r641, %r641, -1; $L__BB148_143: setp.gt.s32 %p92, %r641, 2; @%p92 bra $L__BB148_159; ld.f32 %f814, [%rd14+24]; setp.neu.ftz.f32 %p93, %f814, 0f7F7FFFFF; mov.f32 %f815, 0f00000000; max.ftz.f32 %f200, %f3002, %f815; max.ftz.f32 %f201, %f3001, %f815; add.ftz.f32 %f816, %f201, %f200; max.ftz.f32 %f202, %f3003, %f815; add.ftz.f32 %f817, %f202, %f816; mul.ftz.f32 %f818, %f817, 0f3EAAAAAB; abs.ftz.f32 %f203, %f818; @%p93 bra $L__BB148_152; bra.uni $L__BB148_145; $L__BB148_152: setp.ltu.ftz.f32 %p98, %f203, 0f3727C5AC; mov.u64 %rd704, 0; mov.u64 %rd705, %rd704; @%p98 bra $L__BB148_157; ld.local.u32 %r38, [%rd4+252]; mov.u64 %rd704, 0; setp.eq.s32 %p99, %r38, 0; mov.u64 %rd702, %rd704; mov.u64 %rd703, %rd704; @%p99 bra $L__BB148_155; ld.local.u32 %r212, [%rd4+248]; mul.wide.s32 %rd319, %r212, 80; add.s64 %rd320, %rd1, %rd319; add.s64 %rd702, %rd320, 304; add.s64 %rd321, %rd230, %rd319; add.s64 %rd703, %rd321, 304; mov.b32 %r213, %f202; mov.b32 %r214, %f200; mov.b32 %r215, %f201; mov.u32 %r216, 0; st.local.v4.u32 [%rd320+304], {%r215, %r214, %r213, %r216}; add.s32 %r217, %r212, 1; st.local.u32 [%rd4+248], %r217; add.s32 %r218, %r38, -1; st.local.u32 [%rd4+252], %r218; $L__BB148_155: setp.eq.s64 %p100, %rd703, 0; mov.u64 %rd705, %rd704; @%p100 bra $L__BB148_157; st.local.f32 [%rd702+16], %f203; mov.u64 %rd704, %rd702; mov.u64 %rd705, %rd703; $L__BB148_157: setp.eq.s64 %p101, %rd705, 0; @%p101 bra $L__BB148_159; ld.f32 %f822, [%rd14+-12]; ld.f32 %f823, [%rd14+-8]; ld.f32 %f824, [%rd14+-4]; st.local.f32 [%rd704+20], %f822; st.local.f32 [%rd704+24], %f823; st.local.f32 [%rd704+28], %f824; ld.f32 %f825, [%rd14+24]; st.local.f32 [%rd704+32], %f825; mov.u32 %r220, 5; st.local.u32 [%rd704+12], %r220; mov.u32 %r221, 2; st.local.u32 [%rd704+36], %r221; mov.u32 %r643, 12; bra.uni $L__BB148_159; $L__BB148_373: and.b32 %r419, %r110, 7; setp.eq.s32 %p330, %r419, 0; mov.f32 %f3038, 0f00000000; @%p330 bra $L__BB148_385; setp.eq.s32 %p331, %r110, 4; selp.b32 %r421, %r404, 0, %p331; add.s32 %r422, %r113, %r421; ld.const.u64 %rd513, [kernel_params+2320]; cvta.to.global.u64 %rd514, %rd513; mul.wide.s32 %rd515, %r422, 4; add.s64 %rd516, %rd514, %rd515; ld.global.f32 %f3038, [%rd516]; bra.uni $L__BB148_385; $L__BB148_498: ld.local.u32 %r583, [%rd4+248]; add.s32 %r584, %r583, -1; st.local.u32 [%rd4+248], %r584; add.s32 %r585, %r131, 1; st.local.u32 [%rd4+252], %r585; mov.u64 %rd795, %rd794; $L__BB148_499: st.local.u64 [%rd205+28], %rd795; setp.eq.s64 %p441, %rd795, 0; @%p441 bra $L__BB148_517; ld.local.f32 %f2886, [%rd7]; ld.local.f32 %f2887, [%rd7+4]; ld.local.f32 %f2888, [%rd7+8]; st.local.f32 [%rd794], %f2886; st.local.f32 [%rd794+4], %f2887; st.local.f32 [%rd794+8], %f2888; ld.local.u64 %rd659, [%rd205+28]; mov.u32 %r586, 0; st.u32 [%rd659+12], %r586; st.u32 [%rd659+16], %r586; st.u32 [%rd659+20], %r586; ld.local.u64 %rd660, [%rd205+28]; st.u32 [%rd660+36], %r586; ld.local.v2.f32 {%f2889, %f2890}, [%rd205+36]; setp.neu.ftz.f32 %p442, %f2889, 0f00000000; setp.neu.ftz.f32 %p443, %f2890, 0f00000000; or.pred %p444, %p442, %p443; ld.local.f32 %f2893, [%rd205+44]; setp.neu.ftz.f32 %p445, %f2893, 0f00000000; or.pred %p446, %p445, %p444; @%p446 bra $L__BB148_502; mov.f32 %f2895, 0f3F800000; st.local.v2.f32 [%rd205+36], {%f2895, %f2837}; st.local.u32 [%rd205+44], %r586; $L__BB148_502: mov.u32 %r588, 14; st.local.u32 [%rd205+-8], %r588; ld.local.v2.f32 {%f2896, %f2897}, [%rd205+12]; mov.f32 %f2900, 0f38D1B717; max.ftz.f32 %f2901, %f2896, %f2900; mov.f32 %f2902, 0f3F800000; max.ftz.f32 %f2903, %f2897, %f2900; min.ftz.f32 %f2904, %f2903, %f2902; min.ftz.f32 %f2905, %f2901, %f2902; st.local.v2.f32 [%rd205+12], {%f2905, %f2904}; ld.local.u64 %rd661, [%rd205+28]; ld.f32 %f2906, [%rd661]; cvt.ftz.sat.f32.f32 %f2907, %f2906; ld.f32 %f2908, [%rd661+4]; cvt.ftz.sat.f32.f32 %f2909, %f2908; ld.f32 %f2910, [%rd661+8]; cvt.ftz.sat.f32.f32 %f2911, %f2910; st.f32 [%rd661], %f2907; st.f32 [%rd661+4], %f2909; st.f32 [%rd661+8], %f2911; ld.local.u64 %rd662, [%rd205+28]; ld.f32 %f2912, [%rd662+12]; cvt.ftz.sat.f32.f32 %f2913, %f2912; ld.f32 %f2914, [%rd662+16]; cvt.ftz.sat.f32.f32 %f2915, %f2914; ld.f32 %f2916, [%rd662+20]; cvt.ftz.sat.f32.f32 %f2917, %f2916; st.f32 [%rd662+12], %f2913; st.f32 [%rd662+16], %f2915; st.f32 [%rd662+20], %f2917; ld.local.u32 %r589, [%rd4+40]; or.b32 %r590, %r589, 1036; st.local.u32 [%rd4+40], %r590; bra.uni $L__BB148_517; $L__BB148_371: mad.lo.s32 %r418, %r404, 3, %r113; ld.const.u64 %rd510, [kernel_params+2320]; cvta.to.global.u64 %rd511, %rd510; mul.wide.s32 %rd512, %r418, 4; add.s64 %rd755, %rd511, %rd512; add.s64 %rd756, %rd755, 4; add.s64 %rd757, %rd755, 8; $L__BB148_372: ld.local.f32 %f2509, [%rd4+56]; ld.global.f32 %f2510, [%rd756]; ld.local.f32 %f2511, [%rd4+60]; ld.global.f32 %f2512, [%rd757]; mul.ftz.f32 %f2513, %f2512, %f2511; fma.rn.ftz.f32 %f2514, %f2510, %f2509, %f2513; mov.f32 %f2515, 0f3F800000; sub.ftz.f32 %f2516, %f2515, %f2509; sub.ftz.f32 %f2517, %f2516, %f2511; ld.global.f32 %f2518, [%rd755]; fma.rn.ftz.f32 %f3038, %f2518, %f2517, %f2514; $L__BB148_385: mov.f32 %f2528, 0f00000000; max.ftz.f32 %f399, %f385, %f2528; max.ftz.f32 %f400, %f386, %f2528; add.ftz.f32 %f2529, %f400, %f399; max.ftz.f32 %f401, %f384, %f2528; add.ftz.f32 %f2530, %f401, %f2529; mul.ftz.f32 %f2531, %f2530, 0f3EAAAAAB; abs.ftz.f32 %f402, %f2531; setp.ltu.ftz.f32 %p340, %f402, 0f3727C5AC; mov.u64 %rd760, 0; mov.u64 %rd761, %rd760; @%p340 bra $L__BB148_390; ld.local.u32 %r118, [%rd4+252]; mov.u64 %rd760, 0; setp.eq.s32 %p341, %r118, 0; mov.u64 %rd758, %rd760; mov.u64 %rd759, %rd760; @%p341 bra $L__BB148_388; ld.local.u32 %r435, [%rd4+248]; mul.wide.s32 %rd539, %r435, 80; add.s64 %rd540, %rd1, %rd539; add.s64 %rd758, %rd540, 304; add.s64 %rd541, %rd230, %rd539; add.s64 %rd759, %rd541, 304; mov.b32 %r436, %f401; mov.b32 %r437, %f399; mov.b32 %r438, %f400; mov.u32 %r439, 0; st.local.v4.u32 [%rd540+304], {%r438, %r437, %r436, %r439}; add.s32 %r440, %r435, 1; st.local.u32 [%rd4+248], %r440; add.s32 %r441, %r118, -1; st.local.u32 [%rd4+252], %r441; $L__BB148_388: setp.eq.s64 %p342, %rd759, 0; mov.u64 %rd761, %rd760; @%p342 bra $L__BB148_390; st.local.f32 [%rd758+16], %f402; mov.u64 %rd760, %rd758; mov.u64 %rd761, %rd759; $L__BB148_390: add.s32 %r663, %r142, 4; setp.eq.s64 %p343, %rd761, 0; @%p343 bra $L__BB148_517; ld.local.u32 %r120, [%rd4+252]; setp.lt.s32 %p344, %r120, 1; @%p344 bra $L__BB148_393; add.s32 %r442, %r120, -1; st.local.u32 [%rd4+252], %r442; ld.local.s32 %rd544, [%rd4+248]; cvt.s64.s32 %rd545, %r442; add.s64 %rd546, %rd544, %rd545; mul.lo.s64 %rd547, %rd546, 80; add.s64 %rd548, %rd230, %rd547; add.s64 %rd763, %rd548, 304; add.s64 %rd549, %rd1, %rd547; add.s64 %rd762, %rd549, 304; bra.uni $L__BB148_394; $L__BB148_393: ld.local.u32 %r443, [%rd4+248]; mov.u64 %rd762, 0; add.s32 %r444, %r443, -1; st.local.u32 [%rd4+248], %r444; add.s32 %r445, %r120, 1; st.local.u32 [%rd4+252], %r445; mov.u64 %rd763, %rd762; $L__BB148_394: setp.eq.s64 %p345, %rd763, 0; @%p345 bra $L__BB148_517; setp.eq.s32 %p346, %r91, 255; @%p346 bra $L__BB148_397; bra.uni $L__BB148_396; $L__BB148_397: mov.b32 %f3039, %r382; bra.uni $L__BB148_398; $L__BB148_396: mul.wide.u32 %rd552, %r91, 4; add.s64 %rd553, %rd2, %rd552; ld.local.f32 %f3039, [%rd553]; $L__BB148_398: add.ftz.f32 %f2532, %f3038, 0fBF000000; add.ftz.f32 %f406, %f2532, %f2532; fma.rn.ftz.f32 %f2533, %f406, %f3039, 0f3F800000; mul.ftz.f32 %f407, %f60, %f2533; mul.ftz.f32 %f408, %f63, %f2533; setp.eq.s32 %p347, %r86, 255; @%p347 bra $L__BB148_400; bra.uni $L__BB148_399; $L__BB148_400: mov.b32 %f3040, %r376; bra.uni $L__BB148_401; $L__BB148_399: mul.wide.u32 %rd554, %r86, 4; add.s64 %rd555, %rd2, %rd554; ld.local.f32 %f3040, [%rd555]; $L__BB148_401: max.ftz.f32 %f2535, %f3040, %f2528; mov.f32 %f2536, 0f3F800000; min.ftz.f32 %f2537, %f2535, %f2536; sub.ftz.f32 %f412, %f2536, %f2537; add.s64 %rd153, %rd760, 20; st.local.f32 [%rd760+20], %f55; st.local.f32 [%rd760+24], %f56; st.local.f32 [%rd760+28], %f57; st.local.f32 [%rd760+44], %f407; st.local.v2.f32 [%rd760+48], {%f408, %f3036}; st.local.v2.f32 [%rd760+56], {%f3037, %f412}; st.local.u64 [%rd760+64], %rd763; shr.u32 %r446, %r144, 24; cvt.u16.u32 %rs15, %r446; setp.eq.s16 %p348, %rs15, 0; @%p348 bra $L__BB148_415; setp.eq.s16 %p349, %rs15, 1; @%p349 bra $L__BB148_405; setp.ne.s16 %p350, %rs15, 2; @%p350 bra $L__BB148_416; shr.u32 %r447, %r375, 24; mul.wide.u32 %rd556, %r447, 4; add.s64 %rd557, %rd2, %rd556; ld.local.f32 %f2538, [%rd557+8]; ld.local.f32 %f2539, [%rd557+4]; ld.local.f32 %f2540, [%rd557]; st.local.v2.f32 [%rd153+12], {%f2540, %f2539}; st.local.f32 [%rd153+20], %f2538; bra.uni $L__BB148_417; $L__BB148_415: shr.u32 %r449, %r144, 14; cvt.u64.u32 %rd567, %r449; and.b64 %rd568, %rd567, 1020; add.s64 %rd569, %rd2, %rd568; ld.local.f32 %f2581, [%rd569]; lg2.approx.ftz.f32 %f2582, %f2581; mul.ftz.f32 %f2583, %f2582, 0f3F317218; ld.local.f32 %f2584, [%rd569+4]; lg2.approx.ftz.f32 %f2585, %f2584; mul.ftz.f32 %f2586, %f2585, 0f3F317218; ld.local.f32 %f2587, [%rd569+8]; lg2.approx.ftz.f32 %f2588, %f2587; mul.ftz.f32 %f2589, %f2588, 0f3F317218; fma.rn.ftz.f32 %f2590, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2591, %f408, %f2590, 0fC12BAE14; fma.rn.ftz.f32 %f2592, %f408, %f2591, 0f40220C4A; fma.rn.ftz.f32 %f2593, %f408, %f2592, 0fBE5C28F6; fma.rn.ftz.f32 %f2594, %f408, %f2593, 0f40BF020C; div.approx.ftz.f32 %f2595, %f2583, %f2594; div.approx.ftz.f32 %f2596, %f2586, %f2594; div.approx.ftz.f32 %f2597, %f2589, %f2594; mul.ftz.f32 %f2598, %f2597, %f2597; mul.ftz.f32 %f2599, %f2596, %f2596; mul.ftz.f32 %f2600, %f2595, %f2595; st.local.v2.f32 [%rd153+12], {%f2600, %f2599}; st.local.f32 [%rd153+20], %f2598; bra.uni $L__BB148_417; $L__BB148_405: setp.eq.s32 %p351, %r87, 255; @%p351 bra $L__BB148_407; bra.uni $L__BB148_406; $L__BB148_407: mov.b32 %f3041, %r377; bra.uni $L__BB148_408; $L__BB148_416: mov.f32 %f2601, 0f3F172384; mov.f32 %f2602, 0f3E8D729D; st.local.v2.f32 [%rd153+12], {%f2602, %f2601}; mov.u32 %r450, 1069964113; st.local.u32 [%rd153+20], %r450; bra.uni $L__BB148_417; $L__BB148_406: mul.wide.u32 %rd558, %r87, 4; add.s64 %rd559, %rd2, %rd558; ld.local.f32 %f3041, [%rd559]; $L__BB148_408: setp.eq.s32 %p352, %r88, 255; @%p352 bra $L__BB148_410; bra.uni $L__BB148_409; $L__BB148_410: mov.b32 %f3042, %r378; bra.uni $L__BB148_411; $L__BB148_409: mul.wide.u32 %rd560, %r88, 4; add.s64 %rd561, %rd2, %rd560; ld.local.f32 %f3042, [%rd561]; $L__BB148_411: setp.eq.s32 %p353, %r90, 255; @%p353 bra $L__BB148_413; bra.uni $L__BB148_412; $L__BB148_413: mov.b32 %f3043, %r381; bra.uni $L__BB148_414; $L__BB148_412: mul.wide.u32 %rd562, %r90, 4; add.s64 %rd563, %rd2, %rd562; ld.local.f32 %f3043, [%rd563]; $L__BB148_414: max.ftz.f32 %f2542, %f3043, %f2528; min.ftz.f32 %f2544, %f2542, %f2536; fma.rn.ftz.f32 %f2545, %f406, %f2544, 0f3F800000; mul.ftz.f32 %f2546, %f3041, %f2545; sub.ftz.f32 %f2547, %f2536, %f2546; mov.f32 %f2548, 0f38D1B717; max.ftz.f32 %f2549, %f2547, %f2548; lg2.approx.ftz.f32 %f2550, %f2549; mul.ftz.f32 %f2551, %f2550, 0fBF317218; sub.ftz.f32 %f2552, %f2536, %f3042; mul.ftz.f32 %f2553, %f2552, %f2551; mul.ftz.f32 %f2554, %f3042, %f2551; mul.ftz.f32 %f2555, %f2554, 0f3EAF9DB2; mul.ftz.f32 %f2556, %f2554, 0f3F3BA5E3; mul.ftz.f32 %f2557, %f2554, 0f3FF645A2; fma.rn.ftz.f32 %f2558, %f2553, 0f3F018937, %f2555; fma.rn.ftz.f32 %f2559, %f2553, 0f3F574BC7, %f2556; fma.rn.ftz.f32 %f2560, %f2553, 0f3FD39581, %f2557; shl.b32 %r448, %r379, 2; cvt.u64.u32 %rd564, %r448; and.b64 %rd565, %rd564, 1020; add.s64 %rd566, %rd2, %rd565; ld.local.f32 %f2561, [%rd566]; lg2.approx.ftz.f32 %f2562, %f2561; mul.ftz.f32 %f2563, %f2562, 0f3F317218; ld.local.f32 %f2564, [%rd566+4]; lg2.approx.ftz.f32 %f2565, %f2564; mul.ftz.f32 %f2566, %f2565, 0f3F317218; ld.local.f32 %f2567, [%rd566+8]; lg2.approx.ftz.f32 %f2568, %f2567; mul.ftz.f32 %f2569, %f2568, 0f3F317218; fma.rn.ftz.f32 %f2570, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2571, %f408, %f2570, 0fC12BAE14; fma.rn.ftz.f32 %f2572, %f408, %f2571, 0f40220C4A; fma.rn.ftz.f32 %f2573, %f408, %f2572, 0fBE5C28F6; fma.rn.ftz.f32 %f2574, %f408, %f2573, 0f40BF020C; div.approx.ftz.f32 %f2575, %f2563, %f2574; div.approx.ftz.f32 %f2576, %f2566, %f2574; div.approx.ftz.f32 %f2577, %f2569, %f2574; fma.rn.ftz.f32 %f2578, %f2577, %f2577, %f2560; fma.rn.ftz.f32 %f2579, %f2576, %f2576, %f2559; fma.rn.ftz.f32 %f2580, %f2575, %f2575, %f2558; st.local.v2.f32 [%rd153+12], {%f2580, %f2579}; st.local.f32 [%rd153+20], %f2578; $L__BB148_417: mov.u32 %r451, 29; st.local.u32 [%rd153+-8], %r451; mov.f32 %f2603, 0f3A83126F; max.ftz.f32 %f2604, %f407, %f2603; min.ftz.f32 %f2606, %f2604, %f2536; max.ftz.f32 %f2607, %f408, %f2603; min.ftz.f32 %f2608, %f2607, %f2536; mul.ftz.f32 %f2609, %f2606, %f412; max.ftz.f32 %f2610, %f2609, %f2603; min.ftz.f32 %f2611, %f2610, %f2536; mul.ftz.f32 %f2612, %f2606, %f2606; mul.ftz.f32 %f2613, %f2612, 0f3F4FDF3B; fma.rn.ftz.f32 %f2614, %f2606, 0f3F39DB23, %f2613; mul.ftz.f32 %f2615, %f2612, %f2612; mul.ftz.f32 %f2616, %f2606, %f2615; mul.ftz.f32 %f2617, %f2616, %f2616; mul.ftz.f32 %f2618, %f2617, %f2617; fma.rn.ftz.f32 %f2619, %f2618, 0f406CCCCD, %f2614; mul.ftz.f32 %f2620, %f2619, %f2619; st.local.f32 [%rd153+24], %f2620; mul.ftz.f32 %f2621, %f2608, %f2608; mul.ftz.f32 %f2622, %f2621, 0f3F98D4FE; fma.rn.ftz.f32 %f2623, %f2608, 0f3E87AE14, %f2622; mul.ftz.f32 %f2624, %f2621, %f2621; mul.ftz.f32 %f2625, %f2608, %f2624; mul.ftz.f32 %f2626, %f2625, %f2625; mul.ftz.f32 %f2627, %f2608, %f2626; mul.ftz.f32 %f2628, %f2627, %f2627; fma.rn.ftz.f32 %f2629, %f2628, 0f40ABE76D, %f2623; mul.ftz.f32 %f2630, %f2629, 0f3F206C99; st.local.f32 [%rd153+28], %f2630; mul.ftz.f32 %f2631, %f2611, %f2611; mul.ftz.f32 %f2632, %f2631, 0f3F4FDF3B; fma.rn.ftz.f32 %f2633, %f2611, 0f3F39DB23, %f2632; mul.ftz.f32 %f2634, %f2631, %f2631; mul.ftz.f32 %f2635, %f2611, %f2634; mul.ftz.f32 %f2636, %f2635, %f2635; mul.ftz.f32 %f2637, %f2636, %f2636; fma.rn.ftz.f32 %f2638, %f2637, 0f406CCCCD, %f2633; mul.ftz.f32 %f2639, %f2638, %f2638; st.local.f32 [%rd153+40], %f2639; ld.local.f32 %f3044, [%rd4+104]; ld.local.f32 %f3045, [%rd4+108]; mul.ftz.f32 %f2640, %f3045, %f3045; fma.rn.ftz.f32 %f2641, %f3044, %f3044, %f2640; ld.local.f32 %f3046, [%rd4+112]; fma.rn.ftz.f32 %f2642, %f3046, %f3046, %f2641; sqrt.approx.ftz.f32 %f425, %f2642; setp.neu.ftz.f32 %p354, %f425, 0f00000000; @%p354 bra $L__BB148_418; bra.uni $L__BB148_419; $L__BB148_418: rcp.approx.ftz.f32 %f2643, %f425; mul.ftz.f32 %f3046, %f3046, %f2643; mul.ftz.f32 %f3045, %f3045, %f2643; mul.ftz.f32 %f3044, %f3044, %f2643; $L__BB148_419: ld.local.f32 %f2644, [%rd4+32]; mul.ftz.f32 %f2645, %f3045, %f2644; ld.local.f32 %f2646, [%rd4+28]; mul.ftz.f32 %f2647, %f3046, %f2646; sub.ftz.f32 %f3047, %f2645, %f2647; ld.local.f32 %f2648, [%rd4+24]; mul.ftz.f32 %f2649, %f3046, %f2648; mul.ftz.f32 %f2650, %f3044, %f2644; sub.ftz.f32 %f3048, %f2649, %f2650; mul.ftz.f32 %f2651, %f3044, %f2646; mul.ftz.f32 %f2652, %f3045, %f2648; sub.ftz.f32 %f3049, %f2651, %f2652; mul.ftz.f32 %f2653, %f3048, %f3048; fma.rn.ftz.f32 %f2654, %f3047, %f3047, %f2653; fma.rn.ftz.f32 %f2655, %f3049, %f3049, %f2654; sqrt.approx.ftz.f32 %f441, %f2655; setp.neu.ftz.f32 %p355, %f441, 0f00000000; @%p355 bra $L__BB148_420; bra.uni $L__BB148_421; $L__BB148_420: rcp.approx.ftz.f32 %f2656, %f441; mul.ftz.f32 %f3049, %f3049, %f2656; mul.ftz.f32 %f3048, %f3048, %f2656; mul.ftz.f32 %f3047, %f3047, %f2656; $L__BB148_421: mul.ftz.f32 %f2657, %f3045, %f3049; mul.ftz.f32 %f2658, %f3046, %f3048; sub.ftz.f32 %f3050, %f2657, %f2658; mul.ftz.f32 %f2659, %f3046, %f3047; mul.ftz.f32 %f2660, %f3044, %f3049; sub.ftz.f32 %f3051, %f2659, %f2660; mul.ftz.f32 %f2661, %f3044, %f3048; mul.ftz.f32 %f2662, %f3045, %f3047; sub.ftz.f32 %f3052, %f2661, %f2662; mul.ftz.f32 %f2663, %f3051, %f3051; fma.rn.ftz.f32 %f2664, %f3050, %f3050, %f2663; fma.rn.ftz.f32 %f2665, %f3052, %f3052, %f2664; sqrt.approx.ftz.f32 %f457, %f2665; setp.neu.ftz.f32 %p356, %f457, 0f00000000; @%p356 bra $L__BB148_422; bra.uni $L__BB148_423; $L__BB148_422: rcp.approx.ftz.f32 %f2666, %f457; mul.ftz.f32 %f3052, %f2666, %f3052; mul.ftz.f32 %f3051, %f2666, %f3051; mul.ftz.f32 %f3050, %f2666, %f3050; $L__BB148_423: ld.local.u8 %rs16, [%rd4+52]; and.b16 %rs17, %rs16, 4; setp.eq.s16 %p357, %rs17, 0; @%p357 bra $L__BB148_425; bra.uni $L__BB148_424; $L__BB148_425: ld.local.f32 %f2668, [%rd4+16]; mul.ftz.f32 %f2669, %f3046, %f2668; ld.local.f32 %f2670, [%rd4+20]; mul.ftz.f32 %f2671, %f3045, %f2670; sub.ftz.f32 %f2672, %f2669, %f2671; mul.ftz.f32 %f2673, %f3044, %f2670; ld.local.f32 %f2674, [%rd4+12]; mul.ftz.f32 %f2675, %f3046, %f2674; sub.ftz.f32 %f2676, %f2673, %f2675; mul.ftz.f32 %f2677, %f3045, %f2674; mul.ftz.f32 %f2678, %f3044, %f2668; sub.ftz.f32 %f2679, %f2677, %f2678; mul.ftz.f32 %f2680, %f3051, %f2676; fma.rn.ftz.f32 %f2681, %f3050, %f2672, %f2680; fma.rn.ftz.f32 %f3053, %f3052, %f2679, %f2681; bra.uni $L__BB148_426; $L__BB148_424: ld.local.f32 %f2667, [%rd4+60]; neg.ftz.f32 %f3053, %f2667; $L__BB148_426: st.local.v4.f32 [%rd762], {%f3047, %f3048, %f3049, %f3053}; ld.local.u32 %r452, [%rd4+40]; or.b32 %r453, %r452, 1036; st.local.u32 [%rd4+40], %r453; bra.uni $L__BB148_517; $L__BB148_145: setp.ltu.ftz.f32 %p94, %f203, 0f3727C5AC; mov.u64 %rd700, 0; mov.u64 %rd701, %rd700; @%p94 bra $L__BB148_150; ld.local.u32 %r37, [%rd4+252]; mov.u64 %rd700, 0; setp.eq.s32 %p95, %r37, 0; mov.u64 %rd698, %rd700; mov.u64 %rd699, %rd700; @%p95 bra $L__BB148_148; ld.local.u32 %r203, [%rd4+248]; mul.wide.s32 %rd310, %r203, 80; add.s64 %rd311, %rd1, %rd310; add.s64 %rd698, %rd311, 304; add.s64 %rd312, %rd230, %rd310; add.s64 %rd699, %rd312, 304; mov.b32 %r204, %f202; mov.b32 %r205, %f200; mov.b32 %r206, %f201; mov.u32 %r207, 0; st.local.v4.u32 [%rd311+304], {%r206, %r205, %r204, %r207}; add.s32 %r208, %r203, 1; st.local.u32 [%rd4+248], %r208; add.s32 %r209, %r37, -1; st.local.u32 [%rd4+252], %r209; $L__BB148_148: setp.eq.s64 %p96, %rd699, 0; mov.u64 %rd701, %rd700; @%p96 bra $L__BB148_150; st.local.f32 [%rd698+16], %f203; mov.u64 %rd700, %rd698; mov.u64 %rd701, %rd699; $L__BB148_150: setp.eq.s64 %p97, %rd701, 0; @%p97 bra $L__BB148_159; ld.f32 %f819, [%rd14+-12]; ld.f32 %f820, [%rd14+-8]; ld.f32 %f821, [%rd14+-4]; st.local.f32 [%rd700+20], %f819; st.local.f32 [%rd700+24], %f820; st.local.f32 [%rd700+28], %f821; mov.u32 %r211, 2; st.local.u32 [%rd700+12], %r211; mov.u32 %r643, 12; $L__BB148_159: st.u32 [%rd14+-20], %r149; setp.gt.s32 %p102, %r641, 0; @%p102 bra $L__BB148_161; bra.uni $L__BB148_160; $L__BB148_161: ld.f32 %f826, [%rd14+-28]; ld.f32 %f827, [%rd14+-32]; add.ftz.f32 %f828, %f827, %f826; ld.f32 %f829, [%rd14+-24]; add.ftz.f32 %f830, %f829, %f828; mul.ftz.f32 %f831, %f830, 0f3EAAAAAB; abs.ftz.f32 %f832, %f831; cvt.rn.f32.s32 %f833, %r641; mul.ftz.f32 %f834, %f832, %f833; st.f32 [%rd14+-16], %f834; or.b32 %r223, %r149, 2; setp.eq.s32 %p103, %r223, 34; @%p103 bra $L__BB148_172; bra.uni $L__BB148_162; $L__BB148_172: ld.f32 %f1388, [%rd14]; mul.ftz.f32 %f3007, %f1388, 0f3DA2F983; ld.f32 %f1389, [%rd14+4]; mul.ftz.f32 %f3008, %f1389, 0f3DA2F983; ld.f32 %f1390, [%rd14+8]; mul.ftz.f32 %f3009, %f1390, 0f3DA2F983; bra.uni $L__BB148_173; $L__BB148_160: mov.u32 %r222, 0; st.u32 [%rd14+-16], %r222; bra.uni $L__BB148_174; $L__BB148_162: mov.f32 %f836, 0f3F8147AE; max.ftz.f32 %f837, %f2996, %f836; mov.f32 %f838, 0f40733333; min.ftz.f32 %f839, %f837, %f838; rcp.approx.ftz.f32 %f840, %f839; mov.f32 %f841, 0f3F800000; fma.rn.ftz.f32 %f842, %f840, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f843, %f840, %f842, 0f3F2B020C; fma.rn.ftz.f32 %f844, %f839, 0f3D8240B8, %f843; add.ftz.f32 %f845, %f844, 0f3F800000; mul.ftz.f32 %f846, %f845, 0f3FAAAAAB; sub.ftz.f32 %f847, %f841, %f844; div.approx.ftz.f32 %f204, %f846, %f847; ld.f32 %f205, [%rd14+12]; setp.lt.ftz.f32 %p104, %f205, 0f38D1B717; mov.f32 %f3004, 0f00000000; @%p104 bra $L__BB148_165; setp.ge.ftz.f32 %p105, %f205, 0f3F7EB852; mov.f32 %f3004, 0f3F7FFFEF; @%p105 bra $L__BB148_165; sub.ftz.f32 %f850, %f841, 0f3F000000; mul.ftz.f32 %f851, %f850, 0f40400000; sqrt.approx.ftz.f32 %f852, %f851; mul.ftz.f32 %f853, %f204, %f852; mul.ftz.f32 %f854, %f853, 0fBFB8AA3B; ex2.approx.ftz.f32 %f855, %f854; add.ftz.f32 %f856, %f855, 0f3F800000; mul.ftz.f32 %f857, %f856, 0f3E800000; mul.ftz.f32 %f858, %f852, 0fBFB8AA3B; ex2.approx.ftz.f32 %f859, %f858; mul.ftz.f32 %f860, %f859, %f857; setp.lt.ftz.f32 %p106, %f860, %f205; selp.f32 %f861, 0f3F800000, 0f3F000000, %p106; selp.f32 %f862, 0f3F000000, 0f00000000, %p106; selp.f32 %f863, 0f3F400000, 0f3E800000, %p106; sub.ftz.f32 %f864, %f841, %f863; mul.ftz.f32 %f865, %f864, 0f40400000; sqrt.approx.ftz.f32 %f866, %f865; mul.ftz.f32 %f867, %f863, 0f3F000000; mul.ftz.f32 %f868, %f204, %f866; mul.ftz.f32 %f869, %f868, 0fBFB8AA3B; ex2.approx.ftz.f32 %f870, %f869; add.ftz.f32 %f871, %f870, 0f3F800000; mul.ftz.f32 %f872, %f871, %f867; mul.ftz.f32 %f873, %f866, 0fBFB8AA3B; ex2.approx.ftz.f32 %f874, %f873; mul.ftz.f32 %f875, %f874, %f872; setp.lt.ftz.f32 %p107, %f875, %f205; selp.f32 %f876, %f861, %f863, %p107; selp.f32 %f877, %f863, %f862, %p107; add.ftz.f32 %f878, %f876, %f877; mul.ftz.f32 %f879, %f878, 0f3F000000; sub.ftz.f32 %f880, %f841, %f879; mul.ftz.f32 %f881, %f880, 0f40400000; sqrt.approx.ftz.f32 %f882, %f881; mul.ftz.f32 %f883, %f879, 0f3F000000; mul.ftz.f32 %f884, %f204, %f882; mul.ftz.f32 %f885, %f884, 0fBFB8AA3B; ex2.approx.ftz.f32 %f886, %f885; add.ftz.f32 %f887, %f886, 0f3F800000; mul.ftz.f32 %f888, %f887, %f883; mul.ftz.f32 %f889, %f882, 0fBFB8AA3B; ex2.approx.ftz.f32 %f890, %f889; mul.ftz.f32 %f891, %f890, %f888; setp.lt.ftz.f32 %p108, %f891, %f205; selp.f32 %f892, %f876, %f879, %p108; selp.f32 %f893, %f879, %f877, %p108; add.ftz.f32 %f894, %f892, %f893; mul.ftz.f32 %f895, %f894, 0f3F000000; sub.ftz.f32 %f896, %f841, %f895; mul.ftz.f32 %f897, %f896, 0f40400000; sqrt.approx.ftz.f32 %f898, %f897; mul.ftz.f32 %f899, %f895, 0f3F000000; mul.ftz.f32 %f900, %f204, %f898; mul.ftz.f32 %f901, %f900, 0fBFB8AA3B; ex2.approx.ftz.f32 %f902, %f901; add.ftz.f32 %f903, %f902, 0f3F800000; mul.ftz.f32 %f904, %f903, %f899; mul.ftz.f32 %f905, %f898, 0fBFB8AA3B; ex2.approx.ftz.f32 %f906, %f905; mul.ftz.f32 %f907, %f906, %f904; setp.lt.ftz.f32 %p109, %f907, %f205; selp.f32 %f908, %f892, %f895, %p109; selp.f32 %f909, %f895, %f893, %p109; add.ftz.f32 %f910, %f908, %f909; mul.ftz.f32 %f911, %f910, 0f3F000000; sub.ftz.f32 %f912, %f841, %f911; mul.ftz.f32 %f913, %f912, 0f40400000; sqrt.approx.ftz.f32 %f914, %f913; mul.ftz.f32 %f915, %f911, 0f3F000000; mul.ftz.f32 %f916, %f204, %f914; mul.ftz.f32 %f917, %f916, 0fBFB8AA3B; ex2.approx.ftz.f32 %f918, %f917; add.ftz.f32 %f919, %f918, 0f3F800000; mul.ftz.f32 %f920, %f919, %f915; mul.ftz.f32 %f921, %f914, 0fBFB8AA3B; ex2.approx.ftz.f32 %f922, %f921; mul.ftz.f32 %f923, %f922, %f920; setp.lt.ftz.f32 %p110, %f923, %f205; selp.f32 %f924, %f908, %f911, %p110; selp.f32 %f925, %f911, %f909, %p110; add.ftz.f32 %f926, %f924, %f925; mul.ftz.f32 %f927, %f926, 0f3F000000; sub.ftz.f32 %f928, %f841, %f927; mul.ftz.f32 %f929, %f928, 0f40400000; sqrt.approx.ftz.f32 %f930, %f929; mul.ftz.f32 %f931, %f927, 0f3F000000; mul.ftz.f32 %f932, %f204, %f930; mul.ftz.f32 %f933, %f932, 0fBFB8AA3B; ex2.approx.ftz.f32 %f934, %f933; add.ftz.f32 %f935, %f934, 0f3F800000; mul.ftz.f32 %f936, %f935, %f931; mul.ftz.f32 %f937, %f930, 0fBFB8AA3B; ex2.approx.ftz.f32 %f938, %f937; mul.ftz.f32 %f939, %f938, %f936; setp.lt.ftz.f32 %p111, %f939, %f205; selp.f32 %f940, %f924, %f927, %p111; selp.f32 %f941, %f927, %f925, %p111; add.ftz.f32 %f942, %f940, %f941; mul.ftz.f32 %f943, %f942, 0f3F000000; sub.ftz.f32 %f944, %f841, %f943; mul.ftz.f32 %f945, %f944, 0f40400000; sqrt.approx.ftz.f32 %f946, %f945; mul.ftz.f32 %f947, %f943, 0f3F000000; mul.ftz.f32 %f948, %f204, %f946; mul.ftz.f32 %f949, %f948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f950, %f949; add.ftz.f32 %f951, %f950, 0f3F800000; mul.ftz.f32 %f952, %f951, %f947; mul.ftz.f32 %f953, %f946, 0fBFB8AA3B; ex2.approx.ftz.f32 %f954, %f953; mul.ftz.f32 %f955, %f954, %f952; setp.lt.ftz.f32 %p112, %f955, %f205; selp.f32 %f956, %f940, %f943, %p112; selp.f32 %f957, %f943, %f941, %p112; add.ftz.f32 %f958, %f956, %f957; mul.ftz.f32 %f959, %f958, 0f3F000000; sub.ftz.f32 %f960, %f841, %f959; mul.ftz.f32 %f961, %f960, 0f40400000; sqrt.approx.ftz.f32 %f962, %f961; mul.ftz.f32 %f963, %f959, 0f3F000000; mul.ftz.f32 %f964, %f204, %f962; mul.ftz.f32 %f965, %f964, 0fBFB8AA3B; ex2.approx.ftz.f32 %f966, %f965; add.ftz.f32 %f967, %f966, 0f3F800000; mul.ftz.f32 %f968, %f967, %f963; mul.ftz.f32 %f969, %f962, 0fBFB8AA3B; ex2.approx.ftz.f32 %f970, %f969; mul.ftz.f32 %f971, %f970, %f968; setp.lt.ftz.f32 %p113, %f971, %f205; selp.f32 %f972, %f956, %f959, %p113; selp.f32 %f973, %f959, %f957, %p113; add.ftz.f32 %f974, %f972, %f973; mul.ftz.f32 %f975, %f974, 0f3F000000; sub.ftz.f32 %f976, %f841, %f975; mul.ftz.f32 %f977, %f976, 0f40400000; sqrt.approx.ftz.f32 %f978, %f977; mul.ftz.f32 %f979, %f975, 0f3F000000; mul.ftz.f32 %f980, %f204, %f978; mul.ftz.f32 %f981, %f980, 0fBFB8AA3B; ex2.approx.ftz.f32 %f982, %f981; add.ftz.f32 %f983, %f982, 0f3F800000; mul.ftz.f32 %f984, %f983, %f979; mul.ftz.f32 %f985, %f978, 0fBFB8AA3B; ex2.approx.ftz.f32 %f986, %f985; mul.ftz.f32 %f987, %f986, %f984; setp.lt.ftz.f32 %p114, %f987, %f205; selp.f32 %f988, %f972, %f975, %p114; selp.f32 %f989, %f975, %f973, %p114; add.ftz.f32 %f990, %f988, %f989; mul.ftz.f32 %f991, %f990, 0f3F000000; sub.ftz.f32 %f992, %f841, %f991; mul.ftz.f32 %f993, %f992, 0f40400000; sqrt.approx.ftz.f32 %f994, %f993; mul.ftz.f32 %f995, %f991, 0f3F000000; mul.ftz.f32 %f996, %f204, %f994; mul.ftz.f32 %f997, %f996, 0fBFB8AA3B; ex2.approx.ftz.f32 %f998, %f997; add.ftz.f32 %f999, %f998, 0f3F800000; mul.ftz.f32 %f1000, %f999, %f995; mul.ftz.f32 %f1001, %f994, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1002, %f1001; mul.ftz.f32 %f1003, %f1002, %f1000; setp.lt.ftz.f32 %p115, %f1003, %f205; selp.f32 %f1004, %f988, %f991, %p115; selp.f32 %f1005, %f991, %f989, %p115; add.ftz.f32 %f1006, %f1004, %f1005; mul.ftz.f32 %f1007, %f1006, 0f3F000000; sub.ftz.f32 %f1008, %f841, %f1007; mul.ftz.f32 %f1009, %f1008, 0f40400000; sqrt.approx.ftz.f32 %f1010, %f1009; mul.ftz.f32 %f1011, %f1007, 0f3F000000; mul.ftz.f32 %f1012, %f204, %f1010; mul.ftz.f32 %f1013, %f1012, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1014, %f1013; add.ftz.f32 %f1015, %f1014, 0f3F800000; mul.ftz.f32 %f1016, %f1015, %f1011; mul.ftz.f32 %f1017, %f1010, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1018, %f1017; mul.ftz.f32 %f1019, %f1018, %f1016; setp.lt.ftz.f32 %p116, %f1019, %f205; selp.f32 %f1020, %f1004, %f1007, %p116; selp.f32 %f1021, %f1007, %f1005, %p116; add.ftz.f32 %f1022, %f1020, %f1021; mul.ftz.f32 %f3004, %f1022, 0f3F000000; $L__BB148_165: mov.f32 %f3005, 0f00000000; ld.f32 %f208, [%rd14+16]; setp.lt.ftz.f32 %p117, %f208, 0f38D1B717; @%p117 bra $L__BB148_168; setp.ge.ftz.f32 %p118, %f208, 0f3F7EB852; mov.f32 %f3005, 0f3F7FFFEF; @%p118 bra $L__BB148_168; mov.f32 %f2962, 0f3F800000; sub.ftz.f32 %f1026, %f2962, 0f3F000000; mul.ftz.f32 %f1027, %f1026, 0f40400000; sqrt.approx.ftz.f32 %f1028, %f1027; mul.ftz.f32 %f1029, %f204, %f1028; mul.ftz.f32 %f1030, %f1029, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1031, %f1030; add.ftz.f32 %f1032, %f1031, 0f3F800000; mul.ftz.f32 %f1033, %f1032, 0f3E800000; mul.ftz.f32 %f1034, %f1028, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1035, %f1034; mul.ftz.f32 %f1036, %f1035, %f1033; setp.lt.ftz.f32 %p119, %f1036, %f208; selp.f32 %f1037, 0f3F800000, 0f3F000000, %p119; selp.f32 %f1038, 0f3F000000, 0f00000000, %p119; selp.f32 %f1039, 0f3F400000, 0f3E800000, %p119; sub.ftz.f32 %f1040, %f2962, %f1039; mul.ftz.f32 %f1041, %f1040, 0f40400000; sqrt.approx.ftz.f32 %f1042, %f1041; mul.ftz.f32 %f1043, %f1039, 0f3F000000; mul.ftz.f32 %f1044, %f204, %f1042; mul.ftz.f32 %f1045, %f1044, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1046, %f1045; add.ftz.f32 %f1047, %f1046, 0f3F800000; mul.ftz.f32 %f1048, %f1047, %f1043; mul.ftz.f32 %f1049, %f1042, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1050, %f1049; mul.ftz.f32 %f1051, %f1050, %f1048; setp.lt.ftz.f32 %p120, %f1051, %f208; selp.f32 %f1052, %f1037, %f1039, %p120; selp.f32 %f1053, %f1039, %f1038, %p120; add.ftz.f32 %f1054, %f1052, %f1053; mul.ftz.f32 %f1055, %f1054, 0f3F000000; sub.ftz.f32 %f1056, %f2962, %f1055; mul.ftz.f32 %f1057, %f1056, 0f40400000; sqrt.approx.ftz.f32 %f1058, %f1057; mul.ftz.f32 %f1059, %f1055, 0f3F000000; mul.ftz.f32 %f1060, %f204, %f1058; mul.ftz.f32 %f1061, %f1060, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1062, %f1061; add.ftz.f32 %f1063, %f1062, 0f3F800000; mul.ftz.f32 %f1064, %f1063, %f1059; mul.ftz.f32 %f1065, %f1058, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1066, %f1065; mul.ftz.f32 %f1067, %f1066, %f1064; setp.lt.ftz.f32 %p121, %f1067, %f208; selp.f32 %f1068, %f1052, %f1055, %p121; selp.f32 %f1069, %f1055, %f1053, %p121; add.ftz.f32 %f1070, %f1068, %f1069; mul.ftz.f32 %f1071, %f1070, 0f3F000000; sub.ftz.f32 %f1072, %f2962, %f1071; mul.ftz.f32 %f1073, %f1072, 0f40400000; sqrt.approx.ftz.f32 %f1074, %f1073; mul.ftz.f32 %f1075, %f1071, 0f3F000000; mul.ftz.f32 %f1076, %f204, %f1074; mul.ftz.f32 %f1077, %f1076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1078, %f1077; add.ftz.f32 %f1079, %f1078, 0f3F800000; mul.ftz.f32 %f1080, %f1079, %f1075; mul.ftz.f32 %f1081, %f1074, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1082, %f1081; mul.ftz.f32 %f1083, %f1082, %f1080; setp.lt.ftz.f32 %p122, %f1083, %f208; selp.f32 %f1084, %f1068, %f1071, %p122; selp.f32 %f1085, %f1071, %f1069, %p122; add.ftz.f32 %f1086, %f1084, %f1085; mul.ftz.f32 %f1087, %f1086, 0f3F000000; sub.ftz.f32 %f1088, %f2962, %f1087; mul.ftz.f32 %f1089, %f1088, 0f40400000; sqrt.approx.ftz.f32 %f1090, %f1089; mul.ftz.f32 %f1091, %f1087, 0f3F000000; mul.ftz.f32 %f1092, %f204, %f1090; mul.ftz.f32 %f1093, %f1092, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1094, %f1093; add.ftz.f32 %f1095, %f1094, 0f3F800000; mul.ftz.f32 %f1096, %f1095, %f1091; mul.ftz.f32 %f1097, %f1090, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1098, %f1097; mul.ftz.f32 %f1099, %f1098, %f1096; setp.lt.ftz.f32 %p123, %f1099, %f208; selp.f32 %f1100, %f1084, %f1087, %p123; selp.f32 %f1101, %f1087, %f1085, %p123; add.ftz.f32 %f1102, %f1100, %f1101; mul.ftz.f32 %f1103, %f1102, 0f3F000000; sub.ftz.f32 %f1104, %f2962, %f1103; mul.ftz.f32 %f1105, %f1104, 0f40400000; sqrt.approx.ftz.f32 %f1106, %f1105; mul.ftz.f32 %f1107, %f1103, 0f3F000000; mul.ftz.f32 %f1108, %f204, %f1106; mul.ftz.f32 %f1109, %f1108, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1110, %f1109; add.ftz.f32 %f1111, %f1110, 0f3F800000; mul.ftz.f32 %f1112, %f1111, %f1107; mul.ftz.f32 %f1113, %f1106, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1114, %f1113; mul.ftz.f32 %f1115, %f1114, %f1112; setp.lt.ftz.f32 %p124, %f1115, %f208; selp.f32 %f1116, %f1100, %f1103, %p124; selp.f32 %f1117, %f1103, %f1101, %p124; add.ftz.f32 %f1118, %f1116, %f1117; mul.ftz.f32 %f1119, %f1118, 0f3F000000; sub.ftz.f32 %f1120, %f2962, %f1119; mul.ftz.f32 %f1121, %f1120, 0f40400000; sqrt.approx.ftz.f32 %f1122, %f1121; mul.ftz.f32 %f1123, %f1119, 0f3F000000; mul.ftz.f32 %f1124, %f204, %f1122; mul.ftz.f32 %f1125, %f1124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1126, %f1125; add.ftz.f32 %f1127, %f1126, 0f3F800000; mul.ftz.f32 %f1128, %f1127, %f1123; mul.ftz.f32 %f1129, %f1122, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1130, %f1129; mul.ftz.f32 %f1131, %f1130, %f1128; setp.lt.ftz.f32 %p125, %f1131, %f208; selp.f32 %f1132, %f1116, %f1119, %p125; selp.f32 %f1133, %f1119, %f1117, %p125; add.ftz.f32 %f1134, %f1132, %f1133; mul.ftz.f32 %f1135, %f1134, 0f3F000000; sub.ftz.f32 %f1136, %f2962, %f1135; mul.ftz.f32 %f1137, %f1136, 0f40400000; sqrt.approx.ftz.f32 %f1138, %f1137; mul.ftz.f32 %f1139, %f1135, 0f3F000000; mul.ftz.f32 %f1140, %f204, %f1138; mul.ftz.f32 %f1141, %f1140, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1142, %f1141; add.ftz.f32 %f1143, %f1142, 0f3F800000; mul.ftz.f32 %f1144, %f1143, %f1139; mul.ftz.f32 %f1145, %f1138, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1146, %f1145; mul.ftz.f32 %f1147, %f1146, %f1144; setp.lt.ftz.f32 %p126, %f1147, %f208; selp.f32 %f1148, %f1132, %f1135, %p126; selp.f32 %f1149, %f1135, %f1133, %p126; add.ftz.f32 %f1150, %f1148, %f1149; mul.ftz.f32 %f1151, %f1150, 0f3F000000; sub.ftz.f32 %f1152, %f2962, %f1151; mul.ftz.f32 %f1153, %f1152, 0f40400000; sqrt.approx.ftz.f32 %f1154, %f1153; mul.ftz.f32 %f1155, %f1151, 0f3F000000; mul.ftz.f32 %f1156, %f204, %f1154; mul.ftz.f32 %f1157, %f1156, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1158, %f1157; add.ftz.f32 %f1159, %f1158, 0f3F800000; mul.ftz.f32 %f1160, %f1159, %f1155; mul.ftz.f32 %f1161, %f1154, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1162, %f1161; mul.ftz.f32 %f1163, %f1162, %f1160; setp.lt.ftz.f32 %p127, %f1163, %f208; selp.f32 %f1164, %f1148, %f1151, %p127; selp.f32 %f1165, %f1151, %f1149, %p127; add.ftz.f32 %f1166, %f1164, %f1165; mul.ftz.f32 %f1167, %f1166, 0f3F000000; sub.ftz.f32 %f1168, %f2962, %f1167; mul.ftz.f32 %f1169, %f1168, 0f40400000; sqrt.approx.ftz.f32 %f1170, %f1169; mul.ftz.f32 %f1171, %f1167, 0f3F000000; mul.ftz.f32 %f1172, %f204, %f1170; mul.ftz.f32 %f1173, %f1172, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1174, %f1173; add.ftz.f32 %f1175, %f1174, 0f3F800000; mul.ftz.f32 %f1176, %f1175, %f1171; mul.ftz.f32 %f1177, %f1170, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1178, %f1177; mul.ftz.f32 %f1179, %f1178, %f1176; setp.lt.ftz.f32 %p128, %f1179, %f208; selp.f32 %f1180, %f1164, %f1167, %p128; selp.f32 %f1181, %f1167, %f1165, %p128; add.ftz.f32 %f1182, %f1180, %f1181; mul.ftz.f32 %f1183, %f1182, 0f3F000000; sub.ftz.f32 %f1184, %f2962, %f1183; mul.ftz.f32 %f1185, %f1184, 0f40400000; sqrt.approx.ftz.f32 %f1186, %f1185; mul.ftz.f32 %f1187, %f1183, 0f3F000000; mul.ftz.f32 %f1188, %f204, %f1186; mul.ftz.f32 %f1189, %f1188, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1190, %f1189; add.ftz.f32 %f1191, %f1190, 0f3F800000; mul.ftz.f32 %f1192, %f1191, %f1187; mul.ftz.f32 %f1193, %f1186, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1194, %f1193; mul.ftz.f32 %f1195, %f1194, %f1192; setp.lt.ftz.f32 %p129, %f1195, %f208; selp.f32 %f1196, %f1180, %f1183, %p129; selp.f32 %f1197, %f1183, %f1181, %p129; add.ftz.f32 %f1198, %f1196, %f1197; mul.ftz.f32 %f3005, %f1198, 0f3F000000; $L__BB148_168: mov.f32 %f3006, 0f00000000; ld.f32 %f211, [%rd14+20]; setp.lt.ftz.f32 %p130, %f211, 0f38D1B717; @%p130 bra $L__BB148_171; setp.ge.ftz.f32 %p131, %f211, 0f3F7EB852; mov.f32 %f3006, 0f3F7FFFEF; @%p131 bra $L__BB148_171; mov.f32 %f2957, 0f3F800000; sub.ftz.f32 %f1202, %f2957, 0f3F000000; mul.ftz.f32 %f1203, %f1202, 0f40400000; sqrt.approx.ftz.f32 %f1204, %f1203; mul.ftz.f32 %f1205, %f204, %f1204; mul.ftz.f32 %f1206, %f1205, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1207, %f1206; add.ftz.f32 %f1208, %f1207, 0f3F800000; mul.ftz.f32 %f1209, %f1208, 0f3E800000; mul.ftz.f32 %f1210, %f1204, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1211, %f1210; mul.ftz.f32 %f1212, %f1211, %f1209; setp.lt.ftz.f32 %p132, %f1212, %f211; selp.f32 %f1213, 0f3F800000, 0f3F000000, %p132; selp.f32 %f1214, 0f3F000000, 0f00000000, %p132; selp.f32 %f1215, 0f3F400000, 0f3E800000, %p132; sub.ftz.f32 %f1216, %f2957, %f1215; mul.ftz.f32 %f1217, %f1216, 0f40400000; sqrt.approx.ftz.f32 %f1218, %f1217; mul.ftz.f32 %f1219, %f1215, 0f3F000000; mul.ftz.f32 %f1220, %f204, %f1218; mul.ftz.f32 %f1221, %f1220, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1222, %f1221; add.ftz.f32 %f1223, %f1222, 0f3F800000; mul.ftz.f32 %f1224, %f1223, %f1219; mul.ftz.f32 %f1225, %f1218, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1226, %f1225; mul.ftz.f32 %f1227, %f1226, %f1224; setp.lt.ftz.f32 %p133, %f1227, %f211; selp.f32 %f1228, %f1213, %f1215, %p133; selp.f32 %f1229, %f1215, %f1214, %p133; add.ftz.f32 %f1230, %f1228, %f1229; mul.ftz.f32 %f1231, %f1230, 0f3F000000; sub.ftz.f32 %f1232, %f2957, %f1231; mul.ftz.f32 %f1233, %f1232, 0f40400000; sqrt.approx.ftz.f32 %f1234, %f1233; mul.ftz.f32 %f1235, %f1231, 0f3F000000; mul.ftz.f32 %f1236, %f204, %f1234; mul.ftz.f32 %f1237, %f1236, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1238, %f1237; add.ftz.f32 %f1239, %f1238, 0f3F800000; mul.ftz.f32 %f1240, %f1239, %f1235; mul.ftz.f32 %f1241, %f1234, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1242, %f1241; mul.ftz.f32 %f1243, %f1242, %f1240; setp.lt.ftz.f32 %p134, %f1243, %f211; selp.f32 %f1244, %f1228, %f1231, %p134; selp.f32 %f1245, %f1231, %f1229, %p134; add.ftz.f32 %f1246, %f1244, %f1245; mul.ftz.f32 %f1247, %f1246, 0f3F000000; sub.ftz.f32 %f1248, %f2957, %f1247; mul.ftz.f32 %f1249, %f1248, 0f40400000; sqrt.approx.ftz.f32 %f1250, %f1249; mul.ftz.f32 %f1251, %f1247, 0f3F000000; mul.ftz.f32 %f1252, %f204, %f1250; mul.ftz.f32 %f1253, %f1252, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1254, %f1253; add.ftz.f32 %f1255, %f1254, 0f3F800000; mul.ftz.f32 %f1256, %f1255, %f1251; mul.ftz.f32 %f1257, %f1250, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1258, %f1257; mul.ftz.f32 %f1259, %f1258, %f1256; setp.lt.ftz.f32 %p135, %f1259, %f211; selp.f32 %f1260, %f1244, %f1247, %p135; selp.f32 %f1261, %f1247, %f1245, %p135; add.ftz.f32 %f1262, %f1260, %f1261; mul.ftz.f32 %f1263, %f1262, 0f3F000000; sub.ftz.f32 %f1264, %f2957, %f1263; mul.ftz.f32 %f1265, %f1264, 0f40400000; sqrt.approx.ftz.f32 %f1266, %f1265; mul.ftz.f32 %f1267, %f1263, 0f3F000000; mul.ftz.f32 %f1268, %f204, %f1266; mul.ftz.f32 %f1269, %f1268, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1270, %f1269; add.ftz.f32 %f1271, %f1270, 0f3F800000; mul.ftz.f32 %f1272, %f1271, %f1267; mul.ftz.f32 %f1273, %f1266, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1274, %f1273; mul.ftz.f32 %f1275, %f1274, %f1272; setp.lt.ftz.f32 %p136, %f1275, %f211; selp.f32 %f1276, %f1260, %f1263, %p136; selp.f32 %f1277, %f1263, %f1261, %p136; add.ftz.f32 %f1278, %f1276, %f1277; mul.ftz.f32 %f1279, %f1278, 0f3F000000; sub.ftz.f32 %f1280, %f2957, %f1279; mul.ftz.f32 %f1281, %f1280, 0f40400000; sqrt.approx.ftz.f32 %f1282, %f1281; mul.ftz.f32 %f1283, %f1279, 0f3F000000; mul.ftz.f32 %f1284, %f204, %f1282; mul.ftz.f32 %f1285, %f1284, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1286, %f1285; add.ftz.f32 %f1287, %f1286, 0f3F800000; mul.ftz.f32 %f1288, %f1287, %f1283; mul.ftz.f32 %f1289, %f1282, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1290, %f1289; mul.ftz.f32 %f1291, %f1290, %f1288; setp.lt.ftz.f32 %p137, %f1291, %f211; selp.f32 %f1292, %f1276, %f1279, %p137; selp.f32 %f1293, %f1279, %f1277, %p137; add.ftz.f32 %f1294, %f1292, %f1293; mul.ftz.f32 %f1295, %f1294, 0f3F000000; sub.ftz.f32 %f1296, %f2957, %f1295; mul.ftz.f32 %f1297, %f1296, 0f40400000; sqrt.approx.ftz.f32 %f1298, %f1297; mul.ftz.f32 %f1299, %f1295, 0f3F000000; mul.ftz.f32 %f1300, %f204, %f1298; mul.ftz.f32 %f1301, %f1300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1302, %f1301; add.ftz.f32 %f1303, %f1302, 0f3F800000; mul.ftz.f32 %f1304, %f1303, %f1299; mul.ftz.f32 %f1305, %f1298, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1306, %f1305; mul.ftz.f32 %f1307, %f1306, %f1304; setp.lt.ftz.f32 %p138, %f1307, %f211; selp.f32 %f1308, %f1292, %f1295, %p138; selp.f32 %f1309, %f1295, %f1293, %p138; add.ftz.f32 %f1310, %f1308, %f1309; mul.ftz.f32 %f1311, %f1310, 0f3F000000; sub.ftz.f32 %f1312, %f2957, %f1311; mul.ftz.f32 %f1313, %f1312, 0f40400000; sqrt.approx.ftz.f32 %f1314, %f1313; mul.ftz.f32 %f1315, %f1311, 0f3F000000; mul.ftz.f32 %f1316, %f204, %f1314; mul.ftz.f32 %f1317, %f1316, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1318, %f1317; add.ftz.f32 %f1319, %f1318, 0f3F800000; mul.ftz.f32 %f1320, %f1319, %f1315; mul.ftz.f32 %f1321, %f1314, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1322, %f1321; mul.ftz.f32 %f1323, %f1322, %f1320; setp.lt.ftz.f32 %p139, %f1323, %f211; selp.f32 %f1324, %f1308, %f1311, %p139; selp.f32 %f1325, %f1311, %f1309, %p139; add.ftz.f32 %f1326, %f1324, %f1325; mul.ftz.f32 %f1327, %f1326, 0f3F000000; sub.ftz.f32 %f1328, %f2957, %f1327; mul.ftz.f32 %f1329, %f1328, 0f40400000; sqrt.approx.ftz.f32 %f1330, %f1329; mul.ftz.f32 %f1331, %f1327, 0f3F000000; mul.ftz.f32 %f1332, %f204, %f1330; mul.ftz.f32 %f1333, %f1332, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1334, %f1333; add.ftz.f32 %f1335, %f1334, 0f3F800000; mul.ftz.f32 %f1336, %f1335, %f1331; mul.ftz.f32 %f1337, %f1330, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1338, %f1337; mul.ftz.f32 %f1339, %f1338, %f1336; setp.lt.ftz.f32 %p140, %f1339, %f211; selp.f32 %f1340, %f1324, %f1327, %p140; selp.f32 %f1341, %f1327, %f1325, %p140; add.ftz.f32 %f1342, %f1340, %f1341; mul.ftz.f32 %f1343, %f1342, 0f3F000000; sub.ftz.f32 %f1344, %f2957, %f1343; mul.ftz.f32 %f1345, %f1344, 0f40400000; sqrt.approx.ftz.f32 %f1346, %f1345; mul.ftz.f32 %f1347, %f1343, 0f3F000000; mul.ftz.f32 %f1348, %f204, %f1346; mul.ftz.f32 %f1349, %f1348, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1350, %f1349; add.ftz.f32 %f1351, %f1350, 0f3F800000; mul.ftz.f32 %f1352, %f1351, %f1347; mul.ftz.f32 %f1353, %f1346, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1354, %f1353; mul.ftz.f32 %f1355, %f1354, %f1352; setp.lt.ftz.f32 %p141, %f1355, %f211; selp.f32 %f1356, %f1340, %f1343, %p141; selp.f32 %f1357, %f1343, %f1341, %p141; add.ftz.f32 %f1358, %f1356, %f1357; mul.ftz.f32 %f1359, %f1358, 0f3F000000; sub.ftz.f32 %f1360, %f2957, %f1359; mul.ftz.f32 %f1361, %f1360, 0f40400000; sqrt.approx.ftz.f32 %f1362, %f1361; mul.ftz.f32 %f1363, %f1359, 0f3F000000; mul.ftz.f32 %f1364, %f204, %f1362; mul.ftz.f32 %f1365, %f1364, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1366, %f1365; add.ftz.f32 %f1367, %f1366, 0f3F800000; mul.ftz.f32 %f1368, %f1367, %f1363; mul.ftz.f32 %f1369, %f1362, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f1370, %f1368; setp.lt.ftz.f32 %p142, %f1371, %f211; selp.f32 %f1372, %f1356, %f1359, %p142; selp.f32 %f1373, %f1359, %f1357, %p142; add.ftz.f32 %f1374, %f1372, %f1373; mul.ftz.f32 %f3006, %f1374, 0f3F000000; $L__BB148_171: mov.f32 %f2958, 0f3F800000; sub.ftz.f32 %f1376, %f2958, %f3004; mul.ftz.f32 %f1377, %f1376, 0f40400000; sub.ftz.f32 %f1378, %f2958, %f3005; mul.ftz.f32 %f1379, %f1378, 0f40400000; sub.ftz.f32 %f1380, %f2958, %f3006; mul.ftz.f32 %f1381, %f1380, 0f40400000; sqrt.approx.ftz.f32 %f1382, %f1377; sqrt.approx.ftz.f32 %f1383, %f1379; sqrt.approx.ftz.f32 %f1384, %f1381; ld.f32 %f1385, [%rd14]; mul.ftz.f32 %f3007, %f1382, %f1385; ld.f32 %f1386, [%rd14+4]; mul.ftz.f32 %f3008, %f1383, %f1386; ld.f32 %f1387, [%rd14+8]; mul.ftz.f32 %f3009, %f1384, %f1387; $L__BB148_173: st.f32 [%rd14], %f3007; st.f32 [%rd14+4], %f3008; st.f32 [%rd14+8], %f3009; or.b32 %r643, %r643, 16; $L__BB148_174: ld.local.u32 %r224, [%rd4+40]; or.b32 %r225, %r224, %r643; st.local.u32 [%rd4+40], %r225; $L__BB148_182: setp.leu.ftz.f32 %p147, %f97, 0f3727C5AC; setp.leu.ftz.f32 %p148, %f74, 0f3727C5AC; or.pred %p149, %p148, %p147; @%p149 bra $L__BB148_194; ld.const.v4.f32 {%f1404, %f1405, %f1406, %f1407}, [kernel_params+1184]; mul.ftz.f32 %f1411, %f187, %f1405; fma.rn.ftz.f32 %f1412, %f186, %f1404, %f1411; fma.rn.ftz.f32 %f227, %f188, %f1406, %f1412; setp.leu.ftz.f32 %p150, %f227, 0f00000000; mov.f32 %f1403, 0f3F800000; mov.f32 %f3010, %f1403; mov.f32 %f3011, %f1403; mov.f32 %f3012, %f1403; @%p150 bra $L__BB148_185; div.approx.ftz.f32 %f3012, %f186, %f227; div.approx.ftz.f32 %f3011, %f187, %f227; div.approx.ftz.f32 %f3010, %f188, %f227; $L__BB148_185: sub.ftz.f32 %f1414, %f1403, %f75; fma.rn.ftz.f32 %f1415, %f75, %f3012, %f1414; fma.rn.ftz.f32 %f1416, %f75, %f3011, %f1414; fma.rn.ftz.f32 %f1417, %f75, %f3010, %f1414; mul.ftz.f32 %f1418, %f74, %f177; mul.ftz.f32 %f1419, %f1418, %f1415; mul.ftz.f32 %f1420, %f74, %f178; mul.ftz.f32 %f1421, %f1420, %f1416; mul.ftz.f32 %f1422, %f74, %f179; mul.ftz.f32 %f1423, %f1422, %f1417; mul.ftz.f32 %f1424, %f97, %f1419; mul.ftz.f32 %f1425, %f97, %f1421; mul.ftz.f32 %f1426, %f97, %f1423; mov.f32 %f3013, 0f00000000; max.ftz.f32 %f234, %f1424, %f3013; max.ftz.f32 %f235, %f1425, %f3013; max.ftz.f32 %f236, %f1426, %f3013; add.ftz.f32 %f1428, %f234, %f235; add.ftz.f32 %f1429, %f236, %f1428; mul.ftz.f32 %f1430, %f1429, 0f3EAAAAAB; abs.ftz.f32 %f237, %f1430; setp.ltu.ftz.f32 %p151, %f237, 0f3727C5AC; mov.u64 %rd712, 0; mov.u64 %rd713, %rd712; @%p151 bra $L__BB148_190; ld.local.u32 %r43, [%rd4+252]; mov.u64 %rd712, 0; setp.eq.s32 %p152, %r43, 0; mov.u64 %rd710, %rd712; mov.u64 %rd711, %rd712; @%p152 bra $L__BB148_188; ld.local.u32 %r237, [%rd4+248]; mul.wide.s32 %rd337, %r237, 80; add.s64 %rd338, %rd1, %rd337; add.s64 %rd710, %rd338, 304; add.s64 %rd339, %rd230, %rd337; add.s64 %rd711, %rd339, 304; mov.b32 %r238, %f236; mov.b32 %r239, %f235; mov.b32 %r240, %f234; mov.u32 %r241, 0; st.local.v4.u32 [%rd338+304], {%r240, %r239, %r238, %r241}; add.s32 %r242, %r237, 1; st.local.u32 [%rd4+248], %r242; add.s32 %r243, %r43, -1; st.local.u32 [%rd4+252], %r243; $L__BB148_188: setp.eq.s64 %p153, %rd711, 0; mov.u64 %rd713, %rd712; @%p153 bra $L__BB148_190; st.local.f32 [%rd710+16], %f237; mov.u64 %rd712, %rd710; mov.u64 %rd713, %rd711; $L__BB148_190: setp.eq.s64 %p154, %rd713, 0; @%p154 bra $L__BB148_194; add.s64 %rd45, %rd712, 20; st.local.f32 [%rd712+20], %f55; st.local.f32 [%rd712+24], %f56; st.local.f32 [%rd712+28], %f57; mov.u32 %r244, 6; st.local.u32 [%rd712+12], %r244; ld.local.f32 %f1432, [%rd4+24]; ld.local.f32 %f1433, [%rd4+28]; mul.ftz.f32 %f1434, %f56, %f1433; fma.rn.ftz.f32 %f1435, %f55, %f1432, %f1434; ld.local.f32 %f1436, [%rd4+32]; fma.rn.ftz.f32 %f238, %f57, %f1436, %f1435; setp.lt.ftz.f32 %p155, %f238, 0f00000000; @%p155 bra $L__BB148_193; mov.f32 %f1437, 0f3F800000; sub.ftz.f32 %f1438, %f1437, %f238; mov.f32 %f1439, 0f00000000; max.ftz.f32 %f1440, %f1438, %f1439; min.ftz.f32 %f1441, %f1440, %f1437; mul.ftz.f32 %f1442, %f1441, %f1441; mul.ftz.f32 %f1443, %f1442, %f1442; mul.ftz.f32 %f1444, %f1441, %f1443; mul.ftz.f32 %f3013, %f238, %f1444; $L__BB148_193: st.local.f32 [%rd45+12], %f3013; ld.local.f32 %f1445, [%rd45+-4]; mul.ftz.f32 %f1446, %f3013, %f1445; st.local.f32 [%rd45+-4], %f1446; ld.local.u32 %r245, [%rd4+40]; or.b32 %r246, %r245, 12; st.local.u32 [%rd4+40], %r246; $L__BB148_194: ld.param.u32 %r637, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.const.u32 %r44, [kernel_params+1564]; setp.ne.s32 %p156, %r44, 0; and.b32 %r45, %r637, 8; setp.eq.s32 %p157, %r45, 0; or.pred %p5, %p157, %p156; not.pred %p158, %p5; setp.leu.ftz.f32 %p159, %f99, 0f3727C5AC; or.pred %p160, %p159, %p158; @%p160 bra $L__BB148_225; setp.leu.ftz.f32 %p161, %f70, 0f3727C5AC; setp.leu.ftz.f32 %p162, %f60, 0f3727C5AC; and.pred %p163, %p162, %p161; @%p163 bra $L__BB148_225; mul.ftz.f32 %f1447, %f99, %f177; mov.f32 %f1448, 0f00000000; max.ftz.f32 %f241, %f1447, %f1448; mul.ftz.f32 %f1449, %f99, %f178; max.ftz.f32 %f242, %f1449, %f1448; mul.ftz.f32 %f1450, %f99, %f179; max.ftz.f32 %f243, %f1450, %f1448; add.ftz.f32 %f1451, %f241, %f242; add.ftz.f32 %f1452, %f243, %f1451; mul.ftz.f32 %f1453, %f1452, 0f3EAAAAAB; abs.ftz.f32 %f244, %f1453; setp.ltu.ftz.f32 %p164, %f244, 0f3727C5AC; mov.u64 %rd720, 0; mov.u64 %rd716, %rd720; mov.u64 %rd717, %rd720; @%p164 bra $L__BB148_201; ld.local.u32 %r46, [%rd4+252]; mov.u64 %rd716, 0; setp.eq.s32 %p165, %r46, 0; mov.u64 %rd714, %rd716; mov.u64 %rd715, %rd716; @%p165 bra $L__BB148_199; ld.local.u32 %r247, [%rd4+248]; mul.wide.s32 %rd346, %r247, 80; add.s64 %rd347, %rd1, %rd346; add.s64 %rd714, %rd347, 304; add.s64 %rd348, %rd230, %rd346; add.s64 %rd715, %rd348, 304; mov.b32 %r248, %f243; mov.b32 %r249, %f242; mov.b32 %r250, %f241; mov.u32 %r251, 0; st.local.v4.u32 [%rd347+304], {%r250, %r249, %r248, %r251}; add.s32 %r252, %r247, 1; st.local.u32 [%rd4+248], %r252; add.s32 %r253, %r46, -1; st.local.u32 [%rd4+252], %r253; $L__BB148_199: setp.eq.s64 %p166, %rd715, 0; mov.u64 %rd717, %rd716; @%p166 bra $L__BB148_201; st.local.f32 [%rd714+16], %f244; mov.u64 %rd716, %rd714; mov.u64 %rd717, %rd715; $L__BB148_201: setp.eq.s64 %p167, %rd717, 0; mov.u64 %rd721, %rd720; @%p167 bra $L__BB148_205; ld.local.u32 %r47, [%rd4+252]; setp.lt.s32 %p168, %r47, 1; @%p168 bra $L__BB148_204; add.s32 %r254, %r47, -1; st.local.u32 [%rd4+252], %r254; ld.local.s32 %rd353, [%rd4+248]; cvt.s64.s32 %rd354, %r254; add.s64 %rd355, %rd353, %rd354; mul.lo.s64 %rd356, %rd355, 80; add.s64 %rd357, %rd1, %rd356; add.s64 %rd720, %rd357, 304; add.s64 %rd358, %rd230, %rd356; add.s64 %rd721, %rd358, 304; bra.uni $L__BB148_205; $L__BB148_204: ld.local.u32 %r255, [%rd4+248]; mov.u64 %rd720, 0; add.s32 %r256, %r255, -1; st.local.u32 [%rd4+248], %r256; add.s32 %r257, %r47, 1; st.local.u32 [%rd4+252], %r257; mov.u64 %rd721, %rd720; $L__BB148_205: setp.eq.s64 %p169, %rd721, 0; or.pred %p171, %p167, %p169; @%p171 bra $L__BB148_225; add.s64 %rd60, %rd716, 20; st.local.f32 [%rd716+20], %f55; st.local.f32 [%rd716+24], %f56; st.local.f32 [%rd716+28], %f57; mul.ftz.f32 %f245, %f70, 0f3DA3D70A; max.ftz.f32 %f1458, %f245, %f1448; sqrt.approx.ftz.f32 %f1459, %f1458; mov.f32 %f1456, 0f3F800000; sub.ftz.f32 %f1460, %f1456, %f1459; mov.f32 %f1461, 0f40000000; div.approx.ftz.f32 %f1462, %f1461, %f1460; add.ftz.f32 %f1463, %f1462, 0fBF800000; st.local.f32 [%rd716+40], %f1463; st.local.v2.f32 [%rd716+56], {%f2977, %f2978}; st.local.f32 [%rd716+64], %f2979; st.local.u64 [%rd716+48], %rd721; fma.rn.ftz.f32 %f1464, %f73, 0fBF666666, 0f3F800000; max.ftz.f32 %f1465, %f1464, %f1448; sqrt.approx.ftz.f32 %f1466, %f1465; mul.ftz.f32 %f1467, %f71, %f71; mul.ftz.f32 %f1468, %f1467, %f1466; div.approx.ftz.f32 %f1469, %f1467, %f1466; st.local.v2.f32 [%rd716+32], {%f1469, %f1468}; mul.ftz.f32 %f1470, %f187, 0f3F19999A; fma.rn.ftz.f32 %f1471, %f186, 0f3E99999A, %f1470; fma.rn.ftz.f32 %f246, %f188, 0f3DCCCCCD, %f1471; setp.leu.ftz.f32 %p172, %f246, 0f00000000; mov.f32 %f3014, %f1456; mov.f32 %f3015, %f1456; mov.f32 %f3016, %f1456; @%p172 bra $L__BB148_208; div.approx.ftz.f32 %f3016, %f186, %f246; div.approx.ftz.f32 %f3015, %f187, %f246; div.approx.ftz.f32 %f3014, %f188, %f246; $L__BB148_208: sub.ftz.f32 %f1473, %f1456, %f72; fma.rn.ftz.f32 %f1474, %f72, %f3016, %f1473; fma.rn.ftz.f32 %f1475, %f72, %f3015, %f1473; fma.rn.ftz.f32 %f1476, %f72, %f3014, %f1473; mul.ftz.f32 %f1477, %f245, %f1474; mul.ftz.f32 %f1478, %f245, %f1475; mul.ftz.f32 %f1479, %f245, %f1476; sub.ftz.f32 %f1480, %f1456, %f60; mul.ftz.f32 %f1481, %f1480, %f1477; mul.ftz.f32 %f1482, %f1480, %f1478; mul.ftz.f32 %f1483, %f1480, %f1479; fma.rn.ftz.f32 %f1484, %f60, %f186, %f1481; fma.rn.ftz.f32 %f1485, %f60, %f187, %f1482; fma.rn.ftz.f32 %f1486, %f60, %f188, %f1483; st.local.f32 [%rd720+12], %f1484; st.local.f32 [%rd720+16], %f1485; st.local.f32 [%rd720+20], %f1486; ld.local.u64 %rd361, [%rd60+28]; st.f32 [%rd361], %f186; st.f32 [%rd361+4], %f187; st.f32 [%rd361+8], %f188; mov.u32 %r258, 0; ld.local.u64 %rd362, [%rd60+28]; st.u32 [%rd362+36], %r258; setp.le.ftz.f32 %p173, %f71, 0f3D99999A; setp.eq.s32 %p174, %r148, 26; or.pred %p175, %p174, %p173; @%p175 bra $L__BB148_218; bra.uni $L__BB148_209; $L__BB148_218: ld.local.u64 %rd367, [%rd60+28]; ld.f32 %f1573, [%rd367+12]; cvt.ftz.sat.f32.f32 %f1574, %f1573; ld.f32 %f1575, [%rd367+16]; cvt.ftz.sat.f32.f32 %f1576, %f1575; ld.f32 %f1577, [%rd367+20]; cvt.ftz.sat.f32.f32 %f1578, %f1577; st.f32 [%rd367+12], %f1574; st.f32 [%rd367+16], %f1576; st.f32 [%rd367+20], %f1578; ld.local.v2.f32 {%f1579, %f1580}, [%rd60+12]; cvt.ftz.sat.f32.f32 %f1583, %f1580; cvt.ftz.sat.f32.f32 %f1584, %f1579; st.local.v2.f32 [%rd60+12], {%f1584, %f1583}; mov.u32 %r264, 11; st.local.u32 [%rd60+-8], %r264; ld.local.f32 %f1585, [%rd60+20]; fma.rn.ftz.f32 %f266, %f1585, %f1585, 0fBF800000; mov.f32 %f1572, 0f3F800000; abs.ftz.f32 %f267, %f1572; fma.rn.ftz.f32 %f268, %f267, %f267, %f266; setp.leu.ftz.f32 %p184, %f268, 0f00000000; mov.f32 %f3019, %f1572; @%p184 bra $L__BB148_220; sqrt.approx.ftz.f32 %f1586, %f268; sub.ftz.f32 %f1587, %f1586, %f267; add.ftz.f32 %f1588, %f267, %f1586; div.approx.ftz.f32 %f1589, %f1587, %f1588; fma.rn.ftz.f32 %f1590, %f267, %f1588, 0fBF800000; fma.rn.ftz.f32 %f1591, %f267, %f1587, 0f3F800000; div.approx.ftz.f32 %f1592, %f1590, %f1591; mul.ftz.f32 %f1593, %f1589, 0f3F000000; mul.ftz.f32 %f1594, %f1589, %f1593; fma.rn.ftz.f32 %f1595, %f1592, %f1592, 0f3F800000; mul.ftz.f32 %f3019, %f1594, %f1595; $L__BB148_220: ld.local.u64 %rd368, [%rd60+28]; add.s64 %rd62, %rd368, 12; ld.f32 %f271, [%rd368+12]; ld.f32 %f272, [%rd368+16]; ld.f32 %f273, [%rd368+20]; sub.ftz.f32 %f274, %f1572, %f3019; ld.local.f32 %f1597, [%rd60]; ld.local.f32 %f1598, [%rd4+24]; ld.local.f32 %f1599, [%rd60+4]; ld.local.f32 %f1600, [%rd4+28]; mul.ftz.f32 %f1601, %f1600, %f1599; fma.rn.ftz.f32 %f1602, %f1598, %f1597, %f1601; ld.local.f32 %f1603, [%rd60+8]; ld.local.f32 %f1604, [%rd4+32]; fma.rn.ftz.f32 %f1605, %f1604, %f1603, %f1602; abs.ftz.f32 %f275, %f1605; fma.rn.ftz.f32 %f276, %f275, %f275, %f266; setp.leu.ftz.f32 %p185, %f276, 0f00000000; mov.f32 %f3020, %f274; @%p185 bra $L__BB148_222; sqrt.approx.ftz.f32 %f1606, %f276; sub.ftz.f32 %f1607, %f1606, %f275; add.ftz.f32 %f1608, %f275, %f1606; div.approx.ftz.f32 %f1609, %f1607, %f1608; fma.rn.ftz.f32 %f1610, %f275, %f1608, 0fBF800000; fma.rn.ftz.f32 %f1611, %f275, %f1607, 0f3F800000; div.approx.ftz.f32 %f1612, %f1610, %f1611; mul.ftz.f32 %f1613, %f1609, 0f3F000000; mul.ftz.f32 %f1614, %f1609, %f1613; fma.rn.ftz.f32 %f1615, %f1612, %f1612, 0f3F800000; mul.ftz.f32 %f1616, %f1614, %f1615; sub.ftz.f32 %f3020, %f1616, %f3019; $L__BB148_222: rcp.approx.ftz.f32 %f1617, %f274; mov.f32 %f1618, 0f3F800000; mul.ftz.f32 %f1619, %f1617, %f3020; sub.ftz.f32 %f1620, %f1618, %f1619; fma.rn.ftz.f32 %f1621, %f271, %f1620, %f1619; fma.rn.ftz.f32 %f1622, %f272, %f1620, %f1619; fma.rn.ftz.f32 %f1623, %f273, %f1620, %f1619; st.f32 [%rd62+12], %f1621; st.f32 [%rd62+16], %f1622; st.f32 [%rd62+20], %f1623; ld.local.u32 %r265, [%rd60+-8]; setp.ne.s32 %p186, %r265, 12; @%p186 bra $L__BB148_224; ld.local.u64 %rd369, [%rd60+28]; ld.f32 %f1624, [%rd369+36]; mul.ftz.f32 %f1625, %f1624, 0f3E800000; ld.f32 %f1626, [%rd369+24]; mul.ftz.f32 %f1627, %f1626, %f1625; ld.f32 %f1628, [%rd369+28]; mul.ftz.f32 %f1629, %f1625, %f1628; ld.f32 %f1630, [%rd369+32]; mul.ftz.f32 %f1631, %f1625, %f1630; st.f32 [%rd369+24], %f1627; st.f32 [%rd369+28], %f1629; st.f32 [%rd369+32], %f1631; $L__BB148_224: ld.local.u64 %rd370, [%rd60+28]; ld.f32 %f1632, [%rd370+28]; ld.f32 %f1633, [%rd370+24]; add.ftz.f32 %f1634, %f1633, %f1632; ld.f32 %f1635, [%rd370+32]; add.ftz.f32 %f1636, %f1635, %f1634; mul.ftz.f32 %f1637, %f1636, 0f3EAAAAAB; ld.local.f32 %f1638, [%rd60+-4]; mul.ftz.f32 %f1639, %f1638, %f1637; st.local.f32 [%rd60+-4], %f1639; ld.local.u32 %r266, [%rd4+40]; or.b32 %r267, %r266, 12; st.local.u32 [%rd4+40], %r267; bra.uni $L__BB148_225; $L__BB148_209: ld.local.v2.f32 {%f1487, %f1488}, [%rd60+36]; setp.neu.ftz.f32 %p176, %f1487, 0f00000000; setp.neu.ftz.f32 %p177, %f1488, 0f00000000; or.pred %p178, %p176, %p177; ld.local.f32 %f1491, [%rd60+44]; setp.neu.ftz.f32 %p179, %f1491, 0f00000000; or.pred %p180, %p179, %p178; @%p180 bra $L__BB148_211; mov.f32 %f1493, 0f3F800000; st.local.v2.f32 [%rd60+36], {%f1493, %f1448}; st.local.u32 [%rd60+44], %r258; $L__BB148_211: mov.u32 %r260, 15; st.local.u32 [%rd60+-8], %r260; ld.local.f32 %f1495, [%rd60+20]; fma.rn.ftz.f32 %f253, %f1495, %f1495, 0fBF800000; mov.f32 %f1494, 0f3F800000; abs.ftz.f32 %f254, %f1494; fma.rn.ftz.f32 %f255, %f254, %f254, %f253; setp.leu.ftz.f32 %p181, %f255, 0f00000000; mov.f32 %f3017, %f1494; @%p181 bra $L__BB148_213; sqrt.approx.ftz.f32 %f1496, %f255; sub.ftz.f32 %f1497, %f1496, %f254; add.ftz.f32 %f1498, %f254, %f1496; div.approx.ftz.f32 %f1499, %f1497, %f1498; fma.rn.ftz.f32 %f1500, %f254, %f1498, 0fBF800000; fma.rn.ftz.f32 %f1501, %f254, %f1497, 0f3F800000; div.approx.ftz.f32 %f1502, %f1500, %f1501; mul.ftz.f32 %f1503, %f1499, 0f3F000000; mul.ftz.f32 %f1504, %f1499, %f1503; fma.rn.ftz.f32 %f1505, %f1502, %f1502, 0f3F800000; mul.ftz.f32 %f3017, %f1504, %f1505; $L__BB148_213: ld.local.u64 %rd363, [%rd60+28]; add.s64 %rd61, %rd363, 12; ld.f32 %f258, [%rd363+12]; ld.f32 %f259, [%rd363+16]; ld.f32 %f260, [%rd363+20]; sub.ftz.f32 %f261, %f1494, %f3017; ld.local.f32 %f1507, [%rd60]; ld.local.f32 %f1508, [%rd4+24]; ld.local.f32 %f1509, [%rd60+4]; ld.local.f32 %f1510, [%rd4+28]; mul.ftz.f32 %f1511, %f1510, %f1509; fma.rn.ftz.f32 %f1512, %f1508, %f1507, %f1511; ld.local.f32 %f1513, [%rd60+8]; ld.local.f32 %f1514, [%rd4+32]; fma.rn.ftz.f32 %f1515, %f1514, %f1513, %f1512; abs.ftz.f32 %f262, %f1515; fma.rn.ftz.f32 %f263, %f262, %f262, %f253; setp.leu.ftz.f32 %p182, %f263, 0f00000000; mov.f32 %f3018, %f261; @%p182 bra $L__BB148_215; sqrt.approx.ftz.f32 %f1516, %f263; sub.ftz.f32 %f1517, %f1516, %f262; add.ftz.f32 %f1518, %f262, %f1516; div.approx.ftz.f32 %f1519, %f1517, %f1518; fma.rn.ftz.f32 %f1520, %f262, %f1518, 0fBF800000; fma.rn.ftz.f32 %f1521, %f262, %f1517, 0f3F800000; div.approx.ftz.f32 %f1522, %f1520, %f1521; mul.ftz.f32 %f1523, %f1519, 0f3F000000; mul.ftz.f32 %f1524, %f1519, %f1523; fma.rn.ftz.f32 %f1525, %f1522, %f1522, 0f3F800000; mul.ftz.f32 %f1526, %f1524, %f1525; sub.ftz.f32 %f3018, %f1526, %f3017; $L__BB148_215: rcp.approx.ftz.f32 %f1527, %f261; mov.f32 %f1528, 0f3F800000; mul.ftz.f32 %f1529, %f1527, %f3018; sub.ftz.f32 %f1530, %f1528, %f1529; fma.rn.ftz.f32 %f1531, %f258, %f1530, %f1529; fma.rn.ftz.f32 %f1532, %f259, %f1530, %f1529; fma.rn.ftz.f32 %f1533, %f260, %f1530, %f1529; st.f32 [%rd61+12], %f1531; st.f32 [%rd61+16], %f1532; st.f32 [%rd61+20], %f1533; ld.local.u32 %r261, [%rd60+-8]; setp.ne.s32 %p183, %r261, 12; @%p183 bra $L__BB148_217; ld.local.u64 %rd364, [%rd60+28]; ld.f32 %f1534, [%rd364+36]; mul.ftz.f32 %f1535, %f1534, 0f3E800000; ld.f32 %f1536, [%rd364+24]; mul.ftz.f32 %f1537, %f1536, %f1535; ld.f32 %f1538, [%rd364+28]; mul.ftz.f32 %f1539, %f1535, %f1538; ld.f32 %f1540, [%rd364+32]; mul.ftz.f32 %f1541, %f1535, %f1540; st.f32 [%rd364+24], %f1537; st.f32 [%rd364+28], %f1539; st.f32 [%rd364+32], %f1541; $L__BB148_217: ld.local.u64 %rd365, [%rd60+28]; ld.f32 %f1542, [%rd365+28]; ld.f32 %f1543, [%rd365+24]; add.ftz.f32 %f1544, %f1543, %f1542; ld.f32 %f1545, [%rd365+32]; add.ftz.f32 %f1546, %f1545, %f1544; mul.ftz.f32 %f1547, %f1546, 0f3EAAAAAB; ld.local.f32 %f1548, [%rd60+-4]; mul.ftz.f32 %f1549, %f1548, %f1547; st.local.f32 [%rd60+-4], %f1549; ld.local.v2.f32 {%f1550, %f1551}, [%rd60+12]; mov.f32 %f1554, 0f38D1B717; max.ftz.f32 %f1555, %f1550, %f1554; max.ftz.f32 %f1557, %f1551, %f1554; min.ftz.f32 %f1558, %f1557, %f1528; min.ftz.f32 %f1559, %f1555, %f1528; st.local.v2.f32 [%rd60+12], {%f1559, %f1558}; ld.f32 %f1560, [%rd365]; cvt.ftz.sat.f32.f32 %f1561, %f1560; ld.f32 %f1562, [%rd365+4]; cvt.ftz.sat.f32.f32 %f1563, %f1562; ld.f32 %f1564, [%rd365+8]; cvt.ftz.sat.f32.f32 %f1565, %f1564; st.f32 [%rd365], %f1561; st.f32 [%rd365+4], %f1563; st.f32 [%rd365+8], %f1565; ld.local.u64 %rd366, [%rd60+28]; ld.f32 %f1566, [%rd366+12]; cvt.ftz.sat.f32.f32 %f1567, %f1566; ld.f32 %f1568, [%rd366+16]; cvt.ftz.sat.f32.f32 %f1569, %f1568; ld.f32 %f1570, [%rd366+20]; cvt.ftz.sat.f32.f32 %f1571, %f1570; st.f32 [%rd366+12], %f1567; st.f32 [%rd366+16], %f1569; st.f32 [%rd366+20], %f1571; ld.local.u32 %r262, [%rd4+40]; or.b32 %r263, %r262, 1036; st.local.u32 [%rd4+40], %r263; $L__BB148_225: ld.const.u32 %r48, [kernel_params+1568]; or.b32 %r268, %r48, %r44; setp.eq.s32 %p187, %r268, 0; @%p187 bra $L__BB148_227; setp.gt.ftz.f32 %p188, %f98, 0f3727C5AC; @%p188 bra $L__BB148_228; bra.uni $L__BB148_272; $L__BB148_227: setp.leu.ftz.f32 %p189, %f98, 0f3727C5AC; setp.ne.s32 %p190, %r45, 0; or.pred %p191, %p190, %p189; @%p191 bra $L__BB148_272; $L__BB148_228: mul.ftz.f32 %f279, %f98, %f177; mul.ftz.f32 %f280, %f98, %f178; mul.ftz.f32 %f281, %f98, %f179; mov.f32 %f1640, 0f3F800000; sub.ftz.f32 %f1641, %f1640, %f72; fma.rn.ftz.f32 %f282, %f72, %f186, %f1641; fma.rn.ftz.f32 %f283, %f72, %f187, %f1641; fma.rn.ftz.f32 %f284, %f72, %f188, %f1641; setp.eq.s32 %p192, %r148, 26; setp.le.ftz.f32 %p193, %f71, 0f3D4CCCCD; or.pred %p194, %p192, %p193; @%p194 bra $L__BB148_246; bra.uni $L__BB148_229; $L__BB148_246: @%p158 bra $L__BB148_264; mul.ftz.f32 %f1722, %f2981, %f279; mov.f32 %f1723, 0f00000000; max.ftz.f32 %f302, %f1722, %f1723; mul.ftz.f32 %f1724, %f2981, %f280; max.ftz.f32 %f303, %f1724, %f1723; mul.ftz.f32 %f1725, %f2981, %f281; max.ftz.f32 %f304, %f1725, %f1723; add.ftz.f32 %f1726, %f302, %f303; add.ftz.f32 %f1727, %f304, %f1726; mul.ftz.f32 %f1728, %f1727, 0f3EAAAAAB; abs.ftz.f32 %f305, %f1728; setp.ltu.ftz.f32 %p207, %f305, 0f3727C5AC; mov.u64 %rd732, 0; mov.u64 %rd730, %rd732; mov.u64 %rd731, %rd732; @%p207 bra $L__BB148_252; ld.local.u32 %r51, [%rd4+252]; mov.u64 %rd730, 0; setp.eq.s32 %p208, %r51, 0; mov.u64 %rd728, %rd730; mov.u64 %rd729, %rd730; @%p208 bra $L__BB148_250; ld.local.u32 %r285, [%rd4+248]; mul.wide.s32 %rd401, %r285, 80; add.s64 %rd402, %rd1, %rd401; add.s64 %rd728, %rd402, 304; add.s64 %rd403, %rd230, %rd401; add.s64 %rd729, %rd403, 304; mov.b32 %r286, %f304; mov.b32 %r287, %f303; mov.b32 %r288, %f302; mov.u32 %r289, 0; st.local.v4.u32 [%rd402+304], {%r288, %r287, %r286, %r289}; add.s32 %r290, %r285, 1; st.local.u32 [%rd4+248], %r290; add.s32 %r291, %r51, -1; st.local.u32 [%rd4+252], %r291; $L__BB148_250: setp.eq.s64 %p209, %rd729, 0; mov.u64 %rd731, %rd730; @%p209 bra $L__BB148_252; st.local.f32 [%rd728+16], %f305; mov.u64 %rd730, %rd728; mov.u64 %rd731, %rd729; $L__BB148_252: setp.eq.s64 %p210, %rd731, 0; mov.u64 %rd733, %rd732; @%p210 bra $L__BB148_256; ld.local.u32 %r52, [%rd4+252]; setp.lt.s32 %p211, %r52, 1; @%p211 bra $L__BB148_255; add.s32 %r292, %r52, -1; st.local.u32 [%rd4+252], %r292; ld.local.s32 %rd408, [%rd4+248]; cvt.s64.s32 %rd409, %r292; add.s64 %rd410, %rd408, %rd409; mul.lo.s64 %rd411, %rd410, 80; add.s64 %rd412, %rd1, %rd411; add.s64 %rd732, %rd412, 304; add.s64 %rd413, %rd230, %rd411; add.s64 %rd733, %rd413, 304; bra.uni $L__BB148_256; $L__BB148_229: mov.f32 %f1642, 0f00000000; max.ftz.f32 %f285, %f279, %f1642; max.ftz.f32 %f286, %f280, %f1642; add.ftz.f32 %f1643, %f285, %f286; max.ftz.f32 %f287, %f281, %f1642; add.ftz.f32 %f1644, %f287, %f1643; mul.ftz.f32 %f1645, %f1644, 0f3EAAAAAB; abs.ftz.f32 %f288, %f1645; setp.ltu.ftz.f32 %p195, %f288, 0f3727C5AC; mov.u64 %rd726, 0; mov.u64 %rd724, %rd726; mov.u64 %rd725, %rd726; @%p195 bra $L__BB148_234; ld.local.u32 %r49, [%rd4+252]; mov.u64 %rd724, 0; setp.eq.s32 %p196, %r49, 0; mov.u64 %rd722, %rd724; mov.u64 %rd723, %rd724; @%p196 bra $L__BB148_232; ld.local.u32 %r269, [%rd4+248]; mul.wide.s32 %rd375, %r269, 80; add.s64 %rd376, %rd1, %rd375; add.s64 %rd722, %rd376, 304; add.s64 %rd377, %rd230, %rd375; add.s64 %rd723, %rd377, 304; mov.b32 %r270, %f287; mov.b32 %r271, %f286; mov.b32 %r272, %f285; mov.u32 %r273, 0; st.local.v4.u32 [%rd376+304], {%r272, %r271, %r270, %r273}; add.s32 %r274, %r269, 1; st.local.u32 [%rd4+248], %r274; add.s32 %r275, %r49, -1; st.local.u32 [%rd4+252], %r275; $L__BB148_232: setp.eq.s64 %p197, %rd723, 0; mov.u64 %rd725, %rd724; @%p197 bra $L__BB148_234; st.local.f32 [%rd722+16], %f288; mov.u64 %rd724, %rd722; mov.u64 %rd725, %rd723; $L__BB148_234: setp.eq.s64 %p198, %rd725, 0; mov.u64 %rd727, %rd726; @%p198 bra $L__BB148_238; ld.local.u32 %r50, [%rd4+252]; setp.lt.s32 %p199, %r50, 1; @%p199 bra $L__BB148_237; add.s32 %r276, %r50, -1; st.local.u32 [%rd4+252], %r276; ld.local.s32 %rd382, [%rd4+248]; cvt.s64.s32 %rd383, %r276; add.s64 %rd384, %rd382, %rd383; mul.lo.s64 %rd385, %rd384, 80; add.s64 %rd386, %rd1, %rd385; add.s64 %rd726, %rd386, 304; add.s64 %rd387, %rd230, %rd385; add.s64 %rd727, %rd387, 304; bra.uni $L__BB148_238; $L__BB148_237: ld.local.u32 %r277, [%rd4+248]; add.s32 %r278, %r277, -1; st.local.u32 [%rd4+248], %r278; add.s32 %r279, %r50, 1; st.local.u32 [%rd4+252], %r279; mov.u64 %rd727, %rd726; $L__BB148_238: setp.eq.s64 %p200, %rd727, 0; or.pred %p202, %p198, %p200; @%p202 bra $L__BB148_272; add.s64 %rd73, %rd724, 20; st.local.f32 [%rd724+20], %f55; st.local.f32 [%rd724+24], %f56; st.local.f32 [%rd724+28], %f57; st.local.u64 [%rd724+48], %rd727; mov.u32 %r280, 0; st.local.v2.f32 [%rd724+56], {%f1642, %f1642}; st.local.u32 [%rd724+64], %r280; mul.ftz.f32 %f1648, %f71, %f71; st.local.f32 [%rd724+32], %f1648; st.local.f32 [%rd724+40], %f2980; st.local.f32 [%rd726], %f186; st.local.f32 [%rd726+4], %f187; st.local.f32 [%rd726+8], %f188; ld.local.u64 %rd390, [%rd724+48]; st.f32 [%rd390+12], %f282; st.f32 [%rd390+16], %f283; st.f32 [%rd390+20], %f284; ld.local.u64 %rd391, [%rd724+48]; st.u32 [%rd391+36], %r280; ld.local.f32 %f1649, [%rd724+32]; mov.f32 %f1650, 0f38D1B717; max.ftz.f32 %f1651, %f1649, %f1650; min.ftz.f32 %f1652, %f1651, %f1640; st.local.f32 [%rd724+32], %f1652; st.local.f32 [%rd724+36], %f1652; ld.local.f32 %f1653, [%rd724+40]; max.ftz.f32 %f1654, %f1642, %f1653; st.local.f32 [%rd724+40], %f1654; ld.local.u64 %rd392, [%rd724+48]; ld.f32 %f1655, [%rd392]; cvt.ftz.sat.f32.f32 %f1656, %f1655; ld.f32 %f1657, [%rd392+4]; cvt.ftz.sat.f32.f32 %f1658, %f1657; ld.f32 %f1659, [%rd392+8]; cvt.ftz.sat.f32.f32 %f1660, %f1659; st.f32 [%rd392], %f1656; st.f32 [%rd392+4], %f1658; st.f32 [%rd392+8], %f1660; ld.local.u64 %rd393, [%rd724+48]; ld.f32 %f1661, [%rd393+12]; cvt.ftz.sat.f32.f32 %f1662, %f1661; ld.f32 %f1663, [%rd393+16]; cvt.ftz.sat.f32.f32 %f1664, %f1663; ld.f32 %f1665, [%rd393+20]; cvt.ftz.sat.f32.f32 %f1666, %f1665; st.f32 [%rd393+12], %f1662; st.f32 [%rd393+16], %f1664; st.f32 [%rd393+20], %f1666; mov.u32 %r281, 27; st.local.u32 [%rd724+12], %r281; ld.local.f32 %f1667, [%rd724+40]; fma.rn.ftz.f32 %f289, %f1667, %f1667, 0fBF800000; abs.ftz.f32 %f290, %f1640; fma.rn.ftz.f32 %f291, %f290, %f290, %f289; setp.leu.ftz.f32 %p203, %f291, 0f00000000; mov.f32 %f3021, %f1640; @%p203 bra $L__BB148_241; sqrt.approx.ftz.f32 %f1668, %f291; sub.ftz.f32 %f1669, %f1668, %f290; add.ftz.f32 %f1670, %f290, %f1668; div.approx.ftz.f32 %f1671, %f1669, %f1670; fma.rn.ftz.f32 %f1672, %f290, %f1670, 0fBF800000; fma.rn.ftz.f32 %f1673, %f290, %f1669, 0f3F800000; div.approx.ftz.f32 %f1674, %f1672, %f1673; mul.ftz.f32 %f1675, %f1671, 0f3F000000; mul.ftz.f32 %f1676, %f1671, %f1675; fma.rn.ftz.f32 %f1677, %f1674, %f1674, 0f3F800000; mul.ftz.f32 %f3021, %f1676, %f1677; $L__BB148_241: ld.local.u64 %rd394, [%rd73+28]; add.s64 %rd74, %rd394, 12; ld.f32 %f294, [%rd394+12]; ld.f32 %f295, [%rd394+16]; ld.f32 %f296, [%rd394+20]; sub.ftz.f32 %f297, %f1640, %f3021; ld.local.f32 %f1679, [%rd73]; ld.local.f32 %f1680, [%rd4+24]; ld.local.f32 %f1681, [%rd73+4]; ld.local.f32 %f1682, [%rd4+28]; mul.ftz.f32 %f1683, %f1682, %f1681; fma.rn.ftz.f32 %f1684, %f1680, %f1679, %f1683; ld.local.f32 %f1685, [%rd73+8]; ld.local.f32 %f1686, [%rd4+32]; fma.rn.ftz.f32 %f1687, %f1686, %f1685, %f1684; abs.ftz.f32 %f298, %f1687; fma.rn.ftz.f32 %f299, %f298, %f298, %f289; setp.leu.ftz.f32 %p204, %f299, 0f00000000; mov.f32 %f3022, %f297; @%p204 bra $L__BB148_243; sqrt.approx.ftz.f32 %f1688, %f299; sub.ftz.f32 %f1689, %f1688, %f298; add.ftz.f32 %f1690, %f298, %f1688; div.approx.ftz.f32 %f1691, %f1689, %f1690; fma.rn.ftz.f32 %f1692, %f298, %f1690, 0fBF800000; fma.rn.ftz.f32 %f1693, %f298, %f1689, 0f3F800000; div.approx.ftz.f32 %f1694, %f1692, %f1693; mul.ftz.f32 %f1695, %f1691, 0f3F000000; mul.ftz.f32 %f1696, %f1691, %f1695; fma.rn.ftz.f32 %f1697, %f1694, %f1694, 0f3F800000; mul.ftz.f32 %f1698, %f1696, %f1697; sub.ftz.f32 %f3022, %f1698, %f3021; $L__BB148_243: rcp.approx.ftz.f32 %f1699, %f297; mov.f32 %f1700, 0f3F800000; mul.ftz.f32 %f1701, %f1699, %f3022; sub.ftz.f32 %f1702, %f1700, %f1701; fma.rn.ftz.f32 %f1703, %f294, %f1702, %f1701; fma.rn.ftz.f32 %f1704, %f295, %f1702, %f1701; fma.rn.ftz.f32 %f1705, %f296, %f1702, %f1701; st.f32 [%rd74+12], %f1703; st.f32 [%rd74+16], %f1704; st.f32 [%rd74+20], %f1705; ld.local.u32 %r282, [%rd73+-8]; setp.ne.s32 %p205, %r282, 12; @%p205 bra $L__BB148_245; ld.local.u64 %rd395, [%rd73+28]; ld.f32 %f1706, [%rd395+36]; mul.ftz.f32 %f1707, %f1706, 0f3E800000; ld.f32 %f1708, [%rd395+24]; mul.ftz.f32 %f1709, %f1708, %f1707; ld.f32 %f1710, [%rd395+28]; mul.ftz.f32 %f1711, %f1707, %f1710; ld.f32 %f1712, [%rd395+32]; mul.ftz.f32 %f1713, %f1707, %f1712; st.f32 [%rd395+24], %f1709; st.f32 [%rd395+28], %f1711; st.f32 [%rd395+32], %f1713; $L__BB148_245: ld.local.u64 %rd396, [%rd73+28]; ld.f32 %f1714, [%rd396+28]; ld.f32 %f1715, [%rd396+24]; add.ftz.f32 %f1716, %f1715, %f1714; ld.f32 %f1717, [%rd396+32]; add.ftz.f32 %f1718, %f1717, %f1716; mul.ftz.f32 %f1719, %f1718, 0f3EAAAAAB; ld.local.f32 %f1720, [%rd73+-4]; mul.ftz.f32 %f1721, %f1720, %f1719; st.local.f32 [%rd73+-4], %f1721; ld.local.u32 %r283, [%rd4+40]; or.b32 %r284, %r283, 1036; st.local.u32 [%rd4+40], %r284; bra.uni $L__BB148_272; $L__BB148_255: ld.local.u32 %r293, [%rd4+248]; add.s32 %r294, %r293, -1; st.local.u32 [%rd4+248], %r294; add.s32 %r295, %r52, 1; st.local.u32 [%rd4+252], %r295; mov.u64 %rd733, %rd732; $L__BB148_256: setp.eq.s64 %p212, %rd733, 0; or.pred %p214, %p210, %p212; @%p214 bra $L__BB148_264; add.s64 %rd85, %rd730, 20; st.local.f32 [%rd730+20], %f55; st.local.f32 [%rd730+24], %f56; st.local.f32 [%rd730+28], %f57; mov.u32 %r296, 0; st.local.v2.f32 [%rd730+56], {%f1723, %f1723}; st.local.u32 [%rd730+64], %r296; st.local.u64 [%rd730+48], %rd733; mul.ftz.f32 %f1731, %f71, %f71; st.local.v2.f32 [%rd730+32], {%f1731, %f1731}; st.local.f32 [%rd730+40], %f2980; st.local.f32 [%rd732], %f186; st.local.f32 [%rd732+4], %f187; st.local.f32 [%rd732+8], %f188; ld.local.u64 %rd416, [%rd730+48]; st.f32 [%rd416+12], %f282; st.f32 [%rd416+16], %f283; st.f32 [%rd416+20], %f284; ld.local.u64 %rd417, [%rd730+48]; st.u32 [%rd417+36], %r296; ld.local.u64 %rd418, [%rd730+48]; ld.f32 %f1732, [%rd418+12]; cvt.ftz.sat.f32.f32 %f1733, %f1732; ld.f32 %f1734, [%rd418+16]; cvt.ftz.sat.f32.f32 %f1735, %f1734; ld.f32 %f1736, [%rd418+20]; cvt.ftz.sat.f32.f32 %f1737, %f1736; st.f32 [%rd418+12], %f1733; st.f32 [%rd418+16], %f1735; st.f32 [%rd418+20], %f1737; ld.local.v2.f32 {%f1738, %f1739}, [%rd730+32]; cvt.ftz.sat.f32.f32 %f1742, %f1739; cvt.ftz.sat.f32.f32 %f1743, %f1738; st.local.v2.f32 [%rd730+32], {%f1743, %f1742}; mov.u32 %r297, 11; st.local.u32 [%rd730+12], %r297; ld.local.f32 %f1744, [%rd730+40]; fma.rn.ftz.f32 %f306, %f1744, %f1744, 0fBF800000; mov.f32 %f1729, 0f3F800000; abs.ftz.f32 %f307, %f1729; fma.rn.ftz.f32 %f308, %f307, %f307, %f306; setp.leu.ftz.f32 %p215, %f308, 0f00000000; mov.f32 %f3023, %f1729; @%p215 bra $L__BB148_259; sqrt.approx.ftz.f32 %f1745, %f308; sub.ftz.f32 %f1746, %f1745, %f307; add.ftz.f32 %f1747, %f307, %f1745; div.approx.ftz.f32 %f1748, %f1746, %f1747; fma.rn.ftz.f32 %f1749, %f307, %f1747, 0fBF800000; fma.rn.ftz.f32 %f1750, %f307, %f1746, 0f3F800000; div.approx.ftz.f32 %f1751, %f1749, %f1750; mul.ftz.f32 %f1752, %f1748, 0f3F000000; mul.ftz.f32 %f1753, %f1748, %f1752; fma.rn.ftz.f32 %f1754, %f1751, %f1751, 0f3F800000; mul.ftz.f32 %f3023, %f1753, %f1754; $L__BB148_259: ld.local.u64 %rd419, [%rd85+28]; add.s64 %rd86, %rd419, 12; ld.f32 %f311, [%rd419+12]; ld.f32 %f312, [%rd419+16]; ld.f32 %f313, [%rd419+20]; sub.ftz.f32 %f314, %f1729, %f3023; ld.local.f32 %f1756, [%rd85]; ld.local.f32 %f1757, [%rd4+24]; ld.local.f32 %f1758, [%rd85+4]; ld.local.f32 %f1759, [%rd4+28]; mul.ftz.f32 %f1760, %f1759, %f1758; fma.rn.ftz.f32 %f1761, %f1757, %f1756, %f1760; ld.local.f32 %f1762, [%rd85+8]; ld.local.f32 %f1763, [%rd4+32]; fma.rn.ftz.f32 %f1764, %f1763, %f1762, %f1761; abs.ftz.f32 %f315, %f1764; fma.rn.ftz.f32 %f316, %f315, %f315, %f306; setp.leu.ftz.f32 %p216, %f316, 0f00000000; mov.f32 %f3024, %f314; @%p216 bra $L__BB148_261; sqrt.approx.ftz.f32 %f1765, %f316; sub.ftz.f32 %f1766, %f1765, %f315; add.ftz.f32 %f1767, %f315, %f1765; div.approx.ftz.f32 %f1768, %f1766, %f1767; fma.rn.ftz.f32 %f1769, %f315, %f1767, 0fBF800000; fma.rn.ftz.f32 %f1770, %f315, %f1766, 0f3F800000; div.approx.ftz.f32 %f1771, %f1769, %f1770; mul.ftz.f32 %f1772, %f1768, 0f3F000000; mul.ftz.f32 %f1773, %f1768, %f1772; fma.rn.ftz.f32 %f1774, %f1771, %f1771, 0f3F800000; mul.ftz.f32 %f1775, %f1773, %f1774; sub.ftz.f32 %f3024, %f1775, %f3023; $L__BB148_261: rcp.approx.ftz.f32 %f1776, %f314; mov.f32 %f1777, 0f3F800000; mul.ftz.f32 %f1778, %f1776, %f3024; sub.ftz.f32 %f1779, %f1777, %f1778; fma.rn.ftz.f32 %f1780, %f311, %f1779, %f1778; fma.rn.ftz.f32 %f1781, %f312, %f1779, %f1778; fma.rn.ftz.f32 %f1782, %f313, %f1779, %f1778; st.f32 [%rd86+12], %f1780; st.f32 [%rd86+16], %f1781; st.f32 [%rd86+20], %f1782; ld.local.u32 %r298, [%rd85+-8]; setp.ne.s32 %p217, %r298, 12; @%p217 bra $L__BB148_263; ld.local.u64 %rd420, [%rd85+28]; ld.f32 %f1783, [%rd420+36]; mul.ftz.f32 %f1784, %f1783, 0f3E800000; ld.f32 %f1785, [%rd420+24]; mul.ftz.f32 %f1786, %f1785, %f1784; ld.f32 %f1787, [%rd420+28]; mul.ftz.f32 %f1788, %f1784, %f1787; ld.f32 %f1789, [%rd420+32]; mul.ftz.f32 %f1790, %f1784, %f1789; st.f32 [%rd420+24], %f1786; st.f32 [%rd420+28], %f1788; st.f32 [%rd420+32], %f1790; $L__BB148_263: ld.local.u64 %rd421, [%rd85+28]; ld.f32 %f1791, [%rd421+28]; ld.f32 %f1792, [%rd421+24]; add.ftz.f32 %f1793, %f1792, %f1791; ld.f32 %f1794, [%rd421+32]; add.ftz.f32 %f1795, %f1794, %f1793; mul.ftz.f32 %f1796, %f1795, 0f3EAAAAAB; ld.local.f32 %f1797, [%rd85+-4]; mul.ftz.f32 %f1798, %f1797, %f1796; st.local.f32 [%rd85+-4], %f1798; ld.local.u32 %r299, [%rd4+40]; or.b32 %r300, %r299, 12; st.local.u32 [%rd4+40], %r300; $L__BB148_264: setp.ne.s32 %p218, %r45, 0; setp.eq.s32 %p219, %r48, 0; and.pred %p220, %p218, %p219; @%p220 bra $L__BB148_272; mov.f32 %f1799, 0f3F800000; sub.ftz.f32 %f1800, %f1799, %f2981; mov.f32 %f1801, 0f38D1B717; max.ftz.f32 %f1802, %f1801, %f1800; mul.ftz.f32 %f1803, %f186, %f279; mul.ftz.f32 %f1804, %f1803, %f1802; mul.ftz.f32 %f1805, %f187, %f280; mul.ftz.f32 %f1806, %f1805, %f1802; mul.ftz.f32 %f1807, %f188, %f281; mul.ftz.f32 %f1808, %f1807, %f1802; mov.f32 %f1809, 0f00000000; max.ftz.f32 %f319, %f1804, %f1809; max.ftz.f32 %f320, %f1806, %f1809; max.ftz.f32 %f321, %f1808, %f1809; add.ftz.f32 %f1810, %f319, %f320; add.ftz.f32 %f1811, %f321, %f1810; mul.ftz.f32 %f1812, %f1811, 0f3EAAAAAB; abs.ftz.f32 %f322, %f1812; setp.ltu.ftz.f32 %p221, %f322, 0f3727C5AC; mov.u64 %rd736, 0; mov.u64 %rd737, %rd736; @%p221 bra $L__BB148_270; ld.local.u32 %r53, [%rd4+252]; mov.u64 %rd736, 0; setp.eq.s32 %p222, %r53, 0; mov.u64 %rd734, %rd736; mov.u64 %rd735, %rd736; @%p222 bra $L__BB148_268; ld.local.u32 %r301, [%rd4+248]; mul.wide.s32 %rd426, %r301, 80; add.s64 %rd427, %rd1, %rd426; add.s64 %rd734, %rd427, 304; add.s64 %rd428, %rd230, %rd426; add.s64 %rd735, %rd428, 304; mov.b32 %r302, %f321; mov.b32 %r303, %f320; mov.b32 %r304, %f319; mov.u32 %r305, 0; st.local.v4.u32 [%rd427+304], {%r304, %r303, %r302, %r305}; add.s32 %r306, %r301, 1; st.local.u32 [%rd4+248], %r306; add.s32 %r307, %r53, -1; st.local.u32 [%rd4+252], %r307; $L__BB148_268: setp.eq.s64 %p223, %rd735, 0; mov.u64 %rd737, %rd736; @%p223 bra $L__BB148_270; st.local.f32 [%rd734+16], %f322; mov.u64 %rd736, %rd734; mov.u64 %rd737, %rd735; $L__BB148_270: setp.eq.s64 %p224, %rd737, 0; @%p224 bra $L__BB148_272; mov.u64 %rd431, 0; st.local.f32 [%rd736+20], %f55; st.local.f32 [%rd736+24], %f56; st.local.f32 [%rd736+28], %f57; mov.u32 %r308, 0; st.local.v2.f32 [%rd736+56], {%f1809, %f1809}; st.local.u32 [%rd736+64], %r308; sub.ftz.f32 %f1815, %f1799, %f79; sub.ftz.f32 %f1816, %f1799, %f71; mul.ftz.f32 %f1817, %f1816, %f1815; sub.ftz.f32 %f1818, %f1799, %f1817; selp.f32 %f1819, %f1818, %f71, %p192; mul.ftz.f32 %f1820, %f1819, %f1819; st.local.f32 [%rd736+40], %f2980; st.local.u64 [%rd736+48], %rd431; cvt.ftz.sat.f32.f32 %f1821, %f1820; st.local.v2.f32 [%rd736+32], {%f1821, %f1821}; mov.u32 %r309, 23; st.local.u32 [%rd736+12], %r309; ld.local.u32 %r310, [%rd4+40]; or.b32 %r311, %r310, 12; st.local.u32 [%rd4+40], %r311; $L__BB148_272: ld.param.u32 %r638, [_Z21svm_node_closure_bsdfILj511EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; add.s32 %r663, %r638, 5; setp.leu.ftz.f32 %p226, %f76, 0f3727C5AC; or.pred %p228, %p226, %p158; @%p228 bra $L__BB148_517; mov.f32 %f1822, 0f00000000; max.ftz.f32 %f323, %f177, %f1822; max.ftz.f32 %f324, %f178, %f1822; add.ftz.f32 %f1823, %f323, %f324; max.ftz.f32 %f325, %f179, %f1822; add.ftz.f32 %f1824, %f325, %f1823; mul.ftz.f32 %f1825, %f1824, 0f3EAAAAAB; abs.ftz.f32 %f326, %f1825; setp.ltu.ftz.f32 %p229, %f326, 0f3727C5AC; mov.u64 %rd742, 0; mov.u64 %rd740, %rd742; mov.u64 %rd741, %rd742; @%p229 bra $L__BB148_278; ld.local.u32 %r55, [%rd4+252]; mov.u64 %rd740, 0; setp.eq.s32 %p230, %r55, 0; mov.u64 %rd738, %rd740; mov.u64 %rd739, %rd740; @%p230 bra $L__BB148_276; ld.local.u32 %r312, [%rd4+248]; mul.wide.s32 %rd436, %r312, 80; add.s64 %rd437, %rd1, %rd436; add.s64 %rd738, %rd437, 304; add.s64 %rd438, %rd230, %rd436; add.s64 %rd739, %rd438, 304; mov.b32 %r313, %f325; mov.b32 %r314, %f324; mov.b32 %r315, %f323; mov.u32 %r316, 0; st.local.v4.u32 [%rd437+304], {%r315, %r314, %r313, %r316}; add.s32 %r317, %r312, 1; st.local.u32 [%rd4+248], %r317; add.s32 %r318, %r55, -1; st.local.u32 [%rd4+252], %r318; $L__BB148_276: setp.eq.s64 %p231, %rd739, 0; mov.u64 %rd741, %rd740; @%p231 bra $L__BB148_278; st.local.f32 [%rd738+16], %f326; mov.u64 %rd740, %rd738; mov.u64 %rd741, %rd739; $L__BB148_278: setp.eq.s64 %p232, %rd741, 0; mov.u64 %rd743, %rd742; @%p232 bra $L__BB148_282; ld.local.u32 %r56, [%rd4+252]; setp.lt.s32 %p233, %r56, 1; @%p233 bra $L__BB148_281; add.s32 %r319, %r56, -1; st.local.u32 [%rd4+252], %r319; ld.local.s32 %rd443, [%rd4+248]; cvt.s64.s32 %rd444, %r319; add.s64 %rd445, %rd443, %rd444; mul.lo.s64 %rd446, %rd445, 80; add.s64 %rd447, %rd1, %rd446; add.s64 %rd742, %rd447, 304; add.s64 %rd448, %rd230, %rd446; add.s64 %rd743, %rd448, 304; bra.uni $L__BB148_282; $L__BB148_281: ld.local.u32 %r320, [%rd4+248]; add.s32 %r321, %r320, -1; st.local.u32 [%rd4+248], %r321; add.s32 %r322, %r56, 1; st.local.u32 [%rd4+252], %r322; mov.u64 %rd743, %rd742; $L__BB148_282: setp.eq.s64 %p234, %rd743, 0; or.pred %p236, %p232, %p234; @%p236 bra $L__BB148_517; mov.f32 %f3025, 0f3F800000; add.s64 %rd103, %rd740, 20; st.local.f32 [%rd740+20], %f2990; st.local.f32 [%rd740+24], %f2991; st.local.f32 [%rd740+28], %f2992; mov.u32 %r323, 0; st.local.v2.f32 [%rd740+56], {%f1822, %f1822}; st.local.u32 [%rd740+64], %r323; mov.u32 %r324, 1069547520; st.local.u32 [%rd740+40], %r324; st.local.u64 [%rd740+48], %rd743; mul.ftz.f32 %f1828, %f77, %f77; st.local.v2.f32 [%rd740+32], {%f1828, %f1828}; st.local.u32 [%rd742], %r323; st.local.u32 [%rd742+4], %r323; st.local.u32 [%rd742+8], %r323; ld.local.u64 %rd451, [%rd740+48]; mov.u32 %r325, 1025758986; st.u32 [%rd451+12], %r325; st.u32 [%rd451+16], %r325; st.u32 [%rd451+20], %r325; ld.local.u64 %rd452, [%rd740+48]; st.f32 [%rd452+36], %f76; ld.local.u64 %rd453, [%rd740+48]; ld.f32 %f1829, [%rd453+12]; cvt.ftz.sat.f32.f32 %f1830, %f1829; ld.f32 %f1831, [%rd453+16]; cvt.ftz.sat.f32.f32 %f1832, %f1831; ld.f32 %f1833, [%rd453+20]; cvt.ftz.sat.f32.f32 %f1834, %f1833; st.f32 [%rd453+12], %f1830; st.f32 [%rd453+16], %f1832; st.f32 [%rd453+20], %f1834; ld.local.f32 %f1835, [%rd740+32]; cvt.ftz.sat.f32.f32 %f1836, %f1835; st.local.v2.f32 [%rd740+32], {%f1836, %f1836}; mov.u32 %r326, 12; st.local.u32 [%rd740+12], %r326; ld.local.f32 %f1837, [%rd740+40]; fma.rn.ftz.f32 %f327, %f1837, %f1837, 0fBF800000; abs.ftz.f32 %f328, %f3025; fma.rn.ftz.f32 %f329, %f328, %f328, %f327; setp.leu.ftz.f32 %p237, %f329, 0f00000000; @%p237 bra $L__BB148_285; sqrt.approx.ftz.f32 %f1838, %f329; sub.ftz.f32 %f1839, %f1838, %f328; add.ftz.f32 %f1840, %f328, %f1838; div.approx.ftz.f32 %f1841, %f1839, %f1840; fma.rn.ftz.f32 %f1842, %f328, %f1840, 0fBF800000; fma.rn.ftz.f32 %f1843, %f328, %f1839, 0f3F800000; div.approx.ftz.f32 %f1844, %f1842, %f1843; mul.ftz.f32 %f1845, %f1841, 0f3F000000; mul.ftz.f32 %f1846, %f1841, %f1845; fma.rn.ftz.f32 %f1847, %f1844, %f1844, 0f3F800000; mul.ftz.f32 %f3025, %f1846, %f1847; $L__BB148_285: mov.f32 %f2960, 0f3F800000; ld.local.u64 %rd454, [%rd103+28]; add.s64 %rd104, %rd454, 12; ld.f32 %f332, [%rd454+12]; ld.f32 %f333, [%rd454+16]; ld.f32 %f334, [%rd454+20]; sub.ftz.f32 %f335, %f2960, %f3025; ld.local.f32 %f1849, [%rd103]; ld.local.f32 %f1850, [%rd4+24]; ld.local.f32 %f1851, [%rd103+4]; ld.local.f32 %f1852, [%rd4+28]; mul.ftz.f32 %f1853, %f1852, %f1851; fma.rn.ftz.f32 %f1854, %f1850, %f1849, %f1853; ld.local.f32 %f1855, [%rd103+8]; ld.local.f32 %f1856, [%rd4+32]; fma.rn.ftz.f32 %f1857, %f1856, %f1855, %f1854; abs.ftz.f32 %f336, %f1857; fma.rn.ftz.f32 %f337, %f336, %f336, %f327; setp.leu.ftz.f32 %p238, %f337, 0f00000000; mov.f32 %f3026, %f335; @%p238 bra $L__BB148_287; sqrt.approx.ftz.f32 %f1858, %f337; sub.ftz.f32 %f1859, %f1858, %f336; add.ftz.f32 %f1860, %f336, %f1858; div.approx.ftz.f32 %f1861, %f1859, %f1860; fma.rn.ftz.f32 %f1862, %f336, %f1860, 0fBF800000; fma.rn.ftz.f32 %f1863, %f336, %f1859, 0f3F800000; div.approx.ftz.f32 %f1864, %f1862, %f1863; mul.ftz.f32 %f1865, %f1861, 0f3F000000; mul.ftz.f32 %f1866, %f1861, %f1865; fma.rn.ftz.f32 %f1867, %f1864, %f1864, 0f3F800000; mul.ftz.f32 %f1868, %f1866, %f1867; sub.ftz.f32 %f3026, %f1868, %f3025; $L__BB148_287: rcp.approx.ftz.f32 %f1869, %f335; mov.f32 %f1870, 0f3F800000; mul.ftz.f32 %f1871, %f1869, %f3026; sub.ftz.f32 %f1872, %f1870, %f1871; fma.rn.ftz.f32 %f1873, %f332, %f1872, %f1871; fma.rn.ftz.f32 %f1874, %f333, %f1872, %f1871; fma.rn.ftz.f32 %f1875, %f334, %f1872, %f1871; st.f32 [%rd104+12], %f1873; st.f32 [%rd104+16], %f1874; st.f32 [%rd104+20], %f1875; ld.local.u32 %r327, [%rd103+-8]; setp.ne.s32 %p239, %r327, 12; @%p239 bra $L__BB148_289; ld.local.u64 %rd455, [%rd103+28]; ld.f32 %f1876, [%rd455+36]; mul.ftz.f32 %f1877, %f1876, 0f3E800000; ld.f32 %f1878, [%rd455+24]; mul.ftz.f32 %f1879, %f1878, %f1877; ld.f32 %f1880, [%rd455+28]; mul.ftz.f32 %f1881, %f1877, %f1880; ld.f32 %f1882, [%rd455+32]; mul.ftz.f32 %f1883, %f1877, %f1882; st.f32 [%rd455+24], %f1879; st.f32 [%rd455+28], %f1881; st.f32 [%rd455+32], %f1883; $L__BB148_289: ld.local.u64 %rd456, [%rd103+28]; ld.f32 %f1884, [%rd456+28]; ld.f32 %f1885, [%rd456+24]; add.ftz.f32 %f1886, %f1885, %f1884; ld.f32 %f1887, [%rd456+32]; add.ftz.f32 %f1888, %f1887, %f1886; mul.ftz.f32 %f1889, %f1888, 0f3EAAAAAB; ld.local.f32 %f1890, [%rd103+-4]; mul.ftz.f32 %f1891, %f1890, %f1889; st.local.f32 [%rd103+-4], %f1891; ld.local.u32 %r328, [%rd4+40]; or.b32 %r329, %r328, 12; st.local.u32 [%rd4+40], %r329; bra.uni $L__BB148_517; } .func _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB149_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB149_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB149_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB149_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB149_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB149_11; $L__BB149_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB149_8; add.s32 %r237, %r237, 2; bra.uni $L__BB149_9; $L__BB149_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB149_15; $L__BB149_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB149_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB149_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB149_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB149_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB149_15; $L__BB149_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB149_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB149_17; $L__BB149_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB149_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB149_159; bra.uni $L__BB149_18; $L__BB149_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB149_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB149_185; $L__BB149_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB149_150; bra.uni $L__BB149_19; $L__BB149_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB149_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB149_153; bra.uni $L__BB149_152; $L__BB149_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB149_154; $L__BB149_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB149_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB149_100; bra.uni $L__BB149_21; $L__BB149_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB149_111; bra.uni $L__BB149_101; $L__BB149_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB149_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB149_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB149_120; $L__BB149_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB149_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB149_164; bra.uni $L__BB149_163; $L__BB149_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB149_165; $L__BB149_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB149_136; bra.uni $L__BB149_126; $L__BB149_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB149_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB149_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB149_145; $L__BB149_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB149_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB149_158; bra.uni $L__BB149_155; $L__BB149_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB149_192; $L__BB149_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB149_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB149_192; $L__BB149_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB149_47; bra.uni $L__BB149_22; $L__BB149_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB149_86; bra.uni $L__BB149_48; $L__BB149_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB149_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB149_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB149_95; $L__BB149_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB149_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB149_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB149_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB149_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB149_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB149_184; bra.uni $L__BB149_170; $L__BB149_184: { // callseq 912, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 912 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB149_185; $L__BB149_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB149_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB149_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 906, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 906 bra.uni $L__BB149_145; $L__BB149_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB149_192; $L__BB149_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB149_183; bra.uni $L__BB149_171; $L__BB149_183: { // callseq 911, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 911 mov.f32 %f731, 0f3F800000; bra.uni $L__BB149_185; $L__BB149_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB149_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB149_104; mov.u64 %rd173, 0; { // callseq 905, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 905 bra.uni $L__BB149_120; $L__BB149_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB149_33; bra.uni $L__BB149_23; $L__BB149_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB149_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB149_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB149_42; $L__BB149_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB149_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB149_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB149_133; $L__BB149_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB149_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB149_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB149_108; $L__BB149_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB149_182; bra.uni $L__BB149_172; $L__BB149_182: { // callseq 910, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 910 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB149_185; $L__BB149_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB149_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB149_51; mov.u64 %rd129, 0; { // callseq 904, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 904 bra.uni $L__BB149_95; $L__BB149_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB149_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB149_26; mov.u64 %rd90, 0; { // callseq 903, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 903 bra.uni $L__BB149_42; $L__BB149_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB149_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB149_145; $L__BB149_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB149_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB149_120; $L__BB149_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB149_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB149_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB149_83; $L__BB149_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB149_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB149_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB149_30; $L__BB149_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB149_181; bra.uni $L__BB149_173; $L__BB149_181: { // callseq 909, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 909 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB149_185; $L__BB149_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB149_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB149_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB149_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB149_145; $L__BB149_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB149_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB149_145; $L__BB149_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB149_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB149_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB149_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB149_120; $L__BB149_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB149_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB149_120; $L__BB149_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB149_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB149_95; $L__BB149_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB149_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB149_42; $L__BB149_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB149_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB149_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB149_149; bra.uni $L__BB149_146; $L__BB149_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB149_192; $L__BB149_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB149_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB149_192; $L__BB149_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB149_192; $L__BB149_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB149_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB149_175; bra.uni $L__BB149_178; $L__BB149_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB149_177; bra.uni $L__BB149_176; $L__BB149_177: { // callseq 907, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 907 bra.uni $L__BB149_185; $L__BB149_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB149_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB149_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB149_124; bra.uni $L__BB149_121; $L__BB149_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB149_192; $L__BB149_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB149_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB149_192; $L__BB149_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB149_192; $L__BB149_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB149_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB149_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB149_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB149_95; $L__BB149_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB149_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB149_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB149_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB149_42; $L__BB149_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB149_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB149_95; $L__BB149_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB149_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB149_42; $L__BB149_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB149_82; bra.uni $L__BB149_55; $L__BB149_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB149_83; $L__BB149_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB149_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB149_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB149_46; bra.uni $L__BB149_43; $L__BB149_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB149_192; $L__BB149_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB149_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB149_192; $L__BB149_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB149_192; $L__BB149_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB149_185; $L__BB149_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB149_57; bra.uni $L__BB149_56; $L__BB149_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB149_58; $L__BB149_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB149_180; bra.uni $L__BB149_179; $L__BB149_180: { // callseq 908, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 908 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB149_185; $L__BB149_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB149_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB149_60; bra.uni $L__BB149_59; $L__BB149_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB149_61; $L__BB149_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB149_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB149_63; bra.uni $L__BB149_62; $L__BB149_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB149_64; $L__BB149_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB149_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB149_66; bra.uni $L__BB149_65; $L__BB149_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB149_67; $L__BB149_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB149_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB149_69; bra.uni $L__BB149_68; $L__BB149_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB149_70; $L__BB149_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB149_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB149_72; bra.uni $L__BB149_71; $L__BB149_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB149_73; $L__BB149_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB149_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB149_75; bra.uni $L__BB149_74; $L__BB149_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB149_76; $L__BB149_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB149_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB149_78; bra.uni $L__BB149_77; $L__BB149_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB149_79; $L__BB149_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB149_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB149_81; bra.uni $L__BB149_80; $L__BB149_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB149_83; $L__BB149_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB149_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB149_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB149_99; bra.uni $L__BB149_96; $L__BB149_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB149_192; $L__BB149_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB149_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB149_192; $L__BB149_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB149_192; $L__BB149_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB149_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB149_191; bra.uni $L__BB149_186; $L__BB149_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB149_192; $L__BB149_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB149_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB149_192; $L__BB149_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB149_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB149_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB149_192: ret; } .func _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB150_2; bra.uni $L__BB150_1; $L__BB150_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB150_3; $L__BB150_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB150_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB150_5; bra.uni $L__BB150_4; $L__BB150_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB150_9; bra.uni $L__BB150_6; $L__BB150_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB150_12; bra.uni $L__BB150_10; $L__BB150_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB150_13; $L__BB150_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB150_16; $L__BB150_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB150_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB150_13; $L__BB150_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB150_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB150_15; bra.uni $L__BB150_14; $L__BB150_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB150_16; $L__BB150_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB150_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB150_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB150_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB150_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj511EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB151_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB151_5; bra.uni $L__BB151_2; $L__BB151_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB151_8; bra.uni $L__BB151_6; $L__BB151_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB151_9; $L__BB151_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB151_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB151_9; $L__BB151_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB151_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB151_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB151_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB151_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB151_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB151_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB151_20; $L__BB151_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB151_17; add.s32 %r172, %r172, 2; bra.uni $L__BB151_18; $L__BB151_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB151_24; $L__BB151_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB151_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB151_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB151_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB151_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB151_24; $L__BB151_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB151_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB151_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB151_37; bra.uni $L__BB151_26; $L__BB151_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB151_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB151_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB151_45; $L__BB151_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB151_45; $L__BB151_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB151_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB151_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB151_30; mov.u64 %rd108, 0; { // callseq 913, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 913 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB151_45; $L__BB151_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB151_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB151_9; $L__BB151_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB151_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB151_45; $L__BB151_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB151_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB151_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB151_34; $L__BB151_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB151_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB151_45; $L__BB151_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB151_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB151_45; $L__BB151_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB151_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB151_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB151_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB151_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB151_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB151_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB151_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB151_56; $L__BB151_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB151_53; add.s32 %r187, %r187, 2; bra.uni $L__BB151_54; $L__BB151_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB151_60; $L__BB151_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB151_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB151_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB151_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB151_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB151_60; $L__BB151_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB151_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB151_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB151_72; bra.uni $L__BB151_62; $L__BB151_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB151_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB151_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB151_79; $L__BB151_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB151_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB151_65; mov.u64 %rd160, 0; { // callseq 914, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 914 bra.uni $L__BB151_79; $L__BB151_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB151_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB151_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB151_69; $L__BB151_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB151_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB151_79; $L__BB151_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB151_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB151_79; $L__BB151_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB151_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB151_79; $L__BB151_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB151_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB151_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB151_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB151_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB151_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB151_84; bra.uni $L__BB151_83; $L__BB151_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB151_85; $L__BB151_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB151_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj511EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB152_2; bra.uni $L__BB152_1; $L__BB152_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB152_3; $L__BB152_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB152_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB152_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB152_8; bra.uni $L__BB152_5; $L__BB152_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB152_11; bra.uni $L__BB152_9; $L__BB152_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB152_12; $L__BB152_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB152_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB152_12; $L__BB152_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB152_12: @%p8 bra $L__BB152_14; bra.uni $L__BB152_13; $L__BB152_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB152_15; $L__BB152_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB152_15: @%p8 bra $L__BB152_17; bra.uni $L__BB152_16; $L__BB152_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB152_18; $L__BB152_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB152_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB152_19; bra.uni $L__BB152_20; $L__BB152_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB152_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB152_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB152_22: @%p7 bra $L__BB152_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB152_25; bra.uni $L__BB152_24; $L__BB152_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB152_26; $L__BB152_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB152_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB152_27; bra.uni $L__BB152_34; $L__BB152_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB152_30; bra.uni $L__BB152_28; $L__BB152_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB152_32; bra.uni $L__BB152_31; $L__BB152_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB152_33; $L__BB152_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB152_29; bra.uni $L__BB152_34; $L__BB152_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB152_33; $L__BB152_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB152_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB152_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB152_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB152_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB152_12; } .func _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<55>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.f32 %f20, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u64 %rd6, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; ld.param.u32 %r5, [_Z19svm_node_light_pathILj511EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6]; mov.f32 %f17, 0f00000000; setp.gt.s32 %p1, %r3, 6; @%p1 bra $L__BB153_13; bra.uni $L__BB153_1; $L__BB153_13: setp.gt.s32 %p2, %r3, 10; @%p2 bra $L__BB153_23; setp.gt.s32 %p8, %r3, 8; @%p8 bra $L__BB153_18; setp.eq.s32 %p11, %r3, 7; @%p11 bra $L__BB153_40; setp.eq.s32 %p12, %r3, 8; mov.f32 %f20, %f17; @%p12 bra $L__BB153_17; bra.uni $L__BB153_45; $L__BB153_17: and.b32 %r12, %r2, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB153_45; $L__BB153_1: setp.gt.s32 %p13, %r3, 2; @%p13 bra $L__BB153_6; bra.uni $L__BB153_2; $L__BB153_6: setp.gt.s32 %p14, %r3, 4; @%p14 bra $L__BB153_10; setp.eq.s32 %p17, %r3, 3; @%p17 bra $L__BB153_42; setp.eq.s32 %p18, %r3, 4; mov.f32 %f20, %f17; @%p18 bra $L__BB153_9; bra.uni $L__BB153_45; $L__BB153_9: and.b32 %r16, %r5, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB153_45; $L__BB153_23: setp.gt.s32 %p3, %r3, 12; @%p3 bra $L__BB153_29; setp.eq.s32 %p6, %r3, 11; @%p6 bra $L__BB153_37; setp.eq.s32 %p7, %r3, 12; mov.f32 %f20, %f17; @%p7 bra $L__BB153_26; bra.uni $L__BB153_45; $L__BB153_26: and.b32 %r8, %r5, 768; mov.u64 %rd24, kernel_params; add.s64 %rd52, %rd24, 3400; setp.ne.s32 %p24, %r8, 0; @%p24 bra $L__BB153_28; add.s64 %rd52, %rd24, 2488; $L__BB153_28: ld.const.u64 %rd27, [%rd52]; cvta.to.global.u64 %rd28, %rd27; mul.wide.s32 %rd29, %r1, 2; add.s64 %rd30, %rd28, %rd29; ld.global.u16 %rs3, [%rd30]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB153_45; $L__BB153_2: setp.eq.s32 %p19, %r3, 0; @%p19 bra $L__BB153_44; setp.eq.s32 %p20, %r3, 1; @%p20 bra $L__BB153_43; setp.eq.s32 %p21, %r3, 2; mov.f32 %f20, %f17; @%p21 bra $L__BB153_5; bra.uni $L__BB153_45; $L__BB153_5: and.b32 %r18, %r5, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB153_45; $L__BB153_18: setp.eq.s32 %p9, %r3, 9; @%p9 bra $L__BB153_45; setp.eq.s32 %p10, %r3, 10; mov.f32 %f20, %f17; @%p10 bra $L__BB153_20; bra.uni $L__BB153_45; $L__BB153_20: and.b32 %r10, %r5, 768; mov.u64 %rd40, kernel_params; add.s64 %rd54, %rd40, 3376; setp.ne.s32 %p26, %r10, 0; @%p26 bra $L__BB153_22; add.s64 %rd54, %rd40, 2464; $L__BB153_22: ld.const.u64 %rd43, [%rd54]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r1, 2; add.s64 %rd46, %rd44, %rd45; ld.global.u16 %rs5, [%rd46]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r11, %r5, 525056; setp.eq.s32 %p27, %r11, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB153_45; $L__BB153_10: setp.eq.s32 %p15, %r3, 5; @%p15 bra $L__BB153_41; setp.eq.s32 %p16, %r3, 6; mov.f32 %f20, %f17; @%p16 bra $L__BB153_12; bra.uni $L__BB153_45; $L__BB153_12: and.b32 %r14, %r5, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB153_45; $L__BB153_29: setp.eq.s32 %p4, %r3, 13; @%p4 bra $L__BB153_34; setp.ne.s32 %p5, %r3, 14; mov.f32 %f20, %f17; @%p5 bra $L__BB153_45; and.b32 %r6, %r5, 768; mov.u64 %rd8, kernel_params; add.s64 %rd50, %rd8, 3408; setp.ne.s32 %p22, %r6, 0; @%p22 bra $L__BB153_33; add.s64 %rd50, %rd8, 2496; $L__BB153_33: ld.const.u64 %rd11, [%rd50]; cvta.to.global.u64 %rd12, %rd11; mul.wide.s32 %rd13, %r1, 2; add.s64 %rd14, %rd12, %rd13; ld.global.u16 %rs1, [%rd14]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB153_45; $L__BB153_40: and.b32 %r13, %r5, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB153_45; $L__BB153_42: and.b32 %r17, %r5, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB153_45; $L__BB153_37: and.b32 %r9, %r5, 768; mov.u64 %rd32, kernel_params; add.s64 %rd53, %rd32, 3392; setp.ne.s32 %p25, %r9, 0; @%p25 bra $L__BB153_39; add.s64 %rd53, %rd32, 2480; $L__BB153_39: ld.const.u64 %rd35, [%rd53]; cvta.to.global.u64 %rd36, %rd35; mul.wide.s32 %rd37, %r1, 2; add.s64 %rd38, %rd36, %rd37; ld.global.u16 %rs4, [%rd38]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB153_45; $L__BB153_44: and.b32 %r20, %r5, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB153_45; $L__BB153_43: and.b32 %r19, %r5, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB153_45; $L__BB153_41: and.b32 %r15, %r5, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB153_45; $L__BB153_34: and.b32 %r7, %r5, 768; mov.u64 %rd16, kernel_params; add.s64 %rd51, %rd16, 3384; setp.ne.s32 %p23, %r7, 0; @%p23 bra $L__BB153_36; add.s64 %rd51, %rd16, 2472; $L__BB153_36: ld.const.u64 %rd19, [%rd51]; cvta.to.global.u64 %rd20, %rd19; mul.wide.s32 %rd21, %r1, 2; add.s64 %rd22, %rd20, %rd21; ld.global.u16 %rs2, [%rd22]; cvt.rn.f32.u16 %f20, %rs2; $L__BB153_45: cvta.to.local.u64 %rd47, %rd6; mul.wide.u32 %rd48, %r4, 4; add.s64 %rd49, %rd47, %rd48; st.local.f32 [%rd49], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB154_2; bra.uni $L__BB154_1; $L__BB154_2: mov.b32 %f2764, %r395; bra.uni $L__BB154_3; $L__BB154_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB154_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB154_5; bra.uni $L__BB154_4; $L__BB154_5: mov.b32 %f9, %r396; bra.uni $L__BB154_6; $L__BB154_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB154_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB154_8; bra.uni $L__BB154_7; $L__BB154_8: mov.b32 %f2766, %r397; bra.uni $L__BB154_9; $L__BB154_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB154_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB154_11; bra.uni $L__BB154_10; $L__BB154_11: mov.b32 %f15, %r398; bra.uni $L__BB154_12; $L__BB154_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB154_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB154_14; bra.uni $L__BB154_13; $L__BB154_14: mov.b32 %f2768, %r399; bra.uni $L__BB154_15; $L__BB154_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB154_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB154_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB154_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB154_18; bra.uni $L__BB154_385; $L__BB154_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB154_259; bra.uni $L__BB154_19; $L__BB154_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB154_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB154_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB154_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB154_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB154_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB154_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB154_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB154_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB154_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB154_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB154_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB154_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB154_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB154_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB154_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB154_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB154_362; $L__BB154_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB154_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB154_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB154_40; bra.uni $L__BB154_29; $L__BB154_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB154_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB154_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB154_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB154_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB154_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB154_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB154_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB154_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB154_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB154_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB154_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB154_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB154_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB154_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB154_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB154_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB154_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB154_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB154_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB154_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB154_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB154_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB154_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB154_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB154_136; $L__BB154_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB154_371; bra.uni $L__BB154_366; $L__BB154_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB154_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB154_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB154_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB154_383; $L__BB154_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB154_149; bra.uni $L__BB154_139; $L__BB154_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB154_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB154_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB154_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB154_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB154_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB154_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB154_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB154_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB154_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB154_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB154_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB154_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB154_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB154_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB154_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB154_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB154_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB154_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB154_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB154_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB154_250; $L__BB154_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB154_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB154_253; bra.uni $L__BB154_362; $L__BB154_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB154_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB154_304; @%p171 bra $L__BB154_303; bra.uni $L__BB154_256; $L__BB154_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB154_305; $L__BB154_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB154_305; $L__BB154_256: @%p172 bra $L__BB154_302; bra.uni $L__BB154_257; $L__BB154_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB154_305; $L__BB154_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB154_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB154_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB154_308; bra.uni $L__BB154_306; $L__BB154_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB154_309; $L__BB154_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB154_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB154_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB154_316; @%p171 bra $L__BB154_315; bra.uni $L__BB154_311; $L__BB154_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB154_317; $L__BB154_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB154_317; $L__BB154_311: @%p172 bra $L__BB154_314; bra.uni $L__BB154_312; $L__BB154_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB154_317; $L__BB154_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB154_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB154_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB154_320; bra.uni $L__BB154_318; $L__BB154_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB154_321; $L__BB154_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB154_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB154_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB154_328; @%p171 bra $L__BB154_327; bra.uni $L__BB154_323; $L__BB154_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB154_329; $L__BB154_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB154_329; $L__BB154_323: @%p172 bra $L__BB154_326; bra.uni $L__BB154_324; $L__BB154_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB154_329; $L__BB154_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB154_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB154_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB154_332; bra.uni $L__BB154_330; $L__BB154_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB154_333; $L__BB154_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB154_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB154_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB154_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_362; $L__BB154_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB154_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB154_31; bra.uni $L__BB154_136; $L__BB154_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB154_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB154_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB154_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB154_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB154_107; @%p49 bra $L__BB154_106; bra.uni $L__BB154_37; $L__BB154_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB154_108; $L__BB154_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB154_108; $L__BB154_37: @%p50 bra $L__BB154_105; bra.uni $L__BB154_38; $L__BB154_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB154_108; $L__BB154_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB154_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB154_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB154_111; bra.uni $L__BB154_109; $L__BB154_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB154_112; $L__BB154_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB154_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB154_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB154_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB154_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB154_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB154_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_136; $L__BB154_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB154_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB154_368; bra.uni $L__BB154_383; $L__BB154_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB154_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB154_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB154_383; $L__BB154_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB154_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB154_141; bra.uni $L__BB154_250; $L__BB154_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB154_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB154_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB154_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB154_206; @%p101 bra $L__BB154_205; bra.uni $L__BB154_146; $L__BB154_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB154_207; $L__BB154_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB154_207; $L__BB154_146: @%p102 bra $L__BB154_204; bra.uni $L__BB154_147; $L__BB154_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB154_207; $L__BB154_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB154_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB154_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB154_210; bra.uni $L__BB154_208; $L__BB154_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB154_211; $L__BB154_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB154_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB154_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB154_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB154_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB154_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_250; $L__BB154_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB154_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB154_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB154_295; @%p151 bra $L__BB154_294; bra.uni $L__BB154_290; $L__BB154_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB154_296; $L__BB154_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB154_296; $L__BB154_290: @%p152 bra $L__BB154_293; bra.uni $L__BB154_291; $L__BB154_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB154_296; $L__BB154_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB154_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB154_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB154_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB154_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB154_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB154_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB154_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_362; $L__BB154_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB154_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB154_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB154_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB154_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB154_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB154_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB154_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB154_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB154_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB154_279; bra.uni $L__BB154_362; $L__BB154_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB154_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB154_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB154_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB154_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB154_96; @%p39 bra $L__BB154_95; bra.uni $L__BB154_91; $L__BB154_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB154_97; $L__BB154_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB154_97; $L__BB154_91: @%p40 bra $L__BB154_94; bra.uni $L__BB154_92; $L__BB154_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB154_97; $L__BB154_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB154_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB154_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB154_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB154_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB154_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB154_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB154_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB154_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB154_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_136; $L__BB154_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB154_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB154_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB154_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB154_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB154_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB154_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB154_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB154_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB154_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB154_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB154_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB154_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB154_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB154_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB154_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB154_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB154_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB154_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB154_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB154_76; bra.uni $L__BB154_136; $L__BB154_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB154_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB154_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB154_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB154_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB154_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_383; $L__BB154_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB154_383; $L__BB154_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB154_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB154_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB154_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB154_196; @%p92 bra $L__BB154_195; bra.uni $L__BB154_191; $L__BB154_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB154_197; $L__BB154_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB154_197; $L__BB154_191: @%p93 bra $L__BB154_194; bra.uni $L__BB154_192; $L__BB154_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB154_197; $L__BB154_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB154_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB154_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB154_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB154_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB154_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB154_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB154_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB154_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB154_250; $L__BB154_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB154_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB154_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB154_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB154_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB154_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB154_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB154_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB154_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB154_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB154_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB154_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB154_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB154_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB154_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB154_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB154_179; bra.uni $L__BB154_250; $L__BB154_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB154_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB154_337; @%p188 bra $L__BB154_336; bra.uni $L__BB154_23; $L__BB154_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB154_338; $L__BB154_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB154_338; $L__BB154_23: @%p189 bra $L__BB154_335; bra.uni $L__BB154_24; $L__BB154_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB154_338; $L__BB154_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB154_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB154_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB154_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB154_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB154_347; @%p188 bra $L__BB154_346; bra.uni $L__BB154_342; $L__BB154_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB154_348; $L__BB154_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB154_348; $L__BB154_342: @%p189 bra $L__BB154_345; bra.uni $L__BB154_343; $L__BB154_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB154_348; $L__BB154_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB154_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB154_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB154_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB154_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB154_357; @%p188 bra $L__BB154_356; bra.uni $L__BB154_352; $L__BB154_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB154_358; $L__BB154_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB154_358; $L__BB154_352: @%p189 bra $L__BB154_355; bra.uni $L__BB154_353; $L__BB154_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB154_358; $L__BB154_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB154_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB154_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB154_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB154_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB154_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB154_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB154_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB154_364: mov.f32 %f3392, %f3387; bra.uni $L__BB154_385; $L__BB154_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB154_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB154_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB154_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB154_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB154_128; @%p59 bra $L__BB154_127; bra.uni $L__BB154_123; $L__BB154_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB154_129; $L__BB154_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB154_129; $L__BB154_123: @%p60 bra $L__BB154_126; bra.uni $L__BB154_124; $L__BB154_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB154_129; $L__BB154_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB154_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB154_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB154_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB154_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB154_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB154_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB154_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB154_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB154_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB154_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB154_385; $L__BB154_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB154_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB154_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB154_385; $L__BB154_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB154_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB154_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB154_224; @%p120 bra $L__BB154_223; bra.uni $L__BB154_219; $L__BB154_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB154_225; $L__BB154_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB154_225; $L__BB154_219: @%p121 bra $L__BB154_222; bra.uni $L__BB154_220; $L__BB154_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB154_225; $L__BB154_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB154_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB154_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB154_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB154_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB154_234; @%p120 bra $L__BB154_233; bra.uni $L__BB154_229; $L__BB154_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB154_235; $L__BB154_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB154_235; $L__BB154_229: @%p121 bra $L__BB154_232; bra.uni $L__BB154_230; $L__BB154_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB154_235; $L__BB154_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB154_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB154_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB154_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB154_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB154_244; @%p120 bra $L__BB154_243; bra.uni $L__BB154_239; $L__BB154_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB154_245; $L__BB154_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB154_245; $L__BB154_239: @%p121 bra $L__BB154_242; bra.uni $L__BB154_240; $L__BB154_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB154_245; $L__BB154_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB154_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB154_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB154_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB154_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB154_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB154_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB154_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB154_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB154_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB154_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB154_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB154_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB154_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB154_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB154_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB154_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB154_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB154_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB154_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj511EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b64 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b64 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .align 4 .b8 _Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3[16] ) { .local .align 16 .b8 __local_depot155[928]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<275>; .reg .b16 %rs<41>; .reg .f32 %f<2598>; .reg .b32 %r<612>; .reg .b64 %rd<527>; mov.u64 %SPL, __local_depot155; ld.param.u32 %r209, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd215, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.u64 %rd216, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u32 %r1, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; cvta.to.local.u64 %rd2, %rd215; add.u64 %rd1, %SPL, 768; and.b32 %r2, %r1, 255; cvta.to.local.u64 %rd218, %rd216; shr.u32 %r210, %r1, 6; cvt.u64.u32 %rd219, %r210; and.b64 %rd220, %rd219, 1020; add.s64 %rd221, %rd218, %rd220; ld.local.f32 %f1, [%rd221]; setp.le.ftz.f32 %p4, %f1, 0f00000000; setp.eq.s32 %p5, %r2, 0; or.pred %p6, %p5, %p4; @%p6 bra $L__BB155_218; ld.local.u32 %r3, [%rd2+76]; setp.eq.s32 %p7, %r3, -1; @%p7 bra $L__BB155_218; ld.const.u32 %r211, [kernel_params+1124]; setp.eq.s32 %p8, %r211, 0; @%p8 bra $L__BB155_217; ld.const.u64 %rd222, [kernel_params+2568]; cvta.to.global.u64 %rd223, %rd222; cvt.s64.s32 %rd3, %r209; mul.wide.s32 %rd224, %r209, 4; add.s64 %rd225, %rd223, %rd224; ld.global.f32 %f661, [%rd225]; setp.lt.ftz.f32 %p9, %f661, 0f41000000; @%p9 bra $L__BB155_216; bra.uni $L__BB155_4; $L__BB155_216: ld.local.f32 %f2594, [%rd2+20]; ld.local.f32 %f2593, [%rd2+16]; ld.local.f32 %f2592, [%rd2+12]; bra.uni $L__BB155_219; $L__BB155_218: ld.local.f32 %f2594, [%rd2+20]; ld.local.f32 %f2593, [%rd2+16]; ld.local.f32 %f2592, [%rd2+12]; $L__BB155_219: ld.param.u32 %r534, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; bfe.u32 %r526, %r534, 16, 8; setp.eq.s32 %p265, %r526, 255; @%p265 bra $L__BB155_221; ld.param.u64 %rd466, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; cvta.to.local.u64 %rd465, %rd466; ld.param.u32 %r535, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; shr.u32 %r527, %r535, 14; cvt.u64.u32 %rd458, %r527; and.b64 %rd459, %rd458, 1020; add.s64 %rd460, %rd465, %rd459; ld.local.f32 %f2361, [%rd2+12]; sub.ftz.f32 %f2362, %f2592, %f2361; ld.local.f32 %f2363, [%rd2+16]; sub.ftz.f32 %f2364, %f2593, %f2363; ld.local.f32 %f2365, [%rd2+20]; sub.ftz.f32 %f2366, %f2594, %f2365; ld.local.f32 %f2367, [%rd460]; add.ftz.f32 %f2368, %f2367, %f2362; ld.local.f32 %f2369, [%rd460+4]; add.ftz.f32 %f2370, %f2369, %f2364; ld.local.f32 %f2371, [%rd460+8]; add.ftz.f32 %f2372, %f2371, %f2366; mul.ftz.f32 %f2373, %f2370, %f2370; fma.rn.ftz.f32 %f2374, %f2368, %f2368, %f2373; fma.rn.ftz.f32 %f2375, %f2372, %f2372, %f2374; sqrt.approx.ftz.f32 %f2376, %f2375; div.approx.ftz.f32 %f2592, %f2368, %f2376; div.approx.ftz.f32 %f2593, %f2370, %f2376; div.approx.ftz.f32 %f2594, %f2372, %f2376; $L__BB155_221: ld.param.u64 %rd468, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; cvta.to.local.u64 %rd467, %rd468; ld.param.u32 %r536, [_Z14svm_node_bevelILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; shr.u32 %r528, %r536, 24; mul.wide.u32 %rd463, %r528, 4; add.s64 %rd464, %rd467, %rd463; st.local.f32 [%rd464], %f2592; st.local.f32 [%rd464+4], %f2593; st.local.f32 [%rd464+8], %f2594; ret; $L__BB155_217: ld.local.f32 %f2594, [%rd2+20]; ld.local.f32 %f2593, [%rd2+16]; ld.local.f32 %f2592, [%rd2+12]; bra.uni $L__BB155_219; $L__BB155_4: ld.const.v2.u64 {%rd226, %rd227}, [kernel_params+2528]; mov.u32 %r539, 0; cvta.to.global.u64 %rd230, %rd226; shl.b64 %rd231, %rd3, 2; add.s64 %rd232, %rd230, %rd231; cvta.to.global.u64 %rd233, %rd227; shl.b64 %rd234, %rd3, 1; add.s64 %rd235, %rd233, %rd234; ld.global.u16 %rs19, [%rd235]; cvt.u32.u16 %r213, %rs19; ld.const.u64 %rd236, [kernel_params+2456]; cvta.to.global.u64 %rd237, %rd236; add.s64 %rd238, %rd237, %rd231; ld.global.u32 %r214, [%rd232]; add.s32 %r215, %r214, %r213; ld.global.u32 %r216, [%rd238]; mad.lo.s32 %r217, %r216, 1690739726, %r215; mad.lo.s32 %r573, %r217, 1103515245, 12345; mul.lo.s32 %r5, %r216, %r2; add.s32 %r218, %r213, 5; ld.const.u32 %r6, [kernel_params+1596]; shr.u32 %r219, %r218, 16; xor.b32 %r220, %r219, %r218; mul.lo.s32 %r221, %r220, 569420461; shr.u32 %r222, %r221, 15; xor.b32 %r223, %r222, %r221; mul.lo.s32 %r224, %r223, -749064809; shr.u32 %r225, %r224, 15; xor.b32 %r226, %r214, %r224; xor.b32 %r227, %r226, %r225; xor.b32 %r7, %r227, -419546133; ld.const.v2.u32 {%r228, %r229}, [kernel_params+1600]; mov.b32 %f666, %r229; setp.lt.ftz.f32 %p10, %f666, 0f3F800000; ld.const.u32 %r230, [kernel_params+1576]; selp.b32 %r231, %r230, %r214, %p10; add.s16 %rs20, %rs19, 5; and.b16 %rs21, %rs20, 255; cvt.u32.u16 %r232, %rs21; xor.b32 %r233, %r231, %r232; mul.lo.s32 %r234, %r233, -512718531; shr.u32 %r235, %r231, 16; xor.b32 %r236, %r234, %r235; bfe.u32 %r237, %r236, 4, 4; shr.u32 %r238, %r231, 8; xor.b32 %r239, %r236, %r238; xor.b32 %r240, %r239, %r237; mul.lo.s32 %r241, %r240, 153742143; shr.u32 %r242, %r231, 23; xor.b32 %r243, %r241, %r242; bfe.u32 %r244, %r243, 1, 7; xor.b32 %r245, %r244, %r243; shr.u32 %r246, %r231, 27; or.b32 %r247, %r246, 1; mul.lo.s32 %r248, %r247, %r245; mul.lo.s32 %r249, %r248, -1074439621; bfe.u32 %r250, %r249, 2, 6; xor.b32 %r251, %r250, %r249; mul.lo.s32 %r252, %r251, -1638916925; bfe.u32 %r253, %r252, 2, 6; xor.b32 %r254, %r253, %r252; mul.lo.s32 %r255, %r254, -933190689; and.b32 %r256, %r255, 255; bfe.u32 %r257, %r255, 5, 3; xor.b32 %r258, %r257, %r256; xor.b32 %r259, %r218, 61; xor.b32 %r260, %r231, %r259; mul.lo.s32 %r261, %r260, 9; shr.u32 %r262, %r261, 4; xor.b32 %r263, %r262, %r261; mul.lo.s32 %r10, %r263, 668265261; mul.lo.s32 %r11, %r258, %r228; ld.const.u64 %rd239, [kernel_params+2424]; cvta.to.global.u64 %rd4, %rd239; xor.b32 %r264, %r259, %r214; mul.lo.s32 %r265, %r264, 9; shr.u32 %r266, %r265, 4; xor.b32 %r267, %r266, %r265; mul.lo.s32 %r268, %r267, 668265261; cvt.rn.f32.u32 %f667, %r268; mul.ftz.f32 %f668, %f667, 0f2F7FFFFE; mul.ftz.f32 %f2, %f668, %f666; xor.b32 %r269, %r214, %r218; xor.b32 %r270, %r269, -905047700; mul.lo.s32 %r271, %r270, 9; shr.u32 %r272, %r271, 4; xor.b32 %r273, %r272, %r271; mul.lo.s32 %r274, %r273, 668265261; cvt.rn.f32.u32 %f669, %r274; mul.ftz.f32 %f670, %f669, 0f2F7FFFFE; mul.ftz.f32 %f3, %f670, %f666; ld.const.v2.u64 {%rd240, %rd241}, [kernel_params+2224]; ld.const.u64 %rd242, [kernel_params+2256]; cvta.to.global.u64 %rd7, %rd242; ld.const.v2.u64 {%rd243, %rd244}, [kernel_params+2176]; ld.const.u64 %rd245, [kernel_params+2216]; cvta.to.global.u64 %rd10, %rd245; ld.const.u64 %rd246, [kernel_params+2160]; cvta.to.global.u64 %rd11, %rd246; ld.const.u64 %rd247, [kernel_params+2312]; cvta.to.global.u64 %rd12, %rd247; ld.const.u64 %rd248, [kernel_params+2336]; cvta.to.global.u64 %rd13, %rd248; mul.ftz.f32 %f4, %f1, %f1; mul.ftz.f32 %f671, %f4, %f4; mul.ftz.f32 %f672, %f1, %f671; mul.ftz.f32 %f5, %f672, 0f40490FDB; ld.const.u64 %rd249, [kernel_params+2144]; cvta.to.global.u64 %rd14, %rd249; ld.const.v2.u64 {%rd250, %rd251}, [kernel_params+2096]; ld.const.v2.u64 {%rd252, %rd253}, [kernel_params+2128]; mov.f32 %f2592, 0f00000000; setp.eq.s32 %p11, %r6, 0; mov.u16 %rs22, 0; cvta.to.global.u64 %rd269, %rd243; mov.f32 %f2593, %f2592; mov.f32 %f2594, %f2592; $L__BB155_5: add.s32 %r15, %r539, %r5; brev.b32 %r275, %r15; mul.lo.s32 %r276, %r275, 1025551850; xor.b32 %r16, %r276, %r275; @%p11 bra $L__BB155_8; setp.geu.ftz.f32 %p12, %f666, 0f3F800000; shr.u32 %r277, %r10, 16; or.b32 %r278, %r277, 1; add.s32 %r279, %r10, %r16; mul.lo.s32 %r280, %r278, %r279; mul.lo.s32 %r281, %r280, 89287766; xor.b32 %r282, %r281, %r280; mul.lo.s32 %r283, %r282, 1403136100; xor.b32 %r284, %r283, %r282; brev.b32 %r285, %r284; neg.s32 %r286, %r228; and.b32 %r287, %r15, %r286; add.s32 %r288, %r228, -1; and.b32 %r289, %r285, %r288; or.b32 %r290, %r289, %r287; add.s32 %r291, %r11, %r290; shl.b32 %r292, %r228, 8; rem.u32 %r293, %r291, %r292; shl.b32 %r294, %r293, 1; mul.wide.u32 %rd254, %r294, 4; add.s64 %rd255, %rd4, %rd254; ld.global.f32 %f31, [%rd255]; ld.global.f32 %f2406, [%rd255+4]; @%p12 bra $L__BB155_12; add.ftz.f32 %f674, %f31, %f2; cvt.rmi.ftz.f32.f32 %f675, %f674; sub.ftz.f32 %f31, %f674, %f675; add.ftz.f32 %f676, %f2406, %f3; cvt.rmi.ftz.f32.f32 %f677, %f676; sub.ftz.f32 %f2406, %f676, %f677; bra.uni $L__BB155_12; $L__BB155_8: xor.b32 %r296, %r7, -122820198; add.s32 %r297, %r16, %r296; shr.u32 %r298, %r296, 16; or.b32 %r299, %r298, 1; mul.lo.s32 %r300, %r297, %r299; mul.lo.s32 %r301, %r300, 89287766; xor.b32 %r302, %r301, %r300; mul.lo.s32 %r303, %r302, 1403136100; xor.b32 %r17, %r303, %r302; setp.eq.s32 %p13, %r17, 0; mov.u32 %r544, 0; @%p13 bra $L__BB155_11; mov.u32 %r541, 0; mov.u32 %r542, %r17; mov.u32 %r544, %r541; $L__BB155_10: clz.b32 %r306, %r542; add.s32 %r307, %r306, %r541; mul.wide.u32 %rd256, %r307, 4; mov.u64 %rd257, sobol_burley_table; add.s64 %rd258, %rd257, %rd256; ld.const.u32 %r308, [%rd258+128]; xor.b32 %r544, %r308, %r544; add.s32 %r541, %r307, 1; shl.b32 %r309, %r542, %r306; shl.b32 %r542, %r309, 1; setp.ne.s32 %p14, %r542, 0; @%p14 bra $L__BB155_10; $L__BB155_11: brev.b32 %r310, %r17; xor.b32 %r311, %r7, -525684874; shr.u32 %r312, %r311, 16; or.b32 %r313, %r312, 1; mul.lo.s32 %r314, %r310, 1025551850; xor.b32 %r315, %r314, %r310; add.s32 %r316, %r315, %r311; mul.lo.s32 %r317, %r316, %r313; mul.lo.s32 %r318, %r317, 89287766; xor.b32 %r319, %r318, %r317; mul.lo.s32 %r320, %r544, 1025551850; xor.b32 %r321, %r320, %r544; xor.b32 %r322, %r7, -1802089138; add.s32 %r323, %r321, %r322; shr.u32 %r324, %r322, 16; or.b32 %r325, %r324, 1; mul.lo.s32 %r326, %r323, %r325; mul.lo.s32 %r327, %r326, 89287766; xor.b32 %r328, %r327, %r326; mul.lo.s32 %r329, %r328, 1403136100; xor.b32 %r330, %r329, %r328; brev.b32 %r331, %r330; cvt.rn.f32.u32 %f678, %r331; mul.lo.s32 %r332, %r319, 1403136100; xor.b32 %r333, %r332, %r319; brev.b32 %r334, %r333; cvt.rn.f32.u32 %f679, %r334; mul.ftz.f32 %f2406, %f678, 0f2F7FFFFE; mul.ftz.f32 %f31, %f679, 0f2F7FFFFE; $L__BB155_12: ld.local.v2.f32 {%f2416, %f2415}, [%rd2+24]; setp.neu.ftz.f32 %p15, %f2416, %f2415; ld.local.f32 %f2414, [%rd2+32]; setp.neu.ftz.f32 %p16, %f2416, %f2414; or.pred %p17, %p15, %p16; sub.ftz.f32 %f682, %f2414, %f2415; neg.ftz.f32 %f683, %f2415; selp.f32 %f684, %f2415, %f683, %p17; neg.ftz.f32 %f685, %f2414; selp.f32 %f686, %f685, %f2414, %p17; add.ftz.f32 %f687, %f2416, %f686; sub.ftz.f32 %f688, %f684, %f2416; mul.ftz.f32 %f689, %f687, %f687; fma.rn.ftz.f32 %f690, %f682, %f682, %f689; fma.rn.ftz.f32 %f691, %f688, %f688, %f690; sqrt.approx.ftz.f32 %f692, %f691; div.approx.ftz.f32 %f2413, %f682, %f692; div.approx.ftz.f32 %f2412, %f687, %f692; div.approx.ftz.f32 %f2411, %f688, %f692; mul.ftz.f32 %f693, %f2415, %f2411; mul.ftz.f32 %f694, %f2414, %f2412; sub.ftz.f32 %f39, %f693, %f694; mul.ftz.f32 %f695, %f2414, %f2413; mul.ftz.f32 %f696, %f2416, %f2411; sub.ftz.f32 %f40, %f695, %f696; mul.ftz.f32 %f697, %f2416, %f2412; mul.ftz.f32 %f698, %f2415, %f2413; sub.ftz.f32 %f41, %f697, %f698; setp.lt.ftz.f32 %p18, %f31, 0f3F000000; @%p18 bra $L__BB155_16; bra.uni $L__BB155_13; $L__BB155_16: add.ftz.f32 %f2407, %f31, %f31; mov.f32 %f2418, 0f3E800000; mov.f32 %f2417, 0f3F000000; mov.f32 %f2408, %f2414; mov.f32 %f2409, %f2415; mov.f32 %f2410, %f2416; mov.f32 %f2414, %f41; mov.f32 %f2415, %f40; mov.f32 %f2416, %f39; mov.f32 %f2419, %f2418; bra.uni $L__BB155_17; $L__BB155_13: setp.lt.ftz.f32 %p19, %f31, 0f3F400000; @%p19 bra $L__BB155_15; bra.uni $L__BB155_14; $L__BB155_15: add.ftz.f32 %f706, %f31, 0fBF000000; mov.f32 %f2418, 0f3F000000; mul.ftz.f32 %f2407, %f706, 0f40800000; mov.f32 %f2417, 0f3E800000; mov.f32 %f2408, %f2411; mov.f32 %f2409, %f2412; mov.f32 %f2410, %f2413; mov.f32 %f2411, %f2414; mov.f32 %f2412, %f2415; mov.f32 %f2413, %f2416; mov.f32 %f2414, %f41; mov.f32 %f2415, %f40; mov.f32 %f2416, %f39; mov.f32 %f2419, %f2417; bra.uni $L__BB155_17; $L__BB155_14: add.ftz.f32 %f702, %f31, 0fBF400000; mul.ftz.f32 %f2407, %f702, 0f40800000; mov.f32 %f2419, 0f3F000000; mov.f32 %f2417, 0f3E800000; mov.f32 %f2408, %f41; mov.f32 %f2409, %f40; mov.f32 %f2410, %f39; mov.f32 %f2418, %f2417; $L__BB155_17: mov.f32 %f711, 0f3F200000; sub.ftz.f32 %f712, %f711, 0f3EA00000; add.ftz.f32 %f713, %f712, 0f3D700000; add.ftz.f32 %f714, %f713, 0fBB800000; sub.ftz.f32 %f60, %f714, %f2406; abs.ftz.f32 %f715, %f60; setp.lt.ftz.f32 %p20, %f715, 0f358637BD; mov.f32 %f716, 0f3F800000; sub.ftz.f32 %f717, %f716, 0f3E800000; mul.ftz.f32 %f718, %f717, %f717; mul.ftz.f32 %f719, %f717, 0f3E800000; mul.ftz.f32 %f720, %f719, 0f41A00000; mul.ftz.f32 %f721, %f718, %f720; setp.eq.ftz.f32 %p21, %f721, 0f00000000; or.pred %p22, %p20, %p21; mov.f32 %f2420, 0f3E800000; @%p22 bra $L__BB155_28; div.approx.ftz.f32 %f728, %f60, %f721; mov.f32 %f729, 0f3E800000; sub.ftz.f32 %f730, %f729, %f728; cvt.ftz.sat.f32.f32 %f2420, %f730; mul.ftz.f32 %f731, %f2420, %f2420; mul.ftz.f32 %f732, %f2420, %f731; sub.ftz.f32 %f733, %f716, %f2420; mul.ftz.f32 %f734, %f731, 0f41200000; mul.ftz.f32 %f735, %f732, 0f41A00000; sub.ftz.f32 %f736, %f734, %f735; mul.ftz.f32 %f737, %f731, 0f41700000; fma.rn.ftz.f32 %f738, %f731, %f737, %f736; mul.ftz.f32 %f739, %f731, 0fC0800000; fma.rn.ftz.f32 %f740, %f739, %f732, %f738; sub.ftz.f32 %f62, %f740, %f2406; mul.ftz.f32 %f741, %f2420, %f733; mul.ftz.f32 %f742, %f741, 0f41A00000; mul.ftz.f32 %f743, %f733, %f733; mul.ftz.f32 %f63, %f743, %f742; abs.ftz.f32 %f744, %f62; setp.lt.ftz.f32 %p23, %f744, 0f358637BD; setp.eq.ftz.f32 %p24, %f63, 0f00000000; or.pred %p25, %p23, %p24; @%p25 bra $L__BB155_28; div.approx.ftz.f32 %f745, %f62, %f63; sub.ftz.f32 %f746, %f2420, %f745; cvt.ftz.sat.f32.f32 %f2420, %f746; mul.ftz.f32 %f747, %f2420, %f2420; mul.ftz.f32 %f748, %f2420, %f747; sub.ftz.f32 %f750, %f716, %f2420; mul.ftz.f32 %f751, %f747, 0f41200000; mul.ftz.f32 %f752, %f748, 0f41A00000; sub.ftz.f32 %f753, %f751, %f752; mul.ftz.f32 %f754, %f747, 0f41700000; fma.rn.ftz.f32 %f755, %f747, %f754, %f753; mul.ftz.f32 %f756, %f747, 0fC0800000; fma.rn.ftz.f32 %f757, %f756, %f748, %f755; sub.ftz.f32 %f65, %f757, %f2406; mul.ftz.f32 %f758, %f2420, %f750; mul.ftz.f32 %f759, %f758, 0f41A00000; mul.ftz.f32 %f760, %f750, %f750; mul.ftz.f32 %f66, %f760, %f759; abs.ftz.f32 %f761, %f65; setp.lt.ftz.f32 %p26, %f761, 0f358637BD; setp.eq.ftz.f32 %p27, %f66, 0f00000000; or.pred %p28, %p26, %p27; @%p28 bra $L__BB155_28; div.approx.ftz.f32 %f762, %f65, %f66; sub.ftz.f32 %f763, %f2420, %f762; cvt.ftz.sat.f32.f32 %f2420, %f763; mul.ftz.f32 %f764, %f2420, %f2420; mul.ftz.f32 %f765, %f2420, %f764; sub.ftz.f32 %f767, %f716, %f2420; mul.ftz.f32 %f768, %f764, 0f41200000; mul.ftz.f32 %f769, %f765, 0f41A00000; sub.ftz.f32 %f770, %f768, %f769; mul.ftz.f32 %f771, %f764, 0f41700000; fma.rn.ftz.f32 %f772, %f764, %f771, %f770; mul.ftz.f32 %f773, %f764, 0fC0800000; fma.rn.ftz.f32 %f774, %f773, %f765, %f772; sub.ftz.f32 %f68, %f774, %f2406; mul.ftz.f32 %f775, %f2420, %f767; mul.ftz.f32 %f776, %f775, 0f41A00000; mul.ftz.f32 %f777, %f767, %f767; mul.ftz.f32 %f69, %f777, %f776; abs.ftz.f32 %f778, %f68; setp.lt.ftz.f32 %p29, %f778, 0f358637BD; setp.eq.ftz.f32 %p30, %f69, 0f00000000; or.pred %p31, %p29, %p30; @%p31 bra $L__BB155_28; div.approx.ftz.f32 %f779, %f68, %f69; sub.ftz.f32 %f780, %f2420, %f779; cvt.ftz.sat.f32.f32 %f2420, %f780; mul.ftz.f32 %f781, %f2420, %f2420; mul.ftz.f32 %f782, %f2420, %f781; sub.ftz.f32 %f784, %f716, %f2420; mul.ftz.f32 %f785, %f781, 0f41200000; mul.ftz.f32 %f786, %f782, 0f41A00000; sub.ftz.f32 %f787, %f785, %f786; mul.ftz.f32 %f788, %f781, 0f41700000; fma.rn.ftz.f32 %f789, %f781, %f788, %f787; mul.ftz.f32 %f790, %f781, 0fC0800000; fma.rn.ftz.f32 %f791, %f790, %f782, %f789; sub.ftz.f32 %f71, %f791, %f2406; mul.ftz.f32 %f792, %f2420, %f784; mul.ftz.f32 %f793, %f792, 0f41A00000; mul.ftz.f32 %f794, %f784, %f784; mul.ftz.f32 %f72, %f794, %f793; abs.ftz.f32 %f795, %f71; setp.lt.ftz.f32 %p32, %f795, 0f358637BD; setp.eq.ftz.f32 %p33, %f72, 0f00000000; or.pred %p34, %p32, %p33; @%p34 bra $L__BB155_28; div.approx.ftz.f32 %f796, %f71, %f72; sub.ftz.f32 %f797, %f2420, %f796; cvt.ftz.sat.f32.f32 %f2420, %f797; mul.ftz.f32 %f798, %f2420, %f2420; mul.ftz.f32 %f799, %f2420, %f798; sub.ftz.f32 %f801, %f716, %f2420; mul.ftz.f32 %f802, %f798, 0f41200000; mul.ftz.f32 %f803, %f799, 0f41A00000; sub.ftz.f32 %f804, %f802, %f803; mul.ftz.f32 %f805, %f798, 0f41700000; fma.rn.ftz.f32 %f806, %f798, %f805, %f804; mul.ftz.f32 %f807, %f798, 0fC0800000; fma.rn.ftz.f32 %f808, %f807, %f799, %f806; sub.ftz.f32 %f74, %f808, %f2406; mul.ftz.f32 %f809, %f2420, %f801; mul.ftz.f32 %f810, %f809, 0f41A00000; mul.ftz.f32 %f811, %f801, %f801; mul.ftz.f32 %f75, %f811, %f810; abs.ftz.f32 %f812, %f74; setp.lt.ftz.f32 %p35, %f812, 0f358637BD; setp.eq.ftz.f32 %p36, %f75, 0f00000000; or.pred %p37, %p35, %p36; @%p37 bra $L__BB155_28; div.approx.ftz.f32 %f813, %f74, %f75; sub.ftz.f32 %f814, %f2420, %f813; cvt.ftz.sat.f32.f32 %f2420, %f814; mul.ftz.f32 %f815, %f2420, %f2420; mul.ftz.f32 %f816, %f2420, %f815; sub.ftz.f32 %f818, %f716, %f2420; mul.ftz.f32 %f819, %f815, 0f41200000; mul.ftz.f32 %f820, %f816, 0f41A00000; sub.ftz.f32 %f821, %f819, %f820; mul.ftz.f32 %f822, %f815, 0f41700000; fma.rn.ftz.f32 %f823, %f815, %f822, %f821; mul.ftz.f32 %f824, %f815, 0fC0800000; fma.rn.ftz.f32 %f825, %f824, %f816, %f823; sub.ftz.f32 %f77, %f825, %f2406; mul.ftz.f32 %f826, %f2420, %f818; mul.ftz.f32 %f827, %f826, 0f41A00000; mul.ftz.f32 %f828, %f818, %f818; mul.ftz.f32 %f78, %f828, %f827; abs.ftz.f32 %f829, %f77; setp.lt.ftz.f32 %p38, %f829, 0f358637BD; setp.eq.ftz.f32 %p39, %f78, 0f00000000; or.pred %p40, %p38, %p39; @%p40 bra $L__BB155_28; div.approx.ftz.f32 %f830, %f77, %f78; sub.ftz.f32 %f831, %f2420, %f830; cvt.ftz.sat.f32.f32 %f2420, %f831; mul.ftz.f32 %f832, %f2420, %f2420; mul.ftz.f32 %f833, %f2420, %f832; sub.ftz.f32 %f835, %f716, %f2420; mul.ftz.f32 %f836, %f832, 0f41200000; mul.ftz.f32 %f837, %f833, 0f41A00000; sub.ftz.f32 %f838, %f836, %f837; mul.ftz.f32 %f839, %f832, 0f41700000; fma.rn.ftz.f32 %f840, %f832, %f839, %f838; mul.ftz.f32 %f841, %f832, 0fC0800000; fma.rn.ftz.f32 %f842, %f841, %f833, %f840; sub.ftz.f32 %f80, %f842, %f2406; mul.ftz.f32 %f843, %f2420, %f835; mul.ftz.f32 %f844, %f843, 0f41A00000; mul.ftz.f32 %f845, %f835, %f835; mul.ftz.f32 %f81, %f845, %f844; abs.ftz.f32 %f846, %f80; setp.lt.ftz.f32 %p41, %f846, 0f358637BD; setp.eq.ftz.f32 %p42, %f81, 0f00000000; or.pred %p43, %p41, %p42; @%p43 bra $L__BB155_28; div.approx.ftz.f32 %f847, %f80, %f81; sub.ftz.f32 %f848, %f2420, %f847; cvt.ftz.sat.f32.f32 %f2420, %f848; mul.ftz.f32 %f849, %f2420, %f2420; mul.ftz.f32 %f850, %f2420, %f849; sub.ftz.f32 %f852, %f716, %f2420; mul.ftz.f32 %f853, %f849, 0f41200000; mul.ftz.f32 %f854, %f850, 0f41A00000; sub.ftz.f32 %f855, %f853, %f854; mul.ftz.f32 %f856, %f849, 0f41700000; fma.rn.ftz.f32 %f857, %f849, %f856, %f855; mul.ftz.f32 %f858, %f849, 0fC0800000; fma.rn.ftz.f32 %f859, %f858, %f850, %f857; sub.ftz.f32 %f83, %f859, %f2406; mul.ftz.f32 %f860, %f2420, %f852; mul.ftz.f32 %f861, %f860, 0f41A00000; mul.ftz.f32 %f862, %f852, %f852; mul.ftz.f32 %f84, %f862, %f861; abs.ftz.f32 %f863, %f83; setp.lt.ftz.f32 %p44, %f863, 0f358637BD; setp.eq.ftz.f32 %p45, %f84, 0f00000000; or.pred %p46, %p44, %p45; @%p46 bra $L__BB155_28; div.approx.ftz.f32 %f864, %f83, %f84; sub.ftz.f32 %f865, %f2420, %f864; cvt.ftz.sat.f32.f32 %f2420, %f865; mul.ftz.f32 %f866, %f2420, %f2420; mul.ftz.f32 %f867, %f2420, %f866; sub.ftz.f32 %f869, %f716, %f2420; mul.ftz.f32 %f870, %f866, 0f41200000; mul.ftz.f32 %f871, %f867, 0f41A00000; sub.ftz.f32 %f872, %f870, %f871; mul.ftz.f32 %f873, %f866, 0f41700000; fma.rn.ftz.f32 %f874, %f866, %f873, %f872; mul.ftz.f32 %f875, %f866, 0fC0800000; fma.rn.ftz.f32 %f876, %f875, %f867, %f874; sub.ftz.f32 %f86, %f876, %f2406; mul.ftz.f32 %f877, %f2420, %f869; mul.ftz.f32 %f878, %f877, 0f41A00000; mul.ftz.f32 %f879, %f869, %f869; mul.ftz.f32 %f87, %f879, %f878; abs.ftz.f32 %f880, %f86; setp.lt.ftz.f32 %p47, %f880, 0f358637BD; setp.eq.ftz.f32 %p48, %f87, 0f00000000; or.pred %p49, %p47, %p48; @%p49 bra $L__BB155_28; div.approx.ftz.f32 %f881, %f86, %f87; sub.ftz.f32 %f882, %f2420, %f881; cvt.ftz.sat.f32.f32 %f2420, %f882; $L__BB155_28: mul.ftz.f32 %f883, %f2407, 0f40C90FDB; mul.ftz.f32 %f90, %f1, %f2420; mul.ftz.f32 %f884, %f90, %f90; sub.ftz.f32 %f885, %f4, %f884; mov.f32 %f886, 0f00000000; max.ftz.f32 %f887, %f885, %f886; sqrt.approx.ftz.f32 %f888, %f887; cos.approx.ftz.f32 %f889, %f883; mul.ftz.f32 %f890, %f90, %f889; sin.approx.ftz.f32 %f891, %f883; mul.ftz.f32 %f892, %f90, %f891; mul.ftz.f32 %f893, %f2416, %f892; mul.ftz.f32 %f894, %f2415, %f892; mul.ftz.f32 %f895, %f2414, %f892; fma.rn.ftz.f32 %f896, %f2413, %f890, %f893; fma.rn.ftz.f32 %f897, %f2412, %f890, %f894; fma.rn.ftz.f32 %f898, %f2411, %f890, %f895; ld.local.f32 %f91, [%rd2]; fma.rn.ftz.f32 %f899, %f2410, %f888, %f91; ld.local.f32 %f92, [%rd2+4]; fma.rn.ftz.f32 %f900, %f2409, %f888, %f92; ld.local.f32 %f93, [%rd2+8]; fma.rn.ftz.f32 %f901, %f2408, %f888, %f93; add.ftz.f32 %f2450, %f899, %f896; add.ftz.f32 %f2451, %f900, %f897; add.ftz.f32 %f2452, %f898, %f901; neg.ftz.f32 %f97, %f2410; neg.ftz.f32 %f98, %f2409; neg.ftz.f32 %f99, %f2408; add.ftz.f32 %f100, %f888, %f888; ld.local.f32 %f101, [%rd2+84]; mov.b32 %r25, %f2450; setp.nan.ftz.f32 %p50, %f2450, %f2450; mov.u16 %rs32, %rs22; @%p50 bra $L__BB155_31; or.b32 %r335, %r25, -2147483648; setp.ne.s32 %p51, %r335, -2147483648; add.ftz.f32 %f902, %f2450, %f2450; setp.eq.ftz.f32 %p52, %f902, %f2450; and.pred %p53, %p52, %p51; mov.u16 %rs32, %rs22; @%p53 bra $L__BB155_31; shl.b32 %r336, %r25, 1; setp.lt.u32 %p54, %r336, -16777215; selp.u16 %rs32, 1, 0, %p54; $L__BB155_31: setp.eq.s16 %p55, %rs32, 0; mov.u16 %rs34, %rs22; @%p55 bra $L__BB155_37; mov.b32 %r26, %f97; setp.nan.ftz.f32 %p56, %f2410, %f2410; mov.u16 %rs34, 0; mov.u16 %rs33, %rs34; @%p56 bra $L__BB155_35; or.b32 %r337, %r26, -2147483648; setp.ne.s32 %p57, %r337, -2147483648; mul.ftz.f32 %f903, %f2410, 0fC0000000; setp.eq.ftz.f32 %p58, %f903, %f97; and.pred %p59, %p58, %p57; mov.u16 %rs33, %rs34; @%p59 bra $L__BB155_35; shl.b32 %r338, %r26, 1; setp.lt.u32 %p60, %r338, -16777215; selp.u16 %rs33, 1, 0, %p60; $L__BB155_35: setp.eq.s16 %p61, %rs33, 0; @%p61 bra $L__BB155_37; mul.ftz.f32 %f904, %f2410, %f2410; fma.rn.ftz.f32 %f905, %f2409, %f2409, %f904; fma.rn.ftz.f32 %f906, %f2408, %f2408, %f905; setp.neu.ftz.f32 %p62, %f906, 0f00000000; selp.u16 %rs34, 1, 0, %p62; $L__BB155_37: setp.eq.s16 %p63, %rs34, 0; @%p63 bra $L__BB155_147; ld.const.u32 %r537, [kernel_params+1108]; setp.eq.s32 %p64, %r537, 0; abs.ftz.f32 %f907, %f97; setp.gt.ftz.f32 %p65, %f907, 0f17800000; mov.b32 %r339, %f97; and.b32 %r340, %r339, -2147483648; or.b32 %r341, %r340, 394264576; mov.b32 %f908, %r341; selp.f32 %f2447, %f97, %f908, %p65; abs.ftz.f32 %f909, %f98; setp.gt.ftz.f32 %p66, %f909, 0f17800000; mov.b32 %r342, %f98; and.b32 %r343, %r342, -2147483648; or.b32 %r344, %r343, 394264576; mov.b32 %f910, %r344; selp.f32 %f2448, %f98, %f910, %p66; abs.ftz.f32 %f911, %f99; setp.gt.ftz.f32 %p67, %f911, 0f17800000; mov.b32 %r345, %f99; and.b32 %r346, %r345, -2147483648; or.b32 %r347, %r346, 394264576; mov.b32 %f912, %r347; selp.f32 %f2449, %f99, %f912, %p67; rcp.approx.ftz.f32 %f2444, %f2447; rcp.approx.ftz.f32 %f2445, %f2448; rcp.approx.ftz.f32 %f2446, %f2449; mul.wide.s32 %rd261, %r3, 4; add.s64 %rd20, %rd14, %rd261; cvta.to.global.u64 %rd262, %rd244; add.s64 %rd21, %rd262, %rd261; add.u64 %rd22, %SPL, 0; @%p64 bra $L__BB155_113; mov.u32 %r349, 1985229328; st.local.u32 [%rd22], %r349; ld.global.u32 %r553, [%rd20]; mov.u32 %r574, 0; st.local.u32 [%rd1], %r574; ld.global.u32 %r28, [%rd21]; and.b32 %r351, %r28, 4; setp.ne.s32 %p68, %r351, 0; mov.u32 %r545, -1; @%p68 bra $L__BB155_49; and.b32 %r352, %r28, 2; setp.eq.s32 %p69, %r352, 0; @%p69 bra $L__BB155_47; bra.uni $L__BB155_41; $L__BB155_47: mul.wide.s32 %rd271, %r3, 208; add.s64 %rd272, %rd11, %rd271; ld.global.v4.f32 {%f2432, %f2433, %f2434, %f2435}, [%rd272+48]; ld.global.v4.f32 {%f2436, %f2437, %f2438, %f2439}, [%rd272+64]; ld.global.v4.f32 {%f2440, %f2441, %f2442, %f2443}, [%rd272+80]; bra.uni $L__BB155_48; $L__BB155_147: mov.u32 %r451, 0; st.local.u32 [%rd1], %r451; bra.uni $L__BB155_148; $L__BB155_113: mov.u32 %r417, 1985229328; st.local.u32 [%rd22], %r417; ld.global.u32 %r584, [%rd20]; mov.u32 %r583, 0; st.local.u32 [%rd1], %r583; ld.global.u8 %rs30, [%rd21]; and.b16 %rs31, %rs30, 4; setp.ne.s16 %p155, %rs31, 0; mov.u32 %r576, -1; @%p155 bra $L__BB155_115; mul.wide.s32 %rd339, %r3, 208; add.s64 %rd340, %rd11, %rd339; ld.global.v4.f32 {%f1620, %f1621, %f1622, %f1623}, [%rd340+48]; ld.global.v4.f32 {%f1628, %f1629, %f1630, %f1631}, [%rd340+64]; ld.global.v4.f32 {%f1636, %f1637, %f1638, %f1639}, [%rd340+80]; mul.ftz.f32 %f1644, %f2451, %f1621; fma.rn.ftz.f32 %f1645, %f2450, %f1620, %f1644; fma.rn.ftz.f32 %f1646, %f2452, %f1622, %f1645; add.ftz.f32 %f327, %f1623, %f1646; mul.ftz.f32 %f1647, %f2451, %f1629; fma.rn.ftz.f32 %f1648, %f2450, %f1628, %f1647; fma.rn.ftz.f32 %f1649, %f2452, %f1630, %f1648; add.ftz.f32 %f328, %f1631, %f1649; mul.ftz.f32 %f1650, %f2451, %f1637; fma.rn.ftz.f32 %f1651, %f2450, %f1636, %f1650; fma.rn.ftz.f32 %f1652, %f2452, %f1638, %f1651; add.ftz.f32 %f2452, %f1639, %f1652; mul.ftz.f32 %f1653, %f1621, %f98; fma.rn.ftz.f32 %f1654, %f1620, %f97, %f1653; fma.rn.ftz.f32 %f1655, %f1622, %f99, %f1654; mul.ftz.f32 %f1656, %f1629, %f98; fma.rn.ftz.f32 %f1657, %f1628, %f97, %f1656; fma.rn.ftz.f32 %f1658, %f1630, %f99, %f1657; mul.ftz.f32 %f1659, %f1637, %f98; fma.rn.ftz.f32 %f1660, %f1636, %f97, %f1659; fma.rn.ftz.f32 %f1661, %f1638, %f99, %f1660; abs.ftz.f32 %f1662, %f1655; setp.gt.ftz.f32 %p156, %f1662, 0f17800000; mov.b32 %r419, %f1655; and.b32 %r420, %r419, -2147483648; or.b32 %r421, %r420, 394264576; mov.b32 %f1663, %r421; selp.f32 %f2447, %f1655, %f1663, %p156; abs.ftz.f32 %f1664, %f1658; setp.gt.ftz.f32 %p157, %f1664, 0f17800000; mov.b32 %r422, %f1658; and.b32 %r423, %r422, -2147483648; or.b32 %r424, %r423, 394264576; mov.b32 %f1665, %r424; selp.f32 %f2448, %f1658, %f1665, %p157; abs.ftz.f32 %f1666, %f1661; setp.gt.ftz.f32 %p158, %f1666, 0f17800000; mov.b32 %r425, %f1661; and.b32 %r426, %r425, -2147483648; or.b32 %r427, %r426, 394264576; mov.b32 %f1667, %r427; selp.f32 %f2449, %f1661, %f1667, %p158; rcp.approx.ftz.f32 %f2444, %f2447; rcp.approx.ftz.f32 %f2445, %f2448; rcp.approx.ftz.f32 %f2446, %f2449; mov.f32 %f2450, %f327; mov.f32 %f2451, %f328; mov.u32 %r576, %r3; $L__BB155_115: setp.eq.s32 %p159, %r584, 1985229328; setp.lt.s32 %p160, %r584, 0; or.pred %p161, %p160, %p159; @%p161 bra $L__BB155_127; cvta.to.global.u64 %rd88, %rd250; $L__BB155_117: mul.wide.s32 %rd342, %r584, 16; add.s64 %rd89, %rd88, %rd342; ld.global.v4.f32 {%f1668, %f1669, %f1670, %f1671}, [%rd89]; mov.b32 %r584, %f1671; mov.b32 %r120, %f1670; mov.b32 %r119, %f1669; mov.b32 %r118, %f1668; mov.b64 %rd343, {%r118, %r119}; cvt.u32.u64 %r429, %rd343; and.b32 %r430, %r429, 1024; setp.eq.s32 %p162, %r430, 0; and.b32 %r122, %r118, 2047; @%p162 bra $L__BB155_120; bra.uni $L__BB155_118; $L__BB155_120: setp.ne.s32 %p168, %r122, 0; ld.global.v4.f32 {%f1812, %f1813, %f1814, %f1815}, [%rd89+16]; ld.global.v4.f32 {%f1820, %f1821, %f1822, %f1823}, [%rd89+32]; ld.global.v4.f32 {%f1828, %f1829, %f1830, %f1831}, [%rd89+48]; sub.ftz.f32 %f1836, %f1812, %f2450; mul.ftz.f32 %f1837, %f2444, %f1836; sub.ftz.f32 %f1838, %f1814, %f2450; mul.ftz.f32 %f1839, %f2444, %f1838; sub.ftz.f32 %f1840, %f1820, %f2451; mul.ftz.f32 %f1841, %f2445, %f1840; sub.ftz.f32 %f1842, %f1822, %f2451; mul.ftz.f32 %f1843, %f2445, %f1842; sub.ftz.f32 %f1844, %f1828, %f2452; mul.ftz.f32 %f1845, %f2446, %f1844; sub.ftz.f32 %f1846, %f1830, %f2452; mul.ftz.f32 %f1847, %f2446, %f1846; min.ftz.f32 %f1848, %f1837, %f1839; min.ftz.f32 %f1849, %f1841, %f1843; min.ftz.f32 %f1850, %f1845, %f1847; mov.f32 %f1851, 0f00000000; max.ftz.f32 %f1852, %f1851, %f1848; max.ftz.f32 %f1853, %f1849, %f1850; max.ftz.f32 %f2499, %f1852, %f1853; max.ftz.f32 %f1854, %f1837, %f1839; max.ftz.f32 %f1855, %f1841, %f1843; max.ftz.f32 %f1856, %f1845, %f1847; min.ftz.f32 %f1857, %f100, %f1854; min.ftz.f32 %f1858, %f1855, %f1856; min.ftz.f32 %f1859, %f1857, %f1858; sub.ftz.f32 %f1860, %f1813, %f2450; mul.ftz.f32 %f1861, %f2444, %f1860; sub.ftz.f32 %f1862, %f1815, %f2450; mul.ftz.f32 %f1863, %f2444, %f1862; sub.ftz.f32 %f1864, %f1821, %f2451; mul.ftz.f32 %f1865, %f2445, %f1864; sub.ftz.f32 %f1866, %f1823, %f2451; mul.ftz.f32 %f1867, %f2445, %f1866; sub.ftz.f32 %f1868, %f1829, %f2452; mul.ftz.f32 %f1869, %f2446, %f1868; sub.ftz.f32 %f1870, %f1831, %f2452; mul.ftz.f32 %f1871, %f2446, %f1870; min.ftz.f32 %f1872, %f1861, %f1863; min.ftz.f32 %f1873, %f1865, %f1867; min.ftz.f32 %f1874, %f1869, %f1871; max.ftz.f32 %f1875, %f1851, %f1872; max.ftz.f32 %f1876, %f1873, %f1874; max.ftz.f32 %f2500, %f1875, %f1876; max.ftz.f32 %f1877, %f1861, %f1863; max.ftz.f32 %f1878, %f1865, %f1867; max.ftz.f32 %f1879, %f1869, %f1871; min.ftz.f32 %f1880, %f100, %f1877; min.ftz.f32 %f1881, %f1878, %f1879; min.ftz.f32 %f1882, %f1880, %f1881; setp.ge.ftz.f32 %p169, %f1859, %f2499; and.pred %p170, %p168, %p169; selp.u32 %r434, 1, 0, %p170; setp.ltu.ftz.f32 %p171, %f1882, %f2500; and.b32 %r436, %r119, 2047; setp.eq.s32 %p172, %r436, 0; or.pred %p173, %p172, %p171; selp.b32 %r437, 0, 2, %p173; or.b32 %r582, %r437, %r434; bra.uni $L__BB155_121; $L__BB155_118: setp.ne.s32 %p163, %r122, 0; ld.global.v4.f32 {%f1675, %f1676, %f1677, %f1678}, [%rd89+16]; ld.global.v4.f32 {%f1683, %f1684, %f1685, %f1686}, [%rd89+32]; ld.global.v4.f32 {%f1691, %f1692, %f1693, %f1694}, [%rd89+48]; mul.ftz.f32 %f1699, %f2448, %f1676; fma.rn.ftz.f32 %f1700, %f2447, %f1675, %f1699; fma.rn.ftz.f32 %f1701, %f2449, %f1677, %f1700; mul.ftz.f32 %f1702, %f2448, %f1684; fma.rn.ftz.f32 %f1703, %f2447, %f1683, %f1702; fma.rn.ftz.f32 %f1704, %f2449, %f1685, %f1703; mul.ftz.f32 %f1705, %f2448, %f1692; fma.rn.ftz.f32 %f1706, %f2447, %f1691, %f1705; fma.rn.ftz.f32 %f1707, %f2449, %f1693, %f1706; mul.ftz.f32 %f1708, %f2451, %f1676; fma.rn.ftz.f32 %f1709, %f2450, %f1675, %f1708; fma.rn.ftz.f32 %f1710, %f2452, %f1677, %f1709; add.ftz.f32 %f1711, %f1678, %f1710; mul.ftz.f32 %f1712, %f2451, %f1684; fma.rn.ftz.f32 %f1713, %f2450, %f1683, %f1712; fma.rn.ftz.f32 %f1714, %f2452, %f1685, %f1713; add.ftz.f32 %f1715, %f1686, %f1714; mul.ftz.f32 %f1716, %f2451, %f1692; fma.rn.ftz.f32 %f1717, %f2450, %f1691, %f1716; fma.rn.ftz.f32 %f1718, %f2452, %f1693, %f1717; add.ftz.f32 %f1719, %f1694, %f1718; rcp.approx.ftz.f32 %f1720, %f1701; rcp.approx.ftz.f32 %f1721, %f1704; rcp.approx.ftz.f32 %f1722, %f1707; mul.ftz.f32 %f1723, %f1720, %f1711; neg.ftz.f32 %f1724, %f1723; mul.ftz.f32 %f1725, %f1721, %f1715; neg.ftz.f32 %f1726, %f1725; mul.ftz.f32 %f1727, %f1722, %f1719; neg.ftz.f32 %f1728, %f1727; sub.ftz.f32 %f1729, %f1720, %f1723; sub.ftz.f32 %f1730, %f1721, %f1725; sub.ftz.f32 %f1731, %f1722, %f1727; min.ftz.f32 %f1732, %f1724, %f1729; min.ftz.f32 %f1733, %f1726, %f1730; min.ftz.f32 %f1734, %f1728, %f1731; max.ftz.f32 %f1735, %f1724, %f1729; max.ftz.f32 %f1736, %f1726, %f1730; max.ftz.f32 %f1737, %f1728, %f1731; mov.f32 %f1738, 0f00000000; max.ftz.f32 %f1739, %f1738, %f1732; max.ftz.f32 %f1740, %f1733, %f1734; max.ftz.f32 %f2499, %f1739, %f1740; min.ftz.f32 %f1741, %f100, %f1735; min.ftz.f32 %f1742, %f1736, %f1737; min.ftz.f32 %f1743, %f1741, %f1742; setp.le.ftz.f32 %p164, %f2499, %f1743; and.pred %p165, %p163, %p164; selp.u32 %r582, 1, 0, %p165; ld.global.v4.f32 {%f1744, %f1745, %f1746, %f1747}, [%rd89+64]; ld.global.v4.f32 {%f1752, %f1753, %f1754, %f1755}, [%rd89+80]; ld.global.v4.f32 {%f1760, %f1761, %f1762, %f1763}, [%rd89+96]; mul.ftz.f32 %f1768, %f2448, %f1745; fma.rn.ftz.f32 %f1769, %f2447, %f1744, %f1768; fma.rn.ftz.f32 %f1770, %f2449, %f1746, %f1769; mul.ftz.f32 %f1771, %f2448, %f1753; fma.rn.ftz.f32 %f1772, %f2447, %f1752, %f1771; fma.rn.ftz.f32 %f1773, %f2449, %f1754, %f1772; mul.ftz.f32 %f1774, %f2448, %f1761; fma.rn.ftz.f32 %f1775, %f2447, %f1760, %f1774; fma.rn.ftz.f32 %f1776, %f2449, %f1762, %f1775; mul.ftz.f32 %f1777, %f2451, %f1745; fma.rn.ftz.f32 %f1778, %f2450, %f1744, %f1777; fma.rn.ftz.f32 %f1779, %f2452, %f1746, %f1778; add.ftz.f32 %f1780, %f1747, %f1779; mul.ftz.f32 %f1781, %f2451, %f1753; fma.rn.ftz.f32 %f1782, %f2450, %f1752, %f1781; fma.rn.ftz.f32 %f1783, %f2452, %f1754, %f1782; add.ftz.f32 %f1784, %f1755, %f1783; mul.ftz.f32 %f1785, %f2451, %f1761; fma.rn.ftz.f32 %f1786, %f2450, %f1760, %f1785; fma.rn.ftz.f32 %f1787, %f2452, %f1762, %f1786; add.ftz.f32 %f1788, %f1763, %f1787; rcp.approx.ftz.f32 %f1789, %f1770; rcp.approx.ftz.f32 %f1790, %f1773; rcp.approx.ftz.f32 %f1791, %f1776; mul.ftz.f32 %f1792, %f1789, %f1780; neg.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f1790, %f1784; neg.ftz.f32 %f1795, %f1794; mul.ftz.f32 %f1796, %f1791, %f1788; neg.ftz.f32 %f1797, %f1796; sub.ftz.f32 %f1798, %f1789, %f1792; sub.ftz.f32 %f1799, %f1790, %f1794; sub.ftz.f32 %f1800, %f1791, %f1796; min.ftz.f32 %f1801, %f1793, %f1798; min.ftz.f32 %f1802, %f1795, %f1799; min.ftz.f32 %f1803, %f1797, %f1800; max.ftz.f32 %f1804, %f1793, %f1798; max.ftz.f32 %f1805, %f1795, %f1799; max.ftz.f32 %f1806, %f1797, %f1800; max.ftz.f32 %f1807, %f1738, %f1801; max.ftz.f32 %f1808, %f1802, %f1803; max.ftz.f32 %f2500, %f1807, %f1808; min.ftz.f32 %f1809, %f100, %f1804; min.ftz.f32 %f1810, %f1805, %f1806; min.ftz.f32 %f1811, %f1809, %f1810; setp.gtu.ftz.f32 %p166, %f2500, %f1811; @%p166 bra $L__BB155_121; and.b32 %r432, %r119, 2047; setp.eq.s32 %p167, %r432, 0; or.b32 %r433, %r582, 2; selp.b32 %r582, %r582, %r433, %p167; $L__BB155_121: mul.wide.s32 %rd344, %r583, 4; add.s64 %rd90, %rd22, %rd344; setp.eq.s32 %p174, %r582, 3; @%p174 bra $L__BB155_125; bra.uni $L__BB155_122; $L__BB155_125: setp.lt.ftz.f32 %p177, %f2500, %f2499; selp.b32 %r131, %r584, %r120, %p177; selp.b32 %r438, %r120, %r584, %p177; add.s32 %r583, %r583, 1; st.local.u32 [%rd90+4], %r438; mov.u32 %r584, %r131; bra.uni $L__BB155_126; $L__BB155_122: setp.eq.s32 %p175, %r582, 2; @%p175 bra $L__BB155_126; setp.ne.s32 %p176, %r582, 0; mov.u32 %r584, %r120; @%p176 bra $L__BB155_126; ld.local.u32 %r584, [%rd90]; add.s32 %r583, %r583, -1; $L__BB155_126: setp.ne.s32 %p178, %r584, 1985229328; setp.gt.s32 %p179, %r584, -1; and.pred %p180, %p179, %p178; @%p180 bra $L__BB155_117; $L__BB155_127: mov.u32 %r135, %r583; setp.gt.s32 %p181, %r584, -1; @%p181 bra $L__BB155_146; cvta.to.global.u64 %rd345, %rd251; mul.wide.s32 %rd346, %r584, 16; xor.b64 %rd347, %rd346, -16; add.s64 %rd348, %rd345, %rd347; ld.global.v4.u32 {%r587, %r440, %r441, %r442}, [%rd348]; mul.wide.s32 %rd351, %r135, 4; add.s64 %rd352, %rd22, %rd351; ld.local.u32 %r584, [%rd352]; add.s32 %r583, %r135, -1; and.b32 %r444, %r442, 127; setp.ne.s32 %p182, %r444, 1; setp.ge.s32 %p183, %r587, %r440; or.pred %p184, %p182, %p183; @%p184 bra $L__BB155_146; $L__BB155_129: cvt.s64.s32 %rd91, %r587; setp.ne.s32 %p185, %r576, -1; @%p185 bra $L__BB155_131; cvta.to.global.u64 %rd353, %rd253; shl.b64 %rd354, %rd91, 2; add.s64 %rd355, %rd353, %rd354; ld.global.u32 %r445, [%rd355]; setp.ne.s32 %p186, %r445, %r3; @%p186 bra $L__BB155_145; $L__BB155_131: cvta.to.global.u64 %rd356, %rd252; shl.b64 %rd357, %rd91, 2; add.s64 %rd358, %rd356, %rd357; ld.global.u32 %r143, [%rd358]; setp.eq.s32 %p187, %r143, -1; @%p187 bra $L__BB155_145; cvta.to.global.u64 %rd359, %rd241; mul.wide.s32 %rd360, %r143, 16; add.s64 %rd361, %rd359, %rd360; ld.global.u32 %r446, [%rd361+12]; mul.wide.u32 %rd362, %r446, 12; add.s64 %rd363, %rd7, %rd362; add.s32 %r447, %r446, 1; mul.wide.u32 %rd364, %r447, 12; add.s64 %rd365, %rd7, %rd364; add.s32 %r448, %r446, 2; mul.wide.u32 %rd366, %r448, 12; add.s64 %rd367, %rd7, %rd366; ld.global.f32 %f358, [%rd363]; sub.ftz.f32 %f359, %f358, %f2450; ld.global.f32 %f360, [%rd363+4]; sub.ftz.f32 %f361, %f360, %f2451; ld.global.f32 %f362, [%rd363+8]; sub.ftz.f32 %f363, %f362, %f2452; ld.global.f32 %f364, [%rd365]; sub.ftz.f32 %f1883, %f364, %f2450; ld.global.f32 %f365, [%rd365+4]; sub.ftz.f32 %f1884, %f365, %f2451; ld.global.f32 %f366, [%rd365+8]; sub.ftz.f32 %f1885, %f366, %f2452; ld.global.f32 %f367, [%rd367]; sub.ftz.f32 %f1886, %f367, %f2450; ld.global.f32 %f368, [%rd367+4]; sub.ftz.f32 %f1887, %f368, %f2451; ld.global.f32 %f369, [%rd367+8]; sub.ftz.f32 %f1888, %f369, %f2452; sub.ftz.f32 %f370, %f1886, %f359; sub.ftz.f32 %f371, %f1887, %f361; sub.ftz.f32 %f372, %f1888, %f363; sub.ftz.f32 %f373, %f359, %f1883; sub.ftz.f32 %f374, %f361, %f1884; sub.ftz.f32 %f375, %f363, %f1885; sub.ftz.f32 %f1889, %f1883, %f1886; sub.ftz.f32 %f1890, %f1884, %f1887; sub.ftz.f32 %f1891, %f1885, %f1888; add.ftz.f32 %f1892, %f359, %f1886; add.ftz.f32 %f1893, %f361, %f1887; add.ftz.f32 %f1894, %f363, %f1888; mul.ftz.f32 %f1895, %f371, %f1894; mul.ftz.f32 %f1896, %f1893, %f372; sub.ftz.f32 %f1897, %f1895, %f1896; mul.ftz.f32 %f1898, %f1892, %f372; mul.ftz.f32 %f1899, %f370, %f1894; sub.ftz.f32 %f1900, %f1898, %f1899; mul.ftz.f32 %f1901, %f370, %f1893; mul.ftz.f32 %f1902, %f1892, %f371; sub.ftz.f32 %f1903, %f1901, %f1902; mul.ftz.f32 %f1904, %f2448, %f1900; fma.rn.ftz.f32 %f1905, %f2447, %f1897, %f1904; fma.rn.ftz.f32 %f376, %f2449, %f1903, %f1905; add.ftz.f32 %f1906, %f359, %f1883; add.ftz.f32 %f1907, %f361, %f1884; add.ftz.f32 %f1908, %f363, %f1885; mul.ftz.f32 %f1909, %f374, %f1908; mul.ftz.f32 %f1910, %f1907, %f375; sub.ftz.f32 %f1911, %f1909, %f1910; mul.ftz.f32 %f1912, %f1906, %f375; mul.ftz.f32 %f1913, %f373, %f1908; sub.ftz.f32 %f1914, %f1912, %f1913; mul.ftz.f32 %f1915, %f373, %f1907; mul.ftz.f32 %f1916, %f1906, %f374; sub.ftz.f32 %f1917, %f1915, %f1916; mul.ftz.f32 %f1918, %f2448, %f1914; fma.rn.ftz.f32 %f1919, %f2447, %f1911, %f1918; fma.rn.ftz.f32 %f377, %f2449, %f1917, %f1919; add.ftz.f32 %f1920, %f1883, %f1886; add.ftz.f32 %f1921, %f1884, %f1887; add.ftz.f32 %f1922, %f1885, %f1888; mul.ftz.f32 %f1923, %f1890, %f1922; mul.ftz.f32 %f1924, %f1921, %f1891; sub.ftz.f32 %f1925, %f1923, %f1924; mul.ftz.f32 %f1926, %f1920, %f1891; mul.ftz.f32 %f1927, %f1889, %f1922; sub.ftz.f32 %f1928, %f1926, %f1927; mul.ftz.f32 %f1929, %f1889, %f1921; mul.ftz.f32 %f1930, %f1920, %f1890; sub.ftz.f32 %f1931, %f1929, %f1930; mul.ftz.f32 %f1932, %f2448, %f1928; fma.rn.ftz.f32 %f1933, %f2447, %f1925, %f1932; fma.rn.ftz.f32 %f1934, %f2449, %f1931, %f1933; add.ftz.f32 %f1935, %f377, %f376; add.ftz.f32 %f378, %f1934, %f1935; abs.ftz.f32 %f379, %f378; mul.ftz.f32 %f1936, %f379, 0f34000000; min.ftz.f32 %f1937, %f377, %f1934; min.ftz.f32 %f1938, %f376, %f1937; max.ftz.f32 %f1939, %f377, %f1934; max.ftz.f32 %f1940, %f376, %f1939; mul.ftz.f32 %f1941, %f379, 0fB4000000; setp.ltu.ftz.f32 %p189, %f1938, %f1941; setp.gtu.ftz.f32 %p190, %f1940, %f1936; and.pred %p191, %p189, %p190; mov.pred %p274, -1; @%p191 bra $L__BB155_138; mov.pred %p274, -1; mul.ftz.f32 %f1942, %f374, %f372; mul.ftz.f32 %f1943, %f375, %f371; sub.ftz.f32 %f1944, %f1942, %f1943; mul.ftz.f32 %f1945, %f373, %f372; mul.ftz.f32 %f1946, %f375, %f370; sub.ftz.f32 %f1947, %f1946, %f1945; mul.ftz.f32 %f1948, %f374, %f370; mul.ftz.f32 %f1949, %f373, %f371; sub.ftz.f32 %f1950, %f1949, %f1948; add.ftz.f32 %f380, %f1944, %f1944; add.ftz.f32 %f381, %f1947, %f1947; add.ftz.f32 %f382, %f1950, %f1950; mul.ftz.f32 %f1951, %f2448, %f381; fma.rn.ftz.f32 %f1952, %f2447, %f380, %f1951; fma.rn.ftz.f32 %f383, %f2449, %f382, %f1952; setp.eq.ftz.f32 %p193, %f383, 0f00000000; @%p193 bra $L__BB155_138; mov.pred %p274, -1; sub.ftz.f32 %f2384, %f362, %f2452; sub.ftz.f32 %f2383, %f358, %f2450; sub.ftz.f32 %f2382, %f360, %f2451; mul.ftz.f32 %f1953, %f2382, %f381; fma.rn.ftz.f32 %f1954, %f2383, %f380, %f1953; fma.rn.ftz.f32 %f1955, %f2384, %f382, %f1954; div.approx.ftz.f32 %f384, %f1955, %f383; setp.ltu.ftz.f32 %p195, %f384, 0f00000000; setp.gtu.ftz.f32 %p196, %f384, %f100; or.pred %p197, %p196, %p195; @%p197 bra $L__BB155_138; setp.lt.ftz.f32 %p198, %f379, 0f219392EF; mov.f32 %f2504, 0f00000000; @%p198 bra $L__BB155_137; rcp.approx.ftz.f32 %f2504, %f378; $L__BB155_137: mul.ftz.f32 %f1957, %f376, %f2504; min.ftz.f32 %f2509, %f1957, %f716; mul.ftz.f32 %f1959, %f377, %f2504; min.ftz.f32 %f2508, %f1959, %f716; mov.pred %p274, 0; mov.f32 %f2510, %f384; $L__BB155_138: @%p274 bra $L__BB155_145; ld.local.u32 %r591, [%rd1]; min.s32 %r589, %r591, 4; $L__BB155_140: add.s32 %r147, %r589, -1; setp.gt.s32 %p200, %r589, 0; @%p200 bra $L__BB155_144; bra.uni $L__BB155_141; $L__BB155_144: mul.wide.s32 %rd372, %r147, 24; add.s64 %rd373, %rd1, %rd372; ld.local.f32 %f1982, [%rd373+4]; setp.neu.ftz.f32 %p203, %f1982, %f2510; mov.u32 %r589, %r147; @%p203 bra $L__BB155_140; bra.uni $L__BB155_145; $L__BB155_141: setp.lt.s32 %p201, %r591, 4; add.s32 %r148, %r591, 1; st.local.u32 [%rd1], %r148; @%p201 bra $L__BB155_143; mad.lo.s32 %r573, %r573, 1103515245, 12345; rem.u32 %r591, %r573, %r148; setp.gt.s32 %p202, %r591, 3; @%p202 bra $L__BB155_145; $L__BB155_143: mul.wide.s32 %rd368, %r591, 24; add.s64 %rd369, %rd1, %rd368; st.local.u32 [%rd369+16], %r143; st.local.u32 [%rd369+20], %r3; mov.u32 %r449, 1; st.local.u32 [%rd369+24], %r449; st.local.f32 [%rd369+8], %f2509; st.local.f32 [%rd369+12], %f2508; st.local.f32 [%rd369+4], %f2510; sub.ftz.f32 %f1960, %f369, %f362; sub.ftz.f32 %f1961, %f365, %f360; mul.ftz.f32 %f1962, %f1961, %f1960; sub.ftz.f32 %f1963, %f368, %f360; sub.ftz.f32 %f1964, %f366, %f362; mul.ftz.f32 %f1965, %f1964, %f1963; sub.ftz.f32 %f1966, %f1962, %f1965; sub.ftz.f32 %f1967, %f367, %f358; mul.ftz.f32 %f1968, %f1964, %f1967; sub.ftz.f32 %f1969, %f364, %f358; mul.ftz.f32 %f1970, %f1969, %f1960; sub.ftz.f32 %f1971, %f1968, %f1970; mul.ftz.f32 %f1972, %f1969, %f1963; mul.ftz.f32 %f1973, %f1961, %f1967; sub.ftz.f32 %f1974, %f1972, %f1973; mul.ftz.f32 %f1975, %f1971, %f1971; fma.rn.ftz.f32 %f1976, %f1966, %f1966, %f1975; fma.rn.ftz.f32 %f1977, %f1974, %f1974, %f1976; sqrt.approx.ftz.f32 %f1978, %f1977; div.approx.ftz.f32 %f1979, %f1966, %f1978; div.approx.ftz.f32 %f1980, %f1971, %f1978; div.approx.ftz.f32 %f1981, %f1974, %f1978; mul.wide.s32 %rd370, %r591, 12; add.s64 %rd371, %rd1, %rd370; st.local.f32 [%rd371+100], %f1979; st.local.f32 [%rd371+104], %f1980; st.local.f32 [%rd371+108], %f1981; $L__BB155_145: add.s32 %r583, %r135, -1; cvt.u32.u64 %r450, %rd91; add.s32 %r587, %r450, 1; setp.lt.s32 %p204, %r587, %r440; @%p204 bra $L__BB155_129; $L__BB155_146: setp.ne.s32 %p205, %r584, 1985229328; @%p205 bra $L__BB155_115; bra.uni $L__BB155_148; $L__BB155_41: mul.wide.s32 %rd264, %r3, 208; add.s64 %rd265, %rd11, %rd264; ld.global.u32 %rd266, [%rd265+168]; ld.global.u32 %r353, [%rd265+152]; shl.b32 %r354, %r353, 1; cvt.rn.f32.s32 %f913, %r354; mul.ftz.f32 %f914, %f101, %f913; cvt.rzi.ftz.s32.f32 %r355, %f914; add.s32 %r356, %r354, -1; min.s32 %r357, %r355, %r356; cvt.rn.f32.s32 %f915, %r357; sub.ftz.f32 %f108, %f914, %f915; cvt.s64.s32 %rd267, %r357; add.s64 %rd268, %rd267, %rd266; shl.b64 %rd270, %rd268, 6; add.s64 %rd23, %rd269, %rd270; ld.global.v4.f32 {%f916, %f917, %f918, %f919}, [%rd23]; ld.global.v4.f32 {%f920, %f921, %f922, %f923}, [%rd23+64]; mul.ftz.f32 %f924, %f917, %f921; fma.rn.ftz.f32 %f925, %f916, %f920, %f924; mul.ftz.f32 %f926, %f919, %f923; fma.rn.ftz.f32 %f927, %f918, %f922, %f926; add.ftz.f32 %f117, %f925, %f927; setp.gt.ftz.f32 %p70, %f117, 0f3F7FDF3B; sub.ftz.f32 %f118, %f716, %f108; @%p70 bra $L__BB155_43; bra.uni $L__BB155_42; $L__BB155_43: mul.ftz.f32 %f982, %f916, %f118; fma.rn.ftz.f32 %f983, %f920, %f108, %f982; mul.ftz.f32 %f984, %f917, %f118; fma.rn.ftz.f32 %f985, %f921, %f108, %f984; mul.ftz.f32 %f986, %f918, %f118; fma.rn.ftz.f32 %f987, %f922, %f108, %f986; mul.ftz.f32 %f988, %f919, %f118; fma.rn.ftz.f32 %f989, %f923, %f108, %f988; mul.ftz.f32 %f990, %f985, %f985; fma.rn.ftz.f32 %f991, %f983, %f983, %f990; mul.ftz.f32 %f992, %f989, %f989; fma.rn.ftz.f32 %f993, %f987, %f987, %f992; add.ftz.f32 %f994, %f991, %f993; rsqrt.approx.ftz.f32 %f995, %f994; mul.ftz.f32 %f2424, %f989, %f995; mul.ftz.f32 %f2423, %f987, %f995; mul.ftz.f32 %f2422, %f985, %f995; mul.ftz.f32 %f2421, %f983, %f995; bra.uni $L__BB155_44; $L__BB155_42: mov.f32 %f929, 0fBF800000; max.ftz.f32 %f930, %f117, %f929; min.ftz.f32 %f932, %f930, %f716; abs.ftz.f32 %f933, %f932; sub.ftz.f32 %f934, %f716, %f933; mul.ftz.f32 %f935, %f934, 0f3F000000; sqrt.approx.ftz.f32 %f936, %f935; setp.gt.ftz.f32 %p71, %f933, 0f3F11EB85; selp.f32 %f937, %f936, %f933, %p71; mul.ftz.f32 %f938, %f937, %f937; mov.f32 %f939, 0f3C94D2E9; mov.f32 %f940, 0f3D53F941; fma.rn.ftz.f32 %f941, %f940, %f938, %f939; mov.f32 %f942, 0f3D3F841F; fma.rn.ftz.f32 %f943, %f941, %f938, %f942; mov.f32 %f944, 0f3D994929; fma.rn.ftz.f32 %f945, %f943, %f938, %f944; mov.f32 %f946, 0f3E2AAB94; fma.rn.ftz.f32 %f947, %f945, %f938, %f946; mul.ftz.f32 %f948, %f938, %f947; fma.rn.ftz.f32 %f949, %f948, %f937, %f937; add.ftz.f32 %f950, %f949, %f949; mov.f32 %f951, 0f3FC90FDB; sub.ftz.f32 %f952, %f951, %f949; selp.f32 %f953, %f950, %f952, %p71; setp.lt.ftz.f32 %p72, %f932, 0f00000000; mov.f32 %f954, 0f40490FDB; sub.ftz.f32 %f955, %f954, %f953; selp.f32 %f956, %f955, %f953, %p72; mul.ftz.f32 %f957, %f916, %f117; sub.ftz.f32 %f958, %f920, %f957; mul.ftz.f32 %f959, %f917, %f117; sub.ftz.f32 %f960, %f921, %f959; mul.ftz.f32 %f961, %f918, %f117; sub.ftz.f32 %f962, %f922, %f961; mul.ftz.f32 %f963, %f919, %f117; sub.ftz.f32 %f964, %f923, %f963; mul.ftz.f32 %f965, %f960, %f960; fma.rn.ftz.f32 %f966, %f958, %f958, %f965; mul.ftz.f32 %f967, %f964, %f964; fma.rn.ftz.f32 %f968, %f962, %f962, %f967; add.ftz.f32 %f969, %f966, %f968; rsqrt.approx.ftz.f32 %f970, %f969; mul.ftz.f32 %f971, %f958, %f970; mul.ftz.f32 %f972, %f960, %f970; mul.ftz.f32 %f973, %f962, %f970; mul.ftz.f32 %f974, %f964, %f970; mul.ftz.f32 %f975, %f108, %f956; cos.approx.ftz.f32 %f976, %f975; sin.approx.ftz.f32 %f977, %f975; mul.ftz.f32 %f978, %f971, %f977; mul.ftz.f32 %f979, %f972, %f977; mul.ftz.f32 %f980, %f973, %f977; mul.ftz.f32 %f981, %f974, %f977; fma.rn.ftz.f32 %f2424, %f919, %f976, %f981; fma.rn.ftz.f32 %f2423, %f918, %f976, %f980; fma.rn.ftz.f32 %f2422, %f917, %f976, %f979; fma.rn.ftz.f32 %f2421, %f916, %f976, %f978; $L__BB155_44: ld.global.v4.f32 {%f996, %f997, %f998, %f999}, [%rd23+16]; ld.global.v4.f32 {%f1004, %f1005, %f1006, %f1007}, [%rd23+80]; mul.ftz.f32 %f1012, %f108, %f1004; mul.ftz.f32 %f1013, %f108, %f1005; mul.ftz.f32 %f1014, %f108, %f1006; mul.ftz.f32 %f1015, %f108, %f1007; fma.rn.ftz.f32 %f131, %f118, %f996, %f1012; fma.rn.ftz.f32 %f132, %f118, %f997, %f1013; fma.rn.ftz.f32 %f133, %f118, %f998, %f1014; fma.rn.ftz.f32 %f1016, %f118, %f999, %f1015; ld.global.v4.f32 {%f1017, %f1018, %f1019, %f1020}, [%rd23+32]; ld.global.v4.f32 {%f1025, %f1026, %f1027, %f1028}, [%rd23+96]; mul.ftz.f32 %f1033, %f108, %f1025; mul.ftz.f32 %f1034, %f108, %f1026; mul.ftz.f32 %f1035, %f108, %f1027; mul.ftz.f32 %f1036, %f108, %f1028; fma.rn.ftz.f32 %f1037, %f118, %f1017, %f1033; fma.rn.ftz.f32 %f1038, %f118, %f1018, %f1034; fma.rn.ftz.f32 %f1039, %f118, %f1019, %f1035; fma.rn.ftz.f32 %f1040, %f118, %f1020, %f1036; ld.global.v4.f32 {%f1041, %f1042, %f1043, %f1044}, [%rd23+48]; ld.global.v4.f32 {%f1049, %f1050, %f1051, %f1052}, [%rd23+112]; mul.ftz.f32 %f1057, %f108, %f1049; mul.ftz.f32 %f1058, %f108, %f1050; mul.ftz.f32 %f1059, %f108, %f1051; mul.ftz.f32 %f1060, %f108, %f1052; fma.rn.ftz.f32 %f1061, %f118, %f1041, %f1057; fma.rn.ftz.f32 %f1062, %f118, %f1042, %f1058; fma.rn.ftz.f32 %f1063, %f118, %f1043, %f1059; fma.rn.ftz.f32 %f1064, %f118, %f1044, %f1060; mul.ftz.f32 %f1065, %f2424, 0f3FB504F3; mul.ftz.f32 %f1066, %f2421, 0f3FB504F3; mul.ftz.f32 %f1067, %f2422, 0f3FB504F3; mul.ftz.f32 %f1068, %f2423, 0f3FB504F3; mul.ftz.f32 %f1069, %f1065, %f1066; mul.ftz.f32 %f1070, %f1065, %f1067; mul.ftz.f32 %f1071, %f1065, %f1068; mul.ftz.f32 %f1072, %f1066, %f1066; mul.ftz.f32 %f1073, %f1066, %f1067; mul.ftz.f32 %f1074, %f1066, %f1068; mul.ftz.f32 %f1075, %f1067, %f1067; mul.ftz.f32 %f1076, %f1067, %f1068; mul.ftz.f32 %f1077, %f1068, %f1068; sub.ftz.f32 %f1079, %f716, %f1075; sub.ftz.f32 %f1080, %f1079, %f1077; sub.ftz.f32 %f1081, %f1073, %f1071; add.ftz.f32 %f1082, %f1070, %f1074; add.ftz.f32 %f1083, %f1071, %f1073; sub.ftz.f32 %f1084, %f716, %f1072; sub.ftz.f32 %f1085, %f1084, %f1077; sub.ftz.f32 %f1086, %f1076, %f1069; sub.ftz.f32 %f1087, %f1074, %f1070; add.ftz.f32 %f1088, %f1069, %f1076; sub.ftz.f32 %f1089, %f1084, %f1075; mul.ftz.f32 %f1090, %f1081, %f1039; fma.rn.ftz.f32 %f1091, %f1080, %f1016, %f1090; fma.rn.ftz.f32 %f2430, %f1082, %f1062, %f1091; mul.ftz.f32 %f1092, %f1081, %f1040; fma.rn.ftz.f32 %f1093, %f1080, %f1037, %f1092; fma.rn.ftz.f32 %f135, %f1082, %f1063, %f1093; mul.ftz.f32 %f1094, %f1081, %f1061; fma.rn.ftz.f32 %f1095, %f1080, %f1038, %f1094; fma.rn.ftz.f32 %f136, %f1082, %f1064, %f1095; mul.ftz.f32 %f1096, %f1085, %f1039; fma.rn.ftz.f32 %f1097, %f1083, %f1016, %f1096; fma.rn.ftz.f32 %f137, %f1086, %f1062, %f1097; mul.ftz.f32 %f1098, %f1085, %f1040; fma.rn.ftz.f32 %f1099, %f1083, %f1037, %f1098; fma.rn.ftz.f32 %f2429, %f1086, %f1063, %f1099; mul.ftz.f32 %f1100, %f1085, %f1061; fma.rn.ftz.f32 %f1101, %f1083, %f1038, %f1100; fma.rn.ftz.f32 %f139, %f1086, %f1064, %f1101; mul.ftz.f32 %f1102, %f1088, %f1039; fma.rn.ftz.f32 %f1103, %f1087, %f1016, %f1102; fma.rn.ftz.f32 %f140, %f1089, %f1062, %f1103; mul.ftz.f32 %f1104, %f1088, %f1040; fma.rn.ftz.f32 %f1105, %f1087, %f1037, %f1104; fma.rn.ftz.f32 %f141, %f1089, %f1063, %f1105; mul.ftz.f32 %f1106, %f1088, %f1061; fma.rn.ftz.f32 %f1107, %f1087, %f1038, %f1106; fma.rn.ftz.f32 %f2428, %f1089, %f1064, %f1107; mul.ftz.f32 %f1108, %f2429, %f2428; mul.ftz.f32 %f143, %f141, %f139; sub.ftz.f32 %f2427, %f1108, %f143; mul.ftz.f32 %f145, %f141, %f136; mul.ftz.f32 %f1109, %f135, %f2428; sub.ftz.f32 %f2426, %f145, %f1109; mul.ftz.f32 %f147, %f135, %f139; mul.ftz.f32 %f1110, %f2429, %f136; sub.ftz.f32 %f2425, %f147, %f1110; mul.ftz.f32 %f1111, %f137, %f2426; fma.rn.ftz.f32 %f1112, %f2430, %f2427, %f1111; fma.rn.ftz.f32 %f2431, %f140, %f2425, %f1112; setp.neu.ftz.f32 %p73, %f2431, 0f00000000; @%p73 bra $L__BB155_46; add.ftz.f32 %f2430, %f2430, 0f322BCC77; add.ftz.f32 %f2428, %f2428, 0f322BCC77; add.ftz.f32 %f2429, %f2429, 0f322BCC77; mul.ftz.f32 %f1113, %f2429, %f2428; sub.ftz.f32 %f2427, %f1113, %f143; mul.ftz.f32 %f1114, %f135, %f2428; sub.ftz.f32 %f2426, %f145, %f1114; mul.ftz.f32 %f1115, %f2429, %f136; sub.ftz.f32 %f2425, %f147, %f1115; mul.ftz.f32 %f1116, %f137, %f2426; fma.rn.ftz.f32 %f1117, %f2430, %f2427, %f1116; fma.rn.ftz.f32 %f1118, %f140, %f2425, %f1117; setp.eq.ftz.f32 %p74, %f1118, 0f00000000; selp.f32 %f2431, 0f7F7FFFFF, %f1118, %p74; $L__BB155_46: mul.ftz.f32 %f1119, %f137, %f2428; mul.ftz.f32 %f1120, %f140, %f139; sub.ftz.f32 %f1121, %f1120, %f1119; mul.ftz.f32 %f1122, %f140, %f136; mul.ftz.f32 %f1123, %f2428, %f2430; sub.ftz.f32 %f1124, %f1123, %f1122; mul.ftz.f32 %f1125, %f139, %f2430; mul.ftz.f32 %f1126, %f137, %f136; sub.ftz.f32 %f1127, %f1126, %f1125; div.approx.ftz.f32 %f2436, %f1121, %f2431; div.approx.ftz.f32 %f2437, %f1124, %f2431; div.approx.ftz.f32 %f2438, %f1127, %f2431; mul.ftz.f32 %f1128, %f140, %f2429; mul.ftz.f32 %f1129, %f137, %f141; sub.ftz.f32 %f1130, %f1129, %f1128; mul.ftz.f32 %f1131, %f141, %f2430; mul.ftz.f32 %f1132, %f140, %f135; sub.ftz.f32 %f1133, %f1132, %f1131; mul.ftz.f32 %f1134, %f137, %f135; mul.ftz.f32 %f1135, %f2429, %f2430; sub.ftz.f32 %f1136, %f1135, %f1134; div.approx.ftz.f32 %f2440, %f1130, %f2431; div.approx.ftz.f32 %f2441, %f1133, %f2431; div.approx.ftz.f32 %f2442, %f1136, %f2431; div.approx.ftz.f32 %f2432, %f2427, %f2431; div.approx.ftz.f32 %f2433, %f2426, %f2431; mul.ftz.f32 %f1137, %f132, %f2433; fma.rn.ftz.f32 %f1138, %f131, %f2432, %f1137; div.approx.ftz.f32 %f2434, %f2425, %f2431; fma.rn.ftz.f32 %f1139, %f133, %f2434, %f1138; neg.ftz.f32 %f2435, %f1139; mul.ftz.f32 %f1140, %f132, %f2437; fma.rn.ftz.f32 %f1141, %f131, %f2436, %f1140; fma.rn.ftz.f32 %f1142, %f133, %f2438, %f1141; neg.ftz.f32 %f2439, %f1142; mul.ftz.f32 %f1143, %f132, %f2441; fma.rn.ftz.f32 %f1144, %f131, %f2440, %f1143; fma.rn.ftz.f32 %f1145, %f133, %f2442, %f1144; neg.ftz.f32 %f2443, %f1145; $L__BB155_48: neg.ftz.f32 %f2387, %f2408; neg.ftz.f32 %f2386, %f2409; neg.ftz.f32 %f2385, %f2410; mul.ftz.f32 %f1158, %f2451, %f2433; fma.rn.ftz.f32 %f1159, %f2450, %f2432, %f1158; fma.rn.ftz.f32 %f1160, %f2452, %f2434, %f1159; add.ftz.f32 %f200, %f2435, %f1160; mul.ftz.f32 %f1161, %f2451, %f2437; fma.rn.ftz.f32 %f1162, %f2450, %f2436, %f1161; fma.rn.ftz.f32 %f1163, %f2452, %f2438, %f1162; add.ftz.f32 %f201, %f2439, %f1163; mul.ftz.f32 %f1164, %f2451, %f2441; fma.rn.ftz.f32 %f1165, %f2450, %f2440, %f1164; fma.rn.ftz.f32 %f1166, %f2452, %f2442, %f1165; add.ftz.f32 %f2452, %f2443, %f1166; mul.ftz.f32 %f1167, %f2433, %f2386; fma.rn.ftz.f32 %f1168, %f2432, %f2385, %f1167; fma.rn.ftz.f32 %f1169, %f2434, %f2387, %f1168; mul.ftz.f32 %f1170, %f2437, %f2386; fma.rn.ftz.f32 %f1171, %f2436, %f2385, %f1170; fma.rn.ftz.f32 %f1172, %f2438, %f2387, %f1171; mul.ftz.f32 %f1173, %f2441, %f2386; fma.rn.ftz.f32 %f1174, %f2440, %f2385, %f1173; fma.rn.ftz.f32 %f1175, %f2442, %f2387, %f1174; abs.ftz.f32 %f1176, %f1169; setp.gt.ftz.f32 %p75, %f1176, 0f17800000; mov.b32 %r358, %f1169; and.b32 %r359, %r358, -2147483648; or.b32 %r360, %r359, 394264576; mov.b32 %f1177, %r360; selp.f32 %f2447, %f1169, %f1177, %p75; abs.ftz.f32 %f1178, %f1172; setp.gt.ftz.f32 %p76, %f1178, 0f17800000; mov.b32 %r361, %f1172; and.b32 %r362, %r361, -2147483648; or.b32 %r363, %r362, 394264576; mov.b32 %f1179, %r363; selp.f32 %f2448, %f1172, %f1179, %p76; abs.ftz.f32 %f1180, %f1175; setp.gt.ftz.f32 %p77, %f1180, 0f17800000; mov.b32 %r364, %f1175; and.b32 %r365, %r364, -2147483648; or.b32 %r366, %r365, 394264576; mov.b32 %f1181, %r366; selp.f32 %f2449, %f1175, %f1181, %p77; rcp.approx.ftz.f32 %f2444, %f2447; rcp.approx.ftz.f32 %f2445, %f2448; rcp.approx.ftz.f32 %f2446, %f2449; mov.f32 %f2450, %f200; mov.f32 %f2451, %f201; mov.u32 %r545, %r3; bra.uni $L__BB155_49; $L__BB155_62: cvta.to.global.u64 %rd277, %rd251; mul.wide.s32 %rd278, %r553, 16; xor.b64 %rd279, %rd278, -16; add.s64 %rd280, %rd277, %rd279; ld.global.v4.u32 {%r556, %r379, %r380, %r381}, [%rd280]; mul.wide.s32 %rd283, %r52, 4; add.s64 %rd284, %rd22, %rd283; ld.local.u32 %r553, [%rd284]; cvt.u16.u32 %rs29, %r381; and.b16 %rs28, %rs29, 127; add.s32 %r574, %r52, -1; setp.eq.s16 %p101, %rs28, 1; @%p101 bra $L__BB155_94; setp.ne.s16 %p102, %rs28, 65; @%p102 bra $L__BB155_112; add.s32 %r574, %r52, -1; setp.ge.s32 %p103, %r556, %r379; @%p103 bra $L__BB155_112; $L__BB155_65: cvt.s64.s32 %rd28, %r556; setp.ne.s32 %p104, %r545, -1; @%p104 bra $L__BB155_67; cvta.to.global.u64 %rd285, %rd253; shl.b64 %rd286, %rd28, 2; add.s64 %rd287, %rd285, %rd286; ld.global.u32 %r383, [%rd287]; setp.ne.s32 %p105, %r383, %r3; @%p105 bra $L__BB155_93; $L__BB155_67: cvta.to.global.u64 %rd288, %rd252; shl.b64 %rd289, %rd28, 2; add.s64 %rd290, %rd288, %rd289; ld.global.u32 %r61, [%rd290]; setp.eq.s32 %p106, %r61, -1; @%p106 bra $L__BB155_93; mul.wide.s32 %rd291, %r3, 208; add.s64 %rd292, %rd11, %rd291; ld.global.u32 %r62, [%rd292+156]; ld.global.u32 %r63, [%rd292+152]; shl.b32 %r384, %r63, 1; cvt.rn.f32.s32 %f1397, %r384; mul.ftz.f32 %f234, %f101, %f1397; cvt.rzi.ftz.s32.f32 %r385, %f234; add.s32 %r386, %r384, -1; min.s32 %r64, %r385, %r386; ld.global.u32 %r559, [%rd292+164]; mul.wide.u32 %rd293, %r559, 16; add.s64 %rd294, %rd12, %rd293; ld.global.u16 %rs35, [%rd294+12]; ld.global.u32 %r558, [%rd294+8]; ld.global.u64 %rd469, [%rd294]; setp.eq.s64 %p107, %rd469, 11; @%p107 bra $L__BB155_73; $L__BB155_69: setp.eq.s64 %p108, %rd469, 0; @%p108 bra $L__BB155_71; add.s32 %r559, %r559, 2; bra.uni $L__BB155_72; $L__BB155_71: setp.eq.s16 %p109, %rs35, 0; mov.u32 %r562, -1; mov.u32 %r559, %r558; @%p109 bra $L__BB155_74; $L__BB155_72: mul.wide.u32 %rd295, %r559, 16; add.s64 %rd296, %rd12, %rd295; ld.global.u16 %rs35, [%rd296+12]; ld.global.u32 %r558, [%rd296+8]; ld.global.u64 %rd469, [%rd296]; setp.ne.s64 %p110, %rd469, 11; @%p110 bra $L__BB155_69; $L__BB155_73: setp.eq.s16 %p111, %rs35, 0; selp.b32 %r562, -1, %r558, %p111; $L__BB155_74: cvta.to.global.u64 %rd297, %rd241; mul.wide.s32 %rd298, %r61, 16; add.s64 %rd299, %rd297, %rd298; ld.global.v4.u32 {%r388, %r389, %r390, %r391}, [%rd299]; setp.eq.s32 %p112, %r64, %r63; @%p112 bra $L__BB155_76; bra.uni $L__BB155_75; $L__BB155_76: mul.wide.u32 %rd303, %r391, 12; add.s64 %rd470, %rd7, %rd303; add.s32 %r398, %r391, 1; mul.wide.u32 %rd304, %r398, 12; add.s64 %rd473, %rd7, %rd304; add.s32 %r399, %r391, 2; mul.wide.u32 %rd305, %r399, 12; add.s64 %rd476, %rd7, %rd305; add.s64 %rd471, %rd470, 4; add.s64 %rd472, %rd470, 8; add.s64 %rd474, %rd473, 4; add.s64 %rd475, %rd473, 8; add.s64 %rd477, %rd476, 4; add.s64 %rd478, %rd476, 8; bra.uni $L__BB155_77; $L__BB155_75: setp.gt.s32 %p113, %r64, %r63; selp.b32 %r392, -1, 0, %p113; add.s32 %r393, %r64, %r392; mad.lo.s32 %r394, %r393, %r62, %r562; add.s32 %r395, %r388, %r394; mul.wide.u32 %rd300, %r395, 12; add.s64 %rd470, %rd13, %rd300; add.s32 %r396, %r389, %r394; mul.wide.u32 %rd301, %r396, 12; add.s64 %rd473, %rd13, %rd301; add.s32 %r397, %r390, %r394; mul.wide.u32 %rd302, %r397, 12; add.s64 %rd476, %rd13, %rd302; add.s64 %rd471, %rd470, 4; add.s64 %rd472, %rd470, 8; add.s64 %rd474, %rd473, 4; add.s64 %rd475, %rd473, 8; add.s64 %rd477, %rd476, 4; add.s64 %rd478, %rd476, 8; $L__BB155_77: ld.global.f32 %f235, [%rd478]; ld.global.f32 %f236, [%rd477]; ld.global.f32 %f237, [%rd476]; ld.global.f32 %f238, [%rd475]; ld.global.f32 %f239, [%rd474]; ld.global.f32 %f240, [%rd473]; ld.global.f32 %f241, [%rd472]; ld.global.f32 %f242, [%rd471]; ld.global.f32 %f243, [%rd470]; add.s32 %r79, %r64, 1; setp.eq.s32 %p114, %r79, %r63; @%p114 bra $L__BB155_79; bra.uni $L__BB155_78; $L__BB155_79: mul.wide.u32 %rd309, %r391, 12; add.s64 %rd479, %rd7, %rd309; add.s32 %r406, %r391, 1; mul.wide.u32 %rd310, %r406, 12; add.s64 %rd482, %rd7, %rd310; add.s32 %r407, %r391, 2; mul.wide.u32 %rd311, %r407, 12; add.s64 %rd485, %rd7, %rd311; add.s64 %rd480, %rd479, 4; add.s64 %rd481, %rd479, 8; add.s64 %rd483, %rd482, 4; add.s64 %rd484, %rd482, 8; add.s64 %rd486, %rd485, 4; add.s64 %rd487, %rd485, 8; bra.uni $L__BB155_80; $L__BB155_78: setp.ge.s32 %p115, %r64, %r63; selp.b32 %r400, -1, 0, %p115; add.s32 %r401, %r79, %r400; mad.lo.s32 %r402, %r401, %r62, %r562; add.s32 %r403, %r388, %r402; mul.wide.u32 %rd306, %r403, 12; add.s64 %rd479, %rd13, %rd306; add.s32 %r404, %r389, %r402; mul.wide.u32 %rd307, %r404, 12; add.s64 %rd482, %rd13, %rd307; add.s32 %r405, %r390, %r402; mul.wide.u32 %rd308, %r405, 12; add.s64 %rd485, %rd13, %rd308; add.s64 %rd480, %rd479, 4; add.s64 %rd481, %rd479, 8; add.s64 %rd483, %rd482, 4; add.s64 %rd484, %rd482, 8; add.s64 %rd486, %rd485, 4; add.s64 %rd487, %rd485, 8; $L__BB155_80: cvt.rn.f32.s32 %f1398, %r64; sub.ftz.f32 %f1399, %f234, %f1398; sub.ftz.f32 %f1401, %f716, %f1399; ld.global.f32 %f1402, [%rd479]; mul.ftz.f32 %f1403, %f1399, %f1402; ld.global.f32 %f1404, [%rd480]; mul.ftz.f32 %f1405, %f1399, %f1404; ld.global.f32 %f1406, [%rd481]; mul.ftz.f32 %f1407, %f1399, %f1406; fma.rn.ftz.f32 %f244, %f1401, %f243, %f1403; fma.rn.ftz.f32 %f245, %f1401, %f242, %f1405; fma.rn.ftz.f32 %f246, %f1401, %f241, %f1407; ld.global.f32 %f1408, [%rd482]; mul.ftz.f32 %f1409, %f1399, %f1408; ld.global.f32 %f1410, [%rd483]; mul.ftz.f32 %f1411, %f1399, %f1410; ld.global.f32 %f1412, [%rd484]; mul.ftz.f32 %f1413, %f1399, %f1412; fma.rn.ftz.f32 %f247, %f1401, %f240, %f1409; fma.rn.ftz.f32 %f248, %f1401, %f239, %f1411; fma.rn.ftz.f32 %f249, %f1401, %f238, %f1413; ld.global.f32 %f1414, [%rd485]; mul.ftz.f32 %f1415, %f1399, %f1414; ld.global.f32 %f1416, [%rd486]; mul.ftz.f32 %f1417, %f1399, %f1416; ld.global.f32 %f1418, [%rd487]; mul.ftz.f32 %f1419, %f1399, %f1418; fma.rn.ftz.f32 %f250, %f1401, %f237, %f1415; fma.rn.ftz.f32 %f251, %f1401, %f236, %f1417; fma.rn.ftz.f32 %f252, %f1401, %f235, %f1419; sub.ftz.f32 %f253, %f244, %f2450; sub.ftz.f32 %f254, %f245, %f2451; sub.ftz.f32 %f255, %f246, %f2452; sub.ftz.f32 %f1420, %f247, %f2450; sub.ftz.f32 %f1421, %f248, %f2451; sub.ftz.f32 %f1422, %f249, %f2452; sub.ftz.f32 %f1423, %f250, %f2450; sub.ftz.f32 %f1424, %f251, %f2451; sub.ftz.f32 %f1425, %f252, %f2452; sub.ftz.f32 %f256, %f1423, %f253; sub.ftz.f32 %f257, %f1424, %f254; sub.ftz.f32 %f258, %f1425, %f255; sub.ftz.f32 %f259, %f253, %f1420; sub.ftz.f32 %f260, %f254, %f1421; sub.ftz.f32 %f261, %f255, %f1422; sub.ftz.f32 %f1426, %f1420, %f1423; sub.ftz.f32 %f1427, %f1421, %f1424; sub.ftz.f32 %f1428, %f1422, %f1425; add.ftz.f32 %f1429, %f1423, %f253; add.ftz.f32 %f1430, %f1424, %f254; add.ftz.f32 %f1431, %f1425, %f255; mul.ftz.f32 %f1432, %f1431, %f257; mul.ftz.f32 %f1433, %f258, %f1430; sub.ftz.f32 %f1434, %f1432, %f1433; mul.ftz.f32 %f1435, %f258, %f1429; mul.ftz.f32 %f1436, %f1431, %f256; sub.ftz.f32 %f1437, %f1435, %f1436; mul.ftz.f32 %f1438, %f1430, %f256; mul.ftz.f32 %f1439, %f257, %f1429; sub.ftz.f32 %f1440, %f1438, %f1439; mul.ftz.f32 %f1441, %f2448, %f1437; fma.rn.ftz.f32 %f1442, %f2447, %f1434, %f1441; fma.rn.ftz.f32 %f262, %f2449, %f1440, %f1442; add.ftz.f32 %f1443, %f1420, %f253; add.ftz.f32 %f1444, %f1421, %f254; add.ftz.f32 %f1445, %f1422, %f255; mul.ftz.f32 %f1446, %f1445, %f260; mul.ftz.f32 %f1447, %f261, %f1444; sub.ftz.f32 %f1448, %f1446, %f1447; mul.ftz.f32 %f1449, %f261, %f1443; mul.ftz.f32 %f1450, %f1445, %f259; sub.ftz.f32 %f1451, %f1449, %f1450; mul.ftz.f32 %f1452, %f1444, %f259; mul.ftz.f32 %f1453, %f260, %f1443; sub.ftz.f32 %f1454, %f1452, %f1453; mul.ftz.f32 %f1455, %f2448, %f1451; fma.rn.ftz.f32 %f1456, %f2447, %f1448, %f1455; fma.rn.ftz.f32 %f263, %f2449, %f1454, %f1456; add.ftz.f32 %f1457, %f1423, %f1420; add.ftz.f32 %f1458, %f1424, %f1421; add.ftz.f32 %f1459, %f1425, %f1422; mul.ftz.f32 %f1460, %f1459, %f1427; mul.ftz.f32 %f1461, %f1428, %f1458; sub.ftz.f32 %f1462, %f1460, %f1461; mul.ftz.f32 %f1463, %f1428, %f1457; mul.ftz.f32 %f1464, %f1459, %f1426; sub.ftz.f32 %f1465, %f1463, %f1464; mul.ftz.f32 %f1466, %f1458, %f1426; mul.ftz.f32 %f1467, %f1427, %f1457; sub.ftz.f32 %f1468, %f1466, %f1467; mul.ftz.f32 %f1469, %f2448, %f1465; fma.rn.ftz.f32 %f1470, %f2447, %f1462, %f1469; fma.rn.ftz.f32 %f1471, %f2449, %f1468, %f1470; add.ftz.f32 %f1472, %f262, %f263; add.ftz.f32 %f264, %f1471, %f1472; abs.ftz.f32 %f265, %f264; mul.ftz.f32 %f1473, %f265, 0f34000000; min.ftz.f32 %f1474, %f263, %f1471; min.ftz.f32 %f1475, %f262, %f1474; max.ftz.f32 %f1476, %f263, %f1471; max.ftz.f32 %f1477, %f262, %f1476; mul.ftz.f32 %f1478, %f265, 0fB4000000; setp.ltu.ftz.f32 %p117, %f1475, %f1478; setp.gtu.ftz.f32 %p118, %f1477, %f1473; and.pred %p119, %p117, %p118; mov.pred %p272, -1; @%p119 bra $L__BB155_86; mov.pred %p272, -1; mul.ftz.f32 %f1479, %f258, %f260; mul.ftz.f32 %f1480, %f261, %f257; sub.ftz.f32 %f1481, %f1479, %f1480; mul.ftz.f32 %f1482, %f258, %f259; mul.ftz.f32 %f1483, %f261, %f256; sub.ftz.f32 %f1484, %f1483, %f1482; mul.ftz.f32 %f1485, %f260, %f256; mul.ftz.f32 %f1486, %f257, %f259; sub.ftz.f32 %f1487, %f1486, %f1485; add.ftz.f32 %f266, %f1481, %f1481; add.ftz.f32 %f267, %f1484, %f1484; add.ftz.f32 %f268, %f1487, %f1487; mul.ftz.f32 %f1488, %f2448, %f267; fma.rn.ftz.f32 %f1489, %f2447, %f266, %f1488; fma.rn.ftz.f32 %f269, %f2449, %f268, %f1489; setp.eq.ftz.f32 %p121, %f269, 0f00000000; @%p121 bra $L__BB155_86; mov.pred %p272, -1; mul.ftz.f32 %f1490, %f254, %f267; fma.rn.ftz.f32 %f1491, %f253, %f266, %f1490; fma.rn.ftz.f32 %f1492, %f255, %f268, %f1491; div.approx.ftz.f32 %f270, %f1492, %f269; setp.ltu.ftz.f32 %p123, %f270, 0f00000000; setp.gtu.ftz.f32 %p124, %f270, %f100; or.pred %p125, %p124, %p123; @%p125 bra $L__BB155_86; abs.ftz.f32 %f2381, %f264; mov.f32 %f2464, 0f00000000; setp.lt.ftz.f32 %p126, %f2381, 0f219392EF; @%p126 bra $L__BB155_85; rcp.approx.ftz.f32 %f2464, %f264; $L__BB155_85: mul.ftz.f32 %f1494, %f262, %f2464; min.ftz.f32 %f2485, %f1494, %f716; mul.ftz.f32 %f1496, %f263, %f2464; min.ftz.f32 %f2484, %f1496, %f716; mov.pred %p272, 0; mov.f32 %f2486, %f270; $L__BB155_86: @%p272 bra $L__BB155_93; ld.local.u32 %r565, [%rd1]; min.s32 %r563, %r565, 4; $L__BB155_88: add.s32 %r83, %r563, -1; setp.gt.s32 %p128, %r563, 0; @%p128 bra $L__BB155_92; bra.uni $L__BB155_89; $L__BB155_92: mul.wide.s32 %rd316, %r83, 24; add.s64 %rd317, %rd1, %rd316; ld.local.f32 %f1519, [%rd317+4]; setp.neu.ftz.f32 %p131, %f1519, %f2486; mov.u32 %r563, %r83; @%p131 bra $L__BB155_88; bra.uni $L__BB155_93; $L__BB155_89: setp.lt.s32 %p129, %r565, 4; add.s32 %r84, %r565, 1; st.local.u32 [%rd1], %r84; @%p129 bra $L__BB155_91; mad.lo.s32 %r573, %r573, 1103515245, 12345; rem.u32 %r565, %r573, %r84; setp.gt.s32 %p130, %r565, 3; @%p130 bra $L__BB155_93; $L__BB155_91: mul.wide.s32 %rd312, %r565, 24; add.s64 %rd313, %rd1, %rd312; st.local.f32 [%rd313+4], %f2486; st.local.f32 [%rd313+8], %f2485; st.local.f32 [%rd313+12], %f2484; st.local.u32 [%rd313+16], %r61; st.local.u32 [%rd313+20], %r3; mov.u32 %r408, 65; st.local.u32 [%rd313+24], %r408; sub.ftz.f32 %f1497, %f252, %f246; sub.ftz.f32 %f1498, %f248, %f245; mul.ftz.f32 %f1499, %f1497, %f1498; sub.ftz.f32 %f1500, %f251, %f245; sub.ftz.f32 %f1501, %f249, %f246; mul.ftz.f32 %f1502, %f1501, %f1500; sub.ftz.f32 %f1503, %f1499, %f1502; sub.ftz.f32 %f1504, %f250, %f244; mul.ftz.f32 %f1505, %f1501, %f1504; sub.ftz.f32 %f1506, %f247, %f244; mul.ftz.f32 %f1507, %f1497, %f1506; sub.ftz.f32 %f1508, %f1505, %f1507; mul.ftz.f32 %f1509, %f1500, %f1506; mul.ftz.f32 %f1510, %f1498, %f1504; sub.ftz.f32 %f1511, %f1509, %f1510; mul.ftz.f32 %f1512, %f1508, %f1508; fma.rn.ftz.f32 %f1513, %f1503, %f1503, %f1512; fma.rn.ftz.f32 %f1514, %f1511, %f1511, %f1513; sqrt.approx.ftz.f32 %f1515, %f1514; div.approx.ftz.f32 %f1516, %f1503, %f1515; div.approx.ftz.f32 %f1517, %f1508, %f1515; div.approx.ftz.f32 %f1518, %f1511, %f1515; mul.wide.s32 %rd314, %r565, 12; add.s64 %rd315, %rd1, %rd314; st.local.f32 [%rd315+100], %f1516; st.local.f32 [%rd315+104], %f1517; st.local.f32 [%rd315+108], %f1518; $L__BB155_93: add.s32 %r574, %r52, -1; cvt.u32.u64 %r409, %rd28; add.s32 %r556, %r409, 1; setp.lt.s32 %p132, %r556, %r379; @%p132 bra $L__BB155_65; bra.uni $L__BB155_112; $L__BB155_94: setp.ge.s32 %p133, %r556, %r379; @%p133 bra $L__BB155_112; $L__BB155_95: cvt.s64.s32 %rd86, %r556; setp.ne.s32 %p134, %r545, -1; @%p134 bra $L__BB155_97; cvta.to.global.u64 %rd318, %rd253; shl.b64 %rd319, %rd86, 2; add.s64 %rd320, %rd318, %rd319; ld.global.u32 %r410, [%rd320]; setp.ne.s32 %p135, %r410, %r3; @%p135 bra $L__BB155_111; $L__BB155_97: cvta.to.global.u64 %rd321, %rd252; shl.b64 %rd322, %rd86, 2; add.s64 %rd323, %rd321, %rd322; ld.global.u32 %r95, [%rd323]; setp.eq.s32 %p136, %r95, -1; @%p136 bra $L__BB155_111; cvta.to.global.u64 %rd324, %rd241; mul.wide.s32 %rd325, %r95, 16; add.s64 %rd326, %rd324, %rd325; ld.global.u32 %r411, [%rd326+12]; mul.wide.u32 %rd327, %r411, 12; add.s64 %rd328, %rd7, %rd327; add.s32 %r412, %r411, 1; mul.wide.u32 %rd329, %r412, 12; add.s64 %rd330, %rd7, %rd329; add.s32 %r413, %r411, 2; mul.wide.u32 %rd331, %r413, 12; add.s64 %rd332, %rd7, %rd331; ld.global.f32 %f284, [%rd328]; sub.ftz.f32 %f285, %f284, %f2450; ld.global.f32 %f286, [%rd328+4]; sub.ftz.f32 %f287, %f286, %f2451; ld.global.f32 %f288, [%rd328+8]; sub.ftz.f32 %f289, %f288, %f2452; ld.global.f32 %f290, [%rd330]; sub.ftz.f32 %f1520, %f290, %f2450; ld.global.f32 %f291, [%rd330+4]; sub.ftz.f32 %f1521, %f291, %f2451; ld.global.f32 %f292, [%rd330+8]; sub.ftz.f32 %f1522, %f292, %f2452; ld.global.f32 %f293, [%rd332]; sub.ftz.f32 %f1523, %f293, %f2450; ld.global.f32 %f294, [%rd332+4]; sub.ftz.f32 %f1524, %f294, %f2451; ld.global.f32 %f295, [%rd332+8]; sub.ftz.f32 %f1525, %f295, %f2452; sub.ftz.f32 %f296, %f1523, %f285; sub.ftz.f32 %f297, %f1524, %f287; sub.ftz.f32 %f298, %f1525, %f289; sub.ftz.f32 %f299, %f285, %f1520; sub.ftz.f32 %f300, %f287, %f1521; sub.ftz.f32 %f301, %f289, %f1522; sub.ftz.f32 %f1526, %f1520, %f1523; sub.ftz.f32 %f1527, %f1521, %f1524; sub.ftz.f32 %f1528, %f1522, %f1525; add.ftz.f32 %f1529, %f285, %f1523; add.ftz.f32 %f1530, %f287, %f1524; add.ftz.f32 %f1531, %f289, %f1525; mul.ftz.f32 %f1532, %f297, %f1531; mul.ftz.f32 %f1533, %f1530, %f298; sub.ftz.f32 %f1534, %f1532, %f1533; mul.ftz.f32 %f1535, %f1529, %f298; mul.ftz.f32 %f1536, %f296, %f1531; sub.ftz.f32 %f1537, %f1535, %f1536; mul.ftz.f32 %f1538, %f296, %f1530; mul.ftz.f32 %f1539, %f1529, %f297; sub.ftz.f32 %f1540, %f1538, %f1539; mul.ftz.f32 %f1541, %f2448, %f1537; fma.rn.ftz.f32 %f1542, %f2447, %f1534, %f1541; fma.rn.ftz.f32 %f302, %f2449, %f1540, %f1542; add.ftz.f32 %f1543, %f285, %f1520; add.ftz.f32 %f1544, %f287, %f1521; add.ftz.f32 %f1545, %f289, %f1522; mul.ftz.f32 %f1546, %f300, %f1545; mul.ftz.f32 %f1547, %f1544, %f301; sub.ftz.f32 %f1548, %f1546, %f1547; mul.ftz.f32 %f1549, %f1543, %f301; mul.ftz.f32 %f1550, %f299, %f1545; sub.ftz.f32 %f1551, %f1549, %f1550; mul.ftz.f32 %f1552, %f299, %f1544; mul.ftz.f32 %f1553, %f1543, %f300; sub.ftz.f32 %f1554, %f1552, %f1553; mul.ftz.f32 %f1555, %f2448, %f1551; fma.rn.ftz.f32 %f1556, %f2447, %f1548, %f1555; fma.rn.ftz.f32 %f303, %f2449, %f1554, %f1556; add.ftz.f32 %f1557, %f1520, %f1523; add.ftz.f32 %f1558, %f1521, %f1524; add.ftz.f32 %f1559, %f1522, %f1525; mul.ftz.f32 %f1560, %f1527, %f1559; mul.ftz.f32 %f1561, %f1558, %f1528; sub.ftz.f32 %f1562, %f1560, %f1561; mul.ftz.f32 %f1563, %f1557, %f1528; mul.ftz.f32 %f1564, %f1526, %f1559; sub.ftz.f32 %f1565, %f1563, %f1564; mul.ftz.f32 %f1566, %f1526, %f1558; mul.ftz.f32 %f1567, %f1557, %f1527; sub.ftz.f32 %f1568, %f1566, %f1567; mul.ftz.f32 %f1569, %f2448, %f1565; fma.rn.ftz.f32 %f1570, %f2447, %f1562, %f1569; fma.rn.ftz.f32 %f1571, %f2449, %f1568, %f1570; add.ftz.f32 %f1572, %f303, %f302; add.ftz.f32 %f304, %f1571, %f1572; abs.ftz.f32 %f305, %f304; mul.ftz.f32 %f1573, %f305, 0f34000000; min.ftz.f32 %f1574, %f303, %f1571; min.ftz.f32 %f1575, %f302, %f1574; max.ftz.f32 %f1576, %f303, %f1571; max.ftz.f32 %f1577, %f302, %f1576; mul.ftz.f32 %f1578, %f305, 0fB4000000; setp.ltu.ftz.f32 %p138, %f1575, %f1578; setp.gtu.ftz.f32 %p139, %f1577, %f1573; and.pred %p140, %p138, %p139; mov.pred %p273, -1; @%p140 bra $L__BB155_104; mov.pred %p273, -1; mul.ftz.f32 %f1579, %f300, %f298; mul.ftz.f32 %f1580, %f301, %f297; sub.ftz.f32 %f1581, %f1579, %f1580; mul.ftz.f32 %f1582, %f299, %f298; mul.ftz.f32 %f1583, %f301, %f296; sub.ftz.f32 %f1584, %f1583, %f1582; mul.ftz.f32 %f1585, %f300, %f296; mul.ftz.f32 %f1586, %f299, %f297; sub.ftz.f32 %f1587, %f1586, %f1585; add.ftz.f32 %f306, %f1581, %f1581; add.ftz.f32 %f307, %f1584, %f1584; add.ftz.f32 %f308, %f1587, %f1587; mul.ftz.f32 %f1588, %f2448, %f307; fma.rn.ftz.f32 %f1589, %f2447, %f306, %f1588; fma.rn.ftz.f32 %f309, %f2449, %f308, %f1589; setp.eq.ftz.f32 %p142, %f309, 0f00000000; @%p142 bra $L__BB155_104; mov.pred %p273, -1; sub.ftz.f32 %f2379, %f288, %f2452; sub.ftz.f32 %f2378, %f284, %f2450; sub.ftz.f32 %f2377, %f286, %f2451; mul.ftz.f32 %f1590, %f2377, %f307; fma.rn.ftz.f32 %f1591, %f2378, %f306, %f1590; fma.rn.ftz.f32 %f1592, %f2379, %f308, %f1591; div.approx.ftz.f32 %f310, %f1592, %f309; setp.ltu.ftz.f32 %p144, %f310, 0f00000000; setp.gtu.ftz.f32 %p145, %f310, %f100; or.pred %p146, %p145, %p144; @%p146 bra $L__BB155_104; setp.lt.ftz.f32 %p147, %f305, 0f219392EF; mov.f32 %f2474, 0f00000000; @%p147 bra $L__BB155_103; rcp.approx.ftz.f32 %f2474, %f304; $L__BB155_103: mul.ftz.f32 %f1594, %f302, %f2474; min.ftz.f32 %f2479, %f1594, %f716; mul.ftz.f32 %f1596, %f303, %f2474; min.ftz.f32 %f2478, %f1596, %f716; mov.pred %p273, 0; mov.f32 %f2480, %f310; $L__BB155_104: @%p273 bra $L__BB155_111; ld.local.u32 %r571, [%rd1]; min.s32 %r569, %r571, 4; $L__BB155_106: add.s32 %r99, %r569, -1; setp.gt.s32 %p149, %r569, 0; @%p149 bra $L__BB155_110; bra.uni $L__BB155_107; $L__BB155_110: mul.wide.s32 %rd337, %r99, 24; add.s64 %rd338, %rd1, %rd337; ld.local.f32 %f1619, [%rd338+4]; setp.neu.ftz.f32 %p152, %f1619, %f2480; mov.u32 %r569, %r99; @%p152 bra $L__BB155_106; bra.uni $L__BB155_111; $L__BB155_107: setp.lt.s32 %p150, %r571, 4; add.s32 %r100, %r571, 1; st.local.u32 [%rd1], %r100; @%p150 bra $L__BB155_109; mad.lo.s32 %r573, %r573, 1103515245, 12345; rem.u32 %r571, %r573, %r100; setp.gt.s32 %p151, %r571, 3; @%p151 bra $L__BB155_111; $L__BB155_109: mul.wide.s32 %rd333, %r571, 24; add.s64 %rd334, %rd1, %rd333; st.local.u32 [%rd334+16], %r95; st.local.u32 [%rd334+20], %r3; mov.u32 %r414, 1; st.local.u32 [%rd334+24], %r414; st.local.f32 [%rd334+8], %f2479; st.local.f32 [%rd334+12], %f2478; st.local.f32 [%rd334+4], %f2480; sub.ftz.f32 %f1597, %f295, %f288; sub.ftz.f32 %f1598, %f291, %f286; mul.ftz.f32 %f1599, %f1598, %f1597; sub.ftz.f32 %f1600, %f294, %f286; sub.ftz.f32 %f1601, %f292, %f288; mul.ftz.f32 %f1602, %f1601, %f1600; sub.ftz.f32 %f1603, %f1599, %f1602; sub.ftz.f32 %f1604, %f293, %f284; mul.ftz.f32 %f1605, %f1601, %f1604; sub.ftz.f32 %f1606, %f290, %f284; mul.ftz.f32 %f1607, %f1606, %f1597; sub.ftz.f32 %f1608, %f1605, %f1607; mul.ftz.f32 %f1609, %f1606, %f1600; mul.ftz.f32 %f1610, %f1598, %f1604; sub.ftz.f32 %f1611, %f1609, %f1610; mul.ftz.f32 %f1612, %f1608, %f1608; fma.rn.ftz.f32 %f1613, %f1603, %f1603, %f1612; fma.rn.ftz.f32 %f1614, %f1611, %f1611, %f1613; sqrt.approx.ftz.f32 %f1615, %f1614; div.approx.ftz.f32 %f1616, %f1603, %f1615; div.approx.ftz.f32 %f1617, %f1608, %f1615; div.approx.ftz.f32 %f1618, %f1611, %f1615; mul.wide.s32 %rd335, %r571, 12; add.s64 %rd336, %rd1, %rd335; st.local.f32 [%rd336+100], %f1616; st.local.f32 [%rd336+104], %f1617; st.local.f32 [%rd336+108], %f1618; $L__BB155_111: add.s32 %r574, %r52, -1; cvt.u32.u64 %r415, %rd86; add.s32 %r556, %r415, 1; setp.lt.s32 %p153, %r556, %r379; @%p153 bra $L__BB155_95; bra.uni $L__BB155_112; $L__BB155_49: setp.eq.s32 %p78, %r553, 1985229328; setp.lt.s32 %p79, %r553, 0; or.pred %p80, %p79, %p78; @%p80 bra $L__BB155_61; cvta.to.global.u64 %rd25, %rd250; $L__BB155_51: mul.wide.s32 %rd274, %r553, 16; add.s64 %rd26, %rd25, %rd274; ld.global.v4.f32 {%f1182, %f1183, %f1184, %f1185}, [%rd26]; mov.b32 %r553, %f1185; mov.b32 %r37, %f1184; mov.b32 %r36, %f1183; mov.b32 %r35, %f1182; mov.b64 %rd275, {%r35, %r36}; cvt.u32.u64 %r368, %rd275; and.b32 %r369, %r368, 1024; setp.eq.s32 %p81, %r369, 0; and.b32 %r39, %r35, 2047; @%p81 bra $L__BB155_54; bra.uni $L__BB155_52; $L__BB155_54: setp.ne.s32 %p87, %r39, 0; ld.global.v4.f32 {%f1326, %f1327, %f1328, %f1329}, [%rd26+16]; ld.global.v4.f32 {%f1334, %f1335, %f1336, %f1337}, [%rd26+32]; ld.global.v4.f32 {%f1342, %f1343, %f1344, %f1345}, [%rd26+48]; sub.ftz.f32 %f1350, %f1326, %f2450; mul.ftz.f32 %f1351, %f2444, %f1350; sub.ftz.f32 %f1352, %f1328, %f2450; mul.ftz.f32 %f1353, %f2444, %f1352; sub.ftz.f32 %f1354, %f1334, %f2451; mul.ftz.f32 %f1355, %f2445, %f1354; sub.ftz.f32 %f1356, %f1336, %f2451; mul.ftz.f32 %f1357, %f2445, %f1356; sub.ftz.f32 %f1358, %f1342, %f2452; mul.ftz.f32 %f1359, %f2446, %f1358; sub.ftz.f32 %f1360, %f1344, %f2452; mul.ftz.f32 %f1361, %f2446, %f1360; min.ftz.f32 %f1362, %f1351, %f1353; min.ftz.f32 %f1363, %f1355, %f1357; min.ftz.f32 %f1364, %f1359, %f1361; mov.f32 %f1365, 0f00000000; max.ftz.f32 %f1366, %f1365, %f1362; max.ftz.f32 %f1367, %f1363, %f1364; max.ftz.f32 %f2459, %f1366, %f1367; max.ftz.f32 %f1368, %f1351, %f1353; max.ftz.f32 %f1369, %f1355, %f1357; max.ftz.f32 %f1370, %f1359, %f1361; min.ftz.f32 %f1371, %f100, %f1368; min.ftz.f32 %f1372, %f1369, %f1370; min.ftz.f32 %f1373, %f1371, %f1372; sub.ftz.f32 %f1374, %f1327, %f2450; mul.ftz.f32 %f1375, %f2444, %f1374; sub.ftz.f32 %f1376, %f1329, %f2450; mul.ftz.f32 %f1377, %f2444, %f1376; sub.ftz.f32 %f1378, %f1335, %f2451; mul.ftz.f32 %f1379, %f2445, %f1378; sub.ftz.f32 %f1380, %f1337, %f2451; mul.ftz.f32 %f1381, %f2445, %f1380; sub.ftz.f32 %f1382, %f1343, %f2452; mul.ftz.f32 %f1383, %f2446, %f1382; sub.ftz.f32 %f1384, %f1345, %f2452; mul.ftz.f32 %f1385, %f2446, %f1384; min.ftz.f32 %f1386, %f1375, %f1377; min.ftz.f32 %f1387, %f1379, %f1381; min.ftz.f32 %f1388, %f1383, %f1385; max.ftz.f32 %f1389, %f1365, %f1386; max.ftz.f32 %f1390, %f1387, %f1388; max.ftz.f32 %f2460, %f1389, %f1390; max.ftz.f32 %f1391, %f1375, %f1377; max.ftz.f32 %f1392, %f1379, %f1381; max.ftz.f32 %f1393, %f1383, %f1385; min.ftz.f32 %f1394, %f100, %f1391; min.ftz.f32 %f1395, %f1392, %f1393; min.ftz.f32 %f1396, %f1394, %f1395; setp.ge.ftz.f32 %p88, %f1373, %f2459; and.pred %p89, %p87, %p88; selp.u32 %r373, 1, 0, %p89; setp.ltu.ftz.f32 %p90, %f1396, %f2460; and.b32 %r375, %r36, 2047; setp.eq.s32 %p91, %r375, 0; or.pred %p92, %p91, %p90; selp.b32 %r376, 0, 2, %p92; or.b32 %r551, %r376, %r373; bra.uni $L__BB155_55; $L__BB155_52: setp.ne.s32 %p82, %r39, 0; ld.global.v4.f32 {%f1189, %f1190, %f1191, %f1192}, [%rd26+16]; ld.global.v4.f32 {%f1197, %f1198, %f1199, %f1200}, [%rd26+32]; ld.global.v4.f32 {%f1205, %f1206, %f1207, %f1208}, [%rd26+48]; mul.ftz.f32 %f1213, %f2448, %f1190; fma.rn.ftz.f32 %f1214, %f2447, %f1189, %f1213; fma.rn.ftz.f32 %f1215, %f2449, %f1191, %f1214; mul.ftz.f32 %f1216, %f2448, %f1198; fma.rn.ftz.f32 %f1217, %f2447, %f1197, %f1216; fma.rn.ftz.f32 %f1218, %f2449, %f1199, %f1217; mul.ftz.f32 %f1219, %f2448, %f1206; fma.rn.ftz.f32 %f1220, %f2447, %f1205, %f1219; fma.rn.ftz.f32 %f1221, %f2449, %f1207, %f1220; mul.ftz.f32 %f1222, %f2451, %f1190; fma.rn.ftz.f32 %f1223, %f2450, %f1189, %f1222; fma.rn.ftz.f32 %f1224, %f2452, %f1191, %f1223; add.ftz.f32 %f1225, %f1192, %f1224; mul.ftz.f32 %f1226, %f2451, %f1198; fma.rn.ftz.f32 %f1227, %f2450, %f1197, %f1226; fma.rn.ftz.f32 %f1228, %f2452, %f1199, %f1227; add.ftz.f32 %f1229, %f1200, %f1228; mul.ftz.f32 %f1230, %f2451, %f1206; fma.rn.ftz.f32 %f1231, %f2450, %f1205, %f1230; fma.rn.ftz.f32 %f1232, %f2452, %f1207, %f1231; add.ftz.f32 %f1233, %f1208, %f1232; rcp.approx.ftz.f32 %f1234, %f1215; rcp.approx.ftz.f32 %f1235, %f1218; rcp.approx.ftz.f32 %f1236, %f1221; mul.ftz.f32 %f1237, %f1234, %f1225; neg.ftz.f32 %f1238, %f1237; mul.ftz.f32 %f1239, %f1235, %f1229; neg.ftz.f32 %f1240, %f1239; mul.ftz.f32 %f1241, %f1236, %f1233; neg.ftz.f32 %f1242, %f1241; sub.ftz.f32 %f1243, %f1234, %f1237; sub.ftz.f32 %f1244, %f1235, %f1239; sub.ftz.f32 %f1245, %f1236, %f1241; min.ftz.f32 %f1246, %f1238, %f1243; min.ftz.f32 %f1247, %f1240, %f1244; min.ftz.f32 %f1248, %f1242, %f1245; max.ftz.f32 %f1249, %f1238, %f1243; max.ftz.f32 %f1250, %f1240, %f1244; max.ftz.f32 %f1251, %f1242, %f1245; mov.f32 %f1252, 0f00000000; max.ftz.f32 %f1253, %f1252, %f1246; max.ftz.f32 %f1254, %f1247, %f1248; max.ftz.f32 %f2459, %f1253, %f1254; min.ftz.f32 %f1255, %f100, %f1249; min.ftz.f32 %f1256, %f1250, %f1251; min.ftz.f32 %f1257, %f1255, %f1256; setp.le.ftz.f32 %p83, %f2459, %f1257; and.pred %p84, %p82, %p83; selp.u32 %r551, 1, 0, %p84; ld.global.v4.f32 {%f1258, %f1259, %f1260, %f1261}, [%rd26+64]; ld.global.v4.f32 {%f1266, %f1267, %f1268, %f1269}, [%rd26+80]; ld.global.v4.f32 {%f1274, %f1275, %f1276, %f1277}, [%rd26+96]; mul.ftz.f32 %f1282, %f2448, %f1259; fma.rn.ftz.f32 %f1283, %f2447, %f1258, %f1282; fma.rn.ftz.f32 %f1284, %f2449, %f1260, %f1283; mul.ftz.f32 %f1285, %f2448, %f1267; fma.rn.ftz.f32 %f1286, %f2447, %f1266, %f1285; fma.rn.ftz.f32 %f1287, %f2449, %f1268, %f1286; mul.ftz.f32 %f1288, %f2448, %f1275; fma.rn.ftz.f32 %f1289, %f2447, %f1274, %f1288; fma.rn.ftz.f32 %f1290, %f2449, %f1276, %f1289; mul.ftz.f32 %f1291, %f2451, %f1259; fma.rn.ftz.f32 %f1292, %f2450, %f1258, %f1291; fma.rn.ftz.f32 %f1293, %f2452, %f1260, %f1292; add.ftz.f32 %f1294, %f1261, %f1293; mul.ftz.f32 %f1295, %f2451, %f1267; fma.rn.ftz.f32 %f1296, %f2450, %f1266, %f1295; fma.rn.ftz.f32 %f1297, %f2452, %f1268, %f1296; add.ftz.f32 %f1298, %f1269, %f1297; mul.ftz.f32 %f1299, %f2451, %f1275; fma.rn.ftz.f32 %f1300, %f2450, %f1274, %f1299; fma.rn.ftz.f32 %f1301, %f2452, %f1276, %f1300; add.ftz.f32 %f1302, %f1277, %f1301; rcp.approx.ftz.f32 %f1303, %f1284; rcp.approx.ftz.f32 %f1304, %f1287; rcp.approx.ftz.f32 %f1305, %f1290; mul.ftz.f32 %f1306, %f1303, %f1294; neg.ftz.f32 %f1307, %f1306; mul.ftz.f32 %f1308, %f1304, %f1298; neg.ftz.f32 %f1309, %f1308; mul.ftz.f32 %f1310, %f1305, %f1302; neg.ftz.f32 %f1311, %f1310; sub.ftz.f32 %f1312, %f1303, %f1306; sub.ftz.f32 %f1313, %f1304, %f1308; sub.ftz.f32 %f1314, %f1305, %f1310; min.ftz.f32 %f1315, %f1307, %f1312; min.ftz.f32 %f1316, %f1309, %f1313; min.ftz.f32 %f1317, %f1311, %f1314; max.ftz.f32 %f1318, %f1307, %f1312; max.ftz.f32 %f1319, %f1309, %f1313; max.ftz.f32 %f1320, %f1311, %f1314; max.ftz.f32 %f1321, %f1252, %f1315; max.ftz.f32 %f1322, %f1316, %f1317; max.ftz.f32 %f2460, %f1321, %f1322; min.ftz.f32 %f1323, %f100, %f1318; min.ftz.f32 %f1324, %f1319, %f1320; min.ftz.f32 %f1325, %f1323, %f1324; setp.gtu.ftz.f32 %p85, %f2460, %f1325; @%p85 bra $L__BB155_55; and.b32 %r371, %r36, 2047; setp.eq.s32 %p86, %r371, 0; or.b32 %r372, %r551, 2; selp.b32 %r551, %r551, %r372, %p86; $L__BB155_55: mul.wide.s32 %rd276, %r574, 4; add.s64 %rd27, %rd22, %rd276; setp.eq.s32 %p93, %r551, 3; @%p93 bra $L__BB155_59; bra.uni $L__BB155_56; $L__BB155_59: setp.lt.ftz.f32 %p96, %f2460, %f2459; selp.b32 %r48, %r553, %r37, %p96; selp.b32 %r377, %r37, %r553, %p96; add.s32 %r574, %r574, 1; st.local.u32 [%rd27+4], %r377; mov.u32 %r553, %r48; bra.uni $L__BB155_60; $L__BB155_56: setp.eq.s32 %p94, %r551, 2; @%p94 bra $L__BB155_60; setp.ne.s32 %p95, %r551, 0; mov.u32 %r553, %r37; @%p95 bra $L__BB155_60; ld.local.u32 %r553, [%rd27]; add.s32 %r574, %r574, -1; $L__BB155_60: setp.ne.s32 %p97, %r553, 1985229328; setp.gt.s32 %p98, %r553, -1; and.pred %p99, %p98, %p97; @%p99 bra $L__BB155_51; $L__BB155_61: mov.u32 %r52, %r574; setp.gt.s32 %p100, %r553, -1; @%p100 bra $L__BB155_112; bra.uni $L__BB155_62; $L__BB155_112: setp.ne.s32 %p154, %r553, 1985229328; @%p154 bra $L__BB155_49; $L__BB155_148: ld.local.u32 %r160, [%rd1]; setp.lt.s32 %p206, %r160, 1; @%p206 bra $L__BB155_209; ld.local.u32 %r161, [%rd2+64]; cvt.rn.f32.s32 %f407, %r160; sub.ftz.f32 %f1983, %f1, %f90; mul.ftz.f32 %f1984, %f1983, %f1983; mul.ftz.f32 %f1985, %f1983, %f1984; mul.ftz.f32 %f408, %f1985, 0f41200000; mul.wide.s32 %rd375, %r3, 208; add.s64 %rd376, %rd11, %rd375; add.s64 %rd92, %rd376, 144; cvta.to.global.u64 %rd93, %rd244; min.s32 %r162, %r160, 4; mov.u32 %r599, 0; mov.u32 %r611, %r161; $L__BB155_150: cvt.s64.s32 %rd94, %r599; mul.wide.s32 %rd377, %r599, 24; add.s64 %rd378, %rd1, %rd377; add.s64 %rd95, %rd378, 8; setp.eq.s32 %p207, %r161, 1; @%p207 bra $L__BB155_169; bra.uni $L__BB155_151; $L__BB155_169: ld.local.u32 %r480, [%rd95+8]; cvt.s64.s32 %rd507, %r480; cvta.to.global.u64 %rd402, %rd241; mul.wide.s32 %rd403, %r480, 16; add.s64 %rd404, %rd402, %rd403; ld.global.u32 %r481, [%rd404+12]; mul.wide.u32 %rd405, %r481, 12; add.s64 %rd406, %rd7, %rd405; add.s32 %r482, %r481, 1; mul.wide.u32 %rd407, %r482, 12; add.s64 %rd408, %rd7, %rd407; add.s32 %r483, %r481, 2; mul.wide.u32 %rd409, %r483, 12; add.s64 %rd410, %rd7, %rd409; ld.global.f32 %f2062, [%rd408]; ld.global.f32 %f2063, [%rd406]; sub.ftz.f32 %f2064, %f2062, %f2063; ld.global.f32 %f2065, [%rd408+4]; ld.global.f32 %f2066, [%rd406+4]; sub.ftz.f32 %f2067, %f2065, %f2066; ld.global.f32 %f2068, [%rd408+8]; ld.global.f32 %f2069, [%rd406+8]; sub.ftz.f32 %f2070, %f2068, %f2069; ld.local.f32 %f2071, [%rd95]; fma.rn.ftz.f32 %f2072, %f2071, %f2064, %f2063; fma.rn.ftz.f32 %f2073, %f2071, %f2067, %f2066; fma.rn.ftz.f32 %f2074, %f2071, %f2070, %f2069; ld.global.f32 %f2075, [%rd410]; sub.ftz.f32 %f2076, %f2075, %f2063; ld.global.f32 %f2077, [%rd410+4]; sub.ftz.f32 %f2078, %f2077, %f2066; ld.global.f32 %f2079, [%rd410+8]; sub.ftz.f32 %f2080, %f2079, %f2069; ld.local.f32 %f2081, [%rd95+4]; fma.rn.ftz.f32 %f2561, %f2081, %f2076, %f2072; fma.rn.ftz.f32 %f2560, %f2081, %f2078, %f2073; fma.rn.ftz.f32 %f2559, %f2081, %f2080, %f2074; ld.local.u32 %r185, [%rd2+56]; and.b32 %r484, %r185, 4; setp.ne.s32 %p220, %r484, 0; @%p220 bra $L__BB155_174; and.b32 %r485, %r185, 2; setp.eq.s32 %p221, %r485, 0; @%p221 bra $L__BB155_172; bra.uni $L__BB155_171; $L__BB155_172: ld.global.v4.f32 {%f2544, %f2545, %f2546, %f2547}, [%rd376]; ld.global.v4.f32 {%f2548, %f2549, %f2550, %f2551}, [%rd376+16]; ld.global.v4.f32 {%f2552, %f2553, %f2554, %f2555}, [%rd376+32]; bra.uni $L__BB155_173; $L__BB155_151: setp.eq.s32 %p208, %r611, 65; ld.local.s32 %rd507, [%rd95+8]; @%p208 bra $L__BB155_152; bra.uni $L__BB155_174; $L__BB155_152: ld.global.u32 %r166, [%rd92+12]; ld.global.u32 %r167, [%rd92+8]; shl.b32 %r453, %r167, 1; cvt.rn.f32.s32 %f1986, %r453; mul.ftz.f32 %f415, %f101, %f1986; cvt.rzi.ftz.s32.f32 %r454, %f415; add.s32 %r455, %r453, -1; min.s32 %r168, %r454, %r455; ld.global.u32 %r601, [%rd92+20]; mul.wide.u32 %rd379, %r601, 16; add.s64 %rd380, %rd12, %rd379; ld.global.u16 %rs37, [%rd380+12]; ld.global.u32 %r600, [%rd380+8]; ld.global.u64 %rd488, [%rd380]; setp.eq.s64 %p209, %rd488, 11; @%p209 bra $L__BB155_157; $L__BB155_153: setp.eq.s64 %p210, %rd488, 0; @%p210 bra $L__BB155_155; add.s32 %r601, %r601, 2; bra.uni $L__BB155_156; $L__BB155_155: setp.eq.s16 %p211, %rs37, 0; mov.u32 %r604, -1; mov.u32 %r601, %r600; @%p211 bra $L__BB155_158; $L__BB155_156: mul.wide.u32 %rd381, %r601, 16; add.s64 %rd382, %rd12, %rd381; ld.global.u16 %rs37, [%rd382+12]; ld.global.u32 %r600, [%rd382+8]; ld.global.u64 %rd488, [%rd382]; setp.ne.s64 %p212, %rd488, 11; @%p212 bra $L__BB155_153; $L__BB155_157: setp.eq.s16 %p213, %rs37, 0; selp.b32 %r604, -1, %r600, %p213; $L__BB155_158: cvt.u32.u64 %r457, %rd507; cvta.to.global.u64 %rd383, %rd241; mul.wide.s32 %rd384, %r457, 16; add.s64 %rd385, %rd383, %rd384; ld.global.v4.u32 {%r458, %r459, %r460, %r461}, [%rd385]; setp.eq.s32 %p214, %r168, %r167; @%p214 bra $L__BB155_160; bra.uni $L__BB155_159; $L__BB155_160: mul.wide.u32 %rd389, %r461, 12; add.s64 %rd497, %rd7, %rd389; add.s32 %r468, %r461, 1; mul.wide.u32 %rd390, %r468, 12; add.s64 %rd494, %rd7, %rd390; add.s32 %r469, %r461, 2; mul.wide.u32 %rd391, %r469, 12; add.s64 %rd491, %rd7, %rd391; add.s64 %rd496, %rd497, 4; add.s64 %rd495, %rd497, 8; add.s64 %rd493, %rd494, 4; add.s64 %rd492, %rd494, 8; add.s64 %rd490, %rd491, 4; add.s64 %rd489, %rd491, 8; bra.uni $L__BB155_161; $L__BB155_171: ld.local.v4.f32 {%f2544, %f2545, %f2546, %f2547}, [%rd2+144]; ld.local.v4.f32 {%f2548, %f2549, %f2550, %f2551}, [%rd2+160]; ld.local.v4.f32 {%f2552, %f2553, %f2554, %f2555}, [%rd2+176]; $L__BB155_173: mul.ftz.f32 %f2106, %f2560, %f2545; fma.rn.ftz.f32 %f2107, %f2561, %f2544, %f2106; fma.rn.ftz.f32 %f2108, %f2559, %f2546, %f2107; mul.ftz.f32 %f2109, %f2560, %f2549; fma.rn.ftz.f32 %f2110, %f2561, %f2548, %f2109; fma.rn.ftz.f32 %f2111, %f2559, %f2550, %f2110; mul.ftz.f32 %f2112, %f2560, %f2553; fma.rn.ftz.f32 %f2113, %f2561, %f2552, %f2112; fma.rn.ftz.f32 %f2114, %f2559, %f2554, %f2113; add.ftz.f32 %f2561, %f2547, %f2108; add.ftz.f32 %f2560, %f2551, %f2111; add.ftz.f32 %f2559, %f2555, %f2114; bra.uni $L__BB155_174; $L__BB155_159: setp.gt.s32 %p215, %r168, %r167; selp.b32 %r462, -1, 0, %p215; add.s32 %r463, %r168, %r462; mad.lo.s32 %r464, %r463, %r166, %r604; add.s32 %r465, %r458, %r464; mul.wide.u32 %rd386, %r465, 12; add.s64 %rd497, %rd13, %rd386; add.s32 %r466, %r459, %r464; mul.wide.u32 %rd387, %r466, 12; add.s64 %rd494, %rd13, %rd387; add.s32 %r467, %r460, %r464; mul.wide.u32 %rd388, %r467, 12; add.s64 %rd491, %rd13, %rd388; add.s64 %rd496, %rd497, 4; add.s64 %rd495, %rd497, 8; add.s64 %rd493, %rd494, 4; add.s64 %rd492, %rd494, 8; add.s64 %rd490, %rd491, 4; add.s64 %rd489, %rd491, 8; $L__BB155_161: ld.global.f32 %f416, [%rd497]; ld.global.f32 %f417, [%rd496]; ld.global.f32 %f418, [%rd495]; ld.global.f32 %f419, [%rd494]; ld.global.f32 %f420, [%rd493]; ld.global.f32 %f421, [%rd492]; ld.global.f32 %f422, [%rd491]; ld.global.f32 %f423, [%rd490]; ld.global.f32 %f424, [%rd489]; add.s32 %r183, %r168, 1; setp.eq.s32 %p216, %r183, %r167; @%p216 bra $L__BB155_163; bra.uni $L__BB155_162; $L__BB155_163: mul.wide.u32 %rd395, %r461, 12; add.s64 %rd498, %rd7, %rd395; add.s32 %r476, %r461, 1; mul.wide.u32 %rd396, %r476, 12; add.s64 %rd501, %rd7, %rd396; add.s32 %r477, %r461, 2; mul.wide.u32 %rd397, %r477, 12; add.s64 %rd504, %rd7, %rd397; add.s64 %rd499, %rd498, 4; add.s64 %rd500, %rd498, 8; add.s64 %rd502, %rd501, 4; add.s64 %rd503, %rd501, 8; add.s64 %rd505, %rd504, 4; add.s64 %rd506, %rd504, 8; bra.uni $L__BB155_164; $L__BB155_162: setp.ge.s32 %p217, %r168, %r167; selp.b32 %r470, -1, 0, %p217; add.s32 %r471, %r183, %r470; mad.lo.s32 %r472, %r471, %r166, %r604; add.s32 %r473, %r458, %r472; mul.wide.u32 %rd392, %r473, 12; add.s64 %rd498, %rd13, %rd392; add.s32 %r474, %r459, %r472; mul.wide.u32 %rd393, %r474, 12; add.s64 %rd501, %rd13, %rd393; add.s32 %r475, %r460, %r472; mul.wide.u32 %rd394, %r475, 12; add.s64 %rd504, %rd13, %rd394; add.s64 %rd499, %rd498, 4; add.s64 %rd500, %rd498, 8; add.s64 %rd502, %rd501, 4; add.s64 %rd503, %rd501, 8; add.s64 %rd505, %rd504, 4; add.s64 %rd506, %rd504, 8; $L__BB155_164: cvt.rn.f32.s32 %f1987, %r168; sub.ftz.f32 %f1988, %f415, %f1987; sub.ftz.f32 %f1990, %f716, %f1988; ld.global.f32 %f1991, [%rd498]; mul.ftz.f32 %f1992, %f1988, %f1991; ld.global.f32 %f1993, [%rd499]; mul.ftz.f32 %f1994, %f1988, %f1993; ld.global.f32 %f1995, [%rd500]; mul.ftz.f32 %f1996, %f1988, %f1995; fma.rn.ftz.f32 %f1997, %f1990, %f416, %f1992; fma.rn.ftz.f32 %f1998, %f1990, %f417, %f1994; fma.rn.ftz.f32 %f1999, %f1990, %f418, %f1996; ld.global.f32 %f2000, [%rd501]; mul.ftz.f32 %f2001, %f1988, %f2000; ld.global.f32 %f2002, [%rd502]; mul.ftz.f32 %f2003, %f1988, %f2002; ld.global.f32 %f2004, [%rd503]; mul.ftz.f32 %f2005, %f1988, %f2004; fma.rn.ftz.f32 %f2006, %f1990, %f419, %f2001; fma.rn.ftz.f32 %f2007, %f1990, %f420, %f2003; fma.rn.ftz.f32 %f2008, %f1990, %f421, %f2005; ld.global.f32 %f2009, [%rd504]; mul.ftz.f32 %f2010, %f1988, %f2009; ld.global.f32 %f2011, [%rd505]; mul.ftz.f32 %f2012, %f1988, %f2011; ld.global.f32 %f2013, [%rd506]; mul.ftz.f32 %f2014, %f1988, %f2013; fma.rn.ftz.f32 %f2015, %f1990, %f422, %f2010; fma.rn.ftz.f32 %f2016, %f1990, %f423, %f2012; fma.rn.ftz.f32 %f2017, %f1990, %f424, %f2014; sub.ftz.f32 %f2018, %f2006, %f1997; sub.ftz.f32 %f2019, %f2007, %f1998; sub.ftz.f32 %f2020, %f2008, %f1999; ld.local.f32 %f2021, [%rd95]; fma.rn.ftz.f32 %f2022, %f2021, %f2018, %f1997; fma.rn.ftz.f32 %f2023, %f2021, %f2019, %f1998; fma.rn.ftz.f32 %f2024, %f2020, %f2021, %f1999; sub.ftz.f32 %f2025, %f2015, %f1997; sub.ftz.f32 %f2026, %f2016, %f1998; sub.ftz.f32 %f2027, %f2017, %f1999; ld.local.f32 %f2028, [%rd95+4]; fma.rn.ftz.f32 %f2561, %f2028, %f2025, %f2022; fma.rn.ftz.f32 %f2560, %f2026, %f2028, %f2023; fma.rn.ftz.f32 %f2559, %f2027, %f2028, %f2024; ld.local.u32 %r184, [%rd2+56]; and.b32 %r478, %r184, 4; setp.ne.s32 %p218, %r478, 0; @%p218 bra $L__BB155_174; and.b32 %r479, %r184, 2; setp.eq.s32 %p219, %r479, 0; @%p219 bra $L__BB155_167; bra.uni $L__BB155_166; $L__BB155_167: ld.global.v4.f32 {%f2529, %f2530, %f2531, %f2532}, [%rd376]; ld.global.v4.f32 {%f2533, %f2534, %f2535, %f2536}, [%rd376+16]; ld.global.v4.f32 {%f2537, %f2538, %f2539, %f2540}, [%rd376+32]; bra.uni $L__BB155_168; $L__BB155_166: ld.local.v4.f32 {%f2529, %f2530, %f2531, %f2532}, [%rd2+144]; ld.local.v4.f32 {%f2533, %f2534, %f2535, %f2536}, [%rd2+160]; ld.local.v4.f32 {%f2537, %f2538, %f2539, %f2540}, [%rd2+176]; $L__BB155_168: mul.ftz.f32 %f2053, %f2560, %f2530; fma.rn.ftz.f32 %f2054, %f2561, %f2529, %f2053; fma.rn.ftz.f32 %f2055, %f2559, %f2531, %f2054; mul.ftz.f32 %f2056, %f2560, %f2534; fma.rn.ftz.f32 %f2057, %f2561, %f2533, %f2056; fma.rn.ftz.f32 %f2058, %f2559, %f2535, %f2057; mul.ftz.f32 %f2059, %f2560, %f2538; fma.rn.ftz.f32 %f2060, %f2561, %f2537, %f2059; fma.rn.ftz.f32 %f2061, %f2559, %f2539, %f2060; add.ftz.f32 %f2561, %f2532, %f2055; add.ftz.f32 %f2560, %f2536, %f2058; add.ftz.f32 %f2559, %f2540, %f2061; $L__BB155_174: mul.lo.s64 %rd415, %rd94, 12; add.s64 %rd416, %rd1, %rd415; ld.local.u32 %r486, [%rd95+12]; mul.wide.s32 %rd417, %r486, 4; add.s64 %rd418, %rd93, %rd417; ld.global.u32 %r186, [%rd418]; and.b32 %r487, %r186, 8; setp.eq.s32 %p222, %r487, 0; ld.local.f32 %f2115, [%rd416+100]; neg.ftz.f32 %f2116, %f2115; ld.local.f32 %f2117, [%rd416+104]; neg.ftz.f32 %f2118, %f2117; ld.local.f32 %f2119, [%rd416+108]; neg.ftz.f32 %f2120, %f2119; selp.f32 %f2574, %f2119, %f2120, %p222; selp.f32 %f2575, %f2117, %f2118, %p222; selp.f32 %f2576, %f2115, %f2116, %p222; shl.b64 %rd419, %rd507, 2; add.s64 %rd420, %rd10, %rd419; ld.global.u32 %r488, [%rd420]; setp.gt.s32 %p223, %r488, -1; mov.f32 %f2568, %f2574; mov.f32 %f2569, %f2575; mov.f32 %f2570, %f2576; @%p223 bra $L__BB155_195; ld.local.f32 %f533, [%rd95]; ld.local.f32 %f534, [%rd95+4]; cvta.to.global.u64 %rd421, %rd241; shl.b64 %rd422, %rd507, 4; add.s64 %rd157, %rd421, %rd422; setp.eq.s32 %p224, %r611, 1; @%p224 bra $L__BB155_192; bra.uni $L__BB155_176; $L__BB155_192: ld.global.v4.u32 {%r511, %r512, %r513, %r514}, [%rd157]; cvta.to.global.u64 %rd441, %rd240; mul.wide.u32 %rd442, %r511, 12; add.s64 %rd443, %rd441, %rd442; mul.wide.u32 %rd444, %r512, 12; add.s64 %rd445, %rd441, %rd444; mul.wide.u32 %rd446, %r513, 12; add.s64 %rd447, %rd441, %rd446; sub.ftz.f32 %f2166, %f716, %f533; sub.ftz.f32 %f2167, %f2166, %f534; ld.global.f32 %f2168, [%rd443]; ld.global.f32 %f2169, [%rd443+4]; ld.global.f32 %f2170, [%rd443+8]; ld.global.f32 %f2171, [%rd445]; mul.ftz.f32 %f2172, %f533, %f2171; ld.global.f32 %f2173, [%rd445+4]; mul.ftz.f32 %f2174, %f533, %f2173; ld.global.f32 %f2175, [%rd445+8]; mul.ftz.f32 %f2176, %f533, %f2175; fma.rn.ftz.f32 %f2177, %f2167, %f2168, %f2172; fma.rn.ftz.f32 %f2178, %f2167, %f2169, %f2174; fma.rn.ftz.f32 %f2179, %f2167, %f2170, %f2176; ld.global.f32 %f2180, [%rd447]; ld.global.f32 %f2181, [%rd447+4]; ld.global.f32 %f2182, [%rd447+8]; fma.rn.ftz.f32 %f2565, %f534, %f2180, %f2177; fma.rn.ftz.f32 %f2566, %f534, %f2181, %f2178; fma.rn.ftz.f32 %f2567, %f534, %f2182, %f2179; mul.ftz.f32 %f2183, %f2566, %f2566; fma.rn.ftz.f32 %f2184, %f2565, %f2565, %f2183; fma.rn.ftz.f32 %f2185, %f2567, %f2567, %f2184; sqrt.approx.ftz.f32 %f564, %f2185; setp.neu.ftz.f32 %p241, %f564, 0f00000000; @%p241 bra $L__BB155_193; bra.uni $L__BB155_194; $L__BB155_193: rcp.approx.ftz.f32 %f2186, %f564; mul.ftz.f32 %f2567, %f2567, %f2186; mul.ftz.f32 %f2566, %f2566, %f2186; mul.ftz.f32 %f2565, %f2565, %f2186; $L__BB155_194: setp.eq.ftz.f32 %p242, %f2565, 0f00000000; setp.eq.ftz.f32 %p243, %f2566, 0f00000000; and.pred %p244, %p242, %p243; setp.eq.ftz.f32 %p245, %f2567, 0f00000000; and.pred %p246, %p245, %p244; selp.f32 %f2570, %f2576, %f2565, %p246; selp.f32 %f2569, %f2575, %f2566, %p246; selp.f32 %f2568, %f2574, %f2567, %p246; mov.u32 %r611, 1; bra.uni $L__BB155_195; $L__BB155_176: setp.ne.s32 %p225, %r161, 65; mov.u32 %r611, %r161; mov.f32 %f2568, %f2574; mov.f32 %f2569, %f2575; mov.f32 %f2570, %f2576; @%p225 bra $L__BB155_195; ld.global.u32 %r187, [%rd92+12]; ld.global.u32 %r188, [%rd92+8]; shl.b32 %r489, %r188, 1; cvt.rn.f32.s32 %f2121, %r489; mul.ftz.f32 %f535, %f101, %f2121; cvt.rzi.ftz.s32.f32 %r490, %f535; add.s32 %r491, %r489, -1; min.s32 %r189, %r490, %r491; ld.global.u32 %r606, [%rd92+20]; mul.wide.u32 %rd423, %r606, 16; add.s64 %rd424, %rd12, %rd423; ld.global.u16 %rs39, [%rd424+12]; ld.global.u32 %r605, [%rd424+8]; ld.global.u64 %rd508, [%rd424]; setp.eq.s64 %p226, %rd508, 12; @%p226 bra $L__BB155_182; $L__BB155_178: setp.eq.s64 %p227, %rd508, 0; @%p227 bra $L__BB155_180; add.s32 %r606, %r606, 2; bra.uni $L__BB155_181; $L__BB155_180: setp.eq.s16 %p228, %rs39, 0; mov.u32 %r609, -1; mov.u32 %r606, %r605; @%p228 bra $L__BB155_183; $L__BB155_181: mul.wide.u32 %rd425, %r606, 16; add.s64 %rd426, %rd12, %rd425; ld.global.u16 %rs39, [%rd426+12]; ld.global.u32 %r605, [%rd426+8]; ld.global.u64 %rd508, [%rd426]; setp.ne.s64 %p229, %rd508, 12; @%p229 bra $L__BB155_178; $L__BB155_182: setp.eq.s16 %p230, %rs39, 0; selp.b32 %r609, -1, %r605, %p230; $L__BB155_183: ld.global.v4.u32 {%r493, %r494, %r495, %r496}, [%rd157]; setp.eq.s32 %p231, %r189, %r188; @%p231 bra $L__BB155_185; bra.uni $L__BB155_184; $L__BB155_185: cvta.to.global.u64 %rd430, %rd240; mul.wide.u32 %rd431, %r493, 12; add.s64 %rd509, %rd430, %rd431; mul.wide.u32 %rd432, %r494, 12; add.s64 %rd512, %rd430, %rd432; mul.wide.u32 %rd433, %r495, 12; add.s64 %rd515, %rd430, %rd433; add.s64 %rd510, %rd509, 4; add.s64 %rd511, %rd509, 8; add.s64 %rd513, %rd512, 4; add.s64 %rd514, %rd512, 8; add.s64 %rd516, %rd515, 4; add.s64 %rd517, %rd515, 8; bra.uni $L__BB155_186; $L__BB155_184: setp.gt.s32 %p232, %r189, %r188; selp.b32 %r497, -1, 0, %p232; add.s32 %r498, %r189, %r497; mad.lo.s32 %r499, %r498, %r187, %r609; add.s32 %r500, %r493, %r499; mul.wide.u32 %rd427, %r500, 12; add.s64 %rd509, %rd13, %rd427; add.s32 %r501, %r494, %r499; mul.wide.u32 %rd428, %r501, 12; add.s64 %rd512, %rd13, %rd428; add.s32 %r502, %r495, %r499; mul.wide.u32 %rd429, %r502, 12; add.s64 %rd515, %rd13, %rd429; add.s64 %rd510, %rd509, 4; add.s64 %rd511, %rd509, 8; add.s64 %rd513, %rd512, 4; add.s64 %rd514, %rd512, 8; add.s64 %rd516, %rd515, 4; add.s64 %rd517, %rd515, 8; $L__BB155_186: ld.global.f32 %f536, [%rd517]; ld.global.f32 %f537, [%rd516]; ld.global.f32 %f538, [%rd515]; ld.global.f32 %f539, [%rd514]; ld.global.f32 %f540, [%rd513]; ld.global.f32 %f541, [%rd512]; ld.global.f32 %f542, [%rd511]; ld.global.f32 %f543, [%rd510]; ld.global.f32 %f544, [%rd509]; add.s32 %r203, %r189, 1; setp.eq.s32 %p233, %r203, %r188; @%p233 bra $L__BB155_188; bra.uni $L__BB155_187; $L__BB155_188: cvta.to.global.u64 %rd437, %rd240; mul.wide.u32 %rd438, %r493, 12; add.s64 %rd518, %rd437, %rd438; mul.wide.u32 %rd439, %r494, 12; add.s64 %rd521, %rd437, %rd439; mul.wide.u32 %rd440, %r495, 12; add.s64 %rd524, %rd437, %rd440; add.s64 %rd519, %rd518, 4; add.s64 %rd520, %rd518, 8; add.s64 %rd522, %rd521, 4; add.s64 %rd523, %rd521, 8; add.s64 %rd525, %rd524, 4; add.s64 %rd526, %rd524, 8; bra.uni $L__BB155_189; $L__BB155_187: setp.ge.s32 %p234, %r189, %r188; selp.b32 %r503, -1, 0, %p234; add.s32 %r504, %r203, %r503; mad.lo.s32 %r505, %r504, %r187, %r609; add.s32 %r506, %r493, %r505; mul.wide.u32 %rd434, %r506, 12; add.s64 %rd518, %rd13, %rd434; add.s32 %r507, %r494, %r505; mul.wide.u32 %rd435, %r507, 12; add.s64 %rd521, %rd13, %rd435; add.s32 %r508, %r495, %r505; mul.wide.u32 %rd436, %r508, 12; add.s64 %rd524, %rd13, %rd436; add.s64 %rd519, %rd518, 4; add.s64 %rd520, %rd518, 8; add.s64 %rd522, %rd521, 4; add.s64 %rd523, %rd521, 8; add.s64 %rd525, %rd524, 4; add.s64 %rd526, %rd524, 8; $L__BB155_189: cvt.rn.f32.s32 %f2122, %r189; sub.ftz.f32 %f2123, %f535, %f2122; sub.ftz.f32 %f2125, %f716, %f2123; ld.global.f32 %f2126, [%rd518]; mul.ftz.f32 %f2127, %f2123, %f2126; ld.global.f32 %f2128, [%rd519]; mul.ftz.f32 %f2129, %f2123, %f2128; ld.global.f32 %f2130, [%rd520]; mul.ftz.f32 %f2131, %f2123, %f2130; fma.rn.ftz.f32 %f2132, %f2125, %f544, %f2127; fma.rn.ftz.f32 %f2133, %f2125, %f543, %f2129; fma.rn.ftz.f32 %f2134, %f2125, %f542, %f2131; ld.global.f32 %f2135, [%rd521]; mul.ftz.f32 %f2136, %f2123, %f2135; ld.global.f32 %f2137, [%rd522]; mul.ftz.f32 %f2138, %f2123, %f2137; ld.global.f32 %f2139, [%rd523]; mul.ftz.f32 %f2140, %f2123, %f2139; fma.rn.ftz.f32 %f2141, %f2125, %f541, %f2136; fma.rn.ftz.f32 %f2142, %f2125, %f540, %f2138; fma.rn.ftz.f32 %f2143, %f2125, %f539, %f2140; ld.global.f32 %f2144, [%rd524]; mul.ftz.f32 %f2145, %f2123, %f2144; ld.global.f32 %f2146, [%rd525]; mul.ftz.f32 %f2147, %f2123, %f2146; ld.global.f32 %f2148, [%rd526]; mul.ftz.f32 %f2149, %f2123, %f2148; fma.rn.ftz.f32 %f2150, %f2125, %f538, %f2145; fma.rn.ftz.f32 %f2151, %f2125, %f537, %f2147; fma.rn.ftz.f32 %f2152, %f2125, %f536, %f2149; sub.ftz.f32 %f2153, %f716, %f533; sub.ftz.f32 %f2154, %f2153, %f534; mul.ftz.f32 %f2155, %f533, %f2132; mul.ftz.f32 %f2156, %f533, %f2133; mul.ftz.f32 %f2157, %f533, %f2134; fma.rn.ftz.f32 %f2158, %f534, %f2141, %f2155; fma.rn.ftz.f32 %f2159, %f534, %f2142, %f2156; fma.rn.ftz.f32 %f2160, %f534, %f2143, %f2157; fma.rn.ftz.f32 %f2562, %f2154, %f2150, %f2158; fma.rn.ftz.f32 %f2563, %f2154, %f2151, %f2159; fma.rn.ftz.f32 %f2564, %f2154, %f2152, %f2160; mul.ftz.f32 %f2161, %f2562, %f2562; fma.rn.ftz.f32 %f2162, %f2563, %f2563, %f2161; fma.rn.ftz.f32 %f2163, %f2564, %f2564, %f2162; sqrt.approx.ftz.f32 %f548, %f2163; setp.neu.ftz.f32 %p235, %f548, 0f00000000; @%p235 bra $L__BB155_190; bra.uni $L__BB155_191; $L__BB155_190: rcp.approx.ftz.f32 %f2164, %f548; mul.ftz.f32 %f2564, %f2164, %f2564; mul.ftz.f32 %f2563, %f2164, %f2563; mul.ftz.f32 %f2562, %f2164, %f2562; $L__BB155_191: setp.eq.ftz.f32 %p236, %f2562, 0f00000000; setp.eq.ftz.f32 %p237, %f2563, 0f00000000; and.pred %p238, %p236, %p237; setp.eq.ftz.f32 %p239, %f2564, 0f00000000; and.pred %p240, %p239, %p238; selp.f32 %f2570, %f2576, %f2562, %p240; selp.f32 %f2569, %f2575, %f2563, %p240; selp.f32 %f2568, %f2574, %f2564, %p240; mov.u32 %r611, 65; $L__BB155_195: and.b32 %r520, %r186, 4; setp.ne.s32 %p247, %r520, 0; @%p247 bra $L__BB155_202; ld.local.u32 %r521, [%rd2+56]; and.b32 %r206, %r521, 2; setp.eq.s32 %p248, %r206, 0; @%p248 bra $L__BB155_198; bra.uni $L__BB155_197; $L__BB155_198: ld.global.v4.f32 {%f2221, %f2222, %f2223, %f2224}, [%rd376+48]; ld.global.v4.f32 {%f2228, %f2229, %f2230, %f2231}, [%rd376+64]; ld.global.v4.f32 {%f2235, %f2236, %f2237, %f2238}, [%rd376+80]; mul.ftz.f32 %f2242, %f2569, %f2228; fma.rn.ftz.f32 %f2243, %f2570, %f2221, %f2242; fma.rn.ftz.f32 %f2244, %f2568, %f2235, %f2243; mul.ftz.f32 %f2245, %f2569, %f2229; fma.rn.ftz.f32 %f2246, %f2570, %f2222, %f2245; fma.rn.ftz.f32 %f2247, %f2568, %f2236, %f2246; mul.ftz.f32 %f2248, %f2569, %f2230; fma.rn.ftz.f32 %f2249, %f2570, %f2223, %f2248; fma.rn.ftz.f32 %f2250, %f2568, %f2237, %f2249; mul.ftz.f32 %f2251, %f2247, %f2247; fma.rn.ftz.f32 %f2252, %f2244, %f2244, %f2251; fma.rn.ftz.f32 %f2253, %f2250, %f2250, %f2252; sqrt.approx.ftz.f32 %f2254, %f2253; div.approx.ftz.f32 %f2570, %f2244, %f2254; div.approx.ftz.f32 %f2569, %f2247, %f2254; div.approx.ftz.f32 %f2568, %f2250, %f2254; bra.uni $L__BB155_199; $L__BB155_197: ld.local.v4.f32 {%f2187, %f2188, %f2189, %f2190}, [%rd2+192]; ld.local.v4.f32 {%f2194, %f2195, %f2196, %f2197}, [%rd2+208]; ld.local.v4.f32 {%f2201, %f2202, %f2203, %f2204}, [%rd2+224]; mul.ftz.f32 %f2208, %f2569, %f2194; fma.rn.ftz.f32 %f2209, %f2570, %f2187, %f2208; fma.rn.ftz.f32 %f2210, %f2568, %f2201, %f2209; mul.ftz.f32 %f2211, %f2569, %f2195; fma.rn.ftz.f32 %f2212, %f2570, %f2188, %f2211; fma.rn.ftz.f32 %f2213, %f2568, %f2202, %f2212; mul.ftz.f32 %f2214, %f2569, %f2196; fma.rn.ftz.f32 %f2215, %f2570, %f2189, %f2214; fma.rn.ftz.f32 %f2216, %f2568, %f2203, %f2215; mul.ftz.f32 %f2217, %f2213, %f2213; fma.rn.ftz.f32 %f2218, %f2210, %f2210, %f2217; fma.rn.ftz.f32 %f2219, %f2216, %f2216, %f2218; sqrt.approx.ftz.f32 %f2220, %f2219; div.approx.ftz.f32 %f2570, %f2210, %f2220; div.approx.ftz.f32 %f2569, %f2213, %f2220; div.approx.ftz.f32 %f2568, %f2216, %f2220; $L__BB155_199: @%p248 bra $L__BB155_201; bra.uni $L__BB155_200; $L__BB155_201: ld.global.v4.f32 {%f2289, %f2290, %f2291, %f2292}, [%rd376+48]; ld.global.v4.f32 {%f2296, %f2297, %f2298, %f2299}, [%rd376+64]; ld.global.v4.f32 {%f2303, %f2304, %f2305, %f2306}, [%rd376+80]; mul.ftz.f32 %f2310, %f2575, %f2296; fma.rn.ftz.f32 %f2311, %f2576, %f2289, %f2310; fma.rn.ftz.f32 %f2312, %f2574, %f2303, %f2311; mul.ftz.f32 %f2313, %f2575, %f2297; fma.rn.ftz.f32 %f2314, %f2576, %f2290, %f2313; fma.rn.ftz.f32 %f2315, %f2574, %f2304, %f2314; mul.ftz.f32 %f2316, %f2575, %f2298; fma.rn.ftz.f32 %f2317, %f2576, %f2291, %f2316; fma.rn.ftz.f32 %f2318, %f2574, %f2305, %f2317; mul.ftz.f32 %f2319, %f2315, %f2315; fma.rn.ftz.f32 %f2320, %f2312, %f2312, %f2319; fma.rn.ftz.f32 %f2321, %f2318, %f2318, %f2320; sqrt.approx.ftz.f32 %f2322, %f2321; div.approx.ftz.f32 %f2576, %f2312, %f2322; div.approx.ftz.f32 %f2575, %f2315, %f2322; div.approx.ftz.f32 %f2574, %f2318, %f2322; bra.uni $L__BB155_202; $L__BB155_200: ld.local.v4.f32 {%f2255, %f2256, %f2257, %f2258}, [%rd2+192]; ld.local.v4.f32 {%f2262, %f2263, %f2264, %f2265}, [%rd2+208]; ld.local.v4.f32 {%f2269, %f2270, %f2271, %f2272}, [%rd2+224]; mul.ftz.f32 %f2276, %f2575, %f2262; fma.rn.ftz.f32 %f2277, %f2576, %f2255, %f2276; fma.rn.ftz.f32 %f2278, %f2574, %f2269, %f2277; mul.ftz.f32 %f2279, %f2575, %f2263; fma.rn.ftz.f32 %f2280, %f2576, %f2256, %f2279; fma.rn.ftz.f32 %f2281, %f2574, %f2270, %f2280; mul.ftz.f32 %f2282, %f2575, %f2264; fma.rn.ftz.f32 %f2283, %f2576, %f2257, %f2282; fma.rn.ftz.f32 %f2284, %f2574, %f2271, %f2283; mul.ftz.f32 %f2285, %f2281, %f2281; fma.rn.ftz.f32 %f2286, %f2278, %f2278, %f2285; fma.rn.ftz.f32 %f2287, %f2284, %f2284, %f2286; sqrt.approx.ftz.f32 %f2288, %f2287; div.approx.ftz.f32 %f2576, %f2278, %f2288; div.approx.ftz.f32 %f2575, %f2281, %f2288; div.approx.ftz.f32 %f2574, %f2284, %f2288; $L__BB155_202: mul.ftz.f32 %f2323, %f2410, %f2576; fma.rn.ftz.f32 %f2324, %f2409, %f2575, %f2323; fma.rn.ftz.f32 %f2325, %f2408, %f2574, %f2324; abs.ftz.f32 %f2326, %f2325; mul.ftz.f32 %f2327, %f2417, %f2326; mul.ftz.f32 %f2328, %f2413, %f2576; fma.rn.ftz.f32 %f2329, %f2412, %f2575, %f2328; fma.rn.ftz.f32 %f2330, %f2411, %f2574, %f2329; abs.ftz.f32 %f2331, %f2330; mul.ftz.f32 %f2332, %f2418, %f2331; mul.ftz.f32 %f2333, %f2416, %f2576; fma.rn.ftz.f32 %f2334, %f2415, %f2575, %f2333; fma.rn.ftz.f32 %f2335, %f2414, %f2574, %f2334; abs.ftz.f32 %f2336, %f2335; mul.ftz.f32 %f2337, %f2419, %f2336; mul.ftz.f32 %f2338, %f2332, %f2332; fma.rn.ftz.f32 %f2339, %f2327, %f2327, %f2338; fma.rn.ftz.f32 %f2340, %f2337, %f2337, %f2339; div.approx.ftz.f32 %f2580, %f2327, %f2340; setp.lt.s32 %p250, %r160, 5; @%p250 bra $L__BB155_204; mov.f32 %f2341, 0f40800000; div.approx.ftz.f32 %f2342, %f407, %f2341; mul.ftz.f32 %f2580, %f2580, %f2342; $L__BB155_204: sub.ftz.f32 %f2344, %f2561, %f91; sub.ftz.f32 %f2345, %f2560, %f92; mul.ftz.f32 %f2346, %f2345, %f2345; fma.rn.ftz.f32 %f2347, %f2344, %f2344, %f2346; sub.ftz.f32 %f2348, %f2559, %f93; fma.rn.ftz.f32 %f2349, %f2348, %f2348, %f2347; sqrt.approx.ftz.f32 %f604, %f2349; setp.ge.ftz.f32 %p251, %f604, %f1; mov.f32 %f2582, 0f00000000; mov.f32 %f2581, %f2582; @%p251 bra $L__BB155_206; sub.ftz.f32 %f2350, %f1, %f604; mul.ftz.f32 %f2351, %f2350, %f2350; mul.ftz.f32 %f2352, %f2350, %f2351; mul.ftz.f32 %f2353, %f2352, 0f41200000; div.approx.ftz.f32 %f2581, %f2353, %f5; $L__BB155_206: setp.ge.ftz.f32 %p252, %f90, %f1; @%p252 bra $L__BB155_208; div.approx.ftz.f32 %f2582, %f408, %f5; $L__BB155_208: div.approx.ftz.f32 %f2355, %f2581, %f2582; mul.ftz.f32 %f2356, %f2580, %f2355; fma.rn.ftz.f32 %f2592, %f2570, %f2356, %f2592; fma.rn.ftz.f32 %f2593, %f2569, %f2356, %f2593; fma.rn.ftz.f32 %f2594, %f2568, %f2356, %f2594; cvt.u32.u64 %r522, %rd94; add.s32 %r599, %r522, 1; setp.lt.s32 %p253, %r599, %r162; @%p253 bra $L__BB155_150; $L__BB155_209: add.s32 %r539, %r539, 1; setp.lt.u32 %p254, %r539, %r2; @%p254 bra $L__BB155_5; mul.ftz.f32 %f2357, %f2592, %f2592; fma.rn.ftz.f32 %f2358, %f2593, %f2593, %f2357; fma.rn.ftz.f32 %f2359, %f2594, %f2594, %f2358; sqrt.approx.ftz.f32 %f618, %f2359; setp.neu.ftz.f32 %p255, %f618, 0f00000000; @%p255 bra $L__BB155_211; bra.uni $L__BB155_212; $L__BB155_211: rcp.approx.ftz.f32 %f2360, %f618; mul.ftz.f32 %f2594, %f2594, %f2360; mul.ftz.f32 %f2593, %f2593, %f2360; mul.ftz.f32 %f2592, %f2592, %f2360; $L__BB155_212: setp.eq.ftz.f32 %p256, %f2592, 0f00000000; setp.eq.ftz.f32 %p257, %f2593, 0f00000000; and.pred %p258, %p256, %p257; setp.eq.ftz.f32 %p259, %f2594, 0f00000000; and.pred %p260, %p259, %p258; @%p260 bra $L__BB155_215; bra.uni $L__BB155_213; $L__BB155_215: ld.local.f32 %f2594, [%rd2+20]; ld.local.f32 %f2593, [%rd2+16]; ld.local.f32 %f2592, [%rd2+12]; bra.uni $L__BB155_219; $L__BB155_213: ld.local.u32 %r524, [%rd2+52]; and.b32 %r525, %r524, 1; setp.eq.b32 %p261, %r525, 1; mov.pred %p262, 0; xor.pred %p263, %p261, %p262; not.pred %p264, %p263; @%p264 bra $L__BB155_219; neg.ftz.f32 %f2594, %f2594; neg.ftz.f32 %f2593, %f2593; neg.ftz.f32 %f2592, %f2592; bra.uni $L__BB155_219; } .func _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b64 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b64 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .align 4 .b8 _Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3[16] ) { .local .align 16 .b8 __local_depot156[864]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<174>; .reg .b16 %rs<61>; .reg .f32 %f<1420>; .reg .b32 %r<445>; .reg .b64 %rd<250>; mov.u64 %SPL, __local_depot156; cvta.local.u64 %SP, %SPL; ld.param.u32 %r134, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd88, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.u64 %rd89, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u32 %r138, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+12]; ld.param.u32 %r137, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+8]; ld.param.u32 %r136, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; cvta.to.local.u64 %rd1, %rd89; add.u64 %rd2, %SPL, 0; bfe.u32 %r2, %r136, 8, 8; bfe.u32 %r3, %r136, 16, 8; bfe.u32 %r7, %r137, 16, 8; setp.eq.s32 %p1, %r2, 255; @%p1 bra $L__BB156_2; bra.uni $L__BB156_1; $L__BB156_2: mov.b32 %f1368, %r138; bra.uni $L__BB156_3; $L__BB156_1: mul.wide.u32 %rd91, %r2, 4; add.s64 %rd92, %rd1, %rd91; ld.local.f32 %f1368, [%rd92]; $L__BB156_3: cvta.to.local.u64 %rd93, %rd88; add.s64 %rd3, %rd93, 12; setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB156_5; bra.uni $L__BB156_4; $L__BB156_5: ld.local.f32 %f1371, [%rd3+8]; ld.local.f32 %f1370, [%rd3+4]; ld.local.f32 %f1369, [%rd3]; bra.uni $L__BB156_6; $L__BB156_4: mul.wide.u32 %rd94, %r3, 4; add.s64 %rd95, %rd1, %rd94; ld.local.f32 %f1371, [%rd95+8]; ld.local.f32 %f1370, [%rd95+4]; ld.local.f32 %f1369, [%rd95]; $L__BB156_6: and.b32 %r139, %r136, 4; setp.eq.s32 %p3, %r139, 0; ld.const.f32 %f246, [kernel_params+1540]; selp.f32 %f13, %f1368, %f246, %p3; setp.le.ftz.f32 %p4, %f13, 0f00000000; setp.eq.s32 %p5, %r7, 0; or.pred %p6, %p5, %p4; mov.f32 %f1419, 0f3F800000; @%p6 bra $L__BB156_130; ld.local.u32 %r140, [%rd3+64]; setp.eq.s32 %p7, %r140, -1; ld.const.u32 %r141, [kernel_params+1124]; setp.eq.s32 %p8, %r141, 0; or.pred %p9, %p7, %p8; @%p9 bra $L__BB156_130; and.b32 %r144, %r136, 2; setp.eq.s32 %p10, %r144, 0; mov.u32 %r407, 0; neg.ftz.f32 %f248, %f1371; selp.f32 %f14, %f1371, %f248, %p10; neg.ftz.f32 %f249, %f1370; selp.f32 %f15, %f1370, %f249, %p10; neg.ftz.f32 %f250, %f1369; selp.f32 %f16, %f1369, %f250, %p10; setp.neu.ftz.f32 %p11, %f16, %f15; setp.neu.ftz.f32 %p12, %f16, %f14; or.pred %p13, %p11, %p12; sub.ftz.f32 %f251, %f14, %f15; neg.ftz.f32 %f252, %f15; selp.f32 %f253, %f15, %f252, %p13; neg.ftz.f32 %f254, %f14; selp.f32 %f255, %f254, %f14, %p13; add.ftz.f32 %f256, %f16, %f255; sub.ftz.f32 %f257, %f253, %f16; mul.ftz.f32 %f258, %f256, %f256; fma.rn.ftz.f32 %f259, %f251, %f251, %f258; fma.rn.ftz.f32 %f260, %f257, %f257, %f259; sqrt.approx.ftz.f32 %f261, %f260; div.approx.ftz.f32 %f17, %f251, %f261; div.approx.ftz.f32 %f18, %f256, %f261; div.approx.ftz.f32 %f19, %f257, %f261; mul.ftz.f32 %f262, %f15, %f19; mul.ftz.f32 %f263, %f14, %f18; sub.ftz.f32 %f20, %f262, %f263; mul.ftz.f32 %f264, %f14, %f17; mul.ftz.f32 %f265, %f16, %f19; sub.ftz.f32 %f21, %f264, %f265; mul.ftz.f32 %f266, %f16, %f18; mul.ftz.f32 %f267, %f15, %f17; sub.ftz.f32 %f22, %f266, %f267; ld.const.v2.u64 {%rd96, %rd97}, [kernel_params+2528]; cvta.to.global.u64 %rd100, %rd96; mul.wide.s32 %rd101, %r134, 4; add.s64 %rd102, %rd100, %rd101; cvta.to.global.u64 %rd103, %rd97; mul.wide.s32 %rd104, %r134, 2; add.s64 %rd105, %rd103, %rd104; ld.const.u64 %rd106, [kernel_params+2456]; cvta.to.global.u64 %rd107, %rd106; add.s64 %rd108, %rd107, %rd101; ld.global.u16 %rs24, [%rd105]; cvt.u32.u16 %r145, %rs24; ld.global.u32 %r146, [%rd108]; mul.lo.s32 %r8, %r146, %r7; add.s32 %r147, %r145, 4; ld.const.u32 %r9, [kernel_params+1596]; shr.u32 %r148, %r147, 16; xor.b32 %r149, %r148, %r147; mul.lo.s32 %r150, %r149, 569420461; shr.u32 %r151, %r150, 15; xor.b32 %r152, %r151, %r150; mul.lo.s32 %r153, %r152, -749064809; shr.u32 %r154, %r153, 15; ld.global.u32 %r155, [%rd102]; xor.b32 %r156, %r155, %r153; xor.b32 %r157, %r156, %r154; xor.b32 %r10, %r157, 508809841; shr.u32 %r158, %r10, 16; or.b32 %r11, %r158, 1; xor.b32 %r12, %r157, 106206365; shr.u32 %r159, %r12, 16; or.b32 %r13, %r159, 1; xor.b32 %r14, %r157, 1919448741; ld.const.v2.u32 {%r160, %r161}, [kernel_params+1600]; mov.b32 %f23, %r161; setp.lt.ftz.f32 %p14, %f23, 0f3F800000; ld.const.u32 %r164, [kernel_params+1576]; selp.b32 %r165, %r164, %r155, %p14; add.s16 %rs25, %rs24, 4; and.b16 %rs26, %rs25, 255; cvt.u32.u16 %r166, %rs26; xor.b32 %r167, %r165, %r166; mul.lo.s32 %r168, %r167, -512718531; shr.u32 %r169, %r165, 16; xor.b32 %r170, %r168, %r169; bfe.u32 %r171, %r170, 4, 4; shr.u32 %r172, %r165, 8; xor.b32 %r173, %r170, %r172; xor.b32 %r174, %r173, %r171; mul.lo.s32 %r175, %r174, 153742143; shr.u32 %r176, %r165, 23; xor.b32 %r177, %r175, %r176; bfe.u32 %r178, %r177, 1, 7; xor.b32 %r179, %r178, %r177; shr.u32 %r180, %r165, 27; or.b32 %r181, %r180, 1; mul.lo.s32 %r182, %r181, %r179; mul.lo.s32 %r183, %r182, -1074439621; bfe.u32 %r184, %r183, 2, 6; xor.b32 %r185, %r184, %r183; mul.lo.s32 %r186, %r185, -1638916925; bfe.u32 %r187, %r186, 2, 6; xor.b32 %r188, %r187, %r186; mul.lo.s32 %r189, %r188, -933190689; and.b32 %r190, %r189, 255; bfe.u32 %r191, %r189, 5, 3; xor.b32 %r192, %r191, %r190; xor.b32 %r193, %r147, 61; xor.b32 %r194, %r165, %r193; mul.lo.s32 %r195, %r194, 9; shr.u32 %r196, %r195, 4; xor.b32 %r197, %r196, %r195; mul.lo.s32 %r16, %r197, 668265261; shr.u32 %r198, %r16, 16; or.b32 %r17, %r198, 1; mul.lo.s32 %r19, %r192, %r160; ld.const.u64 %rd109, [kernel_params+2424]; cvta.to.global.u64 %rd4, %rd109; shr.u32 %r199, %r14, 16; or.b32 %r21, %r199, 1; and.b32 %r22, %r136, 1; xor.b32 %r200, %r193, %r155; mul.lo.s32 %r201, %r200, 9; shr.u32 %r202, %r201, 4; xor.b32 %r203, %r202, %r201; mul.lo.s32 %r204, %r203, 668265261; cvt.rn.f32.u32 %f268, %r204; mul.ftz.f32 %f269, %f268, 0f2F7FFFFE; mul.ftz.f32 %f24, %f269, %f23; xor.b32 %r205, %r155, %r147; xor.b32 %r206, %r205, -905047700; mul.lo.s32 %r207, %r206, 9; shr.u32 %r208, %r207, 4; xor.b32 %r209, %r208, %r207; mul.lo.s32 %r210, %r209, 668265261; cvt.rn.f32.u32 %f270, %r210; mul.ftz.f32 %f271, %f270, 0f2F7FFFFE; mul.ftz.f32 %f25, %f271, %f23; ld.const.u32 %r23, [kernel_params+1108]; ld.const.u32 %r24, [kernel_params+1112]; ld.const.u64 %rd110, [kernel_params+2144]; cvta.to.global.u64 %rd5, %rd110; ld.const.v2.u64 {%rd111, %rd112}, [kernel_params+2176]; cvta.to.global.u64 %rd6, %rd112; ld.const.u64 %rd115, [kernel_params+2160]; cvta.to.global.u64 %rd7, %rd115; ld.const.v2.u64 {%rd116, %rd117}, [kernel_params+2096]; cvta.to.global.u64 %rd8, %rd116; cvta.to.global.u64 %rd9, %rd117; ld.const.v2.u64 {%rd120, %rd121}, [kernel_params+2128]; cvta.to.global.u64 %rd10, %rd121; cvta.to.global.u64 %rd11, %rd120; ld.const.u64 %rd124, [kernel_params+2232]; cvta.to.global.u64 %rd12, %rd124; ld.const.u64 %rd125, [kernel_params+2256]; cvta.to.global.u64 %rd13, %rd125; ld.const.u64 %rd126, [kernel_params+2312]; cvta.to.global.u64 %rd14, %rd126; ld.const.u64 %rd127, [kernel_params+2336]; cvta.to.global.u64 %rd15, %rd127; cvta.to.global.u64 %rd16, %rd111; add.u64 %rd128, %SP, 768; setp.eq.s32 %p15, %r9, 0; setp.eq.s32 %p23, %r22, 0; mov.u16 %rs43, 0; mov.u32 %r408, %r407; $L__BB156_9: add.s32 %r27, %r407, %r8; brev.b32 %r211, %r27; mul.lo.s32 %r212, %r211, 1025551850; xor.b32 %r28, %r212, %r211; @%p15 bra $L__BB156_12; shl.b32 %r395, %r160, 8; neg.s32 %r394, %r160; add.s32 %r393, %r160, -1; setp.geu.ftz.f32 %p16, %f23, 0f3F800000; add.s32 %r213, %r16, %r28; mul.lo.s32 %r214, %r17, %r213; mul.lo.s32 %r215, %r214, 89287766; xor.b32 %r216, %r215, %r214; mul.lo.s32 %r217, %r216, 1403136100; xor.b32 %r218, %r217, %r216; brev.b32 %r219, %r218; and.b32 %r220, %r219, %r393; and.b32 %r221, %r27, %r394; or.b32 %r222, %r220, %r221; add.s32 %r223, %r19, %r222; rem.u32 %r224, %r223, %r395; shl.b32 %r225, %r224, 1; mul.wide.u32 %rd129, %r225, 4; add.s64 %rd130, %rd4, %rd129; ld.global.f32 %f1374, [%rd130]; ld.global.f32 %f1375, [%rd130+4]; @%p16 bra $L__BB156_16; add.ftz.f32 %f272, %f1374, %f24; cvt.rmi.ftz.f32.f32 %f273, %f272; sub.ftz.f32 %f1374, %f272, %f273; add.ftz.f32 %f274, %f1375, %f25; cvt.rmi.ftz.f32.f32 %f275, %f274; sub.ftz.f32 %f1375, %f274, %f275; bra.uni $L__BB156_16; $L__BB156_12: xor.b32 %r400, %r157, 508809841; add.s32 %r227, %r28, %r400; mul.lo.s32 %r228, %r227, %r11; mul.lo.s32 %r229, %r228, 89287766; xor.b32 %r230, %r229, %r228; mul.lo.s32 %r231, %r230, 1403136100; xor.b32 %r29, %r231, %r230; setp.eq.s32 %p17, %r29, 0; mov.u32 %r412, 0; @%p17 bra $L__BB156_15; mov.u32 %r409, 0; mov.u32 %r410, %r29; mov.u32 %r412, %r409; $L__BB156_14: clz.b32 %r234, %r410; add.s32 %r235, %r234, %r409; mul.wide.u32 %rd131, %r235, 4; mov.u64 %rd132, sobol_burley_table; add.s64 %rd133, %rd132, %rd131; ld.const.u32 %r236, [%rd133+128]; xor.b32 %r412, %r236, %r412; add.s32 %r409, %r235, 1; shl.b32 %r237, %r410, %r234; shl.b32 %r410, %r237, 1; setp.ne.s32 %p18, %r410, 0; @%p18 bra $L__BB156_14; $L__BB156_15: xor.b32 %r402, %r157, 1919448741; xor.b32 %r401, %r157, 106206365; brev.b32 %r238, %r29; mul.lo.s32 %r239, %r238, 1025551850; xor.b32 %r240, %r239, %r238; add.s32 %r241, %r240, %r401; mul.lo.s32 %r242, %r241, %r13; mul.lo.s32 %r243, %r242, 89287766; xor.b32 %r244, %r243, %r242; mul.lo.s32 %r245, %r412, 1025551850; xor.b32 %r246, %r245, %r412; add.s32 %r247, %r246, %r402; mul.lo.s32 %r248, %r247, %r21; mul.lo.s32 %r249, %r248, 89287766; xor.b32 %r250, %r249, %r248; mul.lo.s32 %r251, %r250, 1403136100; xor.b32 %r252, %r251, %r250; brev.b32 %r253, %r252; cvt.rn.f32.u32 %f276, %r253; mul.lo.s32 %r254, %r244, 1403136100; xor.b32 %r255, %r254, %r244; brev.b32 %r256, %r255; cvt.rn.f32.u32 %f277, %r256; mul.ftz.f32 %f1375, %f276, 0f2F7FFFFE; mul.ftz.f32 %f1374, %f277, 0f2F7FFFFE; $L__BB156_16: fma.rn.ftz.f32 %f38, %f1374, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f1377, %f1375, 0f40000000, 0fBF800000; setp.eq.ftz.f32 %p19, %f38, 0f00000000; setp.eq.ftz.f32 %p20, %f1377, 0f00000000; and.pred %p21, %p19, %p20; mov.f32 %f279, 0f00000000; mov.f32 %f1378, %f279; mov.f32 %f1379, %f279; @%p21 bra $L__BB156_21; mul.ftz.f32 %f280, %f38, %f38; mul.ftz.f32 %f281, %f1377, %f1377; setp.gt.ftz.f32 %p22, %f280, %f281; @%p22 bra $L__BB156_19; bra.uni $L__BB156_18; $L__BB156_19: div.approx.ftz.f32 %f283, %f1377, %f38; mul.ftz.f32 %f1376, %f283, 0f3F490FDB; mov.f32 %f1377, %f38; bra.uni $L__BB156_20; $L__BB156_18: div.approx.ftz.f32 %f282, %f38, %f1377; fma.rn.ftz.f32 %f1376, %f282, 0fBF490FDB, 0f3FC90FDB; $L__BB156_20: cos.approx.ftz.f32 %f284, %f1376; sin.approx.ftz.f32 %f285, %f1376; mul.ftz.f32 %f1379, %f1377, %f285; mul.ftz.f32 %f1378, %f1377, %f284; $L__BB156_21: add.u64 %rd223, %SPL, 768; mul.ftz.f32 %f286, %f1379, %f1379; fma.rn.ftz.f32 %f287, %f1378, %f1378, %f286; mov.f32 %f288, 0f3F800000; sub.ftz.f32 %f289, %f288, %f287; max.ftz.f32 %f291, %f289, %f279; sqrt.approx.ftz.f32 %f292, %f291; ld.local.v4.f32 {%f1406, %f1407, %f1408, %f296}, [%rd93]; mov.u32 %r257, 0; st.local.f32 [%rd223], %f1406; st.local.f32 [%rd223+4], %f1407; st.local.f32 [%rd223+8], %f1408; mul.ftz.f32 %f297, %f20, %f1379; mul.ftz.f32 %f298, %f21, %f1379; mul.ftz.f32 %f299, %f22, %f1379; fma.rn.ftz.f32 %f300, %f17, %f1378, %f297; fma.rn.ftz.f32 %f301, %f18, %f1378, %f298; fma.rn.ftz.f32 %f302, %f19, %f1378, %f299; fma.rn.ftz.f32 %f51, %f16, %f292, %f300; fma.rn.ftz.f32 %f52, %f15, %f292, %f301; fma.rn.ftz.f32 %f53, %f14, %f292, %f302; st.local.f32 [%rd223+12], %f51; st.local.f32 [%rd223+16], %f52; st.local.f32 [%rd223+20], %f53; st.local.u32 [%rd223+24], %r257; st.local.f32 [%rd223+28], %f13; ld.local.f32 %f54, [%rd3+72]; st.local.f32 [%rd223+32], %f54; ld.local.u32 %r37, [%rd3+64]; st.local.u32 [%rd223+40], %r37; ld.local.u32 %r38, [%rd3+48]; st.local.u32 [%rd223+36], %r38; mov.u32 %r413, -1; st.local.u32 [%rd223+48], %r413; st.local.u32 [%rd223+44], %r413; st.local.u32 [%rd223+52], %r257; st.local.u32 [%rd223+56], %r257; mov.b32 %r39, %f1406; @%p23 bra $L__BB156_111; bra.uni $L__BB156_22; $L__BB156_111: setp.nan.ftz.f32 %p153, %f1406, %f1406; mov.u16 %rs57, %rs43; @%p153 bra $L__BB156_114; or.b32 %r370, %r39, -2147483648; setp.ne.s32 %p154, %r370, -2147483648; add.ftz.f32 %f1345, %f1406, %f1406; setp.eq.ftz.f32 %p155, %f1345, %f1406; and.pred %p156, %p155, %p154; mov.u16 %rs57, %rs43; @%p156 bra $L__BB156_114; shl.b32 %r371, %r39, 1; setp.lt.u32 %p157, %r371, -16777215; selp.u16 %rs57, 1, 0, %p157; $L__BB156_114: setp.eq.s16 %p158, %rs57, 0; mov.u16 %rs59, %rs43; @%p158 bra $L__BB156_120; mov.b32 %r131, %f51; setp.nan.ftz.f32 %p159, %f51, %f51; mov.u16 %rs59, 0; mov.u16 %rs58, %rs59; @%p159 bra $L__BB156_118; or.b32 %r372, %r131, -2147483648; setp.ne.s32 %p160, %r372, -2147483648; add.ftz.f32 %f1346, %f51, %f51; setp.eq.ftz.f32 %p161, %f1346, %f51; and.pred %p162, %p161, %p160; mov.u16 %rs58, %rs59; @%p162 bra $L__BB156_118; shl.b32 %r373, %r131, 1; setp.lt.u32 %p163, %r373, -16777215; selp.u16 %rs58, 1, 0, %p163; $L__BB156_118: setp.eq.s16 %p164, %rs58, 0; @%p164 bra $L__BB156_120; mul.ftz.f32 %f1347, %f52, %f52; fma.rn.ftz.f32 %f1348, %f51, %f51, %f1347; fma.rn.ftz.f32 %f1349, %f53, %f53, %f1348; setp.neu.ftz.f32 %p165, %f1349, 0f00000000; selp.u16 %rs59, 1, 0, %p165; $L__BB156_120: setp.eq.s16 %p166, %rs59, 0; mov.u16 %rs60, %rs43; @%p166 bra $L__BB156_128; setp.eq.s32 %p167, %r23, 0; @%p167 bra $L__BB156_125; setp.ne.s32 %p168, %r24, 0; @%p168 bra $L__BB156_124; bra.uni $L__BB156_123; $L__BB156_124: mov.u32 %r376, 256; add.u64 %rd207, %SP, 828; { // callseq 916, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd128; .param .b64 param1; st.param.b64 [param1+0], %rd207; .param .b32 param2; st.param.b32 [param2+0], %r376; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r377, [retval0+0]; } // callseq 916 cvt.u16.u32 %rs60, %r377; bra.uni $L__BB156_128; $L__BB156_22: setp.nan.ftz.f32 %p24, %f1406, %f1406; mov.u16 %rs60, 0; mov.u16 %rs50, %rs60; @%p24 bra $L__BB156_25; or.b32 %r259, %r39, -2147483648; setp.ne.s32 %p25, %r259, -2147483648; add.ftz.f32 %f303, %f1406, %f1406; setp.eq.ftz.f32 %p26, %f303, %f1406; and.pred %p27, %p26, %p25; mov.u16 %rs50, %rs60; @%p27 bra $L__BB156_25; shl.b32 %r260, %r39, 1; setp.lt.u32 %p28, %r260, -16777215; selp.u16 %rs50, 1, 0, %p28; $L__BB156_25: setp.eq.s16 %p29, %rs50, 0; mov.u16 %rs52, %rs60; @%p29 bra $L__BB156_31; mov.b32 %r40, %f51; setp.nan.ftz.f32 %p30, %f51, %f51; mov.u16 %rs52, 0; mov.u16 %rs51, %rs52; @%p30 bra $L__BB156_29; or.b32 %r261, %r40, -2147483648; setp.ne.s32 %p31, %r261, -2147483648; add.ftz.f32 %f304, %f51, %f51; setp.eq.ftz.f32 %p32, %f304, %f51; and.pred %p33, %p32, %p31; mov.u16 %rs51, %rs52; @%p33 bra $L__BB156_29; shl.b32 %r262, %r40, 1; setp.lt.u32 %p34, %r262, -16777215; selp.u16 %rs51, 1, 0, %p34; $L__BB156_29: setp.eq.s16 %p35, %rs51, 0; @%p35 bra $L__BB156_31; mul.ftz.f32 %f305, %f52, %f52; fma.rn.ftz.f32 %f306, %f51, %f51, %f305; fma.rn.ftz.f32 %f307, %f53, %f53, %f306; setp.neu.ftz.f32 %p36, %f307, 0f00000000; selp.u16 %rs52, 1, 0, %p36; $L__BB156_31: setp.eq.s16 %p37, %rs52, 0; @%p37 bra $L__BB156_128; setp.eq.s32 %p38, %r23, 0; abs.ftz.f32 %f308, %f51; setp.gt.ftz.f32 %p39, %f308, 0f17800000; mov.b32 %r263, %f51; and.b32 %r264, %r263, -2147483648; or.b32 %r265, %r264, 394264576; mov.b32 %f309, %r265; selp.f32 %f1403, %f51, %f309, %p39; abs.ftz.f32 %f310, %f52; setp.gt.ftz.f32 %p40, %f310, 0f17800000; mov.b32 %r266, %f52; and.b32 %r267, %r266, -2147483648; or.b32 %r268, %r267, 394264576; mov.b32 %f311, %r268; selp.f32 %f1404, %f52, %f311, %p40; abs.ftz.f32 %f312, %f53; setp.gt.ftz.f32 %p41, %f312, 0f17800000; mov.b32 %r269, %f53; and.b32 %r270, %r269, -2147483648; or.b32 %r271, %r270, 394264576; mov.b32 %f313, %r271; selp.f32 %f1405, %f53, %f313, %p41; mul.wide.s32 %rd135, %r37, 4; add.s64 %rd18, %rd5, %rd135; add.s64 %rd19, %rd6, %rd135; mul.wide.s32 %rd136, %r37, 208; add.s64 %rd137, %rd7, %rd136; add.s64 %rd21, %rd137, 48; @%p38 bra $L__BB156_87; mov.u32 %r273, 1985229328; st.local.u32 [%rd2], %r273; ld.global.u32 %r420, [%rd18]; ld.global.u32 %r42, [%rd19]; and.b32 %r274, %r42, 4; setp.ne.s32 %p42, %r274, 0; @%p42 bra $L__BB156_43; and.b32 %r275, %r42, 2; setp.eq.s32 %p43, %r275, 0; @%p43 bra $L__BB156_41; bra.uni $L__BB156_35; $L__BB156_41: ld.global.v4.f32 {%f1391, %f1392, %f1393, %f1394}, [%rd21]; ld.global.v4.f32 {%f1395, %f1396, %f1397, %f1398}, [%rd21+16]; ld.global.v4.f32 {%f1399, %f1400, %f1401, %f1402}, [%rd21+32]; bra.uni $L__BB156_42; $L__BB156_125: setp.ne.s32 %p169, %r24, 0; @%p169 bra $L__BB156_127; bra.uni $L__BB156_126; $L__BB156_127: mov.u32 %r380, 256; add.u64 %rd211, %SP, 828; { // callseq 918, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd128; .param .b64 param1; st.param.b64 [param1+0], %rd211; .param .b32 param2; st.param.b32 [param2+0], %r380; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r381, [retval0+0]; } // callseq 918 cvt.u16.u32 %rs60, %r381; bra.uni $L__BB156_128; $L__BB156_87: mov.u32 %r338, 1985229328; st.local.u32 [%rd2], %r338; ld.global.u32 %r439, [%rd18]; ld.global.u8 %rs39, [%rd19]; and.b16 %rs40, %rs39, 4; setp.ne.s16 %p111, %rs40, 0; mov.u32 %r432, -1; @%p111 bra $L__BB156_89; ld.global.v4.f32 {%f995, %f996, %f997, %f998}, [%rd21]; ld.global.v4.f32 {%f1003, %f1004, %f1005, %f1006}, [%rd21+16]; ld.global.v4.f32 {%f1011, %f1012, %f1013, %f1014}, [%rd21+32]; mul.ftz.f32 %f1019, %f996, %f1407; fma.rn.ftz.f32 %f1020, %f1406, %f995, %f1019; fma.rn.ftz.f32 %f1021, %f997, %f1408, %f1020; add.ftz.f32 %f208, %f998, %f1021; mul.ftz.f32 %f1022, %f1004, %f1407; fma.rn.ftz.f32 %f1023, %f1406, %f1003, %f1022; fma.rn.ftz.f32 %f1024, %f1005, %f1408, %f1023; add.ftz.f32 %f209, %f1006, %f1024; mul.ftz.f32 %f1025, %f1012, %f1407; fma.rn.ftz.f32 %f1026, %f1406, %f1011, %f1025; fma.rn.ftz.f32 %f1027, %f1013, %f1408, %f1026; add.ftz.f32 %f1408, %f1014, %f1027; mul.ftz.f32 %f1028, %f996, %f52; fma.rn.ftz.f32 %f1029, %f995, %f51, %f1028; fma.rn.ftz.f32 %f1030, %f997, %f53, %f1029; mul.ftz.f32 %f1031, %f1004, %f52; fma.rn.ftz.f32 %f1032, %f1003, %f51, %f1031; fma.rn.ftz.f32 %f1033, %f1005, %f53, %f1032; mul.ftz.f32 %f1034, %f1012, %f52; fma.rn.ftz.f32 %f1035, %f1011, %f51, %f1034; fma.rn.ftz.f32 %f1036, %f1013, %f53, %f1035; abs.ftz.f32 %f1037, %f1030; setp.gt.ftz.f32 %p112, %f1037, 0f17800000; mov.b32 %r339, %f1030; and.b32 %r340, %r339, -2147483648; or.b32 %r341, %r340, 394264576; mov.b32 %f1038, %r341; selp.f32 %f1403, %f1030, %f1038, %p112; abs.ftz.f32 %f1039, %f1033; setp.gt.ftz.f32 %p113, %f1039, 0f17800000; mov.b32 %r342, %f1033; and.b32 %r343, %r342, -2147483648; or.b32 %r344, %r343, 394264576; mov.b32 %f1040, %r344; selp.f32 %f1404, %f1033, %f1040, %p113; abs.ftz.f32 %f1041, %f1036; setp.gt.ftz.f32 %p114, %f1041, 0f17800000; mov.b32 %r345, %f1036; and.b32 %r346, %r345, -2147483648; or.b32 %r347, %r346, 394264576; mov.b32 %f1042, %r347; selp.f32 %f1405, %f1036, %f1042, %p114; mov.f32 %f1406, %f208; mov.f32 %f1407, %f209; mov.u32 %r432, %r37; $L__BB156_89: rcp.approx.ftz.f32 %f220, %f1404; rcp.approx.ftz.f32 %f221, %f1403; rcp.approx.ftz.f32 %f222, %f1405; mov.u32 %r438, 0; $L__BB156_90: setp.eq.s32 %p115, %r439, 1985229328; setp.lt.s32 %p116, %r439, 0; or.pred %p117, %p116, %p115; @%p117 bra $L__BB156_101; $L__BB156_91: mul.wide.s32 %rd184, %r439, 16; add.s64 %rd85, %rd8, %rd184; ld.global.v4.f32 {%f1043, %f1044, %f1045, %f1046}, [%rd85]; mov.b32 %r439, %f1046; mov.b32 %r105, %f1045; mov.b32 %r104, %f1044; mov.b32 %r103, %f1043; mov.b64 %rd185, {%r103, %r104}; cvt.u32.u64 %r349, %rd185; and.b32 %r350, %r349, 1024; setp.eq.s32 %p118, %r350, 0; and.b32 %r107, %r103, 2047; @%p118 bra $L__BB156_94; bra.uni $L__BB156_92; $L__BB156_94: setp.ne.s32 %p124, %r107, 0; ld.global.v4.f32 {%f1187, %f1188, %f1189, %f1190}, [%rd85+16]; ld.global.v4.f32 {%f1195, %f1196, %f1197, %f1198}, [%rd85+32]; ld.global.v4.f32 {%f1203, %f1204, %f1205, %f1206}, [%rd85+48]; sub.ftz.f32 %f1211, %f1187, %f1406; mul.ftz.f32 %f1212, %f221, %f1211; sub.ftz.f32 %f1213, %f1189, %f1406; mul.ftz.f32 %f1214, %f221, %f1213; sub.ftz.f32 %f1215, %f1195, %f1407; mul.ftz.f32 %f1216, %f220, %f1215; sub.ftz.f32 %f1217, %f1197, %f1407; mul.ftz.f32 %f1218, %f220, %f1217; sub.ftz.f32 %f1219, %f1203, %f1408; mul.ftz.f32 %f1220, %f222, %f1219; sub.ftz.f32 %f1221, %f1205, %f1408; mul.ftz.f32 %f1222, %f222, %f1221; min.ftz.f32 %f1223, %f1212, %f1214; min.ftz.f32 %f1224, %f1216, %f1218; min.ftz.f32 %f1225, %f1220, %f1222; mov.f32 %f1226, 0f00000000; max.ftz.f32 %f1227, %f1226, %f1223; max.ftz.f32 %f1228, %f1224, %f1225; max.ftz.f32 %f1417, %f1227, %f1228; max.ftz.f32 %f1229, %f1212, %f1214; max.ftz.f32 %f1230, %f1216, %f1218; max.ftz.f32 %f1231, %f1220, %f1222; min.ftz.f32 %f1232, %f13, %f1229; min.ftz.f32 %f1233, %f1230, %f1231; min.ftz.f32 %f1234, %f1232, %f1233; sub.ftz.f32 %f1235, %f1188, %f1406; mul.ftz.f32 %f1236, %f221, %f1235; sub.ftz.f32 %f1237, %f1190, %f1406; mul.ftz.f32 %f1238, %f221, %f1237; sub.ftz.f32 %f1239, %f1196, %f1407; mul.ftz.f32 %f1240, %f220, %f1239; sub.ftz.f32 %f1241, %f1198, %f1407; mul.ftz.f32 %f1242, %f220, %f1241; sub.ftz.f32 %f1243, %f1204, %f1408; mul.ftz.f32 %f1244, %f222, %f1243; sub.ftz.f32 %f1245, %f1206, %f1408; mul.ftz.f32 %f1246, %f222, %f1245; min.ftz.f32 %f1247, %f1236, %f1238; min.ftz.f32 %f1248, %f1240, %f1242; min.ftz.f32 %f1249, %f1244, %f1246; max.ftz.f32 %f1250, %f1226, %f1247; max.ftz.f32 %f1251, %f1248, %f1249; max.ftz.f32 %f1418, %f1250, %f1251; max.ftz.f32 %f1252, %f1236, %f1238; max.ftz.f32 %f1253, %f1240, %f1242; max.ftz.f32 %f1254, %f1244, %f1246; min.ftz.f32 %f1255, %f13, %f1252; min.ftz.f32 %f1256, %f1253, %f1254; min.ftz.f32 %f1257, %f1255, %f1256; setp.ge.ftz.f32 %p125, %f1234, %f1417; and.pred %p126, %p124, %p125; selp.u32 %r354, 1, 0, %p126; setp.ltu.ftz.f32 %p127, %f1257, %f1418; and.b32 %r356, %r104, 2047; setp.eq.s32 %p128, %r356, 0; or.pred %p129, %p128, %p127; selp.b32 %r357, 0, 2, %p129; or.b32 %r437, %r357, %r354; bra.uni $L__BB156_95; $L__BB156_92: setp.ne.s32 %p119, %r107, 0; ld.global.v4.f32 {%f1050, %f1051, %f1052, %f1053}, [%rd85+16]; ld.global.v4.f32 {%f1058, %f1059, %f1060, %f1061}, [%rd85+32]; ld.global.v4.f32 {%f1066, %f1067, %f1068, %f1069}, [%rd85+48]; mul.ftz.f32 %f1074, %f1404, %f1051; fma.rn.ftz.f32 %f1075, %f1403, %f1050, %f1074; fma.rn.ftz.f32 %f1076, %f1405, %f1052, %f1075; mul.ftz.f32 %f1077, %f1404, %f1059; fma.rn.ftz.f32 %f1078, %f1403, %f1058, %f1077; fma.rn.ftz.f32 %f1079, %f1405, %f1060, %f1078; mul.ftz.f32 %f1080, %f1404, %f1067; fma.rn.ftz.f32 %f1081, %f1403, %f1066, %f1080; fma.rn.ftz.f32 %f1082, %f1405, %f1068, %f1081; mul.ftz.f32 %f1083, %f1407, %f1051; fma.rn.ftz.f32 %f1084, %f1406, %f1050, %f1083; fma.rn.ftz.f32 %f1085, %f1408, %f1052, %f1084; add.ftz.f32 %f1086, %f1053, %f1085; mul.ftz.f32 %f1087, %f1407, %f1059; fma.rn.ftz.f32 %f1088, %f1406, %f1058, %f1087; fma.rn.ftz.f32 %f1089, %f1408, %f1060, %f1088; add.ftz.f32 %f1090, %f1061, %f1089; mul.ftz.f32 %f1091, %f1407, %f1067; fma.rn.ftz.f32 %f1092, %f1406, %f1066, %f1091; fma.rn.ftz.f32 %f1093, %f1408, %f1068, %f1092; add.ftz.f32 %f1094, %f1069, %f1093; rcp.approx.ftz.f32 %f1095, %f1076; rcp.approx.ftz.f32 %f1096, %f1079; rcp.approx.ftz.f32 %f1097, %f1082; mul.ftz.f32 %f1098, %f1095, %f1086; neg.ftz.f32 %f1099, %f1098; mul.ftz.f32 %f1100, %f1096, %f1090; neg.ftz.f32 %f1101, %f1100; mul.ftz.f32 %f1102, %f1097, %f1094; neg.ftz.f32 %f1103, %f1102; sub.ftz.f32 %f1104, %f1095, %f1098; sub.ftz.f32 %f1105, %f1096, %f1100; sub.ftz.f32 %f1106, %f1097, %f1102; min.ftz.f32 %f1107, %f1099, %f1104; min.ftz.f32 %f1108, %f1101, %f1105; min.ftz.f32 %f1109, %f1103, %f1106; max.ftz.f32 %f1110, %f1099, %f1104; max.ftz.f32 %f1111, %f1101, %f1105; max.ftz.f32 %f1112, %f1103, %f1106; mov.f32 %f1113, 0f00000000; max.ftz.f32 %f1114, %f1113, %f1107; max.ftz.f32 %f1115, %f1108, %f1109; max.ftz.f32 %f1417, %f1114, %f1115; min.ftz.f32 %f1116, %f13, %f1110; min.ftz.f32 %f1117, %f1111, %f1112; min.ftz.f32 %f1118, %f1116, %f1117; setp.le.ftz.f32 %p120, %f1417, %f1118; and.pred %p121, %p119, %p120; selp.u32 %r437, 1, 0, %p121; ld.global.v4.f32 {%f1119, %f1120, %f1121, %f1122}, [%rd85+64]; ld.global.v4.f32 {%f1127, %f1128, %f1129, %f1130}, [%rd85+80]; ld.global.v4.f32 {%f1135, %f1136, %f1137, %f1138}, [%rd85+96]; mul.ftz.f32 %f1143, %f1404, %f1120; fma.rn.ftz.f32 %f1144, %f1403, %f1119, %f1143; fma.rn.ftz.f32 %f1145, %f1405, %f1121, %f1144; mul.ftz.f32 %f1146, %f1404, %f1128; fma.rn.ftz.f32 %f1147, %f1403, %f1127, %f1146; fma.rn.ftz.f32 %f1148, %f1405, %f1129, %f1147; mul.ftz.f32 %f1149, %f1404, %f1136; fma.rn.ftz.f32 %f1150, %f1403, %f1135, %f1149; fma.rn.ftz.f32 %f1151, %f1405, %f1137, %f1150; mul.ftz.f32 %f1152, %f1407, %f1120; fma.rn.ftz.f32 %f1153, %f1406, %f1119, %f1152; fma.rn.ftz.f32 %f1154, %f1408, %f1121, %f1153; add.ftz.f32 %f1155, %f1122, %f1154; mul.ftz.f32 %f1156, %f1407, %f1128; fma.rn.ftz.f32 %f1157, %f1406, %f1127, %f1156; fma.rn.ftz.f32 %f1158, %f1408, %f1129, %f1157; add.ftz.f32 %f1159, %f1130, %f1158; mul.ftz.f32 %f1160, %f1407, %f1136; fma.rn.ftz.f32 %f1161, %f1406, %f1135, %f1160; fma.rn.ftz.f32 %f1162, %f1408, %f1137, %f1161; add.ftz.f32 %f1163, %f1138, %f1162; rcp.approx.ftz.f32 %f1164, %f1145; rcp.approx.ftz.f32 %f1165, %f1148; rcp.approx.ftz.f32 %f1166, %f1151; mul.ftz.f32 %f1167, %f1164, %f1155; neg.ftz.f32 %f1168, %f1167; mul.ftz.f32 %f1169, %f1165, %f1159; neg.ftz.f32 %f1170, %f1169; mul.ftz.f32 %f1171, %f1166, %f1163; neg.ftz.f32 %f1172, %f1171; sub.ftz.f32 %f1173, %f1164, %f1167; sub.ftz.f32 %f1174, %f1165, %f1169; sub.ftz.f32 %f1175, %f1166, %f1171; min.ftz.f32 %f1176, %f1168, %f1173; min.ftz.f32 %f1177, %f1170, %f1174; min.ftz.f32 %f1178, %f1172, %f1175; max.ftz.f32 %f1179, %f1168, %f1173; max.ftz.f32 %f1180, %f1170, %f1174; max.ftz.f32 %f1181, %f1172, %f1175; max.ftz.f32 %f1182, %f1113, %f1176; max.ftz.f32 %f1183, %f1177, %f1178; max.ftz.f32 %f1418, %f1182, %f1183; min.ftz.f32 %f1184, %f13, %f1179; min.ftz.f32 %f1185, %f1180, %f1181; min.ftz.f32 %f1186, %f1184, %f1185; setp.gtu.ftz.f32 %p122, %f1418, %f1186; @%p122 bra $L__BB156_95; and.b32 %r352, %r104, 2047; setp.eq.s32 %p123, %r352, 0; or.b32 %r353, %r437, 2; selp.b32 %r437, %r437, %r353, %p123; $L__BB156_95: mul.wide.s32 %rd186, %r438, 4; add.s64 %rd86, %rd2, %rd186; setp.eq.s32 %p130, %r437, 3; @%p130 bra $L__BB156_99; bra.uni $L__BB156_96; $L__BB156_99: setp.lt.ftz.f32 %p133, %f1418, %f1417; selp.b32 %r116, %r439, %r105, %p133; selp.b32 %r358, %r105, %r439, %p133; add.s32 %r438, %r438, 1; st.local.u32 [%rd86+4], %r358; mov.u32 %r439, %r116; bra.uni $L__BB156_100; $L__BB156_96: setp.eq.s32 %p131, %r437, 2; @%p131 bra $L__BB156_100; setp.ne.s32 %p132, %r437, 0; mov.u32 %r439, %r105; @%p132 bra $L__BB156_100; ld.local.u32 %r439, [%rd86]; add.s32 %r438, %r438, -1; $L__BB156_100: setp.ne.s32 %p134, %r439, 1985229328; setp.gt.s32 %p135, %r439, -1; and.pred %p136, %p135, %p134; @%p136 bra $L__BB156_91; $L__BB156_101: mov.u32 %r120, %r438; setp.gt.s32 %p137, %r439, -1; @%p137 bra $L__BB156_110; mul.wide.s32 %rd187, %r439, 16; xor.b64 %rd188, %rd187, -16; add.s64 %rd189, %rd9, %rd188; ld.global.v4.u32 {%r442, %r360, %r361, %r362}, [%rd189]; mul.wide.s32 %rd190, %r120, 4; add.s64 %rd191, %rd2, %rd190; ld.local.u32 %r439, [%rd191]; add.s32 %r438, %r120, -1; and.b32 %r364, %r362, 127; setp.ne.s32 %p138, %r364, 1; setp.ge.s32 %p139, %r442, %r360; or.pred %p140, %p138, %p139; @%p140 bra $L__BB156_110; $L__BB156_103: cvt.s64.s32 %rd87, %r442; setp.ne.s32 %p141, %r432, -1; @%p141 bra $L__BB156_105; shl.b64 %rd192, %rd87, 2; add.s64 %rd193, %rd10, %rd192; ld.global.u32 %r365, [%rd193]; setp.ne.s32 %p142, %r365, %r37; @%p142 bra $L__BB156_109; $L__BB156_105: shl.b64 %rd194, %rd87, 2; add.s64 %rd195, %rd11, %rd194; ld.global.u32 %r127, [%rd195]; setp.eq.s32 %p143, %r38, %r127; @%p143 bra $L__BB156_109; mul.wide.s32 %rd196, %r127, 16; add.s64 %rd197, %rd12, %rd196; ld.global.u32 %r366, [%rd197+12]; mul.wide.u32 %rd198, %r366, 12; add.s64 %rd199, %rd13, %rd198; add.s32 %r367, %r366, 1; mul.wide.u32 %rd200, %r367, 12; add.s64 %rd201, %rd13, %rd200; add.s32 %r368, %r366, 2; mul.wide.u32 %rd202, %r368, 12; add.s64 %rd203, %rd13, %rd202; ld.global.f32 %f1258, [%rd199]; sub.ftz.f32 %f230, %f1258, %f1406; ld.global.f32 %f1259, [%rd199+4]; sub.ftz.f32 %f231, %f1259, %f1407; ld.global.f32 %f1260, [%rd199+8]; sub.ftz.f32 %f232, %f1260, %f1408; ld.global.f32 %f1261, [%rd201]; sub.ftz.f32 %f1262, %f1261, %f1406; ld.global.f32 %f1263, [%rd201+4]; sub.ftz.f32 %f1264, %f1263, %f1407; ld.global.f32 %f1265, [%rd201+8]; sub.ftz.f32 %f1266, %f1265, %f1408; ld.global.f32 %f1267, [%rd203]; sub.ftz.f32 %f1268, %f1267, %f1406; ld.global.f32 %f1269, [%rd203+4]; sub.ftz.f32 %f1270, %f1269, %f1407; ld.global.f32 %f1271, [%rd203+8]; sub.ftz.f32 %f1272, %f1271, %f1408; sub.ftz.f32 %f233, %f1268, %f230; sub.ftz.f32 %f234, %f1270, %f231; sub.ftz.f32 %f235, %f1272, %f232; sub.ftz.f32 %f236, %f230, %f1262; sub.ftz.f32 %f237, %f231, %f1264; sub.ftz.f32 %f238, %f232, %f1266; sub.ftz.f32 %f1273, %f1262, %f1268; sub.ftz.f32 %f1274, %f1264, %f1270; sub.ftz.f32 %f1275, %f1266, %f1272; add.ftz.f32 %f1276, %f230, %f1268; add.ftz.f32 %f1277, %f231, %f1270; add.ftz.f32 %f1278, %f232, %f1272; mul.ftz.f32 %f1279, %f234, %f1278; mul.ftz.f32 %f1280, %f1277, %f235; sub.ftz.f32 %f1281, %f1279, %f1280; mul.ftz.f32 %f1282, %f1276, %f235; mul.ftz.f32 %f1283, %f233, %f1278; sub.ftz.f32 %f1284, %f1282, %f1283; mul.ftz.f32 %f1285, %f233, %f1277; mul.ftz.f32 %f1286, %f1276, %f234; sub.ftz.f32 %f1287, %f1285, %f1286; mul.ftz.f32 %f1288, %f1404, %f1284; fma.rn.ftz.f32 %f1289, %f1403, %f1281, %f1288; fma.rn.ftz.f32 %f1290, %f1405, %f1287, %f1289; add.ftz.f32 %f1291, %f230, %f1262; add.ftz.f32 %f1292, %f231, %f1264; add.ftz.f32 %f1293, %f232, %f1266; mul.ftz.f32 %f1294, %f237, %f1293; mul.ftz.f32 %f1295, %f1292, %f238; sub.ftz.f32 %f1296, %f1294, %f1295; mul.ftz.f32 %f1297, %f1291, %f238; mul.ftz.f32 %f1298, %f236, %f1293; sub.ftz.f32 %f1299, %f1297, %f1298; mul.ftz.f32 %f1300, %f236, %f1292; mul.ftz.f32 %f1301, %f1291, %f237; sub.ftz.f32 %f1302, %f1300, %f1301; mul.ftz.f32 %f1303, %f1404, %f1299; fma.rn.ftz.f32 %f1304, %f1403, %f1296, %f1303; fma.rn.ftz.f32 %f1305, %f1405, %f1302, %f1304; add.ftz.f32 %f1306, %f1262, %f1268; add.ftz.f32 %f1307, %f1264, %f1270; add.ftz.f32 %f1308, %f1266, %f1272; mul.ftz.f32 %f1309, %f1274, %f1308; mul.ftz.f32 %f1310, %f1307, %f1275; sub.ftz.f32 %f1311, %f1309, %f1310; mul.ftz.f32 %f1312, %f1306, %f1275; mul.ftz.f32 %f1313, %f1273, %f1308; sub.ftz.f32 %f1314, %f1312, %f1313; mul.ftz.f32 %f1315, %f1273, %f1307; mul.ftz.f32 %f1316, %f1306, %f1274; sub.ftz.f32 %f1317, %f1315, %f1316; mul.ftz.f32 %f1318, %f1404, %f1314; fma.rn.ftz.f32 %f1319, %f1403, %f1311, %f1318; fma.rn.ftz.f32 %f1320, %f1405, %f1317, %f1319; add.ftz.f32 %f1321, %f1305, %f1290; add.ftz.f32 %f1322, %f1320, %f1321; abs.ftz.f32 %f1323, %f1322; mul.ftz.f32 %f1324, %f1323, 0f34000000; min.ftz.f32 %f1325, %f1305, %f1320; min.ftz.f32 %f1326, %f1290, %f1325; max.ftz.f32 %f1327, %f1305, %f1320; max.ftz.f32 %f1328, %f1290, %f1327; mul.ftz.f32 %f1329, %f1323, 0fB4000000; setp.ltu.ftz.f32 %p144, %f1326, %f1329; setp.gtu.ftz.f32 %p145, %f1328, %f1324; and.pred %p146, %p144, %p145; @%p146 bra $L__BB156_109; mul.ftz.f32 %f1330, %f237, %f235; mul.ftz.f32 %f1331, %f238, %f234; sub.ftz.f32 %f1332, %f1330, %f1331; mul.ftz.f32 %f1333, %f236, %f235; mul.ftz.f32 %f1334, %f238, %f233; sub.ftz.f32 %f1335, %f1334, %f1333; mul.ftz.f32 %f1336, %f237, %f233; mul.ftz.f32 %f1337, %f236, %f234; sub.ftz.f32 %f1338, %f1337, %f1336; add.ftz.f32 %f239, %f1332, %f1332; add.ftz.f32 %f240, %f1335, %f1335; add.ftz.f32 %f241, %f1338, %f1338; mul.ftz.f32 %f1339, %f1404, %f240; fma.rn.ftz.f32 %f1340, %f1403, %f239, %f1339; fma.rn.ftz.f32 %f242, %f1405, %f241, %f1340; setp.eq.ftz.f32 %p147, %f242, 0f00000000; @%p147 bra $L__BB156_109; mul.ftz.f32 %f1341, %f231, %f240; fma.rn.ftz.f32 %f1342, %f230, %f239, %f1341; fma.rn.ftz.f32 %f1343, %f232, %f241, %f1342; div.approx.ftz.f32 %f1344, %f1343, %f242; setp.ge.ftz.f32 %p148, %f1344, 0f00000000; setp.le.ftz.f32 %p149, %f1344, %f13; mov.u16 %rs60, 1; and.pred %p150, %p149, %p148; @%p150 bra $L__BB156_128; $L__BB156_109: add.s32 %r438, %r120, -1; cvt.u32.u64 %r369, %rd87; add.s32 %r442, %r369, 1; setp.lt.s32 %p151, %r442, %r360; @%p151 bra $L__BB156_103; $L__BB156_110: setp.ne.s32 %p152, %r439, 1985229328; mov.u16 %rs60, 0; @%p152 bra $L__BB156_90; bra.uni $L__BB156_128; $L__BB156_123: mov.u32 %r374, 256; add.u64 %rd205, %SP, 828; { // callseq 915, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd128; .param .b64 param1; st.param.b64 [param1+0], %rd205; .param .b32 param2; st.param.b32 [param2+0], %r374; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r375, [retval0+0]; } // callseq 915 cvt.u16.u32 %rs60, %r375; bra.uni $L__BB156_128; $L__BB156_35: ld.global.u32 %rd138, [%rd21+120]; ld.global.u32 %r276, [%rd21+104]; shl.b32 %r277, %r276, 1; cvt.rn.f32.s32 %f314, %r277; mul.ftz.f32 %f315, %f54, %f314; cvt.rzi.ftz.s32.f32 %r278, %f315; add.s32 %r279, %r277, -1; min.s32 %r280, %r278, %r279; cvt.rn.f32.s32 %f316, %r280; sub.ftz.f32 %f58, %f315, %f316; cvt.s64.s32 %rd139, %r280; add.s64 %rd140, %rd139, %rd138; shl.b64 %rd141, %rd140, 6; add.s64 %rd22, %rd16, %rd141; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd22]; ld.global.v4.f32 {%f321, %f322, %f323, %f324}, [%rd22+64]; mul.ftz.f32 %f325, %f318, %f322; fma.rn.ftz.f32 %f326, %f317, %f321, %f325; mul.ftz.f32 %f327, %f320, %f324; fma.rn.ftz.f32 %f328, %f319, %f323, %f327; add.ftz.f32 %f67, %f326, %f328; setp.gt.ftz.f32 %p44, %f67, 0f3F7FDF3B; @%p44 bra $L__BB156_37; bra.uni $L__BB156_36; $L__BB156_37: mov.f32 %f1365, 0f3F800000; sub.ftz.f32 %f1364, %f1365, %f58; mul.ftz.f32 %f383, %f317, %f1364; fma.rn.ftz.f32 %f384, %f321, %f58, %f383; mul.ftz.f32 %f385, %f318, %f1364; fma.rn.ftz.f32 %f386, %f322, %f58, %f385; mul.ftz.f32 %f387, %f319, %f1364; fma.rn.ftz.f32 %f388, %f323, %f58, %f387; mul.ftz.f32 %f389, %f320, %f1364; fma.rn.ftz.f32 %f390, %f324, %f58, %f389; mul.ftz.f32 %f391, %f386, %f386; fma.rn.ftz.f32 %f392, %f384, %f384, %f391; mul.ftz.f32 %f393, %f390, %f390; fma.rn.ftz.f32 %f394, %f388, %f388, %f393; add.ftz.f32 %f395, %f392, %f394; rsqrt.approx.ftz.f32 %f396, %f395; mul.ftz.f32 %f1383, %f390, %f396; mul.ftz.f32 %f1382, %f388, %f396; mul.ftz.f32 %f1381, %f386, %f396; mul.ftz.f32 %f1380, %f384, %f396; bra.uni $L__BB156_38; $L__BB156_126: mov.u32 %r378, 256; add.u64 %rd209, %SP, 828; { // callseq 917, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd128; .param .b64 param1; st.param.b64 [param1+0], %rd209; .param .b32 param2; st.param.b32 [param2+0], %r378; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r379, [retval0+0]; } // callseq 917 cvt.u16.u32 %rs60, %r379; bra.uni $L__BB156_128; $L__BB156_36: mov.f32 %f1360, 0f3F800000; mov.f32 %f330, 0fBF800000; max.ftz.f32 %f331, %f67, %f330; min.ftz.f32 %f333, %f331, %f1360; abs.ftz.f32 %f334, %f333; sub.ftz.f32 %f335, %f1360, %f334; mul.ftz.f32 %f336, %f335, 0f3F000000; sqrt.approx.ftz.f32 %f337, %f336; setp.gt.ftz.f32 %p45, %f334, 0f3F11EB85; selp.f32 %f338, %f337, %f334, %p45; mul.ftz.f32 %f339, %f338, %f338; mov.f32 %f340, 0f3C94D2E9; mov.f32 %f341, 0f3D53F941; fma.rn.ftz.f32 %f342, %f341, %f339, %f340; mov.f32 %f343, 0f3D3F841F; fma.rn.ftz.f32 %f344, %f342, %f339, %f343; mov.f32 %f345, 0f3D994929; fma.rn.ftz.f32 %f346, %f344, %f339, %f345; mov.f32 %f347, 0f3E2AAB94; fma.rn.ftz.f32 %f348, %f346, %f339, %f347; mul.ftz.f32 %f349, %f339, %f348; fma.rn.ftz.f32 %f350, %f349, %f338, %f338; add.ftz.f32 %f351, %f350, %f350; mov.f32 %f352, 0f3FC90FDB; sub.ftz.f32 %f353, %f352, %f350; selp.f32 %f354, %f351, %f353, %p45; setp.lt.ftz.f32 %p46, %f333, 0f00000000; mov.f32 %f355, 0f40490FDB; sub.ftz.f32 %f356, %f355, %f354; selp.f32 %f357, %f356, %f354, %p46; mul.ftz.f32 %f358, %f317, %f67; sub.ftz.f32 %f359, %f321, %f358; mul.ftz.f32 %f360, %f318, %f67; sub.ftz.f32 %f361, %f322, %f360; mul.ftz.f32 %f362, %f319, %f67; sub.ftz.f32 %f363, %f323, %f362; mul.ftz.f32 %f364, %f320, %f67; sub.ftz.f32 %f365, %f324, %f364; mul.ftz.f32 %f366, %f361, %f361; fma.rn.ftz.f32 %f367, %f359, %f359, %f366; mul.ftz.f32 %f368, %f365, %f365; fma.rn.ftz.f32 %f369, %f363, %f363, %f368; add.ftz.f32 %f370, %f367, %f369; rsqrt.approx.ftz.f32 %f371, %f370; mul.ftz.f32 %f372, %f359, %f371; mul.ftz.f32 %f373, %f361, %f371; mul.ftz.f32 %f374, %f363, %f371; mul.ftz.f32 %f375, %f365, %f371; mul.ftz.f32 %f376, %f58, %f357; cos.approx.ftz.f32 %f377, %f376; sin.approx.ftz.f32 %f378, %f376; mul.ftz.f32 %f379, %f372, %f378; mul.ftz.f32 %f380, %f373, %f378; mul.ftz.f32 %f381, %f374, %f378; mul.ftz.f32 %f382, %f375, %f378; fma.rn.ftz.f32 %f1383, %f320, %f377, %f382; fma.rn.ftz.f32 %f1382, %f319, %f377, %f381; fma.rn.ftz.f32 %f1381, %f318, %f377, %f380; fma.rn.ftz.f32 %f1380, %f317, %f377, %f379; $L__BB156_38: mov.f32 %f1363, 0f3F800000; sub.ftz.f32 %f1362, %f1363, %f58; mov.f32 %f1361, 0f3F800000; ld.global.v4.f32 {%f397, %f398, %f399, %f400}, [%rd22+16]; ld.global.v4.f32 {%f405, %f406, %f407, %f408}, [%rd22+80]; mul.ftz.f32 %f413, %f58, %f405; mul.ftz.f32 %f414, %f58, %f406; mul.ftz.f32 %f415, %f58, %f407; mul.ftz.f32 %f416, %f58, %f408; fma.rn.ftz.f32 %f81, %f1362, %f397, %f413; fma.rn.ftz.f32 %f82, %f1362, %f398, %f414; fma.rn.ftz.f32 %f83, %f1362, %f399, %f415; fma.rn.ftz.f32 %f417, %f1362, %f400, %f416; ld.global.v4.f32 {%f418, %f419, %f420, %f421}, [%rd22+32]; ld.global.v4.f32 {%f426, %f427, %f428, %f429}, [%rd22+96]; mul.ftz.f32 %f434, %f58, %f426; mul.ftz.f32 %f435, %f58, %f427; mul.ftz.f32 %f436, %f58, %f428; mul.ftz.f32 %f437, %f58, %f429; fma.rn.ftz.f32 %f438, %f1362, %f418, %f434; fma.rn.ftz.f32 %f439, %f1362, %f419, %f435; fma.rn.ftz.f32 %f440, %f1362, %f420, %f436; fma.rn.ftz.f32 %f441, %f1362, %f421, %f437; ld.global.v4.f32 {%f442, %f443, %f444, %f445}, [%rd22+48]; ld.global.v4.f32 {%f450, %f451, %f452, %f453}, [%rd22+112]; mul.ftz.f32 %f458, %f58, %f450; mul.ftz.f32 %f459, %f58, %f451; mul.ftz.f32 %f460, %f58, %f452; mul.ftz.f32 %f461, %f58, %f453; fma.rn.ftz.f32 %f462, %f1362, %f442, %f458; fma.rn.ftz.f32 %f463, %f1362, %f443, %f459; fma.rn.ftz.f32 %f464, %f1362, %f444, %f460; fma.rn.ftz.f32 %f465, %f1362, %f445, %f461; mul.ftz.f32 %f466, %f1383, 0f3FB504F3; mul.ftz.f32 %f467, %f1380, 0f3FB504F3; mul.ftz.f32 %f468, %f1381, 0f3FB504F3; mul.ftz.f32 %f469, %f1382, 0f3FB504F3; mul.ftz.f32 %f470, %f466, %f467; mul.ftz.f32 %f471, %f466, %f468; mul.ftz.f32 %f472, %f466, %f469; mul.ftz.f32 %f473, %f467, %f467; mul.ftz.f32 %f474, %f467, %f468; mul.ftz.f32 %f475, %f467, %f469; mul.ftz.f32 %f476, %f468, %f468; mul.ftz.f32 %f477, %f468, %f469; mul.ftz.f32 %f478, %f469, %f469; sub.ftz.f32 %f480, %f1361, %f476; sub.ftz.f32 %f481, %f480, %f478; sub.ftz.f32 %f482, %f474, %f472; add.ftz.f32 %f483, %f471, %f475; add.ftz.f32 %f484, %f472, %f474; sub.ftz.f32 %f485, %f1361, %f473; sub.ftz.f32 %f486, %f485, %f478; sub.ftz.f32 %f487, %f477, %f470; sub.ftz.f32 %f488, %f475, %f471; add.ftz.f32 %f489, %f470, %f477; sub.ftz.f32 %f490, %f485, %f476; mul.ftz.f32 %f491, %f482, %f440; fma.rn.ftz.f32 %f492, %f481, %f417, %f491; fma.rn.ftz.f32 %f1389, %f483, %f463, %f492; mul.ftz.f32 %f493, %f482, %f441; fma.rn.ftz.f32 %f494, %f481, %f438, %f493; fma.rn.ftz.f32 %f85, %f483, %f464, %f494; mul.ftz.f32 %f495, %f482, %f462; fma.rn.ftz.f32 %f496, %f481, %f439, %f495; fma.rn.ftz.f32 %f86, %f483, %f465, %f496; mul.ftz.f32 %f497, %f486, %f440; fma.rn.ftz.f32 %f498, %f484, %f417, %f497; fma.rn.ftz.f32 %f87, %f487, %f463, %f498; mul.ftz.f32 %f499, %f486, %f441; fma.rn.ftz.f32 %f500, %f484, %f438, %f499; fma.rn.ftz.f32 %f1388, %f487, %f464, %f500; mul.ftz.f32 %f501, %f486, %f462; fma.rn.ftz.f32 %f502, %f484, %f439, %f501; fma.rn.ftz.f32 %f89, %f487, %f465, %f502; mul.ftz.f32 %f503, %f489, %f440; fma.rn.ftz.f32 %f504, %f488, %f417, %f503; fma.rn.ftz.f32 %f90, %f490, %f463, %f504; mul.ftz.f32 %f505, %f489, %f441; fma.rn.ftz.f32 %f506, %f488, %f438, %f505; fma.rn.ftz.f32 %f91, %f490, %f464, %f506; mul.ftz.f32 %f507, %f489, %f462; fma.rn.ftz.f32 %f508, %f488, %f439, %f507; fma.rn.ftz.f32 %f1387, %f490, %f465, %f508; mul.ftz.f32 %f509, %f1388, %f1387; mul.ftz.f32 %f93, %f91, %f89; sub.ftz.f32 %f1386, %f509, %f93; mul.ftz.f32 %f95, %f91, %f86; mul.ftz.f32 %f510, %f85, %f1387; sub.ftz.f32 %f1385, %f95, %f510; mul.ftz.f32 %f97, %f85, %f89; mul.ftz.f32 %f511, %f1388, %f86; sub.ftz.f32 %f1384, %f97, %f511; mul.ftz.f32 %f512, %f87, %f1385; fma.rn.ftz.f32 %f513, %f1389, %f1386, %f512; fma.rn.ftz.f32 %f1390, %f90, %f1384, %f513; setp.neu.ftz.f32 %p47, %f1390, 0f00000000; @%p47 bra $L__BB156_40; add.ftz.f32 %f1389, %f1389, 0f322BCC77; add.ftz.f32 %f1387, %f1387, 0f322BCC77; add.ftz.f32 %f1388, %f1388, 0f322BCC77; mul.ftz.f32 %f514, %f1388, %f1387; sub.ftz.f32 %f1386, %f514, %f93; mul.ftz.f32 %f515, %f85, %f1387; sub.ftz.f32 %f1385, %f95, %f515; mul.ftz.f32 %f516, %f1388, %f86; sub.ftz.f32 %f1384, %f97, %f516; mul.ftz.f32 %f517, %f87, %f1385; fma.rn.ftz.f32 %f518, %f1389, %f1386, %f517; fma.rn.ftz.f32 %f519, %f90, %f1384, %f518; setp.eq.ftz.f32 %p48, %f519, 0f00000000; selp.f32 %f1390, 0f7F7FFFFF, %f519, %p48; $L__BB156_40: mul.ftz.f32 %f520, %f87, %f1387; mul.ftz.f32 %f521, %f90, %f89; sub.ftz.f32 %f522, %f521, %f520; mul.ftz.f32 %f523, %f90, %f86; mul.ftz.f32 %f524, %f1387, %f1389; sub.ftz.f32 %f525, %f524, %f523; mul.ftz.f32 %f526, %f89, %f1389; mul.ftz.f32 %f527, %f87, %f86; sub.ftz.f32 %f528, %f527, %f526; div.approx.ftz.f32 %f1395, %f522, %f1390; div.approx.ftz.f32 %f1396, %f525, %f1390; div.approx.ftz.f32 %f1397, %f528, %f1390; mul.ftz.f32 %f529, %f90, %f1388; mul.ftz.f32 %f530, %f87, %f91; sub.ftz.f32 %f531, %f530, %f529; mul.ftz.f32 %f532, %f91, %f1389; mul.ftz.f32 %f533, %f90, %f85; sub.ftz.f32 %f534, %f533, %f532; mul.ftz.f32 %f535, %f87, %f85; mul.ftz.f32 %f536, %f1388, %f1389; sub.ftz.f32 %f537, %f536, %f535; div.approx.ftz.f32 %f1399, %f531, %f1390; div.approx.ftz.f32 %f1400, %f534, %f1390; div.approx.ftz.f32 %f1401, %f537, %f1390; div.approx.ftz.f32 %f1391, %f1386, %f1390; div.approx.ftz.f32 %f1392, %f1385, %f1390; mul.ftz.f32 %f538, %f82, %f1392; fma.rn.ftz.f32 %f539, %f81, %f1391, %f538; div.approx.ftz.f32 %f1393, %f1384, %f1390; fma.rn.ftz.f32 %f540, %f83, %f1393, %f539; neg.ftz.f32 %f1394, %f540; mul.ftz.f32 %f541, %f82, %f1396; fma.rn.ftz.f32 %f542, %f81, %f1395, %f541; fma.rn.ftz.f32 %f543, %f83, %f1397, %f542; neg.ftz.f32 %f1398, %f543; mul.ftz.f32 %f544, %f82, %f1400; fma.rn.ftz.f32 %f545, %f81, %f1399, %f544; fma.rn.ftz.f32 %f546, %f83, %f1401, %f545; neg.ftz.f32 %f1402, %f546; $L__BB156_42: mul.ftz.f32 %f559, %f1392, %f1407; fma.rn.ftz.f32 %f560, %f1391, %f1406, %f559; fma.rn.ftz.f32 %f561, %f1393, %f1408, %f560; add.ftz.f32 %f150, %f1394, %f561; mul.ftz.f32 %f562, %f1396, %f1407; fma.rn.ftz.f32 %f563, %f1395, %f1406, %f562; fma.rn.ftz.f32 %f564, %f1397, %f1408, %f563; add.ftz.f32 %f151, %f1398, %f564; mul.ftz.f32 %f565, %f1400, %f1407; fma.rn.ftz.f32 %f566, %f1399, %f1406, %f565; fma.rn.ftz.f32 %f567, %f1401, %f1408, %f566; add.ftz.f32 %f1408, %f1402, %f567; mul.ftz.f32 %f568, %f1392, %f52; fma.rn.ftz.f32 %f569, %f1391, %f51, %f568; fma.rn.ftz.f32 %f570, %f1393, %f53, %f569; mul.ftz.f32 %f571, %f1396, %f52; fma.rn.ftz.f32 %f572, %f1395, %f51, %f571; fma.rn.ftz.f32 %f573, %f1397, %f53, %f572; mul.ftz.f32 %f574, %f1400, %f52; fma.rn.ftz.f32 %f575, %f1399, %f51, %f574; fma.rn.ftz.f32 %f576, %f1401, %f53, %f575; abs.ftz.f32 %f577, %f570; setp.gt.ftz.f32 %p49, %f577, 0f17800000; mov.b32 %r281, %f570; and.b32 %r282, %r281, -2147483648; or.b32 %r283, %r282, 394264576; mov.b32 %f578, %r283; selp.f32 %f1403, %f570, %f578, %p49; abs.ftz.f32 %f579, %f573; setp.gt.ftz.f32 %p50, %f579, 0f17800000; mov.b32 %r284, %f573; and.b32 %r285, %r284, -2147483648; or.b32 %r286, %r285, 394264576; mov.b32 %f580, %r286; selp.f32 %f1404, %f573, %f580, %p50; abs.ftz.f32 %f581, %f576; setp.gt.ftz.f32 %p51, %f581, 0f17800000; mov.b32 %r287, %f576; and.b32 %r288, %r287, -2147483648; or.b32 %r289, %r288, 394264576; mov.b32 %f582, %r289; selp.f32 %f1405, %f576, %f582, %p51; mov.f32 %f1406, %f150; mov.f32 %f1407, %f151; mov.u32 %r413, %r37; $L__BB156_43: mov.u32 %r430, 0; rcp.approx.ftz.f32 %f162, %f1404; rcp.approx.ftz.f32 %f163, %f1403; rcp.approx.ftz.f32 %f164, %f1405; bra.uni $L__BB156_44; $L__BB156_56: mul.wide.s32 %rd145, %r420, 16; xor.b64 %rd146, %rd145, -16; add.s64 %rd147, %rd9, %rd146; ld.global.v4.u32 {%r423, %r302, %r303, %r304}, [%rd147]; mul.wide.s32 %rd148, %r65, 4; add.s64 %rd149, %rd2, %rd148; ld.local.u32 %r420, [%rd149]; cvt.u16.u32 %rs35, %r304; and.b16 %rs34, %rs35, 127; add.s32 %r430, %r65, -1; setp.eq.s16 %p75, %rs34, 1; @%p75 bra $L__BB156_78; setp.ne.s16 %p76, %rs34, 65; @%p76 bra $L__BB156_86; add.s32 %r430, %r65, -1; setp.ge.s32 %p77, %r423, %r302; @%p77 bra $L__BB156_86; $L__BB156_59: cvt.s64.s32 %rd26, %r423; setp.ne.s32 %p78, %r413, -1; @%p78 bra $L__BB156_61; shl.b64 %rd150, %rd26, 2; add.s64 %rd151, %rd10, %rd150; ld.global.u32 %r306, [%rd151]; setp.ne.s32 %p79, %r306, %r37; @%p79 bra $L__BB156_77; $L__BB156_61: shl.b64 %rd152, %rd26, 2; add.s64 %rd153, %rd11, %rd152; ld.global.u32 %r72, [%rd153]; setp.eq.s32 %p80, %r38, %r72; @%p80 bra $L__BB156_77; ld.const.u64 %rd230, [kernel_params+2160]; mul.wide.s32 %rd229, %r37, 208; cvta.to.global.u64 %rd228, %rd230; add.s64 %rd227, %rd228, %rd229; add.s64 %rd226, %rd227, 48; add.s64 %rd225, %rd226, 96; ld.global.u32 %r73, [%rd225+12]; ld.global.u32 %r74, [%rd225+8]; shl.b32 %r307, %r74, 1; cvt.rn.f32.s32 %f798, %r307; mul.ftz.f32 %f172, %f54, %f798; cvt.rzi.ftz.s32.f32 %r308, %f172; add.s32 %r309, %r307, -1; min.s32 %r75, %r308, %r309; ld.global.u32 %r425, [%rd225+20]; mul.wide.u32 %rd154, %r425, 16; add.s64 %rd155, %rd14, %rd154; ld.global.u16 %rs53, [%rd155+12]; ld.global.u32 %r424, [%rd155+8]; ld.global.u64 %rd231, [%rd155]; setp.eq.s64 %p81, %rd231, 11; @%p81 bra $L__BB156_67; $L__BB156_63: setp.eq.s64 %p82, %rd231, 0; @%p82 bra $L__BB156_65; add.s32 %r425, %r425, 2; bra.uni $L__BB156_66; $L__BB156_65: setp.eq.s16 %p83, %rs53, 0; mov.u32 %r428, -1; mov.u32 %r425, %r424; @%p83 bra $L__BB156_68; $L__BB156_66: mul.wide.u32 %rd156, %r425, 16; add.s64 %rd157, %rd14, %rd156; ld.global.u16 %rs53, [%rd157+12]; ld.global.u32 %r424, [%rd157+8]; ld.global.u64 %rd231, [%rd157]; setp.ne.s64 %p84, %rd231, 11; @%p84 bra $L__BB156_63; $L__BB156_67: setp.eq.s16 %p85, %rs53, 0; selp.b32 %r428, -1, %r424, %p85; $L__BB156_68: mul.wide.s32 %rd158, %r72, 16; add.s64 %rd159, %rd12, %rd158; ld.global.v4.u32 {%r311, %r312, %r313, %r314}, [%rd159]; setp.eq.s32 %p86, %r75, %r74; @%p86 bra $L__BB156_70; bra.uni $L__BB156_69; $L__BB156_70: mul.wide.u32 %rd163, %r314, 12; add.s64 %rd232, %rd13, %rd163; add.s32 %r321, %r314, 1; mul.wide.u32 %rd164, %r321, 12; add.s64 %rd235, %rd13, %rd164; add.s32 %r322, %r314, 2; mul.wide.u32 %rd165, %r322, 12; add.s64 %rd238, %rd13, %rd165; add.s64 %rd233, %rd232, 4; add.s64 %rd234, %rd232, 8; add.s64 %rd236, %rd235, 4; add.s64 %rd237, %rd235, 8; add.s64 %rd239, %rd238, 4; add.s64 %rd240, %rd238, 8; bra.uni $L__BB156_71; $L__BB156_69: setp.gt.s32 %p87, %r75, %r74; selp.b32 %r315, -1, 0, %p87; add.s32 %r316, %r75, %r315; mad.lo.s32 %r317, %r316, %r73, %r428; add.s32 %r318, %r311, %r317; mul.wide.u32 %rd160, %r318, 12; add.s64 %rd232, %rd15, %rd160; add.s32 %r319, %r312, %r317; mul.wide.u32 %rd161, %r319, 12; add.s64 %rd235, %rd15, %rd161; add.s32 %r320, %r313, %r317; mul.wide.u32 %rd162, %r320, 12; add.s64 %rd238, %rd15, %rd162; add.s64 %rd233, %rd232, 4; add.s64 %rd234, %rd232, 8; add.s64 %rd236, %rd235, 4; add.s64 %rd237, %rd235, 8; add.s64 %rd239, %rd238, 4; add.s64 %rd240, %rd238, 8; $L__BB156_71: ld.global.f32 %f173, [%rd240]; ld.global.f32 %f174, [%rd239]; ld.global.f32 %f175, [%rd238]; ld.global.f32 %f176, [%rd237]; ld.global.f32 %f177, [%rd236]; ld.global.f32 %f178, [%rd235]; ld.global.f32 %f179, [%rd234]; ld.global.f32 %f180, [%rd233]; ld.global.f32 %f181, [%rd232]; add.s32 %r90, %r75, 1; setp.eq.s32 %p88, %r90, %r74; @%p88 bra $L__BB156_73; bra.uni $L__BB156_72; $L__BB156_73: mul.wide.u32 %rd169, %r314, 12; add.s64 %rd241, %rd13, %rd169; add.s32 %r329, %r314, 1; mul.wide.u32 %rd170, %r329, 12; add.s64 %rd244, %rd13, %rd170; add.s32 %r330, %r314, 2; mul.wide.u32 %rd171, %r330, 12; add.s64 %rd247, %rd13, %rd171; add.s64 %rd242, %rd241, 4; add.s64 %rd243, %rd241, 8; add.s64 %rd245, %rd244, 4; add.s64 %rd246, %rd244, 8; add.s64 %rd248, %rd247, 4; add.s64 %rd249, %rd247, 8; bra.uni $L__BB156_74; $L__BB156_72: setp.ge.s32 %p89, %r75, %r74; selp.b32 %r323, -1, 0, %p89; add.s32 %r324, %r90, %r323; mad.lo.s32 %r325, %r324, %r73, %r428; add.s32 %r326, %r311, %r325; mul.wide.u32 %rd166, %r326, 12; add.s64 %rd241, %rd15, %rd166; add.s32 %r327, %r312, %r325; mul.wide.u32 %rd167, %r327, 12; add.s64 %rd244, %rd15, %rd167; add.s32 %r328, %r313, %r325; mul.wide.u32 %rd168, %r328, 12; add.s64 %rd247, %rd15, %rd168; add.s64 %rd242, %rd241, 4; add.s64 %rd243, %rd241, 8; add.s64 %rd245, %rd244, 4; add.s64 %rd246, %rd244, 8; add.s64 %rd248, %rd247, 4; add.s64 %rd249, %rd247, 8; $L__BB156_74: shl.b32 %r405, %r74, 1; cvt.rn.f32.s32 %f1367, %r405; mul.ftz.f32 %f1366, %f54, %f1367; mov.f32 %f1358, 0f3F800000; cvt.rn.f32.s32 %f799, %r75; sub.ftz.f32 %f800, %f1366, %f799; sub.ftz.f32 %f802, %f1358, %f800; ld.global.f32 %f803, [%rd241]; mul.ftz.f32 %f804, %f800, %f803; ld.global.f32 %f805, [%rd242]; mul.ftz.f32 %f806, %f800, %f805; ld.global.f32 %f807, [%rd243]; mul.ftz.f32 %f808, %f800, %f807; fma.rn.ftz.f32 %f809, %f802, %f181, %f804; fma.rn.ftz.f32 %f810, %f802, %f180, %f806; fma.rn.ftz.f32 %f811, %f802, %f179, %f808; ld.global.f32 %f812, [%rd244]; mul.ftz.f32 %f813, %f800, %f812; ld.global.f32 %f814, [%rd245]; mul.ftz.f32 %f815, %f800, %f814; ld.global.f32 %f816, [%rd246]; mul.ftz.f32 %f817, %f800, %f816; fma.rn.ftz.f32 %f818, %f802, %f178, %f813; fma.rn.ftz.f32 %f819, %f802, %f177, %f815; fma.rn.ftz.f32 %f820, %f802, %f176, %f817; ld.global.f32 %f821, [%rd247]; mul.ftz.f32 %f822, %f800, %f821; ld.global.f32 %f823, [%rd248]; mul.ftz.f32 %f824, %f800, %f823; ld.global.f32 %f825, [%rd249]; mul.ftz.f32 %f826, %f800, %f825; fma.rn.ftz.f32 %f827, %f802, %f175, %f822; fma.rn.ftz.f32 %f828, %f802, %f174, %f824; fma.rn.ftz.f32 %f829, %f802, %f173, %f826; sub.ftz.f32 %f182, %f809, %f1406; sub.ftz.f32 %f183, %f810, %f1407; sub.ftz.f32 %f184, %f811, %f1408; sub.ftz.f32 %f830, %f818, %f1406; sub.ftz.f32 %f831, %f819, %f1407; sub.ftz.f32 %f832, %f820, %f1408; sub.ftz.f32 %f833, %f827, %f1406; sub.ftz.f32 %f834, %f828, %f1407; sub.ftz.f32 %f835, %f829, %f1408; sub.ftz.f32 %f185, %f833, %f182; sub.ftz.f32 %f186, %f834, %f183; sub.ftz.f32 %f187, %f835, %f184; sub.ftz.f32 %f188, %f182, %f830; sub.ftz.f32 %f189, %f183, %f831; sub.ftz.f32 %f190, %f184, %f832; sub.ftz.f32 %f836, %f830, %f833; sub.ftz.f32 %f837, %f831, %f834; sub.ftz.f32 %f838, %f832, %f835; add.ftz.f32 %f839, %f833, %f182; add.ftz.f32 %f840, %f834, %f183; add.ftz.f32 %f841, %f835, %f184; mul.ftz.f32 %f842, %f841, %f186; mul.ftz.f32 %f843, %f187, %f840; sub.ftz.f32 %f844, %f842, %f843; mul.ftz.f32 %f845, %f187, %f839; mul.ftz.f32 %f846, %f841, %f185; sub.ftz.f32 %f847, %f845, %f846; mul.ftz.f32 %f848, %f840, %f185; mul.ftz.f32 %f849, %f186, %f839; sub.ftz.f32 %f850, %f848, %f849; mul.ftz.f32 %f851, %f1404, %f847; fma.rn.ftz.f32 %f852, %f1403, %f844, %f851; fma.rn.ftz.f32 %f853, %f1405, %f850, %f852; add.ftz.f32 %f854, %f830, %f182; add.ftz.f32 %f855, %f831, %f183; add.ftz.f32 %f856, %f832, %f184; mul.ftz.f32 %f857, %f856, %f189; mul.ftz.f32 %f858, %f190, %f855; sub.ftz.f32 %f859, %f857, %f858; mul.ftz.f32 %f860, %f190, %f854; mul.ftz.f32 %f861, %f856, %f188; sub.ftz.f32 %f862, %f860, %f861; mul.ftz.f32 %f863, %f855, %f188; mul.ftz.f32 %f864, %f189, %f854; sub.ftz.f32 %f865, %f863, %f864; mul.ftz.f32 %f866, %f1404, %f862; fma.rn.ftz.f32 %f867, %f1403, %f859, %f866; fma.rn.ftz.f32 %f868, %f1405, %f865, %f867; add.ftz.f32 %f869, %f833, %f830; add.ftz.f32 %f870, %f834, %f831; add.ftz.f32 %f871, %f835, %f832; mul.ftz.f32 %f872, %f871, %f837; mul.ftz.f32 %f873, %f838, %f870; sub.ftz.f32 %f874, %f872, %f873; mul.ftz.f32 %f875, %f838, %f869; mul.ftz.f32 %f876, %f871, %f836; sub.ftz.f32 %f877, %f875, %f876; mul.ftz.f32 %f878, %f870, %f836; mul.ftz.f32 %f879, %f837, %f869; sub.ftz.f32 %f880, %f878, %f879; mul.ftz.f32 %f881, %f1404, %f877; fma.rn.ftz.f32 %f882, %f1403, %f874, %f881; fma.rn.ftz.f32 %f883, %f1405, %f880, %f882; add.ftz.f32 %f884, %f853, %f868; add.ftz.f32 %f885, %f883, %f884; abs.ftz.f32 %f886, %f885; mul.ftz.f32 %f887, %f886, 0f34000000; min.ftz.f32 %f888, %f868, %f883; min.ftz.f32 %f889, %f853, %f888; max.ftz.f32 %f890, %f868, %f883; max.ftz.f32 %f891, %f853, %f890; mul.ftz.f32 %f892, %f886, 0fB4000000; setp.ltu.ftz.f32 %p90, %f889, %f892; setp.gtu.ftz.f32 %p91, %f891, %f887; and.pred %p92, %p90, %p91; @%p92 bra $L__BB156_77; mul.ftz.f32 %f893, %f187, %f189; mul.ftz.f32 %f894, %f190, %f186; sub.ftz.f32 %f895, %f893, %f894; mul.ftz.f32 %f896, %f187, %f188; mul.ftz.f32 %f897, %f190, %f185; sub.ftz.f32 %f898, %f897, %f896; mul.ftz.f32 %f899, %f189, %f185; mul.ftz.f32 %f900, %f186, %f188; sub.ftz.f32 %f901, %f900, %f899; add.ftz.f32 %f191, %f895, %f895; add.ftz.f32 %f192, %f898, %f898; add.ftz.f32 %f193, %f901, %f901; mul.ftz.f32 %f902, %f1404, %f192; fma.rn.ftz.f32 %f903, %f1403, %f191, %f902; fma.rn.ftz.f32 %f194, %f1405, %f193, %f903; setp.eq.ftz.f32 %p93, %f194, 0f00000000; @%p93 bra $L__BB156_77; mul.ftz.f32 %f904, %f183, %f192; fma.rn.ftz.f32 %f905, %f182, %f191, %f904; fma.rn.ftz.f32 %f906, %f184, %f193, %f905; div.approx.ftz.f32 %f907, %f906, %f194; setp.ge.ftz.f32 %p94, %f907, 0f00000000; setp.le.ftz.f32 %p95, %f907, %f13; mov.u16 %rs60, 1; and.pred %p96, %p95, %p94; @%p96 bra $L__BB156_128; $L__BB156_77: add.s32 %r430, %r65, -1; cvt.u32.u64 %r331, %rd26; add.s32 %r423, %r331, 1; setp.lt.s32 %p97, %r423, %r302; @%p97 bra $L__BB156_59; bra.uni $L__BB156_86; $L__BB156_78: setp.ge.s32 %p98, %r423, %r302; @%p98 bra $L__BB156_86; $L__BB156_79: cvt.s64.s32 %rd84, %r423; setp.ne.s32 %p99, %r413, -1; @%p99 bra $L__BB156_81; shl.b64 %rd172, %rd84, 2; add.s64 %rd173, %rd10, %rd172; ld.global.u32 %r332, [%rd173]; setp.ne.s32 %p100, %r332, %r37; @%p100 bra $L__BB156_85; $L__BB156_81: shl.b64 %rd174, %rd84, 2; add.s64 %rd175, %rd11, %rd174; ld.global.u32 %r93, [%rd175]; setp.eq.s32 %p101, %r38, %r93; @%p101 bra $L__BB156_85; mul.wide.s32 %rd176, %r93, 16; add.s64 %rd177, %rd12, %rd176; ld.global.u32 %r333, [%rd177+12]; mul.wide.u32 %rd178, %r333, 12; add.s64 %rd179, %rd13, %rd178; add.s32 %r334, %r333, 1; mul.wide.u32 %rd180, %r334, 12; add.s64 %rd181, %rd13, %rd180; add.s32 %r335, %r333, 2; mul.wide.u32 %rd182, %r335, 12; add.s64 %rd183, %rd13, %rd182; ld.global.f32 %f908, [%rd179]; sub.ftz.f32 %f195, %f908, %f1406; ld.global.f32 %f909, [%rd179+4]; sub.ftz.f32 %f196, %f909, %f1407; ld.global.f32 %f910, [%rd179+8]; sub.ftz.f32 %f197, %f910, %f1408; ld.global.f32 %f911, [%rd181]; sub.ftz.f32 %f912, %f911, %f1406; ld.global.f32 %f913, [%rd181+4]; sub.ftz.f32 %f914, %f913, %f1407; ld.global.f32 %f915, [%rd181+8]; sub.ftz.f32 %f916, %f915, %f1408; ld.global.f32 %f917, [%rd183]; sub.ftz.f32 %f918, %f917, %f1406; ld.global.f32 %f919, [%rd183+4]; sub.ftz.f32 %f920, %f919, %f1407; ld.global.f32 %f921, [%rd183+8]; sub.ftz.f32 %f922, %f921, %f1408; sub.ftz.f32 %f198, %f918, %f195; sub.ftz.f32 %f199, %f920, %f196; sub.ftz.f32 %f200, %f922, %f197; sub.ftz.f32 %f201, %f195, %f912; sub.ftz.f32 %f202, %f196, %f914; sub.ftz.f32 %f203, %f197, %f916; sub.ftz.f32 %f923, %f912, %f918; sub.ftz.f32 %f924, %f914, %f920; sub.ftz.f32 %f925, %f916, %f922; add.ftz.f32 %f926, %f195, %f918; add.ftz.f32 %f927, %f196, %f920; add.ftz.f32 %f928, %f197, %f922; mul.ftz.f32 %f929, %f199, %f928; mul.ftz.f32 %f930, %f927, %f200; sub.ftz.f32 %f931, %f929, %f930; mul.ftz.f32 %f932, %f926, %f200; mul.ftz.f32 %f933, %f198, %f928; sub.ftz.f32 %f934, %f932, %f933; mul.ftz.f32 %f935, %f198, %f927; mul.ftz.f32 %f936, %f926, %f199; sub.ftz.f32 %f937, %f935, %f936; mul.ftz.f32 %f938, %f1404, %f934; fma.rn.ftz.f32 %f939, %f1403, %f931, %f938; fma.rn.ftz.f32 %f940, %f1405, %f937, %f939; add.ftz.f32 %f941, %f195, %f912; add.ftz.f32 %f942, %f196, %f914; add.ftz.f32 %f943, %f197, %f916; mul.ftz.f32 %f944, %f202, %f943; mul.ftz.f32 %f945, %f942, %f203; sub.ftz.f32 %f946, %f944, %f945; mul.ftz.f32 %f947, %f941, %f203; mul.ftz.f32 %f948, %f201, %f943; sub.ftz.f32 %f949, %f947, %f948; mul.ftz.f32 %f950, %f201, %f942; mul.ftz.f32 %f951, %f941, %f202; sub.ftz.f32 %f952, %f950, %f951; mul.ftz.f32 %f953, %f1404, %f949; fma.rn.ftz.f32 %f954, %f1403, %f946, %f953; fma.rn.ftz.f32 %f955, %f1405, %f952, %f954; add.ftz.f32 %f956, %f912, %f918; add.ftz.f32 %f957, %f914, %f920; add.ftz.f32 %f958, %f916, %f922; mul.ftz.f32 %f959, %f924, %f958; mul.ftz.f32 %f960, %f957, %f925; sub.ftz.f32 %f961, %f959, %f960; mul.ftz.f32 %f962, %f956, %f925; mul.ftz.f32 %f963, %f923, %f958; sub.ftz.f32 %f964, %f962, %f963; mul.ftz.f32 %f965, %f923, %f957; mul.ftz.f32 %f966, %f956, %f924; sub.ftz.f32 %f967, %f965, %f966; mul.ftz.f32 %f968, %f1404, %f964; fma.rn.ftz.f32 %f969, %f1403, %f961, %f968; fma.rn.ftz.f32 %f970, %f1405, %f967, %f969; add.ftz.f32 %f971, %f955, %f940; add.ftz.f32 %f972, %f970, %f971; abs.ftz.f32 %f973, %f972; mul.ftz.f32 %f974, %f973, 0f34000000; min.ftz.f32 %f975, %f955, %f970; min.ftz.f32 %f976, %f940, %f975; max.ftz.f32 %f977, %f955, %f970; max.ftz.f32 %f978, %f940, %f977; mul.ftz.f32 %f979, %f973, 0fB4000000; setp.ltu.ftz.f32 %p102, %f976, %f979; setp.gtu.ftz.f32 %p103, %f978, %f974; and.pred %p104, %p102, %p103; @%p104 bra $L__BB156_85; mul.ftz.f32 %f980, %f202, %f200; mul.ftz.f32 %f981, %f203, %f199; sub.ftz.f32 %f982, %f980, %f981; mul.ftz.f32 %f983, %f201, %f200; mul.ftz.f32 %f984, %f203, %f198; sub.ftz.f32 %f985, %f984, %f983; mul.ftz.f32 %f986, %f202, %f198; mul.ftz.f32 %f987, %f201, %f199; sub.ftz.f32 %f988, %f987, %f986; add.ftz.f32 %f204, %f982, %f982; add.ftz.f32 %f205, %f985, %f985; add.ftz.f32 %f206, %f988, %f988; mul.ftz.f32 %f989, %f1404, %f205; fma.rn.ftz.f32 %f990, %f1403, %f204, %f989; fma.rn.ftz.f32 %f207, %f1405, %f206, %f990; setp.eq.ftz.f32 %p105, %f207, 0f00000000; @%p105 bra $L__BB156_85; mul.ftz.f32 %f991, %f196, %f205; fma.rn.ftz.f32 %f992, %f195, %f204, %f991; fma.rn.ftz.f32 %f993, %f197, %f206, %f992; div.approx.ftz.f32 %f994, %f993, %f207; setp.ge.ftz.f32 %p106, %f994, 0f00000000; setp.le.ftz.f32 %p107, %f994, %f13; mov.u16 %rs60, 1; and.pred %p108, %p107, %p106; @%p108 bra $L__BB156_128; $L__BB156_85: add.s32 %r430, %r65, -1; cvt.u32.u64 %r336, %rd84; add.s32 %r423, %r336, 1; setp.lt.s32 %p109, %r423, %r302; @%p109 bra $L__BB156_79; bra.uni $L__BB156_86; $L__BB156_44: setp.eq.s32 %p52, %r420, 1985229328; setp.lt.s32 %p53, %r420, 0; or.pred %p54, %p53, %p52; @%p54 bra $L__BB156_55; $L__BB156_45: mul.wide.s32 %rd142, %r420, 16; add.s64 %rd24, %rd8, %rd142; ld.global.v4.f32 {%f583, %f584, %f585, %f586}, [%rd24]; mov.b32 %r420, %f586; mov.b32 %r50, %f585; mov.b32 %r49, %f584; mov.b32 %r48, %f583; mov.b64 %rd143, {%r48, %r49}; cvt.u32.u64 %r291, %rd143; and.b32 %r292, %r291, 1024; setp.eq.s32 %p55, %r292, 0; and.b32 %r52, %r48, 2047; @%p55 bra $L__BB156_48; bra.uni $L__BB156_46; $L__BB156_48: setp.ne.s32 %p61, %r52, 0; ld.global.v4.f32 {%f727, %f728, %f729, %f730}, [%rd24+16]; ld.global.v4.f32 {%f735, %f736, %f737, %f738}, [%rd24+32]; ld.global.v4.f32 {%f743, %f744, %f745, %f746}, [%rd24+48]; sub.ftz.f32 %f751, %f727, %f1406; mul.ftz.f32 %f752, %f163, %f751; sub.ftz.f32 %f753, %f729, %f1406; mul.ftz.f32 %f754, %f163, %f753; sub.ftz.f32 %f755, %f735, %f1407; mul.ftz.f32 %f756, %f162, %f755; sub.ftz.f32 %f757, %f737, %f1407; mul.ftz.f32 %f758, %f162, %f757; sub.ftz.f32 %f759, %f743, %f1408; mul.ftz.f32 %f760, %f164, %f759; sub.ftz.f32 %f761, %f745, %f1408; mul.ftz.f32 %f762, %f164, %f761; min.ftz.f32 %f763, %f752, %f754; min.ftz.f32 %f764, %f756, %f758; min.ftz.f32 %f765, %f760, %f762; mov.f32 %f766, 0f00000000; max.ftz.f32 %f767, %f766, %f763; max.ftz.f32 %f768, %f764, %f765; max.ftz.f32 %f1409, %f767, %f768; max.ftz.f32 %f769, %f752, %f754; max.ftz.f32 %f770, %f756, %f758; max.ftz.f32 %f771, %f760, %f762; min.ftz.f32 %f772, %f13, %f769; min.ftz.f32 %f773, %f770, %f771; min.ftz.f32 %f774, %f772, %f773; sub.ftz.f32 %f775, %f728, %f1406; mul.ftz.f32 %f776, %f163, %f775; sub.ftz.f32 %f777, %f730, %f1406; mul.ftz.f32 %f778, %f163, %f777; sub.ftz.f32 %f779, %f736, %f1407; mul.ftz.f32 %f780, %f162, %f779; sub.ftz.f32 %f781, %f738, %f1407; mul.ftz.f32 %f782, %f162, %f781; sub.ftz.f32 %f783, %f744, %f1408; mul.ftz.f32 %f784, %f164, %f783; sub.ftz.f32 %f785, %f746, %f1408; mul.ftz.f32 %f786, %f164, %f785; min.ftz.f32 %f787, %f776, %f778; min.ftz.f32 %f788, %f780, %f782; min.ftz.f32 %f789, %f784, %f786; max.ftz.f32 %f790, %f766, %f787; max.ftz.f32 %f791, %f788, %f789; max.ftz.f32 %f1410, %f790, %f791; max.ftz.f32 %f792, %f776, %f778; max.ftz.f32 %f793, %f780, %f782; max.ftz.f32 %f794, %f784, %f786; min.ftz.f32 %f795, %f13, %f792; min.ftz.f32 %f796, %f793, %f794; min.ftz.f32 %f797, %f795, %f796; setp.ge.ftz.f32 %p62, %f774, %f1409; and.pred %p63, %p61, %p62; selp.u32 %r296, 1, 0, %p63; setp.ltu.ftz.f32 %p64, %f797, %f1410; and.b32 %r298, %r49, 2047; setp.eq.s32 %p65, %r298, 0; or.pred %p66, %p65, %p64; selp.b32 %r299, 0, 2, %p66; or.b32 %r418, %r299, %r296; bra.uni $L__BB156_49; $L__BB156_46: setp.ne.s32 %p56, %r52, 0; ld.global.v4.f32 {%f590, %f591, %f592, %f593}, [%rd24+16]; ld.global.v4.f32 {%f598, %f599, %f600, %f601}, [%rd24+32]; ld.global.v4.f32 {%f606, %f607, %f608, %f609}, [%rd24+48]; mul.ftz.f32 %f614, %f1404, %f591; fma.rn.ftz.f32 %f615, %f1403, %f590, %f614; fma.rn.ftz.f32 %f616, %f1405, %f592, %f615; mul.ftz.f32 %f617, %f1404, %f599; fma.rn.ftz.f32 %f618, %f1403, %f598, %f617; fma.rn.ftz.f32 %f619, %f1405, %f600, %f618; mul.ftz.f32 %f620, %f1404, %f607; fma.rn.ftz.f32 %f621, %f1403, %f606, %f620; fma.rn.ftz.f32 %f622, %f1405, %f608, %f621; mul.ftz.f32 %f623, %f1407, %f591; fma.rn.ftz.f32 %f624, %f1406, %f590, %f623; fma.rn.ftz.f32 %f625, %f1408, %f592, %f624; add.ftz.f32 %f626, %f593, %f625; mul.ftz.f32 %f627, %f1407, %f599; fma.rn.ftz.f32 %f628, %f1406, %f598, %f627; fma.rn.ftz.f32 %f629, %f1408, %f600, %f628; add.ftz.f32 %f630, %f601, %f629; mul.ftz.f32 %f631, %f1407, %f607; fma.rn.ftz.f32 %f632, %f1406, %f606, %f631; fma.rn.ftz.f32 %f633, %f1408, %f608, %f632; add.ftz.f32 %f634, %f609, %f633; rcp.approx.ftz.f32 %f635, %f616; rcp.approx.ftz.f32 %f636, %f619; rcp.approx.ftz.f32 %f637, %f622; mul.ftz.f32 %f638, %f635, %f626; neg.ftz.f32 %f639, %f638; mul.ftz.f32 %f640, %f636, %f630; neg.ftz.f32 %f641, %f640; mul.ftz.f32 %f642, %f637, %f634; neg.ftz.f32 %f643, %f642; sub.ftz.f32 %f644, %f635, %f638; sub.ftz.f32 %f645, %f636, %f640; sub.ftz.f32 %f646, %f637, %f642; min.ftz.f32 %f647, %f639, %f644; min.ftz.f32 %f648, %f641, %f645; min.ftz.f32 %f649, %f643, %f646; max.ftz.f32 %f650, %f639, %f644; max.ftz.f32 %f651, %f641, %f645; max.ftz.f32 %f652, %f643, %f646; mov.f32 %f653, 0f00000000; max.ftz.f32 %f654, %f653, %f647; max.ftz.f32 %f655, %f648, %f649; max.ftz.f32 %f1409, %f654, %f655; min.ftz.f32 %f656, %f13, %f650; min.ftz.f32 %f657, %f651, %f652; min.ftz.f32 %f658, %f656, %f657; setp.le.ftz.f32 %p57, %f1409, %f658; and.pred %p58, %p56, %p57; selp.u32 %r418, 1, 0, %p58; ld.global.v4.f32 {%f659, %f660, %f661, %f662}, [%rd24+64]; ld.global.v4.f32 {%f667, %f668, %f669, %f670}, [%rd24+80]; ld.global.v4.f32 {%f675, %f676, %f677, %f678}, [%rd24+96]; mul.ftz.f32 %f683, %f1404, %f660; fma.rn.ftz.f32 %f684, %f1403, %f659, %f683; fma.rn.ftz.f32 %f685, %f1405, %f661, %f684; mul.ftz.f32 %f686, %f1404, %f668; fma.rn.ftz.f32 %f687, %f1403, %f667, %f686; fma.rn.ftz.f32 %f688, %f1405, %f669, %f687; mul.ftz.f32 %f689, %f1404, %f676; fma.rn.ftz.f32 %f690, %f1403, %f675, %f689; fma.rn.ftz.f32 %f691, %f1405, %f677, %f690; mul.ftz.f32 %f692, %f1407, %f660; fma.rn.ftz.f32 %f693, %f1406, %f659, %f692; fma.rn.ftz.f32 %f694, %f1408, %f661, %f693; add.ftz.f32 %f695, %f662, %f694; mul.ftz.f32 %f696, %f1407, %f668; fma.rn.ftz.f32 %f697, %f1406, %f667, %f696; fma.rn.ftz.f32 %f698, %f1408, %f669, %f697; add.ftz.f32 %f699, %f670, %f698; mul.ftz.f32 %f700, %f1407, %f676; fma.rn.ftz.f32 %f701, %f1406, %f675, %f700; fma.rn.ftz.f32 %f702, %f1408, %f677, %f701; add.ftz.f32 %f703, %f678, %f702; rcp.approx.ftz.f32 %f704, %f685; rcp.approx.ftz.f32 %f705, %f688; rcp.approx.ftz.f32 %f706, %f691; mul.ftz.f32 %f707, %f704, %f695; neg.ftz.f32 %f708, %f707; mul.ftz.f32 %f709, %f705, %f699; neg.ftz.f32 %f710, %f709; mul.ftz.f32 %f711, %f706, %f703; neg.ftz.f32 %f712, %f711; sub.ftz.f32 %f713, %f704, %f707; sub.ftz.f32 %f714, %f705, %f709; sub.ftz.f32 %f715, %f706, %f711; min.ftz.f32 %f716, %f708, %f713; min.ftz.f32 %f717, %f710, %f714; min.ftz.f32 %f718, %f712, %f715; max.ftz.f32 %f719, %f708, %f713; max.ftz.f32 %f720, %f710, %f714; max.ftz.f32 %f721, %f712, %f715; max.ftz.f32 %f722, %f653, %f716; max.ftz.f32 %f723, %f717, %f718; max.ftz.f32 %f1410, %f722, %f723; min.ftz.f32 %f724, %f13, %f719; min.ftz.f32 %f725, %f720, %f721; min.ftz.f32 %f726, %f724, %f725; setp.gtu.ftz.f32 %p59, %f1410, %f726; @%p59 bra $L__BB156_49; and.b32 %r294, %r49, 2047; setp.eq.s32 %p60, %r294, 0; or.b32 %r295, %r418, 2; selp.b32 %r418, %r418, %r295, %p60; $L__BB156_49: mul.wide.s32 %rd144, %r430, 4; add.s64 %rd25, %rd2, %rd144; setp.eq.s32 %p67, %r418, 3; @%p67 bra $L__BB156_53; bra.uni $L__BB156_50; $L__BB156_53: setp.lt.ftz.f32 %p70, %f1410, %f1409; selp.b32 %r61, %r420, %r50, %p70; selp.b32 %r300, %r50, %r420, %p70; add.s32 %r430, %r430, 1; st.local.u32 [%rd25+4], %r300; mov.u32 %r420, %r61; bra.uni $L__BB156_54; $L__BB156_50: setp.eq.s32 %p68, %r418, 2; @%p68 bra $L__BB156_54; setp.ne.s32 %p69, %r418, 0; mov.u32 %r420, %r50; @%p69 bra $L__BB156_54; ld.local.u32 %r420, [%rd25]; add.s32 %r430, %r430, -1; $L__BB156_54: setp.ne.s32 %p71, %r420, 1985229328; setp.gt.s32 %p72, %r420, -1; and.pred %p73, %p72, %p71; @%p73 bra $L__BB156_45; $L__BB156_55: mov.u32 %r65, %r430; setp.gt.s32 %p74, %r420, -1; @%p74 bra $L__BB156_86; bra.uni $L__BB156_56; $L__BB156_86: setp.ne.s32 %p110, %r420, 1985229328; mov.u16 %rs60, 0; @%p110 bra $L__BB156_44; $L__BB156_128: setp.eq.s16 %p170, %rs60, 0; selp.u32 %r382, 1, 0, %p170; add.s32 %r408, %r408, %r382; add.s32 %r407, %r407, 1; setp.lt.u32 %p171, %r407, %r7; @%p171 bra $L__BB156_9; cvt.rn.f32.s32 %f1350, %r408; cvt.rn.f32.s32 %f1351, %r7; div.approx.ftz.f32 %f1419, %f1350, %f1351; $L__BB156_130: ld.param.u32 %r385, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; shr.u32 %r384, %r385, 24; setp.eq.s32 %p172, %r384, 255; @%p172 bra $L__BB156_132; ld.param.u32 %r387, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+4]; shr.u32 %r386, %r387, 24; ld.param.u64 %rd220, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; cvta.to.local.u64 %rd219, %rd220; mul.wide.u32 %rd212, %r386, 4; add.s64 %rd213, %rd219, %rd212; st.local.f32 [%rd213], %f1419; $L__BB156_132: ld.param.u32 %r389, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+8]; bfe.u32 %r388, %r389, 8, 8; setp.eq.s32 %p173, %r388, 255; @%p173 bra $L__BB156_134; ld.param.u32 %r392, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+8]; ld.param.u32 %r391, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3+8]; bfe.u32 %r390, %r391, 8, 8; ld.param.u64 %rd222, [_Z11svm_node_aoILj511EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; cvta.to.local.u64 %rd221, %rd222; shl.b32 %r383, %r391, 2; cvt.u64.u32 %rd214, %r383; and.b64 %rd215, %rd214, 1020; add.s64 %rd216, %rd221, %rd215; ld.local.f32 %f1352, [%rd216]; mul.ftz.f32 %f1353, %f1419, %f1352; ld.local.f32 %f1354, [%rd216+4]; mul.ftz.f32 %f1355, %f1419, %f1354; ld.local.f32 %f1356, [%rd216+8]; mul.ftz.f32 %f1357, %f1419, %f1356; mul.wide.u32 %rd217, %r390, 4; add.s64 %rd218, %rd221, %rd217; st.local.f32 [%rd218], %f1353; st.local.f32 [%rd218+4], %f1355; st.local.f32 [%rd218+8], %f1357; $L__BB156_134: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) { .reg .pred %p<463>; .reg .b16 %rs<22>; .reg .f32 %f<3060>; .reg .b32 %r<664>; .reg .b64 %rd<808>; ld.param.u64 %rd230, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd231, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r141, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r142, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; ld.param.u32 %r140, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r139, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r138, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd230; cvta.to.local.u64 %rd2, %rd231; and.b32 %r2, %r138, 255; bfe.u32 %r3, %r138, 8, 8; bfe.u32 %r4, %r138, 16, 8; shr.u32 %r5, %r138, 24; setp.eq.s32 %p7, %r5, 255; mov.f32 %f2, 0f3F800000; @%p7 bra $L__BB157_2; mul.wide.u32 %rd232, %r5, 4; add.s64 %rd233, %rd2, %rd232; ld.local.f32 %f2, [%rd233]; $L__BB157_2: ld.const.u64 %rd234, [kernel_params+2400]; cvta.to.global.u64 %rd235, %rd234; mul.wide.s32 %rd236, %r142, 16; add.s64 %rd3, %rd235, %rd236; ld.global.v4.u32 {%r143, %r144, %r145, %r146}, [%rd3]; add.s32 %r663, %r142, 1; setp.eq.ftz.f32 %p8, %f2, 0f00000000; @%p8 bra $L__BB157_516; bra.uni $L__BB157_3; $L__BB157_516: setp.eq.s32 %p461, %r2, 39; add.s32 %r636, %r142, 5; selp.b32 %r663, %r636, %r663, %p461; $L__BB157_517: st.param.b32 [func_retval0+0], %r663; ret; $L__BB157_3: setp.eq.s32 %p9, %r143, 255; add.s64 %rd4, %rd1, 12; @%p9 bra $L__BB157_5; bra.uni $L__BB157_4; $L__BB157_5: ld.local.f32 %f55, [%rd4]; ld.local.f32 %f56, [%rd4+4]; ld.local.f32 %f57, [%rd4+8]; bra.uni $L__BB157_6; $L__BB157_4: mul.wide.u32 %rd237, %r143, 4; add.s64 %rd238, %rd2, %rd237; ld.local.f32 %f55, [%rd238]; ld.local.f32 %f56, [%rd238+4]; ld.local.f32 %f57, [%rd238+8]; $L__BB157_6: ld.local.u32 %r11, [%rd4+52]; and.b32 %r12, %r11, 6; setp.ne.s32 %p10, %r12, 0; @%p10 bra $L__BB157_15; ld.local.v2.f32 {%f550, %f551}, [%rd4+12]; ld.local.v2.f32 {%f552, %f553}, [%rd4+20]; ld.local.v2.f32 {%f554, %f555}, [%rd4+28]; mul.ftz.f32 %f556, %f56, %f554; fma.rn.ftz.f32 %f557, %f55, %f553, %f556; fma.rn.ftz.f32 %f558, %f57, %f555, %f557; add.ftz.f32 %f559, %f558, %f558; mul.ftz.f32 %f560, %f55, %f559; mul.ftz.f32 %f561, %f56, %f559; mul.ftz.f32 %f562, %f57, %f559; sub.ftz.f32 %f563, %f560, %f553; sub.ftz.f32 %f564, %f561, %f554; sub.ftz.f32 %f565, %f562, %f555; mul.ftz.f32 %f566, %f551, %f554; fma.rn.ftz.f32 %f567, %f550, %f553, %f566; fma.rn.ftz.f32 %f18, %f552, %f555, %f567; mul.ftz.f32 %f568, %f18, 0f3F666666; mov.f32 %f569, 0f3C23D70A; min.ftz.f32 %f19, %f568, %f569; mul.ftz.f32 %f570, %f551, %f564; fma.rn.ftz.f32 %f571, %f550, %f563, %f570; fma.rn.ftz.f32 %f572, %f552, %f565, %f571; setp.ltu.ftz.f32 %p11, %f572, %f19; @%p11 bra $L__BB157_8; bra.uni $L__BB157_15; $L__BB157_8: mul.ftz.f32 %f573, %f56, %f551; fma.rn.ftz.f32 %f574, %f55, %f550, %f573; fma.rn.ftz.f32 %f575, %f57, %f552, %f574; mul.ftz.f32 %f576, %f550, %f575; mul.ftz.f32 %f577, %f551, %f575; mul.ftz.f32 %f578, %f552, %f575; sub.ftz.f32 %f579, %f55, %f576; sub.ftz.f32 %f580, %f56, %f577; sub.ftz.f32 %f581, %f57, %f578; mul.ftz.f32 %f582, %f580, %f580; fma.rn.ftz.f32 %f583, %f579, %f579, %f582; fma.rn.ftz.f32 %f584, %f581, %f581, %f583; sqrt.approx.ftz.f32 %f585, %f584; div.approx.ftz.f32 %f23, %f579, %f585; div.approx.ftz.f32 %f24, %f580, %f585; div.approx.ftz.f32 %f25, %f581, %f585; mul.ftz.f32 %f586, %f554, %f24; fma.rn.ftz.f32 %f587, %f553, %f23, %f586; fma.rn.ftz.f32 %f26, %f555, %f25, %f587; mul.ftz.f32 %f588, %f26, %f26; fma.rn.ftz.f32 %f589, %f18, %f18, %f588; mul.ftz.f32 %f590, %f19, %f19; sub.ftz.f32 %f591, %f589, %f590; mul.ftz.f32 %f592, %f588, %f591; mov.f32 %f593, 0f00000000; max.ftz.f32 %f594, %f592, %f593; sqrt.approx.ftz.f32 %f595, %f594; fma.rn.ftz.f32 %f596, %f19, %f18, %f589; mov.f32 %f597, 0f3F000000; div.approx.ftz.f32 %f598, %f597, %f589; add.ftz.f32 %f599, %f595, %f596; mul.ftz.f32 %f27, %f598, %f599; sub.ftz.f32 %f600, %f596, %f595; mul.ftz.f32 %f28, %f598, %f600; setp.gt.ftz.f32 %p12, %f27, 0f3727C5AC; setp.le.ftz.f32 %p13, %f27, 0f3F800054; and.pred %p1, %p12, %p13; setp.gt.ftz.f32 %p14, %f28, 0f3727C5AC; setp.le.ftz.f32 %p15, %f28, 0f3F800054; and.pred %p2, %p14, %p15; and.pred %p16, %p1, %p2; @%p16 bra $L__BB157_11; bra.uni $L__BB157_9; $L__BB157_11: mov.f32 %f607, 0f3F800000; sub.ftz.f32 %f608, %f607, %f27; max.ftz.f32 %f610, %f608, %f593; sqrt.approx.ftz.f32 %f34, %f610; max.ftz.f32 %f611, %f27, %f593; sqrt.approx.ftz.f32 %f35, %f611; sub.ftz.f32 %f612, %f607, %f28; max.ftz.f32 %f613, %f612, %f593; sqrt.approx.ftz.f32 %f36, %f613; max.ftz.f32 %f614, %f28, %f593; sqrt.approx.ftz.f32 %f37, %f614; mul.ftz.f32 %f615, %f18, %f35; fma.rn.ftz.f32 %f616, %f26, %f34, %f615; add.ftz.f32 %f617, %f616, %f616; mul.ftz.f32 %f618, %f35, %f617; sub.ftz.f32 %f38, %f618, %f18; mul.ftz.f32 %f619, %f18, %f37; fma.rn.ftz.f32 %f620, %f26, %f36, %f619; add.ftz.f32 %f621, %f620, %f620; mul.ftz.f32 %f622, %f37, %f621; sub.ftz.f32 %f39, %f622, %f18; setp.ge.ftz.f32 %p18, %f38, 0f3727C5AC; setp.ge.ftz.f32 %p19, %f39, 0f3727C5AC; and.pred %p20, %p18, %p19; @%p20 bra $L__BB157_13; bra.uni $L__BB157_12; $L__BB157_13: setp.lt.ftz.f32 %p22, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p22; selp.f32 %f2968, %f35, %f37, %p22; bra.uni $L__BB157_14; $L__BB157_9: or.pred %p17, %p1, %p2; mov.f32 %f55, %f550; mov.f32 %f56, %f551; mov.f32 %f57, %f552; @%p17 bra $L__BB157_10; bra.uni $L__BB157_15; $L__BB157_10: selp.f32 %f601, %f27, %f28, %p1; mov.f32 %f602, 0f3F800000; sub.ftz.f32 %f603, %f602, %f601; max.ftz.f32 %f605, %f603, %f593; sqrt.approx.ftz.f32 %f2967, %f605; max.ftz.f32 %f606, %f601, %f593; sqrt.approx.ftz.f32 %f2968, %f606; bra.uni $L__BB157_14; $L__BB157_12: setp.gt.ftz.f32 %p21, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p21; selp.f32 %f2968, %f35, %f37, %p21; $L__BB157_14: mul.ftz.f32 %f623, %f550, %f2968; mul.ftz.f32 %f624, %f551, %f2968; mul.ftz.f32 %f625, %f552, %f2968; fma.rn.ftz.f32 %f57, %f25, %f2967, %f625; fma.rn.ftz.f32 %f56, %f24, %f2967, %f624; fma.rn.ftz.f32 %f55, %f23, %f2967, %f623; $L__BB157_15: setp.eq.s32 %p23, %r3, 255; @%p23 bra $L__BB157_17; bra.uni $L__BB157_16; $L__BB157_17: mov.b32 %f60, %r139; bra.uni $L__BB157_18; $L__BB157_16: mul.wide.u32 %rd239, %r3, 4; add.s64 %rd240, %rd2, %rd239; ld.local.f32 %f60, [%rd240]; $L__BB157_18: setp.eq.s32 %p24, %r4, 255; @%p24 bra $L__BB157_20; bra.uni $L__BB157_19; $L__BB157_20: mov.b32 %f63, %r140; bra.uni $L__BB157_21; $L__BB157_19: mul.wide.u32 %rd241, %r4, 4; add.s64 %rd242, %rd2, %rd241; ld.local.f32 %f63, [%rd242]; $L__BB157_21: cvt.u16.u32 %rs8, %r138; and.b16 %rs7, %rs8, 255; mul.wide.u32 %rd243, %r144, 4; add.s64 %rd5, %rd2, %rd243; mul.wide.u32 %rd244, %r145, 4; add.s64 %rd6, %rd2, %rd244; mul.wide.u32 %rd245, %r146, 4; add.s64 %rd7, %rd2, %rd245; setp.gt.s16 %p25, %rs7, 23; @%p25 bra $L__BB157_63; setp.gt.s16 %p36, %rs7, 15; @%p36 bra $L__BB157_48; setp.gt.s16 %p43, %rs7, 8; @%p43 bra $L__BB157_34; bra.uni $L__BB157_24; $L__BB157_34: add.s16 %rs12, %rs7, -9; setp.lt.u16 %p44, %rs12, 2; @%p44 bra $L__BB157_36; add.s16 %rs13, %rs7, -13; setp.lt.u16 %p45, %rs13, 2; @%p45 bra $L__BB157_36; bra.uni $L__BB157_517; $L__BB157_63: setp.gt.s16 %p26, %rs7, 30; @%p26 bra $L__BB157_88; setp.gt.s16 %p30, %rs7, 27; @%p30 bra $L__BB157_77; add.s16 %rs10, %rs7, -25; setp.lt.u16 %p34, %rs10, 2; @%p34 bra $L__BB157_427; setp.eq.s16 %p35, %rs7, 24; @%p35 bra $L__BB157_67; bra.uni $L__BB157_517; $L__BB157_67: ld.const.u32 %r496, [kernel_params+1564]; setp.ne.s32 %p385, %r496, 0; @%p385 bra $L__BB157_69; ld.const.u32 %r497, [kernel_params+1568]; setp.eq.s32 %p386, %r497, 0; shr.u32 %r498, %r141, 3; and.b32 %r499, %r498, 1; setp.eq.b32 %p387, %r499, 1; and.pred %p388, %p387, %p386; @%p388 bra $L__BB157_517; $L__BB157_69: ld.local.f32 %f2726, [%rd4+256]; mov.u64 %rd774, 0; mul.ftz.f32 %f2727, %f2, %f2726; ld.local.f32 %f2728, [%rd4+260]; mul.ftz.f32 %f2729, %f2, %f2728; ld.local.f32 %f2730, [%rd4+264]; mul.ftz.f32 %f2731, %f2, %f2730; mov.f32 %f2732, 0f00000000; max.ftz.f32 %f493, %f2727, %f2732; max.ftz.f32 %f494, %f2729, %f2732; max.ftz.f32 %f495, %f2731, %f2732; add.ftz.f32 %f2733, %f493, %f494; add.ftz.f32 %f2734, %f495, %f2733; mul.ftz.f32 %f2735, %f2734, 0f3EAAAAAB; abs.ftz.f32 %f496, %f2735; setp.ltu.ftz.f32 %p389, %f496, 0f3727C5AC; mov.u64 %rd775, %rd774; @%p389 bra $L__BB157_74; ld.local.u32 %r125, [%rd4+252]; mov.u64 %rd774, 0; setp.eq.s32 %p390, %r125, 0; mov.u64 %rd772, %rd774; mov.u64 %rd773, %rd774; @%p390 bra $L__BB157_72; ld.local.u32 %r500, [%rd4+248]; mul.wide.s32 %rd596, %r500, 80; add.s64 %rd597, %rd1, %rd596; add.s64 %rd772, %rd597, 304; add.s64 %rd598, %rd230, %rd596; add.s64 %rd773, %rd598, 304; mov.b32 %r501, %f495; mov.b32 %r502, %f494; mov.b32 %r503, %f493; mov.u32 %r504, 0; st.local.v4.u32 [%rd597+304], {%r503, %r502, %r501, %r504}; add.s32 %r505, %r500, 1; st.local.u32 [%rd4+248], %r505; add.s32 %r506, %r125, -1; st.local.u32 [%rd4+252], %r506; $L__BB157_72: setp.eq.s64 %p391, %rd773, 0; mov.u64 %rd775, %rd774; @%p391 bra $L__BB157_74; st.local.f32 [%rd772+16], %f496; mov.u64 %rd774, %rd772; mov.u64 %rd775, %rd773; $L__BB157_74: setp.eq.s64 %p392, %rd775, 0; @%p392 bra $L__BB157_517; ld.local.u32 %r126, [%rd4+252]; setp.lt.s32 %p393, %r126, 1; @%p393 bra $L__BB157_457; add.s32 %r507, %r126, -1; st.local.u32 [%rd4+252], %r507; ld.local.s32 %rd601, [%rd4+248]; cvt.s64.s32 %rd602, %r507; add.s64 %rd603, %rd601, %rd602; mul.lo.s64 %rd604, %rd603, 80; add.s64 %rd605, %rd1, %rd604; add.s64 %rd776, %rd605, 304; add.s64 %rd606, %rd230, %rd604; add.s64 %rd777, %rd606, 304; bra.uni $L__BB157_458; $L__BB157_48: setp.gt.s16 %p37, %rs7, 19; @%p37 bra $L__BB157_61; bra.uni $L__BB157_49; $L__BB157_61: add.s16 %rs11, %rs7, -21; setp.lt.u16 %p38, %rs11, 3; @%p38 bra $L__BB157_462; bra.uni $L__BB157_62; $L__BB157_462: ld.const.u32 %r518, [kernel_params+1568]; setp.eq.s32 %p399, %r518, 0; shr.u32 %r519, %r141, 3; and.b32 %r520, %r519, 1; setp.eq.b32 %p400, %r520, 1; and.pred %p401, %p400, %p399; @%p401 bra $L__BB157_517; ld.local.f32 %f2756, [%rd4+256]; mov.u64 %rd780, 0; mul.ftz.f32 %f2757, %f2, %f2756; ld.local.f32 %f2758, [%rd4+260]; mul.ftz.f32 %f2759, %f2, %f2758; ld.local.f32 %f2760, [%rd4+264]; mul.ftz.f32 %f2761, %f2, %f2760; mov.f32 %f2762, 0f00000000; max.ftz.f32 %f500, %f2757, %f2762; max.ftz.f32 %f501, %f2759, %f2762; max.ftz.f32 %f502, %f2761, %f2762; add.ftz.f32 %f2763, %f500, %f501; add.ftz.f32 %f2764, %f502, %f2763; mul.ftz.f32 %f2765, %f2764, 0f3EAAAAAB; abs.ftz.f32 %f503, %f2765; setp.ltu.ftz.f32 %p402, %f503, 0f3727C5AC; mov.u64 %rd781, %rd780; @%p402 bra $L__BB157_468; ld.local.u32 %r127, [%rd4+252]; mov.u64 %rd780, 0; setp.eq.s32 %p403, %r127, 0; mov.u64 %rd778, %rd780; mov.u64 %rd779, %rd780; @%p403 bra $L__BB157_466; ld.local.u32 %r521, [%rd4+248]; mul.wide.s32 %rd616, %r521, 80; add.s64 %rd617, %rd1, %rd616; add.s64 %rd778, %rd617, 304; add.s64 %rd618, %rd230, %rd616; add.s64 %rd779, %rd618, 304; mov.b32 %r522, %f502; mov.b32 %r523, %f501; mov.b32 %r524, %f500; mov.u32 %r525, 0; st.local.v4.u32 [%rd617+304], {%r524, %r523, %r522, %r525}; add.s32 %r526, %r521, 1; st.local.u32 [%rd4+248], %r526; add.s32 %r527, %r127, -1; st.local.u32 [%rd4+252], %r527; $L__BB157_466: setp.eq.s64 %p404, %rd779, 0; mov.u64 %rd781, %rd780; @%p404 bra $L__BB157_468; st.local.f32 [%rd778+16], %f503; mov.u64 %rd780, %rd778; mov.u64 %rd781, %rd779; $L__BB157_468: setp.eq.s64 %p405, %rd781, 0; @%p405 bra $L__BB157_517; add.s64 %rd185, %rd780, 20; mov.u64 %rd621, 0; st.local.f32 [%rd780+20], %f55; st.local.f32 [%rd780+24], %f56; st.local.f32 [%rd780+28], %f57; mov.u32 %r528, 0; st.local.v2.f32 [%rd780+56], {%f2762, %f2762}; st.local.u32 [%rd780+64], %r528; st.local.u64 [%rd780+48], %rd621; mov.f32 %f2767, 0f3727C5AC; max.ftz.f32 %f3057, %f63, %f2767; ld.local.u32 %r529, [%rd4+40]; and.b32 %r530, %r529, 1; setp.eq.b32 %p406, %r530, 1; mov.pred %p407, 0; xor.pred %p408, %p406, %p407; not.pred %p409, %p408; @%p409 bra $L__BB157_471; rcp.approx.ftz.f32 %f3057, %f3057; $L__BB157_471: setp.eq.s32 %p410, %r2, 21; @%p410 bra $L__BB157_475; bra.uni $L__BB157_472; $L__BB157_475: st.local.v2.f32 [%rd185+12], {%f2762, %f2762}; st.local.f32 [%rd185+20], %f3057; mov.u32 %r537, 21; st.local.u32 [%rd185+-8], %r537; ld.local.u32 %r538, [%rd4+40]; or.b32 %r539, %r538, 4; st.local.u32 [%rd4+40], %r539; bra.uni $L__BB157_517; $L__BB157_88: add.s16 %rs9, %rs7, -32; setp.lt.u16 %p27, %rs9, 3; @%p27 bra $L__BB157_290; bra.uni $L__BB157_89; $L__BB157_290: ld.local.v2.u32 {%r330, %r331}, [%rd4+252]; mov.u64 %rd746, 0; mov.b32 %f1892, %r331; mul.ftz.f32 %f340, %f2, %f1892; ld.local.f32 %f1893, [%rd4+260]; mul.ftz.f32 %f341, %f2, %f1893; ld.local.f32 %f1894, [%rd4+264]; mul.ftz.f32 %f342, %f2, %f1894; setp.eq.s32 %p240, %r330, 0; mov.u64 %rd744, %rd746; mov.u64 %rd745, %rd746; @%p240 bra $L__BB157_292; ld.local.u32 %r333, [%rd4+248]; mul.wide.s32 %rd459, %r333, 80; add.s64 %rd460, %rd1, %rd459; add.s64 %rd744, %rd460, 304; add.s64 %rd461, %rd230, %rd459; add.s64 %rd745, %rd461, 304; mov.b32 %r334, %f342; mov.b32 %r335, %f341; mov.b32 %r336, %f340; mov.u32 %r337, 0; st.local.v4.u32 [%rd460+304], {%r336, %r335, %r334, %r337}; add.s32 %r338, %r333, 1; st.local.u32 [%rd4+248], %r338; add.s32 %r339, %r330, -1; st.local.u32 [%rd4+252], %r339; $L__BB157_292: setp.eq.s64 %p241, %rd745, 0; @%p241 bra $L__BB157_294; add.ftz.f32 %f1895, %f340, %f341; add.ftz.f32 %f1896, %f342, %f1895; mul.ftz.f32 %f1897, %f1896, 0f3EAAAAAB; abs.ftz.f32 %f1898, %f1897; st.local.f32 [%rd744+16], %f1898; setp.ge.ftz.f32 %p242, %f1898, 0f3727C5AC; selp.b64 %rd746, %rd745, 0, %p242; $L__BB157_294: setp.eq.s64 %p243, %rd746, 0; @%p243 bra $L__BB157_517; and.b32 %r341, %r141, 4096; setp.eq.s32 %p244, %r341, 0; selp.f32 %f1900, %f60, 0f00000000, %p244; mov.f32 %f3029, 0f00000000; ld.local.f32 %f1901, [%rd6]; mul.ftz.f32 %f1902, %f1900, %f1901; ld.local.f32 %f1903, [%rd6+4]; mul.ftz.f32 %f343, %f1900, %f1903; ld.local.f32 %f1904, [%rd6+8]; mul.ftz.f32 %f344, %f1900, %f1904; add.s64 %rd111, %rd746, 32; st.f32 [%rd746+32], %f1902; st.f32 [%rd746+36], %f343; st.f32 [%rd746+40], %f344; ld.local.f32 %f1905, [%rd4+256]; ld.local.f32 %f1906, [%rd4+260]; ld.local.f32 %f1907, [%rd4+264]; st.f32 [%rd746+44], %f1905; st.f32 [%rd746+48], %f1906; st.f32 [%rd746+52], %f1907; st.f32 [%rd746+20], %f55; st.f32 [%rd746+24], %f56; st.f32 [%rd746+28], %f57; mov.u32 %r342, 2139095039; st.u32 [%rd746+56], %r342; ld.local.f32 %f1908, [%rd7]; max.ftz.f32 %f1909, %f1908, %f3029; mov.f32 %f1910, 0f3F666666; min.ftz.f32 %f1911, %f1909, %f1910; st.f32 [%rd746+60], %f1911; setp.geu.ftz.f32 %p245, %f1902, 0f322BCC77; mov.u32 %r646, 3; mov.f32 %f3027, %f3029; @%p245 bra $L__BB157_297; ld.f32 %f3027, [%rd111+-32]; mov.u32 %r344, 0; st.u32 [%rd111+-32], %r344; st.u32 [%rd111], %r344; mov.u32 %r646, 2; $L__BB157_297: setp.geu.ftz.f32 %p246, %f343, 0f322BCC77; mov.f32 %f3028, %f3029; @%p246 bra $L__BB157_299; ld.f32 %f3028, [%rd111+-28]; mov.u32 %r345, 0; st.u32 [%rd111+-28], %r345; st.u32 [%rd111+4], %r345; add.s32 %r646, %r646, -1; $L__BB157_299: setp.geu.ftz.f32 %p247, %f344, 0f322BCC77; @%p247 bra $L__BB157_301; ld.f32 %f3029, [%rd111+-24]; mov.u32 %r346, 0; st.u32 [%rd111+-24], %r346; st.u32 [%rd111+8], %r346; add.s32 %r646, %r646, -1; $L__BB157_301: setp.gt.s32 %p248, %r646, 2; mov.u32 %r648, 0; @%p248 bra $L__BB157_309; mov.f32 %f1914, 0f00000000; max.ftz.f32 %f351, %f3027, %f1914; max.ftz.f32 %f352, %f3028, %f1914; add.ftz.f32 %f1915, %f351, %f352; max.ftz.f32 %f353, %f3029, %f1914; add.ftz.f32 %f1916, %f353, %f1915; mul.ftz.f32 %f1917, %f1916, 0f3EAAAAAB; abs.ftz.f32 %f354, %f1917; setp.ltu.ftz.f32 %p249, %f354, 0f3727C5AC; mov.u64 %rd749, 0; mov.u64 %rd750, %rd749; @%p249 bra $L__BB157_307; ld.local.u32 %r63, [%rd4+252]; mov.u64 %rd749, 0; setp.eq.s32 %p250, %r63, 0; mov.u64 %rd747, %rd749; mov.u64 %rd748, %rd749; @%p250 bra $L__BB157_305; ld.local.u32 %r348, [%rd4+248]; mul.wide.s32 %rd467, %r348, 80; add.s64 %rd468, %rd1, %rd467; add.s64 %rd747, %rd468, 304; add.s64 %rd469, %rd230, %rd467; add.s64 %rd748, %rd469, 304; mov.b32 %r349, %f353; mov.b32 %r350, %f352; mov.b32 %r351, %f351; mov.u32 %r352, 0; st.local.v4.u32 [%rd468+304], {%r351, %r350, %r349, %r352}; add.s32 %r353, %r348, 1; st.local.u32 [%rd4+248], %r353; add.s32 %r354, %r63, -1; st.local.u32 [%rd4+252], %r354; $L__BB157_305: setp.eq.s64 %p251, %rd748, 0; mov.u64 %rd750, %rd749; @%p251 bra $L__BB157_307; st.local.f32 [%rd747+16], %f354; mov.u64 %rd749, %rd747; mov.u64 %rd750, %rd748; $L__BB157_307: setp.eq.s64 %p252, %rd750, 0; @%p252 bra $L__BB157_309; ld.f32 %f1918, [%rd111+-12]; ld.f32 %f1919, [%rd111+-8]; ld.f32 %f1920, [%rd111+-4]; st.local.f32 [%rd749+20], %f1918; st.local.f32 [%rd749+24], %f1919; st.local.f32 [%rd749+28], %f1920; mov.u32 %r357, 2; st.local.u32 [%rd749+12], %r357; mov.u32 %r648, 12; $L__BB157_309: st.u32 [%rd111+-20], %r2; setp.gt.s32 %p253, %r646, 0; @%p253 bra $L__BB157_311; bra.uni $L__BB157_310; $L__BB157_311: ld.f32 %f1921, [%rd111+-28]; ld.f32 %f1922, [%rd111+-32]; add.ftz.f32 %f1923, %f1922, %f1921; ld.f32 %f1924, [%rd111+-24]; add.ftz.f32 %f1925, %f1924, %f1923; mul.ftz.f32 %f1926, %f1925, 0f3EAAAAAB; abs.ftz.f32 %f1927, %f1926; cvt.rn.f32.s32 %f1928, %r646; mul.ftz.f32 %f1929, %f1927, %f1928; st.f32 [%rd111+-16], %f1929; or.b32 %r359, %r2, 2; setp.eq.s32 %p254, %r359, 34; @%p254 bra $L__BB157_322; bra.uni $L__BB157_312; $L__BB157_322: ld.f32 %f2483, [%rd111]; mul.ftz.f32 %f3033, %f2483, 0f3DA2F983; ld.f32 %f2484, [%rd111+4]; mul.ftz.f32 %f3034, %f2484, 0f3DA2F983; ld.f32 %f2485, [%rd111+8]; mul.ftz.f32 %f3035, %f2485, 0f3DA2F983; bra.uni $L__BB157_323; $L__BB157_24: setp.eq.s16 %p46, %rs7, 2; @%p46 bra $L__BB157_507; setp.eq.s16 %p47, %rs7, 7; @%p47 bra $L__BB157_53; setp.eq.s16 %p48, %rs7, 8; @%p48 bra $L__BB157_27; bra.uni $L__BB157_517; $L__BB157_27: ld.local.f32 %f2922, [%rd4+256]; mov.u64 %rd798, 0; mul.ftz.f32 %f2923, %f2, %f2922; ld.local.f32 %f2924, [%rd4+260]; mul.ftz.f32 %f2925, %f2, %f2924; ld.local.f32 %f2926, [%rd4+264]; mul.ftz.f32 %f2927, %f2, %f2926; mov.f32 %f2928, 0f00000000; max.ftz.f32 %f535, %f2923, %f2928; max.ftz.f32 %f536, %f2925, %f2928; max.ftz.f32 %f537, %f2927, %f2928; add.ftz.f32 %f2929, %f535, %f536; add.ftz.f32 %f2930, %f537, %f2929; mul.ftz.f32 %f2931, %f2930, 0f3EAAAAAB; abs.ftz.f32 %f538, %f2931; setp.ltu.ftz.f32 %p447, %f538, 0f3727C5AC; mov.u64 %rd799, %rd798; @%p447 bra $L__BB157_32; ld.local.u32 %r132, [%rd4+252]; mov.u64 %rd798, 0; setp.eq.s32 %p448, %r132, 0; mov.u64 %rd796, %rd798; mov.u64 %rd797, %rd798; @%p448 bra $L__BB157_30; ld.local.u32 %r600, [%rd4+248]; mul.wide.s32 %rd668, %r600, 80; add.s64 %rd669, %rd1, %rd668; add.s64 %rd796, %rd669, 304; add.s64 %rd670, %rd230, %rd668; add.s64 %rd797, %rd670, 304; mov.b32 %r601, %f537; mov.b32 %r602, %f536; mov.b32 %r603, %f535; mov.u32 %r604, 0; st.local.v4.u32 [%rd669+304], {%r603, %r602, %r601, %r604}; add.s32 %r605, %r600, 1; st.local.u32 [%rd4+248], %r605; add.s32 %r606, %r132, -1; st.local.u32 [%rd4+252], %r606; $L__BB157_30: setp.eq.s64 %p449, %rd797, 0; mov.u64 %rd799, %rd798; @%p449 bra $L__BB157_32; st.local.f32 [%rd796+16], %f538; mov.u64 %rd798, %rd796; mov.u64 %rd799, %rd797; $L__BB157_32: setp.eq.s64 %p450, %rd799, 0; @%p450 bra $L__BB157_517; st.local.f32 [%rd798+20], %f55; st.local.f32 [%rd798+24], %f56; st.local.f32 [%rd798+28], %f57; mov.u32 %r607, 8; st.local.u32 [%rd798+12], %r607; ld.local.u32 %r608, [%rd4+40]; or.b32 %r609, %r608, 12; st.local.u32 [%rd4+40], %r609; bra.uni $L__BB157_517; $L__BB157_77: setp.eq.s16 %p31, %rs7, 28; @%p31 bra $L__BB157_427; setp.eq.s16 %p32, %rs7, 29; @%p32 bra $L__BB157_341; setp.eq.s16 %p33, %rs7, 30; @%p33 bra $L__BB157_80; bra.uni $L__BB157_517; $L__BB157_49: setp.eq.s16 %p40, %rs7, 16; @%p40 bra $L__BB157_36; setp.eq.s16 %p41, %rs7, 17; @%p41 bra $L__BB157_482; setp.eq.s16 %p42, %rs7, 19; @%p42 bra $L__BB157_52; bra.uni $L__BB157_517; $L__BB157_52: ld.const.u32 %r554, [kernel_params+1564]; setp.eq.s32 %p419, %r554, 0; shr.u32 %r555, %r141, 3; and.b32 %r556, %r555, 1; setp.eq.b32 %p420, %r556, 1; and.pred %p421, %p420, %p419; @%p421 bra $L__BB157_517; $L__BB157_53: ld.local.f32 %f2932, [%rd4+256]; mov.u64 %rd802, 0; mul.ftz.f32 %f2933, %f2, %f2932; ld.local.f32 %f2934, [%rd4+260]; mul.ftz.f32 %f2935, %f2, %f2934; ld.local.f32 %f2936, [%rd4+264]; mul.ftz.f32 %f2937, %f2, %f2936; mov.f32 %f2938, 0f00000000; max.ftz.f32 %f539, %f2933, %f2938; max.ftz.f32 %f540, %f2935, %f2938; max.ftz.f32 %f541, %f2937, %f2938; add.ftz.f32 %f2939, %f539, %f540; add.ftz.f32 %f2940, %f541, %f2939; mul.ftz.f32 %f2941, %f2940, 0f3EAAAAAB; abs.ftz.f32 %f542, %f2941; setp.ltu.ftz.f32 %p451, %f542, 0f3727C5AC; mov.u64 %rd803, %rd802; @%p451 bra $L__BB157_58; ld.local.u32 %r133, [%rd4+252]; mov.u64 %rd802, 0; setp.eq.s32 %p452, %r133, 0; mov.u64 %rd800, %rd802; mov.u64 %rd801, %rd802; @%p452 bra $L__BB157_56; ld.local.u32 %r610, [%rd4+248]; mul.wide.s32 %rd677, %r610, 80; add.s64 %rd678, %rd1, %rd677; add.s64 %rd800, %rd678, 304; add.s64 %rd679, %rd230, %rd677; add.s64 %rd801, %rd679, 304; mov.b32 %r611, %f541; mov.b32 %r612, %f540; mov.b32 %r613, %f539; mov.u32 %r614, 0; st.local.v4.u32 [%rd678+304], {%r613, %r612, %r611, %r614}; add.s32 %r615, %r610, 1; st.local.u32 [%rd4+248], %r615; add.s32 %r616, %r133, -1; st.local.u32 [%rd4+252], %r616; $L__BB157_56: setp.eq.s64 %p453, %rd801, 0; mov.u64 %rd803, %rd802; @%p453 bra $L__BB157_58; st.local.f32 [%rd800+16], %f542; mov.u64 %rd802, %rd800; mov.u64 %rd803, %rd801; $L__BB157_58: setp.eq.s64 %p454, %rd803, 0; @%p454 bra $L__BB157_517; add.s64 %rd222, %rd802, 20; st.local.f32 [%rd802+20], %f55; st.local.f32 [%rd802+24], %f56; st.local.f32 [%rd802+28], %f57; st.local.f32 [%rd802+32], %f60; st.local.f32 [%rd802+36], %f63; cvt.ftz.sat.f32.f32 %f543, %f60; cvt.ftz.sat.f32.f32 %f544, %f63; setp.eq.s32 %p455, %r2, 7; @%p455 bra $L__BB157_506; bra.uni $L__BB157_60; $L__BB157_506: mov.u32 %r620, 7; st.local.u32 [%rd222+-8], %r620; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r621, [%rd4+40]; or.b32 %r622, %r621, 12; st.local.u32 [%rd4+40], %r622; bra.uni $L__BB157_517; $L__BB157_36: ld.const.u32 %r567, [kernel_params+1564]; setp.eq.s32 %p426, %r567, 0; shr.u32 %r568, %r141, 3; and.b32 %r569, %r568, 1; setp.eq.b32 %p427, %r569, 1; and.pred %p428, %p427, %p426; @%p428 bra $L__BB157_517; ld.local.f32 %f2831, [%rd4+256]; mov.u64 %rd792, 0; mul.ftz.f32 %f2832, %f2, %f2831; ld.local.f32 %f2833, [%rd4+260]; mul.ftz.f32 %f2834, %f2, %f2833; ld.local.f32 %f2835, [%rd4+264]; mul.ftz.f32 %f2836, %f2, %f2835; mov.f32 %f2837, 0f00000000; max.ftz.f32 %f519, %f2832, %f2837; max.ftz.f32 %f520, %f2834, %f2837; max.ftz.f32 %f521, %f2836, %f2837; add.ftz.f32 %f2838, %f519, %f520; add.ftz.f32 %f2839, %f521, %f2838; mul.ftz.f32 %f2840, %f2839, 0f3EAAAAAB; abs.ftz.f32 %f522, %f2840; setp.ltu.ftz.f32 %p429, %f522, 0f3727C5AC; mov.u64 %rd793, %rd792; @%p429 bra $L__BB157_42; ld.local.u32 %r130, [%rd4+252]; mov.u64 %rd792, 0; setp.eq.s32 %p430, %r130, 0; mov.u64 %rd790, %rd792; mov.u64 %rd791, %rd792; @%p430 bra $L__BB157_40; ld.local.u32 %r570, [%rd4+248]; mul.wide.s32 %rd645, %r570, 80; add.s64 %rd646, %rd1, %rd645; add.s64 %rd790, %rd646, 304; add.s64 %rd647, %rd230, %rd645; add.s64 %rd791, %rd647, 304; mov.b32 %r571, %f521; mov.b32 %r572, %f520; mov.b32 %r573, %f519; mov.u32 %r574, 0; st.local.v4.u32 [%rd646+304], {%r573, %r572, %r571, %r574}; add.s32 %r575, %r570, 1; st.local.u32 [%rd4+248], %r575; add.s32 %r576, %r130, -1; st.local.u32 [%rd4+252], %r576; $L__BB157_40: setp.eq.s64 %p431, %rd791, 0; mov.u64 %rd793, %rd792; @%p431 bra $L__BB157_42; st.local.f32 [%rd790+16], %f522; mov.u64 %rd792, %rd790; mov.u64 %rd793, %rd791; $L__BB157_42: setp.eq.s64 %p432, %rd793, 0; @%p432 bra $L__BB157_517; mul.ftz.f32 %f3058, %f60, %f60; add.s64 %rd205, %rd792, 20; mov.u64 %rd794, 0; st.local.f32 [%rd792+20], %f55; st.local.f32 [%rd792+24], %f56; st.local.f32 [%rd792+28], %f57; mov.u32 %r577, 1065353216; st.local.u32 [%rd792+40], %r577; st.local.u64 [%rd792+48], %rd794; setp.eq.s32 %p433, %r144, 255; @%p433 bra $L__BB157_490; bra.uni $L__BB157_44; $L__BB157_490: mov.u32 %r578, 0; st.local.v2.f32 [%rd205+36], {%f2837, %f2837}; st.local.u32 [%rd205+44], %r578; st.local.v2.f32 [%rd205+12], {%f3058, %f3058}; mov.f32 %f3059, %f3058; bra.uni $L__BB157_491; $L__BB157_427: ld.const.u32 %r121, [kernel_params+1564]; setp.ne.s32 %p358, %r121, 0; @%p358 bra $L__BB157_429; ld.const.u32 %r454, [kernel_params+1568]; setp.eq.s32 %p359, %r454, 0; shr.u32 %r455, %r141, 3; and.b32 %r456, %r455, 1; setp.eq.b32 %p360, %r456, 1; and.pred %p361, %p360, %p359; @%p361 bra $L__BB157_517; $L__BB157_429: ld.local.f32 %f2682, [%rd4+256]; mul.ftz.f32 %f470, %f2, %f2682; ld.local.f32 %f2683, [%rd4+260]; mul.ftz.f32 %f471, %f2, %f2683; ld.local.f32 %f2684, [%rd4+264]; mul.ftz.f32 %f472, %f2, %f2684; mov.f32 %f2685, 0f3727C5AC; max.ftz.f32 %f3054, %f63, %f2685; ld.local.u32 %r457, [%rd4+40]; and.b32 %r458, %r457, 1; setp.eq.b32 %p362, %r458, 1; mov.pred %p363, 0; xor.pred %p364, %p362, %p363; not.pred %p365, %p364; @%p365 bra $L__BB157_431; rcp.approx.ftz.f32 %f3054, %f3054; $L__BB157_431: ld.local.f32 %f2687, [%rd4+24]; ld.local.f32 %f2688, [%rd4+28]; mul.ftz.f32 %f2689, %f56, %f2688; fma.rn.ftz.f32 %f2690, %f55, %f2687, %f2689; ld.local.f32 %f2691, [%rd4+32]; fma.rn.ftz.f32 %f2692, %f57, %f2691, %f2690; abs.ftz.f32 %f476, %f2692; fma.rn.ftz.f32 %f2693, %f3054, %f3054, 0fBF800000; mov.f32 %f3055, 0f3F800000; fma.rn.ftz.f32 %f477, %f476, %f476, %f2693; setp.leu.ftz.f32 %p366, %f477, 0f00000000; @%p366 bra $L__BB157_433; sqrt.approx.ftz.f32 %f2694, %f477; sub.ftz.f32 %f2695, %f2694, %f476; add.ftz.f32 %f2696, %f476, %f2694; div.approx.ftz.f32 %f2697, %f2695, %f2696; fma.rn.ftz.f32 %f2698, %f476, %f2696, 0fBF800000; fma.rn.ftz.f32 %f2699, %f476, %f2695, 0f3F800000; div.approx.ftz.f32 %f2700, %f2698, %f2699; mul.ftz.f32 %f2701, %f2697, 0f3F000000; mul.ftz.f32 %f2702, %f2697, %f2701; fma.rn.ftz.f32 %f2703, %f2700, %f2700, 0f3F800000; mul.ftz.f32 %f3055, %f2702, %f2703; $L__BB157_433: mul.ftz.f32 %f480, %f60, %f60; and.b32 %r122, %r141, 8; shr.u32 %r459, %r141, 3; and.b32 %r460, %r459, 1; setp.eq.b32 %p367, %r460, 1; setp.eq.s32 %p368, %r121, 0; and.pred %p369, %p367, %p368; @%p369 bra $L__BB157_445; mul.ftz.f32 %f2704, %f470, %f3055; mov.f32 %f2705, 0f00000000; max.ftz.f32 %f481, %f2704, %f2705; mul.ftz.f32 %f2706, %f471, %f3055; max.ftz.f32 %f482, %f2706, %f2705; mul.ftz.f32 %f2707, %f472, %f3055; max.ftz.f32 %f483, %f2707, %f2705; add.ftz.f32 %f2708, %f481, %f482; add.ftz.f32 %f2709, %f483, %f2708; mul.ftz.f32 %f2710, %f2709, 0f3EAAAAAB; abs.ftz.f32 %f484, %f2710; setp.ltu.ftz.f32 %p370, %f484, 0f3727C5AC; mov.u64 %rd766, 0; mov.u64 %rd767, %rd766; @%p370 bra $L__BB157_439; ld.local.u32 %r123, [%rd4+252]; mov.u64 %rd766, 0; setp.eq.s32 %p371, %r123, 0; mov.u64 %rd764, %rd766; mov.u64 %rd765, %rd766; @%p371 bra $L__BB157_437; ld.local.u32 %r461, [%rd4+248]; mul.wide.s32 %rd574, %r461, 80; add.s64 %rd575, %rd1, %rd574; add.s64 %rd764, %rd575, 304; add.s64 %rd576, %rd230, %rd574; add.s64 %rd765, %rd576, 304; mov.b32 %r462, %f483; mov.b32 %r463, %f482; mov.b32 %r464, %f481; mov.u32 %r465, 0; st.local.v4.u32 [%rd575+304], {%r464, %r463, %r462, %r465}; add.s32 %r466, %r461, 1; st.local.u32 [%rd4+248], %r466; add.s32 %r467, %r123, -1; st.local.u32 [%rd4+252], %r467; $L__BB157_437: setp.eq.s64 %p372, %rd765, 0; mov.u64 %rd767, %rd766; @%p372 bra $L__BB157_439; st.local.f32 [%rd764+16], %f484; mov.u64 %rd766, %rd764; mov.u64 %rd767, %rd765; $L__BB157_439: setp.eq.s64 %p373, %rd767, 0; @%p373 bra $L__BB157_445; add.s64 %rd160, %rd766, 20; mov.u64 %rd579, 0; st.local.f32 [%rd766+20], %f55; st.local.f32 [%rd766+24], %f56; st.local.f32 [%rd766+28], %f57; mov.u32 %r468, 0; st.local.v2.f32 [%rd766+56], {%f2705, %f2705}; st.local.u32 [%rd766+64], %r468; st.local.u64 [%rd766+48], %rd579; setp.eq.s32 %p374, %r2, 28; @%p374 bra $L__BB157_444; bra.uni $L__BB157_441; $L__BB157_444: st.local.v2.f32 [%rd160+12], {%f2705, %f2705}; st.local.f32 [%rd160+20], %f3054; mov.u32 %r475, 9; st.local.u32 [%rd160+-8], %r475; ld.local.u32 %r476, [%rd4+40]; or.b32 %r477, %r476, 4; st.local.u32 [%rd4+40], %r477; bra.uni $L__BB157_445; $L__BB157_89: setp.eq.s16 %p28, %rs7, 31; @%p28 bra $L__BB157_325; setp.ne.s16 %p29, %rs7, 39; @%p29 bra $L__BB157_517; ld.global.v4.u32 {%r147, %r148, %r149, %r150}, [%rd3+16]; ld.local.f32 %f2977, [%rd5]; ld.local.f32 %f2978, [%rd5+4]; ld.local.f32 %f2979, [%rd5+8]; shr.u32 %r152, %r147, 24; shl.b32 %r153, %r145, 2; cvt.u64.u32 %rd246, %r153; and.b64 %rd247, %rd246, 1020; add.s64 %rd248, %rd2, %rd247; ld.local.f32 %f70, [%rd248]; shr.u32 %r154, %r145, 6; cvt.u64.u32 %rd249, %r154; and.b64 %rd250, %rd249, 1020; add.s64 %rd251, %rd2, %rd250; ld.local.f32 %f71, [%rd251]; shr.u32 %r155, %r145, 14; cvt.u64.u32 %rd252, %r155; and.b64 %rd253, %rd252, 1020; add.s64 %rd254, %rd2, %rd253; ld.local.f32 %f72, [%rd254]; shr.u32 %r156, %r145, 24; mul.wide.u32 %rd255, %r156, 4; add.s64 %rd256, %rd2, %rd255; ld.local.f32 %f73, [%rd256]; shl.b32 %r157, %r146, 2; cvt.u64.u32 %rd257, %r157; and.b64 %rd258, %rd257, 1020; add.s64 %rd259, %rd2, %rd258; ld.local.f32 %f74, [%rd259]; shr.u32 %r158, %r146, 6; cvt.u64.u32 %rd260, %r158; and.b64 %rd261, %rd260, 1020; add.s64 %rd262, %rd2, %rd261; ld.local.f32 %f75, [%rd262]; shr.u32 %r159, %r146, 14; cvt.u64.u32 %rd263, %r159; and.b64 %rd264, %rd263, 1020; add.s64 %rd265, %rd2, %rd264; ld.local.f32 %f76, [%rd265]; shr.u32 %r160, %r146, 24; mul.wide.u32 %rd266, %r160, 4; add.s64 %rd267, %rd2, %rd266; ld.local.f32 %f77, [%rd267]; shr.u32 %r161, %r147, 6; cvt.u64.u32 %rd268, %r161; and.b64 %rd269, %rd268, 1020; add.s64 %rd270, %rd2, %rd269; ld.local.f32 %f78, [%rd270]; shr.u32 %r162, %r147, 14; cvt.u64.u32 %rd271, %r162; and.b64 %rd272, %rd271, 1020; add.s64 %rd273, %rd2, %rd272; mul.wide.u32 %rd274, %r152, 4; add.s64 %rd275, %rd2, %rd274; ld.local.f32 %f79, [%rd275]; shl.b32 %r163, %r147, 2; cvt.u64.u32 %rd276, %r163; and.b64 %rd277, %rd276, 1020; add.s64 %rd278, %rd2, %rd277; ld.local.f32 %f626, [%rd278]; mov.f32 %f627, 0f3727C5AC; max.ftz.f32 %f2980, %f626, %f627; ld.local.f32 %f81, [%rd273]; setp.eq.ftz.f32 %p49, %f81, 0f00000000; @%p49 bra $L__BB157_93; mul.ftz.f32 %f628, %f81, 0f40C90FDB; cos.approx.ftz.f32 %f629, %f628; sin.approx.ftz.f32 %f630, %f628; mov.f32 %f631, 0f3F800000; sub.ftz.f32 %f632, %f631, %f629; mul.ftz.f32 %f633, %f55, %f632; fma.rn.ftz.f32 %f634, %f55, %f633, %f629; mul.ftz.f32 %f635, %f56, %f633; mul.ftz.f32 %f636, %f57, %f630; sub.ftz.f32 %f637, %f635, %f636; mul.ftz.f32 %f638, %f2978, %f637; fma.rn.ftz.f32 %f639, %f2977, %f634, %f638; mul.ftz.f32 %f640, %f57, %f633; mul.ftz.f32 %f641, %f56, %f630; add.ftz.f32 %f642, %f641, %f640; add.ftz.f32 %f643, %f636, %f635; mul.ftz.f32 %f644, %f56, %f632; fma.rn.ftz.f32 %f645, %f56, %f644, %f629; mul.ftz.f32 %f646, %f2978, %f645; fma.rn.ftz.f32 %f647, %f2977, %f643, %f646; mul.ftz.f32 %f648, %f57, %f644; mul.ftz.f32 %f649, %f55, %f630; sub.ftz.f32 %f650, %f648, %f649; sub.ftz.f32 %f651, %f640, %f641; add.ftz.f32 %f652, %f649, %f648; mul.ftz.f32 %f653, %f2978, %f652; fma.rn.ftz.f32 %f654, %f2977, %f651, %f653; mul.ftz.f32 %f655, %f57, %f632; fma.rn.ftz.f32 %f656, %f57, %f655, %f629; fma.rn.ftz.f32 %f2978, %f2979, %f650, %f647; fma.rn.ftz.f32 %f84, %f2979, %f656, %f654; fma.rn.ftz.f32 %f2977, %f2979, %f642, %f639; mov.f32 %f2979, %f84; $L__BB157_93: ld.local.u32 %r164, [%rd4+40]; and.b32 %r165, %r164, 1; setp.eq.b32 %p50, %r165, 1; mov.pred %p51, 0; xor.pred %p52, %p50, %p51; not.pred %p53, %p52; @%p53 bra $L__BB157_95; rcp.approx.ftz.f32 %f2980, %f2980; $L__BB157_95: ld.local.f32 %f90, [%rd4+24]; ld.local.f32 %f91, [%rd4+28]; mul.ftz.f32 %f658, %f56, %f91; fma.rn.ftz.f32 %f659, %f55, %f90, %f658; ld.local.f32 %f92, [%rd4+32]; fma.rn.ftz.f32 %f660, %f57, %f92, %f659; abs.ftz.f32 %f93, %f660; fma.rn.ftz.f32 %f661, %f2980, %f2980, 0fBF800000; mov.f32 %f657, 0f3F800000; fma.rn.ftz.f32 %f94, %f93, %f93, %f661; setp.leu.ftz.f32 %p54, %f94, 0f00000000; mov.f32 %f2981, %f657; @%p54 bra $L__BB157_97; sqrt.approx.ftz.f32 %f662, %f94; sub.ftz.f32 %f663, %f662, %f93; add.ftz.f32 %f664, %f93, %f662; div.approx.ftz.f32 %f665, %f663, %f664; fma.rn.ftz.f32 %f666, %f93, %f664, 0fBF800000; fma.rn.ftz.f32 %f667, %f93, %f663, 0f3F800000; div.approx.ftz.f32 %f668, %f666, %f667; mul.ftz.f32 %f669, %f665, 0f3F000000; mul.ftz.f32 %f670, %f665, %f669; fma.rn.ftz.f32 %f671, %f668, %f668, 0f3F800000; mul.ftz.f32 %f2981, %f670, %f671; $L__BB157_97: cvt.ftz.sat.f32.f32 %f672, %f60; sub.ftz.f32 %f674, %f657, %f672; cvt.ftz.sat.f32.f32 %f675, %f78; sub.ftz.f32 %f676, %f657, %f675; mul.ftz.f32 %f97, %f674, %f676; mul.ftz.f32 %f98, %f675, %f674; sub.ftz.f32 %f99, %f657, %f98; ld.global.v4.u32 {%r166, %r167, %r168, %r169}, [%rd3+32]; setp.eq.s32 %p55, %r166, 255; @%p55 bra $L__BB157_99; bra.uni $L__BB157_98; $L__BB157_99: mov.b32 %f2982, %r167; mov.b32 %f2983, %r168; mov.b32 %f2984, %r169; bra.uni $L__BB157_100; $L__BB157_62: setp.eq.s16 %p39, %rs7, 20; @%p39 bra $L__BB157_80; bra.uni $L__BB157_517; $L__BB157_80: ld.local.f32 %f2770, [%rd4+256]; mov.u64 %rd784, 0; mul.ftz.f32 %f2771, %f2, %f2770; ld.local.f32 %f2772, [%rd4+260]; mul.ftz.f32 %f2773, %f2, %f2772; ld.local.f32 %f2774, [%rd4+264]; mul.ftz.f32 %f2775, %f2, %f2774; mov.f32 %f2776, 0f00000000; max.ftz.f32 %f509, %f2771, %f2776; max.ftz.f32 %f510, %f2773, %f2776; max.ftz.f32 %f511, %f2775, %f2776; add.ftz.f32 %f2777, %f509, %f510; add.ftz.f32 %f2778, %f511, %f2777; mul.ftz.f32 %f2779, %f2778, 0f3EAAAAAB; abs.ftz.f32 %f512, %f2779; setp.ltu.ftz.f32 %p412, %f512, 0f3727C5AC; mov.u64 %rd785, %rd784; @%p412 bra $L__BB157_85; ld.local.u32 %r128, [%rd4+252]; mov.u64 %rd784, 0; setp.eq.s32 %p413, %r128, 0; mov.u64 %rd782, %rd784; mov.u64 %rd783, %rd784; @%p413 bra $L__BB157_83; ld.local.u32 %r540, [%rd4+248]; mul.wide.s32 %rd627, %r540, 80; add.s64 %rd628, %rd1, %rd627; add.s64 %rd782, %rd628, 304; add.s64 %rd629, %rd230, %rd627; add.s64 %rd783, %rd629, 304; mov.b32 %r541, %f511; mov.b32 %r542, %f510; mov.b32 %r543, %f509; mov.u32 %r544, 0; st.local.v4.u32 [%rd628+304], {%r543, %r542, %r541, %r544}; add.s32 %r545, %r540, 1; st.local.u32 [%rd4+248], %r545; add.s32 %r546, %r128, -1; st.local.u32 [%rd4+252], %r546; $L__BB157_83: setp.eq.s64 %p414, %rd783, 0; mov.u64 %rd785, %rd784; @%p414 bra $L__BB157_85; st.local.f32 [%rd782+16], %f512; mov.u64 %rd784, %rd782; mov.u64 %rd785, %rd783; $L__BB157_85: setp.eq.s64 %p415, %rd785, 0; @%p415 bra $L__BB157_517; add.s64 %rd192, %rd784, 20; st.local.f32 [%rd784+20], %f55; st.local.f32 [%rd784+24], %f56; st.local.f32 [%rd784+28], %f57; st.local.f32 [%rd784+44], %f60; st.local.f32 [%rd784+48], %f63; ld.local.f32 %f2780, [%rd6]; neg.ftz.f32 %f2781, %f2780; st.local.f32 [%rd784+52], %f2781; setp.eq.s32 %p416, %r144, 255; @%p416 bra $L__BB157_476; bra.uni $L__BB157_87; $L__BB157_476: ld.local.u8 %rs18, [%rd4+52]; and.b16 %rs19, %rs18, 6; setp.eq.s16 %p417, %rs19, 0; @%p417 bra $L__BB157_478; ld.local.f32 %f2792, [%rd4+104]; ld.local.f32 %f2793, [%rd4+108]; mul.ftz.f32 %f2794, %f2793, %f2793; fma.rn.ftz.f32 %f2795, %f2792, %f2792, %f2794; ld.local.f32 %f2796, [%rd4+112]; fma.rn.ftz.f32 %f2797, %f2796, %f2796, %f2795; sqrt.approx.ftz.f32 %f2798, %f2797; div.approx.ftz.f32 %f2799, %f2792, %f2798; div.approx.ftz.f32 %f2800, %f2793, %f2798; div.approx.ftz.f32 %f2801, %f2796, %f2798; st.local.f32 [%rd192+12], %f2799; st.local.f32 [%rd192+16], %f2800; st.local.f32 [%rd192+20], %f2801; bra.uni $L__BB157_479; $L__BB157_507: ld.local.f32 %f2942, [%rd4+256]; mov.u64 %rd806, 0; mul.ftz.f32 %f2943, %f2, %f2942; ld.local.f32 %f2944, [%rd4+260]; mul.ftz.f32 %f2945, %f2, %f2944; ld.local.f32 %f2946, [%rd4+264]; mul.ftz.f32 %f2947, %f2, %f2946; mov.f32 %f2948, 0f00000000; max.ftz.f32 %f545, %f2943, %f2948; max.ftz.f32 %f546, %f2945, %f2948; max.ftz.f32 %f547, %f2947, %f2948; add.ftz.f32 %f2949, %f545, %f546; add.ftz.f32 %f2950, %f547, %f2949; mul.ftz.f32 %f2951, %f2950, 0f3EAAAAAB; abs.ftz.f32 %f548, %f2951; setp.ltu.ftz.f32 %p456, %f548, 0f3727C5AC; mov.u64 %rd807, %rd806; @%p456 bra $L__BB157_512; ld.local.u32 %r134, [%rd4+252]; mov.u64 %rd806, 0; setp.eq.s32 %p457, %r134, 0; mov.u64 %rd804, %rd806; mov.u64 %rd805, %rd806; @%p457 bra $L__BB157_510; ld.local.u32 %r623, [%rd4+248]; mul.wide.s32 %rd686, %r623, 80; add.s64 %rd687, %rd1, %rd686; add.s64 %rd804, %rd687, 304; add.s64 %rd688, %rd230, %rd686; add.s64 %rd805, %rd688, 304; mov.b32 %r624, %f547; mov.b32 %r625, %f546; mov.b32 %r626, %f545; mov.u32 %r627, 0; st.local.v4.u32 [%rd687+304], {%r626, %r625, %r624, %r627}; add.s32 %r628, %r623, 1; st.local.u32 [%rd4+248], %r628; add.s32 %r629, %r134, -1; st.local.u32 [%rd4+252], %r629; $L__BB157_510: setp.eq.s64 %p458, %rd805, 0; mov.u64 %rd807, %rd806; @%p458 bra $L__BB157_512; st.local.f32 [%rd804+16], %f548; mov.u64 %rd806, %rd804; mov.u64 %rd807, %rd805; $L__BB157_512: setp.eq.s64 %p459, %rd807, 0; @%p459 bra $L__BB157_517; add.s64 %rd229, %rd806, 20; st.local.f32 [%rd806+20], %f55; st.local.f32 [%rd806+24], %f56; st.local.f32 [%rd806+28], %f57; setp.eq.ftz.f32 %p460, %f60, 0f00000000; @%p460 bra $L__BB157_515; bra.uni $L__BB157_514; $L__BB157_515: mov.u32 %r633, 2; st.local.u32 [%rd229+-8], %r633; ld.local.u32 %r634, [%rd4+40]; or.b32 %r635, %r634, 12; st.local.u32 [%rd4+40], %r635; bra.uni $L__BB157_517; $L__BB157_341: ld.global.v4.u32 {%r375, %r376, %r377, %r378}, [%rd3+16]; ld.global.v4.u32 {%r379, %r380, %r381, %r382}, [%rd3+32]; ld.global.u32 %rd121, [%rd3+52]; ld.local.f32 %f378, [%rd4+256]; ld.local.f32 %f379, [%rd4+260]; ld.local.f32 %f380, [%rd4+264]; bfe.u32 %r82, %r144, 8, 8; and.b32 %r83, %r144, 255; setp.eq.s32 %p304, %r83, 255; @%p304 bra $L__BB157_343; bra.uni $L__BB157_342; $L__BB157_343: mov.b32 %f3036, %r145; bra.uni $L__BB157_344; $L__BB157_482: ld.local.f32 %f2816, [%rd4+256]; mov.u64 %rd788, 0; mul.ftz.f32 %f2817, %f2, %f2816; ld.local.f32 %f2818, [%rd4+260]; mul.ftz.f32 %f2819, %f2, %f2818; ld.local.f32 %f2820, [%rd4+264]; mul.ftz.f32 %f2821, %f2, %f2820; mov.f32 %f2822, 0f00000000; max.ftz.f32 %f515, %f2817, %f2822; max.ftz.f32 %f516, %f2819, %f2822; max.ftz.f32 %f517, %f2821, %f2822; add.ftz.f32 %f2823, %f515, %f516; add.ftz.f32 %f2824, %f517, %f2823; mul.ftz.f32 %f2825, %f2824, 0f3EAAAAAB; abs.ftz.f32 %f518, %f2825; setp.ltu.ftz.f32 %p422, %f518, 0f3727C5AC; mov.u64 %rd789, %rd788; @%p422 bra $L__BB157_487; ld.local.u32 %r129, [%rd4+252]; mov.u64 %rd788, 0; setp.eq.s32 %p423, %r129, 0; mov.u64 %rd786, %rd788; mov.u64 %rd787, %rd788; @%p423 bra $L__BB157_485; ld.local.u32 %r557, [%rd4+248]; mul.wide.s32 %rd636, %r557, 80; add.s64 %rd637, %rd1, %rd636; add.s64 %rd786, %rd637, 304; add.s64 %rd638, %rd230, %rd636; add.s64 %rd787, %rd638, 304; mov.b32 %r558, %f517; mov.b32 %r559, %f516; mov.b32 %r560, %f515; mov.u32 %r561, 0; st.local.v4.u32 [%rd637+304], {%r560, %r559, %r558, %r561}; add.s32 %r562, %r557, 1; st.local.u32 [%rd4+248], %r562; add.s32 %r563, %r129, -1; st.local.u32 [%rd4+252], %r563; $L__BB157_485: setp.eq.s64 %p424, %rd787, 0; mov.u64 %rd789, %rd788; @%p424 bra $L__BB157_487; st.local.f32 [%rd786+16], %f518; mov.u64 %rd788, %rd786; mov.u64 %rd789, %rd787; $L__BB157_487: setp.eq.s64 %p425, %rd789, 0; @%p425 bra $L__BB157_517; st.local.f32 [%rd788+20], %f55; st.local.f32 [%rd788+24], %f56; st.local.f32 [%rd788+28], %f57; cvt.ftz.sat.f32.f32 %f2826, %f60; st.local.f32 [%rd788+32], %f2826; mov.f32 %f2827, 0f3C23D70A; max.ftz.f32 %f2828, %f2826, %f2827; mul.ftz.f32 %f2829, %f2828, %f2828; rcp.approx.ftz.f32 %f2830, %f2829; st.local.f32 [%rd788+36], %f2830; mov.u32 %r564, 17; st.local.u32 [%rd788+12], %r564; ld.local.u32 %r565, [%rd4+40]; or.b32 %r566, %r565, 12; st.local.u32 [%rd4+40], %r566; bra.uni $L__BB157_517; $L__BB157_325: ld.local.f32 %f2486, [%rd4+256]; mul.ftz.f32 %f374, %f2, %f2486; ld.local.f32 %f2487, [%rd4+260]; mul.ftz.f32 %f375, %f2, %f2487; ld.local.f32 %f2488, [%rd4+264]; mul.ftz.f32 %f376, %f2, %f2488; add.ftz.f32 %f2489, %f374, %f375; add.ftz.f32 %f2490, %f376, %f2489; mul.ftz.f32 %f2491, %f2490, 0f3EAAAAAB; abs.ftz.f32 %f377, %f2491; setp.ltu.ftz.f32 %p294, %f377, 0f3727C5AC; @%p294 bra $L__BB157_517; ld.local.u32 %r67, [%rd4+40]; and.b32 %r362, %r67, 512; setp.eq.s32 %p295, %r362, 0; @%p295 bra $L__BB157_332; bra.uni $L__BB157_327; $L__BB157_332: or.b32 %r365, %r67, 516; st.local.u32 [%rd4+40], %r365; st.local.f32 [%rd4+280], %f374; st.local.f32 [%rd4+284], %f375; st.local.f32 [%rd4+288], %f376; and.b32 %r71, %r141, 491520; setp.eq.s32 %p299, %r71, 0; @%p299 bra $L__BB157_334; mov.u32 %r651, 1; st.local.u32 [%rd4+252], %r651; bra.uni $L__BB157_335; $L__BB157_44: ld.local.f32 %f524, [%rd5]; ld.local.f32 %f525, [%rd5+8]; ld.local.f32 %f526, [%rd5+4]; st.local.v2.f32 [%rd205+36], {%f524, %f526}; st.local.f32 [%rd205+44], %f525; ld.local.f32 %f527, [%rd6]; setp.eq.ftz.f32 %p434, %f527, 0f00000000; @%p434 bra $L__BB157_46; mul.ftz.f32 %f2841, %f527, 0f40C90FDB; cos.approx.ftz.f32 %f2842, %f2841; sin.approx.ftz.f32 %f2843, %f2841; mov.f32 %f2844, 0f3F800000; sub.ftz.f32 %f2845, %f2844, %f2842; mul.ftz.f32 %f2846, %f55, %f2845; fma.rn.ftz.f32 %f2847, %f55, %f2846, %f2842; mul.ftz.f32 %f2848, %f56, %f2846; mul.ftz.f32 %f2849, %f57, %f2843; sub.ftz.f32 %f2850, %f2848, %f2849; mul.ftz.f32 %f2851, %f526, %f2850; fma.rn.ftz.f32 %f2852, %f524, %f2847, %f2851; mul.ftz.f32 %f2853, %f57, %f2846; mul.ftz.f32 %f2854, %f56, %f2843; add.ftz.f32 %f2855, %f2854, %f2853; add.ftz.f32 %f2856, %f2849, %f2848; mul.ftz.f32 %f2857, %f56, %f2845; fma.rn.ftz.f32 %f2858, %f56, %f2857, %f2842; mul.ftz.f32 %f2859, %f526, %f2858; fma.rn.ftz.f32 %f2860, %f524, %f2856, %f2859; mul.ftz.f32 %f2861, %f57, %f2857; mul.ftz.f32 %f2862, %f55, %f2843; sub.ftz.f32 %f2863, %f2861, %f2862; sub.ftz.f32 %f2864, %f2853, %f2854; add.ftz.f32 %f2865, %f2862, %f2861; mul.ftz.f32 %f2866, %f526, %f2865; fma.rn.ftz.f32 %f2867, %f524, %f2864, %f2866; mul.ftz.f32 %f2868, %f57, %f2845; fma.rn.ftz.f32 %f2869, %f57, %f2868, %f2842; fma.rn.ftz.f32 %f2870, %f525, %f2869, %f2867; fma.rn.ftz.f32 %f2871, %f525, %f2863, %f2860; fma.rn.ftz.f32 %f2872, %f525, %f2855, %f2852; st.local.v2.f32 [%rd205+36], {%f2872, %f2871}; st.local.f32 [%rd205+44], %f2870; $L__BB157_46: mov.f32 %f2873, 0fBF7D70A4; max.ftz.f32 %f2874, %f63, %f2873; mov.f32 %f2875, 0f3F7D70A4; min.ftz.f32 %f528, %f2874, %f2875; setp.lt.ftz.f32 %p435, %f528, 0f00000000; @%p435 bra $L__BB157_489; bra.uni $L__BB157_47; $L__BB157_489: add.ftz.f32 %f2878, %f528, 0f3F800000; div.approx.ftz.f32 %f3059, %f3058, %f2878; mul.ftz.f32 %f3058, %f3058, %f2878; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; bra.uni $L__BB157_491; $L__BB157_310: mov.u32 %r358, 0; st.u32 [%rd111+-16], %r358; bra.uni $L__BB157_324; $L__BB157_87: ld.local.f32 %f2782, [%rd5]; ld.local.f32 %f2783, [%rd5+4]; mul.ftz.f32 %f2784, %f2783, %f2783; fma.rn.ftz.f32 %f2785, %f2782, %f2782, %f2784; ld.local.f32 %f2786, [%rd5+8]; fma.rn.ftz.f32 %f2787, %f2786, %f2786, %f2785; sqrt.approx.ftz.f32 %f2788, %f2787; div.approx.ftz.f32 %f2789, %f2782, %f2788; div.approx.ftz.f32 %f2790, %f2783, %f2788; div.approx.ftz.f32 %f2791, %f2786, %f2788; st.local.f32 [%rd192+12], %f2789; st.local.f32 [%rd192+16], %f2790; st.local.f32 [%rd192+20], %f2791; bra.uni $L__BB157_479; $L__BB157_312: mov.f32 %f1931, 0f3F8147AE; max.ftz.f32 %f1932, %f63, %f1931; mov.f32 %f1933, 0f40733333; min.ftz.f32 %f1934, %f1932, %f1933; rcp.approx.ftz.f32 %f1935, %f1934; mov.f32 %f1936, 0f3F800000; fma.rn.ftz.f32 %f1937, %f1935, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f1938, %f1935, %f1937, 0f3F2B020C; fma.rn.ftz.f32 %f1939, %f1934, 0f3D8240B8, %f1938; add.ftz.f32 %f1940, %f1939, 0f3F800000; mul.ftz.f32 %f1941, %f1940, 0f3FAAAAAB; sub.ftz.f32 %f1942, %f1936, %f1939; div.approx.ftz.f32 %f355, %f1941, %f1942; ld.f32 %f356, [%rd111+12]; setp.lt.ftz.f32 %p255, %f356, 0f38D1B717; mov.f32 %f3032, 0f00000000; mov.f32 %f3030, %f3032; @%p255 bra $L__BB157_315; setp.ge.ftz.f32 %p256, %f356, 0f3F7EB852; mov.f32 %f3030, 0f3F7FFFEF; @%p256 bra $L__BB157_315; sub.ftz.f32 %f1945, %f1936, 0f3F000000; mul.ftz.f32 %f1946, %f1945, 0f40400000; sqrt.approx.ftz.f32 %f1947, %f1946; mul.ftz.f32 %f1948, %f355, %f1947; mul.ftz.f32 %f1949, %f1948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1950, %f1949; add.ftz.f32 %f1951, %f1950, 0f3F800000; mul.ftz.f32 %f1952, %f1951, 0f3E800000; mul.ftz.f32 %f1953, %f1947, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1954, %f1953; mul.ftz.f32 %f1955, %f1954, %f1952; setp.lt.ftz.f32 %p257, %f1955, %f356; selp.f32 %f1956, 0f3F800000, 0f3F000000, %p257; selp.f32 %f1957, 0f3F000000, 0f00000000, %p257; selp.f32 %f1958, 0f3F400000, 0f3E800000, %p257; sub.ftz.f32 %f1959, %f1936, %f1958; mul.ftz.f32 %f1960, %f1959, 0f40400000; sqrt.approx.ftz.f32 %f1961, %f1960; mul.ftz.f32 %f1962, %f1958, 0f3F000000; mul.ftz.f32 %f1963, %f355, %f1961; mul.ftz.f32 %f1964, %f1963, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1965, 0f3F800000; mul.ftz.f32 %f1967, %f1966, %f1962; mul.ftz.f32 %f1968, %f1961, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1969, %f1968; mul.ftz.f32 %f1970, %f1969, %f1967; setp.lt.ftz.f32 %p258, %f1970, %f356; selp.f32 %f1971, %f1956, %f1958, %p258; selp.f32 %f1972, %f1958, %f1957, %p258; add.ftz.f32 %f1973, %f1971, %f1972; mul.ftz.f32 %f1974, %f1973, 0f3F000000; sub.ftz.f32 %f1975, %f1936, %f1974; mul.ftz.f32 %f1976, %f1975, 0f40400000; sqrt.approx.ftz.f32 %f1977, %f1976; mul.ftz.f32 %f1978, %f1974, 0f3F000000; mul.ftz.f32 %f1979, %f355, %f1977; mul.ftz.f32 %f1980, %f1979, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1981, %f1980; add.ftz.f32 %f1982, %f1981, 0f3F800000; mul.ftz.f32 %f1983, %f1982, %f1978; mul.ftz.f32 %f1984, %f1977, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f1985, %f1983; setp.lt.ftz.f32 %p259, %f1986, %f356; selp.f32 %f1987, %f1971, %f1974, %p259; selp.f32 %f1988, %f1974, %f1972, %p259; add.ftz.f32 %f1989, %f1987, %f1988; mul.ftz.f32 %f1990, %f1989, 0f3F000000; sub.ftz.f32 %f1991, %f1936, %f1990; mul.ftz.f32 %f1992, %f1991, 0f40400000; sqrt.approx.ftz.f32 %f1993, %f1992; mul.ftz.f32 %f1994, %f1990, 0f3F000000; mul.ftz.f32 %f1995, %f355, %f1993; mul.ftz.f32 %f1996, %f1995, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1997, %f1996; add.ftz.f32 %f1998, %f1997, 0f3F800000; mul.ftz.f32 %f1999, %f1998, %f1994; mul.ftz.f32 %f2000, %f1993, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2001, %f2000; mul.ftz.f32 %f2002, %f2001, %f1999; setp.lt.ftz.f32 %p260, %f2002, %f356; selp.f32 %f2003, %f1987, %f1990, %p260; selp.f32 %f2004, %f1990, %f1988, %p260; add.ftz.f32 %f2005, %f2003, %f2004; mul.ftz.f32 %f2006, %f2005, 0f3F000000; sub.ftz.f32 %f2007, %f1936, %f2006; mul.ftz.f32 %f2008, %f2007, 0f40400000; sqrt.approx.ftz.f32 %f2009, %f2008; mul.ftz.f32 %f2010, %f2006, 0f3F000000; mul.ftz.f32 %f2011, %f355, %f2009; mul.ftz.f32 %f2012, %f2011, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2013, %f2012; add.ftz.f32 %f2014, %f2013, 0f3F800000; mul.ftz.f32 %f2015, %f2014, %f2010; mul.ftz.f32 %f2016, %f2009, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2017, %f2016; mul.ftz.f32 %f2018, %f2017, %f2015; setp.lt.ftz.f32 %p261, %f2018, %f356; selp.f32 %f2019, %f2003, %f2006, %p261; selp.f32 %f2020, %f2006, %f2004, %p261; add.ftz.f32 %f2021, %f2019, %f2020; mul.ftz.f32 %f2022, %f2021, 0f3F000000; sub.ftz.f32 %f2023, %f1936, %f2022; mul.ftz.f32 %f2024, %f2023, 0f40400000; sqrt.approx.ftz.f32 %f2025, %f2024; mul.ftz.f32 %f2026, %f2022, 0f3F000000; mul.ftz.f32 %f2027, %f355, %f2025; mul.ftz.f32 %f2028, %f2027, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2029, %f2028; add.ftz.f32 %f2030, %f2029, 0f3F800000; mul.ftz.f32 %f2031, %f2030, %f2026; mul.ftz.f32 %f2032, %f2025, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2033, %f2032; mul.ftz.f32 %f2034, %f2033, %f2031; setp.lt.ftz.f32 %p262, %f2034, %f356; selp.f32 %f2035, %f2019, %f2022, %p262; selp.f32 %f2036, %f2022, %f2020, %p262; add.ftz.f32 %f2037, %f2035, %f2036; mul.ftz.f32 %f2038, %f2037, 0f3F000000; sub.ftz.f32 %f2039, %f1936, %f2038; mul.ftz.f32 %f2040, %f2039, 0f40400000; sqrt.approx.ftz.f32 %f2041, %f2040; mul.ftz.f32 %f2042, %f2038, 0f3F000000; mul.ftz.f32 %f2043, %f355, %f2041; mul.ftz.f32 %f2044, %f2043, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2045, 0f3F800000; mul.ftz.f32 %f2047, %f2046, %f2042; mul.ftz.f32 %f2048, %f2041, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2049, %f2048; mul.ftz.f32 %f2050, %f2049, %f2047; setp.lt.ftz.f32 %p263, %f2050, %f356; selp.f32 %f2051, %f2035, %f2038, %p263; selp.f32 %f2052, %f2038, %f2036, %p263; add.ftz.f32 %f2053, %f2051, %f2052; mul.ftz.f32 %f2054, %f2053, 0f3F000000; sub.ftz.f32 %f2055, %f1936, %f2054; mul.ftz.f32 %f2056, %f2055, 0f40400000; sqrt.approx.ftz.f32 %f2057, %f2056; mul.ftz.f32 %f2058, %f2054, 0f3F000000; mul.ftz.f32 %f2059, %f355, %f2057; mul.ftz.f32 %f2060, %f2059, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2061, %f2060; add.ftz.f32 %f2062, %f2061, 0f3F800000; mul.ftz.f32 %f2063, %f2062, %f2058; mul.ftz.f32 %f2064, %f2057, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2065, %f2064; mul.ftz.f32 %f2066, %f2065, %f2063; setp.lt.ftz.f32 %p264, %f2066, %f356; selp.f32 %f2067, %f2051, %f2054, %p264; selp.f32 %f2068, %f2054, %f2052, %p264; add.ftz.f32 %f2069, %f2067, %f2068; mul.ftz.f32 %f2070, %f2069, 0f3F000000; sub.ftz.f32 %f2071, %f1936, %f2070; mul.ftz.f32 %f2072, %f2071, 0f40400000; sqrt.approx.ftz.f32 %f2073, %f2072; mul.ftz.f32 %f2074, %f2070, 0f3F000000; mul.ftz.f32 %f2075, %f355, %f2073; mul.ftz.f32 %f2076, %f2075, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2077, %f2076; add.ftz.f32 %f2078, %f2077, 0f3F800000; mul.ftz.f32 %f2079, %f2078, %f2074; mul.ftz.f32 %f2080, %f2073, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2081, %f2080; mul.ftz.f32 %f2082, %f2081, %f2079; setp.lt.ftz.f32 %p265, %f2082, %f356; selp.f32 %f2083, %f2067, %f2070, %p265; selp.f32 %f2084, %f2070, %f2068, %p265; add.ftz.f32 %f2085, %f2083, %f2084; mul.ftz.f32 %f2086, %f2085, 0f3F000000; sub.ftz.f32 %f2087, %f1936, %f2086; mul.ftz.f32 %f2088, %f2087, 0f40400000; sqrt.approx.ftz.f32 %f2089, %f2088; mul.ftz.f32 %f2090, %f2086, 0f3F000000; mul.ftz.f32 %f2091, %f355, %f2089; mul.ftz.f32 %f2092, %f2091, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2093, %f2092; add.ftz.f32 %f2094, %f2093, 0f3F800000; mul.ftz.f32 %f2095, %f2094, %f2090; mul.ftz.f32 %f2096, %f2089, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2097, %f2096; mul.ftz.f32 %f2098, %f2097, %f2095; setp.lt.ftz.f32 %p266, %f2098, %f356; selp.f32 %f2099, %f2083, %f2086, %p266; selp.f32 %f2100, %f2086, %f2084, %p266; add.ftz.f32 %f2101, %f2099, %f2100; mul.ftz.f32 %f2102, %f2101, 0f3F000000; sub.ftz.f32 %f2103, %f1936, %f2102; mul.ftz.f32 %f2104, %f2103, 0f40400000; sqrt.approx.ftz.f32 %f2105, %f2104; mul.ftz.f32 %f2106, %f2102, 0f3F000000; mul.ftz.f32 %f2107, %f355, %f2105; mul.ftz.f32 %f2108, %f2107, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2109, %f2108; add.ftz.f32 %f2110, %f2109, 0f3F800000; mul.ftz.f32 %f2111, %f2110, %f2106; mul.ftz.f32 %f2112, %f2105, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2113, %f2112; mul.ftz.f32 %f2114, %f2113, %f2111; setp.lt.ftz.f32 %p267, %f2114, %f356; selp.f32 %f2115, %f2099, %f2102, %p267; selp.f32 %f2116, %f2102, %f2100, %p267; add.ftz.f32 %f2117, %f2115, %f2116; mul.ftz.f32 %f3030, %f2117, 0f3F000000; $L__BB157_315: ld.f32 %f359, [%rd111+16]; setp.lt.ftz.f32 %p268, %f359, 0f38D1B717; mov.f32 %f3031, %f3032; @%p268 bra $L__BB157_318; setp.ge.ftz.f32 %p269, %f359, 0f3F7EB852; mov.f32 %f3031, 0f3F7FFFEF; @%p269 bra $L__BB157_318; sub.ftz.f32 %f2121, %f1936, 0f3F000000; mul.ftz.f32 %f2122, %f2121, 0f40400000; sqrt.approx.ftz.f32 %f2123, %f2122; mul.ftz.f32 %f2124, %f355, %f2123; mul.ftz.f32 %f2125, %f2124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2126, %f2125; add.ftz.f32 %f2127, %f2126, 0f3F800000; mul.ftz.f32 %f2128, %f2127, 0f3E800000; mul.ftz.f32 %f2129, %f2123, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2130, %f2129; mul.ftz.f32 %f2131, %f2130, %f2128; setp.lt.ftz.f32 %p270, %f2131, %f359; selp.f32 %f2132, 0f3F800000, 0f3F000000, %p270; selp.f32 %f2133, 0f3F000000, 0f00000000, %p270; selp.f32 %f2134, 0f3F400000, 0f3E800000, %p270; sub.ftz.f32 %f2135, %f1936, %f2134; mul.ftz.f32 %f2136, %f2135, 0f40400000; sqrt.approx.ftz.f32 %f2137, %f2136; mul.ftz.f32 %f2138, %f2134, 0f3F000000; mul.ftz.f32 %f2139, %f355, %f2137; mul.ftz.f32 %f2140, %f2139, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2141, %f2140; add.ftz.f32 %f2142, %f2141, 0f3F800000; mul.ftz.f32 %f2143, %f2142, %f2138; mul.ftz.f32 %f2144, %f2137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2145, %f2144; mul.ftz.f32 %f2146, %f2145, %f2143; setp.lt.ftz.f32 %p271, %f2146, %f359; selp.f32 %f2147, %f2132, %f2134, %p271; selp.f32 %f2148, %f2134, %f2133, %p271; add.ftz.f32 %f2149, %f2147, %f2148; mul.ftz.f32 %f2150, %f2149, 0f3F000000; sub.ftz.f32 %f2151, %f1936, %f2150; mul.ftz.f32 %f2152, %f2151, 0f40400000; sqrt.approx.ftz.f32 %f2153, %f2152; mul.ftz.f32 %f2154, %f2150, 0f3F000000; mul.ftz.f32 %f2155, %f355, %f2153; mul.ftz.f32 %f2156, %f2155, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2157, %f2156; add.ftz.f32 %f2158, %f2157, 0f3F800000; mul.ftz.f32 %f2159, %f2158, %f2154; mul.ftz.f32 %f2160, %f2153, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2161, %f2160; mul.ftz.f32 %f2162, %f2161, %f2159; setp.lt.ftz.f32 %p272, %f2162, %f359; selp.f32 %f2163, %f2147, %f2150, %p272; selp.f32 %f2164, %f2150, %f2148, %p272; add.ftz.f32 %f2165, %f2163, %f2164; mul.ftz.f32 %f2166, %f2165, 0f3F000000; sub.ftz.f32 %f2167, %f1936, %f2166; mul.ftz.f32 %f2168, %f2167, 0f40400000; sqrt.approx.ftz.f32 %f2169, %f2168; mul.ftz.f32 %f2170, %f2166, 0f3F000000; mul.ftz.f32 %f2171, %f355, %f2169; mul.ftz.f32 %f2172, %f2171, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2173, %f2172; add.ftz.f32 %f2174, %f2173, 0f3F800000; mul.ftz.f32 %f2175, %f2174, %f2170; mul.ftz.f32 %f2176, %f2169, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2177, %f2176; mul.ftz.f32 %f2178, %f2177, %f2175; setp.lt.ftz.f32 %p273, %f2178, %f359; selp.f32 %f2179, %f2163, %f2166, %p273; selp.f32 %f2180, %f2166, %f2164, %p273; add.ftz.f32 %f2181, %f2179, %f2180; mul.ftz.f32 %f2182, %f2181, 0f3F000000; sub.ftz.f32 %f2183, %f1936, %f2182; mul.ftz.f32 %f2184, %f2183, 0f40400000; sqrt.approx.ftz.f32 %f2185, %f2184; mul.ftz.f32 %f2186, %f2182, 0f3F000000; mul.ftz.f32 %f2187, %f355, %f2185; mul.ftz.f32 %f2188, %f2187, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2189, %f2188; add.ftz.f32 %f2190, %f2189, 0f3F800000; mul.ftz.f32 %f2191, %f2190, %f2186; mul.ftz.f32 %f2192, %f2185, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2193, %f2192; mul.ftz.f32 %f2194, %f2193, %f2191; setp.lt.ftz.f32 %p274, %f2194, %f359; selp.f32 %f2195, %f2179, %f2182, %p274; selp.f32 %f2196, %f2182, %f2180, %p274; add.ftz.f32 %f2197, %f2195, %f2196; mul.ftz.f32 %f2198, %f2197, 0f3F000000; sub.ftz.f32 %f2199, %f1936, %f2198; mul.ftz.f32 %f2200, %f2199, 0f40400000; sqrt.approx.ftz.f32 %f2201, %f2200; mul.ftz.f32 %f2202, %f2198, 0f3F000000; mul.ftz.f32 %f2203, %f355, %f2201; mul.ftz.f32 %f2204, %f2203, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2205, %f2204; add.ftz.f32 %f2206, %f2205, 0f3F800000; mul.ftz.f32 %f2207, %f2206, %f2202; mul.ftz.f32 %f2208, %f2201, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2209, %f2208; mul.ftz.f32 %f2210, %f2209, %f2207; setp.lt.ftz.f32 %p275, %f2210, %f359; selp.f32 %f2211, %f2195, %f2198, %p275; selp.f32 %f2212, %f2198, %f2196, %p275; add.ftz.f32 %f2213, %f2211, %f2212; mul.ftz.f32 %f2214, %f2213, 0f3F000000; sub.ftz.f32 %f2215, %f1936, %f2214; mul.ftz.f32 %f2216, %f2215, 0f40400000; sqrt.approx.ftz.f32 %f2217, %f2216; mul.ftz.f32 %f2218, %f2214, 0f3F000000; mul.ftz.f32 %f2219, %f355, %f2217; mul.ftz.f32 %f2220, %f2219, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2221, %f2220; add.ftz.f32 %f2222, %f2221, 0f3F800000; mul.ftz.f32 %f2223, %f2222, %f2218; mul.ftz.f32 %f2224, %f2217, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2225, %f2224; mul.ftz.f32 %f2226, %f2225, %f2223; setp.lt.ftz.f32 %p276, %f2226, %f359; selp.f32 %f2227, %f2211, %f2214, %p276; selp.f32 %f2228, %f2214, %f2212, %p276; add.ftz.f32 %f2229, %f2227, %f2228; mul.ftz.f32 %f2230, %f2229, 0f3F000000; sub.ftz.f32 %f2231, %f1936, %f2230; mul.ftz.f32 %f2232, %f2231, 0f40400000; sqrt.approx.ftz.f32 %f2233, %f2232; mul.ftz.f32 %f2234, %f2230, 0f3F000000; mul.ftz.f32 %f2235, %f355, %f2233; mul.ftz.f32 %f2236, %f2235, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2237, %f2236; add.ftz.f32 %f2238, %f2237, 0f3F800000; mul.ftz.f32 %f2239, %f2238, %f2234; mul.ftz.f32 %f2240, %f2233, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2241, %f2240; mul.ftz.f32 %f2242, %f2241, %f2239; setp.lt.ftz.f32 %p277, %f2242, %f359; selp.f32 %f2243, %f2227, %f2230, %p277; selp.f32 %f2244, %f2230, %f2228, %p277; add.ftz.f32 %f2245, %f2243, %f2244; mul.ftz.f32 %f2246, %f2245, 0f3F000000; sub.ftz.f32 %f2247, %f1936, %f2246; mul.ftz.f32 %f2248, %f2247, 0f40400000; sqrt.approx.ftz.f32 %f2249, %f2248; mul.ftz.f32 %f2250, %f2246, 0f3F000000; mul.ftz.f32 %f2251, %f355, %f2249; mul.ftz.f32 %f2252, %f2251, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2253, %f2252; add.ftz.f32 %f2254, %f2253, 0f3F800000; mul.ftz.f32 %f2255, %f2254, %f2250; mul.ftz.f32 %f2256, %f2249, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2257, %f2256; mul.ftz.f32 %f2258, %f2257, %f2255; setp.lt.ftz.f32 %p278, %f2258, %f359; selp.f32 %f2259, %f2243, %f2246, %p278; selp.f32 %f2260, %f2246, %f2244, %p278; add.ftz.f32 %f2261, %f2259, %f2260; mul.ftz.f32 %f2262, %f2261, 0f3F000000; sub.ftz.f32 %f2263, %f1936, %f2262; mul.ftz.f32 %f2264, %f2263, 0f40400000; sqrt.approx.ftz.f32 %f2265, %f2264; mul.ftz.f32 %f2266, %f2262, 0f3F000000; mul.ftz.f32 %f2267, %f355, %f2265; mul.ftz.f32 %f2268, %f2267, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2269, %f2268; add.ftz.f32 %f2270, %f2269, 0f3F800000; mul.ftz.f32 %f2271, %f2270, %f2266; mul.ftz.f32 %f2272, %f2265, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2273, %f2272; mul.ftz.f32 %f2274, %f2273, %f2271; setp.lt.ftz.f32 %p279, %f2274, %f359; selp.f32 %f2275, %f2259, %f2262, %p279; selp.f32 %f2276, %f2262, %f2260, %p279; add.ftz.f32 %f2277, %f2275, %f2276; mul.ftz.f32 %f2278, %f2277, 0f3F000000; sub.ftz.f32 %f2279, %f1936, %f2278; mul.ftz.f32 %f2280, %f2279, 0f40400000; sqrt.approx.ftz.f32 %f2281, %f2280; mul.ftz.f32 %f2282, %f2278, 0f3F000000; mul.ftz.f32 %f2283, %f355, %f2281; mul.ftz.f32 %f2284, %f2283, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2285, %f2284; add.ftz.f32 %f2286, %f2285, 0f3F800000; mul.ftz.f32 %f2287, %f2286, %f2282; mul.ftz.f32 %f2288, %f2281, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f2289, %f2287; setp.lt.ftz.f32 %p280, %f2290, %f359; selp.f32 %f2291, %f2275, %f2278, %p280; selp.f32 %f2292, %f2278, %f2276, %p280; add.ftz.f32 %f2293, %f2291, %f2292; mul.ftz.f32 %f3031, %f2293, 0f3F000000; $L__BB157_318: ld.f32 %f362, [%rd111+20]; setp.lt.ftz.f32 %p281, %f362, 0f38D1B717; @%p281 bra $L__BB157_321; setp.ge.ftz.f32 %p282, %f362, 0f3F7EB852; mov.f32 %f3032, 0f3F7FFFEF; @%p282 bra $L__BB157_321; sub.ftz.f32 %f2297, %f1936, 0f3F000000; mul.ftz.f32 %f2298, %f2297, 0f40400000; sqrt.approx.ftz.f32 %f2299, %f2298; mul.ftz.f32 %f2300, %f355, %f2299; mul.ftz.f32 %f2301, %f2300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2302, %f2301; add.ftz.f32 %f2303, %f2302, 0f3F800000; mul.ftz.f32 %f2304, %f2303, 0f3E800000; mul.ftz.f32 %f2305, %f2299, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2306, %f2305; mul.ftz.f32 %f2307, %f2306, %f2304; setp.lt.ftz.f32 %p283, %f2307, %f362; selp.f32 %f2308, 0f3F800000, 0f3F000000, %p283; selp.f32 %f2309, 0f3F000000, 0f00000000, %p283; selp.f32 %f2310, 0f3F400000, 0f3E800000, %p283; sub.ftz.f32 %f2311, %f1936, %f2310; mul.ftz.f32 %f2312, %f2311, 0f40400000; sqrt.approx.ftz.f32 %f2313, %f2312; mul.ftz.f32 %f2314, %f2310, 0f3F000000; mul.ftz.f32 %f2315, %f355, %f2313; mul.ftz.f32 %f2316, %f2315, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2317, %f2316; add.ftz.f32 %f2318, %f2317, 0f3F800000; mul.ftz.f32 %f2319, %f2318, %f2314; mul.ftz.f32 %f2320, %f2313, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2321, %f2320; mul.ftz.f32 %f2322, %f2321, %f2319; setp.lt.ftz.f32 %p284, %f2322, %f362; selp.f32 %f2323, %f2308, %f2310, %p284; selp.f32 %f2324, %f2310, %f2309, %p284; add.ftz.f32 %f2325, %f2323, %f2324; mul.ftz.f32 %f2326, %f2325, 0f3F000000; sub.ftz.f32 %f2327, %f1936, %f2326; mul.ftz.f32 %f2328, %f2327, 0f40400000; sqrt.approx.ftz.f32 %f2329, %f2328; mul.ftz.f32 %f2330, %f2326, 0f3F000000; mul.ftz.f32 %f2331, %f355, %f2329; mul.ftz.f32 %f2332, %f2331, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2333, %f2332; add.ftz.f32 %f2334, %f2333, 0f3F800000; mul.ftz.f32 %f2335, %f2334, %f2330; mul.ftz.f32 %f2336, %f2329, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2337, %f2336; mul.ftz.f32 %f2338, %f2337, %f2335; setp.lt.ftz.f32 %p285, %f2338, %f362; selp.f32 %f2339, %f2323, %f2326, %p285; selp.f32 %f2340, %f2326, %f2324, %p285; add.ftz.f32 %f2341, %f2339, %f2340; mul.ftz.f32 %f2342, %f2341, 0f3F000000; sub.ftz.f32 %f2343, %f1936, %f2342; mul.ftz.f32 %f2344, %f2343, 0f40400000; sqrt.approx.ftz.f32 %f2345, %f2344; mul.ftz.f32 %f2346, %f2342, 0f3F000000; mul.ftz.f32 %f2347, %f355, %f2345; mul.ftz.f32 %f2348, %f2347, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2349, %f2348; add.ftz.f32 %f2350, %f2349, 0f3F800000; mul.ftz.f32 %f2351, %f2350, %f2346; mul.ftz.f32 %f2352, %f2345, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2353, %f2352; mul.ftz.f32 %f2354, %f2353, %f2351; setp.lt.ftz.f32 %p286, %f2354, %f362; selp.f32 %f2355, %f2339, %f2342, %p286; selp.f32 %f2356, %f2342, %f2340, %p286; add.ftz.f32 %f2357, %f2355, %f2356; mul.ftz.f32 %f2358, %f2357, 0f3F000000; sub.ftz.f32 %f2359, %f1936, %f2358; mul.ftz.f32 %f2360, %f2359, 0f40400000; sqrt.approx.ftz.f32 %f2361, %f2360; mul.ftz.f32 %f2362, %f2358, 0f3F000000; mul.ftz.f32 %f2363, %f355, %f2361; mul.ftz.f32 %f2364, %f2363, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2365, %f2364; add.ftz.f32 %f2366, %f2365, 0f3F800000; mul.ftz.f32 %f2367, %f2366, %f2362; mul.ftz.f32 %f2368, %f2361, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2369, %f2368; mul.ftz.f32 %f2370, %f2369, %f2367; setp.lt.ftz.f32 %p287, %f2370, %f362; selp.f32 %f2371, %f2355, %f2358, %p287; selp.f32 %f2372, %f2358, %f2356, %p287; add.ftz.f32 %f2373, %f2371, %f2372; mul.ftz.f32 %f2374, %f2373, 0f3F000000; sub.ftz.f32 %f2375, %f1936, %f2374; mul.ftz.f32 %f2376, %f2375, 0f40400000; sqrt.approx.ftz.f32 %f2377, %f2376; mul.ftz.f32 %f2378, %f2374, 0f3F000000; mul.ftz.f32 %f2379, %f355, %f2377; mul.ftz.f32 %f2380, %f2379, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2381, %f2380; add.ftz.f32 %f2382, %f2381, 0f3F800000; mul.ftz.f32 %f2383, %f2382, %f2378; mul.ftz.f32 %f2384, %f2377, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2385, %f2384; mul.ftz.f32 %f2386, %f2385, %f2383; setp.lt.ftz.f32 %p288, %f2386, %f362; selp.f32 %f2387, %f2371, %f2374, %p288; selp.f32 %f2388, %f2374, %f2372, %p288; add.ftz.f32 %f2389, %f2387, %f2388; mul.ftz.f32 %f2390, %f2389, 0f3F000000; sub.ftz.f32 %f2391, %f1936, %f2390; mul.ftz.f32 %f2392, %f2391, 0f40400000; sqrt.approx.ftz.f32 %f2393, %f2392; mul.ftz.f32 %f2394, %f2390, 0f3F000000; mul.ftz.f32 %f2395, %f355, %f2393; mul.ftz.f32 %f2396, %f2395, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2397, %f2396; add.ftz.f32 %f2398, %f2397, 0f3F800000; mul.ftz.f32 %f2399, %f2398, %f2394; mul.ftz.f32 %f2400, %f2393, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f2401, %f2399; setp.lt.ftz.f32 %p289, %f2402, %f362; selp.f32 %f2403, %f2387, %f2390, %p289; selp.f32 %f2404, %f2390, %f2388, %p289; add.ftz.f32 %f2405, %f2403, %f2404; mul.ftz.f32 %f2406, %f2405, 0f3F000000; sub.ftz.f32 %f2407, %f1936, %f2406; mul.ftz.f32 %f2408, %f2407, 0f40400000; sqrt.approx.ftz.f32 %f2409, %f2408; mul.ftz.f32 %f2410, %f2406, 0f3F000000; mul.ftz.f32 %f2411, %f355, %f2409; mul.ftz.f32 %f2412, %f2411, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2413, %f2412; add.ftz.f32 %f2414, %f2413, 0f3F800000; mul.ftz.f32 %f2415, %f2414, %f2410; mul.ftz.f32 %f2416, %f2409, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2417, %f2416; mul.ftz.f32 %f2418, %f2417, %f2415; setp.lt.ftz.f32 %p290, %f2418, %f362; selp.f32 %f2419, %f2403, %f2406, %p290; selp.f32 %f2420, %f2406, %f2404, %p290; add.ftz.f32 %f2421, %f2419, %f2420; mul.ftz.f32 %f2422, %f2421, 0f3F000000; sub.ftz.f32 %f2423, %f1936, %f2422; mul.ftz.f32 %f2424, %f2423, 0f40400000; sqrt.approx.ftz.f32 %f2425, %f2424; mul.ftz.f32 %f2426, %f2422, 0f3F000000; mul.ftz.f32 %f2427, %f355, %f2425; mul.ftz.f32 %f2428, %f2427, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2429, %f2428; add.ftz.f32 %f2430, %f2429, 0f3F800000; mul.ftz.f32 %f2431, %f2430, %f2426; mul.ftz.f32 %f2432, %f2425, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2433, %f2432; mul.ftz.f32 %f2434, %f2433, %f2431; setp.lt.ftz.f32 %p291, %f2434, %f362; selp.f32 %f2435, %f2419, %f2422, %p291; selp.f32 %f2436, %f2422, %f2420, %p291; add.ftz.f32 %f2437, %f2435, %f2436; mul.ftz.f32 %f2438, %f2437, 0f3F000000; sub.ftz.f32 %f2439, %f1936, %f2438; mul.ftz.f32 %f2440, %f2439, 0f40400000; sqrt.approx.ftz.f32 %f2441, %f2440; mul.ftz.f32 %f2442, %f2438, 0f3F000000; mul.ftz.f32 %f2443, %f355, %f2441; mul.ftz.f32 %f2444, %f2443, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2445, %f2444; add.ftz.f32 %f2446, %f2445, 0f3F800000; mul.ftz.f32 %f2447, %f2446, %f2442; mul.ftz.f32 %f2448, %f2441, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2449, %f2448; mul.ftz.f32 %f2450, %f2449, %f2447; setp.lt.ftz.f32 %p292, %f2450, %f362; selp.f32 %f2451, %f2435, %f2438, %p292; selp.f32 %f2452, %f2438, %f2436, %p292; add.ftz.f32 %f2453, %f2451, %f2452; mul.ftz.f32 %f2454, %f2453, 0f3F000000; sub.ftz.f32 %f2455, %f1936, %f2454; mul.ftz.f32 %f2456, %f2455, 0f40400000; sqrt.approx.ftz.f32 %f2457, %f2456; mul.ftz.f32 %f2458, %f2454, 0f3F000000; mul.ftz.f32 %f2459, %f355, %f2457; mul.ftz.f32 %f2460, %f2459, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2461, %f2460; add.ftz.f32 %f2462, %f2461, 0f3F800000; mul.ftz.f32 %f2463, %f2462, %f2458; mul.ftz.f32 %f2464, %f2457, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2465, %f2464; mul.ftz.f32 %f2466, %f2465, %f2463; setp.lt.ftz.f32 %p293, %f2466, %f362; selp.f32 %f2467, %f2451, %f2454, %p293; selp.f32 %f2468, %f2454, %f2452, %p293; add.ftz.f32 %f2469, %f2467, %f2468; mul.ftz.f32 %f3032, %f2469, 0f3F000000; $L__BB157_321: sub.ftz.f32 %f2471, %f1936, %f3030; mul.ftz.f32 %f2472, %f2471, 0f40400000; sub.ftz.f32 %f2473, %f1936, %f3031; mul.ftz.f32 %f2474, %f2473, 0f40400000; sub.ftz.f32 %f2475, %f1936, %f3032; mul.ftz.f32 %f2476, %f2475, 0f40400000; sqrt.approx.ftz.f32 %f2477, %f2472; sqrt.approx.ftz.f32 %f2478, %f2474; sqrt.approx.ftz.f32 %f2479, %f2476; ld.f32 %f2480, [%rd111]; mul.ftz.f32 %f3033, %f2477, %f2480; ld.f32 %f2481, [%rd111+4]; mul.ftz.f32 %f3034, %f2478, %f2481; ld.f32 %f2482, [%rd111+8]; mul.ftz.f32 %f3035, %f2479, %f2482; $L__BB157_323: st.f32 [%rd111], %f3033; st.f32 [%rd111+4], %f3034; st.f32 [%rd111+8], %f3035; or.b32 %r648, %r648, 16; $L__BB157_324: ld.local.u32 %r360, [%rd4+40]; or.b32 %r361, %r360, %r648; st.local.u32 [%rd4+40], %r361; bra.uni $L__BB157_517; $L__BB157_98: mul.wide.u32 %rd279, %r166, 4; add.s64 %rd280, %rd2, %rd279; ld.local.f32 %f2982, [%rd280]; ld.local.f32 %f2983, [%rd280+4]; ld.local.f32 %f2984, [%rd280+8]; $L__BB157_100: ld.global.v4.u32 {%r170, %r171, %r172, %r173}, [%rd3+48]; setp.eq.s32 %p56, %r170, 255; @%p56 bra $L__BB157_102; bra.uni $L__BB157_101; $L__BB157_102: ld.local.f32 %f2992, [%rd4+8]; ld.local.f32 %f2991, [%rd4+4]; ld.local.f32 %f2990, [%rd4]; bra.uni $L__BB157_103; $L__BB157_342: mul.wide.u32 %rd477, %r83, 4; add.s64 %rd478, %rd2, %rd477; ld.local.f32 %f3036, [%rd478]; $L__BB157_344: setp.eq.s32 %p305, %r82, 255; mul.ftz.f32 %f384, %f2, %f380; mul.ftz.f32 %f385, %f2, %f379; mul.ftz.f32 %f386, %f2, %f378; @%p305 bra $L__BB157_346; bra.uni $L__BB157_345; $L__BB157_346: mov.b32 %f3037, %r146; bra.uni $L__BB157_347; $L__BB157_101: mul.wide.u32 %rd281, %r170, 4; add.s64 %rd282, %rd2, %rd281; ld.local.f32 %f2992, [%rd282+8]; ld.local.f32 %f2991, [%rd282+4]; ld.local.f32 %f2990, [%rd282]; $L__BB157_103: @%p10 bra $L__BB157_112; ld.local.v2.f32 {%f677, %f678}, [%rd4+12]; mul.ftz.f32 %f679, %f2991, %f91; fma.rn.ftz.f32 %f680, %f2990, %f90, %f679; fma.rn.ftz.f32 %f681, %f2992, %f92, %f680; add.ftz.f32 %f682, %f681, %f681; mul.ftz.f32 %f683, %f2990, %f682; mul.ftz.f32 %f684, %f2991, %f682; mul.ftz.f32 %f685, %f2992, %f682; sub.ftz.f32 %f686, %f683, %f90; sub.ftz.f32 %f687, %f684, %f91; sub.ftz.f32 %f688, %f685, %f92; mul.ftz.f32 %f689, %f678, %f91; fma.rn.ftz.f32 %f690, %f677, %f90, %f689; ld.local.f32 %f123, [%rd4+20]; fma.rn.ftz.f32 %f124, %f123, %f92, %f690; mul.ftz.f32 %f691, %f124, 0f3F666666; mov.f32 %f692, 0f3C23D70A; min.ftz.f32 %f125, %f691, %f692; mul.ftz.f32 %f693, %f678, %f687; fma.rn.ftz.f32 %f694, %f677, %f686, %f693; fma.rn.ftz.f32 %f695, %f123, %f688, %f694; setp.ltu.ftz.f32 %p58, %f695, %f125; @%p58 bra $L__BB157_105; bra.uni $L__BB157_112; $L__BB157_105: mul.ftz.f32 %f696, %f2991, %f678; fma.rn.ftz.f32 %f697, %f2990, %f677, %f696; fma.rn.ftz.f32 %f698, %f2992, %f123, %f697; mul.ftz.f32 %f699, %f677, %f698; mul.ftz.f32 %f700, %f678, %f698; mul.ftz.f32 %f701, %f123, %f698; sub.ftz.f32 %f702, %f2990, %f699; sub.ftz.f32 %f703, %f2991, %f700; sub.ftz.f32 %f704, %f2992, %f701; mul.ftz.f32 %f705, %f703, %f703; fma.rn.ftz.f32 %f706, %f702, %f702, %f705; fma.rn.ftz.f32 %f707, %f704, %f704, %f706; sqrt.approx.ftz.f32 %f708, %f707; div.approx.ftz.f32 %f129, %f702, %f708; div.approx.ftz.f32 %f130, %f703, %f708; div.approx.ftz.f32 %f131, %f704, %f708; mul.ftz.f32 %f709, %f91, %f130; fma.rn.ftz.f32 %f710, %f90, %f129, %f709; fma.rn.ftz.f32 %f132, %f92, %f131, %f710; mul.ftz.f32 %f711, %f132, %f132; fma.rn.ftz.f32 %f712, %f124, %f124, %f711; mul.ftz.f32 %f713, %f125, %f125; sub.ftz.f32 %f714, %f712, %f713; mul.ftz.f32 %f715, %f711, %f714; mov.f32 %f716, 0f00000000; max.ftz.f32 %f717, %f715, %f716; sqrt.approx.ftz.f32 %f718, %f717; fma.rn.ftz.f32 %f719, %f125, %f124, %f712; mov.f32 %f720, 0f3F000000; div.approx.ftz.f32 %f721, %f720, %f712; add.ftz.f32 %f722, %f718, %f719; mul.ftz.f32 %f133, %f721, %f722; sub.ftz.f32 %f723, %f719, %f718; mul.ftz.f32 %f134, %f721, %f723; setp.gt.ftz.f32 %p59, %f133, 0f3727C5AC; setp.le.ftz.f32 %p60, %f133, 0f3F800054; and.pred %p3, %p59, %p60; setp.gt.ftz.f32 %p61, %f134, 0f3727C5AC; setp.le.ftz.f32 %p62, %f134, 0f3F800054; and.pred %p4, %p61, %p62; and.pred %p63, %p3, %p4; @%p63 bra $L__BB157_108; bra.uni $L__BB157_106; $L__BB157_108: mov.f32 %f730, 0f3F800000; sub.ftz.f32 %f731, %f730, %f133; max.ftz.f32 %f733, %f731, %f716; sqrt.approx.ftz.f32 %f140, %f733; max.ftz.f32 %f734, %f133, %f716; sqrt.approx.ftz.f32 %f141, %f734; sub.ftz.f32 %f735, %f730, %f134; max.ftz.f32 %f736, %f735, %f716; sqrt.approx.ftz.f32 %f142, %f736; max.ftz.f32 %f737, %f134, %f716; sqrt.approx.ftz.f32 %f143, %f737; mul.ftz.f32 %f738, %f124, %f141; fma.rn.ftz.f32 %f739, %f132, %f140, %f738; add.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f141, %f740; sub.ftz.f32 %f144, %f741, %f124; mul.ftz.f32 %f742, %f124, %f143; fma.rn.ftz.f32 %f743, %f132, %f142, %f742; add.ftz.f32 %f744, %f743, %f743; mul.ftz.f32 %f745, %f143, %f744; sub.ftz.f32 %f145, %f745, %f124; setp.ge.ftz.f32 %p65, %f144, 0f3727C5AC; setp.ge.ftz.f32 %p66, %f145, 0f3727C5AC; and.pred %p67, %p65, %p66; @%p67 bra $L__BB157_110; bra.uni $L__BB157_109; $L__BB157_110: setp.lt.ftz.f32 %p69, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p69; selp.f32 %f2989, %f141, %f143, %p69; bra.uni $L__BB157_111; $L__BB157_345: mul.wide.u32 %rd479, %r82, 4; add.s64 %rd480, %rd2, %rd479; ld.local.f32 %f3037, [%rd480]; $L__BB157_347: and.b32 %r86, %r375, 255; bfe.u32 %r87, %r375, 8, 8; bfe.u32 %r88, %r375, 16, 8; bfe.u32 %r89, %r379, 8, 8; bfe.u32 %r90, %r379, 16, 8; shr.u32 %r91, %r379, 24; ld.local.u32 %r92, [%rd4+64]; ld.local.s32 %rd122, [%rd4+48]; setp.eq.s32 %p306, %r92, -1; mov.u32 %r113, -1; mov.u32 %r110, 0; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p306 bra $L__BB157_362; ld.const.u64 %rd481, [kernel_params+2160]; cvta.to.global.u64 %rd482, %rd481; mul.wide.s32 %rd483, %r92, 208; add.s64 %rd484, %rd482, %rd483; ld.global.u32 %r93, [%rd484+164]; and.b32 %r388, %r11, 1; setp.eq.b32 %p307, %r388, 1; mov.pred %p308, 0; xor.pred %p309, %p307, %p308; not.pred %p310, %p309; mov.u32 %r653, 0; @%p310 bra $L__BB157_352; cvt.u32.u64 %r390, %rd122; setp.eq.s32 %p311, %r390, -1; mov.u32 %r652, -1; @%p311 bra $L__BB157_351; ld.const.u64 %rd485, [kernel_params+2240]; cvta.to.global.u64 %rd486, %rd485; shl.b64 %rd487, %rd122, 2; add.s64 %rd488, %rd486, %rd487; ld.global.u32 %r652, [%rd488]; $L__BB157_351: setp.ne.s32 %p312, %r652, -1; selp.u32 %r653, 1, 0, %p312; $L__BB157_352: ld.const.u64 %rd489, [kernel_params+2312]; cvta.to.global.u64 %rd123, %rd489; add.s32 %r655, %r653, %r93; mul.wide.u32 %rd490, %r655, 16; add.s64 %rd491, %rd123, %rd490; add.s64 %rd754, %rd491, 14; ld.global.u16 %rs20, [%rd491+12]; ld.global.u32 %r654, [%rd491+8]; ld.global.u64 %rd752, [%rd491]; setp.eq.s64 %p313, %rd752, %rd121; add.s64 %rd753, %rd491, 15; @%p313 bra $L__BB157_358; $L__BB157_353: setp.eq.s64 %p314, %rd752, 0; @%p314 bra $L__BB157_355; add.s32 %r655, %r655, 2; bra.uni $L__BB157_356; $L__BB157_355: setp.eq.s16 %p315, %rs20, 0; mov.u32 %r110, 0; mov.u32 %r655, %r654; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p315 bra $L__BB157_362; $L__BB157_356: cvt.u64.u32 %rd128, %r655; mul.wide.u32 %rd492, %r655, 16; add.s64 %rd493, %rd123, %rd492; ld.global.u16 %rs20, [%rd493+12]; ld.global.u32 %r654, [%rd493+8]; ld.global.u64 %rd752, [%rd493]; setp.ne.s64 %p316, %rd752, %rd121; @%p316 bra $L__BB157_353; shl.b64 %rd494, %rd128, 4; add.s64 %rd495, %rd123, %rd494; add.s64 %rd754, %rd495, 14; add.s64 %rd753, %rd495, 15; $L__BB157_358: ld.global.u8 %rs5, [%rd754]; ld.global.u8 %rs6, [%rd753]; cvt.u32.u64 %r395, %rd122; setp.ne.s32 %p317, %r395, -1; @%p317 bra $L__BB157_361; mov.u32 %r110, 0; add.s16 %rs14, %rs20, -1; setp.lt.u16 %p318, %rs14, 2; @%p318 bra $L__BB157_361; setp.ne.s16 %p319, %rs20, 1024; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p319 bra $L__BB157_362; $L__BB157_361: setp.eq.s16 %p320, %rs20, 0; selp.b32 %r113, -1, %r654, %p320; cvt.u32.u16 %r110, %rs20; cvt.u32.u16 %r400, %rs6; and.b32 %r660, %r400, 255; cvt.u32.u16 %r401, %rs5; and.b32 %r659, %r401, 255; $L__BB157_362: setp.eq.s32 %p321, %r113, -1; @%p321 bra $L__BB157_382; and.b32 %r402, %r11, 1; setp.eq.b32 %p322, %r402, 1; mov.pred %p323, 0; xor.pred %p324, %p322, %p323; not.pred %p325, %p324; @%p325 bra $L__BB157_375; bra.uni $L__BB157_364; $L__BB157_375: setp.eq.s32 %p332, %r12, 0; @%p332 bra $L__BB157_380; and.b32 %r423, %r110, 768; setp.eq.s32 %p333, %r423, 0; @%p333 bra $L__BB157_378; ld.const.u64 %rd517, [kernel_params+2264]; cvta.to.global.u64 %rd518, %rd517; shl.b64 %rd519, %rd122, 4; add.s64 %rd520, %rd518, %rd519; shr.s32 %r424, %r11, 7; ld.global.u32 %r425, [%rd520+4]; add.s32 %r426, %r424, %r425; ld.const.u64 %rd521, [kernel_params+2320]; cvta.to.global.u64 %rd522, %rd521; add.s32 %r427, %r426, %r113; mul.wide.s32 %rd523, %r427, 4; add.s64 %rd524, %rd522, %rd523; ld.local.f32 %f2520, [%rd4+56]; mov.f32 %f2521, 0f3F800000; sub.ftz.f32 %f2522, %f2521, %f2520; ld.global.f32 %f2523, [%rd524]; mul.ftz.f32 %f2524, %f2523, %f2522; ld.global.f32 %f2525, [%rd524+4]; fma.rn.ftz.f32 %f3038, %f2525, %f2520, %f2524; bra.uni $L__BB157_385; $L__BB157_472: mul.ftz.f32 %f2768, %f60, %f60; st.local.v2.f32 [%rd185+12], {%f2768, %f2768}; st.local.f32 [%rd185+20], %f3057; cvt.ftz.sat.f32.f32 %f507, %f2768; setp.eq.s32 %p411, %r2, 22; @%p411 bra $L__BB157_474; bra.uni $L__BB157_473; $L__BB157_474: st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r534, 22; st.local.u32 [%rd185+-8], %r534; ld.local.u32 %r535, [%rd4+40]; or.b32 %r536, %r535, 12; st.local.u32 [%rd4+40], %r536; bra.uni $L__BB157_517; $L__BB157_60: mov.u32 %r617, 19; st.local.u32 [%rd222+-8], %r617; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r618, [%rd4+40]; or.b32 %r619, %r618, 12; st.local.u32 [%rd4+40], %r619; bra.uni $L__BB157_517; $L__BB157_47: mov.f32 %f2876, 0f3F800000; sub.ftz.f32 %f2877, %f2876, %f528; mul.ftz.f32 %f3059, %f3058, %f2877; div.approx.ftz.f32 %f3058, %f3058, %f2877; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; $L__BB157_491: setp.eq.s32 %p436, %r2, 9; @%p436 bra $L__BB157_505; bra.uni $L__BB157_492; $L__BB157_505: mov.u32 %r597, 9; st.local.u32 [%rd205+-8], %r597; ld.local.u32 %r598, [%rd4+40]; or.b32 %r599, %r598, 4; st.local.u32 [%rd4+40], %r599; bra.uni $L__BB157_517; $L__BB157_492: setp.eq.s32 %p437, %r2, 13; @%p437 bra $L__BB157_504; bra.uni $L__BB157_493; $L__BB157_504: cvt.ftz.sat.f32.f32 %f2920, %f3058; cvt.ftz.sat.f32.f32 %f2921, %f3059; st.local.v2.f32 [%rd205+12], {%f2921, %f2920}; mov.u32 %r594, 13; st.local.u32 [%rd205+-8], %r594; ld.local.u32 %r595, [%rd4+40]; or.b32 %r596, %r595, 12; st.local.u32 [%rd4+40], %r596; bra.uni $L__BB157_517; $L__BB157_493: setp.eq.s32 %p438, %r2, 10; @%p438 bra $L__BB157_503; bra.uni $L__BB157_494; $L__BB157_503: mov.u64 %rd663, 0; st.local.u64 [%rd205+28], %rd663; cvt.ftz.sat.f32.f32 %f2918, %f3058; cvt.ftz.sat.f32.f32 %f2919, %f3059; st.local.v2.f32 [%rd205+12], {%f2919, %f2918}; mov.u32 %r591, 10; st.local.u32 [%rd205+-8], %r591; ld.local.u32 %r592, [%rd4+40]; or.b32 %r593, %r592, 12; st.local.u32 [%rd4+40], %r593; bra.uni $L__BB157_517; $L__BB157_441: setp.eq.s32 %p375, %r2, 25; st.local.v2.f32 [%rd160+12], {%f480, %f480}; st.local.f32 [%rd160+20], %f3054; cvt.ftz.sat.f32.f32 %f485, %f480; @%p375 bra $L__BB157_443; bra.uni $L__BB157_442; $L__BB157_443: st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r472, 13; st.local.u32 [%rd160+-8], %r472; ld.local.u32 %r473, [%rd4+40]; or.b32 %r474, %r473, 12; st.local.u32 [%rd4+40], %r474; bra.uni $L__BB157_445; $L__BB157_382: setp.eq.s32 %p339, %r89, 255; @%p339 bra $L__BB157_384; bra.uni $L__BB157_383; $L__BB157_384: mov.b32 %f3038, %r380; bra.uni $L__BB157_385; $L__BB157_364: cvt.u32.u64 %r404, %rd122; setp.eq.s32 %p326, %r404, -1; mov.u32 %r662, -1; @%p326 bra $L__BB157_366; ld.const.u64 %rd496, [kernel_params+2240]; cvta.to.global.u64 %rd497, %rd496; shl.b64 %rd498, %rd122, 2; add.s64 %rd499, %rd497, %rd498; ld.global.u32 %r662, [%rd499]; $L__BB157_366: setp.eq.s32 %p327, %r662, -1; @%p327 bra $L__BB157_368; mov.u64 %rd500, 0; { // callseq 919, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd230; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r110; st.param.b32 [param1+4], %r659; st.param.b32 [param1+8], %r660; st.param.b32 [param1+12], %r113; .param .b64 param2; st.param.b64 [param2+0], %rd500; .param .b64 param3; st.param.b64 [param3+0], %rd500; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f3038, [retval0+0]; } // callseq 919 bra.uni $L__BB157_385; $L__BB157_327: ld.local.f32 %f2492, [%rd4+280]; add.ftz.f32 %f2493, %f374, %f2492; ld.local.f32 %f2494, [%rd4+284]; add.ftz.f32 %f2495, %f375, %f2494; ld.local.f32 %f2496, [%rd4+288]; add.ftz.f32 %f2497, %f376, %f2496; st.local.f32 [%rd4+280], %f2493; st.local.f32 [%rd4+284], %f2495; st.local.f32 [%rd4+288], %f2497; ld.local.u32 %r68, [%rd4+248]; setp.lt.s32 %p296, %r68, 1; @%p296 bra $L__BB157_517; mov.u32 %r650, 0; $L__BB157_329: mul.wide.s32 %rd472, %r650, 80; add.s64 %rd473, %rd1, %rd472; add.s64 %rd118, %rd473, 316; ld.local.u32 %r364, [%rd473+316]; setp.eq.s32 %p297, %r364, 31; add.s32 %r650, %r650, 1; @%p297 bra $L__BB157_331; setp.lt.s32 %p298, %r650, %r68; @%p298 bra $L__BB157_329; bra.uni $L__BB157_517; $L__BB157_478: ld.local.f32 %f2802, [%rd4+116]; ld.local.f32 %f2803, [%rd4+120]; mul.ftz.f32 %f2804, %f2803, %f2803; fma.rn.ftz.f32 %f2805, %f2802, %f2802, %f2804; ld.local.f32 %f2806, [%rd4+124]; fma.rn.ftz.f32 %f2807, %f2806, %f2806, %f2805; sqrt.approx.ftz.f32 %f2808, %f2807; div.approx.ftz.f32 %f2809, %f2802, %f2808; div.approx.ftz.f32 %f2810, %f2803, %f2808; div.approx.ftz.f32 %f2811, %f2806, %f2808; st.local.f32 [%rd192+12], %f2809; st.local.f32 [%rd192+16], %f2810; st.local.f32 [%rd192+20], %f2811; mov.u32 %r547, 0; st.local.u32 [%rd192+32], %r547; $L__BB157_479: setp.eq.s32 %p418, %r2, 20; mov.f32 %f2812, 0f3A83126F; max.ftz.f32 %f2813, %f60, %f2812; mov.f32 %f2814, 0f3F800000; min.ftz.f32 %f513, %f2813, %f2814; max.ftz.f32 %f2815, %f63, %f2812; min.ftz.f32 %f514, %f2815, %f2814; @%p418 bra $L__BB157_481; bra.uni $L__BB157_480; $L__BB157_481: mov.u32 %r551, 20; st.local.u32 [%rd192+-8], %r551; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r552, [%rd4+40]; or.b32 %r553, %r552, 12; st.local.u32 [%rd4+40], %r553; bra.uni $L__BB157_517; $L__BB157_480: mov.u32 %r548, 30; st.local.u32 [%rd192+-8], %r548; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r549, [%rd4+40]; or.b32 %r550, %r549, 12; st.local.u32 [%rd4+40], %r550; bra.uni $L__BB157_517; $L__BB157_514: st.local.f32 [%rd229+12], %f60; mov.u32 %r630, 3; st.local.u32 [%rd229+-8], %r630; cvt.ftz.sat.f32.f32 %f2952, %f60; fma.rn.ftz.f32 %f2953, %f2952, 0f3F67750B, 0f40490FDB; rcp.approx.ftz.f32 %f2954, %f2953; st.local.f32 [%rd229+16], %f2954; mul.ftz.f32 %f2955, %f2952, %f2954; st.local.f32 [%rd229+20], %f2955; ld.local.u32 %r631, [%rd4+40]; or.b32 %r632, %r631, 12; st.local.u32 [%rd4+40], %r632; bra.uni $L__BB157_517; $L__BB157_473: st.local.u64 [%rd185+28], %rd621; st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r531, 23; st.local.u32 [%rd185+-8], %r531; ld.local.u32 %r532, [%rd4+40]; or.b32 %r533, %r532, 12; st.local.u32 [%rd4+40], %r533; bra.uni $L__BB157_517; $L__BB157_494: setp.eq.s32 %p439, %r2, 14; @%p439 bra $L__BB157_496; bra.uni $L__BB157_495; $L__BB157_496: ld.local.u32 %r131, [%rd4+252]; setp.lt.s32 %p440, %r131, 1; @%p440 bra $L__BB157_498; add.s32 %r582, %r131, -1; st.local.u32 [%rd4+252], %r582; ld.local.s32 %rd651, [%rd4+248]; cvt.s64.s32 %rd652, %r582; add.s64 %rd653, %rd651, %rd652; mul.lo.s64 %rd654, %rd653, 80; add.s64 %rd655, %rd1, %rd654; add.s64 %rd794, %rd655, 304; add.s64 %rd656, %rd230, %rd654; add.s64 %rd795, %rd656, 304; bra.uni $L__BB157_499; $L__BB157_457: ld.local.u32 %r508, [%rd4+248]; mov.u64 %rd776, 0; add.s32 %r509, %r508, -1; st.local.u32 [%rd4+248], %r509; add.s32 %r510, %r126, 1; st.local.u32 [%rd4+252], %r510; mov.u64 %rd777, %rd776; $L__BB157_458: setp.eq.s64 %p394, %rd777, 0; @%p394 bra $L__BB157_517; add.s64 %rd178, %rd774, 20; st.local.f32 [%rd774+20], %f55; st.local.f32 [%rd774+24], %f56; st.local.f32 [%rd774+28], %f57; st.local.u64 [%rd774+48], %rd777; mov.u32 %r511, 0; st.local.v2.f32 [%rd774+56], {%f2732, %f2732}; st.local.u32 [%rd774+64], %r511; mul.ftz.f32 %f2737, %f60, %f60; st.local.v2.f32 [%rd774+32], {%f2737, %f2737}; mov.f32 %f2738, 0f3727C5AC; max.ftz.f32 %f3056, %f63, %f2738; ld.local.u32 %r512, [%rd4+40]; and.b32 %r513, %r512, 1; setp.eq.b32 %p395, %r513, 1; mov.pred %p396, 0; xor.pred %p397, %p395, %p396; not.pred %p398, %p397; @%p398 bra $L__BB157_461; rcp.approx.ftz.f32 %f3056, %f3056; $L__BB157_461: st.local.f32 [%rd178+20], %f3056; ld.local.f32 %f2739, [%rd6]; ld.local.f32 %f2740, [%rd6+4]; ld.local.f32 %f2741, [%rd6+8]; st.local.f32 [%rd776], %f2739; st.local.f32 [%rd776+4], %f2740; st.local.f32 [%rd776+8], %f2741; ld.local.u64 %rd609, [%rd178+28]; st.u32 [%rd609+12], %r511; st.u32 [%rd609+16], %r511; st.u32 [%rd609+20], %r511; ld.local.u64 %rd610, [%rd178+28]; st.u32 [%rd610+36], %r511; ld.local.f32 %f2743, [%rd178+12]; mov.f32 %f2744, 0f38D1B717; max.ftz.f32 %f2745, %f2743, %f2744; mov.f32 %f2746, 0f3F800000; min.ftz.f32 %f2747, %f2745, %f2746; st.local.v2.f32 [%rd178+12], {%f2747, %f2747}; ld.local.f32 %f2748, [%rd178+20]; max.ftz.f32 %f2749, %f2732, %f2748; st.local.f32 [%rd178+20], %f2749; ld.local.u64 %rd611, [%rd178+28]; ld.f32 %f2750, [%rd611]; cvt.ftz.sat.f32.f32 %f2751, %f2750; ld.f32 %f2752, [%rd611+4]; cvt.ftz.sat.f32.f32 %f2753, %f2752; ld.f32 %f2754, [%rd611+8]; cvt.ftz.sat.f32.f32 %f2755, %f2754; st.f32 [%rd611], %f2751; st.f32 [%rd611+4], %f2753; st.f32 [%rd611+8], %f2755; mov.u32 %r515, 24; st.local.u32 [%rd178+-8], %r515; ld.local.u32 %r516, [%rd4+40]; or.b32 %r517, %r516, 1036; st.local.u32 [%rd4+40], %r517; bra.uni $L__BB157_517; $L__BB157_442: st.local.u64 [%rd160+28], %rd579; st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r469, 10; st.local.u32 [%rd160+-8], %r469; ld.local.u32 %r470, [%rd4+40]; or.b32 %r471, %r470, 12; st.local.u32 [%rd4+40], %r471; $L__BB157_445: setp.ne.s32 %p376, %r122, 0; ld.const.u32 %r478, [kernel_params+1568]; setp.eq.s32 %p377, %r478, 0; and.pred %p378, %p376, %p377; @%p378 bra $L__BB157_517; mov.f32 %f2713, 0f3F800000; sub.ftz.f32 %f2714, %f2713, %f3055; mov.f32 %f2715, 0f38D1B717; max.ftz.f32 %f2716, %f2715, %f2714; mul.ftz.f32 %f2717, %f470, %f2716; mul.ftz.f32 %f2718, %f471, %f2716; mul.ftz.f32 %f2719, %f472, %f2716; mov.f32 %f2720, 0f00000000; max.ftz.f32 %f487, %f2717, %f2720; max.ftz.f32 %f488, %f2718, %f2720; max.ftz.f32 %f489, %f2719, %f2720; add.ftz.f32 %f2721, %f487, %f488; add.ftz.f32 %f2722, %f489, %f2721; mul.ftz.f32 %f2723, %f2722, 0f3EAAAAAB; abs.ftz.f32 %f490, %f2723; setp.ltu.ftz.f32 %p379, %f490, 0f3727C5AC; mov.u64 %rd770, 0; mov.u64 %rd771, %rd770; @%p379 bra $L__BB157_451; ld.local.u32 %r124, [%rd4+252]; mov.u64 %rd770, 0; setp.eq.s32 %p380, %r124, 0; mov.u64 %rd768, %rd770; mov.u64 %rd769, %rd770; @%p380 bra $L__BB157_449; ld.local.u32 %r479, [%rd4+248]; mul.wide.s32 %rd585, %r479, 80; add.s64 %rd586, %rd1, %rd585; add.s64 %rd768, %rd586, 304; add.s64 %rd587, %rd230, %rd585; add.s64 %rd769, %rd587, 304; mov.b32 %r480, %f489; mov.b32 %r481, %f488; mov.b32 %r482, %f487; mov.u32 %r483, 0; st.local.v4.u32 [%rd586+304], {%r482, %r481, %r480, %r483}; add.s32 %r484, %r479, 1; st.local.u32 [%rd4+248], %r484; add.s32 %r485, %r124, -1; st.local.u32 [%rd4+252], %r485; $L__BB157_449: setp.eq.s64 %p381, %rd769, 0; mov.u64 %rd771, %rd770; @%p381 bra $L__BB157_451; st.local.f32 [%rd768+16], %f490; mov.u64 %rd770, %rd768; mov.u64 %rd771, %rd769; $L__BB157_451: setp.eq.s64 %p382, %rd771, 0; @%p382 bra $L__BB157_517; add.s64 %rd167, %rd770, 20; mov.u64 %rd590, 0; st.local.f32 [%rd770+20], %f55; st.local.f32 [%rd770+24], %f56; st.local.f32 [%rd770+28], %f57; mov.u32 %r486, 0; st.local.v2.f32 [%rd770+56], {%f2720, %f2720}; st.local.u32 [%rd770+64], %r486; st.local.u64 [%rd770+48], %rd590; setp.eq.s32 %p383, %r2, 28; @%p383 bra $L__BB157_456; bra.uni $L__BB157_453; $L__BB157_456: st.local.v2.f32 [%rd167+12], {%f2720, %f2720}; st.local.f32 [%rd167+20], %f3054; mov.u32 %r493, 21; st.local.u32 [%rd167+-8], %r493; ld.local.u32 %r494, [%rd4+40]; or.b32 %r495, %r494, 4; st.local.u32 [%rd4+40], %r495; bra.uni $L__BB157_517; $L__BB157_453: setp.eq.s32 %p384, %r2, 25; st.local.v2.f32 [%rd167+12], {%f480, %f480}; st.local.f32 [%rd167+20], %f3054; cvt.ftz.sat.f32.f32 %f491, %f480; @%p384 bra $L__BB157_455; bra.uni $L__BB157_454; $L__BB157_455: st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r490, 22; st.local.u32 [%rd167+-8], %r490; ld.local.u32 %r491, [%rd4+40]; or.b32 %r492, %r491, 12; st.local.u32 [%rd4+40], %r492; bra.uni $L__BB157_517; $L__BB157_454: st.local.u64 [%rd167+28], %rd590; st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r487, 23; st.local.u32 [%rd167+-8], %r487; ld.local.u32 %r488, [%rd4+40]; or.b32 %r489, %r488, 12; st.local.u32 [%rd4+40], %r489; bra.uni $L__BB157_517; $L__BB157_383: mul.wide.u32 %rd533, %r89, 4; add.s64 %rd534, %rd2, %rd533; ld.local.f32 %f3038, [%rd534]; bra.uni $L__BB157_385; $L__BB157_380: and.b32 %r432, %r11, 8; setp.eq.s32 %p336, %r432, 0; setp.ne.s32 %p337, %r110, 8; mov.f32 %f3038, 0f00000000; or.pred %p338, %p337, %p336; @%p338 bra $L__BB157_385; cvt.u32.u64 %r433, %rd122; ld.const.u64 %rd529, [kernel_params+2320]; cvta.to.global.u64 %rd530, %rd529; add.s32 %r434, %r433, %r113; mul.wide.s32 %rd531, %r434, 4; add.s64 %rd532, %rd530, %rd531; ld.global.f32 %f3038, [%rd532]; bra.uni $L__BB157_385; $L__BB157_368: and.b32 %r405, %r110, 56; setp.eq.s32 %p328, %r405, 0; @%p328 bra $L__BB157_373; and.b32 %r406, %r110, 24; setp.eq.s32 %p329, %r406, 0; @%p329 bra $L__BB157_371; ld.const.u64 %rd501, [kernel_params+2232]; cvta.to.global.u64 %rd502, %rd501; shl.b64 %rd503, %rd122, 4; add.s64 %rd504, %rd502, %rd503; ld.global.v4.u32 {%r407, %r408, %r409, %r410}, [%rd504]; ld.const.u64 %rd505, [kernel_params+2320]; cvta.to.global.u64 %rd506, %rd505; add.s32 %r414, %r407, %r113; mul.wide.u32 %rd507, %r414, 4; add.s64 %rd755, %rd506, %rd507; add.s32 %r415, %r408, %r113; mul.wide.u32 %rd508, %r415, 4; add.s64 %rd756, %rd506, %rd508; add.s32 %r416, %r409, %r113; mul.wide.u32 %rd509, %r416, 4; add.s64 %rd757, %rd506, %rd509; bra.uni $L__BB157_372; $L__BB157_495: mov.f32 %f2880, 0f38D1B717; max.ftz.f32 %f2881, %f3059, %f2880; mov.f32 %f2882, 0f3F800000; max.ftz.f32 %f2883, %f3058, %f2880; min.ftz.f32 %f2884, %f2883, %f2882; min.ftz.f32 %f2885, %f2881, %f2882; st.local.v2.f32 [%rd205+12], {%f2885, %f2884}; mov.u32 %r579, 16; st.local.u32 [%rd205+-8], %r579; ld.local.u32 %r580, [%rd4+40]; or.b32 %r581, %r580, 12; st.local.u32 [%rd4+40], %r581; bra.uni $L__BB157_517; $L__BB157_334: ld.local.u32 %r651, [%rd4+252]; $L__BB157_335: setp.eq.s32 %p301, %r651, 0; mov.pred %p462, -1; mov.u64 %rd751, 0; @%p301 bra $L__BB157_337; ld.local.u32 %r367, [%rd4+248]; mul.wide.s32 %rd475, %r367, 80; add.s64 %rd476, %rd1, %rd475; add.s64 %rd751, %rd476, 304; mov.b32 %r368, %f376; mov.b32 %r369, %f375; mov.b32 %r370, %f374; mov.u32 %r371, 31; st.local.v4.u32 [%rd476+304], {%r370, %r369, %r368, %r371}; add.s32 %r372, %r367, 1; st.local.u32 [%rd4+248], %r372; add.s32 %r373, %r651, -1; st.local.u32 [%rd4+252], %r373; mov.pred %p462, 0; $L__BB157_337: @%p462 bra $L__BB157_339; bra.uni $L__BB157_338; $L__BB157_339: @%p299 bra $L__BB157_517; mov.u32 %r374, 0; st.local.u32 [%rd4+252], %r374; bra.uni $L__BB157_517; $L__BB157_338: st.local.f32 [%rd751+16], %f377; ld.local.f32 %f2506, [%rd4]; ld.local.f32 %f2507, [%rd4+4]; ld.local.f32 %f2508, [%rd4+8]; st.local.f32 [%rd751+20], %f2506; st.local.f32 [%rd751+24], %f2507; st.local.f32 [%rd751+28], %f2508; bra.uni $L__BB157_517; $L__BB157_106: or.pred %p64, %p3, %p4; mov.f32 %f2990, %f677; mov.f32 %f2991, %f678; mov.f32 %f2992, %f123; @%p64 bra $L__BB157_107; bra.uni $L__BB157_112; $L__BB157_107: selp.f32 %f724, %f133, %f134, %p3; mov.f32 %f725, 0f3F800000; sub.ftz.f32 %f726, %f725, %f724; max.ftz.f32 %f728, %f726, %f716; sqrt.approx.ftz.f32 %f2988, %f728; max.ftz.f32 %f729, %f724, %f716; sqrt.approx.ftz.f32 %f2989, %f729; bra.uni $L__BB157_111; $L__BB157_331: ld.local.f32 %f2498, [%rd118+-12]; add.ftz.f32 %f2499, %f374, %f2498; ld.local.f32 %f2500, [%rd118+-8]; add.ftz.f32 %f2501, %f375, %f2500; ld.local.f32 %f2502, [%rd118+-4]; add.ftz.f32 %f2503, %f376, %f2502; st.local.f32 [%rd118+-12], %f2499; st.local.f32 [%rd118+-8], %f2501; st.local.f32 [%rd118+-4], %f2503; ld.local.f32 %f2504, [%rd118+4]; add.ftz.f32 %f2505, %f377, %f2504; st.local.f32 [%rd118+4], %f2505; bra.uni $L__BB157_517; $L__BB157_378: and.b32 %r428, %r110, 131; setp.eq.s32 %p334, %r428, 0; mov.f32 %f3038, 0f00000000; @%p334 bra $L__BB157_385; cvt.u32.u64 %r429, %rd122; setp.eq.s32 %p335, %r110, 128; selp.b32 %r430, %r429, 0, %p335; add.s32 %r431, %r113, %r430; ld.const.u64 %rd525, [kernel_params+2320]; cvta.to.global.u64 %rd526, %rd525; mul.wide.s32 %rd527, %r431, 4; add.s64 %rd528, %rd526, %rd527; ld.global.f32 %f3038, [%rd528]; bra.uni $L__BB157_385; $L__BB157_109: setp.gt.ftz.f32 %p68, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p68; selp.f32 %f2989, %f141, %f143, %p68; $L__BB157_111: mul.ftz.f32 %f746, %f677, %f2989; mul.ftz.f32 %f747, %f678, %f2989; mul.ftz.f32 %f748, %f123, %f2989; fma.rn.ftz.f32 %f2992, %f131, %f2988, %f748; fma.rn.ftz.f32 %f2991, %f130, %f2988, %f747; fma.rn.ftz.f32 %f2990, %f129, %f2988, %f746; $L__BB157_112: setp.eq.s32 %p70, %r171, 255; mov.f32 %f2993, 0f3F800000; mov.f32 %f2994, %f2993; mov.f32 %f2995, %f2993; @%p70 bra $L__BB157_114; mul.wide.u32 %rd283, %r171, 4; add.s64 %rd284, %rd2, %rd283; ld.local.f32 %f2993, [%rd284]; ld.local.f32 %f2994, [%rd284+4]; ld.local.f32 %f2995, [%rd284+8]; $L__BB157_114: setp.eq.s32 %p71, %r172, 255; mov.f32 %f2996, 0f3FB33333; @%p71 bra $L__BB157_116; mul.wide.u32 %rd285, %r172, 4; add.s64 %rd286, %rd2, %rd285; ld.local.f32 %f2996, [%rd286]; $L__BB157_116: setp.eq.s32 %p72, %r173, 255; mov.f32 %f2997, 0f00000000; @%p72 bra $L__BB157_118; mul.wide.u32 %rd287, %r173, 4; add.s64 %rd288, %rd2, %rd287; ld.local.f32 %f2997, [%rd288]; $L__BB157_118: ld.global.v4.u32 {%r174, %r175, %r176, %r177}, [%rd3+64]; setp.eq.s32 %p73, %r174, 255; @%p73 bra $L__BB157_120; bra.uni $L__BB157_119; $L__BB157_120: mov.b32 %f2998, %r175; mov.b32 %f2999, %r176; mov.b32 %f3000, %r177; bra.uni $L__BB157_121; $L__BB157_119: mul.wide.u32 %rd289, %r174, 4; add.s64 %rd290, %rd2, %rd289; ld.local.f32 %f2998, [%rd290]; ld.local.f32 %f2999, [%rd290+4]; ld.local.f32 %f3000, [%rd290+8]; $L__BB157_121: ld.local.f32 %f754, [%rd4+256]; mul.ftz.f32 %f177, %f2, %f754; ld.local.f32 %f755, [%rd4+260]; mul.ftz.f32 %f178, %f2, %f755; ld.local.f32 %f756, [%rd4+264]; mul.ftz.f32 %f179, %f2, %f756; mov.f32 %f757, 0f3F800000; sub.ftz.f32 %f758, %f757, %f63; mul.ftz.f32 %f759, %f63, %f2998; fma.rn.ftz.f32 %f180, %f758, %f2982, %f759; mul.ftz.f32 %f760, %f63, %f2999; fma.rn.ftz.f32 %f181, %f758, %f2983, %f760; mul.ftz.f32 %f761, %f63, %f3000; fma.rn.ftz.f32 %f182, %f758, %f2984, %f761; mul.ftz.f32 %f762, %f180, %f177; mul.ftz.f32 %f763, %f181, %f178; mul.ftz.f32 %f764, %f182, %f179; mul.ftz.f32 %f183, %f97, %f762; mul.ftz.f32 %f184, %f97, %f763; mul.ftz.f32 %f185, %f97, %f764; and.b32 %r178, %r141, 4096; setp.eq.s32 %p74, %r178, 0; selp.f32 %f186, %f2982, %f180, %p74; selp.f32 %f187, %f2983, %f181, %p74; selp.f32 %f188, %f2984, %f182, %p74; selp.f32 %f189, %f63, 0f00000000, %p74; add.ftz.f32 %f765, %f180, %f181; add.ftz.f32 %f766, %f182, %f765; mul.ftz.f32 %f767, %f766, 0f3EAAAAAB; abs.ftz.f32 %f768, %f767; setp.leu.ftz.f32 %p75, %f768, 0f3727C5AC; @%p75 bra $L__BB157_182; setp.le.ftz.f32 %p76, %f189, 0f3727C5AC; setp.gt.ftz.f32 %p77, %f97, 0f3727C5AC; and.pred %p78, %p76, %p77; @%p78 bra $L__BB157_175; bra.uni $L__BB157_123; $L__BB157_175: mul.ftz.f32 %f1391, %f186, %f177; mul.ftz.f32 %f1392, %f97, %f1391; mul.ftz.f32 %f1393, %f187, %f178; mul.ftz.f32 %f1394, %f97, %f1393; mul.ftz.f32 %f1395, %f188, %f179; mul.ftz.f32 %f1396, %f97, %f1395; mov.f32 %f1397, 0f00000000; max.ftz.f32 %f223, %f1392, %f1397; max.ftz.f32 %f224, %f1394, %f1397; max.ftz.f32 %f225, %f1396, %f1397; add.ftz.f32 %f1398, %f223, %f224; add.ftz.f32 %f1399, %f225, %f1398; mul.ftz.f32 %f1400, %f1399, 0f3EAAAAAB; abs.ftz.f32 %f226, %f1400; setp.ltu.ftz.f32 %p143, %f226, 0f3727C5AC; mov.u64 %rd708, 0; mov.u64 %rd709, %rd708; @%p143 bra $L__BB157_180; ld.local.u32 %r42, [%rd4+252]; mov.u64 %rd708, 0; setp.eq.s32 %p144, %r42, 0; mov.u64 %rd706, %rd708; mov.u64 %rd707, %rd708; @%p144 bra $L__BB157_178; ld.local.u32 %r226, [%rd4+248]; mul.wide.s32 %rd328, %r226, 80; add.s64 %rd329, %rd1, %rd328; add.s64 %rd706, %rd329, 304; add.s64 %rd330, %rd230, %rd328; add.s64 %rd707, %rd330, 304; mov.b32 %r227, %f225; mov.b32 %r228, %f224; mov.b32 %r229, %f223; mov.u32 %r230, 0; st.local.v4.u32 [%rd329+304], {%r229, %r228, %r227, %r230}; add.s32 %r231, %r226, 1; st.local.u32 [%rd4+248], %r231; add.s32 %r232, %r42, -1; st.local.u32 [%rd4+252], %r232; $L__BB157_178: setp.eq.s64 %p145, %rd707, 0; mov.u64 %rd709, %rd708; @%p145 bra $L__BB157_180; st.local.f32 [%rd706+16], %f226; mov.u64 %rd708, %rd706; mov.u64 %rd709, %rd707; $L__BB157_180: setp.eq.s64 %p146, %rd709, 0; @%p146 bra $L__BB157_182; st.local.f32 [%rd708+20], %f55; st.local.f32 [%rd708+24], %f56; st.local.f32 [%rd708+28], %f57; st.local.f32 [%rd708+32], %f71; mov.u32 %r233, 5; st.local.u32 [%rd708+12], %r233; mov.u32 %r234, 1; st.local.u32 [%rd708+36], %r234; ld.local.u32 %r235, [%rd4+40]; or.b32 %r236, %r235, 12; st.local.u32 [%rd4+40], %r236; bra.uni $L__BB157_182; $L__BB157_123: setp.leu.ftz.f32 %p79, %f189, 0f3727C5AC; @%p79 bra $L__BB157_182; ld.local.u32 %r29, [%rd4+252]; mov.u64 %rd693, 0; setp.eq.s32 %p80, %r29, 0; mov.u64 %rd691, %rd693; mov.u64 %rd692, %rd693; @%p80 bra $L__BB157_126; ld.local.u32 %r179, [%rd4+248]; mul.wide.s32 %rd293, %r179, 80; add.s64 %rd294, %rd1, %rd293; add.s64 %rd691, %rd294, 304; add.s64 %rd295, %rd230, %rd293; add.s64 %rd692, %rd295, 304; mov.b32 %r180, %f185; mov.b32 %r181, %f184; mov.b32 %r182, %f183; mov.u32 %r183, 0; st.local.v4.u32 [%rd294+304], {%r182, %r181, %r180, %r183}; add.s32 %r184, %r179, 1; st.local.u32 [%rd4+248], %r184; add.s32 %r185, %r29, -1; st.local.u32 [%rd4+252], %r185; $L__BB157_126: setp.eq.s64 %p81, %rd692, 0; @%p81 bra $L__BB157_128; add.ftz.f32 %f769, %f183, %f184; add.ftz.f32 %f770, %f185, %f769; mul.ftz.f32 %f771, %f770, 0f3EAAAAAB; abs.ftz.f32 %f772, %f771; st.local.f32 [%rd691+16], %f772; setp.ge.ftz.f32 %p82, %f772, 0f3727C5AC; selp.b64 %rd693, %rd692, 0, %p82; $L__BB157_128: setp.eq.s64 %p83, %rd693, 0; @%p83 bra $L__BB157_182; mul.ftz.f32 %f773, %f189, %f2993; add.s64 %rd14, %rd693, 32; st.f32 [%rd693+32], %f773; mul.ftz.f32 %f774, %f189, %f2994; st.f32 [%rd693+36], %f774; mul.ftz.f32 %f775, %f189, %f2995; st.f32 [%rd693+40], %f775; st.f32 [%rd693+44], %f180; st.f32 [%rd693+48], %f181; st.f32 [%rd693+52], %f182; st.f32 [%rd693+20], %f55; st.f32 [%rd693+24], %f56; st.f32 [%rd693+28], %f57; st.f32 [%rd693+56], %f71; mov.f32 %f3003, 0f00000000; max.ftz.f32 %f777, %f2997, %f3003; mov.f32 %f778, 0f3F666666; min.ftz.f32 %f779, %f777, %f778; st.f32 [%rd693+60], %f779; setp.eq.ftz.f32 %p84, %f71, 0f7F7FFFFF; mov.u32 %r643, 0; @%p84 bra $L__BB157_137; ld.f32 %f780, [%rd14+-32]; mov.u64 %rd696, 0; mov.f32 %f781, 0f00000000; max.ftz.f32 %f190, %f780, %f781; ld.f32 %f782, [%rd14+-28]; max.ftz.f32 %f191, %f782, %f781; ld.f32 %f783, [%rd14+-24]; max.ftz.f32 %f192, %f783, %f781; add.ftz.f32 %f784, %f190, %f191; add.ftz.f32 %f785, %f192, %f784; mul.ftz.f32 %f786, %f785, 0f3EAAAAAB; abs.ftz.f32 %f193, %f786; setp.ltu.ftz.f32 %p85, %f193, 0f3727C5AC; mov.u64 %rd697, %rd696; @%p85 bra $L__BB157_135; ld.local.u32 %r30, [%rd4+252]; mov.u64 %rd696, 0; setp.eq.s32 %p86, %r30, 0; mov.u64 %rd694, %rd696; mov.u64 %rd695, %rd696; @%p86 bra $L__BB157_133; ld.local.u32 %r187, [%rd4+248]; mul.wide.s32 %rd301, %r187, 80; add.s64 %rd302, %rd1, %rd301; add.s64 %rd694, %rd302, 304; add.s64 %rd303, %rd230, %rd301; add.s64 %rd695, %rd303, 304; mov.b32 %r188, %f192; mov.b32 %r189, %f191; mov.b32 %r190, %f190; mov.u32 %r191, 0; st.local.v4.u32 [%rd302+304], {%r190, %r189, %r188, %r191}; add.s32 %r192, %r187, 1; st.local.u32 [%rd4+248], %r192; add.s32 %r193, %r30, -1; st.local.u32 [%rd4+252], %r193; $L__BB157_133: setp.eq.s64 %p87, %rd695, 0; mov.u64 %rd697, %rd696; @%p87 bra $L__BB157_135; st.local.f32 [%rd694+16], %f193; mov.u64 %rd696, %rd694; mov.u64 %rd697, %rd695; $L__BB157_135: setp.eq.s64 %p88, %rd697, 0; @%p88 bra $L__BB157_137; ld.f32 %f787, [%rd14+-12]; ld.f32 %f788, [%rd14+-8]; ld.f32 %f789, [%rd14+-4]; st.local.f32 [%rd696+20], %f787; st.local.f32 [%rd696+24], %f788; st.local.f32 [%rd696+28], %f789; ld.f32 %f790, [%rd14+24]; st.local.f32 [%rd696+32], %f790; mov.u32 %r196, 5; st.local.u32 [%rd696+12], %r196; mov.u32 %r197, 8; st.local.u32 [%rd696+36], %r197; ld.local.f32 %f791, [%rd4+24]; ld.local.f32 %f792, [%rd4+28]; mul.ftz.f32 %f793, %f792, %f788; fma.rn.ftz.f32 %f794, %f791, %f787, %f793; ld.local.f32 %f795, [%rd4+32]; fma.rn.ftz.f32 %f796, %f795, %f789, %f794; mov.f32 %f797, 0f3F800000; sub.ftz.f32 %f798, %f797, %f796; max.ftz.f32 %f800, %f798, %f781; min.ftz.f32 %f801, %f800, %f797; mul.ftz.f32 %f802, %f801, %f801; mul.ftz.f32 %f803, %f802, %f802; mul.ftz.f32 %f804, %f801, %f803; mul.ftz.f32 %f805, %f790, %f804; ld.local.f32 %f806, [%rd696+16]; mul.ftz.f32 %f807, %f806, %f805; st.local.f32 [%rd696+16], %f807; mov.u32 %r643, 12; $L__BB157_137: ld.f32 %f809, [%rd14]; setp.geu.ftz.f32 %p89, %f809, 0f322BCC77; mov.u32 %r641, 3; mov.f32 %f3001, %f3003; @%p89 bra $L__BB157_139; ld.f32 %f3001, [%rd14+-32]; mov.u32 %r200, 0; st.u32 [%rd14+-32], %r200; st.u32 [%rd14], %r200; mov.u32 %r641, 2; $L__BB157_139: ld.f32 %f811, [%rd14+4]; setp.geu.ftz.f32 %p90, %f811, 0f322BCC77; mov.f32 %f3002, %f3003; @%p90 bra $L__BB157_141; ld.f32 %f3002, [%rd14+-28]; mov.u32 %r201, 0; st.u32 [%rd14+-28], %r201; st.u32 [%rd14+4], %r201; add.s32 %r641, %r641, -1; $L__BB157_141: ld.f32 %f813, [%rd14+8]; setp.geu.ftz.f32 %p91, %f813, 0f322BCC77; @%p91 bra $L__BB157_143; ld.f32 %f3003, [%rd14+-24]; mov.u32 %r202, 0; st.u32 [%rd14+-24], %r202; st.u32 [%rd14+8], %r202; add.s32 %r641, %r641, -1; $L__BB157_143: setp.gt.s32 %p92, %r641, 2; @%p92 bra $L__BB157_159; ld.f32 %f814, [%rd14+24]; setp.neu.ftz.f32 %p93, %f814, 0f7F7FFFFF; mov.f32 %f815, 0f00000000; max.ftz.f32 %f200, %f3002, %f815; max.ftz.f32 %f201, %f3001, %f815; add.ftz.f32 %f816, %f201, %f200; max.ftz.f32 %f202, %f3003, %f815; add.ftz.f32 %f817, %f202, %f816; mul.ftz.f32 %f818, %f817, 0f3EAAAAAB; abs.ftz.f32 %f203, %f818; @%p93 bra $L__BB157_152; bra.uni $L__BB157_145; $L__BB157_152: setp.ltu.ftz.f32 %p98, %f203, 0f3727C5AC; mov.u64 %rd704, 0; mov.u64 %rd705, %rd704; @%p98 bra $L__BB157_157; ld.local.u32 %r38, [%rd4+252]; mov.u64 %rd704, 0; setp.eq.s32 %p99, %r38, 0; mov.u64 %rd702, %rd704; mov.u64 %rd703, %rd704; @%p99 bra $L__BB157_155; ld.local.u32 %r212, [%rd4+248]; mul.wide.s32 %rd319, %r212, 80; add.s64 %rd320, %rd1, %rd319; add.s64 %rd702, %rd320, 304; add.s64 %rd321, %rd230, %rd319; add.s64 %rd703, %rd321, 304; mov.b32 %r213, %f202; mov.b32 %r214, %f200; mov.b32 %r215, %f201; mov.u32 %r216, 0; st.local.v4.u32 [%rd320+304], {%r215, %r214, %r213, %r216}; add.s32 %r217, %r212, 1; st.local.u32 [%rd4+248], %r217; add.s32 %r218, %r38, -1; st.local.u32 [%rd4+252], %r218; $L__BB157_155: setp.eq.s64 %p100, %rd703, 0; mov.u64 %rd705, %rd704; @%p100 bra $L__BB157_157; st.local.f32 [%rd702+16], %f203; mov.u64 %rd704, %rd702; mov.u64 %rd705, %rd703; $L__BB157_157: setp.eq.s64 %p101, %rd705, 0; @%p101 bra $L__BB157_159; ld.f32 %f822, [%rd14+-12]; ld.f32 %f823, [%rd14+-8]; ld.f32 %f824, [%rd14+-4]; st.local.f32 [%rd704+20], %f822; st.local.f32 [%rd704+24], %f823; st.local.f32 [%rd704+28], %f824; ld.f32 %f825, [%rd14+24]; st.local.f32 [%rd704+32], %f825; mov.u32 %r220, 5; st.local.u32 [%rd704+12], %r220; mov.u32 %r221, 2; st.local.u32 [%rd704+36], %r221; mov.u32 %r643, 12; bra.uni $L__BB157_159; $L__BB157_373: and.b32 %r419, %r110, 7; setp.eq.s32 %p330, %r419, 0; mov.f32 %f3038, 0f00000000; @%p330 bra $L__BB157_385; setp.eq.s32 %p331, %r110, 4; selp.b32 %r421, %r404, 0, %p331; add.s32 %r422, %r113, %r421; ld.const.u64 %rd513, [kernel_params+2320]; cvta.to.global.u64 %rd514, %rd513; mul.wide.s32 %rd515, %r422, 4; add.s64 %rd516, %rd514, %rd515; ld.global.f32 %f3038, [%rd516]; bra.uni $L__BB157_385; $L__BB157_498: ld.local.u32 %r583, [%rd4+248]; add.s32 %r584, %r583, -1; st.local.u32 [%rd4+248], %r584; add.s32 %r585, %r131, 1; st.local.u32 [%rd4+252], %r585; mov.u64 %rd795, %rd794; $L__BB157_499: st.local.u64 [%rd205+28], %rd795; setp.eq.s64 %p441, %rd795, 0; @%p441 bra $L__BB157_517; ld.local.f32 %f2886, [%rd7]; ld.local.f32 %f2887, [%rd7+4]; ld.local.f32 %f2888, [%rd7+8]; st.local.f32 [%rd794], %f2886; st.local.f32 [%rd794+4], %f2887; st.local.f32 [%rd794+8], %f2888; ld.local.u64 %rd659, [%rd205+28]; mov.u32 %r586, 0; st.u32 [%rd659+12], %r586; st.u32 [%rd659+16], %r586; st.u32 [%rd659+20], %r586; ld.local.u64 %rd660, [%rd205+28]; st.u32 [%rd660+36], %r586; ld.local.v2.f32 {%f2889, %f2890}, [%rd205+36]; setp.neu.ftz.f32 %p442, %f2889, 0f00000000; setp.neu.ftz.f32 %p443, %f2890, 0f00000000; or.pred %p444, %p442, %p443; ld.local.f32 %f2893, [%rd205+44]; setp.neu.ftz.f32 %p445, %f2893, 0f00000000; or.pred %p446, %p445, %p444; @%p446 bra $L__BB157_502; mov.f32 %f2895, 0f3F800000; st.local.v2.f32 [%rd205+36], {%f2895, %f2837}; st.local.u32 [%rd205+44], %r586; $L__BB157_502: mov.u32 %r588, 14; st.local.u32 [%rd205+-8], %r588; ld.local.v2.f32 {%f2896, %f2897}, [%rd205+12]; mov.f32 %f2900, 0f38D1B717; max.ftz.f32 %f2901, %f2896, %f2900; mov.f32 %f2902, 0f3F800000; max.ftz.f32 %f2903, %f2897, %f2900; min.ftz.f32 %f2904, %f2903, %f2902; min.ftz.f32 %f2905, %f2901, %f2902; st.local.v2.f32 [%rd205+12], {%f2905, %f2904}; ld.local.u64 %rd661, [%rd205+28]; ld.f32 %f2906, [%rd661]; cvt.ftz.sat.f32.f32 %f2907, %f2906; ld.f32 %f2908, [%rd661+4]; cvt.ftz.sat.f32.f32 %f2909, %f2908; ld.f32 %f2910, [%rd661+8]; cvt.ftz.sat.f32.f32 %f2911, %f2910; st.f32 [%rd661], %f2907; st.f32 [%rd661+4], %f2909; st.f32 [%rd661+8], %f2911; ld.local.u64 %rd662, [%rd205+28]; ld.f32 %f2912, [%rd662+12]; cvt.ftz.sat.f32.f32 %f2913, %f2912; ld.f32 %f2914, [%rd662+16]; cvt.ftz.sat.f32.f32 %f2915, %f2914; ld.f32 %f2916, [%rd662+20]; cvt.ftz.sat.f32.f32 %f2917, %f2916; st.f32 [%rd662+12], %f2913; st.f32 [%rd662+16], %f2915; st.f32 [%rd662+20], %f2917; ld.local.u32 %r589, [%rd4+40]; or.b32 %r590, %r589, 1036; st.local.u32 [%rd4+40], %r590; bra.uni $L__BB157_517; $L__BB157_371: mad.lo.s32 %r418, %r404, 3, %r113; ld.const.u64 %rd510, [kernel_params+2320]; cvta.to.global.u64 %rd511, %rd510; mul.wide.s32 %rd512, %r418, 4; add.s64 %rd755, %rd511, %rd512; add.s64 %rd756, %rd755, 4; add.s64 %rd757, %rd755, 8; $L__BB157_372: ld.local.f32 %f2509, [%rd4+56]; ld.global.f32 %f2510, [%rd756]; ld.local.f32 %f2511, [%rd4+60]; ld.global.f32 %f2512, [%rd757]; mul.ftz.f32 %f2513, %f2512, %f2511; fma.rn.ftz.f32 %f2514, %f2510, %f2509, %f2513; mov.f32 %f2515, 0f3F800000; sub.ftz.f32 %f2516, %f2515, %f2509; sub.ftz.f32 %f2517, %f2516, %f2511; ld.global.f32 %f2518, [%rd755]; fma.rn.ftz.f32 %f3038, %f2518, %f2517, %f2514; $L__BB157_385: mov.f32 %f2528, 0f00000000; max.ftz.f32 %f399, %f385, %f2528; max.ftz.f32 %f400, %f386, %f2528; add.ftz.f32 %f2529, %f400, %f399; max.ftz.f32 %f401, %f384, %f2528; add.ftz.f32 %f2530, %f401, %f2529; mul.ftz.f32 %f2531, %f2530, 0f3EAAAAAB; abs.ftz.f32 %f402, %f2531; setp.ltu.ftz.f32 %p340, %f402, 0f3727C5AC; mov.u64 %rd760, 0; mov.u64 %rd761, %rd760; @%p340 bra $L__BB157_390; ld.local.u32 %r118, [%rd4+252]; mov.u64 %rd760, 0; setp.eq.s32 %p341, %r118, 0; mov.u64 %rd758, %rd760; mov.u64 %rd759, %rd760; @%p341 bra $L__BB157_388; ld.local.u32 %r435, [%rd4+248]; mul.wide.s32 %rd539, %r435, 80; add.s64 %rd540, %rd1, %rd539; add.s64 %rd758, %rd540, 304; add.s64 %rd541, %rd230, %rd539; add.s64 %rd759, %rd541, 304; mov.b32 %r436, %f401; mov.b32 %r437, %f399; mov.b32 %r438, %f400; mov.u32 %r439, 0; st.local.v4.u32 [%rd540+304], {%r438, %r437, %r436, %r439}; add.s32 %r440, %r435, 1; st.local.u32 [%rd4+248], %r440; add.s32 %r441, %r118, -1; st.local.u32 [%rd4+252], %r441; $L__BB157_388: setp.eq.s64 %p342, %rd759, 0; mov.u64 %rd761, %rd760; @%p342 bra $L__BB157_390; st.local.f32 [%rd758+16], %f402; mov.u64 %rd760, %rd758; mov.u64 %rd761, %rd759; $L__BB157_390: add.s32 %r663, %r142, 4; setp.eq.s64 %p343, %rd761, 0; @%p343 bra $L__BB157_517; ld.local.u32 %r120, [%rd4+252]; setp.lt.s32 %p344, %r120, 1; @%p344 bra $L__BB157_393; add.s32 %r442, %r120, -1; st.local.u32 [%rd4+252], %r442; ld.local.s32 %rd544, [%rd4+248]; cvt.s64.s32 %rd545, %r442; add.s64 %rd546, %rd544, %rd545; mul.lo.s64 %rd547, %rd546, 80; add.s64 %rd548, %rd230, %rd547; add.s64 %rd763, %rd548, 304; add.s64 %rd549, %rd1, %rd547; add.s64 %rd762, %rd549, 304; bra.uni $L__BB157_394; $L__BB157_393: ld.local.u32 %r443, [%rd4+248]; mov.u64 %rd762, 0; add.s32 %r444, %r443, -1; st.local.u32 [%rd4+248], %r444; add.s32 %r445, %r120, 1; st.local.u32 [%rd4+252], %r445; mov.u64 %rd763, %rd762; $L__BB157_394: setp.eq.s64 %p345, %rd763, 0; @%p345 bra $L__BB157_517; setp.eq.s32 %p346, %r91, 255; @%p346 bra $L__BB157_397; bra.uni $L__BB157_396; $L__BB157_397: mov.b32 %f3039, %r382; bra.uni $L__BB157_398; $L__BB157_396: mul.wide.u32 %rd552, %r91, 4; add.s64 %rd553, %rd2, %rd552; ld.local.f32 %f3039, [%rd553]; $L__BB157_398: add.ftz.f32 %f2532, %f3038, 0fBF000000; add.ftz.f32 %f406, %f2532, %f2532; fma.rn.ftz.f32 %f2533, %f406, %f3039, 0f3F800000; mul.ftz.f32 %f407, %f60, %f2533; mul.ftz.f32 %f408, %f63, %f2533; setp.eq.s32 %p347, %r86, 255; @%p347 bra $L__BB157_400; bra.uni $L__BB157_399; $L__BB157_400: mov.b32 %f3040, %r376; bra.uni $L__BB157_401; $L__BB157_399: mul.wide.u32 %rd554, %r86, 4; add.s64 %rd555, %rd2, %rd554; ld.local.f32 %f3040, [%rd555]; $L__BB157_401: max.ftz.f32 %f2535, %f3040, %f2528; mov.f32 %f2536, 0f3F800000; min.ftz.f32 %f2537, %f2535, %f2536; sub.ftz.f32 %f412, %f2536, %f2537; add.s64 %rd153, %rd760, 20; st.local.f32 [%rd760+20], %f55; st.local.f32 [%rd760+24], %f56; st.local.f32 [%rd760+28], %f57; st.local.f32 [%rd760+44], %f407; st.local.v2.f32 [%rd760+48], {%f408, %f3036}; st.local.v2.f32 [%rd760+56], {%f3037, %f412}; st.local.u64 [%rd760+64], %rd763; shr.u32 %r446, %r144, 24; cvt.u16.u32 %rs15, %r446; setp.eq.s16 %p348, %rs15, 0; @%p348 bra $L__BB157_415; setp.eq.s16 %p349, %rs15, 1; @%p349 bra $L__BB157_405; setp.ne.s16 %p350, %rs15, 2; @%p350 bra $L__BB157_416; shr.u32 %r447, %r375, 24; mul.wide.u32 %rd556, %r447, 4; add.s64 %rd557, %rd2, %rd556; ld.local.f32 %f2538, [%rd557+8]; ld.local.f32 %f2539, [%rd557+4]; ld.local.f32 %f2540, [%rd557]; st.local.v2.f32 [%rd153+12], {%f2540, %f2539}; st.local.f32 [%rd153+20], %f2538; bra.uni $L__BB157_417; $L__BB157_415: shr.u32 %r449, %r144, 14; cvt.u64.u32 %rd567, %r449; and.b64 %rd568, %rd567, 1020; add.s64 %rd569, %rd2, %rd568; ld.local.f32 %f2581, [%rd569]; lg2.approx.ftz.f32 %f2582, %f2581; mul.ftz.f32 %f2583, %f2582, 0f3F317218; ld.local.f32 %f2584, [%rd569+4]; lg2.approx.ftz.f32 %f2585, %f2584; mul.ftz.f32 %f2586, %f2585, 0f3F317218; ld.local.f32 %f2587, [%rd569+8]; lg2.approx.ftz.f32 %f2588, %f2587; mul.ftz.f32 %f2589, %f2588, 0f3F317218; fma.rn.ftz.f32 %f2590, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2591, %f408, %f2590, 0fC12BAE14; fma.rn.ftz.f32 %f2592, %f408, %f2591, 0f40220C4A; fma.rn.ftz.f32 %f2593, %f408, %f2592, 0fBE5C28F6; fma.rn.ftz.f32 %f2594, %f408, %f2593, 0f40BF020C; div.approx.ftz.f32 %f2595, %f2583, %f2594; div.approx.ftz.f32 %f2596, %f2586, %f2594; div.approx.ftz.f32 %f2597, %f2589, %f2594; mul.ftz.f32 %f2598, %f2597, %f2597; mul.ftz.f32 %f2599, %f2596, %f2596; mul.ftz.f32 %f2600, %f2595, %f2595; st.local.v2.f32 [%rd153+12], {%f2600, %f2599}; st.local.f32 [%rd153+20], %f2598; bra.uni $L__BB157_417; $L__BB157_405: setp.eq.s32 %p351, %r87, 255; @%p351 bra $L__BB157_407; bra.uni $L__BB157_406; $L__BB157_407: mov.b32 %f3041, %r377; bra.uni $L__BB157_408; $L__BB157_416: mov.f32 %f2601, 0f3F172384; mov.f32 %f2602, 0f3E8D729D; st.local.v2.f32 [%rd153+12], {%f2602, %f2601}; mov.u32 %r450, 1069964113; st.local.u32 [%rd153+20], %r450; bra.uni $L__BB157_417; $L__BB157_406: mul.wide.u32 %rd558, %r87, 4; add.s64 %rd559, %rd2, %rd558; ld.local.f32 %f3041, [%rd559]; $L__BB157_408: setp.eq.s32 %p352, %r88, 255; @%p352 bra $L__BB157_410; bra.uni $L__BB157_409; $L__BB157_410: mov.b32 %f3042, %r378; bra.uni $L__BB157_411; $L__BB157_409: mul.wide.u32 %rd560, %r88, 4; add.s64 %rd561, %rd2, %rd560; ld.local.f32 %f3042, [%rd561]; $L__BB157_411: setp.eq.s32 %p353, %r90, 255; @%p353 bra $L__BB157_413; bra.uni $L__BB157_412; $L__BB157_413: mov.b32 %f3043, %r381; bra.uni $L__BB157_414; $L__BB157_412: mul.wide.u32 %rd562, %r90, 4; add.s64 %rd563, %rd2, %rd562; ld.local.f32 %f3043, [%rd563]; $L__BB157_414: max.ftz.f32 %f2542, %f3043, %f2528; min.ftz.f32 %f2544, %f2542, %f2536; fma.rn.ftz.f32 %f2545, %f406, %f2544, 0f3F800000; mul.ftz.f32 %f2546, %f3041, %f2545; sub.ftz.f32 %f2547, %f2536, %f2546; mov.f32 %f2548, 0f38D1B717; max.ftz.f32 %f2549, %f2547, %f2548; lg2.approx.ftz.f32 %f2550, %f2549; mul.ftz.f32 %f2551, %f2550, 0fBF317218; sub.ftz.f32 %f2552, %f2536, %f3042; mul.ftz.f32 %f2553, %f2552, %f2551; mul.ftz.f32 %f2554, %f3042, %f2551; mul.ftz.f32 %f2555, %f2554, 0f3EAF9DB2; mul.ftz.f32 %f2556, %f2554, 0f3F3BA5E3; mul.ftz.f32 %f2557, %f2554, 0f3FF645A2; fma.rn.ftz.f32 %f2558, %f2553, 0f3F018937, %f2555; fma.rn.ftz.f32 %f2559, %f2553, 0f3F574BC7, %f2556; fma.rn.ftz.f32 %f2560, %f2553, 0f3FD39581, %f2557; shl.b32 %r448, %r379, 2; cvt.u64.u32 %rd564, %r448; and.b64 %rd565, %rd564, 1020; add.s64 %rd566, %rd2, %rd565; ld.local.f32 %f2561, [%rd566]; lg2.approx.ftz.f32 %f2562, %f2561; mul.ftz.f32 %f2563, %f2562, 0f3F317218; ld.local.f32 %f2564, [%rd566+4]; lg2.approx.ftz.f32 %f2565, %f2564; mul.ftz.f32 %f2566, %f2565, 0f3F317218; ld.local.f32 %f2567, [%rd566+8]; lg2.approx.ftz.f32 %f2568, %f2567; mul.ftz.f32 %f2569, %f2568, 0f3F317218; fma.rn.ftz.f32 %f2570, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2571, %f408, %f2570, 0fC12BAE14; fma.rn.ftz.f32 %f2572, %f408, %f2571, 0f40220C4A; fma.rn.ftz.f32 %f2573, %f408, %f2572, 0fBE5C28F6; fma.rn.ftz.f32 %f2574, %f408, %f2573, 0f40BF020C; div.approx.ftz.f32 %f2575, %f2563, %f2574; div.approx.ftz.f32 %f2576, %f2566, %f2574; div.approx.ftz.f32 %f2577, %f2569, %f2574; fma.rn.ftz.f32 %f2578, %f2577, %f2577, %f2560; fma.rn.ftz.f32 %f2579, %f2576, %f2576, %f2559; fma.rn.ftz.f32 %f2580, %f2575, %f2575, %f2558; st.local.v2.f32 [%rd153+12], {%f2580, %f2579}; st.local.f32 [%rd153+20], %f2578; $L__BB157_417: mov.u32 %r451, 29; st.local.u32 [%rd153+-8], %r451; mov.f32 %f2603, 0f3A83126F; max.ftz.f32 %f2604, %f407, %f2603; min.ftz.f32 %f2606, %f2604, %f2536; max.ftz.f32 %f2607, %f408, %f2603; min.ftz.f32 %f2608, %f2607, %f2536; mul.ftz.f32 %f2609, %f2606, %f412; max.ftz.f32 %f2610, %f2609, %f2603; min.ftz.f32 %f2611, %f2610, %f2536; mul.ftz.f32 %f2612, %f2606, %f2606; mul.ftz.f32 %f2613, %f2612, 0f3F4FDF3B; fma.rn.ftz.f32 %f2614, %f2606, 0f3F39DB23, %f2613; mul.ftz.f32 %f2615, %f2612, %f2612; mul.ftz.f32 %f2616, %f2606, %f2615; mul.ftz.f32 %f2617, %f2616, %f2616; mul.ftz.f32 %f2618, %f2617, %f2617; fma.rn.ftz.f32 %f2619, %f2618, 0f406CCCCD, %f2614; mul.ftz.f32 %f2620, %f2619, %f2619; st.local.f32 [%rd153+24], %f2620; mul.ftz.f32 %f2621, %f2608, %f2608; mul.ftz.f32 %f2622, %f2621, 0f3F98D4FE; fma.rn.ftz.f32 %f2623, %f2608, 0f3E87AE14, %f2622; mul.ftz.f32 %f2624, %f2621, %f2621; mul.ftz.f32 %f2625, %f2608, %f2624; mul.ftz.f32 %f2626, %f2625, %f2625; mul.ftz.f32 %f2627, %f2608, %f2626; mul.ftz.f32 %f2628, %f2627, %f2627; fma.rn.ftz.f32 %f2629, %f2628, 0f40ABE76D, %f2623; mul.ftz.f32 %f2630, %f2629, 0f3F206C99; st.local.f32 [%rd153+28], %f2630; mul.ftz.f32 %f2631, %f2611, %f2611; mul.ftz.f32 %f2632, %f2631, 0f3F4FDF3B; fma.rn.ftz.f32 %f2633, %f2611, 0f3F39DB23, %f2632; mul.ftz.f32 %f2634, %f2631, %f2631; mul.ftz.f32 %f2635, %f2611, %f2634; mul.ftz.f32 %f2636, %f2635, %f2635; mul.ftz.f32 %f2637, %f2636, %f2636; fma.rn.ftz.f32 %f2638, %f2637, 0f406CCCCD, %f2633; mul.ftz.f32 %f2639, %f2638, %f2638; st.local.f32 [%rd153+40], %f2639; ld.local.f32 %f3044, [%rd4+104]; ld.local.f32 %f3045, [%rd4+108]; mul.ftz.f32 %f2640, %f3045, %f3045; fma.rn.ftz.f32 %f2641, %f3044, %f3044, %f2640; ld.local.f32 %f3046, [%rd4+112]; fma.rn.ftz.f32 %f2642, %f3046, %f3046, %f2641; sqrt.approx.ftz.f32 %f425, %f2642; setp.neu.ftz.f32 %p354, %f425, 0f00000000; @%p354 bra $L__BB157_418; bra.uni $L__BB157_419; $L__BB157_418: rcp.approx.ftz.f32 %f2643, %f425; mul.ftz.f32 %f3046, %f3046, %f2643; mul.ftz.f32 %f3045, %f3045, %f2643; mul.ftz.f32 %f3044, %f3044, %f2643; $L__BB157_419: ld.local.f32 %f2644, [%rd4+32]; mul.ftz.f32 %f2645, %f3045, %f2644; ld.local.f32 %f2646, [%rd4+28]; mul.ftz.f32 %f2647, %f3046, %f2646; sub.ftz.f32 %f3047, %f2645, %f2647; ld.local.f32 %f2648, [%rd4+24]; mul.ftz.f32 %f2649, %f3046, %f2648; mul.ftz.f32 %f2650, %f3044, %f2644; sub.ftz.f32 %f3048, %f2649, %f2650; mul.ftz.f32 %f2651, %f3044, %f2646; mul.ftz.f32 %f2652, %f3045, %f2648; sub.ftz.f32 %f3049, %f2651, %f2652; mul.ftz.f32 %f2653, %f3048, %f3048; fma.rn.ftz.f32 %f2654, %f3047, %f3047, %f2653; fma.rn.ftz.f32 %f2655, %f3049, %f3049, %f2654; sqrt.approx.ftz.f32 %f441, %f2655; setp.neu.ftz.f32 %p355, %f441, 0f00000000; @%p355 bra $L__BB157_420; bra.uni $L__BB157_421; $L__BB157_420: rcp.approx.ftz.f32 %f2656, %f441; mul.ftz.f32 %f3049, %f3049, %f2656; mul.ftz.f32 %f3048, %f3048, %f2656; mul.ftz.f32 %f3047, %f3047, %f2656; $L__BB157_421: mul.ftz.f32 %f2657, %f3045, %f3049; mul.ftz.f32 %f2658, %f3046, %f3048; sub.ftz.f32 %f3050, %f2657, %f2658; mul.ftz.f32 %f2659, %f3046, %f3047; mul.ftz.f32 %f2660, %f3044, %f3049; sub.ftz.f32 %f3051, %f2659, %f2660; mul.ftz.f32 %f2661, %f3044, %f3048; mul.ftz.f32 %f2662, %f3045, %f3047; sub.ftz.f32 %f3052, %f2661, %f2662; mul.ftz.f32 %f2663, %f3051, %f3051; fma.rn.ftz.f32 %f2664, %f3050, %f3050, %f2663; fma.rn.ftz.f32 %f2665, %f3052, %f3052, %f2664; sqrt.approx.ftz.f32 %f457, %f2665; setp.neu.ftz.f32 %p356, %f457, 0f00000000; @%p356 bra $L__BB157_422; bra.uni $L__BB157_423; $L__BB157_422: rcp.approx.ftz.f32 %f2666, %f457; mul.ftz.f32 %f3052, %f2666, %f3052; mul.ftz.f32 %f3051, %f2666, %f3051; mul.ftz.f32 %f3050, %f2666, %f3050; $L__BB157_423: ld.local.u8 %rs16, [%rd4+52]; and.b16 %rs17, %rs16, 4; setp.eq.s16 %p357, %rs17, 0; @%p357 bra $L__BB157_425; bra.uni $L__BB157_424; $L__BB157_425: ld.local.f32 %f2668, [%rd4+16]; mul.ftz.f32 %f2669, %f3046, %f2668; ld.local.f32 %f2670, [%rd4+20]; mul.ftz.f32 %f2671, %f3045, %f2670; sub.ftz.f32 %f2672, %f2669, %f2671; mul.ftz.f32 %f2673, %f3044, %f2670; ld.local.f32 %f2674, [%rd4+12]; mul.ftz.f32 %f2675, %f3046, %f2674; sub.ftz.f32 %f2676, %f2673, %f2675; mul.ftz.f32 %f2677, %f3045, %f2674; mul.ftz.f32 %f2678, %f3044, %f2668; sub.ftz.f32 %f2679, %f2677, %f2678; mul.ftz.f32 %f2680, %f3051, %f2676; fma.rn.ftz.f32 %f2681, %f3050, %f2672, %f2680; fma.rn.ftz.f32 %f3053, %f3052, %f2679, %f2681; bra.uni $L__BB157_426; $L__BB157_424: ld.local.f32 %f2667, [%rd4+60]; neg.ftz.f32 %f3053, %f2667; $L__BB157_426: st.local.v4.f32 [%rd762], {%f3047, %f3048, %f3049, %f3053}; ld.local.u32 %r452, [%rd4+40]; or.b32 %r453, %r452, 1036; st.local.u32 [%rd4+40], %r453; bra.uni $L__BB157_517; $L__BB157_145: setp.ltu.ftz.f32 %p94, %f203, 0f3727C5AC; mov.u64 %rd700, 0; mov.u64 %rd701, %rd700; @%p94 bra $L__BB157_150; ld.local.u32 %r37, [%rd4+252]; mov.u64 %rd700, 0; setp.eq.s32 %p95, %r37, 0; mov.u64 %rd698, %rd700; mov.u64 %rd699, %rd700; @%p95 bra $L__BB157_148; ld.local.u32 %r203, [%rd4+248]; mul.wide.s32 %rd310, %r203, 80; add.s64 %rd311, %rd1, %rd310; add.s64 %rd698, %rd311, 304; add.s64 %rd312, %rd230, %rd310; add.s64 %rd699, %rd312, 304; mov.b32 %r204, %f202; mov.b32 %r205, %f200; mov.b32 %r206, %f201; mov.u32 %r207, 0; st.local.v4.u32 [%rd311+304], {%r206, %r205, %r204, %r207}; add.s32 %r208, %r203, 1; st.local.u32 [%rd4+248], %r208; add.s32 %r209, %r37, -1; st.local.u32 [%rd4+252], %r209; $L__BB157_148: setp.eq.s64 %p96, %rd699, 0; mov.u64 %rd701, %rd700; @%p96 bra $L__BB157_150; st.local.f32 [%rd698+16], %f203; mov.u64 %rd700, %rd698; mov.u64 %rd701, %rd699; $L__BB157_150: setp.eq.s64 %p97, %rd701, 0; @%p97 bra $L__BB157_159; ld.f32 %f819, [%rd14+-12]; ld.f32 %f820, [%rd14+-8]; ld.f32 %f821, [%rd14+-4]; st.local.f32 [%rd700+20], %f819; st.local.f32 [%rd700+24], %f820; st.local.f32 [%rd700+28], %f821; mov.u32 %r211, 2; st.local.u32 [%rd700+12], %r211; mov.u32 %r643, 12; $L__BB157_159: st.u32 [%rd14+-20], %r149; setp.gt.s32 %p102, %r641, 0; @%p102 bra $L__BB157_161; bra.uni $L__BB157_160; $L__BB157_161: ld.f32 %f826, [%rd14+-28]; ld.f32 %f827, [%rd14+-32]; add.ftz.f32 %f828, %f827, %f826; ld.f32 %f829, [%rd14+-24]; add.ftz.f32 %f830, %f829, %f828; mul.ftz.f32 %f831, %f830, 0f3EAAAAAB; abs.ftz.f32 %f832, %f831; cvt.rn.f32.s32 %f833, %r641; mul.ftz.f32 %f834, %f832, %f833; st.f32 [%rd14+-16], %f834; or.b32 %r223, %r149, 2; setp.eq.s32 %p103, %r223, 34; @%p103 bra $L__BB157_172; bra.uni $L__BB157_162; $L__BB157_172: ld.f32 %f1388, [%rd14]; mul.ftz.f32 %f3007, %f1388, 0f3DA2F983; ld.f32 %f1389, [%rd14+4]; mul.ftz.f32 %f3008, %f1389, 0f3DA2F983; ld.f32 %f1390, [%rd14+8]; mul.ftz.f32 %f3009, %f1390, 0f3DA2F983; bra.uni $L__BB157_173; $L__BB157_160: mov.u32 %r222, 0; st.u32 [%rd14+-16], %r222; bra.uni $L__BB157_174; $L__BB157_162: mov.f32 %f836, 0f3F8147AE; max.ftz.f32 %f837, %f2996, %f836; mov.f32 %f838, 0f40733333; min.ftz.f32 %f839, %f837, %f838; rcp.approx.ftz.f32 %f840, %f839; mov.f32 %f841, 0f3F800000; fma.rn.ftz.f32 %f842, %f840, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f843, %f840, %f842, 0f3F2B020C; fma.rn.ftz.f32 %f844, %f839, 0f3D8240B8, %f843; add.ftz.f32 %f845, %f844, 0f3F800000; mul.ftz.f32 %f846, %f845, 0f3FAAAAAB; sub.ftz.f32 %f847, %f841, %f844; div.approx.ftz.f32 %f204, %f846, %f847; ld.f32 %f205, [%rd14+12]; setp.lt.ftz.f32 %p104, %f205, 0f38D1B717; mov.f32 %f3004, 0f00000000; @%p104 bra $L__BB157_165; setp.ge.ftz.f32 %p105, %f205, 0f3F7EB852; mov.f32 %f3004, 0f3F7FFFEF; @%p105 bra $L__BB157_165; sub.ftz.f32 %f850, %f841, 0f3F000000; mul.ftz.f32 %f851, %f850, 0f40400000; sqrt.approx.ftz.f32 %f852, %f851; mul.ftz.f32 %f853, %f204, %f852; mul.ftz.f32 %f854, %f853, 0fBFB8AA3B; ex2.approx.ftz.f32 %f855, %f854; add.ftz.f32 %f856, %f855, 0f3F800000; mul.ftz.f32 %f857, %f856, 0f3E800000; mul.ftz.f32 %f858, %f852, 0fBFB8AA3B; ex2.approx.ftz.f32 %f859, %f858; mul.ftz.f32 %f860, %f859, %f857; setp.lt.ftz.f32 %p106, %f860, %f205; selp.f32 %f861, 0f3F800000, 0f3F000000, %p106; selp.f32 %f862, 0f3F000000, 0f00000000, %p106; selp.f32 %f863, 0f3F400000, 0f3E800000, %p106; sub.ftz.f32 %f864, %f841, %f863; mul.ftz.f32 %f865, %f864, 0f40400000; sqrt.approx.ftz.f32 %f866, %f865; mul.ftz.f32 %f867, %f863, 0f3F000000; mul.ftz.f32 %f868, %f204, %f866; mul.ftz.f32 %f869, %f868, 0fBFB8AA3B; ex2.approx.ftz.f32 %f870, %f869; add.ftz.f32 %f871, %f870, 0f3F800000; mul.ftz.f32 %f872, %f871, %f867; mul.ftz.f32 %f873, %f866, 0fBFB8AA3B; ex2.approx.ftz.f32 %f874, %f873; mul.ftz.f32 %f875, %f874, %f872; setp.lt.ftz.f32 %p107, %f875, %f205; selp.f32 %f876, %f861, %f863, %p107; selp.f32 %f877, %f863, %f862, %p107; add.ftz.f32 %f878, %f876, %f877; mul.ftz.f32 %f879, %f878, 0f3F000000; sub.ftz.f32 %f880, %f841, %f879; mul.ftz.f32 %f881, %f880, 0f40400000; sqrt.approx.ftz.f32 %f882, %f881; mul.ftz.f32 %f883, %f879, 0f3F000000; mul.ftz.f32 %f884, %f204, %f882; mul.ftz.f32 %f885, %f884, 0fBFB8AA3B; ex2.approx.ftz.f32 %f886, %f885; add.ftz.f32 %f887, %f886, 0f3F800000; mul.ftz.f32 %f888, %f887, %f883; mul.ftz.f32 %f889, %f882, 0fBFB8AA3B; ex2.approx.ftz.f32 %f890, %f889; mul.ftz.f32 %f891, %f890, %f888; setp.lt.ftz.f32 %p108, %f891, %f205; selp.f32 %f892, %f876, %f879, %p108; selp.f32 %f893, %f879, %f877, %p108; add.ftz.f32 %f894, %f892, %f893; mul.ftz.f32 %f895, %f894, 0f3F000000; sub.ftz.f32 %f896, %f841, %f895; mul.ftz.f32 %f897, %f896, 0f40400000; sqrt.approx.ftz.f32 %f898, %f897; mul.ftz.f32 %f899, %f895, 0f3F000000; mul.ftz.f32 %f900, %f204, %f898; mul.ftz.f32 %f901, %f900, 0fBFB8AA3B; ex2.approx.ftz.f32 %f902, %f901; add.ftz.f32 %f903, %f902, 0f3F800000; mul.ftz.f32 %f904, %f903, %f899; mul.ftz.f32 %f905, %f898, 0fBFB8AA3B; ex2.approx.ftz.f32 %f906, %f905; mul.ftz.f32 %f907, %f906, %f904; setp.lt.ftz.f32 %p109, %f907, %f205; selp.f32 %f908, %f892, %f895, %p109; selp.f32 %f909, %f895, %f893, %p109; add.ftz.f32 %f910, %f908, %f909; mul.ftz.f32 %f911, %f910, 0f3F000000; sub.ftz.f32 %f912, %f841, %f911; mul.ftz.f32 %f913, %f912, 0f40400000; sqrt.approx.ftz.f32 %f914, %f913; mul.ftz.f32 %f915, %f911, 0f3F000000; mul.ftz.f32 %f916, %f204, %f914; mul.ftz.f32 %f917, %f916, 0fBFB8AA3B; ex2.approx.ftz.f32 %f918, %f917; add.ftz.f32 %f919, %f918, 0f3F800000; mul.ftz.f32 %f920, %f919, %f915; mul.ftz.f32 %f921, %f914, 0fBFB8AA3B; ex2.approx.ftz.f32 %f922, %f921; mul.ftz.f32 %f923, %f922, %f920; setp.lt.ftz.f32 %p110, %f923, %f205; selp.f32 %f924, %f908, %f911, %p110; selp.f32 %f925, %f911, %f909, %p110; add.ftz.f32 %f926, %f924, %f925; mul.ftz.f32 %f927, %f926, 0f3F000000; sub.ftz.f32 %f928, %f841, %f927; mul.ftz.f32 %f929, %f928, 0f40400000; sqrt.approx.ftz.f32 %f930, %f929; mul.ftz.f32 %f931, %f927, 0f3F000000; mul.ftz.f32 %f932, %f204, %f930; mul.ftz.f32 %f933, %f932, 0fBFB8AA3B; ex2.approx.ftz.f32 %f934, %f933; add.ftz.f32 %f935, %f934, 0f3F800000; mul.ftz.f32 %f936, %f935, %f931; mul.ftz.f32 %f937, %f930, 0fBFB8AA3B; ex2.approx.ftz.f32 %f938, %f937; mul.ftz.f32 %f939, %f938, %f936; setp.lt.ftz.f32 %p111, %f939, %f205; selp.f32 %f940, %f924, %f927, %p111; selp.f32 %f941, %f927, %f925, %p111; add.ftz.f32 %f942, %f940, %f941; mul.ftz.f32 %f943, %f942, 0f3F000000; sub.ftz.f32 %f944, %f841, %f943; mul.ftz.f32 %f945, %f944, 0f40400000; sqrt.approx.ftz.f32 %f946, %f945; mul.ftz.f32 %f947, %f943, 0f3F000000; mul.ftz.f32 %f948, %f204, %f946; mul.ftz.f32 %f949, %f948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f950, %f949; add.ftz.f32 %f951, %f950, 0f3F800000; mul.ftz.f32 %f952, %f951, %f947; mul.ftz.f32 %f953, %f946, 0fBFB8AA3B; ex2.approx.ftz.f32 %f954, %f953; mul.ftz.f32 %f955, %f954, %f952; setp.lt.ftz.f32 %p112, %f955, %f205; selp.f32 %f956, %f940, %f943, %p112; selp.f32 %f957, %f943, %f941, %p112; add.ftz.f32 %f958, %f956, %f957; mul.ftz.f32 %f959, %f958, 0f3F000000; sub.ftz.f32 %f960, %f841, %f959; mul.ftz.f32 %f961, %f960, 0f40400000; sqrt.approx.ftz.f32 %f962, %f961; mul.ftz.f32 %f963, %f959, 0f3F000000; mul.ftz.f32 %f964, %f204, %f962; mul.ftz.f32 %f965, %f964, 0fBFB8AA3B; ex2.approx.ftz.f32 %f966, %f965; add.ftz.f32 %f967, %f966, 0f3F800000; mul.ftz.f32 %f968, %f967, %f963; mul.ftz.f32 %f969, %f962, 0fBFB8AA3B; ex2.approx.ftz.f32 %f970, %f969; mul.ftz.f32 %f971, %f970, %f968; setp.lt.ftz.f32 %p113, %f971, %f205; selp.f32 %f972, %f956, %f959, %p113; selp.f32 %f973, %f959, %f957, %p113; add.ftz.f32 %f974, %f972, %f973; mul.ftz.f32 %f975, %f974, 0f3F000000; sub.ftz.f32 %f976, %f841, %f975; mul.ftz.f32 %f977, %f976, 0f40400000; sqrt.approx.ftz.f32 %f978, %f977; mul.ftz.f32 %f979, %f975, 0f3F000000; mul.ftz.f32 %f980, %f204, %f978; mul.ftz.f32 %f981, %f980, 0fBFB8AA3B; ex2.approx.ftz.f32 %f982, %f981; add.ftz.f32 %f983, %f982, 0f3F800000; mul.ftz.f32 %f984, %f983, %f979; mul.ftz.f32 %f985, %f978, 0fBFB8AA3B; ex2.approx.ftz.f32 %f986, %f985; mul.ftz.f32 %f987, %f986, %f984; setp.lt.ftz.f32 %p114, %f987, %f205; selp.f32 %f988, %f972, %f975, %p114; selp.f32 %f989, %f975, %f973, %p114; add.ftz.f32 %f990, %f988, %f989; mul.ftz.f32 %f991, %f990, 0f3F000000; sub.ftz.f32 %f992, %f841, %f991; mul.ftz.f32 %f993, %f992, 0f40400000; sqrt.approx.ftz.f32 %f994, %f993; mul.ftz.f32 %f995, %f991, 0f3F000000; mul.ftz.f32 %f996, %f204, %f994; mul.ftz.f32 %f997, %f996, 0fBFB8AA3B; ex2.approx.ftz.f32 %f998, %f997; add.ftz.f32 %f999, %f998, 0f3F800000; mul.ftz.f32 %f1000, %f999, %f995; mul.ftz.f32 %f1001, %f994, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1002, %f1001; mul.ftz.f32 %f1003, %f1002, %f1000; setp.lt.ftz.f32 %p115, %f1003, %f205; selp.f32 %f1004, %f988, %f991, %p115; selp.f32 %f1005, %f991, %f989, %p115; add.ftz.f32 %f1006, %f1004, %f1005; mul.ftz.f32 %f1007, %f1006, 0f3F000000; sub.ftz.f32 %f1008, %f841, %f1007; mul.ftz.f32 %f1009, %f1008, 0f40400000; sqrt.approx.ftz.f32 %f1010, %f1009; mul.ftz.f32 %f1011, %f1007, 0f3F000000; mul.ftz.f32 %f1012, %f204, %f1010; mul.ftz.f32 %f1013, %f1012, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1014, %f1013; add.ftz.f32 %f1015, %f1014, 0f3F800000; mul.ftz.f32 %f1016, %f1015, %f1011; mul.ftz.f32 %f1017, %f1010, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1018, %f1017; mul.ftz.f32 %f1019, %f1018, %f1016; setp.lt.ftz.f32 %p116, %f1019, %f205; selp.f32 %f1020, %f1004, %f1007, %p116; selp.f32 %f1021, %f1007, %f1005, %p116; add.ftz.f32 %f1022, %f1020, %f1021; mul.ftz.f32 %f3004, %f1022, 0f3F000000; $L__BB157_165: mov.f32 %f3005, 0f00000000; ld.f32 %f208, [%rd14+16]; setp.lt.ftz.f32 %p117, %f208, 0f38D1B717; @%p117 bra $L__BB157_168; setp.ge.ftz.f32 %p118, %f208, 0f3F7EB852; mov.f32 %f3005, 0f3F7FFFEF; @%p118 bra $L__BB157_168; mov.f32 %f2962, 0f3F800000; sub.ftz.f32 %f1026, %f2962, 0f3F000000; mul.ftz.f32 %f1027, %f1026, 0f40400000; sqrt.approx.ftz.f32 %f1028, %f1027; mul.ftz.f32 %f1029, %f204, %f1028; mul.ftz.f32 %f1030, %f1029, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1031, %f1030; add.ftz.f32 %f1032, %f1031, 0f3F800000; mul.ftz.f32 %f1033, %f1032, 0f3E800000; mul.ftz.f32 %f1034, %f1028, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1035, %f1034; mul.ftz.f32 %f1036, %f1035, %f1033; setp.lt.ftz.f32 %p119, %f1036, %f208; selp.f32 %f1037, 0f3F800000, 0f3F000000, %p119; selp.f32 %f1038, 0f3F000000, 0f00000000, %p119; selp.f32 %f1039, 0f3F400000, 0f3E800000, %p119; sub.ftz.f32 %f1040, %f2962, %f1039; mul.ftz.f32 %f1041, %f1040, 0f40400000; sqrt.approx.ftz.f32 %f1042, %f1041; mul.ftz.f32 %f1043, %f1039, 0f3F000000; mul.ftz.f32 %f1044, %f204, %f1042; mul.ftz.f32 %f1045, %f1044, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1046, %f1045; add.ftz.f32 %f1047, %f1046, 0f3F800000; mul.ftz.f32 %f1048, %f1047, %f1043; mul.ftz.f32 %f1049, %f1042, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1050, %f1049; mul.ftz.f32 %f1051, %f1050, %f1048; setp.lt.ftz.f32 %p120, %f1051, %f208; selp.f32 %f1052, %f1037, %f1039, %p120; selp.f32 %f1053, %f1039, %f1038, %p120; add.ftz.f32 %f1054, %f1052, %f1053; mul.ftz.f32 %f1055, %f1054, 0f3F000000; sub.ftz.f32 %f1056, %f2962, %f1055; mul.ftz.f32 %f1057, %f1056, 0f40400000; sqrt.approx.ftz.f32 %f1058, %f1057; mul.ftz.f32 %f1059, %f1055, 0f3F000000; mul.ftz.f32 %f1060, %f204, %f1058; mul.ftz.f32 %f1061, %f1060, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1062, %f1061; add.ftz.f32 %f1063, %f1062, 0f3F800000; mul.ftz.f32 %f1064, %f1063, %f1059; mul.ftz.f32 %f1065, %f1058, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1066, %f1065; mul.ftz.f32 %f1067, %f1066, %f1064; setp.lt.ftz.f32 %p121, %f1067, %f208; selp.f32 %f1068, %f1052, %f1055, %p121; selp.f32 %f1069, %f1055, %f1053, %p121; add.ftz.f32 %f1070, %f1068, %f1069; mul.ftz.f32 %f1071, %f1070, 0f3F000000; sub.ftz.f32 %f1072, %f2962, %f1071; mul.ftz.f32 %f1073, %f1072, 0f40400000; sqrt.approx.ftz.f32 %f1074, %f1073; mul.ftz.f32 %f1075, %f1071, 0f3F000000; mul.ftz.f32 %f1076, %f204, %f1074; mul.ftz.f32 %f1077, %f1076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1078, %f1077; add.ftz.f32 %f1079, %f1078, 0f3F800000; mul.ftz.f32 %f1080, %f1079, %f1075; mul.ftz.f32 %f1081, %f1074, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1082, %f1081; mul.ftz.f32 %f1083, %f1082, %f1080; setp.lt.ftz.f32 %p122, %f1083, %f208; selp.f32 %f1084, %f1068, %f1071, %p122; selp.f32 %f1085, %f1071, %f1069, %p122; add.ftz.f32 %f1086, %f1084, %f1085; mul.ftz.f32 %f1087, %f1086, 0f3F000000; sub.ftz.f32 %f1088, %f2962, %f1087; mul.ftz.f32 %f1089, %f1088, 0f40400000; sqrt.approx.ftz.f32 %f1090, %f1089; mul.ftz.f32 %f1091, %f1087, 0f3F000000; mul.ftz.f32 %f1092, %f204, %f1090; mul.ftz.f32 %f1093, %f1092, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1094, %f1093; add.ftz.f32 %f1095, %f1094, 0f3F800000; mul.ftz.f32 %f1096, %f1095, %f1091; mul.ftz.f32 %f1097, %f1090, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1098, %f1097; mul.ftz.f32 %f1099, %f1098, %f1096; setp.lt.ftz.f32 %p123, %f1099, %f208; selp.f32 %f1100, %f1084, %f1087, %p123; selp.f32 %f1101, %f1087, %f1085, %p123; add.ftz.f32 %f1102, %f1100, %f1101; mul.ftz.f32 %f1103, %f1102, 0f3F000000; sub.ftz.f32 %f1104, %f2962, %f1103; mul.ftz.f32 %f1105, %f1104, 0f40400000; sqrt.approx.ftz.f32 %f1106, %f1105; mul.ftz.f32 %f1107, %f1103, 0f3F000000; mul.ftz.f32 %f1108, %f204, %f1106; mul.ftz.f32 %f1109, %f1108, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1110, %f1109; add.ftz.f32 %f1111, %f1110, 0f3F800000; mul.ftz.f32 %f1112, %f1111, %f1107; mul.ftz.f32 %f1113, %f1106, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1114, %f1113; mul.ftz.f32 %f1115, %f1114, %f1112; setp.lt.ftz.f32 %p124, %f1115, %f208; selp.f32 %f1116, %f1100, %f1103, %p124; selp.f32 %f1117, %f1103, %f1101, %p124; add.ftz.f32 %f1118, %f1116, %f1117; mul.ftz.f32 %f1119, %f1118, 0f3F000000; sub.ftz.f32 %f1120, %f2962, %f1119; mul.ftz.f32 %f1121, %f1120, 0f40400000; sqrt.approx.ftz.f32 %f1122, %f1121; mul.ftz.f32 %f1123, %f1119, 0f3F000000; mul.ftz.f32 %f1124, %f204, %f1122; mul.ftz.f32 %f1125, %f1124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1126, %f1125; add.ftz.f32 %f1127, %f1126, 0f3F800000; mul.ftz.f32 %f1128, %f1127, %f1123; mul.ftz.f32 %f1129, %f1122, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1130, %f1129; mul.ftz.f32 %f1131, %f1130, %f1128; setp.lt.ftz.f32 %p125, %f1131, %f208; selp.f32 %f1132, %f1116, %f1119, %p125; selp.f32 %f1133, %f1119, %f1117, %p125; add.ftz.f32 %f1134, %f1132, %f1133; mul.ftz.f32 %f1135, %f1134, 0f3F000000; sub.ftz.f32 %f1136, %f2962, %f1135; mul.ftz.f32 %f1137, %f1136, 0f40400000; sqrt.approx.ftz.f32 %f1138, %f1137; mul.ftz.f32 %f1139, %f1135, 0f3F000000; mul.ftz.f32 %f1140, %f204, %f1138; mul.ftz.f32 %f1141, %f1140, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1142, %f1141; add.ftz.f32 %f1143, %f1142, 0f3F800000; mul.ftz.f32 %f1144, %f1143, %f1139; mul.ftz.f32 %f1145, %f1138, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1146, %f1145; mul.ftz.f32 %f1147, %f1146, %f1144; setp.lt.ftz.f32 %p126, %f1147, %f208; selp.f32 %f1148, %f1132, %f1135, %p126; selp.f32 %f1149, %f1135, %f1133, %p126; add.ftz.f32 %f1150, %f1148, %f1149; mul.ftz.f32 %f1151, %f1150, 0f3F000000; sub.ftz.f32 %f1152, %f2962, %f1151; mul.ftz.f32 %f1153, %f1152, 0f40400000; sqrt.approx.ftz.f32 %f1154, %f1153; mul.ftz.f32 %f1155, %f1151, 0f3F000000; mul.ftz.f32 %f1156, %f204, %f1154; mul.ftz.f32 %f1157, %f1156, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1158, %f1157; add.ftz.f32 %f1159, %f1158, 0f3F800000; mul.ftz.f32 %f1160, %f1159, %f1155; mul.ftz.f32 %f1161, %f1154, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1162, %f1161; mul.ftz.f32 %f1163, %f1162, %f1160; setp.lt.ftz.f32 %p127, %f1163, %f208; selp.f32 %f1164, %f1148, %f1151, %p127; selp.f32 %f1165, %f1151, %f1149, %p127; add.ftz.f32 %f1166, %f1164, %f1165; mul.ftz.f32 %f1167, %f1166, 0f3F000000; sub.ftz.f32 %f1168, %f2962, %f1167; mul.ftz.f32 %f1169, %f1168, 0f40400000; sqrt.approx.ftz.f32 %f1170, %f1169; mul.ftz.f32 %f1171, %f1167, 0f3F000000; mul.ftz.f32 %f1172, %f204, %f1170; mul.ftz.f32 %f1173, %f1172, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1174, %f1173; add.ftz.f32 %f1175, %f1174, 0f3F800000; mul.ftz.f32 %f1176, %f1175, %f1171; mul.ftz.f32 %f1177, %f1170, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1178, %f1177; mul.ftz.f32 %f1179, %f1178, %f1176; setp.lt.ftz.f32 %p128, %f1179, %f208; selp.f32 %f1180, %f1164, %f1167, %p128; selp.f32 %f1181, %f1167, %f1165, %p128; add.ftz.f32 %f1182, %f1180, %f1181; mul.ftz.f32 %f1183, %f1182, 0f3F000000; sub.ftz.f32 %f1184, %f2962, %f1183; mul.ftz.f32 %f1185, %f1184, 0f40400000; sqrt.approx.ftz.f32 %f1186, %f1185; mul.ftz.f32 %f1187, %f1183, 0f3F000000; mul.ftz.f32 %f1188, %f204, %f1186; mul.ftz.f32 %f1189, %f1188, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1190, %f1189; add.ftz.f32 %f1191, %f1190, 0f3F800000; mul.ftz.f32 %f1192, %f1191, %f1187; mul.ftz.f32 %f1193, %f1186, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1194, %f1193; mul.ftz.f32 %f1195, %f1194, %f1192; setp.lt.ftz.f32 %p129, %f1195, %f208; selp.f32 %f1196, %f1180, %f1183, %p129; selp.f32 %f1197, %f1183, %f1181, %p129; add.ftz.f32 %f1198, %f1196, %f1197; mul.ftz.f32 %f3005, %f1198, 0f3F000000; $L__BB157_168: mov.f32 %f3006, 0f00000000; ld.f32 %f211, [%rd14+20]; setp.lt.ftz.f32 %p130, %f211, 0f38D1B717; @%p130 bra $L__BB157_171; setp.ge.ftz.f32 %p131, %f211, 0f3F7EB852; mov.f32 %f3006, 0f3F7FFFEF; @%p131 bra $L__BB157_171; mov.f32 %f2957, 0f3F800000; sub.ftz.f32 %f1202, %f2957, 0f3F000000; mul.ftz.f32 %f1203, %f1202, 0f40400000; sqrt.approx.ftz.f32 %f1204, %f1203; mul.ftz.f32 %f1205, %f204, %f1204; mul.ftz.f32 %f1206, %f1205, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1207, %f1206; add.ftz.f32 %f1208, %f1207, 0f3F800000; mul.ftz.f32 %f1209, %f1208, 0f3E800000; mul.ftz.f32 %f1210, %f1204, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1211, %f1210; mul.ftz.f32 %f1212, %f1211, %f1209; setp.lt.ftz.f32 %p132, %f1212, %f211; selp.f32 %f1213, 0f3F800000, 0f3F000000, %p132; selp.f32 %f1214, 0f3F000000, 0f00000000, %p132; selp.f32 %f1215, 0f3F400000, 0f3E800000, %p132; sub.ftz.f32 %f1216, %f2957, %f1215; mul.ftz.f32 %f1217, %f1216, 0f40400000; sqrt.approx.ftz.f32 %f1218, %f1217; mul.ftz.f32 %f1219, %f1215, 0f3F000000; mul.ftz.f32 %f1220, %f204, %f1218; mul.ftz.f32 %f1221, %f1220, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1222, %f1221; add.ftz.f32 %f1223, %f1222, 0f3F800000; mul.ftz.f32 %f1224, %f1223, %f1219; mul.ftz.f32 %f1225, %f1218, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1226, %f1225; mul.ftz.f32 %f1227, %f1226, %f1224; setp.lt.ftz.f32 %p133, %f1227, %f211; selp.f32 %f1228, %f1213, %f1215, %p133; selp.f32 %f1229, %f1215, %f1214, %p133; add.ftz.f32 %f1230, %f1228, %f1229; mul.ftz.f32 %f1231, %f1230, 0f3F000000; sub.ftz.f32 %f1232, %f2957, %f1231; mul.ftz.f32 %f1233, %f1232, 0f40400000; sqrt.approx.ftz.f32 %f1234, %f1233; mul.ftz.f32 %f1235, %f1231, 0f3F000000; mul.ftz.f32 %f1236, %f204, %f1234; mul.ftz.f32 %f1237, %f1236, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1238, %f1237; add.ftz.f32 %f1239, %f1238, 0f3F800000; mul.ftz.f32 %f1240, %f1239, %f1235; mul.ftz.f32 %f1241, %f1234, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1242, %f1241; mul.ftz.f32 %f1243, %f1242, %f1240; setp.lt.ftz.f32 %p134, %f1243, %f211; selp.f32 %f1244, %f1228, %f1231, %p134; selp.f32 %f1245, %f1231, %f1229, %p134; add.ftz.f32 %f1246, %f1244, %f1245; mul.ftz.f32 %f1247, %f1246, 0f3F000000; sub.ftz.f32 %f1248, %f2957, %f1247; mul.ftz.f32 %f1249, %f1248, 0f40400000; sqrt.approx.ftz.f32 %f1250, %f1249; mul.ftz.f32 %f1251, %f1247, 0f3F000000; mul.ftz.f32 %f1252, %f204, %f1250; mul.ftz.f32 %f1253, %f1252, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1254, %f1253; add.ftz.f32 %f1255, %f1254, 0f3F800000; mul.ftz.f32 %f1256, %f1255, %f1251; mul.ftz.f32 %f1257, %f1250, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1258, %f1257; mul.ftz.f32 %f1259, %f1258, %f1256; setp.lt.ftz.f32 %p135, %f1259, %f211; selp.f32 %f1260, %f1244, %f1247, %p135; selp.f32 %f1261, %f1247, %f1245, %p135; add.ftz.f32 %f1262, %f1260, %f1261; mul.ftz.f32 %f1263, %f1262, 0f3F000000; sub.ftz.f32 %f1264, %f2957, %f1263; mul.ftz.f32 %f1265, %f1264, 0f40400000; sqrt.approx.ftz.f32 %f1266, %f1265; mul.ftz.f32 %f1267, %f1263, 0f3F000000; mul.ftz.f32 %f1268, %f204, %f1266; mul.ftz.f32 %f1269, %f1268, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1270, %f1269; add.ftz.f32 %f1271, %f1270, 0f3F800000; mul.ftz.f32 %f1272, %f1271, %f1267; mul.ftz.f32 %f1273, %f1266, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1274, %f1273; mul.ftz.f32 %f1275, %f1274, %f1272; setp.lt.ftz.f32 %p136, %f1275, %f211; selp.f32 %f1276, %f1260, %f1263, %p136; selp.f32 %f1277, %f1263, %f1261, %p136; add.ftz.f32 %f1278, %f1276, %f1277; mul.ftz.f32 %f1279, %f1278, 0f3F000000; sub.ftz.f32 %f1280, %f2957, %f1279; mul.ftz.f32 %f1281, %f1280, 0f40400000; sqrt.approx.ftz.f32 %f1282, %f1281; mul.ftz.f32 %f1283, %f1279, 0f3F000000; mul.ftz.f32 %f1284, %f204, %f1282; mul.ftz.f32 %f1285, %f1284, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1286, %f1285; add.ftz.f32 %f1287, %f1286, 0f3F800000; mul.ftz.f32 %f1288, %f1287, %f1283; mul.ftz.f32 %f1289, %f1282, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1290, %f1289; mul.ftz.f32 %f1291, %f1290, %f1288; setp.lt.ftz.f32 %p137, %f1291, %f211; selp.f32 %f1292, %f1276, %f1279, %p137; selp.f32 %f1293, %f1279, %f1277, %p137; add.ftz.f32 %f1294, %f1292, %f1293; mul.ftz.f32 %f1295, %f1294, 0f3F000000; sub.ftz.f32 %f1296, %f2957, %f1295; mul.ftz.f32 %f1297, %f1296, 0f40400000; sqrt.approx.ftz.f32 %f1298, %f1297; mul.ftz.f32 %f1299, %f1295, 0f3F000000; mul.ftz.f32 %f1300, %f204, %f1298; mul.ftz.f32 %f1301, %f1300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1302, %f1301; add.ftz.f32 %f1303, %f1302, 0f3F800000; mul.ftz.f32 %f1304, %f1303, %f1299; mul.ftz.f32 %f1305, %f1298, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1306, %f1305; mul.ftz.f32 %f1307, %f1306, %f1304; setp.lt.ftz.f32 %p138, %f1307, %f211; selp.f32 %f1308, %f1292, %f1295, %p138; selp.f32 %f1309, %f1295, %f1293, %p138; add.ftz.f32 %f1310, %f1308, %f1309; mul.ftz.f32 %f1311, %f1310, 0f3F000000; sub.ftz.f32 %f1312, %f2957, %f1311; mul.ftz.f32 %f1313, %f1312, 0f40400000; sqrt.approx.ftz.f32 %f1314, %f1313; mul.ftz.f32 %f1315, %f1311, 0f3F000000; mul.ftz.f32 %f1316, %f204, %f1314; mul.ftz.f32 %f1317, %f1316, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1318, %f1317; add.ftz.f32 %f1319, %f1318, 0f3F800000; mul.ftz.f32 %f1320, %f1319, %f1315; mul.ftz.f32 %f1321, %f1314, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1322, %f1321; mul.ftz.f32 %f1323, %f1322, %f1320; setp.lt.ftz.f32 %p139, %f1323, %f211; selp.f32 %f1324, %f1308, %f1311, %p139; selp.f32 %f1325, %f1311, %f1309, %p139; add.ftz.f32 %f1326, %f1324, %f1325; mul.ftz.f32 %f1327, %f1326, 0f3F000000; sub.ftz.f32 %f1328, %f2957, %f1327; mul.ftz.f32 %f1329, %f1328, 0f40400000; sqrt.approx.ftz.f32 %f1330, %f1329; mul.ftz.f32 %f1331, %f1327, 0f3F000000; mul.ftz.f32 %f1332, %f204, %f1330; mul.ftz.f32 %f1333, %f1332, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1334, %f1333; add.ftz.f32 %f1335, %f1334, 0f3F800000; mul.ftz.f32 %f1336, %f1335, %f1331; mul.ftz.f32 %f1337, %f1330, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1338, %f1337; mul.ftz.f32 %f1339, %f1338, %f1336; setp.lt.ftz.f32 %p140, %f1339, %f211; selp.f32 %f1340, %f1324, %f1327, %p140; selp.f32 %f1341, %f1327, %f1325, %p140; add.ftz.f32 %f1342, %f1340, %f1341; mul.ftz.f32 %f1343, %f1342, 0f3F000000; sub.ftz.f32 %f1344, %f2957, %f1343; mul.ftz.f32 %f1345, %f1344, 0f40400000; sqrt.approx.ftz.f32 %f1346, %f1345; mul.ftz.f32 %f1347, %f1343, 0f3F000000; mul.ftz.f32 %f1348, %f204, %f1346; mul.ftz.f32 %f1349, %f1348, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1350, %f1349; add.ftz.f32 %f1351, %f1350, 0f3F800000; mul.ftz.f32 %f1352, %f1351, %f1347; mul.ftz.f32 %f1353, %f1346, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1354, %f1353; mul.ftz.f32 %f1355, %f1354, %f1352; setp.lt.ftz.f32 %p141, %f1355, %f211; selp.f32 %f1356, %f1340, %f1343, %p141; selp.f32 %f1357, %f1343, %f1341, %p141; add.ftz.f32 %f1358, %f1356, %f1357; mul.ftz.f32 %f1359, %f1358, 0f3F000000; sub.ftz.f32 %f1360, %f2957, %f1359; mul.ftz.f32 %f1361, %f1360, 0f40400000; sqrt.approx.ftz.f32 %f1362, %f1361; mul.ftz.f32 %f1363, %f1359, 0f3F000000; mul.ftz.f32 %f1364, %f204, %f1362; mul.ftz.f32 %f1365, %f1364, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1366, %f1365; add.ftz.f32 %f1367, %f1366, 0f3F800000; mul.ftz.f32 %f1368, %f1367, %f1363; mul.ftz.f32 %f1369, %f1362, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f1370, %f1368; setp.lt.ftz.f32 %p142, %f1371, %f211; selp.f32 %f1372, %f1356, %f1359, %p142; selp.f32 %f1373, %f1359, %f1357, %p142; add.ftz.f32 %f1374, %f1372, %f1373; mul.ftz.f32 %f3006, %f1374, 0f3F000000; $L__BB157_171: mov.f32 %f2958, 0f3F800000; sub.ftz.f32 %f1376, %f2958, %f3004; mul.ftz.f32 %f1377, %f1376, 0f40400000; sub.ftz.f32 %f1378, %f2958, %f3005; mul.ftz.f32 %f1379, %f1378, 0f40400000; sub.ftz.f32 %f1380, %f2958, %f3006; mul.ftz.f32 %f1381, %f1380, 0f40400000; sqrt.approx.ftz.f32 %f1382, %f1377; sqrt.approx.ftz.f32 %f1383, %f1379; sqrt.approx.ftz.f32 %f1384, %f1381; ld.f32 %f1385, [%rd14]; mul.ftz.f32 %f3007, %f1382, %f1385; ld.f32 %f1386, [%rd14+4]; mul.ftz.f32 %f3008, %f1383, %f1386; ld.f32 %f1387, [%rd14+8]; mul.ftz.f32 %f3009, %f1384, %f1387; $L__BB157_173: st.f32 [%rd14], %f3007; st.f32 [%rd14+4], %f3008; st.f32 [%rd14+8], %f3009; or.b32 %r643, %r643, 16; $L__BB157_174: ld.local.u32 %r224, [%rd4+40]; or.b32 %r225, %r224, %r643; st.local.u32 [%rd4+40], %r225; $L__BB157_182: setp.leu.ftz.f32 %p147, %f97, 0f3727C5AC; setp.leu.ftz.f32 %p148, %f74, 0f3727C5AC; or.pred %p149, %p148, %p147; @%p149 bra $L__BB157_194; ld.const.v4.f32 {%f1404, %f1405, %f1406, %f1407}, [kernel_params+1184]; mul.ftz.f32 %f1411, %f187, %f1405; fma.rn.ftz.f32 %f1412, %f186, %f1404, %f1411; fma.rn.ftz.f32 %f227, %f188, %f1406, %f1412; setp.leu.ftz.f32 %p150, %f227, 0f00000000; mov.f32 %f1403, 0f3F800000; mov.f32 %f3010, %f1403; mov.f32 %f3011, %f1403; mov.f32 %f3012, %f1403; @%p150 bra $L__BB157_185; div.approx.ftz.f32 %f3012, %f186, %f227; div.approx.ftz.f32 %f3011, %f187, %f227; div.approx.ftz.f32 %f3010, %f188, %f227; $L__BB157_185: sub.ftz.f32 %f1414, %f1403, %f75; fma.rn.ftz.f32 %f1415, %f75, %f3012, %f1414; fma.rn.ftz.f32 %f1416, %f75, %f3011, %f1414; fma.rn.ftz.f32 %f1417, %f75, %f3010, %f1414; mul.ftz.f32 %f1418, %f74, %f177; mul.ftz.f32 %f1419, %f1418, %f1415; mul.ftz.f32 %f1420, %f74, %f178; mul.ftz.f32 %f1421, %f1420, %f1416; mul.ftz.f32 %f1422, %f74, %f179; mul.ftz.f32 %f1423, %f1422, %f1417; mul.ftz.f32 %f1424, %f97, %f1419; mul.ftz.f32 %f1425, %f97, %f1421; mul.ftz.f32 %f1426, %f97, %f1423; mov.f32 %f3013, 0f00000000; max.ftz.f32 %f234, %f1424, %f3013; max.ftz.f32 %f235, %f1425, %f3013; max.ftz.f32 %f236, %f1426, %f3013; add.ftz.f32 %f1428, %f234, %f235; add.ftz.f32 %f1429, %f236, %f1428; mul.ftz.f32 %f1430, %f1429, 0f3EAAAAAB; abs.ftz.f32 %f237, %f1430; setp.ltu.ftz.f32 %p151, %f237, 0f3727C5AC; mov.u64 %rd712, 0; mov.u64 %rd713, %rd712; @%p151 bra $L__BB157_190; ld.local.u32 %r43, [%rd4+252]; mov.u64 %rd712, 0; setp.eq.s32 %p152, %r43, 0; mov.u64 %rd710, %rd712; mov.u64 %rd711, %rd712; @%p152 bra $L__BB157_188; ld.local.u32 %r237, [%rd4+248]; mul.wide.s32 %rd337, %r237, 80; add.s64 %rd338, %rd1, %rd337; add.s64 %rd710, %rd338, 304; add.s64 %rd339, %rd230, %rd337; add.s64 %rd711, %rd339, 304; mov.b32 %r238, %f236; mov.b32 %r239, %f235; mov.b32 %r240, %f234; mov.u32 %r241, 0; st.local.v4.u32 [%rd338+304], {%r240, %r239, %r238, %r241}; add.s32 %r242, %r237, 1; st.local.u32 [%rd4+248], %r242; add.s32 %r243, %r43, -1; st.local.u32 [%rd4+252], %r243; $L__BB157_188: setp.eq.s64 %p153, %rd711, 0; mov.u64 %rd713, %rd712; @%p153 bra $L__BB157_190; st.local.f32 [%rd710+16], %f237; mov.u64 %rd712, %rd710; mov.u64 %rd713, %rd711; $L__BB157_190: setp.eq.s64 %p154, %rd713, 0; @%p154 bra $L__BB157_194; add.s64 %rd45, %rd712, 20; st.local.f32 [%rd712+20], %f55; st.local.f32 [%rd712+24], %f56; st.local.f32 [%rd712+28], %f57; mov.u32 %r244, 6; st.local.u32 [%rd712+12], %r244; ld.local.f32 %f1432, [%rd4+24]; ld.local.f32 %f1433, [%rd4+28]; mul.ftz.f32 %f1434, %f56, %f1433; fma.rn.ftz.f32 %f1435, %f55, %f1432, %f1434; ld.local.f32 %f1436, [%rd4+32]; fma.rn.ftz.f32 %f238, %f57, %f1436, %f1435; setp.lt.ftz.f32 %p155, %f238, 0f00000000; @%p155 bra $L__BB157_193; mov.f32 %f1437, 0f3F800000; sub.ftz.f32 %f1438, %f1437, %f238; mov.f32 %f1439, 0f00000000; max.ftz.f32 %f1440, %f1438, %f1439; min.ftz.f32 %f1441, %f1440, %f1437; mul.ftz.f32 %f1442, %f1441, %f1441; mul.ftz.f32 %f1443, %f1442, %f1442; mul.ftz.f32 %f1444, %f1441, %f1443; mul.ftz.f32 %f3013, %f238, %f1444; $L__BB157_193: st.local.f32 [%rd45+12], %f3013; ld.local.f32 %f1445, [%rd45+-4]; mul.ftz.f32 %f1446, %f3013, %f1445; st.local.f32 [%rd45+-4], %f1446; ld.local.u32 %r245, [%rd4+40]; or.b32 %r246, %r245, 12; st.local.u32 [%rd4+40], %r246; $L__BB157_194: ld.param.u32 %r637, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.const.u32 %r44, [kernel_params+1564]; setp.ne.s32 %p156, %r44, 0; and.b32 %r45, %r637, 8; setp.eq.s32 %p157, %r45, 0; or.pred %p5, %p157, %p156; not.pred %p158, %p5; setp.leu.ftz.f32 %p159, %f99, 0f3727C5AC; or.pred %p160, %p159, %p158; @%p160 bra $L__BB157_225; setp.leu.ftz.f32 %p161, %f70, 0f3727C5AC; setp.leu.ftz.f32 %p162, %f60, 0f3727C5AC; and.pred %p163, %p162, %p161; @%p163 bra $L__BB157_225; mul.ftz.f32 %f1447, %f99, %f177; mov.f32 %f1448, 0f00000000; max.ftz.f32 %f241, %f1447, %f1448; mul.ftz.f32 %f1449, %f99, %f178; max.ftz.f32 %f242, %f1449, %f1448; mul.ftz.f32 %f1450, %f99, %f179; max.ftz.f32 %f243, %f1450, %f1448; add.ftz.f32 %f1451, %f241, %f242; add.ftz.f32 %f1452, %f243, %f1451; mul.ftz.f32 %f1453, %f1452, 0f3EAAAAAB; abs.ftz.f32 %f244, %f1453; setp.ltu.ftz.f32 %p164, %f244, 0f3727C5AC; mov.u64 %rd720, 0; mov.u64 %rd716, %rd720; mov.u64 %rd717, %rd720; @%p164 bra $L__BB157_201; ld.local.u32 %r46, [%rd4+252]; mov.u64 %rd716, 0; setp.eq.s32 %p165, %r46, 0; mov.u64 %rd714, %rd716; mov.u64 %rd715, %rd716; @%p165 bra $L__BB157_199; ld.local.u32 %r247, [%rd4+248]; mul.wide.s32 %rd346, %r247, 80; add.s64 %rd347, %rd1, %rd346; add.s64 %rd714, %rd347, 304; add.s64 %rd348, %rd230, %rd346; add.s64 %rd715, %rd348, 304; mov.b32 %r248, %f243; mov.b32 %r249, %f242; mov.b32 %r250, %f241; mov.u32 %r251, 0; st.local.v4.u32 [%rd347+304], {%r250, %r249, %r248, %r251}; add.s32 %r252, %r247, 1; st.local.u32 [%rd4+248], %r252; add.s32 %r253, %r46, -1; st.local.u32 [%rd4+252], %r253; $L__BB157_199: setp.eq.s64 %p166, %rd715, 0; mov.u64 %rd717, %rd716; @%p166 bra $L__BB157_201; st.local.f32 [%rd714+16], %f244; mov.u64 %rd716, %rd714; mov.u64 %rd717, %rd715; $L__BB157_201: setp.eq.s64 %p167, %rd717, 0; mov.u64 %rd721, %rd720; @%p167 bra $L__BB157_205; ld.local.u32 %r47, [%rd4+252]; setp.lt.s32 %p168, %r47, 1; @%p168 bra $L__BB157_204; add.s32 %r254, %r47, -1; st.local.u32 [%rd4+252], %r254; ld.local.s32 %rd353, [%rd4+248]; cvt.s64.s32 %rd354, %r254; add.s64 %rd355, %rd353, %rd354; mul.lo.s64 %rd356, %rd355, 80; add.s64 %rd357, %rd1, %rd356; add.s64 %rd720, %rd357, 304; add.s64 %rd358, %rd230, %rd356; add.s64 %rd721, %rd358, 304; bra.uni $L__BB157_205; $L__BB157_204: ld.local.u32 %r255, [%rd4+248]; mov.u64 %rd720, 0; add.s32 %r256, %r255, -1; st.local.u32 [%rd4+248], %r256; add.s32 %r257, %r47, 1; st.local.u32 [%rd4+252], %r257; mov.u64 %rd721, %rd720; $L__BB157_205: setp.eq.s64 %p169, %rd721, 0; or.pred %p171, %p167, %p169; @%p171 bra $L__BB157_225; add.s64 %rd60, %rd716, 20; st.local.f32 [%rd716+20], %f55; st.local.f32 [%rd716+24], %f56; st.local.f32 [%rd716+28], %f57; mul.ftz.f32 %f245, %f70, 0f3DA3D70A; max.ftz.f32 %f1458, %f245, %f1448; sqrt.approx.ftz.f32 %f1459, %f1458; mov.f32 %f1456, 0f3F800000; sub.ftz.f32 %f1460, %f1456, %f1459; mov.f32 %f1461, 0f40000000; div.approx.ftz.f32 %f1462, %f1461, %f1460; add.ftz.f32 %f1463, %f1462, 0fBF800000; st.local.f32 [%rd716+40], %f1463; st.local.v2.f32 [%rd716+56], {%f2977, %f2978}; st.local.f32 [%rd716+64], %f2979; st.local.u64 [%rd716+48], %rd721; fma.rn.ftz.f32 %f1464, %f73, 0fBF666666, 0f3F800000; max.ftz.f32 %f1465, %f1464, %f1448; sqrt.approx.ftz.f32 %f1466, %f1465; mul.ftz.f32 %f1467, %f71, %f71; mul.ftz.f32 %f1468, %f1467, %f1466; div.approx.ftz.f32 %f1469, %f1467, %f1466; st.local.v2.f32 [%rd716+32], {%f1469, %f1468}; mul.ftz.f32 %f1470, %f187, 0f3F19999A; fma.rn.ftz.f32 %f1471, %f186, 0f3E99999A, %f1470; fma.rn.ftz.f32 %f246, %f188, 0f3DCCCCCD, %f1471; setp.leu.ftz.f32 %p172, %f246, 0f00000000; mov.f32 %f3014, %f1456; mov.f32 %f3015, %f1456; mov.f32 %f3016, %f1456; @%p172 bra $L__BB157_208; div.approx.ftz.f32 %f3016, %f186, %f246; div.approx.ftz.f32 %f3015, %f187, %f246; div.approx.ftz.f32 %f3014, %f188, %f246; $L__BB157_208: sub.ftz.f32 %f1473, %f1456, %f72; fma.rn.ftz.f32 %f1474, %f72, %f3016, %f1473; fma.rn.ftz.f32 %f1475, %f72, %f3015, %f1473; fma.rn.ftz.f32 %f1476, %f72, %f3014, %f1473; mul.ftz.f32 %f1477, %f245, %f1474; mul.ftz.f32 %f1478, %f245, %f1475; mul.ftz.f32 %f1479, %f245, %f1476; sub.ftz.f32 %f1480, %f1456, %f60; mul.ftz.f32 %f1481, %f1480, %f1477; mul.ftz.f32 %f1482, %f1480, %f1478; mul.ftz.f32 %f1483, %f1480, %f1479; fma.rn.ftz.f32 %f1484, %f60, %f186, %f1481; fma.rn.ftz.f32 %f1485, %f60, %f187, %f1482; fma.rn.ftz.f32 %f1486, %f60, %f188, %f1483; st.local.f32 [%rd720+12], %f1484; st.local.f32 [%rd720+16], %f1485; st.local.f32 [%rd720+20], %f1486; ld.local.u64 %rd361, [%rd60+28]; st.f32 [%rd361], %f186; st.f32 [%rd361+4], %f187; st.f32 [%rd361+8], %f188; mov.u32 %r258, 0; ld.local.u64 %rd362, [%rd60+28]; st.u32 [%rd362+36], %r258; setp.le.ftz.f32 %p173, %f71, 0f3D99999A; setp.eq.s32 %p174, %r148, 26; or.pred %p175, %p174, %p173; @%p175 bra $L__BB157_218; bra.uni $L__BB157_209; $L__BB157_218: ld.local.u64 %rd367, [%rd60+28]; ld.f32 %f1573, [%rd367+12]; cvt.ftz.sat.f32.f32 %f1574, %f1573; ld.f32 %f1575, [%rd367+16]; cvt.ftz.sat.f32.f32 %f1576, %f1575; ld.f32 %f1577, [%rd367+20]; cvt.ftz.sat.f32.f32 %f1578, %f1577; st.f32 [%rd367+12], %f1574; st.f32 [%rd367+16], %f1576; st.f32 [%rd367+20], %f1578; ld.local.v2.f32 {%f1579, %f1580}, [%rd60+12]; cvt.ftz.sat.f32.f32 %f1583, %f1580; cvt.ftz.sat.f32.f32 %f1584, %f1579; st.local.v2.f32 [%rd60+12], {%f1584, %f1583}; mov.u32 %r264, 11; st.local.u32 [%rd60+-8], %r264; ld.local.f32 %f1585, [%rd60+20]; fma.rn.ftz.f32 %f266, %f1585, %f1585, 0fBF800000; mov.f32 %f1572, 0f3F800000; abs.ftz.f32 %f267, %f1572; fma.rn.ftz.f32 %f268, %f267, %f267, %f266; setp.leu.ftz.f32 %p184, %f268, 0f00000000; mov.f32 %f3019, %f1572; @%p184 bra $L__BB157_220; sqrt.approx.ftz.f32 %f1586, %f268; sub.ftz.f32 %f1587, %f1586, %f267; add.ftz.f32 %f1588, %f267, %f1586; div.approx.ftz.f32 %f1589, %f1587, %f1588; fma.rn.ftz.f32 %f1590, %f267, %f1588, 0fBF800000; fma.rn.ftz.f32 %f1591, %f267, %f1587, 0f3F800000; div.approx.ftz.f32 %f1592, %f1590, %f1591; mul.ftz.f32 %f1593, %f1589, 0f3F000000; mul.ftz.f32 %f1594, %f1589, %f1593; fma.rn.ftz.f32 %f1595, %f1592, %f1592, 0f3F800000; mul.ftz.f32 %f3019, %f1594, %f1595; $L__BB157_220: ld.local.u64 %rd368, [%rd60+28]; add.s64 %rd62, %rd368, 12; ld.f32 %f271, [%rd368+12]; ld.f32 %f272, [%rd368+16]; ld.f32 %f273, [%rd368+20]; sub.ftz.f32 %f274, %f1572, %f3019; ld.local.f32 %f1597, [%rd60]; ld.local.f32 %f1598, [%rd4+24]; ld.local.f32 %f1599, [%rd60+4]; ld.local.f32 %f1600, [%rd4+28]; mul.ftz.f32 %f1601, %f1600, %f1599; fma.rn.ftz.f32 %f1602, %f1598, %f1597, %f1601; ld.local.f32 %f1603, [%rd60+8]; ld.local.f32 %f1604, [%rd4+32]; fma.rn.ftz.f32 %f1605, %f1604, %f1603, %f1602; abs.ftz.f32 %f275, %f1605; fma.rn.ftz.f32 %f276, %f275, %f275, %f266; setp.leu.ftz.f32 %p185, %f276, 0f00000000; mov.f32 %f3020, %f274; @%p185 bra $L__BB157_222; sqrt.approx.ftz.f32 %f1606, %f276; sub.ftz.f32 %f1607, %f1606, %f275; add.ftz.f32 %f1608, %f275, %f1606; div.approx.ftz.f32 %f1609, %f1607, %f1608; fma.rn.ftz.f32 %f1610, %f275, %f1608, 0fBF800000; fma.rn.ftz.f32 %f1611, %f275, %f1607, 0f3F800000; div.approx.ftz.f32 %f1612, %f1610, %f1611; mul.ftz.f32 %f1613, %f1609, 0f3F000000; mul.ftz.f32 %f1614, %f1609, %f1613; fma.rn.ftz.f32 %f1615, %f1612, %f1612, 0f3F800000; mul.ftz.f32 %f1616, %f1614, %f1615; sub.ftz.f32 %f3020, %f1616, %f3019; $L__BB157_222: rcp.approx.ftz.f32 %f1617, %f274; mov.f32 %f1618, 0f3F800000; mul.ftz.f32 %f1619, %f1617, %f3020; sub.ftz.f32 %f1620, %f1618, %f1619; fma.rn.ftz.f32 %f1621, %f271, %f1620, %f1619; fma.rn.ftz.f32 %f1622, %f272, %f1620, %f1619; fma.rn.ftz.f32 %f1623, %f273, %f1620, %f1619; st.f32 [%rd62+12], %f1621; st.f32 [%rd62+16], %f1622; st.f32 [%rd62+20], %f1623; ld.local.u32 %r265, [%rd60+-8]; setp.ne.s32 %p186, %r265, 12; @%p186 bra $L__BB157_224; ld.local.u64 %rd369, [%rd60+28]; ld.f32 %f1624, [%rd369+36]; mul.ftz.f32 %f1625, %f1624, 0f3E800000; ld.f32 %f1626, [%rd369+24]; mul.ftz.f32 %f1627, %f1626, %f1625; ld.f32 %f1628, [%rd369+28]; mul.ftz.f32 %f1629, %f1625, %f1628; ld.f32 %f1630, [%rd369+32]; mul.ftz.f32 %f1631, %f1625, %f1630; st.f32 [%rd369+24], %f1627; st.f32 [%rd369+28], %f1629; st.f32 [%rd369+32], %f1631; $L__BB157_224: ld.local.u64 %rd370, [%rd60+28]; ld.f32 %f1632, [%rd370+28]; ld.f32 %f1633, [%rd370+24]; add.ftz.f32 %f1634, %f1633, %f1632; ld.f32 %f1635, [%rd370+32]; add.ftz.f32 %f1636, %f1635, %f1634; mul.ftz.f32 %f1637, %f1636, 0f3EAAAAAB; ld.local.f32 %f1638, [%rd60+-4]; mul.ftz.f32 %f1639, %f1638, %f1637; st.local.f32 [%rd60+-4], %f1639; ld.local.u32 %r266, [%rd4+40]; or.b32 %r267, %r266, 12; st.local.u32 [%rd4+40], %r267; bra.uni $L__BB157_225; $L__BB157_209: ld.local.v2.f32 {%f1487, %f1488}, [%rd60+36]; setp.neu.ftz.f32 %p176, %f1487, 0f00000000; setp.neu.ftz.f32 %p177, %f1488, 0f00000000; or.pred %p178, %p176, %p177; ld.local.f32 %f1491, [%rd60+44]; setp.neu.ftz.f32 %p179, %f1491, 0f00000000; or.pred %p180, %p179, %p178; @%p180 bra $L__BB157_211; mov.f32 %f1493, 0f3F800000; st.local.v2.f32 [%rd60+36], {%f1493, %f1448}; st.local.u32 [%rd60+44], %r258; $L__BB157_211: mov.u32 %r260, 15; st.local.u32 [%rd60+-8], %r260; ld.local.f32 %f1495, [%rd60+20]; fma.rn.ftz.f32 %f253, %f1495, %f1495, 0fBF800000; mov.f32 %f1494, 0f3F800000; abs.ftz.f32 %f254, %f1494; fma.rn.ftz.f32 %f255, %f254, %f254, %f253; setp.leu.ftz.f32 %p181, %f255, 0f00000000; mov.f32 %f3017, %f1494; @%p181 bra $L__BB157_213; sqrt.approx.ftz.f32 %f1496, %f255; sub.ftz.f32 %f1497, %f1496, %f254; add.ftz.f32 %f1498, %f254, %f1496; div.approx.ftz.f32 %f1499, %f1497, %f1498; fma.rn.ftz.f32 %f1500, %f254, %f1498, 0fBF800000; fma.rn.ftz.f32 %f1501, %f254, %f1497, 0f3F800000; div.approx.ftz.f32 %f1502, %f1500, %f1501; mul.ftz.f32 %f1503, %f1499, 0f3F000000; mul.ftz.f32 %f1504, %f1499, %f1503; fma.rn.ftz.f32 %f1505, %f1502, %f1502, 0f3F800000; mul.ftz.f32 %f3017, %f1504, %f1505; $L__BB157_213: ld.local.u64 %rd363, [%rd60+28]; add.s64 %rd61, %rd363, 12; ld.f32 %f258, [%rd363+12]; ld.f32 %f259, [%rd363+16]; ld.f32 %f260, [%rd363+20]; sub.ftz.f32 %f261, %f1494, %f3017; ld.local.f32 %f1507, [%rd60]; ld.local.f32 %f1508, [%rd4+24]; ld.local.f32 %f1509, [%rd60+4]; ld.local.f32 %f1510, [%rd4+28]; mul.ftz.f32 %f1511, %f1510, %f1509; fma.rn.ftz.f32 %f1512, %f1508, %f1507, %f1511; ld.local.f32 %f1513, [%rd60+8]; ld.local.f32 %f1514, [%rd4+32]; fma.rn.ftz.f32 %f1515, %f1514, %f1513, %f1512; abs.ftz.f32 %f262, %f1515; fma.rn.ftz.f32 %f263, %f262, %f262, %f253; setp.leu.ftz.f32 %p182, %f263, 0f00000000; mov.f32 %f3018, %f261; @%p182 bra $L__BB157_215; sqrt.approx.ftz.f32 %f1516, %f263; sub.ftz.f32 %f1517, %f1516, %f262; add.ftz.f32 %f1518, %f262, %f1516; div.approx.ftz.f32 %f1519, %f1517, %f1518; fma.rn.ftz.f32 %f1520, %f262, %f1518, 0fBF800000; fma.rn.ftz.f32 %f1521, %f262, %f1517, 0f3F800000; div.approx.ftz.f32 %f1522, %f1520, %f1521; mul.ftz.f32 %f1523, %f1519, 0f3F000000; mul.ftz.f32 %f1524, %f1519, %f1523; fma.rn.ftz.f32 %f1525, %f1522, %f1522, 0f3F800000; mul.ftz.f32 %f1526, %f1524, %f1525; sub.ftz.f32 %f3018, %f1526, %f3017; $L__BB157_215: rcp.approx.ftz.f32 %f1527, %f261; mov.f32 %f1528, 0f3F800000; mul.ftz.f32 %f1529, %f1527, %f3018; sub.ftz.f32 %f1530, %f1528, %f1529; fma.rn.ftz.f32 %f1531, %f258, %f1530, %f1529; fma.rn.ftz.f32 %f1532, %f259, %f1530, %f1529; fma.rn.ftz.f32 %f1533, %f260, %f1530, %f1529; st.f32 [%rd61+12], %f1531; st.f32 [%rd61+16], %f1532; st.f32 [%rd61+20], %f1533; ld.local.u32 %r261, [%rd60+-8]; setp.ne.s32 %p183, %r261, 12; @%p183 bra $L__BB157_217; ld.local.u64 %rd364, [%rd60+28]; ld.f32 %f1534, [%rd364+36]; mul.ftz.f32 %f1535, %f1534, 0f3E800000; ld.f32 %f1536, [%rd364+24]; mul.ftz.f32 %f1537, %f1536, %f1535; ld.f32 %f1538, [%rd364+28]; mul.ftz.f32 %f1539, %f1535, %f1538; ld.f32 %f1540, [%rd364+32]; mul.ftz.f32 %f1541, %f1535, %f1540; st.f32 [%rd364+24], %f1537; st.f32 [%rd364+28], %f1539; st.f32 [%rd364+32], %f1541; $L__BB157_217: ld.local.u64 %rd365, [%rd60+28]; ld.f32 %f1542, [%rd365+28]; ld.f32 %f1543, [%rd365+24]; add.ftz.f32 %f1544, %f1543, %f1542; ld.f32 %f1545, [%rd365+32]; add.ftz.f32 %f1546, %f1545, %f1544; mul.ftz.f32 %f1547, %f1546, 0f3EAAAAAB; ld.local.f32 %f1548, [%rd60+-4]; mul.ftz.f32 %f1549, %f1548, %f1547; st.local.f32 [%rd60+-4], %f1549; ld.local.v2.f32 {%f1550, %f1551}, [%rd60+12]; mov.f32 %f1554, 0f38D1B717; max.ftz.f32 %f1555, %f1550, %f1554; max.ftz.f32 %f1557, %f1551, %f1554; min.ftz.f32 %f1558, %f1557, %f1528; min.ftz.f32 %f1559, %f1555, %f1528; st.local.v2.f32 [%rd60+12], {%f1559, %f1558}; ld.f32 %f1560, [%rd365]; cvt.ftz.sat.f32.f32 %f1561, %f1560; ld.f32 %f1562, [%rd365+4]; cvt.ftz.sat.f32.f32 %f1563, %f1562; ld.f32 %f1564, [%rd365+8]; cvt.ftz.sat.f32.f32 %f1565, %f1564; st.f32 [%rd365], %f1561; st.f32 [%rd365+4], %f1563; st.f32 [%rd365+8], %f1565; ld.local.u64 %rd366, [%rd60+28]; ld.f32 %f1566, [%rd366+12]; cvt.ftz.sat.f32.f32 %f1567, %f1566; ld.f32 %f1568, [%rd366+16]; cvt.ftz.sat.f32.f32 %f1569, %f1568; ld.f32 %f1570, [%rd366+20]; cvt.ftz.sat.f32.f32 %f1571, %f1570; st.f32 [%rd366+12], %f1567; st.f32 [%rd366+16], %f1569; st.f32 [%rd366+20], %f1571; ld.local.u32 %r262, [%rd4+40]; or.b32 %r263, %r262, 1036; st.local.u32 [%rd4+40], %r263; $L__BB157_225: ld.const.u32 %r48, [kernel_params+1568]; or.b32 %r268, %r48, %r44; setp.eq.s32 %p187, %r268, 0; @%p187 bra $L__BB157_227; setp.gt.ftz.f32 %p188, %f98, 0f3727C5AC; @%p188 bra $L__BB157_228; bra.uni $L__BB157_272; $L__BB157_227: setp.leu.ftz.f32 %p189, %f98, 0f3727C5AC; setp.ne.s32 %p190, %r45, 0; or.pred %p191, %p190, %p189; @%p191 bra $L__BB157_272; $L__BB157_228: mul.ftz.f32 %f279, %f98, %f177; mul.ftz.f32 %f280, %f98, %f178; mul.ftz.f32 %f281, %f98, %f179; mov.f32 %f1640, 0f3F800000; sub.ftz.f32 %f1641, %f1640, %f72; fma.rn.ftz.f32 %f282, %f72, %f186, %f1641; fma.rn.ftz.f32 %f283, %f72, %f187, %f1641; fma.rn.ftz.f32 %f284, %f72, %f188, %f1641; setp.eq.s32 %p192, %r148, 26; setp.le.ftz.f32 %p193, %f71, 0f3D4CCCCD; or.pred %p194, %p192, %p193; @%p194 bra $L__BB157_246; bra.uni $L__BB157_229; $L__BB157_246: @%p158 bra $L__BB157_264; mul.ftz.f32 %f1722, %f2981, %f279; mov.f32 %f1723, 0f00000000; max.ftz.f32 %f302, %f1722, %f1723; mul.ftz.f32 %f1724, %f2981, %f280; max.ftz.f32 %f303, %f1724, %f1723; mul.ftz.f32 %f1725, %f2981, %f281; max.ftz.f32 %f304, %f1725, %f1723; add.ftz.f32 %f1726, %f302, %f303; add.ftz.f32 %f1727, %f304, %f1726; mul.ftz.f32 %f1728, %f1727, 0f3EAAAAAB; abs.ftz.f32 %f305, %f1728; setp.ltu.ftz.f32 %p207, %f305, 0f3727C5AC; mov.u64 %rd732, 0; mov.u64 %rd730, %rd732; mov.u64 %rd731, %rd732; @%p207 bra $L__BB157_252; ld.local.u32 %r51, [%rd4+252]; mov.u64 %rd730, 0; setp.eq.s32 %p208, %r51, 0; mov.u64 %rd728, %rd730; mov.u64 %rd729, %rd730; @%p208 bra $L__BB157_250; ld.local.u32 %r285, [%rd4+248]; mul.wide.s32 %rd401, %r285, 80; add.s64 %rd402, %rd1, %rd401; add.s64 %rd728, %rd402, 304; add.s64 %rd403, %rd230, %rd401; add.s64 %rd729, %rd403, 304; mov.b32 %r286, %f304; mov.b32 %r287, %f303; mov.b32 %r288, %f302; mov.u32 %r289, 0; st.local.v4.u32 [%rd402+304], {%r288, %r287, %r286, %r289}; add.s32 %r290, %r285, 1; st.local.u32 [%rd4+248], %r290; add.s32 %r291, %r51, -1; st.local.u32 [%rd4+252], %r291; $L__BB157_250: setp.eq.s64 %p209, %rd729, 0; mov.u64 %rd731, %rd730; @%p209 bra $L__BB157_252; st.local.f32 [%rd728+16], %f305; mov.u64 %rd730, %rd728; mov.u64 %rd731, %rd729; $L__BB157_252: setp.eq.s64 %p210, %rd731, 0; mov.u64 %rd733, %rd732; @%p210 bra $L__BB157_256; ld.local.u32 %r52, [%rd4+252]; setp.lt.s32 %p211, %r52, 1; @%p211 bra $L__BB157_255; add.s32 %r292, %r52, -1; st.local.u32 [%rd4+252], %r292; ld.local.s32 %rd408, [%rd4+248]; cvt.s64.s32 %rd409, %r292; add.s64 %rd410, %rd408, %rd409; mul.lo.s64 %rd411, %rd410, 80; add.s64 %rd412, %rd1, %rd411; add.s64 %rd732, %rd412, 304; add.s64 %rd413, %rd230, %rd411; add.s64 %rd733, %rd413, 304; bra.uni $L__BB157_256; $L__BB157_229: mov.f32 %f1642, 0f00000000; max.ftz.f32 %f285, %f279, %f1642; max.ftz.f32 %f286, %f280, %f1642; add.ftz.f32 %f1643, %f285, %f286; max.ftz.f32 %f287, %f281, %f1642; add.ftz.f32 %f1644, %f287, %f1643; mul.ftz.f32 %f1645, %f1644, 0f3EAAAAAB; abs.ftz.f32 %f288, %f1645; setp.ltu.ftz.f32 %p195, %f288, 0f3727C5AC; mov.u64 %rd726, 0; mov.u64 %rd724, %rd726; mov.u64 %rd725, %rd726; @%p195 bra $L__BB157_234; ld.local.u32 %r49, [%rd4+252]; mov.u64 %rd724, 0; setp.eq.s32 %p196, %r49, 0; mov.u64 %rd722, %rd724; mov.u64 %rd723, %rd724; @%p196 bra $L__BB157_232; ld.local.u32 %r269, [%rd4+248]; mul.wide.s32 %rd375, %r269, 80; add.s64 %rd376, %rd1, %rd375; add.s64 %rd722, %rd376, 304; add.s64 %rd377, %rd230, %rd375; add.s64 %rd723, %rd377, 304; mov.b32 %r270, %f287; mov.b32 %r271, %f286; mov.b32 %r272, %f285; mov.u32 %r273, 0; st.local.v4.u32 [%rd376+304], {%r272, %r271, %r270, %r273}; add.s32 %r274, %r269, 1; st.local.u32 [%rd4+248], %r274; add.s32 %r275, %r49, -1; st.local.u32 [%rd4+252], %r275; $L__BB157_232: setp.eq.s64 %p197, %rd723, 0; mov.u64 %rd725, %rd724; @%p197 bra $L__BB157_234; st.local.f32 [%rd722+16], %f288; mov.u64 %rd724, %rd722; mov.u64 %rd725, %rd723; $L__BB157_234: setp.eq.s64 %p198, %rd725, 0; mov.u64 %rd727, %rd726; @%p198 bra $L__BB157_238; ld.local.u32 %r50, [%rd4+252]; setp.lt.s32 %p199, %r50, 1; @%p199 bra $L__BB157_237; add.s32 %r276, %r50, -1; st.local.u32 [%rd4+252], %r276; ld.local.s32 %rd382, [%rd4+248]; cvt.s64.s32 %rd383, %r276; add.s64 %rd384, %rd382, %rd383; mul.lo.s64 %rd385, %rd384, 80; add.s64 %rd386, %rd1, %rd385; add.s64 %rd726, %rd386, 304; add.s64 %rd387, %rd230, %rd385; add.s64 %rd727, %rd387, 304; bra.uni $L__BB157_238; $L__BB157_237: ld.local.u32 %r277, [%rd4+248]; add.s32 %r278, %r277, -1; st.local.u32 [%rd4+248], %r278; add.s32 %r279, %r50, 1; st.local.u32 [%rd4+252], %r279; mov.u64 %rd727, %rd726; $L__BB157_238: setp.eq.s64 %p200, %rd727, 0; or.pred %p202, %p198, %p200; @%p202 bra $L__BB157_272; add.s64 %rd73, %rd724, 20; st.local.f32 [%rd724+20], %f55; st.local.f32 [%rd724+24], %f56; st.local.f32 [%rd724+28], %f57; st.local.u64 [%rd724+48], %rd727; mov.u32 %r280, 0; st.local.v2.f32 [%rd724+56], {%f1642, %f1642}; st.local.u32 [%rd724+64], %r280; mul.ftz.f32 %f1648, %f71, %f71; st.local.f32 [%rd724+32], %f1648; st.local.f32 [%rd724+40], %f2980; st.local.f32 [%rd726], %f186; st.local.f32 [%rd726+4], %f187; st.local.f32 [%rd726+8], %f188; ld.local.u64 %rd390, [%rd724+48]; st.f32 [%rd390+12], %f282; st.f32 [%rd390+16], %f283; st.f32 [%rd390+20], %f284; ld.local.u64 %rd391, [%rd724+48]; st.u32 [%rd391+36], %r280; ld.local.f32 %f1649, [%rd724+32]; mov.f32 %f1650, 0f38D1B717; max.ftz.f32 %f1651, %f1649, %f1650; min.ftz.f32 %f1652, %f1651, %f1640; st.local.f32 [%rd724+32], %f1652; st.local.f32 [%rd724+36], %f1652; ld.local.f32 %f1653, [%rd724+40]; max.ftz.f32 %f1654, %f1642, %f1653; st.local.f32 [%rd724+40], %f1654; ld.local.u64 %rd392, [%rd724+48]; ld.f32 %f1655, [%rd392]; cvt.ftz.sat.f32.f32 %f1656, %f1655; ld.f32 %f1657, [%rd392+4]; cvt.ftz.sat.f32.f32 %f1658, %f1657; ld.f32 %f1659, [%rd392+8]; cvt.ftz.sat.f32.f32 %f1660, %f1659; st.f32 [%rd392], %f1656; st.f32 [%rd392+4], %f1658; st.f32 [%rd392+8], %f1660; ld.local.u64 %rd393, [%rd724+48]; ld.f32 %f1661, [%rd393+12]; cvt.ftz.sat.f32.f32 %f1662, %f1661; ld.f32 %f1663, [%rd393+16]; cvt.ftz.sat.f32.f32 %f1664, %f1663; ld.f32 %f1665, [%rd393+20]; cvt.ftz.sat.f32.f32 %f1666, %f1665; st.f32 [%rd393+12], %f1662; st.f32 [%rd393+16], %f1664; st.f32 [%rd393+20], %f1666; mov.u32 %r281, 27; st.local.u32 [%rd724+12], %r281; ld.local.f32 %f1667, [%rd724+40]; fma.rn.ftz.f32 %f289, %f1667, %f1667, 0fBF800000; abs.ftz.f32 %f290, %f1640; fma.rn.ftz.f32 %f291, %f290, %f290, %f289; setp.leu.ftz.f32 %p203, %f291, 0f00000000; mov.f32 %f3021, %f1640; @%p203 bra $L__BB157_241; sqrt.approx.ftz.f32 %f1668, %f291; sub.ftz.f32 %f1669, %f1668, %f290; add.ftz.f32 %f1670, %f290, %f1668; div.approx.ftz.f32 %f1671, %f1669, %f1670; fma.rn.ftz.f32 %f1672, %f290, %f1670, 0fBF800000; fma.rn.ftz.f32 %f1673, %f290, %f1669, 0f3F800000; div.approx.ftz.f32 %f1674, %f1672, %f1673; mul.ftz.f32 %f1675, %f1671, 0f3F000000; mul.ftz.f32 %f1676, %f1671, %f1675; fma.rn.ftz.f32 %f1677, %f1674, %f1674, 0f3F800000; mul.ftz.f32 %f3021, %f1676, %f1677; $L__BB157_241: ld.local.u64 %rd394, [%rd73+28]; add.s64 %rd74, %rd394, 12; ld.f32 %f294, [%rd394+12]; ld.f32 %f295, [%rd394+16]; ld.f32 %f296, [%rd394+20]; sub.ftz.f32 %f297, %f1640, %f3021; ld.local.f32 %f1679, [%rd73]; ld.local.f32 %f1680, [%rd4+24]; ld.local.f32 %f1681, [%rd73+4]; ld.local.f32 %f1682, [%rd4+28]; mul.ftz.f32 %f1683, %f1682, %f1681; fma.rn.ftz.f32 %f1684, %f1680, %f1679, %f1683; ld.local.f32 %f1685, [%rd73+8]; ld.local.f32 %f1686, [%rd4+32]; fma.rn.ftz.f32 %f1687, %f1686, %f1685, %f1684; abs.ftz.f32 %f298, %f1687; fma.rn.ftz.f32 %f299, %f298, %f298, %f289; setp.leu.ftz.f32 %p204, %f299, 0f00000000; mov.f32 %f3022, %f297; @%p204 bra $L__BB157_243; sqrt.approx.ftz.f32 %f1688, %f299; sub.ftz.f32 %f1689, %f1688, %f298; add.ftz.f32 %f1690, %f298, %f1688; div.approx.ftz.f32 %f1691, %f1689, %f1690; fma.rn.ftz.f32 %f1692, %f298, %f1690, 0fBF800000; fma.rn.ftz.f32 %f1693, %f298, %f1689, 0f3F800000; div.approx.ftz.f32 %f1694, %f1692, %f1693; mul.ftz.f32 %f1695, %f1691, 0f3F000000; mul.ftz.f32 %f1696, %f1691, %f1695; fma.rn.ftz.f32 %f1697, %f1694, %f1694, 0f3F800000; mul.ftz.f32 %f1698, %f1696, %f1697; sub.ftz.f32 %f3022, %f1698, %f3021; $L__BB157_243: rcp.approx.ftz.f32 %f1699, %f297; mov.f32 %f1700, 0f3F800000; mul.ftz.f32 %f1701, %f1699, %f3022; sub.ftz.f32 %f1702, %f1700, %f1701; fma.rn.ftz.f32 %f1703, %f294, %f1702, %f1701; fma.rn.ftz.f32 %f1704, %f295, %f1702, %f1701; fma.rn.ftz.f32 %f1705, %f296, %f1702, %f1701; st.f32 [%rd74+12], %f1703; st.f32 [%rd74+16], %f1704; st.f32 [%rd74+20], %f1705; ld.local.u32 %r282, [%rd73+-8]; setp.ne.s32 %p205, %r282, 12; @%p205 bra $L__BB157_245; ld.local.u64 %rd395, [%rd73+28]; ld.f32 %f1706, [%rd395+36]; mul.ftz.f32 %f1707, %f1706, 0f3E800000; ld.f32 %f1708, [%rd395+24]; mul.ftz.f32 %f1709, %f1708, %f1707; ld.f32 %f1710, [%rd395+28]; mul.ftz.f32 %f1711, %f1707, %f1710; ld.f32 %f1712, [%rd395+32]; mul.ftz.f32 %f1713, %f1707, %f1712; st.f32 [%rd395+24], %f1709; st.f32 [%rd395+28], %f1711; st.f32 [%rd395+32], %f1713; $L__BB157_245: ld.local.u64 %rd396, [%rd73+28]; ld.f32 %f1714, [%rd396+28]; ld.f32 %f1715, [%rd396+24]; add.ftz.f32 %f1716, %f1715, %f1714; ld.f32 %f1717, [%rd396+32]; add.ftz.f32 %f1718, %f1717, %f1716; mul.ftz.f32 %f1719, %f1718, 0f3EAAAAAB; ld.local.f32 %f1720, [%rd73+-4]; mul.ftz.f32 %f1721, %f1720, %f1719; st.local.f32 [%rd73+-4], %f1721; ld.local.u32 %r283, [%rd4+40]; or.b32 %r284, %r283, 1036; st.local.u32 [%rd4+40], %r284; bra.uni $L__BB157_272; $L__BB157_255: ld.local.u32 %r293, [%rd4+248]; add.s32 %r294, %r293, -1; st.local.u32 [%rd4+248], %r294; add.s32 %r295, %r52, 1; st.local.u32 [%rd4+252], %r295; mov.u64 %rd733, %rd732; $L__BB157_256: setp.eq.s64 %p212, %rd733, 0; or.pred %p214, %p210, %p212; @%p214 bra $L__BB157_264; add.s64 %rd85, %rd730, 20; st.local.f32 [%rd730+20], %f55; st.local.f32 [%rd730+24], %f56; st.local.f32 [%rd730+28], %f57; mov.u32 %r296, 0; st.local.v2.f32 [%rd730+56], {%f1723, %f1723}; st.local.u32 [%rd730+64], %r296; st.local.u64 [%rd730+48], %rd733; mul.ftz.f32 %f1731, %f71, %f71; st.local.v2.f32 [%rd730+32], {%f1731, %f1731}; st.local.f32 [%rd730+40], %f2980; st.local.f32 [%rd732], %f186; st.local.f32 [%rd732+4], %f187; st.local.f32 [%rd732+8], %f188; ld.local.u64 %rd416, [%rd730+48]; st.f32 [%rd416+12], %f282; st.f32 [%rd416+16], %f283; st.f32 [%rd416+20], %f284; ld.local.u64 %rd417, [%rd730+48]; st.u32 [%rd417+36], %r296; ld.local.u64 %rd418, [%rd730+48]; ld.f32 %f1732, [%rd418+12]; cvt.ftz.sat.f32.f32 %f1733, %f1732; ld.f32 %f1734, [%rd418+16]; cvt.ftz.sat.f32.f32 %f1735, %f1734; ld.f32 %f1736, [%rd418+20]; cvt.ftz.sat.f32.f32 %f1737, %f1736; st.f32 [%rd418+12], %f1733; st.f32 [%rd418+16], %f1735; st.f32 [%rd418+20], %f1737; ld.local.v2.f32 {%f1738, %f1739}, [%rd730+32]; cvt.ftz.sat.f32.f32 %f1742, %f1739; cvt.ftz.sat.f32.f32 %f1743, %f1738; st.local.v2.f32 [%rd730+32], {%f1743, %f1742}; mov.u32 %r297, 11; st.local.u32 [%rd730+12], %r297; ld.local.f32 %f1744, [%rd730+40]; fma.rn.ftz.f32 %f306, %f1744, %f1744, 0fBF800000; mov.f32 %f1729, 0f3F800000; abs.ftz.f32 %f307, %f1729; fma.rn.ftz.f32 %f308, %f307, %f307, %f306; setp.leu.ftz.f32 %p215, %f308, 0f00000000; mov.f32 %f3023, %f1729; @%p215 bra $L__BB157_259; sqrt.approx.ftz.f32 %f1745, %f308; sub.ftz.f32 %f1746, %f1745, %f307; add.ftz.f32 %f1747, %f307, %f1745; div.approx.ftz.f32 %f1748, %f1746, %f1747; fma.rn.ftz.f32 %f1749, %f307, %f1747, 0fBF800000; fma.rn.ftz.f32 %f1750, %f307, %f1746, 0f3F800000; div.approx.ftz.f32 %f1751, %f1749, %f1750; mul.ftz.f32 %f1752, %f1748, 0f3F000000; mul.ftz.f32 %f1753, %f1748, %f1752; fma.rn.ftz.f32 %f1754, %f1751, %f1751, 0f3F800000; mul.ftz.f32 %f3023, %f1753, %f1754; $L__BB157_259: ld.local.u64 %rd419, [%rd85+28]; add.s64 %rd86, %rd419, 12; ld.f32 %f311, [%rd419+12]; ld.f32 %f312, [%rd419+16]; ld.f32 %f313, [%rd419+20]; sub.ftz.f32 %f314, %f1729, %f3023; ld.local.f32 %f1756, [%rd85]; ld.local.f32 %f1757, [%rd4+24]; ld.local.f32 %f1758, [%rd85+4]; ld.local.f32 %f1759, [%rd4+28]; mul.ftz.f32 %f1760, %f1759, %f1758; fma.rn.ftz.f32 %f1761, %f1757, %f1756, %f1760; ld.local.f32 %f1762, [%rd85+8]; ld.local.f32 %f1763, [%rd4+32]; fma.rn.ftz.f32 %f1764, %f1763, %f1762, %f1761; abs.ftz.f32 %f315, %f1764; fma.rn.ftz.f32 %f316, %f315, %f315, %f306; setp.leu.ftz.f32 %p216, %f316, 0f00000000; mov.f32 %f3024, %f314; @%p216 bra $L__BB157_261; sqrt.approx.ftz.f32 %f1765, %f316; sub.ftz.f32 %f1766, %f1765, %f315; add.ftz.f32 %f1767, %f315, %f1765; div.approx.ftz.f32 %f1768, %f1766, %f1767; fma.rn.ftz.f32 %f1769, %f315, %f1767, 0fBF800000; fma.rn.ftz.f32 %f1770, %f315, %f1766, 0f3F800000; div.approx.ftz.f32 %f1771, %f1769, %f1770; mul.ftz.f32 %f1772, %f1768, 0f3F000000; mul.ftz.f32 %f1773, %f1768, %f1772; fma.rn.ftz.f32 %f1774, %f1771, %f1771, 0f3F800000; mul.ftz.f32 %f1775, %f1773, %f1774; sub.ftz.f32 %f3024, %f1775, %f3023; $L__BB157_261: rcp.approx.ftz.f32 %f1776, %f314; mov.f32 %f1777, 0f3F800000; mul.ftz.f32 %f1778, %f1776, %f3024; sub.ftz.f32 %f1779, %f1777, %f1778; fma.rn.ftz.f32 %f1780, %f311, %f1779, %f1778; fma.rn.ftz.f32 %f1781, %f312, %f1779, %f1778; fma.rn.ftz.f32 %f1782, %f313, %f1779, %f1778; st.f32 [%rd86+12], %f1780; st.f32 [%rd86+16], %f1781; st.f32 [%rd86+20], %f1782; ld.local.u32 %r298, [%rd85+-8]; setp.ne.s32 %p217, %r298, 12; @%p217 bra $L__BB157_263; ld.local.u64 %rd420, [%rd85+28]; ld.f32 %f1783, [%rd420+36]; mul.ftz.f32 %f1784, %f1783, 0f3E800000; ld.f32 %f1785, [%rd420+24]; mul.ftz.f32 %f1786, %f1785, %f1784; ld.f32 %f1787, [%rd420+28]; mul.ftz.f32 %f1788, %f1784, %f1787; ld.f32 %f1789, [%rd420+32]; mul.ftz.f32 %f1790, %f1784, %f1789; st.f32 [%rd420+24], %f1786; st.f32 [%rd420+28], %f1788; st.f32 [%rd420+32], %f1790; $L__BB157_263: ld.local.u64 %rd421, [%rd85+28]; ld.f32 %f1791, [%rd421+28]; ld.f32 %f1792, [%rd421+24]; add.ftz.f32 %f1793, %f1792, %f1791; ld.f32 %f1794, [%rd421+32]; add.ftz.f32 %f1795, %f1794, %f1793; mul.ftz.f32 %f1796, %f1795, 0f3EAAAAAB; ld.local.f32 %f1797, [%rd85+-4]; mul.ftz.f32 %f1798, %f1797, %f1796; st.local.f32 [%rd85+-4], %f1798; ld.local.u32 %r299, [%rd4+40]; or.b32 %r300, %r299, 12; st.local.u32 [%rd4+40], %r300; $L__BB157_264: setp.ne.s32 %p218, %r45, 0; setp.eq.s32 %p219, %r48, 0; and.pred %p220, %p218, %p219; @%p220 bra $L__BB157_272; mov.f32 %f1799, 0f3F800000; sub.ftz.f32 %f1800, %f1799, %f2981; mov.f32 %f1801, 0f38D1B717; max.ftz.f32 %f1802, %f1801, %f1800; mul.ftz.f32 %f1803, %f186, %f279; mul.ftz.f32 %f1804, %f1803, %f1802; mul.ftz.f32 %f1805, %f187, %f280; mul.ftz.f32 %f1806, %f1805, %f1802; mul.ftz.f32 %f1807, %f188, %f281; mul.ftz.f32 %f1808, %f1807, %f1802; mov.f32 %f1809, 0f00000000; max.ftz.f32 %f319, %f1804, %f1809; max.ftz.f32 %f320, %f1806, %f1809; max.ftz.f32 %f321, %f1808, %f1809; add.ftz.f32 %f1810, %f319, %f320; add.ftz.f32 %f1811, %f321, %f1810; mul.ftz.f32 %f1812, %f1811, 0f3EAAAAAB; abs.ftz.f32 %f322, %f1812; setp.ltu.ftz.f32 %p221, %f322, 0f3727C5AC; mov.u64 %rd736, 0; mov.u64 %rd737, %rd736; @%p221 bra $L__BB157_270; ld.local.u32 %r53, [%rd4+252]; mov.u64 %rd736, 0; setp.eq.s32 %p222, %r53, 0; mov.u64 %rd734, %rd736; mov.u64 %rd735, %rd736; @%p222 bra $L__BB157_268; ld.local.u32 %r301, [%rd4+248]; mul.wide.s32 %rd426, %r301, 80; add.s64 %rd427, %rd1, %rd426; add.s64 %rd734, %rd427, 304; add.s64 %rd428, %rd230, %rd426; add.s64 %rd735, %rd428, 304; mov.b32 %r302, %f321; mov.b32 %r303, %f320; mov.b32 %r304, %f319; mov.u32 %r305, 0; st.local.v4.u32 [%rd427+304], {%r304, %r303, %r302, %r305}; add.s32 %r306, %r301, 1; st.local.u32 [%rd4+248], %r306; add.s32 %r307, %r53, -1; st.local.u32 [%rd4+252], %r307; $L__BB157_268: setp.eq.s64 %p223, %rd735, 0; mov.u64 %rd737, %rd736; @%p223 bra $L__BB157_270; st.local.f32 [%rd734+16], %f322; mov.u64 %rd736, %rd734; mov.u64 %rd737, %rd735; $L__BB157_270: setp.eq.s64 %p224, %rd737, 0; @%p224 bra $L__BB157_272; mov.u64 %rd431, 0; st.local.f32 [%rd736+20], %f55; st.local.f32 [%rd736+24], %f56; st.local.f32 [%rd736+28], %f57; mov.u32 %r308, 0; st.local.v2.f32 [%rd736+56], {%f1809, %f1809}; st.local.u32 [%rd736+64], %r308; sub.ftz.f32 %f1815, %f1799, %f79; sub.ftz.f32 %f1816, %f1799, %f71; mul.ftz.f32 %f1817, %f1816, %f1815; sub.ftz.f32 %f1818, %f1799, %f1817; selp.f32 %f1819, %f1818, %f71, %p192; mul.ftz.f32 %f1820, %f1819, %f1819; st.local.f32 [%rd736+40], %f2980; st.local.u64 [%rd736+48], %rd431; cvt.ftz.sat.f32.f32 %f1821, %f1820; st.local.v2.f32 [%rd736+32], {%f1821, %f1821}; mov.u32 %r309, 23; st.local.u32 [%rd736+12], %r309; ld.local.u32 %r310, [%rd4+40]; or.b32 %r311, %r310, 12; st.local.u32 [%rd4+40], %r311; $L__BB157_272: ld.param.u32 %r638, [_Z21svm_node_closure_bsdfILj33554879EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; add.s32 %r663, %r638, 5; setp.leu.ftz.f32 %p226, %f76, 0f3727C5AC; or.pred %p228, %p226, %p158; @%p228 bra $L__BB157_517; mov.f32 %f1822, 0f00000000; max.ftz.f32 %f323, %f177, %f1822; max.ftz.f32 %f324, %f178, %f1822; add.ftz.f32 %f1823, %f323, %f324; max.ftz.f32 %f325, %f179, %f1822; add.ftz.f32 %f1824, %f325, %f1823; mul.ftz.f32 %f1825, %f1824, 0f3EAAAAAB; abs.ftz.f32 %f326, %f1825; setp.ltu.ftz.f32 %p229, %f326, 0f3727C5AC; mov.u64 %rd742, 0; mov.u64 %rd740, %rd742; mov.u64 %rd741, %rd742; @%p229 bra $L__BB157_278; ld.local.u32 %r55, [%rd4+252]; mov.u64 %rd740, 0; setp.eq.s32 %p230, %r55, 0; mov.u64 %rd738, %rd740; mov.u64 %rd739, %rd740; @%p230 bra $L__BB157_276; ld.local.u32 %r312, [%rd4+248]; mul.wide.s32 %rd436, %r312, 80; add.s64 %rd437, %rd1, %rd436; add.s64 %rd738, %rd437, 304; add.s64 %rd438, %rd230, %rd436; add.s64 %rd739, %rd438, 304; mov.b32 %r313, %f325; mov.b32 %r314, %f324; mov.b32 %r315, %f323; mov.u32 %r316, 0; st.local.v4.u32 [%rd437+304], {%r315, %r314, %r313, %r316}; add.s32 %r317, %r312, 1; st.local.u32 [%rd4+248], %r317; add.s32 %r318, %r55, -1; st.local.u32 [%rd4+252], %r318; $L__BB157_276: setp.eq.s64 %p231, %rd739, 0; mov.u64 %rd741, %rd740; @%p231 bra $L__BB157_278; st.local.f32 [%rd738+16], %f326; mov.u64 %rd740, %rd738; mov.u64 %rd741, %rd739; $L__BB157_278: setp.eq.s64 %p232, %rd741, 0; mov.u64 %rd743, %rd742; @%p232 bra $L__BB157_282; ld.local.u32 %r56, [%rd4+252]; setp.lt.s32 %p233, %r56, 1; @%p233 bra $L__BB157_281; add.s32 %r319, %r56, -1; st.local.u32 [%rd4+252], %r319; ld.local.s32 %rd443, [%rd4+248]; cvt.s64.s32 %rd444, %r319; add.s64 %rd445, %rd443, %rd444; mul.lo.s64 %rd446, %rd445, 80; add.s64 %rd447, %rd1, %rd446; add.s64 %rd742, %rd447, 304; add.s64 %rd448, %rd230, %rd446; add.s64 %rd743, %rd448, 304; bra.uni $L__BB157_282; $L__BB157_281: ld.local.u32 %r320, [%rd4+248]; add.s32 %r321, %r320, -1; st.local.u32 [%rd4+248], %r321; add.s32 %r322, %r56, 1; st.local.u32 [%rd4+252], %r322; mov.u64 %rd743, %rd742; $L__BB157_282: setp.eq.s64 %p234, %rd743, 0; or.pred %p236, %p232, %p234; @%p236 bra $L__BB157_517; mov.f32 %f3025, 0f3F800000; add.s64 %rd103, %rd740, 20; st.local.f32 [%rd740+20], %f2990; st.local.f32 [%rd740+24], %f2991; st.local.f32 [%rd740+28], %f2992; mov.u32 %r323, 0; st.local.v2.f32 [%rd740+56], {%f1822, %f1822}; st.local.u32 [%rd740+64], %r323; mov.u32 %r324, 1069547520; st.local.u32 [%rd740+40], %r324; st.local.u64 [%rd740+48], %rd743; mul.ftz.f32 %f1828, %f77, %f77; st.local.v2.f32 [%rd740+32], {%f1828, %f1828}; st.local.u32 [%rd742], %r323; st.local.u32 [%rd742+4], %r323; st.local.u32 [%rd742+8], %r323; ld.local.u64 %rd451, [%rd740+48]; mov.u32 %r325, 1025758986; st.u32 [%rd451+12], %r325; st.u32 [%rd451+16], %r325; st.u32 [%rd451+20], %r325; ld.local.u64 %rd452, [%rd740+48]; st.f32 [%rd452+36], %f76; ld.local.u64 %rd453, [%rd740+48]; ld.f32 %f1829, [%rd453+12]; cvt.ftz.sat.f32.f32 %f1830, %f1829; ld.f32 %f1831, [%rd453+16]; cvt.ftz.sat.f32.f32 %f1832, %f1831; ld.f32 %f1833, [%rd453+20]; cvt.ftz.sat.f32.f32 %f1834, %f1833; st.f32 [%rd453+12], %f1830; st.f32 [%rd453+16], %f1832; st.f32 [%rd453+20], %f1834; ld.local.f32 %f1835, [%rd740+32]; cvt.ftz.sat.f32.f32 %f1836, %f1835; st.local.v2.f32 [%rd740+32], {%f1836, %f1836}; mov.u32 %r326, 12; st.local.u32 [%rd740+12], %r326; ld.local.f32 %f1837, [%rd740+40]; fma.rn.ftz.f32 %f327, %f1837, %f1837, 0fBF800000; abs.ftz.f32 %f328, %f3025; fma.rn.ftz.f32 %f329, %f328, %f328, %f327; setp.leu.ftz.f32 %p237, %f329, 0f00000000; @%p237 bra $L__BB157_285; sqrt.approx.ftz.f32 %f1838, %f329; sub.ftz.f32 %f1839, %f1838, %f328; add.ftz.f32 %f1840, %f328, %f1838; div.approx.ftz.f32 %f1841, %f1839, %f1840; fma.rn.ftz.f32 %f1842, %f328, %f1840, 0fBF800000; fma.rn.ftz.f32 %f1843, %f328, %f1839, 0f3F800000; div.approx.ftz.f32 %f1844, %f1842, %f1843; mul.ftz.f32 %f1845, %f1841, 0f3F000000; mul.ftz.f32 %f1846, %f1841, %f1845; fma.rn.ftz.f32 %f1847, %f1844, %f1844, 0f3F800000; mul.ftz.f32 %f3025, %f1846, %f1847; $L__BB157_285: mov.f32 %f2960, 0f3F800000; ld.local.u64 %rd454, [%rd103+28]; add.s64 %rd104, %rd454, 12; ld.f32 %f332, [%rd454+12]; ld.f32 %f333, [%rd454+16]; ld.f32 %f334, [%rd454+20]; sub.ftz.f32 %f335, %f2960, %f3025; ld.local.f32 %f1849, [%rd103]; ld.local.f32 %f1850, [%rd4+24]; ld.local.f32 %f1851, [%rd103+4]; ld.local.f32 %f1852, [%rd4+28]; mul.ftz.f32 %f1853, %f1852, %f1851; fma.rn.ftz.f32 %f1854, %f1850, %f1849, %f1853; ld.local.f32 %f1855, [%rd103+8]; ld.local.f32 %f1856, [%rd4+32]; fma.rn.ftz.f32 %f1857, %f1856, %f1855, %f1854; abs.ftz.f32 %f336, %f1857; fma.rn.ftz.f32 %f337, %f336, %f336, %f327; setp.leu.ftz.f32 %p238, %f337, 0f00000000; mov.f32 %f3026, %f335; @%p238 bra $L__BB157_287; sqrt.approx.ftz.f32 %f1858, %f337; sub.ftz.f32 %f1859, %f1858, %f336; add.ftz.f32 %f1860, %f336, %f1858; div.approx.ftz.f32 %f1861, %f1859, %f1860; fma.rn.ftz.f32 %f1862, %f336, %f1860, 0fBF800000; fma.rn.ftz.f32 %f1863, %f336, %f1859, 0f3F800000; div.approx.ftz.f32 %f1864, %f1862, %f1863; mul.ftz.f32 %f1865, %f1861, 0f3F000000; mul.ftz.f32 %f1866, %f1861, %f1865; fma.rn.ftz.f32 %f1867, %f1864, %f1864, 0f3F800000; mul.ftz.f32 %f1868, %f1866, %f1867; sub.ftz.f32 %f3026, %f1868, %f3025; $L__BB157_287: rcp.approx.ftz.f32 %f1869, %f335; mov.f32 %f1870, 0f3F800000; mul.ftz.f32 %f1871, %f1869, %f3026; sub.ftz.f32 %f1872, %f1870, %f1871; fma.rn.ftz.f32 %f1873, %f332, %f1872, %f1871; fma.rn.ftz.f32 %f1874, %f333, %f1872, %f1871; fma.rn.ftz.f32 %f1875, %f334, %f1872, %f1871; st.f32 [%rd104+12], %f1873; st.f32 [%rd104+16], %f1874; st.f32 [%rd104+20], %f1875; ld.local.u32 %r327, [%rd103+-8]; setp.ne.s32 %p239, %r327, 12; @%p239 bra $L__BB157_289; ld.local.u64 %rd455, [%rd103+28]; ld.f32 %f1876, [%rd455+36]; mul.ftz.f32 %f1877, %f1876, 0f3E800000; ld.f32 %f1878, [%rd455+24]; mul.ftz.f32 %f1879, %f1878, %f1877; ld.f32 %f1880, [%rd455+28]; mul.ftz.f32 %f1881, %f1877, %f1880; ld.f32 %f1882, [%rd455+32]; mul.ftz.f32 %f1883, %f1877, %f1882; st.f32 [%rd455+24], %f1879; st.f32 [%rd455+28], %f1881; st.f32 [%rd455+32], %f1883; $L__BB157_289: ld.local.u64 %rd456, [%rd103+28]; ld.f32 %f1884, [%rd456+28]; ld.f32 %f1885, [%rd456+24]; add.ftz.f32 %f1886, %f1885, %f1884; ld.f32 %f1887, [%rd456+32]; add.ftz.f32 %f1888, %f1887, %f1886; mul.ftz.f32 %f1889, %f1888, 0f3EAAAAAB; ld.local.f32 %f1890, [%rd103+-4]; mul.ftz.f32 %f1891, %f1890, %f1889; st.local.f32 [%rd103+-4], %f1891; ld.local.u32 %r328, [%rd4+40]; or.b32 %r329, %r328, 12; st.local.u32 [%rd4+40], %r329; bra.uni $L__BB157_517; } .func _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB158_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB158_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB158_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB158_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB158_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB158_11; $L__BB158_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB158_8; add.s32 %r237, %r237, 2; bra.uni $L__BB158_9; $L__BB158_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB158_15; $L__BB158_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB158_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB158_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB158_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB158_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB158_15; $L__BB158_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB158_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB158_17; $L__BB158_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB158_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB158_159; bra.uni $L__BB158_18; $L__BB158_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB158_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB158_185; $L__BB158_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB158_150; bra.uni $L__BB158_19; $L__BB158_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB158_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB158_153; bra.uni $L__BB158_152; $L__BB158_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB158_154; $L__BB158_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB158_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB158_100; bra.uni $L__BB158_21; $L__BB158_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB158_111; bra.uni $L__BB158_101; $L__BB158_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB158_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB158_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB158_120; $L__BB158_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB158_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB158_164; bra.uni $L__BB158_163; $L__BB158_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB158_165; $L__BB158_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB158_136; bra.uni $L__BB158_126; $L__BB158_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB158_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB158_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB158_145; $L__BB158_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB158_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB158_158; bra.uni $L__BB158_155; $L__BB158_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB158_192; $L__BB158_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB158_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB158_192; $L__BB158_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB158_47; bra.uni $L__BB158_22; $L__BB158_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB158_86; bra.uni $L__BB158_48; $L__BB158_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB158_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB158_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB158_95; $L__BB158_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB158_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB158_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB158_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB158_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB158_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB158_184; bra.uni $L__BB158_170; $L__BB158_184: { // callseq 929, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 929 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB158_185; $L__BB158_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB158_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB158_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 923, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 923 bra.uni $L__BB158_145; $L__BB158_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB158_192; $L__BB158_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB158_183; bra.uni $L__BB158_171; $L__BB158_183: { // callseq 928, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 928 mov.f32 %f731, 0f3F800000; bra.uni $L__BB158_185; $L__BB158_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB158_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB158_104; mov.u64 %rd173, 0; { // callseq 922, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 922 bra.uni $L__BB158_120; $L__BB158_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB158_33; bra.uni $L__BB158_23; $L__BB158_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB158_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB158_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB158_42; $L__BB158_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB158_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB158_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB158_133; $L__BB158_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB158_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB158_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB158_108; $L__BB158_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB158_182; bra.uni $L__BB158_172; $L__BB158_182: { // callseq 927, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 927 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB158_185; $L__BB158_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB158_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB158_51; mov.u64 %rd129, 0; { // callseq 921, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 921 bra.uni $L__BB158_95; $L__BB158_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB158_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB158_26; mov.u64 %rd90, 0; { // callseq 920, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 920 bra.uni $L__BB158_42; $L__BB158_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB158_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB158_145; $L__BB158_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB158_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB158_120; $L__BB158_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB158_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB158_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB158_83; $L__BB158_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB158_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB158_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB158_30; $L__BB158_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB158_181; bra.uni $L__BB158_173; $L__BB158_181: { // callseq 926, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 926 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB158_185; $L__BB158_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB158_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB158_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB158_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB158_145; $L__BB158_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB158_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB158_145; $L__BB158_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB158_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB158_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB158_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB158_120; $L__BB158_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB158_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB158_120; $L__BB158_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB158_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB158_95; $L__BB158_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB158_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB158_42; $L__BB158_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB158_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB158_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB158_149; bra.uni $L__BB158_146; $L__BB158_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB158_192; $L__BB158_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB158_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB158_192; $L__BB158_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB158_192; $L__BB158_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB158_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB158_175; bra.uni $L__BB158_178; $L__BB158_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB158_177; bra.uni $L__BB158_176; $L__BB158_177: { // callseq 924, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 924 bra.uni $L__BB158_185; $L__BB158_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB158_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB158_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB158_124; bra.uni $L__BB158_121; $L__BB158_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB158_192; $L__BB158_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB158_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB158_192; $L__BB158_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB158_192; $L__BB158_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB158_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB158_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB158_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB158_95; $L__BB158_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB158_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB158_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB158_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB158_42; $L__BB158_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB158_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB158_95; $L__BB158_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB158_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB158_42; $L__BB158_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB158_82; bra.uni $L__BB158_55; $L__BB158_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB158_83; $L__BB158_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB158_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB158_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB158_46; bra.uni $L__BB158_43; $L__BB158_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB158_192; $L__BB158_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB158_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB158_192; $L__BB158_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB158_192; $L__BB158_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB158_185; $L__BB158_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB158_57; bra.uni $L__BB158_56; $L__BB158_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB158_58; $L__BB158_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB158_180; bra.uni $L__BB158_179; $L__BB158_180: { // callseq 925, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 925 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB158_185; $L__BB158_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB158_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB158_60; bra.uni $L__BB158_59; $L__BB158_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB158_61; $L__BB158_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB158_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB158_63; bra.uni $L__BB158_62; $L__BB158_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB158_64; $L__BB158_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB158_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB158_66; bra.uni $L__BB158_65; $L__BB158_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB158_67; $L__BB158_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB158_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB158_69; bra.uni $L__BB158_68; $L__BB158_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB158_70; $L__BB158_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB158_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB158_72; bra.uni $L__BB158_71; $L__BB158_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB158_73; $L__BB158_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB158_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB158_75; bra.uni $L__BB158_74; $L__BB158_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB158_76; $L__BB158_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB158_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB158_78; bra.uni $L__BB158_77; $L__BB158_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB158_79; $L__BB158_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB158_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB158_81; bra.uni $L__BB158_80; $L__BB158_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB158_83; $L__BB158_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB158_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB158_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB158_99; bra.uni $L__BB158_96; $L__BB158_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB158_192; $L__BB158_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB158_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB158_192; $L__BB158_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB158_192; $L__BB158_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB158_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB158_191; bra.uni $L__BB158_186; $L__BB158_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB158_192; $L__BB158_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB158_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB158_192; $L__BB158_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB158_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB158_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB158_192: ret; } .func _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB159_2; bra.uni $L__BB159_1; $L__BB159_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB159_3; $L__BB159_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB159_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB159_5; bra.uni $L__BB159_4; $L__BB159_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB159_9; bra.uni $L__BB159_6; $L__BB159_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB159_12; bra.uni $L__BB159_10; $L__BB159_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB159_13; $L__BB159_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB159_16; $L__BB159_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB159_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB159_13; $L__BB159_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB159_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB159_15; bra.uni $L__BB159_14; $L__BB159_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB159_16; $L__BB159_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB159_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB159_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB159_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB159_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB160_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB160_5; bra.uni $L__BB160_2; $L__BB160_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB160_8; bra.uni $L__BB160_6; $L__BB160_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB160_9; $L__BB160_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB160_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB160_9; $L__BB160_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB160_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB160_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB160_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB160_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB160_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB160_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB160_20; $L__BB160_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB160_17; add.s32 %r172, %r172, 2; bra.uni $L__BB160_18; $L__BB160_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB160_24; $L__BB160_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB160_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB160_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB160_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB160_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB160_24; $L__BB160_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB160_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB160_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB160_37; bra.uni $L__BB160_26; $L__BB160_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB160_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB160_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB160_45; $L__BB160_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB160_45; $L__BB160_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB160_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB160_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB160_30; mov.u64 %rd108, 0; { // callseq 930, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 930 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB160_45; $L__BB160_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB160_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB160_9; $L__BB160_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB160_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB160_45; $L__BB160_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB160_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB160_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB160_34; $L__BB160_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB160_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB160_45; $L__BB160_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB160_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB160_45; $L__BB160_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB160_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB160_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB160_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB160_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB160_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB160_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB160_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB160_56; $L__BB160_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB160_53; add.s32 %r187, %r187, 2; bra.uni $L__BB160_54; $L__BB160_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB160_60; $L__BB160_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB160_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB160_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB160_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB160_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB160_60; $L__BB160_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB160_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB160_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB160_72; bra.uni $L__BB160_62; $L__BB160_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB160_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB160_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB160_79; $L__BB160_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB160_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB160_65; mov.u64 %rd160, 0; { // callseq 931, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 931 bra.uni $L__BB160_79; $L__BB160_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB160_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB160_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB160_69; $L__BB160_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB160_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB160_79; $L__BB160_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB160_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB160_79; $L__BB160_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB160_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB160_79; $L__BB160_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB160_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB160_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB160_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB160_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB160_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB160_84; bra.uni $L__BB160_83; $L__BB160_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB160_85; $L__BB160_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB160_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj33554879EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB161_2; bra.uni $L__BB161_1; $L__BB161_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB161_3; $L__BB161_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB161_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB161_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB161_8; bra.uni $L__BB161_5; $L__BB161_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB161_11; bra.uni $L__BB161_9; $L__BB161_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB161_12; $L__BB161_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB161_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB161_12; $L__BB161_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB161_12: @%p8 bra $L__BB161_14; bra.uni $L__BB161_13; $L__BB161_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB161_15; $L__BB161_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB161_15: @%p8 bra $L__BB161_17; bra.uni $L__BB161_16; $L__BB161_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB161_18; $L__BB161_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB161_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB161_19; bra.uni $L__BB161_20; $L__BB161_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB161_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB161_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB161_22: @%p7 bra $L__BB161_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB161_25; bra.uni $L__BB161_24; $L__BB161_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB161_26; $L__BB161_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB161_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB161_27; bra.uni $L__BB161_34; $L__BB161_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB161_30; bra.uni $L__BB161_28; $L__BB161_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB161_32; bra.uni $L__BB161_31; $L__BB161_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB161_33; $L__BB161_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB161_29; bra.uni $L__BB161_34; $L__BB161_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB161_33; $L__BB161_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB161_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB161_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB161_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB161_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB161_12; } .func _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<55>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.f32 %f20, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u64 %rd6, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; ld.param.u32 %r5, [_Z19svm_node_light_pathILj33554879EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6]; mov.f32 %f17, 0f00000000; setp.gt.s32 %p1, %r3, 6; @%p1 bra $L__BB162_13; bra.uni $L__BB162_1; $L__BB162_13: setp.gt.s32 %p2, %r3, 10; @%p2 bra $L__BB162_23; setp.gt.s32 %p8, %r3, 8; @%p8 bra $L__BB162_18; setp.eq.s32 %p11, %r3, 7; @%p11 bra $L__BB162_40; setp.eq.s32 %p12, %r3, 8; mov.f32 %f20, %f17; @%p12 bra $L__BB162_17; bra.uni $L__BB162_45; $L__BB162_17: and.b32 %r12, %r2, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB162_45; $L__BB162_1: setp.gt.s32 %p13, %r3, 2; @%p13 bra $L__BB162_6; bra.uni $L__BB162_2; $L__BB162_6: setp.gt.s32 %p14, %r3, 4; @%p14 bra $L__BB162_10; setp.eq.s32 %p17, %r3, 3; @%p17 bra $L__BB162_42; setp.eq.s32 %p18, %r3, 4; mov.f32 %f20, %f17; @%p18 bra $L__BB162_9; bra.uni $L__BB162_45; $L__BB162_9: and.b32 %r16, %r5, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB162_45; $L__BB162_23: setp.gt.s32 %p3, %r3, 12; @%p3 bra $L__BB162_29; setp.eq.s32 %p6, %r3, 11; @%p6 bra $L__BB162_37; setp.eq.s32 %p7, %r3, 12; mov.f32 %f20, %f17; @%p7 bra $L__BB162_26; bra.uni $L__BB162_45; $L__BB162_26: and.b32 %r8, %r5, 768; mov.u64 %rd24, kernel_params; add.s64 %rd52, %rd24, 3400; setp.ne.s32 %p24, %r8, 0; @%p24 bra $L__BB162_28; add.s64 %rd52, %rd24, 2488; $L__BB162_28: ld.const.u64 %rd27, [%rd52]; cvta.to.global.u64 %rd28, %rd27; mul.wide.s32 %rd29, %r1, 2; add.s64 %rd30, %rd28, %rd29; ld.global.u16 %rs3, [%rd30]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB162_45; $L__BB162_2: setp.eq.s32 %p19, %r3, 0; @%p19 bra $L__BB162_44; setp.eq.s32 %p20, %r3, 1; @%p20 bra $L__BB162_43; setp.eq.s32 %p21, %r3, 2; mov.f32 %f20, %f17; @%p21 bra $L__BB162_5; bra.uni $L__BB162_45; $L__BB162_5: and.b32 %r18, %r5, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB162_45; $L__BB162_18: setp.eq.s32 %p9, %r3, 9; @%p9 bra $L__BB162_45; setp.eq.s32 %p10, %r3, 10; mov.f32 %f20, %f17; @%p10 bra $L__BB162_20; bra.uni $L__BB162_45; $L__BB162_20: and.b32 %r10, %r5, 768; mov.u64 %rd40, kernel_params; add.s64 %rd54, %rd40, 3376; setp.ne.s32 %p26, %r10, 0; @%p26 bra $L__BB162_22; add.s64 %rd54, %rd40, 2464; $L__BB162_22: ld.const.u64 %rd43, [%rd54]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r1, 2; add.s64 %rd46, %rd44, %rd45; ld.global.u16 %rs5, [%rd46]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r11, %r5, 525056; setp.eq.s32 %p27, %r11, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB162_45; $L__BB162_10: setp.eq.s32 %p15, %r3, 5; @%p15 bra $L__BB162_41; setp.eq.s32 %p16, %r3, 6; mov.f32 %f20, %f17; @%p16 bra $L__BB162_12; bra.uni $L__BB162_45; $L__BB162_12: and.b32 %r14, %r5, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB162_45; $L__BB162_29: setp.eq.s32 %p4, %r3, 13; @%p4 bra $L__BB162_34; setp.ne.s32 %p5, %r3, 14; mov.f32 %f20, %f17; @%p5 bra $L__BB162_45; and.b32 %r6, %r5, 768; mov.u64 %rd8, kernel_params; add.s64 %rd50, %rd8, 3408; setp.ne.s32 %p22, %r6, 0; @%p22 bra $L__BB162_33; add.s64 %rd50, %rd8, 2496; $L__BB162_33: ld.const.u64 %rd11, [%rd50]; cvta.to.global.u64 %rd12, %rd11; mul.wide.s32 %rd13, %r1, 2; add.s64 %rd14, %rd12, %rd13; ld.global.u16 %rs1, [%rd14]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB162_45; $L__BB162_40: and.b32 %r13, %r5, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB162_45; $L__BB162_42: and.b32 %r17, %r5, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB162_45; $L__BB162_37: and.b32 %r9, %r5, 768; mov.u64 %rd32, kernel_params; add.s64 %rd53, %rd32, 3392; setp.ne.s32 %p25, %r9, 0; @%p25 bra $L__BB162_39; add.s64 %rd53, %rd32, 2480; $L__BB162_39: ld.const.u64 %rd35, [%rd53]; cvta.to.global.u64 %rd36, %rd35; mul.wide.s32 %rd37, %r1, 2; add.s64 %rd38, %rd36, %rd37; ld.global.u16 %rs4, [%rd38]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB162_45; $L__BB162_44: and.b32 %r20, %r5, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB162_45; $L__BB162_43: and.b32 %r19, %r5, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB162_45; $L__BB162_41: and.b32 %r15, %r5, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB162_45; $L__BB162_34: and.b32 %r7, %r5, 768; mov.u64 %rd16, kernel_params; add.s64 %rd51, %rd16, 3384; setp.ne.s32 %p23, %r7, 0; @%p23 bra $L__BB162_36; add.s64 %rd51, %rd16, 2472; $L__BB162_36: ld.const.u64 %rd19, [%rd51]; cvta.to.global.u64 %rd20, %rd19; mul.wide.s32 %rd21, %r1, 2; add.s64 %rd22, %rd20, %rd21; ld.global.u16 %rs2, [%rd22]; cvt.rn.f32.u16 %f20, %rs2; $L__BB162_45: cvta.to.local.u64 %rd47, %rd6; mul.wide.u32 %rd48, %r4, 4; add.s64 %rd49, %rd47, %rd48; st.local.f32 [%rd49], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB163_2; bra.uni $L__BB163_1; $L__BB163_2: mov.b32 %f2764, %r395; bra.uni $L__BB163_3; $L__BB163_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB163_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB163_5; bra.uni $L__BB163_4; $L__BB163_5: mov.b32 %f9, %r396; bra.uni $L__BB163_6; $L__BB163_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB163_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB163_8; bra.uni $L__BB163_7; $L__BB163_8: mov.b32 %f2766, %r397; bra.uni $L__BB163_9; $L__BB163_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB163_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB163_11; bra.uni $L__BB163_10; $L__BB163_11: mov.b32 %f15, %r398; bra.uni $L__BB163_12; $L__BB163_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB163_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB163_14; bra.uni $L__BB163_13; $L__BB163_14: mov.b32 %f2768, %r399; bra.uni $L__BB163_15; $L__BB163_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB163_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB163_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB163_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB163_18; bra.uni $L__BB163_385; $L__BB163_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB163_259; bra.uni $L__BB163_19; $L__BB163_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB163_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB163_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB163_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB163_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB163_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB163_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB163_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB163_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB163_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB163_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB163_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB163_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB163_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB163_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB163_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB163_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB163_362; $L__BB163_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB163_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB163_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB163_40; bra.uni $L__BB163_29; $L__BB163_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB163_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB163_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB163_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB163_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB163_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB163_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB163_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB163_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB163_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB163_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB163_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB163_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB163_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB163_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB163_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB163_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB163_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB163_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB163_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB163_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB163_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB163_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB163_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB163_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB163_136; $L__BB163_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB163_371; bra.uni $L__BB163_366; $L__BB163_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB163_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB163_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB163_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB163_383; $L__BB163_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB163_149; bra.uni $L__BB163_139; $L__BB163_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB163_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB163_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB163_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB163_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB163_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB163_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB163_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB163_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB163_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB163_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB163_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB163_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB163_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB163_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB163_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB163_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB163_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB163_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB163_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB163_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB163_250; $L__BB163_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB163_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB163_253; bra.uni $L__BB163_362; $L__BB163_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB163_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB163_304; @%p171 bra $L__BB163_303; bra.uni $L__BB163_256; $L__BB163_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB163_305; $L__BB163_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB163_305; $L__BB163_256: @%p172 bra $L__BB163_302; bra.uni $L__BB163_257; $L__BB163_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB163_305; $L__BB163_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB163_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB163_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB163_308; bra.uni $L__BB163_306; $L__BB163_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB163_309; $L__BB163_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB163_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB163_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB163_316; @%p171 bra $L__BB163_315; bra.uni $L__BB163_311; $L__BB163_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB163_317; $L__BB163_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB163_317; $L__BB163_311: @%p172 bra $L__BB163_314; bra.uni $L__BB163_312; $L__BB163_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB163_317; $L__BB163_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB163_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB163_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB163_320; bra.uni $L__BB163_318; $L__BB163_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB163_321; $L__BB163_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB163_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB163_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB163_328; @%p171 bra $L__BB163_327; bra.uni $L__BB163_323; $L__BB163_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB163_329; $L__BB163_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB163_329; $L__BB163_323: @%p172 bra $L__BB163_326; bra.uni $L__BB163_324; $L__BB163_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB163_329; $L__BB163_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB163_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB163_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB163_332; bra.uni $L__BB163_330; $L__BB163_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB163_333; $L__BB163_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB163_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB163_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB163_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_362; $L__BB163_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB163_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB163_31; bra.uni $L__BB163_136; $L__BB163_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB163_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB163_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB163_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB163_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB163_107; @%p49 bra $L__BB163_106; bra.uni $L__BB163_37; $L__BB163_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB163_108; $L__BB163_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB163_108; $L__BB163_37: @%p50 bra $L__BB163_105; bra.uni $L__BB163_38; $L__BB163_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB163_108; $L__BB163_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB163_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB163_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB163_111; bra.uni $L__BB163_109; $L__BB163_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB163_112; $L__BB163_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB163_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB163_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB163_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB163_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB163_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB163_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_136; $L__BB163_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB163_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB163_368; bra.uni $L__BB163_383; $L__BB163_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB163_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB163_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB163_383; $L__BB163_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB163_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB163_141; bra.uni $L__BB163_250; $L__BB163_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB163_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB163_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB163_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB163_206; @%p101 bra $L__BB163_205; bra.uni $L__BB163_146; $L__BB163_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB163_207; $L__BB163_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB163_207; $L__BB163_146: @%p102 bra $L__BB163_204; bra.uni $L__BB163_147; $L__BB163_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB163_207; $L__BB163_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB163_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB163_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB163_210; bra.uni $L__BB163_208; $L__BB163_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB163_211; $L__BB163_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB163_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB163_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB163_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB163_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB163_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_250; $L__BB163_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB163_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB163_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB163_295; @%p151 bra $L__BB163_294; bra.uni $L__BB163_290; $L__BB163_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB163_296; $L__BB163_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB163_296; $L__BB163_290: @%p152 bra $L__BB163_293; bra.uni $L__BB163_291; $L__BB163_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB163_296; $L__BB163_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB163_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB163_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB163_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB163_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB163_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB163_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB163_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_362; $L__BB163_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB163_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB163_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB163_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB163_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB163_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB163_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB163_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB163_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB163_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB163_279; bra.uni $L__BB163_362; $L__BB163_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB163_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB163_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB163_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB163_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB163_96; @%p39 bra $L__BB163_95; bra.uni $L__BB163_91; $L__BB163_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB163_97; $L__BB163_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB163_97; $L__BB163_91: @%p40 bra $L__BB163_94; bra.uni $L__BB163_92; $L__BB163_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB163_97; $L__BB163_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB163_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB163_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB163_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB163_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB163_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB163_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB163_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB163_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB163_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_136; $L__BB163_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB163_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB163_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB163_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB163_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB163_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB163_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB163_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB163_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB163_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB163_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB163_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB163_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB163_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB163_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB163_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB163_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB163_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB163_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB163_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB163_76; bra.uni $L__BB163_136; $L__BB163_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB163_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB163_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB163_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB163_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB163_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_383; $L__BB163_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB163_383; $L__BB163_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB163_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB163_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB163_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB163_196; @%p92 bra $L__BB163_195; bra.uni $L__BB163_191; $L__BB163_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB163_197; $L__BB163_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB163_197; $L__BB163_191: @%p93 bra $L__BB163_194; bra.uni $L__BB163_192; $L__BB163_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB163_197; $L__BB163_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB163_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB163_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB163_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB163_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB163_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB163_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB163_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB163_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB163_250; $L__BB163_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB163_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB163_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB163_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB163_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB163_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB163_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB163_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB163_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB163_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB163_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB163_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB163_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB163_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB163_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB163_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB163_179; bra.uni $L__BB163_250; $L__BB163_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB163_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB163_337; @%p188 bra $L__BB163_336; bra.uni $L__BB163_23; $L__BB163_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB163_338; $L__BB163_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB163_338; $L__BB163_23: @%p189 bra $L__BB163_335; bra.uni $L__BB163_24; $L__BB163_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB163_338; $L__BB163_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB163_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB163_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB163_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB163_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB163_347; @%p188 bra $L__BB163_346; bra.uni $L__BB163_342; $L__BB163_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB163_348; $L__BB163_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB163_348; $L__BB163_342: @%p189 bra $L__BB163_345; bra.uni $L__BB163_343; $L__BB163_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB163_348; $L__BB163_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB163_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB163_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB163_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB163_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB163_357; @%p188 bra $L__BB163_356; bra.uni $L__BB163_352; $L__BB163_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB163_358; $L__BB163_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB163_358; $L__BB163_352: @%p189 bra $L__BB163_355; bra.uni $L__BB163_353; $L__BB163_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB163_358; $L__BB163_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB163_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB163_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB163_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB163_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB163_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB163_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB163_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB163_364: mov.f32 %f3392, %f3387; bra.uni $L__BB163_385; $L__BB163_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB163_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB163_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB163_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB163_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB163_128; @%p59 bra $L__BB163_127; bra.uni $L__BB163_123; $L__BB163_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB163_129; $L__BB163_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB163_129; $L__BB163_123: @%p60 bra $L__BB163_126; bra.uni $L__BB163_124; $L__BB163_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB163_129; $L__BB163_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB163_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB163_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB163_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB163_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB163_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB163_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB163_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB163_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB163_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB163_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB163_385; $L__BB163_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB163_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB163_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB163_385; $L__BB163_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB163_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB163_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB163_224; @%p120 bra $L__BB163_223; bra.uni $L__BB163_219; $L__BB163_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB163_225; $L__BB163_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB163_225; $L__BB163_219: @%p121 bra $L__BB163_222; bra.uni $L__BB163_220; $L__BB163_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB163_225; $L__BB163_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB163_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB163_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB163_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB163_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB163_234; @%p120 bra $L__BB163_233; bra.uni $L__BB163_229; $L__BB163_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB163_235; $L__BB163_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB163_235; $L__BB163_229: @%p121 bra $L__BB163_232; bra.uni $L__BB163_230; $L__BB163_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB163_235; $L__BB163_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB163_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB163_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB163_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB163_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB163_244; @%p120 bra $L__BB163_243; bra.uni $L__BB163_239; $L__BB163_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB163_245; $L__BB163_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB163_245; $L__BB163_239: @%p121 bra $L__BB163_242; bra.uni $L__BB163_240; $L__BB163_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB163_245; $L__BB163_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB163_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB163_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB163_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB163_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB163_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB163_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB163_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB163_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB163_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB163_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB163_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB163_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB163_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB163_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB163_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB163_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB163_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB163_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB163_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj33554879EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj33554879EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB165_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB165_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB165_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB165_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) { .reg .pred %p<2>; .reg .b32 %r<5>; ld.param.u32 %r1, [_Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u8 %r2, [_Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0+4]; setp.eq.s32 %p1, %r2, 39; selp.b32 %r3, 5, 1, %p1; add.s32 %r4, %r3, %r1; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<535>; .reg .b32 %r<242>; .reg .b64 %rd<290>; ld.param.u64 %rd63, [_Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd64, [_Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd65, %rd63; add.s64 %rd1, %rd65, 76; ld.local.u32 %r3, [%rd65+76]; cvt.s64.s32 %rd2, %r3; setp.eq.s32 %p1, %r3, -1; mov.u32 %r29, 0; mov.u32 %r30, %r29; mov.u32 %r236, %r64; mov.u32 %r32, %r29; @%p1 bra $L__BB167_16; cvt.u64.u32 %rd3, %r62; ld.local.s32 %rd4, [%rd1+-16]; ld.const.u64 %rd66, [kernel_params+2160]; cvta.to.global.u64 %rd67, %rd66; mul.lo.s64 %rd68, %rd2, 208; add.s64 %rd69, %rd67, %rd68; ld.global.u32 %r4, [%rd69+164]; ld.local.u32 %r69, [%rd1+-12]; and.b32 %r70, %r69, 1; setp.eq.b32 %p2, %r70, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r225, 0; @%p5 bra $L__BB167_5; cvt.u32.u64 %r72, %rd4; setp.eq.s32 %p6, %r72, -1; mov.u32 %r224, -1; @%p6 bra $L__BB167_4; ld.const.u64 %rd70, [kernel_params+2240]; cvta.to.global.u64 %rd71, %rd70; shl.b64 %rd72, %rd4, 2; add.s64 %rd73, %rd71, %rd72; ld.global.u32 %r224, [%rd73]; $L__BB167_4: setp.ne.s32 %p7, %r224, -1; selp.u32 %r225, 1, 0, %p7; $L__BB167_5: ld.const.u64 %rd74, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd74; add.s32 %r227, %r225, %r4; mul.wide.u32 %rd75, %r227, 16; add.s64 %rd76, %rd5, %rd75; add.s64 %rd276, %rd76, 14; add.s64 %rd277, %rd76, 15; ld.global.u32 %r226, [%rd76+8]; ld.global.u16 %rs37, [%rd76+12]; ld.global.u64 %rd275, [%rd76]; setp.eq.s64 %p8, %rd275, %rd3; @%p8 bra $L__BB167_11; $L__BB167_6: setp.eq.s64 %p9, %rd275, 0; @%p9 bra $L__BB167_8; add.s32 %r227, %r227, 2; bra.uni $L__BB167_9; $L__BB167_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r233, -1; mov.u32 %r230, 0; mov.u32 %r227, %r226; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p10 bra $L__BB167_15; $L__BB167_9: cvt.u64.u32 %rd10, %r227; mul.wide.u32 %rd77, %r227, 16; add.s64 %rd78, %rd5, %rd77; ld.global.u32 %r226, [%rd78+8]; ld.global.u16 %rs37, [%rd78+12]; ld.global.u64 %rd275, [%rd78]; setp.ne.s64 %p11, %rd275, %rd3; @%p11 bra $L__BB167_6; shl.b64 %rd79, %rd10, 4; add.s64 %rd80, %rd5, %rd79; add.s64 %rd276, %rd80, 14; add.s64 %rd277, %rd80, 15; $L__BB167_11: ld.global.u8 %rs5, [%rd277]; ld.global.u8 %rs6, [%rd276]; cvt.u32.u64 %r77, %rd4; setp.ne.s32 %p12, %r77, -1; @%p12 bra $L__BB167_14; mov.u32 %r233, -1; mov.u32 %r230, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB167_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p14 bra $L__BB167_15; $L__BB167_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r233, -1, %r226, %p15; cvt.u32.u16 %r230, %rs37; cvt.u32.u16 %r82, %rs5; and.b32 %r232, %r82, 255; cvt.u32.u16 %r83, %rs6; and.b32 %r231, %r83, 255; $L__BB167_15: setp.eq.s32 %p16, %r233, -1; selp.b32 %r29, 0, %r233, %p16; selp.b32 %r30, 0, %r232, %p16; selp.b32 %r236, %r64, %r231, %p16; selp.b32 %r32, 0, %r230, %p16; $L__BB167_16: setp.eq.s32 %p17, %r32, 0; setp.eq.s32 %p18, %r62, 7; and.pred %p19, %p18, %p17; cvta.to.local.u64 %rd81, %rd64; mul.wide.u32 %rd82, %r63, 4; add.s64 %rd16, %rd81, %rd82; @%p19 bra $L__BB167_148; bra.uni $L__BB167_17; $L__BB167_148: cvt.u32.u64 %r222, %rd2; setp.eq.s32 %p103, %r222, -1; ld.local.v4.f32 {%f534, %f533, %f532, %f426}, [%rd65]; @%p103 bra $L__BB167_152; ld.local.u8 %rs35, [%rd65+56]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p104, %rs36, 0; @%p104 bra $L__BB167_151; bra.uni $L__BB167_150; $L__BB167_151: ld.const.u64 %rd271, [kernel_params+2160]; cvta.to.global.u64 %rd272, %rd271; mul.wide.s32 %rd273, %r3, 208; add.s64 %rd274, %rd272, %rd273; ld.global.v4.f32 {%f460, %f461, %f462, %f463}, [%rd274+48]; ld.global.v4.f32 {%f468, %f469, %f470, %f471}, [%rd274+64]; ld.global.v4.f32 {%f476, %f477, %f478, %f479}, [%rd274+80]; mul.ftz.f32 %f484, %f533, %f461; fma.rn.ftz.f32 %f485, %f534, %f460, %f484; fma.rn.ftz.f32 %f486, %f532, %f462, %f485; add.ftz.f32 %f173, %f463, %f486; mul.ftz.f32 %f487, %f533, %f469; fma.rn.ftz.f32 %f488, %f534, %f468, %f487; fma.rn.ftz.f32 %f489, %f532, %f470, %f488; add.ftz.f32 %f174, %f471, %f489; mul.ftz.f32 %f490, %f533, %f477; fma.rn.ftz.f32 %f491, %f534, %f476, %f490; fma.rn.ftz.f32 %f492, %f532, %f478, %f491; add.ftz.f32 %f532, %f479, %f492; mov.f32 %f533, %f174; mov.f32 %f534, %f173; bra.uni $L__BB167_152; $L__BB167_17: setp.eq.s32 %p20, %r236, 0; @%p20 bra $L__BB167_123; setp.eq.s32 %p21, %r236, 1; @%p21 bra $L__BB167_98; bra.uni $L__BB167_19; $L__BB167_98: mov.u32 %r48, 1; ld.local.u32 %r51, [%rd65+64]; and.b32 %r153, %r51, 1; setp.eq.b32 %p65, %r153, 1; mov.pred %p66, 0; xor.pred %p67, %p65, %p66; not.pred %p68, %p67; @%p68 bra $L__BB167_109; bra.uni $L__BB167_99; $L__BB167_109: and.b32 %r173, %r51, 6; setp.eq.s32 %p75, %r173, 0; @%p75 bra $L__BB167_116; and.b32 %r174, %r32, 768; setp.eq.s32 %p76, %r174, 0; @%p76 bra $L__BB167_112; ld.const.u64 %rd205, [kernel_params+2264]; cvta.to.global.u64 %rd206, %rd205; ld.local.u32 %r175, [%rd65+60]; mul.wide.s32 %rd208, %r175, 16; add.s64 %rd209, %rd206, %rd208; shr.s32 %r176, %r51, 7; ld.global.u32 %r177, [%rd209+4]; add.s32 %r178, %r176, %r177; ld.const.u64 %rd210, [kernel_params+2328]; cvta.to.global.u64 %rd211, %rd210; add.s32 %r179, %r178, %r29; mul.wide.s32 %rd212, %r179, 8; add.s64 %rd213, %rd211, %rd212; ld.global.v2.f32 {%f383, %f384}, [%rd213]; ld.global.v2.f32 {%f387, %f388}, [%rd213+8]; ld.local.f32 %f391, [%rd65+68]; mov.f32 %f392, 0f3F800000; sub.ftz.f32 %f393, %f392, %f391; mul.ftz.f32 %f394, %f383, %f393; mul.ftz.f32 %f395, %f384, %f393; fma.rn.ftz.f32 %f530, %f388, %f391, %f395; fma.rn.ftz.f32 %f529, %f387, %f391, %f394; bra.uni $L__BB167_118; $L__BB167_123: ld.local.u32 %r58, [%rd65+64]; and.b32 %r187, %r58, 1; setp.eq.b32 %p84, %r187, 1; mov.pred %p85, 0; xor.pred %p86, %p84, %p85; not.pred %p87, %p86; @%p87 bra $L__BB167_134; bra.uni $L__BB167_124; $L__BB167_134: and.b32 %r209, %r58, 6; setp.eq.s32 %p94, %r209, 0; @%p94 bra $L__BB167_141; and.b32 %r210, %r32, 768; setp.eq.s32 %p95, %r210, 0; @%p95 bra $L__BB167_137; ld.const.u64 %rd249, [kernel_params+2264]; cvta.to.global.u64 %rd250, %rd249; ld.local.u32 %r211, [%rd65+60]; mul.wide.s32 %rd252, %r211, 16; add.s64 %rd253, %rd250, %rd252; shr.s32 %r212, %r58, 7; ld.global.u32 %r213, [%rd253+4]; add.s32 %r214, %r212, %r213; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r215, %r214, %r29; mul.wide.s32 %rd256, %r215, 4; add.s64 %rd257, %rd255, %rd256; ld.local.f32 %f415, [%rd65+68]; mov.f32 %f416, 0f3F800000; sub.ftz.f32 %f417, %f416, %f415; ld.global.f32 %f418, [%rd257]; mul.ftz.f32 %f419, %f418, %f417; ld.global.f32 %f420, [%rd257+4]; fma.rn.ftz.f32 %f531, %f420, %f415, %f419; bra.uni $L__BB167_143; $L__BB167_150: ld.local.v4.f32 {%f427, %f428, %f429, %f430}, [%rd65+192]; mul.ftz.f32 %f435, %f533, %f428; fma.rn.ftz.f32 %f436, %f534, %f427, %f435; fma.rn.ftz.f32 %f437, %f532, %f429, %f436; add.ftz.f32 %f170, %f430, %f437; ld.local.v4.f32 {%f438, %f439, %f440, %f441}, [%rd65+208]; mul.ftz.f32 %f446, %f533, %f439; fma.rn.ftz.f32 %f447, %f534, %f438, %f446; fma.rn.ftz.f32 %f448, %f532, %f440, %f447; add.ftz.f32 %f171, %f441, %f448; ld.local.v4.f32 {%f449, %f450, %f451, %f452}, [%rd65+224]; mul.ftz.f32 %f457, %f533, %f450; fma.rn.ftz.f32 %f458, %f534, %f449, %f457; fma.rn.ftz.f32 %f459, %f532, %f451, %f458; add.ftz.f32 %f532, %f452, %f459; mov.f32 %f533, %f171; mov.f32 %f534, %f170; $L__BB167_152: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB167_156; bra.uni $L__BB167_153; $L__BB167_156: add.ftz.f32 %f493, %f533, %f534; add.ftz.f32 %f494, %f532, %f493; mul.ftz.f32 %f495, %f494, 0f3EAAAAAB; st.local.f32 [%rd16], %f495; bra.uni $L__BB167_157; $L__BB167_153: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB167_155; mov.u32 %r223, 1065353216; st.local.u32 [%rd16], %r223; bra.uni $L__BB167_157; $L__BB167_19: add.s32 %r87, %r236, -3; setp.lt.u32 %p22, %r87, 2; ld.local.u32 %r41, [%rd65+64]; and.b32 %r42, %r41, 1; @%p22 bra $L__BB167_45; bra.uni $L__BB167_20; $L__BB167_45: setp.eq.s32 %p39, %r42, 0; @%p39 bra $L__BB167_84; bra.uni $L__BB167_46; $L__BB167_84: and.b32 %r141, %r41, 6; setp.eq.s32 %p56, %r141, 0; @%p56 bra $L__BB167_91; and.b32 %r142, %r32, 768; setp.eq.s32 %p57, %r142, 0; @%p57 bra $L__BB167_87; ld.const.u64 %rd158, [kernel_params+2264]; cvta.to.global.u64 %rd159, %rd158; ld.local.u32 %r143, [%rd65+60]; mul.wide.s32 %rd161, %r143, 16; add.s64 %rd162, %rd159, %rd161; shr.s32 %r144, %r41, 7; ld.global.u32 %r145, [%rd162+4]; add.s32 %r146, %r144, %r145; ld.const.u64 %rd163, [kernel_params+2344]; cvta.to.global.u64 %rd164, %rd163; add.s32 %r147, %r146, %r29; mul.wide.s32 %rd165, %r147, 16; add.s64 %rd166, %rd164, %rd165; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd166]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd166+16]; ld.local.f32 %f332, [%rd65+68]; mov.f32 %f333, 0f3F800000; sub.ftz.f32 %f334, %f333, %f332; mul.ftz.f32 %f335, %f316, %f334; mul.ftz.f32 %f336, %f317, %f334; mul.ftz.f32 %f337, %f318, %f334; mul.ftz.f32 %f338, %f319, %f334; fma.rn.ftz.f32 %f522, %f327, %f332, %f338; fma.rn.ftz.f32 %f521, %f326, %f332, %f337; fma.rn.ftz.f32 %f520, %f325, %f332, %f336; fma.rn.ftz.f32 %f519, %f324, %f332, %f335; bra.uni $L__BB167_93; $L__BB167_124: ld.local.s32 %rd53, [%rd65+60]; setp.eq.s64 %p88, %rd53, -1; @%p88 bra $L__BB167_127; ld.const.u64 %rd226, [kernel_params+2240]; cvta.to.global.u64 %rd227, %rd226; shl.b64 %rd228, %rd53, 2; add.s64 %rd229, %rd227, %rd228; ld.global.u32 %r188, [%rd229]; setp.eq.s32 %p89, %r188, -1; @%p89 bra $L__BB167_127; mov.u32 %r189, 0; mov.u64 %rd230, 0; { // callseq 935, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r189; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd230; .param .b64 param3; st.param.b64 [param3+0], %rd230; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f531, [retval0+0]; } // callseq 935 bra.uni $L__BB167_143; $L__BB167_155: st.local.f32 [%rd16], %f534; st.local.f32 [%rd16+4], %f533; st.local.f32 [%rd16+8], %f532; bra.uni $L__BB167_157; $L__BB167_99: ld.local.s32 %rd51, [%rd65+60]; setp.eq.s64 %p69, %rd51, -1; @%p69 bra $L__BB167_102; ld.const.u64 %rd179, [kernel_params+2240]; cvta.to.global.u64 %rd180, %rd179; shl.b64 %rd181, %rd51, 2; add.s64 %rd182, %rd180, %rd181; ld.global.u32 %r154, [%rd182]; setp.eq.s32 %p70, %r154, -1; @%p70 bra $L__BB167_102; mov.u64 %rd183, 0; { // callseq 934, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r48; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd183; .param .b64 param3; st.param.b64 [param3+0], %rd183; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f529, [retval0+0]; ld.param.f32 %f530, [retval0+4]; } // callseq 934 bra.uni $L__BB167_118; $L__BB167_20: setp.eq.s32 %p23, %r42, 0; @%p23 bra $L__BB167_31; bra.uni $L__BB167_21; $L__BB167_31: and.b32 %r109, %r41, 6; setp.eq.s32 %p30, %r109, 0; @%p30 bra $L__BB167_38; and.b32 %r110, %r32, 768; setp.eq.s32 %p31, %r110, 0; @%p31 bra $L__BB167_34; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r111, [%rd65+60]; mul.wide.s32 %rd112, %r111, 16; add.s64 %rd113, %rd110, %rd112; shr.s32 %r112, %r41, 7; ld.global.u32 %r113, [%rd113+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd114, [kernel_params+2336]; cvta.to.global.u64 %rd115, %rd114; add.s32 %r115, %r114, %r29; mul.wide.s32 %rd116, %r115, 12; add.s64 %rd117, %rd115, %rd116; ld.local.f32 %f202, [%rd65+68]; mov.f32 %f203, 0f3F800000; sub.ftz.f32 %f204, %f203, %f202; ld.global.f32 %f205, [%rd117]; mul.ftz.f32 %f206, %f205, %f204; ld.global.f32 %f207, [%rd117+4]; mul.ftz.f32 %f208, %f207, %f204; ld.global.f32 %f209, [%rd117+8]; mul.ftz.f32 %f210, %f209, %f204; ld.global.f32 %f211, [%rd117+12]; ld.global.f32 %f212, [%rd117+16]; ld.global.f32 %f213, [%rd117+20]; fma.rn.ftz.f32 %f498, %f213, %f202, %f210; fma.rn.ftz.f32 %f497, %f212, %f202, %f208; fma.rn.ftz.f32 %f496, %f211, %f202, %f206; bra.uni $L__BB167_40; $L__BB167_127: and.b32 %r190, %r32, 56; setp.eq.s32 %p90, %r190, 0; @%p90 bra $L__BB167_132; and.b32 %r191, %r32, 24; setp.eq.s32 %p91, %r191, 0; @%p91 bra $L__BB167_130; ld.const.u64 %rd231, [kernel_params+2232]; cvta.to.global.u64 %rd232, %rd231; shl.b64 %rd233, %rd53, 4; add.s64 %rd234, %rd232, %rd233; ld.global.v4.u32 {%r192, %r193, %r194, %r195}, [%rd234]; ld.const.u64 %rd235, [kernel_params+2320]; cvta.to.global.u64 %rd236, %rd235; add.s32 %r199, %r192, %r29; mul.wide.u32 %rd237, %r199, 4; add.s64 %rd287, %rd236, %rd237; add.s32 %r200, %r193, %r29; mul.wide.u32 %rd238, %r200, 4; add.s64 %rd288, %rd236, %rd238; add.s32 %r201, %r194, %r29; mul.wide.u32 %rd239, %r201, 4; add.s64 %rd289, %rd236, %rd239; bra.uni $L__BB167_131; $L__BB167_102: and.b32 %r155, %r32, 56; setp.eq.s32 %p71, %r155, 0; @%p71 bra $L__BB167_107; and.b32 %r156, %r32, 24; setp.eq.s32 %p72, %r156, 0; @%p72 bra $L__BB167_105; ld.const.u64 %rd184, [kernel_params+2232]; cvta.to.global.u64 %rd185, %rd184; shl.b64 %rd186, %rd51, 4; add.s64 %rd187, %rd185, %rd186; ld.global.v4.u32 {%r157, %r158, %r159, %r160}, [%rd187]; ld.const.u64 %rd188, [kernel_params+2328]; cvta.to.global.u64 %rd189, %rd188; add.s32 %r164, %r157, %r29; mul.wide.u32 %rd190, %r164, 8; add.s64 %rd191, %rd189, %rd190; ld.global.v2.f32 {%f527, %f528}, [%rd191]; add.s32 %r165, %r158, %r29; mul.wide.u32 %rd192, %r165, 8; add.s64 %rd193, %rd189, %rd192; ld.global.v2.f32 {%f525, %f526}, [%rd193]; add.s32 %r166, %r159, %r29; mul.wide.u32 %rd194, %r166, 8; add.s64 %rd195, %rd189, %rd194; ld.global.v2.f32 {%f523, %f524}, [%rd195]; bra.uni $L__BB167_106; $L__BB167_46: ld.local.s32 %rd47, [%rd65+60]; setp.eq.s64 %p40, %rd47, -1; @%p40 bra $L__BB167_49; ld.const.u64 %rd129, [kernel_params+2240]; cvta.to.global.u64 %rd130, %rd129; shl.b64 %rd131, %rd47, 2; add.s64 %rd132, %rd130, %rd131; ld.global.u32 %r122, [%rd132]; setp.eq.s32 %p41, %r122, -1; @%p41 bra $L__BB167_49; mov.u64 %rd133, 0; { // callseq 933, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd133; .param .b64 param3; st.param.b64 [param3+0], %rd133; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f519, [retval0+0]; ld.param.f32 %f520, [retval0+4]; ld.param.f32 %f521, [retval0+8]; ld.param.f32 %f522, [retval0+12]; } // callseq 933 bra.uni $L__BB167_93; $L__BB167_21: ld.local.s32 %rd18, [%rd65+60]; setp.eq.s64 %p24, %rd18, -1; @%p24 bra $L__BB167_24; ld.const.u64 %rd85, [kernel_params+2240]; cvta.to.global.u64 %rd86, %rd85; shl.b64 %rd87, %rd18, 2; add.s64 %rd88, %rd86, %rd87; ld.global.u32 %r88, [%rd88]; setp.eq.s32 %p25, %r88, -1; @%p25 bra $L__BB167_24; mov.u64 %rd89, 0; { // callseq 932, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd89; .param .b64 param3; st.param.b64 [param3+0], %rd89; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f496, [retval0+0]; ld.param.f32 %f497, [retval0+4]; ld.param.f32 %f498, [retval0+8]; } // callseq 932 bra.uni $L__BB167_40; $L__BB167_141: and.b32 %r218, %r58, 8; setp.eq.s32 %p98, %r218, 0; setp.ne.s32 %p99, %r32, 8; mov.f32 %f531, 0f00000000; or.pred %p100, %p99, %p98; @%p100 bra $L__BB167_143; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; ld.local.u32 %r219, [%rd65+60]; add.s32 %r220, %r219, %r29; mul.wide.s32 %rd266, %r220, 4; add.s64 %rd267, %rd264, %rd266; ld.global.f32 %f531, [%rd267]; bra.uni $L__BB167_143; $L__BB167_116: and.b32 %r182, %r51, 8; setp.eq.s32 %p79, %r182, 0; setp.ne.s32 %p80, %r32, 8; mov.f32 %f529, 0f00000000; or.pred %p81, %p80, %p79; mov.f32 %f530, %f529; @%p81 bra $L__BB167_118; ld.const.u64 %rd219, [kernel_params+2328]; cvta.to.global.u64 %rd220, %rd219; ld.local.u32 %r183, [%rd65+60]; add.s32 %r184, %r183, %r29; mul.wide.s32 %rd222, %r184, 8; add.s64 %rd223, %rd220, %rd222; ld.global.v2.f32 {%f529, %f530}, [%rd223]; bra.uni $L__BB167_118; $L__BB167_49: and.b32 %r123, %r32, 120; setp.eq.s32 %p42, %r123, 0; @%p42 bra $L__BB167_82; and.b32 %r124, %r32, 24; setp.eq.s32 %p43, %r124, 0; @%p43 bra $L__BB167_52; ld.const.u64 %rd134, [kernel_params+2232]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd47, 4; add.s64 %rd137, %rd135, %rd136; ld.global.v4.u32 {%r125, %r126, %r127, %r128}, [%rd137]; ld.const.u64 %rd138, [kernel_params+2344]; cvta.to.global.u64 %rd139, %rd138; add.s32 %r132, %r125, %r29; mul.wide.u32 %rd140, %r132, 16; add.s64 %rd141, %rd139, %rd140; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd141]; add.s32 %r133, %r126, %r29; mul.wide.u32 %rd142, %r133, 16; add.s64 %rd143, %rd139, %rd142; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd143]; add.s32 %r134, %r127, %r29; mul.wide.u32 %rd144, %r134, 16; add.s64 %rd145, %rd139, %rd144; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd145]; bra.uni $L__BB167_81; $L__BB167_24: and.b32 %r89, %r32, 56; setp.eq.s32 %p26, %r89, 0; @%p26 bra $L__BB167_29; and.b32 %r90, %r32, 24; setp.eq.s32 %p27, %r90, 0; @%p27 bra $L__BB167_27; ld.const.u64 %rd90, [kernel_params+2232]; cvta.to.global.u64 %rd91, %rd90; shl.b64 %rd92, %rd18, 4; add.s64 %rd93, %rd91, %rd92; ld.global.v4.u32 {%r91, %r92, %r93, %r94}, [%rd93]; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; add.s32 %r98, %r91, %r29; mul.wide.u32 %rd96, %r98, 12; add.s64 %rd284, %rd95, %rd96; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r99, %r92, %r29; mul.wide.u32 %rd97, %r99, 12; add.s64 %rd281, %rd95, %rd97; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r100, %r93, %r29; mul.wide.u32 %rd98, %r100, 12; add.s64 %rd278, %rd95, %rd98; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; bra.uni $L__BB167_28; $L__BB167_137: and.b32 %r216, %r32, 131; setp.eq.s32 %p96, %r216, 0; mov.f32 %f531, 0f00000000; @%p96 bra $L__BB167_143; setp.ne.s32 %p97, %r32, 128; @%p97 bra $L__BB167_140; ld.local.u32 %r217, [%rd65+60]; add.s32 %r29, %r217, %r29; $L__BB167_140: ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r29, 4; add.s64 %rd262, %rd260, %rd261; ld.global.f32 %f531, [%rd262]; bra.uni $L__BB167_143; $L__BB167_132: and.b32 %r205, %r32, 7; setp.eq.s32 %p92, %r205, 0; mov.f32 %f531, 0f00000000; @%p92 bra $L__BB167_143; cvt.u32.u64 %r206, %rd53; setp.eq.s32 %p93, %r32, 4; selp.b32 %r207, %r206, 0, %p93; add.s32 %r208, %r29, %r207; ld.const.u64 %rd245, [kernel_params+2320]; cvta.to.global.u64 %rd246, %rd245; mul.wide.s32 %rd247, %r208, 4; add.s64 %rd248, %rd246, %rd247; ld.global.f32 %f531, [%rd248]; bra.uni $L__BB167_143; $L__BB167_112: and.b32 %r180, %r32, 131; setp.eq.s32 %p77, %r180, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p77 bra $L__BB167_118; setp.ne.s32 %p78, %r32, 128; @%p78 bra $L__BB167_115; ld.local.u32 %r181, [%rd65+60]; add.s32 %r29, %r181, %r29; $L__BB167_115: ld.const.u64 %rd215, [kernel_params+2328]; cvta.to.global.u64 %rd216, %rd215; mul.wide.s32 %rd217, %r29, 8; add.s64 %rd218, %rd216, %rd217; ld.global.v2.f32 {%f529, %f530}, [%rd218]; bra.uni $L__BB167_118; $L__BB167_107: and.b32 %r169, %r32, 7; setp.eq.s32 %p73, %r169, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p73 bra $L__BB167_118; cvt.u32.u64 %r170, %rd51; setp.eq.s32 %p74, %r32, 4; selp.b32 %r171, %r170, 0, %p74; add.s32 %r172, %r29, %r171; ld.const.u64 %rd201, [kernel_params+2328]; cvta.to.global.u64 %rd202, %rd201; mul.wide.s32 %rd203, %r172, 8; add.s64 %rd204, %rd202, %rd203; ld.global.v2.f32 {%f529, %f530}, [%rd204]; bra.uni $L__BB167_118; $L__BB167_91: and.b32 %r150, %r41, 8; setp.eq.s32 %p60, %r150, 0; setp.ne.s32 %p61, %r32, 8; mov.f32 %f519, 0f00000000; or.pred %p62, %p61, %p60; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p62 bra $L__BB167_93; ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; ld.local.u32 %r151, [%rd65+60]; add.s32 %r152, %r151, %r29; mul.wide.s32 %rd175, %r152, 16; add.s64 %rd176, %rd173, %rd175; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd176]; bra.uni $L__BB167_93; $L__BB167_38: and.b32 %r118, %r41, 8; setp.eq.s32 %p34, %r118, 0; setp.ne.s32 %p35, %r32, 8; mov.f32 %f496, 0f00000000; or.pred %p36, %p35, %p34; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p36 bra $L__BB167_40; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r119, [%rd65+60]; add.s32 %r120, %r119, %r29; mul.wide.s32 %rd126, %r120, 12; add.s64 %rd127, %rd124, %rd126; ld.global.f32 %f498, [%rd127+8]; ld.global.f32 %f497, [%rd127+4]; ld.global.f32 %f496, [%rd127]; bra.uni $L__BB167_40; $L__BB167_130: cvt.u32.u64 %r202, %rd53; mad.lo.s32 %r203, %r202, 3, %r29; ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r203, 4; add.s64 %rd287, %rd241, %rd242; add.s64 %rd288, %rd287, 4; add.s32 %r204, %r203, 2; mul.wide.s32 %rd243, %r204, 4; add.s64 %rd289, %rd241, %rd243; $L__BB167_131: ld.local.f32 %f404, [%rd65+68]; ld.global.f32 %f405, [%rd288]; ld.local.f32 %f406, [%rd65+72]; ld.global.f32 %f407, [%rd289]; mul.ftz.f32 %f408, %f407, %f406; fma.rn.ftz.f32 %f409, %f405, %f404, %f408; mov.f32 %f410, 0f3F800000; sub.ftz.f32 %f411, %f410, %f404; sub.ftz.f32 %f412, %f411, %f406; ld.global.f32 %f413, [%rd287]; fma.rn.ftz.f32 %f531, %f413, %f412, %f409; $L__BB167_143: setp.eq.s32 %p101, %r64, 1; @%p101 bra $L__BB167_147; bra.uni $L__BB167_144; $L__BB167_147: st.local.f32 [%rd16], %f531; bra.uni $L__BB167_157; $L__BB167_144: setp.eq.s32 %p102, %r64, 0; @%p102 bra $L__BB167_146; mov.u32 %r221, 1065353216; st.local.u32 [%rd16], %r221; bra.uni $L__BB167_157; $L__BB167_146: st.local.f32 [%rd16], %f531; st.local.f32 [%rd16+4], %f531; st.local.f32 [%rd16+8], %f531; bra.uni $L__BB167_157; $L__BB167_105: cvt.u32.u64 %r167, %rd51; mad.lo.s32 %r168, %r167, 3, %r29; ld.const.u64 %rd196, [kernel_params+2328]; cvta.to.global.u64 %rd197, %rd196; mul.wide.s32 %rd198, %r168, 8; add.s64 %rd199, %rd197, %rd198; ld.global.v2.f32 {%f527, %f528}, [%rd199]; ld.global.v2.f32 {%f525, %f526}, [%rd199+8]; ld.global.v2.f32 {%f523, %f524}, [%rd199+16]; $L__BB167_106: ld.local.f32 %f370, [%rd65+68]; ld.local.f32 %f371, [%rd65+72]; mul.ftz.f32 %f372, %f523, %f371; mul.ftz.f32 %f373, %f524, %f371; fma.rn.ftz.f32 %f374, %f525, %f370, %f372; fma.rn.ftz.f32 %f375, %f526, %f370, %f373; mov.f32 %f376, 0f3F800000; sub.ftz.f32 %f377, %f376, %f370; sub.ftz.f32 %f378, %f377, %f371; fma.rn.ftz.f32 %f530, %f528, %f378, %f375; fma.rn.ftz.f32 %f529, %f527, %f378, %f374; $L__BB167_118: setp.eq.s32 %p82, %r64, 1; @%p82 bra $L__BB167_122; bra.uni $L__BB167_119; $L__BB167_122: st.local.f32 [%rd16], %f529; bra.uni $L__BB167_157; $L__BB167_119: setp.eq.s32 %p83, %r64, 0; @%p83 bra $L__BB167_121; mov.u32 %r185, 1065353216; st.local.u32 [%rd16], %r185; bra.uni $L__BB167_157; $L__BB167_121: st.local.f32 [%rd16], %f529; st.local.f32 [%rd16+4], %f530; mov.u32 %r186, 0; st.local.u32 [%rd16+8], %r186; bra.uni $L__BB167_157; $L__BB167_87: and.b32 %r148, %r32, 131; setp.eq.s32 %p58, %r148, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p58 bra $L__BB167_93; setp.ne.s32 %p59, %r32, 128; @%p59 bra $L__BB167_90; ld.local.u32 %r149, [%rd65+60]; add.s32 %r29, %r149, %r29; $L__BB167_90: ld.const.u64 %rd168, [kernel_params+2344]; cvta.to.global.u64 %rd169, %rd168; mul.wide.s32 %rd170, %r29, 16; add.s64 %rd171, %rd169, %rd170; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd171]; bra.uni $L__BB167_93; $L__BB167_34: and.b32 %r116, %r32, 131; setp.eq.s32 %p32, %r116, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p32 bra $L__BB167_40; setp.ne.s32 %p33, %r32, 128; @%p33 bra $L__BB167_37; ld.local.u32 %r117, [%rd65+60]; add.s32 %r29, %r117, %r29; $L__BB167_37: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r29, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f498, [%rd122+8]; ld.global.f32 %f497, [%rd122+4]; ld.global.f32 %f496, [%rd122]; bra.uni $L__BB167_40; $L__BB167_82: and.b32 %r137, %r32, 7; setp.eq.s32 %p54, %r137, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p54 bra $L__BB167_93; cvt.u32.u64 %r138, %rd47; setp.eq.s32 %p55, %r32, 4; selp.b32 %r139, %r138, 0, %p55; add.s32 %r140, %r29, %r139; ld.const.u64 %rd154, [kernel_params+2344]; cvta.to.global.u64 %rd155, %rd154; mul.wide.s32 %rd156, %r140, 16; add.s64 %rd157, %rd155, %rd156; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd157]; bra.uni $L__BB167_93; $L__BB167_29: and.b32 %r105, %r32, 7; setp.eq.s32 %p28, %r105, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p28 bra $L__BB167_40; cvt.u32.u64 %r106, %rd18; setp.eq.s32 %p29, %r32, 4; selp.b32 %r107, %r106, 0, %p29; add.s32 %r108, %r29, %r107; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r108, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f498, [%rd108+8]; ld.global.f32 %f497, [%rd108+4]; ld.global.f32 %f496, [%rd108]; bra.uni $L__BB167_40; $L__BB167_52: cvt.u32.u64 %r135, %rd47; mad.lo.s32 %r136, %r135, 3, %r29; cvt.s64.s32 %rd48, %r136; setp.eq.s32 %p44, %r32, 32; @%p44 bra $L__BB167_80; bra.uni $L__BB167_53; $L__BB167_80: ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd48, 4; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd152]; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd152+16]; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd152+32]; bra.uni $L__BB167_81; $L__BB167_27: cvt.u32.u64 %r101, %rd18; mad.lo.s32 %r102, %r101, 3, %r29; ld.const.u64 %rd99, [kernel_params+2336]; cvta.to.global.u64 %rd100, %rd99; mul.wide.s32 %rd101, %r102, 12; add.s64 %rd284, %rd100, %rd101; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r103, %r102, 1; mul.wide.s32 %rd102, %r103, 12; add.s64 %rd281, %rd100, %rd102; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r104, %r102, 2; mul.wide.s32 %rd103, %r104, 12; add.s64 %rd278, %rd100, %rd103; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; $L__BB167_28: ld.local.f32 %f179, [%rd65+68]; ld.global.f32 %f180, [%rd281]; ld.global.f32 %f181, [%rd282]; ld.global.f32 %f182, [%rd283]; ld.local.f32 %f183, [%rd65+72]; ld.global.f32 %f184, [%rd278]; mul.ftz.f32 %f185, %f184, %f183; ld.global.f32 %f186, [%rd279]; mul.ftz.f32 %f187, %f186, %f183; ld.global.f32 %f188, [%rd280]; mul.ftz.f32 %f189, %f188, %f183; fma.rn.ftz.f32 %f190, %f180, %f179, %f185; fma.rn.ftz.f32 %f191, %f181, %f179, %f187; fma.rn.ftz.f32 %f192, %f182, %f179, %f189; mov.f32 %f193, 0f3F800000; sub.ftz.f32 %f194, %f193, %f179; sub.ftz.f32 %f195, %f194, %f183; ld.global.f32 %f196, [%rd284]; ld.global.f32 %f197, [%rd285]; ld.global.f32 %f198, [%rd286]; fma.rn.ftz.f32 %f498, %f198, %f195, %f192; fma.rn.ftz.f32 %f497, %f197, %f195, %f191; fma.rn.ftz.f32 %f496, %f196, %f195, %f190; $L__BB167_40: setp.eq.s32 %p37, %r64, 1; @%p37 bra $L__BB167_44; bra.uni $L__BB167_41; $L__BB167_44: add.ftz.f32 %f220, %f496, %f497; add.ftz.f32 %f221, %f498, %f220; mul.ftz.f32 %f222, %f221, 0f3EAAAAAB; st.local.f32 [%rd16], %f222; bra.uni $L__BB167_157; $L__BB167_41: setp.eq.s32 %p38, %r64, 0; @%p38 bra $L__BB167_43; mov.u32 %r121, 1065353216; st.local.u32 [%rd16], %r121; bra.uni $L__BB167_157; $L__BB167_43: st.local.f32 [%rd16], %f496; st.local.f32 [%rd16+4], %f497; st.local.f32 [%rd16+8], %f498; bra.uni $L__BB167_157; $L__BB167_53: ld.const.u64 %rd146, [kernel_params+2352]; cvta.to.global.u64 %rd147, %rd146; shl.b64 %rd148, %rd48, 2; add.s64 %rd49, %rd147, %rd148; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd49]; cvt.rn.f32.u16 %f235, %rs20; mul.ftz.f32 %f38, %f235, 0f3B808081; cvt.rn.f32.u16 %f236, %rs21; mul.ftz.f32 %f39, %f236, 0f3B808081; cvt.rn.f32.u16 %f237, %rs22; mul.ftz.f32 %f40, %f237, 0f3B808081; setp.lt.ftz.f32 %p45, %f38, 0f3D25AEE6; @%p45 bra $L__BB167_55; bra.uni $L__BB167_54; $L__BB167_55: mul.ftz.f32 %f515, %f38, 0f3D9E8391; bra.uni $L__BB167_56; $L__BB167_54: add.ftz.f32 %f238, %f38, 0f3D6147AE; mul.ftz.f32 %f239, %f238, 0f3F72A76F; lg2.approx.ftz.f32 %f240, %f239; mul.ftz.f32 %f241, %f240, 0f4019999A; ex2.approx.ftz.f32 %f515, %f241; $L__BB167_56: setp.lt.ftz.f32 %p46, %f39, 0f3D25AEE6; @%p46 bra $L__BB167_58; bra.uni $L__BB167_57; $L__BB167_58: mul.ftz.f32 %f516, %f39, 0f3D9E8391; bra.uni $L__BB167_59; $L__BB167_57: add.ftz.f32 %f242, %f39, 0f3D6147AE; mul.ftz.f32 %f243, %f242, 0f3F72A76F; lg2.approx.ftz.f32 %f244, %f243; mul.ftz.f32 %f245, %f244, 0f4019999A; ex2.approx.ftz.f32 %f516, %f245; $L__BB167_59: setp.lt.ftz.f32 %p47, %f40, 0f3D25AEE6; @%p47 bra $L__BB167_61; bra.uni $L__BB167_60; $L__BB167_61: mul.ftz.f32 %f517, %f40, 0f3D9E8391; bra.uni $L__BB167_62; $L__BB167_60: add.ftz.f32 %f246, %f40, 0f3D6147AE; mul.ftz.f32 %f247, %f246, 0f3F72A76F; lg2.approx.ftz.f32 %f248, %f247; mul.ftz.f32 %f249, %f248, 0f4019999A; ex2.approx.ftz.f32 %f517, %f249; $L__BB167_62: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd49+4]; cvt.rn.f32.u16 %f250, %rs24; mul.ftz.f32 %f50, %f250, 0f3B808081; cvt.rn.f32.u16 %f251, %rs25; mul.ftz.f32 %f51, %f251, 0f3B808081; cvt.rn.f32.u16 %f252, %rs26; mul.ftz.f32 %f52, %f252, 0f3B808081; setp.lt.ftz.f32 %p48, %f50, 0f3D25AEE6; @%p48 bra $L__BB167_64; bra.uni $L__BB167_63; $L__BB167_64: mul.ftz.f32 %f511, %f50, 0f3D9E8391; bra.uni $L__BB167_65; $L__BB167_63: add.ftz.f32 %f253, %f50, 0f3D6147AE; mul.ftz.f32 %f254, %f253, 0f3F72A76F; lg2.approx.ftz.f32 %f255, %f254; mul.ftz.f32 %f256, %f255, 0f4019999A; ex2.approx.ftz.f32 %f511, %f256; $L__BB167_65: setp.lt.ftz.f32 %p49, %f51, 0f3D25AEE6; @%p49 bra $L__BB167_67; bra.uni $L__BB167_66; $L__BB167_67: mul.ftz.f32 %f512, %f51, 0f3D9E8391; bra.uni $L__BB167_68; $L__BB167_66: add.ftz.f32 %f257, %f51, 0f3D6147AE; mul.ftz.f32 %f258, %f257, 0f3F72A76F; lg2.approx.ftz.f32 %f259, %f258; mul.ftz.f32 %f260, %f259, 0f4019999A; ex2.approx.ftz.f32 %f512, %f260; $L__BB167_68: setp.lt.ftz.f32 %p50, %f52, 0f3D25AEE6; @%p50 bra $L__BB167_70; bra.uni $L__BB167_69; $L__BB167_70: mul.ftz.f32 %f513, %f52, 0f3D9E8391; bra.uni $L__BB167_71; $L__BB167_69: add.ftz.f32 %f261, %f52, 0f3D6147AE; mul.ftz.f32 %f262, %f261, 0f3F72A76F; lg2.approx.ftz.f32 %f263, %f262; mul.ftz.f32 %f264, %f263, 0f4019999A; ex2.approx.ftz.f32 %f513, %f264; $L__BB167_71: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd49+8]; cvt.rn.f32.u16 %f265, %rs28; mul.ftz.f32 %f62, %f265, 0f3B808081; cvt.rn.f32.u16 %f266, %rs29; mul.ftz.f32 %f63, %f266, 0f3B808081; cvt.rn.f32.u16 %f267, %rs30; mul.ftz.f32 %f64, %f267, 0f3B808081; setp.lt.ftz.f32 %p51, %f62, 0f3D25AEE6; @%p51 bra $L__BB167_73; bra.uni $L__BB167_72; $L__BB167_73: mul.ftz.f32 %f507, %f62, 0f3D9E8391; bra.uni $L__BB167_74; $L__BB167_72: add.ftz.f32 %f268, %f62, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f507, %f271; $L__BB167_74: setp.lt.ftz.f32 %p52, %f63, 0f3D25AEE6; @%p52 bra $L__BB167_76; bra.uni $L__BB167_75; $L__BB167_76: mul.ftz.f32 %f508, %f63, 0f3D9E8391; bra.uni $L__BB167_77; $L__BB167_75: add.ftz.f32 %f272, %f63, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f508, %f275; $L__BB167_77: cvt.rn.f32.u16 %f276, %rs23; mul.ftz.f32 %f518, %f276, 0f3B808081; cvt.rn.f32.u16 %f277, %rs27; mul.ftz.f32 %f514, %f277, 0f3B808081; cvt.rn.f32.u16 %f278, %rs31; mul.ftz.f32 %f510, %f278, 0f3B808081; setp.lt.ftz.f32 %p53, %f64, 0f3D25AEE6; @%p53 bra $L__BB167_79; bra.uni $L__BB167_78; $L__BB167_79: mul.ftz.f32 %f509, %f64, 0f3D9E8391; bra.uni $L__BB167_81; $L__BB167_78: add.ftz.f32 %f279, %f64, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f509, %f282; $L__BB167_81: ld.local.f32 %f295, [%rd65+68]; ld.local.f32 %f296, [%rd65+72]; mul.ftz.f32 %f297, %f507, %f296; mul.ftz.f32 %f298, %f508, %f296; mul.ftz.f32 %f299, %f509, %f296; mul.ftz.f32 %f300, %f510, %f296; fma.rn.ftz.f32 %f301, %f511, %f295, %f297; fma.rn.ftz.f32 %f302, %f512, %f295, %f298; fma.rn.ftz.f32 %f303, %f513, %f295, %f299; fma.rn.ftz.f32 %f304, %f514, %f295, %f300; mov.f32 %f305, 0f3F800000; sub.ftz.f32 %f306, %f305, %f295; sub.ftz.f32 %f307, %f306, %f296; fma.rn.ftz.f32 %f522, %f518, %f307, %f304; fma.rn.ftz.f32 %f521, %f517, %f307, %f303; fma.rn.ftz.f32 %f520, %f516, %f307, %f302; fma.rn.ftz.f32 %f519, %f515, %f307, %f301; $L__BB167_93: setp.eq.s32 %p63, %r64, 1; @%p63 bra $L__BB167_97; bra.uni $L__BB167_94; $L__BB167_97: add.ftz.f32 %f355, %f519, %f520; add.ftz.f32 %f356, %f521, %f355; mul.ftz.f32 %f357, %f356, 0f3EAAAAAB; st.local.f32 [%rd16], %f357; bra.uni $L__BB167_157; $L__BB167_94: setp.eq.s32 %p64, %r64, 0; @%p64 bra $L__BB167_96; st.local.f32 [%rd16], %f522; bra.uni $L__BB167_157; $L__BB167_96: st.local.f32 [%rd16], %f519; st.local.f32 [%rd16+4], %f520; st.local.f32 [%rd16+8], %f521; $L__BB167_157: ret; } .func _Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB168_2; bra.uni $L__BB168_1; $L__BB168_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB168_3; $L__BB168_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB168_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB168_5; bra.uni $L__BB168_4; $L__BB168_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB168_9; bra.uni $L__BB168_6; $L__BB168_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB168_12; bra.uni $L__BB168_10; $L__BB168_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB168_13; $L__BB168_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB168_16; $L__BB168_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB168_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB168_13; $L__BB168_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB168_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB168_15; bra.uni $L__BB168_14; $L__BB168_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB168_16; $L__BB168_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB168_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB168_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB168_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB168_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB169_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB169_5; bra.uni $L__BB169_2; $L__BB169_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB169_8; bra.uni $L__BB169_6; $L__BB169_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB169_9; $L__BB169_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB169_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB169_9; $L__BB169_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB169_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB169_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB169_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB169_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB169_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB169_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB169_20; $L__BB169_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB169_17; add.s32 %r172, %r172, 2; bra.uni $L__BB169_18; $L__BB169_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB169_24; $L__BB169_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB169_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB169_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB169_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB169_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB169_24; $L__BB169_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB169_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB169_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB169_37; bra.uni $L__BB169_26; $L__BB169_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB169_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB169_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB169_45; $L__BB169_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB169_45; $L__BB169_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB169_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB169_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB169_30; mov.u64 %rd108, 0; { // callseq 936, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 936 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB169_45; $L__BB169_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB169_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB169_9; $L__BB169_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB169_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB169_45; $L__BB169_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB169_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB169_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB169_34; $L__BB169_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB169_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB169_45; $L__BB169_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB169_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB169_45; $L__BB169_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB169_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB169_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB169_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB169_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB169_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB169_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB169_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB169_56; $L__BB169_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB169_53; add.s32 %r187, %r187, 2; bra.uni $L__BB169_54; $L__BB169_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB169_60; $L__BB169_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB169_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB169_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB169_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB169_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB169_60; $L__BB169_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB169_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB169_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB169_72; bra.uni $L__BB169_62; $L__BB169_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB169_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB169_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB169_79; $L__BB169_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB169_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB169_65; mov.u64 %rd160, 0; { // callseq 937, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 937 bra.uni $L__BB169_79; $L__BB169_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB169_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB169_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB169_69; $L__BB169_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB169_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB169_79; $L__BB169_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB169_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB169_79; $L__BB169_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB169_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB169_79; $L__BB169_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB169_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB169_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB169_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB169_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB169_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB169_84; bra.uni $L__BB169_83; $L__BB169_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB169_85; $L__BB169_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB169_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB170_2; bra.uni $L__BB170_1; $L__BB170_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB170_3; $L__BB170_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB170_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB170_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB170_8; bra.uni $L__BB170_5; $L__BB170_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB170_11; bra.uni $L__BB170_9; $L__BB170_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB170_12; $L__BB170_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB170_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB170_12; $L__BB170_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB170_12: @%p8 bra $L__BB170_14; bra.uni $L__BB170_13; $L__BB170_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB170_15; $L__BB170_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB170_15: @%p8 bra $L__BB170_17; bra.uni $L__BB170_16; $L__BB170_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB170_18; $L__BB170_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB170_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB170_19; bra.uni $L__BB170_20; $L__BB170_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB170_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB170_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB170_22: @%p7 bra $L__BB170_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB170_25; bra.uni $L__BB170_24; $L__BB170_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB170_26; $L__BB170_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB170_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB170_27; bra.uni $L__BB170_34; $L__BB170_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB170_30; bra.uni $L__BB170_28; $L__BB170_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB170_32; bra.uni $L__BB170_31; $L__BB170_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB170_33; $L__BB170_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB170_29; bra.uni $L__BB170_34; $L__BB170_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB170_33; $L__BB170_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB170_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB170_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB170_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB170_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB170_12; } .func _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4 ) { .reg .pred %p<4>; .reg .f32 %f<6>; .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.f32 %f3, [_Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd1, [_Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; mov.f32 %f5, 0f00000000; setp.eq.s32 %p1, %r2, 9; @%p1 bra $L__BB171_3; setp.ne.s32 %p2, %r2, 8; @%p2 bra $L__BB171_4; and.b32 %r4, %r1, 1; setp.eq.b32 %p3, %r4, 1; selp.f32 %f5, 0f3F800000, 0f00000000, %p3; bra.uni $L__BB171_4; $L__BB171_3: mov.f32 %f5, %f3; $L__BB171_4: cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f5; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB172_2; bra.uni $L__BB172_1; $L__BB172_2: mov.b32 %f2764, %r395; bra.uni $L__BB172_3; $L__BB172_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB172_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB172_5; bra.uni $L__BB172_4; $L__BB172_5: mov.b32 %f9, %r396; bra.uni $L__BB172_6; $L__BB172_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB172_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB172_8; bra.uni $L__BB172_7; $L__BB172_8: mov.b32 %f2766, %r397; bra.uni $L__BB172_9; $L__BB172_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB172_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB172_11; bra.uni $L__BB172_10; $L__BB172_11: mov.b32 %f15, %r398; bra.uni $L__BB172_12; $L__BB172_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB172_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB172_14; bra.uni $L__BB172_13; $L__BB172_14: mov.b32 %f2768, %r399; bra.uni $L__BB172_15; $L__BB172_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB172_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB172_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB172_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB172_18; bra.uni $L__BB172_385; $L__BB172_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB172_259; bra.uni $L__BB172_19; $L__BB172_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB172_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB172_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB172_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB172_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB172_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB172_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB172_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB172_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB172_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB172_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB172_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB172_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB172_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB172_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB172_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB172_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB172_362; $L__BB172_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB172_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB172_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB172_40; bra.uni $L__BB172_29; $L__BB172_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB172_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB172_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB172_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB172_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB172_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB172_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB172_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB172_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB172_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB172_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB172_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB172_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB172_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB172_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB172_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB172_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB172_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB172_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB172_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB172_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB172_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB172_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB172_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB172_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB172_136; $L__BB172_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB172_371; bra.uni $L__BB172_366; $L__BB172_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB172_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB172_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB172_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB172_383; $L__BB172_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB172_149; bra.uni $L__BB172_139; $L__BB172_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB172_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB172_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB172_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB172_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB172_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB172_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB172_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB172_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB172_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB172_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB172_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB172_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB172_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB172_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB172_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB172_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB172_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB172_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB172_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB172_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB172_250; $L__BB172_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB172_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB172_253; bra.uni $L__BB172_362; $L__BB172_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB172_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB172_304; @%p171 bra $L__BB172_303; bra.uni $L__BB172_256; $L__BB172_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB172_305; $L__BB172_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB172_305; $L__BB172_256: @%p172 bra $L__BB172_302; bra.uni $L__BB172_257; $L__BB172_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB172_305; $L__BB172_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB172_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB172_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB172_308; bra.uni $L__BB172_306; $L__BB172_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB172_309; $L__BB172_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB172_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB172_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB172_316; @%p171 bra $L__BB172_315; bra.uni $L__BB172_311; $L__BB172_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB172_317; $L__BB172_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB172_317; $L__BB172_311: @%p172 bra $L__BB172_314; bra.uni $L__BB172_312; $L__BB172_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB172_317; $L__BB172_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB172_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB172_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB172_320; bra.uni $L__BB172_318; $L__BB172_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB172_321; $L__BB172_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB172_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB172_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB172_328; @%p171 bra $L__BB172_327; bra.uni $L__BB172_323; $L__BB172_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB172_329; $L__BB172_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB172_329; $L__BB172_323: @%p172 bra $L__BB172_326; bra.uni $L__BB172_324; $L__BB172_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB172_329; $L__BB172_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB172_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB172_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB172_332; bra.uni $L__BB172_330; $L__BB172_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB172_333; $L__BB172_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB172_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB172_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB172_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_362; $L__BB172_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB172_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB172_31; bra.uni $L__BB172_136; $L__BB172_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB172_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB172_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB172_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB172_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB172_107; @%p49 bra $L__BB172_106; bra.uni $L__BB172_37; $L__BB172_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB172_108; $L__BB172_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB172_108; $L__BB172_37: @%p50 bra $L__BB172_105; bra.uni $L__BB172_38; $L__BB172_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB172_108; $L__BB172_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB172_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB172_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB172_111; bra.uni $L__BB172_109; $L__BB172_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB172_112; $L__BB172_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB172_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB172_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB172_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB172_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB172_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB172_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_136; $L__BB172_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB172_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB172_368; bra.uni $L__BB172_383; $L__BB172_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB172_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB172_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB172_383; $L__BB172_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB172_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB172_141; bra.uni $L__BB172_250; $L__BB172_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB172_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB172_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB172_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB172_206; @%p101 bra $L__BB172_205; bra.uni $L__BB172_146; $L__BB172_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB172_207; $L__BB172_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB172_207; $L__BB172_146: @%p102 bra $L__BB172_204; bra.uni $L__BB172_147; $L__BB172_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB172_207; $L__BB172_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB172_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB172_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB172_210; bra.uni $L__BB172_208; $L__BB172_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB172_211; $L__BB172_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB172_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB172_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB172_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB172_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB172_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_250; $L__BB172_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB172_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB172_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB172_295; @%p151 bra $L__BB172_294; bra.uni $L__BB172_290; $L__BB172_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB172_296; $L__BB172_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB172_296; $L__BB172_290: @%p152 bra $L__BB172_293; bra.uni $L__BB172_291; $L__BB172_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB172_296; $L__BB172_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB172_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB172_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB172_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB172_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB172_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB172_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB172_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_362; $L__BB172_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB172_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB172_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB172_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB172_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB172_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB172_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB172_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB172_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB172_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB172_279; bra.uni $L__BB172_362; $L__BB172_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB172_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB172_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB172_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB172_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB172_96; @%p39 bra $L__BB172_95; bra.uni $L__BB172_91; $L__BB172_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB172_97; $L__BB172_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB172_97; $L__BB172_91: @%p40 bra $L__BB172_94; bra.uni $L__BB172_92; $L__BB172_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB172_97; $L__BB172_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB172_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB172_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB172_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB172_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB172_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB172_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB172_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB172_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB172_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_136; $L__BB172_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB172_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB172_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB172_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB172_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB172_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB172_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB172_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB172_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB172_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB172_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB172_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB172_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB172_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB172_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB172_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB172_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB172_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB172_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB172_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB172_76; bra.uni $L__BB172_136; $L__BB172_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB172_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB172_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB172_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB172_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB172_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_383; $L__BB172_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB172_383; $L__BB172_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB172_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB172_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB172_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB172_196; @%p92 bra $L__BB172_195; bra.uni $L__BB172_191; $L__BB172_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB172_197; $L__BB172_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB172_197; $L__BB172_191: @%p93 bra $L__BB172_194; bra.uni $L__BB172_192; $L__BB172_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB172_197; $L__BB172_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB172_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB172_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB172_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB172_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB172_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB172_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB172_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB172_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB172_250; $L__BB172_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB172_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB172_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB172_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB172_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB172_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB172_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB172_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB172_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB172_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB172_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB172_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB172_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB172_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB172_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB172_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB172_179; bra.uni $L__BB172_250; $L__BB172_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB172_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB172_337; @%p188 bra $L__BB172_336; bra.uni $L__BB172_23; $L__BB172_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB172_338; $L__BB172_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB172_338; $L__BB172_23: @%p189 bra $L__BB172_335; bra.uni $L__BB172_24; $L__BB172_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB172_338; $L__BB172_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB172_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB172_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB172_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB172_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB172_347; @%p188 bra $L__BB172_346; bra.uni $L__BB172_342; $L__BB172_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB172_348; $L__BB172_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB172_348; $L__BB172_342: @%p189 bra $L__BB172_345; bra.uni $L__BB172_343; $L__BB172_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB172_348; $L__BB172_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB172_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB172_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB172_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB172_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB172_357; @%p188 bra $L__BB172_356; bra.uni $L__BB172_352; $L__BB172_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB172_358; $L__BB172_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB172_358; $L__BB172_352: @%p189 bra $L__BB172_355; bra.uni $L__BB172_353; $L__BB172_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB172_358; $L__BB172_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB172_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB172_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB172_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB172_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB172_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB172_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB172_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB172_364: mov.f32 %f3392, %f3387; bra.uni $L__BB172_385; $L__BB172_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB172_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB172_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB172_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB172_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB172_128; @%p59 bra $L__BB172_127; bra.uni $L__BB172_123; $L__BB172_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB172_129; $L__BB172_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB172_129; $L__BB172_123: @%p60 bra $L__BB172_126; bra.uni $L__BB172_124; $L__BB172_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB172_129; $L__BB172_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB172_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB172_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB172_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB172_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB172_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB172_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB172_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB172_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB172_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB172_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB172_385; $L__BB172_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB172_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB172_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB172_385; $L__BB172_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB172_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB172_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB172_224; @%p120 bra $L__BB172_223; bra.uni $L__BB172_219; $L__BB172_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB172_225; $L__BB172_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB172_225; $L__BB172_219: @%p121 bra $L__BB172_222; bra.uni $L__BB172_220; $L__BB172_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB172_225; $L__BB172_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB172_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB172_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB172_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB172_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB172_234; @%p120 bra $L__BB172_233; bra.uni $L__BB172_229; $L__BB172_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB172_235; $L__BB172_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB172_235; $L__BB172_229: @%p121 bra $L__BB172_232; bra.uni $L__BB172_230; $L__BB172_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB172_235; $L__BB172_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB172_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB172_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB172_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB172_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB172_244; @%p120 bra $L__BB172_243; bra.uni $L__BB172_239; $L__BB172_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB172_245; $L__BB172_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB172_245; $L__BB172_239: @%p121 bra $L__BB172_242; bra.uni $L__BB172_240; $L__BB172_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB172_245; $L__BB172_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB172_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB172_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB172_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB172_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB172_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB172_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB172_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB172_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB172_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB172_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB172_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB172_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB172_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB172_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB172_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB172_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB172_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB172_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB172_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB174_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB174_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB174_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB174_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .align 4 .b8 _Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0[16], .param .b32 _Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1 ) { .reg .pred %p<2>; .reg .b32 %r<5>; ld.param.u32 %r1, [_Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u8 %r2, [_Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0+4]; setp.eq.s32 %p1, %r2, 39; selp.b32 %r3, 5, 1, %p1; add.s32 %r4, %r3, %r1; st.param.b32 [func_retval0+0], %r4; ret; } .func _Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<107>; .reg .b16 %rs<39>; .reg .f32 %f<535>; .reg .b32 %r<242>; .reg .b64 %rd<290>; ld.param.u64 %rd63, [_Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd64, [_Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r64, [_Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r63, [_Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r62, [_Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd65, %rd63; add.s64 %rd1, %rd65, 76; ld.local.u32 %r3, [%rd65+76]; cvt.s64.s32 %rd2, %r3; setp.eq.s32 %p1, %r3, -1; mov.u32 %r29, 0; mov.u32 %r30, %r29; mov.u32 %r236, %r64; mov.u32 %r32, %r29; @%p1 bra $L__BB176_16; cvt.u64.u32 %rd3, %r62; ld.local.s32 %rd4, [%rd1+-16]; ld.const.u64 %rd66, [kernel_params+2160]; cvta.to.global.u64 %rd67, %rd66; mul.lo.s64 %rd68, %rd2, 208; add.s64 %rd69, %rd67, %rd68; ld.global.u32 %r4, [%rd69+164]; ld.local.u32 %r69, [%rd1+-12]; and.b32 %r70, %r69, 1; setp.eq.b32 %p2, %r70, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r225, 0; @%p5 bra $L__BB176_5; cvt.u32.u64 %r72, %rd4; setp.eq.s32 %p6, %r72, -1; mov.u32 %r224, -1; @%p6 bra $L__BB176_4; ld.const.u64 %rd70, [kernel_params+2240]; cvta.to.global.u64 %rd71, %rd70; shl.b64 %rd72, %rd4, 2; add.s64 %rd73, %rd71, %rd72; ld.global.u32 %r224, [%rd73]; $L__BB176_4: setp.ne.s32 %p7, %r224, -1; selp.u32 %r225, 1, 0, %p7; $L__BB176_5: ld.const.u64 %rd74, [kernel_params+2312]; cvta.to.global.u64 %rd5, %rd74; add.s32 %r227, %r225, %r4; mul.wide.u32 %rd75, %r227, 16; add.s64 %rd76, %rd5, %rd75; add.s64 %rd276, %rd76, 14; add.s64 %rd277, %rd76, 15; ld.global.u32 %r226, [%rd76+8]; ld.global.u16 %rs37, [%rd76+12]; ld.global.u64 %rd275, [%rd76]; setp.eq.s64 %p8, %rd275, %rd3; @%p8 bra $L__BB176_11; $L__BB176_6: setp.eq.s64 %p9, %rd275, 0; @%p9 bra $L__BB176_8; add.s32 %r227, %r227, 2; bra.uni $L__BB176_9; $L__BB176_8: setp.eq.s16 %p10, %rs37, 0; mov.u32 %r233, -1; mov.u32 %r230, 0; mov.u32 %r227, %r226; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p10 bra $L__BB176_15; $L__BB176_9: cvt.u64.u32 %rd10, %r227; mul.wide.u32 %rd77, %r227, 16; add.s64 %rd78, %rd5, %rd77; ld.global.u32 %r226, [%rd78+8]; ld.global.u16 %rs37, [%rd78+12]; ld.global.u64 %rd275, [%rd78]; setp.ne.s64 %p11, %rd275, %rd3; @%p11 bra $L__BB176_6; shl.b64 %rd79, %rd10, 4; add.s64 %rd80, %rd5, %rd79; add.s64 %rd276, %rd80, 14; add.s64 %rd277, %rd80, 15; $L__BB176_11: ld.global.u8 %rs5, [%rd277]; ld.global.u8 %rs6, [%rd276]; cvt.u32.u64 %r77, %rd4; setp.ne.s32 %p12, %r77, -1; @%p12 bra $L__BB176_14; mov.u32 %r233, -1; mov.u32 %r230, 0; add.s16 %rs19, %rs37, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB176_14; setp.ne.s16 %p14, %rs37, 1024; mov.u32 %r231, %r230; mov.u32 %r232, %r230; @%p14 bra $L__BB176_15; $L__BB176_14: setp.eq.s16 %p15, %rs37, 0; selp.b32 %r233, -1, %r226, %p15; cvt.u32.u16 %r230, %rs37; cvt.u32.u16 %r82, %rs5; and.b32 %r232, %r82, 255; cvt.u32.u16 %r83, %rs6; and.b32 %r231, %r83, 255; $L__BB176_15: setp.eq.s32 %p16, %r233, -1; selp.b32 %r29, 0, %r233, %p16; selp.b32 %r30, 0, %r232, %p16; selp.b32 %r236, %r64, %r231, %p16; selp.b32 %r32, 0, %r230, %p16; $L__BB176_16: setp.eq.s32 %p17, %r32, 0; setp.eq.s32 %p18, %r62, 7; and.pred %p19, %p18, %p17; cvta.to.local.u64 %rd81, %rd64; mul.wide.u32 %rd82, %r63, 4; add.s64 %rd16, %rd81, %rd82; @%p19 bra $L__BB176_148; bra.uni $L__BB176_17; $L__BB176_148: cvt.u32.u64 %r222, %rd2; setp.eq.s32 %p103, %r222, -1; ld.local.v4.f32 {%f534, %f533, %f532, %f426}, [%rd65]; @%p103 bra $L__BB176_152; ld.local.u8 %rs35, [%rd65+56]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p104, %rs36, 0; @%p104 bra $L__BB176_151; bra.uni $L__BB176_150; $L__BB176_151: ld.const.u64 %rd271, [kernel_params+2160]; cvta.to.global.u64 %rd272, %rd271; mul.wide.s32 %rd273, %r3, 208; add.s64 %rd274, %rd272, %rd273; ld.global.v4.f32 {%f460, %f461, %f462, %f463}, [%rd274+48]; ld.global.v4.f32 {%f468, %f469, %f470, %f471}, [%rd274+64]; ld.global.v4.f32 {%f476, %f477, %f478, %f479}, [%rd274+80]; mul.ftz.f32 %f484, %f533, %f461; fma.rn.ftz.f32 %f485, %f534, %f460, %f484; fma.rn.ftz.f32 %f486, %f532, %f462, %f485; add.ftz.f32 %f173, %f463, %f486; mul.ftz.f32 %f487, %f533, %f469; fma.rn.ftz.f32 %f488, %f534, %f468, %f487; fma.rn.ftz.f32 %f489, %f532, %f470, %f488; add.ftz.f32 %f174, %f471, %f489; mul.ftz.f32 %f490, %f533, %f477; fma.rn.ftz.f32 %f491, %f534, %f476, %f490; fma.rn.ftz.f32 %f492, %f532, %f478, %f491; add.ftz.f32 %f532, %f479, %f492; mov.f32 %f533, %f174; mov.f32 %f534, %f173; bra.uni $L__BB176_152; $L__BB176_17: setp.eq.s32 %p20, %r236, 0; @%p20 bra $L__BB176_123; setp.eq.s32 %p21, %r236, 1; @%p21 bra $L__BB176_98; bra.uni $L__BB176_19; $L__BB176_98: mov.u32 %r48, 1; ld.local.u32 %r51, [%rd65+64]; and.b32 %r153, %r51, 1; setp.eq.b32 %p65, %r153, 1; mov.pred %p66, 0; xor.pred %p67, %p65, %p66; not.pred %p68, %p67; @%p68 bra $L__BB176_109; bra.uni $L__BB176_99; $L__BB176_109: and.b32 %r173, %r51, 6; setp.eq.s32 %p75, %r173, 0; @%p75 bra $L__BB176_116; and.b32 %r174, %r32, 768; setp.eq.s32 %p76, %r174, 0; @%p76 bra $L__BB176_112; ld.const.u64 %rd205, [kernel_params+2264]; cvta.to.global.u64 %rd206, %rd205; ld.local.u32 %r175, [%rd65+60]; mul.wide.s32 %rd208, %r175, 16; add.s64 %rd209, %rd206, %rd208; shr.s32 %r176, %r51, 7; ld.global.u32 %r177, [%rd209+4]; add.s32 %r178, %r176, %r177; ld.const.u64 %rd210, [kernel_params+2328]; cvta.to.global.u64 %rd211, %rd210; add.s32 %r179, %r178, %r29; mul.wide.s32 %rd212, %r179, 8; add.s64 %rd213, %rd211, %rd212; ld.global.v2.f32 {%f383, %f384}, [%rd213]; ld.global.v2.f32 {%f387, %f388}, [%rd213+8]; ld.local.f32 %f391, [%rd65+68]; mov.f32 %f392, 0f3F800000; sub.ftz.f32 %f393, %f392, %f391; mul.ftz.f32 %f394, %f383, %f393; mul.ftz.f32 %f395, %f384, %f393; fma.rn.ftz.f32 %f530, %f388, %f391, %f395; fma.rn.ftz.f32 %f529, %f387, %f391, %f394; bra.uni $L__BB176_118; $L__BB176_123: ld.local.u32 %r58, [%rd65+64]; and.b32 %r187, %r58, 1; setp.eq.b32 %p84, %r187, 1; mov.pred %p85, 0; xor.pred %p86, %p84, %p85; not.pred %p87, %p86; @%p87 bra $L__BB176_134; bra.uni $L__BB176_124; $L__BB176_134: and.b32 %r209, %r58, 6; setp.eq.s32 %p94, %r209, 0; @%p94 bra $L__BB176_141; and.b32 %r210, %r32, 768; setp.eq.s32 %p95, %r210, 0; @%p95 bra $L__BB176_137; ld.const.u64 %rd249, [kernel_params+2264]; cvta.to.global.u64 %rd250, %rd249; ld.local.u32 %r211, [%rd65+60]; mul.wide.s32 %rd252, %r211, 16; add.s64 %rd253, %rd250, %rd252; shr.s32 %r212, %r58, 7; ld.global.u32 %r213, [%rd253+4]; add.s32 %r214, %r212, %r213; ld.const.u64 %rd254, [kernel_params+2320]; cvta.to.global.u64 %rd255, %rd254; add.s32 %r215, %r214, %r29; mul.wide.s32 %rd256, %r215, 4; add.s64 %rd257, %rd255, %rd256; ld.local.f32 %f415, [%rd65+68]; mov.f32 %f416, 0f3F800000; sub.ftz.f32 %f417, %f416, %f415; ld.global.f32 %f418, [%rd257]; mul.ftz.f32 %f419, %f418, %f417; ld.global.f32 %f420, [%rd257+4]; fma.rn.ftz.f32 %f531, %f420, %f415, %f419; bra.uni $L__BB176_143; $L__BB176_150: ld.local.v4.f32 {%f427, %f428, %f429, %f430}, [%rd65+192]; mul.ftz.f32 %f435, %f533, %f428; fma.rn.ftz.f32 %f436, %f534, %f427, %f435; fma.rn.ftz.f32 %f437, %f532, %f429, %f436; add.ftz.f32 %f170, %f430, %f437; ld.local.v4.f32 {%f438, %f439, %f440, %f441}, [%rd65+208]; mul.ftz.f32 %f446, %f533, %f439; fma.rn.ftz.f32 %f447, %f534, %f438, %f446; fma.rn.ftz.f32 %f448, %f532, %f440, %f447; add.ftz.f32 %f171, %f441, %f448; ld.local.v4.f32 {%f449, %f450, %f451, %f452}, [%rd65+224]; mul.ftz.f32 %f457, %f533, %f450; fma.rn.ftz.f32 %f458, %f534, %f449, %f457; fma.rn.ftz.f32 %f459, %f532, %f451, %f458; add.ftz.f32 %f532, %f452, %f459; mov.f32 %f533, %f171; mov.f32 %f534, %f170; $L__BB176_152: setp.eq.s32 %p105, %r64, 1; @%p105 bra $L__BB176_156; bra.uni $L__BB176_153; $L__BB176_156: add.ftz.f32 %f493, %f533, %f534; add.ftz.f32 %f494, %f532, %f493; mul.ftz.f32 %f495, %f494, 0f3EAAAAAB; st.local.f32 [%rd16], %f495; bra.uni $L__BB176_157; $L__BB176_153: setp.eq.s32 %p106, %r64, 0; @%p106 bra $L__BB176_155; mov.u32 %r223, 1065353216; st.local.u32 [%rd16], %r223; bra.uni $L__BB176_157; $L__BB176_19: add.s32 %r87, %r236, -3; setp.lt.u32 %p22, %r87, 2; ld.local.u32 %r41, [%rd65+64]; and.b32 %r42, %r41, 1; @%p22 bra $L__BB176_45; bra.uni $L__BB176_20; $L__BB176_45: setp.eq.s32 %p39, %r42, 0; @%p39 bra $L__BB176_84; bra.uni $L__BB176_46; $L__BB176_84: and.b32 %r141, %r41, 6; setp.eq.s32 %p56, %r141, 0; @%p56 bra $L__BB176_91; and.b32 %r142, %r32, 768; setp.eq.s32 %p57, %r142, 0; @%p57 bra $L__BB176_87; ld.const.u64 %rd158, [kernel_params+2264]; cvta.to.global.u64 %rd159, %rd158; ld.local.u32 %r143, [%rd65+60]; mul.wide.s32 %rd161, %r143, 16; add.s64 %rd162, %rd159, %rd161; shr.s32 %r144, %r41, 7; ld.global.u32 %r145, [%rd162+4]; add.s32 %r146, %r144, %r145; ld.const.u64 %rd163, [kernel_params+2344]; cvta.to.global.u64 %rd164, %rd163; add.s32 %r147, %r146, %r29; mul.wide.s32 %rd165, %r147, 16; add.s64 %rd166, %rd164, %rd165; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd166]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd166+16]; ld.local.f32 %f332, [%rd65+68]; mov.f32 %f333, 0f3F800000; sub.ftz.f32 %f334, %f333, %f332; mul.ftz.f32 %f335, %f316, %f334; mul.ftz.f32 %f336, %f317, %f334; mul.ftz.f32 %f337, %f318, %f334; mul.ftz.f32 %f338, %f319, %f334; fma.rn.ftz.f32 %f522, %f327, %f332, %f338; fma.rn.ftz.f32 %f521, %f326, %f332, %f337; fma.rn.ftz.f32 %f520, %f325, %f332, %f336; fma.rn.ftz.f32 %f519, %f324, %f332, %f335; bra.uni $L__BB176_93; $L__BB176_124: ld.local.s32 %rd53, [%rd65+60]; setp.eq.s64 %p88, %rd53, -1; @%p88 bra $L__BB176_127; ld.const.u64 %rd226, [kernel_params+2240]; cvta.to.global.u64 %rd227, %rd226; shl.b64 %rd228, %rd53, 2; add.s64 %rd229, %rd227, %rd228; ld.global.u32 %r188, [%rd229]; setp.eq.s32 %p89, %r188, -1; @%p89 bra $L__BB176_127; mov.u32 %r189, 0; mov.u64 %rd230, 0; { // callseq 941, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r189; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd230; .param .b64 param3; st.param.b64 [param3+0], %rd230; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f531, [retval0+0]; } // callseq 941 bra.uni $L__BB176_143; $L__BB176_155: st.local.f32 [%rd16], %f534; st.local.f32 [%rd16+4], %f533; st.local.f32 [%rd16+8], %f532; bra.uni $L__BB176_157; $L__BB176_99: ld.local.s32 %rd51, [%rd65+60]; setp.eq.s64 %p69, %rd51, -1; @%p69 bra $L__BB176_102; ld.const.u64 %rd179, [kernel_params+2240]; cvta.to.global.u64 %rd180, %rd179; shl.b64 %rd181, %rd51, 2; add.s64 %rd182, %rd180, %rd181; ld.global.u32 %r154, [%rd182]; setp.eq.s32 %p70, %r154, -1; @%p70 bra $L__BB176_102; mov.u64 %rd183, 0; { // callseq 940, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r48; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd183; .param .b64 param3; st.param.b64 [param3+0], %rd183; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f529, [retval0+0]; ld.param.f32 %f530, [retval0+4]; } // callseq 940 bra.uni $L__BB176_118; $L__BB176_20: setp.eq.s32 %p23, %r42, 0; @%p23 bra $L__BB176_31; bra.uni $L__BB176_21; $L__BB176_31: and.b32 %r109, %r41, 6; setp.eq.s32 %p30, %r109, 0; @%p30 bra $L__BB176_38; and.b32 %r110, %r32, 768; setp.eq.s32 %p31, %r110, 0; @%p31 bra $L__BB176_34; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r111, [%rd65+60]; mul.wide.s32 %rd112, %r111, 16; add.s64 %rd113, %rd110, %rd112; shr.s32 %r112, %r41, 7; ld.global.u32 %r113, [%rd113+4]; add.s32 %r114, %r112, %r113; ld.const.u64 %rd114, [kernel_params+2336]; cvta.to.global.u64 %rd115, %rd114; add.s32 %r115, %r114, %r29; mul.wide.s32 %rd116, %r115, 12; add.s64 %rd117, %rd115, %rd116; ld.local.f32 %f202, [%rd65+68]; mov.f32 %f203, 0f3F800000; sub.ftz.f32 %f204, %f203, %f202; ld.global.f32 %f205, [%rd117]; mul.ftz.f32 %f206, %f205, %f204; ld.global.f32 %f207, [%rd117+4]; mul.ftz.f32 %f208, %f207, %f204; ld.global.f32 %f209, [%rd117+8]; mul.ftz.f32 %f210, %f209, %f204; ld.global.f32 %f211, [%rd117+12]; ld.global.f32 %f212, [%rd117+16]; ld.global.f32 %f213, [%rd117+20]; fma.rn.ftz.f32 %f498, %f213, %f202, %f210; fma.rn.ftz.f32 %f497, %f212, %f202, %f208; fma.rn.ftz.f32 %f496, %f211, %f202, %f206; bra.uni $L__BB176_40; $L__BB176_127: and.b32 %r190, %r32, 56; setp.eq.s32 %p90, %r190, 0; @%p90 bra $L__BB176_132; and.b32 %r191, %r32, 24; setp.eq.s32 %p91, %r191, 0; @%p91 bra $L__BB176_130; ld.const.u64 %rd231, [kernel_params+2232]; cvta.to.global.u64 %rd232, %rd231; shl.b64 %rd233, %rd53, 4; add.s64 %rd234, %rd232, %rd233; ld.global.v4.u32 {%r192, %r193, %r194, %r195}, [%rd234]; ld.const.u64 %rd235, [kernel_params+2320]; cvta.to.global.u64 %rd236, %rd235; add.s32 %r199, %r192, %r29; mul.wide.u32 %rd237, %r199, 4; add.s64 %rd287, %rd236, %rd237; add.s32 %r200, %r193, %r29; mul.wide.u32 %rd238, %r200, 4; add.s64 %rd288, %rd236, %rd238; add.s32 %r201, %r194, %r29; mul.wide.u32 %rd239, %r201, 4; add.s64 %rd289, %rd236, %rd239; bra.uni $L__BB176_131; $L__BB176_102: and.b32 %r155, %r32, 56; setp.eq.s32 %p71, %r155, 0; @%p71 bra $L__BB176_107; and.b32 %r156, %r32, 24; setp.eq.s32 %p72, %r156, 0; @%p72 bra $L__BB176_105; ld.const.u64 %rd184, [kernel_params+2232]; cvta.to.global.u64 %rd185, %rd184; shl.b64 %rd186, %rd51, 4; add.s64 %rd187, %rd185, %rd186; ld.global.v4.u32 {%r157, %r158, %r159, %r160}, [%rd187]; ld.const.u64 %rd188, [kernel_params+2328]; cvta.to.global.u64 %rd189, %rd188; add.s32 %r164, %r157, %r29; mul.wide.u32 %rd190, %r164, 8; add.s64 %rd191, %rd189, %rd190; ld.global.v2.f32 {%f527, %f528}, [%rd191]; add.s32 %r165, %r158, %r29; mul.wide.u32 %rd192, %r165, 8; add.s64 %rd193, %rd189, %rd192; ld.global.v2.f32 {%f525, %f526}, [%rd193]; add.s32 %r166, %r159, %r29; mul.wide.u32 %rd194, %r166, 8; add.s64 %rd195, %rd189, %rd194; ld.global.v2.f32 {%f523, %f524}, [%rd195]; bra.uni $L__BB176_106; $L__BB176_46: ld.local.s32 %rd47, [%rd65+60]; setp.eq.s64 %p40, %rd47, -1; @%p40 bra $L__BB176_49; ld.const.u64 %rd129, [kernel_params+2240]; cvta.to.global.u64 %rd130, %rd129; shl.b64 %rd131, %rd47, 2; add.s64 %rd132, %rd130, %rd131; ld.global.u32 %r122, [%rd132]; setp.eq.s32 %p41, %r122, -1; @%p41 bra $L__BB176_49; mov.u64 %rd133, 0; { // callseq 939, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd133; .param .b64 param3; st.param.b64 [param3+0], %rd133; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f519, [retval0+0]; ld.param.f32 %f520, [retval0+4]; ld.param.f32 %f521, [retval0+8]; ld.param.f32 %f522, [retval0+12]; } // callseq 939 bra.uni $L__BB176_93; $L__BB176_21: ld.local.s32 %rd18, [%rd65+60]; setp.eq.s64 %p24, %rd18, -1; @%p24 bra $L__BB176_24; ld.const.u64 %rd85, [kernel_params+2240]; cvta.to.global.u64 %rd86, %rd85; shl.b64 %rd87, %rd18, 2; add.s64 %rd88, %rd86, %rd87; ld.global.u32 %r88, [%rd88]; setp.eq.s32 %p25, %r88, -1; @%p25 bra $L__BB176_24; mov.u64 %rd89, 0; { // callseq 938, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd63; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r32; st.param.b32 [param1+4], %r236; st.param.b32 [param1+8], %r30; st.param.b32 [param1+12], %r29; .param .b64 param2; st.param.b64 [param2+0], %rd89; .param .b64 param3; st.param.b64 [param3+0], %rd89; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f496, [retval0+0]; ld.param.f32 %f497, [retval0+4]; ld.param.f32 %f498, [retval0+8]; } // callseq 938 bra.uni $L__BB176_40; $L__BB176_141: and.b32 %r218, %r58, 8; setp.eq.s32 %p98, %r218, 0; setp.ne.s32 %p99, %r32, 8; mov.f32 %f531, 0f00000000; or.pred %p100, %p99, %p98; @%p100 bra $L__BB176_143; ld.const.u64 %rd263, [kernel_params+2320]; cvta.to.global.u64 %rd264, %rd263; ld.local.u32 %r219, [%rd65+60]; add.s32 %r220, %r219, %r29; mul.wide.s32 %rd266, %r220, 4; add.s64 %rd267, %rd264, %rd266; ld.global.f32 %f531, [%rd267]; bra.uni $L__BB176_143; $L__BB176_116: and.b32 %r182, %r51, 8; setp.eq.s32 %p79, %r182, 0; setp.ne.s32 %p80, %r32, 8; mov.f32 %f529, 0f00000000; or.pred %p81, %p80, %p79; mov.f32 %f530, %f529; @%p81 bra $L__BB176_118; ld.const.u64 %rd219, [kernel_params+2328]; cvta.to.global.u64 %rd220, %rd219; ld.local.u32 %r183, [%rd65+60]; add.s32 %r184, %r183, %r29; mul.wide.s32 %rd222, %r184, 8; add.s64 %rd223, %rd220, %rd222; ld.global.v2.f32 {%f529, %f530}, [%rd223]; bra.uni $L__BB176_118; $L__BB176_49: and.b32 %r123, %r32, 120; setp.eq.s32 %p42, %r123, 0; @%p42 bra $L__BB176_82; and.b32 %r124, %r32, 24; setp.eq.s32 %p43, %r124, 0; @%p43 bra $L__BB176_52; ld.const.u64 %rd134, [kernel_params+2232]; cvta.to.global.u64 %rd135, %rd134; shl.b64 %rd136, %rd47, 4; add.s64 %rd137, %rd135, %rd136; ld.global.v4.u32 {%r125, %r126, %r127, %r128}, [%rd137]; ld.const.u64 %rd138, [kernel_params+2344]; cvta.to.global.u64 %rd139, %rd138; add.s32 %r132, %r125, %r29; mul.wide.u32 %rd140, %r132, 16; add.s64 %rd141, %rd139, %rd140; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd141]; add.s32 %r133, %r126, %r29; mul.wide.u32 %rd142, %r133, 16; add.s64 %rd143, %rd139, %rd142; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd143]; add.s32 %r134, %r127, %r29; mul.wide.u32 %rd144, %r134, 16; add.s64 %rd145, %rd139, %rd144; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd145]; bra.uni $L__BB176_81; $L__BB176_24: and.b32 %r89, %r32, 56; setp.eq.s32 %p26, %r89, 0; @%p26 bra $L__BB176_29; and.b32 %r90, %r32, 24; setp.eq.s32 %p27, %r90, 0; @%p27 bra $L__BB176_27; ld.const.u64 %rd90, [kernel_params+2232]; cvta.to.global.u64 %rd91, %rd90; shl.b64 %rd92, %rd18, 4; add.s64 %rd93, %rd91, %rd92; ld.global.v4.u32 {%r91, %r92, %r93, %r94}, [%rd93]; ld.const.u64 %rd94, [kernel_params+2336]; cvta.to.global.u64 %rd95, %rd94; add.s32 %r98, %r91, %r29; mul.wide.u32 %rd96, %r98, 12; add.s64 %rd284, %rd95, %rd96; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r99, %r92, %r29; mul.wide.u32 %rd97, %r99, 12; add.s64 %rd281, %rd95, %rd97; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r100, %r93, %r29; mul.wide.u32 %rd98, %r100, 12; add.s64 %rd278, %rd95, %rd98; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; bra.uni $L__BB176_28; $L__BB176_137: and.b32 %r216, %r32, 131; setp.eq.s32 %p96, %r216, 0; mov.f32 %f531, 0f00000000; @%p96 bra $L__BB176_143; setp.ne.s32 %p97, %r32, 128; @%p97 bra $L__BB176_140; ld.local.u32 %r217, [%rd65+60]; add.s32 %r29, %r217, %r29; $L__BB176_140: ld.const.u64 %rd259, [kernel_params+2320]; cvta.to.global.u64 %rd260, %rd259; mul.wide.s32 %rd261, %r29, 4; add.s64 %rd262, %rd260, %rd261; ld.global.f32 %f531, [%rd262]; bra.uni $L__BB176_143; $L__BB176_132: and.b32 %r205, %r32, 7; setp.eq.s32 %p92, %r205, 0; mov.f32 %f531, 0f00000000; @%p92 bra $L__BB176_143; cvt.u32.u64 %r206, %rd53; setp.eq.s32 %p93, %r32, 4; selp.b32 %r207, %r206, 0, %p93; add.s32 %r208, %r29, %r207; ld.const.u64 %rd245, [kernel_params+2320]; cvta.to.global.u64 %rd246, %rd245; mul.wide.s32 %rd247, %r208, 4; add.s64 %rd248, %rd246, %rd247; ld.global.f32 %f531, [%rd248]; bra.uni $L__BB176_143; $L__BB176_112: and.b32 %r180, %r32, 131; setp.eq.s32 %p77, %r180, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p77 bra $L__BB176_118; setp.ne.s32 %p78, %r32, 128; @%p78 bra $L__BB176_115; ld.local.u32 %r181, [%rd65+60]; add.s32 %r29, %r181, %r29; $L__BB176_115: ld.const.u64 %rd215, [kernel_params+2328]; cvta.to.global.u64 %rd216, %rd215; mul.wide.s32 %rd217, %r29, 8; add.s64 %rd218, %rd216, %rd217; ld.global.v2.f32 {%f529, %f530}, [%rd218]; bra.uni $L__BB176_118; $L__BB176_107: and.b32 %r169, %r32, 7; setp.eq.s32 %p73, %r169, 0; mov.f32 %f529, 0f00000000; mov.f32 %f530, %f529; @%p73 bra $L__BB176_118; cvt.u32.u64 %r170, %rd51; setp.eq.s32 %p74, %r32, 4; selp.b32 %r171, %r170, 0, %p74; add.s32 %r172, %r29, %r171; ld.const.u64 %rd201, [kernel_params+2328]; cvta.to.global.u64 %rd202, %rd201; mul.wide.s32 %rd203, %r172, 8; add.s64 %rd204, %rd202, %rd203; ld.global.v2.f32 {%f529, %f530}, [%rd204]; bra.uni $L__BB176_118; $L__BB176_91: and.b32 %r150, %r41, 8; setp.eq.s32 %p60, %r150, 0; setp.ne.s32 %p61, %r32, 8; mov.f32 %f519, 0f00000000; or.pred %p62, %p61, %p60; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p62 bra $L__BB176_93; ld.const.u64 %rd172, [kernel_params+2344]; cvta.to.global.u64 %rd173, %rd172; ld.local.u32 %r151, [%rd65+60]; add.s32 %r152, %r151, %r29; mul.wide.s32 %rd175, %r152, 16; add.s64 %rd176, %rd173, %rd175; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd176]; bra.uni $L__BB176_93; $L__BB176_38: and.b32 %r118, %r41, 8; setp.eq.s32 %p34, %r118, 0; setp.ne.s32 %p35, %r32, 8; mov.f32 %f496, 0f00000000; or.pred %p36, %p35, %p34; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p36 bra $L__BB176_40; ld.const.u64 %rd123, [kernel_params+2336]; cvta.to.global.u64 %rd124, %rd123; ld.local.u32 %r119, [%rd65+60]; add.s32 %r120, %r119, %r29; mul.wide.s32 %rd126, %r120, 12; add.s64 %rd127, %rd124, %rd126; ld.global.f32 %f498, [%rd127+8]; ld.global.f32 %f497, [%rd127+4]; ld.global.f32 %f496, [%rd127]; bra.uni $L__BB176_40; $L__BB176_130: cvt.u32.u64 %r202, %rd53; mad.lo.s32 %r203, %r202, 3, %r29; ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r203, 4; add.s64 %rd287, %rd241, %rd242; add.s64 %rd288, %rd287, 4; add.s32 %r204, %r203, 2; mul.wide.s32 %rd243, %r204, 4; add.s64 %rd289, %rd241, %rd243; $L__BB176_131: ld.local.f32 %f404, [%rd65+68]; ld.global.f32 %f405, [%rd288]; ld.local.f32 %f406, [%rd65+72]; ld.global.f32 %f407, [%rd289]; mul.ftz.f32 %f408, %f407, %f406; fma.rn.ftz.f32 %f409, %f405, %f404, %f408; mov.f32 %f410, 0f3F800000; sub.ftz.f32 %f411, %f410, %f404; sub.ftz.f32 %f412, %f411, %f406; ld.global.f32 %f413, [%rd287]; fma.rn.ftz.f32 %f531, %f413, %f412, %f409; $L__BB176_143: setp.eq.s32 %p101, %r64, 1; @%p101 bra $L__BB176_147; bra.uni $L__BB176_144; $L__BB176_147: st.local.f32 [%rd16], %f531; bra.uni $L__BB176_157; $L__BB176_144: setp.eq.s32 %p102, %r64, 0; @%p102 bra $L__BB176_146; mov.u32 %r221, 1065353216; st.local.u32 [%rd16], %r221; bra.uni $L__BB176_157; $L__BB176_146: st.local.f32 [%rd16], %f531; st.local.f32 [%rd16+4], %f531; st.local.f32 [%rd16+8], %f531; bra.uni $L__BB176_157; $L__BB176_105: cvt.u32.u64 %r167, %rd51; mad.lo.s32 %r168, %r167, 3, %r29; ld.const.u64 %rd196, [kernel_params+2328]; cvta.to.global.u64 %rd197, %rd196; mul.wide.s32 %rd198, %r168, 8; add.s64 %rd199, %rd197, %rd198; ld.global.v2.f32 {%f527, %f528}, [%rd199]; ld.global.v2.f32 {%f525, %f526}, [%rd199+8]; ld.global.v2.f32 {%f523, %f524}, [%rd199+16]; $L__BB176_106: ld.local.f32 %f370, [%rd65+68]; ld.local.f32 %f371, [%rd65+72]; mul.ftz.f32 %f372, %f523, %f371; mul.ftz.f32 %f373, %f524, %f371; fma.rn.ftz.f32 %f374, %f525, %f370, %f372; fma.rn.ftz.f32 %f375, %f526, %f370, %f373; mov.f32 %f376, 0f3F800000; sub.ftz.f32 %f377, %f376, %f370; sub.ftz.f32 %f378, %f377, %f371; fma.rn.ftz.f32 %f530, %f528, %f378, %f375; fma.rn.ftz.f32 %f529, %f527, %f378, %f374; $L__BB176_118: setp.eq.s32 %p82, %r64, 1; @%p82 bra $L__BB176_122; bra.uni $L__BB176_119; $L__BB176_122: st.local.f32 [%rd16], %f529; bra.uni $L__BB176_157; $L__BB176_119: setp.eq.s32 %p83, %r64, 0; @%p83 bra $L__BB176_121; mov.u32 %r185, 1065353216; st.local.u32 [%rd16], %r185; bra.uni $L__BB176_157; $L__BB176_121: st.local.f32 [%rd16], %f529; st.local.f32 [%rd16+4], %f530; mov.u32 %r186, 0; st.local.u32 [%rd16+8], %r186; bra.uni $L__BB176_157; $L__BB176_87: and.b32 %r148, %r32, 131; setp.eq.s32 %p58, %r148, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p58 bra $L__BB176_93; setp.ne.s32 %p59, %r32, 128; @%p59 bra $L__BB176_90; ld.local.u32 %r149, [%rd65+60]; add.s32 %r29, %r149, %r29; $L__BB176_90: ld.const.u64 %rd168, [kernel_params+2344]; cvta.to.global.u64 %rd169, %rd168; mul.wide.s32 %rd170, %r29, 16; add.s64 %rd171, %rd169, %rd170; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd171]; bra.uni $L__BB176_93; $L__BB176_34: and.b32 %r116, %r32, 131; setp.eq.s32 %p32, %r116, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p32 bra $L__BB176_40; setp.ne.s32 %p33, %r32, 128; @%p33 bra $L__BB176_37; ld.local.u32 %r117, [%rd65+60]; add.s32 %r29, %r117, %r29; $L__BB176_37: ld.const.u64 %rd119, [kernel_params+2336]; cvta.to.global.u64 %rd120, %rd119; mul.wide.s32 %rd121, %r29, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f498, [%rd122+8]; ld.global.f32 %f497, [%rd122+4]; ld.global.f32 %f496, [%rd122]; bra.uni $L__BB176_40; $L__BB176_82: and.b32 %r137, %r32, 7; setp.eq.s32 %p54, %r137, 0; mov.f32 %f519, 0f00000000; mov.f32 %f520, %f519; mov.f32 %f521, %f519; mov.f32 %f522, %f519; @%p54 bra $L__BB176_93; cvt.u32.u64 %r138, %rd47; setp.eq.s32 %p55, %r32, 4; selp.b32 %r139, %r138, 0, %p55; add.s32 %r140, %r29, %r139; ld.const.u64 %rd154, [kernel_params+2344]; cvta.to.global.u64 %rd155, %rd154; mul.wide.s32 %rd156, %r140, 16; add.s64 %rd157, %rd155, %rd156; ld.global.v4.f32 {%f519, %f520, %f521, %f522}, [%rd157]; bra.uni $L__BB176_93; $L__BB176_29: and.b32 %r105, %r32, 7; setp.eq.s32 %p28, %r105, 0; mov.f32 %f496, 0f00000000; mov.f32 %f497, %f496; mov.f32 %f498, %f496; @%p28 bra $L__BB176_40; cvt.u32.u64 %r106, %rd18; setp.eq.s32 %p29, %r32, 4; selp.b32 %r107, %r106, 0, %p29; add.s32 %r108, %r29, %r107; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r108, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f498, [%rd108+8]; ld.global.f32 %f497, [%rd108+4]; ld.global.f32 %f496, [%rd108]; bra.uni $L__BB176_40; $L__BB176_52: cvt.u32.u64 %r135, %rd47; mad.lo.s32 %r136, %r135, 3, %r29; cvt.s64.s32 %rd48, %r136; setp.eq.s32 %p44, %r32, 32; @%p44 bra $L__BB176_80; bra.uni $L__BB176_53; $L__BB176_80: ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; shl.b64 %rd151, %rd48, 4; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f515, %f516, %f517, %f518}, [%rd152]; ld.global.v4.f32 {%f511, %f512, %f513, %f514}, [%rd152+16]; ld.global.v4.f32 {%f507, %f508, %f509, %f510}, [%rd152+32]; bra.uni $L__BB176_81; $L__BB176_27: cvt.u32.u64 %r101, %rd18; mad.lo.s32 %r102, %r101, 3, %r29; ld.const.u64 %rd99, [kernel_params+2336]; cvta.to.global.u64 %rd100, %rd99; mul.wide.s32 %rd101, %r102, 12; add.s64 %rd284, %rd100, %rd101; add.s64 %rd285, %rd284, 4; add.s64 %rd286, %rd284, 8; add.s32 %r103, %r102, 1; mul.wide.s32 %rd102, %r103, 12; add.s64 %rd281, %rd100, %rd102; add.s64 %rd282, %rd281, 4; add.s64 %rd283, %rd281, 8; add.s32 %r104, %r102, 2; mul.wide.s32 %rd103, %r104, 12; add.s64 %rd278, %rd100, %rd103; add.s64 %rd279, %rd278, 4; add.s64 %rd280, %rd278, 8; $L__BB176_28: ld.local.f32 %f179, [%rd65+68]; ld.global.f32 %f180, [%rd281]; ld.global.f32 %f181, [%rd282]; ld.global.f32 %f182, [%rd283]; ld.local.f32 %f183, [%rd65+72]; ld.global.f32 %f184, [%rd278]; mul.ftz.f32 %f185, %f184, %f183; ld.global.f32 %f186, [%rd279]; mul.ftz.f32 %f187, %f186, %f183; ld.global.f32 %f188, [%rd280]; mul.ftz.f32 %f189, %f188, %f183; fma.rn.ftz.f32 %f190, %f180, %f179, %f185; fma.rn.ftz.f32 %f191, %f181, %f179, %f187; fma.rn.ftz.f32 %f192, %f182, %f179, %f189; mov.f32 %f193, 0f3F800000; sub.ftz.f32 %f194, %f193, %f179; sub.ftz.f32 %f195, %f194, %f183; ld.global.f32 %f196, [%rd284]; ld.global.f32 %f197, [%rd285]; ld.global.f32 %f198, [%rd286]; fma.rn.ftz.f32 %f498, %f198, %f195, %f192; fma.rn.ftz.f32 %f497, %f197, %f195, %f191; fma.rn.ftz.f32 %f496, %f196, %f195, %f190; $L__BB176_40: setp.eq.s32 %p37, %r64, 1; @%p37 bra $L__BB176_44; bra.uni $L__BB176_41; $L__BB176_44: add.ftz.f32 %f220, %f496, %f497; add.ftz.f32 %f221, %f498, %f220; mul.ftz.f32 %f222, %f221, 0f3EAAAAAB; st.local.f32 [%rd16], %f222; bra.uni $L__BB176_157; $L__BB176_41: setp.eq.s32 %p38, %r64, 0; @%p38 bra $L__BB176_43; mov.u32 %r121, 1065353216; st.local.u32 [%rd16], %r121; bra.uni $L__BB176_157; $L__BB176_43: st.local.f32 [%rd16], %f496; st.local.f32 [%rd16+4], %f497; st.local.f32 [%rd16+8], %f498; bra.uni $L__BB176_157; $L__BB176_53: ld.const.u64 %rd146, [kernel_params+2352]; cvta.to.global.u64 %rd147, %rd146; shl.b64 %rd148, %rd48, 2; add.s64 %rd49, %rd147, %rd148; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd49]; cvt.rn.f32.u16 %f235, %rs20; mul.ftz.f32 %f38, %f235, 0f3B808081; cvt.rn.f32.u16 %f236, %rs21; mul.ftz.f32 %f39, %f236, 0f3B808081; cvt.rn.f32.u16 %f237, %rs22; mul.ftz.f32 %f40, %f237, 0f3B808081; setp.lt.ftz.f32 %p45, %f38, 0f3D25AEE6; @%p45 bra $L__BB176_55; bra.uni $L__BB176_54; $L__BB176_55: mul.ftz.f32 %f515, %f38, 0f3D9E8391; bra.uni $L__BB176_56; $L__BB176_54: add.ftz.f32 %f238, %f38, 0f3D6147AE; mul.ftz.f32 %f239, %f238, 0f3F72A76F; lg2.approx.ftz.f32 %f240, %f239; mul.ftz.f32 %f241, %f240, 0f4019999A; ex2.approx.ftz.f32 %f515, %f241; $L__BB176_56: setp.lt.ftz.f32 %p46, %f39, 0f3D25AEE6; @%p46 bra $L__BB176_58; bra.uni $L__BB176_57; $L__BB176_58: mul.ftz.f32 %f516, %f39, 0f3D9E8391; bra.uni $L__BB176_59; $L__BB176_57: add.ftz.f32 %f242, %f39, 0f3D6147AE; mul.ftz.f32 %f243, %f242, 0f3F72A76F; lg2.approx.ftz.f32 %f244, %f243; mul.ftz.f32 %f245, %f244, 0f4019999A; ex2.approx.ftz.f32 %f516, %f245; $L__BB176_59: setp.lt.ftz.f32 %p47, %f40, 0f3D25AEE6; @%p47 bra $L__BB176_61; bra.uni $L__BB176_60; $L__BB176_61: mul.ftz.f32 %f517, %f40, 0f3D9E8391; bra.uni $L__BB176_62; $L__BB176_60: add.ftz.f32 %f246, %f40, 0f3D6147AE; mul.ftz.f32 %f247, %f246, 0f3F72A76F; lg2.approx.ftz.f32 %f248, %f247; mul.ftz.f32 %f249, %f248, 0f4019999A; ex2.approx.ftz.f32 %f517, %f249; $L__BB176_62: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd49+4]; cvt.rn.f32.u16 %f250, %rs24; mul.ftz.f32 %f50, %f250, 0f3B808081; cvt.rn.f32.u16 %f251, %rs25; mul.ftz.f32 %f51, %f251, 0f3B808081; cvt.rn.f32.u16 %f252, %rs26; mul.ftz.f32 %f52, %f252, 0f3B808081; setp.lt.ftz.f32 %p48, %f50, 0f3D25AEE6; @%p48 bra $L__BB176_64; bra.uni $L__BB176_63; $L__BB176_64: mul.ftz.f32 %f511, %f50, 0f3D9E8391; bra.uni $L__BB176_65; $L__BB176_63: add.ftz.f32 %f253, %f50, 0f3D6147AE; mul.ftz.f32 %f254, %f253, 0f3F72A76F; lg2.approx.ftz.f32 %f255, %f254; mul.ftz.f32 %f256, %f255, 0f4019999A; ex2.approx.ftz.f32 %f511, %f256; $L__BB176_65: setp.lt.ftz.f32 %p49, %f51, 0f3D25AEE6; @%p49 bra $L__BB176_67; bra.uni $L__BB176_66; $L__BB176_67: mul.ftz.f32 %f512, %f51, 0f3D9E8391; bra.uni $L__BB176_68; $L__BB176_66: add.ftz.f32 %f257, %f51, 0f3D6147AE; mul.ftz.f32 %f258, %f257, 0f3F72A76F; lg2.approx.ftz.f32 %f259, %f258; mul.ftz.f32 %f260, %f259, 0f4019999A; ex2.approx.ftz.f32 %f512, %f260; $L__BB176_68: setp.lt.ftz.f32 %p50, %f52, 0f3D25AEE6; @%p50 bra $L__BB176_70; bra.uni $L__BB176_69; $L__BB176_70: mul.ftz.f32 %f513, %f52, 0f3D9E8391; bra.uni $L__BB176_71; $L__BB176_69: add.ftz.f32 %f261, %f52, 0f3D6147AE; mul.ftz.f32 %f262, %f261, 0f3F72A76F; lg2.approx.ftz.f32 %f263, %f262; mul.ftz.f32 %f264, %f263, 0f4019999A; ex2.approx.ftz.f32 %f513, %f264; $L__BB176_71: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd49+8]; cvt.rn.f32.u16 %f265, %rs28; mul.ftz.f32 %f62, %f265, 0f3B808081; cvt.rn.f32.u16 %f266, %rs29; mul.ftz.f32 %f63, %f266, 0f3B808081; cvt.rn.f32.u16 %f267, %rs30; mul.ftz.f32 %f64, %f267, 0f3B808081; setp.lt.ftz.f32 %p51, %f62, 0f3D25AEE6; @%p51 bra $L__BB176_73; bra.uni $L__BB176_72; $L__BB176_73: mul.ftz.f32 %f507, %f62, 0f3D9E8391; bra.uni $L__BB176_74; $L__BB176_72: add.ftz.f32 %f268, %f62, 0f3D6147AE; mul.ftz.f32 %f269, %f268, 0f3F72A76F; lg2.approx.ftz.f32 %f270, %f269; mul.ftz.f32 %f271, %f270, 0f4019999A; ex2.approx.ftz.f32 %f507, %f271; $L__BB176_74: setp.lt.ftz.f32 %p52, %f63, 0f3D25AEE6; @%p52 bra $L__BB176_76; bra.uni $L__BB176_75; $L__BB176_76: mul.ftz.f32 %f508, %f63, 0f3D9E8391; bra.uni $L__BB176_77; $L__BB176_75: add.ftz.f32 %f272, %f63, 0f3D6147AE; mul.ftz.f32 %f273, %f272, 0f3F72A76F; lg2.approx.ftz.f32 %f274, %f273; mul.ftz.f32 %f275, %f274, 0f4019999A; ex2.approx.ftz.f32 %f508, %f275; $L__BB176_77: cvt.rn.f32.u16 %f276, %rs23; mul.ftz.f32 %f518, %f276, 0f3B808081; cvt.rn.f32.u16 %f277, %rs27; mul.ftz.f32 %f514, %f277, 0f3B808081; cvt.rn.f32.u16 %f278, %rs31; mul.ftz.f32 %f510, %f278, 0f3B808081; setp.lt.ftz.f32 %p53, %f64, 0f3D25AEE6; @%p53 bra $L__BB176_79; bra.uni $L__BB176_78; $L__BB176_79: mul.ftz.f32 %f509, %f64, 0f3D9E8391; bra.uni $L__BB176_81; $L__BB176_78: add.ftz.f32 %f279, %f64, 0f3D6147AE; mul.ftz.f32 %f280, %f279, 0f3F72A76F; lg2.approx.ftz.f32 %f281, %f280; mul.ftz.f32 %f282, %f281, 0f4019999A; ex2.approx.ftz.f32 %f509, %f282; $L__BB176_81: ld.local.f32 %f295, [%rd65+68]; ld.local.f32 %f296, [%rd65+72]; mul.ftz.f32 %f297, %f507, %f296; mul.ftz.f32 %f298, %f508, %f296; mul.ftz.f32 %f299, %f509, %f296; mul.ftz.f32 %f300, %f510, %f296; fma.rn.ftz.f32 %f301, %f511, %f295, %f297; fma.rn.ftz.f32 %f302, %f512, %f295, %f298; fma.rn.ftz.f32 %f303, %f513, %f295, %f299; fma.rn.ftz.f32 %f304, %f514, %f295, %f300; mov.f32 %f305, 0f3F800000; sub.ftz.f32 %f306, %f305, %f295; sub.ftz.f32 %f307, %f306, %f296; fma.rn.ftz.f32 %f522, %f518, %f307, %f304; fma.rn.ftz.f32 %f521, %f517, %f307, %f303; fma.rn.ftz.f32 %f520, %f516, %f307, %f302; fma.rn.ftz.f32 %f519, %f515, %f307, %f301; $L__BB176_93: setp.eq.s32 %p63, %r64, 1; @%p63 bra $L__BB176_97; bra.uni $L__BB176_94; $L__BB176_97: add.ftz.f32 %f355, %f519, %f520; add.ftz.f32 %f356, %f521, %f355; mul.ftz.f32 %f357, %f356, 0f3EAAAAAB; st.local.f32 [%rd16], %f357; bra.uni $L__BB176_157; $L__BB176_94: setp.eq.s32 %p64, %r64, 0; @%p64 bra $L__BB176_96; st.local.f32 [%rd16], %f522; bra.uni $L__BB176_157; $L__BB176_96: st.local.f32 [%rd16], %f519; st.local.f32 [%rd16+4], %f520; st.local.f32 [%rd16+8], %f521; $L__BB176_157: ret; } .func _Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+8]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .align 4 .b8 _Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1[16], .param .b32 _Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2 ) { .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u32 %r1, [_Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2]; cvta.to.local.u64 %rd2, %rd1; add.s32 %r2, %r1, 1; ld.param.u8 %r3, [_Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1+7]; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r4, 0; st.local.u32 [%rd4], %r4; st.local.u32 [%rd4+4], %r4; st.local.u32 [%rd4+8], %r4; st.param.b32 [func_retval0+0], %r2; ret; } .func _Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1[16] ) { .reg .b32 %r<3>; .reg .b64 %rd<5>; ld.param.u64 %rd1, [_Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u32 %r1, [_Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1+12]; cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; mov.u32 %r2, 0; st.local.u32 [%rd4], %r2; st.local.u32 [%rd4+4], %r2; st.local.u32 [%rd4+8], %r2; ret; } .func _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4 ) { .reg .pred %p<5>; .reg .f32 %f<7>; .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.f32 %f6, [_Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd1, [_Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; mov.f32 %f4, 0f00000000; setp.eq.s32 %p1, %r2, 10; @%p1 bra $L__BB180_4; setp.eq.s32 %p2, %r2, 9; @%p2 bra $L__BB180_5; setp.ne.s32 %p3, %r2, 8; mov.f32 %f6, %f4; @%p3 bra $L__BB180_5; and.b32 %r4, %r1, 1; setp.eq.b32 %p4, %r4, 1; selp.f32 %f6, 0f3F800000, 0f00000000, %p4; bra.uni $L__BB180_5; $L__BB180_4: mov.f32 %f6, 0f3F800000; $L__BB180_5: cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f6; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB181_2; bra.uni $L__BB181_1; $L__BB181_2: mov.b32 %f2764, %r395; bra.uni $L__BB181_3; $L__BB181_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB181_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB181_5; bra.uni $L__BB181_4; $L__BB181_5: mov.b32 %f9, %r396; bra.uni $L__BB181_6; $L__BB181_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB181_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB181_8; bra.uni $L__BB181_7; $L__BB181_8: mov.b32 %f2766, %r397; bra.uni $L__BB181_9; $L__BB181_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB181_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB181_11; bra.uni $L__BB181_10; $L__BB181_11: mov.b32 %f15, %r398; bra.uni $L__BB181_12; $L__BB181_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB181_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB181_14; bra.uni $L__BB181_13; $L__BB181_14: mov.b32 %f2768, %r399; bra.uni $L__BB181_15; $L__BB181_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB181_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB181_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB181_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB181_18; bra.uni $L__BB181_385; $L__BB181_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB181_259; bra.uni $L__BB181_19; $L__BB181_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB181_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB181_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB181_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB181_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB181_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB181_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB181_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB181_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB181_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB181_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB181_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB181_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB181_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB181_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB181_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB181_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB181_362; $L__BB181_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB181_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB181_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB181_40; bra.uni $L__BB181_29; $L__BB181_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB181_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB181_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB181_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB181_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB181_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB181_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB181_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB181_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB181_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB181_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB181_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB181_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB181_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB181_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB181_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB181_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB181_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB181_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB181_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB181_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB181_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB181_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB181_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB181_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB181_136; $L__BB181_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB181_371; bra.uni $L__BB181_366; $L__BB181_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB181_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB181_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB181_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB181_383; $L__BB181_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB181_149; bra.uni $L__BB181_139; $L__BB181_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB181_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB181_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB181_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB181_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB181_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB181_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB181_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB181_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB181_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB181_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB181_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB181_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB181_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB181_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB181_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB181_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB181_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB181_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB181_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB181_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB181_250; $L__BB181_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB181_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB181_253; bra.uni $L__BB181_362; $L__BB181_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB181_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB181_304; @%p171 bra $L__BB181_303; bra.uni $L__BB181_256; $L__BB181_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB181_305; $L__BB181_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB181_305; $L__BB181_256: @%p172 bra $L__BB181_302; bra.uni $L__BB181_257; $L__BB181_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB181_305; $L__BB181_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB181_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB181_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB181_308; bra.uni $L__BB181_306; $L__BB181_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB181_309; $L__BB181_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB181_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB181_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB181_316; @%p171 bra $L__BB181_315; bra.uni $L__BB181_311; $L__BB181_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB181_317; $L__BB181_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB181_317; $L__BB181_311: @%p172 bra $L__BB181_314; bra.uni $L__BB181_312; $L__BB181_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB181_317; $L__BB181_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB181_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB181_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB181_320; bra.uni $L__BB181_318; $L__BB181_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB181_321; $L__BB181_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB181_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB181_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB181_328; @%p171 bra $L__BB181_327; bra.uni $L__BB181_323; $L__BB181_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB181_329; $L__BB181_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB181_329; $L__BB181_323: @%p172 bra $L__BB181_326; bra.uni $L__BB181_324; $L__BB181_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB181_329; $L__BB181_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB181_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB181_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB181_332; bra.uni $L__BB181_330; $L__BB181_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB181_333; $L__BB181_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB181_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB181_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB181_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_362; $L__BB181_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB181_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB181_31; bra.uni $L__BB181_136; $L__BB181_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB181_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB181_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB181_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB181_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB181_107; @%p49 bra $L__BB181_106; bra.uni $L__BB181_37; $L__BB181_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB181_108; $L__BB181_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB181_108; $L__BB181_37: @%p50 bra $L__BB181_105; bra.uni $L__BB181_38; $L__BB181_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB181_108; $L__BB181_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB181_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB181_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB181_111; bra.uni $L__BB181_109; $L__BB181_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB181_112; $L__BB181_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB181_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB181_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB181_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB181_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB181_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB181_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_136; $L__BB181_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB181_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB181_368; bra.uni $L__BB181_383; $L__BB181_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB181_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB181_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB181_383; $L__BB181_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB181_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB181_141; bra.uni $L__BB181_250; $L__BB181_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB181_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB181_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB181_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB181_206; @%p101 bra $L__BB181_205; bra.uni $L__BB181_146; $L__BB181_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB181_207; $L__BB181_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB181_207; $L__BB181_146: @%p102 bra $L__BB181_204; bra.uni $L__BB181_147; $L__BB181_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB181_207; $L__BB181_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB181_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB181_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB181_210; bra.uni $L__BB181_208; $L__BB181_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB181_211; $L__BB181_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB181_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB181_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB181_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB181_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB181_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_250; $L__BB181_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB181_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB181_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB181_295; @%p151 bra $L__BB181_294; bra.uni $L__BB181_290; $L__BB181_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB181_296; $L__BB181_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB181_296; $L__BB181_290: @%p152 bra $L__BB181_293; bra.uni $L__BB181_291; $L__BB181_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB181_296; $L__BB181_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB181_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB181_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB181_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB181_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB181_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB181_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB181_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_362; $L__BB181_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB181_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB181_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB181_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB181_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB181_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB181_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB181_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB181_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB181_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB181_279; bra.uni $L__BB181_362; $L__BB181_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB181_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB181_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB181_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB181_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB181_96; @%p39 bra $L__BB181_95; bra.uni $L__BB181_91; $L__BB181_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB181_97; $L__BB181_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB181_97; $L__BB181_91: @%p40 bra $L__BB181_94; bra.uni $L__BB181_92; $L__BB181_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB181_97; $L__BB181_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB181_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB181_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB181_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB181_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB181_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB181_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB181_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB181_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB181_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_136; $L__BB181_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB181_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB181_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB181_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB181_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB181_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB181_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB181_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB181_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB181_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB181_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB181_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB181_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB181_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB181_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB181_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB181_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB181_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB181_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB181_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB181_76; bra.uni $L__BB181_136; $L__BB181_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB181_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB181_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB181_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB181_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB181_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_383; $L__BB181_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB181_383; $L__BB181_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB181_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB181_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB181_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB181_196; @%p92 bra $L__BB181_195; bra.uni $L__BB181_191; $L__BB181_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB181_197; $L__BB181_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB181_197; $L__BB181_191: @%p93 bra $L__BB181_194; bra.uni $L__BB181_192; $L__BB181_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB181_197; $L__BB181_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB181_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB181_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB181_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB181_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB181_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB181_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB181_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB181_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB181_250; $L__BB181_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB181_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB181_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB181_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB181_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB181_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB181_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB181_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB181_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB181_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB181_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB181_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB181_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB181_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB181_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB181_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB181_179; bra.uni $L__BB181_250; $L__BB181_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB181_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB181_337; @%p188 bra $L__BB181_336; bra.uni $L__BB181_23; $L__BB181_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB181_338; $L__BB181_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB181_338; $L__BB181_23: @%p189 bra $L__BB181_335; bra.uni $L__BB181_24; $L__BB181_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB181_338; $L__BB181_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB181_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB181_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB181_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB181_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB181_347; @%p188 bra $L__BB181_346; bra.uni $L__BB181_342; $L__BB181_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB181_348; $L__BB181_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB181_348; $L__BB181_342: @%p189 bra $L__BB181_345; bra.uni $L__BB181_343; $L__BB181_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB181_348; $L__BB181_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB181_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB181_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB181_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB181_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB181_357; @%p188 bra $L__BB181_356; bra.uni $L__BB181_352; $L__BB181_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB181_358; $L__BB181_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB181_358; $L__BB181_352: @%p189 bra $L__BB181_355; bra.uni $L__BB181_353; $L__BB181_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB181_358; $L__BB181_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB181_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB181_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB181_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB181_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB181_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB181_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB181_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB181_364: mov.f32 %f3392, %f3387; bra.uni $L__BB181_385; $L__BB181_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB181_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB181_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB181_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB181_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB181_128; @%p59 bra $L__BB181_127; bra.uni $L__BB181_123; $L__BB181_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB181_129; $L__BB181_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB181_129; $L__BB181_123: @%p60 bra $L__BB181_126; bra.uni $L__BB181_124; $L__BB181_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB181_129; $L__BB181_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB181_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB181_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB181_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB181_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB181_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB181_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB181_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB181_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB181_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB181_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB181_385; $L__BB181_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB181_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB181_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB181_385; $L__BB181_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB181_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB181_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB181_224; @%p120 bra $L__BB181_223; bra.uni $L__BB181_219; $L__BB181_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB181_225; $L__BB181_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB181_225; $L__BB181_219: @%p121 bra $L__BB181_222; bra.uni $L__BB181_220; $L__BB181_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB181_225; $L__BB181_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB181_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB181_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB181_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB181_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB181_234; @%p120 bra $L__BB181_233; bra.uni $L__BB181_229; $L__BB181_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB181_235; $L__BB181_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB181_235; $L__BB181_229: @%p121 bra $L__BB181_232; bra.uni $L__BB181_230; $L__BB181_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB181_235; $L__BB181_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB181_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB181_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB181_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB181_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB181_244; @%p120 bra $L__BB181_243; bra.uni $L__BB181_239; $L__BB181_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB181_245; $L__BB181_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB181_245; $L__BB181_239: @%p121 bra $L__BB181_242; bra.uni $L__BB181_240; $L__BB181_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB181_245; $L__BB181_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB181_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB181_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB181_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB181_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB181_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB181_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB181_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB181_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB181_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB181_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB181_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB181_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB181_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB181_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB181_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB181_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB181_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB181_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB181_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB183_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB183_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB183_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB183_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3 ) { .reg .pred %p<421>; .reg .b16 %rs<22>; .reg .f32 %f<3055>; .reg .b32 %r<630>; .reg .b64 %rd<808>; ld.param.u64 %rd230, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd231, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r134, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r133, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r132, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r131, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd230; cvta.to.local.u64 %rd2, %rd231; and.b32 %r2, %r131, 255; bfe.u32 %r3, %r131, 8, 8; bfe.u32 %r4, %r131, 16, 8; shr.u32 %r5, %r131, 24; setp.eq.s32 %p6, %r5, 255; mov.f32 %f2, 0f3F800000; @%p6 bra $L__BB184_2; mul.wide.u32 %rd232, %r5, 4; add.s64 %rd233, %rd2, %rd232; ld.local.f32 %f2, [%rd233]; $L__BB184_2: ld.const.u64 %rd234, [kernel_params+2400]; cvta.to.global.u64 %rd235, %rd234; mul.wide.s32 %rd236, %r134, 16; add.s64 %rd3, %rd235, %rd236; ld.global.v4.u32 {%r135, %r136, %r137, %r138}, [%rd3]; add.s32 %r629, %r134, 1; setp.eq.ftz.f32 %p7, %f2, 0f00000000; @%p7 bra $L__BB184_498; bra.uni $L__BB184_3; $L__BB184_498: setp.eq.s32 %p419, %r2, 39; add.s32 %r604, %r134, 5; selp.b32 %r629, %r604, %r629, %p419; $L__BB184_499: st.param.b32 [func_retval0+0], %r629; ret; $L__BB184_3: setp.eq.s32 %p8, %r135, 255; add.s64 %rd4, %rd1, 12; @%p8 bra $L__BB184_5; bra.uni $L__BB184_4; $L__BB184_5: ld.local.f32 %f55, [%rd4]; ld.local.f32 %f56, [%rd4+4]; ld.local.f32 %f57, [%rd4+8]; bra.uni $L__BB184_6; $L__BB184_4: mul.wide.u32 %rd237, %r135, 4; add.s64 %rd238, %rd2, %rd237; ld.local.f32 %f55, [%rd238]; ld.local.f32 %f56, [%rd238+4]; ld.local.f32 %f57, [%rd238+8]; $L__BB184_6: ld.local.u32 %r11, [%rd4+52]; and.b32 %r12, %r11, 6; setp.ne.s32 %p9, %r12, 0; @%p9 bra $L__BB184_15; ld.local.v2.f32 {%f546, %f547}, [%rd4+12]; ld.local.v2.f32 {%f548, %f549}, [%rd4+20]; ld.local.v2.f32 {%f550, %f551}, [%rd4+28]; mul.ftz.f32 %f552, %f56, %f550; fma.rn.ftz.f32 %f553, %f55, %f549, %f552; fma.rn.ftz.f32 %f554, %f57, %f551, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f55, %f555; mul.ftz.f32 %f557, %f56, %f555; mul.ftz.f32 %f558, %f57, %f555; sub.ftz.f32 %f559, %f556, %f549; sub.ftz.f32 %f560, %f557, %f550; sub.ftz.f32 %f561, %f558, %f551; mul.ftz.f32 %f562, %f547, %f550; fma.rn.ftz.f32 %f563, %f546, %f549, %f562; fma.rn.ftz.f32 %f18, %f548, %f551, %f563; mul.ftz.f32 %f564, %f18, 0f3F666666; mov.f32 %f565, 0f3C23D70A; min.ftz.f32 %f19, %f564, %f565; mul.ftz.f32 %f566, %f547, %f560; fma.rn.ftz.f32 %f567, %f546, %f559, %f566; fma.rn.ftz.f32 %f568, %f548, %f561, %f567; setp.ltu.ftz.f32 %p10, %f568, %f19; @%p10 bra $L__BB184_8; bra.uni $L__BB184_15; $L__BB184_8: mul.ftz.f32 %f569, %f56, %f547; fma.rn.ftz.f32 %f570, %f55, %f546, %f569; fma.rn.ftz.f32 %f571, %f57, %f548, %f570; mul.ftz.f32 %f572, %f546, %f571; mul.ftz.f32 %f573, %f547, %f571; mul.ftz.f32 %f574, %f548, %f571; sub.ftz.f32 %f575, %f55, %f572; sub.ftz.f32 %f576, %f56, %f573; sub.ftz.f32 %f577, %f57, %f574; mul.ftz.f32 %f578, %f576, %f576; fma.rn.ftz.f32 %f579, %f575, %f575, %f578; fma.rn.ftz.f32 %f580, %f577, %f577, %f579; sqrt.approx.ftz.f32 %f581, %f580; div.approx.ftz.f32 %f23, %f575, %f581; div.approx.ftz.f32 %f24, %f576, %f581; div.approx.ftz.f32 %f25, %f577, %f581; mul.ftz.f32 %f582, %f550, %f24; fma.rn.ftz.f32 %f583, %f549, %f23, %f582; fma.rn.ftz.f32 %f26, %f551, %f25, %f583; mul.ftz.f32 %f584, %f26, %f26; fma.rn.ftz.f32 %f585, %f18, %f18, %f584; mul.ftz.f32 %f586, %f19, %f19; sub.ftz.f32 %f587, %f585, %f586; mul.ftz.f32 %f588, %f584, %f587; mov.f32 %f589, 0f00000000; max.ftz.f32 %f590, %f588, %f589; sqrt.approx.ftz.f32 %f591, %f590; fma.rn.ftz.f32 %f592, %f19, %f18, %f585; mov.f32 %f593, 0f3F000000; div.approx.ftz.f32 %f594, %f593, %f585; add.ftz.f32 %f595, %f591, %f592; mul.ftz.f32 %f27, %f594, %f595; sub.ftz.f32 %f596, %f592, %f591; mul.ftz.f32 %f28, %f594, %f596; setp.gt.ftz.f32 %p11, %f27, 0f3727C5AC; setp.le.ftz.f32 %p12, %f27, 0f3F800054; and.pred %p1, %p11, %p12; setp.gt.ftz.f32 %p13, %f28, 0f3727C5AC; setp.le.ftz.f32 %p14, %f28, 0f3F800054; and.pred %p2, %p13, %p14; and.pred %p15, %p1, %p2; @%p15 bra $L__BB184_11; bra.uni $L__BB184_9; $L__BB184_11: mov.f32 %f603, 0f3F800000; sub.ftz.f32 %f604, %f603, %f27; max.ftz.f32 %f606, %f604, %f589; sqrt.approx.ftz.f32 %f34, %f606; max.ftz.f32 %f607, %f27, %f589; sqrt.approx.ftz.f32 %f35, %f607; sub.ftz.f32 %f608, %f603, %f28; max.ftz.f32 %f609, %f608, %f589; sqrt.approx.ftz.f32 %f36, %f609; max.ftz.f32 %f610, %f28, %f589; sqrt.approx.ftz.f32 %f37, %f610; mul.ftz.f32 %f611, %f18, %f35; fma.rn.ftz.f32 %f612, %f26, %f34, %f611; add.ftz.f32 %f613, %f612, %f612; mul.ftz.f32 %f614, %f35, %f613; sub.ftz.f32 %f38, %f614, %f18; mul.ftz.f32 %f615, %f18, %f37; fma.rn.ftz.f32 %f616, %f26, %f36, %f615; add.ftz.f32 %f617, %f616, %f616; mul.ftz.f32 %f618, %f37, %f617; sub.ftz.f32 %f39, %f618, %f18; setp.ge.ftz.f32 %p17, %f38, 0f3727C5AC; setp.ge.ftz.f32 %p18, %f39, 0f3727C5AC; and.pred %p19, %p17, %p18; @%p19 bra $L__BB184_13; bra.uni $L__BB184_12; $L__BB184_13: setp.lt.ftz.f32 %p21, %f38, %f39; selp.f32 %f2962, %f34, %f36, %p21; selp.f32 %f2963, %f35, %f37, %p21; bra.uni $L__BB184_14; $L__BB184_9: or.pred %p16, %p1, %p2; mov.f32 %f55, %f546; mov.f32 %f56, %f547; mov.f32 %f57, %f548; @%p16 bra $L__BB184_10; bra.uni $L__BB184_15; $L__BB184_10: selp.f32 %f597, %f27, %f28, %p1; mov.f32 %f598, 0f3F800000; sub.ftz.f32 %f599, %f598, %f597; max.ftz.f32 %f601, %f599, %f589; sqrt.approx.ftz.f32 %f2962, %f601; max.ftz.f32 %f602, %f597, %f589; sqrt.approx.ftz.f32 %f2963, %f602; bra.uni $L__BB184_14; $L__BB184_12: setp.gt.ftz.f32 %p20, %f38, %f39; selp.f32 %f2962, %f34, %f36, %p20; selp.f32 %f2963, %f35, %f37, %p20; $L__BB184_14: mul.ftz.f32 %f619, %f546, %f2963; mul.ftz.f32 %f620, %f547, %f2963; mul.ftz.f32 %f621, %f548, %f2963; fma.rn.ftz.f32 %f57, %f25, %f2962, %f621; fma.rn.ftz.f32 %f56, %f24, %f2962, %f620; fma.rn.ftz.f32 %f55, %f23, %f2962, %f619; $L__BB184_15: setp.eq.s32 %p22, %r3, 255; @%p22 bra $L__BB184_17; bra.uni $L__BB184_16; $L__BB184_17: mov.b32 %f60, %r132; bra.uni $L__BB184_18; $L__BB184_16: mul.wide.u32 %rd239, %r3, 4; add.s64 %rd240, %rd2, %rd239; ld.local.f32 %f60, [%rd240]; $L__BB184_18: setp.eq.s32 %p23, %r4, 255; @%p23 bra $L__BB184_20; bra.uni $L__BB184_19; $L__BB184_20: mov.b32 %f63, %r133; bra.uni $L__BB184_21; $L__BB184_19: mul.wide.u32 %rd241, %r4, 4; add.s64 %rd242, %rd2, %rd241; ld.local.f32 %f63, [%rd242]; $L__BB184_21: cvt.u16.u32 %rs8, %r131; and.b16 %rs7, %rs8, 255; mul.wide.u32 %rd243, %r136, 4; add.s64 %rd5, %rd2, %rd243; mul.wide.u32 %rd244, %r137, 4; add.s64 %rd6, %rd2, %rd244; mul.wide.u32 %rd245, %r138, 4; add.s64 %rd7, %rd2, %rd245; setp.gt.s16 %p24, %rs7, 23; @%p24 bra $L__BB184_61; setp.gt.s16 %p35, %rs7, 15; @%p35 bra $L__BB184_47; setp.gt.s16 %p42, %rs7, 8; @%p42 bra $L__BB184_34; bra.uni $L__BB184_24; $L__BB184_34: add.s16 %rs12, %rs7, -9; setp.lt.u16 %p43, %rs12, 2; @%p43 bra $L__BB184_36; add.s16 %rs13, %rs7, -13; setp.lt.u16 %p44, %rs13, 2; @%p44 bra $L__BB184_36; bra.uni $L__BB184_499; $L__BB184_61: setp.gt.s16 %p25, %rs7, 30; @%p25 bra $L__BB184_84; setp.gt.s16 %p29, %rs7, 27; @%p29 bra $L__BB184_73; add.s16 %rs10, %rs7, -25; setp.lt.u16 %p33, %rs10, 2; @%p33 bra $L__BB184_414; setp.eq.s16 %p34, %rs7, 24; @%p34 bra $L__BB184_65; bra.uni $L__BB184_499; $L__BB184_65: ld.local.f32 %f2721, [%rd4+256]; mov.u64 %rd774, 0; mul.ftz.f32 %f2722, %f2, %f2721; ld.local.f32 %f2723, [%rd4+260]; mul.ftz.f32 %f2724, %f2, %f2723; ld.local.f32 %f2725, [%rd4+264]; mul.ftz.f32 %f2726, %f2, %f2725; mov.f32 %f2727, 0f00000000; max.ftz.f32 %f489, %f2722, %f2727; max.ftz.f32 %f490, %f2724, %f2727; max.ftz.f32 %f491, %f2726, %f2727; add.ftz.f32 %f2728, %f489, %f490; add.ftz.f32 %f2729, %f491, %f2728; mul.ftz.f32 %f2730, %f2729, 0f3EAAAAAB; abs.ftz.f32 %f492, %f2730; setp.ltu.ftz.f32 %p356, %f492, 0f3727C5AC; mov.u64 %rd775, %rd774; @%p356 bra $L__BB184_70; ld.local.u32 %r118, [%rd4+252]; mov.u64 %rd774, 0; setp.eq.s32 %p357, %r118, 0; mov.u64 %rd772, %rd774; mov.u64 %rd773, %rd774; @%p357 bra $L__BB184_68; ld.local.u32 %r477, [%rd4+248]; mul.wide.s32 %rd596, %r477, 80; add.s64 %rd597, %rd1, %rd596; add.s64 %rd772, %rd597, 304; add.s64 %rd598, %rd230, %rd596; add.s64 %rd773, %rd598, 304; mov.b32 %r478, %f491; mov.b32 %r479, %f490; mov.b32 %r480, %f489; mov.u32 %r481, 0; st.local.v4.u32 [%rd597+304], {%r480, %r479, %r478, %r481}; add.s32 %r482, %r477, 1; st.local.u32 [%rd4+248], %r482; add.s32 %r483, %r118, -1; st.local.u32 [%rd4+252], %r483; $L__BB184_68: setp.eq.s64 %p358, %rd773, 0; mov.u64 %rd775, %rd774; @%p358 bra $L__BB184_70; st.local.f32 [%rd772+16], %f492; mov.u64 %rd774, %rd772; mov.u64 %rd775, %rd773; $L__BB184_70: setp.eq.s64 %p359, %rd775, 0; @%p359 bra $L__BB184_499; ld.local.u32 %r119, [%rd4+252]; setp.lt.s32 %p360, %r119, 1; @%p360 bra $L__BB184_440; add.s32 %r484, %r119, -1; st.local.u32 [%rd4+252], %r484; ld.local.s32 %rd601, [%rd4+248]; cvt.s64.s32 %rd602, %r484; add.s64 %rd603, %rd601, %rd602; mul.lo.s64 %rd604, %rd603, 80; add.s64 %rd605, %rd1, %rd604; add.s64 %rd776, %rd605, 304; add.s64 %rd606, %rd230, %rd604; add.s64 %rd777, %rd606, 304; bra.uni $L__BB184_441; $L__BB184_47: setp.gt.s16 %p36, %rs7, 19; @%p36 bra $L__BB184_59; bra.uni $L__BB184_48; $L__BB184_59: add.s16 %rs11, %rs7, -21; setp.lt.u16 %p37, %rs11, 3; @%p37 bra $L__BB184_445; bra.uni $L__BB184_60; $L__BB184_445: ld.local.f32 %f2751, [%rd4+256]; mov.u64 %rd780, 0; mul.ftz.f32 %f2752, %f2, %f2751; ld.local.f32 %f2753, [%rd4+260]; mul.ftz.f32 %f2754, %f2, %f2753; ld.local.f32 %f2755, [%rd4+264]; mul.ftz.f32 %f2756, %f2, %f2755; mov.f32 %f2757, 0f00000000; max.ftz.f32 %f496, %f2752, %f2757; max.ftz.f32 %f497, %f2754, %f2757; max.ftz.f32 %f498, %f2756, %f2757; add.ftz.f32 %f2758, %f496, %f497; add.ftz.f32 %f2759, %f498, %f2758; mul.ftz.f32 %f2760, %f2759, 0f3EAAAAAB; abs.ftz.f32 %f499, %f2760; setp.ltu.ftz.f32 %p366, %f499, 0f3727C5AC; mov.u64 %rd781, %rd780; @%p366 bra $L__BB184_450; ld.local.u32 %r120, [%rd4+252]; mov.u64 %rd780, 0; setp.eq.s32 %p367, %r120, 0; mov.u64 %rd778, %rd780; mov.u64 %rd779, %rd780; @%p367 bra $L__BB184_448; ld.local.u32 %r495, [%rd4+248]; mul.wide.s32 %rd616, %r495, 80; add.s64 %rd617, %rd1, %rd616; add.s64 %rd778, %rd617, 304; add.s64 %rd618, %rd230, %rd616; add.s64 %rd779, %rd618, 304; mov.b32 %r496, %f498; mov.b32 %r497, %f497; mov.b32 %r498, %f496; mov.u32 %r499, 0; st.local.v4.u32 [%rd617+304], {%r498, %r497, %r496, %r499}; add.s32 %r500, %r495, 1; st.local.u32 [%rd4+248], %r500; add.s32 %r501, %r120, -1; st.local.u32 [%rd4+252], %r501; $L__BB184_448: setp.eq.s64 %p368, %rd779, 0; mov.u64 %rd781, %rd780; @%p368 bra $L__BB184_450; st.local.f32 [%rd778+16], %f499; mov.u64 %rd780, %rd778; mov.u64 %rd781, %rd779; $L__BB184_450: setp.eq.s64 %p369, %rd781, 0; @%p369 bra $L__BB184_499; add.s64 %rd185, %rd780, 20; mov.u64 %rd621, 0; st.local.f32 [%rd780+20], %f55; st.local.f32 [%rd780+24], %f56; st.local.f32 [%rd780+28], %f57; mov.u32 %r502, 0; st.local.v2.f32 [%rd780+56], {%f2757, %f2757}; st.local.u32 [%rd780+64], %r502; st.local.u64 [%rd780+48], %rd621; mov.f32 %f2762, 0f3727C5AC; max.ftz.f32 %f3052, %f63, %f2762; ld.local.u32 %r503, [%rd4+40]; and.b32 %r504, %r503, 1; setp.eq.b32 %p370, %r504, 1; mov.pred %p371, 0; xor.pred %p372, %p370, %p371; not.pred %p373, %p372; @%p373 bra $L__BB184_453; rcp.approx.ftz.f32 %f3052, %f3052; $L__BB184_453: setp.eq.s32 %p374, %r2, 21; @%p374 bra $L__BB184_457; bra.uni $L__BB184_454; $L__BB184_457: st.local.v2.f32 [%rd185+12], {%f2757, %f2757}; st.local.f32 [%rd185+20], %f3052; mov.u32 %r511, 21; st.local.u32 [%rd185+-8], %r511; ld.local.u32 %r512, [%rd4+40]; or.b32 %r513, %r512, 4; st.local.u32 [%rd4+40], %r513; bra.uni $L__BB184_499; $L__BB184_84: add.s16 %rs9, %rs7, -32; setp.lt.u16 %p26, %rs9, 3; @%p26 bra $L__BB184_282; bra.uni $L__BB184_85; $L__BB184_282: ld.local.v2.u32 {%r320, %r321}, [%rd4+252]; mov.u64 %rd746, 0; mov.b32 %f1888, %r321; mul.ftz.f32 %f336, %f2, %f1888; ld.local.f32 %f1889, [%rd4+260]; mul.ftz.f32 %f337, %f2, %f1889; ld.local.f32 %f1890, [%rd4+264]; mul.ftz.f32 %f338, %f2, %f1890; setp.eq.s32 %p224, %r320, 0; mov.u64 %rd744, %rd746; mov.u64 %rd745, %rd746; @%p224 bra $L__BB184_284; ld.local.u32 %r323, [%rd4+248]; mul.wide.s32 %rd459, %r323, 80; add.s64 %rd460, %rd1, %rd459; add.s64 %rd744, %rd460, 304; add.s64 %rd461, %rd230, %rd459; add.s64 %rd745, %rd461, 304; mov.b32 %r324, %f338; mov.b32 %r325, %f337; mov.b32 %r326, %f336; mov.u32 %r327, 0; st.local.v4.u32 [%rd460+304], {%r326, %r325, %r324, %r327}; add.s32 %r328, %r323, 1; st.local.u32 [%rd4+248], %r328; add.s32 %r329, %r320, -1; st.local.u32 [%rd4+252], %r329; $L__BB184_284: setp.eq.s64 %p225, %rd745, 0; @%p225 bra $L__BB184_286; add.ftz.f32 %f1891, %f336, %f337; add.ftz.f32 %f1892, %f338, %f1891; mul.ftz.f32 %f1893, %f1892, 0f3EAAAAAB; abs.ftz.f32 %f1894, %f1893; st.local.f32 [%rd744+16], %f1894; setp.ge.ftz.f32 %p226, %f1894, 0f3727C5AC; selp.b64 %rd746, %rd745, 0, %p226; $L__BB184_286: setp.eq.s64 %p227, %rd746, 0; @%p227 bra $L__BB184_499; ld.local.f32 %f1896, [%rd6]; mul.ftz.f32 %f1897, %f60, %f1896; ld.local.f32 %f1898, [%rd6+4]; mul.ftz.f32 %f339, %f60, %f1898; ld.local.f32 %f1899, [%rd6+8]; mul.ftz.f32 %f340, %f60, %f1899; add.s64 %rd111, %rd746, 32; st.f32 [%rd746+32], %f1897; st.f32 [%rd746+36], %f339; st.f32 [%rd746+40], %f340; ld.local.f32 %f1900, [%rd4+256]; ld.local.f32 %f1901, [%rd4+260]; ld.local.f32 %f1902, [%rd4+264]; st.f32 [%rd746+44], %f1900; st.f32 [%rd746+48], %f1901; st.f32 [%rd746+52], %f1902; st.f32 [%rd746+20], %f55; st.f32 [%rd746+24], %f56; st.f32 [%rd746+28], %f57; mov.u32 %r331, 2139095039; st.u32 [%rd746+56], %r331; ld.local.f32 %f1903, [%rd7]; mov.f32 %f3024, 0f00000000; max.ftz.f32 %f1904, %f1903, %f3024; mov.f32 %f1905, 0f3F666666; min.ftz.f32 %f1906, %f1904, %f1905; st.f32 [%rd746+60], %f1906; setp.geu.ftz.f32 %p228, %f1897, 0f322BCC77; mov.u32 %r613, 3; mov.f32 %f3022, %f3024; @%p228 bra $L__BB184_289; ld.f32 %f3022, [%rd111+-32]; mov.u32 %r333, 0; st.u32 [%rd111+-32], %r333; st.u32 [%rd111], %r333; mov.u32 %r613, 2; $L__BB184_289: setp.geu.ftz.f32 %p229, %f339, 0f322BCC77; mov.f32 %f3023, %f3024; @%p229 bra $L__BB184_291; ld.f32 %f3023, [%rd111+-28]; mov.u32 %r334, 0; st.u32 [%rd111+-28], %r334; st.u32 [%rd111+4], %r334; add.s32 %r613, %r613, -1; $L__BB184_291: setp.geu.ftz.f32 %p230, %f340, 0f322BCC77; @%p230 bra $L__BB184_293; ld.f32 %f3024, [%rd111+-24]; mov.u32 %r335, 0; st.u32 [%rd111+-24], %r335; st.u32 [%rd111+8], %r335; add.s32 %r613, %r613, -1; $L__BB184_293: setp.gt.s32 %p231, %r613, 2; mov.u32 %r615, 0; @%p231 bra $L__BB184_301; mov.f32 %f1909, 0f00000000; max.ftz.f32 %f347, %f3022, %f1909; max.ftz.f32 %f348, %f3023, %f1909; add.ftz.f32 %f1910, %f347, %f348; max.ftz.f32 %f349, %f3024, %f1909; add.ftz.f32 %f1911, %f349, %f1910; mul.ftz.f32 %f1912, %f1911, 0f3EAAAAAB; abs.ftz.f32 %f350, %f1912; setp.ltu.ftz.f32 %p232, %f350, 0f3727C5AC; mov.u64 %rd749, 0; mov.u64 %rd750, %rd749; @%p232 bra $L__BB184_299; ld.local.u32 %r60, [%rd4+252]; mov.u64 %rd749, 0; setp.eq.s32 %p233, %r60, 0; mov.u64 %rd747, %rd749; mov.u64 %rd748, %rd749; @%p233 bra $L__BB184_297; ld.local.u32 %r337, [%rd4+248]; mul.wide.s32 %rd467, %r337, 80; add.s64 %rd468, %rd1, %rd467; add.s64 %rd747, %rd468, 304; add.s64 %rd469, %rd230, %rd467; add.s64 %rd748, %rd469, 304; mov.b32 %r338, %f349; mov.b32 %r339, %f348; mov.b32 %r340, %f347; mov.u32 %r341, 0; st.local.v4.u32 [%rd468+304], {%r340, %r339, %r338, %r341}; add.s32 %r342, %r337, 1; st.local.u32 [%rd4+248], %r342; add.s32 %r343, %r60, -1; st.local.u32 [%rd4+252], %r343; $L__BB184_297: setp.eq.s64 %p234, %rd748, 0; mov.u64 %rd750, %rd749; @%p234 bra $L__BB184_299; st.local.f32 [%rd747+16], %f350; mov.u64 %rd749, %rd747; mov.u64 %rd750, %rd748; $L__BB184_299: setp.eq.s64 %p235, %rd750, 0; @%p235 bra $L__BB184_301; ld.f32 %f1913, [%rd111+-12]; ld.f32 %f1914, [%rd111+-8]; ld.f32 %f1915, [%rd111+-4]; st.local.f32 [%rd749+20], %f1913; st.local.f32 [%rd749+24], %f1914; st.local.f32 [%rd749+28], %f1915; mov.u32 %r346, 2; st.local.u32 [%rd749+12], %r346; mov.u32 %r615, 12; $L__BB184_301: st.u32 [%rd111+-20], %r2; setp.gt.s32 %p236, %r613, 0; @%p236 bra $L__BB184_303; bra.uni $L__BB184_302; $L__BB184_303: ld.f32 %f1916, [%rd111+-28]; ld.f32 %f1917, [%rd111+-32]; add.ftz.f32 %f1918, %f1917, %f1916; ld.f32 %f1919, [%rd111+-24]; add.ftz.f32 %f1920, %f1919, %f1918; mul.ftz.f32 %f1921, %f1920, 0f3EAAAAAB; abs.ftz.f32 %f1922, %f1921; cvt.rn.f32.s32 %f1923, %r613; mul.ftz.f32 %f1924, %f1922, %f1923; st.f32 [%rd111+-16], %f1924; or.b32 %r348, %r2, 2; setp.eq.s32 %p237, %r348, 34; @%p237 bra $L__BB184_314; bra.uni $L__BB184_304; $L__BB184_314: ld.f32 %f2478, [%rd111]; mul.ftz.f32 %f3028, %f2478, 0f3DA2F983; ld.f32 %f2479, [%rd111+4]; mul.ftz.f32 %f3029, %f2479, 0f3DA2F983; ld.f32 %f2480, [%rd111+8]; mul.ftz.f32 %f3030, %f2480, 0f3DA2F983; bra.uni $L__BB184_315; $L__BB184_24: setp.eq.s16 %p45, %rs7, 2; @%p45 bra $L__BB184_489; setp.eq.s16 %p46, %rs7, 7; @%p46 bra $L__BB184_51; setp.eq.s16 %p47, %rs7, 8; @%p47 bra $L__BB184_27; bra.uni $L__BB184_499; $L__BB184_27: ld.local.f32 %f2917, [%rd4+256]; mov.u64 %rd798, 0; mul.ftz.f32 %f2918, %f2, %f2917; ld.local.f32 %f2919, [%rd4+260]; mul.ftz.f32 %f2920, %f2, %f2919; ld.local.f32 %f2921, [%rd4+264]; mul.ftz.f32 %f2922, %f2, %f2921; mov.f32 %f2923, 0f00000000; max.ftz.f32 %f531, %f2918, %f2923; max.ftz.f32 %f532, %f2920, %f2923; max.ftz.f32 %f533, %f2922, %f2923; add.ftz.f32 %f2924, %f531, %f532; add.ftz.f32 %f2925, %f533, %f2924; mul.ftz.f32 %f2926, %f2925, 0f3EAAAAAB; abs.ftz.f32 %f534, %f2926; setp.ltu.ftz.f32 %p405, %f534, 0f3727C5AC; mov.u64 %rd799, %rd798; @%p405 bra $L__BB184_32; ld.local.u32 %r125, [%rd4+252]; mov.u64 %rd798, 0; setp.eq.s32 %p406, %r125, 0; mov.u64 %rd796, %rd798; mov.u64 %rd797, %rd798; @%p406 bra $L__BB184_30; ld.local.u32 %r568, [%rd4+248]; mul.wide.s32 %rd668, %r568, 80; add.s64 %rd669, %rd1, %rd668; add.s64 %rd796, %rd669, 304; add.s64 %rd670, %rd230, %rd668; add.s64 %rd797, %rd670, 304; mov.b32 %r569, %f533; mov.b32 %r570, %f532; mov.b32 %r571, %f531; mov.u32 %r572, 0; st.local.v4.u32 [%rd669+304], {%r571, %r570, %r569, %r572}; add.s32 %r573, %r568, 1; st.local.u32 [%rd4+248], %r573; add.s32 %r574, %r125, -1; st.local.u32 [%rd4+252], %r574; $L__BB184_30: setp.eq.s64 %p407, %rd797, 0; mov.u64 %rd799, %rd798; @%p407 bra $L__BB184_32; st.local.f32 [%rd796+16], %f534; mov.u64 %rd798, %rd796; mov.u64 %rd799, %rd797; $L__BB184_32: setp.eq.s64 %p408, %rd799, 0; @%p408 bra $L__BB184_499; st.local.f32 [%rd798+20], %f55; st.local.f32 [%rd798+24], %f56; st.local.f32 [%rd798+28], %f57; mov.u32 %r575, 8; st.local.u32 [%rd798+12], %r575; ld.local.u32 %r576, [%rd4+40]; or.b32 %r577, %r576, 12; st.local.u32 [%rd4+40], %r577; bra.uni $L__BB184_499; $L__BB184_73: setp.eq.s16 %p30, %rs7, 28; @%p30 bra $L__BB184_414; setp.eq.s16 %p31, %rs7, 29; @%p31 bra $L__BB184_328; setp.eq.s16 %p32, %rs7, 30; @%p32 bra $L__BB184_76; bra.uni $L__BB184_499; $L__BB184_48: setp.eq.s16 %p39, %rs7, 16; @%p39 bra $L__BB184_36; setp.eq.s16 %p40, %rs7, 17; @%p40 bra $L__BB184_464; setp.eq.s16 %p41, %rs7, 19; @%p41 bra $L__BB184_51; bra.uni $L__BB184_499; $L__BB184_51: ld.local.f32 %f2927, [%rd4+256]; mov.u64 %rd802, 0; mul.ftz.f32 %f2928, %f2, %f2927; ld.local.f32 %f2929, [%rd4+260]; mul.ftz.f32 %f2930, %f2, %f2929; ld.local.f32 %f2931, [%rd4+264]; mul.ftz.f32 %f2932, %f2, %f2931; mov.f32 %f2933, 0f00000000; max.ftz.f32 %f535, %f2928, %f2933; max.ftz.f32 %f536, %f2930, %f2933; max.ftz.f32 %f537, %f2932, %f2933; add.ftz.f32 %f2934, %f535, %f536; add.ftz.f32 %f2935, %f537, %f2934; mul.ftz.f32 %f2936, %f2935, 0f3EAAAAAB; abs.ftz.f32 %f538, %f2936; setp.ltu.ftz.f32 %p409, %f538, 0f3727C5AC; mov.u64 %rd803, %rd802; @%p409 bra $L__BB184_56; ld.local.u32 %r126, [%rd4+252]; mov.u64 %rd802, 0; setp.eq.s32 %p410, %r126, 0; mov.u64 %rd800, %rd802; mov.u64 %rd801, %rd802; @%p410 bra $L__BB184_54; ld.local.u32 %r578, [%rd4+248]; mul.wide.s32 %rd677, %r578, 80; add.s64 %rd678, %rd1, %rd677; add.s64 %rd800, %rd678, 304; add.s64 %rd679, %rd230, %rd677; add.s64 %rd801, %rd679, 304; mov.b32 %r579, %f537; mov.b32 %r580, %f536; mov.b32 %r581, %f535; mov.u32 %r582, 0; st.local.v4.u32 [%rd678+304], {%r581, %r580, %r579, %r582}; add.s32 %r583, %r578, 1; st.local.u32 [%rd4+248], %r583; add.s32 %r584, %r126, -1; st.local.u32 [%rd4+252], %r584; $L__BB184_54: setp.eq.s64 %p411, %rd801, 0; mov.u64 %rd803, %rd802; @%p411 bra $L__BB184_56; st.local.f32 [%rd800+16], %f538; mov.u64 %rd802, %rd800; mov.u64 %rd803, %rd801; $L__BB184_56: setp.eq.s64 %p412, %rd803, 0; @%p412 bra $L__BB184_499; add.s64 %rd222, %rd802, 20; st.local.f32 [%rd802+20], %f55; st.local.f32 [%rd802+24], %f56; st.local.f32 [%rd802+28], %f57; st.local.f32 [%rd802+32], %f60; st.local.f32 [%rd802+36], %f63; cvt.ftz.sat.f32.f32 %f539, %f60; cvt.ftz.sat.f32.f32 %f540, %f63; setp.eq.s32 %p413, %r2, 7; @%p413 bra $L__BB184_488; bra.uni $L__BB184_58; $L__BB184_488: mov.u32 %r588, 7; st.local.u32 [%rd222+-8], %r588; st.local.f32 [%rd222+12], %f539; st.local.f32 [%rd222+16], %f540; ld.local.u32 %r589, [%rd4+40]; or.b32 %r590, %r589, 12; st.local.u32 [%rd4+40], %r590; bra.uni $L__BB184_499; $L__BB184_36: ld.local.f32 %f2826, [%rd4+256]; mov.u64 %rd792, 0; mul.ftz.f32 %f2827, %f2, %f2826; ld.local.f32 %f2828, [%rd4+260]; mul.ftz.f32 %f2829, %f2, %f2828; ld.local.f32 %f2830, [%rd4+264]; mul.ftz.f32 %f2831, %f2, %f2830; mov.f32 %f2832, 0f00000000; max.ftz.f32 %f515, %f2827, %f2832; max.ftz.f32 %f516, %f2829, %f2832; max.ftz.f32 %f517, %f2831, %f2832; add.ftz.f32 %f2833, %f515, %f516; add.ftz.f32 %f2834, %f517, %f2833; mul.ftz.f32 %f2835, %f2834, 0f3EAAAAAB; abs.ftz.f32 %f518, %f2835; setp.ltu.ftz.f32 %p387, %f518, 0f3727C5AC; mov.u64 %rd793, %rd792; @%p387 bra $L__BB184_41; ld.local.u32 %r123, [%rd4+252]; mov.u64 %rd792, 0; setp.eq.s32 %p388, %r123, 0; mov.u64 %rd790, %rd792; mov.u64 %rd791, %rd792; @%p388 bra $L__BB184_39; ld.local.u32 %r538, [%rd4+248]; mul.wide.s32 %rd645, %r538, 80; add.s64 %rd646, %rd1, %rd645; add.s64 %rd790, %rd646, 304; add.s64 %rd647, %rd230, %rd645; add.s64 %rd791, %rd647, 304; mov.b32 %r539, %f517; mov.b32 %r540, %f516; mov.b32 %r541, %f515; mov.u32 %r542, 0; st.local.v4.u32 [%rd646+304], {%r541, %r540, %r539, %r542}; add.s32 %r543, %r538, 1; st.local.u32 [%rd4+248], %r543; add.s32 %r544, %r123, -1; st.local.u32 [%rd4+252], %r544; $L__BB184_39: setp.eq.s64 %p389, %rd791, 0; mov.u64 %rd793, %rd792; @%p389 bra $L__BB184_41; st.local.f32 [%rd790+16], %f518; mov.u64 %rd792, %rd790; mov.u64 %rd793, %rd791; $L__BB184_41: setp.eq.s64 %p390, %rd793, 0; @%p390 bra $L__BB184_499; mul.ftz.f32 %f3053, %f60, %f60; add.s64 %rd205, %rd792, 20; mov.u64 %rd794, 0; st.local.f32 [%rd792+20], %f55; st.local.f32 [%rd792+24], %f56; st.local.f32 [%rd792+28], %f57; mov.u32 %r545, 1065353216; st.local.u32 [%rd792+40], %r545; st.local.u64 [%rd792+48], %rd794; setp.eq.s32 %p391, %r136, 255; @%p391 bra $L__BB184_472; bra.uni $L__BB184_43; $L__BB184_472: mov.u32 %r546, 0; st.local.v2.f32 [%rd205+36], {%f2832, %f2832}; st.local.u32 [%rd205+44], %r546; st.local.v2.f32 [%rd205+12], {%f3053, %f3053}; mov.f32 %f3054, %f3053; bra.uni $L__BB184_473; $L__BB184_414: ld.local.f32 %f2677, [%rd4+256]; mul.ftz.f32 %f466, %f2, %f2677; ld.local.f32 %f2678, [%rd4+260]; mul.ftz.f32 %f467, %f2, %f2678; ld.local.f32 %f2679, [%rd4+264]; mul.ftz.f32 %f468, %f2, %f2679; mov.f32 %f2680, 0f3727C5AC; max.ftz.f32 %f3049, %f63, %f2680; ld.local.u32 %r441, [%rd4+40]; and.b32 %r442, %r441, 1; setp.eq.b32 %p339, %r442, 1; mov.pred %p340, 0; xor.pred %p341, %p339, %p340; not.pred %p342, %p341; @%p342 bra $L__BB184_416; rcp.approx.ftz.f32 %f3049, %f3049; $L__BB184_416: ld.local.f32 %f2682, [%rd4+24]; ld.local.f32 %f2683, [%rd4+28]; mul.ftz.f32 %f2684, %f56, %f2683; fma.rn.ftz.f32 %f2685, %f55, %f2682, %f2684; ld.local.f32 %f2686, [%rd4+32]; fma.rn.ftz.f32 %f2687, %f57, %f2686, %f2685; abs.ftz.f32 %f472, %f2687; fma.rn.ftz.f32 %f2688, %f3049, %f3049, 0fBF800000; mov.f32 %f3050, 0f3F800000; fma.rn.ftz.f32 %f473, %f472, %f472, %f2688; setp.leu.ftz.f32 %p343, %f473, 0f00000000; @%p343 bra $L__BB184_418; sqrt.approx.ftz.f32 %f2689, %f473; sub.ftz.f32 %f2690, %f2689, %f472; add.ftz.f32 %f2691, %f472, %f2689; div.approx.ftz.f32 %f2692, %f2690, %f2691; fma.rn.ftz.f32 %f2693, %f472, %f2691, 0fBF800000; fma.rn.ftz.f32 %f2694, %f472, %f2690, 0f3F800000; div.approx.ftz.f32 %f2695, %f2693, %f2694; mul.ftz.f32 %f2696, %f2692, 0f3F000000; mul.ftz.f32 %f2697, %f2692, %f2696; fma.rn.ftz.f32 %f2698, %f2695, %f2695, 0f3F800000; mul.ftz.f32 %f3050, %f2697, %f2698; $L__BB184_418: mul.ftz.f32 %f476, %f60, %f60; mul.ftz.f32 %f2699, %f466, %f3050; mov.f32 %f2700, 0f00000000; max.ftz.f32 %f477, %f2699, %f2700; mul.ftz.f32 %f2701, %f467, %f3050; max.ftz.f32 %f478, %f2701, %f2700; mul.ftz.f32 %f2702, %f468, %f3050; max.ftz.f32 %f479, %f2702, %f2700; add.ftz.f32 %f2703, %f477, %f478; add.ftz.f32 %f2704, %f479, %f2703; mul.ftz.f32 %f2705, %f2704, 0f3EAAAAAB; abs.ftz.f32 %f480, %f2705; setp.ltu.ftz.f32 %p344, %f480, 0f3727C5AC; mov.u64 %rd766, 0; mov.u64 %rd767, %rd766; @%p344 bra $L__BB184_423; ld.local.u32 %r116, [%rd4+252]; mov.u64 %rd766, 0; setp.eq.s32 %p345, %r116, 0; mov.u64 %rd764, %rd766; mov.u64 %rd765, %rd766; @%p345 bra $L__BB184_421; ld.local.u32 %r443, [%rd4+248]; mul.wide.s32 %rd574, %r443, 80; add.s64 %rd575, %rd1, %rd574; add.s64 %rd764, %rd575, 304; add.s64 %rd576, %rd230, %rd574; add.s64 %rd765, %rd576, 304; mov.b32 %r444, %f479; mov.b32 %r445, %f478; mov.b32 %r446, %f477; mov.u32 %r447, 0; st.local.v4.u32 [%rd575+304], {%r446, %r445, %r444, %r447}; add.s32 %r448, %r443, 1; st.local.u32 [%rd4+248], %r448; add.s32 %r449, %r116, -1; st.local.u32 [%rd4+252], %r449; $L__BB184_421: setp.eq.s64 %p346, %rd765, 0; mov.u64 %rd767, %rd766; @%p346 bra $L__BB184_423; st.local.f32 [%rd764+16], %f480; mov.u64 %rd766, %rd764; mov.u64 %rd767, %rd765; $L__BB184_423: setp.eq.s64 %p347, %rd767, 0; @%p347 bra $L__BB184_429; add.s64 %rd160, %rd766, 20; mov.u64 %rd579, 0; st.local.f32 [%rd766+20], %f55; st.local.f32 [%rd766+24], %f56; st.local.f32 [%rd766+28], %f57; mov.u32 %r450, 0; st.local.v2.f32 [%rd766+56], {%f2700, %f2700}; st.local.u32 [%rd766+64], %r450; st.local.u64 [%rd766+48], %rd579; setp.eq.s32 %p348, %r2, 28; @%p348 bra $L__BB184_428; bra.uni $L__BB184_425; $L__BB184_428: st.local.v2.f32 [%rd160+12], {%f2700, %f2700}; st.local.f32 [%rd160+20], %f3049; mov.u32 %r457, 9; st.local.u32 [%rd160+-8], %r457; ld.local.u32 %r458, [%rd4+40]; or.b32 %r459, %r458, 4; st.local.u32 [%rd4+40], %r459; bra.uni $L__BB184_429; $L__BB184_85: setp.eq.s16 %p27, %rs7, 31; @%p27 bra $L__BB184_317; setp.ne.s16 %p28, %rs7, 39; @%p28 bra $L__BB184_499; ld.global.v4.u32 {%r139, %r140, %r141, %r142}, [%rd3+16]; ld.local.f32 %f2972, [%rd5]; ld.local.f32 %f2973, [%rd5+4]; ld.local.f32 %f2974, [%rd5+8]; shr.u32 %r144, %r139, 24; shl.b32 %r145, %r137, 2; cvt.u64.u32 %rd246, %r145; and.b64 %rd247, %rd246, 1020; add.s64 %rd248, %rd2, %rd247; ld.local.f32 %f70, [%rd248]; shr.u32 %r146, %r137, 6; cvt.u64.u32 %rd249, %r146; and.b64 %rd250, %rd249, 1020; add.s64 %rd251, %rd2, %rd250; ld.local.f32 %f71, [%rd251]; shr.u32 %r147, %r137, 14; cvt.u64.u32 %rd252, %r147; and.b64 %rd253, %rd252, 1020; add.s64 %rd254, %rd2, %rd253; ld.local.f32 %f72, [%rd254]; shr.u32 %r148, %r137, 24; mul.wide.u32 %rd255, %r148, 4; add.s64 %rd256, %rd2, %rd255; ld.local.f32 %f73, [%rd256]; shl.b32 %r149, %r138, 2; cvt.u64.u32 %rd257, %r149; and.b64 %rd258, %rd257, 1020; add.s64 %rd259, %rd2, %rd258; ld.local.f32 %f74, [%rd259]; shr.u32 %r150, %r138, 6; cvt.u64.u32 %rd260, %r150; and.b64 %rd261, %rd260, 1020; add.s64 %rd262, %rd2, %rd261; ld.local.f32 %f75, [%rd262]; shr.u32 %r151, %r138, 14; cvt.u64.u32 %rd263, %r151; and.b64 %rd264, %rd263, 1020; add.s64 %rd265, %rd2, %rd264; ld.local.f32 %f76, [%rd265]; shr.u32 %r152, %r138, 24; mul.wide.u32 %rd266, %r152, 4; add.s64 %rd267, %rd2, %rd266; ld.local.f32 %f77, [%rd267]; shr.u32 %r153, %r139, 6; cvt.u64.u32 %rd268, %r153; and.b64 %rd269, %rd268, 1020; add.s64 %rd270, %rd2, %rd269; ld.local.f32 %f78, [%rd270]; shr.u32 %r154, %r139, 14; cvt.u64.u32 %rd271, %r154; and.b64 %rd272, %rd271, 1020; add.s64 %rd273, %rd2, %rd272; mul.wide.u32 %rd274, %r144, 4; add.s64 %rd275, %rd2, %rd274; ld.local.f32 %f79, [%rd275]; shl.b32 %r155, %r139, 2; cvt.u64.u32 %rd276, %r155; and.b64 %rd277, %rd276, 1020; add.s64 %rd278, %rd2, %rd277; ld.local.f32 %f622, [%rd278]; mov.f32 %f623, 0f3727C5AC; max.ftz.f32 %f2975, %f622, %f623; ld.local.f32 %f81, [%rd273]; setp.eq.ftz.f32 %p48, %f81, 0f00000000; @%p48 bra $L__BB184_89; mul.ftz.f32 %f624, %f81, 0f40C90FDB; cos.approx.ftz.f32 %f625, %f624; sin.approx.ftz.f32 %f626, %f624; mov.f32 %f627, 0f3F800000; sub.ftz.f32 %f628, %f627, %f625; mul.ftz.f32 %f629, %f55, %f628; fma.rn.ftz.f32 %f630, %f55, %f629, %f625; mul.ftz.f32 %f631, %f56, %f629; mul.ftz.f32 %f632, %f57, %f626; sub.ftz.f32 %f633, %f631, %f632; mul.ftz.f32 %f634, %f2973, %f633; fma.rn.ftz.f32 %f635, %f2972, %f630, %f634; mul.ftz.f32 %f636, %f57, %f629; mul.ftz.f32 %f637, %f56, %f626; add.ftz.f32 %f638, %f637, %f636; add.ftz.f32 %f639, %f632, %f631; mul.ftz.f32 %f640, %f56, %f628; fma.rn.ftz.f32 %f641, %f56, %f640, %f625; mul.ftz.f32 %f642, %f2973, %f641; fma.rn.ftz.f32 %f643, %f2972, %f639, %f642; mul.ftz.f32 %f644, %f57, %f640; mul.ftz.f32 %f645, %f55, %f626; sub.ftz.f32 %f646, %f644, %f645; sub.ftz.f32 %f647, %f636, %f637; add.ftz.f32 %f648, %f645, %f644; mul.ftz.f32 %f649, %f2973, %f648; fma.rn.ftz.f32 %f650, %f2972, %f647, %f649; mul.ftz.f32 %f651, %f57, %f628; fma.rn.ftz.f32 %f652, %f57, %f651, %f625; fma.rn.ftz.f32 %f2973, %f2974, %f646, %f643; fma.rn.ftz.f32 %f84, %f2974, %f652, %f650; fma.rn.ftz.f32 %f2972, %f2974, %f638, %f635; mov.f32 %f2974, %f84; $L__BB184_89: ld.local.u32 %r156, [%rd4+40]; and.b32 %r157, %r156, 1; setp.eq.b32 %p49, %r157, 1; mov.pred %p50, 0; xor.pred %p51, %p49, %p50; not.pred %p52, %p51; @%p52 bra $L__BB184_91; rcp.approx.ftz.f32 %f2975, %f2975; $L__BB184_91: ld.local.f32 %f90, [%rd4+24]; ld.local.f32 %f91, [%rd4+28]; mul.ftz.f32 %f654, %f56, %f91; fma.rn.ftz.f32 %f655, %f55, %f90, %f654; ld.local.f32 %f92, [%rd4+32]; fma.rn.ftz.f32 %f656, %f57, %f92, %f655; abs.ftz.f32 %f93, %f656; fma.rn.ftz.f32 %f657, %f2975, %f2975, 0fBF800000; mov.f32 %f653, 0f3F800000; fma.rn.ftz.f32 %f94, %f93, %f93, %f657; setp.leu.ftz.f32 %p53, %f94, 0f00000000; mov.f32 %f2976, %f653; @%p53 bra $L__BB184_93; sqrt.approx.ftz.f32 %f658, %f94; sub.ftz.f32 %f659, %f658, %f93; add.ftz.f32 %f660, %f93, %f658; div.approx.ftz.f32 %f661, %f659, %f660; fma.rn.ftz.f32 %f662, %f93, %f660, 0fBF800000; fma.rn.ftz.f32 %f663, %f93, %f659, 0f3F800000; div.approx.ftz.f32 %f664, %f662, %f663; mul.ftz.f32 %f665, %f661, 0f3F000000; mul.ftz.f32 %f666, %f661, %f665; fma.rn.ftz.f32 %f667, %f664, %f664, 0f3F800000; mul.ftz.f32 %f2976, %f666, %f667; $L__BB184_93: cvt.ftz.sat.f32.f32 %f668, %f60; sub.ftz.f32 %f670, %f653, %f668; cvt.ftz.sat.f32.f32 %f671, %f78; sub.ftz.f32 %f672, %f653, %f671; mul.ftz.f32 %f97, %f670, %f672; mul.ftz.f32 %f98, %f671, %f670; sub.ftz.f32 %f99, %f653, %f98; ld.global.v4.u32 {%r158, %r159, %r160, %r161}, [%rd3+32]; setp.eq.s32 %p54, %r158, 255; @%p54 bra $L__BB184_95; bra.uni $L__BB184_94; $L__BB184_95: mov.b32 %f2977, %r159; mov.b32 %f2978, %r160; mov.b32 %f2979, %r161; bra.uni $L__BB184_96; $L__BB184_43: ld.local.f32 %f520, [%rd5]; ld.local.f32 %f521, [%rd5+8]; ld.local.f32 %f522, [%rd5+4]; st.local.v2.f32 [%rd205+36], {%f520, %f522}; st.local.f32 [%rd205+44], %f521; ld.local.f32 %f523, [%rd6]; setp.eq.ftz.f32 %p392, %f523, 0f00000000; @%p392 bra $L__BB184_45; mul.ftz.f32 %f2836, %f523, 0f40C90FDB; cos.approx.ftz.f32 %f2837, %f2836; sin.approx.ftz.f32 %f2838, %f2836; mov.f32 %f2839, 0f3F800000; sub.ftz.f32 %f2840, %f2839, %f2837; mul.ftz.f32 %f2841, %f55, %f2840; fma.rn.ftz.f32 %f2842, %f55, %f2841, %f2837; mul.ftz.f32 %f2843, %f56, %f2841; mul.ftz.f32 %f2844, %f57, %f2838; sub.ftz.f32 %f2845, %f2843, %f2844; mul.ftz.f32 %f2846, %f522, %f2845; fma.rn.ftz.f32 %f2847, %f520, %f2842, %f2846; mul.ftz.f32 %f2848, %f57, %f2841; mul.ftz.f32 %f2849, %f56, %f2838; add.ftz.f32 %f2850, %f2849, %f2848; add.ftz.f32 %f2851, %f2844, %f2843; mul.ftz.f32 %f2852, %f56, %f2840; fma.rn.ftz.f32 %f2853, %f56, %f2852, %f2837; mul.ftz.f32 %f2854, %f522, %f2853; fma.rn.ftz.f32 %f2855, %f520, %f2851, %f2854; mul.ftz.f32 %f2856, %f57, %f2852; mul.ftz.f32 %f2857, %f55, %f2838; sub.ftz.f32 %f2858, %f2856, %f2857; sub.ftz.f32 %f2859, %f2848, %f2849; add.ftz.f32 %f2860, %f2857, %f2856; mul.ftz.f32 %f2861, %f522, %f2860; fma.rn.ftz.f32 %f2862, %f520, %f2859, %f2861; mul.ftz.f32 %f2863, %f57, %f2840; fma.rn.ftz.f32 %f2864, %f57, %f2863, %f2837; fma.rn.ftz.f32 %f2865, %f521, %f2864, %f2862; fma.rn.ftz.f32 %f2866, %f521, %f2858, %f2855; fma.rn.ftz.f32 %f2867, %f521, %f2850, %f2847; st.local.v2.f32 [%rd205+36], {%f2867, %f2866}; st.local.f32 [%rd205+44], %f2865; $L__BB184_45: mov.f32 %f2868, 0fBF7D70A4; max.ftz.f32 %f2869, %f63, %f2868; mov.f32 %f2870, 0f3F7D70A4; min.ftz.f32 %f524, %f2869, %f2870; setp.lt.ftz.f32 %p393, %f524, 0f00000000; @%p393 bra $L__BB184_471; bra.uni $L__BB184_46; $L__BB184_471: add.ftz.f32 %f2873, %f524, 0f3F800000; div.approx.ftz.f32 %f3054, %f3053, %f2873; mul.ftz.f32 %f3053, %f3053, %f2873; st.local.v2.f32 [%rd205+12], {%f3054, %f3053}; bra.uni $L__BB184_473; $L__BB184_60: setp.eq.s16 %p38, %rs7, 20; @%p38 bra $L__BB184_76; bra.uni $L__BB184_499; $L__BB184_76: ld.local.f32 %f2765, [%rd4+256]; mov.u64 %rd784, 0; mul.ftz.f32 %f2766, %f2, %f2765; ld.local.f32 %f2767, [%rd4+260]; mul.ftz.f32 %f2768, %f2, %f2767; ld.local.f32 %f2769, [%rd4+264]; mul.ftz.f32 %f2770, %f2, %f2769; mov.f32 %f2771, 0f00000000; max.ftz.f32 %f505, %f2766, %f2771; max.ftz.f32 %f506, %f2768, %f2771; max.ftz.f32 %f507, %f2770, %f2771; add.ftz.f32 %f2772, %f505, %f506; add.ftz.f32 %f2773, %f507, %f2772; mul.ftz.f32 %f2774, %f2773, 0f3EAAAAAB; abs.ftz.f32 %f508, %f2774; setp.ltu.ftz.f32 %p376, %f508, 0f3727C5AC; mov.u64 %rd785, %rd784; @%p376 bra $L__BB184_81; ld.local.u32 %r121, [%rd4+252]; mov.u64 %rd784, 0; setp.eq.s32 %p377, %r121, 0; mov.u64 %rd782, %rd784; mov.u64 %rd783, %rd784; @%p377 bra $L__BB184_79; ld.local.u32 %r514, [%rd4+248]; mul.wide.s32 %rd627, %r514, 80; add.s64 %rd628, %rd1, %rd627; add.s64 %rd782, %rd628, 304; add.s64 %rd629, %rd230, %rd627; add.s64 %rd783, %rd629, 304; mov.b32 %r515, %f507; mov.b32 %r516, %f506; mov.b32 %r517, %f505; mov.u32 %r518, 0; st.local.v4.u32 [%rd628+304], {%r517, %r516, %r515, %r518}; add.s32 %r519, %r514, 1; st.local.u32 [%rd4+248], %r519; add.s32 %r520, %r121, -1; st.local.u32 [%rd4+252], %r520; $L__BB184_79: setp.eq.s64 %p378, %rd783, 0; mov.u64 %rd785, %rd784; @%p378 bra $L__BB184_81; st.local.f32 [%rd782+16], %f508; mov.u64 %rd784, %rd782; mov.u64 %rd785, %rd783; $L__BB184_81: setp.eq.s64 %p379, %rd785, 0; @%p379 bra $L__BB184_499; add.s64 %rd192, %rd784, 20; st.local.f32 [%rd784+20], %f55; st.local.f32 [%rd784+24], %f56; st.local.f32 [%rd784+28], %f57; st.local.f32 [%rd784+44], %f60; st.local.f32 [%rd784+48], %f63; ld.local.f32 %f2775, [%rd6]; neg.ftz.f32 %f2776, %f2775; st.local.f32 [%rd784+52], %f2776; setp.eq.s32 %p380, %r136, 255; @%p380 bra $L__BB184_458; bra.uni $L__BB184_83; $L__BB184_458: ld.local.u8 %rs18, [%rd4+52]; and.b16 %rs19, %rs18, 6; setp.eq.s16 %p381, %rs19, 0; @%p381 bra $L__BB184_460; ld.local.f32 %f2787, [%rd4+104]; ld.local.f32 %f2788, [%rd4+108]; mul.ftz.f32 %f2789, %f2788, %f2788; fma.rn.ftz.f32 %f2790, %f2787, %f2787, %f2789; ld.local.f32 %f2791, [%rd4+112]; fma.rn.ftz.f32 %f2792, %f2791, %f2791, %f2790; sqrt.approx.ftz.f32 %f2793, %f2792; div.approx.ftz.f32 %f2794, %f2787, %f2793; div.approx.ftz.f32 %f2795, %f2788, %f2793; div.approx.ftz.f32 %f2796, %f2791, %f2793; st.local.f32 [%rd192+12], %f2794; st.local.f32 [%rd192+16], %f2795; st.local.f32 [%rd192+20], %f2796; bra.uni $L__BB184_461; $L__BB184_489: ld.local.f32 %f2937, [%rd4+256]; mov.u64 %rd806, 0; mul.ftz.f32 %f2938, %f2, %f2937; ld.local.f32 %f2939, [%rd4+260]; mul.ftz.f32 %f2940, %f2, %f2939; ld.local.f32 %f2941, [%rd4+264]; mul.ftz.f32 %f2942, %f2, %f2941; mov.f32 %f2943, 0f00000000; max.ftz.f32 %f541, %f2938, %f2943; max.ftz.f32 %f542, %f2940, %f2943; max.ftz.f32 %f543, %f2942, %f2943; add.ftz.f32 %f2944, %f541, %f542; add.ftz.f32 %f2945, %f543, %f2944; mul.ftz.f32 %f2946, %f2945, 0f3EAAAAAB; abs.ftz.f32 %f544, %f2946; setp.ltu.ftz.f32 %p414, %f544, 0f3727C5AC; mov.u64 %rd807, %rd806; @%p414 bra $L__BB184_494; ld.local.u32 %r127, [%rd4+252]; mov.u64 %rd806, 0; setp.eq.s32 %p415, %r127, 0; mov.u64 %rd804, %rd806; mov.u64 %rd805, %rd806; @%p415 bra $L__BB184_492; ld.local.u32 %r591, [%rd4+248]; mul.wide.s32 %rd686, %r591, 80; add.s64 %rd687, %rd1, %rd686; add.s64 %rd804, %rd687, 304; add.s64 %rd688, %rd230, %rd686; add.s64 %rd805, %rd688, 304; mov.b32 %r592, %f543; mov.b32 %r593, %f542; mov.b32 %r594, %f541; mov.u32 %r595, 0; st.local.v4.u32 [%rd687+304], {%r594, %r593, %r592, %r595}; add.s32 %r596, %r591, 1; st.local.u32 [%rd4+248], %r596; add.s32 %r597, %r127, -1; st.local.u32 [%rd4+252], %r597; $L__BB184_492: setp.eq.s64 %p416, %rd805, 0; mov.u64 %rd807, %rd806; @%p416 bra $L__BB184_494; st.local.f32 [%rd804+16], %f544; mov.u64 %rd806, %rd804; mov.u64 %rd807, %rd805; $L__BB184_494: setp.eq.s64 %p417, %rd807, 0; @%p417 bra $L__BB184_499; add.s64 %rd229, %rd806, 20; st.local.f32 [%rd806+20], %f55; st.local.f32 [%rd806+24], %f56; st.local.f32 [%rd806+28], %f57; setp.eq.ftz.f32 %p418, %f60, 0f00000000; @%p418 bra $L__BB184_497; bra.uni $L__BB184_496; $L__BB184_497: mov.u32 %r601, 2; st.local.u32 [%rd229+-8], %r601; ld.local.u32 %r602, [%rd4+40]; or.b32 %r603, %r602, 12; st.local.u32 [%rd4+40], %r603; bra.uni $L__BB184_499; $L__BB184_328: ld.global.v4.u32 {%r362, %r363, %r364, %r365}, [%rd3+16]; ld.global.v4.u32 {%r366, %r367, %r368, %r369}, [%rd3+32]; ld.global.u32 %rd121, [%rd3+52]; ld.local.f32 %f374, [%rd4+256]; ld.local.f32 %f375, [%rd4+260]; ld.local.f32 %f376, [%rd4+264]; bfe.u32 %r77, %r136, 8, 8; and.b32 %r78, %r136, 255; setp.eq.s32 %p285, %r78, 255; @%p285 bra $L__BB184_330; bra.uni $L__BB184_329; $L__BB184_330: mov.b32 %f3031, %r137; bra.uni $L__BB184_331; $L__BB184_464: ld.local.f32 %f2811, [%rd4+256]; mov.u64 %rd788, 0; mul.ftz.f32 %f2812, %f2, %f2811; ld.local.f32 %f2813, [%rd4+260]; mul.ftz.f32 %f2814, %f2, %f2813; ld.local.f32 %f2815, [%rd4+264]; mul.ftz.f32 %f2816, %f2, %f2815; mov.f32 %f2817, 0f00000000; max.ftz.f32 %f511, %f2812, %f2817; max.ftz.f32 %f512, %f2814, %f2817; max.ftz.f32 %f513, %f2816, %f2817; add.ftz.f32 %f2818, %f511, %f512; add.ftz.f32 %f2819, %f513, %f2818; mul.ftz.f32 %f2820, %f2819, 0f3EAAAAAB; abs.ftz.f32 %f514, %f2820; setp.ltu.ftz.f32 %p383, %f514, 0f3727C5AC; mov.u64 %rd789, %rd788; @%p383 bra $L__BB184_469; ld.local.u32 %r122, [%rd4+252]; mov.u64 %rd788, 0; setp.eq.s32 %p384, %r122, 0; mov.u64 %rd786, %rd788; mov.u64 %rd787, %rd788; @%p384 bra $L__BB184_467; ld.local.u32 %r528, [%rd4+248]; mul.wide.s32 %rd636, %r528, 80; add.s64 %rd637, %rd1, %rd636; add.s64 %rd786, %rd637, 304; add.s64 %rd638, %rd230, %rd636; add.s64 %rd787, %rd638, 304; mov.b32 %r529, %f513; mov.b32 %r530, %f512; mov.b32 %r531, %f511; mov.u32 %r532, 0; st.local.v4.u32 [%rd637+304], {%r531, %r530, %r529, %r532}; add.s32 %r533, %r528, 1; st.local.u32 [%rd4+248], %r533; add.s32 %r534, %r122, -1; st.local.u32 [%rd4+252], %r534; $L__BB184_467: setp.eq.s64 %p385, %rd787, 0; mov.u64 %rd789, %rd788; @%p385 bra $L__BB184_469; st.local.f32 [%rd786+16], %f514; mov.u64 %rd788, %rd786; mov.u64 %rd789, %rd787; $L__BB184_469: setp.eq.s64 %p386, %rd789, 0; @%p386 bra $L__BB184_499; st.local.f32 [%rd788+20], %f55; st.local.f32 [%rd788+24], %f56; st.local.f32 [%rd788+28], %f57; cvt.ftz.sat.f32.f32 %f2821, %f60; st.local.f32 [%rd788+32], %f2821; mov.f32 %f2822, 0f3C23D70A; max.ftz.f32 %f2823, %f2821, %f2822; mul.ftz.f32 %f2824, %f2823, %f2823; rcp.approx.ftz.f32 %f2825, %f2824; st.local.f32 [%rd788+36], %f2825; mov.u32 %r535, 17; st.local.u32 [%rd788+12], %r535; ld.local.u32 %r536, [%rd4+40]; or.b32 %r537, %r536, 12; st.local.u32 [%rd4+40], %r537; bra.uni $L__BB184_499; $L__BB184_317: ld.local.f32 %f2481, [%rd4+256]; mul.ftz.f32 %f370, %f2, %f2481; ld.local.f32 %f2482, [%rd4+260]; mul.ftz.f32 %f371, %f2, %f2482; ld.local.f32 %f2483, [%rd4+264]; mul.ftz.f32 %f372, %f2, %f2483; add.ftz.f32 %f2484, %f370, %f371; add.ftz.f32 %f2485, %f372, %f2484; mul.ftz.f32 %f2486, %f2485, 0f3EAAAAAB; abs.ftz.f32 %f373, %f2486; setp.ltu.ftz.f32 %p277, %f373, 0f3727C5AC; @%p277 bra $L__BB184_499; ld.local.u32 %r64, [%rd4+40]; and.b32 %r351, %r64, 512; setp.eq.s32 %p278, %r351, 0; @%p278 bra $L__BB184_324; bra.uni $L__BB184_319; $L__BB184_324: or.b32 %r354, %r64, 516; mov.u64 %rd751, 0; st.local.u32 [%rd4+40], %r354; st.local.f32 [%rd4+280], %f370; st.local.f32 [%rd4+284], %f371; st.local.f32 [%rd4+288], %f372; ld.local.u32 %r68, [%rd4+252]; setp.eq.s32 %p283, %r68, 0; mov.pred %p420, -1; @%p283 bra $L__BB184_326; ld.local.u32 %r355, [%rd4+248]; mul.wide.s32 %rd475, %r355, 80; add.s64 %rd476, %rd1, %rd475; add.s64 %rd751, %rd476, 304; mov.b32 %r356, %f372; mov.b32 %r357, %f371; mov.b32 %r358, %f370; mov.u32 %r359, 31; st.local.v4.u32 [%rd476+304], {%r358, %r357, %r356, %r359}; add.s32 %r360, %r355, 1; st.local.u32 [%rd4+248], %r360; add.s32 %r361, %r68, -1; st.local.u32 [%rd4+252], %r361; mov.pred %p420, 0; $L__BB184_326: @%p420 bra $L__BB184_499; st.local.f32 [%rd751+16], %f373; ld.local.f32 %f2501, [%rd4]; ld.local.f32 %f2502, [%rd4+4]; ld.local.f32 %f2503, [%rd4+8]; st.local.f32 [%rd751+20], %f2501; st.local.f32 [%rd751+24], %f2502; st.local.f32 [%rd751+28], %f2503; bra.uni $L__BB184_499; $L__BB184_454: mul.ftz.f32 %f2763, %f60, %f60; st.local.v2.f32 [%rd185+12], {%f2763, %f2763}; st.local.f32 [%rd185+20], %f3052; cvt.ftz.sat.f32.f32 %f503, %f2763; setp.eq.s32 %p375, %r2, 22; @%p375 bra $L__BB184_456; bra.uni $L__BB184_455; $L__BB184_456: st.local.v2.f32 [%rd185+12], {%f503, %f503}; mov.u32 %r508, 22; st.local.u32 [%rd185+-8], %r508; ld.local.u32 %r509, [%rd4+40]; or.b32 %r510, %r509, 12; st.local.u32 [%rd4+40], %r510; bra.uni $L__BB184_499; $L__BB184_425: setp.eq.s32 %p349, %r2, 25; st.local.v2.f32 [%rd160+12], {%f476, %f476}; st.local.f32 [%rd160+20], %f3049; cvt.ftz.sat.f32.f32 %f481, %f476; @%p349 bra $L__BB184_427; bra.uni $L__BB184_426; $L__BB184_427: st.local.v2.f32 [%rd160+12], {%f481, %f481}; mov.u32 %r454, 13; st.local.u32 [%rd160+-8], %r454; ld.local.u32 %r455, [%rd4+40]; or.b32 %r456, %r455, 12; st.local.u32 [%rd4+40], %r456; bra.uni $L__BB184_429; $L__BB184_46: mov.f32 %f2871, 0f3F800000; sub.ftz.f32 %f2872, %f2871, %f524; mul.ftz.f32 %f3054, %f3053, %f2872; div.approx.ftz.f32 %f3053, %f3053, %f2872; st.local.v2.f32 [%rd205+12], {%f3054, %f3053}; $L__BB184_473: setp.eq.s32 %p394, %r2, 9; @%p394 bra $L__BB184_487; bra.uni $L__BB184_474; $L__BB184_487: mov.u32 %r565, 9; st.local.u32 [%rd205+-8], %r565; ld.local.u32 %r566, [%rd4+40]; or.b32 %r567, %r566, 4; st.local.u32 [%rd4+40], %r567; bra.uni $L__BB184_499; $L__BB184_474: setp.eq.s32 %p395, %r2, 13; @%p395 bra $L__BB184_486; bra.uni $L__BB184_475; $L__BB184_486: cvt.ftz.sat.f32.f32 %f2915, %f3053; cvt.ftz.sat.f32.f32 %f2916, %f3054; st.local.v2.f32 [%rd205+12], {%f2916, %f2915}; mov.u32 %r562, 13; st.local.u32 [%rd205+-8], %r562; ld.local.u32 %r563, [%rd4+40]; or.b32 %r564, %r563, 12; st.local.u32 [%rd4+40], %r564; bra.uni $L__BB184_499; $L__BB184_475: setp.eq.s32 %p396, %r2, 10; @%p396 bra $L__BB184_485; bra.uni $L__BB184_476; $L__BB184_485: mov.u64 %rd663, 0; st.local.u64 [%rd205+28], %rd663; cvt.ftz.sat.f32.f32 %f2913, %f3053; cvt.ftz.sat.f32.f32 %f2914, %f3054; st.local.v2.f32 [%rd205+12], {%f2914, %f2913}; mov.u32 %r559, 10; st.local.u32 [%rd205+-8], %r559; ld.local.u32 %r560, [%rd4+40]; or.b32 %r561, %r560, 12; st.local.u32 [%rd4+40], %r561; bra.uni $L__BB184_499; $L__BB184_302: mov.u32 %r347, 0; st.u32 [%rd111+-16], %r347; bra.uni $L__BB184_316; $L__BB184_58: mov.u32 %r585, 19; st.local.u32 [%rd222+-8], %r585; st.local.f32 [%rd222+12], %f539; st.local.f32 [%rd222+16], %f540; ld.local.u32 %r586, [%rd4+40]; or.b32 %r587, %r586, 12; st.local.u32 [%rd4+40], %r587; bra.uni $L__BB184_499; $L__BB184_83: ld.local.f32 %f2777, [%rd5]; ld.local.f32 %f2778, [%rd5+4]; mul.ftz.f32 %f2779, %f2778, %f2778; fma.rn.ftz.f32 %f2780, %f2777, %f2777, %f2779; ld.local.f32 %f2781, [%rd5+8]; fma.rn.ftz.f32 %f2782, %f2781, %f2781, %f2780; sqrt.approx.ftz.f32 %f2783, %f2782; div.approx.ftz.f32 %f2784, %f2777, %f2783; div.approx.ftz.f32 %f2785, %f2778, %f2783; div.approx.ftz.f32 %f2786, %f2781, %f2783; st.local.f32 [%rd192+12], %f2784; st.local.f32 [%rd192+16], %f2785; st.local.f32 [%rd192+20], %f2786; bra.uni $L__BB184_461; $L__BB184_304: mov.f32 %f1926, 0f3F8147AE; max.ftz.f32 %f1927, %f63, %f1926; mov.f32 %f1928, 0f40733333; min.ftz.f32 %f1929, %f1927, %f1928; rcp.approx.ftz.f32 %f1930, %f1929; mov.f32 %f1931, 0f3F800000; fma.rn.ftz.f32 %f1932, %f1930, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f1933, %f1930, %f1932, 0f3F2B020C; fma.rn.ftz.f32 %f1934, %f1929, 0f3D8240B8, %f1933; add.ftz.f32 %f1935, %f1934, 0f3F800000; mul.ftz.f32 %f1936, %f1935, 0f3FAAAAAB; sub.ftz.f32 %f1937, %f1931, %f1934; div.approx.ftz.f32 %f351, %f1936, %f1937; ld.f32 %f352, [%rd111+12]; setp.lt.ftz.f32 %p238, %f352, 0f38D1B717; mov.f32 %f3027, 0f00000000; mov.f32 %f3025, %f3027; @%p238 bra $L__BB184_307; setp.ge.ftz.f32 %p239, %f352, 0f3F7EB852; mov.f32 %f3025, 0f3F7FFFEF; @%p239 bra $L__BB184_307; sub.ftz.f32 %f1940, %f1931, 0f3F000000; mul.ftz.f32 %f1941, %f1940, 0f40400000; sqrt.approx.ftz.f32 %f1942, %f1941; mul.ftz.f32 %f1943, %f351, %f1942; mul.ftz.f32 %f1944, %f1943, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1945, 0f3F800000; mul.ftz.f32 %f1947, %f1946, 0f3E800000; mul.ftz.f32 %f1948, %f1942, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f1949, %f1947; setp.lt.ftz.f32 %p240, %f1950, %f352; selp.f32 %f1951, 0f3F800000, 0f3F000000, %p240; selp.f32 %f1952, 0f3F000000, 0f00000000, %p240; selp.f32 %f1953, 0f3F400000, 0f3E800000, %p240; sub.ftz.f32 %f1954, %f1931, %f1953; mul.ftz.f32 %f1955, %f1954, 0f40400000; sqrt.approx.ftz.f32 %f1956, %f1955; mul.ftz.f32 %f1957, %f1953, 0f3F000000; mul.ftz.f32 %f1958, %f351, %f1956; mul.ftz.f32 %f1959, %f1958, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1960, %f1959; add.ftz.f32 %f1961, %f1960, 0f3F800000; mul.ftz.f32 %f1962, %f1961, %f1957; mul.ftz.f32 %f1963, %f1956, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1964, %f1963; mul.ftz.f32 %f1965, %f1964, %f1962; setp.lt.ftz.f32 %p241, %f1965, %f352; selp.f32 %f1966, %f1951, %f1953, %p241; selp.f32 %f1967, %f1953, %f1952, %p241; add.ftz.f32 %f1968, %f1966, %f1967; mul.ftz.f32 %f1969, %f1968, 0f3F000000; sub.ftz.f32 %f1970, %f1931, %f1969; mul.ftz.f32 %f1971, %f1970, 0f40400000; sqrt.approx.ftz.f32 %f1972, %f1971; mul.ftz.f32 %f1973, %f1969, 0f3F000000; mul.ftz.f32 %f1974, %f351, %f1972; mul.ftz.f32 %f1975, %f1974, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1976, %f1975; add.ftz.f32 %f1977, %f1976, 0f3F800000; mul.ftz.f32 %f1978, %f1977, %f1973; mul.ftz.f32 %f1979, %f1972, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1980, %f1979; mul.ftz.f32 %f1981, %f1980, %f1978; setp.lt.ftz.f32 %p242, %f1981, %f352; selp.f32 %f1982, %f1966, %f1969, %p242; selp.f32 %f1983, %f1969, %f1967, %p242; add.ftz.f32 %f1984, %f1982, %f1983; mul.ftz.f32 %f1985, %f1984, 0f3F000000; sub.ftz.f32 %f1986, %f1931, %f1985; mul.ftz.f32 %f1987, %f1986, 0f40400000; sqrt.approx.ftz.f32 %f1988, %f1987; mul.ftz.f32 %f1989, %f1985, 0f3F000000; mul.ftz.f32 %f1990, %f351, %f1988; mul.ftz.f32 %f1991, %f1990, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1992, 0f3F800000; mul.ftz.f32 %f1994, %f1993, %f1989; mul.ftz.f32 %f1995, %f1988, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f1996, %f1994; setp.lt.ftz.f32 %p243, %f1997, %f352; selp.f32 %f1998, %f1982, %f1985, %p243; selp.f32 %f1999, %f1985, %f1983, %p243; add.ftz.f32 %f2000, %f1998, %f1999; mul.ftz.f32 %f2001, %f2000, 0f3F000000; sub.ftz.f32 %f2002, %f1931, %f2001; mul.ftz.f32 %f2003, %f2002, 0f40400000; sqrt.approx.ftz.f32 %f2004, %f2003; mul.ftz.f32 %f2005, %f2001, 0f3F000000; mul.ftz.f32 %f2006, %f351, %f2004; mul.ftz.f32 %f2007, %f2006, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2008, %f2007; add.ftz.f32 %f2009, %f2008, 0f3F800000; mul.ftz.f32 %f2010, %f2009, %f2005; mul.ftz.f32 %f2011, %f2004, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2012, %f2011; mul.ftz.f32 %f2013, %f2012, %f2010; setp.lt.ftz.f32 %p244, %f2013, %f352; selp.f32 %f2014, %f1998, %f2001, %p244; selp.f32 %f2015, %f2001, %f1999, %p244; add.ftz.f32 %f2016, %f2014, %f2015; mul.ftz.f32 %f2017, %f2016, 0f3F000000; sub.ftz.f32 %f2018, %f1931, %f2017; mul.ftz.f32 %f2019, %f2018, 0f40400000; sqrt.approx.ftz.f32 %f2020, %f2019; mul.ftz.f32 %f2021, %f2017, 0f3F000000; mul.ftz.f32 %f2022, %f351, %f2020; mul.ftz.f32 %f2023, %f2022, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2024, %f2023; add.ftz.f32 %f2025, %f2024, 0f3F800000; mul.ftz.f32 %f2026, %f2025, %f2021; mul.ftz.f32 %f2027, %f2020, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2028, %f2027; mul.ftz.f32 %f2029, %f2028, %f2026; setp.lt.ftz.f32 %p245, %f2029, %f352; selp.f32 %f2030, %f2014, %f2017, %p245; selp.f32 %f2031, %f2017, %f2015, %p245; add.ftz.f32 %f2032, %f2030, %f2031; mul.ftz.f32 %f2033, %f2032, 0f3F000000; sub.ftz.f32 %f2034, %f1931, %f2033; mul.ftz.f32 %f2035, %f2034, 0f40400000; sqrt.approx.ftz.f32 %f2036, %f2035; mul.ftz.f32 %f2037, %f2033, 0f3F000000; mul.ftz.f32 %f2038, %f351, %f2036; mul.ftz.f32 %f2039, %f2038, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2040, %f2039; add.ftz.f32 %f2041, %f2040, 0f3F800000; mul.ftz.f32 %f2042, %f2041, %f2037; mul.ftz.f32 %f2043, %f2036, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2044, %f2043; mul.ftz.f32 %f2045, %f2044, %f2042; setp.lt.ftz.f32 %p246, %f2045, %f352; selp.f32 %f2046, %f2030, %f2033, %p246; selp.f32 %f2047, %f2033, %f2031, %p246; add.ftz.f32 %f2048, %f2046, %f2047; mul.ftz.f32 %f2049, %f2048, 0f3F000000; sub.ftz.f32 %f2050, %f1931, %f2049; mul.ftz.f32 %f2051, %f2050, 0f40400000; sqrt.approx.ftz.f32 %f2052, %f2051; mul.ftz.f32 %f2053, %f2049, 0f3F000000; mul.ftz.f32 %f2054, %f351, %f2052; mul.ftz.f32 %f2055, %f2054, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2056, %f2055; add.ftz.f32 %f2057, %f2056, 0f3F800000; mul.ftz.f32 %f2058, %f2057, %f2053; mul.ftz.f32 %f2059, %f2052, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2060, %f2059; mul.ftz.f32 %f2061, %f2060, %f2058; setp.lt.ftz.f32 %p247, %f2061, %f352; selp.f32 %f2062, %f2046, %f2049, %p247; selp.f32 %f2063, %f2049, %f2047, %p247; add.ftz.f32 %f2064, %f2062, %f2063; mul.ftz.f32 %f2065, %f2064, 0f3F000000; sub.ftz.f32 %f2066, %f1931, %f2065; mul.ftz.f32 %f2067, %f2066, 0f40400000; sqrt.approx.ftz.f32 %f2068, %f2067; mul.ftz.f32 %f2069, %f2065, 0f3F000000; mul.ftz.f32 %f2070, %f351, %f2068; mul.ftz.f32 %f2071, %f2070, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2072, %f2071; add.ftz.f32 %f2073, %f2072, 0f3F800000; mul.ftz.f32 %f2074, %f2073, %f2069; mul.ftz.f32 %f2075, %f2068, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2076, %f2075; mul.ftz.f32 %f2077, %f2076, %f2074; setp.lt.ftz.f32 %p248, %f2077, %f352; selp.f32 %f2078, %f2062, %f2065, %p248; selp.f32 %f2079, %f2065, %f2063, %p248; add.ftz.f32 %f2080, %f2078, %f2079; mul.ftz.f32 %f2081, %f2080, 0f3F000000; sub.ftz.f32 %f2082, %f1931, %f2081; mul.ftz.f32 %f2083, %f2082, 0f40400000; sqrt.approx.ftz.f32 %f2084, %f2083; mul.ftz.f32 %f2085, %f2081, 0f3F000000; mul.ftz.f32 %f2086, %f351, %f2084; mul.ftz.f32 %f2087, %f2086, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2088, %f2087; add.ftz.f32 %f2089, %f2088, 0f3F800000; mul.ftz.f32 %f2090, %f2089, %f2085; mul.ftz.f32 %f2091, %f2084, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2092, %f2091; mul.ftz.f32 %f2093, %f2092, %f2090; setp.lt.ftz.f32 %p249, %f2093, %f352; selp.f32 %f2094, %f2078, %f2081, %p249; selp.f32 %f2095, %f2081, %f2079, %p249; add.ftz.f32 %f2096, %f2094, %f2095; mul.ftz.f32 %f2097, %f2096, 0f3F000000; sub.ftz.f32 %f2098, %f1931, %f2097; mul.ftz.f32 %f2099, %f2098, 0f40400000; sqrt.approx.ftz.f32 %f2100, %f2099; mul.ftz.f32 %f2101, %f2097, 0f3F000000; mul.ftz.f32 %f2102, %f351, %f2100; mul.ftz.f32 %f2103, %f2102, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2104, %f2103; add.ftz.f32 %f2105, %f2104, 0f3F800000; mul.ftz.f32 %f2106, %f2105, %f2101; mul.ftz.f32 %f2107, %f2100, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2108, %f2107; mul.ftz.f32 %f2109, %f2108, %f2106; setp.lt.ftz.f32 %p250, %f2109, %f352; selp.f32 %f2110, %f2094, %f2097, %p250; selp.f32 %f2111, %f2097, %f2095, %p250; add.ftz.f32 %f2112, %f2110, %f2111; mul.ftz.f32 %f3025, %f2112, 0f3F000000; $L__BB184_307: ld.f32 %f355, [%rd111+16]; setp.lt.ftz.f32 %p251, %f355, 0f38D1B717; mov.f32 %f3026, %f3027; @%p251 bra $L__BB184_310; setp.ge.ftz.f32 %p252, %f355, 0f3F7EB852; mov.f32 %f3026, 0f3F7FFFEF; @%p252 bra $L__BB184_310; sub.ftz.f32 %f2116, %f1931, 0f3F000000; mul.ftz.f32 %f2117, %f2116, 0f40400000; sqrt.approx.ftz.f32 %f2118, %f2117; mul.ftz.f32 %f2119, %f351, %f2118; mul.ftz.f32 %f2120, %f2119, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2121, %f2120; add.ftz.f32 %f2122, %f2121, 0f3F800000; mul.ftz.f32 %f2123, %f2122, 0f3E800000; mul.ftz.f32 %f2124, %f2118, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2125, %f2124; mul.ftz.f32 %f2126, %f2125, %f2123; setp.lt.ftz.f32 %p253, %f2126, %f355; selp.f32 %f2127, 0f3F800000, 0f3F000000, %p253; selp.f32 %f2128, 0f3F000000, 0f00000000, %p253; selp.f32 %f2129, 0f3F400000, 0f3E800000, %p253; sub.ftz.f32 %f2130, %f1931, %f2129; mul.ftz.f32 %f2131, %f2130, 0f40400000; sqrt.approx.ftz.f32 %f2132, %f2131; mul.ftz.f32 %f2133, %f2129, 0f3F000000; mul.ftz.f32 %f2134, %f351, %f2132; mul.ftz.f32 %f2135, %f2134, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2136, %f2135; add.ftz.f32 %f2137, %f2136, 0f3F800000; mul.ftz.f32 %f2138, %f2137, %f2133; mul.ftz.f32 %f2139, %f2132, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2140, %f2139; mul.ftz.f32 %f2141, %f2140, %f2138; setp.lt.ftz.f32 %p254, %f2141, %f355; selp.f32 %f2142, %f2127, %f2129, %p254; selp.f32 %f2143, %f2129, %f2128, %p254; add.ftz.f32 %f2144, %f2142, %f2143; mul.ftz.f32 %f2145, %f2144, 0f3F000000; sub.ftz.f32 %f2146, %f1931, %f2145; mul.ftz.f32 %f2147, %f2146, 0f40400000; sqrt.approx.ftz.f32 %f2148, %f2147; mul.ftz.f32 %f2149, %f2145, 0f3F000000; mul.ftz.f32 %f2150, %f351, %f2148; mul.ftz.f32 %f2151, %f2150, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2152, %f2151; add.ftz.f32 %f2153, %f2152, 0f3F800000; mul.ftz.f32 %f2154, %f2153, %f2149; mul.ftz.f32 %f2155, %f2148, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2156, %f2155; mul.ftz.f32 %f2157, %f2156, %f2154; setp.lt.ftz.f32 %p255, %f2157, %f355; selp.f32 %f2158, %f2142, %f2145, %p255; selp.f32 %f2159, %f2145, %f2143, %p255; add.ftz.f32 %f2160, %f2158, %f2159; mul.ftz.f32 %f2161, %f2160, 0f3F000000; sub.ftz.f32 %f2162, %f1931, %f2161; mul.ftz.f32 %f2163, %f2162, 0f40400000; sqrt.approx.ftz.f32 %f2164, %f2163; mul.ftz.f32 %f2165, %f2161, 0f3F000000; mul.ftz.f32 %f2166, %f351, %f2164; mul.ftz.f32 %f2167, %f2166, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2168, %f2167; add.ftz.f32 %f2169, %f2168, 0f3F800000; mul.ftz.f32 %f2170, %f2169, %f2165; mul.ftz.f32 %f2171, %f2164, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2172, %f2171; mul.ftz.f32 %f2173, %f2172, %f2170; setp.lt.ftz.f32 %p256, %f2173, %f355; selp.f32 %f2174, %f2158, %f2161, %p256; selp.f32 %f2175, %f2161, %f2159, %p256; add.ftz.f32 %f2176, %f2174, %f2175; mul.ftz.f32 %f2177, %f2176, 0f3F000000; sub.ftz.f32 %f2178, %f1931, %f2177; mul.ftz.f32 %f2179, %f2178, 0f40400000; sqrt.approx.ftz.f32 %f2180, %f2179; mul.ftz.f32 %f2181, %f2177, 0f3F000000; mul.ftz.f32 %f2182, %f351, %f2180; mul.ftz.f32 %f2183, %f2182, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2184, %f2183; add.ftz.f32 %f2185, %f2184, 0f3F800000; mul.ftz.f32 %f2186, %f2185, %f2181; mul.ftz.f32 %f2187, %f2180, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2188, %f2187; mul.ftz.f32 %f2189, %f2188, %f2186; setp.lt.ftz.f32 %p257, %f2189, %f355; selp.f32 %f2190, %f2174, %f2177, %p257; selp.f32 %f2191, %f2177, %f2175, %p257; add.ftz.f32 %f2192, %f2190, %f2191; mul.ftz.f32 %f2193, %f2192, 0f3F000000; sub.ftz.f32 %f2194, %f1931, %f2193; mul.ftz.f32 %f2195, %f2194, 0f40400000; sqrt.approx.ftz.f32 %f2196, %f2195; mul.ftz.f32 %f2197, %f2193, 0f3F000000; mul.ftz.f32 %f2198, %f351, %f2196; mul.ftz.f32 %f2199, %f2198, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2200, %f2199; add.ftz.f32 %f2201, %f2200, 0f3F800000; mul.ftz.f32 %f2202, %f2201, %f2197; mul.ftz.f32 %f2203, %f2196, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2204, %f2203; mul.ftz.f32 %f2205, %f2204, %f2202; setp.lt.ftz.f32 %p258, %f2205, %f355; selp.f32 %f2206, %f2190, %f2193, %p258; selp.f32 %f2207, %f2193, %f2191, %p258; add.ftz.f32 %f2208, %f2206, %f2207; mul.ftz.f32 %f2209, %f2208, 0f3F000000; sub.ftz.f32 %f2210, %f1931, %f2209; mul.ftz.f32 %f2211, %f2210, 0f40400000; sqrt.approx.ftz.f32 %f2212, %f2211; mul.ftz.f32 %f2213, %f2209, 0f3F000000; mul.ftz.f32 %f2214, %f351, %f2212; mul.ftz.f32 %f2215, %f2214, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2216, %f2215; add.ftz.f32 %f2217, %f2216, 0f3F800000; mul.ftz.f32 %f2218, %f2217, %f2213; mul.ftz.f32 %f2219, %f2212, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2220, %f2219; mul.ftz.f32 %f2221, %f2220, %f2218; setp.lt.ftz.f32 %p259, %f2221, %f355; selp.f32 %f2222, %f2206, %f2209, %p259; selp.f32 %f2223, %f2209, %f2207, %p259; add.ftz.f32 %f2224, %f2222, %f2223; mul.ftz.f32 %f2225, %f2224, 0f3F000000; sub.ftz.f32 %f2226, %f1931, %f2225; mul.ftz.f32 %f2227, %f2226, 0f40400000; sqrt.approx.ftz.f32 %f2228, %f2227; mul.ftz.f32 %f2229, %f2225, 0f3F000000; mul.ftz.f32 %f2230, %f351, %f2228; mul.ftz.f32 %f2231, %f2230, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2232, %f2231; add.ftz.f32 %f2233, %f2232, 0f3F800000; mul.ftz.f32 %f2234, %f2233, %f2229; mul.ftz.f32 %f2235, %f2228, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2236, %f2235; mul.ftz.f32 %f2237, %f2236, %f2234; setp.lt.ftz.f32 %p260, %f2237, %f355; selp.f32 %f2238, %f2222, %f2225, %p260; selp.f32 %f2239, %f2225, %f2223, %p260; add.ftz.f32 %f2240, %f2238, %f2239; mul.ftz.f32 %f2241, %f2240, 0f3F000000; sub.ftz.f32 %f2242, %f1931, %f2241; mul.ftz.f32 %f2243, %f2242, 0f40400000; sqrt.approx.ftz.f32 %f2244, %f2243; mul.ftz.f32 %f2245, %f2241, 0f3F000000; mul.ftz.f32 %f2246, %f351, %f2244; mul.ftz.f32 %f2247, %f2246, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2248, %f2247; add.ftz.f32 %f2249, %f2248, 0f3F800000; mul.ftz.f32 %f2250, %f2249, %f2245; mul.ftz.f32 %f2251, %f2244, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2252, %f2251; mul.ftz.f32 %f2253, %f2252, %f2250; setp.lt.ftz.f32 %p261, %f2253, %f355; selp.f32 %f2254, %f2238, %f2241, %p261; selp.f32 %f2255, %f2241, %f2239, %p261; add.ftz.f32 %f2256, %f2254, %f2255; mul.ftz.f32 %f2257, %f2256, 0f3F000000; sub.ftz.f32 %f2258, %f1931, %f2257; mul.ftz.f32 %f2259, %f2258, 0f40400000; sqrt.approx.ftz.f32 %f2260, %f2259; mul.ftz.f32 %f2261, %f2257, 0f3F000000; mul.ftz.f32 %f2262, %f351, %f2260; mul.ftz.f32 %f2263, %f2262, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2264, %f2263; add.ftz.f32 %f2265, %f2264, 0f3F800000; mul.ftz.f32 %f2266, %f2265, %f2261; mul.ftz.f32 %f2267, %f2260, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2268, %f2267; mul.ftz.f32 %f2269, %f2268, %f2266; setp.lt.ftz.f32 %p262, %f2269, %f355; selp.f32 %f2270, %f2254, %f2257, %p262; selp.f32 %f2271, %f2257, %f2255, %p262; add.ftz.f32 %f2272, %f2270, %f2271; mul.ftz.f32 %f2273, %f2272, 0f3F000000; sub.ftz.f32 %f2274, %f1931, %f2273; mul.ftz.f32 %f2275, %f2274, 0f40400000; sqrt.approx.ftz.f32 %f2276, %f2275; mul.ftz.f32 %f2277, %f2273, 0f3F000000; mul.ftz.f32 %f2278, %f351, %f2276; mul.ftz.f32 %f2279, %f2278, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2280, %f2279; add.ftz.f32 %f2281, %f2280, 0f3F800000; mul.ftz.f32 %f2282, %f2281, %f2277; mul.ftz.f32 %f2283, %f2276, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2284, %f2283; mul.ftz.f32 %f2285, %f2284, %f2282; setp.lt.ftz.f32 %p263, %f2285, %f355; selp.f32 %f2286, %f2270, %f2273, %p263; selp.f32 %f2287, %f2273, %f2271, %p263; add.ftz.f32 %f2288, %f2286, %f2287; mul.ftz.f32 %f3026, %f2288, 0f3F000000; $L__BB184_310: ld.f32 %f358, [%rd111+20]; setp.lt.ftz.f32 %p264, %f358, 0f38D1B717; @%p264 bra $L__BB184_313; setp.ge.ftz.f32 %p265, %f358, 0f3F7EB852; mov.f32 %f3027, 0f3F7FFFEF; @%p265 bra $L__BB184_313; sub.ftz.f32 %f2292, %f1931, 0f3F000000; mul.ftz.f32 %f2293, %f2292, 0f40400000; sqrt.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f351, %f2294; mul.ftz.f32 %f2296, %f2295, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2297, %f2296; add.ftz.f32 %f2298, %f2297, 0f3F800000; mul.ftz.f32 %f2299, %f2298, 0f3E800000; mul.ftz.f32 %f2300, %f2294, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2301, %f2300; mul.ftz.f32 %f2302, %f2301, %f2299; setp.lt.ftz.f32 %p266, %f2302, %f358; selp.f32 %f2303, 0f3F800000, 0f3F000000, %p266; selp.f32 %f2304, 0f3F000000, 0f00000000, %p266; selp.f32 %f2305, 0f3F400000, 0f3E800000, %p266; sub.ftz.f32 %f2306, %f1931, %f2305; mul.ftz.f32 %f2307, %f2306, 0f40400000; sqrt.approx.ftz.f32 %f2308, %f2307; mul.ftz.f32 %f2309, %f2305, 0f3F000000; mul.ftz.f32 %f2310, %f351, %f2308; mul.ftz.f32 %f2311, %f2310, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2312, %f2311; add.ftz.f32 %f2313, %f2312, 0f3F800000; mul.ftz.f32 %f2314, %f2313, %f2309; mul.ftz.f32 %f2315, %f2308, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2316, %f2315; mul.ftz.f32 %f2317, %f2316, %f2314; setp.lt.ftz.f32 %p267, %f2317, %f358; selp.f32 %f2318, %f2303, %f2305, %p267; selp.f32 %f2319, %f2305, %f2304, %p267; add.ftz.f32 %f2320, %f2318, %f2319; mul.ftz.f32 %f2321, %f2320, 0f3F000000; sub.ftz.f32 %f2322, %f1931, %f2321; mul.ftz.f32 %f2323, %f2322, 0f40400000; sqrt.approx.ftz.f32 %f2324, %f2323; mul.ftz.f32 %f2325, %f2321, 0f3F000000; mul.ftz.f32 %f2326, %f351, %f2324; mul.ftz.f32 %f2327, %f2326, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2328, %f2327; add.ftz.f32 %f2329, %f2328, 0f3F800000; mul.ftz.f32 %f2330, %f2329, %f2325; mul.ftz.f32 %f2331, %f2324, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2332, %f2331; mul.ftz.f32 %f2333, %f2332, %f2330; setp.lt.ftz.f32 %p268, %f2333, %f358; selp.f32 %f2334, %f2318, %f2321, %p268; selp.f32 %f2335, %f2321, %f2319, %p268; add.ftz.f32 %f2336, %f2334, %f2335; mul.ftz.f32 %f2337, %f2336, 0f3F000000; sub.ftz.f32 %f2338, %f1931, %f2337; mul.ftz.f32 %f2339, %f2338, 0f40400000; sqrt.approx.ftz.f32 %f2340, %f2339; mul.ftz.f32 %f2341, %f2337, 0f3F000000; mul.ftz.f32 %f2342, %f351, %f2340; mul.ftz.f32 %f2343, %f2342, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2344, %f2343; add.ftz.f32 %f2345, %f2344, 0f3F800000; mul.ftz.f32 %f2346, %f2345, %f2341; mul.ftz.f32 %f2347, %f2340, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2348, %f2347; mul.ftz.f32 %f2349, %f2348, %f2346; setp.lt.ftz.f32 %p269, %f2349, %f358; selp.f32 %f2350, %f2334, %f2337, %p269; selp.f32 %f2351, %f2337, %f2335, %p269; add.ftz.f32 %f2352, %f2350, %f2351; mul.ftz.f32 %f2353, %f2352, 0f3F000000; sub.ftz.f32 %f2354, %f1931, %f2353; mul.ftz.f32 %f2355, %f2354, 0f40400000; sqrt.approx.ftz.f32 %f2356, %f2355; mul.ftz.f32 %f2357, %f2353, 0f3F000000; mul.ftz.f32 %f2358, %f351, %f2356; mul.ftz.f32 %f2359, %f2358, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2360, %f2359; add.ftz.f32 %f2361, %f2360, 0f3F800000; mul.ftz.f32 %f2362, %f2361, %f2357; mul.ftz.f32 %f2363, %f2356, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2364, %f2363; mul.ftz.f32 %f2365, %f2364, %f2362; setp.lt.ftz.f32 %p270, %f2365, %f358; selp.f32 %f2366, %f2350, %f2353, %p270; selp.f32 %f2367, %f2353, %f2351, %p270; add.ftz.f32 %f2368, %f2366, %f2367; mul.ftz.f32 %f2369, %f2368, 0f3F000000; sub.ftz.f32 %f2370, %f1931, %f2369; mul.ftz.f32 %f2371, %f2370, 0f40400000; sqrt.approx.ftz.f32 %f2372, %f2371; mul.ftz.f32 %f2373, %f2369, 0f3F000000; mul.ftz.f32 %f2374, %f351, %f2372; mul.ftz.f32 %f2375, %f2374, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2376, %f2375; add.ftz.f32 %f2377, %f2376, 0f3F800000; mul.ftz.f32 %f2378, %f2377, %f2373; mul.ftz.f32 %f2379, %f2372, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2380, %f2379; mul.ftz.f32 %f2381, %f2380, %f2378; setp.lt.ftz.f32 %p271, %f2381, %f358; selp.f32 %f2382, %f2366, %f2369, %p271; selp.f32 %f2383, %f2369, %f2367, %p271; add.ftz.f32 %f2384, %f2382, %f2383; mul.ftz.f32 %f2385, %f2384, 0f3F000000; sub.ftz.f32 %f2386, %f1931, %f2385; mul.ftz.f32 %f2387, %f2386, 0f40400000; sqrt.approx.ftz.f32 %f2388, %f2387; mul.ftz.f32 %f2389, %f2385, 0f3F000000; mul.ftz.f32 %f2390, %f351, %f2388; mul.ftz.f32 %f2391, %f2390, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2392, %f2391; add.ftz.f32 %f2393, %f2392, 0f3F800000; mul.ftz.f32 %f2394, %f2393, %f2389; mul.ftz.f32 %f2395, %f2388, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f2396, %f2394; setp.lt.ftz.f32 %p272, %f2397, %f358; selp.f32 %f2398, %f2382, %f2385, %p272; selp.f32 %f2399, %f2385, %f2383, %p272; add.ftz.f32 %f2400, %f2398, %f2399; mul.ftz.f32 %f2401, %f2400, 0f3F000000; sub.ftz.f32 %f2402, %f1931, %f2401; mul.ftz.f32 %f2403, %f2402, 0f40400000; sqrt.approx.ftz.f32 %f2404, %f2403; mul.ftz.f32 %f2405, %f2401, 0f3F000000; mul.ftz.f32 %f2406, %f351, %f2404; mul.ftz.f32 %f2407, %f2406, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2408, %f2407; add.ftz.f32 %f2409, %f2408, 0f3F800000; mul.ftz.f32 %f2410, %f2409, %f2405; mul.ftz.f32 %f2411, %f2404, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2412, %f2411; mul.ftz.f32 %f2413, %f2412, %f2410; setp.lt.ftz.f32 %p273, %f2413, %f358; selp.f32 %f2414, %f2398, %f2401, %p273; selp.f32 %f2415, %f2401, %f2399, %p273; add.ftz.f32 %f2416, %f2414, %f2415; mul.ftz.f32 %f2417, %f2416, 0f3F000000; sub.ftz.f32 %f2418, %f1931, %f2417; mul.ftz.f32 %f2419, %f2418, 0f40400000; sqrt.approx.ftz.f32 %f2420, %f2419; mul.ftz.f32 %f2421, %f2417, 0f3F000000; mul.ftz.f32 %f2422, %f351, %f2420; mul.ftz.f32 %f2423, %f2422, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2424, %f2423; add.ftz.f32 %f2425, %f2424, 0f3F800000; mul.ftz.f32 %f2426, %f2425, %f2421; mul.ftz.f32 %f2427, %f2420, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2428, %f2427; mul.ftz.f32 %f2429, %f2428, %f2426; setp.lt.ftz.f32 %p274, %f2429, %f358; selp.f32 %f2430, %f2414, %f2417, %p274; selp.f32 %f2431, %f2417, %f2415, %p274; add.ftz.f32 %f2432, %f2430, %f2431; mul.ftz.f32 %f2433, %f2432, 0f3F000000; sub.ftz.f32 %f2434, %f1931, %f2433; mul.ftz.f32 %f2435, %f2434, 0f40400000; sqrt.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f2433, 0f3F000000; mul.ftz.f32 %f2438, %f351, %f2436; mul.ftz.f32 %f2439, %f2438, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2440, %f2439; add.ftz.f32 %f2441, %f2440, 0f3F800000; mul.ftz.f32 %f2442, %f2441, %f2437; mul.ftz.f32 %f2443, %f2436, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2444, %f2443; mul.ftz.f32 %f2445, %f2444, %f2442; setp.lt.ftz.f32 %p275, %f2445, %f358; selp.f32 %f2446, %f2430, %f2433, %p275; selp.f32 %f2447, %f2433, %f2431, %p275; add.ftz.f32 %f2448, %f2446, %f2447; mul.ftz.f32 %f2449, %f2448, 0f3F000000; sub.ftz.f32 %f2450, %f1931, %f2449; mul.ftz.f32 %f2451, %f2450, 0f40400000; sqrt.approx.ftz.f32 %f2452, %f2451; mul.ftz.f32 %f2453, %f2449, 0f3F000000; mul.ftz.f32 %f2454, %f351, %f2452; mul.ftz.f32 %f2455, %f2454, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2456, %f2455; add.ftz.f32 %f2457, %f2456, 0f3F800000; mul.ftz.f32 %f2458, %f2457, %f2453; mul.ftz.f32 %f2459, %f2452, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2460, %f2459; mul.ftz.f32 %f2461, %f2460, %f2458; setp.lt.ftz.f32 %p276, %f2461, %f358; selp.f32 %f2462, %f2446, %f2449, %p276; selp.f32 %f2463, %f2449, %f2447, %p276; add.ftz.f32 %f2464, %f2462, %f2463; mul.ftz.f32 %f3027, %f2464, 0f3F000000; $L__BB184_313: sub.ftz.f32 %f2466, %f1931, %f3025; mul.ftz.f32 %f2467, %f2466, 0f40400000; sub.ftz.f32 %f2468, %f1931, %f3026; mul.ftz.f32 %f2469, %f2468, 0f40400000; sub.ftz.f32 %f2470, %f1931, %f3027; mul.ftz.f32 %f2471, %f2470, 0f40400000; sqrt.approx.ftz.f32 %f2472, %f2467; sqrt.approx.ftz.f32 %f2473, %f2469; sqrt.approx.ftz.f32 %f2474, %f2471; ld.f32 %f2475, [%rd111]; mul.ftz.f32 %f3028, %f2472, %f2475; ld.f32 %f2476, [%rd111+4]; mul.ftz.f32 %f3029, %f2473, %f2476; ld.f32 %f2477, [%rd111+8]; mul.ftz.f32 %f3030, %f2474, %f2477; $L__BB184_315: st.f32 [%rd111], %f3028; st.f32 [%rd111+4], %f3029; st.f32 [%rd111+8], %f3030; or.b32 %r615, %r615, 16; $L__BB184_316: ld.local.u32 %r349, [%rd4+40]; or.b32 %r350, %r349, %r615; st.local.u32 [%rd4+40], %r350; bra.uni $L__BB184_499; $L__BB184_94: mul.wide.u32 %rd279, %r158, 4; add.s64 %rd280, %rd2, %rd279; ld.local.f32 %f2977, [%rd280]; ld.local.f32 %f2978, [%rd280+4]; ld.local.f32 %f2979, [%rd280+8]; $L__BB184_96: ld.global.v4.u32 {%r162, %r163, %r164, %r165}, [%rd3+48]; setp.eq.s32 %p55, %r162, 255; @%p55 bra $L__BB184_98; bra.uni $L__BB184_97; $L__BB184_98: ld.local.f32 %f2987, [%rd4+8]; ld.local.f32 %f2986, [%rd4+4]; ld.local.f32 %f2985, [%rd4]; bra.uni $L__BB184_99; $L__BB184_329: mul.wide.u32 %rd477, %r78, 4; add.s64 %rd478, %rd2, %rd477; ld.local.f32 %f3031, [%rd478]; $L__BB184_331: setp.eq.s32 %p286, %r77, 255; mul.ftz.f32 %f380, %f2, %f376; mul.ftz.f32 %f381, %f2, %f375; mul.ftz.f32 %f382, %f2, %f374; @%p286 bra $L__BB184_333; bra.uni $L__BB184_332; $L__BB184_333: mov.b32 %f3032, %r138; bra.uni $L__BB184_334; $L__BB184_97: mul.wide.u32 %rd281, %r162, 4; add.s64 %rd282, %rd2, %rd281; ld.local.f32 %f2987, [%rd282+8]; ld.local.f32 %f2986, [%rd282+4]; ld.local.f32 %f2985, [%rd282]; $L__BB184_99: @%p9 bra $L__BB184_108; ld.local.v2.f32 {%f673, %f674}, [%rd4+12]; mul.ftz.f32 %f675, %f2986, %f91; fma.rn.ftz.f32 %f676, %f2985, %f90, %f675; fma.rn.ftz.f32 %f677, %f2987, %f92, %f676; add.ftz.f32 %f678, %f677, %f677; mul.ftz.f32 %f679, %f2985, %f678; mul.ftz.f32 %f680, %f2986, %f678; mul.ftz.f32 %f681, %f2987, %f678; sub.ftz.f32 %f682, %f679, %f90; sub.ftz.f32 %f683, %f680, %f91; sub.ftz.f32 %f684, %f681, %f92; mul.ftz.f32 %f685, %f674, %f91; fma.rn.ftz.f32 %f686, %f673, %f90, %f685; ld.local.f32 %f123, [%rd4+20]; fma.rn.ftz.f32 %f124, %f123, %f92, %f686; mul.ftz.f32 %f687, %f124, 0f3F666666; mov.f32 %f688, 0f3C23D70A; min.ftz.f32 %f125, %f687, %f688; mul.ftz.f32 %f689, %f674, %f683; fma.rn.ftz.f32 %f690, %f673, %f682, %f689; fma.rn.ftz.f32 %f691, %f123, %f684, %f690; setp.ltu.ftz.f32 %p57, %f691, %f125; @%p57 bra $L__BB184_101; bra.uni $L__BB184_108; $L__BB184_101: mul.ftz.f32 %f692, %f2986, %f674; fma.rn.ftz.f32 %f693, %f2985, %f673, %f692; fma.rn.ftz.f32 %f694, %f2987, %f123, %f693; mul.ftz.f32 %f695, %f673, %f694; mul.ftz.f32 %f696, %f674, %f694; mul.ftz.f32 %f697, %f123, %f694; sub.ftz.f32 %f698, %f2985, %f695; sub.ftz.f32 %f699, %f2986, %f696; sub.ftz.f32 %f700, %f2987, %f697; mul.ftz.f32 %f701, %f699, %f699; fma.rn.ftz.f32 %f702, %f698, %f698, %f701; fma.rn.ftz.f32 %f703, %f700, %f700, %f702; sqrt.approx.ftz.f32 %f704, %f703; div.approx.ftz.f32 %f129, %f698, %f704; div.approx.ftz.f32 %f130, %f699, %f704; div.approx.ftz.f32 %f131, %f700, %f704; mul.ftz.f32 %f705, %f91, %f130; fma.rn.ftz.f32 %f706, %f90, %f129, %f705; fma.rn.ftz.f32 %f132, %f92, %f131, %f706; mul.ftz.f32 %f707, %f132, %f132; fma.rn.ftz.f32 %f708, %f124, %f124, %f707; mul.ftz.f32 %f709, %f125, %f125; sub.ftz.f32 %f710, %f708, %f709; mul.ftz.f32 %f711, %f707, %f710; mov.f32 %f712, 0f00000000; max.ftz.f32 %f713, %f711, %f712; sqrt.approx.ftz.f32 %f714, %f713; fma.rn.ftz.f32 %f715, %f125, %f124, %f708; mov.f32 %f716, 0f3F000000; div.approx.ftz.f32 %f717, %f716, %f708; add.ftz.f32 %f718, %f714, %f715; mul.ftz.f32 %f133, %f717, %f718; sub.ftz.f32 %f719, %f715, %f714; mul.ftz.f32 %f134, %f717, %f719; setp.gt.ftz.f32 %p58, %f133, 0f3727C5AC; setp.le.ftz.f32 %p59, %f133, 0f3F800054; and.pred %p3, %p58, %p59; setp.gt.ftz.f32 %p60, %f134, 0f3727C5AC; setp.le.ftz.f32 %p61, %f134, 0f3F800054; and.pred %p4, %p60, %p61; and.pred %p62, %p3, %p4; @%p62 bra $L__BB184_104; bra.uni $L__BB184_102; $L__BB184_104: mov.f32 %f726, 0f3F800000; sub.ftz.f32 %f727, %f726, %f133; max.ftz.f32 %f729, %f727, %f712; sqrt.approx.ftz.f32 %f140, %f729; max.ftz.f32 %f730, %f133, %f712; sqrt.approx.ftz.f32 %f141, %f730; sub.ftz.f32 %f731, %f726, %f134; max.ftz.f32 %f732, %f731, %f712; sqrt.approx.ftz.f32 %f142, %f732; max.ftz.f32 %f733, %f134, %f712; sqrt.approx.ftz.f32 %f143, %f733; mul.ftz.f32 %f734, %f124, %f141; fma.rn.ftz.f32 %f735, %f132, %f140, %f734; add.ftz.f32 %f736, %f735, %f735; mul.ftz.f32 %f737, %f141, %f736; sub.ftz.f32 %f144, %f737, %f124; mul.ftz.f32 %f738, %f124, %f143; fma.rn.ftz.f32 %f739, %f132, %f142, %f738; add.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f143, %f740; sub.ftz.f32 %f145, %f741, %f124; setp.ge.ftz.f32 %p64, %f144, 0f3727C5AC; setp.ge.ftz.f32 %p65, %f145, 0f3727C5AC; and.pred %p66, %p64, %p65; @%p66 bra $L__BB184_106; bra.uni $L__BB184_105; $L__BB184_106: setp.lt.ftz.f32 %p68, %f144, %f145; selp.f32 %f2983, %f140, %f142, %p68; selp.f32 %f2984, %f141, %f143, %p68; bra.uni $L__BB184_107; $L__BB184_332: mul.wide.u32 %rd479, %r77, 4; add.s64 %rd480, %rd2, %rd479; ld.local.f32 %f3032, [%rd480]; $L__BB184_334: and.b32 %r81, %r362, 255; bfe.u32 %r82, %r362, 8, 8; bfe.u32 %r83, %r362, 16, 8; bfe.u32 %r84, %r366, 8, 8; bfe.u32 %r85, %r366, 16, 8; shr.u32 %r86, %r366, 24; ld.local.u32 %r87, [%rd4+64]; ld.local.s32 %rd122, [%rd4+48]; setp.eq.s32 %p287, %r87, -1; mov.u32 %r108, -1; mov.u32 %r105, 0; mov.u32 %r625, %r105; mov.u32 %r626, %r105; @%p287 bra $L__BB184_349; ld.const.u64 %rd481, [kernel_params+2160]; cvta.to.global.u64 %rd482, %rd481; mul.wide.s32 %rd483, %r87, 208; add.s64 %rd484, %rd482, %rd483; ld.global.u32 %r88, [%rd484+164]; and.b32 %r375, %r11, 1; setp.eq.b32 %p288, %r375, 1; mov.pred %p289, 0; xor.pred %p290, %p288, %p289; not.pred %p291, %p290; mov.u32 %r619, 0; @%p291 bra $L__BB184_339; cvt.u32.u64 %r377, %rd122; setp.eq.s32 %p292, %r377, -1; mov.u32 %r618, -1; @%p292 bra $L__BB184_338; ld.const.u64 %rd485, [kernel_params+2240]; cvta.to.global.u64 %rd486, %rd485; shl.b64 %rd487, %rd122, 2; add.s64 %rd488, %rd486, %rd487; ld.global.u32 %r618, [%rd488]; $L__BB184_338: setp.ne.s32 %p293, %r618, -1; selp.u32 %r619, 1, 0, %p293; $L__BB184_339: ld.const.u64 %rd489, [kernel_params+2312]; cvta.to.global.u64 %rd123, %rd489; add.s32 %r621, %r619, %r88; mul.wide.u32 %rd490, %r621, 16; add.s64 %rd491, %rd123, %rd490; add.s64 %rd754, %rd491, 14; ld.global.u16 %rs20, [%rd491+12]; ld.global.u32 %r620, [%rd491+8]; ld.global.u64 %rd752, [%rd491]; setp.eq.s64 %p294, %rd752, %rd121; add.s64 %rd753, %rd491, 15; @%p294 bra $L__BB184_345; $L__BB184_340: setp.eq.s64 %p295, %rd752, 0; @%p295 bra $L__BB184_342; add.s32 %r621, %r621, 2; bra.uni $L__BB184_343; $L__BB184_342: setp.eq.s16 %p296, %rs20, 0; mov.u32 %r105, 0; mov.u32 %r621, %r620; mov.u32 %r625, %r105; mov.u32 %r626, %r105; @%p296 bra $L__BB184_349; $L__BB184_343: cvt.u64.u32 %rd128, %r621; mul.wide.u32 %rd492, %r621, 16; add.s64 %rd493, %rd123, %rd492; ld.global.u16 %rs20, [%rd493+12]; ld.global.u32 %r620, [%rd493+8]; ld.global.u64 %rd752, [%rd493]; setp.ne.s64 %p297, %rd752, %rd121; @%p297 bra $L__BB184_340; shl.b64 %rd494, %rd128, 4; add.s64 %rd495, %rd123, %rd494; add.s64 %rd754, %rd495, 14; add.s64 %rd753, %rd495, 15; $L__BB184_345: ld.global.u8 %rs5, [%rd754]; ld.global.u8 %rs6, [%rd753]; cvt.u32.u64 %r382, %rd122; setp.ne.s32 %p298, %r382, -1; @%p298 bra $L__BB184_348; mov.u32 %r105, 0; add.s16 %rs14, %rs20, -1; setp.lt.u16 %p299, %rs14, 2; @%p299 bra $L__BB184_348; setp.ne.s16 %p300, %rs20, 1024; mov.u32 %r625, %r105; mov.u32 %r626, %r105; @%p300 bra $L__BB184_349; $L__BB184_348: setp.eq.s16 %p301, %rs20, 0; selp.b32 %r108, -1, %r620, %p301; cvt.u32.u16 %r105, %rs20; cvt.u32.u16 %r387, %rs6; and.b32 %r626, %r387, 255; cvt.u32.u16 %r388, %rs5; and.b32 %r625, %r388, 255; $L__BB184_349: setp.eq.s32 %p302, %r108, -1; @%p302 bra $L__BB184_369; and.b32 %r389, %r11, 1; setp.eq.b32 %p303, %r389, 1; mov.pred %p304, 0; xor.pred %p305, %p303, %p304; not.pred %p306, %p305; @%p306 bra $L__BB184_362; bra.uni $L__BB184_351; $L__BB184_362: setp.eq.s32 %p313, %r12, 0; @%p313 bra $L__BB184_367; and.b32 %r410, %r105, 768; setp.eq.s32 %p314, %r410, 0; @%p314 bra $L__BB184_365; ld.const.u64 %rd517, [kernel_params+2264]; cvta.to.global.u64 %rd518, %rd517; shl.b64 %rd519, %rd122, 4; add.s64 %rd520, %rd518, %rd519; shr.s32 %r411, %r11, 7; ld.global.u32 %r412, [%rd520+4]; add.s32 %r413, %r411, %r412; ld.const.u64 %rd521, [kernel_params+2320]; cvta.to.global.u64 %rd522, %rd521; add.s32 %r414, %r413, %r108; mul.wide.s32 %rd523, %r414, 4; add.s64 %rd524, %rd522, %rd523; ld.local.f32 %f2515, [%rd4+56]; mov.f32 %f2516, 0f3F800000; sub.ftz.f32 %f2517, %f2516, %f2515; ld.global.f32 %f2518, [%rd524]; mul.ftz.f32 %f2519, %f2518, %f2517; ld.global.f32 %f2520, [%rd524+4]; fma.rn.ftz.f32 %f3033, %f2520, %f2515, %f2519; bra.uni $L__BB184_372; $L__BB184_455: st.local.u64 [%rd185+28], %rd621; st.local.v2.f32 [%rd185+12], {%f503, %f503}; mov.u32 %r505, 23; st.local.u32 [%rd185+-8], %r505; ld.local.u32 %r506, [%rd4+40]; or.b32 %r507, %r506, 12; st.local.u32 [%rd4+40], %r507; bra.uni $L__BB184_499; $L__BB184_426: st.local.u64 [%rd160+28], %rd579; st.local.v2.f32 [%rd160+12], {%f481, %f481}; mov.u32 %r451, 10; st.local.u32 [%rd160+-8], %r451; ld.local.u32 %r452, [%rd4+40]; or.b32 %r453, %r452, 12; st.local.u32 [%rd4+40], %r453; $L__BB184_429: mov.f32 %f2708, 0f3F800000; sub.ftz.f32 %f2709, %f2708, %f3050; mov.f32 %f2710, 0f38D1B717; max.ftz.f32 %f2711, %f2710, %f2709; mul.ftz.f32 %f2712, %f466, %f2711; mul.ftz.f32 %f2713, %f467, %f2711; mul.ftz.f32 %f2714, %f468, %f2711; max.ftz.f32 %f483, %f2712, %f2700; max.ftz.f32 %f484, %f2713, %f2700; max.ftz.f32 %f485, %f2714, %f2700; add.ftz.f32 %f2716, %f483, %f484; add.ftz.f32 %f2717, %f485, %f2716; mul.ftz.f32 %f2718, %f2717, 0f3EAAAAAB; abs.ftz.f32 %f486, %f2718; setp.ltu.ftz.f32 %p350, %f486, 0f3727C5AC; mov.u64 %rd770, 0; mov.u64 %rd771, %rd770; @%p350 bra $L__BB184_434; ld.local.u32 %r117, [%rd4+252]; mov.u64 %rd770, 0; setp.eq.s32 %p351, %r117, 0; mov.u64 %rd768, %rd770; mov.u64 %rd769, %rd770; @%p351 bra $L__BB184_432; ld.local.u32 %r460, [%rd4+248]; mul.wide.s32 %rd585, %r460, 80; add.s64 %rd586, %rd1, %rd585; add.s64 %rd768, %rd586, 304; add.s64 %rd587, %rd230, %rd585; add.s64 %rd769, %rd587, 304; mov.b32 %r461, %f485; mov.b32 %r462, %f484; mov.b32 %r463, %f483; mov.u32 %r464, 0; st.local.v4.u32 [%rd586+304], {%r463, %r462, %r461, %r464}; add.s32 %r465, %r460, 1; st.local.u32 [%rd4+248], %r465; add.s32 %r466, %r117, -1; st.local.u32 [%rd4+252], %r466; $L__BB184_432: setp.eq.s64 %p352, %rd769, 0; mov.u64 %rd771, %rd770; @%p352 bra $L__BB184_434; st.local.f32 [%rd768+16], %f486; mov.u64 %rd770, %rd768; mov.u64 %rd771, %rd769; $L__BB184_434: setp.eq.s64 %p353, %rd771, 0; @%p353 bra $L__BB184_499; add.s64 %rd167, %rd770, 20; mov.u64 %rd590, 0; st.local.f32 [%rd770+20], %f55; st.local.f32 [%rd770+24], %f56; st.local.f32 [%rd770+28], %f57; mov.u32 %r467, 0; st.local.v2.f32 [%rd770+56], {%f2700, %f2700}; st.local.u32 [%rd770+64], %r467; st.local.u64 [%rd770+48], %rd590; setp.eq.s32 %p354, %r2, 28; @%p354 bra $L__BB184_439; bra.uni $L__BB184_436; $L__BB184_439: st.local.v2.f32 [%rd167+12], {%f2700, %f2700}; st.local.f32 [%rd167+20], %f3049; mov.u32 %r474, 21; st.local.u32 [%rd167+-8], %r474; ld.local.u32 %r475, [%rd4+40]; or.b32 %r476, %r475, 4; st.local.u32 [%rd4+40], %r476; bra.uni $L__BB184_499; $L__BB184_436: setp.eq.s32 %p355, %r2, 25; st.local.v2.f32 [%rd167+12], {%f476, %f476}; st.local.f32 [%rd167+20], %f3049; cvt.ftz.sat.f32.f32 %f487, %f476; @%p355 bra $L__BB184_438; bra.uni $L__BB184_437; $L__BB184_438: st.local.v2.f32 [%rd167+12], {%f487, %f487}; mov.u32 %r471, 22; st.local.u32 [%rd167+-8], %r471; ld.local.u32 %r472, [%rd4+40]; or.b32 %r473, %r472, 12; st.local.u32 [%rd4+40], %r473; bra.uni $L__BB184_499; $L__BB184_437: st.local.u64 [%rd167+28], %rd590; st.local.v2.f32 [%rd167+12], {%f487, %f487}; mov.u32 %r468, 23; st.local.u32 [%rd167+-8], %r468; ld.local.u32 %r469, [%rd4+40]; or.b32 %r470, %r469, 12; st.local.u32 [%rd4+40], %r470; bra.uni $L__BB184_499; $L__BB184_476: setp.eq.s32 %p397, %r2, 14; @%p397 bra $L__BB184_478; bra.uni $L__BB184_477; $L__BB184_478: ld.local.u32 %r124, [%rd4+252]; setp.lt.s32 %p398, %r124, 1; @%p398 bra $L__BB184_480; add.s32 %r550, %r124, -1; st.local.u32 [%rd4+252], %r550; ld.local.s32 %rd651, [%rd4+248]; cvt.s64.s32 %rd652, %r550; add.s64 %rd653, %rd651, %rd652; mul.lo.s64 %rd654, %rd653, 80; add.s64 %rd655, %rd1, %rd654; add.s64 %rd794, %rd655, 304; add.s64 %rd656, %rd230, %rd654; add.s64 %rd795, %rd656, 304; bra.uni $L__BB184_481; $L__BB184_369: setp.eq.s32 %p320, %r84, 255; @%p320 bra $L__BB184_371; bra.uni $L__BB184_370; $L__BB184_371: mov.b32 %f3033, %r367; bra.uni $L__BB184_372; $L__BB184_351: cvt.u32.u64 %r391, %rd122; setp.eq.s32 %p307, %r391, -1; mov.u32 %r628, -1; @%p307 bra $L__BB184_353; ld.const.u64 %rd496, [kernel_params+2240]; cvta.to.global.u64 %rd497, %rd496; shl.b64 %rd498, %rd122, 2; add.s64 %rd499, %rd497, %rd498; ld.global.u32 %r628, [%rd499]; $L__BB184_353: setp.eq.s32 %p308, %r628, -1; @%p308 bra $L__BB184_355; mov.u64 %rd500, 0; { // callseq 942, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd230; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r105; st.param.b32 [param1+4], %r625; st.param.b32 [param1+8], %r626; st.param.b32 [param1+12], %r108; .param .b64 param2; st.param.b64 [param2+0], %rd500; .param .b64 param3; st.param.b64 [param3+0], %rd500; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f3033, [retval0+0]; } // callseq 942 bra.uni $L__BB184_372; $L__BB184_319: ld.local.f32 %f2487, [%rd4+280]; add.ftz.f32 %f2488, %f370, %f2487; ld.local.f32 %f2489, [%rd4+284]; add.ftz.f32 %f2490, %f371, %f2489; ld.local.f32 %f2491, [%rd4+288]; add.ftz.f32 %f2492, %f372, %f2491; st.local.f32 [%rd4+280], %f2488; st.local.f32 [%rd4+284], %f2490; st.local.f32 [%rd4+288], %f2492; ld.local.u32 %r65, [%rd4+248]; setp.lt.s32 %p279, %r65, 1; @%p279 bra $L__BB184_499; mov.u32 %r617, 0; $L__BB184_321: mul.wide.s32 %rd472, %r617, 80; add.s64 %rd473, %rd1, %rd472; add.s64 %rd118, %rd473, 316; ld.local.u32 %r353, [%rd473+316]; setp.eq.s32 %p280, %r353, 31; add.s32 %r617, %r617, 1; @%p280 bra $L__BB184_323; setp.lt.s32 %p281, %r617, %r65; @%p281 bra $L__BB184_321; bra.uni $L__BB184_499; $L__BB184_440: ld.local.u32 %r485, [%rd4+248]; mov.u64 %rd776, 0; add.s32 %r486, %r485, -1; st.local.u32 [%rd4+248], %r486; add.s32 %r487, %r119, 1; st.local.u32 [%rd4+252], %r487; mov.u64 %rd777, %rd776; $L__BB184_441: setp.eq.s64 %p361, %rd777, 0; @%p361 bra $L__BB184_499; add.s64 %rd178, %rd774, 20; st.local.f32 [%rd774+20], %f55; st.local.f32 [%rd774+24], %f56; st.local.f32 [%rd774+28], %f57; st.local.u64 [%rd774+48], %rd777; mov.u32 %r488, 0; st.local.v2.f32 [%rd774+56], {%f2727, %f2727}; st.local.u32 [%rd774+64], %r488; mul.ftz.f32 %f2732, %f60, %f60; st.local.v2.f32 [%rd774+32], {%f2732, %f2732}; mov.f32 %f2733, 0f3727C5AC; max.ftz.f32 %f3051, %f63, %f2733; ld.local.u32 %r489, [%rd4+40]; and.b32 %r490, %r489, 1; setp.eq.b32 %p362, %r490, 1; mov.pred %p363, 0; xor.pred %p364, %p362, %p363; not.pred %p365, %p364; @%p365 bra $L__BB184_444; rcp.approx.ftz.f32 %f3051, %f3051; $L__BB184_444: st.local.f32 [%rd178+20], %f3051; ld.local.f32 %f2734, [%rd6]; ld.local.f32 %f2735, [%rd6+4]; ld.local.f32 %f2736, [%rd6+8]; st.local.f32 [%rd776], %f2734; st.local.f32 [%rd776+4], %f2735; st.local.f32 [%rd776+8], %f2736; ld.local.u64 %rd609, [%rd178+28]; st.u32 [%rd609+12], %r488; st.u32 [%rd609+16], %r488; st.u32 [%rd609+20], %r488; ld.local.u64 %rd610, [%rd178+28]; st.u32 [%rd610+36], %r488; ld.local.f32 %f2738, [%rd178+12]; mov.f32 %f2739, 0f38D1B717; max.ftz.f32 %f2740, %f2738, %f2739; mov.f32 %f2741, 0f3F800000; min.ftz.f32 %f2742, %f2740, %f2741; st.local.v2.f32 [%rd178+12], {%f2742, %f2742}; ld.local.f32 %f2743, [%rd178+20]; max.ftz.f32 %f2744, %f2727, %f2743; st.local.f32 [%rd178+20], %f2744; ld.local.u64 %rd611, [%rd178+28]; ld.f32 %f2745, [%rd611]; cvt.ftz.sat.f32.f32 %f2746, %f2745; ld.f32 %f2747, [%rd611+4]; cvt.ftz.sat.f32.f32 %f2748, %f2747; ld.f32 %f2749, [%rd611+8]; cvt.ftz.sat.f32.f32 %f2750, %f2749; st.f32 [%rd611], %f2746; st.f32 [%rd611+4], %f2748; st.f32 [%rd611+8], %f2750; mov.u32 %r492, 24; st.local.u32 [%rd178+-8], %r492; ld.local.u32 %r493, [%rd4+40]; or.b32 %r494, %r493, 1036; st.local.u32 [%rd4+40], %r494; bra.uni $L__BB184_499; $L__BB184_460: ld.local.f32 %f2797, [%rd4+116]; ld.local.f32 %f2798, [%rd4+120]; mul.ftz.f32 %f2799, %f2798, %f2798; fma.rn.ftz.f32 %f2800, %f2797, %f2797, %f2799; ld.local.f32 %f2801, [%rd4+124]; fma.rn.ftz.f32 %f2802, %f2801, %f2801, %f2800; sqrt.approx.ftz.f32 %f2803, %f2802; div.approx.ftz.f32 %f2804, %f2797, %f2803; div.approx.ftz.f32 %f2805, %f2798, %f2803; div.approx.ftz.f32 %f2806, %f2801, %f2803; st.local.f32 [%rd192+12], %f2804; st.local.f32 [%rd192+16], %f2805; st.local.f32 [%rd192+20], %f2806; mov.u32 %r521, 0; st.local.u32 [%rd192+32], %r521; $L__BB184_461: setp.eq.s32 %p382, %r2, 20; mov.f32 %f2807, 0f3A83126F; max.ftz.f32 %f2808, %f60, %f2807; mov.f32 %f2809, 0f3F800000; min.ftz.f32 %f509, %f2808, %f2809; max.ftz.f32 %f2810, %f63, %f2807; min.ftz.f32 %f510, %f2810, %f2809; @%p382 bra $L__BB184_463; bra.uni $L__BB184_462; $L__BB184_463: mov.u32 %r525, 20; st.local.u32 [%rd192+-8], %r525; st.local.f32 [%rd192+24], %f509; st.local.f32 [%rd192+28], %f510; ld.local.u32 %r526, [%rd4+40]; or.b32 %r527, %r526, 12; st.local.u32 [%rd4+40], %r527; bra.uni $L__BB184_499; $L__BB184_462: mov.u32 %r522, 30; st.local.u32 [%rd192+-8], %r522; st.local.f32 [%rd192+24], %f509; st.local.f32 [%rd192+28], %f510; ld.local.u32 %r523, [%rd4+40]; or.b32 %r524, %r523, 12; st.local.u32 [%rd4+40], %r524; bra.uni $L__BB184_499; $L__BB184_496: st.local.f32 [%rd229+12], %f60; mov.u32 %r598, 3; st.local.u32 [%rd229+-8], %r598; cvt.ftz.sat.f32.f32 %f2947, %f60; fma.rn.ftz.f32 %f2948, %f2947, 0f3F67750B, 0f40490FDB; rcp.approx.ftz.f32 %f2949, %f2948; st.local.f32 [%rd229+16], %f2949; mul.ftz.f32 %f2950, %f2947, %f2949; st.local.f32 [%rd229+20], %f2950; ld.local.u32 %r599, [%rd4+40]; or.b32 %r600, %r599, 12; st.local.u32 [%rd4+40], %r600; bra.uni $L__BB184_499; $L__BB184_477: mov.f32 %f2875, 0f38D1B717; max.ftz.f32 %f2876, %f3054, %f2875; mov.f32 %f2877, 0f3F800000; max.ftz.f32 %f2878, %f3053, %f2875; min.ftz.f32 %f2879, %f2878, %f2877; min.ftz.f32 %f2880, %f2876, %f2877; st.local.v2.f32 [%rd205+12], {%f2880, %f2879}; mov.u32 %r547, 16; st.local.u32 [%rd205+-8], %r547; ld.local.u32 %r548, [%rd4+40]; or.b32 %r549, %r548, 12; st.local.u32 [%rd4+40], %r549; bra.uni $L__BB184_499; $L__BB184_370: mul.wide.u32 %rd533, %r84, 4; add.s64 %rd534, %rd2, %rd533; ld.local.f32 %f3033, [%rd534]; bra.uni $L__BB184_372; $L__BB184_367: and.b32 %r419, %r11, 8; setp.eq.s32 %p317, %r419, 0; setp.ne.s32 %p318, %r105, 8; mov.f32 %f3033, 0f00000000; or.pred %p319, %p318, %p317; @%p319 bra $L__BB184_372; cvt.u32.u64 %r420, %rd122; ld.const.u64 %rd529, [kernel_params+2320]; cvta.to.global.u64 %rd530, %rd529; add.s32 %r421, %r420, %r108; mul.wide.s32 %rd531, %r421, 4; add.s64 %rd532, %rd530, %rd531; ld.global.f32 %f3033, [%rd532]; bra.uni $L__BB184_372; $L__BB184_355: and.b32 %r392, %r105, 56; setp.eq.s32 %p309, %r392, 0; @%p309 bra $L__BB184_360; and.b32 %r393, %r105, 24; setp.eq.s32 %p310, %r393, 0; @%p310 bra $L__BB184_358; ld.const.u64 %rd501, [kernel_params+2232]; cvta.to.global.u64 %rd502, %rd501; shl.b64 %rd503, %rd122, 4; add.s64 %rd504, %rd502, %rd503; ld.global.v4.u32 {%r394, %r395, %r396, %r397}, [%rd504]; ld.const.u64 %rd505, [kernel_params+2320]; cvta.to.global.u64 %rd506, %rd505; add.s32 %r401, %r394, %r108; mul.wide.u32 %rd507, %r401, 4; add.s64 %rd755, %rd506, %rd507; add.s32 %r402, %r395, %r108; mul.wide.u32 %rd508, %r402, 4; add.s64 %rd756, %rd506, %rd508; add.s32 %r403, %r396, %r108; mul.wide.u32 %rd509, %r403, 4; add.s64 %rd757, %rd506, %rd509; bra.uni $L__BB184_359; $L__BB184_102: or.pred %p63, %p3, %p4; mov.f32 %f2985, %f673; mov.f32 %f2986, %f674; mov.f32 %f2987, %f123; @%p63 bra $L__BB184_103; bra.uni $L__BB184_108; $L__BB184_103: selp.f32 %f720, %f133, %f134, %p3; mov.f32 %f721, 0f3F800000; sub.ftz.f32 %f722, %f721, %f720; max.ftz.f32 %f724, %f722, %f712; sqrt.approx.ftz.f32 %f2983, %f724; max.ftz.f32 %f725, %f720, %f712; sqrt.approx.ftz.f32 %f2984, %f725; bra.uni $L__BB184_107; $L__BB184_323: ld.local.f32 %f2493, [%rd118+-12]; add.ftz.f32 %f2494, %f370, %f2493; ld.local.f32 %f2495, [%rd118+-8]; add.ftz.f32 %f2496, %f371, %f2495; ld.local.f32 %f2497, [%rd118+-4]; add.ftz.f32 %f2498, %f372, %f2497; st.local.f32 [%rd118+-12], %f2494; st.local.f32 [%rd118+-8], %f2496; st.local.f32 [%rd118+-4], %f2498; ld.local.f32 %f2499, [%rd118+4]; add.ftz.f32 %f2500, %f373, %f2499; st.local.f32 [%rd118+4], %f2500; bra.uni $L__BB184_499; $L__BB184_365: and.b32 %r415, %r105, 131; setp.eq.s32 %p315, %r415, 0; mov.f32 %f3033, 0f00000000; @%p315 bra $L__BB184_372; cvt.u32.u64 %r416, %rd122; setp.eq.s32 %p316, %r105, 128; selp.b32 %r417, %r416, 0, %p316; add.s32 %r418, %r108, %r417; ld.const.u64 %rd525, [kernel_params+2320]; cvta.to.global.u64 %rd526, %rd525; mul.wide.s32 %rd527, %r418, 4; add.s64 %rd528, %rd526, %rd527; ld.global.f32 %f3033, [%rd528]; bra.uni $L__BB184_372; $L__BB184_480: ld.local.u32 %r551, [%rd4+248]; add.s32 %r552, %r551, -1; st.local.u32 [%rd4+248], %r552; add.s32 %r553, %r124, 1; st.local.u32 [%rd4+252], %r553; mov.u64 %rd795, %rd794; $L__BB184_481: st.local.u64 [%rd205+28], %rd795; setp.eq.s64 %p399, %rd795, 0; @%p399 bra $L__BB184_499; ld.local.f32 %f2881, [%rd7]; ld.local.f32 %f2882, [%rd7+4]; ld.local.f32 %f2883, [%rd7+8]; st.local.f32 [%rd794], %f2881; st.local.f32 [%rd794+4], %f2882; st.local.f32 [%rd794+8], %f2883; ld.local.u64 %rd659, [%rd205+28]; mov.u32 %r554, 0; st.u32 [%rd659+12], %r554; st.u32 [%rd659+16], %r554; st.u32 [%rd659+20], %r554; ld.local.u64 %rd660, [%rd205+28]; st.u32 [%rd660+36], %r554; ld.local.v2.f32 {%f2884, %f2885}, [%rd205+36]; setp.neu.ftz.f32 %p400, %f2884, 0f00000000; setp.neu.ftz.f32 %p401, %f2885, 0f00000000; or.pred %p402, %p400, %p401; ld.local.f32 %f2888, [%rd205+44]; setp.neu.ftz.f32 %p403, %f2888, 0f00000000; or.pred %p404, %p403, %p402; @%p404 bra $L__BB184_484; mov.f32 %f2890, 0f3F800000; st.local.v2.f32 [%rd205+36], {%f2890, %f2832}; st.local.u32 [%rd205+44], %r554; $L__BB184_484: mov.u32 %r556, 14; st.local.u32 [%rd205+-8], %r556; ld.local.v2.f32 {%f2891, %f2892}, [%rd205+12]; mov.f32 %f2895, 0f38D1B717; max.ftz.f32 %f2896, %f2891, %f2895; mov.f32 %f2897, 0f3F800000; max.ftz.f32 %f2898, %f2892, %f2895; min.ftz.f32 %f2899, %f2898, %f2897; min.ftz.f32 %f2900, %f2896, %f2897; st.local.v2.f32 [%rd205+12], {%f2900, %f2899}; ld.local.u64 %rd661, [%rd205+28]; ld.f32 %f2901, [%rd661]; cvt.ftz.sat.f32.f32 %f2902, %f2901; ld.f32 %f2903, [%rd661+4]; cvt.ftz.sat.f32.f32 %f2904, %f2903; ld.f32 %f2905, [%rd661+8]; cvt.ftz.sat.f32.f32 %f2906, %f2905; st.f32 [%rd661], %f2902; st.f32 [%rd661+4], %f2904; st.f32 [%rd661+8], %f2906; ld.local.u64 %rd662, [%rd205+28]; ld.f32 %f2907, [%rd662+12]; cvt.ftz.sat.f32.f32 %f2908, %f2907; ld.f32 %f2909, [%rd662+16]; cvt.ftz.sat.f32.f32 %f2910, %f2909; ld.f32 %f2911, [%rd662+20]; cvt.ftz.sat.f32.f32 %f2912, %f2911; st.f32 [%rd662+12], %f2908; st.f32 [%rd662+16], %f2910; st.f32 [%rd662+20], %f2912; ld.local.u32 %r557, [%rd4+40]; or.b32 %r558, %r557, 1036; st.local.u32 [%rd4+40], %r558; bra.uni $L__BB184_499; $L__BB184_105: setp.gt.ftz.f32 %p67, %f144, %f145; selp.f32 %f2983, %f140, %f142, %p67; selp.f32 %f2984, %f141, %f143, %p67; $L__BB184_107: mul.ftz.f32 %f742, %f673, %f2984; mul.ftz.f32 %f743, %f674, %f2984; mul.ftz.f32 %f744, %f123, %f2984; fma.rn.ftz.f32 %f2987, %f131, %f2983, %f744; fma.rn.ftz.f32 %f2986, %f130, %f2983, %f743; fma.rn.ftz.f32 %f2985, %f129, %f2983, %f742; $L__BB184_108: setp.eq.s32 %p69, %r163, 255; mov.f32 %f2988, 0f3F800000; mov.f32 %f2989, %f2988; mov.f32 %f2990, %f2988; @%p69 bra $L__BB184_110; mul.wide.u32 %rd283, %r163, 4; add.s64 %rd284, %rd2, %rd283; ld.local.f32 %f2988, [%rd284]; ld.local.f32 %f2989, [%rd284+4]; ld.local.f32 %f2990, [%rd284+8]; $L__BB184_110: setp.eq.s32 %p70, %r164, 255; mov.f32 %f2991, 0f3FB33333; @%p70 bra $L__BB184_112; mul.wide.u32 %rd285, %r164, 4; add.s64 %rd286, %rd2, %rd285; ld.local.f32 %f2991, [%rd286]; $L__BB184_112: setp.eq.s32 %p71, %r165, 255; mov.f32 %f2992, 0f00000000; @%p71 bra $L__BB184_114; mul.wide.u32 %rd287, %r165, 4; add.s64 %rd288, %rd2, %rd287; ld.local.f32 %f2992, [%rd288]; $L__BB184_114: ld.global.v4.u32 {%r166, %r167, %r168, %r169}, [%rd3+64]; setp.eq.s32 %p72, %r166, 255; @%p72 bra $L__BB184_116; bra.uni $L__BB184_115; $L__BB184_116: mov.b32 %f2993, %r167; mov.b32 %f2994, %r168; mov.b32 %f2995, %r169; bra.uni $L__BB184_117; $L__BB184_115: mul.wide.u32 %rd289, %r166, 4; add.s64 %rd290, %rd2, %rd289; ld.local.f32 %f2993, [%rd290]; ld.local.f32 %f2994, [%rd290+4]; ld.local.f32 %f2995, [%rd290+8]; $L__BB184_117: ld.local.f32 %f750, [%rd4+256]; mul.ftz.f32 %f177, %f2, %f750; ld.local.f32 %f751, [%rd4+260]; mul.ftz.f32 %f178, %f2, %f751; ld.local.f32 %f752, [%rd4+264]; mul.ftz.f32 %f179, %f2, %f752; mov.f32 %f753, 0f3F800000; sub.ftz.f32 %f754, %f753, %f63; mul.ftz.f32 %f755, %f63, %f2993; fma.rn.ftz.f32 %f180, %f754, %f2977, %f755; mul.ftz.f32 %f756, %f63, %f2994; fma.rn.ftz.f32 %f181, %f754, %f2978, %f756; mul.ftz.f32 %f757, %f63, %f2995; fma.rn.ftz.f32 %f182, %f754, %f2979, %f757; mul.ftz.f32 %f758, %f180, %f177; mul.ftz.f32 %f759, %f181, %f178; mul.ftz.f32 %f760, %f182, %f179; mul.ftz.f32 %f183, %f97, %f758; mul.ftz.f32 %f184, %f97, %f759; mul.ftz.f32 %f185, %f97, %f760; add.ftz.f32 %f761, %f180, %f181; add.ftz.f32 %f762, %f182, %f761; mul.ftz.f32 %f763, %f762, 0f3EAAAAAB; abs.ftz.f32 %f764, %f763; setp.leu.ftz.f32 %p73, %f764, 0f3727C5AC; @%p73 bra $L__BB184_178; setp.le.ftz.f32 %p74, %f63, 0f3727C5AC; setp.gt.ftz.f32 %p75, %f97, 0f3727C5AC; and.pred %p76, %p74, %p75; @%p76 bra $L__BB184_171; bra.uni $L__BB184_119; $L__BB184_171: mul.ftz.f32 %f1387, %f2977, %f177; mul.ftz.f32 %f1388, %f97, %f1387; mul.ftz.f32 %f1389, %f2978, %f178; mul.ftz.f32 %f1390, %f97, %f1389; mul.ftz.f32 %f1391, %f2979, %f179; mul.ftz.f32 %f1392, %f97, %f1391; mov.f32 %f1393, 0f00000000; max.ftz.f32 %f219, %f1388, %f1393; max.ftz.f32 %f220, %f1390, %f1393; max.ftz.f32 %f221, %f1392, %f1393; add.ftz.f32 %f1394, %f219, %f220; add.ftz.f32 %f1395, %f221, %f1394; mul.ftz.f32 %f1396, %f1395, 0f3EAAAAAB; abs.ftz.f32 %f222, %f1396; setp.ltu.ftz.f32 %p141, %f222, 0f3727C5AC; mov.u64 %rd708, 0; mov.u64 %rd709, %rd708; @%p141 bra $L__BB184_176; ld.local.u32 %r42, [%rd4+252]; mov.u64 %rd708, 0; setp.eq.s32 %p142, %r42, 0; mov.u64 %rd706, %rd708; mov.u64 %rd707, %rd708; @%p142 bra $L__BB184_174; ld.local.u32 %r217, [%rd4+248]; mul.wide.s32 %rd328, %r217, 80; add.s64 %rd329, %rd1, %rd328; add.s64 %rd706, %rd329, 304; add.s64 %rd330, %rd230, %rd328; add.s64 %rd707, %rd330, 304; mov.b32 %r218, %f221; mov.b32 %r219, %f220; mov.b32 %r220, %f219; mov.u32 %r221, 0; st.local.v4.u32 [%rd329+304], {%r220, %r219, %r218, %r221}; add.s32 %r222, %r217, 1; st.local.u32 [%rd4+248], %r222; add.s32 %r223, %r42, -1; st.local.u32 [%rd4+252], %r223; $L__BB184_174: setp.eq.s64 %p143, %rd707, 0; mov.u64 %rd709, %rd708; @%p143 bra $L__BB184_176; st.local.f32 [%rd706+16], %f222; mov.u64 %rd708, %rd706; mov.u64 %rd709, %rd707; $L__BB184_176: setp.eq.s64 %p144, %rd709, 0; @%p144 bra $L__BB184_178; st.local.f32 [%rd708+20], %f55; st.local.f32 [%rd708+24], %f56; st.local.f32 [%rd708+28], %f57; st.local.f32 [%rd708+32], %f71; mov.u32 %r224, 5; st.local.u32 [%rd708+12], %r224; mov.u32 %r225, 1; st.local.u32 [%rd708+36], %r225; ld.local.u32 %r226, [%rd4+40]; or.b32 %r227, %r226, 12; st.local.u32 [%rd4+40], %r227; bra.uni $L__BB184_178; $L__BB184_119: setp.leu.ftz.f32 %p77, %f63, 0f3727C5AC; @%p77 bra $L__BB184_178; ld.local.u32 %r29, [%rd4+252]; mov.u64 %rd693, 0; setp.eq.s32 %p78, %r29, 0; mov.u64 %rd691, %rd693; mov.u64 %rd692, %rd693; @%p78 bra $L__BB184_122; ld.local.u32 %r170, [%rd4+248]; mul.wide.s32 %rd293, %r170, 80; add.s64 %rd294, %rd1, %rd293; add.s64 %rd691, %rd294, 304; add.s64 %rd295, %rd230, %rd293; add.s64 %rd692, %rd295, 304; mov.b32 %r171, %f185; mov.b32 %r172, %f184; mov.b32 %r173, %f183; mov.u32 %r174, 0; st.local.v4.u32 [%rd294+304], {%r173, %r172, %r171, %r174}; add.s32 %r175, %r170, 1; st.local.u32 [%rd4+248], %r175; add.s32 %r176, %r29, -1; st.local.u32 [%rd4+252], %r176; $L__BB184_122: setp.eq.s64 %p79, %rd692, 0; @%p79 bra $L__BB184_124; add.ftz.f32 %f765, %f183, %f184; add.ftz.f32 %f766, %f185, %f765; mul.ftz.f32 %f767, %f766, 0f3EAAAAAB; abs.ftz.f32 %f768, %f767; st.local.f32 [%rd691+16], %f768; setp.ge.ftz.f32 %p80, %f768, 0f3727C5AC; selp.b64 %rd693, %rd692, 0, %p80; $L__BB184_124: setp.eq.s64 %p81, %rd693, 0; @%p81 bra $L__BB184_178; mul.ftz.f32 %f769, %f63, %f2988; add.s64 %rd14, %rd693, 32; st.f32 [%rd693+32], %f769; mul.ftz.f32 %f770, %f63, %f2989; st.f32 [%rd693+36], %f770; mul.ftz.f32 %f771, %f63, %f2990; st.f32 [%rd693+40], %f771; st.f32 [%rd693+44], %f180; st.f32 [%rd693+48], %f181; st.f32 [%rd693+52], %f182; st.f32 [%rd693+20], %f55; st.f32 [%rd693+24], %f56; st.f32 [%rd693+28], %f57; st.f32 [%rd693+56], %f71; mov.f32 %f2998, 0f00000000; max.ftz.f32 %f773, %f2992, %f2998; mov.f32 %f774, 0f3F666666; min.ftz.f32 %f775, %f773, %f774; st.f32 [%rd693+60], %f775; setp.eq.ftz.f32 %p82, %f71, 0f7F7FFFFF; mov.u32 %r610, 0; @%p82 bra $L__BB184_133; ld.f32 %f776, [%rd14+-32]; mov.u64 %rd696, 0; mov.f32 %f777, 0f00000000; max.ftz.f32 %f186, %f776, %f777; ld.f32 %f778, [%rd14+-28]; max.ftz.f32 %f187, %f778, %f777; ld.f32 %f779, [%rd14+-24]; max.ftz.f32 %f188, %f779, %f777; add.ftz.f32 %f780, %f186, %f187; add.ftz.f32 %f781, %f188, %f780; mul.ftz.f32 %f782, %f781, 0f3EAAAAAB; abs.ftz.f32 %f189, %f782; setp.ltu.ftz.f32 %p83, %f189, 0f3727C5AC; mov.u64 %rd697, %rd696; @%p83 bra $L__BB184_131; ld.local.u32 %r30, [%rd4+252]; mov.u64 %rd696, 0; setp.eq.s32 %p84, %r30, 0; mov.u64 %rd694, %rd696; mov.u64 %rd695, %rd696; @%p84 bra $L__BB184_129; ld.local.u32 %r178, [%rd4+248]; mul.wide.s32 %rd301, %r178, 80; add.s64 %rd302, %rd1, %rd301; add.s64 %rd694, %rd302, 304; add.s64 %rd303, %rd230, %rd301; add.s64 %rd695, %rd303, 304; mov.b32 %r179, %f188; mov.b32 %r180, %f187; mov.b32 %r181, %f186; mov.u32 %r182, 0; st.local.v4.u32 [%rd302+304], {%r181, %r180, %r179, %r182}; add.s32 %r183, %r178, 1; st.local.u32 [%rd4+248], %r183; add.s32 %r184, %r30, -1; st.local.u32 [%rd4+252], %r184; $L__BB184_129: setp.eq.s64 %p85, %rd695, 0; mov.u64 %rd697, %rd696; @%p85 bra $L__BB184_131; st.local.f32 [%rd694+16], %f189; mov.u64 %rd696, %rd694; mov.u64 %rd697, %rd695; $L__BB184_131: setp.eq.s64 %p86, %rd697, 0; @%p86 bra $L__BB184_133; ld.f32 %f783, [%rd14+-12]; ld.f32 %f784, [%rd14+-8]; ld.f32 %f785, [%rd14+-4]; st.local.f32 [%rd696+20], %f783; st.local.f32 [%rd696+24], %f784; st.local.f32 [%rd696+28], %f785; ld.f32 %f786, [%rd14+24]; st.local.f32 [%rd696+32], %f786; mov.u32 %r187, 5; st.local.u32 [%rd696+12], %r187; mov.u32 %r188, 8; st.local.u32 [%rd696+36], %r188; ld.local.f32 %f787, [%rd4+24]; ld.local.f32 %f788, [%rd4+28]; mul.ftz.f32 %f789, %f788, %f784; fma.rn.ftz.f32 %f790, %f787, %f783, %f789; ld.local.f32 %f791, [%rd4+32]; fma.rn.ftz.f32 %f792, %f791, %f785, %f790; mov.f32 %f793, 0f3F800000; sub.ftz.f32 %f794, %f793, %f792; max.ftz.f32 %f796, %f794, %f777; min.ftz.f32 %f797, %f796, %f793; mul.ftz.f32 %f798, %f797, %f797; mul.ftz.f32 %f799, %f798, %f798; mul.ftz.f32 %f800, %f797, %f799; mul.ftz.f32 %f801, %f786, %f800; ld.local.f32 %f802, [%rd696+16]; mul.ftz.f32 %f803, %f802, %f801; st.local.f32 [%rd696+16], %f803; mov.u32 %r610, 12; $L__BB184_133: ld.f32 %f805, [%rd14]; setp.geu.ftz.f32 %p87, %f805, 0f322BCC77; mov.u32 %r608, 3; mov.f32 %f2996, %f2998; @%p87 bra $L__BB184_135; ld.f32 %f2996, [%rd14+-32]; mov.u32 %r191, 0; st.u32 [%rd14+-32], %r191; st.u32 [%rd14], %r191; mov.u32 %r608, 2; $L__BB184_135: ld.f32 %f807, [%rd14+4]; setp.geu.ftz.f32 %p88, %f807, 0f322BCC77; mov.f32 %f2997, %f2998; @%p88 bra $L__BB184_137; ld.f32 %f2997, [%rd14+-28]; mov.u32 %r192, 0; st.u32 [%rd14+-28], %r192; st.u32 [%rd14+4], %r192; add.s32 %r608, %r608, -1; $L__BB184_137: ld.f32 %f809, [%rd14+8]; setp.geu.ftz.f32 %p89, %f809, 0f322BCC77; @%p89 bra $L__BB184_139; ld.f32 %f2998, [%rd14+-24]; mov.u32 %r193, 0; st.u32 [%rd14+-24], %r193; st.u32 [%rd14+8], %r193; add.s32 %r608, %r608, -1; $L__BB184_139: setp.gt.s32 %p90, %r608, 2; @%p90 bra $L__BB184_155; ld.f32 %f810, [%rd14+24]; setp.neu.ftz.f32 %p91, %f810, 0f7F7FFFFF; mov.f32 %f811, 0f00000000; max.ftz.f32 %f196, %f2997, %f811; max.ftz.f32 %f197, %f2996, %f811; add.ftz.f32 %f812, %f197, %f196; max.ftz.f32 %f198, %f2998, %f811; add.ftz.f32 %f813, %f198, %f812; mul.ftz.f32 %f814, %f813, 0f3EAAAAAB; abs.ftz.f32 %f199, %f814; @%p91 bra $L__BB184_148; bra.uni $L__BB184_141; $L__BB184_148: setp.ltu.ftz.f32 %p96, %f199, 0f3727C5AC; mov.u64 %rd704, 0; mov.u64 %rd705, %rd704; @%p96 bra $L__BB184_153; ld.local.u32 %r38, [%rd4+252]; mov.u64 %rd704, 0; setp.eq.s32 %p97, %r38, 0; mov.u64 %rd702, %rd704; mov.u64 %rd703, %rd704; @%p97 bra $L__BB184_151; ld.local.u32 %r203, [%rd4+248]; mul.wide.s32 %rd319, %r203, 80; add.s64 %rd320, %rd1, %rd319; add.s64 %rd702, %rd320, 304; add.s64 %rd321, %rd230, %rd319; add.s64 %rd703, %rd321, 304; mov.b32 %r204, %f198; mov.b32 %r205, %f196; mov.b32 %r206, %f197; mov.u32 %r207, 0; st.local.v4.u32 [%rd320+304], {%r206, %r205, %r204, %r207}; add.s32 %r208, %r203, 1; st.local.u32 [%rd4+248], %r208; add.s32 %r209, %r38, -1; st.local.u32 [%rd4+252], %r209; $L__BB184_151: setp.eq.s64 %p98, %rd703, 0; mov.u64 %rd705, %rd704; @%p98 bra $L__BB184_153; st.local.f32 [%rd702+16], %f199; mov.u64 %rd704, %rd702; mov.u64 %rd705, %rd703; $L__BB184_153: setp.eq.s64 %p99, %rd705, 0; @%p99 bra $L__BB184_155; ld.f32 %f818, [%rd14+-12]; ld.f32 %f819, [%rd14+-8]; ld.f32 %f820, [%rd14+-4]; st.local.f32 [%rd704+20], %f818; st.local.f32 [%rd704+24], %f819; st.local.f32 [%rd704+28], %f820; ld.f32 %f821, [%rd14+24]; st.local.f32 [%rd704+32], %f821; mov.u32 %r211, 5; st.local.u32 [%rd704+12], %r211; mov.u32 %r212, 2; st.local.u32 [%rd704+36], %r212; mov.u32 %r610, 12; bra.uni $L__BB184_155; $L__BB184_360: and.b32 %r406, %r105, 7; setp.eq.s32 %p311, %r406, 0; mov.f32 %f3033, 0f00000000; @%p311 bra $L__BB184_372; setp.eq.s32 %p312, %r105, 4; selp.b32 %r408, %r391, 0, %p312; add.s32 %r409, %r108, %r408; ld.const.u64 %rd513, [kernel_params+2320]; cvta.to.global.u64 %rd514, %rd513; mul.wide.s32 %rd515, %r409, 4; add.s64 %rd516, %rd514, %rd515; ld.global.f32 %f3033, [%rd516]; bra.uni $L__BB184_372; $L__BB184_358: mad.lo.s32 %r405, %r391, 3, %r108; ld.const.u64 %rd510, [kernel_params+2320]; cvta.to.global.u64 %rd511, %rd510; mul.wide.s32 %rd512, %r405, 4; add.s64 %rd755, %rd511, %rd512; add.s64 %rd756, %rd755, 4; add.s64 %rd757, %rd755, 8; $L__BB184_359: ld.local.f32 %f2504, [%rd4+56]; ld.global.f32 %f2505, [%rd756]; ld.local.f32 %f2506, [%rd4+60]; ld.global.f32 %f2507, [%rd757]; mul.ftz.f32 %f2508, %f2507, %f2506; fma.rn.ftz.f32 %f2509, %f2505, %f2504, %f2508; mov.f32 %f2510, 0f3F800000; sub.ftz.f32 %f2511, %f2510, %f2504; sub.ftz.f32 %f2512, %f2511, %f2506; ld.global.f32 %f2513, [%rd755]; fma.rn.ftz.f32 %f3033, %f2513, %f2512, %f2509; $L__BB184_372: mov.f32 %f2523, 0f00000000; max.ftz.f32 %f395, %f381, %f2523; max.ftz.f32 %f396, %f382, %f2523; add.ftz.f32 %f2524, %f396, %f395; max.ftz.f32 %f397, %f380, %f2523; add.ftz.f32 %f2525, %f397, %f2524; mul.ftz.f32 %f2526, %f2525, 0f3EAAAAAB; abs.ftz.f32 %f398, %f2526; setp.ltu.ftz.f32 %p321, %f398, 0f3727C5AC; mov.u64 %rd760, 0; mov.u64 %rd761, %rd760; @%p321 bra $L__BB184_377; ld.local.u32 %r113, [%rd4+252]; mov.u64 %rd760, 0; setp.eq.s32 %p322, %r113, 0; mov.u64 %rd758, %rd760; mov.u64 %rd759, %rd760; @%p322 bra $L__BB184_375; ld.local.u32 %r422, [%rd4+248]; mul.wide.s32 %rd539, %r422, 80; add.s64 %rd540, %rd1, %rd539; add.s64 %rd758, %rd540, 304; add.s64 %rd541, %rd230, %rd539; add.s64 %rd759, %rd541, 304; mov.b32 %r423, %f397; mov.b32 %r424, %f395; mov.b32 %r425, %f396; mov.u32 %r426, 0; st.local.v4.u32 [%rd540+304], {%r425, %r424, %r423, %r426}; add.s32 %r427, %r422, 1; st.local.u32 [%rd4+248], %r427; add.s32 %r428, %r113, -1; st.local.u32 [%rd4+252], %r428; $L__BB184_375: setp.eq.s64 %p323, %rd759, 0; mov.u64 %rd761, %rd760; @%p323 bra $L__BB184_377; st.local.f32 [%rd758+16], %f398; mov.u64 %rd760, %rd758; mov.u64 %rd761, %rd759; $L__BB184_377: add.s32 %r629, %r134, 4; setp.eq.s64 %p324, %rd761, 0; @%p324 bra $L__BB184_499; ld.local.u32 %r115, [%rd4+252]; setp.lt.s32 %p325, %r115, 1; @%p325 bra $L__BB184_380; add.s32 %r429, %r115, -1; st.local.u32 [%rd4+252], %r429; ld.local.s32 %rd544, [%rd4+248]; cvt.s64.s32 %rd545, %r429; add.s64 %rd546, %rd544, %rd545; mul.lo.s64 %rd547, %rd546, 80; add.s64 %rd548, %rd230, %rd547; add.s64 %rd763, %rd548, 304; add.s64 %rd549, %rd1, %rd547; add.s64 %rd762, %rd549, 304; bra.uni $L__BB184_381; $L__BB184_380: ld.local.u32 %r430, [%rd4+248]; mov.u64 %rd762, 0; add.s32 %r431, %r430, -1; st.local.u32 [%rd4+248], %r431; add.s32 %r432, %r115, 1; st.local.u32 [%rd4+252], %r432; mov.u64 %rd763, %rd762; $L__BB184_381: setp.eq.s64 %p326, %rd763, 0; @%p326 bra $L__BB184_499; setp.eq.s32 %p327, %r86, 255; @%p327 bra $L__BB184_384; bra.uni $L__BB184_383; $L__BB184_384: mov.b32 %f3034, %r369; bra.uni $L__BB184_385; $L__BB184_383: mul.wide.u32 %rd552, %r86, 4; add.s64 %rd553, %rd2, %rd552; ld.local.f32 %f3034, [%rd553]; $L__BB184_385: add.ftz.f32 %f2527, %f3033, 0fBF000000; add.ftz.f32 %f402, %f2527, %f2527; fma.rn.ftz.f32 %f2528, %f402, %f3034, 0f3F800000; mul.ftz.f32 %f403, %f60, %f2528; mul.ftz.f32 %f404, %f63, %f2528; setp.eq.s32 %p328, %r81, 255; @%p328 bra $L__BB184_387; bra.uni $L__BB184_386; $L__BB184_387: mov.b32 %f3035, %r363; bra.uni $L__BB184_388; $L__BB184_386: mul.wide.u32 %rd554, %r81, 4; add.s64 %rd555, %rd2, %rd554; ld.local.f32 %f3035, [%rd555]; $L__BB184_388: max.ftz.f32 %f2530, %f3035, %f2523; mov.f32 %f2531, 0f3F800000; min.ftz.f32 %f2532, %f2530, %f2531; sub.ftz.f32 %f408, %f2531, %f2532; add.s64 %rd153, %rd760, 20; st.local.f32 [%rd760+20], %f55; st.local.f32 [%rd760+24], %f56; st.local.f32 [%rd760+28], %f57; st.local.f32 [%rd760+44], %f403; st.local.v2.f32 [%rd760+48], {%f404, %f3031}; st.local.v2.f32 [%rd760+56], {%f3032, %f408}; st.local.u64 [%rd760+64], %rd763; shr.u32 %r433, %r136, 24; cvt.u16.u32 %rs15, %r433; setp.eq.s16 %p329, %rs15, 0; @%p329 bra $L__BB184_402; setp.eq.s16 %p330, %rs15, 1; @%p330 bra $L__BB184_392; setp.ne.s16 %p331, %rs15, 2; @%p331 bra $L__BB184_403; shr.u32 %r434, %r362, 24; mul.wide.u32 %rd556, %r434, 4; add.s64 %rd557, %rd2, %rd556; ld.local.f32 %f2533, [%rd557+8]; ld.local.f32 %f2534, [%rd557+4]; ld.local.f32 %f2535, [%rd557]; st.local.v2.f32 [%rd153+12], {%f2535, %f2534}; st.local.f32 [%rd153+20], %f2533; bra.uni $L__BB184_404; $L__BB184_402: shr.u32 %r436, %r136, 14; cvt.u64.u32 %rd567, %r436; and.b64 %rd568, %rd567, 1020; add.s64 %rd569, %rd2, %rd568; ld.local.f32 %f2576, [%rd569]; lg2.approx.ftz.f32 %f2577, %f2576; mul.ftz.f32 %f2578, %f2577, 0f3F317218; ld.local.f32 %f2579, [%rd569+4]; lg2.approx.ftz.f32 %f2580, %f2579; mul.ftz.f32 %f2581, %f2580, 0f3F317218; ld.local.f32 %f2582, [%rd569+8]; lg2.approx.ftz.f32 %f2583, %f2582; mul.ftz.f32 %f2584, %f2583, 0f3F317218; fma.rn.ftz.f32 %f2585, %f404, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2586, %f404, %f2585, 0fC12BAE14; fma.rn.ftz.f32 %f2587, %f404, %f2586, 0f40220C4A; fma.rn.ftz.f32 %f2588, %f404, %f2587, 0fBE5C28F6; fma.rn.ftz.f32 %f2589, %f404, %f2588, 0f40BF020C; div.approx.ftz.f32 %f2590, %f2578, %f2589; div.approx.ftz.f32 %f2591, %f2581, %f2589; div.approx.ftz.f32 %f2592, %f2584, %f2589; mul.ftz.f32 %f2593, %f2592, %f2592; mul.ftz.f32 %f2594, %f2591, %f2591; mul.ftz.f32 %f2595, %f2590, %f2590; st.local.v2.f32 [%rd153+12], {%f2595, %f2594}; st.local.f32 [%rd153+20], %f2593; bra.uni $L__BB184_404; $L__BB184_392: setp.eq.s32 %p332, %r82, 255; @%p332 bra $L__BB184_394; bra.uni $L__BB184_393; $L__BB184_394: mov.b32 %f3036, %r364; bra.uni $L__BB184_395; $L__BB184_403: mov.f32 %f2596, 0f3F172384; mov.f32 %f2597, 0f3E8D729D; st.local.v2.f32 [%rd153+12], {%f2597, %f2596}; mov.u32 %r437, 1069964113; st.local.u32 [%rd153+20], %r437; bra.uni $L__BB184_404; $L__BB184_393: mul.wide.u32 %rd558, %r82, 4; add.s64 %rd559, %rd2, %rd558; ld.local.f32 %f3036, [%rd559]; $L__BB184_395: setp.eq.s32 %p333, %r83, 255; @%p333 bra $L__BB184_397; bra.uni $L__BB184_396; $L__BB184_397: mov.b32 %f3037, %r365; bra.uni $L__BB184_398; $L__BB184_396: mul.wide.u32 %rd560, %r83, 4; add.s64 %rd561, %rd2, %rd560; ld.local.f32 %f3037, [%rd561]; $L__BB184_398: setp.eq.s32 %p334, %r85, 255; @%p334 bra $L__BB184_400; bra.uni $L__BB184_399; $L__BB184_400: mov.b32 %f3038, %r368; bra.uni $L__BB184_401; $L__BB184_399: mul.wide.u32 %rd562, %r85, 4; add.s64 %rd563, %rd2, %rd562; ld.local.f32 %f3038, [%rd563]; $L__BB184_401: max.ftz.f32 %f2537, %f3038, %f2523; min.ftz.f32 %f2539, %f2537, %f2531; fma.rn.ftz.f32 %f2540, %f402, %f2539, 0f3F800000; mul.ftz.f32 %f2541, %f3036, %f2540; sub.ftz.f32 %f2542, %f2531, %f2541; mov.f32 %f2543, 0f38D1B717; max.ftz.f32 %f2544, %f2542, %f2543; lg2.approx.ftz.f32 %f2545, %f2544; mul.ftz.f32 %f2546, %f2545, 0fBF317218; sub.ftz.f32 %f2547, %f2531, %f3037; mul.ftz.f32 %f2548, %f2547, %f2546; mul.ftz.f32 %f2549, %f3037, %f2546; mul.ftz.f32 %f2550, %f2549, 0f3EAF9DB2; mul.ftz.f32 %f2551, %f2549, 0f3F3BA5E3; mul.ftz.f32 %f2552, %f2549, 0f3FF645A2; fma.rn.ftz.f32 %f2553, %f2548, 0f3F018937, %f2550; fma.rn.ftz.f32 %f2554, %f2548, 0f3F574BC7, %f2551; fma.rn.ftz.f32 %f2555, %f2548, 0f3FD39581, %f2552; shl.b32 %r435, %r366, 2; cvt.u64.u32 %rd564, %r435; and.b64 %rd565, %rd564, 1020; add.s64 %rd566, %rd2, %rd565; ld.local.f32 %f2556, [%rd566]; lg2.approx.ftz.f32 %f2557, %f2556; mul.ftz.f32 %f2558, %f2557, 0f3F317218; ld.local.f32 %f2559, [%rd566+4]; lg2.approx.ftz.f32 %f2560, %f2559; mul.ftz.f32 %f2561, %f2560, 0f3F317218; ld.local.f32 %f2562, [%rd566+8]; lg2.approx.ftz.f32 %f2563, %f2562; mul.ftz.f32 %f2564, %f2563, 0f3F317218; fma.rn.ftz.f32 %f2565, %f404, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2566, %f404, %f2565, 0fC12BAE14; fma.rn.ftz.f32 %f2567, %f404, %f2566, 0f40220C4A; fma.rn.ftz.f32 %f2568, %f404, %f2567, 0fBE5C28F6; fma.rn.ftz.f32 %f2569, %f404, %f2568, 0f40BF020C; div.approx.ftz.f32 %f2570, %f2558, %f2569; div.approx.ftz.f32 %f2571, %f2561, %f2569; div.approx.ftz.f32 %f2572, %f2564, %f2569; fma.rn.ftz.f32 %f2573, %f2572, %f2572, %f2555; fma.rn.ftz.f32 %f2574, %f2571, %f2571, %f2554; fma.rn.ftz.f32 %f2575, %f2570, %f2570, %f2553; st.local.v2.f32 [%rd153+12], {%f2575, %f2574}; st.local.f32 [%rd153+20], %f2573; $L__BB184_404: mov.u32 %r438, 29; st.local.u32 [%rd153+-8], %r438; mov.f32 %f2598, 0f3A83126F; max.ftz.f32 %f2599, %f403, %f2598; min.ftz.f32 %f2601, %f2599, %f2531; max.ftz.f32 %f2602, %f404, %f2598; min.ftz.f32 %f2603, %f2602, %f2531; mul.ftz.f32 %f2604, %f2601, %f408; max.ftz.f32 %f2605, %f2604, %f2598; min.ftz.f32 %f2606, %f2605, %f2531; mul.ftz.f32 %f2607, %f2601, %f2601; mul.ftz.f32 %f2608, %f2607, 0f3F4FDF3B; fma.rn.ftz.f32 %f2609, %f2601, 0f3F39DB23, %f2608; mul.ftz.f32 %f2610, %f2607, %f2607; mul.ftz.f32 %f2611, %f2601, %f2610; mul.ftz.f32 %f2612, %f2611, %f2611; mul.ftz.f32 %f2613, %f2612, %f2612; fma.rn.ftz.f32 %f2614, %f2613, 0f406CCCCD, %f2609; mul.ftz.f32 %f2615, %f2614, %f2614; st.local.f32 [%rd153+24], %f2615; mul.ftz.f32 %f2616, %f2603, %f2603; mul.ftz.f32 %f2617, %f2616, 0f3F98D4FE; fma.rn.ftz.f32 %f2618, %f2603, 0f3E87AE14, %f2617; mul.ftz.f32 %f2619, %f2616, %f2616; mul.ftz.f32 %f2620, %f2603, %f2619; mul.ftz.f32 %f2621, %f2620, %f2620; mul.ftz.f32 %f2622, %f2603, %f2621; mul.ftz.f32 %f2623, %f2622, %f2622; fma.rn.ftz.f32 %f2624, %f2623, 0f40ABE76D, %f2618; mul.ftz.f32 %f2625, %f2624, 0f3F206C99; st.local.f32 [%rd153+28], %f2625; mul.ftz.f32 %f2626, %f2606, %f2606; mul.ftz.f32 %f2627, %f2626, 0f3F4FDF3B; fma.rn.ftz.f32 %f2628, %f2606, 0f3F39DB23, %f2627; mul.ftz.f32 %f2629, %f2626, %f2626; mul.ftz.f32 %f2630, %f2606, %f2629; mul.ftz.f32 %f2631, %f2630, %f2630; mul.ftz.f32 %f2632, %f2631, %f2631; fma.rn.ftz.f32 %f2633, %f2632, 0f406CCCCD, %f2628; mul.ftz.f32 %f2634, %f2633, %f2633; st.local.f32 [%rd153+40], %f2634; ld.local.f32 %f3039, [%rd4+104]; ld.local.f32 %f3040, [%rd4+108]; mul.ftz.f32 %f2635, %f3040, %f3040; fma.rn.ftz.f32 %f2636, %f3039, %f3039, %f2635; ld.local.f32 %f3041, [%rd4+112]; fma.rn.ftz.f32 %f2637, %f3041, %f3041, %f2636; sqrt.approx.ftz.f32 %f421, %f2637; setp.neu.ftz.f32 %p335, %f421, 0f00000000; @%p335 bra $L__BB184_405; bra.uni $L__BB184_406; $L__BB184_405: rcp.approx.ftz.f32 %f2638, %f421; mul.ftz.f32 %f3041, %f3041, %f2638; mul.ftz.f32 %f3040, %f3040, %f2638; mul.ftz.f32 %f3039, %f3039, %f2638; $L__BB184_406: ld.local.f32 %f2639, [%rd4+32]; mul.ftz.f32 %f2640, %f3040, %f2639; ld.local.f32 %f2641, [%rd4+28]; mul.ftz.f32 %f2642, %f3041, %f2641; sub.ftz.f32 %f3042, %f2640, %f2642; ld.local.f32 %f2643, [%rd4+24]; mul.ftz.f32 %f2644, %f3041, %f2643; mul.ftz.f32 %f2645, %f3039, %f2639; sub.ftz.f32 %f3043, %f2644, %f2645; mul.ftz.f32 %f2646, %f3039, %f2641; mul.ftz.f32 %f2647, %f3040, %f2643; sub.ftz.f32 %f3044, %f2646, %f2647; mul.ftz.f32 %f2648, %f3043, %f3043; fma.rn.ftz.f32 %f2649, %f3042, %f3042, %f2648; fma.rn.ftz.f32 %f2650, %f3044, %f3044, %f2649; sqrt.approx.ftz.f32 %f437, %f2650; setp.neu.ftz.f32 %p336, %f437, 0f00000000; @%p336 bra $L__BB184_407; bra.uni $L__BB184_408; $L__BB184_407: rcp.approx.ftz.f32 %f2651, %f437; mul.ftz.f32 %f3044, %f3044, %f2651; mul.ftz.f32 %f3043, %f3043, %f2651; mul.ftz.f32 %f3042, %f3042, %f2651; $L__BB184_408: mul.ftz.f32 %f2652, %f3040, %f3044; mul.ftz.f32 %f2653, %f3041, %f3043; sub.ftz.f32 %f3045, %f2652, %f2653; mul.ftz.f32 %f2654, %f3041, %f3042; mul.ftz.f32 %f2655, %f3039, %f3044; sub.ftz.f32 %f3046, %f2654, %f2655; mul.ftz.f32 %f2656, %f3039, %f3043; mul.ftz.f32 %f2657, %f3040, %f3042; sub.ftz.f32 %f3047, %f2656, %f2657; mul.ftz.f32 %f2658, %f3046, %f3046; fma.rn.ftz.f32 %f2659, %f3045, %f3045, %f2658; fma.rn.ftz.f32 %f2660, %f3047, %f3047, %f2659; sqrt.approx.ftz.f32 %f453, %f2660; setp.neu.ftz.f32 %p337, %f453, 0f00000000; @%p337 bra $L__BB184_409; bra.uni $L__BB184_410; $L__BB184_409: rcp.approx.ftz.f32 %f2661, %f453; mul.ftz.f32 %f3047, %f2661, %f3047; mul.ftz.f32 %f3046, %f2661, %f3046; mul.ftz.f32 %f3045, %f2661, %f3045; $L__BB184_410: ld.local.u8 %rs16, [%rd4+52]; and.b16 %rs17, %rs16, 4; setp.eq.s16 %p338, %rs17, 0; @%p338 bra $L__BB184_412; bra.uni $L__BB184_411; $L__BB184_412: ld.local.f32 %f2663, [%rd4+16]; mul.ftz.f32 %f2664, %f3041, %f2663; ld.local.f32 %f2665, [%rd4+20]; mul.ftz.f32 %f2666, %f3040, %f2665; sub.ftz.f32 %f2667, %f2664, %f2666; mul.ftz.f32 %f2668, %f3039, %f2665; ld.local.f32 %f2669, [%rd4+12]; mul.ftz.f32 %f2670, %f3041, %f2669; sub.ftz.f32 %f2671, %f2668, %f2670; mul.ftz.f32 %f2672, %f3040, %f2669; mul.ftz.f32 %f2673, %f3039, %f2663; sub.ftz.f32 %f2674, %f2672, %f2673; mul.ftz.f32 %f2675, %f3046, %f2671; fma.rn.ftz.f32 %f2676, %f3045, %f2667, %f2675; fma.rn.ftz.f32 %f3048, %f3047, %f2674, %f2676; bra.uni $L__BB184_413; $L__BB184_411: ld.local.f32 %f2662, [%rd4+60]; neg.ftz.f32 %f3048, %f2662; $L__BB184_413: st.local.v4.f32 [%rd762], {%f3042, %f3043, %f3044, %f3048}; ld.local.u32 %r439, [%rd4+40]; or.b32 %r440, %r439, 1036; st.local.u32 [%rd4+40], %r440; bra.uni $L__BB184_499; $L__BB184_141: setp.ltu.ftz.f32 %p92, %f199, 0f3727C5AC; mov.u64 %rd700, 0; mov.u64 %rd701, %rd700; @%p92 bra $L__BB184_146; ld.local.u32 %r37, [%rd4+252]; mov.u64 %rd700, 0; setp.eq.s32 %p93, %r37, 0; mov.u64 %rd698, %rd700; mov.u64 %rd699, %rd700; @%p93 bra $L__BB184_144; ld.local.u32 %r194, [%rd4+248]; mul.wide.s32 %rd310, %r194, 80; add.s64 %rd311, %rd1, %rd310; add.s64 %rd698, %rd311, 304; add.s64 %rd312, %rd230, %rd310; add.s64 %rd699, %rd312, 304; mov.b32 %r195, %f198; mov.b32 %r196, %f196; mov.b32 %r197, %f197; mov.u32 %r198, 0; st.local.v4.u32 [%rd311+304], {%r197, %r196, %r195, %r198}; add.s32 %r199, %r194, 1; st.local.u32 [%rd4+248], %r199; add.s32 %r200, %r37, -1; st.local.u32 [%rd4+252], %r200; $L__BB184_144: setp.eq.s64 %p94, %rd699, 0; mov.u64 %rd701, %rd700; @%p94 bra $L__BB184_146; st.local.f32 [%rd698+16], %f199; mov.u64 %rd700, %rd698; mov.u64 %rd701, %rd699; $L__BB184_146: setp.eq.s64 %p95, %rd701, 0; @%p95 bra $L__BB184_155; ld.f32 %f815, [%rd14+-12]; ld.f32 %f816, [%rd14+-8]; ld.f32 %f817, [%rd14+-4]; st.local.f32 [%rd700+20], %f815; st.local.f32 [%rd700+24], %f816; st.local.f32 [%rd700+28], %f817; mov.u32 %r202, 2; st.local.u32 [%rd700+12], %r202; mov.u32 %r610, 12; $L__BB184_155: st.u32 [%rd14+-20], %r141; setp.gt.s32 %p100, %r608, 0; @%p100 bra $L__BB184_157; bra.uni $L__BB184_156; $L__BB184_157: ld.f32 %f822, [%rd14+-28]; ld.f32 %f823, [%rd14+-32]; add.ftz.f32 %f824, %f823, %f822; ld.f32 %f825, [%rd14+-24]; add.ftz.f32 %f826, %f825, %f824; mul.ftz.f32 %f827, %f826, 0f3EAAAAAB; abs.ftz.f32 %f828, %f827; cvt.rn.f32.s32 %f829, %r608; mul.ftz.f32 %f830, %f828, %f829; st.f32 [%rd14+-16], %f830; or.b32 %r214, %r141, 2; setp.eq.s32 %p101, %r214, 34; @%p101 bra $L__BB184_168; bra.uni $L__BB184_158; $L__BB184_168: ld.f32 %f1384, [%rd14]; mul.ftz.f32 %f3002, %f1384, 0f3DA2F983; ld.f32 %f1385, [%rd14+4]; mul.ftz.f32 %f3003, %f1385, 0f3DA2F983; ld.f32 %f1386, [%rd14+8]; mul.ftz.f32 %f3004, %f1386, 0f3DA2F983; bra.uni $L__BB184_169; $L__BB184_156: mov.u32 %r213, 0; st.u32 [%rd14+-16], %r213; bra.uni $L__BB184_170; $L__BB184_158: mov.f32 %f832, 0f3F8147AE; max.ftz.f32 %f833, %f2991, %f832; mov.f32 %f834, 0f40733333; min.ftz.f32 %f835, %f833, %f834; rcp.approx.ftz.f32 %f836, %f835; mov.f32 %f837, 0f3F800000; fma.rn.ftz.f32 %f838, %f836, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f839, %f836, %f838, 0f3F2B020C; fma.rn.ftz.f32 %f840, %f835, 0f3D8240B8, %f839; add.ftz.f32 %f841, %f840, 0f3F800000; mul.ftz.f32 %f842, %f841, 0f3FAAAAAB; sub.ftz.f32 %f843, %f837, %f840; div.approx.ftz.f32 %f200, %f842, %f843; ld.f32 %f201, [%rd14+12]; setp.lt.ftz.f32 %p102, %f201, 0f38D1B717; mov.f32 %f2999, 0f00000000; @%p102 bra $L__BB184_161; setp.ge.ftz.f32 %p103, %f201, 0f3F7EB852; mov.f32 %f2999, 0f3F7FFFEF; @%p103 bra $L__BB184_161; sub.ftz.f32 %f846, %f837, 0f3F000000; mul.ftz.f32 %f847, %f846, 0f40400000; sqrt.approx.ftz.f32 %f848, %f847; mul.ftz.f32 %f849, %f200, %f848; mul.ftz.f32 %f850, %f849, 0fBFB8AA3B; ex2.approx.ftz.f32 %f851, %f850; add.ftz.f32 %f852, %f851, 0f3F800000; mul.ftz.f32 %f853, %f852, 0f3E800000; mul.ftz.f32 %f854, %f848, 0fBFB8AA3B; ex2.approx.ftz.f32 %f855, %f854; mul.ftz.f32 %f856, %f855, %f853; setp.lt.ftz.f32 %p104, %f856, %f201; selp.f32 %f857, 0f3F800000, 0f3F000000, %p104; selp.f32 %f858, 0f3F000000, 0f00000000, %p104; selp.f32 %f859, 0f3F400000, 0f3E800000, %p104; sub.ftz.f32 %f860, %f837, %f859; mul.ftz.f32 %f861, %f860, 0f40400000; sqrt.approx.ftz.f32 %f862, %f861; mul.ftz.f32 %f863, %f859, 0f3F000000; mul.ftz.f32 %f864, %f200, %f862; mul.ftz.f32 %f865, %f864, 0fBFB8AA3B; ex2.approx.ftz.f32 %f866, %f865; add.ftz.f32 %f867, %f866, 0f3F800000; mul.ftz.f32 %f868, %f867, %f863; mul.ftz.f32 %f869, %f862, 0fBFB8AA3B; ex2.approx.ftz.f32 %f870, %f869; mul.ftz.f32 %f871, %f870, %f868; setp.lt.ftz.f32 %p105, %f871, %f201; selp.f32 %f872, %f857, %f859, %p105; selp.f32 %f873, %f859, %f858, %p105; add.ftz.f32 %f874, %f872, %f873; mul.ftz.f32 %f875, %f874, 0f3F000000; sub.ftz.f32 %f876, %f837, %f875; mul.ftz.f32 %f877, %f876, 0f40400000; sqrt.approx.ftz.f32 %f878, %f877; mul.ftz.f32 %f879, %f875, 0f3F000000; mul.ftz.f32 %f880, %f200, %f878; mul.ftz.f32 %f881, %f880, 0fBFB8AA3B; ex2.approx.ftz.f32 %f882, %f881; add.ftz.f32 %f883, %f882, 0f3F800000; mul.ftz.f32 %f884, %f883, %f879; mul.ftz.f32 %f885, %f878, 0fBFB8AA3B; ex2.approx.ftz.f32 %f886, %f885; mul.ftz.f32 %f887, %f886, %f884; setp.lt.ftz.f32 %p106, %f887, %f201; selp.f32 %f888, %f872, %f875, %p106; selp.f32 %f889, %f875, %f873, %p106; add.ftz.f32 %f890, %f888, %f889; mul.ftz.f32 %f891, %f890, 0f3F000000; sub.ftz.f32 %f892, %f837, %f891; mul.ftz.f32 %f893, %f892, 0f40400000; sqrt.approx.ftz.f32 %f894, %f893; mul.ftz.f32 %f895, %f891, 0f3F000000; mul.ftz.f32 %f896, %f200, %f894; mul.ftz.f32 %f897, %f896, 0fBFB8AA3B; ex2.approx.ftz.f32 %f898, %f897; add.ftz.f32 %f899, %f898, 0f3F800000; mul.ftz.f32 %f900, %f899, %f895; mul.ftz.f32 %f901, %f894, 0fBFB8AA3B; ex2.approx.ftz.f32 %f902, %f901; mul.ftz.f32 %f903, %f902, %f900; setp.lt.ftz.f32 %p107, %f903, %f201; selp.f32 %f904, %f888, %f891, %p107; selp.f32 %f905, %f891, %f889, %p107; add.ftz.f32 %f906, %f904, %f905; mul.ftz.f32 %f907, %f906, 0f3F000000; sub.ftz.f32 %f908, %f837, %f907; mul.ftz.f32 %f909, %f908, 0f40400000; sqrt.approx.ftz.f32 %f910, %f909; mul.ftz.f32 %f911, %f907, 0f3F000000; mul.ftz.f32 %f912, %f200, %f910; mul.ftz.f32 %f913, %f912, 0fBFB8AA3B; ex2.approx.ftz.f32 %f914, %f913; add.ftz.f32 %f915, %f914, 0f3F800000; mul.ftz.f32 %f916, %f915, %f911; mul.ftz.f32 %f917, %f910, 0fBFB8AA3B; ex2.approx.ftz.f32 %f918, %f917; mul.ftz.f32 %f919, %f918, %f916; setp.lt.ftz.f32 %p108, %f919, %f201; selp.f32 %f920, %f904, %f907, %p108; selp.f32 %f921, %f907, %f905, %p108; add.ftz.f32 %f922, %f920, %f921; mul.ftz.f32 %f923, %f922, 0f3F000000; sub.ftz.f32 %f924, %f837, %f923; mul.ftz.f32 %f925, %f924, 0f40400000; sqrt.approx.ftz.f32 %f926, %f925; mul.ftz.f32 %f927, %f923, 0f3F000000; mul.ftz.f32 %f928, %f200, %f926; mul.ftz.f32 %f929, %f928, 0fBFB8AA3B; ex2.approx.ftz.f32 %f930, %f929; add.ftz.f32 %f931, %f930, 0f3F800000; mul.ftz.f32 %f932, %f931, %f927; mul.ftz.f32 %f933, %f926, 0fBFB8AA3B; ex2.approx.ftz.f32 %f934, %f933; mul.ftz.f32 %f935, %f934, %f932; setp.lt.ftz.f32 %p109, %f935, %f201; selp.f32 %f936, %f920, %f923, %p109; selp.f32 %f937, %f923, %f921, %p109; add.ftz.f32 %f938, %f936, %f937; mul.ftz.f32 %f939, %f938, 0f3F000000; sub.ftz.f32 %f940, %f837, %f939; mul.ftz.f32 %f941, %f940, 0f40400000; sqrt.approx.ftz.f32 %f942, %f941; mul.ftz.f32 %f943, %f939, 0f3F000000; mul.ftz.f32 %f944, %f200, %f942; mul.ftz.f32 %f945, %f944, 0fBFB8AA3B; ex2.approx.ftz.f32 %f946, %f945; add.ftz.f32 %f947, %f946, 0f3F800000; mul.ftz.f32 %f948, %f947, %f943; mul.ftz.f32 %f949, %f942, 0fBFB8AA3B; ex2.approx.ftz.f32 %f950, %f949; mul.ftz.f32 %f951, %f950, %f948; setp.lt.ftz.f32 %p110, %f951, %f201; selp.f32 %f952, %f936, %f939, %p110; selp.f32 %f953, %f939, %f937, %p110; add.ftz.f32 %f954, %f952, %f953; mul.ftz.f32 %f955, %f954, 0f3F000000; sub.ftz.f32 %f956, %f837, %f955; mul.ftz.f32 %f957, %f956, 0f40400000; sqrt.approx.ftz.f32 %f958, %f957; mul.ftz.f32 %f959, %f955, 0f3F000000; mul.ftz.f32 %f960, %f200, %f958; mul.ftz.f32 %f961, %f960, 0fBFB8AA3B; ex2.approx.ftz.f32 %f962, %f961; add.ftz.f32 %f963, %f962, 0f3F800000; mul.ftz.f32 %f964, %f963, %f959; mul.ftz.f32 %f965, %f958, 0fBFB8AA3B; ex2.approx.ftz.f32 %f966, %f965; mul.ftz.f32 %f967, %f966, %f964; setp.lt.ftz.f32 %p111, %f967, %f201; selp.f32 %f968, %f952, %f955, %p111; selp.f32 %f969, %f955, %f953, %p111; add.ftz.f32 %f970, %f968, %f969; mul.ftz.f32 %f971, %f970, 0f3F000000; sub.ftz.f32 %f972, %f837, %f971; mul.ftz.f32 %f973, %f972, 0f40400000; sqrt.approx.ftz.f32 %f974, %f973; mul.ftz.f32 %f975, %f971, 0f3F000000; mul.ftz.f32 %f976, %f200, %f974; mul.ftz.f32 %f977, %f976, 0fBFB8AA3B; ex2.approx.ftz.f32 %f978, %f977; add.ftz.f32 %f979, %f978, 0f3F800000; mul.ftz.f32 %f980, %f979, %f975; mul.ftz.f32 %f981, %f974, 0fBFB8AA3B; ex2.approx.ftz.f32 %f982, %f981; mul.ftz.f32 %f983, %f982, %f980; setp.lt.ftz.f32 %p112, %f983, %f201; selp.f32 %f984, %f968, %f971, %p112; selp.f32 %f985, %f971, %f969, %p112; add.ftz.f32 %f986, %f984, %f985; mul.ftz.f32 %f987, %f986, 0f3F000000; sub.ftz.f32 %f988, %f837, %f987; mul.ftz.f32 %f989, %f988, 0f40400000; sqrt.approx.ftz.f32 %f990, %f989; mul.ftz.f32 %f991, %f987, 0f3F000000; mul.ftz.f32 %f992, %f200, %f990; mul.ftz.f32 %f993, %f992, 0fBFB8AA3B; ex2.approx.ftz.f32 %f994, %f993; add.ftz.f32 %f995, %f994, 0f3F800000; mul.ftz.f32 %f996, %f995, %f991; mul.ftz.f32 %f997, %f990, 0fBFB8AA3B; ex2.approx.ftz.f32 %f998, %f997; mul.ftz.f32 %f999, %f998, %f996; setp.lt.ftz.f32 %p113, %f999, %f201; selp.f32 %f1000, %f984, %f987, %p113; selp.f32 %f1001, %f987, %f985, %p113; add.ftz.f32 %f1002, %f1000, %f1001; mul.ftz.f32 %f1003, %f1002, 0f3F000000; sub.ftz.f32 %f1004, %f837, %f1003; mul.ftz.f32 %f1005, %f1004, 0f40400000; sqrt.approx.ftz.f32 %f1006, %f1005; mul.ftz.f32 %f1007, %f1003, 0f3F000000; mul.ftz.f32 %f1008, %f200, %f1006; mul.ftz.f32 %f1009, %f1008, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1010, %f1009; add.ftz.f32 %f1011, %f1010, 0f3F800000; mul.ftz.f32 %f1012, %f1011, %f1007; mul.ftz.f32 %f1013, %f1006, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1014, %f1013; mul.ftz.f32 %f1015, %f1014, %f1012; setp.lt.ftz.f32 %p114, %f1015, %f201; selp.f32 %f1016, %f1000, %f1003, %p114; selp.f32 %f1017, %f1003, %f1001, %p114; add.ftz.f32 %f1018, %f1016, %f1017; mul.ftz.f32 %f2999, %f1018, 0f3F000000; $L__BB184_161: mov.f32 %f3000, 0f00000000; ld.f32 %f204, [%rd14+16]; setp.lt.ftz.f32 %p115, %f204, 0f38D1B717; @%p115 bra $L__BB184_164; setp.ge.ftz.f32 %p116, %f204, 0f3F7EB852; mov.f32 %f3000, 0f3F7FFFEF; @%p116 bra $L__BB184_164; mov.f32 %f2957, 0f3F800000; sub.ftz.f32 %f1022, %f2957, 0f3F000000; mul.ftz.f32 %f1023, %f1022, 0f40400000; sqrt.approx.ftz.f32 %f1024, %f1023; mul.ftz.f32 %f1025, %f200, %f1024; mul.ftz.f32 %f1026, %f1025, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1027, %f1026; add.ftz.f32 %f1028, %f1027, 0f3F800000; mul.ftz.f32 %f1029, %f1028, 0f3E800000; mul.ftz.f32 %f1030, %f1024, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1031, %f1030; mul.ftz.f32 %f1032, %f1031, %f1029; setp.lt.ftz.f32 %p117, %f1032, %f204; selp.f32 %f1033, 0f3F800000, 0f3F000000, %p117; selp.f32 %f1034, 0f3F000000, 0f00000000, %p117; selp.f32 %f1035, 0f3F400000, 0f3E800000, %p117; sub.ftz.f32 %f1036, %f2957, %f1035; mul.ftz.f32 %f1037, %f1036, 0f40400000; sqrt.approx.ftz.f32 %f1038, %f1037; mul.ftz.f32 %f1039, %f1035, 0f3F000000; mul.ftz.f32 %f1040, %f200, %f1038; mul.ftz.f32 %f1041, %f1040, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1042, %f1041; add.ftz.f32 %f1043, %f1042, 0f3F800000; mul.ftz.f32 %f1044, %f1043, %f1039; mul.ftz.f32 %f1045, %f1038, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1046, %f1045; mul.ftz.f32 %f1047, %f1046, %f1044; setp.lt.ftz.f32 %p118, %f1047, %f204; selp.f32 %f1048, %f1033, %f1035, %p118; selp.f32 %f1049, %f1035, %f1034, %p118; add.ftz.f32 %f1050, %f1048, %f1049; mul.ftz.f32 %f1051, %f1050, 0f3F000000; sub.ftz.f32 %f1052, %f2957, %f1051; mul.ftz.f32 %f1053, %f1052, 0f40400000; sqrt.approx.ftz.f32 %f1054, %f1053; mul.ftz.f32 %f1055, %f1051, 0f3F000000; mul.ftz.f32 %f1056, %f200, %f1054; mul.ftz.f32 %f1057, %f1056, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1058, %f1057; add.ftz.f32 %f1059, %f1058, 0f3F800000; mul.ftz.f32 %f1060, %f1059, %f1055; mul.ftz.f32 %f1061, %f1054, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1062, %f1061; mul.ftz.f32 %f1063, %f1062, %f1060; setp.lt.ftz.f32 %p119, %f1063, %f204; selp.f32 %f1064, %f1048, %f1051, %p119; selp.f32 %f1065, %f1051, %f1049, %p119; add.ftz.f32 %f1066, %f1064, %f1065; mul.ftz.f32 %f1067, %f1066, 0f3F000000; sub.ftz.f32 %f1068, %f2957, %f1067; mul.ftz.f32 %f1069, %f1068, 0f40400000; sqrt.approx.ftz.f32 %f1070, %f1069; mul.ftz.f32 %f1071, %f1067, 0f3F000000; mul.ftz.f32 %f1072, %f200, %f1070; mul.ftz.f32 %f1073, %f1072, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1074, %f1073; add.ftz.f32 %f1075, %f1074, 0f3F800000; mul.ftz.f32 %f1076, %f1075, %f1071; mul.ftz.f32 %f1077, %f1070, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1078, %f1077; mul.ftz.f32 %f1079, %f1078, %f1076; setp.lt.ftz.f32 %p120, %f1079, %f204; selp.f32 %f1080, %f1064, %f1067, %p120; selp.f32 %f1081, %f1067, %f1065, %p120; add.ftz.f32 %f1082, %f1080, %f1081; mul.ftz.f32 %f1083, %f1082, 0f3F000000; sub.ftz.f32 %f1084, %f2957, %f1083; mul.ftz.f32 %f1085, %f1084, 0f40400000; sqrt.approx.ftz.f32 %f1086, %f1085; mul.ftz.f32 %f1087, %f1083, 0f3F000000; mul.ftz.f32 %f1088, %f200, %f1086; mul.ftz.f32 %f1089, %f1088, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1090, %f1089; add.ftz.f32 %f1091, %f1090, 0f3F800000; mul.ftz.f32 %f1092, %f1091, %f1087; mul.ftz.f32 %f1093, %f1086, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1094, %f1093; mul.ftz.f32 %f1095, %f1094, %f1092; setp.lt.ftz.f32 %p121, %f1095, %f204; selp.f32 %f1096, %f1080, %f1083, %p121; selp.f32 %f1097, %f1083, %f1081, %p121; add.ftz.f32 %f1098, %f1096, %f1097; mul.ftz.f32 %f1099, %f1098, 0f3F000000; sub.ftz.f32 %f1100, %f2957, %f1099; mul.ftz.f32 %f1101, %f1100, 0f40400000; sqrt.approx.ftz.f32 %f1102, %f1101; mul.ftz.f32 %f1103, %f1099, 0f3F000000; mul.ftz.f32 %f1104, %f200, %f1102; mul.ftz.f32 %f1105, %f1104, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1106, %f1105; add.ftz.f32 %f1107, %f1106, 0f3F800000; mul.ftz.f32 %f1108, %f1107, %f1103; mul.ftz.f32 %f1109, %f1102, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1110, %f1109; mul.ftz.f32 %f1111, %f1110, %f1108; setp.lt.ftz.f32 %p122, %f1111, %f204; selp.f32 %f1112, %f1096, %f1099, %p122; selp.f32 %f1113, %f1099, %f1097, %p122; add.ftz.f32 %f1114, %f1112, %f1113; mul.ftz.f32 %f1115, %f1114, 0f3F000000; sub.ftz.f32 %f1116, %f2957, %f1115; mul.ftz.f32 %f1117, %f1116, 0f40400000; sqrt.approx.ftz.f32 %f1118, %f1117; mul.ftz.f32 %f1119, %f1115, 0f3F000000; mul.ftz.f32 %f1120, %f200, %f1118; mul.ftz.f32 %f1121, %f1120, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1122, %f1121; add.ftz.f32 %f1123, %f1122, 0f3F800000; mul.ftz.f32 %f1124, %f1123, %f1119; mul.ftz.f32 %f1125, %f1118, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1126, %f1125; mul.ftz.f32 %f1127, %f1126, %f1124; setp.lt.ftz.f32 %p123, %f1127, %f204; selp.f32 %f1128, %f1112, %f1115, %p123; selp.f32 %f1129, %f1115, %f1113, %p123; add.ftz.f32 %f1130, %f1128, %f1129; mul.ftz.f32 %f1131, %f1130, 0f3F000000; sub.ftz.f32 %f1132, %f2957, %f1131; mul.ftz.f32 %f1133, %f1132, 0f40400000; sqrt.approx.ftz.f32 %f1134, %f1133; mul.ftz.f32 %f1135, %f1131, 0f3F000000; mul.ftz.f32 %f1136, %f200, %f1134; mul.ftz.f32 %f1137, %f1136, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1138, %f1137; add.ftz.f32 %f1139, %f1138, 0f3F800000; mul.ftz.f32 %f1140, %f1139, %f1135; mul.ftz.f32 %f1141, %f1134, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1142, %f1141; mul.ftz.f32 %f1143, %f1142, %f1140; setp.lt.ftz.f32 %p124, %f1143, %f204; selp.f32 %f1144, %f1128, %f1131, %p124; selp.f32 %f1145, %f1131, %f1129, %p124; add.ftz.f32 %f1146, %f1144, %f1145; mul.ftz.f32 %f1147, %f1146, 0f3F000000; sub.ftz.f32 %f1148, %f2957, %f1147; mul.ftz.f32 %f1149, %f1148, 0f40400000; sqrt.approx.ftz.f32 %f1150, %f1149; mul.ftz.f32 %f1151, %f1147, 0f3F000000; mul.ftz.f32 %f1152, %f200, %f1150; mul.ftz.f32 %f1153, %f1152, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1154, %f1153; add.ftz.f32 %f1155, %f1154, 0f3F800000; mul.ftz.f32 %f1156, %f1155, %f1151; mul.ftz.f32 %f1157, %f1150, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1158, %f1157; mul.ftz.f32 %f1159, %f1158, %f1156; setp.lt.ftz.f32 %p125, %f1159, %f204; selp.f32 %f1160, %f1144, %f1147, %p125; selp.f32 %f1161, %f1147, %f1145, %p125; add.ftz.f32 %f1162, %f1160, %f1161; mul.ftz.f32 %f1163, %f1162, 0f3F000000; sub.ftz.f32 %f1164, %f2957, %f1163; mul.ftz.f32 %f1165, %f1164, 0f40400000; sqrt.approx.ftz.f32 %f1166, %f1165; mul.ftz.f32 %f1167, %f1163, 0f3F000000; mul.ftz.f32 %f1168, %f200, %f1166; mul.ftz.f32 %f1169, %f1168, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1170, %f1169; add.ftz.f32 %f1171, %f1170, 0f3F800000; mul.ftz.f32 %f1172, %f1171, %f1167; mul.ftz.f32 %f1173, %f1166, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1174, %f1173; mul.ftz.f32 %f1175, %f1174, %f1172; setp.lt.ftz.f32 %p126, %f1175, %f204; selp.f32 %f1176, %f1160, %f1163, %p126; selp.f32 %f1177, %f1163, %f1161, %p126; add.ftz.f32 %f1178, %f1176, %f1177; mul.ftz.f32 %f1179, %f1178, 0f3F000000; sub.ftz.f32 %f1180, %f2957, %f1179; mul.ftz.f32 %f1181, %f1180, 0f40400000; sqrt.approx.ftz.f32 %f1182, %f1181; mul.ftz.f32 %f1183, %f1179, 0f3F000000; mul.ftz.f32 %f1184, %f200, %f1182; mul.ftz.f32 %f1185, %f1184, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1186, %f1185; add.ftz.f32 %f1187, %f1186, 0f3F800000; mul.ftz.f32 %f1188, %f1187, %f1183; mul.ftz.f32 %f1189, %f1182, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1190, %f1189; mul.ftz.f32 %f1191, %f1190, %f1188; setp.lt.ftz.f32 %p127, %f1191, %f204; selp.f32 %f1192, %f1176, %f1179, %p127; selp.f32 %f1193, %f1179, %f1177, %p127; add.ftz.f32 %f1194, %f1192, %f1193; mul.ftz.f32 %f3000, %f1194, 0f3F000000; $L__BB184_164: mov.f32 %f3001, 0f00000000; ld.f32 %f207, [%rd14+20]; setp.lt.ftz.f32 %p128, %f207, 0f38D1B717; @%p128 bra $L__BB184_167; setp.ge.ftz.f32 %p129, %f207, 0f3F7EB852; mov.f32 %f3001, 0f3F7FFFEF; @%p129 bra $L__BB184_167; mov.f32 %f2952, 0f3F800000; sub.ftz.f32 %f1198, %f2952, 0f3F000000; mul.ftz.f32 %f1199, %f1198, 0f40400000; sqrt.approx.ftz.f32 %f1200, %f1199; mul.ftz.f32 %f1201, %f200, %f1200; mul.ftz.f32 %f1202, %f1201, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1203, %f1202; add.ftz.f32 %f1204, %f1203, 0f3F800000; mul.ftz.f32 %f1205, %f1204, 0f3E800000; mul.ftz.f32 %f1206, %f1200, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1207, %f1206; mul.ftz.f32 %f1208, %f1207, %f1205; setp.lt.ftz.f32 %p130, %f1208, %f207; selp.f32 %f1209, 0f3F800000, 0f3F000000, %p130; selp.f32 %f1210, 0f3F000000, 0f00000000, %p130; selp.f32 %f1211, 0f3F400000, 0f3E800000, %p130; sub.ftz.f32 %f1212, %f2952, %f1211; mul.ftz.f32 %f1213, %f1212, 0f40400000; sqrt.approx.ftz.f32 %f1214, %f1213; mul.ftz.f32 %f1215, %f1211, 0f3F000000; mul.ftz.f32 %f1216, %f200, %f1214; mul.ftz.f32 %f1217, %f1216, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1218, %f1217; add.ftz.f32 %f1219, %f1218, 0f3F800000; mul.ftz.f32 %f1220, %f1219, %f1215; mul.ftz.f32 %f1221, %f1214, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1222, %f1221; mul.ftz.f32 %f1223, %f1222, %f1220; setp.lt.ftz.f32 %p131, %f1223, %f207; selp.f32 %f1224, %f1209, %f1211, %p131; selp.f32 %f1225, %f1211, %f1210, %p131; add.ftz.f32 %f1226, %f1224, %f1225; mul.ftz.f32 %f1227, %f1226, 0f3F000000; sub.ftz.f32 %f1228, %f2952, %f1227; mul.ftz.f32 %f1229, %f1228, 0f40400000; sqrt.approx.ftz.f32 %f1230, %f1229; mul.ftz.f32 %f1231, %f1227, 0f3F000000; mul.ftz.f32 %f1232, %f200, %f1230; mul.ftz.f32 %f1233, %f1232, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1234, %f1233; add.ftz.f32 %f1235, %f1234, 0f3F800000; mul.ftz.f32 %f1236, %f1235, %f1231; mul.ftz.f32 %f1237, %f1230, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1238, %f1237; mul.ftz.f32 %f1239, %f1238, %f1236; setp.lt.ftz.f32 %p132, %f1239, %f207; selp.f32 %f1240, %f1224, %f1227, %p132; selp.f32 %f1241, %f1227, %f1225, %p132; add.ftz.f32 %f1242, %f1240, %f1241; mul.ftz.f32 %f1243, %f1242, 0f3F000000; sub.ftz.f32 %f1244, %f2952, %f1243; mul.ftz.f32 %f1245, %f1244, 0f40400000; sqrt.approx.ftz.f32 %f1246, %f1245; mul.ftz.f32 %f1247, %f1243, 0f3F000000; mul.ftz.f32 %f1248, %f200, %f1246; mul.ftz.f32 %f1249, %f1248, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1250, %f1249; add.ftz.f32 %f1251, %f1250, 0f3F800000; mul.ftz.f32 %f1252, %f1251, %f1247; mul.ftz.f32 %f1253, %f1246, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1254, %f1253; mul.ftz.f32 %f1255, %f1254, %f1252; setp.lt.ftz.f32 %p133, %f1255, %f207; selp.f32 %f1256, %f1240, %f1243, %p133; selp.f32 %f1257, %f1243, %f1241, %p133; add.ftz.f32 %f1258, %f1256, %f1257; mul.ftz.f32 %f1259, %f1258, 0f3F000000; sub.ftz.f32 %f1260, %f2952, %f1259; mul.ftz.f32 %f1261, %f1260, 0f40400000; sqrt.approx.ftz.f32 %f1262, %f1261; mul.ftz.f32 %f1263, %f1259, 0f3F000000; mul.ftz.f32 %f1264, %f200, %f1262; mul.ftz.f32 %f1265, %f1264, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1266, %f1265; add.ftz.f32 %f1267, %f1266, 0f3F800000; mul.ftz.f32 %f1268, %f1267, %f1263; mul.ftz.f32 %f1269, %f1262, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1270, %f1269; mul.ftz.f32 %f1271, %f1270, %f1268; setp.lt.ftz.f32 %p134, %f1271, %f207; selp.f32 %f1272, %f1256, %f1259, %p134; selp.f32 %f1273, %f1259, %f1257, %p134; add.ftz.f32 %f1274, %f1272, %f1273; mul.ftz.f32 %f1275, %f1274, 0f3F000000; sub.ftz.f32 %f1276, %f2952, %f1275; mul.ftz.f32 %f1277, %f1276, 0f40400000; sqrt.approx.ftz.f32 %f1278, %f1277; mul.ftz.f32 %f1279, %f1275, 0f3F000000; mul.ftz.f32 %f1280, %f200, %f1278; mul.ftz.f32 %f1281, %f1280, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1282, %f1281; add.ftz.f32 %f1283, %f1282, 0f3F800000; mul.ftz.f32 %f1284, %f1283, %f1279; mul.ftz.f32 %f1285, %f1278, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1286, %f1285; mul.ftz.f32 %f1287, %f1286, %f1284; setp.lt.ftz.f32 %p135, %f1287, %f207; selp.f32 %f1288, %f1272, %f1275, %p135; selp.f32 %f1289, %f1275, %f1273, %p135; add.ftz.f32 %f1290, %f1288, %f1289; mul.ftz.f32 %f1291, %f1290, 0f3F000000; sub.ftz.f32 %f1292, %f2952, %f1291; mul.ftz.f32 %f1293, %f1292, 0f40400000; sqrt.approx.ftz.f32 %f1294, %f1293; mul.ftz.f32 %f1295, %f1291, 0f3F000000; mul.ftz.f32 %f1296, %f200, %f1294; mul.ftz.f32 %f1297, %f1296, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1298, %f1297; add.ftz.f32 %f1299, %f1298, 0f3F800000; mul.ftz.f32 %f1300, %f1299, %f1295; mul.ftz.f32 %f1301, %f1294, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1302, %f1301; mul.ftz.f32 %f1303, %f1302, %f1300; setp.lt.ftz.f32 %p136, %f1303, %f207; selp.f32 %f1304, %f1288, %f1291, %p136; selp.f32 %f1305, %f1291, %f1289, %p136; add.ftz.f32 %f1306, %f1304, %f1305; mul.ftz.f32 %f1307, %f1306, 0f3F000000; sub.ftz.f32 %f1308, %f2952, %f1307; mul.ftz.f32 %f1309, %f1308, 0f40400000; sqrt.approx.ftz.f32 %f1310, %f1309; mul.ftz.f32 %f1311, %f1307, 0f3F000000; mul.ftz.f32 %f1312, %f200, %f1310; mul.ftz.f32 %f1313, %f1312, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1314, %f1313; add.ftz.f32 %f1315, %f1314, 0f3F800000; mul.ftz.f32 %f1316, %f1315, %f1311; mul.ftz.f32 %f1317, %f1310, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1318, %f1317; mul.ftz.f32 %f1319, %f1318, %f1316; setp.lt.ftz.f32 %p137, %f1319, %f207; selp.f32 %f1320, %f1304, %f1307, %p137; selp.f32 %f1321, %f1307, %f1305, %p137; add.ftz.f32 %f1322, %f1320, %f1321; mul.ftz.f32 %f1323, %f1322, 0f3F000000; sub.ftz.f32 %f1324, %f2952, %f1323; mul.ftz.f32 %f1325, %f1324, 0f40400000; sqrt.approx.ftz.f32 %f1326, %f1325; mul.ftz.f32 %f1327, %f1323, 0f3F000000; mul.ftz.f32 %f1328, %f200, %f1326; mul.ftz.f32 %f1329, %f1328, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1330, %f1329; add.ftz.f32 %f1331, %f1330, 0f3F800000; mul.ftz.f32 %f1332, %f1331, %f1327; mul.ftz.f32 %f1333, %f1326, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1334, %f1333; mul.ftz.f32 %f1335, %f1334, %f1332; setp.lt.ftz.f32 %p138, %f1335, %f207; selp.f32 %f1336, %f1320, %f1323, %p138; selp.f32 %f1337, %f1323, %f1321, %p138; add.ftz.f32 %f1338, %f1336, %f1337; mul.ftz.f32 %f1339, %f1338, 0f3F000000; sub.ftz.f32 %f1340, %f2952, %f1339; mul.ftz.f32 %f1341, %f1340, 0f40400000; sqrt.approx.ftz.f32 %f1342, %f1341; mul.ftz.f32 %f1343, %f1339, 0f3F000000; mul.ftz.f32 %f1344, %f200, %f1342; mul.ftz.f32 %f1345, %f1344, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1346, %f1345; add.ftz.f32 %f1347, %f1346, 0f3F800000; mul.ftz.f32 %f1348, %f1347, %f1343; mul.ftz.f32 %f1349, %f1342, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1350, %f1349; mul.ftz.f32 %f1351, %f1350, %f1348; setp.lt.ftz.f32 %p139, %f1351, %f207; selp.f32 %f1352, %f1336, %f1339, %p139; selp.f32 %f1353, %f1339, %f1337, %p139; add.ftz.f32 %f1354, %f1352, %f1353; mul.ftz.f32 %f1355, %f1354, 0f3F000000; sub.ftz.f32 %f1356, %f2952, %f1355; mul.ftz.f32 %f1357, %f1356, 0f40400000; sqrt.approx.ftz.f32 %f1358, %f1357; mul.ftz.f32 %f1359, %f1355, 0f3F000000; mul.ftz.f32 %f1360, %f200, %f1358; mul.ftz.f32 %f1361, %f1360, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1362, %f1361; add.ftz.f32 %f1363, %f1362, 0f3F800000; mul.ftz.f32 %f1364, %f1363, %f1359; mul.ftz.f32 %f1365, %f1358, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1366, %f1365; mul.ftz.f32 %f1367, %f1366, %f1364; setp.lt.ftz.f32 %p140, %f1367, %f207; selp.f32 %f1368, %f1352, %f1355, %p140; selp.f32 %f1369, %f1355, %f1353, %p140; add.ftz.f32 %f1370, %f1368, %f1369; mul.ftz.f32 %f3001, %f1370, 0f3F000000; $L__BB184_167: mov.f32 %f2953, 0f3F800000; sub.ftz.f32 %f1372, %f2953, %f2999; mul.ftz.f32 %f1373, %f1372, 0f40400000; sub.ftz.f32 %f1374, %f2953, %f3000; mul.ftz.f32 %f1375, %f1374, 0f40400000; sub.ftz.f32 %f1376, %f2953, %f3001; mul.ftz.f32 %f1377, %f1376, 0f40400000; sqrt.approx.ftz.f32 %f1378, %f1373; sqrt.approx.ftz.f32 %f1379, %f1375; sqrt.approx.ftz.f32 %f1380, %f1377; ld.f32 %f1381, [%rd14]; mul.ftz.f32 %f3002, %f1378, %f1381; ld.f32 %f1382, [%rd14+4]; mul.ftz.f32 %f3003, %f1379, %f1382; ld.f32 %f1383, [%rd14+8]; mul.ftz.f32 %f3004, %f1380, %f1383; $L__BB184_169: st.f32 [%rd14], %f3002; st.f32 [%rd14+4], %f3003; st.f32 [%rd14+8], %f3004; or.b32 %r610, %r610, 16; $L__BB184_170: ld.local.u32 %r215, [%rd4+40]; or.b32 %r216, %r215, %r610; st.local.u32 [%rd4+40], %r216; $L__BB184_178: setp.leu.ftz.f32 %p145, %f97, 0f3727C5AC; setp.leu.ftz.f32 %p146, %f74, 0f3727C5AC; or.pred %p147, %p146, %p145; @%p147 bra $L__BB184_190; ld.const.v4.f32 {%f1400, %f1401, %f1402, %f1403}, [kernel_params+1184]; mul.ftz.f32 %f1407, %f2978, %f1401; fma.rn.ftz.f32 %f1408, %f2977, %f1400, %f1407; fma.rn.ftz.f32 %f223, %f2979, %f1402, %f1408; setp.leu.ftz.f32 %p148, %f223, 0f00000000; mov.f32 %f1399, 0f3F800000; mov.f32 %f3005, %f1399; mov.f32 %f3006, %f1399; mov.f32 %f3007, %f1399; @%p148 bra $L__BB184_181; div.approx.ftz.f32 %f3007, %f2977, %f223; div.approx.ftz.f32 %f3006, %f2978, %f223; div.approx.ftz.f32 %f3005, %f2979, %f223; $L__BB184_181: sub.ftz.f32 %f1410, %f1399, %f75; fma.rn.ftz.f32 %f1411, %f75, %f3007, %f1410; fma.rn.ftz.f32 %f1412, %f75, %f3006, %f1410; fma.rn.ftz.f32 %f1413, %f75, %f3005, %f1410; mul.ftz.f32 %f1414, %f74, %f177; mul.ftz.f32 %f1415, %f1414, %f1411; mul.ftz.f32 %f1416, %f74, %f178; mul.ftz.f32 %f1417, %f1416, %f1412; mul.ftz.f32 %f1418, %f74, %f179; mul.ftz.f32 %f1419, %f1418, %f1413; mul.ftz.f32 %f1420, %f97, %f1415; mul.ftz.f32 %f1421, %f97, %f1417; mul.ftz.f32 %f1422, %f97, %f1419; mov.f32 %f3008, 0f00000000; max.ftz.f32 %f230, %f1420, %f3008; max.ftz.f32 %f231, %f1421, %f3008; max.ftz.f32 %f232, %f1422, %f3008; add.ftz.f32 %f1424, %f230, %f231; add.ftz.f32 %f1425, %f232, %f1424; mul.ftz.f32 %f1426, %f1425, 0f3EAAAAAB; abs.ftz.f32 %f233, %f1426; setp.ltu.ftz.f32 %p149, %f233, 0f3727C5AC; mov.u64 %rd712, 0; mov.u64 %rd713, %rd712; @%p149 bra $L__BB184_186; ld.local.u32 %r43, [%rd4+252]; mov.u64 %rd712, 0; setp.eq.s32 %p150, %r43, 0; mov.u64 %rd710, %rd712; mov.u64 %rd711, %rd712; @%p150 bra $L__BB184_184; ld.local.u32 %r228, [%rd4+248]; mul.wide.s32 %rd337, %r228, 80; add.s64 %rd338, %rd1, %rd337; add.s64 %rd710, %rd338, 304; add.s64 %rd339, %rd230, %rd337; add.s64 %rd711, %rd339, 304; mov.b32 %r229, %f232; mov.b32 %r230, %f231; mov.b32 %r231, %f230; mov.u32 %r232, 0; st.local.v4.u32 [%rd338+304], {%r231, %r230, %r229, %r232}; add.s32 %r233, %r228, 1; st.local.u32 [%rd4+248], %r233; add.s32 %r234, %r43, -1; st.local.u32 [%rd4+252], %r234; $L__BB184_184: setp.eq.s64 %p151, %rd711, 0; mov.u64 %rd713, %rd712; @%p151 bra $L__BB184_186; st.local.f32 [%rd710+16], %f233; mov.u64 %rd712, %rd710; mov.u64 %rd713, %rd711; $L__BB184_186: setp.eq.s64 %p152, %rd713, 0; @%p152 bra $L__BB184_190; add.s64 %rd45, %rd712, 20; st.local.f32 [%rd712+20], %f55; st.local.f32 [%rd712+24], %f56; st.local.f32 [%rd712+28], %f57; mov.u32 %r235, 6; st.local.u32 [%rd712+12], %r235; ld.local.f32 %f1428, [%rd4+24]; ld.local.f32 %f1429, [%rd4+28]; mul.ftz.f32 %f1430, %f56, %f1429; fma.rn.ftz.f32 %f1431, %f55, %f1428, %f1430; ld.local.f32 %f1432, [%rd4+32]; fma.rn.ftz.f32 %f234, %f57, %f1432, %f1431; setp.lt.ftz.f32 %p153, %f234, 0f00000000; @%p153 bra $L__BB184_189; mov.f32 %f1433, 0f3F800000; sub.ftz.f32 %f1434, %f1433, %f234; mov.f32 %f1435, 0f00000000; max.ftz.f32 %f1436, %f1434, %f1435; min.ftz.f32 %f1437, %f1436, %f1433; mul.ftz.f32 %f1438, %f1437, %f1437; mul.ftz.f32 %f1439, %f1438, %f1438; mul.ftz.f32 %f1440, %f1437, %f1439; mul.ftz.f32 %f3008, %f234, %f1440; $L__BB184_189: st.local.f32 [%rd45+12], %f3008; ld.local.f32 %f1441, [%rd45+-4]; mul.ftz.f32 %f1442, %f3008, %f1441; st.local.f32 [%rd45+-4], %f1442; ld.local.u32 %r236, [%rd4+40]; or.b32 %r237, %r236, 12; st.local.u32 [%rd4+40], %r237; $L__BB184_190: setp.leu.ftz.f32 %p154, %f99, 0f3727C5AC; @%p154 bra $L__BB184_221; setp.leu.ftz.f32 %p155, %f70, 0f3727C5AC; setp.leu.ftz.f32 %p156, %f60, 0f3727C5AC; and.pred %p157, %p156, %p155; @%p157 bra $L__BB184_221; mul.ftz.f32 %f1443, %f99, %f177; mov.f32 %f1444, 0f00000000; max.ftz.f32 %f237, %f1443, %f1444; mul.ftz.f32 %f1445, %f99, %f178; max.ftz.f32 %f238, %f1445, %f1444; mul.ftz.f32 %f1446, %f99, %f179; max.ftz.f32 %f239, %f1446, %f1444; add.ftz.f32 %f1447, %f237, %f238; add.ftz.f32 %f1448, %f239, %f1447; mul.ftz.f32 %f1449, %f1448, 0f3EAAAAAB; abs.ftz.f32 %f240, %f1449; setp.ltu.ftz.f32 %p158, %f240, 0f3727C5AC; mov.u64 %rd720, 0; mov.u64 %rd716, %rd720; mov.u64 %rd717, %rd720; @%p158 bra $L__BB184_197; ld.local.u32 %r44, [%rd4+252]; mov.u64 %rd716, 0; setp.eq.s32 %p159, %r44, 0; mov.u64 %rd714, %rd716; mov.u64 %rd715, %rd716; @%p159 bra $L__BB184_195; ld.local.u32 %r238, [%rd4+248]; mul.wide.s32 %rd346, %r238, 80; add.s64 %rd347, %rd1, %rd346; add.s64 %rd714, %rd347, 304; add.s64 %rd348, %rd230, %rd346; add.s64 %rd715, %rd348, 304; mov.b32 %r239, %f239; mov.b32 %r240, %f238; mov.b32 %r241, %f237; mov.u32 %r242, 0; st.local.v4.u32 [%rd347+304], {%r241, %r240, %r239, %r242}; add.s32 %r243, %r238, 1; st.local.u32 [%rd4+248], %r243; add.s32 %r244, %r44, -1; st.local.u32 [%rd4+252], %r244; $L__BB184_195: setp.eq.s64 %p160, %rd715, 0; mov.u64 %rd717, %rd716; @%p160 bra $L__BB184_197; st.local.f32 [%rd714+16], %f240; mov.u64 %rd716, %rd714; mov.u64 %rd717, %rd715; $L__BB184_197: setp.eq.s64 %p161, %rd717, 0; mov.u64 %rd721, %rd720; @%p161 bra $L__BB184_201; ld.local.u32 %r45, [%rd4+252]; setp.lt.s32 %p162, %r45, 1; @%p162 bra $L__BB184_200; add.s32 %r245, %r45, -1; st.local.u32 [%rd4+252], %r245; ld.local.s32 %rd353, [%rd4+248]; cvt.s64.s32 %rd354, %r245; add.s64 %rd355, %rd353, %rd354; mul.lo.s64 %rd356, %rd355, 80; add.s64 %rd357, %rd1, %rd356; add.s64 %rd720, %rd357, 304; add.s64 %rd358, %rd230, %rd356; add.s64 %rd721, %rd358, 304; bra.uni $L__BB184_201; $L__BB184_200: ld.local.u32 %r246, [%rd4+248]; mov.u64 %rd720, 0; add.s32 %r247, %r246, -1; st.local.u32 [%rd4+248], %r247; add.s32 %r248, %r45, 1; st.local.u32 [%rd4+252], %r248; mov.u64 %rd721, %rd720; $L__BB184_201: setp.eq.s64 %p163, %rd721, 0; or.pred %p165, %p161, %p163; @%p165 bra $L__BB184_221; add.s64 %rd60, %rd716, 20; st.local.f32 [%rd716+20], %f55; st.local.f32 [%rd716+24], %f56; st.local.f32 [%rd716+28], %f57; mul.ftz.f32 %f241, %f70, 0f3DA3D70A; max.ftz.f32 %f1454, %f241, %f1444; sqrt.approx.ftz.f32 %f1455, %f1454; mov.f32 %f1452, 0f3F800000; sub.ftz.f32 %f1456, %f1452, %f1455; mov.f32 %f1457, 0f40000000; div.approx.ftz.f32 %f1458, %f1457, %f1456; add.ftz.f32 %f1459, %f1458, 0fBF800000; st.local.f32 [%rd716+40], %f1459; st.local.v2.f32 [%rd716+56], {%f2972, %f2973}; st.local.f32 [%rd716+64], %f2974; st.local.u64 [%rd716+48], %rd721; fma.rn.ftz.f32 %f1460, %f73, 0fBF666666, 0f3F800000; max.ftz.f32 %f1461, %f1460, %f1444; sqrt.approx.ftz.f32 %f1462, %f1461; mul.ftz.f32 %f1463, %f71, %f71; mul.ftz.f32 %f1464, %f1463, %f1462; div.approx.ftz.f32 %f1465, %f1463, %f1462; st.local.v2.f32 [%rd716+32], {%f1465, %f1464}; mul.ftz.f32 %f1466, %f2978, 0f3F19999A; fma.rn.ftz.f32 %f1467, %f2977, 0f3E99999A, %f1466; fma.rn.ftz.f32 %f242, %f2979, 0f3DCCCCCD, %f1467; setp.leu.ftz.f32 %p166, %f242, 0f00000000; mov.f32 %f3009, %f1452; mov.f32 %f3010, %f1452; mov.f32 %f3011, %f1452; @%p166 bra $L__BB184_204; div.approx.ftz.f32 %f3011, %f2977, %f242; div.approx.ftz.f32 %f3010, %f2978, %f242; div.approx.ftz.f32 %f3009, %f2979, %f242; $L__BB184_204: sub.ftz.f32 %f1469, %f1452, %f72; fma.rn.ftz.f32 %f1470, %f72, %f3011, %f1469; fma.rn.ftz.f32 %f1471, %f72, %f3010, %f1469; fma.rn.ftz.f32 %f1472, %f72, %f3009, %f1469; mul.ftz.f32 %f1473, %f241, %f1470; mul.ftz.f32 %f1474, %f241, %f1471; mul.ftz.f32 %f1475, %f241, %f1472; sub.ftz.f32 %f1476, %f1452, %f60; mul.ftz.f32 %f1477, %f1476, %f1473; mul.ftz.f32 %f1478, %f1476, %f1474; mul.ftz.f32 %f1479, %f1476, %f1475; fma.rn.ftz.f32 %f1480, %f60, %f2977, %f1477; fma.rn.ftz.f32 %f1481, %f60, %f2978, %f1478; fma.rn.ftz.f32 %f1482, %f60, %f2979, %f1479; st.local.f32 [%rd720+12], %f1480; st.local.f32 [%rd720+16], %f1481; st.local.f32 [%rd720+20], %f1482; ld.local.u64 %rd361, [%rd60+28]; st.f32 [%rd361], %f2977; st.f32 [%rd361+4], %f2978; st.f32 [%rd361+8], %f2979; mov.u32 %r249, 0; ld.local.u64 %rd362, [%rd60+28]; st.u32 [%rd362+36], %r249; setp.le.ftz.f32 %p167, %f71, 0f3D99999A; setp.eq.s32 %p168, %r140, 26; or.pred %p169, %p168, %p167; @%p169 bra $L__BB184_214; bra.uni $L__BB184_205; $L__BB184_214: ld.local.u64 %rd367, [%rd60+28]; ld.f32 %f1569, [%rd367+12]; cvt.ftz.sat.f32.f32 %f1570, %f1569; ld.f32 %f1571, [%rd367+16]; cvt.ftz.sat.f32.f32 %f1572, %f1571; ld.f32 %f1573, [%rd367+20]; cvt.ftz.sat.f32.f32 %f1574, %f1573; st.f32 [%rd367+12], %f1570; st.f32 [%rd367+16], %f1572; st.f32 [%rd367+20], %f1574; ld.local.v2.f32 {%f1575, %f1576}, [%rd60+12]; cvt.ftz.sat.f32.f32 %f1579, %f1576; cvt.ftz.sat.f32.f32 %f1580, %f1575; st.local.v2.f32 [%rd60+12], {%f1580, %f1579}; mov.u32 %r255, 11; st.local.u32 [%rd60+-8], %r255; ld.local.f32 %f1581, [%rd60+20]; fma.rn.ftz.f32 %f262, %f1581, %f1581, 0fBF800000; mov.f32 %f1568, 0f3F800000; abs.ftz.f32 %f263, %f1568; fma.rn.ftz.f32 %f264, %f263, %f263, %f262; setp.leu.ftz.f32 %p178, %f264, 0f00000000; mov.f32 %f3014, %f1568; @%p178 bra $L__BB184_216; sqrt.approx.ftz.f32 %f1582, %f264; sub.ftz.f32 %f1583, %f1582, %f263; add.ftz.f32 %f1584, %f263, %f1582; div.approx.ftz.f32 %f1585, %f1583, %f1584; fma.rn.ftz.f32 %f1586, %f263, %f1584, 0fBF800000; fma.rn.ftz.f32 %f1587, %f263, %f1583, 0f3F800000; div.approx.ftz.f32 %f1588, %f1586, %f1587; mul.ftz.f32 %f1589, %f1585, 0f3F000000; mul.ftz.f32 %f1590, %f1585, %f1589; fma.rn.ftz.f32 %f1591, %f1588, %f1588, 0f3F800000; mul.ftz.f32 %f3014, %f1590, %f1591; $L__BB184_216: ld.local.u64 %rd368, [%rd60+28]; add.s64 %rd62, %rd368, 12; ld.f32 %f267, [%rd368+12]; ld.f32 %f268, [%rd368+16]; ld.f32 %f269, [%rd368+20]; sub.ftz.f32 %f270, %f1568, %f3014; ld.local.f32 %f1593, [%rd60]; ld.local.f32 %f1594, [%rd4+24]; ld.local.f32 %f1595, [%rd60+4]; ld.local.f32 %f1596, [%rd4+28]; mul.ftz.f32 %f1597, %f1596, %f1595; fma.rn.ftz.f32 %f1598, %f1594, %f1593, %f1597; ld.local.f32 %f1599, [%rd60+8]; ld.local.f32 %f1600, [%rd4+32]; fma.rn.ftz.f32 %f1601, %f1600, %f1599, %f1598; abs.ftz.f32 %f271, %f1601; fma.rn.ftz.f32 %f272, %f271, %f271, %f262; setp.leu.ftz.f32 %p179, %f272, 0f00000000; mov.f32 %f3015, %f270; @%p179 bra $L__BB184_218; sqrt.approx.ftz.f32 %f1602, %f272; sub.ftz.f32 %f1603, %f1602, %f271; add.ftz.f32 %f1604, %f271, %f1602; div.approx.ftz.f32 %f1605, %f1603, %f1604; fma.rn.ftz.f32 %f1606, %f271, %f1604, 0fBF800000; fma.rn.ftz.f32 %f1607, %f271, %f1603, 0f3F800000; div.approx.ftz.f32 %f1608, %f1606, %f1607; mul.ftz.f32 %f1609, %f1605, 0f3F000000; mul.ftz.f32 %f1610, %f1605, %f1609; fma.rn.ftz.f32 %f1611, %f1608, %f1608, 0f3F800000; mul.ftz.f32 %f1612, %f1610, %f1611; sub.ftz.f32 %f3015, %f1612, %f3014; $L__BB184_218: rcp.approx.ftz.f32 %f1613, %f270; mov.f32 %f1614, 0f3F800000; mul.ftz.f32 %f1615, %f1613, %f3015; sub.ftz.f32 %f1616, %f1614, %f1615; fma.rn.ftz.f32 %f1617, %f267, %f1616, %f1615; fma.rn.ftz.f32 %f1618, %f268, %f1616, %f1615; fma.rn.ftz.f32 %f1619, %f269, %f1616, %f1615; st.f32 [%rd62+12], %f1617; st.f32 [%rd62+16], %f1618; st.f32 [%rd62+20], %f1619; ld.local.u32 %r256, [%rd60+-8]; setp.ne.s32 %p180, %r256, 12; @%p180 bra $L__BB184_220; ld.local.u64 %rd369, [%rd60+28]; ld.f32 %f1620, [%rd369+36]; mul.ftz.f32 %f1621, %f1620, 0f3E800000; ld.f32 %f1622, [%rd369+24]; mul.ftz.f32 %f1623, %f1622, %f1621; ld.f32 %f1624, [%rd369+28]; mul.ftz.f32 %f1625, %f1621, %f1624; ld.f32 %f1626, [%rd369+32]; mul.ftz.f32 %f1627, %f1621, %f1626; st.f32 [%rd369+24], %f1623; st.f32 [%rd369+28], %f1625; st.f32 [%rd369+32], %f1627; $L__BB184_220: ld.local.u64 %rd370, [%rd60+28]; ld.f32 %f1628, [%rd370+28]; ld.f32 %f1629, [%rd370+24]; add.ftz.f32 %f1630, %f1629, %f1628; ld.f32 %f1631, [%rd370+32]; add.ftz.f32 %f1632, %f1631, %f1630; mul.ftz.f32 %f1633, %f1632, 0f3EAAAAAB; ld.local.f32 %f1634, [%rd60+-4]; mul.ftz.f32 %f1635, %f1634, %f1633; st.local.f32 [%rd60+-4], %f1635; ld.local.u32 %r257, [%rd4+40]; or.b32 %r258, %r257, 12; st.local.u32 [%rd4+40], %r258; bra.uni $L__BB184_221; $L__BB184_205: ld.local.v2.f32 {%f1483, %f1484}, [%rd60+36]; setp.neu.ftz.f32 %p170, %f1483, 0f00000000; setp.neu.ftz.f32 %p171, %f1484, 0f00000000; or.pred %p172, %p170, %p171; ld.local.f32 %f1487, [%rd60+44]; setp.neu.ftz.f32 %p173, %f1487, 0f00000000; or.pred %p174, %p173, %p172; @%p174 bra $L__BB184_207; mov.f32 %f1489, 0f3F800000; st.local.v2.f32 [%rd60+36], {%f1489, %f1444}; st.local.u32 [%rd60+44], %r249; $L__BB184_207: mov.u32 %r251, 15; st.local.u32 [%rd60+-8], %r251; ld.local.f32 %f1491, [%rd60+20]; fma.rn.ftz.f32 %f249, %f1491, %f1491, 0fBF800000; mov.f32 %f1490, 0f3F800000; abs.ftz.f32 %f250, %f1490; fma.rn.ftz.f32 %f251, %f250, %f250, %f249; setp.leu.ftz.f32 %p175, %f251, 0f00000000; mov.f32 %f3012, %f1490; @%p175 bra $L__BB184_209; sqrt.approx.ftz.f32 %f1492, %f251; sub.ftz.f32 %f1493, %f1492, %f250; add.ftz.f32 %f1494, %f250, %f1492; div.approx.ftz.f32 %f1495, %f1493, %f1494; fma.rn.ftz.f32 %f1496, %f250, %f1494, 0fBF800000; fma.rn.ftz.f32 %f1497, %f250, %f1493, 0f3F800000; div.approx.ftz.f32 %f1498, %f1496, %f1497; mul.ftz.f32 %f1499, %f1495, 0f3F000000; mul.ftz.f32 %f1500, %f1495, %f1499; fma.rn.ftz.f32 %f1501, %f1498, %f1498, 0f3F800000; mul.ftz.f32 %f3012, %f1500, %f1501; $L__BB184_209: ld.local.u64 %rd363, [%rd60+28]; add.s64 %rd61, %rd363, 12; ld.f32 %f254, [%rd363+12]; ld.f32 %f255, [%rd363+16]; ld.f32 %f256, [%rd363+20]; sub.ftz.f32 %f257, %f1490, %f3012; ld.local.f32 %f1503, [%rd60]; ld.local.f32 %f1504, [%rd4+24]; ld.local.f32 %f1505, [%rd60+4]; ld.local.f32 %f1506, [%rd4+28]; mul.ftz.f32 %f1507, %f1506, %f1505; fma.rn.ftz.f32 %f1508, %f1504, %f1503, %f1507; ld.local.f32 %f1509, [%rd60+8]; ld.local.f32 %f1510, [%rd4+32]; fma.rn.ftz.f32 %f1511, %f1510, %f1509, %f1508; abs.ftz.f32 %f258, %f1511; fma.rn.ftz.f32 %f259, %f258, %f258, %f249; setp.leu.ftz.f32 %p176, %f259, 0f00000000; mov.f32 %f3013, %f257; @%p176 bra $L__BB184_211; sqrt.approx.ftz.f32 %f1512, %f259; sub.ftz.f32 %f1513, %f1512, %f258; add.ftz.f32 %f1514, %f258, %f1512; div.approx.ftz.f32 %f1515, %f1513, %f1514; fma.rn.ftz.f32 %f1516, %f258, %f1514, 0fBF800000; fma.rn.ftz.f32 %f1517, %f258, %f1513, 0f3F800000; div.approx.ftz.f32 %f1518, %f1516, %f1517; mul.ftz.f32 %f1519, %f1515, 0f3F000000; mul.ftz.f32 %f1520, %f1515, %f1519; fma.rn.ftz.f32 %f1521, %f1518, %f1518, 0f3F800000; mul.ftz.f32 %f1522, %f1520, %f1521; sub.ftz.f32 %f3013, %f1522, %f3012; $L__BB184_211: rcp.approx.ftz.f32 %f1523, %f257; mov.f32 %f1524, 0f3F800000; mul.ftz.f32 %f1525, %f1523, %f3013; sub.ftz.f32 %f1526, %f1524, %f1525; fma.rn.ftz.f32 %f1527, %f254, %f1526, %f1525; fma.rn.ftz.f32 %f1528, %f255, %f1526, %f1525; fma.rn.ftz.f32 %f1529, %f256, %f1526, %f1525; st.f32 [%rd61+12], %f1527; st.f32 [%rd61+16], %f1528; st.f32 [%rd61+20], %f1529; ld.local.u32 %r252, [%rd60+-8]; setp.ne.s32 %p177, %r252, 12; @%p177 bra $L__BB184_213; ld.local.u64 %rd364, [%rd60+28]; ld.f32 %f1530, [%rd364+36]; mul.ftz.f32 %f1531, %f1530, 0f3E800000; ld.f32 %f1532, [%rd364+24]; mul.ftz.f32 %f1533, %f1532, %f1531; ld.f32 %f1534, [%rd364+28]; mul.ftz.f32 %f1535, %f1531, %f1534; ld.f32 %f1536, [%rd364+32]; mul.ftz.f32 %f1537, %f1531, %f1536; st.f32 [%rd364+24], %f1533; st.f32 [%rd364+28], %f1535; st.f32 [%rd364+32], %f1537; $L__BB184_213: ld.local.u64 %rd365, [%rd60+28]; ld.f32 %f1538, [%rd365+28]; ld.f32 %f1539, [%rd365+24]; add.ftz.f32 %f1540, %f1539, %f1538; ld.f32 %f1541, [%rd365+32]; add.ftz.f32 %f1542, %f1541, %f1540; mul.ftz.f32 %f1543, %f1542, 0f3EAAAAAB; ld.local.f32 %f1544, [%rd60+-4]; mul.ftz.f32 %f1545, %f1544, %f1543; st.local.f32 [%rd60+-4], %f1545; ld.local.v2.f32 {%f1546, %f1547}, [%rd60+12]; mov.f32 %f1550, 0f38D1B717; max.ftz.f32 %f1551, %f1546, %f1550; max.ftz.f32 %f1553, %f1547, %f1550; min.ftz.f32 %f1554, %f1553, %f1524; min.ftz.f32 %f1555, %f1551, %f1524; st.local.v2.f32 [%rd60+12], {%f1555, %f1554}; ld.f32 %f1556, [%rd365]; cvt.ftz.sat.f32.f32 %f1557, %f1556; ld.f32 %f1558, [%rd365+4]; cvt.ftz.sat.f32.f32 %f1559, %f1558; ld.f32 %f1560, [%rd365+8]; cvt.ftz.sat.f32.f32 %f1561, %f1560; st.f32 [%rd365], %f1557; st.f32 [%rd365+4], %f1559; st.f32 [%rd365+8], %f1561; ld.local.u64 %rd366, [%rd60+28]; ld.f32 %f1562, [%rd366+12]; cvt.ftz.sat.f32.f32 %f1563, %f1562; ld.f32 %f1564, [%rd366+16]; cvt.ftz.sat.f32.f32 %f1565, %f1564; ld.f32 %f1566, [%rd366+20]; cvt.ftz.sat.f32.f32 %f1567, %f1566; st.f32 [%rd366+12], %f1563; st.f32 [%rd366+16], %f1565; st.f32 [%rd366+20], %f1567; ld.local.u32 %r253, [%rd4+40]; or.b32 %r254, %r253, 1036; st.local.u32 [%rd4+40], %r254; $L__BB184_221: setp.leu.ftz.f32 %p181, %f98, 0f3727C5AC; @%p181 bra $L__BB184_264; mul.ftz.f32 %f275, %f98, %f177; mul.ftz.f32 %f276, %f98, %f178; mul.ftz.f32 %f277, %f98, %f179; mov.f32 %f1636, 0f3F800000; sub.ftz.f32 %f1637, %f1636, %f72; fma.rn.ftz.f32 %f278, %f72, %f2977, %f1637; fma.rn.ftz.f32 %f279, %f72, %f2978, %f1637; fma.rn.ftz.f32 %f280, %f72, %f2979, %f1637; setp.eq.s32 %p182, %r140, 26; setp.le.ftz.f32 %p183, %f71, 0f3D4CCCCD; or.pred %p184, %p182, %p183; @%p184 bra $L__BB184_240; bra.uni $L__BB184_223; $L__BB184_240: mul.ftz.f32 %f1718, %f2976, %f275; mov.f32 %f1719, 0f00000000; max.ftz.f32 %f298, %f1718, %f1719; mul.ftz.f32 %f1720, %f2976, %f276; max.ftz.f32 %f299, %f1720, %f1719; mul.ftz.f32 %f1721, %f2976, %f277; max.ftz.f32 %f300, %f1721, %f1719; add.ftz.f32 %f1722, %f298, %f299; add.ftz.f32 %f1723, %f300, %f1722; mul.ftz.f32 %f1724, %f1723, 0f3EAAAAAB; abs.ftz.f32 %f301, %f1724; setp.ltu.ftz.f32 %p196, %f301, 0f3727C5AC; mov.u64 %rd732, 0; mov.u64 %rd730, %rd732; mov.u64 %rd731, %rd732; @%p196 bra $L__BB184_245; ld.local.u32 %r48, [%rd4+252]; mov.u64 %rd730, 0; setp.eq.s32 %p197, %r48, 0; mov.u64 %rd728, %rd730; mov.u64 %rd729, %rd730; @%p197 bra $L__BB184_243; ld.local.u32 %r275, [%rd4+248]; mul.wide.s32 %rd401, %r275, 80; add.s64 %rd402, %rd1, %rd401; add.s64 %rd728, %rd402, 304; add.s64 %rd403, %rd230, %rd401; add.s64 %rd729, %rd403, 304; mov.b32 %r276, %f300; mov.b32 %r277, %f299; mov.b32 %r278, %f298; mov.u32 %r279, 0; st.local.v4.u32 [%rd402+304], {%r278, %r277, %r276, %r279}; add.s32 %r280, %r275, 1; st.local.u32 [%rd4+248], %r280; add.s32 %r281, %r48, -1; st.local.u32 [%rd4+252], %r281; $L__BB184_243: setp.eq.s64 %p198, %rd729, 0; mov.u64 %rd731, %rd730; @%p198 bra $L__BB184_245; st.local.f32 [%rd728+16], %f301; mov.u64 %rd730, %rd728; mov.u64 %rd731, %rd729; $L__BB184_245: setp.eq.s64 %p199, %rd731, 0; mov.u64 %rd733, %rd732; @%p199 bra $L__BB184_249; ld.local.u32 %r49, [%rd4+252]; setp.lt.s32 %p200, %r49, 1; @%p200 bra $L__BB184_248; add.s32 %r282, %r49, -1; st.local.u32 [%rd4+252], %r282; ld.local.s32 %rd408, [%rd4+248]; cvt.s64.s32 %rd409, %r282; add.s64 %rd410, %rd408, %rd409; mul.lo.s64 %rd411, %rd410, 80; add.s64 %rd412, %rd1, %rd411; add.s64 %rd732, %rd412, 304; add.s64 %rd413, %rd230, %rd411; add.s64 %rd733, %rd413, 304; bra.uni $L__BB184_249; $L__BB184_223: mov.f32 %f1638, 0f00000000; max.ftz.f32 %f281, %f275, %f1638; max.ftz.f32 %f282, %f276, %f1638; add.ftz.f32 %f1639, %f281, %f282; max.ftz.f32 %f283, %f277, %f1638; add.ftz.f32 %f1640, %f283, %f1639; mul.ftz.f32 %f1641, %f1640, 0f3EAAAAAB; abs.ftz.f32 %f284, %f1641; setp.ltu.ftz.f32 %p185, %f284, 0f3727C5AC; mov.u64 %rd726, 0; mov.u64 %rd724, %rd726; mov.u64 %rd725, %rd726; @%p185 bra $L__BB184_228; ld.local.u32 %r46, [%rd4+252]; mov.u64 %rd724, 0; setp.eq.s32 %p186, %r46, 0; mov.u64 %rd722, %rd724; mov.u64 %rd723, %rd724; @%p186 bra $L__BB184_226; ld.local.u32 %r259, [%rd4+248]; mul.wide.s32 %rd375, %r259, 80; add.s64 %rd376, %rd1, %rd375; add.s64 %rd722, %rd376, 304; add.s64 %rd377, %rd230, %rd375; add.s64 %rd723, %rd377, 304; mov.b32 %r260, %f283; mov.b32 %r261, %f282; mov.b32 %r262, %f281; mov.u32 %r263, 0; st.local.v4.u32 [%rd376+304], {%r262, %r261, %r260, %r263}; add.s32 %r264, %r259, 1; st.local.u32 [%rd4+248], %r264; add.s32 %r265, %r46, -1; st.local.u32 [%rd4+252], %r265; $L__BB184_226: setp.eq.s64 %p187, %rd723, 0; mov.u64 %rd725, %rd724; @%p187 bra $L__BB184_228; st.local.f32 [%rd722+16], %f284; mov.u64 %rd724, %rd722; mov.u64 %rd725, %rd723; $L__BB184_228: setp.eq.s64 %p188, %rd725, 0; mov.u64 %rd727, %rd726; @%p188 bra $L__BB184_232; ld.local.u32 %r47, [%rd4+252]; setp.lt.s32 %p189, %r47, 1; @%p189 bra $L__BB184_231; add.s32 %r266, %r47, -1; st.local.u32 [%rd4+252], %r266; ld.local.s32 %rd382, [%rd4+248]; cvt.s64.s32 %rd383, %r266; add.s64 %rd384, %rd382, %rd383; mul.lo.s64 %rd385, %rd384, 80; add.s64 %rd386, %rd1, %rd385; add.s64 %rd726, %rd386, 304; add.s64 %rd387, %rd230, %rd385; add.s64 %rd727, %rd387, 304; bra.uni $L__BB184_232; $L__BB184_248: ld.local.u32 %r283, [%rd4+248]; add.s32 %r284, %r283, -1; st.local.u32 [%rd4+248], %r284; add.s32 %r285, %r49, 1; st.local.u32 [%rd4+252], %r285; mov.u64 %rd733, %rd732; $L__BB184_249: setp.eq.s64 %p201, %rd733, 0; or.pred %p203, %p199, %p201; @%p203 bra $L__BB184_257; add.s64 %rd85, %rd730, 20; st.local.f32 [%rd730+20], %f55; st.local.f32 [%rd730+24], %f56; st.local.f32 [%rd730+28], %f57; mov.u32 %r286, 0; st.local.v2.f32 [%rd730+56], {%f1719, %f1719}; st.local.u32 [%rd730+64], %r286; st.local.u64 [%rd730+48], %rd733; mul.ftz.f32 %f1727, %f71, %f71; st.local.v2.f32 [%rd730+32], {%f1727, %f1727}; st.local.f32 [%rd730+40], %f2975; st.local.f32 [%rd732], %f2977; st.local.f32 [%rd732+4], %f2978; st.local.f32 [%rd732+8], %f2979; ld.local.u64 %rd416, [%rd730+48]; st.f32 [%rd416+12], %f278; st.f32 [%rd416+16], %f279; st.f32 [%rd416+20], %f280; ld.local.u64 %rd417, [%rd730+48]; st.u32 [%rd417+36], %r286; ld.local.u64 %rd418, [%rd730+48]; ld.f32 %f1728, [%rd418+12]; cvt.ftz.sat.f32.f32 %f1729, %f1728; ld.f32 %f1730, [%rd418+16]; cvt.ftz.sat.f32.f32 %f1731, %f1730; ld.f32 %f1732, [%rd418+20]; cvt.ftz.sat.f32.f32 %f1733, %f1732; st.f32 [%rd418+12], %f1729; st.f32 [%rd418+16], %f1731; st.f32 [%rd418+20], %f1733; ld.local.v2.f32 {%f1734, %f1735}, [%rd730+32]; cvt.ftz.sat.f32.f32 %f1738, %f1735; cvt.ftz.sat.f32.f32 %f1739, %f1734; st.local.v2.f32 [%rd730+32], {%f1739, %f1738}; mov.u32 %r287, 11; st.local.u32 [%rd730+12], %r287; ld.local.f32 %f1740, [%rd730+40]; fma.rn.ftz.f32 %f302, %f1740, %f1740, 0fBF800000; mov.f32 %f1725, 0f3F800000; abs.ftz.f32 %f303, %f1725; fma.rn.ftz.f32 %f304, %f303, %f303, %f302; setp.leu.ftz.f32 %p204, %f304, 0f00000000; mov.f32 %f3018, %f1725; @%p204 bra $L__BB184_252; sqrt.approx.ftz.f32 %f1741, %f304; sub.ftz.f32 %f1742, %f1741, %f303; add.ftz.f32 %f1743, %f303, %f1741; div.approx.ftz.f32 %f1744, %f1742, %f1743; fma.rn.ftz.f32 %f1745, %f303, %f1743, 0fBF800000; fma.rn.ftz.f32 %f1746, %f303, %f1742, 0f3F800000; div.approx.ftz.f32 %f1747, %f1745, %f1746; mul.ftz.f32 %f1748, %f1744, 0f3F000000; mul.ftz.f32 %f1749, %f1744, %f1748; fma.rn.ftz.f32 %f1750, %f1747, %f1747, 0f3F800000; mul.ftz.f32 %f3018, %f1749, %f1750; $L__BB184_252: ld.local.u64 %rd419, [%rd85+28]; add.s64 %rd86, %rd419, 12; ld.f32 %f307, [%rd419+12]; ld.f32 %f308, [%rd419+16]; ld.f32 %f309, [%rd419+20]; sub.ftz.f32 %f310, %f1725, %f3018; ld.local.f32 %f1752, [%rd85]; ld.local.f32 %f1753, [%rd4+24]; ld.local.f32 %f1754, [%rd85+4]; ld.local.f32 %f1755, [%rd4+28]; mul.ftz.f32 %f1756, %f1755, %f1754; fma.rn.ftz.f32 %f1757, %f1753, %f1752, %f1756; ld.local.f32 %f1758, [%rd85+8]; ld.local.f32 %f1759, [%rd4+32]; fma.rn.ftz.f32 %f1760, %f1759, %f1758, %f1757; abs.ftz.f32 %f311, %f1760; fma.rn.ftz.f32 %f312, %f311, %f311, %f302; setp.leu.ftz.f32 %p205, %f312, 0f00000000; mov.f32 %f3019, %f310; @%p205 bra $L__BB184_254; sqrt.approx.ftz.f32 %f1761, %f312; sub.ftz.f32 %f1762, %f1761, %f311; add.ftz.f32 %f1763, %f311, %f1761; div.approx.ftz.f32 %f1764, %f1762, %f1763; fma.rn.ftz.f32 %f1765, %f311, %f1763, 0fBF800000; fma.rn.ftz.f32 %f1766, %f311, %f1762, 0f3F800000; div.approx.ftz.f32 %f1767, %f1765, %f1766; mul.ftz.f32 %f1768, %f1764, 0f3F000000; mul.ftz.f32 %f1769, %f1764, %f1768; fma.rn.ftz.f32 %f1770, %f1767, %f1767, 0f3F800000; mul.ftz.f32 %f1771, %f1769, %f1770; sub.ftz.f32 %f3019, %f1771, %f3018; $L__BB184_254: rcp.approx.ftz.f32 %f1772, %f310; mov.f32 %f1773, 0f3F800000; mul.ftz.f32 %f1774, %f1772, %f3019; sub.ftz.f32 %f1775, %f1773, %f1774; fma.rn.ftz.f32 %f1776, %f307, %f1775, %f1774; fma.rn.ftz.f32 %f1777, %f308, %f1775, %f1774; fma.rn.ftz.f32 %f1778, %f309, %f1775, %f1774; st.f32 [%rd86+12], %f1776; st.f32 [%rd86+16], %f1777; st.f32 [%rd86+20], %f1778; ld.local.u32 %r288, [%rd85+-8]; setp.ne.s32 %p206, %r288, 12; @%p206 bra $L__BB184_256; ld.local.u64 %rd420, [%rd85+28]; ld.f32 %f1779, [%rd420+36]; mul.ftz.f32 %f1780, %f1779, 0f3E800000; ld.f32 %f1781, [%rd420+24]; mul.ftz.f32 %f1782, %f1781, %f1780; ld.f32 %f1783, [%rd420+28]; mul.ftz.f32 %f1784, %f1780, %f1783; ld.f32 %f1785, [%rd420+32]; mul.ftz.f32 %f1786, %f1780, %f1785; st.f32 [%rd420+24], %f1782; st.f32 [%rd420+28], %f1784; st.f32 [%rd420+32], %f1786; $L__BB184_256: ld.local.u64 %rd421, [%rd85+28]; ld.f32 %f1787, [%rd421+28]; ld.f32 %f1788, [%rd421+24]; add.ftz.f32 %f1789, %f1788, %f1787; ld.f32 %f1790, [%rd421+32]; add.ftz.f32 %f1791, %f1790, %f1789; mul.ftz.f32 %f1792, %f1791, 0f3EAAAAAB; ld.local.f32 %f1793, [%rd85+-4]; mul.ftz.f32 %f1794, %f1793, %f1792; st.local.f32 [%rd85+-4], %f1794; ld.local.u32 %r289, [%rd4+40]; or.b32 %r290, %r289, 12; st.local.u32 [%rd4+40], %r290; $L__BB184_257: mov.f32 %f1795, 0f3F800000; sub.ftz.f32 %f1796, %f1795, %f2976; mov.f32 %f1797, 0f38D1B717; max.ftz.f32 %f1798, %f1797, %f1796; mul.ftz.f32 %f1799, %f2977, %f275; mul.ftz.f32 %f1800, %f1799, %f1798; mul.ftz.f32 %f1801, %f2978, %f276; mul.ftz.f32 %f1802, %f1801, %f1798; mul.ftz.f32 %f1803, %f2979, %f277; mul.ftz.f32 %f1804, %f1803, %f1798; max.ftz.f32 %f315, %f1800, %f1719; max.ftz.f32 %f316, %f1802, %f1719; max.ftz.f32 %f317, %f1804, %f1719; add.ftz.f32 %f1806, %f315, %f316; add.ftz.f32 %f1807, %f317, %f1806; mul.ftz.f32 %f1808, %f1807, 0f3EAAAAAB; abs.ftz.f32 %f318, %f1808; setp.ltu.ftz.f32 %p207, %f318, 0f3727C5AC; mov.u64 %rd736, 0; mov.u64 %rd737, %rd736; @%p207 bra $L__BB184_262; ld.local.u32 %r50, [%rd4+252]; mov.u64 %rd736, 0; setp.eq.s32 %p208, %r50, 0; mov.u64 %rd734, %rd736; mov.u64 %rd735, %rd736; @%p208 bra $L__BB184_260; ld.local.u32 %r291, [%rd4+248]; mul.wide.s32 %rd426, %r291, 80; add.s64 %rd427, %rd1, %rd426; add.s64 %rd734, %rd427, 304; add.s64 %rd428, %rd230, %rd426; add.s64 %rd735, %rd428, 304; mov.b32 %r292, %f317; mov.b32 %r293, %f316; mov.b32 %r294, %f315; mov.u32 %r295, 0; st.local.v4.u32 [%rd427+304], {%r294, %r293, %r292, %r295}; add.s32 %r296, %r291, 1; st.local.u32 [%rd4+248], %r296; add.s32 %r297, %r50, -1; st.local.u32 [%rd4+252], %r297; $L__BB184_260: setp.eq.s64 %p209, %rd735, 0; mov.u64 %rd737, %rd736; @%p209 bra $L__BB184_262; st.local.f32 [%rd734+16], %f318; mov.u64 %rd736, %rd734; mov.u64 %rd737, %rd735; $L__BB184_262: setp.eq.s64 %p210, %rd737, 0; @%p210 bra $L__BB184_264; mov.u64 %rd431, 0; st.local.f32 [%rd736+20], %f55; st.local.f32 [%rd736+24], %f56; st.local.f32 [%rd736+28], %f57; mov.u32 %r298, 0; st.local.v2.f32 [%rd736+56], {%f1719, %f1719}; st.local.u32 [%rd736+64], %r298; sub.ftz.f32 %f1811, %f1795, %f79; sub.ftz.f32 %f1812, %f1795, %f71; mul.ftz.f32 %f1813, %f1812, %f1811; sub.ftz.f32 %f1814, %f1795, %f1813; selp.f32 %f1815, %f1814, %f71, %p182; mul.ftz.f32 %f1816, %f1815, %f1815; st.local.f32 [%rd736+40], %f2975; st.local.u64 [%rd736+48], %rd431; cvt.ftz.sat.f32.f32 %f1817, %f1816; st.local.v2.f32 [%rd736+32], {%f1817, %f1817}; mov.u32 %r299, 23; st.local.u32 [%rd736+12], %r299; ld.local.u32 %r300, [%rd4+40]; or.b32 %r301, %r300, 12; st.local.u32 [%rd4+40], %r301; bra.uni $L__BB184_264; $L__BB184_231: ld.local.u32 %r267, [%rd4+248]; add.s32 %r268, %r267, -1; st.local.u32 [%rd4+248], %r268; add.s32 %r269, %r47, 1; st.local.u32 [%rd4+252], %r269; mov.u64 %rd727, %rd726; $L__BB184_232: setp.eq.s64 %p190, %rd727, 0; or.pred %p192, %p188, %p190; @%p192 bra $L__BB184_264; add.s64 %rd73, %rd724, 20; st.local.f32 [%rd724+20], %f55; st.local.f32 [%rd724+24], %f56; st.local.f32 [%rd724+28], %f57; st.local.u64 [%rd724+48], %rd727; mov.u32 %r270, 0; st.local.v2.f32 [%rd724+56], {%f1638, %f1638}; st.local.u32 [%rd724+64], %r270; mul.ftz.f32 %f1644, %f71, %f71; st.local.f32 [%rd724+32], %f1644; st.local.f32 [%rd724+40], %f2975; st.local.f32 [%rd726], %f2977; st.local.f32 [%rd726+4], %f2978; st.local.f32 [%rd726+8], %f2979; ld.local.u64 %rd390, [%rd724+48]; st.f32 [%rd390+12], %f278; st.f32 [%rd390+16], %f279; st.f32 [%rd390+20], %f280; ld.local.u64 %rd391, [%rd724+48]; st.u32 [%rd391+36], %r270; ld.local.f32 %f1645, [%rd724+32]; mov.f32 %f1646, 0f38D1B717; max.ftz.f32 %f1647, %f1645, %f1646; min.ftz.f32 %f1648, %f1647, %f1636; st.local.f32 [%rd724+32], %f1648; st.local.f32 [%rd724+36], %f1648; ld.local.f32 %f1649, [%rd724+40]; max.ftz.f32 %f1650, %f1638, %f1649; st.local.f32 [%rd724+40], %f1650; ld.local.u64 %rd392, [%rd724+48]; ld.f32 %f1651, [%rd392]; cvt.ftz.sat.f32.f32 %f1652, %f1651; ld.f32 %f1653, [%rd392+4]; cvt.ftz.sat.f32.f32 %f1654, %f1653; ld.f32 %f1655, [%rd392+8]; cvt.ftz.sat.f32.f32 %f1656, %f1655; st.f32 [%rd392], %f1652; st.f32 [%rd392+4], %f1654; st.f32 [%rd392+8], %f1656; ld.local.u64 %rd393, [%rd724+48]; ld.f32 %f1657, [%rd393+12]; cvt.ftz.sat.f32.f32 %f1658, %f1657; ld.f32 %f1659, [%rd393+16]; cvt.ftz.sat.f32.f32 %f1660, %f1659; ld.f32 %f1661, [%rd393+20]; cvt.ftz.sat.f32.f32 %f1662, %f1661; st.f32 [%rd393+12], %f1658; st.f32 [%rd393+16], %f1660; st.f32 [%rd393+20], %f1662; mov.u32 %r271, 27; st.local.u32 [%rd724+12], %r271; ld.local.f32 %f1663, [%rd724+40]; fma.rn.ftz.f32 %f285, %f1663, %f1663, 0fBF800000; abs.ftz.f32 %f286, %f1636; fma.rn.ftz.f32 %f287, %f286, %f286, %f285; setp.leu.ftz.f32 %p193, %f287, 0f00000000; mov.f32 %f3016, %f1636; @%p193 bra $L__BB184_235; sqrt.approx.ftz.f32 %f1664, %f287; sub.ftz.f32 %f1665, %f1664, %f286; add.ftz.f32 %f1666, %f286, %f1664; div.approx.ftz.f32 %f1667, %f1665, %f1666; fma.rn.ftz.f32 %f1668, %f286, %f1666, 0fBF800000; fma.rn.ftz.f32 %f1669, %f286, %f1665, 0f3F800000; div.approx.ftz.f32 %f1670, %f1668, %f1669; mul.ftz.f32 %f1671, %f1667, 0f3F000000; mul.ftz.f32 %f1672, %f1667, %f1671; fma.rn.ftz.f32 %f1673, %f1670, %f1670, 0f3F800000; mul.ftz.f32 %f3016, %f1672, %f1673; $L__BB184_235: ld.local.u64 %rd394, [%rd73+28]; add.s64 %rd74, %rd394, 12; ld.f32 %f290, [%rd394+12]; ld.f32 %f291, [%rd394+16]; ld.f32 %f292, [%rd394+20]; sub.ftz.f32 %f293, %f1636, %f3016; ld.local.f32 %f1675, [%rd73]; ld.local.f32 %f1676, [%rd4+24]; ld.local.f32 %f1677, [%rd73+4]; ld.local.f32 %f1678, [%rd4+28]; mul.ftz.f32 %f1679, %f1678, %f1677; fma.rn.ftz.f32 %f1680, %f1676, %f1675, %f1679; ld.local.f32 %f1681, [%rd73+8]; ld.local.f32 %f1682, [%rd4+32]; fma.rn.ftz.f32 %f1683, %f1682, %f1681, %f1680; abs.ftz.f32 %f294, %f1683; fma.rn.ftz.f32 %f295, %f294, %f294, %f285; setp.leu.ftz.f32 %p194, %f295, 0f00000000; mov.f32 %f3017, %f293; @%p194 bra $L__BB184_237; sqrt.approx.ftz.f32 %f1684, %f295; sub.ftz.f32 %f1685, %f1684, %f294; add.ftz.f32 %f1686, %f294, %f1684; div.approx.ftz.f32 %f1687, %f1685, %f1686; fma.rn.ftz.f32 %f1688, %f294, %f1686, 0fBF800000; fma.rn.ftz.f32 %f1689, %f294, %f1685, 0f3F800000; div.approx.ftz.f32 %f1690, %f1688, %f1689; mul.ftz.f32 %f1691, %f1687, 0f3F000000; mul.ftz.f32 %f1692, %f1687, %f1691; fma.rn.ftz.f32 %f1693, %f1690, %f1690, 0f3F800000; mul.ftz.f32 %f1694, %f1692, %f1693; sub.ftz.f32 %f3017, %f1694, %f3016; $L__BB184_237: rcp.approx.ftz.f32 %f1695, %f293; mov.f32 %f1696, 0f3F800000; mul.ftz.f32 %f1697, %f1695, %f3017; sub.ftz.f32 %f1698, %f1696, %f1697; fma.rn.ftz.f32 %f1699, %f290, %f1698, %f1697; fma.rn.ftz.f32 %f1700, %f291, %f1698, %f1697; fma.rn.ftz.f32 %f1701, %f292, %f1698, %f1697; st.f32 [%rd74+12], %f1699; st.f32 [%rd74+16], %f1700; st.f32 [%rd74+20], %f1701; ld.local.u32 %r272, [%rd73+-8]; setp.ne.s32 %p195, %r272, 12; @%p195 bra $L__BB184_239; ld.local.u64 %rd395, [%rd73+28]; ld.f32 %f1702, [%rd395+36]; mul.ftz.f32 %f1703, %f1702, 0f3E800000; ld.f32 %f1704, [%rd395+24]; mul.ftz.f32 %f1705, %f1704, %f1703; ld.f32 %f1706, [%rd395+28]; mul.ftz.f32 %f1707, %f1703, %f1706; ld.f32 %f1708, [%rd395+32]; mul.ftz.f32 %f1709, %f1703, %f1708; st.f32 [%rd395+24], %f1705; st.f32 [%rd395+28], %f1707; st.f32 [%rd395+32], %f1709; $L__BB184_239: ld.local.u64 %rd396, [%rd73+28]; ld.f32 %f1710, [%rd396+28]; ld.f32 %f1711, [%rd396+24]; add.ftz.f32 %f1712, %f1711, %f1710; ld.f32 %f1713, [%rd396+32]; add.ftz.f32 %f1714, %f1713, %f1712; mul.ftz.f32 %f1715, %f1714, 0f3EAAAAAB; ld.local.f32 %f1716, [%rd73+-4]; mul.ftz.f32 %f1717, %f1716, %f1715; st.local.f32 [%rd73+-4], %f1717; ld.local.u32 %r273, [%rd4+40]; or.b32 %r274, %r273, 1036; st.local.u32 [%rd4+40], %r274; $L__BB184_264: ld.param.u32 %r605, [_Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; add.s32 %r629, %r605, 5; setp.leu.ftz.f32 %p212, %f76, 0f3727C5AC; @%p212 bra $L__BB184_499; mov.f32 %f1818, 0f00000000; max.ftz.f32 %f319, %f177, %f1818; max.ftz.f32 %f320, %f178, %f1818; add.ftz.f32 %f1819, %f319, %f320; max.ftz.f32 %f321, %f179, %f1818; add.ftz.f32 %f1820, %f321, %f1819; mul.ftz.f32 %f1821, %f1820, 0f3EAAAAAB; abs.ftz.f32 %f322, %f1821; setp.ltu.ftz.f32 %p213, %f322, 0f3727C5AC; mov.u64 %rd742, 0; mov.u64 %rd740, %rd742; mov.u64 %rd741, %rd742; @%p213 bra $L__BB184_270; ld.local.u32 %r52, [%rd4+252]; mov.u64 %rd740, 0; setp.eq.s32 %p214, %r52, 0; mov.u64 %rd738, %rd740; mov.u64 %rd739, %rd740; @%p214 bra $L__BB184_268; ld.local.u32 %r302, [%rd4+248]; mul.wide.s32 %rd436, %r302, 80; add.s64 %rd437, %rd1, %rd436; add.s64 %rd738, %rd437, 304; add.s64 %rd438, %rd230, %rd436; add.s64 %rd739, %rd438, 304; mov.b32 %r303, %f321; mov.b32 %r304, %f320; mov.b32 %r305, %f319; mov.u32 %r306, 0; st.local.v4.u32 [%rd437+304], {%r305, %r304, %r303, %r306}; add.s32 %r307, %r302, 1; st.local.u32 [%rd4+248], %r307; add.s32 %r308, %r52, -1; st.local.u32 [%rd4+252], %r308; $L__BB184_268: setp.eq.s64 %p215, %rd739, 0; mov.u64 %rd741, %rd740; @%p215 bra $L__BB184_270; st.local.f32 [%rd738+16], %f322; mov.u64 %rd740, %rd738; mov.u64 %rd741, %rd739; $L__BB184_270: setp.eq.s64 %p216, %rd741, 0; mov.u64 %rd743, %rd742; @%p216 bra $L__BB184_274; ld.local.u32 %r53, [%rd4+252]; setp.lt.s32 %p217, %r53, 1; @%p217 bra $L__BB184_273; add.s32 %r309, %r53, -1; st.local.u32 [%rd4+252], %r309; ld.local.s32 %rd443, [%rd4+248]; cvt.s64.s32 %rd444, %r309; add.s64 %rd445, %rd443, %rd444; mul.lo.s64 %rd446, %rd445, 80; add.s64 %rd447, %rd1, %rd446; add.s64 %rd742, %rd447, 304; add.s64 %rd448, %rd230, %rd446; add.s64 %rd743, %rd448, 304; bra.uni $L__BB184_274; $L__BB184_273: ld.local.u32 %r310, [%rd4+248]; add.s32 %r311, %r310, -1; st.local.u32 [%rd4+248], %r311; add.s32 %r312, %r53, 1; st.local.u32 [%rd4+252], %r312; mov.u64 %rd743, %rd742; $L__BB184_274: setp.eq.s64 %p218, %rd743, 0; or.pred %p220, %p216, %p218; @%p220 bra $L__BB184_499; mov.f32 %f3020, 0f3F800000; add.s64 %rd103, %rd740, 20; st.local.f32 [%rd740+20], %f2985; st.local.f32 [%rd740+24], %f2986; st.local.f32 [%rd740+28], %f2987; mov.u32 %r313, 0; st.local.v2.f32 [%rd740+56], {%f1818, %f1818}; st.local.u32 [%rd740+64], %r313; mov.u32 %r314, 1069547520; st.local.u32 [%rd740+40], %r314; st.local.u64 [%rd740+48], %rd743; mul.ftz.f32 %f1824, %f77, %f77; st.local.v2.f32 [%rd740+32], {%f1824, %f1824}; st.local.u32 [%rd742], %r313; st.local.u32 [%rd742+4], %r313; st.local.u32 [%rd742+8], %r313; ld.local.u64 %rd451, [%rd740+48]; mov.u32 %r315, 1025758986; st.u32 [%rd451+12], %r315; st.u32 [%rd451+16], %r315; st.u32 [%rd451+20], %r315; ld.local.u64 %rd452, [%rd740+48]; st.f32 [%rd452+36], %f76; ld.local.u64 %rd453, [%rd740+48]; ld.f32 %f1825, [%rd453+12]; cvt.ftz.sat.f32.f32 %f1826, %f1825; ld.f32 %f1827, [%rd453+16]; cvt.ftz.sat.f32.f32 %f1828, %f1827; ld.f32 %f1829, [%rd453+20]; cvt.ftz.sat.f32.f32 %f1830, %f1829; st.f32 [%rd453+12], %f1826; st.f32 [%rd453+16], %f1828; st.f32 [%rd453+20], %f1830; ld.local.f32 %f1831, [%rd740+32]; cvt.ftz.sat.f32.f32 %f1832, %f1831; st.local.v2.f32 [%rd740+32], {%f1832, %f1832}; mov.u32 %r316, 12; st.local.u32 [%rd740+12], %r316; ld.local.f32 %f1833, [%rd740+40]; fma.rn.ftz.f32 %f323, %f1833, %f1833, 0fBF800000; abs.ftz.f32 %f324, %f3020; fma.rn.ftz.f32 %f325, %f324, %f324, %f323; setp.leu.ftz.f32 %p221, %f325, 0f00000000; @%p221 bra $L__BB184_277; sqrt.approx.ftz.f32 %f1834, %f325; sub.ftz.f32 %f1835, %f1834, %f324; add.ftz.f32 %f1836, %f324, %f1834; div.approx.ftz.f32 %f1837, %f1835, %f1836; fma.rn.ftz.f32 %f1838, %f324, %f1836, 0fBF800000; fma.rn.ftz.f32 %f1839, %f324, %f1835, 0f3F800000; div.approx.ftz.f32 %f1840, %f1838, %f1839; mul.ftz.f32 %f1841, %f1837, 0f3F000000; mul.ftz.f32 %f1842, %f1837, %f1841; fma.rn.ftz.f32 %f1843, %f1840, %f1840, 0f3F800000; mul.ftz.f32 %f3020, %f1842, %f1843; $L__BB184_277: mov.f32 %f2955, 0f3F800000; ld.local.u64 %rd454, [%rd103+28]; add.s64 %rd104, %rd454, 12; ld.f32 %f328, [%rd454+12]; ld.f32 %f329, [%rd454+16]; ld.f32 %f330, [%rd454+20]; sub.ftz.f32 %f331, %f2955, %f3020; ld.local.f32 %f1845, [%rd103]; ld.local.f32 %f1846, [%rd4+24]; ld.local.f32 %f1847, [%rd103+4]; ld.local.f32 %f1848, [%rd4+28]; mul.ftz.f32 %f1849, %f1848, %f1847; fma.rn.ftz.f32 %f1850, %f1846, %f1845, %f1849; ld.local.f32 %f1851, [%rd103+8]; ld.local.f32 %f1852, [%rd4+32]; fma.rn.ftz.f32 %f1853, %f1852, %f1851, %f1850; abs.ftz.f32 %f332, %f1853; fma.rn.ftz.f32 %f333, %f332, %f332, %f323; setp.leu.ftz.f32 %p222, %f333, 0f00000000; mov.f32 %f3021, %f331; @%p222 bra $L__BB184_279; sqrt.approx.ftz.f32 %f1854, %f333; sub.ftz.f32 %f1855, %f1854, %f332; add.ftz.f32 %f1856, %f332, %f1854; div.approx.ftz.f32 %f1857, %f1855, %f1856; fma.rn.ftz.f32 %f1858, %f332, %f1856, 0fBF800000; fma.rn.ftz.f32 %f1859, %f332, %f1855, 0f3F800000; div.approx.ftz.f32 %f1860, %f1858, %f1859; mul.ftz.f32 %f1861, %f1857, 0f3F000000; mul.ftz.f32 %f1862, %f1857, %f1861; fma.rn.ftz.f32 %f1863, %f1860, %f1860, 0f3F800000; mul.ftz.f32 %f1864, %f1862, %f1863; sub.ftz.f32 %f3021, %f1864, %f3020; $L__BB184_279: rcp.approx.ftz.f32 %f1865, %f331; mov.f32 %f1866, 0f3F800000; mul.ftz.f32 %f1867, %f1865, %f3021; sub.ftz.f32 %f1868, %f1866, %f1867; fma.rn.ftz.f32 %f1869, %f328, %f1868, %f1867; fma.rn.ftz.f32 %f1870, %f329, %f1868, %f1867; fma.rn.ftz.f32 %f1871, %f330, %f1868, %f1867; st.f32 [%rd104+12], %f1869; st.f32 [%rd104+16], %f1870; st.f32 [%rd104+20], %f1871; ld.local.u32 %r317, [%rd103+-8]; setp.ne.s32 %p223, %r317, 12; @%p223 bra $L__BB184_281; ld.local.u64 %rd455, [%rd103+28]; ld.f32 %f1872, [%rd455+36]; mul.ftz.f32 %f1873, %f1872, 0f3E800000; ld.f32 %f1874, [%rd455+24]; mul.ftz.f32 %f1875, %f1874, %f1873; ld.f32 %f1876, [%rd455+28]; mul.ftz.f32 %f1877, %f1873, %f1876; ld.f32 %f1878, [%rd455+32]; mul.ftz.f32 %f1879, %f1873, %f1878; st.f32 [%rd455+24], %f1875; st.f32 [%rd455+28], %f1877; st.f32 [%rd455+32], %f1879; $L__BB184_281: ld.local.u64 %rd456, [%rd103+28]; ld.f32 %f1880, [%rd456+28]; ld.f32 %f1881, [%rd456+24]; add.ftz.f32 %f1882, %f1881, %f1880; ld.f32 %f1883, [%rd456+32]; add.ftz.f32 %f1884, %f1883, %f1882; mul.ftz.f32 %f1885, %f1884, 0f3EAAAAAB; ld.local.f32 %f1886, [%rd103+-4]; mul.ftz.f32 %f1887, %f1886, %f1885; st.local.f32 [%rd103+-4], %f1887; ld.local.u32 %r318, [%rd4+40]; or.b32 %r319, %r318, 12; st.local.u32 [%rd4+40], %r319; bra.uni $L__BB184_499; } .func _Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB185_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB185_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB185_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB185_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB185_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB185_11; $L__BB185_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB185_8; add.s32 %r237, %r237, 2; bra.uni $L__BB185_9; $L__BB185_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB185_15; $L__BB185_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB185_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB185_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB185_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB185_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB185_15; $L__BB185_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB185_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB185_17; $L__BB185_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB185_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB185_159; bra.uni $L__BB185_18; $L__BB185_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB185_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB185_185; $L__BB185_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB185_150; bra.uni $L__BB185_19; $L__BB185_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB185_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB185_153; bra.uni $L__BB185_152; $L__BB185_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB185_154; $L__BB185_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB185_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB185_100; bra.uni $L__BB185_21; $L__BB185_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB185_111; bra.uni $L__BB185_101; $L__BB185_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB185_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB185_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB185_120; $L__BB185_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB185_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB185_164; bra.uni $L__BB185_163; $L__BB185_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB185_165; $L__BB185_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB185_136; bra.uni $L__BB185_126; $L__BB185_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB185_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB185_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB185_145; $L__BB185_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB185_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB185_158; bra.uni $L__BB185_155; $L__BB185_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB185_192; $L__BB185_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB185_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB185_192; $L__BB185_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB185_47; bra.uni $L__BB185_22; $L__BB185_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB185_86; bra.uni $L__BB185_48; $L__BB185_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB185_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB185_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB185_95; $L__BB185_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB185_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB185_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB185_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB185_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB185_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB185_184; bra.uni $L__BB185_170; $L__BB185_184: { // callseq 952, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 952 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB185_185; $L__BB185_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB185_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB185_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 946, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 946 bra.uni $L__BB185_145; $L__BB185_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB185_192; $L__BB185_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB185_183; bra.uni $L__BB185_171; $L__BB185_183: { // callseq 951, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 951 mov.f32 %f731, 0f3F800000; bra.uni $L__BB185_185; $L__BB185_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB185_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB185_104; mov.u64 %rd173, 0; { // callseq 945, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 945 bra.uni $L__BB185_120; $L__BB185_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB185_33; bra.uni $L__BB185_23; $L__BB185_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB185_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB185_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB185_42; $L__BB185_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB185_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB185_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB185_133; $L__BB185_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB185_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB185_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB185_108; $L__BB185_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB185_182; bra.uni $L__BB185_172; $L__BB185_182: { // callseq 950, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 950 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB185_185; $L__BB185_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB185_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB185_51; mov.u64 %rd129, 0; { // callseq 944, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 944 bra.uni $L__BB185_95; $L__BB185_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB185_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB185_26; mov.u64 %rd90, 0; { // callseq 943, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 943 bra.uni $L__BB185_42; $L__BB185_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB185_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB185_145; $L__BB185_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB185_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB185_120; $L__BB185_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB185_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB185_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB185_83; $L__BB185_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB185_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB185_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB185_30; $L__BB185_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB185_181; bra.uni $L__BB185_173; $L__BB185_181: { // callseq 949, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 949 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB185_185; $L__BB185_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB185_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB185_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB185_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB185_145; $L__BB185_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB185_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB185_145; $L__BB185_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB185_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB185_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB185_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB185_120; $L__BB185_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB185_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB185_120; $L__BB185_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB185_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB185_95; $L__BB185_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB185_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB185_42; $L__BB185_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB185_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB185_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB185_149; bra.uni $L__BB185_146; $L__BB185_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB185_192; $L__BB185_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB185_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB185_192; $L__BB185_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB185_192; $L__BB185_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB185_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB185_175; bra.uni $L__BB185_178; $L__BB185_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB185_177; bra.uni $L__BB185_176; $L__BB185_177: { // callseq 947, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 947 bra.uni $L__BB185_185; $L__BB185_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB185_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB185_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB185_124; bra.uni $L__BB185_121; $L__BB185_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB185_192; $L__BB185_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB185_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB185_192; $L__BB185_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB185_192; $L__BB185_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB185_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB185_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB185_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB185_95; $L__BB185_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB185_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB185_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB185_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB185_42; $L__BB185_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB185_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB185_95; $L__BB185_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB185_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB185_42; $L__BB185_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB185_82; bra.uni $L__BB185_55; $L__BB185_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB185_83; $L__BB185_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB185_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB185_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB185_46; bra.uni $L__BB185_43; $L__BB185_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB185_192; $L__BB185_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB185_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB185_192; $L__BB185_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB185_192; $L__BB185_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB185_185; $L__BB185_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB185_57; bra.uni $L__BB185_56; $L__BB185_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB185_58; $L__BB185_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB185_180; bra.uni $L__BB185_179; $L__BB185_180: { // callseq 948, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 948 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB185_185; $L__BB185_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB185_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB185_60; bra.uni $L__BB185_59; $L__BB185_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB185_61; $L__BB185_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB185_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB185_63; bra.uni $L__BB185_62; $L__BB185_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB185_64; $L__BB185_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB185_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB185_66; bra.uni $L__BB185_65; $L__BB185_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB185_67; $L__BB185_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB185_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB185_69; bra.uni $L__BB185_68; $L__BB185_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB185_70; $L__BB185_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB185_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB185_72; bra.uni $L__BB185_71; $L__BB185_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB185_73; $L__BB185_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB185_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB185_75; bra.uni $L__BB185_74; $L__BB185_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB185_76; $L__BB185_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB185_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB185_78; bra.uni $L__BB185_77; $L__BB185_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB185_79; $L__BB185_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB185_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB185_81; bra.uni $L__BB185_80; $L__BB185_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB185_83; $L__BB185_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB185_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB185_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB185_99; bra.uni $L__BB185_96; $L__BB185_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB185_192; $L__BB185_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB185_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB185_192; $L__BB185_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB185_192; $L__BB185_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB185_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB185_191; bra.uni $L__BB185_186; $L__BB185_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB185_192; $L__BB185_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB185_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB185_192; $L__BB185_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB185_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB185_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB185_192: ret; } .func _Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB186_2; bra.uni $L__BB186_1; $L__BB186_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB186_3; $L__BB186_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB186_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB186_5; bra.uni $L__BB186_4; $L__BB186_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB186_9; bra.uni $L__BB186_6; $L__BB186_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB186_12; bra.uni $L__BB186_10; $L__BB186_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB186_13; $L__BB186_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB186_16; $L__BB186_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB186_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB186_13; $L__BB186_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB186_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB186_15; bra.uni $L__BB186_14; $L__BB186_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB186_16; $L__BB186_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB186_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB186_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB186_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB186_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB187_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB187_5; bra.uni $L__BB187_2; $L__BB187_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB187_8; bra.uni $L__BB187_6; $L__BB187_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB187_9; $L__BB187_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB187_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB187_9; $L__BB187_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB187_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB187_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB187_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB187_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB187_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB187_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB187_20; $L__BB187_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB187_17; add.s32 %r172, %r172, 2; bra.uni $L__BB187_18; $L__BB187_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB187_24; $L__BB187_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB187_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB187_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB187_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB187_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB187_24; $L__BB187_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB187_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB187_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB187_37; bra.uni $L__BB187_26; $L__BB187_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB187_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB187_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB187_45; $L__BB187_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB187_45; $L__BB187_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB187_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB187_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB187_30; mov.u64 %rd108, 0; { // callseq 953, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 953 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB187_45; $L__BB187_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB187_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB187_9; $L__BB187_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB187_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB187_45; $L__BB187_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB187_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB187_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB187_34; $L__BB187_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB187_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB187_45; $L__BB187_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB187_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB187_45; $L__BB187_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB187_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB187_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB187_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB187_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB187_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB187_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB187_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB187_56; $L__BB187_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB187_53; add.s32 %r187, %r187, 2; bra.uni $L__BB187_54; $L__BB187_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB187_60; $L__BB187_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB187_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB187_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB187_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB187_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB187_60; $L__BB187_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB187_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB187_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB187_72; bra.uni $L__BB187_62; $L__BB187_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB187_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB187_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB187_79; $L__BB187_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB187_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB187_65; mov.u64 %rd160, 0; { // callseq 954, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 954 bra.uni $L__BB187_79; $L__BB187_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB187_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB187_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB187_69; $L__BB187_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB187_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB187_79; $L__BB187_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB187_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB187_79; $L__BB187_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB187_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB187_79; $L__BB187_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB187_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB187_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB187_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB187_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB187_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB187_84; bra.uni $L__BB187_83; $L__BB187_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB187_85; $L__BB187_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB187_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB188_2; bra.uni $L__BB188_1; $L__BB188_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB188_3; $L__BB188_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB188_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB188_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB188_8; bra.uni $L__BB188_5; $L__BB188_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB188_11; bra.uni $L__BB188_9; $L__BB188_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB188_12; $L__BB188_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB188_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB188_12; $L__BB188_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB188_12: @%p8 bra $L__BB188_14; bra.uni $L__BB188_13; $L__BB188_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB188_15; $L__BB188_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB188_15: @%p8 bra $L__BB188_17; bra.uni $L__BB188_16; $L__BB188_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB188_18; $L__BB188_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB188_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB188_19; bra.uni $L__BB188_20; $L__BB188_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB188_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB188_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB188_22: @%p7 bra $L__BB188_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB188_25; bra.uni $L__BB188_24; $L__BB188_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB188_26; $L__BB188_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB188_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB188_27; bra.uni $L__BB188_34; $L__BB188_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB188_30; bra.uni $L__BB188_28; $L__BB188_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB188_32; bra.uni $L__BB188_31; $L__BB188_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB188_33; $L__BB188_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB188_29; bra.uni $L__BB188_34; $L__BB188_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB188_33; $L__BB188_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB188_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB188_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB188_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB188_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB188_12; } .func _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b64 _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b32 _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4 ) { .reg .pred %p<7>; .reg .f32 %f<8>; .reg .b32 %r<5>; .reg .b64 %rd<5>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.f32 %f7, [_Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.u64 %rd1, [_Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; mov.f32 %f4, 0f00000000; setp.gt.s32 %p1, %r2, 8; @%p1 bra $L__BB189_4; setp.eq.s32 %p4, %r2, 1; @%p4 bra $L__BB189_8; setp.eq.s32 %p5, %r2, 8; mov.f32 %f7, %f4; @%p5 bra $L__BB189_3; bra.uni $L__BB189_7; $L__BB189_3: and.b32 %r4, %r1, 1; setp.eq.b32 %p6, %r4, 1; selp.f32 %f7, 0f3F800000, 0f00000000, %p6; bra.uni $L__BB189_7; $L__BB189_4: setp.eq.s32 %p2, %r2, 9; @%p2 bra $L__BB189_7; setp.eq.s32 %p3, %r2, 10; mov.f32 %f7, %f4; @%p3 bra $L__BB189_6; bra.uni $L__BB189_7; $L__BB189_6: mov.f32 %f7, 0f3F800000; bra.uni $L__BB189_7; $L__BB189_8: mov.f32 %f7, 0f3F800000; $L__BB189_7: cvta.to.local.u64 %rd2, %rd1; mul.wide.u32 %rd3, %r3, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f7; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB190_2; bra.uni $L__BB190_1; $L__BB190_2: mov.b32 %f2764, %r395; bra.uni $L__BB190_3; $L__BB190_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB190_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB190_5; bra.uni $L__BB190_4; $L__BB190_5: mov.b32 %f9, %r396; bra.uni $L__BB190_6; $L__BB190_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB190_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB190_8; bra.uni $L__BB190_7; $L__BB190_8: mov.b32 %f2766, %r397; bra.uni $L__BB190_9; $L__BB190_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB190_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB190_11; bra.uni $L__BB190_10; $L__BB190_11: mov.b32 %f15, %r398; bra.uni $L__BB190_12; $L__BB190_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB190_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB190_14; bra.uni $L__BB190_13; $L__BB190_14: mov.b32 %f2768, %r399; bra.uni $L__BB190_15; $L__BB190_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB190_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB190_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB190_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB190_18; bra.uni $L__BB190_385; $L__BB190_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB190_259; bra.uni $L__BB190_19; $L__BB190_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB190_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB190_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB190_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB190_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB190_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB190_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB190_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB190_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB190_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB190_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB190_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB190_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB190_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB190_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB190_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB190_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB190_362; $L__BB190_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB190_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB190_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB190_40; bra.uni $L__BB190_29; $L__BB190_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB190_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB190_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB190_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB190_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB190_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB190_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB190_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB190_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB190_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB190_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB190_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB190_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB190_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB190_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB190_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB190_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB190_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB190_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB190_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB190_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB190_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB190_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB190_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB190_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB190_136; $L__BB190_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB190_371; bra.uni $L__BB190_366; $L__BB190_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB190_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB190_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB190_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB190_383; $L__BB190_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB190_149; bra.uni $L__BB190_139; $L__BB190_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB190_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB190_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB190_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB190_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB190_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB190_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB190_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB190_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB190_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB190_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB190_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB190_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB190_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB190_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB190_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB190_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB190_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB190_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB190_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB190_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB190_250; $L__BB190_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB190_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB190_253; bra.uni $L__BB190_362; $L__BB190_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB190_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB190_304; @%p171 bra $L__BB190_303; bra.uni $L__BB190_256; $L__BB190_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB190_305; $L__BB190_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB190_305; $L__BB190_256: @%p172 bra $L__BB190_302; bra.uni $L__BB190_257; $L__BB190_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB190_305; $L__BB190_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB190_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB190_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB190_308; bra.uni $L__BB190_306; $L__BB190_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB190_309; $L__BB190_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB190_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB190_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB190_316; @%p171 bra $L__BB190_315; bra.uni $L__BB190_311; $L__BB190_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB190_317; $L__BB190_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB190_317; $L__BB190_311: @%p172 bra $L__BB190_314; bra.uni $L__BB190_312; $L__BB190_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB190_317; $L__BB190_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB190_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB190_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB190_320; bra.uni $L__BB190_318; $L__BB190_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB190_321; $L__BB190_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB190_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB190_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB190_328; @%p171 bra $L__BB190_327; bra.uni $L__BB190_323; $L__BB190_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB190_329; $L__BB190_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB190_329; $L__BB190_323: @%p172 bra $L__BB190_326; bra.uni $L__BB190_324; $L__BB190_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB190_329; $L__BB190_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB190_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB190_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB190_332; bra.uni $L__BB190_330; $L__BB190_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB190_333; $L__BB190_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB190_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB190_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB190_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_362; $L__BB190_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB190_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB190_31; bra.uni $L__BB190_136; $L__BB190_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB190_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB190_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB190_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB190_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB190_107; @%p49 bra $L__BB190_106; bra.uni $L__BB190_37; $L__BB190_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB190_108; $L__BB190_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB190_108; $L__BB190_37: @%p50 bra $L__BB190_105; bra.uni $L__BB190_38; $L__BB190_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB190_108; $L__BB190_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB190_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB190_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB190_111; bra.uni $L__BB190_109; $L__BB190_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB190_112; $L__BB190_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB190_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB190_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB190_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB190_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB190_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB190_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_136; $L__BB190_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB190_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB190_368; bra.uni $L__BB190_383; $L__BB190_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB190_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB190_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB190_383; $L__BB190_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB190_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB190_141; bra.uni $L__BB190_250; $L__BB190_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB190_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB190_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB190_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB190_206; @%p101 bra $L__BB190_205; bra.uni $L__BB190_146; $L__BB190_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB190_207; $L__BB190_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB190_207; $L__BB190_146: @%p102 bra $L__BB190_204; bra.uni $L__BB190_147; $L__BB190_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB190_207; $L__BB190_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB190_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB190_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB190_210; bra.uni $L__BB190_208; $L__BB190_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB190_211; $L__BB190_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB190_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB190_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB190_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB190_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB190_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_250; $L__BB190_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB190_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB190_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB190_295; @%p151 bra $L__BB190_294; bra.uni $L__BB190_290; $L__BB190_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB190_296; $L__BB190_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB190_296; $L__BB190_290: @%p152 bra $L__BB190_293; bra.uni $L__BB190_291; $L__BB190_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB190_296; $L__BB190_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB190_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB190_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB190_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB190_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB190_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB190_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB190_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_362; $L__BB190_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB190_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB190_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB190_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB190_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB190_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB190_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB190_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB190_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB190_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB190_279; bra.uni $L__BB190_362; $L__BB190_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB190_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB190_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB190_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB190_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB190_96; @%p39 bra $L__BB190_95; bra.uni $L__BB190_91; $L__BB190_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB190_97; $L__BB190_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB190_97; $L__BB190_91: @%p40 bra $L__BB190_94; bra.uni $L__BB190_92; $L__BB190_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB190_97; $L__BB190_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB190_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB190_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB190_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB190_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB190_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB190_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB190_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB190_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB190_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_136; $L__BB190_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB190_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB190_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB190_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB190_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB190_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB190_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB190_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB190_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB190_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB190_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB190_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB190_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB190_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB190_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB190_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB190_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB190_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB190_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB190_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB190_76; bra.uni $L__BB190_136; $L__BB190_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB190_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB190_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB190_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB190_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB190_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_383; $L__BB190_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB190_383; $L__BB190_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB190_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB190_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB190_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB190_196; @%p92 bra $L__BB190_195; bra.uni $L__BB190_191; $L__BB190_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB190_197; $L__BB190_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB190_197; $L__BB190_191: @%p93 bra $L__BB190_194; bra.uni $L__BB190_192; $L__BB190_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB190_197; $L__BB190_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB190_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB190_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB190_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB190_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB190_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB190_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB190_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB190_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB190_250; $L__BB190_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB190_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB190_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB190_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB190_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB190_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB190_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB190_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB190_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB190_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB190_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB190_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB190_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB190_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB190_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB190_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB190_179; bra.uni $L__BB190_250; $L__BB190_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB190_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB190_337; @%p188 bra $L__BB190_336; bra.uni $L__BB190_23; $L__BB190_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB190_338; $L__BB190_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB190_338; $L__BB190_23: @%p189 bra $L__BB190_335; bra.uni $L__BB190_24; $L__BB190_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB190_338; $L__BB190_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB190_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB190_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB190_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB190_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB190_347; @%p188 bra $L__BB190_346; bra.uni $L__BB190_342; $L__BB190_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB190_348; $L__BB190_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB190_348; $L__BB190_342: @%p189 bra $L__BB190_345; bra.uni $L__BB190_343; $L__BB190_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB190_348; $L__BB190_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB190_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB190_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB190_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB190_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB190_357; @%p188 bra $L__BB190_356; bra.uni $L__BB190_352; $L__BB190_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB190_358; $L__BB190_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB190_358; $L__BB190_352: @%p189 bra $L__BB190_355; bra.uni $L__BB190_353; $L__BB190_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB190_358; $L__BB190_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB190_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB190_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB190_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB190_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB190_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB190_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB190_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB190_364: mov.f32 %f3392, %f3387; bra.uni $L__BB190_385; $L__BB190_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB190_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB190_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB190_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB190_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB190_128; @%p59 bra $L__BB190_127; bra.uni $L__BB190_123; $L__BB190_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB190_129; $L__BB190_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB190_129; $L__BB190_123: @%p60 bra $L__BB190_126; bra.uni $L__BB190_124; $L__BB190_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB190_129; $L__BB190_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB190_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB190_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB190_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB190_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB190_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB190_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB190_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB190_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB190_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB190_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB190_385; $L__BB190_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB190_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB190_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB190_385; $L__BB190_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB190_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB190_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB190_224; @%p120 bra $L__BB190_223; bra.uni $L__BB190_219; $L__BB190_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB190_225; $L__BB190_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB190_225; $L__BB190_219: @%p121 bra $L__BB190_222; bra.uni $L__BB190_220; $L__BB190_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB190_225; $L__BB190_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB190_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB190_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB190_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB190_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB190_234; @%p120 bra $L__BB190_233; bra.uni $L__BB190_229; $L__BB190_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB190_235; $L__BB190_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB190_235; $L__BB190_229: @%p121 bra $L__BB190_232; bra.uni $L__BB190_230; $L__BB190_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB190_235; $L__BB190_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB190_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB190_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB190_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB190_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB190_244; @%p120 bra $L__BB190_243; bra.uni $L__BB190_239; $L__BB190_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB190_245; $L__BB190_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB190_245; $L__BB190_239: @%p121 bra $L__BB190_242; bra.uni $L__BB190_240; $L__BB190_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB190_245; $L__BB190_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB190_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB190_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB190_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB190_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB190_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB190_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB190_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB190_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB190_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB190_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB190_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB190_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB190_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB190_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB190_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB190_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB190_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB190_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB190_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB192_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB192_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB192_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB192_4: ret; } .func (.param .b32 func_retval0) _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji( .param .b64 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0, .param .b64 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1, .param .align 4 .b8 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2[16], .param .b32 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3, .param .b32 _Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4 ) { .reg .pred %p<463>; .reg .b16 %rs<22>; .reg .f32 %f<3060>; .reg .b32 %r<664>; .reg .b64 %rd<808>; ld.param.u64 %rd230, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_0]; ld.param.u64 %rd231, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_1]; ld.param.u32 %r141, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.param.u32 %r142, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; ld.param.u32 %r140, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+12]; ld.param.u32 %r139, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+8]; ld.param.u32 %r138, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_2+4]; cvta.to.local.u64 %rd1, %rd230; cvta.to.local.u64 %rd2, %rd231; and.b32 %r2, %r138, 255; bfe.u32 %r3, %r138, 8, 8; bfe.u32 %r4, %r138, 16, 8; shr.u32 %r5, %r138, 24; setp.eq.s32 %p7, %r5, 255; mov.f32 %f2, 0f3F800000; @%p7 bra $L__BB193_2; mul.wide.u32 %rd232, %r5, 4; add.s64 %rd233, %rd2, %rd232; ld.local.f32 %f2, [%rd233]; $L__BB193_2: ld.const.u64 %rd234, [kernel_params+2400]; cvta.to.global.u64 %rd235, %rd234; mul.wide.s32 %rd236, %r142, 16; add.s64 %rd3, %rd235, %rd236; ld.global.v4.u32 {%r143, %r144, %r145, %r146}, [%rd3]; add.s32 %r663, %r142, 1; setp.eq.ftz.f32 %p8, %f2, 0f00000000; @%p8 bra $L__BB193_516; bra.uni $L__BB193_3; $L__BB193_516: setp.eq.s32 %p461, %r2, 39; add.s32 %r636, %r142, 5; selp.b32 %r663, %r636, %r663, %p461; $L__BB193_517: st.param.b32 [func_retval0+0], %r663; ret; $L__BB193_3: setp.eq.s32 %p9, %r143, 255; add.s64 %rd4, %rd1, 12; @%p9 bra $L__BB193_5; bra.uni $L__BB193_4; $L__BB193_5: ld.local.f32 %f55, [%rd4]; ld.local.f32 %f56, [%rd4+4]; ld.local.f32 %f57, [%rd4+8]; bra.uni $L__BB193_6; $L__BB193_4: mul.wide.u32 %rd237, %r143, 4; add.s64 %rd238, %rd2, %rd237; ld.local.f32 %f55, [%rd238]; ld.local.f32 %f56, [%rd238+4]; ld.local.f32 %f57, [%rd238+8]; $L__BB193_6: ld.local.u32 %r11, [%rd4+52]; and.b32 %r12, %r11, 6; setp.ne.s32 %p10, %r12, 0; @%p10 bra $L__BB193_15; ld.local.v2.f32 {%f550, %f551}, [%rd4+12]; ld.local.v2.f32 {%f552, %f553}, [%rd4+20]; ld.local.v2.f32 {%f554, %f555}, [%rd4+28]; mul.ftz.f32 %f556, %f56, %f554; fma.rn.ftz.f32 %f557, %f55, %f553, %f556; fma.rn.ftz.f32 %f558, %f57, %f555, %f557; add.ftz.f32 %f559, %f558, %f558; mul.ftz.f32 %f560, %f55, %f559; mul.ftz.f32 %f561, %f56, %f559; mul.ftz.f32 %f562, %f57, %f559; sub.ftz.f32 %f563, %f560, %f553; sub.ftz.f32 %f564, %f561, %f554; sub.ftz.f32 %f565, %f562, %f555; mul.ftz.f32 %f566, %f551, %f554; fma.rn.ftz.f32 %f567, %f550, %f553, %f566; fma.rn.ftz.f32 %f18, %f552, %f555, %f567; mul.ftz.f32 %f568, %f18, 0f3F666666; mov.f32 %f569, 0f3C23D70A; min.ftz.f32 %f19, %f568, %f569; mul.ftz.f32 %f570, %f551, %f564; fma.rn.ftz.f32 %f571, %f550, %f563, %f570; fma.rn.ftz.f32 %f572, %f552, %f565, %f571; setp.ltu.ftz.f32 %p11, %f572, %f19; @%p11 bra $L__BB193_8; bra.uni $L__BB193_15; $L__BB193_8: mul.ftz.f32 %f573, %f56, %f551; fma.rn.ftz.f32 %f574, %f55, %f550, %f573; fma.rn.ftz.f32 %f575, %f57, %f552, %f574; mul.ftz.f32 %f576, %f550, %f575; mul.ftz.f32 %f577, %f551, %f575; mul.ftz.f32 %f578, %f552, %f575; sub.ftz.f32 %f579, %f55, %f576; sub.ftz.f32 %f580, %f56, %f577; sub.ftz.f32 %f581, %f57, %f578; mul.ftz.f32 %f582, %f580, %f580; fma.rn.ftz.f32 %f583, %f579, %f579, %f582; fma.rn.ftz.f32 %f584, %f581, %f581, %f583; sqrt.approx.ftz.f32 %f585, %f584; div.approx.ftz.f32 %f23, %f579, %f585; div.approx.ftz.f32 %f24, %f580, %f585; div.approx.ftz.f32 %f25, %f581, %f585; mul.ftz.f32 %f586, %f554, %f24; fma.rn.ftz.f32 %f587, %f553, %f23, %f586; fma.rn.ftz.f32 %f26, %f555, %f25, %f587; mul.ftz.f32 %f588, %f26, %f26; fma.rn.ftz.f32 %f589, %f18, %f18, %f588; mul.ftz.f32 %f590, %f19, %f19; sub.ftz.f32 %f591, %f589, %f590; mul.ftz.f32 %f592, %f588, %f591; mov.f32 %f593, 0f00000000; max.ftz.f32 %f594, %f592, %f593; sqrt.approx.ftz.f32 %f595, %f594; fma.rn.ftz.f32 %f596, %f19, %f18, %f589; mov.f32 %f597, 0f3F000000; div.approx.ftz.f32 %f598, %f597, %f589; add.ftz.f32 %f599, %f595, %f596; mul.ftz.f32 %f27, %f598, %f599; sub.ftz.f32 %f600, %f596, %f595; mul.ftz.f32 %f28, %f598, %f600; setp.gt.ftz.f32 %p12, %f27, 0f3727C5AC; setp.le.ftz.f32 %p13, %f27, 0f3F800054; and.pred %p1, %p12, %p13; setp.gt.ftz.f32 %p14, %f28, 0f3727C5AC; setp.le.ftz.f32 %p15, %f28, 0f3F800054; and.pred %p2, %p14, %p15; and.pred %p16, %p1, %p2; @%p16 bra $L__BB193_11; bra.uni $L__BB193_9; $L__BB193_11: mov.f32 %f607, 0f3F800000; sub.ftz.f32 %f608, %f607, %f27; max.ftz.f32 %f610, %f608, %f593; sqrt.approx.ftz.f32 %f34, %f610; max.ftz.f32 %f611, %f27, %f593; sqrt.approx.ftz.f32 %f35, %f611; sub.ftz.f32 %f612, %f607, %f28; max.ftz.f32 %f613, %f612, %f593; sqrt.approx.ftz.f32 %f36, %f613; max.ftz.f32 %f614, %f28, %f593; sqrt.approx.ftz.f32 %f37, %f614; mul.ftz.f32 %f615, %f18, %f35; fma.rn.ftz.f32 %f616, %f26, %f34, %f615; add.ftz.f32 %f617, %f616, %f616; mul.ftz.f32 %f618, %f35, %f617; sub.ftz.f32 %f38, %f618, %f18; mul.ftz.f32 %f619, %f18, %f37; fma.rn.ftz.f32 %f620, %f26, %f36, %f619; add.ftz.f32 %f621, %f620, %f620; mul.ftz.f32 %f622, %f37, %f621; sub.ftz.f32 %f39, %f622, %f18; setp.ge.ftz.f32 %p18, %f38, 0f3727C5AC; setp.ge.ftz.f32 %p19, %f39, 0f3727C5AC; and.pred %p20, %p18, %p19; @%p20 bra $L__BB193_13; bra.uni $L__BB193_12; $L__BB193_13: setp.lt.ftz.f32 %p22, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p22; selp.f32 %f2968, %f35, %f37, %p22; bra.uni $L__BB193_14; $L__BB193_9: or.pred %p17, %p1, %p2; mov.f32 %f55, %f550; mov.f32 %f56, %f551; mov.f32 %f57, %f552; @%p17 bra $L__BB193_10; bra.uni $L__BB193_15; $L__BB193_10: selp.f32 %f601, %f27, %f28, %p1; mov.f32 %f602, 0f3F800000; sub.ftz.f32 %f603, %f602, %f601; max.ftz.f32 %f605, %f603, %f593; sqrt.approx.ftz.f32 %f2967, %f605; max.ftz.f32 %f606, %f601, %f593; sqrt.approx.ftz.f32 %f2968, %f606; bra.uni $L__BB193_14; $L__BB193_12: setp.gt.ftz.f32 %p21, %f38, %f39; selp.f32 %f2967, %f34, %f36, %p21; selp.f32 %f2968, %f35, %f37, %p21; $L__BB193_14: mul.ftz.f32 %f623, %f550, %f2968; mul.ftz.f32 %f624, %f551, %f2968; mul.ftz.f32 %f625, %f552, %f2968; fma.rn.ftz.f32 %f57, %f25, %f2967, %f625; fma.rn.ftz.f32 %f56, %f24, %f2967, %f624; fma.rn.ftz.f32 %f55, %f23, %f2967, %f623; $L__BB193_15: setp.eq.s32 %p23, %r3, 255; @%p23 bra $L__BB193_17; bra.uni $L__BB193_16; $L__BB193_17: mov.b32 %f60, %r139; bra.uni $L__BB193_18; $L__BB193_16: mul.wide.u32 %rd239, %r3, 4; add.s64 %rd240, %rd2, %rd239; ld.local.f32 %f60, [%rd240]; $L__BB193_18: setp.eq.s32 %p24, %r4, 255; @%p24 bra $L__BB193_20; bra.uni $L__BB193_19; $L__BB193_20: mov.b32 %f63, %r140; bra.uni $L__BB193_21; $L__BB193_19: mul.wide.u32 %rd241, %r4, 4; add.s64 %rd242, %rd2, %rd241; ld.local.f32 %f63, [%rd242]; $L__BB193_21: cvt.u16.u32 %rs8, %r138; and.b16 %rs7, %rs8, 255; mul.wide.u32 %rd243, %r144, 4; add.s64 %rd5, %rd2, %rd243; mul.wide.u32 %rd244, %r145, 4; add.s64 %rd6, %rd2, %rd244; mul.wide.u32 %rd245, %r146, 4; add.s64 %rd7, %rd2, %rd245; setp.gt.s16 %p25, %rs7, 23; @%p25 bra $L__BB193_63; setp.gt.s16 %p36, %rs7, 15; @%p36 bra $L__BB193_48; setp.gt.s16 %p43, %rs7, 8; @%p43 bra $L__BB193_34; bra.uni $L__BB193_24; $L__BB193_34: add.s16 %rs12, %rs7, -9; setp.lt.u16 %p44, %rs12, 2; @%p44 bra $L__BB193_36; add.s16 %rs13, %rs7, -13; setp.lt.u16 %p45, %rs13, 2; @%p45 bra $L__BB193_36; bra.uni $L__BB193_517; $L__BB193_63: setp.gt.s16 %p26, %rs7, 30; @%p26 bra $L__BB193_88; setp.gt.s16 %p30, %rs7, 27; @%p30 bra $L__BB193_77; add.s16 %rs10, %rs7, -25; setp.lt.u16 %p34, %rs10, 2; @%p34 bra $L__BB193_427; setp.eq.s16 %p35, %rs7, 24; @%p35 bra $L__BB193_67; bra.uni $L__BB193_517; $L__BB193_67: ld.const.u32 %r496, [kernel_params+1564]; setp.ne.s32 %p385, %r496, 0; @%p385 bra $L__BB193_69; ld.const.u32 %r497, [kernel_params+1568]; setp.eq.s32 %p386, %r497, 0; shr.u32 %r498, %r141, 3; and.b32 %r499, %r498, 1; setp.eq.b32 %p387, %r499, 1; and.pred %p388, %p387, %p386; @%p388 bra $L__BB193_517; $L__BB193_69: ld.local.f32 %f2726, [%rd4+256]; mov.u64 %rd774, 0; mul.ftz.f32 %f2727, %f2, %f2726; ld.local.f32 %f2728, [%rd4+260]; mul.ftz.f32 %f2729, %f2, %f2728; ld.local.f32 %f2730, [%rd4+264]; mul.ftz.f32 %f2731, %f2, %f2730; mov.f32 %f2732, 0f00000000; max.ftz.f32 %f493, %f2727, %f2732; max.ftz.f32 %f494, %f2729, %f2732; max.ftz.f32 %f495, %f2731, %f2732; add.ftz.f32 %f2733, %f493, %f494; add.ftz.f32 %f2734, %f495, %f2733; mul.ftz.f32 %f2735, %f2734, 0f3EAAAAAB; abs.ftz.f32 %f496, %f2735; setp.ltu.ftz.f32 %p389, %f496, 0f3727C5AC; mov.u64 %rd775, %rd774; @%p389 bra $L__BB193_74; ld.local.u32 %r125, [%rd4+252]; mov.u64 %rd774, 0; setp.eq.s32 %p390, %r125, 0; mov.u64 %rd772, %rd774; mov.u64 %rd773, %rd774; @%p390 bra $L__BB193_72; ld.local.u32 %r500, [%rd4+248]; mul.wide.s32 %rd596, %r500, 80; add.s64 %rd597, %rd1, %rd596; add.s64 %rd772, %rd597, 304; add.s64 %rd598, %rd230, %rd596; add.s64 %rd773, %rd598, 304; mov.b32 %r501, %f495; mov.b32 %r502, %f494; mov.b32 %r503, %f493; mov.u32 %r504, 0; st.local.v4.u32 [%rd597+304], {%r503, %r502, %r501, %r504}; add.s32 %r505, %r500, 1; st.local.u32 [%rd4+248], %r505; add.s32 %r506, %r125, -1; st.local.u32 [%rd4+252], %r506; $L__BB193_72: setp.eq.s64 %p391, %rd773, 0; mov.u64 %rd775, %rd774; @%p391 bra $L__BB193_74; st.local.f32 [%rd772+16], %f496; mov.u64 %rd774, %rd772; mov.u64 %rd775, %rd773; $L__BB193_74: setp.eq.s64 %p392, %rd775, 0; @%p392 bra $L__BB193_517; ld.local.u32 %r126, [%rd4+252]; setp.lt.s32 %p393, %r126, 1; @%p393 bra $L__BB193_457; add.s32 %r507, %r126, -1; st.local.u32 [%rd4+252], %r507; ld.local.s32 %rd601, [%rd4+248]; cvt.s64.s32 %rd602, %r507; add.s64 %rd603, %rd601, %rd602; mul.lo.s64 %rd604, %rd603, 80; add.s64 %rd605, %rd1, %rd604; add.s64 %rd776, %rd605, 304; add.s64 %rd606, %rd230, %rd604; add.s64 %rd777, %rd606, 304; bra.uni $L__BB193_458; $L__BB193_48: setp.gt.s16 %p37, %rs7, 19; @%p37 bra $L__BB193_61; bra.uni $L__BB193_49; $L__BB193_61: add.s16 %rs11, %rs7, -21; setp.lt.u16 %p38, %rs11, 3; @%p38 bra $L__BB193_462; bra.uni $L__BB193_62; $L__BB193_462: ld.const.u32 %r518, [kernel_params+1568]; setp.eq.s32 %p399, %r518, 0; shr.u32 %r519, %r141, 3; and.b32 %r520, %r519, 1; setp.eq.b32 %p400, %r520, 1; and.pred %p401, %p400, %p399; @%p401 bra $L__BB193_517; ld.local.f32 %f2756, [%rd4+256]; mov.u64 %rd780, 0; mul.ftz.f32 %f2757, %f2, %f2756; ld.local.f32 %f2758, [%rd4+260]; mul.ftz.f32 %f2759, %f2, %f2758; ld.local.f32 %f2760, [%rd4+264]; mul.ftz.f32 %f2761, %f2, %f2760; mov.f32 %f2762, 0f00000000; max.ftz.f32 %f500, %f2757, %f2762; max.ftz.f32 %f501, %f2759, %f2762; max.ftz.f32 %f502, %f2761, %f2762; add.ftz.f32 %f2763, %f500, %f501; add.ftz.f32 %f2764, %f502, %f2763; mul.ftz.f32 %f2765, %f2764, 0f3EAAAAAB; abs.ftz.f32 %f503, %f2765; setp.ltu.ftz.f32 %p402, %f503, 0f3727C5AC; mov.u64 %rd781, %rd780; @%p402 bra $L__BB193_468; ld.local.u32 %r127, [%rd4+252]; mov.u64 %rd780, 0; setp.eq.s32 %p403, %r127, 0; mov.u64 %rd778, %rd780; mov.u64 %rd779, %rd780; @%p403 bra $L__BB193_466; ld.local.u32 %r521, [%rd4+248]; mul.wide.s32 %rd616, %r521, 80; add.s64 %rd617, %rd1, %rd616; add.s64 %rd778, %rd617, 304; add.s64 %rd618, %rd230, %rd616; add.s64 %rd779, %rd618, 304; mov.b32 %r522, %f502; mov.b32 %r523, %f501; mov.b32 %r524, %f500; mov.u32 %r525, 0; st.local.v4.u32 [%rd617+304], {%r524, %r523, %r522, %r525}; add.s32 %r526, %r521, 1; st.local.u32 [%rd4+248], %r526; add.s32 %r527, %r127, -1; st.local.u32 [%rd4+252], %r527; $L__BB193_466: setp.eq.s64 %p404, %rd779, 0; mov.u64 %rd781, %rd780; @%p404 bra $L__BB193_468; st.local.f32 [%rd778+16], %f503; mov.u64 %rd780, %rd778; mov.u64 %rd781, %rd779; $L__BB193_468: setp.eq.s64 %p405, %rd781, 0; @%p405 bra $L__BB193_517; add.s64 %rd185, %rd780, 20; mov.u64 %rd621, 0; st.local.f32 [%rd780+20], %f55; st.local.f32 [%rd780+24], %f56; st.local.f32 [%rd780+28], %f57; mov.u32 %r528, 0; st.local.v2.f32 [%rd780+56], {%f2762, %f2762}; st.local.u32 [%rd780+64], %r528; st.local.u64 [%rd780+48], %rd621; mov.f32 %f2767, 0f3727C5AC; max.ftz.f32 %f3057, %f63, %f2767; ld.local.u32 %r529, [%rd4+40]; and.b32 %r530, %r529, 1; setp.eq.b32 %p406, %r530, 1; mov.pred %p407, 0; xor.pred %p408, %p406, %p407; not.pred %p409, %p408; @%p409 bra $L__BB193_471; rcp.approx.ftz.f32 %f3057, %f3057; $L__BB193_471: setp.eq.s32 %p410, %r2, 21; @%p410 bra $L__BB193_475; bra.uni $L__BB193_472; $L__BB193_475: st.local.v2.f32 [%rd185+12], {%f2762, %f2762}; st.local.f32 [%rd185+20], %f3057; mov.u32 %r537, 21; st.local.u32 [%rd185+-8], %r537; ld.local.u32 %r538, [%rd4+40]; or.b32 %r539, %r538, 4; st.local.u32 [%rd4+40], %r539; bra.uni $L__BB193_517; $L__BB193_88: add.s16 %rs9, %rs7, -32; setp.lt.u16 %p27, %rs9, 3; @%p27 bra $L__BB193_290; bra.uni $L__BB193_89; $L__BB193_290: ld.local.v2.u32 {%r330, %r331}, [%rd4+252]; mov.u64 %rd746, 0; mov.b32 %f1892, %r331; mul.ftz.f32 %f340, %f2, %f1892; ld.local.f32 %f1893, [%rd4+260]; mul.ftz.f32 %f341, %f2, %f1893; ld.local.f32 %f1894, [%rd4+264]; mul.ftz.f32 %f342, %f2, %f1894; setp.eq.s32 %p240, %r330, 0; mov.u64 %rd744, %rd746; mov.u64 %rd745, %rd746; @%p240 bra $L__BB193_292; ld.local.u32 %r333, [%rd4+248]; mul.wide.s32 %rd459, %r333, 80; add.s64 %rd460, %rd1, %rd459; add.s64 %rd744, %rd460, 304; add.s64 %rd461, %rd230, %rd459; add.s64 %rd745, %rd461, 304; mov.b32 %r334, %f342; mov.b32 %r335, %f341; mov.b32 %r336, %f340; mov.u32 %r337, 0; st.local.v4.u32 [%rd460+304], {%r336, %r335, %r334, %r337}; add.s32 %r338, %r333, 1; st.local.u32 [%rd4+248], %r338; add.s32 %r339, %r330, -1; st.local.u32 [%rd4+252], %r339; $L__BB193_292: setp.eq.s64 %p241, %rd745, 0; @%p241 bra $L__BB193_294; add.ftz.f32 %f1895, %f340, %f341; add.ftz.f32 %f1896, %f342, %f1895; mul.ftz.f32 %f1897, %f1896, 0f3EAAAAAB; abs.ftz.f32 %f1898, %f1897; st.local.f32 [%rd744+16], %f1898; setp.ge.ftz.f32 %p242, %f1898, 0f3727C5AC; selp.b64 %rd746, %rd745, 0, %p242; $L__BB193_294: setp.eq.s64 %p243, %rd746, 0; @%p243 bra $L__BB193_517; and.b32 %r341, %r141, 4096; setp.eq.s32 %p244, %r341, 0; selp.f32 %f1900, %f60, 0f00000000, %p244; mov.f32 %f3029, 0f00000000; ld.local.f32 %f1901, [%rd6]; mul.ftz.f32 %f1902, %f1900, %f1901; ld.local.f32 %f1903, [%rd6+4]; mul.ftz.f32 %f343, %f1900, %f1903; ld.local.f32 %f1904, [%rd6+8]; mul.ftz.f32 %f344, %f1900, %f1904; add.s64 %rd111, %rd746, 32; st.f32 [%rd746+32], %f1902; st.f32 [%rd746+36], %f343; st.f32 [%rd746+40], %f344; ld.local.f32 %f1905, [%rd4+256]; ld.local.f32 %f1906, [%rd4+260]; ld.local.f32 %f1907, [%rd4+264]; st.f32 [%rd746+44], %f1905; st.f32 [%rd746+48], %f1906; st.f32 [%rd746+52], %f1907; st.f32 [%rd746+20], %f55; st.f32 [%rd746+24], %f56; st.f32 [%rd746+28], %f57; mov.u32 %r342, 2139095039; st.u32 [%rd746+56], %r342; ld.local.f32 %f1908, [%rd7]; max.ftz.f32 %f1909, %f1908, %f3029; mov.f32 %f1910, 0f3F666666; min.ftz.f32 %f1911, %f1909, %f1910; st.f32 [%rd746+60], %f1911; setp.geu.ftz.f32 %p245, %f1902, 0f322BCC77; mov.u32 %r646, 3; mov.f32 %f3027, %f3029; @%p245 bra $L__BB193_297; ld.f32 %f3027, [%rd111+-32]; mov.u32 %r344, 0; st.u32 [%rd111+-32], %r344; st.u32 [%rd111], %r344; mov.u32 %r646, 2; $L__BB193_297: setp.geu.ftz.f32 %p246, %f343, 0f322BCC77; mov.f32 %f3028, %f3029; @%p246 bra $L__BB193_299; ld.f32 %f3028, [%rd111+-28]; mov.u32 %r345, 0; st.u32 [%rd111+-28], %r345; st.u32 [%rd111+4], %r345; add.s32 %r646, %r646, -1; $L__BB193_299: setp.geu.ftz.f32 %p247, %f344, 0f322BCC77; @%p247 bra $L__BB193_301; ld.f32 %f3029, [%rd111+-24]; mov.u32 %r346, 0; st.u32 [%rd111+-24], %r346; st.u32 [%rd111+8], %r346; add.s32 %r646, %r646, -1; $L__BB193_301: setp.gt.s32 %p248, %r646, 2; mov.u32 %r648, 0; @%p248 bra $L__BB193_309; mov.f32 %f1914, 0f00000000; max.ftz.f32 %f351, %f3027, %f1914; max.ftz.f32 %f352, %f3028, %f1914; add.ftz.f32 %f1915, %f351, %f352; max.ftz.f32 %f353, %f3029, %f1914; add.ftz.f32 %f1916, %f353, %f1915; mul.ftz.f32 %f1917, %f1916, 0f3EAAAAAB; abs.ftz.f32 %f354, %f1917; setp.ltu.ftz.f32 %p249, %f354, 0f3727C5AC; mov.u64 %rd749, 0; mov.u64 %rd750, %rd749; @%p249 bra $L__BB193_307; ld.local.u32 %r63, [%rd4+252]; mov.u64 %rd749, 0; setp.eq.s32 %p250, %r63, 0; mov.u64 %rd747, %rd749; mov.u64 %rd748, %rd749; @%p250 bra $L__BB193_305; ld.local.u32 %r348, [%rd4+248]; mul.wide.s32 %rd467, %r348, 80; add.s64 %rd468, %rd1, %rd467; add.s64 %rd747, %rd468, 304; add.s64 %rd469, %rd230, %rd467; add.s64 %rd748, %rd469, 304; mov.b32 %r349, %f353; mov.b32 %r350, %f352; mov.b32 %r351, %f351; mov.u32 %r352, 0; st.local.v4.u32 [%rd468+304], {%r351, %r350, %r349, %r352}; add.s32 %r353, %r348, 1; st.local.u32 [%rd4+248], %r353; add.s32 %r354, %r63, -1; st.local.u32 [%rd4+252], %r354; $L__BB193_305: setp.eq.s64 %p251, %rd748, 0; mov.u64 %rd750, %rd749; @%p251 bra $L__BB193_307; st.local.f32 [%rd747+16], %f354; mov.u64 %rd749, %rd747; mov.u64 %rd750, %rd748; $L__BB193_307: setp.eq.s64 %p252, %rd750, 0; @%p252 bra $L__BB193_309; ld.f32 %f1918, [%rd111+-12]; ld.f32 %f1919, [%rd111+-8]; ld.f32 %f1920, [%rd111+-4]; st.local.f32 [%rd749+20], %f1918; st.local.f32 [%rd749+24], %f1919; st.local.f32 [%rd749+28], %f1920; mov.u32 %r357, 2; st.local.u32 [%rd749+12], %r357; mov.u32 %r648, 12; $L__BB193_309: st.u32 [%rd111+-20], %r2; setp.gt.s32 %p253, %r646, 0; @%p253 bra $L__BB193_311; bra.uni $L__BB193_310; $L__BB193_311: ld.f32 %f1921, [%rd111+-28]; ld.f32 %f1922, [%rd111+-32]; add.ftz.f32 %f1923, %f1922, %f1921; ld.f32 %f1924, [%rd111+-24]; add.ftz.f32 %f1925, %f1924, %f1923; mul.ftz.f32 %f1926, %f1925, 0f3EAAAAAB; abs.ftz.f32 %f1927, %f1926; cvt.rn.f32.s32 %f1928, %r646; mul.ftz.f32 %f1929, %f1927, %f1928; st.f32 [%rd111+-16], %f1929; or.b32 %r359, %r2, 2; setp.eq.s32 %p254, %r359, 34; @%p254 bra $L__BB193_322; bra.uni $L__BB193_312; $L__BB193_322: ld.f32 %f2483, [%rd111]; mul.ftz.f32 %f3033, %f2483, 0f3DA2F983; ld.f32 %f2484, [%rd111+4]; mul.ftz.f32 %f3034, %f2484, 0f3DA2F983; ld.f32 %f2485, [%rd111+8]; mul.ftz.f32 %f3035, %f2485, 0f3DA2F983; bra.uni $L__BB193_323; $L__BB193_24: setp.eq.s16 %p46, %rs7, 2; @%p46 bra $L__BB193_507; setp.eq.s16 %p47, %rs7, 7; @%p47 bra $L__BB193_53; setp.eq.s16 %p48, %rs7, 8; @%p48 bra $L__BB193_27; bra.uni $L__BB193_517; $L__BB193_27: ld.local.f32 %f2922, [%rd4+256]; mov.u64 %rd798, 0; mul.ftz.f32 %f2923, %f2, %f2922; ld.local.f32 %f2924, [%rd4+260]; mul.ftz.f32 %f2925, %f2, %f2924; ld.local.f32 %f2926, [%rd4+264]; mul.ftz.f32 %f2927, %f2, %f2926; mov.f32 %f2928, 0f00000000; max.ftz.f32 %f535, %f2923, %f2928; max.ftz.f32 %f536, %f2925, %f2928; max.ftz.f32 %f537, %f2927, %f2928; add.ftz.f32 %f2929, %f535, %f536; add.ftz.f32 %f2930, %f537, %f2929; mul.ftz.f32 %f2931, %f2930, 0f3EAAAAAB; abs.ftz.f32 %f538, %f2931; setp.ltu.ftz.f32 %p447, %f538, 0f3727C5AC; mov.u64 %rd799, %rd798; @%p447 bra $L__BB193_32; ld.local.u32 %r132, [%rd4+252]; mov.u64 %rd798, 0; setp.eq.s32 %p448, %r132, 0; mov.u64 %rd796, %rd798; mov.u64 %rd797, %rd798; @%p448 bra $L__BB193_30; ld.local.u32 %r600, [%rd4+248]; mul.wide.s32 %rd668, %r600, 80; add.s64 %rd669, %rd1, %rd668; add.s64 %rd796, %rd669, 304; add.s64 %rd670, %rd230, %rd668; add.s64 %rd797, %rd670, 304; mov.b32 %r601, %f537; mov.b32 %r602, %f536; mov.b32 %r603, %f535; mov.u32 %r604, 0; st.local.v4.u32 [%rd669+304], {%r603, %r602, %r601, %r604}; add.s32 %r605, %r600, 1; st.local.u32 [%rd4+248], %r605; add.s32 %r606, %r132, -1; st.local.u32 [%rd4+252], %r606; $L__BB193_30: setp.eq.s64 %p449, %rd797, 0; mov.u64 %rd799, %rd798; @%p449 bra $L__BB193_32; st.local.f32 [%rd796+16], %f538; mov.u64 %rd798, %rd796; mov.u64 %rd799, %rd797; $L__BB193_32: setp.eq.s64 %p450, %rd799, 0; @%p450 bra $L__BB193_517; st.local.f32 [%rd798+20], %f55; st.local.f32 [%rd798+24], %f56; st.local.f32 [%rd798+28], %f57; mov.u32 %r607, 8; st.local.u32 [%rd798+12], %r607; ld.local.u32 %r608, [%rd4+40]; or.b32 %r609, %r608, 12; st.local.u32 [%rd4+40], %r609; bra.uni $L__BB193_517; $L__BB193_77: setp.eq.s16 %p31, %rs7, 28; @%p31 bra $L__BB193_427; setp.eq.s16 %p32, %rs7, 29; @%p32 bra $L__BB193_341; setp.eq.s16 %p33, %rs7, 30; @%p33 bra $L__BB193_80; bra.uni $L__BB193_517; $L__BB193_49: setp.eq.s16 %p40, %rs7, 16; @%p40 bra $L__BB193_36; setp.eq.s16 %p41, %rs7, 17; @%p41 bra $L__BB193_482; setp.eq.s16 %p42, %rs7, 19; @%p42 bra $L__BB193_52; bra.uni $L__BB193_517; $L__BB193_52: ld.const.u32 %r554, [kernel_params+1564]; setp.eq.s32 %p419, %r554, 0; shr.u32 %r555, %r141, 3; and.b32 %r556, %r555, 1; setp.eq.b32 %p420, %r556, 1; and.pred %p421, %p420, %p419; @%p421 bra $L__BB193_517; $L__BB193_53: ld.local.f32 %f2932, [%rd4+256]; mov.u64 %rd802, 0; mul.ftz.f32 %f2933, %f2, %f2932; ld.local.f32 %f2934, [%rd4+260]; mul.ftz.f32 %f2935, %f2, %f2934; ld.local.f32 %f2936, [%rd4+264]; mul.ftz.f32 %f2937, %f2, %f2936; mov.f32 %f2938, 0f00000000; max.ftz.f32 %f539, %f2933, %f2938; max.ftz.f32 %f540, %f2935, %f2938; max.ftz.f32 %f541, %f2937, %f2938; add.ftz.f32 %f2939, %f539, %f540; add.ftz.f32 %f2940, %f541, %f2939; mul.ftz.f32 %f2941, %f2940, 0f3EAAAAAB; abs.ftz.f32 %f542, %f2941; setp.ltu.ftz.f32 %p451, %f542, 0f3727C5AC; mov.u64 %rd803, %rd802; @%p451 bra $L__BB193_58; ld.local.u32 %r133, [%rd4+252]; mov.u64 %rd802, 0; setp.eq.s32 %p452, %r133, 0; mov.u64 %rd800, %rd802; mov.u64 %rd801, %rd802; @%p452 bra $L__BB193_56; ld.local.u32 %r610, [%rd4+248]; mul.wide.s32 %rd677, %r610, 80; add.s64 %rd678, %rd1, %rd677; add.s64 %rd800, %rd678, 304; add.s64 %rd679, %rd230, %rd677; add.s64 %rd801, %rd679, 304; mov.b32 %r611, %f541; mov.b32 %r612, %f540; mov.b32 %r613, %f539; mov.u32 %r614, 0; st.local.v4.u32 [%rd678+304], {%r613, %r612, %r611, %r614}; add.s32 %r615, %r610, 1; st.local.u32 [%rd4+248], %r615; add.s32 %r616, %r133, -1; st.local.u32 [%rd4+252], %r616; $L__BB193_56: setp.eq.s64 %p453, %rd801, 0; mov.u64 %rd803, %rd802; @%p453 bra $L__BB193_58; st.local.f32 [%rd800+16], %f542; mov.u64 %rd802, %rd800; mov.u64 %rd803, %rd801; $L__BB193_58: setp.eq.s64 %p454, %rd803, 0; @%p454 bra $L__BB193_517; add.s64 %rd222, %rd802, 20; st.local.f32 [%rd802+20], %f55; st.local.f32 [%rd802+24], %f56; st.local.f32 [%rd802+28], %f57; st.local.f32 [%rd802+32], %f60; st.local.f32 [%rd802+36], %f63; cvt.ftz.sat.f32.f32 %f543, %f60; cvt.ftz.sat.f32.f32 %f544, %f63; setp.eq.s32 %p455, %r2, 7; @%p455 bra $L__BB193_506; bra.uni $L__BB193_60; $L__BB193_506: mov.u32 %r620, 7; st.local.u32 [%rd222+-8], %r620; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r621, [%rd4+40]; or.b32 %r622, %r621, 12; st.local.u32 [%rd4+40], %r622; bra.uni $L__BB193_517; $L__BB193_36: ld.const.u32 %r567, [kernel_params+1564]; setp.eq.s32 %p426, %r567, 0; shr.u32 %r568, %r141, 3; and.b32 %r569, %r568, 1; setp.eq.b32 %p427, %r569, 1; and.pred %p428, %p427, %p426; @%p428 bra $L__BB193_517; ld.local.f32 %f2831, [%rd4+256]; mov.u64 %rd792, 0; mul.ftz.f32 %f2832, %f2, %f2831; ld.local.f32 %f2833, [%rd4+260]; mul.ftz.f32 %f2834, %f2, %f2833; ld.local.f32 %f2835, [%rd4+264]; mul.ftz.f32 %f2836, %f2, %f2835; mov.f32 %f2837, 0f00000000; max.ftz.f32 %f519, %f2832, %f2837; max.ftz.f32 %f520, %f2834, %f2837; max.ftz.f32 %f521, %f2836, %f2837; add.ftz.f32 %f2838, %f519, %f520; add.ftz.f32 %f2839, %f521, %f2838; mul.ftz.f32 %f2840, %f2839, 0f3EAAAAAB; abs.ftz.f32 %f522, %f2840; setp.ltu.ftz.f32 %p429, %f522, 0f3727C5AC; mov.u64 %rd793, %rd792; @%p429 bra $L__BB193_42; ld.local.u32 %r130, [%rd4+252]; mov.u64 %rd792, 0; setp.eq.s32 %p430, %r130, 0; mov.u64 %rd790, %rd792; mov.u64 %rd791, %rd792; @%p430 bra $L__BB193_40; ld.local.u32 %r570, [%rd4+248]; mul.wide.s32 %rd645, %r570, 80; add.s64 %rd646, %rd1, %rd645; add.s64 %rd790, %rd646, 304; add.s64 %rd647, %rd230, %rd645; add.s64 %rd791, %rd647, 304; mov.b32 %r571, %f521; mov.b32 %r572, %f520; mov.b32 %r573, %f519; mov.u32 %r574, 0; st.local.v4.u32 [%rd646+304], {%r573, %r572, %r571, %r574}; add.s32 %r575, %r570, 1; st.local.u32 [%rd4+248], %r575; add.s32 %r576, %r130, -1; st.local.u32 [%rd4+252], %r576; $L__BB193_40: setp.eq.s64 %p431, %rd791, 0; mov.u64 %rd793, %rd792; @%p431 bra $L__BB193_42; st.local.f32 [%rd790+16], %f522; mov.u64 %rd792, %rd790; mov.u64 %rd793, %rd791; $L__BB193_42: setp.eq.s64 %p432, %rd793, 0; @%p432 bra $L__BB193_517; mul.ftz.f32 %f3058, %f60, %f60; add.s64 %rd205, %rd792, 20; mov.u64 %rd794, 0; st.local.f32 [%rd792+20], %f55; st.local.f32 [%rd792+24], %f56; st.local.f32 [%rd792+28], %f57; mov.u32 %r577, 1065353216; st.local.u32 [%rd792+40], %r577; st.local.u64 [%rd792+48], %rd794; setp.eq.s32 %p433, %r144, 255; @%p433 bra $L__BB193_490; bra.uni $L__BB193_44; $L__BB193_490: mov.u32 %r578, 0; st.local.v2.f32 [%rd205+36], {%f2837, %f2837}; st.local.u32 [%rd205+44], %r578; st.local.v2.f32 [%rd205+12], {%f3058, %f3058}; mov.f32 %f3059, %f3058; bra.uni $L__BB193_491; $L__BB193_427: ld.const.u32 %r121, [kernel_params+1564]; setp.ne.s32 %p358, %r121, 0; @%p358 bra $L__BB193_429; ld.const.u32 %r454, [kernel_params+1568]; setp.eq.s32 %p359, %r454, 0; shr.u32 %r455, %r141, 3; and.b32 %r456, %r455, 1; setp.eq.b32 %p360, %r456, 1; and.pred %p361, %p360, %p359; @%p361 bra $L__BB193_517; $L__BB193_429: ld.local.f32 %f2682, [%rd4+256]; mul.ftz.f32 %f470, %f2, %f2682; ld.local.f32 %f2683, [%rd4+260]; mul.ftz.f32 %f471, %f2, %f2683; ld.local.f32 %f2684, [%rd4+264]; mul.ftz.f32 %f472, %f2, %f2684; mov.f32 %f2685, 0f3727C5AC; max.ftz.f32 %f3054, %f63, %f2685; ld.local.u32 %r457, [%rd4+40]; and.b32 %r458, %r457, 1; setp.eq.b32 %p362, %r458, 1; mov.pred %p363, 0; xor.pred %p364, %p362, %p363; not.pred %p365, %p364; @%p365 bra $L__BB193_431; rcp.approx.ftz.f32 %f3054, %f3054; $L__BB193_431: ld.local.f32 %f2687, [%rd4+24]; ld.local.f32 %f2688, [%rd4+28]; mul.ftz.f32 %f2689, %f56, %f2688; fma.rn.ftz.f32 %f2690, %f55, %f2687, %f2689; ld.local.f32 %f2691, [%rd4+32]; fma.rn.ftz.f32 %f2692, %f57, %f2691, %f2690; abs.ftz.f32 %f476, %f2692; fma.rn.ftz.f32 %f2693, %f3054, %f3054, 0fBF800000; mov.f32 %f3055, 0f3F800000; fma.rn.ftz.f32 %f477, %f476, %f476, %f2693; setp.leu.ftz.f32 %p366, %f477, 0f00000000; @%p366 bra $L__BB193_433; sqrt.approx.ftz.f32 %f2694, %f477; sub.ftz.f32 %f2695, %f2694, %f476; add.ftz.f32 %f2696, %f476, %f2694; div.approx.ftz.f32 %f2697, %f2695, %f2696; fma.rn.ftz.f32 %f2698, %f476, %f2696, 0fBF800000; fma.rn.ftz.f32 %f2699, %f476, %f2695, 0f3F800000; div.approx.ftz.f32 %f2700, %f2698, %f2699; mul.ftz.f32 %f2701, %f2697, 0f3F000000; mul.ftz.f32 %f2702, %f2697, %f2701; fma.rn.ftz.f32 %f2703, %f2700, %f2700, 0f3F800000; mul.ftz.f32 %f3055, %f2702, %f2703; $L__BB193_433: mul.ftz.f32 %f480, %f60, %f60; and.b32 %r122, %r141, 8; shr.u32 %r459, %r141, 3; and.b32 %r460, %r459, 1; setp.eq.b32 %p367, %r460, 1; setp.eq.s32 %p368, %r121, 0; and.pred %p369, %p367, %p368; @%p369 bra $L__BB193_445; mul.ftz.f32 %f2704, %f470, %f3055; mov.f32 %f2705, 0f00000000; max.ftz.f32 %f481, %f2704, %f2705; mul.ftz.f32 %f2706, %f471, %f3055; max.ftz.f32 %f482, %f2706, %f2705; mul.ftz.f32 %f2707, %f472, %f3055; max.ftz.f32 %f483, %f2707, %f2705; add.ftz.f32 %f2708, %f481, %f482; add.ftz.f32 %f2709, %f483, %f2708; mul.ftz.f32 %f2710, %f2709, 0f3EAAAAAB; abs.ftz.f32 %f484, %f2710; setp.ltu.ftz.f32 %p370, %f484, 0f3727C5AC; mov.u64 %rd766, 0; mov.u64 %rd767, %rd766; @%p370 bra $L__BB193_439; ld.local.u32 %r123, [%rd4+252]; mov.u64 %rd766, 0; setp.eq.s32 %p371, %r123, 0; mov.u64 %rd764, %rd766; mov.u64 %rd765, %rd766; @%p371 bra $L__BB193_437; ld.local.u32 %r461, [%rd4+248]; mul.wide.s32 %rd574, %r461, 80; add.s64 %rd575, %rd1, %rd574; add.s64 %rd764, %rd575, 304; add.s64 %rd576, %rd230, %rd574; add.s64 %rd765, %rd576, 304; mov.b32 %r462, %f483; mov.b32 %r463, %f482; mov.b32 %r464, %f481; mov.u32 %r465, 0; st.local.v4.u32 [%rd575+304], {%r464, %r463, %r462, %r465}; add.s32 %r466, %r461, 1; st.local.u32 [%rd4+248], %r466; add.s32 %r467, %r123, -1; st.local.u32 [%rd4+252], %r467; $L__BB193_437: setp.eq.s64 %p372, %rd765, 0; mov.u64 %rd767, %rd766; @%p372 bra $L__BB193_439; st.local.f32 [%rd764+16], %f484; mov.u64 %rd766, %rd764; mov.u64 %rd767, %rd765; $L__BB193_439: setp.eq.s64 %p373, %rd767, 0; @%p373 bra $L__BB193_445; add.s64 %rd160, %rd766, 20; mov.u64 %rd579, 0; st.local.f32 [%rd766+20], %f55; st.local.f32 [%rd766+24], %f56; st.local.f32 [%rd766+28], %f57; mov.u32 %r468, 0; st.local.v2.f32 [%rd766+56], {%f2705, %f2705}; st.local.u32 [%rd766+64], %r468; st.local.u64 [%rd766+48], %rd579; setp.eq.s32 %p374, %r2, 28; @%p374 bra $L__BB193_444; bra.uni $L__BB193_441; $L__BB193_444: st.local.v2.f32 [%rd160+12], {%f2705, %f2705}; st.local.f32 [%rd160+20], %f3054; mov.u32 %r475, 9; st.local.u32 [%rd160+-8], %r475; ld.local.u32 %r476, [%rd4+40]; or.b32 %r477, %r476, 4; st.local.u32 [%rd4+40], %r477; bra.uni $L__BB193_445; $L__BB193_89: setp.eq.s16 %p28, %rs7, 31; @%p28 bra $L__BB193_325; setp.ne.s16 %p29, %rs7, 39; @%p29 bra $L__BB193_517; ld.global.v4.u32 {%r147, %r148, %r149, %r150}, [%rd3+16]; ld.local.f32 %f2977, [%rd5]; ld.local.f32 %f2978, [%rd5+4]; ld.local.f32 %f2979, [%rd5+8]; shr.u32 %r152, %r147, 24; shl.b32 %r153, %r145, 2; cvt.u64.u32 %rd246, %r153; and.b64 %rd247, %rd246, 1020; add.s64 %rd248, %rd2, %rd247; ld.local.f32 %f70, [%rd248]; shr.u32 %r154, %r145, 6; cvt.u64.u32 %rd249, %r154; and.b64 %rd250, %rd249, 1020; add.s64 %rd251, %rd2, %rd250; ld.local.f32 %f71, [%rd251]; shr.u32 %r155, %r145, 14; cvt.u64.u32 %rd252, %r155; and.b64 %rd253, %rd252, 1020; add.s64 %rd254, %rd2, %rd253; ld.local.f32 %f72, [%rd254]; shr.u32 %r156, %r145, 24; mul.wide.u32 %rd255, %r156, 4; add.s64 %rd256, %rd2, %rd255; ld.local.f32 %f73, [%rd256]; shl.b32 %r157, %r146, 2; cvt.u64.u32 %rd257, %r157; and.b64 %rd258, %rd257, 1020; add.s64 %rd259, %rd2, %rd258; ld.local.f32 %f74, [%rd259]; shr.u32 %r158, %r146, 6; cvt.u64.u32 %rd260, %r158; and.b64 %rd261, %rd260, 1020; add.s64 %rd262, %rd2, %rd261; ld.local.f32 %f75, [%rd262]; shr.u32 %r159, %r146, 14; cvt.u64.u32 %rd263, %r159; and.b64 %rd264, %rd263, 1020; add.s64 %rd265, %rd2, %rd264; ld.local.f32 %f76, [%rd265]; shr.u32 %r160, %r146, 24; mul.wide.u32 %rd266, %r160, 4; add.s64 %rd267, %rd2, %rd266; ld.local.f32 %f77, [%rd267]; shr.u32 %r161, %r147, 6; cvt.u64.u32 %rd268, %r161; and.b64 %rd269, %rd268, 1020; add.s64 %rd270, %rd2, %rd269; ld.local.f32 %f78, [%rd270]; shr.u32 %r162, %r147, 14; cvt.u64.u32 %rd271, %r162; and.b64 %rd272, %rd271, 1020; add.s64 %rd273, %rd2, %rd272; mul.wide.u32 %rd274, %r152, 4; add.s64 %rd275, %rd2, %rd274; ld.local.f32 %f79, [%rd275]; shl.b32 %r163, %r147, 2; cvt.u64.u32 %rd276, %r163; and.b64 %rd277, %rd276, 1020; add.s64 %rd278, %rd2, %rd277; ld.local.f32 %f626, [%rd278]; mov.f32 %f627, 0f3727C5AC; max.ftz.f32 %f2980, %f626, %f627; ld.local.f32 %f81, [%rd273]; setp.eq.ftz.f32 %p49, %f81, 0f00000000; @%p49 bra $L__BB193_93; mul.ftz.f32 %f628, %f81, 0f40C90FDB; cos.approx.ftz.f32 %f629, %f628; sin.approx.ftz.f32 %f630, %f628; mov.f32 %f631, 0f3F800000; sub.ftz.f32 %f632, %f631, %f629; mul.ftz.f32 %f633, %f55, %f632; fma.rn.ftz.f32 %f634, %f55, %f633, %f629; mul.ftz.f32 %f635, %f56, %f633; mul.ftz.f32 %f636, %f57, %f630; sub.ftz.f32 %f637, %f635, %f636; mul.ftz.f32 %f638, %f2978, %f637; fma.rn.ftz.f32 %f639, %f2977, %f634, %f638; mul.ftz.f32 %f640, %f57, %f633; mul.ftz.f32 %f641, %f56, %f630; add.ftz.f32 %f642, %f641, %f640; add.ftz.f32 %f643, %f636, %f635; mul.ftz.f32 %f644, %f56, %f632; fma.rn.ftz.f32 %f645, %f56, %f644, %f629; mul.ftz.f32 %f646, %f2978, %f645; fma.rn.ftz.f32 %f647, %f2977, %f643, %f646; mul.ftz.f32 %f648, %f57, %f644; mul.ftz.f32 %f649, %f55, %f630; sub.ftz.f32 %f650, %f648, %f649; sub.ftz.f32 %f651, %f640, %f641; add.ftz.f32 %f652, %f649, %f648; mul.ftz.f32 %f653, %f2978, %f652; fma.rn.ftz.f32 %f654, %f2977, %f651, %f653; mul.ftz.f32 %f655, %f57, %f632; fma.rn.ftz.f32 %f656, %f57, %f655, %f629; fma.rn.ftz.f32 %f2978, %f2979, %f650, %f647; fma.rn.ftz.f32 %f84, %f2979, %f656, %f654; fma.rn.ftz.f32 %f2977, %f2979, %f642, %f639; mov.f32 %f2979, %f84; $L__BB193_93: ld.local.u32 %r164, [%rd4+40]; and.b32 %r165, %r164, 1; setp.eq.b32 %p50, %r165, 1; mov.pred %p51, 0; xor.pred %p52, %p50, %p51; not.pred %p53, %p52; @%p53 bra $L__BB193_95; rcp.approx.ftz.f32 %f2980, %f2980; $L__BB193_95: ld.local.f32 %f90, [%rd4+24]; ld.local.f32 %f91, [%rd4+28]; mul.ftz.f32 %f658, %f56, %f91; fma.rn.ftz.f32 %f659, %f55, %f90, %f658; ld.local.f32 %f92, [%rd4+32]; fma.rn.ftz.f32 %f660, %f57, %f92, %f659; abs.ftz.f32 %f93, %f660; fma.rn.ftz.f32 %f661, %f2980, %f2980, 0fBF800000; mov.f32 %f657, 0f3F800000; fma.rn.ftz.f32 %f94, %f93, %f93, %f661; setp.leu.ftz.f32 %p54, %f94, 0f00000000; mov.f32 %f2981, %f657; @%p54 bra $L__BB193_97; sqrt.approx.ftz.f32 %f662, %f94; sub.ftz.f32 %f663, %f662, %f93; add.ftz.f32 %f664, %f93, %f662; div.approx.ftz.f32 %f665, %f663, %f664; fma.rn.ftz.f32 %f666, %f93, %f664, 0fBF800000; fma.rn.ftz.f32 %f667, %f93, %f663, 0f3F800000; div.approx.ftz.f32 %f668, %f666, %f667; mul.ftz.f32 %f669, %f665, 0f3F000000; mul.ftz.f32 %f670, %f665, %f669; fma.rn.ftz.f32 %f671, %f668, %f668, 0f3F800000; mul.ftz.f32 %f2981, %f670, %f671; $L__BB193_97: cvt.ftz.sat.f32.f32 %f672, %f60; sub.ftz.f32 %f674, %f657, %f672; cvt.ftz.sat.f32.f32 %f675, %f78; sub.ftz.f32 %f676, %f657, %f675; mul.ftz.f32 %f97, %f674, %f676; mul.ftz.f32 %f98, %f675, %f674; sub.ftz.f32 %f99, %f657, %f98; ld.global.v4.u32 {%r166, %r167, %r168, %r169}, [%rd3+32]; setp.eq.s32 %p55, %r166, 255; @%p55 bra $L__BB193_99; bra.uni $L__BB193_98; $L__BB193_99: mov.b32 %f2982, %r167; mov.b32 %f2983, %r168; mov.b32 %f2984, %r169; bra.uni $L__BB193_100; $L__BB193_62: setp.eq.s16 %p39, %rs7, 20; @%p39 bra $L__BB193_80; bra.uni $L__BB193_517; $L__BB193_80: ld.local.f32 %f2770, [%rd4+256]; mov.u64 %rd784, 0; mul.ftz.f32 %f2771, %f2, %f2770; ld.local.f32 %f2772, [%rd4+260]; mul.ftz.f32 %f2773, %f2, %f2772; ld.local.f32 %f2774, [%rd4+264]; mul.ftz.f32 %f2775, %f2, %f2774; mov.f32 %f2776, 0f00000000; max.ftz.f32 %f509, %f2771, %f2776; max.ftz.f32 %f510, %f2773, %f2776; max.ftz.f32 %f511, %f2775, %f2776; add.ftz.f32 %f2777, %f509, %f510; add.ftz.f32 %f2778, %f511, %f2777; mul.ftz.f32 %f2779, %f2778, 0f3EAAAAAB; abs.ftz.f32 %f512, %f2779; setp.ltu.ftz.f32 %p412, %f512, 0f3727C5AC; mov.u64 %rd785, %rd784; @%p412 bra $L__BB193_85; ld.local.u32 %r128, [%rd4+252]; mov.u64 %rd784, 0; setp.eq.s32 %p413, %r128, 0; mov.u64 %rd782, %rd784; mov.u64 %rd783, %rd784; @%p413 bra $L__BB193_83; ld.local.u32 %r540, [%rd4+248]; mul.wide.s32 %rd627, %r540, 80; add.s64 %rd628, %rd1, %rd627; add.s64 %rd782, %rd628, 304; add.s64 %rd629, %rd230, %rd627; add.s64 %rd783, %rd629, 304; mov.b32 %r541, %f511; mov.b32 %r542, %f510; mov.b32 %r543, %f509; mov.u32 %r544, 0; st.local.v4.u32 [%rd628+304], {%r543, %r542, %r541, %r544}; add.s32 %r545, %r540, 1; st.local.u32 [%rd4+248], %r545; add.s32 %r546, %r128, -1; st.local.u32 [%rd4+252], %r546; $L__BB193_83: setp.eq.s64 %p414, %rd783, 0; mov.u64 %rd785, %rd784; @%p414 bra $L__BB193_85; st.local.f32 [%rd782+16], %f512; mov.u64 %rd784, %rd782; mov.u64 %rd785, %rd783; $L__BB193_85: setp.eq.s64 %p415, %rd785, 0; @%p415 bra $L__BB193_517; add.s64 %rd192, %rd784, 20; st.local.f32 [%rd784+20], %f55; st.local.f32 [%rd784+24], %f56; st.local.f32 [%rd784+28], %f57; st.local.f32 [%rd784+44], %f60; st.local.f32 [%rd784+48], %f63; ld.local.f32 %f2780, [%rd6]; neg.ftz.f32 %f2781, %f2780; st.local.f32 [%rd784+52], %f2781; setp.eq.s32 %p416, %r144, 255; @%p416 bra $L__BB193_476; bra.uni $L__BB193_87; $L__BB193_476: ld.local.u8 %rs18, [%rd4+52]; and.b16 %rs19, %rs18, 6; setp.eq.s16 %p417, %rs19, 0; @%p417 bra $L__BB193_478; ld.local.f32 %f2792, [%rd4+104]; ld.local.f32 %f2793, [%rd4+108]; mul.ftz.f32 %f2794, %f2793, %f2793; fma.rn.ftz.f32 %f2795, %f2792, %f2792, %f2794; ld.local.f32 %f2796, [%rd4+112]; fma.rn.ftz.f32 %f2797, %f2796, %f2796, %f2795; sqrt.approx.ftz.f32 %f2798, %f2797; div.approx.ftz.f32 %f2799, %f2792, %f2798; div.approx.ftz.f32 %f2800, %f2793, %f2798; div.approx.ftz.f32 %f2801, %f2796, %f2798; st.local.f32 [%rd192+12], %f2799; st.local.f32 [%rd192+16], %f2800; st.local.f32 [%rd192+20], %f2801; bra.uni $L__BB193_479; $L__BB193_507: ld.local.f32 %f2942, [%rd4+256]; mov.u64 %rd806, 0; mul.ftz.f32 %f2943, %f2, %f2942; ld.local.f32 %f2944, [%rd4+260]; mul.ftz.f32 %f2945, %f2, %f2944; ld.local.f32 %f2946, [%rd4+264]; mul.ftz.f32 %f2947, %f2, %f2946; mov.f32 %f2948, 0f00000000; max.ftz.f32 %f545, %f2943, %f2948; max.ftz.f32 %f546, %f2945, %f2948; max.ftz.f32 %f547, %f2947, %f2948; add.ftz.f32 %f2949, %f545, %f546; add.ftz.f32 %f2950, %f547, %f2949; mul.ftz.f32 %f2951, %f2950, 0f3EAAAAAB; abs.ftz.f32 %f548, %f2951; setp.ltu.ftz.f32 %p456, %f548, 0f3727C5AC; mov.u64 %rd807, %rd806; @%p456 bra $L__BB193_512; ld.local.u32 %r134, [%rd4+252]; mov.u64 %rd806, 0; setp.eq.s32 %p457, %r134, 0; mov.u64 %rd804, %rd806; mov.u64 %rd805, %rd806; @%p457 bra $L__BB193_510; ld.local.u32 %r623, [%rd4+248]; mul.wide.s32 %rd686, %r623, 80; add.s64 %rd687, %rd1, %rd686; add.s64 %rd804, %rd687, 304; add.s64 %rd688, %rd230, %rd686; add.s64 %rd805, %rd688, 304; mov.b32 %r624, %f547; mov.b32 %r625, %f546; mov.b32 %r626, %f545; mov.u32 %r627, 0; st.local.v4.u32 [%rd687+304], {%r626, %r625, %r624, %r627}; add.s32 %r628, %r623, 1; st.local.u32 [%rd4+248], %r628; add.s32 %r629, %r134, -1; st.local.u32 [%rd4+252], %r629; $L__BB193_510: setp.eq.s64 %p458, %rd805, 0; mov.u64 %rd807, %rd806; @%p458 bra $L__BB193_512; st.local.f32 [%rd804+16], %f548; mov.u64 %rd806, %rd804; mov.u64 %rd807, %rd805; $L__BB193_512: setp.eq.s64 %p459, %rd807, 0; @%p459 bra $L__BB193_517; add.s64 %rd229, %rd806, 20; st.local.f32 [%rd806+20], %f55; st.local.f32 [%rd806+24], %f56; st.local.f32 [%rd806+28], %f57; setp.eq.ftz.f32 %p460, %f60, 0f00000000; @%p460 bra $L__BB193_515; bra.uni $L__BB193_514; $L__BB193_515: mov.u32 %r633, 2; st.local.u32 [%rd229+-8], %r633; ld.local.u32 %r634, [%rd4+40]; or.b32 %r635, %r634, 12; st.local.u32 [%rd4+40], %r635; bra.uni $L__BB193_517; $L__BB193_341: ld.global.v4.u32 {%r375, %r376, %r377, %r378}, [%rd3+16]; ld.global.v4.u32 {%r379, %r380, %r381, %r382}, [%rd3+32]; ld.global.u32 %rd121, [%rd3+52]; ld.local.f32 %f378, [%rd4+256]; ld.local.f32 %f379, [%rd4+260]; ld.local.f32 %f380, [%rd4+264]; bfe.u32 %r82, %r144, 8, 8; and.b32 %r83, %r144, 255; setp.eq.s32 %p304, %r83, 255; @%p304 bra $L__BB193_343; bra.uni $L__BB193_342; $L__BB193_343: mov.b32 %f3036, %r145; bra.uni $L__BB193_344; $L__BB193_482: ld.local.f32 %f2816, [%rd4+256]; mov.u64 %rd788, 0; mul.ftz.f32 %f2817, %f2, %f2816; ld.local.f32 %f2818, [%rd4+260]; mul.ftz.f32 %f2819, %f2, %f2818; ld.local.f32 %f2820, [%rd4+264]; mul.ftz.f32 %f2821, %f2, %f2820; mov.f32 %f2822, 0f00000000; max.ftz.f32 %f515, %f2817, %f2822; max.ftz.f32 %f516, %f2819, %f2822; max.ftz.f32 %f517, %f2821, %f2822; add.ftz.f32 %f2823, %f515, %f516; add.ftz.f32 %f2824, %f517, %f2823; mul.ftz.f32 %f2825, %f2824, 0f3EAAAAAB; abs.ftz.f32 %f518, %f2825; setp.ltu.ftz.f32 %p422, %f518, 0f3727C5AC; mov.u64 %rd789, %rd788; @%p422 bra $L__BB193_487; ld.local.u32 %r129, [%rd4+252]; mov.u64 %rd788, 0; setp.eq.s32 %p423, %r129, 0; mov.u64 %rd786, %rd788; mov.u64 %rd787, %rd788; @%p423 bra $L__BB193_485; ld.local.u32 %r557, [%rd4+248]; mul.wide.s32 %rd636, %r557, 80; add.s64 %rd637, %rd1, %rd636; add.s64 %rd786, %rd637, 304; add.s64 %rd638, %rd230, %rd636; add.s64 %rd787, %rd638, 304; mov.b32 %r558, %f517; mov.b32 %r559, %f516; mov.b32 %r560, %f515; mov.u32 %r561, 0; st.local.v4.u32 [%rd637+304], {%r560, %r559, %r558, %r561}; add.s32 %r562, %r557, 1; st.local.u32 [%rd4+248], %r562; add.s32 %r563, %r129, -1; st.local.u32 [%rd4+252], %r563; $L__BB193_485: setp.eq.s64 %p424, %rd787, 0; mov.u64 %rd789, %rd788; @%p424 bra $L__BB193_487; st.local.f32 [%rd786+16], %f518; mov.u64 %rd788, %rd786; mov.u64 %rd789, %rd787; $L__BB193_487: setp.eq.s64 %p425, %rd789, 0; @%p425 bra $L__BB193_517; st.local.f32 [%rd788+20], %f55; st.local.f32 [%rd788+24], %f56; st.local.f32 [%rd788+28], %f57; cvt.ftz.sat.f32.f32 %f2826, %f60; st.local.f32 [%rd788+32], %f2826; mov.f32 %f2827, 0f3C23D70A; max.ftz.f32 %f2828, %f2826, %f2827; mul.ftz.f32 %f2829, %f2828, %f2828; rcp.approx.ftz.f32 %f2830, %f2829; st.local.f32 [%rd788+36], %f2830; mov.u32 %r564, 17; st.local.u32 [%rd788+12], %r564; ld.local.u32 %r565, [%rd4+40]; or.b32 %r566, %r565, 12; st.local.u32 [%rd4+40], %r566; bra.uni $L__BB193_517; $L__BB193_325: ld.local.f32 %f2486, [%rd4+256]; mul.ftz.f32 %f374, %f2, %f2486; ld.local.f32 %f2487, [%rd4+260]; mul.ftz.f32 %f375, %f2, %f2487; ld.local.f32 %f2488, [%rd4+264]; mul.ftz.f32 %f376, %f2, %f2488; add.ftz.f32 %f2489, %f374, %f375; add.ftz.f32 %f2490, %f376, %f2489; mul.ftz.f32 %f2491, %f2490, 0f3EAAAAAB; abs.ftz.f32 %f377, %f2491; setp.ltu.ftz.f32 %p294, %f377, 0f3727C5AC; @%p294 bra $L__BB193_517; ld.local.u32 %r67, [%rd4+40]; and.b32 %r362, %r67, 512; setp.eq.s32 %p295, %r362, 0; @%p295 bra $L__BB193_332; bra.uni $L__BB193_327; $L__BB193_332: or.b32 %r365, %r67, 516; st.local.u32 [%rd4+40], %r365; st.local.f32 [%rd4+280], %f374; st.local.f32 [%rd4+284], %f375; st.local.f32 [%rd4+288], %f376; and.b32 %r71, %r141, 491520; setp.eq.s32 %p299, %r71, 0; @%p299 bra $L__BB193_334; mov.u32 %r651, 1; st.local.u32 [%rd4+252], %r651; bra.uni $L__BB193_335; $L__BB193_44: ld.local.f32 %f524, [%rd5]; ld.local.f32 %f525, [%rd5+8]; ld.local.f32 %f526, [%rd5+4]; st.local.v2.f32 [%rd205+36], {%f524, %f526}; st.local.f32 [%rd205+44], %f525; ld.local.f32 %f527, [%rd6]; setp.eq.ftz.f32 %p434, %f527, 0f00000000; @%p434 bra $L__BB193_46; mul.ftz.f32 %f2841, %f527, 0f40C90FDB; cos.approx.ftz.f32 %f2842, %f2841; sin.approx.ftz.f32 %f2843, %f2841; mov.f32 %f2844, 0f3F800000; sub.ftz.f32 %f2845, %f2844, %f2842; mul.ftz.f32 %f2846, %f55, %f2845; fma.rn.ftz.f32 %f2847, %f55, %f2846, %f2842; mul.ftz.f32 %f2848, %f56, %f2846; mul.ftz.f32 %f2849, %f57, %f2843; sub.ftz.f32 %f2850, %f2848, %f2849; mul.ftz.f32 %f2851, %f526, %f2850; fma.rn.ftz.f32 %f2852, %f524, %f2847, %f2851; mul.ftz.f32 %f2853, %f57, %f2846; mul.ftz.f32 %f2854, %f56, %f2843; add.ftz.f32 %f2855, %f2854, %f2853; add.ftz.f32 %f2856, %f2849, %f2848; mul.ftz.f32 %f2857, %f56, %f2845; fma.rn.ftz.f32 %f2858, %f56, %f2857, %f2842; mul.ftz.f32 %f2859, %f526, %f2858; fma.rn.ftz.f32 %f2860, %f524, %f2856, %f2859; mul.ftz.f32 %f2861, %f57, %f2857; mul.ftz.f32 %f2862, %f55, %f2843; sub.ftz.f32 %f2863, %f2861, %f2862; sub.ftz.f32 %f2864, %f2853, %f2854; add.ftz.f32 %f2865, %f2862, %f2861; mul.ftz.f32 %f2866, %f526, %f2865; fma.rn.ftz.f32 %f2867, %f524, %f2864, %f2866; mul.ftz.f32 %f2868, %f57, %f2845; fma.rn.ftz.f32 %f2869, %f57, %f2868, %f2842; fma.rn.ftz.f32 %f2870, %f525, %f2869, %f2867; fma.rn.ftz.f32 %f2871, %f525, %f2863, %f2860; fma.rn.ftz.f32 %f2872, %f525, %f2855, %f2852; st.local.v2.f32 [%rd205+36], {%f2872, %f2871}; st.local.f32 [%rd205+44], %f2870; $L__BB193_46: mov.f32 %f2873, 0fBF7D70A4; max.ftz.f32 %f2874, %f63, %f2873; mov.f32 %f2875, 0f3F7D70A4; min.ftz.f32 %f528, %f2874, %f2875; setp.lt.ftz.f32 %p435, %f528, 0f00000000; @%p435 bra $L__BB193_489; bra.uni $L__BB193_47; $L__BB193_489: add.ftz.f32 %f2878, %f528, 0f3F800000; div.approx.ftz.f32 %f3059, %f3058, %f2878; mul.ftz.f32 %f3058, %f3058, %f2878; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; bra.uni $L__BB193_491; $L__BB193_310: mov.u32 %r358, 0; st.u32 [%rd111+-16], %r358; bra.uni $L__BB193_324; $L__BB193_87: ld.local.f32 %f2782, [%rd5]; ld.local.f32 %f2783, [%rd5+4]; mul.ftz.f32 %f2784, %f2783, %f2783; fma.rn.ftz.f32 %f2785, %f2782, %f2782, %f2784; ld.local.f32 %f2786, [%rd5+8]; fma.rn.ftz.f32 %f2787, %f2786, %f2786, %f2785; sqrt.approx.ftz.f32 %f2788, %f2787; div.approx.ftz.f32 %f2789, %f2782, %f2788; div.approx.ftz.f32 %f2790, %f2783, %f2788; div.approx.ftz.f32 %f2791, %f2786, %f2788; st.local.f32 [%rd192+12], %f2789; st.local.f32 [%rd192+16], %f2790; st.local.f32 [%rd192+20], %f2791; bra.uni $L__BB193_479; $L__BB193_312: mov.f32 %f1931, 0f3F8147AE; max.ftz.f32 %f1932, %f63, %f1931; mov.f32 %f1933, 0f40733333; min.ftz.f32 %f1934, %f1932, %f1933; rcp.approx.ftz.f32 %f1935, %f1934; mov.f32 %f1936, 0f3F800000; fma.rn.ftz.f32 %f1937, %f1935, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f1938, %f1935, %f1937, 0f3F2B020C; fma.rn.ftz.f32 %f1939, %f1934, 0f3D8240B8, %f1938; add.ftz.f32 %f1940, %f1939, 0f3F800000; mul.ftz.f32 %f1941, %f1940, 0f3FAAAAAB; sub.ftz.f32 %f1942, %f1936, %f1939; div.approx.ftz.f32 %f355, %f1941, %f1942; ld.f32 %f356, [%rd111+12]; setp.lt.ftz.f32 %p255, %f356, 0f38D1B717; mov.f32 %f3032, 0f00000000; mov.f32 %f3030, %f3032; @%p255 bra $L__BB193_315; setp.ge.ftz.f32 %p256, %f356, 0f3F7EB852; mov.f32 %f3030, 0f3F7FFFEF; @%p256 bra $L__BB193_315; sub.ftz.f32 %f1945, %f1936, 0f3F000000; mul.ftz.f32 %f1946, %f1945, 0f40400000; sqrt.approx.ftz.f32 %f1947, %f1946; mul.ftz.f32 %f1948, %f355, %f1947; mul.ftz.f32 %f1949, %f1948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1950, %f1949; add.ftz.f32 %f1951, %f1950, 0f3F800000; mul.ftz.f32 %f1952, %f1951, 0f3E800000; mul.ftz.f32 %f1953, %f1947, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1954, %f1953; mul.ftz.f32 %f1955, %f1954, %f1952; setp.lt.ftz.f32 %p257, %f1955, %f356; selp.f32 %f1956, 0f3F800000, 0f3F000000, %p257; selp.f32 %f1957, 0f3F000000, 0f00000000, %p257; selp.f32 %f1958, 0f3F400000, 0f3E800000, %p257; sub.ftz.f32 %f1959, %f1936, %f1958; mul.ftz.f32 %f1960, %f1959, 0f40400000; sqrt.approx.ftz.f32 %f1961, %f1960; mul.ftz.f32 %f1962, %f1958, 0f3F000000; mul.ftz.f32 %f1963, %f355, %f1961; mul.ftz.f32 %f1964, %f1963, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1965, 0f3F800000; mul.ftz.f32 %f1967, %f1966, %f1962; mul.ftz.f32 %f1968, %f1961, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1969, %f1968; mul.ftz.f32 %f1970, %f1969, %f1967; setp.lt.ftz.f32 %p258, %f1970, %f356; selp.f32 %f1971, %f1956, %f1958, %p258; selp.f32 %f1972, %f1958, %f1957, %p258; add.ftz.f32 %f1973, %f1971, %f1972; mul.ftz.f32 %f1974, %f1973, 0f3F000000; sub.ftz.f32 %f1975, %f1936, %f1974; mul.ftz.f32 %f1976, %f1975, 0f40400000; sqrt.approx.ftz.f32 %f1977, %f1976; mul.ftz.f32 %f1978, %f1974, 0f3F000000; mul.ftz.f32 %f1979, %f355, %f1977; mul.ftz.f32 %f1980, %f1979, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1981, %f1980; add.ftz.f32 %f1982, %f1981, 0f3F800000; mul.ftz.f32 %f1983, %f1982, %f1978; mul.ftz.f32 %f1984, %f1977, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f1985, %f1983; setp.lt.ftz.f32 %p259, %f1986, %f356; selp.f32 %f1987, %f1971, %f1974, %p259; selp.f32 %f1988, %f1974, %f1972, %p259; add.ftz.f32 %f1989, %f1987, %f1988; mul.ftz.f32 %f1990, %f1989, 0f3F000000; sub.ftz.f32 %f1991, %f1936, %f1990; mul.ftz.f32 %f1992, %f1991, 0f40400000; sqrt.approx.ftz.f32 %f1993, %f1992; mul.ftz.f32 %f1994, %f1990, 0f3F000000; mul.ftz.f32 %f1995, %f355, %f1993; mul.ftz.f32 %f1996, %f1995, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1997, %f1996; add.ftz.f32 %f1998, %f1997, 0f3F800000; mul.ftz.f32 %f1999, %f1998, %f1994; mul.ftz.f32 %f2000, %f1993, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2001, %f2000; mul.ftz.f32 %f2002, %f2001, %f1999; setp.lt.ftz.f32 %p260, %f2002, %f356; selp.f32 %f2003, %f1987, %f1990, %p260; selp.f32 %f2004, %f1990, %f1988, %p260; add.ftz.f32 %f2005, %f2003, %f2004; mul.ftz.f32 %f2006, %f2005, 0f3F000000; sub.ftz.f32 %f2007, %f1936, %f2006; mul.ftz.f32 %f2008, %f2007, 0f40400000; sqrt.approx.ftz.f32 %f2009, %f2008; mul.ftz.f32 %f2010, %f2006, 0f3F000000; mul.ftz.f32 %f2011, %f355, %f2009; mul.ftz.f32 %f2012, %f2011, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2013, %f2012; add.ftz.f32 %f2014, %f2013, 0f3F800000; mul.ftz.f32 %f2015, %f2014, %f2010; mul.ftz.f32 %f2016, %f2009, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2017, %f2016; mul.ftz.f32 %f2018, %f2017, %f2015; setp.lt.ftz.f32 %p261, %f2018, %f356; selp.f32 %f2019, %f2003, %f2006, %p261; selp.f32 %f2020, %f2006, %f2004, %p261; add.ftz.f32 %f2021, %f2019, %f2020; mul.ftz.f32 %f2022, %f2021, 0f3F000000; sub.ftz.f32 %f2023, %f1936, %f2022; mul.ftz.f32 %f2024, %f2023, 0f40400000; sqrt.approx.ftz.f32 %f2025, %f2024; mul.ftz.f32 %f2026, %f2022, 0f3F000000; mul.ftz.f32 %f2027, %f355, %f2025; mul.ftz.f32 %f2028, %f2027, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2029, %f2028; add.ftz.f32 %f2030, %f2029, 0f3F800000; mul.ftz.f32 %f2031, %f2030, %f2026; mul.ftz.f32 %f2032, %f2025, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2033, %f2032; mul.ftz.f32 %f2034, %f2033, %f2031; setp.lt.ftz.f32 %p262, %f2034, %f356; selp.f32 %f2035, %f2019, %f2022, %p262; selp.f32 %f2036, %f2022, %f2020, %p262; add.ftz.f32 %f2037, %f2035, %f2036; mul.ftz.f32 %f2038, %f2037, 0f3F000000; sub.ftz.f32 %f2039, %f1936, %f2038; mul.ftz.f32 %f2040, %f2039, 0f40400000; sqrt.approx.ftz.f32 %f2041, %f2040; mul.ftz.f32 %f2042, %f2038, 0f3F000000; mul.ftz.f32 %f2043, %f355, %f2041; mul.ftz.f32 %f2044, %f2043, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2045, 0f3F800000; mul.ftz.f32 %f2047, %f2046, %f2042; mul.ftz.f32 %f2048, %f2041, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2049, %f2048; mul.ftz.f32 %f2050, %f2049, %f2047; setp.lt.ftz.f32 %p263, %f2050, %f356; selp.f32 %f2051, %f2035, %f2038, %p263; selp.f32 %f2052, %f2038, %f2036, %p263; add.ftz.f32 %f2053, %f2051, %f2052; mul.ftz.f32 %f2054, %f2053, 0f3F000000; sub.ftz.f32 %f2055, %f1936, %f2054; mul.ftz.f32 %f2056, %f2055, 0f40400000; sqrt.approx.ftz.f32 %f2057, %f2056; mul.ftz.f32 %f2058, %f2054, 0f3F000000; mul.ftz.f32 %f2059, %f355, %f2057; mul.ftz.f32 %f2060, %f2059, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2061, %f2060; add.ftz.f32 %f2062, %f2061, 0f3F800000; mul.ftz.f32 %f2063, %f2062, %f2058; mul.ftz.f32 %f2064, %f2057, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2065, %f2064; mul.ftz.f32 %f2066, %f2065, %f2063; setp.lt.ftz.f32 %p264, %f2066, %f356; selp.f32 %f2067, %f2051, %f2054, %p264; selp.f32 %f2068, %f2054, %f2052, %p264; add.ftz.f32 %f2069, %f2067, %f2068; mul.ftz.f32 %f2070, %f2069, 0f3F000000; sub.ftz.f32 %f2071, %f1936, %f2070; mul.ftz.f32 %f2072, %f2071, 0f40400000; sqrt.approx.ftz.f32 %f2073, %f2072; mul.ftz.f32 %f2074, %f2070, 0f3F000000; mul.ftz.f32 %f2075, %f355, %f2073; mul.ftz.f32 %f2076, %f2075, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2077, %f2076; add.ftz.f32 %f2078, %f2077, 0f3F800000; mul.ftz.f32 %f2079, %f2078, %f2074; mul.ftz.f32 %f2080, %f2073, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2081, %f2080; mul.ftz.f32 %f2082, %f2081, %f2079; setp.lt.ftz.f32 %p265, %f2082, %f356; selp.f32 %f2083, %f2067, %f2070, %p265; selp.f32 %f2084, %f2070, %f2068, %p265; add.ftz.f32 %f2085, %f2083, %f2084; mul.ftz.f32 %f2086, %f2085, 0f3F000000; sub.ftz.f32 %f2087, %f1936, %f2086; mul.ftz.f32 %f2088, %f2087, 0f40400000; sqrt.approx.ftz.f32 %f2089, %f2088; mul.ftz.f32 %f2090, %f2086, 0f3F000000; mul.ftz.f32 %f2091, %f355, %f2089; mul.ftz.f32 %f2092, %f2091, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2093, %f2092; add.ftz.f32 %f2094, %f2093, 0f3F800000; mul.ftz.f32 %f2095, %f2094, %f2090; mul.ftz.f32 %f2096, %f2089, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2097, %f2096; mul.ftz.f32 %f2098, %f2097, %f2095; setp.lt.ftz.f32 %p266, %f2098, %f356; selp.f32 %f2099, %f2083, %f2086, %p266; selp.f32 %f2100, %f2086, %f2084, %p266; add.ftz.f32 %f2101, %f2099, %f2100; mul.ftz.f32 %f2102, %f2101, 0f3F000000; sub.ftz.f32 %f2103, %f1936, %f2102; mul.ftz.f32 %f2104, %f2103, 0f40400000; sqrt.approx.ftz.f32 %f2105, %f2104; mul.ftz.f32 %f2106, %f2102, 0f3F000000; mul.ftz.f32 %f2107, %f355, %f2105; mul.ftz.f32 %f2108, %f2107, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2109, %f2108; add.ftz.f32 %f2110, %f2109, 0f3F800000; mul.ftz.f32 %f2111, %f2110, %f2106; mul.ftz.f32 %f2112, %f2105, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2113, %f2112; mul.ftz.f32 %f2114, %f2113, %f2111; setp.lt.ftz.f32 %p267, %f2114, %f356; selp.f32 %f2115, %f2099, %f2102, %p267; selp.f32 %f2116, %f2102, %f2100, %p267; add.ftz.f32 %f2117, %f2115, %f2116; mul.ftz.f32 %f3030, %f2117, 0f3F000000; $L__BB193_315: ld.f32 %f359, [%rd111+16]; setp.lt.ftz.f32 %p268, %f359, 0f38D1B717; mov.f32 %f3031, %f3032; @%p268 bra $L__BB193_318; setp.ge.ftz.f32 %p269, %f359, 0f3F7EB852; mov.f32 %f3031, 0f3F7FFFEF; @%p269 bra $L__BB193_318; sub.ftz.f32 %f2121, %f1936, 0f3F000000; mul.ftz.f32 %f2122, %f2121, 0f40400000; sqrt.approx.ftz.f32 %f2123, %f2122; mul.ftz.f32 %f2124, %f355, %f2123; mul.ftz.f32 %f2125, %f2124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2126, %f2125; add.ftz.f32 %f2127, %f2126, 0f3F800000; mul.ftz.f32 %f2128, %f2127, 0f3E800000; mul.ftz.f32 %f2129, %f2123, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2130, %f2129; mul.ftz.f32 %f2131, %f2130, %f2128; setp.lt.ftz.f32 %p270, %f2131, %f359; selp.f32 %f2132, 0f3F800000, 0f3F000000, %p270; selp.f32 %f2133, 0f3F000000, 0f00000000, %p270; selp.f32 %f2134, 0f3F400000, 0f3E800000, %p270; sub.ftz.f32 %f2135, %f1936, %f2134; mul.ftz.f32 %f2136, %f2135, 0f40400000; sqrt.approx.ftz.f32 %f2137, %f2136; mul.ftz.f32 %f2138, %f2134, 0f3F000000; mul.ftz.f32 %f2139, %f355, %f2137; mul.ftz.f32 %f2140, %f2139, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2141, %f2140; add.ftz.f32 %f2142, %f2141, 0f3F800000; mul.ftz.f32 %f2143, %f2142, %f2138; mul.ftz.f32 %f2144, %f2137, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2145, %f2144; mul.ftz.f32 %f2146, %f2145, %f2143; setp.lt.ftz.f32 %p271, %f2146, %f359; selp.f32 %f2147, %f2132, %f2134, %p271; selp.f32 %f2148, %f2134, %f2133, %p271; add.ftz.f32 %f2149, %f2147, %f2148; mul.ftz.f32 %f2150, %f2149, 0f3F000000; sub.ftz.f32 %f2151, %f1936, %f2150; mul.ftz.f32 %f2152, %f2151, 0f40400000; sqrt.approx.ftz.f32 %f2153, %f2152; mul.ftz.f32 %f2154, %f2150, 0f3F000000; mul.ftz.f32 %f2155, %f355, %f2153; mul.ftz.f32 %f2156, %f2155, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2157, %f2156; add.ftz.f32 %f2158, %f2157, 0f3F800000; mul.ftz.f32 %f2159, %f2158, %f2154; mul.ftz.f32 %f2160, %f2153, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2161, %f2160; mul.ftz.f32 %f2162, %f2161, %f2159; setp.lt.ftz.f32 %p272, %f2162, %f359; selp.f32 %f2163, %f2147, %f2150, %p272; selp.f32 %f2164, %f2150, %f2148, %p272; add.ftz.f32 %f2165, %f2163, %f2164; mul.ftz.f32 %f2166, %f2165, 0f3F000000; sub.ftz.f32 %f2167, %f1936, %f2166; mul.ftz.f32 %f2168, %f2167, 0f40400000; sqrt.approx.ftz.f32 %f2169, %f2168; mul.ftz.f32 %f2170, %f2166, 0f3F000000; mul.ftz.f32 %f2171, %f355, %f2169; mul.ftz.f32 %f2172, %f2171, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2173, %f2172; add.ftz.f32 %f2174, %f2173, 0f3F800000; mul.ftz.f32 %f2175, %f2174, %f2170; mul.ftz.f32 %f2176, %f2169, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2177, %f2176; mul.ftz.f32 %f2178, %f2177, %f2175; setp.lt.ftz.f32 %p273, %f2178, %f359; selp.f32 %f2179, %f2163, %f2166, %p273; selp.f32 %f2180, %f2166, %f2164, %p273; add.ftz.f32 %f2181, %f2179, %f2180; mul.ftz.f32 %f2182, %f2181, 0f3F000000; sub.ftz.f32 %f2183, %f1936, %f2182; mul.ftz.f32 %f2184, %f2183, 0f40400000; sqrt.approx.ftz.f32 %f2185, %f2184; mul.ftz.f32 %f2186, %f2182, 0f3F000000; mul.ftz.f32 %f2187, %f355, %f2185; mul.ftz.f32 %f2188, %f2187, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2189, %f2188; add.ftz.f32 %f2190, %f2189, 0f3F800000; mul.ftz.f32 %f2191, %f2190, %f2186; mul.ftz.f32 %f2192, %f2185, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2193, %f2192; mul.ftz.f32 %f2194, %f2193, %f2191; setp.lt.ftz.f32 %p274, %f2194, %f359; selp.f32 %f2195, %f2179, %f2182, %p274; selp.f32 %f2196, %f2182, %f2180, %p274; add.ftz.f32 %f2197, %f2195, %f2196; mul.ftz.f32 %f2198, %f2197, 0f3F000000; sub.ftz.f32 %f2199, %f1936, %f2198; mul.ftz.f32 %f2200, %f2199, 0f40400000; sqrt.approx.ftz.f32 %f2201, %f2200; mul.ftz.f32 %f2202, %f2198, 0f3F000000; mul.ftz.f32 %f2203, %f355, %f2201; mul.ftz.f32 %f2204, %f2203, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2205, %f2204; add.ftz.f32 %f2206, %f2205, 0f3F800000; mul.ftz.f32 %f2207, %f2206, %f2202; mul.ftz.f32 %f2208, %f2201, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2209, %f2208; mul.ftz.f32 %f2210, %f2209, %f2207; setp.lt.ftz.f32 %p275, %f2210, %f359; selp.f32 %f2211, %f2195, %f2198, %p275; selp.f32 %f2212, %f2198, %f2196, %p275; add.ftz.f32 %f2213, %f2211, %f2212; mul.ftz.f32 %f2214, %f2213, 0f3F000000; sub.ftz.f32 %f2215, %f1936, %f2214; mul.ftz.f32 %f2216, %f2215, 0f40400000; sqrt.approx.ftz.f32 %f2217, %f2216; mul.ftz.f32 %f2218, %f2214, 0f3F000000; mul.ftz.f32 %f2219, %f355, %f2217; mul.ftz.f32 %f2220, %f2219, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2221, %f2220; add.ftz.f32 %f2222, %f2221, 0f3F800000; mul.ftz.f32 %f2223, %f2222, %f2218; mul.ftz.f32 %f2224, %f2217, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2225, %f2224; mul.ftz.f32 %f2226, %f2225, %f2223; setp.lt.ftz.f32 %p276, %f2226, %f359; selp.f32 %f2227, %f2211, %f2214, %p276; selp.f32 %f2228, %f2214, %f2212, %p276; add.ftz.f32 %f2229, %f2227, %f2228; mul.ftz.f32 %f2230, %f2229, 0f3F000000; sub.ftz.f32 %f2231, %f1936, %f2230; mul.ftz.f32 %f2232, %f2231, 0f40400000; sqrt.approx.ftz.f32 %f2233, %f2232; mul.ftz.f32 %f2234, %f2230, 0f3F000000; mul.ftz.f32 %f2235, %f355, %f2233; mul.ftz.f32 %f2236, %f2235, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2237, %f2236; add.ftz.f32 %f2238, %f2237, 0f3F800000; mul.ftz.f32 %f2239, %f2238, %f2234; mul.ftz.f32 %f2240, %f2233, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2241, %f2240; mul.ftz.f32 %f2242, %f2241, %f2239; setp.lt.ftz.f32 %p277, %f2242, %f359; selp.f32 %f2243, %f2227, %f2230, %p277; selp.f32 %f2244, %f2230, %f2228, %p277; add.ftz.f32 %f2245, %f2243, %f2244; mul.ftz.f32 %f2246, %f2245, 0f3F000000; sub.ftz.f32 %f2247, %f1936, %f2246; mul.ftz.f32 %f2248, %f2247, 0f40400000; sqrt.approx.ftz.f32 %f2249, %f2248; mul.ftz.f32 %f2250, %f2246, 0f3F000000; mul.ftz.f32 %f2251, %f355, %f2249; mul.ftz.f32 %f2252, %f2251, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2253, %f2252; add.ftz.f32 %f2254, %f2253, 0f3F800000; mul.ftz.f32 %f2255, %f2254, %f2250; mul.ftz.f32 %f2256, %f2249, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2257, %f2256; mul.ftz.f32 %f2258, %f2257, %f2255; setp.lt.ftz.f32 %p278, %f2258, %f359; selp.f32 %f2259, %f2243, %f2246, %p278; selp.f32 %f2260, %f2246, %f2244, %p278; add.ftz.f32 %f2261, %f2259, %f2260; mul.ftz.f32 %f2262, %f2261, 0f3F000000; sub.ftz.f32 %f2263, %f1936, %f2262; mul.ftz.f32 %f2264, %f2263, 0f40400000; sqrt.approx.ftz.f32 %f2265, %f2264; mul.ftz.f32 %f2266, %f2262, 0f3F000000; mul.ftz.f32 %f2267, %f355, %f2265; mul.ftz.f32 %f2268, %f2267, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2269, %f2268; add.ftz.f32 %f2270, %f2269, 0f3F800000; mul.ftz.f32 %f2271, %f2270, %f2266; mul.ftz.f32 %f2272, %f2265, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2273, %f2272; mul.ftz.f32 %f2274, %f2273, %f2271; setp.lt.ftz.f32 %p279, %f2274, %f359; selp.f32 %f2275, %f2259, %f2262, %p279; selp.f32 %f2276, %f2262, %f2260, %p279; add.ftz.f32 %f2277, %f2275, %f2276; mul.ftz.f32 %f2278, %f2277, 0f3F000000; sub.ftz.f32 %f2279, %f1936, %f2278; mul.ftz.f32 %f2280, %f2279, 0f40400000; sqrt.approx.ftz.f32 %f2281, %f2280; mul.ftz.f32 %f2282, %f2278, 0f3F000000; mul.ftz.f32 %f2283, %f355, %f2281; mul.ftz.f32 %f2284, %f2283, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2285, %f2284; add.ftz.f32 %f2286, %f2285, 0f3F800000; mul.ftz.f32 %f2287, %f2286, %f2282; mul.ftz.f32 %f2288, %f2281, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f2289, %f2287; setp.lt.ftz.f32 %p280, %f2290, %f359; selp.f32 %f2291, %f2275, %f2278, %p280; selp.f32 %f2292, %f2278, %f2276, %p280; add.ftz.f32 %f2293, %f2291, %f2292; mul.ftz.f32 %f3031, %f2293, 0f3F000000; $L__BB193_318: ld.f32 %f362, [%rd111+20]; setp.lt.ftz.f32 %p281, %f362, 0f38D1B717; @%p281 bra $L__BB193_321; setp.ge.ftz.f32 %p282, %f362, 0f3F7EB852; mov.f32 %f3032, 0f3F7FFFEF; @%p282 bra $L__BB193_321; sub.ftz.f32 %f2297, %f1936, 0f3F000000; mul.ftz.f32 %f2298, %f2297, 0f40400000; sqrt.approx.ftz.f32 %f2299, %f2298; mul.ftz.f32 %f2300, %f355, %f2299; mul.ftz.f32 %f2301, %f2300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2302, %f2301; add.ftz.f32 %f2303, %f2302, 0f3F800000; mul.ftz.f32 %f2304, %f2303, 0f3E800000; mul.ftz.f32 %f2305, %f2299, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2306, %f2305; mul.ftz.f32 %f2307, %f2306, %f2304; setp.lt.ftz.f32 %p283, %f2307, %f362; selp.f32 %f2308, 0f3F800000, 0f3F000000, %p283; selp.f32 %f2309, 0f3F000000, 0f00000000, %p283; selp.f32 %f2310, 0f3F400000, 0f3E800000, %p283; sub.ftz.f32 %f2311, %f1936, %f2310; mul.ftz.f32 %f2312, %f2311, 0f40400000; sqrt.approx.ftz.f32 %f2313, %f2312; mul.ftz.f32 %f2314, %f2310, 0f3F000000; mul.ftz.f32 %f2315, %f355, %f2313; mul.ftz.f32 %f2316, %f2315, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2317, %f2316; add.ftz.f32 %f2318, %f2317, 0f3F800000; mul.ftz.f32 %f2319, %f2318, %f2314; mul.ftz.f32 %f2320, %f2313, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2321, %f2320; mul.ftz.f32 %f2322, %f2321, %f2319; setp.lt.ftz.f32 %p284, %f2322, %f362; selp.f32 %f2323, %f2308, %f2310, %p284; selp.f32 %f2324, %f2310, %f2309, %p284; add.ftz.f32 %f2325, %f2323, %f2324; mul.ftz.f32 %f2326, %f2325, 0f3F000000; sub.ftz.f32 %f2327, %f1936, %f2326; mul.ftz.f32 %f2328, %f2327, 0f40400000; sqrt.approx.ftz.f32 %f2329, %f2328; mul.ftz.f32 %f2330, %f2326, 0f3F000000; mul.ftz.f32 %f2331, %f355, %f2329; mul.ftz.f32 %f2332, %f2331, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2333, %f2332; add.ftz.f32 %f2334, %f2333, 0f3F800000; mul.ftz.f32 %f2335, %f2334, %f2330; mul.ftz.f32 %f2336, %f2329, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2337, %f2336; mul.ftz.f32 %f2338, %f2337, %f2335; setp.lt.ftz.f32 %p285, %f2338, %f362; selp.f32 %f2339, %f2323, %f2326, %p285; selp.f32 %f2340, %f2326, %f2324, %p285; add.ftz.f32 %f2341, %f2339, %f2340; mul.ftz.f32 %f2342, %f2341, 0f3F000000; sub.ftz.f32 %f2343, %f1936, %f2342; mul.ftz.f32 %f2344, %f2343, 0f40400000; sqrt.approx.ftz.f32 %f2345, %f2344; mul.ftz.f32 %f2346, %f2342, 0f3F000000; mul.ftz.f32 %f2347, %f355, %f2345; mul.ftz.f32 %f2348, %f2347, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2349, %f2348; add.ftz.f32 %f2350, %f2349, 0f3F800000; mul.ftz.f32 %f2351, %f2350, %f2346; mul.ftz.f32 %f2352, %f2345, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2353, %f2352; mul.ftz.f32 %f2354, %f2353, %f2351; setp.lt.ftz.f32 %p286, %f2354, %f362; selp.f32 %f2355, %f2339, %f2342, %p286; selp.f32 %f2356, %f2342, %f2340, %p286; add.ftz.f32 %f2357, %f2355, %f2356; mul.ftz.f32 %f2358, %f2357, 0f3F000000; sub.ftz.f32 %f2359, %f1936, %f2358; mul.ftz.f32 %f2360, %f2359, 0f40400000; sqrt.approx.ftz.f32 %f2361, %f2360; mul.ftz.f32 %f2362, %f2358, 0f3F000000; mul.ftz.f32 %f2363, %f355, %f2361; mul.ftz.f32 %f2364, %f2363, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2365, %f2364; add.ftz.f32 %f2366, %f2365, 0f3F800000; mul.ftz.f32 %f2367, %f2366, %f2362; mul.ftz.f32 %f2368, %f2361, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2369, %f2368; mul.ftz.f32 %f2370, %f2369, %f2367; setp.lt.ftz.f32 %p287, %f2370, %f362; selp.f32 %f2371, %f2355, %f2358, %p287; selp.f32 %f2372, %f2358, %f2356, %p287; add.ftz.f32 %f2373, %f2371, %f2372; mul.ftz.f32 %f2374, %f2373, 0f3F000000; sub.ftz.f32 %f2375, %f1936, %f2374; mul.ftz.f32 %f2376, %f2375, 0f40400000; sqrt.approx.ftz.f32 %f2377, %f2376; mul.ftz.f32 %f2378, %f2374, 0f3F000000; mul.ftz.f32 %f2379, %f355, %f2377; mul.ftz.f32 %f2380, %f2379, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2381, %f2380; add.ftz.f32 %f2382, %f2381, 0f3F800000; mul.ftz.f32 %f2383, %f2382, %f2378; mul.ftz.f32 %f2384, %f2377, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2385, %f2384; mul.ftz.f32 %f2386, %f2385, %f2383; setp.lt.ftz.f32 %p288, %f2386, %f362; selp.f32 %f2387, %f2371, %f2374, %p288; selp.f32 %f2388, %f2374, %f2372, %p288; add.ftz.f32 %f2389, %f2387, %f2388; mul.ftz.f32 %f2390, %f2389, 0f3F000000; sub.ftz.f32 %f2391, %f1936, %f2390; mul.ftz.f32 %f2392, %f2391, 0f40400000; sqrt.approx.ftz.f32 %f2393, %f2392; mul.ftz.f32 %f2394, %f2390, 0f3F000000; mul.ftz.f32 %f2395, %f355, %f2393; mul.ftz.f32 %f2396, %f2395, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2397, %f2396; add.ftz.f32 %f2398, %f2397, 0f3F800000; mul.ftz.f32 %f2399, %f2398, %f2394; mul.ftz.f32 %f2400, %f2393, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f2401, %f2399; setp.lt.ftz.f32 %p289, %f2402, %f362; selp.f32 %f2403, %f2387, %f2390, %p289; selp.f32 %f2404, %f2390, %f2388, %p289; add.ftz.f32 %f2405, %f2403, %f2404; mul.ftz.f32 %f2406, %f2405, 0f3F000000; sub.ftz.f32 %f2407, %f1936, %f2406; mul.ftz.f32 %f2408, %f2407, 0f40400000; sqrt.approx.ftz.f32 %f2409, %f2408; mul.ftz.f32 %f2410, %f2406, 0f3F000000; mul.ftz.f32 %f2411, %f355, %f2409; mul.ftz.f32 %f2412, %f2411, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2413, %f2412; add.ftz.f32 %f2414, %f2413, 0f3F800000; mul.ftz.f32 %f2415, %f2414, %f2410; mul.ftz.f32 %f2416, %f2409, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2417, %f2416; mul.ftz.f32 %f2418, %f2417, %f2415; setp.lt.ftz.f32 %p290, %f2418, %f362; selp.f32 %f2419, %f2403, %f2406, %p290; selp.f32 %f2420, %f2406, %f2404, %p290; add.ftz.f32 %f2421, %f2419, %f2420; mul.ftz.f32 %f2422, %f2421, 0f3F000000; sub.ftz.f32 %f2423, %f1936, %f2422; mul.ftz.f32 %f2424, %f2423, 0f40400000; sqrt.approx.ftz.f32 %f2425, %f2424; mul.ftz.f32 %f2426, %f2422, 0f3F000000; mul.ftz.f32 %f2427, %f355, %f2425; mul.ftz.f32 %f2428, %f2427, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2429, %f2428; add.ftz.f32 %f2430, %f2429, 0f3F800000; mul.ftz.f32 %f2431, %f2430, %f2426; mul.ftz.f32 %f2432, %f2425, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2433, %f2432; mul.ftz.f32 %f2434, %f2433, %f2431; setp.lt.ftz.f32 %p291, %f2434, %f362; selp.f32 %f2435, %f2419, %f2422, %p291; selp.f32 %f2436, %f2422, %f2420, %p291; add.ftz.f32 %f2437, %f2435, %f2436; mul.ftz.f32 %f2438, %f2437, 0f3F000000; sub.ftz.f32 %f2439, %f1936, %f2438; mul.ftz.f32 %f2440, %f2439, 0f40400000; sqrt.approx.ftz.f32 %f2441, %f2440; mul.ftz.f32 %f2442, %f2438, 0f3F000000; mul.ftz.f32 %f2443, %f355, %f2441; mul.ftz.f32 %f2444, %f2443, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2445, %f2444; add.ftz.f32 %f2446, %f2445, 0f3F800000; mul.ftz.f32 %f2447, %f2446, %f2442; mul.ftz.f32 %f2448, %f2441, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2449, %f2448; mul.ftz.f32 %f2450, %f2449, %f2447; setp.lt.ftz.f32 %p292, %f2450, %f362; selp.f32 %f2451, %f2435, %f2438, %p292; selp.f32 %f2452, %f2438, %f2436, %p292; add.ftz.f32 %f2453, %f2451, %f2452; mul.ftz.f32 %f2454, %f2453, 0f3F000000; sub.ftz.f32 %f2455, %f1936, %f2454; mul.ftz.f32 %f2456, %f2455, 0f40400000; sqrt.approx.ftz.f32 %f2457, %f2456; mul.ftz.f32 %f2458, %f2454, 0f3F000000; mul.ftz.f32 %f2459, %f355, %f2457; mul.ftz.f32 %f2460, %f2459, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2461, %f2460; add.ftz.f32 %f2462, %f2461, 0f3F800000; mul.ftz.f32 %f2463, %f2462, %f2458; mul.ftz.f32 %f2464, %f2457, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2465, %f2464; mul.ftz.f32 %f2466, %f2465, %f2463; setp.lt.ftz.f32 %p293, %f2466, %f362; selp.f32 %f2467, %f2451, %f2454, %p293; selp.f32 %f2468, %f2454, %f2452, %p293; add.ftz.f32 %f2469, %f2467, %f2468; mul.ftz.f32 %f3032, %f2469, 0f3F000000; $L__BB193_321: sub.ftz.f32 %f2471, %f1936, %f3030; mul.ftz.f32 %f2472, %f2471, 0f40400000; sub.ftz.f32 %f2473, %f1936, %f3031; mul.ftz.f32 %f2474, %f2473, 0f40400000; sub.ftz.f32 %f2475, %f1936, %f3032; mul.ftz.f32 %f2476, %f2475, 0f40400000; sqrt.approx.ftz.f32 %f2477, %f2472; sqrt.approx.ftz.f32 %f2478, %f2474; sqrt.approx.ftz.f32 %f2479, %f2476; ld.f32 %f2480, [%rd111]; mul.ftz.f32 %f3033, %f2477, %f2480; ld.f32 %f2481, [%rd111+4]; mul.ftz.f32 %f3034, %f2478, %f2481; ld.f32 %f2482, [%rd111+8]; mul.ftz.f32 %f3035, %f2479, %f2482; $L__BB193_323: st.f32 [%rd111], %f3033; st.f32 [%rd111+4], %f3034; st.f32 [%rd111+8], %f3035; or.b32 %r648, %r648, 16; $L__BB193_324: ld.local.u32 %r360, [%rd4+40]; or.b32 %r361, %r360, %r648; st.local.u32 [%rd4+40], %r361; bra.uni $L__BB193_517; $L__BB193_98: mul.wide.u32 %rd279, %r166, 4; add.s64 %rd280, %rd2, %rd279; ld.local.f32 %f2982, [%rd280]; ld.local.f32 %f2983, [%rd280+4]; ld.local.f32 %f2984, [%rd280+8]; $L__BB193_100: ld.global.v4.u32 {%r170, %r171, %r172, %r173}, [%rd3+48]; setp.eq.s32 %p56, %r170, 255; @%p56 bra $L__BB193_102; bra.uni $L__BB193_101; $L__BB193_102: ld.local.f32 %f2992, [%rd4+8]; ld.local.f32 %f2991, [%rd4+4]; ld.local.f32 %f2990, [%rd4]; bra.uni $L__BB193_103; $L__BB193_342: mul.wide.u32 %rd477, %r83, 4; add.s64 %rd478, %rd2, %rd477; ld.local.f32 %f3036, [%rd478]; $L__BB193_344: setp.eq.s32 %p305, %r82, 255; mul.ftz.f32 %f384, %f2, %f380; mul.ftz.f32 %f385, %f2, %f379; mul.ftz.f32 %f386, %f2, %f378; @%p305 bra $L__BB193_346; bra.uni $L__BB193_345; $L__BB193_346: mov.b32 %f3037, %r146; bra.uni $L__BB193_347; $L__BB193_101: mul.wide.u32 %rd281, %r170, 4; add.s64 %rd282, %rd2, %rd281; ld.local.f32 %f2992, [%rd282+8]; ld.local.f32 %f2991, [%rd282+4]; ld.local.f32 %f2990, [%rd282]; $L__BB193_103: @%p10 bra $L__BB193_112; ld.local.v2.f32 {%f677, %f678}, [%rd4+12]; mul.ftz.f32 %f679, %f2991, %f91; fma.rn.ftz.f32 %f680, %f2990, %f90, %f679; fma.rn.ftz.f32 %f681, %f2992, %f92, %f680; add.ftz.f32 %f682, %f681, %f681; mul.ftz.f32 %f683, %f2990, %f682; mul.ftz.f32 %f684, %f2991, %f682; mul.ftz.f32 %f685, %f2992, %f682; sub.ftz.f32 %f686, %f683, %f90; sub.ftz.f32 %f687, %f684, %f91; sub.ftz.f32 %f688, %f685, %f92; mul.ftz.f32 %f689, %f678, %f91; fma.rn.ftz.f32 %f690, %f677, %f90, %f689; ld.local.f32 %f123, [%rd4+20]; fma.rn.ftz.f32 %f124, %f123, %f92, %f690; mul.ftz.f32 %f691, %f124, 0f3F666666; mov.f32 %f692, 0f3C23D70A; min.ftz.f32 %f125, %f691, %f692; mul.ftz.f32 %f693, %f678, %f687; fma.rn.ftz.f32 %f694, %f677, %f686, %f693; fma.rn.ftz.f32 %f695, %f123, %f688, %f694; setp.ltu.ftz.f32 %p58, %f695, %f125; @%p58 bra $L__BB193_105; bra.uni $L__BB193_112; $L__BB193_105: mul.ftz.f32 %f696, %f2991, %f678; fma.rn.ftz.f32 %f697, %f2990, %f677, %f696; fma.rn.ftz.f32 %f698, %f2992, %f123, %f697; mul.ftz.f32 %f699, %f677, %f698; mul.ftz.f32 %f700, %f678, %f698; mul.ftz.f32 %f701, %f123, %f698; sub.ftz.f32 %f702, %f2990, %f699; sub.ftz.f32 %f703, %f2991, %f700; sub.ftz.f32 %f704, %f2992, %f701; mul.ftz.f32 %f705, %f703, %f703; fma.rn.ftz.f32 %f706, %f702, %f702, %f705; fma.rn.ftz.f32 %f707, %f704, %f704, %f706; sqrt.approx.ftz.f32 %f708, %f707; div.approx.ftz.f32 %f129, %f702, %f708; div.approx.ftz.f32 %f130, %f703, %f708; div.approx.ftz.f32 %f131, %f704, %f708; mul.ftz.f32 %f709, %f91, %f130; fma.rn.ftz.f32 %f710, %f90, %f129, %f709; fma.rn.ftz.f32 %f132, %f92, %f131, %f710; mul.ftz.f32 %f711, %f132, %f132; fma.rn.ftz.f32 %f712, %f124, %f124, %f711; mul.ftz.f32 %f713, %f125, %f125; sub.ftz.f32 %f714, %f712, %f713; mul.ftz.f32 %f715, %f711, %f714; mov.f32 %f716, 0f00000000; max.ftz.f32 %f717, %f715, %f716; sqrt.approx.ftz.f32 %f718, %f717; fma.rn.ftz.f32 %f719, %f125, %f124, %f712; mov.f32 %f720, 0f3F000000; div.approx.ftz.f32 %f721, %f720, %f712; add.ftz.f32 %f722, %f718, %f719; mul.ftz.f32 %f133, %f721, %f722; sub.ftz.f32 %f723, %f719, %f718; mul.ftz.f32 %f134, %f721, %f723; setp.gt.ftz.f32 %p59, %f133, 0f3727C5AC; setp.le.ftz.f32 %p60, %f133, 0f3F800054; and.pred %p3, %p59, %p60; setp.gt.ftz.f32 %p61, %f134, 0f3727C5AC; setp.le.ftz.f32 %p62, %f134, 0f3F800054; and.pred %p4, %p61, %p62; and.pred %p63, %p3, %p4; @%p63 bra $L__BB193_108; bra.uni $L__BB193_106; $L__BB193_108: mov.f32 %f730, 0f3F800000; sub.ftz.f32 %f731, %f730, %f133; max.ftz.f32 %f733, %f731, %f716; sqrt.approx.ftz.f32 %f140, %f733; max.ftz.f32 %f734, %f133, %f716; sqrt.approx.ftz.f32 %f141, %f734; sub.ftz.f32 %f735, %f730, %f134; max.ftz.f32 %f736, %f735, %f716; sqrt.approx.ftz.f32 %f142, %f736; max.ftz.f32 %f737, %f134, %f716; sqrt.approx.ftz.f32 %f143, %f737; mul.ftz.f32 %f738, %f124, %f141; fma.rn.ftz.f32 %f739, %f132, %f140, %f738; add.ftz.f32 %f740, %f739, %f739; mul.ftz.f32 %f741, %f141, %f740; sub.ftz.f32 %f144, %f741, %f124; mul.ftz.f32 %f742, %f124, %f143; fma.rn.ftz.f32 %f743, %f132, %f142, %f742; add.ftz.f32 %f744, %f743, %f743; mul.ftz.f32 %f745, %f143, %f744; sub.ftz.f32 %f145, %f745, %f124; setp.ge.ftz.f32 %p65, %f144, 0f3727C5AC; setp.ge.ftz.f32 %p66, %f145, 0f3727C5AC; and.pred %p67, %p65, %p66; @%p67 bra $L__BB193_110; bra.uni $L__BB193_109; $L__BB193_110: setp.lt.ftz.f32 %p69, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p69; selp.f32 %f2989, %f141, %f143, %p69; bra.uni $L__BB193_111; $L__BB193_345: mul.wide.u32 %rd479, %r82, 4; add.s64 %rd480, %rd2, %rd479; ld.local.f32 %f3037, [%rd480]; $L__BB193_347: and.b32 %r86, %r375, 255; bfe.u32 %r87, %r375, 8, 8; bfe.u32 %r88, %r375, 16, 8; bfe.u32 %r89, %r379, 8, 8; bfe.u32 %r90, %r379, 16, 8; shr.u32 %r91, %r379, 24; ld.local.u32 %r92, [%rd4+64]; ld.local.s32 %rd122, [%rd4+48]; setp.eq.s32 %p306, %r92, -1; mov.u32 %r113, -1; mov.u32 %r110, 0; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p306 bra $L__BB193_362; ld.const.u64 %rd481, [kernel_params+2160]; cvta.to.global.u64 %rd482, %rd481; mul.wide.s32 %rd483, %r92, 208; add.s64 %rd484, %rd482, %rd483; ld.global.u32 %r93, [%rd484+164]; and.b32 %r388, %r11, 1; setp.eq.b32 %p307, %r388, 1; mov.pred %p308, 0; xor.pred %p309, %p307, %p308; not.pred %p310, %p309; mov.u32 %r653, 0; @%p310 bra $L__BB193_352; cvt.u32.u64 %r390, %rd122; setp.eq.s32 %p311, %r390, -1; mov.u32 %r652, -1; @%p311 bra $L__BB193_351; ld.const.u64 %rd485, [kernel_params+2240]; cvta.to.global.u64 %rd486, %rd485; shl.b64 %rd487, %rd122, 2; add.s64 %rd488, %rd486, %rd487; ld.global.u32 %r652, [%rd488]; $L__BB193_351: setp.ne.s32 %p312, %r652, -1; selp.u32 %r653, 1, 0, %p312; $L__BB193_352: ld.const.u64 %rd489, [kernel_params+2312]; cvta.to.global.u64 %rd123, %rd489; add.s32 %r655, %r653, %r93; mul.wide.u32 %rd490, %r655, 16; add.s64 %rd491, %rd123, %rd490; add.s64 %rd754, %rd491, 14; ld.global.u16 %rs20, [%rd491+12]; ld.global.u32 %r654, [%rd491+8]; ld.global.u64 %rd752, [%rd491]; setp.eq.s64 %p313, %rd752, %rd121; add.s64 %rd753, %rd491, 15; @%p313 bra $L__BB193_358; $L__BB193_353: setp.eq.s64 %p314, %rd752, 0; @%p314 bra $L__BB193_355; add.s32 %r655, %r655, 2; bra.uni $L__BB193_356; $L__BB193_355: setp.eq.s16 %p315, %rs20, 0; mov.u32 %r110, 0; mov.u32 %r655, %r654; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p315 bra $L__BB193_362; $L__BB193_356: cvt.u64.u32 %rd128, %r655; mul.wide.u32 %rd492, %r655, 16; add.s64 %rd493, %rd123, %rd492; ld.global.u16 %rs20, [%rd493+12]; ld.global.u32 %r654, [%rd493+8]; ld.global.u64 %rd752, [%rd493]; setp.ne.s64 %p316, %rd752, %rd121; @%p316 bra $L__BB193_353; shl.b64 %rd494, %rd128, 4; add.s64 %rd495, %rd123, %rd494; add.s64 %rd754, %rd495, 14; add.s64 %rd753, %rd495, 15; $L__BB193_358: ld.global.u8 %rs5, [%rd754]; ld.global.u8 %rs6, [%rd753]; cvt.u32.u64 %r395, %rd122; setp.ne.s32 %p317, %r395, -1; @%p317 bra $L__BB193_361; mov.u32 %r110, 0; add.s16 %rs14, %rs20, -1; setp.lt.u16 %p318, %rs14, 2; @%p318 bra $L__BB193_361; setp.ne.s16 %p319, %rs20, 1024; mov.u32 %r659, %r110; mov.u32 %r660, %r110; @%p319 bra $L__BB193_362; $L__BB193_361: setp.eq.s16 %p320, %rs20, 0; selp.b32 %r113, -1, %r654, %p320; cvt.u32.u16 %r110, %rs20; cvt.u32.u16 %r400, %rs6; and.b32 %r660, %r400, 255; cvt.u32.u16 %r401, %rs5; and.b32 %r659, %r401, 255; $L__BB193_362: setp.eq.s32 %p321, %r113, -1; @%p321 bra $L__BB193_382; and.b32 %r402, %r11, 1; setp.eq.b32 %p322, %r402, 1; mov.pred %p323, 0; xor.pred %p324, %p322, %p323; not.pred %p325, %p324; @%p325 bra $L__BB193_375; bra.uni $L__BB193_364; $L__BB193_375: setp.eq.s32 %p332, %r12, 0; @%p332 bra $L__BB193_380; and.b32 %r423, %r110, 768; setp.eq.s32 %p333, %r423, 0; @%p333 bra $L__BB193_378; ld.const.u64 %rd517, [kernel_params+2264]; cvta.to.global.u64 %rd518, %rd517; shl.b64 %rd519, %rd122, 4; add.s64 %rd520, %rd518, %rd519; shr.s32 %r424, %r11, 7; ld.global.u32 %r425, [%rd520+4]; add.s32 %r426, %r424, %r425; ld.const.u64 %rd521, [kernel_params+2320]; cvta.to.global.u64 %rd522, %rd521; add.s32 %r427, %r426, %r113; mul.wide.s32 %rd523, %r427, 4; add.s64 %rd524, %rd522, %rd523; ld.local.f32 %f2520, [%rd4+56]; mov.f32 %f2521, 0f3F800000; sub.ftz.f32 %f2522, %f2521, %f2520; ld.global.f32 %f2523, [%rd524]; mul.ftz.f32 %f2524, %f2523, %f2522; ld.global.f32 %f2525, [%rd524+4]; fma.rn.ftz.f32 %f3038, %f2525, %f2520, %f2524; bra.uni $L__BB193_385; $L__BB193_472: mul.ftz.f32 %f2768, %f60, %f60; st.local.v2.f32 [%rd185+12], {%f2768, %f2768}; st.local.f32 [%rd185+20], %f3057; cvt.ftz.sat.f32.f32 %f507, %f2768; setp.eq.s32 %p411, %r2, 22; @%p411 bra $L__BB193_474; bra.uni $L__BB193_473; $L__BB193_474: st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r534, 22; st.local.u32 [%rd185+-8], %r534; ld.local.u32 %r535, [%rd4+40]; or.b32 %r536, %r535, 12; st.local.u32 [%rd4+40], %r536; bra.uni $L__BB193_517; $L__BB193_60: mov.u32 %r617, 19; st.local.u32 [%rd222+-8], %r617; st.local.f32 [%rd222+12], %f543; st.local.f32 [%rd222+16], %f544; ld.local.u32 %r618, [%rd4+40]; or.b32 %r619, %r618, 12; st.local.u32 [%rd4+40], %r619; bra.uni $L__BB193_517; $L__BB193_47: mov.f32 %f2876, 0f3F800000; sub.ftz.f32 %f2877, %f2876, %f528; mul.ftz.f32 %f3059, %f3058, %f2877; div.approx.ftz.f32 %f3058, %f3058, %f2877; st.local.v2.f32 [%rd205+12], {%f3059, %f3058}; $L__BB193_491: setp.eq.s32 %p436, %r2, 9; @%p436 bra $L__BB193_505; bra.uni $L__BB193_492; $L__BB193_505: mov.u32 %r597, 9; st.local.u32 [%rd205+-8], %r597; ld.local.u32 %r598, [%rd4+40]; or.b32 %r599, %r598, 4; st.local.u32 [%rd4+40], %r599; bra.uni $L__BB193_517; $L__BB193_492: setp.eq.s32 %p437, %r2, 13; @%p437 bra $L__BB193_504; bra.uni $L__BB193_493; $L__BB193_504: cvt.ftz.sat.f32.f32 %f2920, %f3058; cvt.ftz.sat.f32.f32 %f2921, %f3059; st.local.v2.f32 [%rd205+12], {%f2921, %f2920}; mov.u32 %r594, 13; st.local.u32 [%rd205+-8], %r594; ld.local.u32 %r595, [%rd4+40]; or.b32 %r596, %r595, 12; st.local.u32 [%rd4+40], %r596; bra.uni $L__BB193_517; $L__BB193_493: setp.eq.s32 %p438, %r2, 10; @%p438 bra $L__BB193_503; bra.uni $L__BB193_494; $L__BB193_503: mov.u64 %rd663, 0; st.local.u64 [%rd205+28], %rd663; cvt.ftz.sat.f32.f32 %f2918, %f3058; cvt.ftz.sat.f32.f32 %f2919, %f3059; st.local.v2.f32 [%rd205+12], {%f2919, %f2918}; mov.u32 %r591, 10; st.local.u32 [%rd205+-8], %r591; ld.local.u32 %r592, [%rd4+40]; or.b32 %r593, %r592, 12; st.local.u32 [%rd4+40], %r593; bra.uni $L__BB193_517; $L__BB193_441: setp.eq.s32 %p375, %r2, 25; st.local.v2.f32 [%rd160+12], {%f480, %f480}; st.local.f32 [%rd160+20], %f3054; cvt.ftz.sat.f32.f32 %f485, %f480; @%p375 bra $L__BB193_443; bra.uni $L__BB193_442; $L__BB193_443: st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r472, 13; st.local.u32 [%rd160+-8], %r472; ld.local.u32 %r473, [%rd4+40]; or.b32 %r474, %r473, 12; st.local.u32 [%rd4+40], %r474; bra.uni $L__BB193_445; $L__BB193_382: setp.eq.s32 %p339, %r89, 255; @%p339 bra $L__BB193_384; bra.uni $L__BB193_383; $L__BB193_384: mov.b32 %f3038, %r380; bra.uni $L__BB193_385; $L__BB193_364: cvt.u32.u64 %r404, %rd122; setp.eq.s32 %p326, %r404, -1; mov.u32 %r662, -1; @%p326 bra $L__BB193_366; ld.const.u64 %rd496, [kernel_params+2240]; cvta.to.global.u64 %rd497, %rd496; shl.b64 %rd498, %rd122, 2; add.s64 %rd499, %rd497, %rd498; ld.global.u32 %r662, [%rd499]; $L__BB193_366: setp.eq.s32 %p327, %r662, -1; @%p327 bra $L__BB193_368; mov.u64 %rd500, 0; { // callseq 955, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd230; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r110; st.param.b32 [param1+4], %r659; st.param.b32 [param1+8], %r660; st.param.b32 [param1+12], %r113; .param .b64 param2; st.param.b64 [param2+0], %rd500; .param .b64 param3; st.param.b64 [param3+0], %rd500; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f3038, [retval0+0]; } // callseq 955 bra.uni $L__BB193_385; $L__BB193_327: ld.local.f32 %f2492, [%rd4+280]; add.ftz.f32 %f2493, %f374, %f2492; ld.local.f32 %f2494, [%rd4+284]; add.ftz.f32 %f2495, %f375, %f2494; ld.local.f32 %f2496, [%rd4+288]; add.ftz.f32 %f2497, %f376, %f2496; st.local.f32 [%rd4+280], %f2493; st.local.f32 [%rd4+284], %f2495; st.local.f32 [%rd4+288], %f2497; ld.local.u32 %r68, [%rd4+248]; setp.lt.s32 %p296, %r68, 1; @%p296 bra $L__BB193_517; mov.u32 %r650, 0; $L__BB193_329: mul.wide.s32 %rd472, %r650, 80; add.s64 %rd473, %rd1, %rd472; add.s64 %rd118, %rd473, 316; ld.local.u32 %r364, [%rd473+316]; setp.eq.s32 %p297, %r364, 31; add.s32 %r650, %r650, 1; @%p297 bra $L__BB193_331; setp.lt.s32 %p298, %r650, %r68; @%p298 bra $L__BB193_329; bra.uni $L__BB193_517; $L__BB193_478: ld.local.f32 %f2802, [%rd4+116]; ld.local.f32 %f2803, [%rd4+120]; mul.ftz.f32 %f2804, %f2803, %f2803; fma.rn.ftz.f32 %f2805, %f2802, %f2802, %f2804; ld.local.f32 %f2806, [%rd4+124]; fma.rn.ftz.f32 %f2807, %f2806, %f2806, %f2805; sqrt.approx.ftz.f32 %f2808, %f2807; div.approx.ftz.f32 %f2809, %f2802, %f2808; div.approx.ftz.f32 %f2810, %f2803, %f2808; div.approx.ftz.f32 %f2811, %f2806, %f2808; st.local.f32 [%rd192+12], %f2809; st.local.f32 [%rd192+16], %f2810; st.local.f32 [%rd192+20], %f2811; mov.u32 %r547, 0; st.local.u32 [%rd192+32], %r547; $L__BB193_479: setp.eq.s32 %p418, %r2, 20; mov.f32 %f2812, 0f3A83126F; max.ftz.f32 %f2813, %f60, %f2812; mov.f32 %f2814, 0f3F800000; min.ftz.f32 %f513, %f2813, %f2814; max.ftz.f32 %f2815, %f63, %f2812; min.ftz.f32 %f514, %f2815, %f2814; @%p418 bra $L__BB193_481; bra.uni $L__BB193_480; $L__BB193_481: mov.u32 %r551, 20; st.local.u32 [%rd192+-8], %r551; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r552, [%rd4+40]; or.b32 %r553, %r552, 12; st.local.u32 [%rd4+40], %r553; bra.uni $L__BB193_517; $L__BB193_480: mov.u32 %r548, 30; st.local.u32 [%rd192+-8], %r548; st.local.f32 [%rd192+24], %f513; st.local.f32 [%rd192+28], %f514; ld.local.u32 %r549, [%rd4+40]; or.b32 %r550, %r549, 12; st.local.u32 [%rd4+40], %r550; bra.uni $L__BB193_517; $L__BB193_514: st.local.f32 [%rd229+12], %f60; mov.u32 %r630, 3; st.local.u32 [%rd229+-8], %r630; cvt.ftz.sat.f32.f32 %f2952, %f60; fma.rn.ftz.f32 %f2953, %f2952, 0f3F67750B, 0f40490FDB; rcp.approx.ftz.f32 %f2954, %f2953; st.local.f32 [%rd229+16], %f2954; mul.ftz.f32 %f2955, %f2952, %f2954; st.local.f32 [%rd229+20], %f2955; ld.local.u32 %r631, [%rd4+40]; or.b32 %r632, %r631, 12; st.local.u32 [%rd4+40], %r632; bra.uni $L__BB193_517; $L__BB193_473: st.local.u64 [%rd185+28], %rd621; st.local.v2.f32 [%rd185+12], {%f507, %f507}; mov.u32 %r531, 23; st.local.u32 [%rd185+-8], %r531; ld.local.u32 %r532, [%rd4+40]; or.b32 %r533, %r532, 12; st.local.u32 [%rd4+40], %r533; bra.uni $L__BB193_517; $L__BB193_494: setp.eq.s32 %p439, %r2, 14; @%p439 bra $L__BB193_496; bra.uni $L__BB193_495; $L__BB193_496: ld.local.u32 %r131, [%rd4+252]; setp.lt.s32 %p440, %r131, 1; @%p440 bra $L__BB193_498; add.s32 %r582, %r131, -1; st.local.u32 [%rd4+252], %r582; ld.local.s32 %rd651, [%rd4+248]; cvt.s64.s32 %rd652, %r582; add.s64 %rd653, %rd651, %rd652; mul.lo.s64 %rd654, %rd653, 80; add.s64 %rd655, %rd1, %rd654; add.s64 %rd794, %rd655, 304; add.s64 %rd656, %rd230, %rd654; add.s64 %rd795, %rd656, 304; bra.uni $L__BB193_499; $L__BB193_457: ld.local.u32 %r508, [%rd4+248]; mov.u64 %rd776, 0; add.s32 %r509, %r508, -1; st.local.u32 [%rd4+248], %r509; add.s32 %r510, %r126, 1; st.local.u32 [%rd4+252], %r510; mov.u64 %rd777, %rd776; $L__BB193_458: setp.eq.s64 %p394, %rd777, 0; @%p394 bra $L__BB193_517; add.s64 %rd178, %rd774, 20; st.local.f32 [%rd774+20], %f55; st.local.f32 [%rd774+24], %f56; st.local.f32 [%rd774+28], %f57; st.local.u64 [%rd774+48], %rd777; mov.u32 %r511, 0; st.local.v2.f32 [%rd774+56], {%f2732, %f2732}; st.local.u32 [%rd774+64], %r511; mul.ftz.f32 %f2737, %f60, %f60; st.local.v2.f32 [%rd774+32], {%f2737, %f2737}; mov.f32 %f2738, 0f3727C5AC; max.ftz.f32 %f3056, %f63, %f2738; ld.local.u32 %r512, [%rd4+40]; and.b32 %r513, %r512, 1; setp.eq.b32 %p395, %r513, 1; mov.pred %p396, 0; xor.pred %p397, %p395, %p396; not.pred %p398, %p397; @%p398 bra $L__BB193_461; rcp.approx.ftz.f32 %f3056, %f3056; $L__BB193_461: st.local.f32 [%rd178+20], %f3056; ld.local.f32 %f2739, [%rd6]; ld.local.f32 %f2740, [%rd6+4]; ld.local.f32 %f2741, [%rd6+8]; st.local.f32 [%rd776], %f2739; st.local.f32 [%rd776+4], %f2740; st.local.f32 [%rd776+8], %f2741; ld.local.u64 %rd609, [%rd178+28]; st.u32 [%rd609+12], %r511; st.u32 [%rd609+16], %r511; st.u32 [%rd609+20], %r511; ld.local.u64 %rd610, [%rd178+28]; st.u32 [%rd610+36], %r511; ld.local.f32 %f2743, [%rd178+12]; mov.f32 %f2744, 0f38D1B717; max.ftz.f32 %f2745, %f2743, %f2744; mov.f32 %f2746, 0f3F800000; min.ftz.f32 %f2747, %f2745, %f2746; st.local.v2.f32 [%rd178+12], {%f2747, %f2747}; ld.local.f32 %f2748, [%rd178+20]; max.ftz.f32 %f2749, %f2732, %f2748; st.local.f32 [%rd178+20], %f2749; ld.local.u64 %rd611, [%rd178+28]; ld.f32 %f2750, [%rd611]; cvt.ftz.sat.f32.f32 %f2751, %f2750; ld.f32 %f2752, [%rd611+4]; cvt.ftz.sat.f32.f32 %f2753, %f2752; ld.f32 %f2754, [%rd611+8]; cvt.ftz.sat.f32.f32 %f2755, %f2754; st.f32 [%rd611], %f2751; st.f32 [%rd611+4], %f2753; st.f32 [%rd611+8], %f2755; mov.u32 %r515, 24; st.local.u32 [%rd178+-8], %r515; ld.local.u32 %r516, [%rd4+40]; or.b32 %r517, %r516, 1036; st.local.u32 [%rd4+40], %r517; bra.uni $L__BB193_517; $L__BB193_442: st.local.u64 [%rd160+28], %rd579; st.local.v2.f32 [%rd160+12], {%f485, %f485}; mov.u32 %r469, 10; st.local.u32 [%rd160+-8], %r469; ld.local.u32 %r470, [%rd4+40]; or.b32 %r471, %r470, 12; st.local.u32 [%rd4+40], %r471; $L__BB193_445: setp.ne.s32 %p376, %r122, 0; ld.const.u32 %r478, [kernel_params+1568]; setp.eq.s32 %p377, %r478, 0; and.pred %p378, %p376, %p377; @%p378 bra $L__BB193_517; mov.f32 %f2713, 0f3F800000; sub.ftz.f32 %f2714, %f2713, %f3055; mov.f32 %f2715, 0f38D1B717; max.ftz.f32 %f2716, %f2715, %f2714; mul.ftz.f32 %f2717, %f470, %f2716; mul.ftz.f32 %f2718, %f471, %f2716; mul.ftz.f32 %f2719, %f472, %f2716; mov.f32 %f2720, 0f00000000; max.ftz.f32 %f487, %f2717, %f2720; max.ftz.f32 %f488, %f2718, %f2720; max.ftz.f32 %f489, %f2719, %f2720; add.ftz.f32 %f2721, %f487, %f488; add.ftz.f32 %f2722, %f489, %f2721; mul.ftz.f32 %f2723, %f2722, 0f3EAAAAAB; abs.ftz.f32 %f490, %f2723; setp.ltu.ftz.f32 %p379, %f490, 0f3727C5AC; mov.u64 %rd770, 0; mov.u64 %rd771, %rd770; @%p379 bra $L__BB193_451; ld.local.u32 %r124, [%rd4+252]; mov.u64 %rd770, 0; setp.eq.s32 %p380, %r124, 0; mov.u64 %rd768, %rd770; mov.u64 %rd769, %rd770; @%p380 bra $L__BB193_449; ld.local.u32 %r479, [%rd4+248]; mul.wide.s32 %rd585, %r479, 80; add.s64 %rd586, %rd1, %rd585; add.s64 %rd768, %rd586, 304; add.s64 %rd587, %rd230, %rd585; add.s64 %rd769, %rd587, 304; mov.b32 %r480, %f489; mov.b32 %r481, %f488; mov.b32 %r482, %f487; mov.u32 %r483, 0; st.local.v4.u32 [%rd586+304], {%r482, %r481, %r480, %r483}; add.s32 %r484, %r479, 1; st.local.u32 [%rd4+248], %r484; add.s32 %r485, %r124, -1; st.local.u32 [%rd4+252], %r485; $L__BB193_449: setp.eq.s64 %p381, %rd769, 0; mov.u64 %rd771, %rd770; @%p381 bra $L__BB193_451; st.local.f32 [%rd768+16], %f490; mov.u64 %rd770, %rd768; mov.u64 %rd771, %rd769; $L__BB193_451: setp.eq.s64 %p382, %rd771, 0; @%p382 bra $L__BB193_517; add.s64 %rd167, %rd770, 20; mov.u64 %rd590, 0; st.local.f32 [%rd770+20], %f55; st.local.f32 [%rd770+24], %f56; st.local.f32 [%rd770+28], %f57; mov.u32 %r486, 0; st.local.v2.f32 [%rd770+56], {%f2720, %f2720}; st.local.u32 [%rd770+64], %r486; st.local.u64 [%rd770+48], %rd590; setp.eq.s32 %p383, %r2, 28; @%p383 bra $L__BB193_456; bra.uni $L__BB193_453; $L__BB193_456: st.local.v2.f32 [%rd167+12], {%f2720, %f2720}; st.local.f32 [%rd167+20], %f3054; mov.u32 %r493, 21; st.local.u32 [%rd167+-8], %r493; ld.local.u32 %r494, [%rd4+40]; or.b32 %r495, %r494, 4; st.local.u32 [%rd4+40], %r495; bra.uni $L__BB193_517; $L__BB193_453: setp.eq.s32 %p384, %r2, 25; st.local.v2.f32 [%rd167+12], {%f480, %f480}; st.local.f32 [%rd167+20], %f3054; cvt.ftz.sat.f32.f32 %f491, %f480; @%p384 bra $L__BB193_455; bra.uni $L__BB193_454; $L__BB193_455: st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r490, 22; st.local.u32 [%rd167+-8], %r490; ld.local.u32 %r491, [%rd4+40]; or.b32 %r492, %r491, 12; st.local.u32 [%rd4+40], %r492; bra.uni $L__BB193_517; $L__BB193_454: st.local.u64 [%rd167+28], %rd590; st.local.v2.f32 [%rd167+12], {%f491, %f491}; mov.u32 %r487, 23; st.local.u32 [%rd167+-8], %r487; ld.local.u32 %r488, [%rd4+40]; or.b32 %r489, %r488, 12; st.local.u32 [%rd4+40], %r489; bra.uni $L__BB193_517; $L__BB193_383: mul.wide.u32 %rd533, %r89, 4; add.s64 %rd534, %rd2, %rd533; ld.local.f32 %f3038, [%rd534]; bra.uni $L__BB193_385; $L__BB193_380: and.b32 %r432, %r11, 8; setp.eq.s32 %p336, %r432, 0; setp.ne.s32 %p337, %r110, 8; mov.f32 %f3038, 0f00000000; or.pred %p338, %p337, %p336; @%p338 bra $L__BB193_385; cvt.u32.u64 %r433, %rd122; ld.const.u64 %rd529, [kernel_params+2320]; cvta.to.global.u64 %rd530, %rd529; add.s32 %r434, %r433, %r113; mul.wide.s32 %rd531, %r434, 4; add.s64 %rd532, %rd530, %rd531; ld.global.f32 %f3038, [%rd532]; bra.uni $L__BB193_385; $L__BB193_368: and.b32 %r405, %r110, 56; setp.eq.s32 %p328, %r405, 0; @%p328 bra $L__BB193_373; and.b32 %r406, %r110, 24; setp.eq.s32 %p329, %r406, 0; @%p329 bra $L__BB193_371; ld.const.u64 %rd501, [kernel_params+2232]; cvta.to.global.u64 %rd502, %rd501; shl.b64 %rd503, %rd122, 4; add.s64 %rd504, %rd502, %rd503; ld.global.v4.u32 {%r407, %r408, %r409, %r410}, [%rd504]; ld.const.u64 %rd505, [kernel_params+2320]; cvta.to.global.u64 %rd506, %rd505; add.s32 %r414, %r407, %r113; mul.wide.u32 %rd507, %r414, 4; add.s64 %rd755, %rd506, %rd507; add.s32 %r415, %r408, %r113; mul.wide.u32 %rd508, %r415, 4; add.s64 %rd756, %rd506, %rd508; add.s32 %r416, %r409, %r113; mul.wide.u32 %rd509, %r416, 4; add.s64 %rd757, %rd506, %rd509; bra.uni $L__BB193_372; $L__BB193_495: mov.f32 %f2880, 0f38D1B717; max.ftz.f32 %f2881, %f3059, %f2880; mov.f32 %f2882, 0f3F800000; max.ftz.f32 %f2883, %f3058, %f2880; min.ftz.f32 %f2884, %f2883, %f2882; min.ftz.f32 %f2885, %f2881, %f2882; st.local.v2.f32 [%rd205+12], {%f2885, %f2884}; mov.u32 %r579, 16; st.local.u32 [%rd205+-8], %r579; ld.local.u32 %r580, [%rd4+40]; or.b32 %r581, %r580, 12; st.local.u32 [%rd4+40], %r581; bra.uni $L__BB193_517; $L__BB193_334: ld.local.u32 %r651, [%rd4+252]; $L__BB193_335: setp.eq.s32 %p301, %r651, 0; mov.pred %p462, -1; mov.u64 %rd751, 0; @%p301 bra $L__BB193_337; ld.local.u32 %r367, [%rd4+248]; mul.wide.s32 %rd475, %r367, 80; add.s64 %rd476, %rd1, %rd475; add.s64 %rd751, %rd476, 304; mov.b32 %r368, %f376; mov.b32 %r369, %f375; mov.b32 %r370, %f374; mov.u32 %r371, 31; st.local.v4.u32 [%rd476+304], {%r370, %r369, %r368, %r371}; add.s32 %r372, %r367, 1; st.local.u32 [%rd4+248], %r372; add.s32 %r373, %r651, -1; st.local.u32 [%rd4+252], %r373; mov.pred %p462, 0; $L__BB193_337: @%p462 bra $L__BB193_339; bra.uni $L__BB193_338; $L__BB193_339: @%p299 bra $L__BB193_517; mov.u32 %r374, 0; st.local.u32 [%rd4+252], %r374; bra.uni $L__BB193_517; $L__BB193_338: st.local.f32 [%rd751+16], %f377; ld.local.f32 %f2506, [%rd4]; ld.local.f32 %f2507, [%rd4+4]; ld.local.f32 %f2508, [%rd4+8]; st.local.f32 [%rd751+20], %f2506; st.local.f32 [%rd751+24], %f2507; st.local.f32 [%rd751+28], %f2508; bra.uni $L__BB193_517; $L__BB193_106: or.pred %p64, %p3, %p4; mov.f32 %f2990, %f677; mov.f32 %f2991, %f678; mov.f32 %f2992, %f123; @%p64 bra $L__BB193_107; bra.uni $L__BB193_112; $L__BB193_107: selp.f32 %f724, %f133, %f134, %p3; mov.f32 %f725, 0f3F800000; sub.ftz.f32 %f726, %f725, %f724; max.ftz.f32 %f728, %f726, %f716; sqrt.approx.ftz.f32 %f2988, %f728; max.ftz.f32 %f729, %f724, %f716; sqrt.approx.ftz.f32 %f2989, %f729; bra.uni $L__BB193_111; $L__BB193_331: ld.local.f32 %f2498, [%rd118+-12]; add.ftz.f32 %f2499, %f374, %f2498; ld.local.f32 %f2500, [%rd118+-8]; add.ftz.f32 %f2501, %f375, %f2500; ld.local.f32 %f2502, [%rd118+-4]; add.ftz.f32 %f2503, %f376, %f2502; st.local.f32 [%rd118+-12], %f2499; st.local.f32 [%rd118+-8], %f2501; st.local.f32 [%rd118+-4], %f2503; ld.local.f32 %f2504, [%rd118+4]; add.ftz.f32 %f2505, %f377, %f2504; st.local.f32 [%rd118+4], %f2505; bra.uni $L__BB193_517; $L__BB193_378: and.b32 %r428, %r110, 131; setp.eq.s32 %p334, %r428, 0; mov.f32 %f3038, 0f00000000; @%p334 bra $L__BB193_385; cvt.u32.u64 %r429, %rd122; setp.eq.s32 %p335, %r110, 128; selp.b32 %r430, %r429, 0, %p335; add.s32 %r431, %r113, %r430; ld.const.u64 %rd525, [kernel_params+2320]; cvta.to.global.u64 %rd526, %rd525; mul.wide.s32 %rd527, %r431, 4; add.s64 %rd528, %rd526, %rd527; ld.global.f32 %f3038, [%rd528]; bra.uni $L__BB193_385; $L__BB193_109: setp.gt.ftz.f32 %p68, %f144, %f145; selp.f32 %f2988, %f140, %f142, %p68; selp.f32 %f2989, %f141, %f143, %p68; $L__BB193_111: mul.ftz.f32 %f746, %f677, %f2989; mul.ftz.f32 %f747, %f678, %f2989; mul.ftz.f32 %f748, %f123, %f2989; fma.rn.ftz.f32 %f2992, %f131, %f2988, %f748; fma.rn.ftz.f32 %f2991, %f130, %f2988, %f747; fma.rn.ftz.f32 %f2990, %f129, %f2988, %f746; $L__BB193_112: setp.eq.s32 %p70, %r171, 255; mov.f32 %f2993, 0f3F800000; mov.f32 %f2994, %f2993; mov.f32 %f2995, %f2993; @%p70 bra $L__BB193_114; mul.wide.u32 %rd283, %r171, 4; add.s64 %rd284, %rd2, %rd283; ld.local.f32 %f2993, [%rd284]; ld.local.f32 %f2994, [%rd284+4]; ld.local.f32 %f2995, [%rd284+8]; $L__BB193_114: setp.eq.s32 %p71, %r172, 255; mov.f32 %f2996, 0f3FB33333; @%p71 bra $L__BB193_116; mul.wide.u32 %rd285, %r172, 4; add.s64 %rd286, %rd2, %rd285; ld.local.f32 %f2996, [%rd286]; $L__BB193_116: setp.eq.s32 %p72, %r173, 255; mov.f32 %f2997, 0f00000000; @%p72 bra $L__BB193_118; mul.wide.u32 %rd287, %r173, 4; add.s64 %rd288, %rd2, %rd287; ld.local.f32 %f2997, [%rd288]; $L__BB193_118: ld.global.v4.u32 {%r174, %r175, %r176, %r177}, [%rd3+64]; setp.eq.s32 %p73, %r174, 255; @%p73 bra $L__BB193_120; bra.uni $L__BB193_119; $L__BB193_120: mov.b32 %f2998, %r175; mov.b32 %f2999, %r176; mov.b32 %f3000, %r177; bra.uni $L__BB193_121; $L__BB193_119: mul.wide.u32 %rd289, %r174, 4; add.s64 %rd290, %rd2, %rd289; ld.local.f32 %f2998, [%rd290]; ld.local.f32 %f2999, [%rd290+4]; ld.local.f32 %f3000, [%rd290+8]; $L__BB193_121: ld.local.f32 %f754, [%rd4+256]; mul.ftz.f32 %f177, %f2, %f754; ld.local.f32 %f755, [%rd4+260]; mul.ftz.f32 %f178, %f2, %f755; ld.local.f32 %f756, [%rd4+264]; mul.ftz.f32 %f179, %f2, %f756; mov.f32 %f757, 0f3F800000; sub.ftz.f32 %f758, %f757, %f63; mul.ftz.f32 %f759, %f63, %f2998; fma.rn.ftz.f32 %f180, %f758, %f2982, %f759; mul.ftz.f32 %f760, %f63, %f2999; fma.rn.ftz.f32 %f181, %f758, %f2983, %f760; mul.ftz.f32 %f761, %f63, %f3000; fma.rn.ftz.f32 %f182, %f758, %f2984, %f761; mul.ftz.f32 %f762, %f180, %f177; mul.ftz.f32 %f763, %f181, %f178; mul.ftz.f32 %f764, %f182, %f179; mul.ftz.f32 %f183, %f97, %f762; mul.ftz.f32 %f184, %f97, %f763; mul.ftz.f32 %f185, %f97, %f764; and.b32 %r178, %r141, 4096; setp.eq.s32 %p74, %r178, 0; selp.f32 %f186, %f2982, %f180, %p74; selp.f32 %f187, %f2983, %f181, %p74; selp.f32 %f188, %f2984, %f182, %p74; selp.f32 %f189, %f63, 0f00000000, %p74; add.ftz.f32 %f765, %f180, %f181; add.ftz.f32 %f766, %f182, %f765; mul.ftz.f32 %f767, %f766, 0f3EAAAAAB; abs.ftz.f32 %f768, %f767; setp.leu.ftz.f32 %p75, %f768, 0f3727C5AC; @%p75 bra $L__BB193_182; setp.le.ftz.f32 %p76, %f189, 0f3727C5AC; setp.gt.ftz.f32 %p77, %f97, 0f3727C5AC; and.pred %p78, %p76, %p77; @%p78 bra $L__BB193_175; bra.uni $L__BB193_123; $L__BB193_175: mul.ftz.f32 %f1391, %f186, %f177; mul.ftz.f32 %f1392, %f97, %f1391; mul.ftz.f32 %f1393, %f187, %f178; mul.ftz.f32 %f1394, %f97, %f1393; mul.ftz.f32 %f1395, %f188, %f179; mul.ftz.f32 %f1396, %f97, %f1395; mov.f32 %f1397, 0f00000000; max.ftz.f32 %f223, %f1392, %f1397; max.ftz.f32 %f224, %f1394, %f1397; max.ftz.f32 %f225, %f1396, %f1397; add.ftz.f32 %f1398, %f223, %f224; add.ftz.f32 %f1399, %f225, %f1398; mul.ftz.f32 %f1400, %f1399, 0f3EAAAAAB; abs.ftz.f32 %f226, %f1400; setp.ltu.ftz.f32 %p143, %f226, 0f3727C5AC; mov.u64 %rd708, 0; mov.u64 %rd709, %rd708; @%p143 bra $L__BB193_180; ld.local.u32 %r42, [%rd4+252]; mov.u64 %rd708, 0; setp.eq.s32 %p144, %r42, 0; mov.u64 %rd706, %rd708; mov.u64 %rd707, %rd708; @%p144 bra $L__BB193_178; ld.local.u32 %r226, [%rd4+248]; mul.wide.s32 %rd328, %r226, 80; add.s64 %rd329, %rd1, %rd328; add.s64 %rd706, %rd329, 304; add.s64 %rd330, %rd230, %rd328; add.s64 %rd707, %rd330, 304; mov.b32 %r227, %f225; mov.b32 %r228, %f224; mov.b32 %r229, %f223; mov.u32 %r230, 0; st.local.v4.u32 [%rd329+304], {%r229, %r228, %r227, %r230}; add.s32 %r231, %r226, 1; st.local.u32 [%rd4+248], %r231; add.s32 %r232, %r42, -1; st.local.u32 [%rd4+252], %r232; $L__BB193_178: setp.eq.s64 %p145, %rd707, 0; mov.u64 %rd709, %rd708; @%p145 bra $L__BB193_180; st.local.f32 [%rd706+16], %f226; mov.u64 %rd708, %rd706; mov.u64 %rd709, %rd707; $L__BB193_180: setp.eq.s64 %p146, %rd709, 0; @%p146 bra $L__BB193_182; st.local.f32 [%rd708+20], %f55; st.local.f32 [%rd708+24], %f56; st.local.f32 [%rd708+28], %f57; st.local.f32 [%rd708+32], %f71; mov.u32 %r233, 5; st.local.u32 [%rd708+12], %r233; mov.u32 %r234, 1; st.local.u32 [%rd708+36], %r234; ld.local.u32 %r235, [%rd4+40]; or.b32 %r236, %r235, 12; st.local.u32 [%rd4+40], %r236; bra.uni $L__BB193_182; $L__BB193_123: setp.leu.ftz.f32 %p79, %f189, 0f3727C5AC; @%p79 bra $L__BB193_182; ld.local.u32 %r29, [%rd4+252]; mov.u64 %rd693, 0; setp.eq.s32 %p80, %r29, 0; mov.u64 %rd691, %rd693; mov.u64 %rd692, %rd693; @%p80 bra $L__BB193_126; ld.local.u32 %r179, [%rd4+248]; mul.wide.s32 %rd293, %r179, 80; add.s64 %rd294, %rd1, %rd293; add.s64 %rd691, %rd294, 304; add.s64 %rd295, %rd230, %rd293; add.s64 %rd692, %rd295, 304; mov.b32 %r180, %f185; mov.b32 %r181, %f184; mov.b32 %r182, %f183; mov.u32 %r183, 0; st.local.v4.u32 [%rd294+304], {%r182, %r181, %r180, %r183}; add.s32 %r184, %r179, 1; st.local.u32 [%rd4+248], %r184; add.s32 %r185, %r29, -1; st.local.u32 [%rd4+252], %r185; $L__BB193_126: setp.eq.s64 %p81, %rd692, 0; @%p81 bra $L__BB193_128; add.ftz.f32 %f769, %f183, %f184; add.ftz.f32 %f770, %f185, %f769; mul.ftz.f32 %f771, %f770, 0f3EAAAAAB; abs.ftz.f32 %f772, %f771; st.local.f32 [%rd691+16], %f772; setp.ge.ftz.f32 %p82, %f772, 0f3727C5AC; selp.b64 %rd693, %rd692, 0, %p82; $L__BB193_128: setp.eq.s64 %p83, %rd693, 0; @%p83 bra $L__BB193_182; mul.ftz.f32 %f773, %f189, %f2993; add.s64 %rd14, %rd693, 32; st.f32 [%rd693+32], %f773; mul.ftz.f32 %f774, %f189, %f2994; st.f32 [%rd693+36], %f774; mul.ftz.f32 %f775, %f189, %f2995; st.f32 [%rd693+40], %f775; st.f32 [%rd693+44], %f180; st.f32 [%rd693+48], %f181; st.f32 [%rd693+52], %f182; st.f32 [%rd693+20], %f55; st.f32 [%rd693+24], %f56; st.f32 [%rd693+28], %f57; st.f32 [%rd693+56], %f71; mov.f32 %f3003, 0f00000000; max.ftz.f32 %f777, %f2997, %f3003; mov.f32 %f778, 0f3F666666; min.ftz.f32 %f779, %f777, %f778; st.f32 [%rd693+60], %f779; setp.eq.ftz.f32 %p84, %f71, 0f7F7FFFFF; mov.u32 %r643, 0; @%p84 bra $L__BB193_137; ld.f32 %f780, [%rd14+-32]; mov.u64 %rd696, 0; mov.f32 %f781, 0f00000000; max.ftz.f32 %f190, %f780, %f781; ld.f32 %f782, [%rd14+-28]; max.ftz.f32 %f191, %f782, %f781; ld.f32 %f783, [%rd14+-24]; max.ftz.f32 %f192, %f783, %f781; add.ftz.f32 %f784, %f190, %f191; add.ftz.f32 %f785, %f192, %f784; mul.ftz.f32 %f786, %f785, 0f3EAAAAAB; abs.ftz.f32 %f193, %f786; setp.ltu.ftz.f32 %p85, %f193, 0f3727C5AC; mov.u64 %rd697, %rd696; @%p85 bra $L__BB193_135; ld.local.u32 %r30, [%rd4+252]; mov.u64 %rd696, 0; setp.eq.s32 %p86, %r30, 0; mov.u64 %rd694, %rd696; mov.u64 %rd695, %rd696; @%p86 bra $L__BB193_133; ld.local.u32 %r187, [%rd4+248]; mul.wide.s32 %rd301, %r187, 80; add.s64 %rd302, %rd1, %rd301; add.s64 %rd694, %rd302, 304; add.s64 %rd303, %rd230, %rd301; add.s64 %rd695, %rd303, 304; mov.b32 %r188, %f192; mov.b32 %r189, %f191; mov.b32 %r190, %f190; mov.u32 %r191, 0; st.local.v4.u32 [%rd302+304], {%r190, %r189, %r188, %r191}; add.s32 %r192, %r187, 1; st.local.u32 [%rd4+248], %r192; add.s32 %r193, %r30, -1; st.local.u32 [%rd4+252], %r193; $L__BB193_133: setp.eq.s64 %p87, %rd695, 0; mov.u64 %rd697, %rd696; @%p87 bra $L__BB193_135; st.local.f32 [%rd694+16], %f193; mov.u64 %rd696, %rd694; mov.u64 %rd697, %rd695; $L__BB193_135: setp.eq.s64 %p88, %rd697, 0; @%p88 bra $L__BB193_137; ld.f32 %f787, [%rd14+-12]; ld.f32 %f788, [%rd14+-8]; ld.f32 %f789, [%rd14+-4]; st.local.f32 [%rd696+20], %f787; st.local.f32 [%rd696+24], %f788; st.local.f32 [%rd696+28], %f789; ld.f32 %f790, [%rd14+24]; st.local.f32 [%rd696+32], %f790; mov.u32 %r196, 5; st.local.u32 [%rd696+12], %r196; mov.u32 %r197, 8; st.local.u32 [%rd696+36], %r197; ld.local.f32 %f791, [%rd4+24]; ld.local.f32 %f792, [%rd4+28]; mul.ftz.f32 %f793, %f792, %f788; fma.rn.ftz.f32 %f794, %f791, %f787, %f793; ld.local.f32 %f795, [%rd4+32]; fma.rn.ftz.f32 %f796, %f795, %f789, %f794; mov.f32 %f797, 0f3F800000; sub.ftz.f32 %f798, %f797, %f796; max.ftz.f32 %f800, %f798, %f781; min.ftz.f32 %f801, %f800, %f797; mul.ftz.f32 %f802, %f801, %f801; mul.ftz.f32 %f803, %f802, %f802; mul.ftz.f32 %f804, %f801, %f803; mul.ftz.f32 %f805, %f790, %f804; ld.local.f32 %f806, [%rd696+16]; mul.ftz.f32 %f807, %f806, %f805; st.local.f32 [%rd696+16], %f807; mov.u32 %r643, 12; $L__BB193_137: ld.f32 %f809, [%rd14]; setp.geu.ftz.f32 %p89, %f809, 0f322BCC77; mov.u32 %r641, 3; mov.f32 %f3001, %f3003; @%p89 bra $L__BB193_139; ld.f32 %f3001, [%rd14+-32]; mov.u32 %r200, 0; st.u32 [%rd14+-32], %r200; st.u32 [%rd14], %r200; mov.u32 %r641, 2; $L__BB193_139: ld.f32 %f811, [%rd14+4]; setp.geu.ftz.f32 %p90, %f811, 0f322BCC77; mov.f32 %f3002, %f3003; @%p90 bra $L__BB193_141; ld.f32 %f3002, [%rd14+-28]; mov.u32 %r201, 0; st.u32 [%rd14+-28], %r201; st.u32 [%rd14+4], %r201; add.s32 %r641, %r641, -1; $L__BB193_141: ld.f32 %f813, [%rd14+8]; setp.geu.ftz.f32 %p91, %f813, 0f322BCC77; @%p91 bra $L__BB193_143; ld.f32 %f3003, [%rd14+-24]; mov.u32 %r202, 0; st.u32 [%rd14+-24], %r202; st.u32 [%rd14+8], %r202; add.s32 %r641, %r641, -1; $L__BB193_143: setp.gt.s32 %p92, %r641, 2; @%p92 bra $L__BB193_159; ld.f32 %f814, [%rd14+24]; setp.neu.ftz.f32 %p93, %f814, 0f7F7FFFFF; mov.f32 %f815, 0f00000000; max.ftz.f32 %f200, %f3002, %f815; max.ftz.f32 %f201, %f3001, %f815; add.ftz.f32 %f816, %f201, %f200; max.ftz.f32 %f202, %f3003, %f815; add.ftz.f32 %f817, %f202, %f816; mul.ftz.f32 %f818, %f817, 0f3EAAAAAB; abs.ftz.f32 %f203, %f818; @%p93 bra $L__BB193_152; bra.uni $L__BB193_145; $L__BB193_152: setp.ltu.ftz.f32 %p98, %f203, 0f3727C5AC; mov.u64 %rd704, 0; mov.u64 %rd705, %rd704; @%p98 bra $L__BB193_157; ld.local.u32 %r38, [%rd4+252]; mov.u64 %rd704, 0; setp.eq.s32 %p99, %r38, 0; mov.u64 %rd702, %rd704; mov.u64 %rd703, %rd704; @%p99 bra $L__BB193_155; ld.local.u32 %r212, [%rd4+248]; mul.wide.s32 %rd319, %r212, 80; add.s64 %rd320, %rd1, %rd319; add.s64 %rd702, %rd320, 304; add.s64 %rd321, %rd230, %rd319; add.s64 %rd703, %rd321, 304; mov.b32 %r213, %f202; mov.b32 %r214, %f200; mov.b32 %r215, %f201; mov.u32 %r216, 0; st.local.v4.u32 [%rd320+304], {%r215, %r214, %r213, %r216}; add.s32 %r217, %r212, 1; st.local.u32 [%rd4+248], %r217; add.s32 %r218, %r38, -1; st.local.u32 [%rd4+252], %r218; $L__BB193_155: setp.eq.s64 %p100, %rd703, 0; mov.u64 %rd705, %rd704; @%p100 bra $L__BB193_157; st.local.f32 [%rd702+16], %f203; mov.u64 %rd704, %rd702; mov.u64 %rd705, %rd703; $L__BB193_157: setp.eq.s64 %p101, %rd705, 0; @%p101 bra $L__BB193_159; ld.f32 %f822, [%rd14+-12]; ld.f32 %f823, [%rd14+-8]; ld.f32 %f824, [%rd14+-4]; st.local.f32 [%rd704+20], %f822; st.local.f32 [%rd704+24], %f823; st.local.f32 [%rd704+28], %f824; ld.f32 %f825, [%rd14+24]; st.local.f32 [%rd704+32], %f825; mov.u32 %r220, 5; st.local.u32 [%rd704+12], %r220; mov.u32 %r221, 2; st.local.u32 [%rd704+36], %r221; mov.u32 %r643, 12; bra.uni $L__BB193_159; $L__BB193_373: and.b32 %r419, %r110, 7; setp.eq.s32 %p330, %r419, 0; mov.f32 %f3038, 0f00000000; @%p330 bra $L__BB193_385; setp.eq.s32 %p331, %r110, 4; selp.b32 %r421, %r404, 0, %p331; add.s32 %r422, %r113, %r421; ld.const.u64 %rd513, [kernel_params+2320]; cvta.to.global.u64 %rd514, %rd513; mul.wide.s32 %rd515, %r422, 4; add.s64 %rd516, %rd514, %rd515; ld.global.f32 %f3038, [%rd516]; bra.uni $L__BB193_385; $L__BB193_498: ld.local.u32 %r583, [%rd4+248]; add.s32 %r584, %r583, -1; st.local.u32 [%rd4+248], %r584; add.s32 %r585, %r131, 1; st.local.u32 [%rd4+252], %r585; mov.u64 %rd795, %rd794; $L__BB193_499: st.local.u64 [%rd205+28], %rd795; setp.eq.s64 %p441, %rd795, 0; @%p441 bra $L__BB193_517; ld.local.f32 %f2886, [%rd7]; ld.local.f32 %f2887, [%rd7+4]; ld.local.f32 %f2888, [%rd7+8]; st.local.f32 [%rd794], %f2886; st.local.f32 [%rd794+4], %f2887; st.local.f32 [%rd794+8], %f2888; ld.local.u64 %rd659, [%rd205+28]; mov.u32 %r586, 0; st.u32 [%rd659+12], %r586; st.u32 [%rd659+16], %r586; st.u32 [%rd659+20], %r586; ld.local.u64 %rd660, [%rd205+28]; st.u32 [%rd660+36], %r586; ld.local.v2.f32 {%f2889, %f2890}, [%rd205+36]; setp.neu.ftz.f32 %p442, %f2889, 0f00000000; setp.neu.ftz.f32 %p443, %f2890, 0f00000000; or.pred %p444, %p442, %p443; ld.local.f32 %f2893, [%rd205+44]; setp.neu.ftz.f32 %p445, %f2893, 0f00000000; or.pred %p446, %p445, %p444; @%p446 bra $L__BB193_502; mov.f32 %f2895, 0f3F800000; st.local.v2.f32 [%rd205+36], {%f2895, %f2837}; st.local.u32 [%rd205+44], %r586; $L__BB193_502: mov.u32 %r588, 14; st.local.u32 [%rd205+-8], %r588; ld.local.v2.f32 {%f2896, %f2897}, [%rd205+12]; mov.f32 %f2900, 0f38D1B717; max.ftz.f32 %f2901, %f2896, %f2900; mov.f32 %f2902, 0f3F800000; max.ftz.f32 %f2903, %f2897, %f2900; min.ftz.f32 %f2904, %f2903, %f2902; min.ftz.f32 %f2905, %f2901, %f2902; st.local.v2.f32 [%rd205+12], {%f2905, %f2904}; ld.local.u64 %rd661, [%rd205+28]; ld.f32 %f2906, [%rd661]; cvt.ftz.sat.f32.f32 %f2907, %f2906; ld.f32 %f2908, [%rd661+4]; cvt.ftz.sat.f32.f32 %f2909, %f2908; ld.f32 %f2910, [%rd661+8]; cvt.ftz.sat.f32.f32 %f2911, %f2910; st.f32 [%rd661], %f2907; st.f32 [%rd661+4], %f2909; st.f32 [%rd661+8], %f2911; ld.local.u64 %rd662, [%rd205+28]; ld.f32 %f2912, [%rd662+12]; cvt.ftz.sat.f32.f32 %f2913, %f2912; ld.f32 %f2914, [%rd662+16]; cvt.ftz.sat.f32.f32 %f2915, %f2914; ld.f32 %f2916, [%rd662+20]; cvt.ftz.sat.f32.f32 %f2917, %f2916; st.f32 [%rd662+12], %f2913; st.f32 [%rd662+16], %f2915; st.f32 [%rd662+20], %f2917; ld.local.u32 %r589, [%rd4+40]; or.b32 %r590, %r589, 1036; st.local.u32 [%rd4+40], %r590; bra.uni $L__BB193_517; $L__BB193_371: mad.lo.s32 %r418, %r404, 3, %r113; ld.const.u64 %rd510, [kernel_params+2320]; cvta.to.global.u64 %rd511, %rd510; mul.wide.s32 %rd512, %r418, 4; add.s64 %rd755, %rd511, %rd512; add.s64 %rd756, %rd755, 4; add.s64 %rd757, %rd755, 8; $L__BB193_372: ld.local.f32 %f2509, [%rd4+56]; ld.global.f32 %f2510, [%rd756]; ld.local.f32 %f2511, [%rd4+60]; ld.global.f32 %f2512, [%rd757]; mul.ftz.f32 %f2513, %f2512, %f2511; fma.rn.ftz.f32 %f2514, %f2510, %f2509, %f2513; mov.f32 %f2515, 0f3F800000; sub.ftz.f32 %f2516, %f2515, %f2509; sub.ftz.f32 %f2517, %f2516, %f2511; ld.global.f32 %f2518, [%rd755]; fma.rn.ftz.f32 %f3038, %f2518, %f2517, %f2514; $L__BB193_385: mov.f32 %f2528, 0f00000000; max.ftz.f32 %f399, %f385, %f2528; max.ftz.f32 %f400, %f386, %f2528; add.ftz.f32 %f2529, %f400, %f399; max.ftz.f32 %f401, %f384, %f2528; add.ftz.f32 %f2530, %f401, %f2529; mul.ftz.f32 %f2531, %f2530, 0f3EAAAAAB; abs.ftz.f32 %f402, %f2531; setp.ltu.ftz.f32 %p340, %f402, 0f3727C5AC; mov.u64 %rd760, 0; mov.u64 %rd761, %rd760; @%p340 bra $L__BB193_390; ld.local.u32 %r118, [%rd4+252]; mov.u64 %rd760, 0; setp.eq.s32 %p341, %r118, 0; mov.u64 %rd758, %rd760; mov.u64 %rd759, %rd760; @%p341 bra $L__BB193_388; ld.local.u32 %r435, [%rd4+248]; mul.wide.s32 %rd539, %r435, 80; add.s64 %rd540, %rd1, %rd539; add.s64 %rd758, %rd540, 304; add.s64 %rd541, %rd230, %rd539; add.s64 %rd759, %rd541, 304; mov.b32 %r436, %f401; mov.b32 %r437, %f399; mov.b32 %r438, %f400; mov.u32 %r439, 0; st.local.v4.u32 [%rd540+304], {%r438, %r437, %r436, %r439}; add.s32 %r440, %r435, 1; st.local.u32 [%rd4+248], %r440; add.s32 %r441, %r118, -1; st.local.u32 [%rd4+252], %r441; $L__BB193_388: setp.eq.s64 %p342, %rd759, 0; mov.u64 %rd761, %rd760; @%p342 bra $L__BB193_390; st.local.f32 [%rd758+16], %f402; mov.u64 %rd760, %rd758; mov.u64 %rd761, %rd759; $L__BB193_390: add.s32 %r663, %r142, 4; setp.eq.s64 %p343, %rd761, 0; @%p343 bra $L__BB193_517; ld.local.u32 %r120, [%rd4+252]; setp.lt.s32 %p344, %r120, 1; @%p344 bra $L__BB193_393; add.s32 %r442, %r120, -1; st.local.u32 [%rd4+252], %r442; ld.local.s32 %rd544, [%rd4+248]; cvt.s64.s32 %rd545, %r442; add.s64 %rd546, %rd544, %rd545; mul.lo.s64 %rd547, %rd546, 80; add.s64 %rd548, %rd230, %rd547; add.s64 %rd763, %rd548, 304; add.s64 %rd549, %rd1, %rd547; add.s64 %rd762, %rd549, 304; bra.uni $L__BB193_394; $L__BB193_393: ld.local.u32 %r443, [%rd4+248]; mov.u64 %rd762, 0; add.s32 %r444, %r443, -1; st.local.u32 [%rd4+248], %r444; add.s32 %r445, %r120, 1; st.local.u32 [%rd4+252], %r445; mov.u64 %rd763, %rd762; $L__BB193_394: setp.eq.s64 %p345, %rd763, 0; @%p345 bra $L__BB193_517; setp.eq.s32 %p346, %r91, 255; @%p346 bra $L__BB193_397; bra.uni $L__BB193_396; $L__BB193_397: mov.b32 %f3039, %r382; bra.uni $L__BB193_398; $L__BB193_396: mul.wide.u32 %rd552, %r91, 4; add.s64 %rd553, %rd2, %rd552; ld.local.f32 %f3039, [%rd553]; $L__BB193_398: add.ftz.f32 %f2532, %f3038, 0fBF000000; add.ftz.f32 %f406, %f2532, %f2532; fma.rn.ftz.f32 %f2533, %f406, %f3039, 0f3F800000; mul.ftz.f32 %f407, %f60, %f2533; mul.ftz.f32 %f408, %f63, %f2533; setp.eq.s32 %p347, %r86, 255; @%p347 bra $L__BB193_400; bra.uni $L__BB193_399; $L__BB193_400: mov.b32 %f3040, %r376; bra.uni $L__BB193_401; $L__BB193_399: mul.wide.u32 %rd554, %r86, 4; add.s64 %rd555, %rd2, %rd554; ld.local.f32 %f3040, [%rd555]; $L__BB193_401: max.ftz.f32 %f2535, %f3040, %f2528; mov.f32 %f2536, 0f3F800000; min.ftz.f32 %f2537, %f2535, %f2536; sub.ftz.f32 %f412, %f2536, %f2537; add.s64 %rd153, %rd760, 20; st.local.f32 [%rd760+20], %f55; st.local.f32 [%rd760+24], %f56; st.local.f32 [%rd760+28], %f57; st.local.f32 [%rd760+44], %f407; st.local.v2.f32 [%rd760+48], {%f408, %f3036}; st.local.v2.f32 [%rd760+56], {%f3037, %f412}; st.local.u64 [%rd760+64], %rd763; shr.u32 %r446, %r144, 24; cvt.u16.u32 %rs15, %r446; setp.eq.s16 %p348, %rs15, 0; @%p348 bra $L__BB193_415; setp.eq.s16 %p349, %rs15, 1; @%p349 bra $L__BB193_405; setp.ne.s16 %p350, %rs15, 2; @%p350 bra $L__BB193_416; shr.u32 %r447, %r375, 24; mul.wide.u32 %rd556, %r447, 4; add.s64 %rd557, %rd2, %rd556; ld.local.f32 %f2538, [%rd557+8]; ld.local.f32 %f2539, [%rd557+4]; ld.local.f32 %f2540, [%rd557]; st.local.v2.f32 [%rd153+12], {%f2540, %f2539}; st.local.f32 [%rd153+20], %f2538; bra.uni $L__BB193_417; $L__BB193_415: shr.u32 %r449, %r144, 14; cvt.u64.u32 %rd567, %r449; and.b64 %rd568, %rd567, 1020; add.s64 %rd569, %rd2, %rd568; ld.local.f32 %f2581, [%rd569]; lg2.approx.ftz.f32 %f2582, %f2581; mul.ftz.f32 %f2583, %f2582, 0f3F317218; ld.local.f32 %f2584, [%rd569+4]; lg2.approx.ftz.f32 %f2585, %f2584; mul.ftz.f32 %f2586, %f2585, 0f3F317218; ld.local.f32 %f2587, [%rd569+8]; lg2.approx.ftz.f32 %f2588, %f2587; mul.ftz.f32 %f2589, %f2588, 0f3F317218; fma.rn.ftz.f32 %f2590, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2591, %f408, %f2590, 0fC12BAE14; fma.rn.ftz.f32 %f2592, %f408, %f2591, 0f40220C4A; fma.rn.ftz.f32 %f2593, %f408, %f2592, 0fBE5C28F6; fma.rn.ftz.f32 %f2594, %f408, %f2593, 0f40BF020C; div.approx.ftz.f32 %f2595, %f2583, %f2594; div.approx.ftz.f32 %f2596, %f2586, %f2594; div.approx.ftz.f32 %f2597, %f2589, %f2594; mul.ftz.f32 %f2598, %f2597, %f2597; mul.ftz.f32 %f2599, %f2596, %f2596; mul.ftz.f32 %f2600, %f2595, %f2595; st.local.v2.f32 [%rd153+12], {%f2600, %f2599}; st.local.f32 [%rd153+20], %f2598; bra.uni $L__BB193_417; $L__BB193_405: setp.eq.s32 %p351, %r87, 255; @%p351 bra $L__BB193_407; bra.uni $L__BB193_406; $L__BB193_407: mov.b32 %f3041, %r377; bra.uni $L__BB193_408; $L__BB193_416: mov.f32 %f2601, 0f3F172384; mov.f32 %f2602, 0f3E8D729D; st.local.v2.f32 [%rd153+12], {%f2602, %f2601}; mov.u32 %r450, 1069964113; st.local.u32 [%rd153+20], %r450; bra.uni $L__BB193_417; $L__BB193_406: mul.wide.u32 %rd558, %r87, 4; add.s64 %rd559, %rd2, %rd558; ld.local.f32 %f3041, [%rd559]; $L__BB193_408: setp.eq.s32 %p352, %r88, 255; @%p352 bra $L__BB193_410; bra.uni $L__BB193_409; $L__BB193_410: mov.b32 %f3042, %r378; bra.uni $L__BB193_411; $L__BB193_409: mul.wide.u32 %rd560, %r88, 4; add.s64 %rd561, %rd2, %rd560; ld.local.f32 %f3042, [%rd561]; $L__BB193_411: setp.eq.s32 %p353, %r90, 255; @%p353 bra $L__BB193_413; bra.uni $L__BB193_412; $L__BB193_413: mov.b32 %f3043, %r381; bra.uni $L__BB193_414; $L__BB193_412: mul.wide.u32 %rd562, %r90, 4; add.s64 %rd563, %rd2, %rd562; ld.local.f32 %f3043, [%rd563]; $L__BB193_414: max.ftz.f32 %f2542, %f3043, %f2528; min.ftz.f32 %f2544, %f2542, %f2536; fma.rn.ftz.f32 %f2545, %f406, %f2544, 0f3F800000; mul.ftz.f32 %f2546, %f3041, %f2545; sub.ftz.f32 %f2547, %f2536, %f2546; mov.f32 %f2548, 0f38D1B717; max.ftz.f32 %f2549, %f2547, %f2548; lg2.approx.ftz.f32 %f2550, %f2549; mul.ftz.f32 %f2551, %f2550, 0fBF317218; sub.ftz.f32 %f2552, %f2536, %f3042; mul.ftz.f32 %f2553, %f2552, %f2551; mul.ftz.f32 %f2554, %f3042, %f2551; mul.ftz.f32 %f2555, %f2554, 0f3EAF9DB2; mul.ftz.f32 %f2556, %f2554, 0f3F3BA5E3; mul.ftz.f32 %f2557, %f2554, 0f3FF645A2; fma.rn.ftz.f32 %f2558, %f2553, 0f3F018937, %f2555; fma.rn.ftz.f32 %f2559, %f2553, 0f3F574BC7, %f2556; fma.rn.ftz.f32 %f2560, %f2553, 0f3FD39581, %f2557; shl.b32 %r448, %r379, 2; cvt.u64.u32 %rd564, %r448; and.b64 %rd565, %rd564, 1020; add.s64 %rd566, %rd2, %rd565; ld.local.f32 %f2561, [%rd566]; lg2.approx.ftz.f32 %f2562, %f2561; mul.ftz.f32 %f2563, %f2562, 0f3F317218; ld.local.f32 %f2564, [%rd566+4]; lg2.approx.ftz.f32 %f2565, %f2564; mul.ftz.f32 %f2566, %f2565, 0f3F317218; ld.local.f32 %f2567, [%rd566+8]; lg2.approx.ftz.f32 %f2568, %f2567; mul.ftz.f32 %f2569, %f2568, 0f3F317218; fma.rn.ftz.f32 %f2570, %f408, 0f3E7AE148, 0f40B25E35; fma.rn.ftz.f32 %f2571, %f408, %f2570, 0fC12BAE14; fma.rn.ftz.f32 %f2572, %f408, %f2571, 0f40220C4A; fma.rn.ftz.f32 %f2573, %f408, %f2572, 0fBE5C28F6; fma.rn.ftz.f32 %f2574, %f408, %f2573, 0f40BF020C; div.approx.ftz.f32 %f2575, %f2563, %f2574; div.approx.ftz.f32 %f2576, %f2566, %f2574; div.approx.ftz.f32 %f2577, %f2569, %f2574; fma.rn.ftz.f32 %f2578, %f2577, %f2577, %f2560; fma.rn.ftz.f32 %f2579, %f2576, %f2576, %f2559; fma.rn.ftz.f32 %f2580, %f2575, %f2575, %f2558; st.local.v2.f32 [%rd153+12], {%f2580, %f2579}; st.local.f32 [%rd153+20], %f2578; $L__BB193_417: mov.u32 %r451, 29; st.local.u32 [%rd153+-8], %r451; mov.f32 %f2603, 0f3A83126F; max.ftz.f32 %f2604, %f407, %f2603; min.ftz.f32 %f2606, %f2604, %f2536; max.ftz.f32 %f2607, %f408, %f2603; min.ftz.f32 %f2608, %f2607, %f2536; mul.ftz.f32 %f2609, %f2606, %f412; max.ftz.f32 %f2610, %f2609, %f2603; min.ftz.f32 %f2611, %f2610, %f2536; mul.ftz.f32 %f2612, %f2606, %f2606; mul.ftz.f32 %f2613, %f2612, 0f3F4FDF3B; fma.rn.ftz.f32 %f2614, %f2606, 0f3F39DB23, %f2613; mul.ftz.f32 %f2615, %f2612, %f2612; mul.ftz.f32 %f2616, %f2606, %f2615; mul.ftz.f32 %f2617, %f2616, %f2616; mul.ftz.f32 %f2618, %f2617, %f2617; fma.rn.ftz.f32 %f2619, %f2618, 0f406CCCCD, %f2614; mul.ftz.f32 %f2620, %f2619, %f2619; st.local.f32 [%rd153+24], %f2620; mul.ftz.f32 %f2621, %f2608, %f2608; mul.ftz.f32 %f2622, %f2621, 0f3F98D4FE; fma.rn.ftz.f32 %f2623, %f2608, 0f3E87AE14, %f2622; mul.ftz.f32 %f2624, %f2621, %f2621; mul.ftz.f32 %f2625, %f2608, %f2624; mul.ftz.f32 %f2626, %f2625, %f2625; mul.ftz.f32 %f2627, %f2608, %f2626; mul.ftz.f32 %f2628, %f2627, %f2627; fma.rn.ftz.f32 %f2629, %f2628, 0f40ABE76D, %f2623; mul.ftz.f32 %f2630, %f2629, 0f3F206C99; st.local.f32 [%rd153+28], %f2630; mul.ftz.f32 %f2631, %f2611, %f2611; mul.ftz.f32 %f2632, %f2631, 0f3F4FDF3B; fma.rn.ftz.f32 %f2633, %f2611, 0f3F39DB23, %f2632; mul.ftz.f32 %f2634, %f2631, %f2631; mul.ftz.f32 %f2635, %f2611, %f2634; mul.ftz.f32 %f2636, %f2635, %f2635; mul.ftz.f32 %f2637, %f2636, %f2636; fma.rn.ftz.f32 %f2638, %f2637, 0f406CCCCD, %f2633; mul.ftz.f32 %f2639, %f2638, %f2638; st.local.f32 [%rd153+40], %f2639; ld.local.f32 %f3044, [%rd4+104]; ld.local.f32 %f3045, [%rd4+108]; mul.ftz.f32 %f2640, %f3045, %f3045; fma.rn.ftz.f32 %f2641, %f3044, %f3044, %f2640; ld.local.f32 %f3046, [%rd4+112]; fma.rn.ftz.f32 %f2642, %f3046, %f3046, %f2641; sqrt.approx.ftz.f32 %f425, %f2642; setp.neu.ftz.f32 %p354, %f425, 0f00000000; @%p354 bra $L__BB193_418; bra.uni $L__BB193_419; $L__BB193_418: rcp.approx.ftz.f32 %f2643, %f425; mul.ftz.f32 %f3046, %f3046, %f2643; mul.ftz.f32 %f3045, %f3045, %f2643; mul.ftz.f32 %f3044, %f3044, %f2643; $L__BB193_419: ld.local.f32 %f2644, [%rd4+32]; mul.ftz.f32 %f2645, %f3045, %f2644; ld.local.f32 %f2646, [%rd4+28]; mul.ftz.f32 %f2647, %f3046, %f2646; sub.ftz.f32 %f3047, %f2645, %f2647; ld.local.f32 %f2648, [%rd4+24]; mul.ftz.f32 %f2649, %f3046, %f2648; mul.ftz.f32 %f2650, %f3044, %f2644; sub.ftz.f32 %f3048, %f2649, %f2650; mul.ftz.f32 %f2651, %f3044, %f2646; mul.ftz.f32 %f2652, %f3045, %f2648; sub.ftz.f32 %f3049, %f2651, %f2652; mul.ftz.f32 %f2653, %f3048, %f3048; fma.rn.ftz.f32 %f2654, %f3047, %f3047, %f2653; fma.rn.ftz.f32 %f2655, %f3049, %f3049, %f2654; sqrt.approx.ftz.f32 %f441, %f2655; setp.neu.ftz.f32 %p355, %f441, 0f00000000; @%p355 bra $L__BB193_420; bra.uni $L__BB193_421; $L__BB193_420: rcp.approx.ftz.f32 %f2656, %f441; mul.ftz.f32 %f3049, %f3049, %f2656; mul.ftz.f32 %f3048, %f3048, %f2656; mul.ftz.f32 %f3047, %f3047, %f2656; $L__BB193_421: mul.ftz.f32 %f2657, %f3045, %f3049; mul.ftz.f32 %f2658, %f3046, %f3048; sub.ftz.f32 %f3050, %f2657, %f2658; mul.ftz.f32 %f2659, %f3046, %f3047; mul.ftz.f32 %f2660, %f3044, %f3049; sub.ftz.f32 %f3051, %f2659, %f2660; mul.ftz.f32 %f2661, %f3044, %f3048; mul.ftz.f32 %f2662, %f3045, %f3047; sub.ftz.f32 %f3052, %f2661, %f2662; mul.ftz.f32 %f2663, %f3051, %f3051; fma.rn.ftz.f32 %f2664, %f3050, %f3050, %f2663; fma.rn.ftz.f32 %f2665, %f3052, %f3052, %f2664; sqrt.approx.ftz.f32 %f457, %f2665; setp.neu.ftz.f32 %p356, %f457, 0f00000000; @%p356 bra $L__BB193_422; bra.uni $L__BB193_423; $L__BB193_422: rcp.approx.ftz.f32 %f2666, %f457; mul.ftz.f32 %f3052, %f2666, %f3052; mul.ftz.f32 %f3051, %f2666, %f3051; mul.ftz.f32 %f3050, %f2666, %f3050; $L__BB193_423: ld.local.u8 %rs16, [%rd4+52]; and.b16 %rs17, %rs16, 4; setp.eq.s16 %p357, %rs17, 0; @%p357 bra $L__BB193_425; bra.uni $L__BB193_424; $L__BB193_425: ld.local.f32 %f2668, [%rd4+16]; mul.ftz.f32 %f2669, %f3046, %f2668; ld.local.f32 %f2670, [%rd4+20]; mul.ftz.f32 %f2671, %f3045, %f2670; sub.ftz.f32 %f2672, %f2669, %f2671; mul.ftz.f32 %f2673, %f3044, %f2670; ld.local.f32 %f2674, [%rd4+12]; mul.ftz.f32 %f2675, %f3046, %f2674; sub.ftz.f32 %f2676, %f2673, %f2675; mul.ftz.f32 %f2677, %f3045, %f2674; mul.ftz.f32 %f2678, %f3044, %f2668; sub.ftz.f32 %f2679, %f2677, %f2678; mul.ftz.f32 %f2680, %f3051, %f2676; fma.rn.ftz.f32 %f2681, %f3050, %f2672, %f2680; fma.rn.ftz.f32 %f3053, %f3052, %f2679, %f2681; bra.uni $L__BB193_426; $L__BB193_424: ld.local.f32 %f2667, [%rd4+60]; neg.ftz.f32 %f3053, %f2667; $L__BB193_426: st.local.v4.f32 [%rd762], {%f3047, %f3048, %f3049, %f3053}; ld.local.u32 %r452, [%rd4+40]; or.b32 %r453, %r452, 1036; st.local.u32 [%rd4+40], %r453; bra.uni $L__BB193_517; $L__BB193_145: setp.ltu.ftz.f32 %p94, %f203, 0f3727C5AC; mov.u64 %rd700, 0; mov.u64 %rd701, %rd700; @%p94 bra $L__BB193_150; ld.local.u32 %r37, [%rd4+252]; mov.u64 %rd700, 0; setp.eq.s32 %p95, %r37, 0; mov.u64 %rd698, %rd700; mov.u64 %rd699, %rd700; @%p95 bra $L__BB193_148; ld.local.u32 %r203, [%rd4+248]; mul.wide.s32 %rd310, %r203, 80; add.s64 %rd311, %rd1, %rd310; add.s64 %rd698, %rd311, 304; add.s64 %rd312, %rd230, %rd310; add.s64 %rd699, %rd312, 304; mov.b32 %r204, %f202; mov.b32 %r205, %f200; mov.b32 %r206, %f201; mov.u32 %r207, 0; st.local.v4.u32 [%rd311+304], {%r206, %r205, %r204, %r207}; add.s32 %r208, %r203, 1; st.local.u32 [%rd4+248], %r208; add.s32 %r209, %r37, -1; st.local.u32 [%rd4+252], %r209; $L__BB193_148: setp.eq.s64 %p96, %rd699, 0; mov.u64 %rd701, %rd700; @%p96 bra $L__BB193_150; st.local.f32 [%rd698+16], %f203; mov.u64 %rd700, %rd698; mov.u64 %rd701, %rd699; $L__BB193_150: setp.eq.s64 %p97, %rd701, 0; @%p97 bra $L__BB193_159; ld.f32 %f819, [%rd14+-12]; ld.f32 %f820, [%rd14+-8]; ld.f32 %f821, [%rd14+-4]; st.local.f32 [%rd700+20], %f819; st.local.f32 [%rd700+24], %f820; st.local.f32 [%rd700+28], %f821; mov.u32 %r211, 2; st.local.u32 [%rd700+12], %r211; mov.u32 %r643, 12; $L__BB193_159: st.u32 [%rd14+-20], %r149; setp.gt.s32 %p102, %r641, 0; @%p102 bra $L__BB193_161; bra.uni $L__BB193_160; $L__BB193_161: ld.f32 %f826, [%rd14+-28]; ld.f32 %f827, [%rd14+-32]; add.ftz.f32 %f828, %f827, %f826; ld.f32 %f829, [%rd14+-24]; add.ftz.f32 %f830, %f829, %f828; mul.ftz.f32 %f831, %f830, 0f3EAAAAAB; abs.ftz.f32 %f832, %f831; cvt.rn.f32.s32 %f833, %r641; mul.ftz.f32 %f834, %f832, %f833; st.f32 [%rd14+-16], %f834; or.b32 %r223, %r149, 2; setp.eq.s32 %p103, %r223, 34; @%p103 bra $L__BB193_172; bra.uni $L__BB193_162; $L__BB193_172: ld.f32 %f1388, [%rd14]; mul.ftz.f32 %f3007, %f1388, 0f3DA2F983; ld.f32 %f1389, [%rd14+4]; mul.ftz.f32 %f3008, %f1389, 0f3DA2F983; ld.f32 %f1390, [%rd14+8]; mul.ftz.f32 %f3009, %f1390, 0f3DA2F983; bra.uni $L__BB193_173; $L__BB193_160: mov.u32 %r222, 0; st.u32 [%rd14+-16], %r222; bra.uni $L__BB193_174; $L__BB193_162: mov.f32 %f836, 0f3F8147AE; max.ftz.f32 %f837, %f2996, %f836; mov.f32 %f838, 0f40733333; min.ftz.f32 %f839, %f837, %f838; rcp.approx.ftz.f32 %f840, %f839; mov.f32 %f841, 0f3F800000; fma.rn.ftz.f32 %f842, %f840, 0fBFB851EC, 0f3F35C28F; fma.rn.ftz.f32 %f843, %f840, %f842, 0f3F2B020C; fma.rn.ftz.f32 %f844, %f839, 0f3D8240B8, %f843; add.ftz.f32 %f845, %f844, 0f3F800000; mul.ftz.f32 %f846, %f845, 0f3FAAAAAB; sub.ftz.f32 %f847, %f841, %f844; div.approx.ftz.f32 %f204, %f846, %f847; ld.f32 %f205, [%rd14+12]; setp.lt.ftz.f32 %p104, %f205, 0f38D1B717; mov.f32 %f3004, 0f00000000; @%p104 bra $L__BB193_165; setp.ge.ftz.f32 %p105, %f205, 0f3F7EB852; mov.f32 %f3004, 0f3F7FFFEF; @%p105 bra $L__BB193_165; sub.ftz.f32 %f850, %f841, 0f3F000000; mul.ftz.f32 %f851, %f850, 0f40400000; sqrt.approx.ftz.f32 %f852, %f851; mul.ftz.f32 %f853, %f204, %f852; mul.ftz.f32 %f854, %f853, 0fBFB8AA3B; ex2.approx.ftz.f32 %f855, %f854; add.ftz.f32 %f856, %f855, 0f3F800000; mul.ftz.f32 %f857, %f856, 0f3E800000; mul.ftz.f32 %f858, %f852, 0fBFB8AA3B; ex2.approx.ftz.f32 %f859, %f858; mul.ftz.f32 %f860, %f859, %f857; setp.lt.ftz.f32 %p106, %f860, %f205; selp.f32 %f861, 0f3F800000, 0f3F000000, %p106; selp.f32 %f862, 0f3F000000, 0f00000000, %p106; selp.f32 %f863, 0f3F400000, 0f3E800000, %p106; sub.ftz.f32 %f864, %f841, %f863; mul.ftz.f32 %f865, %f864, 0f40400000; sqrt.approx.ftz.f32 %f866, %f865; mul.ftz.f32 %f867, %f863, 0f3F000000; mul.ftz.f32 %f868, %f204, %f866; mul.ftz.f32 %f869, %f868, 0fBFB8AA3B; ex2.approx.ftz.f32 %f870, %f869; add.ftz.f32 %f871, %f870, 0f3F800000; mul.ftz.f32 %f872, %f871, %f867; mul.ftz.f32 %f873, %f866, 0fBFB8AA3B; ex2.approx.ftz.f32 %f874, %f873; mul.ftz.f32 %f875, %f874, %f872; setp.lt.ftz.f32 %p107, %f875, %f205; selp.f32 %f876, %f861, %f863, %p107; selp.f32 %f877, %f863, %f862, %p107; add.ftz.f32 %f878, %f876, %f877; mul.ftz.f32 %f879, %f878, 0f3F000000; sub.ftz.f32 %f880, %f841, %f879; mul.ftz.f32 %f881, %f880, 0f40400000; sqrt.approx.ftz.f32 %f882, %f881; mul.ftz.f32 %f883, %f879, 0f3F000000; mul.ftz.f32 %f884, %f204, %f882; mul.ftz.f32 %f885, %f884, 0fBFB8AA3B; ex2.approx.ftz.f32 %f886, %f885; add.ftz.f32 %f887, %f886, 0f3F800000; mul.ftz.f32 %f888, %f887, %f883; mul.ftz.f32 %f889, %f882, 0fBFB8AA3B; ex2.approx.ftz.f32 %f890, %f889; mul.ftz.f32 %f891, %f890, %f888; setp.lt.ftz.f32 %p108, %f891, %f205; selp.f32 %f892, %f876, %f879, %p108; selp.f32 %f893, %f879, %f877, %p108; add.ftz.f32 %f894, %f892, %f893; mul.ftz.f32 %f895, %f894, 0f3F000000; sub.ftz.f32 %f896, %f841, %f895; mul.ftz.f32 %f897, %f896, 0f40400000; sqrt.approx.ftz.f32 %f898, %f897; mul.ftz.f32 %f899, %f895, 0f3F000000; mul.ftz.f32 %f900, %f204, %f898; mul.ftz.f32 %f901, %f900, 0fBFB8AA3B; ex2.approx.ftz.f32 %f902, %f901; add.ftz.f32 %f903, %f902, 0f3F800000; mul.ftz.f32 %f904, %f903, %f899; mul.ftz.f32 %f905, %f898, 0fBFB8AA3B; ex2.approx.ftz.f32 %f906, %f905; mul.ftz.f32 %f907, %f906, %f904; setp.lt.ftz.f32 %p109, %f907, %f205; selp.f32 %f908, %f892, %f895, %p109; selp.f32 %f909, %f895, %f893, %p109; add.ftz.f32 %f910, %f908, %f909; mul.ftz.f32 %f911, %f910, 0f3F000000; sub.ftz.f32 %f912, %f841, %f911; mul.ftz.f32 %f913, %f912, 0f40400000; sqrt.approx.ftz.f32 %f914, %f913; mul.ftz.f32 %f915, %f911, 0f3F000000; mul.ftz.f32 %f916, %f204, %f914; mul.ftz.f32 %f917, %f916, 0fBFB8AA3B; ex2.approx.ftz.f32 %f918, %f917; add.ftz.f32 %f919, %f918, 0f3F800000; mul.ftz.f32 %f920, %f919, %f915; mul.ftz.f32 %f921, %f914, 0fBFB8AA3B; ex2.approx.ftz.f32 %f922, %f921; mul.ftz.f32 %f923, %f922, %f920; setp.lt.ftz.f32 %p110, %f923, %f205; selp.f32 %f924, %f908, %f911, %p110; selp.f32 %f925, %f911, %f909, %p110; add.ftz.f32 %f926, %f924, %f925; mul.ftz.f32 %f927, %f926, 0f3F000000; sub.ftz.f32 %f928, %f841, %f927; mul.ftz.f32 %f929, %f928, 0f40400000; sqrt.approx.ftz.f32 %f930, %f929; mul.ftz.f32 %f931, %f927, 0f3F000000; mul.ftz.f32 %f932, %f204, %f930; mul.ftz.f32 %f933, %f932, 0fBFB8AA3B; ex2.approx.ftz.f32 %f934, %f933; add.ftz.f32 %f935, %f934, 0f3F800000; mul.ftz.f32 %f936, %f935, %f931; mul.ftz.f32 %f937, %f930, 0fBFB8AA3B; ex2.approx.ftz.f32 %f938, %f937; mul.ftz.f32 %f939, %f938, %f936; setp.lt.ftz.f32 %p111, %f939, %f205; selp.f32 %f940, %f924, %f927, %p111; selp.f32 %f941, %f927, %f925, %p111; add.ftz.f32 %f942, %f940, %f941; mul.ftz.f32 %f943, %f942, 0f3F000000; sub.ftz.f32 %f944, %f841, %f943; mul.ftz.f32 %f945, %f944, 0f40400000; sqrt.approx.ftz.f32 %f946, %f945; mul.ftz.f32 %f947, %f943, 0f3F000000; mul.ftz.f32 %f948, %f204, %f946; mul.ftz.f32 %f949, %f948, 0fBFB8AA3B; ex2.approx.ftz.f32 %f950, %f949; add.ftz.f32 %f951, %f950, 0f3F800000; mul.ftz.f32 %f952, %f951, %f947; mul.ftz.f32 %f953, %f946, 0fBFB8AA3B; ex2.approx.ftz.f32 %f954, %f953; mul.ftz.f32 %f955, %f954, %f952; setp.lt.ftz.f32 %p112, %f955, %f205; selp.f32 %f956, %f940, %f943, %p112; selp.f32 %f957, %f943, %f941, %p112; add.ftz.f32 %f958, %f956, %f957; mul.ftz.f32 %f959, %f958, 0f3F000000; sub.ftz.f32 %f960, %f841, %f959; mul.ftz.f32 %f961, %f960, 0f40400000; sqrt.approx.ftz.f32 %f962, %f961; mul.ftz.f32 %f963, %f959, 0f3F000000; mul.ftz.f32 %f964, %f204, %f962; mul.ftz.f32 %f965, %f964, 0fBFB8AA3B; ex2.approx.ftz.f32 %f966, %f965; add.ftz.f32 %f967, %f966, 0f3F800000; mul.ftz.f32 %f968, %f967, %f963; mul.ftz.f32 %f969, %f962, 0fBFB8AA3B; ex2.approx.ftz.f32 %f970, %f969; mul.ftz.f32 %f971, %f970, %f968; setp.lt.ftz.f32 %p113, %f971, %f205; selp.f32 %f972, %f956, %f959, %p113; selp.f32 %f973, %f959, %f957, %p113; add.ftz.f32 %f974, %f972, %f973; mul.ftz.f32 %f975, %f974, 0f3F000000; sub.ftz.f32 %f976, %f841, %f975; mul.ftz.f32 %f977, %f976, 0f40400000; sqrt.approx.ftz.f32 %f978, %f977; mul.ftz.f32 %f979, %f975, 0f3F000000; mul.ftz.f32 %f980, %f204, %f978; mul.ftz.f32 %f981, %f980, 0fBFB8AA3B; ex2.approx.ftz.f32 %f982, %f981; add.ftz.f32 %f983, %f982, 0f3F800000; mul.ftz.f32 %f984, %f983, %f979; mul.ftz.f32 %f985, %f978, 0fBFB8AA3B; ex2.approx.ftz.f32 %f986, %f985; mul.ftz.f32 %f987, %f986, %f984; setp.lt.ftz.f32 %p114, %f987, %f205; selp.f32 %f988, %f972, %f975, %p114; selp.f32 %f989, %f975, %f973, %p114; add.ftz.f32 %f990, %f988, %f989; mul.ftz.f32 %f991, %f990, 0f3F000000; sub.ftz.f32 %f992, %f841, %f991; mul.ftz.f32 %f993, %f992, 0f40400000; sqrt.approx.ftz.f32 %f994, %f993; mul.ftz.f32 %f995, %f991, 0f3F000000; mul.ftz.f32 %f996, %f204, %f994; mul.ftz.f32 %f997, %f996, 0fBFB8AA3B; ex2.approx.ftz.f32 %f998, %f997; add.ftz.f32 %f999, %f998, 0f3F800000; mul.ftz.f32 %f1000, %f999, %f995; mul.ftz.f32 %f1001, %f994, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1002, %f1001; mul.ftz.f32 %f1003, %f1002, %f1000; setp.lt.ftz.f32 %p115, %f1003, %f205; selp.f32 %f1004, %f988, %f991, %p115; selp.f32 %f1005, %f991, %f989, %p115; add.ftz.f32 %f1006, %f1004, %f1005; mul.ftz.f32 %f1007, %f1006, 0f3F000000; sub.ftz.f32 %f1008, %f841, %f1007; mul.ftz.f32 %f1009, %f1008, 0f40400000; sqrt.approx.ftz.f32 %f1010, %f1009; mul.ftz.f32 %f1011, %f1007, 0f3F000000; mul.ftz.f32 %f1012, %f204, %f1010; mul.ftz.f32 %f1013, %f1012, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1014, %f1013; add.ftz.f32 %f1015, %f1014, 0f3F800000; mul.ftz.f32 %f1016, %f1015, %f1011; mul.ftz.f32 %f1017, %f1010, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1018, %f1017; mul.ftz.f32 %f1019, %f1018, %f1016; setp.lt.ftz.f32 %p116, %f1019, %f205; selp.f32 %f1020, %f1004, %f1007, %p116; selp.f32 %f1021, %f1007, %f1005, %p116; add.ftz.f32 %f1022, %f1020, %f1021; mul.ftz.f32 %f3004, %f1022, 0f3F000000; $L__BB193_165: mov.f32 %f3005, 0f00000000; ld.f32 %f208, [%rd14+16]; setp.lt.ftz.f32 %p117, %f208, 0f38D1B717; @%p117 bra $L__BB193_168; setp.ge.ftz.f32 %p118, %f208, 0f3F7EB852; mov.f32 %f3005, 0f3F7FFFEF; @%p118 bra $L__BB193_168; mov.f32 %f2962, 0f3F800000; sub.ftz.f32 %f1026, %f2962, 0f3F000000; mul.ftz.f32 %f1027, %f1026, 0f40400000; sqrt.approx.ftz.f32 %f1028, %f1027; mul.ftz.f32 %f1029, %f204, %f1028; mul.ftz.f32 %f1030, %f1029, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1031, %f1030; add.ftz.f32 %f1032, %f1031, 0f3F800000; mul.ftz.f32 %f1033, %f1032, 0f3E800000; mul.ftz.f32 %f1034, %f1028, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1035, %f1034; mul.ftz.f32 %f1036, %f1035, %f1033; setp.lt.ftz.f32 %p119, %f1036, %f208; selp.f32 %f1037, 0f3F800000, 0f3F000000, %p119; selp.f32 %f1038, 0f3F000000, 0f00000000, %p119; selp.f32 %f1039, 0f3F400000, 0f3E800000, %p119; sub.ftz.f32 %f1040, %f2962, %f1039; mul.ftz.f32 %f1041, %f1040, 0f40400000; sqrt.approx.ftz.f32 %f1042, %f1041; mul.ftz.f32 %f1043, %f1039, 0f3F000000; mul.ftz.f32 %f1044, %f204, %f1042; mul.ftz.f32 %f1045, %f1044, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1046, %f1045; add.ftz.f32 %f1047, %f1046, 0f3F800000; mul.ftz.f32 %f1048, %f1047, %f1043; mul.ftz.f32 %f1049, %f1042, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1050, %f1049; mul.ftz.f32 %f1051, %f1050, %f1048; setp.lt.ftz.f32 %p120, %f1051, %f208; selp.f32 %f1052, %f1037, %f1039, %p120; selp.f32 %f1053, %f1039, %f1038, %p120; add.ftz.f32 %f1054, %f1052, %f1053; mul.ftz.f32 %f1055, %f1054, 0f3F000000; sub.ftz.f32 %f1056, %f2962, %f1055; mul.ftz.f32 %f1057, %f1056, 0f40400000; sqrt.approx.ftz.f32 %f1058, %f1057; mul.ftz.f32 %f1059, %f1055, 0f3F000000; mul.ftz.f32 %f1060, %f204, %f1058; mul.ftz.f32 %f1061, %f1060, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1062, %f1061; add.ftz.f32 %f1063, %f1062, 0f3F800000; mul.ftz.f32 %f1064, %f1063, %f1059; mul.ftz.f32 %f1065, %f1058, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1066, %f1065; mul.ftz.f32 %f1067, %f1066, %f1064; setp.lt.ftz.f32 %p121, %f1067, %f208; selp.f32 %f1068, %f1052, %f1055, %p121; selp.f32 %f1069, %f1055, %f1053, %p121; add.ftz.f32 %f1070, %f1068, %f1069; mul.ftz.f32 %f1071, %f1070, 0f3F000000; sub.ftz.f32 %f1072, %f2962, %f1071; mul.ftz.f32 %f1073, %f1072, 0f40400000; sqrt.approx.ftz.f32 %f1074, %f1073; mul.ftz.f32 %f1075, %f1071, 0f3F000000; mul.ftz.f32 %f1076, %f204, %f1074; mul.ftz.f32 %f1077, %f1076, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1078, %f1077; add.ftz.f32 %f1079, %f1078, 0f3F800000; mul.ftz.f32 %f1080, %f1079, %f1075; mul.ftz.f32 %f1081, %f1074, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1082, %f1081; mul.ftz.f32 %f1083, %f1082, %f1080; setp.lt.ftz.f32 %p122, %f1083, %f208; selp.f32 %f1084, %f1068, %f1071, %p122; selp.f32 %f1085, %f1071, %f1069, %p122; add.ftz.f32 %f1086, %f1084, %f1085; mul.ftz.f32 %f1087, %f1086, 0f3F000000; sub.ftz.f32 %f1088, %f2962, %f1087; mul.ftz.f32 %f1089, %f1088, 0f40400000; sqrt.approx.ftz.f32 %f1090, %f1089; mul.ftz.f32 %f1091, %f1087, 0f3F000000; mul.ftz.f32 %f1092, %f204, %f1090; mul.ftz.f32 %f1093, %f1092, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1094, %f1093; add.ftz.f32 %f1095, %f1094, 0f3F800000; mul.ftz.f32 %f1096, %f1095, %f1091; mul.ftz.f32 %f1097, %f1090, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1098, %f1097; mul.ftz.f32 %f1099, %f1098, %f1096; setp.lt.ftz.f32 %p123, %f1099, %f208; selp.f32 %f1100, %f1084, %f1087, %p123; selp.f32 %f1101, %f1087, %f1085, %p123; add.ftz.f32 %f1102, %f1100, %f1101; mul.ftz.f32 %f1103, %f1102, 0f3F000000; sub.ftz.f32 %f1104, %f2962, %f1103; mul.ftz.f32 %f1105, %f1104, 0f40400000; sqrt.approx.ftz.f32 %f1106, %f1105; mul.ftz.f32 %f1107, %f1103, 0f3F000000; mul.ftz.f32 %f1108, %f204, %f1106; mul.ftz.f32 %f1109, %f1108, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1110, %f1109; add.ftz.f32 %f1111, %f1110, 0f3F800000; mul.ftz.f32 %f1112, %f1111, %f1107; mul.ftz.f32 %f1113, %f1106, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1114, %f1113; mul.ftz.f32 %f1115, %f1114, %f1112; setp.lt.ftz.f32 %p124, %f1115, %f208; selp.f32 %f1116, %f1100, %f1103, %p124; selp.f32 %f1117, %f1103, %f1101, %p124; add.ftz.f32 %f1118, %f1116, %f1117; mul.ftz.f32 %f1119, %f1118, 0f3F000000; sub.ftz.f32 %f1120, %f2962, %f1119; mul.ftz.f32 %f1121, %f1120, 0f40400000; sqrt.approx.ftz.f32 %f1122, %f1121; mul.ftz.f32 %f1123, %f1119, 0f3F000000; mul.ftz.f32 %f1124, %f204, %f1122; mul.ftz.f32 %f1125, %f1124, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1126, %f1125; add.ftz.f32 %f1127, %f1126, 0f3F800000; mul.ftz.f32 %f1128, %f1127, %f1123; mul.ftz.f32 %f1129, %f1122, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1130, %f1129; mul.ftz.f32 %f1131, %f1130, %f1128; setp.lt.ftz.f32 %p125, %f1131, %f208; selp.f32 %f1132, %f1116, %f1119, %p125; selp.f32 %f1133, %f1119, %f1117, %p125; add.ftz.f32 %f1134, %f1132, %f1133; mul.ftz.f32 %f1135, %f1134, 0f3F000000; sub.ftz.f32 %f1136, %f2962, %f1135; mul.ftz.f32 %f1137, %f1136, 0f40400000; sqrt.approx.ftz.f32 %f1138, %f1137; mul.ftz.f32 %f1139, %f1135, 0f3F000000; mul.ftz.f32 %f1140, %f204, %f1138; mul.ftz.f32 %f1141, %f1140, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1142, %f1141; add.ftz.f32 %f1143, %f1142, 0f3F800000; mul.ftz.f32 %f1144, %f1143, %f1139; mul.ftz.f32 %f1145, %f1138, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1146, %f1145; mul.ftz.f32 %f1147, %f1146, %f1144; setp.lt.ftz.f32 %p126, %f1147, %f208; selp.f32 %f1148, %f1132, %f1135, %p126; selp.f32 %f1149, %f1135, %f1133, %p126; add.ftz.f32 %f1150, %f1148, %f1149; mul.ftz.f32 %f1151, %f1150, 0f3F000000; sub.ftz.f32 %f1152, %f2962, %f1151; mul.ftz.f32 %f1153, %f1152, 0f40400000; sqrt.approx.ftz.f32 %f1154, %f1153; mul.ftz.f32 %f1155, %f1151, 0f3F000000; mul.ftz.f32 %f1156, %f204, %f1154; mul.ftz.f32 %f1157, %f1156, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1158, %f1157; add.ftz.f32 %f1159, %f1158, 0f3F800000; mul.ftz.f32 %f1160, %f1159, %f1155; mul.ftz.f32 %f1161, %f1154, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1162, %f1161; mul.ftz.f32 %f1163, %f1162, %f1160; setp.lt.ftz.f32 %p127, %f1163, %f208; selp.f32 %f1164, %f1148, %f1151, %p127; selp.f32 %f1165, %f1151, %f1149, %p127; add.ftz.f32 %f1166, %f1164, %f1165; mul.ftz.f32 %f1167, %f1166, 0f3F000000; sub.ftz.f32 %f1168, %f2962, %f1167; mul.ftz.f32 %f1169, %f1168, 0f40400000; sqrt.approx.ftz.f32 %f1170, %f1169; mul.ftz.f32 %f1171, %f1167, 0f3F000000; mul.ftz.f32 %f1172, %f204, %f1170; mul.ftz.f32 %f1173, %f1172, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1174, %f1173; add.ftz.f32 %f1175, %f1174, 0f3F800000; mul.ftz.f32 %f1176, %f1175, %f1171; mul.ftz.f32 %f1177, %f1170, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1178, %f1177; mul.ftz.f32 %f1179, %f1178, %f1176; setp.lt.ftz.f32 %p128, %f1179, %f208; selp.f32 %f1180, %f1164, %f1167, %p128; selp.f32 %f1181, %f1167, %f1165, %p128; add.ftz.f32 %f1182, %f1180, %f1181; mul.ftz.f32 %f1183, %f1182, 0f3F000000; sub.ftz.f32 %f1184, %f2962, %f1183; mul.ftz.f32 %f1185, %f1184, 0f40400000; sqrt.approx.ftz.f32 %f1186, %f1185; mul.ftz.f32 %f1187, %f1183, 0f3F000000; mul.ftz.f32 %f1188, %f204, %f1186; mul.ftz.f32 %f1189, %f1188, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1190, %f1189; add.ftz.f32 %f1191, %f1190, 0f3F800000; mul.ftz.f32 %f1192, %f1191, %f1187; mul.ftz.f32 %f1193, %f1186, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1194, %f1193; mul.ftz.f32 %f1195, %f1194, %f1192; setp.lt.ftz.f32 %p129, %f1195, %f208; selp.f32 %f1196, %f1180, %f1183, %p129; selp.f32 %f1197, %f1183, %f1181, %p129; add.ftz.f32 %f1198, %f1196, %f1197; mul.ftz.f32 %f3005, %f1198, 0f3F000000; $L__BB193_168: mov.f32 %f3006, 0f00000000; ld.f32 %f211, [%rd14+20]; setp.lt.ftz.f32 %p130, %f211, 0f38D1B717; @%p130 bra $L__BB193_171; setp.ge.ftz.f32 %p131, %f211, 0f3F7EB852; mov.f32 %f3006, 0f3F7FFFEF; @%p131 bra $L__BB193_171; mov.f32 %f2957, 0f3F800000; sub.ftz.f32 %f1202, %f2957, 0f3F000000; mul.ftz.f32 %f1203, %f1202, 0f40400000; sqrt.approx.ftz.f32 %f1204, %f1203; mul.ftz.f32 %f1205, %f204, %f1204; mul.ftz.f32 %f1206, %f1205, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1207, %f1206; add.ftz.f32 %f1208, %f1207, 0f3F800000; mul.ftz.f32 %f1209, %f1208, 0f3E800000; mul.ftz.f32 %f1210, %f1204, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1211, %f1210; mul.ftz.f32 %f1212, %f1211, %f1209; setp.lt.ftz.f32 %p132, %f1212, %f211; selp.f32 %f1213, 0f3F800000, 0f3F000000, %p132; selp.f32 %f1214, 0f3F000000, 0f00000000, %p132; selp.f32 %f1215, 0f3F400000, 0f3E800000, %p132; sub.ftz.f32 %f1216, %f2957, %f1215; mul.ftz.f32 %f1217, %f1216, 0f40400000; sqrt.approx.ftz.f32 %f1218, %f1217; mul.ftz.f32 %f1219, %f1215, 0f3F000000; mul.ftz.f32 %f1220, %f204, %f1218; mul.ftz.f32 %f1221, %f1220, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1222, %f1221; add.ftz.f32 %f1223, %f1222, 0f3F800000; mul.ftz.f32 %f1224, %f1223, %f1219; mul.ftz.f32 %f1225, %f1218, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1226, %f1225; mul.ftz.f32 %f1227, %f1226, %f1224; setp.lt.ftz.f32 %p133, %f1227, %f211; selp.f32 %f1228, %f1213, %f1215, %p133; selp.f32 %f1229, %f1215, %f1214, %p133; add.ftz.f32 %f1230, %f1228, %f1229; mul.ftz.f32 %f1231, %f1230, 0f3F000000; sub.ftz.f32 %f1232, %f2957, %f1231; mul.ftz.f32 %f1233, %f1232, 0f40400000; sqrt.approx.ftz.f32 %f1234, %f1233; mul.ftz.f32 %f1235, %f1231, 0f3F000000; mul.ftz.f32 %f1236, %f204, %f1234; mul.ftz.f32 %f1237, %f1236, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1238, %f1237; add.ftz.f32 %f1239, %f1238, 0f3F800000; mul.ftz.f32 %f1240, %f1239, %f1235; mul.ftz.f32 %f1241, %f1234, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1242, %f1241; mul.ftz.f32 %f1243, %f1242, %f1240; setp.lt.ftz.f32 %p134, %f1243, %f211; selp.f32 %f1244, %f1228, %f1231, %p134; selp.f32 %f1245, %f1231, %f1229, %p134; add.ftz.f32 %f1246, %f1244, %f1245; mul.ftz.f32 %f1247, %f1246, 0f3F000000; sub.ftz.f32 %f1248, %f2957, %f1247; mul.ftz.f32 %f1249, %f1248, 0f40400000; sqrt.approx.ftz.f32 %f1250, %f1249; mul.ftz.f32 %f1251, %f1247, 0f3F000000; mul.ftz.f32 %f1252, %f204, %f1250; mul.ftz.f32 %f1253, %f1252, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1254, %f1253; add.ftz.f32 %f1255, %f1254, 0f3F800000; mul.ftz.f32 %f1256, %f1255, %f1251; mul.ftz.f32 %f1257, %f1250, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1258, %f1257; mul.ftz.f32 %f1259, %f1258, %f1256; setp.lt.ftz.f32 %p135, %f1259, %f211; selp.f32 %f1260, %f1244, %f1247, %p135; selp.f32 %f1261, %f1247, %f1245, %p135; add.ftz.f32 %f1262, %f1260, %f1261; mul.ftz.f32 %f1263, %f1262, 0f3F000000; sub.ftz.f32 %f1264, %f2957, %f1263; mul.ftz.f32 %f1265, %f1264, 0f40400000; sqrt.approx.ftz.f32 %f1266, %f1265; mul.ftz.f32 %f1267, %f1263, 0f3F000000; mul.ftz.f32 %f1268, %f204, %f1266; mul.ftz.f32 %f1269, %f1268, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1270, %f1269; add.ftz.f32 %f1271, %f1270, 0f3F800000; mul.ftz.f32 %f1272, %f1271, %f1267; mul.ftz.f32 %f1273, %f1266, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1274, %f1273; mul.ftz.f32 %f1275, %f1274, %f1272; setp.lt.ftz.f32 %p136, %f1275, %f211; selp.f32 %f1276, %f1260, %f1263, %p136; selp.f32 %f1277, %f1263, %f1261, %p136; add.ftz.f32 %f1278, %f1276, %f1277; mul.ftz.f32 %f1279, %f1278, 0f3F000000; sub.ftz.f32 %f1280, %f2957, %f1279; mul.ftz.f32 %f1281, %f1280, 0f40400000; sqrt.approx.ftz.f32 %f1282, %f1281; mul.ftz.f32 %f1283, %f1279, 0f3F000000; mul.ftz.f32 %f1284, %f204, %f1282; mul.ftz.f32 %f1285, %f1284, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1286, %f1285; add.ftz.f32 %f1287, %f1286, 0f3F800000; mul.ftz.f32 %f1288, %f1287, %f1283; mul.ftz.f32 %f1289, %f1282, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1290, %f1289; mul.ftz.f32 %f1291, %f1290, %f1288; setp.lt.ftz.f32 %p137, %f1291, %f211; selp.f32 %f1292, %f1276, %f1279, %p137; selp.f32 %f1293, %f1279, %f1277, %p137; add.ftz.f32 %f1294, %f1292, %f1293; mul.ftz.f32 %f1295, %f1294, 0f3F000000; sub.ftz.f32 %f1296, %f2957, %f1295; mul.ftz.f32 %f1297, %f1296, 0f40400000; sqrt.approx.ftz.f32 %f1298, %f1297; mul.ftz.f32 %f1299, %f1295, 0f3F000000; mul.ftz.f32 %f1300, %f204, %f1298; mul.ftz.f32 %f1301, %f1300, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1302, %f1301; add.ftz.f32 %f1303, %f1302, 0f3F800000; mul.ftz.f32 %f1304, %f1303, %f1299; mul.ftz.f32 %f1305, %f1298, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1306, %f1305; mul.ftz.f32 %f1307, %f1306, %f1304; setp.lt.ftz.f32 %p138, %f1307, %f211; selp.f32 %f1308, %f1292, %f1295, %p138; selp.f32 %f1309, %f1295, %f1293, %p138; add.ftz.f32 %f1310, %f1308, %f1309; mul.ftz.f32 %f1311, %f1310, 0f3F000000; sub.ftz.f32 %f1312, %f2957, %f1311; mul.ftz.f32 %f1313, %f1312, 0f40400000; sqrt.approx.ftz.f32 %f1314, %f1313; mul.ftz.f32 %f1315, %f1311, 0f3F000000; mul.ftz.f32 %f1316, %f204, %f1314; mul.ftz.f32 %f1317, %f1316, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1318, %f1317; add.ftz.f32 %f1319, %f1318, 0f3F800000; mul.ftz.f32 %f1320, %f1319, %f1315; mul.ftz.f32 %f1321, %f1314, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1322, %f1321; mul.ftz.f32 %f1323, %f1322, %f1320; setp.lt.ftz.f32 %p139, %f1323, %f211; selp.f32 %f1324, %f1308, %f1311, %p139; selp.f32 %f1325, %f1311, %f1309, %p139; add.ftz.f32 %f1326, %f1324, %f1325; mul.ftz.f32 %f1327, %f1326, 0f3F000000; sub.ftz.f32 %f1328, %f2957, %f1327; mul.ftz.f32 %f1329, %f1328, 0f40400000; sqrt.approx.ftz.f32 %f1330, %f1329; mul.ftz.f32 %f1331, %f1327, 0f3F000000; mul.ftz.f32 %f1332, %f204, %f1330; mul.ftz.f32 %f1333, %f1332, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1334, %f1333; add.ftz.f32 %f1335, %f1334, 0f3F800000; mul.ftz.f32 %f1336, %f1335, %f1331; mul.ftz.f32 %f1337, %f1330, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1338, %f1337; mul.ftz.f32 %f1339, %f1338, %f1336; setp.lt.ftz.f32 %p140, %f1339, %f211; selp.f32 %f1340, %f1324, %f1327, %p140; selp.f32 %f1341, %f1327, %f1325, %p140; add.ftz.f32 %f1342, %f1340, %f1341; mul.ftz.f32 %f1343, %f1342, 0f3F000000; sub.ftz.f32 %f1344, %f2957, %f1343; mul.ftz.f32 %f1345, %f1344, 0f40400000; sqrt.approx.ftz.f32 %f1346, %f1345; mul.ftz.f32 %f1347, %f1343, 0f3F000000; mul.ftz.f32 %f1348, %f204, %f1346; mul.ftz.f32 %f1349, %f1348, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1350, %f1349; add.ftz.f32 %f1351, %f1350, 0f3F800000; mul.ftz.f32 %f1352, %f1351, %f1347; mul.ftz.f32 %f1353, %f1346, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1354, %f1353; mul.ftz.f32 %f1355, %f1354, %f1352; setp.lt.ftz.f32 %p141, %f1355, %f211; selp.f32 %f1356, %f1340, %f1343, %p141; selp.f32 %f1357, %f1343, %f1341, %p141; add.ftz.f32 %f1358, %f1356, %f1357; mul.ftz.f32 %f1359, %f1358, 0f3F000000; sub.ftz.f32 %f1360, %f2957, %f1359; mul.ftz.f32 %f1361, %f1360, 0f40400000; sqrt.approx.ftz.f32 %f1362, %f1361; mul.ftz.f32 %f1363, %f1359, 0f3F000000; mul.ftz.f32 %f1364, %f204, %f1362; mul.ftz.f32 %f1365, %f1364, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1366, %f1365; add.ftz.f32 %f1367, %f1366, 0f3F800000; mul.ftz.f32 %f1368, %f1367, %f1363; mul.ftz.f32 %f1369, %f1362, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f1370, %f1368; setp.lt.ftz.f32 %p142, %f1371, %f211; selp.f32 %f1372, %f1356, %f1359, %p142; selp.f32 %f1373, %f1359, %f1357, %p142; add.ftz.f32 %f1374, %f1372, %f1373; mul.ftz.f32 %f3006, %f1374, 0f3F000000; $L__BB193_171: mov.f32 %f2958, 0f3F800000; sub.ftz.f32 %f1376, %f2958, %f3004; mul.ftz.f32 %f1377, %f1376, 0f40400000; sub.ftz.f32 %f1378, %f2958, %f3005; mul.ftz.f32 %f1379, %f1378, 0f40400000; sub.ftz.f32 %f1380, %f2958, %f3006; mul.ftz.f32 %f1381, %f1380, 0f40400000; sqrt.approx.ftz.f32 %f1382, %f1377; sqrt.approx.ftz.f32 %f1383, %f1379; sqrt.approx.ftz.f32 %f1384, %f1381; ld.f32 %f1385, [%rd14]; mul.ftz.f32 %f3007, %f1382, %f1385; ld.f32 %f1386, [%rd14+4]; mul.ftz.f32 %f3008, %f1383, %f1386; ld.f32 %f1387, [%rd14+8]; mul.ftz.f32 %f3009, %f1384, %f1387; $L__BB193_173: st.f32 [%rd14], %f3007; st.f32 [%rd14+4], %f3008; st.f32 [%rd14+8], %f3009; or.b32 %r643, %r643, 16; $L__BB193_174: ld.local.u32 %r224, [%rd4+40]; or.b32 %r225, %r224, %r643; st.local.u32 [%rd4+40], %r225; $L__BB193_182: setp.leu.ftz.f32 %p147, %f97, 0f3727C5AC; setp.leu.ftz.f32 %p148, %f74, 0f3727C5AC; or.pred %p149, %p148, %p147; @%p149 bra $L__BB193_194; ld.const.v4.f32 {%f1404, %f1405, %f1406, %f1407}, [kernel_params+1184]; mul.ftz.f32 %f1411, %f187, %f1405; fma.rn.ftz.f32 %f1412, %f186, %f1404, %f1411; fma.rn.ftz.f32 %f227, %f188, %f1406, %f1412; setp.leu.ftz.f32 %p150, %f227, 0f00000000; mov.f32 %f1403, 0f3F800000; mov.f32 %f3010, %f1403; mov.f32 %f3011, %f1403; mov.f32 %f3012, %f1403; @%p150 bra $L__BB193_185; div.approx.ftz.f32 %f3012, %f186, %f227; div.approx.ftz.f32 %f3011, %f187, %f227; div.approx.ftz.f32 %f3010, %f188, %f227; $L__BB193_185: sub.ftz.f32 %f1414, %f1403, %f75; fma.rn.ftz.f32 %f1415, %f75, %f3012, %f1414; fma.rn.ftz.f32 %f1416, %f75, %f3011, %f1414; fma.rn.ftz.f32 %f1417, %f75, %f3010, %f1414; mul.ftz.f32 %f1418, %f74, %f177; mul.ftz.f32 %f1419, %f1418, %f1415; mul.ftz.f32 %f1420, %f74, %f178; mul.ftz.f32 %f1421, %f1420, %f1416; mul.ftz.f32 %f1422, %f74, %f179; mul.ftz.f32 %f1423, %f1422, %f1417; mul.ftz.f32 %f1424, %f97, %f1419; mul.ftz.f32 %f1425, %f97, %f1421; mul.ftz.f32 %f1426, %f97, %f1423; mov.f32 %f3013, 0f00000000; max.ftz.f32 %f234, %f1424, %f3013; max.ftz.f32 %f235, %f1425, %f3013; max.ftz.f32 %f236, %f1426, %f3013; add.ftz.f32 %f1428, %f234, %f235; add.ftz.f32 %f1429, %f236, %f1428; mul.ftz.f32 %f1430, %f1429, 0f3EAAAAAB; abs.ftz.f32 %f237, %f1430; setp.ltu.ftz.f32 %p151, %f237, 0f3727C5AC; mov.u64 %rd712, 0; mov.u64 %rd713, %rd712; @%p151 bra $L__BB193_190; ld.local.u32 %r43, [%rd4+252]; mov.u64 %rd712, 0; setp.eq.s32 %p152, %r43, 0; mov.u64 %rd710, %rd712; mov.u64 %rd711, %rd712; @%p152 bra $L__BB193_188; ld.local.u32 %r237, [%rd4+248]; mul.wide.s32 %rd337, %r237, 80; add.s64 %rd338, %rd1, %rd337; add.s64 %rd710, %rd338, 304; add.s64 %rd339, %rd230, %rd337; add.s64 %rd711, %rd339, 304; mov.b32 %r238, %f236; mov.b32 %r239, %f235; mov.b32 %r240, %f234; mov.u32 %r241, 0; st.local.v4.u32 [%rd338+304], {%r240, %r239, %r238, %r241}; add.s32 %r242, %r237, 1; st.local.u32 [%rd4+248], %r242; add.s32 %r243, %r43, -1; st.local.u32 [%rd4+252], %r243; $L__BB193_188: setp.eq.s64 %p153, %rd711, 0; mov.u64 %rd713, %rd712; @%p153 bra $L__BB193_190; st.local.f32 [%rd710+16], %f237; mov.u64 %rd712, %rd710; mov.u64 %rd713, %rd711; $L__BB193_190: setp.eq.s64 %p154, %rd713, 0; @%p154 bra $L__BB193_194; add.s64 %rd45, %rd712, 20; st.local.f32 [%rd712+20], %f55; st.local.f32 [%rd712+24], %f56; st.local.f32 [%rd712+28], %f57; mov.u32 %r244, 6; st.local.u32 [%rd712+12], %r244; ld.local.f32 %f1432, [%rd4+24]; ld.local.f32 %f1433, [%rd4+28]; mul.ftz.f32 %f1434, %f56, %f1433; fma.rn.ftz.f32 %f1435, %f55, %f1432, %f1434; ld.local.f32 %f1436, [%rd4+32]; fma.rn.ftz.f32 %f238, %f57, %f1436, %f1435; setp.lt.ftz.f32 %p155, %f238, 0f00000000; @%p155 bra $L__BB193_193; mov.f32 %f1437, 0f3F800000; sub.ftz.f32 %f1438, %f1437, %f238; mov.f32 %f1439, 0f00000000; max.ftz.f32 %f1440, %f1438, %f1439; min.ftz.f32 %f1441, %f1440, %f1437; mul.ftz.f32 %f1442, %f1441, %f1441; mul.ftz.f32 %f1443, %f1442, %f1442; mul.ftz.f32 %f1444, %f1441, %f1443; mul.ftz.f32 %f3013, %f238, %f1444; $L__BB193_193: st.local.f32 [%rd45+12], %f3013; ld.local.f32 %f1445, [%rd45+-4]; mul.ftz.f32 %f1446, %f3013, %f1445; st.local.f32 [%rd45+-4], %f1446; ld.local.u32 %r245, [%rd4+40]; or.b32 %r246, %r245, 12; st.local.u32 [%rd4+40], %r246; $L__BB193_194: ld.param.u32 %r637, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_3]; ld.const.u32 %r44, [kernel_params+1564]; setp.ne.s32 %p156, %r44, 0; and.b32 %r45, %r637, 8; setp.eq.s32 %p157, %r45, 0; or.pred %p5, %p157, %p156; not.pred %p158, %p5; setp.leu.ftz.f32 %p159, %f99, 0f3727C5AC; or.pred %p160, %p159, %p158; @%p160 bra $L__BB193_225; setp.leu.ftz.f32 %p161, %f70, 0f3727C5AC; setp.leu.ftz.f32 %p162, %f60, 0f3727C5AC; and.pred %p163, %p162, %p161; @%p163 bra $L__BB193_225; mul.ftz.f32 %f1447, %f99, %f177; mov.f32 %f1448, 0f00000000; max.ftz.f32 %f241, %f1447, %f1448; mul.ftz.f32 %f1449, %f99, %f178; max.ftz.f32 %f242, %f1449, %f1448; mul.ftz.f32 %f1450, %f99, %f179; max.ftz.f32 %f243, %f1450, %f1448; add.ftz.f32 %f1451, %f241, %f242; add.ftz.f32 %f1452, %f243, %f1451; mul.ftz.f32 %f1453, %f1452, 0f3EAAAAAB; abs.ftz.f32 %f244, %f1453; setp.ltu.ftz.f32 %p164, %f244, 0f3727C5AC; mov.u64 %rd720, 0; mov.u64 %rd716, %rd720; mov.u64 %rd717, %rd720; @%p164 bra $L__BB193_201; ld.local.u32 %r46, [%rd4+252]; mov.u64 %rd716, 0; setp.eq.s32 %p165, %r46, 0; mov.u64 %rd714, %rd716; mov.u64 %rd715, %rd716; @%p165 bra $L__BB193_199; ld.local.u32 %r247, [%rd4+248]; mul.wide.s32 %rd346, %r247, 80; add.s64 %rd347, %rd1, %rd346; add.s64 %rd714, %rd347, 304; add.s64 %rd348, %rd230, %rd346; add.s64 %rd715, %rd348, 304; mov.b32 %r248, %f243; mov.b32 %r249, %f242; mov.b32 %r250, %f241; mov.u32 %r251, 0; st.local.v4.u32 [%rd347+304], {%r250, %r249, %r248, %r251}; add.s32 %r252, %r247, 1; st.local.u32 [%rd4+248], %r252; add.s32 %r253, %r46, -1; st.local.u32 [%rd4+252], %r253; $L__BB193_199: setp.eq.s64 %p166, %rd715, 0; mov.u64 %rd717, %rd716; @%p166 bra $L__BB193_201; st.local.f32 [%rd714+16], %f244; mov.u64 %rd716, %rd714; mov.u64 %rd717, %rd715; $L__BB193_201: setp.eq.s64 %p167, %rd717, 0; mov.u64 %rd721, %rd720; @%p167 bra $L__BB193_205; ld.local.u32 %r47, [%rd4+252]; setp.lt.s32 %p168, %r47, 1; @%p168 bra $L__BB193_204; add.s32 %r254, %r47, -1; st.local.u32 [%rd4+252], %r254; ld.local.s32 %rd353, [%rd4+248]; cvt.s64.s32 %rd354, %r254; add.s64 %rd355, %rd353, %rd354; mul.lo.s64 %rd356, %rd355, 80; add.s64 %rd357, %rd1, %rd356; add.s64 %rd720, %rd357, 304; add.s64 %rd358, %rd230, %rd356; add.s64 %rd721, %rd358, 304; bra.uni $L__BB193_205; $L__BB193_204: ld.local.u32 %r255, [%rd4+248]; mov.u64 %rd720, 0; add.s32 %r256, %r255, -1; st.local.u32 [%rd4+248], %r256; add.s32 %r257, %r47, 1; st.local.u32 [%rd4+252], %r257; mov.u64 %rd721, %rd720; $L__BB193_205: setp.eq.s64 %p169, %rd721, 0; or.pred %p171, %p167, %p169; @%p171 bra $L__BB193_225; add.s64 %rd60, %rd716, 20; st.local.f32 [%rd716+20], %f55; st.local.f32 [%rd716+24], %f56; st.local.f32 [%rd716+28], %f57; mul.ftz.f32 %f245, %f70, 0f3DA3D70A; max.ftz.f32 %f1458, %f245, %f1448; sqrt.approx.ftz.f32 %f1459, %f1458; mov.f32 %f1456, 0f3F800000; sub.ftz.f32 %f1460, %f1456, %f1459; mov.f32 %f1461, 0f40000000; div.approx.ftz.f32 %f1462, %f1461, %f1460; add.ftz.f32 %f1463, %f1462, 0fBF800000; st.local.f32 [%rd716+40], %f1463; st.local.v2.f32 [%rd716+56], {%f2977, %f2978}; st.local.f32 [%rd716+64], %f2979; st.local.u64 [%rd716+48], %rd721; fma.rn.ftz.f32 %f1464, %f73, 0fBF666666, 0f3F800000; max.ftz.f32 %f1465, %f1464, %f1448; sqrt.approx.ftz.f32 %f1466, %f1465; mul.ftz.f32 %f1467, %f71, %f71; mul.ftz.f32 %f1468, %f1467, %f1466; div.approx.ftz.f32 %f1469, %f1467, %f1466; st.local.v2.f32 [%rd716+32], {%f1469, %f1468}; mul.ftz.f32 %f1470, %f187, 0f3F19999A; fma.rn.ftz.f32 %f1471, %f186, 0f3E99999A, %f1470; fma.rn.ftz.f32 %f246, %f188, 0f3DCCCCCD, %f1471; setp.leu.ftz.f32 %p172, %f246, 0f00000000; mov.f32 %f3014, %f1456; mov.f32 %f3015, %f1456; mov.f32 %f3016, %f1456; @%p172 bra $L__BB193_208; div.approx.ftz.f32 %f3016, %f186, %f246; div.approx.ftz.f32 %f3015, %f187, %f246; div.approx.ftz.f32 %f3014, %f188, %f246; $L__BB193_208: sub.ftz.f32 %f1473, %f1456, %f72; fma.rn.ftz.f32 %f1474, %f72, %f3016, %f1473; fma.rn.ftz.f32 %f1475, %f72, %f3015, %f1473; fma.rn.ftz.f32 %f1476, %f72, %f3014, %f1473; mul.ftz.f32 %f1477, %f245, %f1474; mul.ftz.f32 %f1478, %f245, %f1475; mul.ftz.f32 %f1479, %f245, %f1476; sub.ftz.f32 %f1480, %f1456, %f60; mul.ftz.f32 %f1481, %f1480, %f1477; mul.ftz.f32 %f1482, %f1480, %f1478; mul.ftz.f32 %f1483, %f1480, %f1479; fma.rn.ftz.f32 %f1484, %f60, %f186, %f1481; fma.rn.ftz.f32 %f1485, %f60, %f187, %f1482; fma.rn.ftz.f32 %f1486, %f60, %f188, %f1483; st.local.f32 [%rd720+12], %f1484; st.local.f32 [%rd720+16], %f1485; st.local.f32 [%rd720+20], %f1486; ld.local.u64 %rd361, [%rd60+28]; st.f32 [%rd361], %f186; st.f32 [%rd361+4], %f187; st.f32 [%rd361+8], %f188; mov.u32 %r258, 0; ld.local.u64 %rd362, [%rd60+28]; st.u32 [%rd362+36], %r258; setp.le.ftz.f32 %p173, %f71, 0f3D99999A; setp.eq.s32 %p174, %r148, 26; or.pred %p175, %p174, %p173; @%p175 bra $L__BB193_218; bra.uni $L__BB193_209; $L__BB193_218: ld.local.u64 %rd367, [%rd60+28]; ld.f32 %f1573, [%rd367+12]; cvt.ftz.sat.f32.f32 %f1574, %f1573; ld.f32 %f1575, [%rd367+16]; cvt.ftz.sat.f32.f32 %f1576, %f1575; ld.f32 %f1577, [%rd367+20]; cvt.ftz.sat.f32.f32 %f1578, %f1577; st.f32 [%rd367+12], %f1574; st.f32 [%rd367+16], %f1576; st.f32 [%rd367+20], %f1578; ld.local.v2.f32 {%f1579, %f1580}, [%rd60+12]; cvt.ftz.sat.f32.f32 %f1583, %f1580; cvt.ftz.sat.f32.f32 %f1584, %f1579; st.local.v2.f32 [%rd60+12], {%f1584, %f1583}; mov.u32 %r264, 11; st.local.u32 [%rd60+-8], %r264; ld.local.f32 %f1585, [%rd60+20]; fma.rn.ftz.f32 %f266, %f1585, %f1585, 0fBF800000; mov.f32 %f1572, 0f3F800000; abs.ftz.f32 %f267, %f1572; fma.rn.ftz.f32 %f268, %f267, %f267, %f266; setp.leu.ftz.f32 %p184, %f268, 0f00000000; mov.f32 %f3019, %f1572; @%p184 bra $L__BB193_220; sqrt.approx.ftz.f32 %f1586, %f268; sub.ftz.f32 %f1587, %f1586, %f267; add.ftz.f32 %f1588, %f267, %f1586; div.approx.ftz.f32 %f1589, %f1587, %f1588; fma.rn.ftz.f32 %f1590, %f267, %f1588, 0fBF800000; fma.rn.ftz.f32 %f1591, %f267, %f1587, 0f3F800000; div.approx.ftz.f32 %f1592, %f1590, %f1591; mul.ftz.f32 %f1593, %f1589, 0f3F000000; mul.ftz.f32 %f1594, %f1589, %f1593; fma.rn.ftz.f32 %f1595, %f1592, %f1592, 0f3F800000; mul.ftz.f32 %f3019, %f1594, %f1595; $L__BB193_220: ld.local.u64 %rd368, [%rd60+28]; add.s64 %rd62, %rd368, 12; ld.f32 %f271, [%rd368+12]; ld.f32 %f272, [%rd368+16]; ld.f32 %f273, [%rd368+20]; sub.ftz.f32 %f274, %f1572, %f3019; ld.local.f32 %f1597, [%rd60]; ld.local.f32 %f1598, [%rd4+24]; ld.local.f32 %f1599, [%rd60+4]; ld.local.f32 %f1600, [%rd4+28]; mul.ftz.f32 %f1601, %f1600, %f1599; fma.rn.ftz.f32 %f1602, %f1598, %f1597, %f1601; ld.local.f32 %f1603, [%rd60+8]; ld.local.f32 %f1604, [%rd4+32]; fma.rn.ftz.f32 %f1605, %f1604, %f1603, %f1602; abs.ftz.f32 %f275, %f1605; fma.rn.ftz.f32 %f276, %f275, %f275, %f266; setp.leu.ftz.f32 %p185, %f276, 0f00000000; mov.f32 %f3020, %f274; @%p185 bra $L__BB193_222; sqrt.approx.ftz.f32 %f1606, %f276; sub.ftz.f32 %f1607, %f1606, %f275; add.ftz.f32 %f1608, %f275, %f1606; div.approx.ftz.f32 %f1609, %f1607, %f1608; fma.rn.ftz.f32 %f1610, %f275, %f1608, 0fBF800000; fma.rn.ftz.f32 %f1611, %f275, %f1607, 0f3F800000; div.approx.ftz.f32 %f1612, %f1610, %f1611; mul.ftz.f32 %f1613, %f1609, 0f3F000000; mul.ftz.f32 %f1614, %f1609, %f1613; fma.rn.ftz.f32 %f1615, %f1612, %f1612, 0f3F800000; mul.ftz.f32 %f1616, %f1614, %f1615; sub.ftz.f32 %f3020, %f1616, %f3019; $L__BB193_222: rcp.approx.ftz.f32 %f1617, %f274; mov.f32 %f1618, 0f3F800000; mul.ftz.f32 %f1619, %f1617, %f3020; sub.ftz.f32 %f1620, %f1618, %f1619; fma.rn.ftz.f32 %f1621, %f271, %f1620, %f1619; fma.rn.ftz.f32 %f1622, %f272, %f1620, %f1619; fma.rn.ftz.f32 %f1623, %f273, %f1620, %f1619; st.f32 [%rd62+12], %f1621; st.f32 [%rd62+16], %f1622; st.f32 [%rd62+20], %f1623; ld.local.u32 %r265, [%rd60+-8]; setp.ne.s32 %p186, %r265, 12; @%p186 bra $L__BB193_224; ld.local.u64 %rd369, [%rd60+28]; ld.f32 %f1624, [%rd369+36]; mul.ftz.f32 %f1625, %f1624, 0f3E800000; ld.f32 %f1626, [%rd369+24]; mul.ftz.f32 %f1627, %f1626, %f1625; ld.f32 %f1628, [%rd369+28]; mul.ftz.f32 %f1629, %f1625, %f1628; ld.f32 %f1630, [%rd369+32]; mul.ftz.f32 %f1631, %f1625, %f1630; st.f32 [%rd369+24], %f1627; st.f32 [%rd369+28], %f1629; st.f32 [%rd369+32], %f1631; $L__BB193_224: ld.local.u64 %rd370, [%rd60+28]; ld.f32 %f1632, [%rd370+28]; ld.f32 %f1633, [%rd370+24]; add.ftz.f32 %f1634, %f1633, %f1632; ld.f32 %f1635, [%rd370+32]; add.ftz.f32 %f1636, %f1635, %f1634; mul.ftz.f32 %f1637, %f1636, 0f3EAAAAAB; ld.local.f32 %f1638, [%rd60+-4]; mul.ftz.f32 %f1639, %f1638, %f1637; st.local.f32 [%rd60+-4], %f1639; ld.local.u32 %r266, [%rd4+40]; or.b32 %r267, %r266, 12; st.local.u32 [%rd4+40], %r267; bra.uni $L__BB193_225; $L__BB193_209: ld.local.v2.f32 {%f1487, %f1488}, [%rd60+36]; setp.neu.ftz.f32 %p176, %f1487, 0f00000000; setp.neu.ftz.f32 %p177, %f1488, 0f00000000; or.pred %p178, %p176, %p177; ld.local.f32 %f1491, [%rd60+44]; setp.neu.ftz.f32 %p179, %f1491, 0f00000000; or.pred %p180, %p179, %p178; @%p180 bra $L__BB193_211; mov.f32 %f1493, 0f3F800000; st.local.v2.f32 [%rd60+36], {%f1493, %f1448}; st.local.u32 [%rd60+44], %r258; $L__BB193_211: mov.u32 %r260, 15; st.local.u32 [%rd60+-8], %r260; ld.local.f32 %f1495, [%rd60+20]; fma.rn.ftz.f32 %f253, %f1495, %f1495, 0fBF800000; mov.f32 %f1494, 0f3F800000; abs.ftz.f32 %f254, %f1494; fma.rn.ftz.f32 %f255, %f254, %f254, %f253; setp.leu.ftz.f32 %p181, %f255, 0f00000000; mov.f32 %f3017, %f1494; @%p181 bra $L__BB193_213; sqrt.approx.ftz.f32 %f1496, %f255; sub.ftz.f32 %f1497, %f1496, %f254; add.ftz.f32 %f1498, %f254, %f1496; div.approx.ftz.f32 %f1499, %f1497, %f1498; fma.rn.ftz.f32 %f1500, %f254, %f1498, 0fBF800000; fma.rn.ftz.f32 %f1501, %f254, %f1497, 0f3F800000; div.approx.ftz.f32 %f1502, %f1500, %f1501; mul.ftz.f32 %f1503, %f1499, 0f3F000000; mul.ftz.f32 %f1504, %f1499, %f1503; fma.rn.ftz.f32 %f1505, %f1502, %f1502, 0f3F800000; mul.ftz.f32 %f3017, %f1504, %f1505; $L__BB193_213: ld.local.u64 %rd363, [%rd60+28]; add.s64 %rd61, %rd363, 12; ld.f32 %f258, [%rd363+12]; ld.f32 %f259, [%rd363+16]; ld.f32 %f260, [%rd363+20]; sub.ftz.f32 %f261, %f1494, %f3017; ld.local.f32 %f1507, [%rd60]; ld.local.f32 %f1508, [%rd4+24]; ld.local.f32 %f1509, [%rd60+4]; ld.local.f32 %f1510, [%rd4+28]; mul.ftz.f32 %f1511, %f1510, %f1509; fma.rn.ftz.f32 %f1512, %f1508, %f1507, %f1511; ld.local.f32 %f1513, [%rd60+8]; ld.local.f32 %f1514, [%rd4+32]; fma.rn.ftz.f32 %f1515, %f1514, %f1513, %f1512; abs.ftz.f32 %f262, %f1515; fma.rn.ftz.f32 %f263, %f262, %f262, %f253; setp.leu.ftz.f32 %p182, %f263, 0f00000000; mov.f32 %f3018, %f261; @%p182 bra $L__BB193_215; sqrt.approx.ftz.f32 %f1516, %f263; sub.ftz.f32 %f1517, %f1516, %f262; add.ftz.f32 %f1518, %f262, %f1516; div.approx.ftz.f32 %f1519, %f1517, %f1518; fma.rn.ftz.f32 %f1520, %f262, %f1518, 0fBF800000; fma.rn.ftz.f32 %f1521, %f262, %f1517, 0f3F800000; div.approx.ftz.f32 %f1522, %f1520, %f1521; mul.ftz.f32 %f1523, %f1519, 0f3F000000; mul.ftz.f32 %f1524, %f1519, %f1523; fma.rn.ftz.f32 %f1525, %f1522, %f1522, 0f3F800000; mul.ftz.f32 %f1526, %f1524, %f1525; sub.ftz.f32 %f3018, %f1526, %f3017; $L__BB193_215: rcp.approx.ftz.f32 %f1527, %f261; mov.f32 %f1528, 0f3F800000; mul.ftz.f32 %f1529, %f1527, %f3018; sub.ftz.f32 %f1530, %f1528, %f1529; fma.rn.ftz.f32 %f1531, %f258, %f1530, %f1529; fma.rn.ftz.f32 %f1532, %f259, %f1530, %f1529; fma.rn.ftz.f32 %f1533, %f260, %f1530, %f1529; st.f32 [%rd61+12], %f1531; st.f32 [%rd61+16], %f1532; st.f32 [%rd61+20], %f1533; ld.local.u32 %r261, [%rd60+-8]; setp.ne.s32 %p183, %r261, 12; @%p183 bra $L__BB193_217; ld.local.u64 %rd364, [%rd60+28]; ld.f32 %f1534, [%rd364+36]; mul.ftz.f32 %f1535, %f1534, 0f3E800000; ld.f32 %f1536, [%rd364+24]; mul.ftz.f32 %f1537, %f1536, %f1535; ld.f32 %f1538, [%rd364+28]; mul.ftz.f32 %f1539, %f1535, %f1538; ld.f32 %f1540, [%rd364+32]; mul.ftz.f32 %f1541, %f1535, %f1540; st.f32 [%rd364+24], %f1537; st.f32 [%rd364+28], %f1539; st.f32 [%rd364+32], %f1541; $L__BB193_217: ld.local.u64 %rd365, [%rd60+28]; ld.f32 %f1542, [%rd365+28]; ld.f32 %f1543, [%rd365+24]; add.ftz.f32 %f1544, %f1543, %f1542; ld.f32 %f1545, [%rd365+32]; add.ftz.f32 %f1546, %f1545, %f1544; mul.ftz.f32 %f1547, %f1546, 0f3EAAAAAB; ld.local.f32 %f1548, [%rd60+-4]; mul.ftz.f32 %f1549, %f1548, %f1547; st.local.f32 [%rd60+-4], %f1549; ld.local.v2.f32 {%f1550, %f1551}, [%rd60+12]; mov.f32 %f1554, 0f38D1B717; max.ftz.f32 %f1555, %f1550, %f1554; max.ftz.f32 %f1557, %f1551, %f1554; min.ftz.f32 %f1558, %f1557, %f1528; min.ftz.f32 %f1559, %f1555, %f1528; st.local.v2.f32 [%rd60+12], {%f1559, %f1558}; ld.f32 %f1560, [%rd365]; cvt.ftz.sat.f32.f32 %f1561, %f1560; ld.f32 %f1562, [%rd365+4]; cvt.ftz.sat.f32.f32 %f1563, %f1562; ld.f32 %f1564, [%rd365+8]; cvt.ftz.sat.f32.f32 %f1565, %f1564; st.f32 [%rd365], %f1561; st.f32 [%rd365+4], %f1563; st.f32 [%rd365+8], %f1565; ld.local.u64 %rd366, [%rd60+28]; ld.f32 %f1566, [%rd366+12]; cvt.ftz.sat.f32.f32 %f1567, %f1566; ld.f32 %f1568, [%rd366+16]; cvt.ftz.sat.f32.f32 %f1569, %f1568; ld.f32 %f1570, [%rd366+20]; cvt.ftz.sat.f32.f32 %f1571, %f1570; st.f32 [%rd366+12], %f1567; st.f32 [%rd366+16], %f1569; st.f32 [%rd366+20], %f1571; ld.local.u32 %r262, [%rd4+40]; or.b32 %r263, %r262, 1036; st.local.u32 [%rd4+40], %r263; $L__BB193_225: ld.const.u32 %r48, [kernel_params+1568]; or.b32 %r268, %r48, %r44; setp.eq.s32 %p187, %r268, 0; @%p187 bra $L__BB193_227; setp.gt.ftz.f32 %p188, %f98, 0f3727C5AC; @%p188 bra $L__BB193_228; bra.uni $L__BB193_272; $L__BB193_227: setp.leu.ftz.f32 %p189, %f98, 0f3727C5AC; setp.ne.s32 %p190, %r45, 0; or.pred %p191, %p190, %p189; @%p191 bra $L__BB193_272; $L__BB193_228: mul.ftz.f32 %f279, %f98, %f177; mul.ftz.f32 %f280, %f98, %f178; mul.ftz.f32 %f281, %f98, %f179; mov.f32 %f1640, 0f3F800000; sub.ftz.f32 %f1641, %f1640, %f72; fma.rn.ftz.f32 %f282, %f72, %f186, %f1641; fma.rn.ftz.f32 %f283, %f72, %f187, %f1641; fma.rn.ftz.f32 %f284, %f72, %f188, %f1641; setp.eq.s32 %p192, %r148, 26; setp.le.ftz.f32 %p193, %f71, 0f3D4CCCCD; or.pred %p194, %p192, %p193; @%p194 bra $L__BB193_246; bra.uni $L__BB193_229; $L__BB193_246: @%p158 bra $L__BB193_264; mul.ftz.f32 %f1722, %f2981, %f279; mov.f32 %f1723, 0f00000000; max.ftz.f32 %f302, %f1722, %f1723; mul.ftz.f32 %f1724, %f2981, %f280; max.ftz.f32 %f303, %f1724, %f1723; mul.ftz.f32 %f1725, %f2981, %f281; max.ftz.f32 %f304, %f1725, %f1723; add.ftz.f32 %f1726, %f302, %f303; add.ftz.f32 %f1727, %f304, %f1726; mul.ftz.f32 %f1728, %f1727, 0f3EAAAAAB; abs.ftz.f32 %f305, %f1728; setp.ltu.ftz.f32 %p207, %f305, 0f3727C5AC; mov.u64 %rd732, 0; mov.u64 %rd730, %rd732; mov.u64 %rd731, %rd732; @%p207 bra $L__BB193_252; ld.local.u32 %r51, [%rd4+252]; mov.u64 %rd730, 0; setp.eq.s32 %p208, %r51, 0; mov.u64 %rd728, %rd730; mov.u64 %rd729, %rd730; @%p208 bra $L__BB193_250; ld.local.u32 %r285, [%rd4+248]; mul.wide.s32 %rd401, %r285, 80; add.s64 %rd402, %rd1, %rd401; add.s64 %rd728, %rd402, 304; add.s64 %rd403, %rd230, %rd401; add.s64 %rd729, %rd403, 304; mov.b32 %r286, %f304; mov.b32 %r287, %f303; mov.b32 %r288, %f302; mov.u32 %r289, 0; st.local.v4.u32 [%rd402+304], {%r288, %r287, %r286, %r289}; add.s32 %r290, %r285, 1; st.local.u32 [%rd4+248], %r290; add.s32 %r291, %r51, -1; st.local.u32 [%rd4+252], %r291; $L__BB193_250: setp.eq.s64 %p209, %rd729, 0; mov.u64 %rd731, %rd730; @%p209 bra $L__BB193_252; st.local.f32 [%rd728+16], %f305; mov.u64 %rd730, %rd728; mov.u64 %rd731, %rd729; $L__BB193_252: setp.eq.s64 %p210, %rd731, 0; mov.u64 %rd733, %rd732; @%p210 bra $L__BB193_256; ld.local.u32 %r52, [%rd4+252]; setp.lt.s32 %p211, %r52, 1; @%p211 bra $L__BB193_255; add.s32 %r292, %r52, -1; st.local.u32 [%rd4+252], %r292; ld.local.s32 %rd408, [%rd4+248]; cvt.s64.s32 %rd409, %r292; add.s64 %rd410, %rd408, %rd409; mul.lo.s64 %rd411, %rd410, 80; add.s64 %rd412, %rd1, %rd411; add.s64 %rd732, %rd412, 304; add.s64 %rd413, %rd230, %rd411; add.s64 %rd733, %rd413, 304; bra.uni $L__BB193_256; $L__BB193_229: mov.f32 %f1642, 0f00000000; max.ftz.f32 %f285, %f279, %f1642; max.ftz.f32 %f286, %f280, %f1642; add.ftz.f32 %f1643, %f285, %f286; max.ftz.f32 %f287, %f281, %f1642; add.ftz.f32 %f1644, %f287, %f1643; mul.ftz.f32 %f1645, %f1644, 0f3EAAAAAB; abs.ftz.f32 %f288, %f1645; setp.ltu.ftz.f32 %p195, %f288, 0f3727C5AC; mov.u64 %rd726, 0; mov.u64 %rd724, %rd726; mov.u64 %rd725, %rd726; @%p195 bra $L__BB193_234; ld.local.u32 %r49, [%rd4+252]; mov.u64 %rd724, 0; setp.eq.s32 %p196, %r49, 0; mov.u64 %rd722, %rd724; mov.u64 %rd723, %rd724; @%p196 bra $L__BB193_232; ld.local.u32 %r269, [%rd4+248]; mul.wide.s32 %rd375, %r269, 80; add.s64 %rd376, %rd1, %rd375; add.s64 %rd722, %rd376, 304; add.s64 %rd377, %rd230, %rd375; add.s64 %rd723, %rd377, 304; mov.b32 %r270, %f287; mov.b32 %r271, %f286; mov.b32 %r272, %f285; mov.u32 %r273, 0; st.local.v4.u32 [%rd376+304], {%r272, %r271, %r270, %r273}; add.s32 %r274, %r269, 1; st.local.u32 [%rd4+248], %r274; add.s32 %r275, %r49, -1; st.local.u32 [%rd4+252], %r275; $L__BB193_232: setp.eq.s64 %p197, %rd723, 0; mov.u64 %rd725, %rd724; @%p197 bra $L__BB193_234; st.local.f32 [%rd722+16], %f288; mov.u64 %rd724, %rd722; mov.u64 %rd725, %rd723; $L__BB193_234: setp.eq.s64 %p198, %rd725, 0; mov.u64 %rd727, %rd726; @%p198 bra $L__BB193_238; ld.local.u32 %r50, [%rd4+252]; setp.lt.s32 %p199, %r50, 1; @%p199 bra $L__BB193_237; add.s32 %r276, %r50, -1; st.local.u32 [%rd4+252], %r276; ld.local.s32 %rd382, [%rd4+248]; cvt.s64.s32 %rd383, %r276; add.s64 %rd384, %rd382, %rd383; mul.lo.s64 %rd385, %rd384, 80; add.s64 %rd386, %rd1, %rd385; add.s64 %rd726, %rd386, 304; add.s64 %rd387, %rd230, %rd385; add.s64 %rd727, %rd387, 304; bra.uni $L__BB193_238; $L__BB193_237: ld.local.u32 %r277, [%rd4+248]; add.s32 %r278, %r277, -1; st.local.u32 [%rd4+248], %r278; add.s32 %r279, %r50, 1; st.local.u32 [%rd4+252], %r279; mov.u64 %rd727, %rd726; $L__BB193_238: setp.eq.s64 %p200, %rd727, 0; or.pred %p202, %p198, %p200; @%p202 bra $L__BB193_272; add.s64 %rd73, %rd724, 20; st.local.f32 [%rd724+20], %f55; st.local.f32 [%rd724+24], %f56; st.local.f32 [%rd724+28], %f57; st.local.u64 [%rd724+48], %rd727; mov.u32 %r280, 0; st.local.v2.f32 [%rd724+56], {%f1642, %f1642}; st.local.u32 [%rd724+64], %r280; mul.ftz.f32 %f1648, %f71, %f71; st.local.f32 [%rd724+32], %f1648; st.local.f32 [%rd724+40], %f2980; st.local.f32 [%rd726], %f186; st.local.f32 [%rd726+4], %f187; st.local.f32 [%rd726+8], %f188; ld.local.u64 %rd390, [%rd724+48]; st.f32 [%rd390+12], %f282; st.f32 [%rd390+16], %f283; st.f32 [%rd390+20], %f284; ld.local.u64 %rd391, [%rd724+48]; st.u32 [%rd391+36], %r280; ld.local.f32 %f1649, [%rd724+32]; mov.f32 %f1650, 0f38D1B717; max.ftz.f32 %f1651, %f1649, %f1650; min.ftz.f32 %f1652, %f1651, %f1640; st.local.f32 [%rd724+32], %f1652; st.local.f32 [%rd724+36], %f1652; ld.local.f32 %f1653, [%rd724+40]; max.ftz.f32 %f1654, %f1642, %f1653; st.local.f32 [%rd724+40], %f1654; ld.local.u64 %rd392, [%rd724+48]; ld.f32 %f1655, [%rd392]; cvt.ftz.sat.f32.f32 %f1656, %f1655; ld.f32 %f1657, [%rd392+4]; cvt.ftz.sat.f32.f32 %f1658, %f1657; ld.f32 %f1659, [%rd392+8]; cvt.ftz.sat.f32.f32 %f1660, %f1659; st.f32 [%rd392], %f1656; st.f32 [%rd392+4], %f1658; st.f32 [%rd392+8], %f1660; ld.local.u64 %rd393, [%rd724+48]; ld.f32 %f1661, [%rd393+12]; cvt.ftz.sat.f32.f32 %f1662, %f1661; ld.f32 %f1663, [%rd393+16]; cvt.ftz.sat.f32.f32 %f1664, %f1663; ld.f32 %f1665, [%rd393+20]; cvt.ftz.sat.f32.f32 %f1666, %f1665; st.f32 [%rd393+12], %f1662; st.f32 [%rd393+16], %f1664; st.f32 [%rd393+20], %f1666; mov.u32 %r281, 27; st.local.u32 [%rd724+12], %r281; ld.local.f32 %f1667, [%rd724+40]; fma.rn.ftz.f32 %f289, %f1667, %f1667, 0fBF800000; abs.ftz.f32 %f290, %f1640; fma.rn.ftz.f32 %f291, %f290, %f290, %f289; setp.leu.ftz.f32 %p203, %f291, 0f00000000; mov.f32 %f3021, %f1640; @%p203 bra $L__BB193_241; sqrt.approx.ftz.f32 %f1668, %f291; sub.ftz.f32 %f1669, %f1668, %f290; add.ftz.f32 %f1670, %f290, %f1668; div.approx.ftz.f32 %f1671, %f1669, %f1670; fma.rn.ftz.f32 %f1672, %f290, %f1670, 0fBF800000; fma.rn.ftz.f32 %f1673, %f290, %f1669, 0f3F800000; div.approx.ftz.f32 %f1674, %f1672, %f1673; mul.ftz.f32 %f1675, %f1671, 0f3F000000; mul.ftz.f32 %f1676, %f1671, %f1675; fma.rn.ftz.f32 %f1677, %f1674, %f1674, 0f3F800000; mul.ftz.f32 %f3021, %f1676, %f1677; $L__BB193_241: ld.local.u64 %rd394, [%rd73+28]; add.s64 %rd74, %rd394, 12; ld.f32 %f294, [%rd394+12]; ld.f32 %f295, [%rd394+16]; ld.f32 %f296, [%rd394+20]; sub.ftz.f32 %f297, %f1640, %f3021; ld.local.f32 %f1679, [%rd73]; ld.local.f32 %f1680, [%rd4+24]; ld.local.f32 %f1681, [%rd73+4]; ld.local.f32 %f1682, [%rd4+28]; mul.ftz.f32 %f1683, %f1682, %f1681; fma.rn.ftz.f32 %f1684, %f1680, %f1679, %f1683; ld.local.f32 %f1685, [%rd73+8]; ld.local.f32 %f1686, [%rd4+32]; fma.rn.ftz.f32 %f1687, %f1686, %f1685, %f1684; abs.ftz.f32 %f298, %f1687; fma.rn.ftz.f32 %f299, %f298, %f298, %f289; setp.leu.ftz.f32 %p204, %f299, 0f00000000; mov.f32 %f3022, %f297; @%p204 bra $L__BB193_243; sqrt.approx.ftz.f32 %f1688, %f299; sub.ftz.f32 %f1689, %f1688, %f298; add.ftz.f32 %f1690, %f298, %f1688; div.approx.ftz.f32 %f1691, %f1689, %f1690; fma.rn.ftz.f32 %f1692, %f298, %f1690, 0fBF800000; fma.rn.ftz.f32 %f1693, %f298, %f1689, 0f3F800000; div.approx.ftz.f32 %f1694, %f1692, %f1693; mul.ftz.f32 %f1695, %f1691, 0f3F000000; mul.ftz.f32 %f1696, %f1691, %f1695; fma.rn.ftz.f32 %f1697, %f1694, %f1694, 0f3F800000; mul.ftz.f32 %f1698, %f1696, %f1697; sub.ftz.f32 %f3022, %f1698, %f3021; $L__BB193_243: rcp.approx.ftz.f32 %f1699, %f297; mov.f32 %f1700, 0f3F800000; mul.ftz.f32 %f1701, %f1699, %f3022; sub.ftz.f32 %f1702, %f1700, %f1701; fma.rn.ftz.f32 %f1703, %f294, %f1702, %f1701; fma.rn.ftz.f32 %f1704, %f295, %f1702, %f1701; fma.rn.ftz.f32 %f1705, %f296, %f1702, %f1701; st.f32 [%rd74+12], %f1703; st.f32 [%rd74+16], %f1704; st.f32 [%rd74+20], %f1705; ld.local.u32 %r282, [%rd73+-8]; setp.ne.s32 %p205, %r282, 12; @%p205 bra $L__BB193_245; ld.local.u64 %rd395, [%rd73+28]; ld.f32 %f1706, [%rd395+36]; mul.ftz.f32 %f1707, %f1706, 0f3E800000; ld.f32 %f1708, [%rd395+24]; mul.ftz.f32 %f1709, %f1708, %f1707; ld.f32 %f1710, [%rd395+28]; mul.ftz.f32 %f1711, %f1707, %f1710; ld.f32 %f1712, [%rd395+32]; mul.ftz.f32 %f1713, %f1707, %f1712; st.f32 [%rd395+24], %f1709; st.f32 [%rd395+28], %f1711; st.f32 [%rd395+32], %f1713; $L__BB193_245: ld.local.u64 %rd396, [%rd73+28]; ld.f32 %f1714, [%rd396+28]; ld.f32 %f1715, [%rd396+24]; add.ftz.f32 %f1716, %f1715, %f1714; ld.f32 %f1717, [%rd396+32]; add.ftz.f32 %f1718, %f1717, %f1716; mul.ftz.f32 %f1719, %f1718, 0f3EAAAAAB; ld.local.f32 %f1720, [%rd73+-4]; mul.ftz.f32 %f1721, %f1720, %f1719; st.local.f32 [%rd73+-4], %f1721; ld.local.u32 %r283, [%rd4+40]; or.b32 %r284, %r283, 1036; st.local.u32 [%rd4+40], %r284; bra.uni $L__BB193_272; $L__BB193_255: ld.local.u32 %r293, [%rd4+248]; add.s32 %r294, %r293, -1; st.local.u32 [%rd4+248], %r294; add.s32 %r295, %r52, 1; st.local.u32 [%rd4+252], %r295; mov.u64 %rd733, %rd732; $L__BB193_256: setp.eq.s64 %p212, %rd733, 0; or.pred %p214, %p210, %p212; @%p214 bra $L__BB193_264; add.s64 %rd85, %rd730, 20; st.local.f32 [%rd730+20], %f55; st.local.f32 [%rd730+24], %f56; st.local.f32 [%rd730+28], %f57; mov.u32 %r296, 0; st.local.v2.f32 [%rd730+56], {%f1723, %f1723}; st.local.u32 [%rd730+64], %r296; st.local.u64 [%rd730+48], %rd733; mul.ftz.f32 %f1731, %f71, %f71; st.local.v2.f32 [%rd730+32], {%f1731, %f1731}; st.local.f32 [%rd730+40], %f2980; st.local.f32 [%rd732], %f186; st.local.f32 [%rd732+4], %f187; st.local.f32 [%rd732+8], %f188; ld.local.u64 %rd416, [%rd730+48]; st.f32 [%rd416+12], %f282; st.f32 [%rd416+16], %f283; st.f32 [%rd416+20], %f284; ld.local.u64 %rd417, [%rd730+48]; st.u32 [%rd417+36], %r296; ld.local.u64 %rd418, [%rd730+48]; ld.f32 %f1732, [%rd418+12]; cvt.ftz.sat.f32.f32 %f1733, %f1732; ld.f32 %f1734, [%rd418+16]; cvt.ftz.sat.f32.f32 %f1735, %f1734; ld.f32 %f1736, [%rd418+20]; cvt.ftz.sat.f32.f32 %f1737, %f1736; st.f32 [%rd418+12], %f1733; st.f32 [%rd418+16], %f1735; st.f32 [%rd418+20], %f1737; ld.local.v2.f32 {%f1738, %f1739}, [%rd730+32]; cvt.ftz.sat.f32.f32 %f1742, %f1739; cvt.ftz.sat.f32.f32 %f1743, %f1738; st.local.v2.f32 [%rd730+32], {%f1743, %f1742}; mov.u32 %r297, 11; st.local.u32 [%rd730+12], %r297; ld.local.f32 %f1744, [%rd730+40]; fma.rn.ftz.f32 %f306, %f1744, %f1744, 0fBF800000; mov.f32 %f1729, 0f3F800000; abs.ftz.f32 %f307, %f1729; fma.rn.ftz.f32 %f308, %f307, %f307, %f306; setp.leu.ftz.f32 %p215, %f308, 0f00000000; mov.f32 %f3023, %f1729; @%p215 bra $L__BB193_259; sqrt.approx.ftz.f32 %f1745, %f308; sub.ftz.f32 %f1746, %f1745, %f307; add.ftz.f32 %f1747, %f307, %f1745; div.approx.ftz.f32 %f1748, %f1746, %f1747; fma.rn.ftz.f32 %f1749, %f307, %f1747, 0fBF800000; fma.rn.ftz.f32 %f1750, %f307, %f1746, 0f3F800000; div.approx.ftz.f32 %f1751, %f1749, %f1750; mul.ftz.f32 %f1752, %f1748, 0f3F000000; mul.ftz.f32 %f1753, %f1748, %f1752; fma.rn.ftz.f32 %f1754, %f1751, %f1751, 0f3F800000; mul.ftz.f32 %f3023, %f1753, %f1754; $L__BB193_259: ld.local.u64 %rd419, [%rd85+28]; add.s64 %rd86, %rd419, 12; ld.f32 %f311, [%rd419+12]; ld.f32 %f312, [%rd419+16]; ld.f32 %f313, [%rd419+20]; sub.ftz.f32 %f314, %f1729, %f3023; ld.local.f32 %f1756, [%rd85]; ld.local.f32 %f1757, [%rd4+24]; ld.local.f32 %f1758, [%rd85+4]; ld.local.f32 %f1759, [%rd4+28]; mul.ftz.f32 %f1760, %f1759, %f1758; fma.rn.ftz.f32 %f1761, %f1757, %f1756, %f1760; ld.local.f32 %f1762, [%rd85+8]; ld.local.f32 %f1763, [%rd4+32]; fma.rn.ftz.f32 %f1764, %f1763, %f1762, %f1761; abs.ftz.f32 %f315, %f1764; fma.rn.ftz.f32 %f316, %f315, %f315, %f306; setp.leu.ftz.f32 %p216, %f316, 0f00000000; mov.f32 %f3024, %f314; @%p216 bra $L__BB193_261; sqrt.approx.ftz.f32 %f1765, %f316; sub.ftz.f32 %f1766, %f1765, %f315; add.ftz.f32 %f1767, %f315, %f1765; div.approx.ftz.f32 %f1768, %f1766, %f1767; fma.rn.ftz.f32 %f1769, %f315, %f1767, 0fBF800000; fma.rn.ftz.f32 %f1770, %f315, %f1766, 0f3F800000; div.approx.ftz.f32 %f1771, %f1769, %f1770; mul.ftz.f32 %f1772, %f1768, 0f3F000000; mul.ftz.f32 %f1773, %f1768, %f1772; fma.rn.ftz.f32 %f1774, %f1771, %f1771, 0f3F800000; mul.ftz.f32 %f1775, %f1773, %f1774; sub.ftz.f32 %f3024, %f1775, %f3023; $L__BB193_261: rcp.approx.ftz.f32 %f1776, %f314; mov.f32 %f1777, 0f3F800000; mul.ftz.f32 %f1778, %f1776, %f3024; sub.ftz.f32 %f1779, %f1777, %f1778; fma.rn.ftz.f32 %f1780, %f311, %f1779, %f1778; fma.rn.ftz.f32 %f1781, %f312, %f1779, %f1778; fma.rn.ftz.f32 %f1782, %f313, %f1779, %f1778; st.f32 [%rd86+12], %f1780; st.f32 [%rd86+16], %f1781; st.f32 [%rd86+20], %f1782; ld.local.u32 %r298, [%rd85+-8]; setp.ne.s32 %p217, %r298, 12; @%p217 bra $L__BB193_263; ld.local.u64 %rd420, [%rd85+28]; ld.f32 %f1783, [%rd420+36]; mul.ftz.f32 %f1784, %f1783, 0f3E800000; ld.f32 %f1785, [%rd420+24]; mul.ftz.f32 %f1786, %f1785, %f1784; ld.f32 %f1787, [%rd420+28]; mul.ftz.f32 %f1788, %f1784, %f1787; ld.f32 %f1789, [%rd420+32]; mul.ftz.f32 %f1790, %f1784, %f1789; st.f32 [%rd420+24], %f1786; st.f32 [%rd420+28], %f1788; st.f32 [%rd420+32], %f1790; $L__BB193_263: ld.local.u64 %rd421, [%rd85+28]; ld.f32 %f1791, [%rd421+28]; ld.f32 %f1792, [%rd421+24]; add.ftz.f32 %f1793, %f1792, %f1791; ld.f32 %f1794, [%rd421+32]; add.ftz.f32 %f1795, %f1794, %f1793; mul.ftz.f32 %f1796, %f1795, 0f3EAAAAAB; ld.local.f32 %f1797, [%rd85+-4]; mul.ftz.f32 %f1798, %f1797, %f1796; st.local.f32 [%rd85+-4], %f1798; ld.local.u32 %r299, [%rd4+40]; or.b32 %r300, %r299, 12; st.local.u32 [%rd4+40], %r300; $L__BB193_264: setp.ne.s32 %p218, %r45, 0; setp.eq.s32 %p219, %r48, 0; and.pred %p220, %p218, %p219; @%p220 bra $L__BB193_272; mov.f32 %f1799, 0f3F800000; sub.ftz.f32 %f1800, %f1799, %f2981; mov.f32 %f1801, 0f38D1B717; max.ftz.f32 %f1802, %f1801, %f1800; mul.ftz.f32 %f1803, %f186, %f279; mul.ftz.f32 %f1804, %f1803, %f1802; mul.ftz.f32 %f1805, %f187, %f280; mul.ftz.f32 %f1806, %f1805, %f1802; mul.ftz.f32 %f1807, %f188, %f281; mul.ftz.f32 %f1808, %f1807, %f1802; mov.f32 %f1809, 0f00000000; max.ftz.f32 %f319, %f1804, %f1809; max.ftz.f32 %f320, %f1806, %f1809; max.ftz.f32 %f321, %f1808, %f1809; add.ftz.f32 %f1810, %f319, %f320; add.ftz.f32 %f1811, %f321, %f1810; mul.ftz.f32 %f1812, %f1811, 0f3EAAAAAB; abs.ftz.f32 %f322, %f1812; setp.ltu.ftz.f32 %p221, %f322, 0f3727C5AC; mov.u64 %rd736, 0; mov.u64 %rd737, %rd736; @%p221 bra $L__BB193_270; ld.local.u32 %r53, [%rd4+252]; mov.u64 %rd736, 0; setp.eq.s32 %p222, %r53, 0; mov.u64 %rd734, %rd736; mov.u64 %rd735, %rd736; @%p222 bra $L__BB193_268; ld.local.u32 %r301, [%rd4+248]; mul.wide.s32 %rd426, %r301, 80; add.s64 %rd427, %rd1, %rd426; add.s64 %rd734, %rd427, 304; add.s64 %rd428, %rd230, %rd426; add.s64 %rd735, %rd428, 304; mov.b32 %r302, %f321; mov.b32 %r303, %f320; mov.b32 %r304, %f319; mov.u32 %r305, 0; st.local.v4.u32 [%rd427+304], {%r304, %r303, %r302, %r305}; add.s32 %r306, %r301, 1; st.local.u32 [%rd4+248], %r306; add.s32 %r307, %r53, -1; st.local.u32 [%rd4+252], %r307; $L__BB193_268: setp.eq.s64 %p223, %rd735, 0; mov.u64 %rd737, %rd736; @%p223 bra $L__BB193_270; st.local.f32 [%rd734+16], %f322; mov.u64 %rd736, %rd734; mov.u64 %rd737, %rd735; $L__BB193_270: setp.eq.s64 %p224, %rd737, 0; @%p224 bra $L__BB193_272; mov.u64 %rd431, 0; st.local.f32 [%rd736+20], %f55; st.local.f32 [%rd736+24], %f56; st.local.f32 [%rd736+28], %f57; mov.u32 %r308, 0; st.local.v2.f32 [%rd736+56], {%f1809, %f1809}; st.local.u32 [%rd736+64], %r308; sub.ftz.f32 %f1815, %f1799, %f79; sub.ftz.f32 %f1816, %f1799, %f71; mul.ftz.f32 %f1817, %f1816, %f1815; sub.ftz.f32 %f1818, %f1799, %f1817; selp.f32 %f1819, %f1818, %f71, %p192; mul.ftz.f32 %f1820, %f1819, %f1819; st.local.f32 [%rd736+40], %f2980; st.local.u64 [%rd736+48], %rd431; cvt.ftz.sat.f32.f32 %f1821, %f1820; st.local.v2.f32 [%rd736+32], {%f1821, %f1821}; mov.u32 %r309, 23; st.local.u32 [%rd736+12], %r309; ld.local.u32 %r310, [%rd4+40]; or.b32 %r311, %r310, 12; st.local.u32 [%rd4+40], %r311; $L__BB193_272: ld.param.u32 %r638, [_Z21svm_node_closure_bsdfILj447EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji_param_4]; add.s32 %r663, %r638, 5; setp.leu.ftz.f32 %p226, %f76, 0f3727C5AC; or.pred %p228, %p226, %p158; @%p228 bra $L__BB193_517; mov.f32 %f1822, 0f00000000; max.ftz.f32 %f323, %f177, %f1822; max.ftz.f32 %f324, %f178, %f1822; add.ftz.f32 %f1823, %f323, %f324; max.ftz.f32 %f325, %f179, %f1822; add.ftz.f32 %f1824, %f325, %f1823; mul.ftz.f32 %f1825, %f1824, 0f3EAAAAAB; abs.ftz.f32 %f326, %f1825; setp.ltu.ftz.f32 %p229, %f326, 0f3727C5AC; mov.u64 %rd742, 0; mov.u64 %rd740, %rd742; mov.u64 %rd741, %rd742; @%p229 bra $L__BB193_278; ld.local.u32 %r55, [%rd4+252]; mov.u64 %rd740, 0; setp.eq.s32 %p230, %r55, 0; mov.u64 %rd738, %rd740; mov.u64 %rd739, %rd740; @%p230 bra $L__BB193_276; ld.local.u32 %r312, [%rd4+248]; mul.wide.s32 %rd436, %r312, 80; add.s64 %rd437, %rd1, %rd436; add.s64 %rd738, %rd437, 304; add.s64 %rd438, %rd230, %rd436; add.s64 %rd739, %rd438, 304; mov.b32 %r313, %f325; mov.b32 %r314, %f324; mov.b32 %r315, %f323; mov.u32 %r316, 0; st.local.v4.u32 [%rd437+304], {%r315, %r314, %r313, %r316}; add.s32 %r317, %r312, 1; st.local.u32 [%rd4+248], %r317; add.s32 %r318, %r55, -1; st.local.u32 [%rd4+252], %r318; $L__BB193_276: setp.eq.s64 %p231, %rd739, 0; mov.u64 %rd741, %rd740; @%p231 bra $L__BB193_278; st.local.f32 [%rd738+16], %f326; mov.u64 %rd740, %rd738; mov.u64 %rd741, %rd739; $L__BB193_278: setp.eq.s64 %p232, %rd741, 0; mov.u64 %rd743, %rd742; @%p232 bra $L__BB193_282; ld.local.u32 %r56, [%rd4+252]; setp.lt.s32 %p233, %r56, 1; @%p233 bra $L__BB193_281; add.s32 %r319, %r56, -1; st.local.u32 [%rd4+252], %r319; ld.local.s32 %rd443, [%rd4+248]; cvt.s64.s32 %rd444, %r319; add.s64 %rd445, %rd443, %rd444; mul.lo.s64 %rd446, %rd445, 80; add.s64 %rd447, %rd1, %rd446; add.s64 %rd742, %rd447, 304; add.s64 %rd448, %rd230, %rd446; add.s64 %rd743, %rd448, 304; bra.uni $L__BB193_282; $L__BB193_281: ld.local.u32 %r320, [%rd4+248]; add.s32 %r321, %r320, -1; st.local.u32 [%rd4+248], %r321; add.s32 %r322, %r56, 1; st.local.u32 [%rd4+252], %r322; mov.u64 %rd743, %rd742; $L__BB193_282: setp.eq.s64 %p234, %rd743, 0; or.pred %p236, %p232, %p234; @%p236 bra $L__BB193_517; mov.f32 %f3025, 0f3F800000; add.s64 %rd103, %rd740, 20; st.local.f32 [%rd740+20], %f2990; st.local.f32 [%rd740+24], %f2991; st.local.f32 [%rd740+28], %f2992; mov.u32 %r323, 0; st.local.v2.f32 [%rd740+56], {%f1822, %f1822}; st.local.u32 [%rd740+64], %r323; mov.u32 %r324, 1069547520; st.local.u32 [%rd740+40], %r324; st.local.u64 [%rd740+48], %rd743; mul.ftz.f32 %f1828, %f77, %f77; st.local.v2.f32 [%rd740+32], {%f1828, %f1828}; st.local.u32 [%rd742], %r323; st.local.u32 [%rd742+4], %r323; st.local.u32 [%rd742+8], %r323; ld.local.u64 %rd451, [%rd740+48]; mov.u32 %r325, 1025758986; st.u32 [%rd451+12], %r325; st.u32 [%rd451+16], %r325; st.u32 [%rd451+20], %r325; ld.local.u64 %rd452, [%rd740+48]; st.f32 [%rd452+36], %f76; ld.local.u64 %rd453, [%rd740+48]; ld.f32 %f1829, [%rd453+12]; cvt.ftz.sat.f32.f32 %f1830, %f1829; ld.f32 %f1831, [%rd453+16]; cvt.ftz.sat.f32.f32 %f1832, %f1831; ld.f32 %f1833, [%rd453+20]; cvt.ftz.sat.f32.f32 %f1834, %f1833; st.f32 [%rd453+12], %f1830; st.f32 [%rd453+16], %f1832; st.f32 [%rd453+20], %f1834; ld.local.f32 %f1835, [%rd740+32]; cvt.ftz.sat.f32.f32 %f1836, %f1835; st.local.v2.f32 [%rd740+32], {%f1836, %f1836}; mov.u32 %r326, 12; st.local.u32 [%rd740+12], %r326; ld.local.f32 %f1837, [%rd740+40]; fma.rn.ftz.f32 %f327, %f1837, %f1837, 0fBF800000; abs.ftz.f32 %f328, %f3025; fma.rn.ftz.f32 %f329, %f328, %f328, %f327; setp.leu.ftz.f32 %p237, %f329, 0f00000000; @%p237 bra $L__BB193_285; sqrt.approx.ftz.f32 %f1838, %f329; sub.ftz.f32 %f1839, %f1838, %f328; add.ftz.f32 %f1840, %f328, %f1838; div.approx.ftz.f32 %f1841, %f1839, %f1840; fma.rn.ftz.f32 %f1842, %f328, %f1840, 0fBF800000; fma.rn.ftz.f32 %f1843, %f328, %f1839, 0f3F800000; div.approx.ftz.f32 %f1844, %f1842, %f1843; mul.ftz.f32 %f1845, %f1841, 0f3F000000; mul.ftz.f32 %f1846, %f1841, %f1845; fma.rn.ftz.f32 %f1847, %f1844, %f1844, 0f3F800000; mul.ftz.f32 %f3025, %f1846, %f1847; $L__BB193_285: mov.f32 %f2960, 0f3F800000; ld.local.u64 %rd454, [%rd103+28]; add.s64 %rd104, %rd454, 12; ld.f32 %f332, [%rd454+12]; ld.f32 %f333, [%rd454+16]; ld.f32 %f334, [%rd454+20]; sub.ftz.f32 %f335, %f2960, %f3025; ld.local.f32 %f1849, [%rd103]; ld.local.f32 %f1850, [%rd4+24]; ld.local.f32 %f1851, [%rd103+4]; ld.local.f32 %f1852, [%rd4+28]; mul.ftz.f32 %f1853, %f1852, %f1851; fma.rn.ftz.f32 %f1854, %f1850, %f1849, %f1853; ld.local.f32 %f1855, [%rd103+8]; ld.local.f32 %f1856, [%rd4+32]; fma.rn.ftz.f32 %f1857, %f1856, %f1855, %f1854; abs.ftz.f32 %f336, %f1857; fma.rn.ftz.f32 %f337, %f336, %f336, %f327; setp.leu.ftz.f32 %p238, %f337, 0f00000000; mov.f32 %f3026, %f335; @%p238 bra $L__BB193_287; sqrt.approx.ftz.f32 %f1858, %f337; sub.ftz.f32 %f1859, %f1858, %f336; add.ftz.f32 %f1860, %f336, %f1858; div.approx.ftz.f32 %f1861, %f1859, %f1860; fma.rn.ftz.f32 %f1862, %f336, %f1860, 0fBF800000; fma.rn.ftz.f32 %f1863, %f336, %f1859, 0f3F800000; div.approx.ftz.f32 %f1864, %f1862, %f1863; mul.ftz.f32 %f1865, %f1861, 0f3F000000; mul.ftz.f32 %f1866, %f1861, %f1865; fma.rn.ftz.f32 %f1867, %f1864, %f1864, 0f3F800000; mul.ftz.f32 %f1868, %f1866, %f1867; sub.ftz.f32 %f3026, %f1868, %f3025; $L__BB193_287: rcp.approx.ftz.f32 %f1869, %f335; mov.f32 %f1870, 0f3F800000; mul.ftz.f32 %f1871, %f1869, %f3026; sub.ftz.f32 %f1872, %f1870, %f1871; fma.rn.ftz.f32 %f1873, %f332, %f1872, %f1871; fma.rn.ftz.f32 %f1874, %f333, %f1872, %f1871; fma.rn.ftz.f32 %f1875, %f334, %f1872, %f1871; st.f32 [%rd104+12], %f1873; st.f32 [%rd104+16], %f1874; st.f32 [%rd104+20], %f1875; ld.local.u32 %r327, [%rd103+-8]; setp.ne.s32 %p239, %r327, 12; @%p239 bra $L__BB193_289; ld.local.u64 %rd455, [%rd103+28]; ld.f32 %f1876, [%rd455+36]; mul.ftz.f32 %f1877, %f1876, 0f3E800000; ld.f32 %f1878, [%rd455+24]; mul.ftz.f32 %f1879, %f1878, %f1877; ld.f32 %f1880, [%rd455+28]; mul.ftz.f32 %f1881, %f1877, %f1880; ld.f32 %f1882, [%rd455+32]; mul.ftz.f32 %f1883, %f1877, %f1882; st.f32 [%rd455+24], %f1879; st.f32 [%rd455+28], %f1881; st.f32 [%rd455+32], %f1883; $L__BB193_289: ld.local.u64 %rd456, [%rd103+28]; ld.f32 %f1884, [%rd456+28]; ld.f32 %f1885, [%rd456+24]; add.ftz.f32 %f1886, %f1885, %f1884; ld.f32 %f1887, [%rd456+32]; add.ftz.f32 %f1888, %f1887, %f1886; mul.ftz.f32 %f1889, %f1888, 0f3EAAAAAB; ld.local.f32 %f1890, [%rd103+-4]; mul.ftz.f32 %f1891, %f1890, %f1889; st.local.f32 [%rd103+-4], %f1891; ld.local.u32 %r328, [%rd4+40]; or.b32 %r329, %r328, 12; st.local.u32 [%rd4+40], %r329; bra.uni $L__BB193_517; } .func _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<121>; .reg .b16 %rs<41>; .reg .f32 %f<735>; .reg .b32 %r<254>; .reg .b64 %rd<282>; ld.param.u64 %rd66, [_Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd67, [_Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r69, [_Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r68, [_Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r67, [_Z13svm_node_attrILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd68, %rd66; add.s64 %rd1, %rd68, 76; ld.local.s32 %rd3, [%rd68+76]; setp.eq.s64 %p1, %rd3, -1; @%p1 bra $L__BB194_16; cvt.u64.u32 %rd4, %r67; ld.local.s32 %rd5, [%rd1+-16]; ld.const.u64 %rd69, [kernel_params+2160]; cvta.to.global.u64 %rd70, %rd69; mul.lo.s64 %rd71, %rd3, 208; add.s64 %rd72, %rd70, %rd71; ld.global.u32 %r3, [%rd72+164]; ld.local.u32 %r30, [%rd1+-12]; and.b32 %r71, %r30, 1; setp.eq.b32 %p2, %r71, 1; mov.pred %p3, 0; xor.pred %p4, %p2, %p3; not.pred %p5, %p4; mov.u32 %r235, 0; @%p5 bra $L__BB194_5; cvt.u32.u64 %r73, %rd5; setp.eq.s32 %p6, %r73, -1; mov.u32 %r234, -1; @%p6 bra $L__BB194_4; ld.const.u64 %rd73, [kernel_params+2240]; cvta.to.global.u64 %rd74, %rd73; shl.b64 %rd75, %rd5, 2; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r234, [%rd76]; $L__BB194_4: setp.ne.s32 %p7, %r234, -1; selp.u32 %r235, 1, 0, %p7; $L__BB194_5: ld.const.u64 %rd77, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd77; add.s32 %r237, %r235, %r3; mul.wide.u32 %rd78, %r237, 16; add.s64 %rd79, %rd6, %rd78; add.s64 %rd268, %rd79, 14; add.s64 %rd269, %rd79, 15; ld.global.u32 %r236, [%rd79+8]; ld.global.u16 %rs39, [%rd79+12]; ld.global.u64 %rd267, [%rd79]; setp.eq.s64 %p8, %rd267, %rd4; @%p8 bra $L__BB194_11; $L__BB194_6: setp.eq.s64 %p9, %rd267, 0; @%p9 bra $L__BB194_8; add.s32 %r237, %r237, 2; bra.uni $L__BB194_9; $L__BB194_8: setp.eq.s16 %p10, %rs39, 0; mov.u32 %r243, -1; mov.u32 %r240, 0; mov.u32 %r237, %r236; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p10 bra $L__BB194_15; $L__BB194_9: cvt.u64.u32 %rd11, %r237; mul.wide.u32 %rd80, %r237, 16; add.s64 %rd81, %rd6, %rd80; ld.global.u32 %r236, [%rd81+8]; ld.global.u16 %rs39, [%rd81+12]; ld.global.u64 %rd267, [%rd81]; setp.ne.s64 %p11, %rd267, %rd4; @%p11 bra $L__BB194_6; shl.b64 %rd82, %rd11, 4; add.s64 %rd83, %rd6, %rd82; add.s64 %rd268, %rd83, 14; add.s64 %rd269, %rd83, 15; $L__BB194_11: ld.global.u8 %rs5, [%rd269]; ld.global.u8 %rs6, [%rd268]; cvt.u32.u64 %r78, %rd5; setp.ne.s32 %p12, %r78, -1; @%p12 bra $L__BB194_14; mov.u32 %r243, -1; mov.u32 %r240, 0; add.s16 %rs19, %rs39, -1; setp.lt.u16 %p13, %rs19, 2; @%p13 bra $L__BB194_14; setp.ne.s16 %p14, %rs39, 1024; mov.u32 %r241, %r240; mov.u32 %r242, %r240; @%p14 bra $L__BB194_15; $L__BB194_14: setp.eq.s16 %p15, %rs39, 0; selp.b32 %r243, -1, %r236, %p15; cvt.u32.u16 %r240, %rs39; cvt.u32.u16 %r83, %rs5; and.b32 %r242, %r83, 255; cvt.u32.u16 %r84, %rs6; and.b32 %r241, %r84, 255; $L__BB194_15: setp.eq.s32 %p16, %r243, -1; selp.b32 %r31, 0, %r243, %p16; selp.b32 %r32, 0, %r242, %p16; selp.b32 %r247, %r69, %r241, %p16; selp.b32 %r34, 0, %r240, %p16; bra.uni $L__BB194_17; $L__BB194_16: ld.local.u32 %r30, [%rd1+-12]; mov.u32 %r31, 0; mov.u32 %r32, %r31; mov.u32 %r247, %r69; mov.u32 %r34, %r31; $L__BB194_17: cvta.to.local.u64 %rd84, %rd67; mul.wide.u32 %rd85, %r68, 4; add.s64 %rd17, %rd84, %rd85; setp.eq.s32 %p17, %r30, 16; @%p17 bra $L__BB194_159; bra.uni $L__BB194_18; $L__BB194_159: and.b32 %r227, %r34, 3; setp.eq.s32 %p102, %r227, 0; @%p102 bra $L__BB194_161; ld.const.u64 %rd253, [kernel_params+2344]; cvta.to.global.u64 %rd254, %rd253; mul.wide.s32 %rd255, %r31, 16; add.s64 %rd256, %rd254, %rd255; ld.global.v4.f32 {%f728, %f729, %f730, %f731}, [%rd256]; bra.uni $L__BB194_185; $L__BB194_18: setp.eq.s32 %p18, %r67, 7; setp.eq.s32 %p19, %r34, 0; and.pred %p20, %p18, %p19; @%p20 bra $L__BB194_150; bra.uni $L__BB194_19; $L__BB194_150: cvt.u32.u64 %r224, %rd3; setp.eq.s32 %p98, %r224, -1; ld.local.v4.f32 {%f721, %f720, %f719, %f496}, [%rd68]; @%p98 bra $L__BB194_154; ld.local.u8 %rs35, [%rd1+-20]; and.b16 %rs36, %rs35, 2; setp.eq.s16 %p99, %rs36, 0; @%p99 bra $L__BB194_153; bra.uni $L__BB194_152; $L__BB194_153: ld.const.u64 %rd249, [kernel_params+2160]; cvta.to.global.u64 %rd250, %rd249; mul.wide.s32 %rd251, %r224, 208; add.s64 %rd252, %rd250, %rd251; ld.global.v4.f32 {%f530, %f531, %f532, %f533}, [%rd252+48]; ld.global.v4.f32 {%f538, %f539, %f540, %f541}, [%rd252+64]; ld.global.v4.f32 {%f546, %f547, %f548, %f549}, [%rd252+80]; mul.ftz.f32 %f554, %f720, %f531; fma.rn.ftz.f32 %f555, %f721, %f530, %f554; fma.rn.ftz.f32 %f556, %f719, %f532, %f555; add.ftz.f32 %f173, %f533, %f556; mul.ftz.f32 %f557, %f720, %f539; fma.rn.ftz.f32 %f558, %f721, %f538, %f557; fma.rn.ftz.f32 %f559, %f719, %f540, %f558; add.ftz.f32 %f174, %f541, %f559; mul.ftz.f32 %f560, %f720, %f547; fma.rn.ftz.f32 %f561, %f721, %f546, %f560; fma.rn.ftz.f32 %f562, %f719, %f548, %f561; add.ftz.f32 %f719, %f549, %f562; mov.f32 %f720, %f174; mov.f32 %f721, %f173; bra.uni $L__BB194_154; $L__BB194_19: setp.eq.s32 %p21, %r247, 0; and.b32 %r39, %r30, 1; @%p21 bra $L__BB194_125; setp.eq.s32 %p22, %r247, 1; @%p22 bra $L__BB194_100; bra.uni $L__BB194_21; $L__BB194_100: setp.eq.s32 %p66, %r39, 0; mov.u32 %r49, 1; @%p66 bra $L__BB194_111; bra.uni $L__BB194_101; $L__BB194_111: and.b32 %r176, %r30, 6; setp.eq.s32 %p73, %r176, 0; @%p73 bra $L__BB194_118; and.b32 %r177, %r34, 768; setp.eq.s32 %p74, %r177, 0; @%p74 bra $L__BB194_114; ld.const.u64 %rd194, [kernel_params+2264]; cvta.to.global.u64 %rd195, %rd194; ld.local.u32 %r178, [%rd1+-16]; mul.wide.s32 %rd196, %r178, 16; add.s64 %rd197, %rd195, %rd196; shr.s32 %r179, %r30, 7; ld.global.u32 %r180, [%rd197+4]; add.s32 %r181, %r179, %r180; ld.const.u64 %rd198, [kernel_params+2328]; cvta.to.global.u64 %rd199, %rd198; add.s32 %r182, %r181, %r31; mul.wide.s32 %rd200, %r182, 8; add.s64 %rd201, %rd199, %rd200; ld.global.v2.f32 {%f453, %f454}, [%rd201]; ld.global.v2.f32 {%f457, %f458}, [%rd201+8]; ld.local.f32 %f461, [%rd1+-8]; mov.f32 %f462, 0f3F800000; sub.ftz.f32 %f463, %f462, %f461; mul.ftz.f32 %f464, %f453, %f463; mul.ftz.f32 %f465, %f454, %f463; fma.rn.ftz.f32 %f717, %f458, %f461, %f465; fma.rn.ftz.f32 %f716, %f457, %f461, %f464; bra.uni $L__BB194_120; $L__BB194_161: setp.ne.s32 %p103, %r34, 1024; mov.f32 %f728, 0f00000000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; mov.f32 %f731, %f728; @%p103 bra $L__BB194_185; ld.local.f32 %f183, [%rd1+-76]; ld.local.f32 %f184, [%rd1+-72]; ld.local.f32 %f185, [%rd1+-68]; ld.local.u8 %rs37, [%rd1+-20]; and.b16 %rs38, %rs37, 2; setp.eq.s16 %p104, %rs38, 0; @%p104 bra $L__BB194_164; bra.uni $L__BB194_163; $L__BB194_164: ld.const.u64 %rd257, [kernel_params+2160]; cvta.to.global.u64 %rd258, %rd257; mul.lo.s64 %rd259, %rd3, 208; add.s64 %rd260, %rd258, %rd259; ld.global.v4.f32 {%f607, %f608, %f609, %f610}, [%rd260+48]; ld.global.v4.f32 {%f615, %f616, %f617, %f618}, [%rd260+64]; ld.global.v4.f32 {%f623, %f624, %f625, %f626}, [%rd260+80]; mul.ftz.f32 %f631, %f184, %f608; fma.rn.ftz.f32 %f632, %f183, %f607, %f631; fma.rn.ftz.f32 %f633, %f185, %f609, %f632; add.ftz.f32 %f198, %f610, %f633; mul.ftz.f32 %f634, %f184, %f616; fma.rn.ftz.f32 %f635, %f183, %f615, %f634; fma.rn.ftz.f32 %f636, %f185, %f617, %f635; add.ftz.f32 %f199, %f618, %f636; mul.ftz.f32 %f637, %f184, %f624; fma.rn.ftz.f32 %f638, %f183, %f623, %f637; fma.rn.ftz.f32 %f639, %f185, %f625, %f638; add.ftz.f32 %f200, %f626, %f639; bra.uni $L__BB194_165; $L__BB194_125: setp.eq.s32 %p82, %r39, 0; @%p82 bra $L__BB194_136; bra.uni $L__BB194_126; $L__BB194_136: and.b32 %r211, %r30, 6; setp.eq.s32 %p89, %r211, 0; @%p89 bra $L__BB194_143; and.b32 %r212, %r34, 768; setp.eq.s32 %p90, %r212, 0; @%p90 bra $L__BB194_139; ld.const.u64 %rd232, [kernel_params+2264]; cvta.to.global.u64 %rd233, %rd232; ld.local.u32 %r213, [%rd1+-16]; mul.wide.s32 %rd234, %r213, 16; add.s64 %rd235, %rd233, %rd234; shr.s32 %r214, %r30, 7; ld.global.u32 %r215, [%rd235+4]; add.s32 %r216, %r214, %r215; ld.const.u64 %rd236, [kernel_params+2320]; cvta.to.global.u64 %rd237, %rd236; add.s32 %r217, %r216, %r31; mul.wide.s32 %rd238, %r217, 4; add.s64 %rd239, %rd237, %rd238; ld.local.f32 %f485, [%rd1+-8]; mov.f32 %f486, 0f3F800000; sub.ftz.f32 %f487, %f486, %f485; ld.global.f32 %f488, [%rd239]; mul.ftz.f32 %f489, %f488, %f487; ld.global.f32 %f490, [%rd239+4]; fma.rn.ftz.f32 %f718, %f490, %f485, %f489; bra.uni $L__BB194_145; $L__BB194_152: ld.local.v4.f32 {%f497, %f498, %f499, %f500}, [%rd1+116]; mul.ftz.f32 %f505, %f720, %f498; fma.rn.ftz.f32 %f506, %f721, %f497, %f505; fma.rn.ftz.f32 %f507, %f719, %f499, %f506; add.ftz.f32 %f170, %f500, %f507; ld.local.v4.f32 {%f508, %f509, %f510, %f511}, [%rd1+132]; mul.ftz.f32 %f516, %f720, %f509; fma.rn.ftz.f32 %f517, %f721, %f508, %f516; fma.rn.ftz.f32 %f518, %f719, %f510, %f517; add.ftz.f32 %f171, %f511, %f518; ld.local.v4.f32 {%f519, %f520, %f521, %f522}, [%rd1+148]; mul.ftz.f32 %f527, %f720, %f520; fma.rn.ftz.f32 %f528, %f721, %f519, %f527; fma.rn.ftz.f32 %f529, %f719, %f521, %f528; add.ftz.f32 %f719, %f522, %f529; mov.f32 %f720, %f171; mov.f32 %f721, %f170; $L__BB194_154: setp.eq.s32 %p100, %r69, 1; @%p100 bra $L__BB194_158; bra.uni $L__BB194_155; $L__BB194_158: add.ftz.f32 %f563, %f720, %f721; add.ftz.f32 %f564, %f719, %f563; mul.ftz.f32 %f565, %f564, 0f3EAAAAAB; st.local.f32 [%rd17], %f565; bra.uni $L__BB194_192; $L__BB194_155: setp.eq.s32 %p101, %r69, 0; @%p101 bra $L__BB194_157; mov.u32 %r226, 1065353216; st.local.u32 [%rd17], %r226; bra.uni $L__BB194_192; $L__BB194_21: add.s32 %r91, %r247, -3; setp.lt.u32 %p23, %r91, 2; @%p23 bra $L__BB194_47; bra.uni $L__BB194_22; $L__BB194_47: setp.eq.s32 %p40, %r39, 0; @%p40 bra $L__BB194_86; bra.uni $L__BB194_48; $L__BB194_86: and.b32 %r145, %r30, 6; setp.eq.s32 %p57, %r145, 0; @%p57 bra $L__BB194_93; and.b32 %r146, %r34, 768; setp.eq.s32 %p58, %r146, 0; @%p58 bra $L__BB194_89; ld.const.u64 %rd153, [kernel_params+2264]; cvta.to.global.u64 %rd154, %rd153; ld.local.u32 %r147, [%rd1+-16]; mul.wide.s32 %rd155, %r147, 16; add.s64 %rd156, %rd154, %rd155; shr.s32 %r148, %r30, 7; ld.global.u32 %r149, [%rd156+4]; add.s32 %r150, %r148, %r149; ld.const.u64 %rd157, [kernel_params+2344]; cvta.to.global.u64 %rd158, %rd157; add.s32 %r151, %r150, %r31; mul.wide.s32 %rd159, %r151, 16; add.s64 %rd160, %rd158, %rd159; ld.global.v4.f32 {%f386, %f387, %f388, %f389}, [%rd160]; ld.global.v4.f32 {%f394, %f395, %f396, %f397}, [%rd160+16]; ld.local.f32 %f402, [%rd1+-8]; mov.f32 %f403, 0f3F800000; sub.ftz.f32 %f404, %f403, %f402; mul.ftz.f32 %f405, %f386, %f404; mul.ftz.f32 %f406, %f387, %f404; mul.ftz.f32 %f407, %f388, %f404; mul.ftz.f32 %f408, %f389, %f404; fma.rn.ftz.f32 %f709, %f397, %f402, %f408; fma.rn.ftz.f32 %f708, %f396, %f402, %f407; fma.rn.ftz.f32 %f707, %f395, %f402, %f406; fma.rn.ftz.f32 %f706, %f394, %f402, %f405; bra.uni $L__BB194_95; $L__BB194_163: ld.local.v4.f32 {%f574, %f575, %f576, %f577}, [%rd1+116]; mul.ftz.f32 %f582, %f184, %f575; fma.rn.ftz.f32 %f583, %f183, %f574, %f582; fma.rn.ftz.f32 %f584, %f185, %f576, %f583; add.ftz.f32 %f198, %f577, %f584; ld.local.v4.f32 {%f585, %f586, %f587, %f588}, [%rd1+132]; mul.ftz.f32 %f593, %f184, %f586; fma.rn.ftz.f32 %f594, %f183, %f585, %f593; fma.rn.ftz.f32 %f595, %f185, %f587, %f594; add.ftz.f32 %f199, %f588, %f595; ld.local.v4.f32 {%f596, %f597, %f598, %f599}, [%rd1+148]; mul.ftz.f32 %f604, %f184, %f597; fma.rn.ftz.f32 %f605, %f183, %f596, %f604; fma.rn.ftz.f32 %f606, %f185, %f598, %f605; add.ftz.f32 %f200, %f599, %f606; $L__BB194_165: ld.local.u32 %r60, [%rd1+-24]; ld.const.u64 %rd261, [kernel_params+2432]; cvta.to.global.u64 %rd262, %rd261; mul.wide.s32 %rd263, %r31, 96; add.s64 %rd64, %rd261, %rd263; add.s64 %rd264, %rd262, %rd263; add.s64 %rd65, %rd264, 32; ld.global.u32 %r228, [%rd264+32]; setp.eq.s32 %p105, %r228, 0; @%p105 bra $L__BB194_167; ld.global.v4.f32 {%f640, %f641, %f642, %f643}, [%rd65+16]; mul.ftz.f32 %f648, %f199, %f641; fma.rn.ftz.f32 %f649, %f198, %f640, %f648; fma.rn.ftz.f32 %f650, %f200, %f642, %f649; add.ftz.f32 %f195, %f643, %f650; ld.global.v4.f32 {%f651, %f652, %f653, %f654}, [%rd65+32]; mul.ftz.f32 %f659, %f199, %f652; fma.rn.ftz.f32 %f660, %f198, %f651, %f659; fma.rn.ftz.f32 %f661, %f200, %f653, %f660; add.ftz.f32 %f196, %f654, %f661; ld.global.v4.f32 {%f662, %f663, %f664, %f665}, [%rd65+48]; mul.ftz.f32 %f670, %f199, %f663; fma.rn.ftz.f32 %f671, %f198, %f662, %f670; fma.rn.ftz.f32 %f672, %f200, %f664, %f671; add.ftz.f32 %f200, %f665, %f672; mov.f32 %f198, %f195; mov.f32 %f199, %f196; $L__BB194_167: and.b32 %r229, %r60, 16777216; setp.ne.s32 %p106, %r229, 0; setp.eq.s32 %p107, %r229, 0; selp.b32 %r63, -1, 2, %p107; @%p106 bra $L__BB194_169; ld.global.u32 %r63, [%rd65+-20]; $L__BB194_169: ld.global.u32 %r64, [%rd65+-24]; setp.eq.s32 %p108, %r64, 8; @%p108 bra $L__BB194_184; bra.uni $L__BB194_170; $L__BB194_184: { // callseq 965, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 965 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB194_185; $L__BB194_126: ld.local.s32 %rd54, [%rd1+-16]; setp.eq.s64 %p83, %rd54, -1; @%p83 bra $L__BB194_129; ld.const.u64 %rd210, [kernel_params+2240]; cvta.to.global.u64 %rd211, %rd210; shl.b64 %rd212, %rd54, 2; add.s64 %rd213, %rd211, %rd212; ld.global.u32 %r190, [%rd213]; setp.eq.s32 %p84, %r190, -1; @%p84 bra $L__BB194_129; mov.u32 %r191, 0; mov.u64 %rd214, 0; { // callseq 959, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd214; .param .b64 param3; st.param.b64 [param3+0], %rd214; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f718, [retval0+0]; } // callseq 959 bra.uni $L__BB194_145; $L__BB194_157: st.local.f32 [%rd17], %f721; st.local.f32 [%rd17+4], %f720; st.local.f32 [%rd17+8], %f719; bra.uni $L__BB194_192; $L__BB194_170: setp.eq.s32 %p109, %r64, 9; @%p109 bra $L__BB194_183; bra.uni $L__BB194_171; $L__BB194_183: { // callseq 964, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; } // callseq 964 mov.f32 %f731, 0f3F800000; bra.uni $L__BB194_185; $L__BB194_101: ld.local.s32 %rd52, [%rd1+-16]; setp.eq.s64 %p67, %rd52, -1; @%p67 bra $L__BB194_104; ld.const.u64 %rd169, [kernel_params+2240]; cvta.to.global.u64 %rd170, %rd169; shl.b64 %rd171, %rd52, 2; add.s64 %rd172, %rd170, %rd171; ld.global.u32 %r157, [%rd172]; setp.eq.s32 %p68, %r157, -1; @%p68 bra $L__BB194_104; mov.u64 %rd173, 0; { // callseq 958, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r49; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd173; .param .b64 param3; st.param.b64 [param3+0], %rd173; .param .align 4 .b8 retval0[8]; call.uni (retval0), _Z30subd_triangle_attribute_float2PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float2S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f716, [retval0+0]; ld.param.f32 %f717, [retval0+4]; } // callseq 958 bra.uni $L__BB194_120; $L__BB194_22: setp.eq.s32 %p24, %r39, 0; @%p24 bra $L__BB194_33; bra.uni $L__BB194_23; $L__BB194_33: and.b32 %r113, %r30, 6; setp.eq.s32 %p31, %r113, 0; @%p31 bra $L__BB194_40; and.b32 %r114, %r34, 768; setp.eq.s32 %p32, %r114, 0; @%p32 bra $L__BB194_36; ld.const.u64 %rd109, [kernel_params+2264]; cvta.to.global.u64 %rd110, %rd109; ld.local.u32 %r115, [%rd1+-16]; mul.wide.s32 %rd111, %r115, 16; add.s64 %rd112, %rd110, %rd111; shr.s32 %r116, %r30, 7; ld.global.u32 %r117, [%rd112+4]; add.s32 %r118, %r116, %r117; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r119, %r118, %r31; mul.wide.s32 %rd115, %r119, 12; add.s64 %rd116, %rd114, %rd115; ld.local.f32 %f272, [%rd1+-8]; mov.f32 %f273, 0f3F800000; sub.ftz.f32 %f274, %f273, %f272; ld.global.f32 %f275, [%rd116]; mul.ftz.f32 %f276, %f275, %f274; ld.global.f32 %f277, [%rd116+4]; mul.ftz.f32 %f278, %f277, %f274; ld.global.f32 %f279, [%rd116+8]; mul.ftz.f32 %f280, %f279, %f274; ld.global.f32 %f281, [%rd116+12]; ld.global.f32 %f282, [%rd116+16]; ld.global.f32 %f283, [%rd116+20]; fma.rn.ftz.f32 %f685, %f283, %f272, %f280; fma.rn.ftz.f32 %f684, %f282, %f272, %f278; fma.rn.ftz.f32 %f683, %f281, %f272, %f276; bra.uni $L__BB194_42; $L__BB194_129: and.b32 %r192, %r34, 56; setp.eq.s32 %p85, %r192, 0; @%p85 bra $L__BB194_134; and.b32 %r193, %r34, 24; setp.eq.s32 %p86, %r193, 0; @%p86 bra $L__BB194_132; ld.const.u64 %rd215, [kernel_params+2232]; cvta.to.global.u64 %rd216, %rd215; shl.b64 %rd217, %rd54, 4; add.s64 %rd218, %rd216, %rd217; ld.global.v4.u32 {%r194, %r195, %r196, %r197}, [%rd218]; ld.const.u64 %rd219, [kernel_params+2320]; cvta.to.global.u64 %rd220, %rd219; add.s32 %r201, %r194, %r31; mul.wide.u32 %rd221, %r201, 4; add.s64 %rd279, %rd220, %rd221; add.s32 %r202, %r195, %r31; mul.wide.u32 %rd222, %r202, 4; add.s64 %rd280, %rd220, %rd222; add.s32 %r203, %r196, %r31; mul.wide.u32 %rd223, %r203, 4; add.s64 %rd281, %rd220, %rd223; bra.uni $L__BB194_133; $L__BB194_104: and.b32 %r158, %r34, 56; setp.eq.s32 %p69, %r158, 0; @%p69 bra $L__BB194_109; and.b32 %r159, %r34, 24; setp.eq.s32 %p70, %r159, 0; @%p70 bra $L__BB194_107; ld.const.u64 %rd174, [kernel_params+2232]; cvta.to.global.u64 %rd175, %rd174; shl.b64 %rd176, %rd52, 4; add.s64 %rd177, %rd175, %rd176; ld.global.v4.u32 {%r160, %r161, %r162, %r163}, [%rd177]; ld.const.u64 %rd178, [kernel_params+2328]; cvta.to.global.u64 %rd179, %rd178; add.s32 %r167, %r160, %r31; mul.wide.u32 %rd180, %r167, 8; add.s64 %rd181, %rd179, %rd180; ld.global.v2.f32 {%f714, %f715}, [%rd181]; add.s32 %r168, %r161, %r31; mul.wide.u32 %rd182, %r168, 8; add.s64 %rd183, %rd179, %rd182; ld.global.v2.f32 {%f712, %f713}, [%rd183]; add.s32 %r169, %r162, %r31; mul.wide.u32 %rd184, %r169, 8; add.s64 %rd185, %rd179, %rd184; ld.global.v2.f32 {%f710, %f711}, [%rd185]; bra.uni $L__BB194_108; $L__BB194_171: setp.eq.s32 %p110, %r64, 10; @%p110 bra $L__BB194_182; bra.uni $L__BB194_172; $L__BB194_182: { // callseq 963, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 963 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB194_185; $L__BB194_48: ld.local.s32 %rd48, [%rd1+-16]; setp.eq.s64 %p41, %rd48, -1; @%p41 bra $L__BB194_51; ld.const.u64 %rd125, [kernel_params+2240]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd48, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r126, [%rd128]; setp.eq.s32 %p42, %r126, -1; @%p42 bra $L__BB194_51; mov.u64 %rd129, 0; { // callseq 957, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd129; .param .b64 param3; st.param.b64 [param3+0], %rd129; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z30subd_triangle_attribute_float4PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float4S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f706, [retval0+0]; ld.param.f32 %f707, [retval0+4]; ld.param.f32 %f708, [retval0+8]; ld.param.f32 %f709, [retval0+12]; } // callseq 957 bra.uni $L__BB194_95; $L__BB194_23: ld.local.s32 %rd19, [%rd1+-16]; setp.eq.s64 %p25, %rd19, -1; @%p25 bra $L__BB194_26; ld.const.u64 %rd86, [kernel_params+2240]; cvta.to.global.u64 %rd87, %rd86; shl.b64 %rd88, %rd19, 2; add.s64 %rd89, %rd87, %rd88; ld.global.u32 %r92, [%rd89]; setp.eq.s32 %p26, %r92, -1; @%p26 bra $L__BB194_26; mov.u64 %rd90, 0; { // callseq 956, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd66; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r34; st.param.b32 [param1+4], %r247; st.param.b32 [param1+8], %r32; st.param.b32 [param1+12], %r31; .param .b64 param2; st.param.b64 [param2+0], %rd90; .param .b64 param3; st.param.b64 [param3+0], %rd90; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f683, [retval0+0]; ld.param.f32 %f684, [retval0+4]; ld.param.f32 %f685, [retval0+8]; } // callseq 956 bra.uni $L__BB194_42; $L__BB194_143: and.b32 %r220, %r30, 8; setp.eq.s32 %p93, %r220, 0; setp.ne.s32 %p94, %r34, 8; mov.f32 %f718, 0f00000000; or.pred %p95, %p94, %p93; @%p95 bra $L__BB194_145; ld.const.u64 %rd244, [kernel_params+2320]; cvta.to.global.u64 %rd245, %rd244; ld.local.u32 %r221, [%rd1+-16]; add.s32 %r222, %r221, %r31; mul.wide.s32 %rd246, %r222, 4; add.s64 %rd247, %rd245, %rd246; ld.global.f32 %f718, [%rd247]; bra.uni $L__BB194_145; $L__BB194_118: and.b32 %r185, %r30, 8; setp.eq.s32 %p77, %r185, 0; setp.ne.s32 %p78, %r34, 8; mov.f32 %f716, 0f00000000; or.pred %p79, %p78, %p77; mov.f32 %f717, %f716; @%p79 bra $L__BB194_120; ld.const.u64 %rd206, [kernel_params+2328]; cvta.to.global.u64 %rd207, %rd206; ld.local.u32 %r186, [%rd1+-16]; add.s32 %r187, %r186, %r31; mul.wide.s32 %rd208, %r187, 8; add.s64 %rd209, %rd207, %rd208; ld.global.v2.f32 {%f716, %f717}, [%rd209]; bra.uni $L__BB194_120; $L__BB194_51: and.b32 %r127, %r34, 120; setp.eq.s32 %p43, %r127, 0; @%p43 bra $L__BB194_84; and.b32 %r128, %r34, 24; setp.eq.s32 %p44, %r128, 0; @%p44 bra $L__BB194_54; ld.const.u64 %rd130, [kernel_params+2232]; cvta.to.global.u64 %rd131, %rd130; shl.b64 %rd132, %rd48, 4; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r129, %r130, %r131, %r132}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2344]; cvta.to.global.u64 %rd135, %rd134; add.s32 %r136, %r129, %r31; mul.wide.u32 %rd136, %r136, 16; add.s64 %rd137, %rd135, %rd136; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd137]; add.s32 %r137, %r130, %r31; mul.wide.u32 %rd138, %r137, 16; add.s64 %rd139, %rd135, %rd138; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd139]; add.s32 %r138, %r131, %r31; mul.wide.u32 %rd140, %r138, 16; add.s64 %rd141, %rd135, %rd140; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd141]; bra.uni $L__BB194_83; $L__BB194_26: and.b32 %r93, %r34, 56; setp.eq.s32 %p27, %r93, 0; @%p27 bra $L__BB194_31; and.b32 %r94, %r34, 24; setp.eq.s32 %p28, %r94, 0; @%p28 bra $L__BB194_29; ld.const.u64 %rd91, [kernel_params+2232]; cvta.to.global.u64 %rd92, %rd91; shl.b64 %rd93, %rd19, 4; add.s64 %rd94, %rd92, %rd93; ld.global.v4.u32 {%r95, %r96, %r97, %r98}, [%rd94]; ld.const.u64 %rd95, [kernel_params+2336]; cvta.to.global.u64 %rd96, %rd95; add.s32 %r102, %r95, %r31; mul.wide.u32 %rd97, %r102, 12; add.s64 %rd276, %rd96, %rd97; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r103, %r96, %r31; mul.wide.u32 %rd98, %r103, 12; add.s64 %rd273, %rd96, %rd98; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r104, %r97, %r31; mul.wide.u32 %rd99, %r104, 12; add.s64 %rd270, %rd96, %rd99; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; bra.uni $L__BB194_30; $L__BB194_172: setp.eq.s32 %p111, %r64, 11; @%p111 bra $L__BB194_181; bra.uni $L__BB194_173; $L__BB194_181: { // callseq 962, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f728, [retval0+0]; } // callseq 962 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB194_185; $L__BB194_139: and.b32 %r218, %r34, 131; setp.eq.s32 %p91, %r218, 0; mov.f32 %f718, 0f00000000; @%p91 bra $L__BB194_145; setp.ne.s32 %p92, %r34, 128; @%p92 bra $L__BB194_142; ld.local.u32 %r219, [%rd1+-16]; add.s32 %r31, %r219, %r31; $L__BB194_142: ld.const.u64 %rd240, [kernel_params+2320]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r31, 4; add.s64 %rd243, %rd241, %rd242; ld.global.f32 %f718, [%rd243]; bra.uni $L__BB194_145; $L__BB194_134: and.b32 %r207, %r34, 7; setp.eq.s32 %p87, %r207, 0; mov.f32 %f718, 0f00000000; @%p87 bra $L__BB194_145; cvt.u32.u64 %r208, %rd54; setp.eq.s32 %p88, %r34, 4; selp.b32 %r209, %r208, 0, %p88; add.s32 %r210, %r31, %r209; ld.const.u64 %rd228, [kernel_params+2320]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r210, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f718, [%rd231]; bra.uni $L__BB194_145; $L__BB194_114: and.b32 %r183, %r34, 131; setp.eq.s32 %p75, %r183, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p75 bra $L__BB194_120; setp.ne.s32 %p76, %r34, 128; @%p76 bra $L__BB194_117; ld.local.u32 %r184, [%rd1+-16]; add.s32 %r31, %r184, %r31; $L__BB194_117: ld.const.u64 %rd202, [kernel_params+2328]; cvta.to.global.u64 %rd203, %rd202; mul.wide.s32 %rd204, %r31, 8; add.s64 %rd205, %rd203, %rd204; ld.global.v2.f32 {%f716, %f717}, [%rd205]; bra.uni $L__BB194_120; $L__BB194_109: and.b32 %r172, %r34, 7; setp.eq.s32 %p71, %r172, 0; mov.f32 %f716, 0f00000000; mov.f32 %f717, %f716; @%p71 bra $L__BB194_120; cvt.u32.u64 %r173, %rd52; setp.eq.s32 %p72, %r34, 4; selp.b32 %r174, %r173, 0, %p72; add.s32 %r175, %r31, %r174; ld.const.u64 %rd190, [kernel_params+2328]; cvta.to.global.u64 %rd191, %rd190; mul.wide.s32 %rd192, %r175, 8; add.s64 %rd193, %rd191, %rd192; ld.global.v2.f32 {%f716, %f717}, [%rd193]; bra.uni $L__BB194_120; $L__BB194_93: and.b32 %r154, %r30, 8; setp.eq.s32 %p61, %r154, 0; setp.ne.s32 %p62, %r34, 8; mov.f32 %f706, 0f00000000; or.pred %p63, %p62, %p61; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p63 bra $L__BB194_95; ld.const.u64 %rd165, [kernel_params+2344]; cvta.to.global.u64 %rd166, %rd165; ld.local.u32 %r155, [%rd1+-16]; add.s32 %r156, %r155, %r31; mul.wide.s32 %rd167, %r156, 16; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd168]; bra.uni $L__BB194_95; $L__BB194_40: and.b32 %r122, %r30, 8; setp.eq.s32 %p35, %r122, 0; setp.ne.s32 %p36, %r34, 8; mov.f32 %f683, 0f00000000; or.pred %p37, %p36, %p35; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p37 bra $L__BB194_42; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; ld.local.u32 %r123, [%rd1+-16]; add.s32 %r124, %r123, %r31; mul.wide.s32 %rd123, %r124, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f685, [%rd124+8]; ld.global.f32 %f684, [%rd124+4]; ld.global.f32 %f683, [%rd124]; bra.uni $L__BB194_42; $L__BB194_132: cvt.u32.u64 %r204, %rd54; mad.lo.s32 %r205, %r204, 3, %r31; ld.const.u64 %rd224, [kernel_params+2320]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r205, 4; add.s64 %rd279, %rd225, %rd226; add.s64 %rd280, %rd279, 4; add.s32 %r206, %r205, 2; mul.wide.s32 %rd227, %r206, 4; add.s64 %rd281, %rd225, %rd227; $L__BB194_133: ld.local.f32 %f474, [%rd1+-8]; ld.global.f32 %f475, [%rd280]; ld.local.f32 %f476, [%rd1+-4]; ld.global.f32 %f477, [%rd281]; mul.ftz.f32 %f478, %f477, %f476; fma.rn.ftz.f32 %f479, %f475, %f474, %f478; mov.f32 %f480, 0f3F800000; sub.ftz.f32 %f481, %f480, %f474; sub.ftz.f32 %f482, %f481, %f476; ld.global.f32 %f483, [%rd279]; fma.rn.ftz.f32 %f718, %f483, %f482, %f479; $L__BB194_145: setp.eq.s32 %p96, %r69, 1; @%p96 bra $L__BB194_149; bra.uni $L__BB194_146; $L__BB194_149: st.local.f32 [%rd17], %f718; bra.uni $L__BB194_192; $L__BB194_146: setp.eq.s32 %p97, %r69, 0; @%p97 bra $L__BB194_148; mov.u32 %r223, 1065353216; st.local.u32 [%rd17], %r223; bra.uni $L__BB194_192; $L__BB194_148: st.local.f32 [%rd17], %f718; st.local.f32 [%rd17+4], %f718; st.local.f32 [%rd17+8], %f718; bra.uni $L__BB194_192; $L__BB194_173: or.b32 %r65, %r63, 1; setp.gt.u32 %p112, %r64, 6; @%p112 bra $L__BB194_178; mov.u32 %r231, 1; shl.b32 %r232, %r231, %r64; and.b32 %r233, %r232, 71; setp.ne.s32 %p113, %r233, 0; @%p113 bra $L__BB194_175; bra.uni $L__BB194_178; $L__BB194_175: setp.eq.s32 %p114, %r65, 3; @%p114 bra $L__BB194_177; bra.uni $L__BB194_176; $L__BB194_177: { // callseq 960, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; ld.param.f32 %f729, [retval0+4]; ld.param.f32 %f730, [retval0+8]; ld.param.f32 %f731, [retval0+12]; } // callseq 960 bra.uni $L__BB194_185; $L__BB194_107: cvt.u32.u64 %r170, %rd52; mad.lo.s32 %r171, %r170, 3, %r31; ld.const.u64 %rd186, [kernel_params+2328]; cvta.to.global.u64 %rd187, %rd186; mul.wide.s32 %rd188, %r171, 8; add.s64 %rd189, %rd187, %rd188; ld.global.v2.f32 {%f714, %f715}, [%rd189]; ld.global.v2.f32 {%f712, %f713}, [%rd189+8]; ld.global.v2.f32 {%f710, %f711}, [%rd189+16]; $L__BB194_108: ld.local.f32 %f440, [%rd1+-8]; ld.local.f32 %f441, [%rd1+-4]; mul.ftz.f32 %f442, %f710, %f441; mul.ftz.f32 %f443, %f711, %f441; fma.rn.ftz.f32 %f444, %f712, %f440, %f442; fma.rn.ftz.f32 %f445, %f713, %f440, %f443; mov.f32 %f446, 0f3F800000; sub.ftz.f32 %f447, %f446, %f440; sub.ftz.f32 %f448, %f447, %f441; fma.rn.ftz.f32 %f717, %f715, %f448, %f445; fma.rn.ftz.f32 %f716, %f714, %f448, %f444; $L__BB194_120: setp.eq.s32 %p80, %r69, 1; @%p80 bra $L__BB194_124; bra.uni $L__BB194_121; $L__BB194_124: st.local.f32 [%rd17], %f716; bra.uni $L__BB194_192; $L__BB194_121: setp.eq.s32 %p81, %r69, 0; @%p81 bra $L__BB194_123; mov.u32 %r188, 1065353216; st.local.u32 [%rd17], %r188; bra.uni $L__BB194_192; $L__BB194_123: st.local.f32 [%rd17], %f716; st.local.f32 [%rd17+4], %f717; mov.u32 %r189, 0; st.local.u32 [%rd17+8], %r189; bra.uni $L__BB194_192; $L__BB194_89: and.b32 %r152, %r34, 131; setp.eq.s32 %p59, %r152, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p59 bra $L__BB194_95; setp.ne.s32 %p60, %r34, 128; @%p60 bra $L__BB194_92; ld.local.u32 %r153, [%rd1+-16]; add.s32 %r31, %r153, %r31; $L__BB194_92: ld.const.u64 %rd161, [kernel_params+2344]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r31, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd164]; bra.uni $L__BB194_95; $L__BB194_36: and.b32 %r120, %r34, 131; setp.eq.s32 %p33, %r120, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p33 bra $L__BB194_42; setp.ne.s32 %p34, %r34, 128; @%p34 bra $L__BB194_39; ld.local.u32 %r121, [%rd1+-16]; add.s32 %r31, %r121, %r31; $L__BB194_39: ld.const.u64 %rd117, [kernel_params+2336]; cvta.to.global.u64 %rd118, %rd117; mul.wide.s32 %rd119, %r31, 12; add.s64 %rd120, %rd118, %rd119; ld.global.f32 %f685, [%rd120+8]; ld.global.f32 %f684, [%rd120+4]; ld.global.f32 %f683, [%rd120]; bra.uni $L__BB194_42; $L__BB194_84: and.b32 %r141, %r34, 7; setp.eq.s32 %p55, %r141, 0; mov.f32 %f706, 0f00000000; mov.f32 %f707, %f706; mov.f32 %f708, %f706; mov.f32 %f709, %f706; @%p55 bra $L__BB194_95; cvt.u32.u64 %r142, %rd48; setp.eq.s32 %p56, %r34, 4; selp.b32 %r143, %r142, 0, %p56; add.s32 %r144, %r31, %r143; ld.const.u64 %rd149, [kernel_params+2344]; cvta.to.global.u64 %rd150, %rd149; mul.wide.s32 %rd151, %r144, 16; add.s64 %rd152, %rd150, %rd151; ld.global.v4.f32 {%f706, %f707, %f708, %f709}, [%rd152]; bra.uni $L__BB194_95; $L__BB194_31: and.b32 %r109, %r34, 7; setp.eq.s32 %p29, %r109, 0; mov.f32 %f683, 0f00000000; mov.f32 %f684, %f683; mov.f32 %f685, %f683; @%p29 bra $L__BB194_42; cvt.u32.u64 %r110, %rd19; setp.eq.s32 %p30, %r34, 4; selp.b32 %r111, %r110, 0, %p30; add.s32 %r112, %r31, %r111; ld.const.u64 %rd105, [kernel_params+2336]; cvta.to.global.u64 %rd106, %rd105; mul.wide.s32 %rd107, %r112, 12; add.s64 %rd108, %rd106, %rd107; ld.global.f32 %f685, [%rd108+8]; ld.global.f32 %f684, [%rd108+4]; ld.global.f32 %f683, [%rd108]; bra.uni $L__BB194_42; $L__BB194_54: cvt.u32.u64 %r139, %rd48; mad.lo.s32 %r140, %r139, 3, %r31; cvt.s64.s32 %rd49, %r140; setp.eq.s32 %p45, %r34, 32; @%p45 bra $L__BB194_82; bra.uni $L__BB194_55; $L__BB194_82: ld.const.u64 %rd145, [kernel_params+2344]; cvta.to.global.u64 %rd146, %rd145; shl.b64 %rd147, %rd49, 4; add.s64 %rd148, %rd146, %rd147; ld.global.v4.f32 {%f702, %f703, %f704, %f705}, [%rd148]; ld.global.v4.f32 {%f698, %f699, %f700, %f701}, [%rd148+16]; ld.global.v4.f32 {%f694, %f695, %f696, %f697}, [%rd148+32]; bra.uni $L__BB194_83; $L__BB194_29: cvt.u32.u64 %r105, %rd19; mad.lo.s32 %r106, %r105, 3, %r31; ld.const.u64 %rd100, [kernel_params+2336]; cvta.to.global.u64 %rd101, %rd100; mul.wide.s32 %rd102, %r106, 12; add.s64 %rd276, %rd101, %rd102; add.s64 %rd277, %rd276, 4; add.s64 %rd278, %rd276, 8; add.s32 %r107, %r106, 1; mul.wide.s32 %rd103, %r107, 12; add.s64 %rd273, %rd101, %rd103; add.s64 %rd274, %rd273, 4; add.s64 %rd275, %rd273, 8; add.s32 %r108, %r106, 2; mul.wide.s32 %rd104, %r108, 12; add.s64 %rd270, %rd101, %rd104; add.s64 %rd271, %rd270, 4; add.s64 %rd272, %rd270, 8; $L__BB194_30: ld.local.f32 %f249, [%rd1+-8]; ld.global.f32 %f250, [%rd273]; ld.global.f32 %f251, [%rd274]; ld.global.f32 %f252, [%rd275]; ld.local.f32 %f253, [%rd1+-4]; ld.global.f32 %f254, [%rd270]; mul.ftz.f32 %f255, %f254, %f253; ld.global.f32 %f256, [%rd271]; mul.ftz.f32 %f257, %f256, %f253; ld.global.f32 %f258, [%rd272]; mul.ftz.f32 %f259, %f258, %f253; fma.rn.ftz.f32 %f260, %f250, %f249, %f255; fma.rn.ftz.f32 %f261, %f251, %f249, %f257; fma.rn.ftz.f32 %f262, %f252, %f249, %f259; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f249; sub.ftz.f32 %f265, %f264, %f253; ld.global.f32 %f266, [%rd276]; ld.global.f32 %f267, [%rd277]; ld.global.f32 %f268, [%rd278]; fma.rn.ftz.f32 %f685, %f268, %f265, %f262; fma.rn.ftz.f32 %f684, %f267, %f265, %f261; fma.rn.ftz.f32 %f683, %f266, %f265, %f260; $L__BB194_42: setp.eq.s32 %p38, %r69, 1; @%p38 bra $L__BB194_46; bra.uni $L__BB194_43; $L__BB194_46: add.ftz.f32 %f290, %f683, %f684; add.ftz.f32 %f291, %f685, %f290; mul.ftz.f32 %f292, %f291, 0f3EAAAAAB; st.local.f32 [%rd17], %f292; bra.uni $L__BB194_192; $L__BB194_43: setp.eq.s32 %p39, %r69, 0; @%p39 bra $L__BB194_45; mov.u32 %r125, 1065353216; st.local.u32 [%rd17], %r125; bra.uni $L__BB194_192; $L__BB194_45: st.local.f32 [%rd17], %f683; st.local.f32 [%rd17+4], %f684; st.local.f32 [%rd17+8], %f685; bra.uni $L__BB194_192; $L__BB194_176: ld.global.u64 %rd265, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f729, %f730, %f731}, [%rd265, {%f198, %f199, %f200, %f200}]; bra.uni $L__BB194_185; $L__BB194_55: ld.const.u64 %rd142, [kernel_params+2352]; cvta.to.global.u64 %rd143, %rd142; shl.b64 %rd144, %rd49, 2; add.s64 %rd50, %rd143, %rd144; ld.global.v4.u8 {%rs20, %rs21, %rs22, %rs23}, [%rd50]; cvt.rn.f32.u16 %f305, %rs20; mul.ftz.f32 %f38, %f305, 0f3B808081; cvt.rn.f32.u16 %f306, %rs21; mul.ftz.f32 %f39, %f306, 0f3B808081; cvt.rn.f32.u16 %f307, %rs22; mul.ftz.f32 %f40, %f307, 0f3B808081; setp.lt.ftz.f32 %p46, %f38, 0f3D25AEE6; @%p46 bra $L__BB194_57; bra.uni $L__BB194_56; $L__BB194_57: mul.ftz.f32 %f702, %f38, 0f3D9E8391; bra.uni $L__BB194_58; $L__BB194_178: setp.eq.s32 %p115, %r65, 3; @%p115 bra $L__BB194_180; bra.uni $L__BB194_179; $L__BB194_180: { // callseq 961, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd64; .param .b32 param1; st.param.f32 [param1+0], %f198; .param .b32 param2; st.param.f32 [param2+0], %f199; .param .b32 param3; st.param.f32 [param3+0], %f200; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f728, [retval0+0]; } // callseq 961 mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; bra.uni $L__BB194_185; $L__BB194_56: add.ftz.f32 %f308, %f38, 0f3D6147AE; mul.ftz.f32 %f309, %f308, 0f3F72A76F; lg2.approx.ftz.f32 %f310, %f309; mul.ftz.f32 %f311, %f310, 0f4019999A; ex2.approx.ftz.f32 %f702, %f311; $L__BB194_58: setp.lt.ftz.f32 %p47, %f39, 0f3D25AEE6; @%p47 bra $L__BB194_60; bra.uni $L__BB194_59; $L__BB194_60: mul.ftz.f32 %f703, %f39, 0f3D9E8391; bra.uni $L__BB194_61; $L__BB194_59: add.ftz.f32 %f312, %f39, 0f3D6147AE; mul.ftz.f32 %f313, %f312, 0f3F72A76F; lg2.approx.ftz.f32 %f314, %f313; mul.ftz.f32 %f315, %f314, 0f4019999A; ex2.approx.ftz.f32 %f703, %f315; $L__BB194_61: setp.lt.ftz.f32 %p48, %f40, 0f3D25AEE6; @%p48 bra $L__BB194_63; bra.uni $L__BB194_62; $L__BB194_63: mul.ftz.f32 %f704, %f40, 0f3D9E8391; bra.uni $L__BB194_64; $L__BB194_62: add.ftz.f32 %f316, %f40, 0f3D6147AE; mul.ftz.f32 %f317, %f316, 0f3F72A76F; lg2.approx.ftz.f32 %f318, %f317; mul.ftz.f32 %f319, %f318, 0f4019999A; ex2.approx.ftz.f32 %f704, %f319; $L__BB194_64: ld.global.v4.u8 {%rs24, %rs25, %rs26, %rs27}, [%rd50+4]; cvt.rn.f32.u16 %f320, %rs24; mul.ftz.f32 %f50, %f320, 0f3B808081; cvt.rn.f32.u16 %f321, %rs25; mul.ftz.f32 %f51, %f321, 0f3B808081; cvt.rn.f32.u16 %f322, %rs26; mul.ftz.f32 %f52, %f322, 0f3B808081; setp.lt.ftz.f32 %p49, %f50, 0f3D25AEE6; @%p49 bra $L__BB194_66; bra.uni $L__BB194_65; $L__BB194_66: mul.ftz.f32 %f698, %f50, 0f3D9E8391; bra.uni $L__BB194_67; $L__BB194_65: add.ftz.f32 %f323, %f50, 0f3D6147AE; mul.ftz.f32 %f324, %f323, 0f3F72A76F; lg2.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f325, 0f4019999A; ex2.approx.ftz.f32 %f698, %f326; $L__BB194_67: setp.lt.ftz.f32 %p50, %f51, 0f3D25AEE6; @%p50 bra $L__BB194_69; bra.uni $L__BB194_68; $L__BB194_69: mul.ftz.f32 %f699, %f51, 0f3D9E8391; bra.uni $L__BB194_70; $L__BB194_68: add.ftz.f32 %f327, %f51, 0f3D6147AE; mul.ftz.f32 %f328, %f327, 0f3F72A76F; lg2.approx.ftz.f32 %f329, %f328; mul.ftz.f32 %f330, %f329, 0f4019999A; ex2.approx.ftz.f32 %f699, %f330; $L__BB194_70: setp.lt.ftz.f32 %p51, %f52, 0f3D25AEE6; @%p51 bra $L__BB194_72; bra.uni $L__BB194_71; $L__BB194_72: mul.ftz.f32 %f700, %f52, 0f3D9E8391; bra.uni $L__BB194_73; $L__BB194_71: add.ftz.f32 %f331, %f52, 0f3D6147AE; mul.ftz.f32 %f332, %f331, 0f3F72A76F; lg2.approx.ftz.f32 %f333, %f332; mul.ftz.f32 %f334, %f333, 0f4019999A; ex2.approx.ftz.f32 %f700, %f334; $L__BB194_73: ld.global.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd50+8]; cvt.rn.f32.u16 %f335, %rs28; mul.ftz.f32 %f62, %f335, 0f3B808081; cvt.rn.f32.u16 %f336, %rs29; mul.ftz.f32 %f63, %f336, 0f3B808081; cvt.rn.f32.u16 %f337, %rs30; mul.ftz.f32 %f64, %f337, 0f3B808081; setp.lt.ftz.f32 %p52, %f62, 0f3D25AEE6; @%p52 bra $L__BB194_75; bra.uni $L__BB194_74; $L__BB194_75: mul.ftz.f32 %f694, %f62, 0f3D9E8391; bra.uni $L__BB194_76; $L__BB194_74: add.ftz.f32 %f338, %f62, 0f3D6147AE; mul.ftz.f32 %f339, %f338, 0f3F72A76F; lg2.approx.ftz.f32 %f340, %f339; mul.ftz.f32 %f341, %f340, 0f4019999A; ex2.approx.ftz.f32 %f694, %f341; $L__BB194_76: setp.lt.ftz.f32 %p53, %f63, 0f3D25AEE6; @%p53 bra $L__BB194_78; bra.uni $L__BB194_77; $L__BB194_78: mul.ftz.f32 %f695, %f63, 0f3D9E8391; bra.uni $L__BB194_79; $L__BB194_77: add.ftz.f32 %f342, %f63, 0f3D6147AE; mul.ftz.f32 %f343, %f342, 0f3F72A76F; lg2.approx.ftz.f32 %f344, %f343; mul.ftz.f32 %f345, %f344, 0f4019999A; ex2.approx.ftz.f32 %f695, %f345; $L__BB194_79: cvt.rn.f32.u16 %f346, %rs23; mul.ftz.f32 %f705, %f346, 0f3B808081; cvt.rn.f32.u16 %f347, %rs27; mul.ftz.f32 %f701, %f347, 0f3B808081; cvt.rn.f32.u16 %f348, %rs31; mul.ftz.f32 %f697, %f348, 0f3B808081; setp.lt.ftz.f32 %p54, %f64, 0f3D25AEE6; @%p54 bra $L__BB194_81; bra.uni $L__BB194_80; $L__BB194_81: mul.ftz.f32 %f696, %f64, 0f3D9E8391; bra.uni $L__BB194_83; $L__BB194_80: add.ftz.f32 %f349, %f64, 0f3D6147AE; mul.ftz.f32 %f350, %f349, 0f3F72A76F; lg2.approx.ftz.f32 %f351, %f350; mul.ftz.f32 %f352, %f351, 0f4019999A; ex2.approx.ftz.f32 %f696, %f352; $L__BB194_83: ld.local.f32 %f365, [%rd1+-8]; ld.local.f32 %f366, [%rd1+-4]; mul.ftz.f32 %f367, %f694, %f366; mul.ftz.f32 %f368, %f695, %f366; mul.ftz.f32 %f369, %f696, %f366; mul.ftz.f32 %f370, %f697, %f366; fma.rn.ftz.f32 %f371, %f698, %f365, %f367; fma.rn.ftz.f32 %f372, %f699, %f365, %f368; fma.rn.ftz.f32 %f373, %f700, %f365, %f369; fma.rn.ftz.f32 %f374, %f701, %f365, %f370; mov.f32 %f375, 0f3F800000; sub.ftz.f32 %f376, %f375, %f365; sub.ftz.f32 %f377, %f376, %f366; fma.rn.ftz.f32 %f709, %f705, %f377, %f374; fma.rn.ftz.f32 %f708, %f704, %f377, %f373; fma.rn.ftz.f32 %f707, %f703, %f377, %f372; fma.rn.ftz.f32 %f706, %f702, %f377, %f371; $L__BB194_95: setp.eq.s32 %p64, %r69, 1; @%p64 bra $L__BB194_99; bra.uni $L__BB194_96; $L__BB194_99: add.ftz.f32 %f425, %f706, %f707; add.ftz.f32 %f426, %f708, %f425; mul.ftz.f32 %f427, %f426, 0f3EAAAAAB; st.local.f32 [%rd17], %f427; bra.uni $L__BB194_192; $L__BB194_96: setp.eq.s32 %p65, %r69, 0; @%p65 bra $L__BB194_98; st.local.f32 [%rd17], %f709; bra.uni $L__BB194_192; $L__BB194_98: st.local.f32 [%rd17], %f706; st.local.f32 [%rd17+4], %f707; st.local.f32 [%rd17+8], %f708; bra.uni $L__BB194_192; $L__BB194_179: ld.global.u64 %rd266, [%rd65+-32]; tex.3d.v4.f32.f32 {%f728, %f673, %f674, %f675}, [%rd266, {%f198, %f199, %f200, %f200}]; mov.f32 %f731, 0f3F800000; mov.f32 %f729, %f728; mov.f32 %f730, %f728; $L__BB194_185: setp.eq.s32 %p116, %r69, 1; @%p116 bra $L__BB194_191; bra.uni $L__BB194_186; $L__BB194_191: add.ftz.f32 %f676, %f728, %f729; add.ftz.f32 %f677, %f730, %f676; mul.ftz.f32 %f678, %f677, 0f3EAAAAAB; st.local.f32 [%rd17], %f678; bra.uni $L__BB194_192; $L__BB194_186: setp.eq.s32 %p117, %r69, 0; @%p117 bra $L__BB194_188; st.local.f32 [%rd17], %f731; bra.uni $L__BB194_192; $L__BB194_188: setp.leu.ftz.f32 %p118, %f731, 0f358637BD; setp.eq.ftz.f32 %p119, %f731, 0f3F800000; or.pred %p120, %p118, %p119; @%p120 bra $L__BB194_190; div.approx.ftz.f32 %f730, %f730, %f731; div.approx.ftz.f32 %f729, %f729, %f731; div.approx.ftz.f32 %f728, %f728, %f731; $L__BB194_190: st.local.f32 [%rd17], %f728; st.local.f32 [%rd17+4], %f729; st.local.f32 [%rd17+8], %f730; $L__BB194_192: ret; } .func _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<9>; .reg .f32 %f<212>; .reg .b32 %r<16>; .reg .b64 %rd<33>; ld.param.u64 %rd5, [_Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd6, [_Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r6, [_Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r5, [_Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r4, [_Z21svm_node_displacementILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd7, %rd5; shr.u32 %r1, %r4, 24; cvta.to.local.u64 %rd1, %rd6; shl.b32 %r7, %r4, 2; cvt.u64.u32 %rd8, %r7; and.b64 %rd9, %rd8, 1020; add.s64 %rd10, %rd1, %rd9; ld.local.f32 %f1, [%rd10]; shr.u32 %r8, %r4, 6; cvt.u64.u32 %rd11, %r8; and.b64 %rd12, %rd11, 1020; add.s64 %rd13, %rd1, %rd12; ld.local.f32 %f2, [%rd13]; shr.u32 %r9, %r4, 14; cvt.u64.u32 %rd14, %r9; and.b64 %rd15, %rd14, 1020; add.s64 %rd16, %rd1, %rd15; ld.local.f32 %f3, [%rd16]; setp.eq.s32 %p1, %r1, 255; add.s64 %rd2, %rd7, 12; @%p1 bra $L__BB195_2; bra.uni $L__BB195_1; $L__BB195_2: ld.local.f32 %f206, [%rd2+8]; ld.local.f32 %f207, [%rd2+4]; ld.local.f32 %f208, [%rd2]; bra.uni $L__BB195_3; $L__BB195_1: mul.wide.u32 %rd17, %r1, 4; add.s64 %rd18, %rd1, %rd17; ld.local.f32 %f206, [%rd18+8]; ld.local.f32 %f207, [%rd18+4]; ld.local.f32 %f208, [%rd18]; $L__BB195_3: setp.eq.s32 %p2, %r6, 1; @%p2 bra $L__BB195_5; bra.uni $L__BB195_4; $L__BB195_5: ld.local.u32 %r10, [%rd2+44]; and.b32 %r2, %r10, 2; setp.eq.s32 %p3, %r2, 0; ld.local.s32 %rd4, [%rd2+64]; @%p3 bra $L__BB195_9; bra.uni $L__BB195_6; $L__BB195_9: cvt.u32.u64 %r13, %rd4; setp.ne.s32 %p6, %r13, -1; @%p6 bra $L__BB195_12; bra.uni $L__BB195_10; $L__BB195_12: ld.const.u64 %rd23, [kernel_params+2160]; cvta.to.global.u64 %rd24, %rd23; mul.lo.s64 %rd25, %rd4, 208; add.s64 %rd26, %rd24, %rd25; ld.global.v4.f32 {%f113, %f114, %f115, %f116}, [%rd26]; ld.global.v4.f32 {%f120, %f121, %f122, %f123}, [%rd26+16]; ld.global.v4.f32 {%f127, %f128, %f129, %f130}, [%rd26+32]; mul.ftz.f32 %f134, %f207, %f120; fma.rn.ftz.f32 %f135, %f208, %f113, %f134; fma.rn.ftz.f32 %f136, %f206, %f127, %f135; mul.ftz.f32 %f137, %f207, %f121; fma.rn.ftz.f32 %f138, %f208, %f114, %f137; fma.rn.ftz.f32 %f139, %f206, %f128, %f138; mul.ftz.f32 %f140, %f207, %f122; fma.rn.ftz.f32 %f141, %f208, %f115, %f140; fma.rn.ftz.f32 %f142, %f206, %f129, %f141; mul.ftz.f32 %f143, %f139, %f139; fma.rn.ftz.f32 %f144, %f136, %f136, %f143; fma.rn.ftz.f32 %f145, %f142, %f142, %f144; sqrt.approx.ftz.f32 %f146, %f145; div.approx.ftz.f32 %f208, %f136, %f146; div.approx.ftz.f32 %f207, %f139, %f146; div.approx.ftz.f32 %f206, %f142, %f146; bra.uni $L__BB195_13; $L__BB195_4: sub.ftz.f32 %f43, %f1, %f2; mul.ftz.f32 %f44, %f43, %f3; mul.ftz.f32 %f211, %f44, %f208; mul.ftz.f32 %f210, %f44, %f207; mul.ftz.f32 %f209, %f44, %f206; bra.uni $L__BB195_16; $L__BB195_6: cvt.u32.u64 %r11, %rd4; setp.ne.s32 %p4, %r11, -1; @%p4 bra $L__BB195_8; ld.local.u32 %r12, [%rd2+52]; setp.ne.s32 %p5, %r12, 32; @%p5 bra $L__BB195_13; $L__BB195_8: ld.local.v4.f32 {%f45, %f46, %f47, %f48}, [%rd2+132]; ld.local.v4.f32 {%f52, %f53, %f54, %f55}, [%rd2+148]; ld.local.v4.f32 {%f59, %f60, %f61, %f62}, [%rd2+164]; mul.ftz.f32 %f66, %f207, %f52; fma.rn.ftz.f32 %f67, %f208, %f45, %f66; fma.rn.ftz.f32 %f68, %f206, %f59, %f67; mul.ftz.f32 %f69, %f207, %f53; fma.rn.ftz.f32 %f70, %f208, %f46, %f69; fma.rn.ftz.f32 %f71, %f206, %f60, %f70; mul.ftz.f32 %f72, %f207, %f54; fma.rn.ftz.f32 %f73, %f208, %f47, %f72; fma.rn.ftz.f32 %f74, %f206, %f61, %f73; mul.ftz.f32 %f75, %f71, %f71; fma.rn.ftz.f32 %f76, %f68, %f68, %f75; fma.rn.ftz.f32 %f77, %f74, %f74, %f76; sqrt.approx.ftz.f32 %f78, %f77; div.approx.ftz.f32 %f208, %f68, %f78; div.approx.ftz.f32 %f207, %f71, %f78; div.approx.ftz.f32 %f206, %f74, %f78; $L__BB195_13: sub.ftz.f32 %f147, %f1, %f2; mul.ftz.f32 %f148, %f147, %f3; mul.ftz.f32 %f31, %f148, %f208; mul.ftz.f32 %f32, %f148, %f207; mul.ftz.f32 %f33, %f148, %f206; @%p3 bra $L__BB195_15; bra.uni $L__BB195_14; $L__BB195_15: ld.const.u64 %rd27, [kernel_params+2160]; cvta.to.global.u64 %rd28, %rd27; mul.lo.s64 %rd29, %rd4, 208; add.s64 %rd30, %rd28, %rd29; ld.global.v4.f32 {%f176, %f177, %f178, %f179}, [%rd30]; ld.global.v4.f32 {%f183, %f184, %f185, %f186}, [%rd30+16]; ld.global.v4.f32 {%f190, %f191, %f192, %f193}, [%rd30+32]; mul.ftz.f32 %f197, %f32, %f177; fma.rn.ftz.f32 %f198, %f31, %f176, %f197; fma.rn.ftz.f32 %f211, %f33, %f178, %f198; mul.ftz.f32 %f199, %f32, %f184; fma.rn.ftz.f32 %f200, %f31, %f183, %f199; fma.rn.ftz.f32 %f210, %f33, %f185, %f200; mul.ftz.f32 %f201, %f32, %f191; fma.rn.ftz.f32 %f202, %f31, %f190, %f201; fma.rn.ftz.f32 %f209, %f33, %f192, %f202; bra.uni $L__BB195_16; $L__BB195_14: ld.local.v4.f32 {%f149, %f150, %f151, %f152}, [%rd2+132]; mul.ftz.f32 %f156, %f32, %f150; fma.rn.ftz.f32 %f157, %f31, %f149, %f156; fma.rn.ftz.f32 %f211, %f33, %f151, %f157; ld.local.v4.f32 {%f158, %f159, %f160, %f161}, [%rd2+148]; mul.ftz.f32 %f165, %f32, %f159; fma.rn.ftz.f32 %f166, %f31, %f158, %f165; fma.rn.ftz.f32 %f210, %f33, %f160, %f166; ld.local.v4.f32 {%f167, %f168, %f169, %f170}, [%rd2+164]; mul.ftz.f32 %f174, %f32, %f168; fma.rn.ftz.f32 %f175, %f31, %f167, %f174; fma.rn.ftz.f32 %f209, %f33, %f169, %f175; $L__BB195_16: mul.wide.u32 %rd31, %r5, 4; add.s64 %rd32, %rd1, %rd31; st.local.f32 [%rd32], %f211; st.local.f32 [%rd32+4], %f210; st.local.f32 [%rd32+8], %f209; ret; $L__BB195_10: ld.local.u32 %r14, [%rd2+52]; setp.ne.s32 %p7, %r14, 32; @%p7 bra $L__BB195_13; ld.const.u64 %rd19, [kernel_params+2368]; cvta.to.global.u64 %rd20, %rd19; ld.local.u32 %r15, [%rd2+68]; mul.wide.s32 %rd21, %r15, 192; add.s64 %rd22, %rd20, %rd21; ld.global.v4.f32 {%f79, %f80, %f81, %f82}, [%rd22+48]; ld.global.v4.f32 {%f86, %f87, %f88, %f89}, [%rd22+64]; ld.global.v4.f32 {%f93, %f94, %f95, %f96}, [%rd22+80]; mul.ftz.f32 %f100, %f207, %f86; fma.rn.ftz.f32 %f101, %f208, %f79, %f100; fma.rn.ftz.f32 %f102, %f206, %f93, %f101; mul.ftz.f32 %f103, %f207, %f87; fma.rn.ftz.f32 %f104, %f208, %f80, %f103; fma.rn.ftz.f32 %f105, %f206, %f94, %f104; mul.ftz.f32 %f106, %f207, %f88; fma.rn.ftz.f32 %f107, %f208, %f81, %f106; fma.rn.ftz.f32 %f108, %f206, %f95, %f107; mul.ftz.f32 %f109, %f105, %f105; fma.rn.ftz.f32 %f110, %f102, %f102, %f109; fma.rn.ftz.f32 %f111, %f108, %f108, %f110; sqrt.approx.ftz.f32 %f112, %f111; div.approx.ftz.f32 %f208, %f102, %f112; div.approx.ftz.f32 %f207, %f105, %f112; div.approx.ftz.f32 %f206, %f108, %f112; bra.uni $L__BB195_13; } .func (.param .b32 func_retval0) _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i( .param .b64 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0, .param .b64 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1, .param .align 4 .b8 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2[16], .param .b32 _Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3 ) { .reg .pred %p<75>; .reg .b16 %rs<23>; .reg .f32 %f<349>; .reg .b32 %r<194>; .reg .b64 %rd<217>; ld.param.u64 %rd65, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_0]; ld.param.u64 %rd66, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_1]; ld.param.u32 %r65, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_3]; ld.param.u32 %r64, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+12]; ld.param.u32 %r63, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+8]; ld.param.u32 %r62, [_Z28svm_node_vector_displacementILj447EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i_param_2+4]; cvta.to.local.u64 %rd67, %rd65; ld.const.u64 %rd68, [kernel_params+2400]; cvta.to.global.u64 %rd69, %rd68; mul.wide.s32 %rd70, %r65, 16; add.s64 %rd71, %rd69, %rd70; cvta.to.local.u64 %rd1, %rd66; shl.b32 %r66, %r62, 2; cvt.u64.u32 %rd72, %r66; and.b64 %rd73, %rd72, 1020; add.s64 %rd74, %rd1, %rd73; shr.u32 %r67, %r62, 6; cvt.u64.u32 %rd75, %r67; and.b64 %rd76, %rd75, 1020; add.s64 %rd77, %rd1, %rd76; shr.u32 %r68, %r62, 14; cvt.u64.u32 %rd78, %r68; and.b64 %rd79, %rd78, 1020; add.s64 %rd80, %rd1, %rd79; ld.local.f32 %f80, [%rd77]; ld.local.f32 %f81, [%rd74]; sub.ftz.f32 %f82, %f81, %f80; ld.local.f32 %f83, [%rd74+4]; sub.ftz.f32 %f84, %f83, %f80; ld.local.f32 %f85, [%rd74+8]; sub.ftz.f32 %f86, %f85, %f80; ld.local.f32 %f87, [%rd80]; mul.ftz.f32 %f346, %f82, %f87; mul.ftz.f32 %f347, %f84, %f87; mul.ftz.f32 %f348, %f86, %f87; ld.global.u32 %r2, [%rd71]; setp.ne.s32 %p1, %r2, 0; add.s64 %rd2, %rd67, 12; @%p1 bra $L__BB196_81; ld.local.f32 %f332, [%rd2]; ld.local.f32 %f331, [%rd2+4]; ld.local.f32 %f330, [%rd2+8]; ld.local.u8 %rs13, [%rd2+44]; and.b16 %rs14, %rs13, 2; setp.eq.s16 %p2, %rs14, 0; ld.local.u32 %r181, [%rd2+64]; cvt.s64.s32 %rd3, %r181; @%p2 bra $L__BB196_5; bra.uni $L__BB196_2; $L__BB196_5: setp.ne.s32 %p5, %r181, -1; @%p5 bra $L__BB196_8; bra.uni $L__BB196_6; $L__BB196_8: ld.const.u64 %rd85, [kernel_params+2160]; cvta.to.global.u64 %rd86, %rd85; mul.lo.s64 %rd87, %rd3, 208; add.s64 %rd88, %rd86, %rd87; ld.global.v4.f32 {%f156, %f157, %f158, %f159}, [%rd88]; ld.global.v4.f32 {%f163, %f164, %f165, %f166}, [%rd88+16]; ld.global.v4.f32 {%f170, %f171, %f172, %f173}, [%rd88+32]; mul.ftz.f32 %f177, %f331, %f163; fma.rn.ftz.f32 %f178, %f332, %f156, %f177; fma.rn.ftz.f32 %f179, %f330, %f170, %f178; mul.ftz.f32 %f180, %f331, %f164; fma.rn.ftz.f32 %f181, %f332, %f157, %f180; fma.rn.ftz.f32 %f182, %f330, %f171, %f181; mul.ftz.f32 %f183, %f331, %f165; fma.rn.ftz.f32 %f184, %f332, %f158, %f183; fma.rn.ftz.f32 %f185, %f330, %f172, %f184; mul.ftz.f32 %f186, %f182, %f182; fma.rn.ftz.f32 %f187, %f179, %f179, %f186; fma.rn.ftz.f32 %f188, %f185, %f185, %f187; sqrt.approx.ftz.f32 %f189, %f188; div.approx.ftz.f32 %f332, %f179, %f189; div.approx.ftz.f32 %f331, %f182, %f189; div.approx.ftz.f32 %f330, %f185, %f189; bra.uni $L__BB196_9; $L__BB196_2: setp.ne.s32 %p3, %r181, -1; @%p3 bra $L__BB196_4; ld.local.u32 %r69, [%rd2+52]; setp.ne.s32 %p4, %r69, 32; @%p4 bra $L__BB196_9; $L__BB196_4: ld.local.v4.f32 {%f88, %f89, %f90, %f91}, [%rd2+132]; ld.local.v4.f32 {%f95, %f96, %f97, %f98}, [%rd2+148]; ld.local.v4.f32 {%f102, %f103, %f104, %f105}, [%rd2+164]; mul.ftz.f32 %f109, %f331, %f95; fma.rn.ftz.f32 %f110, %f332, %f88, %f109; fma.rn.ftz.f32 %f111, %f330, %f102, %f110; mul.ftz.f32 %f112, %f331, %f96; fma.rn.ftz.f32 %f113, %f332, %f89, %f112; fma.rn.ftz.f32 %f114, %f330, %f103, %f113; mul.ftz.f32 %f115, %f331, %f97; fma.rn.ftz.f32 %f116, %f332, %f90, %f115; fma.rn.ftz.f32 %f117, %f330, %f104, %f116; mul.ftz.f32 %f118, %f114, %f114; fma.rn.ftz.f32 %f119, %f111, %f111, %f118; fma.rn.ftz.f32 %f120, %f117, %f117, %f119; sqrt.approx.ftz.f32 %f121, %f120; div.approx.ftz.f32 %f332, %f111, %f121; div.approx.ftz.f32 %f331, %f114, %f121; div.approx.ftz.f32 %f330, %f117, %f121; $L__BB196_9: cvt.u64.u32 %rd4, %r63; ld.local.u32 %r180, [%rd2+48]; cvt.s64.s32 %rd5, %r180; setp.eq.s32 %p7, %r181, -1; mov.u32 %r25, -1; mov.u32 %r22, 0; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p7 bra $L__BB196_24; ld.const.u64 %rd89, [kernel_params+2160]; cvta.to.global.u64 %rd90, %rd89; mul.lo.s64 %rd91, %rd3, 208; add.s64 %rd92, %rd90, %rd91; ld.global.u32 %r5, [%rd92+164]; ld.local.u32 %r77, [%rd2+52]; and.b32 %r78, %r77, 1; setp.eq.b32 %p8, %r78, 1; mov.pred %p9, 0; xor.pred %p10, %p8, %p9; not.pred %p11, %p10; mov.u32 %r170, 0; @%p11 bra $L__BB196_14; setp.eq.s32 %p12, %r180, -1; mov.u32 %r169, -1; @%p12 bra $L__BB196_13; ld.const.u64 %rd93, [kernel_params+2240]; cvta.to.global.u64 %rd94, %rd93; shl.b64 %rd95, %rd5, 2; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r169, [%rd96]; $L__BB196_13: setp.ne.s32 %p13, %r169, -1; selp.u32 %r170, 1, 0, %p13; $L__BB196_14: ld.const.u64 %rd97, [kernel_params+2312]; cvta.to.global.u64 %rd6, %rd97; add.s32 %r172, %r170, %r5; mul.wide.u32 %rd98, %r172, 16; add.s64 %rd99, %rd6, %rd98; add.s64 %rd201, %rd99, 14; ld.global.u16 %rs19, [%rd99+12]; ld.global.u32 %r171, [%rd99+8]; ld.global.u64 %rd199, [%rd99]; setp.eq.s64 %p14, %rd199, %rd4; add.s64 %rd200, %rd99, 15; @%p14 bra $L__BB196_20; $L__BB196_15: setp.eq.s64 %p15, %rd199, 0; @%p15 bra $L__BB196_17; add.s32 %r172, %r172, 2; bra.uni $L__BB196_18; $L__BB196_17: setp.eq.s16 %p16, %rs19, 0; mov.u32 %r22, 0; mov.u32 %r172, %r171; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p16 bra $L__BB196_24; $L__BB196_18: cvt.u64.u32 %rd11, %r172; mul.wide.u32 %rd100, %r172, 16; add.s64 %rd101, %rd6, %rd100; ld.global.u16 %rs19, [%rd101+12]; ld.global.u32 %r171, [%rd101+8]; ld.global.u64 %rd199, [%rd101]; setp.ne.s64 %p17, %rd199, %rd4; @%p17 bra $L__BB196_15; shl.b64 %rd102, %rd11, 4; add.s64 %rd103, %rd6, %rd102; add.s64 %rd201, %rd103, 14; add.s64 %rd200, %rd103, 15; $L__BB196_20: ld.global.u8 %rs5, [%rd201]; ld.global.u8 %rs6, [%rd200]; setp.ne.s32 %p18, %r180, -1; @%p18 bra $L__BB196_23; mov.u32 %r22, 0; add.s16 %rs15, %rs19, -1; setp.lt.u16 %p19, %rs15, 2; @%p19 bra $L__BB196_23; setp.ne.s16 %p20, %rs19, 1024; mov.u32 %r176, %r22; mov.u32 %r177, %r22; @%p20 bra $L__BB196_24; $L__BB196_23: setp.eq.s16 %p21, %rs19, 0; selp.b32 %r25, -1, %r171, %p21; cvt.u32.u16 %r22, %rs19; cvt.u32.u16 %r88, %rs6; and.b32 %r177, %r88, 255; cvt.u32.u16 %r89, %rs5; and.b32 %r176, %r89, 255; $L__BB196_24: setp.eq.s32 %p22, %r25, -1; @%p22 bra $L__BB196_44; ld.local.u32 %r28, [%rd2+52]; and.b32 %r90, %r28, 1; setp.eq.b32 %p23, %r90, 1; mov.pred %p24, 0; xor.pred %p25, %p23, %p24; not.pred %p26, %p25; @%p26 bra $L__BB196_37; bra.uni $L__BB196_26; $L__BB196_37: and.b32 %r108, %r28, 6; setp.eq.s32 %p33, %r108, 0; @%p33 bra $L__BB196_42; and.b32 %r109, %r22, 768; setp.eq.s32 %p34, %r109, 0; @%p34 bra $L__BB196_40; ld.const.u64 %rd125, [kernel_params+2264]; cvta.to.global.u64 %rd126, %rd125; shl.b64 %rd127, %rd5, 4; add.s64 %rd128, %rd126, %rd127; shr.s32 %r110, %r28, 7; ld.global.u32 %r111, [%rd128+4]; add.s32 %r112, %r110, %r111; ld.const.u64 %rd129, [kernel_params+2336]; cvta.to.global.u64 %rd130, %rd129; add.s32 %r113, %r112, %r25; mul.wide.s32 %rd131, %r113, 12; add.s64 %rd132, %rd130, %rd131; ld.local.f32 %f213, [%rd2+56]; mov.f32 %f214, 0f3F800000; sub.ftz.f32 %f215, %f214, %f213; ld.global.f32 %f216, [%rd132]; mul.ftz.f32 %f217, %f216, %f215; ld.global.f32 %f218, [%rd132+4]; mul.ftz.f32 %f219, %f218, %f215; ld.global.f32 %f220, [%rd132+8]; mul.ftz.f32 %f221, %f220, %f215; ld.global.f32 %f222, [%rd132+12]; ld.global.f32 %f223, [%rd132+16]; ld.global.f32 %f224, [%rd132+20]; fma.rn.ftz.f32 %f335, %f224, %f213, %f221; fma.rn.ftz.f32 %f334, %f223, %f213, %f219; fma.rn.ftz.f32 %f333, %f222, %f213, %f217; bra.uni $L__BB196_45; $L__BB196_44: ld.local.f32 %f231, [%rd2+104]; ld.local.f32 %f232, [%rd2+108]; mul.ftz.f32 %f233, %f232, %f232; fma.rn.ftz.f32 %f234, %f231, %f231, %f233; ld.local.f32 %f235, [%rd2+112]; fma.rn.ftz.f32 %f236, %f235, %f235, %f234; sqrt.approx.ftz.f32 %f237, %f236; div.approx.ftz.f32 %f333, %f231, %f237; div.approx.ftz.f32 %f334, %f232, %f237; div.approx.ftz.f32 %f335, %f235, %f237; bra.uni $L__BB196_45; $L__BB196_26: setp.eq.s32 %p27, %r180, -1; mov.u32 %r179, -1; @%p27 bra $L__BB196_28; ld.const.u64 %rd104, [kernel_params+2240]; cvta.to.global.u64 %rd105, %rd104; shl.b64 %rd106, %rd5, 2; add.s64 %rd107, %rd105, %rd106; ld.global.u32 %r179, [%rd107]; $L__BB196_28: setp.eq.s32 %p28, %r179, -1; @%p28 bra $L__BB196_30; mov.u64 %rd108, 0; { // callseq 966, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r22; st.param.b32 [param1+4], %r176; st.param.b32 [param1+8], %r177; st.param.b32 [param1+12], %r25; .param .b64 param2; st.param.b64 [param2+0], %rd108; .param .b64 param3; st.param.b64 [param3+0], %rd108; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z30subd_triangle_attribute_float3PK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorP6float3S7_, ( param0, param1, param2, param3 ); ld.param.f32 %f333, [retval0+0]; ld.param.f32 %f334, [retval0+4]; ld.param.f32 %f335, [retval0+8]; } // callseq 966 ld.local.u32 %r181, [%rd2+64]; ld.local.u32 %r180, [%rd2+48]; bra.uni $L__BB196_45; $L__BB196_6: ld.local.u32 %r70, [%rd2+52]; setp.ne.s32 %p6, %r70, 32; @%p6 bra $L__BB196_9; ld.const.u64 %rd81, [kernel_params+2368]; cvta.to.global.u64 %rd82, %rd81; ld.local.u32 %r71, [%rd2+68]; mul.wide.s32 %rd83, %r71, 192; add.s64 %rd84, %rd82, %rd83; ld.global.v4.f32 {%f122, %f123, %f124, %f125}, [%rd84+48]; ld.global.v4.f32 {%f129, %f130, %f131, %f132}, [%rd84+64]; ld.global.v4.f32 {%f136, %f137, %f138, %f139}, [%rd84+80]; mul.ftz.f32 %f143, %f331, %f129; fma.rn.ftz.f32 %f144, %f332, %f122, %f143; fma.rn.ftz.f32 %f145, %f330, %f136, %f144; mul.ftz.f32 %f146, %f331, %f130; fma.rn.ftz.f32 %f147, %f332, %f123, %f146; fma.rn.ftz.f32 %f148, %f330, %f137, %f147; mul.ftz.f32 %f149, %f331, %f131; fma.rn.ftz.f32 %f150, %f332, %f124, %f149; fma.rn.ftz.f32 %f151, %f330, %f138, %f150; mul.ftz.f32 %f152, %f148, %f148; fma.rn.ftz.f32 %f153, %f145, %f145, %f152; fma.rn.ftz.f32 %f154, %f151, %f151, %f153; sqrt.approx.ftz.f32 %f155, %f154; div.approx.ftz.f32 %f332, %f145, %f155; div.approx.ftz.f32 %f331, %f148, %f155; div.approx.ftz.f32 %f330, %f151, %f155; bra.uni $L__BB196_9; $L__BB196_42: and.b32 %r117, %r28, 8; setp.eq.s32 %p37, %r117, 0; setp.ne.s32 %p38, %r22, 8; mov.f32 %f333, 0f00000000; or.pred %p39, %p38, %p37; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p39 bra $L__BB196_45; ld.const.u64 %rd137, [kernel_params+2336]; cvta.to.global.u64 %rd138, %rd137; add.s32 %r118, %r180, %r25; mul.wide.s32 %rd139, %r118, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f335, [%rd140+8]; ld.global.f32 %f334, [%rd140+4]; ld.global.f32 %f333, [%rd140]; bra.uni $L__BB196_45; $L__BB196_30: and.b32 %r92, %r22, 56; setp.eq.s32 %p29, %r92, 0; @%p29 bra $L__BB196_35; and.b32 %r93, %r22, 24; setp.eq.s32 %p30, %r93, 0; @%p30 bra $L__BB196_33; ld.const.u64 %rd109, [kernel_params+2232]; cvta.to.global.u64 %rd110, %rd109; shl.b64 %rd111, %rd5, 4; add.s64 %rd112, %rd110, %rd111; ld.global.v4.u32 {%r94, %r95, %r96, %r97}, [%rd112]; ld.const.u64 %rd113, [kernel_params+2336]; cvta.to.global.u64 %rd114, %rd113; add.s32 %r101, %r94, %r25; mul.wide.u32 %rd115, %r101, 12; add.s64 %rd208, %rd114, %rd115; add.s32 %r102, %r95, %r25; mul.wide.u32 %rd116, %r102, 12; add.s64 %rd205, %rd114, %rd116; add.s32 %r103, %r96, %r25; mul.wide.u32 %rd117, %r103, 12; add.s64 %rd202, %rd114, %rd117; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd206, %rd205, 4; add.s64 %rd207, %rd205, 8; add.s64 %rd203, %rd202, 4; add.s64 %rd204, %rd202, 8; bra.uni $L__BB196_34; $L__BB196_40: and.b32 %r114, %r22, 131; setp.eq.s32 %p35, %r114, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p35 bra $L__BB196_45; setp.eq.s32 %p36, %r22, 128; selp.b32 %r115, %r180, 0, %p36; add.s32 %r116, %r25, %r115; ld.const.u64 %rd133, [kernel_params+2336]; cvta.to.global.u64 %rd134, %rd133; mul.wide.s32 %rd135, %r116, 12; add.s64 %rd136, %rd134, %rd135; ld.global.f32 %f335, [%rd136+8]; ld.global.f32 %f334, [%rd136+4]; ld.global.f32 %f333, [%rd136]; bra.uni $L__BB196_45; $L__BB196_35: and.b32 %r105, %r22, 7; setp.eq.s32 %p31, %r105, 0; mov.f32 %f333, 0f00000000; mov.f32 %f334, %f333; mov.f32 %f335, %f333; @%p31 bra $L__BB196_45; setp.eq.s32 %p32, %r22, 4; selp.b32 %r106, %r180, 0, %p32; add.s32 %r107, %r25, %r106; ld.const.u64 %rd121, [kernel_params+2336]; cvta.to.global.u64 %rd122, %rd121; mul.wide.s32 %rd123, %r107, 12; add.s64 %rd124, %rd122, %rd123; ld.global.f32 %f335, [%rd124+8]; ld.global.f32 %f334, [%rd124+4]; ld.global.f32 %f333, [%rd124]; bra.uni $L__BB196_45; $L__BB196_33: mad.lo.s32 %r104, %r180, 3, %r25; ld.const.u64 %rd118, [kernel_params+2336]; cvta.to.global.u64 %rd119, %rd118; mul.wide.s32 %rd120, %r104, 12; add.s64 %rd208, %rd119, %rd120; add.s64 %rd209, %rd208, 4; add.s64 %rd210, %rd208, 8; add.s64 %rd205, %rd208, 12; add.s64 %rd206, %rd208, 16; add.s64 %rd207, %rd208, 20; add.s64 %rd202, %rd208, 24; add.s64 %rd203, %rd208, 28; add.s64 %rd204, %rd208, 32; $L__BB196_34: ld.local.f32 %f190, [%rd2+56]; ld.global.f32 %f191, [%rd205]; ld.global.f32 %f192, [%rd206]; ld.global.f32 %f193, [%rd207]; ld.local.f32 %f194, [%rd2+60]; ld.global.f32 %f195, [%rd202]; mul.ftz.f32 %f196, %f195, %f194; ld.global.f32 %f197, [%rd203]; mul.ftz.f32 %f198, %f197, %f194; ld.global.f32 %f199, [%rd204]; mul.ftz.f32 %f200, %f199, %f194; fma.rn.ftz.f32 %f201, %f191, %f190, %f196; fma.rn.ftz.f32 %f202, %f192, %f190, %f198; fma.rn.ftz.f32 %f203, %f193, %f190, %f200; mov.f32 %f204, 0f3F800000; sub.ftz.f32 %f205, %f204, %f190; sub.ftz.f32 %f206, %f205, %f194; ld.global.f32 %f207, [%rd208]; ld.global.f32 %f208, [%rd209]; ld.global.f32 %f209, [%rd210]; fma.rn.ftz.f32 %f335, %f209, %f206, %f203; fma.rn.ftz.f32 %f334, %f208, %f206, %f202; fma.rn.ftz.f32 %f333, %f207, %f206, %f201; $L__BB196_45: mul.ftz.f32 %f238, %f330, %f334; mul.ftz.f32 %f239, %f331, %f335; sub.ftz.f32 %f240, %f239, %f238; mul.ftz.f32 %f241, %f332, %f335; mul.ftz.f32 %f242, %f330, %f333; sub.ftz.f32 %f243, %f242, %f241; mul.ftz.f32 %f244, %f331, %f333; mul.ftz.f32 %f245, %f332, %f334; sub.ftz.f32 %f246, %f245, %f244; mul.ftz.f32 %f247, %f243, %f243; fma.rn.ftz.f32 %f248, %f240, %f240, %f247; fma.rn.ftz.f32 %f249, %f246, %f246, %f248; sqrt.approx.ftz.f32 %f250, %f249; div.approx.ftz.f32 %f342, %f240, %f250; div.approx.ftz.f32 %f341, %f243, %f250; div.approx.ftz.f32 %f340, %f246, %f250; cvt.u64.u32 %rd44, %r64; setp.eq.s32 %p40, %r181, -1; mov.u32 %r57, -1; mov.u32 %r54, 0; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p40 bra $L__BB196_60; ld.const.u64 %rd141, [kernel_params+2160]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r181, 208; add.s64 %rd144, %rd142, %rd143; ld.global.u32 %r37, [%rd144+164]; ld.local.u32 %r124, [%rd2+52]; and.b32 %r125, %r124, 1; setp.eq.b32 %p41, %r125, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; mov.u32 %r185, 0; @%p44 bra $L__BB196_50; setp.eq.s32 %p45, %r180, -1; mov.u32 %r184, -1; @%p45 bra $L__BB196_49; ld.const.u64 %rd145, [kernel_params+2240]; cvta.to.global.u64 %rd146, %rd145; mul.wide.s32 %rd147, %r180, 4; add.s64 %rd148, %rd146, %rd147; ld.global.u32 %r184, [%rd148]; $L__BB196_49: setp.ne.s32 %p46, %r184, -1; selp.u32 %r185, 1, 0, %p46; $L__BB196_50: ld.const.u64 %rd149, [kernel_params+2312]; cvta.to.global.u64 %rd45, %rd149; add.s32 %r187, %r185, %r37; mul.wide.u32 %rd150, %r187, 16; add.s64 %rd151, %rd45, %rd150; add.s64 %rd213, %rd151, 14; ld.global.u16 %rs21, [%rd151+12]; ld.global.u32 %r186, [%rd151+8]; ld.global.u64 %rd211, [%rd151]; setp.eq.s64 %p47, %rd211, %rd44; add.s64 %rd212, %rd151, 15; @%p47 bra $L__BB196_56; $L__BB196_51: setp.eq.s64 %p48, %rd211, 0; @%p48 bra $L__BB196_53; add.s32 %r187, %r187, 2; bra.uni $L__BB196_54; $L__BB196_53: setp.eq.s16 %p49, %rs21, 0; mov.u32 %r54, 0; mov.u32 %r187, %r186; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p49 bra $L__BB196_60; $L__BB196_54: cvt.u64.u32 %rd50, %r187; mul.wide.u32 %rd152, %r187, 16; add.s64 %rd153, %rd45, %rd152; ld.global.u16 %rs21, [%rd153+12]; ld.global.u32 %r186, [%rd153+8]; ld.global.u64 %rd211, [%rd153]; setp.ne.s64 %p50, %rd211, %rd44; @%p50 bra $L__BB196_51; shl.b64 %rd154, %rd50, 4; add.s64 %rd155, %rd45, %rd154; add.s64 %rd213, %rd155, 14; add.s64 %rd212, %rd155, 15; $L__BB196_56: ld.global.u8 %rs11, [%rd213]; ld.global.u8 %rs12, [%rd212]; setp.ne.s32 %p51, %r180, -1; @%p51 bra $L__BB196_59; mov.u32 %r54, 0; add.s16 %rs16, %rs21, -1; setp.lt.u16 %p52, %rs16, 2; @%p52 bra $L__BB196_59; setp.ne.s16 %p53, %rs21, 1024; mov.u32 %r191, %r54; mov.u32 %r192, %r54; @%p53 bra $L__BB196_60; $L__BB196_59: setp.eq.s16 %p54, %rs21, 0; selp.b32 %r57, -1, %r186, %p54; cvt.u32.u16 %r54, %rs21; cvt.u32.u16 %r135, %rs12; and.b32 %r192, %r135, 255; cvt.u32.u16 %r136, %rs11; and.b32 %r191, %r136, 255; $L__BB196_60: setp.eq.s32 %p55, %r57, -1; @%p55 bra $L__BB196_80; ld.local.u32 %r60, [%rd2+52]; and.b32 %r137, %r60, 1; setp.eq.b32 %p56, %r137, 1; mov.pred %p57, 0; xor.pred %p58, %p56, %p57; not.pred %p59, %p58; @%p59 bra $L__BB196_72; bra.uni $L__BB196_62; $L__BB196_72: and.b32 %r155, %r60, 6; setp.eq.s32 %p66, %r155, 0; @%p66 bra $L__BB196_77; and.b32 %r156, %r54, 768; setp.eq.s32 %p67, %r156, 0; @%p67 bra $L__BB196_75; ld.const.u64 %rd177, [kernel_params+2264]; cvta.to.global.u64 %rd178, %rd177; mul.wide.s32 %rd179, %r180, 16; add.s64 %rd180, %rd178, %rd179; shr.s32 %r157, %r60, 7; ld.global.u32 %r158, [%rd180+4]; add.s32 %r159, %r157, %r158; ld.const.u64 %rd181, [kernel_params+2320]; cvta.to.global.u64 %rd182, %rd181; add.s32 %r160, %r159, %r57; mul.wide.s32 %rd183, %r160, 4; add.s64 %rd184, %rd182, %rd183; ld.local.f32 %f262, [%rd2+56]; mov.f32 %f263, 0f3F800000; sub.ftz.f32 %f264, %f263, %f262; ld.global.f32 %f265, [%rd184]; mul.ftz.f32 %f266, %f265, %f264; ld.global.f32 %f267, [%rd184+4]; fma.rn.ftz.f32 %f339, %f267, %f262, %f266; bra.uni $L__BB196_79; $L__BB196_62: setp.eq.s32 %p60, %r180, -1; @%p60 bra $L__BB196_65; ld.const.u64 %rd156, [kernel_params+2240]; cvta.to.global.u64 %rd157, %rd156; mul.wide.s32 %rd158, %r180, 4; add.s64 %rd159, %rd157, %rd158; ld.global.u32 %r138, [%rd159]; setp.eq.s32 %p61, %r138, -1; @%p61 bra $L__BB196_65; mov.u64 %rd160, 0; { // callseq 967, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd65; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r54; st.param.b32 [param1+4], %r191; st.param.b32 [param1+8], %r192; st.param.b32 [param1+12], %r57; .param .b64 param2; st.param.b64 [param2+0], %rd160; .param .b64 param3; st.param.b64 [param3+0], %rd160; .param .b32 retval0; call.uni (retval0), _Z29subd_triangle_attribute_floatPK16KernelGlobalsGPUPK10ShaderData19AttributeDescriptorPfS6_, ( param0, param1, param2, param3 ); ld.param.f32 %f339, [retval0+0]; } // callseq 967 bra.uni $L__BB196_79; $L__BB196_65: and.b32 %r139, %r54, 56; setp.eq.s32 %p62, %r139, 0; @%p62 bra $L__BB196_70; and.b32 %r140, %r54, 24; setp.eq.s32 %p63, %r140, 0; @%p63 bra $L__BB196_68; ld.const.u64 %rd161, [kernel_params+2232]; cvta.to.global.u64 %rd162, %rd161; mul.wide.s32 %rd163, %r180, 16; add.s64 %rd164, %rd162, %rd163; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd164]; ld.const.u64 %rd165, [kernel_params+2320]; cvta.to.global.u64 %rd166, %rd165; add.s32 %r148, %r141, %r57; mul.wide.u32 %rd167, %r148, 4; add.s64 %rd214, %rd166, %rd167; add.s32 %r149, %r142, %r57; mul.wide.u32 %rd168, %r149, 4; add.s64 %rd215, %rd166, %rd168; add.s32 %r150, %r143, %r57; mul.wide.u32 %rd169, %r150, 4; add.s64 %rd216, %rd166, %rd169; bra.uni $L__BB196_69; $L__BB196_77: and.b32 %r164, %r60, 8; setp.eq.s32 %p70, %r164, 0; setp.ne.s32 %p71, %r54, 8; mov.f32 %f339, 0f00000000; or.pred %p72, %p71, %p70; @%p72 bra $L__BB196_79; ld.const.u64 %rd189, [kernel_params+2320]; cvta.to.global.u64 %rd190, %rd189; add.s32 %r165, %r180, %r57; mul.wide.s32 %rd191, %r165, 4; add.s64 %rd192, %rd190, %rd191; ld.global.f32 %f339, [%rd192]; bra.uni $L__BB196_79; $L__BB196_75: and.b32 %r161, %r54, 131; setp.eq.s32 %p68, %r161, 0; mov.f32 %f339, 0f00000000; @%p68 bra $L__BB196_79; setp.eq.s32 %p69, %r54, 128; selp.b32 %r162, %r180, 0, %p69; add.s32 %r163, %r57, %r162; ld.const.u64 %rd185, [kernel_params+2320]; cvta.to.global.u64 %rd186, %rd185; mul.wide.s32 %rd187, %r163, 4; add.s64 %rd188, %rd186, %rd187; ld.global.f32 %f339, [%rd188]; bra.uni $L__BB196_79; $L__BB196_70: and.b32 %r152, %r54, 7; setp.eq.s32 %p64, %r152, 0; mov.f32 %f339, 0f00000000; @%p64 bra $L__BB196_79; setp.eq.s32 %p65, %r54, 4; selp.b32 %r153, %r180, 0, %p65; add.s32 %r154, %r57, %r153; ld.const.u64 %rd173, [kernel_params+2320]; cvta.to.global.u64 %rd174, %rd173; mul.wide.s32 %rd175, %r154, 4; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f339, [%rd176]; bra.uni $L__BB196_79; $L__BB196_68: mad.lo.s32 %r151, %r180, 3, %r57; ld.const.u64 %rd170, [kernel_params+2320]; cvta.to.global.u64 %rd171, %rd170; mul.wide.s32 %rd172, %r151, 4; add.s64 %rd214, %rd171, %rd172; add.s64 %rd215, %rd214, 4; add.s64 %rd216, %rd214, 8; $L__BB196_69: ld.local.f32 %f251, [%rd2+56]; ld.global.f32 %f252, [%rd215]; ld.local.f32 %f253, [%rd2+60]; ld.global.f32 %f254, [%rd216]; mul.ftz.f32 %f255, %f254, %f253; fma.rn.ftz.f32 %f256, %f252, %f251, %f255; mov.f32 %f257, 0f3F800000; sub.ftz.f32 %f258, %f257, %f251; sub.ftz.f32 %f259, %f258, %f253; ld.global.f32 %f260, [%rd214]; fma.rn.ftz.f32 %f339, %f260, %f259, %f256; $L__BB196_79: mul.ftz.f32 %f342, %f342, %f339; mul.ftz.f32 %f341, %f341, %f339; mul.ftz.f32 %f340, %f340, %f339; $L__BB196_80: mul.ftz.f32 %f270, %f346, %f333; fma.rn.ftz.f32 %f271, %f347, %f332, %f270; mul.ftz.f32 %f272, %f346, %f334; fma.rn.ftz.f32 %f273, %f347, %f331, %f272; mul.ftz.f32 %f274, %f346, %f335; fma.rn.ftz.f32 %f275, %f347, %f330, %f274; fma.rn.ftz.f32 %f346, %f348, %f342, %f271; fma.rn.ftz.f32 %f347, %f348, %f341, %f273; fma.rn.ftz.f32 %f348, %f348, %f340, %f275; $L__BB196_81: setp.eq.s32 %p73, %r2, 2; @%p73 bra $L__BB196_85; ld.local.u8 %rs17, [%rd2+44]; and.b16 %rs18, %rs17, 2; setp.eq.s16 %p74, %rs18, 0; @%p74 bra $L__BB196_84; bra.uni $L__BB196_83; $L__BB196_84: ld.const.u64 %rd193, [kernel_params+2160]; cvta.to.global.u64 %rd194, %rd193; ld.local.u32 %r166, [%rd2+64]; mul.wide.s32 %rd195, %r166, 208; add.s64 %rd196, %rd194, %rd195; ld.global.v4.f32 {%f303, %f304, %f305, %f306}, [%rd196]; ld.global.v4.f32 {%f310, %f311, %f312, %f313}, [%rd196+16]; ld.global.v4.f32 {%f317, %f318, %f319, %f320}, [%rd196+32]; mul.ftz.f32 %f324, %f347, %f304; fma.rn.ftz.f32 %f325, %f346, %f303, %f324; fma.rn.ftz.f32 %f74, %f348, %f305, %f325; mul.ftz.f32 %f326, %f347, %f311; fma.rn.ftz.f32 %f327, %f346, %f310, %f326; fma.rn.ftz.f32 %f75, %f348, %f312, %f327; mul.ftz.f32 %f328, %f347, %f318; fma.rn.ftz.f32 %f329, %f346, %f317, %f328; fma.rn.ftz.f32 %f348, %f348, %f319, %f329; mov.f32 %f346, %f74; mov.f32 %f347, %f75; bra.uni $L__BB196_85; $L__BB196_83: ld.local.v4.f32 {%f276, %f277, %f278, %f279}, [%rd2+132]; mul.ftz.f32 %f283, %f347, %f277; fma.rn.ftz.f32 %f284, %f346, %f276, %f283; fma.rn.ftz.f32 %f71, %f348, %f278, %f284; ld.local.v4.f32 {%f285, %f286, %f287, %f288}, [%rd2+148]; mul.ftz.f32 %f292, %f347, %f286; fma.rn.ftz.f32 %f293, %f346, %f285, %f292; fma.rn.ftz.f32 %f72, %f348, %f287, %f293; ld.local.v4.f32 {%f294, %f295, %f296, %f297}, [%rd2+164]; mul.ftz.f32 %f301, %f347, %f295; fma.rn.ftz.f32 %f302, %f346, %f294, %f301; fma.rn.ftz.f32 %f348, %f348, %f296, %f302; mov.f32 %f346, %f71; mov.f32 %f347, %f72; $L__BB196_85: shr.u32 %r167, %r62, 24; mul.wide.u32 %rd197, %r167, 4; add.s64 %rd198, %rd1, %rd197; st.local.f32 [%rd198], %f346; st.local.f32 [%rd198+4], %f347; st.local.f32 [%rd198+8], %f348; add.s32 %r168, %r65, 1; st.param.b32 [func_retval0+0], %r168; ret; } .func _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4( .param .b64 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0, .param .b64 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1, .param .align 4 .b8 _Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2[16] ) { .reg .pred %p<37>; .reg .b16 %rs<3>; .reg .f32 %f<592>; .reg .b32 %r<22>; .reg .b64 %rd<45>; ld.param.u64 %rd4, [_Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_0]; ld.param.u64 %rd5, [_Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_1]; ld.param.u32 %r8, [_Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+12]; ld.param.u32 %r7, [_Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+8]; ld.param.u32 %r6, [_Z17svm_node_set_bumpILj447EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4_param_2+4]; cvta.to.local.u64 %rd6, %rd4; cvta.to.local.u64 %rd1, %rd5; and.b32 %r3, %r6, 255; setp.eq.s32 %p3, %r3, 255; add.s64 %rd2, %rd6, 12; @%p3 bra $L__BB197_2; bra.uni $L__BB197_1; $L__BB197_2: ld.local.f32 %f7, [%rd2]; ld.local.f32 %f8, [%rd2+4]; ld.local.f32 %f9, [%rd2+8]; bra.uni $L__BB197_3; $L__BB197_1: mul.wide.u32 %rd7, %r3, 4; add.s64 %rd8, %rd1, %rd7; ld.local.f32 %f7, [%rd8]; ld.local.f32 %f8, [%rd8+4]; ld.local.f32 %f9, [%rd8+8]; $L__BB197_3: ld.local.v2.f32 {%f124, %f125}, [%rd2+12]; setp.neu.ftz.f32 %p4, %f124, %f125; ld.local.f32 %f12, [%rd2+20]; setp.neu.ftz.f32 %p5, %f124, %f12; or.pred %p6, %p4, %p5; sub.ftz.f32 %f126, %f12, %f125; neg.ftz.f32 %f127, %f125; neg.ftz.f32 %f128, %f12; selp.f32 %f129, %f128, %f12, %p6; add.ftz.f32 %f130, %f124, %f129; selp.f32 %f131, %f125, %f127, %p6; sub.ftz.f32 %f132, %f131, %f124; mul.ftz.f32 %f133, %f130, %f130; fma.rn.ftz.f32 %f134, %f126, %f126, %f133; fma.rn.ftz.f32 %f135, %f132, %f132, %f134; sqrt.approx.ftz.f32 %f136, %f135; div.approx.ftz.f32 %f137, %f126, %f136; div.approx.ftz.f32 %f138, %f130, %f136; div.approx.ftz.f32 %f139, %f132, %f136; mul.ftz.f32 %f140, %f125, %f139; mul.ftz.f32 %f141, %f12, %f138; sub.ftz.f32 %f142, %f140, %f141; mul.ftz.f32 %f143, %f12, %f137; mul.ftz.f32 %f144, %f124, %f139; sub.ftz.f32 %f145, %f143, %f144; mul.ftz.f32 %f146, %f124, %f138; mul.ftz.f32 %f147, %f125, %f137; sub.ftz.f32 %f148, %f146, %f147; ld.local.f32 %f149, [%rd2+80]; mul.ftz.f32 %f569, %f149, %f137; mul.ftz.f32 %f570, %f149, %f138; mul.ftz.f32 %f571, %f149, %f139; mul.ftz.f32 %f572, %f149, %f142; mul.ftz.f32 %f573, %f149, %f145; mul.ftz.f32 %f574, %f149, %f148; setp.lt.u32 %p7, %r6, 16777216; @%p7 bra $L__BB197_18; ld.local.u32 %r9, [%rd2+44]; and.b32 %r4, %r9, 2; setp.eq.s32 %p8, %r4, 0; ld.local.s32 %rd3, [%rd2+64]; @%p8 bra $L__BB197_8; bra.uni $L__BB197_5; $L__BB197_8: cvt.u32.u64 %r12, %rd3; setp.ne.s32 %p11, %r12, -1; @%p11 bra $L__BB197_11; bra.uni $L__BB197_9; $L__BB197_11: ld.const.u64 %rd13, [kernel_params+2160]; cvta.to.global.u64 %rd14, %rd13; mul.lo.s64 %rd15, %rd3, 208; add.s64 %rd16, %rd14, %rd15; ld.global.v4.f32 {%f218, %f219, %f220, %f221}, [%rd16]; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd16+16]; ld.global.v4.f32 {%f232, %f233, %f234, %f235}, [%rd16+32]; mul.ftz.f32 %f239, %f8, %f225; fma.rn.ftz.f32 %f240, %f7, %f218, %f239; fma.rn.ftz.f32 %f241, %f9, %f232, %f240; mul.ftz.f32 %f242, %f8, %f226; fma.rn.ftz.f32 %f243, %f7, %f219, %f242; fma.rn.ftz.f32 %f244, %f9, %f233, %f243; mul.ftz.f32 %f245, %f8, %f227; fma.rn.ftz.f32 %f246, %f7, %f220, %f245; fma.rn.ftz.f32 %f247, %f9, %f234, %f246; mul.ftz.f32 %f248, %f244, %f244; fma.rn.ftz.f32 %f249, %f241, %f241, %f248; fma.rn.ftz.f32 %f250, %f247, %f247, %f249; sqrt.approx.ftz.f32 %f251, %f250; div.approx.ftz.f32 %f7, %f241, %f251; div.approx.ftz.f32 %f8, %f244, %f251; div.approx.ftz.f32 %f9, %f247, %f251; bra.uni $L__BB197_12; $L__BB197_5: cvt.u32.u64 %r10, %rd3; setp.ne.s32 %p9, %r10, -1; @%p9 bra $L__BB197_7; ld.local.u32 %r11, [%rd2+52]; setp.ne.s32 %p10, %r11, 32; @%p10 bra $L__BB197_12; $L__BB197_7: ld.local.v4.f32 {%f150, %f151, %f152, %f153}, [%rd2+132]; ld.local.v4.f32 {%f157, %f158, %f159, %f160}, [%rd2+148]; ld.local.v4.f32 {%f164, %f165, %f166, %f167}, [%rd2+164]; mul.ftz.f32 %f171, %f8, %f157; fma.rn.ftz.f32 %f172, %f7, %f150, %f171; fma.rn.ftz.f32 %f173, %f9, %f164, %f172; mul.ftz.f32 %f174, %f8, %f158; fma.rn.ftz.f32 %f175, %f7, %f151, %f174; fma.rn.ftz.f32 %f176, %f9, %f165, %f175; mul.ftz.f32 %f177, %f8, %f159; fma.rn.ftz.f32 %f178, %f7, %f152, %f177; fma.rn.ftz.f32 %f179, %f9, %f166, %f178; mul.ftz.f32 %f180, %f176, %f176; fma.rn.ftz.f32 %f181, %f173, %f173, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; sqrt.approx.ftz.f32 %f183, %f182; div.approx.ftz.f32 %f7, %f173, %f183; div.approx.ftz.f32 %f8, %f176, %f183; div.approx.ftz.f32 %f9, %f179, %f183; $L__BB197_12: @%p8 bra $L__BB197_14; bra.uni $L__BB197_13; $L__BB197_14: ld.const.u64 %rd17, [kernel_params+2160]; cvta.to.global.u64 %rd18, %rd17; mul.lo.s64 %rd19, %rd3, 208; add.s64 %rd20, %rd18, %rd19; ld.global.v4.f32 {%f279, %f280, %f281, %f282}, [%rd20+48]; ld.global.v4.f32 {%f286, %f287, %f288, %f289}, [%rd20+64]; ld.global.v4.f32 {%f293, %f294, %f295, %f296}, [%rd20+80]; mul.ftz.f32 %f300, %f570, %f280; fma.rn.ftz.f32 %f301, %f569, %f279, %f300; fma.rn.ftz.f32 %f566, %f571, %f281, %f301; mul.ftz.f32 %f302, %f570, %f287; fma.rn.ftz.f32 %f303, %f569, %f286, %f302; fma.rn.ftz.f32 %f567, %f571, %f288, %f303; mul.ftz.f32 %f304, %f570, %f294; fma.rn.ftz.f32 %f305, %f569, %f293, %f304; fma.rn.ftz.f32 %f571, %f571, %f295, %f305; bra.uni $L__BB197_15; $L__BB197_13: ld.local.v4.f32 {%f252, %f253, %f254, %f255}, [%rd2+180]; mul.ftz.f32 %f259, %f570, %f253; fma.rn.ftz.f32 %f260, %f569, %f252, %f259; fma.rn.ftz.f32 %f566, %f571, %f254, %f260; ld.local.v4.f32 {%f261, %f262, %f263, %f264}, [%rd2+196]; mul.ftz.f32 %f268, %f570, %f262; fma.rn.ftz.f32 %f269, %f569, %f261, %f268; fma.rn.ftz.f32 %f567, %f571, %f263, %f269; ld.local.v4.f32 {%f270, %f271, %f272, %f273}, [%rd2+212]; mul.ftz.f32 %f277, %f570, %f271; fma.rn.ftz.f32 %f278, %f569, %f270, %f277; fma.rn.ftz.f32 %f571, %f571, %f272, %f278; $L__BB197_15: @%p8 bra $L__BB197_17; bra.uni $L__BB197_16; $L__BB197_17: ld.const.u64 %rd21, [kernel_params+2160]; cvta.to.global.u64 %rd22, %rd21; mul.lo.s64 %rd23, %rd3, 208; add.s64 %rd24, %rd22, %rd23; ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [%rd24+48]; ld.global.v4.f32 {%f340, %f341, %f342, %f343}, [%rd24+64]; ld.global.v4.f32 {%f347, %f348, %f349, %f350}, [%rd24+80]; mul.ftz.f32 %f354, %f573, %f334; fma.rn.ftz.f32 %f355, %f572, %f333, %f354; fma.rn.ftz.f32 %f43, %f574, %f335, %f355; mul.ftz.f32 %f356, %f573, %f341; fma.rn.ftz.f32 %f357, %f572, %f340, %f356; fma.rn.ftz.f32 %f44, %f574, %f342, %f357; mul.ftz.f32 %f358, %f573, %f348; fma.rn.ftz.f32 %f359, %f572, %f347, %f358; fma.rn.ftz.f32 %f574, %f574, %f349, %f359; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f43; mov.f32 %f573, %f44; bra.uni $L__BB197_18; $L__BB197_16: ld.local.v4.f32 {%f306, %f307, %f308, %f309}, [%rd2+180]; mul.ftz.f32 %f313, %f573, %f307; fma.rn.ftz.f32 %f314, %f572, %f306, %f313; fma.rn.ftz.f32 %f40, %f574, %f308, %f314; ld.local.v4.f32 {%f315, %f316, %f317, %f318}, [%rd2+196]; mul.ftz.f32 %f322, %f573, %f316; fma.rn.ftz.f32 %f323, %f572, %f315, %f322; fma.rn.ftz.f32 %f41, %f574, %f317, %f323; ld.local.v4.f32 {%f324, %f325, %f326, %f327}, [%rd2+212]; mul.ftz.f32 %f331, %f573, %f325; fma.rn.ftz.f32 %f332, %f572, %f324, %f331; fma.rn.ftz.f32 %f574, %f574, %f326, %f332; mov.f32 %f569, %f566; mov.f32 %f570, %f567; mov.f32 %f572, %f40; mov.f32 %f573, %f41; $L__BB197_18: mul.ftz.f32 %f360, %f574, %f8; mul.ftz.f32 %f361, %f573, %f9; sub.ftz.f32 %f362, %f361, %f360; mul.ftz.f32 %f363, %f572, %f9; mul.ftz.f32 %f364, %f574, %f7; sub.ftz.f32 %f365, %f364, %f363; mul.ftz.f32 %f366, %f573, %f7; mul.ftz.f32 %f367, %f572, %f8; sub.ftz.f32 %f368, %f367, %f366; mul.ftz.f32 %f369, %f570, %f9; mul.ftz.f32 %f370, %f571, %f8; sub.ftz.f32 %f371, %f370, %f369; mul.ftz.f32 %f372, %f571, %f7; mul.ftz.f32 %f373, %f569, %f9; sub.ftz.f32 %f374, %f373, %f372; mul.ftz.f32 %f375, %f569, %f8; mul.ftz.f32 %f376, %f570, %f7; sub.ftz.f32 %f377, %f376, %f375; shl.b32 %r15, %r7, 2; cvt.u64.u32 %rd25, %r15; and.b64 %rd26, %rd25, 1020; add.s64 %rd27, %rd1, %rd26; shr.u32 %r16, %r7, 6; cvt.u64.u32 %rd28, %r16; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd1, %rd29; shr.u32 %r17, %r7, 14; cvt.u64.u32 %rd31, %r17; and.b64 %rd32, %rd31, 1020; add.s64 %rd33, %rd1, %rd32; mul.ftz.f32 %f378, %f570, %f365; fma.rn.ftz.f32 %f379, %f569, %f362, %f378; fma.rn.ftz.f32 %f380, %f571, %f368, %f379; ld.local.f32 %f381, [%rd30]; ld.local.f32 %f382, [%rd27]; sub.ftz.f32 %f383, %f381, %f382; ld.local.f32 %f384, [%rd33]; sub.ftz.f32 %f385, %f384, %f382; mul.ftz.f32 %f386, %f371, %f385; mul.ftz.f32 %f387, %f374, %f385; mul.ftz.f32 %f388, %f377, %f385; fma.rn.ftz.f32 %f389, %f362, %f383, %f386; fma.rn.ftz.f32 %f390, %f365, %f383, %f387; fma.rn.ftz.f32 %f391, %f368, %f383, %f388; abs.ftz.f32 %f392, %f380; shr.u32 %r18, %r7, 24; mul.wide.u32 %rd34, %r18, 4; add.s64 %rd35, %rd1, %rd34; shr.u32 %r19, %r6, 6; cvt.u64.u32 %rd36, %r19; and.b64 %rd37, %rd36, 1020; add.s64 %rd38, %rd1, %rd37; ld.local.f32 %f393, [%rd38]; neg.ftz.f32 %f394, %f393; and.b32 %r20, %r6, 16711680; setp.eq.s32 %p15, %r20, 0; selp.f32 %f395, %f393, %f394, %p15; ld.local.f32 %f55, [%rd35]; mul.ftz.f32 %f396, %f7, %f392; mul.ftz.f32 %f397, %f8, %f392; mul.ftz.f32 %f398, %f9, %f392; setp.lt.ftz.f32 %p16, %f380, 0f00000000; selp.f32 %f399, 0fBF800000, 0f3F800000, %p16; mul.ftz.f32 %f400, %f395, %f399; mul.ftz.f32 %f401, %f389, %f400; mul.ftz.f32 %f402, %f390, %f400; mul.ftz.f32 %f403, %f391, %f400; sub.ftz.f32 %f578, %f396, %f401; sub.ftz.f32 %f579, %f397, %f402; sub.ftz.f32 %f580, %f398, %f403; mul.ftz.f32 %f404, %f579, %f579; fma.rn.ftz.f32 %f405, %f578, %f578, %f404; fma.rn.ftz.f32 %f406, %f580, %f580, %f405; sqrt.approx.ftz.f32 %f59, %f406; setp.neu.ftz.f32 %p17, %f59, 0f00000000; @%p17 bra $L__BB197_19; bra.uni $L__BB197_20; $L__BB197_19: rcp.approx.ftz.f32 %f407, %f59; mul.ftz.f32 %f580, %f407, %f580; mul.ftz.f32 %f579, %f407, %f579; mul.ftz.f32 %f578, %f407, %f578; $L__BB197_20: setp.eq.ftz.f32 %p18, %f578, 0f00000000; setp.eq.ftz.f32 %p19, %f579, 0f00000000; and.pred %p20, %p18, %p19; setp.eq.ftz.f32 %p21, %f580, 0f00000000; and.pred %p22, %p21, %p20; @%p22 bra $L__BB197_22; mov.f32 %f408, 0f00000000; max.ftz.f32 %f409, %f55, %f408; mul.ftz.f32 %f410, %f409, %f578; mov.f32 %f411, 0f3F800000; sub.ftz.f32 %f412, %f411, %f409; fma.rn.ftz.f32 %f413, %f7, %f412, %f410; mul.ftz.f32 %f414, %f409, %f579; fma.rn.ftz.f32 %f415, %f8, %f412, %f414; mul.ftz.f32 %f416, %f409, %f580; fma.rn.ftz.f32 %f417, %f9, %f412, %f416; mul.ftz.f32 %f418, %f415, %f415; fma.rn.ftz.f32 %f419, %f413, %f413, %f418; fma.rn.ftz.f32 %f420, %f417, %f417, %f419; sqrt.approx.ftz.f32 %f421, %f420; div.approx.ftz.f32 %f7, %f413, %f421; div.approx.ftz.f32 %f8, %f415, %f421; div.approx.ftz.f32 %f9, %f417, %f421; $L__BB197_22: @%p7 bra $L__BB197_26; ld.local.u8 %rs1, [%rd2+44]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p24, %rs2, 0; @%p24 bra $L__BB197_25; bra.uni $L__BB197_24; $L__BB197_25: ld.const.u64 %rd39, [kernel_params+2160]; cvta.to.global.u64 %rd40, %rd39; ld.local.u32 %r21, [%rd2+64]; mul.wide.s32 %rd41, %r21, 208; add.s64 %rd42, %rd40, %rd41; ld.global.v4.f32 {%f456, %f457, %f458, %f459}, [%rd42+48]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd42+64]; ld.global.v4.f32 {%f470, %f471, %f472, %f473}, [%rd42+80]; mul.ftz.f32 %f477, %f8, %f463; fma.rn.ftz.f32 %f478, %f7, %f456, %f477; fma.rn.ftz.f32 %f479, %f9, %f470, %f478; mul.ftz.f32 %f480, %f8, %f464; fma.rn.ftz.f32 %f481, %f7, %f457, %f480; fma.rn.ftz.f32 %f482, %f9, %f471, %f481; mul.ftz.f32 %f483, %f8, %f465; fma.rn.ftz.f32 %f484, %f7, %f458, %f483; fma.rn.ftz.f32 %f485, %f9, %f472, %f484; mul.ftz.f32 %f486, %f482, %f482; fma.rn.ftz.f32 %f487, %f479, %f479, %f486; fma.rn.ftz.f32 %f488, %f485, %f485, %f487; sqrt.approx.ftz.f32 %f489, %f488; div.approx.ftz.f32 %f7, %f479, %f489; div.approx.ftz.f32 %f8, %f482, %f489; div.approx.ftz.f32 %f9, %f485, %f489; bra.uni $L__BB197_26; $L__BB197_24: ld.local.v4.f32 {%f422, %f423, %f424, %f425}, [%rd2+180]; ld.local.v4.f32 {%f429, %f430, %f431, %f432}, [%rd2+196]; ld.local.v4.f32 {%f436, %f437, %f438, %f439}, [%rd2+212]; mul.ftz.f32 %f443, %f8, %f429; fma.rn.ftz.f32 %f444, %f7, %f422, %f443; fma.rn.ftz.f32 %f445, %f9, %f436, %f444; mul.ftz.f32 %f446, %f8, %f430; fma.rn.ftz.f32 %f447, %f7, %f423, %f446; fma.rn.ftz.f32 %f448, %f9, %f437, %f447; mul.ftz.f32 %f449, %f8, %f431; fma.rn.ftz.f32 %f450, %f7, %f424, %f449; fma.rn.ftz.f32 %f451, %f9, %f438, %f450; mul.ftz.f32 %f452, %f448, %f448; fma.rn.ftz.f32 %f453, %f445, %f445, %f452; fma.rn.ftz.f32 %f454, %f451, %f451, %f453; sqrt.approx.ftz.f32 %f455, %f454; div.approx.ftz.f32 %f7, %f445, %f455; div.approx.ftz.f32 %f8, %f448, %f455; div.approx.ftz.f32 %f9, %f451, %f455; $L__BB197_26: ld.local.f32 %f87, [%rd2+24]; ld.local.f32 %f88, [%rd2+28]; mul.ftz.f32 %f490, %f8, %f88; fma.rn.ftz.f32 %f491, %f7, %f87, %f490; ld.local.f32 %f89, [%rd2+32]; fma.rn.ftz.f32 %f492, %f9, %f89, %f491; add.ftz.f32 %f493, %f492, %f492; mul.ftz.f32 %f494, %f7, %f493; mul.ftz.f32 %f495, %f8, %f493; mul.ftz.f32 %f496, %f9, %f493; sub.ftz.f32 %f497, %f494, %f87; sub.ftz.f32 %f498, %f495, %f88; sub.ftz.f32 %f499, %f496, %f89; mul.ftz.f32 %f500, %f125, %f88; fma.rn.ftz.f32 %f501, %f124, %f87, %f500; fma.rn.ftz.f32 %f90, %f12, %f89, %f501; mul.ftz.f32 %f502, %f90, 0f3F666666; mov.f32 %f503, 0f3C23D70A; min.ftz.f32 %f91, %f502, %f503; mul.ftz.f32 %f504, %f125, %f498; fma.rn.ftz.f32 %f505, %f124, %f497, %f504; fma.rn.ftz.f32 %f506, %f12, %f499, %f505; setp.ltu.ftz.f32 %p25, %f506, %f91; @%p25 bra $L__BB197_27; bra.uni $L__BB197_34; $L__BB197_27: mul.ftz.f32 %f507, %f8, %f125; fma.rn.ftz.f32 %f508, %f7, %f124, %f507; fma.rn.ftz.f32 %f509, %f9, %f12, %f508; mul.ftz.f32 %f510, %f124, %f509; mul.ftz.f32 %f511, %f125, %f509; mul.ftz.f32 %f512, %f12, %f509; sub.ftz.f32 %f513, %f7, %f510; sub.ftz.f32 %f514, %f8, %f511; sub.ftz.f32 %f515, %f9, %f512; mul.ftz.f32 %f516, %f514, %f514; fma.rn.ftz.f32 %f517, %f513, %f513, %f516; fma.rn.ftz.f32 %f518, %f515, %f515, %f517; sqrt.approx.ftz.f32 %f519, %f518; div.approx.ftz.f32 %f95, %f513, %f519; div.approx.ftz.f32 %f96, %f514, %f519; div.approx.ftz.f32 %f97, %f515, %f519; mul.ftz.f32 %f520, %f88, %f96; fma.rn.ftz.f32 %f521, %f87, %f95, %f520; fma.rn.ftz.f32 %f98, %f89, %f97, %f521; mul.ftz.f32 %f522, %f98, %f98; fma.rn.ftz.f32 %f523, %f90, %f90, %f522; mul.ftz.f32 %f524, %f91, %f91; sub.ftz.f32 %f525, %f523, %f524; mul.ftz.f32 %f526, %f522, %f525; mov.f32 %f527, 0f00000000; max.ftz.f32 %f528, %f526, %f527; sqrt.approx.ftz.f32 %f529, %f528; fma.rn.ftz.f32 %f530, %f91, %f90, %f523; mov.f32 %f531, 0f3F000000; div.approx.ftz.f32 %f532, %f531, %f523; add.ftz.f32 %f533, %f529, %f530; mul.ftz.f32 %f99, %f532, %f533; sub.ftz.f32 %f534, %f530, %f529; mul.ftz.f32 %f100, %f532, %f534; setp.gt.ftz.f32 %p26, %f99, 0f3727C5AC; setp.le.ftz.f32 %p27, %f99, 0f3F800054; and.pred %p1, %p26, %p27; setp.gt.ftz.f32 %p28, %f100, 0f3727C5AC; setp.le.ftz.f32 %p29, %f100, 0f3F800054; and.pred %p2, %p28, %p29; and.pred %p30, %p1, %p2; @%p30 bra $L__BB197_30; bra.uni $L__BB197_28; $L__BB197_30: mov.f32 %f541, 0f3F800000; sub.ftz.f32 %f542, %f541, %f99; max.ftz.f32 %f544, %f542, %f527; sqrt.approx.ftz.f32 %f106, %f544; max.ftz.f32 %f545, %f99, %f527; sqrt.approx.ftz.f32 %f107, %f545; sub.ftz.f32 %f546, %f541, %f100; max.ftz.f32 %f547, %f546, %f527; sqrt.approx.ftz.f32 %f108, %f547; max.ftz.f32 %f548, %f100, %f527; sqrt.approx.ftz.f32 %f109, %f548; mul.ftz.f32 %f549, %f90, %f107; fma.rn.ftz.f32 %f550, %f98, %f106, %f549; add.ftz.f32 %f551, %f550, %f550; mul.ftz.f32 %f552, %f107, %f551; sub.ftz.f32 %f110, %f552, %f90; mul.ftz.f32 %f553, %f90, %f109; fma.rn.ftz.f32 %f554, %f98, %f108, %f553; add.ftz.f32 %f555, %f554, %f554; mul.ftz.f32 %f556, %f109, %f555; sub.ftz.f32 %f111, %f556, %f90; setp.ge.ftz.f32 %p32, %f110, 0f3727C5AC; setp.ge.ftz.f32 %p33, %f111, 0f3727C5AC; and.pred %p34, %p32, %p33; @%p34 bra $L__BB197_32; bra.uni $L__BB197_31; $L__BB197_32: setp.lt.ftz.f32 %p36, %f110, %f111; selp.f32 %f587, %f106, %f108, %p36; selp.f32 %f588, %f107, %f109, %p36; bra.uni $L__BB197_33; $L__BB197_28: or.pred %p31, %p1, %p2; mov.f32 %f7, %f124; mov.f32 %f8, %f125; mov.f32 %f9, %f12; @%p31 bra $L__BB197_29; bra.uni $L__BB197_34; $L__BB197_29: selp.f32 %f535, %f99, %f100, %p1; mov.f32 %f536, 0f3F800000; sub.ftz.f32 %f537, %f536, %f535; max.ftz.f32 %f539, %f537, %f527; sqrt.approx.ftz.f32 %f587, %f539; max.ftz.f32 %f540, %f535, %f527; sqrt.approx.ftz.f32 %f588, %f540; bra.uni $L__BB197_33; $L__BB197_31: setp.gt.ftz.f32 %p35, %f110, %f111; selp.f32 %f587, %f106, %f108, %p35; selp.f32 %f588, %f107, %f109, %p35; $L__BB197_33: mul.ftz.f32 %f557, %f124, %f588; mul.ftz.f32 %f558, %f125, %f588; mul.ftz.f32 %f559, %f12, %f588; fma.rn.ftz.f32 %f9, %f97, %f587, %f559; fma.rn.ftz.f32 %f8, %f96, %f587, %f558; fma.rn.ftz.f32 %f7, %f95, %f587, %f557; $L__BB197_34: mul.wide.u32 %rd43, %r8, 4; add.s64 %rd44, %rd1, %rd43; st.local.f32 [%rd44], %f7; st.local.f32 [%rd44+4], %f8; st.local.f32 [%rd44+8], %f9; ret; $L__BB197_9: ld.local.u32 %r13, [%rd2+52]; setp.ne.s32 %p12, %r13, 32; @%p12 bra $L__BB197_12; ld.const.u64 %rd9, [kernel_params+2368]; cvta.to.global.u64 %rd10, %rd9; ld.local.u32 %r14, [%rd2+68]; mul.wide.s32 %rd11, %r14, 192; add.s64 %rd12, %rd10, %rd11; ld.global.v4.f32 {%f184, %f185, %f186, %f187}, [%rd12+48]; ld.global.v4.f32 {%f191, %f192, %f193, %f194}, [%rd12+64]; ld.global.v4.f32 {%f198, %f199, %f200, %f201}, [%rd12+80]; mul.ftz.f32 %f205, %f8, %f191; fma.rn.ftz.f32 %f206, %f7, %f184, %f205; fma.rn.ftz.f32 %f207, %f9, %f198, %f206; mul.ftz.f32 %f208, %f8, %f192; fma.rn.ftz.f32 %f209, %f7, %f185, %f208; fma.rn.ftz.f32 %f210, %f9, %f199, %f209; mul.ftz.f32 %f211, %f8, %f193; fma.rn.ftz.f32 %f212, %f7, %f186, %f211; fma.rn.ftz.f32 %f213, %f9, %f200, %f212; mul.ftz.f32 %f214, %f210, %f210; fma.rn.ftz.f32 %f215, %f207, %f207, %f214; fma.rn.ftz.f32 %f216, %f213, %f213, %f215; sqrt.approx.ftz.f32 %f217, %f216; div.approx.ftz.f32 %f7, %f207, %f217; div.approx.ftz.f32 %f8, %f210, %f217; div.approx.ftz.f32 %f9, %f213, %f217; bra.uni $L__BB197_12; } .func _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj( .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2, .param .b64 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5, .param .b32 _Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6 ) { .reg .pred %p<37>; .reg .b16 %rs<6>; .reg .f32 %f<21>; .reg .b32 %r<21>; .reg .b64 %rd<55>; ld.param.u32 %r1, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_0]; ld.param.u32 %r2, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_1]; ld.param.f32 %f20, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_2]; ld.param.u64 %rd6, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_3]; ld.param.u32 %r3, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_4]; ld.param.u32 %r4, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_5]; ld.param.u32 %r5, [_Z19svm_node_light_pathILj447EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj_param_6]; mov.f32 %f17, 0f00000000; setp.gt.s32 %p1, %r3, 6; @%p1 bra $L__BB198_13; bra.uni $L__BB198_1; $L__BB198_13: setp.gt.s32 %p2, %r3, 10; @%p2 bra $L__BB198_23; setp.gt.s32 %p8, %r3, 8; @%p8 bra $L__BB198_18; setp.eq.s32 %p11, %r3, 7; @%p11 bra $L__BB198_40; setp.eq.s32 %p12, %r3, 8; mov.f32 %f20, %f17; @%p12 bra $L__BB198_17; bra.uni $L__BB198_45; $L__BB198_17: and.b32 %r12, %r2, 1; setp.eq.b32 %p28, %r12, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p28; bra.uni $L__BB198_45; $L__BB198_1: setp.gt.s32 %p13, %r3, 2; @%p13 bra $L__BB198_6; bra.uni $L__BB198_2; $L__BB198_6: setp.gt.s32 %p14, %r3, 4; @%p14 bra $L__BB198_10; setp.eq.s32 %p17, %r3, 3; @%p17 bra $L__BB198_42; setp.eq.s32 %p18, %r3, 4; mov.f32 %f20, %f17; @%p18 bra $L__BB198_9; bra.uni $L__BB198_45; $L__BB198_9: and.b32 %r16, %r5, 32; setp.eq.s32 %p32, %r16, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p32; bra.uni $L__BB198_45; $L__BB198_23: setp.gt.s32 %p3, %r3, 12; @%p3 bra $L__BB198_29; setp.eq.s32 %p6, %r3, 11; @%p6 bra $L__BB198_37; setp.eq.s32 %p7, %r3, 12; mov.f32 %f20, %f17; @%p7 bra $L__BB198_26; bra.uni $L__BB198_45; $L__BB198_26: and.b32 %r8, %r5, 768; mov.u64 %rd24, kernel_params; add.s64 %rd52, %rd24, 3400; setp.ne.s32 %p24, %r8, 0; @%p24 bra $L__BB198_28; add.s64 %rd52, %rd24, 2488; $L__BB198_28: ld.const.u64 %rd27, [%rd52]; cvta.to.global.u64 %rd28, %rd27; mul.wide.s32 %rd29, %r1, 2; add.s64 %rd30, %rd28, %rd29; ld.global.u16 %rs3, [%rd30]; cvt.rn.f32.u16 %f20, %rs3; bra.uni $L__BB198_45; $L__BB198_2: setp.eq.s32 %p19, %r3, 0; @%p19 bra $L__BB198_44; setp.eq.s32 %p20, %r3, 1; @%p20 bra $L__BB198_43; setp.eq.s32 %p21, %r3, 2; mov.f32 %f20, %f17; @%p21 bra $L__BB198_5; bra.uni $L__BB198_45; $L__BB198_5: and.b32 %r18, %r5, 8; setp.eq.s32 %p34, %r18, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p34; bra.uni $L__BB198_45; $L__BB198_18: setp.eq.s32 %p9, %r3, 9; @%p9 bra $L__BB198_45; setp.eq.s32 %p10, %r3, 10; mov.f32 %f20, %f17; @%p10 bra $L__BB198_20; bra.uni $L__BB198_45; $L__BB198_20: and.b32 %r10, %r5, 768; mov.u64 %rd40, kernel_params; add.s64 %rd54, %rd40, 3376; setp.ne.s32 %p26, %r10, 0; @%p26 bra $L__BB198_22; add.s64 %rd54, %rd40, 2464; $L__BB198_22: ld.const.u64 %rd43, [%rd54]; cvta.to.global.u64 %rd44, %rd43; mul.wide.s32 %rd45, %r1, 2; add.s64 %rd46, %rd44, %rd45; ld.global.u16 %rs5, [%rd46]; cvt.rn.f32.u16 %f18, %rs5; and.b32 %r11, %r5, 525056; setp.eq.s32 %p27, %r11, 0; add.ftz.f32 %f19, %f18, 0f3F800000; selp.f32 %f20, %f18, %f19, %p27; bra.uni $L__BB198_45; $L__BB198_10: setp.eq.s32 %p15, %r3, 5; @%p15 bra $L__BB198_41; setp.eq.s32 %p16, %r3, 6; mov.f32 %f20, %f17; @%p16 bra $L__BB198_12; bra.uni $L__BB198_45; $L__BB198_12: and.b32 %r14, %r5, 4; setp.eq.s32 %p30, %r14, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p30; bra.uni $L__BB198_45; $L__BB198_29: setp.eq.s32 %p4, %r3, 13; @%p4 bra $L__BB198_34; setp.ne.s32 %p5, %r3, 14; mov.f32 %f20, %f17; @%p5 bra $L__BB198_45; and.b32 %r6, %r5, 768; mov.u64 %rd8, kernel_params; add.s64 %rd50, %rd8, 3408; setp.ne.s32 %p22, %r6, 0; @%p22 bra $L__BB198_33; add.s64 %rd50, %rd8, 2496; $L__BB198_33: ld.const.u64 %rd11, [%rd50]; cvta.to.global.u64 %rd12, %rd11; mul.wide.s32 %rd13, %r1, 2; add.s64 %rd14, %rd12, %rd13; ld.global.u16 %rs1, [%rd14]; cvt.rn.f32.u16 %f20, %rs1; bra.uni $L__BB198_45; $L__BB198_40: and.b32 %r13, %r5, 128; setp.eq.s32 %p29, %r13, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p29; bra.uni $L__BB198_45; $L__BB198_42: and.b32 %r17, %r5, 16; setp.eq.s32 %p33, %r17, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p33; bra.uni $L__BB198_45; $L__BB198_37: and.b32 %r9, %r5, 768; mov.u64 %rd32, kernel_params; add.s64 %rd53, %rd32, 3392; setp.ne.s32 %p25, %r9, 0; @%p25 bra $L__BB198_39; add.s64 %rd53, %rd32, 2480; $L__BB198_39: ld.const.u64 %rd35, [%rd53]; cvta.to.global.u64 %rd36, %rd35; mul.wide.s32 %rd37, %r1, 2; add.s64 %rd38, %rd36, %rd37; ld.global.u16 %rs4, [%rd38]; cvt.rn.f32.u16 %f20, %rs4; bra.uni $L__BB198_45; $L__BB198_44: and.b32 %r20, %r5, 1; setp.eq.b32 %p36, %r20, 1; selp.f32 %f20, 0f3F800000, 0f00000000, %p36; bra.uni $L__BB198_45; $L__BB198_43: and.b32 %r19, %r5, 768; setp.eq.s32 %p35, %r19, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p35; bra.uni $L__BB198_45; $L__BB198_41: and.b32 %r15, %r5, 2; setp.eq.s32 %p31, %r15, 0; selp.f32 %f20, 0f00000000, 0f3F800000, %p31; bra.uni $L__BB198_45; $L__BB198_34: and.b32 %r7, %r5, 768; mov.u64 %rd16, kernel_params; add.s64 %rd51, %rd16, 3384; setp.ne.s32 %p23, %r7, 0; @%p23 bra $L__BB198_36; add.s64 %rd51, %rd16, 2472; $L__BB198_36: ld.const.u64 %rd19, [%rd51]; cvta.to.global.u64 %rd20, %rd19; mul.wide.s32 %rd21, %r1, 2; add.s64 %rd22, %rd20, %rd21; ld.global.u16 %rs2, [%rd22]; cvt.rn.f32.u16 %f20, %rs2; $L__BB198_45: cvta.to.local.u64 %rd47, %rd6; mul.wide.u32 %rd48, %r4, 4; add.s64 %rd49, %rd47, %rd48; st.local.f32 [%rd49], %f20; ret; } .func (.param .b32 func_retval0) _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji( .param .b64 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3, .param .b32 _Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4 ) { .reg .pred %p<226>; .reg .f32 %f<3395>; .reg .b32 %r<4380>; .reg .b64 %rd<46>; ld.param.u64 %rd1, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_0]; ld.param.u32 %r389, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_2]; ld.param.u32 %r391, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; cvta.to.local.u64 %rd2, %rd1; ld.const.u64 %rd3, [kernel_params+2400]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r391, 16; add.s64 %rd6, %rd4, %rd5; ld.global.v4.u32 {%r392, %r393, %r394, %r395}, [%rd6]; ld.global.v4.u32 {%r396, %r397, %r398, %r399}, [%rd6+16]; bfe.u32 %r11, %r392, 8, 8; shl.b32 %r400, %r392, 2; cvt.u64.u32 %rd7, %r400; and.b64 %rd8, %rd7, 1020; add.s64 %rd9, %rd2, %rd8; ld.local.f32 %f1, [%rd9]; ld.local.f32 %f2, [%rd9+4]; ld.local.f32 %f3, [%rd9+8]; setp.eq.s32 %p1, %r11, 255; @%p1 bra $L__BB199_2; bra.uni $L__BB199_1; $L__BB199_2: mov.b32 %f2764, %r395; bra.uni $L__BB199_3; $L__BB199_1: mul.wide.u32 %rd11, %r11, 4; add.s64 %rd12, %rd2, %rd11; ld.local.f32 %f2764, [%rd12]; $L__BB199_3: bfe.u32 %r4304, %r392, 16, 8; setp.eq.s32 %p2, %r4304, 255; @%p2 bra $L__BB199_5; bra.uni $L__BB199_4; $L__BB199_5: mov.b32 %f9, %r396; bra.uni $L__BB199_6; $L__BB199_4: bfe.u32 %r4305, %r392, 16, 8; mul.wide.u32 %rd14, %r4305, 4; add.s64 %rd15, %rd2, %rd14; ld.local.f32 %f9, [%rd15]; $L__BB199_6: shr.u32 %r4306, %r392, 24; setp.eq.s32 %p3, %r4306, 255; @%p3 bra $L__BB199_8; bra.uni $L__BB199_7; $L__BB199_8: mov.b32 %f2766, %r397; bra.uni $L__BB199_9; $L__BB199_7: shr.u32 %r4307, %r392, 24; mul.wide.u32 %rd17, %r4307, 4; add.s64 %rd18, %rd2, %rd17; ld.local.f32 %f2766, [%rd18]; $L__BB199_9: and.b32 %r401, %r393, 255; setp.eq.s32 %p4, %r401, 255; @%p4 bra $L__BB199_11; bra.uni $L__BB199_10; $L__BB199_11: mov.b32 %f15, %r398; bra.uni $L__BB199_12; $L__BB199_10: shl.b32 %r402, %r393, 2; cvt.u64.u32 %rd20, %r402; and.b64 %rd21, %rd20, 1020; add.s64 %rd22, %rd2, %rd21; ld.local.f32 %f15, [%rd22]; $L__BB199_12: bfe.u32 %r403, %r393, 8, 8; setp.eq.s32 %p5, %r403, 255; @%p5 bra $L__BB199_14; bra.uni $L__BB199_13; $L__BB199_14: mov.b32 %f2768, %r399; bra.uni $L__BB199_15; $L__BB199_13: shr.u32 %r404, %r393, 6; cvt.u64.u32 %rd24, %r404; and.b64 %rd25, %rd24, 1020; add.s64 %rd26, %rd2, %rd25; ld.local.f32 %f2768, [%rd26]; $L__BB199_15: ld.param.u32 %r4308, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; mov.f32 %f1181, 0f00000000; max.ftz.f32 %f1182, %f2768, %f1181; mov.f32 %f1183, 0f3F800000; min.ftz.f32 %f19, %f1182, %f1183; mov.f32 %f1184, 0f40000000; div.approx.ftz.f32 %f1185, %f2766, %f1184; max.ftz.f32 %f1186, %f1185, %f1181; mov.f32 %f1187, 0f3F000000; min.ftz.f32 %f20, %f1186, %f1187; mul.ftz.f32 %f21, %f2764, %f9; mul.ftz.f32 %f22, %f1, %f9; mul.ftz.f32 %f23, %f2, %f9; mul.ftz.f32 %f24, %f3, %f9; setp.gt.s32 %p6, %r4308, 2; @%p6 bra $L__BB199_26; ld.param.u32 %r4318, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p9, %r4318, 1; @%p9 bra $L__BB199_365; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4319, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p10, %r4319, 2; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p10 bra $L__BB199_18; bra.uni $L__BB199_385; $L__BB199_18: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p127, %r389, 1; @%p127 bra $L__BB199_259; bra.uni $L__BB199_19; $L__BB199_259: setp.eq.s32 %p128, %r389, 2; @%p128 bra $L__BB199_286; setp.eq.s32 %p129, %r389, 3; @%p129 bra $L__BB199_276; setp.ne.s32 %p130, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p130 bra $L__BB199_362; cvt.rmi.ftz.f32.f32 %f824, %f22; sub.ftz.f32 %f825, %f22, %f824; cvt.rmi.ftz.f32.f32 %f826, %f23; sub.ftz.f32 %f827, %f23, %f826; add.ftz.f32 %f2093, %f824, 0fBF800000; mov.b32 %r2985, %f2093; add.s32 %r305, %r2985, -559038716; add.s32 %r306, %r2985, -559038712; add.ftz.f32 %f2094, %f824, 0f00000000; mov.f32 %f3254, 0f00000000; mov.b32 %r2986, %f2094; add.s32 %r307, %r2986, -559038716; add.s32 %r308, %r2986, -559038712; add.ftz.f32 %f2095, %f824, 0f3F800000; mov.b32 %r2987, %f2095; add.s32 %r309, %r2987, -559038716; add.s32 %r310, %r2987, -559038712; mov.f32 %f3255, 0f41000000; mov.u32 %r4371, -2; mov.f32 %f3253, %f3254; mov.f32 %f3252, %f3254; mov.f32 %f3251, %f3254; $L__BB199_263: add.s32 %r4371, %r4371, 1; cvt.rn.f32.s32 %f833, %r4371; add.ftz.f32 %f2096, %f826, %f833; mov.b32 %r2988, %f2096; add.s32 %r313, %r2988, -559038716; xor.b32 %r2989, %r313, -559038716; shf.l.wrap.b32 %r2990, %r313, %r313, 14; sub.s32 %r314, %r2989, %r2990; shf.l.wrap.b32 %r315, %r314, %r314, 11; add.s32 %r2991, %r2988, -559038712; xor.b32 %r2992, %r2991, 506314504; shf.l.wrap.b32 %r2993, %r2991, %r2991, 14; sub.s32 %r316, %r2992, %r2993; shf.l.wrap.b32 %r317, %r316, %r316, 11; xor.b32 %r2994, %r305, %r314; sub.s32 %r2995, %r2994, %r315; xor.b32 %r2996, %r2995, %r313; shf.l.wrap.b32 %r2997, %r2995, %r2995, 25; sub.s32 %r2998, %r2996, %r2997; xor.b32 %r2999, %r2998, %r314; shf.l.wrap.b32 %r3000, %r2998, %r2998, 16; sub.s32 %r3001, %r2999, %r3000; xor.b32 %r3002, %r3001, %r2995; shf.l.wrap.b32 %r3003, %r3001, %r3001, 4; sub.s32 %r3004, %r3002, %r3003; xor.b32 %r3005, %r3004, %r2998; shf.l.wrap.b32 %r3006, %r3004, %r3004, 14; sub.s32 %r3007, %r3005, %r3006; xor.b32 %r3008, %r3007, %r3001; shf.l.wrap.b32 %r3009, %r3007, %r3007, 24; sub.s32 %r3010, %r3008, %r3009; cvt.rn.f32.u32 %f2097, %r3010; mul.ftz.f32 %f2098, %f2097, 0f2F800000; xor.b32 %r3011, %r306, %r316; sub.s32 %r3012, %r3011, %r317; xor.b32 %r3013, %r3012, %r2991; shf.l.wrap.b32 %r3014, %r3012, %r3012, 25; sub.s32 %r3015, %r3013, %r3014; xor.b32 %r3016, %r3015, %r316; shf.l.wrap.b32 %r3017, %r3015, %r3015, 16; sub.s32 %r3018, %r3016, %r3017; xor.b32 %r3019, %r3018, %r3012; shf.l.wrap.b32 %r3020, %r3018, %r3018, 4; sub.s32 %r3021, %r3019, %r3020; xor.b32 %r3022, %r3021, %r3015; shf.l.wrap.b32 %r3023, %r3021, %r3021, 14; sub.s32 %r3024, %r3022, %r3023; xor.b32 %r3025, %r3024, %r3018; shf.l.wrap.b32 %r3026, %r3024, %r3024, 24; sub.s32 %r3027, %r3025, %r3026; cvt.rn.f32.u32 %f2099, %r3027; mul.ftz.f32 %f2100, %f2099, 0f2F800000; fma.rn.ftz.f32 %f834, %f19, %f2098, 0fBF800000; fma.rn.ftz.f32 %f835, %f19, %f2100, %f833; sub.ftz.f32 %f2101, %f834, %f825; sub.ftz.f32 %f2102, %f835, %f827; mul.ftz.f32 %f2103, %f2102, %f2102; fma.rn.ftz.f32 %f2104, %f2101, %f2101, %f2103; sqrt.approx.ftz.f32 %f836, %f2104; setp.geu.ftz.f32 %p133, %f836, %f3255; @%p133 bra $L__BB199_265; mov.f32 %f3251, 0fBF800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f834; mov.f32 %f3254, %f835; mov.f32 %f3255, %f836; $L__BB199_265: xor.b32 %r3028, %r307, %r314; sub.s32 %r3029, %r3028, %r315; xor.b32 %r3030, %r3029, %r313; shf.l.wrap.b32 %r3031, %r3029, %r3029, 25; sub.s32 %r3032, %r3030, %r3031; xor.b32 %r3033, %r3032, %r314; shf.l.wrap.b32 %r3034, %r3032, %r3032, 16; sub.s32 %r3035, %r3033, %r3034; xor.b32 %r3036, %r3035, %r3029; shf.l.wrap.b32 %r3037, %r3035, %r3035, 4; sub.s32 %r3038, %r3036, %r3037; xor.b32 %r3039, %r3038, %r3032; shf.l.wrap.b32 %r3040, %r3038, %r3038, 14; sub.s32 %r3041, %r3039, %r3040; xor.b32 %r3042, %r3041, %r3035; shf.l.wrap.b32 %r3043, %r3041, %r3041, 24; sub.s32 %r3044, %r3042, %r3043; cvt.rn.f32.u32 %f2106, %r3044; mul.ftz.f32 %f2107, %f2106, 0f2F800000; xor.b32 %r3045, %r308, %r316; sub.s32 %r3046, %r3045, %r317; add.s32 %r3047, %r313, 4; xor.b32 %r3048, %r3046, %r3047; shf.l.wrap.b32 %r3049, %r3046, %r3046, 25; sub.s32 %r3050, %r3048, %r3049; xor.b32 %r3051, %r3050, %r316; shf.l.wrap.b32 %r3052, %r3050, %r3050, 16; sub.s32 %r3053, %r3051, %r3052; xor.b32 %r3054, %r3053, %r3046; shf.l.wrap.b32 %r3055, %r3053, %r3053, 4; sub.s32 %r3056, %r3054, %r3055; xor.b32 %r3057, %r3056, %r3050; shf.l.wrap.b32 %r3058, %r3056, %r3056, 14; sub.s32 %r3059, %r3057, %r3058; xor.b32 %r3060, %r3059, %r3053; shf.l.wrap.b32 %r3061, %r3059, %r3059, 24; sub.s32 %r3062, %r3060, %r3061; cvt.rn.f32.u32 %f2108, %r3062; mul.ftz.f32 %f2109, %f2108, 0f2F800000; fma.rn.ftz.f32 %f842, %f19, %f2107, 0f00000000; fma.rn.ftz.f32 %f843, %f19, %f2109, %f833; sub.ftz.f32 %f2110, %f842, %f825; sub.ftz.f32 %f2111, %f843, %f827; mul.ftz.f32 %f2112, %f2111, %f2111; fma.rn.ftz.f32 %f2113, %f2110, %f2110, %f2112; sqrt.approx.ftz.f32 %f844, %f2113; setp.geu.ftz.f32 %p134, %f844, %f3255; @%p134 bra $L__BB199_267; mov.f32 %f3251, 0f00000000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f842; mov.f32 %f3254, %f843; mov.f32 %f3255, %f844; $L__BB199_267: xor.b32 %r3063, %r309, %r314; sub.s32 %r3064, %r3063, %r315; xor.b32 %r3065, %r3064, %r313; shf.l.wrap.b32 %r3066, %r3064, %r3064, 25; sub.s32 %r3067, %r3065, %r3066; xor.b32 %r3068, %r3067, %r314; shf.l.wrap.b32 %r3069, %r3067, %r3067, 16; sub.s32 %r3070, %r3068, %r3069; xor.b32 %r3071, %r3070, %r3064; shf.l.wrap.b32 %r3072, %r3070, %r3070, 4; sub.s32 %r3073, %r3071, %r3072; xor.b32 %r3074, %r3073, %r3067; shf.l.wrap.b32 %r3075, %r3073, %r3073, 14; sub.s32 %r3076, %r3074, %r3075; xor.b32 %r3077, %r3076, %r3070; shf.l.wrap.b32 %r3078, %r3076, %r3076, 24; sub.s32 %r3079, %r3077, %r3078; cvt.rn.f32.u32 %f2115, %r3079; mul.ftz.f32 %f2116, %f2115, 0f2F800000; xor.b32 %r3080, %r310, %r316; sub.s32 %r3081, %r3080, %r317; xor.b32 %r3083, %r3081, %r3047; shf.l.wrap.b32 %r3084, %r3081, %r3081, 25; sub.s32 %r3085, %r3083, %r3084; xor.b32 %r3086, %r3085, %r316; shf.l.wrap.b32 %r3087, %r3085, %r3085, 16; sub.s32 %r3088, %r3086, %r3087; xor.b32 %r3089, %r3088, %r3081; shf.l.wrap.b32 %r3090, %r3088, %r3088, 4; sub.s32 %r3091, %r3089, %r3090; xor.b32 %r3092, %r3091, %r3085; shf.l.wrap.b32 %r3093, %r3091, %r3091, 14; sub.s32 %r3094, %r3092, %r3093; xor.b32 %r3095, %r3094, %r3088; shf.l.wrap.b32 %r3096, %r3094, %r3094, 24; sub.s32 %r3097, %r3095, %r3096; cvt.rn.f32.u32 %f2117, %r3097; mul.ftz.f32 %f2118, %f2117, 0f2F800000; fma.rn.ftz.f32 %f850, %f19, %f2116, 0f3F800000; fma.rn.ftz.f32 %f851, %f19, %f2118, %f833; sub.ftz.f32 %f2119, %f850, %f825; sub.ftz.f32 %f2120, %f851, %f827; mul.ftz.f32 %f2121, %f2120, %f2120; fma.rn.ftz.f32 %f2122, %f2119, %f2119, %f2121; sqrt.approx.ftz.f32 %f852, %f2122; setp.geu.ftz.f32 %p135, %f852, %f3255; @%p135 bra $L__BB199_269; mov.f32 %f3251, 0f3F800000; mov.f32 %f3252, %f833; mov.f32 %f3253, %f850; mov.f32 %f3254, %f851; mov.f32 %f3255, %f852; $L__BB199_269: setp.lt.s32 %p136, %r4371, 1; @%p136 bra $L__BB199_263; add.ftz.f32 %f858, %f3251, 0fBF800000; add.ftz.f32 %f2127, %f824, %f858; mov.b32 %r3099, %f2127; add.s32 %r318, %r3099, -559038716; add.s32 %r319, %r3099, -559038712; add.ftz.f32 %f859, %f3251, 0f3F800000; add.ftz.f32 %f2128, %f824, %f859; mov.b32 %r3100, %f2128; add.s32 %r320, %r3100, -559038716; add.s32 %r321, %r3100, -559038712; add.ftz.f32 %f860, %f3251, 0f00000000; mov.f32 %f3267, 0f00000000; add.ftz.f32 %f2129, %f824, %f860; mov.b32 %r3101, %f2129; add.s32 %r322, %r3101, -559038716; add.s32 %r323, %r3101, -559038712; mov.f32 %f3266, 0f41000000; mov.u32 %r4372, -2; mov.f32 %f3268, %f3267; $L__BB199_271: mov.u32 %r324, %r4372; add.s32 %r4372, %r324, 1; cvt.rn.f32.s32 %f2130, %r4372; add.ftz.f32 %f864, %f3252, %f2130; add.ftz.f32 %f2131, %f826, %f864; mov.b32 %r3102, %f2131; add.s32 %r326, %r3102, -559038716; xor.b32 %r3103, %r326, -559038716; shf.l.wrap.b32 %r3104, %r326, %r326, 14; sub.s32 %r327, %r3103, %r3104; shf.l.wrap.b32 %r328, %r327, %r327, 11; add.s32 %r3105, %r3102, -559038712; xor.b32 %r3106, %r3105, 506314504; shf.l.wrap.b32 %r3107, %r3105, %r3105, 14; sub.s32 %r329, %r3106, %r3107; shf.l.wrap.b32 %r330, %r329, %r329, 11; xor.b32 %r3108, %r318, %r327; sub.s32 %r3109, %r3108, %r328; xor.b32 %r3110, %r3109, %r326; shf.l.wrap.b32 %r3111, %r3109, %r3109, 25; sub.s32 %r3112, %r3110, %r3111; xor.b32 %r3113, %r3112, %r327; shf.l.wrap.b32 %r3114, %r3112, %r3112, 16; sub.s32 %r3115, %r3113, %r3114; xor.b32 %r3116, %r3115, %r3109; shf.l.wrap.b32 %r3117, %r3115, %r3115, 4; sub.s32 %r3118, %r3116, %r3117; xor.b32 %r3119, %r3118, %r3112; shf.l.wrap.b32 %r3120, %r3118, %r3118, 14; sub.s32 %r3121, %r3119, %r3120; xor.b32 %r3122, %r3121, %r3115; shf.l.wrap.b32 %r3123, %r3121, %r3121, 24; sub.s32 %r3124, %r3122, %r3123; cvt.rn.f32.u32 %f2132, %r3124; mul.ftz.f32 %f2133, %f2132, 0f2F800000; xor.b32 %r3125, %r319, %r329; sub.s32 %r3126, %r3125, %r330; xor.b32 %r3127, %r3126, %r3105; shf.l.wrap.b32 %r3128, %r3126, %r3126, 25; sub.s32 %r3129, %r3127, %r3128; xor.b32 %r3130, %r3129, %r329; shf.l.wrap.b32 %r3131, %r3129, %r3129, 16; sub.s32 %r3132, %r3130, %r3131; xor.b32 %r3133, %r3132, %r3126; shf.l.wrap.b32 %r3134, %r3132, %r3132, 4; sub.s32 %r3135, %r3133, %r3134; xor.b32 %r3136, %r3135, %r3129; shf.l.wrap.b32 %r3137, %r3135, %r3135, 14; sub.s32 %r3138, %r3136, %r3137; xor.b32 %r3139, %r3138, %r3132; shf.l.wrap.b32 %r3140, %r3138, %r3138, 24; sub.s32 %r3141, %r3139, %r3140; cvt.rn.f32.u32 %f2134, %r3141; mul.ftz.f32 %f2135, %f2134, 0f2F800000; fma.rn.ftz.f32 %f2136, %f19, %f2133, %f858; fma.rn.ftz.f32 %f2137, %f19, %f2135, %f864; sub.ftz.f32 %f2138, %f3253, %f2136; sub.ftz.f32 %f2139, %f3254, %f2137; mul.ftz.f32 %f2140, %f2139, %f2139; fma.rn.ftz.f32 %f2141, %f2138, %f2138, %f2140; sqrt.approx.ftz.f32 %f2142, %f2141; setp.lt.ftz.f32 %p137, %f2142, %f3266; selp.f32 %f3269, %f2136, %f3268, %p137; selp.f32 %f3270, %f2137, %f3267, %p137; selp.f32 %f3271, %f2142, %f3266, %p137; setp.eq.s32 %p138, %r324, -1; @%p138 bra $L__BB199_274; xor.b32 %r3142, %r322, %r327; sub.s32 %r3143, %r3142, %r328; xor.b32 %r3144, %r3143, %r326; shf.l.wrap.b32 %r3145, %r3143, %r3143, 25; sub.s32 %r3146, %r3144, %r3145; xor.b32 %r3147, %r3146, %r327; shf.l.wrap.b32 %r3148, %r3146, %r3146, 16; sub.s32 %r3149, %r3147, %r3148; xor.b32 %r3150, %r3149, %r3143; shf.l.wrap.b32 %r3151, %r3149, %r3149, 4; sub.s32 %r3152, %r3150, %r3151; xor.b32 %r3153, %r3152, %r3146; shf.l.wrap.b32 %r3154, %r3152, %r3152, 14; sub.s32 %r3155, %r3153, %r3154; xor.b32 %r3156, %r3155, %r3149; shf.l.wrap.b32 %r3157, %r3155, %r3155, 24; sub.s32 %r3158, %r3156, %r3157; cvt.rn.f32.u32 %f2143, %r3158; mul.ftz.f32 %f2144, %f2143, 0f2F800000; xor.b32 %r3159, %r323, %r329; sub.s32 %r3160, %r3159, %r330; add.s32 %r3161, %r326, 4; xor.b32 %r3162, %r3160, %r3161; shf.l.wrap.b32 %r3163, %r3160, %r3160, 25; sub.s32 %r3164, %r3162, %r3163; xor.b32 %r3165, %r3164, %r329; shf.l.wrap.b32 %r3166, %r3164, %r3164, 16; sub.s32 %r3167, %r3165, %r3166; xor.b32 %r3168, %r3167, %r3160; shf.l.wrap.b32 %r3169, %r3167, %r3167, 4; sub.s32 %r3170, %r3168, %r3169; xor.b32 %r3171, %r3170, %r3164; shf.l.wrap.b32 %r3172, %r3170, %r3170, 14; sub.s32 %r3173, %r3171, %r3172; xor.b32 %r3174, %r3173, %r3167; shf.l.wrap.b32 %r3175, %r3173, %r3173, 24; sub.s32 %r3176, %r3174, %r3175; cvt.rn.f32.u32 %f2145, %r3176; mul.ftz.f32 %f2146, %f2145, 0f2F800000; fma.rn.ftz.f32 %f868, %f19, %f2144, %f860; fma.rn.ftz.f32 %f869, %f19, %f2146, %f864; sub.ftz.f32 %f2147, %f3253, %f868; sub.ftz.f32 %f2148, %f3254, %f869; mul.ftz.f32 %f2149, %f2148, %f2148; fma.rn.ftz.f32 %f2150, %f2147, %f2147, %f2149; sqrt.approx.ftz.f32 %f870, %f2150; setp.geu.ftz.f32 %p139, %f870, %f3271; @%p139 bra $L__BB199_274; mov.f32 %f3269, %f868; mov.f32 %f3270, %f869; mov.f32 %f3271, %f870; $L__BB199_274: xor.b32 %r3177, %r320, %r327; sub.s32 %r3178, %r3177, %r328; xor.b32 %r3179, %r3178, %r326; shf.l.wrap.b32 %r3180, %r3178, %r3178, 25; sub.s32 %r3181, %r3179, %r3180; xor.b32 %r3182, %r3181, %r327; shf.l.wrap.b32 %r3183, %r3181, %r3181, 16; sub.s32 %r3184, %r3182, %r3183; xor.b32 %r3185, %r3184, %r3178; shf.l.wrap.b32 %r3186, %r3184, %r3184, 4; sub.s32 %r3187, %r3185, %r3186; xor.b32 %r3188, %r3187, %r3181; shf.l.wrap.b32 %r3189, %r3187, %r3187, 14; sub.s32 %r3190, %r3188, %r3189; xor.b32 %r3191, %r3190, %r3184; shf.l.wrap.b32 %r3192, %r3190, %r3190, 24; sub.s32 %r3193, %r3191, %r3192; cvt.rn.f32.u32 %f2151, %r3193; mul.ftz.f32 %f2152, %f2151, 0f2F800000; xor.b32 %r3194, %r321, %r329; sub.s32 %r3195, %r3194, %r330; add.s32 %r3196, %r326, 4; xor.b32 %r3197, %r3195, %r3196; shf.l.wrap.b32 %r3198, %r3195, %r3195, 25; sub.s32 %r3199, %r3197, %r3198; xor.b32 %r3200, %r3199, %r329; shf.l.wrap.b32 %r3201, %r3199, %r3199, 16; sub.s32 %r3202, %r3200, %r3201; xor.b32 %r3203, %r3202, %r3195; shf.l.wrap.b32 %r3204, %r3202, %r3202, 4; sub.s32 %r3205, %r3203, %r3204; xor.b32 %r3206, %r3205, %r3199; shf.l.wrap.b32 %r3207, %r3205, %r3205, 14; sub.s32 %r3208, %r3206, %r3207; xor.b32 %r3209, %r3208, %r3202; shf.l.wrap.b32 %r3210, %r3208, %r3208, 24; sub.s32 %r3211, %r3209, %r3210; cvt.rn.f32.u32 %f2153, %r3211; mul.ftz.f32 %f2154, %f2153, 0f2F800000; fma.rn.ftz.f32 %f2155, %f19, %f2152, %f859; fma.rn.ftz.f32 %f2156, %f19, %f2154, %f864; sub.ftz.f32 %f2157, %f3253, %f2155; sub.ftz.f32 %f2158, %f3254, %f2156; mul.ftz.f32 %f2159, %f2158, %f2158; fma.rn.ftz.f32 %f2160, %f2157, %f2157, %f2159; sqrt.approx.ftz.f32 %f2161, %f2160; setp.lt.ftz.f32 %p140, %f2161, %f3271; selp.f32 %f3268, %f2155, %f3269, %p140; selp.f32 %f3267, %f2156, %f3270, %p140; selp.f32 %f3266, %f2161, %f3271, %p140; setp.lt.s32 %p141, %r4372, 1; @%p141 bra $L__BB199_271; sub.ftz.f32 %f2168, %f3268, %f3253; sub.ftz.f32 %f2169, %f3267, %f3254; mul.ftz.f32 %f2170, %f2169, %f2169; fma.rn.ftz.f32 %f2171, %f2168, %f2168, %f2170; sqrt.approx.ftz.f32 %f2172, %f2171; div.approx.ftz.f32 %f3388, %f2172, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3364, %f3386; mov.f32 %f3365, %f3386; bra.uni $L__BB199_362; $L__BB199_26: ld.param.u32 %r4309, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.eq.s32 %p7, %r4309, 3; @%p7 bra $L__BB199_138; mov.f32 %f3386, 0f00000000; ld.param.u32 %r4310, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_1]; setp.ne.s32 %p8, %r4310, 4; mov.f32 %f3387, %f3386; mov.f32 %f3388, %f3386; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3392, %f3386; mov.f32 %f3393, %f3386; mov.f32 %f3394, %f3386; @%p8 bra $L__BB199_385; mov.f32 %f3394, 0f00000000; setp.gt.s32 %p11, %r389, 1; @%p11 bra $L__BB199_40; bra.uni $L__BB199_29; $L__BB199_40: setp.eq.s32 %p12, %r389, 2; @%p12 bra $L__BB199_85; setp.eq.s32 %p13, %r389, 3; @%p13 bra $L__BB199_65; setp.ne.s32 %p14, %r389, 4; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p14 bra $L__BB199_136; cvt.rmi.ftz.f32.f32 %f25, %f22; sub.ftz.f32 %f26, %f22, %f25; cvt.rmi.ftz.f32.f32 %f27, %f23; sub.ftz.f32 %f28, %f23, %f27; cvt.rmi.ftz.f32.f32 %f29, %f24; sub.ftz.f32 %f30, %f24, %f29; cvt.rmi.ftz.f32.f32 %f31, %f21; sub.ftz.f32 %f32, %f21, %f31; mov.f32 %f2812, 0f00000000; mov.u32 %r405, -1; mov.f32 %f2813, 0f41000000; mov.u32 %r4323, %r405; mov.f32 %f2811, %f2812; mov.f32 %f2810, %f2812; mov.f32 %f2809, %f2812; mov.f32 %f2808, %f2812; mov.f32 %f2807, %f2812; mov.f32 %f2806, %f2812; mov.f32 %f2805, %f2812; $L__BB199_44: cvt.rn.f32.s32 %f42, %r4323; add.ftz.f32 %f1203, %f31, %f42; mov.b32 %r15, %f1203; add.s32 %r16, %r15, -559038708; shf.l.wrap.b32 %r17, %r16, %r16, 4; mov.u32 %r4324, %r405; $L__BB199_45: cvt.rn.f32.s32 %f52, %r4324; add.ftz.f32 %f1204, %f29, %f52; mov.b32 %r19, %f1204; add.s32 %r20, %r19, -559038708; shf.l.wrap.b32 %r21, %r20, %r20, 4; add.s32 %r22, %r20, %r16; mov.u32 %r4325, %r405; $L__BB199_46: cvt.rn.f32.s32 %f62, %r4325; add.ftz.f32 %f1205, %f27, %f62; mov.b32 %r24, %f1205; add.s32 %r25, %r24, -559038708; add.s32 %r26, %r25, %r20; sub.s32 %r409, %r15, %r24; shf.l.wrap.b32 %r410, %r25, %r25, 4; xor.b32 %r27, %r410, %r409; shf.l.wrap.b32 %r28, %r27, %r27, 6; sub.s32 %r411, %r24, %r15; xor.b32 %r412, %r411, %r17; sub.s32 %r413, %r20, %r412; shf.l.wrap.b32 %r414, %r412, %r412, 6; xor.b32 %r415, %r414, %r413; add.s32 %r416, %r412, %r22; sub.s32 %r417, %r22, %r415; shf.l.wrap.b32 %r418, %r415, %r415, 8; xor.b32 %r419, %r418, %r417; add.s32 %r420, %r415, %r416; sub.s32 %r421, %r416, %r419; shf.l.wrap.b32 %r422, %r419, %r419, 16; xor.b32 %r423, %r422, %r421; add.s32 %r424, %r419, %r420; sub.s32 %r425, %r420, %r423; shf.l.wrap.b32 %r426, %r423, %r423, 19; xor.b32 %r427, %r426, %r425; add.s32 %r29, %r423, %r424; sub.s32 %r428, %r424, %r427; shf.l.wrap.b32 %r429, %r427, %r427, 4; add.s32 %r30, %r427, %r29; xor.b32 %r430, %r30, %r428; xor.b32 %r431, %r430, %r429; shf.l.wrap.b32 %r432, %r30, %r30, 14; sub.s32 %r31, %r431, %r432; shf.l.wrap.b32 %r32, %r31, %r31, 11; mov.u32 %r4326, %r405; $L__BB199_47: cvt.rn.f32.s32 %f72, %r4326; add.ftz.f32 %f1206, %f25, %f72; mov.b32 %r433, %f1206; sub.s32 %r434, %r433, %r19; xor.b32 %r435, %r434, %r21; sub.s32 %r436, %r25, %r435; shf.l.wrap.b32 %r437, %r435, %r435, 6; xor.b32 %r438, %r437, %r436; add.s32 %r439, %r435, %r26; sub.s32 %r440, %r26, %r438; shf.l.wrap.b32 %r441, %r438, %r438, 8; xor.b32 %r442, %r441, %r440; add.s32 %r443, %r438, %r439; sub.s32 %r444, %r439, %r442; shf.l.wrap.b32 %r445, %r442, %r442, 16; xor.b32 %r446, %r445, %r444; add.s32 %r447, %r442, %r443; sub.s32 %r448, %r443, %r446; shf.l.wrap.b32 %r449, %r446, %r446, 19; xor.b32 %r450, %r449, %r448; add.s32 %r451, %r446, %r447; sub.s32 %r452, %r447, %r450; shf.l.wrap.b32 %r453, %r450, %r450, 4; add.s32 %r454, %r450, %r451; add.s32 %r455, %r451, %r15; xor.b32 %r456, %r454, %r452; xor.b32 %r457, %r456, %r453; shf.l.wrap.b32 %r458, %r454, %r454, 14; sub.s32 %r459, %r457, %r458; xor.b32 %r460, %r459, %r455; shf.l.wrap.b32 %r461, %r459, %r459, 11; sub.s32 %r462, %r460, %r461; xor.b32 %r463, %r462, %r454; shf.l.wrap.b32 %r464, %r462, %r462, 25; sub.s32 %r465, %r463, %r464; xor.b32 %r466, %r465, %r459; shf.l.wrap.b32 %r467, %r465, %r465, 16; sub.s32 %r468, %r466, %r467; xor.b32 %r469, %r468, %r462; shf.l.wrap.b32 %r470, %r468, %r468, 4; sub.s32 %r471, %r469, %r470; xor.b32 %r472, %r471, %r465; shf.l.wrap.b32 %r473, %r471, %r471, 14; sub.s32 %r474, %r472, %r473; xor.b32 %r475, %r474, %r468; shf.l.wrap.b32 %r476, %r474, %r474, 24; sub.s32 %r477, %r475, %r476; cvt.rn.f32.u32 %f1207, %r477; mul.ftz.f32 %f1208, %f1207, 0f2F800000; add.s32 %r478, %r433, -559038708; add.s32 %r479, %r478, %r25; sub.s32 %r480, %r478, %r27; xor.b32 %r481, %r480, %r28; add.s32 %r482, %r479, %r27; sub.s32 %r483, %r479, %r481; shf.l.wrap.b32 %r484, %r481, %r481, 8; xor.b32 %r485, %r484, %r483; add.s32 %r486, %r481, %r482; sub.s32 %r487, %r482, %r485; shf.l.wrap.b32 %r488, %r485, %r485, 16; xor.b32 %r489, %r488, %r487; add.s32 %r490, %r485, %r486; sub.s32 %r491, %r486, %r489; shf.l.wrap.b32 %r492, %r489, %r489, 19; xor.b32 %r493, %r492, %r491; add.s32 %r494, %r489, %r490; sub.s32 %r495, %r490, %r493; shf.l.wrap.b32 %r496, %r493, %r493, 4; add.s32 %r497, %r493, %r494; add.s32 %r498, %r494, %r19; xor.b32 %r499, %r497, %r495; xor.b32 %r500, %r499, %r496; shf.l.wrap.b32 %r501, %r497, %r497, 14; sub.s32 %r502, %r500, %r501; xor.b32 %r503, %r502, %r498; shf.l.wrap.b32 %r504, %r502, %r502, 11; sub.s32 %r505, %r503, %r504; xor.b32 %r506, %r505, %r497; shf.l.wrap.b32 %r507, %r505, %r505, 25; sub.s32 %r508, %r506, %r507; xor.b32 %r509, %r508, %r502; shf.l.wrap.b32 %r510, %r508, %r508, 16; sub.s32 %r511, %r509, %r510; xor.b32 %r512, %r511, %r505; shf.l.wrap.b32 %r513, %r511, %r511, 4; sub.s32 %r514, %r512, %r513; xor.b32 %r515, %r514, %r508; shf.l.wrap.b32 %r516, %r514, %r514, 14; sub.s32 %r517, %r515, %r516; xor.b32 %r518, %r517, %r511; shf.l.wrap.b32 %r519, %r517, %r517, 24; sub.s32 %r520, %r518, %r519; cvt.rn.f32.u32 %f1209, %r520; mul.ftz.f32 %f1210, %f1209, 0f2F800000; sub.s32 %r521, %r19, %r433; shf.l.wrap.b32 %r522, %r478, %r478, 4; xor.b32 %r523, %r522, %r521; add.s32 %r524, %r478, %r16; sub.s32 %r525, %r16, %r523; shf.l.wrap.b32 %r526, %r523, %r523, 6; xor.b32 %r527, %r526, %r525; add.s32 %r528, %r523, %r524; sub.s32 %r529, %r524, %r527; shf.l.wrap.b32 %r530, %r527, %r527, 8; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r527, %r528; sub.s32 %r533, %r528, %r531; shf.l.wrap.b32 %r534, %r531, %r531, 16; xor.b32 %r535, %r534, %r533; add.s32 %r536, %r531, %r532; sub.s32 %r537, %r532, %r535; shf.l.wrap.b32 %r538, %r535, %r535, 19; xor.b32 %r539, %r538, %r537; add.s32 %r540, %r535, %r536; sub.s32 %r541, %r536, %r539; shf.l.wrap.b32 %r542, %r539, %r539, 4; add.s32 %r543, %r539, %r540; add.s32 %r544, %r540, %r24; xor.b32 %r545, %r543, %r541; xor.b32 %r546, %r545, %r542; shf.l.wrap.b32 %r547, %r543, %r543, 14; sub.s32 %r548, %r546, %r547; xor.b32 %r549, %r548, %r544; shf.l.wrap.b32 %r550, %r548, %r548, 11; sub.s32 %r551, %r549, %r550; xor.b32 %r552, %r551, %r543; shf.l.wrap.b32 %r553, %r551, %r551, 25; sub.s32 %r554, %r552, %r553; xor.b32 %r555, %r554, %r548; shf.l.wrap.b32 %r556, %r554, %r554, 16; sub.s32 %r557, %r555, %r556; xor.b32 %r558, %r557, %r551; shf.l.wrap.b32 %r559, %r557, %r557, 4; sub.s32 %r560, %r558, %r559; xor.b32 %r561, %r560, %r554; shf.l.wrap.b32 %r562, %r560, %r560, 14; sub.s32 %r563, %r561, %r562; xor.b32 %r564, %r563, %r557; shf.l.wrap.b32 %r565, %r563, %r563, 24; sub.s32 %r566, %r564, %r565; cvt.rn.f32.u32 %f1211, %r566; mul.ftz.f32 %f1212, %f1211, 0f2F800000; add.s32 %r567, %r29, %r433; xor.b32 %r568, %r567, %r31; sub.s32 %r569, %r568, %r32; xor.b32 %r570, %r569, %r30; shf.l.wrap.b32 %r571, %r569, %r569, 25; sub.s32 %r572, %r570, %r571; xor.b32 %r573, %r572, %r31; shf.l.wrap.b32 %r574, %r572, %r572, 16; sub.s32 %r575, %r573, %r574; xor.b32 %r576, %r575, %r569; shf.l.wrap.b32 %r577, %r575, %r575, 4; sub.s32 %r578, %r576, %r577; xor.b32 %r579, %r578, %r572; shf.l.wrap.b32 %r580, %r578, %r578, 14; sub.s32 %r581, %r579, %r580; xor.b32 %r582, %r581, %r575; shf.l.wrap.b32 %r583, %r581, %r581, 24; sub.s32 %r584, %r582, %r583; cvt.rn.f32.u32 %f1213, %r584; mul.ftz.f32 %f1214, %f1213, 0f2F800000; fma.rn.ftz.f32 %f73, %f19, %f1208, %f72; fma.rn.ftz.f32 %f74, %f19, %f1210, %f62; fma.rn.ftz.f32 %f75, %f19, %f1212, %f52; fma.rn.ftz.f32 %f76, %f19, %f1214, %f42; sub.ftz.f32 %f1215, %f73, %f26; sub.ftz.f32 %f1216, %f74, %f28; sub.ftz.f32 %f1217, %f75, %f30; sub.ftz.f32 %f1218, %f76, %f32; mul.ftz.f32 %f1219, %f1215, %f1215; fma.rn.ftz.f32 %f1220, %f1216, %f1216, %f1219; mul.ftz.f32 %f1221, %f1217, %f1217; fma.rn.ftz.f32 %f1222, %f1218, %f1218, %f1221; add.ftz.f32 %f1223, %f1220, %f1222; sqrt.approx.ftz.f32 %f77, %f1223; setp.geu.ftz.f32 %p17, %f77, %f2813; @%p17 bra $L__BB199_49; mov.f32 %f2805, %f72; mov.f32 %f2806, %f62; mov.f32 %f2807, %f52; mov.f32 %f2808, %f42; mov.f32 %f2809, %f73; mov.f32 %f2810, %f74; mov.f32 %f2811, %f75; mov.f32 %f2812, %f76; mov.f32 %f2813, %f77; $L__BB199_49: add.s32 %r34, %r4326, 1; setp.lt.s32 %p18, %r4326, 1; mov.u32 %r4326, %r34; @%p18 bra $L__BB199_47; add.s32 %r35, %r4325, 1; setp.lt.s32 %p19, %r4325, 1; mov.u32 %r4325, %r35; @%p19 bra $L__BB199_46; add.s32 %r36, %r4324, 1; setp.lt.s32 %p20, %r4324, 1; mov.u32 %r4324, %r36; @%p20 bra $L__BB199_45; add.s32 %r37, %r4323, 1; setp.lt.s32 %p21, %r4323, 1; mov.u32 %r4323, %r37; @%p21 bra $L__BB199_44; mov.f32 %f2837, 0f00000000; mov.u32 %r585, -1; mov.f32 %f2838, 0f41000000; mov.u32 %r4327, %r585; mov.f32 %f2836, %f2837; mov.f32 %f2835, %f2837; mov.f32 %f2834, %f2837; $L__BB199_54: cvt.rn.f32.s32 %f1229, %r4327; add.ftz.f32 %f92, %f2808, %f1229; add.ftz.f32 %f1230, %f31, %f92; mov.b32 %r39, %f1230; add.s32 %r40, %r39, -559038708; shf.l.wrap.b32 %r41, %r40, %r40, 4; mov.u32 %r4328, %r585; $L__BB199_55: or.b32 %r43, %r4328, %r4327; cvt.rn.f32.s32 %f1231, %r4328; add.ftz.f32 %f98, %f2807, %f1231; add.ftz.f32 %f1232, %f29, %f98; mov.b32 %r44, %f1232; add.s32 %r45, %r44, -559038708; shf.l.wrap.b32 %r46, %r45, %r45, 4; add.s32 %r47, %r45, %r40; mov.u32 %r4329, %r585; $L__BB199_56: or.b32 %r49, %r43, %r4329; cvt.rn.f32.s32 %f1233, %r4329; add.ftz.f32 %f104, %f2806, %f1233; add.ftz.f32 %f1234, %f27, %f104; mov.b32 %r50, %f1234; add.s32 %r51, %r50, -559038708; add.s32 %r52, %r51, %r45; sub.s32 %r589, %r39, %r50; shf.l.wrap.b32 %r590, %r51, %r51, 4; xor.b32 %r53, %r590, %r589; shf.l.wrap.b32 %r54, %r53, %r53, 6; sub.s32 %r591, %r50, %r39; xor.b32 %r592, %r591, %r41; sub.s32 %r593, %r45, %r592; shf.l.wrap.b32 %r594, %r592, %r592, 6; xor.b32 %r595, %r594, %r593; add.s32 %r596, %r592, %r47; sub.s32 %r597, %r47, %r595; shf.l.wrap.b32 %r598, %r595, %r595, 8; xor.b32 %r599, %r598, %r597; add.s32 %r600, %r595, %r596; sub.s32 %r601, %r596, %r599; shf.l.wrap.b32 %r602, %r599, %r599, 16; xor.b32 %r603, %r602, %r601; add.s32 %r604, %r599, %r600; sub.s32 %r605, %r600, %r603; shf.l.wrap.b32 %r606, %r603, %r603, 19; xor.b32 %r607, %r606, %r605; add.s32 %r55, %r603, %r604; sub.s32 %r608, %r604, %r607; shf.l.wrap.b32 %r609, %r607, %r607, 4; add.s32 %r56, %r607, %r55; xor.b32 %r610, %r56, %r608; xor.b32 %r611, %r610, %r609; shf.l.wrap.b32 %r612, %r56, %r56, 14; sub.s32 %r57, %r611, %r612; shf.l.wrap.b32 %r58, %r57, %r57, 11; mov.u32 %r4330, %r585; $L__BB199_57: mov.u32 %r59, %r4330; or.b32 %r613, %r49, %r59; setp.eq.s32 %p22, %r613, 0; @%p22 bra $L__BB199_60; cvt.rn.f32.s32 %f1235, %r59; add.ftz.f32 %f1236, %f2805, %f1235; add.ftz.f32 %f1237, %f25, %f1236; mov.b32 %r614, %f1237; sub.s32 %r615, %r614, %r44; xor.b32 %r616, %r615, %r46; sub.s32 %r617, %r51, %r616; shf.l.wrap.b32 %r618, %r616, %r616, 6; xor.b32 %r619, %r618, %r617; add.s32 %r620, %r616, %r52; sub.s32 %r621, %r52, %r619; shf.l.wrap.b32 %r622, %r619, %r619, 8; xor.b32 %r623, %r622, %r621; add.s32 %r624, %r619, %r620; sub.s32 %r625, %r620, %r623; shf.l.wrap.b32 %r626, %r623, %r623, 16; xor.b32 %r627, %r626, %r625; add.s32 %r628, %r623, %r624; sub.s32 %r629, %r624, %r627; shf.l.wrap.b32 %r630, %r627, %r627, 19; xor.b32 %r631, %r630, %r629; add.s32 %r632, %r627, %r628; sub.s32 %r633, %r628, %r631; shf.l.wrap.b32 %r634, %r631, %r631, 4; add.s32 %r635, %r631, %r632; add.s32 %r636, %r632, %r39; xor.b32 %r637, %r635, %r633; xor.b32 %r638, %r637, %r634; shf.l.wrap.b32 %r639, %r635, %r635, 14; sub.s32 %r640, %r638, %r639; xor.b32 %r641, %r640, %r636; shf.l.wrap.b32 %r642, %r640, %r640, 11; sub.s32 %r643, %r641, %r642; xor.b32 %r644, %r643, %r635; shf.l.wrap.b32 %r645, %r643, %r643, 25; sub.s32 %r646, %r644, %r645; xor.b32 %r647, %r646, %r640; shf.l.wrap.b32 %r648, %r646, %r646, 16; sub.s32 %r649, %r647, %r648; xor.b32 %r650, %r649, %r643; shf.l.wrap.b32 %r651, %r649, %r649, 4; sub.s32 %r652, %r650, %r651; xor.b32 %r653, %r652, %r646; shf.l.wrap.b32 %r654, %r652, %r652, 14; sub.s32 %r655, %r653, %r654; xor.b32 %r656, %r655, %r649; shf.l.wrap.b32 %r657, %r655, %r655, 24; sub.s32 %r658, %r656, %r657; cvt.rn.f32.u32 %f1238, %r658; mul.ftz.f32 %f1239, %f1238, 0f2F800000; add.s32 %r659, %r614, -559038708; add.s32 %r660, %r659, %r51; sub.s32 %r661, %r659, %r53; xor.b32 %r662, %r661, %r54; add.s32 %r663, %r660, %r53; sub.s32 %r664, %r660, %r662; shf.l.wrap.b32 %r665, %r662, %r662, 8; xor.b32 %r666, %r665, %r664; add.s32 %r667, %r662, %r663; sub.s32 %r668, %r663, %r666; shf.l.wrap.b32 %r669, %r666, %r666, 16; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r666, %r667; sub.s32 %r672, %r667, %r670; shf.l.wrap.b32 %r673, %r670, %r670, 19; xor.b32 %r674, %r673, %r672; add.s32 %r675, %r670, %r671; sub.s32 %r676, %r671, %r674; shf.l.wrap.b32 %r677, %r674, %r674, 4; add.s32 %r678, %r674, %r675; add.s32 %r679, %r675, %r44; xor.b32 %r680, %r678, %r676; xor.b32 %r681, %r680, %r677; shf.l.wrap.b32 %r682, %r678, %r678, 14; sub.s32 %r683, %r681, %r682; xor.b32 %r684, %r683, %r679; shf.l.wrap.b32 %r685, %r683, %r683, 11; sub.s32 %r686, %r684, %r685; xor.b32 %r687, %r686, %r678; shf.l.wrap.b32 %r688, %r686, %r686, 25; sub.s32 %r689, %r687, %r688; xor.b32 %r690, %r689, %r683; shf.l.wrap.b32 %r691, %r689, %r689, 16; sub.s32 %r692, %r690, %r691; xor.b32 %r693, %r692, %r686; shf.l.wrap.b32 %r694, %r692, %r692, 4; sub.s32 %r695, %r693, %r694; xor.b32 %r696, %r695, %r689; shf.l.wrap.b32 %r697, %r695, %r695, 14; sub.s32 %r698, %r696, %r697; xor.b32 %r699, %r698, %r692; shf.l.wrap.b32 %r700, %r698, %r698, 24; sub.s32 %r701, %r699, %r700; cvt.rn.f32.u32 %f1240, %r701; mul.ftz.f32 %f1241, %f1240, 0f2F800000; sub.s32 %r702, %r44, %r614; shf.l.wrap.b32 %r703, %r659, %r659, 4; xor.b32 %r704, %r703, %r702; add.s32 %r705, %r659, %r40; sub.s32 %r706, %r40, %r704; shf.l.wrap.b32 %r707, %r704, %r704, 6; xor.b32 %r708, %r707, %r706; add.s32 %r709, %r704, %r705; sub.s32 %r710, %r705, %r708; shf.l.wrap.b32 %r711, %r708, %r708, 8; xor.b32 %r712, %r711, %r710; add.s32 %r713, %r708, %r709; sub.s32 %r714, %r709, %r712; shf.l.wrap.b32 %r715, %r712, %r712, 16; xor.b32 %r716, %r715, %r714; add.s32 %r717, %r712, %r713; sub.s32 %r718, %r713, %r716; shf.l.wrap.b32 %r719, %r716, %r716, 19; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r716, %r717; sub.s32 %r722, %r717, %r720; shf.l.wrap.b32 %r723, %r720, %r720, 4; add.s32 %r724, %r720, %r721; add.s32 %r725, %r721, %r50; xor.b32 %r726, %r724, %r722; xor.b32 %r727, %r726, %r723; shf.l.wrap.b32 %r728, %r724, %r724, 14; sub.s32 %r729, %r727, %r728; xor.b32 %r730, %r729, %r725; shf.l.wrap.b32 %r731, %r729, %r729, 11; sub.s32 %r732, %r730, %r731; xor.b32 %r733, %r732, %r724; shf.l.wrap.b32 %r734, %r732, %r732, 25; sub.s32 %r735, %r733, %r734; xor.b32 %r736, %r735, %r729; shf.l.wrap.b32 %r737, %r735, %r735, 16; sub.s32 %r738, %r736, %r737; xor.b32 %r739, %r738, %r732; shf.l.wrap.b32 %r740, %r738, %r738, 4; sub.s32 %r741, %r739, %r740; xor.b32 %r742, %r741, %r735; shf.l.wrap.b32 %r743, %r741, %r741, 14; sub.s32 %r744, %r742, %r743; xor.b32 %r745, %r744, %r738; shf.l.wrap.b32 %r746, %r744, %r744, 24; sub.s32 %r747, %r745, %r746; cvt.rn.f32.u32 %f1242, %r747; mul.ftz.f32 %f1243, %f1242, 0f2F800000; add.s32 %r748, %r55, %r614; xor.b32 %r749, %r748, %r57; sub.s32 %r750, %r749, %r58; xor.b32 %r751, %r750, %r56; shf.l.wrap.b32 %r752, %r750, %r750, 25; sub.s32 %r753, %r751, %r752; xor.b32 %r754, %r753, %r57; shf.l.wrap.b32 %r755, %r753, %r753, 16; sub.s32 %r756, %r754, %r755; xor.b32 %r757, %r756, %r750; shf.l.wrap.b32 %r758, %r756, %r756, 4; sub.s32 %r759, %r757, %r758; xor.b32 %r760, %r759, %r753; shf.l.wrap.b32 %r761, %r759, %r759, 14; sub.s32 %r762, %r760, %r761; xor.b32 %r763, %r762, %r756; shf.l.wrap.b32 %r764, %r762, %r762, 24; sub.s32 %r765, %r763, %r764; cvt.rn.f32.u32 %f1244, %r765; mul.ftz.f32 %f1245, %f1244, 0f2F800000; fma.rn.ftz.f32 %f110, %f19, %f1239, %f1236; fma.rn.ftz.f32 %f111, %f19, %f1241, %f104; fma.rn.ftz.f32 %f112, %f19, %f1243, %f98; fma.rn.ftz.f32 %f113, %f19, %f1245, %f92; sub.ftz.f32 %f1246, %f2809, %f110; sub.ftz.f32 %f1247, %f2810, %f111; sub.ftz.f32 %f1248, %f2811, %f112; sub.ftz.f32 %f1249, %f2812, %f113; mul.ftz.f32 %f1250, %f1246, %f1246; fma.rn.ftz.f32 %f1251, %f1247, %f1247, %f1250; mul.ftz.f32 %f1252, %f1248, %f1248; fma.rn.ftz.f32 %f1253, %f1249, %f1249, %f1252; add.ftz.f32 %f1254, %f1251, %f1253; sqrt.approx.ftz.f32 %f114, %f1254; setp.geu.ftz.f32 %p23, %f114, %f2838; @%p23 bra $L__BB199_60; mov.f32 %f2834, %f110; mov.f32 %f2835, %f111; mov.f32 %f2836, %f112; mov.f32 %f2837, %f113; mov.f32 %f2838, %f114; $L__BB199_60: add.s32 %r4330, %r59, 1; setp.lt.s32 %p24, %r59, 1; @%p24 bra $L__BB199_57; add.s32 %r61, %r4329, 1; setp.lt.s32 %p25, %r4329, 1; mov.u32 %r4329, %r61; @%p25 bra $L__BB199_56; add.s32 %r62, %r4328, 1; setp.lt.s32 %p26, %r4328, 1; mov.u32 %r4328, %r62; @%p26 bra $L__BB199_55; add.s32 %r63, %r4327, 1; setp.lt.s32 %p27, %r4327, 1; mov.u32 %r4327, %r63; @%p27 bra $L__BB199_54; sub.ftz.f32 %f1260, %f2834, %f2809; sub.ftz.f32 %f1261, %f2835, %f2810; mul.ftz.f32 %f1262, %f1261, %f1261; fma.rn.ftz.f32 %f1263, %f1260, %f1260, %f1262; sub.ftz.f32 %f1264, %f2836, %f2811; sub.ftz.f32 %f1265, %f2837, %f2812; mul.ftz.f32 %f1266, %f1265, %f1265; fma.rn.ftz.f32 %f1267, %f1264, %f1264, %f1266; add.ftz.f32 %f1268, %f1263, %f1267; sqrt.approx.ftz.f32 %f1269, %f1268; div.approx.ftz.f32 %f3388, %f1269, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; bra.uni $L__BB199_136; $L__BB199_365: mov.f32 %f3392, 0f00000000; setp.gt.s32 %p194, %r389, 1; @%p194 bra $L__BB199_371; bra.uni $L__BB199_366; $L__BB199_371: setp.eq.s32 %p195, %r389, 2; @%p195 bra $L__BB199_376; setp.eq.s32 %p196, %r389, 3; @%p196 bra $L__BB199_375; setp.ne.s32 %p197, %r389, 4; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p197 bra $L__BB199_383; cvt.rmi.ftz.f32.f32 %f2581, %f21; sub.ftz.f32 %f2582, %f21, %f2581; add.ftz.f32 %f2583, %f2581, 0fBF800000; mov.b32 %r3865, %f2583; add.s32 %r3866, %r3865, -559038720; xor.b32 %r3867, %r3866, -1874868139; add.s32 %r3868, %r3867, 29184895; xor.b32 %r3869, %r3868, -559038720; shf.l.wrap.b32 %r3870, %r3868, %r3868, 25; sub.s32 %r3871, %r3869, %r3870; xor.b32 %r3872, %r3871, -1874868139; shf.l.wrap.b32 %r3873, %r3871, %r3871, 16; sub.s32 %r3874, %r3872, %r3873; xor.b32 %r3875, %r3874, %r3868; shf.l.wrap.b32 %r3876, %r3874, %r3874, 4; sub.s32 %r3877, %r3875, %r3876; xor.b32 %r3878, %r3877, %r3871; shf.l.wrap.b32 %r3879, %r3877, %r3877, 14; sub.s32 %r3880, %r3878, %r3879; xor.b32 %r3881, %r3880, %r3874; shf.l.wrap.b32 %r3882, %r3880, %r3880, 24; sub.s32 %r3883, %r3881, %r3882; cvt.rn.f32.u32 %f2584, %r3883; mul.ftz.f32 %f2585, %f2584, 0f2F800000; fma.rn.ftz.f32 %f2586, %f19, %f2585, 0fBF800000; sub.ftz.f32 %f2587, %f2586, %f2582; abs.ftz.f32 %f2588, %f2587; setp.lt.ftz.f32 %p200, %f2588, 0f41000000; selp.f32 %f2589, %f2586, 0f00000000, %p200; selp.f32 %f2590, %f2588, 0f41000000, %p200; add.ftz.f32 %f2591, %f2581, 0f00000000; mov.b32 %r3884, %f2591; add.s32 %r3885, %r3884, -559038720; xor.b32 %r3886, %r3885, -1874868139; add.s32 %r3887, %r3886, 29184895; xor.b32 %r3888, %r3887, -559038720; shf.l.wrap.b32 %r3889, %r3887, %r3887, 25; sub.s32 %r3890, %r3888, %r3889; xor.b32 %r3891, %r3890, -1874868139; shf.l.wrap.b32 %r3892, %r3890, %r3890, 16; sub.s32 %r3893, %r3891, %r3892; xor.b32 %r3894, %r3893, %r3887; shf.l.wrap.b32 %r3895, %r3893, %r3893, 4; sub.s32 %r3896, %r3894, %r3895; xor.b32 %r3897, %r3896, %r3890; shf.l.wrap.b32 %r3898, %r3896, %r3896, 14; sub.s32 %r3899, %r3897, %r3898; xor.b32 %r3900, %r3899, %r3893; shf.l.wrap.b32 %r3901, %r3899, %r3899, 24; sub.s32 %r3902, %r3900, %r3901; cvt.rn.f32.u32 %f2592, %r3902; mul.ftz.f32 %f2593, %f2592, 0f2F800000; fma.rn.ftz.f32 %f2594, %f19, %f2593, 0f00000000; sub.ftz.f32 %f2595, %f2594, %f2582; abs.ftz.f32 %f2596, %f2595; setp.lt.ftz.f32 %p201, %f2596, %f2590; selp.f32 %f2597, %f2594, %f2589, %p201; setp.geu.ftz.f32 %p202, %f2596, %f2590; and.pred %p203, %p200, %p202; selp.f32 %f2598, 0fBF800000, 0f00000000, %p203; selp.f32 %f2599, %f2596, %f2590, %p201; add.ftz.f32 %f2600, %f2581, 0f3F800000; mov.b32 %r3903, %f2600; add.s32 %r3904, %r3903, -559038720; xor.b32 %r3905, %r3904, -1874868139; add.s32 %r3906, %r3905, 29184895; xor.b32 %r3907, %r3906, -559038720; shf.l.wrap.b32 %r3908, %r3906, %r3906, 25; sub.s32 %r3909, %r3907, %r3908; xor.b32 %r3910, %r3909, -1874868139; shf.l.wrap.b32 %r3911, %r3909, %r3909, 16; sub.s32 %r3912, %r3910, %r3911; xor.b32 %r3913, %r3912, %r3906; shf.l.wrap.b32 %r3914, %r3912, %r3912, 4; sub.s32 %r3915, %r3913, %r3914; xor.b32 %r3916, %r3915, %r3909; shf.l.wrap.b32 %r3917, %r3915, %r3915, 14; sub.s32 %r3918, %r3916, %r3917; xor.b32 %r3919, %r3918, %r3912; shf.l.wrap.b32 %r3920, %r3918, %r3918, 24; sub.s32 %r3921, %r3919, %r3920; cvt.rn.f32.u32 %f2601, %r3921; mul.ftz.f32 %f2602, %f2601, 0f2F800000; fma.rn.ftz.f32 %f2603, %f19, %f2602, 0f3F800000; sub.ftz.f32 %f2604, %f2603, %f2582; abs.ftz.f32 %f2605, %f2604; setp.lt.ftz.f32 %p204, %f2605, %f2599; selp.f32 %f2606, %f2603, %f2597, %p204; selp.f32 %f2607, 0f3F800000, %f2598, %p204; add.ftz.f32 %f2608, %f2607, 0fBF800000; add.ftz.f32 %f2609, %f2581, %f2608; mov.b32 %r3922, %f2609; add.s32 %r3923, %r3922, -559038720; xor.b32 %r3924, %r3923, -1874868139; add.s32 %r3925, %r3924, 29184895; xor.b32 %r3926, %r3925, -559038720; shf.l.wrap.b32 %r3927, %r3925, %r3925, 25; sub.s32 %r3928, %r3926, %r3927; xor.b32 %r3929, %r3928, -1874868139; shf.l.wrap.b32 %r3930, %r3928, %r3928, 16; sub.s32 %r3931, %r3929, %r3930; xor.b32 %r3932, %r3931, %r3925; shf.l.wrap.b32 %r3933, %r3931, %r3931, 4; sub.s32 %r3934, %r3932, %r3933; xor.b32 %r3935, %r3934, %r3928; shf.l.wrap.b32 %r3936, %r3934, %r3934, 14; sub.s32 %r3937, %r3935, %r3936; xor.b32 %r3938, %r3937, %r3931; shf.l.wrap.b32 %r3939, %r3937, %r3937, 24; sub.s32 %r3940, %r3938, %r3939; cvt.rn.f32.u32 %f2610, %r3940; mul.ftz.f32 %f2611, %f2610, 0f2F800000; fma.rn.ftz.f32 %f2612, %f19, %f2611, %f2608; sub.ftz.f32 %f2613, %f2606, %f2612; abs.ftz.f32 %f2614, %f2613; setp.lt.ftz.f32 %p205, %f2614, 0f41000000; min.ftz.f32 %f2615, %f2614, 0f41000000; selp.f32 %f2616, %f2612, 0f00000000, %p205; add.ftz.f32 %f2617, %f2607, 0f3F800000; add.ftz.f32 %f2618, %f2581, %f2617; mov.b32 %r3941, %f2618; add.s32 %r3942, %r3941, -559038720; xor.b32 %r3943, %r3942, -1874868139; add.s32 %r3944, %r3943, 29184895; xor.b32 %r3945, %r3944, -559038720; shf.l.wrap.b32 %r3946, %r3944, %r3944, 25; sub.s32 %r3947, %r3945, %r3946; xor.b32 %r3948, %r3947, -1874868139; shf.l.wrap.b32 %r3949, %r3947, %r3947, 16; sub.s32 %r3950, %r3948, %r3949; xor.b32 %r3951, %r3950, %r3944; shf.l.wrap.b32 %r3952, %r3950, %r3950, 4; sub.s32 %r3953, %r3951, %r3952; xor.b32 %r3954, %r3953, %r3947; shf.l.wrap.b32 %r3955, %r3953, %r3953, 14; sub.s32 %r3956, %r3954, %r3955; xor.b32 %r3957, %r3956, %r3950; shf.l.wrap.b32 %r3958, %r3956, %r3956, 24; sub.s32 %r3959, %r3957, %r3958; cvt.rn.f32.u32 %f2619, %r3959; mul.ftz.f32 %f2620, %f2619, 0f2F800000; fma.rn.ftz.f32 %f2621, %f19, %f2620, %f2617; sub.ftz.f32 %f2622, %f2606, %f2621; abs.ftz.f32 %f2623, %f2622; setp.lt.ftz.f32 %p206, %f2623, %f2615; selp.f32 %f2624, %f2621, %f2616, %p206; sub.ftz.f32 %f2625, %f2624, %f2606; abs.ftz.f32 %f2626, %f2625; div.approx.ftz.f32 %f3388, %f2626, %f1184; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB199_383; $L__BB199_138: mov.f32 %f3387, 0f00000000; setp.gt.s32 %p68, %r389, 1; @%p68 bra $L__BB199_149; bra.uni $L__BB199_139; $L__BB199_149: setp.eq.s32 %p69, %r389, 2; @%p69 bra $L__BB199_186; setp.eq.s32 %p70, %r389, 3; @%p70 bra $L__BB199_170; setp.ne.s32 %p71, %r389, 4; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p71 bra $L__BB199_250; cvt.rmi.ftz.f32.f32 %f480, %f22; sub.ftz.f32 %f481, %f22, %f480; cvt.rmi.ftz.f32.f32 %f482, %f23; sub.ftz.f32 %f483, %f23, %f482; cvt.rmi.ftz.f32.f32 %f484, %f24; sub.ftz.f32 %f485, %f24, %f484; mov.f32 %f3079, 0f00000000; mov.u32 %r2034, -1; mov.f32 %f3080, 0f41000000; mov.u32 %r4351, %r2034; mov.f32 %f3078, %f3079; mov.f32 %f3077, %f3079; mov.f32 %f3076, %f3079; mov.f32 %f3075, %f3079; mov.f32 %f3074, %f3079; $L__BB199_153: cvt.rn.f32.s32 %f493, %r4351; add.ftz.f32 %f1652, %f484, %f493; mov.b32 %r193, %f1652; add.s32 %r194, %r193, -559038712; add.s32 %r195, %r193, -559038708; shf.l.wrap.b32 %r196, %r195, %r195, 4; mov.u32 %r4352, %r2034; $L__BB199_154: cvt.rn.f32.s32 %f501, %r4352; add.ftz.f32 %f1653, %f482, %f501; mov.b32 %r2037, %f1653; add.s32 %r198, %r2037, -559038712; xor.b32 %r2038, %r198, %r194; shf.l.wrap.b32 %r2039, %r198, %r198, 14; sub.s32 %r199, %r2038, %r2039; shf.l.wrap.b32 %r200, %r199, %r199, 11; add.s32 %r201, %r2037, -559038708; add.s32 %r202, %r201, %r195; mov.u32 %r4353, %r2034; $L__BB199_155: cvt.rn.f32.s32 %f509, %r4353; add.ftz.f32 %f1654, %f480, %f509; mov.b32 %r2040, %f1654; add.s32 %r2041, %r2040, -559038712; xor.b32 %r2042, %r2041, %r199; sub.s32 %r2043, %r2042, %r200; xor.b32 %r2044, %r2043, %r198; shf.l.wrap.b32 %r2045, %r2043, %r2043, 25; sub.s32 %r2046, %r2044, %r2045; xor.b32 %r2047, %r2046, %r199; shf.l.wrap.b32 %r2048, %r2046, %r2046, 16; sub.s32 %r2049, %r2047, %r2048; xor.b32 %r2050, %r2049, %r2043; shf.l.wrap.b32 %r2051, %r2049, %r2049, 4; sub.s32 %r2052, %r2050, %r2051; xor.b32 %r2053, %r2052, %r2046; shf.l.wrap.b32 %r2054, %r2052, %r2052, 14; sub.s32 %r2055, %r2053, %r2054; xor.b32 %r2056, %r2055, %r2049; shf.l.wrap.b32 %r2057, %r2055, %r2055, 24; sub.s32 %r2058, %r2056, %r2057; cvt.rn.f32.u32 %f1655, %r2058; mul.ftz.f32 %f1656, %f1655, 0f2F800000; sub.s32 %r2059, %r2040, %r193; xor.b32 %r2060, %r2059, %r196; sub.s32 %r2061, %r201, %r2060; shf.l.wrap.b32 %r2062, %r2060, %r2060, 6; xor.b32 %r2063, %r2062, %r2061; add.s32 %r2064, %r2060, %r202; sub.s32 %r2065, %r202, %r2063; shf.l.wrap.b32 %r2066, %r2063, %r2063, 8; xor.b32 %r2067, %r2066, %r2065; add.s32 %r2068, %r2063, %r2064; sub.s32 %r2069, %r2064, %r2067; shf.l.wrap.b32 %r2070, %r2067, %r2067, 16; xor.b32 %r2071, %r2070, %r2069; add.s32 %r2072, %r2067, %r2068; sub.s32 %r2073, %r2068, %r2071; shf.l.wrap.b32 %r2074, %r2071, %r2071, 19; xor.b32 %r2075, %r2074, %r2073; add.s32 %r2076, %r2071, %r2072; sub.s32 %r2077, %r2072, %r2075; shf.l.wrap.b32 %r2078, %r2075, %r2075, 4; add.s32 %r2079, %r2075, %r2076; add.s32 %r2080, %r2076, 1065353216; xor.b32 %r2081, %r2079, %r2077; xor.b32 %r2082, %r2081, %r2078; shf.l.wrap.b32 %r2083, %r2079, %r2079, 14; sub.s32 %r2084, %r2082, %r2083; xor.b32 %r2085, %r2084, %r2080; shf.l.wrap.b32 %r2086, %r2084, %r2084, 11; sub.s32 %r2087, %r2085, %r2086; xor.b32 %r2088, %r2087, %r2079; shf.l.wrap.b32 %r2089, %r2087, %r2087, 25; sub.s32 %r2090, %r2088, %r2089; xor.b32 %r2091, %r2090, %r2084; shf.l.wrap.b32 %r2092, %r2090, %r2090, 16; sub.s32 %r2093, %r2091, %r2092; xor.b32 %r2094, %r2093, %r2087; shf.l.wrap.b32 %r2095, %r2093, %r2093, 4; sub.s32 %r2096, %r2094, %r2095; xor.b32 %r2097, %r2096, %r2090; shf.l.wrap.b32 %r2098, %r2096, %r2096, 14; sub.s32 %r2099, %r2097, %r2098; xor.b32 %r2100, %r2099, %r2093; shf.l.wrap.b32 %r2101, %r2099, %r2099, 24; sub.s32 %r2102, %r2100, %r2101; cvt.rn.f32.u32 %f1657, %r2102; mul.ftz.f32 %f1658, %f1657, 0f2F800000; add.s32 %r2103, %r2076, 1073741824; xor.b32 %r2104, %r2084, %r2103; sub.s32 %r2105, %r2104, %r2086; xor.b32 %r2106, %r2105, %r2079; shf.l.wrap.b32 %r2107, %r2105, %r2105, 25; sub.s32 %r2108, %r2106, %r2107; xor.b32 %r2109, %r2108, %r2084; shf.l.wrap.b32 %r2110, %r2108, %r2108, 16; sub.s32 %r2111, %r2109, %r2110; xor.b32 %r2112, %r2111, %r2105; shf.l.wrap.b32 %r2113, %r2111, %r2111, 4; sub.s32 %r2114, %r2112, %r2113; xor.b32 %r2115, %r2114, %r2108; shf.l.wrap.b32 %r2116, %r2114, %r2114, 14; sub.s32 %r2117, %r2115, %r2116; xor.b32 %r2118, %r2117, %r2111; shf.l.wrap.b32 %r2119, %r2117, %r2117, 24; sub.s32 %r2120, %r2118, %r2119; cvt.rn.f32.u32 %f1659, %r2120; mul.ftz.f32 %f1660, %f1659, 0f2F800000; fma.rn.ftz.f32 %f510, %f19, %f1656, %f509; fma.rn.ftz.f32 %f511, %f19, %f1658, %f501; fma.rn.ftz.f32 %f512, %f19, %f1660, %f493; sub.ftz.f32 %f1661, %f510, %f481; sub.ftz.f32 %f1662, %f511, %f483; sub.ftz.f32 %f1663, %f512, %f485; mul.ftz.f32 %f1664, %f1662, %f1662; fma.rn.ftz.f32 %f1665, %f1661, %f1661, %f1664; fma.rn.ftz.f32 %f1666, %f1663, %f1663, %f1665; sqrt.approx.ftz.f32 %f513, %f1666; setp.geu.ftz.f32 %p74, %f513, %f3080; @%p74 bra $L__BB199_157; mov.f32 %f3074, %f509; mov.f32 %f3075, %f501; mov.f32 %f3076, %f493; mov.f32 %f3077, %f510; mov.f32 %f3078, %f511; mov.f32 %f3079, %f512; mov.f32 %f3080, %f513; $L__BB199_157: add.s32 %r204, %r4353, 1; setp.lt.s32 %p75, %r4353, 1; mov.u32 %r4353, %r204; @%p75 bra $L__BB199_155; add.s32 %r205, %r4352, 1; setp.lt.s32 %p76, %r4352, 1; mov.u32 %r4352, %r205; @%p76 bra $L__BB199_154; add.s32 %r206, %r4351, 1; setp.lt.s32 %p77, %r4351, 1; mov.u32 %r4351, %r206; @%p77 bra $L__BB199_153; mov.f32 %f3095, 0f00000000; mov.u32 %r2121, -1; mov.f32 %f3096, 0f41000000; mov.u32 %r4354, %r2121; mov.f32 %f3094, %f3095; mov.f32 %f3093, %f3095; $L__BB199_161: cvt.rn.f32.s32 %f1671, %r4354; add.ftz.f32 %f525, %f3076, %f1671; add.ftz.f32 %f1672, %f484, %f525; mov.b32 %r208, %f1672; add.s32 %r209, %r208, -559038712; add.s32 %r210, %r208, -559038708; shf.l.wrap.b32 %r211, %r210, %r210, 4; mov.u32 %r4355, %r2121; $L__BB199_162: or.b32 %r213, %r4355, %r4354; cvt.rn.f32.s32 %f1673, %r4355; add.ftz.f32 %f530, %f3075, %f1673; add.ftz.f32 %f1674, %f482, %f530; mov.b32 %r2124, %f1674; add.s32 %r214, %r2124, -559038712; xor.b32 %r2125, %r214, %r209; shf.l.wrap.b32 %r2126, %r214, %r214, 14; sub.s32 %r215, %r2125, %r2126; shf.l.wrap.b32 %r216, %r215, %r215, 11; add.s32 %r217, %r2124, -559038708; add.s32 %r218, %r217, %r210; mov.u32 %r4356, %r2121; $L__BB199_163: mov.u32 %r219, %r4356; or.b32 %r2127, %r213, %r219; setp.eq.s32 %p78, %r2127, 0; @%p78 bra $L__BB199_166; cvt.rn.f32.s32 %f1675, %r219; add.ftz.f32 %f1676, %f3074, %f1675; add.ftz.f32 %f1677, %f480, %f1676; mov.b32 %r2128, %f1677; add.s32 %r2129, %r2128, -559038712; xor.b32 %r2130, %r2129, %r215; sub.s32 %r2131, %r2130, %r216; xor.b32 %r2132, %r2131, %r214; shf.l.wrap.b32 %r2133, %r2131, %r2131, 25; sub.s32 %r2134, %r2132, %r2133; xor.b32 %r2135, %r2134, %r215; shf.l.wrap.b32 %r2136, %r2134, %r2134, 16; sub.s32 %r2137, %r2135, %r2136; xor.b32 %r2138, %r2137, %r2131; shf.l.wrap.b32 %r2139, %r2137, %r2137, 4; sub.s32 %r2140, %r2138, %r2139; xor.b32 %r2141, %r2140, %r2134; shf.l.wrap.b32 %r2142, %r2140, %r2140, 14; sub.s32 %r2143, %r2141, %r2142; xor.b32 %r2144, %r2143, %r2137; shf.l.wrap.b32 %r2145, %r2143, %r2143, 24; sub.s32 %r2146, %r2144, %r2145; cvt.rn.f32.u32 %f1678, %r2146; mul.ftz.f32 %f1679, %f1678, 0f2F800000; sub.s32 %r2147, %r2128, %r208; xor.b32 %r2148, %r2147, %r211; sub.s32 %r2149, %r217, %r2148; shf.l.wrap.b32 %r2150, %r2148, %r2148, 6; xor.b32 %r2151, %r2150, %r2149; add.s32 %r2152, %r2148, %r218; sub.s32 %r2153, %r218, %r2151; shf.l.wrap.b32 %r2154, %r2151, %r2151, 8; xor.b32 %r2155, %r2154, %r2153; add.s32 %r2156, %r2151, %r2152; sub.s32 %r2157, %r2152, %r2155; shf.l.wrap.b32 %r2158, %r2155, %r2155, 16; xor.b32 %r2159, %r2158, %r2157; add.s32 %r2160, %r2155, %r2156; sub.s32 %r2161, %r2156, %r2159; shf.l.wrap.b32 %r2162, %r2159, %r2159, 19; xor.b32 %r2163, %r2162, %r2161; add.s32 %r2164, %r2159, %r2160; sub.s32 %r2165, %r2160, %r2163; shf.l.wrap.b32 %r2166, %r2163, %r2163, 4; add.s32 %r2167, %r2163, %r2164; add.s32 %r2168, %r2164, 1065353216; xor.b32 %r2169, %r2167, %r2165; xor.b32 %r2170, %r2169, %r2166; shf.l.wrap.b32 %r2171, %r2167, %r2167, 14; sub.s32 %r2172, %r2170, %r2171; xor.b32 %r2173, %r2172, %r2168; shf.l.wrap.b32 %r2174, %r2172, %r2172, 11; sub.s32 %r2175, %r2173, %r2174; xor.b32 %r2176, %r2175, %r2167; shf.l.wrap.b32 %r2177, %r2175, %r2175, 25; sub.s32 %r2178, %r2176, %r2177; xor.b32 %r2179, %r2178, %r2172; shf.l.wrap.b32 %r2180, %r2178, %r2178, 16; sub.s32 %r2181, %r2179, %r2180; xor.b32 %r2182, %r2181, %r2175; shf.l.wrap.b32 %r2183, %r2181, %r2181, 4; sub.s32 %r2184, %r2182, %r2183; xor.b32 %r2185, %r2184, %r2178; shf.l.wrap.b32 %r2186, %r2184, %r2184, 14; sub.s32 %r2187, %r2185, %r2186; xor.b32 %r2188, %r2187, %r2181; shf.l.wrap.b32 %r2189, %r2187, %r2187, 24; sub.s32 %r2190, %r2188, %r2189; cvt.rn.f32.u32 %f1680, %r2190; mul.ftz.f32 %f1681, %f1680, 0f2F800000; add.s32 %r2191, %r2164, 1073741824; xor.b32 %r2192, %r2172, %r2191; sub.s32 %r2193, %r2192, %r2174; xor.b32 %r2194, %r2193, %r2167; shf.l.wrap.b32 %r2195, %r2193, %r2193, 25; sub.s32 %r2196, %r2194, %r2195; xor.b32 %r2197, %r2196, %r2172; shf.l.wrap.b32 %r2198, %r2196, %r2196, 16; sub.s32 %r2199, %r2197, %r2198; xor.b32 %r2200, %r2199, %r2193; shf.l.wrap.b32 %r2201, %r2199, %r2199, 4; sub.s32 %r2202, %r2200, %r2201; xor.b32 %r2203, %r2202, %r2196; shf.l.wrap.b32 %r2204, %r2202, %r2202, 14; sub.s32 %r2205, %r2203, %r2204; xor.b32 %r2206, %r2205, %r2199; shf.l.wrap.b32 %r2207, %r2205, %r2205, 24; sub.s32 %r2208, %r2206, %r2207; cvt.rn.f32.u32 %f1682, %r2208; mul.ftz.f32 %f1683, %f1682, 0f2F800000; fma.rn.ftz.f32 %f535, %f19, %f1679, %f1676; fma.rn.ftz.f32 %f536, %f19, %f1681, %f530; fma.rn.ftz.f32 %f537, %f19, %f1683, %f525; sub.ftz.f32 %f1684, %f3077, %f535; sub.ftz.f32 %f1685, %f3078, %f536; sub.ftz.f32 %f1686, %f3079, %f537; mul.ftz.f32 %f1687, %f1685, %f1685; fma.rn.ftz.f32 %f1688, %f1684, %f1684, %f1687; fma.rn.ftz.f32 %f1689, %f1686, %f1686, %f1688; sqrt.approx.ftz.f32 %f538, %f1689; setp.geu.ftz.f32 %p79, %f538, %f3096; @%p79 bra $L__BB199_166; mov.f32 %f3093, %f535; mov.f32 %f3094, %f536; mov.f32 %f3095, %f537; mov.f32 %f3096, %f538; $L__BB199_166: add.s32 %r4356, %r219, 1; setp.lt.s32 %p80, %r219, 1; @%p80 bra $L__BB199_163; add.s32 %r221, %r4355, 1; setp.lt.s32 %p81, %r4355, 1; mov.u32 %r4355, %r221; @%p81 bra $L__BB199_162; add.s32 %r222, %r4354, 1; setp.lt.s32 %p82, %r4354, 1; mov.u32 %r4354, %r222; @%p82 bra $L__BB199_161; sub.ftz.f32 %f1697, %f3093, %f3077; sub.ftz.f32 %f1698, %f3094, %f3078; mul.ftz.f32 %f1699, %f1698, %f1698; fma.rn.ftz.f32 %f1700, %f1697, %f1697, %f1699; sub.ftz.f32 %f1701, %f3095, %f3079; fma.rn.ftz.f32 %f1702, %f1701, %f1701, %f1700; sqrt.approx.ftz.f32 %f1703, %f1702; div.approx.ftz.f32 %f3388, %f1703, %f1184; mov.f32 %f3386, 0f00000000; mov.f32 %f3389, %f3386; mov.f32 %f3390, %f3386; mov.f32 %f3391, %f3386; mov.f32 %f3240, %f3386; mov.f32 %f3241, %f3386; mov.f32 %f3242, %f3386; bra.uni $L__BB199_250; $L__BB199_19: setp.eq.s32 %p131, %r389, 0; @%p131 bra $L__BB199_20; setp.eq.s32 %p132, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3364, %f3387; mov.f32 %f3365, %f3387; @%p132 bra $L__BB199_253; bra.uni $L__BB199_362; $L__BB199_253: ld.param.u32 %r4321, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f949, %f22; sub.ftz.f32 %f950, %f22, %f949; cvt.rmi.ftz.f32.f32 %f951, %f23; sub.ftz.f32 %f952, %f23, %f951; rcp.approx.ftz.f32 %f953, %f15; add.ftz.f32 %f2354, %f949, 0fBF800000; mov.b32 %r3502, %f2354; add.s32 %r359, %r3502, -559038716; add.s32 %r360, %r3502, -559038712; add.ftz.f32 %f2355, %f949, 0f00000000; mov.b32 %r3503, %f2355; add.s32 %r361, %r3503, -559038716; add.s32 %r362, %r3503, -559038712; add.ftz.f32 %f2356, %f949, 0f3F800000; mov.b32 %r3504, %f2356; add.s32 %r363, %r3504, -559038716; add.s32 %r364, %r3504, -559038712; mov.f32 %f3334, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4377, -1; setp.eq.s32 %p158, %r4321, 0; setp.eq.s32 %p171, %r4321, 1; setp.eq.s32 %p172, %r4321, 2; setp.ne.s32 %p173, %r4321, 3; mov.f32 %f3335, %f3334; mov.f32 %f3332, %f3334; mov.f32 %f3333, %f3334; mov.f32 %f3330, %f3386; mov.f32 %f3329, %f3334; mov.f32 %f3328, %f3334; mov.f32 %f3327, %f3334; mov.f32 %f3326, %f3334; $L__BB199_254: cvt.rn.f32.s32 %f964, %r4377; add.ftz.f32 %f2357, %f951, %f964; mov.b32 %r3505, %f2357; add.s32 %r366, %r3505, -559038716; xor.b32 %r3506, %r366, -559038716; shf.l.wrap.b32 %r3507, %r366, %r366, 14; sub.s32 %r367, %r3506, %r3507; shf.l.wrap.b32 %r368, %r367, %r367, 11; add.s32 %r3508, %r3505, -559038712; xor.b32 %r3509, %r3508, 506314504; shf.l.wrap.b32 %r3510, %r3508, %r3508, 14; sub.s32 %r369, %r3509, %r3510; shf.l.wrap.b32 %r370, %r369, %r369, 11; xor.b32 %r3511, %r359, %r367; sub.s32 %r3512, %r3511, %r368; xor.b32 %r3513, %r3512, %r366; shf.l.wrap.b32 %r3514, %r3512, %r3512, 25; sub.s32 %r3515, %r3513, %r3514; xor.b32 %r3516, %r3515, %r367; shf.l.wrap.b32 %r3517, %r3515, %r3515, 16; sub.s32 %r3518, %r3516, %r3517; xor.b32 %r3519, %r3518, %r3512; shf.l.wrap.b32 %r3520, %r3518, %r3518, 4; sub.s32 %r3521, %r3519, %r3520; xor.b32 %r3522, %r3521, %r3515; shf.l.wrap.b32 %r3523, %r3521, %r3521, 14; sub.s32 %r3524, %r3522, %r3523; xor.b32 %r3525, %r3524, %r3518; shf.l.wrap.b32 %r3526, %r3524, %r3524, 24; sub.s32 %r3527, %r3525, %r3526; cvt.rn.f32.u32 %f2358, %r3527; mul.ftz.f32 %f2359, %f2358, 0f2F800000; xor.b32 %r3528, %r360, %r369; sub.s32 %r3529, %r3528, %r370; xor.b32 %r3530, %r3529, %r3508; shf.l.wrap.b32 %r3531, %r3529, %r3529, 25; sub.s32 %r3532, %r3530, %r3531; xor.b32 %r3533, %r3532, %r369; shf.l.wrap.b32 %r3534, %r3532, %r3532, 16; sub.s32 %r3535, %r3533, %r3534; xor.b32 %r3536, %r3535, %r3529; shf.l.wrap.b32 %r3537, %r3535, %r3535, 4; sub.s32 %r3538, %r3536, %r3537; xor.b32 %r3539, %r3538, %r3532; shf.l.wrap.b32 %r3540, %r3538, %r3538, 14; sub.s32 %r3541, %r3539, %r3540; xor.b32 %r3542, %r3541, %r3535; shf.l.wrap.b32 %r3543, %r3541, %r3541, 24; sub.s32 %r3544, %r3542, %r3543; cvt.rn.f32.u32 %f2360, %r3544; mul.ftz.f32 %f2361, %f2360, 0f2F800000; fma.rn.ftz.f32 %f965, %f19, %f2359, 0fBF800000; fma.rn.ftz.f32 %f966, %f19, %f2361, %f964; @%p158 bra $L__BB199_304; @%p171 bra $L__BB199_303; bra.uni $L__BB199_256; $L__BB199_303: sub.ftz.f32 %f2380, %f965, %f950; abs.ftz.f32 %f2381, %f2380; sub.ftz.f32 %f2382, %f966, %f952; abs.ftz.f32 %f2383, %f2382; add.ftz.f32 %f3303, %f2381, %f2383; bra.uni $L__BB199_305; $L__BB199_304: sub.ftz.f32 %f2384, %f965, %f950; sub.ftz.f32 %f2385, %f966, %f952; mul.ftz.f32 %f2386, %f2385, %f2385; fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386; sqrt.approx.ftz.f32 %f3303, %f2387; bra.uni $L__BB199_305; $L__BB199_256: @%p172 bra $L__BB199_302; bra.uni $L__BB199_257; $L__BB199_302: sub.ftz.f32 %f2376, %f965, %f950; abs.ftz.f32 %f2377, %f2376; sub.ftz.f32 %f2378, %f966, %f952; abs.ftz.f32 %f2379, %f2378; max.ftz.f32 %f3303, %f2377, %f2379; bra.uni $L__BB199_305; $L__BB199_257: mov.f32 %f3303, 0f00000000; @%p173 bra $L__BB199_305; sub.ftz.f32 %f2363, %f965, %f950; abs.ftz.f32 %f2364, %f2363; lg2.approx.ftz.f32 %f2365, %f2364; mul.ftz.f32 %f2366, %f15, %f2365; ex2.approx.ftz.f32 %f2367, %f2366; sub.ftz.f32 %f2368, %f966, %f952; abs.ftz.f32 %f2369, %f2368; lg2.approx.ftz.f32 %f2370, %f2369; mul.ftz.f32 %f2371, %f15, %f2370; ex2.approx.ftz.f32 %f2372, %f2371; add.ftz.f32 %f2373, %f2367, %f2372; lg2.approx.ftz.f32 %f2374, %f2373; mul.ftz.f32 %f2375, %f953, %f2374; ex2.approx.ftz.f32 %f3303, %f2375; $L__BB199_305: setp.lt.ftz.f32 %p162, %f3303, %f3330; @%p162 bra $L__BB199_308; bra.uni $L__BB199_306; $L__BB199_308: mov.f32 %f974, 0fBF800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3308, %f3303; mov.f32 %f3386, %f3330; mov.f32 %f3332, %f974; mov.f32 %f3333, %f964; mov.f32 %f3334, %f965; mov.f32 %f3335, %f966; bra.uni $L__BB199_309; $L__BB199_306: setp.geu.ftz.f32 %p163, %f3303, %f3386; mov.f32 %f3308, %f3330; @%p163 bra $L__BB199_309; mov.f32 %f3328, 0fBF800000; mov.f32 %f3326, %f965; mov.f32 %f3327, %f966; mov.f32 %f3329, %f964; mov.f32 %f3308, %f3330; mov.f32 %f3386, %f3303; $L__BB199_309: xor.b32 %r3545, %r361, %r367; sub.s32 %r3546, %r3545, %r368; xor.b32 %r3547, %r3546, %r366; shf.l.wrap.b32 %r3548, %r3546, %r3546, 25; sub.s32 %r3549, %r3547, %r3548; xor.b32 %r3550, %r3549, %r367; shf.l.wrap.b32 %r3551, %r3549, %r3549, 16; sub.s32 %r3552, %r3550, %r3551; xor.b32 %r3553, %r3552, %r3546; shf.l.wrap.b32 %r3554, %r3552, %r3552, 4; sub.s32 %r3555, %r3553, %r3554; xor.b32 %r3556, %r3555, %r3549; shf.l.wrap.b32 %r3557, %r3555, %r3555, 14; sub.s32 %r3558, %r3556, %r3557; xor.b32 %r3559, %r3558, %r3552; shf.l.wrap.b32 %r3560, %r3558, %r3558, 24; sub.s32 %r3561, %r3559, %r3560; cvt.rn.f32.u32 %f2389, %r3561; mul.ftz.f32 %f2390, %f2389, 0f2F800000; xor.b32 %r3562, %r362, %r369; sub.s32 %r3563, %r3562, %r370; add.s32 %r3564, %r366, 4; xor.b32 %r3565, %r3563, %r3564; shf.l.wrap.b32 %r3566, %r3563, %r3563, 25; sub.s32 %r3567, %r3565, %r3566; xor.b32 %r3568, %r3567, %r369; shf.l.wrap.b32 %r3569, %r3567, %r3567, 16; sub.s32 %r3570, %r3568, %r3569; xor.b32 %r3571, %r3570, %r3563; shf.l.wrap.b32 %r3572, %r3570, %r3570, 4; sub.s32 %r3573, %r3571, %r3572; xor.b32 %r3574, %r3573, %r3567; shf.l.wrap.b32 %r3575, %r3573, %r3573, 14; sub.s32 %r3576, %r3574, %r3575; xor.b32 %r3577, %r3576, %r3570; shf.l.wrap.b32 %r3578, %r3576, %r3576, 24; sub.s32 %r3579, %r3577, %r3578; cvt.rn.f32.u32 %f2391, %r3579; mul.ftz.f32 %f2392, %f2391, 0f2F800000; fma.rn.ftz.f32 %f990, %f19, %f2390, 0f00000000; fma.rn.ftz.f32 %f991, %f19, %f2392, %f964; @%p158 bra $L__BB199_316; @%p171 bra $L__BB199_315; bra.uni $L__BB199_311; $L__BB199_315: sub.ftz.f32 %f2411, %f990, %f950; abs.ftz.f32 %f2412, %f2411; sub.ftz.f32 %f2413, %f991, %f952; abs.ftz.f32 %f2414, %f2413; add.ftz.f32 %f3314, %f2412, %f2414; bra.uni $L__BB199_317; $L__BB199_316: sub.ftz.f32 %f2415, %f990, %f950; sub.ftz.f32 %f2416, %f991, %f952; mul.ftz.f32 %f2417, %f2416, %f2416; fma.rn.ftz.f32 %f2418, %f2415, %f2415, %f2417; sqrt.approx.ftz.f32 %f3314, %f2418; bra.uni $L__BB199_317; $L__BB199_311: @%p172 bra $L__BB199_314; bra.uni $L__BB199_312; $L__BB199_314: sub.ftz.f32 %f2407, %f990, %f950; abs.ftz.f32 %f2408, %f2407; sub.ftz.f32 %f2409, %f991, %f952; abs.ftz.f32 %f2410, %f2409; max.ftz.f32 %f3314, %f2408, %f2410; bra.uni $L__BB199_317; $L__BB199_312: mov.f32 %f3314, 0f00000000; @%p173 bra $L__BB199_317; sub.ftz.f32 %f2394, %f990, %f950; abs.ftz.f32 %f2395, %f2394; lg2.approx.ftz.f32 %f2396, %f2395; mul.ftz.f32 %f2397, %f15, %f2396; ex2.approx.ftz.f32 %f2398, %f2397; sub.ftz.f32 %f2399, %f991, %f952; abs.ftz.f32 %f2400, %f2399; lg2.approx.ftz.f32 %f2401, %f2400; mul.ftz.f32 %f2402, %f15, %f2401; ex2.approx.ftz.f32 %f2403, %f2402; add.ftz.f32 %f2404, %f2398, %f2403; lg2.approx.ftz.f32 %f2405, %f2404; mul.ftz.f32 %f2406, %f953, %f2405; ex2.approx.ftz.f32 %f3314, %f2406; $L__BB199_317: setp.lt.ftz.f32 %p168, %f3314, %f3308; @%p168 bra $L__BB199_320; bra.uni $L__BB199_318; $L__BB199_320: mov.f32 %f999, 0f00000000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3319, %f3314; mov.f32 %f3386, %f3308; mov.f32 %f3332, %f999; mov.f32 %f3333, %f964; mov.f32 %f3334, %f990; mov.f32 %f3335, %f991; bra.uni $L__BB199_321; $L__BB199_318: setp.geu.ftz.f32 %p169, %f3314, %f3386; mov.f32 %f3319, %f3308; @%p169 bra $L__BB199_321; mov.f32 %f3328, 0f00000000; mov.f32 %f3326, %f990; mov.f32 %f3327, %f991; mov.f32 %f3329, %f964; mov.f32 %f3319, %f3308; mov.f32 %f3386, %f3314; $L__BB199_321: xor.b32 %r3580, %r363, %r367; sub.s32 %r3581, %r3580, %r368; xor.b32 %r3582, %r3581, %r366; shf.l.wrap.b32 %r3583, %r3581, %r3581, 25; sub.s32 %r3584, %r3582, %r3583; xor.b32 %r3585, %r3584, %r367; shf.l.wrap.b32 %r3586, %r3584, %r3584, 16; sub.s32 %r3587, %r3585, %r3586; xor.b32 %r3588, %r3587, %r3581; shf.l.wrap.b32 %r3589, %r3587, %r3587, 4; sub.s32 %r3590, %r3588, %r3589; xor.b32 %r3591, %r3590, %r3584; shf.l.wrap.b32 %r3592, %r3590, %r3590, 14; sub.s32 %r3593, %r3591, %r3592; xor.b32 %r3594, %r3593, %r3587; shf.l.wrap.b32 %r3595, %r3593, %r3593, 24; sub.s32 %r3596, %r3594, %r3595; cvt.rn.f32.u32 %f2420, %r3596; mul.ftz.f32 %f2421, %f2420, 0f2F800000; xor.b32 %r3597, %r364, %r369; sub.s32 %r3598, %r3597, %r370; xor.b32 %r3600, %r3598, %r3564; shf.l.wrap.b32 %r3601, %r3598, %r3598, 25; sub.s32 %r3602, %r3600, %r3601; xor.b32 %r3603, %r3602, %r369; shf.l.wrap.b32 %r3604, %r3602, %r3602, 16; sub.s32 %r3605, %r3603, %r3604; xor.b32 %r3606, %r3605, %r3598; shf.l.wrap.b32 %r3607, %r3605, %r3605, 4; sub.s32 %r3608, %r3606, %r3607; xor.b32 %r3609, %r3608, %r3602; shf.l.wrap.b32 %r3610, %r3608, %r3608, 14; sub.s32 %r3611, %r3609, %r3610; xor.b32 %r3612, %r3611, %r3605; shf.l.wrap.b32 %r3613, %r3611, %r3611, 24; sub.s32 %r3614, %r3612, %r3613; cvt.rn.f32.u32 %f2422, %r3614; mul.ftz.f32 %f2423, %f2422, 0f2F800000; fma.rn.ftz.f32 %f1015, %f19, %f2421, 0f3F800000; fma.rn.ftz.f32 %f1016, %f19, %f2423, %f964; @%p158 bra $L__BB199_328; @%p171 bra $L__BB199_327; bra.uni $L__BB199_323; $L__BB199_327: sub.ftz.f32 %f2442, %f1015, %f950; abs.ftz.f32 %f2443, %f2442; sub.ftz.f32 %f2444, %f1016, %f952; abs.ftz.f32 %f2445, %f2444; add.ftz.f32 %f3325, %f2443, %f2445; bra.uni $L__BB199_329; $L__BB199_328: sub.ftz.f32 %f2446, %f1015, %f950; sub.ftz.f32 %f2447, %f1016, %f952; mul.ftz.f32 %f2448, %f2447, %f2447; fma.rn.ftz.f32 %f2449, %f2446, %f2446, %f2448; sqrt.approx.ftz.f32 %f3325, %f2449; bra.uni $L__BB199_329; $L__BB199_323: @%p172 bra $L__BB199_326; bra.uni $L__BB199_324; $L__BB199_326: sub.ftz.f32 %f2438, %f1015, %f950; abs.ftz.f32 %f2439, %f2438; sub.ftz.f32 %f2440, %f1016, %f952; abs.ftz.f32 %f2441, %f2440; max.ftz.f32 %f3325, %f2439, %f2441; bra.uni $L__BB199_329; $L__BB199_324: mov.f32 %f3325, 0f00000000; @%p173 bra $L__BB199_329; sub.ftz.f32 %f2425, %f1015, %f950; abs.ftz.f32 %f2426, %f2425; lg2.approx.ftz.f32 %f2427, %f2426; mul.ftz.f32 %f2428, %f15, %f2427; ex2.approx.ftz.f32 %f2429, %f2428; sub.ftz.f32 %f2430, %f1016, %f952; abs.ftz.f32 %f2431, %f2430; lg2.approx.ftz.f32 %f2432, %f2431; mul.ftz.f32 %f2433, %f15, %f2432; ex2.approx.ftz.f32 %f2434, %f2433; add.ftz.f32 %f2435, %f2429, %f2434; lg2.approx.ftz.f32 %f2436, %f2435; mul.ftz.f32 %f2437, %f953, %f2436; ex2.approx.ftz.f32 %f3325, %f2437; $L__BB199_329: setp.lt.ftz.f32 %p174, %f3325, %f3319; @%p174 bra $L__BB199_332; bra.uni $L__BB199_330; $L__BB199_332: mov.f32 %f1024, 0f3F800000; mov.f32 %f3326, %f3334; mov.f32 %f3327, %f3335; mov.f32 %f3328, %f3332; mov.f32 %f3329, %f3333; mov.f32 %f3330, %f3325; mov.f32 %f3386, %f3319; mov.f32 %f3332, %f1024; mov.f32 %f3333, %f964; mov.f32 %f3334, %f1015; mov.f32 %f3335, %f1016; bra.uni $L__BB199_333; $L__BB199_330: setp.geu.ftz.f32 %p175, %f3325, %f3386; mov.f32 %f3330, %f3319; @%p175 bra $L__BB199_333; mov.f32 %f3328, 0f3F800000; mov.f32 %f3326, %f1015; mov.f32 %f3327, %f1016; mov.f32 %f3329, %f964; mov.f32 %f3330, %f3319; mov.f32 %f3386, %f3325; $L__BB199_333: add.s32 %r4377, %r4377, 1; setp.ne.s32 %p176, %r4377, 2; @%p176 bra $L__BB199_254; add.ftz.f32 %f2452, %f949, %f3328; mov.b32 %r3615, %f2452; add.ftz.f32 %f2453, %f951, %f3329; mov.b32 %r3616, %f2453; add.s32 %r3617, %r3616, -559038716; add.s32 %r3618, %r3615, -559038716; xor.b32 %r3619, %r3617, -559038716; shf.l.wrap.b32 %r3620, %r3617, %r3617, 14; sub.s32 %r3621, %r3619, %r3620; xor.b32 %r3622, %r3621, %r3618; shf.l.wrap.b32 %r3623, %r3621, %r3621, 11; sub.s32 %r3624, %r3622, %r3623; xor.b32 %r3625, %r3624, %r3617; shf.l.wrap.b32 %r3626, %r3624, %r3624, 25; sub.s32 %r3627, %r3625, %r3626; xor.b32 %r3628, %r3627, %r3621; shf.l.wrap.b32 %r3629, %r3627, %r3627, 16; sub.s32 %r3630, %r3628, %r3629; xor.b32 %r3631, %r3630, %r3624; shf.l.wrap.b32 %r3632, %r3630, %r3630, 4; sub.s32 %r3633, %r3631, %r3632; xor.b32 %r3634, %r3633, %r3627; shf.l.wrap.b32 %r3635, %r3633, %r3633, 14; sub.s32 %r3636, %r3634, %r3635; xor.b32 %r3637, %r3636, %r3630; shf.l.wrap.b32 %r3638, %r3636, %r3636, 24; sub.s32 %r3639, %r3637, %r3638; cvt.rn.f32.u32 %f2454, %r3639; mul.ftz.f32 %f3391, %f2454, 0f2F800000; add.s32 %r3640, %r3616, -559038712; add.s32 %r3641, %r3615, -559038712; xor.b32 %r3642, %r3640, 506314504; shf.l.wrap.b32 %r3643, %r3640, %r3640, 14; sub.s32 %r3644, %r3642, %r3643; xor.b32 %r3645, %r3644, %r3641; shf.l.wrap.b32 %r3646, %r3644, %r3644, 11; sub.s32 %r3647, %r3645, %r3646; xor.b32 %r3648, %r3647, %r3640; shf.l.wrap.b32 %r3649, %r3647, %r3647, 25; sub.s32 %r3650, %r3648, %r3649; xor.b32 %r3651, %r3650, %r3644; shf.l.wrap.b32 %r3652, %r3650, %r3650, 16; sub.s32 %r3653, %r3651, %r3652; xor.b32 %r3654, %r3653, %r3647; shf.l.wrap.b32 %r3655, %r3653, %r3653, 4; sub.s32 %r3656, %r3654, %r3655; xor.b32 %r3657, %r3656, %r3650; shf.l.wrap.b32 %r3658, %r3656, %r3656, 14; sub.s32 %r3659, %r3657, %r3658; xor.b32 %r3660, %r3659, %r3653; shf.l.wrap.b32 %r3661, %r3659, %r3659, 24; sub.s32 %r3662, %r3660, %r3661; cvt.rn.f32.u32 %f2455, %r3662; mul.ftz.f32 %f3390, %f2455, 0f2F800000; xor.b32 %r3663, %r3640, 514703112; sub.s32 %r3664, %r3663, %r3643; xor.b32 %r3665, %r3664, %r3641; shf.l.wrap.b32 %r3666, %r3664, %r3664, 11; sub.s32 %r3667, %r3665, %r3666; xor.b32 %r3668, %r3667, %r3640; shf.l.wrap.b32 %r3669, %r3667, %r3667, 25; sub.s32 %r3670, %r3668, %r3669; xor.b32 %r3671, %r3670, %r3664; shf.l.wrap.b32 %r3672, %r3670, %r3670, 16; sub.s32 %r3673, %r3671, %r3672; xor.b32 %r3674, %r3673, %r3667; shf.l.wrap.b32 %r3675, %r3673, %r3673, 4; sub.s32 %r3676, %r3674, %r3675; xor.b32 %r3677, %r3676, %r3670; shf.l.wrap.b32 %r3678, %r3676, %r3676, 14; sub.s32 %r3679, %r3677, %r3678; xor.b32 %r3680, %r3679, %r3673; shf.l.wrap.b32 %r3681, %r3679, %r3679, 24; sub.s32 %r3682, %r3680, %r3681; cvt.rn.f32.u32 %f2456, %r3682; mul.ftz.f32 %f3389, %f2456, 0f2F800000; add.ftz.f32 %f3365, %f949, %f3326; add.ftz.f32 %f3364, %f951, %f3327; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_362; $L__BB199_29: setp.eq.s32 %p15, %r389, 0; @%p15 bra $L__BB199_117; setp.eq.s32 %p16, %r389, 1; mov.f32 %f3386, %f3394; mov.f32 %f3388, %f3394; mov.f32 %f3389, %f3394; mov.f32 %f3390, %f3394; mov.f32 %f3391, %f3394; @%p16 bra $L__BB199_31; bra.uni $L__BB199_136; $L__BB199_31: ld.param.u32 %r4313, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f250, %f22; sub.ftz.f32 %f251, %f22, %f250; cvt.rmi.ftz.f32.f32 %f252, %f23; sub.ftz.f32 %f253, %f23, %f252; cvt.rmi.ftz.f32.f32 %f254, %f24; sub.ftz.f32 %f255, %f24, %f254; cvt.rmi.ftz.f32.f32 %f256, %f21; sub.ftz.f32 %f257, %f21, %f256; rcp.approx.ftz.f32 %f258, %f15; mov.f32 %f2990, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1384, -1; setp.eq.s32 %p48, %r4313, 0; setp.eq.s32 %p49, %r4313, 1; setp.eq.s32 %p50, %r4313, 2; setp.ne.s32 %p51, %r4313, 3; mov.u32 %r4343, %r1384; mov.f32 %f2991, %f2990; mov.f32 %f2992, %f2990; mov.f32 %f2993, %f2990; mov.f32 %f2986, %f2990; mov.f32 %f2987, %f2990; mov.f32 %f2988, %f2990; mov.f32 %f2989, %f2990; mov.f32 %f2984, %f3386; mov.f32 %f2983, %f2990; mov.f32 %f2982, %f2990; mov.f32 %f2981, %f2990; mov.f32 %f2980, %f2990; mov.f32 %f2979, %f2990; mov.f32 %f2978, %f2990; mov.f32 %f2977, %f2990; mov.f32 %f2976, %f2990; $L__BB199_32: cvt.rn.f32.s32 %f277, %r4343; add.ftz.f32 %f1473, %f256, %f277; mov.b32 %r145, %f1473; add.s32 %r146, %r145, -559038708; shf.l.wrap.b32 %r147, %r146, %r146, 4; mov.u32 %r4344, %r1384; $L__BB199_33: cvt.rn.f32.s32 %f296, %r4344; add.ftz.f32 %f1474, %f254, %f296; mov.b32 %r149, %f1474; add.s32 %r150, %r149, -559038708; shf.l.wrap.b32 %r151, %r150, %r150, 4; add.s32 %r152, %r150, %r146; mov.u32 %r4345, %r1384; $L__BB199_34: cvt.rn.f32.s32 %f315, %r4345; add.ftz.f32 %f1475, %f252, %f315; mov.b32 %r154, %f1475; add.s32 %r155, %r154, -559038708; add.s32 %r156, %r155, %r150; sub.s32 %r1388, %r145, %r154; shf.l.wrap.b32 %r1389, %r155, %r155, 4; xor.b32 %r157, %r1389, %r1388; shf.l.wrap.b32 %r158, %r157, %r157, 6; sub.s32 %r1390, %r154, %r145; xor.b32 %r1391, %r1390, %r147; sub.s32 %r1392, %r150, %r1391; shf.l.wrap.b32 %r1393, %r1391, %r1391, 6; xor.b32 %r1394, %r1393, %r1392; add.s32 %r1395, %r1391, %r152; sub.s32 %r1396, %r152, %r1394; shf.l.wrap.b32 %r1397, %r1394, %r1394, 8; xor.b32 %r1398, %r1397, %r1396; add.s32 %r1399, %r1394, %r1395; sub.s32 %r1400, %r1395, %r1398; shf.l.wrap.b32 %r1401, %r1398, %r1398, 16; xor.b32 %r1402, %r1401, %r1400; add.s32 %r1403, %r1398, %r1399; sub.s32 %r1404, %r1399, %r1402; shf.l.wrap.b32 %r1405, %r1402, %r1402, 19; xor.b32 %r1406, %r1405, %r1404; add.s32 %r159, %r1402, %r1403; sub.s32 %r1407, %r1403, %r1406; shf.l.wrap.b32 %r1408, %r1406, %r1406, 4; add.s32 %r160, %r1406, %r159; xor.b32 %r1409, %r160, %r1407; xor.b32 %r1410, %r1409, %r1408; shf.l.wrap.b32 %r1411, %r160, %r160, 14; sub.s32 %r161, %r1410, %r1411; shf.l.wrap.b32 %r162, %r161, %r161, 11; mov.u32 %r4346, %r1384; $L__BB199_35: mov.f32 %f325, %f2984; mov.u32 %r163, %r4346; cvt.rn.f32.s32 %f334, %r163; add.ftz.f32 %f1476, %f250, %f334; mov.b32 %r1412, %f1476; sub.s32 %r1413, %r1412, %r149; xor.b32 %r1414, %r1413, %r151; sub.s32 %r1415, %r155, %r1414; shf.l.wrap.b32 %r1416, %r1414, %r1414, 6; xor.b32 %r1417, %r1416, %r1415; add.s32 %r1418, %r1414, %r156; sub.s32 %r1419, %r156, %r1417; shf.l.wrap.b32 %r1420, %r1417, %r1417, 8; xor.b32 %r1421, %r1420, %r1419; add.s32 %r1422, %r1417, %r1418; sub.s32 %r1423, %r1418, %r1421; shf.l.wrap.b32 %r1424, %r1421, %r1421, 16; xor.b32 %r1425, %r1424, %r1423; add.s32 %r1426, %r1421, %r1422; sub.s32 %r1427, %r1422, %r1425; shf.l.wrap.b32 %r1428, %r1425, %r1425, 19; xor.b32 %r1429, %r1428, %r1427; add.s32 %r1430, %r1425, %r1426; sub.s32 %r1431, %r1426, %r1429; shf.l.wrap.b32 %r1432, %r1429, %r1429, 4; add.s32 %r1433, %r1429, %r1430; add.s32 %r1434, %r1430, %r145; xor.b32 %r1435, %r1433, %r1431; xor.b32 %r1436, %r1435, %r1432; shf.l.wrap.b32 %r1437, %r1433, %r1433, 14; sub.s32 %r1438, %r1436, %r1437; xor.b32 %r1439, %r1438, %r1434; shf.l.wrap.b32 %r1440, %r1438, %r1438, 11; sub.s32 %r1441, %r1439, %r1440; xor.b32 %r1442, %r1441, %r1433; shf.l.wrap.b32 %r1443, %r1441, %r1441, 25; sub.s32 %r1444, %r1442, %r1443; xor.b32 %r1445, %r1444, %r1438; shf.l.wrap.b32 %r1446, %r1444, %r1444, 16; sub.s32 %r1447, %r1445, %r1446; xor.b32 %r1448, %r1447, %r1441; shf.l.wrap.b32 %r1449, %r1447, %r1447, 4; sub.s32 %r1450, %r1448, %r1449; xor.b32 %r1451, %r1450, %r1444; shf.l.wrap.b32 %r1452, %r1450, %r1450, 14; sub.s32 %r1453, %r1451, %r1452; xor.b32 %r1454, %r1453, %r1447; shf.l.wrap.b32 %r1455, %r1453, %r1453, 24; sub.s32 %r1456, %r1454, %r1455; cvt.rn.f32.u32 %f1477, %r1456; mul.ftz.f32 %f1478, %f1477, 0f2F800000; add.s32 %r1457, %r1412, -559038708; add.s32 %r1458, %r1457, %r155; sub.s32 %r1459, %r1457, %r157; xor.b32 %r1460, %r1459, %r158; add.s32 %r1461, %r1458, %r157; sub.s32 %r1462, %r1458, %r1460; shf.l.wrap.b32 %r1463, %r1460, %r1460, 8; xor.b32 %r1464, %r1463, %r1462; add.s32 %r1465, %r1460, %r1461; sub.s32 %r1466, %r1461, %r1464; shf.l.wrap.b32 %r1467, %r1464, %r1464, 16; xor.b32 %r1468, %r1467, %r1466; add.s32 %r1469, %r1464, %r1465; sub.s32 %r1470, %r1465, %r1468; shf.l.wrap.b32 %r1471, %r1468, %r1468, 19; xor.b32 %r1472, %r1471, %r1470; add.s32 %r1473, %r1468, %r1469; sub.s32 %r1474, %r1469, %r1472; shf.l.wrap.b32 %r1475, %r1472, %r1472, 4; add.s32 %r1476, %r1472, %r1473; add.s32 %r1477, %r1473, %r149; xor.b32 %r1478, %r1476, %r1474; xor.b32 %r1479, %r1478, %r1475; shf.l.wrap.b32 %r1480, %r1476, %r1476, 14; sub.s32 %r1481, %r1479, %r1480; xor.b32 %r1482, %r1481, %r1477; shf.l.wrap.b32 %r1483, %r1481, %r1481, 11; sub.s32 %r1484, %r1482, %r1483; xor.b32 %r1485, %r1484, %r1476; shf.l.wrap.b32 %r1486, %r1484, %r1484, 25; sub.s32 %r1487, %r1485, %r1486; xor.b32 %r1488, %r1487, %r1481; shf.l.wrap.b32 %r1489, %r1487, %r1487, 16; sub.s32 %r1490, %r1488, %r1489; xor.b32 %r1491, %r1490, %r1484; shf.l.wrap.b32 %r1492, %r1490, %r1490, 4; sub.s32 %r1493, %r1491, %r1492; xor.b32 %r1494, %r1493, %r1487; shf.l.wrap.b32 %r1495, %r1493, %r1493, 14; sub.s32 %r1496, %r1494, %r1495; xor.b32 %r1497, %r1496, %r1490; shf.l.wrap.b32 %r1498, %r1496, %r1496, 24; sub.s32 %r1499, %r1497, %r1498; cvt.rn.f32.u32 %f1479, %r1499; mul.ftz.f32 %f1480, %f1479, 0f2F800000; sub.s32 %r1500, %r149, %r1412; shf.l.wrap.b32 %r1501, %r1457, %r1457, 4; xor.b32 %r1502, %r1501, %r1500; add.s32 %r1503, %r1457, %r146; sub.s32 %r1504, %r146, %r1502; shf.l.wrap.b32 %r1505, %r1502, %r1502, 6; xor.b32 %r1506, %r1505, %r1504; add.s32 %r1507, %r1502, %r1503; sub.s32 %r1508, %r1503, %r1506; shf.l.wrap.b32 %r1509, %r1506, %r1506, 8; xor.b32 %r1510, %r1509, %r1508; add.s32 %r1511, %r1506, %r1507; sub.s32 %r1512, %r1507, %r1510; shf.l.wrap.b32 %r1513, %r1510, %r1510, 16; xor.b32 %r1514, %r1513, %r1512; add.s32 %r1515, %r1510, %r1511; sub.s32 %r1516, %r1511, %r1514; shf.l.wrap.b32 %r1517, %r1514, %r1514, 19; xor.b32 %r1518, %r1517, %r1516; add.s32 %r1519, %r1514, %r1515; sub.s32 %r1520, %r1515, %r1518; shf.l.wrap.b32 %r1521, %r1518, %r1518, 4; add.s32 %r1522, %r1518, %r1519; add.s32 %r1523, %r1519, %r154; xor.b32 %r1524, %r1522, %r1520; xor.b32 %r1525, %r1524, %r1521; shf.l.wrap.b32 %r1526, %r1522, %r1522, 14; sub.s32 %r1527, %r1525, %r1526; xor.b32 %r1528, %r1527, %r1523; shf.l.wrap.b32 %r1529, %r1527, %r1527, 11; sub.s32 %r1530, %r1528, %r1529; xor.b32 %r1531, %r1530, %r1522; shf.l.wrap.b32 %r1532, %r1530, %r1530, 25; sub.s32 %r1533, %r1531, %r1532; xor.b32 %r1534, %r1533, %r1527; shf.l.wrap.b32 %r1535, %r1533, %r1533, 16; sub.s32 %r1536, %r1534, %r1535; xor.b32 %r1537, %r1536, %r1530; shf.l.wrap.b32 %r1538, %r1536, %r1536, 4; sub.s32 %r1539, %r1537, %r1538; xor.b32 %r1540, %r1539, %r1533; shf.l.wrap.b32 %r1541, %r1539, %r1539, 14; sub.s32 %r1542, %r1540, %r1541; xor.b32 %r1543, %r1542, %r1536; shf.l.wrap.b32 %r1544, %r1542, %r1542, 24; sub.s32 %r1545, %r1543, %r1544; cvt.rn.f32.u32 %f1481, %r1545; mul.ftz.f32 %f1482, %f1481, 0f2F800000; add.s32 %r1546, %r159, %r1412; xor.b32 %r1547, %r1546, %r161; sub.s32 %r1548, %r1547, %r162; xor.b32 %r1549, %r1548, %r160; shf.l.wrap.b32 %r1550, %r1548, %r1548, 25; sub.s32 %r1551, %r1549, %r1550; xor.b32 %r1552, %r1551, %r161; shf.l.wrap.b32 %r1553, %r1551, %r1551, 16; sub.s32 %r1554, %r1552, %r1553; xor.b32 %r1555, %r1554, %r1548; shf.l.wrap.b32 %r1556, %r1554, %r1554, 4; sub.s32 %r1557, %r1555, %r1556; xor.b32 %r1558, %r1557, %r1551; shf.l.wrap.b32 %r1559, %r1557, %r1557, 14; sub.s32 %r1560, %r1558, %r1559; xor.b32 %r1561, %r1560, %r1554; shf.l.wrap.b32 %r1562, %r1560, %r1560, 24; sub.s32 %r1563, %r1561, %r1562; cvt.rn.f32.u32 %f1483, %r1563; mul.ftz.f32 %f1484, %f1483, 0f2F800000; fma.rn.ftz.f32 %f335, %f19, %f1478, %f334; fma.rn.ftz.f32 %f336, %f19, %f1480, %f315; fma.rn.ftz.f32 %f337, %f19, %f1482, %f296; fma.rn.ftz.f32 %f338, %f19, %f1484, %f277; @%p48 bra $L__BB199_107; @%p49 bra $L__BB199_106; bra.uni $L__BB199_37; $L__BB199_106: sub.ftz.f32 %f1521, %f335, %f251; abs.ftz.f32 %f1522, %f1521; sub.ftz.f32 %f1523, %f336, %f253; abs.ftz.f32 %f1524, %f1523; add.ftz.f32 %f1525, %f1522, %f1524; sub.ftz.f32 %f1526, %f337, %f255; abs.ftz.f32 %f1527, %f1526; add.ftz.f32 %f1528, %f1525, %f1527; sub.ftz.f32 %f1529, %f338, %f257; abs.ftz.f32 %f1530, %f1529; add.ftz.f32 %f2975, %f1528, %f1530; bra.uni $L__BB199_108; $L__BB199_107: sub.ftz.f32 %f1531, %f335, %f251; mul.ftz.f32 %f1532, %f1531, %f1531; sub.ftz.f32 %f1533, %f336, %f253; fma.rn.ftz.f32 %f1534, %f1533, %f1533, %f1532; sub.ftz.f32 %f1535, %f337, %f255; mul.ftz.f32 %f1536, %f1535, %f1535; sub.ftz.f32 %f1537, %f338, %f257; fma.rn.ftz.f32 %f1538, %f1537, %f1537, %f1536; add.ftz.f32 %f1539, %f1534, %f1538; sqrt.approx.ftz.f32 %f2975, %f1539; bra.uni $L__BB199_108; $L__BB199_37: @%p50 bra $L__BB199_105; bra.uni $L__BB199_38; $L__BB199_105: sub.ftz.f32 %f1511, %f335, %f251; abs.ftz.f32 %f1512, %f1511; sub.ftz.f32 %f1513, %f336, %f253; abs.ftz.f32 %f1514, %f1513; sub.ftz.f32 %f1515, %f337, %f255; abs.ftz.f32 %f1516, %f1515; sub.ftz.f32 %f1517, %f338, %f257; abs.ftz.f32 %f1518, %f1517; max.ftz.f32 %f1519, %f1516, %f1518; max.ftz.f32 %f1520, %f1514, %f1519; max.ftz.f32 %f2975, %f1512, %f1520; bra.uni $L__BB199_108; $L__BB199_38: mov.f32 %f2975, 0f00000000; @%p51 bra $L__BB199_108; sub.ftz.f32 %f1486, %f335, %f251; abs.ftz.f32 %f1487, %f1486; lg2.approx.ftz.f32 %f1488, %f1487; mul.ftz.f32 %f1489, %f15, %f1488; ex2.approx.ftz.f32 %f1490, %f1489; sub.ftz.f32 %f1491, %f336, %f253; abs.ftz.f32 %f1492, %f1491; lg2.approx.ftz.f32 %f1493, %f1492; mul.ftz.f32 %f1494, %f15, %f1493; ex2.approx.ftz.f32 %f1495, %f1494; add.ftz.f32 %f1496, %f1490, %f1495; sub.ftz.f32 %f1497, %f337, %f255; abs.ftz.f32 %f1498, %f1497; lg2.approx.ftz.f32 %f1499, %f1498; mul.ftz.f32 %f1500, %f15, %f1499; ex2.approx.ftz.f32 %f1501, %f1500; add.ftz.f32 %f1502, %f1496, %f1501; sub.ftz.f32 %f1503, %f338, %f257; abs.ftz.f32 %f1504, %f1503; lg2.approx.ftz.f32 %f1505, %f1504; mul.ftz.f32 %f1506, %f15, %f1505; ex2.approx.ftz.f32 %f1507, %f1506; add.ftz.f32 %f1508, %f1502, %f1507; lg2.approx.ftz.f32 %f1509, %f1508; mul.ftz.f32 %f1510, %f258, %f1509; ex2.approx.ftz.f32 %f2975, %f1510; $L__BB199_108: setp.lt.ftz.f32 %p52, %f2975, %f325; @%p52 bra $L__BB199_111; bra.uni $L__BB199_109; $L__BB199_111: mov.f32 %f2976, %f2990; mov.f32 %f2977, %f2991; mov.f32 %f2978, %f2992; mov.f32 %f2979, %f2993; mov.f32 %f2980, %f2986; mov.f32 %f2981, %f2987; mov.f32 %f2982, %f2988; mov.f32 %f2983, %f2989; mov.f32 %f2984, %f2975; mov.f32 %f3386, %f325; mov.f32 %f2986, %f334; mov.f32 %f2987, %f315; mov.f32 %f2988, %f296; mov.f32 %f2989, %f277; mov.f32 %f2990, %f335; mov.f32 %f2991, %f336; mov.f32 %f2992, %f337; mov.f32 %f2993, %f338; bra.uni $L__BB199_112; $L__BB199_109: setp.geu.ftz.f32 %p53, %f2975, %f3386; mov.f32 %f2984, %f325; @%p53 bra $L__BB199_112; mov.f32 %f2976, %f335; mov.f32 %f2977, %f336; mov.f32 %f2978, %f337; mov.f32 %f2979, %f338; mov.f32 %f2980, %f334; mov.f32 %f2981, %f315; mov.f32 %f2982, %f296; mov.f32 %f2983, %f277; mov.f32 %f2984, %f325; mov.f32 %f3386, %f2975; $L__BB199_112: add.s32 %r4346, %r163, 1; setp.lt.s32 %p54, %r163, 1; @%p54 bra $L__BB199_35; add.s32 %r165, %r4345, 1; setp.lt.s32 %p55, %r4345, 1; mov.u32 %r4345, %r165; @%p55 bra $L__BB199_34; add.s32 %r166, %r4344, 1; setp.lt.s32 %p56, %r4344, 1; mov.u32 %r4344, %r166; @%p56 bra $L__BB199_33; add.s32 %r167, %r4343, 1; setp.lt.s32 %p57, %r4343, 1; mov.u32 %r4343, %r167; @%p57 bra $L__BB199_32; add.ftz.f32 %f1541, %f250, %f2980; mov.b32 %r1564, %f1541; add.ftz.f32 %f1542, %f252, %f2981; mov.b32 %r1565, %f1542; add.ftz.f32 %f1543, %f254, %f2982; mov.b32 %r1566, %f1543; add.ftz.f32 %f1544, %f256, %f2983; mov.b32 %r1567, %f1544; add.s32 %r1568, %r1565, -559038708; add.s32 %r1569, %r1566, -559038708; sub.s32 %r1570, %r1564, %r1566; shf.l.wrap.b32 %r1571, %r1569, %r1569, 4; xor.b32 %r1572, %r1571, %r1570; add.s32 %r1573, %r1569, %r1568; sub.s32 %r1574, %r1568, %r1572; shf.l.wrap.b32 %r1575, %r1572, %r1572, 6; xor.b32 %r1576, %r1575, %r1574; add.s32 %r1577, %r1572, %r1573; sub.s32 %r1578, %r1573, %r1576; shf.l.wrap.b32 %r1579, %r1576, %r1576, 8; xor.b32 %r1580, %r1579, %r1578; add.s32 %r1581, %r1576, %r1577; sub.s32 %r1582, %r1577, %r1580; shf.l.wrap.b32 %r1583, %r1580, %r1580, 16; xor.b32 %r1584, %r1583, %r1582; add.s32 %r1585, %r1580, %r1581; sub.s32 %r1586, %r1581, %r1584; shf.l.wrap.b32 %r1587, %r1584, %r1584, 19; xor.b32 %r1588, %r1587, %r1586; add.s32 %r1589, %r1584, %r1585; sub.s32 %r1590, %r1585, %r1588; shf.l.wrap.b32 %r1591, %r1588, %r1588, 4; add.s32 %r1592, %r1588, %r1589; add.s32 %r1593, %r1589, %r1567; xor.b32 %r1594, %r1592, %r1590; xor.b32 %r1595, %r1594, %r1591; shf.l.wrap.b32 %r1596, %r1592, %r1592, 14; sub.s32 %r1597, %r1595, %r1596; xor.b32 %r1598, %r1597, %r1593; shf.l.wrap.b32 %r1599, %r1597, %r1597, 11; sub.s32 %r1600, %r1598, %r1599; xor.b32 %r1601, %r1600, %r1592; shf.l.wrap.b32 %r1602, %r1600, %r1600, 25; sub.s32 %r1603, %r1601, %r1602; xor.b32 %r1604, %r1603, %r1597; shf.l.wrap.b32 %r1605, %r1603, %r1603, 16; sub.s32 %r1606, %r1604, %r1605; xor.b32 %r1607, %r1606, %r1600; shf.l.wrap.b32 %r1608, %r1606, %r1606, 4; sub.s32 %r1609, %r1607, %r1608; xor.b32 %r1610, %r1609, %r1603; shf.l.wrap.b32 %r1611, %r1609, %r1609, 14; sub.s32 %r1612, %r1610, %r1611; xor.b32 %r1613, %r1612, %r1606; shf.l.wrap.b32 %r1614, %r1612, %r1612, 24; sub.s32 %r1615, %r1613, %r1614; cvt.rn.f32.u32 %f1545, %r1615; mul.ftz.f32 %f3391, %f1545, 0f2F800000; add.s32 %r1616, %r1564, -559038708; add.s32 %r1617, %r1567, -559038708; sub.s32 %r1618, %r1566, %r1567; shf.l.wrap.b32 %r1619, %r1617, %r1617, 4; xor.b32 %r1620, %r1619, %r1618; add.s32 %r1621, %r1617, %r1616; sub.s32 %r1622, %r1616, %r1620; shf.l.wrap.b32 %r1623, %r1620, %r1620, 6; xor.b32 %r1624, %r1623, %r1622; add.s32 %r1625, %r1620, %r1621; sub.s32 %r1626, %r1621, %r1624; shf.l.wrap.b32 %r1627, %r1624, %r1624, 8; xor.b32 %r1628, %r1627, %r1626; add.s32 %r1629, %r1624, %r1625; sub.s32 %r1630, %r1625, %r1628; shf.l.wrap.b32 %r1631, %r1628, %r1628, 16; xor.b32 %r1632, %r1631, %r1630; add.s32 %r1633, %r1628, %r1629; sub.s32 %r1634, %r1629, %r1632; shf.l.wrap.b32 %r1635, %r1632, %r1632, 19; xor.b32 %r1636, %r1635, %r1634; add.s32 %r1637, %r1632, %r1633; sub.s32 %r1638, %r1633, %r1636; shf.l.wrap.b32 %r1639, %r1636, %r1636, 4; add.s32 %r1640, %r1636, %r1637; add.s32 %r1641, %r1637, %r1565; xor.b32 %r1642, %r1640, %r1638; xor.b32 %r1643, %r1642, %r1639; shf.l.wrap.b32 %r1644, %r1640, %r1640, 14; sub.s32 %r1645, %r1643, %r1644; xor.b32 %r1646, %r1645, %r1641; shf.l.wrap.b32 %r1647, %r1645, %r1645, 11; sub.s32 %r1648, %r1646, %r1647; xor.b32 %r1649, %r1648, %r1640; shf.l.wrap.b32 %r1650, %r1648, %r1648, 25; sub.s32 %r1651, %r1649, %r1650; xor.b32 %r1652, %r1651, %r1645; shf.l.wrap.b32 %r1653, %r1651, %r1651, 16; sub.s32 %r1654, %r1652, %r1653; xor.b32 %r1655, %r1654, %r1648; shf.l.wrap.b32 %r1656, %r1654, %r1654, 4; sub.s32 %r1657, %r1655, %r1656; xor.b32 %r1658, %r1657, %r1651; shf.l.wrap.b32 %r1659, %r1657, %r1657, 14; sub.s32 %r1660, %r1658, %r1659; xor.b32 %r1661, %r1660, %r1654; shf.l.wrap.b32 %r1662, %r1660, %r1660, 24; sub.s32 %r1663, %r1661, %r1662; cvt.rn.f32.u32 %f1546, %r1663; mul.ftz.f32 %f3390, %f1546, 0f2F800000; sub.s32 %r1664, %r1567, %r1565; shf.l.wrap.b32 %r1665, %r1568, %r1568, 4; xor.b32 %r1666, %r1665, %r1664; sub.s32 %r1667, %r1569, %r1666; shf.l.wrap.b32 %r1668, %r1666, %r1666, 6; xor.b32 %r1669, %r1668, %r1667; add.s32 %r1670, %r1666, %r1573; sub.s32 %r1671, %r1573, %r1669; shf.l.wrap.b32 %r1672, %r1669, %r1669, 8; xor.b32 %r1673, %r1672, %r1671; add.s32 %r1674, %r1669, %r1670; sub.s32 %r1675, %r1670, %r1673; shf.l.wrap.b32 %r1676, %r1673, %r1673, 16; xor.b32 %r1677, %r1676, %r1675; add.s32 %r1678, %r1673, %r1674; sub.s32 %r1679, %r1674, %r1677; shf.l.wrap.b32 %r1680, %r1677, %r1677, 19; xor.b32 %r1681, %r1680, %r1679; add.s32 %r1682, %r1677, %r1678; sub.s32 %r1683, %r1678, %r1681; shf.l.wrap.b32 %r1684, %r1681, %r1681, 4; add.s32 %r1685, %r1681, %r1682; add.s32 %r1686, %r1682, %r1564; xor.b32 %r1687, %r1685, %r1683; xor.b32 %r1688, %r1687, %r1684; shf.l.wrap.b32 %r1689, %r1685, %r1685, 14; sub.s32 %r1690, %r1688, %r1689; xor.b32 %r1691, %r1690, %r1686; shf.l.wrap.b32 %r1692, %r1690, %r1690, 11; sub.s32 %r1693, %r1691, %r1692; xor.b32 %r1694, %r1693, %r1685; shf.l.wrap.b32 %r1695, %r1693, %r1693, 25; sub.s32 %r1696, %r1694, %r1695; xor.b32 %r1697, %r1696, %r1690; shf.l.wrap.b32 %r1698, %r1696, %r1696, 16; sub.s32 %r1699, %r1697, %r1698; xor.b32 %r1700, %r1699, %r1693; shf.l.wrap.b32 %r1701, %r1699, %r1699, 4; sub.s32 %r1702, %r1700, %r1701; xor.b32 %r1703, %r1702, %r1696; shf.l.wrap.b32 %r1704, %r1702, %r1702, 14; sub.s32 %r1705, %r1703, %r1704; xor.b32 %r1706, %r1705, %r1699; shf.l.wrap.b32 %r1707, %r1705, %r1705, 24; sub.s32 %r1708, %r1706, %r1707; cvt.rn.f32.u32 %f1547, %r1708; mul.ftz.f32 %f3389, %f1547, 0f2F800000; add.ftz.f32 %f3048, %f250, %f2976; add.ftz.f32 %f3047, %f252, %f2977; add.ftz.f32 %f3046, %f254, %f2978; add.ftz.f32 %f3045, %f256, %f2979; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_136; $L__BB199_366: setp.eq.s32 %p198, %r389, 0; @%p198 bra $L__BB199_382; setp.eq.s32 %p199, %r389, 1; mov.f32 %f3386, %f3392; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; @%p199 bra $L__BB199_368; bra.uni $L__BB199_383; $L__BB199_368: cvt.rmi.ftz.f32.f32 %f1132, %f21; sub.ftz.f32 %f1133, %f21, %f1132; add.ftz.f32 %f2696, %f1132, 0fBF800000; mov.b32 %r4071, %f2696; add.s32 %r4072, %r4071, -559038720; xor.b32 %r4073, %r4072, -1874868139; add.s32 %r4074, %r4073, 29184895; xor.b32 %r4075, %r4074, -559038720; shf.l.wrap.b32 %r4076, %r4074, %r4074, 25; sub.s32 %r4077, %r4075, %r4076; xor.b32 %r4078, %r4077, -1874868139; shf.l.wrap.b32 %r4079, %r4077, %r4077, 16; sub.s32 %r4080, %r4078, %r4079; xor.b32 %r4081, %r4080, %r4074; shf.l.wrap.b32 %r4082, %r4080, %r4080, 4; sub.s32 %r4083, %r4081, %r4082; xor.b32 %r4084, %r4083, %r4077; shf.l.wrap.b32 %r4085, %r4083, %r4083, 14; sub.s32 %r4086, %r4084, %r4085; xor.b32 %r4087, %r4086, %r4080; shf.l.wrap.b32 %r4088, %r4086, %r4086, 24; sub.s32 %r4089, %r4087, %r4088; cvt.rn.f32.u32 %f2697, %r4089; mul.ftz.f32 %f2698, %f2697, 0f2F800000; fma.rn.ftz.f32 %f2699, %f19, %f2698, 0fBF800000; sub.ftz.f32 %f2700, %f1133, %f2699; abs.ftz.f32 %f2701, %f2700; setp.lt.ftz.f32 %p210, %f2701, 0f41000000; min.ftz.f32 %f3375, %f2701, 0f41000000; selp.f32 %f3378, 0fBF800000, 0f00000000, %p210; mov.f32 %f3388, 0f00000000; selp.f32 %f3379, %f2699, 0f00000000, %p210; add.ftz.f32 %f2702, %f1132, 0f00000000; mov.b32 %r4090, %f2702; add.s32 %r4091, %r4090, -559038720; xor.b32 %r4092, %r4091, -1874868139; add.s32 %r4093, %r4092, 29184895; xor.b32 %r4094, %r4093, -559038720; shf.l.wrap.b32 %r4095, %r4093, %r4093, 25; sub.s32 %r4096, %r4094, %r4095; xor.b32 %r4097, %r4096, -1874868139; shf.l.wrap.b32 %r4098, %r4096, %r4096, 16; sub.s32 %r4099, %r4097, %r4098; xor.b32 %r4100, %r4099, %r4093; shf.l.wrap.b32 %r4101, %r4099, %r4099, 4; sub.s32 %r4102, %r4100, %r4101; xor.b32 %r4103, %r4102, %r4096; shf.l.wrap.b32 %r4104, %r4102, %r4102, 14; sub.s32 %r4105, %r4103, %r4104; xor.b32 %r4106, %r4105, %r4099; shf.l.wrap.b32 %r4107, %r4105, %r4105, 24; sub.s32 %r4108, %r4106, %r4107; cvt.rn.f32.u32 %f2703, %r4108; mul.ftz.f32 %f2704, %f2703, 0f2F800000; fma.rn.ftz.f32 %f3377, %f19, %f2704, 0f00000000; sub.ftz.f32 %f2705, %f1133, %f3377; abs.ftz.f32 %f3374, %f2705; setp.lt.ftz.f32 %p211, %f3374, %f3375; mov.f32 %f3376, %f3388; @%p211 bra $L__BB199_370; setp.lt.ftz.f32 %p212, %f3374, 0f41000000; min.ftz.f32 %f1139, %f3374, 0f41000000; selp.f32 %f1140, %f3377, 0f00000000, %p212; mov.f32 %f2706, 0f00000000; mov.f32 %f3374, %f3375; mov.f32 %f3375, %f1139; mov.f32 %f3376, %f3378; mov.f32 %f3377, %f3379; mov.f32 %f3378, %f2706; mov.f32 %f3379, %f1140; $L__BB199_370: add.ftz.f32 %f2708, %f1132, 0f3F800000; mov.b32 %r4109, %f2708; add.s32 %r4110, %r4109, -559038720; xor.b32 %r4111, %r4110, -1874868139; add.s32 %r4112, %r4111, 29184895; xor.b32 %r4113, %r4112, -559038720; shf.l.wrap.b32 %r4114, %r4112, %r4112, 25; sub.s32 %r4115, %r4113, %r4114; xor.b32 %r4116, %r4115, -1874868139; shf.l.wrap.b32 %r4117, %r4115, %r4115, 16; sub.s32 %r4118, %r4116, %r4117; xor.b32 %r4119, %r4118, %r4112; shf.l.wrap.b32 %r4120, %r4118, %r4118, 4; sub.s32 %r4121, %r4119, %r4120; xor.b32 %r4122, %r4121, %r4115; shf.l.wrap.b32 %r4123, %r4121, %r4121, 14; sub.s32 %r4124, %r4122, %r4123; xor.b32 %r4125, %r4124, %r4118; shf.l.wrap.b32 %r4126, %r4124, %r4124, 24; sub.s32 %r4127, %r4125, %r4126; cvt.rn.f32.u32 %f2709, %r4127; mul.ftz.f32 %f2710, %f2709, 0f2F800000; fma.rn.ftz.f32 %f2711, %f19, %f2710, 0f3F800000; sub.ftz.f32 %f2712, %f1133, %f2711; abs.ftz.f32 %f2713, %f2712; setp.lt.ftz.f32 %p213, %f2713, %f3374; setp.lt.ftz.f32 %p214, %f2713, %f3375; selp.f32 %f2714, %f2713, %f3375, %p214; selp.f32 %f2715, 0f3F800000, %f3378, %p214; selp.f32 %f2716, %f2711, %f3379, %p214; selp.f32 %f3386, %f3374, %f2714, %p213; selp.f32 %f2717, %f3376, %f2715, %p213; selp.f32 %f2718, %f3377, %f2716, %p213; add.ftz.f32 %f2719, %f1132, %f2717; mov.b32 %r4128, %f2719; add.s32 %r4129, %r4128, -559038720; xor.b32 %r4130, %r4129, -1874868139; add.s32 %r4131, %r4130, 29184895; xor.b32 %r4132, %r4131, -559038720; shf.l.wrap.b32 %r4133, %r4131, %r4131, 25; sub.s32 %r4134, %r4132, %r4133; xor.b32 %r4135, %r4134, -1874868139; shf.l.wrap.b32 %r4136, %r4134, %r4134, 16; sub.s32 %r4137, %r4135, %r4136; xor.b32 %r4138, %r4137, %r4131; shf.l.wrap.b32 %r4139, %r4137, %r4137, 4; sub.s32 %r4140, %r4138, %r4139; xor.b32 %r4141, %r4140, %r4134; shf.l.wrap.b32 %r4142, %r4140, %r4140, 14; sub.s32 %r4143, %r4141, %r4142; xor.b32 %r4144, %r4143, %r4137; shf.l.wrap.b32 %r4145, %r4143, %r4143, 24; sub.s32 %r4146, %r4144, %r4145; cvt.rn.f32.u32 %f2720, %r4146; mul.ftz.f32 %f3391, %f2720, 0f2F800000; add.s32 %r4147, %r4128, -559038716; xor.b32 %r4148, %r4147, 1354692725; add.s32 %r4149, %r4148, 138171771; xor.b32 %r4150, %r4149, 506314500; shf.l.wrap.b32 %r4151, %r4149, %r4149, 25; sub.s32 %r4152, %r4150, %r4151; xor.b32 %r4153, %r4152, 1354692725; shf.l.wrap.b32 %r4154, %r4152, %r4152, 16; sub.s32 %r4155, %r4153, %r4154; xor.b32 %r4156, %r4155, %r4149; shf.l.wrap.b32 %r4157, %r4155, %r4155, 4; sub.s32 %r4158, %r4156, %r4157; xor.b32 %r4159, %r4158, %r4152; shf.l.wrap.b32 %r4160, %r4158, %r4158, 14; sub.s32 %r4161, %r4159, %r4160; xor.b32 %r4162, %r4161, %r4155; shf.l.wrap.b32 %r4163, %r4161, %r4161, 24; sub.s32 %r4164, %r4162, %r4163; cvt.rn.f32.u32 %f2721, %r4164; mul.ftz.f32 %f3390, %f2721, 0f2F800000; xor.b32 %r4165, %r4147, 1346304085; add.s32 %r4166, %r4165, 138237311; xor.b32 %r4167, %r4166, 514703108; shf.l.wrap.b32 %r4168, %r4166, %r4166, 25; sub.s32 %r4169, %r4167, %r4168; xor.b32 %r4170, %r4169, 1346304085; shf.l.wrap.b32 %r4171, %r4169, %r4169, 16; sub.s32 %r4172, %r4170, %r4171; xor.b32 %r4173, %r4172, %r4166; shf.l.wrap.b32 %r4174, %r4172, %r4172, 4; sub.s32 %r4175, %r4173, %r4174; xor.b32 %r4176, %r4175, %r4169; shf.l.wrap.b32 %r4177, %r4175, %r4175, 14; sub.s32 %r4178, %r4176, %r4177; xor.b32 %r4179, %r4178, %r4172; shf.l.wrap.b32 %r4180, %r4178, %r4178, 24; sub.s32 %r4181, %r4179, %r4180; cvt.rn.f32.u32 %f2722, %r4181; mul.ftz.f32 %f3389, %f2722, 0f2F800000; add.ftz.f32 %f3381, %f1132, %f2718; bra.uni $L__BB199_383; $L__BB199_139: setp.eq.s32 %p72, %r389, 0; @%p72 bra $L__BB199_215; setp.eq.s32 %p73, %r389, 1; mov.f32 %f3386, %f3387; mov.f32 %f3388, %f3387; mov.f32 %f3389, %f3387; mov.f32 %f3390, %f3387; mov.f32 %f3391, %f3387; mov.f32 %f3240, %f3387; mov.f32 %f3241, %f3387; mov.f32 %f3242, %f3387; @%p73 bra $L__BB199_141; bra.uni $L__BB199_250; $L__BB199_141: ld.param.u32 %r4316, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f641, %f22; sub.ftz.f32 %f642, %f22, %f641; cvt.rmi.ftz.f32.f32 %f643, %f23; sub.ftz.f32 %f644, %f23, %f643; cvt.rmi.ftz.f32.f32 %f645, %f24; sub.ftz.f32 %f646, %f24, %f645; rcp.approx.ftz.f32 %f647, %f15; mov.f32 %f3194, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2470, -1; setp.eq.s32 %p100, %r4316, 0; setp.eq.s32 %p101, %r4316, 1; setp.eq.s32 %p102, %r4316, 2; setp.ne.s32 %p103, %r4316, 3; mov.u32 %r4366, %r2470; mov.f32 %f3195, %f3194; mov.f32 %f3196, %f3194; mov.f32 %f3191, %f3194; mov.f32 %f3192, %f3194; mov.f32 %f3193, %f3194; mov.f32 %f3189, %f3386; mov.f32 %f3188, %f3194; mov.f32 %f3187, %f3194; mov.f32 %f3186, %f3194; mov.f32 %f3185, %f3194; mov.f32 %f3184, %f3194; mov.f32 %f3183, %f3194; $L__BB199_142: cvt.rn.f32.s32 %f662, %r4366; add.ftz.f32 %f1861, %f645, %f662; mov.b32 %r269, %f1861; add.s32 %r270, %r269, -559038712; add.s32 %r271, %r269, -559038708; shf.l.wrap.b32 %r272, %r271, %r271, 4; mov.u32 %r4367, %r2470; $L__BB199_143: cvt.rn.f32.s32 %f677, %r4367; add.ftz.f32 %f1862, %f643, %f677; mov.b32 %r2473, %f1862; add.s32 %r274, %r2473, -559038712; xor.b32 %r2474, %r274, %r270; shf.l.wrap.b32 %r2475, %r274, %r274, 14; sub.s32 %r275, %r2474, %r2475; shf.l.wrap.b32 %r276, %r275, %r275, 11; add.s32 %r277, %r2473, -559038708; add.s32 %r278, %r277, %r271; mov.u32 %r4368, %r2470; $L__BB199_144: mov.f32 %f685, %f3189; mov.u32 %r279, %r4368; cvt.rn.f32.s32 %f692, %r279; add.ftz.f32 %f1863, %f641, %f692; mov.b32 %r2476, %f1863; add.s32 %r2477, %r2476, -559038712; xor.b32 %r2478, %r2477, %r275; sub.s32 %r2479, %r2478, %r276; xor.b32 %r2480, %r2479, %r274; shf.l.wrap.b32 %r2481, %r2479, %r2479, 25; sub.s32 %r2482, %r2480, %r2481; xor.b32 %r2483, %r2482, %r275; shf.l.wrap.b32 %r2484, %r2482, %r2482, 16; sub.s32 %r2485, %r2483, %r2484; xor.b32 %r2486, %r2485, %r2479; shf.l.wrap.b32 %r2487, %r2485, %r2485, 4; sub.s32 %r2488, %r2486, %r2487; xor.b32 %r2489, %r2488, %r2482; shf.l.wrap.b32 %r2490, %r2488, %r2488, 14; sub.s32 %r2491, %r2489, %r2490; xor.b32 %r2492, %r2491, %r2485; shf.l.wrap.b32 %r2493, %r2491, %r2491, 24; sub.s32 %r2494, %r2492, %r2493; cvt.rn.f32.u32 %f1864, %r2494; mul.ftz.f32 %f1865, %f1864, 0f2F800000; sub.s32 %r2495, %r2476, %r269; xor.b32 %r2496, %r2495, %r272; sub.s32 %r2497, %r277, %r2496; shf.l.wrap.b32 %r2498, %r2496, %r2496, 6; xor.b32 %r2499, %r2498, %r2497; add.s32 %r2500, %r2496, %r278; sub.s32 %r2501, %r278, %r2499; shf.l.wrap.b32 %r2502, %r2499, %r2499, 8; xor.b32 %r2503, %r2502, %r2501; add.s32 %r2504, %r2499, %r2500; sub.s32 %r2505, %r2500, %r2503; shf.l.wrap.b32 %r2506, %r2503, %r2503, 16; xor.b32 %r2507, %r2506, %r2505; add.s32 %r2508, %r2503, %r2504; sub.s32 %r2509, %r2504, %r2507; shf.l.wrap.b32 %r2510, %r2507, %r2507, 19; xor.b32 %r2511, %r2510, %r2509; add.s32 %r2512, %r2507, %r2508; sub.s32 %r2513, %r2508, %r2511; shf.l.wrap.b32 %r2514, %r2511, %r2511, 4; add.s32 %r2515, %r2511, %r2512; add.s32 %r2516, %r2512, 1065353216; xor.b32 %r2517, %r2515, %r2513; xor.b32 %r2518, %r2517, %r2514; shf.l.wrap.b32 %r2519, %r2515, %r2515, 14; sub.s32 %r2520, %r2518, %r2519; xor.b32 %r2521, %r2520, %r2516; shf.l.wrap.b32 %r2522, %r2520, %r2520, 11; sub.s32 %r2523, %r2521, %r2522; xor.b32 %r2524, %r2523, %r2515; shf.l.wrap.b32 %r2525, %r2523, %r2523, 25; sub.s32 %r2526, %r2524, %r2525; xor.b32 %r2527, %r2526, %r2520; shf.l.wrap.b32 %r2528, %r2526, %r2526, 16; sub.s32 %r2529, %r2527, %r2528; xor.b32 %r2530, %r2529, %r2523; shf.l.wrap.b32 %r2531, %r2529, %r2529, 4; sub.s32 %r2532, %r2530, %r2531; xor.b32 %r2533, %r2532, %r2526; shf.l.wrap.b32 %r2534, %r2532, %r2532, 14; sub.s32 %r2535, %r2533, %r2534; xor.b32 %r2536, %r2535, %r2529; shf.l.wrap.b32 %r2537, %r2535, %r2535, 24; sub.s32 %r2538, %r2536, %r2537; cvt.rn.f32.u32 %f1866, %r2538; mul.ftz.f32 %f1867, %f1866, 0f2F800000; add.s32 %r2539, %r2512, 1073741824; xor.b32 %r2540, %r2520, %r2539; sub.s32 %r2541, %r2540, %r2522; xor.b32 %r2542, %r2541, %r2515; shf.l.wrap.b32 %r2543, %r2541, %r2541, 25; sub.s32 %r2544, %r2542, %r2543; xor.b32 %r2545, %r2544, %r2520; shf.l.wrap.b32 %r2546, %r2544, %r2544, 16; sub.s32 %r2547, %r2545, %r2546; xor.b32 %r2548, %r2547, %r2541; shf.l.wrap.b32 %r2549, %r2547, %r2547, 4; sub.s32 %r2550, %r2548, %r2549; xor.b32 %r2551, %r2550, %r2544; shf.l.wrap.b32 %r2552, %r2550, %r2550, 14; sub.s32 %r2553, %r2551, %r2552; xor.b32 %r2554, %r2553, %r2547; shf.l.wrap.b32 %r2555, %r2553, %r2553, 24; sub.s32 %r2556, %r2554, %r2555; cvt.rn.f32.u32 %f1868, %r2556; mul.ftz.f32 %f1869, %f1868, 0f2F800000; fma.rn.ftz.f32 %f693, %f19, %f1865, %f692; fma.rn.ftz.f32 %f694, %f19, %f1867, %f677; fma.rn.ftz.f32 %f695, %f19, %f1869, %f662; @%p100 bra $L__BB199_206; @%p101 bra $L__BB199_205; bra.uni $L__BB199_146; $L__BB199_205: sub.ftz.f32 %f1897, %f693, %f642; abs.ftz.f32 %f1898, %f1897; sub.ftz.f32 %f1899, %f694, %f644; abs.ftz.f32 %f1900, %f1899; add.ftz.f32 %f1901, %f1898, %f1900; sub.ftz.f32 %f1902, %f695, %f646; abs.ftz.f32 %f1903, %f1902; add.ftz.f32 %f3182, %f1901, %f1903; bra.uni $L__BB199_207; $L__BB199_206: sub.ftz.f32 %f1904, %f693, %f642; sub.ftz.f32 %f1905, %f694, %f644; mul.ftz.f32 %f1906, %f1905, %f1905; fma.rn.ftz.f32 %f1907, %f1904, %f1904, %f1906; sub.ftz.f32 %f1908, %f695, %f646; fma.rn.ftz.f32 %f1909, %f1908, %f1908, %f1907; sqrt.approx.ftz.f32 %f3182, %f1909; bra.uni $L__BB199_207; $L__BB199_146: @%p102 bra $L__BB199_204; bra.uni $L__BB199_147; $L__BB199_204: sub.ftz.f32 %f1890, %f693, %f642; abs.ftz.f32 %f1891, %f1890; sub.ftz.f32 %f1892, %f694, %f644; abs.ftz.f32 %f1893, %f1892; sub.ftz.f32 %f1894, %f695, %f646; abs.ftz.f32 %f1895, %f1894; max.ftz.f32 %f1896, %f1893, %f1895; max.ftz.f32 %f3182, %f1891, %f1896; bra.uni $L__BB199_207; $L__BB199_147: mov.f32 %f3182, 0f00000000; @%p103 bra $L__BB199_207; sub.ftz.f32 %f1871, %f693, %f642; abs.ftz.f32 %f1872, %f1871; lg2.approx.ftz.f32 %f1873, %f1872; mul.ftz.f32 %f1874, %f15, %f1873; ex2.approx.ftz.f32 %f1875, %f1874; sub.ftz.f32 %f1876, %f694, %f644; abs.ftz.f32 %f1877, %f1876; lg2.approx.ftz.f32 %f1878, %f1877; mul.ftz.f32 %f1879, %f15, %f1878; ex2.approx.ftz.f32 %f1880, %f1879; add.ftz.f32 %f1881, %f1875, %f1880; sub.ftz.f32 %f1882, %f695, %f646; abs.ftz.f32 %f1883, %f1882; lg2.approx.ftz.f32 %f1884, %f1883; mul.ftz.f32 %f1885, %f15, %f1884; ex2.approx.ftz.f32 %f1886, %f1885; add.ftz.f32 %f1887, %f1881, %f1886; lg2.approx.ftz.f32 %f1888, %f1887; mul.ftz.f32 %f1889, %f647, %f1888; ex2.approx.ftz.f32 %f3182, %f1889; $L__BB199_207: setp.lt.ftz.f32 %p104, %f3182, %f685; @%p104 bra $L__BB199_210; bra.uni $L__BB199_208; $L__BB199_210: mov.f32 %f3183, %f3194; mov.f32 %f3184, %f3195; mov.f32 %f3185, %f3196; mov.f32 %f3186, %f3191; mov.f32 %f3187, %f3192; mov.f32 %f3188, %f3193; mov.f32 %f3189, %f3182; mov.f32 %f3386, %f685; mov.f32 %f3191, %f692; mov.f32 %f3192, %f677; mov.f32 %f3193, %f662; mov.f32 %f3194, %f693; mov.f32 %f3195, %f694; mov.f32 %f3196, %f695; bra.uni $L__BB199_211; $L__BB199_208: setp.geu.ftz.f32 %p105, %f3182, %f3386; mov.f32 %f3189, %f685; @%p105 bra $L__BB199_211; mov.f32 %f3183, %f693; mov.f32 %f3184, %f694; mov.f32 %f3185, %f695; mov.f32 %f3186, %f692; mov.f32 %f3187, %f677; mov.f32 %f3188, %f662; mov.f32 %f3189, %f685; mov.f32 %f3386, %f3182; $L__BB199_211: add.s32 %r4368, %r279, 1; setp.lt.s32 %p106, %r279, 1; @%p106 bra $L__BB199_144; add.s32 %r281, %r4367, 1; setp.lt.s32 %p107, %r4367, 1; mov.u32 %r4367, %r281; @%p107 bra $L__BB199_143; add.s32 %r282, %r4366, 1; setp.lt.s32 %p108, %r4366, 1; mov.u32 %r4366, %r282; @%p108 bra $L__BB199_142; add.ftz.f32 %f1911, %f641, %f3186; mov.b32 %r2557, %f1911; add.ftz.f32 %f1912, %f643, %f3187; mov.b32 %r2558, %f1912; add.ftz.f32 %f1913, %f645, %f3188; mov.b32 %r2559, %f1913; add.s32 %r2560, %r2559, -559038712; add.s32 %r2561, %r2558, -559038712; add.s32 %r2562, %r2557, -559038712; xor.b32 %r2563, %r2560, %r2561; shf.l.wrap.b32 %r2564, %r2561, %r2561, 14; sub.s32 %r2565, %r2563, %r2564; xor.b32 %r2566, %r2565, %r2562; shf.l.wrap.b32 %r2567, %r2565, %r2565, 11; sub.s32 %r2568, %r2566, %r2567; xor.b32 %r2569, %r2568, %r2561; shf.l.wrap.b32 %r2570, %r2568, %r2568, 25; sub.s32 %r2571, %r2569, %r2570; xor.b32 %r2572, %r2571, %r2565; shf.l.wrap.b32 %r2573, %r2571, %r2571, 16; sub.s32 %r2574, %r2572, %r2573; xor.b32 %r2575, %r2574, %r2568; shf.l.wrap.b32 %r2576, %r2574, %r2574, 4; sub.s32 %r2577, %r2575, %r2576; xor.b32 %r2578, %r2577, %r2571; shf.l.wrap.b32 %r2579, %r2577, %r2577, 14; sub.s32 %r2580, %r2578, %r2579; xor.b32 %r2581, %r2580, %r2574; shf.l.wrap.b32 %r2582, %r2580, %r2580, 24; sub.s32 %r2583, %r2581, %r2582; cvt.rn.f32.u32 %f1914, %r2583; mul.ftz.f32 %f3391, %f1914, 0f2F800000; add.s32 %r2584, %r2558, -559038708; add.s32 %r2585, %r2559, -559038708; sub.s32 %r2586, %r2557, %r2559; shf.l.wrap.b32 %r2587, %r2585, %r2585, 4; xor.b32 %r2588, %r2587, %r2586; add.s32 %r2589, %r2585, %r2584; sub.s32 %r2590, %r2584, %r2588; shf.l.wrap.b32 %r2591, %r2588, %r2588, 6; xor.b32 %r2592, %r2591, %r2590; add.s32 %r2593, %r2588, %r2589; sub.s32 %r2594, %r2589, %r2592; shf.l.wrap.b32 %r2595, %r2592, %r2592, 8; xor.b32 %r2596, %r2595, %r2594; add.s32 %r2597, %r2592, %r2593; sub.s32 %r2598, %r2593, %r2596; shf.l.wrap.b32 %r2599, %r2596, %r2596, 16; xor.b32 %r2600, %r2599, %r2598; add.s32 %r2601, %r2596, %r2597; sub.s32 %r2602, %r2597, %r2600; shf.l.wrap.b32 %r2603, %r2600, %r2600, 19; xor.b32 %r2604, %r2603, %r2602; add.s32 %r2605, %r2600, %r2601; sub.s32 %r2606, %r2601, %r2604; shf.l.wrap.b32 %r2607, %r2604, %r2604, 4; add.s32 %r2608, %r2604, %r2605; add.s32 %r2609, %r2605, 1065353216; xor.b32 %r2610, %r2608, %r2606; xor.b32 %r2611, %r2610, %r2607; shf.l.wrap.b32 %r2612, %r2608, %r2608, 14; sub.s32 %r2613, %r2611, %r2612; xor.b32 %r2614, %r2613, %r2609; shf.l.wrap.b32 %r2615, %r2613, %r2613, 11; sub.s32 %r2616, %r2614, %r2615; xor.b32 %r2617, %r2616, %r2608; shf.l.wrap.b32 %r2618, %r2616, %r2616, 25; sub.s32 %r2619, %r2617, %r2618; xor.b32 %r2620, %r2619, %r2613; shf.l.wrap.b32 %r2621, %r2619, %r2619, 16; sub.s32 %r2622, %r2620, %r2621; xor.b32 %r2623, %r2622, %r2616; shf.l.wrap.b32 %r2624, %r2622, %r2622, 4; sub.s32 %r2625, %r2623, %r2624; xor.b32 %r2626, %r2625, %r2619; shf.l.wrap.b32 %r2627, %r2625, %r2625, 14; sub.s32 %r2628, %r2626, %r2627; xor.b32 %r2629, %r2628, %r2622; shf.l.wrap.b32 %r2630, %r2628, %r2628, 24; sub.s32 %r2631, %r2629, %r2630; cvt.rn.f32.u32 %f1915, %r2631; mul.ftz.f32 %f3390, %f1915, 0f2F800000; add.s32 %r2632, %r2605, 1073741824; xor.b32 %r2633, %r2613, %r2632; sub.s32 %r2634, %r2633, %r2615; xor.b32 %r2635, %r2634, %r2608; shf.l.wrap.b32 %r2636, %r2634, %r2634, 25; sub.s32 %r2637, %r2635, %r2636; xor.b32 %r2638, %r2637, %r2613; shf.l.wrap.b32 %r2639, %r2637, %r2637, 16; sub.s32 %r2640, %r2638, %r2639; xor.b32 %r2641, %r2640, %r2634; shf.l.wrap.b32 %r2642, %r2640, %r2640, 4; sub.s32 %r2643, %r2641, %r2642; xor.b32 %r2644, %r2643, %r2637; shf.l.wrap.b32 %r2645, %r2643, %r2643, 14; sub.s32 %r2646, %r2644, %r2645; xor.b32 %r2647, %r2646, %r2640; shf.l.wrap.b32 %r2648, %r2646, %r2646, 24; sub.s32 %r2649, %r2647, %r2648; cvt.rn.f32.u32 %f1916, %r2649; mul.ftz.f32 %f3389, %f1916, 0f2F800000; add.ftz.f32 %f3242, %f641, %f3183; add.ftz.f32 %f3241, %f643, %f3184; add.ftz.f32 %f3240, %f645, %f3185; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_250; $L__BB199_286: ld.param.u32 %r4320, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f910, %f22; sub.ftz.f32 %f911, %f22, %f910; cvt.rmi.ftz.f32.f32 %f912, %f23; sub.ftz.f32 %f913, %f23, %f912; fma.rn.ftz.f32 %f914, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f915, %f15; mov.f32 %f3286, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4375, -2; mov.u32 %r3441, -3; setp.eq.s32 %p150, %r4320, 0; setp.eq.s32 %p151, %r4320, 1; setp.eq.s32 %p152, %r4320, 2; setp.ne.s32 %p153, %r4320, 3; mov.f32 %f3287, %f3286; mov.f32 %f3391, %f3286; mov.f32 %f3390, %f3286; mov.f32 %f3389, %f3286; $L__BB199_287: cvt.rn.f32.s32 %f922, %r4375; add.ftz.f32 %f2281, %f912, %f922; mov.b32 %r3442, %f2281; add.s32 %r347, %r3442, -559038716; xor.b32 %r3443, %r347, -559038716; shf.l.wrap.b32 %r3444, %r347, %r347, 14; sub.s32 %r348, %r3443, %r3444; shf.l.wrap.b32 %r349, %r348, %r348, 11; add.s32 %r350, %r3442, -559038712; xor.b32 %r3445, %r350, 506314504; shf.l.wrap.b32 %r3446, %r350, %r350, 14; sub.s32 %r351, %r3445, %r3446; shf.l.wrap.b32 %r352, %r351, %r351, 11; xor.b32 %r3447, %r350, 514703112; sub.s32 %r353, %r3447, %r3446; shf.l.wrap.b32 %r354, %r353, %r353, 11; mov.u32 %r4376, %r3441; $L__BB199_288: add.s32 %r4376, %r4376, 1; cvt.rn.f32.s32 %f2282, %r4376; add.ftz.f32 %f2283, %f910, %f2282; mov.b32 %r3448, %f2283; add.s32 %r3449, %r3448, -559038716; xor.b32 %r3450, %r3449, %r348; sub.s32 %r3451, %r3450, %r349; xor.b32 %r3452, %r3451, %r347; shf.l.wrap.b32 %r3453, %r3451, %r3451, 25; sub.s32 %r3454, %r3452, %r3453; xor.b32 %r3455, %r3454, %r348; shf.l.wrap.b32 %r3456, %r3454, %r3454, 16; sub.s32 %r3457, %r3455, %r3456; xor.b32 %r3458, %r3457, %r3451; shf.l.wrap.b32 %r3459, %r3457, %r3457, 4; sub.s32 %r3460, %r3458, %r3459; xor.b32 %r3461, %r3460, %r3454; shf.l.wrap.b32 %r3462, %r3460, %r3460, 14; sub.s32 %r3463, %r3461, %r3462; xor.b32 %r3464, %r3463, %r3457; shf.l.wrap.b32 %r3465, %r3463, %r3463, 24; sub.s32 %r3466, %r3464, %r3465; cvt.rn.f32.u32 %f2284, %r3466; mul.ftz.f32 %f929, %f2284, 0f2F800000; add.s32 %r357, %r3448, -559038712; xor.b32 %r3467, %r357, %r351; sub.s32 %r3468, %r3467, %r352; xor.b32 %r3469, %r3468, %r350; shf.l.wrap.b32 %r3470, %r3468, %r3468, 25; sub.s32 %r3471, %r3469, %r3470; xor.b32 %r3472, %r3471, %r351; shf.l.wrap.b32 %r3473, %r3471, %r3471, 16; sub.s32 %r3474, %r3472, %r3473; xor.b32 %r3475, %r3474, %r3468; shf.l.wrap.b32 %r3476, %r3474, %r3474, 4; sub.s32 %r3477, %r3475, %r3476; xor.b32 %r3478, %r3477, %r3471; shf.l.wrap.b32 %r3479, %r3477, %r3477, 14; sub.s32 %r3480, %r3478, %r3479; xor.b32 %r3481, %r3480, %r3474; shf.l.wrap.b32 %r3482, %r3480, %r3480, 24; sub.s32 %r3483, %r3481, %r3482; cvt.rn.f32.u32 %f2285, %r3483; mul.ftz.f32 %f930, %f2285, 0f2F800000; fma.rn.ftz.f32 %f931, %f19, %f929, %f2282; fma.rn.ftz.f32 %f932, %f19, %f930, %f922; @%p150 bra $L__BB199_295; @%p151 bra $L__BB199_294; bra.uni $L__BB199_290; $L__BB199_294: sub.ftz.f32 %f2304, %f931, %f911; abs.ftz.f32 %f2305, %f2304; sub.ftz.f32 %f2306, %f932, %f913; abs.ftz.f32 %f2307, %f2306; add.ftz.f32 %f3291, %f2305, %f2307; bra.uni $L__BB199_296; $L__BB199_295: sub.ftz.f32 %f2308, %f931, %f911; sub.ftz.f32 %f2309, %f932, %f913; mul.ftz.f32 %f2310, %f2309, %f2309; fma.rn.ftz.f32 %f2311, %f2308, %f2308, %f2310; sqrt.approx.ftz.f32 %f3291, %f2311; bra.uni $L__BB199_296; $L__BB199_290: @%p152 bra $L__BB199_293; bra.uni $L__BB199_291; $L__BB199_293: sub.ftz.f32 %f2300, %f931, %f911; abs.ftz.f32 %f2301, %f2300; sub.ftz.f32 %f2302, %f932, %f913; abs.ftz.f32 %f2303, %f2302; max.ftz.f32 %f3291, %f2301, %f2303; bra.uni $L__BB199_296; $L__BB199_291: mov.f32 %f3291, 0f00000000; @%p153 bra $L__BB199_296; sub.ftz.f32 %f2287, %f931, %f911; abs.ftz.f32 %f2288, %f2287; lg2.approx.ftz.f32 %f2289, %f2288; mul.ftz.f32 %f2290, %f15, %f2289; ex2.approx.ftz.f32 %f2291, %f2290; sub.ftz.f32 %f2292, %f932, %f913; abs.ftz.f32 %f2293, %f2292; lg2.approx.ftz.f32 %f2294, %f2293; mul.ftz.f32 %f2295, %f15, %f2294; ex2.approx.ftz.f32 %f2296, %f2295; add.ftz.f32 %f2297, %f2291, %f2296; lg2.approx.ftz.f32 %f2298, %f2297; mul.ftz.f32 %f2299, %f915, %f2298; ex2.approx.ftz.f32 %f3291, %f2299; $L__BB199_296: sub.ftz.f32 %f2313, %f3386, %f3291; mul.ftz.f32 %f2314, %f2313, 0f3F000000; div.approx.ftz.f32 %f2315, %f2314, %f20; add.ftz.f32 %f938, %f2315, 0f3F000000; setp.lt.ftz.f32 %p154, %f938, 0f00000000; mov.f32 %f3292, 0f00000000; @%p154 bra $L__BB199_299; setp.ge.ftz.f32 %p155, %f938, 0f3F800000; mov.f32 %f3292, 0f3F800000; @%p155 bra $L__BB199_299; mov.f32 %f2317, 0f3F800000; sub.ftz.f32 %f2318, %f2317, 0f00000000; div.approx.ftz.f32 %f2319, %f938, %f2318; add.ftz.f32 %f2320, %f2319, %f2319; mov.f32 %f2321, 0f40400000; sub.ftz.f32 %f2322, %f2321, %f2320; mul.ftz.f32 %f2323, %f2319, %f2319; mul.ftz.f32 %f3292, %f2323, %f2322; $L__BB199_299: mov.f32 %f2324, 0f3F800000; sub.ftz.f32 %f2325, %f2324, %f3292; mul.ftz.f32 %f2326, %f20, %f3292; mul.ftz.f32 %f2327, %f2326, %f2325; sub.ftz.f32 %f2328, %f3291, %f3386; fma.rn.ftz.f32 %f2329, %f2328, %f3292, %f3386; sub.ftz.f32 %f3386, %f2329, %f2327; div.approx.ftz.f32 %f2330, %f2327, %f914; xor.b32 %r3484, %r357, %r353; sub.s32 %r3485, %r3484, %r354; xor.b32 %r3486, %r3485, %r350; shf.l.wrap.b32 %r3487, %r3485, %r3485, 25; sub.s32 %r3488, %r3486, %r3487; xor.b32 %r3489, %r3488, %r353; shf.l.wrap.b32 %r3490, %r3488, %r3488, 16; sub.s32 %r3491, %r3489, %r3490; xor.b32 %r3492, %r3491, %r3485; shf.l.wrap.b32 %r3493, %r3491, %r3491, 4; sub.s32 %r3494, %r3492, %r3493; xor.b32 %r3495, %r3494, %r3488; shf.l.wrap.b32 %r3496, %r3494, %r3494, 14; sub.s32 %r3497, %r3495, %r3496; xor.b32 %r3498, %r3497, %r3491; shf.l.wrap.b32 %r3499, %r3497, %r3497, 24; sub.s32 %r3500, %r3498, %r3499; cvt.rn.f32.u32 %f2331, %r3500; mul.ftz.f32 %f2332, %f2331, 0f2F800000; sub.ftz.f32 %f2333, %f2332, %f3389; sub.ftz.f32 %f2334, %f929, %f3391; sub.ftz.f32 %f2335, %f930, %f3390; fma.rn.ftz.f32 %f2336, %f2334, %f3292, %f3391; fma.rn.ftz.f32 %f2337, %f2335, %f3292, %f3390; fma.rn.ftz.f32 %f2338, %f2333, %f3292, %f3389; sub.ftz.f32 %f3391, %f2336, %f2330; sub.ftz.f32 %f3390, %f2337, %f2330; sub.ftz.f32 %f3389, %f2338, %f2330; sub.ftz.f32 %f2339, %f931, %f3286; sub.ftz.f32 %f2340, %f932, %f3287; fma.rn.ftz.f32 %f2341, %f2339, %f3292, %f3286; fma.rn.ftz.f32 %f2342, %f2340, %f3292, %f3287; sub.ftz.f32 %f3286, %f2341, %f2330; sub.ftz.f32 %f3287, %f2342, %f2330; setp.lt.s32 %p156, %r4376, 2; @%p156 bra $L__BB199_288; add.s32 %r358, %r4375, 1; setp.lt.s32 %p157, %r4375, 2; mov.u32 %r4375, %r358; @%p157 bra $L__BB199_287; add.ftz.f32 %f3365, %f910, %f3286; add.ftz.f32 %f3364, %f912, %f3287; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_362; $L__BB199_276: cvt.rmi.ftz.f32.f32 %f2177, %f22; sub.ftz.f32 %f878, %f22, %f2177; cvt.rmi.ftz.f32.f32 %f879, %f23; sub.ftz.f32 %f880, %f23, %f879; add.ftz.f32 %f2178, %f2177, 0fBF800000; mov.b32 %r3213, %f2178; add.s32 %r331, %r3213, -559038716; add.ftz.f32 %f2179, %f2177, 0f3F800000; mov.b32 %r3214, %f2179; add.s32 %r332, %r3214, -559038716; add.s32 %r333, %r3213, -559038712; add.ftz.f32 %f2180, %f2177, 0f00000000; mov.f32 %f3273, 0f00000000; mov.b32 %r3215, %f2180; add.s32 %r334, %r3215, -559038716; add.s32 %r335, %r3215, -559038712; add.s32 %r336, %r3214, -559038712; mov.f32 %f3272, 0f41000000; mov.u32 %r4373, -2; mov.f32 %f3274, %f3273; $L__BB199_277: add.s32 %r4373, %r4373, 1; cvt.rn.f32.s32 %f2181, %r4373; add.ftz.f32 %f2182, %f879, %f2181; mov.b32 %r3216, %f2182; add.s32 %r3217, %r3216, -559038716; xor.b32 %r3218, %r3217, -559038716; shf.l.wrap.b32 %r3219, %r3217, %r3217, 14; sub.s32 %r3220, %r3218, %r3219; shf.l.wrap.b32 %r3221, %r3220, %r3220, 11; add.s32 %r3222, %r3216, -559038712; xor.b32 %r3223, %r3222, 506314504; shf.l.wrap.b32 %r3224, %r3222, %r3222, 14; sub.s32 %r3225, %r3223, %r3224; shf.l.wrap.b32 %r3226, %r3225, %r3225, 11; xor.b32 %r3227, %r331, %r3220; sub.s32 %r3228, %r3227, %r3221; xor.b32 %r3229, %r3228, %r3217; shf.l.wrap.b32 %r3230, %r3228, %r3228, 25; sub.s32 %r3231, %r3229, %r3230; xor.b32 %r3232, %r3231, %r3220; shf.l.wrap.b32 %r3233, %r3231, %r3231, 16; sub.s32 %r3234, %r3232, %r3233; xor.b32 %r3235, %r3234, %r3228; shf.l.wrap.b32 %r3236, %r3234, %r3234, 4; sub.s32 %r3237, %r3235, %r3236; xor.b32 %r3238, %r3237, %r3231; shf.l.wrap.b32 %r3239, %r3237, %r3237, 14; sub.s32 %r3240, %r3238, %r3239; xor.b32 %r3241, %r3240, %r3234; shf.l.wrap.b32 %r3242, %r3240, %r3240, 24; sub.s32 %r3243, %r3241, %r3242; cvt.rn.f32.u32 %f2183, %r3243; mul.ftz.f32 %f2184, %f2183, 0f2F800000; xor.b32 %r3244, %r333, %r3225; sub.s32 %r3245, %r3244, %r3226; xor.b32 %r3246, %r3245, %r3222; shf.l.wrap.b32 %r3247, %r3245, %r3245, 25; sub.s32 %r3248, %r3246, %r3247; xor.b32 %r3249, %r3248, %r3225; shf.l.wrap.b32 %r3250, %r3248, %r3248, 16; sub.s32 %r3251, %r3249, %r3250; xor.b32 %r3252, %r3251, %r3245; shf.l.wrap.b32 %r3253, %r3251, %r3251, 4; sub.s32 %r3254, %r3252, %r3253; xor.b32 %r3255, %r3254, %r3248; shf.l.wrap.b32 %r3256, %r3254, %r3254, 14; sub.s32 %r3257, %r3255, %r3256; xor.b32 %r3258, %r3257, %r3251; shf.l.wrap.b32 %r3259, %r3257, %r3257, 24; sub.s32 %r3260, %r3258, %r3259; cvt.rn.f32.u32 %f2185, %r3260; mul.ftz.f32 %f2186, %f2185, 0f2F800000; fma.rn.ftz.f32 %f2187, %f19, %f2184, 0fBF800000; fma.rn.ftz.f32 %f2188, %f19, %f2186, %f2181; sub.ftz.f32 %f2189, %f2187, %f878; sub.ftz.f32 %f2190, %f2188, %f880; mul.ftz.f32 %f2191, %f2190, %f2190; fma.rn.ftz.f32 %f2192, %f2189, %f2189, %f2191; setp.lt.ftz.f32 %p142, %f2192, %f3272; selp.f32 %f2193, %f2189, %f3274, %p142; selp.f32 %f2194, %f2190, %f3273, %p142; selp.f32 %f2195, %f2192, %f3272, %p142; xor.b32 %r3261, %r334, %r3220; sub.s32 %r3262, %r3261, %r3221; xor.b32 %r3263, %r3262, %r3217; shf.l.wrap.b32 %r3264, %r3262, %r3262, 25; sub.s32 %r3265, %r3263, %r3264; xor.b32 %r3266, %r3265, %r3220; shf.l.wrap.b32 %r3267, %r3265, %r3265, 16; sub.s32 %r3268, %r3266, %r3267; xor.b32 %r3269, %r3268, %r3262; shf.l.wrap.b32 %r3270, %r3268, %r3268, 4; sub.s32 %r3271, %r3269, %r3270; xor.b32 %r3272, %r3271, %r3265; shf.l.wrap.b32 %r3273, %r3271, %r3271, 14; sub.s32 %r3274, %r3272, %r3273; xor.b32 %r3275, %r3274, %r3268; shf.l.wrap.b32 %r3276, %r3274, %r3274, 24; sub.s32 %r3277, %r3275, %r3276; cvt.rn.f32.u32 %f2196, %r3277; mul.ftz.f32 %f2197, %f2196, 0f2F800000; xor.b32 %r3278, %r335, %r3225; sub.s32 %r3279, %r3278, %r3226; xor.b32 %r3280, %r3279, %r3222; shf.l.wrap.b32 %r3281, %r3279, %r3279, 25; sub.s32 %r3282, %r3280, %r3281; xor.b32 %r3283, %r3282, %r3225; shf.l.wrap.b32 %r3284, %r3282, %r3282, 16; sub.s32 %r3285, %r3283, %r3284; xor.b32 %r3286, %r3285, %r3279; shf.l.wrap.b32 %r3287, %r3285, %r3285, 4; sub.s32 %r3288, %r3286, %r3287; xor.b32 %r3289, %r3288, %r3282; shf.l.wrap.b32 %r3290, %r3288, %r3288, 14; sub.s32 %r3291, %r3289, %r3290; xor.b32 %r3292, %r3291, %r3285; shf.l.wrap.b32 %r3293, %r3291, %r3291, 24; sub.s32 %r3294, %r3292, %r3293; cvt.rn.f32.u32 %f2198, %r3294; mul.ftz.f32 %f2199, %f2198, 0f2F800000; fma.rn.ftz.f32 %f2200, %f19, %f2197, 0f00000000; fma.rn.ftz.f32 %f2201, %f19, %f2199, %f2181; sub.ftz.f32 %f2202, %f2200, %f878; sub.ftz.f32 %f2203, %f2201, %f880; mul.ftz.f32 %f2204, %f2203, %f2203; fma.rn.ftz.f32 %f2205, %f2202, %f2202, %f2204; setp.lt.ftz.f32 %p143, %f2205, %f2195; selp.f32 %f2206, %f2202, %f2193, %p143; selp.f32 %f2207, %f2203, %f2194, %p143; selp.f32 %f2208, %f2205, %f2195, %p143; xor.b32 %r3295, %r332, %r3220; sub.s32 %r3296, %r3295, %r3221; xor.b32 %r3297, %r3296, %r3217; shf.l.wrap.b32 %r3298, %r3296, %r3296, 25; sub.s32 %r3299, %r3297, %r3298; xor.b32 %r3300, %r3299, %r3220; shf.l.wrap.b32 %r3301, %r3299, %r3299, 16; sub.s32 %r3302, %r3300, %r3301; xor.b32 %r3303, %r3302, %r3296; shf.l.wrap.b32 %r3304, %r3302, %r3302, 4; sub.s32 %r3305, %r3303, %r3304; xor.b32 %r3306, %r3305, %r3299; shf.l.wrap.b32 %r3307, %r3305, %r3305, 14; sub.s32 %r3308, %r3306, %r3307; xor.b32 %r3309, %r3308, %r3302; shf.l.wrap.b32 %r3310, %r3308, %r3308, 24; sub.s32 %r3311, %r3309, %r3310; cvt.rn.f32.u32 %f2209, %r3311; mul.ftz.f32 %f2210, %f2209, 0f2F800000; xor.b32 %r3312, %r336, %r3225; sub.s32 %r3313, %r3312, %r3226; xor.b32 %r3314, %r3313, %r3222; shf.l.wrap.b32 %r3315, %r3313, %r3313, 25; sub.s32 %r3316, %r3314, %r3315; xor.b32 %r3317, %r3316, %r3225; shf.l.wrap.b32 %r3318, %r3316, %r3316, 16; sub.s32 %r3319, %r3317, %r3318; xor.b32 %r3320, %r3319, %r3313; shf.l.wrap.b32 %r3321, %r3319, %r3319, 4; sub.s32 %r3322, %r3320, %r3321; xor.b32 %r3323, %r3322, %r3316; shf.l.wrap.b32 %r3324, %r3322, %r3322, 14; sub.s32 %r3325, %r3323, %r3324; xor.b32 %r3326, %r3325, %r3319; shf.l.wrap.b32 %r3327, %r3325, %r3325, 24; sub.s32 %r3328, %r3326, %r3327; cvt.rn.f32.u32 %f2211, %r3328; mul.ftz.f32 %f2212, %f2211, 0f2F800000; fma.rn.ftz.f32 %f2213, %f19, %f2210, 0f3F800000; fma.rn.ftz.f32 %f2214, %f19, %f2212, %f2181; sub.ftz.f32 %f2215, %f2213, %f878; sub.ftz.f32 %f2216, %f2214, %f880; mul.ftz.f32 %f2217, %f2216, %f2216; fma.rn.ftz.f32 %f2218, %f2215, %f2215, %f2217; setp.lt.ftz.f32 %p144, %f2218, %f2208; selp.f32 %f3274, %f2215, %f2206, %p144; selp.f32 %f3273, %f2216, %f2207, %p144; selp.f32 %f3272, %f2218, %f2208, %p144; setp.lt.s32 %p145, %r4373, 1; @%p145 bra $L__BB199_277; mov.f32 %f3386, 0f41000000; mov.u32 %r4374, -1; $L__BB199_279: cvt.rn.f32.s32 %f888, %r4374; add.ftz.f32 %f2220, %f879, %f888; mov.b32 %r3330, %f2220; add.s32 %r340, %r3330, -559038716; xor.b32 %r3331, %r340, -559038716; shf.l.wrap.b32 %r3332, %r340, %r340, 14; sub.s32 %r341, %r3331, %r3332; shf.l.wrap.b32 %r342, %r341, %r341, 11; add.s32 %r3333, %r3330, -559038712; xor.b32 %r3334, %r3333, 506314504; shf.l.wrap.b32 %r3335, %r3333, %r3333, 14; sub.s32 %r343, %r3334, %r3335; shf.l.wrap.b32 %r344, %r343, %r343, 11; xor.b32 %r3336, %r331, %r341; sub.s32 %r3337, %r3336, %r342; xor.b32 %r3338, %r3337, %r340; shf.l.wrap.b32 %r3339, %r3337, %r3337, 25; sub.s32 %r3340, %r3338, %r3339; xor.b32 %r3341, %r3340, %r341; shf.l.wrap.b32 %r3342, %r3340, %r3340, 16; sub.s32 %r3343, %r3341, %r3342; xor.b32 %r3344, %r3343, %r3337; shf.l.wrap.b32 %r3345, %r3343, %r3343, 4; sub.s32 %r3346, %r3344, %r3345; xor.b32 %r3347, %r3346, %r3340; shf.l.wrap.b32 %r3348, %r3346, %r3346, 14; sub.s32 %r3349, %r3347, %r3348; xor.b32 %r3350, %r3349, %r3343; shf.l.wrap.b32 %r3351, %r3349, %r3349, 24; sub.s32 %r3352, %r3350, %r3351; cvt.rn.f32.u32 %f2221, %r3352; mul.ftz.f32 %f2222, %f2221, 0f2F800000; xor.b32 %r3353, %r333, %r343; sub.s32 %r3354, %r3353, %r344; xor.b32 %r3355, %r3354, %r3333; shf.l.wrap.b32 %r3356, %r3354, %r3354, 25; sub.s32 %r3357, %r3355, %r3356; xor.b32 %r3358, %r3357, %r343; shf.l.wrap.b32 %r3359, %r3357, %r3357, 16; sub.s32 %r3360, %r3358, %r3359; xor.b32 %r3361, %r3360, %r3354; shf.l.wrap.b32 %r3362, %r3360, %r3360, 4; sub.s32 %r3363, %r3361, %r3362; xor.b32 %r3364, %r3363, %r3357; shf.l.wrap.b32 %r3365, %r3363, %r3363, 14; sub.s32 %r3366, %r3364, %r3365; xor.b32 %r3367, %r3366, %r3360; shf.l.wrap.b32 %r3368, %r3366, %r3366, 24; sub.s32 %r3369, %r3367, %r3368; cvt.rn.f32.u32 %f2223, %r3369; mul.ftz.f32 %f2224, %f2223, 0f2F800000; fma.rn.ftz.f32 %f2225, %f19, %f2222, 0fBF800000; fma.rn.ftz.f32 %f2226, %f19, %f2224, %f888; sub.ftz.f32 %f889, %f2225, %f878; sub.ftz.f32 %f890, %f2226, %f880; sub.ftz.f32 %f891, %f889, %f3274; sub.ftz.f32 %f892, %f890, %f3273; mul.ftz.f32 %f2227, %f892, %f892; fma.rn.ftz.f32 %f893, %f891, %f891, %f2227; setp.leu.ftz.f32 %p146, %f893, 0f38D1B717; @%p146 bra $L__BB199_281; add.ftz.f32 %f2228, %f3274, %f889; mul.ftz.f32 %f2229, %f2228, 0f3F000000; add.ftz.f32 %f2230, %f3273, %f890; mul.ftz.f32 %f2231, %f2230, 0f3F000000; rsqrt.approx.ftz.f32 %f2232, %f893; mul.ftz.f32 %f2233, %f891, %f2232; mul.ftz.f32 %f2234, %f892, %f2232; mul.ftz.f32 %f2235, %f2231, %f2234; fma.rn.ftz.f32 %f2236, %f2229, %f2233, %f2235; min.ftz.f32 %f3386, %f3386, %f2236; $L__BB199_281: xor.b32 %r3370, %r334, %r341; sub.s32 %r3371, %r3370, %r342; xor.b32 %r3372, %r3371, %r340; shf.l.wrap.b32 %r3373, %r3371, %r3371, 25; sub.s32 %r3374, %r3372, %r3373; xor.b32 %r3375, %r3374, %r341; shf.l.wrap.b32 %r3376, %r3374, %r3374, 16; sub.s32 %r3377, %r3375, %r3376; xor.b32 %r3378, %r3377, %r3371; shf.l.wrap.b32 %r3379, %r3377, %r3377, 4; sub.s32 %r3380, %r3378, %r3379; xor.b32 %r3381, %r3380, %r3374; shf.l.wrap.b32 %r3382, %r3380, %r3380, 14; sub.s32 %r3383, %r3381, %r3382; xor.b32 %r3384, %r3383, %r3377; shf.l.wrap.b32 %r3385, %r3383, %r3383, 24; sub.s32 %r3386, %r3384, %r3385; cvt.rn.f32.u32 %f2237, %r3386; mul.ftz.f32 %f2238, %f2237, 0f2F800000; xor.b32 %r3387, %r335, %r343; sub.s32 %r3388, %r3387, %r344; add.s32 %r3389, %r340, 4; xor.b32 %r3390, %r3388, %r3389; shf.l.wrap.b32 %r3391, %r3388, %r3388, 25; sub.s32 %r3392, %r3390, %r3391; xor.b32 %r3393, %r3392, %r343; shf.l.wrap.b32 %r3394, %r3392, %r3392, 16; sub.s32 %r3395, %r3393, %r3394; xor.b32 %r3396, %r3395, %r3388; shf.l.wrap.b32 %r3397, %r3395, %r3395, 4; sub.s32 %r3398, %r3396, %r3397; xor.b32 %r3399, %r3398, %r3392; shf.l.wrap.b32 %r3400, %r3398, %r3398, 14; sub.s32 %r3401, %r3399, %r3400; xor.b32 %r3402, %r3401, %r3395; shf.l.wrap.b32 %r3403, %r3401, %r3401, 24; sub.s32 %r3404, %r3402, %r3403; cvt.rn.f32.u32 %f2239, %r3404; mul.ftz.f32 %f2240, %f2239, 0f2F800000; fma.rn.ftz.f32 %f2241, %f19, %f2238, 0f00000000; fma.rn.ftz.f32 %f2242, %f19, %f2240, %f888; sub.ftz.f32 %f896, %f2241, %f878; sub.ftz.f32 %f897, %f2242, %f880; sub.ftz.f32 %f898, %f896, %f3274; sub.ftz.f32 %f899, %f897, %f3273; mul.ftz.f32 %f2243, %f899, %f899; fma.rn.ftz.f32 %f900, %f898, %f898, %f2243; setp.leu.ftz.f32 %p147, %f900, 0f38D1B717; @%p147 bra $L__BB199_283; add.ftz.f32 %f2244, %f3274, %f896; mul.ftz.f32 %f2245, %f2244, 0f3F000000; add.ftz.f32 %f2246, %f3273, %f897; mul.ftz.f32 %f2247, %f2246, 0f3F000000; rsqrt.approx.ftz.f32 %f2248, %f900; mul.ftz.f32 %f2249, %f898, %f2248; mul.ftz.f32 %f2250, %f899, %f2248; mul.ftz.f32 %f2251, %f2247, %f2250; fma.rn.ftz.f32 %f2252, %f2245, %f2249, %f2251; min.ftz.f32 %f3386, %f3386, %f2252; $L__BB199_283: xor.b32 %r3405, %r332, %r341; sub.s32 %r3406, %r3405, %r342; xor.b32 %r3407, %r3406, %r340; shf.l.wrap.b32 %r3408, %r3406, %r3406, 25; sub.s32 %r3409, %r3407, %r3408; xor.b32 %r3410, %r3409, %r341; shf.l.wrap.b32 %r3411, %r3409, %r3409, 16; sub.s32 %r3412, %r3410, %r3411; xor.b32 %r3413, %r3412, %r3406; shf.l.wrap.b32 %r3414, %r3412, %r3412, 4; sub.s32 %r3415, %r3413, %r3414; xor.b32 %r3416, %r3415, %r3409; shf.l.wrap.b32 %r3417, %r3415, %r3415, 14; sub.s32 %r3418, %r3416, %r3417; xor.b32 %r3419, %r3418, %r3412; shf.l.wrap.b32 %r3420, %r3418, %r3418, 24; sub.s32 %r3421, %r3419, %r3420; cvt.rn.f32.u32 %f2253, %r3421; mul.ftz.f32 %f2254, %f2253, 0f2F800000; xor.b32 %r3422, %r336, %r343; sub.s32 %r3423, %r3422, %r344; xor.b32 %r3425, %r3423, %r3389; shf.l.wrap.b32 %r3426, %r3423, %r3423, 25; sub.s32 %r3427, %r3425, %r3426; xor.b32 %r3428, %r3427, %r343; shf.l.wrap.b32 %r3429, %r3427, %r3427, 16; sub.s32 %r3430, %r3428, %r3429; xor.b32 %r3431, %r3430, %r3423; shf.l.wrap.b32 %r3432, %r3430, %r3430, 4; sub.s32 %r3433, %r3431, %r3432; xor.b32 %r3434, %r3433, %r3427; shf.l.wrap.b32 %r3435, %r3433, %r3433, 14; sub.s32 %r3436, %r3434, %r3435; xor.b32 %r3437, %r3436, %r3430; shf.l.wrap.b32 %r3438, %r3436, %r3436, 24; sub.s32 %r3439, %r3437, %r3438; cvt.rn.f32.u32 %f2255, %r3439; mul.ftz.f32 %f2256, %f2255, 0f2F800000; fma.rn.ftz.f32 %f2257, %f19, %f2254, 0f3F800000; fma.rn.ftz.f32 %f2258, %f19, %f2256, %f888; sub.ftz.f32 %f903, %f2257, %f878; sub.ftz.f32 %f904, %f2258, %f880; sub.ftz.f32 %f905, %f903, %f3274; sub.ftz.f32 %f906, %f904, %f3273; mul.ftz.f32 %f2259, %f906, %f906; fma.rn.ftz.f32 %f907, %f905, %f905, %f2259; setp.leu.ftz.f32 %p148, %f907, 0f38D1B717; @%p148 bra $L__BB199_285; add.ftz.f32 %f2260, %f3274, %f903; mul.ftz.f32 %f2261, %f2260, 0f3F000000; add.ftz.f32 %f2262, %f3273, %f904; mul.ftz.f32 %f2263, %f2262, 0f3F000000; rsqrt.approx.ftz.f32 %f2264, %f907; mul.ftz.f32 %f2265, %f905, %f2264; mul.ftz.f32 %f2266, %f906, %f2264; mul.ftz.f32 %f2267, %f2263, %f2266; fma.rn.ftz.f32 %f2268, %f2261, %f2265, %f2267; min.ftz.f32 %f3386, %f3386, %f2268; $L__BB199_285: add.s32 %r345, %r4374, 1; setp.lt.s32 %p149, %r4374, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4374, %r345; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3364, %f3388; mov.f32 %f3365, %f3388; @%p149 bra $L__BB199_279; bra.uni $L__BB199_362; $L__BB199_85: ld.param.u32 %r4312, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f180, %f22; sub.ftz.f32 %f181, %f22, %f180; cvt.rmi.ftz.f32.f32 %f182, %f23; sub.ftz.f32 %f183, %f23, %f182; cvt.rmi.ftz.f32.f32 %f184, %f24; sub.ftz.f32 %f185, %f24, %f184; cvt.rmi.ftz.f32.f32 %f186, %f21; sub.ftz.f32 %f187, %f21, %f186; fma.rn.ftz.f32 %f188, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f189, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1126, -2; setp.eq.s32 %p38, %r4312, 0; setp.eq.s32 %p39, %r4312, 1; setp.eq.s32 %p40, %r4312, 2; setp.ne.s32 %p41, %r4312, 3; mov.u32 %r4339, %r1126; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f2897, %f3389; mov.f32 %f2898, %f3389; mov.f32 %f2899, %f3389; mov.f32 %f2900, %f3389; $L__BB199_86: cvt.rn.f32.s32 %f198, %r4339; add.ftz.f32 %f1350, %f186, %f198; mov.b32 %r113, %f1350; add.s32 %r114, %r113, -559038708; shf.l.wrap.b32 %r115, %r114, %r114, 4; mov.u32 %r4340, %r1126; $L__BB199_87: cvt.rn.f32.s32 %f207, %r4340; add.ftz.f32 %f1351, %f184, %f207; mov.b32 %r117, %f1351; add.s32 %r118, %r117, -559038708; shf.l.wrap.b32 %r119, %r118, %r118, 4; add.s32 %r120, %r118, %r114; sub.s32 %r1129, %r117, %r113; xor.b32 %r121, %r1129, %r115; shf.l.wrap.b32 %r122, %r121, %r121, 6; mov.u32 %r4341, %r1126; $L__BB199_88: cvt.rn.f32.s32 %f216, %r4341; add.ftz.f32 %f1352, %f182, %f216; mov.b32 %r124, %f1352; add.s32 %r125, %r124, -559038708; add.s32 %r126, %r125, %r118; sub.s32 %r1131, %r113, %r124; shf.l.wrap.b32 %r1132, %r125, %r125, 4; xor.b32 %r127, %r1132, %r1131; shf.l.wrap.b32 %r128, %r127, %r127, 6; sub.s32 %r1133, %r124, %r113; xor.b32 %r1134, %r1133, %r115; sub.s32 %r1135, %r118, %r1134; shf.l.wrap.b32 %r1136, %r1134, %r1134, 6; xor.b32 %r1137, %r1136, %r1135; add.s32 %r1138, %r1134, %r120; sub.s32 %r1139, %r120, %r1137; shf.l.wrap.b32 %r1140, %r1137, %r1137, 8; xor.b32 %r1141, %r1140, %r1139; add.s32 %r1142, %r1137, %r1138; sub.s32 %r1143, %r1138, %r1141; shf.l.wrap.b32 %r1144, %r1141, %r1141, 16; xor.b32 %r1145, %r1144, %r1143; add.s32 %r1146, %r1141, %r1142; sub.s32 %r1147, %r1142, %r1145; shf.l.wrap.b32 %r1148, %r1145, %r1145, 19; xor.b32 %r1149, %r1148, %r1147; add.s32 %r129, %r1145, %r1146; sub.s32 %r1150, %r1146, %r1149; shf.l.wrap.b32 %r1151, %r1149, %r1149, 4; add.s32 %r130, %r1149, %r129; xor.b32 %r1152, %r130, %r1150; xor.b32 %r1153, %r1152, %r1151; shf.l.wrap.b32 %r1154, %r130, %r130, 14; sub.s32 %r131, %r1153, %r1154; shf.l.wrap.b32 %r132, %r131, %r131, 11; sub.s32 %r1155, %r118, %r127; xor.b32 %r1156, %r128, %r1155; add.s32 %r1157, %r127, %r126; sub.s32 %r1158, %r126, %r1156; shf.l.wrap.b32 %r1159, %r1156, %r1156, 8; xor.b32 %r1160, %r1159, %r1158; add.s32 %r1161, %r1156, %r1157; sub.s32 %r1162, %r1157, %r1160; shf.l.wrap.b32 %r1163, %r1160, %r1160, 16; xor.b32 %r1164, %r1163, %r1162; add.s32 %r1165, %r1160, %r1161; sub.s32 %r1166, %r1161, %r1164; shf.l.wrap.b32 %r1167, %r1164, %r1164, 19; xor.b32 %r1168, %r1167, %r1166; add.s32 %r133, %r1164, %r1165; sub.s32 %r1169, %r1165, %r1168; shf.l.wrap.b32 %r1170, %r1168, %r1168, 4; add.s32 %r134, %r1168, %r133; xor.b32 %r1171, %r134, %r1169; xor.b32 %r1172, %r1171, %r1170; shf.l.wrap.b32 %r1173, %r134, %r134, 14; sub.s32 %r135, %r1172, %r1173; shf.l.wrap.b32 %r136, %r135, %r135, 11; mov.u32 %r4342, %r1126; $L__BB199_89: mov.u32 %r137, %r4342; cvt.rn.f32.s32 %f1353, %r137; add.ftz.f32 %f1354, %f180, %f1353; mov.b32 %r138, %f1354; sub.s32 %r1174, %r138, %r117; xor.b32 %r1175, %r1174, %r119; sub.s32 %r1176, %r125, %r1175; shf.l.wrap.b32 %r1177, %r1175, %r1175, 6; xor.b32 %r1178, %r1177, %r1176; add.s32 %r1179, %r1175, %r126; sub.s32 %r1180, %r126, %r1178; shf.l.wrap.b32 %r1181, %r1178, %r1178, 8; xor.b32 %r1182, %r1181, %r1180; add.s32 %r1183, %r1178, %r1179; sub.s32 %r1184, %r1179, %r1182; shf.l.wrap.b32 %r1185, %r1182, %r1182, 16; xor.b32 %r1186, %r1185, %r1184; add.s32 %r1187, %r1182, %r1183; sub.s32 %r1188, %r1183, %r1186; shf.l.wrap.b32 %r1189, %r1186, %r1186, 19; xor.b32 %r1190, %r1189, %r1188; add.s32 %r1191, %r1186, %r1187; sub.s32 %r1192, %r1187, %r1190; shf.l.wrap.b32 %r1193, %r1190, %r1190, 4; add.s32 %r1194, %r1190, %r1191; add.s32 %r1195, %r1191, %r113; xor.b32 %r1196, %r1194, %r1192; xor.b32 %r1197, %r1196, %r1193; shf.l.wrap.b32 %r1198, %r1194, %r1194, 14; sub.s32 %r1199, %r1197, %r1198; xor.b32 %r1200, %r1199, %r1195; shf.l.wrap.b32 %r1201, %r1199, %r1199, 11; sub.s32 %r1202, %r1200, %r1201; xor.b32 %r1203, %r1202, %r1194; shf.l.wrap.b32 %r1204, %r1202, %r1202, 25; sub.s32 %r1205, %r1203, %r1204; xor.b32 %r1206, %r1205, %r1199; shf.l.wrap.b32 %r1207, %r1205, %r1205, 16; sub.s32 %r1208, %r1206, %r1207; xor.b32 %r1209, %r1208, %r1202; shf.l.wrap.b32 %r1210, %r1208, %r1208, 4; sub.s32 %r1211, %r1209, %r1210; xor.b32 %r1212, %r1211, %r1205; shf.l.wrap.b32 %r1213, %r1211, %r1211, 14; sub.s32 %r1214, %r1212, %r1213; xor.b32 %r1215, %r1214, %r1208; shf.l.wrap.b32 %r1216, %r1214, %r1214, 24; sub.s32 %r1217, %r1215, %r1216; cvt.rn.f32.u32 %f1355, %r1217; mul.ftz.f32 %f225, %f1355, 0f2F800000; add.s32 %r1218, %r138, -559038708; add.s32 %r1219, %r1218, %r125; sub.s32 %r1220, %r1218, %r127; xor.b32 %r1221, %r1220, %r128; add.s32 %r1222, %r1219, %r127; sub.s32 %r1223, %r1219, %r1221; shf.l.wrap.b32 %r1224, %r1221, %r1221, 8; xor.b32 %r1225, %r1224, %r1223; add.s32 %r1226, %r1221, %r1222; sub.s32 %r1227, %r1222, %r1225; shf.l.wrap.b32 %r1228, %r1225, %r1225, 16; xor.b32 %r1229, %r1228, %r1227; add.s32 %r1230, %r1225, %r1226; sub.s32 %r1231, %r1226, %r1229; shf.l.wrap.b32 %r1232, %r1229, %r1229, 19; xor.b32 %r1233, %r1232, %r1231; add.s32 %r1234, %r1229, %r1230; sub.s32 %r1235, %r1230, %r1233; shf.l.wrap.b32 %r1236, %r1233, %r1233, 4; add.s32 %r1237, %r1233, %r1234; add.s32 %r1238, %r1234, %r117; xor.b32 %r1239, %r1237, %r1235; xor.b32 %r1240, %r1239, %r1236; shf.l.wrap.b32 %r1241, %r1237, %r1237, 14; sub.s32 %r1242, %r1240, %r1241; xor.b32 %r1243, %r1242, %r1238; shf.l.wrap.b32 %r1244, %r1242, %r1242, 11; sub.s32 %r1245, %r1243, %r1244; xor.b32 %r1246, %r1245, %r1237; shf.l.wrap.b32 %r1247, %r1245, %r1245, 25; sub.s32 %r1248, %r1246, %r1247; xor.b32 %r1249, %r1248, %r1242; shf.l.wrap.b32 %r1250, %r1248, %r1248, 16; sub.s32 %r1251, %r1249, %r1250; xor.b32 %r1252, %r1251, %r1245; shf.l.wrap.b32 %r1253, %r1251, %r1251, 4; sub.s32 %r1254, %r1252, %r1253; xor.b32 %r1255, %r1254, %r1248; shf.l.wrap.b32 %r1256, %r1254, %r1254, 14; sub.s32 %r1257, %r1255, %r1256; xor.b32 %r1258, %r1257, %r1251; shf.l.wrap.b32 %r1259, %r1257, %r1257, 24; sub.s32 %r1260, %r1258, %r1259; cvt.rn.f32.u32 %f1356, %r1260; mul.ftz.f32 %f1357, %f1356, 0f2F800000; sub.s32 %r1261, %r117, %r138; shf.l.wrap.b32 %r1262, %r1218, %r1218, 4; xor.b32 %r1263, %r1262, %r1261; add.s32 %r139, %r1218, %r114; sub.s32 %r1264, %r114, %r1263; shf.l.wrap.b32 %r1265, %r1263, %r1263, 6; xor.b32 %r1266, %r1265, %r1264; add.s32 %r1267, %r1263, %r139; sub.s32 %r1268, %r139, %r1266; shf.l.wrap.b32 %r1269, %r1266, %r1266, 8; xor.b32 %r1270, %r1269, %r1268; add.s32 %r1271, %r1266, %r1267; sub.s32 %r1272, %r1267, %r1270; shf.l.wrap.b32 %r1273, %r1270, %r1270, 16; xor.b32 %r1274, %r1273, %r1272; add.s32 %r1275, %r1270, %r1271; sub.s32 %r1276, %r1271, %r1274; shf.l.wrap.b32 %r1277, %r1274, %r1274, 19; xor.b32 %r1278, %r1277, %r1276; add.s32 %r1279, %r1274, %r1275; sub.s32 %r1280, %r1275, %r1278; shf.l.wrap.b32 %r1281, %r1278, %r1278, 4; add.s32 %r1282, %r1278, %r1279; add.s32 %r1283, %r1279, %r124; xor.b32 %r1284, %r1282, %r1280; xor.b32 %r1285, %r1284, %r1281; shf.l.wrap.b32 %r1286, %r1282, %r1282, 14; sub.s32 %r1287, %r1285, %r1286; xor.b32 %r1288, %r1287, %r1283; shf.l.wrap.b32 %r1289, %r1287, %r1287, 11; sub.s32 %r1290, %r1288, %r1289; xor.b32 %r1291, %r1290, %r1282; shf.l.wrap.b32 %r1292, %r1290, %r1290, 25; sub.s32 %r1293, %r1291, %r1292; xor.b32 %r1294, %r1293, %r1287; shf.l.wrap.b32 %r1295, %r1293, %r1293, 16; sub.s32 %r1296, %r1294, %r1295; xor.b32 %r1297, %r1296, %r1290; shf.l.wrap.b32 %r1298, %r1296, %r1296, 4; sub.s32 %r1299, %r1297, %r1298; xor.b32 %r1300, %r1299, %r1293; shf.l.wrap.b32 %r1301, %r1299, %r1299, 14; sub.s32 %r1302, %r1300, %r1301; xor.b32 %r1303, %r1302, %r1296; shf.l.wrap.b32 %r1304, %r1302, %r1302, 24; sub.s32 %r1305, %r1303, %r1304; cvt.rn.f32.u32 %f1358, %r1305; mul.ftz.f32 %f1359, %f1358, 0f2F800000; add.s32 %r1306, %r129, %r138; xor.b32 %r1307, %r1306, %r131; sub.s32 %r1308, %r1307, %r132; xor.b32 %r1309, %r1308, %r130; shf.l.wrap.b32 %r1310, %r1308, %r1308, 25; sub.s32 %r1311, %r1309, %r1310; xor.b32 %r1312, %r1311, %r131; shf.l.wrap.b32 %r1313, %r1311, %r1311, 16; sub.s32 %r1314, %r1312, %r1313; xor.b32 %r1315, %r1314, %r1308; shf.l.wrap.b32 %r1316, %r1314, %r1314, 4; sub.s32 %r1317, %r1315, %r1316; xor.b32 %r1318, %r1317, %r1311; shf.l.wrap.b32 %r1319, %r1317, %r1317, 14; sub.s32 %r1320, %r1318, %r1319; xor.b32 %r1321, %r1320, %r1314; shf.l.wrap.b32 %r1322, %r1320, %r1320, 24; sub.s32 %r1323, %r1321, %r1322; cvt.rn.f32.u32 %f1360, %r1323; mul.ftz.f32 %f1361, %f1360, 0f2F800000; fma.rn.ftz.f32 %f226, %f19, %f225, %f1353; fma.rn.ftz.f32 %f227, %f19, %f1357, %f216; fma.rn.ftz.f32 %f228, %f19, %f1359, %f207; fma.rn.ftz.f32 %f229, %f19, %f1361, %f198; @%p38 bra $L__BB199_96; @%p39 bra $L__BB199_95; bra.uni $L__BB199_91; $L__BB199_95: sub.ftz.f32 %f1398, %f226, %f181; abs.ftz.f32 %f1399, %f1398; sub.ftz.f32 %f1400, %f227, %f183; abs.ftz.f32 %f1401, %f1400; add.ftz.f32 %f1402, %f1399, %f1401; sub.ftz.f32 %f1403, %f228, %f185; abs.ftz.f32 %f1404, %f1403; add.ftz.f32 %f1405, %f1402, %f1404; sub.ftz.f32 %f1406, %f229, %f187; abs.ftz.f32 %f1407, %f1406; add.ftz.f32 %f2901, %f1405, %f1407; bra.uni $L__BB199_97; $L__BB199_96: sub.ftz.f32 %f1408, %f226, %f181; mul.ftz.f32 %f1409, %f1408, %f1408; sub.ftz.f32 %f1410, %f227, %f183; fma.rn.ftz.f32 %f1411, %f1410, %f1410, %f1409; sub.ftz.f32 %f1412, %f228, %f185; mul.ftz.f32 %f1413, %f1412, %f1412; sub.ftz.f32 %f1414, %f229, %f187; fma.rn.ftz.f32 %f1415, %f1414, %f1414, %f1413; add.ftz.f32 %f1416, %f1411, %f1415; sqrt.approx.ftz.f32 %f2901, %f1416; bra.uni $L__BB199_97; $L__BB199_91: @%p40 bra $L__BB199_94; bra.uni $L__BB199_92; $L__BB199_94: sub.ftz.f32 %f1388, %f226, %f181; abs.ftz.f32 %f1389, %f1388; sub.ftz.f32 %f1390, %f227, %f183; abs.ftz.f32 %f1391, %f1390; sub.ftz.f32 %f1392, %f228, %f185; abs.ftz.f32 %f1393, %f1392; sub.ftz.f32 %f1394, %f229, %f187; abs.ftz.f32 %f1395, %f1394; max.ftz.f32 %f1396, %f1393, %f1395; max.ftz.f32 %f1397, %f1391, %f1396; max.ftz.f32 %f2901, %f1389, %f1397; bra.uni $L__BB199_97; $L__BB199_92: mov.f32 %f2901, 0f00000000; @%p41 bra $L__BB199_97; sub.ftz.f32 %f1363, %f226, %f181; abs.ftz.f32 %f1364, %f1363; lg2.approx.ftz.f32 %f1365, %f1364; mul.ftz.f32 %f1366, %f15, %f1365; ex2.approx.ftz.f32 %f1367, %f1366; sub.ftz.f32 %f1368, %f227, %f183; abs.ftz.f32 %f1369, %f1368; lg2.approx.ftz.f32 %f1370, %f1369; mul.ftz.f32 %f1371, %f15, %f1370; ex2.approx.ftz.f32 %f1372, %f1371; add.ftz.f32 %f1373, %f1367, %f1372; sub.ftz.f32 %f1374, %f228, %f185; abs.ftz.f32 %f1375, %f1374; lg2.approx.ftz.f32 %f1376, %f1375; mul.ftz.f32 %f1377, %f15, %f1376; ex2.approx.ftz.f32 %f1378, %f1377; add.ftz.f32 %f1379, %f1373, %f1378; sub.ftz.f32 %f1380, %f229, %f187; abs.ftz.f32 %f1381, %f1380; lg2.approx.ftz.f32 %f1382, %f1381; mul.ftz.f32 %f1383, %f15, %f1382; ex2.approx.ftz.f32 %f1384, %f1383; add.ftz.f32 %f1385, %f1379, %f1384; lg2.approx.ftz.f32 %f1386, %f1385; mul.ftz.f32 %f1387, %f189, %f1386; ex2.approx.ftz.f32 %f2901, %f1387; $L__BB199_97: sub.ftz.f32 %f1418, %f3386, %f2901; mul.ftz.f32 %f1419, %f1418, 0f3F000000; div.approx.ftz.f32 %f1420, %f1419, %f20; add.ftz.f32 %f235, %f1420, 0f3F000000; setp.lt.ftz.f32 %p42, %f235, 0f00000000; mov.f32 %f2902, 0f00000000; @%p42 bra $L__BB199_100; setp.ge.ftz.f32 %p43, %f235, 0f3F800000; mov.f32 %f2902, %f1183; @%p43 bra $L__BB199_100; mov.f32 %f1422, 0f3F800000; sub.ftz.f32 %f1423, %f1422, 0f00000000; div.approx.ftz.f32 %f1424, %f235, %f1423; add.ftz.f32 %f1425, %f1424, %f1424; mov.f32 %f1426, 0f40400000; sub.ftz.f32 %f1427, %f1426, %f1425; mul.ftz.f32 %f1428, %f1424, %f1424; mul.ftz.f32 %f2902, %f1428, %f1427; $L__BB199_100: sub.ftz.f32 %f1430, %f1183, %f2902; mul.ftz.f32 %f1431, %f20, %f2902; mul.ftz.f32 %f1432, %f1431, %f1430; sub.ftz.f32 %f1433, %f2901, %f3386; fma.rn.ftz.f32 %f1434, %f1433, %f2902, %f3386; sub.ftz.f32 %f3386, %f1434, %f1432; div.approx.ftz.f32 %f1435, %f1432, %f188; sub.s32 %r1325, %r1218, %r121; xor.b32 %r1326, %r1325, %r122; sub.s32 %r1327, %r139, %r1326; shf.l.wrap.b32 %r1328, %r1326, %r1326, 8; xor.b32 %r1329, %r1328, %r1327; add.s32 %r1330, %r139, %r121; add.s32 %r1331, %r1326, %r1330; sub.s32 %r1332, %r1330, %r1329; shf.l.wrap.b32 %r1333, %r1329, %r1329, 16; xor.b32 %r1334, %r1333, %r1332; add.s32 %r1335, %r1329, %r1331; sub.s32 %r1336, %r1331, %r1334; shf.l.wrap.b32 %r1337, %r1334, %r1334, 19; xor.b32 %r1338, %r1337, %r1336; add.s32 %r1339, %r1334, %r1335; sub.s32 %r1340, %r1335, %r1338; shf.l.wrap.b32 %r1341, %r1338, %r1338, 4; add.s32 %r1342, %r1338, %r1339; add.s32 %r1343, %r1339, %r124; xor.b32 %r1344, %r1342, %r1340; xor.b32 %r1345, %r1344, %r1341; shf.l.wrap.b32 %r1346, %r1342, %r1342, 14; sub.s32 %r1347, %r1345, %r1346; xor.b32 %r1348, %r1347, %r1343; shf.l.wrap.b32 %r1349, %r1347, %r1347, 11; sub.s32 %r1350, %r1348, %r1349; xor.b32 %r1351, %r1350, %r1342; shf.l.wrap.b32 %r1352, %r1350, %r1350, 25; sub.s32 %r1353, %r1351, %r1352; xor.b32 %r1354, %r1353, %r1347; shf.l.wrap.b32 %r1355, %r1353, %r1353, 16; sub.s32 %r1356, %r1354, %r1355; xor.b32 %r1357, %r1356, %r1350; shf.l.wrap.b32 %r1358, %r1356, %r1356, 4; sub.s32 %r1359, %r1357, %r1358; xor.b32 %r1360, %r1359, %r1353; shf.l.wrap.b32 %r1361, %r1359, %r1359, 14; sub.s32 %r1362, %r1360, %r1361; xor.b32 %r1363, %r1362, %r1356; shf.l.wrap.b32 %r1364, %r1362, %r1362, 24; sub.s32 %r1365, %r1363, %r1364; cvt.rn.f32.u32 %f1436, %r1365; mul.ftz.f32 %f1437, %f1436, 0f2F800000; add.s32 %r1366, %r133, %r138; xor.b32 %r1367, %r1366, %r135; sub.s32 %r1368, %r1367, %r136; xor.b32 %r1369, %r1368, %r134; shf.l.wrap.b32 %r1370, %r1368, %r1368, 25; sub.s32 %r1371, %r1369, %r1370; xor.b32 %r1372, %r1371, %r135; shf.l.wrap.b32 %r1373, %r1371, %r1371, 16; sub.s32 %r1374, %r1372, %r1373; xor.b32 %r1375, %r1374, %r1368; shf.l.wrap.b32 %r1376, %r1374, %r1374, 4; sub.s32 %r1377, %r1375, %r1376; xor.b32 %r1378, %r1377, %r1371; shf.l.wrap.b32 %r1379, %r1377, %r1377, 14; sub.s32 %r1380, %r1378, %r1379; xor.b32 %r1381, %r1380, %r1374; shf.l.wrap.b32 %r1382, %r1380, %r1380, 24; sub.s32 %r1383, %r1381, %r1382; cvt.rn.f32.u32 %f1438, %r1383; mul.ftz.f32 %f1439, %f1438, 0f2F800000; sub.ftz.f32 %f1440, %f1437, %f3390; sub.ftz.f32 %f1441, %f1439, %f3389; sub.ftz.f32 %f1442, %f225, %f3391; fma.rn.ftz.f32 %f1443, %f1442, %f2902, %f3391; fma.rn.ftz.f32 %f1444, %f1440, %f2902, %f3390; fma.rn.ftz.f32 %f1445, %f1441, %f2902, %f3389; sub.ftz.f32 %f3391, %f1443, %f1435; sub.ftz.f32 %f3390, %f1444, %f1435; sub.ftz.f32 %f3389, %f1445, %f1435; sub.ftz.f32 %f1446, %f226, %f2900; sub.ftz.f32 %f1447, %f227, %f2899; sub.ftz.f32 %f1448, %f228, %f2898; sub.ftz.f32 %f1449, %f229, %f2897; fma.rn.ftz.f32 %f1450, %f1446, %f2902, %f2900; fma.rn.ftz.f32 %f1451, %f1447, %f2902, %f2899; fma.rn.ftz.f32 %f1452, %f1448, %f2902, %f2898; fma.rn.ftz.f32 %f1453, %f1449, %f2902, %f2897; sub.ftz.f32 %f2900, %f1450, %f1435; sub.ftz.f32 %f2899, %f1451, %f1435; sub.ftz.f32 %f2898, %f1452, %f1435; sub.ftz.f32 %f2897, %f1453, %f1435; add.s32 %r4342, %r137, 1; setp.lt.s32 %p44, %r137, 2; @%p44 bra $L__BB199_89; add.s32 %r141, %r4341, 1; setp.lt.s32 %p45, %r4341, 2; mov.u32 %r4341, %r141; @%p45 bra $L__BB199_88; add.s32 %r142, %r4340, 1; setp.lt.s32 %p46, %r4340, 2; mov.u32 %r4340, %r142; @%p46 bra $L__BB199_87; add.s32 %r143, %r4339, 1; setp.lt.s32 %p47, %r4339, 2; mov.u32 %r4339, %r143; @%p47 bra $L__BB199_86; add.ftz.f32 %f3048, %f180, %f2900; add.ftz.f32 %f3047, %f182, %f2899; add.ftz.f32 %f3046, %f184, %f2898; add.ftz.f32 %f3045, %f186, %f2897; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_136; $L__BB199_65: cvt.rmi.ftz.f32.f32 %f121, %f22; sub.ftz.f32 %f122, %f22, %f121; cvt.rmi.ftz.f32.f32 %f123, %f23; sub.ftz.f32 %f124, %f23, %f123; cvt.rmi.ftz.f32.f32 %f125, %f24; sub.ftz.f32 %f126, %f24, %f125; cvt.rmi.ftz.f32.f32 %f127, %f21; sub.ftz.f32 %f128, %f21, %f127; mov.f32 %f2862, 0f00000000; mov.u32 %r766, -1; mov.f32 %f2863, 0f41000000; mov.u32 %r4331, %r766; mov.f32 %f2861, %f2862; mov.f32 %f2860, %f2862; mov.f32 %f2859, %f2862; $L__BB199_66: cvt.rn.f32.s32 %f134, %r4331; add.ftz.f32 %f1276, %f127, %f134; mov.b32 %r65, %f1276; add.s32 %r66, %r65, -559038708; shf.l.wrap.b32 %r67, %r66, %r66, 4; mov.u32 %r4332, %r766; $L__BB199_67: cvt.rn.f32.s32 %f140, %r4332; add.ftz.f32 %f1277, %f125, %f140; mov.b32 %r69, %f1277; add.s32 %r70, %r69, -559038708; shf.l.wrap.b32 %r71, %r70, %r70, 4; add.s32 %r72, %r70, %r66; mov.u32 %r4333, %r766; $L__BB199_68: cvt.rn.f32.s32 %f146, %r4333; add.ftz.f32 %f1278, %f123, %f146; mov.b32 %r74, %f1278; add.s32 %r75, %r74, -559038708; add.s32 %r76, %r75, %r70; sub.s32 %r770, %r65, %r74; shf.l.wrap.b32 %r771, %r75, %r75, 4; xor.b32 %r77, %r771, %r770; shf.l.wrap.b32 %r78, %r77, %r77, 6; sub.s32 %r772, %r74, %r65; xor.b32 %r773, %r772, %r67; sub.s32 %r774, %r70, %r773; shf.l.wrap.b32 %r775, %r773, %r773, 6; xor.b32 %r776, %r775, %r774; add.s32 %r777, %r773, %r72; sub.s32 %r778, %r72, %r776; shf.l.wrap.b32 %r779, %r776, %r776, 8; xor.b32 %r780, %r779, %r778; add.s32 %r781, %r776, %r777; sub.s32 %r782, %r777, %r780; shf.l.wrap.b32 %r783, %r780, %r780, 16; xor.b32 %r784, %r783, %r782; add.s32 %r785, %r780, %r781; sub.s32 %r786, %r781, %r784; shf.l.wrap.b32 %r787, %r784, %r784, 19; xor.b32 %r788, %r787, %r786; add.s32 %r79, %r784, %r785; sub.s32 %r789, %r785, %r788; shf.l.wrap.b32 %r790, %r788, %r788, 4; add.s32 %r80, %r788, %r79; xor.b32 %r791, %r80, %r789; xor.b32 %r792, %r791, %r790; shf.l.wrap.b32 %r793, %r80, %r80, 14; sub.s32 %r81, %r792, %r793; shf.l.wrap.b32 %r82, %r81, %r81, 11; mov.u32 %r4334, %r766; $L__BB199_69: cvt.rn.f32.s32 %f1279, %r4334; add.ftz.f32 %f1280, %f121, %f1279; mov.b32 %r794, %f1280; sub.s32 %r795, %r794, %r69; xor.b32 %r796, %r795, %r71; sub.s32 %r797, %r75, %r796; shf.l.wrap.b32 %r798, %r796, %r796, 6; xor.b32 %r799, %r798, %r797; add.s32 %r800, %r796, %r76; sub.s32 %r801, %r76, %r799; shf.l.wrap.b32 %r802, %r799, %r799, 8; xor.b32 %r803, %r802, %r801; add.s32 %r804, %r799, %r800; sub.s32 %r805, %r800, %r803; shf.l.wrap.b32 %r806, %r803, %r803, 16; xor.b32 %r807, %r806, %r805; add.s32 %r808, %r803, %r804; sub.s32 %r809, %r804, %r807; shf.l.wrap.b32 %r810, %r807, %r807, 19; xor.b32 %r811, %r810, %r809; add.s32 %r812, %r807, %r808; sub.s32 %r813, %r808, %r811; shf.l.wrap.b32 %r814, %r811, %r811, 4; add.s32 %r815, %r811, %r812; add.s32 %r816, %r812, %r65; xor.b32 %r817, %r815, %r813; xor.b32 %r818, %r817, %r814; shf.l.wrap.b32 %r819, %r815, %r815, 14; sub.s32 %r820, %r818, %r819; xor.b32 %r821, %r820, %r816; shf.l.wrap.b32 %r822, %r820, %r820, 11; sub.s32 %r823, %r821, %r822; xor.b32 %r824, %r823, %r815; shf.l.wrap.b32 %r825, %r823, %r823, 25; sub.s32 %r826, %r824, %r825; xor.b32 %r827, %r826, %r820; shf.l.wrap.b32 %r828, %r826, %r826, 16; sub.s32 %r829, %r827, %r828; xor.b32 %r830, %r829, %r823; shf.l.wrap.b32 %r831, %r829, %r829, 4; sub.s32 %r832, %r830, %r831; xor.b32 %r833, %r832, %r826; shf.l.wrap.b32 %r834, %r832, %r832, 14; sub.s32 %r835, %r833, %r834; xor.b32 %r836, %r835, %r829; shf.l.wrap.b32 %r837, %r835, %r835, 24; sub.s32 %r838, %r836, %r837; cvt.rn.f32.u32 %f1281, %r838; mul.ftz.f32 %f1282, %f1281, 0f2F800000; add.s32 %r839, %r794, -559038708; add.s32 %r840, %r839, %r75; sub.s32 %r841, %r839, %r77; xor.b32 %r842, %r841, %r78; add.s32 %r843, %r840, %r77; sub.s32 %r844, %r840, %r842; shf.l.wrap.b32 %r845, %r842, %r842, 8; xor.b32 %r846, %r845, %r844; add.s32 %r847, %r842, %r843; sub.s32 %r848, %r843, %r846; shf.l.wrap.b32 %r849, %r846, %r846, 16; xor.b32 %r850, %r849, %r848; add.s32 %r851, %r846, %r847; sub.s32 %r852, %r847, %r850; shf.l.wrap.b32 %r853, %r850, %r850, 19; xor.b32 %r854, %r853, %r852; add.s32 %r855, %r850, %r851; sub.s32 %r856, %r851, %r854; shf.l.wrap.b32 %r857, %r854, %r854, 4; add.s32 %r858, %r854, %r855; add.s32 %r859, %r855, %r69; xor.b32 %r860, %r858, %r856; xor.b32 %r861, %r860, %r857; shf.l.wrap.b32 %r862, %r858, %r858, 14; sub.s32 %r863, %r861, %r862; xor.b32 %r864, %r863, %r859; shf.l.wrap.b32 %r865, %r863, %r863, 11; sub.s32 %r866, %r864, %r865; xor.b32 %r867, %r866, %r858; shf.l.wrap.b32 %r868, %r866, %r866, 25; sub.s32 %r869, %r867, %r868; xor.b32 %r870, %r869, %r863; shf.l.wrap.b32 %r871, %r869, %r869, 16; sub.s32 %r872, %r870, %r871; xor.b32 %r873, %r872, %r866; shf.l.wrap.b32 %r874, %r872, %r872, 4; sub.s32 %r875, %r873, %r874; xor.b32 %r876, %r875, %r869; shf.l.wrap.b32 %r877, %r875, %r875, 14; sub.s32 %r878, %r876, %r877; xor.b32 %r879, %r878, %r872; shf.l.wrap.b32 %r880, %r878, %r878, 24; sub.s32 %r881, %r879, %r880; cvt.rn.f32.u32 %f1283, %r881; mul.ftz.f32 %f1284, %f1283, 0f2F800000; sub.s32 %r882, %r69, %r794; shf.l.wrap.b32 %r883, %r839, %r839, 4; xor.b32 %r884, %r883, %r882; add.s32 %r885, %r839, %r66; sub.s32 %r886, %r66, %r884; shf.l.wrap.b32 %r887, %r884, %r884, 6; xor.b32 %r888, %r887, %r886; add.s32 %r889, %r884, %r885; sub.s32 %r890, %r885, %r888; shf.l.wrap.b32 %r891, %r888, %r888, 8; xor.b32 %r892, %r891, %r890; add.s32 %r893, %r888, %r889; sub.s32 %r894, %r889, %r892; shf.l.wrap.b32 %r895, %r892, %r892, 16; xor.b32 %r896, %r895, %r894; add.s32 %r897, %r892, %r893; sub.s32 %r898, %r893, %r896; shf.l.wrap.b32 %r899, %r896, %r896, 19; xor.b32 %r900, %r899, %r898; add.s32 %r901, %r896, %r897; sub.s32 %r902, %r897, %r900; shf.l.wrap.b32 %r903, %r900, %r900, 4; add.s32 %r904, %r900, %r901; add.s32 %r905, %r901, %r74; xor.b32 %r906, %r904, %r902; xor.b32 %r907, %r906, %r903; shf.l.wrap.b32 %r908, %r904, %r904, 14; sub.s32 %r909, %r907, %r908; xor.b32 %r910, %r909, %r905; shf.l.wrap.b32 %r911, %r909, %r909, 11; sub.s32 %r912, %r910, %r911; xor.b32 %r913, %r912, %r904; shf.l.wrap.b32 %r914, %r912, %r912, 25; sub.s32 %r915, %r913, %r914; xor.b32 %r916, %r915, %r909; shf.l.wrap.b32 %r917, %r915, %r915, 16; sub.s32 %r918, %r916, %r917; xor.b32 %r919, %r918, %r912; shf.l.wrap.b32 %r920, %r918, %r918, 4; sub.s32 %r921, %r919, %r920; xor.b32 %r922, %r921, %r915; shf.l.wrap.b32 %r923, %r921, %r921, 14; sub.s32 %r924, %r922, %r923; xor.b32 %r925, %r924, %r918; shf.l.wrap.b32 %r926, %r924, %r924, 24; sub.s32 %r927, %r925, %r926; cvt.rn.f32.u32 %f1285, %r927; mul.ftz.f32 %f1286, %f1285, 0f2F800000; add.s32 %r928, %r79, %r794; xor.b32 %r929, %r928, %r81; sub.s32 %r930, %r929, %r82; xor.b32 %r931, %r930, %r80; shf.l.wrap.b32 %r932, %r930, %r930, 25; sub.s32 %r933, %r931, %r932; xor.b32 %r934, %r933, %r81; shf.l.wrap.b32 %r935, %r933, %r933, 16; sub.s32 %r936, %r934, %r935; xor.b32 %r937, %r936, %r930; shf.l.wrap.b32 %r938, %r936, %r936, 4; sub.s32 %r939, %r937, %r938; xor.b32 %r940, %r939, %r933; shf.l.wrap.b32 %r941, %r939, %r939, 14; sub.s32 %r942, %r940, %r941; xor.b32 %r943, %r942, %r936; shf.l.wrap.b32 %r944, %r942, %r942, 24; sub.s32 %r945, %r943, %r944; cvt.rn.f32.u32 %f1287, %r945; mul.ftz.f32 %f1288, %f1287, 0f2F800000; fma.rn.ftz.f32 %f1289, %f19, %f1282, %f1279; fma.rn.ftz.f32 %f1290, %f19, %f1284, %f146; fma.rn.ftz.f32 %f1291, %f19, %f1286, %f140; fma.rn.ftz.f32 %f1292, %f19, %f1288, %f134; sub.ftz.f32 %f152, %f1289, %f122; sub.ftz.f32 %f153, %f1290, %f124; sub.ftz.f32 %f154, %f1291, %f126; sub.ftz.f32 %f155, %f1292, %f128; mul.ftz.f32 %f1293, %f152, %f152; fma.rn.ftz.f32 %f1294, %f153, %f153, %f1293; mul.ftz.f32 %f1295, %f154, %f154; fma.rn.ftz.f32 %f1296, %f155, %f155, %f1295; add.ftz.f32 %f156, %f1294, %f1296; setp.geu.ftz.f32 %p28, %f156, %f2863; @%p28 bra $L__BB199_71; mov.f32 %f2859, %f152; mov.f32 %f2860, %f153; mov.f32 %f2861, %f154; mov.f32 %f2862, %f155; mov.f32 %f2863, %f156; $L__BB199_71: add.s32 %r84, %r4334, 1; setp.lt.s32 %p29, %r4334, 1; mov.u32 %r4334, %r84; @%p29 bra $L__BB199_69; add.s32 %r85, %r4333, 1; setp.lt.s32 %p30, %r4333, 1; mov.u32 %r4333, %r85; @%p30 bra $L__BB199_68; add.s32 %r86, %r4332, 1; setp.lt.s32 %p31, %r4332, 1; mov.u32 %r4332, %r86; @%p31 bra $L__BB199_67; add.s32 %r87, %r4331, 1; setp.lt.s32 %p32, %r4331, 1; mov.u32 %r4331, %r87; @%p32 bra $L__BB199_66; mov.u32 %r946, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4335, %r946; $L__BB199_76: cvt.rn.f32.s32 %f163, %r4335; add.ftz.f32 %f1298, %f127, %f163; mov.b32 %r89, %f1298; add.s32 %r90, %r89, -559038708; shf.l.wrap.b32 %r91, %r90, %r90, 4; mov.u32 %r4336, %r946; $L__BB199_77: cvt.rn.f32.s32 %f165, %r4336; add.ftz.f32 %f1299, %f125, %f165; mov.b32 %r93, %f1299; add.s32 %r94, %r93, -559038708; shf.l.wrap.b32 %r95, %r94, %r94, 4; add.s32 %r96, %r94, %r90; mov.u32 %r4337, %r946; $L__BB199_78: cvt.rn.f32.s32 %f167, %r4337; add.ftz.f32 %f1300, %f123, %f167; mov.b32 %r98, %f1300; add.s32 %r99, %r98, -559038708; add.s32 %r100, %r99, %r94; sub.s32 %r950, %r89, %r98; shf.l.wrap.b32 %r951, %r99, %r99, 4; xor.b32 %r101, %r951, %r950; shf.l.wrap.b32 %r102, %r101, %r101, 6; sub.s32 %r952, %r98, %r89; xor.b32 %r953, %r952, %r91; sub.s32 %r954, %r94, %r953; shf.l.wrap.b32 %r955, %r953, %r953, 6; xor.b32 %r956, %r955, %r954; add.s32 %r957, %r953, %r96; sub.s32 %r958, %r96, %r956; shf.l.wrap.b32 %r959, %r956, %r956, 8; xor.b32 %r960, %r959, %r958; add.s32 %r961, %r956, %r957; sub.s32 %r962, %r957, %r960; shf.l.wrap.b32 %r963, %r960, %r960, 16; xor.b32 %r964, %r963, %r962; add.s32 %r965, %r960, %r961; sub.s32 %r966, %r961, %r964; shf.l.wrap.b32 %r967, %r964, %r964, 19; xor.b32 %r968, %r967, %r966; add.s32 %r103, %r964, %r965; sub.s32 %r969, %r965, %r968; shf.l.wrap.b32 %r970, %r968, %r968, 4; add.s32 %r104, %r968, %r103; xor.b32 %r971, %r104, %r969; xor.b32 %r972, %r971, %r970; shf.l.wrap.b32 %r973, %r104, %r104, 14; sub.s32 %r105, %r972, %r973; shf.l.wrap.b32 %r106, %r105, %r105, 11; mov.u32 %r4338, %r946; $L__BB199_79: cvt.rn.f32.s32 %f1301, %r4338; add.ftz.f32 %f1302, %f121, %f1301; mov.b32 %r974, %f1302; sub.s32 %r975, %r974, %r93; xor.b32 %r976, %r975, %r95; sub.s32 %r977, %r99, %r976; shf.l.wrap.b32 %r978, %r976, %r976, 6; xor.b32 %r979, %r978, %r977; add.s32 %r980, %r976, %r100; sub.s32 %r981, %r100, %r979; shf.l.wrap.b32 %r982, %r979, %r979, 8; xor.b32 %r983, %r982, %r981; add.s32 %r984, %r979, %r980; sub.s32 %r985, %r980, %r983; shf.l.wrap.b32 %r986, %r983, %r983, 16; xor.b32 %r987, %r986, %r985; add.s32 %r988, %r983, %r984; sub.s32 %r989, %r984, %r987; shf.l.wrap.b32 %r990, %r987, %r987, 19; xor.b32 %r991, %r990, %r989; add.s32 %r992, %r987, %r988; sub.s32 %r993, %r988, %r991; shf.l.wrap.b32 %r994, %r991, %r991, 4; add.s32 %r995, %r991, %r992; add.s32 %r996, %r992, %r89; xor.b32 %r997, %r995, %r993; xor.b32 %r998, %r997, %r994; shf.l.wrap.b32 %r999, %r995, %r995, 14; sub.s32 %r1000, %r998, %r999; xor.b32 %r1001, %r1000, %r996; shf.l.wrap.b32 %r1002, %r1000, %r1000, 11; sub.s32 %r1003, %r1001, %r1002; xor.b32 %r1004, %r1003, %r995; shf.l.wrap.b32 %r1005, %r1003, %r1003, 25; sub.s32 %r1006, %r1004, %r1005; xor.b32 %r1007, %r1006, %r1000; shf.l.wrap.b32 %r1008, %r1006, %r1006, 16; sub.s32 %r1009, %r1007, %r1008; xor.b32 %r1010, %r1009, %r1003; shf.l.wrap.b32 %r1011, %r1009, %r1009, 4; sub.s32 %r1012, %r1010, %r1011; xor.b32 %r1013, %r1012, %r1006; shf.l.wrap.b32 %r1014, %r1012, %r1012, 14; sub.s32 %r1015, %r1013, %r1014; xor.b32 %r1016, %r1015, %r1009; shf.l.wrap.b32 %r1017, %r1015, %r1015, 24; sub.s32 %r1018, %r1016, %r1017; cvt.rn.f32.u32 %f1303, %r1018; mul.ftz.f32 %f1304, %f1303, 0f2F800000; add.s32 %r1019, %r974, -559038708; add.s32 %r1020, %r1019, %r99; sub.s32 %r1021, %r1019, %r101; xor.b32 %r1022, %r1021, %r102; add.s32 %r1023, %r1020, %r101; sub.s32 %r1024, %r1020, %r1022; shf.l.wrap.b32 %r1025, %r1022, %r1022, 8; xor.b32 %r1026, %r1025, %r1024; add.s32 %r1027, %r1022, %r1023; sub.s32 %r1028, %r1023, %r1026; shf.l.wrap.b32 %r1029, %r1026, %r1026, 16; xor.b32 %r1030, %r1029, %r1028; add.s32 %r1031, %r1026, %r1027; sub.s32 %r1032, %r1027, %r1030; shf.l.wrap.b32 %r1033, %r1030, %r1030, 19; xor.b32 %r1034, %r1033, %r1032; add.s32 %r1035, %r1030, %r1031; sub.s32 %r1036, %r1031, %r1034; shf.l.wrap.b32 %r1037, %r1034, %r1034, 4; add.s32 %r1038, %r1034, %r1035; add.s32 %r1039, %r1035, %r93; xor.b32 %r1040, %r1038, %r1036; xor.b32 %r1041, %r1040, %r1037; shf.l.wrap.b32 %r1042, %r1038, %r1038, 14; sub.s32 %r1043, %r1041, %r1042; xor.b32 %r1044, %r1043, %r1039; shf.l.wrap.b32 %r1045, %r1043, %r1043, 11; sub.s32 %r1046, %r1044, %r1045; xor.b32 %r1047, %r1046, %r1038; shf.l.wrap.b32 %r1048, %r1046, %r1046, 25; sub.s32 %r1049, %r1047, %r1048; xor.b32 %r1050, %r1049, %r1043; shf.l.wrap.b32 %r1051, %r1049, %r1049, 16; sub.s32 %r1052, %r1050, %r1051; xor.b32 %r1053, %r1052, %r1046; shf.l.wrap.b32 %r1054, %r1052, %r1052, 4; sub.s32 %r1055, %r1053, %r1054; xor.b32 %r1056, %r1055, %r1049; shf.l.wrap.b32 %r1057, %r1055, %r1055, 14; sub.s32 %r1058, %r1056, %r1057; xor.b32 %r1059, %r1058, %r1052; shf.l.wrap.b32 %r1060, %r1058, %r1058, 24; sub.s32 %r1061, %r1059, %r1060; cvt.rn.f32.u32 %f1305, %r1061; mul.ftz.f32 %f1306, %f1305, 0f2F800000; sub.s32 %r1062, %r93, %r974; shf.l.wrap.b32 %r1063, %r1019, %r1019, 4; xor.b32 %r1064, %r1063, %r1062; add.s32 %r1065, %r1019, %r90; sub.s32 %r1066, %r90, %r1064; shf.l.wrap.b32 %r1067, %r1064, %r1064, 6; xor.b32 %r1068, %r1067, %r1066; add.s32 %r1069, %r1064, %r1065; sub.s32 %r1070, %r1065, %r1068; shf.l.wrap.b32 %r1071, %r1068, %r1068, 8; xor.b32 %r1072, %r1071, %r1070; add.s32 %r1073, %r1068, %r1069; sub.s32 %r1074, %r1069, %r1072; shf.l.wrap.b32 %r1075, %r1072, %r1072, 16; xor.b32 %r1076, %r1075, %r1074; add.s32 %r1077, %r1072, %r1073; sub.s32 %r1078, %r1073, %r1076; shf.l.wrap.b32 %r1079, %r1076, %r1076, 19; xor.b32 %r1080, %r1079, %r1078; add.s32 %r1081, %r1076, %r1077; sub.s32 %r1082, %r1077, %r1080; shf.l.wrap.b32 %r1083, %r1080, %r1080, 4; add.s32 %r1084, %r1080, %r1081; add.s32 %r1085, %r1081, %r98; xor.b32 %r1086, %r1084, %r1082; xor.b32 %r1087, %r1086, %r1083; shf.l.wrap.b32 %r1088, %r1084, %r1084, 14; sub.s32 %r1089, %r1087, %r1088; xor.b32 %r1090, %r1089, %r1085; shf.l.wrap.b32 %r1091, %r1089, %r1089, 11; sub.s32 %r1092, %r1090, %r1091; xor.b32 %r1093, %r1092, %r1084; shf.l.wrap.b32 %r1094, %r1092, %r1092, 25; sub.s32 %r1095, %r1093, %r1094; xor.b32 %r1096, %r1095, %r1089; shf.l.wrap.b32 %r1097, %r1095, %r1095, 16; sub.s32 %r1098, %r1096, %r1097; xor.b32 %r1099, %r1098, %r1092; shf.l.wrap.b32 %r1100, %r1098, %r1098, 4; sub.s32 %r1101, %r1099, %r1100; xor.b32 %r1102, %r1101, %r1095; shf.l.wrap.b32 %r1103, %r1101, %r1101, 14; sub.s32 %r1104, %r1102, %r1103; xor.b32 %r1105, %r1104, %r1098; shf.l.wrap.b32 %r1106, %r1104, %r1104, 24; sub.s32 %r1107, %r1105, %r1106; cvt.rn.f32.u32 %f1307, %r1107; mul.ftz.f32 %f1308, %f1307, 0f2F800000; add.s32 %r1108, %r103, %r974; xor.b32 %r1109, %r1108, %r105; sub.s32 %r1110, %r1109, %r106; xor.b32 %r1111, %r1110, %r104; shf.l.wrap.b32 %r1112, %r1110, %r1110, 25; sub.s32 %r1113, %r1111, %r1112; xor.b32 %r1114, %r1113, %r105; shf.l.wrap.b32 %r1115, %r1113, %r1113, 16; sub.s32 %r1116, %r1114, %r1115; xor.b32 %r1117, %r1116, %r1110; shf.l.wrap.b32 %r1118, %r1116, %r1116, 4; sub.s32 %r1119, %r1117, %r1118; xor.b32 %r1120, %r1119, %r1113; shf.l.wrap.b32 %r1121, %r1119, %r1119, 14; sub.s32 %r1122, %r1120, %r1121; xor.b32 %r1123, %r1122, %r1116; shf.l.wrap.b32 %r1124, %r1122, %r1122, 24; sub.s32 %r1125, %r1123, %r1124; cvt.rn.f32.u32 %f1309, %r1125; mul.ftz.f32 %f1310, %f1309, 0f2F800000; fma.rn.ftz.f32 %f1311, %f19, %f1304, %f1301; fma.rn.ftz.f32 %f1312, %f19, %f1306, %f167; fma.rn.ftz.f32 %f1313, %f19, %f1308, %f165; fma.rn.ftz.f32 %f1314, %f19, %f1310, %f163; sub.ftz.f32 %f169, %f1311, %f122; sub.ftz.f32 %f170, %f1312, %f124; sub.ftz.f32 %f171, %f1313, %f126; sub.ftz.f32 %f172, %f1314, %f128; sub.ftz.f32 %f173, %f169, %f2859; sub.ftz.f32 %f174, %f170, %f2860; sub.ftz.f32 %f175, %f171, %f2861; sub.ftz.f32 %f176, %f172, %f2862; mul.ftz.f32 %f1315, %f173, %f173; fma.rn.ftz.f32 %f1316, %f174, %f174, %f1315; mul.ftz.f32 %f1317, %f175, %f175; fma.rn.ftz.f32 %f1318, %f176, %f176, %f1317; add.ftz.f32 %f177, %f1316, %f1318; setp.leu.ftz.f32 %p33, %f177, 0f38D1B717; @%p33 bra $L__BB199_81; add.ftz.f32 %f1319, %f2859, %f169; mul.ftz.f32 %f1320, %f1319, 0f3F000000; add.ftz.f32 %f1321, %f2860, %f170; mul.ftz.f32 %f1322, %f1321, 0f3F000000; add.ftz.f32 %f1323, %f2861, %f171; mul.ftz.f32 %f1324, %f1323, 0f3F000000; add.ftz.f32 %f1325, %f2862, %f172; mul.ftz.f32 %f1326, %f1325, 0f3F000000; rsqrt.approx.ftz.f32 %f1327, %f177; mul.ftz.f32 %f1328, %f173, %f1327; mul.ftz.f32 %f1329, %f174, %f1327; mul.ftz.f32 %f1330, %f175, %f1327; mul.ftz.f32 %f1331, %f176, %f1327; mul.ftz.f32 %f1332, %f1322, %f1329; fma.rn.ftz.f32 %f1333, %f1320, %f1328, %f1332; mul.ftz.f32 %f1334, %f1326, %f1331; fma.rn.ftz.f32 %f1335, %f1324, %f1330, %f1334; add.ftz.f32 %f1336, %f1333, %f1335; min.ftz.f32 %f3386, %f3386, %f1336; $L__BB199_81: add.s32 %r108, %r4338, 1; setp.lt.s32 %p34, %r4338, 1; mov.u32 %r4338, %r108; @%p34 bra $L__BB199_79; add.s32 %r109, %r4337, 1; setp.lt.s32 %p35, %r4337, 1; mov.u32 %r4337, %r109; @%p35 bra $L__BB199_78; add.s32 %r110, %r4336, 1; setp.lt.s32 %p36, %r4336, 1; mov.u32 %r4336, %r110; @%p36 bra $L__BB199_77; add.s32 %r111, %r4335, 1; setp.lt.s32 %p37, %r4335, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4335, %r111; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; @%p37 bra $L__BB199_76; bra.uni $L__BB199_136; $L__BB199_376: cvt.rmi.ftz.f32.f32 %f1112, %f21; sub.ftz.f32 %f1113, %f21, %f1112; fma.rn.ftz.f32 %f1114, %f20, 0f40400000, 0f3F800000; mov.f32 %f2658, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r4379, -2; mov.f32 %f3368, %f2658; mov.f32 %f3391, %f2658; mov.f32 %f3390, %f2658; mov.f32 %f3389, %f2658; $L__BB199_377: cvt.rn.f32.s32 %f2660, %r4379; add.ftz.f32 %f2661, %f1112, %f2660; mov.b32 %r386, %f2661; add.s32 %r4018, %r386, -559038720; xor.b32 %r4019, %r4018, -1874868139; add.s32 %r4020, %r4019, 29184895; xor.b32 %r4021, %r4020, -559038720; shf.l.wrap.b32 %r4022, %r4020, %r4020, 25; sub.s32 %r4023, %r4021, %r4022; xor.b32 %r4024, %r4023, -1874868139; shf.l.wrap.b32 %r4025, %r4023, %r4023, 16; sub.s32 %r4026, %r4024, %r4025; xor.b32 %r4027, %r4026, %r4020; shf.l.wrap.b32 %r4028, %r4026, %r4026, 4; sub.s32 %r4029, %r4027, %r4028; xor.b32 %r4030, %r4029, %r4023; shf.l.wrap.b32 %r4031, %r4029, %r4029, 14; sub.s32 %r4032, %r4030, %r4031; xor.b32 %r4033, %r4032, %r4026; shf.l.wrap.b32 %r4034, %r4032, %r4032, 24; sub.s32 %r4035, %r4033, %r4034; cvt.rn.f32.u32 %f2662, %r4035; mul.ftz.f32 %f1120, %f2662, 0f2F800000; fma.rn.ftz.f32 %f1121, %f19, %f1120, %f2660; sub.ftz.f32 %f2663, %f1113, %f1121; abs.ftz.f32 %f1122, %f2663; sub.ftz.f32 %f2664, %f3386, %f1122; mul.ftz.f32 %f2665, %f2664, 0f3F000000; div.approx.ftz.f32 %f2666, %f2665, %f20; add.ftz.f32 %f1123, %f2666, 0f3F000000; setp.lt.ftz.f32 %p207, %f1123, 0f00000000; mov.f32 %f3373, %f2658; @%p207 bra $L__BB199_380; setp.ge.ftz.f32 %p208, %f1123, 0f3F800000; mov.f32 %f3373, 0f3F800000; @%p208 bra $L__BB199_380; mov.f32 %f2668, 0f3F800000; sub.ftz.f32 %f2669, %f2668, 0f00000000; div.approx.ftz.f32 %f2670, %f1123, %f2669; add.ftz.f32 %f2671, %f2670, %f2670; mov.f32 %f2672, 0f40400000; sub.ftz.f32 %f2673, %f2672, %f2671; mul.ftz.f32 %f2674, %f2670, %f2670; mul.ftz.f32 %f3373, %f2674, %f2673; $L__BB199_380: mov.f32 %f2675, 0f3F800000; sub.ftz.f32 %f2676, %f2675, %f3373; mul.ftz.f32 %f2677, %f20, %f3373; mul.ftz.f32 %f2678, %f2677, %f2676; sub.ftz.f32 %f2679, %f1122, %f3386; fma.rn.ftz.f32 %f2680, %f2679, %f3373, %f3386; sub.ftz.f32 %f3386, %f2680, %f2678; div.approx.ftz.f32 %f2681, %f2678, %f1114; add.s32 %r4036, %r386, -559038716; xor.b32 %r4037, %r4036, 1354692725; add.s32 %r4038, %r4037, 138171771; xor.b32 %r4039, %r4038, 506314500; shf.l.wrap.b32 %r4040, %r4038, %r4038, 25; sub.s32 %r4041, %r4039, %r4040; xor.b32 %r4042, %r4041, 1354692725; shf.l.wrap.b32 %r4043, %r4041, %r4041, 16; sub.s32 %r4044, %r4042, %r4043; xor.b32 %r4045, %r4044, %r4038; shf.l.wrap.b32 %r4046, %r4044, %r4044, 4; sub.s32 %r4047, %r4045, %r4046; xor.b32 %r4048, %r4047, %r4041; shf.l.wrap.b32 %r4049, %r4047, %r4047, 14; sub.s32 %r4050, %r4048, %r4049; xor.b32 %r4051, %r4050, %r4044; shf.l.wrap.b32 %r4052, %r4050, %r4050, 24; sub.s32 %r4053, %r4051, %r4052; cvt.rn.f32.u32 %f2682, %r4053; mul.ftz.f32 %f2683, %f2682, 0f2F800000; xor.b32 %r4054, %r4036, 1346304085; add.s32 %r4055, %r4054, 138237311; xor.b32 %r4056, %r4055, 514703108; shf.l.wrap.b32 %r4057, %r4055, %r4055, 25; sub.s32 %r4058, %r4056, %r4057; xor.b32 %r4059, %r4058, 1346304085; shf.l.wrap.b32 %r4060, %r4058, %r4058, 16; sub.s32 %r4061, %r4059, %r4060; xor.b32 %r4062, %r4061, %r4055; shf.l.wrap.b32 %r4063, %r4061, %r4061, 4; sub.s32 %r4064, %r4062, %r4063; xor.b32 %r4065, %r4064, %r4058; shf.l.wrap.b32 %r4066, %r4064, %r4064, 14; sub.s32 %r4067, %r4065, %r4066; xor.b32 %r4068, %r4067, %r4061; shf.l.wrap.b32 %r4069, %r4067, %r4067, 24; sub.s32 %r4070, %r4068, %r4069; cvt.rn.f32.u32 %f2684, %r4070; mul.ftz.f32 %f2685, %f2684, 0f2F800000; sub.ftz.f32 %f2686, %f2683, %f3390; sub.ftz.f32 %f2687, %f2685, %f3389; sub.ftz.f32 %f2688, %f1120, %f3391; fma.rn.ftz.f32 %f2689, %f2688, %f3373, %f3391; fma.rn.ftz.f32 %f2690, %f2686, %f3373, %f3390; fma.rn.ftz.f32 %f2691, %f2687, %f3373, %f3389; sub.ftz.f32 %f3391, %f2689, %f2681; sub.ftz.f32 %f3390, %f2690, %f2681; sub.ftz.f32 %f3389, %f2691, %f2681; sub.ftz.f32 %f2692, %f1121, %f3368; fma.rn.ftz.f32 %f2693, %f2692, %f3373, %f3368; sub.ftz.f32 %f3368, %f2693, %f2681; add.s32 %r4379, %r4379, 1; setp.ne.s32 %p209, %r4379, 3; @%p209 bra $L__BB199_377; add.ftz.f32 %f3381, %f1112, %f3368; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_383; $L__BB199_375: cvt.rmi.ftz.f32.f32 %f2633, %f21; sub.ftz.f32 %f2634, %f21, %f2633; mov.b32 %r3960, %f2633; add.s32 %r3961, %r3960, -559038720; xor.b32 %r3962, %r3961, -1874868139; add.s32 %r3963, %r3962, 29184895; xor.b32 %r3964, %r3963, -559038720; shf.l.wrap.b32 %r3965, %r3963, %r3963, 25; sub.s32 %r3966, %r3964, %r3965; xor.b32 %r3967, %r3966, -1874868139; shf.l.wrap.b32 %r3968, %r3966, %r3966, 16; sub.s32 %r3969, %r3967, %r3968; xor.b32 %r3970, %r3969, %r3963; shf.l.wrap.b32 %r3971, %r3969, %r3969, 4; sub.s32 %r3972, %r3970, %r3971; xor.b32 %r3973, %r3972, %r3966; shf.l.wrap.b32 %r3974, %r3972, %r3972, 14; sub.s32 %r3975, %r3973, %r3974; xor.b32 %r3976, %r3975, %r3969; shf.l.wrap.b32 %r3977, %r3975, %r3975, 24; sub.s32 %r3978, %r3976, %r3977; cvt.rn.f32.u32 %f2635, %r3978; mul.ftz.f32 %f2636, %f2635, 0f2F800000; add.ftz.f32 %f2637, %f2633, 0fBF800000; mov.b32 %r3979, %f2637; add.s32 %r3980, %r3979, -559038720; xor.b32 %r3981, %r3980, -1874868139; add.s32 %r3982, %r3981, 29184895; xor.b32 %r3983, %r3982, -559038720; shf.l.wrap.b32 %r3984, %r3982, %r3982, 25; sub.s32 %r3985, %r3983, %r3984; xor.b32 %r3986, %r3985, -1874868139; shf.l.wrap.b32 %r3987, %r3985, %r3985, 16; sub.s32 %r3988, %r3986, %r3987; xor.b32 %r3989, %r3988, %r3982; shf.l.wrap.b32 %r3990, %r3988, %r3988, 4; sub.s32 %r3991, %r3989, %r3990; xor.b32 %r3992, %r3991, %r3985; shf.l.wrap.b32 %r3993, %r3991, %r3991, 14; sub.s32 %r3994, %r3992, %r3993; xor.b32 %r3995, %r3994, %r3988; shf.l.wrap.b32 %r3996, %r3994, %r3994, 24; sub.s32 %r3997, %r3995, %r3996; cvt.rn.f32.u32 %f2638, %r3997; mul.ftz.f32 %f2639, %f2638, 0f2F800000; fma.rn.ftz.f32 %f2640, %f19, %f2639, 0fBF800000; add.ftz.f32 %f2641, %f2633, 0f3F800000; mov.b32 %r3998, %f2641; add.s32 %r3999, %r3998, -559038720; xor.b32 %r4000, %r3999, -1874868139; add.s32 %r4001, %r4000, 29184895; xor.b32 %r4002, %r4001, -559038720; shf.l.wrap.b32 %r4003, %r4001, %r4001, 25; sub.s32 %r4004, %r4002, %r4003; xor.b32 %r4005, %r4004, -1874868139; shf.l.wrap.b32 %r4006, %r4004, %r4004, 16; sub.s32 %r4007, %r4005, %r4006; xor.b32 %r4008, %r4007, %r4001; shf.l.wrap.b32 %r4009, %r4007, %r4007, 4; sub.s32 %r4010, %r4008, %r4009; xor.b32 %r4011, %r4010, %r4004; shf.l.wrap.b32 %r4012, %r4010, %r4010, 14; sub.s32 %r4013, %r4011, %r4012; xor.b32 %r4014, %r4013, %r4007; shf.l.wrap.b32 %r4015, %r4013, %r4013, 24; sub.s32 %r4016, %r4014, %r4015; cvt.rn.f32.u32 %f2642, %r4016; mul.ftz.f32 %f2643, %f2642, 0f2F800000; fma.rn.ftz.f32 %f2644, %f19, %f2643, 0f3F800000; fma.rn.ftz.f32 %f2645, %f19, %f2636, %f2640; div.approx.ftz.f32 %f2647, %f2645, %f1184; sub.ftz.f32 %f2648, %f2647, %f2634; abs.ftz.f32 %f2649, %f2648; fma.rn.ftz.f32 %f2650, %f19, %f2636, %f2644; div.approx.ftz.f32 %f2651, %f2650, %f1184; sub.ftz.f32 %f2652, %f2651, %f2634; abs.ftz.f32 %f2653, %f2652; min.ftz.f32 %f3386, %f2649, %f2653; mov.f32 %f3381, %f3392; mov.f32 %f3388, %f3392; mov.f32 %f3389, %f3392; mov.f32 %f3390, %f3392; mov.f32 %f3391, %f3392; bra.uni $L__BB199_383; $L__BB199_186: ld.param.u32 %r4315, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f586, %f22; sub.ftz.f32 %f587, %f22, %f586; cvt.rmi.ftz.f32.f32 %f588, %f23; sub.ftz.f32 %f589, %f23, %f588; cvt.rmi.ftz.f32.f32 %f590, %f24; sub.ftz.f32 %f591, %f24, %f590; fma.rn.ftz.f32 %f592, %f20, 0f40400000, 0f3F800000; rcp.approx.ftz.f32 %f593, %f15; mov.f32 %f3389, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r2383, -2; setp.eq.s32 %p91, %r4315, 0; setp.eq.s32 %p92, %r4315, 1; setp.eq.s32 %p93, %r4315, 2; setp.ne.s32 %p94, %r4315, 3; mov.u32 %r4363, %r2383; mov.f32 %f3390, %f3389; mov.f32 %f3391, %f3389; mov.f32 %f3135, %f3389; mov.f32 %f3136, %f3389; mov.f32 %f3137, %f3389; $L__BB199_187: cvt.rn.f32.s32 %f601, %r4363; add.ftz.f32 %f1768, %f590, %f601; mov.b32 %r254, %f1768; add.s32 %r255, %r254, -559038712; add.s32 %r256, %r254, -559038708; shf.l.wrap.b32 %r257, %r256, %r256, 4; mov.u32 %r4364, %r2383; $L__BB199_188: cvt.rn.f32.s32 %f609, %r4364; add.ftz.f32 %f1769, %f588, %f609; mov.b32 %r2386, %f1769; add.s32 %r259, %r2386, -559038712; xor.b32 %r2387, %r259, %r255; shf.l.wrap.b32 %r2388, %r259, %r259, 14; sub.s32 %r260, %r2387, %r2388; shf.l.wrap.b32 %r261, %r260, %r260, 11; add.s32 %r262, %r2386, -559038708; add.s32 %r263, %r262, %r256; mov.u32 %r4365, %r2383; $L__BB199_189: mov.u32 %r264, %r4365; cvt.rn.f32.s32 %f1770, %r264; add.ftz.f32 %f1771, %f586, %f1770; mov.b32 %r2389, %f1771; add.s32 %r2390, %r2389, -559038712; xor.b32 %r2391, %r2390, %r260; sub.s32 %r2392, %r2391, %r261; xor.b32 %r2393, %r2392, %r259; shf.l.wrap.b32 %r2394, %r2392, %r2392, 25; sub.s32 %r2395, %r2393, %r2394; xor.b32 %r2396, %r2395, %r260; shf.l.wrap.b32 %r2397, %r2395, %r2395, 16; sub.s32 %r2398, %r2396, %r2397; xor.b32 %r2399, %r2398, %r2392; shf.l.wrap.b32 %r2400, %r2398, %r2398, 4; sub.s32 %r2401, %r2399, %r2400; xor.b32 %r2402, %r2401, %r2395; shf.l.wrap.b32 %r2403, %r2401, %r2401, 14; sub.s32 %r2404, %r2402, %r2403; xor.b32 %r2405, %r2404, %r2398; shf.l.wrap.b32 %r2406, %r2404, %r2404, 24; sub.s32 %r2407, %r2405, %r2406; cvt.rn.f32.u32 %f1772, %r2407; mul.ftz.f32 %f617, %f1772, 0f2F800000; sub.s32 %r2408, %r2389, %r254; xor.b32 %r2409, %r2408, %r257; sub.s32 %r2410, %r262, %r2409; shf.l.wrap.b32 %r2411, %r2409, %r2409, 6; xor.b32 %r2412, %r2411, %r2410; add.s32 %r2413, %r2409, %r263; sub.s32 %r2414, %r263, %r2412; shf.l.wrap.b32 %r2415, %r2412, %r2412, 8; xor.b32 %r2416, %r2415, %r2414; add.s32 %r2417, %r2412, %r2413; sub.s32 %r2418, %r2413, %r2416; shf.l.wrap.b32 %r2419, %r2416, %r2416, 16; xor.b32 %r2420, %r2419, %r2418; add.s32 %r2421, %r2416, %r2417; sub.s32 %r2422, %r2417, %r2420; shf.l.wrap.b32 %r2423, %r2420, %r2420, 19; xor.b32 %r2424, %r2423, %r2422; add.s32 %r2425, %r2420, %r2421; sub.s32 %r2426, %r2421, %r2424; shf.l.wrap.b32 %r2427, %r2424, %r2424, 4; add.s32 %r2428, %r2424, %r2425; add.s32 %r2429, %r2425, 1065353216; xor.b32 %r2430, %r2428, %r2426; xor.b32 %r2431, %r2430, %r2427; shf.l.wrap.b32 %r2432, %r2428, %r2428, 14; sub.s32 %r2433, %r2431, %r2432; xor.b32 %r2434, %r2433, %r2429; shf.l.wrap.b32 %r2435, %r2433, %r2433, 11; sub.s32 %r2436, %r2434, %r2435; xor.b32 %r2437, %r2436, %r2428; shf.l.wrap.b32 %r2438, %r2436, %r2436, 25; sub.s32 %r2439, %r2437, %r2438; xor.b32 %r2440, %r2439, %r2433; shf.l.wrap.b32 %r2441, %r2439, %r2439, 16; sub.s32 %r2442, %r2440, %r2441; xor.b32 %r2443, %r2442, %r2436; shf.l.wrap.b32 %r2444, %r2442, %r2442, 4; sub.s32 %r2445, %r2443, %r2444; xor.b32 %r2446, %r2445, %r2439; shf.l.wrap.b32 %r2447, %r2445, %r2445, 14; sub.s32 %r2448, %r2446, %r2447; xor.b32 %r2449, %r2448, %r2442; shf.l.wrap.b32 %r2450, %r2448, %r2448, 24; sub.s32 %r2451, %r2449, %r2450; cvt.rn.f32.u32 %f1773, %r2451; mul.ftz.f32 %f618, %f1773, 0f2F800000; add.s32 %r2452, %r2425, 1073741824; xor.b32 %r2453, %r2433, %r2452; sub.s32 %r2454, %r2453, %r2435; xor.b32 %r2455, %r2454, %r2428; shf.l.wrap.b32 %r2456, %r2454, %r2454, 25; sub.s32 %r2457, %r2455, %r2456; xor.b32 %r2458, %r2457, %r2433; shf.l.wrap.b32 %r2459, %r2457, %r2457, 16; sub.s32 %r2460, %r2458, %r2459; xor.b32 %r2461, %r2460, %r2454; shf.l.wrap.b32 %r2462, %r2460, %r2460, 4; sub.s32 %r2463, %r2461, %r2462; xor.b32 %r2464, %r2463, %r2457; shf.l.wrap.b32 %r2465, %r2463, %r2463, 14; sub.s32 %r2466, %r2464, %r2465; xor.b32 %r2467, %r2466, %r2460; shf.l.wrap.b32 %r2468, %r2466, %r2466, 24; sub.s32 %r2469, %r2467, %r2468; cvt.rn.f32.u32 %f1774, %r2469; mul.ftz.f32 %f619, %f1774, 0f2F800000; fma.rn.ftz.f32 %f620, %f19, %f617, %f1770; fma.rn.ftz.f32 %f621, %f19, %f618, %f609; fma.rn.ftz.f32 %f622, %f19, %f619, %f601; @%p91 bra $L__BB199_196; @%p92 bra $L__BB199_195; bra.uni $L__BB199_191; $L__BB199_195: sub.ftz.f32 %f1802, %f620, %f587; abs.ftz.f32 %f1803, %f1802; sub.ftz.f32 %f1804, %f621, %f589; abs.ftz.f32 %f1805, %f1804; add.ftz.f32 %f1806, %f1803, %f1805; sub.ftz.f32 %f1807, %f622, %f591; abs.ftz.f32 %f1808, %f1807; add.ftz.f32 %f3138, %f1806, %f1808; bra.uni $L__BB199_197; $L__BB199_196: sub.ftz.f32 %f1809, %f620, %f587; sub.ftz.f32 %f1810, %f621, %f589; mul.ftz.f32 %f1811, %f1810, %f1810; fma.rn.ftz.f32 %f1812, %f1809, %f1809, %f1811; sub.ftz.f32 %f1813, %f622, %f591; fma.rn.ftz.f32 %f1814, %f1813, %f1813, %f1812; sqrt.approx.ftz.f32 %f3138, %f1814; bra.uni $L__BB199_197; $L__BB199_191: @%p93 bra $L__BB199_194; bra.uni $L__BB199_192; $L__BB199_194: sub.ftz.f32 %f1795, %f620, %f587; abs.ftz.f32 %f1796, %f1795; sub.ftz.f32 %f1797, %f621, %f589; abs.ftz.f32 %f1798, %f1797; sub.ftz.f32 %f1799, %f622, %f591; abs.ftz.f32 %f1800, %f1799; max.ftz.f32 %f1801, %f1798, %f1800; max.ftz.f32 %f3138, %f1796, %f1801; bra.uni $L__BB199_197; $L__BB199_192: mov.f32 %f3138, 0f00000000; @%p94 bra $L__BB199_197; sub.ftz.f32 %f1776, %f620, %f587; abs.ftz.f32 %f1777, %f1776; lg2.approx.ftz.f32 %f1778, %f1777; mul.ftz.f32 %f1779, %f15, %f1778; ex2.approx.ftz.f32 %f1780, %f1779; sub.ftz.f32 %f1781, %f621, %f589; abs.ftz.f32 %f1782, %f1781; lg2.approx.ftz.f32 %f1783, %f1782; mul.ftz.f32 %f1784, %f15, %f1783; ex2.approx.ftz.f32 %f1785, %f1784; add.ftz.f32 %f1786, %f1780, %f1785; sub.ftz.f32 %f1787, %f622, %f591; abs.ftz.f32 %f1788, %f1787; lg2.approx.ftz.f32 %f1789, %f1788; mul.ftz.f32 %f1790, %f15, %f1789; ex2.approx.ftz.f32 %f1791, %f1790; add.ftz.f32 %f1792, %f1786, %f1791; lg2.approx.ftz.f32 %f1793, %f1792; mul.ftz.f32 %f1794, %f593, %f1793; ex2.approx.ftz.f32 %f3138, %f1794; $L__BB199_197: sub.ftz.f32 %f1816, %f3386, %f3138; mul.ftz.f32 %f1817, %f1816, 0f3F000000; div.approx.ftz.f32 %f1818, %f1817, %f20; add.ftz.f32 %f628, %f1818, 0f3F000000; setp.lt.ftz.f32 %p95, %f628, 0f00000000; mov.f32 %f3139, 0f00000000; @%p95 bra $L__BB199_200; setp.ge.ftz.f32 %p96, %f628, 0f3F800000; mov.f32 %f3139, 0f3F800000; @%p96 bra $L__BB199_200; mov.f32 %f1820, 0f3F800000; sub.ftz.f32 %f1821, %f1820, 0f00000000; div.approx.ftz.f32 %f1822, %f628, %f1821; add.ftz.f32 %f1823, %f1822, %f1822; mov.f32 %f1824, 0f40400000; sub.ftz.f32 %f1825, %f1824, %f1823; mul.ftz.f32 %f1826, %f1822, %f1822; mul.ftz.f32 %f3139, %f1826, %f1825; $L__BB199_200: mov.f32 %f1827, 0f3F800000; sub.ftz.f32 %f1828, %f1827, %f3139; mul.ftz.f32 %f1829, %f20, %f3139; mul.ftz.f32 %f1830, %f1829, %f1828; sub.ftz.f32 %f1831, %f3138, %f3386; fma.rn.ftz.f32 %f1832, %f1831, %f3139, %f3386; sub.ftz.f32 %f3386, %f1832, %f1830; div.approx.ftz.f32 %f1833, %f1830, %f592; sub.ftz.f32 %f1834, %f617, %f3391; sub.ftz.f32 %f1835, %f618, %f3390; sub.ftz.f32 %f1836, %f619, %f3389; fma.rn.ftz.f32 %f1837, %f1834, %f3139, %f3391; fma.rn.ftz.f32 %f1838, %f1835, %f3139, %f3390; fma.rn.ftz.f32 %f1839, %f1836, %f3139, %f3389; sub.ftz.f32 %f3391, %f1837, %f1833; sub.ftz.f32 %f3390, %f1838, %f1833; sub.ftz.f32 %f3389, %f1839, %f1833; sub.ftz.f32 %f1840, %f620, %f3137; sub.ftz.f32 %f1841, %f621, %f3136; sub.ftz.f32 %f1842, %f622, %f3135; fma.rn.ftz.f32 %f1843, %f1840, %f3139, %f3137; fma.rn.ftz.f32 %f1844, %f1841, %f3139, %f3136; fma.rn.ftz.f32 %f1845, %f1842, %f3139, %f3135; sub.ftz.f32 %f3137, %f1843, %f1833; sub.ftz.f32 %f3136, %f1844, %f1833; sub.ftz.f32 %f3135, %f1845, %f1833; add.s32 %r4365, %r264, 1; setp.lt.s32 %p97, %r264, 2; @%p97 bra $L__BB199_189; add.s32 %r266, %r4364, 1; setp.lt.s32 %p98, %r4364, 2; mov.u32 %r4364, %r266; @%p98 bra $L__BB199_188; add.s32 %r267, %r4363, 1; setp.lt.s32 %p99, %r4363, 2; mov.u32 %r4363, %r267; @%p99 bra $L__BB199_187; add.ftz.f32 %f3242, %f586, %f3137; add.ftz.f32 %f3241, %f588, %f3136; add.ftz.f32 %f3240, %f590, %f3135; mov.f32 %f3388, 0f00000000; bra.uni $L__BB199_250; $L__BB199_170: cvt.rmi.ftz.f32.f32 %f544, %f22; sub.ftz.f32 %f545, %f22, %f544; cvt.rmi.ftz.f32.f32 %f546, %f23; sub.ftz.f32 %f547, %f23, %f546; cvt.rmi.ftz.f32.f32 %f548, %f24; sub.ftz.f32 %f549, %f24, %f548; mov.f32 %f3111, 0f00000000; mov.u32 %r2209, -1; mov.f32 %f3112, 0f41000000; mov.u32 %r4357, %r2209; mov.f32 %f3110, %f3111; mov.f32 %f3109, %f3111; $L__BB199_171: cvt.rn.f32.s32 %f554, %r4357; add.ftz.f32 %f1709, %f548, %f554; mov.b32 %r224, %f1709; add.s32 %r225, %r224, -559038712; add.s32 %r226, %r224, -559038708; shf.l.wrap.b32 %r227, %r226, %r226, 4; mov.u32 %r4358, %r2209; $L__BB199_172: cvt.rn.f32.s32 %f559, %r4358; add.ftz.f32 %f1710, %f546, %f559; mov.b32 %r2212, %f1710; add.s32 %r229, %r2212, -559038712; xor.b32 %r2213, %r229, %r225; shf.l.wrap.b32 %r2214, %r229, %r229, 14; sub.s32 %r230, %r2213, %r2214; shf.l.wrap.b32 %r231, %r230, %r230, 11; add.s32 %r232, %r2212, -559038708; add.s32 %r233, %r232, %r226; mov.u32 %r4359, %r2209; $L__BB199_173: cvt.rn.f32.s32 %f1711, %r4359; add.ftz.f32 %f1712, %f544, %f1711; mov.b32 %r2215, %f1712; add.s32 %r2216, %r2215, -559038712; xor.b32 %r2217, %r2216, %r230; sub.s32 %r2218, %r2217, %r231; xor.b32 %r2219, %r2218, %r229; shf.l.wrap.b32 %r2220, %r2218, %r2218, 25; sub.s32 %r2221, %r2219, %r2220; xor.b32 %r2222, %r2221, %r230; shf.l.wrap.b32 %r2223, %r2221, %r2221, 16; sub.s32 %r2224, %r2222, %r2223; xor.b32 %r2225, %r2224, %r2218; shf.l.wrap.b32 %r2226, %r2224, %r2224, 4; sub.s32 %r2227, %r2225, %r2226; xor.b32 %r2228, %r2227, %r2221; shf.l.wrap.b32 %r2229, %r2227, %r2227, 14; sub.s32 %r2230, %r2228, %r2229; xor.b32 %r2231, %r2230, %r2224; shf.l.wrap.b32 %r2232, %r2230, %r2230, 24; sub.s32 %r2233, %r2231, %r2232; cvt.rn.f32.u32 %f1713, %r2233; mul.ftz.f32 %f1714, %f1713, 0f2F800000; sub.s32 %r2234, %r2215, %r224; xor.b32 %r2235, %r2234, %r227; sub.s32 %r2236, %r232, %r2235; shf.l.wrap.b32 %r2237, %r2235, %r2235, 6; xor.b32 %r2238, %r2237, %r2236; add.s32 %r2239, %r2235, %r233; sub.s32 %r2240, %r233, %r2238; shf.l.wrap.b32 %r2241, %r2238, %r2238, 8; xor.b32 %r2242, %r2241, %r2240; add.s32 %r2243, %r2238, %r2239; sub.s32 %r2244, %r2239, %r2242; shf.l.wrap.b32 %r2245, %r2242, %r2242, 16; xor.b32 %r2246, %r2245, %r2244; add.s32 %r2247, %r2242, %r2243; sub.s32 %r2248, %r2243, %r2246; shf.l.wrap.b32 %r2249, %r2246, %r2246, 19; xor.b32 %r2250, %r2249, %r2248; add.s32 %r2251, %r2246, %r2247; sub.s32 %r2252, %r2247, %r2250; shf.l.wrap.b32 %r2253, %r2250, %r2250, 4; add.s32 %r2254, %r2250, %r2251; add.s32 %r2255, %r2251, 1065353216; xor.b32 %r2256, %r2254, %r2252; xor.b32 %r2257, %r2256, %r2253; shf.l.wrap.b32 %r2258, %r2254, %r2254, 14; sub.s32 %r2259, %r2257, %r2258; xor.b32 %r2260, %r2259, %r2255; shf.l.wrap.b32 %r2261, %r2259, %r2259, 11; sub.s32 %r2262, %r2260, %r2261; xor.b32 %r2263, %r2262, %r2254; shf.l.wrap.b32 %r2264, %r2262, %r2262, 25; sub.s32 %r2265, %r2263, %r2264; xor.b32 %r2266, %r2265, %r2259; shf.l.wrap.b32 %r2267, %r2265, %r2265, 16; sub.s32 %r2268, %r2266, %r2267; xor.b32 %r2269, %r2268, %r2262; shf.l.wrap.b32 %r2270, %r2268, %r2268, 4; sub.s32 %r2271, %r2269, %r2270; xor.b32 %r2272, %r2271, %r2265; shf.l.wrap.b32 %r2273, %r2271, %r2271, 14; sub.s32 %r2274, %r2272, %r2273; xor.b32 %r2275, %r2274, %r2268; shf.l.wrap.b32 %r2276, %r2274, %r2274, 24; sub.s32 %r2277, %r2275, %r2276; cvt.rn.f32.u32 %f1715, %r2277; mul.ftz.f32 %f1716, %f1715, 0f2F800000; add.s32 %r2278, %r2251, 1073741824; xor.b32 %r2279, %r2259, %r2278; sub.s32 %r2280, %r2279, %r2261; xor.b32 %r2281, %r2280, %r2254; shf.l.wrap.b32 %r2282, %r2280, %r2280, 25; sub.s32 %r2283, %r2281, %r2282; xor.b32 %r2284, %r2283, %r2259; shf.l.wrap.b32 %r2285, %r2283, %r2283, 16; sub.s32 %r2286, %r2284, %r2285; xor.b32 %r2287, %r2286, %r2280; shf.l.wrap.b32 %r2288, %r2286, %r2286, 4; sub.s32 %r2289, %r2287, %r2288; xor.b32 %r2290, %r2289, %r2283; shf.l.wrap.b32 %r2291, %r2289, %r2289, 14; sub.s32 %r2292, %r2290, %r2291; xor.b32 %r2293, %r2292, %r2286; shf.l.wrap.b32 %r2294, %r2292, %r2292, 24; sub.s32 %r2295, %r2293, %r2294; cvt.rn.f32.u32 %f1717, %r2295; mul.ftz.f32 %f1718, %f1717, 0f2F800000; fma.rn.ftz.f32 %f1719, %f19, %f1714, %f1711; fma.rn.ftz.f32 %f1720, %f19, %f1716, %f559; fma.rn.ftz.f32 %f1721, %f19, %f1718, %f554; sub.ftz.f32 %f564, %f1719, %f545; sub.ftz.f32 %f565, %f1720, %f547; sub.ftz.f32 %f566, %f1721, %f549; mul.ftz.f32 %f1722, %f565, %f565; fma.rn.ftz.f32 %f1723, %f564, %f564, %f1722; fma.rn.ftz.f32 %f567, %f566, %f566, %f1723; setp.geu.ftz.f32 %p83, %f567, %f3112; @%p83 bra $L__BB199_175; mov.f32 %f3109, %f564; mov.f32 %f3110, %f565; mov.f32 %f3111, %f566; mov.f32 %f3112, %f567; $L__BB199_175: add.s32 %r235, %r4359, 1; setp.lt.s32 %p84, %r4359, 1; mov.u32 %r4359, %r235; @%p84 bra $L__BB199_173; add.s32 %r236, %r4358, 1; setp.lt.s32 %p85, %r4358, 1; mov.u32 %r4358, %r236; @%p85 bra $L__BB199_172; add.s32 %r237, %r4357, 1; setp.lt.s32 %p86, %r4357, 1; mov.u32 %r4357, %r237; @%p86 bra $L__BB199_171; mov.u32 %r2296, -1; mov.f32 %f3386, 0f41000000; mov.u32 %r4360, %r2296; $L__BB199_179: cvt.rn.f32.s32 %f573, %r4360; add.ftz.f32 %f1725, %f548, %f573; mov.b32 %r239, %f1725; add.s32 %r240, %r239, -559038712; add.s32 %r241, %r239, -559038708; shf.l.wrap.b32 %r242, %r241, %r241, 4; mov.u32 %r4361, %r2296; $L__BB199_180: cvt.rn.f32.s32 %f575, %r4361; add.ftz.f32 %f1726, %f546, %f575; mov.b32 %r2299, %f1726; add.s32 %r244, %r2299, -559038712; xor.b32 %r2300, %r244, %r240; shf.l.wrap.b32 %r2301, %r244, %r244, 14; sub.s32 %r245, %r2300, %r2301; shf.l.wrap.b32 %r246, %r245, %r245, 11; add.s32 %r247, %r2299, -559038708; add.s32 %r248, %r247, %r241; mov.u32 %r4362, %r2296; $L__BB199_181: cvt.rn.f32.s32 %f1727, %r4362; add.ftz.f32 %f1728, %f544, %f1727; mov.b32 %r2302, %f1728; add.s32 %r2303, %r2302, -559038712; xor.b32 %r2304, %r2303, %r245; sub.s32 %r2305, %r2304, %r246; xor.b32 %r2306, %r2305, %r244; shf.l.wrap.b32 %r2307, %r2305, %r2305, 25; sub.s32 %r2308, %r2306, %r2307; xor.b32 %r2309, %r2308, %r245; shf.l.wrap.b32 %r2310, %r2308, %r2308, 16; sub.s32 %r2311, %r2309, %r2310; xor.b32 %r2312, %r2311, %r2305; shf.l.wrap.b32 %r2313, %r2311, %r2311, 4; sub.s32 %r2314, %r2312, %r2313; xor.b32 %r2315, %r2314, %r2308; shf.l.wrap.b32 %r2316, %r2314, %r2314, 14; sub.s32 %r2317, %r2315, %r2316; xor.b32 %r2318, %r2317, %r2311; shf.l.wrap.b32 %r2319, %r2317, %r2317, 24; sub.s32 %r2320, %r2318, %r2319; cvt.rn.f32.u32 %f1729, %r2320; mul.ftz.f32 %f1730, %f1729, 0f2F800000; sub.s32 %r2321, %r2302, %r239; xor.b32 %r2322, %r2321, %r242; sub.s32 %r2323, %r247, %r2322; shf.l.wrap.b32 %r2324, %r2322, %r2322, 6; xor.b32 %r2325, %r2324, %r2323; add.s32 %r2326, %r2322, %r248; sub.s32 %r2327, %r248, %r2325; shf.l.wrap.b32 %r2328, %r2325, %r2325, 8; xor.b32 %r2329, %r2328, %r2327; add.s32 %r2330, %r2325, %r2326; sub.s32 %r2331, %r2326, %r2329; shf.l.wrap.b32 %r2332, %r2329, %r2329, 16; xor.b32 %r2333, %r2332, %r2331; add.s32 %r2334, %r2329, %r2330; sub.s32 %r2335, %r2330, %r2333; shf.l.wrap.b32 %r2336, %r2333, %r2333, 19; xor.b32 %r2337, %r2336, %r2335; add.s32 %r2338, %r2333, %r2334; sub.s32 %r2339, %r2334, %r2337; shf.l.wrap.b32 %r2340, %r2337, %r2337, 4; add.s32 %r2341, %r2337, %r2338; add.s32 %r2342, %r2338, 1065353216; xor.b32 %r2343, %r2341, %r2339; xor.b32 %r2344, %r2343, %r2340; shf.l.wrap.b32 %r2345, %r2341, %r2341, 14; sub.s32 %r2346, %r2344, %r2345; xor.b32 %r2347, %r2346, %r2342; shf.l.wrap.b32 %r2348, %r2346, %r2346, 11; sub.s32 %r2349, %r2347, %r2348; xor.b32 %r2350, %r2349, %r2341; shf.l.wrap.b32 %r2351, %r2349, %r2349, 25; sub.s32 %r2352, %r2350, %r2351; xor.b32 %r2353, %r2352, %r2346; shf.l.wrap.b32 %r2354, %r2352, %r2352, 16; sub.s32 %r2355, %r2353, %r2354; xor.b32 %r2356, %r2355, %r2349; shf.l.wrap.b32 %r2357, %r2355, %r2355, 4; sub.s32 %r2358, %r2356, %r2357; xor.b32 %r2359, %r2358, %r2352; shf.l.wrap.b32 %r2360, %r2358, %r2358, 14; sub.s32 %r2361, %r2359, %r2360; xor.b32 %r2362, %r2361, %r2355; shf.l.wrap.b32 %r2363, %r2361, %r2361, 24; sub.s32 %r2364, %r2362, %r2363; cvt.rn.f32.u32 %f1731, %r2364; mul.ftz.f32 %f1732, %f1731, 0f2F800000; add.s32 %r2365, %r2338, 1073741824; xor.b32 %r2366, %r2346, %r2365; sub.s32 %r2367, %r2366, %r2348; xor.b32 %r2368, %r2367, %r2341; shf.l.wrap.b32 %r2369, %r2367, %r2367, 25; sub.s32 %r2370, %r2368, %r2369; xor.b32 %r2371, %r2370, %r2346; shf.l.wrap.b32 %r2372, %r2370, %r2370, 16; sub.s32 %r2373, %r2371, %r2372; xor.b32 %r2374, %r2373, %r2367; shf.l.wrap.b32 %r2375, %r2373, %r2373, 4; sub.s32 %r2376, %r2374, %r2375; xor.b32 %r2377, %r2376, %r2370; shf.l.wrap.b32 %r2378, %r2376, %r2376, 14; sub.s32 %r2379, %r2377, %r2378; xor.b32 %r2380, %r2379, %r2373; shf.l.wrap.b32 %r2381, %r2379, %r2379, 24; sub.s32 %r2382, %r2380, %r2381; cvt.rn.f32.u32 %f1733, %r2382; mul.ftz.f32 %f1734, %f1733, 0f2F800000; fma.rn.ftz.f32 %f1735, %f19, %f1730, %f1727; fma.rn.ftz.f32 %f1736, %f19, %f1732, %f575; fma.rn.ftz.f32 %f1737, %f19, %f1734, %f573; sub.ftz.f32 %f577, %f1735, %f545; sub.ftz.f32 %f578, %f1736, %f547; sub.ftz.f32 %f579, %f1737, %f549; sub.ftz.f32 %f580, %f577, %f3109; sub.ftz.f32 %f581, %f578, %f3110; sub.ftz.f32 %f582, %f579, %f3111; mul.ftz.f32 %f1738, %f581, %f581; fma.rn.ftz.f32 %f1739, %f580, %f580, %f1738; fma.rn.ftz.f32 %f583, %f582, %f582, %f1739; setp.leu.ftz.f32 %p87, %f583, 0f38D1B717; @%p87 bra $L__BB199_183; add.ftz.f32 %f1740, %f3109, %f577; div.approx.ftz.f32 %f1742, %f1740, %f1184; add.ftz.f32 %f1743, %f3110, %f578; div.approx.ftz.f32 %f1744, %f1743, %f1184; add.ftz.f32 %f1745, %f3111, %f579; div.approx.ftz.f32 %f1746, %f1745, %f1184; sqrt.approx.ftz.f32 %f1747, %f583; div.approx.ftz.f32 %f1748, %f580, %f1747; div.approx.ftz.f32 %f1749, %f581, %f1747; div.approx.ftz.f32 %f1750, %f582, %f1747; mul.ftz.f32 %f1751, %f1744, %f1749; fma.rn.ftz.f32 %f1752, %f1742, %f1748, %f1751; fma.rn.ftz.f32 %f1753, %f1746, %f1750, %f1752; min.ftz.f32 %f3386, %f3386, %f1753; $L__BB199_183: add.s32 %r250, %r4362, 1; setp.lt.s32 %p88, %r4362, 1; mov.u32 %r4362, %r250; @%p88 bra $L__BB199_181; add.s32 %r251, %r4361, 1; setp.lt.s32 %p89, %r4361, 1; mov.u32 %r4361, %r251; @%p89 bra $L__BB199_180; add.s32 %r252, %r4360, 1; setp.lt.s32 %p90, %r4360, 1; mov.f32 %f3388, 0f00000000; mov.u32 %r4360, %r252; mov.f32 %f3389, %f3388; mov.f32 %f3390, %f3388; mov.f32 %f3391, %f3388; mov.f32 %f3240, %f3388; mov.f32 %f3241, %f3388; mov.f32 %f3242, %f3388; @%p90 bra $L__BB199_179; bra.uni $L__BB199_250; $L__BB199_20: ld.param.u32 %r4322, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f1045, %f22; sub.ftz.f32 %f1046, %f22, %f1045; cvt.rmi.ftz.f32.f32 %f1047, %f23; sub.ftz.f32 %f1048, %f23, %f1047; rcp.approx.ftz.f32 %f1049, %f15; add.ftz.f32 %f2462, %f1045, 0fBF800000; mov.b32 %r3684, %f2462; add.s32 %r372, %r3684, -559038716; add.s32 %r373, %r3684, -559038712; add.ftz.f32 %f2463, %f1045, 0f00000000; mov.f32 %f3345, 0f00000000; mov.b32 %r3685, %f2463; add.s32 %r374, %r3685, -559038716; add.s32 %r375, %r3685, -559038712; add.ftz.f32 %f2464, %f1045, 0f3F800000; mov.b32 %r3686, %f2464; add.s32 %r376, %r3686, -559038716; add.s32 %r377, %r3686, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4378, -1; setp.eq.s32 %p177, %r4322, 0; setp.eq.s32 %p188, %r4322, 1; setp.eq.s32 %p189, %r4322, 2; setp.ne.s32 %p190, %r4322, 3; mov.f32 %f3344, %f3345; mov.f32 %f3343, %f3345; mov.f32 %f3342, %f3345; $L__BB199_21: cvt.rn.f32.s32 %f1055, %r4378; add.ftz.f32 %f2465, %f1047, %f1055; mov.b32 %r3687, %f2465; add.s32 %r379, %r3687, -559038716; xor.b32 %r3688, %r379, -559038716; shf.l.wrap.b32 %r3689, %r379, %r379, 14; sub.s32 %r380, %r3688, %r3689; shf.l.wrap.b32 %r381, %r380, %r380, 11; add.s32 %r3690, %r3687, -559038712; xor.b32 %r3691, %r3690, 506314504; shf.l.wrap.b32 %r3692, %r3690, %r3690, 14; sub.s32 %r382, %r3691, %r3692; shf.l.wrap.b32 %r383, %r382, %r382, 11; xor.b32 %r3693, %r372, %r380; sub.s32 %r3694, %r3693, %r381; xor.b32 %r3695, %r3694, %r379; shf.l.wrap.b32 %r3696, %r3694, %r3694, 25; sub.s32 %r3697, %r3695, %r3696; xor.b32 %r3698, %r3697, %r380; shf.l.wrap.b32 %r3699, %r3697, %r3697, 16; sub.s32 %r3700, %r3698, %r3699; xor.b32 %r3701, %r3700, %r3694; shf.l.wrap.b32 %r3702, %r3700, %r3700, 4; sub.s32 %r3703, %r3701, %r3702; xor.b32 %r3704, %r3703, %r3697; shf.l.wrap.b32 %r3705, %r3703, %r3703, 14; sub.s32 %r3706, %r3704, %r3705; xor.b32 %r3707, %r3706, %r3700; shf.l.wrap.b32 %r3708, %r3706, %r3706, 24; sub.s32 %r3709, %r3707, %r3708; cvt.rn.f32.u32 %f2466, %r3709; mul.ftz.f32 %f2467, %f2466, 0f2F800000; xor.b32 %r3710, %r373, %r382; sub.s32 %r3711, %r3710, %r383; xor.b32 %r3712, %r3711, %r3690; shf.l.wrap.b32 %r3713, %r3711, %r3711, 25; sub.s32 %r3714, %r3712, %r3713; xor.b32 %r3715, %r3714, %r382; shf.l.wrap.b32 %r3716, %r3714, %r3714, 16; sub.s32 %r3717, %r3715, %r3716; xor.b32 %r3718, %r3717, %r3711; shf.l.wrap.b32 %r3719, %r3717, %r3717, 4; sub.s32 %r3720, %r3718, %r3719; xor.b32 %r3721, %r3720, %r3714; shf.l.wrap.b32 %r3722, %r3720, %r3720, 14; sub.s32 %r3723, %r3721, %r3722; xor.b32 %r3724, %r3723, %r3717; shf.l.wrap.b32 %r3725, %r3723, %r3723, 24; sub.s32 %r3726, %r3724, %r3725; cvt.rn.f32.u32 %f2468, %r3726; mul.ftz.f32 %f2469, %f2468, 0f2F800000; fma.rn.ftz.f32 %f1056, %f19, %f2467, 0fBF800000; fma.rn.ftz.f32 %f1057, %f19, %f2469, %f1055; @%p177 bra $L__BB199_337; @%p188 bra $L__BB199_336; bra.uni $L__BB199_23; $L__BB199_336: sub.ftz.f32 %f2488, %f1056, %f1046; abs.ftz.f32 %f2489, %f2488; sub.ftz.f32 %f2490, %f1057, %f1048; abs.ftz.f32 %f2491, %f2490; add.ftz.f32 %f3341, %f2489, %f2491; bra.uni $L__BB199_338; $L__BB199_337: sub.ftz.f32 %f2492, %f1056, %f1046; sub.ftz.f32 %f2493, %f1057, %f1048; mul.ftz.f32 %f2494, %f2493, %f2493; fma.rn.ftz.f32 %f2495, %f2492, %f2492, %f2494; sqrt.approx.ftz.f32 %f3341, %f2495; bra.uni $L__BB199_338; $L__BB199_23: @%p189 bra $L__BB199_335; bra.uni $L__BB199_24; $L__BB199_335: sub.ftz.f32 %f2484, %f1056, %f1046; abs.ftz.f32 %f2485, %f2484; sub.ftz.f32 %f2486, %f1057, %f1048; abs.ftz.f32 %f2487, %f2486; max.ftz.f32 %f3341, %f2485, %f2487; bra.uni $L__BB199_338; $L__BB199_24: mov.f32 %f3341, 0f00000000; @%p190 bra $L__BB199_338; sub.ftz.f32 %f2471, %f1056, %f1046; abs.ftz.f32 %f2472, %f2471; lg2.approx.ftz.f32 %f2473, %f2472; mul.ftz.f32 %f2474, %f15, %f2473; ex2.approx.ftz.f32 %f2475, %f2474; sub.ftz.f32 %f2476, %f1057, %f1048; abs.ftz.f32 %f2477, %f2476; lg2.approx.ftz.f32 %f2478, %f2477; mul.ftz.f32 %f2479, %f15, %f2478; ex2.approx.ftz.f32 %f2480, %f2479; add.ftz.f32 %f2481, %f2475, %f2480; lg2.approx.ftz.f32 %f2482, %f2481; mul.ftz.f32 %f2483, %f1049, %f2482; ex2.approx.ftz.f32 %f3341, %f2483; $L__BB199_338: setp.geu.ftz.f32 %p181, %f3341, %f3386; @%p181 bra $L__BB199_340; mov.f32 %f3343, 0fBF800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1057; mov.f32 %f3345, %f1056; mov.f32 %f3386, %f3341; $L__BB199_340: xor.b32 %r3727, %r374, %r380; sub.s32 %r3728, %r3727, %r381; xor.b32 %r3729, %r3728, %r379; shf.l.wrap.b32 %r3730, %r3728, %r3728, 25; sub.s32 %r3731, %r3729, %r3730; xor.b32 %r3732, %r3731, %r380; shf.l.wrap.b32 %r3733, %r3731, %r3731, 16; sub.s32 %r3734, %r3732, %r3733; xor.b32 %r3735, %r3734, %r3728; shf.l.wrap.b32 %r3736, %r3734, %r3734, 4; sub.s32 %r3737, %r3735, %r3736; xor.b32 %r3738, %r3737, %r3731; shf.l.wrap.b32 %r3739, %r3737, %r3737, 14; sub.s32 %r3740, %r3738, %r3739; xor.b32 %r3741, %r3740, %r3734; shf.l.wrap.b32 %r3742, %r3740, %r3740, 24; sub.s32 %r3743, %r3741, %r3742; cvt.rn.f32.u32 %f2497, %r3743; mul.ftz.f32 %f2498, %f2497, 0f2F800000; xor.b32 %r3744, %r375, %r382; sub.s32 %r3745, %r3744, %r383; add.s32 %r3746, %r379, 4; xor.b32 %r3747, %r3745, %r3746; shf.l.wrap.b32 %r3748, %r3745, %r3745, 25; sub.s32 %r3749, %r3747, %r3748; xor.b32 %r3750, %r3749, %r382; shf.l.wrap.b32 %r3751, %r3749, %r3749, 16; sub.s32 %r3752, %r3750, %r3751; xor.b32 %r3753, %r3752, %r3745; shf.l.wrap.b32 %r3754, %r3752, %r3752, 4; sub.s32 %r3755, %r3753, %r3754; xor.b32 %r3756, %r3755, %r3749; shf.l.wrap.b32 %r3757, %r3755, %r3755, 14; sub.s32 %r3758, %r3756, %r3757; xor.b32 %r3759, %r3758, %r3752; shf.l.wrap.b32 %r3760, %r3758, %r3758, 24; sub.s32 %r3761, %r3759, %r3760; cvt.rn.f32.u32 %f2499, %r3761; mul.ftz.f32 %f2500, %f2499, 0f2F800000; fma.rn.ftz.f32 %f1068, %f19, %f2498, 0f00000000; fma.rn.ftz.f32 %f1069, %f19, %f2500, %f1055; @%p177 bra $L__BB199_347; @%p188 bra $L__BB199_346; bra.uni $L__BB199_342; $L__BB199_346: sub.ftz.f32 %f2519, %f1068, %f1046; abs.ftz.f32 %f2520, %f2519; sub.ftz.f32 %f2521, %f1069, %f1048; abs.ftz.f32 %f2522, %f2521; add.ftz.f32 %f3347, %f2520, %f2522; bra.uni $L__BB199_348; $L__BB199_347: sub.ftz.f32 %f2523, %f1068, %f1046; sub.ftz.f32 %f2524, %f1069, %f1048; mul.ftz.f32 %f2525, %f2524, %f2524; fma.rn.ftz.f32 %f2526, %f2523, %f2523, %f2525; sqrt.approx.ftz.f32 %f3347, %f2526; bra.uni $L__BB199_348; $L__BB199_342: @%p189 bra $L__BB199_345; bra.uni $L__BB199_343; $L__BB199_345: sub.ftz.f32 %f2515, %f1068, %f1046; abs.ftz.f32 %f2516, %f2515; sub.ftz.f32 %f2517, %f1069, %f1048; abs.ftz.f32 %f2518, %f2517; max.ftz.f32 %f3347, %f2516, %f2518; bra.uni $L__BB199_348; $L__BB199_343: mov.f32 %f3347, 0f00000000; @%p190 bra $L__BB199_348; sub.ftz.f32 %f2502, %f1068, %f1046; abs.ftz.f32 %f2503, %f2502; lg2.approx.ftz.f32 %f2504, %f2503; mul.ftz.f32 %f2505, %f15, %f2504; ex2.approx.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f1069, %f1048; abs.ftz.f32 %f2508, %f2507; lg2.approx.ftz.f32 %f2509, %f2508; mul.ftz.f32 %f2510, %f15, %f2509; ex2.approx.ftz.f32 %f2511, %f2510; add.ftz.f32 %f2512, %f2506, %f2511; lg2.approx.ftz.f32 %f2513, %f2512; mul.ftz.f32 %f2514, %f1049, %f2513; ex2.approx.ftz.f32 %f3347, %f2514; $L__BB199_348: setp.geu.ftz.f32 %p186, %f3347, %f3386; @%p186 bra $L__BB199_350; mov.f32 %f3343, 0f00000000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1069; mov.f32 %f3345, %f1068; mov.f32 %f3386, %f3347; $L__BB199_350: xor.b32 %r3762, %r376, %r380; sub.s32 %r3763, %r3762, %r381; xor.b32 %r3764, %r3763, %r379; shf.l.wrap.b32 %r3765, %r3763, %r3763, 25; sub.s32 %r3766, %r3764, %r3765; xor.b32 %r3767, %r3766, %r380; shf.l.wrap.b32 %r3768, %r3766, %r3766, 16; sub.s32 %r3769, %r3767, %r3768; xor.b32 %r3770, %r3769, %r3763; shf.l.wrap.b32 %r3771, %r3769, %r3769, 4; sub.s32 %r3772, %r3770, %r3771; xor.b32 %r3773, %r3772, %r3766; shf.l.wrap.b32 %r3774, %r3772, %r3772, 14; sub.s32 %r3775, %r3773, %r3774; xor.b32 %r3776, %r3775, %r3769; shf.l.wrap.b32 %r3777, %r3775, %r3775, 24; sub.s32 %r3778, %r3776, %r3777; cvt.rn.f32.u32 %f2528, %r3778; mul.ftz.f32 %f2529, %f2528, 0f2F800000; xor.b32 %r3779, %r377, %r382; sub.s32 %r3780, %r3779, %r383; xor.b32 %r3782, %r3780, %r3746; shf.l.wrap.b32 %r3783, %r3780, %r3780, 25; sub.s32 %r3784, %r3782, %r3783; xor.b32 %r3785, %r3784, %r382; shf.l.wrap.b32 %r3786, %r3784, %r3784, 16; sub.s32 %r3787, %r3785, %r3786; xor.b32 %r3788, %r3787, %r3780; shf.l.wrap.b32 %r3789, %r3787, %r3787, 4; sub.s32 %r3790, %r3788, %r3789; xor.b32 %r3791, %r3790, %r3784; shf.l.wrap.b32 %r3792, %r3790, %r3790, 14; sub.s32 %r3793, %r3791, %r3792; xor.b32 %r3794, %r3793, %r3787; shf.l.wrap.b32 %r3795, %r3793, %r3793, 24; sub.s32 %r3796, %r3794, %r3795; cvt.rn.f32.u32 %f2530, %r3796; mul.ftz.f32 %f2531, %f2530, 0f2F800000; fma.rn.ftz.f32 %f1080, %f19, %f2529, 0f3F800000; fma.rn.ftz.f32 %f1081, %f19, %f2531, %f1055; @%p177 bra $L__BB199_357; @%p188 bra $L__BB199_356; bra.uni $L__BB199_352; $L__BB199_356: sub.ftz.f32 %f2550, %f1080, %f1046; abs.ftz.f32 %f2551, %f2550; sub.ftz.f32 %f2552, %f1081, %f1048; abs.ftz.f32 %f2553, %f2552; add.ftz.f32 %f3353, %f2551, %f2553; bra.uni $L__BB199_358; $L__BB199_357: sub.ftz.f32 %f2554, %f1080, %f1046; sub.ftz.f32 %f2555, %f1081, %f1048; mul.ftz.f32 %f2556, %f2555, %f2555; fma.rn.ftz.f32 %f2557, %f2554, %f2554, %f2556; sqrt.approx.ftz.f32 %f3353, %f2557; bra.uni $L__BB199_358; $L__BB199_352: @%p189 bra $L__BB199_355; bra.uni $L__BB199_353; $L__BB199_355: sub.ftz.f32 %f2546, %f1080, %f1046; abs.ftz.f32 %f2547, %f2546; sub.ftz.f32 %f2548, %f1081, %f1048; abs.ftz.f32 %f2549, %f2548; max.ftz.f32 %f3353, %f2547, %f2549; bra.uni $L__BB199_358; $L__BB199_353: mov.f32 %f3353, 0f00000000; @%p190 bra $L__BB199_358; sub.ftz.f32 %f2533, %f1080, %f1046; abs.ftz.f32 %f2534, %f2533; lg2.approx.ftz.f32 %f2535, %f2534; mul.ftz.f32 %f2536, %f15, %f2535; ex2.approx.ftz.f32 %f2537, %f2536; sub.ftz.f32 %f2538, %f1081, %f1048; abs.ftz.f32 %f2539, %f2538; lg2.approx.ftz.f32 %f2540, %f2539; mul.ftz.f32 %f2541, %f15, %f2540; ex2.approx.ftz.f32 %f2542, %f2541; add.ftz.f32 %f2543, %f2537, %f2542; lg2.approx.ftz.f32 %f2544, %f2543; mul.ftz.f32 %f2545, %f1049, %f2544; ex2.approx.ftz.f32 %f3353, %f2545; $L__BB199_358: setp.geu.ftz.f32 %p191, %f3353, %f3386; @%p191 bra $L__BB199_360; mov.f32 %f3343, 0f3F800000; mov.f32 %f3342, %f1055; mov.f32 %f3344, %f1081; mov.f32 %f3345, %f1080; mov.f32 %f3386, %f3353; $L__BB199_360: add.s32 %r4378, %r4378, 1; setp.ne.s32 %p192, %r4378, 2; @%p192 bra $L__BB199_21; add.ftz.f32 %f2560, %f1045, %f3343; mov.b32 %r3797, %f2560; add.ftz.f32 %f2561, %f1047, %f3342; mov.b32 %r3798, %f2561; add.s32 %r3799, %r3798, -559038716; add.s32 %r3800, %r3797, -559038716; xor.b32 %r3801, %r3799, -559038716; shf.l.wrap.b32 %r3802, %r3799, %r3799, 14; sub.s32 %r3803, %r3801, %r3802; xor.b32 %r3804, %r3803, %r3800; shf.l.wrap.b32 %r3805, %r3803, %r3803, 11; sub.s32 %r3806, %r3804, %r3805; xor.b32 %r3807, %r3806, %r3799; shf.l.wrap.b32 %r3808, %r3806, %r3806, 25; sub.s32 %r3809, %r3807, %r3808; xor.b32 %r3810, %r3809, %r3803; shf.l.wrap.b32 %r3811, %r3809, %r3809, 16; sub.s32 %r3812, %r3810, %r3811; xor.b32 %r3813, %r3812, %r3806; shf.l.wrap.b32 %r3814, %r3812, %r3812, 4; sub.s32 %r3815, %r3813, %r3814; xor.b32 %r3816, %r3815, %r3809; shf.l.wrap.b32 %r3817, %r3815, %r3815, 14; sub.s32 %r3818, %r3816, %r3817; xor.b32 %r3819, %r3818, %r3812; shf.l.wrap.b32 %r3820, %r3818, %r3818, 24; sub.s32 %r3821, %r3819, %r3820; cvt.rn.f32.u32 %f2562, %r3821; mul.ftz.f32 %f3391, %f2562, 0f2F800000; add.s32 %r3822, %r3798, -559038712; add.s32 %r3823, %r3797, -559038712; xor.b32 %r3824, %r3822, 506314504; shf.l.wrap.b32 %r3825, %r3822, %r3822, 14; sub.s32 %r3826, %r3824, %r3825; xor.b32 %r3827, %r3826, %r3823; shf.l.wrap.b32 %r3828, %r3826, %r3826, 11; sub.s32 %r3829, %r3827, %r3828; xor.b32 %r3830, %r3829, %r3822; shf.l.wrap.b32 %r3831, %r3829, %r3829, 25; sub.s32 %r3832, %r3830, %r3831; xor.b32 %r3833, %r3832, %r3826; shf.l.wrap.b32 %r3834, %r3832, %r3832, 16; sub.s32 %r3835, %r3833, %r3834; xor.b32 %r3836, %r3835, %r3829; shf.l.wrap.b32 %r3837, %r3835, %r3835, 4; sub.s32 %r3838, %r3836, %r3837; xor.b32 %r3839, %r3838, %r3832; shf.l.wrap.b32 %r3840, %r3838, %r3838, 14; sub.s32 %r3841, %r3839, %r3840; xor.b32 %r3842, %r3841, %r3835; shf.l.wrap.b32 %r3843, %r3841, %r3841, 24; sub.s32 %r3844, %r3842, %r3843; cvt.rn.f32.u32 %f2563, %r3844; mul.ftz.f32 %f3390, %f2563, 0f2F800000; xor.b32 %r3845, %r3822, 514703112; sub.s32 %r3846, %r3845, %r3825; xor.b32 %r3847, %r3846, %r3823; shf.l.wrap.b32 %r3848, %r3846, %r3846, 11; sub.s32 %r3849, %r3847, %r3848; xor.b32 %r3850, %r3849, %r3822; shf.l.wrap.b32 %r3851, %r3849, %r3849, 25; sub.s32 %r3852, %r3850, %r3851; xor.b32 %r3853, %r3852, %r3846; shf.l.wrap.b32 %r3854, %r3852, %r3852, 16; sub.s32 %r3855, %r3853, %r3854; xor.b32 %r3856, %r3855, %r3849; shf.l.wrap.b32 %r3857, %r3855, %r3855, 4; sub.s32 %r3858, %r3856, %r3857; xor.b32 %r3859, %r3858, %r3852; shf.l.wrap.b32 %r3860, %r3858, %r3858, 14; sub.s32 %r3861, %r3859, %r3860; xor.b32 %r3862, %r3861, %r3855; shf.l.wrap.b32 %r3863, %r3861, %r3861, 24; sub.s32 %r3864, %r3862, %r3863; cvt.rn.f32.u32 %f2564, %r3864; mul.ftz.f32 %f3389, %f2564, 0f2F800000; add.ftz.f32 %f3365, %f1045, %f3345; add.ftz.f32 %f3364, %f1047, %f3344; mov.f32 %f3388, 0f00000000; $L__BB199_362: setp.eq.ftz.f32 %p193, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; @%p193 bra $L__BB199_364; rcp.approx.ftz.f32 %f2567, %f9; mul.ftz.f32 %f3393, %f2567, %f3364; mul.ftz.f32 %f3394, %f2567, %f3365; $L__BB199_364: mov.f32 %f3392, %f3387; bra.uni $L__BB199_385; $L__BB199_117: ld.param.u32 %r4314, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f385, %f22; sub.ftz.f32 %f386, %f22, %f385; cvt.rmi.ftz.f32.f32 %f387, %f23; sub.ftz.f32 %f388, %f23, %f387; cvt.rmi.ftz.f32.f32 %f389, %f24; sub.ftz.f32 %f390, %f24, %f389; cvt.rmi.ftz.f32.f32 %f391, %f21; sub.ftz.f32 %f392, %f21, %f391; rcp.approx.ftz.f32 %f393, %f15; mov.f32 %f3038, 0f00000000; mov.f32 %f3386, 0f41000000; mov.u32 %r1709, -1; setp.eq.s32 %p58, %r4314, 0; setp.eq.s32 %p59, %r4314, 1; setp.eq.s32 %p60, %r4314, 2; setp.ne.s32 %p61, %r4314, 3; mov.u32 %r4347, %r1709; mov.f32 %f3037, %f3038; mov.f32 %f3036, %f3038; mov.f32 %f3035, %f3038; mov.f32 %f3034, %f3038; mov.f32 %f3033, %f3038; mov.f32 %f3032, %f3038; mov.f32 %f3031, %f3038; $L__BB199_118: cvt.rn.f32.s32 %f403, %r4347; add.ftz.f32 %f1557, %f391, %f403; mov.b32 %r169, %f1557; add.s32 %r170, %r169, -559038708; shf.l.wrap.b32 %r171, %r170, %r170, 4; mov.u32 %r4348, %r1709; $L__BB199_119: cvt.rn.f32.s32 %f413, %r4348; add.ftz.f32 %f1558, %f389, %f413; mov.b32 %r173, %f1558; add.s32 %r174, %r173, -559038708; shf.l.wrap.b32 %r175, %r174, %r174, 4; add.s32 %r176, %r174, %r170; mov.u32 %r4349, %r1709; $L__BB199_120: cvt.rn.f32.s32 %f423, %r4349; add.ftz.f32 %f1559, %f387, %f423; mov.b32 %r178, %f1559; add.s32 %r179, %r178, -559038708; add.s32 %r180, %r179, %r174; sub.s32 %r1713, %r169, %r178; shf.l.wrap.b32 %r1714, %r179, %r179, 4; xor.b32 %r181, %r1714, %r1713; shf.l.wrap.b32 %r182, %r181, %r181, 6; sub.s32 %r1715, %r178, %r169; xor.b32 %r1716, %r1715, %r171; sub.s32 %r1717, %r174, %r1716; shf.l.wrap.b32 %r1718, %r1716, %r1716, 6; xor.b32 %r1719, %r1718, %r1717; add.s32 %r1720, %r1716, %r176; sub.s32 %r1721, %r176, %r1719; shf.l.wrap.b32 %r1722, %r1719, %r1719, 8; xor.b32 %r1723, %r1722, %r1721; add.s32 %r1724, %r1719, %r1720; sub.s32 %r1725, %r1720, %r1723; shf.l.wrap.b32 %r1726, %r1723, %r1723, 16; xor.b32 %r1727, %r1726, %r1725; add.s32 %r1728, %r1723, %r1724; sub.s32 %r1729, %r1724, %r1727; shf.l.wrap.b32 %r1730, %r1727, %r1727, 19; xor.b32 %r1731, %r1730, %r1729; add.s32 %r183, %r1727, %r1728; sub.s32 %r1732, %r1728, %r1731; shf.l.wrap.b32 %r1733, %r1731, %r1731, 4; add.s32 %r184, %r1731, %r183; xor.b32 %r1734, %r184, %r1732; xor.b32 %r1735, %r1734, %r1733; shf.l.wrap.b32 %r1736, %r184, %r184, 14; sub.s32 %r185, %r1735, %r1736; shf.l.wrap.b32 %r186, %r185, %r185, 11; mov.u32 %r4350, %r1709; $L__BB199_121: cvt.rn.f32.s32 %f433, %r4350; add.ftz.f32 %f1560, %f385, %f433; mov.b32 %r1737, %f1560; sub.s32 %r1738, %r1737, %r173; xor.b32 %r1739, %r1738, %r175; sub.s32 %r1740, %r179, %r1739; shf.l.wrap.b32 %r1741, %r1739, %r1739, 6; xor.b32 %r1742, %r1741, %r1740; add.s32 %r1743, %r1739, %r180; sub.s32 %r1744, %r180, %r1742; shf.l.wrap.b32 %r1745, %r1742, %r1742, 8; xor.b32 %r1746, %r1745, %r1744; add.s32 %r1747, %r1742, %r1743; sub.s32 %r1748, %r1743, %r1746; shf.l.wrap.b32 %r1749, %r1746, %r1746, 16; xor.b32 %r1750, %r1749, %r1748; add.s32 %r1751, %r1746, %r1747; sub.s32 %r1752, %r1747, %r1750; shf.l.wrap.b32 %r1753, %r1750, %r1750, 19; xor.b32 %r1754, %r1753, %r1752; add.s32 %r1755, %r1750, %r1751; sub.s32 %r1756, %r1751, %r1754; shf.l.wrap.b32 %r1757, %r1754, %r1754, 4; add.s32 %r1758, %r1754, %r1755; add.s32 %r1759, %r1755, %r169; xor.b32 %r1760, %r1758, %r1756; xor.b32 %r1761, %r1760, %r1757; shf.l.wrap.b32 %r1762, %r1758, %r1758, 14; sub.s32 %r1763, %r1761, %r1762; xor.b32 %r1764, %r1763, %r1759; shf.l.wrap.b32 %r1765, %r1763, %r1763, 11; sub.s32 %r1766, %r1764, %r1765; xor.b32 %r1767, %r1766, %r1758; shf.l.wrap.b32 %r1768, %r1766, %r1766, 25; sub.s32 %r1769, %r1767, %r1768; xor.b32 %r1770, %r1769, %r1763; shf.l.wrap.b32 %r1771, %r1769, %r1769, 16; sub.s32 %r1772, %r1770, %r1771; xor.b32 %r1773, %r1772, %r1766; shf.l.wrap.b32 %r1774, %r1772, %r1772, 4; sub.s32 %r1775, %r1773, %r1774; xor.b32 %r1776, %r1775, %r1769; shf.l.wrap.b32 %r1777, %r1775, %r1775, 14; sub.s32 %r1778, %r1776, %r1777; xor.b32 %r1779, %r1778, %r1772; shf.l.wrap.b32 %r1780, %r1778, %r1778, 24; sub.s32 %r1781, %r1779, %r1780; cvt.rn.f32.u32 %f1561, %r1781; mul.ftz.f32 %f1562, %f1561, 0f2F800000; add.s32 %r1782, %r1737, -559038708; add.s32 %r1783, %r1782, %r179; sub.s32 %r1784, %r1782, %r181; xor.b32 %r1785, %r1784, %r182; add.s32 %r1786, %r1783, %r181; sub.s32 %r1787, %r1783, %r1785; shf.l.wrap.b32 %r1788, %r1785, %r1785, 8; xor.b32 %r1789, %r1788, %r1787; add.s32 %r1790, %r1785, %r1786; sub.s32 %r1791, %r1786, %r1789; shf.l.wrap.b32 %r1792, %r1789, %r1789, 16; xor.b32 %r1793, %r1792, %r1791; add.s32 %r1794, %r1789, %r1790; sub.s32 %r1795, %r1790, %r1793; shf.l.wrap.b32 %r1796, %r1793, %r1793, 19; xor.b32 %r1797, %r1796, %r1795; add.s32 %r1798, %r1793, %r1794; sub.s32 %r1799, %r1794, %r1797; shf.l.wrap.b32 %r1800, %r1797, %r1797, 4; add.s32 %r1801, %r1797, %r1798; add.s32 %r1802, %r1798, %r173; xor.b32 %r1803, %r1801, %r1799; xor.b32 %r1804, %r1803, %r1800; shf.l.wrap.b32 %r1805, %r1801, %r1801, 14; sub.s32 %r1806, %r1804, %r1805; xor.b32 %r1807, %r1806, %r1802; shf.l.wrap.b32 %r1808, %r1806, %r1806, 11; sub.s32 %r1809, %r1807, %r1808; xor.b32 %r1810, %r1809, %r1801; shf.l.wrap.b32 %r1811, %r1809, %r1809, 25; sub.s32 %r1812, %r1810, %r1811; xor.b32 %r1813, %r1812, %r1806; shf.l.wrap.b32 %r1814, %r1812, %r1812, 16; sub.s32 %r1815, %r1813, %r1814; xor.b32 %r1816, %r1815, %r1809; shf.l.wrap.b32 %r1817, %r1815, %r1815, 4; sub.s32 %r1818, %r1816, %r1817; xor.b32 %r1819, %r1818, %r1812; shf.l.wrap.b32 %r1820, %r1818, %r1818, 14; sub.s32 %r1821, %r1819, %r1820; xor.b32 %r1822, %r1821, %r1815; shf.l.wrap.b32 %r1823, %r1821, %r1821, 24; sub.s32 %r1824, %r1822, %r1823; cvt.rn.f32.u32 %f1563, %r1824; mul.ftz.f32 %f1564, %f1563, 0f2F800000; sub.s32 %r1825, %r173, %r1737; shf.l.wrap.b32 %r1826, %r1782, %r1782, 4; xor.b32 %r1827, %r1826, %r1825; add.s32 %r1828, %r1782, %r170; sub.s32 %r1829, %r170, %r1827; shf.l.wrap.b32 %r1830, %r1827, %r1827, 6; xor.b32 %r1831, %r1830, %r1829; add.s32 %r1832, %r1827, %r1828; sub.s32 %r1833, %r1828, %r1831; shf.l.wrap.b32 %r1834, %r1831, %r1831, 8; xor.b32 %r1835, %r1834, %r1833; add.s32 %r1836, %r1831, %r1832; sub.s32 %r1837, %r1832, %r1835; shf.l.wrap.b32 %r1838, %r1835, %r1835, 16; xor.b32 %r1839, %r1838, %r1837; add.s32 %r1840, %r1835, %r1836; sub.s32 %r1841, %r1836, %r1839; shf.l.wrap.b32 %r1842, %r1839, %r1839, 19; xor.b32 %r1843, %r1842, %r1841; add.s32 %r1844, %r1839, %r1840; sub.s32 %r1845, %r1840, %r1843; shf.l.wrap.b32 %r1846, %r1843, %r1843, 4; add.s32 %r1847, %r1843, %r1844; add.s32 %r1848, %r1844, %r178; xor.b32 %r1849, %r1847, %r1845; xor.b32 %r1850, %r1849, %r1846; shf.l.wrap.b32 %r1851, %r1847, %r1847, 14; sub.s32 %r1852, %r1850, %r1851; xor.b32 %r1853, %r1852, %r1848; shf.l.wrap.b32 %r1854, %r1852, %r1852, 11; sub.s32 %r1855, %r1853, %r1854; xor.b32 %r1856, %r1855, %r1847; shf.l.wrap.b32 %r1857, %r1855, %r1855, 25; sub.s32 %r1858, %r1856, %r1857; xor.b32 %r1859, %r1858, %r1852; shf.l.wrap.b32 %r1860, %r1858, %r1858, 16; sub.s32 %r1861, %r1859, %r1860; xor.b32 %r1862, %r1861, %r1855; shf.l.wrap.b32 %r1863, %r1861, %r1861, 4; sub.s32 %r1864, %r1862, %r1863; xor.b32 %r1865, %r1864, %r1858; shf.l.wrap.b32 %r1866, %r1864, %r1864, 14; sub.s32 %r1867, %r1865, %r1866; xor.b32 %r1868, %r1867, %r1861; shf.l.wrap.b32 %r1869, %r1867, %r1867, 24; sub.s32 %r1870, %r1868, %r1869; cvt.rn.f32.u32 %f1565, %r1870; mul.ftz.f32 %f1566, %f1565, 0f2F800000; add.s32 %r1871, %r183, %r1737; xor.b32 %r1872, %r1871, %r185; sub.s32 %r1873, %r1872, %r186; xor.b32 %r1874, %r1873, %r184; shf.l.wrap.b32 %r1875, %r1873, %r1873, 25; sub.s32 %r1876, %r1874, %r1875; xor.b32 %r1877, %r1876, %r185; shf.l.wrap.b32 %r1878, %r1876, %r1876, 16; sub.s32 %r1879, %r1877, %r1878; xor.b32 %r1880, %r1879, %r1873; shf.l.wrap.b32 %r1881, %r1879, %r1879, 4; sub.s32 %r1882, %r1880, %r1881; xor.b32 %r1883, %r1882, %r1876; shf.l.wrap.b32 %r1884, %r1882, %r1882, 14; sub.s32 %r1885, %r1883, %r1884; xor.b32 %r1886, %r1885, %r1879; shf.l.wrap.b32 %r1887, %r1885, %r1885, 24; sub.s32 %r1888, %r1886, %r1887; cvt.rn.f32.u32 %f1567, %r1888; mul.ftz.f32 %f1568, %f1567, 0f2F800000; fma.rn.ftz.f32 %f434, %f19, %f1562, %f433; fma.rn.ftz.f32 %f435, %f19, %f1564, %f423; fma.rn.ftz.f32 %f436, %f19, %f1566, %f413; fma.rn.ftz.f32 %f437, %f19, %f1568, %f403; @%p58 bra $L__BB199_128; @%p59 bra $L__BB199_127; bra.uni $L__BB199_123; $L__BB199_127: sub.ftz.f32 %f1605, %f434, %f386; abs.ftz.f32 %f1606, %f1605; sub.ftz.f32 %f1607, %f435, %f388; abs.ftz.f32 %f1608, %f1607; add.ftz.f32 %f1609, %f1606, %f1608; sub.ftz.f32 %f1610, %f436, %f390; abs.ftz.f32 %f1611, %f1610; add.ftz.f32 %f1612, %f1609, %f1611; sub.ftz.f32 %f1613, %f437, %f392; abs.ftz.f32 %f1614, %f1613; add.ftz.f32 %f3030, %f1612, %f1614; bra.uni $L__BB199_129; $L__BB199_128: sub.ftz.f32 %f1615, %f434, %f386; mul.ftz.f32 %f1616, %f1615, %f1615; sub.ftz.f32 %f1617, %f435, %f388; fma.rn.ftz.f32 %f1618, %f1617, %f1617, %f1616; sub.ftz.f32 %f1619, %f436, %f390; mul.ftz.f32 %f1620, %f1619, %f1619; sub.ftz.f32 %f1621, %f437, %f392; fma.rn.ftz.f32 %f1622, %f1621, %f1621, %f1620; add.ftz.f32 %f1623, %f1618, %f1622; sqrt.approx.ftz.f32 %f3030, %f1623; bra.uni $L__BB199_129; $L__BB199_123: @%p60 bra $L__BB199_126; bra.uni $L__BB199_124; $L__BB199_126: sub.ftz.f32 %f1595, %f434, %f386; abs.ftz.f32 %f1596, %f1595; sub.ftz.f32 %f1597, %f435, %f388; abs.ftz.f32 %f1598, %f1597; sub.ftz.f32 %f1599, %f436, %f390; abs.ftz.f32 %f1600, %f1599; sub.ftz.f32 %f1601, %f437, %f392; abs.ftz.f32 %f1602, %f1601; max.ftz.f32 %f1603, %f1600, %f1602; max.ftz.f32 %f1604, %f1598, %f1603; max.ftz.f32 %f3030, %f1596, %f1604; bra.uni $L__BB199_129; $L__BB199_124: mov.f32 %f3030, 0f00000000; @%p61 bra $L__BB199_129; sub.ftz.f32 %f1570, %f434, %f386; abs.ftz.f32 %f1571, %f1570; lg2.approx.ftz.f32 %f1572, %f1571; mul.ftz.f32 %f1573, %f15, %f1572; ex2.approx.ftz.f32 %f1574, %f1573; sub.ftz.f32 %f1575, %f435, %f388; abs.ftz.f32 %f1576, %f1575; lg2.approx.ftz.f32 %f1577, %f1576; mul.ftz.f32 %f1578, %f15, %f1577; ex2.approx.ftz.f32 %f1579, %f1578; add.ftz.f32 %f1580, %f1574, %f1579; sub.ftz.f32 %f1581, %f436, %f390; abs.ftz.f32 %f1582, %f1581; lg2.approx.ftz.f32 %f1583, %f1582; mul.ftz.f32 %f1584, %f15, %f1583; ex2.approx.ftz.f32 %f1585, %f1584; add.ftz.f32 %f1586, %f1580, %f1585; sub.ftz.f32 %f1587, %f437, %f392; abs.ftz.f32 %f1588, %f1587; lg2.approx.ftz.f32 %f1589, %f1588; mul.ftz.f32 %f1590, %f15, %f1589; ex2.approx.ftz.f32 %f1591, %f1590; add.ftz.f32 %f1592, %f1586, %f1591; lg2.approx.ftz.f32 %f1593, %f1592; mul.ftz.f32 %f1594, %f393, %f1593; ex2.approx.ftz.f32 %f3030, %f1594; $L__BB199_129: setp.geu.ftz.f32 %p62, %f3030, %f3386; @%p62 bra $L__BB199_131; mov.f32 %f3031, %f434; mov.f32 %f3032, %f435; mov.f32 %f3033, %f436; mov.f32 %f3034, %f437; mov.f32 %f3035, %f433; mov.f32 %f3036, %f423; mov.f32 %f3037, %f413; mov.f32 %f3038, %f403; mov.f32 %f3386, %f3030; $L__BB199_131: add.s32 %r188, %r4350, 1; setp.lt.s32 %p63, %r4350, 1; mov.u32 %r4350, %r188; @%p63 bra $L__BB199_121; add.s32 %r189, %r4349, 1; setp.lt.s32 %p64, %r4349, 1; mov.u32 %r4349, %r189; @%p64 bra $L__BB199_120; add.s32 %r190, %r4348, 1; setp.lt.s32 %p65, %r4348, 1; mov.u32 %r4348, %r190; @%p65 bra $L__BB199_119; add.s32 %r191, %r4347, 1; setp.lt.s32 %p66, %r4347, 1; mov.u32 %r4347, %r191; @%p66 bra $L__BB199_118; add.ftz.f32 %f1625, %f385, %f3035; mov.b32 %r1889, %f1625; add.ftz.f32 %f1626, %f387, %f3036; mov.b32 %r1890, %f1626; add.ftz.f32 %f1627, %f389, %f3037; mov.b32 %r1891, %f1627; add.ftz.f32 %f1628, %f391, %f3038; mov.b32 %r1892, %f1628; add.s32 %r1893, %r1890, -559038708; add.s32 %r1894, %r1891, -559038708; sub.s32 %r1895, %r1889, %r1891; shf.l.wrap.b32 %r1896, %r1894, %r1894, 4; xor.b32 %r1897, %r1896, %r1895; add.s32 %r1898, %r1894, %r1893; sub.s32 %r1899, %r1893, %r1897; shf.l.wrap.b32 %r1900, %r1897, %r1897, 6; xor.b32 %r1901, %r1900, %r1899; add.s32 %r1902, %r1897, %r1898; sub.s32 %r1903, %r1898, %r1901; shf.l.wrap.b32 %r1904, %r1901, %r1901, 8; xor.b32 %r1905, %r1904, %r1903; add.s32 %r1906, %r1901, %r1902; sub.s32 %r1907, %r1902, %r1905; shf.l.wrap.b32 %r1908, %r1905, %r1905, 16; xor.b32 %r1909, %r1908, %r1907; add.s32 %r1910, %r1905, %r1906; sub.s32 %r1911, %r1906, %r1909; shf.l.wrap.b32 %r1912, %r1909, %r1909, 19; xor.b32 %r1913, %r1912, %r1911; add.s32 %r1914, %r1909, %r1910; sub.s32 %r1915, %r1910, %r1913; shf.l.wrap.b32 %r1916, %r1913, %r1913, 4; add.s32 %r1917, %r1913, %r1914; add.s32 %r1918, %r1914, %r1892; xor.b32 %r1919, %r1917, %r1915; xor.b32 %r1920, %r1919, %r1916; shf.l.wrap.b32 %r1921, %r1917, %r1917, 14; sub.s32 %r1922, %r1920, %r1921; xor.b32 %r1923, %r1922, %r1918; shf.l.wrap.b32 %r1924, %r1922, %r1922, 11; sub.s32 %r1925, %r1923, %r1924; xor.b32 %r1926, %r1925, %r1917; shf.l.wrap.b32 %r1927, %r1925, %r1925, 25; sub.s32 %r1928, %r1926, %r1927; xor.b32 %r1929, %r1928, %r1922; shf.l.wrap.b32 %r1930, %r1928, %r1928, 16; sub.s32 %r1931, %r1929, %r1930; xor.b32 %r1932, %r1931, %r1925; shf.l.wrap.b32 %r1933, %r1931, %r1931, 4; sub.s32 %r1934, %r1932, %r1933; xor.b32 %r1935, %r1934, %r1928; shf.l.wrap.b32 %r1936, %r1934, %r1934, 14; sub.s32 %r1937, %r1935, %r1936; xor.b32 %r1938, %r1937, %r1931; shf.l.wrap.b32 %r1939, %r1937, %r1937, 24; sub.s32 %r1940, %r1938, %r1939; cvt.rn.f32.u32 %f1629, %r1940; mul.ftz.f32 %f3391, %f1629, 0f2F800000; add.s32 %r1941, %r1889, -559038708; add.s32 %r1942, %r1892, -559038708; sub.s32 %r1943, %r1891, %r1892; shf.l.wrap.b32 %r1944, %r1942, %r1942, 4; xor.b32 %r1945, %r1944, %r1943; add.s32 %r1946, %r1942, %r1941; sub.s32 %r1947, %r1941, %r1945; shf.l.wrap.b32 %r1948, %r1945, %r1945, 6; xor.b32 %r1949, %r1948, %r1947; add.s32 %r1950, %r1945, %r1946; sub.s32 %r1951, %r1946, %r1949; shf.l.wrap.b32 %r1952, %r1949, %r1949, 8; xor.b32 %r1953, %r1952, %r1951; add.s32 %r1954, %r1949, %r1950; sub.s32 %r1955, %r1950, %r1953; shf.l.wrap.b32 %r1956, %r1953, %r1953, 16; xor.b32 %r1957, %r1956, %r1955; add.s32 %r1958, %r1953, %r1954; sub.s32 %r1959, %r1954, %r1957; shf.l.wrap.b32 %r1960, %r1957, %r1957, 19; xor.b32 %r1961, %r1960, %r1959; add.s32 %r1962, %r1957, %r1958; sub.s32 %r1963, %r1958, %r1961; shf.l.wrap.b32 %r1964, %r1961, %r1961, 4; add.s32 %r1965, %r1961, %r1962; add.s32 %r1966, %r1962, %r1890; xor.b32 %r1967, %r1965, %r1963; xor.b32 %r1968, %r1967, %r1964; shf.l.wrap.b32 %r1969, %r1965, %r1965, 14; sub.s32 %r1970, %r1968, %r1969; xor.b32 %r1971, %r1970, %r1966; shf.l.wrap.b32 %r1972, %r1970, %r1970, 11; sub.s32 %r1973, %r1971, %r1972; xor.b32 %r1974, %r1973, %r1965; shf.l.wrap.b32 %r1975, %r1973, %r1973, 25; sub.s32 %r1976, %r1974, %r1975; xor.b32 %r1977, %r1976, %r1970; shf.l.wrap.b32 %r1978, %r1976, %r1976, 16; sub.s32 %r1979, %r1977, %r1978; xor.b32 %r1980, %r1979, %r1973; shf.l.wrap.b32 %r1981, %r1979, %r1979, 4; sub.s32 %r1982, %r1980, %r1981; xor.b32 %r1983, %r1982, %r1976; shf.l.wrap.b32 %r1984, %r1982, %r1982, 14; sub.s32 %r1985, %r1983, %r1984; xor.b32 %r1986, %r1985, %r1979; shf.l.wrap.b32 %r1987, %r1985, %r1985, 24; sub.s32 %r1988, %r1986, %r1987; cvt.rn.f32.u32 %f1630, %r1988; mul.ftz.f32 %f3390, %f1630, 0f2F800000; sub.s32 %r1989, %r1892, %r1890; shf.l.wrap.b32 %r1990, %r1893, %r1893, 4; xor.b32 %r1991, %r1990, %r1989; sub.s32 %r1992, %r1894, %r1991; shf.l.wrap.b32 %r1993, %r1991, %r1991, 6; xor.b32 %r1994, %r1993, %r1992; add.s32 %r1995, %r1991, %r1898; sub.s32 %r1996, %r1898, %r1994; shf.l.wrap.b32 %r1997, %r1994, %r1994, 8; xor.b32 %r1998, %r1997, %r1996; add.s32 %r1999, %r1994, %r1995; sub.s32 %r2000, %r1995, %r1998; shf.l.wrap.b32 %r2001, %r1998, %r1998, 16; xor.b32 %r2002, %r2001, %r2000; add.s32 %r2003, %r1998, %r1999; sub.s32 %r2004, %r1999, %r2002; shf.l.wrap.b32 %r2005, %r2002, %r2002, 19; xor.b32 %r2006, %r2005, %r2004; add.s32 %r2007, %r2002, %r2003; sub.s32 %r2008, %r2003, %r2006; shf.l.wrap.b32 %r2009, %r2006, %r2006, 4; add.s32 %r2010, %r2006, %r2007; add.s32 %r2011, %r2007, %r1889; xor.b32 %r2012, %r2010, %r2008; xor.b32 %r2013, %r2012, %r2009; shf.l.wrap.b32 %r2014, %r2010, %r2010, 14; sub.s32 %r2015, %r2013, %r2014; xor.b32 %r2016, %r2015, %r2011; shf.l.wrap.b32 %r2017, %r2015, %r2015, 11; sub.s32 %r2018, %r2016, %r2017; xor.b32 %r2019, %r2018, %r2010; shf.l.wrap.b32 %r2020, %r2018, %r2018, 25; sub.s32 %r2021, %r2019, %r2020; xor.b32 %r2022, %r2021, %r2015; shf.l.wrap.b32 %r2023, %r2021, %r2021, 16; sub.s32 %r2024, %r2022, %r2023; xor.b32 %r2025, %r2024, %r2018; shf.l.wrap.b32 %r2026, %r2024, %r2024, 4; sub.s32 %r2027, %r2025, %r2026; xor.b32 %r2028, %r2027, %r2021; shf.l.wrap.b32 %r2029, %r2027, %r2027, 14; sub.s32 %r2030, %r2028, %r2029; xor.b32 %r2031, %r2030, %r2024; shf.l.wrap.b32 %r2032, %r2030, %r2030, 24; sub.s32 %r2033, %r2031, %r2032; cvt.rn.f32.u32 %f1631, %r2033; mul.ftz.f32 %f3389, %f1631, 0f2F800000; add.ftz.f32 %f3048, %f385, %f3031; add.ftz.f32 %f3047, %f387, %f3032; add.ftz.f32 %f3046, %f389, %f3033; add.ftz.f32 %f3045, %f391, %f3034; mov.f32 %f3388, 0f00000000; $L__BB199_136: setp.eq.ftz.f32 %p67, %f9, 0f00000000; mov.f32 %f3393, %f3394; mov.f32 %f3392, %f3394; mov.f32 %f3387, %f3394; @%p67 bra $L__BB199_385; rcp.approx.ftz.f32 %f1636, %f9; mul.ftz.f32 %f3387, %f1636, %f3045; mul.ftz.f32 %f3392, %f1636, %f3046; mul.ftz.f32 %f3393, %f1636, %f3047; mul.ftz.f32 %f3394, %f1636, %f3048; bra.uni $L__BB199_385; $L__BB199_382: cvt.rmi.ftz.f32.f32 %f2724, %f21; sub.ftz.f32 %f2725, %f21, %f2724; add.ftz.f32 %f2726, %f2724, 0fBF800000; mov.b32 %r4182, %f2726; add.s32 %r4183, %r4182, -559038720; xor.b32 %r4184, %r4183, -1874868139; add.s32 %r4185, %r4184, 29184895; xor.b32 %r4186, %r4185, -559038720; shf.l.wrap.b32 %r4187, %r4185, %r4185, 25; sub.s32 %r4188, %r4186, %r4187; xor.b32 %r4189, %r4188, -1874868139; shf.l.wrap.b32 %r4190, %r4188, %r4188, 16; sub.s32 %r4191, %r4189, %r4190; xor.b32 %r4192, %r4191, %r4185; shf.l.wrap.b32 %r4193, %r4191, %r4191, 4; sub.s32 %r4194, %r4192, %r4193; xor.b32 %r4195, %r4194, %r4188; shf.l.wrap.b32 %r4196, %r4194, %r4194, 14; sub.s32 %r4197, %r4195, %r4196; xor.b32 %r4198, %r4197, %r4191; shf.l.wrap.b32 %r4199, %r4197, %r4197, 24; sub.s32 %r4200, %r4198, %r4199; cvt.rn.f32.u32 %f2727, %r4200; mul.ftz.f32 %f2728, %f2727, 0f2F800000; fma.rn.ftz.f32 %f2729, %f19, %f2728, 0fBF800000; sub.ftz.f32 %f2730, %f2725, %f2729; abs.ftz.f32 %f2731, %f2730; setp.lt.ftz.f32 %p215, %f2731, 0f41000000; selp.f32 %f2732, %f2731, 0f41000000, %p215; selp.f32 %f2733, %f2729, 0f00000000, %p215; add.ftz.f32 %f2734, %f2724, 0f00000000; mov.b32 %r4201, %f2734; add.s32 %r4202, %r4201, -559038720; xor.b32 %r4203, %r4202, -1874868139; add.s32 %r4204, %r4203, 29184895; xor.b32 %r4205, %r4204, -559038720; shf.l.wrap.b32 %r4206, %r4204, %r4204, 25; sub.s32 %r4207, %r4205, %r4206; xor.b32 %r4208, %r4207, -1874868139; shf.l.wrap.b32 %r4209, %r4207, %r4207, 16; sub.s32 %r4210, %r4208, %r4209; xor.b32 %r4211, %r4210, %r4204; shf.l.wrap.b32 %r4212, %r4210, %r4210, 4; sub.s32 %r4213, %r4211, %r4212; xor.b32 %r4214, %r4213, %r4207; shf.l.wrap.b32 %r4215, %r4213, %r4213, 14; sub.s32 %r4216, %r4214, %r4215; xor.b32 %r4217, %r4216, %r4210; shf.l.wrap.b32 %r4218, %r4216, %r4216, 24; sub.s32 %r4219, %r4217, %r4218; cvt.rn.f32.u32 %f2735, %r4219; mul.ftz.f32 %f2736, %f2735, 0f2F800000; fma.rn.ftz.f32 %f2737, %f19, %f2736, 0f00000000; sub.ftz.f32 %f2738, %f2725, %f2737; abs.ftz.f32 %f2739, %f2738; setp.lt.ftz.f32 %p216, %f2739, %f2732; selp.f32 %f2740, %f2739, %f2732, %p216; setp.geu.ftz.f32 %p217, %f2739, %f2732; and.pred %p218, %p215, %p217; selp.f32 %f2741, 0fBF800000, 0f00000000, %p218; selp.f32 %f2742, %f2737, %f2733, %p216; add.ftz.f32 %f2743, %f2724, 0f3F800000; mov.b32 %r4220, %f2743; add.s32 %r4221, %r4220, -559038720; xor.b32 %r4222, %r4221, -1874868139; add.s32 %r4223, %r4222, 29184895; xor.b32 %r4224, %r4223, -559038720; shf.l.wrap.b32 %r4225, %r4223, %r4223, 25; sub.s32 %r4226, %r4224, %r4225; xor.b32 %r4227, %r4226, -1874868139; shf.l.wrap.b32 %r4228, %r4226, %r4226, 16; sub.s32 %r4229, %r4227, %r4228; xor.b32 %r4230, %r4229, %r4223; shf.l.wrap.b32 %r4231, %r4229, %r4229, 4; sub.s32 %r4232, %r4230, %r4231; xor.b32 %r4233, %r4232, %r4226; shf.l.wrap.b32 %r4234, %r4232, %r4232, 14; sub.s32 %r4235, %r4233, %r4234; xor.b32 %r4236, %r4235, %r4229; shf.l.wrap.b32 %r4237, %r4235, %r4235, 24; sub.s32 %r4238, %r4236, %r4237; cvt.rn.f32.u32 %f2744, %r4238; mul.ftz.f32 %f2745, %f2744, 0f2F800000; fma.rn.ftz.f32 %f2746, %f19, %f2745, 0f3F800000; sub.ftz.f32 %f2747, %f2725, %f2746; abs.ftz.f32 %f2748, %f2747; setp.lt.ftz.f32 %p219, %f2748, %f2740; selp.f32 %f3386, %f2748, %f2740, %p219; selp.f32 %f2749, 0f3F800000, %f2741, %p219; selp.f32 %f2750, %f2746, %f2742, %p219; add.ftz.f32 %f2751, %f2724, %f2749; mov.b32 %r4239, %f2751; add.s32 %r4240, %r4239, -559038720; xor.b32 %r4241, %r4240, -1874868139; add.s32 %r4242, %r4241, 29184895; xor.b32 %r4243, %r4242, -559038720; shf.l.wrap.b32 %r4244, %r4242, %r4242, 25; sub.s32 %r4245, %r4243, %r4244; xor.b32 %r4246, %r4245, -1874868139; shf.l.wrap.b32 %r4247, %r4245, %r4245, 16; sub.s32 %r4248, %r4246, %r4247; xor.b32 %r4249, %r4248, %r4242; shf.l.wrap.b32 %r4250, %r4248, %r4248, 4; sub.s32 %r4251, %r4249, %r4250; xor.b32 %r4252, %r4251, %r4245; shf.l.wrap.b32 %r4253, %r4251, %r4251, 14; sub.s32 %r4254, %r4252, %r4253; xor.b32 %r4255, %r4254, %r4248; shf.l.wrap.b32 %r4256, %r4254, %r4254, 24; sub.s32 %r4257, %r4255, %r4256; cvt.rn.f32.u32 %f2752, %r4257; mul.ftz.f32 %f3391, %f2752, 0f2F800000; add.s32 %r4258, %r4239, -559038716; xor.b32 %r4259, %r4258, 1354692725; add.s32 %r4260, %r4259, 138171771; xor.b32 %r4261, %r4260, 506314500; shf.l.wrap.b32 %r4262, %r4260, %r4260, 25; sub.s32 %r4263, %r4261, %r4262; xor.b32 %r4264, %r4263, 1354692725; shf.l.wrap.b32 %r4265, %r4263, %r4263, 16; sub.s32 %r4266, %r4264, %r4265; xor.b32 %r4267, %r4266, %r4260; shf.l.wrap.b32 %r4268, %r4266, %r4266, 4; sub.s32 %r4269, %r4267, %r4268; xor.b32 %r4270, %r4269, %r4263; shf.l.wrap.b32 %r4271, %r4269, %r4269, 14; sub.s32 %r4272, %r4270, %r4271; xor.b32 %r4273, %r4272, %r4266; shf.l.wrap.b32 %r4274, %r4272, %r4272, 24; sub.s32 %r4275, %r4273, %r4274; cvt.rn.f32.u32 %f2753, %r4275; mul.ftz.f32 %f3390, %f2753, 0f2F800000; xor.b32 %r4276, %r4258, 1346304085; add.s32 %r4277, %r4276, 138237311; xor.b32 %r4278, %r4277, 514703108; shf.l.wrap.b32 %r4279, %r4277, %r4277, 25; sub.s32 %r4280, %r4278, %r4279; xor.b32 %r4281, %r4280, 1346304085; shf.l.wrap.b32 %r4282, %r4280, %r4280, 16; sub.s32 %r4283, %r4281, %r4282; xor.b32 %r4284, %r4283, %r4277; shf.l.wrap.b32 %r4285, %r4283, %r4283, 4; sub.s32 %r4286, %r4284, %r4285; xor.b32 %r4287, %r4286, %r4280; shf.l.wrap.b32 %r4288, %r4286, %r4286, 14; sub.s32 %r4289, %r4287, %r4288; xor.b32 %r4290, %r4289, %r4283; shf.l.wrap.b32 %r4291, %r4289, %r4289, 24; sub.s32 %r4292, %r4290, %r4291; cvt.rn.f32.u32 %f2754, %r4292; mul.ftz.f32 %f3389, %f2754, 0f2F800000; add.ftz.f32 %f3381, %f2724, %f2750; mov.f32 %f3388, %f3392; $L__BB199_383: setp.eq.ftz.f32 %p220, %f9, 0f00000000; mov.f32 %f3387, %f3392; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; @%p220 bra $L__BB199_385; div.approx.ftz.f32 %f3387, %f3381, %f9; mov.f32 %f3393, %f3392; mov.f32 %f3394, %f3392; bra.uni $L__BB199_385; $L__BB199_215: ld.param.u32 %r4317, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_3]; cvt.rmi.ftz.f32.f32 %f733, %f22; sub.ftz.f32 %f734, %f22, %f733; cvt.rmi.ftz.f32.f32 %f735, %f23; sub.ftz.f32 %f736, %f23, %f735; cvt.rmi.ftz.f32.f32 %f737, %f24; sub.ftz.f32 %f738, %f24, %f737; rcp.approx.ftz.f32 %f739, %f15; add.ftz.f32 %f1924, %f733, 0fBF800000; mov.b32 %r283, %f1924; add.s32 %r284, %r283, -559038712; add.ftz.f32 %f1925, %f733, 0f00000000; mov.f32 %f3217, 0f00000000; mov.b32 %r285, %f1925; add.s32 %r286, %r285, -559038712; add.ftz.f32 %f1926, %f733, 0f3F800000; mov.b32 %r287, %f1926; add.s32 %r288, %r287, -559038712; mov.f32 %f3386, 0f41000000; mov.u32 %r4369, -1; setp.eq.s32 %p109, %r4317, 0; setp.eq.s32 %p120, %r4317, 1; setp.eq.s32 %p121, %r4317, 2; setp.ne.s32 %p122, %r4317, 3; mov.f32 %f3216, %f3217; mov.f32 %f3215, %f3217; mov.f32 %f3214, %f3217; mov.f32 %f3213, %f3217; mov.f32 %f3212, %f3217; $L__BB199_216: cvt.rn.f32.s32 %f747, %r4369; add.ftz.f32 %f1927, %f737, %f747; mov.b32 %r2652, %f1927; add.s32 %r290, %r2652, -559038712; add.s32 %r291, %r2652, -559038708; shf.l.wrap.b32 %r2653, %r291, %r291, 4; sub.s32 %r2654, %r283, %r2652; xor.b32 %r292, %r2654, %r2653; shf.l.wrap.b32 %r293, %r292, %r292, 6; sub.s32 %r2655, %r285, %r2652; xor.b32 %r294, %r2655, %r2653; shf.l.wrap.b32 %r295, %r294, %r294, 6; sub.s32 %r2656, %r287, %r2652; xor.b32 %r296, %r2656, %r2653; shf.l.wrap.b32 %r297, %r296, %r296, 6; mov.u32 %r4370, -2; $L__BB199_217: add.s32 %r4370, %r4370, 1; cvt.rn.f32.s32 %f755, %r4370; add.ftz.f32 %f1928, %f735, %f755; mov.b32 %r2657, %f1928; add.s32 %r300, %r2657, -559038712; xor.b32 %r2658, %r300, %r290; shf.l.wrap.b32 %r2659, %r300, %r300, 14; sub.s32 %r301, %r2658, %r2659; shf.l.wrap.b32 %r302, %r301, %r301, 11; add.s32 %r2660, %r2657, -559038708; add.s32 %r303, %r2660, %r291; xor.b32 %r2661, %r284, %r301; sub.s32 %r2662, %r2661, %r302; xor.b32 %r2663, %r2662, %r300; shf.l.wrap.b32 %r2664, %r2662, %r2662, 25; sub.s32 %r2665, %r2663, %r2664; xor.b32 %r2666, %r2665, %r301; shf.l.wrap.b32 %r2667, %r2665, %r2665, 16; sub.s32 %r2668, %r2666, %r2667; xor.b32 %r2669, %r2668, %r2662; shf.l.wrap.b32 %r2670, %r2668, %r2668, 4; sub.s32 %r2671, %r2669, %r2670; xor.b32 %r2672, %r2671, %r2665; shf.l.wrap.b32 %r2673, %r2671, %r2671, 14; sub.s32 %r2674, %r2672, %r2673; xor.b32 %r2675, %r2674, %r2668; shf.l.wrap.b32 %r2676, %r2674, %r2674, 24; sub.s32 %r2677, %r2675, %r2676; cvt.rn.f32.u32 %f1929, %r2677; mul.ftz.f32 %f1930, %f1929, 0f2F800000; sub.s32 %r2678, %r2660, %r292; xor.b32 %r2679, %r293, %r2678; add.s32 %r2680, %r292, %r303; sub.s32 %r2681, %r303, %r2679; shf.l.wrap.b32 %r2682, %r2679, %r2679, 8; xor.b32 %r2683, %r2682, %r2681; add.s32 %r2684, %r2679, %r2680; sub.s32 %r2685, %r2680, %r2683; shf.l.wrap.b32 %r2686, %r2683, %r2683, 16; xor.b32 %r2687, %r2686, %r2685; add.s32 %r2688, %r2683, %r2684; sub.s32 %r2689, %r2684, %r2687; shf.l.wrap.b32 %r2690, %r2687, %r2687, 19; xor.b32 %r2691, %r2690, %r2689; add.s32 %r2692, %r2687, %r2688; sub.s32 %r2693, %r2688, %r2691; shf.l.wrap.b32 %r2694, %r2691, %r2691, 4; add.s32 %r2695, %r2691, %r2692; add.s32 %r2696, %r2692, 1065353216; xor.b32 %r2697, %r2695, %r2693; xor.b32 %r2698, %r2697, %r2694; shf.l.wrap.b32 %r2699, %r2695, %r2695, 14; sub.s32 %r2700, %r2698, %r2699; xor.b32 %r2701, %r2700, %r2696; shf.l.wrap.b32 %r2702, %r2700, %r2700, 11; sub.s32 %r2703, %r2701, %r2702; xor.b32 %r2704, %r2703, %r2695; shf.l.wrap.b32 %r2705, %r2703, %r2703, 25; sub.s32 %r2706, %r2704, %r2705; xor.b32 %r2707, %r2706, %r2700; shf.l.wrap.b32 %r2708, %r2706, %r2706, 16; sub.s32 %r2709, %r2707, %r2708; xor.b32 %r2710, %r2709, %r2703; shf.l.wrap.b32 %r2711, %r2709, %r2709, 4; sub.s32 %r2712, %r2710, %r2711; xor.b32 %r2713, %r2712, %r2706; shf.l.wrap.b32 %r2714, %r2712, %r2712, 14; sub.s32 %r2715, %r2713, %r2714; xor.b32 %r2716, %r2715, %r2709; shf.l.wrap.b32 %r2717, %r2715, %r2715, 24; sub.s32 %r2718, %r2716, %r2717; cvt.rn.f32.u32 %f1931, %r2718; mul.ftz.f32 %f1932, %f1931, 0f2F800000; add.s32 %r2719, %r2692, 1073741824; xor.b32 %r2720, %r2700, %r2719; sub.s32 %r2721, %r2720, %r2702; xor.b32 %r2722, %r2721, %r2695; shf.l.wrap.b32 %r2723, %r2721, %r2721, 25; sub.s32 %r2724, %r2722, %r2723; xor.b32 %r2725, %r2724, %r2700; shf.l.wrap.b32 %r2726, %r2724, %r2724, 16; sub.s32 %r2727, %r2725, %r2726; xor.b32 %r2728, %r2727, %r2721; shf.l.wrap.b32 %r2729, %r2727, %r2727, 4; sub.s32 %r2730, %r2728, %r2729; xor.b32 %r2731, %r2730, %r2724; shf.l.wrap.b32 %r2732, %r2730, %r2730, 14; sub.s32 %r2733, %r2731, %r2732; xor.b32 %r2734, %r2733, %r2727; shf.l.wrap.b32 %r2735, %r2733, %r2733, 24; sub.s32 %r2736, %r2734, %r2735; cvt.rn.f32.u32 %f1933, %r2736; mul.ftz.f32 %f1934, %f1933, 0f2F800000; fma.rn.ftz.f32 %f756, %f19, %f1930, 0fBF800000; fma.rn.ftz.f32 %f757, %f19, %f1932, %f755; fma.rn.ftz.f32 %f758, %f19, %f1934, %f747; @%p109 bra $L__BB199_224; @%p120 bra $L__BB199_223; bra.uni $L__BB199_219; $L__BB199_223: sub.ftz.f32 %f1962, %f756, %f734; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f757, %f736; abs.ftz.f32 %f1965, %f1964; add.ftz.f32 %f1966, %f1963, %f1965; sub.ftz.f32 %f1967, %f758, %f738; abs.ftz.f32 %f1968, %f1967; add.ftz.f32 %f3211, %f1966, %f1968; bra.uni $L__BB199_225; $L__BB199_224: sub.ftz.f32 %f1969, %f756, %f734; sub.ftz.f32 %f1970, %f757, %f736; mul.ftz.f32 %f1971, %f1970, %f1970; fma.rn.ftz.f32 %f1972, %f1969, %f1969, %f1971; sub.ftz.f32 %f1973, %f758, %f738; fma.rn.ftz.f32 %f1974, %f1973, %f1973, %f1972; sqrt.approx.ftz.f32 %f3211, %f1974; bra.uni $L__BB199_225; $L__BB199_219: @%p121 bra $L__BB199_222; bra.uni $L__BB199_220; $L__BB199_222: sub.ftz.f32 %f1955, %f756, %f734; abs.ftz.f32 %f1956, %f1955; sub.ftz.f32 %f1957, %f757, %f736; abs.ftz.f32 %f1958, %f1957; sub.ftz.f32 %f1959, %f758, %f738; abs.ftz.f32 %f1960, %f1959; max.ftz.f32 %f1961, %f1958, %f1960; max.ftz.f32 %f3211, %f1956, %f1961; bra.uni $L__BB199_225; $L__BB199_220: mov.f32 %f3211, 0f00000000; @%p122 bra $L__BB199_225; sub.ftz.f32 %f1936, %f756, %f734; abs.ftz.f32 %f1937, %f1936; lg2.approx.ftz.f32 %f1938, %f1937; mul.ftz.f32 %f1939, %f15, %f1938; ex2.approx.ftz.f32 %f1940, %f1939; sub.ftz.f32 %f1941, %f757, %f736; abs.ftz.f32 %f1942, %f1941; lg2.approx.ftz.f32 %f1943, %f1942; mul.ftz.f32 %f1944, %f15, %f1943; ex2.approx.ftz.f32 %f1945, %f1944; add.ftz.f32 %f1946, %f1940, %f1945; sub.ftz.f32 %f1947, %f758, %f738; abs.ftz.f32 %f1948, %f1947; lg2.approx.ftz.f32 %f1949, %f1948; mul.ftz.f32 %f1950, %f15, %f1949; ex2.approx.ftz.f32 %f1951, %f1950; add.ftz.f32 %f1952, %f1946, %f1951; lg2.approx.ftz.f32 %f1953, %f1952; mul.ftz.f32 %f1954, %f739, %f1953; ex2.approx.ftz.f32 %f3211, %f1954; $L__BB199_225: setp.geu.ftz.f32 %p113, %f3211, %f3386; @%p113 bra $L__BB199_227; mov.f32 %f3215, 0fBF800000; mov.f32 %f3212, %f756; mov.f32 %f3213, %f757; mov.f32 %f3214, %f758; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3211; $L__BB199_227: xor.b32 %r2737, %r286, %r301; sub.s32 %r2738, %r2737, %r302; xor.b32 %r2739, %r2738, %r300; shf.l.wrap.b32 %r2740, %r2738, %r2738, 25; sub.s32 %r2741, %r2739, %r2740; xor.b32 %r2742, %r2741, %r301; shf.l.wrap.b32 %r2743, %r2741, %r2741, 16; sub.s32 %r2744, %r2742, %r2743; xor.b32 %r2745, %r2744, %r2738; shf.l.wrap.b32 %r2746, %r2744, %r2744, 4; sub.s32 %r2747, %r2745, %r2746; xor.b32 %r2748, %r2747, %r2741; shf.l.wrap.b32 %r2749, %r2747, %r2747, 14; sub.s32 %r2750, %r2748, %r2749; xor.b32 %r2751, %r2750, %r2744; shf.l.wrap.b32 %r2752, %r2750, %r2750, 24; sub.s32 %r2753, %r2751, %r2752; cvt.rn.f32.u32 %f1976, %r2753; mul.ftz.f32 %f1977, %f1976, 0f2F800000; add.s32 %r2754, %r300, 4; sub.s32 %r2755, %r2754, %r294; xor.b32 %r2756, %r295, %r2755; sub.s32 %r2757, %r303, %r2756; shf.l.wrap.b32 %r2758, %r2756, %r2756, 8; xor.b32 %r2759, %r2758, %r2757; add.s32 %r2760, %r294, %r303; add.s32 %r2761, %r2756, %r2760; sub.s32 %r2762, %r2760, %r2759; shf.l.wrap.b32 %r2763, %r2759, %r2759, 16; xor.b32 %r2764, %r2763, %r2762; add.s32 %r2765, %r2759, %r2761; sub.s32 %r2766, %r2761, %r2764; shf.l.wrap.b32 %r2767, %r2764, %r2764, 19; xor.b32 %r2768, %r2767, %r2766; add.s32 %r2769, %r2764, %r2765; sub.s32 %r2770, %r2765, %r2768; shf.l.wrap.b32 %r2771, %r2768, %r2768, 4; add.s32 %r2772, %r2768, %r2769; add.s32 %r2773, %r2769, 1065353216; xor.b32 %r2774, %r2772, %r2770; xor.b32 %r2775, %r2774, %r2771; shf.l.wrap.b32 %r2776, %r2772, %r2772, 14; sub.s32 %r2777, %r2775, %r2776; xor.b32 %r2778, %r2777, %r2773; shf.l.wrap.b32 %r2779, %r2777, %r2777, 11; sub.s32 %r2780, %r2778, %r2779; xor.b32 %r2781, %r2780, %r2772; shf.l.wrap.b32 %r2782, %r2780, %r2780, 25; sub.s32 %r2783, %r2781, %r2782; xor.b32 %r2784, %r2783, %r2777; shf.l.wrap.b32 %r2785, %r2783, %r2783, 16; sub.s32 %r2786, %r2784, %r2785; xor.b32 %r2787, %r2786, %r2780; shf.l.wrap.b32 %r2788, %r2786, %r2786, 4; sub.s32 %r2789, %r2787, %r2788; xor.b32 %r2790, %r2789, %r2783; shf.l.wrap.b32 %r2791, %r2789, %r2789, 14; sub.s32 %r2792, %r2790, %r2791; xor.b32 %r2793, %r2792, %r2786; shf.l.wrap.b32 %r2794, %r2792, %r2792, 24; sub.s32 %r2795, %r2793, %r2794; cvt.rn.f32.u32 %f1978, %r2795; mul.ftz.f32 %f1979, %f1978, 0f2F800000; add.s32 %r2796, %r2769, 1073741824; xor.b32 %r2797, %r2777, %r2796; sub.s32 %r2798, %r2797, %r2779; xor.b32 %r2799, %r2798, %r2772; shf.l.wrap.b32 %r2800, %r2798, %r2798, 25; sub.s32 %r2801, %r2799, %r2800; xor.b32 %r2802, %r2801, %r2777; shf.l.wrap.b32 %r2803, %r2801, %r2801, 16; sub.s32 %r2804, %r2802, %r2803; xor.b32 %r2805, %r2804, %r2798; shf.l.wrap.b32 %r2806, %r2804, %r2804, 4; sub.s32 %r2807, %r2805, %r2806; xor.b32 %r2808, %r2807, %r2801; shf.l.wrap.b32 %r2809, %r2807, %r2807, 14; sub.s32 %r2810, %r2808, %r2809; xor.b32 %r2811, %r2810, %r2804; shf.l.wrap.b32 %r2812, %r2810, %r2810, 24; sub.s32 %r2813, %r2811, %r2812; cvt.rn.f32.u32 %f1980, %r2813; mul.ftz.f32 %f1981, %f1980, 0f2F800000; fma.rn.ftz.f32 %f771, %f19, %f1977, 0f00000000; fma.rn.ftz.f32 %f772, %f19, %f1979, %f755; fma.rn.ftz.f32 %f773, %f19, %f1981, %f747; @%p109 bra $L__BB199_234; @%p120 bra $L__BB199_233; bra.uni $L__BB199_229; $L__BB199_233: sub.ftz.f32 %f2009, %f771, %f734; abs.ftz.f32 %f2010, %f2009; sub.ftz.f32 %f2011, %f772, %f736; abs.ftz.f32 %f2012, %f2011; add.ftz.f32 %f2013, %f2010, %f2012; sub.ftz.f32 %f2014, %f773, %f738; abs.ftz.f32 %f2015, %f2014; add.ftz.f32 %f3219, %f2013, %f2015; bra.uni $L__BB199_235; $L__BB199_234: sub.ftz.f32 %f2016, %f771, %f734; sub.ftz.f32 %f2017, %f772, %f736; mul.ftz.f32 %f2018, %f2017, %f2017; fma.rn.ftz.f32 %f2019, %f2016, %f2016, %f2018; sub.ftz.f32 %f2020, %f773, %f738; fma.rn.ftz.f32 %f2021, %f2020, %f2020, %f2019; sqrt.approx.ftz.f32 %f3219, %f2021; bra.uni $L__BB199_235; $L__BB199_229: @%p121 bra $L__BB199_232; bra.uni $L__BB199_230; $L__BB199_232: sub.ftz.f32 %f2002, %f771, %f734; abs.ftz.f32 %f2003, %f2002; sub.ftz.f32 %f2004, %f772, %f736; abs.ftz.f32 %f2005, %f2004; sub.ftz.f32 %f2006, %f773, %f738; abs.ftz.f32 %f2007, %f2006; max.ftz.f32 %f2008, %f2005, %f2007; max.ftz.f32 %f3219, %f2003, %f2008; bra.uni $L__BB199_235; $L__BB199_230: mov.f32 %f3219, 0f00000000; @%p122 bra $L__BB199_235; sub.ftz.f32 %f1983, %f771, %f734; abs.ftz.f32 %f1984, %f1983; lg2.approx.ftz.f32 %f1985, %f1984; mul.ftz.f32 %f1986, %f15, %f1985; ex2.approx.ftz.f32 %f1987, %f1986; sub.ftz.f32 %f1988, %f772, %f736; abs.ftz.f32 %f1989, %f1988; lg2.approx.ftz.f32 %f1990, %f1989; mul.ftz.f32 %f1991, %f15, %f1990; ex2.approx.ftz.f32 %f1992, %f1991; add.ftz.f32 %f1993, %f1987, %f1992; sub.ftz.f32 %f1994, %f773, %f738; abs.ftz.f32 %f1995, %f1994; lg2.approx.ftz.f32 %f1996, %f1995; mul.ftz.f32 %f1997, %f15, %f1996; ex2.approx.ftz.f32 %f1998, %f1997; add.ftz.f32 %f1999, %f1993, %f1998; lg2.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f739, %f2000; ex2.approx.ftz.f32 %f3219, %f2001; $L__BB199_235: setp.geu.ftz.f32 %p118, %f3219, %f3386; @%p118 bra $L__BB199_237; mov.f32 %f3215, 0f00000000; mov.f32 %f3212, %f771; mov.f32 %f3213, %f772; mov.f32 %f3214, %f773; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3219; $L__BB199_237: xor.b32 %r2814, %r288, %r301; sub.s32 %r2815, %r2814, %r302; xor.b32 %r2816, %r2815, %r300; shf.l.wrap.b32 %r2817, %r2815, %r2815, 25; sub.s32 %r2818, %r2816, %r2817; xor.b32 %r2819, %r2818, %r301; shf.l.wrap.b32 %r2820, %r2818, %r2818, 16; sub.s32 %r2821, %r2819, %r2820; xor.b32 %r2822, %r2821, %r2815; shf.l.wrap.b32 %r2823, %r2821, %r2821, 4; sub.s32 %r2824, %r2822, %r2823; xor.b32 %r2825, %r2824, %r2818; shf.l.wrap.b32 %r2826, %r2824, %r2824, 14; sub.s32 %r2827, %r2825, %r2826; xor.b32 %r2828, %r2827, %r2821; shf.l.wrap.b32 %r2829, %r2827, %r2827, 24; sub.s32 %r2830, %r2828, %r2829; cvt.rn.f32.u32 %f2023, %r2830; mul.ftz.f32 %f2024, %f2023, 0f2F800000; sub.s32 %r2832, %r2754, %r296; xor.b32 %r2833, %r297, %r2832; sub.s32 %r2834, %r303, %r2833; shf.l.wrap.b32 %r2835, %r2833, %r2833, 8; xor.b32 %r2836, %r2835, %r2834; add.s32 %r2837, %r296, %r303; add.s32 %r2838, %r2833, %r2837; sub.s32 %r2839, %r2837, %r2836; shf.l.wrap.b32 %r2840, %r2836, %r2836, 16; xor.b32 %r2841, %r2840, %r2839; add.s32 %r2842, %r2836, %r2838; sub.s32 %r2843, %r2838, %r2841; shf.l.wrap.b32 %r2844, %r2841, %r2841, 19; xor.b32 %r2845, %r2844, %r2843; add.s32 %r2846, %r2841, %r2842; sub.s32 %r2847, %r2842, %r2845; shf.l.wrap.b32 %r2848, %r2845, %r2845, 4; add.s32 %r2849, %r2845, %r2846; add.s32 %r2850, %r2846, 1065353216; xor.b32 %r2851, %r2849, %r2847; xor.b32 %r2852, %r2851, %r2848; shf.l.wrap.b32 %r2853, %r2849, %r2849, 14; sub.s32 %r2854, %r2852, %r2853; xor.b32 %r2855, %r2854, %r2850; shf.l.wrap.b32 %r2856, %r2854, %r2854, 11; sub.s32 %r2857, %r2855, %r2856; xor.b32 %r2858, %r2857, %r2849; shf.l.wrap.b32 %r2859, %r2857, %r2857, 25; sub.s32 %r2860, %r2858, %r2859; xor.b32 %r2861, %r2860, %r2854; shf.l.wrap.b32 %r2862, %r2860, %r2860, 16; sub.s32 %r2863, %r2861, %r2862; xor.b32 %r2864, %r2863, %r2857; shf.l.wrap.b32 %r2865, %r2863, %r2863, 4; sub.s32 %r2866, %r2864, %r2865; xor.b32 %r2867, %r2866, %r2860; shf.l.wrap.b32 %r2868, %r2866, %r2866, 14; sub.s32 %r2869, %r2867, %r2868; xor.b32 %r2870, %r2869, %r2863; shf.l.wrap.b32 %r2871, %r2869, %r2869, 24; sub.s32 %r2872, %r2870, %r2871; cvt.rn.f32.u32 %f2025, %r2872; mul.ftz.f32 %f2026, %f2025, 0f2F800000; add.s32 %r2873, %r2846, 1073741824; xor.b32 %r2874, %r2854, %r2873; sub.s32 %r2875, %r2874, %r2856; xor.b32 %r2876, %r2875, %r2849; shf.l.wrap.b32 %r2877, %r2875, %r2875, 25; sub.s32 %r2878, %r2876, %r2877; xor.b32 %r2879, %r2878, %r2854; shf.l.wrap.b32 %r2880, %r2878, %r2878, 16; sub.s32 %r2881, %r2879, %r2880; xor.b32 %r2882, %r2881, %r2875; shf.l.wrap.b32 %r2883, %r2881, %r2881, 4; sub.s32 %r2884, %r2882, %r2883; xor.b32 %r2885, %r2884, %r2878; shf.l.wrap.b32 %r2886, %r2884, %r2884, 14; sub.s32 %r2887, %r2885, %r2886; xor.b32 %r2888, %r2887, %r2881; shf.l.wrap.b32 %r2889, %r2887, %r2887, 24; sub.s32 %r2890, %r2888, %r2889; cvt.rn.f32.u32 %f2027, %r2890; mul.ftz.f32 %f2028, %f2027, 0f2F800000; fma.rn.ftz.f32 %f786, %f19, %f2024, 0f3F800000; fma.rn.ftz.f32 %f787, %f19, %f2026, %f755; fma.rn.ftz.f32 %f788, %f19, %f2028, %f747; @%p109 bra $L__BB199_244; @%p120 bra $L__BB199_243; bra.uni $L__BB199_239; $L__BB199_243: sub.ftz.f32 %f2056, %f786, %f734; abs.ftz.f32 %f2057, %f2056; sub.ftz.f32 %f2058, %f787, %f736; abs.ftz.f32 %f2059, %f2058; add.ftz.f32 %f2060, %f2057, %f2059; sub.ftz.f32 %f2061, %f788, %f738; abs.ftz.f32 %f2062, %f2061; add.ftz.f32 %f3227, %f2060, %f2062; bra.uni $L__BB199_245; $L__BB199_244: sub.ftz.f32 %f2063, %f786, %f734; sub.ftz.f32 %f2064, %f787, %f736; mul.ftz.f32 %f2065, %f2064, %f2064; fma.rn.ftz.f32 %f2066, %f2063, %f2063, %f2065; sub.ftz.f32 %f2067, %f788, %f738; fma.rn.ftz.f32 %f2068, %f2067, %f2067, %f2066; sqrt.approx.ftz.f32 %f3227, %f2068; bra.uni $L__BB199_245; $L__BB199_239: @%p121 bra $L__BB199_242; bra.uni $L__BB199_240; $L__BB199_242: sub.ftz.f32 %f2049, %f786, %f734; abs.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f787, %f736; abs.ftz.f32 %f2052, %f2051; sub.ftz.f32 %f2053, %f788, %f738; abs.ftz.f32 %f2054, %f2053; max.ftz.f32 %f2055, %f2052, %f2054; max.ftz.f32 %f3227, %f2050, %f2055; bra.uni $L__BB199_245; $L__BB199_240: mov.f32 %f3227, 0f00000000; @%p122 bra $L__BB199_245; sub.ftz.f32 %f2030, %f786, %f734; abs.ftz.f32 %f2031, %f2030; lg2.approx.ftz.f32 %f2032, %f2031; mul.ftz.f32 %f2033, %f15, %f2032; ex2.approx.ftz.f32 %f2034, %f2033; sub.ftz.f32 %f2035, %f787, %f736; abs.ftz.f32 %f2036, %f2035; lg2.approx.ftz.f32 %f2037, %f2036; mul.ftz.f32 %f2038, %f15, %f2037; ex2.approx.ftz.f32 %f2039, %f2038; add.ftz.f32 %f2040, %f2034, %f2039; sub.ftz.f32 %f2041, %f788, %f738; abs.ftz.f32 %f2042, %f2041; lg2.approx.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f15, %f2043; ex2.approx.ftz.f32 %f2045, %f2044; add.ftz.f32 %f2046, %f2040, %f2045; lg2.approx.ftz.f32 %f2047, %f2046; mul.ftz.f32 %f2048, %f739, %f2047; ex2.approx.ftz.f32 %f3227, %f2048; $L__BB199_245: setp.geu.ftz.f32 %p123, %f3227, %f3386; @%p123 bra $L__BB199_247; mov.f32 %f3215, 0f3F800000; mov.f32 %f3212, %f786; mov.f32 %f3213, %f787; mov.f32 %f3214, %f788; mov.f32 %f3216, %f755; mov.f32 %f3217, %f747; mov.f32 %f3386, %f3227; $L__BB199_247: setp.lt.s32 %p124, %r4370, 1; @%p124 bra $L__BB199_217; add.s32 %r304, %r4369, 1; setp.lt.s32 %p125, %r4369, 1; mov.u32 %r4369, %r304; @%p125 bra $L__BB199_216; add.ftz.f32 %f2071, %f733, %f3215; mov.b32 %r2891, %f2071; add.ftz.f32 %f2072, %f735, %f3216; mov.b32 %r2892, %f2072; add.ftz.f32 %f2073, %f737, %f3217; mov.b32 %r2893, %f2073; add.s32 %r2894, %r2893, -559038712; add.s32 %r2895, %r2892, -559038712; add.s32 %r2896, %r2891, -559038712; xor.b32 %r2897, %r2894, %r2895; shf.l.wrap.b32 %r2898, %r2895, %r2895, 14; sub.s32 %r2899, %r2897, %r2898; xor.b32 %r2900, %r2899, %r2896; shf.l.wrap.b32 %r2901, %r2899, %r2899, 11; sub.s32 %r2902, %r2900, %r2901; xor.b32 %r2903, %r2902, %r2895; shf.l.wrap.b32 %r2904, %r2902, %r2902, 25; sub.s32 %r2905, %r2903, %r2904; xor.b32 %r2906, %r2905, %r2899; shf.l.wrap.b32 %r2907, %r2905, %r2905, 16; sub.s32 %r2908, %r2906, %r2907; xor.b32 %r2909, %r2908, %r2902; shf.l.wrap.b32 %r2910, %r2908, %r2908, 4; sub.s32 %r2911, %r2909, %r2910; xor.b32 %r2912, %r2911, %r2905; shf.l.wrap.b32 %r2913, %r2911, %r2911, 14; sub.s32 %r2914, %r2912, %r2913; xor.b32 %r2915, %r2914, %r2908; shf.l.wrap.b32 %r2916, %r2914, %r2914, 24; sub.s32 %r2917, %r2915, %r2916; cvt.rn.f32.u32 %f2074, %r2917; mul.ftz.f32 %f3391, %f2074, 0f2F800000; add.s32 %r2918, %r2892, -559038708; add.s32 %r2919, %r2893, -559038708; sub.s32 %r2920, %r2891, %r2893; shf.l.wrap.b32 %r2921, %r2919, %r2919, 4; xor.b32 %r2922, %r2921, %r2920; add.s32 %r2923, %r2919, %r2918; sub.s32 %r2924, %r2918, %r2922; shf.l.wrap.b32 %r2925, %r2922, %r2922, 6; xor.b32 %r2926, %r2925, %r2924; add.s32 %r2927, %r2922, %r2923; sub.s32 %r2928, %r2923, %r2926; shf.l.wrap.b32 %r2929, %r2926, %r2926, 8; xor.b32 %r2930, %r2929, %r2928; add.s32 %r2931, %r2926, %r2927; sub.s32 %r2932, %r2927, %r2930; shf.l.wrap.b32 %r2933, %r2930, %r2930, 16; xor.b32 %r2934, %r2933, %r2932; add.s32 %r2935, %r2930, %r2931; sub.s32 %r2936, %r2931, %r2934; shf.l.wrap.b32 %r2937, %r2934, %r2934, 19; xor.b32 %r2938, %r2937, %r2936; add.s32 %r2939, %r2934, %r2935; sub.s32 %r2940, %r2935, %r2938; shf.l.wrap.b32 %r2941, %r2938, %r2938, 4; add.s32 %r2942, %r2938, %r2939; add.s32 %r2943, %r2939, 1065353216; xor.b32 %r2944, %r2942, %r2940; xor.b32 %r2945, %r2944, %r2941; shf.l.wrap.b32 %r2946, %r2942, %r2942, 14; sub.s32 %r2947, %r2945, %r2946; xor.b32 %r2948, %r2947, %r2943; shf.l.wrap.b32 %r2949, %r2947, %r2947, 11; sub.s32 %r2950, %r2948, %r2949; xor.b32 %r2951, %r2950, %r2942; shf.l.wrap.b32 %r2952, %r2950, %r2950, 25; sub.s32 %r2953, %r2951, %r2952; xor.b32 %r2954, %r2953, %r2947; shf.l.wrap.b32 %r2955, %r2953, %r2953, 16; sub.s32 %r2956, %r2954, %r2955; xor.b32 %r2957, %r2956, %r2950; shf.l.wrap.b32 %r2958, %r2956, %r2956, 4; sub.s32 %r2959, %r2957, %r2958; xor.b32 %r2960, %r2959, %r2953; shf.l.wrap.b32 %r2961, %r2959, %r2959, 14; sub.s32 %r2962, %r2960, %r2961; xor.b32 %r2963, %r2962, %r2956; shf.l.wrap.b32 %r2964, %r2962, %r2962, 24; sub.s32 %r2965, %r2963, %r2964; cvt.rn.f32.u32 %f2075, %r2965; mul.ftz.f32 %f3390, %f2075, 0f2F800000; add.s32 %r2966, %r2939, 1073741824; xor.b32 %r2967, %r2947, %r2966; sub.s32 %r2968, %r2967, %r2949; xor.b32 %r2969, %r2968, %r2942; shf.l.wrap.b32 %r2970, %r2968, %r2968, 25; sub.s32 %r2971, %r2969, %r2970; xor.b32 %r2972, %r2971, %r2947; shf.l.wrap.b32 %r2973, %r2971, %r2971, 16; sub.s32 %r2974, %r2972, %r2973; xor.b32 %r2975, %r2974, %r2968; shf.l.wrap.b32 %r2976, %r2974, %r2974, 4; sub.s32 %r2977, %r2975, %r2976; xor.b32 %r2978, %r2977, %r2971; shf.l.wrap.b32 %r2979, %r2977, %r2977, 14; sub.s32 %r2980, %r2978, %r2979; xor.b32 %r2981, %r2980, %r2974; shf.l.wrap.b32 %r2982, %r2980, %r2980, 24; sub.s32 %r2983, %r2981, %r2982; cvt.rn.f32.u32 %f2076, %r2983; mul.ftz.f32 %f3389, %f2076, 0f2F800000; add.ftz.f32 %f3242, %f733, %f3212; add.ftz.f32 %f3241, %f735, %f3213; add.ftz.f32 %f3240, %f737, %f3214; mov.f32 %f3388, 0f00000000; $L__BB199_250: setp.eq.ftz.f32 %p126, %f9, 0f00000000; mov.f32 %f3394, %f3387; mov.f32 %f3393, %f3387; mov.f32 %f3392, %f3387; @%p126 bra $L__BB199_385; div.approx.ftz.f32 %f3392, %f3240, %f9; div.approx.ftz.f32 %f3393, %f3241, %f9; div.approx.ftz.f32 %f3394, %f3242, %f9; $L__BB199_385: bfe.u32 %r4293, %r393, 16, 8; setp.eq.s32 %p221, %r4293, 255; @%p221 bra $L__BB199_387; shr.u32 %r4294, %r393, 14; cvt.u64.u32 %rd28, %r4294; and.b64 %rd29, %rd28, 1020; add.s64 %rd30, %rd2, %rd29; st.local.f32 [%rd30], %f3386; $L__BB199_387: shr.u32 %r4295, %r393, 24; setp.eq.s32 %p222, %r4295, 255; @%p222 bra $L__BB199_389; mul.wide.u32 %rd32, %r4295, 4; add.s64 %rd33, %rd2, %rd32; st.local.f32 [%rd33], %f3391; st.local.f32 [%rd33+4], %f3390; st.local.f32 [%rd33+8], %f3389; $L__BB199_389: and.b32 %r4297, %r394, 255; setp.eq.s32 %p223, %r4297, 255; @%p223 bra $L__BB199_391; shl.b32 %r4298, %r394, 2; cvt.u64.u32 %rd35, %r4298; and.b64 %rd36, %rd35, 1020; add.s64 %rd37, %rd2, %rd36; st.local.f32 [%rd37], %f3394; st.local.f32 [%rd37+4], %f3393; st.local.f32 [%rd37+8], %f3392; $L__BB199_391: bfe.u32 %r4299, %r394, 8, 8; setp.eq.s32 %p224, %r4299, 255; @%p224 bra $L__BB199_393; shr.u32 %r4300, %r394, 6; cvt.u64.u32 %rd39, %r4300; and.b64 %rd40, %rd39, 1020; add.s64 %rd41, %rd2, %rd40; st.local.f32 [%rd41], %f3387; $L__BB199_393: bfe.u32 %r4301, %r394, 16, 8; setp.eq.s32 %p225, %r4301, 255; @%p225 bra $L__BB199_395; shr.u32 %r4302, %r394, 14; cvt.u64.u32 %rd43, %r4302; and.b64 %rd44, %rd43, 1020; add.s64 %rd45, %rd2, %rd44; st.local.f32 [%rd45], %f3388; $L__BB199_395: ld.param.u32 %r4311, [_Z20svm_node_tex_voronoiILj447EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji_param_4]; add.s32 %r4303, %r4311, 2; st.param.b32 [func_retval0+0], %r4303; ret; } .func _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1, .param .b32 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2, .param .b64 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3, .param .align 4 .b8 _Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4[16] ) { .reg .f32 %f<4>; .reg .b32 %r<2>; .reg .b64 %rd<5>; ld.param.f32 %f1, [_Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.f32 %f2, [_Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1]; ld.param.f32 %f3, [_Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_2]; ld.param.u64 %rd1, [_Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_3]; cvta.to.local.u64 %rd2, %rd1; ld.param.u8 %r1, [_Z14svm_node_bevelILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_4+7]; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd2, %rd3; st.local.f32 [%rd4], %f1; st.local.f32 [%rd4+4], %f2; st.local.f32 [%rd4+8], %f3; ret; } .func _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4( .param .b64 _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0, .param .align 4 .b8 _Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1[16] ) { .reg .pred %p<3>; .reg .f32 %f<4>; .reg .b32 %r<7>; .reg .b64 %rd<10>; ld.param.u64 %rd2, [_Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_0]; ld.param.u32 %r2, [_Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+8]; ld.param.u32 %r4, [_Z11svm_node_aoILj447EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4_param_1+4]; cvta.to.local.u64 %rd1, %rd2; shr.u32 %r1, %r4, 24; bfe.u32 %r3, %r2, 8, 8; setp.eq.s32 %p1, %r1, 255; @%p1 bra $L__BB201_2; mul.wide.u32 %rd3, %r1, 4; add.s64 %rd4, %rd1, %rd3; mov.u32 %r5, 1065353216; st.local.u32 [%rd4], %r5; $L__BB201_2: setp.eq.s32 %p2, %r3, 255; @%p2 bra $L__BB201_4; shl.b32 %r6, %r2, 2; cvt.u64.u32 %rd5, %r6; and.b64 %rd6, %rd5, 1020; add.s64 %rd7, %rd1, %rd6; ld.local.f32 %f1, [%rd7]; ld.local.f32 %f2, [%rd7+4]; ld.local.f32 %f3, [%rd7+8]; mul.wide.u32 %rd8, %r3, 4; add.s64 %rd9, %rd1, %rd8; st.local.f32 [%rd9], %f1; st.local.f32 [%rd9+4], %f2; st.local.f32 [%rd9+8], %f3; $L__BB201_4: ret; } // .globl kernel_gpu_integrator_reset .visible .entry kernel_gpu_integrator_reset( .param .u32 kernel_gpu_integrator_reset_param_0 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<2>; .reg .b16 %rs<2>; .reg .b32 %r<6>; .reg .b64 %rd<8>; ld.param.u32 %r2, [kernel_gpu_integrator_reset_param_0]; mov.u32 %r3, %ctaid.x; mov.u32 %r4, %ntid.x; mov.u32 %r5, %tid.x; mad.lo.s32 %r1, %r3, %r4, %r5; setp.ge.s32 %p1, %r1, %r2; @%p1 bra $L__BB202_2; ld.const.u64 %rd1, [kernel_params+2520]; cvta.to.global.u64 %rd2, %rd1; mul.wide.s32 %rd3, %r1, 2; add.s64 %rd4, %rd2, %rd3; mov.u16 %rs1, 0; st.global.u16 [%rd4], %rs1; ld.const.u64 %rd5, [kernel_params+3416]; cvta.to.global.u64 %rd6, %rd5; add.s64 %rd7, %rd6, %rd3; st.global.u16 [%rd7], %rs1; $L__BB202_2: ret; } // .globl kernel_gpu_integrator_init_from_camera .visible .entry kernel_gpu_integrator_init_from_camera( .param .u64 kernel_gpu_integrator_init_from_camera_param_0, .param .u32 kernel_gpu_integrator_init_from_camera_param_1, .param .u64 kernel_gpu_integrator_init_from_camera_param_2, .param .u32 kernel_gpu_integrator_init_from_camera_param_3 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<244>; .reg .b16 %rs<11>; .reg .f32 %f<3999>; .reg .b32 %r<501>; .reg .b64 %rd<203>; ld.param.u64 %rd13, [kernel_gpu_integrator_init_from_camera_param_0]; ld.param.u32 %r52, [kernel_gpu_integrator_init_from_camera_param_1]; ld.param.u32 %r51, [kernel_gpu_integrator_init_from_camera_param_3]; mov.u32 %r53, %ntid.x; mov.u32 %r54, %ctaid.x; mov.u32 %r55, %tid.x; mad.lo.s32 %r1, %r54, %r53, %r55; mul.lo.s32 %r56, %r51, %r52; setp.ge.s32 %p1, %r1, %r56; @%p1 bra $L__BB203_263; div.s32 %r57, %r1, %r51; mul.lo.s32 %r58, %r57, %r51; sub.s32 %r2, %r1, %r58; cvta.to.global.u64 %rd15, %rd13; mul.wide.s32 %rd16, %r57, 44; add.s64 %rd17, %rd15, %rd16; add.s64 %rd2, %rd17, 40; ld.global.u32 %r59, [%rd17+40]; setp.ge.s32 %p2, %r2, %r59; @%p2 bra $L__BB203_263; ld.global.u32 %r3, [%rd2+-4]; ld.const.f32 %f1, [kernel_params+1604]; setp.lt.ftz.f32 %p3, %f1, 0f3F666666; @%p3 bra $L__BB203_4; bra.uni $L__BB203_3; $L__BB203_4: ld.global.u32 %r62, [%rd2+-28]; ld.global.u32 %r489, [%rd2+-32]; mul.lo.s32 %r63, %r62, %r489; div.u32 %r490, %r2, %r63; mul.lo.s32 %r64, %r490, %r63; sub.s32 %r491, %r2, %r64; bra.uni $L__BB203_5; $L__BB203_3: ld.global.u32 %r60, [%rd2+-20]; div.u32 %r491, %r2, %r60; mul.lo.s32 %r61, %r491, %r60; sub.s32 %r490, %r2, %r61; ld.global.u32 %r489, [%rd2+-32]; $L__BB203_5: div.u32 %r65, %r491, %r489; mul.lo.s32 %r66, %r65, %r489; sub.s32 %r67, %r491, %r66; ld.global.u32 %r68, [%rd2+-40]; add.s32 %r13, %r67, %r68; ld.global.u32 %r69, [%rd2+-36]; add.s32 %r14, %r69, %r65; ld.global.u32 %r70, [%rd2+-24]; add.s32 %r492, %r70, %r490; ld.global.u32 %r71, [%rd2+-12]; add.s32 %r72, %r71, %r13; ld.global.u32 %r73, [%rd2+-8]; mad.lo.s32 %r74, %r73, %r14, %r72; ld.const.u64 %rd18, [kernel_params+2448]; cvta.to.global.u64 %rd19, %rd18; add.s32 %r75, %r3, %r2; cvt.s64.s32 %rd3, %r75; mul.wide.s32 %rd20, %r75, 4; add.s64 %rd4, %rd19, %rd20; st.global.u32 [%rd4], %r74; ld.const.u64 %rd21, [kernel_params+2520]; cvta.to.global.u64 %rd22, %rd21; mul.wide.s32 %rd23, %r75, 2; add.s64 %rd5, %rd22, %rd23; mov.u16 %rs4, 0; st.global.u16 [%rd5], %rs4; ld.const.u32 %r16, [kernel_params+1400]; setp.eq.s32 %p4, %r16, -1; mov.u16 %rs10, 1; @%p4 bra $L__BB203_7; ld.param.u64 %rd200, [kernel_gpu_integrator_init_from_camera_param_2]; cvta.to.global.u64 %rd199, %rd200; ld.global.u32 %rd24, [%rd4]; ld.const.s32 %rd25, [kernel_params+1264]; mul.lo.s64 %rd26, %rd25, %rd24; add.s32 %r76, %r16, 3; cvt.u64.u32 %rd27, %r76; add.s64 %rd28, %rd26, %rd27; shl.b64 %rd29, %rd28, 2; add.s64 %rd30, %rd199, %rd29; ld.global.f32 %f961, [%rd30]; setp.eq.ftz.f32 %p5, %f961, 0f00000000; selp.u16 %rs10, 1, 0, %p5; $L__BB203_7: setp.eq.s16 %p6, %rs10, 0; @%p6 bra $L__BB203_263; ld.const.u32 %r17, [kernel_params+1404]; setp.eq.s32 %p7, %r17, -1; @%p7 bra $L__BB203_10; ld.param.u64 %rd202, [kernel_gpu_integrator_init_from_camera_param_2]; cvta.to.global.u64 %rd201, %rd202; ld.global.u32 %r77, [%rd2+-16]; ld.global.u32 %rd31, [%rd4]; ld.const.s32 %rd32, [kernel_params+1264]; mul.lo.s64 %rd33, %rd32, %rd31; cvt.s64.s32 %rd34, %r17; add.s64 %rd35, %rd33, %rd34; shl.b64 %rd36, %rd35, 2; add.s64 %rd37, %rd201, %rd36; atom.global.add.u32 %r78, [%rd37], 1; add.s32 %r492, %r78, %r77; $L__BB203_10: shr.u32 %r79, %r13, 1; xor.b32 %r80, %r79, %r14; mul.lo.s32 %r81, %r80, 1103515245; shr.u32 %r82, %r14, 1; xor.b32 %r83, %r82, %r13; mul.lo.s32 %r84, %r83, 1103515245; shr.u32 %r85, %r84, 3; xor.b32 %r86, %r85, %r81; mul.lo.s32 %r87, %r86, 1103515245; ld.const.u32 %r20, [kernel_params+1576]; xor.b32 %r21, %r20, %r87; setp.eq.s32 %p8, %r492, 0; mov.f32 %f3763, 0f3F000000; mov.f32 %f3762, %f3763; @%p8 bra $L__BB203_18; ld.const.u32 %r88, [kernel_params+1596]; setp.eq.s32 %p9, %r88, 0; brev.b32 %r89, %r492; mul.lo.s32 %r90, %r89, 1025551850; xor.b32 %r22, %r90, %r89; @%p9 bra $L__BB203_14; ld.const.f32 %f3707, [kernel_params+1604]; setp.geu.ftz.f32 %p10, %f3707, 0f3F800000; setp.lt.ftz.f32 %p11, %f3707, 0f3F800000; selp.b32 %r91, %r20, %r21, %p11; mul.lo.s32 %r92, %r91, -512718531; shr.u32 %r93, %r91, 16; xor.b32 %r94, %r92, %r93; bfe.u32 %r95, %r94, 4, 4; shr.u32 %r96, %r91, 8; xor.b32 %r97, %r94, %r96; xor.b32 %r98, %r97, %r95; mul.lo.s32 %r99, %r98, 153742143; shr.u32 %r100, %r91, 23; xor.b32 %r101, %r99, %r100; bfe.u32 %r102, %r101, 1, 7; xor.b32 %r103, %r102, %r101; shr.u32 %r104, %r91, 27; or.b32 %r105, %r104, 1; mul.lo.s32 %r106, %r105, %r103; mul.lo.s32 %r107, %r106, -1074439621; bfe.u32 %r108, %r107, 2, 6; xor.b32 %r109, %r108, %r107; mul.lo.s32 %r110, %r109, -1638916925; bfe.u32 %r111, %r110, 2, 6; xor.b32 %r112, %r111, %r110; mul.lo.s32 %r113, %r112, -933190689; and.b32 %r114, %r113, 255; bfe.u32 %r115, %r113, 5, 3; xor.b32 %r116, %r115, %r114; ld.const.u32 %r117, [kernel_params+1600]; add.s32 %r118, %r117, -1; xor.b32 %r119, %r91, 61; mul.lo.s32 %r120, %r119, 9; shr.u32 %r121, %r120, 4; xor.b32 %r122, %r121, %r120; mul.lo.s32 %r123, %r122, 668265261; add.s32 %r124, %r123, %r22; shr.u32 %r125, %r123, 16; or.b32 %r126, %r125, 1; mul.lo.s32 %r127, %r126, %r124; mul.lo.s32 %r128, %r127, 89287766; xor.b32 %r129, %r128, %r127; mul.lo.s32 %r130, %r129, 1403136100; xor.b32 %r131, %r130, %r129; brev.b32 %r132, %r131; neg.s32 %r133, %r117; and.b32 %r134, %r492, %r133; and.b32 %r135, %r132, %r118; or.b32 %r136, %r135, %r134; mad.lo.s32 %r137, %r116, %r117, %r136; shl.b32 %r138, %r117, 8; rem.u32 %r139, %r137, %r138; ld.const.u64 %rd38, [kernel_params+2424]; cvta.to.global.u64 %rd39, %rd38; shl.b32 %r140, %r139, 1; mul.wide.u32 %rd40, %r140, 4; add.s64 %rd41, %rd39, %rd40; ld.global.f32 %f3762, [%rd41]; ld.global.f32 %f3763, [%rd41+4]; @%p10 bra $L__BB203_18; ld.const.f32 %f3708, [kernel_params+1604]; xor.b32 %r141, %r21, 61; mul.lo.s32 %r142, %r141, 9; shr.u32 %r143, %r142, 4; xor.b32 %r144, %r143, %r142; mul.lo.s32 %r145, %r144, 668265261; cvt.rn.f32.u32 %f964, %r145; mul.ftz.f32 %f965, %f964, 0f2F7FFFFE; xor.b32 %r146, %r21, -905047700; mul.lo.s32 %r147, %r146, 9; shr.u32 %r148, %r147, 4; xor.b32 %r149, %r148, %r147; mul.lo.s32 %r150, %r149, 668265261; cvt.rn.f32.u32 %f966, %r150; mul.ftz.f32 %f967, %f966, 0f2F7FFFFE; fma.rn.ftz.f32 %f968, %f965, %f3708, %f3762; fma.rn.ftz.f32 %f969, %f967, %f3708, %f3763; cvt.rmi.ftz.f32.f32 %f970, %f968; sub.ftz.f32 %f3762, %f968, %f970; cvt.rmi.ftz.f32.f32 %f971, %f969; sub.ftz.f32 %f3763, %f969, %f971; bra.uni $L__BB203_18; $L__BB203_14: xor.b32 %r152, %r21, 508809841; add.s32 %r153, %r152, %r22; shr.u32 %r154, %r152, 16; or.b32 %r155, %r154, 1; mul.lo.s32 %r156, %r155, %r153; mul.lo.s32 %r157, %r156, 89287766; xor.b32 %r158, %r157, %r156; mul.lo.s32 %r159, %r158, 1403136100; xor.b32 %r23, %r159, %r158; setp.eq.s32 %p12, %r23, 0; mov.u32 %r496, 0; @%p12 bra $L__BB203_17; mov.u32 %r493, 0; mov.u64 %rd43, sobol_burley_table; mov.u32 %r494, %r23; mov.u32 %r496, %r493; $L__BB203_16: clz.b32 %r162, %r494; add.s32 %r163, %r162, %r493; mul.wide.u32 %rd42, %r163, 4; add.s64 %rd44, %rd43, %rd42; ld.const.u32 %r164, [%rd44+128]; xor.b32 %r496, %r164, %r496; add.s32 %r493, %r163, 1; shl.b32 %r165, %r494, %r162; shl.b32 %r494, %r165, 1; setp.ne.s32 %p13, %r494, 0; @%p13 bra $L__BB203_16; $L__BB203_17: brev.b32 %r166, %r23; xor.b32 %r167, %r21, 106206365; shr.u32 %r168, %r167, 16; or.b32 %r169, %r168, 1; mul.lo.s32 %r170, %r166, 1025551850; xor.b32 %r171, %r170, %r166; add.s32 %r172, %r171, %r167; mul.lo.s32 %r173, %r172, %r169; mul.lo.s32 %r174, %r173, 89287766; xor.b32 %r175, %r174, %r173; xor.b32 %r176, %r21, 1919448741; mul.lo.s32 %r177, %r496, 1025551850; xor.b32 %r178, %r177, %r496; add.s32 %r179, %r178, %r176; shr.u32 %r180, %r176, 16; or.b32 %r181, %r180, 1; mul.lo.s32 %r182, %r179, %r181; mul.lo.s32 %r183, %r182, 89287766; xor.b32 %r184, %r183, %r182; mul.lo.s32 %r185, %r184, 1403136100; xor.b32 %r186, %r185, %r184; brev.b32 %r187, %r186; cvt.rn.f32.u32 %f972, %r187; mul.lo.s32 %r188, %r175, 1403136100; xor.b32 %r189, %r188, %r175; brev.b32 %r190, %r189; cvt.rn.f32.u32 %f973, %r190; mul.ftz.f32 %f3763, %f972, 0f2F7FFFFE; mul.ftz.f32 %f3762, %f973, 0f2F7FFFFE; $L__BB203_18: ld.const.f32 %f18, [kernel_params+240]; setp.leu.ftz.f32 %p14, %f18, 0f00000000; mov.f32 %f33, 0f00000000; mov.f32 %f34, %f33; @%p14 bra $L__BB203_26; ld.const.u32 %r191, [kernel_params+1596]; setp.eq.s32 %p15, %r191, 0; brev.b32 %r192, %r492; mul.lo.s32 %r193, %r192, 1025551850; xor.b32 %r31, %r193, %r192; @%p15 bra $L__BB203_22; ld.const.u32 %r488, [kernel_params+1576]; ld.const.f32 %f3711, [kernel_params+1604]; setp.geu.ftz.f32 %p16, %f3711, 0f3F800000; setp.lt.ftz.f32 %p17, %f3711, 0f3F800000; selp.b32 %r194, %r488, %r21, %p17; xor.b32 %r195, %r194, 1; mul.lo.s32 %r196, %r195, -512718531; shr.u32 %r197, %r194, 16; xor.b32 %r198, %r196, %r197; bfe.u32 %r199, %r198, 4, 4; shr.u32 %r200, %r194, 8; xor.b32 %r201, %r198, %r200; xor.b32 %r202, %r201, %r199; mul.lo.s32 %r203, %r202, 153742143; shr.u32 %r204, %r194, 23; xor.b32 %r205, %r203, %r204; bfe.u32 %r206, %r205, 1, 7; xor.b32 %r207, %r206, %r205; shr.u32 %r208, %r194, 27; or.b32 %r209, %r208, 1; mul.lo.s32 %r210, %r209, %r207; mul.lo.s32 %r211, %r210, -1074439621; bfe.u32 %r212, %r211, 2, 6; xor.b32 %r213, %r212, %r211; mul.lo.s32 %r214, %r213, -1638916925; bfe.u32 %r215, %r214, 2, 6; xor.b32 %r216, %r215, %r214; mul.lo.s32 %r217, %r216, -933190689; and.b32 %r218, %r217, 255; bfe.u32 %r219, %r217, 5, 3; xor.b32 %r220, %r219, %r218; ld.const.u32 %r221, [kernel_params+1600]; add.s32 %r222, %r221, -1; xor.b32 %r223, %r194, 60; mul.lo.s32 %r224, %r223, 9; shr.u32 %r225, %r224, 4; xor.b32 %r226, %r225, %r224; mul.lo.s32 %r227, %r226, 668265261; add.s32 %r228, %r227, %r31; shr.u32 %r229, %r227, 16; or.b32 %r230, %r229, 1; mul.lo.s32 %r231, %r230, %r228; mul.lo.s32 %r232, %r231, 89287766; xor.b32 %r233, %r232, %r231; mul.lo.s32 %r234, %r233, 1403136100; xor.b32 %r235, %r234, %r233; brev.b32 %r236, %r235; neg.s32 %r237, %r221; and.b32 %r238, %r492, %r237; and.b32 %r239, %r236, %r222; or.b32 %r240, %r239, %r238; mad.lo.s32 %r241, %r220, %r221, %r240; shl.b32 %r242, %r221, 8; rem.u32 %r243, %r241, %r242; ld.const.u64 %rd45, [kernel_params+2424]; cvta.to.global.u64 %rd46, %rd45; shl.b32 %r244, %r243, 1; mul.wide.u32 %rd47, %r244, 4; add.s64 %rd48, %rd46, %rd47; ld.global.f32 %f34, [%rd48]; ld.global.f32 %f33, [%rd48+4]; @%p16 bra $L__BB203_26; ld.const.f32 %f3712, [kernel_params+1604]; xor.b32 %r245, %r21, 60; mul.lo.s32 %r246, %r245, 9; shr.u32 %r247, %r246, 4; xor.b32 %r248, %r247, %r246; mul.lo.s32 %r249, %r248, 668265261; cvt.rn.f32.u32 %f976, %r249; mul.ftz.f32 %f977, %f976, 0f2F7FFFFE; xor.b32 %r250, %r21, -905047699; mul.lo.s32 %r251, %r250, 9; shr.u32 %r252, %r251, 4; xor.b32 %r253, %r252, %r251; mul.lo.s32 %r254, %r253, 668265261; cvt.rn.f32.u32 %f978, %r254; mul.ftz.f32 %f979, %f978, 0f2F7FFFFE; fma.rn.ftz.f32 %f980, %f977, %f3712, %f34; fma.rn.ftz.f32 %f981, %f979, %f3712, %f33; cvt.rmi.ftz.f32.f32 %f982, %f980; sub.ftz.f32 %f34, %f980, %f982; cvt.rmi.ftz.f32.f32 %f983, %f981; sub.ftz.f32 %f33, %f981, %f983; bra.uni $L__BB203_26; $L__BB203_22: xor.b32 %r256, %r21, 411052034; add.s32 %r257, %r256, %r31; shr.u32 %r258, %r256, 16; or.b32 %r259, %r258, 1; mul.lo.s32 %r260, %r259, %r257; mul.lo.s32 %r261, %r260, 89287766; xor.b32 %r262, %r261, %r260; mul.lo.s32 %r263, %r262, 1403136100; xor.b32 %r32, %r263, %r262; setp.eq.s32 %p18, %r32, 0; mov.u32 %r500, 0; @%p18 bra $L__BB203_25; mov.u32 %r497, 0; mov.u64 %rd50, sobol_burley_table; mov.u32 %r498, %r32; mov.u32 %r500, %r497; $L__BB203_24: clz.b32 %r266, %r498; add.s32 %r267, %r266, %r497; mul.wide.u32 %rd49, %r267, 4; add.s64 %rd51, %rd50, %rd49; ld.const.u32 %r268, [%rd51+128]; xor.b32 %r500, %r268, %r500; add.s32 %r497, %r267, 1; shl.b32 %r269, %r498, %r266; shl.b32 %r498, %r269, 1; setp.ne.s32 %p19, %r498, 0; @%p19 bra $L__BB203_24; $L__BB203_25: brev.b32 %r270, %r32; xor.b32 %r271, %r21, 8875758; shr.u32 %r272, %r271, 16; or.b32 %r273, %r272, 1; mul.lo.s32 %r274, %r270, 1025551850; xor.b32 %r275, %r274, %r270; add.s32 %r276, %r275, %r271; mul.lo.s32 %r277, %r276, %r273; mul.lo.s32 %r278, %r277, 89287766; xor.b32 %r279, %r278, %r277; xor.b32 %r280, %r21, 1958448342; mul.lo.s32 %r281, %r500, 1025551850; xor.b32 %r282, %r281, %r500; add.s32 %r283, %r282, %r280; shr.u32 %r284, %r280, 16; or.b32 %r285, %r284, 1; mul.lo.s32 %r286, %r283, %r285; mul.lo.s32 %r287, %r286, 89287766; xor.b32 %r288, %r287, %r286; mul.lo.s32 %r289, %r288, 1403136100; xor.b32 %r290, %r289, %r288; brev.b32 %r291, %r290; cvt.rn.f32.u32 %f984, %r291; mul.lo.s32 %r292, %r279, 1403136100; xor.b32 %r293, %r292, %r279; brev.b32 %r294, %r293; cvt.rn.f32.u32 %f985, %r294; mul.ftz.f32 %f33, %f984, 0f2F7FFFFE; mul.ftz.f32 %f34, %f985, 0f2F7FFFFE; $L__BB203_26: mov.f32 %f3772, 0f00000000; ld.const.f32 %f35, [kernel_params+256]; setp.eq.ftz.f32 %p20, %f35, 0fBF800000; @%p20 bra $L__BB203_31; ld.const.u32 %r295, [kernel_params+1596]; setp.eq.s32 %p21, %r295, 0; brev.b32 %r296, %r492; mul.lo.s32 %r297, %r296, 1025551850; xor.b32 %r40, %r297, %r296; @%p21 bra $L__BB203_30; ld.const.u32 %r487, [kernel_params+1576]; ld.const.f32 %f3709, [kernel_params+1604]; setp.geu.ftz.f32 %p22, %f3709, 0f3F800000; setp.lt.ftz.f32 %p23, %f3709, 0f3F800000; selp.b32 %r298, %r487, %r21, %p23; xor.b32 %r299, %r298, 2; mul.lo.s32 %r300, %r299, -512718531; shr.u32 %r301, %r298, 16; xor.b32 %r302, %r300, %r301; bfe.u32 %r303, %r302, 4, 4; shr.u32 %r304, %r298, 8; xor.b32 %r305, %r302, %r304; xor.b32 %r306, %r305, %r303; mul.lo.s32 %r307, %r306, 153742143; shr.u32 %r308, %r298, 23; xor.b32 %r309, %r307, %r308; bfe.u32 %r310, %r309, 1, 7; xor.b32 %r311, %r310, %r309; shr.u32 %r312, %r298, 27; or.b32 %r313, %r312, 1; mul.lo.s32 %r314, %r313, %r311; mul.lo.s32 %r315, %r314, -1074439621; bfe.u32 %r316, %r315, 2, 6; xor.b32 %r317, %r316, %r315; mul.lo.s32 %r318, %r317, -1638916925; bfe.u32 %r319, %r318, 2, 6; xor.b32 %r320, %r319, %r318; mul.lo.s32 %r321, %r320, -933190689; and.b32 %r322, %r321, 255; bfe.u32 %r323, %r321, 5, 3; xor.b32 %r324, %r323, %r322; ld.const.u32 %r325, [kernel_params+1600]; add.s32 %r326, %r325, -1; xor.b32 %r327, %r298, 63; mul.lo.s32 %r328, %r327, 9; shr.u32 %r329, %r328, 4; xor.b32 %r330, %r329, %r328; mul.lo.s32 %r331, %r330, 668265261; add.s32 %r332, %r331, %r40; shr.u32 %r333, %r331, 16; or.b32 %r334, %r333, 1; mul.lo.s32 %r335, %r334, %r332; mul.lo.s32 %r336, %r335, 89287766; xor.b32 %r337, %r336, %r335; mul.lo.s32 %r338, %r337, 1403136100; xor.b32 %r339, %r338, %r337; brev.b32 %r340, %r339; neg.s32 %r341, %r325; and.b32 %r342, %r492, %r341; and.b32 %r343, %r340, %r326; or.b32 %r344, %r343, %r342; mad.lo.s32 %r345, %r324, %r325, %r344; shl.b32 %r346, %r325, 8; rem.u32 %r347, %r345, %r346; ld.const.u64 %rd52, [kernel_params+2424]; cvta.to.global.u64 %rd53, %rd52; shl.b32 %r348, %r347, 1; mul.wide.u32 %rd54, %r348, 4; add.s64 %rd55, %rd53, %rd54; ld.global.f32 %f3772, [%rd55]; @%p22 bra $L__BB203_31; ld.const.f32 %f3710, [kernel_params+1604]; xor.b32 %r349, %r21, 63; mul.lo.s32 %r350, %r349, 9; shr.u32 %r351, %r350, 4; xor.b32 %r352, %r351, %r350; mul.lo.s32 %r353, %r352, 668265261; cvt.rn.f32.u32 %f987, %r353; mul.ftz.f32 %f988, %f987, 0f2F7FFFFE; fma.rn.ftz.f32 %f989, %f988, %f3710, %f3772; cvt.rmi.ftz.f32.f32 %f990, %f989; sub.ftz.f32 %f3772, %f989, %f990; bra.uni $L__BB203_31; $L__BB203_30: xor.b32 %r354, %r21, 1419810034; add.s32 %r355, %r354, %r40; shr.u32 %r356, %r354, 16; or.b32 %r357, %r356, 1; mul.lo.s32 %r358, %r357, %r355; mul.lo.s32 %r359, %r358, 89287766; xor.b32 %r360, %r359, %r358; mul.lo.s32 %r361, %r360, 1403136100; xor.b32 %r362, %r361, %r360; brev.b32 %r363, %r362; mul.lo.s32 %r364, %r363, 1025551850; xor.b32 %r365, %r364, %r363; xor.b32 %r366, %r21, -2012890959; add.s32 %r367, %r365, %r366; shr.u32 %r368, %r366, 16; or.b32 %r369, %r368, 1; mul.lo.s32 %r370, %r367, %r369; mul.lo.s32 %r371, %r370, 89287766; xor.b32 %r372, %r371, %r370; mul.lo.s32 %r373, %r372, 1403136100; xor.b32 %r374, %r373, %r372; brev.b32 %r375, %r374; cvt.rn.f32.u32 %f991, %r375; mul.ftz.f32 %f3772, %f991, 0f2F7FFFFE; $L__BB203_31: cvt.ftz.sat.f32.f32 %f992, %f3762; mul.ftz.f32 %f993, %f992, 0f447FC000; cvt.rzi.ftz.s32.f32 %r376, %f993; min.s32 %r41, %r376, 1023; cvt.rn.f32.s32 %f994, %r41; sub.ftz.f32 %f40, %f993, %f994; ld.const.u64 %rd56, [kernel_params+2416]; cvta.to.global.u64 %rd6, %rd56; ld.const.u32 %r42, [kernel_params+996]; add.s32 %r377, %r41, %r42; mul.wide.s32 %rd57, %r377, 4; add.s64 %rd58, %rd6, %rd57; ld.global.f32 %f3773, [%rd58]; setp.eq.ftz.f32 %p24, %f40, 0f00000000; @%p24 bra $L__BB203_33; add.s32 %r378, %r41, 1; min.s32 %r379, %r378, 1023; add.s32 %r380, %r379, %r42; mul.wide.s32 %rd59, %r380, 4; add.s64 %rd60, %rd6, %rd59; mov.f32 %f995, 0f3F800000; sub.ftz.f32 %f996, %f995, %f40; ld.global.f32 %f997, [%rd60]; mul.ftz.f32 %f998, %f40, %f997; fma.rn.ftz.f32 %f3773, %f3773, %f996, %f998; $L__BB203_33: cvt.rn.f32.s32 %f999, %r13; add.ftz.f32 %f44, %f3773, %f999; cvt.rn.f32.s32 %f45, %r14; cvt.ftz.sat.f32.f32 %f1000, %f3763; mul.ftz.f32 %f1001, %f1000, 0f447FC000; cvt.rzi.ftz.s32.f32 %r381, %f1001; min.s32 %r43, %r381, 1023; cvt.rn.f32.s32 %f1002, %r43; sub.ftz.f32 %f46, %f1001, %f1002; add.s32 %r382, %r43, %r42; mul.wide.s32 %rd61, %r382, 4; add.s64 %rd62, %rd6, %rd61; ld.global.f32 %f3774, [%rd62]; setp.eq.ftz.f32 %p25, %f46, 0f00000000; @%p25 bra $L__BB203_35; add.s32 %r383, %r43, 1; min.s32 %r384, %r383, 1023; add.s32 %r385, %r384, %r42; mul.wide.s32 %rd63, %r385, 4; add.s64 %rd64, %rd6, %rd63; mov.f32 %f1003, 0f3F800000; sub.ftz.f32 %f1004, %f1003, %f46; ld.global.f32 %f1005, [%rd64]; mul.ftz.f32 %f1006, %f46, %f1005; fma.rn.ftz.f32 %f3774, %f3774, %f1004, %f1006; $L__BB203_35: ld.const.f32 %f3718, [kernel_params+256]; setp.eq.ftz.f32 %p239, %f3718, 0fBF800000; add.ftz.f32 %f50, %f3774, %f45; mov.f32 %f3776, 0f3F000000; @%p239 bra $L__BB203_41; bra.uni $L__BB203_36; $L__BB203_41: ld.const.u32 %r392, [kernel_params+16]; setp.eq.s32 %p30, %r392, 0; @%p30 bra $L__BB203_61; setp.ne.s32 %p31, %r392, 1; @%p31 bra $L__BB203_117; ld.const.v4.f32 {%f1026, %f1027, %f1028, %f1029}, [kernel_params+144]; ld.const.v4.f32 {%f1034, %f1035, %f1036, %f1037}, [kernel_params+160]; ld.const.v4.f32 {%f1042, %f1043, %f1044, %f1045}, [kernel_params+176]; ld.const.v4.f32 {%f1050, %f1051, %f1052, %f1053}, [kernel_params+192]; mul.ftz.f32 %f1058, %f50, %f1027; fma.rn.ftz.f32 %f1059, %f44, %f1026, %f1058; fma.rn.ftz.f32 %f1060, %f1028, 0f00000000, %f1029; add.ftz.f32 %f59, %f1059, %f1060; mul.ftz.f32 %f1061, %f50, %f1035; fma.rn.ftz.f32 %f1062, %f44, %f1034, %f1061; fma.rn.ftz.f32 %f1063, %f1036, 0f00000000, %f1037; add.ftz.f32 %f60, %f1062, %f1063; mul.ftz.f32 %f1064, %f50, %f1043; fma.rn.ftz.f32 %f1065, %f44, %f1042, %f1064; fma.rn.ftz.f32 %f1066, %f1044, 0f00000000, %f1045; add.ftz.f32 %f61, %f1065, %f1066; mul.ftz.f32 %f1067, %f50, %f1051; fma.rn.ftz.f32 %f1068, %f44, %f1050, %f1067; fma.rn.ftz.f32 %f1069, %f1052, 0f00000000, %f1053; add.ftz.f32 %f62, %f1068, %f1069; setp.eq.ftz.f32 %p32, %f62, 0f00000000; mov.f32 %f3777, 0f00000000; mov.f32 %f3778, %f3777; mov.f32 %f3779, %f3777; @%p32 bra $L__BB203_45; div.approx.ftz.f32 %f3779, %f61, %f62; div.approx.ftz.f32 %f3778, %f60, %f62; div.approx.ftz.f32 %f3777, %f59, %f62; $L__BB203_45: ld.const.f32 %f3714, [kernel_params+240]; setp.gt.ftz.f32 %p33, %f3714, 0f00000000; @%p33 bra $L__BB203_47; bra.uni $L__BB203_46; $L__BB203_47: ld.const.f32 %f75, [kernel_params+244]; setp.eq.ftz.f32 %p34, %f75, 0f00000000; @%p34 bra $L__BB203_49; bra.uni $L__BB203_48; $L__BB203_49: fma.rn.ftz.f32 %f78, %f34, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f3781, %f33, 0f40000000, 0fBF800000; setp.eq.ftz.f32 %p35, %f78, 0f00000000; setp.eq.ftz.f32 %p36, %f3781, 0f00000000; and.pred %p37, %p36, %p35; mov.f32 %f3784, 0f00000000; mov.f32 %f3785, %f3784; @%p37 bra $L__BB203_54; mul.ftz.f32 %f1098, %f78, %f78; mul.ftz.f32 %f1099, %f3781, %f3781; setp.gt.ftz.f32 %p38, %f1098, %f1099; @%p38 bra $L__BB203_52; bra.uni $L__BB203_51; $L__BB203_52: div.approx.ftz.f32 %f1101, %f3781, %f78; mul.ftz.f32 %f3780, %f1101, 0f3F490FDB; mov.f32 %f3781, %f78; bra.uni $L__BB203_53; $L__BB203_36: ld.const.u32 %r44, [kernel_params+960]; cvt.ftz.sat.f32.f32 %f1008, %f3772; mul.ftz.f32 %f1009, %f1008, 0f437F0000; cvt.rzi.ftz.s32.f32 %r386, %f1009; min.s32 %r45, %r386, 255; cvt.rn.f32.s32 %f1010, %r45; sub.ftz.f32 %f51, %f1009, %f1010; add.s32 %r387, %r45, %r44; mul.wide.s32 %rd65, %r387, 4; add.s64 %rd66, %rd6, %rd65; ld.global.f32 %f3775, [%rd66]; setp.eq.ftz.f32 %p27, %f51, 0f00000000; @%p27 bra $L__BB203_38; add.s32 %r388, %r45, 1; min.s32 %r389, %r388, 255; add.s32 %r390, %r389, %r44; mul.wide.s32 %rd67, %r390, 4; add.s64 %rd68, %rd6, %rd67; mov.f32 %f1011, 0f3F800000; sub.ftz.f32 %f1012, %f1011, %f51; ld.global.f32 %f1013, [%rd68]; mul.ftz.f32 %f1014, %f51, %f1013; fma.rn.ftz.f32 %f3775, %f3775, %f1012, %f1014; $L__BB203_38: ld.const.u32 %r391, [kernel_params+964]; setp.eq.s32 %p28, %r391, 0; mov.f32 %f3776, %f3775; @%p28 bra $L__BB203_41; ld.const.f32 %f1015, [kernel_params+288]; div.approx.ftz.f32 %f1016, %f45, %f1015; mov.f32 %f1017, 0f3F800000; sub.ftz.f32 %f3776, %f1017, %f1016; ld.const.f32 %f56, [kernel_params+968]; setp.eq.ftz.f32 %p29, %f56, 0f00000000; @%p29 bra $L__BB203_41; add.ftz.f32 %f1018, %f3775, 0fBF000000; sub.ftz.f32 %f1020, %f1017, %f56; add.ftz.f32 %f1021, %f3776, 0fBF000000; fma.rn.ftz.f32 %f1022, %f1021, %f1020, 0f3F000000; fma.rn.ftz.f32 %f3776, %f1018, %f56, %f1022; bra.uni $L__BB203_41; $L__BB203_61: ld.const.v4.f32 {%f1361, %f1362, %f1363, %f1364}, [kernel_params+144]; ld.const.v4.f32 {%f1365, %f1366, %f1367, %f1368}, [kernel_params+160]; ld.const.v4.f32 {%f1369, %f1370, %f1371, %f1372}, [kernel_params+176]; ld.const.v4.f32 {%f1373, %f1374, %f1375, %f1376}, [kernel_params+192]; mul.ftz.f32 %f187, %f44, %f1361; mul.ftz.f32 %f188, %f50, %f1362; add.ftz.f32 %f189, %f187, %f188; fma.rn.ftz.f32 %f190, %f1363, 0f00000000, %f1364; mul.ftz.f32 %f191, %f44, %f1365; mul.ftz.f32 %f192, %f50, %f1366; add.ftz.f32 %f193, %f191, %f192; fma.rn.ftz.f32 %f194, %f1367, 0f00000000, %f1368; mul.ftz.f32 %f195, %f44, %f1369; mul.ftz.f32 %f196, %f50, %f1370; add.ftz.f32 %f197, %f195, %f196; fma.rn.ftz.f32 %f198, %f1371, 0f00000000, %f1372; mul.ftz.f32 %f199, %f44, %f1373; mul.ftz.f32 %f200, %f50, %f1374; add.ftz.f32 %f1377, %f199, %f200; fma.rn.ftz.f32 %f201, %f1375, 0f00000000, %f1376; add.ftz.f32 %f202, %f1377, %f201; setp.eq.ftz.f32 %p43, %f202, 0f00000000; mov.f32 %f3817, 0f00000000; mov.f32 %f3818, %f3817; mov.f32 %f3819, %f3817; @%p43 bra $L__BB203_63; add.ftz.f32 %f1378, %f189, %f190; add.ftz.f32 %f1379, %f193, %f194; add.ftz.f32 %f1380, %f197, %f198; div.approx.ftz.f32 %f3819, %f1380, %f202; div.approx.ftz.f32 %f3818, %f1379, %f202; div.approx.ftz.f32 %f3817, %f1378, %f202; $L__BB203_63: ld.const.u32 %r397, [kernel_params+264]; setp.eq.s32 %p44, %r397, 0; @%p44 bra $L__BB203_71; setp.lt.ftz.f32 %p45, %f3776, 0f3F000000; @%p45 bra $L__BB203_68; bra.uni $L__BB203_65; $L__BB203_68: ld.const.v4.f32 {%f1436, %f1437, %f1438, %f1439}, [kernel_params+736]; ld.const.v4.f32 {%f1444, %f1445, %f1446, %f1447}, [kernel_params+752]; ld.const.v4.f32 {%f1452, %f1453, %f1454, %f1455}, [kernel_params+768]; ld.const.v4.f32 {%f1460, %f1461, %f1462, %f1463}, [kernel_params+784]; mul.ftz.f32 %f1468, %f50, %f1437; fma.rn.ftz.f32 %f1469, %f44, %f1436, %f1468; fma.rn.ftz.f32 %f1470, %f1438, 0f00000000, %f1439; add.ftz.f32 %f225, %f1469, %f1470; mul.ftz.f32 %f1471, %f50, %f1445; fma.rn.ftz.f32 %f1472, %f44, %f1444, %f1471; fma.rn.ftz.f32 %f1473, %f1446, 0f00000000, %f1447; add.ftz.f32 %f226, %f1472, %f1473; mul.ftz.f32 %f1474, %f50, %f1453; fma.rn.ftz.f32 %f1475, %f44, %f1452, %f1474; fma.rn.ftz.f32 %f1476, %f1454, 0f00000000, %f1455; add.ftz.f32 %f227, %f1475, %f1476; mul.ftz.f32 %f1477, %f50, %f1461; fma.rn.ftz.f32 %f1478, %f44, %f1460, %f1477; fma.rn.ftz.f32 %f1479, %f1462, 0f00000000, %f1463; add.ftz.f32 %f228, %f1478, %f1479; setp.eq.ftz.f32 %p47, %f228, 0f00000000; mov.f32 %f3814, 0f00000000; mov.f32 %f3815, %f3814; mov.f32 %f3816, %f3814; @%p47 bra $L__BB203_70; div.approx.ftz.f32 %f3816, %f227, %f228; div.approx.ftz.f32 %f3815, %f226, %f228; div.approx.ftz.f32 %f3814, %f225, %f228; $L__BB203_70: sub.ftz.f32 %f1480, %f3817, %f3814; sub.ftz.f32 %f1481, %f3818, %f3815; sub.ftz.f32 %f1482, %f3819, %f3816; add.ftz.f32 %f1483, %f3776, %f3776; fma.rn.ftz.f32 %f3817, %f1483, %f1480, %f3814; fma.rn.ftz.f32 %f3818, %f1483, %f1481, %f3815; fma.rn.ftz.f32 %f3819, %f1483, %f1482, %f3816; bra.uni $L__BB203_71; $L__BB203_117: ld.const.v4.f32 {%f2147, %f2148, %f2149, %f2150}, [kernel_params+144]; ld.const.v4.f32 {%f2153, %f2154, %f2155, %f2156}, [kernel_params+160]; ld.const.v4.f32 {%f2159, %f2160, %f2161, %f2162}, [kernel_params+176]; ld.const.v4.f32 {%f2165, %f2166, %f2167, %f2168}, [kernel_params+192]; mul.ftz.f32 %f466, %f44, %f2147; mul.ftz.f32 %f467, %f50, %f2148; add.ftz.f32 %f468, %f466, %f467; fma.rn.ftz.f32 %f469, %f2149, 0f00000000, %f2150; mul.ftz.f32 %f470, %f44, %f2153; mul.ftz.f32 %f471, %f50, %f2154; add.ftz.f32 %f472, %f470, %f471; fma.rn.ftz.f32 %f473, %f2155, 0f00000000, %f2156; mul.ftz.f32 %f474, %f44, %f2159; mul.ftz.f32 %f475, %f50, %f2160; add.ftz.f32 %f476, %f474, %f475; fma.rn.ftz.f32 %f477, %f2161, 0f00000000, %f2162; mul.ftz.f32 %f478, %f44, %f2165; mul.ftz.f32 %f479, %f50, %f2166; add.ftz.f32 %f2171, %f478, %f479; fma.rn.ftz.f32 %f480, %f2167, 0f00000000, %f2168; add.ftz.f32 %f481, %f2171, %f480; setp.eq.ftz.f32 %p85, %f481, 0f00000000; mov.f32 %f3943, 0f00000000; mov.f32 %f3944, %f3943; mov.f32 %f3945, %f3943; @%p85 bra $L__BB203_119; add.ftz.f32 %f2172, %f468, %f469; add.ftz.f32 %f2173, %f472, %f473; add.ftz.f32 %f2174, %f476, %f477; div.approx.ftz.f32 %f3945, %f2174, %f481; div.approx.ftz.f32 %f3944, %f2173, %f481; div.approx.ftz.f32 %f3943, %f2172, %f481; $L__BB203_119: ld.const.u32 %r48, [kernel_params+20]; add.ftz.f32 %f489, %f3943, 0fBF000000; add.ftz.f32 %f491, %f3944, 0fBF000000; setp.gt.s32 %p86, %r48, 2; @%p86 bra $L__BB203_126; bra.uni $L__BB203_120; $L__BB203_126: setp.eq.s32 %p87, %r48, 3; @%p87 bra $L__BB203_133; setp.eq.s32 %p88, %r48, 5; @%p88 bra $L__BB203_135; setp.ne.s32 %p89, %r48, 4; @%p89 bra $L__BB203_137; ld.const.f32 %f2178, [kernel_params+48]; ld.const.v4.f32 {%f2179, %f2180, %f2181, %f2182}, [kernel_params+64]; ld.const.f32 %f2187, [kernel_params+276]; mul.ftz.f32 %f492, %f489, %f2187; ld.const.f32 %f2188, [kernel_params+280]; mul.ftz.f32 %f493, %f491, %f2188; mul.ftz.f32 %f2189, %f493, %f493; fma.rn.ftz.f32 %f2190, %f492, %f492, %f2189; sqrt.approx.ftz.f32 %f494, %f2190; mul.ftz.f32 %f2191, %f494, %f494; mul.ftz.f32 %f2192, %f494, %f2191; mul.ftz.f32 %f2193, %f2191, %f2191; mul.ftz.f32 %f2194, %f2180, %f2191; fma.rn.ftz.f32 %f2195, %f2179, %f494, %f2194; mul.ftz.f32 %f2196, %f2182, %f2193; fma.rn.ftz.f32 %f2197, %f2181, %f2192, %f2196; add.ftz.f32 %f2198, %f2195, %f2197; add.ftz.f32 %f2199, %f2178, %f2198; neg.ftz.f32 %f495, %f2199; abs.ftz.f32 %f2200, %f495; ld.const.f32 %f2201, [kernel_params+24]; mul.ftz.f32 %f2202, %f2201, 0f3F000000; setp.gt.ftz.f32 %p92, %f2200, %f2202; mov.f32 %f3889, 0f00000000; mov.f32 %f3890, %f3889; mov.f32 %f3891, %f3889; @%p92 bra $L__BB203_141; setp.eq.ftz.f32 %p93, %f494, 0f00000000; mov.f32 %f3880, 0f00000000; @%p93 bra $L__BB203_132; div.approx.ftz.f32 %f3880, %f492, %f494; $L__BB203_132: mov.f32 %f2204, 0fBF800000; max.ftz.f32 %f2205, %f3880, %f2204; mov.f32 %f2206, 0f3F800000; min.ftz.f32 %f2207, %f2205, %f2206; abs.ftz.f32 %f2208, %f2207; sub.ftz.f32 %f2209, %f2206, %f2208; mul.ftz.f32 %f2210, %f2209, 0f3F000000; sqrt.approx.ftz.f32 %f2211, %f2210; setp.gt.ftz.f32 %p94, %f2208, 0f3F11EB85; selp.f32 %f2212, %f2211, %f2208, %p94; mul.ftz.f32 %f2213, %f2212, %f2212; mov.f32 %f2214, 0f3C94D2E9; mov.f32 %f2215, 0f3D53F941; fma.rn.ftz.f32 %f2216, %f2215, %f2213, %f2214; mov.f32 %f2217, 0f3D3F841F; fma.rn.ftz.f32 %f2218, %f2216, %f2213, %f2217; mov.f32 %f2219, 0f3D994929; fma.rn.ftz.f32 %f2220, %f2218, %f2213, %f2219; mov.f32 %f2221, 0f3E2AAB94; fma.rn.ftz.f32 %f2222, %f2220, %f2213, %f2221; mul.ftz.f32 %f2223, %f2213, %f2222; fma.rn.ftz.f32 %f2224, %f2223, %f2212, %f2212; add.ftz.f32 %f2225, %f2224, %f2224; mov.f32 %f2226, 0f3FC90FDB; sub.ftz.f32 %f2227, %f2226, %f2224; selp.f32 %f2228, %f2225, %f2227, %p94; setp.lt.ftz.f32 %p95, %f2207, 0f00000000; mov.f32 %f2229, 0f40490FDB; sub.ftz.f32 %f2230, %f2229, %f2228; selp.f32 %f2231, %f2230, %f2228, %p95; neg.ftz.f32 %f2232, %f2231; setp.lt.ftz.f32 %p96, %f493, 0f00000000; selp.f32 %f2233, %f2232, %f2231, %p96; cos.approx.ftz.f32 %f2234, %f2233; sin.approx.ftz.f32 %f2235, %f495; mul.ftz.f32 %f2236, %f2234, %f2235; sin.approx.ftz.f32 %f2237, %f2233; mul.ftz.f32 %f3891, %f2235, %f2237; neg.ftz.f32 %f3890, %f2236; cos.approx.ftz.f32 %f3889, %f495; bra.uni $L__BB203_141; $L__BB203_46: mov.f32 %f3788, 0f3F800000; mov.f32 %f3786, 0f00000000; mov.f32 %f3787, %f3786; bra.uni $L__BB203_55; $L__BB203_120: setp.eq.s32 %p90, %r48, 0; @%p90 bra $L__BB203_136; setp.eq.s32 %p91, %r48, 1; @%p91 bra $L__BB203_122; bra.uni $L__BB203_137; $L__BB203_122: add.ftz.f32 %f511, %f489, %f489; add.ftz.f32 %f512, %f491, %f491; mul.ftz.f32 %f2262, %f512, %f512; fma.rn.ftz.f32 %f2263, %f511, %f511, %f2262; sqrt.approx.ftz.f32 %f513, %f2263; setp.gt.ftz.f32 %p98, %f513, 0f3F800000; mov.f32 %f3889, 0f00000000; mov.f32 %f3890, %f3889; mov.f32 %f3891, %f3889; @%p98 bra $L__BB203_141; ld.const.f32 %f514, [kernel_params+24]; setp.eq.ftz.f32 %p99, %f513, 0f00000000; mov.f32 %f3881, 0f00000000; @%p99 bra $L__BB203_125; div.approx.ftz.f32 %f3881, %f511, %f513; $L__BB203_125: mov.f32 %f2265, 0fBF800000; max.ftz.f32 %f2266, %f3881, %f2265; mov.f32 %f2267, 0f3F800000; min.ftz.f32 %f2268, %f2266, %f2267; abs.ftz.f32 %f2269, %f2268; sub.ftz.f32 %f2270, %f2267, %f2269; mul.ftz.f32 %f2271, %f2270, 0f3F000000; sqrt.approx.ftz.f32 %f2272, %f2271; setp.gt.ftz.f32 %p100, %f2269, 0f3F11EB85; selp.f32 %f2273, %f2272, %f2269, %p100; mul.ftz.f32 %f2274, %f2273, %f2273; mov.f32 %f2275, 0f3C94D2E9; mov.f32 %f2276, 0f3D53F941; fma.rn.ftz.f32 %f2277, %f2276, %f2274, %f2275; mov.f32 %f2278, 0f3D3F841F; fma.rn.ftz.f32 %f2279, %f2277, %f2274, %f2278; mov.f32 %f2280, 0f3D994929; fma.rn.ftz.f32 %f2281, %f2279, %f2274, %f2280; mov.f32 %f2282, 0f3E2AAB94; fma.rn.ftz.f32 %f2283, %f2281, %f2274, %f2282; mul.ftz.f32 %f2284, %f2274, %f2283; fma.rn.ftz.f32 %f2285, %f2284, %f2273, %f2273; add.ftz.f32 %f2286, %f2285, %f2285; mov.f32 %f2287, 0f3FC90FDB; sub.ftz.f32 %f2288, %f2287, %f2285; selp.f32 %f2289, %f2286, %f2288, %p100; setp.lt.ftz.f32 %p101, %f2268, 0f00000000; mov.f32 %f2290, 0f40490FDB; sub.ftz.f32 %f2291, %f2290, %f2289; selp.f32 %f2292, %f2291, %f2289, %p101; mul.ftz.f32 %f2293, %f514, %f513; mul.ftz.f32 %f2294, %f2293, 0f3F000000; neg.ftz.f32 %f2295, %f2292; setp.lt.ftz.f32 %p102, %f512, 0f00000000; selp.f32 %f2296, %f2295, %f2292, %p102; cos.approx.ftz.f32 %f2297, %f2296; sin.approx.ftz.f32 %f2298, %f2294; mul.ftz.f32 %f2299, %f2297, %f2298; sin.approx.ftz.f32 %f2300, %f2296; mul.ftz.f32 %f3891, %f2298, %f2300; neg.ftz.f32 %f3890, %f2299; cos.approx.ftz.f32 %f3889, %f2294; bra.uni $L__BB203_141; $L__BB203_65: ld.const.v4.f32 {%f1384, %f1385, %f1386, %f1387}, [kernel_params+800]; ld.const.v4.f32 {%f1392, %f1393, %f1394, %f1395}, [kernel_params+816]; ld.const.v4.f32 {%f1400, %f1401, %f1402, %f1403}, [kernel_params+832]; ld.const.v4.f32 {%f1408, %f1409, %f1410, %f1411}, [kernel_params+848]; mul.ftz.f32 %f1416, %f50, %f1385; fma.rn.ftz.f32 %f1417, %f44, %f1384, %f1416; fma.rn.ftz.f32 %f1418, %f1386, 0f00000000, %f1387; add.ftz.f32 %f212, %f1417, %f1418; mul.ftz.f32 %f1419, %f50, %f1393; fma.rn.ftz.f32 %f1420, %f44, %f1392, %f1419; fma.rn.ftz.f32 %f1421, %f1394, 0f00000000, %f1395; add.ftz.f32 %f213, %f1420, %f1421; mul.ftz.f32 %f1422, %f50, %f1401; fma.rn.ftz.f32 %f1423, %f44, %f1400, %f1422; fma.rn.ftz.f32 %f1424, %f1402, 0f00000000, %f1403; add.ftz.f32 %f214, %f1423, %f1424; mul.ftz.f32 %f1425, %f50, %f1409; fma.rn.ftz.f32 %f1426, %f44, %f1408, %f1425; fma.rn.ftz.f32 %f1427, %f1410, 0f00000000, %f1411; add.ftz.f32 %f215, %f1426, %f1427; setp.eq.ftz.f32 %p46, %f215, 0f00000000; mov.f32 %f3811, 0f00000000; mov.f32 %f3812, %f3811; mov.f32 %f3813, %f3811; @%p46 bra $L__BB203_67; div.approx.ftz.f32 %f3813, %f214, %f215; div.approx.ftz.f32 %f3812, %f213, %f215; div.approx.ftz.f32 %f3811, %f212, %f215; $L__BB203_67: add.ftz.f32 %f1428, %f3776, 0fBF000000; add.ftz.f32 %f1429, %f1428, %f1428; sub.ftz.f32 %f1430, %f3811, %f3817; sub.ftz.f32 %f1431, %f3812, %f3818; sub.ftz.f32 %f1432, %f3813, %f3819; fma.rn.ftz.f32 %f3817, %f1429, %f1430, %f3817; fma.rn.ftz.f32 %f3818, %f1429, %f1431, %f3818; fma.rn.ftz.f32 %f3819, %f1429, %f1432, %f3819; $L__BB203_71: ld.const.f32 %f3716, [kernel_params+240]; setp.leu.ftz.f32 %p238, %f3716, 0f00000000; mov.f32 %f3829, 0f00000000; mov.f32 %f3826, %f3817; mov.f32 %f3827, %f3818; mov.f32 %f3828, %f3819; mov.f32 %f3830, %f3829; @%p238 bra $L__BB203_80; ld.const.f32 %f241, [kernel_params+244]; setp.eq.ftz.f32 %p49, %f241, 0f00000000; @%p49 bra $L__BB203_74; bra.uni $L__BB203_73; $L__BB203_74: fma.rn.ftz.f32 %f244, %f34, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f3821, %f33, 0f40000000, 0fBF800000; setp.eq.ftz.f32 %p50, %f244, 0f00000000; setp.eq.ftz.f32 %p51, %f3821, 0f00000000; and.pred %p52, %p51, %p50; mov.f32 %f3824, 0f00000000; mov.f32 %f3825, %f3824; @%p52 bra $L__BB203_79; mul.ftz.f32 %f1511, %f244, %f244; mul.ftz.f32 %f1512, %f3821, %f3821; setp.gt.ftz.f32 %p53, %f1511, %f1512; @%p53 bra $L__BB203_77; bra.uni $L__BB203_76; $L__BB203_77: div.approx.ftz.f32 %f1514, %f3821, %f244; mul.ftz.f32 %f3820, %f1514, 0f3F490FDB; mov.f32 %f3821, %f244; bra.uni $L__BB203_78; $L__BB203_48: ld.const.f32 %f1073, [kernel_params+248]; mul.ftz.f32 %f1074, %f34, %f75; cvt.rmi.ftz.f32.f32 %f1075, %f1074; sub.ftz.f32 %f1076, %f1074, %f1075; sqrt.approx.ftz.f32 %f1077, %f1076; mul.ftz.f32 %f1078, %f33, %f1077; mov.f32 %f1079, 0f3F800000; sub.ftz.f32 %f1080, %f1079, %f1077; add.ftz.f32 %f1081, %f1080, %f1078; mov.f32 %f1082, 0f40490FDB; div.approx.ftz.f32 %f1083, %f1082, %f75; cos.approx.ftz.f32 %f1084, %f1083; mul.ftz.f32 %f1085, %f1081, %f1084; sub.ftz.f32 %f1086, %f1080, %f1078; sin.approx.ftz.f32 %f1087, %f1083; mul.ftz.f32 %f1088, %f1086, %f1087; add.ftz.f32 %f1089, %f1075, %f1075; fma.rn.ftz.f32 %f1090, %f1089, %f1083, %f1073; cos.approx.ftz.f32 %f1091, %f1090; sin.approx.ftz.f32 %f1092, %f1090; mul.ftz.f32 %f1093, %f1085, %f1091; mul.ftz.f32 %f1094, %f1088, %f1092; sub.ftz.f32 %f3784, %f1093, %f1094; mul.ftz.f32 %f1095, %f1085, %f1092; fma.rn.ftz.f32 %f3785, %f1091, %f1088, %f1095; bra.uni $L__BB203_54; $L__BB203_73: ld.const.f32 %f1486, [kernel_params+248]; mul.ftz.f32 %f1487, %f34, %f241; cvt.rmi.ftz.f32.f32 %f1488, %f1487; sub.ftz.f32 %f1489, %f1487, %f1488; sqrt.approx.ftz.f32 %f1490, %f1489; mul.ftz.f32 %f1491, %f33, %f1490; mov.f32 %f1492, 0f3F800000; sub.ftz.f32 %f1493, %f1492, %f1490; add.ftz.f32 %f1494, %f1493, %f1491; mov.f32 %f1495, 0f40490FDB; div.approx.ftz.f32 %f1496, %f1495, %f241; cos.approx.ftz.f32 %f1497, %f1496; mul.ftz.f32 %f1498, %f1494, %f1497; sub.ftz.f32 %f1499, %f1493, %f1491; sin.approx.ftz.f32 %f1500, %f1496; mul.ftz.f32 %f1501, %f1499, %f1500; add.ftz.f32 %f1502, %f1488, %f1488; fma.rn.ftz.f32 %f1503, %f1502, %f1496, %f1486; cos.approx.ftz.f32 %f1504, %f1503; sin.approx.ftz.f32 %f1505, %f1503; mul.ftz.f32 %f1506, %f1498, %f1504; mul.ftz.f32 %f1507, %f1501, %f1505; sub.ftz.f32 %f3824, %f1506, %f1507; mul.ftz.f32 %f1508, %f1498, %f1505; fma.rn.ftz.f32 %f3825, %f1504, %f1501, %f1508; bra.uni $L__BB203_79; $L__BB203_133: fma.rn.ftz.f32 %f501, %f3943, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f502, %f3944, 0f40000000, 0fBF800000; mul.ftz.f32 %f503, %f501, %f501; mul.ftz.f32 %f504, %f502, %f502; add.ftz.f32 %f2241, %f503, %f504; setp.gt.ftz.f32 %p97, %f2241, 0f3F800000; mov.f32 %f3889, 0f00000000; mov.f32 %f3890, %f3889; mov.f32 %f3891, %f3889; @%p97 bra $L__BB203_141; mov.f32 %f2242, 0f3F800000; sub.ftz.f32 %f2243, %f2242, %f503; sub.ftz.f32 %f2244, %f2243, %f504; mov.f32 %f2245, 0f00000000; max.ftz.f32 %f2246, %f2244, %f2245; sqrt.approx.ftz.f32 %f2247, %f2246; fma.rn.ftz.f32 %f2248, %f501, 0f00000000, %f2247; fma.rn.ftz.f32 %f2249, %f502, 0f00000000, %f2248; add.ftz.f32 %f2250, %f2249, %f2249; mul.ftz.f32 %f2251, %f2247, %f2250; sub.ftz.f32 %f3890, %f2242, %f2251; mul.ftz.f32 %f3891, %f502, %f2250; mul.ftz.f32 %f3889, %f501, %f2250; bra.uni $L__BB203_141; $L__BB203_135: mov.f32 %f3889, 0f3F800000; sub.ftz.f32 %f2252, %f3889, %f3943; fma.rn.ftz.f32 %f2253, %f2252, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f2254, %f2253; cos.approx.ftz.f32 %f2255, %f2253; fma.rn.ftz.f32 %f2256, %f3944, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f2257, %f2256; cos.approx.ftz.f32 %f2258, %f2256; div.approx.ftz.f32 %f3890, %f2254, %f2255; div.approx.ftz.f32 %f3891, %f2257, %f2258; bra.uni $L__BB203_141; $L__BB203_136: ld.const.v4.f32 {%f2301, %f2302, %f2303, %f2304}, [kernel_params+32]; fma.rn.ftz.f32 %f2309, %f3943, %f2301, %f2302; fma.rn.ftz.f32 %f2310, %f3944, %f2303, %f2304; sin.approx.ftz.f32 %f2311, %f2310; cos.approx.ftz.f32 %f2312, %f2309; sin.approx.ftz.f32 %f2313, %f2309; mul.ftz.f32 %f3890, %f2311, %f2313; mul.ftz.f32 %f3889, %f2311, %f2312; cos.approx.ftz.f32 %f3891, %f2310; bra.uni $L__BB203_141; $L__BB203_137: ld.const.v2.f32 {%f2317, %f2318}, [kernel_params+24]; ld.const.f32 %f2321, [kernel_params+276]; mul.ftz.f32 %f529, %f489, %f2321; ld.const.f32 %f2322, [kernel_params+280]; mul.ftz.f32 %f530, %f491, %f2322; add.ftz.f32 %f531, %f2318, %f2318; mul.ftz.f32 %f2323, %f2317, 0f3E800000; sin.approx.ftz.f32 %f2324, %f2323; mul.ftz.f32 %f2325, %f531, %f2324; mul.ftz.f32 %f2326, %f530, %f530; fma.rn.ftz.f32 %f2327, %f529, %f529, %f2326; sqrt.approx.ftz.f32 %f532, %f2327; setp.gt.ftz.f32 %p103, %f532, %f2325; mov.f32 %f3889, 0f00000000; mov.f32 %f3890, %f3889; mov.f32 %f3891, %f3889; @%p103 bra $L__BB203_141; setp.eq.ftz.f32 %p104, %f532, 0f00000000; mov.f32 %f3885, 0f00000000; @%p104 bra $L__BB203_140; div.approx.ftz.f32 %f3885, %f529, %f532; $L__BB203_140: mov.f32 %f2329, 0fBF800000; max.ftz.f32 %f2330, %f3885, %f2329; mov.f32 %f2331, 0f3F800000; min.ftz.f32 %f2332, %f2330, %f2331; abs.ftz.f32 %f2333, %f2332; sub.ftz.f32 %f2334, %f2331, %f2333; mul.ftz.f32 %f2335, %f2334, 0f3F000000; sqrt.approx.ftz.f32 %f2336, %f2335; setp.gt.ftz.f32 %p105, %f2333, 0f3F11EB85; selp.f32 %f2337, %f2336, %f2333, %p105; mul.ftz.f32 %f2338, %f2337, %f2337; mov.f32 %f2339, 0f3C94D2E9; mov.f32 %f2340, 0f3D53F941; fma.rn.ftz.f32 %f2341, %f2340, %f2338, %f2339; mov.f32 %f2342, 0f3D3F841F; fma.rn.ftz.f32 %f2343, %f2341, %f2338, %f2342; mov.f32 %f2344, 0f3D994929; fma.rn.ftz.f32 %f2345, %f2343, %f2338, %f2344; mov.f32 %f2346, 0f3E2AAB94; fma.rn.ftz.f32 %f2347, %f2345, %f2338, %f2346; mul.ftz.f32 %f2348, %f2338, %f2347; fma.rn.ftz.f32 %f2349, %f2348, %f2337, %f2337; add.ftz.f32 %f2350, %f2349, %f2349; mov.f32 %f2351, 0f3FC90FDB; sub.ftz.f32 %f2352, %f2351, %f2349; selp.f32 %f2353, %f2350, %f2352, %p105; setp.lt.ftz.f32 %p106, %f2332, 0f00000000; mov.f32 %f2354, 0f40490FDB; sub.ftz.f32 %f2355, %f2354, %f2353; selp.f32 %f2356, %f2355, %f2353, %p106; div.approx.ftz.f32 %f2357, %f532, %f531; abs.ftz.f32 %f2358, %f2357; sub.ftz.f32 %f2359, %f2331, %f2358; mul.ftz.f32 %f2360, %f2359, 0f3F000000; sqrt.approx.ftz.f32 %f2361, %f2360; setp.gt.ftz.f32 %p107, %f2358, 0f3F11EB85; selp.f32 %f2362, %f2361, %f2358, %p107; mul.ftz.f32 %f2363, %f2362, %f2362; fma.rn.ftz.f32 %f2364, %f2340, %f2363, %f2339; fma.rn.ftz.f32 %f2365, %f2364, %f2363, %f2342; fma.rn.ftz.f32 %f2366, %f2365, %f2363, %f2344; fma.rn.ftz.f32 %f2367, %f2366, %f2363, %f2346; mul.ftz.f32 %f2368, %f2363, %f2367; fma.rn.ftz.f32 %f2369, %f2368, %f2362, %f2362; mov.f32 %f2370, 0fC0000000; fma.rn.ftz.f32 %f2371, %f2370, %f2369, %f2351; selp.f32 %f2372, %f2371, %f2369, %p107; setp.le.ftz.f32 %p108, %f2372, 0f7F800000; mov.b32 %r418, %f2372; mov.b32 %r419, %f2357; and.b32 %r420, %r419, -2147483648; or.b32 %r421, %r420, %r418; mov.b32 %f2373, %r421; selp.f32 %f2374, %f2373, %f2372, %p108; add.ftz.f32 %f2375, %f2374, %f2374; neg.ftz.f32 %f2376, %f2356; setp.lt.ftz.f32 %p109, %f530, 0f00000000; selp.f32 %f2377, %f2376, %f2356, %p109; cos.approx.ftz.f32 %f2378, %f2377; sin.approx.ftz.f32 %f2379, %f2375; mul.ftz.f32 %f2380, %f2378, %f2379; sin.approx.ftz.f32 %f2381, %f2377; mul.ftz.f32 %f3891, %f2379, %f2381; neg.ftz.f32 %f3890, %f2380; cos.approx.ftz.f32 %f3889, %f2375; $L__BB203_141: setp.eq.ftz.f32 %p110, %f3889, 0f00000000; mov.f32 %f3992, 0f00000000; setp.eq.ftz.f32 %p111, %f3890, 0f00000000; and.pred %p112, %p110, %p111; setp.eq.ftz.f32 %p113, %f3891, 0f00000000; and.pred %p114, %p113, %p112; @%p114 bra $L__BB203_255; ld.const.f32 %f3694, [kernel_params+240]; setp.leu.ftz.f32 %p233, %f3694, 0f00000000; mov.f32 %f3924, 0f00000000; mov.f32 %f3925, %f3924; mov.f32 %f3926, %f3924; @%p233 bra $L__BB203_151; ld.const.f32 %f550, [kernel_params+244]; setp.eq.ftz.f32 %p116, %f550, 0f00000000; @%p116 bra $L__BB203_145; bra.uni $L__BB203_144; $L__BB203_145: fma.rn.ftz.f32 %f553, %f34, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f3893, %f33, 0f40000000, 0fBF800000; setp.eq.ftz.f32 %p117, %f553, 0f00000000; setp.eq.ftz.f32 %p118, %f3893, 0f00000000; and.pred %p119, %p118, %p117; mov.f32 %f3896, 0f00000000; mov.f32 %f3897, %f3896; @%p119 bra $L__BB203_150; mul.ftz.f32 %f2412, %f553, %f553; mul.ftz.f32 %f2413, %f3893, %f3893; setp.gt.ftz.f32 %p120, %f2412, %f2413; @%p120 bra $L__BB203_148; bra.uni $L__BB203_147; $L__BB203_148: div.approx.ftz.f32 %f2415, %f3893, %f553; mul.ftz.f32 %f3892, %f2415, 0f3F490FDB; mov.f32 %f3893, %f553; bra.uni $L__BB203_149; $L__BB203_144: ld.const.f32 %f2387, [kernel_params+248]; mul.ftz.f32 %f2388, %f34, %f550; cvt.rmi.ftz.f32.f32 %f2389, %f2388; sub.ftz.f32 %f2390, %f2388, %f2389; sqrt.approx.ftz.f32 %f2391, %f2390; mul.ftz.f32 %f2392, %f33, %f2391; mov.f32 %f2393, 0f3F800000; sub.ftz.f32 %f2394, %f2393, %f2391; add.ftz.f32 %f2395, %f2394, %f2392; mov.f32 %f2396, 0f40490FDB; div.approx.ftz.f32 %f2397, %f2396, %f550; cos.approx.ftz.f32 %f2398, %f2397; mul.ftz.f32 %f2399, %f2395, %f2398; sub.ftz.f32 %f2400, %f2394, %f2392; sin.approx.ftz.f32 %f2401, %f2397; mul.ftz.f32 %f2402, %f2400, %f2401; add.ftz.f32 %f2403, %f2389, %f2389; fma.rn.ftz.f32 %f2404, %f2403, %f2397, %f2387; cos.approx.ftz.f32 %f2405, %f2404; sin.approx.ftz.f32 %f2406, %f2404; mul.ftz.f32 %f2407, %f2399, %f2405; mul.ftz.f32 %f2408, %f2402, %f2406; sub.ftz.f32 %f3896, %f2407, %f2408; mul.ftz.f32 %f2409, %f2399, %f2406; fma.rn.ftz.f32 %f3897, %f2405, %f2402, %f2409; bra.uni $L__BB203_150; $L__BB203_51: div.approx.ftz.f32 %f1100, %f78, %f3781; fma.rn.ftz.f32 %f3780, %f1100, 0fBF490FDB, 0f3FC90FDB; $L__BB203_53: cos.approx.ftz.f32 %f1102, %f3780; sin.approx.ftz.f32 %f1103, %f3780; mul.ftz.f32 %f3785, %f3781, %f1103; mul.ftz.f32 %f3784, %f3781, %f1102; $L__BB203_54: ld.const.f32 %f3715, [kernel_params+240]; ld.const.f32 %f1104, [kernel_params+296]; mul.ftz.f32 %f1105, %f3784, %f1104; mul.ftz.f32 %f1106, %f3715, %f1105; ld.const.f32 %f1107, [kernel_params+252]; mul.ftz.f32 %f1108, %f1107, 0f00000000; mul.ftz.f32 %f1109, %f3715, %f3785; add.ftz.f32 %f3778, %f3778, %f1109; add.ftz.f32 %f3777, %f3777, %f1106; add.ftz.f32 %f3779, %f3779, 0f00000000; sub.ftz.f32 %f1110, %f1108, %f1106; sub.ftz.f32 %f1111, %f1108, %f1109; mul.ftz.f32 %f1112, %f1111, %f1111; fma.rn.ftz.f32 %f1113, %f1110, %f1110, %f1112; fma.rn.ftz.f32 %f1114, %f1107, %f1107, %f1113; sqrt.approx.ftz.f32 %f1115, %f1114; div.approx.ftz.f32 %f3786, %f1110, %f1115; div.approx.ftz.f32 %f3787, %f1111, %f1115; div.approx.ftz.f32 %f3788, %f1107, %f1115; $L__BB203_55: ld.const.v4.f32 {%f3796, %f3797, %f3798, %f3799}, [kernel_params+96]; ld.const.v4.f32 {%f3800, %f3801, %f3802, %f3803}, [kernel_params+112]; ld.const.v4.f32 {%f3804, %f3805, %f3806, %f3807}, [kernel_params+128]; ld.const.u32 %r46, [kernel_params+260]; setp.eq.s32 %p39, %r46, 0; @%p39 bra $L__BB203_60; ld.const.u64 %rd69, [kernel_params+2208]; cvta.to.global.u64 %rd70, %rd69; add.s32 %r393, %r46, -1; cvt.rn.f32.s32 %f1128, %r393; mul.ftz.f32 %f1129, %f3776, %f1128; cvt.rzi.ftz.s32.f32 %r394, %f1129; add.s32 %r395, %r46, -2; min.s32 %r396, %r394, %r395; cvt.rn.f32.s32 %f1130, %r396; sub.ftz.f32 %f116, %f1129, %f1130; mul.wide.s32 %rd71, %r396, 64; add.s64 %rd7, %rd70, %rd71; ld.global.v4.f32 {%f1131, %f1132, %f1133, %f1134}, [%rd7]; ld.global.v4.f32 {%f1135, %f1136, %f1137, %f1138}, [%rd7+64]; mul.ftz.f32 %f1139, %f1132, %f1136; fma.rn.ftz.f32 %f1140, %f1131, %f1135, %f1139; mul.ftz.f32 %f1141, %f1134, %f1138; fma.rn.ftz.f32 %f1142, %f1133, %f1137, %f1141; add.ftz.f32 %f125, %f1140, %f1142; setp.gt.ftz.f32 %p40, %f125, 0f3F7FDF3B; mov.f32 %f1143, 0f3F800000; sub.ftz.f32 %f126, %f1143, %f116; @%p40 bra $L__BB203_58; bra.uni $L__BB203_57; $L__BB203_58: mul.ftz.f32 %f1197, %f1131, %f126; fma.rn.ftz.f32 %f1198, %f1135, %f116, %f1197; mul.ftz.f32 %f1199, %f1132, %f126; fma.rn.ftz.f32 %f1200, %f1136, %f116, %f1199; mul.ftz.f32 %f1201, %f1133, %f126; fma.rn.ftz.f32 %f1202, %f1137, %f116, %f1201; mul.ftz.f32 %f1203, %f1134, %f126; fma.rn.ftz.f32 %f1204, %f1138, %f116, %f1203; mul.ftz.f32 %f1205, %f1200, %f1200; fma.rn.ftz.f32 %f1206, %f1198, %f1198, %f1205; mul.ftz.f32 %f1207, %f1204, %f1204; fma.rn.ftz.f32 %f1208, %f1202, %f1202, %f1207; add.ftz.f32 %f1209, %f1206, %f1208; rsqrt.approx.ftz.f32 %f1210, %f1209; mul.ftz.f32 %f3795, %f1204, %f1210; mul.ftz.f32 %f3794, %f1202, %f1210; mul.ftz.f32 %f3793, %f1200, %f1210; mul.ftz.f32 %f3792, %f1198, %f1210; bra.uni $L__BB203_59; $L__BB203_57: mov.f32 %f1144, 0fBF800000; max.ftz.f32 %f1145, %f125, %f1144; min.ftz.f32 %f1147, %f1145, %f1143; abs.ftz.f32 %f1148, %f1147; sub.ftz.f32 %f1149, %f1143, %f1148; mul.ftz.f32 %f1150, %f1149, 0f3F000000; sqrt.approx.ftz.f32 %f1151, %f1150; setp.gt.ftz.f32 %p41, %f1148, 0f3F11EB85; selp.f32 %f1152, %f1151, %f1148, %p41; mul.ftz.f32 %f1153, %f1152, %f1152; mov.f32 %f1154, 0f3C94D2E9; mov.f32 %f1155, 0f3D53F941; fma.rn.ftz.f32 %f1156, %f1155, %f1153, %f1154; mov.f32 %f1157, 0f3D3F841F; fma.rn.ftz.f32 %f1158, %f1156, %f1153, %f1157; mov.f32 %f1159, 0f3D994929; fma.rn.ftz.f32 %f1160, %f1158, %f1153, %f1159; mov.f32 %f1161, 0f3E2AAB94; fma.rn.ftz.f32 %f1162, %f1160, %f1153, %f1161; mul.ftz.f32 %f1163, %f1153, %f1162; fma.rn.ftz.f32 %f1164, %f1163, %f1152, %f1152; add.ftz.f32 %f1165, %f1164, %f1164; mov.f32 %f1166, 0f3FC90FDB; sub.ftz.f32 %f1167, %f1166, %f1164; selp.f32 %f1168, %f1165, %f1167, %p41; setp.lt.ftz.f32 %p42, %f1147, 0f00000000; mov.f32 %f1169, 0f40490FDB; sub.ftz.f32 %f1170, %f1169, %f1168; selp.f32 %f1171, %f1170, %f1168, %p42; mul.ftz.f32 %f1172, %f1131, %f125; sub.ftz.f32 %f1173, %f1135, %f1172; mul.ftz.f32 %f1174, %f1132, %f125; sub.ftz.f32 %f1175, %f1136, %f1174; mul.ftz.f32 %f1176, %f1133, %f125; sub.ftz.f32 %f1177, %f1137, %f1176; mul.ftz.f32 %f1178, %f1134, %f125; sub.ftz.f32 %f1179, %f1138, %f1178; mul.ftz.f32 %f1180, %f1175, %f1175; fma.rn.ftz.f32 %f1181, %f1173, %f1173, %f1180; mul.ftz.f32 %f1182, %f1179, %f1179; fma.rn.ftz.f32 %f1183, %f1177, %f1177, %f1182; add.ftz.f32 %f1184, %f1181, %f1183; rsqrt.approx.ftz.f32 %f1185, %f1184; mul.ftz.f32 %f1186, %f1173, %f1185; mul.ftz.f32 %f1187, %f1175, %f1185; mul.ftz.f32 %f1188, %f1177, %f1185; mul.ftz.f32 %f1189, %f1179, %f1185; mul.ftz.f32 %f1190, %f116, %f1171; cos.approx.ftz.f32 %f1191, %f1190; sin.approx.ftz.f32 %f1192, %f1190; mul.ftz.f32 %f1193, %f1186, %f1192; mul.ftz.f32 %f1194, %f1187, %f1192; mul.ftz.f32 %f1195, %f1188, %f1192; mul.ftz.f32 %f1196, %f1189, %f1192; fma.rn.ftz.f32 %f3795, %f1134, %f1191, %f1196; fma.rn.ftz.f32 %f3794, %f1133, %f1191, %f1195; fma.rn.ftz.f32 %f3793, %f1132, %f1191, %f1194; fma.rn.ftz.f32 %f3792, %f1131, %f1191, %f1193; $L__BB203_59: ld.global.v4.f32 {%f1211, %f1212, %f1213, %f1214}, [%rd7+16]; ld.global.v4.f32 {%f1219, %f1220, %f1221, %f1222}, [%rd7+80]; mul.ftz.f32 %f1227, %f116, %f1219; mul.ftz.f32 %f1228, %f116, %f1220; mul.ftz.f32 %f1229, %f116, %f1221; mul.ftz.f32 %f1230, %f116, %f1222; fma.rn.ftz.f32 %f3799, %f126, %f1211, %f1227; fma.rn.ftz.f32 %f3803, %f126, %f1212, %f1228; fma.rn.ftz.f32 %f3807, %f126, %f1213, %f1229; fma.rn.ftz.f32 %f1231, %f126, %f1214, %f1230; ld.global.v4.f32 {%f1232, %f1233, %f1234, %f1235}, [%rd7+32]; ld.global.v4.f32 {%f1240, %f1241, %f1242, %f1243}, [%rd7+96]; mul.ftz.f32 %f1248, %f116, %f1240; mul.ftz.f32 %f1249, %f116, %f1241; mul.ftz.f32 %f1250, %f116, %f1242; mul.ftz.f32 %f1251, %f116, %f1243; fma.rn.ftz.f32 %f1252, %f126, %f1232, %f1248; fma.rn.ftz.f32 %f1253, %f126, %f1233, %f1249; fma.rn.ftz.f32 %f1254, %f126, %f1234, %f1250; fma.rn.ftz.f32 %f1255, %f126, %f1235, %f1251; ld.global.v4.f32 {%f1256, %f1257, %f1258, %f1259}, [%rd7+48]; ld.global.v4.f32 {%f1264, %f1265, %f1266, %f1267}, [%rd7+112]; mul.ftz.f32 %f1272, %f116, %f1264; mul.ftz.f32 %f1273, %f116, %f1265; mul.ftz.f32 %f1274, %f116, %f1266; mul.ftz.f32 %f1275, %f116, %f1267; fma.rn.ftz.f32 %f1276, %f126, %f1256, %f1272; fma.rn.ftz.f32 %f1277, %f126, %f1257, %f1273; fma.rn.ftz.f32 %f1278, %f126, %f1258, %f1274; fma.rn.ftz.f32 %f1279, %f126, %f1259, %f1275; mul.ftz.f32 %f1280, %f3795, 0f3FB504F3; mul.ftz.f32 %f1281, %f3792, 0f3FB504F3; mul.ftz.f32 %f1282, %f3793, 0f3FB504F3; mul.ftz.f32 %f1283, %f3794, 0f3FB504F3; mul.ftz.f32 %f1284, %f1280, %f1281; mul.ftz.f32 %f1285, %f1280, %f1282; mul.ftz.f32 %f1286, %f1280, %f1283; mul.ftz.f32 %f1287, %f1281, %f1281; mul.ftz.f32 %f1288, %f1281, %f1282; mul.ftz.f32 %f1289, %f1281, %f1283; mul.ftz.f32 %f1290, %f1282, %f1282; mul.ftz.f32 %f1291, %f1282, %f1283; mul.ftz.f32 %f1292, %f1283, %f1283; sub.ftz.f32 %f1294, %f1143, %f1290; sub.ftz.f32 %f1295, %f1294, %f1292; sub.ftz.f32 %f1296, %f1288, %f1286; add.ftz.f32 %f1297, %f1285, %f1289; add.ftz.f32 %f1298, %f1286, %f1288; sub.ftz.f32 %f1299, %f1143, %f1287; sub.ftz.f32 %f1300, %f1299, %f1292; sub.ftz.f32 %f1301, %f1291, %f1284; sub.ftz.f32 %f1302, %f1289, %f1285; add.ftz.f32 %f1303, %f1284, %f1291; sub.ftz.f32 %f1304, %f1299, %f1290; mul.ftz.f32 %f1305, %f1296, %f1254; fma.rn.ftz.f32 %f1306, %f1295, %f1231, %f1305; fma.rn.ftz.f32 %f3796, %f1297, %f1277, %f1306; mul.ftz.f32 %f1307, %f1296, %f1255; fma.rn.ftz.f32 %f1308, %f1295, %f1252, %f1307; fma.rn.ftz.f32 %f3797, %f1297, %f1278, %f1308; mul.ftz.f32 %f1309, %f1296, %f1276; fma.rn.ftz.f32 %f1310, %f1295, %f1253, %f1309; fma.rn.ftz.f32 %f3798, %f1297, %f1279, %f1310; mul.ftz.f32 %f1311, %f1300, %f1254; fma.rn.ftz.f32 %f1312, %f1298, %f1231, %f1311; fma.rn.ftz.f32 %f3800, %f1301, %f1277, %f1312; mul.ftz.f32 %f1313, %f1300, %f1255; fma.rn.ftz.f32 %f1314, %f1298, %f1252, %f1313; fma.rn.ftz.f32 %f3801, %f1301, %f1278, %f1314; mul.ftz.f32 %f1315, %f1300, %f1276; fma.rn.ftz.f32 %f1316, %f1298, %f1253, %f1315; fma.rn.ftz.f32 %f3802, %f1301, %f1279, %f1316; mul.ftz.f32 %f1317, %f1303, %f1254; fma.rn.ftz.f32 %f1318, %f1302, %f1231, %f1317; fma.rn.ftz.f32 %f3804, %f1304, %f1277, %f1318; mul.ftz.f32 %f1319, %f1303, %f1255; fma.rn.ftz.f32 %f1320, %f1302, %f1252, %f1319; fma.rn.ftz.f32 %f3805, %f1304, %f1278, %f1320; mul.ftz.f32 %f1321, %f1303, %f1276; fma.rn.ftz.f32 %f1322, %f1302, %f1253, %f1321; fma.rn.ftz.f32 %f3806, %f1304, %f1279, %f1322; $L__BB203_60: mul.ftz.f32 %f1324, %f3778, %f3797; fma.rn.ftz.f32 %f1325, %f3777, %f3796, %f1324; fma.rn.ftz.f32 %f1326, %f3779, %f3798, %f1325; add.ftz.f32 %f3998, %f3799, %f1326; mul.ftz.f32 %f1327, %f3778, %f3801; fma.rn.ftz.f32 %f1328, %f3777, %f3800, %f1327; fma.rn.ftz.f32 %f1329, %f3779, %f3802, %f1328; add.ftz.f32 %f3997, %f3803, %f1329; mul.ftz.f32 %f1330, %f3778, %f3805; fma.rn.ftz.f32 %f1331, %f3777, %f3804, %f1330; fma.rn.ftz.f32 %f1332, %f3779, %f3806, %f1331; add.ftz.f32 %f3996, %f3807, %f1332; mul.ftz.f32 %f1333, %f3787, %f3797; fma.rn.ftz.f32 %f1334, %f3786, %f3796, %f1333; fma.rn.ftz.f32 %f1335, %f3788, %f3798, %f1334; mul.ftz.f32 %f1336, %f3787, %f3801; fma.rn.ftz.f32 %f1337, %f3786, %f3800, %f1336; fma.rn.ftz.f32 %f1338, %f3788, %f3802, %f1337; mul.ftz.f32 %f1339, %f3787, %f3805; fma.rn.ftz.f32 %f1340, %f3786, %f3804, %f1339; fma.rn.ftz.f32 %f1341, %f3788, %f3806, %f1340; mul.ftz.f32 %f1342, %f1338, %f1338; fma.rn.ftz.f32 %f1343, %f1335, %f1335, %f1342; fma.rn.ftz.f32 %f1344, %f1341, %f1341, %f1343; sqrt.approx.ftz.f32 %f1345, %f1344; div.approx.ftz.f32 %f3995, %f1335, %f1345; div.approx.ftz.f32 %f3994, %f1338, %f1345; div.approx.ftz.f32 %f3993, %f1341, %f1345; ld.const.v4.f32 {%f1346, %f1347, %f1348, %f1349}, [kernel_params+208]; mul.ftz.f32 %f1353, %f1347, %f1347; fma.rn.ftz.f32 %f1354, %f1346, %f1346, %f1353; fma.rn.ftz.f32 %f1355, %f1348, %f1348, %f1354; sqrt.approx.ftz.f32 %f1356, %f1355; add.ftz.f32 %f1357, %f1356, %f1356; mul.ftz.f32 %f3991, %f1357, 0f3F000000; mov.f32 %f3990, 0f00000000; ld.const.f32 %f3992, [kernel_params+272]; bra.uni $L__BB203_255; $L__BB203_76: div.approx.ftz.f32 %f1513, %f244, %f3821; fma.rn.ftz.f32 %f3820, %f1513, 0fBF490FDB, 0f3FC90FDB; $L__BB203_78: cos.approx.ftz.f32 %f1515, %f3820; sin.approx.ftz.f32 %f1516, %f3820; mul.ftz.f32 %f3825, %f3821, %f1516; mul.ftz.f32 %f3824, %f3821, %f1515; $L__BB203_79: ld.const.f32 %f3717, [kernel_params+240]; ld.const.f32 %f1517, [kernel_params+296]; mul.ftz.f32 %f1518, %f3824, %f1517; mul.ftz.f32 %f3829, %f3717, %f1518; ld.const.f32 %f1519, [kernel_params+252]; div.approx.ftz.f32 %f1520, %f1519, %f3819; mul.ftz.f32 %f1521, %f3817, %f1520; mul.ftz.f32 %f1522, %f3818, %f1520; mul.ftz.f32 %f1523, %f3819, %f1520; sub.ftz.f32 %f1524, %f1521, %f3829; mul.ftz.f32 %f3830, %f3717, %f3825; sub.ftz.f32 %f1525, %f1522, %f3830; mul.ftz.f32 %f1526, %f1525, %f1525; fma.rn.ftz.f32 %f1527, %f1524, %f1524, %f1526; fma.rn.ftz.f32 %f1528, %f1523, %f1523, %f1527; sqrt.approx.ftz.f32 %f1529, %f1528; div.approx.ftz.f32 %f3826, %f1524, %f1529; div.approx.ftz.f32 %f3827, %f1525, %f1529; div.approx.ftz.f32 %f3828, %f1523, %f1529; $L__BB203_80: ld.const.v4.f32 {%f3835, %f3836, %f3837, %f3838}, [kernel_params+96]; ld.const.v4.f32 {%f3839, %f3840, %f3841, %f3842}, [kernel_params+112]; ld.const.v4.f32 {%f3843, %f3844, %f3845, %f3846}, [kernel_params+128]; ld.const.u32 %r47, [kernel_params+260]; setp.eq.s32 %p54, %r47, 0; @%p54 bra $L__BB203_85; ld.const.u64 %rd72, [kernel_params+2208]; cvta.to.global.u64 %rd73, %rd72; add.s32 %r398, %r47, -1; cvt.rn.f32.s32 %f1542, %r398; mul.ftz.f32 %f1543, %f3776, %f1542; cvt.rzi.ftz.s32.f32 %r399, %f1543; add.s32 %r400, %r47, -2; min.s32 %r401, %r399, %r400; cvt.rn.f32.s32 %f1544, %r401; sub.ftz.f32 %f280, %f1543, %f1544; mul.wide.s32 %rd74, %r401, 64; add.s64 %rd8, %rd73, %rd74; ld.global.v4.f32 {%f1545, %f1546, %f1547, %f1548}, [%rd8]; ld.global.v4.f32 {%f1549, %f1550, %f1551, %f1552}, [%rd8+64]; mul.ftz.f32 %f1553, %f1546, %f1550; fma.rn.ftz.f32 %f1554, %f1545, %f1549, %f1553; mul.ftz.f32 %f1555, %f1548, %f1552; fma.rn.ftz.f32 %f1556, %f1547, %f1551, %f1555; add.ftz.f32 %f289, %f1554, %f1556; setp.gt.ftz.f32 %p55, %f289, 0f3F7FDF3B; mov.f32 %f1557, 0f3F800000; sub.ftz.f32 %f290, %f1557, %f280; @%p55 bra $L__BB203_83; bra.uni $L__BB203_82; $L__BB203_83: mul.ftz.f32 %f1611, %f1545, %f290; fma.rn.ftz.f32 %f1612, %f1549, %f280, %f1611; mul.ftz.f32 %f1613, %f1546, %f290; fma.rn.ftz.f32 %f1614, %f1550, %f280, %f1613; mul.ftz.f32 %f1615, %f1547, %f290; fma.rn.ftz.f32 %f1616, %f1551, %f280, %f1615; mul.ftz.f32 %f1617, %f1548, %f290; fma.rn.ftz.f32 %f1618, %f1552, %f280, %f1617; mul.ftz.f32 %f1619, %f1614, %f1614; fma.rn.ftz.f32 %f1620, %f1612, %f1612, %f1619; mul.ftz.f32 %f1621, %f1618, %f1618; fma.rn.ftz.f32 %f1622, %f1616, %f1616, %f1621; add.ftz.f32 %f1623, %f1620, %f1622; rsqrt.approx.ftz.f32 %f1624, %f1623; mul.ftz.f32 %f3834, %f1618, %f1624; mul.ftz.f32 %f3833, %f1616, %f1624; mul.ftz.f32 %f3832, %f1614, %f1624; mul.ftz.f32 %f3831, %f1612, %f1624; bra.uni $L__BB203_84; $L__BB203_82: mov.f32 %f1558, 0fBF800000; max.ftz.f32 %f1559, %f289, %f1558; min.ftz.f32 %f1561, %f1559, %f1557; abs.ftz.f32 %f1562, %f1561; sub.ftz.f32 %f1563, %f1557, %f1562; mul.ftz.f32 %f1564, %f1563, 0f3F000000; sqrt.approx.ftz.f32 %f1565, %f1564; setp.gt.ftz.f32 %p56, %f1562, 0f3F11EB85; selp.f32 %f1566, %f1565, %f1562, %p56; mul.ftz.f32 %f1567, %f1566, %f1566; mov.f32 %f1568, 0f3C94D2E9; mov.f32 %f1569, 0f3D53F941; fma.rn.ftz.f32 %f1570, %f1569, %f1567, %f1568; mov.f32 %f1571, 0f3D3F841F; fma.rn.ftz.f32 %f1572, %f1570, %f1567, %f1571; mov.f32 %f1573, 0f3D994929; fma.rn.ftz.f32 %f1574, %f1572, %f1567, %f1573; mov.f32 %f1575, 0f3E2AAB94; fma.rn.ftz.f32 %f1576, %f1574, %f1567, %f1575; mul.ftz.f32 %f1577, %f1567, %f1576; fma.rn.ftz.f32 %f1578, %f1577, %f1566, %f1566; add.ftz.f32 %f1579, %f1578, %f1578; mov.f32 %f1580, 0f3FC90FDB; sub.ftz.f32 %f1581, %f1580, %f1578; selp.f32 %f1582, %f1579, %f1581, %p56; setp.lt.ftz.f32 %p57, %f1561, 0f00000000; mov.f32 %f1583, 0f40490FDB; sub.ftz.f32 %f1584, %f1583, %f1582; selp.f32 %f1585, %f1584, %f1582, %p57; mul.ftz.f32 %f1586, %f1545, %f289; sub.ftz.f32 %f1587, %f1549, %f1586; mul.ftz.f32 %f1588, %f1546, %f289; sub.ftz.f32 %f1589, %f1550, %f1588; mul.ftz.f32 %f1590, %f1547, %f289; sub.ftz.f32 %f1591, %f1551, %f1590; mul.ftz.f32 %f1592, %f1548, %f289; sub.ftz.f32 %f1593, %f1552, %f1592; mul.ftz.f32 %f1594, %f1589, %f1589; fma.rn.ftz.f32 %f1595, %f1587, %f1587, %f1594; mul.ftz.f32 %f1596, %f1593, %f1593; fma.rn.ftz.f32 %f1597, %f1591, %f1591, %f1596; add.ftz.f32 %f1598, %f1595, %f1597; rsqrt.approx.ftz.f32 %f1599, %f1598; mul.ftz.f32 %f1600, %f1587, %f1599; mul.ftz.f32 %f1601, %f1589, %f1599; mul.ftz.f32 %f1602, %f1591, %f1599; mul.ftz.f32 %f1603, %f1593, %f1599; mul.ftz.f32 %f1604, %f280, %f1585; cos.approx.ftz.f32 %f1605, %f1604; sin.approx.ftz.f32 %f1606, %f1604; mul.ftz.f32 %f1607, %f1600, %f1606; mul.ftz.f32 %f1608, %f1601, %f1606; mul.ftz.f32 %f1609, %f1602, %f1606; mul.ftz.f32 %f1610, %f1603, %f1606; fma.rn.ftz.f32 %f3834, %f1548, %f1605, %f1610; fma.rn.ftz.f32 %f3833, %f1547, %f1605, %f1609; fma.rn.ftz.f32 %f3832, %f1546, %f1605, %f1608; fma.rn.ftz.f32 %f3831, %f1545, %f1605, %f1607; $L__BB203_84: ld.global.v4.f32 {%f1625, %f1626, %f1627, %f1628}, [%rd8+16]; ld.global.v4.f32 {%f1633, %f1634, %f1635, %f1636}, [%rd8+80]; mul.ftz.f32 %f1641, %f280, %f1633; mul.ftz.f32 %f1642, %f280, %f1634; mul.ftz.f32 %f1643, %f280, %f1635; mul.ftz.f32 %f1644, %f280, %f1636; fma.rn.ftz.f32 %f3838, %f290, %f1625, %f1641; fma.rn.ftz.f32 %f3842, %f290, %f1626, %f1642; fma.rn.ftz.f32 %f3846, %f290, %f1627, %f1643; fma.rn.ftz.f32 %f1645, %f290, %f1628, %f1644; ld.global.v4.f32 {%f1646, %f1647, %f1648, %f1649}, [%rd8+32]; ld.global.v4.f32 {%f1654, %f1655, %f1656, %f1657}, [%rd8+96]; mul.ftz.f32 %f1662, %f280, %f1654; mul.ftz.f32 %f1663, %f280, %f1655; mul.ftz.f32 %f1664, %f280, %f1656; mul.ftz.f32 %f1665, %f280, %f1657; fma.rn.ftz.f32 %f1666, %f290, %f1646, %f1662; fma.rn.ftz.f32 %f1667, %f290, %f1647, %f1663; fma.rn.ftz.f32 %f1668, %f290, %f1648, %f1664; fma.rn.ftz.f32 %f1669, %f290, %f1649, %f1665; ld.global.v4.f32 {%f1670, %f1671, %f1672, %f1673}, [%rd8+48]; ld.global.v4.f32 {%f1678, %f1679, %f1680, %f1681}, [%rd8+112]; mul.ftz.f32 %f1686, %f280, %f1678; mul.ftz.f32 %f1687, %f280, %f1679; mul.ftz.f32 %f1688, %f280, %f1680; mul.ftz.f32 %f1689, %f280, %f1681; fma.rn.ftz.f32 %f1690, %f290, %f1670, %f1686; fma.rn.ftz.f32 %f1691, %f290, %f1671, %f1687; fma.rn.ftz.f32 %f1692, %f290, %f1672, %f1688; fma.rn.ftz.f32 %f1693, %f290, %f1673, %f1689; mul.ftz.f32 %f1694, %f3834, 0f3FB504F3; mul.ftz.f32 %f1695, %f3831, 0f3FB504F3; mul.ftz.f32 %f1696, %f3832, 0f3FB504F3; mul.ftz.f32 %f1697, %f3833, 0f3FB504F3; mul.ftz.f32 %f1698, %f1694, %f1695; mul.ftz.f32 %f1699, %f1694, %f1696; mul.ftz.f32 %f1700, %f1694, %f1697; mul.ftz.f32 %f1701, %f1695, %f1695; mul.ftz.f32 %f1702, %f1695, %f1696; mul.ftz.f32 %f1703, %f1695, %f1697; mul.ftz.f32 %f1704, %f1696, %f1696; mul.ftz.f32 %f1705, %f1696, %f1697; mul.ftz.f32 %f1706, %f1697, %f1697; sub.ftz.f32 %f1708, %f1557, %f1704; sub.ftz.f32 %f1709, %f1708, %f1706; sub.ftz.f32 %f1710, %f1702, %f1700; add.ftz.f32 %f1711, %f1699, %f1703; add.ftz.f32 %f1712, %f1700, %f1702; sub.ftz.f32 %f1713, %f1557, %f1701; sub.ftz.f32 %f1714, %f1713, %f1706; sub.ftz.f32 %f1715, %f1705, %f1698; sub.ftz.f32 %f1716, %f1703, %f1699; add.ftz.f32 %f1717, %f1698, %f1705; sub.ftz.f32 %f1718, %f1713, %f1704; mul.ftz.f32 %f1719, %f1710, %f1668; fma.rn.ftz.f32 %f1720, %f1709, %f1645, %f1719; fma.rn.ftz.f32 %f3835, %f1711, %f1691, %f1720; mul.ftz.f32 %f1721, %f1710, %f1669; fma.rn.ftz.f32 %f1722, %f1709, %f1666, %f1721; fma.rn.ftz.f32 %f3836, %f1711, %f1692, %f1722; mul.ftz.f32 %f1723, %f1710, %f1690; fma.rn.ftz.f32 %f1724, %f1709, %f1667, %f1723; fma.rn.ftz.f32 %f3837, %f1711, %f1693, %f1724; mul.ftz.f32 %f1725, %f1714, %f1668; fma.rn.ftz.f32 %f1726, %f1712, %f1645, %f1725; fma.rn.ftz.f32 %f3839, %f1715, %f1691, %f1726; mul.ftz.f32 %f1727, %f1714, %f1669; fma.rn.ftz.f32 %f1728, %f1712, %f1666, %f1727; fma.rn.ftz.f32 %f3840, %f1715, %f1692, %f1728; mul.ftz.f32 %f1729, %f1714, %f1690; fma.rn.ftz.f32 %f1730, %f1712, %f1667, %f1729; fma.rn.ftz.f32 %f3841, %f1715, %f1693, %f1730; mul.ftz.f32 %f1731, %f1717, %f1668; fma.rn.ftz.f32 %f1732, %f1716, %f1645, %f1731; fma.rn.ftz.f32 %f3843, %f1718, %f1691, %f1732; mul.ftz.f32 %f1733, %f1717, %f1669; fma.rn.ftz.f32 %f1734, %f1716, %f1666, %f1733; fma.rn.ftz.f32 %f3844, %f1718, %f1692, %f1734; mul.ftz.f32 %f1735, %f1717, %f1690; fma.rn.ftz.f32 %f1736, %f1716, %f1667, %f1735; fma.rn.ftz.f32 %f3845, %f1718, %f1693, %f1736; $L__BB203_85: mul.ftz.f32 %f1737, %f3830, %f3836; fma.rn.ftz.f32 %f1738, %f3829, %f3835, %f1737; mul.ftz.f32 %f327, %f3837, 0f00000000; add.ftz.f32 %f1739, %f1738, %f327; add.ftz.f32 %f3876, %f3838, %f1739; mul.ftz.f32 %f1740, %f3830, %f3840; fma.rn.ftz.f32 %f1741, %f3829, %f3839, %f1740; mul.ftz.f32 %f329, %f3841, 0f00000000; add.ftz.f32 %f1742, %f1741, %f329; add.ftz.f32 %f3875, %f3842, %f1742; mul.ftz.f32 %f1743, %f3830, %f3844; fma.rn.ftz.f32 %f1744, %f3829, %f3843, %f1743; mul.ftz.f32 %f331, %f3845, 0f00000000; add.ftz.f32 %f1745, %f1744, %f331; add.ftz.f32 %f3874, %f3846, %f1745; mul.ftz.f32 %f1746, %f3827, %f3836; fma.rn.ftz.f32 %f1747, %f3826, %f3835, %f1746; fma.rn.ftz.f32 %f1748, %f3828, %f3837, %f1747; mul.ftz.f32 %f1749, %f3827, %f3840; fma.rn.ftz.f32 %f1750, %f3826, %f3839, %f1749; fma.rn.ftz.f32 %f1751, %f3828, %f3841, %f1750; mul.ftz.f32 %f1752, %f3827, %f3844; fma.rn.ftz.f32 %f1753, %f3826, %f3843, %f1752; fma.rn.ftz.f32 %f1754, %f3828, %f3845, %f1753; mul.ftz.f32 %f1755, %f1751, %f1751; fma.rn.ftz.f32 %f1756, %f1748, %f1748, %f1755; fma.rn.ftz.f32 %f1757, %f1754, %f1754, %f1756; sqrt.approx.ftz.f32 %f1758, %f1757; div.approx.ftz.f32 %f3995, %f1748, %f1758; div.approx.ftz.f32 %f3994, %f1751, %f1758; div.approx.ftz.f32 %f3993, %f1754, %f1758; ld.const.f32 %f3851, [kernel_params+80]; setp.eq.ftz.f32 %p58, %f3851, 0f00000000; @%p58 bra $L__BB203_115; bra.uni $L__BB203_86; $L__BB203_115: mul.ftz.f32 %f2071, %f3818, %f3836; fma.rn.ftz.f32 %f2072, %f3817, %f3835, %f2071; fma.rn.ftz.f32 %f2073, %f3819, %f3837, %f2072; mul.ftz.f32 %f2074, %f3818, %f3840; fma.rn.ftz.f32 %f2075, %f3817, %f3839, %f2074; fma.rn.ftz.f32 %f2076, %f3819, %f3841, %f2075; mul.ftz.f32 %f2077, %f3818, %f3844; fma.rn.ftz.f32 %f2078, %f3817, %f3843, %f2077; fma.rn.ftz.f32 %f2079, %f3819, %f3845, %f2078; mul.ftz.f32 %f2080, %f2076, %f2076; fma.rn.ftz.f32 %f2081, %f2073, %f2073, %f2080; fma.rn.ftz.f32 %f2082, %f2079, %f2079, %f2081; sqrt.approx.ftz.f32 %f2083, %f2082; div.approx.ftz.f32 %f2084, %f2073, %f2083; div.approx.ftz.f32 %f2085, %f2076, %f2083; div.approx.ftz.f32 %f2086, %f2079, %f2083; ld.const.v4.f32 {%f2087, %f2088, %f2089, %f2090}, [kernel_params+208]; add.ftz.f32 %f2094, %f2073, %f2087; add.ftz.f32 %f2095, %f2076, %f2088; add.ftz.f32 %f2096, %f2079, %f2089; mul.ftz.f32 %f2097, %f2095, %f2095; fma.rn.ftz.f32 %f2098, %f2094, %f2094, %f2097; fma.rn.ftz.f32 %f2099, %f2096, %f2096, %f2098; sqrt.approx.ftz.f32 %f2100, %f2099; div.approx.ftz.f32 %f2101, %f2094, %f2100; div.approx.ftz.f32 %f2102, %f2095, %f2100; div.approx.ftz.f32 %f2103, %f2096, %f2100; sub.ftz.f32 %f2104, %f2101, %f2084; sub.ftz.f32 %f2105, %f2102, %f2085; sub.ftz.f32 %f2106, %f2103, %f2086; ld.const.v4.f32 {%f2107, %f2108, %f2109, %f2110}, [kernel_params+224]; add.ftz.f32 %f2114, %f2073, %f2107; add.ftz.f32 %f2115, %f2076, %f2108; add.ftz.f32 %f2116, %f2079, %f2109; mul.ftz.f32 %f2117, %f2115, %f2115; fma.rn.ftz.f32 %f2118, %f2114, %f2114, %f2117; fma.rn.ftz.f32 %f2119, %f2116, %f2116, %f2118; sqrt.approx.ftz.f32 %f2120, %f2119; div.approx.ftz.f32 %f2121, %f2114, %f2120; div.approx.ftz.f32 %f2122, %f2115, %f2120; div.approx.ftz.f32 %f2123, %f2116, %f2120; sub.ftz.f32 %f2124, %f2121, %f2084; sub.ftz.f32 %f2125, %f2122, %f2085; sub.ftz.f32 %f2126, %f2123, %f2086; mul.ftz.f32 %f2127, %f2105, %f2105; fma.rn.ftz.f32 %f2128, %f2104, %f2104, %f2127; fma.rn.ftz.f32 %f2129, %f2106, %f2106, %f2128; sqrt.approx.ftz.f32 %f2130, %f2129; mul.ftz.f32 %f2131, %f2125, %f2125; fma.rn.ftz.f32 %f2132, %f2124, %f2124, %f2131; fma.rn.ftz.f32 %f2133, %f2126, %f2126, %f2132; sqrt.approx.ftz.f32 %f2134, %f2133; add.ftz.f32 %f3869, %f2130, %f2134; mov.f32 %f3870, 0f00000000; bra.uni $L__BB203_116; $L__BB203_86: ld.const.f32 %f337, [kernel_params+92]; setp.leu.ftz.f32 %p59, %f337, 0f00000000; mov.f32 %f3847, %f3851; @%p59 bra $L__BB203_90; mov.f32 %f1760, 0fBF800000; max.ftz.f32 %f1761, %f3993, %f1760; mov.f32 %f1762, 0f3F800000; min.ftz.f32 %f1763, %f1761, %f1762; abs.ftz.f32 %f1764, %f1763; sub.ftz.f32 %f1765, %f1762, %f1764; mul.ftz.f32 %f1766, %f1765, 0f3F000000; sqrt.approx.ftz.f32 %f1767, %f1766; setp.gt.ftz.f32 %p60, %f1764, 0f3F11EB85; selp.f32 %f1768, %f1767, %f1764, %p60; mul.ftz.f32 %f1769, %f1768, %f1768; mov.f32 %f1770, 0f3C94D2E9; mov.f32 %f1771, 0f3D53F941; fma.rn.ftz.f32 %f1772, %f1771, %f1769, %f1770; mov.f32 %f1773, 0f3D3F841F; fma.rn.ftz.f32 %f1774, %f1772, %f1769, %f1773; mov.f32 %f1775, 0f3D994929; fma.rn.ftz.f32 %f1776, %f1774, %f1769, %f1775; mov.f32 %f1777, 0f3E2AAB94; fma.rn.ftz.f32 %f1778, %f1776, %f1769, %f1777; mul.ftz.f32 %f1779, %f1769, %f1778; fma.rn.ftz.f32 %f1780, %f1779, %f1768, %f1768; mov.f32 %f1781, 0f3FC90FDB; mov.f32 %f1782, 0fC0000000; fma.rn.ftz.f32 %f1783, %f1782, %f1780, %f1781; selp.f32 %f1784, %f1783, %f1780, %p60; setp.le.ftz.f32 %p61, %f1784, 0f7F800000; mov.b32 %r402, %f1784; mov.b32 %r403, %f1763; and.b32 %r404, %r403, -2147483648; or.b32 %r405, %r404, %r402; mov.b32 %f1785, %r405; selp.f32 %f1786, %f1785, %f1784, %p61; abs.ftz.f32 %f339, %f1786; setp.gt.ftz.f32 %p62, %f339, %f337; mov.f32 %f3847, 0f00000000; @%p62 bra $L__BB203_90; ld.const.f32 %f3719, [kernel_params+88]; setp.leu.ftz.f32 %p63, %f339, %f3719; mov.f32 %f3847, %f3851; @%p63 bra $L__BB203_90; ld.const.f32 %f3721, [kernel_params+88]; ld.const.f32 %f3720, [kernel_params+92]; sub.ftz.f32 %f1787, %f339, %f3721; sub.ftz.f32 %f1788, %f3720, %f3721; div.approx.ftz.f32 %f1789, %f1787, %f1788; mul.ftz.f32 %f1790, %f1789, 0f3FC90FDB; cos.approx.ftz.f32 %f1791, %f1790; mul.ftz.f32 %f3847, %f3851, %f1791; $L__BB203_90: mul.ftz.f32 %f1792, %f3993, 0f00000000; sub.ftz.f32 %f1793, %f3994, %f1792; sub.ftz.f32 %f1794, %f1792, %f3995; mul.ftz.f32 %f1795, %f3994, 0f00000000; mul.ftz.f32 %f1796, %f3995, 0f00000000; sub.ftz.f32 %f1797, %f1796, %f1795; mul.ftz.f32 %f1798, %f1794, %f1794; fma.rn.ftz.f32 %f1799, %f1793, %f1793, %f1798; fma.rn.ftz.f32 %f1800, %f1797, %f1797, %f1799; sqrt.approx.ftz.f32 %f1801, %f1800; div.approx.ftz.f32 %f1802, %f1793, %f1801; div.approx.ftz.f32 %f1803, %f1794, %f1801; div.approx.ftz.f32 %f1804, %f1797, %f1801; mul.ftz.f32 %f342, %f3847, %f1802; mul.ftz.f32 %f343, %f3847, %f1803; mul.ftz.f32 %f344, %f3847, %f1804; add.ftz.f32 %f3876, %f3876, %f342; add.ftz.f32 %f3875, %f3875, %f343; add.ftz.f32 %f3874, %f3874, %f344; ld.const.f32 %f348, [kernel_params+84]; setp.eq.ftz.f32 %p64, %f348, 0f7F7FFFFF; @%p64 bra $L__BB203_92; mul.ftz.f32 %f1805, %f3995, %f348; sub.ftz.f32 %f1806, %f1805, %f342; mul.ftz.f32 %f1807, %f3994, %f348; sub.ftz.f32 %f1808, %f1807, %f343; mul.ftz.f32 %f1809, %f3993, %f348; sub.ftz.f32 %f1810, %f1809, %f344; mul.ftz.f32 %f1811, %f1808, %f1808; fma.rn.ftz.f32 %f1812, %f1806, %f1806, %f1811; fma.rn.ftz.f32 %f1813, %f1810, %f1810, %f1812; sqrt.approx.ftz.f32 %f1814, %f1813; div.approx.ftz.f32 %f3995, %f1806, %f1814; div.approx.ftz.f32 %f3994, %f1808, %f1814; div.approx.ftz.f32 %f3993, %f1810, %f1814; $L__BB203_92: mul.ftz.f32 %f3724, %f3845, 0f00000000; mul.ftz.f32 %f3723, %f3841, 0f00000000; mul.ftz.f32 %f3722, %f3837, 0f00000000; mul.ftz.f32 %f1815, %f3836, 0f00000000; fma.rn.ftz.f32 %f1816, %f3835, 0f00000000, %f1815; add.ftz.f32 %f1817, %f1816, %f3722; add.ftz.f32 %f355, %f3838, %f1817; mul.ftz.f32 %f1818, %f3840, 0f00000000; fma.rn.ftz.f32 %f1819, %f3839, 0f00000000, %f1818; add.ftz.f32 %f1820, %f1819, %f3723; add.ftz.f32 %f356, %f3842, %f1820; mul.ftz.f32 %f1821, %f3844, 0f00000000; fma.rn.ftz.f32 %f1822, %f3843, 0f00000000, %f1821; add.ftz.f32 %f1823, %f1822, %f3724; add.ftz.f32 %f357, %f3846, %f1823; mul.ftz.f32 %f1824, %f3818, %f3836; fma.rn.ftz.f32 %f1825, %f3817, %f3835, %f1824; fma.rn.ftz.f32 %f1826, %f3819, %f3837, %f1825; mul.ftz.f32 %f1827, %f3818, %f3840; fma.rn.ftz.f32 %f1828, %f3817, %f3839, %f1827; fma.rn.ftz.f32 %f1829, %f3819, %f3841, %f1828; mul.ftz.f32 %f1830, %f3818, %f3844; fma.rn.ftz.f32 %f1831, %f3817, %f3843, %f1830; fma.rn.ftz.f32 %f1832, %f3819, %f3845, %f1831; mul.ftz.f32 %f1833, %f1829, %f1829; fma.rn.ftz.f32 %f1834, %f1826, %f1826, %f1833; fma.rn.ftz.f32 %f1835, %f1832, %f1832, %f1834; sqrt.approx.ftz.f32 %f1836, %f1835; div.approx.ftz.f32 %f3854, %f1826, %f1836; div.approx.ftz.f32 %f3853, %f1829, %f1836; div.approx.ftz.f32 %f3852, %f1832, %f1836; @%p59 bra $L__BB203_96; ld.const.f32 %f3737, [kernel_params+92]; mov.f32 %f1838, 0fBF800000; max.ftz.f32 %f1839, %f3852, %f1838; mov.f32 %f1840, 0f3F800000; min.ftz.f32 %f1841, %f1839, %f1840; abs.ftz.f32 %f1842, %f1841; sub.ftz.f32 %f1843, %f1840, %f1842; mul.ftz.f32 %f1844, %f1843, 0f3F000000; sqrt.approx.ftz.f32 %f1845, %f1844; setp.gt.ftz.f32 %p66, %f1842, 0f3F11EB85; selp.f32 %f1846, %f1845, %f1842, %p66; mul.ftz.f32 %f1847, %f1846, %f1846; mov.f32 %f1848, 0f3C94D2E9; mov.f32 %f1849, 0f3D53F941; fma.rn.ftz.f32 %f1850, %f1849, %f1847, %f1848; mov.f32 %f1851, 0f3D3F841F; fma.rn.ftz.f32 %f1852, %f1850, %f1847, %f1851; mov.f32 %f1853, 0f3D994929; fma.rn.ftz.f32 %f1854, %f1852, %f1847, %f1853; mov.f32 %f1855, 0f3E2AAB94; fma.rn.ftz.f32 %f1856, %f1854, %f1847, %f1855; mul.ftz.f32 %f1857, %f1847, %f1856; fma.rn.ftz.f32 %f1858, %f1857, %f1846, %f1846; mov.f32 %f1859, 0f3FC90FDB; mov.f32 %f1860, 0fC0000000; fma.rn.ftz.f32 %f1861, %f1860, %f1858, %f1859; selp.f32 %f1862, %f1861, %f1858, %p66; setp.le.ftz.f32 %p67, %f1862, 0f7F800000; mov.b32 %r406, %f1862; mov.b32 %r407, %f1841; and.b32 %r408, %r407, -2147483648; or.b32 %r409, %r408, %r406; mov.b32 %f1863, %r409; selp.f32 %f1864, %f1863, %f1862, %p67; abs.ftz.f32 %f362, %f1864; setp.gt.ftz.f32 %p68, %f362, %f3737; mov.f32 %f3851, 0f00000000; @%p68 bra $L__BB203_96; ld.const.f32 %f3740, [kernel_params+88]; ld.const.f32 %f3851, [kernel_params+80]; setp.leu.ftz.f32 %p69, %f362, %f3740; @%p69 bra $L__BB203_96; ld.const.f32 %f3742, [kernel_params+88]; ld.const.f32 %f3741, [kernel_params+80]; ld.const.f32 %f3738, [kernel_params+92]; sub.ftz.f32 %f1865, %f362, %f3742; sub.ftz.f32 %f1866, %f3738, %f3742; div.approx.ftz.f32 %f1867, %f1865, %f1866; mul.ftz.f32 %f1868, %f1867, 0f3FC90FDB; cos.approx.ftz.f32 %f1869, %f1868; mul.ftz.f32 %f3851, %f3741, %f1869; $L__BB203_96: mul.ftz.f32 %f1870, %f3852, 0f00000000; sub.ftz.f32 %f1871, %f3853, %f1870; sub.ftz.f32 %f1872, %f1870, %f3854; mul.ftz.f32 %f1873, %f3853, 0f00000000; mul.ftz.f32 %f1874, %f3854, 0f00000000; sub.ftz.f32 %f1875, %f1874, %f1873; mul.ftz.f32 %f1876, %f1872, %f1872; fma.rn.ftz.f32 %f1877, %f1871, %f1871, %f1876; fma.rn.ftz.f32 %f1878, %f1875, %f1875, %f1877; sqrt.approx.ftz.f32 %f1879, %f1878; div.approx.ftz.f32 %f1880, %f1871, %f1879; div.approx.ftz.f32 %f1881, %f1872, %f1879; div.approx.ftz.f32 %f1882, %f1875, %f1879; mul.ftz.f32 %f365, %f3851, %f1880; mul.ftz.f32 %f366, %f3851, %f1881; mul.ftz.f32 %f367, %f3851, %f1882; add.ftz.f32 %f368, %f355, %f365; add.ftz.f32 %f369, %f356, %f366; add.ftz.f32 %f370, %f357, %f367; @%p64 bra $L__BB203_98; ld.const.f32 %f3753, [kernel_params+84]; mul.ftz.f32 %f1883, %f3854, %f3753; sub.ftz.f32 %f1884, %f1883, %f365; mul.ftz.f32 %f1885, %f3853, %f3753; sub.ftz.f32 %f1886, %f1885, %f366; mul.ftz.f32 %f1887, %f3852, %f3753; sub.ftz.f32 %f1888, %f1887, %f367; mul.ftz.f32 %f1889, %f1886, %f1886; fma.rn.ftz.f32 %f1890, %f1884, %f1884, %f1889; fma.rn.ftz.f32 %f1891, %f1888, %f1888, %f1890; sqrt.approx.ftz.f32 %f1892, %f1891; div.approx.ftz.f32 %f3854, %f1884, %f1892; div.approx.ftz.f32 %f3853, %f1886, %f1892; div.approx.ftz.f32 %f3852, %f1888, %f1892; $L__BB203_98: mul.ftz.f32 %f3728, %f50, %f1374; mul.ftz.f32 %f3727, %f50, %f1370; mul.ftz.f32 %f3726, %f50, %f1366; mul.ftz.f32 %f3725, %f50, %f1362; add.ftz.f32 %f1896, %f44, 0f3F800000; fma.rn.ftz.f32 %f377, %f1896, %f1361, %f3725; fma.rn.ftz.f32 %f378, %f1896, %f1365, %f3726; fma.rn.ftz.f32 %f379, %f1896, %f1369, %f3727; fma.rn.ftz.f32 %f1897, %f1896, %f1373, %f3728; add.ftz.f32 %f380, %f1897, %f201; setp.eq.ftz.f32 %p71, %f380, 0f00000000; mov.f32 %f3855, 0f00000000; mov.f32 %f3856, %f3855; mov.f32 %f3857, %f3855; @%p71 bra $L__BB203_100; add.ftz.f32 %f1898, %f377, %f190; add.ftz.f32 %f1899, %f378, %f194; add.ftz.f32 %f1900, %f379, %f198; div.approx.ftz.f32 %f3857, %f1900, %f380; div.approx.ftz.f32 %f3856, %f1899, %f380; div.approx.ftz.f32 %f3855, %f1898, %f380; $L__BB203_100: ld.const.f32 %f3744, [kernel_params+92]; setp.leu.ftz.f32 %p240, %f3744, 0f00000000; ld.const.f32 %f3858, [kernel_params+80]; mul.ftz.f32 %f1901, %f3836, %f3856; fma.rn.ftz.f32 %f1902, %f3835, %f3855, %f1901; fma.rn.ftz.f32 %f1903, %f3837, %f3857, %f1902; mul.ftz.f32 %f1904, %f3840, %f3856; fma.rn.ftz.f32 %f1905, %f3839, %f3855, %f1904; fma.rn.ftz.f32 %f1906, %f3841, %f3857, %f1905; mul.ftz.f32 %f1907, %f3844, %f3856; fma.rn.ftz.f32 %f1908, %f3843, %f3855, %f1907; fma.rn.ftz.f32 %f1909, %f3845, %f3857, %f1908; mul.ftz.f32 %f1910, %f1906, %f1906; fma.rn.ftz.f32 %f1911, %f1903, %f1903, %f1910; fma.rn.ftz.f32 %f1912, %f1909, %f1909, %f1911; sqrt.approx.ftz.f32 %f1913, %f1912; div.approx.ftz.f32 %f3861, %f1903, %f1913; div.approx.ftz.f32 %f3860, %f1906, %f1913; div.approx.ftz.f32 %f3859, %f1909, %f1913; @%p240 bra $L__BB203_104; ld.const.f32 %f3735, [kernel_params+92]; mov.f32 %f1915, 0fBF800000; max.ftz.f32 %f1916, %f3859, %f1915; mov.f32 %f1917, 0f3F800000; min.ftz.f32 %f1918, %f1916, %f1917; abs.ftz.f32 %f1919, %f1918; sub.ftz.f32 %f1920, %f1917, %f1919; mul.ftz.f32 %f1921, %f1920, 0f3F000000; sqrt.approx.ftz.f32 %f1922, %f1921; setp.gt.ftz.f32 %p73, %f1919, 0f3F11EB85; selp.f32 %f1923, %f1922, %f1919, %p73; mul.ftz.f32 %f1924, %f1923, %f1923; mov.f32 %f1925, 0f3C94D2E9; mov.f32 %f1926, 0f3D53F941; fma.rn.ftz.f32 %f1927, %f1926, %f1924, %f1925; mov.f32 %f1928, 0f3D3F841F; fma.rn.ftz.f32 %f1929, %f1927, %f1924, %f1928; mov.f32 %f1930, 0f3D994929; fma.rn.ftz.f32 %f1931, %f1929, %f1924, %f1930; mov.f32 %f1932, 0f3E2AAB94; fma.rn.ftz.f32 %f1933, %f1931, %f1924, %f1932; mul.ftz.f32 %f1934, %f1924, %f1933; fma.rn.ftz.f32 %f1935, %f1934, %f1923, %f1923; mov.f32 %f1936, 0f3FC90FDB; mov.f32 %f1937, 0fC0000000; fma.rn.ftz.f32 %f1938, %f1937, %f1935, %f1936; selp.f32 %f1939, %f1938, %f1935, %p73; setp.le.ftz.f32 %p74, %f1939, 0f7F800000; mov.b32 %r410, %f1939; mov.b32 %r411, %f1918; and.b32 %r412, %r411, -2147483648; or.b32 %r413, %r412, %r410; mov.b32 %f1940, %r413; selp.f32 %f1941, %f1940, %f1939, %p74; abs.ftz.f32 %f391, %f1941; setp.gt.ftz.f32 %p75, %f391, %f3735; mov.f32 %f3858, 0f00000000; @%p75 bra $L__BB203_104; ld.const.f32 %f3754, [kernel_params+88]; ld.const.f32 %f3858, [kernel_params+80]; setp.leu.ftz.f32 %p76, %f391, %f3754; @%p76 bra $L__BB203_104; ld.const.f32 %f3755, [kernel_params+88]; ld.const.f32 %f3752, [kernel_params+80]; ld.const.f32 %f3736, [kernel_params+92]; sub.ftz.f32 %f1942, %f391, %f3755; sub.ftz.f32 %f1943, %f3736, %f3755; div.approx.ftz.f32 %f1944, %f1942, %f1943; mul.ftz.f32 %f1945, %f1944, 0f3FC90FDB; cos.approx.ftz.f32 %f1946, %f1945; mul.ftz.f32 %f3858, %f3752, %f1946; $L__BB203_104: ld.const.f32 %f3756, [kernel_params+84]; setp.eq.ftz.f32 %p242, %f3756, 0f7F7FFFFF; mul.ftz.f32 %f1947, %f3859, 0f00000000; sub.ftz.f32 %f1948, %f3860, %f1947; sub.ftz.f32 %f1949, %f1947, %f3861; mul.ftz.f32 %f1950, %f3860, 0f00000000; mul.ftz.f32 %f1951, %f3861, 0f00000000; sub.ftz.f32 %f1952, %f1951, %f1950; mul.ftz.f32 %f1953, %f1949, %f1949; fma.rn.ftz.f32 %f1954, %f1948, %f1948, %f1953; fma.rn.ftz.f32 %f1955, %f1952, %f1952, %f1954; sqrt.approx.ftz.f32 %f1956, %f1955; div.approx.ftz.f32 %f1957, %f1948, %f1956; div.approx.ftz.f32 %f1958, %f1949, %f1956; div.approx.ftz.f32 %f1959, %f1952, %f1956; mul.ftz.f32 %f394, %f3858, %f1957; mul.ftz.f32 %f395, %f3858, %f1958; mul.ftz.f32 %f396, %f3858, %f1959; add.ftz.f32 %f397, %f355, %f394; add.ftz.f32 %f398, %f356, %f395; add.ftz.f32 %f399, %f357, %f396; @%p242 bra $L__BB203_106; ld.const.f32 %f3750, [kernel_params+84]; mul.ftz.f32 %f1960, %f3861, %f3750; sub.ftz.f32 %f1961, %f1960, %f394; mul.ftz.f32 %f1962, %f3860, %f3750; sub.ftz.f32 %f1963, %f1962, %f395; mul.ftz.f32 %f1964, %f3859, %f3750; sub.ftz.f32 %f1965, %f1964, %f396; mul.ftz.f32 %f1966, %f1963, %f1963; fma.rn.ftz.f32 %f1967, %f1961, %f1961, %f1966; fma.rn.ftz.f32 %f1968, %f1965, %f1965, %f1967; sqrt.approx.ftz.f32 %f1969, %f1968; div.approx.ftz.f32 %f3861, %f1961, %f1969; div.approx.ftz.f32 %f3860, %f1963, %f1969; div.approx.ftz.f32 %f3859, %f1965, %f1969; $L__BB203_106: mul.ftz.f32 %f3732, %f44, %f1373; mul.ftz.f32 %f3731, %f44, %f1369; mul.ftz.f32 %f3730, %f44, %f1365; mul.ftz.f32 %f3729, %f44, %f1361; sub.ftz.f32 %f406, %f397, %f368; sub.ftz.f32 %f407, %f398, %f369; sub.ftz.f32 %f408, %f399, %f370; sub.ftz.f32 %f409, %f3861, %f3854; sub.ftz.f32 %f410, %f3860, %f3853; sub.ftz.f32 %f411, %f3859, %f3852; add.ftz.f32 %f1973, %f50, 0f3F800000; fma.rn.ftz.f32 %f1974, %f1973, %f1362, %f3729; add.ftz.f32 %f412, %f1974, %f190; fma.rn.ftz.f32 %f1975, %f1973, %f1366, %f3730; add.ftz.f32 %f413, %f1975, %f194; fma.rn.ftz.f32 %f1976, %f1973, %f1370, %f3731; add.ftz.f32 %f414, %f1976, %f198; fma.rn.ftz.f32 %f1977, %f1973, %f1374, %f3732; add.ftz.f32 %f415, %f1977, %f201; setp.eq.ftz.f32 %p78, %f415, 0f00000000; mov.f32 %f3862, 0f00000000; mov.f32 %f3863, %f3862; mov.f32 %f3864, %f3862; @%p78 bra $L__BB203_108; div.approx.ftz.f32 %f3864, %f414, %f415; div.approx.ftz.f32 %f3863, %f413, %f415; div.approx.ftz.f32 %f3862, %f412, %f415; $L__BB203_108: ld.const.f32 %f3746, [kernel_params+92]; setp.leu.ftz.f32 %p241, %f3746, 0f00000000; ld.const.f32 %f3865, [kernel_params+80]; mul.ftz.f32 %f1978, %f3836, %f3863; fma.rn.ftz.f32 %f1979, %f3835, %f3862, %f1978; fma.rn.ftz.f32 %f1980, %f3837, %f3864, %f1979; mul.ftz.f32 %f1981, %f3840, %f3863; fma.rn.ftz.f32 %f1982, %f3839, %f3862, %f1981; fma.rn.ftz.f32 %f1983, %f3841, %f3864, %f1982; mul.ftz.f32 %f1984, %f3844, %f3863; fma.rn.ftz.f32 %f1985, %f3843, %f3862, %f1984; fma.rn.ftz.f32 %f1986, %f3845, %f3864, %f1985; mul.ftz.f32 %f1987, %f1983, %f1983; fma.rn.ftz.f32 %f1988, %f1980, %f1980, %f1987; fma.rn.ftz.f32 %f1989, %f1986, %f1986, %f1988; sqrt.approx.ftz.f32 %f1990, %f1989; div.approx.ftz.f32 %f3868, %f1980, %f1990; div.approx.ftz.f32 %f3867, %f1983, %f1990; div.approx.ftz.f32 %f3866, %f1986, %f1990; @%p241 bra $L__BB203_112; ld.const.f32 %f3733, [kernel_params+92]; ld.const.f32 %f425, [kernel_params+88]; mov.f32 %f1992, 0fBF800000; max.ftz.f32 %f1993, %f3866, %f1992; mov.f32 %f1994, 0f3F800000; min.ftz.f32 %f1995, %f1993, %f1994; abs.ftz.f32 %f1996, %f1995; sub.ftz.f32 %f1997, %f1994, %f1996; mul.ftz.f32 %f1998, %f1997, 0f3F000000; sqrt.approx.ftz.f32 %f1999, %f1998; setp.gt.ftz.f32 %p80, %f1996, 0f3F11EB85; selp.f32 %f2000, %f1999, %f1996, %p80; mul.ftz.f32 %f2001, %f2000, %f2000; mov.f32 %f2002, 0f3C94D2E9; mov.f32 %f2003, 0f3D53F941; fma.rn.ftz.f32 %f2004, %f2003, %f2001, %f2002; mov.f32 %f2005, 0f3D3F841F; fma.rn.ftz.f32 %f2006, %f2004, %f2001, %f2005; mov.f32 %f2007, 0f3D994929; fma.rn.ftz.f32 %f2008, %f2006, %f2001, %f2007; mov.f32 %f2009, 0f3E2AAB94; fma.rn.ftz.f32 %f2010, %f2008, %f2001, %f2009; mul.ftz.f32 %f2011, %f2001, %f2010; fma.rn.ftz.f32 %f2012, %f2011, %f2000, %f2000; mov.f32 %f2013, 0f3FC90FDB; mov.f32 %f2014, 0fC0000000; fma.rn.ftz.f32 %f2015, %f2014, %f2012, %f2013; selp.f32 %f2016, %f2015, %f2012, %p80; setp.le.ftz.f32 %p81, %f2016, 0f7F800000; mov.b32 %r414, %f2016; mov.b32 %r415, %f1995; and.b32 %r416, %r415, -2147483648; or.b32 %r417, %r416, %r414; mov.b32 %f2017, %r417; selp.f32 %f2018, %f2017, %f2016, %p81; abs.ftz.f32 %f426, %f2018; setp.gt.ftz.f32 %p82, %f426, %f3733; mov.f32 %f3865, 0f00000000; @%p82 bra $L__BB203_112; ld.const.f32 %f3865, [kernel_params+80]; setp.leu.ftz.f32 %p83, %f426, %f425; @%p83 bra $L__BB203_112; ld.const.f32 %f3749, [kernel_params+80]; ld.const.f32 %f3734, [kernel_params+92]; sub.ftz.f32 %f2019, %f426, %f425; sub.ftz.f32 %f2020, %f3734, %f425; div.approx.ftz.f32 %f2021, %f2019, %f2020; mul.ftz.f32 %f2022, %f2021, 0f3FC90FDB; cos.approx.ftz.f32 %f2023, %f2022; mul.ftz.f32 %f3865, %f3749, %f2023; $L__BB203_112: ld.const.f32 %f3757, [kernel_params+84]; setp.eq.ftz.f32 %p243, %f3757, 0f7F7FFFFF; mul.ftz.f32 %f2024, %f3866, 0f00000000; sub.ftz.f32 %f2025, %f3867, %f2024; sub.ftz.f32 %f2026, %f2024, %f3868; mul.ftz.f32 %f2027, %f3867, 0f00000000; mul.ftz.f32 %f2028, %f3868, 0f00000000; sub.ftz.f32 %f2029, %f2028, %f2027; mul.ftz.f32 %f2030, %f2026, %f2026; fma.rn.ftz.f32 %f2031, %f2025, %f2025, %f2030; fma.rn.ftz.f32 %f2032, %f2029, %f2029, %f2031; sqrt.approx.ftz.f32 %f2033, %f2032; div.approx.ftz.f32 %f2034, %f2025, %f2033; div.approx.ftz.f32 %f2035, %f2026, %f2033; div.approx.ftz.f32 %f2036, %f2029, %f2033; mul.ftz.f32 %f429, %f3865, %f2034; mul.ftz.f32 %f430, %f3865, %f2035; mul.ftz.f32 %f431, %f3865, %f2036; add.ftz.f32 %f432, %f355, %f429; add.ftz.f32 %f433, %f356, %f430; add.ftz.f32 %f434, %f357, %f431; @%p243 bra $L__BB203_114; ld.const.f32 %f3747, [kernel_params+84]; mul.ftz.f32 %f2037, %f3868, %f3747; sub.ftz.f32 %f2038, %f2037, %f429; mul.ftz.f32 %f2039, %f3867, %f3747; sub.ftz.f32 %f2040, %f2039, %f430; mul.ftz.f32 %f2041, %f3866, %f3747; sub.ftz.f32 %f2042, %f2041, %f431; mul.ftz.f32 %f2043, %f2040, %f2040; fma.rn.ftz.f32 %f2044, %f2038, %f2038, %f2043; fma.rn.ftz.f32 %f2045, %f2042, %f2042, %f2044; sqrt.approx.ftz.f32 %f2046, %f2045; div.approx.ftz.f32 %f3868, %f2038, %f2046; div.approx.ftz.f32 %f3867, %f2040, %f2046; div.approx.ftz.f32 %f3866, %f2042, %f2046; $L__BB203_114: mul.ftz.f32 %f2047, %f409, %f409; fma.rn.ftz.f32 %f2048, %f410, %f410, %f2047; fma.rn.ftz.f32 %f2049, %f411, %f411, %f2048; sqrt.approx.ftz.f32 %f2050, %f2049; sub.ftz.f32 %f2051, %f3868, %f3854; mul.ftz.f32 %f2052, %f2051, %f2051; sub.ftz.f32 %f2053, %f3867, %f3853; fma.rn.ftz.f32 %f2054, %f2053, %f2053, %f2052; sub.ftz.f32 %f2055, %f3866, %f3852; fma.rn.ftz.f32 %f2056, %f2055, %f2055, %f2054; sqrt.approx.ftz.f32 %f2057, %f2056; add.ftz.f32 %f3869, %f2050, %f2057; mul.ftz.f32 %f2058, %f407, %f407; fma.rn.ftz.f32 %f2059, %f406, %f406, %f2058; fma.rn.ftz.f32 %f2060, %f408, %f408, %f2059; sqrt.approx.ftz.f32 %f2061, %f2060; sub.ftz.f32 %f2062, %f432, %f368; sub.ftz.f32 %f2063, %f433, %f369; mul.ftz.f32 %f2064, %f2063, %f2063; fma.rn.ftz.f32 %f2065, %f2062, %f2062, %f2064; sub.ftz.f32 %f2066, %f434, %f370; fma.rn.ftz.f32 %f2067, %f2066, %f2066, %f2065; sqrt.approx.ftz.f32 %f2068, %f2067; add.ftz.f32 %f2069, %f2061, %f2068; mul.ftz.f32 %f3870, %f2069, 0f3F000000; $L__BB203_116: mul.ftz.f32 %f2135, %f3818, %f3818; fma.rn.ftz.f32 %f2136, %f3817, %f3817, %f2135; fma.rn.ftz.f32 %f2137, %f3819, %f3819, %f2136; sqrt.approx.ftz.f32 %f2138, %f2137; div.approx.ftz.f32 %f2139, %f3819, %f2138; rcp.approx.ftz.f32 %f2140, %f2139; ld.const.f32 %f2141, [kernel_params+268]; mul.ftz.f32 %f2142, %f2141, %f2140; fma.rn.ftz.f32 %f3998, %f3995, %f2142, %f3876; fma.rn.ftz.f32 %f3997, %f3994, %f2142, %f3875; fma.rn.ftz.f32 %f3996, %f3993, %f2142, %f3874; mul.ftz.f32 %f3990, %f3869, 0f3F000000; fma.rn.ftz.f32 %f3991, %f3990, %f2142, %f3870; ld.const.f32 %f2143, [kernel_params+272]; mul.ftz.f32 %f3992, %f2140, %f2143; bra.uni $L__BB203_255; $L__BB203_147: div.approx.ftz.f32 %f2414, %f553, %f3893; fma.rn.ftz.f32 %f3892, %f2414, 0fBF490FDB, 0f3FC90FDB; $L__BB203_149: cos.approx.ftz.f32 %f2416, %f3892; sin.approx.ftz.f32 %f2417, %f3892; mul.ftz.f32 %f3897, %f3893, %f2417; mul.ftz.f32 %f3896, %f3893, %f2416; $L__BB203_150: ld.const.f32 %f3695, [kernel_params+240]; ld.const.f32 %f2418, [kernel_params+296]; mul.ftz.f32 %f2419, %f3896, %f2418; mul.ftz.f32 %f2420, %f3695, %f2419; mul.ftz.f32 %f2421, %f3890, %f3890; fma.rn.ftz.f32 %f2422, %f3889, %f3889, %f2421; fma.rn.ftz.f32 %f2423, %f3891, %f3891, %f2422; sqrt.approx.ftz.f32 %f2424, %f2423; div.approx.ftz.f32 %f2425, %f3889, %f2424; div.approx.ftz.f32 %f2426, %f3890, %f2424; div.approx.ftz.f32 %f2427, %f3891, %f2424; ld.const.f32 %f2428, [kernel_params+252]; mul.ftz.f32 %f2429, %f2425, %f2428; mul.ftz.f32 %f2430, %f2426, %f2428; mul.ftz.f32 %f2431, %f2427, %f2428; mul.ftz.f32 %f2432, %f2425, %f2425; mul.ftz.f32 %f2433, %f2425, %f2426; mul.ftz.f32 %f2434, %f2425, %f2427; mov.f32 %f2435, 0f3F800000; sub.ftz.f32 %f2436, %f2435, %f2432; mov.f32 %f2437, 0f00000000; sub.ftz.f32 %f2438, %f2437, %f2433; sub.ftz.f32 %f2439, %f2437, %f2434; mul.ftz.f32 %f2440, %f2438, %f2438; fma.rn.ftz.f32 %f2441, %f2436, %f2436, %f2440; fma.rn.ftz.f32 %f2442, %f2439, %f2439, %f2441; sqrt.approx.ftz.f32 %f2443, %f2442; div.approx.ftz.f32 %f2444, %f2436, %f2443; div.approx.ftz.f32 %f2445, %f2438, %f2443; div.approx.ftz.f32 %f2446, %f2439, %f2443; mul.ftz.f32 %f2447, %f2426, %f2446; mul.ftz.f32 %f2448, %f2427, %f2445; sub.ftz.f32 %f2449, %f2447, %f2448; mul.ftz.f32 %f2450, %f2427, %f2444; mul.ftz.f32 %f2451, %f2425, %f2446; sub.ftz.f32 %f2452, %f2450, %f2451; mul.ftz.f32 %f2453, %f2425, %f2445; mul.ftz.f32 %f2454, %f2426, %f2444; sub.ftz.f32 %f2455, %f2453, %f2454; mul.ftz.f32 %f2456, %f2452, %f2452; fma.rn.ftz.f32 %f2457, %f2449, %f2449, %f2456; fma.rn.ftz.f32 %f2458, %f2455, %f2455, %f2457; sqrt.approx.ftz.f32 %f2459, %f2458; div.approx.ftz.f32 %f2460, %f2449, %f2459; div.approx.ftz.f32 %f2461, %f2452, %f2459; div.approx.ftz.f32 %f2462, %f2455, %f2459; mul.ftz.f32 %f2463, %f3695, %f3897; mul.ftz.f32 %f2464, %f2463, %f2460; mul.ftz.f32 %f2465, %f2463, %f2461; mul.ftz.f32 %f2466, %f2463, %f2462; fma.rn.ftz.f32 %f3924, %f2420, %f2444, %f2464; fma.rn.ftz.f32 %f3925, %f2420, %f2445, %f2465; fma.rn.ftz.f32 %f3926, %f2420, %f2446, %f2466; sub.ftz.f32 %f2467, %f2429, %f3924; sub.ftz.f32 %f2468, %f2430, %f3925; sub.ftz.f32 %f2469, %f2431, %f3926; mul.ftz.f32 %f2470, %f2468, %f2468; fma.rn.ftz.f32 %f2471, %f2467, %f2467, %f2470; fma.rn.ftz.f32 %f2472, %f2469, %f2469, %f2471; sqrt.approx.ftz.f32 %f2473, %f2472; div.approx.ftz.f32 %f3889, %f2467, %f2473; div.approx.ftz.f32 %f3890, %f2468, %f2473; div.approx.ftz.f32 %f3891, %f2469, %f2473; $L__BB203_151: ld.const.v4.f32 {%f3908, %f3909, %f3910, %f3911}, [kernel_params+96]; ld.const.v4.f32 {%f3912, %f3913, %f3914, %f3915}, [kernel_params+112]; ld.const.v4.f32 {%f3916, %f3917, %f3918, %f3919}, [kernel_params+128]; ld.const.u32 %r49, [kernel_params+260]; setp.eq.s32 %p121, %r49, 0; @%p121 bra $L__BB203_156; ld.const.u64 %rd198, [kernel_params+2208]; add.s32 %r422, %r49, -1; cvt.rn.f32.s32 %f2486, %r422; mul.ftz.f32 %f2487, %f3776, %f2486; cvt.rzi.ftz.s32.f32 %r423, %f2487; add.s32 %r424, %r49, -2; min.s32 %r425, %r423, %r424; cvt.rn.f32.s32 %f2488, %r425; sub.ftz.f32 %f591, %f2487, %f2488; cvta.to.global.u64 %rd75, %rd198; mul.wide.s32 %rd76, %r425, 64; add.s64 %rd10, %rd75, %rd76; ld.global.v4.f32 {%f2489, %f2490, %f2491, %f2492}, [%rd10]; ld.global.v4.f32 {%f2493, %f2494, %f2495, %f2496}, [%rd10+64]; mul.ftz.f32 %f2497, %f2490, %f2494; fma.rn.ftz.f32 %f2498, %f2489, %f2493, %f2497; mul.ftz.f32 %f2499, %f2492, %f2496; fma.rn.ftz.f32 %f2500, %f2491, %f2495, %f2499; add.ftz.f32 %f600, %f2498, %f2500; setp.gt.ftz.f32 %p122, %f600, 0f3F7FDF3B; @%p122 bra $L__BB203_154; bra.uni $L__BB203_153; $L__BB203_154: mov.f32 %f3700, 0f3F800000; sub.ftz.f32 %f3699, %f3700, %f591; mul.ftz.f32 %f2555, %f2489, %f3699; fma.rn.ftz.f32 %f2556, %f2493, %f591, %f2555; mul.ftz.f32 %f2557, %f2490, %f3699; fma.rn.ftz.f32 %f2558, %f2494, %f591, %f2557; mul.ftz.f32 %f2559, %f2491, %f3699; fma.rn.ftz.f32 %f2560, %f2495, %f591, %f2559; mul.ftz.f32 %f2561, %f2492, %f3699; fma.rn.ftz.f32 %f2562, %f2496, %f591, %f2561; mul.ftz.f32 %f2563, %f2558, %f2558; fma.rn.ftz.f32 %f2564, %f2556, %f2556, %f2563; mul.ftz.f32 %f2565, %f2562, %f2562; fma.rn.ftz.f32 %f2566, %f2560, %f2560, %f2565; add.ftz.f32 %f2567, %f2564, %f2566; rsqrt.approx.ftz.f32 %f2568, %f2567; mul.ftz.f32 %f3907, %f2562, %f2568; mul.ftz.f32 %f3906, %f2560, %f2568; mul.ftz.f32 %f3905, %f2558, %f2568; mul.ftz.f32 %f3904, %f2556, %f2568; bra.uni $L__BB203_155; $L__BB203_153: mov.f32 %f3696, 0f3F800000; mov.f32 %f2502, 0fBF800000; max.ftz.f32 %f2503, %f600, %f2502; min.ftz.f32 %f2505, %f2503, %f3696; abs.ftz.f32 %f2506, %f2505; sub.ftz.f32 %f2507, %f3696, %f2506; mul.ftz.f32 %f2508, %f2507, 0f3F000000; sqrt.approx.ftz.f32 %f2509, %f2508; setp.gt.ftz.f32 %p123, %f2506, 0f3F11EB85; selp.f32 %f2510, %f2509, %f2506, %p123; mul.ftz.f32 %f2511, %f2510, %f2510; mov.f32 %f2512, 0f3C94D2E9; mov.f32 %f2513, 0f3D53F941; fma.rn.ftz.f32 %f2514, %f2513, %f2511, %f2512; mov.f32 %f2515, 0f3D3F841F; fma.rn.ftz.f32 %f2516, %f2514, %f2511, %f2515; mov.f32 %f2517, 0f3D994929; fma.rn.ftz.f32 %f2518, %f2516, %f2511, %f2517; mov.f32 %f2519, 0f3E2AAB94; fma.rn.ftz.f32 %f2520, %f2518, %f2511, %f2519; mul.ftz.f32 %f2521, %f2511, %f2520; fma.rn.ftz.f32 %f2522, %f2521, %f2510, %f2510; add.ftz.f32 %f2523, %f2522, %f2522; mov.f32 %f2524, 0f3FC90FDB; sub.ftz.f32 %f2525, %f2524, %f2522; selp.f32 %f2526, %f2523, %f2525, %p123; setp.lt.ftz.f32 %p124, %f2505, 0f00000000; mov.f32 %f2527, 0f40490FDB; sub.ftz.f32 %f2528, %f2527, %f2526; selp.f32 %f2529, %f2528, %f2526, %p124; mul.ftz.f32 %f2530, %f2489, %f600; sub.ftz.f32 %f2531, %f2493, %f2530; mul.ftz.f32 %f2532, %f2490, %f600; sub.ftz.f32 %f2533, %f2494, %f2532; mul.ftz.f32 %f2534, %f2491, %f600; sub.ftz.f32 %f2535, %f2495, %f2534; mul.ftz.f32 %f2536, %f2492, %f600; sub.ftz.f32 %f2537, %f2496, %f2536; mul.ftz.f32 %f2538, %f2533, %f2533; fma.rn.ftz.f32 %f2539, %f2531, %f2531, %f2538; mul.ftz.f32 %f2540, %f2537, %f2537; fma.rn.ftz.f32 %f2541, %f2535, %f2535, %f2540; add.ftz.f32 %f2542, %f2539, %f2541; rsqrt.approx.ftz.f32 %f2543, %f2542; mul.ftz.f32 %f2544, %f2531, %f2543; mul.ftz.f32 %f2545, %f2533, %f2543; mul.ftz.f32 %f2546, %f2535, %f2543; mul.ftz.f32 %f2547, %f2537, %f2543; mul.ftz.f32 %f2548, %f591, %f2529; cos.approx.ftz.f32 %f2549, %f2548; sin.approx.ftz.f32 %f2550, %f2548; mul.ftz.f32 %f2551, %f2544, %f2550; mul.ftz.f32 %f2552, %f2545, %f2550; mul.ftz.f32 %f2553, %f2546, %f2550; mul.ftz.f32 %f2554, %f2547, %f2550; fma.rn.ftz.f32 %f3907, %f2492, %f2549, %f2554; fma.rn.ftz.f32 %f3906, %f2491, %f2549, %f2553; fma.rn.ftz.f32 %f3905, %f2490, %f2549, %f2552; fma.rn.ftz.f32 %f3904, %f2489, %f2549, %f2551; $L__BB203_155: mov.f32 %f3698, 0f3F800000; sub.ftz.f32 %f3697, %f3698, %f591; ld.global.v4.f32 {%f2569, %f2570, %f2571, %f2572}, [%rd10+16]; ld.global.v4.f32 {%f2577, %f2578, %f2579, %f2580}, [%rd10+80]; mul.ftz.f32 %f2585, %f591, %f2577; mul.ftz.f32 %f2586, %f591, %f2578; mul.ftz.f32 %f2587, %f591, %f2579; mul.ftz.f32 %f2588, %f591, %f2580; fma.rn.ftz.f32 %f3911, %f3697, %f2569, %f2585; fma.rn.ftz.f32 %f3915, %f3697, %f2570, %f2586; fma.rn.ftz.f32 %f3919, %f3697, %f2571, %f2587; fma.rn.ftz.f32 %f2589, %f3697, %f2572, %f2588; ld.global.v4.f32 {%f2590, %f2591, %f2592, %f2593}, [%rd10+32]; ld.global.v4.f32 {%f2598, %f2599, %f2600, %f2601}, [%rd10+96]; mul.ftz.f32 %f2606, %f591, %f2598; mul.ftz.f32 %f2607, %f591, %f2599; mul.ftz.f32 %f2608, %f591, %f2600; mul.ftz.f32 %f2609, %f591, %f2601; fma.rn.ftz.f32 %f2610, %f3697, %f2590, %f2606; fma.rn.ftz.f32 %f2611, %f3697, %f2591, %f2607; fma.rn.ftz.f32 %f2612, %f3697, %f2592, %f2608; fma.rn.ftz.f32 %f2613, %f3697, %f2593, %f2609; ld.global.v4.f32 {%f2614, %f2615, %f2616, %f2617}, [%rd10+48]; ld.global.v4.f32 {%f2622, %f2623, %f2624, %f2625}, [%rd10+112]; mul.ftz.f32 %f2630, %f591, %f2622; mul.ftz.f32 %f2631, %f591, %f2623; mul.ftz.f32 %f2632, %f591, %f2624; mul.ftz.f32 %f2633, %f591, %f2625; fma.rn.ftz.f32 %f2634, %f3697, %f2614, %f2630; fma.rn.ftz.f32 %f2635, %f3697, %f2615, %f2631; fma.rn.ftz.f32 %f2636, %f3697, %f2616, %f2632; fma.rn.ftz.f32 %f2637, %f3697, %f2617, %f2633; mul.ftz.f32 %f2638, %f3907, 0f3FB504F3; mul.ftz.f32 %f2639, %f3904, 0f3FB504F3; mul.ftz.f32 %f2640, %f3905, 0f3FB504F3; mul.ftz.f32 %f2641, %f3906, 0f3FB504F3; mul.ftz.f32 %f2642, %f2638, %f2639; mul.ftz.f32 %f2643, %f2638, %f2640; mul.ftz.f32 %f2644, %f2638, %f2641; mul.ftz.f32 %f2645, %f2639, %f2639; mul.ftz.f32 %f2646, %f2639, %f2640; mul.ftz.f32 %f2647, %f2639, %f2641; mul.ftz.f32 %f2648, %f2640, %f2640; mul.ftz.f32 %f2649, %f2640, %f2641; mul.ftz.f32 %f2650, %f2641, %f2641; sub.ftz.f32 %f2652, %f3698, %f2648; sub.ftz.f32 %f2653, %f2652, %f2650; sub.ftz.f32 %f2654, %f2646, %f2644; add.ftz.f32 %f2655, %f2643, %f2647; add.ftz.f32 %f2656, %f2644, %f2646; sub.ftz.f32 %f2657, %f3698, %f2645; sub.ftz.f32 %f2658, %f2657, %f2650; sub.ftz.f32 %f2659, %f2649, %f2642; sub.ftz.f32 %f2660, %f2647, %f2643; add.ftz.f32 %f2661, %f2642, %f2649; sub.ftz.f32 %f2662, %f2657, %f2648; mul.ftz.f32 %f2663, %f2654, %f2612; fma.rn.ftz.f32 %f2664, %f2653, %f2589, %f2663; fma.rn.ftz.f32 %f3908, %f2655, %f2635, %f2664; mul.ftz.f32 %f2665, %f2654, %f2613; fma.rn.ftz.f32 %f2666, %f2653, %f2610, %f2665; fma.rn.ftz.f32 %f3909, %f2655, %f2636, %f2666; mul.ftz.f32 %f2667, %f2654, %f2634; fma.rn.ftz.f32 %f2668, %f2653, %f2611, %f2667; fma.rn.ftz.f32 %f3910, %f2655, %f2637, %f2668; mul.ftz.f32 %f2669, %f2658, %f2612; fma.rn.ftz.f32 %f2670, %f2656, %f2589, %f2669; fma.rn.ftz.f32 %f3912, %f2659, %f2635, %f2670; mul.ftz.f32 %f2671, %f2658, %f2613; fma.rn.ftz.f32 %f2672, %f2656, %f2610, %f2671; fma.rn.ftz.f32 %f3913, %f2659, %f2636, %f2672; mul.ftz.f32 %f2673, %f2658, %f2634; fma.rn.ftz.f32 %f2674, %f2656, %f2611, %f2673; fma.rn.ftz.f32 %f3914, %f2659, %f2637, %f2674; mul.ftz.f32 %f2675, %f2661, %f2612; fma.rn.ftz.f32 %f2676, %f2660, %f2589, %f2675; fma.rn.ftz.f32 %f3916, %f2662, %f2635, %f2676; mul.ftz.f32 %f2677, %f2661, %f2613; fma.rn.ftz.f32 %f2678, %f2660, %f2610, %f2677; fma.rn.ftz.f32 %f3917, %f2662, %f2636, %f2678; mul.ftz.f32 %f2679, %f2661, %f2634; fma.rn.ftz.f32 %f2680, %f2660, %f2611, %f2679; fma.rn.ftz.f32 %f3918, %f2662, %f2637, %f2680; $L__BB203_156: ld.const.f32 %f3920, [kernel_params+80]; setp.eq.ftz.f32 %p125, %f3920, 0f00000000; @%p125 bra $L__BB203_163; ld.const.f32 %f639, [kernel_params+92]; setp.leu.ftz.f32 %p126, %f639, 0f00000000; @%p126 bra $L__BB203_161; mov.f32 %f2682, 0fBF800000; max.ftz.f32 %f2683, %f3891, %f2682; mov.f32 %f2684, 0f3F800000; min.ftz.f32 %f2685, %f2683, %f2684; abs.ftz.f32 %f2686, %f2685; sub.ftz.f32 %f2687, %f2684, %f2686; mul.ftz.f32 %f2688, %f2687, 0f3F000000; sqrt.approx.ftz.f32 %f2689, %f2688; setp.gt.ftz.f32 %p127, %f2686, 0f3F11EB85; selp.f32 %f2690, %f2689, %f2686, %p127; mul.ftz.f32 %f2691, %f2690, %f2690; mov.f32 %f2692, 0f3C94D2E9; mov.f32 %f2693, 0f3D53F941; fma.rn.ftz.f32 %f2694, %f2693, %f2691, %f2692; mov.f32 %f2695, 0f3D3F841F; fma.rn.ftz.f32 %f2696, %f2694, %f2691, %f2695; mov.f32 %f2697, 0f3D994929; fma.rn.ftz.f32 %f2698, %f2696, %f2691, %f2697; mov.f32 %f2699, 0f3E2AAB94; fma.rn.ftz.f32 %f2700, %f2698, %f2691, %f2699; mul.ftz.f32 %f2701, %f2691, %f2700; fma.rn.ftz.f32 %f2702, %f2701, %f2690, %f2690; mov.f32 %f2703, 0f3FC90FDB; mov.f32 %f2704, 0fC0000000; fma.rn.ftz.f32 %f2705, %f2704, %f2702, %f2703; selp.f32 %f2706, %f2705, %f2702, %p127; setp.le.ftz.f32 %p128, %f2706, 0f7F800000; mov.b32 %r426, %f2706; mov.b32 %r427, %f2685; and.b32 %r428, %r427, -2147483648; or.b32 %r429, %r428, %r426; mov.b32 %f2707, %r429; selp.f32 %f2708, %f2707, %f2706, %p128; abs.ftz.f32 %f641, %f2708; setp.gt.ftz.f32 %p129, %f641, %f639; mov.f32 %f3920, 0f00000000; @%p129 bra $L__BB203_161; ld.const.f32 %f3920, [kernel_params+80]; ld.const.f32 %f3668, [kernel_params+88]; setp.leu.ftz.f32 %p130, %f641, %f3668; @%p130 bra $L__BB203_161; ld.const.f32 %f3759, [kernel_params+80]; ld.const.f32 %f3670, [kernel_params+88]; ld.const.f32 %f3669, [kernel_params+92]; sub.ftz.f32 %f2709, %f641, %f3670; sub.ftz.f32 %f2710, %f3669, %f3670; div.approx.ftz.f32 %f2711, %f2709, %f2710; mul.ftz.f32 %f2712, %f2711, 0f3FC90FDB; cos.approx.ftz.f32 %f2713, %f2712; mul.ftz.f32 %f3920, %f3759, %f2713; $L__BB203_161: mul.ftz.f32 %f2714, %f3891, 0f00000000; sub.ftz.f32 %f2715, %f3890, %f2714; sub.ftz.f32 %f2716, %f2714, %f3889; mul.ftz.f32 %f2717, %f3890, 0f00000000; mul.ftz.f32 %f2718, %f3889, 0f00000000; sub.ftz.f32 %f2719, %f2718, %f2717; mul.ftz.f32 %f2720, %f2716, %f2716; fma.rn.ftz.f32 %f2721, %f2715, %f2715, %f2720; fma.rn.ftz.f32 %f2722, %f2719, %f2719, %f2721; sqrt.approx.ftz.f32 %f2723, %f2722; div.approx.ftz.f32 %f2724, %f2715, %f2723; div.approx.ftz.f32 %f2725, %f2716, %f2723; div.approx.ftz.f32 %f2726, %f2719, %f2723; mul.ftz.f32 %f644, %f3920, %f2724; mul.ftz.f32 %f645, %f3920, %f2725; mul.ftz.f32 %f646, %f3920, %f2726; add.ftz.f32 %f3924, %f3924, %f644; add.ftz.f32 %f3925, %f3925, %f645; add.ftz.f32 %f3926, %f3926, %f646; ld.const.f32 %f650, [kernel_params+84]; setp.eq.ftz.f32 %p131, %f650, 0f7F7FFFFF; @%p131 bra $L__BB203_163; mul.ftz.f32 %f2727, %f3889, %f650; sub.ftz.f32 %f2728, %f2727, %f644; mul.ftz.f32 %f2729, %f3890, %f650; sub.ftz.f32 %f2730, %f2729, %f645; mul.ftz.f32 %f2731, %f3891, %f650; sub.ftz.f32 %f2732, %f2731, %f646; mul.ftz.f32 %f2733, %f2730, %f2730; fma.rn.ftz.f32 %f2734, %f2728, %f2728, %f2733; fma.rn.ftz.f32 %f2735, %f2732, %f2732, %f2734; sqrt.approx.ftz.f32 %f2736, %f2735; div.approx.ftz.f32 %f3889, %f2728, %f2736; div.approx.ftz.f32 %f3890, %f2730, %f2736; div.approx.ftz.f32 %f3891, %f2732, %f2736; $L__BB203_163: ld.const.u32 %r486, [kernel_params+20]; setp.gt.s32 %p234, %r486, 2; mul.ftz.f32 %f2737, %f3909, %f3925; fma.rn.ftz.f32 %f2738, %f3908, %f3924, %f2737; fma.rn.ftz.f32 %f660, %f3910, %f3926, %f2738; mul.ftz.f32 %f2739, %f3913, %f3925; fma.rn.ftz.f32 %f2740, %f3912, %f3924, %f2739; fma.rn.ftz.f32 %f661, %f3914, %f3926, %f2740; mul.ftz.f32 %f2741, %f3917, %f3925; fma.rn.ftz.f32 %f2742, %f3916, %f3924, %f2741; fma.rn.ftz.f32 %f662, %f3918, %f3926, %f2742; mul.ftz.f32 %f2743, %f3909, %f3890; fma.rn.ftz.f32 %f2744, %f3908, %f3889, %f2743; fma.rn.ftz.f32 %f2745, %f3910, %f3891, %f2744; mul.ftz.f32 %f2746, %f3913, %f3890; fma.rn.ftz.f32 %f2747, %f3912, %f3889, %f2746; fma.rn.ftz.f32 %f2748, %f3914, %f3891, %f2747; mul.ftz.f32 %f2749, %f3917, %f3890; fma.rn.ftz.f32 %f2750, %f3916, %f3889, %f2749; fma.rn.ftz.f32 %f2751, %f3918, %f3891, %f2750; mul.ftz.f32 %f2752, %f2748, %f2748; fma.rn.ftz.f32 %f2753, %f2745, %f2745, %f2752; fma.rn.ftz.f32 %f2754, %f2751, %f2751, %f2753; sqrt.approx.ftz.f32 %f2755, %f2754; div.approx.ftz.f32 %f3995, %f2745, %f2755; div.approx.ftz.f32 %f3994, %f2748, %f2755; div.approx.ftz.f32 %f3993, %f2751, %f2755; @%p234 bra $L__BB203_170; bra.uni $L__BB203_164; $L__BB203_170: ld.const.u32 %r481, [kernel_params+20]; setp.eq.s32 %p133, %r481, 3; @%p133 bra $L__BB203_177; ld.const.u32 %r482, [kernel_params+20]; setp.eq.s32 %p134, %r482, 5; @%p134 bra $L__BB203_179; ld.const.u32 %r483, [kernel_params+20]; setp.ne.s32 %p135, %r483, 4; @%p135 bra $L__BB203_181; add.ftz.f32 %f3682, %f3944, 0fBF000000; add.ftz.f32 %f3681, %f3943, 0fBF000000; ld.const.f32 %f2759, [kernel_params+48]; ld.const.v4.f32 {%f2760, %f2761, %f2762, %f2763}, [kernel_params+64]; ld.const.f32 %f2768, [kernel_params+276]; mul.ftz.f32 %f666, %f3681, %f2768; ld.const.f32 %f2769, [kernel_params+280]; mul.ftz.f32 %f667, %f3682, %f2769; mul.ftz.f32 %f2770, %f667, %f667; fma.rn.ftz.f32 %f2771, %f666, %f666, %f2770; sqrt.approx.ftz.f32 %f668, %f2771; mul.ftz.f32 %f2772, %f668, %f668; mul.ftz.f32 %f2773, %f668, %f2772; mul.ftz.f32 %f2774, %f2772, %f2772; mul.ftz.f32 %f2775, %f2761, %f2772; fma.rn.ftz.f32 %f2776, %f2760, %f668, %f2775; mul.ftz.f32 %f2777, %f2763, %f2774; fma.rn.ftz.f32 %f2778, %f2762, %f2773, %f2777; add.ftz.f32 %f2779, %f2776, %f2778; add.ftz.f32 %f2780, %f2759, %f2779; neg.ftz.f32 %f669, %f2780; abs.ftz.f32 %f2781, %f669; ld.const.f32 %f2782, [kernel_params+24]; mul.ftz.f32 %f2783, %f2782, 0f3F000000; setp.gt.ftz.f32 %p138, %f2781, %f2783; mov.f32 %f3936, 0f00000000; mov.f32 %f3937, %f3936; mov.f32 %f3938, %f3936; @%p138 bra $L__BB203_185; setp.eq.ftz.f32 %p139, %f668, 0f00000000; mov.f32 %f3927, 0f00000000; @%p139 bra $L__BB203_176; div.approx.ftz.f32 %f3927, %f666, %f668; $L__BB203_176: mov.f32 %f2785, 0fBF800000; max.ftz.f32 %f2786, %f3927, %f2785; mov.f32 %f2787, 0f3F800000; min.ftz.f32 %f2788, %f2786, %f2787; abs.ftz.f32 %f2789, %f2788; sub.ftz.f32 %f2790, %f2787, %f2789; mul.ftz.f32 %f2791, %f2790, 0f3F000000; sqrt.approx.ftz.f32 %f2792, %f2791; setp.gt.ftz.f32 %p140, %f2789, 0f3F11EB85; selp.f32 %f2793, %f2792, %f2789, %p140; mul.ftz.f32 %f2794, %f2793, %f2793; mov.f32 %f2795, 0f3C94D2E9; mov.f32 %f2796, 0f3D53F941; fma.rn.ftz.f32 %f2797, %f2796, %f2794, %f2795; mov.f32 %f2798, 0f3D3F841F; fma.rn.ftz.f32 %f2799, %f2797, %f2794, %f2798; mov.f32 %f2800, 0f3D994929; fma.rn.ftz.f32 %f2801, %f2799, %f2794, %f2800; mov.f32 %f2802, 0f3E2AAB94; fma.rn.ftz.f32 %f2803, %f2801, %f2794, %f2802; mul.ftz.f32 %f2804, %f2794, %f2803; fma.rn.ftz.f32 %f2805, %f2804, %f2793, %f2793; add.ftz.f32 %f2806, %f2805, %f2805; mov.f32 %f2807, 0f3FC90FDB; sub.ftz.f32 %f2808, %f2807, %f2805; selp.f32 %f2809, %f2806, %f2808, %p140; setp.lt.ftz.f32 %p141, %f2788, 0f00000000; mov.f32 %f2810, 0f40490FDB; sub.ftz.f32 %f2811, %f2810, %f2809; selp.f32 %f2812, %f2811, %f2809, %p141; neg.ftz.f32 %f2813, %f2812; setp.lt.ftz.f32 %p142, %f667, 0f00000000; selp.f32 %f2814, %f2813, %f2812, %p142; cos.approx.ftz.f32 %f2815, %f2814; sin.approx.ftz.f32 %f2816, %f669; mul.ftz.f32 %f2817, %f2815, %f2816; sin.approx.ftz.f32 %f2818, %f2814; mul.ftz.f32 %f3938, %f2816, %f2818; neg.ftz.f32 %f3937, %f2817; cos.approx.ftz.f32 %f3936, %f669; bra.uni $L__BB203_185; $L__BB203_164: ld.const.u32 %r484, [kernel_params+20]; setp.eq.s32 %p136, %r484, 0; @%p136 bra $L__BB203_180; ld.const.u32 %r485, [kernel_params+20]; setp.eq.s32 %p137, %r485, 1; @%p137 bra $L__BB203_166; bra.uni $L__BB203_181; $L__BB203_166: add.ftz.f32 %f3684, %f3944, 0fBF000000; add.ftz.f32 %f3683, %f3943, 0fBF000000; add.ftz.f32 %f685, %f3683, %f3683; add.ftz.f32 %f686, %f3684, %f3684; mul.ftz.f32 %f2843, %f686, %f686; fma.rn.ftz.f32 %f2844, %f685, %f685, %f2843; sqrt.approx.ftz.f32 %f687, %f2844; setp.gt.ftz.f32 %p144, %f687, 0f3F800000; mov.f32 %f3936, 0f00000000; mov.f32 %f3937, %f3936; mov.f32 %f3938, %f3936; @%p144 bra $L__BB203_185; ld.const.f32 %f688, [kernel_params+24]; setp.eq.ftz.f32 %p145, %f687, 0f00000000; mov.f32 %f3928, 0f00000000; @%p145 bra $L__BB203_169; div.approx.ftz.f32 %f3928, %f685, %f687; $L__BB203_169: mov.f32 %f2846, 0fBF800000; max.ftz.f32 %f2847, %f3928, %f2846; mov.f32 %f2848, 0f3F800000; min.ftz.f32 %f2849, %f2847, %f2848; abs.ftz.f32 %f2850, %f2849; sub.ftz.f32 %f2851, %f2848, %f2850; mul.ftz.f32 %f2852, %f2851, 0f3F000000; sqrt.approx.ftz.f32 %f2853, %f2852; setp.gt.ftz.f32 %p146, %f2850, 0f3F11EB85; selp.f32 %f2854, %f2853, %f2850, %p146; mul.ftz.f32 %f2855, %f2854, %f2854; mov.f32 %f2856, 0f3C94D2E9; mov.f32 %f2857, 0f3D53F941; fma.rn.ftz.f32 %f2858, %f2857, %f2855, %f2856; mov.f32 %f2859, 0f3D3F841F; fma.rn.ftz.f32 %f2860, %f2858, %f2855, %f2859; mov.f32 %f2861, 0f3D994929; fma.rn.ftz.f32 %f2862, %f2860, %f2855, %f2861; mov.f32 %f2863, 0f3E2AAB94; fma.rn.ftz.f32 %f2864, %f2862, %f2855, %f2863; mul.ftz.f32 %f2865, %f2855, %f2864; fma.rn.ftz.f32 %f2866, %f2865, %f2854, %f2854; add.ftz.f32 %f2867, %f2866, %f2866; mov.f32 %f2868, 0f3FC90FDB; sub.ftz.f32 %f2869, %f2868, %f2866; selp.f32 %f2870, %f2867, %f2869, %p146; setp.lt.ftz.f32 %p147, %f2849, 0f00000000; mov.f32 %f2871, 0f40490FDB; sub.ftz.f32 %f2872, %f2871, %f2870; selp.f32 %f2873, %f2872, %f2870, %p147; mul.ftz.f32 %f2874, %f688, %f687; mul.ftz.f32 %f2875, %f2874, 0f3F000000; neg.ftz.f32 %f2876, %f2873; setp.lt.ftz.f32 %p148, %f686, 0f00000000; selp.f32 %f2877, %f2876, %f2873, %p148; cos.approx.ftz.f32 %f2878, %f2877; sin.approx.ftz.f32 %f2879, %f2875; mul.ftz.f32 %f2880, %f2878, %f2879; sin.approx.ftz.f32 %f2881, %f2877; mul.ftz.f32 %f3938, %f2879, %f2881; neg.ftz.f32 %f3937, %f2880; cos.approx.ftz.f32 %f3936, %f2875; bra.uni $L__BB203_185; $L__BB203_177: fma.rn.ftz.f32 %f675, %f3943, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f676, %f3944, 0f40000000, 0fBF800000; mul.ftz.f32 %f677, %f675, %f675; mul.ftz.f32 %f678, %f676, %f676; add.ftz.f32 %f2822, %f677, %f678; setp.gt.ftz.f32 %p143, %f2822, 0f3F800000; mov.f32 %f3936, 0f00000000; mov.f32 %f3937, %f3936; mov.f32 %f3938, %f3936; @%p143 bra $L__BB203_185; mov.f32 %f2823, 0f3F800000; sub.ftz.f32 %f2824, %f2823, %f677; sub.ftz.f32 %f2825, %f2824, %f678; mov.f32 %f2826, 0f00000000; max.ftz.f32 %f2827, %f2825, %f2826; sqrt.approx.ftz.f32 %f2828, %f2827; fma.rn.ftz.f32 %f2829, %f675, 0f00000000, %f2828; fma.rn.ftz.f32 %f2830, %f676, 0f00000000, %f2829; add.ftz.f32 %f2831, %f2830, %f2830; mul.ftz.f32 %f2832, %f2828, %f2831; sub.ftz.f32 %f3937, %f2823, %f2832; mul.ftz.f32 %f3938, %f676, %f2831; mul.ftz.f32 %f3936, %f675, %f2831; bra.uni $L__BB203_185; $L__BB203_179: mov.f32 %f3936, 0f3F800000; sub.ftz.f32 %f2833, %f3936, %f3943; fma.rn.ftz.f32 %f2834, %f2833, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f2835, %f2834; cos.approx.ftz.f32 %f2836, %f2834; fma.rn.ftz.f32 %f2837, %f3944, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f2838, %f2837; cos.approx.ftz.f32 %f2839, %f2837; div.approx.ftz.f32 %f3937, %f2835, %f2836; div.approx.ftz.f32 %f3938, %f2838, %f2839; bra.uni $L__BB203_185; $L__BB203_180: ld.const.v4.f32 {%f2882, %f2883, %f2884, %f2885}, [kernel_params+32]; fma.rn.ftz.f32 %f2890, %f3943, %f2882, %f2883; fma.rn.ftz.f32 %f2891, %f3944, %f2884, %f2885; sin.approx.ftz.f32 %f2892, %f2891; cos.approx.ftz.f32 %f2893, %f2890; sin.approx.ftz.f32 %f2894, %f2890; mul.ftz.f32 %f3937, %f2892, %f2894; mul.ftz.f32 %f3936, %f2892, %f2893; cos.approx.ftz.f32 %f3938, %f2891; bra.uni $L__BB203_185; $L__BB203_181: add.ftz.f32 %f3672, %f3944, 0fBF000000; add.ftz.f32 %f3671, %f3943, 0fBF000000; ld.const.v2.f32 {%f2898, %f2899}, [kernel_params+24]; ld.const.f32 %f2902, [kernel_params+276]; mul.ftz.f32 %f703, %f3671, %f2902; ld.const.f32 %f2903, [kernel_params+280]; mul.ftz.f32 %f704, %f3672, %f2903; add.ftz.f32 %f705, %f2899, %f2899; mul.ftz.f32 %f2904, %f2898, 0f3E800000; sin.approx.ftz.f32 %f2905, %f2904; mul.ftz.f32 %f2906, %f705, %f2905; mul.ftz.f32 %f2907, %f704, %f704; fma.rn.ftz.f32 %f2908, %f703, %f703, %f2907; sqrt.approx.ftz.f32 %f706, %f2908; setp.gt.ftz.f32 %p149, %f706, %f2906; mov.f32 %f3936, 0f00000000; mov.f32 %f3937, %f3936; mov.f32 %f3938, %f3936; @%p149 bra $L__BB203_185; setp.eq.ftz.f32 %p150, %f706, 0f00000000; mov.f32 %f3932, 0f00000000; @%p150 bra $L__BB203_184; div.approx.ftz.f32 %f3932, %f703, %f706; $L__BB203_184: mov.f32 %f2910, 0fBF800000; max.ftz.f32 %f2911, %f3932, %f2910; mov.f32 %f2912, 0f3F800000; min.ftz.f32 %f2913, %f2911, %f2912; abs.ftz.f32 %f2914, %f2913; sub.ftz.f32 %f2915, %f2912, %f2914; mul.ftz.f32 %f2916, %f2915, 0f3F000000; sqrt.approx.ftz.f32 %f2917, %f2916; setp.gt.ftz.f32 %p151, %f2914, 0f3F11EB85; selp.f32 %f2918, %f2917, %f2914, %p151; mul.ftz.f32 %f2919, %f2918, %f2918; mov.f32 %f2920, 0f3C94D2E9; mov.f32 %f2921, 0f3D53F941; fma.rn.ftz.f32 %f2922, %f2921, %f2919, %f2920; mov.f32 %f2923, 0f3D3F841F; fma.rn.ftz.f32 %f2924, %f2922, %f2919, %f2923; mov.f32 %f2925, 0f3D994929; fma.rn.ftz.f32 %f2926, %f2924, %f2919, %f2925; mov.f32 %f2927, 0f3E2AAB94; fma.rn.ftz.f32 %f2928, %f2926, %f2919, %f2927; mul.ftz.f32 %f2929, %f2919, %f2928; fma.rn.ftz.f32 %f2930, %f2929, %f2918, %f2918; add.ftz.f32 %f2931, %f2930, %f2930; mov.f32 %f2932, 0f3FC90FDB; sub.ftz.f32 %f2933, %f2932, %f2930; selp.f32 %f2934, %f2931, %f2933, %p151; setp.lt.ftz.f32 %p152, %f2913, 0f00000000; mov.f32 %f2935, 0f40490FDB; sub.ftz.f32 %f2936, %f2935, %f2934; selp.f32 %f2937, %f2936, %f2934, %p152; div.approx.ftz.f32 %f2938, %f706, %f705; abs.ftz.f32 %f2939, %f2938; sub.ftz.f32 %f2940, %f2912, %f2939; mul.ftz.f32 %f2941, %f2940, 0f3F000000; sqrt.approx.ftz.f32 %f2942, %f2941; setp.gt.ftz.f32 %p153, %f2939, 0f3F11EB85; selp.f32 %f2943, %f2942, %f2939, %p153; mul.ftz.f32 %f2944, %f2943, %f2943; fma.rn.ftz.f32 %f2945, %f2921, %f2944, %f2920; fma.rn.ftz.f32 %f2946, %f2945, %f2944, %f2923; fma.rn.ftz.f32 %f2947, %f2946, %f2944, %f2925; fma.rn.ftz.f32 %f2948, %f2947, %f2944, %f2927; mul.ftz.f32 %f2949, %f2944, %f2948; fma.rn.ftz.f32 %f2950, %f2949, %f2943, %f2943; mov.f32 %f2951, 0fC0000000; fma.rn.ftz.f32 %f2952, %f2951, %f2950, %f2932; selp.f32 %f2953, %f2952, %f2950, %p153; setp.le.ftz.f32 %p154, %f2953, 0f7F800000; mov.b32 %r430, %f2953; mov.b32 %r431, %f2938; and.b32 %r432, %r431, -2147483648; or.b32 %r433, %r432, %r430; mov.b32 %f2954, %r433; selp.f32 %f2955, %f2954, %f2953, %p154; add.ftz.f32 %f2956, %f2955, %f2955; neg.ftz.f32 %f2957, %f2937; setp.lt.ftz.f32 %p155, %f704, 0f00000000; selp.f32 %f2958, %f2957, %f2937, %p155; cos.approx.ftz.f32 %f2959, %f2958; sin.approx.ftz.f32 %f2960, %f2956; mul.ftz.f32 %f2961, %f2959, %f2960; sin.approx.ftz.f32 %f2962, %f2958; mul.ftz.f32 %f3938, %f2960, %f2962; neg.ftz.f32 %f3937, %f2961; cos.approx.ftz.f32 %f3936, %f2956; $L__BB203_185: ld.const.f32 %f3706, [kernel_params+80]; setp.eq.ftz.f32 %p237, %f3706, 0f00000000; @%p237 bra $L__BB203_192; ld.const.f32 %f3939, [kernel_params+80]; ld.const.f32 %f725, [kernel_params+92]; setp.leu.ftz.f32 %p157, %f725, 0f00000000; @%p157 bra $L__BB203_190; mov.f32 %f2964, 0fBF800000; max.ftz.f32 %f2965, %f3938, %f2964; mov.f32 %f2966, 0f3F800000; min.ftz.f32 %f2967, %f2965, %f2966; abs.ftz.f32 %f2968, %f2967; sub.ftz.f32 %f2969, %f2966, %f2968; mul.ftz.f32 %f2970, %f2969, 0f3F000000; sqrt.approx.ftz.f32 %f2971, %f2970; setp.gt.ftz.f32 %p158, %f2968, 0f3F11EB85; selp.f32 %f2972, %f2971, %f2968, %p158; mul.ftz.f32 %f2973, %f2972, %f2972; mov.f32 %f2974, 0f3C94D2E9; mov.f32 %f2975, 0f3D53F941; fma.rn.ftz.f32 %f2976, %f2975, %f2973, %f2974; mov.f32 %f2977, 0f3D3F841F; fma.rn.ftz.f32 %f2978, %f2976, %f2973, %f2977; mov.f32 %f2979, 0f3D994929; fma.rn.ftz.f32 %f2980, %f2978, %f2973, %f2979; mov.f32 %f2981, 0f3E2AAB94; fma.rn.ftz.f32 %f2982, %f2980, %f2973, %f2981; mul.ftz.f32 %f2983, %f2973, %f2982; fma.rn.ftz.f32 %f2984, %f2983, %f2972, %f2972; mov.f32 %f2985, 0f3FC90FDB; mov.f32 %f2986, 0fC0000000; fma.rn.ftz.f32 %f2987, %f2986, %f2984, %f2985; selp.f32 %f2988, %f2987, %f2984, %p158; setp.le.ftz.f32 %p159, %f2988, 0f7F800000; mov.b32 %r434, %f2988; mov.b32 %r435, %f2967; and.b32 %r436, %r435, -2147483648; or.b32 %r437, %r436, %r434; mov.b32 %f2989, %r437; selp.f32 %f2990, %f2989, %f2988, %p159; abs.ftz.f32 %f727, %f2990; setp.gt.ftz.f32 %p160, %f727, %f725; mov.f32 %f3939, 0f00000000; @%p160 bra $L__BB203_190; ld.const.f32 %f3701, [kernel_params+88]; ld.const.f32 %f3939, [kernel_params+80]; setp.leu.ftz.f32 %p161, %f727, %f3701; @%p161 bra $L__BB203_190; ld.const.f32 %f3703, [kernel_params+88]; ld.const.f32 %f3702, [kernel_params+92]; ld.const.f32 %f3693, [kernel_params+80]; sub.ftz.f32 %f2991, %f727, %f3703; sub.ftz.f32 %f2992, %f3702, %f3703; div.approx.ftz.f32 %f2993, %f2991, %f2992; mul.ftz.f32 %f2994, %f2993, 0f3FC90FDB; cos.approx.ftz.f32 %f2995, %f2994; mul.ftz.f32 %f3939, %f3693, %f2995; $L__BB203_190: mul.ftz.f32 %f2996, %f3938, 0f00000000; sub.ftz.f32 %f2997, %f3937, %f2996; sub.ftz.f32 %f2998, %f2996, %f3936; mul.ftz.f32 %f2999, %f3937, 0f00000000; mul.ftz.f32 %f3000, %f3936, 0f00000000; sub.ftz.f32 %f3001, %f3000, %f2999; mul.ftz.f32 %f3002, %f2998, %f2998; fma.rn.ftz.f32 %f3003, %f2997, %f2997, %f3002; fma.rn.ftz.f32 %f3004, %f3001, %f3001, %f3003; sqrt.approx.ftz.f32 %f3005, %f3004; div.approx.ftz.f32 %f3006, %f2997, %f3005; div.approx.ftz.f32 %f3007, %f2998, %f3005; div.approx.ftz.f32 %f3008, %f3001, %f3005; mul.ftz.f32 %f730, %f3939, %f3006; mul.ftz.f32 %f731, %f3939, %f3007; mul.ftz.f32 %f732, %f3939, %f3008; add.ftz.f32 %f3943, %f3943, %f730; add.ftz.f32 %f3944, %f3944, %f731; add.ftz.f32 %f3945, %f3945, %f732; ld.const.f32 %f736, [kernel_params+84]; setp.eq.ftz.f32 %p162, %f736, 0f7F7FFFFF; @%p162 bra $L__BB203_192; mul.ftz.f32 %f3009, %f3936, %f736; sub.ftz.f32 %f3010, %f3009, %f730; mul.ftz.f32 %f3011, %f3937, %f736; sub.ftz.f32 %f3012, %f3011, %f731; mul.ftz.f32 %f3013, %f3938, %f736; sub.ftz.f32 %f3014, %f3013, %f732; mul.ftz.f32 %f3015, %f3012, %f3012; fma.rn.ftz.f32 %f3016, %f3010, %f3010, %f3015; fma.rn.ftz.f32 %f3017, %f3014, %f3014, %f3016; sqrt.approx.ftz.f32 %f3018, %f3017; div.approx.ftz.f32 %f3936, %f3010, %f3018; div.approx.ftz.f32 %f3937, %f3012, %f3018; div.approx.ftz.f32 %f3938, %f3014, %f3018; $L__BB203_192: mul.ftz.f32 %f3676, %f50, %f2166; mul.ftz.f32 %f3675, %f50, %f2160; mul.ftz.f32 %f3674, %f50, %f2154; mul.ftz.f32 %f3673, %f50, %f2148; mul.ftz.f32 %f3022, %f3909, %f3944; fma.rn.ftz.f32 %f3023, %f3908, %f3943, %f3022; fma.rn.ftz.f32 %f3024, %f3910, %f3945, %f3023; add.ftz.f32 %f746, %f3911, %f3024; mul.ftz.f32 %f3025, %f3913, %f3944; fma.rn.ftz.f32 %f3026, %f3912, %f3943, %f3025; fma.rn.ftz.f32 %f3027, %f3914, %f3945, %f3026; add.ftz.f32 %f747, %f3915, %f3027; mul.ftz.f32 %f3028, %f3917, %f3944; fma.rn.ftz.f32 %f3029, %f3916, %f3943, %f3028; fma.rn.ftz.f32 %f3030, %f3918, %f3945, %f3029; add.ftz.f32 %f748, %f3919, %f3030; mul.ftz.f32 %f3031, %f3909, %f3937; fma.rn.ftz.f32 %f3032, %f3908, %f3936, %f3031; fma.rn.ftz.f32 %f3033, %f3910, %f3938, %f3032; mul.ftz.f32 %f3034, %f3913, %f3937; fma.rn.ftz.f32 %f3035, %f3912, %f3936, %f3034; fma.rn.ftz.f32 %f3036, %f3914, %f3938, %f3035; mul.ftz.f32 %f3037, %f3917, %f3937; fma.rn.ftz.f32 %f3038, %f3916, %f3936, %f3037; fma.rn.ftz.f32 %f3039, %f3918, %f3938, %f3038; mul.ftz.f32 %f3040, %f3036, %f3036; fma.rn.ftz.f32 %f3041, %f3033, %f3033, %f3040; fma.rn.ftz.f32 %f3042, %f3039, %f3039, %f3041; sqrt.approx.ftz.f32 %f3043, %f3042; div.approx.ftz.f32 %f749, %f3033, %f3043; div.approx.ftz.f32 %f750, %f3036, %f3043; div.approx.ftz.f32 %f751, %f3039, %f3043; add.ftz.f32 %f3044, %f44, 0f3F800000; fma.rn.ftz.f32 %f752, %f3044, %f2147, %f3673; fma.rn.ftz.f32 %f753, %f3044, %f2153, %f3674; fma.rn.ftz.f32 %f754, %f3044, %f2159, %f3675; fma.rn.ftz.f32 %f3045, %f3044, %f2165, %f3676; add.ftz.f32 %f755, %f3045, %f480; setp.eq.ftz.f32 %p163, %f755, 0f00000000; mov.f32 %f3965, 0f00000000; mov.f32 %f3966, %f3965; mov.f32 %f3967, %f3965; @%p163 bra $L__BB203_194; add.ftz.f32 %f3046, %f752, %f469; add.ftz.f32 %f3047, %f753, %f473; add.ftz.f32 %f3048, %f754, %f477; div.approx.ftz.f32 %f3967, %f3048, %f755; div.approx.ftz.f32 %f3966, %f3047, %f755; div.approx.ftz.f32 %f3965, %f3046, %f755; $L__BB203_194: ld.const.u32 %r469, [kernel_params+20]; setp.gt.s32 %p231, %r469, 2; add.ftz.f32 %f763, %f3965, 0fBF000000; add.ftz.f32 %f765, %f3966, 0fBF000000; @%p231 bra $L__BB203_201; bra.uni $L__BB203_195; $L__BB203_201: ld.const.u32 %r470, [kernel_params+20]; setp.eq.s32 %p165, %r470, 3; @%p165 bra $L__BB203_208; ld.const.u32 %r471, [kernel_params+20]; setp.eq.s32 %p166, %r471, 5; @%p166 bra $L__BB203_210; ld.const.u32 %r472, [kernel_params+20]; setp.ne.s32 %p167, %r472, 4; @%p167 bra $L__BB203_212; ld.const.f32 %f3052, [kernel_params+48]; ld.const.v4.f32 {%f3053, %f3054, %f3055, %f3056}, [kernel_params+64]; ld.const.f32 %f3061, [kernel_params+276]; mul.ftz.f32 %f766, %f763, %f3061; ld.const.f32 %f3062, [kernel_params+280]; mul.ftz.f32 %f767, %f765, %f3062; mul.ftz.f32 %f3063, %f767, %f767; fma.rn.ftz.f32 %f3064, %f766, %f766, %f3063; sqrt.approx.ftz.f32 %f768, %f3064; mul.ftz.f32 %f3065, %f768, %f768; mul.ftz.f32 %f3066, %f768, %f3065; mul.ftz.f32 %f3067, %f3065, %f3065; mul.ftz.f32 %f3068, %f3054, %f3065; fma.rn.ftz.f32 %f3069, %f3053, %f768, %f3068; mul.ftz.f32 %f3070, %f3056, %f3067; fma.rn.ftz.f32 %f3071, %f3055, %f3066, %f3070; add.ftz.f32 %f3072, %f3069, %f3071; add.ftz.f32 %f3073, %f3052, %f3072; neg.ftz.f32 %f769, %f3073; abs.ftz.f32 %f3074, %f769; ld.const.f32 %f3075, [kernel_params+24]; mul.ftz.f32 %f3076, %f3075, 0f3F000000; setp.gt.ftz.f32 %p170, %f3074, %f3076; mov.f32 %f3958, 0f00000000; mov.f32 %f3959, %f3958; mov.f32 %f3960, %f3958; @%p170 bra $L__BB203_216; setp.eq.ftz.f32 %p171, %f768, 0f00000000; mov.f32 %f3949, 0f00000000; @%p171 bra $L__BB203_207; div.approx.ftz.f32 %f3949, %f766, %f768; $L__BB203_207: mov.f32 %f3078, 0fBF800000; max.ftz.f32 %f3079, %f3949, %f3078; mov.f32 %f3080, 0f3F800000; min.ftz.f32 %f3081, %f3079, %f3080; abs.ftz.f32 %f3082, %f3081; sub.ftz.f32 %f3083, %f3080, %f3082; mul.ftz.f32 %f3084, %f3083, 0f3F000000; sqrt.approx.ftz.f32 %f3085, %f3084; setp.gt.ftz.f32 %p172, %f3082, 0f3F11EB85; selp.f32 %f3086, %f3085, %f3082, %p172; mul.ftz.f32 %f3087, %f3086, %f3086; mov.f32 %f3088, 0f3C94D2E9; mov.f32 %f3089, 0f3D53F941; fma.rn.ftz.f32 %f3090, %f3089, %f3087, %f3088; mov.f32 %f3091, 0f3D3F841F; fma.rn.ftz.f32 %f3092, %f3090, %f3087, %f3091; mov.f32 %f3093, 0f3D994929; fma.rn.ftz.f32 %f3094, %f3092, %f3087, %f3093; mov.f32 %f3095, 0f3E2AAB94; fma.rn.ftz.f32 %f3096, %f3094, %f3087, %f3095; mul.ftz.f32 %f3097, %f3087, %f3096; fma.rn.ftz.f32 %f3098, %f3097, %f3086, %f3086; add.ftz.f32 %f3099, %f3098, %f3098; mov.f32 %f3100, 0f3FC90FDB; sub.ftz.f32 %f3101, %f3100, %f3098; selp.f32 %f3102, %f3099, %f3101, %p172; setp.lt.ftz.f32 %p173, %f3081, 0f00000000; mov.f32 %f3103, 0f40490FDB; sub.ftz.f32 %f3104, %f3103, %f3102; selp.f32 %f3105, %f3104, %f3102, %p173; neg.ftz.f32 %f3106, %f3105; setp.lt.ftz.f32 %p174, %f767, 0f00000000; selp.f32 %f3107, %f3106, %f3105, %p174; cos.approx.ftz.f32 %f3108, %f3107; sin.approx.ftz.f32 %f3109, %f769; mul.ftz.f32 %f3110, %f3108, %f3109; sin.approx.ftz.f32 %f3111, %f3107; mul.ftz.f32 %f3960, %f3109, %f3111; neg.ftz.f32 %f3959, %f3110; cos.approx.ftz.f32 %f3958, %f769; bra.uni $L__BB203_216; $L__BB203_195: ld.const.u32 %r479, [kernel_params+20]; setp.eq.s32 %p168, %r479, 0; @%p168 bra $L__BB203_211; ld.const.u32 %r480, [kernel_params+20]; setp.eq.s32 %p169, %r480, 1; @%p169 bra $L__BB203_197; bra.uni $L__BB203_212; $L__BB203_197: add.ftz.f32 %f785, %f763, %f763; add.ftz.f32 %f786, %f765, %f765; mul.ftz.f32 %f3136, %f786, %f786; fma.rn.ftz.f32 %f3137, %f785, %f785, %f3136; sqrt.approx.ftz.f32 %f787, %f3137; setp.gt.ftz.f32 %p176, %f787, 0f3F800000; mov.f32 %f3958, 0f00000000; mov.f32 %f3959, %f3958; mov.f32 %f3960, %f3958; @%p176 bra $L__BB203_216; ld.const.f32 %f788, [kernel_params+24]; setp.eq.ftz.f32 %p177, %f787, 0f00000000; mov.f32 %f3950, 0f00000000; @%p177 bra $L__BB203_200; div.approx.ftz.f32 %f3950, %f785, %f787; $L__BB203_200: mov.f32 %f3139, 0fBF800000; max.ftz.f32 %f3140, %f3950, %f3139; mov.f32 %f3141, 0f3F800000; min.ftz.f32 %f3142, %f3140, %f3141; abs.ftz.f32 %f3143, %f3142; sub.ftz.f32 %f3144, %f3141, %f3143; mul.ftz.f32 %f3145, %f3144, 0f3F000000; sqrt.approx.ftz.f32 %f3146, %f3145; setp.gt.ftz.f32 %p178, %f3143, 0f3F11EB85; selp.f32 %f3147, %f3146, %f3143, %p178; mul.ftz.f32 %f3148, %f3147, %f3147; mov.f32 %f3149, 0f3C94D2E9; mov.f32 %f3150, 0f3D53F941; fma.rn.ftz.f32 %f3151, %f3150, %f3148, %f3149; mov.f32 %f3152, 0f3D3F841F; fma.rn.ftz.f32 %f3153, %f3151, %f3148, %f3152; mov.f32 %f3154, 0f3D994929; fma.rn.ftz.f32 %f3155, %f3153, %f3148, %f3154; mov.f32 %f3156, 0f3E2AAB94; fma.rn.ftz.f32 %f3157, %f3155, %f3148, %f3156; mul.ftz.f32 %f3158, %f3148, %f3157; fma.rn.ftz.f32 %f3159, %f3158, %f3147, %f3147; add.ftz.f32 %f3160, %f3159, %f3159; mov.f32 %f3161, 0f3FC90FDB; sub.ftz.f32 %f3162, %f3161, %f3159; selp.f32 %f3163, %f3160, %f3162, %p178; setp.lt.ftz.f32 %p179, %f3142, 0f00000000; mov.f32 %f3164, 0f40490FDB; sub.ftz.f32 %f3165, %f3164, %f3163; selp.f32 %f3166, %f3165, %f3163, %p179; mul.ftz.f32 %f3167, %f788, %f787; mul.ftz.f32 %f3168, %f3167, 0f3F000000; neg.ftz.f32 %f3169, %f3166; setp.lt.ftz.f32 %p180, %f786, 0f00000000; selp.f32 %f3170, %f3169, %f3166, %p180; cos.approx.ftz.f32 %f3171, %f3170; sin.approx.ftz.f32 %f3172, %f3168; mul.ftz.f32 %f3173, %f3171, %f3172; sin.approx.ftz.f32 %f3174, %f3170; mul.ftz.f32 %f3960, %f3172, %f3174; neg.ftz.f32 %f3959, %f3173; cos.approx.ftz.f32 %f3958, %f3168; bra.uni $L__BB203_216; $L__BB203_208: fma.rn.ftz.f32 %f775, %f3965, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f776, %f3966, 0f40000000, 0fBF800000; mul.ftz.f32 %f777, %f775, %f775; mul.ftz.f32 %f778, %f776, %f776; add.ftz.f32 %f3115, %f777, %f778; setp.gt.ftz.f32 %p175, %f3115, 0f3F800000; mov.f32 %f3958, 0f00000000; mov.f32 %f3959, %f3958; mov.f32 %f3960, %f3958; @%p175 bra $L__BB203_216; mov.f32 %f3116, 0f3F800000; sub.ftz.f32 %f3117, %f3116, %f777; sub.ftz.f32 %f3118, %f3117, %f778; mov.f32 %f3119, 0f00000000; max.ftz.f32 %f3120, %f3118, %f3119; sqrt.approx.ftz.f32 %f3121, %f3120; fma.rn.ftz.f32 %f3122, %f775, 0f00000000, %f3121; fma.rn.ftz.f32 %f3123, %f776, 0f00000000, %f3122; add.ftz.f32 %f3124, %f3123, %f3123; mul.ftz.f32 %f3125, %f3121, %f3124; sub.ftz.f32 %f3959, %f3116, %f3125; mul.ftz.f32 %f3960, %f776, %f3124; mul.ftz.f32 %f3958, %f775, %f3124; bra.uni $L__BB203_216; $L__BB203_210: mov.f32 %f3958, 0f3F800000; sub.ftz.f32 %f3126, %f3958, %f3965; fma.rn.ftz.f32 %f3127, %f3126, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f3128, %f3127; cos.approx.ftz.f32 %f3129, %f3127; fma.rn.ftz.f32 %f3130, %f3966, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f3131, %f3130; cos.approx.ftz.f32 %f3132, %f3130; div.approx.ftz.f32 %f3959, %f3128, %f3129; div.approx.ftz.f32 %f3960, %f3131, %f3132; bra.uni $L__BB203_216; $L__BB203_211: ld.const.v4.f32 {%f3175, %f3176, %f3177, %f3178}, [kernel_params+32]; fma.rn.ftz.f32 %f3183, %f3965, %f3175, %f3176; fma.rn.ftz.f32 %f3184, %f3966, %f3177, %f3178; sin.approx.ftz.f32 %f3185, %f3184; cos.approx.ftz.f32 %f3186, %f3183; sin.approx.ftz.f32 %f3187, %f3183; mul.ftz.f32 %f3959, %f3185, %f3187; mul.ftz.f32 %f3958, %f3185, %f3186; cos.approx.ftz.f32 %f3960, %f3184; bra.uni $L__BB203_216; $L__BB203_212: ld.const.v2.f32 {%f3191, %f3192}, [kernel_params+24]; ld.const.f32 %f3195, [kernel_params+276]; mul.ftz.f32 %f803, %f763, %f3195; ld.const.f32 %f3196, [kernel_params+280]; mul.ftz.f32 %f804, %f765, %f3196; add.ftz.f32 %f805, %f3192, %f3192; mul.ftz.f32 %f3197, %f3191, 0f3E800000; sin.approx.ftz.f32 %f3198, %f3197; mul.ftz.f32 %f3199, %f805, %f3198; mul.ftz.f32 %f3200, %f804, %f804; fma.rn.ftz.f32 %f3201, %f803, %f803, %f3200; sqrt.approx.ftz.f32 %f806, %f3201; setp.gt.ftz.f32 %p181, %f806, %f3199; mov.f32 %f3958, 0f00000000; mov.f32 %f3959, %f3958; mov.f32 %f3960, %f3958; @%p181 bra $L__BB203_216; setp.eq.ftz.f32 %p182, %f806, 0f00000000; mov.f32 %f3954, 0f00000000; @%p182 bra $L__BB203_215; div.approx.ftz.f32 %f3954, %f803, %f806; $L__BB203_215: mov.f32 %f3203, 0fBF800000; max.ftz.f32 %f3204, %f3954, %f3203; mov.f32 %f3205, 0f3F800000; min.ftz.f32 %f3206, %f3204, %f3205; abs.ftz.f32 %f3207, %f3206; sub.ftz.f32 %f3208, %f3205, %f3207; mul.ftz.f32 %f3209, %f3208, 0f3F000000; sqrt.approx.ftz.f32 %f3210, %f3209; setp.gt.ftz.f32 %p183, %f3207, 0f3F11EB85; selp.f32 %f3211, %f3210, %f3207, %p183; mul.ftz.f32 %f3212, %f3211, %f3211; mov.f32 %f3213, 0f3C94D2E9; mov.f32 %f3214, 0f3D53F941; fma.rn.ftz.f32 %f3215, %f3214, %f3212, %f3213; mov.f32 %f3216, 0f3D3F841F; fma.rn.ftz.f32 %f3217, %f3215, %f3212, %f3216; mov.f32 %f3218, 0f3D994929; fma.rn.ftz.f32 %f3219, %f3217, %f3212, %f3218; mov.f32 %f3220, 0f3E2AAB94; fma.rn.ftz.f32 %f3221, %f3219, %f3212, %f3220; mul.ftz.f32 %f3222, %f3212, %f3221; fma.rn.ftz.f32 %f3223, %f3222, %f3211, %f3211; add.ftz.f32 %f3224, %f3223, %f3223; mov.f32 %f3225, 0f3FC90FDB; sub.ftz.f32 %f3226, %f3225, %f3223; selp.f32 %f3227, %f3224, %f3226, %p183; setp.lt.ftz.f32 %p184, %f3206, 0f00000000; mov.f32 %f3228, 0f40490FDB; sub.ftz.f32 %f3229, %f3228, %f3227; selp.f32 %f3230, %f3229, %f3227, %p184; div.approx.ftz.f32 %f3231, %f806, %f805; abs.ftz.f32 %f3232, %f3231; sub.ftz.f32 %f3233, %f3205, %f3232; mul.ftz.f32 %f3234, %f3233, 0f3F000000; sqrt.approx.ftz.f32 %f3235, %f3234; setp.gt.ftz.f32 %p185, %f3232, 0f3F11EB85; selp.f32 %f3236, %f3235, %f3232, %p185; mul.ftz.f32 %f3237, %f3236, %f3236; fma.rn.ftz.f32 %f3238, %f3214, %f3237, %f3213; fma.rn.ftz.f32 %f3239, %f3238, %f3237, %f3216; fma.rn.ftz.f32 %f3240, %f3239, %f3237, %f3218; fma.rn.ftz.f32 %f3241, %f3240, %f3237, %f3220; mul.ftz.f32 %f3242, %f3237, %f3241; fma.rn.ftz.f32 %f3243, %f3242, %f3236, %f3236; mov.f32 %f3244, 0fC0000000; fma.rn.ftz.f32 %f3245, %f3244, %f3243, %f3225; selp.f32 %f3246, %f3245, %f3243, %p185; setp.le.ftz.f32 %p186, %f3246, 0f7F800000; mov.b32 %r438, %f3246; mov.b32 %r439, %f3231; and.b32 %r440, %r439, -2147483648; or.b32 %r441, %r440, %r438; mov.b32 %f3247, %r441; selp.f32 %f3248, %f3247, %f3246, %p186; add.ftz.f32 %f3249, %f3248, %f3248; neg.ftz.f32 %f3250, %f3230; setp.lt.ftz.f32 %p187, %f804, 0f00000000; selp.f32 %f3251, %f3250, %f3230, %p187; cos.approx.ftz.f32 %f3252, %f3251; sin.approx.ftz.f32 %f3253, %f3249; mul.ftz.f32 %f3254, %f3252, %f3253; sin.approx.ftz.f32 %f3255, %f3251; mul.ftz.f32 %f3960, %f3253, %f3255; neg.ftz.f32 %f3959, %f3254; cos.approx.ftz.f32 %f3958, %f3249; $L__BB203_216: ld.const.f32 %f3704, [kernel_params+80]; setp.eq.ftz.f32 %p235, %f3704, 0f00000000; @%p235 bra $L__BB203_223; ld.const.f32 %f3961, [kernel_params+80]; ld.const.f32 %f825, [kernel_params+92]; setp.leu.ftz.f32 %p189, %f825, 0f00000000; @%p189 bra $L__BB203_221; ld.const.f32 %f826, [kernel_params+88]; mov.f32 %f3257, 0fBF800000; max.ftz.f32 %f3258, %f3960, %f3257; mov.f32 %f3259, 0f3F800000; min.ftz.f32 %f3260, %f3258, %f3259; abs.ftz.f32 %f3261, %f3260; sub.ftz.f32 %f3262, %f3259, %f3261; mul.ftz.f32 %f3263, %f3262, 0f3F000000; sqrt.approx.ftz.f32 %f3264, %f3263; setp.gt.ftz.f32 %p190, %f3261, 0f3F11EB85; selp.f32 %f3265, %f3264, %f3261, %p190; mul.ftz.f32 %f3266, %f3265, %f3265; mov.f32 %f3267, 0f3C94D2E9; mov.f32 %f3268, 0f3D53F941; fma.rn.ftz.f32 %f3269, %f3268, %f3266, %f3267; mov.f32 %f3270, 0f3D3F841F; fma.rn.ftz.f32 %f3271, %f3269, %f3266, %f3270; mov.f32 %f3272, 0f3D994929; fma.rn.ftz.f32 %f3273, %f3271, %f3266, %f3272; mov.f32 %f3274, 0f3E2AAB94; fma.rn.ftz.f32 %f3275, %f3273, %f3266, %f3274; mul.ftz.f32 %f3276, %f3266, %f3275; fma.rn.ftz.f32 %f3277, %f3276, %f3265, %f3265; mov.f32 %f3278, 0f3FC90FDB; mov.f32 %f3279, 0fC0000000; fma.rn.ftz.f32 %f3280, %f3279, %f3277, %f3278; selp.f32 %f3281, %f3280, %f3277, %p190; setp.le.ftz.f32 %p191, %f3281, 0f7F800000; mov.b32 %r442, %f3281; mov.b32 %r443, %f3260; and.b32 %r444, %r443, -2147483648; or.b32 %r445, %r444, %r442; mov.b32 %f3282, %r445; selp.f32 %f3283, %f3282, %f3281, %p191; abs.ftz.f32 %f827, %f3283; setp.gt.ftz.f32 %p192, %f827, %f825; mov.f32 %f3961, 0f00000000; @%p192 bra $L__BB203_221; ld.const.f32 %f3961, [kernel_params+80]; setp.leu.ftz.f32 %p193, %f827, %f826; @%p193 bra $L__BB203_221; ld.const.f32 %f3690, [kernel_params+80]; sub.ftz.f32 %f3284, %f827, %f826; sub.ftz.f32 %f3285, %f825, %f826; div.approx.ftz.f32 %f3286, %f3284, %f3285; mul.ftz.f32 %f3287, %f3286, 0f3FC90FDB; cos.approx.ftz.f32 %f3288, %f3287; mul.ftz.f32 %f3961, %f3690, %f3288; $L__BB203_221: mul.ftz.f32 %f3289, %f3960, 0f00000000; sub.ftz.f32 %f3290, %f3959, %f3289; sub.ftz.f32 %f3291, %f3289, %f3958; mul.ftz.f32 %f3292, %f3959, 0f00000000; mul.ftz.f32 %f3293, %f3958, 0f00000000; sub.ftz.f32 %f3294, %f3293, %f3292; mul.ftz.f32 %f3295, %f3291, %f3291; fma.rn.ftz.f32 %f3296, %f3290, %f3290, %f3295; fma.rn.ftz.f32 %f3297, %f3294, %f3294, %f3296; sqrt.approx.ftz.f32 %f3298, %f3297; div.approx.ftz.f32 %f3299, %f3290, %f3298; div.approx.ftz.f32 %f3300, %f3291, %f3298; div.approx.ftz.f32 %f3301, %f3294, %f3298; mul.ftz.f32 %f830, %f3961, %f3299; mul.ftz.f32 %f831, %f3961, %f3300; mul.ftz.f32 %f832, %f3961, %f3301; add.ftz.f32 %f3965, %f3965, %f830; add.ftz.f32 %f3966, %f3966, %f831; add.ftz.f32 %f3967, %f3967, %f832; ld.const.f32 %f836, [kernel_params+84]; setp.eq.ftz.f32 %p194, %f836, 0f7F7FFFFF; @%p194 bra $L__BB203_223; mul.ftz.f32 %f3302, %f3958, %f836; sub.ftz.f32 %f3303, %f3302, %f830; mul.ftz.f32 %f3304, %f3959, %f836; sub.ftz.f32 %f3305, %f3304, %f831; mul.ftz.f32 %f3306, %f3960, %f836; sub.ftz.f32 %f3307, %f3306, %f832; mul.ftz.f32 %f3308, %f3305, %f3305; fma.rn.ftz.f32 %f3309, %f3303, %f3303, %f3308; fma.rn.ftz.f32 %f3310, %f3307, %f3307, %f3309; sqrt.approx.ftz.f32 %f3311, %f3310; div.approx.ftz.f32 %f3958, %f3303, %f3311; div.approx.ftz.f32 %f3959, %f3305, %f3311; div.approx.ftz.f32 %f3960, %f3307, %f3311; $L__BB203_223: mul.ftz.f32 %f3680, %f44, %f2165; mul.ftz.f32 %f3679, %f44, %f2159; mul.ftz.f32 %f3678, %f44, %f2153; mul.ftz.f32 %f3677, %f44, %f2147; mul.ftz.f32 %f3315, %f3909, %f3966; fma.rn.ftz.f32 %f3316, %f3908, %f3965, %f3315; fma.rn.ftz.f32 %f3317, %f3910, %f3967, %f3316; add.ftz.f32 %f3318, %f3911, %f3317; mul.ftz.f32 %f3319, %f3913, %f3966; fma.rn.ftz.f32 %f3320, %f3912, %f3965, %f3319; fma.rn.ftz.f32 %f3321, %f3914, %f3967, %f3320; add.ftz.f32 %f3322, %f3915, %f3321; mul.ftz.f32 %f3323, %f3917, %f3966; fma.rn.ftz.f32 %f3324, %f3916, %f3965, %f3323; fma.rn.ftz.f32 %f3325, %f3918, %f3967, %f3324; add.ftz.f32 %f3326, %f3919, %f3325; mul.ftz.f32 %f3327, %f3909, %f3959; fma.rn.ftz.f32 %f3328, %f3908, %f3958, %f3327; fma.rn.ftz.f32 %f3329, %f3910, %f3960, %f3328; mul.ftz.f32 %f3330, %f3913, %f3959; fma.rn.ftz.f32 %f3331, %f3912, %f3958, %f3330; fma.rn.ftz.f32 %f3332, %f3914, %f3960, %f3331; mul.ftz.f32 %f3333, %f3917, %f3959; fma.rn.ftz.f32 %f3334, %f3916, %f3958, %f3333; fma.rn.ftz.f32 %f3335, %f3918, %f3960, %f3334; mul.ftz.f32 %f3336, %f3332, %f3332; fma.rn.ftz.f32 %f3337, %f3329, %f3329, %f3336; fma.rn.ftz.f32 %f3338, %f3335, %f3335, %f3337; sqrt.approx.ftz.f32 %f3339, %f3338; div.approx.ftz.f32 %f3340, %f3329, %f3339; div.approx.ftz.f32 %f3341, %f3332, %f3339; div.approx.ftz.f32 %f3342, %f3335, %f3339; sub.ftz.f32 %f846, %f3318, %f746; sub.ftz.f32 %f847, %f3322, %f747; sub.ftz.f32 %f848, %f3326, %f748; sub.ftz.f32 %f849, %f3340, %f749; sub.ftz.f32 %f850, %f3341, %f750; sub.ftz.f32 %f851, %f3342, %f751; add.ftz.f32 %f3343, %f50, 0f3F800000; fma.rn.ftz.f32 %f3344, %f3343, %f2148, %f3677; add.ftz.f32 %f852, %f3344, %f469; fma.rn.ftz.f32 %f3345, %f3343, %f2154, %f3678; add.ftz.f32 %f853, %f3345, %f473; fma.rn.ftz.f32 %f3346, %f3343, %f2160, %f3679; add.ftz.f32 %f854, %f3346, %f477; fma.rn.ftz.f32 %f3347, %f3343, %f2166, %f3680; add.ftz.f32 %f855, %f3347, %f480; setp.eq.ftz.f32 %p195, %f855, 0f00000000; mov.f32 %f3987, 0f00000000; mov.f32 %f3988, %f3987; mov.f32 %f3989, %f3987; @%p195 bra $L__BB203_225; div.approx.ftz.f32 %f3989, %f854, %f855; div.approx.ftz.f32 %f3988, %f853, %f855; div.approx.ftz.f32 %f3987, %f852, %f855; $L__BB203_225: ld.const.u32 %r473, [kernel_params+20]; setp.gt.s32 %p232, %r473, 2; add.ftz.f32 %f864, %f3987, 0fBF000000; add.ftz.f32 %f866, %f3988, 0fBF000000; @%p232 bra $L__BB203_232; bra.uni $L__BB203_226; $L__BB203_232: ld.const.u32 %r474, [kernel_params+20]; setp.eq.s32 %p197, %r474, 3; @%p197 bra $L__BB203_239; ld.const.u32 %r475, [kernel_params+20]; setp.eq.s32 %p198, %r475, 5; @%p198 bra $L__BB203_241; ld.const.u32 %r476, [kernel_params+20]; setp.ne.s32 %p199, %r476, 4; @%p199 bra $L__BB203_243; ld.const.f32 %f3351, [kernel_params+48]; ld.const.v4.f32 {%f3352, %f3353, %f3354, %f3355}, [kernel_params+64]; ld.const.f32 %f3360, [kernel_params+276]; mul.ftz.f32 %f867, %f864, %f3360; ld.const.f32 %f3361, [kernel_params+280]; mul.ftz.f32 %f868, %f866, %f3361; mul.ftz.f32 %f3362, %f868, %f868; fma.rn.ftz.f32 %f3363, %f867, %f867, %f3362; sqrt.approx.ftz.f32 %f869, %f3363; mul.ftz.f32 %f3364, %f869, %f869; mul.ftz.f32 %f3365, %f869, %f3364; mul.ftz.f32 %f3366, %f3364, %f3364; mul.ftz.f32 %f3367, %f3353, %f3364; fma.rn.ftz.f32 %f3368, %f3352, %f869, %f3367; mul.ftz.f32 %f3369, %f3355, %f3366; fma.rn.ftz.f32 %f3370, %f3354, %f3365, %f3369; add.ftz.f32 %f3371, %f3368, %f3370; add.ftz.f32 %f3372, %f3351, %f3371; neg.ftz.f32 %f870, %f3372; abs.ftz.f32 %f3373, %f870; ld.const.f32 %f3374, [kernel_params+24]; mul.ftz.f32 %f3375, %f3374, 0f3F000000; setp.gt.ftz.f32 %p202, %f3373, %f3375; mov.f32 %f3980, 0f00000000; mov.f32 %f3981, %f3980; mov.f32 %f3982, %f3980; @%p202 bra $L__BB203_247; setp.eq.ftz.f32 %p203, %f869, 0f00000000; mov.f32 %f3971, 0f00000000; @%p203 bra $L__BB203_238; div.approx.ftz.f32 %f3971, %f867, %f869; $L__BB203_238: mov.f32 %f3377, 0fBF800000; max.ftz.f32 %f3378, %f3971, %f3377; mov.f32 %f3379, 0f3F800000; min.ftz.f32 %f3380, %f3378, %f3379; abs.ftz.f32 %f3381, %f3380; sub.ftz.f32 %f3382, %f3379, %f3381; mul.ftz.f32 %f3383, %f3382, 0f3F000000; sqrt.approx.ftz.f32 %f3384, %f3383; setp.gt.ftz.f32 %p204, %f3381, 0f3F11EB85; selp.f32 %f3385, %f3384, %f3381, %p204; mul.ftz.f32 %f3386, %f3385, %f3385; mov.f32 %f3387, 0f3C94D2E9; mov.f32 %f3388, 0f3D53F941; fma.rn.ftz.f32 %f3389, %f3388, %f3386, %f3387; mov.f32 %f3390, 0f3D3F841F; fma.rn.ftz.f32 %f3391, %f3389, %f3386, %f3390; mov.f32 %f3392, 0f3D994929; fma.rn.ftz.f32 %f3393, %f3391, %f3386, %f3392; mov.f32 %f3394, 0f3E2AAB94; fma.rn.ftz.f32 %f3395, %f3393, %f3386, %f3394; mul.ftz.f32 %f3396, %f3386, %f3395; fma.rn.ftz.f32 %f3397, %f3396, %f3385, %f3385; add.ftz.f32 %f3398, %f3397, %f3397; mov.f32 %f3399, 0f3FC90FDB; sub.ftz.f32 %f3400, %f3399, %f3397; selp.f32 %f3401, %f3398, %f3400, %p204; setp.lt.ftz.f32 %p205, %f3380, 0f00000000; mov.f32 %f3402, 0f40490FDB; sub.ftz.f32 %f3403, %f3402, %f3401; selp.f32 %f3404, %f3403, %f3401, %p205; neg.ftz.f32 %f3405, %f3404; setp.lt.ftz.f32 %p206, %f868, 0f00000000; selp.f32 %f3406, %f3405, %f3404, %p206; cos.approx.ftz.f32 %f3407, %f3406; sin.approx.ftz.f32 %f3408, %f870; mul.ftz.f32 %f3409, %f3407, %f3408; sin.approx.ftz.f32 %f3410, %f3406; mul.ftz.f32 %f3982, %f3408, %f3410; neg.ftz.f32 %f3981, %f3409; cos.approx.ftz.f32 %f3980, %f870; bra.uni $L__BB203_247; $L__BB203_226: ld.const.u32 %r477, [kernel_params+20]; setp.eq.s32 %p200, %r477, 0; @%p200 bra $L__BB203_242; ld.const.u32 %r478, [kernel_params+20]; setp.eq.s32 %p201, %r478, 1; @%p201 bra $L__BB203_228; bra.uni $L__BB203_243; $L__BB203_228: add.ftz.f32 %f886, %f864, %f864; add.ftz.f32 %f887, %f866, %f866; mul.ftz.f32 %f3435, %f887, %f887; fma.rn.ftz.f32 %f3436, %f886, %f886, %f3435; sqrt.approx.ftz.f32 %f888, %f3436; setp.gt.ftz.f32 %p208, %f888, 0f3F800000; mov.f32 %f3980, 0f00000000; mov.f32 %f3981, %f3980; mov.f32 %f3982, %f3980; @%p208 bra $L__BB203_247; ld.const.f32 %f889, [kernel_params+24]; setp.eq.ftz.f32 %p209, %f888, 0f00000000; mov.f32 %f3972, 0f00000000; @%p209 bra $L__BB203_231; div.approx.ftz.f32 %f3972, %f886, %f888; $L__BB203_231: mov.f32 %f3438, 0fBF800000; max.ftz.f32 %f3439, %f3972, %f3438; mov.f32 %f3440, 0f3F800000; min.ftz.f32 %f3441, %f3439, %f3440; abs.ftz.f32 %f3442, %f3441; sub.ftz.f32 %f3443, %f3440, %f3442; mul.ftz.f32 %f3444, %f3443, 0f3F000000; sqrt.approx.ftz.f32 %f3445, %f3444; setp.gt.ftz.f32 %p210, %f3442, 0f3F11EB85; selp.f32 %f3446, %f3445, %f3442, %p210; mul.ftz.f32 %f3447, %f3446, %f3446; mov.f32 %f3448, 0f3C94D2E9; mov.f32 %f3449, 0f3D53F941; fma.rn.ftz.f32 %f3450, %f3449, %f3447, %f3448; mov.f32 %f3451, 0f3D3F841F; fma.rn.ftz.f32 %f3452, %f3450, %f3447, %f3451; mov.f32 %f3453, 0f3D994929; fma.rn.ftz.f32 %f3454, %f3452, %f3447, %f3453; mov.f32 %f3455, 0f3E2AAB94; fma.rn.ftz.f32 %f3456, %f3454, %f3447, %f3455; mul.ftz.f32 %f3457, %f3447, %f3456; fma.rn.ftz.f32 %f3458, %f3457, %f3446, %f3446; add.ftz.f32 %f3459, %f3458, %f3458; mov.f32 %f3460, 0f3FC90FDB; sub.ftz.f32 %f3461, %f3460, %f3458; selp.f32 %f3462, %f3459, %f3461, %p210; setp.lt.ftz.f32 %p211, %f3441, 0f00000000; mov.f32 %f3463, 0f40490FDB; sub.ftz.f32 %f3464, %f3463, %f3462; selp.f32 %f3465, %f3464, %f3462, %p211; mul.ftz.f32 %f3466, %f889, %f888; mul.ftz.f32 %f3467, %f3466, 0f3F000000; neg.ftz.f32 %f3468, %f3465; setp.lt.ftz.f32 %p212, %f887, 0f00000000; selp.f32 %f3469, %f3468, %f3465, %p212; cos.approx.ftz.f32 %f3470, %f3469; sin.approx.ftz.f32 %f3471, %f3467; mul.ftz.f32 %f3472, %f3470, %f3471; sin.approx.ftz.f32 %f3473, %f3469; mul.ftz.f32 %f3982, %f3471, %f3473; neg.ftz.f32 %f3981, %f3472; cos.approx.ftz.f32 %f3980, %f3467; bra.uni $L__BB203_247; $L__BB203_239: fma.rn.ftz.f32 %f876, %f3987, 0f40000000, 0fBF800000; fma.rn.ftz.f32 %f877, %f3988, 0f40000000, 0fBF800000; mul.ftz.f32 %f878, %f876, %f876; mul.ftz.f32 %f879, %f877, %f877; add.ftz.f32 %f3414, %f878, %f879; setp.gt.ftz.f32 %p207, %f3414, 0f3F800000; mov.f32 %f3980, 0f00000000; mov.f32 %f3981, %f3980; mov.f32 %f3982, %f3980; @%p207 bra $L__BB203_247; mov.f32 %f3415, 0f3F800000; sub.ftz.f32 %f3416, %f3415, %f878; sub.ftz.f32 %f3417, %f3416, %f879; mov.f32 %f3418, 0f00000000; max.ftz.f32 %f3419, %f3417, %f3418; sqrt.approx.ftz.f32 %f3420, %f3419; fma.rn.ftz.f32 %f3421, %f876, 0f00000000, %f3420; fma.rn.ftz.f32 %f3422, %f877, 0f00000000, %f3421; add.ftz.f32 %f3423, %f3422, %f3422; mul.ftz.f32 %f3424, %f3420, %f3423; sub.ftz.f32 %f3981, %f3415, %f3424; mul.ftz.f32 %f3982, %f877, %f3423; mul.ftz.f32 %f3980, %f876, %f3423; bra.uni $L__BB203_247; $L__BB203_241: mov.f32 %f3980, 0f3F800000; sub.ftz.f32 %f3425, %f3980, %f3987; fma.rn.ftz.f32 %f3426, %f3425, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f3427, %f3426; cos.approx.ftz.f32 %f3428, %f3426; fma.rn.ftz.f32 %f3429, %f3988, 0f3FC90FDB, 0fBF490FDB; sin.approx.ftz.f32 %f3430, %f3429; cos.approx.ftz.f32 %f3431, %f3429; div.approx.ftz.f32 %f3981, %f3427, %f3428; div.approx.ftz.f32 %f3982, %f3430, %f3431; bra.uni $L__BB203_247; $L__BB203_242: ld.const.v4.f32 {%f3474, %f3475, %f3476, %f3477}, [kernel_params+32]; fma.rn.ftz.f32 %f3482, %f3987, %f3474, %f3475; fma.rn.ftz.f32 %f3483, %f3988, %f3476, %f3477; sin.approx.ftz.f32 %f3484, %f3483; cos.approx.ftz.f32 %f3485, %f3482; sin.approx.ftz.f32 %f3486, %f3482; mul.ftz.f32 %f3981, %f3484, %f3486; mul.ftz.f32 %f3980, %f3484, %f3485; cos.approx.ftz.f32 %f3982, %f3483; bra.uni $L__BB203_247; $L__BB203_243: ld.const.v2.f32 {%f3490, %f3491}, [kernel_params+24]; ld.const.f32 %f3494, [kernel_params+276]; mul.ftz.f32 %f904, %f864, %f3494; ld.const.f32 %f3495, [kernel_params+280]; mul.ftz.f32 %f905, %f866, %f3495; add.ftz.f32 %f906, %f3491, %f3491; mul.ftz.f32 %f3496, %f3490, 0f3E800000; sin.approx.ftz.f32 %f3497, %f3496; mul.ftz.f32 %f3498, %f906, %f3497; mul.ftz.f32 %f3499, %f905, %f905; fma.rn.ftz.f32 %f3500, %f904, %f904, %f3499; sqrt.approx.ftz.f32 %f907, %f3500; setp.gt.ftz.f32 %p213, %f907, %f3498; mov.f32 %f3980, 0f00000000; mov.f32 %f3981, %f3980; mov.f32 %f3982, %f3980; @%p213 bra $L__BB203_247; setp.eq.ftz.f32 %p214, %f907, 0f00000000; mov.f32 %f3976, 0f00000000; @%p214 bra $L__BB203_246; div.approx.ftz.f32 %f3976, %f904, %f907; $L__BB203_246: mov.f32 %f3502, 0fBF800000; max.ftz.f32 %f3503, %f3976, %f3502; mov.f32 %f3504, 0f3F800000; min.ftz.f32 %f3505, %f3503, %f3504; abs.ftz.f32 %f3506, %f3505; sub.ftz.f32 %f3507, %f3504, %f3506; mul.ftz.f32 %f3508, %f3507, 0f3F000000; sqrt.approx.ftz.f32 %f3509, %f3508; setp.gt.ftz.f32 %p215, %f3506, 0f3F11EB85; selp.f32 %f3510, %f3509, %f3506, %p215; mul.ftz.f32 %f3511, %f3510, %f3510; mov.f32 %f3512, 0f3C94D2E9; mov.f32 %f3513, 0f3D53F941; fma.rn.ftz.f32 %f3514, %f3513, %f3511, %f3512; mov.f32 %f3515, 0f3D3F841F; fma.rn.ftz.f32 %f3516, %f3514, %f3511, %f3515; mov.f32 %f3517, 0f3D994929; fma.rn.ftz.f32 %f3518, %f3516, %f3511, %f3517; mov.f32 %f3519, 0f3E2AAB94; fma.rn.ftz.f32 %f3520, %f3518, %f3511, %f3519; mul.ftz.f32 %f3521, %f3511, %f3520; fma.rn.ftz.f32 %f3522, %f3521, %f3510, %f3510; add.ftz.f32 %f3523, %f3522, %f3522; mov.f32 %f3524, 0f3FC90FDB; sub.ftz.f32 %f3525, %f3524, %f3522; selp.f32 %f3526, %f3523, %f3525, %p215; setp.lt.ftz.f32 %p216, %f3505, 0f00000000; mov.f32 %f3527, 0f40490FDB; sub.ftz.f32 %f3528, %f3527, %f3526; selp.f32 %f3529, %f3528, %f3526, %p216; div.approx.ftz.f32 %f3530, %f907, %f906; abs.ftz.f32 %f3531, %f3530; sub.ftz.f32 %f3532, %f3504, %f3531; mul.ftz.f32 %f3533, %f3532, 0f3F000000; sqrt.approx.ftz.f32 %f3534, %f3533; setp.gt.ftz.f32 %p217, %f3531, 0f3F11EB85; selp.f32 %f3535, %f3534, %f3531, %p217; mul.ftz.f32 %f3536, %f3535, %f3535; fma.rn.ftz.f32 %f3537, %f3513, %f3536, %f3512; fma.rn.ftz.f32 %f3538, %f3537, %f3536, %f3515; fma.rn.ftz.f32 %f3539, %f3538, %f3536, %f3517; fma.rn.ftz.f32 %f3540, %f3539, %f3536, %f3519; mul.ftz.f32 %f3541, %f3536, %f3540; fma.rn.ftz.f32 %f3542, %f3541, %f3535, %f3535; mov.f32 %f3543, 0fC0000000; fma.rn.ftz.f32 %f3544, %f3543, %f3542, %f3524; selp.f32 %f3545, %f3544, %f3542, %p217; setp.le.ftz.f32 %p218, %f3545, 0f7F800000; mov.b32 %r446, %f3545; mov.b32 %r447, %f3530; and.b32 %r448, %r447, -2147483648; or.b32 %r449, %r448, %r446; mov.b32 %f3546, %r449; selp.f32 %f3547, %f3546, %f3545, %p218; add.ftz.f32 %f3548, %f3547, %f3547; neg.ftz.f32 %f3549, %f3529; setp.lt.ftz.f32 %p219, %f905, 0f00000000; selp.f32 %f3550, %f3549, %f3529, %p219; cos.approx.ftz.f32 %f3551, %f3550; sin.approx.ftz.f32 %f3552, %f3548; mul.ftz.f32 %f3553, %f3551, %f3552; sin.approx.ftz.f32 %f3554, %f3550; mul.ftz.f32 %f3982, %f3552, %f3554; neg.ftz.f32 %f3981, %f3553; cos.approx.ftz.f32 %f3980, %f3548; $L__BB203_247: ld.const.f32 %f3705, [kernel_params+80]; setp.eq.ftz.f32 %p236, %f3705, 0f00000000; @%p236 bra $L__BB203_254; ld.const.f32 %f3983, [kernel_params+80]; ld.const.f32 %f925, [kernel_params+92]; setp.leu.ftz.f32 %p221, %f925, 0f00000000; @%p221 bra $L__BB203_252; ld.const.f32 %f926, [kernel_params+88]; mov.f32 %f3556, 0fBF800000; max.ftz.f32 %f3557, %f3982, %f3556; mov.f32 %f3558, 0f3F800000; min.ftz.f32 %f3559, %f3557, %f3558; abs.ftz.f32 %f3560, %f3559; sub.ftz.f32 %f3561, %f3558, %f3560; mul.ftz.f32 %f3562, %f3561, 0f3F000000; sqrt.approx.ftz.f32 %f3563, %f3562; setp.gt.ftz.f32 %p222, %f3560, 0f3F11EB85; selp.f32 %f3564, %f3563, %f3560, %p222; mul.ftz.f32 %f3565, %f3564, %f3564; mov.f32 %f3566, 0f3C94D2E9; mov.f32 %f3567, 0f3D53F941; fma.rn.ftz.f32 %f3568, %f3567, %f3565, %f3566; mov.f32 %f3569, 0f3D3F841F; fma.rn.ftz.f32 %f3570, %f3568, %f3565, %f3569; mov.f32 %f3571, 0f3D994929; fma.rn.ftz.f32 %f3572, %f3570, %f3565, %f3571; mov.f32 %f3573, 0f3E2AAB94; fma.rn.ftz.f32 %f3574, %f3572, %f3565, %f3573; mul.ftz.f32 %f3575, %f3565, %f3574; fma.rn.ftz.f32 %f3576, %f3575, %f3564, %f3564; mov.f32 %f3577, 0f3FC90FDB; mov.f32 %f3578, 0fC0000000; fma.rn.ftz.f32 %f3579, %f3578, %f3576, %f3577; selp.f32 %f3580, %f3579, %f3576, %p222; setp.le.ftz.f32 %p223, %f3580, 0f7F800000; mov.b32 %r450, %f3580; mov.b32 %r451, %f3559; and.b32 %r452, %r451, -2147483648; or.b32 %r453, %r452, %r450; mov.b32 %f3581, %r453; selp.f32 %f3582, %f3581, %f3580, %p223; abs.ftz.f32 %f927, %f3582; setp.gt.ftz.f32 %p224, %f927, %f925; mov.f32 %f3983, 0f00000000; @%p224 bra $L__BB203_252; ld.const.f32 %f3983, [kernel_params+80]; setp.leu.ftz.f32 %p225, %f927, %f926; @%p225 bra $L__BB203_252; ld.const.f32 %f3687, [kernel_params+80]; sub.ftz.f32 %f3583, %f927, %f926; sub.ftz.f32 %f3584, %f925, %f926; div.approx.ftz.f32 %f3585, %f3583, %f3584; mul.ftz.f32 %f3586, %f3585, 0f3FC90FDB; cos.approx.ftz.f32 %f3587, %f3586; mul.ftz.f32 %f3983, %f3687, %f3587; $L__BB203_252: mul.ftz.f32 %f3588, %f3982, 0f00000000; sub.ftz.f32 %f3589, %f3981, %f3588; sub.ftz.f32 %f3590, %f3588, %f3980; mul.ftz.f32 %f3591, %f3981, 0f00000000; mul.ftz.f32 %f3592, %f3980, 0f00000000; sub.ftz.f32 %f3593, %f3592, %f3591; mul.ftz.f32 %f3594, %f3590, %f3590; fma.rn.ftz.f32 %f3595, %f3589, %f3589, %f3594; fma.rn.ftz.f32 %f3596, %f3593, %f3593, %f3595; sqrt.approx.ftz.f32 %f3597, %f3596; div.approx.ftz.f32 %f3598, %f3589, %f3597; div.approx.ftz.f32 %f3599, %f3590, %f3597; div.approx.ftz.f32 %f3600, %f3593, %f3597; mul.ftz.f32 %f930, %f3983, %f3598; mul.ftz.f32 %f931, %f3983, %f3599; mul.ftz.f32 %f932, %f3983, %f3600; add.ftz.f32 %f3987, %f3987, %f930; add.ftz.f32 %f3988, %f3988, %f931; add.ftz.f32 %f3989, %f3989, %f932; ld.const.f32 %f936, [kernel_params+84]; setp.eq.ftz.f32 %p226, %f936, 0f7F7FFFFF; @%p226 bra $L__BB203_254; mul.ftz.f32 %f3601, %f3980, %f936; sub.ftz.f32 %f3602, %f3601, %f930; mul.ftz.f32 %f3603, %f3981, %f936; sub.ftz.f32 %f3604, %f3603, %f931; mul.ftz.f32 %f3605, %f3982, %f936; sub.ftz.f32 %f3606, %f3605, %f932; mul.ftz.f32 %f3607, %f3604, %f3604; fma.rn.ftz.f32 %f3608, %f3602, %f3602, %f3607; fma.rn.ftz.f32 %f3609, %f3606, %f3606, %f3608; sqrt.approx.ftz.f32 %f3610, %f3609; div.approx.ftz.f32 %f3980, %f3602, %f3610; div.approx.ftz.f32 %f3981, %f3604, %f3610; div.approx.ftz.f32 %f3982, %f3606, %f3610; $L__BB203_254: mul.ftz.f32 %f3611, %f3909, %f3988; fma.rn.ftz.f32 %f3612, %f3908, %f3987, %f3611; fma.rn.ftz.f32 %f3613, %f3910, %f3989, %f3612; add.ftz.f32 %f3614, %f3911, %f3613; mul.ftz.f32 %f3615, %f3913, %f3988; fma.rn.ftz.f32 %f3616, %f3912, %f3987, %f3615; fma.rn.ftz.f32 %f3617, %f3914, %f3989, %f3616; add.ftz.f32 %f3618, %f3915, %f3617; mul.ftz.f32 %f3619, %f3917, %f3988; fma.rn.ftz.f32 %f3620, %f3916, %f3987, %f3619; fma.rn.ftz.f32 %f3621, %f3918, %f3989, %f3620; add.ftz.f32 %f3622, %f3919, %f3621; mul.ftz.f32 %f3623, %f3909, %f3981; fma.rn.ftz.f32 %f3624, %f3908, %f3980, %f3623; fma.rn.ftz.f32 %f3625, %f3910, %f3982, %f3624; mul.ftz.f32 %f3626, %f3913, %f3981; fma.rn.ftz.f32 %f3627, %f3912, %f3980, %f3626; fma.rn.ftz.f32 %f3628, %f3914, %f3982, %f3627; mul.ftz.f32 %f3629, %f3917, %f3981; fma.rn.ftz.f32 %f3630, %f3916, %f3980, %f3629; fma.rn.ftz.f32 %f3631, %f3918, %f3982, %f3630; mul.ftz.f32 %f3632, %f3628, %f3628; fma.rn.ftz.f32 %f3633, %f3625, %f3625, %f3632; fma.rn.ftz.f32 %f3634, %f3631, %f3631, %f3633; sqrt.approx.ftz.f32 %f3635, %f3634; div.approx.ftz.f32 %f3636, %f3625, %f3635; div.approx.ftz.f32 %f3637, %f3628, %f3635; div.approx.ftz.f32 %f3638, %f3631, %f3635; sub.ftz.f32 %f3639, %f3614, %f746; sub.ftz.f32 %f3640, %f3618, %f747; sub.ftz.f32 %f3641, %f3622, %f748; sub.ftz.f32 %f3642, %f3636, %f749; sub.ftz.f32 %f3643, %f3637, %f750; sub.ftz.f32 %f3644, %f3638, %f751; mul.ftz.f32 %f3645, %f850, %f850; fma.rn.ftz.f32 %f3646, %f849, %f849, %f3645; fma.rn.ftz.f32 %f3647, %f851, %f851, %f3646; sqrt.approx.ftz.f32 %f3648, %f3647; mul.ftz.f32 %f3649, %f3643, %f3643; fma.rn.ftz.f32 %f3650, %f3642, %f3642, %f3649; fma.rn.ftz.f32 %f3651, %f3644, %f3644, %f3650; sqrt.approx.ftz.f32 %f3652, %f3651; add.ftz.f32 %f3653, %f3648, %f3652; mul.ftz.f32 %f3990, %f3653, 0f3F000000; mul.ftz.f32 %f3654, %f847, %f847; fma.rn.ftz.f32 %f3655, %f846, %f846, %f3654; fma.rn.ftz.f32 %f3656, %f848, %f848, %f3655; sqrt.approx.ftz.f32 %f3657, %f3656; mul.ftz.f32 %f3658, %f3640, %f3640; fma.rn.ftz.f32 %f3659, %f3639, %f3639, %f3658; fma.rn.ftz.f32 %f3660, %f3641, %f3641, %f3659; sqrt.approx.ftz.f32 %f3661, %f3660; add.ftz.f32 %f3662, %f3657, %f3661; ld.const.f32 %f3663, [kernel_params+268]; add.ftz.f32 %f3664, %f3911, %f660; fma.rn.ftz.f32 %f3998, %f3995, %f3663, %f3664; add.ftz.f32 %f3665, %f3915, %f661; fma.rn.ftz.f32 %f3997, %f3994, %f3663, %f3665; add.ftz.f32 %f3666, %f3919, %f662; fma.rn.ftz.f32 %f3996, %f3993, %f3663, %f3666; mul.ftz.f32 %f3667, %f3990, %f3663; fma.rn.ftz.f32 %f3991, %f3662, 0f3F000000, %f3667; ld.const.f32 %f3992, [kernel_params+272]; $L__BB203_255: setp.eq.ftz.f32 %p227, %f3992, 0f00000000; @%p227 bra $L__BB203_263; mov.u16 %rs9, 0; ld.const.u64 %rd77, [kernel_params+2632]; cvta.to.global.u64 %rd78, %rd77; mul.lo.s64 %rd79, %rd3, 12; add.s64 %rd80, %rd78, %rd79; st.global.f32 [%rd80], %f3998; st.global.f32 [%rd80+4], %f3997; st.global.f32 [%rd80+8], %f3996; ld.const.v2.u64 {%rd81, %rd82}, [kernel_params+2640]; mov.u32 %r454, 0; cvta.to.global.u64 %rd85, %rd81; add.s64 %rd86, %rd85, %rd79; st.global.f32 [%rd86], %f3995; st.global.f32 [%rd86+4], %f3994; st.global.f32 [%rd86+8], %f3993; cvta.to.global.u64 %rd87, %rd82; shl.b64 %rd88, %rd3, 2; add.s64 %rd89, %rd87, %rd88; st.global.u32 [%rd89], %r454; ld.const.v2.u64 {%rd90, %rd91}, [kernel_params+2656]; cvta.to.global.u64 %rd94, %rd90; add.s64 %rd95, %rd94, %rd88; st.global.f32 [%rd95], %f3992; cvta.to.global.u64 %rd96, %rd91; add.s64 %rd97, %rd96, %rd88; st.global.f32 [%rd97], %f3776; ld.const.v2.u64 {%rd98, %rd99}, [kernel_params+2672]; cvta.to.global.u64 %rd102, %rd98; add.s64 %rd103, %rd102, %rd88; st.global.f32 [%rd103], %f3991; cvta.to.global.u64 %rd104, %rd99; add.s64 %rd105, %rd104, %rd88; st.global.f32 [%rd105], %f3990; ld.const.u64 %rd106, [kernel_params+2456]; cvta.to.global.u64 %rd107, %rd106; add.s64 %rd108, %rd107, %rd88; st.global.u32 [%rd108], %r492; ld.const.v2.u64 {%rd109, %rd110}, [kernel_params+2464]; cvta.to.global.u64 %rd113, %rd109; shl.b64 %rd114, %rd3, 1; add.s64 %rd115, %rd113, %rd114; st.global.u16 [%rd115], %rs9; ld.const.v2.u64 {%rd116, %rd117}, [kernel_params+2480]; cvta.to.global.u64 %rd120, %rd116; add.s64 %rd121, %rd120, %rd114; st.global.u16 [%rd121], %rs9; cvta.to.global.u64 %rd122, %rd117; add.s64 %rd123, %rd122, %rd114; st.global.u16 [%rd123], %rs9; ld.const.v2.u64 {%rd124, %rd125}, [kernel_params+2496]; cvta.to.global.u64 %rd128, %rd124; add.s64 %rd129, %rd128, %rd114; st.global.u16 [%rd129], %rs9; cvta.to.global.u64 %rd130, %rd110; add.s64 %rd131, %rd130, %rd114; st.global.u16 [%rd131], %rs9; cvta.to.global.u64 %rd132, %rd125; add.s64 %rd133, %rd132, %rd114; st.global.u16 [%rd133], %rs9; ld.const.u64 %rd134, [kernel_params+2512]; cvta.to.global.u64 %rd135, %rd134; add.s64 %rd136, %rd135, %rd114; st.global.u16 [%rd136], %rs9; ld.const.v2.u64 {%rd137, %rd138}, [kernel_params+2528]; cvta.to.global.u64 %rd141, %rd137; add.s64 %rd142, %rd141, %rd88; st.global.u32 [%rd142], %r21; cvta.to.global.u64 %rd143, %rd138; add.s64 %rd144, %rd143, %rd114; mov.u16 %rs6, 16; st.global.u16 [%rd144], %rs6; ld.const.v2.u64 {%rd145, %rd146}, [kernel_params+2544]; cvta.to.global.u64 %rd149, %rd145; add.s64 %rd11, %rd149, %rd88; mov.u32 %r455, 18433; st.global.u32 [%rd11], %r455; ld.const.v2.u64 {%rd150, %rd151}, [kernel_params+2560]; cvta.to.global.u64 %rd154, %rd150; add.s64 %rd155, %rd154, %rd88; st.global.u32 [%rd155], %r454; cvta.to.global.u64 %rd156, %rd151; add.s64 %rd157, %rd156, %rd88; mov.u32 %r456, 2139095039; st.global.u32 [%rd157], %r456; ld.const.v2.u64 {%rd158, %rd159}, [kernel_params+2576]; cvta.to.global.u64 %rd162, %rd158; add.s64 %rd163, %rd162, %rd88; mov.u32 %r457, 1065353216; st.global.u32 [%rd163], %r457; cvta.to.global.u64 %rd164, %rd159; add.s64 %rd165, %rd164, %rd79; st.global.u32 [%rd165], %r457; st.global.u32 [%rd165+4], %r457; st.global.u32 [%rd165+8], %r457; cvta.to.global.u64 %rd166, %rd146; add.s64 %rd167, %rd166, %rd3; st.global.u8 [%rd167], %rs9; ld.const.u64 %rd168, [kernel_params+2720]; cvta.to.global.u64 %rd169, %rd168; add.s64 %rd170, %rd169, %rd88; mov.u32 %r458, -1; st.global.u32 [%rd170], %r458; ld.const.u64 %rd171, [kernel_params+2712]; cvta.to.global.u64 %rd172, %rd171; add.s64 %rd173, %rd172, %rd88; st.global.u32 [%rd173], %r458; ld.const.u32 %r50, [kernel_params]; and.b32 %r459, %r50, 131072; setp.eq.s32 %p228, %r459, 0; @%p228 bra $L__BB203_258; ld.const.v2.u64 {%rd174, %rd175}, [kernel_params+2768]; cvta.to.global.u64 %rd178, %rd174; add.s64 %rd180, %rd178, %rd88; st.global.u32 [%rd180], %r458; ld.const.u32 %r461, [kernel_params+1028]; cvta.to.global.u64 %rd181, %rd175; add.s64 %rd182, %rd181, %rd88; st.global.u32 [%rd182], %r461; ld.const.v2.u64 {%rd183, %rd184}, [kernel_params+2784]; cvta.to.global.u64 %rd187, %rd183; add.s64 %rd188, %rd187, %rd88; st.global.u32 [%rd188], %r458; cvta.to.global.u64 %rd189, %rd184; add.s64 %rd190, %rd189, %rd88; st.global.u32 [%rd190], %r458; $L__BB203_258: and.b32 %r462, %r50, 512; setp.eq.s32 %p229, %r462, 0; @%p229 bra $L__BB203_260; ld.global.u32 %r463, [%rd11]; or.b32 %r464, %r463, 16777216; st.global.u32 [%rd11], %r464; ld.const.u64 %rd191, [kernel_params+2616]; cvta.to.global.u64 %rd192, %rd191; add.s64 %rd194, %rd192, %rd79; st.global.u32 [%rd194], %r457; st.global.u32 [%rd194+4], %r457; st.global.u32 [%rd194+8], %r457; $L__BB203_260: ld.const.u32 %r466, [kernel_params+300]; setp.eq.s32 %p230, %r466, 0; ld.const.u64 %rd195, [kernel_params+4256]; cvta.to.global.u64 %rd12, %rd195; @%p230 bra $L__BB203_262; add.s64 %rd196, %rd12, 20; atom.global.add.u32 %r467, [%rd196], 1; mov.u16 %rs7, 5; st.global.u16 [%rd5], %rs7; bra.uni $L__BB203_263; $L__BB203_262: add.s64 %rd197, %rd12, 8; atom.global.add.u32 %r468, [%rd197], 1; mov.u16 %rs8, 2; st.global.u16 [%rd5], %rs8; $L__BB203_263: ret; } // .globl kernel_gpu_integrator_init_from_bake .visible .entry kernel_gpu_integrator_init_from_bake( .param .u64 kernel_gpu_integrator_init_from_bake_param_0, .param .u32 kernel_gpu_integrator_init_from_bake_param_1, .param .u64 kernel_gpu_integrator_init_from_bake_param_2, .param .u32 kernel_gpu_integrator_init_from_bake_param_3 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<54>; .reg .b16 %rs<14>; .reg .f32 %f<604>; .reg .b32 %r<275>; .reg .b64 %rd<339>; ld.param.u64 %rd19, [kernel_gpu_integrator_init_from_bake_param_0]; ld.param.u32 %r43, [kernel_gpu_integrator_init_from_bake_param_1]; ld.param.u64 %rd20, [kernel_gpu_integrator_init_from_bake_param_2]; ld.param.u32 %r42, [kernel_gpu_integrator_init_from_bake_param_3]; cvta.to.global.u64 %rd1, %rd20; mov.u32 %r44, %ntid.x; mov.u32 %r45, %ctaid.x; mov.u32 %r46, %tid.x; mad.lo.s32 %r1, %r45, %r44, %r46; mul.lo.s32 %r47, %r42, %r43; setp.ge.s32 %p1, %r1, %r47; @%p1 bra $L__BB204_69; div.s32 %r48, %r1, %r42; mul.lo.s32 %r49, %r48, %r42; sub.s32 %r2, %r1, %r49; cvta.to.global.u64 %rd21, %rd19; mul.wide.s32 %rd22, %r48, 44; add.s64 %rd23, %rd21, %rd22; add.s64 %rd2, %rd23, 40; ld.global.u32 %r50, [%rd23+40]; setp.ge.s32 %p2, %r2, %r50; @%p2 bra $L__BB204_69; ld.global.u32 %r51, [%rd2+-4]; add.s32 %r3, %r51, %r2; ld.const.f32 %f1, [kernel_params+1604]; setp.lt.ftz.f32 %p3, %f1, 0f3F666666; @%p3 bra $L__BB204_4; bra.uni $L__BB204_3; $L__BB204_4: ld.global.u32 %r54, [%rd2+-28]; ld.global.u32 %r267, [%rd2+-32]; mul.lo.s32 %r55, %r54, %r267; div.u32 %r268, %r2, %r55; mul.lo.s32 %r56, %r268, %r55; sub.s32 %r269, %r2, %r56; bra.uni $L__BB204_5; $L__BB204_3: ld.global.u32 %r52, [%rd2+-20]; div.u32 %r269, %r2, %r52; mul.lo.s32 %r53, %r269, %r52; sub.s32 %r268, %r2, %r53; ld.global.u32 %r267, [%rd2+-32]; $L__BB204_5: div.u32 %r57, %r269, %r267; mul.lo.s32 %r58, %r57, %r267; ld.global.u32 %r59, [%rd2+-36]; add.s32 %r60, %r59, %r57; ld.global.u32 %r61, [%rd2+-24]; add.s32 %r270, %r61, %r268; ld.global.u32 %r62, [%rd2+-8]; ld.global.u32 %r63, [%rd2+-40]; add.s32 %r64, %r63, %r269; sub.s32 %r65, %r64, %r58; ld.global.u32 %r66, [%rd2+-12]; add.s32 %r67, %r65, %r66; mad.lo.s32 %r68, %r62, %r60, %r67; ld.const.u64 %rd24, [kernel_params+2448]; cvta.to.global.u64 %rd25, %rd24; cvt.s64.s32 %rd3, %r3; mul.wide.s32 %rd26, %r3, 4; add.s64 %rd4, %rd25, %rd26; st.global.u32 [%rd4], %r68; ld.const.u64 %rd27, [kernel_params+2520]; cvta.to.global.u64 %rd28, %rd27; mul.wide.s32 %rd29, %r3, 2; add.s64 %rd5, %rd28, %rd29; mov.u16 %rs4, 0; st.global.u16 [%rd5], %rs4; ld.const.u32 %r14, [kernel_params+1400]; setp.eq.s32 %p4, %r14, -1; mov.u16 %rs13, 1; @%p4 bra $L__BB204_7; ld.global.u32 %rd30, [%rd4]; ld.const.s32 %rd31, [kernel_params+1264]; mul.lo.s64 %rd32, %rd31, %rd30; add.s32 %r69, %r14, 3; cvt.u64.u32 %rd33, %r69; add.s64 %rd34, %rd32, %rd33; shl.b64 %rd35, %rd34, 2; add.s64 %rd36, %rd1, %rd35; ld.global.f32 %f161, [%rd36]; setp.eq.ftz.f32 %p5, %f161, 0f00000000; selp.u16 %rs13, 1, 0, %p5; $L__BB204_7: setp.eq.s16 %p6, %rs13, 0; @%p6 bra $L__BB204_69; ld.const.u32 %r15, [kernel_params+1404]; setp.eq.s32 %p7, %r15, -1; @%p7 bra $L__BB204_10; ld.global.u32 %r70, [%rd2+-16]; ld.global.u32 %rd37, [%rd4]; ld.const.s32 %rd338, [kernel_params+1264]; mul.lo.s64 %rd38, %rd338, %rd37; cvt.s64.s32 %rd39, %r15; add.s64 %rd40, %rd38, %rd39; shl.b64 %rd41, %rd40, 2; add.s64 %rd42, %rd1, %rd41; atom.global.add.u32 %r71, [%rd42], 1; add.s32 %r270, %r71, %r70; bra.uni $L__BB204_11; $L__BB204_10: ld.const.s32 %rd338, [kernel_params+1264]; $L__BB204_11: ld.global.s32 %rd43, [%rd4]; mul.lo.s64 %rd9, %rd338, %rd43; ld.const.v2.u32 {%r72, %r73}, [kernel_params+1448]; cvt.s64.s32 %rd44, %r72; add.s64 %rd45, %rd9, %rd44; shl.b64 %rd46, %rd45, 2; add.s64 %rd10, %rd1, %rd46; ld.global.u32 %r20, [%rd10+4]; setp.eq.s32 %p8, %r20, -1; @%p8 bra $L__BB204_65; cvt.s64.s32 %rd47, %r73; add.s64 %rd48, %rd9, %rd47; shl.b64 %rd49, %rd48, 2; add.s64 %rd11, %rd1, %rd49; ld.global.u32 %r74, [%rd10]; ld.const.u32 %r75, [kernel_params+984]; add.s32 %r21, %r75, %r20; add.s32 %r76, %r74, -559038720; xor.b32 %r77, %r76, -1874868139; add.s32 %r78, %r77, 29184895; xor.b32 %r79, %r78, -559038720; shf.l.wrap.b32 %r80, %r78, %r78, 25; sub.s32 %r81, %r79, %r80; xor.b32 %r82, %r81, -1874868139; shf.l.wrap.b32 %r83, %r81, %r81, 16; sub.s32 %r84, %r82, %r83; xor.b32 %r85, %r84, %r78; shf.l.wrap.b32 %r86, %r84, %r84, 4; sub.s32 %r87, %r85, %r86; xor.b32 %r88, %r87, %r81; shf.l.wrap.b32 %r89, %r87, %r87, 14; sub.s32 %r90, %r88, %r89; xor.b32 %r91, %r90, %r84; shf.l.wrap.b32 %r92, %r90, %r90, 24; sub.s32 %r93, %r91, %r92; ld.const.u32 %r22, [kernel_params+1576]; xor.b32 %r23, %r93, %r22; setp.eq.s32 %p9, %r270, 0; mov.f32 %f558, 0f3F000000; mov.f32 %f559, %f558; @%p9 bra $L__BB204_20; ld.const.u32 %r94, [kernel_params+1596]; setp.eq.s32 %p10, %r94, 0; brev.b32 %r95, %r270; mul.lo.s32 %r96, %r95, 1025551850; xor.b32 %r24, %r96, %r95; @%p10 bra $L__BB204_16; setp.geu.ftz.f32 %p11, %f1, 0f3F800000; setp.lt.ftz.f32 %p12, %f1, 0f3F800000; selp.b32 %r97, %r22, %r23, %p12; mul.lo.s32 %r98, %r97, -512718531; shr.u32 %r99, %r97, 16; xor.b32 %r100, %r98, %r99; bfe.u32 %r101, %r100, 4, 4; shr.u32 %r102, %r97, 8; xor.b32 %r103, %r100, %r102; xor.b32 %r104, %r103, %r101; mul.lo.s32 %r105, %r104, 153742143; shr.u32 %r106, %r97, 23; xor.b32 %r107, %r105, %r106; bfe.u32 %r108, %r107, 1, 7; xor.b32 %r109, %r108, %r107; shr.u32 %r110, %r97, 27; or.b32 %r111, %r110, 1; mul.lo.s32 %r112, %r111, %r109; mul.lo.s32 %r113, %r112, -1074439621; bfe.u32 %r114, %r113, 2, 6; xor.b32 %r115, %r114, %r113; mul.lo.s32 %r116, %r115, -1638916925; bfe.u32 %r117, %r116, 2, 6; xor.b32 %r118, %r117, %r116; mul.lo.s32 %r119, %r118, -933190689; and.b32 %r120, %r119, 255; bfe.u32 %r121, %r119, 5, 3; xor.b32 %r122, %r121, %r120; ld.const.u32 %r123, [kernel_params+1600]; add.s32 %r124, %r123, -1; xor.b32 %r125, %r97, 61; mul.lo.s32 %r126, %r125, 9; shr.u32 %r127, %r126, 4; xor.b32 %r128, %r127, %r126; mul.lo.s32 %r129, %r128, 668265261; add.s32 %r130, %r129, %r24; shr.u32 %r131, %r129, 16; or.b32 %r132, %r131, 1; mul.lo.s32 %r133, %r132, %r130; mul.lo.s32 %r134, %r133, 89287766; xor.b32 %r135, %r134, %r133; mul.lo.s32 %r136, %r135, 1403136100; xor.b32 %r137, %r136, %r135; brev.b32 %r138, %r137; neg.s32 %r139, %r123; and.b32 %r140, %r270, %r139; and.b32 %r141, %r138, %r124; or.b32 %r142, %r141, %r140; mad.lo.s32 %r143, %r122, %r123, %r142; shl.b32 %r144, %r123, 8; rem.u32 %r145, %r143, %r144; ld.const.u64 %rd50, [kernel_params+2424]; cvta.to.global.u64 %rd51, %rd50; shl.b32 %r146, %r145, 1; mul.wide.u32 %rd52, %r146, 4; add.s64 %rd53, %rd51, %rd52; ld.global.f32 %f558, [%rd53]; or.b32 %r147, %r146, 1; mul.wide.u32 %rd54, %r147, 4; add.s64 %rd55, %rd51, %rd54; ld.global.f32 %f559, [%rd55]; @%p11 bra $L__BB204_20; xor.b32 %r148, %r23, 61; mul.lo.s32 %r149, %r148, 9; shr.u32 %r150, %r149, 4; xor.b32 %r151, %r150, %r149; mul.lo.s32 %r152, %r151, 668265261; cvt.rn.f32.u32 %f164, %r152; mul.ftz.f32 %f165, %f164, 0f2F7FFFFE; xor.b32 %r153, %r23, -905047700; mul.lo.s32 %r154, %r153, 9; shr.u32 %r155, %r154, 4; xor.b32 %r156, %r155, %r154; mul.lo.s32 %r157, %r156, 668265261; cvt.rn.f32.u32 %f166, %r157; mul.ftz.f32 %f167, %f166, 0f2F7FFFFE; fma.rn.ftz.f32 %f168, %f165, %f1, %f558; fma.rn.ftz.f32 %f169, %f167, %f1, %f559; cvt.rmi.ftz.f32.f32 %f170, %f168; sub.ftz.f32 %f558, %f168, %f170; cvt.rmi.ftz.f32.f32 %f171, %f169; sub.ftz.f32 %f559, %f169, %f171; bra.uni $L__BB204_20; $L__BB204_65: ld.const.u8 %rs11, [kernel_params+1260]; and.b16 %rs12, %rs11, 2; setp.eq.s16 %p52, %rs12, 0; @%p52 bra $L__BB204_67; ld.const.s32 %rd324, [kernel_params+1268]; add.s64 %rd325, %rd324, %rd9; shl.b64 %rd326, %rd325, 2; add.s64 %rd327, %rd1, %rd326; add.s64 %rd328, %rd327, 12; atom.global.add.f32 %f554, [%rd328], 0f3F800000; $L__BB204_67: ld.const.u32 %r261, [kernel_params+1620]; setp.eq.s32 %p53, %r261, 0; @%p53 bra $L__BB204_69; ld.const.s32 %rd329, [kernel_params+1384]; add.s64 %rd330, %rd329, %rd9; shl.b64 %rd331, %rd330, 2; add.s64 %rd332, %rd1, %rd331; add.s64 %rd333, %rd332, 12; atom.global.add.f32 %f555, [%rd333], 0f3F800000; bra.uni $L__BB204_69; $L__BB204_16: xor.b32 %r159, %r23, 508809841; add.s32 %r160, %r159, %r24; shr.u32 %r161, %r159, 16; or.b32 %r162, %r161, 1; mul.lo.s32 %r163, %r162, %r160; mul.lo.s32 %r164, %r163, 89287766; xor.b32 %r165, %r164, %r163; mul.lo.s32 %r166, %r165, 1403136100; xor.b32 %r25, %r166, %r165; setp.eq.s32 %p13, %r25, 0; mov.u32 %r274, 0; @%p13 bra $L__BB204_19; mov.u32 %r271, 0; mov.u64 %rd57, sobol_burley_table; mov.u32 %r272, %r25; mov.u32 %r274, %r271; $L__BB204_18: clz.b32 %r169, %r272; add.s32 %r170, %r169, %r271; mul.wide.u32 %rd56, %r170, 4; add.s64 %rd58, %rd57, %rd56; ld.const.u32 %r171, [%rd58+128]; xor.b32 %r274, %r171, %r274; add.s32 %r271, %r170, 1; shl.b32 %r172, %r272, %r169; shl.b32 %r272, %r172, 1; setp.ne.s32 %p14, %r272, 0; @%p14 bra $L__BB204_18; $L__BB204_19: brev.b32 %r173, %r25; xor.b32 %r174, %r23, 106206365; shr.u32 %r175, %r174, 16; or.b32 %r176, %r175, 1; mul.lo.s32 %r177, %r173, 1025551850; xor.b32 %r178, %r177, %r173; add.s32 %r179, %r178, %r174; mul.lo.s32 %r180, %r179, %r176; mul.lo.s32 %r181, %r180, 89287766; xor.b32 %r182, %r181, %r180; xor.b32 %r183, %r23, 1919448741; mul.lo.s32 %r184, %r274, 1025551850; xor.b32 %r185, %r184, %r274; add.s32 %r186, %r185, %r183; shr.u32 %r187, %r183, 16; or.b32 %r188, %r187, 1; mul.lo.s32 %r189, %r186, %r188; mul.lo.s32 %r190, %r189, 89287766; xor.b32 %r191, %r190, %r189; mul.lo.s32 %r192, %r191, 1403136100; xor.b32 %r193, %r192, %r191; brev.b32 %r194, %r193; cvt.rn.f32.u32 %f172, %r194; mul.lo.s32 %r195, %r182, 1403136100; xor.b32 %r196, %r195, %r182; brev.b32 %r197, %r196; cvt.rn.f32.u32 %f173, %r197; mul.ftz.f32 %f559, %f172, 0f2F7FFFFE; mul.ftz.f32 %f558, %f173, 0f2F7FFFFE; $L__BB204_20: ld.const.u64 %rd59, [kernel_params+2456]; cvta.to.global.u64 %rd60, %rd59; shl.b64 %rd61, %rd3, 2; add.s64 %rd62, %rd60, %rd61; st.global.u32 [%rd62], %r270; ld.const.v2.u64 {%rd63, %rd64}, [kernel_params+2464]; mov.u32 %r198, 0; cvta.to.global.u64 %rd67, %rd63; shl.b64 %rd68, %rd3, 1; add.s64 %rd69, %rd67, %rd68; st.global.u16 [%rd69], %rs4; ld.const.v2.u64 {%rd70, %rd71}, [kernel_params+2480]; cvta.to.global.u64 %rd74, %rd70; add.s64 %rd75, %rd74, %rd68; st.global.u16 [%rd75], %rs4; cvta.to.global.u64 %rd76, %rd71; add.s64 %rd77, %rd76, %rd68; st.global.u16 [%rd77], %rs4; ld.const.v2.u64 {%rd78, %rd79}, [kernel_params+2496]; cvta.to.global.u64 %rd82, %rd78; add.s64 %rd83, %rd82, %rd68; st.global.u16 [%rd83], %rs4; cvta.to.global.u64 %rd84, %rd64; add.s64 %rd85, %rd84, %rd68; st.global.u16 [%rd85], %rs4; cvta.to.global.u64 %rd86, %rd79; add.s64 %rd87, %rd86, %rd68; st.global.u16 [%rd87], %rs4; ld.const.u64 %rd88, [kernel_params+2512]; cvta.to.global.u64 %rd89, %rd88; add.s64 %rd90, %rd89, %rd68; st.global.u16 [%rd90], %rs4; ld.const.v2.u64 {%rd91, %rd92}, [kernel_params+2528]; cvta.to.global.u64 %rd95, %rd91; add.s64 %rd96, %rd95, %rd61; st.global.u32 [%rd96], %r23; cvta.to.global.u64 %rd97, %rd92; add.s64 %rd98, %rd97, %rd68; mov.u16 %rs6, 16; st.global.u16 [%rd98], %rs6; ld.const.v2.u64 {%rd99, %rd100}, [kernel_params+2544]; cvta.to.global.u64 %rd103, %rd99; add.s64 %rd12, %rd103, %rd61; mov.u32 %r199, 18433; st.global.u32 [%rd12], %r199; ld.const.v2.u64 {%rd104, %rd105}, [kernel_params+2560]; cvta.to.global.u64 %rd108, %rd104; add.s64 %rd109, %rd108, %rd61; st.global.u32 [%rd109], %r198; cvta.to.global.u64 %rd110, %rd105; add.s64 %rd111, %rd110, %rd61; mov.u32 %r200, 2139095039; st.global.u32 [%rd111], %r200; ld.const.v2.u64 {%rd112, %rd113}, [kernel_params+2576]; cvta.to.global.u64 %rd116, %rd112; add.s64 %rd117, %rd116, %rd61; mov.u32 %r201, 1065353216; st.global.u32 [%rd117], %r201; cvta.to.global.u64 %rd118, %rd113; mul.lo.s64 %rd119, %rd3, 12; add.s64 %rd120, %rd118, %rd119; st.global.u32 [%rd120], %r201; st.global.u32 [%rd120+4], %r201; st.global.u32 [%rd120+8], %r201; cvta.to.global.u64 %rd121, %rd100; add.s64 %rd122, %rd121, %rd3; st.global.u8 [%rd122], %rs4; ld.const.u64 %rd123, [kernel_params+2720]; cvta.to.global.u64 %rd124, %rd123; add.s64 %rd13, %rd124, %rd61; mov.u32 %r202, -1; st.global.u32 [%rd13], %r202; ld.const.u64 %rd125, [kernel_params+2712]; cvta.to.global.u64 %rd126, %rd125; add.s64 %rd14, %rd126, %rd61; st.global.u32 [%rd14], %r202; ld.const.u32 %r33, [kernel_params]; and.b32 %r203, %r33, 131072; setp.eq.s32 %p15, %r203, 0; @%p15 bra $L__BB204_22; ld.const.v2.u64 {%rd127, %rd128}, [kernel_params+2768]; cvta.to.global.u64 %rd131, %rd127; add.s64 %rd133, %rd131, %rd61; st.global.u32 [%rd133], %r202; ld.const.u32 %r205, [kernel_params+1028]; cvta.to.global.u64 %rd134, %rd128; add.s64 %rd135, %rd134, %rd61; st.global.u32 [%rd135], %r205; ld.const.v2.u64 {%rd136, %rd137}, [kernel_params+2784]; cvta.to.global.u64 %rd140, %rd136; add.s64 %rd141, %rd140, %rd61; st.global.u32 [%rd141], %r202; cvta.to.global.u64 %rd142, %rd137; add.s64 %rd143, %rd142, %rd61; st.global.u32 [%rd143], %r202; $L__BB204_22: and.b32 %r206, %r33, 512; setp.eq.s32 %p16, %r206, 0; @%p16 bra $L__BB204_24; ld.global.u32 %r207, [%rd12]; or.b32 %r208, %r207, 16777216; st.global.u32 [%rd12], %r208; ld.const.u64 %rd144, [kernel_params+2616]; cvta.to.global.u64 %rd145, %rd144; add.s64 %rd147, %rd145, %rd119; st.global.u32 [%rd147], %r201; st.global.u32 [%rd147+4], %r201; st.global.u32 [%rd147+8], %r201; $L__BB204_24: ld.global.f32 %f565, [%rd10+12]; ld.global.f32 %f19, [%rd11]; ld.global.f32 %f20, [%rd11+4]; ld.global.f32 %f21, [%rd11+8]; ld.global.f32 %f22, [%rd11+12]; ld.global.f32 %f564, [%rd10+8]; setp.neu.ftz.f32 %p17, %f564, 0f00000000; setp.neu.ftz.f32 %p18, %f564, 0f3F800000; and.pred %p19, %p17, %p18; @%p19 bra $L__BB204_28; setp.neu.ftz.f32 %p20, %f565, 0f00000000; setp.neu.ftz.f32 %p21, %f565, 0f3F800000; and.pred %p22, %p20, %p21; @%p22 bra $L__BB204_28; ld.const.u64 %rd148, [kernel_params+2232]; cvta.to.global.u64 %rd149, %rd148; mul.wide.s32 %rd150, %r21, 16; add.s64 %rd151, %rd149, %rd150; ld.const.u64 %rd152, [kernel_params+2256]; cvta.to.global.u64 %rd153, %rd152; ld.global.u32 %r210, [%rd151+12]; mul.wide.u32 %rd154, %r210, 12; add.s64 %rd155, %rd153, %rd154; add.s32 %r211, %r210, 1; mul.wide.u32 %rd156, %r211, 12; add.s64 %rd157, %rd153, %rd156; add.s32 %r212, %r210, 2; mul.wide.u32 %rd158, %r212, 12; add.s64 %rd159, %rd153, %rd158; ld.global.f32 %f176, [%rd155]; ld.global.f32 %f177, [%rd155+4]; ld.global.f32 %f178, [%rd155+8]; ld.global.f32 %f179, [%rd157]; mul.ftz.f32 %f180, %f565, %f179; ld.global.f32 %f181, [%rd157+4]; mul.ftz.f32 %f182, %f565, %f181; ld.global.f32 %f183, [%rd157+8]; mul.ftz.f32 %f184, %f565, %f183; fma.rn.ftz.f32 %f185, %f564, %f176, %f180; fma.rn.ftz.f32 %f186, %f564, %f177, %f182; fma.rn.ftz.f32 %f187, %f564, %f178, %f184; mov.f32 %f188, 0f3F800000; sub.ftz.f32 %f189, %f188, %f564; sub.ftz.f32 %f190, %f189, %f565; ld.global.f32 %f191, [%rd159]; ld.global.f32 %f192, [%rd159+4]; ld.global.f32 %f193, [%rd159+8]; fma.rn.ftz.f32 %f194, %f190, %f191, %f185; fma.rn.ftz.f32 %f195, %f190, %f192, %f186; fma.rn.ftz.f32 %f196, %f190, %f193, %f187; add.ftz.f32 %f197, %f176, %f179; add.ftz.f32 %f198, %f177, %f181; add.ftz.f32 %f199, %f178, %f183; add.ftz.f32 %f200, %f197, %f191; add.ftz.f32 %f201, %f198, %f192; add.ftz.f32 %f202, %f199, %f193; mov.f32 %f203, 0f40400000; div.approx.ftz.f32 %f204, %f200, %f203; div.approx.ftz.f32 %f205, %f201, %f203; div.approx.ftz.f32 %f206, %f202, %f203; sub.ftz.f32 %f207, %f204, %f194; sub.ftz.f32 %f208, %f205, %f195; sub.ftz.f32 %f209, %f206, %f196; mul.ftz.f32 %f210, %f208, %f208; fma.rn.ftz.f32 %f211, %f207, %f207, %f210; fma.rn.ftz.f32 %f212, %f209, %f209, %f211; sqrt.approx.ftz.f32 %f213, %f212; div.approx.ftz.f32 %f214, %f207, %f213; div.approx.ftz.f32 %f215, %f208, %f213; div.approx.ftz.f32 %f216, %f209, %f213; abs.ftz.f32 %f217, %f194; abs.ftz.f32 %f218, %f195; abs.ftz.f32 %f219, %f196; max.ftz.f32 %f220, %f217, %f218; max.ftz.f32 %f221, %f220, %f219; max.ftz.f32 %f222, %f221, %f188; mul.ftz.f32 %f223, %f222, 0f38D1B717; min.ftz.f32 %f224, %f213, %f223; fma.rn.ftz.f32 %f225, %f214, %f224, %f194; fma.rn.ftz.f32 %f226, %f215, %f224, %f195; fma.rn.ftz.f32 %f227, %f216, %f224, %f196; sub.ftz.f32 %f228, %f179, %f176; sub.ftz.f32 %f229, %f181, %f177; sub.ftz.f32 %f230, %f183, %f178; sub.ftz.f32 %f231, %f191, %f176; sub.ftz.f32 %f232, %f192, %f177; sub.ftz.f32 %f233, %f193, %f178; sub.ftz.f32 %f234, %f225, %f176; sub.ftz.f32 %f235, %f226, %f177; sub.ftz.f32 %f236, %f227, %f178; mul.ftz.f32 %f237, %f229, %f229; fma.rn.ftz.f32 %f238, %f228, %f228, %f237; fma.rn.ftz.f32 %f24, %f230, %f230, %f238; mul.ftz.f32 %f239, %f229, %f232; fma.rn.ftz.f32 %f240, %f228, %f231, %f239; fma.rn.ftz.f32 %f25, %f230, %f233, %f240; mul.ftz.f32 %f241, %f232, %f232; fma.rn.ftz.f32 %f242, %f231, %f231, %f241; fma.rn.ftz.f32 %f26, %f233, %f233, %f242; mul.ftz.f32 %f243, %f229, %f235; fma.rn.ftz.f32 %f244, %f228, %f234, %f243; fma.rn.ftz.f32 %f27, %f230, %f236, %f244; mul.ftz.f32 %f245, %f232, %f235; fma.rn.ftz.f32 %f246, %f231, %f234, %f245; fma.rn.ftz.f32 %f28, %f233, %f236, %f246; mul.ftz.f32 %f247, %f24, %f26; mul.ftz.f32 %f248, %f25, %f25; sub.ftz.f32 %f29, %f247, %f248; setp.eq.ftz.f32 %p23, %f29, 0f00000000; mov.f32 %f564, 0f00000000; mov.f32 %f565, %f564; @%p23 bra $L__BB204_28; mul.ftz.f32 %f249, %f26, %f27; mul.ftz.f32 %f250, %f25, %f28; sub.ftz.f32 %f251, %f249, %f250; div.approx.ftz.f32 %f252, %f251, %f29; mov.f32 %f253, 0f3727C5AC; max.ftz.f32 %f254, %f252, %f253; mov.f32 %f255, 0f3F7FFF58; mul.ftz.f32 %f256, %f25, %f27; mul.ftz.f32 %f257, %f24, %f28; sub.ftz.f32 %f258, %f257, %f256; div.approx.ftz.f32 %f259, %f258, %f29; max.ftz.f32 %f260, %f259, %f253; min.ftz.f32 %f261, %f260, %f255; min.ftz.f32 %f565, %f254, %f255; sub.ftz.f32 %f263, %f188, %f565; sub.ftz.f32 %f264, %f263, %f261; max.ftz.f32 %f265, %f264, %f253; min.ftz.f32 %f564, %f265, %f255; $L__BB204_28: setp.gt.s32 %p24, %r270, 0; @%p24 bra $L__BB204_30; bra.uni $L__BB204_29; $L__BB204_30: add.ftz.f32 %f267, %f558, 0fBF000000; fma.rn.ftz.f32 %f268, %f267, %f19, %f564; add.ftz.f32 %f269, %f559, 0fBF000000; fma.rn.ftz.f32 %f270, %f269, %f20, %f268; mov.f32 %f271, 0f3F800000; div.approx.ftz.f32 %f272, %f270, %f271; cvt.rmi.ftz.f32.f32 %f273, %f272; sub.ftz.f32 %f274, %f272, %f273; cvt.rzi.ftz.s32.f32 %r213, %f273; and.b32 %r214, %r213, 1; setp.eq.b32 %p25, %r214, 1; sub.ftz.f32 %f275, %f271, %f274; selp.f32 %f276, %f275, %f274, %p25; fma.rn.ftz.f32 %f277, %f267, %f21, %f565; fma.rn.ftz.f32 %f278, %f269, %f22, %f277; sub.ftz.f32 %f566, %f271, %f276; div.approx.ftz.f32 %f279, %f278, %f566; cvt.rmi.ftz.f32.f32 %f280, %f279; sub.ftz.f32 %f281, %f279, %f280; cvt.rzi.ftz.s32.f32 %r215, %f280; and.b32 %r216, %r215, 1; setp.eq.b32 %p26, %r216, 1; sub.ftz.f32 %f282, %f271, %f281; selp.f32 %f283, %f282, %f281, %p26; mul.ftz.f32 %f565, %f566, %f283; bra.uni $L__BB204_31; $L__BB204_29: mov.f32 %f266, 0f3F800000; sub.ftz.f32 %f566, %f266, %f564; $L__BB204_31: ld.const.u64 %rd160, [kernel_params+2232]; cvta.to.global.u64 %rd161, %rd160; cvt.s64.s32 %rd15, %r21; mul.wide.s32 %rd162, %r21, 16; add.s64 %rd163, %rd161, %rd162; add.s64 %rd16, %rd163, 12; ld.const.u64 %rd164, [kernel_params+2256]; cvta.to.global.u64 %rd165, %rd164; ld.global.u32 %r217, [%rd163+12]; mul.wide.u32 %rd166, %r217, 12; add.s64 %rd167, %rd165, %rd166; add.s32 %r218, %r217, 1; mul.wide.u32 %rd168, %r218, 12; add.s64 %rd169, %rd165, %rd168; add.s32 %r219, %r217, 2; mul.wide.u32 %rd170, %r219, 12; add.s64 %rd171, %rd165, %rd170; mov.f32 %f284, 0f3F800000; sub.ftz.f32 %f285, %f284, %f565; sub.ftz.f32 %f43, %f566, %f565; sub.ftz.f32 %f286, %f285, %f43; ld.global.f32 %f287, [%rd167]; ld.global.f32 %f288, [%rd167+4]; ld.global.f32 %f289, [%rd167+8]; ld.global.f32 %f290, [%rd169]; mul.ftz.f32 %f291, %f565, %f290; ld.global.f32 %f292, [%rd169+4]; mul.ftz.f32 %f293, %f565, %f292; ld.global.f32 %f294, [%rd169+8]; mul.ftz.f32 %f295, %f565, %f294; fma.rn.ftz.f32 %f296, %f286, %f287, %f291; fma.rn.ftz.f32 %f297, %f286, %f288, %f293; fma.rn.ftz.f32 %f298, %f286, %f289, %f295; ld.global.f32 %f299, [%rd171]; ld.global.f32 %f300, [%rd171+4]; ld.global.f32 %f301, [%rd171+8]; fma.rn.ftz.f32 %f71, %f43, %f299, %f296; fma.rn.ftz.f32 %f72, %f43, %f300, %f297; fma.rn.ftz.f32 %f73, %f43, %f301, %f298; ld.const.u64 %rd172, [kernel_params+2184]; cvta.to.global.u64 %rd173, %rd172; ld.const.u32 %r34, [kernel_params+980]; mul.wide.s32 %rd174, %r34, 4; add.s64 %rd175, %rd173, %rd174; ld.global.u32 %r35, [%rd175]; and.b32 %r220, %r35, 8; setp.eq.s32 %p27, %r220, 0; sub.ftz.f32 %f598, %f299, %f287; sub.ftz.f32 %f599, %f300, %f288; sub.ftz.f32 %f600, %f301, %f289; sub.ftz.f32 %f601, %f290, %f287; sub.ftz.f32 %f602, %f292, %f288; sub.ftz.f32 %f603, %f294, %f289; mul.ftz.f32 %f53, %f603, %f599; mul.ftz.f32 %f54, %f602, %f600; mul.ftz.f32 %f55, %f601, %f600; mul.ftz.f32 %f56, %f603, %f598; mul.ftz.f32 %f57, %f602, %f598; mul.ftz.f32 %f58, %f601, %f599; @%p27 bra $L__BB204_33; bra.uni $L__BB204_32; $L__BB204_33: sub.ftz.f32 %f309, %f54, %f53; sub.ftz.f32 %f310, %f56, %f55; mul.ftz.f32 %f311, %f310, %f310; fma.rn.ftz.f32 %f312, %f309, %f309, %f311; sub.ftz.f32 %f313, %f58, %f57; fma.rn.ftz.f32 %f314, %f313, %f313, %f312; sqrt.approx.ftz.f32 %f315, %f314; div.approx.ftz.f32 %f568, %f309, %f315; div.approx.ftz.f32 %f569, %f310, %f315; div.approx.ftz.f32 %f570, %f313, %f315; bra.uni $L__BB204_34; $L__BB204_32: sub.ftz.f32 %f302, %f53, %f54; sub.ftz.f32 %f303, %f55, %f56; mul.ftz.f32 %f304, %f303, %f303; fma.rn.ftz.f32 %f305, %f302, %f302, %f304; sub.ftz.f32 %f306, %f57, %f58; fma.rn.ftz.f32 %f307, %f306, %f306, %f305; sqrt.approx.ftz.f32 %f308, %f307; div.approx.ftz.f32 %f568, %f302, %f308; div.approx.ftz.f32 %f569, %f303, %f308; div.approx.ftz.f32 %f570, %f306, %f308; $L__BB204_34: ld.const.u64 %rd176, [kernel_params+2216]; cvta.to.global.u64 %rd177, %rd176; shl.b64 %rd178, %rd15, 2; add.s64 %rd179, %rd177, %rd178; ld.global.u32 %r36, [%rd179]; and.b32 %r221, %r35, 4; setp.ne.s32 %p28, %r221, 0; @%p28 bra $L__BB204_36; ld.const.u64 %rd180, [kernel_params+2160]; cvta.to.global.u64 %rd181, %rd180; mul.wide.s32 %rd182, %r34, 208; add.s64 %rd183, %rd181, %rd182; ld.global.v4.f32 {%f316, %f317, %f318, %f319}, [%rd183]; ld.global.v4.f32 {%f324, %f325, %f326, %f327}, [%rd183+16]; ld.global.v4.f32 {%f332, %f333, %f334, %f335}, [%rd183+32]; mul.ftz.f32 %f340, %f72, %f317; fma.rn.ftz.f32 %f341, %f71, %f316, %f340; fma.rn.ftz.f32 %f342, %f73, %f318, %f341; add.ftz.f32 %f68, %f319, %f342; mul.ftz.f32 %f343, %f72, %f325; fma.rn.ftz.f32 %f344, %f71, %f324, %f343; fma.rn.ftz.f32 %f345, %f73, %f326, %f344; add.ftz.f32 %f69, %f327, %f345; mul.ftz.f32 %f346, %f72, %f333; fma.rn.ftz.f32 %f347, %f71, %f332, %f346; fma.rn.ftz.f32 %f348, %f73, %f334, %f347; add.ftz.f32 %f73, %f335, %f348; mov.f32 %f71, %f68; mov.f32 %f72, %f69; $L__BB204_36: ld.const.u32 %r222, [kernel_params+1356]; setp.eq.s32 %p29, %r222, -1; @%p29 bra $L__BB204_38; mov.u32 %r263, 2139095039; mov.u32 %r262, 0; mul.ftz.f32 %f349, %f72, %f72; fma.rn.ftz.f32 %f350, %f71, %f71, %f349; fma.rn.ftz.f32 %f351, %f73, %f73, %f350; sqrt.approx.ftz.f32 %f352, %f351; div.approx.ftz.f32 %f353, %f71, %f352; div.approx.ftz.f32 %f354, %f72, %f352; div.approx.ftz.f32 %f355, %f73, %f352; ld.const.u64 %rd184, [kernel_params+2632]; cvta.to.global.u64 %rd185, %rd184; add.s64 %rd187, %rd185, %rd119; st.global.u32 [%rd187], %r262; st.global.u32 [%rd187+4], %r262; st.global.u32 [%rd187+8], %r262; ld.const.v2.u64 {%rd188, %rd189}, [kernel_params+2640]; cvta.to.global.u64 %rd192, %rd188; add.s64 %rd193, %rd192, %rd119; st.global.f32 [%rd193], %f353; st.global.f32 [%rd193+4], %f354; st.global.f32 [%rd193+8], %f355; cvta.to.global.u64 %rd194, %rd189; add.s64 %rd196, %rd194, %rd61; st.global.u32 [%rd196], %r262; ld.const.v2.u64 {%rd197, %rd198}, [kernel_params+2656]; cvta.to.global.u64 %rd201, %rd197; add.s64 %rd202, %rd201, %rd61; st.global.u32 [%rd202], %r263; cvta.to.global.u64 %rd203, %rd198; add.s64 %rd204, %rd203, %rd61; mov.u32 %r225, 1056964608; st.global.u32 [%rd204], %r225; ld.const.v2.u64 {%rd205, %rd206}, [kernel_params+2672]; cvta.to.global.u64 %rd209, %rd205; add.s64 %rd210, %rd209, %rd61; st.global.u32 [%rd210], %r262; cvta.to.global.u64 %rd211, %rd206; add.s64 %rd212, %rd211, %rd61; st.global.u32 [%rd212], %r262; ld.const.u64 %rd213, [kernel_params+4256]; cvta.to.global.u64 %rd214, %rd213; add.s64 %rd215, %rd214, 24; atom.global.add.u32 %r226, [%rd215], 1; mov.u16 %rs7, 6; st.global.u16 [%rd5], %rs7; bra.uni $L__BB204_69; $L__BB204_38: setp.gt.s32 %p30, %r36, -1; mov.f32 %f580, %f568; mov.f32 %f581, %f569; mov.f32 %f582, %f570; @%p30 bra $L__BB204_42; ld.global.v4.u32 {%r227, %r228, %r229, %r230}, [%rd16+-12]; ld.const.u64 %rd216, [kernel_params+2224]; cvta.to.global.u64 %rd217, %rd216; mul.wide.u32 %rd218, %r227, 12; add.s64 %rd219, %rd217, %rd218; mul.wide.u32 %rd220, %r228, 12; add.s64 %rd221, %rd217, %rd220; mul.wide.u32 %rd222, %r229, 12; add.s64 %rd223, %rd217, %rd222; ld.global.f32 %f359, [%rd219]; ld.global.f32 %f360, [%rd219+4]; ld.global.f32 %f361, [%rd219+8]; ld.global.f32 %f362, [%rd221]; mul.ftz.f32 %f363, %f565, %f362; ld.global.f32 %f364, [%rd221+4]; mul.ftz.f32 %f365, %f565, %f364; ld.global.f32 %f366, [%rd221+8]; mul.ftz.f32 %f367, %f565, %f366; fma.rn.ftz.f32 %f368, %f286, %f359, %f363; fma.rn.ftz.f32 %f369, %f286, %f360, %f365; fma.rn.ftz.f32 %f370, %f286, %f361, %f367; ld.global.f32 %f371, [%rd223]; ld.global.f32 %f372, [%rd223+4]; ld.global.f32 %f373, [%rd223+8]; fma.rn.ftz.f32 %f574, %f43, %f371, %f368; fma.rn.ftz.f32 %f575, %f43, %f372, %f369; fma.rn.ftz.f32 %f576, %f43, %f373, %f370; mul.ftz.f32 %f374, %f575, %f575; fma.rn.ftz.f32 %f375, %f574, %f574, %f374; fma.rn.ftz.f32 %f376, %f576, %f576, %f375; sqrt.approx.ftz.f32 %f77, %f376; setp.neu.ftz.f32 %p31, %f77, 0f00000000; @%p31 bra $L__BB204_40; bra.uni $L__BB204_41; $L__BB204_40: rcp.approx.ftz.f32 %f377, %f77; mul.ftz.f32 %f576, %f576, %f377; mul.ftz.f32 %f575, %f575, %f377; mul.ftz.f32 %f574, %f574, %f377; $L__BB204_41: setp.eq.ftz.f32 %p32, %f574, 0f00000000; setp.eq.ftz.f32 %p33, %f575, 0f00000000; and.pred %p34, %p32, %p33; setp.eq.ftz.f32 %p35, %f576, 0f00000000; and.pred %p36, %p35, %p34; selp.f32 %f580, %f568, %f574, %p36; selp.f32 %f581, %f569, %f575, %p36; selp.f32 %f582, %f570, %f576, %p36; $L__BB204_42: @%p28 bra $L__BB204_44; ld.const.u64 %rd224, [kernel_params+2160]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r34, 208; add.s64 %rd227, %rd225, %rd226; ld.global.v4.f32 {%f378, %f379, %f380, %f381}, [%rd227+48]; ld.global.v4.f32 {%f385, %f386, %f387, %f388}, [%rd227+64]; ld.global.v4.f32 {%f392, %f393, %f394, %f395}, [%rd227+80]; mul.ftz.f32 %f399, %f581, %f385; fma.rn.ftz.f32 %f400, %f580, %f378, %f399; fma.rn.ftz.f32 %f401, %f582, %f392, %f400; mul.ftz.f32 %f402, %f581, %f386; fma.rn.ftz.f32 %f403, %f580, %f379, %f402; fma.rn.ftz.f32 %f404, %f582, %f393, %f403; mul.ftz.f32 %f405, %f581, %f387; fma.rn.ftz.f32 %f406, %f580, %f380, %f405; fma.rn.ftz.f32 %f407, %f582, %f394, %f406; mul.ftz.f32 %f408, %f404, %f404; fma.rn.ftz.f32 %f409, %f401, %f401, %f408; fma.rn.ftz.f32 %f410, %f407, %f407, %f409; sqrt.approx.ftz.f32 %f411, %f410; div.approx.ftz.f32 %f580, %f401, %f411; div.approx.ftz.f32 %f581, %f404, %f411; div.approx.ftz.f32 %f582, %f407, %f411; mul.ftz.f32 %f412, %f569, %f385; fma.rn.ftz.f32 %f413, %f568, %f378, %f412; fma.rn.ftz.f32 %f414, %f570, %f392, %f413; mul.ftz.f32 %f415, %f569, %f386; fma.rn.ftz.f32 %f416, %f568, %f379, %f415; fma.rn.ftz.f32 %f417, %f570, %f393, %f416; mul.ftz.f32 %f418, %f569, %f387; fma.rn.ftz.f32 %f419, %f568, %f380, %f418; fma.rn.ftz.f32 %f420, %f570, %f394, %f419; mul.ftz.f32 %f421, %f417, %f417; fma.rn.ftz.f32 %f422, %f414, %f414, %f421; fma.rn.ftz.f32 %f423, %f420, %f420, %f422; sqrt.approx.ftz.f32 %f424, %f423; div.approx.ftz.f32 %f568, %f414, %f424; div.approx.ftz.f32 %f569, %f417, %f424; div.approx.ftz.f32 %f570, %f420, %f424; $L__BB204_44: ld.const.u64 %rd228, [kernel_params+2408]; cvta.to.global.u64 %rd229, %rd228; and.b32 %r37, %r36, 4194303; shl.b32 %r235, %r36, 5; cvt.u64.u32 %rd230, %r235; and.b64 %rd231, %rd230, 134217696; add.s64 %rd232, %rd229, %rd231; ld.global.u32 %r38, [%rd232+16]; ld.const.u32 %r39, [kernel_params+1276]; setp.eq.s32 %p38, %r39, -1; @%p38 bra $L__BB204_46; ld.param.u64 %rd335, [kernel_gpu_integrator_init_from_bake_param_2]; cvta.to.global.u64 %rd334, %rd335; cvt.s64.s32 %rd233, %r39; add.s64 %rd234, %rd9, %rd233; shl.b64 %rd236, %rd234, 2; add.s64 %rd237, %rd334, %rd236; add.s64 %rd238, %rd237, 4; add.s64 %rd239, %rd237, 8; atom.global.add.f32 %f425, [%rd237], %f71; atom.global.add.f32 %f426, [%rd238], %f72; atom.global.add.f32 %f427, [%rd239], %f73; bra.uni $L__BB204_69; $L__BB204_46: ld.const.u32 %r40, [kernel_params+1280]; setp.ne.s32 %p39, %r40, -1; and.b32 %r236, %r38, 33554432; setp.eq.s32 %p40, %r236, 0; and.pred %p41, %p40, %p39; @%p41 bra $L__BB204_64; bra.uni $L__BB204_47; $L__BB204_64: ld.param.u64 %rd337, [kernel_gpu_integrator_init_from_bake_param_2]; cvta.to.global.u64 %rd336, %rd337; cvt.s64.s32 %rd317, %r40; add.s64 %rd318, %rd9, %rd317; shl.b64 %rd320, %rd318, 2; add.s64 %rd321, %rd336, %rd320; add.s64 %rd322, %rd321, 4; add.s64 %rd323, %rd321, 8; atom.global.add.f32 %f551, [%rd321], %f580; atom.global.add.f32 %f552, [%rd322], %f581; atom.global.add.f32 %f553, [%rd323], %f582; bra.uni $L__BB204_69; $L__BB204_47: ld.const.u32 %r237, [kernel_params+988]; setp.eq.s32 %p42, %r237, 0; mov.f32 %f590, %f580; mov.f32 %f589, %f581; mov.f32 %f591, %f582; @%p42 bra $L__BB204_57; ld.const.v4.f32 {%f428, %f429, %f430, %f431}, [kernel_params+96]; ld.const.v4.f32 {%f432, %f433, %f434, %f435}, [kernel_params+112]; ld.const.v4.f32 {%f436, %f437, %f438, %f439}, [kernel_params+128]; ld.const.u32 %r238, [kernel_params+16]; setp.eq.s32 %p43, %r238, 1; @%p43 bra $L__BB204_50; bra.uni $L__BB204_49; $L__BB204_50: neg.ftz.f32 %f591, %f438; neg.ftz.f32 %f589, %f434; neg.ftz.f32 %f590, %f430; bra.uni $L__BB204_51; $L__BB204_49: sub.ftz.f32 %f440, %f431, %f71; sub.ftz.f32 %f441, %f435, %f72; sub.ftz.f32 %f442, %f439, %f73; mul.ftz.f32 %f443, %f441, %f441; fma.rn.ftz.f32 %f444, %f440, %f440, %f443; fma.rn.ftz.f32 %f445, %f442, %f442, %f444; sqrt.approx.ftz.f32 %f446, %f445; div.approx.ftz.f32 %f591, %f442, %f446; div.approx.ftz.f32 %f589, %f441, %f446; div.approx.ftz.f32 %f590, %f440, %f446; $L__BB204_51: mul.ftz.f32 %f447, %f581, %f589; fma.rn.ftz.f32 %f448, %f580, %f590, %f447; fma.rn.ftz.f32 %f129, %f582, %f591, %f448; setp.gtu.ftz.f32 %p44, %f129, 0f3E4CCCCD; @%p44 bra $L__BB204_53; add.ftz.f32 %f449, %f129, 0fBE4CCCCD; mul.ftz.f32 %f450, %f580, %f449; mul.ftz.f32 %f451, %f581, %f449; mul.ftz.f32 %f452, %f582, %f449; sub.ftz.f32 %f453, %f590, %f450; sub.ftz.f32 %f454, %f589, %f451; sub.ftz.f32 %f455, %f591, %f452; mul.ftz.f32 %f456, %f454, %f454; fma.rn.ftz.f32 %f457, %f453, %f453, %f456; fma.rn.ftz.f32 %f458, %f455, %f455, %f457; sqrt.approx.ftz.f32 %f459, %f458; div.approx.ftz.f32 %f590, %f453, %f459; div.approx.ftz.f32 %f589, %f454, %f459; div.approx.ftz.f32 %f591, %f455, %f459; $L__BB204_53: setp.gtu.ftz.f32 %p45, %f129, 0f00000000; @%p45 bra $L__BB204_55; sub.ftz.f32 %f460, %f580, %f590; mul.ftz.f32 %f461, %f460, %f129; sub.ftz.f32 %f462, %f581, %f589; mul.ftz.f32 %f463, %f462, %f129; sub.ftz.f32 %f464, %f582, %f591; mul.ftz.f32 %f465, %f464, %f129; sub.ftz.f32 %f466, %f590, %f461; sub.ftz.f32 %f467, %f589, %f463; sub.ftz.f32 %f468, %f591, %f465; mul.ftz.f32 %f469, %f466, %f466; fma.rn.ftz.f32 %f470, %f467, %f467, %f469; fma.rn.ftz.f32 %f471, %f468, %f468, %f470; sqrt.approx.ftz.f32 %f472, %f471; div.approx.ftz.f32 %f590, %f466, %f472; div.approx.ftz.f32 %f589, %f467, %f472; div.approx.ftz.f32 %f591, %f468, %f472; $L__BB204_55: mul.ftz.f32 %f473, %f568, %f590; fma.rn.ftz.f32 %f474, %f569, %f589, %f473; fma.rn.ftz.f32 %f142, %f570, %f591, %f474; setp.gtu.ftz.f32 %p46, %f142, 0f3727C5AC; @%p46 bra $L__BB204_57; add.ftz.f32 %f475, %f142, 0fB727C5AC; mul.ftz.f32 %f476, %f568, %f475; mul.ftz.f32 %f477, %f569, %f475; mul.ftz.f32 %f478, %f570, %f475; sub.ftz.f32 %f479, %f590, %f476; sub.ftz.f32 %f480, %f589, %f477; sub.ftz.f32 %f481, %f591, %f478; mul.ftz.f32 %f482, %f480, %f480; fma.rn.ftz.f32 %f483, %f479, %f479, %f482; fma.rn.ftz.f32 %f484, %f481, %f481, %f483; sqrt.approx.ftz.f32 %f485, %f484; div.approx.ftz.f32 %f590, %f479, %f485; div.approx.ftz.f32 %f589, %f480, %f485; div.approx.ftz.f32 %f591, %f481, %f485; $L__BB204_57: @%p28 bra $L__BB204_59; ld.const.u64 %rd240, [kernel_params+2160]; cvta.to.global.u64 %rd241, %rd240; mul.wide.s32 %rd242, %r34, 208; add.s64 %rd243, %rd241, %rd242; ld.global.v4.f32 {%f486, %f487, %f488, %f489}, [%rd243]; ld.global.v4.f32 {%f493, %f494, %f495, %f496}, [%rd243+16]; ld.global.v4.f32 {%f500, %f501, %f502, %f503}, [%rd243+32]; mul.ftz.f32 %f507, %f602, %f487; fma.rn.ftz.f32 %f508, %f601, %f486, %f507; fma.rn.ftz.f32 %f149, %f603, %f488, %f508; mul.ftz.f32 %f509, %f602, %f494; fma.rn.ftz.f32 %f510, %f601, %f493, %f509; fma.rn.ftz.f32 %f150, %f603, %f495, %f510; mul.ftz.f32 %f511, %f602, %f501; fma.rn.ftz.f32 %f512, %f601, %f500, %f511; fma.rn.ftz.f32 %f603, %f603, %f502, %f512; mul.ftz.f32 %f513, %f599, %f487; fma.rn.ftz.f32 %f514, %f598, %f486, %f513; fma.rn.ftz.f32 %f152, %f600, %f488, %f514; mul.ftz.f32 %f515, %f599, %f494; fma.rn.ftz.f32 %f516, %f598, %f493, %f515; fma.rn.ftz.f32 %f153, %f600, %f495, %f516; mul.ftz.f32 %f517, %f599, %f501; fma.rn.ftz.f32 %f518, %f598, %f500, %f517; fma.rn.ftz.f32 %f600, %f600, %f502, %f518; mov.f32 %f598, %f152; mov.f32 %f599, %f153; mov.f32 %f601, %f149; mov.f32 %f602, %f150; $L__BB204_59: mov.u32 %r266, 2139095039; mov.u32 %r265, 0; mov.u32 %r264, 1065353216; neg.ftz.f32 %f519, %f19; sub.ftz.f32 %f520, %f519, %f21; neg.ftz.f32 %f521, %f20; sub.ftz.f32 %f522, %f521, %f22; mul.ftz.f32 %f523, %f21, %f601; fma.rn.ftz.f32 %f524, %f520, %f598, %f523; mul.ftz.f32 %f525, %f21, %f602; fma.rn.ftz.f32 %f526, %f520, %f599, %f525; mul.ftz.f32 %f527, %f21, %f603; fma.rn.ftz.f32 %f528, %f520, %f600, %f527; mul.ftz.f32 %f529, %f22, %f601; fma.rn.ftz.f32 %f530, %f522, %f598, %f529; mul.ftz.f32 %f531, %f22, %f602; fma.rn.ftz.f32 %f532, %f522, %f599, %f531; mul.ftz.f32 %f533, %f22, %f603; fma.rn.ftz.f32 %f534, %f522, %f600, %f533; mul.ftz.f32 %f535, %f526, %f526; fma.rn.ftz.f32 %f536, %f524, %f524, %f535; fma.rn.ftz.f32 %f537, %f528, %f528, %f536; sqrt.approx.ftz.f32 %f538, %f537; mul.ftz.f32 %f539, %f532, %f532; fma.rn.ftz.f32 %f540, %f530, %f530, %f539; fma.rn.ftz.f32 %f541, %f534, %f534, %f540; sqrt.approx.ftz.f32 %f542, %f541; add.ftz.f32 %f543, %f538, %f542; mul.ftz.f32 %f544, %f543, 0f3F000000; mov.u32 %r240, 1056964608; ld.const.u64 %rd244, [kernel_params+2632]; cvta.to.global.u64 %rd245, %rd244; add.s64 %rd247, %rd245, %rd119; add.ftz.f32 %f545, %f71, %f590; st.global.f32 [%rd247], %f545; add.ftz.f32 %f546, %f72, %f589; st.global.f32 [%rd247+4], %f546; add.ftz.f32 %f547, %f73, %f591; st.global.f32 [%rd247+8], %f547; ld.const.v2.u64 {%rd248, %rd249}, [kernel_params+2640]; mov.u32 %r242, 1; cvta.to.global.u64 %rd252, %rd248; add.s64 %rd253, %rd252, %rd119; neg.ftz.f32 %f548, %f590; st.global.f32 [%rd253], %f548; neg.ftz.f32 %f549, %f589; st.global.f32 [%rd253+4], %f549; neg.ftz.f32 %f550, %f591; st.global.f32 [%rd253+8], %f550; cvta.to.global.u64 %rd254, %rd249; add.s64 %rd256, %rd254, %rd61; st.global.u32 [%rd256], %r265; ld.const.v2.u64 {%rd257, %rd258}, [kernel_params+2656]; cvta.to.global.u64 %rd261, %rd257; add.s64 %rd262, %rd261, %rd61; st.global.u32 [%rd262], %r266; cvta.to.global.u64 %rd263, %rd258; add.s64 %rd264, %rd263, %rd61; st.global.u32 [%rd264], %r240; ld.const.v2.u64 {%rd265, %rd266}, [kernel_params+2672]; cvta.to.global.u64 %rd269, %rd265; add.s64 %rd270, %rd269, %rd61; st.global.f32 [%rd270], %f544; cvta.to.global.u64 %rd271, %rd266; add.s64 %rd272, %rd271, %rd61; st.global.u32 [%rd272], %r265; ld.const.v2.u64 {%rd273, %rd274}, [kernel_params+2688]; cvta.to.global.u64 %rd277, %rd273; add.s64 %rd278, %rd277, %rd61; st.global.u32 [%rd278], %r264; cvta.to.global.u64 %rd279, %rd274; add.s64 %rd280, %rd279, %rd61; st.global.f32 [%rd280], %f565; ld.const.u64 %rd281, [kernel_params+2704]; cvta.to.global.u64 %rd282, %rd281; add.s64 %rd283, %rd282, %rd61; st.global.f32 [%rd283], %f43; st.global.u32 [%rd13], %r34; st.global.u32 [%rd14], %r21; ld.const.u64 %rd284, [kernel_params+2728]; cvta.to.global.u64 %rd285, %rd284; add.s64 %rd286, %rd285, %rd61; st.global.u32 [%rd286], %r242; ld.const.u32 %r245, [kernel_params+1592]; setp.ne.s32 %p48, %r245, 0; and.b32 %r246, %r35, 1536; setp.ne.s32 %p49, %r246, 0; and.pred %p50, %p49, %p48; @%p50 bra $L__BB204_63; bra.uni $L__BB204_60; $L__BB204_63: ld.const.u32 %r255, [kernel_params+8]; ld.const.u32 %r256, [kernel_params+4392]; div.u32 %r257, %r3, %r256; mad.lo.s32 %r258, %r257, %r255, %r37; ld.const.u64 %rd306, [kernel_params+4256]; cvta.to.global.u64 %rd307, %rd306; add.s64 %rd308, %rd307, 40; atom.global.add.u32 %r259, [%rd308], 1; mov.u16 %rs10, 10; st.global.u16 [%rd5], %rs10; ld.const.u64 %rd309, [kernel_params+2624]; cvta.to.global.u64 %rd310, %rd309; add.s64 %rd312, %rd310, %rd61; st.global.u32 [%rd312], %r258; ld.const.u64 %rd313, [kernel_params+4344]; cvta.to.global.u64 %rd314, %rd313; mul.wide.s32 %rd315, %r258, 4; add.s64 %rd316, %rd314, %rd315; atom.global.add.u32 %r260, [%rd316], 1; bra.uni $L__BB204_69; $L__BB204_60: and.b32 %r247, %r38, 1073741824; setp.eq.s32 %p51, %r247, 0; ld.const.u32 %r248, [kernel_params+4392]; div.u32 %r249, %r3, %r248; ld.const.u32 %r250, [kernel_params+8]; mad.lo.s32 %r41, %r249, %r250, %r37; ld.const.u64 %rd287, [kernel_params+4256]; cvta.to.global.u64 %rd17, %rd287; cvt.s64.s32 %rd18, %r41; @%p51 bra $L__BB204_62; bra.uni $L__BB204_61; $L__BB204_62: add.s64 %rd297, %rd17, 32; atom.global.add.u32 %r253, [%rd297], 1; mov.u16 %rs9, 8; st.global.u16 [%rd5], %rs9; ld.const.u64 %rd298, [kernel_params+2624]; cvta.to.global.u64 %rd299, %rd298; add.s64 %rd301, %rd299, %rd61; st.global.u32 [%rd301], %r41; ld.const.u64 %rd302, [kernel_params+4328]; cvta.to.global.u64 %rd303, %rd302; shl.b64 %rd304, %rd18, 2; add.s64 %rd305, %rd303, %rd304; atom.global.add.u32 %r254, [%rd305], 1; bra.uni $L__BB204_69; $L__BB204_61: add.s64 %rd288, %rd17, 36; atom.global.add.u32 %r251, [%rd288], 1; mov.u16 %rs8, 9; st.global.u16 [%rd5], %rs8; ld.const.u64 %rd289, [kernel_params+2624]; cvta.to.global.u64 %rd290, %rd289; add.s64 %rd292, %rd290, %rd61; st.global.u32 [%rd292], %r41; ld.const.u64 %rd293, [kernel_params+4336]; cvta.to.global.u64 %rd294, %rd293; shl.b64 %rd295, %rd18, 2; add.s64 %rd296, %rd294, %rd295; atom.global.add.u32 %r252, [%rd296], 1; $L__BB204_69: ret; } // .globl kernel_gpu_integrator_intersect_closest .visible .entry kernel_gpu_integrator_intersect_closest( .param .u64 kernel_gpu_integrator_intersect_closest_param_0, .param .u64 kernel_gpu_integrator_intersect_closest_param_1, .param .u32 kernel_gpu_integrator_intersect_closest_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 4 .b8 __local_depot205[84]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<274>; .reg .b16 %rs<103>; .reg .f32 %f<332>; .reg .b32 %r<476>; .reg .b64 %rd<842>; mov.u64 %SPL, __local_depot205; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd127, [kernel_gpu_integrator_intersect_closest_param_0]; ld.param.u32 %r93, [kernel_gpu_integrator_intersect_closest_param_2]; mov.u32 %r94, %ntid.x; mov.u32 %r95, %ctaid.x; mov.u32 %r96, %tid.x; mad.lo.s32 %r3, %r95, %r94, %r96; setp.ge.s32 %p10, %r3, %r93; @%p10 bra $L__BB205_297; setp.eq.s64 %p11, %rd127, 0; @%p11 bra $L__BB205_3; cvta.to.global.u64 %rd129, %rd127; mul.wide.s32 %rd130, %r3, 4; add.s64 %rd131, %rd129, %rd130; ld.global.u32 %r3, [%rd131]; $L__BB205_3: ld.const.u64 %rd1, [kernel_params+2632]; cvta.to.global.u64 %rd2, %rd1; cvt.s64.s32 %rd3, %r3; mul.wide.s32 %rd132, %r3, 12; add.s64 %rd4, %rd2, %rd132; ld.global.f32 %f1, [%rd4]; ld.global.f32 %f2, [%rd4+4]; ld.global.f32 %f3, [%rd4+8]; add.u64 %rd5, %SPL, 0; st.local.f32 [%rd5], %f1; st.local.f32 [%rd5+4], %f2; st.local.f32 [%rd5+8], %f3; ld.const.v2.u64 {%rd134, %rd135}, [kernel_params+2640]; cvta.to.global.u64 %rd136, %rd134; add.s64 %rd137, %rd136, %rd132; ld.global.f32 %f4, [%rd137]; ld.global.f32 %f5, [%rd137+4]; ld.global.f32 %f6, [%rd137+8]; st.local.f32 [%rd5+12], %f4; st.local.f32 [%rd5+16], %f5; st.local.f32 [%rd5+20], %f6; cvta.to.global.u64 %rd138, %rd135; mul.wide.s32 %rd139, %r3, 4; add.s64 %rd8, %rd138, %rd139; ld.global.f32 %f7, [%rd8]; st.local.f32 [%rd5+24], %f7; ld.const.v2.u64 {%rd140, %rd141}, [kernel_params+2656]; cvta.to.global.u64 %rd142, %rd140; add.s64 %rd11, %rd142, %rd139; ld.global.f32 %f313, [%rd11]; st.local.f32 [%rd5+28], %f313; cvta.to.global.u64 %rd143, %rd141; add.s64 %rd144, %rd143, %rd139; ld.global.f32 %f86, [%rd144]; st.local.f32 [%rd5+32], %f86; ld.const.v2.u64 {%rd145, %rd146}, [kernel_params+2672]; cvta.to.global.u64 %rd147, %rd145; add.s64 %rd148, %rd147, %rd139; ld.global.f32 %f87, [%rd148]; st.local.f32 [%rd5+52], %f87; cvta.to.global.u64 %rd149, %rd146; add.s64 %rd14, %rd149, %rd139; ld.global.f32 %f88, [%rd14]; st.local.f32 [%rd5+56], %f88; ld.const.u64 %rd15, [kernel_params+2544]; cvta.to.global.u64 %rd150, %rd15; add.s64 %rd16, %rd150, %rd139; ld.global.u32 %r97, [%rd16]; and.b32 %r98, %r97, 4; setp.eq.s32 %p12, %r98, 0; selp.b32 %r99, 2047, 2023, %p12; and.b32 %r100, %r99, %r97; and.b32 %r101, %r97, 128; shr.u32 %r102, %r101, 4; or.b32 %r103, %r100, %r102; and.b32 %r104, %r97, 1073741824; setp.eq.s32 %p13, %r104, 0; shl.b32 %r105, %r103, 16; selp.b32 %r4, %r103, %r105, %p13; ld.const.u64 %rd17, [kernel_params+2712]; cvta.to.global.u64 %rd18, %rd17; add.s64 %rd19, %rd18, %rd139; ld.global.u32 %r5, [%rd19]; ld.const.u64 %rd20, [kernel_params+2720]; cvta.to.global.u64 %rd21, %rd20; add.s64 %rd22, %rd21, %rd139; ld.global.u32 %r6, [%rd22]; cvt.s64.s32 %rd23, %r6; ld.const.u32 %r7, [kernel_params+1536]; setp.eq.s32 %p14, %r7, 0; mov.u16 %rs90, 0; @%p14 bra $L__BB205_5; ld.const.v2.u64 {%rd151, %rd152}, [kernel_params+2464]; cvta.to.global.u64 %rd154, %rd151; shl.b64 %rd155, %rd3, 1; add.s64 %rd156, %rd154, %rd155; ld.global.u16 %r106, [%rd156]; ld.const.u64 %rd157, [kernel_params+2496]; cvta.to.global.u64 %rd158, %rd157; add.s64 %rd159, %rd158, %rd155; ld.global.u16 %r107, [%rd159]; sub.s32 %r108, %r106, %r107; ld.const.v2.u64 {%rd160, %rd161}, [kernel_params+2480]; cvta.to.global.u64 %rd163, %rd161; add.s64 %rd164, %rd163, %rd155; ld.global.u16 %rs33, [%rd164]; setp.ne.s16 %p15, %rs33, 0; selp.b32 %r109, -1, 0, %p15; add.s32 %r110, %r108, %r109; setp.ge.s32 %p16, %r110, %r7; selp.u16 %rs90, 1, 0, %p16; $L__BB205_5: setp.eq.s16 %p17, %rs90, 0; @%p17 bra $L__BB205_9; add.u64 %rd810, %SPL, 0; ld.const.f32 %f313, [kernel_params+1540]; st.local.f32 [%rd810+28], %f313; setp.eq.s32 %p18, %r6, -1; @%p18 bra $L__BB205_9; ld.const.u64 %rd165, [kernel_params+2160]; cvta.to.global.u64 %rd166, %rd165; mul.lo.s64 %rd167, %rd23, 208; add.s64 %rd168, %rd166, %rd167; ld.global.f32 %f10, [%rd168+188]; setp.eq.ftz.f32 %p19, %f10, 0f00000000; @%p19 bra $L__BB205_9; add.u64 %rd812, %SPL, 0; st.local.f32 [%rd812+28], %f10; mov.f32 %f313, %f10; $L__BB205_9: add.u64 %rd814, %SPL, 0; add.u64 %rd170, %SPL, 60; add.s64 %rd24, %rd170, 16; mov.u32 %r111, -1; st.local.u32 [%rd170+16], %r111; st.local.u32 [%rd170+12], %r111; st.local.u32 [%rd814+40], %r6; st.local.u32 [%rd814+36], %r5; st.local.u32 [%rd814+48], %r111; st.local.u32 [%rd814+44], %r111; mov.b32 %r8, %f1; setp.nan.ftz.f32 %p20, %f1, %f1; mov.u16 %rs91, 0; @%p20 bra $L__BB205_12; mov.u16 %rs91, 0; or.b32 %r112, %r8, -2147483648; setp.ne.s32 %p21, %r112, -2147483648; add.ftz.f32 %f89, %f1, %f1; setp.eq.ftz.f32 %p22, %f89, %f1; and.pred %p23, %p22, %p21; @%p23 bra $L__BB205_12; shl.b32 %r113, %r8, 1; setp.lt.u32 %p24, %r113, -16777215; selp.u16 %rs91, 1, 0, %p24; $L__BB205_12: mov.u16 %rs93, 0; setp.eq.s16 %p25, %rs91, 0; @%p25 bra $L__BB205_18; mov.b32 %r9, %f4; setp.nan.ftz.f32 %p26, %f4, %f4; mov.u16 %rs93, 0; mov.u16 %rs92, %rs93; @%p26 bra $L__BB205_16; or.b32 %r114, %r9, -2147483648; setp.ne.s32 %p27, %r114, -2147483648; add.ftz.f32 %f90, %f4, %f4; setp.eq.ftz.f32 %p28, %f90, %f4; and.pred %p29, %p28, %p27; mov.u16 %rs92, %rs93; @%p29 bra $L__BB205_16; shl.b32 %r115, %r9, 1; setp.lt.u32 %p30, %r115, -16777215; selp.u16 %rs92, 1, 0, %p30; $L__BB205_16: setp.eq.s16 %p31, %rs92, 0; @%p31 bra $L__BB205_18; mul.ftz.f32 %f91, %f5, %f5; fma.rn.ftz.f32 %f92, %f4, %f4, %f91; fma.rn.ftz.f32 %f93, %f6, %f6, %f92; setp.neu.ftz.f32 %p32, %f93, 0f00000000; selp.u16 %rs93, 1, 0, %p32; $L__BB205_18: mov.u16 %rs94, 0; setp.eq.s16 %p33, %rs93, 0; @%p33 bra $L__BB205_26; ld.const.u32 %r116, [kernel_params+1108]; setp.eq.s32 %p34, %r116, 0; ld.const.u32 %r10, [kernel_params+1112]; @%p34 bra $L__BB205_23; setp.ne.s32 %p35, %r10, 0; @%p35 bra $L__BB205_22; bra.uni $L__BB205_21; $L__BB205_22: add.u64 %rd826, %SP, 60; add.u64 %rd817, %SP, 0; { // callseq 969, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd817; .param .b64 param1; st.param.b64 [param1+0], %rd826; .param .b32 param2; st.param.b32 [param2+0], %r4; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r118, [retval0+0]; } // callseq 969 cvt.u16.u32 %rs94, %r118; bra.uni $L__BB205_26; $L__BB205_23: setp.ne.s32 %p36, %r10, 0; @%p36 bra $L__BB205_25; bra.uni $L__BB205_24; $L__BB205_25: add.u64 %rd828, %SP, 60; add.u64 %rd819, %SP, 0; { // callseq 971, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd819; .param .b64 param1; st.param.b64 [param1+0], %rd828; .param .b32 param2; st.param.b32 [param2+0], %r4; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r120, [retval0+0]; } // callseq 971 cvt.u16.u32 %rs94, %r120; bra.uni $L__BB205_26; $L__BB205_21: add.u64 %rd825, %SP, 60; add.u64 %rd816, %SP, 0; { // callseq 968, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd816; .param .b64 param1; st.param.b64 [param1+0], %rd825; .param .b32 param2; st.param.b32 [param2+0], %r4; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r117, [retval0+0]; } // callseq 968 cvt.u16.u32 %rs94, %r117; bra.uni $L__BB205_26; $L__BB205_24: add.u64 %rd827, %SP, 60; add.u64 %rd818, %SP, 0; { // callseq 970, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd818; .param .b64 param1; st.param.b64 [param1+0], %rd827; .param .b32 param2; st.param.b32 [param2+0], %r4; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r119, [retval0+0]; } // callseq 970 cvt.u16.u32 %rs94, %r119; $L__BB205_26: setp.ne.s16 %p37, %rs94, 0; @%p37 bra $L__BB205_28; mov.u32 %r454, -1; st.local.u32 [%rd24+-4], %r454; $L__BB205_28: ld.global.u32 %r11, [%rd16]; ld.const.u32 %r12, [kernel_params+1592]; setp.eq.s32 %p38, %r12, 0; @%p38 bra $L__BB205_35; and.b32 %r122, %r11, 1; setp.eq.b32 %p40, %r122, 1; setp.eq.s32 %p41, %r6, -1; mov.u16 %rs95, 0; mov.pred %p269, 0; or.pred %p42, %p41, %p40; @%p42 bra $L__BB205_31; cvt.s64.s32 %rd824, %r6; ld.const.u64 %rd179, [kernel_params+2184]; cvta.to.global.u64 %rd180, %rd179; shl.b64 %rd181, %rd824, 2; add.s64 %rd182, %rd180, %rd181; ld.global.u32 %r123, [%rd182]; bfe.u32 %r124, %r123, 10, 1; cvt.u16.u32 %rs95, %r124; shr.u32 %r125, %r123, 9; and.b32 %r126, %r125, 1; setp.eq.b32 %p269, %r126, 1; $L__BB205_31: ld.const.u64 %rd183, [kernel_params+2552]; cvta.to.global.u64 %rd184, %rd183; add.s64 %rd25, %rd184, %rd3; ld.global.u8 %rs16, [%rd25]; and.b16 %rs42, %rs16, 2; setp.eq.s16 %p43, %rs42, 0; and.b16 %rs96, %rs16, 251; st.global.u8 [%rd25], %rs96; not.pred %p44, %p269; or.pred %p45, %p44, %p43; @%p45 bra $L__BB205_33; or.b16 %rs96, %rs16, 4; st.global.u8 [%rd25], %rs96; $L__BB205_33: setp.eq.s16 %p46, %rs95, 0; @%p46 bra $L__BB205_35; ld.const.u64 %rd835, [kernel_params+2552]; cvta.to.global.u64 %rd834, %rd835; add.s64 %rd833, %rd834, %rd3; or.b16 %rs43, %rs96, 2; st.global.u8 [%rd833], %rs43; $L__BB205_35: ld.const.u32 %r127, [kernel_params+1588]; setp.eq.s32 %p47, %r127, 0; @%p47 bra $L__BB205_74; ld.const.u64 %rd26, [kernel_params+2728]; ld.const.u32 %r13, [kernel_params+1496]; setp.gt.s32 %p48, %r13, 0; @%p48 bra $L__BB205_38; bra.uni $L__BB205_37; $L__BB205_38: cvta.to.global.u64 %rd185, %rd26; shl.b64 %rd186, %rd3, 2; add.s64 %rd187, %rd185, %rd186; ld.global.u32 %r15, [%rd187]; ld.const.u64 %rd188, [kernel_params+2368]; cvta.to.global.u64 %rd27, %rd188; and.b32 %r16, %r11, 1; and.b32 %r17, %r11, 1073741824; ld.const.u64 %rd189, [kernel_params+2552]; cvta.to.global.u64 %rd190, %rd189; add.s64 %rd28, %rd190, %rd3; ld.local.f32 %f327, [%rd24+-16]; ld.local.f32 %f326, [%rd24+-12]; ld.local.f32 %f325, [%rd24+-8]; ld.local.u32 %r462, [%rd24+4]; ld.local.u32 %r36, [%rd24+-4]; ld.local.u32 %r460, [%rd24]; mov.u32 %r459, 0; setp.eq.s32 %p49, %r16, 0; setp.ne.s32 %p54, %r17, 0; mov.pred %p91, -1; setp.eq.s32 %p102, %r6, -1; $L__BB205_39: mul.wide.s32 %rd191, %r459, 192; add.s64 %rd192, %rd27, %rd191; add.s64 %rd30, %rd192, 16; ld.global.u32 %r25, [%rd192+16]; @%p49 bra $L__BB205_41; bra.uni $L__BB205_40; $L__BB205_41: and.b32 %r130, %r25, 268435456; setp.eq.s32 %p51, %r130, 0; @%p51 bra $L__BB205_71; ld.global.u8 %rs44, [%rd28]; and.b16 %rs45, %rs44, 4; setp.eq.s16 %p52, %rs45, 0; @%p52 bra $L__BB205_44; ld.global.u32 %r131, [%rd30+24]; setp.ne.s32 %p53, %r131, 0; @%p53 bra $L__BB205_71; bra.uni $L__BB205_44; $L__BB205_40: and.b32 %r129, %r25, 16777216; setp.eq.s32 %p50, %r129, 0; @%p50 bra $L__BB205_44; bra.uni $L__BB205_71; $L__BB205_44: shr.u32 %r132, %r25, 22; and.b32 %r133, %r132, 1; setp.eq.b32 %p55, %r133, 1; and.pred %p56, %p54, %p55; @%p56 bra $L__BB205_71; ld.global.u32 %r26, [%rd30+-16]; setp.eq.s32 %p57, %r26, 4; @%p57 bra $L__BB205_62; bra.uni $L__BB205_46; $L__BB205_62: ld.global.f32 %f46, [%rd30+-12]; ld.global.f32 %f47, [%rd30+-8]; ld.global.f32 %f48, [%rd30+-4]; ld.global.f32 %f49, [%rd30+128]; setp.eq.ftz.f32 %p89, %f49, 0f00000000; @%p89 bra $L__BB205_71; sub.ftz.f32 %f186, %f1, %f46; sub.ftz.f32 %f50, %f2, %f47; mul.ftz.f32 %f187, %f50, %f50; fma.rn.ftz.f32 %f188, %f186, %f186, %f187; sub.ftz.f32 %f51, %f3, %f48; fma.rn.ftz.f32 %f189, %f51, %f51, %f188; sqrt.approx.ftz.f32 %f190, %f189; div.approx.ftz.f32 %f52, %f186, %f190; div.approx.ftz.f32 %f53, %f50, %f190; div.approx.ftz.f32 %f54, %f51, %f190; mul.ftz.f32 %f55, %f52, %f4; mul.ftz.f32 %f56, %f53, %f5; add.ftz.f32 %f191, %f55, %f56; mul.ftz.f32 %f57, %f54, %f6; add.ftz.f32 %f192, %f191, %f57; setp.ge.ftz.f32 %p90, %f192, 0f00000000; @%p90 bra $L__BB205_71; mul.ftz.f32 %f195, %f53, %f50; fma.rn.ftz.f32 %f196, %f52, %f186, %f195; fma.rn.ftz.f32 %f58, %f54, %f51, %f196; neg.ftz.f32 %f197, %f56; sub.ftz.f32 %f198, %f197, %f55; sub.ftz.f32 %f59, %f198, %f57; mul.ftz.f32 %f199, %f58, %f59; setp.leu.ftz.f32 %p92, %f199, 0f00000000; mov.f32 %f319, 0f00000000; mov.pred %p272, %p91; @%p92 bra $L__BB205_68; mov.f32 %f319, 0f00000000; div.approx.ftz.f32 %f60, %f58, %f59; setp.lt.ftz.f32 %p94, %f60, 0f00000000; mov.pred %p272, %p91; @%p94 bra $L__BB205_68; mov.f32 %f319, 0f00000000; fma.rn.ftz.f32 %f202, %f4, %f60, %f1; fma.rn.ftz.f32 %f203, %f5, %f60, %f2; fma.rn.ftz.f32 %f204, %f6, %f60, %f3; sub.ftz.f32 %f205, %f202, %f46; sub.ftz.f32 %f206, %f203, %f47; sub.ftz.f32 %f207, %f204, %f48; mul.ftz.f32 %f208, %f206, %f206; fma.rn.ftz.f32 %f209, %f205, %f205, %f208; fma.rn.ftz.f32 %f210, %f207, %f207, %f209; mul.ftz.f32 %f211, %f49, %f49; setp.geu.ftz.f32 %p96, %f210, %f211; mov.pred %p272, %p91; @%p96 bra $L__BB205_68; setp.gt.ftz.f32 %p97, %f60, %f7; setp.lt.ftz.f32 %p98, %f60, %f313; and.pred %p99, %p97, %p98; selp.f32 %f319, %f60, 0f00000000, %p99; not.pred %p272, %p99; $L__BB205_68: mov.f32 %f317, 0f00000000; mov.f32 %f318, %f317; @%p272 bra $L__BB205_71; bra.uni $L__BB205_69; $L__BB205_46: setp.eq.s32 %p58, %r26, 0; @%p58 bra $L__BB205_56; setp.ne.s32 %p59, %r26, 3; @%p59 bra $L__BB205_71; ld.global.f32 %f18, [%rd30+140]; abs.ftz.f32 %f94, %f18; setp.eq.ftz.f32 %p60, %f94, 0f00000000; @%p60 bra $L__BB205_71; add.s64 %rd31, %rd30, 128; ld.global.v4.f32 {%f95, %f96, %f97, %f98}, [%rd30+160]; mul.ftz.f32 %f99, %f96, %f5; fma.rn.ftz.f32 %f100, %f95, %f4, %f99; fma.rn.ftz.f32 %f22, %f97, %f6, %f100; setp.ge.ftz.f32 %p61, %f22, 0f00000000; @%p61 bra $L__BB205_71; mul.ftz.f32 %f104, %f96, %f2; fma.rn.ftz.f32 %f105, %f95, %f1, %f104; fma.rn.ftz.f32 %f106, %f97, %f3, %f105; ld.global.f32 %f23, [%rd30+-12]; ld.global.f32 %f24, [%rd30+-8]; mul.ftz.f32 %f107, %f96, %f24; fma.rn.ftz.f32 %f108, %f95, %f23, %f107; ld.global.f32 %f25, [%rd30+-4]; fma.rn.ftz.f32 %f109, %f97, %f25, %f108; sub.ftz.f32 %f110, %f106, %f109; neg.ftz.f32 %f111, %f110; div.approx.ftz.f32 %f26, %f111, %f22; setp.leu.ftz.f32 %p63, %f26, %f7; setp.geu.ftz.f32 %p64, %f26, %f313; or.pred %p65, %p64, %p63; mov.pred %p270, -1; mov.f32 %f317, 0f00000000; mov.f32 %f318, %f317; mov.f32 %f319, %f317; @%p65 bra $L__BB205_55; mov.pred %p270, -1; mov.f32 %f317, 0f00000000; ld.global.v4.f32 {%f115, %f116, %f117, %f118}, [%rd31]; fma.rn.ftz.f32 %f122, %f4, %f26, %f1; fma.rn.ftz.f32 %f123, %f5, %f26, %f2; fma.rn.ftz.f32 %f124, %f6, %f26, %f3; sub.ftz.f32 %f27, %f122, %f23; sub.ftz.f32 %f28, %f123, %f24; sub.ftz.f32 %f29, %f124, %f25; mul.ftz.f32 %f125, %f116, %f28; fma.rn.ftz.f32 %f126, %f115, %f27, %f125; fma.rn.ftz.f32 %f127, %f117, %f29, %f126; mul.ftz.f32 %f128, %f116, %f116; fma.rn.ftz.f32 %f129, %f115, %f115, %f128; fma.rn.ftz.f32 %f130, %f117, %f117, %f129; div.approx.ftz.f32 %f30, %f127, %f130; setp.lt.ftz.f32 %p67, %f30, 0fBF000000; setp.gt.ftz.f32 %p68, %f30, 0f3F000000; or.pred %p69, %p67, %p68; mov.f32 %f318, %f317; mov.f32 %f319, %f317; @%p69 bra $L__BB205_55; mov.pred %p270, -1; mov.f32 %f317, 0f00000000; ld.global.v4.f32 {%f134, %f135, %f136, %f137}, [%rd31+16]; mul.ftz.f32 %f141, %f135, %f28; fma.rn.ftz.f32 %f142, %f134, %f27, %f141; fma.rn.ftz.f32 %f143, %f136, %f29, %f142; mul.ftz.f32 %f144, %f135, %f135; fma.rn.ftz.f32 %f145, %f134, %f134, %f144; fma.rn.ftz.f32 %f146, %f136, %f136, %f145; div.approx.ftz.f32 %f31, %f143, %f146; setp.lt.ftz.f32 %p71, %f31, 0fBF000000; setp.gt.ftz.f32 %p72, %f31, 0f3F000000; or.pred %p73, %p71, %p72; mov.f32 %f318, %f317; mov.f32 %f319, %f317; @%p73 bra $L__BB205_55; mov.pred %p270, -1; mov.f32 %f317, 0f00000000; setp.lt.ftz.f32 %p75, %f18, 0f00000000; mul.ftz.f32 %f150, %f31, %f31; fma.rn.ftz.f32 %f151, %f30, %f30, %f150; setp.gt.ftz.f32 %p76, %f151, 0f3E800000; and.pred %p77, %p75, %p76; mov.f32 %f318, %f317; mov.f32 %f319, %f317; @%p77 bra $L__BB205_55; add.ftz.f32 %f318, %f31, 0f3F000000; neg.ftz.f32 %f152, %f30; sub.ftz.f32 %f317, %f152, %f31; mov.pred %p270, 0; mov.f32 %f319, %f26; $L__BB205_55: @%p270 bra $L__BB205_71; bra.uni $L__BB205_69; $L__BB205_56: ld.global.f32 %f37, [%rd30+-12]; ld.global.f32 %f38, [%rd30+-8]; ld.global.f32 %f39, [%rd30+-4]; ld.global.f32 %f40, [%rd30+128]; setp.eq.ftz.f32 %p79, %f40, 0f00000000; @%p79 bra $L__BB205_71; sub.ftz.f32 %f154, %f1, %f37; sub.ftz.f32 %f155, %f2, %f38; mul.ftz.f32 %f156, %f155, %f155; fma.rn.ftz.f32 %f157, %f154, %f154, %f156; sub.ftz.f32 %f158, %f3, %f39; fma.rn.ftz.f32 %f159, %f158, %f158, %f157; sqrt.approx.ftz.f32 %f160, %f159; div.approx.ftz.f32 %f161, %f154, %f160; div.approx.ftz.f32 %f162, %f155, %f160; div.approx.ftz.f32 %f163, %f158, %f160; mul.ftz.f32 %f164, %f162, %f155; fma.rn.ftz.f32 %f165, %f161, %f154, %f164; fma.rn.ftz.f32 %f41, %f163, %f158, %f165; mul.ftz.f32 %f166, %f161, %f4; mul.ftz.f32 %f167, %f162, %f5; neg.ftz.f32 %f168, %f167; sub.ftz.f32 %f169, %f168, %f166; mul.ftz.f32 %f170, %f163, %f6; sub.ftz.f32 %f42, %f169, %f170; mul.ftz.f32 %f171, %f41, %f42; setp.leu.ftz.f32 %p81, %f171, 0f00000000; mov.pred %p271, -1; mov.f32 %f319, 0f00000000; @%p81 bra $L__BB205_61; mov.pred %p271, -1; mov.f32 %f319, 0f00000000; div.approx.ftz.f32 %f43, %f41, %f42; setp.lt.ftz.f32 %p83, %f43, 0f00000000; @%p83 bra $L__BB205_61; mov.pred %p271, -1; mov.f32 %f319, 0f00000000; fma.rn.ftz.f32 %f174, %f4, %f43, %f1; fma.rn.ftz.f32 %f175, %f5, %f43, %f2; fma.rn.ftz.f32 %f176, %f6, %f43, %f3; sub.ftz.f32 %f177, %f174, %f37; sub.ftz.f32 %f178, %f175, %f38; sub.ftz.f32 %f179, %f176, %f39; mul.ftz.f32 %f180, %f178, %f178; fma.rn.ftz.f32 %f181, %f177, %f177, %f180; fma.rn.ftz.f32 %f182, %f179, %f179, %f181; mul.ftz.f32 %f183, %f40, %f40; setp.geu.ftz.f32 %p85, %f182, %f183; @%p85 bra $L__BB205_61; setp.gt.ftz.f32 %p86, %f43, %f7; setp.lt.ftz.f32 %p87, %f43, %f313; and.pred %p88, %p86, %p87; selp.f32 %f319, %f43, 0f00000000, %p88; not.pred %p271, %p88; $L__BB205_61: mov.f32 %f317, 0f00000000; mov.f32 %f318, %f317; @%p271 bra $L__BB205_71; $L__BB205_69: setp.geu.ftz.f32 %p100, %f319, %f327; @%p100 bra $L__BB205_71; cvt.s64.s32 %rd796, %r459; cvt.u32.u64 %r134, %rd796; setp.eq.s32 %p101, %r134, %r5; and.pred %p103, %p102, %p101; setp.eq.s32 %p104, %r15, 32; and.pred %p105, %p104, %p103; selp.b32 %r460, %r460, -1, %p105; selp.b32 %r36, %r36, %r134, %p105; selp.b32 %r462, %r462, 32, %p105; selp.f32 %f325, %f325, %f317, %p105; selp.f32 %f326, %f326, %f318, %p105; selp.f32 %f327, %f327, %f319, %p105; $L__BB205_71: cvt.s64.s32 %rd797, %r459; cvt.u32.u64 %r135, %rd797; add.s32 %r459, %r135, 1; setp.lt.s32 %p106, %r459, %r13; @%p106 bra $L__BB205_39; add.u64 %rd789, %SPL, 60; st.local.f32 [%rd789], %f327; st.local.f32 [%rd789+4], %f326; st.local.f32 [%rd789+8], %f325; st.local.u32 [%rd789+20], %r462; st.local.u32 [%rd789+12], %r36; st.local.u32 [%rd789+16], %r460; bra.uni $L__BB205_73; $L__BB205_74: ld.local.u32 %r36, [%rd24+-4]; ld.const.u64 %rd33, [kernel_params+2728]; bra.uni $L__BB205_75; $L__BB205_37: ld.local.u32 %r36, [%rd24+-4]; $L__BB205_73: ld.const.u64 %rd33, [kernel_params+2728]; setp.ne.s32 %p107, %r36, -1; selp.u16 %rs46, 1, 0, %p107; or.b16 %rs47, %rs94, %rs46; setp.ne.s16 %p108, %rs47, 0; selp.u16 %rs94, 1, 0, %p108; $L__BB205_75: ld.const.u64 %rd832, [kernel_params+2712]; mul.wide.s32 %rd831, %r3, 4; cvta.to.global.u64 %rd830, %rd832; add.s64 %rd829, %rd830, %rd831; ld.const.u64 %rd823, [kernel_params+2720]; mul.wide.s32 %rd822, %r3, 4; cvta.to.global.u64 %rd821, %rd823; add.s64 %rd820, %rd821, %rd822; add.u64 %rd791, %SPL, 60; ld.local.f32 %f214, [%rd791]; ld.const.v2.u64 {%rd197, %rd198}, [kernel_params+2688]; cvta.to.global.u64 %rd36, %rd197; shl.b64 %rd199, %rd3, 2; add.s64 %rd37, %rd36, %rd199; st.global.f32 [%rd37], %f214; ld.local.f32 %f215, [%rd791+4]; cvta.to.global.u64 %rd38, %rd198; add.s64 %rd39, %rd38, %rd199; st.global.f32 [%rd39], %f215; ld.local.f32 %f216, [%rd791+8]; ld.const.u64 %rd40, [kernel_params+2704]; cvta.to.global.u64 %rd41, %rd40; add.s64 %rd42, %rd41, %rd199; st.global.f32 [%rd42], %f216; ld.local.u32 %r37, [%rd791+16]; st.global.u32 [%rd820], %r37; st.global.u32 [%rd829], %r36; ld.local.u32 %r38, [%rd791+20]; cvta.to.global.u64 %rd43, %rd33; add.s64 %rd44, %rd43, %rd199; st.global.u32 [%rd44], %r38; ld.const.u32 %r136, [kernel_params]; and.b32 %r39, %r136, 131072; setp.eq.s32 %p109, %r39, 0; mov.u16 %rs98, 1; @%p109 bra $L__BB205_77; ld.const.u64 %rd200, [kernel_params+2776]; cvta.to.global.u64 %rd201, %rd200; add.s64 %rd203, %rd201, %rd199; ld.global.u32 %r137, [%rd203]; setp.eq.s32 %p110, %r137, -1; selp.u16 %rs98, 1, 0, %p110; $L__BB205_77: setp.eq.s16 %p111, %rs98, 0; @%p111 bra $L__BB205_262; setp.ne.s16 %p112, %rs94, 0; @%p112 bra $L__BB205_80; bra.uni $L__BB205_79; $L__BB205_80: and.b32 %r140, %r38, 32; setp.eq.s32 %p113, %r140, 0; @%p113 bra $L__BB205_82; bra.uni $L__BB205_81; $L__BB205_82: and.b32 %r40, %r38, 1; setp.eq.s32 %p114, %r40, 0; @%p114 bra $L__BB205_84; bra.uni $L__BB205_83; $L__BB205_84: and.b32 %r143, %r38, 8; setp.eq.s32 %p115, %r143, 0; @%p115 bra $L__BB205_86; bra.uni $L__BB205_85; $L__BB205_86: and.b32 %r145, %r38, 6; setp.eq.s32 %p116, %r145, 0; mov.u32 %r465, 0; @%p116 bra $L__BB205_88; ld.const.u64 %rd228, [kernel_params+2264]; cvta.to.global.u64 %rd229, %rd228; mul.wide.s32 %rd230, %r36, 16; add.s64 %rd231, %rd229, %rd230; ld.global.u32 %r465, [%rd231]; bra.uni $L__BB205_88; $L__BB205_262: setp.ne.s16 %p223, %rs94, 0; and.b32 %r324, %r38, 32; setp.eq.s32 %p224, %r324, 0; and.pred %p8, %p223, %p224; mov.u64 %rd841, -1; not.pred %p225, %p8; @%p225 bra $L__BB205_270; and.b32 %r325, %r38, 1; setp.eq.b32 %p226, %r325, 1; mov.pred %p227, 0; xor.pred %p228, %p226, %p227; not.pred %p229, %p228; @%p229 bra $L__BB205_265; bra.uni $L__BB205_264; $L__BB205_265: and.b32 %r326, %r38, 8; setp.eq.s32 %p230, %r326, 0; @%p230 bra $L__BB205_267; bra.uni $L__BB205_266; $L__BB205_267: and.b32 %r328, %r38, 6; setp.eq.s32 %p231, %r328, 0; mov.u32 %r474, 0; @%p231 bra $L__BB205_269; ld.const.u64 %rd659, [kernel_params+2264]; cvta.to.global.u64 %rd660, %rd659; mul.wide.s32 %rd661, %r36, 16; add.s64 %rd662, %rd660, %rd661; ld.global.u32 %r474, [%rd662]; bra.uni $L__BB205_269; $L__BB205_79: ld.const.u64 %rd204, [kernel_params+4256]; cvta.to.global.u64 %rd205, %rd204; add.s64 %rd206, %rd205, 8; atom.global.add.u32 %r138, [%rd206], -1; add.s64 %rd207, %rd205, 24; atom.global.add.u32 %r139, [%rd207], 1; ld.const.u64 %rd208, [kernel_params+2520]; cvta.to.global.u64 %rd209, %rd208; shl.b64 %rd210, %rd3, 1; add.s64 %rd211, %rd209, %rd210; mov.u16 %rs49, 6; st.global.u16 [%rd211], %rs49; bra.uni $L__BB205_297; $L__BB205_81: ld.const.u64 %rd212, [kernel_params+4256]; cvta.to.global.u64 %rd213, %rd212; add.s64 %rd214, %rd213, 8; atom.global.add.u32 %r141, [%rd214], -1; add.s64 %rd215, %rd213, 28; atom.global.add.u32 %r142, [%rd215], 1; ld.const.u64 %rd216, [kernel_params+2520]; cvta.to.global.u64 %rd217, %rd216; shl.b64 %rd218, %rd3, 1; add.s64 %rd219, %rd217, %rd218; mov.u16 %rs50, 7; st.global.u16 [%rd219], %rs50; bra.uni $L__BB205_297; $L__BB205_83: ld.const.u64 %rd220, [kernel_params+2216]; cvta.to.global.u64 %rd221, %rd220; mul.wide.s32 %rd222, %r36, 4; add.s64 %rd223, %rd221, %rd222; ld.global.u32 %r465, [%rd223]; bra.uni $L__BB205_88; $L__BB205_264: ld.const.u64 %rd651, [kernel_params+2216]; cvta.to.global.u64 %rd652, %rd651; mul.wide.s32 %rd653, %r36, 4; add.s64 %rd654, %rd652, %rd653; ld.global.u32 %r474, [%rd654]; bra.uni $L__BB205_269; $L__BB205_85: ld.const.u64 %rd224, [kernel_params+2304]; cvta.to.global.u64 %rd225, %rd224; mul.wide.s32 %rd226, %r36, 4; add.s64 %rd227, %rd225, %rd226; ld.global.u32 %r465, [%rd227]; $L__BB205_88: ld.const.u32 %r451, [kernel_params+1536]; setp.eq.s32 %p264, %r451, 0; ld.const.u64 %rd232, [kernel_params+2408]; cvta.to.global.u64 %rd45, %rd232; and.b32 %r45, %r465, 4194303; shl.b32 %r146, %r465, 5; cvt.u64.u32 %rd233, %r146; and.b64 %rd234, %rd233, 134217696; add.s64 %rd235, %rd45, %rd234; ld.global.u32 %r46, [%rd235+16]; mov.u16 %rs99, 0; @%p264 bra $L__BB205_90; ld.const.u32 %r449, [kernel_params+1536]; ld.const.v2.u64 {%rd236, %rd237}, [kernel_params+2464]; cvta.to.global.u64 %rd239, %rd236; shl.b64 %rd240, %rd3, 1; add.s64 %rd241, %rd239, %rd240; ld.global.u16 %r147, [%rd241]; ld.const.u64 %rd242, [kernel_params+2496]; cvta.to.global.u64 %rd243, %rd242; add.s64 %rd244, %rd243, %rd240; ld.global.u16 %r148, [%rd244]; sub.s32 %r149, %r147, %r148; ld.const.v2.u64 {%rd245, %rd246}, [kernel_params+2480]; cvta.to.global.u64 %rd248, %rd246; add.s64 %rd249, %rd248, %rd240; ld.global.u16 %rs52, [%rd249]; setp.ne.s16 %p118, %rs52, 0; selp.b32 %r150, -1, 0, %p118; add.s32 %r151, %r149, %r150; setp.ge.s32 %p119, %r151, %r449; selp.u16 %rs99, 1, 0, %p119; $L__BB205_90: setp.eq.s16 %p120, %rs99, 0; @%p120 bra $L__BB205_97; and.b32 %r152, %r46, 537001984; setp.eq.s32 %p121, %r152, 0; @%p121 bra $L__BB205_93; ld.global.u32 %r153, [%rd16]; or.b32 %r466, %r153, 131072; st.global.u32 [%rd16], %r466; bra.uni $L__BB205_98; $L__BB205_97: ld.global.u32 %r466, [%rd16]; $L__BB205_98: ld.const.v2.u64 {%rd254, %rd255}, [kernel_params+2528]; cvta.to.global.u64 %rd48, %rd254; add.s64 %rd49, %rd48, %rd199; ld.global.u32 %r51, [%rd49]; cvta.to.global.u64 %rd50, %rd255; shl.b64 %rd257, %rd3, 1; add.s64 %rd51, %rd50, %rd257; ld.global.u16 %r52, [%rd51]; ld.const.u64 %rd52, [kernel_params+2456]; cvta.to.global.u64 %rd53, %rd52; add.s64 %rd54, %rd53, %rd199; ld.global.u32 %r53, [%rd54]; and.b32 %r156, %r466, 64; setp.eq.s32 %p125, %r156, 0; @%p125 bra $L__BB205_100; bra.uni $L__BB205_99; $L__BB205_100: ld.const.u64 %rd262, [kernel_params+2464]; cvta.to.global.u64 %rd263, %rd262; add.s64 %rd265, %rd263, %rd257; ld.global.u16 %r159, [%rd265]; ld.const.u32 %r160, [kernel_params+1512]; setp.ge.u32 %p127, %r160, %r159; mov.f32 %f328, 0f3F800000; @%p127 bra $L__BB205_102; bra.uni $L__BB205_101; $L__BB205_99: ld.const.u64 %rd258, [kernel_params+2472]; cvta.to.global.u64 %rd259, %rd258; add.s64 %rd261, %rd259, %rd257; ld.global.u16 %r157, [%rd261]; ld.const.u32 %r158, [kernel_params+1552]; setp.lt.u32 %p126, %r158, %r157; mov.f32 %f328, 0f3F800000; @%p126 bra $L__BB205_101; bra.uni $L__BB205_102; $L__BB205_101: ld.const.u64 %rd266, [kernel_params+2584]; cvta.to.global.u64 %rd267, %rd266; mul.lo.s64 %rd268, %rd3, 12; add.s64 %rd269, %rd267, %rd268; ld.global.f32 %f219, [%rd269]; abs.ftz.f32 %f220, %f219; ld.global.f32 %f221, [%rd269+4]; abs.ftz.f32 %f222, %f221; ld.global.f32 %f223, [%rd269+8]; abs.ftz.f32 %f224, %f223; max.ftz.f32 %f225, %f220, %f222; max.ftz.f32 %f226, %f225, %f224; sqrt.approx.ftz.f32 %f227, %f226; mov.f32 %f228, 0f3F800000; min.ftz.f32 %f328, %f227, %f228; $L__BB205_102: ld.const.u64 %rd55, [kernel_params+2576]; cvta.to.global.u64 %rd56, %rd55; add.s64 %rd57, %rd56, %rd199; st.global.f32 [%rd57], %f328; setp.eq.ftz.f32 %p128, %f328, 0f3F800000; @%p128 bra $L__BB205_115; ld.const.u32 %r161, [kernel_params+1596]; setp.eq.s32 %p129, %r161, 0; brev.b32 %r162, %r53; mul.lo.s32 %r163, %r162, 1025551850; xor.b32 %r54, %r163, %r162; @%p129 bra $L__BB205_106; ld.const.v2.u32 {%r164, %r165}, [kernel_params+1600]; mov.b32 %f74, %r165; setp.geu.ftz.f32 %p130, %f74, 0f3F800000; setp.lt.ftz.f32 %p131, %f74, 0f3F800000; ld.const.u32 %r168, [kernel_params+1576]; selp.b32 %r169, %r168, %r51, %p131; and.b32 %r170, %r52, 255; xor.b32 %r171, %r169, %r170; mul.lo.s32 %r172, %r171, -512718531; shr.u32 %r173, %r169, 16; xor.b32 %r174, %r172, %r173; bfe.u32 %r175, %r174, 4, 4; shr.u32 %r176, %r169, 8; xor.b32 %r177, %r174, %r176; xor.b32 %r178, %r177, %r175; mul.lo.s32 %r179, %r178, 153742143; shr.u32 %r180, %r169, 23; xor.b32 %r181, %r179, %r180; bfe.u32 %r182, %r181, 1, 7; xor.b32 %r183, %r182, %r181; shr.u32 %r184, %r169, 27; or.b32 %r185, %r184, 1; mul.lo.s32 %r186, %r185, %r183; mul.lo.s32 %r187, %r186, -1074439621; bfe.u32 %r188, %r187, 2, 6; xor.b32 %r189, %r188, %r187; mul.lo.s32 %r190, %r189, -1638916925; bfe.u32 %r191, %r190, 2, 6; xor.b32 %r192, %r191, %r190; mul.lo.s32 %r193, %r192, -933190689; and.b32 %r194, %r193, 255; bfe.u32 %r195, %r193, 5, 3; xor.b32 %r196, %r195, %r194; add.s32 %r197, %r164, -1; xor.b32 %r55, %r52, 61; xor.b32 %r198, %r169, %r55; mul.lo.s32 %r199, %r198, 9; shr.u32 %r200, %r199, 4; xor.b32 %r201, %r200, %r199; mul.lo.s32 %r202, %r201, 668265261; add.s32 %r203, %r202, %r54; shr.u32 %r204, %r202, 16; or.b32 %r205, %r204, 1; mul.lo.s32 %r206, %r205, %r203; mul.lo.s32 %r207, %r206, 89287766; xor.b32 %r208, %r207, %r206; mul.lo.s32 %r209, %r208, 1403136100; xor.b32 %r210, %r209, %r208; brev.b32 %r211, %r210; neg.s32 %r212, %r164; and.b32 %r213, %r53, %r212; and.b32 %r214, %r211, %r197; or.b32 %r215, %r214, %r213; mad.lo.s32 %r216, %r196, %r164, %r215; shl.b32 %r217, %r164, 8; rem.u32 %r218, %r216, %r217; ld.const.u64 %rd271, [kernel_params+2424]; cvta.to.global.u64 %rd272, %rd271; shl.b32 %r219, %r218, 1; mul.wide.u32 %rd273, %r219, 4; add.s64 %rd274, %rd272, %rd273; ld.global.f32 %f329, [%rd274]; @%p130 bra $L__BB205_107; xor.b32 %r220, %r55, %r51; mul.lo.s32 %r221, %r220, 9; shr.u32 %r222, %r221, 4; xor.b32 %r223, %r222, %r221; mul.lo.s32 %r224, %r223, 668265261; cvt.rn.f32.u32 %f229, %r224; mul.ftz.f32 %f230, %f229, 0f2F7FFFFE; fma.rn.ftz.f32 %f231, %f230, %f74, %f329; cvt.rmi.ftz.f32.f32 %f232, %f231; sub.ftz.f32 %f329, %f231, %f232; bra.uni $L__BB205_107; $L__BB205_266: ld.const.u64 %rd655, [kernel_params+2304]; cvta.to.global.u64 %rd656, %rd655; mul.wide.s32 %rd657, %r36, 4; add.s64 %rd658, %rd656, %rd657; ld.global.u32 %r474, [%rd658]; $L__BB205_269: and.b32 %r329, %r474, 4194303; cvt.u64.u32 %rd841, %r329; $L__BB205_270: mov.u32 %r475, 0; @%p225 bra $L__BB205_272; ld.const.u64 %rd663, [kernel_params+2408]; cvta.to.global.u64 %rd664, %rd663; shl.b64 %rd665, %rd841, 5; add.s64 %rd666, %rd664, %rd665; ld.global.u32 %r475, [%rd666+16]; $L__BB205_272: ld.const.u32 %r452, [kernel_params+1536]; setp.eq.s32 %p265, %r452, 0; @%p265 bra $L__BB205_279; ld.const.u32 %r450, [kernel_params+1536]; ld.const.v2.u64 {%rd667, %rd668}, [kernel_params+2464]; cvta.to.global.u64 %rd670, %rd667; shl.b64 %rd671, %rd3, 1; add.s64 %rd672, %rd670, %rd671; ld.global.u16 %r331, [%rd672]; ld.const.u64 %rd673, [kernel_params+2496]; cvta.to.global.u64 %rd674, %rd673; add.s64 %rd675, %rd674, %rd671; ld.global.u16 %r332, [%rd675]; sub.s32 %r333, %r331, %r332; ld.const.v2.u64 {%rd676, %rd677}, [kernel_params+2480]; cvta.to.global.u64 %rd679, %rd677; add.s64 %rd680, %rd679, %rd671; ld.global.u16 %rs82, [%rd680]; setp.ne.s16 %p234, %rs82, 0; selp.b32 %r334, -1, 0, %p234; add.s32 %r335, %r333, %r334; setp.lt.s32 %p235, %r335, %r450; @%p235 bra $L__BB205_279; and.b32 %r336, %r475, 537001984; setp.eq.s32 %p236, %r336, 0; @%p236 bra $L__BB205_276; ld.global.u32 %r337, [%rd16]; or.b32 %r338, %r337, 131072; st.global.u32 [%rd16], %r338; $L__BB205_279: ld.const.v2.u64 {%rd685, %rd686}, [kernel_params+2528]; cvta.to.global.u64 %rd689, %rd685; add.s64 %rd691, %rd689, %rd199; ld.global.u32 %r88, [%rd691]; cvta.to.global.u64 %rd692, %rd686; shl.b64 %rd693, %rd3, 1; add.s64 %rd694, %rd692, %rd693; ld.global.u16 %r89, [%rd694]; ld.const.u64 %rd695, [kernel_params+2456]; cvta.to.global.u64 %rd696, %rd695; add.s64 %rd697, %rd696, %rd199; ld.global.u32 %r90, [%rd697]; ld.global.u8 %rs83, [%rd16]; and.b16 %rs84, %rs83, 64; setp.eq.s16 %p241, %rs84, 0; @%p241 bra $L__BB205_281; bra.uni $L__BB205_280; $L__BB205_281: ld.const.u64 %rd702, [kernel_params+2464]; cvta.to.global.u64 %rd703, %rd702; add.s64 %rd705, %rd703, %rd693; ld.global.u16 %r344, [%rd705]; ld.const.u32 %r345, [kernel_params+1512]; setp.ge.u32 %p243, %r345, %r344; mov.f32 %f330, 0f3F800000; @%p243 bra $L__BB205_283; bra.uni $L__BB205_282; $L__BB205_280: ld.const.u64 %rd698, [kernel_params+2472]; cvta.to.global.u64 %rd699, %rd698; add.s64 %rd701, %rd699, %rd693; ld.global.u16 %r342, [%rd701]; ld.const.u32 %r343, [kernel_params+1552]; setp.lt.u32 %p242, %r343, %r342; mov.f32 %f330, 0f3F800000; @%p242 bra $L__BB205_282; bra.uni $L__BB205_283; $L__BB205_282: ld.const.u64 %rd706, [kernel_params+2584]; cvta.to.global.u64 %rd707, %rd706; mul.lo.s64 %rd708, %rd3, 12; add.s64 %rd709, %rd707, %rd708; ld.global.f32 %f289, [%rd709]; abs.ftz.f32 %f290, %f289; ld.global.f32 %f291, [%rd709+4]; abs.ftz.f32 %f292, %f291; ld.global.f32 %f293, [%rd709+8]; abs.ftz.f32 %f294, %f293; max.ftz.f32 %f295, %f290, %f292; max.ftz.f32 %f296, %f295, %f294; sqrt.approx.ftz.f32 %f297, %f296; mov.f32 %f298, 0f3F800000; min.ftz.f32 %f330, %f297, %f298; $L__BB205_283: ld.const.u64 %rd710, [kernel_params+2576]; cvta.to.global.u64 %rd711, %rd710; add.s64 %rd713, %rd711, %rd199; st.global.f32 [%rd713], %f330; setp.eq.ftz.f32 %p245, %f330, 0f3F800000; mov.pred %p244, -1; mov.pred %p273, %p244; @%p245 bra $L__BB205_294; ld.const.u32 %r346, [kernel_params+1596]; setp.eq.s32 %p246, %r346, 0; brev.b32 %r347, %r90; mul.lo.s32 %r348, %r347, 1025551850; xor.b32 %r91, %r348, %r347; @%p246 bra $L__BB205_287; ld.const.v2.u32 {%r349, %r350}, [kernel_params+1600]; mov.b32 %f81, %r350; setp.geu.ftz.f32 %p247, %f81, 0f3F800000; setp.lt.ftz.f32 %p248, %f81, 0f3F800000; ld.const.u32 %r353, [kernel_params+1576]; selp.b32 %r354, %r353, %r88, %p248; and.b32 %r355, %r89, 255; xor.b32 %r356, %r354, %r355; mul.lo.s32 %r357, %r356, -512718531; shr.u32 %r358, %r354, 16; xor.b32 %r359, %r357, %r358; bfe.u32 %r360, %r359, 4, 4; shr.u32 %r361, %r354, 8; xor.b32 %r362, %r359, %r361; xor.b32 %r363, %r362, %r360; mul.lo.s32 %r364, %r363, 153742143; shr.u32 %r365, %r354, 23; xor.b32 %r366, %r364, %r365; bfe.u32 %r367, %r366, 1, 7; xor.b32 %r368, %r367, %r366; shr.u32 %r369, %r354, 27; or.b32 %r370, %r369, 1; mul.lo.s32 %r371, %r370, %r368; mul.lo.s32 %r372, %r371, -1074439621; bfe.u32 %r373, %r372, 2, 6; xor.b32 %r374, %r373, %r372; mul.lo.s32 %r375, %r374, -1638916925; bfe.u32 %r376, %r375, 2, 6; xor.b32 %r377, %r376, %r375; mul.lo.s32 %r378, %r377, -933190689; and.b32 %r379, %r378, 255; bfe.u32 %r380, %r378, 5, 3; xor.b32 %r381, %r380, %r379; add.s32 %r382, %r349, -1; xor.b32 %r92, %r89, 61; xor.b32 %r383, %r354, %r92; mul.lo.s32 %r384, %r383, 9; shr.u32 %r385, %r384, 4; xor.b32 %r386, %r385, %r384; mul.lo.s32 %r387, %r386, 668265261; add.s32 %r388, %r387, %r91; shr.u32 %r389, %r387, 16; or.b32 %r390, %r389, 1; mul.lo.s32 %r391, %r390, %r388; mul.lo.s32 %r392, %r391, 89287766; xor.b32 %r393, %r392, %r391; mul.lo.s32 %r394, %r393, 1403136100; xor.b32 %r395, %r394, %r393; brev.b32 %r396, %r395; neg.s32 %r397, %r349; and.b32 %r398, %r90, %r397; and.b32 %r399, %r396, %r382; or.b32 %r400, %r399, %r398; mad.lo.s32 %r401, %r381, %r349, %r400; shl.b32 %r402, %r349, 8; rem.u32 %r403, %r401, %r402; ld.const.u64 %rd714, [kernel_params+2424]; cvta.to.global.u64 %rd715, %rd714; shl.b32 %r404, %r403, 1; mul.wide.u32 %rd716, %r404, 4; add.s64 %rd717, %rd715, %rd716; ld.global.f32 %f331, [%rd717]; @%p247 bra $L__BB205_288; xor.b32 %r405, %r92, %r88; mul.lo.s32 %r406, %r405, 9; shr.u32 %r407, %r406, 4; xor.b32 %r408, %r407, %r406; mul.lo.s32 %r409, %r408, 668265261; cvt.rn.f32.u32 %f299, %r409; mul.ftz.f32 %f300, %f299, 0f2F7FFFFE; fma.rn.ftz.f32 %f301, %f300, %f81, %f331; cvt.rmi.ftz.f32.f32 %f302, %f301; sub.ftz.f32 %f331, %f301, %f302; bra.uni $L__BB205_288; $L__BB205_287: mul.lo.s32 %r410, %r89, 569420461; shr.u32 %r411, %r410, 15; xor.b32 %r412, %r411, %r410; mul.lo.s32 %r413, %r412, -749064809; shr.u32 %r414, %r413, 15; xor.b32 %r415, %r88, %r413; xor.b32 %r416, %r415, %r414; xor.b32 %r417, %r416, 1493655573; add.s32 %r418, %r417, %r91; shr.u32 %r419, %r417, 16; or.b32 %r420, %r419, 1; mul.lo.s32 %r421, %r420, %r418; mul.lo.s32 %r422, %r421, 89287766; xor.b32 %r423, %r422, %r421; mul.lo.s32 %r424, %r423, 1403136100; xor.b32 %r425, %r424, %r423; xor.b32 %r426, %r416, -2052961194; brev.b32 %r427, %r425; mul.lo.s32 %r428, %r427, 1025551850; xor.b32 %r429, %r428, %r427; add.s32 %r430, %r429, %r426; shr.u32 %r431, %r426, 16; or.b32 %r432, %r431, 1; mul.lo.s32 %r433, %r430, %r432; mul.lo.s32 %r434, %r433, 89287766; xor.b32 %r435, %r434, %r433; mul.lo.s32 %r436, %r435, 1403136100; xor.b32 %r437, %r436, %r435; brev.b32 %r438, %r437; cvt.rn.f32.u32 %f303, %r438; mul.ftz.f32 %f331, %f303, 0f2F7FFFFE; $L__BB205_288: setp.ltu.ftz.f32 %p250, %f331, %f330; setp.neu.ftz.f32 %p251, %f330, 0f00000000; and.pred %p252, %p251, %p250; mov.pred %p273, %p244; @%p252 bra $L__BB205_294; and.b32 %r439, %r475, 536870912; setp.eq.s32 %p253, %r439, 0; @%p253 bra $L__BB205_291; bra.uni $L__BB205_290; $L__BB205_291: mov.pred %p273, 0; @%p109 bra $L__BB205_294; ld.const.u64 %rd718, [kernel_params+2776]; cvta.to.global.u64 %rd719, %rd718; add.s64 %rd721, %rd719, %rd199; ld.global.u32 %r442, [%rd721]; setp.eq.s32 %p258, %r442, -1; @%p258 bra $L__BB205_294; ld.global.u32 %r443, [%rd16]; or.b32 %r444, %r443, 65536; st.global.u32 [%rd16], %r444; mov.pred %p273, %p244; bra.uni $L__BB205_294; $L__BB205_290: ld.global.u32 %r440, [%rd16]; or.b32 %r441, %r440, 32768; st.global.u32 [%rd16], %r441; mov.pred %p273, %p244; bra.uni $L__BB205_294; $L__BB205_93: mov.u16 %rs100, 1; @%p109 bra $L__BB205_95; ld.const.u64 %rd250, [kernel_params+2776]; cvta.to.global.u64 %rd251, %rd250; add.s64 %rd253, %rd251, %rd199; ld.global.u32 %r154, [%rd253]; setp.eq.s32 %p123, %r154, -1; selp.u16 %rs100, 1, 0, %p123; $L__BB205_95: setp.ne.s16 %p124, %rs100, 0; @%p124 bra $L__BB205_113; ld.global.u32 %r155, [%rd16]; or.b32 %r466, %r155, 262144; st.global.u32 [%rd16], %r466; bra.uni $L__BB205_98; $L__BB205_276: mov.pred %p273, 0; @%p109 bra $L__BB205_294; ld.const.u64 %rd681, [kernel_params+2776]; cvta.to.global.u64 %rd682, %rd681; add.s64 %rd684, %rd682, %rd199; ld.global.u32 %r339, [%rd684]; setp.eq.s32 %p240, %r339, -1; @%p240 bra $L__BB205_294; ld.global.u32 %r340, [%rd16]; or.b32 %r341, %r340, 262144; st.global.u32 [%rd16], %r341; bra.uni $L__BB205_279; $L__BB205_294: ld.const.u64 %rd722, [kernel_params+4256]; cvta.to.global.u64 %rd126, %rd722; add.s64 %rd723, %rd126, 8; atom.global.add.u32 %r445, [%rd723], -1; @%p273 bra $L__BB205_296; bra.uni $L__BB205_295; $L__BB205_296: add.s64 %rd728, %rd126, 44; atom.global.add.u32 %r446, [%rd728], 1; ld.const.u64 %rd729, [kernel_params+2520]; cvta.to.global.u64 %rd730, %rd729; shl.b64 %rd731, %rd3, 1; add.s64 %rd732, %rd730, %rd731; mov.u16 %rs86, 11; st.global.u16 [%rd732], %rs86; bra.uni $L__BB205_297; $L__BB205_295: ld.const.u64 %rd724, [kernel_params+2520]; cvta.to.global.u64 %rd725, %rd724; shl.b64 %rd726, %rd3, 1; add.s64 %rd727, %rd725, %rd726; mov.u16 %rs85, 0; st.global.u16 [%rd727], %rs85; bra.uni $L__BB205_297; $L__BB205_106: mul.lo.s32 %r225, %r52, 569420461; shr.u32 %r226, %r225, 15; xor.b32 %r227, %r226, %r225; mul.lo.s32 %r228, %r227, -749064809; shr.u32 %r229, %r228, 15; xor.b32 %r230, %r51, %r228; xor.b32 %r231, %r230, %r229; xor.b32 %r232, %r231, 1493655573; add.s32 %r233, %r232, %r54; shr.u32 %r234, %r232, 16; or.b32 %r235, %r234, 1; mul.lo.s32 %r236, %r235, %r233; mul.lo.s32 %r237, %r236, 89287766; xor.b32 %r238, %r237, %r236; mul.lo.s32 %r239, %r238, 1403136100; xor.b32 %r240, %r239, %r238; xor.b32 %r241, %r231, -2052961194; brev.b32 %r242, %r240; mul.lo.s32 %r243, %r242, 1025551850; xor.b32 %r244, %r243, %r242; add.s32 %r245, %r244, %r241; shr.u32 %r246, %r241, 16; or.b32 %r247, %r246, 1; mul.lo.s32 %r248, %r245, %r247; mul.lo.s32 %r249, %r248, 89287766; xor.b32 %r250, %r249, %r248; mul.lo.s32 %r251, %r250, 1403136100; xor.b32 %r252, %r251, %r250; brev.b32 %r253, %r252; cvt.rn.f32.u32 %f233, %r253; mul.ftz.f32 %f329, %f233, 0f2F7FFFFE; $L__BB205_107: setp.ltu.ftz.f32 %p132, %f329, %f328; setp.neu.ftz.f32 %p133, %f328, 0f00000000; and.pred %p134, %p133, %p132; @%p134 bra $L__BB205_115; and.b32 %r254, %r46, 536870912; setp.eq.s32 %p135, %r254, 0; @%p135 bra $L__BB205_110; bra.uni $L__BB205_109; $L__BB205_110: mov.u16 %rs101, 1; @%p109 bra $L__BB205_112; ld.const.u64 %rd275, [kernel_params+2776]; cvta.to.global.u64 %rd276, %rd275; add.s64 %rd278, %rd276, %rd199; ld.global.u32 %r257, [%rd278]; setp.eq.s32 %p137, %r257, -1; selp.u16 %rs101, 1, 0, %p137; $L__BB205_112: setp.eq.s16 %p138, %rs101, 0; @%p138 bra $L__BB205_114; $L__BB205_113: ld.const.u64 %rd279, [kernel_params+4256]; cvta.to.global.u64 %rd280, %rd279; add.s64 %rd281, %rd280, 8; atom.global.add.u32 %r258, [%rd281], -1; ld.const.u64 %rd282, [kernel_params+2520]; cvta.to.global.u64 %rd283, %rd282; shl.b64 %rd284, %rd3, 1; add.s64 %rd285, %rd283, %rd284; mov.u16 %rs55, 0; st.global.u16 [%rd285], %rs55; bra.uni $L__BB205_297; $L__BB205_109: ld.global.u32 %r255, [%rd16]; or.b32 %r256, %r255, 32768; st.global.u32 [%rd16], %r256; bra.uni $L__BB205_115; $L__BB205_114: ld.global.u32 %r259, [%rd16]; or.b32 %r260, %r259, 65536; st.global.u32 [%rd16], %r260; $L__BB205_115: ld.const.u32 %r447, [kernel_params+1592]; ld.const.u64 %rd286, [kernel_params+2184]; cvta.to.global.u64 %rd287, %rd286; mul.wide.s32 %rd288, %r37, 4; add.s64 %rd58, %rd287, %rd288; ld.global.u8 %rs56, [%rd58+1]; and.b16 %rs57, %rs56, 6; setp.ne.s16 %p139, %rs57, 0; setp.ne.s32 %p140, %r447, 0; and.pred %p141, %p139, %p140; @%p141 bra $L__BB205_119; bra.uni $L__BB205_116; $L__BB205_119: cvt.u32.u64 %r269, %rd3; ld.const.u32 %r467, [kernel_params+4392]; div.u32 %r270, %r269, %r467; ld.const.u32 %r468, [kernel_params+8]; mad.lo.s32 %r271, %r270, %r468, %r45; ld.const.u64 %rd69, [kernel_params+4256]; cvta.to.global.u64 %rd312, %rd69; add.s64 %rd313, %rd312, 8; atom.global.add.u32 %r272, [%rd313], -1; add.s64 %rd314, %rd312, 40; atom.global.add.u32 %r273, [%rd314], 1; ld.const.u64 %rd71, [kernel_params+2520]; cvta.to.global.u64 %rd315, %rd71; add.s64 %rd317, %rd315, %rd257; mov.u16 %rs60, 10; st.global.u16 [%rd317], %rs60; ld.const.u64 %rd70, [kernel_params+2624]; cvta.to.global.u64 %rd318, %rd70; add.s64 %rd320, %rd318, %rd199; st.global.u32 [%rd320], %r271; ld.const.u64 %rd321, [kernel_params+4344]; cvta.to.global.u64 %rd322, %rd321; mul.wide.s32 %rd323, %r271, 4; add.s64 %rd324, %rd322, %rd323; atom.global.add.u32 %r274, [%rd324], 1; bra.uni $L__BB205_120; $L__BB205_116: cvt.u32.u64 %r261, %rd3; and.b32 %r262, %r46, 1073741824; setp.eq.s32 %p142, %r262, 0; ld.const.u32 %r467, [kernel_params+4392]; div.u32 %r263, %r261, %r467; ld.const.u32 %r468, [kernel_params+8]; mad.lo.s32 %r58, %r263, %r468, %r45; ld.const.u64 %rd69, [kernel_params+4256]; cvta.to.global.u64 %rd60, %rd69; add.s64 %rd289, %rd60, 8; atom.global.add.u32 %r264, [%rd289], -1; cvt.s64.s32 %rd61, %r58; @%p142 bra $L__BB205_118; bra.uni $L__BB205_117; $L__BB205_118: add.s64 %rd301, %rd60, 32; atom.global.add.u32 %r267, [%rd301], 1; ld.const.u64 %rd71, [kernel_params+2520]; cvta.to.global.u64 %rd302, %rd71; add.s64 %rd304, %rd302, %rd257; mov.u16 %rs59, 8; st.global.u16 [%rd304], %rs59; ld.const.u64 %rd70, [kernel_params+2624]; cvta.to.global.u64 %rd305, %rd70; add.s64 %rd307, %rd305, %rd199; st.global.u32 [%rd307], %r58; ld.const.u64 %rd308, [kernel_params+4328]; cvta.to.global.u64 %rd309, %rd308; shl.b64 %rd310, %rd61, 2; add.s64 %rd311, %rd309, %rd310; atom.global.add.u32 %r268, [%rd311], 1; bra.uni $L__BB205_120; $L__BB205_117: add.s64 %rd290, %rd60, 36; atom.global.add.u32 %r265, [%rd290], 1; ld.const.u64 %rd71, [kernel_params+2520]; cvta.to.global.u64 %rd291, %rd71; add.s64 %rd293, %rd291, %rd257; mov.u16 %rs58, 9; st.global.u16 [%rd293], %rs58; ld.const.u64 %rd70, [kernel_params+2624]; cvta.to.global.u64 %rd294, %rd70; add.s64 %rd296, %rd294, %rd199; st.global.u32 [%rd296], %r58; ld.const.u64 %rd297, [kernel_params+4336]; cvta.to.global.u64 %rd298, %rd297; shl.b64 %rd299, %rd61, 2; add.s64 %rd300, %rd298, %rd299; atom.global.add.u32 %r266, [%rd300], 1; $L__BB205_120: ld.const.u32 %r275, [kernel_params+1620]; setp.eq.s32 %p143, %r275, 0; ld.global.u32 %r63, [%rd58]; and.b32 %r276, %r63, 129; setp.ne.s32 %p144, %r276, 128; or.pred %p145, %p144, %p143; mov.u16 %rs102, 0; @%p145 bra $L__BB205_123; ld.global.u32 %r64, [%rd16]; and.b32 %r277, %r64, 16384; setp.eq.s32 %p146, %r277, 0; @%p146 bra $L__BB205_123; and.b32 %r278, %r64, 536870912; setp.eq.s32 %p147, %r278, 0; selp.u16 %rs102, 1, 0, %p147; $L__BB205_123: setp.eq.s16 %p148, %rs102, 0; @%p148 bra $L__BB205_297; ld.param.u64 %rd760, [kernel_gpu_integrator_intersect_closest_param_1]; ld.const.u64 %rd325, [kernel_params+2448]; cvta.to.global.u64 %rd72, %rd325; add.s64 %rd73, %rd72, %rd199; ld.global.u32 %rd327, [%rd73]; ld.const.s32 %rd328, [kernel_params+1264]; mul.lo.s64 %rd329, %rd328, %rd327; ld.const.s32 %rd330, [kernel_params+1380]; add.s64 %rd331, %rd329, %rd330; cvta.to.global.u64 %rd332, %rd760; shl.b64 %rd333, %rd331, 2; add.s64 %rd334, %rd332, %rd333; atom.global.add.f32 %f234, [%rd334], 0f3F800000; ld.const.u64 %rd74, [kernel_params+2584]; cvta.to.global.u64 %rd75, %rd74; mul.lo.s64 %rd335, %rd3, 12; add.s64 %rd76, %rd75, %rd335; ld.const.s32 %rd336, [kernel_params+1384]; add.s64 %rd337, %rd336, %rd329; shl.b64 %rd338, %rd337, 2; add.s64 %rd339, %rd332, %rd338; add.s64 %rd340, %rd339, 12; ld.global.f32 %f235, [%rd76+4]; ld.global.f32 %f236, [%rd76]; add.ftz.f32 %f237, %f236, %f235; ld.global.f32 %f238, [%rd76+8]; add.ftz.f32 %f239, %f238, %f237; mul.ftz.f32 %f240, %f239, 0f3EAAAAAB; atom.global.add.f32 %f241, [%rd340], %f240; ld.global.u32 %r279, [%rd16]; or.b32 %r280, %r279, 536870912; st.global.u32 [%rd16], %r280; ld.const.u64 %rd341, [kernel_params+4384]; cvta.to.global.u64 %rd342, %rd341; atom.global.add.u32 %r65, [%rd342], 1; setp.eq.s64 %p149, %rd325, 0; @%p149 bra $L__BB205_126; ld.global.u32 %r281, [%rd73]; mul.wide.s32 %rd343, %r65, 4; add.s64 %rd344, %rd72, %rd343; st.global.u32 [%rd344], %r281; $L__BB205_126: ld.const.u64 %rd733, [kernel_params+2456]; setp.eq.s64 %p150, %rd733, 0; @%p150 bra $L__BB205_128; ld.const.u64 %rd765, [kernel_params+2456]; cvta.to.global.u64 %rd764, %rd765; add.s64 %rd763, %rd764, %rd199; ld.global.u32 %r282, [%rd763]; mul.wide.s32 %rd345, %r65, 4; add.s64 %rd346, %rd764, %rd345; st.global.u32 [%rd346], %r282; $L__BB205_128: ld.const.u64 %rd77, [kernel_params+2464]; setp.eq.s64 %p151, %rd77, 0; @%p151 bra $L__BB205_130; cvta.to.global.u64 %rd347, %rd77; add.s64 %rd349, %rd347, %rd257; ld.global.u16 %rs63, [%rd349]; mul.wide.s32 %rd350, %r65, 2; add.s64 %rd351, %rd347, %rd350; st.global.u16 [%rd351], %rs63; $L__BB205_130: ld.const.u64 %rd78, [kernel_params+2472]; setp.eq.s64 %p152, %rd78, 0; @%p152 bra $L__BB205_132; cvta.to.global.u64 %rd352, %rd78; add.s64 %rd354, %rd352, %rd257; ld.global.u16 %rs64, [%rd354]; mul.wide.s32 %rd355, %r65, 2; add.s64 %rd356, %rd352, %rd355; st.global.u16 [%rd356], %rs64; $L__BB205_132: ld.const.u64 %rd79, [kernel_params+2480]; setp.eq.s64 %p153, %rd79, 0; @%p153 bra $L__BB205_134; cvta.to.global.u64 %rd357, %rd79; add.s64 %rd359, %rd357, %rd257; ld.global.u16 %rs65, [%rd359]; mul.wide.s32 %rd360, %r65, 2; add.s64 %rd361, %rd357, %rd360; st.global.u16 [%rd361], %rs65; $L__BB205_134: ld.const.u64 %rd80, [kernel_params+2488]; setp.eq.s64 %p154, %rd80, 0; @%p154 bra $L__BB205_136; cvta.to.global.u64 %rd362, %rd80; add.s64 %rd364, %rd362, %rd257; ld.global.u16 %rs66, [%rd364]; mul.wide.s32 %rd365, %r65, 2; add.s64 %rd366, %rd362, %rd365; st.global.u16 [%rd366], %rs66; $L__BB205_136: ld.const.u64 %rd81, [kernel_params+2496]; setp.eq.s64 %p155, %rd81, 0; @%p155 bra $L__BB205_138; cvta.to.global.u64 %rd367, %rd81; add.s64 %rd369, %rd367, %rd257; ld.global.u16 %rs67, [%rd369]; mul.wide.s32 %rd370, %r65, 2; add.s64 %rd371, %rd367, %rd370; st.global.u16 [%rd371], %rs67; $L__BB205_138: ld.const.u64 %rd82, [kernel_params+2504]; setp.eq.s64 %p156, %rd82, 0; @%p156 bra $L__BB205_140; cvta.to.global.u64 %rd372, %rd82; add.s64 %rd374, %rd372, %rd257; ld.global.u16 %rs68, [%rd374]; mul.wide.s32 %rd375, %r65, 2; add.s64 %rd376, %rd372, %rd375; st.global.u16 [%rd376], %rs68; $L__BB205_140: ld.const.u64 %rd83, [kernel_params+2512]; setp.eq.s64 %p157, %rd83, 0; @%p157 bra $L__BB205_142; cvta.to.global.u64 %rd377, %rd83; add.s64 %rd379, %rd377, %rd257; ld.global.u16 %rs69, [%rd379]; mul.wide.s32 %rd380, %r65, 2; add.s64 %rd381, %rd377, %rd380; st.global.u16 [%rd381], %rs69; $L__BB205_142: setp.eq.s64 %p158, %rd71, 0; @%p158 bra $L__BB205_144; cvta.to.global.u64 %rd382, %rd71; add.s64 %rd384, %rd382, %rd257; ld.global.u16 %rs70, [%rd384]; mul.wide.s32 %rd385, %r65, 2; add.s64 %rd386, %rd382, %rd385; st.global.u16 [%rd386], %rs70; $L__BB205_144: setp.eq.s64 %p159, %rd254, 0; @%p159 bra $L__BB205_146; cvta.to.global.u64 %rd735, %rd254; add.s64 %rd734, %rd735, %rd199; ld.global.u32 %r283, [%rd734]; mul.wide.s32 %rd387, %r65, 4; add.s64 %rd388, %rd735, %rd387; st.global.u32 [%rd388], %r283; $L__BB205_146: setp.eq.s64 %p160, %rd255, 0; @%p160 bra $L__BB205_148; cvta.to.global.u64 %rd762, %rd255; add.s64 %rd761, %rd762, %rd257; cvta.to.global.u64 %rd736, %rd255; ld.global.u16 %rs71, [%rd761]; mul.wide.s32 %rd389, %r65, 2; add.s64 %rd390, %rd736, %rd389; st.global.u16 [%rd390], %rs71; $L__BB205_148: ld.const.u64 %rd767, [kernel_params+2544]; cvta.to.global.u64 %rd766, %rd767; ld.const.u64 %rd737, [kernel_params+2544]; setp.eq.s64 %p161, %rd737, 0; cvt.s64.s32 %rd84, %r65; mul.wide.s32 %rd392, %r65, 4; add.s64 %rd85, %rd766, %rd392; @%p161 bra $L__BB205_150; ld.global.u32 %r284, [%rd16]; st.global.u32 [%rd85], %r284; $L__BB205_150: ld.const.u64 %rd86, [kernel_params+2552]; setp.eq.s64 %p162, %rd86, 0; @%p162 bra $L__BB205_152; cvta.to.global.u64 %rd393, %rd86; add.s64 %rd394, %rd393, %rd3; ld.global.u8 %rs72, [%rd394]; add.s64 %rd396, %rd393, %rd84; st.global.u8 [%rd396], %rs72; $L__BB205_152: ld.const.u64 %rd87, [kernel_params+2560]; setp.eq.s64 %p163, %rd87, 0; @%p163 bra $L__BB205_154; cvta.to.global.u64 %rd397, %rd87; add.s64 %rd399, %rd397, %rd199; ld.global.f32 %f242, [%rd399]; add.s64 %rd401, %rd397, %rd392; st.global.f32 [%rd401], %f242; $L__BB205_154: ld.const.u64 %rd88, [kernel_params+2568]; setp.eq.s64 %p164, %rd88, 0; @%p164 bra $L__BB205_156; cvta.to.global.u64 %rd402, %rd88; add.s64 %rd404, %rd402, %rd199; ld.global.f32 %f243, [%rd404]; add.s64 %rd406, %rd402, %rd392; st.global.f32 [%rd406], %f243; $L__BB205_156: ld.const.u64 %rd738, [kernel_params+2576]; setp.eq.s64 %p165, %rd738, 0; @%p165 bra $L__BB205_158; ld.const.u64 %rd770, [kernel_params+2576]; cvta.to.global.u64 %rd769, %rd770; add.s64 %rd768, %rd769, %rd199; ld.const.u64 %rd740, [kernel_params+2576]; cvta.to.global.u64 %rd739, %rd740; ld.global.f32 %f244, [%rd768]; add.s64 %rd408, %rd739, %rd392; st.global.f32 [%rd408], %f244; $L__BB205_158: ld.const.u64 %rd741, [kernel_params+2584]; setp.eq.s64 %p166, %rd741, 0; @%p166 bra $L__BB205_160; ld.const.u64 %rd772, [kernel_params+2584]; cvta.to.global.u64 %rd771, %rd772; ld.global.f32 %f245, [%rd76]; ld.global.f32 %f246, [%rd76+4]; ld.global.f32 %f247, [%rd76+8]; mul.wide.s32 %rd409, %r65, 12; add.s64 %rd410, %rd771, %rd409; st.global.f32 [%rd410], %f245; st.global.f32 [%rd410+4], %f246; st.global.f32 [%rd410+8], %f247; $L__BB205_160: ld.const.u64 %rd89, [kernel_params+2592]; setp.eq.s64 %p167, %rd89, 0; @%p167 bra $L__BB205_162; cvta.to.global.u64 %rd411, %rd89; add.s64 %rd413, %rd411, %rd199; ld.global.f32 %f248, [%rd413]; add.s64 %rd415, %rd411, %rd392; st.global.f32 [%rd415], %f248; $L__BB205_162: ld.const.u64 %rd90, [kernel_params+2600]; setp.eq.s64 %p168, %rd90, 0; @%p168 bra $L__BB205_164; cvta.to.global.u64 %rd416, %rd90; add.s64 %rd418, %rd416, %rd335; ld.global.f32 %f249, [%rd418]; ld.global.f32 %f250, [%rd418+4]; ld.global.f32 %f251, [%rd418+8]; mul.wide.s32 %rd419, %r65, 12; add.s64 %rd420, %rd416, %rd419; st.global.f32 [%rd420], %f249; st.global.f32 [%rd420+4], %f250; st.global.f32 [%rd420+8], %f251; $L__BB205_164: ld.const.u64 %rd91, [kernel_params+2608]; setp.eq.s64 %p169, %rd91, 0; @%p169 bra $L__BB205_166; cvta.to.global.u64 %rd421, %rd91; add.s64 %rd423, %rd421, %rd335; ld.global.f32 %f252, [%rd423]; ld.global.f32 %f253, [%rd423+4]; ld.global.f32 %f254, [%rd423+8]; mul.wide.s32 %rd424, %r65, 12; add.s64 %rd425, %rd421, %rd424; st.global.f32 [%rd425], %f252; st.global.f32 [%rd425+4], %f253; st.global.f32 [%rd425+8], %f254; $L__BB205_166: ld.const.u64 %rd92, [kernel_params+2616]; setp.eq.s64 %p170, %rd92, 0; @%p170 bra $L__BB205_168; cvta.to.global.u64 %rd426, %rd92; add.s64 %rd428, %rd426, %rd335; ld.global.f32 %f255, [%rd428]; ld.global.f32 %f256, [%rd428+4]; ld.global.f32 %f257, [%rd428+8]; mul.wide.s32 %rd429, %r65, 12; add.s64 %rd430, %rd426, %rd429; st.global.f32 [%rd430], %f255; st.global.f32 [%rd430+4], %f256; st.global.f32 [%rd430+8], %f257; $L__BB205_168: setp.eq.s64 %p171, %rd70, 0; @%p171 bra $L__BB205_170; cvta.to.global.u64 %rd431, %rd70; add.s64 %rd433, %rd431, %rd199; ld.global.u32 %r285, [%rd433]; add.s64 %rd435, %rd431, %rd392; st.global.u32 [%rd435], %r285; $L__BB205_170: ld.const.u64 %rd742, [kernel_params+2632]; setp.eq.s64 %p172, %rd742, 0; @%p172 bra $L__BB205_172; ld.const.u64 %rd801, [kernel_params+2632]; mul.wide.s32 %rd800, %r3, 12; cvta.to.global.u64 %rd799, %rd801; add.s64 %rd798, %rd799, %rd800; ld.const.u64 %rd774, [kernel_params+2632]; cvta.to.global.u64 %rd773, %rd774; ld.global.f32 %f258, [%rd798]; ld.global.f32 %f259, [%rd798+4]; ld.global.f32 %f260, [%rd798+8]; mul.wide.s32 %rd436, %r65, 12; add.s64 %rd437, %rd773, %rd436; st.global.f32 [%rd437], %f258; st.global.f32 [%rd437+4], %f259; st.global.f32 [%rd437+8], %f260; $L__BB205_172: setp.eq.s64 %p173, %rd134, 0; @%p173 bra $L__BB205_174; cvta.to.global.u64 %rd743, %rd134; add.s64 %rd440, %rd743, %rd335; ld.global.f32 %f261, [%rd440]; ld.global.f32 %f262, [%rd440+4]; ld.global.f32 %f263, [%rd440+8]; mul.wide.s32 %rd441, %r65, 12; add.s64 %rd442, %rd743, %rd441; st.global.f32 [%rd442], %f261; st.global.f32 [%rd442+4], %f262; st.global.f32 [%rd442+8], %f263; $L__BB205_174: setp.eq.s64 %p174, %rd135, 0; @%p174 bra $L__BB205_176; mul.wide.s32 %rd777, %r3, 4; cvta.to.global.u64 %rd776, %rd135; add.s64 %rd775, %rd776, %rd777; cvta.to.global.u64 %rd744, %rd135; ld.global.f32 %f264, [%rd775]; add.s64 %rd445, %rd744, %rd392; st.global.f32 [%rd445], %f264; $L__BB205_176: setp.eq.s64 %p175, %rd140, 0; @%p175 bra $L__BB205_178; mul.wide.s32 %rd780, %r3, 4; cvta.to.global.u64 %rd779, %rd140; add.s64 %rd778, %rd779, %rd780; cvta.to.global.u64 %rd745, %rd140; ld.global.f32 %f265, [%rd778]; add.s64 %rd448, %rd745, %rd392; st.global.f32 [%rd448], %f265; $L__BB205_178: setp.eq.s64 %p176, %rd141, 0; @%p176 bra $L__BB205_180; cvta.to.global.u64 %rd746, %rd141; add.s64 %rd451, %rd746, %rd199; ld.global.f32 %f266, [%rd451]; add.s64 %rd453, %rd746, %rd392; st.global.f32 [%rd453], %f266; $L__BB205_180: setp.eq.s64 %p177, %rd145, 0; @%p177 bra $L__BB205_182; cvta.to.global.u64 %rd747, %rd145; add.s64 %rd456, %rd747, %rd199; ld.global.f32 %f267, [%rd456]; add.s64 %rd458, %rd747, %rd392; st.global.f32 [%rd458], %f267; $L__BB205_182: setp.eq.s64 %p178, %rd146, 0; @%p178 bra $L__BB205_184; mul.wide.s32 %rd795, %r3, 4; cvta.to.global.u64 %rd794, %rd146; add.s64 %rd793, %rd794, %rd795; cvta.to.global.u64 %rd748, %rd146; ld.global.f32 %f268, [%rd793]; add.s64 %rd461, %rd748, %rd392; st.global.f32 [%rd461], %f268; $L__BB205_184: setp.eq.s64 %p179, %rd197, 0; @%p179 bra $L__BB205_186; cvta.to.global.u64 %rd750, %rd197; add.s64 %rd749, %rd750, %rd199; ld.global.f32 %f269, [%rd749]; add.s64 %rd463, %rd750, %rd392; st.global.f32 [%rd463], %f269; $L__BB205_186: setp.eq.s64 %p180, %rd198, 0; @%p180 bra $L__BB205_188; cvta.to.global.u64 %rd752, %rd198; add.s64 %rd751, %rd752, %rd199; ld.global.f32 %f270, [%rd751]; add.s64 %rd465, %rd752, %rd392; st.global.f32 [%rd465], %f270; $L__BB205_188: ld.const.u64 %rd753, [kernel_params+2704]; setp.eq.s64 %p181, %rd753, 0; @%p181 bra $L__BB205_190; ld.const.u64 %rd783, [kernel_params+2704]; cvta.to.global.u64 %rd782, %rd783; add.s64 %rd781, %rd782, %rd199; ld.global.f32 %f271, [%rd781]; add.s64 %rd467, %rd782, %rd392; st.global.f32 [%rd467], %f271; $L__BB205_190: ld.const.u64 %rd754, [kernel_params+2712]; setp.eq.s64 %p182, %rd754, 0; @%p182 bra $L__BB205_192; ld.const.u64 %rd805, [kernel_params+2712]; mul.wide.s32 %rd804, %r3, 4; cvta.to.global.u64 %rd803, %rd805; add.s64 %rd802, %rd803, %rd804; ld.const.u64 %rd756, [kernel_params+2712]; cvta.to.global.u64 %rd755, %rd756; ld.global.u32 %r286, [%rd802]; add.s64 %rd469, %rd755, %rd392; st.global.u32 [%rd469], %r286; $L__BB205_192: ld.const.u64 %rd757, [kernel_params+2720]; setp.eq.s64 %p183, %rd757, 0; @%p183 bra $L__BB205_194; ld.const.u64 %rd809, [kernel_params+2720]; mul.wide.s32 %rd808, %r3, 4; cvta.to.global.u64 %rd807, %rd809; add.s64 %rd806, %rd807, %rd808; ld.const.u64 %rd785, [kernel_params+2720]; cvta.to.global.u64 %rd784, %rd785; ld.global.u32 %r287, [%rd806]; add.s64 %rd471, %rd784, %rd392; st.global.u32 [%rd471], %r287; $L__BB205_194: setp.eq.s64 %p184, %rd33, 0; @%p184 bra $L__BB205_196; cvta.to.global.u64 %rd759, %rd33; add.s64 %rd758, %rd759, %rd199; ld.global.u32 %r288, [%rd758]; add.s64 %rd473, %rd759, %rd392; st.global.u32 [%rd473], %r288; $L__BB205_196: ld.const.u64 %rd93, [kernel_params+2736]; setp.eq.s64 %p185, %rd93, 0; @%p185 bra $L__BB205_198; cvta.to.global.u64 %rd474, %rd93; add.s64 %rd476, %rd474, %rd335; ld.global.f32 %f272, [%rd476]; ld.global.f32 %f273, [%rd476+4]; ld.global.f32 %f274, [%rd476+8]; mul.wide.s32 %rd477, %r65, 12; add.s64 %rd478, %rd474, %rd477; st.global.f32 [%rd478], %f272; st.global.f32 [%rd478+4], %f273; st.global.f32 [%rd478+8], %f274; $L__BB205_198: ld.const.u64 %rd94, [kernel_params+2744]; setp.eq.s64 %p186, %rd94, 0; @%p186 bra $L__BB205_200; cvta.to.global.u64 %rd479, %rd94; add.s64 %rd481, %rd479, %rd335; ld.global.f32 %f275, [%rd481]; ld.global.f32 %f276, [%rd481+4]; ld.global.f32 %f277, [%rd481+8]; mul.wide.s32 %rd482, %r65, 12; add.s64 %rd483, %rd479, %rd482; st.global.f32 [%rd483], %f275; st.global.f32 [%rd483+4], %f276; st.global.f32 [%rd483+8], %f277; $L__BB205_200: ld.const.u64 %rd95, [kernel_params+2752]; setp.eq.s64 %p187, %rd95, 0; @%p187 bra $L__BB205_202; cvta.to.global.u64 %rd484, %rd95; add.s64 %rd486, %rd484, %rd199; ld.global.f32 %f278, [%rd486]; shl.b64 %rd487, %rd84, 2; add.s64 %rd488, %rd484, %rd487; st.global.f32 [%rd488], %f278; $L__BB205_202: ld.const.u64 %rd96, [kernel_params+2760]; setp.eq.s64 %p188, %rd96, 0; @%p188 bra $L__BB205_204; cvta.to.global.u64 %rd489, %rd96; add.s64 %rd491, %rd489, %rd335; ld.global.f32 %f279, [%rd491]; ld.global.f32 %f280, [%rd491+4]; ld.global.f32 %f281, [%rd491+8]; mul.lo.s64 %rd492, %rd84, 12; add.s64 %rd493, %rd489, %rd492; st.global.f32 [%rd493], %f279; st.global.f32 [%rd493+4], %f280; st.global.f32 [%rd493+8], %f281; $L__BB205_204: ld.const.u32 %r290, [kernel_params+12]; max.u32 %r66, %r290, 1; add.s32 %r291, %r66, -1; and.b32 %r472, %r66, 3; setp.lt.u32 %p189, %r291, 3; mov.u32 %r471, 0; @%p189 bra $L__BB205_223; sub.s32 %r470, %r66, %r472; mov.u32 %r471, 0; mov.u64 %rd495, kernel_params; $L__BB205_206: cvt.s64.s32 %rd97, %r471; mul.wide.s32 %rd494, %r471, 16; add.s64 %rd496, %rd495, %rd494; add.s64 %rd98, %rd496, 2768; ld.const.u64 %rd99, [%rd496+2768]; setp.eq.s64 %p190, %rd99, 0; @%p190 bra $L__BB205_208; cvta.to.global.u64 %rd497, %rd99; add.s64 %rd499, %rd497, %rd199; ld.global.u32 %r293, [%rd499]; shl.b64 %rd500, %rd84, 2; add.s64 %rd501, %rd497, %rd500; st.global.u32 [%rd501], %r293; $L__BB205_208: ld.const.u64 %rd100, [%rd98+8]; setp.eq.s64 %p191, %rd100, 0; @%p191 bra $L__BB205_210; cvta.to.global.u64 %rd502, %rd100; add.s64 %rd504, %rd502, %rd199; ld.global.u32 %r294, [%rd504]; shl.b64 %rd505, %rd84, 2; add.s64 %rd506, %rd502, %rd505; st.global.u32 [%rd506], %r294; $L__BB205_210: ld.const.u64 %rd101, [%rd98+16]; setp.eq.s64 %p192, %rd101, 0; @%p192 bra $L__BB205_212; cvta.to.global.u64 %rd507, %rd101; add.s64 %rd509, %rd507, %rd199; ld.global.u32 %r295, [%rd509]; shl.b64 %rd510, %rd84, 2; add.s64 %rd511, %rd507, %rd510; st.global.u32 [%rd511], %r295; $L__BB205_212: ld.const.u64 %rd102, [%rd98+24]; setp.eq.s64 %p193, %rd102, 0; @%p193 bra $L__BB205_214; cvta.to.global.u64 %rd512, %rd102; add.s64 %rd514, %rd512, %rd199; ld.global.u32 %r296, [%rd514]; shl.b64 %rd515, %rd84, 2; add.s64 %rd516, %rd512, %rd515; st.global.u32 [%rd516], %r296; $L__BB205_214: ld.const.u64 %rd103, [%rd98+32]; setp.eq.s64 %p194, %rd103, 0; @%p194 bra $L__BB205_216; cvta.to.global.u64 %rd517, %rd103; add.s64 %rd519, %rd517, %rd199; ld.global.u32 %r297, [%rd519]; shl.b64 %rd520, %rd84, 2; add.s64 %rd521, %rd517, %rd520; st.global.u32 [%rd521], %r297; $L__BB205_216: ld.const.u64 %rd104, [%rd98+40]; setp.eq.s64 %p195, %rd104, 0; @%p195 bra $L__BB205_218; cvta.to.global.u64 %rd522, %rd104; add.s64 %rd524, %rd522, %rd199; ld.global.u32 %r298, [%rd524]; shl.b64 %rd525, %rd84, 2; add.s64 %rd526, %rd522, %rd525; st.global.u32 [%rd526], %r298; $L__BB205_218: ld.const.u64 %rd105, [%rd98+48]; setp.eq.s64 %p196, %rd105, 0; @%p196 bra $L__BB205_220; cvta.to.global.u64 %rd527, %rd105; add.s64 %rd529, %rd527, %rd199; ld.global.u32 %r299, [%rd529]; shl.b64 %rd530, %rd84, 2; add.s64 %rd531, %rd527, %rd530; st.global.u32 [%rd531], %r299; $L__BB205_220: ld.const.u64 %rd106, [%rd98+56]; setp.eq.s64 %p197, %rd106, 0; @%p197 bra $L__BB205_222; cvta.to.global.u64 %rd532, %rd106; add.s64 %rd534, %rd532, %rd199; ld.global.u32 %r300, [%rd534]; shl.b64 %rd535, %rd84, 2; add.s64 %rd536, %rd532, %rd535; st.global.u32 [%rd536], %r300; $L__BB205_222: cvt.u32.u64 %r301, %rd97; add.s32 %r471, %r301, 4; add.s32 %r470, %r470, -4; setp.ne.s32 %p198, %r470, 0; @%p198 bra $L__BB205_206; $L__BB205_223: setp.eq.s32 %p199, %r472, 0; @%p199 bra $L__BB205_230; mul.wide.s32 %rd537, %r471, 16; mov.u64 %rd538, kernel_params; add.s64 %rd539, %rd538, %rd537; add.s64 %rd840, %rd539, 2776; $L__BB205_225: .pragma "nounroll"; ld.const.u64 %rd110, [%rd840+-8]; setp.eq.s64 %p200, %rd110, 0; @%p200 bra $L__BB205_227; cvta.to.global.u64 %rd540, %rd110; add.s64 %rd542, %rd540, %rd199; ld.global.u32 %r302, [%rd542]; shl.b64 %rd543, %rd84, 2; add.s64 %rd544, %rd540, %rd543; st.global.u32 [%rd544], %r302; $L__BB205_227: ld.const.u64 %rd111, [%rd840]; setp.eq.s64 %p201, %rd111, 0; @%p201 bra $L__BB205_229; cvta.to.global.u64 %rd545, %rd111; add.s64 %rd547, %rd545, %rd199; ld.global.u32 %r303, [%rd547]; shl.b64 %rd548, %rd84, 2; add.s64 %rd549, %rd545, %rd548; st.global.u32 [%rd549], %r303; $L__BB205_229: add.s32 %r472, %r472, -1; add.s64 %rd840, %rd840, 16; setp.ne.s32 %p202, %r472, 0; @%p202 bra $L__BB205_225; $L__BB205_230: ld.const.u64 %rd113, [kernel_params+3280]; setp.eq.s64 %p203, %rd113, 0; @%p203 bra $L__BB205_232; cvta.to.global.u64 %rd550, %rd113; shl.b64 %rd551, %rd3, 3; add.s64 %rd552, %rd550, %rd551; ld.global.u64 %rd553, [%rd552]; shl.b64 %rd554, %rd84, 3; add.s64 %rd555, %rd550, %rd554; st.global.u64 [%rd555], %rd553; $L__BB205_232: ld.const.u64 %rd114, [kernel_params+3288]; setp.eq.s64 %p204, %rd114, 0; @%p204 bra $L__BB205_234; cvta.to.global.u64 %rd556, %rd114; add.s64 %rd557, %rd556, %rd3; ld.global.u8 %rs73, [%rd557]; add.s64 %rd558, %rd556, %rd84; st.global.u8 [%rd558], %rs73; $L__BB205_234: ld.const.u64 %rd115, [kernel_params+3296]; setp.eq.s64 %p205, %rd115, 0; @%p205 bra $L__BB205_236; cvta.to.global.u64 %rd559, %rd115; add.s64 %rd561, %rd559, %rd199; ld.global.f32 %f282, [%rd561]; shl.b64 %rd562, %rd84, 2; add.s64 %rd563, %rd559, %rd562; st.global.f32 [%rd563], %f282; $L__BB205_236: ld.const.u64 %rd116, [kernel_params+3304]; setp.eq.s64 %p206, %rd116, 0; @%p206 bra $L__BB205_238; cvta.to.global.u64 %rd564, %rd116; add.s64 %rd566, %rd564, %rd199; ld.global.f32 %f283, [%rd566]; shl.b64 %rd567, %rd84, 2; add.s64 %rd568, %rd564, %rd567; st.global.f32 [%rd568], %f283; $L__BB205_238: ld.const.u64 %rd117, [kernel_params+3312]; setp.eq.s64 %p207, %rd117, 0; @%p207 bra $L__BB205_240; cvta.to.global.u64 %rd569, %rd117; add.s64 %rd571, %rd569, %rd199; ld.global.f32 %f284, [%rd571]; shl.b64 %rd572, %rd84, 2; add.s64 %rd573, %rd569, %rd572; st.global.f32 [%rd573], %f284; $L__BB205_240: ld.const.u64 %rd118, [kernel_params+3320]; setp.eq.s64 %p208, %rd118, 0; @%p208 bra $L__BB205_242; cvta.to.global.u64 %rd574, %rd118; add.s64 %rd575, %rd574, %rd3; ld.global.u8 %rs74, [%rd575]; add.s64 %rd576, %rd574, %rd84; st.global.u8 [%rd576], %rs74; $L__BB205_242: ld.const.u64 %rd119, [kernel_params+3328]; setp.eq.s64 %p209, %rd119, 0; @%p209 bra $L__BB205_244; cvta.to.global.u64 %rd577, %rd119; add.s64 %rd579, %rd577, %rd199; ld.global.f32 %f285, [%rd579]; shl.b64 %rd580, %rd84, 2; add.s64 %rd581, %rd577, %rd580; st.global.f32 [%rd581], %f285; $L__BB205_244: ld.const.u64 %rd120, [kernel_params+3336]; setp.eq.s64 %p210, %rd120, 0; @%p210 bra $L__BB205_246; cvta.to.global.u64 %rd582, %rd120; add.s64 %rd584, %rd582, %rd199; ld.global.f32 %f286, [%rd584]; shl.b64 %rd585, %rd84, 2; add.s64 %rd586, %rd582, %rd585; st.global.f32 [%rd586], %f286; $L__BB205_246: ld.global.u32 %r76, [%rd85]; or.b32 %r304, %r76, 1073741824; st.global.u32 [%rd85], %r304; ld.const.u32 %r305, [kernel_params+1356]; setp.ne.s32 %p211, %r305, -1; ld.const.u32 %r306, [kernel_params+1036]; setp.eq.s32 %p212, %r306, 0; and.pred %p213, %p211, %p212; @%p213 bra $L__BB205_261; bra.uni $L__BB205_247; $L__BB205_261: or.b32 %r322, %r76, -1073741824; st.global.u32 [%rd85], %r322; cvta.to.global.u64 %rd645, %rd69; add.s64 %rd646, %rd645, 24; atom.global.add.u32 %r323, [%rd646], 1; cvta.to.global.u64 %rd647, %rd71; shl.b64 %rd648, %rd84, 1; add.s64 %rd649, %rd647, %rd648; mov.u16 %rs81, 6; st.global.u16 [%rd649], %rs81; bra.uni $L__BB205_297; $L__BB205_247: @%p109 bra $L__BB205_250; ld.const.u64 %rd587, [kernel_params+2776]; cvta.to.global.u64 %rd588, %rd587; shl.b64 %rd589, %rd84, 2; add.s64 %rd590, %rd588, %rd589; ld.global.u32 %r307, [%rd590]; setp.eq.s32 %p215, %r307, -1; @%p215 bra $L__BB205_250; cvta.to.global.u64 %rd591, %rd69; add.s64 %rd592, %rd591, 20; atom.global.add.u32 %r308, [%rd592], 1; cvta.to.global.u64 %rd593, %rd71; shl.b64 %rd594, %rd84, 1; add.s64 %rd595, %rd593, %rd594; mov.u16 %rs75, 5; st.global.u16 [%rd595], %rs75; bra.uni $L__BB205_297; $L__BB205_250: and.b32 %r453, %r38, 1; setp.eq.s32 %p268, %r453, 0; @%p268 bra $L__BB205_252; bra.uni $L__BB205_251; $L__BB205_252: and.b32 %r309, %r38, 8; setp.eq.s32 %p217, %r309, 0; @%p217 bra $L__BB205_254; bra.uni $L__BB205_253; $L__BB205_254: and.b32 %r311, %r38, 6; setp.eq.s32 %p218, %r311, 0; mov.u32 %r473, 0; @%p218 bra $L__BB205_256; ld.const.u64 %rd604, [kernel_params+2264]; cvta.to.global.u64 %rd605, %rd604; mul.wide.s32 %rd606, %r36, 16; add.s64 %rd607, %rd605, %rd606; ld.global.u32 %r473, [%rd607]; bra.uni $L__BB205_256; $L__BB205_251: ld.const.u64 %rd596, [kernel_params+2216]; cvta.to.global.u64 %rd597, %rd596; mul.wide.s32 %rd598, %r36, 4; add.s64 %rd599, %rd597, %rd598; ld.global.u32 %r473, [%rd599]; bra.uni $L__BB205_256; $L__BB205_253: ld.const.u64 %rd600, [kernel_params+2304]; cvta.to.global.u64 %rd601, %rd600; mul.wide.s32 %rd602, %r36, 4; add.s64 %rd603, %rd601, %rd602; ld.global.u32 %r473, [%rd603]; $L__BB205_256: ld.const.u32 %r448, [kernel_params+1592]; setp.ne.s32 %p260, %r448, 0; ld.const.u64 %rd787, [kernel_params+2408]; cvta.to.global.u64 %rd786, %rd787; and.b32 %r312, %r473, 4194303; shl.b32 %r313, %r473, 5; cvt.u64.u32 %rd608, %r313; and.b64 %rd609, %rd608, 134217696; add.s64 %rd610, %rd786, %rd609; add.s64 %rd121, %rd610, 16; and.b32 %r314, %r63, 1536; setp.ne.s32 %p219, %r314, 0; and.pred %p221, %p219, %p260; div.u32 %r315, %r65, %r467; mad.lo.s32 %r81, %r315, %r468, %r312; cvt.s64.s32 %rd122, %r81; @%p221 bra $L__BB205_260; bra.uni $L__BB205_257; $L__BB205_260: cvta.to.global.u64 %rd633, %rd69; add.s64 %rd634, %rd633, 40; atom.global.add.u32 %r320, [%rd634], 1; cvta.to.global.u64 %rd635, %rd71; shl.b64 %rd636, %rd84, 1; add.s64 %rd637, %rd635, %rd636; mov.u16 %rs80, 10; st.global.u16 [%rd637], %rs80; cvta.to.global.u64 %rd638, %rd70; shl.b64 %rd639, %rd84, 2; add.s64 %rd640, %rd638, %rd639; st.global.u32 [%rd640], %r81; ld.const.u64 %rd641, [kernel_params+4344]; cvta.to.global.u64 %rd642, %rd641; shl.b64 %rd643, %rd122, 2; add.s64 %rd644, %rd642, %rd643; atom.global.add.u32 %r321, [%rd644], 1; bra.uni $L__BB205_297; $L__BB205_257: ld.global.u8 %rs76, [%rd121+3]; and.b16 %rs77, %rs76, 64; setp.eq.s16 %p222, %rs77, 0; cvta.to.global.u64 %rd123, %rd69; @%p222 bra $L__BB205_259; bra.uni $L__BB205_258; $L__BB205_259: add.s64 %rd622, %rd123, 32; atom.global.add.u32 %r318, [%rd622], 1; cvta.to.global.u64 %rd623, %rd71; shl.b64 %rd624, %rd84, 1; add.s64 %rd625, %rd623, %rd624; mov.u16 %rs79, 8; st.global.u16 [%rd625], %rs79; cvta.to.global.u64 %rd626, %rd70; shl.b64 %rd627, %rd84, 2; add.s64 %rd628, %rd626, %rd627; st.global.u32 [%rd628], %r81; ld.const.u64 %rd629, [kernel_params+4328]; cvta.to.global.u64 %rd630, %rd629; shl.b64 %rd631, %rd122, 2; add.s64 %rd632, %rd630, %rd631; atom.global.add.u32 %r319, [%rd632], 1; bra.uni $L__BB205_297; $L__BB205_258: add.s64 %rd611, %rd123, 36; atom.global.add.u32 %r316, [%rd611], 1; cvta.to.global.u64 %rd612, %rd71; shl.b64 %rd613, %rd84, 1; add.s64 %rd614, %rd612, %rd613; mov.u16 %rs78, 9; st.global.u16 [%rd614], %rs78; cvta.to.global.u64 %rd615, %rd70; shl.b64 %rd616, %rd84, 2; add.s64 %rd617, %rd615, %rd616; st.global.u32 [%rd617], %r81; ld.const.u64 %rd618, [kernel_params+4336]; cvta.to.global.u64 %rd619, %rd618; shl.b64 %rd620, %rd122, 2; add.s64 %rd621, %rd619, %rd620; atom.global.add.u32 %r317, [%rd621], 1; $L__BB205_297: ret; } // .globl kernel_gpu_integrator_intersect_shadow .visible .entry kernel_gpu_integrator_intersect_shadow( .param .u64 kernel_gpu_integrator_intersect_shadow_param_0, .param .u32 kernel_gpu_integrator_intersect_shadow_param_1 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 4 .b8 __local_depot206[84]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<23>; .reg .b16 %rs<20>; .reg .f32 %f<16>; .reg .b32 %r<33>; .reg .b64 %rd<72>; mov.u64 %SPL, __local_depot206; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd3, [kernel_gpu_integrator_intersect_shadow_param_0]; ld.param.u32 %r7, [kernel_gpu_integrator_intersect_shadow_param_1]; mov.u32 %r8, %ntid.x; mov.u32 %r9, %ctaid.x; mov.u32 %r10, %tid.x; mad.lo.s32 %r32, %r9, %r8, %r10; setp.ge.s32 %p1, %r32, %r7; @%p1 bra $L__BB206_22; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB206_3; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r32, 4; add.s64 %rd6, %rd4, %rd5; ld.global.u32 %r32, [%rd6]; $L__BB206_3: ld.const.u64 %rd7, [kernel_params+3496]; cvta.to.global.u64 %rd8, %rd7; cvt.s64.s32 %rd1, %r32; mul.wide.s32 %rd9, %r32, 12; add.s64 %rd10, %rd8, %rd9; ld.global.f32 %f1, [%rd10]; ld.global.f32 %f5, [%rd10+4]; ld.global.f32 %f6, [%rd10+8]; add.u64 %rd11, %SP, 24; add.u64 %rd12, %SPL, 24; st.local.f32 [%rd12], %f1; st.local.f32 [%rd12+4], %f5; st.local.f32 [%rd12+8], %f6; ld.const.v2.u64 {%rd13, %rd14}, [kernel_params+3504]; mov.u32 %r11, 0; cvta.to.global.u64 %rd17, %rd13; add.s64 %rd18, %rd17, %rd9; ld.global.f32 %f2, [%rd18]; ld.global.f32 %f3, [%rd18+4]; ld.global.f32 %f4, [%rd18+8]; st.local.f32 [%rd12+12], %f2; st.local.f32 [%rd12+16], %f3; st.local.f32 [%rd12+20], %f4; cvta.to.global.u64 %rd19, %rd14; mul.wide.s32 %rd20, %r32, 4; add.s64 %rd21, %rd19, %rd20; ld.global.f32 %f7, [%rd21]; st.local.f32 [%rd12+24], %f7; ld.const.v2.u64 {%rd22, %rd23}, [kernel_params+3520]; cvta.to.global.u64 %rd26, %rd22; add.s64 %rd27, %rd26, %rd20; ld.global.f32 %f8, [%rd27]; st.local.f32 [%rd12+28], %f8; cvta.to.global.u64 %rd28, %rd23; add.s64 %rd29, %rd28, %rd20; ld.global.f32 %f9, [%rd29]; st.local.f32 [%rd12+32], %f9; ld.const.u64 %rd30, [kernel_params+3536]; cvta.to.global.u64 %rd31, %rd30; add.s64 %rd32, %rd31, %rd20; ld.global.f32 %f10, [%rd32]; st.local.f32 [%rd12+52], %f10; st.local.u32 [%rd12+56], %r11; ld.const.u64 %rd33, [kernel_params+3584]; cvta.to.global.u64 %rd34, %rd33; add.s64 %rd35, %rd34, %rd20; ld.global.u32 %r12, [%rd35]; st.local.u32 [%rd12+40], %r12; ld.const.u64 %rd36, [kernel_params+3576]; cvta.to.global.u64 %rd37, %rd36; add.s64 %rd38, %rd37, %rd20; ld.global.u32 %r13, [%rd38]; st.local.u32 [%rd12+36], %r13; ld.const.u64 %rd39, [kernel_params+3632]; cvta.to.global.u64 %rd40, %rd39; add.s64 %rd41, %rd40, %rd20; ld.global.u32 %r14, [%rd41]; st.local.u32 [%rd12+48], %r14; ld.const.u64 %rd42, [kernel_params+3624]; cvta.to.global.u64 %rd43, %rd42; add.s64 %rd44, %rd43, %rd20; ld.global.u32 %r15, [%rd44]; st.local.u32 [%rd12+44], %r15; ld.const.u64 %rd45, [kernel_params+3424]; cvta.to.global.u64 %rd46, %rd45; add.s64 %rd47, %rd46, %rd20; ld.global.u8 %rs8, [%rd47+3]; and.b16 %rs9, %rs8, 64; setp.eq.s16 %p3, %rs9, 0; selp.b32 %r4, 768, 50331648, %p3; ld.const.u32 %r16, [kernel_params+1560]; setp.eq.s32 %p4, %r16, 0; @%p4 bra $L__BB206_5; cvt.u32.u64 %r17, %rd1; { // callseq 972, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r17; .param .b64 param1; st.param.b64 [param1+0], %rd11; .param .b32 param2; st.param.b32 [param2+0], %r4; .param .b32 retval0; call.uni (retval0), _Z38integrate_intersect_shadow_transparentPK16KernelGlobalsGPUiPK3Rayj, ( param0, param1, param2 ); ld.param.b32 %r18, [retval0+0]; } // callseq 972 cvt.u16.u32 %rs18, %r18; bra.uni $L__BB206_19; $L__BB206_5: and.b32 %r5, %r4, 16777472; setp.nan.ftz.f32 %p5, %f1, %f1; mov.u16 %rs18, 0; @%p5 bra $L__BB206_17; mov.b32 %r19, %f1; or.b32 %r20, %r19, -2147483648; setp.ne.s32 %p6, %r20, -2147483648; add.ftz.f32 %f11, %f1, %f1; setp.eq.ftz.f32 %p7, %f11, %f1; and.pred %p8, %p6, %p7; shl.b32 %r21, %r19, 1; setp.gt.u32 %p9, %r21, -16777216; or.pred %p10, %p9, %p8; @%p10 bra $L__BB206_17; setp.nan.ftz.f32 %p11, %f2, %f2; @%p11 bra $L__BB206_17; mov.b32 %r22, %f2; or.b32 %r23, %r22, -2147483648; setp.ne.s32 %p12, %r23, -2147483648; add.ftz.f32 %f12, %f2, %f2; setp.eq.ftz.f32 %p13, %f12, %f2; and.pred %p14, %p12, %p13; shl.b32 %r24, %r22, 1; setp.gt.u32 %p15, %r24, -16777216; or.pred %p16, %p15, %p14; @%p16 bra $L__BB206_17; mul.ftz.f32 %f13, %f3, %f3; fma.rn.ftz.f32 %f14, %f2, %f2, %f13; fma.rn.ftz.f32 %f15, %f4, %f4, %f14; setp.eq.ftz.f32 %p17, %f15, 0f00000000; @%p17 bra $L__BB206_17; bra.uni $L__BB206_10; $L__BB206_17: setp.ne.s16 %p21, %rs18, 0; @%p21 bra $L__BB206_19; ld.const.u64 %rd57, [kernel_params+3464]; cvta.to.global.u64 %rd58, %rd57; shl.b64 %rd59, %rd1, 1; add.s64 %rd60, %rd58, %rd59; mov.u16 %rs15, 0; st.global.u16 [%rd60], %rs15; $L__BB206_19: ld.const.u64 %rd61, [kernel_params+4256]; cvta.to.global.u64 %rd2, %rd61; add.s64 %rd62, %rd2, 12; atom.global.add.u32 %r30, [%rd62], -1; setp.eq.s16 %p22, %rs18, 0; @%p22 bra $L__BB206_21; ld.const.u64 %rd63, [kernel_params+3416]; cvta.to.global.u64 %rd64, %rd63; shl.b64 %rd65, %rd1, 1; add.s64 %rd66, %rd64, %rd65; mov.u16 %rs16, 0; st.global.u16 [%rd66], %rs16; bra.uni $L__BB206_22; $L__BB206_21: add.s64 %rd67, %rd2, 48; atom.global.add.u32 %r31, [%rd67], 1; ld.const.u64 %rd68, [kernel_params+3416]; cvta.to.global.u64 %rd69, %rd68; shl.b64 %rd70, %rd1, 1; add.s64 %rd71, %rd69, %rd70; mov.u16 %rs17, 12; st.global.u16 [%rd71], %rs17; $L__BB206_22: ret; $L__BB206_10: ld.const.u32 %r25, [kernel_params+1108]; setp.eq.s32 %p18, %r25, 0; ld.const.u32 %r6, [kernel_params+1112]; @%p18 bra $L__BB206_14; setp.ne.s32 %p19, %r6, 0; @%p19 bra $L__BB206_13; bra.uni $L__BB206_12; $L__BB206_13: add.u64 %rd52, %SP, 0; { // callseq 974, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd11; .param .b64 param1; st.param.b64 [param1+0], %rd52; .param .b32 param2; st.param.b32 [param2+0], %r5; .param .b32 retval0; call.uni (retval0), _Z29BVH_bvh_intersect_hair_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r27, [retval0+0]; } // callseq 974 cvt.u16.u32 %rs18, %r27; bra.uni $L__BB206_17; $L__BB206_14: setp.ne.s32 %p20, %r6, 0; @%p20 bra $L__BB206_16; bra.uni $L__BB206_15; $L__BB206_16: add.u64 %rd56, %SP, 0; { // callseq 976, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd11; .param .b64 param1; st.param.b64 [param1+0], %rd56; .param .b32 param2; st.param.b32 [param2+0], %r5; .param .b32 retval0; call.uni (retval0), _Z22BVH_bvh_intersect_hairPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r29, [retval0+0]; } // callseq 976 cvt.u16.u32 %rs18, %r29; bra.uni $L__BB206_17; $L__BB206_12: add.u64 %rd50, %SP, 0; { // callseq 973, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd11; .param .b64 param1; st.param.b64 [param1+0], %rd50; .param .b32 param2; st.param.b32 [param2+0], %r5; .param .b32 retval0; call.uni (retval0), _Z24BVH_bvh_intersect_motionPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r26, [retval0+0]; } // callseq 973 cvt.u16.u32 %rs18, %r26; bra.uni $L__BB206_17; $L__BB206_15: add.u64 %rd54, %SP, 0; { // callseq 975, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd11; .param .b64 param1; st.param.b64 [param1+0], %rd54; .param .b32 param2; st.param.b32 [param2+0], %r5; .param .b32 retval0; call.uni (retval0), _Z17BVH_bvh_intersectPK16KernelGlobalsGPUPK3RayP12Intersectionj, ( param0, param1, param2 ); ld.param.b32 %r28, [retval0+0]; } // callseq 975 cvt.u16.u32 %rs18, %r28; bra.uni $L__BB206_17; } // .globl kernel_gpu_integrator_intersect_subsurface .visible .entry kernel_gpu_integrator_intersect_subsurface( .param .u64 kernel_gpu_integrator_intersect_subsurface_param_0, .param .u32 kernel_gpu_integrator_intersect_subsurface_param_1 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot207[976]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<285>; .reg .b16 %rs<54>; .reg .f32 %f<2577>; .reg .b32 %r<1016>; .reg .b64 %rd<474>; mov.u64 %SPL, __local_depot207; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd93, [kernel_gpu_integrator_intersect_subsurface_param_0]; ld.param.u32 %r179, [kernel_gpu_integrator_intersect_subsurface_param_1]; mov.u32 %r180, %ntid.x; mov.u32 %r181, %ctaid.x; mov.u32 %r182, %tid.x; mad.lo.s32 %r951, %r181, %r180, %r182; setp.ge.s32 %p9, %r951, %r179; @%p9 bra $L__BB207_229; setp.eq.s64 %p10, %rd93, 0; @%p10 bra $L__BB207_3; cvta.to.global.u64 %rd94, %rd93; mul.wide.s32 %rd95, %r951, 4; add.s64 %rd96, %rd94, %rd95; ld.global.u32 %r951, [%rd96]; $L__BB207_3: ld.const.v2.u64 {%rd97, %rd98}, [kernel_params+2528]; cvta.to.global.u64 %rd101, %rd97; cvt.s64.s32 %rd1, %r951; mul.wide.s32 %rd102, %r951, 4; add.s64 %rd103, %rd101, %rd102; cvta.to.global.u64 %rd104, %rd98; mul.wide.s32 %rd105, %r951, 2; add.s64 %rd2, %rd104, %rd105; ld.global.u16 %r959, [%rd2]; ld.const.u64 %rd106, [kernel_params+2456]; cvta.to.global.u64 %rd107, %rd106; add.s64 %rd108, %rd107, %rd102; ld.const.u64 %rd109, [kernel_params+2544]; cvta.to.global.u64 %rd110, %rd109; add.s64 %rd111, %rd110, %rd102; ld.global.u8 %rs16, [%rd111+2]; and.b16 %rs17, %rs16, 16; setp.eq.s16 %p11, %rs17, 0; ld.global.u32 %r5, [%rd103]; ld.global.u32 %r6, [%rd108]; @%p11 bra $L__BB207_212; bra.uni $L__BB207_4; $L__BB207_212: cvt.u32.u64 %r911, %rd1; add.u64 %rd299, %SP, 768; add.u64 %rd300, %SP, 828; { // callseq 977, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r911; .param .align 4 .b8 param1[12]; st.param.b32 [param1+0], %r5; st.param.b32 [param1+4], %r959; st.param.b32 [param1+8], %r6; .param .b64 param2; st.param.b64 [param2+0], %rd299; .param .b64 param3; st.param.b64 [param3+0], %rd300; .param .b32 retval0; call.uni (retval0), _Z15subsurface_diskPK16KernelGlobalsGPUi8RNGStateR3RayR17LocalIntersection, ( param0, param1, param2, param3 ); ld.param.b32 %r912, [retval0+0]; } // callseq 977 cvt.u16.u32 %rs33, %r912; setp.eq.s16 %p254, %rs33, 0; mov.pred %p284, -1; @%p254 bra $L__BB207_227; $L__BB207_213: ld.const.u32 %r913, [kernel_params+1608]; setp.eq.s32 %p255, %r913, 0; @%p255 bra $L__BB207_216; add.u64 %rd302, %SPL, 828; ld.const.u64 %rd303, [kernel_params+2184]; cvta.to.global.u64 %rd304, %rd303; ld.local.u32 %r914, [%rd302+20]; mul.wide.s32 %rd305, %r914, 4; add.s64 %rd306, %rd304, %rd305; ld.global.u8 %rs34, [%rd306]; and.b16 %rs35, %rs34, 32; setp.eq.s16 %p256, %rs35, 0; @%p256 bra $L__BB207_216; cvt.u32.u64 %r915, %rd1; ld.const.u64 %rd307, [kernel_params+2632]; cvta.to.global.u64 %rd308, %rd307; mul.lo.s64 %rd309, %rd1, 12; add.s64 %rd310, %rd308, %rd309; ld.global.f32 %f2327, [%rd310+8]; ld.global.f32 %f2328, [%rd310+4]; ld.global.f32 %f2329, [%rd310]; add.u64 %rd312, %SPL, 768; ld.local.f32 %f2330, [%rd312+8]; ld.local.f32 %f2331, [%rd312+4]; ld.local.f32 %f2332, [%rd312]; { // callseq 978, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r915; .param .align 4 .b8 param1[12]; st.param.f32 [param1+0], %f2329; st.param.f32 [param1+4], %f2328; st.param.f32 [param1+8], %f2327; .param .align 4 .b8 param2[12]; st.param.f32 [param2+0], %f2332; st.param.f32 [param2+4], %f2331; st.param.f32 [param2+8], %f2330; call.uni _Z45integrator_volume_stack_update_for_subsurfacePK16KernelGlobalsGPUi6float3S2_, ( param0, param1, param2 ); } // callseq 978 $L__BB207_216: add.u64 %rd314, %SPL, 768; ld.local.f32 %f2333, [%rd314+12]; ld.local.f32 %f2334, [%rd314+28]; mul.ftz.f32 %f2335, %f2334, %f2333; ld.local.f32 %f2336, [%rd314+16]; mul.ftz.f32 %f2337, %f2334, %f2336; ld.local.f32 %f2338, [%rd314+20]; mul.ftz.f32 %f2339, %f2334, %f2338; ld.local.f32 %f2340, [%rd314]; fma.rn.ftz.f32 %f2341, %f2335, 0f40000000, %f2340; ld.local.f32 %f2342, [%rd314+4]; fma.rn.ftz.f32 %f2343, %f2337, 0f40000000, %f2342; ld.local.f32 %f2344, [%rd314+8]; fma.rn.ftz.f32 %f2345, %f2339, 0f40000000, %f2344; st.local.f32 [%rd314], %f2341; st.local.f32 [%rd314+4], %f2343; st.local.f32 [%rd314+8], %f2345; neg.ftz.f32 %f2346, %f2333; neg.ftz.f32 %f2347, %f2336; neg.ftz.f32 %f2348, %f2338; st.local.f32 [%rd314+12], %f2346; st.local.f32 [%rd314+16], %f2347; st.local.f32 [%rd314+20], %f2348; add.u64 %rd316, %SPL, 828; ld.local.f32 %f2349, [%rd316+4]; ld.const.v2.u64 {%rd317, %rd318}, [kernel_params+2688]; cvta.to.global.u64 %rd321, %rd317; shl.b64 %rd322, %rd1, 2; add.s64 %rd323, %rd321, %rd322; st.global.f32 [%rd323], %f2349; ld.local.f32 %f2350, [%rd316+8]; cvta.to.global.u64 %rd324, %rd318; add.s64 %rd325, %rd324, %rd322; st.global.f32 [%rd325], %f2350; ld.local.f32 %f2351, [%rd316+12]; ld.const.v2.u64 {%rd326, %rd327}, [kernel_params+2704]; cvta.to.global.u64 %rd330, %rd326; add.s64 %rd331, %rd330, %rd322; st.global.f32 [%rd331], %f2351; ld.local.u32 %r170, [%rd316+20]; ld.const.v2.u64 {%rd332, %rd333}, [kernel_params+2720]; cvta.to.global.u64 %rd336, %rd332; add.s64 %rd337, %rd336, %rd322; st.global.u32 [%rd337], %r170; ld.local.u32 %r171, [%rd316+16]; cvt.s64.s32 %rd90, %r171; cvta.to.global.u64 %rd338, %rd327; add.s64 %rd339, %rd338, %rd322; st.global.u32 [%rd339], %r171; ld.local.u32 %r172, [%rd316+24]; cvta.to.global.u64 %rd340, %rd333; add.s64 %rd341, %rd340, %rd322; st.global.u32 [%rd341], %r172; ld.const.u64 %rd342, [kernel_params+2632]; cvta.to.global.u64 %rd343, %rd342; mul.lo.s64 %rd344, %rd1, 12; add.s64 %rd345, %rd343, %rd344; st.global.f32 [%rd345], %f2341; st.global.f32 [%rd345+4], %f2343; st.global.f32 [%rd345+8], %f2345; ld.const.v2.u64 {%rd346, %rd347}, [kernel_params+2640]; cvta.to.global.u64 %rd350, %rd346; add.s64 %rd351, %rd350, %rd344; st.global.f32 [%rd351], %f2346; st.global.f32 [%rd351+4], %f2347; st.global.f32 [%rd351+8], %f2348; ld.local.f32 %f2352, [%rd314+24]; cvta.to.global.u64 %rd352, %rd347; add.s64 %rd353, %rd352, %rd322; st.global.f32 [%rd353], %f2352; ld.const.v2.u64 {%rd354, %rd355}, [kernel_params+2656]; cvta.to.global.u64 %rd358, %rd354; add.s64 %rd359, %rd358, %rd322; st.global.f32 [%rd359], %f2334; ld.local.f32 %f2353, [%rd314+32]; cvta.to.global.u64 %rd360, %rd355; add.s64 %rd361, %rd360, %rd322; st.global.f32 [%rd361], %f2353; ld.local.f32 %f2354, [%rd314+52]; ld.const.v2.u64 {%rd362, %rd363}, [kernel_params+2672]; cvta.to.global.u64 %rd366, %rd362; add.s64 %rd367, %rd366, %rd322; st.global.f32 [%rd367], %f2354; ld.local.f32 %f2355, [%rd314+56]; cvta.to.global.u64 %rd368, %rd363; add.s64 %rd369, %rd368, %rd322; st.global.f32 [%rd369], %f2355; ld.global.u16 %rs36, [%rd2]; add.s16 %rs37, %rs36, 16; st.global.u16 [%rd2], %rs37; and.b32 %r916, %r172, 1; setp.eq.b32 %p257, %r916, 1; mov.pred %p284, 0; xor.pred %p259, %p257, %p284; not.pred %p260, %p259; @%p260 bra $L__BB207_218; bra.uni $L__BB207_217; $L__BB207_218: and.b32 %r917, %r172, 8; setp.eq.s32 %p261, %r917, 0; @%p261 bra $L__BB207_220; bra.uni $L__BB207_219; $L__BB207_220: and.b32 %r920, %r172, 6; setp.eq.s32 %p262, %r920, 0; mov.u32 %r1015, 0; @%p262 bra $L__BB207_222; ld.const.u64 %rd378, [kernel_params+2264]; cvta.to.global.u64 %rd379, %rd378; mul.wide.s32 %rd380, %r171, 16; add.s64 %rd381, %rd379, %rd380; ld.global.u32 %r1015, [%rd381]; bra.uni $L__BB207_222; $L__BB207_4: ld.const.u32 %r7, [kernel_params+1596]; setp.eq.s32 %p12, %r7, 0; brev.b32 %r183, %r6; mul.lo.s32 %r184, %r183, 1025551850; xor.b32 %r8, %r184, %r183; @%p12 bra $L__BB207_7; ld.const.v2.u32 {%r185, %r186}, [kernel_params+1600]; mov.b32 %f1, %r186; setp.geu.ftz.f32 %p13, %f1, 0f3F800000; setp.lt.ftz.f32 %p14, %f1, 0f3F800000; ld.const.u32 %r189, [kernel_params+1576]; selp.b32 %r190, %r189, %r5, %p14; and.b32 %r191, %r959, 255; xor.b32 %r192, %r190, %r191; mul.lo.s32 %r193, %r192, -512718531; shr.u32 %r194, %r190, 16; xor.b32 %r195, %r193, %r194; bfe.u32 %r196, %r195, 4, 4; shr.u32 %r197, %r190, 8; xor.b32 %r198, %r195, %r197; xor.b32 %r199, %r198, %r196; mul.lo.s32 %r200, %r199, 153742143; shr.u32 %r201, %r190, 23; xor.b32 %r202, %r200, %r201; bfe.u32 %r203, %r202, 1, 7; xor.b32 %r204, %r203, %r202; shr.u32 %r205, %r190, 27; or.b32 %r206, %r205, 1; mul.lo.s32 %r207, %r206, %r204; mul.lo.s32 %r208, %r207, -1074439621; bfe.u32 %r209, %r208, 2, 6; xor.b32 %r210, %r209, %r208; mul.lo.s32 %r211, %r210, -1638916925; bfe.u32 %r212, %r211, 2, 6; xor.b32 %r213, %r212, %r211; mul.lo.s32 %r214, %r213, -933190689; and.b32 %r215, %r214, 255; bfe.u32 %r216, %r214, 5, 3; xor.b32 %r217, %r216, %r215; add.s32 %r218, %r185, -1; xor.b32 %r9, %r959, 61; xor.b32 %r219, %r190, %r9; mul.lo.s32 %r220, %r219, 9; shr.u32 %r221, %r220, 4; xor.b32 %r222, %r221, %r220; mul.lo.s32 %r223, %r222, 668265261; add.s32 %r224, %r223, %r8; shr.u32 %r225, %r223, 16; or.b32 %r226, %r225, 1; mul.lo.s32 %r227, %r226, %r224; mul.lo.s32 %r228, %r227, 89287766; xor.b32 %r229, %r228, %r227; mul.lo.s32 %r230, %r229, 1403136100; xor.b32 %r231, %r230, %r229; brev.b32 %r232, %r231; neg.s32 %r233, %r185; and.b32 %r234, %r6, %r233; and.b32 %r235, %r232, %r218; or.b32 %r236, %r235, %r234; mad.lo.s32 %r237, %r217, %r185, %r236; shl.b32 %r238, %r185, 8; rem.u32 %r239, %r237, %r238; ld.const.u64 %rd112, [kernel_params+2424]; cvta.to.global.u64 %rd113, %rd112; shl.b32 %r240, %r239, 1; mul.wide.u32 %rd114, %r240, 4; add.s64 %rd115, %rd113, %rd114; ld.global.f32 %f2369, [%rd115]; ld.global.f32 %f2370, [%rd115+4]; @%p13 bra $L__BB207_11; xor.b32 %r241, %r9, %r5; mul.lo.s32 %r242, %r241, 9; shr.u32 %r243, %r242, 4; xor.b32 %r244, %r243, %r242; mul.lo.s32 %r245, %r244, 668265261; cvt.rn.f32.u32 %f606, %r245; mul.ftz.f32 %f607, %f606, 0f2F7FFFFE; xor.b32 %r246, %r5, %r959; xor.b32 %r247, %r246, -905047700; mul.lo.s32 %r248, %r247, 9; shr.u32 %r249, %r248, 4; xor.b32 %r250, %r249, %r248; mul.lo.s32 %r251, %r250, 668265261; cvt.rn.f32.u32 %f608, %r251; mul.ftz.f32 %f609, %f608, 0f2F7FFFFE; fma.rn.ftz.f32 %f610, %f607, %f1, %f2369; fma.rn.ftz.f32 %f611, %f609, %f1, %f2370; cvt.rmi.ftz.f32.f32 %f612, %f610; sub.ftz.f32 %f2369, %f610, %f612; cvt.rmi.ftz.f32.f32 %f613, %f611; sub.ftz.f32 %f2370, %f611, %f613; bra.uni $L__BB207_11; $L__BB207_7: mul.lo.s32 %r253, %r959, 569420461; shr.u32 %r254, %r253, 15; xor.b32 %r255, %r254, %r253; mul.lo.s32 %r256, %r255, -749064809; shr.u32 %r257, %r256, 15; xor.b32 %r258, %r5, %r256; xor.b32 %r259, %r258, %r257; xor.b32 %r10, %r259, -419546133; xor.b32 %r260, %r259, 508809841; add.s32 %r261, %r260, %r8; shr.u32 %r262, %r260, 16; or.b32 %r263, %r262, 1; mul.lo.s32 %r264, %r263, %r261; mul.lo.s32 %r265, %r264, 89287766; xor.b32 %r266, %r265, %r264; mul.lo.s32 %r267, %r266, 1403136100; xor.b32 %r11, %r267, %r266; setp.eq.s32 %p15, %r11, 0; mov.u32 %r955, 0; @%p15 bra $L__BB207_10; mov.u32 %r952, 0; mov.u64 %rd117, sobol_burley_table; mov.u32 %r953, %r11; mov.u32 %r955, %r952; $L__BB207_9: clz.b32 %r270, %r953; add.s32 %r271, %r270, %r952; mul.wide.u32 %rd116, %r271, 4; add.s64 %rd118, %rd117, %rd116; ld.const.u32 %r272, [%rd118+128]; xor.b32 %r955, %r272, %r955; add.s32 %r952, %r271, 1; shl.b32 %r273, %r953, %r270; shl.b32 %r953, %r273, 1; setp.ne.s32 %p16, %r953, 0; @%p16 bra $L__BB207_9; $L__BB207_10: xor.b32 %r274, %r10, -525684874; brev.b32 %r275, %r11; shr.u32 %r276, %r274, 16; or.b32 %r277, %r276, 1; mul.lo.s32 %r278, %r275, 1025551850; xor.b32 %r279, %r278, %r275; add.s32 %r280, %r279, %r274; mul.lo.s32 %r281, %r280, %r277; mul.lo.s32 %r282, %r281, 89287766; xor.b32 %r283, %r282, %r281; xor.b32 %r284, %r10, -1802089138; mul.lo.s32 %r285, %r955, 1025551850; xor.b32 %r286, %r285, %r955; add.s32 %r287, %r286, %r284; shr.u32 %r288, %r284, 16; or.b32 %r289, %r288, 1; mul.lo.s32 %r290, %r287, %r289; mul.lo.s32 %r291, %r290, 89287766; xor.b32 %r292, %r291, %r290; mul.lo.s32 %r293, %r292, 1403136100; xor.b32 %r294, %r293, %r292; brev.b32 %r295, %r294; cvt.rn.f32.u32 %f614, %r295; mul.lo.s32 %r296, %r283, 1403136100; xor.b32 %r297, %r296, %r283; brev.b32 %r298, %r297; cvt.rn.f32.u32 %f615, %r298; mul.ftz.f32 %f2370, %f614, 0f2F7FFFFE; mul.ftz.f32 %f2369, %f615, 0f2F7FFFFE; $L__BB207_11: ld.const.u64 %rd119, [kernel_params+2632]; cvta.to.global.u64 %rd120, %rd119; mul.lo.s64 %rd121, %rd1, 12; add.s64 %rd122, %rd120, %rd121; ld.global.f32 %f14, [%rd122]; ld.global.f32 %f15, [%rd122+4]; ld.global.f32 %f16, [%rd122+8]; ld.const.u64 %rd123, [kernel_params+2640]; cvta.to.global.u64 %rd124, %rd123; add.s64 %rd125, %rd124, %rd121; ld.const.u64 %rd126, [kernel_params+2760]; cvta.to.global.u64 %rd127, %rd126; add.s64 %rd128, %rd127, %rd121; ld.const.u64 %rd129, [kernel_params+2720]; cvta.to.global.u64 %rd130, %rd129; shl.b64 %rd131, %rd1, 2; add.s64 %rd132, %rd130, %rd131; ld.global.u32 %r19, [%rd132]; ld.global.f32 %f17, [%rd125]; ld.global.f32 %f18, [%rd125+4]; neg.ftz.f32 %f19, %f18; ld.global.f32 %f20, [%rd125+8]; neg.ftz.f32 %f21, %f20; mul.ftz.f32 %f616, %f2369, 0f40C90FDB; sqrt.approx.ftz.f32 %f617, %f2370; cos.approx.ftz.f32 %f618, %f616; mul.ftz.f32 %f619, %f617, %f618; sin.approx.ftz.f32 %f620, %f616; mul.ftz.f32 %f621, %f617, %f620; mul.ftz.f32 %f622, %f619, %f619; mov.f32 %f623, 0f3F800000; sub.ftz.f32 %f624, %f623, %f622; mul.ftz.f32 %f625, %f621, %f621; sub.ftz.f32 %f626, %f624, %f625; mov.f32 %f2384, 0f00000000; max.ftz.f32 %f628, %f626, %f2384; sqrt.approx.ftz.f32 %f629, %f628; setp.neu.ftz.f32 %p18, %f17, %f18; setp.neu.ftz.f32 %p19, %f17, %f20; or.pred %p1, %p18, %p19; sub.ftz.f32 %f630, %f18, %f20; selp.f32 %f631, %f20, %f21, %p1; selp.f32 %f632, %f19, %f18, %p1; add.ftz.f32 %f633, %f17, %f632; sub.ftz.f32 %f634, %f631, %f17; mul.ftz.f32 %f635, %f634, %f634; fma.rn.ftz.f32 %f636, %f630, %f630, %f635; fma.rn.ftz.f32 %f637, %f633, %f633, %f636; sqrt.approx.ftz.f32 %f638, %f637; div.approx.ftz.f32 %f639, %f630, %f638; div.approx.ftz.f32 %f640, %f634, %f638; div.approx.ftz.f32 %f641, %f633, %f638; mul.ftz.f32 %f642, %f641, %f18; mul.ftz.f32 %f643, %f640, %f20; sub.ftz.f32 %f644, %f643, %f642; mul.ftz.f32 %f645, %f639, %f20; mul.ftz.f32 %f646, %f641, %f17; sub.ftz.f32 %f647, %f646, %f645; mul.ftz.f32 %f648, %f640, %f17; mul.ftz.f32 %f649, %f639, %f18; sub.ftz.f32 %f650, %f649, %f648; mul.ftz.f32 %f651, %f621, %f644; mul.ftz.f32 %f652, %f621, %f647; mul.ftz.f32 %f653, %f621, %f650; fma.rn.ftz.f32 %f654, %f619, %f639, %f651; fma.rn.ftz.f32 %f655, %f619, %f640, %f652; fma.rn.ftz.f32 %f656, %f619, %f641, %f653; mul.ftz.f32 %f657, %f629, %f17; mul.ftz.f32 %f658, %f629, %f18; mul.ftz.f32 %f659, %f629, %f20; sub.ftz.f32 %f2429, %f654, %f657; sub.ftz.f32 %f2428, %f655, %f658; sub.ftz.f32 %f2427, %f656, %f659; ld.global.f32 %f660, [%rd128]; mul.ftz.f32 %f661, %f660, %f2429; ld.global.f32 %f662, [%rd128+4]; mul.ftz.f32 %f663, %f662, %f2428; neg.ftz.f32 %f664, %f663; sub.ftz.f32 %f665, %f664, %f661; ld.global.f32 %f666, [%rd128+8]; mul.ftz.f32 %f667, %f666, %f2427; sub.ftz.f32 %f668, %f665, %f667; setp.le.ftz.f32 %p20, %f668, 0f00000000; mov.pred %p284, -1; @%p20 bra $L__BB207_227; add.u64 %rd134, %SPL, 768; ld.const.u64 %rd135, [kernel_params+2672]; ld.const.u64 %rd136, [kernel_params+2664]; ld.const.u64 %rd137, [kernel_params+2712]; cvta.to.global.u64 %rd138, %rd137; add.s64 %rd140, %rd138, %rd131; ld.global.u32 %r957, [%rd140]; cvta.to.global.u64 %rd141, %rd136; add.s64 %rd142, %rd141, %rd131; ld.global.f32 %f25, [%rd142]; cvta.to.global.u64 %rd143, %rd135; add.s64 %rd144, %rd143, %rd131; ld.global.f32 %f669, [%rd144]; st.local.f32 [%rd134], %f14; st.local.f32 [%rd134+4], %f15; st.local.f32 [%rd134+8], %f16; st.local.f32 [%rd134+12], %f2429; st.local.f32 [%rd134+16], %f2428; st.local.f32 [%rd134+20], %f2427; mov.u32 %r958, 0; st.local.u32 [%rd134+24], %r958; mov.u32 %r300, 2139095039; st.local.u32 [%rd134+28], %r300; st.local.f32 [%rd134+32], %f25; st.local.f32 [%rd134+52], %f669; st.local.u32 [%rd134+56], %r958; st.local.u32 [%rd134+40], %r19; st.local.u32 [%rd134+36], %r957; mov.u32 %r301, -1; st.local.u32 [%rd134+48], %r301; st.local.u32 [%rd134+44], %r301; ld.const.v2.u64 {%rd145, %rd146}, [kernel_params+2736]; cvta.to.global.u64 %rd149, %rd145; add.s64 %rd151, %rd149, %rd121; ld.global.f32 %f26, [%rd151+4]; ld.global.f32 %f27, [%rd151+8]; cvta.to.global.u64 %rd152, %rd146; add.s64 %rd153, %rd152, %rd121; ld.global.f32 %f28, [%rd153]; ld.global.f32 %f29, [%rd153+4]; ld.global.f32 %f30, [%rd153+8]; ld.const.u64 %rd154, [kernel_params+2752]; cvta.to.global.u64 %rd155, %rd154; add.s64 %rd156, %rd155, %rd131; ld.const.u64 %rd157, [kernel_params+2584]; cvta.to.global.u64 %rd158, %rd157; add.s64 %rd3, %rd158, %rd121; ld.global.f32 %f31, [%rd3]; ld.global.f32 %f32, [%rd3+4]; ld.global.f32 %f33, [%rd3+8]; ld.global.f32 %f34, [%rd156]; mul.ftz.f32 %f670, %f34, %f34; mul.ftz.f32 %f671, %f34, %f670; mul.ftz.f32 %f672, %f34, %f671; mul.ftz.f32 %f673, %f34, %f672; mul.ftz.f32 %f674, %f34, %f673; mul.ftz.f32 %f675, %f34, %f674; fma.rn.ftz.f32 %f676, %f34, 0fBFA46AD8, 0f3FE9BC16; fma.rn.ftz.f32 %f677, %f670, 0fBFE64726, %f676; fma.rn.ftz.f32 %f678, %f671, 0f41131A59, %f677; fma.rn.ftz.f32 %f679, %f672, 0fC1B6928D, %f678; fma.rn.ftz.f32 %f680, %f673, 0f4200180D, %f679; fma.rn.ftz.f32 %f681, %f674, 0fC1BD0308, %f680; fma.rn.ftz.f32 %f35, %f675, 0f40E6BDCF, %f681; mul.ftz.f32 %f682, %f670, 0fC349D8D6; fma.rn.ftz.f32 %f683, %f34, 0f41F9317A, %f682; fma.rn.ftz.f32 %f684, %f671, 0f445264DD, %f683; fma.rn.ftz.f32 %f685, %f672, 0fC4FC42F9, %f684; fma.rn.ftz.f32 %f686, %f673, 0f452ABB73, %f685; fma.rn.ftz.f32 %f687, %f674, 0fC4F1ED41, %f686; fma.rn.ftz.f32 %f688, %f675, 0f440BC094, %f687; mul.ftz.f32 %f689, %f688, 0f3FB8AA3B; ex2.approx.ftz.f32 %f690, %f689; fma.rn.ftz.f32 %f36, %f690, 0f3E02699A, 0f409F8609; fma.rn.ftz.f32 %f691, %f34, 0fBECA16A7, 0f3F8C65F1; fma.rn.ftz.f32 %f692, %f670, 0f3F86BAFB, %f691; fma.rn.ftz.f32 %f693, %f671, 0fC10D6F27, %f692; fma.rn.ftz.f32 %f694, %f672, 0f41E6EA22, %f693; fma.rn.ftz.f32 %f695, %f673, 0fC23B856B, %f694; fma.rn.ftz.f32 %f696, %f674, 0f421A2940, %f695; fma.rn.ftz.f32 %f37, %f675, 0fC14B7D5B, %f696; fma.rn.ftz.f32 %f697, %f34, 0f3EB86522, 0f3EFE1C5F; fma.rn.ftz.f32 %f698, %f670, 0fC009B06D, %f697; fma.rn.ftz.f32 %f699, %f671, 0f418F1E16, %f698; fma.rn.ftz.f32 %f700, %f672, 0fC25D3190, %f699; fma.rn.ftz.f32 %f701, %f673, 0f42A421C8, %f700; fma.rn.ftz.f32 %f702, %f674, 0fC26A0ADB, %f701; fma.rn.ftz.f32 %f38, %f675, 0f417D90B6, %f702; mul.ftz.f32 %f703, %f670, 0fC41496D5; fma.rn.ftz.f32 %f704, %f34, 0f42997698, %f703; fma.rn.ftz.f32 %f705, %f671, 0f45190E22, %f704; fma.rn.ftz.f32 %f706, %f672, 0fC5AE457B, %f705; fma.rn.ftz.f32 %f707, %f673, 0f45DE64D0, %f706; fma.rn.ftz.f32 %f708, %f674, 0fC594DC5C, %f707; fma.rn.ftz.f32 %f709, %f675, 0f44A2F105, %f708; mul.ftz.f32 %f710, %f709, 0f3FB8AA3B; ex2.approx.ftz.f32 %f711, %f710; fma.rn.ftz.f32 %f39, %f711, 0f3B4B8EB2, 0f40876BC0; fma.rn.ftz.f32 %f712, %f34, 0fC0212958, 0f4019FC65; fma.rn.ftz.f32 %f713, %f670, 0f4112F58D, %f712; fma.rn.ftz.f32 %f714, %f671, 0fC29E7037, %f713; fma.rn.ftz.f32 %f715, %f672, 0f43818A9B, %f714; fma.rn.ftz.f32 %f716, %f673, 0fC3C9CE91, %f715; fma.rn.ftz.f32 %f717, %f674, 0f43976DB6, %f716; fma.rn.ftz.f32 %f40, %f675, 0fC2AEDFC5, %f717; sub.ftz.f32 %f41, %f623, %f34; ld.global.f32 %f42, [%rd151]; mul.ftz.f32 %f43, %f42, %f36; abs.ftz.f32 %f44, %f43; setp.leu.ftz.f32 %p21, %f44, 0f3F800000; mov.f32 %f2371, %f44; @%p21 bra $L__BB207_14; rcp.approx.ftz.f32 %f2371, %f44; $L__BB207_14: mul.ftz.f32 %f47, %f42, %f39; abs.ftz.f32 %f48, %f47; setp.leu.ftz.f32 %p22, %f48, 0f3F800000; mov.f32 %f2372, %f48; @%p22 bra $L__BB207_16; rcp.approx.ftz.f32 %f2372, %f48; $L__BB207_16: lg2.approx.ftz.f32 %f719, %f42; mul.ftz.f32 %f720, %f719, 0f3E800000; ex2.approx.ftz.f32 %f721, %f720; mul.rn.ftz.f32 %f722, %f2371, %f2371; add.ftz.f32 %f723, %f722, 0f41355DC0; mov.f32 %f724, 0f41E6BD60; fma.rn.ftz.f32 %f725, %f723, %f722, %f724; mov.f32 %f726, 0f419D92C8; fma.rn.ftz.f32 %f727, %f725, %f722, %f726; rcp.approx.ftz.f32 %f728, %f727; mov.f32 %f730, 0fC0B59883; mov.f32 %f731, 0fBF52C7EA; fma.rn.ftz.f32 %f732, %f722, %f731, %f730; mov.f32 %f733, 0fC0D21907; fma.rn.ftz.f32 %f734, %f732, %f722, %f733; mul.ftz.f32 %f735, %f722, %f734; mul.ftz.f32 %f736, %f2371, %f735; fma.rn.ftz.f32 %f737, %f736, %f728, %f2371; setp.gt.ftz.f32 %p23, %f44, 0f3F800000; mov.f32 %f738, 0f3FC90FDB; sub.ftz.f32 %f739, %f738, %f737; selp.f32 %f740, %f739, %f737, %p23; mul.rn.ftz.f32 %f741, %f2372, %f2372; fma.rn.ftz.f32 %f742, %f741, %f731, %f730; fma.rn.ftz.f32 %f743, %f742, %f741, %f733; mul.ftz.f32 %f744, %f741, %f743; mul.ftz.f32 %f745, %f2372, %f744; add.ftz.f32 %f746, %f741, 0f41355DC0; fma.rn.ftz.f32 %f747, %f746, %f741, %f724; fma.rn.ftz.f32 %f748, %f747, %f741, %f726; rcp.approx.ftz.f32 %f749, %f748; fma.rn.ftz.f32 %f750, %f745, %f749, %f2372; sub.ftz.f32 %f751, %f738, %f750; setp.gt.ftz.f32 %p24, %f48, 0f3F800000; selp.f32 %f752, %f751, %f750, %p24; mov.b32 %r302, %f752; mov.b32 %r303, %f47; and.b32 %r304, %r303, -2147483648; or.b32 %r305, %r304, %r302; mov.b32 %f753, %r305; setp.le.ftz.f32 %p25, %f48, 0f7F800000; selp.f32 %f754, %f753, %f752, %p25; lg2.approx.ftz.f32 %f755, %f754; mul.ftz.f32 %f756, %f40, %f755; ex2.approx.ftz.f32 %f757, %f756; mul.ftz.f32 %f758, %f38, %f721; mul.ftz.f32 %f759, %f758, %f757; sub.ftz.f32 %f760, %f623, %f721; mul.ftz.f32 %f761, %f35, %f760; setp.le.ftz.f32 %p26, %f44, 0f7F800000; mov.b32 %r306, %f43; and.b32 %r307, %r306, -2147483648; mov.b32 %r308, %f740; or.b32 %r309, %r307, %r308; mov.b32 %f762, %r309; selp.f32 %f763, %f762, %f740, %p26; lg2.approx.ftz.f32 %f764, %f763; mul.ftz.f32 %f765, %f37, %f764; ex2.approx.ftz.f32 %f766, %f765; fma.rn.ftz.f32 %f767, %f761, %f766, %f759; max.ftz.f32 %f769, %f767, %f2384; mov.f32 %f770, 0f3F7FFFEF; min.ftz.f32 %f2381, %f769, %f770; mov.f32 %f771, 0f24E69595; max.ftz.f32 %f772, %f28, %f771; rcp.approx.ftz.f32 %f773, %f772; div.approx.ftz.f32 %f52, %f773, %f41; mul.ftz.f32 %f53, %f26, %f36; abs.ftz.f32 %f54, %f53; setp.leu.ftz.f32 %p27, %f54, 0f3F800000; mov.f32 %f2373, %f54; @%p27 bra $L__BB207_18; rcp.approx.ftz.f32 %f2373, %f54; $L__BB207_18: mul.ftz.f32 %f57, %f26, %f39; abs.ftz.f32 %f58, %f57; setp.leu.ftz.f32 %p28, %f58, 0f3F800000; mov.f32 %f2374, %f58; @%p28 bra $L__BB207_20; rcp.approx.ftz.f32 %f2374, %f58; $L__BB207_20: lg2.approx.ftz.f32 %f774, %f26; mul.ftz.f32 %f775, %f774, 0f3E800000; ex2.approx.ftz.f32 %f776, %f775; mul.rn.ftz.f32 %f777, %f2373, %f2373; add.ftz.f32 %f778, %f777, 0f41355DC0; fma.rn.ftz.f32 %f780, %f778, %f777, %f724; fma.rn.ftz.f32 %f782, %f780, %f777, %f726; rcp.approx.ftz.f32 %f783, %f782; fma.rn.ftz.f32 %f787, %f777, %f731, %f730; fma.rn.ftz.f32 %f789, %f787, %f777, %f733; mul.ftz.f32 %f790, %f777, %f789; mul.ftz.f32 %f791, %f2373, %f790; fma.rn.ftz.f32 %f792, %f791, %f783, %f2373; setp.gt.ftz.f32 %p29, %f54, 0f3F800000; sub.ftz.f32 %f794, %f738, %f792; selp.f32 %f795, %f794, %f792, %p29; mul.rn.ftz.f32 %f796, %f2374, %f2374; fma.rn.ftz.f32 %f797, %f796, %f731, %f730; fma.rn.ftz.f32 %f798, %f797, %f796, %f733; mul.ftz.f32 %f799, %f796, %f798; mul.ftz.f32 %f800, %f2374, %f799; add.ftz.f32 %f801, %f796, 0f41355DC0; fma.rn.ftz.f32 %f802, %f801, %f796, %f724; fma.rn.ftz.f32 %f803, %f802, %f796, %f726; rcp.approx.ftz.f32 %f804, %f803; fma.rn.ftz.f32 %f805, %f800, %f804, %f2374; sub.ftz.f32 %f806, %f738, %f805; setp.gt.ftz.f32 %p30, %f58, 0f3F800000; selp.f32 %f807, %f806, %f805, %p30; mov.b32 %r310, %f807; mov.b32 %r311, %f57; and.b32 %r312, %r311, -2147483648; or.b32 %r313, %r312, %r310; mov.b32 %f808, %r313; setp.le.ftz.f32 %p31, %f58, 0f7F800000; selp.f32 %f809, %f808, %f807, %p31; lg2.approx.ftz.f32 %f810, %f809; mul.ftz.f32 %f811, %f40, %f810; ex2.approx.ftz.f32 %f812, %f811; mul.ftz.f32 %f813, %f38, %f776; mul.ftz.f32 %f814, %f813, %f812; sub.ftz.f32 %f815, %f623, %f776; mul.ftz.f32 %f816, %f35, %f815; setp.le.ftz.f32 %p32, %f54, 0f7F800000; mov.b32 %r314, %f53; and.b32 %r315, %r314, -2147483648; mov.b32 %r316, %f795; or.b32 %r317, %r315, %r316; mov.b32 %f817, %r317; selp.f32 %f818, %f817, %f795, %p32; lg2.approx.ftz.f32 %f819, %f818; mul.ftz.f32 %f820, %f37, %f819; ex2.approx.ftz.f32 %f821, %f820; fma.rn.ftz.f32 %f822, %f816, %f821, %f814; max.ftz.f32 %f824, %f822, %f2384; min.ftz.f32 %f2383, %f824, %f770; max.ftz.f32 %f827, %f29, %f771; rcp.approx.ftz.f32 %f828, %f827; div.approx.ftz.f32 %f62, %f828, %f41; mul.ftz.f32 %f63, %f27, %f36; abs.ftz.f32 %f64, %f63; setp.leu.ftz.f32 %p33, %f64, 0f3F800000; mov.f32 %f2375, %f64; @%p33 bra $L__BB207_22; rcp.approx.ftz.f32 %f2375, %f64; $L__BB207_22: cvt.s64.s32 %rd4, %r19; lg2.approx.ftz.f32 %f829, %f27; mul.ftz.f32 %f830, %f829, 0f3E800000; ex2.approx.ftz.f32 %f831, %f830; mul.rn.ftz.f32 %f832, %f2375, %f2375; fma.rn.ftz.f32 %f835, %f832, %f731, %f730; fma.rn.ftz.f32 %f837, %f835, %f832, %f733; mul.ftz.f32 %f838, %f832, %f837; mul.ftz.f32 %f839, %f2375, %f838; add.ftz.f32 %f840, %f832, 0f41355DC0; fma.rn.ftz.f32 %f842, %f840, %f832, %f724; fma.rn.ftz.f32 %f844, %f842, %f832, %f726; rcp.approx.ftz.f32 %f845, %f844; fma.rn.ftz.f32 %f847, %f839, %f845, %f2375; sub.ftz.f32 %f849, %f738, %f847; setp.gt.ftz.f32 %p34, %f64, 0f3F800000; selp.f32 %f850, %f849, %f847, %p34; mov.b32 %r318, %f850; mov.b32 %r319, %f63; and.b32 %r320, %r319, -2147483648; or.b32 %r321, %r320, %r318; mov.b32 %f851, %r321; setp.le.ftz.f32 %p35, %f64, 0f7F800000; selp.f32 %f852, %f851, %f850, %p35; lg2.approx.ftz.f32 %f853, %f852; mul.ftz.f32 %f854, %f37, %f853; ex2.approx.ftz.f32 %f855, %f854; sub.ftz.f32 %f856, %f623, %f831; mul.ftz.f32 %f857, %f35, %f856; mul.ftz.f32 %f67, %f857, %f855; mul.ftz.f32 %f68, %f38, %f831; mul.ftz.f32 %f69, %f27, %f39; abs.ftz.f32 %f70, %f69; setp.leu.ftz.f32 %p36, %f70, 0f3F800000; mov.f32 %f2376, %f70; @%p36 bra $L__BB207_24; rcp.approx.ftz.f32 %f2376, %f70; $L__BB207_24: mul.rn.ftz.f32 %f859, %f2376, %f2376; fma.rn.ftz.f32 %f862, %f859, %f731, %f730; fma.rn.ftz.f32 %f864, %f862, %f859, %f733; mul.ftz.f32 %f865, %f859, %f864; mul.ftz.f32 %f866, %f2376, %f865; add.ftz.f32 %f867, %f859, 0f41355DC0; fma.rn.ftz.f32 %f869, %f867, %f859, %f724; fma.rn.ftz.f32 %f871, %f869, %f859, %f726; rcp.approx.ftz.f32 %f872, %f871; fma.rn.ftz.f32 %f873, %f866, %f872, %f2376; sub.ftz.f32 %f875, %f738, %f873; setp.gt.ftz.f32 %p37, %f70, 0f3F800000; selp.f32 %f876, %f875, %f873, %p37; mov.b32 %r322, %f876; mov.b32 %r323, %f69; and.b32 %r324, %r323, -2147483648; or.b32 %r325, %r324, %r322; mov.b32 %f877, %r325; setp.le.ftz.f32 %p38, %f70, 0f7F800000; selp.f32 %f878, %f877, %f876, %p38; lg2.approx.ftz.f32 %f879, %f878; mul.ftz.f32 %f880, %f40, %f879; ex2.approx.ftz.f32 %f881, %f880; fma.rn.ftz.f32 %f882, %f68, %f881, %f67; max.ftz.f32 %f883, %f882, %f2384; min.ftz.f32 %f2385, %f883, %f770; max.ftz.f32 %f886, %f30, %f771; rcp.approx.ftz.f32 %f887, %f886; div.approx.ftz.f32 %f74, %f887, %f41; setp.eq.ftz.f32 %p39, %f42, 0f00000000; mov.f32 %f2380, %f2384; @%p39 bra $L__BB207_26; div.approx.ftz.f32 %f2380, %f31, %f42; $L__BB207_26: setp.eq.ftz.f32 %p40, %f26, 0f00000000; mov.f32 %f2382, %f2384; @%p40 bra $L__BB207_28; div.approx.ftz.f32 %f2382, %f32, %f26; $L__BB207_28: setp.eq.ftz.f32 %p41, %f27, 0f00000000; @%p41 bra $L__BB207_30; div.approx.ftz.f32 %f2384, %f33, %f27; $L__BB207_30: setp.geu.ftz.f32 %p42, %f2381, 0f3E4CCCCD; @%p42 bra $L__BB207_32; mov.f32 %f890, 0f3E4CCCCD; div.approx.ftz.f32 %f891, %f2381, %f890; mul.ftz.f32 %f2380, %f891, %f2380; mov.f32 %f2381, %f890; $L__BB207_32: setp.geu.ftz.f32 %p43, %f2383, 0f3E4CCCCD; @%p43 bra $L__BB207_34; mov.f32 %f892, 0f3E4CCCCD; div.approx.ftz.f32 %f893, %f2383, %f892; mul.ftz.f32 %f2382, %f893, %f2382; mov.f32 %f2383, %f892; $L__BB207_34: setp.geu.ftz.f32 %p44, %f2385, 0f3E4CCCCD; @%p44 bra $L__BB207_36; mov.f32 %f894, 0f3E4CCCCD; div.approx.ftz.f32 %f895, %f2385, %f894; mul.ftz.f32 %f2384, %f895, %f2384; mov.f32 %f2385, %f894; $L__BB207_36: mul.ftz.f32 %f90, %f52, %f2381; mul.ftz.f32 %f91, %f62, %f2383; mul.ftz.f32 %f92, %f74, %f2385; max.ftz.f32 %f896, %f2381, %f2383; max.ftz.f32 %f897, %f896, %f2385; fma.rn.ftz.f32 %f898, %f897, 0fBCB0CB44, 0f401C5921; mov.f32 %f899, 0f3F14218E; div.approx.ftz.f32 %f900, %f899, %f897; add.ftz.f32 %f901, %f900, %f898; lg2.approx.ftz.f32 %f902, %f897; mul.ftz.f32 %f903, %f902, %f901; ex2.approx.ftz.f32 %f904, %f903; sub.ftz.f32 %f906, %f623, %f904; rsqrt.approx.ftz.f32 %f93, %f906; setp.eq.ftz.f32 %p46, %f93, 0f3F800000; @%p46 bra $L__BB207_227; add.ftz.f32 %f909, %f93, 0f3F800000; add.ftz.f32 %f911, %f93, 0fBF800000; div.approx.ftz.f32 %f912, %f909, %f911; lg2.approx.ftz.f32 %f913, %f912; mul.ftz.f32 %f94, %f913, 0f3F317218; add.s32 %r327, %r5, %r959; xor.b32 %r328, %r327, 701873903; shr.u32 %r329, %r328, 16; xor.b32 %r330, %r329, %r328; mul.lo.s32 %r331, %r330, 569420461; shr.u32 %r332, %r331, 15; xor.b32 %r333, %r332, %r331; mul.lo.s32 %r21, %r333, -749064809; shr.u32 %r334, %r21, 15; xor.b32 %r335, %r21, %r334; xor.b32 %r336, %r335, -419546133; abs.ftz.f32 %f914, %f34; lg2.approx.ftz.f32 %f915, %f914; mul.ftz.f32 %f916, %f915, 0f3E000000; ex2.approx.ftz.f32 %f917, %f916; mov.f32 %f918, 0f3F000000; max.ftz.f32 %f919, %f918, %f917; sub.ftz.f32 %f95, %f623, %f919; mul.ftz.f32 %f96, %f41, %f90; mul.ftz.f32 %f97, %f41, %f91; mul.ftz.f32 %f98, %f41, %f92; ld.const.v2.u32 {%r337, %r338}, [kernel_params+1600]; mov.b32 %f920, %r338; setp.lt.ftz.f32 %p47, %f920, 0f3F800000; ld.const.u32 %r339, [kernel_params+1576]; selp.b32 %r24, %r339, %r336, %p47; shr.u32 %r25, %r24, 16; shr.u32 %r26, %r24, 8; shr.u32 %r27, %r24, 23; shr.u32 %r340, %r24, 27; or.b32 %r341, %r340, 1; mul.lo.s32 %r28, %r341, -1074439621; neg.s32 %r342, %r337; and.b32 %r29, %r6, %r342; ld.const.u64 %rd159, [kernel_params+2424]; cvta.to.global.u64 %rd5, %rd159; selp.f32 %f921, %f21, %f20, %p1; add.ftz.f32 %f922, %f17, %f921; selp.f32 %f923, %f18, %f19, %p1; sub.ftz.f32 %f924, %f923, %f17; sub.ftz.f32 %f925, %f20, %f18; mul.ftz.f32 %f926, %f922, %f922; fma.rn.ftz.f32 %f927, %f925, %f925, %f926; fma.rn.ftz.f32 %f99, %f924, %f924, %f927; ld.const.u64 %rd160, [kernel_params+2144]; cvta.to.global.u64 %rd161, %rd160; shl.b64 %rd162, %rd4, 2; add.s64 %rd6, %rd161, %rd162; ld.const.v2.u64 {%rd163, %rd164}, [kernel_params+2176]; ld.const.u64 %rd165, [kernel_params+2160]; cvta.to.global.u64 %rd166, %rd165; mul.lo.s64 %rd167, %rd4, 208; add.s64 %rd168, %rd166, %rd167; add.s64 %rd9, %rd168, 48; ld.const.v2.u64 {%rd169, %rd170}, [kernel_params+2096]; ld.const.v2.u64 {%rd171, %rd172}, [kernel_params+2128]; ld.const.u64 %rd173, [kernel_params+2232]; cvta.to.global.u64 %rd14, %rd173; ld.const.u64 %rd174, [kernel_params+2256]; cvta.to.global.u64 %rd15, %rd174; ld.const.u64 %rd175, [kernel_params+2312]; cvta.to.global.u64 %rd16, %rd175; ld.const.u64 %rd176, [kernel_params+2336]; cvta.to.global.u64 %rd17, %rd176; mov.u16 %rs18, 0; mov.f32 %f2561, 0f00000000; cvta.to.global.u64 %rd206, %rd163; sqrt.approx.ftz.f32 %f1089, %f99; mov.u32 %r956, %r19; mov.f32 %f2390, %f16; mov.f32 %f2391, %f15; mov.f32 %f2392, %f14; mov.u16 %rs53, %rs18; $L__BB207_38: setp.lt.u32 %p48, %r958, 10; selp.f32 %f120, %f34, 0f00000000, %p48; @%p12 bra $L__BB207_41; setp.geu.ftz.f32 %p50, %f920, 0f3F800000; add.s32 %r343, %r959, 17; and.b32 %r344, %r343, 255; xor.b32 %r345, %r24, %r344; mul.lo.s32 %r346, %r345, -512718531; xor.b32 %r347, %r346, %r25; bfe.u32 %r348, %r347, 4, 4; xor.b32 %r349, %r347, %r26; xor.b32 %r350, %r349, %r348; mul.lo.s32 %r351, %r350, 153742143; xor.b32 %r352, %r351, %r27; bfe.u32 %r353, %r352, 1, 7; xor.b32 %r354, %r353, %r352; mul.lo.s32 %r355, %r28, %r354; bfe.u32 %r356, %r355, 2, 6; xor.b32 %r357, %r356, %r355; mul.lo.s32 %r358, %r357, -1638916925; bfe.u32 %r359, %r358, 2, 6; xor.b32 %r360, %r359, %r358; mul.lo.s32 %r361, %r360, -933190689; and.b32 %r362, %r361, 255; bfe.u32 %r363, %r361, 5, 3; xor.b32 %r364, %r363, %r362; xor.b32 %r35, %r343, 61; xor.b32 %r365, %r24, %r35; mul.lo.s32 %r366, %r365, 9; shr.u32 %r367, %r366, 4; xor.b32 %r368, %r367, %r366; mul.lo.s32 %r369, %r368, 668265261; add.s32 %r370, %r369, %r8; shr.u32 %r371, %r369, 16; or.b32 %r372, %r371, 1; mul.lo.s32 %r373, %r372, %r370; mul.lo.s32 %r374, %r373, 89287766; xor.b32 %r375, %r374, %r373; mul.lo.s32 %r376, %r375, 1403136100; xor.b32 %r377, %r376, %r375; brev.b32 %r378, %r377; add.s32 %r379, %r337, -1; and.b32 %r380, %r378, %r379; or.b32 %r381, %r380, %r29; mad.lo.s32 %r382, %r364, %r337, %r381; shl.b32 %r383, %r337, 8; rem.u32 %r384, %r382, %r383; shl.b32 %r385, %r384, 1; mul.wide.u32 %rd177, %r385, 4; add.s64 %rd178, %rd5, %rd177; ld.global.f32 %f2406, [%rd178]; @%p50 bra $L__BB207_42; xor.b32 %r388, %r35, %r335; xor.b32 %r389, %r388, -419546133; mul.lo.s32 %r390, %r389, 9; shr.u32 %r391, %r390, 4; xor.b32 %r392, %r391, %r390; mul.lo.s32 %r393, %r392, 668265261; cvt.rn.f32.u32 %f929, %r393; mul.ftz.f32 %f930, %f929, 0f2F7FFFFE; fma.rn.ftz.f32 %f932, %f930, %f920, %f2406; cvt.rmi.ftz.f32.f32 %f933, %f932; sub.ftz.f32 %f2406, %f932, %f933; bra.uni $L__BB207_42; $L__BB207_41: add.s32 %r394, %r959, 17; shr.u32 %r395, %r394, 16; xor.b32 %r396, %r395, %r394; mul.lo.s32 %r397, %r396, 569420461; shr.u32 %r398, %r397, 15; xor.b32 %r399, %r398, %r397; mul.lo.s32 %r400, %r399, -749064809; shr.u32 %r401, %r400, 15; xor.b32 %r403, %r334, %r21; xor.b32 %r404, %r403, %r400; xor.b32 %r405, %r404, %r401; xor.b32 %r406, %r405, -1074177026; add.s32 %r407, %r406, %r8; shr.u32 %r408, %r406, 16; or.b32 %r409, %r408, 1; mul.lo.s32 %r410, %r409, %r407; mul.lo.s32 %r411, %r410, 89287766; xor.b32 %r412, %r411, %r410; mul.lo.s32 %r413, %r412, 1403136100; xor.b32 %r414, %r413, %r412; xor.b32 %r415, %r405, 1667004349; brev.b32 %r416, %r414; mul.lo.s32 %r417, %r416, 1025551850; xor.b32 %r418, %r417, %r416; add.s32 %r419, %r418, %r415; shr.u32 %r420, %r415, 16; or.b32 %r421, %r420, 1; mul.lo.s32 %r422, %r419, %r421; mul.lo.s32 %r423, %r422, 89287766; xor.b32 %r424, %r423, %r422; mul.lo.s32 %r425, %r424, 1403136100; xor.b32 %r426, %r425, %r424; brev.b32 %r427, %r426; cvt.rn.f32.u32 %f934, %r427; mul.ftz.f32 %f2406, %f934, 0f2F7FFFFE; $L__BB207_42: mul.ftz.f32 %f938, %f2381, %f2380; abs.ftz.f32 %f125, %f938; mul.ftz.f32 %f939, %f2383, %f2382; abs.ftz.f32 %f126, %f939; mul.ftz.f32 %f940, %f2385, %f2384; abs.ftz.f32 %f127, %f940; add.ftz.f32 %f941, %f125, %f126; add.ftz.f32 %f128, %f127, %f941; setp.leu.ftz.f32 %p51, %f128, 0f00000000; mov.f32 %f2407, 0f3EAAAAAB; mov.f32 %f2408, %f2407; mov.f32 %f2409, %f2407; @%p51 bra $L__BB207_44; div.approx.ftz.f32 %f2407, %f125, %f128; div.approx.ftz.f32 %f2408, %f126, %f128; div.approx.ftz.f32 %f2409, %f127, %f128; $L__BB207_44: add.ftz.f32 %f135, %f2407, 0f00000000; setp.gt.ftz.f32 %p52, %f135, %f2406; mov.u64 %rd454, 0; @%p52 bra $L__BB207_47; add.ftz.f32 %f942, %f135, %f2408; setp.gt.ftz.f32 %p53, %f942, %f2406; mov.u64 %rd454, 1; @%p53 bra $L__BB207_47; mov.u64 %rd454, 2; $L__BB207_47: add.u64 %rd183, %SPL, 0; sub.ftz.f32 %f943, %f52, %f90; add.ftz.f32 %f944, %f96, %f943; selp.f32 %f945, %f52, %f944, %p48; st.local.f32 [%rd183], %f945; sub.ftz.f32 %f946, %f62, %f91; add.ftz.f32 %f947, %f97, %f946; selp.f32 %f948, %f62, %f947, %p48; st.local.f32 [%rd183+4], %f948; sub.ftz.f32 %f949, %f74, %f92; add.ftz.f32 %f950, %f98, %f949; selp.f32 %f951, %f74, %f950, %p48; st.local.f32 [%rd183+8], %f951; shl.b64 %rd184, %rd454, 2; add.s64 %rd185, %rd183, %rd184; ld.local.f32 %f2430, [%rd185]; add.s32 %r36, %r959, 18; @%p12 bra $L__BB207_50; setp.geu.ftz.f32 %p56, %f920, 0f3F800000; and.b32 %r428, %r36, 255; xor.b32 %r429, %r24, %r428; mul.lo.s32 %r430, %r429, -512718531; xor.b32 %r431, %r430, %r25; bfe.u32 %r432, %r431, 4, 4; xor.b32 %r433, %r431, %r26; xor.b32 %r434, %r433, %r432; mul.lo.s32 %r435, %r434, 153742143; xor.b32 %r436, %r435, %r27; bfe.u32 %r437, %r436, 1, 7; xor.b32 %r438, %r437, %r436; mul.lo.s32 %r439, %r28, %r438; bfe.u32 %r440, %r439, 2, 6; xor.b32 %r441, %r440, %r439; mul.lo.s32 %r442, %r441, -1638916925; bfe.u32 %r443, %r442, 2, 6; xor.b32 %r444, %r443, %r442; mul.lo.s32 %r445, %r444, -933190689; and.b32 %r446, %r445, 255; bfe.u32 %r447, %r445, 5, 3; xor.b32 %r448, %r447, %r446; xor.b32 %r37, %r36, 61; xor.b32 %r449, %r24, %r37; mul.lo.s32 %r450, %r449, 9; shr.u32 %r451, %r450, 4; xor.b32 %r452, %r451, %r450; mul.lo.s32 %r453, %r452, 668265261; add.s32 %r454, %r453, %r8; shr.u32 %r455, %r453, 16; or.b32 %r456, %r455, 1; mul.lo.s32 %r457, %r456, %r454; mul.lo.s32 %r458, %r457, 89287766; xor.b32 %r459, %r458, %r457; mul.lo.s32 %r460, %r459, 1403136100; xor.b32 %r461, %r460, %r459; brev.b32 %r462, %r461; add.s32 %r463, %r337, -1; and.b32 %r464, %r462, %r463; or.b32 %r465, %r464, %r29; mad.lo.s32 %r466, %r448, %r337, %r465; shl.b32 %r467, %r337, 8; rem.u32 %r468, %r466, %r467; shl.b32 %r469, %r468, 1; mul.wide.u32 %rd186, %r469, 4; add.s64 %rd187, %rd5, %rd186; ld.global.f32 %f2410, [%rd187]; @%p56 bra $L__BB207_51; xor.b32 %r472, %r37, %r335; xor.b32 %r473, %r472, -419546133; mul.lo.s32 %r474, %r473, 9; shr.u32 %r475, %r474, 4; xor.b32 %r476, %r475, %r474; mul.lo.s32 %r477, %r476, 668265261; cvt.rn.f32.u32 %f953, %r477; mul.ftz.f32 %f954, %f953, 0f2F7FFFFE; fma.rn.ftz.f32 %f956, %f954, %f920, %f2410; cvt.rmi.ftz.f32.f32 %f957, %f956; sub.ftz.f32 %f2410, %f956, %f957; bra.uni $L__BB207_51; $L__BB207_50: shr.u32 %r478, %r36, 16; xor.b32 %r479, %r478, %r36; mul.lo.s32 %r480, %r479, 569420461; shr.u32 %r481, %r480, 15; xor.b32 %r482, %r481, %r480; mul.lo.s32 %r483, %r482, -749064809; shr.u32 %r484, %r483, 15; xor.b32 %r486, %r334, %r21; xor.b32 %r487, %r486, %r483; xor.b32 %r488, %r487, %r484; xor.b32 %r489, %r488, -1074177026; add.s32 %r490, %r489, %r8; shr.u32 %r491, %r489, 16; or.b32 %r492, %r491, 1; mul.lo.s32 %r493, %r492, %r490; mul.lo.s32 %r494, %r493, 89287766; xor.b32 %r495, %r494, %r493; mul.lo.s32 %r496, %r495, 1403136100; xor.b32 %r497, %r496, %r495; xor.b32 %r498, %r488, 1667004349; brev.b32 %r499, %r497; mul.lo.s32 %r500, %r499, 1025551850; xor.b32 %r501, %r500, %r499; add.s32 %r502, %r501, %r498; shr.u32 %r503, %r498, 16; or.b32 %r504, %r503, 1; mul.lo.s32 %r505, %r502, %r504; mul.lo.s32 %r506, %r505, 89287766; xor.b32 %r507, %r506, %r505; mul.lo.s32 %r508, %r507, 1403136100; xor.b32 %r509, %r508, %r507; brev.b32 %r510, %r509; cvt.rn.f32.u32 %f958, %r510; mul.ftz.f32 %f2410, %f958, 0f2F7FFFFE; $L__BB207_51: setp.eq.s32 %p57, %r958, 0; mov.f32 %f2431, 0f00000000; mov.f32 %f2432, %f2431; mov.f32 %f2433, %f623; mov.f32 %f2434, %f2431; mov.f32 %f2435, %f623; @%p57 bra $L__BB207_80; add.s32 %r38, %r959, 19; @%p12 bra $L__BB207_55; setp.geu.ftz.f32 %p59, %f920, 0f3F800000; and.b32 %r511, %r38, 255; xor.b32 %r512, %r24, %r511; mul.lo.s32 %r513, %r512, -512718531; xor.b32 %r514, %r513, %r25; bfe.u32 %r515, %r514, 4, 4; xor.b32 %r516, %r514, %r26; xor.b32 %r517, %r516, %r515; mul.lo.s32 %r518, %r517, 153742143; xor.b32 %r519, %r518, %r27; bfe.u32 %r520, %r519, 1, 7; xor.b32 %r521, %r520, %r519; mul.lo.s32 %r522, %r28, %r521; bfe.u32 %r523, %r522, 2, 6; xor.b32 %r524, %r523, %r522; mul.lo.s32 %r525, %r524, -1638916925; bfe.u32 %r526, %r525, 2, 6; xor.b32 %r527, %r526, %r525; mul.lo.s32 %r528, %r527, -933190689; and.b32 %r529, %r528, 255; bfe.u32 %r530, %r528, 5, 3; xor.b32 %r531, %r530, %r529; xor.b32 %r39, %r38, 61; xor.b32 %r532, %r24, %r39; mul.lo.s32 %r533, %r532, 9; shr.u32 %r534, %r533, 4; xor.b32 %r535, %r534, %r533; mul.lo.s32 %r536, %r535, 668265261; add.s32 %r537, %r536, %r8; shr.u32 %r538, %r536, 16; or.b32 %r539, %r538, 1; mul.lo.s32 %r540, %r539, %r537; mul.lo.s32 %r541, %r540, 89287766; xor.b32 %r542, %r541, %r540; mul.lo.s32 %r543, %r542, 1403136100; xor.b32 %r544, %r543, %r542; brev.b32 %r545, %r544; add.s32 %r546, %r337, -1; and.b32 %r547, %r545, %r546; or.b32 %r548, %r547, %r29; mad.lo.s32 %r549, %r531, %r337, %r548; shl.b32 %r550, %r337, 8; rem.u32 %r551, %r549, %r550; shl.b32 %r552, %r551, 1; mul.wide.u32 %rd188, %r552, 4; add.s64 %rd189, %rd5, %rd188; ld.global.f32 %f2411, [%rd189]; @%p59 bra $L__BB207_56; xor.b32 %r555, %r39, %r335; xor.b32 %r556, %r555, -419546133; mul.lo.s32 %r557, %r556, 9; shr.u32 %r558, %r557, 4; xor.b32 %r559, %r558, %r557; mul.lo.s32 %r560, %r559, 668265261; cvt.rn.f32.u32 %f965, %r560; mul.ftz.f32 %f966, %f965, 0f2F7FFFFE; fma.rn.ftz.f32 %f968, %f966, %f920, %f2411; cvt.rmi.ftz.f32.f32 %f969, %f968; sub.ftz.f32 %f2411, %f968, %f969; bra.uni $L__BB207_56; $L__BB207_55: shr.u32 %r561, %r38, 16; xor.b32 %r562, %r561, %r38; mul.lo.s32 %r563, %r562, 569420461; shr.u32 %r564, %r563, 15; xor.b32 %r565, %r564, %r563; mul.lo.s32 %r566, %r565, -749064809; shr.u32 %r567, %r566, 15; xor.b32 %r569, %r334, %r21; xor.b32 %r570, %r569, %r566; xor.b32 %r571, %r570, %r567; xor.b32 %r572, %r571, -1074177026; add.s32 %r573, %r572, %r8; shr.u32 %r574, %r572, 16; or.b32 %r575, %r574, 1; mul.lo.s32 %r576, %r575, %r573; mul.lo.s32 %r577, %r576, 89287766; xor.b32 %r578, %r577, %r576; mul.lo.s32 %r579, %r578, 1403136100; xor.b32 %r580, %r579, %r578; xor.b32 %r581, %r571, 1667004349; brev.b32 %r582, %r580; mul.lo.s32 %r583, %r582, 1025551850; xor.b32 %r584, %r583, %r582; add.s32 %r585, %r584, %r581; shr.u32 %r586, %r581, 16; or.b32 %r587, %r586, 1; mul.lo.s32 %r588, %r585, %r587; mul.lo.s32 %r589, %r588, 89287766; xor.b32 %r590, %r589, %r588; mul.lo.s32 %r591, %r590, 1403136100; xor.b32 %r592, %r591, %r590; brev.b32 %r593, %r592; cvt.rn.f32.u32 %f970, %r593; mul.ftz.f32 %f2411, %f970, 0f2F7FFFFE; $L__BB207_56: and.b16 %rs20, %rs53, 255; setp.eq.s16 %p60, %rs20, 0; mov.u16 %rs46, 0; mov.f32 %f2431, 0f00000000; @%p60 bra $L__BB207_62; neg.ftz.f32 %f972, %f17; sub.ftz.f32 %f973, %f2392, %f14; mul.ftz.f32 %f974, %f973, %f972; sub.ftz.f32 %f975, %f2391, %f15; mul.ftz.f32 %f976, %f975, %f18; sub.ftz.f32 %f977, %f974, %f976; sub.ftz.f32 %f978, %f2390, %f16; mul.ftz.f32 %f979, %f978, %f20; sub.ftz.f32 %f980, %f977, %f979; mov.f32 %f981, 0f00000000; max.ftz.f32 %f982, %f980, %f981; min.ftz.f32 %f983, %f982, %f2561; add.ftz.f32 %f984, %f983, %f983; sub.ftz.f32 %f985, %f2561, %f984; div.approx.ftz.f32 %f986, %f985, %f93; mul.ftz.f32 %f987, %f986, 0f3FB8AA3B; ex2.approx.ftz.f32 %f988, %f987; add.ftz.f32 %f989, %f988, 0f3F800000; rcp.approx.ftz.f32 %f2431, %f989; add.s32 %r40, %r959, 20; @%p12 bra $L__BB207_60; setp.geu.ftz.f32 %p62, %f920, 0f3F800000; and.b32 %r594, %r40, 255; xor.b32 %r595, %r24, %r594; mul.lo.s32 %r596, %r595, -512718531; xor.b32 %r597, %r596, %r25; bfe.u32 %r598, %r597, 4, 4; xor.b32 %r599, %r597, %r26; xor.b32 %r600, %r599, %r598; mul.lo.s32 %r601, %r600, 153742143; xor.b32 %r602, %r601, %r27; bfe.u32 %r603, %r602, 1, 7; xor.b32 %r604, %r603, %r602; mul.lo.s32 %r605, %r28, %r604; bfe.u32 %r606, %r605, 2, 6; xor.b32 %r607, %r606, %r605; mul.lo.s32 %r608, %r607, -1638916925; bfe.u32 %r609, %r608, 2, 6; xor.b32 %r610, %r609, %r608; mul.lo.s32 %r611, %r610, -933190689; and.b32 %r612, %r611, 255; bfe.u32 %r613, %r611, 5, 3; xor.b32 %r614, %r613, %r612; xor.b32 %r41, %r40, 61; xor.b32 %r615, %r24, %r41; mul.lo.s32 %r616, %r615, 9; shr.u32 %r617, %r616, 4; xor.b32 %r618, %r617, %r616; mul.lo.s32 %r619, %r618, 668265261; add.s32 %r620, %r619, %r8; shr.u32 %r621, %r619, 16; or.b32 %r622, %r621, 1; mul.lo.s32 %r623, %r622, %r620; mul.lo.s32 %r624, %r623, 89287766; xor.b32 %r625, %r624, %r623; mul.lo.s32 %r626, %r625, 1403136100; xor.b32 %r627, %r626, %r625; brev.b32 %r628, %r627; add.s32 %r629, %r337, -1; and.b32 %r630, %r628, %r629; or.b32 %r631, %r630, %r29; mad.lo.s32 %r632, %r614, %r337, %r631; shl.b32 %r633, %r337, 8; rem.u32 %r634, %r632, %r633; shl.b32 %r635, %r634, 1; mul.wide.u32 %rd190, %r635, 4; add.s64 %rd191, %rd5, %rd190; ld.global.f32 %f2412, [%rd191]; @%p62 bra $L__BB207_61; xor.b32 %r638, %r41, %r335; xor.b32 %r639, %r638, -419546133; mul.lo.s32 %r640, %r639, 9; shr.u32 %r641, %r640, 4; xor.b32 %r642, %r641, %r640; mul.lo.s32 %r643, %r642, 668265261; cvt.rn.f32.u32 %f991, %r643; mul.ftz.f32 %f992, %f991, 0f2F7FFFFE; fma.rn.ftz.f32 %f994, %f992, %f920, %f2412; cvt.rmi.ftz.f32.f32 %f995, %f994; sub.ftz.f32 %f2412, %f994, %f995; bra.uni $L__BB207_61; $L__BB207_60: shr.u32 %r644, %r40, 16; xor.b32 %r645, %r644, %r40; mul.lo.s32 %r646, %r645, 569420461; shr.u32 %r647, %r646, 15; xor.b32 %r648, %r647, %r646; mul.lo.s32 %r649, %r648, -749064809; shr.u32 %r650, %r649, 15; xor.b32 %r652, %r334, %r21; xor.b32 %r653, %r652, %r649; xor.b32 %r654, %r653, %r650; xor.b32 %r655, %r654, -1074177026; add.s32 %r656, %r655, %r8; shr.u32 %r657, %r655, 16; or.b32 %r658, %r657, 1; mul.lo.s32 %r659, %r658, %r656; mul.lo.s32 %r660, %r659, 89287766; xor.b32 %r661, %r660, %r659; mul.lo.s32 %r662, %r661, 1403136100; xor.b32 %r663, %r662, %r661; xor.b32 %r664, %r654, 1667004349; brev.b32 %r665, %r663; mul.lo.s32 %r666, %r665, 1025551850; xor.b32 %r667, %r666, %r665; add.s32 %r668, %r667, %r664; shr.u32 %r669, %r664, 16; or.b32 %r670, %r669, 1; mul.lo.s32 %r671, %r668, %r670; mul.lo.s32 %r672, %r671, 89287766; xor.b32 %r673, %r672, %r671; mul.lo.s32 %r674, %r673, 1403136100; xor.b32 %r675, %r674, %r673; brev.b32 %r676, %r675; cvt.rn.f32.u32 %f996, %r676; mul.ftz.f32 %f2412, %f996, 0f2F7FFFFE; $L__BB207_61: setp.lt.ftz.f32 %p63, %f2412, %f2431; selp.u16 %rs46, 1, 0, %p63; $L__BB207_62: @%p12 bra $L__BB207_65; setp.geu.ftz.f32 %p65, %f920, 0f3F800000; add.s32 %r677, %r959, 16; and.b32 %r678, %r677, 255; xor.b32 %r679, %r24, %r678; mul.lo.s32 %r680, %r679, -512718531; xor.b32 %r681, %r680, %r25; bfe.u32 %r682, %r681, 4, 4; xor.b32 %r683, %r681, %r26; xor.b32 %r684, %r683, %r682; mul.lo.s32 %r685, %r684, 153742143; xor.b32 %r686, %r685, %r27; bfe.u32 %r687, %r686, 1, 7; xor.b32 %r688, %r687, %r686; mul.lo.s32 %r689, %r28, %r688; bfe.u32 %r690, %r689, 2, 6; xor.b32 %r691, %r690, %r689; mul.lo.s32 %r692, %r691, -1638916925; bfe.u32 %r693, %r692, 2, 6; xor.b32 %r694, %r693, %r692; mul.lo.s32 %r695, %r694, -933190689; and.b32 %r696, %r695, 255; bfe.u32 %r697, %r695, 5, 3; xor.b32 %r698, %r697, %r696; xor.b32 %r42, %r677, 61; xor.b32 %r699, %r24, %r42; mul.lo.s32 %r700, %r699, 9; shr.u32 %r701, %r700, 4; xor.b32 %r702, %r701, %r700; mul.lo.s32 %r703, %r702, 668265261; add.s32 %r704, %r703, %r8; shr.u32 %r705, %r703, 16; or.b32 %r706, %r705, 1; mul.lo.s32 %r707, %r706, %r704; mul.lo.s32 %r708, %r707, 89287766; xor.b32 %r709, %r708, %r707; mul.lo.s32 %r710, %r709, 1403136100; xor.b32 %r711, %r710, %r709; brev.b32 %r712, %r711; add.s32 %r713, %r337, -1; and.b32 %r714, %r712, %r713; or.b32 %r715, %r714, %r29; mad.lo.s32 %r716, %r698, %r337, %r715; shl.b32 %r717, %r337, 8; rem.u32 %r718, %r716, %r717; shl.b32 %r719, %r718, 1; mul.wide.u32 %rd192, %r719, 4; add.s64 %rd193, %rd5, %rd192; ld.global.f32 %f161, [%rd193]; ld.global.f32 %f2417, [%rd193+4]; @%p65 bra $L__BB207_69; xor.b32 %r947, %r335, -419546133; xor.b32 %r723, %r42, %r947; mul.lo.s32 %r724, %r723, 9; shr.u32 %r725, %r724, 4; xor.b32 %r726, %r725, %r724; mul.lo.s32 %r727, %r726, 668265261; cvt.rn.f32.u32 %f998, %r727; mul.ftz.f32 %f999, %f998, 0f2F7FFFFE; xor.b32 %r729, %r947, %r677; xor.b32 %r730, %r729, -905047700; mul.lo.s32 %r731, %r730, 9; shr.u32 %r732, %r731, 4; xor.b32 %r733, %r732, %r731; mul.lo.s32 %r734, %r733, 668265261; cvt.rn.f32.u32 %f1001, %r734; mul.ftz.f32 %f1002, %f1001, 0f2F7FFFFE; fma.rn.ftz.f32 %f1003, %f999, %f920, %f161; fma.rn.ftz.f32 %f1004, %f1002, %f920, %f2417; cvt.rmi.ftz.f32.f32 %f1005, %f1003; sub.ftz.f32 %f161, %f1003, %f1005; cvt.rmi.ftz.f32.f32 %f1006, %f1004; sub.ftz.f32 %f2417, %f1004, %f1006; bra.uni $L__BB207_69; $L__BB207_65: add.s32 %r736, %r959, 16; shr.u32 %r737, %r736, 16; xor.b32 %r738, %r737, %r736; mul.lo.s32 %r739, %r738, 569420461; shr.u32 %r740, %r739, 15; xor.b32 %r741, %r740, %r739; mul.lo.s32 %r742, %r741, -749064809; shr.u32 %r743, %r742, 15; xor.b32 %r745, %r334, %r21; xor.b32 %r746, %r745, %r742; xor.b32 %r43, %r746, %r743; xor.b32 %r747, %r43, -122820198; add.s32 %r748, %r747, %r8; shr.u32 %r749, %r747, 16; or.b32 %r750, %r749, 1; mul.lo.s32 %r751, %r750, %r748; mul.lo.s32 %r752, %r751, 89287766; xor.b32 %r753, %r752, %r751; mul.lo.s32 %r754, %r753, 1403136100; xor.b32 %r44, %r754, %r753; setp.eq.s32 %p66, %r44, 0; mov.u32 %r963, 0; @%p66 bra $L__BB207_68; mov.u32 %r960, 0; mov.u32 %r961, %r44; mov.u32 %r963, %r960; $L__BB207_67: clz.b32 %r757, %r961; add.s32 %r758, %r757, %r960; mul.wide.u32 %rd194, %r758, 4; mov.u64 %rd195, sobol_burley_table; add.s64 %rd196, %rd195, %rd194; ld.const.u32 %r759, [%rd196+128]; xor.b32 %r963, %r759, %r963; add.s32 %r960, %r758, 1; shl.b32 %r760, %r961, %r757; shl.b32 %r961, %r760, 1; setp.ne.s32 %p67, %r961, 0; @%p67 bra $L__BB207_67; $L__BB207_68: xor.b32 %r761, %r43, -525684874; brev.b32 %r762, %r44; shr.u32 %r763, %r761, 16; or.b32 %r764, %r763, 1; mul.lo.s32 %r765, %r762, 1025551850; xor.b32 %r766, %r765, %r762; add.s32 %r767, %r766, %r761; mul.lo.s32 %r768, %r767, %r764; mul.lo.s32 %r769, %r768, 89287766; xor.b32 %r770, %r769, %r768; xor.b32 %r771, %r43, -1802089138; mul.lo.s32 %r772, %r963, 1025551850; xor.b32 %r773, %r772, %r963; add.s32 %r774, %r773, %r771; shr.u32 %r775, %r771, 16; or.b32 %r776, %r775, 1; mul.lo.s32 %r777, %r774, %r776; mul.lo.s32 %r778, %r777, 89287766; xor.b32 %r779, %r778, %r777; mul.lo.s32 %r780, %r779, 1403136100; xor.b32 %r781, %r780, %r779; brev.b32 %r782, %r781; cvt.rn.f32.u32 %f1007, %r782; mul.lo.s32 %r783, %r770, 1403136100; xor.b32 %r784, %r783, %r770; brev.b32 %r785, %r784; cvt.rn.f32.u32 %f1008, %r785; mul.ftz.f32 %f2417, %f1007, 0f2F7FFFFE; mul.ftz.f32 %f161, %f1008, 0f2F7FFFFE; $L__BB207_69: selp.f32 %f1009, %f95, 0f3F400000, %p48; setp.lt.ftz.f32 %p69, %f2411, %f1009; @%p69 bra $L__BB207_76; bra.uni $L__BB207_70; $L__BB207_76: add.ftz.f32 %f2360, %f93, 0f3F800000; sub.ftz.f32 %f2359, %f20, %f18; mul.ftz.f32 %f1071, %f94, %f161; mul.ftz.f32 %f1072, %f1071, 0fBFB8AA3B; ex2.approx.ftz.f32 %f1073, %f1072; mov.f32 %f1075, 0f3F800000; mul.ftz.f32 %f1076, %f2360, %f1073; sub.ftz.f32 %f1077, %f93, %f1076; neg.ftz.f32 %f1078, %f1077; setp.eq.s16 %p75, %rs46, 0; selp.f32 %f2426, %f1077, %f1078, %p75; mul.ftz.f32 %f1079, %f2426, %f2426; sub.ftz.f32 %f1080, %f1075, %f1079; mov.f32 %f2421, 0f00000000; max.ftz.f32 %f1081, %f1080, %f2421; sqrt.approx.ftz.f32 %f1082, %f1081; mul.ftz.f32 %f1083, %f2417, 0f40C90FDB; cos.approx.ftz.f32 %f1084, %f1083; mul.ftz.f32 %f1085, %f1082, %f1084; sin.approx.ftz.f32 %f1086, %f1083; mul.ftz.f32 %f1087, %f1082, %f1086; div.approx.ftz.f32 %f1090, %f2359, %f1089; div.approx.ftz.f32 %f1094, %f922, %f1089; div.approx.ftz.f32 %f1098, %f924, %f1089; mul.ftz.f32 %f1099, %f18, %f1098; mul.ftz.f32 %f1100, %f20, %f1094; sub.ftz.f32 %f1101, %f1099, %f1100; mul.ftz.f32 %f1102, %f20, %f1090; mul.ftz.f32 %f1103, %f17, %f1098; sub.ftz.f32 %f1104, %f1102, %f1103; mul.ftz.f32 %f1105, %f17, %f1094; mul.ftz.f32 %f1106, %f18, %f1090; sub.ftz.f32 %f1107, %f1105, %f1106; mul.ftz.f32 %f1108, %f1087, %f1101; mul.ftz.f32 %f1109, %f1087, %f1104; mul.ftz.f32 %f1110, %f1087, %f1107; fma.rn.ftz.f32 %f1111, %f1085, %f1090, %f1108; fma.rn.ftz.f32 %f1112, %f1085, %f1094, %f1109; fma.rn.ftz.f32 %f1113, %f1085, %f1098, %f1110; fma.rn.ftz.f32 %f2422, %f17, %f2426, %f1111; fma.rn.ftz.f32 %f2423, %f18, %f2426, %f1112; fma.rn.ftz.f32 %f2424, %f20, %f2426, %f1113; mul.ftz.f32 %f1114, %f2428, %f2423; fma.rn.ftz.f32 %f1115, %f2429, %f2422, %f1114; fma.rn.ftz.f32 %f1116, %f2427, %f2424, %f1115; mul.ftz.f32 %f182, %f120, %f120; add.ftz.f32 %f1117, %f182, 0f3F800000; mul.ftz.f32 %f1118, %f120, 0fC0000000; fma.rn.ftz.f32 %f183, %f1118, %f1116, %f1117; setp.le.ftz.f32 %p79, %f183, 0f00000000; @%p79 bra $L__BB207_78; lg2.approx.ftz.f32 %f1119, %f183; mul.ftz.f32 %f1120, %f1119, 0f3FC00000; ex2.approx.ftz.f32 %f2421, %f1120; $L__BB207_78: sub.ftz.f32 %f1122, %f1075, %f182; div.approx.ftz.f32 %f1123, %f1122, %f2421; mul.ftz.f32 %f2425, %f1123, 0f3DA2F983; bra.uni $L__BB207_79; $L__BB207_70: abs.ftz.f32 %f1010, %f120; setp.lt.ftz.f32 %p70, %f1010, 0f3A83126F; @%p70 bra $L__BB207_74; bra.uni $L__BB207_71; $L__BB207_74: add.ftz.f32 %f1026, %f161, %f161; mov.f32 %f1027, 0f3F800000; sub.ftz.f32 %f2420, %f1027, %f1026; mov.f32 %f2425, 0f3DA2F983; bra.uni $L__BB207_75; $L__BB207_71: mul.ftz.f32 %f1012, %f120, %f120; mov.f32 %f1013, 0f3F800000; sub.ftz.f32 %f165, %f1013, %f1012; add.ftz.f32 %f1014, %f120, %f120; sub.ftz.f32 %f1015, %f1013, %f120; fma.rn.ftz.f32 %f1016, %f1014, %f161, %f1015; div.approx.ftz.f32 %f1017, %f165, %f1016; add.ftz.f32 %f1018, %f1012, 0f3F800000; mul.ftz.f32 %f1019, %f1017, %f1017; sub.ftz.f32 %f1020, %f1018, %f1019; div.approx.ftz.f32 %f2420, %f1020, %f1014; mul.ftz.f32 %f1021, %f1014, %f2420; sub.ftz.f32 %f167, %f1018, %f1021; setp.le.ftz.f32 %p71, %f167, 0f00000000; mov.f32 %f2418, 0f00000000; @%p71 bra $L__BB207_73; lg2.approx.ftz.f32 %f1022, %f167; mul.ftz.f32 %f1023, %f1022, 0f3FC00000; ex2.approx.ftz.f32 %f2418, %f1023; $L__BB207_73: div.approx.ftz.f32 %f1024, %f165, %f2418; mul.ftz.f32 %f2425, %f1024, 0f3DA2F983; $L__BB207_75: mul.ftz.f32 %f1028, %f2420, %f2420; mov.f32 %f1029, 0f3F800000; sub.ftz.f32 %f1030, %f1029, %f1028; mov.f32 %f1031, 0f00000000; max.ftz.f32 %f1032, %f1030, %f1031; sqrt.approx.ftz.f32 %f1033, %f1032; mul.ftz.f32 %f1034, %f2417, 0f40C90FDB; cos.approx.ftz.f32 %f1035, %f1034; mul.ftz.f32 %f1036, %f1033, %f1035; sin.approx.ftz.f32 %f1037, %f1034; mul.ftz.f32 %f1038, %f1033, %f1037; setp.neu.ftz.f32 %p72, %f2429, %f2427; setp.neu.ftz.f32 %p73, %f2429, %f2428; or.pred %p74, %p73, %p72; neg.ftz.f32 %f1039, %f2427; selp.f32 %f1040, %f1039, %f2427, %p74; add.ftz.f32 %f1041, %f2429, %f1040; neg.ftz.f32 %f1042, %f2428; selp.f32 %f1043, %f2428, %f1042, %p74; sub.ftz.f32 %f1044, %f1043, %f2429; sub.ftz.f32 %f1045, %f2427, %f2428; mul.ftz.f32 %f1046, %f1041, %f1041; fma.rn.ftz.f32 %f1047, %f1045, %f1045, %f1046; fma.rn.ftz.f32 %f1048, %f1044, %f1044, %f1047; sqrt.approx.ftz.f32 %f1049, %f1048; div.approx.ftz.f32 %f1050, %f1045, %f1049; div.approx.ftz.f32 %f1051, %f1041, %f1049; div.approx.ftz.f32 %f1052, %f1044, %f1049; mul.ftz.f32 %f1053, %f2428, %f1052; mul.ftz.f32 %f1054, %f2427, %f1051; sub.ftz.f32 %f1055, %f1053, %f1054; mul.ftz.f32 %f1056, %f2427, %f1050; mul.ftz.f32 %f1057, %f2429, %f1052; sub.ftz.f32 %f1058, %f1056, %f1057; mul.ftz.f32 %f1059, %f2429, %f1051; mul.ftz.f32 %f1060, %f2428, %f1050; sub.ftz.f32 %f1061, %f1059, %f1060; mul.ftz.f32 %f1062, %f1038, %f1055; mul.ftz.f32 %f1063, %f1038, %f1058; mul.ftz.f32 %f1064, %f1038, %f1061; fma.rn.ftz.f32 %f1065, %f1036, %f1050, %f1062; fma.rn.ftz.f32 %f1066, %f1036, %f1051, %f1063; fma.rn.ftz.f32 %f1067, %f1036, %f1052, %f1064; fma.rn.ftz.f32 %f2422, %f2429, %f2420, %f1065; fma.rn.ftz.f32 %f2423, %f2428, %f2420, %f1066; fma.rn.ftz.f32 %f2424, %f2427, %f2420, %f1067; mul.ftz.f32 %f1068, %f18, %f2423; fma.rn.ftz.f32 %f1069, %f17, %f2422, %f1068; fma.rn.ftz.f32 %f2426, %f20, %f2424, %f1069; $L__BB207_79: sub.ftz.f32 %f1125, %f93, %f2426; mul.ftz.f32 %f1126, %f94, %f1125; rcp.approx.ftz.f32 %f1127, %f1126; mov.f32 %f1128, 0f3F800000; mul.ftz.f32 %f1129, %f1127, 0f3E22F983; div.approx.ftz.f32 %f2432, %f1129, %f2425; add.ftz.f32 %f1130, %f93, %f2426; mul.ftz.f32 %f1131, %f94, %f1130; rcp.approx.ftz.f32 %f1132, %f1131; mul.ftz.f32 %f1133, %f1132, 0f3E22F983; div.approx.ftz.f32 %f2434, %f1133, %f2425; div.approx.ftz.f32 %f1134, %f2426, %f93; sub.ftz.f32 %f2433, %f1128, %f1134; add.ftz.f32 %f2435, %f1134, 0f3F800000; setp.eq.s16 %p82, %rs46, 0; selp.f32 %f1135, %f2433, %f2435, %p82; mul.ftz.f32 %f1136, %f2430, %f1135; selp.f32 %f2430, %f1136, %f2430, %p69; mov.f32 %f2427, %f2424; mov.f32 %f2428, %f2423; mov.f32 %f2429, %f2422; $L__BB207_80: mov.u32 %r964, -1; sub.ftz.f32 %f1138, %f623, %f2410; lg2.approx.ftz.f32 %f1139, %f1138; mul.ftz.f32 %f1140, %f1139, 0fBF317218; div.approx.ftz.f32 %f206, %f1140, %f2430; setp.ne.s32 %p83, %r958, 0; mov.f32 %f2436, %f206; mov.u32 %r965, %r964; @%p83 bra $L__BB207_82; min.ftz.f32 %f1147, %f945, %f948; min.ftz.f32 %f1151, %f1147, %f951; mov.f32 %f1152, 0f41200000; div.approx.ftz.f32 %f1153, %f1152, %f1151; max.ftz.f32 %f2436, %f206, %f1153; mov.u32 %r964, %r956; mov.u32 %r965, %r957; $L__BB207_82: mov.b32 %r54, %f2392; setp.nan.ftz.f32 %p85, %f2392, %f2392; mov.u16 %rs47, %rs18; @%p85 bra $L__BB207_85; or.b32 %r788, %r54, -2147483648; setp.ne.s32 %p86, %r788, -2147483648; add.ftz.f32 %f1154, %f2392, %f2392; setp.eq.ftz.f32 %p87, %f1154, %f2392; and.pred %p88, %p87, %p86; mov.u16 %rs47, %rs18; @%p88 bra $L__BB207_85; shl.b32 %r789, %r54, 1; setp.lt.u32 %p89, %r789, -16777215; selp.u16 %rs47, 1, 0, %p89; $L__BB207_85: setp.eq.s16 %p90, %rs47, 0; mov.u16 %rs49, %rs18; @%p90 bra $L__BB207_91; mov.b32 %r55, %f2429; setp.nan.ftz.f32 %p91, %f2429, %f2429; mov.u16 %rs49, 0; mov.u16 %rs48, %rs49; @%p91 bra $L__BB207_89; or.b32 %r790, %r55, -2147483648; setp.ne.s32 %p92, %r790, -2147483648; add.ftz.f32 %f1155, %f2429, %f2429; setp.eq.ftz.f32 %p93, %f1155, %f2429; and.pred %p94, %p93, %p92; mov.u16 %rs48, %rs49; @%p94 bra $L__BB207_89; shl.b32 %r791, %r55, 1; setp.lt.u32 %p95, %r791, -16777215; selp.u16 %rs48, 1, 0, %p95; $L__BB207_89: setp.eq.s16 %p96, %rs48, 0; @%p96 bra $L__BB207_91; mul.ftz.f32 %f1156, %f2428, %f2428; fma.rn.ftz.f32 %f1157, %f2429, %f2429, %f1156; fma.rn.ftz.f32 %f1158, %f2427, %f2427, %f1157; setp.neu.ftz.f32 %p97, %f1158, 0f00000000; selp.u16 %rs49, 1, 0, %p97; $L__BB207_91: setp.eq.s16 %p98, %rs49, 0; @%p98 bra $L__BB207_191; ld.const.u32 %r949, [kernel_params+1108]; setp.eq.s32 %p99, %r949, 0; abs.ftz.f32 %f1159, %f2429; setp.gt.ftz.f32 %p100, %f1159, 0f17800000; mov.b32 %r792, %f2429; and.b32 %r793, %r792, -2147483648; or.b32 %r794, %r793, 394264576; mov.b32 %f1160, %r794; selp.f32 %f2460, %f2429, %f1160, %p100; abs.ftz.f32 %f1161, %f2428; setp.gt.ftz.f32 %p101, %f1161, 0f17800000; mov.b32 %r795, %f2428; and.b32 %r796, %r795, -2147483648; or.b32 %r797, %r796, 394264576; mov.b32 %f1162, %r797; selp.f32 %f2461, %f2428, %f1162, %p101; abs.ftz.f32 %f1163, %f2427; setp.gt.ftz.f32 %p102, %f1163, 0f17800000; mov.b32 %r798, %f2427; and.b32 %r799, %r798, -2147483648; or.b32 %r800, %r799, 394264576; mov.b32 %f1164, %r800; selp.f32 %f2462, %f2427, %f1164, %p102; @%p99 bra $L__BB207_160; mov.u32 %r802, 1985229328; st.local.u32 [%rd183], %r802; ld.global.u32 %r975, [%rd6]; add.u64 %rd199, %SPL, 828; mov.u32 %r987, 0; st.local.u32 [%rd199], %r987; cvta.to.global.u64 %rd200, %rd164; mul.wide.s32 %rd201, %r19, 4; add.s64 %rd202, %rd200, %rd201; ld.global.u32 %r57, [%rd202]; and.b32 %r804, %r57, 4; setp.ne.s32 %p103, %r804, 0; mov.u32 %r966, -1; mov.f32 %f2463, %f2392; mov.f32 %f2464, %f2391; mov.f32 %f2465, %f2390; @%p103 bra $L__BB207_103; and.b32 %r805, %r57, 2; setp.eq.s32 %p104, %r805, 0; @%p104 bra $L__BB207_101; bra.uni $L__BB207_95; $L__BB207_101: ld.global.v4.f32 {%f2448, %f2449, %f2450, %f2451}, [%rd9]; ld.global.v4.f32 {%f2452, %f2453, %f2454, %f2455}, [%rd9+16]; ld.global.v4.f32 {%f2456, %f2457, %f2458, %f2459}, [%rd9+32]; bra.uni $L__BB207_102; $L__BB207_191: add.u64 %rd296, %SPL, 828; mov.u32 %r987, 0; st.local.u32 [%rd296], %r987; bra.uni $L__BB207_192; $L__BB207_160: mov.u32 %r875, 1985229328; st.local.u32 [%rd183], %r875; ld.global.u32 %r1005, [%rd6]; add.u64 %rd263, %SPL, 828; mov.u32 %r987, 0; st.local.u32 [%rd263], %r987; cvta.to.global.u64 %rd264, %rd164; mul.wide.s32 %rd265, %r19, 4; add.s64 %rd266, %rd264, %rd265; ld.global.u8 %rs29, [%rd266]; and.b16 %rs30, %rs29, 4; setp.ne.s16 %p188, %rs30, 0; mov.u32 %r997, -1; mov.f32 %f2521, %f2392; mov.f32 %f2522, %f2391; mov.f32 %f2523, %f2390; @%p188 bra $L__BB207_162; ld.global.v4.f32 {%f1870, %f1871, %f1872, %f1873}, [%rd9]; ld.global.v4.f32 {%f1878, %f1879, %f1880, %f1881}, [%rd9+16]; ld.global.v4.f32 {%f1886, %f1887, %f1888, %f1889}, [%rd9+32]; mul.ftz.f32 %f1894, %f1871, %f2391; fma.rn.ftz.f32 %f1895, %f2392, %f1870, %f1894; fma.rn.ftz.f32 %f1896, %f1872, %f2390, %f1895; add.ftz.f32 %f2521, %f1873, %f1896; mul.ftz.f32 %f1897, %f1879, %f2391; fma.rn.ftz.f32 %f1898, %f2392, %f1878, %f1897; fma.rn.ftz.f32 %f1899, %f1880, %f2390, %f1898; add.ftz.f32 %f2522, %f1881, %f1899; mul.ftz.f32 %f1900, %f1887, %f2391; fma.rn.ftz.f32 %f1901, %f2392, %f1886, %f1900; fma.rn.ftz.f32 %f1902, %f1888, %f2390, %f1901; add.ftz.f32 %f2523, %f1889, %f1902; mul.ftz.f32 %f1903, %f1871, %f2428; fma.rn.ftz.f32 %f1904, %f1870, %f2429, %f1903; fma.rn.ftz.f32 %f1905, %f1872, %f2427, %f1904; mul.ftz.f32 %f1906, %f1879, %f2428; fma.rn.ftz.f32 %f1907, %f1878, %f2429, %f1906; fma.rn.ftz.f32 %f1908, %f1880, %f2427, %f1907; mul.ftz.f32 %f1909, %f1887, %f2428; fma.rn.ftz.f32 %f1910, %f1886, %f2429, %f1909; fma.rn.ftz.f32 %f1911, %f1888, %f2427, %f1910; abs.ftz.f32 %f1912, %f1905; setp.gt.ftz.f32 %p189, %f1912, 0f17800000; mov.b32 %r877, %f1905; and.b32 %r878, %r877, -2147483648; or.b32 %r879, %r878, 394264576; mov.b32 %f1913, %r879; selp.f32 %f2460, %f1905, %f1913, %p189; abs.ftz.f32 %f1914, %f1908; setp.gt.ftz.f32 %p190, %f1914, 0f17800000; mov.b32 %r880, %f1908; and.b32 %r881, %r880, -2147483648; or.b32 %r882, %r881, 394264576; mov.b32 %f1915, %r882; selp.f32 %f2461, %f1908, %f1915, %p190; abs.ftz.f32 %f1916, %f1911; setp.gt.ftz.f32 %p191, %f1916, 0f17800000; mov.b32 %r883, %f1911; and.b32 %r884, %r883, -2147483648; or.b32 %r885, %r884, 394264576; mov.b32 %f1917, %r885; selp.f32 %f2462, %f1911, %f1917, %p191; mov.u32 %r997, %r19; $L__BB207_162: rcp.approx.ftz.f32 %f458, %f2461; rcp.approx.ftz.f32 %f459, %f2460; rcp.approx.ftz.f32 %f460, %f2462; mov.u32 %r1004, %r987; $L__BB207_163: setp.eq.s32 %p192, %r1005, 1985229328; setp.lt.s32 %p193, %r1005, 0; or.pred %p194, %p193, %p192; @%p194 bra $L__BB207_175; cvta.to.global.u64 %rd86, %rd169; $L__BB207_165: mul.wide.s32 %rd268, %r1005, 16; add.s64 %rd87, %rd86, %rd268; ld.global.v4.f32 {%f1918, %f1919, %f1920, %f1921}, [%rd87]; mov.b32 %r1005, %f1921; mov.b32 %r137, %f1920; mov.b32 %r136, %f1919; mov.b32 %r135, %f1918; mov.b64 %rd269, {%r135, %r136}; cvt.u32.u64 %r888, %rd269; and.b32 %r889, %r888, 1024; setp.eq.s32 %p195, %r889, 0; and.b32 %r139, %r135, 2047; @%p195 bra $L__BB207_168; bra.uni $L__BB207_166; $L__BB207_168: setp.ne.s32 %p201, %r139, 0; ld.global.v4.f32 {%f2062, %f2063, %f2064, %f2065}, [%rd87+16]; ld.global.v4.f32 {%f2070, %f2071, %f2072, %f2073}, [%rd87+32]; ld.global.v4.f32 {%f2078, %f2079, %f2080, %f2081}, [%rd87+48]; sub.ftz.f32 %f2086, %f2062, %f2521; mul.ftz.f32 %f2087, %f459, %f2086; sub.ftz.f32 %f2088, %f2064, %f2521; mul.ftz.f32 %f2089, %f459, %f2088; sub.ftz.f32 %f2090, %f2070, %f2522; mul.ftz.f32 %f2091, %f458, %f2090; sub.ftz.f32 %f2092, %f2072, %f2522; mul.ftz.f32 %f2093, %f458, %f2092; sub.ftz.f32 %f2094, %f2078, %f2523; mul.ftz.f32 %f2095, %f460, %f2094; sub.ftz.f32 %f2096, %f2080, %f2523; mul.ftz.f32 %f2097, %f460, %f2096; min.ftz.f32 %f2098, %f2087, %f2089; min.ftz.f32 %f2099, %f2091, %f2093; min.ftz.f32 %f2100, %f2095, %f2097; mov.f32 %f2101, 0f00000000; max.ftz.f32 %f2102, %f2101, %f2098; max.ftz.f32 %f2103, %f2099, %f2100; max.ftz.f32 %f2529, %f2102, %f2103; max.ftz.f32 %f2104, %f2087, %f2089; max.ftz.f32 %f2105, %f2091, %f2093; max.ftz.f32 %f2106, %f2095, %f2097; min.ftz.f32 %f2107, %f2436, %f2104; min.ftz.f32 %f2108, %f2105, %f2106; min.ftz.f32 %f2109, %f2107, %f2108; sub.ftz.f32 %f2110, %f2063, %f2521; mul.ftz.f32 %f2111, %f459, %f2110; sub.ftz.f32 %f2112, %f2065, %f2521; mul.ftz.f32 %f2113, %f459, %f2112; sub.ftz.f32 %f2114, %f2071, %f2522; mul.ftz.f32 %f2115, %f458, %f2114; sub.ftz.f32 %f2116, %f2073, %f2522; mul.ftz.f32 %f2117, %f458, %f2116; sub.ftz.f32 %f2118, %f2079, %f2523; mul.ftz.f32 %f2119, %f460, %f2118; sub.ftz.f32 %f2120, %f2081, %f2523; mul.ftz.f32 %f2121, %f460, %f2120; min.ftz.f32 %f2122, %f2111, %f2113; min.ftz.f32 %f2123, %f2115, %f2117; min.ftz.f32 %f2124, %f2119, %f2121; max.ftz.f32 %f2125, %f2101, %f2122; max.ftz.f32 %f2126, %f2123, %f2124; max.ftz.f32 %f2530, %f2125, %f2126; max.ftz.f32 %f2127, %f2111, %f2113; max.ftz.f32 %f2128, %f2115, %f2117; max.ftz.f32 %f2129, %f2119, %f2121; min.ftz.f32 %f2130, %f2436, %f2127; min.ftz.f32 %f2131, %f2128, %f2129; min.ftz.f32 %f2132, %f2130, %f2131; setp.ge.ftz.f32 %p202, %f2109, %f2529; and.pred %p203, %p201, %p202; selp.u32 %r893, 1, 0, %p203; setp.ltu.ftz.f32 %p204, %f2132, %f2530; and.b32 %r895, %r136, 2047; setp.eq.s32 %p205, %r895, 0; or.pred %p206, %p205, %p204; selp.b32 %r896, 0, 2, %p206; or.b32 %r1003, %r896, %r893; bra.uni $L__BB207_169; $L__BB207_166: setp.ne.s32 %p196, %r139, 0; ld.global.v4.f32 {%f1925, %f1926, %f1927, %f1928}, [%rd87+16]; ld.global.v4.f32 {%f1933, %f1934, %f1935, %f1936}, [%rd87+32]; ld.global.v4.f32 {%f1941, %f1942, %f1943, %f1944}, [%rd87+48]; mul.ftz.f32 %f1949, %f2461, %f1926; fma.rn.ftz.f32 %f1950, %f2460, %f1925, %f1949; fma.rn.ftz.f32 %f1951, %f2462, %f1927, %f1950; mul.ftz.f32 %f1952, %f2461, %f1934; fma.rn.ftz.f32 %f1953, %f2460, %f1933, %f1952; fma.rn.ftz.f32 %f1954, %f2462, %f1935, %f1953; mul.ftz.f32 %f1955, %f2461, %f1942; fma.rn.ftz.f32 %f1956, %f2460, %f1941, %f1955; fma.rn.ftz.f32 %f1957, %f2462, %f1943, %f1956; mul.ftz.f32 %f1958, %f2522, %f1926; fma.rn.ftz.f32 %f1959, %f2521, %f1925, %f1958; fma.rn.ftz.f32 %f1960, %f2523, %f1927, %f1959; add.ftz.f32 %f1961, %f1928, %f1960; mul.ftz.f32 %f1962, %f2522, %f1934; fma.rn.ftz.f32 %f1963, %f2521, %f1933, %f1962; fma.rn.ftz.f32 %f1964, %f2523, %f1935, %f1963; add.ftz.f32 %f1965, %f1936, %f1964; mul.ftz.f32 %f1966, %f2522, %f1942; fma.rn.ftz.f32 %f1967, %f2521, %f1941, %f1966; fma.rn.ftz.f32 %f1968, %f2523, %f1943, %f1967; add.ftz.f32 %f1969, %f1944, %f1968; rcp.approx.ftz.f32 %f1970, %f1951; rcp.approx.ftz.f32 %f1971, %f1954; rcp.approx.ftz.f32 %f1972, %f1957; mul.ftz.f32 %f1973, %f1970, %f1961; neg.ftz.f32 %f1974, %f1973; mul.ftz.f32 %f1975, %f1971, %f1965; neg.ftz.f32 %f1976, %f1975; mul.ftz.f32 %f1977, %f1972, %f1969; neg.ftz.f32 %f1978, %f1977; sub.ftz.f32 %f1979, %f1970, %f1973; sub.ftz.f32 %f1980, %f1971, %f1975; sub.ftz.f32 %f1981, %f1972, %f1977; min.ftz.f32 %f1982, %f1974, %f1979; min.ftz.f32 %f1983, %f1976, %f1980; min.ftz.f32 %f1984, %f1978, %f1981; max.ftz.f32 %f1985, %f1974, %f1979; max.ftz.f32 %f1986, %f1976, %f1980; max.ftz.f32 %f1987, %f1978, %f1981; mov.f32 %f1988, 0f00000000; max.ftz.f32 %f1989, %f1988, %f1982; max.ftz.f32 %f1990, %f1983, %f1984; max.ftz.f32 %f2529, %f1989, %f1990; min.ftz.f32 %f1991, %f2436, %f1985; min.ftz.f32 %f1992, %f1986, %f1987; min.ftz.f32 %f1993, %f1991, %f1992; setp.le.ftz.f32 %p197, %f2529, %f1993; and.pred %p198, %p196, %p197; selp.u32 %r1003, 1, 0, %p198; ld.global.v4.f32 {%f1994, %f1995, %f1996, %f1997}, [%rd87+64]; ld.global.v4.f32 {%f2002, %f2003, %f2004, %f2005}, [%rd87+80]; ld.global.v4.f32 {%f2010, %f2011, %f2012, %f2013}, [%rd87+96]; mul.ftz.f32 %f2018, %f2461, %f1995; fma.rn.ftz.f32 %f2019, %f2460, %f1994, %f2018; fma.rn.ftz.f32 %f2020, %f2462, %f1996, %f2019; mul.ftz.f32 %f2021, %f2461, %f2003; fma.rn.ftz.f32 %f2022, %f2460, %f2002, %f2021; fma.rn.ftz.f32 %f2023, %f2462, %f2004, %f2022; mul.ftz.f32 %f2024, %f2461, %f2011; fma.rn.ftz.f32 %f2025, %f2460, %f2010, %f2024; fma.rn.ftz.f32 %f2026, %f2462, %f2012, %f2025; mul.ftz.f32 %f2027, %f2522, %f1995; fma.rn.ftz.f32 %f2028, %f2521, %f1994, %f2027; fma.rn.ftz.f32 %f2029, %f2523, %f1996, %f2028; add.ftz.f32 %f2030, %f1997, %f2029; mul.ftz.f32 %f2031, %f2522, %f2003; fma.rn.ftz.f32 %f2032, %f2521, %f2002, %f2031; fma.rn.ftz.f32 %f2033, %f2523, %f2004, %f2032; add.ftz.f32 %f2034, %f2005, %f2033; mul.ftz.f32 %f2035, %f2522, %f2011; fma.rn.ftz.f32 %f2036, %f2521, %f2010, %f2035; fma.rn.ftz.f32 %f2037, %f2523, %f2012, %f2036; add.ftz.f32 %f2038, %f2013, %f2037; rcp.approx.ftz.f32 %f2039, %f2020; rcp.approx.ftz.f32 %f2040, %f2023; rcp.approx.ftz.f32 %f2041, %f2026; mul.ftz.f32 %f2042, %f2039, %f2030; neg.ftz.f32 %f2043, %f2042; mul.ftz.f32 %f2044, %f2040, %f2034; neg.ftz.f32 %f2045, %f2044; mul.ftz.f32 %f2046, %f2041, %f2038; neg.ftz.f32 %f2047, %f2046; sub.ftz.f32 %f2048, %f2039, %f2042; sub.ftz.f32 %f2049, %f2040, %f2044; sub.ftz.f32 %f2050, %f2041, %f2046; min.ftz.f32 %f2051, %f2043, %f2048; min.ftz.f32 %f2052, %f2045, %f2049; min.ftz.f32 %f2053, %f2047, %f2050; max.ftz.f32 %f2054, %f2043, %f2048; max.ftz.f32 %f2055, %f2045, %f2049; max.ftz.f32 %f2056, %f2047, %f2050; max.ftz.f32 %f2057, %f1988, %f2051; max.ftz.f32 %f2058, %f2052, %f2053; max.ftz.f32 %f2530, %f2057, %f2058; min.ftz.f32 %f2059, %f2436, %f2054; min.ftz.f32 %f2060, %f2055, %f2056; min.ftz.f32 %f2061, %f2059, %f2060; setp.gtu.ftz.f32 %p199, %f2530, %f2061; @%p199 bra $L__BB207_169; and.b32 %r891, %r136, 2047; setp.eq.s32 %p200, %r891, 0; or.b32 %r892, %r1003, 2; selp.b32 %r1003, %r1003, %r892, %p200; $L__BB207_169: mul.wide.s32 %rd270, %r1004, 4; add.s64 %rd88, %rd183, %rd270; setp.eq.s32 %p207, %r1003, 3; @%p207 bra $L__BB207_173; bra.uni $L__BB207_170; $L__BB207_173: setp.lt.ftz.f32 %p210, %f2530, %f2529; selp.b32 %r148, %r1005, %r137, %p210; selp.b32 %r897, %r137, %r1005, %p210; add.s32 %r1004, %r1004, 1; st.local.u32 [%rd88+4], %r897; mov.u32 %r1005, %r148; bra.uni $L__BB207_174; $L__BB207_170: setp.eq.s32 %p208, %r1003, 2; @%p208 bra $L__BB207_174; setp.ne.s32 %p209, %r1003, 0; mov.u32 %r1005, %r137; @%p209 bra $L__BB207_174; ld.local.u32 %r1005, [%rd88]; add.s32 %r1004, %r1004, -1; $L__BB207_174: setp.ne.s32 %p211, %r1005, 1985229328; setp.gt.s32 %p212, %r1005, -1; and.pred %p213, %p212, %p211; @%p213 bra $L__BB207_165; $L__BB207_175: mov.u32 %r152, %r1004; setp.gt.s32 %p214, %r1005, -1; @%p214 bra $L__BB207_190; cvta.to.global.u64 %rd271, %rd170; mul.wide.s32 %rd272, %r1005, 16; xor.b64 %rd273, %rd272, -16; add.s64 %rd274, %rd271, %rd273; ld.global.v4.u32 {%r1009, %r899, %r900, %r901}, [%rd274]; mul.wide.s32 %rd277, %r152, 4; add.s64 %rd278, %rd183, %rd277; ld.local.u32 %r1005, [%rd278]; add.s32 %r1004, %r152, -1; and.b32 %r903, %r901, 127; setp.ne.s32 %p215, %r903, 1; setp.ge.s32 %p216, %r1009, %r899; or.pred %p217, %p215, %p216; @%p217 bra $L__BB207_190; $L__BB207_177: cvt.s64.s32 %rd89, %r1009; setp.ne.s32 %p218, %r997, -1; @%p218 bra $L__BB207_179; cvta.to.global.u64 %rd279, %rd172; shl.b64 %rd280, %rd89, 2; add.s64 %rd281, %rd279, %rd280; ld.global.u32 %r904, [%rd281]; setp.ne.s32 %p219, %r904, %r19; @%p219 bra $L__BB207_189; $L__BB207_179: cvta.to.global.u64 %rd282, %rd171; shl.b64 %rd283, %rd89, 2; add.s64 %rd284, %rd282, %rd283; ld.global.u32 %r160, [%rd284]; setp.eq.s32 %p220, %r965, %r160; @%p220 bra $L__BB207_189; mul.wide.s32 %rd285, %r160, 16; add.s64 %rd286, %rd14, %rd285; ld.global.u32 %r905, [%rd286+12]; mul.wide.u32 %rd287, %r905, 12; add.s64 %rd288, %rd15, %rd287; add.s32 %r906, %r905, 1; mul.wide.u32 %rd289, %r906, 12; add.s64 %rd290, %rd15, %rd289; add.s32 %r907, %r905, 2; mul.wide.u32 %rd291, %r907, 12; add.s64 %rd292, %rd15, %rd291; ld.global.f32 %f478, [%rd288]; sub.ftz.f32 %f479, %f478, %f2521; ld.global.f32 %f480, [%rd288+4]; sub.ftz.f32 %f481, %f480, %f2522; ld.global.f32 %f482, [%rd288+8]; sub.ftz.f32 %f483, %f482, %f2523; ld.global.f32 %f484, [%rd290]; sub.ftz.f32 %f2133, %f484, %f2521; ld.global.f32 %f485, [%rd290+4]; sub.ftz.f32 %f2134, %f485, %f2522; ld.global.f32 %f486, [%rd290+8]; sub.ftz.f32 %f2135, %f486, %f2523; ld.global.f32 %f487, [%rd292]; sub.ftz.f32 %f2136, %f487, %f2521; ld.global.f32 %f488, [%rd292+4]; sub.ftz.f32 %f2137, %f488, %f2522; ld.global.f32 %f489, [%rd292+8]; sub.ftz.f32 %f2138, %f489, %f2523; sub.ftz.f32 %f490, %f2136, %f479; sub.ftz.f32 %f491, %f2137, %f481; sub.ftz.f32 %f492, %f2138, %f483; sub.ftz.f32 %f493, %f479, %f2133; sub.ftz.f32 %f494, %f481, %f2134; sub.ftz.f32 %f495, %f483, %f2135; sub.ftz.f32 %f2139, %f2133, %f2136; sub.ftz.f32 %f2140, %f2134, %f2137; sub.ftz.f32 %f2141, %f2135, %f2138; add.ftz.f32 %f2142, %f479, %f2136; add.ftz.f32 %f2143, %f481, %f2137; add.ftz.f32 %f2144, %f483, %f2138; mul.ftz.f32 %f2145, %f491, %f2144; mul.ftz.f32 %f2146, %f2143, %f492; sub.ftz.f32 %f2147, %f2145, %f2146; mul.ftz.f32 %f2148, %f2142, %f492; mul.ftz.f32 %f2149, %f490, %f2144; sub.ftz.f32 %f2150, %f2148, %f2149; mul.ftz.f32 %f2151, %f490, %f2143; mul.ftz.f32 %f2152, %f2142, %f491; sub.ftz.f32 %f2153, %f2151, %f2152; mul.ftz.f32 %f2154, %f2461, %f2150; fma.rn.ftz.f32 %f2155, %f2460, %f2147, %f2154; fma.rn.ftz.f32 %f496, %f2462, %f2153, %f2155; add.ftz.f32 %f2156, %f479, %f2133; add.ftz.f32 %f2157, %f481, %f2134; add.ftz.f32 %f2158, %f483, %f2135; mul.ftz.f32 %f2159, %f494, %f2158; mul.ftz.f32 %f2160, %f2157, %f495; sub.ftz.f32 %f2161, %f2159, %f2160; mul.ftz.f32 %f2162, %f2156, %f495; mul.ftz.f32 %f2163, %f493, %f2158; sub.ftz.f32 %f2164, %f2162, %f2163; mul.ftz.f32 %f2165, %f493, %f2157; mul.ftz.f32 %f2166, %f2156, %f494; sub.ftz.f32 %f2167, %f2165, %f2166; mul.ftz.f32 %f2168, %f2461, %f2164; fma.rn.ftz.f32 %f2169, %f2460, %f2161, %f2168; fma.rn.ftz.f32 %f497, %f2462, %f2167, %f2169; add.ftz.f32 %f2170, %f2133, %f2136; add.ftz.f32 %f2171, %f2134, %f2137; add.ftz.f32 %f2172, %f2135, %f2138; mul.ftz.f32 %f2173, %f2140, %f2172; mul.ftz.f32 %f2174, %f2171, %f2141; sub.ftz.f32 %f2175, %f2173, %f2174; mul.ftz.f32 %f2176, %f2170, %f2141; mul.ftz.f32 %f2177, %f2139, %f2172; sub.ftz.f32 %f2178, %f2176, %f2177; mul.ftz.f32 %f2179, %f2139, %f2171; mul.ftz.f32 %f2180, %f2170, %f2140; sub.ftz.f32 %f2181, %f2179, %f2180; mul.ftz.f32 %f2182, %f2461, %f2178; fma.rn.ftz.f32 %f2183, %f2460, %f2175, %f2182; fma.rn.ftz.f32 %f2184, %f2462, %f2181, %f2183; add.ftz.f32 %f2185, %f497, %f496; add.ftz.f32 %f498, %f2184, %f2185; abs.ftz.f32 %f499, %f498; mul.ftz.f32 %f2186, %f499, 0f34000000; min.ftz.f32 %f2187, %f497, %f2184; min.ftz.f32 %f2188, %f496, %f2187; max.ftz.f32 %f2189, %f497, %f2184; max.ftz.f32 %f2190, %f496, %f2189; mul.ftz.f32 %f2191, %f499, 0fB4000000; setp.ltu.ftz.f32 %p222, %f2188, %f2191; setp.gtu.ftz.f32 %p223, %f2190, %f2186; and.pred %p224, %p222, %p223; mov.pred %p282, -1; @%p224 bra $L__BB207_186; mov.pred %p282, -1; mul.ftz.f32 %f2192, %f494, %f492; mul.ftz.f32 %f2193, %f495, %f491; sub.ftz.f32 %f2194, %f2192, %f2193; mul.ftz.f32 %f2195, %f493, %f492; mul.ftz.f32 %f2196, %f495, %f490; sub.ftz.f32 %f2197, %f2196, %f2195; mul.ftz.f32 %f2198, %f494, %f490; mul.ftz.f32 %f2199, %f493, %f491; sub.ftz.f32 %f2200, %f2199, %f2198; add.ftz.f32 %f500, %f2194, %f2194; add.ftz.f32 %f501, %f2197, %f2197; add.ftz.f32 %f502, %f2200, %f2200; mul.ftz.f32 %f2201, %f2461, %f501; fma.rn.ftz.f32 %f2202, %f2460, %f500, %f2201; fma.rn.ftz.f32 %f503, %f2462, %f502, %f2202; setp.eq.ftz.f32 %p226, %f503, 0f00000000; @%p226 bra $L__BB207_186; mov.pred %p282, -1; sub.ftz.f32 %f2366, %f482, %f2523; sub.ftz.f32 %f2365, %f478, %f2521; sub.ftz.f32 %f2364, %f480, %f2522; mul.ftz.f32 %f2203, %f2364, %f501; fma.rn.ftz.f32 %f2204, %f2365, %f500, %f2203; fma.rn.ftz.f32 %f2205, %f2366, %f502, %f2204; div.approx.ftz.f32 %f504, %f2205, %f503; setp.ltu.ftz.f32 %p228, %f504, 0f00000000; setp.gtu.ftz.f32 %p229, %f504, %f2436; or.pred %p230, %p229, %p228; @%p230 bra $L__BB207_186; setp.lt.ftz.f32 %p231, %f499, 0f219392EF; mov.f32 %f2536, 0f00000000; @%p231 bra $L__BB207_185; rcp.approx.ftz.f32 %f2536, %f498; $L__BB207_185: mul.ftz.f32 %f2207, %f496, %f2536; mov.f32 %f2208, 0f3F800000; min.ftz.f32 %f2543, %f2207, %f2208; mul.ftz.f32 %f2209, %f497, %f2536; min.ftz.f32 %f2542, %f2209, %f2208; mov.pred %p282, 0; mov.f32 %f2544, %f504; $L__BB207_186: @%p282 bra $L__BB207_189; setp.ne.s32 %p233, %r987, 0; setp.gt.ftz.f32 %p234, %f2544, %f2487; and.pred %p235, %p233, %p234; @%p235 bra $L__BB207_189; add.u64 %rd452, %SPL, 828; mov.u32 %r987, 1; st.local.u32 [%rd452], %r987; st.local.u32 [%rd452+16], %r160; st.local.u32 [%rd452+20], %r19; st.local.u32 [%rd452+24], %r987; st.local.f32 [%rd452+8], %f2543; st.local.f32 [%rd452+12], %f2542; st.local.f32 [%rd452+4], %f2544; sub.ftz.f32 %f2210, %f489, %f482; sub.ftz.f32 %f2211, %f485, %f480; mul.ftz.f32 %f2212, %f2211, %f2210; sub.ftz.f32 %f2213, %f488, %f480; sub.ftz.f32 %f2214, %f486, %f482; mul.ftz.f32 %f2215, %f2214, %f2213; sub.ftz.f32 %f2216, %f2212, %f2215; sub.ftz.f32 %f2217, %f487, %f478; mul.ftz.f32 %f2218, %f2214, %f2217; sub.ftz.f32 %f2219, %f484, %f478; mul.ftz.f32 %f2220, %f2219, %f2210; sub.ftz.f32 %f2221, %f2218, %f2220; mul.ftz.f32 %f2222, %f2219, %f2213; mul.ftz.f32 %f2223, %f2211, %f2217; sub.ftz.f32 %f2224, %f2222, %f2223; mul.ftz.f32 %f2225, %f2221, %f2221; fma.rn.ftz.f32 %f2226, %f2216, %f2216, %f2225; fma.rn.ftz.f32 %f2227, %f2224, %f2224, %f2226; sqrt.approx.ftz.f32 %f2228, %f2227; div.approx.ftz.f32 %f2229, %f2216, %f2228; div.approx.ftz.f32 %f2230, %f2221, %f2228; div.approx.ftz.f32 %f2231, %f2224, %f2228; st.local.f32 [%rd452+100], %f2229; st.local.f32 [%rd452+104], %f2230; st.local.f32 [%rd452+108], %f2231; mov.f32 %f2487, %f2544; $L__BB207_189: add.s32 %r1004, %r152, -1; cvt.u32.u64 %r909, %rd89; add.s32 %r1009, %r909, 1; setp.lt.s32 %p236, %r1009, %r899; @%p236 bra $L__BB207_177; $L__BB207_190: setp.ne.s32 %p237, %r1005, 1985229328; @%p237 bra $L__BB207_163; bra.uni $L__BB207_192; $L__BB207_95: ld.global.u32 %rd203, [%rd9+120]; ld.global.u32 %r806, [%rd9+104]; shl.b32 %r807, %r806, 1; cvt.rn.f32.s32 %f1165, %r807; mul.ftz.f32 %f1166, %f25, %f1165; cvt.rzi.ftz.s32.f32 %r808, %f1166; add.s32 %r809, %r807, -1; min.s32 %r810, %r808, %r809; cvt.rn.f32.s32 %f1167, %r810; sub.ftz.f32 %f212, %f1166, %f1167; cvt.s64.s32 %rd204, %r810; add.s64 %rd205, %rd204, %rd203; shl.b64 %rd207, %rd205, 6; add.s64 %rd21, %rd206, %rd207; ld.global.v4.f32 {%f1168, %f1169, %f1170, %f1171}, [%rd21]; ld.global.v4.f32 {%f1172, %f1173, %f1174, %f1175}, [%rd21+64]; mul.ftz.f32 %f1176, %f1169, %f1173; fma.rn.ftz.f32 %f1177, %f1168, %f1172, %f1176; mul.ftz.f32 %f1178, %f1171, %f1175; fma.rn.ftz.f32 %f1179, %f1170, %f1174, %f1178; add.ftz.f32 %f221, %f1177, %f1179; setp.gt.ftz.f32 %p105, %f221, 0f3F7FDF3B; mov.f32 %f1180, 0f3F800000; sub.ftz.f32 %f222, %f1180, %f212; @%p105 bra $L__BB207_97; bra.uni $L__BB207_96; $L__BB207_97: mul.ftz.f32 %f1234, %f1168, %f222; fma.rn.ftz.f32 %f1235, %f1172, %f212, %f1234; mul.ftz.f32 %f1236, %f1169, %f222; fma.rn.ftz.f32 %f1237, %f1173, %f212, %f1236; mul.ftz.f32 %f1238, %f1170, %f222; fma.rn.ftz.f32 %f1239, %f1174, %f212, %f1238; mul.ftz.f32 %f1240, %f1171, %f222; fma.rn.ftz.f32 %f1241, %f1175, %f212, %f1240; mul.ftz.f32 %f1242, %f1237, %f1237; fma.rn.ftz.f32 %f1243, %f1235, %f1235, %f1242; mul.ftz.f32 %f1244, %f1241, %f1241; fma.rn.ftz.f32 %f1245, %f1239, %f1239, %f1244; add.ftz.f32 %f1246, %f1243, %f1245; rsqrt.approx.ftz.f32 %f1247, %f1246; mul.ftz.f32 %f2440, %f1241, %f1247; mul.ftz.f32 %f2439, %f1239, %f1247; mul.ftz.f32 %f2438, %f1237, %f1247; mul.ftz.f32 %f2437, %f1235, %f1247; bra.uni $L__BB207_98; $L__BB207_96: mov.f32 %f2361, 0f3FC90FDB; mov.f32 %f1181, 0fBF800000; max.ftz.f32 %f1182, %f221, %f1181; min.ftz.f32 %f1184, %f1182, %f1180; abs.ftz.f32 %f1185, %f1184; sub.ftz.f32 %f1186, %f1180, %f1185; mul.ftz.f32 %f1187, %f1186, 0f3F000000; sqrt.approx.ftz.f32 %f1188, %f1187; setp.gt.ftz.f32 %p106, %f1185, 0f3F11EB85; selp.f32 %f1189, %f1188, %f1185, %p106; mul.ftz.f32 %f1190, %f1189, %f1189; mov.f32 %f1191, 0f3C94D2E9; mov.f32 %f1192, 0f3D53F941; fma.rn.ftz.f32 %f1193, %f1192, %f1190, %f1191; mov.f32 %f1194, 0f3D3F841F; fma.rn.ftz.f32 %f1195, %f1193, %f1190, %f1194; mov.f32 %f1196, 0f3D994929; fma.rn.ftz.f32 %f1197, %f1195, %f1190, %f1196; mov.f32 %f1198, 0f3E2AAB94; fma.rn.ftz.f32 %f1199, %f1197, %f1190, %f1198; mul.ftz.f32 %f1200, %f1190, %f1199; fma.rn.ftz.f32 %f1201, %f1200, %f1189, %f1189; add.ftz.f32 %f1202, %f1201, %f1201; sub.ftz.f32 %f1204, %f2361, %f1201; selp.f32 %f1205, %f1202, %f1204, %p106; setp.lt.ftz.f32 %p107, %f1184, 0f00000000; mov.f32 %f1206, 0f40490FDB; sub.ftz.f32 %f1207, %f1206, %f1205; selp.f32 %f1208, %f1207, %f1205, %p107; mul.ftz.f32 %f1209, %f1168, %f221; sub.ftz.f32 %f1210, %f1172, %f1209; mul.ftz.f32 %f1211, %f1169, %f221; sub.ftz.f32 %f1212, %f1173, %f1211; mul.ftz.f32 %f1213, %f1170, %f221; sub.ftz.f32 %f1214, %f1174, %f1213; mul.ftz.f32 %f1215, %f1171, %f221; sub.ftz.f32 %f1216, %f1175, %f1215; mul.ftz.f32 %f1217, %f1212, %f1212; fma.rn.ftz.f32 %f1218, %f1210, %f1210, %f1217; mul.ftz.f32 %f1219, %f1216, %f1216; fma.rn.ftz.f32 %f1220, %f1214, %f1214, %f1219; add.ftz.f32 %f1221, %f1218, %f1220; rsqrt.approx.ftz.f32 %f1222, %f1221; mul.ftz.f32 %f1223, %f1210, %f1222; mul.ftz.f32 %f1224, %f1212, %f1222; mul.ftz.f32 %f1225, %f1214, %f1222; mul.ftz.f32 %f1226, %f1216, %f1222; mul.ftz.f32 %f1227, %f212, %f1208; cos.approx.ftz.f32 %f1228, %f1227; sin.approx.ftz.f32 %f1229, %f1227; mul.ftz.f32 %f1230, %f1223, %f1229; mul.ftz.f32 %f1231, %f1224, %f1229; mul.ftz.f32 %f1232, %f1225, %f1229; mul.ftz.f32 %f1233, %f1226, %f1229; fma.rn.ftz.f32 %f2440, %f1171, %f1228, %f1233; fma.rn.ftz.f32 %f2439, %f1170, %f1228, %f1232; fma.rn.ftz.f32 %f2438, %f1169, %f1228, %f1231; fma.rn.ftz.f32 %f2437, %f1168, %f1228, %f1230; $L__BB207_98: ld.global.v4.f32 {%f1248, %f1249, %f1250, %f1251}, [%rd21+16]; ld.global.v4.f32 {%f1256, %f1257, %f1258, %f1259}, [%rd21+80]; mul.ftz.f32 %f1264, %f212, %f1256; mul.ftz.f32 %f1265, %f212, %f1257; mul.ftz.f32 %f1266, %f212, %f1258; mul.ftz.f32 %f1267, %f212, %f1259; fma.rn.ftz.f32 %f235, %f222, %f1248, %f1264; fma.rn.ftz.f32 %f236, %f222, %f1249, %f1265; fma.rn.ftz.f32 %f237, %f222, %f1250, %f1266; fma.rn.ftz.f32 %f1268, %f222, %f1251, %f1267; ld.global.v4.f32 {%f1269, %f1270, %f1271, %f1272}, [%rd21+32]; ld.global.v4.f32 {%f1277, %f1278, %f1279, %f1280}, [%rd21+96]; mul.ftz.f32 %f1285, %f212, %f1277; mul.ftz.f32 %f1286, %f212, %f1278; mul.ftz.f32 %f1287, %f212, %f1279; mul.ftz.f32 %f1288, %f212, %f1280; fma.rn.ftz.f32 %f1289, %f222, %f1269, %f1285; fma.rn.ftz.f32 %f1290, %f222, %f1270, %f1286; fma.rn.ftz.f32 %f1291, %f222, %f1271, %f1287; fma.rn.ftz.f32 %f1292, %f222, %f1272, %f1288; ld.global.v4.f32 {%f1293, %f1294, %f1295, %f1296}, [%rd21+48]; ld.global.v4.f32 {%f1301, %f1302, %f1303, %f1304}, [%rd21+112]; mul.ftz.f32 %f1309, %f212, %f1301; mul.ftz.f32 %f1310, %f212, %f1302; mul.ftz.f32 %f1311, %f212, %f1303; mul.ftz.f32 %f1312, %f212, %f1304; fma.rn.ftz.f32 %f1313, %f222, %f1293, %f1309; fma.rn.ftz.f32 %f1314, %f222, %f1294, %f1310; fma.rn.ftz.f32 %f1315, %f222, %f1295, %f1311; fma.rn.ftz.f32 %f1316, %f222, %f1296, %f1312; mul.ftz.f32 %f1317, %f2440, 0f3FB504F3; mul.ftz.f32 %f1318, %f2437, 0f3FB504F3; mul.ftz.f32 %f1319, %f2438, 0f3FB504F3; mul.ftz.f32 %f1320, %f2439, 0f3FB504F3; mul.ftz.f32 %f1321, %f1317, %f1318; mul.ftz.f32 %f1322, %f1317, %f1319; mul.ftz.f32 %f1323, %f1317, %f1320; mul.ftz.f32 %f1324, %f1318, %f1318; mul.ftz.f32 %f1325, %f1318, %f1319; mul.ftz.f32 %f1326, %f1318, %f1320; mul.ftz.f32 %f1327, %f1319, %f1319; mul.ftz.f32 %f1328, %f1319, %f1320; mul.ftz.f32 %f1329, %f1320, %f1320; sub.ftz.f32 %f1331, %f1180, %f1327; sub.ftz.f32 %f1332, %f1331, %f1329; sub.ftz.f32 %f1333, %f1325, %f1323; add.ftz.f32 %f1334, %f1322, %f1326; add.ftz.f32 %f1335, %f1323, %f1325; sub.ftz.f32 %f1336, %f1180, %f1324; sub.ftz.f32 %f1337, %f1336, %f1329; sub.ftz.f32 %f1338, %f1328, %f1321; sub.ftz.f32 %f1339, %f1326, %f1322; add.ftz.f32 %f1340, %f1321, %f1328; sub.ftz.f32 %f1341, %f1336, %f1327; mul.ftz.f32 %f1342, %f1333, %f1291; fma.rn.ftz.f32 %f1343, %f1332, %f1268, %f1342; fma.rn.ftz.f32 %f2446, %f1334, %f1314, %f1343; mul.ftz.f32 %f1344, %f1333, %f1292; fma.rn.ftz.f32 %f1345, %f1332, %f1289, %f1344; fma.rn.ftz.f32 %f239, %f1334, %f1315, %f1345; mul.ftz.f32 %f1346, %f1333, %f1313; fma.rn.ftz.f32 %f1347, %f1332, %f1290, %f1346; fma.rn.ftz.f32 %f240, %f1334, %f1316, %f1347; mul.ftz.f32 %f1348, %f1337, %f1291; fma.rn.ftz.f32 %f1349, %f1335, %f1268, %f1348; fma.rn.ftz.f32 %f241, %f1338, %f1314, %f1349; mul.ftz.f32 %f1350, %f1337, %f1292; fma.rn.ftz.f32 %f1351, %f1335, %f1289, %f1350; fma.rn.ftz.f32 %f2445, %f1338, %f1315, %f1351; mul.ftz.f32 %f1352, %f1337, %f1313; fma.rn.ftz.f32 %f1353, %f1335, %f1290, %f1352; fma.rn.ftz.f32 %f243, %f1338, %f1316, %f1353; mul.ftz.f32 %f1354, %f1340, %f1291; fma.rn.ftz.f32 %f1355, %f1339, %f1268, %f1354; fma.rn.ftz.f32 %f244, %f1341, %f1314, %f1355; mul.ftz.f32 %f1356, %f1340, %f1292; fma.rn.ftz.f32 %f1357, %f1339, %f1289, %f1356; fma.rn.ftz.f32 %f245, %f1341, %f1315, %f1357; mul.ftz.f32 %f1358, %f1340, %f1313; fma.rn.ftz.f32 %f1359, %f1339, %f1290, %f1358; fma.rn.ftz.f32 %f2444, %f1341, %f1316, %f1359; mul.ftz.f32 %f1360, %f2445, %f2444; mul.ftz.f32 %f247, %f245, %f243; sub.ftz.f32 %f2443, %f1360, %f247; mul.ftz.f32 %f249, %f245, %f240; mul.ftz.f32 %f1361, %f239, %f2444; sub.ftz.f32 %f2442, %f249, %f1361; mul.ftz.f32 %f251, %f239, %f243; mul.ftz.f32 %f1362, %f2445, %f240; sub.ftz.f32 %f2441, %f251, %f1362; mul.ftz.f32 %f1363, %f241, %f2442; fma.rn.ftz.f32 %f1364, %f2446, %f2443, %f1363; fma.rn.ftz.f32 %f2447, %f244, %f2441, %f1364; setp.neu.ftz.f32 %p108, %f2447, 0f00000000; @%p108 bra $L__BB207_100; add.ftz.f32 %f2446, %f2446, 0f322BCC77; add.ftz.f32 %f2444, %f2444, 0f322BCC77; add.ftz.f32 %f2445, %f2445, 0f322BCC77; mul.ftz.f32 %f1365, %f2445, %f2444; sub.ftz.f32 %f2443, %f1365, %f247; mul.ftz.f32 %f1366, %f239, %f2444; sub.ftz.f32 %f2442, %f249, %f1366; mul.ftz.f32 %f1367, %f2445, %f240; sub.ftz.f32 %f2441, %f251, %f1367; mul.ftz.f32 %f1368, %f241, %f2442; fma.rn.ftz.f32 %f1369, %f2446, %f2443, %f1368; fma.rn.ftz.f32 %f1370, %f244, %f2441, %f1369; setp.eq.ftz.f32 %p109, %f1370, 0f00000000; selp.f32 %f2447, 0f7F7FFFFF, %f1370, %p109; $L__BB207_100: mul.ftz.f32 %f1371, %f241, %f2444; mul.ftz.f32 %f1372, %f244, %f243; sub.ftz.f32 %f1373, %f1372, %f1371; mul.ftz.f32 %f1374, %f244, %f240; mul.ftz.f32 %f1375, %f2444, %f2446; sub.ftz.f32 %f1376, %f1375, %f1374; mul.ftz.f32 %f1377, %f243, %f2446; mul.ftz.f32 %f1378, %f241, %f240; sub.ftz.f32 %f1379, %f1378, %f1377; div.approx.ftz.f32 %f2452, %f1373, %f2447; div.approx.ftz.f32 %f2453, %f1376, %f2447; div.approx.ftz.f32 %f2454, %f1379, %f2447; mul.ftz.f32 %f1380, %f244, %f2445; mul.ftz.f32 %f1381, %f241, %f245; sub.ftz.f32 %f1382, %f1381, %f1380; mul.ftz.f32 %f1383, %f245, %f2446; mul.ftz.f32 %f1384, %f244, %f239; sub.ftz.f32 %f1385, %f1384, %f1383; mul.ftz.f32 %f1386, %f241, %f239; mul.ftz.f32 %f1387, %f2445, %f2446; sub.ftz.f32 %f1388, %f1387, %f1386; div.approx.ftz.f32 %f2456, %f1382, %f2447; div.approx.ftz.f32 %f2457, %f1385, %f2447; div.approx.ftz.f32 %f2458, %f1388, %f2447; div.approx.ftz.f32 %f2448, %f2443, %f2447; div.approx.ftz.f32 %f2449, %f2442, %f2447; mul.ftz.f32 %f1389, %f236, %f2449; fma.rn.ftz.f32 %f1390, %f235, %f2448, %f1389; div.approx.ftz.f32 %f2450, %f2441, %f2447; fma.rn.ftz.f32 %f1391, %f237, %f2450, %f1390; neg.ftz.f32 %f2451, %f1391; mul.ftz.f32 %f1392, %f236, %f2453; fma.rn.ftz.f32 %f1393, %f235, %f2452, %f1392; fma.rn.ftz.f32 %f1394, %f237, %f2454, %f1393; neg.ftz.f32 %f2455, %f1394; mul.ftz.f32 %f1395, %f236, %f2457; fma.rn.ftz.f32 %f1396, %f235, %f2456, %f1395; fma.rn.ftz.f32 %f1397, %f237, %f2458, %f1396; neg.ftz.f32 %f2459, %f1397; $L__BB207_102: mul.ftz.f32 %f1410, %f2449, %f2391; fma.rn.ftz.f32 %f1411, %f2448, %f2392, %f1410; fma.rn.ftz.f32 %f1412, %f2450, %f2390, %f1411; add.ftz.f32 %f2463, %f2451, %f1412; mul.ftz.f32 %f1413, %f2453, %f2391; fma.rn.ftz.f32 %f1414, %f2452, %f2392, %f1413; fma.rn.ftz.f32 %f1415, %f2454, %f2390, %f1414; add.ftz.f32 %f2464, %f2455, %f1415; mul.ftz.f32 %f1416, %f2457, %f2391; fma.rn.ftz.f32 %f1417, %f2456, %f2392, %f1416; fma.rn.ftz.f32 %f1418, %f2458, %f2390, %f1417; add.ftz.f32 %f2465, %f2459, %f1418; mul.ftz.f32 %f1419, %f2449, %f2428; fma.rn.ftz.f32 %f1420, %f2448, %f2429, %f1419; fma.rn.ftz.f32 %f1421, %f2450, %f2427, %f1420; mul.ftz.f32 %f1422, %f2453, %f2428; fma.rn.ftz.f32 %f1423, %f2452, %f2429, %f1422; fma.rn.ftz.f32 %f1424, %f2454, %f2427, %f1423; mul.ftz.f32 %f1425, %f2457, %f2428; fma.rn.ftz.f32 %f1426, %f2456, %f2429, %f1425; fma.rn.ftz.f32 %f1427, %f2458, %f2427, %f1426; abs.ftz.f32 %f1428, %f1421; setp.gt.ftz.f32 %p110, %f1428, 0f17800000; mov.b32 %r811, %f1421; and.b32 %r812, %r811, -2147483648; or.b32 %r813, %r812, 394264576; mov.b32 %f1429, %r813; selp.f32 %f2460, %f1421, %f1429, %p110; abs.ftz.f32 %f1430, %f1424; setp.gt.ftz.f32 %p111, %f1430, 0f17800000; mov.b32 %r814, %f1424; and.b32 %r815, %r814, -2147483648; or.b32 %r816, %r815, 394264576; mov.b32 %f1431, %r816; selp.f32 %f2461, %f1424, %f1431, %p111; abs.ftz.f32 %f1432, %f1427; setp.gt.ftz.f32 %p112, %f1432, 0f17800000; mov.b32 %r817, %f1427; and.b32 %r818, %r817, -2147483648; or.b32 %r819, %r818, 394264576; mov.b32 %f1433, %r819; selp.f32 %f2462, %f1427, %f1433, %p112; mov.u32 %r966, %r19; $L__BB207_103: rcp.approx.ftz.f32 %f316, %f2461; rcp.approx.ftz.f32 %f317, %f2460; rcp.approx.ftz.f32 %f318, %f2462; mov.u32 %r995, %r987; bra.uni $L__BB207_104; $L__BB207_117: cvta.to.global.u64 %rd212, %rd170; mul.wide.s32 %rd213, %r975, 16; xor.b64 %rd214, %rd213, -16; add.s64 %rd215, %rd212, %rd214; ld.global.v4.u32 {%r980, %r834, %r835, %r836}, [%rd215]; mul.wide.s32 %rd218, %r82, 4; add.s64 %rd219, %rd183, %rd218; ld.local.u32 %r975, [%rd219]; cvt.u16.u32 %rs28, %r836; and.b16 %rs27, %rs28, 127; add.s32 %r995, %r82, -1; setp.eq.s16 %p136, %rs27, 1; @%p136 bra $L__BB207_145; setp.ne.s16 %p137, %rs27, 65; @%p137 bra $L__BB207_159; add.s32 %r995, %r82, -1; setp.ge.s32 %p138, %r980, %r834; @%p138 bra $L__BB207_159; $L__BB207_120: cvt.s64.s32 %rd26, %r980; setp.ne.s32 %p139, %r966, -1; @%p139 bra $L__BB207_122; cvta.to.global.u64 %rd220, %rd172; shl.b64 %rd221, %rd26, 2; add.s64 %rd222, %rd220, %rd221; ld.global.u32 %r838, [%rd222]; setp.ne.s32 %p140, %r838, %r19; @%p140 bra $L__BB207_144; $L__BB207_122: cvta.to.global.u64 %rd223, %rd171; shl.b64 %rd224, %rd26, 2; add.s64 %rd225, %rd223, %rd224; ld.global.u32 %r92, [%rd225]; setp.eq.s32 %p141, %r965, %r92; @%p141 bra $L__BB207_144; ld.global.u32 %r93, [%rd9+108]; ld.global.u32 %r94, [%rd9+104]; shl.b32 %r839, %r94, 1; cvt.rn.f32.s32 %f1649, %r839; mul.ftz.f32 %f341, %f25, %f1649; cvt.rzi.ftz.s32.f32 %r840, %f341; add.s32 %r841, %r839, -1; min.s32 %r95, %r840, %r841; ld.global.u32 %r982, [%rd9+116]; mul.wide.u32 %rd226, %r982, 16; add.s64 %rd227, %rd16, %rd226; ld.global.u16 %rs50, [%rd227+12]; ld.global.u32 %r981, [%rd227+8]; ld.global.u64 %rd455, [%rd227]; setp.eq.s64 %p142, %rd455, 11; @%p142 bra $L__BB207_128; $L__BB207_124: setp.eq.s64 %p143, %rd455, 0; @%p143 bra $L__BB207_126; add.s32 %r982, %r982, 2; bra.uni $L__BB207_127; $L__BB207_126: setp.eq.s16 %p144, %rs50, 0; mov.u32 %r985, -1; mov.u32 %r982, %r981; @%p144 bra $L__BB207_129; $L__BB207_127: mul.wide.u32 %rd228, %r982, 16; add.s64 %rd229, %rd16, %rd228; ld.global.u16 %rs50, [%rd229+12]; ld.global.u32 %r981, [%rd229+8]; ld.global.u64 %rd455, [%rd229]; setp.ne.s64 %p145, %rd455, 11; @%p145 bra $L__BB207_124; $L__BB207_128: setp.eq.s16 %p146, %rs50, 0; selp.b32 %r985, -1, %r981, %p146; $L__BB207_129: mul.wide.s32 %rd230, %r92, 16; add.s64 %rd231, %rd14, %rd230; ld.global.v4.u32 {%r843, %r844, %r845, %r846}, [%rd231]; setp.eq.s32 %p147, %r95, %r94; @%p147 bra $L__BB207_131; bra.uni $L__BB207_130; $L__BB207_131: mul.wide.u32 %rd235, %r846, 12; add.s64 %rd456, %rd15, %rd235; add.s32 %r853, %r846, 1; mul.wide.u32 %rd236, %r853, 12; add.s64 %rd459, %rd15, %rd236; add.s32 %r854, %r846, 2; mul.wide.u32 %rd237, %r854, 12; add.s64 %rd462, %rd15, %rd237; add.s64 %rd457, %rd456, 4; add.s64 %rd458, %rd456, 8; add.s64 %rd460, %rd459, 4; add.s64 %rd461, %rd459, 8; add.s64 %rd463, %rd462, 4; add.s64 %rd464, %rd462, 8; bra.uni $L__BB207_132; $L__BB207_130: setp.gt.s32 %p148, %r95, %r94; selp.b32 %r847, -1, 0, %p148; add.s32 %r848, %r95, %r847; mad.lo.s32 %r849, %r848, %r93, %r985; add.s32 %r850, %r843, %r849; mul.wide.u32 %rd232, %r850, 12; add.s64 %rd456, %rd17, %rd232; add.s32 %r851, %r844, %r849; mul.wide.u32 %rd233, %r851, 12; add.s64 %rd459, %rd17, %rd233; add.s32 %r852, %r845, %r849; mul.wide.u32 %rd234, %r852, 12; add.s64 %rd462, %rd17, %rd234; add.s64 %rd457, %rd456, 4; add.s64 %rd458, %rd456, 8; add.s64 %rd460, %rd459, 4; add.s64 %rd461, %rd459, 8; add.s64 %rd463, %rd462, 4; add.s64 %rd464, %rd462, 8; $L__BB207_132: ld.global.f32 %f342, [%rd464]; ld.global.f32 %f343, [%rd463]; ld.global.f32 %f344, [%rd462]; ld.global.f32 %f345, [%rd461]; ld.global.f32 %f346, [%rd460]; ld.global.f32 %f347, [%rd459]; ld.global.f32 %f348, [%rd458]; ld.global.f32 %f349, [%rd457]; ld.global.f32 %f350, [%rd456]; add.s32 %r110, %r95, 1; setp.eq.s32 %p149, %r110, %r94; @%p149 bra $L__BB207_134; bra.uni $L__BB207_133; $L__BB207_134: mul.wide.u32 %rd241, %r846, 12; add.s64 %rd465, %rd15, %rd241; add.s32 %r861, %r846, 1; mul.wide.u32 %rd242, %r861, 12; add.s64 %rd468, %rd15, %rd242; add.s32 %r862, %r846, 2; mul.wide.u32 %rd243, %r862, 12; add.s64 %rd471, %rd15, %rd243; add.s64 %rd466, %rd465, 4; add.s64 %rd467, %rd465, 8; add.s64 %rd469, %rd468, 4; add.s64 %rd470, %rd468, 8; add.s64 %rd472, %rd471, 4; add.s64 %rd473, %rd471, 8; bra.uni $L__BB207_135; $L__BB207_133: setp.ge.s32 %p150, %r95, %r94; selp.b32 %r855, -1, 0, %p150; add.s32 %r856, %r110, %r855; mad.lo.s32 %r857, %r856, %r93, %r985; add.s32 %r858, %r843, %r857; mul.wide.u32 %rd238, %r858, 12; add.s64 %rd465, %rd17, %rd238; add.s32 %r859, %r844, %r857; mul.wide.u32 %rd239, %r859, 12; add.s64 %rd468, %rd17, %rd239; add.s32 %r860, %r845, %r857; mul.wide.u32 %rd240, %r860, 12; add.s64 %rd471, %rd17, %rd240; add.s64 %rd466, %rd465, 4; add.s64 %rd467, %rd465, 8; add.s64 %rd469, %rd468, 4; add.s64 %rd470, %rd468, 8; add.s64 %rd472, %rd471, 4; add.s64 %rd473, %rd471, 8; $L__BB207_135: cvt.rn.f32.s32 %f1650, %r95; sub.ftz.f32 %f1651, %f341, %f1650; mov.f32 %f1652, 0f3F800000; sub.ftz.f32 %f1653, %f1652, %f1651; ld.global.f32 %f1654, [%rd465]; mul.ftz.f32 %f1655, %f1651, %f1654; ld.global.f32 %f1656, [%rd466]; mul.ftz.f32 %f1657, %f1651, %f1656; ld.global.f32 %f1658, [%rd467]; mul.ftz.f32 %f1659, %f1651, %f1658; fma.rn.ftz.f32 %f351, %f1653, %f350, %f1655; fma.rn.ftz.f32 %f352, %f1653, %f349, %f1657; fma.rn.ftz.f32 %f353, %f1653, %f348, %f1659; ld.global.f32 %f1660, [%rd468]; mul.ftz.f32 %f1661, %f1651, %f1660; ld.global.f32 %f1662, [%rd469]; mul.ftz.f32 %f1663, %f1651, %f1662; ld.global.f32 %f1664, [%rd470]; mul.ftz.f32 %f1665, %f1651, %f1664; fma.rn.ftz.f32 %f354, %f1653, %f347, %f1661; fma.rn.ftz.f32 %f355, %f1653, %f346, %f1663; fma.rn.ftz.f32 %f356, %f1653, %f345, %f1665; ld.global.f32 %f1666, [%rd471]; mul.ftz.f32 %f1667, %f1651, %f1666; ld.global.f32 %f1668, [%rd472]; mul.ftz.f32 %f1669, %f1651, %f1668; ld.global.f32 %f1670, [%rd473]; mul.ftz.f32 %f1671, %f1651, %f1670; fma.rn.ftz.f32 %f357, %f1653, %f344, %f1667; fma.rn.ftz.f32 %f358, %f1653, %f343, %f1669; fma.rn.ftz.f32 %f359, %f1653, %f342, %f1671; sub.ftz.f32 %f360, %f351, %f2463; sub.ftz.f32 %f361, %f352, %f2464; sub.ftz.f32 %f362, %f353, %f2465; sub.ftz.f32 %f1672, %f354, %f2463; sub.ftz.f32 %f1673, %f355, %f2464; sub.ftz.f32 %f1674, %f356, %f2465; sub.ftz.f32 %f1675, %f357, %f2463; sub.ftz.f32 %f1676, %f358, %f2464; sub.ftz.f32 %f1677, %f359, %f2465; sub.ftz.f32 %f363, %f1675, %f360; sub.ftz.f32 %f364, %f1676, %f361; sub.ftz.f32 %f365, %f1677, %f362; sub.ftz.f32 %f366, %f360, %f1672; sub.ftz.f32 %f367, %f361, %f1673; sub.ftz.f32 %f368, %f362, %f1674; sub.ftz.f32 %f1678, %f1672, %f1675; sub.ftz.f32 %f1679, %f1673, %f1676; sub.ftz.f32 %f1680, %f1674, %f1677; add.ftz.f32 %f1681, %f1675, %f360; add.ftz.f32 %f1682, %f1676, %f361; add.ftz.f32 %f1683, %f1677, %f362; mul.ftz.f32 %f1684, %f1683, %f364; mul.ftz.f32 %f1685, %f365, %f1682; sub.ftz.f32 %f1686, %f1684, %f1685; mul.ftz.f32 %f1687, %f365, %f1681; mul.ftz.f32 %f1688, %f1683, %f363; sub.ftz.f32 %f1689, %f1687, %f1688; mul.ftz.f32 %f1690, %f1682, %f363; mul.ftz.f32 %f1691, %f364, %f1681; sub.ftz.f32 %f1692, %f1690, %f1691; mul.ftz.f32 %f1693, %f2461, %f1689; fma.rn.ftz.f32 %f1694, %f2460, %f1686, %f1693; fma.rn.ftz.f32 %f369, %f2462, %f1692, %f1694; add.ftz.f32 %f1695, %f1672, %f360; add.ftz.f32 %f1696, %f1673, %f361; add.ftz.f32 %f1697, %f1674, %f362; mul.ftz.f32 %f1698, %f1697, %f367; mul.ftz.f32 %f1699, %f368, %f1696; sub.ftz.f32 %f1700, %f1698, %f1699; mul.ftz.f32 %f1701, %f368, %f1695; mul.ftz.f32 %f1702, %f1697, %f366; sub.ftz.f32 %f1703, %f1701, %f1702; mul.ftz.f32 %f1704, %f1696, %f366; mul.ftz.f32 %f1705, %f367, %f1695; sub.ftz.f32 %f1706, %f1704, %f1705; mul.ftz.f32 %f1707, %f2461, %f1703; fma.rn.ftz.f32 %f1708, %f2460, %f1700, %f1707; fma.rn.ftz.f32 %f370, %f2462, %f1706, %f1708; add.ftz.f32 %f1709, %f1675, %f1672; add.ftz.f32 %f1710, %f1676, %f1673; add.ftz.f32 %f1711, %f1677, %f1674; mul.ftz.f32 %f1712, %f1711, %f1679; mul.ftz.f32 %f1713, %f1680, %f1710; sub.ftz.f32 %f1714, %f1712, %f1713; mul.ftz.f32 %f1715, %f1680, %f1709; mul.ftz.f32 %f1716, %f1711, %f1678; sub.ftz.f32 %f1717, %f1715, %f1716; mul.ftz.f32 %f1718, %f1710, %f1678; mul.ftz.f32 %f1719, %f1679, %f1709; sub.ftz.f32 %f1720, %f1718, %f1719; mul.ftz.f32 %f1721, %f2461, %f1717; fma.rn.ftz.f32 %f1722, %f2460, %f1714, %f1721; fma.rn.ftz.f32 %f1723, %f2462, %f1720, %f1722; add.ftz.f32 %f1724, %f369, %f370; add.ftz.f32 %f371, %f1723, %f1724; abs.ftz.f32 %f372, %f371; mul.ftz.f32 %f1725, %f372, 0f34000000; min.ftz.f32 %f1726, %f370, %f1723; min.ftz.f32 %f1727, %f369, %f1726; max.ftz.f32 %f1728, %f370, %f1723; max.ftz.f32 %f1729, %f369, %f1728; mul.ftz.f32 %f1730, %f372, 0fB4000000; setp.ltu.ftz.f32 %p152, %f1727, %f1730; setp.gtu.ftz.f32 %p153, %f1729, %f1725; and.pred %p154, %p152, %p153; mov.pred %p280, -1; @%p154 bra $L__BB207_141; mov.pred %p280, -1; mul.ftz.f32 %f1731, %f365, %f367; mul.ftz.f32 %f1732, %f368, %f364; sub.ftz.f32 %f1733, %f1731, %f1732; mul.ftz.f32 %f1734, %f365, %f366; mul.ftz.f32 %f1735, %f368, %f363; sub.ftz.f32 %f1736, %f1735, %f1734; mul.ftz.f32 %f1737, %f367, %f363; mul.ftz.f32 %f1738, %f364, %f366; sub.ftz.f32 %f1739, %f1738, %f1737; add.ftz.f32 %f373, %f1733, %f1733; add.ftz.f32 %f374, %f1736, %f1736; add.ftz.f32 %f375, %f1739, %f1739; mul.ftz.f32 %f1740, %f2461, %f374; fma.rn.ftz.f32 %f1741, %f2460, %f373, %f1740; fma.rn.ftz.f32 %f376, %f2462, %f375, %f1741; setp.eq.ftz.f32 %p156, %f376, 0f00000000; @%p156 bra $L__BB207_141; mov.pred %p280, -1; mul.ftz.f32 %f1742, %f361, %f374; fma.rn.ftz.f32 %f1743, %f360, %f373, %f1742; fma.rn.ftz.f32 %f1744, %f362, %f375, %f1743; div.approx.ftz.f32 %f377, %f1744, %f376; setp.ltu.ftz.f32 %p158, %f377, 0f00000000; setp.gtu.ftz.f32 %p159, %f377, %f2436; or.pred %p160, %p159, %p158; @%p160 bra $L__BB207_141; abs.ftz.f32 %f2362, %f371; setp.lt.ftz.f32 %p161, %f2362, 0f219392EF; mov.f32 %f2483, 0f00000000; @%p161 bra $L__BB207_140; rcp.approx.ftz.f32 %f2483, %f371; $L__BB207_140: mov.f32 %f2363, 0f3F800000; mul.ftz.f32 %f1746, %f369, %f2483; min.ftz.f32 %f2491, %f1746, %f2363; mul.ftz.f32 %f1748, %f370, %f2483; min.ftz.f32 %f2490, %f1748, %f2363; mov.pred %p280, 0; mov.f32 %f2492, %f377; $L__BB207_141: @%p280 bra $L__BB207_144; setp.ne.s32 %p163, %r987, 0; setp.gt.ftz.f32 %p164, %f2492, %f2487; and.pred %p165, %p163, %p164; @%p165 bra $L__BB207_144; add.u64 %rd443, %SPL, 828; mov.u32 %r987, 1; st.local.u32 [%rd443], %r987; st.local.f32 [%rd443+4], %f2492; st.local.f32 [%rd443+8], %f2491; st.local.f32 [%rd443+12], %f2490; st.local.u32 [%rd443+16], %r92; st.local.u32 [%rd443+20], %r19; mov.u32 %r865, 65; st.local.u32 [%rd443+24], %r865; sub.ftz.f32 %f1749, %f359, %f353; sub.ftz.f32 %f1750, %f355, %f352; mul.ftz.f32 %f1751, %f1749, %f1750; sub.ftz.f32 %f1752, %f358, %f352; sub.ftz.f32 %f1753, %f356, %f353; mul.ftz.f32 %f1754, %f1753, %f1752; sub.ftz.f32 %f1755, %f1751, %f1754; sub.ftz.f32 %f1756, %f357, %f351; mul.ftz.f32 %f1757, %f1753, %f1756; sub.ftz.f32 %f1758, %f354, %f351; mul.ftz.f32 %f1759, %f1749, %f1758; sub.ftz.f32 %f1760, %f1757, %f1759; mul.ftz.f32 %f1761, %f1752, %f1758; mul.ftz.f32 %f1762, %f1750, %f1756; sub.ftz.f32 %f1763, %f1761, %f1762; mul.ftz.f32 %f1764, %f1760, %f1760; fma.rn.ftz.f32 %f1765, %f1755, %f1755, %f1764; fma.rn.ftz.f32 %f1766, %f1763, %f1763, %f1765; sqrt.approx.ftz.f32 %f1767, %f1766; div.approx.ftz.f32 %f1768, %f1755, %f1767; div.approx.ftz.f32 %f1769, %f1760, %f1767; div.approx.ftz.f32 %f1770, %f1763, %f1767; st.local.f32 [%rd443+100], %f1768; st.local.f32 [%rd443+104], %f1769; st.local.f32 [%rd443+108], %f1770; mov.f32 %f2487, %f2492; $L__BB207_144: add.s32 %r995, %r82, -1; cvt.u32.u64 %r866, %rd26; add.s32 %r980, %r866, 1; setp.lt.s32 %p166, %r980, %r834; @%p166 bra $L__BB207_120; bra.uni $L__BB207_159; $L__BB207_145: setp.ge.s32 %p167, %r980, %r834; @%p167 bra $L__BB207_159; $L__BB207_146: cvt.s64.s32 %rd84, %r980; setp.ne.s32 %p168, %r966, -1; @%p168 bra $L__BB207_148; cvta.to.global.u64 %rd246, %rd172; shl.b64 %rd247, %rd84, 2; add.s64 %rd248, %rd246, %rd247; ld.global.u32 %r867, [%rd248]; setp.ne.s32 %p169, %r867, %r19; @%p169 bra $L__BB207_158; $L__BB207_148: cvta.to.global.u64 %rd249, %rd171; shl.b64 %rd250, %rd84, 2; add.s64 %rd251, %rd249, %rd250; ld.global.u32 %r119, [%rd251]; setp.eq.s32 %p170, %r965, %r119; @%p170 bra $L__BB207_158; mul.wide.s32 %rd252, %r119, 16; add.s64 %rd253, %rd14, %rd252; ld.global.u32 %r868, [%rd253+12]; mul.wide.u32 %rd254, %r868, 12; add.s64 %rd255, %rd15, %rd254; add.s32 %r869, %r868, 1; mul.wide.u32 %rd256, %r869, 12; add.s64 %rd257, %rd15, %rd256; add.s32 %r870, %r868, 2; mul.wide.u32 %rd258, %r870, 12; add.s64 %rd259, %rd15, %rd258; ld.global.f32 %f397, [%rd255]; sub.ftz.f32 %f398, %f397, %f2463; ld.global.f32 %f399, [%rd255+4]; sub.ftz.f32 %f400, %f399, %f2464; ld.global.f32 %f401, [%rd255+8]; sub.ftz.f32 %f402, %f401, %f2465; ld.global.f32 %f403, [%rd257]; sub.ftz.f32 %f1771, %f403, %f2463; ld.global.f32 %f404, [%rd257+4]; sub.ftz.f32 %f1772, %f404, %f2464; ld.global.f32 %f405, [%rd257+8]; sub.ftz.f32 %f1773, %f405, %f2465; ld.global.f32 %f406, [%rd259]; sub.ftz.f32 %f1774, %f406, %f2463; ld.global.f32 %f407, [%rd259+4]; sub.ftz.f32 %f1775, %f407, %f2464; ld.global.f32 %f408, [%rd259+8]; sub.ftz.f32 %f1776, %f408, %f2465; sub.ftz.f32 %f409, %f1774, %f398; sub.ftz.f32 %f410, %f1775, %f400; sub.ftz.f32 %f411, %f1776, %f402; sub.ftz.f32 %f412, %f398, %f1771; sub.ftz.f32 %f413, %f400, %f1772; sub.ftz.f32 %f414, %f402, %f1773; sub.ftz.f32 %f1777, %f1771, %f1774; sub.ftz.f32 %f1778, %f1772, %f1775; sub.ftz.f32 %f1779, %f1773, %f1776; add.ftz.f32 %f1780, %f398, %f1774; add.ftz.f32 %f1781, %f400, %f1775; add.ftz.f32 %f1782, %f402, %f1776; mul.ftz.f32 %f1783, %f410, %f1782; mul.ftz.f32 %f1784, %f1781, %f411; sub.ftz.f32 %f1785, %f1783, %f1784; mul.ftz.f32 %f1786, %f1780, %f411; mul.ftz.f32 %f1787, %f409, %f1782; sub.ftz.f32 %f1788, %f1786, %f1787; mul.ftz.f32 %f1789, %f409, %f1781; mul.ftz.f32 %f1790, %f1780, %f410; sub.ftz.f32 %f1791, %f1789, %f1790; mul.ftz.f32 %f1792, %f2461, %f1788; fma.rn.ftz.f32 %f1793, %f2460, %f1785, %f1792; fma.rn.ftz.f32 %f415, %f2462, %f1791, %f1793; add.ftz.f32 %f1794, %f398, %f1771; add.ftz.f32 %f1795, %f400, %f1772; add.ftz.f32 %f1796, %f402, %f1773; mul.ftz.f32 %f1797, %f413, %f1796; mul.ftz.f32 %f1798, %f1795, %f414; sub.ftz.f32 %f1799, %f1797, %f1798; mul.ftz.f32 %f1800, %f1794, %f414; mul.ftz.f32 %f1801, %f412, %f1796; sub.ftz.f32 %f1802, %f1800, %f1801; mul.ftz.f32 %f1803, %f412, %f1795; mul.ftz.f32 %f1804, %f1794, %f413; sub.ftz.f32 %f1805, %f1803, %f1804; mul.ftz.f32 %f1806, %f2461, %f1802; fma.rn.ftz.f32 %f1807, %f2460, %f1799, %f1806; fma.rn.ftz.f32 %f416, %f2462, %f1805, %f1807; add.ftz.f32 %f1808, %f1771, %f1774; add.ftz.f32 %f1809, %f1772, %f1775; add.ftz.f32 %f1810, %f1773, %f1776; mul.ftz.f32 %f1811, %f1778, %f1810; mul.ftz.f32 %f1812, %f1809, %f1779; sub.ftz.f32 %f1813, %f1811, %f1812; mul.ftz.f32 %f1814, %f1808, %f1779; mul.ftz.f32 %f1815, %f1777, %f1810; sub.ftz.f32 %f1816, %f1814, %f1815; mul.ftz.f32 %f1817, %f1777, %f1809; mul.ftz.f32 %f1818, %f1808, %f1778; sub.ftz.f32 %f1819, %f1817, %f1818; mul.ftz.f32 %f1820, %f2461, %f1816; fma.rn.ftz.f32 %f1821, %f2460, %f1813, %f1820; fma.rn.ftz.f32 %f1822, %f2462, %f1819, %f1821; add.ftz.f32 %f1823, %f416, %f415; add.ftz.f32 %f417, %f1822, %f1823; abs.ftz.f32 %f418, %f417; mul.ftz.f32 %f1824, %f418, 0f34000000; min.ftz.f32 %f1825, %f416, %f1822; min.ftz.f32 %f1826, %f415, %f1825; max.ftz.f32 %f1827, %f416, %f1822; max.ftz.f32 %f1828, %f415, %f1827; mul.ftz.f32 %f1829, %f418, 0fB4000000; setp.ltu.ftz.f32 %p172, %f1826, %f1829; setp.gtu.ftz.f32 %p173, %f1828, %f1824; and.pred %p174, %p172, %p173; mov.pred %p281, -1; @%p174 bra $L__BB207_155; mov.pred %p281, -1; mul.ftz.f32 %f1830, %f413, %f411; mul.ftz.f32 %f1831, %f414, %f410; sub.ftz.f32 %f1832, %f1830, %f1831; mul.ftz.f32 %f1833, %f412, %f411; mul.ftz.f32 %f1834, %f414, %f409; sub.ftz.f32 %f1835, %f1834, %f1833; mul.ftz.f32 %f1836, %f413, %f409; mul.ftz.f32 %f1837, %f412, %f410; sub.ftz.f32 %f1838, %f1837, %f1836; add.ftz.f32 %f419, %f1832, %f1832; add.ftz.f32 %f420, %f1835, %f1835; add.ftz.f32 %f421, %f1838, %f1838; mul.ftz.f32 %f1839, %f2461, %f420; fma.rn.ftz.f32 %f1840, %f2460, %f419, %f1839; fma.rn.ftz.f32 %f422, %f2462, %f421, %f1840; setp.eq.ftz.f32 %p176, %f422, 0f00000000; @%p176 bra $L__BB207_155; mov.pred %p281, -1; sub.ftz.f32 %f2358, %f401, %f2465; sub.ftz.f32 %f2357, %f397, %f2463; sub.ftz.f32 %f2356, %f399, %f2464; mul.ftz.f32 %f1841, %f2356, %f420; fma.rn.ftz.f32 %f1842, %f2357, %f419, %f1841; fma.rn.ftz.f32 %f1843, %f2358, %f421, %f1842; div.approx.ftz.f32 %f423, %f1843, %f422; setp.ltu.ftz.f32 %p178, %f423, 0f00000000; setp.gtu.ftz.f32 %p179, %f423, %f2436; or.pred %p180, %p179, %p178; @%p180 bra $L__BB207_155; setp.lt.ftz.f32 %p181, %f418, 0f219392EF; mov.f32 %f2499, 0f00000000; @%p181 bra $L__BB207_154; rcp.approx.ftz.f32 %f2499, %f417; $L__BB207_154: mul.ftz.f32 %f1845, %f415, %f2499; mov.f32 %f1846, 0f3F800000; min.ftz.f32 %f2507, %f1845, %f1846; mul.ftz.f32 %f1847, %f416, %f2499; min.ftz.f32 %f2506, %f1847, %f1846; mov.pred %p281, 0; mov.f32 %f2508, %f423; $L__BB207_155: @%p281 bra $L__BB207_158; setp.ne.s32 %p183, %r987, 0; setp.gt.ftz.f32 %p184, %f2508, %f2487; and.pred %p185, %p183, %p184; @%p185 bra $L__BB207_158; add.u64 %rd441, %SPL, 828; mov.u32 %r987, 1; st.local.u32 [%rd441], %r987; st.local.u32 [%rd441+16], %r119; st.local.u32 [%rd441+20], %r19; st.local.u32 [%rd441+24], %r987; st.local.f32 [%rd441+8], %f2507; st.local.f32 [%rd441+12], %f2506; st.local.f32 [%rd441+4], %f2508; sub.ftz.f32 %f1848, %f408, %f401; sub.ftz.f32 %f1849, %f404, %f399; mul.ftz.f32 %f1850, %f1849, %f1848; sub.ftz.f32 %f1851, %f407, %f399; sub.ftz.f32 %f1852, %f405, %f401; mul.ftz.f32 %f1853, %f1852, %f1851; sub.ftz.f32 %f1854, %f1850, %f1853; sub.ftz.f32 %f1855, %f406, %f397; mul.ftz.f32 %f1856, %f1852, %f1855; sub.ftz.f32 %f1857, %f403, %f397; mul.ftz.f32 %f1858, %f1857, %f1848; sub.ftz.f32 %f1859, %f1856, %f1858; mul.ftz.f32 %f1860, %f1857, %f1851; mul.ftz.f32 %f1861, %f1849, %f1855; sub.ftz.f32 %f1862, %f1860, %f1861; mul.ftz.f32 %f1863, %f1859, %f1859; fma.rn.ftz.f32 %f1864, %f1854, %f1854, %f1863; fma.rn.ftz.f32 %f1865, %f1862, %f1862, %f1864; sqrt.approx.ftz.f32 %f1866, %f1865; div.approx.ftz.f32 %f1867, %f1854, %f1866; div.approx.ftz.f32 %f1868, %f1859, %f1866; div.approx.ftz.f32 %f1869, %f1862, %f1866; st.local.f32 [%rd441+100], %f1867; st.local.f32 [%rd441+104], %f1868; st.local.f32 [%rd441+108], %f1869; mov.f32 %f2487, %f2508; $L__BB207_158: add.s32 %r995, %r82, -1; cvt.u32.u64 %r873, %rd84; add.s32 %r980, %r873, 1; setp.lt.s32 %p186, %r980, %r834; @%p186 bra $L__BB207_146; bra.uni $L__BB207_159; $L__BB207_104: setp.eq.s32 %p113, %r975, 1985229328; setp.lt.s32 %p114, %r975, 0; or.pred %p115, %p114, %p113; @%p115 bra $L__BB207_116; cvta.to.global.u64 %rd23, %rd169; $L__BB207_106: mul.wide.s32 %rd209, %r975, 16; add.s64 %rd24, %rd23, %rd209; ld.global.v4.f32 {%f1434, %f1435, %f1436, %f1437}, [%rd24]; mov.b32 %r975, %f1437; mov.b32 %r67, %f1436; mov.b32 %r66, %f1435; mov.b32 %r65, %f1434; mov.b64 %rd210, {%r65, %r66}; cvt.u32.u64 %r823, %rd210; and.b32 %r824, %r823, 1024; setp.eq.s32 %p116, %r824, 0; and.b32 %r69, %r65, 2047; @%p116 bra $L__BB207_109; bra.uni $L__BB207_107; $L__BB207_109: setp.ne.s32 %p122, %r69, 0; ld.global.v4.f32 {%f1578, %f1579, %f1580, %f1581}, [%rd24+16]; ld.global.v4.f32 {%f1586, %f1587, %f1588, %f1589}, [%rd24+32]; ld.global.v4.f32 {%f1594, %f1595, %f1596, %f1597}, [%rd24+48]; sub.ftz.f32 %f1602, %f1578, %f2463; mul.ftz.f32 %f1603, %f317, %f1602; sub.ftz.f32 %f1604, %f1580, %f2463; mul.ftz.f32 %f1605, %f317, %f1604; sub.ftz.f32 %f1606, %f1586, %f2464; mul.ftz.f32 %f1607, %f316, %f1606; sub.ftz.f32 %f1608, %f1588, %f2464; mul.ftz.f32 %f1609, %f316, %f1608; sub.ftz.f32 %f1610, %f1594, %f2465; mul.ftz.f32 %f1611, %f318, %f1610; sub.ftz.f32 %f1612, %f1596, %f2465; mul.ftz.f32 %f1613, %f318, %f1612; min.ftz.f32 %f1614, %f1603, %f1605; min.ftz.f32 %f1615, %f1607, %f1609; min.ftz.f32 %f1616, %f1611, %f1613; mov.f32 %f1617, 0f00000000; max.ftz.f32 %f1618, %f1617, %f1614; max.ftz.f32 %f1619, %f1615, %f1616; max.ftz.f32 %f2475, %f1618, %f1619; max.ftz.f32 %f1620, %f1603, %f1605; max.ftz.f32 %f1621, %f1607, %f1609; max.ftz.f32 %f1622, %f1611, %f1613; min.ftz.f32 %f1623, %f2436, %f1620; min.ftz.f32 %f1624, %f1621, %f1622; min.ftz.f32 %f1625, %f1623, %f1624; sub.ftz.f32 %f1626, %f1579, %f2463; mul.ftz.f32 %f1627, %f317, %f1626; sub.ftz.f32 %f1628, %f1581, %f2463; mul.ftz.f32 %f1629, %f317, %f1628; sub.ftz.f32 %f1630, %f1587, %f2464; mul.ftz.f32 %f1631, %f316, %f1630; sub.ftz.f32 %f1632, %f1589, %f2464; mul.ftz.f32 %f1633, %f316, %f1632; sub.ftz.f32 %f1634, %f1595, %f2465; mul.ftz.f32 %f1635, %f318, %f1634; sub.ftz.f32 %f1636, %f1597, %f2465; mul.ftz.f32 %f1637, %f318, %f1636; min.ftz.f32 %f1638, %f1627, %f1629; min.ftz.f32 %f1639, %f1631, %f1633; min.ftz.f32 %f1640, %f1635, %f1637; max.ftz.f32 %f1641, %f1617, %f1638; max.ftz.f32 %f1642, %f1639, %f1640; max.ftz.f32 %f2476, %f1641, %f1642; max.ftz.f32 %f1643, %f1627, %f1629; max.ftz.f32 %f1644, %f1631, %f1633; max.ftz.f32 %f1645, %f1635, %f1637; min.ftz.f32 %f1646, %f2436, %f1643; min.ftz.f32 %f1647, %f1644, %f1645; min.ftz.f32 %f1648, %f1646, %f1647; setp.ge.ftz.f32 %p123, %f1625, %f2475; and.pred %p124, %p122, %p123; selp.u32 %r828, 1, 0, %p124; setp.ltu.ftz.f32 %p125, %f1648, %f2476; and.b32 %r830, %r66, 2047; setp.eq.s32 %p126, %r830, 0; or.pred %p127, %p126, %p125; selp.b32 %r831, 0, 2, %p127; or.b32 %r973, %r831, %r828; bra.uni $L__BB207_110; $L__BB207_107: setp.ne.s32 %p117, %r69, 0; ld.global.v4.f32 {%f1441, %f1442, %f1443, %f1444}, [%rd24+16]; ld.global.v4.f32 {%f1449, %f1450, %f1451, %f1452}, [%rd24+32]; ld.global.v4.f32 {%f1457, %f1458, %f1459, %f1460}, [%rd24+48]; mul.ftz.f32 %f1465, %f2461, %f1442; fma.rn.ftz.f32 %f1466, %f2460, %f1441, %f1465; fma.rn.ftz.f32 %f1467, %f2462, %f1443, %f1466; mul.ftz.f32 %f1468, %f2461, %f1450; fma.rn.ftz.f32 %f1469, %f2460, %f1449, %f1468; fma.rn.ftz.f32 %f1470, %f2462, %f1451, %f1469; mul.ftz.f32 %f1471, %f2461, %f1458; fma.rn.ftz.f32 %f1472, %f2460, %f1457, %f1471; fma.rn.ftz.f32 %f1473, %f2462, %f1459, %f1472; mul.ftz.f32 %f1474, %f2464, %f1442; fma.rn.ftz.f32 %f1475, %f2463, %f1441, %f1474; fma.rn.ftz.f32 %f1476, %f2465, %f1443, %f1475; add.ftz.f32 %f1477, %f1444, %f1476; mul.ftz.f32 %f1478, %f2464, %f1450; fma.rn.ftz.f32 %f1479, %f2463, %f1449, %f1478; fma.rn.ftz.f32 %f1480, %f2465, %f1451, %f1479; add.ftz.f32 %f1481, %f1452, %f1480; mul.ftz.f32 %f1482, %f2464, %f1458; fma.rn.ftz.f32 %f1483, %f2463, %f1457, %f1482; fma.rn.ftz.f32 %f1484, %f2465, %f1459, %f1483; add.ftz.f32 %f1485, %f1460, %f1484; rcp.approx.ftz.f32 %f1486, %f1467; rcp.approx.ftz.f32 %f1487, %f1470; rcp.approx.ftz.f32 %f1488, %f1473; mul.ftz.f32 %f1489, %f1486, %f1477; neg.ftz.f32 %f1490, %f1489; mul.ftz.f32 %f1491, %f1487, %f1481; neg.ftz.f32 %f1492, %f1491; mul.ftz.f32 %f1493, %f1488, %f1485; neg.ftz.f32 %f1494, %f1493; sub.ftz.f32 %f1495, %f1486, %f1489; sub.ftz.f32 %f1496, %f1487, %f1491; sub.ftz.f32 %f1497, %f1488, %f1493; min.ftz.f32 %f1498, %f1490, %f1495; min.ftz.f32 %f1499, %f1492, %f1496; min.ftz.f32 %f1500, %f1494, %f1497; max.ftz.f32 %f1501, %f1490, %f1495; max.ftz.f32 %f1502, %f1492, %f1496; max.ftz.f32 %f1503, %f1494, %f1497; mov.f32 %f1504, 0f00000000; max.ftz.f32 %f1505, %f1504, %f1498; max.ftz.f32 %f1506, %f1499, %f1500; max.ftz.f32 %f2475, %f1505, %f1506; min.ftz.f32 %f1507, %f2436, %f1501; min.ftz.f32 %f1508, %f1502, %f1503; min.ftz.f32 %f1509, %f1507, %f1508; setp.le.ftz.f32 %p118, %f2475, %f1509; and.pred %p119, %p117, %p118; selp.u32 %r973, 1, 0, %p119; ld.global.v4.f32 {%f1510, %f1511, %f1512, %f1513}, [%rd24+64]; ld.global.v4.f32 {%f1518, %f1519, %f1520, %f1521}, [%rd24+80]; ld.global.v4.f32 {%f1526, %f1527, %f1528, %f1529}, [%rd24+96]; mul.ftz.f32 %f1534, %f2461, %f1511; fma.rn.ftz.f32 %f1535, %f2460, %f1510, %f1534; fma.rn.ftz.f32 %f1536, %f2462, %f1512, %f1535; mul.ftz.f32 %f1537, %f2461, %f1519; fma.rn.ftz.f32 %f1538, %f2460, %f1518, %f1537; fma.rn.ftz.f32 %f1539, %f2462, %f1520, %f1538; mul.ftz.f32 %f1540, %f2461, %f1527; fma.rn.ftz.f32 %f1541, %f2460, %f1526, %f1540; fma.rn.ftz.f32 %f1542, %f2462, %f1528, %f1541; mul.ftz.f32 %f1543, %f2464, %f1511; fma.rn.ftz.f32 %f1544, %f2463, %f1510, %f1543; fma.rn.ftz.f32 %f1545, %f2465, %f1512, %f1544; add.ftz.f32 %f1546, %f1513, %f1545; mul.ftz.f32 %f1547, %f2464, %f1519; fma.rn.ftz.f32 %f1548, %f2463, %f1518, %f1547; fma.rn.ftz.f32 %f1549, %f2465, %f1520, %f1548; add.ftz.f32 %f1550, %f1521, %f1549; mul.ftz.f32 %f1551, %f2464, %f1527; fma.rn.ftz.f32 %f1552, %f2463, %f1526, %f1551; fma.rn.ftz.f32 %f1553, %f2465, %f1528, %f1552; add.ftz.f32 %f1554, %f1529, %f1553; rcp.approx.ftz.f32 %f1555, %f1536; rcp.approx.ftz.f32 %f1556, %f1539; rcp.approx.ftz.f32 %f1557, %f1542; mul.ftz.f32 %f1558, %f1555, %f1546; neg.ftz.f32 %f1559, %f1558; mul.ftz.f32 %f1560, %f1556, %f1550; neg.ftz.f32 %f1561, %f1560; mul.ftz.f32 %f1562, %f1557, %f1554; neg.ftz.f32 %f1563, %f1562; sub.ftz.f32 %f1564, %f1555, %f1558; sub.ftz.f32 %f1565, %f1556, %f1560; sub.ftz.f32 %f1566, %f1557, %f1562; min.ftz.f32 %f1567, %f1559, %f1564; min.ftz.f32 %f1568, %f1561, %f1565; min.ftz.f32 %f1569, %f1563, %f1566; max.ftz.f32 %f1570, %f1559, %f1564; max.ftz.f32 %f1571, %f1561, %f1565; max.ftz.f32 %f1572, %f1563, %f1566; max.ftz.f32 %f1573, %f1504, %f1567; max.ftz.f32 %f1574, %f1568, %f1569; max.ftz.f32 %f2476, %f1573, %f1574; min.ftz.f32 %f1575, %f2436, %f1570; min.ftz.f32 %f1576, %f1571, %f1572; min.ftz.f32 %f1577, %f1575, %f1576; setp.gtu.ftz.f32 %p120, %f2476, %f1577; @%p120 bra $L__BB207_110; and.b32 %r826, %r66, 2047; setp.eq.s32 %p121, %r826, 0; or.b32 %r827, %r973, 2; selp.b32 %r973, %r973, %r827, %p121; $L__BB207_110: mul.wide.s32 %rd211, %r995, 4; add.s64 %rd25, %rd183, %rd211; setp.eq.s32 %p128, %r973, 3; @%p128 bra $L__BB207_114; bra.uni $L__BB207_111; $L__BB207_114: setp.lt.ftz.f32 %p131, %f2476, %f2475; selp.b32 %r78, %r975, %r67, %p131; selp.b32 %r832, %r67, %r975, %p131; add.s32 %r995, %r995, 1; st.local.u32 [%rd25+4], %r832; mov.u32 %r975, %r78; bra.uni $L__BB207_115; $L__BB207_111: setp.eq.s32 %p129, %r973, 2; @%p129 bra $L__BB207_115; setp.ne.s32 %p130, %r973, 0; mov.u32 %r975, %r67; @%p130 bra $L__BB207_115; ld.local.u32 %r975, [%rd25]; add.s32 %r995, %r995, -1; $L__BB207_115: setp.ne.s32 %p132, %r975, 1985229328; setp.gt.s32 %p133, %r975, -1; and.pred %p134, %p133, %p132; @%p134 bra $L__BB207_106; $L__BB207_116: mov.u32 %r82, %r995; setp.gt.s32 %p135, %r975, -1; @%p135 bra $L__BB207_159; bra.uni $L__BB207_117; $L__BB207_159: setp.ne.s32 %p187, %r975, 1985229328; @%p187 bra $L__BB207_104; $L__BB207_192: setp.ne.s32 %p273, %r958, 0; setp.gt.s32 %p283, %r987, 0; selp.f32 %f532, %f2487, %f2436, %p283; @%p273 bra $L__BB207_196; setp.lt.s32 %p239, %r987, 1; @%p239 bra $L__BB207_195; mov.u16 %rs53, 1; fma.rn.ftz.f32 %f2232, %f532, %f2429, %f2392; fma.rn.ftz.f32 %f2233, %f532, %f2428, %f2391; fma.rn.ftz.f32 %f2234, %f532, %f2427, %f2390; sub.ftz.f32 %f2235, %f2232, %f14; sub.ftz.f32 %f2236, %f2233, %f15; sub.ftz.f32 %f2237, %f2234, %f16; neg.ftz.f32 %f2238, %f17; mul.ftz.f32 %f2239, %f2235, %f2238; mul.ftz.f32 %f2240, %f2236, %f18; sub.ftz.f32 %f2241, %f2239, %f2240; mul.ftz.f32 %f2242, %f2237, %f20; sub.ftz.f32 %f2561, %f2241, %f2242; $L__BB207_195: setp.lt.ftz.f32 %p283, %f532, %f206; $L__BB207_196: selp.f32 %f536, %f532, %f206, %p283; fma.rn.ftz.f32 %f2392, %f2429, %f536, %f2392; fma.rn.ftz.f32 %f2391, %f536, %f2428, %f2391; fma.rn.ftz.f32 %f2390, %f536, %f2427, %f2390; mul.ftz.f32 %f2246, %f945, %f536; mul.ftz.f32 %f2250, %f948, %f536; mul.ftz.f32 %f2254, %f951, %f536; mul.ftz.f32 %f2255, %f2246, 0fBFB8AA3B; ex2.approx.ftz.f32 %f540, %f2255; mul.ftz.f32 %f2256, %f2250, 0fBFB8AA3B; ex2.approx.ftz.f32 %f541, %f2256; mul.ftz.f32 %f2257, %f2254, 0fBFB8AA3B; ex2.approx.ftz.f32 %f542, %f2257; @%p283 bra $L__BB207_198; bra.uni $L__BB207_197; $L__BB207_198: mov.f32 %f2576, %f540; mov.f32 %f2575, %f541; mov.f32 %f2574, %f542; bra.uni $L__BB207_199; $L__BB207_197: mul.ftz.f32 %f2574, %f951, %f542; mul.ftz.f32 %f2575, %f948, %f541; mul.ftz.f32 %f2576, %f945, %f540; $L__BB207_199: setp.eq.s32 %p274, %r958, 0; @%p274 bra $L__BB207_208; mul.ftz.f32 %f555, %f945, %f2433; mul.ftz.f32 %f556, %f948, %f2433; mul.ftz.f32 %f557, %f951, %f2433; mul.ftz.f32 %f2276, %f555, %f536; mul.ftz.f32 %f2277, %f556, %f536; mul.ftz.f32 %f2278, %f557, %f536; mul.ftz.f32 %f2279, %f2276, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2565, %f2279; mul.ftz.f32 %f2280, %f2277, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2566, %f2280; mul.ftz.f32 %f2281, %f2278, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2567, %f2281; @%p283 bra $L__BB207_202; mul.ftz.f32 %f2567, %f557, %f2567; mul.ftz.f32 %f2566, %f556, %f2566; mul.ftz.f32 %f2565, %f555, %f2565; $L__BB207_202: and.b16 %rs32, %rs53, 255; setp.eq.s16 %p244, %rs32, 0; @%p244 bra $L__BB207_206; mul.ftz.f32 %f573, %f945, %f2435; mul.ftz.f32 %f574, %f948, %f2435; mul.ftz.f32 %f575, %f951, %f2435; mul.ftz.f32 %f2291, %f573, %f536; mul.ftz.f32 %f2292, %f574, %f536; mul.ftz.f32 %f2293, %f575, %f536; mul.ftz.f32 %f2294, %f2291, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2568, %f2294; mul.ftz.f32 %f2295, %f2292, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2569, %f2295; mul.ftz.f32 %f2296, %f2293, 0fBFB8AA3B; ex2.approx.ftz.f32 %f2570, %f2296; @%p283 bra $L__BB207_205; mul.ftz.f32 %f2570, %f575, %f2570; mul.ftz.f32 %f2569, %f574, %f2569; mul.ftz.f32 %f2568, %f573, %f2568; $L__BB207_205: mul.ftz.f32 %f2297, %f2434, %f2568; mul.ftz.f32 %f2298, %f2434, %f2569; mul.ftz.f32 %f2299, %f2434, %f2570; mul.ftz.f32 %f2300, %f2432, %f2565; sub.ftz.f32 %f2301, %f2297, %f2300; mul.ftz.f32 %f2302, %f2432, %f2566; sub.ftz.f32 %f2303, %f2298, %f2302; mul.ftz.f32 %f2304, %f2432, %f2567; sub.ftz.f32 %f2305, %f2299, %f2304; fma.rn.ftz.f32 %f2573, %f2431, %f2301, %f2300; fma.rn.ftz.f32 %f2572, %f2431, %f2303, %f2302; fma.rn.ftz.f32 %f2571, %f2431, %f2305, %f2304; bra.uni $L__BB207_207; $L__BB207_206: mul.ftz.f32 %f2573, %f2432, %f2565; mul.ftz.f32 %f2572, %f2432, %f2566; mul.ftz.f32 %f2571, %f2432, %f2567; $L__BB207_207: selp.f32 %f2306, %f95, 0f3F400000, %p48; sub.ftz.f32 %f2307, %f2573, %f2576; sub.ftz.f32 %f2308, %f2572, %f2575; sub.ftz.f32 %f2309, %f2571, %f2574; fma.rn.ftz.f32 %f2576, %f2306, %f2307, %f2576; fma.rn.ftz.f32 %f2575, %f2306, %f2308, %f2575; fma.rn.ftz.f32 %f2574, %f2306, %f2309, %f2574; $L__BB207_208: selp.f32 %f2310, %f90, %f96, %p48; mul.ftz.f32 %f2311, %f2310, %f540; selp.f32 %f2312, %f91, %f97, %p48; mul.ftz.f32 %f2313, %f2312, %f541; selp.f32 %f2314, %f92, %f98, %p48; mul.ftz.f32 %f2315, %f2314, %f542; selp.f32 %f2316, %f542, %f2315, %p283; selp.f32 %f2317, %f541, %f2313, %p283; selp.f32 %f2318, %f540, %f2311, %p283; mul.ftz.f32 %f2319, %f2575, %f2408; fma.rn.ftz.f32 %f2320, %f2576, %f2407, %f2319; fma.rn.ftz.f32 %f2321, %f2574, %f2409, %f2320; div.approx.ftz.f32 %f2322, %f2318, %f2321; div.approx.ftz.f32 %f2323, %f2317, %f2321; div.approx.ftz.f32 %f2324, %f2316, %f2321; mul.ftz.f32 %f2380, %f2322, %f2380; mul.ftz.f32 %f2382, %f2323, %f2382; mul.ftz.f32 %f2384, %f2324, %f2384; @%p283 bra $L__BB207_210; max.ftz.f32 %f2325, %f2380, %f2382; max.ftz.f32 %f2326, %f2325, %f2384; setp.geu.ftz.f32 %p248, %f2326, 0f358637BD; add.s32 %r958, %r958, 1; setp.lt.u32 %p249, %r958, 256; and.pred %p250, %p248, %p249; add.s32 %r959, %r959, 16; mov.u32 %r956, %r964; mov.u32 %r957, %r965; @%p250 bra $L__BB207_38; $L__BB207_210: mov.pred %p284, -1; add.u64 %rd445, %SPL, 768; st.local.u32 [%rd445+36], %r965; st.local.f32 [%rd445], %f2392; st.local.f32 [%rd445+4], %f2391; st.local.f32 [%rd445+8], %f2390; st.local.f32 [%rd445+12], %f2429; st.local.f32 [%rd445+16], %f2428; st.local.f32 [%rd445+20], %f2427; st.local.f32 [%rd445+28], %f532; st.local.u32 [%rd445+40], %r964; not.pred %p252, %p283; @%p252 bra $L__BB207_227; ld.const.u64 %rd451, [kernel_params+2584]; cvt.s64.s32 %rd450, %r951; mul.lo.s64 %rd449, %rd450, 12; cvta.to.global.u64 %rd448, %rd451; add.s64 %rd447, %rd448, %rd449; st.global.f32 [%rd447], %f2380; st.global.f32 [%rd447+4], %f2382; st.global.f32 [%rd447+8], %f2384; bra.uni $L__BB207_213; $L__BB207_217: ld.const.u64 %rd370, [kernel_params+2216]; cvta.to.global.u64 %rd371, %rd370; shl.b64 %rd372, %rd90, 2; add.s64 %rd373, %rd371, %rd372; ld.global.u32 %r1015, [%rd373]; bra.uni $L__BB207_222; $L__BB207_219: cvt.u32.u64 %r918, %rd90; ld.const.u64 %rd374, [kernel_params+2304]; cvta.to.global.u64 %rd375, %rd374; mul.wide.s32 %rd376, %r918, 4; add.s64 %rd377, %rd375, %rd376; ld.global.u32 %r1015, [%rd377]; $L__BB207_222: and.b32 %r177, %r1015, 4194303; ld.const.u64 %rd382, [kernel_params+2184]; cvta.to.global.u64 %rd383, %rd382; mul.wide.s32 %rd384, %r170, 4; add.s64 %rd385, %rd383, %rd384; ld.const.u32 %r921, [kernel_params+1592]; setp.ne.s32 %p263, %r921, 0; ld.global.u8 %rs38, [%rd385+1]; and.b16 %rs39, %rs38, 6; setp.ne.s16 %p264, %rs39, 0; and.pred %p265, %p264, %p263; @%p265 bra $L__BB207_226; bra.uni $L__BB207_223; $L__BB207_226: cvt.u32.u64 %r933, %rd1; ld.const.u32 %r934, [kernel_params+4392]; div.u32 %r935, %r933, %r934; ld.const.u32 %r936, [kernel_params+8]; mad.lo.s32 %r937, %r935, %r936, %r177; ld.const.u64 %rd418, [kernel_params+4256]; cvta.to.global.u64 %rd419, %rd418; add.s64 %rd420, %rd419, 16; atom.global.add.u32 %r938, [%rd420], -1; add.s64 %rd421, %rd419, 40; atom.global.add.u32 %r939, [%rd421], 1; ld.const.u64 %rd422, [kernel_params+2520]; cvta.to.global.u64 %rd423, %rd422; shl.b64 %rd424, %rd1, 1; add.s64 %rd425, %rd423, %rd424; mov.u16 %rs42, 10; st.global.u16 [%rd425], %rs42; ld.const.u64 %rd426, [kernel_params+2624]; cvta.to.global.u64 %rd427, %rd426; add.s64 %rd429, %rd427, %rd322; st.global.u32 [%rd429], %r937; ld.const.u64 %rd430, [kernel_params+4344]; cvta.to.global.u64 %rd431, %rd430; mul.wide.s32 %rd432, %r937, 4; add.s64 %rd433, %rd431, %rd432; atom.global.add.u32 %r940, [%rd433], 1; bra.uni $L__BB207_227; $L__BB207_223: ld.const.u64 %rd386, [kernel_params+2408]; cvt.u32.u64 %r922, %rd1; cvta.to.global.u64 %rd387, %rd386; mul.wide.u32 %rd388, %r177, 32; add.s64 %rd389, %rd387, %rd388; ld.global.u32 %r923, [%rd389+16]; and.b32 %r924, %r923, 1073741824; setp.eq.s32 %p266, %r924, 0; ld.const.u32 %r925, [kernel_params+4392]; div.u32 %r926, %r922, %r925; ld.const.u32 %r927, [kernel_params+8]; mad.lo.s32 %r178, %r926, %r927, %r177; ld.const.u64 %rd390, [kernel_params+4256]; cvta.to.global.u64 %rd91, %rd390; add.s64 %rd391, %rd91, 16; atom.global.add.u32 %r928, [%rd391], -1; cvt.s64.s32 %rd92, %r178; @%p266 bra $L__BB207_225; bra.uni $L__BB207_224; $L__BB207_225: add.s64 %rd405, %rd91, 32; atom.global.add.u32 %r931, [%rd405], 1; ld.const.u64 %rd406, [kernel_params+2520]; cvta.to.global.u64 %rd407, %rd406; shl.b64 %rd408, %rd1, 1; add.s64 %rd409, %rd407, %rd408; mov.u16 %rs41, 8; st.global.u16 [%rd409], %rs41; ld.const.u64 %rd410, [kernel_params+2624]; cvta.to.global.u64 %rd411, %rd410; add.s64 %rd413, %rd411, %rd322; st.global.u32 [%rd413], %r178; ld.const.u64 %rd414, [kernel_params+4328]; cvta.to.global.u64 %rd415, %rd414; shl.b64 %rd416, %rd92, 2; add.s64 %rd417, %rd415, %rd416; atom.global.add.u32 %r932, [%rd417], 1; bra.uni $L__BB207_227; $L__BB207_224: add.s64 %rd392, %rd91, 36; atom.global.add.u32 %r929, [%rd392], 1; ld.const.u64 %rd393, [kernel_params+2520]; cvta.to.global.u64 %rd394, %rd393; shl.b64 %rd395, %rd1, 1; add.s64 %rd396, %rd394, %rd395; mov.u16 %rs40, 9; st.global.u16 [%rd396], %rs40; ld.const.u64 %rd397, [kernel_params+2624]; cvta.to.global.u64 %rd398, %rd397; add.s64 %rd400, %rd398, %rd322; st.global.u32 [%rd400], %r178; ld.const.u64 %rd401, [kernel_params+4336]; cvta.to.global.u64 %rd402, %rd401; shl.b64 %rd403, %rd92, 2; add.s64 %rd404, %rd402, %rd403; atom.global.add.u32 %r930, [%rd404], 1; $L__BB207_227: not.pred %p270, %p284; @%p270 bra $L__BB207_229; ld.const.u64 %rd434, [kernel_params+4256]; cvta.to.global.u64 %rd435, %rd434; add.s64 %rd436, %rd435, 16; atom.global.add.u32 %r941, [%rd436], -1; ld.const.u64 %rd437, [kernel_params+2520]; cvta.to.global.u64 %rd438, %rd437; shl.b64 %rd439, %rd1, 1; add.s64 %rd440, %rd438, %rd439; mov.u16 %rs43, 0; st.global.u16 [%rd440], %rs43; $L__BB207_229: ret; } // .globl kernel_gpu_integrator_intersect_volume_stack .visible .entry kernel_gpu_integrator_intersect_volume_stack( .param .u64 kernel_gpu_integrator_intersect_volume_stack_param_0, .param .u32 kernel_gpu_integrator_intersect_volume_stack_param_1 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot208[1200]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<287>; .reg .b16 %rs<49>; .reg .f32 %f<3573>; .reg .b32 %r<649>; .reg .b64 %rd<880>; mov.u64 %SPL, __local_depot208; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd132, [kernel_gpu_integrator_intersect_volume_stack_param_0]; ld.param.u32 %r259, [kernel_gpu_integrator_intersect_volume_stack_param_1]; add.u64 %rd1, %SPL, 0; mov.u32 %r260, %ntid.x; mov.u32 %r261, %ctaid.x; mov.u32 %r262, %tid.x; mad.lo.s32 %r533, %r261, %r260, %r262; setp.ge.s32 %p10, %r533, %r259; @%p10 bra $L__BB208_245; setp.eq.s64 %p11, %rd132, 0; @%p11 bra $L__BB208_3; cvta.to.global.u64 %rd134, %rd132; mul.wide.s32 %rd135, %r533, 4; add.s64 %rd136, %rd134, %rd135; ld.global.u32 %r533, [%rd136]; $L__BB208_3: ld.const.u64 %rd137, [kernel_params+2632]; cvta.to.global.u64 %rd138, %rd137; cvt.s64.s32 %rd2, %r533; mul.wide.s32 %rd139, %r533, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f1, [%rd140]; ld.global.f32 %f2, [%rd140+4]; ld.global.f32 %f3, [%rd140+8]; ld.const.u64 %rd141, [kernel_params+2664]; cvta.to.global.u64 %rd142, %rd141; mul.wide.s32 %rd143, %r533, 4; add.s64 %rd144, %rd142, %rd143; ld.global.f32 %f4, [%rd144]; ld.const.v2.u64 {%rd145, %rd146}, [kernel_params+2672]; mov.u32 %r647, 0; cvta.to.global.u64 %rd149, %rd145; add.s64 %rd150, %rd149, %rd143; ld.global.f32 %f5, [%rd150]; cvta.to.global.u64 %rd151, %rd146; add.s64 %rd152, %rd151, %rd143; ld.global.f32 %f6, [%rd152]; ld.const.u64 %rd153, [kernel_params+2544]; cvta.to.global.u64 %rd154, %rd153; add.s64 %rd3, %rd154, %rd143; ld.global.u32 %r264, [%rd3]; and.b32 %r265, %r264, 1073741824; shr.u32 %r266, %r264, 30; and.b32 %r267, %r266, 1; setp.eq.b32 %p12, %r267, 1; setp.eq.s32 %p13, %r265, 0; selp.b32 %r4, 1, 65536, %p13; ld.const.u32 %r5, [kernel_params+1028]; setp.eq.s32 %p14, %r5, -1; or.pred %p15, %p14, %p12; @%p15 bra $L__BB208_5; ld.const.v2.u64 {%rd155, %rd156}, [kernel_params+2768]; mov.u32 %r647, 1; cvta.to.global.u64 %rd159, %rd155; shl.b64 %rd160, %rd2, 2; add.s64 %rd161, %rd159, %rd160; mov.u32 %r269, -1; st.global.u32 [%rd161], %r269; cvta.to.global.u64 %rd162, %rd156; add.s64 %rd163, %rd162, %rd160; st.global.u32 [%rd163], %r5; $L__BB208_5: ld.const.u32 %r7, [kernel_params+12]; add.s32 %r270, %r7, -1; setp.ge.u32 %p16, %r647, %r270; shl.b32 %r271, %r7, 1; setp.eq.s32 %p17, %r271, 0; or.pred %p18, %p17, %p16; @%p18 bra $L__BB208_232; mov.b32 %r277, %f1; or.b32 %r278, %r277, -2147483648; setp.eq.s32 %p19, %r278, -2147483648; add.ftz.f32 %f908, %f1, %f1; setp.neu.ftz.f32 %p20, %f908, %f1; or.pred %p21, %p20, %p19; shl.b32 %r279, %r277, 1; setp.lt.u32 %p22, %r279, -16777215; ld.const.v2.u32 {%r280, %r281}, [kernel_params+1104]; ld.const.v2.u64 {%rd164, %rd165}, [kernel_params+2176]; cvta.to.global.u64 %rd6, %rd165; ld.const.v2.u64 {%rd166, %rd167}, [kernel_params+2096]; mov.u32 %r536, 0; cvta.to.global.u64 %rd9, %rd166; ld.const.v2.u64 {%rd168, %rd169}, [kernel_params+2128]; cvta.to.global.u64 %rd10, %rd169; cvta.to.global.u64 %rd11, %rd168; ld.const.v2.u64 {%rd172, %rd173}, [kernel_params+2224]; cvta.to.global.u64 %rd14, %rd173; ld.const.v2.u64 {%rd174, %rd175}, [kernel_params+2256]; cvta.to.global.u64 %rd17, %rd174; ld.const.u64 %rd176, [kernel_params+2120]; cvta.to.global.u64 %rd18, %rd176; ld.const.u64 %rd177, [kernel_params+2160]; cvta.to.global.u64 %rd19, %rd177; ld.const.u64 %rd178, [kernel_params+2144]; cvta.to.global.u64 %rd20, %rd178; ld.const.v2.u64 {%rd179, %rd180}, [kernel_params+2304]; cvta.to.global.u64 %rd23, %rd180; ld.const.v2.u64 {%rd181, %rd182}, [kernel_params+2336]; cvta.to.global.u64 %rd26, %rd181; ld.const.u64 %rd183, [kernel_params+2408]; cvta.to.global.u64 %rd27, %rd183; ld.const.u64 %rd184, [kernel_params+2296]; cvta.to.global.u64 %rd28, %rd184; ld.const.u64 %rd185, [kernel_params+2272]; cvta.to.global.u64 %rd29, %rd185; ld.const.u64 %rd186, [kernel_params+2216]; cvta.to.global.u64 %rd30, %rd186; and.pred %p23, %p21, %p22; setp.num.ftz.f32 %p24, %f1, %f1; mov.f32 %f3217, 0f00000000; and.pred %p1, %p24, %p23; mov.u32 %r630, -1; not.pred %p26, %p1; setp.eq.s32 %p27, %r281, 0; mov.pred %p154, -1; cvta.to.global.u64 %rd519, %rd172; cvta.to.global.u64 %rd445, %rd179; mov.u32 %r537, %r536; mov.u32 %r629, %r630; $L__BB208_7: mov.pred %p285, 0; @%p26 bra $L__BB208_117; mov.f32 %f3335, 0f00000000; abs.ftz.f32 %f910, %f3335; setp.gt.ftz.f32 %p28, %f910, 0f17800000; selp.f32 %f20, 0f00000000, 0f17800000, %p28; mov.f32 %f911, 0f3F800000; abs.ftz.f32 %f912, %f911; setp.gt.ftz.f32 %p29, %f912, 0f17800000; selp.f32 %f21, 0f3F800000, 0f17800000, %p29; rcp.approx.ftz.f32 %f22, %f20; rcp.approx.ftz.f32 %f23, %f21; @%p27 bra $L__BB208_81; mov.u32 %r286, 1985229328; st.local.u32 [%rd1], %r286; mov.u32 %r584, -1; mov.u32 %r586, 0; mov.f32 %f3337, 0f7F7FFFFF; mov.f32 %f3326, %f3; mov.f32 %f3327, %f2; mov.f32 %f3328, %f1; mov.f32 %f3329, %f21; mov.f32 %f3330, %f20; mov.f32 %f3331, %f20; mov.f32 %f3332, %f23; mov.f32 %f3333, %f22; mov.f32 %f3334, %f22; mov.u32 %r585, %r584; mov.f32 %f3336, %f3335; mov.u32 %r559, %r280; mov.u32 %r588, %r584; bra.uni $L__BB208_10; $L__BB208_22: cvta.to.global.u64 %rd190, %rd167; mul.wide.s32 %rd191, %r559, 16; xor.b64 %rd192, %rd191, -16; add.s64 %rd193, %rd190, %rd192; ld.global.v4.u32 {%r562, %r298, %r299, %r300}, [%rd193]; setp.gt.s32 %p53, %r562, -1; mul.wide.s32 %rd194, %r50, 4; add.s64 %rd34, %rd1, %rd194; @%p53 bra $L__BB208_34; bra.uni $L__BB208_23; $L__BB208_34: ld.local.u32 %r559, [%rd34]; cvt.u16.u32 %rs18, %r300; and.b16 %rs17, %rs18, 127; add.s32 %r586, %r50, -1; setp.eq.s16 %p64, %rs17, 1; @%p64 bra $L__BB208_62; setp.ne.s16 %p65, %rs17, 65; @%p65 bra $L__BB208_76; add.s32 %r586, %r50, -1; setp.le.s32 %p66, %r298, %r562; @%p66 bra $L__BB208_76; $L__BB208_37: cvt.s64.s32 %rd38, %r562; setp.ne.s32 %p67, %r588, -1; mov.u32 %r566, %r588; @%p67 bra $L__BB208_39; shl.b64 %rd209, %rd38, 2; add.s64 %rd210, %rd10, %rd209; ld.global.u32 %r566, [%rd210]; $L__BB208_39: shl.b64 %rd211, %rd38, 2; add.s64 %rd212, %rd11, %rd211; ld.global.u32 %r319, [%rd212]; cvt.s64.s32 %rd39, %r319; setp.eq.s32 %p68, %r630, %r319; setp.eq.s32 %p69, %r629, %r566; and.pred %p70, %p69, %p68; @%p70 bra $L__BB208_61; cvt.s64.s32 %rd40, %r566; mul.wide.s32 %rd213, %r566, 4; add.s64 %rd214, %rd6, %rd213; ld.global.u8 %rs19, [%rd214]; and.b16 %rs20, %rs19, 16; setp.eq.s16 %p71, %rs20, 0; @%p71 bra $L__BB208_61; mul.lo.s64 %rd215, %rd40, 208; add.s64 %rd216, %rd19, %rd215; ld.global.u32 %r72, [%rd216+156]; ld.global.u32 %r73, [%rd216+152]; shl.b32 %r320, %r73, 1; cvt.rn.f32.s32 %f1398, %r320; mul.ftz.f32 %f174, %f4, %f1398; cvt.rzi.ftz.s32.f32 %r321, %f174; add.s32 %r322, %r320, -1; min.s32 %r74, %r321, %r322; ld.global.u32 %r568, [%rd216+164]; mul.wide.u32 %rd217, %r568, 16; add.s64 %rd218, %rd23, %rd217; ld.global.u16 %rs40, [%rd218+12]; ld.global.u32 %r567, [%rd218+8]; ld.global.u64 %rd857, [%rd218]; setp.eq.s64 %p72, %rd857, 11; @%p72 bra $L__BB208_46; $L__BB208_42: setp.eq.s64 %p73, %rd857, 0; @%p73 bra $L__BB208_44; add.s32 %r568, %r568, 2; bra.uni $L__BB208_45; $L__BB208_44: setp.eq.s16 %p74, %rs40, 0; mov.u32 %r571, -1; mov.u32 %r568, %r567; @%p74 bra $L__BB208_47; $L__BB208_45: mul.wide.u32 %rd219, %r568, 16; add.s64 %rd220, %rd23, %rd219; ld.global.u16 %rs40, [%rd220+12]; ld.global.u32 %r567, [%rd220+8]; ld.global.u64 %rd857, [%rd220]; setp.ne.s64 %p75, %rd857, 11; @%p75 bra $L__BB208_42; $L__BB208_46: setp.eq.s16 %p76, %rs40, 0; selp.b32 %r571, -1, %r567, %p76; $L__BB208_47: shl.b64 %rd221, %rd39, 4; add.s64 %rd222, %rd14, %rd221; ld.global.v4.u32 {%r324, %r325, %r326, %r327}, [%rd222]; setp.eq.s32 %p77, %r74, %r73; @%p77 bra $L__BB208_49; bra.uni $L__BB208_48; $L__BB208_49: mul.wide.u32 %rd226, %r327, 12; add.s64 %rd858, %rd17, %rd226; add.s32 %r334, %r327, 1; mul.wide.u32 %rd227, %r334, 12; add.s64 %rd861, %rd17, %rd227; add.s32 %r335, %r327, 2; mul.wide.u32 %rd228, %r335, 12; add.s64 %rd864, %rd17, %rd228; add.s64 %rd859, %rd858, 4; add.s64 %rd860, %rd858, 8; add.s64 %rd862, %rd861, 4; add.s64 %rd863, %rd861, 8; add.s64 %rd865, %rd864, 4; add.s64 %rd866, %rd864, 8; bra.uni $L__BB208_50; $L__BB208_48: setp.gt.s32 %p78, %r74, %r73; selp.b32 %r328, -1, 0, %p78; add.s32 %r329, %r74, %r328; mad.lo.s32 %r330, %r329, %r72, %r571; add.s32 %r331, %r324, %r330; mul.wide.u32 %rd223, %r331, 12; add.s64 %rd858, %rd26, %rd223; add.s32 %r332, %r325, %r330; mul.wide.u32 %rd224, %r332, 12; add.s64 %rd861, %rd26, %rd224; add.s32 %r333, %r326, %r330; mul.wide.u32 %rd225, %r333, 12; add.s64 %rd864, %rd26, %rd225; add.s64 %rd859, %rd858, 4; add.s64 %rd860, %rd858, 8; add.s64 %rd862, %rd861, 4; add.s64 %rd863, %rd861, 8; add.s64 %rd865, %rd864, 4; add.s64 %rd866, %rd864, 8; $L__BB208_50: ld.global.f32 %f175, [%rd866]; ld.global.f32 %f176, [%rd865]; ld.global.f32 %f177, [%rd864]; ld.global.f32 %f178, [%rd863]; ld.global.f32 %f179, [%rd862]; ld.global.f32 %f180, [%rd861]; ld.global.f32 %f181, [%rd860]; ld.global.f32 %f182, [%rd859]; ld.global.f32 %f183, [%rd858]; add.s32 %r89, %r74, 1; setp.eq.s32 %p79, %r89, %r73; @%p79 bra $L__BB208_52; bra.uni $L__BB208_51; $L__BB208_52: mul.wide.u32 %rd232, %r327, 12; add.s64 %rd867, %rd17, %rd232; add.s32 %r342, %r327, 1; mul.wide.u32 %rd233, %r342, 12; add.s64 %rd870, %rd17, %rd233; add.s32 %r343, %r327, 2; mul.wide.u32 %rd234, %r343, 12; add.s64 %rd873, %rd17, %rd234; add.s64 %rd868, %rd867, 4; add.s64 %rd869, %rd867, 8; add.s64 %rd871, %rd870, 4; add.s64 %rd872, %rd870, 8; add.s64 %rd874, %rd873, 4; add.s64 %rd875, %rd873, 8; bra.uni $L__BB208_53; $L__BB208_51: setp.ge.s32 %p80, %r74, %r73; selp.b32 %r336, -1, 0, %p80; add.s32 %r337, %r89, %r336; mad.lo.s32 %r338, %r337, %r72, %r571; add.s32 %r339, %r324, %r338; mul.wide.u32 %rd229, %r339, 12; add.s64 %rd867, %rd26, %rd229; add.s32 %r340, %r325, %r338; mul.wide.u32 %rd230, %r340, 12; add.s64 %rd870, %rd26, %rd230; add.s32 %r341, %r326, %r338; mul.wide.u32 %rd231, %r341, 12; add.s64 %rd873, %rd26, %rd231; add.s64 %rd868, %rd867, 4; add.s64 %rd869, %rd867, 8; add.s64 %rd871, %rd870, 4; add.s64 %rd872, %rd870, 8; add.s64 %rd874, %rd873, 4; add.s64 %rd875, %rd873, 8; $L__BB208_53: mov.f32 %f3207, 0f3F800000; cvt.rn.f32.s32 %f1399, %r74; sub.ftz.f32 %f1400, %f174, %f1399; sub.ftz.f32 %f1402, %f3207, %f1400; ld.global.f32 %f1403, [%rd867]; mul.ftz.f32 %f1404, %f1400, %f1403; ld.global.f32 %f1405, [%rd868]; mul.ftz.f32 %f1406, %f1400, %f1405; ld.global.f32 %f1407, [%rd869]; mul.ftz.f32 %f1408, %f1400, %f1407; fma.rn.ftz.f32 %f1409, %f1402, %f183, %f1404; fma.rn.ftz.f32 %f1410, %f1402, %f182, %f1406; fma.rn.ftz.f32 %f1411, %f1402, %f181, %f1408; ld.global.f32 %f1412, [%rd870]; mul.ftz.f32 %f1413, %f1400, %f1412; ld.global.f32 %f1414, [%rd871]; mul.ftz.f32 %f1415, %f1400, %f1414; ld.global.f32 %f1416, [%rd872]; mul.ftz.f32 %f1417, %f1400, %f1416; fma.rn.ftz.f32 %f1418, %f1402, %f180, %f1413; fma.rn.ftz.f32 %f1419, %f1402, %f179, %f1415; fma.rn.ftz.f32 %f1420, %f1402, %f178, %f1417; ld.global.f32 %f1421, [%rd873]; mul.ftz.f32 %f1422, %f1400, %f1421; ld.global.f32 %f1423, [%rd874]; mul.ftz.f32 %f1424, %f1400, %f1423; ld.global.f32 %f1425, [%rd875]; mul.ftz.f32 %f1426, %f1400, %f1425; fma.rn.ftz.f32 %f1427, %f1402, %f177, %f1422; fma.rn.ftz.f32 %f1428, %f1402, %f176, %f1424; fma.rn.ftz.f32 %f1429, %f1402, %f175, %f1426; sub.ftz.f32 %f184, %f1409, %f3328; sub.ftz.f32 %f185, %f1410, %f3327; sub.ftz.f32 %f186, %f1411, %f3326; sub.ftz.f32 %f1430, %f1418, %f3328; sub.ftz.f32 %f1431, %f1419, %f3327; sub.ftz.f32 %f1432, %f1420, %f3326; sub.ftz.f32 %f1433, %f1427, %f3328; sub.ftz.f32 %f1434, %f1428, %f3327; sub.ftz.f32 %f1435, %f1429, %f3326; sub.ftz.f32 %f187, %f1433, %f184; sub.ftz.f32 %f188, %f1434, %f185; sub.ftz.f32 %f189, %f1435, %f186; sub.ftz.f32 %f190, %f184, %f1430; sub.ftz.f32 %f191, %f185, %f1431; sub.ftz.f32 %f192, %f186, %f1432; sub.ftz.f32 %f1436, %f1430, %f1433; sub.ftz.f32 %f1437, %f1431, %f1434; sub.ftz.f32 %f1438, %f1432, %f1435; add.ftz.f32 %f1439, %f1433, %f184; add.ftz.f32 %f1440, %f1434, %f185; add.ftz.f32 %f1441, %f1435, %f186; mul.ftz.f32 %f1442, %f1441, %f188; mul.ftz.f32 %f1443, %f189, %f1440; sub.ftz.f32 %f1444, %f1442, %f1443; mul.ftz.f32 %f1445, %f189, %f1439; mul.ftz.f32 %f1446, %f1441, %f187; sub.ftz.f32 %f1447, %f1445, %f1446; mul.ftz.f32 %f1448, %f1440, %f187; mul.ftz.f32 %f1449, %f188, %f1439; sub.ftz.f32 %f1450, %f1448, %f1449; mul.ftz.f32 %f1451, %f3330, %f1447; fma.rn.ftz.f32 %f1452, %f3331, %f1444, %f1451; fma.rn.ftz.f32 %f193, %f3329, %f1450, %f1452; add.ftz.f32 %f1453, %f1430, %f184; add.ftz.f32 %f1454, %f1431, %f185; add.ftz.f32 %f1455, %f1432, %f186; mul.ftz.f32 %f1456, %f1455, %f191; mul.ftz.f32 %f1457, %f192, %f1454; sub.ftz.f32 %f1458, %f1456, %f1457; mul.ftz.f32 %f1459, %f192, %f1453; mul.ftz.f32 %f1460, %f1455, %f190; sub.ftz.f32 %f1461, %f1459, %f1460; mul.ftz.f32 %f1462, %f1454, %f190; mul.ftz.f32 %f1463, %f191, %f1453; sub.ftz.f32 %f1464, %f1462, %f1463; mul.ftz.f32 %f1465, %f3330, %f1461; fma.rn.ftz.f32 %f1466, %f3331, %f1458, %f1465; fma.rn.ftz.f32 %f194, %f3329, %f1464, %f1466; add.ftz.f32 %f1467, %f1433, %f1430; add.ftz.f32 %f1468, %f1434, %f1431; add.ftz.f32 %f1469, %f1435, %f1432; mul.ftz.f32 %f1470, %f1469, %f1437; mul.ftz.f32 %f1471, %f1438, %f1468; sub.ftz.f32 %f1472, %f1470, %f1471; mul.ftz.f32 %f1473, %f1438, %f1467; mul.ftz.f32 %f1474, %f1469, %f1436; sub.ftz.f32 %f1475, %f1473, %f1474; mul.ftz.f32 %f1476, %f1468, %f1436; mul.ftz.f32 %f1477, %f1437, %f1467; sub.ftz.f32 %f1478, %f1476, %f1477; mul.ftz.f32 %f1479, %f3330, %f1475; fma.rn.ftz.f32 %f1480, %f3331, %f1472, %f1479; fma.rn.ftz.f32 %f1481, %f3329, %f1478, %f1480; add.ftz.f32 %f1482, %f193, %f194; add.ftz.f32 %f195, %f1481, %f1482; abs.ftz.f32 %f196, %f195; mul.ftz.f32 %f1483, %f196, 0f34000000; min.ftz.f32 %f1484, %f194, %f1481; min.ftz.f32 %f1485, %f193, %f1484; max.ftz.f32 %f1486, %f194, %f1481; max.ftz.f32 %f1487, %f193, %f1486; mul.ftz.f32 %f1488, %f196, 0fB4000000; setp.ltu.ftz.f32 %p82, %f1485, %f1488; setp.gtu.ftz.f32 %p83, %f1487, %f1483; and.pred %p84, %p82, %p83; mov.pred %p282, -1; @%p84 bra $L__BB208_59; mov.pred %p282, -1; mul.ftz.f32 %f1489, %f189, %f191; mul.ftz.f32 %f1490, %f192, %f188; sub.ftz.f32 %f1491, %f1489, %f1490; mul.ftz.f32 %f1492, %f189, %f190; mul.ftz.f32 %f1493, %f192, %f187; sub.ftz.f32 %f1494, %f1493, %f1492; mul.ftz.f32 %f1495, %f191, %f187; mul.ftz.f32 %f1496, %f188, %f190; sub.ftz.f32 %f1497, %f1496, %f1495; add.ftz.f32 %f197, %f1491, %f1491; add.ftz.f32 %f198, %f1494, %f1494; add.ftz.f32 %f199, %f1497, %f1497; mul.ftz.f32 %f1498, %f3330, %f198; fma.rn.ftz.f32 %f1499, %f3331, %f197, %f1498; fma.rn.ftz.f32 %f200, %f3329, %f199, %f1499; setp.eq.ftz.f32 %p86, %f200, 0f00000000; @%p86 bra $L__BB208_59; mov.pred %p282, -1; mul.ftz.f32 %f1500, %f185, %f198; fma.rn.ftz.f32 %f1501, %f184, %f197, %f1500; fma.rn.ftz.f32 %f1502, %f186, %f199, %f1501; div.approx.ftz.f32 %f201, %f1502, %f200; setp.ltu.ftz.f32 %p88, %f201, %f3217; setp.gtu.ftz.f32 %p89, %f201, %f3337; or.pred %p90, %p89, %p88; @%p90 bra $L__BB208_59; abs.ftz.f32 %f3210, %f195; setp.lt.ftz.f32 %p91, %f3210, 0f219392EF; mov.f32 %f3294, 0f00000000; @%p91 bra $L__BB208_58; rcp.approx.ftz.f32 %f3294, %f195; $L__BB208_58: mov.f32 %f3208, 0f3F800000; mul.ftz.f32 %f1504, %f193, %f3294; min.ftz.f32 %f3299, %f1504, %f3208; mul.ftz.f32 %f1506, %f194, %f3294; min.ftz.f32 %f3298, %f1506, %f3208; mov.pred %p282, 0; mov.f32 %f3300, %f201; $L__BB208_59: @%p282 bra $L__BB208_61; cvt.s64.s32 %rd834, %r562; shl.b64 %rd833, %rd834, 2; cvt.u32.u64 %r344, %rd39; add.s64 %rd236, %rd18, %rd833; ld.global.u32 %r345, [%rd236]; and.b32 %r346, %r345, %r4; setp.eq.s32 %p93, %r346, 0; selp.b32 %r583, %r583, 65, %p93; selp.b32 %r584, %r584, %r566, %p93; selp.b32 %r585, %r585, %r344, %p93; selp.f32 %f3335, %f3335, %f3298, %p93; selp.f32 %f3336, %f3336, %f3299, %p93; selp.f32 %f3337, %f3337, %f3300, %p93; $L__BB208_61: add.s32 %r586, %r50, -1; cvt.u32.u64 %r347, %rd38; add.s32 %r562, %r347, 1; setp.lt.s32 %p94, %r562, %r298; @%p94 bra $L__BB208_37; bra.uni $L__BB208_76; $L__BB208_23: mul.wide.s32 %rd195, %r562, 4; xor.b64 %rd196, %rd195, -4; add.s64 %rd197, %rd10, %rd196; ld.global.u32 %r588, [%rd197]; cvt.s64.s32 %rd35, %r588; mul.wide.s32 %rd198, %r588, 4; add.s64 %rd199, %rd6, %rd198; ld.global.u32 %r58, [%rd199]; and.b32 %r301, %r58, 16; setp.eq.s32 %p54, %r301, 0; @%p54 bra $L__BB208_33; bra.uni $L__BB208_24; $L__BB208_33: ld.local.u32 %r559, [%rd34]; add.s32 %r586, %r50, -1; mov.u32 %r588, -1; bra.uni $L__BB208_76; $L__BB208_62: setp.le.s32 %p95, %r298, %r562; @%p95 bra $L__BB208_76; $L__BB208_63: cvt.s64.s32 %rd98, %r562; setp.ne.s32 %p96, %r588, -1; mov.u32 %r579, %r588; @%p96 bra $L__BB208_65; shl.b64 %rd237, %rd98, 2; add.s64 %rd238, %rd10, %rd237; ld.global.u32 %r579, [%rd238]; $L__BB208_65: cvt.s64.s32 %rd832, %r562; shl.b64 %rd239, %rd832, 2; add.s64 %rd240, %rd11, %rd239; ld.global.u32 %r348, [%rd240]; cvt.s64.s32 %rd99, %r348; setp.eq.s32 %p97, %r630, %r348; setp.eq.s32 %p98, %r629, %r579; and.pred %p99, %p98, %p97; @%p99 bra $L__BB208_75; mul.wide.s32 %rd241, %r579, 4; add.s64 %rd242, %rd6, %rd241; ld.global.u8 %rs21, [%rd242]; and.b16 %rs22, %rs21, 16; setp.eq.s16 %p100, %rs22, 0; @%p100 bra $L__BB208_75; shl.b64 %rd243, %rd99, 4; add.s64 %rd244, %rd14, %rd243; ld.global.u32 %r349, [%rd244+12]; mul.wide.u32 %rd245, %r349, 12; add.s64 %rd246, %rd17, %rd245; add.s32 %r350, %r349, 1; mul.wide.u32 %rd247, %r350, 12; add.s64 %rd248, %rd17, %rd247; add.s32 %r351, %r349, 2; mul.wide.u32 %rd249, %r351, 12; add.s64 %rd250, %rd17, %rd249; ld.global.f32 %f1507, [%rd246]; sub.ftz.f32 %f224, %f1507, %f3328; ld.global.f32 %f1508, [%rd246+4]; sub.ftz.f32 %f225, %f1508, %f3327; ld.global.f32 %f1509, [%rd246+8]; sub.ftz.f32 %f226, %f1509, %f3326; ld.global.f32 %f1510, [%rd248]; sub.ftz.f32 %f1511, %f1510, %f3328; ld.global.f32 %f1512, [%rd248+4]; sub.ftz.f32 %f1513, %f1512, %f3327; ld.global.f32 %f1514, [%rd248+8]; sub.ftz.f32 %f1515, %f1514, %f3326; ld.global.f32 %f1516, [%rd250]; sub.ftz.f32 %f1517, %f1516, %f3328; ld.global.f32 %f1518, [%rd250+4]; sub.ftz.f32 %f1519, %f1518, %f3327; ld.global.f32 %f1520, [%rd250+8]; sub.ftz.f32 %f1521, %f1520, %f3326; sub.ftz.f32 %f227, %f1517, %f224; sub.ftz.f32 %f228, %f1519, %f225; sub.ftz.f32 %f229, %f1521, %f226; sub.ftz.f32 %f230, %f224, %f1511; sub.ftz.f32 %f231, %f225, %f1513; sub.ftz.f32 %f232, %f226, %f1515; sub.ftz.f32 %f1522, %f1511, %f1517; sub.ftz.f32 %f1523, %f1513, %f1519; sub.ftz.f32 %f1524, %f1515, %f1521; add.ftz.f32 %f1525, %f224, %f1517; add.ftz.f32 %f1526, %f225, %f1519; add.ftz.f32 %f1527, %f226, %f1521; mul.ftz.f32 %f1528, %f228, %f1527; mul.ftz.f32 %f1529, %f1526, %f229; sub.ftz.f32 %f1530, %f1528, %f1529; mul.ftz.f32 %f1531, %f1525, %f229; mul.ftz.f32 %f1532, %f227, %f1527; sub.ftz.f32 %f1533, %f1531, %f1532; mul.ftz.f32 %f1534, %f227, %f1526; mul.ftz.f32 %f1535, %f1525, %f228; sub.ftz.f32 %f1536, %f1534, %f1535; mul.ftz.f32 %f1537, %f3330, %f1533; fma.rn.ftz.f32 %f1538, %f3331, %f1530, %f1537; fma.rn.ftz.f32 %f233, %f3329, %f1536, %f1538; add.ftz.f32 %f1539, %f224, %f1511; add.ftz.f32 %f1540, %f225, %f1513; add.ftz.f32 %f1541, %f226, %f1515; mul.ftz.f32 %f1542, %f231, %f1541; mul.ftz.f32 %f1543, %f1540, %f232; sub.ftz.f32 %f1544, %f1542, %f1543; mul.ftz.f32 %f1545, %f1539, %f232; mul.ftz.f32 %f1546, %f230, %f1541; sub.ftz.f32 %f1547, %f1545, %f1546; mul.ftz.f32 %f1548, %f230, %f1540; mul.ftz.f32 %f1549, %f1539, %f231; sub.ftz.f32 %f1550, %f1548, %f1549; mul.ftz.f32 %f1551, %f3330, %f1547; fma.rn.ftz.f32 %f1552, %f3331, %f1544, %f1551; fma.rn.ftz.f32 %f234, %f3329, %f1550, %f1552; add.ftz.f32 %f1553, %f1511, %f1517; add.ftz.f32 %f1554, %f1513, %f1519; add.ftz.f32 %f1555, %f1515, %f1521; mul.ftz.f32 %f1556, %f1523, %f1555; mul.ftz.f32 %f1557, %f1554, %f1524; sub.ftz.f32 %f1558, %f1556, %f1557; mul.ftz.f32 %f1559, %f1553, %f1524; mul.ftz.f32 %f1560, %f1522, %f1555; sub.ftz.f32 %f1561, %f1559, %f1560; mul.ftz.f32 %f1562, %f1522, %f1554; mul.ftz.f32 %f1563, %f1553, %f1523; sub.ftz.f32 %f1564, %f1562, %f1563; mul.ftz.f32 %f1565, %f3330, %f1561; fma.rn.ftz.f32 %f1566, %f3331, %f1558, %f1565; fma.rn.ftz.f32 %f1567, %f3329, %f1564, %f1566; add.ftz.f32 %f1568, %f234, %f233; add.ftz.f32 %f235, %f1567, %f1568; abs.ftz.f32 %f236, %f235; mul.ftz.f32 %f1569, %f236, 0f34000000; min.ftz.f32 %f1570, %f234, %f1567; min.ftz.f32 %f1571, %f233, %f1570; max.ftz.f32 %f1572, %f234, %f1567; max.ftz.f32 %f1573, %f233, %f1572; mul.ftz.f32 %f1574, %f236, 0fB4000000; setp.ltu.ftz.f32 %p102, %f1571, %f1574; setp.gtu.ftz.f32 %p103, %f1573, %f1569; and.pred %p104, %p102, %p103; mov.pred %p283, -1; @%p104 bra $L__BB208_73; mov.pred %p283, -1; mul.ftz.f32 %f1575, %f231, %f229; mul.ftz.f32 %f1576, %f232, %f228; sub.ftz.f32 %f1577, %f1575, %f1576; mul.ftz.f32 %f1578, %f230, %f229; mul.ftz.f32 %f1579, %f232, %f227; sub.ftz.f32 %f1580, %f1579, %f1578; mul.ftz.f32 %f1581, %f231, %f227; mul.ftz.f32 %f1582, %f230, %f228; sub.ftz.f32 %f1583, %f1582, %f1581; add.ftz.f32 %f237, %f1577, %f1577; add.ftz.f32 %f238, %f1580, %f1580; add.ftz.f32 %f239, %f1583, %f1583; mul.ftz.f32 %f1584, %f3330, %f238; fma.rn.ftz.f32 %f1585, %f3331, %f237, %f1584; fma.rn.ftz.f32 %f240, %f3329, %f239, %f1585; setp.eq.ftz.f32 %p106, %f240, 0f00000000; @%p106 bra $L__BB208_73; mov.pred %p283, -1; mul.ftz.f32 %f1586, %f225, %f238; fma.rn.ftz.f32 %f1587, %f224, %f237, %f1586; fma.rn.ftz.f32 %f1588, %f226, %f239, %f1587; div.approx.ftz.f32 %f241, %f1588, %f240; setp.ltu.ftz.f32 %p108, %f241, %f3217; setp.gtu.ftz.f32 %p109, %f241, %f3337; or.pred %p110, %p109, %p108; @%p110 bra $L__BB208_73; abs.ftz.f32 %f3202, %f235; setp.lt.ftz.f32 %p111, %f3202, 0f219392EF; mov.f32 %f3310, 0f00000000; @%p111 bra $L__BB208_72; rcp.approx.ftz.f32 %f3310, %f235; $L__BB208_72: mov.f32 %f3203, 0f3F800000; mul.ftz.f32 %f1590, %f233, %f3310; min.ftz.f32 %f3315, %f1590, %f3203; mul.ftz.f32 %f1592, %f234, %f3310; min.ftz.f32 %f3314, %f1592, %f3203; mov.pred %p283, 0; mov.f32 %f3316, %f241; $L__BB208_73: @%p283 bra $L__BB208_75; cvt.s64.s32 %rd827, %r562; shl.b64 %rd826, %rd827, 2; cvt.u32.u64 %r352, %rd99; add.s64 %rd252, %rd18, %rd826; ld.global.u32 %r353, [%rd252]; and.b32 %r354, %r353, %r4; setp.eq.s32 %p113, %r354, 0; selp.b32 %r583, %r583, 1, %p113; selp.b32 %r584, %r584, %r579, %p113; selp.b32 %r585, %r585, %r352, %p113; selp.f32 %f3335, %f3335, %f3314, %p113; selp.f32 %f3336, %f3336, %f3315, %p113; selp.f32 %f3337, %f3337, %f3316, %p113; $L__BB208_75: add.s32 %r586, %r50, -1; cvt.u32.u64 %r355, %rd98; add.s32 %r562, %r355, 1; setp.lt.s32 %p114, %r562, %r298; @%p114 bra $L__BB208_63; bra.uni $L__BB208_76; $L__BB208_24: and.b32 %r302, %r58, 2; setp.eq.s32 %p55, %r302, 0; mul.lo.s64 %rd200, %rd35, 208; add.s64 %rd201, %rd19, %rd200; add.s64 %rd36, %rd201, 48; @%p55 bra $L__BB208_31; bra.uni $L__BB208_25; $L__BB208_31: ld.global.v4.f32 {%f3276, %f3277, %f3278, %f3279}, [%rd36]; ld.global.v4.f32 {%f3280, %f3281, %f3282, %f3283}, [%rd36+16]; ld.global.v4.f32 {%f3284, %f3285, %f3286, %f3287}, [%rd36+32]; bra.uni $L__BB208_32; $L__BB208_25: mov.f32 %f3204, 0f3F800000; ld.global.u32 %rd202, [%rd36+120]; ld.global.u32 %r303, [%rd36+104]; shl.b32 %r304, %r303, 1; cvt.rn.f32.s32 %f1129, %r304; mul.ftz.f32 %f1130, %f4, %f1129; cvt.rzi.ftz.s32.f32 %r305, %f1130; add.s32 %r306, %r304, -1; min.s32 %r307, %r305, %r306; cvt.rn.f32.s32 %f1131, %r307; sub.ftz.f32 %f67, %f1130, %f1131; cvt.s64.s32 %rd203, %r307; add.s64 %rd204, %rd203, %rd202; cvta.to.global.u64 %rd205, %rd164; shl.b64 %rd206, %rd204, 6; add.s64 %rd37, %rd205, %rd206; ld.global.v4.f32 {%f1132, %f1133, %f1134, %f1135}, [%rd37]; ld.global.v4.f32 {%f1136, %f1137, %f1138, %f1139}, [%rd37+64]; mul.ftz.f32 %f1140, %f1133, %f1137; fma.rn.ftz.f32 %f1141, %f1132, %f1136, %f1140; mul.ftz.f32 %f1142, %f1135, %f1139; fma.rn.ftz.f32 %f1143, %f1134, %f1138, %f1142; add.ftz.f32 %f76, %f1141, %f1143; setp.gt.ftz.f32 %p56, %f76, 0f3F7FDF3B; sub.ftz.f32 %f77, %f3204, %f67; @%p56 bra $L__BB208_27; bra.uni $L__BB208_26; $L__BB208_27: mul.ftz.f32 %f1198, %f1132, %f77; fma.rn.ftz.f32 %f1199, %f1136, %f67, %f1198; mul.ftz.f32 %f1200, %f1133, %f77; fma.rn.ftz.f32 %f1201, %f1137, %f67, %f1200; mul.ftz.f32 %f1202, %f1134, %f77; fma.rn.ftz.f32 %f1203, %f1138, %f67, %f1202; mul.ftz.f32 %f1204, %f1135, %f77; fma.rn.ftz.f32 %f1205, %f1139, %f67, %f1204; mul.ftz.f32 %f1206, %f1201, %f1201; fma.rn.ftz.f32 %f1207, %f1199, %f1199, %f1206; mul.ftz.f32 %f1208, %f1205, %f1205; fma.rn.ftz.f32 %f1209, %f1203, %f1203, %f1208; add.ftz.f32 %f1210, %f1207, %f1209; rsqrt.approx.ftz.f32 %f1211, %f1210; mul.ftz.f32 %f3268, %f1205, %f1211; mul.ftz.f32 %f3267, %f1203, %f1211; mul.ftz.f32 %f3266, %f1201, %f1211; mul.ftz.f32 %f3265, %f1199, %f1211; bra.uni $L__BB208_28; $L__BB208_26: mov.f32 %f3205, 0f3F800000; mov.f32 %f1145, 0fBF800000; max.ftz.f32 %f1146, %f76, %f1145; min.ftz.f32 %f1148, %f1146, %f3205; abs.ftz.f32 %f1149, %f1148; sub.ftz.f32 %f1150, %f3205, %f1149; mul.ftz.f32 %f1151, %f1150, 0f3F000000; sqrt.approx.ftz.f32 %f1152, %f1151; setp.gt.ftz.f32 %p57, %f1149, 0f3F11EB85; selp.f32 %f1153, %f1152, %f1149, %p57; mul.ftz.f32 %f1154, %f1153, %f1153; mov.f32 %f1155, 0f3C94D2E9; mov.f32 %f1156, 0f3D53F941; fma.rn.ftz.f32 %f1157, %f1156, %f1154, %f1155; mov.f32 %f1158, 0f3D3F841F; fma.rn.ftz.f32 %f1159, %f1157, %f1154, %f1158; mov.f32 %f1160, 0f3D994929; fma.rn.ftz.f32 %f1161, %f1159, %f1154, %f1160; mov.f32 %f1162, 0f3E2AAB94; fma.rn.ftz.f32 %f1163, %f1161, %f1154, %f1162; mul.ftz.f32 %f1164, %f1154, %f1163; fma.rn.ftz.f32 %f1165, %f1164, %f1153, %f1153; add.ftz.f32 %f1166, %f1165, %f1165; mov.f32 %f1167, 0f3FC90FDB; sub.ftz.f32 %f1168, %f1167, %f1165; selp.f32 %f1169, %f1166, %f1168, %p57; setp.lt.ftz.f32 %p58, %f1148, 0f00000000; mov.f32 %f1170, 0f40490FDB; sub.ftz.f32 %f1171, %f1170, %f1169; selp.f32 %f1172, %f1171, %f1169, %p58; mul.ftz.f32 %f1173, %f1132, %f76; sub.ftz.f32 %f1174, %f1136, %f1173; mul.ftz.f32 %f1175, %f1133, %f76; sub.ftz.f32 %f1176, %f1137, %f1175; mul.ftz.f32 %f1177, %f1134, %f76; sub.ftz.f32 %f1178, %f1138, %f1177; mul.ftz.f32 %f1179, %f1135, %f76; sub.ftz.f32 %f1180, %f1139, %f1179; mul.ftz.f32 %f1181, %f1176, %f1176; fma.rn.ftz.f32 %f1182, %f1174, %f1174, %f1181; mul.ftz.f32 %f1183, %f1180, %f1180; fma.rn.ftz.f32 %f1184, %f1178, %f1178, %f1183; add.ftz.f32 %f1185, %f1182, %f1184; rsqrt.approx.ftz.f32 %f1186, %f1185; mul.ftz.f32 %f1187, %f1174, %f1186; mul.ftz.f32 %f1188, %f1176, %f1186; mul.ftz.f32 %f1189, %f1178, %f1186; mul.ftz.f32 %f1190, %f1180, %f1186; mul.ftz.f32 %f1191, %f67, %f1172; cos.approx.ftz.f32 %f1192, %f1191; sin.approx.ftz.f32 %f1193, %f1191; mul.ftz.f32 %f1194, %f1187, %f1193; mul.ftz.f32 %f1195, %f1188, %f1193; mul.ftz.f32 %f1196, %f1189, %f1193; mul.ftz.f32 %f1197, %f1190, %f1193; fma.rn.ftz.f32 %f3268, %f1135, %f1192, %f1197; fma.rn.ftz.f32 %f3267, %f1134, %f1192, %f1196; fma.rn.ftz.f32 %f3266, %f1133, %f1192, %f1195; fma.rn.ftz.f32 %f3265, %f1132, %f1192, %f1194; $L__BB208_28: mov.f32 %f3206, 0f3F800000; ld.global.v4.f32 {%f1212, %f1213, %f1214, %f1215}, [%rd37+16]; ld.global.v4.f32 {%f1220, %f1221, %f1222, %f1223}, [%rd37+80]; mul.ftz.f32 %f1228, %f67, %f1220; mul.ftz.f32 %f1229, %f67, %f1221; mul.ftz.f32 %f1230, %f67, %f1222; mul.ftz.f32 %f1231, %f67, %f1223; fma.rn.ftz.f32 %f90, %f77, %f1212, %f1228; fma.rn.ftz.f32 %f91, %f77, %f1213, %f1229; fma.rn.ftz.f32 %f92, %f77, %f1214, %f1230; fma.rn.ftz.f32 %f1232, %f77, %f1215, %f1231; ld.global.v4.f32 {%f1233, %f1234, %f1235, %f1236}, [%rd37+32]; ld.global.v4.f32 {%f1241, %f1242, %f1243, %f1244}, [%rd37+96]; mul.ftz.f32 %f1249, %f67, %f1241; mul.ftz.f32 %f1250, %f67, %f1242; mul.ftz.f32 %f1251, %f67, %f1243; mul.ftz.f32 %f1252, %f67, %f1244; fma.rn.ftz.f32 %f1253, %f77, %f1233, %f1249; fma.rn.ftz.f32 %f1254, %f77, %f1234, %f1250; fma.rn.ftz.f32 %f1255, %f77, %f1235, %f1251; fma.rn.ftz.f32 %f1256, %f77, %f1236, %f1252; ld.global.v4.f32 {%f1257, %f1258, %f1259, %f1260}, [%rd37+48]; ld.global.v4.f32 {%f1265, %f1266, %f1267, %f1268}, [%rd37+112]; mul.ftz.f32 %f1273, %f67, %f1265; mul.ftz.f32 %f1274, %f67, %f1266; mul.ftz.f32 %f1275, %f67, %f1267; mul.ftz.f32 %f1276, %f67, %f1268; fma.rn.ftz.f32 %f1277, %f77, %f1257, %f1273; fma.rn.ftz.f32 %f1278, %f77, %f1258, %f1274; fma.rn.ftz.f32 %f1279, %f77, %f1259, %f1275; fma.rn.ftz.f32 %f1280, %f77, %f1260, %f1276; mul.ftz.f32 %f1281, %f3268, 0f3FB504F3; mul.ftz.f32 %f1282, %f3265, 0f3FB504F3; mul.ftz.f32 %f1283, %f3266, 0f3FB504F3; mul.ftz.f32 %f1284, %f3267, 0f3FB504F3; mul.ftz.f32 %f1285, %f1281, %f1282; mul.ftz.f32 %f1286, %f1281, %f1283; mul.ftz.f32 %f1287, %f1281, %f1284; mul.ftz.f32 %f1288, %f1282, %f1282; mul.ftz.f32 %f1289, %f1282, %f1283; mul.ftz.f32 %f1290, %f1282, %f1284; mul.ftz.f32 %f1291, %f1283, %f1283; mul.ftz.f32 %f1292, %f1283, %f1284; mul.ftz.f32 %f1293, %f1284, %f1284; sub.ftz.f32 %f1295, %f3206, %f1291; sub.ftz.f32 %f1296, %f1295, %f1293; sub.ftz.f32 %f1297, %f1289, %f1287; add.ftz.f32 %f1298, %f1286, %f1290; add.ftz.f32 %f1299, %f1287, %f1289; sub.ftz.f32 %f1300, %f3206, %f1288; sub.ftz.f32 %f1301, %f1300, %f1293; sub.ftz.f32 %f1302, %f1292, %f1285; sub.ftz.f32 %f1303, %f1290, %f1286; add.ftz.f32 %f1304, %f1285, %f1292; sub.ftz.f32 %f1305, %f1300, %f1291; mul.ftz.f32 %f1306, %f1297, %f1255; fma.rn.ftz.f32 %f1307, %f1296, %f1232, %f1306; fma.rn.ftz.f32 %f3274, %f1298, %f1278, %f1307; mul.ftz.f32 %f1308, %f1297, %f1256; fma.rn.ftz.f32 %f1309, %f1296, %f1253, %f1308; fma.rn.ftz.f32 %f94, %f1298, %f1279, %f1309; mul.ftz.f32 %f1310, %f1297, %f1277; fma.rn.ftz.f32 %f1311, %f1296, %f1254, %f1310; fma.rn.ftz.f32 %f95, %f1298, %f1280, %f1311; mul.ftz.f32 %f1312, %f1301, %f1255; fma.rn.ftz.f32 %f1313, %f1299, %f1232, %f1312; fma.rn.ftz.f32 %f96, %f1302, %f1278, %f1313; mul.ftz.f32 %f1314, %f1301, %f1256; fma.rn.ftz.f32 %f1315, %f1299, %f1253, %f1314; fma.rn.ftz.f32 %f3273, %f1302, %f1279, %f1315; mul.ftz.f32 %f1316, %f1301, %f1277; fma.rn.ftz.f32 %f1317, %f1299, %f1254, %f1316; fma.rn.ftz.f32 %f98, %f1302, %f1280, %f1317; mul.ftz.f32 %f1318, %f1304, %f1255; fma.rn.ftz.f32 %f1319, %f1303, %f1232, %f1318; fma.rn.ftz.f32 %f99, %f1305, %f1278, %f1319; mul.ftz.f32 %f1320, %f1304, %f1256; fma.rn.ftz.f32 %f1321, %f1303, %f1253, %f1320; fma.rn.ftz.f32 %f100, %f1305, %f1279, %f1321; mul.ftz.f32 %f1322, %f1304, %f1277; fma.rn.ftz.f32 %f1323, %f1303, %f1254, %f1322; fma.rn.ftz.f32 %f3272, %f1305, %f1280, %f1323; mul.ftz.f32 %f1324, %f3273, %f3272; mul.ftz.f32 %f102, %f100, %f98; sub.ftz.f32 %f3271, %f1324, %f102; mul.ftz.f32 %f104, %f100, %f95; mul.ftz.f32 %f1325, %f94, %f3272; sub.ftz.f32 %f3270, %f104, %f1325; mul.ftz.f32 %f106, %f94, %f98; mul.ftz.f32 %f1326, %f3273, %f95; sub.ftz.f32 %f3269, %f106, %f1326; mul.ftz.f32 %f1327, %f96, %f3270; fma.rn.ftz.f32 %f1328, %f3274, %f3271, %f1327; fma.rn.ftz.f32 %f3275, %f99, %f3269, %f1328; setp.neu.ftz.f32 %p59, %f3275, 0f00000000; @%p59 bra $L__BB208_30; add.ftz.f32 %f3274, %f3274, 0f322BCC77; add.ftz.f32 %f3272, %f3272, 0f322BCC77; add.ftz.f32 %f3273, %f3273, 0f322BCC77; mul.ftz.f32 %f1329, %f3273, %f3272; sub.ftz.f32 %f3271, %f1329, %f102; mul.ftz.f32 %f1330, %f94, %f3272; sub.ftz.f32 %f3270, %f104, %f1330; mul.ftz.f32 %f1331, %f3273, %f95; sub.ftz.f32 %f3269, %f106, %f1331; mul.ftz.f32 %f1332, %f96, %f3270; fma.rn.ftz.f32 %f1333, %f3274, %f3271, %f1332; fma.rn.ftz.f32 %f1334, %f99, %f3269, %f1333; setp.eq.ftz.f32 %p60, %f1334, 0f00000000; selp.f32 %f3275, 0f7F7FFFFF, %f1334, %p60; $L__BB208_30: mul.ftz.f32 %f1335, %f96, %f3272; mul.ftz.f32 %f1336, %f99, %f98; sub.ftz.f32 %f1337, %f1336, %f1335; mul.ftz.f32 %f1338, %f99, %f95; mul.ftz.f32 %f1339, %f3272, %f3274; sub.ftz.f32 %f1340, %f1339, %f1338; mul.ftz.f32 %f1341, %f98, %f3274; mul.ftz.f32 %f1342, %f96, %f95; sub.ftz.f32 %f1343, %f1342, %f1341; div.approx.ftz.f32 %f3280, %f1337, %f3275; div.approx.ftz.f32 %f3281, %f1340, %f3275; div.approx.ftz.f32 %f3282, %f1343, %f3275; mul.ftz.f32 %f1344, %f99, %f3273; mul.ftz.f32 %f1345, %f96, %f100; sub.ftz.f32 %f1346, %f1345, %f1344; mul.ftz.f32 %f1347, %f100, %f3274; mul.ftz.f32 %f1348, %f99, %f94; sub.ftz.f32 %f1349, %f1348, %f1347; mul.ftz.f32 %f1350, %f96, %f94; mul.ftz.f32 %f1351, %f3273, %f3274; sub.ftz.f32 %f1352, %f1351, %f1350; div.approx.ftz.f32 %f3284, %f1346, %f3275; div.approx.ftz.f32 %f3285, %f1349, %f3275; div.approx.ftz.f32 %f3286, %f1352, %f3275; div.approx.ftz.f32 %f3276, %f3271, %f3275; div.approx.ftz.f32 %f3277, %f3270, %f3275; mul.ftz.f32 %f1353, %f91, %f3277; fma.rn.ftz.f32 %f1354, %f90, %f3276, %f1353; div.approx.ftz.f32 %f3278, %f3269, %f3275; fma.rn.ftz.f32 %f1355, %f92, %f3278, %f1354; neg.ftz.f32 %f3279, %f1355; mul.ftz.f32 %f1356, %f91, %f3281; fma.rn.ftz.f32 %f1357, %f90, %f3280, %f1356; fma.rn.ftz.f32 %f1358, %f92, %f3282, %f1357; neg.ftz.f32 %f3283, %f1358; mul.ftz.f32 %f1359, %f91, %f3285; fma.rn.ftz.f32 %f1360, %f90, %f3284, %f1359; fma.rn.ftz.f32 %f1361, %f92, %f3286, %f1360; neg.ftz.f32 %f3287, %f1361; $L__BB208_32: cvt.s64.s32 %rd854, %r588; mul.ftz.f32 %f1374, %f2, %f3277; fma.rn.ftz.f32 %f1375, %f1, %f3276, %f1374; fma.rn.ftz.f32 %f1376, %f3, %f3278, %f1375; add.ftz.f32 %f3328, %f3279, %f1376; mul.ftz.f32 %f1377, %f2, %f3281; fma.rn.ftz.f32 %f1378, %f1, %f3280, %f1377; fma.rn.ftz.f32 %f1379, %f3, %f3282, %f1378; add.ftz.f32 %f3327, %f3283, %f1379; mul.ftz.f32 %f1380, %f2, %f3285; fma.rn.ftz.f32 %f1381, %f1, %f3284, %f1380; fma.rn.ftz.f32 %f1382, %f3, %f3286, %f1381; add.ftz.f32 %f3326, %f3287, %f1382; mul.ftz.f32 %f1383, %f3277, 0f00000000; fma.rn.ftz.f32 %f1384, %f3276, 0f00000000, %f1383; add.ftz.f32 %f1385, %f3278, %f1384; mul.ftz.f32 %f1386, %f3281, 0f00000000; fma.rn.ftz.f32 %f1387, %f3280, 0f00000000, %f1386; add.ftz.f32 %f1388, %f3282, %f1387; mul.ftz.f32 %f1389, %f3285, 0f00000000; fma.rn.ftz.f32 %f1390, %f3284, 0f00000000, %f1389; add.ftz.f32 %f1391, %f3286, %f1390; abs.ftz.f32 %f1392, %f1385; setp.gt.ftz.f32 %p61, %f1392, 0f17800000; mov.b32 %r308, %f1385; and.b32 %r309, %r308, -2147483648; or.b32 %r310, %r309, 394264576; mov.b32 %f1393, %r310; selp.f32 %f3331, %f1385, %f1393, %p61; abs.ftz.f32 %f1394, %f1388; setp.gt.ftz.f32 %p62, %f1394, 0f17800000; mov.b32 %r311, %f1388; and.b32 %r312, %r311, -2147483648; or.b32 %r313, %r312, 394264576; mov.b32 %f1395, %r313; selp.f32 %f3330, %f1388, %f1395, %p62; abs.ftz.f32 %f1396, %f1391; setp.gt.ftz.f32 %p63, %f1396, 0f17800000; mov.b32 %r314, %f1391; and.b32 %r315, %r314, -2147483648; or.b32 %r316, %r315, 394264576; mov.b32 %f1397, %r316; selp.f32 %f3329, %f1391, %f1397, %p63; rcp.approx.ftz.f32 %f3334, %f3331; rcp.approx.ftz.f32 %f3333, %f3330; rcp.approx.ftz.f32 %f3332, %f3329; add.s32 %r586, %r50, 1; mov.u32 %r317, 1985229328; st.local.u32 [%rd34+4], %r317; shl.b64 %rd207, %rd854, 2; add.s64 %rd208, %rd20, %rd207; ld.global.u32 %r559, [%rd208]; bra.uni $L__BB208_76; $L__BB208_10: setp.eq.s32 %p30, %r559, 1985229328; setp.lt.s32 %p31, %r559, 0; or.pred %p32, %p31, %p30; @%p32 bra $L__BB208_21; $L__BB208_11: mul.wide.s32 %rd187, %r559, 16; add.s64 %rd32, %rd9, %rd187; ld.global.v4.f32 {%f916, %f917, %f918, %f919}, [%rd32]; mov.b32 %r559, %f919; mov.b32 %r35, %f918; mov.b32 %r34, %f917; mov.b32 %r33, %f916; mov.b64 %rd188, {%r33, %r34}; cvt.u32.u64 %r287, %rd188; and.b32 %r288, %r287, 1024; setp.eq.s32 %p33, %r288, 0; and.b32 %r37, %r4, %r33; @%p33 bra $L__BB208_14; bra.uni $L__BB208_12; $L__BB208_14: setp.ne.s32 %p39, %r37, 0; ld.global.v4.f32 {%f1059, %f1060, %f1061, %f1062}, [%rd32+16]; ld.global.v4.f32 {%f1067, %f1068, %f1069, %f1070}, [%rd32+32]; ld.global.v4.f32 {%f1075, %f1076, %f1077, %f1078}, [%rd32+48]; sub.ftz.f32 %f1083, %f1059, %f3328; mul.ftz.f32 %f1084, %f3334, %f1083; sub.ftz.f32 %f1085, %f1061, %f3328; mul.ftz.f32 %f1086, %f3334, %f1085; sub.ftz.f32 %f1087, %f1067, %f3327; mul.ftz.f32 %f1088, %f3333, %f1087; sub.ftz.f32 %f1089, %f1069, %f3327; mul.ftz.f32 %f1090, %f3333, %f1089; sub.ftz.f32 %f1091, %f1075, %f3326; mul.ftz.f32 %f1092, %f3332, %f1091; sub.ftz.f32 %f1093, %f1077, %f3326; mul.ftz.f32 %f1094, %f3332, %f1093; min.ftz.f32 %f1095, %f1084, %f1086; min.ftz.f32 %f1096, %f1088, %f1090; min.ftz.f32 %f1097, %f1092, %f1094; max.ftz.f32 %f1098, %f3217, %f1095; max.ftz.f32 %f1099, %f1096, %f1097; max.ftz.f32 %f3264, %f1098, %f1099; max.ftz.f32 %f1100, %f1084, %f1086; max.ftz.f32 %f1101, %f1088, %f1090; max.ftz.f32 %f1102, %f1092, %f1094; min.ftz.f32 %f1103, %f3337, %f1100; min.ftz.f32 %f1104, %f1101, %f1102; min.ftz.f32 %f1105, %f1103, %f1104; sub.ftz.f32 %f1106, %f1060, %f3328; mul.ftz.f32 %f1107, %f3334, %f1106; sub.ftz.f32 %f1108, %f1062, %f3328; mul.ftz.f32 %f1109, %f3334, %f1108; sub.ftz.f32 %f1110, %f1068, %f3327; mul.ftz.f32 %f1111, %f3333, %f1110; sub.ftz.f32 %f1112, %f1070, %f3327; mul.ftz.f32 %f1113, %f3333, %f1112; sub.ftz.f32 %f1114, %f1076, %f3326; mul.ftz.f32 %f1115, %f3332, %f1114; sub.ftz.f32 %f1116, %f1078, %f3326; mul.ftz.f32 %f1117, %f3332, %f1116; min.ftz.f32 %f1118, %f1107, %f1109; min.ftz.f32 %f1119, %f1111, %f1113; min.ftz.f32 %f1120, %f1115, %f1117; max.ftz.f32 %f1121, %f3217, %f1118; max.ftz.f32 %f1122, %f1119, %f1120; max.ftz.f32 %f3263, %f1121, %f1122; max.ftz.f32 %f1123, %f1107, %f1109; max.ftz.f32 %f1124, %f1111, %f1113; max.ftz.f32 %f1125, %f1115, %f1117; min.ftz.f32 %f1126, %f3337, %f1123; min.ftz.f32 %f1127, %f1124, %f1125; min.ftz.f32 %f1128, %f1126, %f1127; setp.ge.ftz.f32 %p40, %f1105, %f3264; and.pred %p41, %p39, %p40; selp.u32 %r292, 1, 0, %p41; setp.ltu.ftz.f32 %p42, %f1128, %f3263; and.b32 %r294, %r4, %r34; setp.eq.s32 %p43, %r294, 0; or.pred %p44, %p43, %p42; selp.b32 %r295, 0, 2, %p44; or.b32 %r557, %r295, %r292; bra.uni $L__BB208_15; $L__BB208_12: setp.ne.s32 %p34, %r37, 0; ld.global.v4.f32 {%f923, %f924, %f925, %f926}, [%rd32+16]; ld.global.v4.f32 {%f931, %f932, %f933, %f934}, [%rd32+32]; ld.global.v4.f32 {%f939, %f940, %f941, %f942}, [%rd32+48]; mul.ftz.f32 %f947, %f3330, %f924; fma.rn.ftz.f32 %f948, %f3331, %f923, %f947; fma.rn.ftz.f32 %f949, %f3329, %f925, %f948; mul.ftz.f32 %f950, %f3330, %f932; fma.rn.ftz.f32 %f951, %f3331, %f931, %f950; fma.rn.ftz.f32 %f952, %f3329, %f933, %f951; mul.ftz.f32 %f953, %f3330, %f940; fma.rn.ftz.f32 %f954, %f3331, %f939, %f953; fma.rn.ftz.f32 %f955, %f3329, %f941, %f954; mul.ftz.f32 %f956, %f3327, %f924; fma.rn.ftz.f32 %f957, %f3328, %f923, %f956; fma.rn.ftz.f32 %f958, %f3326, %f925, %f957; add.ftz.f32 %f959, %f926, %f958; mul.ftz.f32 %f960, %f3327, %f932; fma.rn.ftz.f32 %f961, %f3328, %f931, %f960; fma.rn.ftz.f32 %f962, %f3326, %f933, %f961; add.ftz.f32 %f963, %f934, %f962; mul.ftz.f32 %f964, %f3327, %f940; fma.rn.ftz.f32 %f965, %f3328, %f939, %f964; fma.rn.ftz.f32 %f966, %f3326, %f941, %f965; add.ftz.f32 %f967, %f942, %f966; rcp.approx.ftz.f32 %f968, %f949; rcp.approx.ftz.f32 %f969, %f952; rcp.approx.ftz.f32 %f970, %f955; mul.ftz.f32 %f971, %f968, %f959; neg.ftz.f32 %f972, %f971; mul.ftz.f32 %f973, %f969, %f963; neg.ftz.f32 %f974, %f973; mul.ftz.f32 %f975, %f970, %f967; neg.ftz.f32 %f976, %f975; sub.ftz.f32 %f977, %f968, %f971; sub.ftz.f32 %f978, %f969, %f973; sub.ftz.f32 %f979, %f970, %f975; min.ftz.f32 %f980, %f972, %f977; min.ftz.f32 %f981, %f974, %f978; min.ftz.f32 %f982, %f976, %f979; max.ftz.f32 %f983, %f972, %f977; max.ftz.f32 %f984, %f974, %f978; max.ftz.f32 %f985, %f976, %f979; max.ftz.f32 %f986, %f3217, %f980; max.ftz.f32 %f987, %f981, %f982; max.ftz.f32 %f3264, %f986, %f987; min.ftz.f32 %f988, %f3337, %f983; min.ftz.f32 %f989, %f984, %f985; min.ftz.f32 %f990, %f988, %f989; setp.le.ftz.f32 %p35, %f3264, %f990; and.pred %p36, %p34, %p35; selp.u32 %r557, 1, 0, %p36; ld.global.v4.f32 {%f991, %f992, %f993, %f994}, [%rd32+64]; ld.global.v4.f32 {%f999, %f1000, %f1001, %f1002}, [%rd32+80]; ld.global.v4.f32 {%f1007, %f1008, %f1009, %f1010}, [%rd32+96]; mul.ftz.f32 %f1015, %f3330, %f992; fma.rn.ftz.f32 %f1016, %f3331, %f991, %f1015; fma.rn.ftz.f32 %f1017, %f3329, %f993, %f1016; mul.ftz.f32 %f1018, %f3330, %f1000; fma.rn.ftz.f32 %f1019, %f3331, %f999, %f1018; fma.rn.ftz.f32 %f1020, %f3329, %f1001, %f1019; mul.ftz.f32 %f1021, %f3330, %f1008; fma.rn.ftz.f32 %f1022, %f3331, %f1007, %f1021; fma.rn.ftz.f32 %f1023, %f3329, %f1009, %f1022; mul.ftz.f32 %f1024, %f3327, %f992; fma.rn.ftz.f32 %f1025, %f3328, %f991, %f1024; fma.rn.ftz.f32 %f1026, %f3326, %f993, %f1025; add.ftz.f32 %f1027, %f994, %f1026; mul.ftz.f32 %f1028, %f3327, %f1000; fma.rn.ftz.f32 %f1029, %f3328, %f999, %f1028; fma.rn.ftz.f32 %f1030, %f3326, %f1001, %f1029; add.ftz.f32 %f1031, %f1002, %f1030; mul.ftz.f32 %f1032, %f3327, %f1008; fma.rn.ftz.f32 %f1033, %f3328, %f1007, %f1032; fma.rn.ftz.f32 %f1034, %f3326, %f1009, %f1033; add.ftz.f32 %f1035, %f1010, %f1034; rcp.approx.ftz.f32 %f1036, %f1017; rcp.approx.ftz.f32 %f1037, %f1020; rcp.approx.ftz.f32 %f1038, %f1023; mul.ftz.f32 %f1039, %f1036, %f1027; neg.ftz.f32 %f1040, %f1039; mul.ftz.f32 %f1041, %f1037, %f1031; neg.ftz.f32 %f1042, %f1041; mul.ftz.f32 %f1043, %f1038, %f1035; neg.ftz.f32 %f1044, %f1043; sub.ftz.f32 %f1045, %f1036, %f1039; sub.ftz.f32 %f1046, %f1037, %f1041; sub.ftz.f32 %f1047, %f1038, %f1043; min.ftz.f32 %f1048, %f1040, %f1045; min.ftz.f32 %f1049, %f1042, %f1046; min.ftz.f32 %f1050, %f1044, %f1047; max.ftz.f32 %f1051, %f1040, %f1045; max.ftz.f32 %f1052, %f1042, %f1046; max.ftz.f32 %f1053, %f1044, %f1047; max.ftz.f32 %f1054, %f3217, %f1048; max.ftz.f32 %f1055, %f1049, %f1050; max.ftz.f32 %f3263, %f1054, %f1055; min.ftz.f32 %f1056, %f3337, %f1051; min.ftz.f32 %f1057, %f1052, %f1053; min.ftz.f32 %f1058, %f1056, %f1057; setp.gtu.ftz.f32 %p37, %f3263, %f1058; @%p37 bra $L__BB208_15; and.b32 %r290, %r4, %r34; setp.eq.s32 %p38, %r290, 0; or.b32 %r291, %r557, 2; selp.b32 %r557, %r557, %r291, %p38; $L__BB208_15: mul.wide.s32 %rd189, %r586, 4; add.s64 %rd33, %rd1, %rd189; setp.eq.s32 %p45, %r557, 3; @%p45 bra $L__BB208_19; bra.uni $L__BB208_16; $L__BB208_19: setp.lt.ftz.f32 %p48, %f3263, %f3264; selp.b32 %r46, %r559, %r35, %p48; selp.b32 %r296, %r35, %r559, %p48; add.s32 %r586, %r586, 1; st.local.u32 [%rd33+4], %r296; mov.u32 %r559, %r46; bra.uni $L__BB208_20; $L__BB208_16: setp.eq.s32 %p46, %r557, 2; @%p46 bra $L__BB208_20; setp.ne.s32 %p47, %r557, 0; mov.u32 %r559, %r35; @%p47 bra $L__BB208_20; ld.local.u32 %r559, [%rd33]; add.s32 %r586, %r586, -1; $L__BB208_20: setp.ne.s32 %p49, %r559, 1985229328; setp.gt.s32 %p50, %r559, -1; and.pred %p51, %p50, %p49; @%p51 bra $L__BB208_11; $L__BB208_21: mov.u32 %r50, %r586; setp.gt.s32 %p52, %r559, -1; @%p52 bra $L__BB208_76; bra.uni $L__BB208_22; $L__BB208_76: setp.ne.s32 %p115, %r559, 1985229328; @%p115 bra $L__BB208_10; mov.u32 %r559, 1985229328; setp.lt.s32 %p116, %r586, 0; @%p116 bra $L__BB208_79; mul.wide.s32 %rd253, %r586, 4; add.s64 %rd254, %rd1, %rd253; ld.local.u32 %r559, [%rd254]; add.s32 %r586, %r586, -1; mov.u32 %r588, -1; mov.f32 %f3326, %f3; mov.f32 %f3327, %f2; mov.f32 %f3328, %f1; mov.f32 %f3329, %f21; mov.f32 %f3330, %f20; mov.f32 %f3331, %f20; mov.f32 %f3332, %f23; mov.f32 %f3333, %f22; mov.f32 %f3334, %f22; $L__BB208_79: setp.ne.s32 %p117, %r559, 1985229328; @%p117 bra $L__BB208_10; setp.ne.s32 %p285, %r585, -1; mov.u32 %r629, %r584; mov.u32 %r630, %r585; bra.uni $L__BB208_117; $L__BB208_81: mov.u32 %r362, 1985229328; st.local.u32 [%rd1], %r362; mov.u32 %r620, -1; mov.u32 %r607, 0; mov.f32 %f3337, 0f7F7FFFFF; mov.f32 %f3335, 0f00000000; mov.f32 %f3395, %f22; mov.f32 %f3396, %f22; mov.f32 %f3397, %f23; mov.f32 %f3398, %f20; mov.f32 %f3399, %f20; mov.f32 %f3400, %f21; mov.f32 %f3401, %f1; mov.f32 %f3402, %f2; mov.f32 %f3403, %f3; mov.u32 %r621, %r620; mov.f32 %f3336, %f3335; mov.u32 %r608, %r280; mov.u32 %r624, %r620; bra.uni $L__BB208_82; $L__BB208_94: cvta.to.global.u64 %rd258, %rd167; mul.wide.s32 %rd259, %r608, 16; xor.b64 %rd260, %rd259, -16; add.s64 %rd261, %rd258, %rd260; ld.global.v4.u32 {%r611, %r374, %r375, %r376}, [%rd261]; setp.gt.s32 %p141, %r611, -1; mul.wide.s32 %rd262, %r153, 4; add.s64 %rd102, %rd1, %rd262; @%p141 bra $L__BB208_98; bra.uni $L__BB208_95; $L__BB208_98: ld.local.u32 %r608, [%rd102]; add.s32 %r607, %r153, -1; and.b32 %r388, %r376, 127; setp.ne.s32 %p146, %r388, 1; setp.le.s32 %p147, %r374, %r611; or.pred %p148, %p146, %p147; @%p148 bra $L__BB208_112; $L__BB208_99: cvt.s64.s32 %rd104, %r611; setp.ne.s32 %p149, %r624, -1; mov.u32 %r615, %r624; @%p149 bra $L__BB208_101; shl.b64 %rd272, %rd104, 2; add.s64 %rd273, %rd10, %rd272; ld.global.u32 %r615, [%rd273]; $L__BB208_101: cvt.s64.s32 %rd837, %r611; shl.b64 %rd274, %rd837, 2; add.s64 %rd275, %rd11, %rd274; ld.global.u32 %r389, [%rd275]; cvt.s64.s32 %rd105, %r389; setp.eq.s32 %p150, %r630, %r389; setp.eq.s32 %p151, %r629, %r615; and.pred %p152, %p151, %p150; @%p152 bra $L__BB208_111; mul.wide.s32 %rd276, %r615, 4; add.s64 %rd277, %rd6, %rd276; ld.global.u8 %rs25, [%rd277]; and.b16 %rs26, %rs25, 16; setp.eq.s16 %p153, %rs26, 0; @%p153 bra $L__BB208_111; shl.b64 %rd278, %rd105, 4; add.s64 %rd279, %rd14, %rd278; ld.global.u32 %r390, [%rd279+12]; mul.wide.u32 %rd280, %r390, 12; add.s64 %rd281, %rd17, %rd280; add.s32 %r391, %r390, 1; mul.wide.u32 %rd282, %r391, 12; add.s64 %rd283, %rd17, %rd282; add.s32 %r392, %r390, 2; mul.wide.u32 %rd284, %r392, 12; add.s64 %rd285, %rd17, %rd284; ld.global.f32 %f1857, [%rd281]; sub.ftz.f32 %f337, %f1857, %f3401; ld.global.f32 %f1858, [%rd281+4]; sub.ftz.f32 %f338, %f1858, %f3402; ld.global.f32 %f1859, [%rd281+8]; sub.ftz.f32 %f339, %f1859, %f3403; ld.global.f32 %f1860, [%rd283]; sub.ftz.f32 %f1861, %f1860, %f3401; ld.global.f32 %f1862, [%rd283+4]; sub.ftz.f32 %f1863, %f1862, %f3402; ld.global.f32 %f1864, [%rd283+8]; sub.ftz.f32 %f1865, %f1864, %f3403; ld.global.f32 %f1866, [%rd285]; sub.ftz.f32 %f1867, %f1866, %f3401; ld.global.f32 %f1868, [%rd285+4]; sub.ftz.f32 %f1869, %f1868, %f3402; ld.global.f32 %f1870, [%rd285+8]; sub.ftz.f32 %f1871, %f1870, %f3403; sub.ftz.f32 %f340, %f1867, %f337; sub.ftz.f32 %f341, %f1869, %f338; sub.ftz.f32 %f342, %f1871, %f339; sub.ftz.f32 %f343, %f337, %f1861; sub.ftz.f32 %f344, %f338, %f1863; sub.ftz.f32 %f345, %f339, %f1865; sub.ftz.f32 %f1872, %f1861, %f1867; sub.ftz.f32 %f1873, %f1863, %f1869; sub.ftz.f32 %f1874, %f1865, %f1871; add.ftz.f32 %f1875, %f337, %f1867; add.ftz.f32 %f1876, %f338, %f1869; add.ftz.f32 %f1877, %f339, %f1871; mul.ftz.f32 %f1878, %f341, %f1877; mul.ftz.f32 %f1879, %f1876, %f342; sub.ftz.f32 %f1880, %f1878, %f1879; mul.ftz.f32 %f1881, %f1875, %f342; mul.ftz.f32 %f1882, %f340, %f1877; sub.ftz.f32 %f1883, %f1881, %f1882; mul.ftz.f32 %f1884, %f340, %f1876; mul.ftz.f32 %f1885, %f1875, %f341; sub.ftz.f32 %f1886, %f1884, %f1885; mul.ftz.f32 %f1887, %f3399, %f1883; fma.rn.ftz.f32 %f1888, %f3398, %f1880, %f1887; fma.rn.ftz.f32 %f346, %f3400, %f1886, %f1888; add.ftz.f32 %f1889, %f337, %f1861; add.ftz.f32 %f1890, %f338, %f1863; add.ftz.f32 %f1891, %f339, %f1865; mul.ftz.f32 %f1892, %f344, %f1891; mul.ftz.f32 %f1893, %f1890, %f345; sub.ftz.f32 %f1894, %f1892, %f1893; mul.ftz.f32 %f1895, %f1889, %f345; mul.ftz.f32 %f1896, %f343, %f1891; sub.ftz.f32 %f1897, %f1895, %f1896; mul.ftz.f32 %f1898, %f343, %f1890; mul.ftz.f32 %f1899, %f1889, %f344; sub.ftz.f32 %f1900, %f1898, %f1899; mul.ftz.f32 %f1901, %f3399, %f1897; fma.rn.ftz.f32 %f1902, %f3398, %f1894, %f1901; fma.rn.ftz.f32 %f347, %f3400, %f1900, %f1902; add.ftz.f32 %f1903, %f1861, %f1867; add.ftz.f32 %f1904, %f1863, %f1869; add.ftz.f32 %f1905, %f1865, %f1871; mul.ftz.f32 %f1906, %f1873, %f1905; mul.ftz.f32 %f1907, %f1904, %f1874; sub.ftz.f32 %f1908, %f1906, %f1907; mul.ftz.f32 %f1909, %f1903, %f1874; mul.ftz.f32 %f1910, %f1872, %f1905; sub.ftz.f32 %f1911, %f1909, %f1910; mul.ftz.f32 %f1912, %f1872, %f1904; mul.ftz.f32 %f1913, %f1903, %f1873; sub.ftz.f32 %f1914, %f1912, %f1913; mul.ftz.f32 %f1915, %f3399, %f1911; fma.rn.ftz.f32 %f1916, %f3398, %f1908, %f1915; fma.rn.ftz.f32 %f1917, %f3400, %f1914, %f1916; add.ftz.f32 %f1918, %f347, %f346; add.ftz.f32 %f348, %f1917, %f1918; abs.ftz.f32 %f349, %f348; mul.ftz.f32 %f1919, %f349, 0f34000000; min.ftz.f32 %f1920, %f347, %f1917; min.ftz.f32 %f1921, %f346, %f1920; max.ftz.f32 %f1922, %f347, %f1917; max.ftz.f32 %f1923, %f346, %f1922; mul.ftz.f32 %f1924, %f349, 0fB4000000; setp.ltu.ftz.f32 %p155, %f1921, %f1924; setp.gtu.ftz.f32 %p156, %f1923, %f1919; and.pred %p157, %p155, %p156; mov.pred %p284, %p154; @%p157 bra $L__BB208_109; mul.ftz.f32 %f1925, %f344, %f342; mul.ftz.f32 %f1926, %f345, %f341; sub.ftz.f32 %f1927, %f1925, %f1926; mul.ftz.f32 %f1928, %f343, %f342; mul.ftz.f32 %f1929, %f345, %f340; sub.ftz.f32 %f1930, %f1929, %f1928; mul.ftz.f32 %f1931, %f344, %f340; mul.ftz.f32 %f1932, %f343, %f341; sub.ftz.f32 %f1933, %f1932, %f1931; add.ftz.f32 %f350, %f1927, %f1927; add.ftz.f32 %f351, %f1930, %f1930; add.ftz.f32 %f352, %f1933, %f1933; mul.ftz.f32 %f1934, %f3399, %f351; fma.rn.ftz.f32 %f1935, %f3398, %f350, %f1934; fma.rn.ftz.f32 %f353, %f3400, %f352, %f1935; setp.eq.ftz.f32 %p159, %f353, 0f00000000; mov.pred %p284, %p154; @%p159 bra $L__BB208_109; mul.ftz.f32 %f1936, %f338, %f351; fma.rn.ftz.f32 %f1937, %f337, %f350, %f1936; fma.rn.ftz.f32 %f1938, %f339, %f352, %f1937; div.approx.ftz.f32 %f354, %f1938, %f353; setp.ltu.ftz.f32 %p161, %f354, %f3217; setp.gtu.ftz.f32 %p162, %f354, %f3337; or.pred %p163, %p162, %p161; mov.pred %p284, %p154; @%p163 bra $L__BB208_109; abs.ftz.f32 %f3211, %f348; setp.lt.ftz.f32 %p164, %f3211, 0f219392EF; mov.f32 %f3385, 0f00000000; @%p164 bra $L__BB208_108; rcp.approx.ftz.f32 %f3385, %f348; $L__BB208_108: mov.f32 %f3209, 0f3F800000; mul.ftz.f32 %f1940, %f346, %f3385; min.ftz.f32 %f3390, %f1940, %f3209; mul.ftz.f32 %f1942, %f347, %f3385; min.ftz.f32 %f3389, %f1942, %f3209; mov.pred %p284, 0; mov.f32 %f3391, %f354; $L__BB208_109: @%p284 bra $L__BB208_111; cvt.s64.s32 %rd836, %r611; shl.b64 %rd835, %rd836, 2; cvt.u32.u64 %r393, %rd105; add.s64 %rd287, %rd18, %rd835; ld.global.u32 %r394, [%rd287]; and.b32 %r395, %r394, %r4; setp.eq.s32 %p166, %r395, 0; selp.b32 %r583, %r583, 1, %p166; selp.b32 %r620, %r620, %r615, %p166; selp.b32 %r621, %r621, %r393, %p166; selp.f32 %f3335, %f3335, %f3389, %p166; selp.f32 %f3336, %f3336, %f3390, %p166; selp.f32 %f3337, %f3337, %f3391, %p166; $L__BB208_111: add.s32 %r607, %r153, -1; cvt.u32.u64 %r396, %rd104; add.s32 %r611, %r396, 1; setp.lt.s32 %p167, %r611, %r374; @%p167 bra $L__BB208_99; bra.uni $L__BB208_112; $L__BB208_95: mul.wide.s32 %rd263, %r611, 4; xor.b64 %rd264, %rd263, -4; add.s64 %rd265, %rd10, %rd264; ld.global.u32 %r624, [%rd265]; cvt.s64.s32 %rd103, %r624; mul.wide.s32 %rd266, %r624, 4; add.s64 %rd267, %rd6, %rd266; ld.global.u8 %rs23, [%rd267]; and.b16 %rs24, %rs23, 16; setp.eq.s16 %p142, %rs24, 0; @%p142 bra $L__BB208_97; bra.uni $L__BB208_96; $L__BB208_97: ld.local.u32 %r608, [%rd102]; add.s32 %r607, %r153, -1; mov.u32 %r624, -1; bra.uni $L__BB208_112; $L__BB208_96: mul.lo.s64 %rd268, %rd103, 208; add.s64 %rd269, %rd19, %rd268; ld.global.v4.f32 {%f1809, %f1810, %f1811, %f1812}, [%rd269+48]; ld.global.v4.f32 {%f1817, %f1818, %f1819, %f1820}, [%rd269+64]; ld.global.v4.f32 {%f1825, %f1826, %f1827, %f1828}, [%rd269+80]; mul.ftz.f32 %f1833, %f2, %f1810; fma.rn.ftz.f32 %f1834, %f1, %f1809, %f1833; fma.rn.ftz.f32 %f1835, %f3, %f1811, %f1834; add.ftz.f32 %f3401, %f1812, %f1835; mul.ftz.f32 %f1836, %f2, %f1818; fma.rn.ftz.f32 %f1837, %f1, %f1817, %f1836; fma.rn.ftz.f32 %f1838, %f3, %f1819, %f1837; add.ftz.f32 %f3402, %f1820, %f1838; mul.ftz.f32 %f1839, %f2, %f1826; fma.rn.ftz.f32 %f1840, %f1, %f1825, %f1839; fma.rn.ftz.f32 %f1841, %f3, %f1827, %f1840; add.ftz.f32 %f3403, %f1828, %f1841; mul.ftz.f32 %f1842, %f1810, 0f00000000; fma.rn.ftz.f32 %f1843, %f1809, 0f00000000, %f1842; add.ftz.f32 %f1844, %f1811, %f1843; mul.ftz.f32 %f1845, %f1818, 0f00000000; fma.rn.ftz.f32 %f1846, %f1817, 0f00000000, %f1845; add.ftz.f32 %f1847, %f1819, %f1846; mul.ftz.f32 %f1848, %f1826, 0f00000000; fma.rn.ftz.f32 %f1849, %f1825, 0f00000000, %f1848; add.ftz.f32 %f1850, %f1827, %f1849; abs.ftz.f32 %f1851, %f1844; setp.gt.ftz.f32 %p143, %f1851, 0f17800000; mov.b32 %r377, %f1844; and.b32 %r378, %r377, -2147483648; or.b32 %r379, %r378, 394264576; mov.b32 %f1852, %r379; selp.f32 %f3398, %f1844, %f1852, %p143; abs.ftz.f32 %f1853, %f1847; setp.gt.ftz.f32 %p144, %f1853, 0f17800000; mov.b32 %r380, %f1847; and.b32 %r381, %r380, -2147483648; or.b32 %r382, %r381, 394264576; mov.b32 %f1854, %r382; selp.f32 %f3399, %f1847, %f1854, %p144; abs.ftz.f32 %f1855, %f1850; setp.gt.ftz.f32 %p145, %f1855, 0f17800000; mov.b32 %r383, %f1850; and.b32 %r384, %r383, -2147483648; or.b32 %r385, %r384, 394264576; mov.b32 %f1856, %r385; selp.f32 %f3400, %f1850, %f1856, %p145; rcp.approx.ftz.f32 %f3395, %f3398; rcp.approx.ftz.f32 %f3396, %f3399; rcp.approx.ftz.f32 %f3397, %f3400; add.s32 %r607, %r153, 1; mov.u32 %r386, 1985229328; st.local.u32 [%rd102+4], %r386; shl.b64 %rd270, %rd103, 2; add.s64 %rd271, %rd20, %rd270; ld.global.u32 %r608, [%rd271]; bra.uni $L__BB208_112; $L__BB208_82: setp.eq.s32 %p118, %r608, 1985229328; setp.lt.s32 %p119, %r608, 0; or.pred %p120, %p119, %p118; @%p120 bra $L__BB208_93; $L__BB208_83: mul.wide.s32 %rd255, %r608, 16; add.s64 %rd100, %rd9, %rd255; ld.global.v4.f32 {%f1596, %f1597, %f1598, %f1599}, [%rd100]; mov.b32 %r608, %f1599; mov.b32 %r138, %f1598; mov.b32 %r137, %f1597; mov.b32 %r136, %f1596; mov.b64 %rd256, {%r136, %r137}; cvt.u32.u64 %r363, %rd256; and.b32 %r364, %r363, 1024; setp.eq.s32 %p121, %r364, 0; and.b32 %r140, %r4, %r136; @%p121 bra $L__BB208_86; bra.uni $L__BB208_84; $L__BB208_86: setp.ne.s32 %p127, %r140, 0; ld.global.v4.f32 {%f1739, %f1740, %f1741, %f1742}, [%rd100+16]; ld.global.v4.f32 {%f1747, %f1748, %f1749, %f1750}, [%rd100+32]; ld.global.v4.f32 {%f1755, %f1756, %f1757, %f1758}, [%rd100+48]; sub.ftz.f32 %f1763, %f1739, %f3401; mul.ftz.f32 %f1764, %f3395, %f1763; sub.ftz.f32 %f1765, %f1741, %f3401; mul.ftz.f32 %f1766, %f3395, %f1765; sub.ftz.f32 %f1767, %f1747, %f3402; mul.ftz.f32 %f1768, %f3396, %f1767; sub.ftz.f32 %f1769, %f1749, %f3402; mul.ftz.f32 %f1770, %f3396, %f1769; sub.ftz.f32 %f1771, %f1755, %f3403; mul.ftz.f32 %f1772, %f3397, %f1771; sub.ftz.f32 %f1773, %f1757, %f3403; mul.ftz.f32 %f1774, %f3397, %f1773; min.ftz.f32 %f1775, %f1764, %f1766; min.ftz.f32 %f1776, %f1768, %f1770; min.ftz.f32 %f1777, %f1772, %f1774; max.ftz.f32 %f1778, %f3217, %f1775; max.ftz.f32 %f1779, %f1776, %f1777; max.ftz.f32 %f3378, %f1778, %f1779; max.ftz.f32 %f1780, %f1764, %f1766; max.ftz.f32 %f1781, %f1768, %f1770; max.ftz.f32 %f1782, %f1772, %f1774; min.ftz.f32 %f1783, %f3337, %f1780; min.ftz.f32 %f1784, %f1781, %f1782; min.ftz.f32 %f1785, %f1783, %f1784; sub.ftz.f32 %f1786, %f1740, %f3401; mul.ftz.f32 %f1787, %f3395, %f1786; sub.ftz.f32 %f1788, %f1742, %f3401; mul.ftz.f32 %f1789, %f3395, %f1788; sub.ftz.f32 %f1790, %f1748, %f3402; mul.ftz.f32 %f1791, %f3396, %f1790; sub.ftz.f32 %f1792, %f1750, %f3402; mul.ftz.f32 %f1793, %f3396, %f1792; sub.ftz.f32 %f1794, %f1756, %f3403; mul.ftz.f32 %f1795, %f3397, %f1794; sub.ftz.f32 %f1796, %f1758, %f3403; mul.ftz.f32 %f1797, %f3397, %f1796; min.ftz.f32 %f1798, %f1787, %f1789; min.ftz.f32 %f1799, %f1791, %f1793; min.ftz.f32 %f1800, %f1795, %f1797; max.ftz.f32 %f1801, %f3217, %f1798; max.ftz.f32 %f1802, %f1799, %f1800; max.ftz.f32 %f3377, %f1801, %f1802; max.ftz.f32 %f1803, %f1787, %f1789; max.ftz.f32 %f1804, %f1791, %f1793; max.ftz.f32 %f1805, %f1795, %f1797; min.ftz.f32 %f1806, %f3337, %f1803; min.ftz.f32 %f1807, %f1804, %f1805; min.ftz.f32 %f1808, %f1806, %f1807; setp.ge.ftz.f32 %p128, %f1785, %f3378; and.pred %p129, %p127, %p128; selp.u32 %r368, 1, 0, %p129; setp.ltu.ftz.f32 %p130, %f1808, %f3377; and.b32 %r370, %r4, %r137; setp.eq.s32 %p131, %r370, 0; or.pred %p132, %p131, %p130; selp.b32 %r371, 0, 2, %p132; or.b32 %r606, %r371, %r368; bra.uni $L__BB208_87; $L__BB208_84: setp.ne.s32 %p122, %r140, 0; ld.global.v4.f32 {%f1603, %f1604, %f1605, %f1606}, [%rd100+16]; ld.global.v4.f32 {%f1611, %f1612, %f1613, %f1614}, [%rd100+32]; ld.global.v4.f32 {%f1619, %f1620, %f1621, %f1622}, [%rd100+48]; mul.ftz.f32 %f1627, %f3399, %f1604; fma.rn.ftz.f32 %f1628, %f3398, %f1603, %f1627; fma.rn.ftz.f32 %f1629, %f3400, %f1605, %f1628; mul.ftz.f32 %f1630, %f3399, %f1612; fma.rn.ftz.f32 %f1631, %f3398, %f1611, %f1630; fma.rn.ftz.f32 %f1632, %f3400, %f1613, %f1631; mul.ftz.f32 %f1633, %f3399, %f1620; fma.rn.ftz.f32 %f1634, %f3398, %f1619, %f1633; fma.rn.ftz.f32 %f1635, %f3400, %f1621, %f1634; mul.ftz.f32 %f1636, %f3402, %f1604; fma.rn.ftz.f32 %f1637, %f3401, %f1603, %f1636; fma.rn.ftz.f32 %f1638, %f3403, %f1605, %f1637; add.ftz.f32 %f1639, %f1606, %f1638; mul.ftz.f32 %f1640, %f3402, %f1612; fma.rn.ftz.f32 %f1641, %f3401, %f1611, %f1640; fma.rn.ftz.f32 %f1642, %f3403, %f1613, %f1641; add.ftz.f32 %f1643, %f1614, %f1642; mul.ftz.f32 %f1644, %f3402, %f1620; fma.rn.ftz.f32 %f1645, %f3401, %f1619, %f1644; fma.rn.ftz.f32 %f1646, %f3403, %f1621, %f1645; add.ftz.f32 %f1647, %f1622, %f1646; rcp.approx.ftz.f32 %f1648, %f1629; rcp.approx.ftz.f32 %f1649, %f1632; rcp.approx.ftz.f32 %f1650, %f1635; mul.ftz.f32 %f1651, %f1648, %f1639; neg.ftz.f32 %f1652, %f1651; mul.ftz.f32 %f1653, %f1649, %f1643; neg.ftz.f32 %f1654, %f1653; mul.ftz.f32 %f1655, %f1650, %f1647; neg.ftz.f32 %f1656, %f1655; sub.ftz.f32 %f1657, %f1648, %f1651; sub.ftz.f32 %f1658, %f1649, %f1653; sub.ftz.f32 %f1659, %f1650, %f1655; min.ftz.f32 %f1660, %f1652, %f1657; min.ftz.f32 %f1661, %f1654, %f1658; min.ftz.f32 %f1662, %f1656, %f1659; max.ftz.f32 %f1663, %f1652, %f1657; max.ftz.f32 %f1664, %f1654, %f1658; max.ftz.f32 %f1665, %f1656, %f1659; max.ftz.f32 %f1666, %f3217, %f1660; max.ftz.f32 %f1667, %f1661, %f1662; max.ftz.f32 %f3378, %f1666, %f1667; min.ftz.f32 %f1668, %f3337, %f1663; min.ftz.f32 %f1669, %f1664, %f1665; min.ftz.f32 %f1670, %f1668, %f1669; setp.le.ftz.f32 %p123, %f3378, %f1670; and.pred %p124, %p122, %p123; selp.u32 %r606, 1, 0, %p124; ld.global.v4.f32 {%f1671, %f1672, %f1673, %f1674}, [%rd100+64]; ld.global.v4.f32 {%f1679, %f1680, %f1681, %f1682}, [%rd100+80]; ld.global.v4.f32 {%f1687, %f1688, %f1689, %f1690}, [%rd100+96]; mul.ftz.f32 %f1695, %f3399, %f1672; fma.rn.ftz.f32 %f1696, %f3398, %f1671, %f1695; fma.rn.ftz.f32 %f1697, %f3400, %f1673, %f1696; mul.ftz.f32 %f1698, %f3399, %f1680; fma.rn.ftz.f32 %f1699, %f3398, %f1679, %f1698; fma.rn.ftz.f32 %f1700, %f3400, %f1681, %f1699; mul.ftz.f32 %f1701, %f3399, %f1688; fma.rn.ftz.f32 %f1702, %f3398, %f1687, %f1701; fma.rn.ftz.f32 %f1703, %f3400, %f1689, %f1702; mul.ftz.f32 %f1704, %f3402, %f1672; fma.rn.ftz.f32 %f1705, %f3401, %f1671, %f1704; fma.rn.ftz.f32 %f1706, %f3403, %f1673, %f1705; add.ftz.f32 %f1707, %f1674, %f1706; mul.ftz.f32 %f1708, %f3402, %f1680; fma.rn.ftz.f32 %f1709, %f3401, %f1679, %f1708; fma.rn.ftz.f32 %f1710, %f3403, %f1681, %f1709; add.ftz.f32 %f1711, %f1682, %f1710; mul.ftz.f32 %f1712, %f3402, %f1688; fma.rn.ftz.f32 %f1713, %f3401, %f1687, %f1712; fma.rn.ftz.f32 %f1714, %f3403, %f1689, %f1713; add.ftz.f32 %f1715, %f1690, %f1714; rcp.approx.ftz.f32 %f1716, %f1697; rcp.approx.ftz.f32 %f1717, %f1700; rcp.approx.ftz.f32 %f1718, %f1703; mul.ftz.f32 %f1719, %f1716, %f1707; neg.ftz.f32 %f1720, %f1719; mul.ftz.f32 %f1721, %f1717, %f1711; neg.ftz.f32 %f1722, %f1721; mul.ftz.f32 %f1723, %f1718, %f1715; neg.ftz.f32 %f1724, %f1723; sub.ftz.f32 %f1725, %f1716, %f1719; sub.ftz.f32 %f1726, %f1717, %f1721; sub.ftz.f32 %f1727, %f1718, %f1723; min.ftz.f32 %f1728, %f1720, %f1725; min.ftz.f32 %f1729, %f1722, %f1726; min.ftz.f32 %f1730, %f1724, %f1727; max.ftz.f32 %f1731, %f1720, %f1725; max.ftz.f32 %f1732, %f1722, %f1726; max.ftz.f32 %f1733, %f1724, %f1727; max.ftz.f32 %f1734, %f3217, %f1728; max.ftz.f32 %f1735, %f1729, %f1730; max.ftz.f32 %f3377, %f1734, %f1735; min.ftz.f32 %f1736, %f3337, %f1731; min.ftz.f32 %f1737, %f1732, %f1733; min.ftz.f32 %f1738, %f1736, %f1737; setp.gtu.ftz.f32 %p125, %f3377, %f1738; @%p125 bra $L__BB208_87; and.b32 %r366, %r4, %r137; setp.eq.s32 %p126, %r366, 0; or.b32 %r367, %r606, 2; selp.b32 %r606, %r606, %r367, %p126; $L__BB208_87: mul.wide.s32 %rd257, %r607, 4; add.s64 %rd101, %rd1, %rd257; setp.eq.s32 %p133, %r606, 3; @%p133 bra $L__BB208_91; bra.uni $L__BB208_88; $L__BB208_91: setp.lt.ftz.f32 %p136, %f3377, %f3378; selp.b32 %r149, %r608, %r138, %p136; selp.b32 %r372, %r138, %r608, %p136; add.s32 %r607, %r607, 1; st.local.u32 [%rd101+4], %r372; mov.u32 %r608, %r149; bra.uni $L__BB208_92; $L__BB208_88: setp.eq.s32 %p134, %r606, 2; @%p134 bra $L__BB208_92; setp.ne.s32 %p135, %r606, 0; mov.u32 %r608, %r138; @%p135 bra $L__BB208_92; ld.local.u32 %r608, [%rd101]; add.s32 %r607, %r607, -1; $L__BB208_92: setp.ne.s32 %p137, %r608, 1985229328; setp.gt.s32 %p138, %r608, -1; and.pred %p139, %p138, %p137; @%p139 bra $L__BB208_83; $L__BB208_93: mov.u32 %r153, %r607; setp.gt.s32 %p140, %r608, -1; @%p140 bra $L__BB208_112; bra.uni $L__BB208_94; $L__BB208_112: setp.ne.s32 %p168, %r608, 1985229328; @%p168 bra $L__BB208_82; mov.u32 %r608, 1985229328; setp.lt.s32 %p169, %r607, 0; @%p169 bra $L__BB208_115; mul.wide.s32 %rd288, %r607, 4; add.s64 %rd289, %rd1, %rd288; ld.local.u32 %r608, [%rd289]; add.s32 %r607, %r607, -1; mov.u32 %r624, -1; mov.f32 %f3395, %f22; mov.f32 %f3396, %f22; mov.f32 %f3397, %f23; mov.f32 %f3398, %f20; mov.f32 %f3399, %f20; mov.f32 %f3400, %f21; mov.f32 %f3401, %f1; mov.f32 %f3402, %f2; mov.f32 %f3403, %f3; $L__BB208_115: setp.ne.s32 %p170, %r608, 1985229328; @%p170 bra $L__BB208_82; setp.ne.s32 %p285, %r621, -1; mov.u32 %r629, %r620; mov.u32 %r630, %r621; $L__BB208_117: not.pred %p171, %p285; @%p171 bra $L__BB208_232; add.u64 %rd290, %SP, 768; add.s64 %rd291, %rd290, 88; cvta.to.local.u64 %rd292, %rd291; st.local.f32 [%rd292], %f3337; mov.u32 %r399, 0; mov.b32 %r400, %f3335; mov.b32 %r401, %f3336; add.s64 %rd293, %rd290, 64; cvta.to.local.u64 %rd294, %rd293; st.local.v4.u32 [%rd294], {%r583, %r401, %r400, %r629}; mul.wide.s32 %rd295, %r629, 4; add.s64 %rd296, %rd6, %rd295; ld.global.u32 %r641, [%rd296]; add.s64 %rd297, %rd290, 56; cvta.to.local.u64 %rd298, %rd297; st.local.u32 [%rd298], %r641; add.s64 %rd299, %rd290, 60; cvta.to.local.u64 %rd300, %rd299; st.local.u32 [%rd300], %r630; add.s64 %rd301, %rd290, 52; cvta.to.local.u64 %rd302, %rd301; st.local.u32 [%rd302], %r399; mov.b32 %r402, %f4; mov.u32 %r403, -1; add.s64 %rd303, %rd290, 80; cvta.to.local.u64 %rd304, %rd303; st.local.v2.u32 [%rd304], {%r403, %r402}; and.b32 %r196, %r641, 2; setp.eq.s32 %p172, %r196, 0; mul.wide.s32 %rd305, %r629, 208; add.s64 %rd106, %rd19, %rd305; @%p172 bra $L__BB208_125; ld.global.u32 %rd306, [%rd106+168]; ld.global.u32 %r404, [%rd106+152]; shl.b32 %r405, %r404, 1; cvt.rn.f32.s32 %f1943, %r405; mul.ftz.f32 %f1944, %f4, %f1943; cvt.rzi.ftz.s32.f32 %r406, %f1944; add.s32 %r407, %r405, -1; min.s32 %r408, %r406, %r407; cvt.rn.f32.s32 %f1945, %r408; sub.ftz.f32 %f407, %f1944, %f1945; cvt.s64.s32 %rd307, %r408; add.s64 %rd308, %rd307, %rd306; cvta.to.global.u64 %rd309, %rd164; shl.b64 %rd310, %rd308, 6; add.s64 %rd107, %rd309, %rd310; ld.global.v4.f32 {%f1946, %f1947, %f1948, %f1949}, [%rd107]; ld.global.v4.f32 {%f1950, %f1951, %f1952, %f1953}, [%rd107+64]; mul.ftz.f32 %f1954, %f1947, %f1951; fma.rn.ftz.f32 %f1955, %f1946, %f1950, %f1954; mul.ftz.f32 %f1956, %f1949, %f1953; fma.rn.ftz.f32 %f1957, %f1948, %f1952, %f1956; add.ftz.f32 %f416, %f1955, %f1957; setp.gt.ftz.f32 %p173, %f416, 0f3F7FDF3B; mov.f32 %f1958, 0f3F800000; sub.ftz.f32 %f417, %f1958, %f407; @%p173 bra $L__BB208_121; bra.uni $L__BB208_120; $L__BB208_121: mul.ftz.f32 %f2012, %f1946, %f417; fma.rn.ftz.f32 %f2013, %f1950, %f407, %f2012; mul.ftz.f32 %f2014, %f1947, %f417; fma.rn.ftz.f32 %f2015, %f1951, %f407, %f2014; mul.ftz.f32 %f2016, %f1948, %f417; fma.rn.ftz.f32 %f2017, %f1952, %f407, %f2016; mul.ftz.f32 %f2018, %f1949, %f417; fma.rn.ftz.f32 %f2019, %f1953, %f407, %f2018; mul.ftz.f32 %f2020, %f2015, %f2015; fma.rn.ftz.f32 %f2021, %f2013, %f2013, %f2020; mul.ftz.f32 %f2022, %f2019, %f2019; fma.rn.ftz.f32 %f2023, %f2017, %f2017, %f2022; add.ftz.f32 %f2024, %f2021, %f2023; rsqrt.approx.ftz.f32 %f2025, %f2024; mul.ftz.f32 %f3434, %f2019, %f2025; mul.ftz.f32 %f3433, %f2017, %f2025; mul.ftz.f32 %f3432, %f2015, %f2025; mul.ftz.f32 %f3431, %f2013, %f2025; bra.uni $L__BB208_122; $L__BB208_120: mov.f32 %f1959, 0fBF800000; max.ftz.f32 %f1960, %f416, %f1959; min.ftz.f32 %f1962, %f1960, %f1958; abs.ftz.f32 %f1963, %f1962; sub.ftz.f32 %f1964, %f1958, %f1963; mul.ftz.f32 %f1965, %f1964, 0f3F000000; sqrt.approx.ftz.f32 %f1966, %f1965; setp.gt.ftz.f32 %p174, %f1963, 0f3F11EB85; selp.f32 %f1967, %f1966, %f1963, %p174; mul.ftz.f32 %f1968, %f1967, %f1967; mov.f32 %f1969, 0f3C94D2E9; mov.f32 %f1970, 0f3D53F941; fma.rn.ftz.f32 %f1971, %f1970, %f1968, %f1969; mov.f32 %f1972, 0f3D3F841F; fma.rn.ftz.f32 %f1973, %f1971, %f1968, %f1972; mov.f32 %f1974, 0f3D994929; fma.rn.ftz.f32 %f1975, %f1973, %f1968, %f1974; mov.f32 %f1976, 0f3E2AAB94; fma.rn.ftz.f32 %f1977, %f1975, %f1968, %f1976; mul.ftz.f32 %f1978, %f1968, %f1977; fma.rn.ftz.f32 %f1979, %f1978, %f1967, %f1967; add.ftz.f32 %f1980, %f1979, %f1979; mov.f32 %f1981, 0f3FC90FDB; sub.ftz.f32 %f1982, %f1981, %f1979; selp.f32 %f1983, %f1980, %f1982, %p174; setp.lt.ftz.f32 %p175, %f1962, 0f00000000; mov.f32 %f1984, 0f40490FDB; sub.ftz.f32 %f1985, %f1984, %f1983; selp.f32 %f1986, %f1985, %f1983, %p175; mul.ftz.f32 %f1987, %f1946, %f416; sub.ftz.f32 %f1988, %f1950, %f1987; mul.ftz.f32 %f1989, %f1947, %f416; sub.ftz.f32 %f1990, %f1951, %f1989; mul.ftz.f32 %f1991, %f1948, %f416; sub.ftz.f32 %f1992, %f1952, %f1991; mul.ftz.f32 %f1993, %f1949, %f416; sub.ftz.f32 %f1994, %f1953, %f1993; mul.ftz.f32 %f1995, %f1990, %f1990; fma.rn.ftz.f32 %f1996, %f1988, %f1988, %f1995; mul.ftz.f32 %f1997, %f1994, %f1994; fma.rn.ftz.f32 %f1998, %f1992, %f1992, %f1997; add.ftz.f32 %f1999, %f1996, %f1998; rsqrt.approx.ftz.f32 %f2000, %f1999; mul.ftz.f32 %f2001, %f1988, %f2000; mul.ftz.f32 %f2002, %f1990, %f2000; mul.ftz.f32 %f2003, %f1992, %f2000; mul.ftz.f32 %f2004, %f1994, %f2000; mul.ftz.f32 %f2005, %f407, %f1986; cos.approx.ftz.f32 %f2006, %f2005; sin.approx.ftz.f32 %f2007, %f2005; mul.ftz.f32 %f2008, %f2001, %f2007; mul.ftz.f32 %f2009, %f2002, %f2007; mul.ftz.f32 %f2010, %f2003, %f2007; mul.ftz.f32 %f2011, %f2004, %f2007; fma.rn.ftz.f32 %f3434, %f1949, %f2006, %f2011; fma.rn.ftz.f32 %f3433, %f1948, %f2006, %f2010; fma.rn.ftz.f32 %f3432, %f1947, %f2006, %f2009; fma.rn.ftz.f32 %f3431, %f1946, %f2006, %f2008; $L__BB208_122: ld.global.v4.f32 {%f2026, %f2027, %f2028, %f2029}, [%rd107+16]; ld.global.v4.f32 {%f2034, %f2035, %f2036, %f2037}, [%rd107+80]; mul.ftz.f32 %f2042, %f407, %f2034; mul.ftz.f32 %f2043, %f407, %f2035; mul.ftz.f32 %f2044, %f407, %f2036; mul.ftz.f32 %f2045, %f407, %f2037; fma.rn.ftz.f32 %f430, %f417, %f2026, %f2042; fma.rn.ftz.f32 %f431, %f417, %f2027, %f2043; fma.rn.ftz.f32 %f432, %f417, %f2028, %f2044; fma.rn.ftz.f32 %f2046, %f417, %f2029, %f2045; ld.global.v4.f32 {%f2047, %f2048, %f2049, %f2050}, [%rd107+32]; ld.global.v4.f32 {%f2055, %f2056, %f2057, %f2058}, [%rd107+96]; mul.ftz.f32 %f2063, %f407, %f2055; mul.ftz.f32 %f2064, %f407, %f2056; mul.ftz.f32 %f2065, %f407, %f2057; mul.ftz.f32 %f2066, %f407, %f2058; fma.rn.ftz.f32 %f2067, %f417, %f2047, %f2063; fma.rn.ftz.f32 %f2068, %f417, %f2048, %f2064; fma.rn.ftz.f32 %f2069, %f417, %f2049, %f2065; fma.rn.ftz.f32 %f2070, %f417, %f2050, %f2066; ld.global.v4.f32 {%f2071, %f2072, %f2073, %f2074}, [%rd107+48]; ld.global.v4.f32 {%f2079, %f2080, %f2081, %f2082}, [%rd107+112]; mul.ftz.f32 %f2087, %f407, %f2079; mul.ftz.f32 %f2088, %f407, %f2080; mul.ftz.f32 %f2089, %f407, %f2081; mul.ftz.f32 %f2090, %f407, %f2082; fma.rn.ftz.f32 %f2091, %f417, %f2071, %f2087; fma.rn.ftz.f32 %f2092, %f417, %f2072, %f2088; fma.rn.ftz.f32 %f2093, %f417, %f2073, %f2089; fma.rn.ftz.f32 %f2094, %f417, %f2074, %f2090; mul.ftz.f32 %f2095, %f3434, 0f3FB504F3; mul.ftz.f32 %f2096, %f3431, 0f3FB504F3; mul.ftz.f32 %f2097, %f3432, 0f3FB504F3; mul.ftz.f32 %f2098, %f3433, 0f3FB504F3; mul.ftz.f32 %f2099, %f2095, %f2096; mul.ftz.f32 %f2100, %f2095, %f2097; mul.ftz.f32 %f2101, %f2095, %f2098; mul.ftz.f32 %f2102, %f2096, %f2096; mul.ftz.f32 %f2103, %f2096, %f2097; mul.ftz.f32 %f2104, %f2096, %f2098; mul.ftz.f32 %f2105, %f2097, %f2097; mul.ftz.f32 %f2106, %f2097, %f2098; mul.ftz.f32 %f2107, %f2098, %f2098; sub.ftz.f32 %f2109, %f1958, %f2105; sub.ftz.f32 %f2110, %f2109, %f2107; sub.ftz.f32 %f2111, %f2103, %f2101; add.ftz.f32 %f2112, %f2100, %f2104; add.ftz.f32 %f2113, %f2101, %f2103; sub.ftz.f32 %f2114, %f1958, %f2102; sub.ftz.f32 %f2115, %f2114, %f2107; sub.ftz.f32 %f2116, %f2106, %f2099; sub.ftz.f32 %f2117, %f2104, %f2100; add.ftz.f32 %f2118, %f2099, %f2106; sub.ftz.f32 %f2119, %f2114, %f2105; mul.ftz.f32 %f2120, %f2111, %f2069; fma.rn.ftz.f32 %f2121, %f2110, %f2046, %f2120; fma.rn.ftz.f32 %f3440, %f2112, %f2092, %f2121; mul.ftz.f32 %f2122, %f2111, %f2070; fma.rn.ftz.f32 %f2123, %f2110, %f2067, %f2122; fma.rn.ftz.f32 %f434, %f2112, %f2093, %f2123; mul.ftz.f32 %f2124, %f2111, %f2091; fma.rn.ftz.f32 %f2125, %f2110, %f2068, %f2124; fma.rn.ftz.f32 %f435, %f2112, %f2094, %f2125; mul.ftz.f32 %f2126, %f2115, %f2069; fma.rn.ftz.f32 %f2127, %f2113, %f2046, %f2126; fma.rn.ftz.f32 %f436, %f2116, %f2092, %f2127; mul.ftz.f32 %f2128, %f2115, %f2070; fma.rn.ftz.f32 %f2129, %f2113, %f2067, %f2128; fma.rn.ftz.f32 %f3439, %f2116, %f2093, %f2129; mul.ftz.f32 %f2130, %f2115, %f2091; fma.rn.ftz.f32 %f2131, %f2113, %f2068, %f2130; fma.rn.ftz.f32 %f438, %f2116, %f2094, %f2131; mul.ftz.f32 %f2132, %f2118, %f2069; fma.rn.ftz.f32 %f2133, %f2117, %f2046, %f2132; fma.rn.ftz.f32 %f439, %f2119, %f2092, %f2133; mul.ftz.f32 %f2134, %f2118, %f2070; fma.rn.ftz.f32 %f2135, %f2117, %f2067, %f2134; fma.rn.ftz.f32 %f440, %f2119, %f2093, %f2135; mul.ftz.f32 %f2136, %f2118, %f2091; fma.rn.ftz.f32 %f2137, %f2117, %f2068, %f2136; fma.rn.ftz.f32 %f3438, %f2119, %f2094, %f2137; add.s64 %rd312, %rd290, 144; cvta.to.local.u64 %rd313, %rd312; st.local.v4.f32 [%rd313], {%f3440, %f434, %f435, %f430}; add.s64 %rd314, %rd290, 160; cvta.to.local.u64 %rd315, %rd314; st.local.v4.f32 [%rd315], {%f436, %f3439, %f438, %f431}; add.s64 %rd316, %rd290, 176; cvta.to.local.u64 %rd317, %rd316; st.local.v4.f32 [%rd317], {%f439, %f440, %f3438, %f432}; mul.ftz.f32 %f2138, %f3439, %f3438; mul.ftz.f32 %f442, %f440, %f438; sub.ftz.f32 %f3437, %f2138, %f442; mul.ftz.f32 %f444, %f440, %f435; mul.ftz.f32 %f2139, %f434, %f3438; sub.ftz.f32 %f3436, %f444, %f2139; mul.ftz.f32 %f446, %f434, %f438; mul.ftz.f32 %f2140, %f3439, %f435; sub.ftz.f32 %f3435, %f446, %f2140; mul.ftz.f32 %f2141, %f436, %f3436; fma.rn.ftz.f32 %f2142, %f3440, %f3437, %f2141; fma.rn.ftz.f32 %f3441, %f439, %f3435, %f2142; setp.neu.ftz.f32 %p176, %f3441, 0f00000000; @%p176 bra $L__BB208_124; add.ftz.f32 %f3440, %f3440, 0f322BCC77; add.ftz.f32 %f3438, %f3438, 0f322BCC77; add.ftz.f32 %f3439, %f3439, 0f322BCC77; mul.ftz.f32 %f2143, %f3439, %f3438; sub.ftz.f32 %f3437, %f2143, %f442; mul.ftz.f32 %f2144, %f434, %f3438; sub.ftz.f32 %f3436, %f444, %f2144; mul.ftz.f32 %f2145, %f3439, %f435; sub.ftz.f32 %f3435, %f446, %f2145; mul.ftz.f32 %f2146, %f436, %f3436; fma.rn.ftz.f32 %f2147, %f3440, %f3437, %f2146; fma.rn.ftz.f32 %f2148, %f439, %f3435, %f2147; setp.eq.ftz.f32 %p177, %f2148, 0f00000000; selp.f32 %f3441, 0f7F7FFFFF, %f2148, %p177; $L__BB208_124: mul.ftz.f32 %f2149, %f436, %f3438; mul.ftz.f32 %f2150, %f439, %f438; sub.ftz.f32 %f2151, %f2150, %f2149; mul.ftz.f32 %f2152, %f439, %f435; mul.ftz.f32 %f2153, %f3438, %f3440; sub.ftz.f32 %f2154, %f2153, %f2152; mul.ftz.f32 %f2155, %f438, %f3440; mul.ftz.f32 %f2156, %f436, %f435; sub.ftz.f32 %f2157, %f2156, %f2155; mul.ftz.f32 %f2158, %f439, %f3439; mul.ftz.f32 %f2159, %f436, %f440; sub.ftz.f32 %f2160, %f2159, %f2158; mul.ftz.f32 %f2161, %f440, %f3440; mul.ftz.f32 %f2162, %f439, %f434; sub.ftz.f32 %f2163, %f2162, %f2161; mul.ftz.f32 %f2164, %f436, %f434; mul.ftz.f32 %f2165, %f3439, %f3440; sub.ftz.f32 %f2166, %f2165, %f2164; div.approx.ftz.f32 %f2167, %f3436, %f3441; mul.ftz.f32 %f2168, %f431, %f2167; div.approx.ftz.f32 %f2169, %f3437, %f3441; fma.rn.ftz.f32 %f2170, %f430, %f2169, %f2168; div.approx.ftz.f32 %f2171, %f3435, %f3441; fma.rn.ftz.f32 %f2172, %f432, %f2171, %f2170; div.approx.ftz.f32 %f2173, %f2154, %f3441; mul.ftz.f32 %f2174, %f431, %f2173; div.approx.ftz.f32 %f2175, %f2151, %f3441; fma.rn.ftz.f32 %f2176, %f430, %f2175, %f2174; div.approx.ftz.f32 %f2177, %f2157, %f3441; fma.rn.ftz.f32 %f2178, %f432, %f2177, %f2176; div.approx.ftz.f32 %f2179, %f2163, %f3441; mul.ftz.f32 %f2180, %f431, %f2179; div.approx.ftz.f32 %f2181, %f2160, %f3441; fma.rn.ftz.f32 %f2182, %f430, %f2181, %f2180; div.approx.ftz.f32 %f2183, %f2166, %f3441; fma.rn.ftz.f32 %f2184, %f432, %f2183, %f2182; neg.ftz.f32 %f2185, %f2172; add.s64 %rd319, %rd290, 192; cvta.to.local.u64 %rd320, %rd319; st.local.v4.f32 [%rd320], {%f2169, %f2167, %f2171, %f2185}; neg.ftz.f32 %f2186, %f2178; add.s64 %rd321, %rd290, 208; cvta.to.local.u64 %rd322, %rd321; st.local.v4.f32 [%rd322], {%f2175, %f2173, %f2177, %f2186}; neg.ftz.f32 %f2187, %f2184; add.s64 %rd323, %rd290, 224; cvta.to.local.u64 %rd324, %rd323; st.local.v4.f32 [%rd324], {%f2181, %f2179, %f2183, %f2187}; $L__BB208_125: add.s64 %rd326, %rd290, 36; cvta.to.local.u64 %rd327, %rd326; mov.u32 %r409, -2147483648; st.local.u32 [%rd327], %r409; add.s64 %rd328, %rd290, 40; cvta.to.local.u64 %rd329, %rd328; st.local.u32 [%rd329], %r409; add.s64 %rd330, %rd290, 44; cvta.to.local.u64 %rd331, %rd330; mov.u32 %r410, -1082130432; st.local.u32 [%rd331], %r410; and.b32 %r411, %r583, 6; setp.eq.s32 %p178, %r411, 0; @%p178 bra $L__BB208_163; and.b32 %r197, %r641, 4; setp.ne.s32 %p179, %r197, 0; mov.f32 %f3460, 0f3F800000; mov.f32 %f3458, 0f00000000; mov.f32 %f3457, %f3337; mov.f32 %f3459, %f3458; mov.f32 %f3461, %f1; mov.f32 %f3462, %f2; mov.f32 %f3463, %f3; @%p179 bra $L__BB208_132; @%p172 bra $L__BB208_129; bra.uni $L__BB208_128; $L__BB208_129: ld.global.v4.f32 {%f3442, %f3443, %f3444, %f3445}, [%rd106+48]; ld.global.v4.f32 {%f3446, %f3447, %f3448, %f3449}, [%rd106+64]; ld.global.v4.f32 {%f3450, %f3451, %f3452, %f3453}, [%rd106+80]; bra.uni $L__BB208_130; $L__BB208_163: and.b32 %r438, %r583, 8; setp.eq.s32 %p203, %r438, 0; @%p203 bra $L__BB208_184; bra.uni $L__BB208_164; $L__BB208_184: setp.eq.s32 %p216, %r583, 1; @%p216 bra $L__BB208_186; bra.uni $L__BB208_185; $L__BB208_186: cvt.s64.s32 %rd125, %r630; mul.wide.s32 %rd493, %r630, 16; add.s64 %rd494, %rd14, %rd493; add.s64 %rd126, %rd494, 12; ld.global.u32 %r455, [%rd494+12]; mul.wide.u32 %rd495, %r455, 12; add.s64 %rd496, %rd17, %rd495; add.s32 %r456, %r455, 1; mul.wide.u32 %rd497, %r456, 12; add.s64 %rd498, %rd17, %rd497; add.s32 %r457, %r455, 2; mul.wide.u32 %rd499, %r457, 12; add.s64 %rd500, %rd17, %rd499; and.b32 %r458, %r641, 8; setp.eq.s32 %p217, %r458, 0; ld.global.f32 %f2745, [%rd500]; ld.global.f32 %f737, [%rd496]; sub.ftz.f32 %f738, %f2745, %f737; ld.global.f32 %f2746, [%rd500+4]; ld.global.f32 %f739, [%rd496+4]; sub.ftz.f32 %f740, %f2746, %f739; ld.global.f32 %f2747, [%rd500+8]; ld.global.f32 %f741, [%rd496+8]; sub.ftz.f32 %f742, %f2747, %f741; ld.global.f32 %f2748, [%rd498]; sub.ftz.f32 %f743, %f2748, %f737; ld.global.f32 %f2749, [%rd498+4]; sub.ftz.f32 %f744, %f2749, %f739; ld.global.f32 %f2750, [%rd498+8]; sub.ftz.f32 %f745, %f2750, %f741; mul.ftz.f32 %f746, %f745, %f740; mul.ftz.f32 %f747, %f744, %f742; mul.ftz.f32 %f748, %f743, %f742; mul.ftz.f32 %f749, %f745, %f738; mul.ftz.f32 %f750, %f744, %f738; mul.ftz.f32 %f751, %f743, %f740; @%p217 bra $L__BB208_188; bra.uni $L__BB208_187; $L__BB208_188: sub.ftz.f32 %f2758, %f747, %f746; sub.ftz.f32 %f2759, %f749, %f748; mul.ftz.f32 %f2760, %f2759, %f2759; fma.rn.ftz.f32 %f2761, %f2758, %f2758, %f2760; sub.ftz.f32 %f2762, %f751, %f750; fma.rn.ftz.f32 %f2763, %f2762, %f2762, %f2761; sqrt.approx.ftz.f32 %f2764, %f2763; div.approx.ftz.f32 %f3528, %f2762, %f2764; div.approx.ftz.f32 %f3527, %f2759, %f2764; div.approx.ftz.f32 %f3526, %f2758, %f2764; bra.uni $L__BB208_189; $L__BB208_164: cvt.s64.s32 %rd115, %r630; mul.wide.s32 %rd446, %r630, 4; add.s64 %rd447, %rd445, %rd446; ld.global.u32 %r439, [%rd447]; add.s64 %rd449, %rd290, 48; cvta.to.local.u64 %rd450, %rd449; st.local.u32 [%rd450], %r439; fma.rn.ftz.f32 %f700, %f3337, 0f00000000, %f1; fma.rn.ftz.f32 %f701, %f3337, 0f00000000, %f2; cvta.to.local.u64 %rd451, %rd290; st.local.v2.f32 [%rd451], {%f700, %f701}; or.b64 %rd452, %rd290, 8; cvta.to.local.u64 %rd453, %rd452; add.ftz.f32 %f702, %f3, %f3337; st.local.f32 [%rd453], %f702; and.b32 %r440, %r583, 64; setp.eq.s32 %p204, %r440, 0; @%p204 bra $L__BB208_178; bra.uni $L__BB208_165; $L__BB208_178: shl.b64 %rd464, %rd115, 4; add.s64 %rd465, %rd28, %rd464; ld.global.v4.f32 {%f3523, %f3524, %f3525, %f2670}, [%rd465]; bra.uni $L__BB208_179; $L__BB208_128: add.s64 %rd333, %rd290, 192; cvta.to.local.u64 %rd334, %rd333; ld.local.v4.f32 {%f3442, %f3443, %f3444, %f3445}, [%rd334]; add.s64 %rd335, %rd290, 208; cvta.to.local.u64 %rd336, %rd335; ld.local.v4.f32 {%f3446, %f3447, %f3448, %f3449}, [%rd336]; add.s64 %rd337, %rd290, 224; cvta.to.local.u64 %rd338, %rd337; ld.local.v4.f32 {%f3450, %f3451, %f3452, %f3453}, [%rd338]; $L__BB208_130: mul.ftz.f32 %f2215, %f2, %f3443; fma.rn.ftz.f32 %f2216, %f1, %f3442, %f2215; fma.rn.ftz.f32 %f2217, %f3, %f3444, %f2216; add.ftz.f32 %f3461, %f3445, %f2217; mul.ftz.f32 %f2218, %f2, %f3447; fma.rn.ftz.f32 %f2219, %f1, %f3446, %f2218; fma.rn.ftz.f32 %f2220, %f3, %f3448, %f2219; add.ftz.f32 %f3462, %f3449, %f2220; mul.ftz.f32 %f2221, %f2, %f3451; fma.rn.ftz.f32 %f2222, %f1, %f3450, %f2221; fma.rn.ftz.f32 %f2223, %f3, %f3452, %f2222; add.ftz.f32 %f3463, %f3453, %f2223; mul.ftz.f32 %f2224, %f3337, 0f00000000; mul.ftz.f32 %f2225, %f2224, %f3443; fma.rn.ftz.f32 %f2226, %f2224, %f3442, %f2225; fma.rn.ftz.f32 %f3458, %f3337, %f3444, %f2226; mul.ftz.f32 %f2227, %f2224, %f3447; fma.rn.ftz.f32 %f2228, %f2224, %f3446, %f2227; fma.rn.ftz.f32 %f3459, %f3337, %f3448, %f2228; mul.ftz.f32 %f2229, %f2224, %f3451; fma.rn.ftz.f32 %f2230, %f2224, %f3450, %f2229; fma.rn.ftz.f32 %f3460, %f3337, %f3452, %f2230; mul.ftz.f32 %f2231, %f3459, %f3459; fma.rn.ftz.f32 %f2232, %f3458, %f3458, %f2231; fma.rn.ftz.f32 %f2233, %f3460, %f3460, %f2232; sqrt.approx.ftz.f32 %f3457, %f2233; setp.neu.ftz.f32 %p181, %f3457, 0f00000000; @%p181 bra $L__BB208_131; bra.uni $L__BB208_132; $L__BB208_131: div.approx.ftz.f32 %f3460, %f3460, %f3457; div.approx.ftz.f32 %f3459, %f3459, %f3457; div.approx.ftz.f32 %f3458, %f3458, %f3457; $L__BB208_132: cvta.to.global.u64 %rd851, %rd175; mul.wide.s32 %rd340, %r630, 16; add.s64 %rd341, %rd851, %rd340; shr.s32 %r412, %r583, 7; ld.global.u32 %r413, [%rd341+4]; add.s32 %r198, %r412, %r413; add.s32 %r414, %r198, -1; max.s32 %r199, %r414, %r413; add.s32 %r415, %r198, 2; ld.global.u32 %r416, [%rd341+8]; add.s32 %r417, %r413, %r416; add.s32 %r418, %r417, -1; min.s32 %r200, %r415, %r418; and.b32 %r419, %r583, 64; setp.eq.s32 %p182, %r419, 0; mul.wide.s32 %rd342, %r199, 16; add.s64 %rd109, %rd29, %rd342; mul.wide.s32 %rd343, %r198, 16; add.s64 %rd110, %rd29, %rd343; mul.wide.s32 %rd344, %r200, 16; add.s64 %rd111, %rd29, %rd344; @%p182 bra $L__BB208_146; bra.uni $L__BB208_133; $L__BB208_146: ld.global.v4.f32 {%f3488, %f3489, %f3490, %f2318}, [%rd109]; ld.global.v4.f32 {%f3491, %f3492, %f3493, %f2322}, [%rd110]; ld.global.v4.f32 {%f3494, %f3495, %f3496, %f2326}, [%rd110+16]; ld.global.v4.f32 {%f3497, %f3498, %f3499, %f2330}, [%rd111]; bra.uni $L__BB208_147; $L__BB208_133: ld.global.u32 %r201, [%rd106+148]; ld.global.u32 %r202, [%rd106+152]; shl.b32 %r420, %r202, 1; cvt.rn.f32.s32 %f2234, %r420; mul.ftz.f32 %f525, %f4, %f2234; cvt.rzi.ftz.s32.f32 %r421, %f525; add.s32 %r422, %r420, -1; min.s32 %r203, %r421, %r422; ld.global.u32 %r632, [%rd106+164]; mul.wide.u32 %rd345, %r632, 16; add.s64 %rd346, %rd23, %rd345; ld.global.u16 %rs42, [%rd346+12]; ld.global.u32 %r631, [%rd346+8]; ld.global.u64 %rd876, [%rd346]; setp.eq.s64 %p183, %rd876, 11; @%p183 bra $L__BB208_138; $L__BB208_134: setp.eq.s64 %p184, %rd876, 0; @%p184 bra $L__BB208_136; add.s32 %r632, %r632, 2; bra.uni $L__BB208_137; $L__BB208_136: mov.u32 %r635, -1; setp.eq.s16 %p185, %rs42, 0; mov.u32 %r632, %r631; @%p185 bra $L__BB208_139; $L__BB208_137: mul.wide.u32 %rd347, %r632, 16; add.s64 %rd348, %rd23, %rd347; ld.global.u16 %rs42, [%rd348+12]; ld.global.u32 %r631, [%rd348+8]; ld.global.u64 %rd876, [%rd348]; setp.ne.s64 %p186, %rd876, 11; @%p186 bra $L__BB208_134; $L__BB208_138: setp.eq.s16 %p187, %rs42, 0; selp.b32 %r635, -1, %r631, %p187; $L__BB208_139: setp.eq.s32 %p188, %r203, %r202; @%p188 bra $L__BB208_141; bra.uni $L__BB208_140; $L__BB208_141: ld.global.v4.f32 {%f3464, %f3465, %f3466, %f2254}, [%rd109]; ld.global.v4.f32 {%f3467, %f3468, %f3469, %f2258}, [%rd110]; ld.global.v4.f32 {%f3470, %f3471, %f3472, %f2262}, [%rd110+16]; ld.global.v4.f32 {%f3473, %f3474, %f3475, %f2266}, [%rd111]; bra.uni $L__BB208_142; $L__BB208_140: shr.s32 %r525, %r583, 7; add.s32 %r524, %r525, %r413; setp.gt.s32 %p189, %r203, %r202; selp.b32 %r424, -1, 0, %p189; add.s32 %r425, %r203, %r424; mad.lo.s32 %r426, %r425, %r201, %r635; add.s32 %r427, %r426, %r199; cvta.to.global.u64 %rd349, %rd182; mul.wide.s32 %rd350, %r427, 16; add.s64 %rd351, %rd349, %rd350; ld.global.v4.f32 {%f3464, %f3465, %f3466, %f2238}, [%rd351]; add.s32 %r428, %r426, %r524; mul.wide.s32 %rd352, %r428, 16; add.s64 %rd353, %rd349, %rd352; ld.global.v4.f32 {%f3467, %f3468, %f3469, %f2242}, [%rd353]; ld.global.v4.f32 {%f3470, %f3471, %f3472, %f2246}, [%rd353+16]; add.s32 %r429, %r426, %r200; mul.wide.s32 %rd354, %r429, 16; add.s64 %rd355, %rd349, %rd354; ld.global.v4.f32 {%f3473, %f3474, %f3475, %f2250}, [%rd355]; $L__BB208_142: add.s32 %r214, %r203, 1; setp.eq.s32 %p190, %r214, %r202; @%p190 bra $L__BB208_144; bra.uni $L__BB208_143; $L__BB208_144: ld.global.v4.f32 {%f3476, %f3477, %f3478, %f2286}, [%rd109]; ld.global.v4.f32 {%f3479, %f3480, %f3481, %f2290}, [%rd110]; ld.global.v4.f32 {%f3482, %f3483, %f3484, %f2294}, [%rd110+16]; ld.global.v4.f32 {%f3485, %f3486, %f3487, %f2298}, [%rd111]; bra.uni $L__BB208_145; $L__BB208_143: shr.s32 %r527, %r583, 7; add.s32 %r526, %r527, %r413; setp.ge.s32 %p191, %r203, %r202; selp.b32 %r430, -1, 0, %p191; add.s32 %r431, %r214, %r430; mad.lo.s32 %r432, %r431, %r201, %r635; add.s32 %r433, %r432, %r199; cvta.to.global.u64 %rd356, %rd182; mul.wide.s32 %rd357, %r433, 16; add.s64 %rd358, %rd356, %rd357; ld.global.v4.f32 {%f3476, %f3477, %f3478, %f2270}, [%rd358]; add.s32 %r434, %r432, %r526; mul.wide.s32 %rd359, %r434, 16; add.s64 %rd360, %rd356, %rd359; ld.global.v4.f32 {%f3479, %f3480, %f3481, %f2274}, [%rd360]; ld.global.v4.f32 {%f3482, %f3483, %f3484, %f2278}, [%rd360+16]; add.s32 %r435, %r432, %r200; mul.wide.s32 %rd361, %r435, 16; add.s64 %rd362, %rd356, %rd361; ld.global.v4.f32 {%f3485, %f3486, %f3487, %f2282}, [%rd362]; $L__BB208_145: shl.b32 %r528, %r202, 1; cvt.rn.f32.s32 %f3213, %r528; mul.ftz.f32 %f3212, %f4, %f3213; cvt.rn.f32.s32 %f2299, %r203; sub.ftz.f32 %f2300, %f3212, %f2299; mov.f32 %f2301, 0f3F800000; sub.ftz.f32 %f2302, %f2301, %f2300; mul.ftz.f32 %f2303, %f2300, %f3476; fma.rn.ftz.f32 %f3488, %f2302, %f3464, %f2303; mul.ftz.f32 %f2304, %f2300, %f3477; fma.rn.ftz.f32 %f3489, %f2302, %f3465, %f2304; mul.ftz.f32 %f2305, %f2300, %f3478; fma.rn.ftz.f32 %f3490, %f2302, %f3466, %f2305; mul.ftz.f32 %f2306, %f2300, %f3479; fma.rn.ftz.f32 %f3491, %f2302, %f3467, %f2306; mul.ftz.f32 %f2307, %f2300, %f3480; fma.rn.ftz.f32 %f3492, %f2302, %f3468, %f2307; mul.ftz.f32 %f2308, %f2300, %f3481; fma.rn.ftz.f32 %f3493, %f2302, %f3469, %f2308; mul.ftz.f32 %f2309, %f2300, %f3482; fma.rn.ftz.f32 %f3494, %f2302, %f3470, %f2309; mul.ftz.f32 %f2310, %f2300, %f3483; fma.rn.ftz.f32 %f3495, %f2302, %f3471, %f2310; mul.ftz.f32 %f2311, %f2300, %f3484; fma.rn.ftz.f32 %f3496, %f2302, %f3472, %f2311; mul.ftz.f32 %f2312, %f2300, %f3485; fma.rn.ftz.f32 %f3497, %f2302, %f3473, %f2312; mul.ftz.f32 %f2313, %f2300, %f3486; fma.rn.ftz.f32 %f3498, %f2302, %f3474, %f2313; mul.ftz.f32 %f2314, %f2300, %f3487; fma.rn.ftz.f32 %f3499, %f2302, %f3475, %f2314; $L__BB208_147: fma.rn.ftz.f32 %f3517, %f3457, %f3458, %f3461; fma.rn.ftz.f32 %f3518, %f3457, %f3459, %f3462; fma.rn.ftz.f32 %f3519, %f3457, %f3460, %f3463; mov.f32 %f3503, 0f3F800000; sub.ftz.f32 %f637, %f3503, %f3336; mul.ftz.f32 %f638, %f637, %f637; add.ftz.f32 %f2332, %f637, %f637; mul.ftz.f32 %f2333, %f3336, %f2332; sub.ftz.f32 %f2334, %f2333, %f638; mul.ftz.f32 %f2335, %f3336, 0f40400000; add.ftz.f32 %f639, %f2335, 0fC0A00000; add.ftz.f32 %f2336, %f3336, %f3336; mul.ftz.f32 %f2337, %f2336, %f639; fma.rn.ftz.f32 %f2338, %f3336, %f2335, %f2337; add.ftz.f32 %f2339, %f2335, 0f40000000; mul.ftz.f32 %f2340, %f2332, %f2339; mul.ftz.f32 %f640, %f637, 0f40400000; mul.ftz.f32 %f2341, %f637, %f640; sub.ftz.f32 %f2342, %f2340, %f2341; mul.ftz.f32 %f2343, %f637, 0fC0000000; mul.ftz.f32 %f641, %f3336, %f3336; fma.rn.ftz.f32 %f2344, %f3336, %f2343, %f641; mul.ftz.f32 %f2345, %f3491, %f2338; mul.ftz.f32 %f2346, %f3492, %f2338; mul.ftz.f32 %f2347, %f3493, %f2338; fma.rn.ftz.f32 %f2348, %f3488, %f2334, %f2345; fma.rn.ftz.f32 %f2349, %f3489, %f2334, %f2346; fma.rn.ftz.f32 %f2350, %f3490, %f2334, %f2347; fma.rn.ftz.f32 %f2351, %f3494, %f2342, %f2348; fma.rn.ftz.f32 %f2352, %f3495, %f2342, %f2349; fma.rn.ftz.f32 %f2353, %f3496, %f2342, %f2350; fma.rn.ftz.f32 %f2354, %f3497, %f2344, %f2351; fma.rn.ftz.f32 %f2355, %f3498, %f2344, %f2352; fma.rn.ftz.f32 %f2356, %f3499, %f2344, %f2353; mul.ftz.f32 %f3512, %f2354, 0f3F000000; mul.ftz.f32 %f3514, %f2355, 0f3F000000; mul.ftz.f32 %f3513, %f2356, 0f3F000000; and.b32 %r215, %r583, 4; setp.eq.s32 %p192, %r215, 0; @%p192 bra $L__BB208_149; bra.uni $L__BB208_148; $L__BB208_149: mul.ftz.f32 %f2418, %f3336, %f637; mul.ftz.f32 %f2419, %f637, %f2418; fma.rn.ftz.f32 %f2420, %f641, %f639, 0f40000000; add.ftz.f32 %f2421, %f640, 0fC0A00000; fma.rn.ftz.f32 %f2422, %f638, %f2421, 0f40000000; mul.ftz.f32 %f2423, %f3336, %f2418; mul.ftz.f32 %f2424, %f3488, %f2419; mul.ftz.f32 %f2425, %f3489, %f2419; mul.ftz.f32 %f2426, %f3490, %f2419; mul.ftz.f32 %f2427, %f3491, %f2420; mul.ftz.f32 %f2428, %f3492, %f2420; mul.ftz.f32 %f2429, %f3493, %f2420; sub.ftz.f32 %f2430, %f2427, %f2424; sub.ftz.f32 %f2431, %f2428, %f2425; sub.ftz.f32 %f2432, %f2429, %f2426; fma.rn.ftz.f32 %f2433, %f3494, %f2422, %f2430; fma.rn.ftz.f32 %f2434, %f3495, %f2422, %f2431; fma.rn.ftz.f32 %f2435, %f3496, %f2422, %f2432; mul.ftz.f32 %f2436, %f3497, %f2423; mul.ftz.f32 %f2437, %f3498, %f2423; mul.ftz.f32 %f2438, %f3499, %f2423; sub.ftz.f32 %f2439, %f2433, %f2436; sub.ftz.f32 %f2440, %f2434, %f2437; sub.ftz.f32 %f2441, %f2435, %f2438; mul.ftz.f32 %f648, %f2439, 0f3F000000; mul.ftz.f32 %f649, %f2440, 0f3F000000; mul.ftz.f32 %f650, %f2441, 0f3F000000; setp.eq.ftz.f32 %p193, %f3517, %f648; setp.eq.ftz.f32 %p194, %f3518, %f649; and.pred %p195, %p193, %p194; setp.eq.ftz.f32 %p196, %f3519, %f650; and.pred %p197, %p196, %p195; mov.f32 %f3516, 0f00000000; mov.f32 %f3515, %f3516; @%p197 bra $L__BB208_151; sub.ftz.f32 %f2442, %f3517, %f648; sub.ftz.f32 %f2443, %f3518, %f649; mul.ftz.f32 %f2444, %f2443, %f2443; fma.rn.ftz.f32 %f2445, %f2442, %f2442, %f2444; sub.ftz.f32 %f2446, %f3519, %f650; fma.rn.ftz.f32 %f2447, %f2446, %f2446, %f2445; sqrt.approx.ftz.f32 %f2448, %f2447; div.approx.ftz.f32 %f3503, %f2446, %f2448; div.approx.ftz.f32 %f3515, %f2443, %f2448; div.approx.ftz.f32 %f3516, %f2442, %f2448; $L__BB208_151: mov.u32 %r531, 0; or.b64 %rd371, %rd290, 12; cvta.to.local.u64 %rd372, %rd371; st.local.f32 [%rd372], %f3516; add.s64 %rd373, %rd290, 16; cvta.to.local.u64 %rd374, %rd373; st.local.f32 [%rd374], %f3515; add.s64 %rd375, %rd290, 20; cvta.to.local.u64 %rd376, %rd375; st.local.f32 [%rd376], %f3503; add.s64 %rd377, %rd290, 72; cvta.to.local.u64 %rd378, %rd377; st.local.u32 [%rd378], %r531; bra.uni $L__BB208_152; $L__BB208_148: mul.ftz.f32 %f2357, %f3514, %f3514; fma.rn.ftz.f32 %f2358, %f3512, %f3512, %f2357; fma.rn.ftz.f32 %f2359, %f3513, %f3513, %f2358; sqrt.approx.ftz.f32 %f2360, %f2359; div.approx.ftz.f32 %f2361, %f3512, %f2360; div.approx.ftz.f32 %f2362, %f3514, %f2360; div.approx.ftz.f32 %f2363, %f3513, %f2360; mul.ftz.f32 %f2364, %f2362, %f3460; mul.ftz.f32 %f2365, %f2363, %f3459; sub.ftz.f32 %f2366, %f2365, %f2364; mul.ftz.f32 %f2367, %f2363, %f3458; mul.ftz.f32 %f2368, %f2361, %f3460; sub.ftz.f32 %f2369, %f2368, %f2367; mul.ftz.f32 %f2370, %f2361, %f3459; mul.ftz.f32 %f2371, %f2362, %f3458; sub.ftz.f32 %f2372, %f2371, %f2370; mul.ftz.f32 %f2373, %f2369, %f2369; fma.rn.ftz.f32 %f2374, %f2366, %f2366, %f2373; fma.rn.ftz.f32 %f2375, %f2372, %f2372, %f2374; sqrt.approx.ftz.f32 %f2376, %f2375; div.approx.ftz.f32 %f2377, %f2366, %f2376; div.approx.ftz.f32 %f2378, %f2369, %f2376; div.approx.ftz.f32 %f2379, %f2372, %f2376; mul.ftz.f32 %f2380, %f3335, %f3335; sub.ftz.f32 %f2382, %f3503, %f2380; mov.f32 %f2383, 0f00000000; max.ftz.f32 %f2384, %f2382, %f2383; sqrt.approx.ftz.f32 %f2385, %f2384; mul.ftz.f32 %f2386, %f2377, %f3335; mul.ftz.f32 %f2387, %f2378, %f3335; mul.ftz.f32 %f2388, %f2379, %f3335; mul.ftz.f32 %f2389, %f2362, %f2379; mul.ftz.f32 %f2390, %f2363, %f2378; sub.ftz.f32 %f2391, %f2389, %f2390; mul.ftz.f32 %f2392, %f2363, %f2377; mul.ftz.f32 %f2393, %f2361, %f2379; sub.ftz.f32 %f2394, %f2392, %f2393; mul.ftz.f32 %f2395, %f2361, %f2378; mul.ftz.f32 %f2396, %f2362, %f2377; sub.ftz.f32 %f2397, %f2395, %f2396; mul.ftz.f32 %f2398, %f2394, %f2394; fma.rn.ftz.f32 %f2399, %f2391, %f2391, %f2398; fma.rn.ftz.f32 %f2400, %f2397, %f2397, %f2399; sqrt.approx.ftz.f32 %f2401, %f2400; div.approx.ftz.f32 %f2402, %f2391, %f2401; div.approx.ftz.f32 %f2403, %f2394, %f2401; div.approx.ftz.f32 %f2404, %f2397, %f2401; mul.ftz.f32 %f2405, %f2385, %f2402; mul.ftz.f32 %f2406, %f2385, %f2403; mul.ftz.f32 %f2407, %f2385, %f2404; sub.ftz.f32 %f2408, %f2386, %f2405; sub.ftz.f32 %f2409, %f2387, %f2406; sub.ftz.f32 %f2410, %f2388, %f2407; mul.ftz.f32 %f2411, %f2409, %f2409; fma.rn.ftz.f32 %f2412, %f2408, %f2408, %f2411; fma.rn.ftz.f32 %f2413, %f2410, %f2410, %f2412; sqrt.approx.ftz.f32 %f2414, %f2413; div.approx.ftz.f32 %f3516, %f2408, %f2414; div.approx.ftz.f32 %f3515, %f2409, %f2414; div.approx.ftz.f32 %f3503, %f2410, %f2414; or.b64 %rd364, %rd290, 12; cvta.to.local.u64 %rd365, %rd364; st.local.f32 [%rd365], %f3516; add.s64 %rd366, %rd290, 16; cvta.to.local.u64 %rd367, %rd366; st.local.f32 [%rd367], %f3515; add.s64 %rd368, %rd290, 20; cvta.to.local.u64 %rd369, %rd368; st.local.f32 [%rd369], %f3503; $L__BB208_152: and.b32 %r529, %r641, 4; setp.ne.s32 %p280, %r529, 0; add.s64 %rd380, %rd290, 116; cvta.to.local.u64 %rd381, %rd380; st.local.f32 [%rd381], %f3512; add.s64 %rd382, %rd290, 120; cvta.to.local.u64 %rd383, %rd382; st.local.f32 [%rd383], %f3514; add.s64 %rd384, %rd290, 124; cvta.to.local.u64 %rd385, %rd384; st.local.f32 [%rd385], %f3513; @%p280 bra $L__BB208_162; @%p172 bra $L__BB208_155; bra.uni $L__BB208_154; $L__BB208_155: ld.global.v4.f32 {%f2482, %f2483, %f2484, %f2485}, [%rd106]; ld.global.v4.f32 {%f2490, %f2491, %f2492, %f2493}, [%rd106+16]; ld.global.v4.f32 {%f2498, %f2499, %f2500, %f2501}, [%rd106+32]; mul.ftz.f32 %f2506, %f3518, %f2483; fma.rn.ftz.f32 %f2507, %f3517, %f2482, %f2506; fma.rn.ftz.f32 %f2508, %f3519, %f2484, %f2507; add.ftz.f32 %f3506, %f2485, %f2508; mul.ftz.f32 %f2509, %f3518, %f2491; fma.rn.ftz.f32 %f2510, %f3517, %f2490, %f2509; fma.rn.ftz.f32 %f2511, %f3519, %f2492, %f2510; add.ftz.f32 %f3507, %f2493, %f2511; mul.ftz.f32 %f2512, %f3518, %f2499; fma.rn.ftz.f32 %f2513, %f3517, %f2498, %f2512; fma.rn.ftz.f32 %f2514, %f3519, %f2500, %f2513; add.ftz.f32 %f3519, %f2501, %f2514; bra.uni $L__BB208_156; $L__BB208_154: add.s64 %rd387, %rd290, 144; cvta.to.local.u64 %rd388, %rd387; ld.local.v4.f32 {%f2449, %f2450, %f2451, %f2452}, [%rd388]; mul.ftz.f32 %f2457, %f3518, %f2450; fma.rn.ftz.f32 %f2458, %f3517, %f2449, %f2457; fma.rn.ftz.f32 %f2459, %f3519, %f2451, %f2458; add.ftz.f32 %f3506, %f2452, %f2459; add.s64 %rd389, %rd290, 160; cvta.to.local.u64 %rd390, %rd389; ld.local.v4.f32 {%f2460, %f2461, %f2462, %f2463}, [%rd390]; mul.ftz.f32 %f2468, %f3518, %f2461; fma.rn.ftz.f32 %f2469, %f3517, %f2460, %f2468; fma.rn.ftz.f32 %f2470, %f3519, %f2462, %f2469; add.ftz.f32 %f3507, %f2463, %f2470; add.s64 %rd391, %rd290, 176; cvta.to.local.u64 %rd392, %rd391; ld.local.v4.f32 {%f2471, %f2472, %f2473, %f2474}, [%rd392]; mul.ftz.f32 %f2479, %f3518, %f2472; fma.rn.ftz.f32 %f2480, %f3517, %f2471, %f2479; fma.rn.ftz.f32 %f2481, %f3519, %f2473, %f2480; add.ftz.f32 %f3519, %f2474, %f2481; $L__BB208_156: @%p172 bra $L__BB208_158; bra.uni $L__BB208_157; $L__BB208_158: ld.global.v4.f32 {%f2549, %f2550, %f2551, %f2552}, [%rd106+48]; ld.global.v4.f32 {%f2556, %f2557, %f2558, %f2559}, [%rd106+64]; ld.global.v4.f32 {%f2563, %f2564, %f2565, %f2566}, [%rd106+80]; mul.ftz.f32 %f2570, %f2556, %f3515; fma.rn.ftz.f32 %f2571, %f2549, %f3516, %f2570; fma.rn.ftz.f32 %f2572, %f2563, %f3503, %f2571; mul.ftz.f32 %f2573, %f2557, %f3515; fma.rn.ftz.f32 %f2574, %f2550, %f3516, %f2573; fma.rn.ftz.f32 %f2575, %f2564, %f3503, %f2574; mul.ftz.f32 %f2576, %f2558, %f3515; fma.rn.ftz.f32 %f2577, %f2551, %f3516, %f2576; fma.rn.ftz.f32 %f2578, %f2565, %f3503, %f2577; mul.ftz.f32 %f2579, %f2575, %f2575; fma.rn.ftz.f32 %f2580, %f2572, %f2572, %f2579; fma.rn.ftz.f32 %f2581, %f2578, %f2578, %f2580; sqrt.approx.ftz.f32 %f2582, %f2581; div.approx.ftz.f32 %f3516, %f2572, %f2582; div.approx.ftz.f32 %f3515, %f2575, %f2582; div.approx.ftz.f32 %f3509, %f2578, %f2582; bra.uni $L__BB208_159; $L__BB208_157: add.s64 %rd394, %rd290, 192; cvta.to.local.u64 %rd395, %rd394; ld.local.v4.f32 {%f2515, %f2516, %f2517, %f2518}, [%rd395]; add.s64 %rd396, %rd290, 208; cvta.to.local.u64 %rd397, %rd396; ld.local.v4.f32 {%f2522, %f2523, %f2524, %f2525}, [%rd397]; add.s64 %rd398, %rd290, 224; cvta.to.local.u64 %rd399, %rd398; ld.local.v4.f32 {%f2529, %f2530, %f2531, %f2532}, [%rd399]; mul.ftz.f32 %f2536, %f3515, %f2522; fma.rn.ftz.f32 %f2537, %f3516, %f2515, %f2536; fma.rn.ftz.f32 %f2538, %f3503, %f2529, %f2537; mul.ftz.f32 %f2539, %f3515, %f2523; fma.rn.ftz.f32 %f2540, %f3516, %f2516, %f2539; fma.rn.ftz.f32 %f2541, %f3503, %f2530, %f2540; mul.ftz.f32 %f2542, %f3515, %f2524; fma.rn.ftz.f32 %f2543, %f3516, %f2517, %f2542; fma.rn.ftz.f32 %f2544, %f3503, %f2531, %f2543; mul.ftz.f32 %f2545, %f2541, %f2541; fma.rn.ftz.f32 %f2546, %f2538, %f2538, %f2545; fma.rn.ftz.f32 %f2547, %f2544, %f2544, %f2546; sqrt.approx.ftz.f32 %f2548, %f2547; div.approx.ftz.f32 %f3516, %f2538, %f2548; div.approx.ftz.f32 %f3515, %f2541, %f2548; div.approx.ftz.f32 %f3509, %f2544, %f2548; $L__BB208_159: or.b64 %rd401, %rd290, 12; cvta.to.local.u64 %rd402, %rd401; st.local.f32 [%rd402], %f3516; add.s64 %rd403, %rd290, 16; cvta.to.local.u64 %rd404, %rd403; st.local.f32 [%rd404], %f3515; add.s64 %rd405, %rd290, 20; cvta.to.local.u64 %rd406, %rd405; st.local.f32 [%rd406], %f3509; @%p172 bra $L__BB208_161; bra.uni $L__BB208_160; $L__BB208_161: ld.global.v4.f32 {%f2610, %f2611, %f2612, %f2613}, [%rd106]; ld.global.v4.f32 {%f2617, %f2618, %f2619, %f2620}, [%rd106+16]; ld.global.v4.f32 {%f2624, %f2625, %f2626, %f2627}, [%rd106+32]; mul.ftz.f32 %f2631, %f2611, %f3514; fma.rn.ftz.f32 %f2632, %f3512, %f2610, %f2631; fma.rn.ftz.f32 %f684, %f2612, %f3513, %f2632; mul.ftz.f32 %f2633, %f2618, %f3514; fma.rn.ftz.f32 %f2634, %f3512, %f2617, %f2633; fma.rn.ftz.f32 %f685, %f2619, %f3513, %f2634; mul.ftz.f32 %f2635, %f2625, %f3514; fma.rn.ftz.f32 %f2636, %f3512, %f2624, %f2635; fma.rn.ftz.f32 %f3513, %f2626, %f3513, %f2636; st.local.f32 [%rd381], %f684; st.local.f32 [%rd383], %f685; st.local.f32 [%rd385], %f3513; mov.f32 %f3512, %f684; mov.f32 %f3514, %f685; mov.f32 %f3517, %f3506; mov.f32 %f3518, %f3507; bra.uni $L__BB208_162; $L__BB208_160: add.s64 %rd408, %rd290, 144; cvta.to.local.u64 %rd409, %rd408; ld.local.v4.f32 {%f2583, %f2584, %f2585, %f2586}, [%rd409]; mul.ftz.f32 %f2590, %f3514, %f2584; fma.rn.ftz.f32 %f2591, %f3512, %f2583, %f2590; fma.rn.ftz.f32 %f681, %f3513, %f2585, %f2591; add.s64 %rd410, %rd290, 160; cvta.to.local.u64 %rd411, %rd410; ld.local.v4.f32 {%f2592, %f2593, %f2594, %f2595}, [%rd411]; mul.ftz.f32 %f2599, %f3514, %f2593; fma.rn.ftz.f32 %f2600, %f3512, %f2592, %f2599; fma.rn.ftz.f32 %f682, %f3513, %f2594, %f2600; add.s64 %rd412, %rd290, 176; cvta.to.local.u64 %rd413, %rd412; ld.local.v4.f32 {%f2601, %f2602, %f2603, %f2604}, [%rd413]; mul.ftz.f32 %f2608, %f3514, %f2602; fma.rn.ftz.f32 %f2609, %f3512, %f2601, %f2608; fma.rn.ftz.f32 %f3513, %f3513, %f2603, %f2609; st.local.f32 [%rd381], %f681; st.local.f32 [%rd383], %f682; st.local.f32 [%rd385], %f3513; mov.f32 %f3512, %f681; mov.f32 %f3514, %f682; mov.f32 %f3517, %f3506; mov.f32 %f3518, %f3507; $L__BB208_162: and.b32 %r532, %r583, 4; setp.eq.s32 %p281, %r532, 0; mul.wide.s32 %rd842, %r630, 16; cvta.to.global.u64 %rd841, %rd175; add.s64 %rd840, %rd841, %rd842; add.s64 %rd839, %rd840, 4; add.s64 %rd838, %rd290, 44; add.s64 %rd428, %rd290, 20; cvta.to.local.u64 %rd430, %rd290; st.local.v2.f32 [%rd430], {%f3517, %f3518}; or.b64 %rd431, %rd290, 8; cvta.to.local.u64 %rd432, %rd431; st.local.f32 [%rd432], %f3519; selp.f32 %f3556, %f3516, 0f80000000, %p281; selp.f32 %f3557, %f3515, 0f80000000, %p281; selp.b64 %rd433, %rd428, %rd838, %p281; cvta.to.local.u64 %rd434, %rd433; ld.local.f32 %f3558, [%rd434]; add.s64 %rd435, %rd290, 24; cvta.to.local.u64 %rd436, %rd435; st.local.v2.f32 [%rd436], {%f3556, %f3557}; add.s64 %rd437, %rd290, 32; cvta.to.local.u64 %rd438, %rd437; st.local.f32 [%rd438], %f3558; mul.ftz.f32 %f2637, %f3514, %f3558; mul.ftz.f32 %f2638, %f3513, %f3557; mul.ftz.f32 %f2639, %f3513, %f3556; mul.ftz.f32 %f2640, %f3512, %f3558; mul.ftz.f32 %f2641, %f3512, %f3557; mul.ftz.f32 %f2642, %f3514, %f3556; sub.ftz.f32 %f2643, %f2641, %f2642; sub.ftz.f32 %f2644, %f2639, %f2640; sub.ftz.f32 %f2645, %f2637, %f2638; add.s64 %rd439, %rd290, 128; cvta.to.local.u64 %rd440, %rd439; st.local.v2.f32 [%rd440], {%f2645, %f2644}; add.s64 %rd441, %rd290, 136; cvta.to.local.u64 %rd442, %rd441; st.local.f32 [%rd442], %f2643; ld.global.u32 %r437, [%rd839+-4]; add.s64 %rd443, %rd290, 48; cvta.to.local.u64 %rd444, %rd443; st.local.u32 [%rd444], %r437; bra.uni $L__BB208_213; $L__BB208_185: add.s64 %rd856, %rd290, 56; cvta.to.local.u64 %rd855, %rd856; { // callseq 979, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd290; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 979 ld.local.u32 %r641, [%rd855]; bra.uni $L__BB208_199; $L__BB208_165: ld.global.u32 %r216, [%rd106+148]; ld.global.u32 %r217, [%rd106+152]; shl.b32 %r441, %r217, 1; cvt.rn.f32.s32 %f2646, %r441; mul.ftz.f32 %f703, %f4, %f2646; cvt.rzi.ftz.s32.f32 %r442, %f703; add.s32 %r443, %r441, -1; min.s32 %r218, %r442, %r443; ld.global.u32 %r637, [%rd106+164]; mul.wide.u32 %rd454, %r637, 16; add.s64 %rd455, %rd23, %rd454; ld.global.u16 %rs44, [%rd455+12]; ld.global.u32 %r636, [%rd455+8]; ld.global.u64 %rd877, [%rd455]; setp.eq.s64 %p205, %rd877, 11; @%p205 bra $L__BB208_170; $L__BB208_166: setp.eq.s64 %p206, %rd877, 0; @%p206 bra $L__BB208_168; add.s32 %r637, %r637, 2; bra.uni $L__BB208_169; $L__BB208_168: setp.eq.s16 %p207, %rs44, 0; mov.u32 %r640, -1; mov.u32 %r637, %r636; @%p207 bra $L__BB208_171; $L__BB208_169: mul.wide.u32 %rd456, %r637, 16; add.s64 %rd457, %rd23, %rd456; ld.global.u16 %rs44, [%rd457+12]; ld.global.u32 %r636, [%rd457+8]; ld.global.u64 %rd877, [%rd457]; setp.ne.s64 %p208, %rd877, 11; @%p208 bra $L__BB208_166; $L__BB208_170: setp.eq.s16 %p209, %rs44, 0; selp.b32 %r640, -1, %r636, %p209; $L__BB208_171: setp.eq.s32 %p210, %r218, %r217; @%p210 bra $L__BB208_173; bra.uni $L__BB208_172; $L__BB208_173: shl.b64 %rd460, %rd115, 4; add.s64 %rd878, %rd28, %rd460; bra.uni $L__BB208_174; $L__BB208_187: sub.ftz.f32 %f2751, %f746, %f747; sub.ftz.f32 %f2752, %f748, %f749; mul.ftz.f32 %f2753, %f2752, %f2752; fma.rn.ftz.f32 %f2754, %f2751, %f2751, %f2753; sub.ftz.f32 %f2755, %f750, %f751; fma.rn.ftz.f32 %f2756, %f2755, %f2755, %f2754; sqrt.approx.ftz.f32 %f2757, %f2756; div.approx.ftz.f32 %f3528, %f2755, %f2757; div.approx.ftz.f32 %f3527, %f2752, %f2757; div.approx.ftz.f32 %f3526, %f2751, %f2757; $L__BB208_189: shl.b64 %rd501, %rd125, 2; add.s64 %rd502, %rd30, %rd501; ld.global.u32 %r231, [%rd502]; add.s64 %rd504, %rd290, 48; cvta.to.local.u64 %rd505, %rd504; st.local.u32 [%rd505], %r231; fma.rn.ftz.f32 %f2765, %f3336, %f743, %f737; fma.rn.ftz.f32 %f2766, %f3336, %f744, %f739; fma.rn.ftz.f32 %f2767, %f3336, %f745, %f741; fma.rn.ftz.f32 %f3541, %f3335, %f738, %f2765; fma.rn.ftz.f32 %f3542, %f3335, %f740, %f2766; fma.rn.ftz.f32 %f3543, %f3335, %f742, %f2767; and.b32 %r459, %r641, 4; setp.ne.s32 %p218, %r459, 0; @%p218 bra $L__BB208_194; @%p172 bra $L__BB208_192; bra.uni $L__BB208_191; $L__BB208_192: ld.global.v4.f32 {%f3529, %f3530, %f3531, %f3532}, [%rd106]; ld.global.v4.f32 {%f3533, %f3534, %f3535, %f3536}, [%rd106+16]; ld.global.v4.f32 {%f3537, %f3538, %f3539, %f3540}, [%rd106+32]; bra.uni $L__BB208_193; $L__BB208_172: setp.gt.s32 %p211, %r218, %r217; selp.b32 %r445, -1, 0, %p211; add.s32 %r446, %r218, %r445; mad.lo.s32 %r447, %r446, %r216, %r630; add.s32 %r448, %r447, %r640; cvta.to.global.u64 %rd458, %rd182; mul.wide.s32 %rd459, %r448, 16; add.s64 %rd878, %rd458, %rd459; $L__BB208_174: ld.global.v4.f32 {%f2647, %f2648, %f2649, %f2650}, [%rd878]; add.s32 %r229, %r218, 1; setp.eq.s32 %p212, %r229, %r217; @%p212 bra $L__BB208_176; bra.uni $L__BB208_175; $L__BB208_176: shl.b64 %rd463, %rd115, 4; add.s64 %rd879, %rd28, %rd463; bra.uni $L__BB208_177; $L__BB208_175: setp.ge.s32 %p213, %r218, %r217; selp.b32 %r449, -1, 0, %p213; add.s32 %r450, %r229, %r449; mad.lo.s32 %r451, %r450, %r216, %r630; add.s32 %r452, %r451, %r640; cvta.to.global.u64 %rd461, %rd182; mul.wide.s32 %rd462, %r452, 16; add.s64 %rd879, %rd461, %rd462; $L__BB208_177: cvt.rn.f32.s32 %f2651, %r218; sub.ftz.f32 %f2652, %f703, %f2651; ld.global.v4.f32 {%f2653, %f2654, %f2655, %f2656}, [%rd879]; mov.f32 %f2661, 0f3F800000; sub.ftz.f32 %f2662, %f2661, %f2652; mul.ftz.f32 %f2663, %f2652, %f2653; mul.ftz.f32 %f2664, %f2652, %f2654; mul.ftz.f32 %f2665, %f2652, %f2655; fma.rn.ftz.f32 %f3525, %f2662, %f2649, %f2665; fma.rn.ftz.f32 %f3524, %f2662, %f2648, %f2664; fma.rn.ftz.f32 %f3523, %f2662, %f2647, %f2663; $L__BB208_179: and.b32 %r453, %r641, 4; setp.ne.s32 %p214, %r453, 0; @%p214 bra $L__BB208_183; @%p172 bra $L__BB208_182; bra.uni $L__BB208_181; $L__BB208_182: ld.global.v4.f32 {%f2704, %f2705, %f2706, %f2707}, [%rd106]; ld.global.v4.f32 {%f2712, %f2713, %f2714, %f2715}, [%rd106+16]; ld.global.v4.f32 {%f2720, %f2721, %f2722, %f2723}, [%rd106+32]; mul.ftz.f32 %f2728, %f3524, %f2705; fma.rn.ftz.f32 %f2729, %f3523, %f2704, %f2728; fma.rn.ftz.f32 %f2730, %f3525, %f2706, %f2729; add.ftz.f32 %f726, %f2707, %f2730; mul.ftz.f32 %f2731, %f3524, %f2713; fma.rn.ftz.f32 %f2732, %f3523, %f2712, %f2731; fma.rn.ftz.f32 %f2733, %f3525, %f2714, %f2732; add.ftz.f32 %f727, %f2715, %f2733; mul.ftz.f32 %f2734, %f3524, %f2721; fma.rn.ftz.f32 %f2735, %f3523, %f2720, %f2734; fma.rn.ftz.f32 %f2736, %f3525, %f2722, %f2735; add.ftz.f32 %f3525, %f2723, %f2736; mov.f32 %f3523, %f726; mov.f32 %f3524, %f727; bra.uni $L__BB208_183; $L__BB208_181: add.s64 %rd467, %rd290, 144; cvta.to.local.u64 %rd468, %rd467; ld.local.v4.f32 {%f2671, %f2672, %f2673, %f2674}, [%rd468]; mul.ftz.f32 %f2679, %f3524, %f2672; fma.rn.ftz.f32 %f2680, %f3523, %f2671, %f2679; fma.rn.ftz.f32 %f2681, %f3525, %f2673, %f2680; add.ftz.f32 %f723, %f2674, %f2681; add.s64 %rd469, %rd290, 160; cvta.to.local.u64 %rd470, %rd469; ld.local.v4.f32 {%f2682, %f2683, %f2684, %f2685}, [%rd470]; mul.ftz.f32 %f2690, %f3524, %f2683; fma.rn.ftz.f32 %f2691, %f3523, %f2682, %f2690; fma.rn.ftz.f32 %f2692, %f3525, %f2684, %f2691; add.ftz.f32 %f724, %f2685, %f2692; add.s64 %rd471, %rd290, 176; cvta.to.local.u64 %rd472, %rd471; ld.local.v4.f32 {%f2693, %f2694, %f2695, %f2696}, [%rd472]; mul.ftz.f32 %f2701, %f3524, %f2694; fma.rn.ftz.f32 %f2702, %f3523, %f2693, %f2701; fma.rn.ftz.f32 %f2703, %f3525, %f2695, %f2702; add.ftz.f32 %f3525, %f2696, %f2703; mov.f32 %f3523, %f723; mov.f32 %f3524, %f724; $L__BB208_183: sub.ftz.f32 %f2737, %f700, %f3523; sub.ftz.f32 %f2738, %f701, %f3524; mul.ftz.f32 %f2739, %f2738, %f2738; fma.rn.ftz.f32 %f2740, %f2737, %f2737, %f2739; sub.ftz.f32 %f2741, %f702, %f3525; fma.rn.ftz.f32 %f2742, %f2741, %f2741, %f2740; sqrt.approx.ftz.f32 %f2743, %f2742; div.approx.ftz.f32 %f3556, %f2737, %f2743; div.approx.ftz.f32 %f3557, %f2738, %f2743; div.approx.ftz.f32 %f3558, %f2741, %f2743; add.s64 %rd474, %rd290, 32; cvta.to.local.u64 %rd475, %rd474; st.local.f32 [%rd475], %f3558; or.b64 %rd476, %rd290, 12; cvta.to.local.u64 %rd477, %rd476; st.local.f32 [%rd477], %f3556; add.s64 %rd478, %rd290, 16; cvta.to.local.u64 %rd479, %rd478; st.local.v4.f32 [%rd479], {%f3557, %f3558, %f3556, %f3557}; add.s64 %rd480, %rd290, 116; cvta.to.local.u64 %rd481, %rd480; mov.u32 %r454, 0; st.local.u32 [%rd481], %r454; add.s64 %rd482, %rd290, 120; cvta.to.local.u64 %rd483, %rd482; st.local.u32 [%rd483], %r454; add.s64 %rd484, %rd290, 124; cvta.to.local.u64 %rd485, %rd484; st.local.u32 [%rd485], %r454; add.s64 %rd486, %rd290, 128; cvta.to.local.u64 %rd487, %rd486; mov.f32 %f2744, 0f00000000; st.local.v2.f32 [%rd487], {%f2744, %f2744}; add.s64 %rd488, %rd290, 136; cvta.to.local.u64 %rd489, %rd488; st.local.u32 [%rd489], %r454; bra.uni $L__BB208_213; $L__BB208_191: add.s64 %rd507, %rd290, 144; cvta.to.local.u64 %rd508, %rd507; ld.local.v4.f32 {%f3529, %f3530, %f3531, %f3532}, [%rd508]; add.s64 %rd509, %rd290, 160; cvta.to.local.u64 %rd510, %rd509; ld.local.v4.f32 {%f3533, %f3534, %f3535, %f3536}, [%rd510]; add.s64 %rd511, %rd290, 176; cvta.to.local.u64 %rd512, %rd511; ld.local.v4.f32 {%f3537, %f3538, %f3539, %f3540}, [%rd512]; $L__BB208_193: mul.ftz.f32 %f2792, %f3542, %f3530; fma.rn.ftz.f32 %f2793, %f3541, %f3529, %f2792; fma.rn.ftz.f32 %f2794, %f3543, %f3531, %f2793; mul.ftz.f32 %f2795, %f3542, %f3534; fma.rn.ftz.f32 %f2796, %f3541, %f3533, %f2795; fma.rn.ftz.f32 %f2797, %f3543, %f3535, %f2796; mul.ftz.f32 %f2798, %f3542, %f3538; fma.rn.ftz.f32 %f2799, %f3541, %f3537, %f2798; fma.rn.ftz.f32 %f2800, %f3543, %f3539, %f2799; add.ftz.f32 %f3541, %f3532, %f2794; add.ftz.f32 %f3542, %f3536, %f2797; add.ftz.f32 %f3543, %f3540, %f2800; $L__BB208_194: add.s64 %rd514, %rd290, 32; cvta.to.local.u64 %rd515, %rd514; st.local.f32 [%rd515], %f3528; cvta.to.local.u64 %rd516, %rd290; st.local.v4.f32 [%rd516], {%f3541, %f3542, %f3543, %f3526}; add.s64 %rd517, %rd290, 16; cvta.to.local.u64 %rd518, %rd517; st.local.v4.f32 [%rd518], {%f3527, %f3528, %f3526, %f3527}; setp.gt.s32 %p220, %r231, -1; @%p220 bra $L__BB208_198; ld.global.v4.u32 {%r460, %r461, %r462, %r463}, [%rd126+-12]; mul.wide.u32 %rd520, %r460, 12; add.s64 %rd521, %rd519, %rd520; mul.wide.u32 %rd522, %r461, 12; add.s64 %rd523, %rd519, %rd522; mul.wide.u32 %rd524, %r462, 12; add.s64 %rd525, %rd519, %rd524; mov.f32 %f2801, 0f3F800000; sub.ftz.f32 %f2802, %f2801, %f3336; sub.ftz.f32 %f2803, %f2802, %f3335; ld.global.f32 %f2804, [%rd521]; ld.global.f32 %f2805, [%rd521+4]; ld.global.f32 %f2806, [%rd521+8]; ld.global.f32 %f2807, [%rd523]; mul.ftz.f32 %f2808, %f3336, %f2807; ld.global.f32 %f2809, [%rd523+4]; mul.ftz.f32 %f2810, %f3336, %f2809; ld.global.f32 %f2811, [%rd523+8]; mul.ftz.f32 %f2812, %f3336, %f2811; fma.rn.ftz.f32 %f2813, %f2803, %f2804, %f2808; fma.rn.ftz.f32 %f2814, %f2803, %f2805, %f2810; fma.rn.ftz.f32 %f2815, %f2803, %f2806, %f2812; ld.global.f32 %f2816, [%rd525]; ld.global.f32 %f2817, [%rd525+4]; ld.global.f32 %f2818, [%rd525+8]; fma.rn.ftz.f32 %f3544, %f3335, %f2816, %f2813; fma.rn.ftz.f32 %f3545, %f3335, %f2817, %f2814; fma.rn.ftz.f32 %f3546, %f3335, %f2818, %f2815; mul.ftz.f32 %f2819, %f3545, %f3545; fma.rn.ftz.f32 %f2820, %f3544, %f3544, %f2819; fma.rn.ftz.f32 %f2821, %f3546, %f3546, %f2820; sqrt.approx.ftz.f32 %f812, %f2821; setp.neu.ftz.f32 %p221, %f812, 0f00000000; @%p221 bra $L__BB208_196; bra.uni $L__BB208_197; $L__BB208_196: rcp.approx.ftz.f32 %f2822, %f812; mul.ftz.f32 %f3546, %f3546, %f2822; mul.ftz.f32 %f3545, %f3545, %f2822; mul.ftz.f32 %f3544, %f3544, %f2822; $L__BB208_197: setp.eq.ftz.f32 %p222, %f3544, 0f00000000; setp.eq.ftz.f32 %p223, %f3545, 0f00000000; and.pred %p224, %p222, %p223; setp.eq.ftz.f32 %p225, %f3546, 0f00000000; and.pred %p226, %p225, %p224; selp.f32 %f2823, %f3528, %f3546, %p226; selp.f32 %f2824, %f3527, %f3545, %p226; selp.f32 %f2825, %f3526, %f3544, %p226; or.b64 %rd527, %rd290, 12; cvta.to.local.u64 %rd528, %rd527; st.local.f32 [%rd528], %f2825; st.local.f32 [%rd518], %f2824; add.s64 %rd531, %rd290, 20; cvta.to.local.u64 %rd532, %rd531; st.local.f32 [%rd532], %f2823; $L__BB208_198: add.s64 %rd534, %rd290, 116; cvta.to.local.u64 %rd535, %rd534; st.local.f32 [%rd535], %f743; add.s64 %rd536, %rd290, 120; cvta.to.local.u64 %rd537, %rd536; st.local.f32 [%rd537], %f744; add.s64 %rd538, %rd290, 124; cvta.to.local.u64 %rd539, %rd538; st.local.f32 [%rd539], %f745; add.s64 %rd540, %rd290, 128; cvta.to.local.u64 %rd541, %rd540; st.local.v2.f32 [%rd541], {%f738, %f740}; add.s64 %rd542, %rd290, 136; cvta.to.local.u64 %rd543, %rd542; st.local.f32 [%rd543], %f742; $L__BB208_199: and.b32 %r467, %r641, 4; setp.eq.s32 %p227, %r467, 0; @%p227 bra $L__BB208_201; bra.uni $L__BB208_200; $L__BB208_201: and.b32 %r233, %r641, 2; setp.eq.s32 %p228, %r233, 0; @%p228 bra $L__BB208_203; bra.uni $L__BB208_202; $L__BB208_203: add.s64 %rd561, %rd290, 16; add.s64 %rd562, %rd290, 76; cvta.to.local.u64 %rd563, %rd562; ld.local.u32 %r468, [%rd563]; mul.wide.s32 %rd564, %r468, 208; add.s64 %rd565, %rd19, %rd564; ld.global.v4.f32 {%f2867, %f2868, %f2869, %f2870}, [%rd565+48]; ld.global.v4.f32 {%f2874, %f2875, %f2876, %f2877}, [%rd565+64]; ld.global.v4.f32 {%f2881, %f2882, %f2883, %f2884}, [%rd565+80]; or.b64 %rd566, %rd290, 12; cvta.to.local.u64 %rd567, %rd566; cvta.to.local.u64 %rd568, %rd561; ld.local.v2.f32 {%f2888, %f2889}, [%rd568]; ld.local.f32 %f2892, [%rd567]; mul.ftz.f32 %f2893, %f2874, %f2888; fma.rn.ftz.f32 %f2894, %f2867, %f2892, %f2893; fma.rn.ftz.f32 %f2895, %f2881, %f2889, %f2894; mul.ftz.f32 %f2896, %f2875, %f2888; fma.rn.ftz.f32 %f2897, %f2868, %f2892, %f2896; fma.rn.ftz.f32 %f2898, %f2882, %f2889, %f2897; mul.ftz.f32 %f2899, %f2876, %f2888; fma.rn.ftz.f32 %f2900, %f2869, %f2892, %f2899; fma.rn.ftz.f32 %f2901, %f2883, %f2889, %f2900; mul.ftz.f32 %f2902, %f2898, %f2898; fma.rn.ftz.f32 %f2903, %f2895, %f2895, %f2902; fma.rn.ftz.f32 %f2904, %f2901, %f2901, %f2903; sqrt.approx.ftz.f32 %f2905, %f2904; div.approx.ftz.f32 %f3549, %f2895, %f2905; div.approx.ftz.f32 %f3548, %f2898, %f2905; div.approx.ftz.f32 %f3547, %f2901, %f2905; bra.uni $L__BB208_204; $L__BB208_200: add.s64 %rd545, %rd290, 24; cvta.to.local.u64 %rd546, %rd545; ld.local.v2.f32 {%f3556, %f3557}, [%rd546]; add.s64 %rd547, %rd290, 32; cvta.to.local.u64 %rd548, %rd547; ld.local.f32 %f3558, [%rd548]; bra.uni $L__BB208_213; $L__BB208_202: add.s64 %rd550, %rd290, 16; or.b64 %rd551, %rd290, 12; cvta.to.local.u64 %rd552, %rd551; cvta.to.local.u64 %rd553, %rd550; ld.local.v2.f32 {%f2828, %f2829}, [%rd553]; add.s64 %rd554, %rd290, 192; cvta.to.local.u64 %rd555, %rd554; ld.local.v4.f32 {%f2832, %f2833, %f2834, %f2835}, [%rd555]; add.s64 %rd556, %rd290, 208; cvta.to.local.u64 %rd557, %rd556; ld.local.v4.f32 {%f2839, %f2840, %f2841, %f2842}, [%rd557]; add.s64 %rd558, %rd290, 224; cvta.to.local.u64 %rd559, %rd558; ld.local.v4.f32 {%f2846, %f2847, %f2848, %f2849}, [%rd559]; ld.local.f32 %f2853, [%rd552]; mul.ftz.f32 %f2854, %f2828, %f2839; fma.rn.ftz.f32 %f2855, %f2853, %f2832, %f2854; fma.rn.ftz.f32 %f2856, %f2829, %f2846, %f2855; mul.ftz.f32 %f2857, %f2828, %f2840; fma.rn.ftz.f32 %f2858, %f2853, %f2833, %f2857; fma.rn.ftz.f32 %f2859, %f2829, %f2847, %f2858; mul.ftz.f32 %f2860, %f2828, %f2841; fma.rn.ftz.f32 %f2861, %f2853, %f2834, %f2860; fma.rn.ftz.f32 %f2862, %f2829, %f2848, %f2861; mul.ftz.f32 %f2863, %f2859, %f2859; fma.rn.ftz.f32 %f2864, %f2856, %f2856, %f2863; fma.rn.ftz.f32 %f2865, %f2862, %f2862, %f2864; sqrt.approx.ftz.f32 %f2866, %f2865; div.approx.ftz.f32 %f3549, %f2856, %f2866; div.approx.ftz.f32 %f3548, %f2859, %f2866; div.approx.ftz.f32 %f3547, %f2862, %f2866; $L__BB208_204: or.b64 %rd570, %rd290, 12; cvta.to.local.u64 %rd571, %rd570; st.local.f32 [%rd571], %f3549; add.s64 %rd572, %rd290, 16; cvta.to.local.u64 %rd573, %rd572; st.local.f32 [%rd573], %f3548; add.s64 %rd574, %rd290, 20; cvta.to.local.u64 %rd575, %rd574; st.local.f32 [%rd575], %f3547; @%p228 bra $L__BB208_206; bra.uni $L__BB208_205; $L__BB208_206: add.s64 %rd588, %rd290, 24; add.s64 %rd589, %rd290, 76; cvta.to.local.u64 %rd590, %rd589; ld.local.u32 %r469, [%rd590]; mul.wide.s32 %rd591, %r469, 208; add.s64 %rd592, %rd19, %rd591; ld.global.v4.f32 {%f2945, %f2946, %f2947, %f2948}, [%rd592+48]; ld.global.v4.f32 {%f2952, %f2953, %f2954, %f2955}, [%rd592+64]; ld.global.v4.f32 {%f2959, %f2960, %f2961, %f2962}, [%rd592+80]; cvta.to.local.u64 %rd593, %rd588; ld.local.v2.f32 {%f2966, %f2967}, [%rd593]; add.s64 %rd594, %rd290, 32; cvta.to.local.u64 %rd595, %rd594; mul.ftz.f32 %f2970, %f2952, %f2967; fma.rn.ftz.f32 %f2971, %f2945, %f2966, %f2970; ld.local.f32 %f2972, [%rd595]; fma.rn.ftz.f32 %f2973, %f2959, %f2972, %f2971; mul.ftz.f32 %f2974, %f2953, %f2967; fma.rn.ftz.f32 %f2975, %f2946, %f2966, %f2974; fma.rn.ftz.f32 %f2976, %f2960, %f2972, %f2975; mul.ftz.f32 %f2977, %f2954, %f2967; fma.rn.ftz.f32 %f2978, %f2947, %f2966, %f2977; fma.rn.ftz.f32 %f2979, %f2961, %f2972, %f2978; mul.ftz.f32 %f2980, %f2976, %f2976; fma.rn.ftz.f32 %f2981, %f2973, %f2973, %f2980; fma.rn.ftz.f32 %f2982, %f2979, %f2979, %f2981; sqrt.approx.ftz.f32 %f2983, %f2982; div.approx.ftz.f32 %f3556, %f2973, %f2983; div.approx.ftz.f32 %f3557, %f2976, %f2983; div.approx.ftz.f32 %f3558, %f2979, %f2983; bra.uni $L__BB208_207; $L__BB208_205: add.s64 %rd577, %rd290, 24; cvta.to.local.u64 %rd578, %rd577; ld.local.v2.f32 {%f2906, %f2907}, [%rd578]; add.s64 %rd579, %rd290, 32; cvta.to.local.u64 %rd580, %rd579; add.s64 %rd581, %rd290, 192; cvta.to.local.u64 %rd582, %rd581; ld.local.v4.f32 {%f2910, %f2911, %f2912, %f2913}, [%rd582]; add.s64 %rd583, %rd290, 208; cvta.to.local.u64 %rd584, %rd583; ld.local.v4.f32 {%f2917, %f2918, %f2919, %f2920}, [%rd584]; add.s64 %rd585, %rd290, 224; cvta.to.local.u64 %rd586, %rd585; ld.local.v4.f32 {%f2924, %f2925, %f2926, %f2927}, [%rd586]; mul.ftz.f32 %f2931, %f2907, %f2917; fma.rn.ftz.f32 %f2932, %f2906, %f2910, %f2931; ld.local.f32 %f2933, [%rd580]; fma.rn.ftz.f32 %f2934, %f2933, %f2924, %f2932; mul.ftz.f32 %f2935, %f2907, %f2918; fma.rn.ftz.f32 %f2936, %f2906, %f2911, %f2935; fma.rn.ftz.f32 %f2937, %f2933, %f2925, %f2936; mul.ftz.f32 %f2938, %f2907, %f2919; fma.rn.ftz.f32 %f2939, %f2906, %f2912, %f2938; fma.rn.ftz.f32 %f2940, %f2933, %f2926, %f2939; mul.ftz.f32 %f2941, %f2937, %f2937; fma.rn.ftz.f32 %f2942, %f2934, %f2934, %f2941; fma.rn.ftz.f32 %f2943, %f2940, %f2940, %f2942; sqrt.approx.ftz.f32 %f2944, %f2943; div.approx.ftz.f32 %f3556, %f2934, %f2944; div.approx.ftz.f32 %f3557, %f2937, %f2944; div.approx.ftz.f32 %f3558, %f2940, %f2944; $L__BB208_207: add.s64 %rd597, %rd290, 24; cvta.to.local.u64 %rd598, %rd597; st.local.v2.f32 [%rd598], {%f3556, %f3557}; add.s64 %rd599, %rd290, 32; cvta.to.local.u64 %rd600, %rd599; st.local.f32 [%rd600], %f3558; @%p228 bra $L__BB208_209; bra.uni $L__BB208_208; $L__BB208_209: add.s64 %rd613, %rd290, 120; add.s64 %rd614, %rd290, 76; cvta.to.local.u64 %rd615, %rd614; ld.local.u32 %r470, [%rd615]; mul.wide.s32 %rd616, %r470, 208; add.s64 %rd617, %rd19, %rd616; ld.global.v4.f32 {%f3016, %f3017, %f3018, %f3019}, [%rd617]; ld.global.v4.f32 {%f3023, %f3024, %f3025, %f3026}, [%rd617+16]; ld.global.v4.f32 {%f3030, %f3031, %f3032, %f3033}, [%rd617+32]; add.s64 %rd618, %rd290, 116; cvta.to.local.u64 %rd619, %rd618; cvta.to.local.u64 %rd620, %rd613; ld.local.v2.f32 {%f3037, %f3038}, [%rd620]; ld.local.f32 %f3041, [%rd619]; mul.ftz.f32 %f3042, %f3017, %f3037; fma.rn.ftz.f32 %f3043, %f3041, %f3016, %f3042; fma.rn.ftz.f32 %f3555, %f3018, %f3038, %f3043; mul.ftz.f32 %f3044, %f3024, %f3037; fma.rn.ftz.f32 %f3045, %f3041, %f3023, %f3044; fma.rn.ftz.f32 %f3554, %f3025, %f3038, %f3045; mul.ftz.f32 %f3046, %f3031, %f3037; fma.rn.ftz.f32 %f3047, %f3041, %f3030, %f3046; fma.rn.ftz.f32 %f3553, %f3032, %f3038, %f3047; bra.uni $L__BB208_210; $L__BB208_208: add.s64 %rd602, %rd290, 120; add.s64 %rd603, %rd290, 116; cvta.to.local.u64 %rd604, %rd603; cvta.to.local.u64 %rd605, %rd602; ld.local.v2.f32 {%f2984, %f2985}, [%rd605]; add.s64 %rd606, %rd290, 144; cvta.to.local.u64 %rd607, %rd606; ld.local.v4.f32 {%f2988, %f2989, %f2990, %f2991}, [%rd607]; ld.local.f32 %f2995, [%rd604]; mul.ftz.f32 %f2996, %f2984, %f2989; fma.rn.ftz.f32 %f2997, %f2995, %f2988, %f2996; fma.rn.ftz.f32 %f3555, %f2985, %f2990, %f2997; add.s64 %rd608, %rd290, 160; cvta.to.local.u64 %rd609, %rd608; ld.local.v4.f32 {%f2998, %f2999, %f3000, %f3001}, [%rd609]; mul.ftz.f32 %f3005, %f2984, %f2999; fma.rn.ftz.f32 %f3006, %f2995, %f2998, %f3005; fma.rn.ftz.f32 %f3554, %f2985, %f3000, %f3006; add.s64 %rd610, %rd290, 176; cvta.to.local.u64 %rd611, %rd610; ld.local.v4.f32 {%f3007, %f3008, %f3009, %f3010}, [%rd611]; mul.ftz.f32 %f3014, %f2984, %f3008; fma.rn.ftz.f32 %f3015, %f2995, %f3007, %f3014; fma.rn.ftz.f32 %f3553, %f2985, %f3009, %f3015; $L__BB208_210: add.s64 %rd622, %rd290, 116; cvta.to.local.u64 %rd623, %rd622; st.local.f32 [%rd623], %f3555; add.s64 %rd624, %rd290, 120; cvta.to.local.u64 %rd625, %rd624; st.local.f32 [%rd625], %f3554; add.s64 %rd626, %rd290, 124; cvta.to.local.u64 %rd627, %rd626; st.local.f32 [%rd627], %f3553; @%p228 bra $L__BB208_212; bra.uni $L__BB208_211; $L__BB208_212: add.s64 %rd640, %rd290, 128; add.s64 %rd641, %rd290, 76; cvta.to.local.u64 %rd642, %rd641; ld.local.u32 %r471, [%rd642]; mul.wide.s32 %rd643, %r471, 208; add.s64 %rd644, %rd19, %rd643; ld.global.v4.f32 {%f3085, %f3086, %f3087, %f3088}, [%rd644]; ld.global.v4.f32 {%f3092, %f3093, %f3094, %f3095}, [%rd644+16]; ld.global.v4.f32 {%f3099, %f3100, %f3101, %f3102}, [%rd644+32]; cvta.to.local.u64 %rd645, %rd640; ld.local.v4.f32 {%f3106, %f3107, %f3108, %f3109}, [%rd645]; mul.ftz.f32 %f3113, %f3086, %f3107; fma.rn.ftz.f32 %f3114, %f3106, %f3085, %f3113; mul.ftz.f32 %f3115, %f3093, %f3107; fma.rn.ftz.f32 %f3116, %f3106, %f3092, %f3115; mul.ftz.f32 %f3117, %f3100, %f3107; fma.rn.ftz.f32 %f3118, %f3106, %f3099, %f3117; fma.rn.ftz.f32 %f3119, %f3101, %f3108, %f3118; fma.rn.ftz.f32 %f3120, %f3094, %f3108, %f3116; fma.rn.ftz.f32 %f3121, %f3087, %f3108, %f3114; st.local.v2.f32 [%rd645], {%f3121, %f3120}; add.s64 %rd646, %rd290, 136; cvta.to.local.u64 %rd647, %rd646; st.local.f32 [%rd647], %f3119; bra.uni $L__BB208_213; $L__BB208_211: add.s64 %rd629, %rd290, 128; cvta.to.local.u64 %rd630, %rd629; ld.local.v4.f32 {%f3048, %f3049, %f3050, %f3051}, [%rd630]; add.s64 %rd631, %rd290, 144; cvta.to.local.u64 %rd632, %rd631; ld.local.v4.f32 {%f3055, %f3056, %f3057, %f3058}, [%rd632]; mul.ftz.f32 %f3062, %f3049, %f3056; fma.rn.ftz.f32 %f3063, %f3048, %f3055, %f3062; add.s64 %rd633, %rd290, 160; cvta.to.local.u64 %rd634, %rd633; ld.local.v4.f32 {%f3064, %f3065, %f3066, %f3067}, [%rd634]; mul.ftz.f32 %f3071, %f3049, %f3065; fma.rn.ftz.f32 %f3072, %f3048, %f3064, %f3071; add.s64 %rd635, %rd290, 176; cvta.to.local.u64 %rd636, %rd635; ld.local.v4.f32 {%f3073, %f3074, %f3075, %f3076}, [%rd636]; mul.ftz.f32 %f3080, %f3049, %f3074; fma.rn.ftz.f32 %f3081, %f3048, %f3073, %f3080; fma.rn.ftz.f32 %f3082, %f3050, %f3075, %f3081; fma.rn.ftz.f32 %f3083, %f3050, %f3066, %f3072; fma.rn.ftz.f32 %f3084, %f3050, %f3057, %f3063; st.local.v2.f32 [%rd630], {%f3084, %f3083}; add.s64 %rd637, %rd290, 136; cvta.to.local.u64 %rd638, %rd637; st.local.f32 [%rd638], %f3082; $L__BB208_213: add.s64 %rd848, %rd290, 40; cvta.to.local.u64 %rd847, %rd848; add.s64 %rd846, %rd290, 36; cvta.to.local.u64 %rd845, %rd846; add.s64 %rd844, %rd290, 52; cvta.to.local.u64 %rd843, %rd844; add.s64 %rd649, %rd290, 48; cvta.to.local.u64 %rd650, %rd649; ld.local.u32 %r234, [%rd650]; shl.b32 %r472, %r234, 5; cvt.u64.u32 %rd651, %r472; and.b64 %rd652, %rd651, 134217696; add.s64 %rd653, %rd27, %rd652; ld.global.u32 %r642, [%rd653+16]; st.local.u32 [%rd843], %r642; ld.local.f32 %f3122, [%rd845]; ld.local.v2.f32 {%f3123, %f3124}, [%rd847]; mul.ftz.f32 %f3127, %f3557, %f3123; fma.rn.ftz.f32 %f3128, %f3556, %f3122, %f3127; fma.rn.ftz.f32 %f3129, %f3558, %f3124, %f3128; setp.lt.ftz.f32 %p232, %f3129, 0f00000000; @%p232 bra $L__BB208_215; bra.uni $L__BB208_214; $L__BB208_215: add.s64 %rd853, %rd290, 52; cvta.to.local.u64 %rd852, %rd853; add.s64 %rd666, %rd290, 128; or.b32 %r642, %r642, 1; st.local.u32 [%rd852], %r642; add.s64 %rd669, %rd290, 32; cvta.to.local.u64 %rd670, %rd669; neg.ftz.f32 %f3558, %f3558; st.local.f32 [%rd670], %f3558; or.b64 %rd671, %rd290, 12; cvta.to.local.u64 %rd672, %rd671; ld.local.f32 %f3130, [%rd672]; neg.ftz.f32 %f3131, %f3130; add.s64 %rd673, %rd290, 16; cvta.to.local.u64 %rd674, %rd673; ld.local.v2.f32 {%f3132, %f3133}, [%rd674]; st.local.f32 [%rd672], %f3131; neg.ftz.f32 %f3136, %f3133; neg.ftz.f32 %f3137, %f3132; neg.ftz.f32 %f3557, %f3557; neg.ftz.f32 %f3556, %f3556; st.local.v4.f32 [%rd674], {%f3137, %f3136, %f3556, %f3557}; add.s64 %rd675, %rd290, 116; cvta.to.local.u64 %rd676, %rd675; ld.local.f32 %f3138, [%rd676]; neg.ftz.f32 %f3570, %f3138; add.s64 %rd677, %rd290, 120; cvta.to.local.u64 %rd678, %rd677; ld.local.v2.f32 {%f3139, %f3140}, [%rd678]; neg.ftz.f32 %f3571, %f3139; neg.ftz.f32 %f3143, %f3140; st.local.f32 [%rd676], %f3570; st.local.f32 [%rd678], %f3571; add.s64 %rd679, %rd290, 124; cvta.to.local.u64 %rd680, %rd679; st.local.f32 [%rd680], %f3143; cvta.to.local.u64 %rd681, %rd666; ld.local.v4.f32 {%f3144, %f3145, %f3146, %f3147}, [%rd681]; neg.ftz.f32 %f3151, %f3146; neg.ftz.f32 %f3152, %f3145; neg.ftz.f32 %f3153, %f3144; st.local.v2.f32 [%rd681], {%f3153, %f3152}; add.s64 %rd682, %rd290, 136; cvta.to.local.u64 %rd683, %rd682; st.local.f32 [%rd683], %f3151; bra.uni $L__BB208_216; $L__BB208_214: add.s64 %rd661, %rd290, 116; cvta.to.local.u64 %rd662, %rd661; ld.local.f32 %f3570, [%rd662]; add.s64 %rd663, %rd290, 120; cvta.to.local.u64 %rd664, %rd663; ld.local.f32 %f3571, [%rd664]; $L__BB208_216: add.s64 %rd850, %rd290, 88; cvta.to.local.u64 %rd849, %rd850; add.s64 %rd685, %rd290, 128; ld.local.f32 %f3154, [%rd849]; fma.rn.ftz.f32 %f3155, %f6, %f3154, %f5; add.s64 %rd688, %rd290, 92; cvta.to.local.u64 %rd689, %rd688; st.local.f32 [%rd689], %f3155; add.s64 %rd690, %rd290, 96; cvta.to.local.u64 %rd691, %rd690; st.local.f32 [%rd691], %f6; cvta.to.local.u64 %rd692, %rd685; ld.local.v4.f32 {%f3568, %f3569, %f3158, %f3159}, [%rd692]; setp.neu.ftz.f32 %p233, %f3556, %f3557; setp.neu.ftz.f32 %p234, %f3556, %f3558; or.pred %p235, %p233, %p234; sub.ftz.f32 %f3160, %f3558, %f3557; neg.ftz.f32 %f3161, %f3557; neg.ftz.f32 %f3162, %f3558; selp.f32 %f3163, %f3162, %f3558, %p235; add.ftz.f32 %f3164, %f3556, %f3163; selp.f32 %f3165, %f3557, %f3161, %p235; sub.ftz.f32 %f3166, %f3165, %f3556; mul.ftz.f32 %f3167, %f3164, %f3164; fma.rn.ftz.f32 %f3168, %f3160, %f3160, %f3167; fma.rn.ftz.f32 %f3169, %f3166, %f3166, %f3168; sqrt.approx.ftz.f32 %f3170, %f3169; div.approx.ftz.f32 %f3171, %f3160, %f3170; div.approx.ftz.f32 %f3172, %f3164, %f3170; div.approx.ftz.f32 %f3173, %f3166, %f3170; mul.ftz.f32 %f3174, %f3557, %f3173; mul.ftz.f32 %f3175, %f3558, %f3172; sub.ftz.f32 %f3176, %f3174, %f3175; mul.ftz.f32 %f3177, %f3558, %f3171; mul.ftz.f32 %f3178, %f3556, %f3173; sub.ftz.f32 %f3179, %f3177, %f3178; mul.ftz.f32 %f3180, %f3556, %f3172; mul.ftz.f32 %f3181, %f3557, %f3171; sub.ftz.f32 %f3182, %f3180, %f3181; mul.ftz.f32 %f3564, %f3155, %f3171; mul.ftz.f32 %f3565, %f3155, %f3172; mul.ftz.f32 %f881, %f3155, %f3173; mul.ftz.f32 %f3566, %f3155, %f3176; mul.ftz.f32 %f3567, %f3155, %f3179; mul.ftz.f32 %f884, %f3155, %f3182; abs.ftz.f32 %f885, %f3556; abs.ftz.f32 %f886, %f3557; abs.ftz.f32 %f887, %f3558; setp.geu.ftz.f32 %p236, %f887, %f885; setp.geu.ftz.f32 %p237, %f887, %f886; and.pred %p238, %p236, %p237; @%p238 bra $L__BB208_218; add.s64 %rd694, %rd290, 124; cvta.to.local.u64 %rd695, %rd694; ld.local.f32 %f889, [%rd695]; setp.lt.ftz.f32 %p239, %f886, %f887; setp.lt.ftz.f32 %p240, %f886, %f885; or.pred %p241, %p240, %p239; selp.f32 %f3564, %f3565, %f3564, %p241; selp.f32 %f3566, %f3567, %f3566, %p241; selp.f32 %f3568, %f3569, %f3568, %p241; selp.f32 %f3570, %f3571, %f3570, %p241; mov.f32 %f3565, %f881; mov.f32 %f3567, %f884; mov.f32 %f3569, %f3158; mov.f32 %f3571, %f889; $L__BB208_218: mul.ftz.f32 %f3183, %f3568, %f3571; mul.ftz.f32 %f3184, %f3569, %f3570; sub.ftz.f32 %f3572, %f3184, %f3183; setp.eq.ftz.f32 %p242, %f3572, 0f00000000; @%p242 bra $L__BB208_220; rcp.approx.ftz.f32 %f3572, %f3572; $L__BB208_220: mul.ftz.f32 %f3185, %f3565, %f3568; mul.ftz.f32 %f3186, %f3564, %f3569; sub.ftz.f32 %f3187, %f3186, %f3185; mul.ftz.f32 %f3188, %f3187, %f3572; add.s64 %rd697, %rd290, 100; cvta.to.local.u64 %rd698, %rd697; st.local.f32 [%rd698], %f3188; mul.ftz.f32 %f3189, %f3564, %f3571; mul.ftz.f32 %f3190, %f3565, %f3570; sub.ftz.f32 %f3191, %f3190, %f3189; mul.ftz.f32 %f3192, %f3567, %f3568; mul.ftz.f32 %f3193, %f3566, %f3569; sub.ftz.f32 %f3194, %f3193, %f3192; mul.ftz.f32 %f3195, %f3191, %f3572; mul.ftz.f32 %f3196, %f3194, %f3572; add.s64 %rd699, %rd290, 104; cvta.to.local.u64 %rd700, %rd699; st.local.v2.f32 [%rd700], {%f3196, %f3195}; mul.ftz.f32 %f3197, %f3566, %f3571; mul.ftz.f32 %f3198, %f3567, %f3570; sub.ftz.f32 %f3199, %f3198, %f3197; mul.ftz.f32 %f3200, %f3199, %f3572; add.s64 %rd701, %rd290, 112; cvta.to.local.u64 %rd702, %rd701; st.local.f32 [%rd702], %f3200; and.b32 %r473, %r642, 1; setp.eq.b32 %p243, %r473, 1; mov.pred %p244, 0; xor.pred %p245, %p243, %p244; not.pred %p246, %p245; cvta.to.local.u64 %rd703, %rd290; add.s64 %rd127, %rd703, 76; @%p246 bra $L__BB208_230; bra.uni $L__BB208_221; $L__BB208_230: ld.local.u32 %r479, [%rd127]; add.s32 %r245, %r536, 1; add.u64 %rd727, %SPL, 1072; mul.wide.s32 %rd728, %r536, 4; add.s64 %rd729, %rd727, %rd728; st.local.u32 [%rd729], %r479; mov.u32 %r536, %r245; bra.uni $L__BB208_231; $L__BB208_221: setp.lt.s32 %p248, %r536, 1; mov.pred %p286, -1; mov.u16 %rs47, 1; @%p248 bra $L__BB208_224; ld.local.u32 %r238, [%rd127]; add.u64 %rd128, %SPL, 1072; mov.u16 %rs47, 1; mov.u32 %r643, 0; $L__BB208_223: mul.wide.s32 %rd705, %r643, 4; add.s64 %rd706, %rd128, %rd705; ld.local.u32 %r475, [%rd706]; setp.eq.s32 %p249, %r475, %r238; selp.b16 %rs47, 0, %rs47, %p249; and.b16 %rs29, %rs47, 255; add.s32 %r643, %r643, 1; setp.lt.s32 %p250, %r643, %r536; setp.ne.s16 %p286, %rs29, 0; and.pred %p251, %p286, %p250; @%p251 bra $L__BB208_223; $L__BB208_224: setp.lt.s32 %p252, %r647, 1; not.pred %p253, %p286; or.pred %p254, %p253, %p252; mov.u16 %rs48, %rs47; @%p254 bra $L__BB208_228; mov.u32 %r644, 0; ld.local.u32 %r241, [%rd127]; $L__BB208_226: mul.wide.s32 %rd707, %r644, 16; mov.u64 %rd708, kernel_params; add.s64 %rd709, %rd708, %rd707; ld.const.u64 %rd710, [%rd709+2768]; cvta.to.global.u64 %rd711, %rd710; shl.b64 %rd712, %rd2, 2; add.s64 %rd713, %rd711, %rd712; ld.global.u32 %r477, [%rd713]; setp.eq.s32 %p255, %r477, %r241; add.s32 %r644, %r644, 1; mov.u16 %rs48, 0; @%p255 bra $L__BB208_228; setp.lt.s32 %p256, %r644, %r647; mov.u16 %rs48, %rs47; @%p256 bra $L__BB208_226; $L__BB208_228: and.b16 %rs31, %rs48, 255; setp.eq.s16 %p257, %rs31, 0; @%p257 bra $L__BB208_231; ld.local.u32 %r478, [%rd127]; mul.wide.s32 %rd714, %r647, 16; mov.u64 %rd715, kernel_params; add.s64 %rd716, %rd715, %rd714; ld.const.v2.u64 {%rd717, %rd718}, [%rd716+2768]; cvta.to.global.u64 %rd721, %rd717; shl.b64 %rd722, %rd2, 2; add.s64 %rd723, %rd721, %rd722; st.global.u32 [%rd723], %r478; cvta.to.global.u64 %rd724, %rd718; add.s64 %rd725, %rd724, %rd722; st.global.u32 [%rd725], %r234; add.s32 %r647, %r647, 1; $L__BB208_231: ld.const.u32 %r520, [kernel_params+12]; shl.b32 %r519, %r520, 1; add.s32 %r518, %r520, -1; mov.b32 %r480, %f3337; add.s32 %r481, %r480, 1; mov.b32 %f3201, %r481; setp.eq.ftz.f32 %p258, %f3337, 0f00000000; selp.f32 %f3217, 0f00000001, %f3201, %p258; setp.lt.u32 %p259, %r647, %r518; setp.lt.s32 %p260, %r536, 31; and.pred %p261, %p260, %p259; add.s32 %r537, %r537, 1; setp.lt.u32 %p262, %r537, %r519; and.pred %p263, %p262, %p261; @%p263 bra $L__BB208_7; $L__BB208_232: ld.const.u64 %rd831, [kernel_params+2544]; mul.wide.s32 %rd830, %r533, 4; cvta.to.global.u64 %rd829, %rd831; add.s64 %rd828, %rd829, %rd830; mul.wide.s32 %rd730, %r647, 16; mov.u64 %rd731, kernel_params; add.s64 %rd732, %rd731, %rd730; ld.const.v2.u64 {%rd733, %rd734}, [%rd732+2768]; cvta.to.global.u64 %rd737, %rd733; shl.b64 %rd738, %rd2, 2; add.s64 %rd739, %rd737, %rd738; mov.u32 %r484, -1; st.global.u32 [%rd739], %r484; cvta.to.global.u64 %rd740, %rd734; add.s64 %rd741, %rd740, %rd738; st.global.u32 [%rd741], %r484; ld.global.u8 %rs32, [%rd828+3]; and.b16 %rs33, %rs32, 64; setp.eq.s16 %p264, %rs33, 0; @%p264 bra $L__BB208_244; bra.uni $L__BB208_233; $L__BB208_244: ld.const.u64 %rd818, [kernel_params+4256]; cvta.to.global.u64 %rd819, %rd818; add.s64 %rd820, %rd819, 20; atom.global.add.u32 %r510, [%rd820], -1; add.s64 %rd821, %rd819, 8; atom.global.add.u32 %r511, [%rd821], 1; ld.const.u64 %rd822, [kernel_params+2520]; cvta.to.global.u64 %rd823, %rd822; shl.b64 %rd824, %rd2, 1; add.s64 %rd825, %rd823, %rd824; mov.u16 %rs39, 2; st.global.u16 [%rd825], %rs39; bra.uni $L__BB208_245; $L__BB208_233: ld.const.u64 %rd742, [kernel_params+2712]; cvta.to.global.u64 %rd743, %rd742; add.s64 %rd745, %rd743, %rd738; ld.global.u32 %r250, [%rd745]; cvt.s64.s32 %rd129, %r250; ld.const.v2.u64 {%rd746, %rd747}, [kernel_params+2720]; cvta.to.global.u64 %rd750, %rd746; add.s64 %rd751, %rd750, %rd738; ld.global.u32 %r251, [%rd751]; cvta.to.global.u64 %rd752, %rd747; add.s64 %rd753, %rd752, %rd738; ld.global.u32 %r252, [%rd753]; and.b32 %r485, %r252, 1; setp.eq.b32 %p265, %r485, 1; mov.pred %p266, 0; xor.pred %p267, %p265, %p266; not.pred %p268, %p267; @%p268 bra $L__BB208_235; bra.uni $L__BB208_234; $L__BB208_235: and.b32 %r486, %r252, 8; setp.eq.s32 %p269, %r486, 0; @%p269 bra $L__BB208_237; bra.uni $L__BB208_236; $L__BB208_237: and.b32 %r488, %r252, 6; setp.eq.s32 %p270, %r488, 0; mov.u32 %r648, 0; @%p270 bra $L__BB208_239; cvt.u32.u64 %r489, %rd129; ld.const.u64 %rd762, [kernel_params+2264]; cvta.to.global.u64 %rd763, %rd762; mul.wide.s32 %rd764, %r489, 16; add.s64 %rd765, %rd763, %rd764; ld.global.u32 %r648, [%rd765]; bra.uni $L__BB208_239; $L__BB208_234: ld.const.u64 %rd754, [kernel_params+2216]; cvta.to.global.u64 %rd755, %rd754; shl.b64 %rd756, %rd129, 2; add.s64 %rd757, %rd755, %rd756; ld.global.u32 %r648, [%rd757]; bra.uni $L__BB208_239; $L__BB208_236: ld.const.u64 %rd758, [kernel_params+2304]; cvta.to.global.u64 %rd759, %rd758; mul.wide.s32 %rd760, %r250, 4; add.s64 %rd761, %rd759, %rd760; ld.global.u32 %r648, [%rd761]; $L__BB208_239: and.b32 %r257, %r648, 4194303; ld.const.u64 %rd766, [kernel_params+2184]; cvta.to.global.u64 %rd767, %rd766; mul.wide.s32 %rd768, %r251, 4; add.s64 %rd769, %rd767, %rd768; ld.const.u32 %r490, [kernel_params+1592]; setp.ne.s32 %p271, %r490, 0; ld.global.u8 %rs34, [%rd769+1]; and.b16 %rs35, %rs34, 6; setp.ne.s16 %p272, %rs35, 0; and.pred %p273, %p272, %p271; @%p273 bra $L__BB208_243; bra.uni $L__BB208_240; $L__BB208_243: cvt.u32.u64 %r502, %rd2; ld.const.u32 %r503, [kernel_params+4392]; div.u32 %r504, %r502, %r503; ld.const.u32 %r505, [kernel_params+8]; mad.lo.s32 %r506, %r504, %r505, %r257; ld.const.u64 %rd802, [kernel_params+4256]; cvta.to.global.u64 %rd803, %rd802; add.s64 %rd804, %rd803, 20; atom.global.add.u32 %r507, [%rd804], -1; add.s64 %rd805, %rd803, 40; atom.global.add.u32 %r508, [%rd805], 1; ld.const.u64 %rd806, [kernel_params+2520]; cvta.to.global.u64 %rd807, %rd806; shl.b64 %rd808, %rd2, 1; add.s64 %rd809, %rd807, %rd808; mov.u16 %rs38, 10; st.global.u16 [%rd809], %rs38; ld.const.u64 %rd810, [kernel_params+2624]; cvta.to.global.u64 %rd811, %rd810; add.s64 %rd813, %rd811, %rd738; st.global.u32 [%rd813], %r506; ld.const.u64 %rd814, [kernel_params+4344]; cvta.to.global.u64 %rd815, %rd814; mul.wide.s32 %rd816, %r506, 4; add.s64 %rd817, %rd815, %rd816; atom.global.add.u32 %r509, [%rd817], 1; bra.uni $L__BB208_245; $L__BB208_240: ld.const.u64 %rd770, [kernel_params+2408]; cvt.u32.u64 %r491, %rd2; cvta.to.global.u64 %rd771, %rd770; mul.wide.u32 %rd772, %r257, 32; add.s64 %rd773, %rd771, %rd772; ld.global.u32 %r492, [%rd773+16]; and.b32 %r493, %r492, 1073741824; setp.eq.s32 %p274, %r493, 0; ld.const.u32 %r494, [kernel_params+4392]; div.u32 %r495, %r491, %r494; ld.const.u32 %r496, [kernel_params+8]; mad.lo.s32 %r258, %r495, %r496, %r257; ld.const.u64 %rd774, [kernel_params+4256]; cvta.to.global.u64 %rd130, %rd774; add.s64 %rd775, %rd130, 20; atom.global.add.u32 %r497, [%rd775], -1; cvt.s64.s32 %rd131, %r258; @%p274 bra $L__BB208_242; bra.uni $L__BB208_241; $L__BB208_242: add.s64 %rd789, %rd130, 32; atom.global.add.u32 %r500, [%rd789], 1; ld.const.u64 %rd790, [kernel_params+2520]; cvta.to.global.u64 %rd791, %rd790; shl.b64 %rd792, %rd2, 1; add.s64 %rd793, %rd791, %rd792; mov.u16 %rs37, 8; st.global.u16 [%rd793], %rs37; ld.const.u64 %rd794, [kernel_params+2624]; cvta.to.global.u64 %rd795, %rd794; add.s64 %rd797, %rd795, %rd738; st.global.u32 [%rd797], %r258; ld.const.u64 %rd798, [kernel_params+4328]; cvta.to.global.u64 %rd799, %rd798; shl.b64 %rd800, %rd131, 2; add.s64 %rd801, %rd799, %rd800; atom.global.add.u32 %r501, [%rd801], 1; bra.uni $L__BB208_245; $L__BB208_241: add.s64 %rd776, %rd130, 36; atom.global.add.u32 %r498, [%rd776], 1; ld.const.u64 %rd777, [kernel_params+2520]; cvta.to.global.u64 %rd778, %rd777; shl.b64 %rd779, %rd2, 1; add.s64 %rd780, %rd778, %rd779; mov.u16 %rs36, 9; st.global.u16 [%rd780], %rs36; ld.const.u64 %rd781, [kernel_params+2624]; cvta.to.global.u64 %rd782, %rd781; add.s64 %rd784, %rd782, %rd738; st.global.u32 [%rd784], %r258; ld.const.u64 %rd785, [kernel_params+4336]; cvta.to.global.u64 %rd786, %rd785; shl.b64 %rd787, %rd131, 2; add.s64 %rd788, %rd786, %rd787; atom.global.add.u32 %r499, [%rd788], 1; $L__BB208_245: ret; } // .globl kernel_gpu_integrator_shade_background .visible .entry kernel_gpu_integrator_shade_background( .param .u64 kernel_gpu_integrator_shade_background_param_0, .param .u64 kernel_gpu_integrator_shade_background_param_1, .param .u32 kernel_gpu_integrator_shade_background_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot209[1328]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<530>; .reg .b16 %rs<64>; .reg .f32 %f<957>; .reg .b32 %r<662>; .reg .b64 %rd<806>; mov.u64 %SPL, __local_depot209; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd72, [kernel_gpu_integrator_shade_background_param_0]; ld.param.u64 %rd73, [kernel_gpu_integrator_shade_background_param_1]; ld.param.u32 %r279, [kernel_gpu_integrator_shade_background_param_2]; cvta.to.global.u64 %rd1, %rd73; mov.u32 %r280, %ntid.x; mov.u32 %r281, %ctaid.x; mov.u32 %r282, %tid.x; mad.lo.s32 %r646, %r281, %r280, %r282; setp.ge.s32 %p6, %r646, %r279; @%p6 bra $L__BB209_627; setp.eq.s64 %p7, %rd72, 0; @%p7 bra $L__BB209_3; cvta.to.global.u64 %rd74, %rd72; mul.wide.s32 %rd75, %r646, 4; add.s64 %rd76, %rd74, %rd75; ld.global.u32 %r646, [%rd76]; $L__BB209_3: ld.const.u64 %rd77, [kernel_params+2640]; cvta.to.global.u64 %rd78, %rd77; cvt.s64.s32 %rd2, %r646; mul.wide.s32 %rd79, %r646, 12; add.s64 %rd3, %rd78, %rd79; ld.global.f32 %f886, [%rd3]; ld.global.f32 %f885, [%rd3+4]; ld.global.f32 %f884, [%rd3+8]; ld.const.u64 %rd80, [kernel_params+2664]; cvta.to.global.u64 %rd81, %rd80; mul.wide.s32 %rd82, %r646, 4; add.s64 %rd4, %rd81, %rd82; ld.const.u32 %r4, [kernel_params+1496]; setp.lt.s32 %p8, %r4, 1; @%p8 bra $L__BB209_277; ld.global.u32 %r5, [%rd4]; ld.const.u64 %rd83, [kernel_params+2368]; cvta.to.global.u64 %rd5, %rd83; ld.const.v2.u64 {%rd84, %rd85}, [kernel_params+2544]; mov.u32 %r647, 0; cvta.to.global.u64 %rd88, %rd84; shl.b64 %rd89, %rd2, 2; add.s64 %rd6, %rd88, %rd89; neg.ftz.f32 %f874, %f886; neg.ftz.f32 %f875, %f885; neg.ftz.f32 %f890, %f884; ld.const.f32 %f7, [kernel_params+1504]; cvta.to.global.u64 %rd90, %rd85; add.s64 %rd7, %rd90, %rd2; ld.const.v2.u64 {%rd91, %rd92}, [kernel_params+2400]; ld.const.u64 %rd93, [kernel_params+2560]; cvta.to.global.u64 %rd94, %rd93; add.s64 %rd10, %rd94, %rd89; ld.const.u64 %rd95, [kernel_params+2584]; cvta.to.global.u64 %rd96, %rd95; mul.lo.s64 %rd97, %rd2, 12; add.s64 %rd11, %rd96, %rd97; ld.const.u64 %rd98, [kernel_params+2464]; cvta.to.global.u64 %rd99, %rd98; shl.b64 %rd100, %rd2, 1; add.s64 %rd12, %rd99, %rd100; ld.const.f32 %f8, [kernel_params+1584]; ld.const.f32 %f9, [kernel_params+1580]; ld.const.v2.u64 {%rd101, %rd102}, [kernel_params+2448]; cvta.to.global.u64 %rd105, %rd101; add.s64 %rd13, %rd105, %rd89; ld.const.v2.u32 {%r285, %r286}, [kernel_params+1264]; cvta.to.global.u64 %rd106, %rd102; add.s64 %rd14, %rd106, %rd89; ld.const.u32 %r8, [kernel_params+1620]; ld.const.u32 %r9, [kernel_params+1260]; ld.const.v4.u32 {%r287, %r288, %r289, %r290}, [kernel_params+1344]; ld.const.v4.u32 {%r291, %r292, %r293, %r294}, [kernel_params+1376]; ld.const.u32 %r295, [kernel_params+1076]; setp.eq.s32 %p9, %r295, -1; ld.const.s32 %rd18, [kernel_params+1444]; setp.eq.s64 %p10, %rd18, -1; or.pred %p1, %p9, %p10; ld.const.v2.u64 {%rd107, %rd108}, [kernel_params+2608]; cvta.to.global.u64 %rd111, %rd108; add.s64 %rd15, %rd111, %rd97; ld.const.s32 %rd16, [kernel_params+1428]; ld.const.s32 %rd17, [kernel_params+1400]; mul.lo.s32 %r296, %r295, 3; cvt.s64.s32 %rd19, %r296; ld.const.u32 %r297, [kernel_params]; and.b32 %r18, %r297, 2097152; ld.const.u64 %rd112, [kernel_params+2600]; cvta.to.global.u64 %rd113, %rd112; add.s64 %rd20, %rd113, %rd97; cvta.to.global.u64 %rd114, %rd107; add.s64 %rd21, %rd114, %rd97; ld.const.v4.u32 {%r298, %r299, %r300, %r301}, [kernel_params+1328]; ld.const.v2.u32 {%r302, %r303}, [kernel_params+1320]; mov.pred %p11, -1; mov.pred %p19, 0; cvta.to.global.u64 %rd116, %rd92; cvta.to.global.u64 %rd27, %rd91; setp.eq.s32 %p193, %r8, 0; bra.uni $L__BB209_5; $L__BB209_266: ld.global.f32 %f95, [%rd20]; ld.global.f32 %f96, [%rd20+4]; ld.global.f32 %f97, [%rd20+8]; ld.global.f32 %f98, [%rd21]; ld.global.f32 %f99, [%rd21+4]; ld.global.f32 %f100, [%rd21+8]; setp.eq.s16 %p210, %rs1, 1; selp.b32 %r130, %r301, %r303, %p210; setp.eq.s32 %p211, %r130, -1; @%p211 bra $L__BB209_268; cvt.s64.s32 %rd318, %r130; add.s64 %rd319, %rd34, %rd318; shl.b64 %rd320, %rd319, 2; add.s64 %rd321, %rd1, %rd320; add.s64 %rd322, %rd321, 4; add.s64 %rd323, %rd321, 8; mul.ftz.f32 %f361, %f915, %f98; atom.global.add.f32 %f362, [%rd321], %f361; mul.ftz.f32 %f363, %f916, %f99; atom.global.add.f32 %f364, [%rd322], %f363; mul.ftz.f32 %f365, %f917, %f100; atom.global.add.f32 %f366, [%rd323], %f365; $L__BB209_268: setp.eq.s16 %p525, %rs1, 1; selp.b32 %r131, %r287, %r298, %p525; setp.eq.s32 %p213, %r131, -1; @%p213 bra $L__BB209_270; mov.f32 %f367, 0f3F800000; sub.ftz.f32 %f368, %f367, %f95; sub.ftz.f32 %f369, %f368, %f98; sub.ftz.f32 %f370, %f367, %f96; sub.ftz.f32 %f371, %f370, %f99; sub.ftz.f32 %f372, %f367, %f97; sub.ftz.f32 %f373, %f372, %f100; cvt.s64.s32 %rd324, %r131; add.s64 %rd325, %rd34, %rd324; shl.b64 %rd326, %rd325, 2; add.s64 %rd327, %rd1, %rd326; mul.ftz.f32 %f374, %f915, %f369; mul.ftz.f32 %f375, %f916, %f371; mul.ftz.f32 %f376, %f917, %f373; add.s64 %rd328, %rd327, 4; add.s64 %rd329, %rd327, 8; atom.global.add.f32 %f377, [%rd327], %f374; atom.global.add.f32 %f378, [%rd328], %f375; atom.global.add.f32 %f379, [%rd329], %f376; $L__BB209_270: setp.eq.s16 %p526, %rs1, 1; mov.u32 %r653, -1; selp.b32 %r132, %r300, %r302, %p526; setp.eq.s32 %p215, %r132, -1; @%p215 bra $L__BB209_274; mul.ftz.f32 %f915, %f915, %f95; mul.ftz.f32 %f916, %f916, %f96; mul.ftz.f32 %f917, %f917, %f97; mov.u32 %r653, %r132; bra.uni $L__BB209_274; $L__BB209_5: mul.wide.s32 %rd115, %r647, 192; add.s64 %rd22, %rd5, %rd115; ld.global.u32 %r304, [%rd22]; setp.ne.s32 %p12, %r304, 1; mov.pred %p527, %p11; @%p12 bra $L__BB209_9; ld.global.u32 %r28, [%rd22+16]; and.b32 %r305, %r28, 268435456; setp.eq.s32 %p14, %r305, 0; ld.global.f32 %f243, [%rd22+144]; setp.eq.ftz.f32 %p15, %f243, 0f00000000; or.pred %p16, %p15, %p14; mov.pred %p527, %p11; @%p16 bra $L__BB209_9; ld.global.f32 %f244, [%rd22+4]; mul.ftz.f32 %f245, %f886, %f244; ld.global.f32 %f246, [%rd22+8]; mul.ftz.f32 %f247, %f885, %f246; neg.ftz.f32 %f248, %f247; sub.ftz.f32 %f249, %f248, %f245; ld.global.f32 %f250, [%rd22+12]; mul.ftz.f32 %f251, %f884, %f250; sub.ftz.f32 %f21, %f249, %f251; ld.global.f32 %f252, [%rd22+148]; setp.lt.ftz.f32 %p18, %f21, %f252; mov.pred %p527, %p11; @%p18 bra $L__BB209_9; mul.ftz.f32 %f253, %f21, %f21; mul.ftz.f32 %f254, %f21, %f253; ld.global.f32 %f255, [%rd22+152]; div.approx.ftz.f32 %f882, %f255, %f254; mul.ftz.f32 %f883, %f882, %f7; mov.u32 %r650, %r647; mov.u32 %r651, %r28; mov.f32 %f876, %f890; mov.pred %p527, %p19; $L__BB209_9: @%p527 bra $L__BB209_276; ld.global.u32 %r31, [%rd6]; and.b32 %r306, %r651, 264241152; setp.eq.s32 %p20, %r306, 0; @%p20 bra $L__BB209_16; shr.u32 %r307, %r651, 27; and.b32 %r308, %r307, 1; setp.eq.b32 %p21, %r308, 1; shr.u32 %r309, %r31, 3; and.b32 %r310, %r309, 1; setp.eq.b32 %p22, %r310, 1; and.pred %p23, %p21, %p22; @%p23 bra $L__BB209_277; shr.u32 %r311, %r651, 26; and.b32 %r312, %r311, 1; setp.eq.b32 %p24, %r312, 1; and.b32 %r313, %r31, 18; setp.eq.s32 %p25, %r313, 18; and.pred %p26, %p24, %p25; @%p26 bra $L__BB209_277; shr.u32 %r314, %r651, 25; and.b32 %r315, %r314, 1; setp.eq.b32 %p27, %r315, 1; shr.u32 %r316, %r31, 2; and.b32 %r317, %r316, 1; setp.eq.b32 %p28, %r317, 1; and.pred %p29, %p27, %p28; @%p29 bra $L__BB209_277; shr.u32 %r318, %r651, 24; and.b32 %r319, %r318, 1; setp.eq.b32 %p30, %r319, 1; and.b32 %r320, %r31, 1; setp.eq.b32 %p31, %r320, 1; and.pred %p32, %p30, %p31; @%p32 bra $L__BB209_277; shr.u32 %r321, %r651, 23; and.b32 %r322, %r321, 1; setp.eq.b32 %p33, %r322, 1; shr.u32 %r323, %r31, 7; and.b32 %r324, %r323, 1; setp.eq.b32 %p34, %r324, 1; and.pred %p35, %p33, %p34; @%p35 bra $L__BB209_277; $L__BB209_16: ld.global.u8 %rs13, [%rd7]; and.b16 %rs14, %rs13, 4; setp.eq.s16 %p36, %rs14, 0; @%p36 bra $L__BB209_18; ld.global.u32 %r325, [%rd22+40]; setp.ne.s32 %p37, %r325, 0; @%p37 bra $L__BB209_277; $L__BB209_18: shl.b32 %r326, %r651, 5; cvt.u64.u32 %rd117, %r326; and.b64 %rd118, %rd117, 134217696; add.s64 %rd119, %rd116, %rd118; ld.global.u32 %r33, [%rd119+16]; and.b32 %r327, %r33, 134217728; setp.eq.s32 %p39, %r327, 0; mov.f32 %f899, 0f00000000; mov.f32 %f900, %f899; mov.f32 %f901, %f899; mov.pred %p528, %p11; @%p39 bra $L__BB209_20; shl.b32 %r644, %r651, 5; cvt.u64.u32 %rd805, %r644; and.b64 %rd804, %rd805, 134217696; add.s64 %rd803, %rd116, %rd804; add.s64 %rd802, %rd803, 16; ld.global.f32 %f901, [%rd802+-16]; ld.global.f32 %f900, [%rd802+-12]; ld.global.f32 %f899, [%rd802+-8]; mov.pred %p528, 0; $L__BB209_20: not.pred %p41, %p528; @%p41 bra $L__BB209_231; add.u64 %rd120, %SP, 0; add.u64 %rd24, %SPL, 0; st.local.v4.f32 [%rd24], {%f874, %f875, %f890, %f874}; add.s64 %rd121, %rd120, 16; cvta.to.local.u64 %rd122, %rd121; st.local.v4.f32 [%rd122], {%f875, %f876, %f874, %f875}; neg.ftz.f32 %f48, %f884; neg.ftz.f32 %f47, %f885; neg.ftz.f32 %f46, %f886; add.s64 %rd123, %rd120, 32; cvta.to.local.u64 %rd25, %rd123; st.local.v4.f32 [%rd25], {%f876, %f46, %f47, %f48}; setp.eq.s32 %p42, %r650, -1; selp.b32 %r328, 0, 32, %p42; mov.u32 %r329, -1; mov.u32 %r330, 0; add.s64 %rd124, %rd120, 64; cvta.to.local.u64 %rd125, %rd124; st.local.v4.u32 [%rd125], {%r328, %r330, %r330, %r329}; add.s64 %rd126, %rd120, 80; cvta.to.local.u64 %rd127, %rd126; st.local.v2.u32 [%rd127], {%r329, %r5}; add.s64 %rd128, %rd120, 88; cvta.to.local.u64 %rd129, %rd128; mov.u32 %r331, 2139095039; st.local.u32 [%rd129], %r331; add.s64 %rd130, %rd120, 48; cvta.to.local.u64 %rd131, %rd130; st.local.v4.u32 [%rd131], {%r651, %r33, %r330, %r329}; @%p42 bra $L__BB209_23; st.local.u32 [%rd127], %r650; $L__BB209_23: and.b32 %r652, %r651, 4194303; add.s64 %rd136, %rd120, 128; cvta.to.local.u64 %rd137, %rd136; mov.f32 %f259, 0f00000000; st.local.v2.f32 [%rd137], {%f259, %f259}; add.s64 %rd138, %rd120, 136; cvta.to.local.u64 %rd139, %rd138; st.local.u32 [%rd139], %r330; add.s64 %rd140, %rd120, 92; cvta.to.local.u64 %rd141, %rd140; st.local.u32 [%rd141], %r330; add.s64 %rd142, %rd120, 96; cvta.to.local.u64 %rd143, %rd142; st.local.v4.f32 [%rd143], {%f259, %f259, %f259, %f259}; add.s64 %rd144, %rd120, 112; cvta.to.local.u64 %rd145, %rd144; st.local.v4.f32 [%rd145], {%f259, %f259, %f259, %f259}; add.s64 %rd146, %rd120, 260; cvta.to.local.u64 %rd147, %rd146; st.local.u32 [%rd147], %r330; add.s64 %rd148, %rd120, 264; cvta.to.local.u64 %rd149, %rd148; st.local.u32 [%rd149], %r330; add.u64 %rd150, %SP, 304; mov.b32 %r333, %f876; mov.b64 %rd151, {%r333, %r334}; cvt.u32.u64 %r335, %rd151; mov.b32 %f876, %r335; bra.uni $L__BB209_24; $L__BB209_176: { // callseq 980, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 980 $L__BB209_24: mov.u32 %r34, %r652; add.u64 %rd777, %SPL, 304; mul.wide.s32 %rd152, %r34, 16; add.s64 %rd28, %rd27, %rd152; ld.global.v4.u32 {%r337, %r338, %r339, %r340}, [%rd28]; mul.wide.u32 %rd153, %r338, 4; add.s64 %rd29, %rd777, %rd153; mul.wide.u32 %rd154, %r339, 4; add.s64 %rd30, %rd777, %rd154; add.s32 %r652, %r34, 1; setp.gt.s32 %p43, %r337, 36; @%p43 bra $L__BB209_73; setp.gt.s32 %p125, %r337, 25; @%p125 bra $L__BB209_69; bra.uni $L__BB209_26; $L__BB209_69: add.s32 %r342, %r337, -27; setp.lt.u32 %p126, %r342, 9; @%p126 bra $L__BB209_24; setp.eq.s32 %p127, %r337, 26; @%p127 bra $L__BB209_215; setp.eq.s32 %p128, %r337, 36; @%p128 bra $L__BB209_72; bra.uni $L__BB209_229; $L__BB209_72: mov.u32 %r384, 36; { // callseq 1033, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r384; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1033 bra.uni $L__BB209_24; $L__BB209_73: setp.gt.s32 %p44, %r337, 68; @%p44 bra $L__BB209_128; bra.uni $L__BB209_74; $L__BB209_128: setp.gt.s32 %p45, %r337, 88; @%p45 bra $L__BB209_164; bra.uni $L__BB209_129; $L__BB209_164: setp.gt.s32 %p46, %r337, 93; @%p46 bra $L__BB209_169; bra.uni $L__BB209_165; $L__BB209_169: setp.gt.s32 %p47, %r337, 95; @%p47 bra $L__BB209_173; bra.uni $L__BB209_170; $L__BB209_173: setp.eq.s32 %p48, %r337, 98; @%p48 bra $L__BB209_228; setp.eq.s32 %p49, %r337, 97; @%p49 bra $L__BB209_178; setp.ne.s32 %p50, %r337, 96; @%p50 bra $L__BB209_229; bra.uni $L__BB209_176; $L__BB209_74: setp.gt.s32 %p83, %r337, 48; @%p83 bra $L__BB209_93; bra.uni $L__BB209_75; $L__BB209_93: setp.gt.s32 %p84, %r337, 58; @%p84 bra $L__BB209_111; setp.gt.s32 %p98, %r337, 53; @%p98 bra $L__BB209_103; setp.gt.s32 %p105, %r337, 50; @%p105 bra $L__BB209_99; bra.uni $L__BB209_96; $L__BB209_99: setp.eq.s32 %p106, %r337, 51; @%p106 bra $L__BB209_207; setp.eq.s32 %p107, %r337, 52; @%p107 bra $L__BB209_206; setp.eq.s32 %p108, %r337, 53; @%p108 bra $L__BB209_102; bra.uni $L__BB209_229; $L__BB209_102: { // callseq 1018, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1018 bra.uni $L__BB209_24; $L__BB209_26: setp.gt.s32 %p129, %r337, 17; @%p129 bra $L__BB209_60; bra.uni $L__BB209_27; $L__BB209_60: setp.lt.s32 %p130, %r337, 21; @%p130 bra $L__BB209_24; setp.gt.s32 %p131, %r337, 22; @%p131 bra $L__BB209_65; bra.uni $L__BB209_62; $L__BB209_65: setp.eq.s32 %p132, %r337, 23; @%p132 bra $L__BB209_217; setp.eq.s32 %p133, %r337, 24; @%p133 bra $L__BB209_216; setp.eq.s32 %p134, %r337, 25; @%p134 bra $L__BB209_68; bra.uni $L__BB209_229; $L__BB209_68: add.s32 %r386, %r34, 1; { // callseq 1035, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r386; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1035 bra.uni $L__BB209_24; $L__BB209_129: setp.gt.s32 %p56, %r337, 78; @%p56 bra $L__BB209_147; setp.gt.s32 %p70, %r337, 73; @%p70 bra $L__BB209_139; setp.gt.s32 %p77, %r337, 70; @%p77 bra $L__BB209_135; bra.uni $L__BB209_132; $L__BB209_135: setp.eq.s32 %p78, %r337, 71; @%p78 bra $L__BB209_195; setp.eq.s32 %p79, %r337, 72; @%p79 bra $L__BB209_194; setp.eq.s32 %p80, %r337, 73; @%p80 bra $L__BB209_138; bra.uni $L__BB209_229; $L__BB209_138: add.s32 %r355, %r34, 1; { // callseq 998, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r355; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 998 bra.uni $L__BB209_24; $L__BB209_75: setp.gt.s32 %p111, %r337, 41; @%p111 bra $L__BB209_81; setp.gt.s32 %p121, %r337, 39; @%p121 bra $L__BB209_24; setp.eq.s32 %p122, %r337, 37; @%p122 bra $L__BB209_214; setp.eq.s32 %p123, %r337, 38; @%p123 bra $L__BB209_213; setp.eq.s32 %p124, %r337, 39; @%p124 bra $L__BB209_80; bra.uni $L__BB209_229; $L__BB209_80: mov.u32 %r382, 39; { // callseq 1030, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r382, %r338, %r339, %r340}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1030 bra.uni $L__BB209_24; $L__BB209_27: setp.gt.s32 %p137, %r337, 8; @%p137 bra $L__BB209_43; bra.uni $L__BB209_28; $L__BB209_43: setp.gt.s32 %p138, %r337, 12; @%p138 bra $L__BB209_52; bra.uni $L__BB209_44; $L__BB209_52: setp.gt.s32 %p139, %r337, 14; @%p139 bra $L__BB209_56; bra.uni $L__BB209_53; $L__BB209_56: setp.eq.s32 %p140, %r337, 15; @%p140 bra $L__BB209_220; setp.eq.s32 %p141, %r337, 16; @%p141 bra $L__BB209_219; setp.eq.s32 %p142, %r337, 17; @%p142 bra $L__BB209_59; bra.uni $L__BB209_229; $L__BB209_59: { // callseq 1040, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; .param .b32 param4; st.param.b32 [param4+0], %r340; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1040 bra.uni $L__BB209_24; $L__BB209_111: setp.gt.s32 %p85, %r337, 63; @%p85 bra $L__BB209_120; setp.gt.s32 %p92, %r337, 60; @%p92 bra $L__BB209_116; bra.uni $L__BB209_113; $L__BB209_116: setp.eq.s32 %p93, %r337, 61; @%p93 bra $L__BB209_201; setp.eq.s32 %p94, %r337, 62; @%p94 bra $L__BB209_200; setp.eq.s32 %p95, %r337, 63; @%p95 bra $L__BB209_119; bra.uni $L__BB209_229; $L__BB209_119: mov.u32 %r365, 63; { // callseq 1008, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r365; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1008 bra.uni $L__BB209_24; $L__BB209_147: setp.gt.s32 %p57, %r337, 83; @%p57 bra $L__BB209_156; setp.gt.s32 %p64, %r337, 80; @%p64 bra $L__BB209_152; bra.uni $L__BB209_149; $L__BB209_152: setp.eq.s32 %p65, %r337, 81; @%p65 bra $L__BB209_183; setp.eq.s32 %p66, %r337, 82; @%p66 bra $L__BB209_182; setp.eq.s32 %p67, %r337, 83; @%p67 bra $L__BB209_155; bra.uni $L__BB209_229; $L__BB209_155: { // callseq 990, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 990 bra.uni $L__BB209_24; $L__BB209_28: setp.gt.s32 %p150, %r337, 4; @%p150 bra $L__BB209_36; setp.gt.s32 %p156, %r337, 2; @%p156 bra $L__BB209_33; setp.eq.s32 %p159, %r337, 1; mov.u32 %r652, %r338; @%p159 bra $L__BB209_24; setp.eq.s32 %p160, %r337, 2; @%p160 bra $L__BB209_32; bra.uni $L__BB209_229; $L__BB209_32: mov.u32 %r410, 2; add.s32 %r411, %r34, 1; { // callseq 1049, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r410; st.param.b32 [param0+4], %r338; st.param.b32 [param0+8], %r339; st.param.b32 [param0+12], %r340; .param .b32 param1; st.param.b32 [param1+0], %r411; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1049 bra.uni $L__BB209_24; $L__BB209_81: setp.gt.s32 %p112, %r337, 44; @%p112 bra $L__BB209_86; bra.uni $L__BB209_82; $L__BB209_86: setp.gt.s32 %p113, %r337, 46; @%p113 bra $L__BB209_90; setp.eq.s32 %p116, %r337, 45; @%p116 bra $L__BB209_210; setp.eq.s32 %p117, %r337, 46; @%p117 bra $L__BB209_89; bra.uni $L__BB209_229; $L__BB209_89: { // callseq 1025, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1025 bra.uni $L__BB209_24; $L__BB209_170: setp.eq.s32 %p51, %r337, 94; @%p51 bra $L__BB209_177; setp.eq.s32 %p52, %r337, 95; @%p52 bra $L__BB209_172; bra.uni $L__BB209_229; $L__BB209_172: { // callseq 981, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 981 bra.uni $L__BB209_24; $L__BB209_165: add.s32 %r341, %r337, -92; setp.lt.u32 %p53, %r341, 2; @%p53 bra $L__BB209_24; setp.eq.s32 %p54, %r337, 90; @%p54 bra $L__BB209_24; setp.eq.s32 %p55, %r337, 89; @%p55 bra $L__BB209_168; bra.uni $L__BB209_229; $L__BB209_168: mov.u32 %r345, 89; { // callseq 984, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r345; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 984 bra.uni $L__BB209_24; $L__BB209_103: setp.gt.s32 %p99, %r337, 55; @%p99 bra $L__BB209_107; bra.uni $L__BB209_104; $L__BB209_107: setp.eq.s32 %p100, %r337, 56; @%p100 bra $L__BB209_204; setp.eq.s32 %p101, %r337, 57; @%p101 bra $L__BB209_203; setp.eq.s32 %p102, %r337, 58; @%p102 bra $L__BB209_110; bra.uni $L__BB209_229; $L__BB209_110: mov.u32 %r372, 58; { // callseq 1013, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r372; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1013 bra.uni $L__BB209_24; $L__BB209_139: setp.gt.s32 %p71, %r337, 75; @%p71 bra $L__BB209_143; bra.uni $L__BB209_140; $L__BB209_143: setp.eq.s32 %p72, %r337, 76; @%p72 bra $L__BB209_187; setp.eq.s32 %p73, %r337, 77; @%p73 bra $L__BB209_185; setp.eq.s32 %p74, %r337, 78; @%p74 bra $L__BB209_146; bra.uni $L__BB209_229; $L__BB209_146: add.s32 %r353, %r34, 1; { // callseq 995, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r353; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 995 bra.uni $L__BB209_24; $L__BB209_120: setp.gt.s32 %p86, %r337, 65; @%p86 bra $L__BB209_124; bra.uni $L__BB209_121; $L__BB209_124: setp.eq.s32 %p87, %r337, 66; @%p87 bra $L__BB209_198; setp.eq.s32 %p88, %r337, 67; @%p88 bra $L__BB209_197; setp.eq.s32 %p89, %r337, 68; @%p89 bra $L__BB209_127; bra.uni $L__BB209_229; $L__BB209_127: mov.u32 %r360, 68; { // callseq 1003, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r360; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1003 bra.uni $L__BB209_24; $L__BB209_156: setp.gt.s32 %p58, %r337, 85; @%p58 bra $L__BB209_160; bra.uni $L__BB209_157; $L__BB209_160: setp.eq.s32 %p59, %r337, 86; @%p59 bra $L__BB209_180; setp.eq.s32 %p60, %r337, 87; @%p60 bra $L__BB209_179; setp.eq.s32 %p61, %r337, 88; @%p61 bra $L__BB209_163; bra.uni $L__BB209_229; $L__BB209_163: cvta.to.local.u64 %rd780, %rd120; mov.u32 %r346, 88; ld.local.f32 %f260, [%rd780+12]; ld.local.v2.f32 {%f261, %f262}, [%rd780+16]; { // callseq 985, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f260; .param .b32 param1; st.param.f32 [param1+0], %f261; .param .b32 param2; st.param.f32 [param2+0], %f262; .param .b64 param3; st.param.b64 [param3+0], %rd150; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r346; st.param.b32 [param4+4], %r338; st.param.b32 [param4+8], %r339; st.param.b32 [param4+12], %r340; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 985 bra.uni $L__BB209_24; $L__BB209_44: setp.gt.s32 %p145, %r337, 10; @%p145 bra $L__BB209_49; setp.eq.s32 %p148, %r337, 9; @%p148 bra $L__BB209_223; setp.eq.s32 %p149, %r337, 10; @%p149 bra $L__BB209_47; bra.uni $L__BB209_229; $L__BB209_47: ld.local.f32 %f276, [%rd30]; setp.ltu.ftz.f32 %p165, %f276, 0f3F800000; @%p165 bra $L__BB209_24; add.s32 %r404, %r338, %r34; add.s32 %r652, %r404, 1; bra.uni $L__BB209_24; $L__BB209_36: setp.gt.s32 %p151, %r337, 6; @%p151 bra $L__BB209_40; setp.eq.s32 %p154, %r337, 5; @%p154 bra $L__BB209_226; setp.eq.s32 %p155, %r337, 6; @%p155 bra $L__BB209_39; bra.uni $L__BB209_229; $L__BB209_39: ld.local.f32 %f278, [%rd29]; ld.local.f32 %f279, [%rd29+4]; ld.local.f32 %f280, [%rd29+8]; add.s64 %rd251, %rd120, 268; cvta.to.local.u64 %rd252, %rd251; st.local.f32 [%rd252], %f278; add.s64 %rd253, %rd120, 272; cvta.to.local.u64 %rd254, %rd253; st.local.f32 [%rd254], %f279; add.s64 %rd255, %rd120, 276; cvta.to.local.u64 %rd256, %rd255; st.local.f32 [%rd256], %f280; bra.uni $L__BB209_24; $L__BB209_82: setp.eq.s32 %p118, %r337, 42; @%p118 bra $L__BB209_212; setp.eq.s32 %p119, %r337, 43; @%p119 bra $L__BB209_211; setp.eq.s32 %p120, %r337, 44; @%p120 bra $L__BB209_85; bra.uni $L__BB209_229; $L__BB209_85: mov.u32 %r379, 44; add.s32 %r380, %r34, 1; { // callseq 1027, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r379; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r380; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1027 bra.uni $L__BB209_24; $L__BB209_228: { // callseq 1050, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1050 bra.uni $L__BB209_24; $L__BB209_178: { // callseq 983, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 983 bra.uni $L__BB209_24; $L__BB209_177: mov.u32 %r343, 94; add.s32 %r344, %r34, 1; { // callseq 982, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r343; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r344; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 982 bra.uni $L__BB209_24; $L__BB209_62: setp.eq.s32 %p135, %r337, 21; @%p135 bra $L__BB209_218; setp.eq.s32 %p136, %r337, 22; @%p136 bra $L__BB209_64; bra.uni $L__BB209_229; $L__BB209_64: mov.u32 %r390, 22; add.s32 %r391, %r34, 1; { // callseq 1038, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r390; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r391; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1038 bra.uni $L__BB209_24; $L__BB209_53: setp.eq.s32 %p143, %r337, 13; @%p143 bra $L__BB209_221; setp.eq.s32 %p144, %r337, 14; @%p144 bra $L__BB209_55; bra.uni $L__BB209_229; $L__BB209_55: st.local.u32 [%rd30], %r338; bra.uni $L__BB209_24; $L__BB209_33: setp.eq.s32 %p157, %r337, 3; @%p157 bra $L__BB209_227; setp.eq.s32 %p158, %r337, 4; @%p158 bra $L__BB209_35; bra.uni $L__BB209_229; $L__BB209_35: mov.u32 %r408, 4; { // callseq 1047, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r408, %r338, %r339, %r340}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1047 bra.uni $L__BB209_24; $L__BB209_49: setp.eq.s32 %p146, %r337, 11; @%p146 bra $L__BB209_222; setp.eq.s32 %p147, %r337, 12; @%p147 bra $L__BB209_51; bra.uni $L__BB209_229; $L__BB209_51: { // callseq 1043, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1043 bra.uni $L__BB209_24; $L__BB209_40: setp.eq.s32 %p152, %r337, 7; @%p152 bra $L__BB209_225; setp.eq.s32 %p153, %r337, 8; @%p153 bra $L__BB209_42; bra.uni $L__BB209_229; $L__BB209_42: mov.u32 %r406, 8; { // callseq 1045, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r406; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1045 bra.uni $L__BB209_24; $L__BB209_96: setp.eq.s32 %p109, %r337, 49; @%p109 bra $L__BB209_208; setp.eq.s32 %p110, %r337, 50; @%p110 bra $L__BB209_98; bra.uni $L__BB209_229; $L__BB209_98: { // callseq 1021, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1021 bra.uni $L__BB209_24; $L__BB209_132: setp.eq.s32 %p81, %r337, 69; @%p81 bra $L__BB209_196; setp.eq.s32 %p82, %r337, 70; @%p82 bra $L__BB209_134; bra.uni $L__BB209_229; $L__BB209_134: mov.u32 %r357, 70; { // callseq 1001, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r357; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1001 bra.uni $L__BB209_24; $L__BB209_113: setp.eq.s32 %p96, %r337, 59; @%p96 bra $L__BB209_202; setp.eq.s32 %p97, %r337, 60; @%p97 bra $L__BB209_115; bra.uni $L__BB209_229; $L__BB209_115: add.s32 %r370, %r34, 1; { // callseq 1011, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r370; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1011 bra.uni $L__BB209_24; $L__BB209_149: setp.eq.s32 %p68, %r337, 79; @%p68 bra $L__BB209_184; setp.eq.s32 %p69, %r337, 80; @%p69 bra $L__BB209_151; bra.uni $L__BB209_229; $L__BB209_151: { // callseq 993, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 993 bra.uni $L__BB209_24; $L__BB209_90: setp.eq.s32 %p114, %r337, 47; @%p114 bra $L__BB209_209; setp.eq.s32 %p115, %r337, 48; @%p115 bra $L__BB209_92; bra.uni $L__BB209_229; $L__BB209_92: { // callseq 1023, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1023 bra.uni $L__BB209_24; $L__BB209_104: setp.eq.s32 %p103, %r337, 54; @%p103 bra $L__BB209_205; setp.eq.s32 %p104, %r337, 55; @%p104 bra $L__BB209_106; bra.uni $L__BB209_229; $L__BB209_106: cvta.to.local.u64 %rd785, %rd120; ld.local.v4.f32 {%f269, %f270, %f271, %f272}, [%rd785]; { // callseq 1016, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f269; st.param.f32 [param0+4], %f270; st.param.f32 [param0+8], %f271; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; .param .b32 param4; st.param.b32 [param4+0], %r340; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1016 bra.uni $L__BB209_24; $L__BB209_140: setp.eq.s32 %p75, %r337, 74; @%p75 bra $L__BB209_193; setp.eq.s32 %p76, %r337, 75; @%p76 bra $L__BB209_142; bra.uni $L__BB209_229; $L__BB209_142: { // callseq 996, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 996 bra.uni $L__BB209_24; $L__BB209_121: setp.eq.s32 %p90, %r337, 64; @%p90 bra $L__BB209_199; setp.eq.s32 %p91, %r337, 65; @%p91 bra $L__BB209_123; bra.uni $L__BB209_229; $L__BB209_123: { // callseq 1006, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1006 bra.uni $L__BB209_24; $L__BB209_157: setp.eq.s32 %p62, %r337, 84; @%p62 bra $L__BB209_181; setp.eq.s32 %p63, %r337, 85; @%p63 bra $L__BB209_159; bra.uni $L__BB209_229; $L__BB209_159: add.s32 %r349, %r34, 1; { // callseq 988, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r349; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 988 bra.uni $L__BB209_24; $L__BB209_217: mov.u32 %r388, 23; add.s32 %r389, %r34, 1; { // callseq 1037, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r388; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r389; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1037 bra.uni $L__BB209_24; $L__BB209_216: mov.u32 %r387, 24; { // callseq 1036, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r387; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1036 bra.uni $L__BB209_24; $L__BB209_220: mul.wide.s32 %rd787, %r34, 16; add.s64 %rd786, %rd27, %rd787; ld.global.v4.u32 {%r394, %r395, %r396, %r397}, [%rd786+16]; add.s32 %r652, %r34, 2; st.local.u32 [%rd29], %r395; st.local.u32 [%rd29+4], %r396; st.local.u32 [%rd29+8], %r397; bra.uni $L__BB209_24; $L__BB209_219: mov.u32 %r393, 16; { // callseq 1041, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r393; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1041 bra.uni $L__BB209_24; $L__BB209_223: ld.local.f32 %f277, [%rd30]; setp.gtu.ftz.f32 %p166, %f277, 0f00000000; @%p166 bra $L__BB209_24; add.s32 %r405, %r338, %r34; add.s32 %r652, %r405, 1; bra.uni $L__BB209_24; $L__BB209_226: add.s64 %rd258, %rd120, 268; cvta.to.local.u64 %rd259, %rd258; st.local.u32 [%rd259], %r338; add.s64 %rd260, %rd120, 272; cvta.to.local.u64 %rd261, %rd260; st.local.u32 [%rd261], %r339; add.s64 %rd262, %rd120, 276; cvta.to.local.u64 %rd263, %rd262; st.local.u32 [%rd263], %r340; bra.uni $L__BB209_24; $L__BB209_218: mov.u32 %r392, 21; { // callseq 1039, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r392; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1039 bra.uni $L__BB209_24; $L__BB209_221: mov.u32 %r401, 13; add.s32 %r402, %r34, 1; mov.u32 %r403, 524288; { // callseq 1042, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b32 param1; st.param.b32 [param1+0], %r403; .param .b64 param2; st.param.b64 [param2+0], %rd150; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r401; st.param.b32 [param3+4], %r338; st.param.b32 [param3+8], %r339; st.param.b32 [param3+12], %r340; .param .b32 param4; st.param.b32 [param4+0], %r402; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1042 bra.uni $L__BB209_24; $L__BB209_227: mov.u32 %r409, 3; { // callseq 1048, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r409, %r338, %r339, %r340}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1048 bra.uni $L__BB209_24; $L__BB209_222: { // callseq 1044, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1044 bra.uni $L__BB209_24; $L__BB209_225: mov.u32 %r407, 7; { // callseq 1046, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r407; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1046 bra.uni $L__BB209_24; $L__BB209_207: { // callseq 1020, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1020 bra.uni $L__BB209_24; $L__BB209_206: add.s32 %r377, %r34, 1; { // callseq 1019, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r377; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1019 bra.uni $L__BB209_24; $L__BB209_195: mov.u32 %r356, 71; { // callseq 1000, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r356; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1000 bra.uni $L__BB209_24; $L__BB209_194: { // callseq 999, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 999 bra.uni $L__BB209_24; $L__BB209_214: mov.u32 %r383, 37; { // callseq 1032, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r383, %r338, %r339, %r340}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1032 bra.uni $L__BB209_24; $L__BB209_213: { // callseq 1031, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; .param .b32 param4; st.param.b32 [param4+0], %r340; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1031 bra.uni $L__BB209_24; $L__BB209_201: mov.u32 %r368, 61; add.s32 %r369, %r34, 1; { // callseq 1010, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r368; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r369; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1010 bra.uni $L__BB209_24; $L__BB209_200: mov.u32 %r366, 62; add.s32 %r367, %r34, 1; { // callseq 1009, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r366; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r367; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1009 bra.uni $L__BB209_24; $L__BB209_183: mov.u32 %r351, 81; { // callseq 992, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r351; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 992 bra.uni $L__BB209_24; $L__BB209_182: mov.u32 %r350, 82; { // callseq 991, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r350; st.param.b32 [param2+4], %r338; st.param.b32 [param2+8], %r339; st.param.b32 [param2+12], %r340; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 991 bra.uni $L__BB209_24; $L__BB209_210: { // callseq 1026, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1026 bra.uni $L__BB209_24; $L__BB209_204: mov.u32 %r375, 56; { // callseq 1015, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r375; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1015 bra.uni $L__BB209_24; $L__BB209_203: mov.u32 %r373, 57; add.s32 %r374, %r34, 1; { // callseq 1014, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r373; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r374; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1014 bra.uni $L__BB209_24; $L__BB209_187: setp.eq.s32 %p162, %r340, 255; @%p162 bra $L__BB209_24; add.u64 %rd783, %SPL, 304; setp.eq.s32 %p163, %r339, 0; mul.wide.u32 %rd181, %r340, 4; add.s64 %rd33, %rd783, %rd181; @%p163 bra $L__BB209_192; setp.eq.s32 %p164, %r339, 1; @%p164 bra $L__BB209_191; bra.uni $L__BB209_190; $L__BB209_191: mul.wide.u32 %rd800, %r338, 4; add.u64 %rd799, %SPL, 304; add.s64 %rd798, %rd799, %rd800; add.s64 %rd797, %rd798, 4; ld.local.f32 %f267, [%rd797]; st.local.f32 [%rd33], %f267; bra.uni $L__BB209_24; $L__BB209_185: setp.eq.s32 %p161, %r340, 255; @%p161 bra $L__BB209_24; add.u64 %rd781, %SPL, 304; ld.local.f32 %f265, [%rd29]; add.s32 %r354, %r340, %r339; mul.wide.u32 %rd177, %r354, 4; add.s64 %rd178, %rd781, %rd177; st.local.f32 [%rd178], %f265; bra.uni $L__BB209_24; $L__BB209_198: add.s32 %r362, %r34, 1; { // callseq 1005, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r362; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1005 bra.uni $L__BB209_24; $L__BB209_197: mov.u32 %r361, 67; { // callseq 1004, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r361, %r338, %r339, %r340}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1004 bra.uni $L__BB209_24; $L__BB209_180: add.s32 %r348, %r34, 1; { // callseq 987, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r348; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 987 bra.uni $L__BB209_24; $L__BB209_179: add.s32 %r347, %r34, 1; { // callseq 986, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r347; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 986 bra.uni $L__BB209_24; $L__BB209_212: { // callseq 1029, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1029 bra.uni $L__BB209_24; $L__BB209_211: add.s32 %r381, %r34, 1; { // callseq 1028, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r381; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1028 bra.uni $L__BB209_24; $L__BB209_208: { // callseq 1022, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd120; .param .b64 param1; st.param.b64 [param1+0], %rd150; .param .b32 param2; st.param.b32 [param2+0], %r338; .param .b32 param3; st.param.b32 [param3+0], %r339; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1022 bra.uni $L__BB209_24; $L__BB209_196: mov.u32 %r358, 69; add.s32 %r359, %r34, 1; { // callseq 1002, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r358; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r359; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1002 bra.uni $L__BB209_24; $L__BB209_202: add.s32 %r371, %r34, 1; { // callseq 1012, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r371; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1012 bra.uni $L__BB209_24; $L__BB209_184: add.s32 %r352, %r34, 1; { // callseq 994, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; .param .b32 param4; st.param.b32 [param4+0], %r352; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r652, [retval0+0]; } // callseq 994 bra.uni $L__BB209_24; $L__BB209_209: cvt.u32.u64 %r378, %rd2; { // callseq 1024, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r378; .param .b64 param1; st.param.b64 [param1+0], %rd120; .param .b64 param2; st.param.b64 [param2+0], %rd150; .param .b32 param3; st.param.b32 [param3+0], %r338; .param .b32 param4; st.param.b32 [param4+0], %r339; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1024 bra.uni $L__BB209_24; $L__BB209_205: add.s32 %r376, %r34, 1; { // callseq 1017, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r376; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1017 bra.uni $L__BB209_24; $L__BB209_193: { // callseq 997, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; .param .b32 param3; st.param.b32 [param3+0], %r340; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 997 bra.uni $L__BB209_24; $L__BB209_199: mov.u32 %r363, 64; add.s32 %r364, %r34, 1; { // callseq 1007, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r363; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; .param .b32 param2; st.param.b32 [param2+0], %r364; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r652, [retval0+0]; } // callseq 1007 bra.uni $L__BB209_24; $L__BB209_181: { // callseq 989, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .b32 param1; st.param.b32 [param1+0], %r338; .param .b32 param2; st.param.b32 [param2+0], %r339; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 989 bra.uni $L__BB209_24; $L__BB209_215: mov.u32 %r385, 26; { // callseq 1034, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd150; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r385; st.param.b32 [param1+4], %r338; st.param.b32 [param1+8], %r339; st.param.b32 [param1+12], %r340; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1034 bra.uni $L__BB209_24; $L__BB209_192: ld.local.f32 %f268, [%rd29]; st.local.f32 [%rd33], %f268; bra.uni $L__BB209_24; $L__BB209_190: mul.wide.u32 %rd795, %r338, 4; add.u64 %rd794, %SPL, 304; add.s64 %rd793, %rd794, %rd795; add.s64 %rd792, %rd793, 8; ld.local.f32 %f266, [%rd792]; st.local.f32 [%rd33], %f266; bra.uni $L__BB209_24; $L__BB209_229: mov.f32 %f901, 0f00000000; add.s64 %rd270, %rd120, 52; cvta.to.local.u64 %rd271, %rd270; ld.local.u8 %rs15, [%rd271]; and.b16 %rs16, %rs15, 2; setp.eq.s16 %p167, %rs16, 0; mov.f32 %f900, %f901; mov.f32 %f899, %f901; @%p167 bra $L__BB209_231; add.s64 %rd776, %rd120, 32; cvta.to.local.u64 %rd775, %rd776; add.s64 %rd273, %rd120, 24; cvta.to.local.u64 %rd274, %rd273; ld.local.v2.f32 {%f284, %f285}, [%rd274]; ld.local.v4.f32 {%f288, %f289, %f290, %f291}, [%rd775]; mul.ftz.f32 %f296, %f285, %f290; fma.rn.ftz.f32 %f297, %f284, %f289, %f296; fma.rn.ftz.f32 %f298, %f288, %f291, %f297; abs.ftz.f32 %f299, %f298; setp.gt.ftz.f32 %p168, %f299, 0f00000000; selp.f32 %f300, 0f3F800000, 0f00000000, %p168; add.s64 %rd275, %rd120, 280; cvta.to.local.u64 %rd276, %rd275; ld.local.v2.f32 {%f301, %f302}, [%rd276]; add.s64 %rd277, %rd120, 288; cvta.to.local.u64 %rd278, %rd277; ld.local.f32 %f305, [%rd278]; mul.ftz.f32 %f899, %f300, %f305; mul.ftz.f32 %f900, %f300, %f302; mul.ftz.f32 %f901, %f301, %f300; $L__BB209_231: mul.ftz.f32 %f907, %f882, %f901; mul.ftz.f32 %f906, %f882, %f900; mul.ftz.f32 %f905, %f882, %f899; setp.eq.s32 %p169, %r650, -1; @%p169 bra $L__BB209_233; mul.wide.s32 %rd279, %r650, 192; add.s64 %rd280, %rd5, %rd279; ld.global.f32 %f306, [%rd280+28]; mul.ftz.f32 %f907, %f907, %f306; ld.global.f32 %f307, [%rd280+32]; mul.ftz.f32 %f906, %f906, %f307; ld.global.f32 %f308, [%rd280+36]; mul.ftz.f32 %f905, %f905, %f308; $L__BB209_233: setp.eq.ftz.f32 %p170, %f906, 0f00000000; setp.eq.ftz.f32 %p171, %f907, 0f00000000; and.pred %p172, %p170, %p171; setp.eq.ftz.f32 %p173, %f905, 0f00000000; and.pred %p174, %p173, %p172; @%p174 bra $L__BB209_277; and.b32 %r412, %r31, 2048; setp.ne.s32 %p175, %r412, 0; mov.f32 %f908, 0f3F800000; @%p175 bra $L__BB209_236; ld.global.f32 %f310, [%rd10]; mul.ftz.f32 %f311, %f310, %f310; fma.rn.ftz.f32 %f312, %f883, %f883, %f311; div.approx.ftz.f32 %f908, %f311, %f312; $L__BB209_236: ld.global.f32 %f314, [%rd11]; mul.ftz.f32 %f315, %f907, %f314; ld.global.f32 %f316, [%rd11+4]; mul.ftz.f32 %f317, %f906, %f316; ld.global.f32 %f318, [%rd11+8]; mul.ftz.f32 %f319, %f905, %f318; mul.ftz.f32 %f78, %f908, %f315; mul.ftz.f32 %f79, %f908, %f317; mul.ftz.f32 %f80, %f908, %f319; ld.global.u16 %rs1, [%rd12]; mov.b32 %r124, %f78; setp.nan.ftz.f32 %p176, %f78, %f78; mov.f32 %f915, 0f00000000; @%p176 bra $L__BB209_239; mov.f32 %f915, 0f00000000; or.b32 %r413, %r124, -2147483648; setp.ne.s32 %p177, %r413, -2147483648; add.ftz.f32 %f321, %f78, %f78; setp.eq.ftz.f32 %p178, %f321, %f78; and.pred %p179, %p178, %p177; @%p179 bra $L__BB209_239; shl.b32 %r414, %r124, 1; setp.lt.u32 %p180, %r414, -16777215; selp.f32 %f915, %f78, 0f00000000, %p180; $L__BB209_239: mov.f32 %f916, 0f00000000; mov.b32 %r125, %f79; setp.nan.ftz.f32 %p181, %f79, %f79; @%p181 bra $L__BB209_242; mov.f32 %f916, 0f00000000; or.b32 %r415, %r125, -2147483648; setp.ne.s32 %p182, %r415, -2147483648; add.ftz.f32 %f324, %f79, %f79; setp.eq.ftz.f32 %p183, %f324, %f79; and.pred %p184, %p183, %p182; @%p184 bra $L__BB209_242; shl.b32 %r416, %r125, 1; setp.lt.u32 %p185, %r416, -16777215; selp.f32 %f916, %f79, 0f00000000, %p185; $L__BB209_242: mov.f32 %f917, 0f00000000; mov.b32 %r126, %f80; setp.nan.ftz.f32 %p186, %f80, %f80; @%p186 bra $L__BB209_245; mov.f32 %f917, 0f00000000; or.b32 %r417, %r126, -2147483648; setp.ne.s32 %p187, %r417, -2147483648; add.ftz.f32 %f327, %f80, %f80; setp.eq.ftz.f32 %p188, %f327, %f80; and.pred %p189, %p188, %p187; @%p189 bra $L__BB209_245; shl.b32 %r418, %r126, 1; setp.lt.u32 %p190, %r418, -16777215; selp.f32 %f917, %f80, 0f00000000, %p190; $L__BB209_245: setp.gt.u16 %p191, %rs1, 1; selp.f32 %f87, %f8, %f9, %p191; abs.ftz.f32 %f328, %f916; abs.ftz.f32 %f329, %f915; add.ftz.f32 %f330, %f329, %f328; abs.ftz.f32 %f331, %f917; add.ftz.f32 %f88, %f331, %f330; setp.leu.ftz.f32 %p192, %f88, %f87; @%p192 bra $L__BB209_247; div.approx.ftz.f32 %f332, %f87, %f88; mul.ftz.f32 %f915, %f915, %f332; mul.ftz.f32 %f916, %f916, %f332; mul.ftz.f32 %f917, %f917, %f332; $L__BB209_247: ld.global.u32 %rd281, [%rd13]; cvt.s64.s32 %rd282, %r285; mul.lo.s64 %rd34, %rd282, %rd281; ld.global.u32 %r127, [%rd6]; ld.global.u32 %r128, [%rd14]; @%p193 bra $L__BB209_252; and.b32 %r419, %r127, 536870912; setp.ne.s32 %p194, %r419, 0; @%p194 bra $L__BB209_250; cvt.s64.s32 %rd283, %r293; add.s64 %rd284, %rd34, %rd283; shl.b64 %rd285, %rd284, 2; add.s64 %rd286, %rd1, %rd285; add.s64 %rd287, %rd286, 4; add.s64 %rd288, %rd286, 8; atom.global.add.f32 %f333, [%rd286], %f915; atom.global.add.f32 %f334, [%rd287], %f916; atom.global.add.f32 %f335, [%rd288], %f917; $L__BB209_250: and.b32 %r420, %r127, 1073741824; setp.eq.s32 %p195, %r420, 0; @%p195 bra $L__BB209_252; bra.uni $L__BB209_251; $L__BB209_252: and.b32 %r421, %r9, 2; setp.eq.s32 %p196, %r421, 0; @%p196 bra $L__BB209_254; cvt.s64.s32 %rd295, %r286; add.s64 %rd296, %rd34, %rd295; shl.b64 %rd297, %rd296, 2; add.s64 %rd298, %rd1, %rd297; add.s64 %rd299, %rd298, 4; add.s64 %rd300, %rd298, 8; atom.global.add.f32 %f339, [%rd298], %f915; atom.global.add.f32 %f340, [%rd299], %f916; atom.global.add.f32 %f341, [%rd300], %f917; $L__BB209_254: cvt.u32.u64 %r422, %rd17; setp.eq.s32 %p197, %r422, -1; @%p197 bra $L__BB209_257; and.b32 %r423, %r128, -1431655766; popc.b32 %r424, %r423; and.b32 %r425, %r424, 1; setp.eq.b32 %p198, %r425, 1; mov.pred %p199, 0; xor.pred %p200, %p198, %p199; not.pred %p201, %p200; @%p201 bra $L__BB209_257; add.s64 %rd301, %rd34, %rd17; shl.b64 %rd302, %rd301, 2; add.s64 %rd303, %rd1, %rd302; add.s64 %rd304, %rd303, 4; add.s64 %rd305, %rd303, 8; add.s64 %rd306, %rd303, 12; add.ftz.f32 %f342, %f915, %f915; atom.global.add.f32 %f343, [%rd303], %f342; add.ftz.f32 %f344, %f916, %f916; atom.global.add.f32 %f345, [%rd304], %f344; add.ftz.f32 %f346, %f917, %f917; atom.global.add.f32 %f347, [%rd305], %f346; atom.global.add.f32 %f348, [%rd306], 0f00000000; bra.uni $L__BB209_257; $L__BB209_251: cvt.s64.s32 %rd289, %r291; add.s64 %rd290, %rd34, %rd289; shl.b64 %rd291, %rd290, 2; add.s64 %rd292, %rd1, %rd291; add.s64 %rd293, %rd292, 4; add.s64 %rd294, %rd292, 8; atom.global.add.f32 %f336, [%rd292], %f915; atom.global.add.f32 %f337, [%rd293], %f916; atom.global.add.f32 %f338, [%rd294], %f917; $L__BB209_257: setp.eq.s32 %p202, %r9, 0; @%p202 bra $L__BB209_276; cvt.u32.u64 %r426, %rd16; setp.eq.s32 %p203, %r426, -1; and.b32 %r427, %r127, 16777216; setp.eq.s32 %p204, %r427, 0; or.pred %p205, %p204, %p203; @%p205 bra $L__BB209_260; ld.global.f32 %f349, [%rd15]; mul.ftz.f32 %f350, %f915, %f349; ld.global.f32 %f351, [%rd15+4]; mul.ftz.f32 %f352, %f916, %f351; ld.global.f32 %f353, [%rd15+8]; mul.ftz.f32 %f354, %f917, %f353; add.s64 %rd307, %rd34, %rd16; shl.b64 %rd308, %rd307, 2; add.s64 %rd309, %rd1, %rd308; add.s64 %rd310, %rd309, 4; add.s64 %rd311, %rd309, 8; atom.global.add.f32 %f355, [%rd309], %f350; atom.global.add.f32 %f356, [%rd310], %f352; atom.global.add.f32 %f357, [%rd311], %f354; $L__BB209_260: @%p1 bra $L__BB209_262; add.s64 %rd312, %rd34, %rd19; add.s64 %rd313, %rd312, %rd18; shl.b64 %rd314, %rd313, 2; add.s64 %rd315, %rd1, %rd314; add.s64 %rd316, %rd315, 4; add.s64 %rd317, %rd315, 8; atom.global.add.f32 %f358, [%rd315], %f915; atom.global.add.f32 %f359, [%rd316], %f916; atom.global.add.f32 %f360, [%rd317], %f917; $L__BB209_262: and.b32 %r428, %r127, 100663296; setp.eq.s32 %p206, %r428, 0; mov.u32 %r653, %r289; @%p206 bra $L__BB209_274; setp.eq.s32 %p207, %r18, 0; mov.u32 %r653, -1; @%p207 bra $L__BB209_274; bra.uni $L__BB209_264; $L__BB209_274: setp.eq.s32 %p218, %r653, -1; @%p218 bra $L__BB209_276; cvt.s64.s32 %rd330, %r653; add.s64 %rd331, %rd34, %rd330; shl.b64 %rd332, %rd331, 2; add.s64 %rd333, %rd1, %rd332; add.s64 %rd334, %rd333, 4; add.s64 %rd335, %rd333, 8; atom.global.add.f32 %f380, [%rd333], %f915; atom.global.add.f32 %f381, [%rd334], %f916; atom.global.add.f32 %f382, [%rd335], %f917; bra.uni $L__BB209_276; $L__BB209_264: and.b32 %r430, %r127, 536870912; setp.ne.s32 %p208, %r430, 0; @%p208 bra $L__BB209_276; and.b32 %r431, %r127, 33554432; setp.eq.s32 %p209, %r431, 0; @%p209 bra $L__BB209_272; bra.uni $L__BB209_266; $L__BB209_272: mov.u32 %r653, -1; and.b32 %r434, %r127, 67108864; setp.eq.s32 %p216, %r434, 0; @%p216 bra $L__BB209_274; setp.eq.s16 %p217, %rs1, 1; selp.b32 %r653, %r288, %r299, %p217; bra.uni $L__BB209_274; $L__BB209_276: add.s32 %r647, %r647, 1; setp.lt.s32 %p219, %r647, %r4; @%p219 bra $L__BB209_5; $L__BB209_277: ld.const.u32 %r435, [kernel_params+1036]; setp.eq.s32 %p220, %r435, 0; mov.u16 %rs57, 0; @%p220 bra $L__BB209_279; ld.const.u64 %rd336, [kernel_params+2544]; cvta.to.global.u64 %rd337, %rd336; shl.b64 %rd338, %rd2, 2; add.s64 %rd339, %rd337, %rd338; ld.global.u8 %rs18, [%rd339+1]; and.b16 %rs19, %rs18, 64; shr.u16 %rs57, %rs19, 6; $L__BB209_279: setp.eq.s16 %p221, %rs57, 0; mov.f32 %f111, 0f00000000; mov.u16 %rs58, 1; @%p221 bra $L__BB209_281; ld.const.u64 %rd340, [kernel_params+2584]; cvta.to.global.u64 %rd341, %rd340; mul.lo.s64 %rd342, %rd2, 12; add.s64 %rd343, %rd341, %rd342; ld.global.f32 %f384, [%rd343+4]; ld.global.f32 %f385, [%rd343]; add.ftz.f32 %f386, %f385, %f384; ld.global.f32 %f387, [%rd343+8]; add.ftz.f32 %f388, %f387, %f386; mul.ftz.f32 %f111, %f388, 0f3EAAAAAB; ld.const.u8 %rs21, [kernel_params+1260]; and.b16 %rs22, %rs21, 8; shr.u16 %rs58, %rs22, 3; $L__BB209_281: setp.lt.s32 %p524, %r4, 1; ld.const.u64 %rd344, [kernel_params+2552]; cvta.to.global.u64 %rd345, %rd344; add.s64 %rd346, %rd345, %rd2; ld.global.u8 %rs23, [%rd346]; and.b16 %rs24, %rs23, 4; setp.eq.s16 %p222, %rs24, 0; ld.const.u32 %r136, [kernel_params+1072]; setp.eq.s32 %p223, %r136, 0; or.pred %p224, %p223, %p222; or.pred %p226, %p224, %p524; mov.u16 %rs59, %rs58; @%p226 bra $L__BB209_286; ld.const.u64 %rd347, [kernel_params+2368]; cvta.to.global.u64 %rd35, %rd347; mov.u32 %r654, 0; mov.u16 %rs25, 0; $L__BB209_283: cvt.s64.s32 %rd36, %r654; mul.wide.s32 %rd348, %r654, 192; add.s64 %rd37, %rd35, %rd348; ld.global.u32 %r437, [%rd37]; setp.ne.s32 %p227, %r437, 2; @%p227 bra $L__BB209_285; ld.global.u32 %r438, [%rd37+40]; setp.ne.s32 %p228, %r438, 0; mov.u16 %rs59, %rs25; @%p228 bra $L__BB209_286; $L__BB209_285: cvt.u32.u64 %r439, %rd36; add.s32 %r654, %r439, 1; setp.lt.s32 %p229, %r654, %r4; mov.u16 %rs59, %rs58; @%p229 bra $L__BB209_283; $L__BB209_286: setp.eq.s16 %p230, %rs59, 0; mov.f32 %f956, 0f00000000; mov.f32 %f945, %f956; mov.f32 %f946, %f956; mov.f32 %f947, %f956; @%p230 bra $L__BB209_552; ld.const.u64 %rd349, [kernel_params+2544]; cvta.to.global.u64 %rd350, %rd349; shl.b64 %rd351, %rd2, 2; add.s64 %rd38, %rd350, %rd351; ld.global.u32 %r139, [%rd38]; ld.const.u32 %r140, [kernel_params+1024]; and.b32 %r440, %r140, 264241152; setp.eq.s32 %p231, %r440, 0; @%p231 bra $L__BB209_289; shr.u32 %r441, %r140, 27; and.b32 %r442, %r441, 1; setp.eq.b32 %p232, %r442, 1; shr.u32 %r443, %r139, 3; and.b32 %r444, %r443, 1; setp.eq.b32 %p233, %r444, 1; and.pred %p234, %p232, %p233; shr.u32 %r445, %r140, 26; and.b32 %r446, %r445, 1; setp.eq.b32 %p235, %r446, 1; and.b32 %r447, %r139, 18; setp.eq.s32 %p236, %r447, 18; and.pred %p237, %p235, %p236; or.pred %p238, %p234, %p237; shr.u32 %r448, %r140, 25; and.b32 %r449, %r448, 1; setp.eq.b32 %p239, %r449, 1; shr.u32 %r450, %r139, 2; and.b32 %r451, %r450, 1; setp.eq.b32 %p240, %r451, 1; and.pred %p241, %p239, %p240; or.pred %p242, %p241, %p238; shr.u32 %r452, %r140, 24; and.b32 %r453, %r452, 1; setp.eq.b32 %p243, %r453, 1; and.b32 %r454, %r139, 1; setp.eq.b32 %p244, %r454, 1; and.pred %p245, %p243, %p244; or.pred %p246, %p245, %p242; shr.u32 %r455, %r140, 23; and.b32 %r456, %r455, 1; setp.eq.b32 %p247, %r456, 1; shr.u32 %r457, %r139, 7; and.b32 %r458, %r457, 1; setp.eq.b32 %p248, %r458, 1; and.pred %p249, %p247, %p248; or.pred %p250, %p249, %p246; mov.f32 %f925, 0f00000000; mov.f32 %f926, %f925; mov.f32 %f927, %f925; @%p250 bra $L__BB209_509; $L__BB209_289: and.b32 %r655, %r140, 4194303; ld.const.u64 %rd352, [kernel_params+2408]; cvta.to.global.u64 %rd353, %rd352; shl.b32 %r459, %r140, 5; cvt.u64.u32 %rd354, %r459; and.b64 %rd355, %rd354, 134217696; add.s64 %rd356, %rd353, %rd355; add.s64 %rd39, %rd356, 16; ld.global.u32 %r142, [%rd356+16]; and.b32 %r460, %r142, 134217728; setp.eq.s32 %p252, %r460, 0; mov.pred %p529, -1; mov.f32 %f925, 0f00000000; mov.f32 %f926, %f925; mov.f32 %f927, %f925; @%p252 bra $L__BB209_291; ld.global.f32 %f925, [%rd39+-16]; ld.global.f32 %f926, [%rd39+-12]; ld.global.f32 %f927, [%rd39+-8]; mov.pred %p529, 0; $L__BB209_291: @%p529 bra $L__BB209_292; bra.uni $L__BB209_509; $L__BB209_292: ld.const.u64 %rd791, [kernel_params+2664]; mul.wide.s32 %rd790, %r646, 4; cvta.to.global.u64 %rd789, %rd791; add.s64 %rd788, %rd789, %rd790; ld.param.u64 %rd779, [kernel_gpu_integrator_shade_background_param_1]; ld.const.u64 %rd357, [kernel_params+2632]; cvta.to.global.u64 %rd358, %rd357; mul.lo.s64 %rd359, %rd2, 12; add.s64 %rd360, %rd358, %rd359; ld.global.f32 %f398, [%rd360+8]; add.u64 %rd361, %SP, 0; add.s64 %rd362, %rd361, 248; cvta.to.local.u64 %rd363, %rd362; ld.global.f32 %f399, [%rd360+4]; ld.global.f32 %f400, [%rd360]; add.s64 %rd364, %rd361, 240; cvta.to.local.u64 %rd365, %rd364; ld.global.f32 %f401, [%rd3+8]; ld.global.f32 %f402, [%rd3+4]; ld.global.f32 %f403, [%rd3]; ld.global.u32 %r461, [%rd788]; st.local.v2.f32 [%rd365], {%f400, %f399}; st.local.f32 [%rd363], %f398; add.u64 %rd40, %SPL, 0; neg.ftz.f32 %f404, %f403; st.local.v4.f32 [%rd40], {%f403, %f402, %f401, %f404}; add.s64 %rd366, %rd361, 16; cvta.to.local.u64 %rd367, %rd366; neg.ftz.f32 %f405, %f401; neg.ftz.f32 %f406, %f402; st.local.v4.f32 [%rd367], {%f406, %f405, %f404, %f406}; add.s64 %rd368, %rd361, 32; cvta.to.local.u64 %rd369, %rd368; st.local.v4.f32 [%rd369], {%f405, %f404, %f406, %f405}; add.s64 %rd41, %rd361, 52; add.s64 %rd370, %rd361, 76; cvta.to.local.u64 %rd371, %rd370; mov.u32 %r462, -1; st.local.u32 [%rd371], %r462; add.s64 %rd372, %rd361, 48; cvta.to.local.u64 %rd373, %rd372; mov.u32 %r463, 0; st.local.v4.u32 [%rd373], {%r140, %r142, %r463, %r462}; add.s64 %rd374, %rd361, 68; cvta.to.local.u64 %rd375, %rd374; st.local.u32 [%rd375], %r463; add.s64 %rd376, %rd361, 72; cvta.to.local.u64 %rd377, %rd376; st.local.u32 [%rd377], %r463; add.s64 %rd378, %rd361, 136; cvta.to.local.u64 %rd379, %rd378; add.s64 %rd380, %rd361, 128; cvta.to.local.u64 %rd381, %rd380; mov.f32 %f925, 0f00000000; st.local.v2.f32 [%rd381], {%f925, %f925}; st.local.u32 [%rd379], %r463; add.s64 %rd382, %rd361, 80; cvta.to.local.u64 %rd383, %rd382; st.local.v4.u32 [%rd383], {%r462, %r461, %r463, %r463}; add.s64 %rd384, %rd361, 96; cvta.to.local.u64 %rd385, %rd384; st.local.v4.f32 [%rd385], {%f925, %f925, %f925, %f925}; add.s64 %rd386, %rd361, 112; cvta.to.local.u64 %rd387, %rd386; st.local.v4.f32 [%rd387], {%f925, %f925, %f925, %f925}; or.b32 %r143, %r139, 524288; add.s64 %rd388, %rd361, 260; cvta.to.local.u64 %rd389, %rd388; st.local.u32 [%rd389], %r463; add.s64 %rd390, %rd361, 264; cvta.to.local.u64 %rd391, %rd390; st.local.u32 [%rd391], %r463; ld.const.u64 %rd392, [kernel_params+2400]; cvta.to.global.u64 %rd42, %rd392; add.s64 %rd393, %rd361, 268; cvta.to.local.u64 %rd46, %rd393; add.s64 %rd394, %rd361, 272; cvta.to.local.u64 %rd47, %rd394; add.s64 %rd395, %rd361, 276; cvta.to.local.u64 %rd48, %rd395; and.b32 %r464, %r139, 24576; setp.ne.s32 %p254, %r464, 16384; setp.eq.s64 %p255, %rd779, 0; or.pred %p5, %p255, %p254; ld.const.u64 %rd396, [kernel_params+2448]; cvta.to.global.u64 %rd397, %rd396; add.s64 %rd49, %rd397, %rd351; ld.const.s32 %rd50, [kernel_params+1264]; ld.const.s32 %rd51, [kernel_params+1436]; ld.const.s32 %rd52, [kernel_params+1440]; add.u64 %rd399, %SP, 304; add.u64 %rd53, %SPL, 304; bra.uni $L__BB209_293; $L__BB209_436: add.s32 %r475, %r144, 1; { // callseq 1057, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r475; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1057 $L__BB209_293: mov.u32 %r144, %r655; mul.wide.s32 %rd400, %r144, 16; add.s64 %rd54, %rd42, %rd400; ld.global.v4.u32 {%r466, %r467, %r468, %r469}, [%rd54]; mul.wide.u32 %rd401, %r467, 4; add.s64 %rd55, %rd53, %rd401; mul.wide.u32 %rd402, %r468, 4; add.s64 %rd56, %rd53, %rd402; add.s32 %r655, %r144, 1; setp.gt.s32 %p256, %r466, 35; @%p256 bra $L__BB209_341; setp.gt.s32 %p345, %r466, 24; @%p345 bra $L__BB209_337; bra.uni $L__BB209_295; $L__BB209_337: add.s32 %r470, %r466, -27; setp.lt.u32 %p346, %r470, 9; @%p346 bra $L__BB209_293; setp.eq.s32 %p347, %r466, 25; @%p347 bra $L__BB209_493; setp.eq.s32 %p348, %r466, 26; @%p348 bra $L__BB209_340; bra.uni $L__BB209_507; $L__BB209_340: mov.u32 %r513, 26; { // callseq 1105, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r513; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z17svm_node_set_bumpILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1105 bra.uni $L__BB209_293; $L__BB209_341: setp.gt.s32 %p257, %r466, 64; @%p257 bra $L__BB209_393; bra.uni $L__BB209_342; $L__BB209_393: setp.gt.s32 %p258, %r466, 83; @%p258 bra $L__BB209_428; bra.uni $L__BB209_394; $L__BB209_428: setp.gt.s32 %p259, %r466, 90; @%p259 bra $L__BB209_441; setp.gt.s32 %p271, %r466, 87; @%p271 bra $L__BB209_437; setp.gt.s32 %p275, %r466, 85; @%p275 bra $L__BB209_434; setp.eq.s32 %p278, %r466, 84; @%p278 bra $L__BB209_460; setp.eq.s32 %p279, %r466, 85; @%p279 bra $L__BB209_433; bra.uni $L__BB209_507; $L__BB209_433: add.s32 %r477, %r144, 1; { // callseq 1059, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r477; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1059 bra.uni $L__BB209_293; $L__BB209_295: setp.gt.s32 %p349, %r466, 17; @%p349 bra $L__BB209_329; bra.uni $L__BB209_296; $L__BB209_329: setp.lt.s32 %p350, %r466, 21; @%p350 bra $L__BB209_293; setp.gt.s32 %p351, %r466, 22; @%p351 bra $L__BB209_334; setp.eq.s32 %p354, %r466, 21; @%p354 bra $L__BB209_495; setp.eq.s32 %p355, %r466, 22; @%p355 bra $L__BB209_333; bra.uni $L__BB209_507; $L__BB209_333: mov.u32 %r518, 22; add.s32 %r519, %r144, 1; { // callseq 1109, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r518; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r519; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj418EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1109 bra.uni $L__BB209_293; $L__BB209_342: setp.gt.s32 %p306, %r466, 45; @%p306 bra $L__BB209_359; setp.gt.s32 %p333, %r466, 39; @%p333 bra $L__BB209_351; bra.uni $L__BB209_344; $L__BB209_351: setp.lt.s32 %p334, %r466, 42; @%p334 bra $L__BB209_293; setp.gt.s32 %p335, %r466, 43; @%p335 bra $L__BB209_356; setp.eq.s32 %p338, %r466, 42; @%p338 bra $L__BB209_490; setp.eq.s32 %p339, %r466, 43; @%p339 bra $L__BB209_355; bra.uni $L__BB209_507; $L__BB209_355: add.s32 %r509, %r144, 1; { // callseq 1099, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r509; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1099 bra.uni $L__BB209_293; $L__BB209_394: setp.gt.s32 %p280, %r466, 73; @%p280 bra $L__BB209_411; bra.uni $L__BB209_395; $L__BB209_411: setp.gt.s32 %p281, %r466, 78; @%p281 bra $L__BB209_420; setp.gt.s32 %p288, %r466, 75; @%p288 bra $L__BB209_416; bra.uni $L__BB209_413; $L__BB209_416: setp.eq.s32 %p289, %r466, 76; @%p289 bra $L__BB209_466; setp.eq.s32 %p290, %r466, 77; @%p290 bra $L__BB209_464; setp.eq.s32 %p291, %r466, 78; @%p291 bra $L__BB209_419; bra.uni $L__BB209_507; $L__BB209_419: add.s32 %r481, %r144, 1; { // callseq 1066, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r481; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1066 bra.uni $L__BB209_293; $L__BB209_359: setp.gt.s32 %p307, %r466, 54; @%p307 bra $L__BB209_376; bra.uni $L__BB209_360; $L__BB209_376: setp.gt.s32 %p308, %r466, 59; @%p308 bra $L__BB209_385; setp.gt.s32 %p315, %r466, 56; @%p315 bra $L__BB209_381; bra.uni $L__BB209_378; $L__BB209_381: setp.eq.s32 %p316, %r466, 57; @%p316 bra $L__BB209_482; setp.eq.s32 %p317, %r466, 58; @%p317 bra $L__BB209_481; setp.eq.s32 %p318, %r466, 59; @%p318 bra $L__BB209_384; bra.uni $L__BB209_507; $L__BB209_384: add.s32 %r499, %r144, 1; { // callseq 1083, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r499; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj418EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1083 bra.uni $L__BB209_293; $L__BB209_296: setp.gt.s32 %p356, %r466, 8; @%p356 bra $L__BB209_312; bra.uni $L__BB209_297; $L__BB209_312: setp.gt.s32 %p357, %r466, 12; @%p357 bra $L__BB209_321; bra.uni $L__BB209_313; $L__BB209_321: setp.gt.s32 %p358, %r466, 14; @%p358 bra $L__BB209_325; bra.uni $L__BB209_322; $L__BB209_325: setp.eq.s32 %p359, %r466, 15; @%p359 bra $L__BB209_497; setp.eq.s32 %p360, %r466, 16; @%p360 bra $L__BB209_496; setp.eq.s32 %p361, %r466, 17; @%p361 bra $L__BB209_328; bra.uni $L__BB209_507; $L__BB209_328: { // callseq 1111, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; .param .b32 param4; st.param.b32 [param4+0], %r469; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1111 bra.uni $L__BB209_293; $L__BB209_395: setp.gt.s32 %p294, %r466, 68; @%p294 bra $L__BB209_403; bra.uni $L__BB209_396; $L__BB209_403: setp.gt.s32 %p295, %r466, 70; @%p295 bra $L__BB209_407; bra.uni $L__BB209_404; $L__BB209_407: setp.eq.s32 %p296, %r466, 71; @%p296 bra $L__BB209_474; setp.eq.s32 %p297, %r466, 72; @%p297 bra $L__BB209_473; setp.eq.s32 %p298, %r466, 73; @%p298 bra $L__BB209_410; bra.uni $L__BB209_507; $L__BB209_410: add.s32 %r483, %r144, 1; { // callseq 1069, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r483; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1069 bra.uni $L__BB209_293; $L__BB209_360: setp.gt.s32 %p321, %r466, 49; @%p321 bra $L__BB209_368; bra.uni $L__BB209_361; $L__BB209_368: setp.gt.s32 %p322, %r466, 51; @%p322 bra $L__BB209_372; bra.uni $L__BB209_369; $L__BB209_372: setp.eq.s32 %p323, %r466, 52; @%p323 bra $L__BB209_485; setp.eq.s32 %p324, %r466, 53; @%p324 bra $L__BB209_484; setp.eq.s32 %p325, %r466, 54; @%p325 bra $L__BB209_375; bra.uni $L__BB209_507; $L__BB209_375: add.s32 %r504, %r144, 1; { // callseq 1088, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r504; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1088 bra.uni $L__BB209_293; $L__BB209_297: setp.gt.s32 %p369, %r466, 4; @%p369 bra $L__BB209_305; setp.gt.s32 %p375, %r466, 2; @%p375 bra $L__BB209_302; setp.eq.s32 %p378, %r466, 1; mov.u32 %r655, %r467; @%p378 bra $L__BB209_293; setp.eq.s32 %p379, %r466, 2; @%p379 bra $L__BB209_301; bra.uni $L__BB209_507; $L__BB209_301: mov.u32 %r537, 2; add.s32 %r538, %r144, 1; { // callseq 1120, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r537; st.param.b32 [param0+4], %r467; st.param.b32 [param0+8], %r468; st.param.b32 [param0+12], %r469; .param .b32 param1; st.param.b32 [param1+0], %r538; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj418EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1120 bra.uni $L__BB209_293; $L__BB209_441: setp.gt.s32 %p260, %r466, 94; @%p260 bra $L__BB209_449; setp.gt.s32 %p266, %r466, 92; @%p266 bra $L__BB209_446; setp.eq.s32 %p269, %r466, 91; @%p269 bra $L__BB209_506; setp.eq.s32 %p270, %r466, 92; @%p270 bra $L__BB209_445; bra.uni $L__BB209_507; $L__BB209_445: ld.local.f32 %f410, [%rd55]; ld.local.f32 %f411, [%rd55+4]; ld.local.f32 %f412, [%rd55+8]; ld.global.u32 %rd414, [%rd49]; mul.lo.s64 %rd415, %rd50, %rd414; cvt.s64.s32 %rd416, %r468; add.s64 %rd417, %rd51, %rd416; add.s64 %rd418, %rd417, %rd415; shl.b64 %rd419, %rd418, 2; add.s64 %rd420, %rd1, %rd419; add.s64 %rd421, %rd420, 4; add.s64 %rd422, %rd420, 8; add.s64 %rd423, %rd420, 12; atom.global.add.f32 %f413, [%rd420], %f410; atom.global.add.f32 %f414, [%rd421], %f411; atom.global.add.f32 %f415, [%rd422], %f412; atom.global.add.f32 %f416, [%rd423], 0f3F800000; bra.uni $L__BB209_293; $L__BB209_437: setp.eq.s32 %p272, %r466, 90; @%p272 bra $L__BB209_293; setp.eq.s32 %p273, %r466, 88; @%p273 bra $L__BB209_458; setp.eq.s32 %p274, %r466, 89; @%p274 bra $L__BB209_440; bra.uni $L__BB209_507; $L__BB209_440: mov.u32 %r473, 89; { // callseq 1055, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r473; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z11svm_node_aoILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1055 bra.uni $L__BB209_293; $L__BB209_420: setp.gt.s32 %p282, %r466, 80; @%p282 bra $L__BB209_424; bra.uni $L__BB209_421; $L__BB209_424: setp.eq.s32 %p283, %r466, 81; @%p283 bra $L__BB209_462; setp.eq.s32 %p284, %r466, 82; @%p284 bra $L__BB209_461; setp.eq.s32 %p285, %r466, 83; @%p285 bra $L__BB209_427; bra.uni $L__BB209_507; $L__BB209_427: { // callseq 1061, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1061 bra.uni $L__BB209_293; $L__BB209_385: setp.gt.s32 %p309, %r466, 61; @%p309 bra $L__BB209_389; bra.uni $L__BB209_386; $L__BB209_389: setp.eq.s32 %p310, %r466, 62; @%p310 bra $L__BB209_479; setp.eq.s32 %p311, %r466, 63; @%p311 bra $L__BB209_478; setp.eq.s32 %p312, %r466, 64; @%p312 bra $L__BB209_392; bra.uni $L__BB209_507; $L__BB209_392: mov.u32 %r491, 64; add.s32 %r492, %r144, 1; { // callseq 1078, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r491; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r492; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1078 bra.uni $L__BB209_293; $L__BB209_313: setp.gt.s32 %p364, %r466, 10; @%p364 bra $L__BB209_318; setp.eq.s32 %p367, %r466, 9; @%p367 bra $L__BB209_500; setp.eq.s32 %p368, %r466, 10; @%p368 bra $L__BB209_316; bra.uni $L__BB209_507; $L__BB209_316: ld.local.f32 %f433, [%rd56]; setp.ltu.ftz.f32 %p384, %f433, 0f3F800000; @%p384 bra $L__BB209_293; add.s32 %r531, %r467, %r144; add.s32 %r655, %r531, 1; bra.uni $L__BB209_293; $L__BB209_305: setp.gt.s32 %p370, %r466, 6; @%p370 bra $L__BB209_309; setp.eq.s32 %p373, %r466, 5; @%p373 bra $L__BB209_503; setp.eq.s32 %p374, %r466, 6; @%p374 bra $L__BB209_308; bra.uni $L__BB209_507; $L__BB209_308: ld.local.f32 %f435, [%rd55]; ld.local.f32 %f436, [%rd55+4]; ld.local.f32 %f437, [%rd55+8]; st.local.f32 [%rd46], %f435; st.local.f32 [%rd47], %f436; st.local.f32 [%rd48], %f437; bra.uni $L__BB209_293; $L__BB209_344: setp.gt.s32 %p340, %r466, 37; @%p340 bra $L__BB209_348; setp.eq.s32 %p343, %r466, 36; @%p343 bra $L__BB209_492; setp.eq.s32 %p344, %r466, 37; @%p344 bra $L__BB209_347; bra.uni $L__BB209_507; $L__BB209_347: mov.u32 %r511, 37; { // callseq 1103, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r511, %r467, %r468, %r469}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1103 bra.uni $L__BB209_293; $L__BB209_396: setp.gt.s32 %p301, %r466, 66; @%p301 bra $L__BB209_400; setp.eq.s32 %p304, %r466, 65; @%p304 bra $L__BB209_477; setp.eq.s32 %p305, %r466, 66; @%p305 bra $L__BB209_399; bra.uni $L__BB209_507; $L__BB209_399: add.s32 %r490, %r144, 1; { // callseq 1076, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r490; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1076 bra.uni $L__BB209_293; $L__BB209_361: setp.gt.s32 %p328, %r466, 47; @%p328 bra $L__BB209_365; setp.eq.s32 %p331, %r466, 46; @%p331 bra $L__BB209_488; setp.eq.s32 %p332, %r466, 47; @%p332 bra $L__BB209_364; bra.uni $L__BB209_507; $L__BB209_364: cvt.u32.u64 %r506, %rd2; { // callseq 1095, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r506; .param .b64 param1; st.param.b64 [param1+0], %rd361; .param .b64 param2; st.param.b64 [param2+0], %rd399; .param .b32 param3; st.param.b32 [param3+0], %r467; .param .b32 param4; st.param.b32 [param4+0], %r468; .param .b32 param5; st.param.b32 [param5+0], %r143; call.uni _Z19svm_node_light_pathILj418EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 1095 bra.uni $L__BB209_293; $L__BB209_449: setp.gt.s32 %p261, %r466, 96; @%p261 bra $L__BB209_453; setp.eq.s32 %p264, %r466, 95; @%p264 bra $L__BB209_456; setp.eq.s32 %p265, %r466, 96; @%p265 bra $L__BB209_452; bra.uni $L__BB209_507; $L__BB209_452: { // callseq 1052, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1052 bra.uni $L__BB209_293; $L__BB209_434: setp.eq.s32 %p276, %r466, 86; @%p276 bra $L__BB209_459; setp.eq.s32 %p277, %r466, 87; @%p277 bra $L__BB209_436; bra.uni $L__BB209_507; $L__BB209_334: setp.eq.s32 %p352, %r466, 23; @%p352 bra $L__BB209_494; setp.eq.s32 %p353, %r466, 24; @%p353 bra $L__BB209_336; bra.uni $L__BB209_507; $L__BB209_336: mov.u32 %r515, 24; { // callseq 1107, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r515; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1107 bra.uni $L__BB209_293; $L__BB209_322: setp.eq.s32 %p362, %r466, 13; @%p362 bra $L__BB209_498; setp.eq.s32 %p363, %r466, 14; @%p363 bra $L__BB209_324; bra.uni $L__BB209_507; $L__BB209_324: st.local.u32 [%rd56], %r467; bra.uni $L__BB209_293; $L__BB209_302: setp.eq.s32 %p376, %r466, 3; @%p376 bra $L__BB209_504; setp.eq.s32 %p377, %r466, 4; @%p377 bra $L__BB209_304; bra.uni $L__BB209_507; $L__BB209_304: mov.u32 %r535, 4; { // callseq 1118, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r535, %r467, %r468, %r469}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1118 bra.uni $L__BB209_293; $L__BB209_318: setp.eq.s32 %p365, %r466, 11; @%p365 bra $L__BB209_499; setp.eq.s32 %p366, %r466, 12; @%p366 bra $L__BB209_320; bra.uni $L__BB209_507; $L__BB209_320: { // callseq 1114, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1114 bra.uni $L__BB209_293; $L__BB209_309: setp.eq.s32 %p371, %r466, 7; @%p371 bra $L__BB209_502; setp.eq.s32 %p372, %r466, 8; @%p372 bra $L__BB209_311; bra.uni $L__BB209_507; $L__BB209_311: mov.u32 %r533, 8; { // callseq 1116, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r533; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1116 bra.uni $L__BB209_293; $L__BB209_348: setp.eq.s32 %p341, %r466, 38; @%p341 bra $L__BB209_491; setp.eq.s32 %p342, %r466, 39; @%p342 bra $L__BB209_350; bra.uni $L__BB209_507; $L__BB209_350: mov.u32 %r510, 39; { // callseq 1101, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r510, %r467, %r468, %r469}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1101 bra.uni $L__BB209_293; $L__BB209_356: setp.eq.s32 %p336, %r466, 44; @%p336 bra $L__BB209_489; setp.eq.s32 %p337, %r466, 45; @%p337 bra $L__BB209_358; bra.uni $L__BB209_507; $L__BB209_358: { // callseq 1097, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1097 bra.uni $L__BB209_293; $L__BB209_413: setp.eq.s32 %p292, %r466, 74; @%p292 bra $L__BB209_472; setp.eq.s32 %p293, %r466, 75; @%p293 bra $L__BB209_415; bra.uni $L__BB209_507; $L__BB209_415: { // callseq 1067, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1067 bra.uni $L__BB209_293; $L__BB209_378: setp.eq.s32 %p319, %r466, 55; @%p319 bra $L__BB209_483; setp.eq.s32 %p320, %r466, 56; @%p320 bra $L__BB209_380; bra.uni $L__BB209_507; $L__BB209_380: mov.u32 %r503, 56; { // callseq 1086, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r503; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1086 bra.uni $L__BB209_293; $L__BB209_404: setp.eq.s32 %p299, %r466, 69; @%p299 bra $L__BB209_475; setp.eq.s32 %p300, %r466, 70; @%p300 bra $L__BB209_406; bra.uni $L__BB209_507; $L__BB209_406: mov.u32 %r485, 70; { // callseq 1072, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r485; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1072 bra.uni $L__BB209_293; $L__BB209_369: setp.eq.s32 %p326, %r466, 50; @%p326 bra $L__BB209_486; setp.eq.s32 %p327, %r466, 51; @%p327 bra $L__BB209_371; bra.uni $L__BB209_507; $L__BB209_371: { // callseq 1091, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1091 bra.uni $L__BB209_293; $L__BB209_446: setp.eq.s32 %p267, %r466, 93; @%p267 bra $L__BB209_457; setp.eq.s32 %p268, %r466, 94; @%p268 bra $L__BB209_448; bra.uni $L__BB209_507; $L__BB209_448: mov.u32 %r471, 94; add.s32 %r472, %r144, 1; { // callseq 1054, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r471; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r472; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1054 bra.uni $L__BB209_293; $L__BB209_421: setp.eq.s32 %p286, %r466, 79; @%p286 bra $L__BB209_463; setp.eq.s32 %p287, %r466, 80; @%p287 bra $L__BB209_423; bra.uni $L__BB209_507; $L__BB209_423: { // callseq 1064, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1064 bra.uni $L__BB209_293; $L__BB209_386: setp.eq.s32 %p313, %r466, 60; @%p313 bra $L__BB209_480; setp.eq.s32 %p314, %r466, 61; @%p314 bra $L__BB209_388; bra.uni $L__BB209_507; $L__BB209_388: mov.u32 %r496, 61; add.s32 %r497, %r144, 1; { // callseq 1081, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r496; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r497; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1081 bra.uni $L__BB209_293; $L__BB209_400: setp.eq.s32 %p302, %r466, 67; @%p302 bra $L__BB209_476; setp.eq.s32 %p303, %r466, 68; @%p303 bra $L__BB209_402; bra.uni $L__BB209_507; $L__BB209_402: mov.u32 %r488, 68; { // callseq 1074, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r488; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1074 bra.uni $L__BB209_293; $L__BB209_365: setp.eq.s32 %p329, %r466, 48; @%p329 bra $L__BB209_487; setp.eq.s32 %p330, %r466, 49; @%p330 bra $L__BB209_367; bra.uni $L__BB209_507; $L__BB209_367: { // callseq 1093, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1093 bra.uni $L__BB209_293; $L__BB209_453: setp.eq.s32 %p262, %r466, 98; @%p262 bra $L__BB209_505; setp.ne.s32 %p263, %r466, 97; @%p263 bra $L__BB209_507; { // callseq 1051, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1051 bra.uni $L__BB209_293; $L__BB209_460: { // callseq 1060, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1060 bra.uni $L__BB209_293; $L__BB209_459: add.s32 %r476, %r144, 1; { // callseq 1058, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r476; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1058 bra.uni $L__BB209_293; $L__BB209_495: mov.u32 %r520, 21; { // callseq 1110, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r520; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z21svm_node_displacementILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1110 bra.uni $L__BB209_293; $L__BB209_497: ld.global.v4.u32 {%r522, %r523, %r524, %r525}, [%rd54+16]; add.s32 %r655, %r144, 2; st.local.u32 [%rd55], %r523; st.local.u32 [%rd55+4], %r524; st.local.u32 [%rd55+8], %r525; bra.uni $L__BB209_293; $L__BB209_496: mov.u32 %r521, 16; { // callseq 1112, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r521; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z13svm_node_attrILj418EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1112 bra.uni $L__BB209_293; $L__BB209_500: ld.local.f32 %f434, [%rd56]; setp.gtu.ftz.f32 %p385, %f434, 0f00000000; @%p385 bra $L__BB209_293; add.s32 %r532, %r467, %r144; add.s32 %r655, %r532, 1; bra.uni $L__BB209_293; $L__BB209_503: st.local.u32 [%rd46], %r467; st.local.u32 [%rd47], %r468; st.local.u32 [%rd48], %r469; bra.uni $L__BB209_293; $L__BB209_494: mov.u32 %r516, 23; add.s32 %r517, %r144, 1; { // callseq 1108, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r516; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r517; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1108 bra.uni $L__BB209_293; $L__BB209_498: mov.u32 %r529, 13; add.s32 %r530, %r144, 1; { // callseq 1113, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b32 param1; st.param.b32 [param1+0], %r143; .param .b64 param2; st.param.b64 [param2+0], %rd399; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r529; st.param.b32 [param3+4], %r467; st.param.b32 [param3+8], %r468; st.param.b32 [param3+12], %r469; .param .b32 param4; st.param.b32 [param4+0], %r530; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1113 bra.uni $L__BB209_293; $L__BB209_504: mov.u32 %r536, 3; { // callseq 1119, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r536, %r467, %r468, %r469}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1119 bra.uni $L__BB209_293; $L__BB209_499: { // callseq 1115, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1115 bra.uni $L__BB209_293; $L__BB209_502: mov.u32 %r534, 7; { // callseq 1117, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r534; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1117 bra.uni $L__BB209_293; $L__BB209_490: { // callseq 1100, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1100 bra.uni $L__BB209_293; $L__BB209_466: setp.eq.s32 %p381, %r469, 255; @%p381 bra $L__BB209_293; add.s64 %rd57, %rd55, 4; add.s64 %rd58, %rd55, 8; setp.eq.s32 %p382, %r468, 0; mul.wide.u32 %rd446, %r469, 4; add.s64 %rd59, %rd53, %rd446; @%p382 bra $L__BB209_471; setp.eq.s32 %p383, %r468, 1; @%p383 bra $L__BB209_470; bra.uni $L__BB209_469; $L__BB209_470: ld.local.f32 %f424, [%rd57]; st.local.f32 [%rd59], %f424; bra.uni $L__BB209_293; $L__BB209_464: setp.eq.s32 %p380, %r469, 255; @%p380 bra $L__BB209_293; ld.local.f32 %f422, [%rd55]; add.s32 %r482, %r469, %r468; mul.wide.u32 %rd442, %r482, 4; add.s64 %rd443, %rd53, %rd442; st.local.f32 [%rd443], %f422; bra.uni $L__BB209_293; $L__BB209_482: mov.u32 %r501, 57; add.s32 %r502, %r144, 1; { // callseq 1085, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r501; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r502; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1085 bra.uni $L__BB209_293; $L__BB209_481: mov.u32 %r500, 58; { // callseq 1084, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r500; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1084 bra.uni $L__BB209_293; $L__BB209_474: mov.u32 %r484, 71; { // callseq 1071, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r484; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1071 bra.uni $L__BB209_293; $L__BB209_473: { // callseq 1070, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1070 bra.uni $L__BB209_293; $L__BB209_485: add.s32 %r505, %r144, 1; { // callseq 1090, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r505; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1090 bra.uni $L__BB209_293; $L__BB209_484: { // callseq 1089, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1089 bra.uni $L__BB209_293; $L__BB209_506: @%p5 bra $L__BB209_507; bra.uni $L__BB209_293; $L__BB209_462: mov.u32 %r479, 81; { // callseq 1063, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r479; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1063 bra.uni $L__BB209_293; $L__BB209_461: mov.u32 %r478, 82; { // callseq 1062, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r478; st.param.b32 [param2+4], %r467; st.param.b32 [param2+8], %r468; st.param.b32 [param2+12], %r469; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1062 bra.uni $L__BB209_293; $L__BB209_479: mov.u32 %r494, 62; add.s32 %r495, %r144, 1; { // callseq 1080, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r494; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r495; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1080 bra.uni $L__BB209_293; $L__BB209_478: mov.u32 %r493, 63; { // callseq 1079, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r493; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1079 bra.uni $L__BB209_293; $L__BB209_492: mov.u32 %r512, 36; { // callseq 1104, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r512; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1104 bra.uni $L__BB209_293; $L__BB209_477: { // callseq 1077, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1077 bra.uni $L__BB209_293; $L__BB209_488: { // callseq 1096, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1096 bra.uni $L__BB209_293; $L__BB209_456: { // callseq 1053, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1053 bra.uni $L__BB209_293; $L__BB209_491: { // callseq 1102, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; .param .b32 param4; st.param.b32 [param4+0], %r469; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1102 bra.uni $L__BB209_293; $L__BB209_489: mov.u32 %r507, 44; add.s32 %r508, %r144, 1; { // callseq 1098, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r507; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r508; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1098 bra.uni $L__BB209_293; $L__BB209_472: { // callseq 1068, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1068 bra.uni $L__BB209_293; $L__BB209_483: ld.local.v4.f32 {%f426, %f427, %f428, %f429}, [%rd40]; { // callseq 1087, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f426; st.param.f32 [param0+4], %f427; st.param.f32 [param0+8], %f428; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; .param .b32 param4; st.param.b32 [param4+0], %r469; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1087 bra.uni $L__BB209_293; $L__BB209_475: mov.u32 %r486, 69; add.s32 %r487, %r144, 1; { // callseq 1073, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r486; st.param.b32 [param1+4], %r467; st.param.b32 [param1+8], %r468; st.param.b32 [param1+12], %r469; .param .b32 param2; st.param.b32 [param2+0], %r487; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1073 bra.uni $L__BB209_293; $L__BB209_486: { // callseq 1092, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1092 bra.uni $L__BB209_293; $L__BB209_457: ld.local.f32 %f408, [%rd55]; ld.global.u32 %rd407, [%rd49]; mul.lo.s64 %rd408, %rd50, %rd407; cvt.s64.s32 %rd409, %r468; add.s64 %rd410, %rd52, %rd409; add.s64 %rd411, %rd410, %rd408; shl.b64 %rd412, %rd411, 2; add.s64 %rd413, %rd1, %rd412; atom.global.add.f32 %f409, [%rd413], %f408; bra.uni $L__BB209_293; $L__BB209_463: add.s32 %r480, %r144, 1; { // callseq 1065, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r480; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1065 bra.uni $L__BB209_293; $L__BB209_480: add.s32 %r498, %r144, 1; { // callseq 1082, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r498; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1082 bra.uni $L__BB209_293; $L__BB209_476: mov.u32 %r489, 67; { // callseq 1075, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r489, %r467, %r468, %r469}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1075 bra.uni $L__BB209_293; $L__BB209_487: { // callseq 1094, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd361; .param .b64 param1; st.param.b64 [param1+0], %rd399; .param .b32 param2; st.param.b32 [param2+0], %r467; .param .b32 param3; st.param.b32 [param3+0], %r468; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1094 bra.uni $L__BB209_293; $L__BB209_505: { // callseq 1121, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1121 bra.uni $L__BB209_293; $L__BB209_493: add.s32 %r514, %r144, 1; { // callseq 1106, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd399; .param .b32 param1; st.param.b32 [param1+0], %r467; .param .b32 param2; st.param.b32 [param2+0], %r468; .param .b32 param3; st.param.b32 [param3+0], %r469; .param .b32 param4; st.param.b32 [param4+0], %r514; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r655, [retval0+0]; } // callseq 1106 bra.uni $L__BB209_293; $L__BB209_458: mov.u32 %r474, 88; ld.local.f32 %f417, [%rd40+12]; ld.local.v2.f32 {%f418, %f419}, [%rd40+16]; { // callseq 1056, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f417; .param .b32 param1; st.param.f32 [param1+0], %f418; .param .b32 param2; st.param.f32 [param2+0], %f419; .param .b64 param3; st.param.b64 [param3+0], %rd399; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r474; st.param.b32 [param4+4], %r467; st.param.b32 [param4+8], %r468; st.param.b32 [param4+12], %r469; call.uni _Z14svm_node_bevelILj418EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1056 bra.uni $L__BB209_293; $L__BB209_471: ld.local.f32 %f425, [%rd55]; st.local.f32 [%rd59], %f425; bra.uni $L__BB209_293; $L__BB209_469: ld.local.f32 %f423, [%rd58]; st.local.f32 [%rd59], %f423; bra.uni $L__BB209_293; $L__BB209_507: cvta.to.local.u64 %rd520, %rd41; ld.local.u8 %rs26, [%rd520]; and.b16 %rs27, %rs26, 2; setp.eq.s16 %p386, %rs27, 0; mov.f32 %f926, %f925; mov.f32 %f927, %f925; @%p386 bra $L__BB209_509; add.s64 %rd522, %rd361, 280; cvta.to.local.u64 %rd523, %rd522; ld.local.v2.f32 {%f925, %f926}, [%rd523]; add.s64 %rd524, %rd361, 288; cvta.to.local.u64 %rd525, %rd524; ld.local.f32 %f927, [%rd525]; $L__BB209_509: ld.const.u32 %r237, [kernel_params+1536]; setp.eq.s32 %p387, %r237, 0; mov.u16 %rs60, 0; @%p387 bra $L__BB209_511; ld.const.v2.u64 {%rd526, %rd527}, [kernel_params+2464]; cvta.to.global.u64 %rd529, %rd526; shl.b64 %rd530, %rd2, 1; add.s64 %rd531, %rd529, %rd530; ld.global.u16 %r539, [%rd531]; ld.const.u64 %rd532, [kernel_params+2496]; cvta.to.global.u64 %rd533, %rd532; add.s64 %rd534, %rd533, %rd530; ld.global.u16 %r540, [%rd534]; sub.s32 %r541, %r539, %r540; ld.const.v2.u64 {%rd535, %rd536}, [kernel_params+2480]; cvta.to.global.u64 %rd538, %rd536; add.s64 %rd539, %rd538, %rd530; ld.global.u16 %rs29, [%rd539]; setp.ne.s16 %p388, %rs29, 0; selp.b32 %r542, -1, 0, %p388; add.s32 %r543, %r541, %r542; setp.ge.s32 %p389, %r543, %r237; selp.u16 %rs60, 1, 0, %p389; $L__BB209_511: ld.const.f32 %f127, [kernel_params+1544]; ld.global.u8 %rs30, [%rd38+1]; shr.u16 %rs31, %rs30, 3; and.b16 %rs32, %rs31, 1; setp.eq.b16 %p390, %rs32, 1; or.pred %p392, %p390, %p223; mov.f32 %f944, 0f3F800000; @%p392 bra $L__BB209_551; ld.const.u64 %rd540, [kernel_params+2632]; cvta.to.global.u64 %rd541, %rd540; mul.lo.s64 %rd542, %rd2, 12; add.s64 %rd543, %rd541, %rd542; ld.global.f32 %f128, [%rd543]; ld.global.f32 %f129, [%rd543+4]; ld.global.f32 %f130, [%rd543+8]; ld.global.f32 %f131, [%rd3]; ld.global.f32 %f132, [%rd3+4]; ld.global.f32 %f133, [%rd3+8]; ld.const.u64 %rd544, [kernel_params+2560]; cvta.to.global.u64 %rd545, %rd544; add.s64 %rd547, %rd545, %rd351; ld.global.f32 %f134, [%rd547]; ld.const.v2.f32 {%f445, %f446}, [kernel_params+1056]; ld.const.f32 %f934, [kernel_params+1044]; setp.leu.ftz.f32 %p393, %f934, 0f00000000; mov.f32 %f935, 0f00000000; @%p393 bra $L__BB209_531; ld.const.u32 %r238, [kernel_params+1048]; setp.lt.s32 %p394, %r238, 1; mov.u32 %r658, 0; mov.f32 %f935, 0f00000000; mov.u16 %rs62, 0; mov.f32 %f931, %f935; @%p394 bra $L__BB209_528; ld.const.u32 %r239, [kernel_params+1052]; ld.const.u64 %rd548, [kernel_params+2368]; cvta.to.global.u64 %rd60, %rd548; mov.u16 %rs62, 0; mov.f32 %f931, 0f00000000; mov.u32 %r656, 0; mov.u32 %r658, %r656; $L__BB209_515: add.s32 %r547, %r239, %r656; mul.wide.s32 %rd549, %r547, 192; add.s64 %rd550, %rd60, %rd549; add.s64 %rd61, %rd550, 4; ld.global.v4.f32 {%f449, %f450, %f451, %f452}, [%rd550+176]; ld.global.f32 %f142, [%rd550+4]; sub.ftz.f32 %f453, %f128, %f142; ld.global.f32 %f143, [%rd550+8]; sub.ftz.f32 %f454, %f129, %f143; ld.global.f32 %f144, [%rd550+12]; sub.ftz.f32 %f455, %f130, %f144; mul.ftz.f32 %f456, %f454, %f450; fma.rn.ftz.f32 %f457, %f453, %f449, %f456; fma.rn.ftz.f32 %f458, %f455, %f451, %f457; setp.leu.ftz.f32 %p395, %f458, 0f38D1B717; @%p395 bra $L__BB209_527; add.s32 %r658, %r658, 1; ld.global.v4.f32 {%f459, %f460, %f461, %f462}, [%rd61+140]; ld.global.v4.f32 {%f463, %f464, %f465, %f466}, [%rd61+156]; mul.ftz.f32 %f467, %f129, %f450; fma.rn.ftz.f32 %f468, %f128, %f449, %f467; fma.rn.ftz.f32 %f469, %f130, %f451, %f468; mul.ftz.f32 %f470, %f143, %f450; fma.rn.ftz.f32 %f471, %f142, %f449, %f470; fma.rn.ftz.f32 %f472, %f144, %f451, %f471; sub.ftz.f32 %f473, %f469, %f472; neg.ftz.f32 %f474, %f473; mul.ftz.f32 %f475, %f132, %f450; fma.rn.ftz.f32 %f476, %f131, %f449, %f475; fma.rn.ftz.f32 %f477, %f133, %f451, %f476; div.approx.ftz.f32 %f152, %f474, %f477; setp.leu.ftz.f32 %p396, %f152, 0f38D1B717; setp.geu.ftz.f32 %p397, %f152, 0f7F7FFFFF; or.pred %p398, %p397, %p396; mov.u16 %rs62, 1; @%p398 bra $L__BB209_527; fma.rn.ftz.f32 %f478, %f131, %f152, %f128; fma.rn.ftz.f32 %f479, %f132, %f152, %f129; fma.rn.ftz.f32 %f480, %f133, %f152, %f130; sub.ftz.f32 %f153, %f478, %f142; sub.ftz.f32 %f154, %f479, %f143; sub.ftz.f32 %f155, %f480, %f144; mul.ftz.f32 %f481, %f460, %f154; fma.rn.ftz.f32 %f482, %f459, %f153, %f481; fma.rn.ftz.f32 %f483, %f461, %f155, %f482; mul.ftz.f32 %f484, %f460, %f460; fma.rn.ftz.f32 %f485, %f459, %f459, %f484; fma.rn.ftz.f32 %f156, %f461, %f461, %f485; div.approx.ftz.f32 %f157, %f483, %f156; setp.lt.ftz.f32 %p399, %f157, 0fBF000000; setp.gt.ftz.f32 %p400, %f157, 0f3F000000; or.pred %p401, %p399, %p400; @%p401 bra $L__BB209_527; mul.ftz.f32 %f486, %f464, %f154; fma.rn.ftz.f32 %f487, %f463, %f153, %f486; fma.rn.ftz.f32 %f488, %f465, %f155, %f487; mul.ftz.f32 %f489, %f464, %f464; fma.rn.ftz.f32 %f490, %f463, %f463, %f489; fma.rn.ftz.f32 %f158, %f465, %f465, %f490; div.approx.ftz.f32 %f159, %f488, %f158; setp.lt.ftz.f32 %p402, %f159, 0fBF000000; setp.gt.ftz.f32 %p403, %f159, 0f3F000000; or.pred %p404, %p402, %p403; @%p404 bra $L__BB209_527; setp.lt.ftz.f32 %p405, %f462, 0f00000000; mul.ftz.f32 %f491, %f159, %f159; fma.rn.ftz.f32 %f492, %f157, %f157, %f491; setp.gt.ftz.f32 %p406, %f492, 0f3E800000; and.pred %p407, %p405, %p406; @%p407 bra $L__BB209_527; @%p405 bra $L__BB209_524; bra.uni $L__BB209_521; $L__BB209_524: sub.ftz.f32 %f668, %f142, %f128; sub.ftz.f32 %f669, %f143, %f129; mul.ftz.f32 %f670, %f669, %f669; fma.rn.ftz.f32 %f671, %f668, %f668, %f670; sub.ftz.f32 %f672, %f144, %f130; fma.rn.ftz.f32 %f673, %f672, %f672, %f671; sqrt.approx.ftz.f32 %f164, %f673; rsqrt.approx.ftz.f32 %f674, %f673; mul.ftz.f32 %f675, %f668, %f674; mul.ftz.f32 %f676, %f669, %f674; mul.ftz.f32 %f677, %f672, %f674; mul.ftz.f32 %f678, %f449, %f675; mul.ftz.f32 %f679, %f450, %f676; neg.ftz.f32 %f680, %f679; sub.ftz.f32 %f681, %f680, %f678; mul.ftz.f32 %f682, %f451, %f677; sub.ftz.f32 %f165, %f681, %f682; setp.le.ftz.f32 %p419, %f165, 0f00000000; mov.f32 %f930, 0f00000000; @%p419 bra $L__BB209_526; mul.ftz.f32 %f683, %f164, %f164; div.approx.ftz.f32 %f930, %f683, %f165; $L__BB209_526: abs.ftz.f32 %f684, %f462; fma.rn.ftz.f32 %f931, %f684, %f930, %f931; bra.uni $L__BB209_527; $L__BB209_521: mul.ftz.f32 %f494, %f459, 0f3F000000; sub.ftz.f32 %f495, %f142, %f494; mul.ftz.f32 %f496, %f460, 0f3F000000; sub.ftz.f32 %f497, %f143, %f496; mul.ftz.f32 %f498, %f461, 0f3F000000; sub.ftz.f32 %f499, %f144, %f498; mul.ftz.f32 %f500, %f463, 0f3F000000; sub.ftz.f32 %f501, %f495, %f500; mul.ftz.f32 %f502, %f464, 0f3F000000; sub.ftz.f32 %f503, %f497, %f502; mul.ftz.f32 %f504, %f465, 0f3F000000; sub.ftz.f32 %f505, %f499, %f504; sqrt.approx.ftz.f32 %f506, %f156; rsqrt.approx.ftz.f32 %f507, %f156; mov.f32 %f508, 0f3F800000; mul.ftz.f32 %f509, %f459, %f507; mul.ftz.f32 %f510, %f460, %f507; mul.ftz.f32 %f511, %f461, %f507; sqrt.approx.ftz.f32 %f512, %f158; rsqrt.approx.ftz.f32 %f513, %f158; mul.ftz.f32 %f514, %f463, %f513; mul.ftz.f32 %f515, %f464, %f513; mul.ftz.f32 %f516, %f465, %f513; mul.ftz.f32 %f517, %f510, %f516; mul.ftz.f32 %f518, %f511, %f515; sub.ftz.f32 %f519, %f517, %f518; mul.ftz.f32 %f520, %f511, %f514; mul.ftz.f32 %f521, %f509, %f516; sub.ftz.f32 %f522, %f520, %f521; mul.ftz.f32 %f523, %f509, %f515; mul.ftz.f32 %f524, %f510, %f514; sub.ftz.f32 %f525, %f523, %f524; sub.ftz.f32 %f526, %f501, %f128; sub.ftz.f32 %f527, %f503, %f129; sub.ftz.f32 %f528, %f505, %f130; mul.ftz.f32 %f529, %f527, %f522; fma.rn.ftz.f32 %f530, %f526, %f519, %f529; fma.rn.ftz.f32 %f531, %f528, %f525, %f530; setp.gt.ftz.f32 %p409, %f531, 0f00000000; mov.f32 %f929, 0f00000000; neg.ftz.f32 %f532, %f531; selp.f32 %f533, %f532, %f531, %p409; mul.ftz.f32 %f534, %f527, %f510; fma.rn.ftz.f32 %f535, %f526, %f509, %f534; fma.rn.ftz.f32 %f536, %f528, %f511, %f535; mul.ftz.f32 %f537, %f527, %f515; fma.rn.ftz.f32 %f538, %f526, %f514, %f537; fma.rn.ftz.f32 %f539, %f528, %f516, %f538; add.ftz.f32 %f540, %f506, %f536; add.ftz.f32 %f541, %f512, %f539; sub.ftz.f32 %f542, %f536, %f540; sub.ftz.f32 %f543, %f541, %f539; sub.ftz.f32 %f544, %f540, %f536; sub.ftz.f32 %f545, %f539, %f541; mul.ftz.f32 %f546, %f539, %f542; mul.ftz.f32 %f547, %f540, %f543; mul.ftz.f32 %f548, %f541, %f544; mul.ftz.f32 %f549, %f536, %f545; mul.ftz.f32 %f550, %f533, %f533; mul.ftz.f32 %f551, %f542, %f550; mul.ftz.f32 %f552, %f543, %f550; mul.ftz.f32 %f553, %f544, %f550; mul.ftz.f32 %f554, %f545, %f550; mul.ftz.f32 %f555, %f542, %f551; mul.ftz.f32 %f556, %f543, %f552; mul.ftz.f32 %f557, %f544, %f553; mul.ftz.f32 %f558, %f545, %f554; fma.rn.ftz.f32 %f559, %f546, %f546, %f555; fma.rn.ftz.f32 %f560, %f547, %f547, %f556; fma.rn.ftz.f32 %f561, %f548, %f548, %f557; fma.rn.ftz.f32 %f562, %f549, %f549, %f558; sqrt.approx.ftz.f32 %f563, %f559; sqrt.approx.ftz.f32 %f564, %f560; sqrt.approx.ftz.f32 %f565, %f561; sqrt.approx.ftz.f32 %f566, %f562; div.approx.ftz.f32 %f567, %f546, %f563; div.approx.ftz.f32 %f568, %f547, %f564; div.approx.ftz.f32 %f569, %f548, %f565; div.approx.ftz.f32 %f570, %f549, %f566; mul.ftz.f32 %f571, %f567, %f568; neg.ftz.f32 %f572, %f571; mov.f32 %f573, 0fBF800000; max.ftz.f32 %f574, %f572, %f573; min.ftz.f32 %f575, %f574, %f508; abs.ftz.f32 %f576, %f575; sub.ftz.f32 %f577, %f508, %f576; mul.ftz.f32 %f578, %f577, 0f3F000000; sqrt.approx.ftz.f32 %f579, %f578; setp.gt.ftz.f32 %p410, %f576, 0f3F11EB85; selp.f32 %f580, %f579, %f576, %p410; mul.ftz.f32 %f581, %f580, %f580; mov.f32 %f582, 0f3C94D2E9; mov.f32 %f583, 0f3D53F941; fma.rn.ftz.f32 %f584, %f583, %f581, %f582; mov.f32 %f585, 0f3D3F841F; fma.rn.ftz.f32 %f586, %f584, %f581, %f585; mov.f32 %f587, 0f3D994929; fma.rn.ftz.f32 %f588, %f586, %f581, %f587; mov.f32 %f589, 0f3E2AAB94; fma.rn.ftz.f32 %f590, %f588, %f581, %f589; mul.ftz.f32 %f591, %f581, %f590; fma.rn.ftz.f32 %f592, %f591, %f580, %f580; add.ftz.f32 %f593, %f592, %f592; mov.f32 %f594, 0f3FC90FDB; sub.ftz.f32 %f595, %f594, %f592; selp.f32 %f596, %f593, %f595, %p410; setp.lt.ftz.f32 %p411, %f575, 0f00000000; mov.f32 %f597, 0f40490FDB; sub.ftz.f32 %f598, %f597, %f596; selp.f32 %f599, %f598, %f596, %p411; mul.ftz.f32 %f600, %f568, %f569; neg.ftz.f32 %f601, %f600; max.ftz.f32 %f602, %f601, %f573; min.ftz.f32 %f603, %f602, %f508; abs.ftz.f32 %f604, %f603; sub.ftz.f32 %f605, %f508, %f604; mul.ftz.f32 %f606, %f605, 0f3F000000; sqrt.approx.ftz.f32 %f607, %f606; setp.gt.ftz.f32 %p412, %f604, 0f3F11EB85; selp.f32 %f608, %f607, %f604, %p412; mul.ftz.f32 %f609, %f608, %f608; fma.rn.ftz.f32 %f610, %f583, %f609, %f582; fma.rn.ftz.f32 %f611, %f610, %f609, %f585; fma.rn.ftz.f32 %f612, %f611, %f609, %f587; fma.rn.ftz.f32 %f613, %f612, %f609, %f589; mul.ftz.f32 %f614, %f609, %f613; fma.rn.ftz.f32 %f615, %f614, %f608, %f608; add.ftz.f32 %f616, %f615, %f615; sub.ftz.f32 %f617, %f594, %f615; selp.f32 %f618, %f616, %f617, %p412; setp.lt.ftz.f32 %p413, %f603, 0f00000000; sub.ftz.f32 %f619, %f597, %f618; selp.f32 %f620, %f619, %f618, %p413; mul.ftz.f32 %f621, %f569, %f570; neg.ftz.f32 %f622, %f621; max.ftz.f32 %f623, %f622, %f573; min.ftz.f32 %f624, %f623, %f508; abs.ftz.f32 %f625, %f624; sub.ftz.f32 %f626, %f508, %f625; mul.ftz.f32 %f627, %f626, 0f3F000000; sqrt.approx.ftz.f32 %f628, %f627; setp.gt.ftz.f32 %p414, %f625, 0f3F11EB85; selp.f32 %f629, %f628, %f625, %p414; mul.ftz.f32 %f630, %f629, %f629; fma.rn.ftz.f32 %f631, %f583, %f630, %f582; fma.rn.ftz.f32 %f632, %f631, %f630, %f585; fma.rn.ftz.f32 %f633, %f632, %f630, %f587; fma.rn.ftz.f32 %f634, %f633, %f630, %f589; mul.ftz.f32 %f635, %f630, %f634; fma.rn.ftz.f32 %f636, %f635, %f629, %f629; add.ftz.f32 %f637, %f636, %f636; sub.ftz.f32 %f638, %f594, %f636; selp.f32 %f639, %f637, %f638, %p414; setp.lt.ftz.f32 %p415, %f624, 0f00000000; sub.ftz.f32 %f640, %f597, %f639; selp.f32 %f641, %f640, %f639, %p415; mul.ftz.f32 %f642, %f567, %f570; neg.ftz.f32 %f643, %f642; max.ftz.f32 %f644, %f643, %f573; min.ftz.f32 %f645, %f644, %f508; abs.ftz.f32 %f646, %f645; sub.ftz.f32 %f647, %f508, %f646; mul.ftz.f32 %f648, %f647, 0f3F000000; sqrt.approx.ftz.f32 %f649, %f648; setp.gt.ftz.f32 %p416, %f646, 0f3F11EB85; selp.f32 %f650, %f649, %f646, %p416; mul.ftz.f32 %f651, %f650, %f650; fma.rn.ftz.f32 %f652, %f583, %f651, %f582; fma.rn.ftz.f32 %f653, %f652, %f651, %f585; fma.rn.ftz.f32 %f654, %f653, %f651, %f587; fma.rn.ftz.f32 %f655, %f654, %f651, %f589; mul.ftz.f32 %f656, %f651, %f655; fma.rn.ftz.f32 %f657, %f656, %f650, %f650; add.ftz.f32 %f658, %f657, %f657; sub.ftz.f32 %f659, %f594, %f657; selp.f32 %f660, %f658, %f659, %p416; setp.lt.ftz.f32 %p417, %f645, 0f00000000; sub.ftz.f32 %f661, %f597, %f660; selp.f32 %f662, %f661, %f660, %p417; mov.f32 %f663, 0f40C90FDB; sub.ftz.f32 %f664, %f663, %f641; sub.ftz.f32 %f665, %f664, %f662; add.ftz.f32 %f666, %f599, %f620; sub.ftz.f32 %f160, %f666, %f665; setp.eq.ftz.f32 %p418, %f160, 0f00000000; @%p418 bra $L__BB209_523; rcp.approx.ftz.f32 %f929, %f160; $L__BB209_523: add.ftz.f32 %f931, %f931, %f929; $L__BB209_527: add.s32 %r656, %r656, 1; setp.lt.s32 %p420, %r656, %r238; @%p420 bra $L__BB209_515; $L__BB209_528: setp.lt.s32 %p421, %r658, 1; @%p421 bra $L__BB209_530; cvt.rn.f32.s32 %f686, %r658; div.approx.ftz.f32 %f935, %f931, %f686; $L__BB209_530: and.b16 %rs41, %rs62, 255; setp.eq.s16 %p422, %rs41, 0; selp.f32 %f934, 0f00000000, %f934, %p422; $L__BB209_531: add.ftz.f32 %f687, %f445, %f934; add.ftz.f32 %f176, %f446, %f687; setp.eq.ftz.f32 %p423, %f176, 0f00000000; @%p423 bra $L__BB209_549; bra.uni $L__BB209_532; $L__BB209_549: ld.const.f32 %f782, [kernel_params+1504]; mov.f32 %f783, 0f41490FDB; div.approx.ftz.f32 %f943, %f782, %f783; bra.uni $L__BB209_550; $L__BB209_532: rcp.approx.ftz.f32 %f688, %f176; mul.ftz.f32 %f689, %f934, %f688; mul.ftz.f32 %f177, %f445, %f688; mul.ftz.f32 %f178, %f446, %f688; mul.ftz.f32 %f942, %f935, %f689; setp.eq.ftz.f32 %p424, %f177, 0f00000000; @%p424 bra $L__BB209_536; ld.const.v4.f32 {%f691, %f692, %f693, %f694}, [kernel_params+1008]; ld.const.f32 %f698, [kernel_params+1020]; cos.approx.ftz.f32 %f180, %f698; mul.ftz.f32 %f699, %f132, %f692; fma.rn.ftz.f32 %f700, %f131, %f691, %f699; fma.rn.ftz.f32 %f701, %f133, %f693, %f700; setp.leu.ftz.f32 %p425, %f701, %f180; mov.f32 %f936, 0f00000000; @%p425 bra $L__BB209_535; mov.f32 %f702, 0f3F800000; sub.ftz.f32 %f703, %f702, %f180; mov.f32 %f704, 0f3E22F983; div.approx.ftz.f32 %f936, %f704, %f703; $L__BB209_535: fma.rn.ftz.f32 %f942, %f177, %f936, %f942; $L__BB209_536: setp.eq.ftz.f32 %p426, %f178, 0f00000000; @%p426 bra $L__BB209_548; setp.eq.ftz.f32 %p427, %f131, 0f00000000; setp.eq.ftz.f32 %p428, %f132, 0f00000000; and.pred %p429, %p427, %p428; setp.eq.ftz.f32 %p430, %f133, 0f00000000; and.pred %p431, %p430, %p429; mov.f32 %f941, 0f00000000; mov.f32 %f939, %f941; mov.f32 %f940, %f941; @%p431 bra $L__BB209_544; abs.ftz.f32 %f185, %f131; setp.eq.ftz.f32 %p432, %f185, 0f00000000; abs.ftz.f32 %f186, %f132; setp.eq.ftz.f32 %p433, %f186, 0f00000000; and.pred %p434, %p432, %p433; mov.b32 %r246, %f131; mov.b32 %r548, %f132; and.b32 %r247, %r548, -2147483648; @%p434 bra $L__BB209_542; bra.uni $L__BB209_539; $L__BB209_542: shr.s32 %r553, %r246, 31; and.b32 %r554, %r553, 1078530011; or.b32 %r555, %r554, %r247; mov.b32 %f938, %r555; bra.uni $L__BB209_543; $L__BB209_539: setp.eq.ftz.f32 %p435, %f185, 0f7F800000; setp.eq.ftz.f32 %p436, %f186, 0f7F800000; and.pred %p437, %p435, %p436; @%p437 bra $L__BB209_541; bra.uni $L__BB209_540; $L__BB209_541: setp.lt.s32 %p441, %r246, 0; selp.b32 %r551, 1075235812, 1061752795, %p441; or.b32 %r552, %r551, %r247; mov.b32 %f938, %r552; bra.uni $L__BB209_543; $L__BB209_540: setp.lt.s32 %p438, %r246, 0; min.ftz.f32 %f707, %f186, %f185; max.ftz.f32 %f708, %f186, %f185; div.full.ftz.f32 %f709, %f707, %f708; mul.rn.ftz.f32 %f710, %f709, %f709; mov.f32 %f711, 0fC0B59883; mov.f32 %f712, 0fBF52C7EA; fma.rn.ftz.f32 %f713, %f710, %f712, %f711; mov.f32 %f714, 0fC0D21907; fma.rn.ftz.f32 %f715, %f713, %f710, %f714; mul.ftz.f32 %f716, %f710, %f715; mul.ftz.f32 %f717, %f709, %f716; add.ftz.f32 %f718, %f710, 0f41355DC0; mov.f32 %f719, 0f41E6BD60; fma.rn.ftz.f32 %f720, %f718, %f710, %f719; mov.f32 %f721, 0f419D92C8; fma.rn.ftz.f32 %f722, %f720, %f710, %f721; rcp.approx.ftz.f32 %f723, %f722; fma.rn.ftz.f32 %f724, %f717, %f723, %f709; mov.f32 %f725, 0f3FC90FDB; sub.ftz.f32 %f726, %f725, %f724; setp.gt.ftz.f32 %p439, %f186, %f185; selp.f32 %f727, %f726, %f724, %p439; mov.f32 %f728, 0f40490FDB; sub.ftz.f32 %f729, %f728, %f727; selp.f32 %f730, %f729, %f727, %p438; mov.b32 %r549, %f730; or.b32 %r550, %r247, %r549; mov.b32 %f731, %r550; add.ftz.f32 %f732, %f185, %f186; setp.le.ftz.f32 %p440, %f732, 0f7F800000; selp.f32 %f938, %f731, %f732, %p440; $L__BB209_543: add.ftz.f32 %f733, %f938, 0fC0490FDB; mov.f32 %f734, 0fC0490FDB; mov.f32 %f735, 0f40490FDB; mov.f32 %f736, 0fC0C90FDB; mul.ftz.f32 %f737, %f132, %f132; fma.rn.ftz.f32 %f738, %f131, %f131, %f737; fma.rn.ftz.f32 %f739, %f133, %f133, %f738; sqrt.approx.ftz.f32 %f740, %f739; div.approx.ftz.f32 %f741, %f133, %f740; abs.ftz.f32 %f742, %f741; mov.f32 %f743, 0f3F800000; sub.ftz.f32 %f744, %f743, %f742; mul.ftz.f32 %f745, %f744, 0f3F000000; sqrt.approx.ftz.f32 %f746, %f745; setp.gt.ftz.f32 %p442, %f742, 0f3F11EB85; selp.f32 %f747, %f746, %f742, %p442; mul.ftz.f32 %f748, %f747, %f747; mov.f32 %f749, 0f3C94D2E9; mov.f32 %f750, 0f3D53F941; fma.rn.ftz.f32 %f751, %f750, %f748, %f749; mov.f32 %f752, 0f3D3F841F; fma.rn.ftz.f32 %f753, %f751, %f748, %f752; mov.f32 %f754, 0f3D994929; fma.rn.ftz.f32 %f755, %f753, %f748, %f754; mov.f32 %f756, 0f3E2AAB94; fma.rn.ftz.f32 %f757, %f755, %f748, %f756; mul.ftz.f32 %f758, %f748, %f757; fma.rn.ftz.f32 %f759, %f758, %f747, %f747; add.ftz.f32 %f760, %f759, %f759; mov.f32 %f761, 0f3FC90FDB; sub.ftz.f32 %f762, %f761, %f759; selp.f32 %f763, %f760, %f762, %p442; setp.lt.ftz.f32 %p443, %f741, 0f00000000; sub.ftz.f32 %f764, %f735, %f763; selp.f32 %f765, %f764, %f763, %p443; add.ftz.f32 %f766, %f765, 0fC0490FDB; div.approx.ftz.f32 %f940, %f766, %f734; div.approx.ftz.f32 %f939, %f733, %f736; $L__BB209_544: mul.ftz.f32 %f768, %f940, 0f40490FDB; sin.approx.ftz.f32 %f196, %f768; setp.eq.ftz.f32 %p444, %f196, 0f00000000; @%p444 bra $L__BB209_547; ld.const.v2.u32 {%r556, %r557}, [kernel_params+1064]; add.s32 %r560, %r556, 1; cvt.rn.f32.s32 %f770, %r556; mul.ftz.f32 %f771, %f939, %f770; cvt.rzi.ftz.s32.f32 %r561, %f771; max.s32 %r562, %r561, 0; add.s32 %r563, %r556, -1; min.s32 %r248, %r562, %r563; cvt.rn.f32.s32 %f772, %r557; mul.ftz.f32 %f773, %f940, %f772; cvt.rzi.ftz.s32.f32 %r564, %f773; max.s32 %r565, %r564, 0; add.s32 %r566, %r557, -1; min.s32 %r249, %r565, %r566; ld.const.u64 %rd551, [kernel_params+2384]; cvta.to.global.u64 %rd62, %rd551; mul.lo.s32 %r250, %r249, %r560; add.s32 %r567, %r250, %r556; mul.wide.s32 %rd552, %r567, 8; add.s64 %rd553, %rd62, %rd552; ld.const.u64 %rd554, [kernel_params+2376]; cvta.to.global.u64 %rd63, %rd554; mul.wide.s32 %rd555, %r557, 8; add.s64 %rd556, %rd63, %rd555; mul.ftz.f32 %f774, %f196, 0f419DE9E7; ld.global.f32 %f775, [%rd553]; mul.ftz.f32 %f776, %f774, %f775; ld.global.f32 %f777, [%rd556]; mul.ftz.f32 %f197, %f776, %f777; setp.eq.ftz.f32 %p445, %f197, 0f00000000; @%p445 bra $L__BB209_547; add.s32 %r568, %r250, %r248; mul.wide.s32 %rd557, %r568, 8; add.s64 %rd558, %rd62, %rd557; mul.wide.s32 %rd559, %r249, 8; add.s64 %rd560, %rd63, %rd559; ld.global.f32 %f778, [%rd560]; ld.global.f32 %f779, [%rd558]; mul.ftz.f32 %f780, %f779, %f778; div.approx.ftz.f32 %f941, %f780, %f197; $L__BB209_547: fma.rn.ftz.f32 %f942, %f178, %f941, %f942; $L__BB209_548: ld.const.f32 %f781, [kernel_params+1504]; mul.ftz.f32 %f943, %f942, %f781; $L__BB209_550: mul.ftz.f32 %f784, %f134, %f134; fma.rn.ftz.f32 %f785, %f943, %f943, %f784; div.approx.ftz.f32 %f944, %f784, %f785; $L__BB209_551: setp.eq.s16 %p446, %rs60, 0; mul.ftz.f32 %f786, %f925, %f127; selp.f32 %f787, %f925, %f786, %p446; mul.ftz.f32 %f945, %f787, %f944; mul.ftz.f32 %f788, %f926, %f127; selp.f32 %f789, %f926, %f788, %p446; mul.ftz.f32 %f946, %f789, %f944; mul.ftz.f32 %f790, %f927, %f127; selp.f32 %f791, %f927, %f790, %p446; mul.ftz.f32 %f947, %f791, %f944; $L__BB209_552: ld.const.u64 %rd561, [kernel_params+2584]; cvta.to.global.u64 %rd562, %rd561; mul.lo.s64 %rd563, %rd2, 12; add.s64 %rd564, %rd562, %rd563; ld.global.f32 %f793, [%rd564]; mul.ftz.f32 %f213, %f945, %f793; ld.global.f32 %f794, [%rd564+4]; mul.ftz.f32 %f214, %f946, %f794; ld.global.f32 %f795, [%rd564+8]; mul.ftz.f32 %f215, %f947, %f795; ld.const.u64 %rd565, [kernel_params+2464]; cvta.to.global.u64 %rd566, %rd565; shl.b64 %rd567, %rd2, 1; add.s64 %rd568, %rd566, %rd567; ld.global.u16 %rs12, [%rd568]; mov.b32 %r251, %f213; setp.nan.ftz.f32 %p447, %f213, %f213; mov.f32 %f954, %f956; @%p447 bra $L__BB209_555; or.b32 %r569, %r251, -2147483648; setp.ne.s32 %p448, %r569, -2147483648; add.ftz.f32 %f797, %f213, %f213; setp.eq.ftz.f32 %p449, %f797, %f213; and.pred %p450, %p449, %p448; mov.f32 %f954, %f956; @%p450 bra $L__BB209_555; shl.b32 %r570, %r251, 1; setp.lt.u32 %p451, %r570, -16777215; selp.f32 %f954, %f213, 0f00000000, %p451; $L__BB209_555: mov.b32 %r252, %f214; setp.nan.ftz.f32 %p452, %f214, %f214; mov.f32 %f955, %f956; @%p452 bra $L__BB209_558; or.b32 %r571, %r252, -2147483648; setp.ne.s32 %p453, %r571, -2147483648; add.ftz.f32 %f800, %f214, %f214; setp.eq.ftz.f32 %p454, %f800, %f214; and.pred %p455, %p454, %p453; mov.f32 %f955, %f956; @%p455 bra $L__BB209_558; shl.b32 %r572, %r252, 1; setp.lt.u32 %p456, %r572, -16777215; selp.f32 %f955, %f214, 0f00000000, %p456; $L__BB209_558: mov.b32 %r253, %f215; setp.nan.ftz.f32 %p457, %f215, %f215; @%p457 bra $L__BB209_561; or.b32 %r573, %r253, -2147483648; setp.ne.s32 %p458, %r573, -2147483648; add.ftz.f32 %f803, %f215, %f215; setp.eq.ftz.f32 %p459, %f803, %f215; and.pred %p460, %p459, %p458; @%p460 bra $L__BB209_561; shl.b32 %r574, %r253, 1; setp.lt.u32 %p461, %r574, -16777215; selp.f32 %f956, %f215, 0f00000000, %p461; $L__BB209_561: ld.const.f32 %f804, [kernel_params+1580]; ld.const.f32 %f805, [kernel_params+1584]; setp.gt.u16 %p462, %rs12, 1; selp.f32 %f222, %f805, %f804, %p462; abs.ftz.f32 %f806, %f955; abs.ftz.f32 %f807, %f954; add.ftz.f32 %f808, %f807, %f806; abs.ftz.f32 %f809, %f956; add.ftz.f32 %f223, %f809, %f808; setp.leu.ftz.f32 %p463, %f223, %f222; @%p463 bra $L__BB209_563; div.approx.ftz.f32 %f810, %f222, %f223; mul.ftz.f32 %f954, %f954, %f810; mul.ftz.f32 %f955, %f955, %f810; mul.ftz.f32 %f956, %f956, %f810; $L__BB209_563: ld.const.u64 %rd569, [kernel_params+2448]; cvta.to.global.u64 %rd570, %rd569; shl.b64 %rd571, %rd2, 2; add.s64 %rd64, %rd570, %rd571; ld.global.u32 %rd572, [%rd64]; ld.const.s32 %rd65, [kernel_params+1264]; mul.lo.s64 %rd66, %rd65, %rd572; ld.const.u64 %rd573, [kernel_params+2544]; cvta.to.global.u64 %rd574, %rd573; add.s64 %rd67, %rd574, %rd571; ld.global.u32 %r254, [%rd67]; @%p221 bra $L__BB209_568; ld.const.u8 %rs42, [kernel_params+1260]; and.b16 %rs43, %rs42, 2; setp.eq.s16 %p465, %rs43, 0; @%p465 bra $L__BB209_566; ld.const.s32 %rd575, [kernel_params+1268]; add.s64 %rd576, %rd575, %rd66; shl.b64 %rd577, %rd576, 2; add.s64 %rd578, %rd1, %rd577; add.s64 %rd579, %rd578, 12; atom.global.add.f32 %f811, [%rd579], %f111; $L__BB209_566: ld.const.u32 %r575, [kernel_params+1620]; setp.eq.s32 %p466, %r575, 0; shr.u32 %r576, %r254, 29; and.b32 %r577, %r576, 1; setp.eq.b32 %p467, %r577, 1; or.pred %p468, %p467, %p466; @%p468 bra $L__BB209_579; ld.const.s32 %rd580, [kernel_params+1384]; add.s64 %rd581, %rd580, %rd66; shl.b64 %rd582, %rd581, 2; add.s64 %rd583, %rd1, %rd582; add.s64 %rd584, %rd583, 12; atom.global.add.f32 %f812, [%rd584], %f111; bra.uni $L__BB209_579; $L__BB209_568: ld.const.u64 %rd585, [kernel_params+2456]; cvta.to.global.u64 %rd586, %rd585; add.s64 %rd588, %rd586, %rd571; ld.global.u32 %r255, [%rd588]; ld.const.u32 %r578, [kernel_params+1620]; setp.eq.s32 %p469, %r578, 0; @%p469 bra $L__BB209_574; setp.lt.s32 %p470, %r254, 0; @%p470 bra $L__BB209_579; and.b32 %r579, %r254, 536870912; setp.ne.s32 %p471, %r579, 0; @%p471 bra $L__BB209_572; ld.const.s32 %rd589, [kernel_params+1384]; add.s64 %rd590, %rd66, %rd589; shl.b64 %rd591, %rd590, 2; add.s64 %rd592, %rd1, %rd591; add.s64 %rd593, %rd592, 4; add.s64 %rd594, %rd592, 8; add.s64 %rd595, %rd592, 12; atom.global.add.f32 %f813, [%rd592], %f954; atom.global.add.f32 %f814, [%rd593], %f955; atom.global.add.f32 %f815, [%rd594], %f956; atom.global.add.f32 %f816, [%rd595], %f111; $L__BB209_572: and.b32 %r580, %r254, 1073741824; setp.eq.s32 %p472, %r580, 0; @%p472 bra $L__BB209_574; bra.uni $L__BB209_573; $L__BB209_574: ld.const.u8 %rs44, [kernel_params+1260]; and.b16 %rs45, %rs44, 2; setp.eq.s16 %p473, %rs45, 0; @%p473 bra $L__BB209_576; ld.const.s32 %rd602, [kernel_params+1268]; add.s64 %rd603, %rd66, %rd602; shl.b64 %rd604, %rd603, 2; add.s64 %rd605, %rd1, %rd604; add.s64 %rd606, %rd605, 4; add.s64 %rd607, %rd605, 8; add.s64 %rd608, %rd605, 12; atom.global.add.f32 %f820, [%rd605], %f954; atom.global.add.f32 %f821, [%rd606], %f955; atom.global.add.f32 %f822, [%rd607], %f956; atom.global.add.f32 %f823, [%rd608], %f111; $L__BB209_576: ld.const.u32 %r256, [kernel_params+1400]; setp.eq.s32 %p474, %r256, -1; @%p474 bra $L__BB209_579; and.b32 %r581, %r255, -1431655766; popc.b32 %r582, %r581; and.b32 %r583, %r582, 1; setp.eq.b32 %p475, %r583, 1; mov.pred %p476, 0; xor.pred %p477, %p475, %p476; not.pred %p478, %p477; @%p478 bra $L__BB209_579; cvt.s64.s32 %rd609, %r256; add.s64 %rd610, %rd66, %rd609; shl.b64 %rd611, %rd610, 2; add.s64 %rd612, %rd1, %rd611; add.s64 %rd613, %rd612, 4; add.s64 %rd614, %rd612, 8; add.s64 %rd615, %rd612, 12; add.ftz.f32 %f824, %f954, %f954; atom.global.add.f32 %f825, [%rd612], %f824; add.ftz.f32 %f826, %f955, %f955; atom.global.add.f32 %f827, [%rd613], %f826; add.ftz.f32 %f828, %f956, %f956; atom.global.add.f32 %f829, [%rd614], %f828; atom.global.add.f32 %f830, [%rd615], 0f00000000; bra.uni $L__BB209_579; $L__BB209_573: ld.const.s32 %rd596, [kernel_params+1376]; add.s64 %rd597, %rd66, %rd596; shl.b64 %rd598, %rd597, 2; add.s64 %rd599, %rd1, %rd598; add.s64 %rd600, %rd599, 4; add.s64 %rd601, %rd599, 8; atom.global.add.f32 %f817, [%rd599], %f954; atom.global.add.f32 %f818, [%rd600], %f955; atom.global.add.f32 %f819, [%rd601], %f956; $L__BB209_579: ld.const.u32 %r257, [kernel_params+1076]; ld.const.u32 %r584, [kernel_params+1260]; setp.eq.s32 %p479, %r584, 0; @%p479 bra $L__BB209_600; ld.const.u32 %r660, [kernel_params+1356]; and.b32 %r585, %r254, 16777216; setp.eq.s32 %p480, %r585, 0; ld.const.u32 %r259, [kernel_params+1428]; setp.eq.s32 %p481, %r259, -1; or.pred %p482, %p480, %p481; @%p482 bra $L__BB209_582; ld.const.u64 %rd616, [kernel_params+2616]; cvta.to.global.u64 %rd617, %rd616; add.s64 %rd619, %rd617, %rd563; ld.global.f32 %f831, [%rd619]; mul.ftz.f32 %f832, %f954, %f831; ld.global.f32 %f833, [%rd619+4]; mul.ftz.f32 %f834, %f955, %f833; ld.global.f32 %f835, [%rd619+8]; mul.ftz.f32 %f836, %f956, %f835; cvt.s64.s32 %rd620, %r259; add.s64 %rd621, %rd66, %rd620; shl.b64 %rd622, %rd621, 2; add.s64 %rd623, %rd1, %rd622; add.s64 %rd624, %rd623, 4; add.s64 %rd625, %rd623, 8; atom.global.add.f32 %f837, [%rd623], %f832; atom.global.add.f32 %f838, [%rd624], %f834; atom.global.add.f32 %f839, [%rd625], %f836; $L__BB209_582: setp.eq.s32 %p483, %r257, -1; ld.const.u32 %r260, [kernel_params+1444]; setp.eq.s32 %p484, %r260, -1; or.pred %p485, %p483, %p484; @%p485 bra $L__BB209_584; mul.lo.s32 %r586, %r257, 3; cvt.s64.s32 %rd626, %r586; add.s64 %rd627, %rd66, %rd626; cvt.s64.s32 %rd628, %r260; add.s64 %rd629, %rd627, %rd628; shl.b64 %rd630, %rd629, 2; add.s64 %rd631, %rd1, %rd630; add.s64 %rd632, %rd631, 4; add.s64 %rd633, %rd631, 8; atom.global.add.f32 %f840, [%rd631], %f954; atom.global.add.f32 %f841, [%rd632], %f955; atom.global.add.f32 %f842, [%rd633], %f956; $L__BB209_584: and.b32 %r587, %r254, 100663296; setp.eq.s32 %p486, %r587, 0; @%p486 bra $L__BB209_598; ld.const.u8 %rs46, [kernel_params+2]; and.b16 %rs47, %rs46, 32; setp.eq.s16 %p487, %rs47, 0; mov.u32 %r660, -1; @%p487 bra $L__BB209_598; bra.uni $L__BB209_586; $L__BB209_598: setp.eq.s32 %p498, %r660, -1; @%p498 bra $L__BB209_600; cvt.s64.s32 %rd653, %r660; add.s64 %rd654, %rd66, %rd653; shl.b64 %rd655, %rd654, 2; add.s64 %rd656, %rd1, %rd655; add.s64 %rd657, %rd656, 4; add.s64 %rd658, %rd656, 8; atom.global.add.f32 %f862, [%rd656], %f954; atom.global.add.f32 %f863, [%rd657], %f955; atom.global.add.f32 %f864, [%rd658], %f956; bra.uni $L__BB209_600; $L__BB209_586: and.b32 %r589, %r254, 536870912; setp.ne.s32 %p488, %r589, 0; @%p488 bra $L__BB209_600; and.b32 %r590, %r254, 33554432; setp.eq.s32 %p489, %r590, 0; @%p489 bra $L__BB209_594; bra.uni $L__BB209_588; $L__BB209_594: and.b32 %r599, %r254, 67108864; setp.eq.s32 %p496, %r599, 0; @%p496 bra $L__BB209_598; setp.eq.s16 %p497, %rs12, 1; @%p497 bra $L__BB209_597; bra.uni $L__BB209_596; $L__BB209_597: ld.const.u32 %r660, [kernel_params+1348]; bra.uni $L__BB209_598; $L__BB209_600: ld.global.u32 %r267, [%rd67]; and.b32 %r600, %r267, 1073758208; setp.ne.s32 %p499, %r600, 16384; @%p499 bra $L__BB209_611; ld.const.u32 %r268, [kernel_params+1256]; setp.eq.s32 %p500, %r268, 0; shr.u32 %r601, %r267, 13; and.b32 %r602, %r601, 1; setp.eq.b32 %p501, %r602, 1; or.pred %p502, %p501, %p500; @%p502 bra $L__BB209_611; ld.global.u32 %rd659, [%rd64]; mul.lo.s64 %rd68, %rd65, %rd659; ld.const.u64 %rd660, [kernel_params+2456]; cvta.to.global.u64 %rd661, %rd660; add.s64 %rd663, %rd661, %rd571; ld.global.u32 %r603, [%rd663]; setp.ne.s32 %p503, %r603, 0; @%p503 bra $L__BB209_611; and.b32 %r604, %r268, 1; setp.eq.b32 %p504, %r604, 1; mov.pred %p505, 0; xor.pred %p506, %p504, %p505; not.pred %p507, %p506; @%p507 bra $L__BB209_605; ld.const.s32 %rd664, [kernel_params+1272]; add.s64 %rd665, %rd68, %rd664; shl.b64 %rd666, %rd665, 2; add.s64 %rd667, %rd1, %rd666; mov.u32 %r605, 0; st.global.u32 [%rd667], %r605; $L__BB209_605: and.b32 %r606, %r268, 32; setp.eq.s32 %p508, %r606, 0; @%p508 bra $L__BB209_607; ld.const.s32 %rd668, [kernel_params+1300]; add.s64 %rd669, %rd68, %rd668; shl.b64 %rd670, %rd669, 2; add.s64 %rd671, %rd1, %rd670; mov.u32 %r607, 0; st.global.u32 [%rd671], %r607; $L__BB209_607: and.b32 %r608, %r268, 64; setp.eq.s32 %p509, %r608, 0; @%p509 bra $L__BB209_609; ld.const.s32 %rd672, [kernel_params+1304]; add.s64 %rd673, %rd68, %rd672; shl.b64 %rd674, %rd673, 2; add.s64 %rd675, %rd1, %rd674; mov.u32 %r609, 0; st.global.u32 [%rd675], %r609; $L__BB209_609: and.b32 %r610, %r268, 2; setp.eq.s32 %p510, %r610, 0; @%p510 bra $L__BB209_611; ld.const.s32 %rd676, [kernel_params+1276]; add.s64 %rd677, %rd68, %rd676; shl.b64 %rd678, %rd677, 2; add.s64 %rd679, %rd1, %rd678; mov.u32 %r611, 0; st.global.u32 [%rd679], %r611; st.global.u32 [%rd679+4], %r611; st.global.u32 [%rd679+8], %r611; $L__BB209_611: ld.global.u32 %r269, [%rd67]; setp.lt.s32 %p511, %r269, 0; @%p511 bra $L__BB209_613; ld.const.u64 %rd680, [kernel_params+4256]; cvta.to.global.u64 %rd681, %rd680; add.s64 %rd682, %rd681, 24; atom.global.add.u32 %r612, [%rd682], -1; ld.const.u64 %rd683, [kernel_params+2520]; cvta.to.global.u64 %rd684, %rd683; add.s64 %rd686, %rd684, %rd567; mov.u16 %rs48, 0; st.global.u16 [%rd686], %rs48; $L__BB209_627: ret; $L__BB209_613: and.b32 %r613, %r269, 2147483647; st.global.u32 [%rd67], %r613; ld.const.u8 %rs49, [kernel_params+2]; and.b16 %rs50, %rs49, 2; setp.eq.s16 %p512, %rs50, 0; @%p512 bra $L__BB209_616; ld.const.u64 %rd687, [kernel_params+2776]; cvta.to.global.u64 %rd688, %rd687; add.s64 %rd690, %rd688, %rd571; ld.global.u32 %r614, [%rd690]; setp.eq.s32 %p513, %r614, -1; @%p513 bra $L__BB209_616; ld.const.u64 %rd691, [kernel_params+4256]; cvta.to.global.u64 %rd692, %rd691; add.s64 %rd693, %rd692, 24; atom.global.add.u32 %r615, [%rd693], -1; add.s64 %rd694, %rd692, 20; atom.global.add.u32 %r616, [%rd694], 1; ld.const.u64 %rd695, [kernel_params+2520]; cvta.to.global.u64 %rd696, %rd695; add.s64 %rd698, %rd696, %rd567; mov.u16 %rs51, 5; st.global.u16 [%rd698], %rs51; bra.uni $L__BB209_627; $L__BB209_616: ld.const.u64 %rd699, [kernel_params+2712]; cvta.to.global.u64 %rd700, %rd699; add.s64 %rd702, %rd700, %rd571; ld.global.u32 %r270, [%rd702]; cvt.s64.s32 %rd69, %r270; ld.const.v2.u64 {%rd703, %rd704}, [kernel_params+2720]; cvta.to.global.u64 %rd707, %rd703; add.s64 %rd708, %rd707, %rd571; ld.global.u32 %r271, [%rd708]; cvta.to.global.u64 %rd709, %rd704; add.s64 %rd710, %rd709, %rd571; ld.global.u32 %r272, [%rd710]; and.b32 %r617, %r272, 1; setp.eq.b32 %p514, %r617, 1; mov.pred %p515, 0; xor.pred %p516, %p514, %p515; not.pred %p517, %p516; @%p517 bra $L__BB209_618; bra.uni $L__BB209_617; $L__BB209_618: and.b32 %r618, %r272, 8; setp.eq.s32 %p518, %r618, 0; @%p518 bra $L__BB209_620; bra.uni $L__BB209_619; $L__BB209_620: and.b32 %r620, %r272, 6; setp.eq.s32 %p519, %r620, 0; mov.u32 %r661, 0; @%p519 bra $L__BB209_622; cvt.u32.u64 %r621, %rd69; ld.const.u64 %rd719, [kernel_params+2264]; cvta.to.global.u64 %rd720, %rd719; mul.wide.s32 %rd721, %r621, 16; add.s64 %rd722, %rd720, %rd721; ld.global.u32 %r661, [%rd722]; bra.uni $L__BB209_622; $L__BB209_617: ld.const.u64 %rd711, [kernel_params+2216]; cvta.to.global.u64 %rd712, %rd711; shl.b64 %rd713, %rd69, 2; add.s64 %rd714, %rd712, %rd713; ld.global.u32 %r661, [%rd714]; bra.uni $L__BB209_622; $L__BB209_619: ld.const.u64 %rd715, [kernel_params+2304]; cvta.to.global.u64 %rd716, %rd715; mul.wide.s32 %rd717, %r270, 4; add.s64 %rd718, %rd716, %rd717; ld.global.u32 %r661, [%rd718]; $L__BB209_622: and.b32 %r277, %r661, 4194303; ld.const.u64 %rd723, [kernel_params+2184]; cvta.to.global.u64 %rd724, %rd723; mul.wide.s32 %rd725, %r271, 4; add.s64 %rd726, %rd724, %rd725; ld.const.u32 %r622, [kernel_params+1592]; setp.ne.s32 %p520, %r622, 0; ld.global.u8 %rs52, [%rd726+1]; and.b16 %rs53, %rs52, 6; setp.ne.s16 %p521, %rs53, 0; and.pred %p522, %p521, %p520; @%p522 bra $L__BB209_626; bra.uni $L__BB209_623; $L__BB209_626: cvt.u32.u64 %r634, %rd2; ld.const.u32 %r635, [kernel_params+4392]; div.u32 %r636, %r634, %r635; ld.const.u32 %r637, [kernel_params+8]; mad.lo.s32 %r638, %r636, %r637, %r277; ld.const.u64 %rd759, [kernel_params+4256]; cvta.to.global.u64 %rd760, %rd759; add.s64 %rd761, %rd760, 24; atom.global.add.u32 %r639, [%rd761], -1; add.s64 %rd762, %rd760, 40; atom.global.add.u32 %r640, [%rd762], 1; ld.const.u64 %rd763, [kernel_params+2520]; cvta.to.global.u64 %rd764, %rd763; add.s64 %rd766, %rd764, %rd567; mov.u16 %rs56, 10; st.global.u16 [%rd766], %rs56; ld.const.u64 %rd767, [kernel_params+2624]; cvta.to.global.u64 %rd768, %rd767; add.s64 %rd770, %rd768, %rd571; st.global.u32 [%rd770], %r638; ld.const.u64 %rd771, [kernel_params+4344]; cvta.to.global.u64 %rd772, %rd771; mul.wide.s32 %rd773, %r638, 4; add.s64 %rd774, %rd772, %rd773; atom.global.add.u32 %r641, [%rd774], 1; bra.uni $L__BB209_627; $L__BB209_623: ld.const.u64 %rd727, [kernel_params+2408]; cvt.u32.u64 %r623, %rd2; cvta.to.global.u64 %rd728, %rd727; mul.wide.u32 %rd729, %r277, 32; add.s64 %rd730, %rd728, %rd729; ld.global.u32 %r624, [%rd730+16]; and.b32 %r625, %r624, 1073741824; setp.eq.s32 %p523, %r625, 0; ld.const.u32 %r626, [kernel_params+4392]; div.u32 %r627, %r623, %r626; ld.const.u32 %r628, [kernel_params+8]; mad.lo.s32 %r278, %r627, %r628, %r277; ld.const.u64 %rd731, [kernel_params+4256]; cvta.to.global.u64 %rd70, %rd731; add.s64 %rd732, %rd70, 24; atom.global.add.u32 %r629, [%rd732], -1; cvt.s64.s32 %rd71, %r278; @%p523 bra $L__BB209_625; bra.uni $L__BB209_624; $L__BB209_625: add.s64 %rd746, %rd70, 32; atom.global.add.u32 %r632, [%rd746], 1; ld.const.u64 %rd747, [kernel_params+2520]; cvta.to.global.u64 %rd748, %rd747; add.s64 %rd750, %rd748, %rd567; mov.u16 %rs55, 8; st.global.u16 [%rd750], %rs55; ld.const.u64 %rd751, [kernel_params+2624]; cvta.to.global.u64 %rd752, %rd751; add.s64 %rd754, %rd752, %rd571; st.global.u32 [%rd754], %r278; ld.const.u64 %rd755, [kernel_params+4328]; cvta.to.global.u64 %rd756, %rd755; shl.b64 %rd757, %rd71, 2; add.s64 %rd758, %rd756, %rd757; atom.global.add.u32 %r633, [%rd758], 1; bra.uni $L__BB209_627; $L__BB209_624: add.s64 %rd733, %rd70, 36; atom.global.add.u32 %r630, [%rd733], 1; ld.const.u64 %rd734, [kernel_params+2520]; cvta.to.global.u64 %rd735, %rd734; add.s64 %rd737, %rd735, %rd567; mov.u16 %rs54, 9; st.global.u16 [%rd737], %rs54; ld.const.u64 %rd738, [kernel_params+2624]; cvta.to.global.u64 %rd739, %rd738; add.s64 %rd741, %rd739, %rd571; st.global.u32 [%rd741], %r278; ld.const.u64 %rd742, [kernel_params+4336]; cvta.to.global.u64 %rd743, %rd742; shl.b64 %rd744, %rd71, 2; add.s64 %rd745, %rd743, %rd744; atom.global.add.u32 %r631, [%rd745], 1; bra.uni $L__BB209_627; $L__BB209_588: ld.const.u64 %rd634, [kernel_params+2600]; cvta.to.global.u64 %rd635, %rd634; add.s64 %rd637, %rd635, %rd563; ld.global.f32 %f230, [%rd637]; ld.global.f32 %f231, [%rd637+4]; ld.global.f32 %f232, [%rd637+8]; ld.const.u64 %rd638, [kernel_params+2608]; cvta.to.global.u64 %rd639, %rd638; add.s64 %rd640, %rd639, %rd563; ld.global.f32 %f233, [%rd640]; ld.global.f32 %f234, [%rd640+4]; ld.global.f32 %f235, [%rd640+8]; ld.const.u32 %r591, [kernel_params+1324]; ld.const.u32 %r592, [kernel_params+1340]; setp.eq.s16 %p490, %rs12, 1; selp.b32 %r261, %r592, %r591, %p490; setp.eq.s32 %p491, %r261, -1; @%p491 bra $L__BB209_590; cvt.s64.s32 %rd641, %r261; add.s64 %rd642, %rd66, %rd641; shl.b64 %rd643, %rd642, 2; add.s64 %rd644, %rd1, %rd643; add.s64 %rd645, %rd644, 4; add.s64 %rd646, %rd644, 8; mul.ftz.f32 %f843, %f954, %f233; atom.global.add.f32 %f844, [%rd644], %f843; mul.ftz.f32 %f845, %f955, %f234; atom.global.add.f32 %f846, [%rd645], %f845; mul.ftz.f32 %f847, %f956, %f235; atom.global.add.f32 %f848, [%rd646], %f847; $L__BB209_590: ld.const.u32 %r593, [kernel_params+1328]; ld.const.u32 %r594, [kernel_params+1344]; selp.b32 %r262, %r594, %r593, %p490; setp.eq.s32 %p493, %r262, -1; @%p493 bra $L__BB209_592; mov.f32 %f849, 0f3F800000; sub.ftz.f32 %f850, %f849, %f230; sub.ftz.f32 %f851, %f850, %f233; sub.ftz.f32 %f852, %f849, %f231; sub.ftz.f32 %f853, %f852, %f234; sub.ftz.f32 %f854, %f849, %f232; sub.ftz.f32 %f855, %f854, %f235; cvt.s64.s32 %rd647, %r262; add.s64 %rd648, %rd66, %rd647; shl.b64 %rd649, %rd648, 2; add.s64 %rd650, %rd1, %rd649; mul.ftz.f32 %f856, %f954, %f851; mul.ftz.f32 %f857, %f955, %f853; mul.ftz.f32 %f858, %f956, %f855; add.s64 %rd651, %rd650, 4; add.s64 %rd652, %rd650, 8; atom.global.add.f32 %f859, [%rd650], %f856; atom.global.add.f32 %f860, [%rd651], %f857; atom.global.add.f32 %f861, [%rd652], %f858; $L__BB209_592: ld.const.u32 %r596, [kernel_params+1320]; ld.const.u32 %r597, [kernel_params+1336]; selp.b32 %r263, %r597, %r596, %p490; setp.eq.s32 %p495, %r263, -1; @%p495 bra $L__BB209_598; mul.ftz.f32 %f954, %f954, %f230; mul.ftz.f32 %f955, %f955, %f231; mul.ftz.f32 %f956, %f956, %f232; mov.u32 %r660, %r263; bra.uni $L__BB209_598; $L__BB209_596: ld.const.u32 %r660, [kernel_params+1332]; bra.uni $L__BB209_598; } // .globl kernel_gpu_integrator_shade_light .visible .entry kernel_gpu_integrator_shade_light( .param .u64 kernel_gpu_integrator_shade_light_param_0, .param .u64 kernel_gpu_integrator_shade_light_param_1, .param .u32 kernel_gpu_integrator_shade_light_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot210[1328]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<307>; .reg .b16 %rs<6>; .reg .f32 %f<1494>; .reg .b32 %r<328>; .reg .b64 %rd<520>; mov.u64 %SPL, __local_depot210; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd35, [kernel_gpu_integrator_shade_light_param_0]; ld.param.u64 %rd36, [kernel_gpu_integrator_shade_light_param_1]; ld.param.u32 %r133, [kernel_gpu_integrator_shade_light_param_2]; mov.u32 %r134, %ntid.x; mov.u32 %r135, %ctaid.x; mov.u32 %r136, %tid.x; mad.lo.s32 %r321, %r135, %r134, %r136; setp.ge.s32 %p4, %r321, %r133; @%p4 bra $L__BB210_356; setp.eq.s64 %p5, %rd35, 0; @%p5 bra $L__BB210_3; cvta.to.global.u64 %rd37, %rd35; mul.wide.s32 %rd38, %r321, 4; add.s64 %rd39, %rd37, %rd38; ld.global.u32 %r321, [%rd39]; $L__BB210_3: ld.const.v2.u64 {%rd40, %rd41}, [kernel_params+2704]; cvta.to.global.u64 %rd44, %rd41; cvt.s64.s32 %rd1, %r321; mul.wide.s32 %rd45, %r321, 4; add.s64 %rd46, %rd44, %rd45; ld.global.u32 %r4, [%rd46]; ld.const.u64 %rd47, [kernel_params+2720]; cvta.to.global.u64 %rd48, %rd47; add.s64 %rd49, %rd48, %rd45; ld.global.s32 %rd2, [%rd49]; ld.const.v2.u64 {%rd50, %rd51}, [kernel_params+2688]; cvta.to.global.u64 %rd54, %rd51; add.s64 %rd55, %rd54, %rd45; ld.global.f32 %f1418, [%rd55]; cvta.to.global.u64 %rd56, %rd40; add.s64 %rd57, %rd56, %rd45; ld.global.f32 %f1417, [%rd57]; cvta.to.global.u64 %rd58, %rd50; add.s64 %rd59, %rd58, %rd45; ld.const.u64 %rd60, [kernel_params+2632]; cvta.to.global.u64 %rd61, %rd60; mul.wide.s32 %rd62, %r321, 12; add.s64 %rd63, %rd61, %rd62; ld.global.f32 %f3, [%rd63]; ld.global.f32 %f4, [%rd63+4]; ld.global.f32 %f5, [%rd63+8]; ld.const.v2.u64 {%rd64, %rd65}, [kernel_params+2640]; cvta.to.global.u64 %rd68, %rd64; add.s64 %rd69, %rd68, %rd62; ld.global.f32 %f6, [%rd69]; ld.global.f32 %f7, [%rd69+4]; ld.global.f32 %f8, [%rd69+8]; ld.const.v2.u64 {%rd70, %rd71}, [kernel_params+2656]; cvta.to.global.u64 %rd73, %rd71; add.s64 %rd74, %rd73, %rd45; ld.global.f32 %f9, [%rd74]; ld.global.f32 %f10, [%rd59]; setp.eq.ftz.f32 %p6, %f10, 0f00000000; mov.b32 %r138, %f10; add.s32 %r139, %r138, 1; mov.b32 %f377, %r139; selp.f32 %f378, 0f00000001, %f377, %p6; cvta.to.global.u64 %rd75, %rd65; add.s64 %rd76, %rd75, %rd45; st.global.f32 [%rd76], %f378; ld.const.u64 %rd77, [kernel_params+2368]; cvta.to.global.u64 %rd78, %rd77; mul.wide.s32 %rd79, %r4, 192; add.s64 %rd3, %rd78, %rd79; ld.global.u32 %r5, [%rd3]; ld.global.u32 %r6, [%rd3+16]; setp.eq.s32 %p7, %r4, -1; mov.u32 %r322, -1; @%p7 bra $L__BB210_5; ld.global.u32 %r322, [%rd3+44]; $L__BB210_5: setp.eq.s32 %p8, %r5, 4; @%p8 bra $L__BB210_34; bra.uni $L__BB210_6; $L__BB210_34: ld.global.v4.f32 {%f730, %f731, %f732, %f733}, [%rd3+160]; ld.global.f32 %f737, [%rd3+4]; sub.ftz.f32 %f738, %f3, %f737; ld.global.f32 %f739, [%rd3+8]; sub.ftz.f32 %f740, %f4, %f739; ld.global.f32 %f741, [%rd3+12]; sub.ftz.f32 %f742, %f5, %f741; mul.ftz.f32 %f743, %f740, %f740; fma.rn.ftz.f32 %f744, %f738, %f738, %f743; fma.rn.ftz.f32 %f745, %f742, %f742, %f744; sqrt.approx.ftz.f32 %f746, %f745; div.approx.ftz.f32 %f124, %f738, %f746; div.approx.ftz.f32 %f125, %f740, %f746; div.approx.ftz.f32 %f126, %f742, %f746; ld.global.f32 %f1416, [%rd3+148]; ld.global.v2.f32 {%f747, %f748}, [%rd3+152]; neg.ftz.f32 %f1421, %f6; mul.ftz.f32 %f749, %f730, %f1421; neg.ftz.f32 %f1420, %f7; mul.ftz.f32 %f750, %f731, %f7; sub.ftz.f32 %f751, %f749, %f750; neg.ftz.f32 %f1419, %f8; mul.ftz.f32 %f752, %f732, %f8; sub.ftz.f32 %f133, %f751, %f752; setp.le.ftz.f32 %p54, %f133, %f747; mov.f32 %f1402, 0f00000000; @%p54 bra $L__BB210_37; sub.ftz.f32 %f134, %f133, %f747; setp.geu.ftz.f32 %p55, %f134, %f748; setp.eq.ftz.f32 %p56, %f748, 0f00000000; or.pred %p57, %p55, %p56; mov.f32 %f1402, %f133; @%p57 bra $L__BB210_37; div.approx.ftz.f32 %f753, %f134, %f748; mul.ftz.f32 %f754, %f753, %f753; mul.ftz.f32 %f755, %f754, 0fC0000000; mul.ftz.f32 %f756, %f753, %f755; fma.rn.ftz.f32 %f757, %f754, 0f40400000, %f756; mul.ftz.f32 %f1402, %f133, %f757; $L__BB210_37: mul.ftz.f32 %f759, %f1416, 0f3DA2F983; mul.ftz.f32 %f1415, %f759, %f1402; setp.eq.ftz.f32 %p59, %f1415, 0f00000000; mov.pred %p305, -1; @%p59 bra $L__BB210_52; mul.ftz.f32 %f762, %f7, %f7; fma.rn.ftz.f32 %f763, %f6, %f6, %f762; fma.rn.ftz.f32 %f138, %f8, %f8, %f763; setp.leu.ftz.f32 %p60, %f138, 0f00000000; mov.f32 %f1409, 0f00000000; mov.f32 %f1418, %f1409; mov.f32 %f1417, %f1409; @%p60 bra $L__BB210_47; setp.eq.ftz.f32 %p61, %f6, 0f80000000; mov.f32 %f1418, 0f00000000; setp.eq.ftz.f32 %p62, %f7, 0f80000000; and.pred %p63, %p61, %p62; @%p63 bra $L__BB210_46; abs.ftz.f32 %f139, %f1420; setp.eq.ftz.f32 %p64, %f139, 0f00000000; abs.ftz.f32 %f140, %f1421; setp.eq.ftz.f32 %p65, %f140, 0f00000000; and.pred %p66, %p64, %p65; mov.b32 %r11, %f1420; mov.b32 %r148, %f1421; and.b32 %r12, %r148, -2147483648; @%p66 bra $L__BB210_44; bra.uni $L__BB210_41; $L__BB210_44: shr.s32 %r153, %r11, 31; and.b32 %r154, %r153, 1078530011; or.b32 %r155, %r154, %r12; mov.b32 %f1403, %r155; bra.uni $L__BB210_45; $L__BB210_6: setp.eq.s32 %p9, %r5, 0; @%p9 bra $L__BB210_20; setp.ne.s32 %p11, %r5, 3; mov.pred %p305, -1; @%p11 bra $L__BB210_52; ld.global.v4.f32 {%f1388, %f1389, %f1390, %f383}, [%rd3+144]; abs.ftz.f32 %f18, %f383; ld.global.v4.f32 {%f1385, %f1386, %f1387, %f387}, [%rd3+160]; ld.global.v4.f32 {%f1421, %f1420, %f1419, %f391}, [%rd3+176]; ld.global.f32 %f1391, [%rd3+4]; ld.global.f32 %f1392, [%rd3+8]; ld.global.f32 %f1393, [%rd3+12]; setp.lt.ftz.f32 %p12, %f383, 0f00000000; @%p12 bra $L__BB210_15; bra.uni $L__BB210_9; $L__BB210_15: mul.ftz.f32 %f630, %f6, %f1421; mul.ftz.f32 %f631, %f7, %f1420; neg.ftz.f32 %f632, %f631; sub.ftz.f32 %f633, %f632, %f630; mul.ftz.f32 %f634, %f8, %f1419; sub.ftz.f32 %f85, %f633, %f634; setp.le.ftz.f32 %p30, %f85, 0f00000000; mov.f32 %f1394, 0f00000000; @%p30 bra $L__BB210_17; mul.ftz.f32 %f635, %f10, %f10; div.approx.ftz.f32 %f1394, %f635, %f85; $L__BB210_17: mul.ftz.f32 %f1409, %f18, %f1394; ld.global.f32 %f1395, [%rd3+172]; bra.uni $L__BB210_18; $L__BB210_20: ld.global.f32 %f653, [%rd3+4]; sub.ftz.f32 %f654, %f3, %f653; ld.global.f32 %f655, [%rd3+8]; sub.ftz.f32 %f656, %f4, %f655; ld.global.f32 %f657, [%rd3+12]; sub.ftz.f32 %f658, %f5, %f657; mul.ftz.f32 %f659, %f656, %f656; fma.rn.ftz.f32 %f660, %f654, %f654, %f659; fma.rn.ftz.f32 %f661, %f658, %f658, %f660; sqrt.approx.ftz.f32 %f662, %f661; div.approx.ftz.f32 %f100, %f654, %f662; div.approx.ftz.f32 %f101, %f656, %f662; div.approx.ftz.f32 %f102, %f658, %f662; neg.ftz.f32 %f1421, %f6; neg.ftz.f32 %f1420, %f7; neg.ftz.f32 %f1419, %f8; ld.global.f32 %f1416, [%rd3+148]; mul.ftz.f32 %f1415, %f1416, 0f3DA2F983; setp.eq.ftz.f32 %p35, %f1415, 0f00000000; mov.pred %p305, -1; @%p35 bra $L__BB210_52; mul.ftz.f32 %f665, %f7, %f7; fma.rn.ftz.f32 %f666, %f6, %f6, %f665; fma.rn.ftz.f32 %f108, %f8, %f8, %f666; setp.leu.ftz.f32 %p36, %f108, 0f00000000; mov.f32 %f1409, 0f00000000; mov.f32 %f1418, %f1409; mov.f32 %f1417, %f1409; @%p36 bra $L__BB210_30; setp.eq.ftz.f32 %p37, %f6, 0f80000000; mov.f32 %f1418, 0f00000000; setp.eq.ftz.f32 %p38, %f7, 0f80000000; and.pred %p39, %p37, %p38; @%p39 bra $L__BB210_29; abs.ftz.f32 %f109, %f1420; setp.eq.ftz.f32 %p40, %f109, 0f00000000; abs.ftz.f32 %f110, %f1421; setp.eq.ftz.f32 %p41, %f110, 0f00000000; and.pred %p42, %p40, %p41; mov.b32 %r9, %f1420; mov.b32 %r140, %f1421; and.b32 %r10, %r140, -2147483648; @%p42 bra $L__BB210_27; bra.uni $L__BB210_24; $L__BB210_27: shr.s32 %r145, %r9, 31; and.b32 %r146, %r145, 1078530011; or.b32 %r147, %r146, %r10; mov.b32 %f1397, %r147; bra.uni $L__BB210_28; $L__BB210_9: ld.global.f32 %f1395, [%rd3+172]; setp.leu.ftz.f32 %p13, %f1395, 0f00000000; @%p13 bra $L__BB210_13; sub.ftz.f32 %f392, %f3, %f1391; sub.ftz.f32 %f393, %f4, %f1392; mul.ftz.f32 %f394, %f1420, %f393; fma.rn.ftz.f32 %f395, %f1421, %f392, %f394; sub.ftz.f32 %f396, %f5, %f1393; fma.rn.ftz.f32 %f397, %f1419, %f396, %f395; mul.ftz.f32 %f398, %f1421, %f397; mul.ftz.f32 %f399, %f1420, %f397; mul.ftz.f32 %f400, %f1419, %f397; sub.ftz.f32 %f401, %f3, %f398; sub.ftz.f32 %f402, %f4, %f399; sub.ftz.f32 %f403, %f5, %f400; sub.ftz.f32 %f404, %f401, %f3; sub.ftz.f32 %f405, %f402, %f4; sub.ftz.f32 %f406, %f403, %f5; mul.ftz.f32 %f407, %f405, %f405; fma.rn.ftz.f32 %f408, %f404, %f404, %f407; fma.rn.ftz.f32 %f409, %f406, %f406, %f408; sqrt.approx.ftz.f32 %f410, %f409; div.approx.ftz.f32 %f411, %f410, %f1395; mul.ftz.f32 %f412, %f1389, %f1389; fma.rn.ftz.f32 %f413, %f1388, %f1388, %f412; fma.rn.ftz.f32 %f414, %f1390, %f1390, %f413; sqrt.approx.ftz.f32 %f415, %f414; rsqrt.approx.ftz.f32 %f416, %f414; mul.ftz.f32 %f40, %f1388, %f416; mul.ftz.f32 %f41, %f1389, %f416; mul.ftz.f32 %f42, %f1390, %f416; mul.ftz.f32 %f417, %f1386, %f1386; fma.rn.ftz.f32 %f418, %f1385, %f1385, %f417; fma.rn.ftz.f32 %f419, %f1387, %f1387, %f418; sqrt.approx.ftz.f32 %f420, %f419; rsqrt.approx.ftz.f32 %f421, %f419; mul.ftz.f32 %f46, %f1385, %f421; mul.ftz.f32 %f47, %f1386, %f421; mul.ftz.f32 %f48, %f1387, %f421; sub.ftz.f32 %f422, %f401, %f1391; sub.ftz.f32 %f423, %f402, %f1392; sub.ftz.f32 %f424, %f403, %f1393; mul.ftz.f32 %f425, %f423, %f41; fma.rn.ftz.f32 %f426, %f422, %f40, %f425; fma.rn.ftz.f32 %f427, %f424, %f42, %f426; mul.ftz.f32 %f428, %f423, %f47; fma.rn.ftz.f32 %f429, %f422, %f46, %f428; fma.rn.ftz.f32 %f430, %f424, %f48, %f429; sub.ftz.f32 %f431, %f427, %f411; mul.ftz.f32 %f432, %f415, 0fBF000000; max.ftz.f32 %f49, %f431, %f432; add.ftz.f32 %f433, %f411, %f427; mul.ftz.f32 %f434, %f415, 0f3F000000; min.ftz.f32 %f50, %f433, %f434; sub.ftz.f32 %f435, %f430, %f411; mul.ftz.f32 %f436, %f420, 0fBF000000; max.ftz.f32 %f51, %f435, %f436; add.ftz.f32 %f437, %f411, %f430; mul.ftz.f32 %f438, %f420, 0f3F000000; min.ftz.f32 %f52, %f437, %f438; setp.ge.ftz.f32 %p15, %f49, %f50; setp.ge.ftz.f32 %p16, %f51, %f52; or.pred %p17, %p15, %p16; mov.pred %p305, -1; mov.pred %p304, %p305; @%p17 bra $L__BB210_12; add.ftz.f32 %f439, %f49, %f50; mul.ftz.f32 %f440, %f439, 0f3F000000; add.ftz.f32 %f441, %f51, %f52; mul.ftz.f32 %f442, %f441, 0f3F000000; fma.rn.ftz.f32 %f443, %f40, %f440, %f1391; fma.rn.ftz.f32 %f444, %f41, %f440, %f1392; fma.rn.ftz.f32 %f445, %f42, %f440, %f1393; fma.rn.ftz.f32 %f1391, %f46, %f442, %f443; fma.rn.ftz.f32 %f1392, %f47, %f442, %f444; fma.rn.ftz.f32 %f1393, %f48, %f442, %f445; sub.ftz.f32 %f446, %f50, %f49; mul.ftz.f32 %f1388, %f40, %f446; mul.ftz.f32 %f1389, %f41, %f446; mul.ftz.f32 %f1390, %f42, %f446; sub.ftz.f32 %f447, %f52, %f51; mul.ftz.f32 %f1385, %f46, %f447; mul.ftz.f32 %f1386, %f47, %f447; mul.ftz.f32 %f1387, %f48, %f447; mov.pred %p304, 0; $L__BB210_12: @%p304 bra $L__BB210_52; $L__BB210_13: mul.ftz.f32 %f450, %f1388, 0f3F000000; sub.ftz.f32 %f451, %f1391, %f450; mul.ftz.f32 %f452, %f1389, 0f3F000000; sub.ftz.f32 %f453, %f1392, %f452; mul.ftz.f32 %f454, %f1390, 0f3F000000; sub.ftz.f32 %f455, %f1393, %f454; mul.ftz.f32 %f456, %f1385, 0f3F000000; sub.ftz.f32 %f457, %f451, %f456; mul.ftz.f32 %f458, %f1386, 0f3F000000; sub.ftz.f32 %f459, %f453, %f458; mul.ftz.f32 %f460, %f1387, 0f3F000000; sub.ftz.f32 %f461, %f455, %f460; mul.ftz.f32 %f462, %f1389, %f1389; fma.rn.ftz.f32 %f463, %f1388, %f1388, %f462; fma.rn.ftz.f32 %f464, %f1390, %f1390, %f463; sqrt.approx.ftz.f32 %f465, %f464; rsqrt.approx.ftz.f32 %f466, %f464; mov.f32 %f467, 0f3F800000; mul.ftz.f32 %f468, %f1388, %f466; mul.ftz.f32 %f469, %f1389, %f466; mul.ftz.f32 %f470, %f1390, %f466; mul.ftz.f32 %f471, %f1386, %f1386; fma.rn.ftz.f32 %f472, %f1385, %f1385, %f471; fma.rn.ftz.f32 %f473, %f1387, %f1387, %f472; sqrt.approx.ftz.f32 %f474, %f473; rsqrt.approx.ftz.f32 %f475, %f473; mul.ftz.f32 %f476, %f1385, %f475; mul.ftz.f32 %f477, %f1386, %f475; mul.ftz.f32 %f478, %f1387, %f475; mul.ftz.f32 %f479, %f469, %f478; mul.ftz.f32 %f480, %f470, %f477; sub.ftz.f32 %f481, %f479, %f480; mul.ftz.f32 %f482, %f470, %f476; mul.ftz.f32 %f483, %f468, %f478; sub.ftz.f32 %f484, %f482, %f483; mul.ftz.f32 %f485, %f468, %f477; mul.ftz.f32 %f486, %f469, %f476; sub.ftz.f32 %f487, %f485, %f486; sub.ftz.f32 %f488, %f457, %f3; sub.ftz.f32 %f489, %f459, %f4; sub.ftz.f32 %f490, %f461, %f5; mul.ftz.f32 %f491, %f489, %f484; fma.rn.ftz.f32 %f492, %f488, %f481, %f491; fma.rn.ftz.f32 %f493, %f490, %f487, %f492; setp.gt.ftz.f32 %p20, %f493, 0f00000000; mov.f32 %f1409, 0f00000000; neg.ftz.f32 %f494, %f493; selp.f32 %f495, %f494, %f493, %p20; mul.ftz.f32 %f496, %f489, %f469; fma.rn.ftz.f32 %f497, %f488, %f468, %f496; fma.rn.ftz.f32 %f498, %f490, %f470, %f497; mul.ftz.f32 %f499, %f489, %f477; fma.rn.ftz.f32 %f500, %f488, %f476, %f499; fma.rn.ftz.f32 %f501, %f490, %f478, %f500; add.ftz.f32 %f502, %f465, %f498; add.ftz.f32 %f503, %f474, %f501; sub.ftz.f32 %f504, %f498, %f502; sub.ftz.f32 %f505, %f503, %f501; sub.ftz.f32 %f506, %f502, %f498; sub.ftz.f32 %f507, %f501, %f503; mul.ftz.f32 %f508, %f501, %f504; mul.ftz.f32 %f509, %f502, %f505; mul.ftz.f32 %f510, %f503, %f506; mul.ftz.f32 %f511, %f498, %f507; mul.ftz.f32 %f512, %f495, %f495; mul.ftz.f32 %f513, %f504, %f512; mul.ftz.f32 %f514, %f505, %f512; mul.ftz.f32 %f515, %f506, %f512; mul.ftz.f32 %f516, %f507, %f512; mul.ftz.f32 %f517, %f504, %f513; mul.ftz.f32 %f518, %f505, %f514; mul.ftz.f32 %f519, %f506, %f515; mul.ftz.f32 %f520, %f507, %f516; fma.rn.ftz.f32 %f521, %f508, %f508, %f517; fma.rn.ftz.f32 %f522, %f509, %f509, %f518; fma.rn.ftz.f32 %f523, %f510, %f510, %f519; fma.rn.ftz.f32 %f524, %f511, %f511, %f520; sqrt.approx.ftz.f32 %f525, %f521; sqrt.approx.ftz.f32 %f526, %f522; sqrt.approx.ftz.f32 %f527, %f523; sqrt.approx.ftz.f32 %f528, %f524; div.approx.ftz.f32 %f529, %f508, %f525; div.approx.ftz.f32 %f530, %f509, %f526; div.approx.ftz.f32 %f531, %f510, %f527; div.approx.ftz.f32 %f532, %f511, %f528; mul.ftz.f32 %f533, %f529, %f530; neg.ftz.f32 %f534, %f533; mov.f32 %f535, 0fBF800000; max.ftz.f32 %f536, %f534, %f535; min.ftz.f32 %f537, %f536, %f467; abs.ftz.f32 %f538, %f537; sub.ftz.f32 %f539, %f467, %f538; mul.ftz.f32 %f540, %f539, 0f3F000000; sqrt.approx.ftz.f32 %f541, %f540; setp.gt.ftz.f32 %p21, %f538, 0f3F11EB85; selp.f32 %f542, %f541, %f538, %p21; mul.ftz.f32 %f543, %f542, %f542; mov.f32 %f544, 0f3C94D2E9; mov.f32 %f545, 0f3D53F941; fma.rn.ftz.f32 %f546, %f545, %f543, %f544; mov.f32 %f547, 0f3D3F841F; fma.rn.ftz.f32 %f548, %f546, %f543, %f547; mov.f32 %f549, 0f3D994929; fma.rn.ftz.f32 %f550, %f548, %f543, %f549; mov.f32 %f551, 0f3E2AAB94; fma.rn.ftz.f32 %f552, %f550, %f543, %f551; mul.ftz.f32 %f553, %f543, %f552; fma.rn.ftz.f32 %f554, %f553, %f542, %f542; add.ftz.f32 %f555, %f554, %f554; mov.f32 %f556, 0f3FC90FDB; sub.ftz.f32 %f557, %f556, %f554; selp.f32 %f558, %f555, %f557, %p21; setp.lt.ftz.f32 %p22, %f537, 0f00000000; mov.f32 %f559, 0f40490FDB; sub.ftz.f32 %f560, %f559, %f558; selp.f32 %f561, %f560, %f558, %p22; mul.ftz.f32 %f562, %f530, %f531; neg.ftz.f32 %f563, %f562; max.ftz.f32 %f564, %f563, %f535; min.ftz.f32 %f565, %f564, %f467; abs.ftz.f32 %f566, %f565; sub.ftz.f32 %f567, %f467, %f566; mul.ftz.f32 %f568, %f567, 0f3F000000; sqrt.approx.ftz.f32 %f569, %f568; setp.gt.ftz.f32 %p23, %f566, 0f3F11EB85; selp.f32 %f570, %f569, %f566, %p23; mul.ftz.f32 %f571, %f570, %f570; fma.rn.ftz.f32 %f572, %f545, %f571, %f544; fma.rn.ftz.f32 %f573, %f572, %f571, %f547; fma.rn.ftz.f32 %f574, %f573, %f571, %f549; fma.rn.ftz.f32 %f575, %f574, %f571, %f551; mul.ftz.f32 %f576, %f571, %f575; fma.rn.ftz.f32 %f577, %f576, %f570, %f570; add.ftz.f32 %f578, %f577, %f577; sub.ftz.f32 %f579, %f556, %f577; selp.f32 %f580, %f578, %f579, %p23; setp.lt.ftz.f32 %p24, %f565, 0f00000000; sub.ftz.f32 %f581, %f559, %f580; selp.f32 %f582, %f581, %f580, %p24; mul.ftz.f32 %f583, %f531, %f532; neg.ftz.f32 %f584, %f583; max.ftz.f32 %f585, %f584, %f535; min.ftz.f32 %f586, %f585, %f467; abs.ftz.f32 %f587, %f586; sub.ftz.f32 %f588, %f467, %f587; mul.ftz.f32 %f589, %f588, 0f3F000000; sqrt.approx.ftz.f32 %f590, %f589; setp.gt.ftz.f32 %p25, %f587, 0f3F11EB85; selp.f32 %f591, %f590, %f587, %p25; mul.ftz.f32 %f592, %f591, %f591; fma.rn.ftz.f32 %f593, %f545, %f592, %f544; fma.rn.ftz.f32 %f594, %f593, %f592, %f547; fma.rn.ftz.f32 %f595, %f594, %f592, %f549; fma.rn.ftz.f32 %f596, %f595, %f592, %f551; mul.ftz.f32 %f597, %f592, %f596; fma.rn.ftz.f32 %f598, %f597, %f591, %f591; add.ftz.f32 %f599, %f598, %f598; sub.ftz.f32 %f600, %f556, %f598; selp.f32 %f601, %f599, %f600, %p25; setp.lt.ftz.f32 %p26, %f586, 0f00000000; sub.ftz.f32 %f602, %f559, %f601; selp.f32 %f603, %f602, %f601, %p26; mul.ftz.f32 %f604, %f529, %f532; neg.ftz.f32 %f605, %f604; max.ftz.f32 %f606, %f605, %f535; min.ftz.f32 %f607, %f606, %f467; abs.ftz.f32 %f608, %f607; sub.ftz.f32 %f609, %f467, %f608; mul.ftz.f32 %f610, %f609, 0f3F000000; sqrt.approx.ftz.f32 %f611, %f610; setp.gt.ftz.f32 %p27, %f608, 0f3F11EB85; selp.f32 %f612, %f611, %f608, %p27; mul.ftz.f32 %f613, %f612, %f612; fma.rn.ftz.f32 %f614, %f545, %f613, %f544; fma.rn.ftz.f32 %f615, %f614, %f613, %f547; fma.rn.ftz.f32 %f616, %f615, %f613, %f549; fma.rn.ftz.f32 %f617, %f616, %f613, %f551; mul.ftz.f32 %f618, %f613, %f617; fma.rn.ftz.f32 %f619, %f618, %f612, %f612; add.ftz.f32 %f620, %f619, %f619; sub.ftz.f32 %f621, %f556, %f619; selp.f32 %f622, %f620, %f621, %p27; setp.lt.ftz.f32 %p28, %f607, 0f00000000; sub.ftz.f32 %f623, %f559, %f622; selp.f32 %f624, %f623, %f622, %p28; mov.f32 %f625, 0f40C90FDB; sub.ftz.f32 %f626, %f625, %f603; sub.ftz.f32 %f627, %f626, %f624; add.ftz.f32 %f628, %f561, %f582; sub.ftz.f32 %f83, %f628, %f627; setp.eq.ftz.f32 %p29, %f83, 0f00000000; @%p29 bra $L__BB210_18; rcp.approx.ftz.f32 %f1409, %f83; $L__BB210_18: mul.ftz.f32 %f1415, %f18, 0f3E800000; setp.leu.ftz.f32 %p31, %f1395, 0f00000000; @%p31 bra $L__BB210_51; ld.const.u64 %rd519, [kernel_params+2368]; mul.wide.s32 %rd518, %r4, 192; cvta.to.global.u64 %rd517, %rd519; add.s64 %rd516, %rd517, %rd518; mov.pred %p305, -1; mul.ftz.f32 %f636, %f7, %f1420; fma.rn.ftz.f32 %f637, %f6, %f1421, %f636; fma.rn.ftz.f32 %f638, %f8, %f1419, %f637; neg.ftz.f32 %f639, %f638; mul.ftz.f32 %f640, %f638, %f638; mov.f32 %f641, 0f3F800000; sub.ftz.f32 %f642, %f641, %f640; mov.f32 %f643, 0f00000000; max.ftz.f32 %f644, %f642, %f643; sqrt.approx.ftz.f32 %f645, %f644; div.approx.ftz.f32 %f646, %f645, %f639; mul.ftz.f32 %f647, %f1395, %f646; sub.ftz.f32 %f648, %f641, %f647; ld.global.f32 %f649, [%rd516+188]; mul.ftz.f32 %f650, %f649, %f648; max.ftz.f32 %f651, %f650, %f643; mul.ftz.f32 %f1415, %f1415, %f651; setp.eq.ftz.f32 %p33, %f1415, 0f00000000; mov.f32 %f1416, %f1409; @%p33 bra $L__BB210_52; bra.uni $L__BB210_51; $L__BB210_41: setp.eq.ftz.f32 %p67, %f139, 0f7F800000; setp.eq.ftz.f32 %p68, %f140, 0f7F800000; and.pred %p69, %p67, %p68; @%p69 bra $L__BB210_43; bra.uni $L__BB210_42; $L__BB210_43: setp.lt.s32 %p73, %r11, 0; selp.b32 %r151, 1075235812, 1061752795, %p73; or.b32 %r152, %r151, %r12; mov.b32 %f1403, %r152; bra.uni $L__BB210_45; $L__BB210_42: setp.lt.s32 %p70, %r11, 0; min.ftz.f32 %f765, %f140, %f139; max.ftz.f32 %f766, %f140, %f139; div.full.ftz.f32 %f767, %f765, %f766; mul.rn.ftz.f32 %f768, %f767, %f767; mov.f32 %f769, 0fC0B59883; mov.f32 %f770, 0fBF52C7EA; fma.rn.ftz.f32 %f771, %f768, %f770, %f769; mov.f32 %f772, 0fC0D21907; fma.rn.ftz.f32 %f773, %f771, %f768, %f772; mul.ftz.f32 %f774, %f768, %f773; mul.ftz.f32 %f775, %f767, %f774; add.ftz.f32 %f776, %f768, 0f41355DC0; mov.f32 %f777, 0f41E6BD60; fma.rn.ftz.f32 %f778, %f776, %f768, %f777; mov.f32 %f779, 0f419D92C8; fma.rn.ftz.f32 %f780, %f778, %f768, %f779; rcp.approx.ftz.f32 %f781, %f780; fma.rn.ftz.f32 %f782, %f775, %f781, %f767; mov.f32 %f783, 0f3FC90FDB; sub.ftz.f32 %f784, %f783, %f782; setp.gt.ftz.f32 %p71, %f140, %f139; selp.f32 %f785, %f784, %f782, %p71; mov.f32 %f786, 0f40490FDB; sub.ftz.f32 %f787, %f786, %f785; selp.f32 %f788, %f787, %f785, %p70; mov.b32 %r149, %f788; or.b32 %r150, %r12, %r149; mov.b32 %f789, %r150; add.ftz.f32 %f790, %f139, %f140; setp.le.ftz.f32 %p72, %f790, 0f7F800000; selp.f32 %f1403, %f789, %f790, %p72; $L__BB210_45: fma.rn.ftz.f32 %f1418, %f1403, 0fBE22F983, 0f3F000000; $L__BB210_46: sqrt.approx.ftz.f32 %f791, %f138; div.approx.ftz.f32 %f792, %f1419, %f791; mov.f32 %f793, 0fBF800000; max.ftz.f32 %f794, %f792, %f793; mov.f32 %f795, 0f3F800000; min.ftz.f32 %f796, %f794, %f795; abs.ftz.f32 %f797, %f796; sub.ftz.f32 %f798, %f795, %f797; mul.ftz.f32 %f799, %f798, 0f3F000000; sqrt.approx.ftz.f32 %f800, %f799; setp.gt.ftz.f32 %p74, %f797, 0f3F11EB85; selp.f32 %f801, %f800, %f797, %p74; mul.ftz.f32 %f802, %f801, %f801; mov.f32 %f803, 0f3C94D2E9; mov.f32 %f804, 0f3D53F941; fma.rn.ftz.f32 %f805, %f804, %f802, %f803; mov.f32 %f806, 0f3D3F841F; fma.rn.ftz.f32 %f807, %f805, %f802, %f806; mov.f32 %f808, 0f3D994929; fma.rn.ftz.f32 %f809, %f807, %f802, %f808; mov.f32 %f810, 0f3E2AAB94; fma.rn.ftz.f32 %f811, %f809, %f802, %f810; mul.ftz.f32 %f812, %f802, %f811; fma.rn.ftz.f32 %f813, %f812, %f801, %f801; add.ftz.f32 %f814, %f813, %f813; mov.f32 %f815, 0f3FC90FDB; sub.ftz.f32 %f816, %f815, %f813; selp.f32 %f817, %f814, %f816, %p74; setp.lt.ftz.f32 %p75, %f796, 0f00000000; mov.f32 %f818, 0f40490FDB; sub.ftz.f32 %f819, %f818, %f817; selp.f32 %f820, %f819, %f817, %p75; fma.rn.ftz.f32 %f1417, %f820, 0fBEA2F983, 0f3F800000; $L__BB210_47: setp.eq.ftz.f32 %p76, %f10, 0f7F7FFFFF; @%p76 bra $L__BB210_51; mul.ftz.f32 %f823, %f125, %f1420; fma.rn.ftz.f32 %f824, %f124, %f1421, %f823; fma.rn.ftz.f32 %f150, %f126, %f1419, %f824; setp.le.ftz.f32 %p77, %f150, 0f00000000; mov.f32 %f1407, 0f00000000; @%p77 bra $L__BB210_50; mul.ftz.f32 %f825, %f10, %f10; div.approx.ftz.f32 %f1407, %f825, %f150; $L__BB210_50: mul.ftz.f32 %f1409, %f1416, %f1407; bra.uni $L__BB210_51; $L__BB210_24: setp.eq.ftz.f32 %p43, %f109, 0f7F800000; setp.eq.ftz.f32 %p44, %f110, 0f7F800000; and.pred %p45, %p43, %p44; @%p45 bra $L__BB210_26; bra.uni $L__BB210_25; $L__BB210_26: setp.lt.s32 %p49, %r9, 0; selp.b32 %r143, 1075235812, 1061752795, %p49; or.b32 %r144, %r143, %r10; mov.b32 %f1397, %r144; bra.uni $L__BB210_28; $L__BB210_25: setp.lt.s32 %p46, %r9, 0; min.ftz.f32 %f668, %f110, %f109; max.ftz.f32 %f669, %f110, %f109; div.full.ftz.f32 %f670, %f668, %f669; mul.rn.ftz.f32 %f671, %f670, %f670; mov.f32 %f672, 0fC0B59883; mov.f32 %f673, 0fBF52C7EA; fma.rn.ftz.f32 %f674, %f671, %f673, %f672; mov.f32 %f675, 0fC0D21907; fma.rn.ftz.f32 %f676, %f674, %f671, %f675; mul.ftz.f32 %f677, %f671, %f676; mul.ftz.f32 %f678, %f670, %f677; add.ftz.f32 %f679, %f671, 0f41355DC0; mov.f32 %f680, 0f41E6BD60; fma.rn.ftz.f32 %f681, %f679, %f671, %f680; mov.f32 %f682, 0f419D92C8; fma.rn.ftz.f32 %f683, %f681, %f671, %f682; rcp.approx.ftz.f32 %f684, %f683; fma.rn.ftz.f32 %f685, %f678, %f684, %f670; mov.f32 %f686, 0f3FC90FDB; sub.ftz.f32 %f687, %f686, %f685; setp.gt.ftz.f32 %p47, %f110, %f109; selp.f32 %f688, %f687, %f685, %p47; mov.f32 %f689, 0f40490FDB; sub.ftz.f32 %f690, %f689, %f688; selp.f32 %f691, %f690, %f688, %p46; mov.b32 %r141, %f691; or.b32 %r142, %r10, %r141; mov.b32 %f692, %r142; add.ftz.f32 %f693, %f109, %f110; setp.le.ftz.f32 %p48, %f693, 0f7F800000; selp.f32 %f1397, %f692, %f693, %p48; $L__BB210_28: fma.rn.ftz.f32 %f1418, %f1397, 0fBE22F983, 0f3F000000; $L__BB210_29: sqrt.approx.ftz.f32 %f694, %f108; div.approx.ftz.f32 %f695, %f1419, %f694; mov.f32 %f696, 0fBF800000; max.ftz.f32 %f697, %f695, %f696; mov.f32 %f698, 0f3F800000; min.ftz.f32 %f699, %f697, %f698; abs.ftz.f32 %f700, %f699; sub.ftz.f32 %f701, %f698, %f700; mul.ftz.f32 %f702, %f701, 0f3F000000; sqrt.approx.ftz.f32 %f703, %f702; setp.gt.ftz.f32 %p50, %f700, 0f3F11EB85; selp.f32 %f704, %f703, %f700, %p50; mul.ftz.f32 %f705, %f704, %f704; mov.f32 %f706, 0f3C94D2E9; mov.f32 %f707, 0f3D53F941; fma.rn.ftz.f32 %f708, %f707, %f705, %f706; mov.f32 %f709, 0f3D3F841F; fma.rn.ftz.f32 %f710, %f708, %f705, %f709; mov.f32 %f711, 0f3D994929; fma.rn.ftz.f32 %f712, %f710, %f705, %f711; mov.f32 %f713, 0f3E2AAB94; fma.rn.ftz.f32 %f714, %f712, %f705, %f713; mul.ftz.f32 %f715, %f705, %f714; fma.rn.ftz.f32 %f716, %f715, %f704, %f704; add.ftz.f32 %f717, %f716, %f716; mov.f32 %f718, 0f3FC90FDB; sub.ftz.f32 %f719, %f718, %f716; selp.f32 %f720, %f717, %f719, %p50; setp.lt.ftz.f32 %p51, %f699, 0f00000000; mov.f32 %f721, 0f40490FDB; sub.ftz.f32 %f722, %f721, %f720; selp.f32 %f723, %f722, %f720, %p51; fma.rn.ftz.f32 %f1417, %f723, 0fBEA2F983, 0f3F800000; $L__BB210_30: setp.eq.ftz.f32 %p52, %f10, 0f7F7FFFFF; @%p52 bra $L__BB210_51; mul.ftz.f32 %f726, %f101, %f1420; fma.rn.ftz.f32 %f727, %f100, %f1421, %f726; fma.rn.ftz.f32 %f120, %f102, %f1419, %f727; setp.le.ftz.f32 %p53, %f120, 0f00000000; mov.f32 %f1401, 0f00000000; @%p53 bra $L__BB210_33; mul.ftz.f32 %f728, %f10, %f10; div.approx.ftz.f32 %f1401, %f728, %f120; $L__BB210_33: mul.ftz.f32 %f1409, %f1416, %f1401; $L__BB210_51: ld.const.f32 %f826, [kernel_params+1504]; mul.ftz.f32 %f1416, %f1409, %f826; mov.pred %p305, 0; $L__BB210_52: @%p305 bra $L__BB210_353; ld.const.u64 %rd80, [kernel_params+2544]; cvta.to.global.u64 %rd81, %rd80; shl.b64 %rd82, %rd1, 2; add.s64 %rd4, %rd81, %rd82; ld.global.u32 %r13, [%rd4]; and.b32 %r156, %r6, 264241152; setp.eq.s32 %p79, %r156, 0; @%p79 bra $L__BB210_58; shr.u32 %r157, %r6, 27; and.b32 %r158, %r157, 1; setp.eq.b32 %p80, %r158, 1; shr.u32 %r159, %r13, 3; and.b32 %r160, %r159, 1; setp.eq.b32 %p81, %r160, 1; and.pred %p82, %p80, %p81; @%p82 bra $L__BB210_353; shr.u32 %r161, %r6, 26; and.b32 %r162, %r161, 1; setp.eq.b32 %p83, %r162, 1; and.b32 %r163, %r13, 18; setp.eq.s32 %p84, %r163, 18; and.pred %p85, %p83, %p84; @%p85 bra $L__BB210_353; shr.u32 %r164, %r6, 25; and.b32 %r165, %r164, 1; setp.eq.b32 %p86, %r165, 1; shr.u32 %r166, %r13, 2; and.b32 %r167, %r166, 1; setp.eq.b32 %p87, %r167, 1; and.pred %p88, %p86, %p87; @%p88 bra $L__BB210_353; shr.u32 %r168, %r6, 23; and.b32 %r169, %r168, 1; setp.eq.b32 %p89, %r169, 1; shr.u32 %r170, %r13, 7; and.b32 %r171, %r170, 1; setp.eq.b32 %p90, %r171, 1; and.pred %p91, %p89, %p90; @%p91 bra $L__BB210_353; $L__BB210_58: ld.const.u64 %rd83, [kernel_params+2408]; cvta.to.global.u64 %rd5, %rd83; shl.b32 %r172, %r6, 5; cvt.u64.u32 %rd84, %r172; and.b64 %rd85, %rd84, 134217696; add.s64 %rd86, %rd5, %rd85; add.s64 %rd6, %rd86, 16; ld.global.u32 %r15, [%rd86+16]; and.b32 %r173, %r15, 134217728; setp.eq.s32 %p93, %r173, 0; mov.pred %p306, -1; mov.f32 %f1478, 0f00000000; mov.f32 %f1479, %f1478; mov.f32 %f1480, %f1478; @%p93 bra $L__BB210_60; ld.global.f32 %f1478, [%rd6+-16]; ld.global.f32 %f1479, [%rd6+-12]; ld.global.f32 %f1480, [%rd6+-8]; mov.pred %p306, 0; $L__BB210_60: not.pred %p95, %p306; @%p95 bra $L__BB210_306; setp.eq.s32 %p96, %r5, 2; neg.ftz.f32 %f175, %f6; neg.ftz.f32 %f176, %f7; neg.ftz.f32 %f177, %f8; add.u64 %rd87, %SP, 1024; or.b64 %rd7, %rd87, 12; mov.b32 %r16, %f9; @%p96 bra $L__BB210_94; bra.uni $L__BB210_62; $L__BB210_94: fma.rn.ftz.f32 %f1375, %f10, %f8, %f5; fma.rn.ftz.f32 %f1374, %f10, %f7, %f4; fma.rn.ftz.f32 %f1373, %f10, %f6, %f3; add.s64 %rd220, %rd7, 36; add.s64 %rd221, %rd7, 20; add.s64 %rd222, %rd7, 4; add.s64 %rd224, %rd87, 72; add.s64 %rd225, %rd87, 68; add.s64 %rd226, %rd87, 76; add.s64 %rd227, %rd87, 248; cvta.to.local.u64 %rd228, %rd227; add.s64 %rd229, %rd87, 240; cvta.to.local.u64 %rd230, %rd229; st.local.v2.f32 [%rd230], {%f1373, %f1374}; st.local.f32 [%rd228], %f1375; cvta.to.local.u64 %rd231, %rd87; st.local.v4.f32 [%rd231], {%f6, %f7, %f8, %f175}; cvta.to.local.u64 %rd232, %rd222; st.local.v4.f32 [%rd232], {%f176, %f177, %f175, %f176}; cvta.to.local.u64 %rd233, %rd221; st.local.v4.f32 [%rd233], {%f177, %f175, %f176, %f177}; ld.const.u32 %r201, [kernel_params+1024]; and.b32 %r325, %r201, 4194303; shl.b32 %r202, %r201, 5; cvt.u64.u32 %rd234, %r202; and.b64 %rd235, %rd234, 134217696; add.s64 %rd236, %rd5, %rd235; cvta.to.local.u64 %rd237, %rd226; ld.global.u32 %r203, [%rd236+16]; mov.u32 %r204, -1; st.local.u32 [%rd237], %r204; mov.u32 %r205, 0; cvta.to.local.u64 %rd238, %rd220; st.local.v4.u32 [%rd238], {%r201, %r203, %r205, %r204}; cvta.to.local.u64 %rd239, %rd225; st.local.u32 [%rd239], %r205; cvta.to.local.u64 %rd240, %rd224; st.local.u32 [%rd240], %r205; add.s64 %rd241, %rd87, 136; cvta.to.local.u64 %rd242, %rd241; add.s64 %rd243, %rd87, 128; cvta.to.local.u64 %rd244, %rd243; mov.f32 %f1239, 0f00000000; st.local.v2.f32 [%rd244], {%f1239, %f1239}; st.local.u32 [%rd242], %r205; add.s64 %rd245, %rd7, 68; cvta.to.local.u64 %rd246, %rd245; st.local.v4.u32 [%rd246], {%r204, %r16, %r205, %r205}; add.s64 %rd247, %rd87, 96; cvta.to.local.u64 %rd248, %rd247; st.local.v4.f32 [%rd248], {%f1239, %f1239, %f1239, %f1239}; add.s64 %rd249, %rd87, 112; cvta.to.local.u64 %rd250, %rd249; st.local.v4.f32 [%rd250], {%f1239, %f1239, %f1239, %f1239}; bra.uni $L__BB210_95; $L__BB210_62: fma.rn.ftz.f32 %f1372, %f10, %f8, %f5; fma.rn.ftz.f32 %f1371, %f10, %f7, %f4; fma.rn.ftz.f32 %f1370, %f10, %f6, %f3; add.s64 %rd88, %rd7, 36; add.s64 %rd89, %rd7, 20; add.s64 %rd90, %rd7, 4; add.s64 %rd92, %rd87, 88; cvta.to.local.u64 %rd93, %rd87; st.local.v4.f32 [%rd93], {%f1370, %f1371, %f1372, %f1421}; cvta.to.local.u64 %rd8, %rd90; st.local.v4.f32 [%rd8], {%f1420, %f1419, %f1421, %f1420}; cvta.to.local.u64 %rd94, %rd89; st.local.v4.f32 [%rd94], {%f1419, %f175, %f176, %f177}; setp.ne.s32 %p97, %r4, -1; cvt.u32.u64 %r174, %rd2; mov.b32 %r175, %f1417; mov.b32 %r176, %f1418; selp.u32 %r177, 1, 0, %p97; add.s64 %rd95, %rd87, 64; cvta.to.local.u64 %rd96, %rd95; st.local.v4.u32 [%rd96], {%r177, %r176, %r175, %r174}; add.s64 %rd97, %rd7, 68; cvta.to.local.u64 %rd98, %rd97; mov.u32 %r178, -1; st.local.v2.u32 [%rd98], {%r178, %r16}; cvta.to.local.u64 %rd99, %rd92; st.local.f32 [%rd99], %f10; mov.u32 %r179, 0; cvta.to.local.u64 %rd100, %rd88; st.local.v4.u32 [%rd100], {%r6, %r15, %r179, %r4}; setp.eq.s32 %p98, %r174, -1; @%p98 bra $L__BB210_87; add.s64 %rd101, %rd7, 44; ld.const.u64 %rd102, [kernel_params+2184]; cvta.to.global.u64 %rd103, %rd102; shl.b64 %rd104, %rd2, 2; add.s64 %rd105, %rd103, %rd104; ld.global.u32 %r17, [%rd105]; cvta.to.local.u64 %rd106, %rd101; st.local.u32 [%rd106], %r17; and.b32 %r18, %r17, 2; setp.eq.s32 %p99, %r18, 0; @%p99 bra $L__BB210_70; ld.const.u64 %rd107, [kernel_params+2160]; cvta.to.global.u64 %rd108, %rd107; mul.lo.s64 %rd109, %rd2, 208; add.s64 %rd110, %rd108, %rd109; ld.const.u64 %rd111, [kernel_params+2176]; cvta.to.global.u64 %rd112, %rd111; ld.global.u32 %rd113, [%rd110+168]; ld.global.u32 %r180, [%rd110+152]; shl.b32 %r181, %r180, 1; cvt.rn.f32.s32 %f831, %r181; mul.ftz.f32 %f832, %f9, %f831; cvt.rzi.ftz.s32.f32 %r182, %f832; add.s32 %r183, %r181, -1; min.s32 %r184, %r182, %r183; cvt.rn.f32.s32 %f833, %r184; sub.ftz.f32 %f178, %f832, %f833; cvt.s64.s32 %rd114, %r184; add.s64 %rd115, %rd114, %rd113; shl.b64 %rd116, %rd115, 6; add.s64 %rd9, %rd112, %rd116; ld.global.v4.f32 {%f834, %f835, %f836, %f837}, [%rd9]; ld.global.v4.f32 {%f838, %f839, %f840, %f841}, [%rd9+64]; mul.ftz.f32 %f842, %f835, %f839; fma.rn.ftz.f32 %f843, %f834, %f838, %f842; mul.ftz.f32 %f844, %f837, %f841; fma.rn.ftz.f32 %f845, %f836, %f840, %f844; add.ftz.f32 %f187, %f843, %f845; setp.gt.ftz.f32 %p100, %f187, 0f3F7FDF3B; mov.f32 %f846, 0f3F800000; sub.ftz.f32 %f188, %f846, %f178; @%p100 bra $L__BB210_66; bra.uni $L__BB210_65; $L__BB210_66: mul.ftz.f32 %f900, %f834, %f188; fma.rn.ftz.f32 %f901, %f838, %f178, %f900; mul.ftz.f32 %f902, %f835, %f188; fma.rn.ftz.f32 %f903, %f839, %f178, %f902; mul.ftz.f32 %f904, %f836, %f188; fma.rn.ftz.f32 %f905, %f840, %f178, %f904; mul.ftz.f32 %f906, %f837, %f188; fma.rn.ftz.f32 %f907, %f841, %f178, %f906; mul.ftz.f32 %f908, %f903, %f903; fma.rn.ftz.f32 %f909, %f901, %f901, %f908; mul.ftz.f32 %f910, %f907, %f907; fma.rn.ftz.f32 %f911, %f905, %f905, %f910; add.ftz.f32 %f912, %f909, %f911; rsqrt.approx.ftz.f32 %f913, %f912; mul.ftz.f32 %f1428, %f907, %f913; mul.ftz.f32 %f1427, %f905, %f913; mul.ftz.f32 %f1426, %f903, %f913; mul.ftz.f32 %f1425, %f901, %f913; bra.uni $L__BB210_67; $L__BB210_87: @%p7 bra $L__BB210_89; st.local.u32 [%rd98], %r4; $L__BB210_89: add.s64 %rd187, %rd87, 116; cvta.to.local.u64 %rd188, %rd187; mov.u32 %r198, 0; st.local.u32 [%rd188], %r198; add.s64 %rd189, %rd87, 120; cvta.to.local.u64 %rd190, %rd189; st.local.u32 [%rd190], %r198; add.s64 %rd191, %rd87, 124; cvta.to.local.u64 %rd192, %rd191; st.local.u32 [%rd192], %r198; add.s64 %rd193, %rd87, 136; cvta.to.local.u64 %rd194, %rd193; add.s64 %rd195, %rd87, 128; cvta.to.local.u64 %rd196, %rd195; mov.f32 %f1463, 0f00000000; st.local.v2.f32 [%rd196], {%f1463, %f1463}; st.local.u32 [%rd194], %r198; mov.f32 %f1464, %f1463; mov.f32 %f1465, %f1463; mov.f32 %f1466, %f1463; mov.f32 %f1467, %f1463; mov.f32 %f1468, %f1463; mov.f32 %f1469, %f1419; mov.f32 %f1470, %f1420; mov.f32 %f1471, %f1421; bra.uni $L__BB210_90; $L__BB210_65: mov.f32 %f847, 0fBF800000; max.ftz.f32 %f848, %f187, %f847; min.ftz.f32 %f850, %f848, %f846; abs.ftz.f32 %f851, %f850; sub.ftz.f32 %f852, %f846, %f851; mul.ftz.f32 %f853, %f852, 0f3F000000; sqrt.approx.ftz.f32 %f854, %f853; setp.gt.ftz.f32 %p101, %f851, 0f3F11EB85; selp.f32 %f855, %f854, %f851, %p101; mul.ftz.f32 %f856, %f855, %f855; mov.f32 %f857, 0f3C94D2E9; mov.f32 %f858, 0f3D53F941; fma.rn.ftz.f32 %f859, %f858, %f856, %f857; mov.f32 %f860, 0f3D3F841F; fma.rn.ftz.f32 %f861, %f859, %f856, %f860; mov.f32 %f862, 0f3D994929; fma.rn.ftz.f32 %f863, %f861, %f856, %f862; mov.f32 %f864, 0f3E2AAB94; fma.rn.ftz.f32 %f865, %f863, %f856, %f864; mul.ftz.f32 %f866, %f856, %f865; fma.rn.ftz.f32 %f867, %f866, %f855, %f855; add.ftz.f32 %f868, %f867, %f867; mov.f32 %f869, 0f3FC90FDB; sub.ftz.f32 %f870, %f869, %f867; selp.f32 %f871, %f868, %f870, %p101; setp.lt.ftz.f32 %p102, %f850, 0f00000000; mov.f32 %f872, 0f40490FDB; sub.ftz.f32 %f873, %f872, %f871; selp.f32 %f874, %f873, %f871, %p102; mul.ftz.f32 %f875, %f834, %f187; sub.ftz.f32 %f876, %f838, %f875; mul.ftz.f32 %f877, %f835, %f187; sub.ftz.f32 %f878, %f839, %f877; mul.ftz.f32 %f879, %f836, %f187; sub.ftz.f32 %f880, %f840, %f879; mul.ftz.f32 %f881, %f837, %f187; sub.ftz.f32 %f882, %f841, %f881; mul.ftz.f32 %f883, %f878, %f878; fma.rn.ftz.f32 %f884, %f876, %f876, %f883; mul.ftz.f32 %f885, %f882, %f882; fma.rn.ftz.f32 %f886, %f880, %f880, %f885; add.ftz.f32 %f887, %f884, %f886; rsqrt.approx.ftz.f32 %f888, %f887; mul.ftz.f32 %f889, %f876, %f888; mul.ftz.f32 %f890, %f878, %f888; mul.ftz.f32 %f891, %f880, %f888; mul.ftz.f32 %f892, %f882, %f888; mul.ftz.f32 %f893, %f178, %f874; cos.approx.ftz.f32 %f894, %f893; sin.approx.ftz.f32 %f895, %f893; mul.ftz.f32 %f896, %f889, %f895; mul.ftz.f32 %f897, %f890, %f895; mul.ftz.f32 %f898, %f891, %f895; mul.ftz.f32 %f899, %f892, %f895; fma.rn.ftz.f32 %f1428, %f837, %f894, %f899; fma.rn.ftz.f32 %f1427, %f836, %f894, %f898; fma.rn.ftz.f32 %f1426, %f835, %f894, %f897; fma.rn.ftz.f32 %f1425, %f834, %f894, %f896; $L__BB210_67: ld.global.v4.f32 {%f914, %f915, %f916, %f917}, [%rd9+16]; ld.global.v4.f32 {%f922, %f923, %f924, %f925}, [%rd9+80]; mul.ftz.f32 %f930, %f178, %f922; mul.ftz.f32 %f931, %f178, %f923; mul.ftz.f32 %f932, %f178, %f924; mul.ftz.f32 %f933, %f178, %f925; fma.rn.ftz.f32 %f201, %f188, %f914, %f930; fma.rn.ftz.f32 %f202, %f188, %f915, %f931; fma.rn.ftz.f32 %f203, %f188, %f916, %f932; fma.rn.ftz.f32 %f934, %f188, %f917, %f933; ld.global.v4.f32 {%f935, %f936, %f937, %f938}, [%rd9+32]; ld.global.v4.f32 {%f943, %f944, %f945, %f946}, [%rd9+96]; mul.ftz.f32 %f951, %f178, %f943; mul.ftz.f32 %f952, %f178, %f944; mul.ftz.f32 %f953, %f178, %f945; mul.ftz.f32 %f954, %f178, %f946; fma.rn.ftz.f32 %f955, %f188, %f935, %f951; fma.rn.ftz.f32 %f956, %f188, %f936, %f952; fma.rn.ftz.f32 %f957, %f188, %f937, %f953; fma.rn.ftz.f32 %f958, %f188, %f938, %f954; ld.global.v4.f32 {%f959, %f960, %f961, %f962}, [%rd9+48]; ld.global.v4.f32 {%f967, %f968, %f969, %f970}, [%rd9+112]; mul.ftz.f32 %f975, %f178, %f967; mul.ftz.f32 %f976, %f178, %f968; mul.ftz.f32 %f977, %f178, %f969; mul.ftz.f32 %f978, %f178, %f970; fma.rn.ftz.f32 %f979, %f188, %f959, %f975; fma.rn.ftz.f32 %f980, %f188, %f960, %f976; fma.rn.ftz.f32 %f981, %f188, %f961, %f977; fma.rn.ftz.f32 %f982, %f188, %f962, %f978; mul.ftz.f32 %f983, %f1428, 0f3FB504F3; mul.ftz.f32 %f984, %f1425, 0f3FB504F3; mul.ftz.f32 %f985, %f1426, 0f3FB504F3; mul.ftz.f32 %f986, %f1427, 0f3FB504F3; mul.ftz.f32 %f987, %f983, %f984; mul.ftz.f32 %f988, %f983, %f985; mul.ftz.f32 %f989, %f983, %f986; mul.ftz.f32 %f990, %f984, %f984; mul.ftz.f32 %f991, %f984, %f985; mul.ftz.f32 %f992, %f984, %f986; mul.ftz.f32 %f993, %f985, %f985; mul.ftz.f32 %f994, %f985, %f986; mul.ftz.f32 %f995, %f986, %f986; sub.ftz.f32 %f997, %f846, %f993; sub.ftz.f32 %f998, %f997, %f995; sub.ftz.f32 %f999, %f991, %f989; add.ftz.f32 %f1000, %f988, %f992; add.ftz.f32 %f1001, %f989, %f991; sub.ftz.f32 %f1002, %f846, %f990; sub.ftz.f32 %f1003, %f1002, %f995; sub.ftz.f32 %f1004, %f994, %f987; sub.ftz.f32 %f1005, %f992, %f988; add.ftz.f32 %f1006, %f987, %f994; sub.ftz.f32 %f1007, %f1002, %f993; mul.ftz.f32 %f1008, %f999, %f957; fma.rn.ftz.f32 %f1009, %f998, %f934, %f1008; fma.rn.ftz.f32 %f1444, %f1000, %f980, %f1009; mul.ftz.f32 %f1010, %f999, %f958; fma.rn.ftz.f32 %f1011, %f998, %f955, %f1010; fma.rn.ftz.f32 %f1443, %f1000, %f981, %f1011; mul.ftz.f32 %f1012, %f999, %f979; fma.rn.ftz.f32 %f1013, %f998, %f956, %f1012; fma.rn.ftz.f32 %f1442, %f1000, %f982, %f1013; mul.ftz.f32 %f1014, %f1003, %f957; fma.rn.ftz.f32 %f1015, %f1001, %f934, %f1014; fma.rn.ftz.f32 %f1441, %f1004, %f980, %f1015; mul.ftz.f32 %f1016, %f1003, %f958; fma.rn.ftz.f32 %f1017, %f1001, %f955, %f1016; fma.rn.ftz.f32 %f1440, %f1004, %f981, %f1017; mul.ftz.f32 %f1018, %f1003, %f979; fma.rn.ftz.f32 %f1019, %f1001, %f956, %f1018; fma.rn.ftz.f32 %f1439, %f1004, %f982, %f1019; mul.ftz.f32 %f1020, %f1006, %f957; fma.rn.ftz.f32 %f1021, %f1005, %f934, %f1020; fma.rn.ftz.f32 %f1438, %f1007, %f980, %f1021; mul.ftz.f32 %f1022, %f1006, %f958; fma.rn.ftz.f32 %f1023, %f1005, %f955, %f1022; fma.rn.ftz.f32 %f1437, %f1007, %f981, %f1023; mul.ftz.f32 %f1024, %f1006, %f979; fma.rn.ftz.f32 %f1025, %f1005, %f956, %f1024; fma.rn.ftz.f32 %f1436, %f1007, %f982, %f1025; add.s64 %rd118, %rd87, 144; cvta.to.local.u64 %rd119, %rd118; st.local.v4.f32 [%rd119], {%f1444, %f1443, %f1442, %f201}; add.s64 %rd120, %rd87, 160; cvta.to.local.u64 %rd121, %rd120; st.local.v4.f32 [%rd121], {%f1441, %f1440, %f1439, %f202}; add.s64 %rd122, %rd87, 176; cvta.to.local.u64 %rd123, %rd122; st.local.v4.f32 [%rd123], {%f1438, %f1437, %f1436, %f203}; mul.ftz.f32 %f1026, %f1440, %f1436; mul.ftz.f32 %f213, %f1437, %f1439; sub.ftz.f32 %f1431, %f1026, %f213; mul.ftz.f32 %f215, %f1437, %f1442; mul.ftz.f32 %f1027, %f1443, %f1436; sub.ftz.f32 %f1430, %f215, %f1027; mul.ftz.f32 %f217, %f1443, %f1439; mul.ftz.f32 %f1028, %f1440, %f1442; sub.ftz.f32 %f1429, %f217, %f1028; mul.ftz.f32 %f1029, %f1441, %f1430; fma.rn.ftz.f32 %f1030, %f1444, %f1431, %f1029; fma.rn.ftz.f32 %f1435, %f1438, %f1429, %f1030; setp.neu.ftz.f32 %p103, %f1435, 0f00000000; mov.f32 %f1432, %f1436; mov.f32 %f1433, %f1440; mov.f32 %f1434, %f1444; @%p103 bra $L__BB210_69; add.ftz.f32 %f1434, %f1444, 0f322BCC77; add.ftz.f32 %f1432, %f1436, 0f322BCC77; add.ftz.f32 %f1433, %f1440, 0f322BCC77; mul.ftz.f32 %f1031, %f1433, %f1432; sub.ftz.f32 %f1431, %f1031, %f213; mul.ftz.f32 %f1032, %f1443, %f1432; sub.ftz.f32 %f1430, %f215, %f1032; mul.ftz.f32 %f1033, %f1433, %f1442; sub.ftz.f32 %f1429, %f217, %f1033; mul.ftz.f32 %f1034, %f1441, %f1430; fma.rn.ftz.f32 %f1035, %f1434, %f1431, %f1034; fma.rn.ftz.f32 %f1036, %f1438, %f1429, %f1035; setp.eq.ftz.f32 %p104, %f1036, 0f00000000; selp.f32 %f1435, 0f7F7FFFFF, %f1036, %p104; $L__BB210_69: mul.ftz.f32 %f1037, %f1441, %f1432; mul.ftz.f32 %f1038, %f1438, %f1439; sub.ftz.f32 %f1039, %f1038, %f1037; mul.ftz.f32 %f1040, %f1438, %f1442; mul.ftz.f32 %f1041, %f1432, %f1434; sub.ftz.f32 %f1042, %f1041, %f1040; mul.ftz.f32 %f1043, %f1439, %f1434; mul.ftz.f32 %f1044, %f1441, %f1442; sub.ftz.f32 %f1045, %f1044, %f1043; div.approx.ftz.f32 %f1452, %f1039, %f1435; div.approx.ftz.f32 %f1449, %f1042, %f1435; div.approx.ftz.f32 %f1446, %f1045, %f1435; mul.ftz.f32 %f1046, %f1438, %f1433; mul.ftz.f32 %f1047, %f1441, %f1437; sub.ftz.f32 %f1048, %f1047, %f1046; mul.ftz.f32 %f1049, %f1437, %f1434; mul.ftz.f32 %f1050, %f1438, %f1443; sub.ftz.f32 %f1051, %f1050, %f1049; mul.ftz.f32 %f1052, %f1441, %f1443; mul.ftz.f32 %f1053, %f1433, %f1434; sub.ftz.f32 %f1054, %f1053, %f1052; div.approx.ftz.f32 %f1451, %f1048, %f1435; div.approx.ftz.f32 %f1448, %f1051, %f1435; div.approx.ftz.f32 %f1445, %f1054, %f1435; div.approx.ftz.f32 %f1453, %f1431, %f1435; div.approx.ftz.f32 %f1450, %f1430, %f1435; mul.ftz.f32 %f1055, %f202, %f1450; fma.rn.ftz.f32 %f1056, %f201, %f1453, %f1055; div.approx.ftz.f32 %f1447, %f1429, %f1435; fma.rn.ftz.f32 %f1057, %f203, %f1447, %f1056; mul.ftz.f32 %f1058, %f202, %f1449; fma.rn.ftz.f32 %f1059, %f201, %f1452, %f1058; fma.rn.ftz.f32 %f1060, %f203, %f1446, %f1059; mul.ftz.f32 %f1061, %f202, %f1448; fma.rn.ftz.f32 %f1062, %f201, %f1451, %f1061; fma.rn.ftz.f32 %f1063, %f203, %f1445, %f1062; neg.ftz.f32 %f1064, %f1057; add.s64 %rd125, %rd87, 192; cvta.to.local.u64 %rd126, %rd125; st.local.v4.f32 [%rd126], {%f1453, %f1450, %f1447, %f1064}; neg.ftz.f32 %f1065, %f1060; add.s64 %rd127, %rd87, 208; cvta.to.local.u64 %rd128, %rd127; st.local.v4.f32 [%rd128], {%f1452, %f1449, %f1446, %f1065}; neg.ftz.f32 %f1066, %f1063; add.s64 %rd129, %rd87, 224; cvta.to.local.u64 %rd130, %rd129; st.local.v4.f32 [%rd130], {%f1451, %f1448, %f1445, %f1066}; $L__BB210_70: @%p7 bra $L__BB210_86; setp.lt.s32 %p106, %r6, 0; ld.const.u64 %rd10, [kernel_params+2232]; @%p106 bra $L__BB210_73; and.b32 %r323, %r17, 4; mov.f32 %f1469, %f1419; mov.f32 %f1470, %f1420; mov.f32 %f1471, %f1421; bra.uni $L__BB210_79; $L__BB210_86: add.s64 %rd174, %rd87, 116; cvta.to.local.u64 %rd175, %rd174; mov.u32 %r197, 0; st.local.u32 [%rd175], %r197; add.s64 %rd176, %rd87, 120; cvta.to.local.u64 %rd177, %rd176; st.local.u32 [%rd177], %r197; add.s64 %rd178, %rd87, 124; cvta.to.local.u64 %rd179, %rd178; st.local.u32 [%rd179], %r197; add.s64 %rd180, %rd87, 136; cvta.to.local.u64 %rd181, %rd180; add.s64 %rd182, %rd87, 128; cvta.to.local.u64 %rd183, %rd182; mov.f32 %f1463, 0f00000000; st.local.v2.f32 [%rd183], {%f1463, %f1463}; st.local.u32 [%rd181], %r197; mov.f32 %f1464, %f1463; mov.f32 %f1465, %f1463; mov.f32 %f1466, %f1463; mov.f32 %f1467, %f1463; mov.f32 %f1468, %f1463; mov.f32 %f1469, %f1419; mov.f32 %f1470, %f1420; mov.f32 %f1471, %f1421; bra.uni $L__BB210_90; $L__BB210_73: cvta.to.global.u64 %rd131, %rd10; mul.wide.s32 %rd132, %r4, 16; add.s64 %rd133, %rd131, %rd132; ld.global.v4.u32 {%r185, %r186, %r187, %r188}, [%rd133]; ld.const.u64 %rd134, [kernel_params+2224]; cvta.to.global.u64 %rd135, %rd134; mul.wide.u32 %rd136, %r185, 12; add.s64 %rd137, %rd135, %rd136; mul.wide.u32 %rd138, %r186, 12; add.s64 %rd139, %rd135, %rd138; mul.wide.u32 %rd140, %r187, 12; add.s64 %rd141, %rd135, %rd140; mov.f32 %f1067, 0f3F800000; sub.ftz.f32 %f1068, %f1067, %f1418; sub.ftz.f32 %f1069, %f1068, %f1417; ld.global.f32 %f1070, [%rd137]; ld.global.f32 %f1071, [%rd137+4]; ld.global.f32 %f1072, [%rd137+8]; ld.global.f32 %f1073, [%rd139]; mul.ftz.f32 %f1074, %f1418, %f1073; ld.global.f32 %f1075, [%rd139+4]; mul.ftz.f32 %f1076, %f1418, %f1075; ld.global.f32 %f1077, [%rd139+8]; mul.ftz.f32 %f1078, %f1418, %f1077; fma.rn.ftz.f32 %f1079, %f1069, %f1070, %f1074; fma.rn.ftz.f32 %f1080, %f1069, %f1071, %f1076; fma.rn.ftz.f32 %f1081, %f1069, %f1072, %f1078; ld.global.f32 %f1082, [%rd141]; ld.global.f32 %f1083, [%rd141+4]; ld.global.f32 %f1084, [%rd141+8]; fma.rn.ftz.f32 %f1454, %f1417, %f1082, %f1079; fma.rn.ftz.f32 %f1455, %f1417, %f1083, %f1080; fma.rn.ftz.f32 %f1456, %f1417, %f1084, %f1081; mul.ftz.f32 %f1085, %f1455, %f1455; fma.rn.ftz.f32 %f1086, %f1454, %f1454, %f1085; fma.rn.ftz.f32 %f1087, %f1456, %f1456, %f1086; sqrt.approx.ftz.f32 %f264, %f1087; setp.neu.ftz.f32 %p107, %f264, 0f00000000; @%p107 bra $L__BB210_74; bra.uni $L__BB210_75; $L__BB210_74: rcp.approx.ftz.f32 %f1088, %f264; mul.ftz.f32 %f1456, %f1456, %f1088; mul.ftz.f32 %f1455, %f1455, %f1088; mul.ftz.f32 %f1454, %f1454, %f1088; $L__BB210_75: add.u64 %rd515, %SP, 1024; or.b64 %rd514, %rd515, 12; add.s64 %rd142, %rd514, 8; setp.eq.ftz.f32 %p108, %f1454, 0f00000000; setp.eq.ftz.f32 %p109, %f1455, 0f00000000; and.pred %p110, %p108, %p109; setp.eq.ftz.f32 %p111, %f1456, 0f00000000; and.pred %p112, %p111, %p110; selp.f32 %f1469, %f1419, %f1456, %p112; selp.f32 %f1470, %f1420, %f1455, %p112; selp.f32 %f1471, %f1421, %f1454, %p112; cvta.to.local.u64 %rd11, %rd514; st.local.f32 [%rd11], %f1471; st.local.f32 [%rd8], %f1470; cvta.to.local.u64 %rd13, %rd142; st.local.f32 [%rd13], %f1469; and.b32 %r323, %r17, 4; setp.ne.s32 %p113, %r323, 0; @%p113 bra $L__BB210_79; mov.u32 %r323, 0; @%p99 bra $L__BB210_78; bra.uni $L__BB210_77; $L__BB210_78: ld.const.u64 %rd144, [kernel_params+2160]; cvta.to.global.u64 %rd145, %rd144; mul.lo.s64 %rd146, %rd2, 208; add.s64 %rd147, %rd145, %rd146; ld.global.v4.f32 {%f1102, %f1103, %f1104, %f1105}, [%rd147+48]; ld.global.v4.f32 {%f1109, %f1110, %f1111, %f1112}, [%rd147+64]; ld.global.v4.f32 {%f1116, %f1117, %f1118, %f1119}, [%rd147+80]; mul.ftz.f32 %f1123, %f1109, %f1470; fma.rn.ftz.f32 %f1124, %f1102, %f1471, %f1123; fma.rn.ftz.f32 %f1125, %f1116, %f1469, %f1124; mul.ftz.f32 %f1126, %f1110, %f1470; fma.rn.ftz.f32 %f1127, %f1103, %f1471, %f1126; fma.rn.ftz.f32 %f1128, %f1117, %f1469, %f1127; mul.ftz.f32 %f1129, %f1111, %f1470; fma.rn.ftz.f32 %f1130, %f1104, %f1471, %f1129; fma.rn.ftz.f32 %f1131, %f1118, %f1469, %f1130; mul.ftz.f32 %f1132, %f1128, %f1128; fma.rn.ftz.f32 %f1133, %f1125, %f1125, %f1132; fma.rn.ftz.f32 %f1134, %f1131, %f1131, %f1133; sqrt.approx.ftz.f32 %f1135, %f1134; div.approx.ftz.f32 %f1471, %f1125, %f1135; div.approx.ftz.f32 %f1470, %f1128, %f1135; div.approx.ftz.f32 %f1469, %f1131, %f1135; st.local.f32 [%rd11], %f1471; st.local.f32 [%rd8], %f1470; st.local.f32 [%rd13], %f1469; bra.uni $L__BB210_79; $L__BB210_77: mul.ftz.f32 %f1089, %f1470, %f1452; fma.rn.ftz.f32 %f1090, %f1471, %f1453, %f1089; fma.rn.ftz.f32 %f1091, %f1469, %f1451, %f1090; mul.ftz.f32 %f1092, %f1470, %f1449; fma.rn.ftz.f32 %f1093, %f1471, %f1450, %f1092; fma.rn.ftz.f32 %f1094, %f1469, %f1448, %f1093; mul.ftz.f32 %f1095, %f1470, %f1446; fma.rn.ftz.f32 %f1096, %f1471, %f1447, %f1095; fma.rn.ftz.f32 %f1097, %f1469, %f1445, %f1096; mul.ftz.f32 %f1098, %f1094, %f1094; fma.rn.ftz.f32 %f1099, %f1091, %f1091, %f1098; fma.rn.ftz.f32 %f1100, %f1097, %f1097, %f1099; sqrt.approx.ftz.f32 %f1101, %f1100; div.approx.ftz.f32 %f1471, %f1091, %f1101; div.approx.ftz.f32 %f1470, %f1094, %f1101; div.approx.ftz.f32 %f1469, %f1097, %f1101; st.local.f32 [%rd11], %f1471; st.local.f32 [%rd8], %f1470; st.local.f32 [%rd13], %f1469; $L__BB210_79: cvta.to.global.u64 %rd148, %rd10; mul.wide.s32 %rd149, %r4, 16; add.s64 %rd150, %rd148, %rd149; ld.const.u64 %rd151, [kernel_params+2256]; cvta.to.global.u64 %rd152, %rd151; ld.global.u32 %r194, [%rd150+12]; mul.wide.u32 %rd153, %r194, 12; add.s64 %rd154, %rd152, %rd153; add.s32 %r195, %r194, 1; mul.wide.u32 %rd155, %r195, 12; add.s64 %rd156, %rd152, %rd155; add.s32 %r196, %r194, 2; mul.wide.u32 %rd157, %r196, 12; add.s64 %rd158, %rd152, %rd157; ld.global.f32 %f1136, [%rd158]; ld.global.f32 %f1137, [%rd158+4]; ld.global.f32 %f1138, [%rd158+8]; ld.global.f32 %f1139, [%rd156]; ld.global.f32 %f1140, [%rd154]; sub.ftz.f32 %f1468, %f1139, %f1140; ld.global.f32 %f1141, [%rd156+4]; ld.global.f32 %f1142, [%rd154+4]; sub.ftz.f32 %f1467, %f1141, %f1142; ld.global.f32 %f1143, [%rd156+8]; ld.global.f32 %f1144, [%rd154+8]; sub.ftz.f32 %f1466, %f1143, %f1144; add.s64 %rd160, %rd87, 116; cvta.to.local.u64 %rd14, %rd160; st.local.f32 [%rd14], %f1468; add.s64 %rd161, %rd87, 120; cvta.to.local.u64 %rd15, %rd161; st.local.f32 [%rd15], %f1467; add.s64 %rd162, %rd87, 124; cvta.to.local.u64 %rd16, %rd162; st.local.f32 [%rd16], %f1466; sub.ftz.f32 %f1465, %f1136, %f1140; sub.ftz.f32 %f1464, %f1137, %f1142; sub.ftz.f32 %f1463, %f1138, %f1144; add.s64 %rd163, %rd87, 128; cvta.to.local.u64 %rd17, %rd163; st.local.v2.f32 [%rd17], {%f1465, %f1464}; add.s64 %rd164, %rd87, 136; cvta.to.local.u64 %rd18, %rd164; st.local.f32 [%rd18], %f1463; setp.ne.s32 %p115, %r323, 0; @%p115 bra $L__BB210_90; @%p99 bra $L__BB210_82; bra.uni $L__BB210_81; $L__BB210_82: ld.const.u64 %rd165, [kernel_params+2160]; cvta.to.global.u64 %rd166, %rd165; mul.lo.s64 %rd167, %rd2, 208; add.s64 %rd168, %rd166, %rd167; ld.global.v4.f32 {%f1151, %f1152, %f1153, %f1154}, [%rd168]; ld.global.v4.f32 {%f1158, %f1159, %f1160, %f1161}, [%rd168+16]; ld.global.v4.f32 {%f1165, %f1166, %f1167, %f1168}, [%rd168+32]; mul.ftz.f32 %f1172, %f1152, %f1467; fma.rn.ftz.f32 %f1173, %f1468, %f1151, %f1172; fma.rn.ftz.f32 %f1462, %f1153, %f1466, %f1173; mul.ftz.f32 %f1174, %f1159, %f1467; fma.rn.ftz.f32 %f1175, %f1468, %f1158, %f1174; fma.rn.ftz.f32 %f1461, %f1160, %f1466, %f1175; mul.ftz.f32 %f1176, %f1166, %f1467; fma.rn.ftz.f32 %f1177, %f1468, %f1165, %f1176; fma.rn.ftz.f32 %f1466, %f1167, %f1466, %f1177; bra.uni $L__BB210_83; $L__BB210_81: mul.ftz.f32 %f1145, %f1467, %f1443; fma.rn.ftz.f32 %f1146, %f1468, %f1444, %f1145; fma.rn.ftz.f32 %f1462, %f1466, %f1442, %f1146; mul.ftz.f32 %f1147, %f1467, %f1440; fma.rn.ftz.f32 %f1148, %f1468, %f1441, %f1147; fma.rn.ftz.f32 %f1461, %f1466, %f1439, %f1148; mul.ftz.f32 %f1149, %f1467, %f1437; fma.rn.ftz.f32 %f1150, %f1468, %f1438, %f1149; fma.rn.ftz.f32 %f1466, %f1466, %f1436, %f1150; $L__BB210_83: st.local.f32 [%rd14], %f1462; st.local.f32 [%rd15], %f1461; st.local.f32 [%rd16], %f1466; @%p99 bra $L__BB210_85; bra.uni $L__BB210_84; $L__BB210_85: ld.const.u64 %rd169, [kernel_params+2160]; cvta.to.global.u64 %rd170, %rd169; mul.lo.s64 %rd171, %rd2, 208; add.s64 %rd172, %rd170, %rd171; ld.global.v4.f32 {%f1184, %f1185, %f1186, %f1187}, [%rd172]; ld.global.v4.f32 {%f1191, %f1192, %f1193, %f1194}, [%rd172+16]; ld.global.v4.f32 {%f1198, %f1199, %f1200, %f1201}, [%rd172+32]; mul.ftz.f32 %f1205, %f1185, %f1464; fma.rn.ftz.f32 %f1206, %f1465, %f1184, %f1205; fma.rn.ftz.f32 %f304, %f1186, %f1463, %f1206; mul.ftz.f32 %f1207, %f1192, %f1464; fma.rn.ftz.f32 %f1208, %f1465, %f1191, %f1207; fma.rn.ftz.f32 %f305, %f1193, %f1463, %f1208; mul.ftz.f32 %f1209, %f1199, %f1464; fma.rn.ftz.f32 %f1210, %f1465, %f1198, %f1209; fma.rn.ftz.f32 %f1463, %f1200, %f1463, %f1210; st.local.v2.f32 [%rd17], {%f304, %f305}; st.local.f32 [%rd18], %f1463; mov.f32 %f1464, %f305; mov.f32 %f1465, %f304; mov.f32 %f1467, %f1461; mov.f32 %f1468, %f1462; bra.uni $L__BB210_90; $L__BB210_84: mul.ftz.f32 %f1178, %f1464, %f1443; fma.rn.ftz.f32 %f1179, %f1465, %f1444, %f1178; fma.rn.ftz.f32 %f301, %f1463, %f1442, %f1179; mul.ftz.f32 %f1180, %f1464, %f1440; fma.rn.ftz.f32 %f1181, %f1465, %f1441, %f1180; fma.rn.ftz.f32 %f302, %f1463, %f1439, %f1181; mul.ftz.f32 %f1182, %f1464, %f1437; fma.rn.ftz.f32 %f1183, %f1465, %f1438, %f1182; fma.rn.ftz.f32 %f1463, %f1463, %f1436, %f1183; st.local.v2.f32 [%rd17], {%f301, %f302}; st.local.f32 [%rd18], %f1463; mov.f32 %f1464, %f302; mov.f32 %f1465, %f301; mov.f32 %f1467, %f1461; mov.f32 %f1468, %f1462; $L__BB210_90: @%p7 bra $L__BB210_93; mul.ftz.f32 %f1223, %f1420, %f176; fma.rn.ftz.f32 %f1224, %f1421, %f175, %f1223; fma.rn.ftz.f32 %f1225, %f1419, %f177, %f1224; setp.geu.ftz.f32 %p120, %f1225, 0f00000000; @%p120 bra $L__BB210_93; add.u64 %rd500, %SP, 1024; or.b64 %rd499, %rd500, 12; add.s64 %rd498, %rd499, 20; cvta.to.local.u64 %rd497, %rd498; add.s64 %rd198, %rd499, 40; cvta.to.local.u64 %rd199, %rd198; or.b32 %r199, %r15, 1; st.local.u32 [%rd199], %r199; neg.ftz.f32 %f1226, %f1419; st.local.f32 [%rd497], %f1226; cvta.to.local.u64 %rd201, %rd499; neg.ftz.f32 %f1227, %f1471; st.local.f32 [%rd201], %f1227; neg.ftz.f32 %f1228, %f1420; neg.ftz.f32 %f1229, %f1421; neg.ftz.f32 %f1230, %f1469; neg.ftz.f32 %f1231, %f1470; st.local.v4.f32 [%rd8], {%f1231, %f1230, %f1229, %f1228}; add.s64 %rd203, %rd500, 116; add.s64 %rd204, %rd500, 120; add.s64 %rd205, %rd500, 124; cvta.to.local.u64 %rd206, %rd203; neg.ftz.f32 %f1232, %f1468; st.local.f32 [%rd206], %f1232; cvta.to.local.u64 %rd207, %rd204; neg.ftz.f32 %f1233, %f1467; st.local.f32 [%rd207], %f1233; cvta.to.local.u64 %rd208, %rd205; neg.ftz.f32 %f1234, %f1466; st.local.f32 [%rd208], %f1234; add.s64 %rd209, %rd500, 136; cvta.to.local.u64 %rd210, %rd209; neg.ftz.f32 %f1235, %f1464; neg.ftz.f32 %f1236, %f1465; add.s64 %rd211, %rd500, 128; cvta.to.local.u64 %rd212, %rd211; st.local.v2.f32 [%rd212], {%f1236, %f1235}; neg.ftz.f32 %f1237, %f1463; st.local.f32 [%rd210], %f1237; $L__BB210_93: and.b32 %r325, %r6, 4194303; add.s64 %rd214, %rd87, 92; cvta.to.local.u64 %rd215, %rd214; mov.u32 %r200, 0; st.local.u32 [%rd215], %r200; add.s64 %rd216, %rd87, 96; cvta.to.local.u64 %rd217, %rd216; mov.f32 %f1238, 0f00000000; st.local.v4.f32 [%rd217], {%f1238, %f1238, %f1238, %f1238}; add.s64 %rd218, %rd87, 112; cvta.to.local.u64 %rd219, %rd218; st.local.u32 [%rd219], %r200; $L__BB210_95: add.s64 %rd252, %rd87, 260; cvta.to.local.u64 %rd253, %rd252; mov.u32 %r206, 0; st.local.u32 [%rd253], %r206; add.s64 %rd254, %rd87, 264; cvta.to.local.u64 %rd255, %rd254; st.local.u32 [%rd255], %r206; ld.const.u64 %rd256, [kernel_params+2400]; cvta.to.global.u64 %rd19, %rd256; add.s64 %rd257, %rd87, 268; cvta.to.local.u64 %rd23, %rd257; add.s64 %rd258, %rd87, 272; cvta.to.local.u64 %rd24, %rd258; add.s64 %rd259, %rd87, 276; cvta.to.local.u64 %rd25, %rd259; add.u64 %rd260, %SP, 0; add.u64 %rd26, %SPL, 0; bra.uni $L__BB210_96; $L__BB210_248: { // callseq 1122, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1122 $L__BB210_96: mov.u32 %r24, %r325; mul.wide.s32 %rd261, %r24, 16; add.s64 %rd27, %rd19, %rd261; ld.global.v4.u32 {%r208, %r209, %r210, %r211}, [%rd27]; mul.wide.u32 %rd262, %r209, 4; add.s64 %rd28, %rd26, %rd262; mul.wide.u32 %rd263, %r210, 4; add.s64 %rd29, %rd26, %rd263; add.s32 %r325, %r24, 1; setp.gt.s32 %p121, %r208, 36; @%p121 bra $L__BB210_145; setp.gt.s32 %p203, %r208, 25; @%p203 bra $L__BB210_141; bra.uni $L__BB210_98; $L__BB210_141: add.s32 %r213, %r208, -27; setp.lt.u32 %p204, %r213, 9; @%p204 bra $L__BB210_96; setp.eq.s32 %p205, %r208, 26; @%p205 bra $L__BB210_287; setp.eq.s32 %p206, %r208, 36; @%p206 bra $L__BB210_144; bra.uni $L__BB210_301; $L__BB210_144: mov.u32 %r255, 36; { // callseq 1175, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r255; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1175 bra.uni $L__BB210_96; $L__BB210_145: setp.gt.s32 %p122, %r208, 68; @%p122 bra $L__BB210_200; bra.uni $L__BB210_146; $L__BB210_200: setp.gt.s32 %p123, %r208, 88; @%p123 bra $L__BB210_236; bra.uni $L__BB210_201; $L__BB210_236: setp.gt.s32 %p124, %r208, 93; @%p124 bra $L__BB210_241; bra.uni $L__BB210_237; $L__BB210_241: setp.gt.s32 %p125, %r208, 95; @%p125 bra $L__BB210_245; bra.uni $L__BB210_242; $L__BB210_245: setp.eq.s32 %p126, %r208, 98; @%p126 bra $L__BB210_300; setp.eq.s32 %p127, %r208, 97; @%p127 bra $L__BB210_250; setp.ne.s32 %p128, %r208, 96; @%p128 bra $L__BB210_301; bra.uni $L__BB210_248; $L__BB210_146: setp.gt.s32 %p161, %r208, 48; @%p161 bra $L__BB210_165; bra.uni $L__BB210_147; $L__BB210_165: setp.gt.s32 %p162, %r208, 58; @%p162 bra $L__BB210_183; setp.gt.s32 %p176, %r208, 53; @%p176 bra $L__BB210_175; setp.gt.s32 %p183, %r208, 50; @%p183 bra $L__BB210_171; bra.uni $L__BB210_168; $L__BB210_171: setp.eq.s32 %p184, %r208, 51; @%p184 bra $L__BB210_279; setp.eq.s32 %p185, %r208, 52; @%p185 bra $L__BB210_278; setp.eq.s32 %p186, %r208, 53; @%p186 bra $L__BB210_174; bra.uni $L__BB210_301; $L__BB210_174: { // callseq 1160, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1160 bra.uni $L__BB210_96; $L__BB210_98: setp.gt.s32 %p207, %r208, 17; @%p207 bra $L__BB210_132; bra.uni $L__BB210_99; $L__BB210_132: setp.lt.s32 %p208, %r208, 21; @%p208 bra $L__BB210_96; setp.gt.s32 %p209, %r208, 22; @%p209 bra $L__BB210_137; bra.uni $L__BB210_134; $L__BB210_137: setp.eq.s32 %p210, %r208, 23; @%p210 bra $L__BB210_289; setp.eq.s32 %p211, %r208, 24; @%p211 bra $L__BB210_288; setp.eq.s32 %p212, %r208, 25; @%p212 bra $L__BB210_140; bra.uni $L__BB210_301; $L__BB210_140: add.s32 %r257, %r24, 1; { // callseq 1177, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r257; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1177 bra.uni $L__BB210_96; $L__BB210_201: setp.gt.s32 %p134, %r208, 78; @%p134 bra $L__BB210_219; setp.gt.s32 %p148, %r208, 73; @%p148 bra $L__BB210_211; setp.gt.s32 %p155, %r208, 70; @%p155 bra $L__BB210_207; bra.uni $L__BB210_204; $L__BB210_207: setp.eq.s32 %p156, %r208, 71; @%p156 bra $L__BB210_267; setp.eq.s32 %p157, %r208, 72; @%p157 bra $L__BB210_266; setp.eq.s32 %p158, %r208, 73; @%p158 bra $L__BB210_210; bra.uni $L__BB210_301; $L__BB210_210: add.s32 %r226, %r24, 1; { // callseq 1140, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r226; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1140 bra.uni $L__BB210_96; $L__BB210_147: setp.gt.s32 %p189, %r208, 41; @%p189 bra $L__BB210_153; setp.gt.s32 %p199, %r208, 39; @%p199 bra $L__BB210_96; setp.eq.s32 %p200, %r208, 37; @%p200 bra $L__BB210_286; setp.eq.s32 %p201, %r208, 38; @%p201 bra $L__BB210_285; setp.eq.s32 %p202, %r208, 39; @%p202 bra $L__BB210_152; bra.uni $L__BB210_301; $L__BB210_152: mov.u32 %r253, 39; { // callseq 1172, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r253, %r209, %r210, %r211}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1172 bra.uni $L__BB210_96; $L__BB210_99: setp.gt.s32 %p215, %r208, 8; @%p215 bra $L__BB210_115; bra.uni $L__BB210_100; $L__BB210_115: setp.gt.s32 %p216, %r208, 12; @%p216 bra $L__BB210_124; bra.uni $L__BB210_116; $L__BB210_124: setp.gt.s32 %p217, %r208, 14; @%p217 bra $L__BB210_128; bra.uni $L__BB210_125; $L__BB210_128: setp.eq.s32 %p218, %r208, 15; @%p218 bra $L__BB210_292; setp.eq.s32 %p219, %r208, 16; @%p219 bra $L__BB210_291; setp.eq.s32 %p220, %r208, 17; @%p220 bra $L__BB210_131; bra.uni $L__BB210_301; $L__BB210_131: { // callseq 1182, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; .param .b32 param4; st.param.b32 [param4+0], %r211; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1182 bra.uni $L__BB210_96; $L__BB210_183: setp.gt.s32 %p163, %r208, 63; @%p163 bra $L__BB210_192; setp.gt.s32 %p170, %r208, 60; @%p170 bra $L__BB210_188; bra.uni $L__BB210_185; $L__BB210_188: setp.eq.s32 %p171, %r208, 61; @%p171 bra $L__BB210_273; setp.eq.s32 %p172, %r208, 62; @%p172 bra $L__BB210_272; setp.eq.s32 %p173, %r208, 63; @%p173 bra $L__BB210_191; bra.uni $L__BB210_301; $L__BB210_191: mov.u32 %r236, 63; { // callseq 1150, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r236; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1150 bra.uni $L__BB210_96; $L__BB210_219: setp.gt.s32 %p135, %r208, 83; @%p135 bra $L__BB210_228; setp.gt.s32 %p142, %r208, 80; @%p142 bra $L__BB210_224; bra.uni $L__BB210_221; $L__BB210_224: setp.eq.s32 %p143, %r208, 81; @%p143 bra $L__BB210_255; setp.eq.s32 %p144, %r208, 82; @%p144 bra $L__BB210_254; setp.eq.s32 %p145, %r208, 83; @%p145 bra $L__BB210_227; bra.uni $L__BB210_301; $L__BB210_227: { // callseq 1132, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1132 bra.uni $L__BB210_96; $L__BB210_100: setp.gt.s32 %p228, %r208, 4; @%p228 bra $L__BB210_108; setp.gt.s32 %p234, %r208, 2; @%p234 bra $L__BB210_105; setp.eq.s32 %p237, %r208, 1; mov.u32 %r325, %r209; @%p237 bra $L__BB210_96; setp.eq.s32 %p238, %r208, 2; @%p238 bra $L__BB210_104; bra.uni $L__BB210_301; $L__BB210_104: mov.u32 %r281, 2; add.s32 %r282, %r24, 1; { // callseq 1191, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r281; st.param.b32 [param0+4], %r209; st.param.b32 [param0+8], %r210; st.param.b32 [param0+12], %r211; .param .b32 param1; st.param.b32 [param1+0], %r282; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1191 bra.uni $L__BB210_96; $L__BB210_153: setp.gt.s32 %p190, %r208, 44; @%p190 bra $L__BB210_158; bra.uni $L__BB210_154; $L__BB210_158: setp.gt.s32 %p191, %r208, 46; @%p191 bra $L__BB210_162; setp.eq.s32 %p194, %r208, 45; @%p194 bra $L__BB210_282; setp.eq.s32 %p195, %r208, 46; @%p195 bra $L__BB210_161; bra.uni $L__BB210_301; $L__BB210_161: { // callseq 1167, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1167 bra.uni $L__BB210_96; $L__BB210_242: setp.eq.s32 %p129, %r208, 94; @%p129 bra $L__BB210_249; setp.eq.s32 %p130, %r208, 95; @%p130 bra $L__BB210_244; bra.uni $L__BB210_301; $L__BB210_244: { // callseq 1123, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1123 bra.uni $L__BB210_96; $L__BB210_237: add.s32 %r212, %r208, -92; setp.lt.u32 %p131, %r212, 2; @%p131 bra $L__BB210_96; setp.eq.s32 %p132, %r208, 90; @%p132 bra $L__BB210_96; setp.eq.s32 %p133, %r208, 89; @%p133 bra $L__BB210_240; bra.uni $L__BB210_301; $L__BB210_240: mov.u32 %r216, 89; { // callseq 1126, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r216; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1126 bra.uni $L__BB210_96; $L__BB210_175: setp.gt.s32 %p177, %r208, 55; @%p177 bra $L__BB210_179; bra.uni $L__BB210_176; $L__BB210_179: setp.eq.s32 %p178, %r208, 56; @%p178 bra $L__BB210_276; setp.eq.s32 %p179, %r208, 57; @%p179 bra $L__BB210_275; setp.eq.s32 %p180, %r208, 58; @%p180 bra $L__BB210_182; bra.uni $L__BB210_301; $L__BB210_182: mov.u32 %r243, 58; { // callseq 1155, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r243; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1155 bra.uni $L__BB210_96; $L__BB210_211: setp.gt.s32 %p149, %r208, 75; @%p149 bra $L__BB210_215; bra.uni $L__BB210_212; $L__BB210_215: setp.eq.s32 %p150, %r208, 76; @%p150 bra $L__BB210_259; setp.eq.s32 %p151, %r208, 77; @%p151 bra $L__BB210_257; setp.eq.s32 %p152, %r208, 78; @%p152 bra $L__BB210_218; bra.uni $L__BB210_301; $L__BB210_218: add.s32 %r224, %r24, 1; { // callseq 1137, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r224; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1137 bra.uni $L__BB210_96; $L__BB210_192: setp.gt.s32 %p164, %r208, 65; @%p164 bra $L__BB210_196; bra.uni $L__BB210_193; $L__BB210_196: setp.eq.s32 %p165, %r208, 66; @%p165 bra $L__BB210_270; setp.eq.s32 %p166, %r208, 67; @%p166 bra $L__BB210_269; setp.eq.s32 %p167, %r208, 68; @%p167 bra $L__BB210_199; bra.uni $L__BB210_301; $L__BB210_199: mov.u32 %r231, 68; { // callseq 1145, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r231; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1145 bra.uni $L__BB210_96; $L__BB210_228: setp.gt.s32 %p136, %r208, 85; @%p136 bra $L__BB210_232; bra.uni $L__BB210_229; $L__BB210_232: setp.eq.s32 %p137, %r208, 86; @%p137 bra $L__BB210_252; setp.eq.s32 %p138, %r208, 87; @%p138 bra $L__BB210_251; setp.eq.s32 %p139, %r208, 88; @%p139 bra $L__BB210_235; bra.uni $L__BB210_301; $L__BB210_235: mov.u32 %r217, 88; cvta.to.local.u64 %rd270, %rd87; ld.local.f32 %f1240, [%rd270+12]; ld.local.v2.f32 {%f1241, %f1242}, [%rd270+16]; { // callseq 1127, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1240; .param .b32 param1; st.param.f32 [param1+0], %f1241; .param .b32 param2; st.param.f32 [param2+0], %f1242; .param .b64 param3; st.param.b64 [param3+0], %rd260; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r217; st.param.b32 [param4+4], %r209; st.param.b32 [param4+8], %r210; st.param.b32 [param4+12], %r211; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1127 bra.uni $L__BB210_96; $L__BB210_116: setp.gt.s32 %p223, %r208, 10; @%p223 bra $L__BB210_121; setp.eq.s32 %p226, %r208, 9; @%p226 bra $L__BB210_295; setp.eq.s32 %p227, %r208, 10; @%p227 bra $L__BB210_119; bra.uni $L__BB210_301; $L__BB210_119: ld.local.f32 %f1256, [%rd29]; setp.ltu.ftz.f32 %p243, %f1256, 0f3F800000; @%p243 bra $L__BB210_96; add.s32 %r275, %r209, %r24; add.s32 %r325, %r275, 1; bra.uni $L__BB210_96; $L__BB210_108: setp.gt.s32 %p229, %r208, 6; @%p229 bra $L__BB210_112; setp.eq.s32 %p232, %r208, 5; @%p232 bra $L__BB210_298; setp.eq.s32 %p233, %r208, 6; @%p233 bra $L__BB210_111; bra.uni $L__BB210_301; $L__BB210_111: ld.local.f32 %f1258, [%rd28]; ld.local.f32 %f1259, [%rd28+4]; ld.local.f32 %f1260, [%rd28+8]; st.local.f32 [%rd23], %f1258; st.local.f32 [%rd24], %f1259; st.local.f32 [%rd25], %f1260; bra.uni $L__BB210_96; $L__BB210_154: setp.eq.s32 %p196, %r208, 42; @%p196 bra $L__BB210_284; setp.eq.s32 %p197, %r208, 43; @%p197 bra $L__BB210_283; setp.eq.s32 %p198, %r208, 44; @%p198 bra $L__BB210_157; bra.uni $L__BB210_301; $L__BB210_157: mov.u32 %r250, 44; add.s32 %r251, %r24, 1; { // callseq 1169, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r250; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r251; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1169 bra.uni $L__BB210_96; $L__BB210_300: { // callseq 1192, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1192 bra.uni $L__BB210_96; $L__BB210_250: { // callseq 1125, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1125 bra.uni $L__BB210_96; $L__BB210_249: mov.u32 %r214, 94; add.s32 %r215, %r24, 1; { // callseq 1124, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r214; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r215; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1124 bra.uni $L__BB210_96; $L__BB210_134: setp.eq.s32 %p213, %r208, 21; @%p213 bra $L__BB210_290; setp.eq.s32 %p214, %r208, 22; @%p214 bra $L__BB210_136; bra.uni $L__BB210_301; $L__BB210_136: mov.u32 %r261, 22; add.s32 %r262, %r24, 1; { // callseq 1180, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r261; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r262; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1180 bra.uni $L__BB210_96; $L__BB210_125: setp.eq.s32 %p221, %r208, 13; @%p221 bra $L__BB210_293; setp.eq.s32 %p222, %r208, 14; @%p222 bra $L__BB210_127; bra.uni $L__BB210_301; $L__BB210_127: st.local.u32 [%rd29], %r209; bra.uni $L__BB210_96; $L__BB210_105: setp.eq.s32 %p235, %r208, 3; @%p235 bra $L__BB210_299; setp.eq.s32 %p236, %r208, 4; @%p236 bra $L__BB210_107; bra.uni $L__BB210_301; $L__BB210_107: mov.u32 %r279, 4; { // callseq 1189, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r279, %r209, %r210, %r211}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1189 bra.uni $L__BB210_96; $L__BB210_121: setp.eq.s32 %p224, %r208, 11; @%p224 bra $L__BB210_294; setp.eq.s32 %p225, %r208, 12; @%p225 bra $L__BB210_123; bra.uni $L__BB210_301; $L__BB210_123: { // callseq 1185, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1185 bra.uni $L__BB210_96; $L__BB210_112: setp.eq.s32 %p230, %r208, 7; @%p230 bra $L__BB210_297; setp.eq.s32 %p231, %r208, 8; @%p231 bra $L__BB210_114; bra.uni $L__BB210_301; $L__BB210_114: mov.u32 %r277, 8; { // callseq 1187, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r277; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1187 bra.uni $L__BB210_96; $L__BB210_168: setp.eq.s32 %p187, %r208, 49; @%p187 bra $L__BB210_280; setp.eq.s32 %p188, %r208, 50; @%p188 bra $L__BB210_170; bra.uni $L__BB210_301; $L__BB210_170: { // callseq 1163, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1163 bra.uni $L__BB210_96; $L__BB210_204: setp.eq.s32 %p159, %r208, 69; @%p159 bra $L__BB210_268; setp.eq.s32 %p160, %r208, 70; @%p160 bra $L__BB210_206; bra.uni $L__BB210_301; $L__BB210_206: mov.u32 %r228, 70; { // callseq 1143, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r228; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1143 bra.uni $L__BB210_96; $L__BB210_185: setp.eq.s32 %p174, %r208, 59; @%p174 bra $L__BB210_274; setp.eq.s32 %p175, %r208, 60; @%p175 bra $L__BB210_187; bra.uni $L__BB210_301; $L__BB210_187: add.s32 %r241, %r24, 1; { // callseq 1153, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r241; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1153 bra.uni $L__BB210_96; $L__BB210_221: setp.eq.s32 %p146, %r208, 79; @%p146 bra $L__BB210_256; setp.eq.s32 %p147, %r208, 80; @%p147 bra $L__BB210_223; bra.uni $L__BB210_301; $L__BB210_223: { // callseq 1135, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1135 bra.uni $L__BB210_96; $L__BB210_162: setp.eq.s32 %p192, %r208, 47; @%p192 bra $L__BB210_281; setp.eq.s32 %p193, %r208, 48; @%p193 bra $L__BB210_164; bra.uni $L__BB210_301; $L__BB210_164: { // callseq 1165, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1165 bra.uni $L__BB210_96; $L__BB210_176: setp.eq.s32 %p181, %r208, 54; @%p181 bra $L__BB210_277; setp.eq.s32 %p182, %r208, 55; @%p182 bra $L__BB210_178; bra.uni $L__BB210_301; $L__BB210_178: cvta.to.local.u64 %rd315, %rd87; ld.local.v4.f32 {%f1249, %f1250, %f1251, %f1252}, [%rd315]; { // callseq 1158, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1249; st.param.f32 [param0+4], %f1250; st.param.f32 [param0+8], %f1251; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; .param .b32 param4; st.param.b32 [param4+0], %r211; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1158 bra.uni $L__BB210_96; $L__BB210_212: setp.eq.s32 %p153, %r208, 74; @%p153 bra $L__BB210_265; setp.eq.s32 %p154, %r208, 75; @%p154 bra $L__BB210_214; bra.uni $L__BB210_301; $L__BB210_214: { // callseq 1138, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1138 bra.uni $L__BB210_96; $L__BB210_193: setp.eq.s32 %p168, %r208, 64; @%p168 bra $L__BB210_271; setp.eq.s32 %p169, %r208, 65; @%p169 bra $L__BB210_195; bra.uni $L__BB210_301; $L__BB210_195: { // callseq 1148, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1148 bra.uni $L__BB210_96; $L__BB210_229: setp.eq.s32 %p140, %r208, 84; @%p140 bra $L__BB210_253; setp.eq.s32 %p141, %r208, 85; @%p141 bra $L__BB210_231; bra.uni $L__BB210_301; $L__BB210_231: add.s32 %r220, %r24, 1; { // callseq 1130, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r220; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1130 bra.uni $L__BB210_96; $L__BB210_289: mov.u32 %r259, 23; add.s32 %r260, %r24, 1; { // callseq 1179, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r259; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r260; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1179 bra.uni $L__BB210_96; $L__BB210_288: mov.u32 %r258, 24; { // callseq 1178, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r258; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1178 bra.uni $L__BB210_96; $L__BB210_292: ld.global.v4.u32 {%r265, %r266, %r267, %r268}, [%rd27+16]; add.s32 %r325, %r24, 2; st.local.u32 [%rd28], %r266; st.local.u32 [%rd28+4], %r267; st.local.u32 [%rd28+8], %r268; bra.uni $L__BB210_96; $L__BB210_291: mov.u32 %r264, 16; { // callseq 1183, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r264; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1183 bra.uni $L__BB210_96; $L__BB210_295: ld.local.f32 %f1257, [%rd29]; setp.gtu.ftz.f32 %p244, %f1257, 0f00000000; @%p244 bra $L__BB210_96; add.s32 %r276, %r209, %r24; add.s32 %r325, %r276, 1; bra.uni $L__BB210_96; $L__BB210_298: st.local.u32 [%rd23], %r209; st.local.u32 [%rd24], %r210; st.local.u32 [%rd25], %r211; bra.uni $L__BB210_96; $L__BB210_290: mov.u32 %r263, 21; { // callseq 1181, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r263; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1181 bra.uni $L__BB210_96; $L__BB210_293: mov.u32 %r272, 13; add.s32 %r273, %r24, 1; mov.u32 %r274, 524288; { // callseq 1184, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b32 param1; st.param.b32 [param1+0], %r274; .param .b64 param2; st.param.b64 [param2+0], %rd260; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r272; st.param.b32 [param3+4], %r209; st.param.b32 [param3+8], %r210; st.param.b32 [param3+12], %r211; .param .b32 param4; st.param.b32 [param4+0], %r273; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1184 bra.uni $L__BB210_96; $L__BB210_299: mov.u32 %r280, 3; { // callseq 1190, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r280, %r209, %r210, %r211}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1190 bra.uni $L__BB210_96; $L__BB210_294: { // callseq 1186, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1186 bra.uni $L__BB210_96; $L__BB210_297: mov.u32 %r278, 7; { // callseq 1188, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r278; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1188 bra.uni $L__BB210_96; $L__BB210_279: { // callseq 1162, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1162 bra.uni $L__BB210_96; $L__BB210_278: add.s32 %r248, %r24, 1; { // callseq 1161, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r248; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1161 bra.uni $L__BB210_96; $L__BB210_267: mov.u32 %r227, 71; { // callseq 1142, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r227; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1142 bra.uni $L__BB210_96; $L__BB210_266: { // callseq 1141, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1141 bra.uni $L__BB210_96; $L__BB210_286: mov.u32 %r254, 37; { // callseq 1174, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r254, %r209, %r210, %r211}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1174 bra.uni $L__BB210_96; $L__BB210_285: { // callseq 1173, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; .param .b32 param4; st.param.b32 [param4+0], %r211; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1173 bra.uni $L__BB210_96; $L__BB210_273: mov.u32 %r239, 61; add.s32 %r240, %r24, 1; { // callseq 1152, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r239; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r240; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1152 bra.uni $L__BB210_96; $L__BB210_272: mov.u32 %r237, 62; add.s32 %r238, %r24, 1; { // callseq 1151, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r237; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r238; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1151 bra.uni $L__BB210_96; $L__BB210_255: mov.u32 %r222, 81; { // callseq 1134, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r222; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1134 bra.uni $L__BB210_96; $L__BB210_254: mov.u32 %r221, 82; { // callseq 1133, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r221; st.param.b32 [param2+4], %r209; st.param.b32 [param2+8], %r210; st.param.b32 [param2+12], %r211; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1133 bra.uni $L__BB210_96; $L__BB210_282: { // callseq 1168, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1168 bra.uni $L__BB210_96; $L__BB210_276: mov.u32 %r246, 56; { // callseq 1157, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r246; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1157 bra.uni $L__BB210_96; $L__BB210_275: mov.u32 %r244, 57; add.s32 %r245, %r24, 1; { // callseq 1156, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r244; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r245; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1156 bra.uni $L__BB210_96; $L__BB210_259: setp.eq.s32 %p240, %r211, 255; @%p240 bra $L__BB210_96; add.s64 %rd30, %rd28, 4; add.s64 %rd31, %rd28, 8; setp.eq.s32 %p241, %r210, 0; mul.wide.u32 %rd290, %r211, 4; add.s64 %rd32, %rd26, %rd290; @%p241 bra $L__BB210_264; setp.eq.s32 %p242, %r210, 1; @%p242 bra $L__BB210_263; bra.uni $L__BB210_262; $L__BB210_263: ld.local.f32 %f1247, [%rd30]; st.local.f32 [%rd32], %f1247; bra.uni $L__BB210_96; $L__BB210_257: setp.eq.s32 %p239, %r211, 255; @%p239 bra $L__BB210_96; ld.local.f32 %f1245, [%rd28]; add.s32 %r225, %r211, %r210; mul.wide.u32 %rd286, %r225, 4; add.s64 %rd287, %rd26, %rd286; st.local.f32 [%rd287], %f1245; bra.uni $L__BB210_96; $L__BB210_270: add.s32 %r233, %r24, 1; { // callseq 1147, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r233; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1147 bra.uni $L__BB210_96; $L__BB210_269: mov.u32 %r232, 67; { // callseq 1146, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r232, %r209, %r210, %r211}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1146 bra.uni $L__BB210_96; $L__BB210_252: add.s32 %r219, %r24, 1; { // callseq 1129, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r219; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1129 bra.uni $L__BB210_96; $L__BB210_251: add.s32 %r218, %r24, 1; { // callseq 1128, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r218; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1128 bra.uni $L__BB210_96; $L__BB210_284: { // callseq 1171, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1171 bra.uni $L__BB210_96; $L__BB210_283: add.s32 %r252, %r24, 1; { // callseq 1170, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r252; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1170 bra.uni $L__BB210_96; $L__BB210_280: { // callseq 1164, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd87; .param .b64 param1; st.param.b64 [param1+0], %rd260; .param .b32 param2; st.param.b32 [param2+0], %r209; .param .b32 param3; st.param.b32 [param3+0], %r210; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1164 bra.uni $L__BB210_96; $L__BB210_268: mov.u32 %r229, 69; add.s32 %r230, %r24, 1; { // callseq 1144, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r229; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r230; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1144 bra.uni $L__BB210_96; $L__BB210_274: add.s32 %r242, %r24, 1; { // callseq 1154, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r242; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1154 bra.uni $L__BB210_96; $L__BB210_256: add.s32 %r223, %r24, 1; { // callseq 1136, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; .param .b32 param4; st.param.b32 [param4+0], %r223; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1136 bra.uni $L__BB210_96; $L__BB210_281: cvt.u32.u64 %r249, %rd1; { // callseq 1166, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r249; .param .b64 param1; st.param.b64 [param1+0], %rd87; .param .b64 param2; st.param.b64 [param2+0], %rd260; .param .b32 param3; st.param.b32 [param3+0], %r209; .param .b32 param4; st.param.b32 [param4+0], %r210; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1166 bra.uni $L__BB210_96; $L__BB210_277: add.s32 %r247, %r24, 1; { // callseq 1159, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r247; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1159 bra.uni $L__BB210_96; $L__BB210_265: { // callseq 1139, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; .param .b32 param3; st.param.b32 [param3+0], %r211; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1139 bra.uni $L__BB210_96; $L__BB210_271: mov.u32 %r234, 64; add.s32 %r235, %r24, 1; { // callseq 1149, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r234; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; .param .b32 param2; st.param.b32 [param2+0], %r235; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r325, [retval0+0]; } // callseq 1149 bra.uni $L__BB210_96; $L__BB210_253: { // callseq 1131, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .b32 param1; st.param.b32 [param1+0], %r209; .param .b32 param2; st.param.b32 [param2+0], %r210; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1131 bra.uni $L__BB210_96; $L__BB210_287: mov.u32 %r256, 26; { // callseq 1176, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd260; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r256; st.param.b32 [param1+4], %r209; st.param.b32 [param1+8], %r210; st.param.b32 [param1+12], %r211; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1176 bra.uni $L__BB210_96; $L__BB210_264: ld.local.f32 %f1248, [%rd28]; st.local.f32 [%rd32], %f1248; bra.uni $L__BB210_96; $L__BB210_262: ld.local.f32 %f1246, [%rd31]; st.local.f32 [%rd32], %f1246; bra.uni $L__BB210_96; $L__BB210_301: add.u64 %rd502, %SP, 1024; or.b64 %rd501, %rd502, 12; add.s64 %rd366, %rd501, 40; cvta.to.local.u64 %rd367, %rd366; ld.local.u32 %r283, [%rd367]; and.b32 %r114, %r283, 2; @%p96 bra $L__BB210_304; bra.uni $L__BB210_302; $L__BB210_304: setp.eq.s32 %p248, %r114, 0; mov.f32 %f1478, 0f00000000; mov.f32 %f1479, %f1478; mov.f32 %f1480, %f1478; @%p248 bra $L__BB210_306; add.s64 %rd378, %rd87, 280; cvta.to.local.u64 %rd379, %rd378; ld.local.v2.f32 {%f1478, %f1479}, [%rd379]; add.s64 %rd380, %rd87, 288; cvta.to.local.u64 %rd381, %rd380; ld.local.f32 %f1480, [%rd381]; bra.uni $L__BB210_306; $L__BB210_302: setp.eq.s32 %p246, %r114, 0; mov.f32 %f1478, 0f00000000; mov.f32 %f1479, %f1478; mov.f32 %f1480, %f1478; @%p246 bra $L__BB210_306; add.u64 %rd504, %SP, 1024; or.b64 %rd503, %rd504, 12; add.s64 %rd368, %rd503, 20; add.s64 %rd369, %rd503, 12; cvta.to.local.u64 %rd370, %rd369; ld.local.v2.f32 {%f1264, %f1265}, [%rd370]; cvta.to.local.u64 %rd371, %rd368; ld.local.v4.f32 {%f1268, %f1269, %f1270, %f1271}, [%rd371]; mul.ftz.f32 %f1276, %f1265, %f1270; fma.rn.ftz.f32 %f1277, %f1264, %f1269, %f1276; fma.rn.ftz.f32 %f1278, %f1268, %f1271, %f1277; abs.ftz.f32 %f1279, %f1278; setp.gt.ftz.f32 %p247, %f1279, 0f00000000; selp.f32 %f1280, 0f3F800000, 0f00000000, %p247; add.s64 %rd373, %rd504, 280; cvta.to.local.u64 %rd374, %rd373; ld.local.v2.f32 {%f1281, %f1282}, [%rd374]; add.s64 %rd375, %rd504, 288; cvta.to.local.u64 %rd376, %rd375; ld.local.f32 %f1285, [%rd376]; mul.ftz.f32 %f1480, %f1280, %f1285; mul.ftz.f32 %f1479, %f1280, %f1282; mul.ftz.f32 %f1478, %f1281, %f1280; $L__BB210_306: mul.ftz.f32 %f1481, %f1415, %f1478; mul.ftz.f32 %f1482, %f1415, %f1479; mul.ftz.f32 %f1483, %f1415, %f1480; @%p7 bra $L__BB210_308; ld.const.u64 %rd508, [kernel_params+2368]; mul.wide.s32 %rd507, %r4, 192; cvta.to.global.u64 %rd506, %rd508; add.s64 %rd505, %rd506, %rd507; ld.global.f32 %f1291, [%rd505+28]; mul.ftz.f32 %f1481, %f1481, %f1291; ld.global.f32 %f1292, [%rd505+32]; mul.ftz.f32 %f1482, %f1482, %f1292; ld.global.f32 %f1293, [%rd505+36]; mul.ftz.f32 %f1483, %f1483, %f1293; $L__BB210_308: setp.eq.ftz.f32 %p250, %f1482, 0f00000000; setp.eq.ftz.f32 %p251, %f1481, 0f00000000; and.pred %p252, %p251, %p250; setp.eq.ftz.f32 %p253, %f1483, 0f00000000; and.pred %p254, %p253, %p252; @%p254 bra $L__BB210_353; and.b32 %r284, %r13, 2048; setp.ne.s32 %p255, %r284, 0; mov.f32 %f1484, 0f3F800000; @%p255 bra $L__BB210_311; shl.b64 %rd513, %rd1, 2; ld.const.u64 %rd382, [kernel_params+2560]; cvta.to.global.u64 %rd383, %rd382; add.s64 %rd385, %rd383, %rd513; ld.global.f32 %f1295, [%rd385]; mul.ftz.f32 %f1296, %f1295, %f1295; fma.rn.ftz.f32 %f1297, %f1416, %f1416, %f1296; div.approx.ftz.f32 %f1484, %f1296, %f1297; $L__BB210_311: ld.const.u64 %rd386, [kernel_params+2584]; cvta.to.global.u64 %rd387, %rd386; mul.lo.s64 %rd388, %rd1, 12; add.s64 %rd389, %rd387, %rd388; ld.global.f32 %f1299, [%rd389]; mul.ftz.f32 %f1300, %f1481, %f1299; ld.global.f32 %f1301, [%rd389+4]; mul.ftz.f32 %f1302, %f1482, %f1301; ld.global.f32 %f1303, [%rd389+8]; mul.ftz.f32 %f1304, %f1483, %f1303; mul.ftz.f32 %f348, %f1484, %f1300; mul.ftz.f32 %f349, %f1484, %f1302; mul.ftz.f32 %f350, %f1484, %f1304; ld.const.u64 %rd390, [kernel_params+2464]; cvta.to.global.u64 %rd391, %rd390; shl.b64 %rd392, %rd1, 1; add.s64 %rd393, %rd391, %rd392; ld.global.u16 %rs1, [%rd393]; mov.b32 %r115, %f348; setp.nan.ftz.f32 %p256, %f348, %f348; mov.f32 %f1493, 0f00000000; mov.f32 %f1491, %f1493; @%p256 bra $L__BB210_314; or.b32 %r285, %r115, -2147483648; setp.ne.s32 %p257, %r285, -2147483648; add.ftz.f32 %f1306, %f348, %f348; setp.eq.ftz.f32 %p258, %f1306, %f348; and.pred %p259, %p258, %p257; mov.f32 %f1491, %f1493; @%p259 bra $L__BB210_314; shl.b32 %r286, %r115, 1; setp.lt.u32 %p260, %r286, -16777215; selp.f32 %f1491, %f348, 0f00000000, %p260; $L__BB210_314: mov.b32 %r116, %f349; setp.nan.ftz.f32 %p261, %f349, %f349; mov.f32 %f1492, %f1493; @%p261 bra $L__BB210_317; or.b32 %r287, %r116, -2147483648; setp.ne.s32 %p262, %r287, -2147483648; add.ftz.f32 %f1309, %f349, %f349; setp.eq.ftz.f32 %p263, %f1309, %f349; and.pred %p264, %p263, %p262; mov.f32 %f1492, %f1493; @%p264 bra $L__BB210_317; shl.b32 %r288, %r116, 1; setp.lt.u32 %p265, %r288, -16777215; selp.f32 %f1492, %f349, 0f00000000, %p265; $L__BB210_317: mov.b32 %r117, %f350; setp.nan.ftz.f32 %p266, %f350, %f350; @%p266 bra $L__BB210_320; or.b32 %r289, %r117, -2147483648; setp.ne.s32 %p267, %r289, -2147483648; add.ftz.f32 %f1312, %f350, %f350; setp.eq.ftz.f32 %p268, %f1312, %f350; and.pred %p269, %p268, %p267; @%p269 bra $L__BB210_320; shl.b32 %r290, %r117, 1; setp.lt.u32 %p270, %r290, -16777215; selp.f32 %f1493, %f350, 0f00000000, %p270; $L__BB210_320: ld.const.f32 %f1313, [kernel_params+1580]; ld.const.f32 %f1314, [kernel_params+1584]; setp.gt.u16 %p271, %rs1, 1; selp.f32 %f357, %f1314, %f1313, %p271; abs.ftz.f32 %f1315, %f1492; abs.ftz.f32 %f1316, %f1491; add.ftz.f32 %f1317, %f1316, %f1315; abs.ftz.f32 %f1318, %f1493; add.ftz.f32 %f358, %f1318, %f1317; setp.leu.ftz.f32 %p272, %f358, %f357; @%p272 bra $L__BB210_322; div.approx.ftz.f32 %f1319, %f357, %f358; mul.ftz.f32 %f1491, %f1491, %f1319; mul.ftz.f32 %f1492, %f1492, %f1319; mul.ftz.f32 %f1493, %f1493, %f1319; $L__BB210_322: shl.b64 %rd512, %rd1, 2; ld.const.u64 %rd511, [kernel_params+2544]; cvta.to.global.u64 %rd510, %rd511; add.s64 %rd509, %rd510, %rd512; ld.const.v2.u64 {%rd394, %rd395}, [kernel_params+2448]; cvta.to.global.u64 %rd398, %rd394; add.s64 %rd400, %rd398, %rd512; ld.global.u32 %rd401, [%rd400]; ld.const.s32 %rd402, [kernel_params+1264]; mul.lo.s64 %rd33, %rd402, %rd401; ld.global.u32 %r118, [%rd509]; cvta.to.global.u64 %rd403, %rd395; add.s64 %rd404, %rd403, %rd512; ld.global.u32 %r119, [%rd404]; ld.const.u32 %r291, [kernel_params+1620]; setp.eq.s32 %p273, %r291, 0; @%p273 bra $L__BB210_327; and.b32 %r292, %r118, 536870912; setp.ne.s32 %p274, %r292, 0; @%p274 bra $L__BB210_325; ld.const.s32 %rd405, [kernel_params+1384]; add.s64 %rd406, %rd33, %rd405; cvta.to.global.u64 %rd407, %rd36; shl.b64 %rd408, %rd406, 2; add.s64 %rd409, %rd407, %rd408; add.s64 %rd410, %rd409, 4; add.s64 %rd411, %rd409, 8; atom.global.add.f32 %f1320, [%rd409], %f1491; atom.global.add.f32 %f1321, [%rd410], %f1492; atom.global.add.f32 %f1322, [%rd411], %f1493; $L__BB210_325: and.b32 %r293, %r118, 1073741824; setp.eq.s32 %p275, %r293, 0; @%p275 bra $L__BB210_327; bra.uni $L__BB210_326; $L__BB210_327: ld.const.u32 %r326, [kernel_params+1260]; and.b32 %r294, %r326, 2; setp.eq.s32 %p276, %r294, 0; @%p276 bra $L__BB210_329; ld.const.s32 %rd419, [kernel_params+1268]; add.s64 %rd420, %rd33, %rd419; cvta.to.global.u64 %rd421, %rd36; shl.b64 %rd422, %rd420, 2; add.s64 %rd423, %rd421, %rd422; add.s64 %rd424, %rd423, 4; add.s64 %rd425, %rd423, 8; atom.global.add.f32 %f1326, [%rd423], %f1491; atom.global.add.f32 %f1327, [%rd424], %f1492; atom.global.add.f32 %f1328, [%rd425], %f1493; $L__BB210_329: ld.const.u32 %r122, [kernel_params+1400]; setp.eq.s32 %p277, %r122, -1; @%p277 bra $L__BB210_332; and.b32 %r295, %r119, -1431655766; popc.b32 %r296, %r295; and.b32 %r297, %r296, 1; setp.eq.b32 %p278, %r297, 1; mov.pred %p279, 0; xor.pred %p280, %p278, %p279; not.pred %p281, %p280; @%p281 bra $L__BB210_332; cvt.s64.s32 %rd426, %r122; add.s64 %rd427, %rd33, %rd426; cvta.to.global.u64 %rd428, %rd36; shl.b64 %rd429, %rd427, 2; add.s64 %rd430, %rd428, %rd429; add.s64 %rd431, %rd430, 4; add.s64 %rd432, %rd430, 8; add.s64 %rd433, %rd430, 12; add.ftz.f32 %f1329, %f1491, %f1491; atom.global.add.f32 %f1330, [%rd430], %f1329; add.ftz.f32 %f1331, %f1492, %f1492; atom.global.add.f32 %f1332, [%rd431], %f1331; add.ftz.f32 %f1333, %f1493, %f1493; atom.global.add.f32 %f1334, [%rd432], %f1333; atom.global.add.f32 %f1335, [%rd433], 0f00000000; bra.uni $L__BB210_332; $L__BB210_326: ld.const.s32 %rd412, [kernel_params+1376]; add.s64 %rd413, %rd33, %rd412; cvta.to.global.u64 %rd414, %rd36; shl.b64 %rd415, %rd413, 2; add.s64 %rd416, %rd414, %rd415; add.s64 %rd417, %rd416, 4; add.s64 %rd418, %rd416, 8; atom.global.add.f32 %f1323, [%rd416], %f1491; atom.global.add.f32 %f1324, [%rd417], %f1492; atom.global.add.f32 %f1325, [%rd418], %f1493; ld.const.u32 %r326, [kernel_params+1260]; $L__BB210_332: setp.eq.s32 %p282, %r326, 0; @%p282 bra $L__BB210_353; ld.const.u32 %r327, [kernel_params+1352]; and.b32 %r298, %r118, 16777216; setp.eq.s32 %p283, %r298, 0; ld.const.u32 %r125, [kernel_params+1428]; setp.eq.s32 %p284, %r125, -1; or.pred %p285, %p283, %p284; @%p285 bra $L__BB210_335; ld.const.u64 %rd434, [kernel_params+2616]; cvta.to.global.u64 %rd435, %rd434; add.s64 %rd437, %rd435, %rd388; ld.global.f32 %f1336, [%rd437]; mul.ftz.f32 %f1337, %f1491, %f1336; ld.global.f32 %f1338, [%rd437+4]; mul.ftz.f32 %f1339, %f1492, %f1338; ld.global.f32 %f1340, [%rd437+8]; mul.ftz.f32 %f1341, %f1493, %f1340; cvta.to.global.u64 %rd438, %rd36; cvt.s64.s32 %rd439, %r125; add.s64 %rd440, %rd33, %rd439; shl.b64 %rd441, %rd440, 2; add.s64 %rd442, %rd438, %rd441; add.s64 %rd443, %rd442, 4; add.s64 %rd444, %rd442, 8; atom.global.add.f32 %f1342, [%rd442], %f1337; atom.global.add.f32 %f1343, [%rd443], %f1339; atom.global.add.f32 %f1344, [%rd444], %f1341; $L__BB210_335: setp.eq.s32 %p286, %r322, -1; ld.const.u32 %r126, [kernel_params+1444]; setp.eq.s32 %p287, %r126, -1; or.pred %p288, %p286, %p287; @%p288 bra $L__BB210_337; mul.lo.s32 %r299, %r322, 3; cvt.s64.s32 %rd445, %r299; add.s64 %rd446, %rd33, %rd445; cvt.s64.s32 %rd447, %r126; add.s64 %rd448, %rd446, %rd447; cvta.to.global.u64 %rd449, %rd36; shl.b64 %rd450, %rd448, 2; add.s64 %rd451, %rd449, %rd450; add.s64 %rd452, %rd451, 4; add.s64 %rd453, %rd451, 8; atom.global.add.f32 %f1345, [%rd451], %f1491; atom.global.add.f32 %f1346, [%rd452], %f1492; atom.global.add.f32 %f1347, [%rd453], %f1493; $L__BB210_337: and.b32 %r300, %r118, 100663296; setp.eq.s32 %p289, %r300, 0; @%p289 bra $L__BB210_351; ld.const.u8 %rs2, [kernel_params+2]; and.b16 %rs3, %rs2, 32; setp.eq.s16 %p290, %rs3, 0; mov.u32 %r327, -1; @%p290 bra $L__BB210_351; bra.uni $L__BB210_339; $L__BB210_351: setp.eq.s32 %p301, %r327, -1; @%p301 bra $L__BB210_353; cvt.s64.s32 %rd475, %r327; add.s64 %rd476, %rd33, %rd475; cvta.to.global.u64 %rd477, %rd36; shl.b64 %rd478, %rd476, 2; add.s64 %rd479, %rd477, %rd478; add.s64 %rd480, %rd479, 4; add.s64 %rd481, %rd479, 8; atom.global.add.f32 %f1367, [%rd479], %f1491; atom.global.add.f32 %f1368, [%rd480], %f1492; atom.global.add.f32 %f1369, [%rd481], %f1493; bra.uni $L__BB210_353; $L__BB210_339: and.b32 %r302, %r118, 536870912; setp.ne.s32 %p291, %r302, 0; @%p291 bra $L__BB210_353; and.b32 %r303, %r118, 33554432; setp.eq.s32 %p292, %r303, 0; @%p292 bra $L__BB210_347; bra.uni $L__BB210_341; $L__BB210_347: and.b32 %r312, %r118, 67108864; setp.eq.s32 %p299, %r312, 0; @%p299 bra $L__BB210_351; setp.eq.s16 %p300, %rs1, 1; @%p300 bra $L__BB210_350; bra.uni $L__BB210_349; $L__BB210_350: ld.const.u32 %r327, [kernel_params+1348]; bra.uni $L__BB210_351; $L__BB210_353: ld.const.u64 %rd482, [kernel_params+2472]; cvta.to.global.u64 %rd483, %rd482; shl.b64 %rd484, %rd1, 1; add.s64 %rd485, %rd483, %rd484; ld.global.u16 %r313, [%rd485]; add.s32 %r314, %r313, 1; st.global.u16 [%rd485], %r314; ld.const.u32 %r315, [kernel_params+1556]; setp.lt.u32 %p302, %r314, %r315; ld.const.u64 %rd486, [kernel_params+4256]; cvta.to.global.u64 %rd34, %rd486; add.s64 %rd487, %rd34, 28; atom.global.add.u32 %r316, [%rd487], -1; @%p302 bra $L__BB210_355; bra.uni $L__BB210_354; $L__BB210_355: add.s64 %rd492, %rd34, 8; atom.global.add.u32 %r317, [%rd492], 1; ld.const.u64 %rd493, [kernel_params+2520]; cvta.to.global.u64 %rd494, %rd493; add.s64 %rd496, %rd494, %rd484; mov.u16 %rs5, 2; st.global.u16 [%rd496], %rs5; bra.uni $L__BB210_356; $L__BB210_354: ld.const.u64 %rd488, [kernel_params+2520]; cvta.to.global.u64 %rd489, %rd488; add.s64 %rd491, %rd489, %rd484; mov.u16 %rs4, 0; st.global.u16 [%rd491], %rs4; $L__BB210_356: ret; $L__BB210_341: ld.const.u64 %rd454, [kernel_params+2600]; cvta.to.global.u64 %rd455, %rd454; add.s64 %rd457, %rd455, %rd388; ld.global.f32 %f365, [%rd457]; ld.global.f32 %f366, [%rd457+4]; ld.global.f32 %f367, [%rd457+8]; ld.const.u64 %rd458, [kernel_params+2608]; cvta.to.global.u64 %rd459, %rd458; add.s64 %rd460, %rd459, %rd388; ld.global.f32 %f368, [%rd460]; ld.global.f32 %f369, [%rd460+4]; ld.global.f32 %f370, [%rd460+8]; ld.const.u32 %r304, [kernel_params+1324]; ld.const.u32 %r305, [kernel_params+1340]; setp.eq.s16 %p293, %rs1, 1; selp.b32 %r127, %r305, %r304, %p293; setp.eq.s32 %p294, %r127, -1; @%p294 bra $L__BB210_343; cvt.s64.s32 %rd461, %r127; add.s64 %rd462, %rd33, %rd461; cvta.to.global.u64 %rd463, %rd36; shl.b64 %rd464, %rd462, 2; add.s64 %rd465, %rd463, %rd464; add.s64 %rd466, %rd465, 4; add.s64 %rd467, %rd465, 8; mul.ftz.f32 %f1348, %f1491, %f368; atom.global.add.f32 %f1349, [%rd465], %f1348; mul.ftz.f32 %f1350, %f1492, %f369; atom.global.add.f32 %f1351, [%rd466], %f1350; mul.ftz.f32 %f1352, %f1493, %f370; atom.global.add.f32 %f1353, [%rd467], %f1352; $L__BB210_343: ld.const.u32 %r306, [kernel_params+1328]; ld.const.u32 %r307, [kernel_params+1344]; selp.b32 %r128, %r307, %r306, %p293; setp.eq.s32 %p296, %r128, -1; @%p296 bra $L__BB210_345; mov.f32 %f1354, 0f3F800000; sub.ftz.f32 %f1355, %f1354, %f365; sub.ftz.f32 %f1356, %f1355, %f368; sub.ftz.f32 %f1357, %f1354, %f366; sub.ftz.f32 %f1358, %f1357, %f369; sub.ftz.f32 %f1359, %f1354, %f367; sub.ftz.f32 %f1360, %f1359, %f370; cvt.s64.s32 %rd468, %r128; add.s64 %rd469, %rd33, %rd468; cvta.to.global.u64 %rd470, %rd36; shl.b64 %rd471, %rd469, 2; add.s64 %rd472, %rd470, %rd471; mul.ftz.f32 %f1361, %f1491, %f1356; mul.ftz.f32 %f1362, %f1492, %f1358; mul.ftz.f32 %f1363, %f1493, %f1360; add.s64 %rd473, %rd472, 4; add.s64 %rd474, %rd472, 8; atom.global.add.f32 %f1364, [%rd472], %f1361; atom.global.add.f32 %f1365, [%rd473], %f1362; atom.global.add.f32 %f1366, [%rd474], %f1363; $L__BB210_345: ld.const.u32 %r309, [kernel_params+1320]; ld.const.u32 %r310, [kernel_params+1336]; selp.b32 %r129, %r310, %r309, %p293; setp.eq.s32 %p298, %r129, -1; @%p298 bra $L__BB210_351; mul.ftz.f32 %f1491, %f1491, %f365; mul.ftz.f32 %f1492, %f1492, %f366; mul.ftz.f32 %f1493, %f1493, %f367; mov.u32 %r327, %r129; bra.uni $L__BB210_351; $L__BB210_349: ld.const.u32 %r327, [kernel_params+1332]; bra.uni $L__BB210_351; } // .globl kernel_gpu_integrator_shade_shadow .visible .entry kernel_gpu_integrator_shade_shadow( .param .u64 kernel_gpu_integrator_shade_shadow_param_0, .param .u64 kernel_gpu_integrator_shade_shadow_param_1, .param .u32 kernel_gpu_integrator_shade_shadow_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot211[1328]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<535>; .reg .b16 %rs<52>; .reg .f32 %f<3177>; .reg .b32 %r<912>; .reg .b64 %rd<1206>; mov.u64 %SPL, __local_depot211; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd144, [kernel_gpu_integrator_shade_shadow_param_0]; ld.param.u64 %rd145, [kernel_gpu_integrator_shade_shadow_param_1]; ld.param.u32 %r361, [kernel_gpu_integrator_shade_shadow_param_2]; add.u64 %rd146, %SP, 304; add.u64 %rd1, %SPL, 304; mov.u32 %r362, %ntid.x; mov.u32 %r363, %ctaid.x; mov.u32 %r364, %tid.x; mad.lo.s32 %r871, %r363, %r362, %r364; setp.ge.s32 %p2, %r871, %r361; @%p2 bra $L__BB211_769; setp.eq.s64 %p3, %rd144, 0; @%p3 bra $L__BB211_3; cvta.to.global.u64 %rd147, %rd144; mul.wide.s32 %rd148, %r871, 4; add.s64 %rd149, %rd147, %rd148; ld.global.u32 %r871, [%rd149]; $L__BB211_3: ld.const.u64 %rd150, [kernel_params+3464]; cvta.to.global.u64 %rd151, %rd150; cvt.s64.s32 %rd2, %r871; mul.wide.s32 %rd152, %r871, 2; add.s64 %rd153, %rd151, %rd152; ld.global.u16 %rs1, [%rd153]; ld.const.v4.u32 {%r366, %r367, %r368, %r369}, [kernel_params]; ld.const.u64 %rd154, [kernel_params+3752]; cvta.to.global.u64 %rd155, %rd154; mul.wide.s32 %rd156, %r871, 4; add.s64 %rd3, %rd155, %rd156; ld.const.u64 %rd157, [kernel_params+3496]; cvta.to.global.u64 %rd158, %rd157; mul.wide.s32 %rd159, %r871, 12; add.s64 %rd4, %rd158, %rd159; ld.const.u64 %rd160, [kernel_params+3504]; cvta.to.global.u64 %rd161, %rd160; add.s64 %rd5, %rd161, %rd159; ld.const.u64 %rd162, [kernel_params+3528]; cvta.to.global.u64 %rd163, %rd162; add.s64 %rd6, %rd163, %rd156; ld.const.u64 %rd164, [kernel_params+3536]; cvta.to.global.u64 %rd165, %rd164; add.s64 %rd7, %rd165, %rd156; add.u64 %rd166, %SP, 0; add.s64 %rd167, %rd166, 76; ld.const.v2.u64 {%rd168, %rd169}, [kernel_params+2176]; cvta.to.global.u64 %rd10, %rd169; ld.const.u64 %rd170, [kernel_params+3432]; cvta.to.global.u64 %rd171, %rd170; add.s64 %rd11, %rd171, %rd159; or.b64 %rd173, %rd166, 8; add.s64 %rd174, %rd166, 48; add.s64 %rd175, %rd166, 52; add.s64 %rd176, %rd166, 56; add.s64 %rd177, %rd166, 84; add.s64 %rd179, %rd166, 80; add.s64 %rd180, %rd166, 60; add.s64 %rd181, %rd166, 64; ld.const.u64 %rd182, [kernel_params+2160]; cvta.to.global.u64 %rd12, %rd182; ld.const.v2.u64 {%rd183, %rd184}, [kernel_params+3360]; mov.u32 %r872, 0; cvta.to.global.u64 %rd187, %rd183; add.s64 %rd13, %rd187, %rd156; cvta.to.global.u64 %rd188, %rd184; add.s64 %rd14, %rd188, %rd152; ld.const.u64 %rd189, [kernel_params+3352]; cvta.to.global.u64 %rd190, %rd189; add.s64 %rd15, %rd190, %rd156; ld.const.v2.u64 {%rd191, %rd192}, [kernel_params+2400]; cvta.to.global.u64 %rd16, %rd192; ld.const.u64 %rd195, [kernel_params+2192]; cvta.to.global.u64 %rd17, %rd195; cvta.to.local.u64 %rd18, %rd174; add.s64 %rd197, %rd166, 32; cvta.to.local.u64 %rd19, %rd197; add.s64 %rd198, %rd166, 116; add.s64 %rd199, %rd166, 120; add.s64 %rd200, %rd166, 124; ld.const.u32 %r8, [kernel_params+1612]; cvt.rn.f32.s32 %f3, %r8; ld.const.v2.u64 {%rd201, %rd202}, [kernel_params+2304]; or.b64 %rd203, %rd166, 12; cvta.to.local.u64 %rd22, %rd203; add.s64 %rd204, %rd166, 16; add.s64 %rd205, %rd166, 20; cvta.to.local.u64 %rd23, %rd198; cvta.to.local.u64 %rd24, %rd199; cvta.to.local.u64 %rd25, %rd200; ld.const.v2.u64 {%rd206, %rd207}, [kernel_params+2256]; cvta.to.local.u64 %rd28, %rd173; add.s64 %rd208, %rd166, 260; cvta.to.local.u64 %rd29, %rd208; add.s64 %rd209, %rd166, 264; cvta.to.local.u64 %rd30, %rd209; cvta.to.local.u64 %rd31, %rd175; cvta.to.local.u64 %rd32, %rd176; cvta.to.local.u64 %rd33, %rd167; cvta.to.local.u64 %rd34, %rd179; cvta.to.global.u64 %rd36, %rd191; add.s64 %rd210, %rd166, 268; cvta.to.local.u64 %rd40, %rd210; add.s64 %rd211, %rd166, 272; cvta.to.local.u64 %rd41, %rd211; add.s64 %rd212, %rd166, 276; cvta.to.local.u64 %rd42, %rd212; cvta.to.local.u64 %rd43, %rd177; add.s64 %rd44, %rd166, 144; cvta.to.local.u64 %rd45, %rd180; add.s64 %rd46, %rd166, 176; add.s64 %rd47, %rd166, 160; add.s64 %rd48, %rd166, 192; ld.const.u32 %r9, [kernel_params+972]; setp.eq.s32 %p4, %r9, 1; selp.f32 %f4, 0f3F000000, 0f00000000, %p4; cvta.to.local.u64 %rd49, %rd181; add.s64 %rd50, %rd166, 224; add.s64 %rd51, %rd166, 208; ld.const.f32 %f860, [kernel_params+256]; mov.f32 %f861, 0f3F800000; sub.ftz.f32 %f5, %f861, %f860; cvta.to.global.u64 %rd52, %rd202; ld.const.u64 %rd213, [kernel_params+2240]; cvta.to.global.u64 %rd53, %rd213; ld.const.u64 %rd214, [kernel_params+2344]; cvta.to.global.u64 %rd54, %rd214; ld.const.u64 %rd55, [kernel_params+2432]; cvta.to.global.u64 %rd56, %rd55; add.s64 %rd215, %rd166, 292; cvta.to.local.u64 %rd57, %rd215; ld.const.v2.u64 {%rd216, %rd217}, [kernel_params+2224]; ld.const.u64 %rd218, [kernel_params+2296]; cvta.to.global.u64 %rd60, %rd218; ld.const.u64 %rd219, [kernel_params+2272]; cvta.to.global.u64 %rd61, %rd219; ld.const.v2.u32 {%r370, %r371}, [kernel_params+1600]; ld.const.u64 %rd220, [kernel_params+2424]; cvta.to.global.u64 %rd62, %rd220; ld.const.u64 %rd221, [kernel_params+2216]; cvta.to.global.u64 %rd63, %rd221; cvta.to.local.u64 %rd64, %rd204; cvta.to.local.u64 %rd65, %rd205; ld.const.u64 %rd225, [kernel_params+3384]; cvta.to.global.u64 %rd226, %rd225; add.s64 %rd70, %rd226, %rd152; add.u64 %rd71, %SPL, 0; setp.gt.u16 %p6, %rs1, 3; mov.pred %p222, 0; setp.eq.s32 %p69, %r9, 2; cvta.to.global.u64 %rd641, %rd217; cvta.to.global.u64 %rd644, %rd206; cvta.to.global.u64 %rd664, %rd216; cvta.to.global.u64 %rd607, %rd201; add.s32 %r793, %r369, -1; cvta.to.global.u64 %rd547, %rd207; bra.uni $L__BB211_4; $L__BB211_452: and.b32 %r785, %r324, 1; setp.eq.b32 %p441, %r785, 1; mov.pred %p442, 0; xor.pred %p443, %p441, %p442; not.pred %p444, %p443; ld.global.u32 %r904, [%rd3]; @%p444 bra $L__BB211_690; bra.uni $L__BB211_453; $L__BB211_690: mov.u64 %rd1181, kernel_params; add.s64 %rd1205, %rd1181, 3752; add.s64 %rd1204, %rd1181, 3744; setp.eq.s32 %p451, %r904, -1; mov.u32 %r909, 0; @%p451 bra $L__BB211_694; mov.u64 %rd1182, kernel_params; ld.local.u32 %r335, [%rd33]; add.s64 %rd1204, %rd1182, 3744; ld.local.u32 %r336, [%rd18]; mov.u32 %r909, 0; $L__BB211_692: shl.b64 %rd1183, %rd2, 2; ld.const.u64 %rd964, [%rd1204]; cvta.to.global.u64 %rd965, %rd964; add.s64 %rd967, %rd965, %rd1183; ld.global.u32 %r792, [%rd967]; setp.eq.s32 %p452, %r792, %r335; setp.eq.s32 %p453, %r904, %r336; and.pred %p454, %p452, %p453; @%p454 bra $L__BB211_696; shl.b64 %rd1185, %rd2, 2; mov.u64 %rd1184, kernel_params; add.s32 %r909, %r909, 1; mul.wide.s32 %rd968, %r909, 16; add.s64 %rd970, %rd1184, %rd968; add.s64 %rd1204, %rd970, 3744; add.s64 %rd1205, %rd970, 3752; ld.const.u64 %rd971, [%rd970+3752]; cvta.to.global.u64 %rd972, %rd971; add.s64 %rd974, %rd972, %rd1185; ld.global.u32 %r904, [%rd974]; setp.ne.s32 %p455, %r904, -1; @%p455 bra $L__BB211_692; $L__BB211_694: setp.ge.u32 %p456, %r909, %r793; @%p456 bra $L__BB211_696; shl.b64 %rd1187, %rd2, 2; mov.u64 %rd1186, kernel_params; ld.local.u32 %r794, [%rd33]; ld.local.u32 %r795, [%rd18]; ld.const.u64 %rd975, [%rd1204]; cvta.to.global.u64 %rd976, %rd975; add.s64 %rd978, %rd976, %rd1187; st.global.u32 [%rd978], %r794; ld.const.u64 %rd979, [%rd1205]; cvta.to.global.u64 %rd980, %rd979; add.s64 %rd981, %rd980, %rd1187; st.global.u32 [%rd981], %r795; add.s32 %r796, %r909, 1; mul.wide.s32 %rd982, %r796, 16; add.s64 %rd984, %rd1186, %rd982; ld.const.v2.u64 {%rd985, %rd986}, [%rd984+3744]; cvta.to.global.u64 %rd989, %rd985; add.s64 %rd990, %rd989, %rd1187; mov.u32 %r797, -1; st.global.u32 [%rd990], %r797; cvta.to.global.u64 %rd991, %rd986; add.s64 %rd992, %rd991, %rd1187; st.global.u32 [%rd992], %r797; bra.uni $L__BB211_696; $L__BB211_453: setp.eq.s32 %p445, %r904, -1; @%p445 bra $L__BB211_696; mov.u64 %rd1175, kernel_params; mov.u32 %r905, 0; ld.local.u32 %r326, [%rd33]; add.s64 %rd1200, %rd1175, 3744; ld.local.u32 %r327, [%rd18]; $L__BB211_455: shl.b64 %rd1177, %rd2, 2; mov.u64 %rd1176, kernel_params; ld.const.u64 %rd929, [%rd1200]; cvta.to.global.u64 %rd930, %rd929; add.s64 %rd932, %rd930, %rd1177; ld.global.u32 %r787, [%rd932]; setp.eq.s32 %p446, %r787, %r326; setp.eq.s32 %p447, %r904, %r327; and.pred %p448, %p446, %p447; add.s32 %r330, %r905, 1; mul.wide.s32 %rd933, %r330, 16; add.s64 %rd935, %rd1176, %rd933; add.s64 %rd130, %rd935, 3752; @%p448 bra $L__BB211_688; shl.b64 %rd1178, %rd2, 2; add.s64 %rd1200, %rd935, 3744; ld.const.u64 %rd939, [%rd130]; cvta.to.global.u64 %rd940, %rd939; add.s64 %rd942, %rd940, %rd1178; ld.global.u32 %r904, [%rd942]; setp.eq.s32 %p449, %r904, -1; mov.u32 %r905, %r330; @%p449 bra $L__BB211_696; bra.uni $L__BB211_455; $L__BB211_688: ld.const.v2.u64 {%rd1202, %rd1201}, [%rd130+-24]; $L__BB211_689: shl.b64 %rd1180, %rd2, 2; mov.u64 %rd1179, kernel_params; add.s32 %r905, %r905, 1; mul.wide.s32 %rd945, %r905, 16; add.s64 %rd947, %rd1179, %rd945; ld.const.v2.u64 {%rd948, %rd949}, [%rd947+3744]; cvta.to.global.u64 %rd950, %rd948; add.s64 %rd952, %rd950, %rd1180; ld.global.u32 %r788, [%rd952]; cvta.to.global.u64 %rd953, %rd949; add.s64 %rd954, %rd953, %rd1180; ld.global.u32 %r789, [%rd954]; cvta.to.global.u64 %rd955, %rd1202; add.s64 %rd956, %rd955, %rd1180; st.global.u32 [%rd956], %r788; cvta.to.global.u64 %rd957, %rd1201; add.s64 %rd958, %rd957, %rd1180; st.global.u32 [%rd958], %r789; setp.eq.s32 %p450, %r789, -1; mov.u64 %rd1201, %rd949; mov.u64 %rd1202, %rd948; @%p450 bra $L__BB211_696; bra.uni $L__BB211_689; $L__BB211_4: cvt.u32.u16 %r372, %rs1; min.u32 %r373, %r372, 4; setp.ge.u32 %p5, %r872, %r373; and.pred %p7, %p6, %p5; @%p7 bra $L__BB211_342; mov.u16 %rs44, 1; and.b32 %r374, %r366, 131072; setp.eq.s32 %p8, %r374, 0; @%p8 bra $L__BB211_7; ld.global.u32 %r375, [%rd3]; setp.eq.s32 %p9, %r375, -1; selp.u16 %rs44, 1, 0, %p9; $L__BB211_7: setp.ne.s16 %p10, %rs44, 0; @%p10 bra $L__BB211_342; cvt.u32.u16 %r854, %rs1; min.u32 %r853, %r854, 4; ld.global.f32 %f6, [%rd11]; ld.global.f32 %f7, [%rd11+4]; ld.global.f32 %f8, [%rd11+8]; add.s32 %r376, %r872, -1; mov.u64 %rd227, kernel_params; add.s64 %rd228, %rd227, 3552; mul.wide.s32 %rd229, %r376, 48; add.s64 %rd230, %rd228, %rd229; setp.eq.s32 %p11, %r872, 0; add.s64 %rd231, %rd227, 3512; selp.b64 %rd232, %rd231, %rd230, %p11; ld.const.u64 %rd233, [%rd232]; cvta.to.global.u64 %rd234, %rd233; shl.b64 %rd235, %rd2, 2; add.s64 %rd236, %rd234, %rd235; setp.lt.s32 %p12, %r872, %r853; mul.wide.s32 %rd237, %r872, 48; add.s64 %rd238, %rd228, %rd237; add.s64 %rd239, %rd227, 3520; selp.b64 %rd240, %rd238, %rd239, %p12; ld.const.u64 %rd241, [%rd240]; cvta.to.global.u64 %rd242, %rd241; add.s64 %rd243, %rd242, %rd235; ld.global.f32 %f9, [%rd243]; ld.global.f32 %f10, [%rd236]; ld.global.f32 %f11, [%rd5]; ld.global.f32 %f12, [%rd5+4]; ld.global.f32 %f13, [%rd5+8]; ld.global.f32 %f14, [%rd4]; ld.global.f32 %f15, [%rd4+4]; ld.global.f32 %f16, [%rd4+8]; fma.rn.ftz.f32 %f863, %f13, %f10, %f16; fma.rn.ftz.f32 %f864, %f12, %f10, %f15; fma.rn.ftz.f32 %f865, %f11, %f10, %f14; neg.ftz.f32 %f866, %f11; ld.global.u32 %r379, [%rd6]; st.local.v4.f32 [%rd71], {%f865, %f864, %f863, %f866}; neg.ftz.f32 %f867, %f13; neg.ftz.f32 %f868, %f12; st.local.v4.f32 [%rd64], {%f868, %f867, %f866, %f868}; st.local.v4.f32 [%rd19], {%f867, %f866, %f868, %f867}; mov.u32 %r874, 0; mov.u32 %r381, -1; st.local.v4.u32 [%rd18], {%r381, %r874, %r874, %r381}; mov.u32 %r382, 16; st.local.v4.u32 [%rd49], {%r382, %r874, %r874, %r381}; add.s64 %rd253, %rd166, 128; cvta.to.local.u64 %rd254, %rd253; mov.f32 %f2956, 0f00000000; st.local.v2.f32 [%rd254], {%f2956, %f2956}; add.s64 %rd255, %rd166, 136; cvta.to.local.u64 %rd256, %rd255; st.local.u32 [%rd256], %r874; st.local.v4.u32 [%rd34], {%r381, %r379, %r874, %r874}; add.s64 %rd259, %rd166, 96; cvta.to.local.u64 %rd260, %rd259; st.local.v4.f32 [%rd260], {%f2956, %f2956, %f2956, %f2956}; add.s64 %rd261, %rd166, 112; cvta.to.local.u64 %rd262, %rd261; st.local.v4.f32 [%rd262], {%f2956, %f2956, %f2956, %f2956}; add.s64 %rd263, %rd166, 240; cvta.to.local.u64 %rd264, %rd263; st.local.v2.f32 [%rd264], {%f14, %f15}; add.s64 %rd265, %rd166, 248; cvta.to.local.u64 %rd266, %rd265; st.local.f32 [%rd266], %f16; ld.global.u32 %r873, [%rd3]; setp.eq.s32 %p13, %r873, -1; mov.f32 %f2952, 0f7F7FFFFF; @%p13 bra $L__BB211_18; mov.u64 %rd1167, kernel_params; add.s64 %rd1193, %rd1167, 3744; mov.f32 %f2952, 0f7F7FFFFF; $L__BB211_10: shl.b64 %rd1168, %rd2, 2; ld.const.u64 %rd269, [%rd1193]; cvta.to.global.u64 %rd270, %rd269; add.s64 %rd272, %rd270, %rd1168; ld.global.s32 %rd74, [%rd272]; shl.b32 %r384, %r873, 5; cvt.u64.u32 %rd273, %r384; and.b64 %rd274, %rd273, 134217696; add.s64 %rd275, %rd16, %rd274; ld.global.u32 %r18, [%rd275+16]; and.b32 %r385, %r18, 1048576; setp.ne.s32 %p14, %r385, 0; mov.u16 %rs45, 1; @%p14 bra $L__BB211_13; cvt.u32.u64 %r386, %rd74; and.b32 %r387, %r18, 268435456; setp.eq.s32 %p15, %r387, 0; setp.eq.s32 %p16, %r386, -1; or.pred %p17, %p16, %p15; mov.u16 %rs45, 0; @%p17 bra $L__BB211_13; shl.b64 %rd276, %rd74, 2; add.s64 %rd277, %rd10, %rd276; ld.global.u8 %rs26, [%rd277+1]; and.b16 %rs45, %rs26, 1; $L__BB211_13: setp.eq.s16 %p18, %rs45, 0; @%p18 bra $L__BB211_17; ld.const.f32 %f2951, [kernel_params+1032]; cvt.u32.u64 %r388, %rd74; setp.eq.s32 %p19, %r388, -1; @%p19 bra $L__BB211_16; shl.b64 %rd278, %rd74, 2; add.s64 %rd279, %rd17, %rd278; ld.global.f32 %f2951, [%rd279]; $L__BB211_16: ld.const.f32 %f2938, [kernel_params+1616]; mul.ftz.f32 %f871, %f2951, %f2938; min.ftz.f32 %f2952, %f871, %f2952; $L__BB211_17: shl.b64 %rd1170, %rd2, 2; mov.u64 %rd1169, kernel_params; add.s32 %r874, %r874, 1; mul.wide.s32 %rd280, %r874, 16; add.s64 %rd282, %rd1169, %rd280; add.s64 %rd1193, %rd282, 3744; ld.const.u64 %rd283, [%rd282+3752]; cvta.to.global.u64 %rd284, %rd283; add.s64 %rd286, %rd284, %rd1170; ld.global.u32 %r873, [%rd286]; setp.ne.s32 %p20, %r873, -1; @%p20 bra $L__BB211_10; $L__BB211_18: ld.global.u32 %r21, [%rd13]; ld.global.u16 %rs6, [%rd14]; ld.global.u32 %r22, [%rd15]; sub.ftz.f32 %f2955, %f9, %f10; setp.eq.ftz.f32 %p21, %f2952, 0f7F7FFFFF; mov.u32 %r875, 1; @%p21 bra $L__BB211_25; bra.uni $L__BB211_19; $L__BB211_25: setp.lt.s32 %p26, %r875, 1; mov.f32 %f3017, %f6; mov.f32 %f3018, %f7; mov.f32 %f3019, %f8; @%p26 bra $L__BB211_340; mov.f32 %f3013, 0f00000000; mov.u32 %r876, 0; mov.f32 %f2957, %f10; mov.f32 %f3017, %f6; mov.f32 %f3018, %f7; mov.f32 %f3019, %f8; mov.f32 %f3012, %f3013; mov.f32 %f3011, %f3013; $L__BB211_27: mov.u32 %r857, 0; cvt.rn.f32.s32 %f884, %r876; add.ftz.f32 %f885, %f884, 0f3F800000; fma.rn.ftz.f32 %f886, %f2955, %f885, %f10; min.ftz.f32 %f42, %f9, %f886; sub.ftz.f32 %f43, %f42, %f2957; fma.rn.ftz.f32 %f887, %f2956, %f43, %f2957; fma.rn.ftz.f32 %f888, %f13, %f887, %f16; fma.rn.ftz.f32 %f889, %f12, %f887, %f15; fma.rn.ftz.f32 %f890, %f11, %f887, %f14; st.local.v2.f32 [%rd71], {%f890, %f889}; st.local.f32 [%rd28], %f888; st.local.u32 [%rd29], %r857; st.local.u32 [%rd30], %r857; st.local.u32 [%rd31], %r857; st.local.u32 [%rd32], %r857; ld.global.u32 %r877, [%rd3]; setp.eq.s32 %p27, %r877, -1; @%p27 bra $L__BB211_330; mov.u64 %rd1171, kernel_params; add.s64 %rd1194, %rd1171, 3744; mov.u32 %r878, 0; $L__BB211_29: shl.b64 %rd1166, %rd2, 2; mov.u32 %r850, -1; ld.const.u64 %rd292, [%rd1194]; cvta.to.global.u64 %rd293, %rd292; add.s64 %rd295, %rd293, %rd1166; ld.global.u32 %r31, [%rd295]; st.local.u32 [%rd33], %r31; st.local.u32 [%rd34], %r850; ld.local.u32 %r484, [%rd31]; and.b32 %r485, %r484, -2147418113; and.b32 %r890, %r877, 4194303; shl.b32 %r486, %r877, 5; cvt.u64.u32 %rd296, %r486; and.b64 %rd297, %rd296, 134217696; add.s64 %rd298, %rd16, %rd297; ld.global.u32 %r33, [%rd298+16]; or.b32 %r487, %r485, %r33; st.local.v2.u32 [%rd18], {%r877, %r487}; ld.local.u32 %r488, [%rd32]; and.b32 %r34, %r488, -4032; st.local.u32 [%rd32], %r34; setp.eq.s32 %p28, %r31, -1; @%p28 bra $L__BB211_116; mul.wide.s32 %rd302, %r31, 4; add.s64 %rd303, %rd10, %rd302; ld.global.u32 %r35, [%rd303]; or.b32 %r489, %r34, %r35; st.local.u32 [%rd32], %r489; and.b32 %r36, %r35, 2; setp.eq.s32 %p29, %r36, 0; mul.wide.s32 %rd304, %r31, 208; add.s64 %rd305, %rd12, %rd304; add.s64 %rd78, %rd305, 48; @%p29 bra $L__BB211_37; ld.local.f32 %f891, [%rd43]; ld.global.u32 %rd306, [%rd78+120]; ld.global.u32 %r490, [%rd78+104]; shl.b32 %r491, %r490, 1; cvt.rn.f32.s32 %f892, %r491; mul.ftz.f32 %f893, %f891, %f892; cvt.rzi.ftz.s32.f32 %r492, %f893; add.s32 %r493, %r491, -1; min.s32 %r494, %r492, %r493; cvt.rn.f32.s32 %f894, %r494; sub.ftz.f32 %f44, %f893, %f894; cvt.s64.s32 %rd307, %r494; add.s64 %rd308, %rd307, %rd306; cvta.to.global.u64 %rd309, %rd168; shl.b64 %rd310, %rd308, 6; add.s64 %rd79, %rd309, %rd310; ld.global.v4.f32 {%f895, %f896, %f897, %f898}, [%rd79]; ld.global.v4.f32 {%f899, %f900, %f901, %f902}, [%rd79+64]; mul.ftz.f32 %f903, %f896, %f900; fma.rn.ftz.f32 %f904, %f895, %f899, %f903; mul.ftz.f32 %f905, %f898, %f902; fma.rn.ftz.f32 %f906, %f897, %f901, %f905; add.ftz.f32 %f53, %f904, %f906; setp.gt.ftz.f32 %p30, %f53, 0f3F7FDF3B; @%p30 bra $L__BB211_33; bra.uni $L__BB211_32; $L__BB211_33: mov.f32 %f2943, 0f3F800000; sub.ftz.f32 %f2942, %f2943, %f44; mul.ftz.f32 %f961, %f895, %f2942; fma.rn.ftz.f32 %f962, %f899, %f44, %f961; mul.ftz.f32 %f963, %f896, %f2942; fma.rn.ftz.f32 %f964, %f900, %f44, %f963; mul.ftz.f32 %f965, %f897, %f2942; fma.rn.ftz.f32 %f966, %f901, %f44, %f965; mul.ftz.f32 %f967, %f898, %f2942; fma.rn.ftz.f32 %f968, %f902, %f44, %f967; mul.ftz.f32 %f969, %f964, %f964; fma.rn.ftz.f32 %f970, %f962, %f962, %f969; mul.ftz.f32 %f971, %f968, %f968; fma.rn.ftz.f32 %f972, %f966, %f966, %f971; add.ftz.f32 %f973, %f970, %f972; rsqrt.approx.ftz.f32 %f974, %f973; mul.ftz.f32 %f2967, %f968, %f974; mul.ftz.f32 %f2966, %f966, %f974; mul.ftz.f32 %f2965, %f964, %f974; mul.ftz.f32 %f2964, %f962, %f974; bra.uni $L__BB211_34; $L__BB211_32: mov.f32 %f2939, 0f3F800000; mov.f32 %f908, 0fBF800000; max.ftz.f32 %f909, %f53, %f908; min.ftz.f32 %f911, %f909, %f2939; abs.ftz.f32 %f912, %f911; sub.ftz.f32 %f913, %f2939, %f912; mul.ftz.f32 %f914, %f913, 0f3F000000; sqrt.approx.ftz.f32 %f915, %f914; setp.gt.ftz.f32 %p31, %f912, 0f3F11EB85; selp.f32 %f916, %f915, %f912, %p31; mul.ftz.f32 %f917, %f916, %f916; mov.f32 %f918, 0f3C94D2E9; mov.f32 %f919, 0f3D53F941; fma.rn.ftz.f32 %f920, %f919, %f917, %f918; mov.f32 %f921, 0f3D3F841F; fma.rn.ftz.f32 %f922, %f920, %f917, %f921; mov.f32 %f923, 0f3D994929; fma.rn.ftz.f32 %f924, %f922, %f917, %f923; mov.f32 %f925, 0f3E2AAB94; fma.rn.ftz.f32 %f926, %f924, %f917, %f925; mul.ftz.f32 %f927, %f917, %f926; fma.rn.ftz.f32 %f928, %f927, %f916, %f916; add.ftz.f32 %f929, %f928, %f928; mov.f32 %f930, 0f3FC90FDB; sub.ftz.f32 %f931, %f930, %f928; selp.f32 %f932, %f929, %f931, %p31; setp.lt.ftz.f32 %p32, %f911, 0f00000000; mov.f32 %f933, 0f40490FDB; sub.ftz.f32 %f934, %f933, %f932; selp.f32 %f935, %f934, %f932, %p32; mul.ftz.f32 %f936, %f895, %f53; sub.ftz.f32 %f937, %f899, %f936; mul.ftz.f32 %f938, %f896, %f53; sub.ftz.f32 %f939, %f900, %f938; mul.ftz.f32 %f940, %f897, %f53; sub.ftz.f32 %f941, %f901, %f940; mul.ftz.f32 %f942, %f898, %f53; sub.ftz.f32 %f943, %f902, %f942; mul.ftz.f32 %f944, %f939, %f939; fma.rn.ftz.f32 %f945, %f937, %f937, %f944; mul.ftz.f32 %f946, %f943, %f943; fma.rn.ftz.f32 %f947, %f941, %f941, %f946; add.ftz.f32 %f948, %f945, %f947; rsqrt.approx.ftz.f32 %f949, %f948; mul.ftz.f32 %f950, %f937, %f949; mul.ftz.f32 %f951, %f939, %f949; mul.ftz.f32 %f952, %f941, %f949; mul.ftz.f32 %f953, %f943, %f949; mul.ftz.f32 %f954, %f44, %f935; cos.approx.ftz.f32 %f955, %f954; sin.approx.ftz.f32 %f956, %f954; mul.ftz.f32 %f957, %f950, %f956; mul.ftz.f32 %f958, %f951, %f956; mul.ftz.f32 %f959, %f952, %f956; mul.ftz.f32 %f960, %f953, %f956; fma.rn.ftz.f32 %f2967, %f898, %f955, %f960; fma.rn.ftz.f32 %f2966, %f897, %f955, %f959; fma.rn.ftz.f32 %f2965, %f896, %f955, %f958; fma.rn.ftz.f32 %f2964, %f895, %f955, %f957; $L__BB211_34: mov.f32 %f2941, 0f3F800000; sub.ftz.f32 %f2940, %f2941, %f44; ld.global.v4.f32 {%f975, %f976, %f977, %f978}, [%rd79+16]; ld.global.v4.f32 {%f983, %f984, %f985, %f986}, [%rd79+80]; mul.ftz.f32 %f991, %f44, %f983; mul.ftz.f32 %f992, %f44, %f984; mul.ftz.f32 %f993, %f44, %f985; mul.ftz.f32 %f994, %f44, %f986; fma.rn.ftz.f32 %f67, %f2940, %f975, %f991; fma.rn.ftz.f32 %f68, %f2940, %f976, %f992; fma.rn.ftz.f32 %f69, %f2940, %f977, %f993; fma.rn.ftz.f32 %f995, %f2940, %f978, %f994; ld.global.v4.f32 {%f996, %f997, %f998, %f999}, [%rd79+32]; ld.global.v4.f32 {%f1004, %f1005, %f1006, %f1007}, [%rd79+96]; mul.ftz.f32 %f1012, %f44, %f1004; mul.ftz.f32 %f1013, %f44, %f1005; mul.ftz.f32 %f1014, %f44, %f1006; mul.ftz.f32 %f1015, %f44, %f1007; fma.rn.ftz.f32 %f1016, %f2940, %f996, %f1012; fma.rn.ftz.f32 %f1017, %f2940, %f997, %f1013; fma.rn.ftz.f32 %f1018, %f2940, %f998, %f1014; fma.rn.ftz.f32 %f1019, %f2940, %f999, %f1015; ld.global.v4.f32 {%f1020, %f1021, %f1022, %f1023}, [%rd79+48]; ld.global.v4.f32 {%f1028, %f1029, %f1030, %f1031}, [%rd79+112]; mul.ftz.f32 %f1036, %f44, %f1028; mul.ftz.f32 %f1037, %f44, %f1029; mul.ftz.f32 %f1038, %f44, %f1030; mul.ftz.f32 %f1039, %f44, %f1031; fma.rn.ftz.f32 %f1040, %f2940, %f1020, %f1036; fma.rn.ftz.f32 %f1041, %f2940, %f1021, %f1037; fma.rn.ftz.f32 %f1042, %f2940, %f1022, %f1038; fma.rn.ftz.f32 %f1043, %f2940, %f1023, %f1039; mul.ftz.f32 %f1044, %f2967, 0f3FB504F3; mul.ftz.f32 %f1045, %f2964, 0f3FB504F3; mul.ftz.f32 %f1046, %f2965, 0f3FB504F3; mul.ftz.f32 %f1047, %f2966, 0f3FB504F3; mul.ftz.f32 %f1048, %f1044, %f1045; mul.ftz.f32 %f1049, %f1044, %f1046; mul.ftz.f32 %f1050, %f1044, %f1047; mul.ftz.f32 %f1051, %f1045, %f1045; mul.ftz.f32 %f1052, %f1045, %f1046; mul.ftz.f32 %f1053, %f1045, %f1047; mul.ftz.f32 %f1054, %f1046, %f1046; mul.ftz.f32 %f1055, %f1046, %f1047; mul.ftz.f32 %f1056, %f1047, %f1047; sub.ftz.f32 %f1058, %f2941, %f1054; sub.ftz.f32 %f1059, %f1058, %f1056; sub.ftz.f32 %f1060, %f1052, %f1050; add.ftz.f32 %f1061, %f1049, %f1053; add.ftz.f32 %f1062, %f1050, %f1052; sub.ftz.f32 %f1063, %f2941, %f1051; sub.ftz.f32 %f1064, %f1063, %f1056; sub.ftz.f32 %f1065, %f1055, %f1048; sub.ftz.f32 %f1066, %f1053, %f1049; add.ftz.f32 %f1067, %f1048, %f1055; sub.ftz.f32 %f1068, %f1063, %f1054; mul.ftz.f32 %f1069, %f1060, %f1018; fma.rn.ftz.f32 %f1070, %f1059, %f995, %f1069; fma.rn.ftz.f32 %f2973, %f1061, %f1041, %f1070; mul.ftz.f32 %f1071, %f1060, %f1019; fma.rn.ftz.f32 %f1072, %f1059, %f1016, %f1071; fma.rn.ftz.f32 %f71, %f1061, %f1042, %f1072; mul.ftz.f32 %f1073, %f1060, %f1040; fma.rn.ftz.f32 %f1074, %f1059, %f1017, %f1073; fma.rn.ftz.f32 %f72, %f1061, %f1043, %f1074; mul.ftz.f32 %f1075, %f1064, %f1018; fma.rn.ftz.f32 %f1076, %f1062, %f995, %f1075; fma.rn.ftz.f32 %f73, %f1065, %f1041, %f1076; mul.ftz.f32 %f1077, %f1064, %f1019; fma.rn.ftz.f32 %f1078, %f1062, %f1016, %f1077; fma.rn.ftz.f32 %f2972, %f1065, %f1042, %f1078; mul.ftz.f32 %f1079, %f1064, %f1040; fma.rn.ftz.f32 %f1080, %f1062, %f1017, %f1079; fma.rn.ftz.f32 %f75, %f1065, %f1043, %f1080; mul.ftz.f32 %f1081, %f1067, %f1018; fma.rn.ftz.f32 %f1082, %f1066, %f995, %f1081; fma.rn.ftz.f32 %f76, %f1068, %f1041, %f1082; mul.ftz.f32 %f1083, %f1067, %f1019; fma.rn.ftz.f32 %f1084, %f1066, %f1016, %f1083; fma.rn.ftz.f32 %f77, %f1068, %f1042, %f1084; mul.ftz.f32 %f1085, %f1067, %f1040; fma.rn.ftz.f32 %f1086, %f1066, %f1017, %f1085; fma.rn.ftz.f32 %f2971, %f1068, %f1043, %f1086; cvta.to.local.u64 %rd311, %rd44; st.local.v4.f32 [%rd311], {%f2973, %f71, %f72, %f67}; cvta.to.local.u64 %rd312, %rd47; st.local.v4.f32 [%rd312], {%f73, %f2972, %f75, %f68}; cvta.to.local.u64 %rd313, %rd46; st.local.v4.f32 [%rd313], {%f76, %f77, %f2971, %f69}; mul.ftz.f32 %f1087, %f2972, %f2971; mul.ftz.f32 %f79, %f77, %f75; sub.ftz.f32 %f2970, %f1087, %f79; mul.ftz.f32 %f81, %f77, %f72; mul.ftz.f32 %f1088, %f71, %f2971; sub.ftz.f32 %f2969, %f81, %f1088; mul.ftz.f32 %f83, %f71, %f75; mul.ftz.f32 %f1089, %f2972, %f72; sub.ftz.f32 %f2968, %f83, %f1089; mul.ftz.f32 %f1090, %f73, %f2969; fma.rn.ftz.f32 %f1091, %f2973, %f2970, %f1090; fma.rn.ftz.f32 %f2974, %f76, %f2968, %f1091; setp.neu.ftz.f32 %p33, %f2974, 0f00000000; @%p33 bra $L__BB211_36; add.ftz.f32 %f2973, %f2973, 0f322BCC77; add.ftz.f32 %f2971, %f2971, 0f322BCC77; add.ftz.f32 %f2972, %f2972, 0f322BCC77; mul.ftz.f32 %f1092, %f2972, %f2971; sub.ftz.f32 %f2970, %f1092, %f79; mul.ftz.f32 %f1093, %f71, %f2971; sub.ftz.f32 %f2969, %f81, %f1093; mul.ftz.f32 %f1094, %f2972, %f72; sub.ftz.f32 %f2968, %f83, %f1094; mul.ftz.f32 %f1095, %f73, %f2969; fma.rn.ftz.f32 %f1096, %f2973, %f2970, %f1095; fma.rn.ftz.f32 %f1097, %f76, %f2968, %f1096; setp.eq.ftz.f32 %p34, %f1097, 0f00000000; selp.f32 %f2974, 0f7F7FFFFF, %f1097, %p34; $L__BB211_36: mul.ftz.f32 %f1098, %f73, %f2971; mul.ftz.f32 %f1099, %f76, %f75; sub.ftz.f32 %f1100, %f1099, %f1098; mul.ftz.f32 %f1101, %f76, %f72; mul.ftz.f32 %f1102, %f2971, %f2973; sub.ftz.f32 %f1103, %f1102, %f1101; mul.ftz.f32 %f1104, %f75, %f2973; mul.ftz.f32 %f1105, %f73, %f72; sub.ftz.f32 %f1106, %f1105, %f1104; mul.ftz.f32 %f1107, %f76, %f2972; mul.ftz.f32 %f1108, %f73, %f77; sub.ftz.f32 %f1109, %f1108, %f1107; mul.ftz.f32 %f1110, %f77, %f2973; mul.ftz.f32 %f1111, %f76, %f71; sub.ftz.f32 %f1112, %f1111, %f1110; mul.ftz.f32 %f1113, %f73, %f71; mul.ftz.f32 %f1114, %f2972, %f2973; sub.ftz.f32 %f1115, %f1114, %f1113; div.approx.ftz.f32 %f1116, %f2969, %f2974; mul.ftz.f32 %f1117, %f68, %f1116; div.approx.ftz.f32 %f1118, %f2970, %f2974; fma.rn.ftz.f32 %f1119, %f67, %f1118, %f1117; div.approx.ftz.f32 %f1120, %f2968, %f2974; fma.rn.ftz.f32 %f1121, %f69, %f1120, %f1119; div.approx.ftz.f32 %f1122, %f1103, %f2974; mul.ftz.f32 %f1123, %f68, %f1122; div.approx.ftz.f32 %f1124, %f1100, %f2974; fma.rn.ftz.f32 %f1125, %f67, %f1124, %f1123; div.approx.ftz.f32 %f1126, %f1106, %f2974; fma.rn.ftz.f32 %f1127, %f69, %f1126, %f1125; div.approx.ftz.f32 %f1128, %f1112, %f2974; mul.ftz.f32 %f1129, %f68, %f1128; div.approx.ftz.f32 %f1130, %f1109, %f2974; fma.rn.ftz.f32 %f1131, %f67, %f1130, %f1129; div.approx.ftz.f32 %f1132, %f1115, %f2974; fma.rn.ftz.f32 %f1133, %f69, %f1132, %f1131; neg.ftz.f32 %f1134, %f1121; cvta.to.local.u64 %rd314, %rd48; st.local.v4.f32 [%rd314], {%f1118, %f1116, %f1120, %f1134}; neg.ftz.f32 %f1135, %f1127; cvta.to.local.u64 %rd315, %rd51; st.local.v4.f32 [%rd315], {%f1124, %f1122, %f1126, %f1135}; neg.ftz.f32 %f1136, %f1133; cvta.to.local.u64 %rd316, %rd50; st.local.v4.f32 [%rd316], {%f1130, %f1128, %f1132, %f1136}; $L__BB211_37: and.b32 %r495, %r35, 2048; setp.eq.s32 %p35, %r495, 0; @%p35 bra $L__BB211_116; ld.local.s32 %rd80, [%rd45]; ld.global.u32 %r37, [%rd78+116]; ld.local.u32 %r38, [%rd49]; and.b32 %r497, %r38, 1; setp.eq.b32 %p36, %r497, 1; xor.pred %p38, %p36, %p222; not.pred %p39, %p38; mov.u32 %r880, 0; @%p39 bra $L__BB211_42; mov.u32 %r879, -1; cvt.u32.u64 %r499, %rd80; setp.eq.s32 %p40, %r499, -1; @%p40 bra $L__BB211_41; shl.b64 %rd317, %rd80, 2; add.s64 %rd318, %rd53, %rd317; ld.global.u32 %r879, [%rd318]; $L__BB211_41: setp.ne.s32 %p41, %r879, -1; selp.u32 %r880, 1, 0, %p41; $L__BB211_42: add.s32 %r882, %r880, %r37; mul.wide.u32 %rd319, %r882, 16; add.s64 %rd320, %rd52, %rd319; ld.global.u16 %rs46, [%rd320+12]; ld.global.u32 %r881, [%rd320+8]; ld.global.u64 %rd1195, [%rd320]; setp.eq.s64 %p42, %rd1195, 25; @%p42 bra $L__BB211_47; $L__BB211_43: setp.eq.s64 %p43, %rd1195, 0; @%p43 bra $L__BB211_45; add.s32 %r882, %r882, 2; bra.uni $L__BB211_46; $L__BB211_45: setp.eq.s16 %p44, %rs46, 0; mov.u32 %r886, -1; mov.u32 %r885, 0; mov.u32 %r882, %r881; @%p44 bra $L__BB211_51; $L__BB211_46: mul.wide.u32 %rd321, %r882, 16; add.s64 %rd322, %rd52, %rd321; ld.global.u16 %rs46, [%rd322+12]; ld.global.u32 %r881, [%rd322+8]; ld.global.u64 %rd1195, [%rd322]; setp.ne.s64 %p45, %rd1195, 25; @%p45 bra $L__BB211_43; $L__BB211_47: cvt.u32.u64 %r504, %rd80; setp.ne.s32 %p46, %r504, -1; @%p46 bra $L__BB211_50; mov.u32 %r886, -1; mov.u32 %r885, 0; add.s16 %rs29, %rs46, -1; setp.lt.u16 %p47, %rs29, 2; @%p47 bra $L__BB211_50; setp.ne.s16 %p48, %rs46, 1024; @%p48 bra $L__BB211_51; $L__BB211_50: setp.eq.s16 %p49, %rs46, 0; selp.b32 %r886, -1, %r881, %p49; cvt.u32.u16 %r885, %rs46; $L__BB211_51: ld.local.v4.f32 {%f1140, %f1141, %f1142, %f1143}, [%rd71]; ld.global.f32 %f103, [%rd78+156]; ld.local.f32 %f104, [%rd43]; setp.ne.s32 %p50, %r38, 16; mov.f32 %f2981, 0f00000000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; @%p50 bra $L__BB211_80; and.b32 %r511, %r885, 3; setp.eq.s32 %p51, %r511, 0; @%p51 bra $L__BB211_54; mul.wide.s32 %rd325, %r886, 16; add.s64 %rd326, %rd54, %rd325; ld.global.v4.f32 {%f2981, %f2982, %f2983, %f2984}, [%rd326]; bra.uni $L__BB211_78; $L__BB211_54: setp.ne.s32 %p52, %r885, 1024; mov.f32 %f2981, 0f00000000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; mov.f32 %f2984, %f2981; @%p52 bra $L__BB211_78; and.b32 %r849, %r35, 2; setp.eq.s32 %p531, %r849, 0; @%p531 bra $L__BB211_57; bra.uni $L__BB211_56; $L__BB211_57: ld.global.v4.f32 {%f1185, %f1186, %f1187, %f1188}, [%rd78]; ld.global.v4.f32 {%f1193, %f1194, %f1195, %f1196}, [%rd78+16]; ld.global.v4.f32 {%f1201, %f1202, %f1203, %f1204}, [%rd78+32]; mul.ftz.f32 %f1209, %f1141, %f1186; fma.rn.ftz.f32 %f1210, %f1140, %f1185, %f1209; fma.rn.ftz.f32 %f1211, %f1142, %f1187, %f1210; add.ftz.f32 %f121, %f1188, %f1211; mul.ftz.f32 %f1212, %f1141, %f1194; fma.rn.ftz.f32 %f1213, %f1140, %f1193, %f1212; fma.rn.ftz.f32 %f1214, %f1142, %f1195, %f1213; add.ftz.f32 %f122, %f1196, %f1214; mul.ftz.f32 %f1215, %f1141, %f1202; fma.rn.ftz.f32 %f1216, %f1140, %f1201, %f1215; fma.rn.ftz.f32 %f1217, %f1142, %f1203, %f1216; add.ftz.f32 %f123, %f1204, %f1217; bra.uni $L__BB211_58; $L__BB211_56: cvta.to.local.u64 %rd327, %rd48; ld.local.v4.f32 {%f1152, %f1153, %f1154, %f1155}, [%rd327]; mul.ftz.f32 %f1160, %f1141, %f1153; fma.rn.ftz.f32 %f1161, %f1140, %f1152, %f1160; fma.rn.ftz.f32 %f1162, %f1142, %f1154, %f1161; add.ftz.f32 %f121, %f1155, %f1162; cvta.to.local.u64 %rd328, %rd51; ld.local.v4.f32 {%f1163, %f1164, %f1165, %f1166}, [%rd328]; mul.ftz.f32 %f1171, %f1141, %f1164; fma.rn.ftz.f32 %f1172, %f1140, %f1163, %f1171; fma.rn.ftz.f32 %f1173, %f1142, %f1165, %f1172; add.ftz.f32 %f122, %f1166, %f1173; cvta.to.local.u64 %rd329, %rd50; ld.local.v4.f32 {%f1174, %f1175, %f1176, %f1177}, [%rd329]; mul.ftz.f32 %f1182, %f1141, %f1175; fma.rn.ftz.f32 %f1183, %f1140, %f1174, %f1182; fma.rn.ftz.f32 %f1184, %f1142, %f1176, %f1183; add.ftz.f32 %f123, %f1177, %f1184; $L__BB211_58: mul.wide.s32 %rd330, %r886, 96; add.s64 %rd91, %rd55, %rd330; add.s64 %rd331, %rd56, %rd330; add.s64 %rd92, %rd331, 32; ld.global.u32 %r512, [%rd331+32]; setp.eq.s32 %p54, %r512, 0; @%p54 bra $L__BB211_60; ld.global.v4.f32 {%f1218, %f1219, %f1220, %f1221}, [%rd92+16]; mul.ftz.f32 %f1226, %f122, %f1219; fma.rn.ftz.f32 %f1227, %f121, %f1218, %f1226; fma.rn.ftz.f32 %f1228, %f123, %f1220, %f1227; add.ftz.f32 %f118, %f1221, %f1228; ld.global.v4.f32 {%f1229, %f1230, %f1231, %f1232}, [%rd92+32]; mul.ftz.f32 %f1237, %f122, %f1230; fma.rn.ftz.f32 %f1238, %f121, %f1229, %f1237; fma.rn.ftz.f32 %f1239, %f123, %f1231, %f1238; add.ftz.f32 %f119, %f1232, %f1239; ld.global.v4.f32 {%f1240, %f1241, %f1242, %f1243}, [%rd92+48]; mul.ftz.f32 %f1248, %f122, %f1241; fma.rn.ftz.f32 %f1249, %f121, %f1240, %f1248; fma.rn.ftz.f32 %f1250, %f123, %f1242, %f1249; add.ftz.f32 %f123, %f1243, %f1250; mov.f32 %f121, %f118; mov.f32 %f122, %f119; $L__BB211_60: and.b32 %r513, %r33, 16777216; setp.ne.s32 %p55, %r513, 0; setp.eq.s32 %p56, %r513, 0; selp.b32 %r63, -1, 2, %p56; @%p55 bra $L__BB211_62; ld.global.u32 %r63, [%rd92+-20]; $L__BB211_62: ld.global.u32 %r64, [%rd92+-24]; setp.eq.s32 %p57, %r64, 8; @%p57 bra $L__BB211_77; bra.uni $L__BB211_63; $L__BB211_77: { // callseq 1198, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2981, [retval0+0]; } // callseq 1198 mov.f32 %f2984, 0f3F800000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; bra.uni $L__BB211_78; $L__BB211_63: setp.eq.s32 %p58, %r64, 9; @%p58 bra $L__BB211_76; bra.uni $L__BB211_64; $L__BB211_76: { // callseq 1197, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2981, [retval0+0]; ld.param.f32 %f2982, [retval0+4]; ld.param.f32 %f2983, [retval0+8]; } // callseq 1197 mov.f32 %f2984, 0f3F800000; bra.uni $L__BB211_78; $L__BB211_64: setp.eq.s32 %p59, %r64, 10; @%p59 bra $L__BB211_75; bra.uni $L__BB211_65; $L__BB211_75: { // callseq 1196, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2981, [retval0+0]; } // callseq 1196 mov.f32 %f2984, 0f3F800000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; bra.uni $L__BB211_78; $L__BB211_65: setp.eq.s32 %p60, %r64, 11; @%p60 bra $L__BB211_74; bra.uni $L__BB211_66; $L__BB211_74: { // callseq 1195, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .b32 param4; st.param.b32 [param4+0], %r63; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2981, [retval0+0]; } // callseq 1195 mov.f32 %f2984, 0f3F800000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; bra.uni $L__BB211_78; $L__BB211_66: or.b32 %r65, %r63, 1; setp.gt.u32 %p61, %r64, 6; @%p61 bra $L__BB211_71; mov.u32 %r515, 1; shl.b32 %r516, %r515, %r64; and.b32 %r517, %r516, 71; setp.ne.s32 %p62, %r517, 0; @%p62 bra $L__BB211_68; bra.uni $L__BB211_71; $L__BB211_68: setp.eq.s32 %p63, %r65, 3; @%p63 bra $L__BB211_70; bra.uni $L__BB211_69; $L__BB211_70: { // callseq 1193, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f2981, [retval0+0]; ld.param.f32 %f2982, [retval0+4]; ld.param.f32 %f2983, [retval0+8]; ld.param.f32 %f2984, [retval0+12]; } // callseq 1193 bra.uni $L__BB211_78; $L__BB211_69: ld.global.u64 %rd332, [%rd92+-32]; tex.3d.v4.f32.f32 {%f2981, %f2982, %f2983, %f2984}, [%rd332, {%f121, %f122, %f123, %f123}]; bra.uni $L__BB211_78; $L__BB211_71: setp.eq.s32 %p64, %r65, 3; @%p64 bra $L__BB211_73; bra.uni $L__BB211_72; $L__BB211_73: { // callseq 1194, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd91; .param .b32 param1; st.param.f32 [param1+0], %f121; .param .b32 param2; st.param.f32 [param2+0], %f122; .param .b32 param3; st.param.f32 [param3+0], %f123; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f2981, [retval0+0]; } // callseq 1194 mov.f32 %f2984, 0f3F800000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; bra.uni $L__BB211_78; $L__BB211_72: ld.global.u64 %rd333, [%rd92+-32]; tex.3d.v4.f32.f32 {%f2981, %f1251, %f1252, %f1253}, [%rd333, {%f121, %f122, %f123, %f123}]; mov.f32 %f2984, 0f3F800000; mov.f32 %f2982, %f2981; mov.f32 %f2983, %f2981; $L__BB211_78: setp.leu.ftz.f32 %p65, %f2984, 0f358637BD; setp.eq.ftz.f32 %p66, %f2984, 0f3F800000; or.pred %p67, %p65, %p66; @%p67 bra $L__BB211_80; div.approx.ftz.f32 %f2983, %f2983, %f2984; div.approx.ftz.f32 %f2982, %f2982, %f2984; div.approx.ftz.f32 %f2981, %f2981, %f2984; $L__BB211_80: ld.local.u32 %r518, [%rd32]; and.b32 %r66, %r518, 2; setp.eq.s32 %p68, %r66, 0; @%p68 bra $L__BB211_82; bra.uni $L__BB211_81; $L__BB211_82: ld.local.u32 %r519, [%rd33]; mul.wide.s32 %rd337, %r519, 208; add.s64 %rd338, %rd12, %rd337; ld.global.v4.f32 {%f1281, %f1282, %f1283, %f1284}, [%rd338]; ld.global.v4.f32 {%f1288, %f1289, %f1290, %f1291}, [%rd338+16]; ld.global.v4.f32 {%f1295, %f1296, %f1297, %f1298}, [%rd338+32]; mul.ftz.f32 %f1302, %f2982, %f1282; fma.rn.ftz.f32 %f1303, %f2981, %f1281, %f1302; fma.rn.ftz.f32 %f2988, %f2983, %f1283, %f1303; mul.ftz.f32 %f1304, %f2982, %f1289; fma.rn.ftz.f32 %f1305, %f2981, %f1288, %f1304; fma.rn.ftz.f32 %f2989, %f2983, %f1290, %f1305; mul.ftz.f32 %f1306, %f2982, %f1296; fma.rn.ftz.f32 %f1307, %f2981, %f1295, %f1306; fma.rn.ftz.f32 %f2990, %f2983, %f1297, %f1307; bra.uni $L__BB211_83; $L__BB211_81: cvta.to.local.u64 %rd334, %rd44; ld.local.v4.f32 {%f1254, %f1255, %f1256, %f1257}, [%rd334]; mul.ftz.f32 %f1261, %f2982, %f1255; fma.rn.ftz.f32 %f1262, %f2981, %f1254, %f1261; fma.rn.ftz.f32 %f2988, %f2983, %f1256, %f1262; cvta.to.local.u64 %rd335, %rd47; ld.local.v4.f32 {%f1263, %f1264, %f1265, %f1266}, [%rd335]; mul.ftz.f32 %f1270, %f2982, %f1264; fma.rn.ftz.f32 %f1271, %f2981, %f1263, %f1270; fma.rn.ftz.f32 %f2989, %f2983, %f1265, %f1271; cvta.to.local.u64 %rd336, %rd46; ld.local.v4.f32 {%f1272, %f1273, %f1274, %f1275}, [%rd336]; mul.ftz.f32 %f1279, %f2982, %f1273; fma.rn.ftz.f32 %f1280, %f2981, %f1272, %f1279; fma.rn.ftz.f32 %f2990, %f2983, %f1274, %f1280; $L__BB211_83: add.ftz.f32 %f1311, %f5, %f104; selp.f32 %f1312, %f1311, %f104, %p69; sub.ftz.f32 %f1313, %f1312, %f4; mul.ftz.f32 %f184, %f103, %f1313; mul.ftz.f32 %f1314, %f184, %f2988; mul.ftz.f32 %f1315, %f184, %f2989; mul.ftz.f32 %f1316, %f184, %f2990; sub.ftz.f32 %f185, %f1140, %f1314; sub.ftz.f32 %f186, %f1141, %f1315; sub.ftz.f32 %f187, %f1142, %f1316; st.local.v2.f32 [%rd71], {%f185, %f186}; st.local.f32 [%rd28], %f187; ld.local.u32 %r520, [%rd49]; setp.ne.s32 %p70, %r520, 16; mov.f32 %f2997, 0f00000000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; @%p70 bra $L__BB211_112; and.b32 %r521, %r885, 3; setp.eq.s32 %p71, %r521, 0; @%p71 bra $L__BB211_86; mul.wide.s32 %rd339, %r886, 16; add.s64 %rd340, %rd54, %rd339; ld.global.v4.f32 {%f2997, %f2998, %f2999, %f3000}, [%rd340]; bra.uni $L__BB211_110; $L__BB211_86: setp.ne.s32 %p72, %r885, 1024; mov.f32 %f2997, 0f00000000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; mov.f32 %f3000, %f2997; @%p72 bra $L__BB211_110; @%p68 bra $L__BB211_89; bra.uni $L__BB211_88; $L__BB211_89: ld.local.u32 %r522, [%rd33]; mul.wide.s32 %rd344, %r522, 208; add.s64 %rd345, %rd12, %rd344; ld.global.v4.f32 {%f1358, %f1359, %f1360, %f1361}, [%rd345+48]; ld.global.v4.f32 {%f1366, %f1367, %f1368, %f1369}, [%rd345+64]; ld.global.v4.f32 {%f1374, %f1375, %f1376, %f1377}, [%rd345+80]; mul.ftz.f32 %f1382, %f186, %f1359; fma.rn.ftz.f32 %f1383, %f185, %f1358, %f1382; fma.rn.ftz.f32 %f1384, %f187, %f1360, %f1383; add.ftz.f32 %f204, %f1361, %f1384; mul.ftz.f32 %f1385, %f186, %f1367; fma.rn.ftz.f32 %f1386, %f185, %f1366, %f1385; fma.rn.ftz.f32 %f1387, %f187, %f1368, %f1386; add.ftz.f32 %f205, %f1369, %f1387; mul.ftz.f32 %f1388, %f186, %f1375; fma.rn.ftz.f32 %f1389, %f185, %f1374, %f1388; fma.rn.ftz.f32 %f1390, %f187, %f1376, %f1389; add.ftz.f32 %f206, %f1377, %f1390; bra.uni $L__BB211_90; $L__BB211_88: cvta.to.local.u64 %rd341, %rd48; ld.local.v4.f32 {%f1325, %f1326, %f1327, %f1328}, [%rd341]; mul.ftz.f32 %f1333, %f186, %f1326; fma.rn.ftz.f32 %f1334, %f185, %f1325, %f1333; fma.rn.ftz.f32 %f1335, %f187, %f1327, %f1334; add.ftz.f32 %f204, %f1328, %f1335; cvta.to.local.u64 %rd342, %rd51; ld.local.v4.f32 {%f1336, %f1337, %f1338, %f1339}, [%rd342]; mul.ftz.f32 %f1344, %f186, %f1337; fma.rn.ftz.f32 %f1345, %f185, %f1336, %f1344; fma.rn.ftz.f32 %f1346, %f187, %f1338, %f1345; add.ftz.f32 %f205, %f1339, %f1346; cvta.to.local.u64 %rd343, %rd50; ld.local.v4.f32 {%f1347, %f1348, %f1349, %f1350}, [%rd343]; mul.ftz.f32 %f1355, %f186, %f1348; fma.rn.ftz.f32 %f1356, %f185, %f1347, %f1355; fma.rn.ftz.f32 %f1357, %f187, %f1349, %f1356; add.ftz.f32 %f206, %f1350, %f1357; $L__BB211_90: ld.local.u32 %r69, [%rd31]; mul.wide.s32 %rd346, %r886, 96; add.s64 %rd93, %rd55, %rd346; add.s64 %rd347, %rd56, %rd346; add.s64 %rd94, %rd347, 32; ld.global.u32 %r523, [%rd347+32]; setp.eq.s32 %p74, %r523, 0; @%p74 bra $L__BB211_92; ld.global.v4.f32 {%f1391, %f1392, %f1393, %f1394}, [%rd94+16]; mul.ftz.f32 %f1399, %f205, %f1392; fma.rn.ftz.f32 %f1400, %f204, %f1391, %f1399; fma.rn.ftz.f32 %f1401, %f206, %f1393, %f1400; add.ftz.f32 %f201, %f1394, %f1401; ld.global.v4.f32 {%f1402, %f1403, %f1404, %f1405}, [%rd94+32]; mul.ftz.f32 %f1410, %f205, %f1403; fma.rn.ftz.f32 %f1411, %f204, %f1402, %f1410; fma.rn.ftz.f32 %f1412, %f206, %f1404, %f1411; add.ftz.f32 %f202, %f1405, %f1412; ld.global.v4.f32 {%f1413, %f1414, %f1415, %f1416}, [%rd94+48]; mul.ftz.f32 %f1421, %f205, %f1414; fma.rn.ftz.f32 %f1422, %f204, %f1413, %f1421; fma.rn.ftz.f32 %f1423, %f206, %f1415, %f1422; add.ftz.f32 %f206, %f1416, %f1423; mov.f32 %f204, %f201; mov.f32 %f205, %f202; $L__BB211_92: and.b32 %r524, %r69, 16777216; setp.ne.s32 %p75, %r524, 0; setp.eq.s32 %p76, %r524, 0; selp.b32 %r72, -1, 2, %p76; @%p75 bra $L__BB211_94; ld.global.u32 %r72, [%rd94+-20]; $L__BB211_94: ld.global.u32 %r73, [%rd94+-24]; setp.eq.s32 %p77, %r73, 8; @%p77 bra $L__BB211_109; bra.uni $L__BB211_95; $L__BB211_109: { // callseq 1204, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .b32 param4; st.param.b32 [param4+0], %r72; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2997, [retval0+0]; } // callseq 1204 mov.f32 %f3000, 0f3F800000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; bra.uni $L__BB211_110; $L__BB211_95: setp.eq.s32 %p78, %r73, 9; @%p78 bra $L__BB211_108; bra.uni $L__BB211_96; $L__BB211_108: { // callseq 1203, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .b32 param4; st.param.b32 [param4+0], %r72; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2997, [retval0+0]; ld.param.f32 %f2998, [retval0+4]; ld.param.f32 %f2999, [retval0+8]; } // callseq 1203 mov.f32 %f3000, 0f3F800000; bra.uni $L__BB211_110; $L__BB211_96: setp.eq.s32 %p79, %r73, 10; @%p79 bra $L__BB211_107; bra.uni $L__BB211_97; $L__BB211_107: { // callseq 1202, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .b32 param4; st.param.b32 [param4+0], %r72; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2997, [retval0+0]; } // callseq 1202 mov.f32 %f3000, 0f3F800000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; bra.uni $L__BB211_110; $L__BB211_97: setp.eq.s32 %p80, %r73, 11; @%p80 bra $L__BB211_106; bra.uni $L__BB211_98; $L__BB211_106: { // callseq 1201, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .b32 param4; st.param.b32 [param4+0], %r72; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f2997, [retval0+0]; } // callseq 1201 mov.f32 %f3000, 0f3F800000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; bra.uni $L__BB211_110; $L__BB211_98: or.b32 %r74, %r72, 1; setp.gt.u32 %p81, %r73, 6; @%p81 bra $L__BB211_103; mov.u32 %r526, 1; shl.b32 %r527, %r526, %r73; and.b32 %r528, %r527, 71; setp.ne.s32 %p82, %r528, 0; @%p82 bra $L__BB211_100; bra.uni $L__BB211_103; $L__BB211_100: setp.eq.s32 %p83, %r74, 3; @%p83 bra $L__BB211_102; bra.uni $L__BB211_101; $L__BB211_102: { // callseq 1199, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f2997, [retval0+0]; ld.param.f32 %f2998, [retval0+4]; ld.param.f32 %f2999, [retval0+8]; ld.param.f32 %f3000, [retval0+12]; } // callseq 1199 bra.uni $L__BB211_110; $L__BB211_101: ld.global.u64 %rd348, [%rd94+-32]; tex.3d.v4.f32.f32 {%f2997, %f2998, %f2999, %f3000}, [%rd348, {%f204, %f205, %f206, %f206}]; bra.uni $L__BB211_110; $L__BB211_103: setp.eq.s32 %p84, %r74, 3; @%p84 bra $L__BB211_105; bra.uni $L__BB211_104; $L__BB211_105: { // callseq 1200, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd93; .param .b32 param1; st.param.f32 [param1+0], %f204; .param .b32 param2; st.param.f32 [param2+0], %f205; .param .b32 param3; st.param.f32 [param3+0], %f206; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f2997, [retval0+0]; } // callseq 1200 mov.f32 %f3000, 0f3F800000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; bra.uni $L__BB211_110; $L__BB211_104: ld.global.u64 %rd349, [%rd94+-32]; tex.3d.v4.f32.f32 {%f2997, %f1424, %f1425, %f1426}, [%rd349, {%f204, %f205, %f206, %f206}]; mov.f32 %f3000, 0f3F800000; mov.f32 %f2998, %f2997; mov.f32 %f2999, %f2997; $L__BB211_110: setp.leu.ftz.f32 %p85, %f3000, 0f358637BD; setp.eq.ftz.f32 %p86, %f3000, 0f3F800000; or.pred %p87, %p85, %p86; @%p87 bra $L__BB211_112; div.approx.ftz.f32 %f2999, %f2999, %f3000; div.approx.ftz.f32 %f2998, %f2998, %f3000; div.approx.ftz.f32 %f2997, %f2997, %f3000; $L__BB211_112: ld.local.u8 %rs30, [%rd32]; and.b16 %rs31, %rs30, 2; setp.eq.s16 %p88, %rs31, 0; @%p88 bra $L__BB211_114; bra.uni $L__BB211_113; $L__BB211_114: ld.local.u32 %r529, [%rd33]; mul.wide.s32 %rd353, %r529, 208; add.s64 %rd354, %rd12, %rd353; ld.global.v4.f32 {%f1454, %f1455, %f1456, %f1457}, [%rd354]; ld.global.v4.f32 {%f1461, %f1462, %f1463, %f1464}, [%rd354+16]; ld.global.v4.f32 {%f1468, %f1469, %f1470, %f1471}, [%rd354+32]; mul.ftz.f32 %f1475, %f2998, %f1455; fma.rn.ftz.f32 %f1476, %f2997, %f1454, %f1475; fma.rn.ftz.f32 %f3004, %f2999, %f1456, %f1476; mul.ftz.f32 %f1477, %f2998, %f1462; fma.rn.ftz.f32 %f1478, %f2997, %f1461, %f1477; fma.rn.ftz.f32 %f3005, %f2999, %f1463, %f1478; mul.ftz.f32 %f1479, %f2998, %f1469; fma.rn.ftz.f32 %f1480, %f2997, %f1468, %f1479; fma.rn.ftz.f32 %f3006, %f2999, %f1470, %f1480; bra.uni $L__BB211_115; $L__BB211_113: cvta.to.local.u64 %rd350, %rd44; ld.local.v4.f32 {%f1427, %f1428, %f1429, %f1430}, [%rd350]; mul.ftz.f32 %f1434, %f2998, %f1428; fma.rn.ftz.f32 %f1435, %f2997, %f1427, %f1434; fma.rn.ftz.f32 %f3004, %f2999, %f1429, %f1435; cvta.to.local.u64 %rd351, %rd47; ld.local.v4.f32 {%f1436, %f1437, %f1438, %f1439}, [%rd351]; mul.ftz.f32 %f1443, %f2998, %f1437; fma.rn.ftz.f32 %f1444, %f2997, %f1436, %f1443; fma.rn.ftz.f32 %f3005, %f2999, %f1438, %f1444; cvta.to.local.u64 %rd352, %rd46; ld.local.v4.f32 {%f1445, %f1446, %f1447, %f1448}, [%rd352]; mul.ftz.f32 %f1452, %f2998, %f1446; fma.rn.ftz.f32 %f1453, %f2997, %f1445, %f1452; fma.rn.ftz.f32 %f3006, %f2999, %f1447, %f1453; $L__BB211_115: mul.ftz.f32 %f1481, %f184, %f3004; mul.ftz.f32 %f1482, %f184, %f3005; mul.ftz.f32 %f1483, %f184, %f3006; sub.ftz.f32 %f1484, %f1142, %f1483; sub.ftz.f32 %f1485, %f1141, %f1482; sub.ftz.f32 %f1486, %f1140, %f1481; st.local.v2.f32 [%rd71], {%f1486, %f1485}; st.local.f32 [%rd28], %f1484; ld.local.u32 %r530, [%rd18]; and.b32 %r890, %r530, 4194303; bra.uni $L__BB211_116; $L__BB211_159: mov.u32 %r589, 24; { // callseq 1264, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r589; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1264 $L__BB211_116: mov.u32 %r77, %r890; mul.wide.s32 %rd355, %r77, 16; add.s64 %rd95, %rd36, %rd355; ld.global.v4.u32 {%r532, %r533, %r534, %r535}, [%rd95]; mul.wide.u32 %rd356, %r533, 4; add.s64 %rd96, %rd1, %rd356; mul.wide.u32 %rd357, %r534, 4; add.s64 %rd97, %rd1, %rd357; add.s32 %r890, %r77, 1; setp.gt.s32 %p89, %r532, 35; @%p89 bra $L__BB211_164; setp.gt.s32 %p177, %r532, 24; @%p177 bra $L__BB211_160; bra.uni $L__BB211_118; $L__BB211_160: add.s32 %r537, %r532, -27; setp.lt.u32 %p178, %r537, 9; @%p178 bra $L__BB211_116; setp.eq.s32 %p179, %r532, 25; @%p179 bra $L__BB211_316; setp.eq.s32 %p180, %r532, 26; @%p180 bra $L__BB211_163; bra.uni $L__BB211_329; $L__BB211_163: mov.u32 %r587, 26; { // callseq 1262, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r587; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1262 bra.uni $L__BB211_116; $L__BB211_164: setp.gt.s32 %p90, %r532, 71; @%p90 bra $L__BB211_232; setp.gt.s32 %p126, %r532, 53; @%p126 bra $L__BB211_199; setp.gt.s32 %p152, %r532, 44; @%p152 bra $L__BB211_183; setp.gt.s32 %p165, %r532, 39; @%p165 bra $L__BB211_175; bra.uni $L__BB211_168; $L__BB211_175: setp.gt.s32 %p166, %r532, 41; @%p166 bra $L__BB211_179; bra.uni $L__BB211_176; $L__BB211_179: setp.eq.s32 %p167, %r532, 42; @%p167 bra $L__BB211_312; setp.eq.s32 %p168, %r532, 43; @%p168 bra $L__BB211_311; setp.eq.s32 %p169, %r532, 44; @%p169 bra $L__BB211_182; bra.uni $L__BB211_329; $L__BB211_182: mov.u32 %r577, 44; add.s32 %r578, %r77, 1; { // callseq 1253, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r577; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r578; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1253 bra.uni $L__BB211_116; $L__BB211_118: setp.gt.s32 %p181, %r532, 17; @%p181 bra $L__BB211_152; bra.uni $L__BB211_119; $L__BB211_152: setp.lt.s32 %p182, %r532, 21; @%p182 bra $L__BB211_116; setp.gt.s32 %p183, %r532, 22; @%p183 bra $L__BB211_157; setp.eq.s32 %p186, %r532, 21; @%p186 bra $L__BB211_318; setp.eq.s32 %p187, %r532, 22; @%p187 bra $L__BB211_156; bra.uni $L__BB211_329; $L__BB211_156: mov.u32 %r592, 22; add.s32 %r593, %r77, 1; { // callseq 1266, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r592; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r593; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1266 bra.uni $L__BB211_116; $L__BB211_232: setp.gt.s32 %p91, %r532, 89; @%p91 bra $L__BB211_267; setp.gt.s32 %p101, %r532, 80; @%p101 bra $L__BB211_251; setp.gt.s32 %p114, %r532, 75; @%p114 bra $L__BB211_242; bra.uni $L__BB211_235; $L__BB211_242: setp.gt.s32 %p115, %r532, 77; @%p115 bra $L__BB211_247; bra.uni $L__BB211_243; $L__BB211_247: setp.eq.s32 %p116, %r532, 78; @%p116 bra $L__BB211_287; setp.eq.s32 %p117, %r532, 79; @%p117 bra $L__BB211_286; setp.eq.s32 %p118, %r532, 80; @%p118 bra $L__BB211_250; bra.uni $L__BB211_329; $L__BB211_250: { // callseq 1219, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1219 bra.uni $L__BB211_116; $L__BB211_199: setp.gt.s32 %p127, %r532, 62; @%p127 bra $L__BB211_216; setp.gt.s32 %p140, %r532, 57; @%p140 bra $L__BB211_208; bra.uni $L__BB211_201; $L__BB211_208: setp.gt.s32 %p141, %r532, 59; @%p141 bra $L__BB211_212; bra.uni $L__BB211_209; $L__BB211_212: setp.eq.s32 %p142, %r532, 60; @%p142 bra $L__BB211_302; setp.eq.s32 %p143, %r532, 61; @%p143 bra $L__BB211_301; setp.eq.s32 %p144, %r532, 62; @%p144 bra $L__BB211_215; bra.uni $L__BB211_329; $L__BB211_215: mov.u32 %r563, 62; add.s32 %r564, %r77, 1; { // callseq 1235, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r563; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r564; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1235 bra.uni $L__BB211_116; $L__BB211_119: setp.gt.s32 %p188, %r532, 8; @%p188 bra $L__BB211_135; bra.uni $L__BB211_120; $L__BB211_135: setp.gt.s32 %p189, %r532, 12; @%p189 bra $L__BB211_144; bra.uni $L__BB211_136; $L__BB211_144: setp.gt.s32 %p190, %r532, 14; @%p190 bra $L__BB211_148; bra.uni $L__BB211_145; $L__BB211_148: setp.eq.s32 %p191, %r532, 15; @%p191 bra $L__BB211_320; setp.eq.s32 %p192, %r532, 16; @%p192 bra $L__BB211_319; setp.eq.s32 %p193, %r532, 17; @%p193 bra $L__BB211_151; bra.uni $L__BB211_329; $L__BB211_151: { // callseq 1268, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; .param .b32 param4; st.param.b32 [param4+0], %r535; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1268 bra.uni $L__BB211_116; $L__BB211_267: setp.gt.s32 %p92, %r532, 93; @%p92 bra $L__BB211_271; add.s32 %r536, %r532, -92; setp.lt.u32 %p99, %r536, 2; @%p99 bra $L__BB211_116; setp.eq.s32 %p100, %r532, 90; @%p100 bra $L__BB211_270; bra.uni $L__BB211_329; $L__BB211_270: mov.u32 %r540, 90; add.s32 %r541, %r77, 1; { // callseq 1209, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r540; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; .param .b32 param3; st.param.b32 [param3+0], %r541; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1209 bra.uni $L__BB211_116; $L__BB211_183: setp.gt.s32 %p153, %r532, 48; @%p153 bra $L__BB211_191; bra.uni $L__BB211_184; $L__BB211_191: setp.gt.s32 %p154, %r532, 50; @%p154 bra $L__BB211_195; bra.uni $L__BB211_192; $L__BB211_195: setp.eq.s32 %p155, %r532, 51; @%p155 bra $L__BB211_307; setp.eq.s32 %p156, %r532, 52; @%p156 bra $L__BB211_306; setp.eq.s32 %p157, %r532, 53; @%p157 bra $L__BB211_198; bra.uni $L__BB211_329; $L__BB211_198: { // callseq 1244, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1244 bra.uni $L__BB211_116; $L__BB211_251: setp.gt.s32 %p102, %r532, 84; @%p102 bra $L__BB211_259; bra.uni $L__BB211_252; $L__BB211_259: setp.gt.s32 %p103, %r532, 86; @%p103 bra $L__BB211_263; bra.uni $L__BB211_260; $L__BB211_263: setp.eq.s32 %p104, %r532, 87; @%p104 bra $L__BB211_282; setp.eq.s32 %p105, %r532, 88; @%p105 bra $L__BB211_281; setp.eq.s32 %p106, %r532, 89; @%p106 bra $L__BB211_266; bra.uni $L__BB211_329; $L__BB211_266: mov.u32 %r542, 89; { // callseq 1210, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r542; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1210 bra.uni $L__BB211_116; $L__BB211_216: setp.gt.s32 %p128, %r532, 66; @%p128 bra $L__BB211_224; bra.uni $L__BB211_217; $L__BB211_224: setp.gt.s32 %p129, %r532, 68; @%p129 bra $L__BB211_228; bra.uni $L__BB211_225; $L__BB211_228: setp.eq.s32 %p130, %r532, 69; @%p130 bra $L__BB211_297; setp.eq.s32 %p131, %r532, 70; @%p131 bra $L__BB211_296; setp.eq.s32 %p132, %r532, 71; @%p132 bra $L__BB211_231; bra.uni $L__BB211_329; $L__BB211_231: mov.u32 %r553, 71; { // callseq 1226, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r553; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1226 bra.uni $L__BB211_116; $L__BB211_120: setp.gt.s32 %p201, %r532, 4; @%p201 bra $L__BB211_128; setp.gt.s32 %p207, %r532, 2; @%p207 bra $L__BB211_125; setp.eq.s32 %p210, %r532, 1; mov.u32 %r890, %r534; @%p210 bra $L__BB211_116; setp.eq.s32 %p211, %r532, 2; @%p211 bra $L__BB211_124; bra.uni $L__BB211_329; $L__BB211_124: mov.u32 %r612, 2; add.s32 %r613, %r77, 1; { // callseq 1277, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r612; st.param.b32 [param0+4], %r533; st.param.b32 [param0+8], %r534; st.param.b32 [param0+12], %r535; .param .b32 param1; st.param.b32 [param1+0], %r613; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1277 bra.uni $L__BB211_116; $L__BB211_271: setp.gt.s32 %p93, %r532, 95; @%p93 bra $L__BB211_275; bra.uni $L__BB211_272; $L__BB211_275: setp.eq.s32 %p94, %r532, 98; @%p94 bra $L__BB211_328; setp.eq.s32 %p95, %r532, 97; @%p95 bra $L__BB211_280; setp.ne.s32 %p96, %r532, 96; @%p96 bra $L__BB211_329; { // callseq 1205, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1205 bra.uni $L__BB211_116; $L__BB211_136: setp.gt.s32 %p196, %r532, 10; @%p196 bra $L__BB211_141; setp.eq.s32 %p199, %r532, 9; @%p199 bra $L__BB211_323; setp.eq.s32 %p200, %r532, 10; @%p200 bra $L__BB211_139; bra.uni $L__BB211_329; $L__BB211_139: ld.local.f32 %f1503, [%rd97]; setp.ltu.ftz.f32 %p216, %f1503, 0f3F800000; @%p216 bra $L__BB211_116; add.s32 %r606, %r533, %r77; add.s32 %r890, %r606, 1; bra.uni $L__BB211_116; $L__BB211_128: setp.gt.s32 %p202, %r532, 6; @%p202 bra $L__BB211_132; setp.eq.s32 %p205, %r532, 5; @%p205 bra $L__BB211_326; setp.eq.s32 %p206, %r532, 6; @%p206 bra $L__BB211_131; bra.uni $L__BB211_329; $L__BB211_131: ld.local.f32 %f1505, [%rd96]; ld.local.f32 %f1506, [%rd96+4]; ld.local.f32 %f1507, [%rd96+8]; st.local.f32 [%rd40], %f1505; st.local.f32 [%rd41], %f1506; st.local.f32 [%rd42], %f1507; bra.uni $L__BB211_116; $L__BB211_168: setp.gt.s32 %p172, %r532, 37; @%p172 bra $L__BB211_172; setp.eq.s32 %p175, %r532, 36; @%p175 bra $L__BB211_315; setp.eq.s32 %p176, %r532, 37; @%p176 bra $L__BB211_171; bra.uni $L__BB211_329; $L__BB211_171: mov.u32 %r585, 37; { // callseq 1260, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r585, %r533, %r534, %r535}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1260 bra.uni $L__BB211_116; $L__BB211_235: setp.gt.s32 %p121, %r532, 73; @%p121 bra $L__BB211_239; setp.eq.s32 %p124, %r532, 72; @%p124 bra $L__BB211_295; setp.eq.s32 %p125, %r532, 73; @%p125 bra $L__BB211_238; bra.uni $L__BB211_329; $L__BB211_238: add.s32 %r552, %r77, 1; { // callseq 1224, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r552; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1224 bra.uni $L__BB211_116; $L__BB211_201: setp.gt.s32 %p147, %r532, 55; @%p147 bra $L__BB211_205; setp.eq.s32 %p150, %r532, 54; @%p150 bra $L__BB211_305; setp.eq.s32 %p151, %r532, 55; @%p151 bra $L__BB211_204; bra.uni $L__BB211_329; $L__BB211_204: ld.local.v4.f32 {%f1496, %f1497, %f1498, %f1499}, [%rd71]; { // callseq 1242, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1496; st.param.f32 [param0+4], %f1497; st.param.f32 [param0+8], %f1498; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; .param .b32 param4; st.param.b32 [param4+0], %r535; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1242 bra.uni $L__BB211_116; $L__BB211_184: setp.gt.s32 %p160, %r532, 46; @%p160 bra $L__BB211_188; setp.eq.s32 %p163, %r532, 45; @%p163 bra $L__BB211_310; setp.eq.s32 %p164, %r532, 46; @%p164 bra $L__BB211_187; bra.uni $L__BB211_329; $L__BB211_187: { // callseq 1251, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1251 bra.uni $L__BB211_116; $L__BB211_252: setp.gt.s32 %p109, %r532, 82; @%p109 bra $L__BB211_256; setp.eq.s32 %p112, %r532, 81; @%p112 bra $L__BB211_285; setp.eq.s32 %p113, %r532, 82; @%p113 bra $L__BB211_255; bra.uni $L__BB211_329; $L__BB211_255: mov.u32 %r547, 82; { // callseq 1217, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r547; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1217 bra.uni $L__BB211_116; $L__BB211_217: setp.gt.s32 %p135, %r532, 64; @%p135 bra $L__BB211_221; setp.eq.s32 %p138, %r532, 63; @%p138 bra $L__BB211_300; setp.eq.s32 %p139, %r532, 64; @%p139 bra $L__BB211_220; bra.uni $L__BB211_329; $L__BB211_220: mov.u32 %r560, 64; add.s32 %r561, %r77, 1; { // callseq 1233, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r560; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r561; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1233 bra.uni $L__BB211_116; $L__BB211_157: setp.eq.s32 %p184, %r532, 23; @%p184 bra $L__BB211_317; setp.eq.s32 %p185, %r532, 24; @%p185 bra $L__BB211_159; bra.uni $L__BB211_329; $L__BB211_145: setp.eq.s32 %p194, %r532, 13; @%p194 bra $L__BB211_321; setp.eq.s32 %p195, %r532, 14; @%p195 bra $L__BB211_147; bra.uni $L__BB211_329; $L__BB211_147: st.local.u32 [%rd97], %r533; bra.uni $L__BB211_116; $L__BB211_125: setp.eq.s32 %p208, %r532, 3; @%p208 bra $L__BB211_327; setp.eq.s32 %p209, %r532, 4; @%p209 bra $L__BB211_127; bra.uni $L__BB211_329; $L__BB211_127: mov.u32 %r610, 4; { // callseq 1275, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r610, %r533, %r534, %r535}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1275 bra.uni $L__BB211_116; $L__BB211_141: setp.eq.s32 %p197, %r532, 11; @%p197 bra $L__BB211_322; setp.eq.s32 %p198, %r532, 12; @%p198 bra $L__BB211_143; bra.uni $L__BB211_329; $L__BB211_143: { // callseq 1271, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1271 bra.uni $L__BB211_116; $L__BB211_132: setp.eq.s32 %p203, %r532, 7; @%p203 bra $L__BB211_325; setp.eq.s32 %p204, %r532, 8; @%p204 bra $L__BB211_134; bra.uni $L__BB211_329; $L__BB211_134: mov.u32 %r608, 8; { // callseq 1273, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r608; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1273 bra.uni $L__BB211_116; $L__BB211_272: setp.eq.s32 %p97, %r532, 94; @%p97 bra $L__BB211_279; setp.eq.s32 %p98, %r532, 95; @%p98 bra $L__BB211_274; bra.uni $L__BB211_329; $L__BB211_274: { // callseq 1206, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1206 bra.uni $L__BB211_116; $L__BB211_176: setp.eq.s32 %p170, %r532, 40; @%p170 bra $L__BB211_313; setp.eq.s32 %p171, %r532, 41; @%p171 bra $L__BB211_178; bra.uni $L__BB211_329; $L__BB211_178: mov.u32 %r580, 41; add.s32 %r581, %r77, 1; mov.u32 %r582, 768; { // callseq 1256, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r580; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; .param .b32 param3; st.param.b32 [param3+0], %r582; .param .b32 param4; st.param.b32 [param4+0], %r581; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1256 bra.uni $L__BB211_116; $L__BB211_243: setp.eq.s32 %p119, %r532, 76; @%p119 bra $L__BB211_288; setp.eq.s32 %p120, %r532, 77; @%p120 bra $L__BB211_245; bra.uni $L__BB211_329; $L__BB211_245: setp.eq.s32 %p212, %r535, 255; @%p212 bra $L__BB211_116; ld.local.f32 %f1492, [%rd96]; add.s32 %r551, %r535, %r534; mul.wide.u32 %rd378, %r551, 4; add.s64 %rd379, %rd1, %rd378; st.local.f32 [%rd379], %f1492; bra.uni $L__BB211_116; $L__BB211_209: setp.eq.s32 %p145, %r532, 58; @%p145 bra $L__BB211_303; setp.eq.s32 %p146, %r532, 59; @%p146 bra $L__BB211_211; bra.uni $L__BB211_329; $L__BB211_211: add.s32 %r568, %r77, 1; { // callseq 1238, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r568; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1238 bra.uni $L__BB211_116; $L__BB211_192: setp.eq.s32 %p158, %r532, 49; @%p158 bra $L__BB211_308; setp.eq.s32 %p159, %r532, 50; @%p159 bra $L__BB211_194; bra.uni $L__BB211_329; $L__BB211_194: { // callseq 1247, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1247 bra.uni $L__BB211_116; $L__BB211_260: setp.eq.s32 %p107, %r532, 85; @%p107 bra $L__BB211_283; setp.eq.s32 %p108, %r532, 86; @%p108 bra $L__BB211_262; bra.uni $L__BB211_329; $L__BB211_262: add.s32 %r545, %r77, 1; { // callseq 1213, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r545; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1213 bra.uni $L__BB211_116; $L__BB211_225: setp.eq.s32 %p133, %r532, 67; @%p133 bra $L__BB211_298; setp.eq.s32 %p134, %r532, 68; @%p134 bra $L__BB211_227; bra.uni $L__BB211_329; $L__BB211_227: mov.u32 %r557, 68; { // callseq 1229, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r557; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1229 bra.uni $L__BB211_116; $L__BB211_172: setp.eq.s32 %p173, %r532, 38; @%p173 bra $L__BB211_314; setp.eq.s32 %p174, %r532, 39; @%p174 bra $L__BB211_174; bra.uni $L__BB211_329; $L__BB211_174: mov.u32 %r584, 39; { // callseq 1258, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r584, %r533, %r534, %r535}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1258 bra.uni $L__BB211_116; $L__BB211_239: setp.eq.s32 %p122, %r532, 74; @%p122 bra $L__BB211_294; setp.eq.s32 %p123, %r532, 75; @%p123 bra $L__BB211_241; bra.uni $L__BB211_329; $L__BB211_241: { // callseq 1222, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1222 bra.uni $L__BB211_116; $L__BB211_205: setp.eq.s32 %p148, %r532, 56; @%p148 bra $L__BB211_304; setp.eq.s32 %p149, %r532, 57; @%p149 bra $L__BB211_207; bra.uni $L__BB211_329; $L__BB211_207: mov.u32 %r570, 57; add.s32 %r571, %r77, 1; { // callseq 1240, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r570; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r571; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1240 bra.uni $L__BB211_116; $L__BB211_188: setp.eq.s32 %p161, %r532, 47; @%p161 bra $L__BB211_309; setp.eq.s32 %p162, %r532, 48; @%p162 bra $L__BB211_190; bra.uni $L__BB211_329; $L__BB211_190: { // callseq 1249, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1249 bra.uni $L__BB211_116; $L__BB211_256: setp.eq.s32 %p110, %r532, 83; @%p110 bra $L__BB211_284; setp.eq.s32 %p111, %r532, 84; @%p111 bra $L__BB211_258; bra.uni $L__BB211_329; $L__BB211_258: { // callseq 1215, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1215 bra.uni $L__BB211_116; $L__BB211_221: setp.eq.s32 %p136, %r532, 65; @%p136 bra $L__BB211_299; setp.eq.s32 %p137, %r532, 66; @%p137 bra $L__BB211_223; bra.uni $L__BB211_329; $L__BB211_223: add.s32 %r559, %r77, 1; { // callseq 1231, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r559; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1231 bra.uni $L__BB211_116; $L__BB211_318: mov.u32 %r594, 21; { // callseq 1267, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r594; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1267 bra.uni $L__BB211_116; $L__BB211_320: ld.global.v4.u32 {%r596, %r597, %r598, %r599}, [%rd95+16]; add.s32 %r890, %r77, 2; st.local.u32 [%rd96], %r597; st.local.u32 [%rd96+4], %r598; st.local.u32 [%rd96+8], %r599; bra.uni $L__BB211_116; $L__BB211_319: mov.u32 %r860, 16; { // callseq 1269, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r860; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1269 bra.uni $L__BB211_116; $L__BB211_323: ld.local.f32 %f1504, [%rd97]; setp.gtu.ftz.f32 %p217, %f1504, 0f00000000; @%p217 bra $L__BB211_116; add.s32 %r607, %r533, %r77; add.s32 %r890, %r607, 1; bra.uni $L__BB211_116; $L__BB211_326: st.local.u32 [%rd40], %r533; st.local.u32 [%rd41], %r534; st.local.u32 [%rd42], %r535; bra.uni $L__BB211_116; $L__BB211_317: mov.u32 %r590, 23; add.s32 %r591, %r77, 1; { // callseq 1265, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r590; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r591; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1265 bra.uni $L__BB211_116; $L__BB211_321: mov.u32 %r603, 13; add.s32 %r604, %r77, 1; mov.u32 %r605, 768; { // callseq 1270, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b32 param1; st.param.b32 [param1+0], %r605; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r603; st.param.b32 [param3+4], %r533; st.param.b32 [param3+8], %r534; st.param.b32 [param3+12], %r535; .param .b32 param4; st.param.b32 [param4+0], %r604; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1270 bra.uni $L__BB211_116; $L__BB211_327: mov.u32 %r611, 3; { // callseq 1276, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r611, %r533, %r534, %r535}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1276 bra.uni $L__BB211_116; $L__BB211_322: { // callseq 1272, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1272 bra.uni $L__BB211_116; $L__BB211_325: mov.u32 %r609, 7; { // callseq 1274, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r609; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1274 bra.uni $L__BB211_116; $L__BB211_312: { // callseq 1255, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1255 bra.uni $L__BB211_116; $L__BB211_311: add.s32 %r579, %r77, 1; { // callseq 1254, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r579; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1254 bra.uni $L__BB211_116; $L__BB211_287: add.s32 %r550, %r77, 1; { // callseq 1221, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r550; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1221 bra.uni $L__BB211_116; $L__BB211_286: add.s32 %r549, %r77, 1; { // callseq 1220, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r549; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1220 bra.uni $L__BB211_116; $L__BB211_302: add.s32 %r567, %r77, 1; { // callseq 1237, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r567; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1237 bra.uni $L__BB211_116; $L__BB211_301: mov.u32 %r565, 61; add.s32 %r566, %r77, 1; { // callseq 1236, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r565; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r566; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1236 bra.uni $L__BB211_116; $L__BB211_307: { // callseq 1246, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1246 bra.uni $L__BB211_116; $L__BB211_306: add.s32 %r574, %r77, 1; { // callseq 1245, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r574; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1245 bra.uni $L__BB211_116; $L__BB211_282: add.s32 %r544, %r77, 1; { // callseq 1212, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r544; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1212 bra.uni $L__BB211_116; $L__BB211_281: mov.u32 %r543, 88; ld.local.f32 %f1487, [%rd71+12]; ld.local.v2.f32 {%f1488, %f1489}, [%rd71+16]; { // callseq 1211, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1487; .param .b32 param1; st.param.f32 [param1+0], %f1488; .param .b32 param2; st.param.f32 [param2+0], %f1489; .param .b64 param3; st.param.b64 [param3+0], %rd146; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r543; st.param.b32 [param4+4], %r533; st.param.b32 [param4+8], %r534; st.param.b32 [param4+12], %r535; call.uni _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1211 bra.uni $L__BB211_116; $L__BB211_297: mov.u32 %r555, 69; add.s32 %r556, %r77, 1; { // callseq 1228, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r555; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r556; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1228 bra.uni $L__BB211_116; $L__BB211_296: mov.u32 %r554, 70; { // callseq 1227, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r554; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1227 bra.uni $L__BB211_116; $L__BB211_328: { // callseq 1278, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1278 bra.uni $L__BB211_116; $L__BB211_280: { // callseq 1208, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1208 bra.uni $L__BB211_116; $L__BB211_315: mov.u32 %r586, 36; { // callseq 1261, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r586; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1261 bra.uni $L__BB211_116; $L__BB211_295: { // callseq 1225, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1225 bra.uni $L__BB211_116; $L__BB211_305: add.s32 %r573, %r77, 1; { // callseq 1243, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r573; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1243 bra.uni $L__BB211_116; $L__BB211_310: { // callseq 1252, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1252 bra.uni $L__BB211_116; $L__BB211_285: mov.u32 %r548, 81; { // callseq 1218, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r548; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1218 bra.uni $L__BB211_116; $L__BB211_300: mov.u32 %r562, 63; { // callseq 1234, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r562; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1234 bra.uni $L__BB211_116; $L__BB211_279: mov.u32 %r538, 94; add.s32 %r539, %r77, 1; { // callseq 1207, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r538; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; .param .b32 param2; st.param.b32 [param2+0], %r539; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1207 bra.uni $L__BB211_116; $L__BB211_313: mov.u32 %r583, 40; { // callseq 1257, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r583; st.param.b32 [param2+4], %r533; st.param.b32 [param2+8], %r534; st.param.b32 [param2+12], %r535; call.uni _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1257 bra.uni $L__BB211_116; $L__BB211_288: setp.eq.s32 %p213, %r535, 255; @%p213 bra $L__BB211_116; add.s64 %rd98, %rd96, 4; add.s64 %rd99, %rd96, 8; setp.eq.s32 %p214, %r534, 0; mul.wide.u32 %rd380, %r535, 4; add.s64 %rd100, %rd1, %rd380; @%p214 bra $L__BB211_293; setp.eq.s32 %p215, %r534, 1; @%p215 bra $L__BB211_292; bra.uni $L__BB211_291; $L__BB211_292: ld.local.f32 %f1494, [%rd98]; st.local.f32 [%rd100], %f1494; bra.uni $L__BB211_116; $L__BB211_303: mov.u32 %r569, 58; { // callseq 1239, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r569; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1239 bra.uni $L__BB211_116; $L__BB211_308: { // callseq 1248, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1248 bra.uni $L__BB211_116; $L__BB211_283: add.s32 %r546, %r77, 1; { // callseq 1214, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r546; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1214 bra.uni $L__BB211_116; $L__BB211_298: mov.u32 %r558, 67; { // callseq 1230, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r558, %r533, %r534, %r535}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1230 bra.uni $L__BB211_116; $L__BB211_314: { // callseq 1259, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r533; .param .b32 param3; st.param.b32 [param3+0], %r534; .param .b32 param4; st.param.b32 [param4+0], %r535; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1259 bra.uni $L__BB211_116; $L__BB211_294: { // callseq 1223, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1223 bra.uni $L__BB211_116; $L__BB211_304: mov.u32 %r572, 56; { // callseq 1241, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r572; st.param.b32 [param1+4], %r533; st.param.b32 [param1+8], %r534; st.param.b32 [param1+12], %r535; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1241 bra.uni $L__BB211_116; $L__BB211_309: cvt.u32.u64 %r575, %rd2; mov.u32 %r576, 768; { // callseq 1250, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r575; .param .b64 param1; st.param.b64 [param1+0], %rd166; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .b32 param3; st.param.b32 [param3+0], %r533; .param .b32 param4; st.param.b32 [param4+0], %r534; .param .b32 param5; st.param.b32 [param5+0], %r576; call.uni _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 1250 bra.uni $L__BB211_116; $L__BB211_284: { // callseq 1216, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1216 bra.uni $L__BB211_116; $L__BB211_299: { // callseq 1232, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1232 bra.uni $L__BB211_116; $L__BB211_316: add.s32 %r588, %r77, 1; { // callseq 1263, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r533; .param .b32 param2; st.param.b32 [param2+0], %r534; .param .b32 param3; st.param.b32 [param3+0], %r535; .param .b32 param4; st.param.b32 [param4+0], %r588; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r890, [retval0+0]; } // callseq 1263 bra.uni $L__BB211_116; $L__BB211_293: ld.local.f32 %f1495, [%rd96]; st.local.f32 [%rd100], %f1495; bra.uni $L__BB211_116; $L__BB211_291: ld.local.f32 %f1493, [%rd99]; st.local.f32 [%rd100], %f1493; bra.uni $L__BB211_116; $L__BB211_329: shl.b64 %rd1165, %rd2, 2; mov.u64 %rd1164, kernel_params; add.s32 %r878, %r878, 1; mul.wide.s32 %rd458, %r878, 16; add.s64 %rd460, %rd1164, %rd458; add.s64 %rd1194, %rd460, 3744; ld.const.u64 %rd461, [%rd460+3752]; cvta.to.global.u64 %rd462, %rd461; add.s64 %rd464, %rd462, %rd1165; ld.global.u32 %r877, [%rd464]; setp.ne.s32 %p218, %r877, -1; @%p218 bra $L__BB211_29; $L__BB211_330: mov.pred %p534, -1; ld.local.u8 %rs32, [%rd31]; and.b16 %rs33, %rs32, 64; setp.eq.s16 %p220, %rs33, 0; mov.f32 %f3008, 0f00000000; mov.f32 %f3009, %f3008; mov.f32 %f3010, %f3008; @%p220 bra $L__BB211_334; ld.local.u32 %r172, [%rd33]; setp.eq.s32 %p221, %r172, -1; mov.f32 %f3007, 0f3F800000; @%p221 bra $L__BB211_333; mul.wide.s32 %rd465, %r172, 208; add.s64 %rd466, %rd12, %rd465; ld.global.f32 %f3007, [%rd466+96]; $L__BB211_333: ld.local.f32 %f1512, [%rd57]; mul.ftz.f32 %f3010, %f3007, %f1512; add.s64 %rd468, %rd166, 296; cvta.to.local.u64 %rd469, %rd468; ld.local.v2.f32 {%f1513, %f1514}, [%rd469]; mul.ftz.f32 %f3009, %f3007, %f1513; mul.ftz.f32 %f3008, %f3007, %f1514; mov.pred %p534, %p222; $L__BB211_334: @%p534 bra $L__BB211_337; mul.ftz.f32 %f1517, %f43, %f3010; sub.ftz.f32 %f3013, %f3013, %f1517; mul.ftz.f32 %f1518, %f43, %f3009; sub.ftz.f32 %f3012, %f3012, %f1518; mul.ftz.f32 %f1519, %f43, %f3008; sub.ftz.f32 %f3011, %f3011, %f1519; and.b32 %r614, %r876, 7; setp.ne.s32 %p223, %r614, 0; @%p223 bra $L__BB211_337; mul.ftz.f32 %f1520, %f3013, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1521, %f1520; mul.ftz.f32 %f1522, %f3012, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1523, %f1522; mul.ftz.f32 %f1524, %f3011, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1525, %f1524; mul.ftz.f32 %f3019, %f8, %f1525; mul.ftz.f32 %f3018, %f7, %f1523; mul.ftz.f32 %f3017, %f6, %f1521; max.ftz.f32 %f1526, %f3017, %f3018; max.ftz.f32 %f1527, %f1526, %f3019; setp.lt.ftz.f32 %p224, %f1527, 0f358637BD; @%p224 bra $L__BB211_340; $L__BB211_337: add.s32 %r876, %r876, 1; setp.eq.ftz.f32 %p225, %f42, %f9; @%p225 bra $L__BB211_339; setp.lt.s32 %p226, %r876, %r875; mov.f32 %f2957, %f42; @%p226 bra $L__BB211_27; bra.uni $L__BB211_340; $L__BB211_19: min.ftz.f32 %f2954, %f2952, %f2955; mul.ftz.f32 %f873, %f2954, %f3; setp.leu.ftz.f32 %p22, %f2955, %f873; @%p22 bra $L__BB211_21; div.approx.ftz.f32 %f2954, %f2955, %f3; $L__BB211_21: ld.const.u32 %r855, [kernel_params+1596]; cvt.u32.u16 %r390, %rs6; add.s32 %r23, %r390, 7; brev.b32 %r391, %r22; mul.lo.s32 %r392, %r391, 1025551850; xor.b32 %r24, %r392, %r391; setp.eq.s32 %p23, %r855, 0; @%p23 bra $L__BB211_24; ld.const.u32 %r856, [kernel_params+1576]; mov.b32 %f874, %r371; setp.geu.ftz.f32 %p24, %f874, 0f3F800000; setp.lt.ftz.f32 %p25, %f874, 0f3F800000; selp.b32 %r393, %r856, %r21, %p25; add.s16 %rs27, %rs6, 7; and.b16 %rs28, %rs27, 255; cvt.u32.u16 %r394, %rs28; xor.b32 %r395, %r393, %r394; mul.lo.s32 %r396, %r395, -512718531; shr.u32 %r397, %r393, 16; xor.b32 %r398, %r396, %r397; bfe.u32 %r399, %r398, 4, 4; shr.u32 %r400, %r393, 8; xor.b32 %r401, %r398, %r400; xor.b32 %r402, %r401, %r399; mul.lo.s32 %r403, %r402, 153742143; shr.u32 %r404, %r393, 23; xor.b32 %r405, %r403, %r404; bfe.u32 %r406, %r405, 1, 7; xor.b32 %r407, %r406, %r405; shr.u32 %r408, %r393, 27; or.b32 %r409, %r408, 1; mul.lo.s32 %r410, %r409, %r407; mul.lo.s32 %r411, %r410, -1074439621; bfe.u32 %r412, %r411, 2, 6; xor.b32 %r413, %r412, %r411; mul.lo.s32 %r414, %r413, -1638916925; bfe.u32 %r415, %r414, 2, 6; xor.b32 %r416, %r415, %r414; mul.lo.s32 %r417, %r416, -933190689; and.b32 %r418, %r417, 255; bfe.u32 %r419, %r417, 5, 3; xor.b32 %r420, %r419, %r418; xor.b32 %r25, %r23, 61; xor.b32 %r421, %r393, %r25; mul.lo.s32 %r422, %r421, 9; shr.u32 %r423, %r422, 4; xor.b32 %r424, %r423, %r422; mul.lo.s32 %r425, %r424, 668265261; add.s32 %r426, %r425, %r24; shr.u32 %r427, %r425, 16; or.b32 %r428, %r427, 1; mul.lo.s32 %r429, %r428, %r426; mul.lo.s32 %r430, %r429, 89287766; xor.b32 %r431, %r430, %r429; mul.lo.s32 %r432, %r431, 1403136100; xor.b32 %r433, %r432, %r431; brev.b32 %r434, %r433; neg.s32 %r435, %r370; and.b32 %r436, %r22, %r435; add.s32 %r437, %r370, -1; and.b32 %r438, %r434, %r437; or.b32 %r439, %r438, %r436; mad.lo.s32 %r440, %r420, %r370, %r439; shl.b32 %r441, %r370, 8; rem.u32 %r442, %r440, %r441; shl.b32 %r443, %r442, 1; mul.wide.u32 %rd287, %r443, 4; add.s64 %rd288, %rd62, %rd287; ld.global.f32 %f2956, [%rd288]; mov.u32 %r875, %r8; mov.f32 %f2955, %f2954; @%p24 bra $L__BB211_25; xor.b32 %r444, %r25, %r21; mul.lo.s32 %r445, %r444, 9; shr.u32 %r446, %r445, 4; xor.b32 %r447, %r446, %r445; mul.lo.s32 %r448, %r447, 668265261; cvt.rn.f32.u32 %f875, %r448; mul.ftz.f32 %f876, %f875, 0f2F7FFFFE; fma.rn.ftz.f32 %f878, %f876, %f874, %f2956; cvt.rmi.ftz.f32.f32 %f879, %f878; sub.ftz.f32 %f2956, %f878, %f879; mov.u32 %r875, %r8; mov.f32 %f2955, %f2954; bra.uni $L__BB211_25; $L__BB211_339: mul.ftz.f32 %f1528, %f3013, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1529, %f1528; mul.ftz.f32 %f1530, %f3012, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1531, %f1530; mul.ftz.f32 %f1532, %f3011, 0f3FB8AA3B; ex2.approx.ftz.f32 %f1533, %f1532; mul.ftz.f32 %f3019, %f8, %f1533; mul.ftz.f32 %f3018, %f7, %f1531; mul.ftz.f32 %f3017, %f6, %f1529; $L__BB211_340: setp.eq.ftz.f32 %p227, %f3017, 0f00000000; setp.eq.ftz.f32 %p228, %f3018, 0f00000000; and.pred %p229, %p227, %p228; setp.eq.ftz.f32 %p230, %f3019, 0f00000000; and.pred %p231, %p230, %p229; @%p231 bra $L__BB211_768; st.global.f32 [%rd11], %f3017; st.global.f32 [%rd11+4], %f3018; st.global.f32 [%rd11+8], %f3019; $L__BB211_342: @%p5 bra $L__BB211_701; mul.wide.s32 %rd470, %r872, 48; mov.u64 %rd471, kernel_params; add.s64 %rd472, %rd471, %rd470; ld.const.v2.u64 {%rd473, %rd474}, [%rd472+3568]; mov.u32 %r617, 0; cvta.to.global.u64 %rd477, %rd474; shl.b64 %rd478, %rd2, 2; add.s64 %rd479, %rd477, %rd478; ld.global.u32 %r618, [%rd479]; cvt.s64.s32 %rd102, %r618; ld.const.v2.u64 {%rd480, %rd481}, [%rd472+3584]; cvta.to.global.u64 %rd484, %rd480; add.s64 %rd485, %rd484, %rd478; cvta.to.global.u64 %rd486, %rd481; add.s64 %rd487, %rd486, %rd478; ld.global.u32 %r174, [%rd487]; ld.const.v2.u64 {%rd488, %rd489}, [%rd472+3552]; cvta.to.global.u64 %rd492, %rd489; add.s64 %rd493, %rd492, %rd478; ld.global.f32 %f296, [%rd493]; cvta.to.global.u64 %rd494, %rd473; add.s64 %rd495, %rd494, %rd478; ld.global.f32 %f297, [%rd495]; cvta.to.global.u64 %rd496, %rd488; add.s64 %rd497, %rd496, %rd478; ld.global.f32 %f3046, [%rd497]; ld.global.f32 %f3050, [%rd4]; ld.global.f32 %f3051, [%rd4+4]; ld.global.f32 %f3052, [%rd4+8]; ld.global.f32 %f302, [%rd5]; ld.global.f32 %f303, [%rd5+4]; ld.global.f32 %f3092, [%rd5+8]; ld.global.f32 %f305, [%rd6]; ld.global.f32 %f306, [%rd7]; add.s64 %rd499, %rd166, 88; cvta.to.local.u64 %rd500, %rd499; ld.global.u32 %r619, [%rd485]; st.local.f32 [%rd500], %f3046; mov.b32 %r620, %f297; mov.b32 %r621, %f296; st.local.v4.u32 [%rd49], {%r174, %r621, %r620, %r619}; mul.wide.s32 %rd503, %r619, 4; add.s64 %rd504, %rd10, %rd503; ld.global.u32 %r901, [%rd504]; st.local.u32 [%rd32], %r901; st.local.u32 [%rd45], %r618; st.local.u32 [%rd31], %r617; mov.b32 %r622, %f305; mov.u32 %r623, -1; st.local.v2.u32 [%rd34], {%r623, %r622}; and.b32 %r176, %r901, 2; setp.eq.s32 %p233, %r176, 0; mul.wide.s32 %rd513, %r619, 208; add.s64 %rd103, %rd12, %rd513; @%p233 bra $L__BB211_350; ld.global.u32 %rd514, [%rd103+168]; ld.global.u32 %r624, [%rd103+152]; shl.b32 %r625, %r624, 1; cvt.rn.f32.s32 %f1534, %r625; mul.ftz.f32 %f1535, %f305, %f1534; cvt.rzi.ftz.s32.f32 %r626, %f1535; add.s32 %r627, %r625, -1; min.s32 %r628, %r626, %r627; cvt.rn.f32.s32 %f1536, %r628; sub.ftz.f32 %f307, %f1535, %f1536; cvt.s64.s32 %rd515, %r628; add.s64 %rd516, %rd515, %rd514; cvta.to.global.u64 %rd517, %rd168; shl.b64 %rd518, %rd516, 6; add.s64 %rd104, %rd517, %rd518; ld.global.v4.f32 {%f1537, %f1538, %f1539, %f1540}, [%rd104]; ld.global.v4.f32 {%f1541, %f1542, %f1543, %f1544}, [%rd104+64]; mul.ftz.f32 %f1545, %f1538, %f1542; fma.rn.ftz.f32 %f1546, %f1537, %f1541, %f1545; mul.ftz.f32 %f1547, %f1540, %f1544; fma.rn.ftz.f32 %f1548, %f1539, %f1543, %f1547; add.ftz.f32 %f316, %f1546, %f1548; setp.gt.ftz.f32 %p234, %f316, 0f3F7FDF3B; @%p234 bra $L__BB211_346; bra.uni $L__BB211_345; $L__BB211_346: sub.ftz.f32 %f2949, %f861, %f307; mul.ftz.f32 %f1603, %f1537, %f2949; fma.rn.ftz.f32 %f1604, %f1541, %f307, %f1603; mul.ftz.f32 %f1605, %f1538, %f2949; fma.rn.ftz.f32 %f1606, %f1542, %f307, %f1605; mul.ftz.f32 %f1607, %f1539, %f2949; fma.rn.ftz.f32 %f1608, %f1543, %f307, %f1607; mul.ftz.f32 %f1609, %f1540, %f2949; fma.rn.ftz.f32 %f1610, %f1544, %f307, %f1609; mul.ftz.f32 %f1611, %f1606, %f1606; fma.rn.ftz.f32 %f1612, %f1604, %f1604, %f1611; mul.ftz.f32 %f1613, %f1610, %f1610; fma.rn.ftz.f32 %f1614, %f1608, %f1608, %f1613; add.ftz.f32 %f1615, %f1612, %f1614; rsqrt.approx.ftz.f32 %f1616, %f1615; mul.ftz.f32 %f3023, %f1610, %f1616; mul.ftz.f32 %f3022, %f1608, %f1616; mul.ftz.f32 %f3021, %f1606, %f1616; mul.ftz.f32 %f3020, %f1604, %f1616; bra.uni $L__BB211_347; $L__BB211_345: mov.f32 %f1550, 0fBF800000; max.ftz.f32 %f1551, %f316, %f1550; min.ftz.f32 %f1553, %f1551, %f861; abs.ftz.f32 %f1554, %f1553; sub.ftz.f32 %f1555, %f861, %f1554; mul.ftz.f32 %f1556, %f1555, 0f3F000000; sqrt.approx.ftz.f32 %f1557, %f1556; setp.gt.ftz.f32 %p235, %f1554, 0f3F11EB85; selp.f32 %f1558, %f1557, %f1554, %p235; mul.ftz.f32 %f1559, %f1558, %f1558; mov.f32 %f1560, 0f3C94D2E9; mov.f32 %f1561, 0f3D53F941; fma.rn.ftz.f32 %f1562, %f1561, %f1559, %f1560; mov.f32 %f1563, 0f3D3F841F; fma.rn.ftz.f32 %f1564, %f1562, %f1559, %f1563; mov.f32 %f1565, 0f3D994929; fma.rn.ftz.f32 %f1566, %f1564, %f1559, %f1565; mov.f32 %f1567, 0f3E2AAB94; fma.rn.ftz.f32 %f1568, %f1566, %f1559, %f1567; mul.ftz.f32 %f1569, %f1559, %f1568; fma.rn.ftz.f32 %f1570, %f1569, %f1558, %f1558; add.ftz.f32 %f1571, %f1570, %f1570; mov.f32 %f1572, 0f3FC90FDB; sub.ftz.f32 %f1573, %f1572, %f1570; selp.f32 %f1574, %f1571, %f1573, %p235; setp.lt.ftz.f32 %p236, %f1553, 0f00000000; mov.f32 %f1575, 0f40490FDB; sub.ftz.f32 %f1576, %f1575, %f1574; selp.f32 %f1577, %f1576, %f1574, %p236; mul.ftz.f32 %f1578, %f1537, %f316; sub.ftz.f32 %f1579, %f1541, %f1578; mul.ftz.f32 %f1580, %f1538, %f316; sub.ftz.f32 %f1581, %f1542, %f1580; mul.ftz.f32 %f1582, %f1539, %f316; sub.ftz.f32 %f1583, %f1543, %f1582; mul.ftz.f32 %f1584, %f1540, %f316; sub.ftz.f32 %f1585, %f1544, %f1584; mul.ftz.f32 %f1586, %f1581, %f1581; fma.rn.ftz.f32 %f1587, %f1579, %f1579, %f1586; mul.ftz.f32 %f1588, %f1585, %f1585; fma.rn.ftz.f32 %f1589, %f1583, %f1583, %f1588; add.ftz.f32 %f1590, %f1587, %f1589; rsqrt.approx.ftz.f32 %f1591, %f1590; mul.ftz.f32 %f1592, %f1579, %f1591; mul.ftz.f32 %f1593, %f1581, %f1591; mul.ftz.f32 %f1594, %f1583, %f1591; mul.ftz.f32 %f1595, %f1585, %f1591; mul.ftz.f32 %f1596, %f307, %f1577; cos.approx.ftz.f32 %f1597, %f1596; sin.approx.ftz.f32 %f1598, %f1596; mul.ftz.f32 %f1599, %f1592, %f1598; mul.ftz.f32 %f1600, %f1593, %f1598; mul.ftz.f32 %f1601, %f1594, %f1598; mul.ftz.f32 %f1602, %f1595, %f1598; fma.rn.ftz.f32 %f3023, %f1540, %f1597, %f1602; fma.rn.ftz.f32 %f3022, %f1539, %f1597, %f1601; fma.rn.ftz.f32 %f3021, %f1538, %f1597, %f1600; fma.rn.ftz.f32 %f3020, %f1537, %f1597, %f1599; $L__BB211_347: sub.ftz.f32 %f2948, %f861, %f307; ld.global.v4.f32 {%f1617, %f1618, %f1619, %f1620}, [%rd104+16]; ld.global.v4.f32 {%f1625, %f1626, %f1627, %f1628}, [%rd104+80]; mul.ftz.f32 %f1633, %f307, %f1625; mul.ftz.f32 %f1634, %f307, %f1626; mul.ftz.f32 %f1635, %f307, %f1627; mul.ftz.f32 %f1636, %f307, %f1628; fma.rn.ftz.f32 %f330, %f2948, %f1617, %f1633; fma.rn.ftz.f32 %f331, %f2948, %f1618, %f1634; fma.rn.ftz.f32 %f332, %f2948, %f1619, %f1635; fma.rn.ftz.f32 %f1637, %f2948, %f1620, %f1636; ld.global.v4.f32 {%f1638, %f1639, %f1640, %f1641}, [%rd104+32]; ld.global.v4.f32 {%f1646, %f1647, %f1648, %f1649}, [%rd104+96]; mul.ftz.f32 %f1654, %f307, %f1646; mul.ftz.f32 %f1655, %f307, %f1647; mul.ftz.f32 %f1656, %f307, %f1648; mul.ftz.f32 %f1657, %f307, %f1649; fma.rn.ftz.f32 %f1658, %f2948, %f1638, %f1654; fma.rn.ftz.f32 %f1659, %f2948, %f1639, %f1655; fma.rn.ftz.f32 %f1660, %f2948, %f1640, %f1656; fma.rn.ftz.f32 %f1661, %f2948, %f1641, %f1657; ld.global.v4.f32 {%f1662, %f1663, %f1664, %f1665}, [%rd104+48]; ld.global.v4.f32 {%f1670, %f1671, %f1672, %f1673}, [%rd104+112]; mul.ftz.f32 %f1678, %f307, %f1670; mul.ftz.f32 %f1679, %f307, %f1671; mul.ftz.f32 %f1680, %f307, %f1672; mul.ftz.f32 %f1681, %f307, %f1673; fma.rn.ftz.f32 %f1682, %f2948, %f1662, %f1678; fma.rn.ftz.f32 %f1683, %f2948, %f1663, %f1679; fma.rn.ftz.f32 %f1684, %f2948, %f1664, %f1680; fma.rn.ftz.f32 %f1685, %f2948, %f1665, %f1681; mul.ftz.f32 %f1686, %f3023, 0f3FB504F3; mul.ftz.f32 %f1687, %f3020, 0f3FB504F3; mul.ftz.f32 %f1688, %f3021, 0f3FB504F3; mul.ftz.f32 %f1689, %f3022, 0f3FB504F3; mul.ftz.f32 %f1690, %f1686, %f1687; mul.ftz.f32 %f1691, %f1686, %f1688; mul.ftz.f32 %f1692, %f1686, %f1689; mul.ftz.f32 %f1693, %f1687, %f1687; mul.ftz.f32 %f1694, %f1687, %f1688; mul.ftz.f32 %f1695, %f1687, %f1689; mul.ftz.f32 %f1696, %f1688, %f1688; mul.ftz.f32 %f1697, %f1688, %f1689; mul.ftz.f32 %f1698, %f1689, %f1689; sub.ftz.f32 %f1700, %f861, %f1696; sub.ftz.f32 %f1701, %f1700, %f1698; sub.ftz.f32 %f1702, %f1694, %f1692; add.ftz.f32 %f1703, %f1691, %f1695; add.ftz.f32 %f1704, %f1692, %f1694; sub.ftz.f32 %f1705, %f861, %f1693; sub.ftz.f32 %f1706, %f1705, %f1698; sub.ftz.f32 %f1707, %f1697, %f1690; sub.ftz.f32 %f1708, %f1695, %f1691; add.ftz.f32 %f1709, %f1690, %f1697; sub.ftz.f32 %f1710, %f1705, %f1696; mul.ftz.f32 %f1711, %f1702, %f1660; fma.rn.ftz.f32 %f1712, %f1701, %f1637, %f1711; fma.rn.ftz.f32 %f3029, %f1703, %f1683, %f1712; mul.ftz.f32 %f1713, %f1702, %f1661; fma.rn.ftz.f32 %f1714, %f1701, %f1658, %f1713; fma.rn.ftz.f32 %f334, %f1703, %f1684, %f1714; mul.ftz.f32 %f1715, %f1702, %f1682; fma.rn.ftz.f32 %f1716, %f1701, %f1659, %f1715; fma.rn.ftz.f32 %f335, %f1703, %f1685, %f1716; mul.ftz.f32 %f1717, %f1706, %f1660; fma.rn.ftz.f32 %f1718, %f1704, %f1637, %f1717; fma.rn.ftz.f32 %f336, %f1707, %f1683, %f1718; mul.ftz.f32 %f1719, %f1706, %f1661; fma.rn.ftz.f32 %f1720, %f1704, %f1658, %f1719; fma.rn.ftz.f32 %f3028, %f1707, %f1684, %f1720; mul.ftz.f32 %f1721, %f1706, %f1682; fma.rn.ftz.f32 %f1722, %f1704, %f1659, %f1721; fma.rn.ftz.f32 %f338, %f1707, %f1685, %f1722; mul.ftz.f32 %f1723, %f1709, %f1660; fma.rn.ftz.f32 %f1724, %f1708, %f1637, %f1723; fma.rn.ftz.f32 %f339, %f1710, %f1683, %f1724; mul.ftz.f32 %f1725, %f1709, %f1661; fma.rn.ftz.f32 %f1726, %f1708, %f1658, %f1725; fma.rn.ftz.f32 %f340, %f1710, %f1684, %f1726; mul.ftz.f32 %f1727, %f1709, %f1682; fma.rn.ftz.f32 %f1728, %f1708, %f1659, %f1727; fma.rn.ftz.f32 %f3027, %f1710, %f1685, %f1728; cvta.to.local.u64 %rd521, %rd44; st.local.v4.f32 [%rd521], {%f3029, %f334, %f335, %f330}; cvta.to.local.u64 %rd523, %rd47; st.local.v4.f32 [%rd523], {%f336, %f3028, %f338, %f331}; cvta.to.local.u64 %rd525, %rd46; st.local.v4.f32 [%rd525], {%f339, %f340, %f3027, %f332}; mul.ftz.f32 %f1729, %f3028, %f3027; mul.ftz.f32 %f342, %f340, %f338; sub.ftz.f32 %f3026, %f1729, %f342; mul.ftz.f32 %f344, %f340, %f335; mul.ftz.f32 %f1730, %f334, %f3027; sub.ftz.f32 %f3025, %f344, %f1730; mul.ftz.f32 %f346, %f334, %f338; mul.ftz.f32 %f1731, %f3028, %f335; sub.ftz.f32 %f3024, %f346, %f1731; mul.ftz.f32 %f1732, %f336, %f3025; fma.rn.ftz.f32 %f1733, %f3029, %f3026, %f1732; fma.rn.ftz.f32 %f3030, %f339, %f3024, %f1733; setp.neu.ftz.f32 %p237, %f3030, 0f00000000; @%p237 bra $L__BB211_349; add.ftz.f32 %f3029, %f3029, 0f322BCC77; add.ftz.f32 %f3027, %f3027, 0f322BCC77; add.ftz.f32 %f3028, %f3028, 0f322BCC77; mul.ftz.f32 %f1734, %f3028, %f3027; sub.ftz.f32 %f3026, %f1734, %f342; mul.ftz.f32 %f1735, %f334, %f3027; sub.ftz.f32 %f3025, %f344, %f1735; mul.ftz.f32 %f1736, %f3028, %f335; sub.ftz.f32 %f3024, %f346, %f1736; mul.ftz.f32 %f1737, %f336, %f3025; fma.rn.ftz.f32 %f1738, %f3029, %f3026, %f1737; fma.rn.ftz.f32 %f1739, %f339, %f3024, %f1738; setp.eq.ftz.f32 %p238, %f1739, 0f00000000; selp.f32 %f3030, 0f7F7FFFFF, %f1739, %p238; $L__BB211_349: mul.ftz.f32 %f1740, %f336, %f3027; mul.ftz.f32 %f1741, %f339, %f338; sub.ftz.f32 %f1742, %f1741, %f1740; mul.ftz.f32 %f1743, %f339, %f335; mul.ftz.f32 %f1744, %f3027, %f3029; sub.ftz.f32 %f1745, %f1744, %f1743; mul.ftz.f32 %f1746, %f338, %f3029; mul.ftz.f32 %f1747, %f336, %f335; sub.ftz.f32 %f1748, %f1747, %f1746; mul.ftz.f32 %f1749, %f339, %f3028; mul.ftz.f32 %f1750, %f336, %f340; sub.ftz.f32 %f1751, %f1750, %f1749; mul.ftz.f32 %f1752, %f340, %f3029; mul.ftz.f32 %f1753, %f339, %f334; sub.ftz.f32 %f1754, %f1753, %f1752; mul.ftz.f32 %f1755, %f336, %f334; mul.ftz.f32 %f1756, %f3028, %f3029; sub.ftz.f32 %f1757, %f1756, %f1755; div.approx.ftz.f32 %f1758, %f3025, %f3030; mul.ftz.f32 %f1759, %f331, %f1758; div.approx.ftz.f32 %f1760, %f3026, %f3030; fma.rn.ftz.f32 %f1761, %f330, %f1760, %f1759; div.approx.ftz.f32 %f1762, %f3024, %f3030; fma.rn.ftz.f32 %f1763, %f332, %f1762, %f1761; div.approx.ftz.f32 %f1764, %f1745, %f3030; mul.ftz.f32 %f1765, %f331, %f1764; div.approx.ftz.f32 %f1766, %f1742, %f3030; fma.rn.ftz.f32 %f1767, %f330, %f1766, %f1765; div.approx.ftz.f32 %f1768, %f1748, %f3030; fma.rn.ftz.f32 %f1769, %f332, %f1768, %f1767; div.approx.ftz.f32 %f1770, %f1754, %f3030; mul.ftz.f32 %f1771, %f331, %f1770; div.approx.ftz.f32 %f1772, %f1751, %f3030; fma.rn.ftz.f32 %f1773, %f330, %f1772, %f1771; div.approx.ftz.f32 %f1774, %f1757, %f3030; fma.rn.ftz.f32 %f1775, %f332, %f1774, %f1773; neg.ftz.f32 %f1776, %f1763; cvta.to.local.u64 %rd528, %rd48; st.local.v4.f32 [%rd528], {%f1760, %f1758, %f1762, %f1776}; neg.ftz.f32 %f1777, %f1769; cvta.to.local.u64 %rd530, %rd51; st.local.v4.f32 [%rd530], {%f1766, %f1764, %f1768, %f1777}; neg.ftz.f32 %f1778, %f1775; cvta.to.local.u64 %rd532, %rd50; st.local.v4.f32 [%rd532], {%f1772, %f1770, %f1774, %f1778}; $L__BB211_350: neg.ftz.f32 %f363, %f302; add.s64 %rd534, %rd166, 36; cvta.to.local.u64 %rd535, %rd534; st.local.f32 [%rd535], %f363; add.s64 %rd536, %rd166, 40; cvta.to.local.u64 %rd537, %rd536; neg.ftz.f32 %f364, %f303; st.local.f32 [%rd537], %f364; add.s64 %rd538, %rd166, 44; cvta.to.local.u64 %rd539, %rd538; neg.ftz.f32 %f1779, %f3092; st.local.f32 [%rd539], %f1779; and.b32 %r629, %r174, 6; setp.eq.s32 %p239, %r629, 0; @%p239 bra $L__BB211_390; and.b32 %r177, %r901, 4; setp.ne.s32 %p240, %r177, 0; mov.f32 %f3047, %f302; mov.f32 %f3048, %f303; mov.f32 %f3049, %f3092; @%p240 bra $L__BB211_358; @%p233 bra $L__BB211_354; bra.uni $L__BB211_353; $L__BB211_354: ld.global.v4.f32 {%f3031, %f3032, %f3033, %f3034}, [%rd103+48]; ld.global.v4.f32 {%f3035, %f3036, %f3037, %f3038}, [%rd103+64]; ld.global.v4.f32 {%f3039, %f3040, %f3041, %f3042}, [%rd103+80]; bra.uni $L__BB211_355; $L__BB211_390: and.b32 %r656, %r174, 8; setp.eq.s32 %p264, %r656, 0; @%p264 bra $L__BB211_411; bra.uni $L__BB211_391; $L__BB211_411: setp.eq.s32 %p277, %r174, 1; @%p277 bra $L__BB211_413; bra.uni $L__BB211_412; $L__BB211_413: cvt.u32.u64 %r675, %rd102; mul.wide.s32 %rd642, %r675, 16; add.s64 %rd643, %rd641, %rd642; add.s64 %rd122, %rd643, 12; ld.global.u32 %r676, [%rd643+12]; mul.wide.u32 %rd645, %r676, 12; add.s64 %rd646, %rd644, %rd645; add.s32 %r677, %r676, 1; mul.wide.u32 %rd647, %r677, 12; add.s64 %rd648, %rd644, %rd647; add.s32 %r678, %r676, 2; mul.wide.u32 %rd649, %r678, 12; add.s64 %rd650, %rd644, %rd649; and.b32 %r679, %r901, 8; setp.eq.s32 %p278, %r679, 0; ld.global.f32 %f2333, [%rd650]; ld.global.f32 %f642, [%rd646]; sub.ftz.f32 %f643, %f2333, %f642; ld.global.f32 %f2334, [%rd650+4]; ld.global.f32 %f644, [%rd646+4]; sub.ftz.f32 %f645, %f2334, %f644; ld.global.f32 %f2335, [%rd650+8]; ld.global.f32 %f646, [%rd646+8]; sub.ftz.f32 %f647, %f2335, %f646; ld.global.f32 %f2336, [%rd648]; sub.ftz.f32 %f648, %f2336, %f642; ld.global.f32 %f2337, [%rd648+4]; sub.ftz.f32 %f649, %f2337, %f644; ld.global.f32 %f2338, [%rd648+8]; sub.ftz.f32 %f650, %f2338, %f646; mul.ftz.f32 %f651, %f650, %f645; mul.ftz.f32 %f652, %f649, %f647; mul.ftz.f32 %f653, %f648, %f647; mul.ftz.f32 %f654, %f650, %f643; mul.ftz.f32 %f655, %f649, %f643; mul.ftz.f32 %f656, %f648, %f645; @%p278 bra $L__BB211_415; bra.uni $L__BB211_414; $L__BB211_415: sub.ftz.f32 %f2346, %f652, %f651; sub.ftz.f32 %f2347, %f654, %f653; mul.ftz.f32 %f2348, %f2347, %f2347; fma.rn.ftz.f32 %f2349, %f2346, %f2346, %f2348; sub.ftz.f32 %f2350, %f656, %f655; fma.rn.ftz.f32 %f2351, %f2350, %f2350, %f2349; sqrt.approx.ftz.f32 %f2352, %f2351; div.approx.ftz.f32 %f3117, %f2350, %f2352; div.approx.ftz.f32 %f3116, %f2347, %f2352; div.approx.ftz.f32 %f3115, %f2346, %f2352; bra.uni $L__BB211_416; $L__BB211_391: shl.b64 %rd608, %rd102, 2; add.s64 %rd609, %rd607, %rd608; ld.global.u32 %r657, [%rd609]; st.local.u32 [%rd18], %r657; fma.rn.ftz.f32 %f605, %f3046, %f302, %f3050; fma.rn.ftz.f32 %f606, %f3046, %f303, %f3051; fma.rn.ftz.f32 %f607, %f3046, %f3092, %f3052; st.local.v2.f32 [%rd71], {%f605, %f606}; st.local.f32 [%rd28], %f607; and.b32 %r658, %r174, 64; setp.eq.s32 %p265, %r658, 0; @%p265 bra $L__BB211_405; bra.uni $L__BB211_392; $L__BB211_405: shl.b64 %rd622, %rd102, 4; add.s64 %rd623, %rd60, %rd622; ld.global.v4.f32 {%f3112, %f3113, %f3114, %f2258}, [%rd623]; bra.uni $L__BB211_406; $L__BB211_353: cvta.to.local.u64 %rd542, %rd48; ld.local.v4.f32 {%f3031, %f3032, %f3033, %f3034}, [%rd542]; cvta.to.local.u64 %rd544, %rd51; ld.local.v4.f32 {%f3035, %f3036, %f3037, %f3038}, [%rd544]; cvta.to.local.u64 %rd546, %rd50; ld.local.v4.f32 {%f3039, %f3040, %f3041, %f3042}, [%rd546]; $L__BB211_355: mul.ftz.f32 %f1804, %f3051, %f3032; fma.rn.ftz.f32 %f1805, %f3050, %f3031, %f1804; fma.rn.ftz.f32 %f1806, %f3052, %f3033, %f1805; add.ftz.f32 %f401, %f3034, %f1806; mul.ftz.f32 %f1807, %f3051, %f3036; fma.rn.ftz.f32 %f1808, %f3050, %f3035, %f1807; fma.rn.ftz.f32 %f1809, %f3052, %f3037, %f1808; add.ftz.f32 %f402, %f3038, %f1809; mul.ftz.f32 %f1810, %f3051, %f3040; fma.rn.ftz.f32 %f1811, %f3050, %f3039, %f1810; fma.rn.ftz.f32 %f1812, %f3052, %f3041, %f1811; add.ftz.f32 %f3052, %f3042, %f1812; mul.ftz.f32 %f1813, %f3046, %f302; mul.ftz.f32 %f1814, %f3046, %f303; mul.ftz.f32 %f1815, %f1814, %f3032; fma.rn.ftz.f32 %f1816, %f1813, %f3031, %f1815; mul.ftz.f32 %f1817, %f3046, %f3092; fma.rn.ftz.f32 %f3047, %f1817, %f3033, %f1816; mul.ftz.f32 %f1818, %f1814, %f3036; fma.rn.ftz.f32 %f1819, %f1813, %f3035, %f1818; fma.rn.ftz.f32 %f3048, %f1817, %f3037, %f1819; mul.ftz.f32 %f1820, %f1814, %f3040; fma.rn.ftz.f32 %f1821, %f1813, %f3039, %f1820; fma.rn.ftz.f32 %f3049, %f1817, %f3041, %f1821; mul.ftz.f32 %f1822, %f3048, %f3048; fma.rn.ftz.f32 %f1823, %f3047, %f3047, %f1822; fma.rn.ftz.f32 %f1824, %f3049, %f3049, %f1823; sqrt.approx.ftz.f32 %f3046, %f1824; setp.neu.ftz.f32 %p242, %f3046, 0f00000000; @%p242 bra $L__BB211_356; bra.uni $L__BB211_357; $L__BB211_356: div.approx.ftz.f32 %f3049, %f3049, %f3046; div.approx.ftz.f32 %f3048, %f3048, %f3046; div.approx.ftz.f32 %f3047, %f3047, %f3046; $L__BB211_357: mov.f32 %f3050, %f401; mov.f32 %f3051, %f402; $L__BB211_358: shl.b64 %rd548, %rd102, 4; add.s64 %rd549, %rd547, %rd548; add.s64 %rd105, %rd549, 4; shr.s32 %r630, %r174, 7; ld.global.u32 %r631, [%rd549+4]; add.s32 %r178, %r630, %r631; add.s32 %r632, %r178, -1; max.s32 %r179, %r632, %r631; add.s32 %r633, %r178, 2; ld.global.u32 %r634, [%rd549+8]; add.s32 %r635, %r631, %r634; add.s32 %r636, %r635, -1; min.s32 %r180, %r633, %r636; and.b32 %r637, %r174, 64; setp.eq.s32 %p243, %r637, 0; mul.wide.s32 %rd550, %r179, 16; add.s64 %rd106, %rd61, %rd550; mul.wide.s32 %rd551, %r178, 16; add.s64 %rd107, %rd61, %rd551; mul.wide.s32 %rd552, %r180, 16; add.s64 %rd108, %rd61, %rd552; @%p243 bra $L__BB211_372; bra.uni $L__BB211_359; $L__BB211_372: ld.global.v4.f32 {%f3077, %f3078, %f3079, %f1909}, [%rd106]; ld.global.v4.f32 {%f3080, %f3081, %f3082, %f1913}, [%rd107]; ld.global.v4.f32 {%f3083, %f3084, %f3085, %f1917}, [%rd107+16]; ld.global.v4.f32 {%f3086, %f3087, %f3088, %f1921}, [%rd108]; bra.uni $L__BB211_373; $L__BB211_359: ld.global.u32 %r181, [%rd103+148]; ld.global.u32 %r182, [%rd103+152]; shl.b32 %r638, %r182, 1; cvt.rn.f32.s32 %f1825, %r638; mul.ftz.f32 %f427, %f305, %f1825; cvt.rzi.ftz.s32.f32 %r639, %f427; add.s32 %r640, %r638, -1; min.s32 %r183, %r639, %r640; ld.global.u32 %r892, [%rd103+164]; mul.wide.u32 %rd553, %r892, 16; add.s64 %rd554, %rd52, %rd553; ld.global.u16 %rs48, [%rd554+12]; ld.global.u32 %r891, [%rd554+8]; ld.global.u64 %rd1196, [%rd554]; setp.eq.s64 %p244, %rd1196, 11; @%p244 bra $L__BB211_364; $L__BB211_360: setp.eq.s64 %p245, %rd1196, 0; @%p245 bra $L__BB211_362; add.s32 %r892, %r892, 2; bra.uni $L__BB211_363; $L__BB211_362: mov.u32 %r895, -1; setp.eq.s16 %p246, %rs48, 0; mov.u32 %r892, %r891; @%p246 bra $L__BB211_365; $L__BB211_363: mul.wide.u32 %rd555, %r892, 16; add.s64 %rd556, %rd52, %rd555; ld.global.u16 %rs48, [%rd556+12]; ld.global.u32 %r891, [%rd556+8]; ld.global.u64 %rd1196, [%rd556]; setp.ne.s64 %p247, %rd1196, 11; @%p247 bra $L__BB211_360; $L__BB211_364: setp.eq.s16 %p248, %rs48, 0; selp.b32 %r895, -1, %r891, %p248; $L__BB211_365: setp.eq.s32 %p249, %r183, %r182; @%p249 bra $L__BB211_367; bra.uni $L__BB211_366; $L__BB211_367: ld.global.v4.f32 {%f3053, %f3054, %f3055, %f1845}, [%rd106]; ld.global.v4.f32 {%f3056, %f3057, %f3058, %f1849}, [%rd107]; ld.global.v4.f32 {%f3059, %f3060, %f3061, %f1853}, [%rd107+16]; ld.global.v4.f32 {%f3062, %f3063, %f3064, %f1857}, [%rd108]; bra.uni $L__BB211_368; $L__BB211_366: shr.s32 %r866, %r174, 7; add.s32 %r865, %r866, %r631; setp.gt.s32 %p250, %r183, %r182; selp.b32 %r642, -1, 0, %p250; add.s32 %r643, %r183, %r642; mad.lo.s32 %r644, %r643, %r181, %r895; add.s32 %r645, %r644, %r179; mul.wide.s32 %rd557, %r645, 16; add.s64 %rd558, %rd54, %rd557; ld.global.v4.f32 {%f3053, %f3054, %f3055, %f1829}, [%rd558]; add.s32 %r646, %r644, %r865; mul.wide.s32 %rd559, %r646, 16; add.s64 %rd560, %rd54, %rd559; ld.global.v4.f32 {%f3056, %f3057, %f3058, %f1833}, [%rd560]; ld.global.v4.f32 {%f3059, %f3060, %f3061, %f1837}, [%rd560+16]; add.s32 %r647, %r644, %r180; mul.wide.s32 %rd561, %r647, 16; add.s64 %rd562, %rd54, %rd561; ld.global.v4.f32 {%f3062, %f3063, %f3064, %f1841}, [%rd562]; $L__BB211_368: add.s32 %r194, %r183, 1; setp.eq.s32 %p251, %r194, %r182; @%p251 bra $L__BB211_370; bra.uni $L__BB211_369; $L__BB211_370: ld.global.v4.f32 {%f3065, %f3066, %f3067, %f1877}, [%rd106]; ld.global.v4.f32 {%f3068, %f3069, %f3070, %f1881}, [%rd107]; ld.global.v4.f32 {%f3071, %f3072, %f3073, %f1885}, [%rd107+16]; ld.global.v4.f32 {%f3074, %f3075, %f3076, %f1889}, [%rd108]; bra.uni $L__BB211_371; $L__BB211_369: shr.s32 %r868, %r174, 7; add.s32 %r867, %r868, %r631; setp.ge.s32 %p252, %r183, %r182; selp.b32 %r648, -1, 0, %p252; add.s32 %r649, %r194, %r648; mad.lo.s32 %r650, %r649, %r181, %r895; add.s32 %r651, %r650, %r179; mul.wide.s32 %rd563, %r651, 16; add.s64 %rd564, %rd54, %rd563; ld.global.v4.f32 {%f3065, %f3066, %f3067, %f1861}, [%rd564]; add.s32 %r652, %r650, %r867; mul.wide.s32 %rd565, %r652, 16; add.s64 %rd566, %rd54, %rd565; ld.global.v4.f32 {%f3068, %f3069, %f3070, %f1865}, [%rd566]; ld.global.v4.f32 {%f3071, %f3072, %f3073, %f1869}, [%rd566+16]; add.s32 %r653, %r650, %r180; mul.wide.s32 %rd567, %r653, 16; add.s64 %rd568, %rd54, %rd567; ld.global.v4.f32 {%f3074, %f3075, %f3076, %f1873}, [%rd568]; $L__BB211_371: shl.b32 %r869, %r182, 1; cvt.rn.f32.s32 %f2945, %r869; mul.ftz.f32 %f2944, %f305, %f2945; cvt.rn.f32.s32 %f1890, %r183; sub.ftz.f32 %f1891, %f2944, %f1890; sub.ftz.f32 %f1893, %f861, %f1891; mul.ftz.f32 %f1894, %f1891, %f3065; fma.rn.ftz.f32 %f3077, %f1893, %f3053, %f1894; mul.ftz.f32 %f1895, %f1891, %f3066; fma.rn.ftz.f32 %f3078, %f1893, %f3054, %f1895; mul.ftz.f32 %f1896, %f1891, %f3067; fma.rn.ftz.f32 %f3079, %f1893, %f3055, %f1896; mul.ftz.f32 %f1897, %f1891, %f3068; fma.rn.ftz.f32 %f3080, %f1893, %f3056, %f1897; mul.ftz.f32 %f1898, %f1891, %f3069; fma.rn.ftz.f32 %f3081, %f1893, %f3057, %f1898; mul.ftz.f32 %f1899, %f1891, %f3070; fma.rn.ftz.f32 %f3082, %f1893, %f3058, %f1899; mul.ftz.f32 %f1900, %f1891, %f3071; fma.rn.ftz.f32 %f3083, %f1893, %f3059, %f1900; mul.ftz.f32 %f1901, %f1891, %f3072; fma.rn.ftz.f32 %f3084, %f1893, %f3060, %f1901; mul.ftz.f32 %f1902, %f1891, %f3073; fma.rn.ftz.f32 %f3085, %f1893, %f3061, %f1902; mul.ftz.f32 %f1903, %f1891, %f3074; fma.rn.ftz.f32 %f3086, %f1893, %f3062, %f1903; mul.ftz.f32 %f1904, %f1891, %f3075; fma.rn.ftz.f32 %f3087, %f1893, %f3063, %f1904; mul.ftz.f32 %f1905, %f1891, %f3076; fma.rn.ftz.f32 %f3088, %f1893, %f3064, %f1905; $L__BB211_373: fma.rn.ftz.f32 %f3106, %f3046, %f3047, %f3050; fma.rn.ftz.f32 %f3107, %f3046, %f3048, %f3051; fma.rn.ftz.f32 %f3108, %f3046, %f3049, %f3052; sub.ftz.f32 %f539, %f861, %f296; mul.ftz.f32 %f540, %f539, %f539; add.ftz.f32 %f1923, %f539, %f539; mul.ftz.f32 %f1924, %f296, %f1923; sub.ftz.f32 %f1925, %f1924, %f540; mul.ftz.f32 %f1926, %f296, 0f40400000; add.ftz.f32 %f541, %f1926, 0fC0A00000; add.ftz.f32 %f1927, %f296, %f296; mul.ftz.f32 %f1928, %f1927, %f541; fma.rn.ftz.f32 %f1929, %f296, %f1926, %f1928; add.ftz.f32 %f1930, %f1926, 0f40000000; mul.ftz.f32 %f1931, %f1923, %f1930; mul.ftz.f32 %f542, %f539, 0f40400000; mul.ftz.f32 %f1932, %f539, %f542; sub.ftz.f32 %f1933, %f1931, %f1932; mul.ftz.f32 %f1934, %f539, 0fC0000000; mul.ftz.f32 %f543, %f296, %f296; fma.rn.ftz.f32 %f1935, %f296, %f1934, %f543; mul.ftz.f32 %f1936, %f3080, %f1929; mul.ftz.f32 %f1937, %f3081, %f1929; mul.ftz.f32 %f1938, %f3082, %f1929; fma.rn.ftz.f32 %f1939, %f3077, %f1925, %f1936; fma.rn.ftz.f32 %f1940, %f3078, %f1925, %f1937; fma.rn.ftz.f32 %f1941, %f3079, %f1925, %f1938; fma.rn.ftz.f32 %f1942, %f3083, %f1933, %f1939; fma.rn.ftz.f32 %f1943, %f3084, %f1933, %f1940; fma.rn.ftz.f32 %f1944, %f3085, %f1933, %f1941; fma.rn.ftz.f32 %f1945, %f3086, %f1935, %f1942; fma.rn.ftz.f32 %f1946, %f3087, %f1935, %f1943; fma.rn.ftz.f32 %f1947, %f3088, %f1935, %f1944; mul.ftz.f32 %f3101, %f1945, 0f3F000000; mul.ftz.f32 %f3103, %f1946, 0f3F000000; mul.ftz.f32 %f3102, %f1947, 0f3F000000; and.b32 %r195, %r174, 4; setp.eq.s32 %p253, %r195, 0; @%p253 bra $L__BB211_375; bra.uni $L__BB211_374; $L__BB211_375: mul.ftz.f32 %f2006, %f296, %f539; mul.ftz.f32 %f2007, %f539, %f2006; fma.rn.ftz.f32 %f2008, %f543, %f541, 0f40000000; add.ftz.f32 %f2009, %f542, 0fC0A00000; fma.rn.ftz.f32 %f2010, %f540, %f2009, 0f40000000; mul.ftz.f32 %f2011, %f296, %f2006; mul.ftz.f32 %f2012, %f3077, %f2007; mul.ftz.f32 %f2013, %f3078, %f2007; mul.ftz.f32 %f2014, %f3079, %f2007; mul.ftz.f32 %f2015, %f3080, %f2008; mul.ftz.f32 %f2016, %f3081, %f2008; mul.ftz.f32 %f2017, %f3082, %f2008; sub.ftz.f32 %f2018, %f2015, %f2012; sub.ftz.f32 %f2019, %f2016, %f2013; sub.ftz.f32 %f2020, %f2017, %f2014; fma.rn.ftz.f32 %f2021, %f3083, %f2010, %f2018; fma.rn.ftz.f32 %f2022, %f3084, %f2010, %f2019; fma.rn.ftz.f32 %f2023, %f3085, %f2010, %f2020; mul.ftz.f32 %f2024, %f3086, %f2011; mul.ftz.f32 %f2025, %f3087, %f2011; mul.ftz.f32 %f2026, %f3088, %f2011; sub.ftz.f32 %f2027, %f2021, %f2024; sub.ftz.f32 %f2028, %f2022, %f2025; sub.ftz.f32 %f2029, %f2023, %f2026; mul.ftz.f32 %f550, %f2027, 0f3F000000; mul.ftz.f32 %f551, %f2028, 0f3F000000; mul.ftz.f32 %f552, %f2029, 0f3F000000; setp.eq.ftz.f32 %p254, %f3106, %f550; setp.eq.ftz.f32 %p255, %f3107, %f551; and.pred %p256, %p254, %p255; setp.eq.ftz.f32 %p257, %f3108, %f552; and.pred %p258, %p257, %p256; @%p258 bra $L__BB211_377; bra.uni $L__BB211_376; $L__BB211_377: mov.f32 %f3105, %f302; mov.f32 %f3104, %f303; bra.uni $L__BB211_378; $L__BB211_374: mul.ftz.f32 %f1948, %f3103, %f3103; fma.rn.ftz.f32 %f1949, %f3101, %f3101, %f1948; fma.rn.ftz.f32 %f1950, %f3102, %f3102, %f1949; sqrt.approx.ftz.f32 %f1951, %f1950; div.approx.ftz.f32 %f1952, %f3101, %f1951; div.approx.ftz.f32 %f1953, %f3103, %f1951; div.approx.ftz.f32 %f1954, %f3102, %f1951; mul.ftz.f32 %f1955, %f1953, %f3049; mul.ftz.f32 %f1956, %f1954, %f3048; sub.ftz.f32 %f1957, %f1956, %f1955; mul.ftz.f32 %f1958, %f1954, %f3047; mul.ftz.f32 %f1959, %f1952, %f3049; sub.ftz.f32 %f1960, %f1959, %f1958; mul.ftz.f32 %f1961, %f1952, %f3048; mul.ftz.f32 %f1962, %f1953, %f3047; sub.ftz.f32 %f1963, %f1962, %f1961; mul.ftz.f32 %f1964, %f1960, %f1960; fma.rn.ftz.f32 %f1965, %f1957, %f1957, %f1964; fma.rn.ftz.f32 %f1966, %f1963, %f1963, %f1965; sqrt.approx.ftz.f32 %f1967, %f1966; div.approx.ftz.f32 %f1968, %f1957, %f1967; div.approx.ftz.f32 %f1969, %f1960, %f1967; div.approx.ftz.f32 %f1970, %f1963, %f1967; mul.ftz.f32 %f1971, %f297, %f297; sub.ftz.f32 %f1973, %f861, %f1971; mov.f32 %f1974, 0f00000000; max.ftz.f32 %f1975, %f1973, %f1974; sqrt.approx.ftz.f32 %f1976, %f1975; mul.ftz.f32 %f1977, %f1968, %f297; mul.ftz.f32 %f1978, %f1969, %f297; mul.ftz.f32 %f1979, %f1970, %f297; mul.ftz.f32 %f1980, %f1953, %f1970; mul.ftz.f32 %f1981, %f1954, %f1969; sub.ftz.f32 %f1982, %f1980, %f1981; mul.ftz.f32 %f1983, %f1954, %f1968; mul.ftz.f32 %f1984, %f1952, %f1970; sub.ftz.f32 %f1985, %f1983, %f1984; mul.ftz.f32 %f1986, %f1952, %f1969; mul.ftz.f32 %f1987, %f1953, %f1968; sub.ftz.f32 %f1988, %f1986, %f1987; mul.ftz.f32 %f1989, %f1985, %f1985; fma.rn.ftz.f32 %f1990, %f1982, %f1982, %f1989; fma.rn.ftz.f32 %f1991, %f1988, %f1988, %f1990; sqrt.approx.ftz.f32 %f1992, %f1991; div.approx.ftz.f32 %f1993, %f1982, %f1992; div.approx.ftz.f32 %f1994, %f1985, %f1992; div.approx.ftz.f32 %f1995, %f1988, %f1992; mul.ftz.f32 %f1996, %f1976, %f1993; mul.ftz.f32 %f1997, %f1976, %f1994; mul.ftz.f32 %f1998, %f1976, %f1995; sub.ftz.f32 %f1999, %f1977, %f1996; sub.ftz.f32 %f2000, %f1978, %f1997; sub.ftz.f32 %f2001, %f1979, %f1998; mul.ftz.f32 %f2002, %f2000, %f2000; fma.rn.ftz.f32 %f2003, %f1999, %f1999, %f2002; fma.rn.ftz.f32 %f2004, %f2001, %f2001, %f2003; sqrt.approx.ftz.f32 %f2005, %f2004; div.approx.ftz.f32 %f3105, %f1999, %f2005; div.approx.ftz.f32 %f3104, %f2000, %f2005; div.approx.ftz.f32 %f3092, %f2001, %f2005; st.local.f32 [%rd22], %f3105; st.local.f32 [%rd64], %f3104; st.local.f32 [%rd65], %f3092; bra.uni $L__BB211_379; $L__BB211_376: sub.ftz.f32 %f2030, %f3106, %f550; sub.ftz.f32 %f2031, %f3107, %f551; mul.ftz.f32 %f2032, %f2031, %f2031; fma.rn.ftz.f32 %f2033, %f2030, %f2030, %f2032; sub.ftz.f32 %f2034, %f3108, %f552; fma.rn.ftz.f32 %f2035, %f2034, %f2034, %f2033; sqrt.approx.ftz.f32 %f2036, %f2035; div.approx.ftz.f32 %f3092, %f2034, %f2036; div.approx.ftz.f32 %f3104, %f2031, %f2036; div.approx.ftz.f32 %f3105, %f2030, %f2036; $L__BB211_378: mov.u32 %r864, 0; st.local.f32 [%rd22], %f3105; st.local.f32 [%rd64], %f3104; st.local.f32 [%rd65], %f3092; add.s64 %rd570, %rd166, 72; cvta.to.local.u64 %rd571, %rd570; st.local.u32 [%rd571], %r864; $L__BB211_379: and.b32 %r870, %r901, 4; setp.ne.s32 %p533, %r870, 0; st.local.f32 [%rd23], %f3101; st.local.f32 [%rd24], %f3103; st.local.f32 [%rd25], %f3102; @%p533 bra $L__BB211_389; @%p233 bra $L__BB211_382; bra.uni $L__BB211_381; $L__BB211_382: ld.global.v4.f32 {%f2070, %f2071, %f2072, %f2073}, [%rd103]; ld.global.v4.f32 {%f2078, %f2079, %f2080, %f2081}, [%rd103+16]; ld.global.v4.f32 {%f2086, %f2087, %f2088, %f2089}, [%rd103+32]; mul.ftz.f32 %f2094, %f3107, %f2071; fma.rn.ftz.f32 %f2095, %f3106, %f2070, %f2094; fma.rn.ftz.f32 %f2096, %f3108, %f2072, %f2095; add.ftz.f32 %f3095, %f2073, %f2096; mul.ftz.f32 %f2097, %f3107, %f2079; fma.rn.ftz.f32 %f2098, %f3106, %f2078, %f2097; fma.rn.ftz.f32 %f2099, %f3108, %f2080, %f2098; add.ftz.f32 %f3096, %f2081, %f2099; mul.ftz.f32 %f2100, %f3107, %f2087; fma.rn.ftz.f32 %f2101, %f3106, %f2086, %f2100; fma.rn.ftz.f32 %f2102, %f3108, %f2088, %f2101; add.ftz.f32 %f3108, %f2089, %f2102; bra.uni $L__BB211_383; $L__BB211_381: cvta.to.local.u64 %rd574, %rd44; ld.local.v4.f32 {%f2037, %f2038, %f2039, %f2040}, [%rd574]; mul.ftz.f32 %f2045, %f3107, %f2038; fma.rn.ftz.f32 %f2046, %f3106, %f2037, %f2045; fma.rn.ftz.f32 %f2047, %f3108, %f2039, %f2046; add.ftz.f32 %f3095, %f2040, %f2047; cvta.to.local.u64 %rd576, %rd47; ld.local.v4.f32 {%f2048, %f2049, %f2050, %f2051}, [%rd576]; mul.ftz.f32 %f2056, %f3107, %f2049; fma.rn.ftz.f32 %f2057, %f3106, %f2048, %f2056; fma.rn.ftz.f32 %f2058, %f3108, %f2050, %f2057; add.ftz.f32 %f3096, %f2051, %f2058; cvta.to.local.u64 %rd578, %rd46; ld.local.v4.f32 {%f2059, %f2060, %f2061, %f2062}, [%rd578]; mul.ftz.f32 %f2067, %f3107, %f2060; fma.rn.ftz.f32 %f2068, %f3106, %f2059, %f2067; fma.rn.ftz.f32 %f2069, %f3108, %f2061, %f2068; add.ftz.f32 %f3108, %f2062, %f2069; $L__BB211_383: @%p233 bra $L__BB211_385; bra.uni $L__BB211_384; $L__BB211_385: ld.global.v4.f32 {%f2137, %f2138, %f2139, %f2140}, [%rd103+48]; ld.global.v4.f32 {%f2144, %f2145, %f2146, %f2147}, [%rd103+64]; ld.global.v4.f32 {%f2151, %f2152, %f2153, %f2154}, [%rd103+80]; mul.ftz.f32 %f2158, %f2144, %f3104; fma.rn.ftz.f32 %f2159, %f2137, %f3105, %f2158; fma.rn.ftz.f32 %f2160, %f2151, %f3092, %f2159; mul.ftz.f32 %f2161, %f2145, %f3104; fma.rn.ftz.f32 %f2162, %f2138, %f3105, %f2161; fma.rn.ftz.f32 %f2163, %f2152, %f3092, %f2162; mul.ftz.f32 %f2164, %f2146, %f3104; fma.rn.ftz.f32 %f2165, %f2139, %f3105, %f2164; fma.rn.ftz.f32 %f2166, %f2153, %f3092, %f2165; mul.ftz.f32 %f2167, %f2163, %f2163; fma.rn.ftz.f32 %f2168, %f2160, %f2160, %f2167; fma.rn.ftz.f32 %f2169, %f2166, %f2166, %f2168; sqrt.approx.ftz.f32 %f2170, %f2169; div.approx.ftz.f32 %f3105, %f2160, %f2170; div.approx.ftz.f32 %f3104, %f2163, %f2170; div.approx.ftz.f32 %f3098, %f2166, %f2170; bra.uni $L__BB211_386; $L__BB211_384: cvta.to.local.u64 %rd581, %rd48; ld.local.v4.f32 {%f2103, %f2104, %f2105, %f2106}, [%rd581]; cvta.to.local.u64 %rd583, %rd51; ld.local.v4.f32 {%f2110, %f2111, %f2112, %f2113}, [%rd583]; cvta.to.local.u64 %rd585, %rd50; ld.local.v4.f32 {%f2117, %f2118, %f2119, %f2120}, [%rd585]; mul.ftz.f32 %f2124, %f3104, %f2110; fma.rn.ftz.f32 %f2125, %f3105, %f2103, %f2124; fma.rn.ftz.f32 %f2126, %f3092, %f2117, %f2125; mul.ftz.f32 %f2127, %f3104, %f2111; fma.rn.ftz.f32 %f2128, %f3105, %f2104, %f2127; fma.rn.ftz.f32 %f2129, %f3092, %f2118, %f2128; mul.ftz.f32 %f2130, %f3104, %f2112; fma.rn.ftz.f32 %f2131, %f3105, %f2105, %f2130; fma.rn.ftz.f32 %f2132, %f3092, %f2119, %f2131; mul.ftz.f32 %f2133, %f2129, %f2129; fma.rn.ftz.f32 %f2134, %f2126, %f2126, %f2133; fma.rn.ftz.f32 %f2135, %f2132, %f2132, %f2134; sqrt.approx.ftz.f32 %f2136, %f2135; div.approx.ftz.f32 %f3105, %f2126, %f2136; div.approx.ftz.f32 %f3104, %f2129, %f2136; div.approx.ftz.f32 %f3098, %f2132, %f2136; $L__BB211_386: st.local.f32 [%rd22], %f3105; st.local.f32 [%rd64], %f3104; st.local.f32 [%rd65], %f3098; @%p233 bra $L__BB211_388; bra.uni $L__BB211_387; $L__BB211_388: ld.global.v4.f32 {%f2198, %f2199, %f2200, %f2201}, [%rd103]; ld.global.v4.f32 {%f2205, %f2206, %f2207, %f2208}, [%rd103+16]; ld.global.v4.f32 {%f2212, %f2213, %f2214, %f2215}, [%rd103+32]; mul.ftz.f32 %f2219, %f2199, %f3103; fma.rn.ftz.f32 %f2220, %f3101, %f2198, %f2219; fma.rn.ftz.f32 %f589, %f2200, %f3102, %f2220; mul.ftz.f32 %f2221, %f2206, %f3103; fma.rn.ftz.f32 %f2222, %f3101, %f2205, %f2221; fma.rn.ftz.f32 %f590, %f2207, %f3102, %f2222; mul.ftz.f32 %f2223, %f2213, %f3103; fma.rn.ftz.f32 %f2224, %f3101, %f2212, %f2223; fma.rn.ftz.f32 %f3102, %f2214, %f3102, %f2224; st.local.f32 [%rd23], %f589; st.local.f32 [%rd24], %f590; st.local.f32 [%rd25], %f3102; mov.f32 %f3101, %f589; mov.f32 %f3103, %f590; mov.f32 %f3106, %f3095; mov.f32 %f3107, %f3096; bra.uni $L__BB211_389; $L__BB211_387: cvta.to.local.u64 %rd588, %rd44; ld.local.v4.f32 {%f2171, %f2172, %f2173, %f2174}, [%rd588]; mul.ftz.f32 %f2178, %f3103, %f2172; fma.rn.ftz.f32 %f2179, %f3101, %f2171, %f2178; fma.rn.ftz.f32 %f586, %f3102, %f2173, %f2179; cvta.to.local.u64 %rd590, %rd47; ld.local.v4.f32 {%f2180, %f2181, %f2182, %f2183}, [%rd590]; mul.ftz.f32 %f2187, %f3103, %f2181; fma.rn.ftz.f32 %f2188, %f3101, %f2180, %f2187; fma.rn.ftz.f32 %f587, %f3102, %f2182, %f2188; cvta.to.local.u64 %rd592, %rd46; ld.local.v4.f32 {%f2189, %f2190, %f2191, %f2192}, [%rd592]; mul.ftz.f32 %f2196, %f3103, %f2190; fma.rn.ftz.f32 %f2197, %f3101, %f2189, %f2196; fma.rn.ftz.f32 %f3102, %f3102, %f2191, %f2197; st.local.f32 [%rd23], %f586; st.local.f32 [%rd24], %f587; st.local.f32 [%rd25], %f3102; mov.f32 %f3101, %f586; mov.f32 %f3103, %f587; mov.f32 %f3106, %f3095; mov.f32 %f3107, %f3096; $L__BB211_389: add.s64 %rd1188, %rd166, 44; neg.ftz.f32 %f2947, %f303; neg.ftz.f32 %f2946, %f302; add.s64 %rd1172, %rd166, 20; st.local.v2.f32 [%rd71], {%f3106, %f3107}; st.local.f32 [%rd28], %f3108; selp.f32 %f3145, %f3105, %f2946, %p253; selp.f32 %f3146, %f3104, %f2947, %p253; selp.b64 %rd599, %rd1172, %rd1188, %p253; cvta.to.local.u64 %rd600, %rd599; ld.local.f32 %f3147, [%rd600]; add.s64 %rd601, %rd166, 24; cvta.to.local.u64 %rd602, %rd601; st.local.v2.f32 [%rd602], {%f3145, %f3146}; st.local.f32 [%rd19], %f3147; mul.ftz.f32 %f2225, %f3103, %f3147; mul.ftz.f32 %f2226, %f3102, %f3146; mul.ftz.f32 %f2227, %f3102, %f3145; mul.ftz.f32 %f2228, %f3101, %f3147; mul.ftz.f32 %f2229, %f3101, %f3146; mul.ftz.f32 %f2230, %f3103, %f3145; sub.ftz.f32 %f2231, %f2229, %f2230; sub.ftz.f32 %f2232, %f2227, %f2228; sub.ftz.f32 %f2233, %f2225, %f2226; add.s64 %rd603, %rd166, 128; cvta.to.local.u64 %rd604, %rd603; st.local.v2.f32 [%rd604], {%f2233, %f2232}; add.s64 %rd605, %rd166, 136; cvta.to.local.u64 %rd606, %rd605; st.local.f32 [%rd606], %f2231; ld.global.u32 %r655, [%rd105+-4]; st.local.u32 [%rd18], %r655; bra.uni $L__BB211_440; $L__BB211_412: { // callseq 1279, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; call.uni _Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib, ( param0 ); } // callseq 1279 ld.local.u32 %r901, [%rd32]; bra.uni $L__BB211_426; $L__BB211_392: ld.global.u32 %r196, [%rd103+148]; ld.global.u32 %r197, [%rd103+152]; shl.b32 %r659, %r197, 1; cvt.rn.f32.s32 %f2234, %r659; mul.ftz.f32 %f608, %f305, %f2234; cvt.rzi.ftz.s32.f32 %r660, %f608; add.s32 %r661, %r659, -1; min.s32 %r198, %r660, %r661; ld.global.u32 %r897, [%rd103+164]; mul.wide.u32 %rd614, %r897, 16; add.s64 %rd615, %rd52, %rd614; ld.global.u16 %rs50, [%rd615+12]; ld.global.u32 %r896, [%rd615+8]; ld.global.u64 %rd1197, [%rd615]; setp.eq.s64 %p266, %rd1197, 11; @%p266 bra $L__BB211_397; $L__BB211_393: setp.eq.s64 %p267, %rd1197, 0; @%p267 bra $L__BB211_395; add.s32 %r897, %r897, 2; bra.uni $L__BB211_396; $L__BB211_395: setp.eq.s16 %p268, %rs50, 0; mov.u32 %r900, -1; mov.u32 %r897, %r896; @%p268 bra $L__BB211_398; $L__BB211_396: mul.wide.u32 %rd616, %r897, 16; add.s64 %rd617, %rd52, %rd616; ld.global.u16 %rs50, [%rd617+12]; ld.global.u32 %r896, [%rd617+8]; ld.global.u64 %rd1197, [%rd617]; setp.ne.s64 %p269, %rd1197, 11; @%p269 bra $L__BB211_393; $L__BB211_397: setp.eq.s16 %p270, %rs50, 0; selp.b32 %r900, -1, %r896, %p270; $L__BB211_398: setp.eq.s32 %p271, %r198, %r197; @%p271 bra $L__BB211_400; bra.uni $L__BB211_399; $L__BB211_400: shl.b64 %rd619, %rd102, 4; add.s64 %rd1198, %rd60, %rd619; bra.uni $L__BB211_401; $L__BB211_24: shr.u32 %r449, %r23, 16; xor.b32 %r450, %r449, %r23; mul.lo.s32 %r451, %r450, 569420461; shr.u32 %r452, %r451, 15; xor.b32 %r453, %r452, %r451; mul.lo.s32 %r454, %r453, -749064809; shr.u32 %r455, %r454, 15; xor.b32 %r456, %r21, %r454; xor.b32 %r457, %r456, %r455; xor.b32 %r458, %r457, 1493655573; add.s32 %r459, %r458, %r24; shr.u32 %r460, %r458, 16; or.b32 %r461, %r460, 1; mul.lo.s32 %r462, %r461, %r459; mul.lo.s32 %r463, %r462, 89287766; xor.b32 %r464, %r463, %r462; mul.lo.s32 %r465, %r464, 1403136100; xor.b32 %r466, %r465, %r464; xor.b32 %r467, %r457, -2052961194; brev.b32 %r468, %r466; mul.lo.s32 %r469, %r468, 1025551850; xor.b32 %r470, %r469, %r468; add.s32 %r471, %r470, %r467; shr.u32 %r472, %r467, 16; or.b32 %r473, %r472, 1; mul.lo.s32 %r474, %r471, %r473; mul.lo.s32 %r475, %r474, 89287766; xor.b32 %r476, %r475, %r474; mul.lo.s32 %r477, %r476, 1403136100; xor.b32 %r478, %r477, %r476; brev.b32 %r479, %r478; cvt.rn.f32.u32 %f880, %r479; mul.ftz.f32 %f2956, %f880, 0f2F7FFFFE; mov.u32 %r875, %r8; mov.f32 %f2955, %f2954; bra.uni $L__BB211_25; $L__BB211_414: sub.ftz.f32 %f2339, %f651, %f652; sub.ftz.f32 %f2340, %f653, %f654; mul.ftz.f32 %f2341, %f2340, %f2340; fma.rn.ftz.f32 %f2342, %f2339, %f2339, %f2341; sub.ftz.f32 %f2343, %f655, %f656; fma.rn.ftz.f32 %f2344, %f2343, %f2343, %f2342; sqrt.approx.ftz.f32 %f2345, %f2344; div.approx.ftz.f32 %f3117, %f2343, %f2345; div.approx.ftz.f32 %f3116, %f2340, %f2345; div.approx.ftz.f32 %f3115, %f2339, %f2345; $L__BB211_416: shl.b64 %rd651, %rd102, 2; add.s64 %rd652, %rd63, %rd651; ld.global.u32 %r211, [%rd652]; st.local.u32 [%rd18], %r211; fma.rn.ftz.f32 %f2353, %f296, %f648, %f642; fma.rn.ftz.f32 %f2354, %f296, %f649, %f644; fma.rn.ftz.f32 %f2355, %f296, %f650, %f646; fma.rn.ftz.f32 %f3130, %f297, %f643, %f2353; fma.rn.ftz.f32 %f3131, %f297, %f645, %f2354; fma.rn.ftz.f32 %f3132, %f297, %f647, %f2355; and.b32 %r680, %r901, 4; setp.ne.s32 %p279, %r680, 0; @%p279 bra $L__BB211_421; @%p233 bra $L__BB211_419; bra.uni $L__BB211_418; $L__BB211_419: ld.global.v4.f32 {%f3118, %f3119, %f3120, %f3121}, [%rd103]; ld.global.v4.f32 {%f3122, %f3123, %f3124, %f3125}, [%rd103+16]; ld.global.v4.f32 {%f3126, %f3127, %f3128, %f3129}, [%rd103+32]; bra.uni $L__BB211_420; $L__BB211_399: cvt.u32.u64 %r663, %rd102; setp.gt.s32 %p272, %r198, %r197; selp.b32 %r664, -1, 0, %p272; add.s32 %r665, %r198, %r664; mad.lo.s32 %r666, %r665, %r196, %r663; add.s32 %r667, %r666, %r900; mul.wide.s32 %rd618, %r667, 16; add.s64 %rd1198, %rd54, %rd618; $L__BB211_401: ld.global.v4.f32 {%f2235, %f2236, %f2237, %f2238}, [%rd1198]; add.s32 %r209, %r198, 1; setp.eq.s32 %p273, %r209, %r197; @%p273 bra $L__BB211_403; bra.uni $L__BB211_402; $L__BB211_403: shl.b64 %rd621, %rd102, 4; add.s64 %rd1199, %rd60, %rd621; bra.uni $L__BB211_404; $L__BB211_402: cvt.u32.u64 %r668, %rd102; setp.ge.s32 %p274, %r198, %r197; selp.b32 %r669, -1, 0, %p274; add.s32 %r670, %r209, %r669; mad.lo.s32 %r671, %r670, %r196, %r668; add.s32 %r672, %r671, %r900; mul.wide.s32 %rd620, %r672, 16; add.s64 %rd1199, %rd54, %rd620; $L__BB211_404: cvt.rn.f32.s32 %f2239, %r198; sub.ftz.f32 %f2240, %f608, %f2239; ld.global.v4.f32 {%f2241, %f2242, %f2243, %f2244}, [%rd1199]; mov.f32 %f2249, 0f3F800000; sub.ftz.f32 %f2250, %f2249, %f2240; mul.ftz.f32 %f2251, %f2240, %f2241; mul.ftz.f32 %f2252, %f2240, %f2242; mul.ftz.f32 %f2253, %f2240, %f2243; fma.rn.ftz.f32 %f3114, %f2250, %f2237, %f2253; fma.rn.ftz.f32 %f3113, %f2250, %f2236, %f2252; fma.rn.ftz.f32 %f3112, %f2250, %f2235, %f2251; $L__BB211_406: and.b32 %r673, %r901, 4; setp.ne.s32 %p275, %r673, 0; @%p275 bra $L__BB211_410; @%p233 bra $L__BB211_409; bra.uni $L__BB211_408; $L__BB211_409: ld.global.v4.f32 {%f2292, %f2293, %f2294, %f2295}, [%rd103]; ld.global.v4.f32 {%f2300, %f2301, %f2302, %f2303}, [%rd103+16]; ld.global.v4.f32 {%f2308, %f2309, %f2310, %f2311}, [%rd103+32]; mul.ftz.f32 %f2316, %f3113, %f2293; fma.rn.ftz.f32 %f2317, %f3112, %f2292, %f2316; fma.rn.ftz.f32 %f2318, %f3114, %f2294, %f2317; add.ftz.f32 %f631, %f2295, %f2318; mul.ftz.f32 %f2319, %f3113, %f2301; fma.rn.ftz.f32 %f2320, %f3112, %f2300, %f2319; fma.rn.ftz.f32 %f2321, %f3114, %f2302, %f2320; add.ftz.f32 %f632, %f2303, %f2321; mul.ftz.f32 %f2322, %f3113, %f2309; fma.rn.ftz.f32 %f2323, %f3112, %f2308, %f2322; fma.rn.ftz.f32 %f2324, %f3114, %f2310, %f2323; add.ftz.f32 %f3114, %f2311, %f2324; mov.f32 %f3112, %f631; mov.f32 %f3113, %f632; bra.uni $L__BB211_410; $L__BB211_408: cvta.to.local.u64 %rd626, %rd44; ld.local.v4.f32 {%f2259, %f2260, %f2261, %f2262}, [%rd626]; mul.ftz.f32 %f2267, %f3113, %f2260; fma.rn.ftz.f32 %f2268, %f3112, %f2259, %f2267; fma.rn.ftz.f32 %f2269, %f3114, %f2261, %f2268; add.ftz.f32 %f628, %f2262, %f2269; cvta.to.local.u64 %rd628, %rd47; ld.local.v4.f32 {%f2270, %f2271, %f2272, %f2273}, [%rd628]; mul.ftz.f32 %f2278, %f3113, %f2271; fma.rn.ftz.f32 %f2279, %f3112, %f2270, %f2278; fma.rn.ftz.f32 %f2280, %f3114, %f2272, %f2279; add.ftz.f32 %f629, %f2273, %f2280; cvta.to.local.u64 %rd630, %rd46; ld.local.v4.f32 {%f2281, %f2282, %f2283, %f2284}, [%rd630]; mul.ftz.f32 %f2289, %f3113, %f2282; fma.rn.ftz.f32 %f2290, %f3112, %f2281, %f2289; fma.rn.ftz.f32 %f2291, %f3114, %f2283, %f2290; add.ftz.f32 %f3114, %f2284, %f2291; mov.f32 %f3112, %f628; mov.f32 %f3113, %f629; $L__BB211_410: sub.ftz.f32 %f2325, %f605, %f3112; sub.ftz.f32 %f2326, %f606, %f3113; mul.ftz.f32 %f2327, %f2326, %f2326; fma.rn.ftz.f32 %f2328, %f2325, %f2325, %f2327; sub.ftz.f32 %f2329, %f607, %f3114; fma.rn.ftz.f32 %f2330, %f2329, %f2329, %f2328; sqrt.approx.ftz.f32 %f2331, %f2330; div.approx.ftz.f32 %f3145, %f2325, %f2331; div.approx.ftz.f32 %f3146, %f2326, %f2331; div.approx.ftz.f32 %f3147, %f2329, %f2331; st.local.f32 [%rd19], %f3147; st.local.f32 [%rd22], %f3145; st.local.v4.f32 [%rd64], {%f3146, %f3147, %f3145, %f3146}; mov.u32 %r674, 0; st.local.u32 [%rd23], %r674; st.local.u32 [%rd24], %r674; st.local.u32 [%rd25], %r674; add.s64 %rd634, %rd166, 128; cvta.to.local.u64 %rd635, %rd634; mov.f32 %f2332, 0f00000000; st.local.v2.f32 [%rd635], {%f2332, %f2332}; add.s64 %rd636, %rd166, 136; cvta.to.local.u64 %rd637, %rd636; st.local.u32 [%rd637], %r674; bra.uni $L__BB211_440; $L__BB211_418: cvta.to.local.u64 %rd655, %rd44; ld.local.v4.f32 {%f3118, %f3119, %f3120, %f3121}, [%rd655]; cvta.to.local.u64 %rd657, %rd47; ld.local.v4.f32 {%f3122, %f3123, %f3124, %f3125}, [%rd657]; cvta.to.local.u64 %rd659, %rd46; ld.local.v4.f32 {%f3126, %f3127, %f3128, %f3129}, [%rd659]; $L__BB211_420: mul.ftz.f32 %f2380, %f3131, %f3119; fma.rn.ftz.f32 %f2381, %f3130, %f3118, %f2380; fma.rn.ftz.f32 %f2382, %f3132, %f3120, %f2381; mul.ftz.f32 %f2383, %f3131, %f3123; fma.rn.ftz.f32 %f2384, %f3130, %f3122, %f2383; fma.rn.ftz.f32 %f2385, %f3132, %f3124, %f2384; mul.ftz.f32 %f2386, %f3131, %f3127; fma.rn.ftz.f32 %f2387, %f3130, %f3126, %f2386; fma.rn.ftz.f32 %f2388, %f3132, %f3128, %f2387; add.ftz.f32 %f3130, %f3121, %f2382; add.ftz.f32 %f3131, %f3125, %f2385; add.ftz.f32 %f3132, %f3129, %f2388; $L__BB211_421: st.local.f32 [%rd19], %f3117; st.local.v4.f32 [%rd71], {%f3130, %f3131, %f3132, %f3115}; st.local.v4.f32 [%rd64], {%f3116, %f3117, %f3115, %f3116}; setp.gt.s32 %p281, %r211, -1; @%p281 bra $L__BB211_425; ld.global.v4.u32 {%r681, %r682, %r683, %r684}, [%rd122+-12]; mul.wide.u32 %rd665, %r681, 12; add.s64 %rd666, %rd664, %rd665; mul.wide.u32 %rd667, %r682, 12; add.s64 %rd668, %rd664, %rd667; mul.wide.u32 %rd669, %r683, 12; add.s64 %rd670, %rd664, %rd669; mov.f32 %f2389, 0f3F800000; sub.ftz.f32 %f2390, %f2389, %f296; sub.ftz.f32 %f2391, %f2390, %f297; ld.global.f32 %f2392, [%rd666]; ld.global.f32 %f2393, [%rd666+4]; ld.global.f32 %f2394, [%rd666+8]; ld.global.f32 %f2395, [%rd668]; mul.ftz.f32 %f2396, %f296, %f2395; ld.global.f32 %f2397, [%rd668+4]; mul.ftz.f32 %f2398, %f296, %f2397; ld.global.f32 %f2399, [%rd668+8]; mul.ftz.f32 %f2400, %f296, %f2399; fma.rn.ftz.f32 %f2401, %f2391, %f2392, %f2396; fma.rn.ftz.f32 %f2402, %f2391, %f2393, %f2398; fma.rn.ftz.f32 %f2403, %f2391, %f2394, %f2400; ld.global.f32 %f2404, [%rd670]; ld.global.f32 %f2405, [%rd670+4]; ld.global.f32 %f2406, [%rd670+8]; fma.rn.ftz.f32 %f3133, %f297, %f2404, %f2401; fma.rn.ftz.f32 %f3134, %f297, %f2405, %f2402; fma.rn.ftz.f32 %f3135, %f297, %f2406, %f2403; mul.ftz.f32 %f2407, %f3134, %f3134; fma.rn.ftz.f32 %f2408, %f3133, %f3133, %f2407; fma.rn.ftz.f32 %f2409, %f3135, %f3135, %f2408; sqrt.approx.ftz.f32 %f717, %f2409; setp.neu.ftz.f32 %p282, %f717, 0f00000000; @%p282 bra $L__BB211_423; bra.uni $L__BB211_424; $L__BB211_423: rcp.approx.ftz.f32 %f2410, %f717; mul.ftz.f32 %f3135, %f3135, %f2410; mul.ftz.f32 %f3134, %f3134, %f2410; mul.ftz.f32 %f3133, %f3133, %f2410; $L__BB211_424: setp.eq.ftz.f32 %p283, %f3133, 0f00000000; setp.eq.ftz.f32 %p284, %f3134, 0f00000000; and.pred %p285, %p283, %p284; setp.eq.ftz.f32 %p286, %f3135, 0f00000000; and.pred %p287, %p286, %p285; selp.f32 %f2411, %f3117, %f3135, %p287; selp.f32 %f2412, %f3116, %f3134, %p287; selp.f32 %f2413, %f3115, %f3133, %p287; st.local.f32 [%rd22], %f2413; st.local.f32 [%rd64], %f2412; st.local.f32 [%rd65], %f2411; $L__BB211_425: st.local.f32 [%rd23], %f648; st.local.f32 [%rd24], %f649; st.local.f32 [%rd25], %f650; add.s64 %rd672, %rd166, 128; cvta.to.local.u64 %rd673, %rd672; st.local.v2.f32 [%rd673], {%f643, %f645}; add.s64 %rd674, %rd166, 136; cvta.to.local.u64 %rd675, %rd674; st.local.f32 [%rd675], %f647; $L__BB211_426: and.b32 %r688, %r901, 4; setp.eq.s32 %p288, %r688, 0; @%p288 bra $L__BB211_428; bra.uni $L__BB211_427; $L__BB211_428: and.b32 %r213, %r901, 2; setp.eq.s32 %p289, %r213, 0; @%p289 bra $L__BB211_430; bra.uni $L__BB211_429; $L__BB211_430: ld.local.u32 %r689, [%rd33]; mul.wide.s32 %rd690, %r689, 208; add.s64 %rd691, %rd12, %rd690; ld.global.v4.f32 {%f2455, %f2456, %f2457, %f2458}, [%rd691+48]; ld.global.v4.f32 {%f2462, %f2463, %f2464, %f2465}, [%rd691+64]; ld.global.v4.f32 {%f2469, %f2470, %f2471, %f2472}, [%rd691+80]; ld.local.v2.f32 {%f2476, %f2477}, [%rd64]; ld.local.f32 %f2480, [%rd22]; mul.ftz.f32 %f2481, %f2462, %f2476; fma.rn.ftz.f32 %f2482, %f2455, %f2480, %f2481; fma.rn.ftz.f32 %f2483, %f2469, %f2477, %f2482; mul.ftz.f32 %f2484, %f2463, %f2476; fma.rn.ftz.f32 %f2485, %f2456, %f2480, %f2484; fma.rn.ftz.f32 %f2486, %f2470, %f2477, %f2485; mul.ftz.f32 %f2487, %f2464, %f2476; fma.rn.ftz.f32 %f2488, %f2457, %f2480, %f2487; fma.rn.ftz.f32 %f2489, %f2471, %f2477, %f2488; mul.ftz.f32 %f2490, %f2486, %f2486; fma.rn.ftz.f32 %f2491, %f2483, %f2483, %f2490; fma.rn.ftz.f32 %f2492, %f2489, %f2489, %f2491; sqrt.approx.ftz.f32 %f2493, %f2492; div.approx.ftz.f32 %f3138, %f2483, %f2493; div.approx.ftz.f32 %f3137, %f2486, %f2493; div.approx.ftz.f32 %f3136, %f2489, %f2493; bra.uni $L__BB211_431; $L__BB211_427: add.s64 %rd677, %rd166, 24; cvta.to.local.u64 %rd678, %rd677; ld.local.v2.f32 {%f3145, %f3146}, [%rd678]; ld.local.f32 %f3147, [%rd19]; bra.uni $L__BB211_440; $L__BB211_429: ld.local.v2.f32 {%f2416, %f2417}, [%rd64]; cvta.to.local.u64 %rd683, %rd48; ld.local.v4.f32 {%f2420, %f2421, %f2422, %f2423}, [%rd683]; cvta.to.local.u64 %rd685, %rd51; ld.local.v4.f32 {%f2427, %f2428, %f2429, %f2430}, [%rd685]; cvta.to.local.u64 %rd687, %rd50; ld.local.v4.f32 {%f2434, %f2435, %f2436, %f2437}, [%rd687]; ld.local.f32 %f2441, [%rd22]; mul.ftz.f32 %f2442, %f2416, %f2427; fma.rn.ftz.f32 %f2443, %f2441, %f2420, %f2442; fma.rn.ftz.f32 %f2444, %f2417, %f2434, %f2443; mul.ftz.f32 %f2445, %f2416, %f2428; fma.rn.ftz.f32 %f2446, %f2441, %f2421, %f2445; fma.rn.ftz.f32 %f2447, %f2417, %f2435, %f2446; mul.ftz.f32 %f2448, %f2416, %f2429; fma.rn.ftz.f32 %f2449, %f2441, %f2422, %f2448; fma.rn.ftz.f32 %f2450, %f2417, %f2436, %f2449; mul.ftz.f32 %f2451, %f2447, %f2447; fma.rn.ftz.f32 %f2452, %f2444, %f2444, %f2451; fma.rn.ftz.f32 %f2453, %f2450, %f2450, %f2452; sqrt.approx.ftz.f32 %f2454, %f2453; div.approx.ftz.f32 %f3138, %f2444, %f2454; div.approx.ftz.f32 %f3137, %f2447, %f2454; div.approx.ftz.f32 %f3136, %f2450, %f2454; $L__BB211_431: st.local.f32 [%rd22], %f3138; st.local.f32 [%rd64], %f3137; st.local.f32 [%rd65], %f3136; @%p289 bra $L__BB211_433; bra.uni $L__BB211_432; $L__BB211_433: add.s64 %rd703, %rd166, 24; ld.local.u32 %r690, [%rd33]; mul.wide.s32 %rd704, %r690, 208; add.s64 %rd705, %rd12, %rd704; ld.global.v4.f32 {%f2533, %f2534, %f2535, %f2536}, [%rd705+48]; ld.global.v4.f32 {%f2540, %f2541, %f2542, %f2543}, [%rd705+64]; ld.global.v4.f32 {%f2547, %f2548, %f2549, %f2550}, [%rd705+80]; cvta.to.local.u64 %rd706, %rd703; ld.local.v2.f32 {%f2554, %f2555}, [%rd706]; mul.ftz.f32 %f2558, %f2540, %f2555; fma.rn.ftz.f32 %f2559, %f2533, %f2554, %f2558; ld.local.f32 %f2560, [%rd19]; fma.rn.ftz.f32 %f2561, %f2547, %f2560, %f2559; mul.ftz.f32 %f2562, %f2541, %f2555; fma.rn.ftz.f32 %f2563, %f2534, %f2554, %f2562; fma.rn.ftz.f32 %f2564, %f2548, %f2560, %f2563; mul.ftz.f32 %f2565, %f2542, %f2555; fma.rn.ftz.f32 %f2566, %f2535, %f2554, %f2565; fma.rn.ftz.f32 %f2567, %f2549, %f2560, %f2566; mul.ftz.f32 %f2568, %f2564, %f2564; fma.rn.ftz.f32 %f2569, %f2561, %f2561, %f2568; fma.rn.ftz.f32 %f2570, %f2567, %f2567, %f2569; sqrt.approx.ftz.f32 %f2571, %f2570; div.approx.ftz.f32 %f3145, %f2561, %f2571; div.approx.ftz.f32 %f3146, %f2564, %f2571; div.approx.ftz.f32 %f3147, %f2567, %f2571; bra.uni $L__BB211_434; $L__BB211_432: add.s64 %rd694, %rd166, 24; cvta.to.local.u64 %rd695, %rd694; ld.local.v2.f32 {%f2494, %f2495}, [%rd695]; cvta.to.local.u64 %rd697, %rd48; ld.local.v4.f32 {%f2498, %f2499, %f2500, %f2501}, [%rd697]; cvta.to.local.u64 %rd699, %rd51; ld.local.v4.f32 {%f2505, %f2506, %f2507, %f2508}, [%rd699]; cvta.to.local.u64 %rd701, %rd50; ld.local.v4.f32 {%f2512, %f2513, %f2514, %f2515}, [%rd701]; mul.ftz.f32 %f2519, %f2495, %f2505; fma.rn.ftz.f32 %f2520, %f2494, %f2498, %f2519; ld.local.f32 %f2521, [%rd19]; fma.rn.ftz.f32 %f2522, %f2521, %f2512, %f2520; mul.ftz.f32 %f2523, %f2495, %f2506; fma.rn.ftz.f32 %f2524, %f2494, %f2499, %f2523; fma.rn.ftz.f32 %f2525, %f2521, %f2513, %f2524; mul.ftz.f32 %f2526, %f2495, %f2507; fma.rn.ftz.f32 %f2527, %f2494, %f2500, %f2526; fma.rn.ftz.f32 %f2528, %f2521, %f2514, %f2527; mul.ftz.f32 %f2529, %f2525, %f2525; fma.rn.ftz.f32 %f2530, %f2522, %f2522, %f2529; fma.rn.ftz.f32 %f2531, %f2528, %f2528, %f2530; sqrt.approx.ftz.f32 %f2532, %f2531; div.approx.ftz.f32 %f3145, %f2522, %f2532; div.approx.ftz.f32 %f3146, %f2525, %f2532; div.approx.ftz.f32 %f3147, %f2528, %f2532; $L__BB211_434: add.s64 %rd708, %rd166, 24; cvta.to.local.u64 %rd709, %rd708; st.local.v2.f32 [%rd709], {%f3145, %f3146}; st.local.f32 [%rd19], %f3147; @%p289 bra $L__BB211_436; bra.uni $L__BB211_435; $L__BB211_436: ld.local.u32 %r691, [%rd33]; mul.wide.s32 %rd721, %r691, 208; add.s64 %rd722, %rd12, %rd721; ld.global.v4.f32 {%f2604, %f2605, %f2606, %f2607}, [%rd722]; ld.global.v4.f32 {%f2611, %f2612, %f2613, %f2614}, [%rd722+16]; ld.global.v4.f32 {%f2618, %f2619, %f2620, %f2621}, [%rd722+32]; ld.local.v2.f32 {%f2625, %f2626}, [%rd24]; ld.local.f32 %f2629, [%rd23]; mul.ftz.f32 %f2630, %f2605, %f2625; fma.rn.ftz.f32 %f2631, %f2629, %f2604, %f2630; fma.rn.ftz.f32 %f3144, %f2606, %f2626, %f2631; mul.ftz.f32 %f2632, %f2612, %f2625; fma.rn.ftz.f32 %f2633, %f2629, %f2611, %f2632; fma.rn.ftz.f32 %f3143, %f2613, %f2626, %f2633; mul.ftz.f32 %f2634, %f2619, %f2625; fma.rn.ftz.f32 %f2635, %f2629, %f2618, %f2634; fma.rn.ftz.f32 %f3142, %f2620, %f2626, %f2635; bra.uni $L__BB211_437; $L__BB211_435: ld.local.v2.f32 {%f2572, %f2573}, [%rd24]; cvta.to.local.u64 %rd714, %rd44; ld.local.v4.f32 {%f2576, %f2577, %f2578, %f2579}, [%rd714]; ld.local.f32 %f2583, [%rd23]; mul.ftz.f32 %f2584, %f2572, %f2577; fma.rn.ftz.f32 %f2585, %f2583, %f2576, %f2584; fma.rn.ftz.f32 %f3144, %f2573, %f2578, %f2585; cvta.to.local.u64 %rd716, %rd47; ld.local.v4.f32 {%f2586, %f2587, %f2588, %f2589}, [%rd716]; mul.ftz.f32 %f2593, %f2572, %f2587; fma.rn.ftz.f32 %f2594, %f2583, %f2586, %f2593; fma.rn.ftz.f32 %f3143, %f2573, %f2588, %f2594; cvta.to.local.u64 %rd718, %rd46; ld.local.v4.f32 {%f2595, %f2596, %f2597, %f2598}, [%rd718]; mul.ftz.f32 %f2602, %f2572, %f2596; fma.rn.ftz.f32 %f2603, %f2583, %f2595, %f2602; fma.rn.ftz.f32 %f3142, %f2573, %f2597, %f2603; $L__BB211_437: st.local.f32 [%rd23], %f3144; st.local.f32 [%rd24], %f3143; st.local.f32 [%rd25], %f3142; @%p289 bra $L__BB211_439; bra.uni $L__BB211_438; $L__BB211_439: add.s64 %rd736, %rd166, 128; ld.local.u32 %r692, [%rd33]; mul.wide.s32 %rd737, %r692, 208; add.s64 %rd738, %rd12, %rd737; ld.global.v4.f32 {%f2673, %f2674, %f2675, %f2676}, [%rd738]; ld.global.v4.f32 {%f2680, %f2681, %f2682, %f2683}, [%rd738+16]; ld.global.v4.f32 {%f2687, %f2688, %f2689, %f2690}, [%rd738+32]; cvta.to.local.u64 %rd739, %rd736; ld.local.v4.f32 {%f2694, %f2695, %f2696, %f2697}, [%rd739]; mul.ftz.f32 %f2701, %f2674, %f2695; fma.rn.ftz.f32 %f2702, %f2694, %f2673, %f2701; mul.ftz.f32 %f2703, %f2681, %f2695; fma.rn.ftz.f32 %f2704, %f2694, %f2680, %f2703; mul.ftz.f32 %f2705, %f2688, %f2695; fma.rn.ftz.f32 %f2706, %f2694, %f2687, %f2705; fma.rn.ftz.f32 %f2707, %f2689, %f2696, %f2706; fma.rn.ftz.f32 %f2708, %f2682, %f2696, %f2704; fma.rn.ftz.f32 %f2709, %f2675, %f2696, %f2702; st.local.v2.f32 [%rd739], {%f2709, %f2708}; add.s64 %rd740, %rd166, 136; cvta.to.local.u64 %rd741, %rd740; st.local.f32 [%rd741], %f2707; bra.uni $L__BB211_440; $L__BB211_438: add.s64 %rd725, %rd166, 128; cvta.to.local.u64 %rd726, %rd725; ld.local.v4.f32 {%f2636, %f2637, %f2638, %f2639}, [%rd726]; cvta.to.local.u64 %rd728, %rd44; ld.local.v4.f32 {%f2643, %f2644, %f2645, %f2646}, [%rd728]; mul.ftz.f32 %f2650, %f2637, %f2644; fma.rn.ftz.f32 %f2651, %f2636, %f2643, %f2650; cvta.to.local.u64 %rd730, %rd47; ld.local.v4.f32 {%f2652, %f2653, %f2654, %f2655}, [%rd730]; mul.ftz.f32 %f2659, %f2637, %f2653; fma.rn.ftz.f32 %f2660, %f2636, %f2652, %f2659; cvta.to.local.u64 %rd732, %rd46; ld.local.v4.f32 {%f2661, %f2662, %f2663, %f2664}, [%rd732]; mul.ftz.f32 %f2668, %f2637, %f2662; fma.rn.ftz.f32 %f2669, %f2636, %f2661, %f2668; fma.rn.ftz.f32 %f2670, %f2638, %f2663, %f2669; fma.rn.ftz.f32 %f2671, %f2638, %f2654, %f2660; fma.rn.ftz.f32 %f2672, %f2638, %f2645, %f2651; st.local.v2.f32 [%rd726], {%f2672, %f2671}; add.s64 %rd733, %rd166, 136; cvta.to.local.u64 %rd734, %rd733; st.local.f32 [%rd734], %f2670; $L__BB211_440: add.s64 %rd1192, %rd166, 40; cvta.to.local.u64 %rd1191, %rd1192; add.s64 %rd1190, %rd166, 36; cvta.to.local.u64 %rd1189, %rd1190; ld.local.u32 %rd742, [%rd18]; cvt.u32.u64 %r693, %rd742; and.b32 %r903, %r693, 4194303; shl.b64 %rd743, %rd742, 5; and.b64 %rd744, %rd743, 134217696; add.s64 %rd745, %rd16, %rd744; ld.global.u32 %r902, [%rd745+16]; st.local.u32 [%rd31], %r902; ld.local.f32 %f2710, [%rd1189]; ld.local.v2.f32 {%f2711, %f2712}, [%rd1191]; mul.ftz.f32 %f2715, %f3146, %f2711; fma.rn.ftz.f32 %f2716, %f3145, %f2710, %f2715; fma.rn.ftz.f32 %f2717, %f3147, %f2712, %f2716; setp.lt.ftz.f32 %p293, %f2717, 0f00000000; @%p293 bra $L__BB211_442; bra.uni $L__BB211_441; $L__BB211_442: add.s64 %rd754, %rd166, 128; or.b32 %r902, %r902, 1; st.local.u32 [%rd31], %r902; neg.ftz.f32 %f3147, %f3147; st.local.f32 [%rd19], %f3147; ld.local.f32 %f2718, [%rd22]; neg.ftz.f32 %f2719, %f2718; ld.local.v2.f32 {%f2720, %f2721}, [%rd64]; st.local.f32 [%rd22], %f2719; neg.ftz.f32 %f2724, %f2721; neg.ftz.f32 %f2725, %f2720; neg.ftz.f32 %f3146, %f3146; neg.ftz.f32 %f3145, %f3145; st.local.v4.f32 [%rd64], {%f2725, %f2724, %f3145, %f3146}; ld.local.f32 %f2726, [%rd23]; neg.ftz.f32 %f3159, %f2726; ld.local.v2.f32 {%f2727, %f2728}, [%rd24]; neg.ftz.f32 %f3160, %f2727; neg.ftz.f32 %f2731, %f2728; st.local.f32 [%rd23], %f3159; st.local.f32 [%rd24], %f3160; st.local.f32 [%rd25], %f2731; cvta.to.local.u64 %rd761, %rd754; ld.local.v4.f32 {%f2732, %f2733, %f2734, %f2735}, [%rd761]; neg.ftz.f32 %f2739, %f2734; neg.ftz.f32 %f2740, %f2733; neg.ftz.f32 %f2741, %f2732; st.local.v2.f32 [%rd761], {%f2741, %f2740}; add.s64 %rd762, %rd166, 136; cvta.to.local.u64 %rd763, %rd762; st.local.f32 [%rd763], %f2739; bra.uni $L__BB211_443; $L__BB211_441: ld.local.f32 %f3159, [%rd23]; ld.local.f32 %f3160, [%rd24]; $L__BB211_443: add.s64 %rd1174, %rd166, 88; cvta.to.local.u64 %rd1173, %rd1174; mov.u32 %r862, 0; add.s64 %rd765, %rd166, 128; ld.local.f32 %f2742, [%rd1173]; fma.rn.ftz.f32 %f2743, %f2742, 0f00000000, %f306; add.s64 %rd768, %rd166, 92; cvta.to.local.u64 %rd769, %rd768; st.local.f32 [%rd769], %f2743; add.s64 %rd770, %rd166, 96; cvta.to.local.u64 %rd771, %rd770; st.local.u32 [%rd771], %r862; cvta.to.local.u64 %rd772, %rd765; ld.local.v4.f32 {%f3157, %f3158, %f2746, %f2747}, [%rd772]; setp.neu.ftz.f32 %p294, %f3145, %f3146; setp.neu.ftz.f32 %p295, %f3145, %f3147; or.pred %p296, %p294, %p295; sub.ftz.f32 %f2748, %f3147, %f3146; neg.ftz.f32 %f2749, %f3146; neg.ftz.f32 %f2750, %f3147; selp.f32 %f2751, %f2750, %f3147, %p296; add.ftz.f32 %f2752, %f3145, %f2751; selp.f32 %f2753, %f3146, %f2749, %p296; sub.ftz.f32 %f2754, %f2753, %f3145; mul.ftz.f32 %f2755, %f2752, %f2752; fma.rn.ftz.f32 %f2756, %f2748, %f2748, %f2755; fma.rn.ftz.f32 %f2757, %f2754, %f2754, %f2756; sqrt.approx.ftz.f32 %f2758, %f2757; div.approx.ftz.f32 %f2759, %f2748, %f2758; div.approx.ftz.f32 %f2760, %f2752, %f2758; div.approx.ftz.f32 %f2761, %f2754, %f2758; mul.ftz.f32 %f2762, %f3146, %f2761; mul.ftz.f32 %f2763, %f3147, %f2760; sub.ftz.f32 %f2764, %f2762, %f2763; mul.ftz.f32 %f2765, %f3147, %f2759; mul.ftz.f32 %f2766, %f3145, %f2761; sub.ftz.f32 %f2767, %f2765, %f2766; mul.ftz.f32 %f2768, %f3145, %f2760; mul.ftz.f32 %f2769, %f3146, %f2759; sub.ftz.f32 %f2770, %f2768, %f2769; mul.ftz.f32 %f3153, %f2743, %f2759; mul.ftz.f32 %f3154, %f2743, %f2760; mul.ftz.f32 %f786, %f2743, %f2761; mul.ftz.f32 %f3155, %f2743, %f2764; mul.ftz.f32 %f3156, %f2743, %f2767; mul.ftz.f32 %f789, %f2743, %f2770; abs.ftz.f32 %f790, %f3145; abs.ftz.f32 %f791, %f3146; abs.ftz.f32 %f792, %f3147; setp.geu.ftz.f32 %p297, %f792, %f790; setp.geu.ftz.f32 %p298, %f792, %f791; and.pred %p299, %p297, %p298; @%p299 bra $L__BB211_445; ld.local.f32 %f794, [%rd25]; setp.lt.ftz.f32 %p300, %f791, %f792; setp.lt.ftz.f32 %p301, %f791, %f790; or.pred %p302, %p301, %p300; selp.f32 %f3153, %f3154, %f3153, %p302; selp.f32 %f3155, %f3156, %f3155, %p302; selp.f32 %f3157, %f3158, %f3157, %p302; selp.f32 %f3159, %f3160, %f3159, %p302; mov.f32 %f3154, %f786; mov.f32 %f3156, %f789; mov.f32 %f3158, %f2746; mov.f32 %f3160, %f794; $L__BB211_445: mul.ftz.f32 %f2771, %f3157, %f3160; mul.ftz.f32 %f2772, %f3158, %f3159; sub.ftz.f32 %f3161, %f2772, %f2771; setp.eq.ftz.f32 %p303, %f3161, 0f00000000; @%p303 bra $L__BB211_447; rcp.approx.ftz.f32 %f3161, %f3161; $L__BB211_447: mul.ftz.f32 %f2773, %f3154, %f3157; mul.ftz.f32 %f2774, %f3153, %f3158; sub.ftz.f32 %f2775, %f2774, %f2773; mul.ftz.f32 %f2776, %f2775, %f3161; add.s64 %rd774, %rd166, 100; cvta.to.local.u64 %rd775, %rd774; st.local.f32 [%rd775], %f2776; mul.ftz.f32 %f2777, %f3153, %f3160; mul.ftz.f32 %f2778, %f3154, %f3159; sub.ftz.f32 %f2779, %f2778, %f2777; mul.ftz.f32 %f2780, %f3156, %f3157; mul.ftz.f32 %f2781, %f3155, %f3158; sub.ftz.f32 %f2782, %f2781, %f2780; mul.ftz.f32 %f2783, %f2779, %f3161; mul.ftz.f32 %f2784, %f2782, %f3161; add.s64 %rd776, %rd166, 104; cvta.to.local.u64 %rd777, %rd776; st.local.v2.f32 [%rd777], {%f2784, %f2783}; mul.ftz.f32 %f2785, %f3155, %f3160; mul.ftz.f32 %f2786, %f3156, %f3159; sub.ftz.f32 %f2787, %f2786, %f2785; mul.ftz.f32 %f2788, %f2787, %f3161; add.s64 %rd778, %rd166, 112; cvta.to.local.u64 %rd779, %rd778; st.local.f32 [%rd779], %f2788; and.b32 %r695, %r902, 524288; setp.ne.s32 %p304, %r695, 0; @%p304 bra $L__BB211_451; mov.u32 %r696, 0; st.local.u32 [%rd29], %r696; st.local.u32 [%rd30], %r696; bra.uni $L__BB211_449; $L__BB211_619: { // callseq 1280, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1280 add.s32 %r903, %r218, 1; $L__BB211_449: mov.u32 %r218, %r903; mul.wide.s32 %rd785, %r218, 16; add.s64 %rd123, %rd36, %rd785; ld.global.v4.u32 {%r698, %r699, %r700, %r701}, [%rd123]; mul.wide.u32 %rd786, %r699, 4; add.s64 %rd124, %rd1, %rd786; mul.wide.u32 %rd787, %r700, 4; add.s64 %rd125, %rd1, %rd787; setp.gt.s32 %p305, %r698, 1; @%p305 bra $L__BB211_457; setp.eq.s32 %p433, %r698, 1; mov.u32 %r903, %r699; @%p433 bra $L__BB211_449; bra.uni $L__BB211_451; $L__BB211_457: setp.gt.s32 %p306, %r698, 49; @%p306 bra $L__BB211_538; setp.gt.s32 %p370, %r698, 25; @%p370 bra $L__BB211_499; setp.gt.s32 %p402, %r698, 13; @%p402 bra $L__BB211_480; setp.gt.s32 %p418, %r698, 7; @%p418 bra $L__BB211_470; setp.gt.s32 %p426, %r698, 4; @%p426 bra $L__BB211_466; setp.eq.s32 %p430, %r698, 2; @%p430 bra $L__BB211_686; setp.eq.s32 %p431, %r698, 3; @%p431 bra $L__BB211_685; setp.eq.s32 %p432, %r698, 4; @%p432 bra $L__BB211_465; bra.uni $L__BB211_451; $L__BB211_465: mov.u32 %r779, 4; { // callseq 1359, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r779, %r699, %r700, %r701}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1359 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_538: setp.gt.s32 %p307, %r698, 73; @%p307 bra $L__BB211_578; setp.gt.s32 %p339, %r698, 61; @%p339 bra $L__BB211_559; setp.gt.s32 %p355, %r698, 55; @%p355 bra $L__BB211_550; setp.gt.s32 %p363, %r698, 52; @%p363 bra $L__BB211_546; setp.eq.s32 %p367, %r698, 50; @%p367 bra $L__BB211_653; setp.eq.s32 %p368, %r698, 51; @%p368 bra $L__BB211_652; setp.eq.s32 %p369, %r698, 52; @%p369 bra $L__BB211_545; bra.uni $L__BB211_451; $L__BB211_545: add.s32 %r738, %r218, 1; { // callseq 1320, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r738; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1320 bra.uni $L__BB211_449; $L__BB211_499: setp.gt.s32 %p371, %r698, 37; @%p371 bra $L__BB211_519; setp.gt.s32 %p387, %r698, 31; @%p387 bra $L__BB211_510; setp.gt.s32 %p395, %r698, 28; @%p395 bra $L__BB211_506; setp.eq.s32 %p399, %r698, 26; @%p399 bra $L__BB211_669; setp.eq.s32 %p400, %r698, 27; @%p400 bra $L__BB211_668; setp.eq.s32 %p401, %r698, 28; @%p401 bra $L__BB211_505; bra.uni $L__BB211_451; $L__BB211_505: mov.u32 %r754, 28; { // callseq 1342, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r754; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1342 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_578: setp.gt.s32 %p308, %r698, 85; @%p308 bra $L__BB211_601; setp.gt.s32 %p324, %r698, 79; @%p324 bra $L__BB211_592; setp.gt.s32 %p332, %r698, 76; @%p332 bra $L__BB211_588; setp.eq.s32 %p336, %r698, 74; @%p336 bra $L__BB211_637; setp.eq.s32 %p337, %r698, 75; @%p337 bra $L__BB211_636; setp.eq.s32 %p338, %r698, 76; @%p338 bra $L__BB211_584; bra.uni $L__BB211_451; $L__BB211_584: add.s32 %r903, %r218, 1; setp.eq.s32 %p435, %r701, 255; @%p435 bra $L__BB211_449; add.s64 %rd126, %rd124, 4; add.s64 %rd127, %rd124, 8; setp.eq.s32 %p436, %r700, 0; mul.wide.u32 %rd812, %r701, 4; add.s64 %rd128, %rd1, %rd812; @%p436 bra $L__BB211_635; setp.eq.s32 %p437, %r700, 1; @%p437 bra $L__BB211_634; bra.uni $L__BB211_587; $L__BB211_634: ld.local.f32 %f2796, [%rd126]; st.local.f32 [%rd128], %f2796; bra.uni $L__BB211_449; $L__BB211_480: setp.gt.s32 %p403, %r698, 19; @%p403 bra $L__BB211_490; setp.gt.s32 %p411, %r698, 16; @%p411 bra $L__BB211_486; setp.eq.s32 %p415, %r698, 14; @%p415 bra $L__BB211_677; setp.eq.s32 %p416, %r698, 15; @%p416 bra $L__BB211_676; setp.eq.s32 %p417, %r698, 16; @%p417 bra $L__BB211_485; bra.uni $L__BB211_451; $L__BB211_485: mov.u32 %r764, 16; { // callseq 1353, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r764; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z13svm_node_attrILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1353 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_559: setp.gt.s32 %p340, %r698, 67; @%p340 bra $L__BB211_569; setp.gt.s32 %p348, %r698, 64; @%p348 bra $L__BB211_565; setp.eq.s32 %p352, %r698, 62; @%p352 bra $L__BB211_645; setp.eq.s32 %p353, %r698, 63; @%p353 bra $L__BB211_644; setp.eq.s32 %p354, %r698, 64; @%p354 bra $L__BB211_564; bra.uni $L__BB211_451; $L__BB211_564: mov.u32 %r724, 64; add.s32 %r725, %r218, 1; { // callseq 1308, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r724; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r725; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1308 bra.uni $L__BB211_449; $L__BB211_519: setp.gt.s32 %p372, %r698, 43; @%p372 bra $L__BB211_529; setp.gt.s32 %p380, %r698, 40; @%p380 bra $L__BB211_525; setp.eq.s32 %p384, %r698, 38; @%p384 bra $L__BB211_661; setp.eq.s32 %p385, %r698, 39; @%p385 bra $L__BB211_660; setp.eq.s32 %p386, %r698, 40; @%p386 bra $L__BB211_524; bra.uni $L__BB211_451; $L__BB211_524: add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_601: setp.gt.s32 %p309, %r698, 92; @%p309 bra $L__BB211_611; setp.gt.s32 %p317, %r698, 88; @%p317 bra $L__BB211_607; setp.eq.s32 %p321, %r698, 86; @%p321 bra $L__BB211_626; setp.eq.s32 %p322, %r698, 87; @%p322 bra $L__BB211_625; setp.eq.s32 %p323, %r698, 88; @%p323 bra $L__BB211_606; bra.uni $L__BB211_451; $L__BB211_606: mov.u32 %r707, 88; ld.local.f32 %f2789, [%rd71+12]; ld.local.v2.f32 {%f2790, %f2791}, [%rd71+16]; { // callseq 1286, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f2789; .param .b32 param1; st.param.f32 [param1+0], %f2790; .param .b32 param2; st.param.f32 [param2+0], %f2791; .param .b64 param3; st.param.b64 [param3+0], %rd146; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r707; st.param.b32 [param4+4], %r699; st.param.b32 [param4+8], %r700; st.param.b32 [param4+12], %r701; call.uni _Z14svm_node_bevelILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1286 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_470: setp.gt.s32 %p419, %r698, 10; @%p419 bra $L__BB211_476; setp.eq.s32 %p423, %r698, 8; @%p423 bra $L__BB211_682; setp.eq.s32 %p424, %r698, 9; @%p424 bra $L__BB211_680; setp.eq.s32 %p425, %r698, 10; @%p425 bra $L__BB211_474; bra.uni $L__BB211_451; $L__BB211_474: ld.local.f32 %f2821, [%rd125]; setp.ltu.ftz.f32 %p438, %f2821, 0f3F800000; add.s32 %r903, %r218, 1; @%p438 bra $L__BB211_449; add.s32 %r775, %r699, %r218; add.s32 %r903, %r775, 1; bra.uni $L__BB211_449; $L__BB211_550: setp.gt.s32 %p356, %r698, 58; @%p356 bra $L__BB211_555; setp.eq.s32 %p360, %r698, 56; @%p360 bra $L__BB211_649; setp.eq.s32 %p361, %r698, 57; @%p361 bra $L__BB211_648; setp.eq.s32 %p362, %r698, 58; @%p362 bra $L__BB211_554; bra.uni $L__BB211_451; $L__BB211_554: mov.u32 %r733, 58; { // callseq 1314, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r733; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1314 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_510: setp.gt.s32 %p388, %r698, 34; @%p388 bra $L__BB211_515; setp.eq.s32 %p392, %r698, 32; @%p392 bra $L__BB211_665; setp.eq.s32 %p393, %r698, 33; @%p393 bra $L__BB211_664; setp.eq.s32 %p394, %r698, 34; @%p394 bra $L__BB211_514; bra.uni $L__BB211_451; $L__BB211_514: { // callseq 1337, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1337 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_592: setp.gt.s32 %p325, %r698, 82; @%p325 bra $L__BB211_597; setp.eq.s32 %p329, %r698, 80; @%p329 bra $L__BB211_630; setp.eq.s32 %p330, %r698, 81; @%p330 bra $L__BB211_629; setp.eq.s32 %p331, %r698, 82; @%p331 bra $L__BB211_596; bra.uni $L__BB211_451; $L__BB211_596: mov.u32 %r711, 82; { // callseq 1292, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r711; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1292 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_490: setp.gt.s32 %p404, %r698, 22; @%p404 bra $L__BB211_495; setp.eq.s32 %p408, %r698, 20; @%p408 bra $L__BB211_673; setp.eq.s32 %p409, %r698, 21; @%p409 bra $L__BB211_672; setp.eq.s32 %p410, %r698, 22; @%p410 bra $L__BB211_494; bra.uni $L__BB211_451; $L__BB211_494: mov.u32 %r761, 22; add.s32 %r762, %r218, 1; { // callseq 1348, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r761; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; .param .b32 param3; st.param.b32 [param3+0], %r762; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj319EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1348 bra.uni $L__BB211_449; $L__BB211_569: setp.gt.s32 %p341, %r698, 70; @%p341 bra $L__BB211_574; setp.eq.s32 %p345, %r698, 68; @%p345 bra $L__BB211_641; setp.eq.s32 %p346, %r698, 69; @%p346 bra $L__BB211_640; setp.eq.s32 %p347, %r698, 70; @%p347 bra $L__BB211_573; bra.uni $L__BB211_451; $L__BB211_573: mov.u32 %r718, 70; { // callseq 1302, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r718; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1302 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_529: setp.gt.s32 %p373, %r698, 46; @%p373 bra $L__BB211_534; setp.eq.s32 %p377, %r698, 44; @%p377 bra $L__BB211_657; setp.eq.s32 %p378, %r698, 45; @%p378 bra $L__BB211_656; setp.eq.s32 %p379, %r698, 46; @%p379 bra $L__BB211_533; bra.uni $L__BB211_451; $L__BB211_533: { // callseq 1326, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1326 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_611: setp.gt.s32 %p310, %r698, 95; @%p310 bra $L__BB211_616; setp.eq.s32 %p314, %r698, 93; @%p314 bra $L__BB211_621; setp.eq.s32 %p315, %r698, 94; @%p315 bra $L__BB211_620; setp.eq.s32 %p316, %r698, 95; @%p316 bra $L__BB211_615; bra.uni $L__BB211_451; $L__BB211_615: { // callseq 1281, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1281 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_466: setp.eq.s32 %p427, %r698, 5; @%p427 bra $L__BB211_684; setp.eq.s32 %p428, %r698, 6; @%p428 bra $L__BB211_683; setp.eq.s32 %p429, %r698, 7; @%p429 bra $L__BB211_469; bra.uni $L__BB211_451; $L__BB211_469: mov.u32 %r778, 7; { // callseq 1358, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r778; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1358 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_546: setp.eq.s32 %p364, %r698, 53; @%p364 bra $L__BB211_651; setp.eq.s32 %p365, %r698, 54; @%p365 bra $L__BB211_650; setp.eq.s32 %p366, %r698, 55; @%p366 bra $L__BB211_549; bra.uni $L__BB211_451; $L__BB211_549: ld.local.v4.f32 {%f2798, %f2799, %f2800, %f2801}, [%rd71]; { // callseq 1317, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f2798; st.param.f32 [param0+4], %f2799; st.param.f32 [param0+8], %f2800; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; .param .b32 param4; st.param.b32 [param4+0], %r701; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1317 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_506: setp.eq.s32 %p396, %r698, 29; @%p396 bra $L__BB211_667; setp.eq.s32 %p397, %r698, 30; @%p397 bra $L__BB211_666; setp.eq.s32 %p398, %r698, 31; @%p398 bra $L__BB211_509; bra.uni $L__BB211_451; $L__BB211_509: mov.u32 %r751, 31; add.s32 %r752, %r218, 1; mov.u32 %r753, 768; { // callseq 1339, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b32 param1; st.param.b32 [param1+0], %r753; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r751; st.param.b32 [param3+4], %r699; st.param.b32 [param3+8], %r700; st.param.b32 [param3+12], %r701; .param .b32 param4; st.param.b32 [param4+0], %r752; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1339 bra.uni $L__BB211_449; $L__BB211_588: setp.eq.s32 %p333, %r698, 77; @%p333 bra $L__BB211_632; setp.eq.s32 %p334, %r698, 78; @%p334 bra $L__BB211_631; setp.eq.s32 %p335, %r698, 79; @%p335 bra $L__BB211_591; bra.uni $L__BB211_451; $L__BB211_591: add.s32 %r713, %r218, 1; { // callseq 1295, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r713; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1295 bra.uni $L__BB211_449; $L__BB211_486: setp.eq.s32 %p412, %r698, 17; @%p412 bra $L__BB211_675; setp.eq.s32 %p413, %r698, 18; @%p413 bra $L__BB211_674; setp.eq.s32 %p414, %r698, 19; @%p414 bra $L__BB211_489; bra.uni $L__BB211_451; $L__BB211_489: { // callseq 1350, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1350 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_565: setp.eq.s32 %p349, %r698, 65; @%p349 bra $L__BB211_643; setp.eq.s32 %p350, %r698, 66; @%p350 bra $L__BB211_642; setp.eq.s32 %p351, %r698, 67; @%p351 bra $L__BB211_568; bra.uni $L__BB211_451; $L__BB211_568: mov.u32 %r722, 67; { // callseq 1305, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r722, %r699, %r700, %r701}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1305 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_525: setp.eq.s32 %p381, %r698, 41; @%p381 bra $L__BB211_659; setp.eq.s32 %p382, %r698, 42; @%p382 bra $L__BB211_658; setp.eq.s32 %p383, %r698, 43; @%p383 bra $L__BB211_528; bra.uni $L__BB211_451; $L__BB211_528: add.s32 %r743, %r218, 1; { // callseq 1329, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r743; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1329 bra.uni $L__BB211_449; $L__BB211_607: setp.eq.s32 %p318, %r698, 89; @%p318 bra $L__BB211_624; setp.eq.s32 %p319, %r698, 90; @%p319 bra $L__BB211_623; setp.eq.s32 %p320, %r698, 92; @%p320 bra $L__BB211_610; bra.uni $L__BB211_451; $L__BB211_610: add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_476: setp.eq.s32 %p420, %r698, 11; @%p420 bra $L__BB211_679; setp.eq.s32 %p421, %r698, 12; @%p421 bra $L__BB211_678; setp.eq.s32 %p422, %r698, 13; @%p422 bra $L__BB211_479; bra.uni $L__BB211_451; $L__BB211_479: mov.u32 %r772, 13; add.s32 %r773, %r218, 1; mov.u32 %r774, 768; { // callseq 1354, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b32 param1; st.param.b32 [param1+0], %r774; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r772; st.param.b32 [param3+4], %r699; st.param.b32 [param3+8], %r700; st.param.b32 [param3+12], %r701; .param .b32 param4; st.param.b32 [param4+0], %r773; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1354 bra.uni $L__BB211_449; $L__BB211_555: setp.eq.s32 %p357, %r698, 59; @%p357 bra $L__BB211_647; setp.eq.s32 %p358, %r698, 60; @%p358 bra $L__BB211_646; setp.eq.s32 %p359, %r698, 61; @%p359 bra $L__BB211_558; bra.uni $L__BB211_451; $L__BB211_558: mov.u32 %r729, 61; add.s32 %r730, %r218, 1; { // callseq 1311, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r729; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r730; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1311 bra.uni $L__BB211_449; $L__BB211_515: setp.eq.s32 %p389, %r698, 35; @%p389 bra $L__BB211_663; setp.eq.s32 %p390, %r698, 36; @%p390 bra $L__BB211_662; setp.eq.s32 %p391, %r698, 37; @%p391 bra $L__BB211_518; bra.uni $L__BB211_451; $L__BB211_518: mov.u32 %r746, 37; { // callseq 1334, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r746, %r699, %r700, %r701}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1334 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_597: setp.eq.s32 %p326, %r698, 83; @%p326 bra $L__BB211_628; setp.eq.s32 %p327, %r698, 84; @%p327 bra $L__BB211_627; setp.eq.s32 %p328, %r698, 85; @%p328 bra $L__BB211_600; bra.uni $L__BB211_451; $L__BB211_600: add.s32 %r710, %r218, 1; { // callseq 1289, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r710; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1289 bra.uni $L__BB211_449; $L__BB211_495: setp.eq.s32 %p405, %r698, 23; @%p405 bra $L__BB211_671; setp.eq.s32 %p406, %r698, 24; @%p406 bra $L__BB211_670; setp.eq.s32 %p407, %r698, 25; @%p407 bra $L__BB211_498; bra.uni $L__BB211_451; $L__BB211_498: add.s32 %r757, %r218, 1; { // callseq 1345, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r757; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1345 bra.uni $L__BB211_449; $L__BB211_574: setp.eq.s32 %p342, %r698, 71; @%p342 bra $L__BB211_639; setp.eq.s32 %p343, %r698, 72; @%p343 bra $L__BB211_638; setp.eq.s32 %p344, %r698, 73; @%p344 bra $L__BB211_577; bra.uni $L__BB211_451; $L__BB211_577: add.s32 %r716, %r218, 1; { // callseq 1299, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r716; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1299 bra.uni $L__BB211_449; $L__BB211_534: setp.eq.s32 %p374, %r698, 47; @%p374 bra $L__BB211_655; setp.eq.s32 %p375, %r698, 48; @%p375 bra $L__BB211_654; setp.eq.s32 %p376, %r698, 49; @%p376 bra $L__BB211_537; bra.uni $L__BB211_451; $L__BB211_537: { // callseq 1323, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1323 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_616: setp.eq.s32 %p311, %r698, 98; @%p311 bra $L__BB211_687; setp.eq.s32 %p312, %r698, 97; @%p312 bra $L__BB211_622; setp.ne.s32 %p313, %r698, 96; @%p313 bra $L__BB211_451; bra.uni $L__BB211_619; $L__BB211_686: mov.u32 %r781, 2; add.s32 %r782, %r218, 1; mov.u32 %r783, 768; { // callseq 1361, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r781; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; .param .b32 param3; st.param.b32 [param3+0], %r783; .param .b32 param4; st.param.b32 [param4+0], %r782; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj319EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1361 bra.uni $L__BB211_449; $L__BB211_685: mov.u32 %r780, 3; { // callseq 1360, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r780, %r699, %r700, %r701}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1360 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_653: { // callseq 1322, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1322 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_652: { // callseq 1321, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1321 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_669: mov.u32 %r756, 26; { // callseq 1344, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r756; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z17svm_node_set_bumpILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1344 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_668: mov.u32 %r755, 27; { // callseq 1343, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r755; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1343 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_637: { // callseq 1298, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1298 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_636: { // callseq 1297, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1297 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_677: st.local.u32 [%rd125], %r699; add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_676: ld.global.v4.u32 {%r765, %r766, %r767, %r768}, [%rd123+16]; add.s32 %r903, %r218, 2; st.local.u32 [%rd124], %r766; st.local.u32 [%rd124+4], %r767; st.local.u32 [%rd124+8], %r768; bra.uni $L__BB211_449; $L__BB211_645: mov.u32 %r727, 62; add.s32 %r728, %r218, 1; { // callseq 1310, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r727; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r728; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1310 bra.uni $L__BB211_449; $L__BB211_644: mov.u32 %r726, 63; { // callseq 1309, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r726; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1309 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_661: { // callseq 1333, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; .param .b32 param4; st.param.b32 [param4+0], %r701; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1333 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_660: mov.u32 %r745, 39; { // callseq 1332, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r745, %r699, %r700, %r701}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1332 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_626: add.s32 %r709, %r218, 1; { // callseq 1288, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r709; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1288 bra.uni $L__BB211_449; $L__BB211_625: add.s32 %r708, %r218, 1; { // callseq 1287, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r708; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1287 bra.uni $L__BB211_449; $L__BB211_682: mov.u32 %r777, 8; { // callseq 1357, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r777; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1357 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_680: ld.local.f32 %f2822, [%rd125]; setp.gtu.ftz.f32 %p439, %f2822, 0f00000000; add.s32 %r903, %r218, 1; @%p439 bra $L__BB211_449; add.s32 %r776, %r699, %r218; add.s32 %r903, %r776, 1; bra.uni $L__BB211_449; $L__BB211_649: mov.u32 %r736, 56; { // callseq 1316, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r736; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1316 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_648: mov.u32 %r734, 57; add.s32 %r735, %r218, 1; { // callseq 1315, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r734; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r735; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1315 bra.uni $L__BB211_449; $L__BB211_665: mov.u32 %r748, 32; add.s32 %r749, %r218, 1; mov.u32 %r750, 768; { // callseq 1338, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b32 param1; st.param.b32 [param1+0], %r750; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r748; st.param.b32 [param3+4], %r699; st.param.b32 [param3+8], %r700; st.param.b32 [param3+12], %r701; .param .b32 param4; st.param.b32 [param4+0], %r749; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1338 bra.uni $L__BB211_449; $L__BB211_664: ld.local.f32 %f2805, [%rd124]; ld.local.f32 %f2806, [%rd124+4]; ld.local.f32 %f2807, [%rd124+8]; st.local.f32 [%rd22], %f2805; st.local.f32 [%rd64], %f2806; st.local.f32 [%rd65], %f2807; st.local.f32 [%rd125], %f2805; st.local.f32 [%rd125+4], %f2806; st.local.f32 [%rd125+8], %f2807; add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_630: { // callseq 1294, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1294 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_629: mov.u32 %r712, 81; { // callseq 1293, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r712; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1293 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_673: ld.local.f32 %f2808, [%rd124]; ld.local.v4.f32 {%f2809, %f2810, %f2811, %f2812}, [%rd71]; ld.local.f32 %f2816, [%rd124+4]; ld.local.f32 %f2817, [%rd124+8]; add.ftz.f32 %f2818, %f2817, %f2811; add.ftz.f32 %f2819, %f2816, %f2810; add.ftz.f32 %f2820, %f2808, %f2809; st.local.v2.f32 [%rd71], {%f2820, %f2819}; st.local.f32 [%rd28], %f2818; add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_672: mov.u32 %r763, 21; { // callseq 1349, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r763; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z21svm_node_displacementILj319EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1349 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_641: mov.u32 %r721, 68; { // callseq 1304, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r721; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1304 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_640: mov.u32 %r719, 69; add.s32 %r720, %r218, 1; { // callseq 1303, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r719; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r720; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1303 bra.uni $L__BB211_449; $L__BB211_657: mov.u32 %r741, 44; add.s32 %r742, %r218, 1; { // callseq 1328, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r741; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r742; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1328 bra.uni $L__BB211_449; $L__BB211_656: { // callseq 1327, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1327 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_621: add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_620: mov.u32 %r702, 94; add.s32 %r703, %r218, 1; { // callseq 1282, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r702; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r703; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1282 bra.uni $L__BB211_449; $L__BB211_684: st.local.u32 [%rd40], %r699; st.local.u32 [%rd41], %r700; st.local.u32 [%rd42], %r701; add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_683: ld.local.f32 %f2823, [%rd124]; ld.local.f32 %f2824, [%rd124+4]; ld.local.f32 %f2825, [%rd124+8]; st.local.f32 [%rd40], %f2823; st.local.f32 [%rd41], %f2824; st.local.f32 [%rd42], %f2825; add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_651: { // callseq 1319, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1319 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_650: add.s32 %r737, %r218, 1; { // callseq 1318, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r737; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1318 bra.uni $L__BB211_449; $L__BB211_667: { // callseq 1341, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; .param .b32 param4; st.param.b32 [param4+0], %r701; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1341 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_666: { // callseq 1340, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; .param .b32 param4; st.param.b32 [param4+0], %r701; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1340 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_632: setp.eq.s32 %p434, %r701, 255; add.s32 %r903, %r218, 1; @%p434 bra $L__BB211_449; ld.local.f32 %f2794, [%rd124]; add.s32 %r715, %r701, %r700; mul.wide.u32 %rd810, %r715, 4; add.s64 %rd811, %rd1, %rd810; st.local.f32 [%rd811], %f2794; bra.uni $L__BB211_449; $L__BB211_631: add.s32 %r714, %r218, 1; { // callseq 1296, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r714; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1296 bra.uni $L__BB211_449; $L__BB211_675: { // callseq 1352, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; .param .b32 param4; st.param.b32 [param4+0], %r701; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1352 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_674: { // callseq 1351, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1351 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_643: { // callseq 1307, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1307 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_642: add.s32 %r723, %r218, 1; { // callseq 1306, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r723; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1306 bra.uni $L__BB211_449; $L__BB211_659: add.s32 %r744, %r218, 1; { // callseq 1331, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r744; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1331 bra.uni $L__BB211_449; $L__BB211_658: { // callseq 1330, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1330 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_624: mov.u32 %r706, 89; { // callseq 1285, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r706; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z11svm_node_aoILj319EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1285 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_623: mov.u32 %r704, 90; add.s32 %r705, %r218, 1; { // callseq 1284, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r704; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; .param .b32 param3; st.param.b32 [param3+0], %r705; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1284 bra.uni $L__BB211_449; $L__BB211_679: { // callseq 1356, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1356 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_678: { // callseq 1355, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1355 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_647: add.s32 %r732, %r218, 1; { // callseq 1313, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r732; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj319EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1313 bra.uni $L__BB211_449; $L__BB211_646: add.s32 %r731, %r218, 1; { // callseq 1312, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; .param .b32 param4; st.param.b32 [param4+0], %r731; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1312 bra.uni $L__BB211_449; $L__BB211_663: { // callseq 1336, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1336 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_662: mov.u32 %r747, 36; { // callseq 1335, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r747; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1335 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_628: { // callseq 1291, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1291 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_627: { // callseq 1290, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1290 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_671: mov.u32 %r759, 23; add.s32 %r760, %r218, 1; { // callseq 1347, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r759; st.param.b32 [param1+4], %r699; st.param.b32 [param1+8], %r700; st.param.b32 [param1+12], %r701; .param .b32 param2; st.param.b32 [param2+0], %r760; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r903, [retval0+0]; } // callseq 1347 bra.uni $L__BB211_449; $L__BB211_670: mov.u32 %r758, 24; { // callseq 1346, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r758; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1346 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_639: mov.u32 %r717, 71; { // callseq 1301, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r717; st.param.b32 [param2+4], %r699; st.param.b32 [param2+8], %r700; st.param.b32 [param2+12], %r701; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1301 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_638: { // callseq 1300, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; .param .b32 param3; st.param.b32 [param3+0], %r701; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1300 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_655: cvt.u32.u64 %r739, %rd2; mov.u32 %r740, 768; { // callseq 1325, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r739; .param .b64 param1; st.param.b64 [param1+0], %rd166; .param .b64 param2; st.param.b64 [param2+0], %rd146; .param .b32 param3; st.param.b32 [param3+0], %r699; .param .b32 param4; st.param.b32 [param4+0], %r700; .param .b32 param5; st.param.b32 [param5+0], %r740; call.uni _Z19svm_node_light_pathILj319EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 1325 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_654: { // callseq 1324, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd166; .param .b64 param1; st.param.b64 [param1+0], %rd146; .param .b32 param2; st.param.b32 [param2+0], %r699; .param .b32 param3; st.param.b32 [param3+0], %r700; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1324 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_687: { // callseq 1362, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1362 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_622: { // callseq 1283, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd146; .param .b32 param1; st.param.b32 [param1+0], %r699; .param .b32 param2; st.param.b32 [param2+0], %r700; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1283 add.s32 %r903, %r218, 1; bra.uni $L__BB211_449; $L__BB211_635: ld.local.f32 %f2797, [%rd124]; st.local.f32 [%rd128], %f2797; bra.uni $L__BB211_449; $L__BB211_587: ld.local.f32 %f2795, [%rd127]; st.local.f32 [%rd128], %f2795; bra.uni $L__BB211_449; $L__BB211_451: ld.local.u32 %r324, [%rd31]; and.b32 %r784, %r324, 262144; setp.eq.s32 %p440, %r784, 0; @%p440 bra $L__BB211_696; bra.uni $L__BB211_452; $L__BB211_696: and.b32 %r798, %r324, 524288; setp.ne.s32 %p457, %r798, 0; mov.f32 %f3162, %f861; mov.f32 %f3163, %f861; mov.f32 %f3164, %f861; @%p457 bra $L__BB211_699; and.b32 %r799, %r324, 512; setp.eq.s32 %p458, %r799, 0; mov.f32 %f3162, 0f00000000; mov.f32 %f3163, %f3162; mov.f32 %f3164, %f3162; @%p458 bra $L__BB211_699; add.s64 %rd996, %rd166, 296; cvta.to.local.u64 %rd997, %rd996; ld.local.v2.f32 {%f3163, %f3164}, [%rd997]; ld.local.f32 %f3162, [%rd57]; $L__BB211_699: ld.global.f32 %f2834, [%rd11]; mul.ftz.f32 %f816, %f3162, %f2834; ld.global.f32 %f2835, [%rd11+4]; mul.ftz.f32 %f817, %f3163, %f2835; ld.global.f32 %f2836, [%rd11+8]; mul.ftz.f32 %f818, %f3164, %f2836; setp.eq.ftz.f32 %p459, %f816, 0f00000000; setp.eq.ftz.f32 %p460, %f817, 0f00000000; and.pred %p461, %p459, %p460; setp.eq.ftz.f32 %p462, %f818, 0f00000000; and.pred %p463, %p462, %p461; @%p463 bra $L__BB211_768; st.global.f32 [%rd11], %f816; st.global.f32 [%rd11+4], %f817; st.global.f32 [%rd11+8], %f818; ld.global.u16 %rs34, [%rd70]; add.s16 %rs35, %rs34, 1; st.global.u16 [%rd70], %rs35; ld.global.u16 %rs36, [%rd14]; add.s16 %rs37, %rs36, 16; st.global.u16 [%rd14], %rs37; $L__BB211_701: cvt.u32.u16 %r852, %rs1; min.u32 %r851, %r852, 4; add.s32 %r872, %r872, 1; setp.le.u32 %p464, %r872, %r851; @%p464 bra $L__BB211_4; setp.lt.u16 %p465, %rs1, 4; @%p465 bra $L__BB211_704; cvt.u32.u16 %r859, %rs1; min.u32 %r858, %r859, 4; add.s32 %r804, %r858, -1; mul.wide.u32 %rd998, %r804, 48; mov.u64 %rd999, kernel_params; add.s64 %rd1000, %rd999, %rd998; ld.const.u64 %rd1001, [%rd1000+3552]; cvta.to.global.u64 %rd1002, %rd1001; shl.b64 %rd1003, %rd2, 2; add.s64 %rd1004, %rd1002, %rd1003; ld.global.f32 %f2837, [%rd1004]; setp.eq.ftz.f32 %p466, %f2837, 0f00000000; mov.b32 %r805, %f2837; add.s32 %r806, %r805, 1; mov.b32 %f2838, %r806; selp.f32 %f2839, 0f00000001, %f2838, %p466; ld.const.u64 %rd1005, [kernel_params+3512]; cvta.to.global.u64 %rd1006, %rd1005; add.s64 %rd1007, %rd1006, %rd1003; st.global.f32 [%rd1007], %f2839; $L__BB211_704: @%p465 bra $L__BB211_706; bra.uni $L__BB211_705; $L__BB211_706: ld.global.f32 %f819, [%rd11]; ld.global.f32 %f820, [%rd11+4]; ld.global.f32 %f821, [%rd11+8]; ld.const.u64 %rd1016, [kernel_params+3376]; cvta.to.global.u64 %rd1017, %rd1016; shl.b64 %rd1018, %rd2, 1; add.s64 %rd1019, %rd1017, %rd1018; ld.global.u16 %rs21, [%rd1019]; mov.b32 %r343, %f819; setp.nan.ftz.f32 %p468, %f819, %f819; mov.f32 %f3170, 0f00000000; mov.f32 %f3165, %f3170; @%p468 bra $L__BB211_709; or.b32 %r809, %r343, -2147483648; setp.ne.s32 %p469, %r809, -2147483648; add.ftz.f32 %f2842, %f819, %f819; setp.eq.ftz.f32 %p470, %f2842, %f819; and.pred %p471, %p470, %p469; mov.f32 %f3165, %f3170; @%p471 bra $L__BB211_709; shl.b32 %r810, %r343, 1; setp.lt.u32 %p472, %r810, -16777215; selp.f32 %f3165, %f819, 0f00000000, %p472; $L__BB211_709: mov.b32 %r344, %f820; setp.nan.ftz.f32 %p473, %f820, %f820; mov.f32 %f3166, %f3170; @%p473 bra $L__BB211_712; or.b32 %r811, %r344, -2147483648; setp.ne.s32 %p474, %r811, -2147483648; add.ftz.f32 %f2845, %f820, %f820; setp.eq.ftz.f32 %p475, %f2845, %f820; and.pred %p476, %p475, %p474; mov.f32 %f3166, %f3170; @%p476 bra $L__BB211_712; shl.b32 %r812, %r344, 1; setp.lt.u32 %p477, %r812, -16777215; selp.f32 %f3166, %f820, 0f00000000, %p477; $L__BB211_712: mov.b32 %r345, %f821; setp.nan.ftz.f32 %p478, %f821, %f821; @%p478 bra $L__BB211_715; or.b32 %r813, %r345, -2147483648; setp.ne.s32 %p479, %r813, -2147483648; add.ftz.f32 %f2848, %f821, %f821; setp.eq.ftz.f32 %p480, %f2848, %f821; and.pred %p481, %p480, %p479; @%p481 bra $L__BB211_715; shl.b32 %r814, %r345, 1; setp.lt.u32 %p482, %r814, -16777215; selp.f32 %f3170, %f821, 0f00000000, %p482; $L__BB211_715: ld.const.f32 %f2849, [kernel_params+1584]; ld.const.f32 %f2850, [kernel_params+1580]; setp.eq.s16 %p483, %rs21, 0; selp.f32 %f828, %f2850, %f2849, %p483; abs.ftz.f32 %f2851, %f3166; abs.ftz.f32 %f2852, %f3165; add.ftz.f32 %f2853, %f2852, %f2851; abs.ftz.f32 %f2854, %f3170; add.ftz.f32 %f829, %f2854, %f2853; setp.leu.ftz.f32 %p484, %f829, %f828; @%p484 bra $L__BB211_717; div.approx.ftz.f32 %f2855, %f828, %f829; mul.ftz.f32 %f3165, %f3165, %f2855; mul.ftz.f32 %f3166, %f3166, %f2855; mul.ftz.f32 %f3170, %f3170, %f2855; $L__BB211_717: ld.const.u64 %rd1020, [kernel_params+3344]; cvta.to.global.u64 %rd1021, %rd1020; shl.b64 %rd1022, %rd2, 2; add.s64 %rd1023, %rd1021, %rd1022; ld.global.u32 %rd1024, [%rd1023]; ld.const.s32 %rd1025, [kernel_params+1264]; mul.lo.s64 %rd143, %rd1025, %rd1024; ld.const.u64 %rd1026, [kernel_params+3424]; cvta.to.global.u64 %rd1027, %rd1026; add.s64 %rd1028, %rd1027, %rd1022; ld.global.u32 %r346, [%rd15]; ld.global.u32 %r347, [%rd1028]; and.b32 %r815, %r347, 268435456; setp.eq.s32 %p485, %r815, 0; @%p485 bra $L__BB211_731; bra.uni $L__BB211_718; $L__BB211_731: ld.const.u32 %r825, [kernel_params+1620]; setp.eq.s32 %p500, %r825, 0; @%p500 bra $L__BB211_736; and.b32 %r826, %r347, 536870912; setp.ne.s32 %p501, %r826, 0; @%p501 bra $L__BB211_734; ld.const.s32 %rd1069, [kernel_params+1384]; add.s64 %rd1070, %rd143, %rd1069; cvta.to.global.u64 %rd1071, %rd145; shl.b64 %rd1072, %rd1070, 2; add.s64 %rd1073, %rd1071, %rd1072; add.s64 %rd1074, %rd1073, 4; add.s64 %rd1075, %rd1073, 8; atom.global.add.f32 %f2878, [%rd1073], %f3165; atom.global.add.f32 %f2879, [%rd1074], %f3166; atom.global.add.f32 %f2880, [%rd1075], %f3170; $L__BB211_734: and.b32 %r827, %r347, 1073741824; setp.eq.s32 %p502, %r827, 0; @%p502 bra $L__BB211_736; bra.uni $L__BB211_735; $L__BB211_736: ld.const.u32 %r910, [kernel_params+1260]; and.b32 %r828, %r910, 2; setp.eq.s32 %p503, %r828, 0; @%p503 bra $L__BB211_738; ld.const.s32 %rd1083, [kernel_params+1268]; add.s64 %rd1084, %rd143, %rd1083; cvta.to.global.u64 %rd1085, %rd145; shl.b64 %rd1086, %rd1084, 2; add.s64 %rd1087, %rd1085, %rd1086; add.s64 %rd1088, %rd1087, 4; add.s64 %rd1089, %rd1087, 8; atom.global.add.f32 %f2884, [%rd1087], %f3165; atom.global.add.f32 %f2885, [%rd1088], %f3166; atom.global.add.f32 %f2886, [%rd1089], %f3170; $L__BB211_738: ld.const.u32 %r351, [kernel_params+1400]; setp.eq.s32 %p504, %r351, -1; @%p504 bra $L__BB211_741; and.b32 %r829, %r346, -1431655766; popc.b32 %r830, %r829; and.b32 %r831, %r830, 1; setp.eq.b32 %p505, %r831, 1; mov.pred %p506, 0; xor.pred %p507, %p505, %p506; not.pred %p508, %p507; @%p508 bra $L__BB211_741; cvt.s64.s32 %rd1090, %r351; add.s64 %rd1091, %rd143, %rd1090; cvta.to.global.u64 %rd1092, %rd145; shl.b64 %rd1093, %rd1091, 2; add.s64 %rd1094, %rd1092, %rd1093; add.s64 %rd1095, %rd1094, 4; add.s64 %rd1096, %rd1094, 8; add.s64 %rd1097, %rd1094, 12; add.ftz.f32 %f2887, %f3165, %f3165; atom.global.add.f32 %f2888, [%rd1094], %f2887; add.ftz.f32 %f2889, %f3166, %f3166; atom.global.add.f32 %f2890, [%rd1095], %f2889; add.ftz.f32 %f2891, %f3170, %f3170; atom.global.add.f32 %f2892, [%rd1096], %f2891; atom.global.add.f32 %f2893, [%rd1097], 0f00000000; bra.uni $L__BB211_741; $L__BB211_768: ld.const.u64 %rd1157, [kernel_params+4256]; cvta.to.global.u64 %rd1158, %rd1157; add.s64 %rd1159, %rd1158, 48; atom.global.add.u32 %r847, [%rd1159], -1; ld.const.u64 %rd1160, [kernel_params+3416]; cvta.to.global.u64 %rd1161, %rd1160; shl.b64 %rd1162, %rd2, 1; add.s64 %rd1163, %rd1161, %rd1162; mov.u16 %rs42, 0; st.global.u16 [%rd1163], %rs42; bra.uni $L__BB211_769; $L__BB211_705: ld.const.u64 %rd1008, [kernel_params+4256]; cvta.to.global.u64 %rd1009, %rd1008; add.s64 %rd1010, %rd1009, 48; atom.global.add.u32 %r807, [%rd1010], -1; add.s64 %rd1011, %rd1009, 12; atom.global.add.u32 %r808, [%rd1011], 1; ld.const.u64 %rd1012, [kernel_params+3416]; cvta.to.global.u64 %rd1013, %rd1012; shl.b64 %rd1014, %rd2, 1; add.s64 %rd1015, %rd1013, %rd1014; mov.u16 %rs38, 3; st.global.u16 [%rd1015], %rs38; bra.uni $L__BB211_769; $L__BB211_718: and.b32 %r816, %r366, 8388608; setp.eq.s32 %p486, %r816, 0; and.b32 %r817, %r347, 1; setp.eq.b32 %p487, %r817, 1; not.pred %p488, %p487; or.pred %p489, %p488, %p486; @%p489 bra $L__BB211_720; ld.const.s32 %rd1029, [kernel_params+1360]; add.s64 %rd1030, %rd143, %rd1029; cvta.to.global.u64 %rd1031, %rd145; shl.b64 %rd1032, %rd1030, 2; add.s64 %rd1033, %rd1031, %rd1032; add.s64 %rd1034, %rd1033, 4; add.s64 %rd1035, %rd1033, 8; atom.global.add.f32 %f2856, [%rd1033], %f3165; atom.global.add.f32 %f2857, [%rd1034], %f3166; atom.global.add.f32 %f2858, [%rd1035], %f3170; $L__BB211_720: and.b32 %r818, %r366, 16777216; setp.eq.s32 %p490, %r818, 0; @%p490 bra $L__BB211_767; ld.const.u64 %rd1036, [kernel_params+3440]; cvta.to.global.u64 %rd1037, %rd1036; mul.lo.s64 %rd1038, %rd2, 12; add.s64 %rd1039, %rd1037, %rd1038; ld.global.f32 %f2859, [%rd1039]; mul.ftz.f32 %f836, %f3165, %f2859; ld.global.f32 %f2860, [%rd1039+4]; mul.ftz.f32 %f837, %f3166, %f2860; ld.global.f32 %f2861, [%rd1039+8]; mul.ftz.f32 %f838, %f3170, %f2861; ld.const.u32 %r819, [kernel_params+1620]; setp.eq.s32 %p491, %r819, 0; @%p491 bra $L__BB211_726; and.b32 %r820, %r347, 536870912; setp.ne.s32 %p492, %r820, 0; @%p492 bra $L__BB211_724; ld.const.s32 %rd1040, [kernel_params+1384]; add.s64 %rd1041, %rd143, %rd1040; cvta.to.global.u64 %rd1042, %rd145; shl.b64 %rd1043, %rd1041, 2; add.s64 %rd1044, %rd1042, %rd1043; add.s64 %rd1045, %rd1044, 4; add.s64 %rd1046, %rd1044, 8; atom.global.add.f32 %f2862, [%rd1044], %f836; atom.global.add.f32 %f2863, [%rd1045], %f837; atom.global.add.f32 %f2864, [%rd1046], %f838; $L__BB211_724: and.b32 %r821, %r347, 1073741824; setp.eq.s32 %p493, %r821, 0; @%p493 bra $L__BB211_726; bra.uni $L__BB211_725; $L__BB211_726: ld.const.u8 %rs39, [kernel_params+1260]; and.b16 %rs40, %rs39, 2; setp.eq.s16 %p494, %rs40, 0; @%p494 bra $L__BB211_728; ld.const.s32 %rd1054, [kernel_params+1268]; add.s64 %rd1055, %rd143, %rd1054; cvta.to.global.u64 %rd1056, %rd145; shl.b64 %rd1057, %rd1055, 2; add.s64 %rd1058, %rd1056, %rd1057; add.s64 %rd1059, %rd1058, 4; add.s64 %rd1060, %rd1058, 8; atom.global.add.f32 %f2868, [%rd1058], %f836; atom.global.add.f32 %f2869, [%rd1059], %f837; atom.global.add.f32 %f2870, [%rd1060], %f838; $L__BB211_728: ld.const.u32 %r348, [kernel_params+1400]; setp.eq.s32 %p495, %r348, -1; @%p495 bra $L__BB211_767; and.b32 %r822, %r346, -1431655766; popc.b32 %r823, %r822; and.b32 %r824, %r823, 1; setp.eq.b32 %p496, %r824, 1; mov.pred %p497, 0; xor.pred %p498, %p496, %p497; not.pred %p499, %p498; @%p499 bra $L__BB211_767; cvt.s64.s32 %rd1061, %r348; add.s64 %rd1062, %rd143, %rd1061; cvta.to.global.u64 %rd1063, %rd145; shl.b64 %rd1064, %rd1062, 2; add.s64 %rd1065, %rd1063, %rd1064; add.s64 %rd1066, %rd1065, 4; add.s64 %rd1067, %rd1065, 8; add.s64 %rd1068, %rd1065, 12; add.ftz.f32 %f2871, %f836, %f836; atom.global.add.f32 %f2872, [%rd1065], %f2871; add.ftz.f32 %f2873, %f837, %f837; atom.global.add.f32 %f2874, [%rd1066], %f2873; add.ftz.f32 %f2875, %f838, %f838; atom.global.add.f32 %f2876, [%rd1067], %f2875; atom.global.add.f32 %f2877, [%rd1068], 0f00000000; bra.uni $L__BB211_767; $L__BB211_735: ld.const.s32 %rd1076, [kernel_params+1376]; add.s64 %rd1077, %rd143, %rd1076; cvta.to.global.u64 %rd1078, %rd145; shl.b64 %rd1079, %rd1077, 2; add.s64 %rd1080, %rd1078, %rd1079; add.s64 %rd1081, %rd1080, 4; add.s64 %rd1082, %rd1080, 8; atom.global.add.f32 %f2881, [%rd1080], %f3165; atom.global.add.f32 %f2882, [%rd1081], %f3166; atom.global.add.f32 %f2883, [%rd1082], %f3170; ld.const.u32 %r910, [kernel_params+1260]; $L__BB211_741: setp.eq.s32 %p509, %r910, 0; @%p509 bra $L__BB211_767; and.b32 %r832, %r347, 536870912; setp.ne.s32 %p510, %r832, 0; @%p510 bra $L__BB211_767; ld.const.u64 %rd1098, [kernel_params+3472]; cvta.to.global.u64 %rd1099, %rd1098; add.s64 %rd1100, %rd1099, %rd2; ld.global.u8 %rs22, [%rd1100]; setp.eq.s16 %p511, %rs22, 0; ld.const.u32 %r353, [kernel_params+1444]; setp.eq.s32 %p512, %r353, -1; or.pred %p513, %p511, %p512; @%p513 bra $L__BB211_745; cvt.u32.u16 %r833, %rs22; mul.wide.u32 %rd1101, %r833, 3; cvt.s64.s32 %rd1102, %r353; cvta.to.global.u64 %rd1103, %rd145; add.s64 %rd1104, %rd1102, %rd1101; add.s64 %rd1105, %rd1104, %rd143; shl.b64 %rd1106, %rd1105, 2; add.s64 %rd1107, %rd1103, %rd1106; add.s64 %rd1108, %rd1107, -12; add.s64 %rd1109, %rd1107, -8; add.s64 %rd1110, %rd1107, -4; atom.global.add.f32 %f2894, [%rd1108], %f3165; atom.global.add.f32 %f2895, [%rd1109], %f3166; atom.global.add.f32 %f2896, [%rd1110], %f3170; $L__BB211_745: and.b32 %r834, %r366, 2097152; setp.eq.s32 %p514, %r834, 0; @%p514 bra $L__BB211_759; and.b32 %r835, %r347, 33554432; setp.eq.s32 %p515, %r835, 0; @%p515 bra $L__BB211_753; bra.uni $L__BB211_747; $L__BB211_753: and.b32 %r844, %r347, 67108864; setp.eq.s32 %p522, %r844, 0; mov.u32 %r911, -1; @%p522 bra $L__BB211_757; @%p483 bra $L__BB211_756; ld.const.u32 %r911, [kernel_params+1332]; bra.uni $L__BB211_757; $L__BB211_725: ld.const.s32 %rd1047, [kernel_params+1376]; add.s64 %rd1048, %rd143, %rd1047; cvta.to.global.u64 %rd1049, %rd145; shl.b64 %rd1050, %rd1048, 2; add.s64 %rd1051, %rd1049, %rd1050; add.s64 %rd1052, %rd1051, 4; add.s64 %rd1053, %rd1051, 8; atom.global.add.f32 %f2865, [%rd1051], %f836; atom.global.add.f32 %f2866, [%rd1052], %f837; atom.global.add.f32 %f2867, [%rd1053], %f838; bra.uni $L__BB211_767; $L__BB211_747: ld.const.u64 %rd1111, [kernel_params+3448]; cvta.to.global.u64 %rd1112, %rd1111; mul.lo.s64 %rd1113, %rd2, 12; add.s64 %rd1114, %rd1112, %rd1113; ld.global.f32 %f839, [%rd1114]; ld.global.f32 %f840, [%rd1114+4]; ld.global.f32 %f841, [%rd1114+8]; ld.const.u64 %rd1115, [kernel_params+3456]; cvta.to.global.u64 %rd1116, %rd1115; add.s64 %rd1117, %rd1116, %rd1113; ld.global.f32 %f842, [%rd1117]; ld.global.f32 %f843, [%rd1117+4]; ld.global.f32 %f844, [%rd1117+8]; ld.const.u32 %r836, [kernel_params+1324]; ld.const.u32 %r837, [kernel_params+1340]; selp.b32 %r354, %r837, %r836, %p483; setp.eq.s32 %p517, %r354, -1; @%p517 bra $L__BB211_749; cvt.s64.s32 %rd1118, %r354; add.s64 %rd1119, %rd143, %rd1118; cvta.to.global.u64 %rd1120, %rd145; shl.b64 %rd1121, %rd1119, 2; add.s64 %rd1122, %rd1120, %rd1121; add.s64 %rd1123, %rd1122, 4; add.s64 %rd1124, %rd1122, 8; mul.ftz.f32 %f2897, %f3165, %f842; atom.global.add.f32 %f2898, [%rd1122], %f2897; mul.ftz.f32 %f2899, %f3166, %f843; atom.global.add.f32 %f2900, [%rd1123], %f2899; mul.ftz.f32 %f2901, %f3170, %f844; atom.global.add.f32 %f2902, [%rd1124], %f2901; $L__BB211_749: ld.const.u32 %r838, [kernel_params+1328]; ld.const.u32 %r839, [kernel_params+1344]; selp.b32 %r355, %r839, %r838, %p483; setp.eq.s32 %p519, %r355, -1; @%p519 bra $L__BB211_751; sub.ftz.f32 %f2904, %f861, %f839; sub.ftz.f32 %f2905, %f2904, %f842; sub.ftz.f32 %f2906, %f861, %f840; sub.ftz.f32 %f2907, %f2906, %f843; sub.ftz.f32 %f2908, %f861, %f841; sub.ftz.f32 %f2909, %f2908, %f844; cvt.s64.s32 %rd1125, %r355; add.s64 %rd1126, %rd143, %rd1125; cvta.to.global.u64 %rd1127, %rd145; shl.b64 %rd1128, %rd1126, 2; add.s64 %rd1129, %rd1127, %rd1128; mul.ftz.f32 %f2910, %f3165, %f2905; mul.ftz.f32 %f2911, %f3166, %f2907; mul.ftz.f32 %f2912, %f3170, %f2909; add.s64 %rd1130, %rd1129, 4; add.s64 %rd1131, %rd1129, 8; atom.global.add.f32 %f2913, [%rd1129], %f2910; atom.global.add.f32 %f2914, [%rd1130], %f2911; atom.global.add.f32 %f2915, [%rd1131], %f2912; $L__BB211_751: ld.const.u32 %r841, [kernel_params+1320]; ld.const.u32 %r842, [kernel_params+1336]; selp.b32 %r356, %r842, %r841, %p483; setp.eq.s32 %p521, %r356, -1; mov.u32 %r911, -1; @%p521 bra $L__BB211_757; mul.ftz.f32 %f3165, %f3165, %f839; mul.ftz.f32 %f3166, %f3166, %f840; mul.ftz.f32 %f3170, %f3170, %f841; mov.u32 %r911, %r356; bra.uni $L__BB211_757; $L__BB211_756: ld.const.u32 %r911, [kernel_params+1348]; $L__BB211_757: setp.eq.s32 %p524, %r911, -1; @%p524 bra $L__BB211_759; cvt.s64.s32 %rd1132, %r911; add.s64 %rd1133, %rd143, %rd1132; cvta.to.global.u64 %rd1134, %rd145; shl.b64 %rd1135, %rd1133, 2; add.s64 %rd1136, %rd1134, %rd1135; add.s64 %rd1137, %rd1136, 4; add.s64 %rd1138, %rd1136, 8; atom.global.add.f32 %f2916, [%rd1136], %f3165; atom.global.add.f32 %f2917, [%rd1137], %f3166; atom.global.add.f32 %f2918, [%rd1138], %f3170; $L__BB211_759: ld.const.u32 %r360, [kernel_params+1368]; setp.eq.s32 %p525, %r360, -1; and.b32 %r845, %r347, 134234112; setp.ne.s32 %p526, %r845, 134234112; or.pred %p527, %p526, %p525; @%p527 bra $L__BB211_767; ld.const.u64 %rd1139, [kernel_params+3440]; cvta.to.global.u64 %rd1140, %rd1139; mul.lo.s64 %rd1141, %rd2, 12; add.s64 %rd1142, %rd1140, %rd1141; ld.global.f32 %f851, [%rd1142+4]; ld.global.f32 %f852, [%rd1142+8]; ld.global.f32 %f853, [%rd1142]; setp.eq.ftz.f32 %p528, %f853, 0f00000000; mov.f32 %f3176, 0f00000000; mov.f32 %f3174, %f3176; @%p528 bra $L__BB211_762; div.approx.ftz.f32 %f3174, %f819, %f853; $L__BB211_762: setp.eq.ftz.f32 %p529, %f851, 0f00000000; mov.f32 %f3175, %f3176; @%p529 bra $L__BB211_764; div.approx.ftz.f32 %f3175, %f820, %f851; $L__BB211_764: setp.eq.ftz.f32 %p530, %f852, 0f00000000; @%p530 bra $L__BB211_766; div.approx.ftz.f32 %f3176, %f821, %f852; $L__BB211_766: ld.const.f32 %f2922, [kernel_params+1372]; mul.ftz.f32 %f2923, %f3174, %f2922; mul.ftz.f32 %f2924, %f3175, %f2922; mul.ftz.f32 %f2925, %f3176, %f2922; cvta.to.global.u64 %rd1143, %rd145; cvt.s64.s32 %rd1144, %r360; add.s64 %rd1145, %rd143, %rd1144; shl.b64 %rd1146, %rd1145, 2; add.s64 %rd1147, %rd1143, %rd1146; add.s64 %rd1148, %rd1147, 4; add.s64 %rd1149, %rd1147, 8; atom.global.add.f32 %f2926, [%rd1147], %f2923; atom.global.add.f32 %f2927, [%rd1148], %f2924; atom.global.add.f32 %f2928, [%rd1149], %f2925; $L__BB211_767: ld.const.u64 %rd1150, [kernel_params+4256]; cvta.to.global.u64 %rd1151, %rd1150; add.s64 %rd1152, %rd1151, 48; atom.global.add.u32 %r846, [%rd1152], -1; ld.const.u64 %rd1153, [kernel_params+3416]; cvta.to.global.u64 %rd1154, %rd1153; add.s64 %rd1156, %rd1154, %rd1018; mov.u16 %rs41, 0; st.global.u16 [%rd1156], %rs41; $L__BB211_769: ret; } // .globl kernel_gpu_integrator_shade_surface .visible .entry kernel_gpu_integrator_shade_surface( .param .u64 kernel_gpu_integrator_shade_surface_param_0, .param .u64 kernel_gpu_integrator_shade_surface_param_1, .param .u32 kernel_gpu_integrator_shade_surface_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<5>; .reg .b16 %rs<5>; .reg .b32 %r<16>; .reg .b64 %rd<31>; ld.param.u64 %rd3, [kernel_gpu_integrator_shade_surface_param_0]; ld.param.u64 %rd4, [kernel_gpu_integrator_shade_surface_param_1]; ld.param.u32 %r4, [kernel_gpu_integrator_shade_surface_param_2]; mov.u32 %r5, %ntid.x; mov.u32 %r6, %ctaid.x; mov.u32 %r7, %tid.x; mad.lo.s32 %r15, %r6, %r5, %r7; setp.ge.s32 %p1, %r15, %r4; @%p1 bra $L__BB212_8; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB212_3; cvta.to.global.u64 %rd5, %rd3; mul.wide.s32 %rd6, %r15, 4; add.s64 %rd7, %rd5, %rd6; ld.global.u32 %r15, [%rd7]; $L__BB212_3: { // callseq 1363, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r15; .param .b64 param1; st.param.b64 [param1+0], %rd4; .param .b32 retval0; call.uni (retval0), _Z17integrate_surfaceILj447EEbPK16KernelGlobalsGPUiPf, ( param0, param1 ); ld.param.b32 %r8, [retval0+0]; } // callseq 1363 cvt.u16.u32 %rs1, %r8; setp.eq.s16 %p3, %rs1, 0; cvt.s64.s32 %rd1, %r15; @%p3 bra $L__BB212_7; ld.const.u64 %rd8, [kernel_params+2544]; cvta.to.global.u64 %rd9, %rd8; shl.b64 %rd10, %rd1, 2; add.s64 %rd11, %rd9, %rd10; ld.global.u32 %r9, [%rd11]; and.b32 %r10, %r9, 15728640; setp.eq.s32 %p4, %r10, 0; ld.const.u64 %rd12, [kernel_params+4256]; cvta.to.global.u64 %rd2, %rd12; add.s64 %rd13, %rd2, 32; atom.global.add.u32 %r11, [%rd13], -1; @%p4 bra $L__BB212_6; add.s64 %rd14, %rd2, 16; atom.global.add.u32 %r12, [%rd14], 1; ld.const.u64 %rd15, [kernel_params+2520]; cvta.to.global.u64 %rd16, %rd15; shl.b64 %rd17, %rd1, 1; add.s64 %rd18, %rd16, %rd17; mov.u16 %rs2, 4; st.global.u16 [%rd18], %rs2; bra.uni $L__BB212_8; $L__BB212_7: ld.const.u64 %rd24, [kernel_params+4256]; cvta.to.global.u64 %rd25, %rd24; add.s64 %rd26, %rd25, 32; atom.global.add.u32 %r14, [%rd26], -1; ld.const.u64 %rd27, [kernel_params+2520]; cvta.to.global.u64 %rd28, %rd27; shl.b64 %rd29, %rd1, 1; add.s64 %rd30, %rd28, %rd29; mov.u16 %rs4, 0; st.global.u16 [%rd30], %rs4; bra.uni $L__BB212_8; $L__BB212_6: add.s64 %rd19, %rd2, 8; atom.global.add.u32 %r13, [%rd19], 1; ld.const.u64 %rd20, [kernel_params+2520]; cvta.to.global.u64 %rd21, %rd20; shl.b64 %rd22, %rd1, 1; add.s64 %rd23, %rd21, %rd22; mov.u16 %rs3, 2; st.global.u16 [%rd23], %rs3; $L__BB212_8: ret; } // .globl kernel_gpu_integrator_shade_surface_raytrace .visible .entry kernel_gpu_integrator_shade_surface_raytrace( .param .u64 kernel_gpu_integrator_shade_surface_raytrace_param_0, .param .u64 kernel_gpu_integrator_shade_surface_raytrace_param_1, .param .u32 kernel_gpu_integrator_shade_surface_raytrace_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<5>; .reg .b16 %rs<5>; .reg .b32 %r<16>; .reg .b64 %rd<31>; ld.param.u64 %rd3, [kernel_gpu_integrator_shade_surface_raytrace_param_0]; ld.param.u64 %rd4, [kernel_gpu_integrator_shade_surface_raytrace_param_1]; ld.param.u32 %r4, [kernel_gpu_integrator_shade_surface_raytrace_param_2]; mov.u32 %r5, %ntid.x; mov.u32 %r6, %ctaid.x; mov.u32 %r7, %tid.x; mad.lo.s32 %r15, %r6, %r5, %r7; setp.ge.s32 %p1, %r15, %r4; @%p1 bra $L__BB213_8; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB213_3; cvta.to.global.u64 %rd5, %rd3; mul.wide.s32 %rd6, %r15, 4; add.s64 %rd7, %rd5, %rd6; ld.global.u32 %r15, [%rd7]; $L__BB213_3: { // callseq 1364, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r15; .param .b64 param1; st.param.b64 [param1+0], %rd4; .param .b32 retval0; call.uni (retval0), _Z17integrate_surfaceILj511EEbPK16KernelGlobalsGPUiPf, ( param0, param1 ); ld.param.b32 %r8, [retval0+0]; } // callseq 1364 cvt.u16.u32 %rs1, %r8; setp.eq.s16 %p3, %rs1, 0; cvt.s64.s32 %rd1, %r15; @%p3 bra $L__BB213_7; ld.const.u64 %rd8, [kernel_params+2544]; cvta.to.global.u64 %rd9, %rd8; shl.b64 %rd10, %rd1, 2; add.s64 %rd11, %rd9, %rd10; ld.global.u32 %r9, [%rd11]; and.b32 %r10, %r9, 15728640; setp.eq.s32 %p4, %r10, 0; ld.const.u64 %rd12, [kernel_params+4256]; cvta.to.global.u64 %rd2, %rd12; add.s64 %rd13, %rd2, 36; atom.global.add.u32 %r11, [%rd13], -1; @%p4 bra $L__BB213_6; add.s64 %rd14, %rd2, 16; atom.global.add.u32 %r12, [%rd14], 1; ld.const.u64 %rd15, [kernel_params+2520]; cvta.to.global.u64 %rd16, %rd15; shl.b64 %rd17, %rd1, 1; add.s64 %rd18, %rd16, %rd17; mov.u16 %rs2, 4; st.global.u16 [%rd18], %rs2; bra.uni $L__BB213_8; $L__BB213_7: ld.const.u64 %rd24, [kernel_params+4256]; cvta.to.global.u64 %rd25, %rd24; add.s64 %rd26, %rd25, 36; atom.global.add.u32 %r14, [%rd26], -1; ld.const.u64 %rd27, [kernel_params+2520]; cvta.to.global.u64 %rd28, %rd27; shl.b64 %rd29, %rd1, 1; add.s64 %rd30, %rd28, %rd29; mov.u16 %rs4, 0; st.global.u16 [%rd30], %rs4; bra.uni $L__BB213_8; $L__BB213_6: add.s64 %rd19, %rd2, 8; atom.global.add.u32 %r13, [%rd19], 1; ld.const.u64 %rd20, [kernel_params+2520]; cvta.to.global.u64 %rd21, %rd20; shl.b64 %rd22, %rd1, 1; add.s64 %rd23, %rd21, %rd22; mov.u16 %rs3, 2; st.global.u16 [%rd23], %rs3; $L__BB213_8: ret; } // .globl kernel_gpu_integrator_shade_surface_mnee .visible .entry kernel_gpu_integrator_shade_surface_mnee( .param .u64 kernel_gpu_integrator_shade_surface_mnee_param_0, .param .u64 kernel_gpu_integrator_shade_surface_mnee_param_1, .param .u32 kernel_gpu_integrator_shade_surface_mnee_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<5>; .reg .b16 %rs<5>; .reg .b32 %r<16>; .reg .b64 %rd<31>; ld.param.u64 %rd3, [kernel_gpu_integrator_shade_surface_mnee_param_0]; ld.param.u64 %rd4, [kernel_gpu_integrator_shade_surface_mnee_param_1]; ld.param.u32 %r4, [kernel_gpu_integrator_shade_surface_mnee_param_2]; mov.u32 %r5, %ntid.x; mov.u32 %r6, %ctaid.x; mov.u32 %r7, %tid.x; mad.lo.s32 %r15, %r6, %r5, %r7; setp.ge.s32 %p1, %r15, %r4; @%p1 bra $L__BB214_8; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB214_3; cvta.to.global.u64 %rd5, %rd3; mul.wide.s32 %rd6, %r15, 4; add.s64 %rd7, %rd5, %rd6; ld.global.u32 %r15, [%rd7]; $L__BB214_3: { // callseq 1365, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r15; .param .b64 param1; st.param.b64 [param1+0], %rd4; .param .b32 retval0; call.uni (retval0), _Z17integrate_surfaceILj33554879EEbPK16KernelGlobalsGPUiPf, ( param0, param1 ); ld.param.b32 %r8, [retval0+0]; } // callseq 1365 cvt.u16.u32 %rs1, %r8; setp.eq.s16 %p3, %rs1, 0; cvt.s64.s32 %rd1, %r15; @%p3 bra $L__BB214_7; ld.const.u64 %rd8, [kernel_params+2544]; cvta.to.global.u64 %rd9, %rd8; shl.b64 %rd10, %rd1, 2; add.s64 %rd11, %rd9, %rd10; ld.global.u32 %r9, [%rd11]; and.b32 %r10, %r9, 15728640; setp.eq.s32 %p4, %r10, 0; ld.const.u64 %rd12, [kernel_params+4256]; cvta.to.global.u64 %rd2, %rd12; add.s64 %rd13, %rd2, 40; atom.global.add.u32 %r11, [%rd13], -1; @%p4 bra $L__BB214_6; add.s64 %rd14, %rd2, 16; atom.global.add.u32 %r12, [%rd14], 1; ld.const.u64 %rd15, [kernel_params+2520]; cvta.to.global.u64 %rd16, %rd15; shl.b64 %rd17, %rd1, 1; add.s64 %rd18, %rd16, %rd17; mov.u16 %rs2, 4; st.global.u16 [%rd18], %rs2; bra.uni $L__BB214_8; $L__BB214_7: ld.const.u64 %rd24, [kernel_params+4256]; cvta.to.global.u64 %rd25, %rd24; add.s64 %rd26, %rd25, 40; atom.global.add.u32 %r14, [%rd26], -1; ld.const.u64 %rd27, [kernel_params+2520]; cvta.to.global.u64 %rd28, %rd27; shl.b64 %rd29, %rd1, 1; add.s64 %rd30, %rd28, %rd29; mov.u16 %rs4, 0; st.global.u16 [%rd30], %rs4; bra.uni $L__BB214_8; $L__BB214_6: add.s64 %rd19, %rd2, 8; atom.global.add.u32 %r13, [%rd19], 1; ld.const.u64 %rd20, [kernel_params+2520]; cvta.to.global.u64 %rd21, %rd20; shl.b64 %rd22, %rd1, 1; add.s64 %rd23, %rd21, %rd22; mov.u16 %rs3, 2; st.global.u16 [%rd23], %rs3; $L__BB214_8: ret; } // .globl kernel_gpu_integrator_shade_volume .visible .entry kernel_gpu_integrator_shade_volume( .param .u64 kernel_gpu_integrator_shade_volume_param_0, .param .u64 kernel_gpu_integrator_shade_volume_param_1, .param .u32 kernel_gpu_integrator_shade_volume_param_2 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot215[7200]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<801>; .reg .b16 %rs<116>; .reg .f32 %f<3363>; .reg .b32 %r<2246>; .reg .b64 %rd<1753>; mov.u64 %SPL, __local_depot215; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd225, [kernel_gpu_integrator_shade_volume_param_0]; ld.param.u64 %rd226, [kernel_gpu_integrator_shade_volume_param_1]; ld.param.u32 %r482, [kernel_gpu_integrator_shade_volume_param_2]; add.u64 %rd227, %SP, 1328; add.u64 %rd1, %SPL, 1328; mov.u32 %r483, %ntid.x; mov.u32 %r484, %ctaid.x; mov.u32 %r485, %tid.x; mad.lo.s32 %r2152, %r484, %r483, %r485; setp.ge.s32 %p4, %r2152, %r482; @%p4 bra $L__BB215_1097; setp.eq.s64 %p5, %rd225, 0; @%p5 bra $L__BB215_3; cvta.to.global.u64 %rd228, %rd225; mul.wide.s32 %rd229, %r2152, 4; add.s64 %rd230, %rd228, %rd229; ld.global.u32 %r2152, [%rd230]; $L__BB215_3: ld.const.u64 %rd2, [kernel_params+2632]; cvta.to.global.u64 %rd3, %rd2; cvt.s64.s32 %rd4, %r2152; mul.wide.s32 %rd231, %r2152, 12; add.s64 %rd5, %rd3, %rd231; ld.global.f32 %f1, [%rd5]; ld.global.f32 %f2, [%rd5+4]; ld.global.f32 %f3, [%rd5+8]; ld.const.v2.u64 {%rd232, %rd233}, [kernel_params+2640]; cvta.to.global.u64 %rd234, %rd232; add.s64 %rd235, %rd234, %rd231; ld.global.f32 %f4, [%rd235]; ld.global.f32 %f5, [%rd235+4]; ld.global.f32 %f6, [%rd235+8]; cvta.to.global.u64 %rd236, %rd233; mul.wide.s32 %rd237, %r2152, 4; add.s64 %rd238, %rd236, %rd237; ld.global.f32 %f7, [%rd238]; ld.const.v2.u64 {%rd239, %rd240}, [kernel_params+2656]; cvta.to.global.u64 %rd11, %rd240; add.s64 %rd12, %rd11, %rd237; ld.global.f32 %f8, [%rd12]; ld.const.u64 %rd13, [kernel_params+2712]; cvta.to.global.u64 %rd14, %rd13; add.s64 %rd15, %rd14, %rd237; ld.global.s32 %rd16, [%rd15]; ld.const.v2.u64 {%rd241, %rd242}, [kernel_params+2720]; cvta.to.global.u64 %rd243, %rd241; add.s64 %rd244, %rd243, %rd237; ld.global.u32 %r4, [%rd244]; cvta.to.global.u64 %rd245, %rd242; add.s64 %rd246, %rd245, %rd237; ld.global.u32 %r5, [%rd246]; ld.const.u64 %rd19, [kernel_params+2688]; cvta.to.global.u64 %rd20, %rd19; add.s64 %rd21, %rd20, %rd237; setp.ne.s64 %p6, %rd16, -1; setp.eq.s64 %p7, %rd16, -1; ld.global.f32 %f1004, [%rd21]; selp.f32 %f9, 0f7F7FFFFF, %f1004, %p7; @%p6 bra $L__BB215_7; ld.const.u32 %r486, [kernel_params+1028]; setp.eq.s32 %p8, %r486, -1; @%p8 bra $L__BB215_6; ld.const.u64 %rd247, [kernel_params+2792]; cvta.to.global.u64 %rd248, %rd247; shl.b64 %rd249, %rd4, 2; add.s64 %rd250, %rd248, %rd249; mov.u32 %r487, -1; st.global.u32 [%rd250], %r487; bra.uni $L__BB215_7; $L__BB215_6: ld.const.u64 %rd251, [kernel_params+2776]; cvta.to.global.u64 %rd252, %rd251; shl.b64 %rd253, %rd4, 2; add.s64 %rd254, %rd252, %rd253; mov.u32 %r488, -1; st.global.u32 [%rd254], %r488; $L__BB215_7: fma.rn.ftz.f32 %f10, %f4, %f7, %f1; fma.rn.ftz.f32 %f11, %f5, %f7, %f2; fma.rn.ftz.f32 %f12, %f6, %f7, %f3; mov.u32 %r2161, 1; neg.ftz.f32 %f1005, %f4; st.local.v4.f32 [%rd1], {%f10, %f11, %f12, %f1005}; add.s64 %rd23, %rd1, 16; neg.ftz.f32 %f1006, %f6; neg.ftz.f32 %f1007, %f5; st.local.v4.f32 [%rd1+16], {%f1007, %f1006, %f1005, %f1007}; st.local.v4.f32 [%rd1+32], {%f1006, %f1005, %f1007, %f1006}; mov.u32 %r2163, 0; mov.u32 %r491, -1; st.local.v4.u32 [%rd1+48], {%r491, %r2163, %r2163, %r491}; mov.u32 %r492, 16; st.local.v4.u32 [%rd1+64], {%r492, %r2163, %r2163, %r491}; mov.f32 %f3167, 0f00000000; st.local.v2.f32 [%rd1+128], {%f3167, %f3167}; st.local.u32 [%rd1+136], %r2163; mov.b32 %r493, %f8; st.local.v4.u32 [%rd1+80], {%r491, %r493, %r2163, %r2163}; st.local.v4.f32 [%rd1+96], {%f3167, %f3167, %f3167, %f3167}; st.local.v4.f32 [%rd1+112], {%f3167, %f3167, %f3167, %f3167}; st.local.v2.f32 [%rd1+240], {%f1, %f2}; st.local.f32 [%rd1+248], %f3; ld.const.v2.u64 {%rd257, %rd258}, [kernel_params+2528]; cvta.to.global.u64 %rd259, %rd257; shl.b64 %rd260, %rd4, 2; add.s64 %rd261, %rd259, %rd260; ld.global.u32 %r6, [%rd261]; cvta.to.global.u64 %rd262, %rd258; shl.b64 %rd263, %rd4, 1; add.s64 %rd264, %rd262, %rd263; ld.global.u16 %rs1, [%rd264]; cvt.u32.u16 %r7, %rs1; ld.const.u64 %rd26, [kernel_params+2456]; cvta.to.global.u64 %rd27, %rd26; add.s64 %rd28, %rd27, %rd260; ld.global.u32 %r8, [%rd28]; ld.const.u64 %rd29, [kernel_params+2544]; cvta.to.global.u64 %rd265, %rd29; add.s64 %rd30, %rd265, %rd260; ld.global.u32 %r9, [%rd30]; and.b32 %r494, %r9, 491520; setp.ne.s32 %p9, %r494, 0; add.u64 %rd266, %SP, 6752; add.u64 %rd31, %SPL, 6752; @%p9 bra $L__BB215_26; ld.const.u32 %r495, [kernel_params+1488]; setp.eq.s32 %p10, %r495, 0; mov.u16 %rs109, 0; @%p10 bra $L__BB215_18; ld.const.u64 %rd267, [kernel_params+2464]; cvta.to.global.u64 %rd268, %rd267; add.s64 %rd270, %rd268, %rd263; ld.global.u16 %rs2, [%rd270]; add.s32 %r10, %r7, 1; ld.const.u32 %r496, [kernel_params+1596]; setp.eq.s32 %p11, %r496, 0; brev.b32 %r497, %r8; mul.lo.s32 %r498, %r497, 1025551850; xor.b32 %r11, %r498, %r497; @%p11 bra $L__BB215_12; ld.const.v2.u32 {%r499, %r500}, [kernel_params+1600]; mov.b32 %f13, %r500; setp.geu.ftz.f32 %p12, %f13, 0f3F800000; setp.lt.ftz.f32 %p13, %f13, 0f3F800000; ld.const.u32 %r503, [kernel_params+1576]; selp.b32 %r504, %r503, %r6, %p13; add.s16 %rs25, %rs1, 1; and.b16 %rs26, %rs25, 255; cvt.u32.u16 %r505, %rs26; xor.b32 %r506, %r504, %r505; mul.lo.s32 %r507, %r506, -512718531; shr.u32 %r508, %r504, 16; xor.b32 %r509, %r507, %r508; bfe.u32 %r510, %r509, 4, 4; shr.u32 %r511, %r504, 8; xor.b32 %r512, %r509, %r511; xor.b32 %r513, %r512, %r510; mul.lo.s32 %r514, %r513, 153742143; shr.u32 %r515, %r504, 23; xor.b32 %r516, %r514, %r515; bfe.u32 %r517, %r516, 1, 7; xor.b32 %r518, %r517, %r516; shr.u32 %r519, %r504, 27; or.b32 %r520, %r519, 1; mul.lo.s32 %r521, %r520, %r518; mul.lo.s32 %r522, %r521, -1074439621; bfe.u32 %r523, %r522, 2, 6; xor.b32 %r524, %r523, %r522; mul.lo.s32 %r525, %r524, -1638916925; bfe.u32 %r526, %r525, 2, 6; xor.b32 %r527, %r526, %r525; mul.lo.s32 %r528, %r527, -933190689; and.b32 %r529, %r528, 255; bfe.u32 %r530, %r528, 5, 3; xor.b32 %r531, %r530, %r529; add.s32 %r532, %r499, -1; xor.b32 %r12, %r10, 61; xor.b32 %r533, %r504, %r12; mul.lo.s32 %r534, %r533, 9; shr.u32 %r535, %r534, 4; xor.b32 %r536, %r535, %r534; mul.lo.s32 %r537, %r536, 668265261; add.s32 %r538, %r537, %r11; shr.u32 %r539, %r537, 16; or.b32 %r540, %r539, 1; mul.lo.s32 %r541, %r540, %r538; mul.lo.s32 %r542, %r541, 89287766; xor.b32 %r543, %r542, %r541; mul.lo.s32 %r544, %r543, 1403136100; xor.b32 %r545, %r544, %r543; brev.b32 %r546, %r545; neg.s32 %r547, %r499; and.b32 %r548, %r8, %r547; and.b32 %r549, %r546, %r532; or.b32 %r550, %r549, %r548; mad.lo.s32 %r551, %r531, %r499, %r550; shl.b32 %r552, %r499, 8; rem.u32 %r553, %r551, %r552; ld.const.u64 %rd271, [kernel_params+2424]; cvta.to.global.u64 %rd272, %rd271; shl.b32 %r554, %r553, 1; mul.wide.u32 %rd273, %r554, 4; add.s64 %rd274, %rd272, %rd273; ld.global.f32 %f3011, [%rd274]; ld.global.f32 %f3012, [%rd274+4]; @%p12 bra $L__BB215_16; xor.b32 %r555, %r12, %r6; mul.lo.s32 %r556, %r555, 9; shr.u32 %r557, %r556, 4; xor.b32 %r558, %r557, %r556; mul.lo.s32 %r559, %r558, 668265261; cvt.rn.f32.u32 %f1009, %r559; mul.ftz.f32 %f1010, %f1009, 0f2F7FFFFE; xor.b32 %r560, %r6, %r10; xor.b32 %r561, %r560, -905047700; mul.lo.s32 %r562, %r561, 9; shr.u32 %r563, %r562, 4; xor.b32 %r564, %r563, %r562; mul.lo.s32 %r565, %r564, 668265261; cvt.rn.f32.u32 %f1011, %r565; mul.ftz.f32 %f1012, %f1011, 0f2F7FFFFE; fma.rn.ftz.f32 %f1013, %f1010, %f13, %f3011; fma.rn.ftz.f32 %f1014, %f1012, %f13, %f3012; cvt.rmi.ftz.f32.f32 %f1015, %f1013; sub.ftz.f32 %f3011, %f1013, %f1015; cvt.rmi.ftz.f32.f32 %f1016, %f1014; sub.ftz.f32 %f3012, %f1014, %f1016; bra.uni $L__BB215_16; $L__BB215_12: shr.u32 %r567, %r10, 16; xor.b32 %r568, %r567, %r10; mul.lo.s32 %r569, %r568, 569420461; shr.u32 %r570, %r569, 15; xor.b32 %r571, %r570, %r569; mul.lo.s32 %r572, %r571, -749064809; shr.u32 %r573, %r572, 15; xor.b32 %r574, %r6, %r572; xor.b32 %r575, %r574, %r573; xor.b32 %r13, %r575, -419546133; xor.b32 %r576, %r575, 508809841; add.s32 %r577, %r576, %r11; shr.u32 %r578, %r576, 16; or.b32 %r579, %r578, 1; mul.lo.s32 %r580, %r579, %r577; mul.lo.s32 %r581, %r580, 89287766; xor.b32 %r582, %r581, %r580; mul.lo.s32 %r583, %r582, 1403136100; xor.b32 %r14, %r583, %r582; setp.eq.s32 %p14, %r14, 0; mov.u32 %r2156, 0; @%p14 bra $L__BB215_15; mov.u32 %r2153, 0; mov.u64 %rd276, sobol_burley_table; mov.u32 %r2154, %r14; mov.u32 %r2156, %r2153; $L__BB215_14: clz.b32 %r586, %r2154; add.s32 %r587, %r586, %r2153; mul.wide.u32 %rd275, %r587, 4; add.s64 %rd277, %rd276, %rd275; ld.const.u32 %r588, [%rd277+128]; xor.b32 %r2156, %r588, %r2156; add.s32 %r2153, %r587, 1; shl.b32 %r589, %r2154, %r586; shl.b32 %r2154, %r589, 1; setp.ne.s32 %p15, %r2154, 0; @%p15 bra $L__BB215_14; $L__BB215_15: xor.b32 %r590, %r13, -525684874; brev.b32 %r591, %r14; shr.u32 %r592, %r590, 16; or.b32 %r593, %r592, 1; mul.lo.s32 %r594, %r591, 1025551850; xor.b32 %r595, %r594, %r591; add.s32 %r596, %r595, %r590; mul.lo.s32 %r597, %r596, %r593; mul.lo.s32 %r598, %r597, 89287766; xor.b32 %r599, %r598, %r597; xor.b32 %r600, %r13, -1802089138; mul.lo.s32 %r601, %r2156, 1025551850; xor.b32 %r602, %r601, %r2156; add.s32 %r603, %r602, %r600; shr.u32 %r604, %r600, 16; or.b32 %r605, %r604, 1; mul.lo.s32 %r606, %r603, %r605; mul.lo.s32 %r607, %r606, 89287766; xor.b32 %r608, %r607, %r606; mul.lo.s32 %r609, %r608, 1403136100; xor.b32 %r610, %r609, %r608; brev.b32 %r611, %r610; cvt.rn.f32.u32 %f1017, %r611; mul.lo.s32 %r612, %r599, 1403136100; xor.b32 %r613, %r612, %r599; brev.b32 %r614, %r613; cvt.rn.f32.u32 %f1018, %r614; mul.ftz.f32 %f3012, %f1017, 0f2F7FFFFE; mul.ftz.f32 %f3011, %f1018, 0f2F7FFFFE; $L__BB215_16: cvt.u32.u16 %r615, %rs2; { // callseq 1366, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f3011; .param .b32 param1; st.param.f32 [param1+0], %f3012; .param .b32 param2; st.param.f32 [param2+0], %f8; .param .align 4 .b8 param3[12]; st.param.f32 [param3+0], %f10; st.param.f32 [param3+4], %f11; st.param.f32 [param3+8], %f12; .param .b32 param4; st.param.b32 [param4+0], %r615; .param .b32 param5; st.param.b32 [param5+0], %r9; .param .b64 param6; st.param.b64 [param6+0], %rd266; .param .b32 retval0; call.uni (retval0), _Z25light_distribution_sampleILb1EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample, ( param0, param1, param2, param3, param4, param5, param6 ); ld.param.b32 %r616, [retval0+0]; } // callseq 1366 cvt.u16.u32 %rs28, %r616; setp.eq.s16 %p16, %rs28, 0; @%p16 bra $L__BB215_18; ld.local.u8 %rs29, [%rd31+66]; and.b16 %rs30, %rs29, 128; setp.eq.s16 %p17, %rs30, 0; selp.u16 %rs109, 1, 0, %p17; $L__BB215_18: ld.local.f32 %f1019, [%rd31+36]; setp.eq.ftz.f32 %p18, %f1019, 0f7F7FFFFF; setp.eq.s16 %p19, %rs109, 0; or.pred %p20, %p19, %p18; @%p20 bra $L__BB215_26; ld.const.u64 %rd279, [kernel_params+2776]; cvta.to.global.u64 %rd280, %rd279; add.s64 %rd282, %rd280, %rd260; ld.global.u32 %r2157, [%rd282]; setp.eq.s32 %p21, %r2157, -1; mov.u32 %r2161, 0; @%p21 bra $L__BB215_26; ld.const.u64 %rd283, [kernel_params+2408]; cvta.to.global.u64 %rd32, %rd283; mov.u32 %r2158, 0; mov.u32 %r621, 3; mov.u64 %rd288, kernel_params; mov.u32 %r2159, %r2158; $L__BB215_21: shl.b32 %r622, %r2157, 5; cvt.u64.u32 %rd284, %r622; and.b64 %rd285, %rd284, 134217696; add.s64 %rd286, %rd32, %rd285; ld.global.u32 %r26, [%rd286+16]; and.b32 %r623, %r26, 8388608; setp.ne.s32 %p22, %r623, 0; mov.u32 %r2161, %r621; @%p22 bra $L__BB215_26; and.b32 %r624, %r26, 4194304; setp.eq.s32 %p23, %r624, 0; @%p23 bra $L__BB215_24; bra.uni $L__BB215_23; $L__BB215_24: setp.eq.s32 %p25, %r2159, 2; mov.u32 %r2159, 1; mov.u32 %r2161, %r621; @%p25 bra $L__BB215_26; bra.uni $L__BB215_25; $L__BB215_23: setp.eq.s32 %p24, %r2159, 1; mov.u32 %r2159, 2; mov.u32 %r2161, %r621; @%p24 bra $L__BB215_26; $L__BB215_25: add.s32 %r2158, %r2158, 1; mul.wide.s32 %rd287, %r2158, 16; add.s64 %rd289, %rd288, %rd287; ld.const.u64 %rd290, [%rd289+2776]; cvta.to.global.u64 %rd291, %rd290; add.s64 %rd293, %rd291, %rd260; ld.global.u32 %r2157, [%rd293]; setp.ne.s32 %p26, %r2157, -1; mov.u32 %r2161, %r2159; @%p26 bra $L__BB215_21; $L__BB215_26: ld.const.u64 %rd33, [kernel_params+2776]; cvta.to.global.u64 %rd34, %rd33; add.s64 %rd295, %rd34, %rd260; ld.global.u32 %r2162, [%rd295]; setp.eq.s32 %p27, %r2162, -1; mov.f32 %f3015, 0f7F7FFFFF; @%p27 bra $L__BB215_36; mov.u64 %rd297, kernel_params; ld.const.u64 %rd298, [kernel_params+2408]; cvta.to.global.u64 %rd35, %rd298; ld.const.u64 %rd299, [kernel_params+2184]; cvta.to.global.u64 %rd36, %rd299; ld.const.f32 %f26, [kernel_params+1032]; ld.const.f32 %f27, [kernel_params+1616]; ld.const.u64 %rd300, [kernel_params+2192]; add.s64 %rd1742, %rd297, 2768; cvta.to.global.u64 %rd37, %rd300; mov.f32 %f3015, 0f7F7FFFFF; mov.u16 %rs31, 1; mov.u16 %rs32, 0; $L__BB215_28: ld.const.u64 %rd301, [%rd1742]; cvta.to.global.u64 %rd302, %rd301; add.s64 %rd304, %rd302, %rd260; ld.global.s32 %rd39, [%rd304]; shl.b32 %r630, %r2162, 5; cvt.u64.u32 %rd305, %r630; and.b64 %rd306, %rd305, 134217696; add.s64 %rd307, %rd35, %rd306; ld.global.u32 %r34, [%rd307+16]; and.b32 %r631, %r34, 1048576; setp.ne.s32 %p28, %r631, 0; mov.u16 %rs110, %rs31; @%p28 bra $L__BB215_31; cvt.u32.u64 %r632, %rd39; and.b32 %r633, %r34, 268435456; setp.eq.s32 %p29, %r633, 0; setp.eq.s32 %p30, %r632, -1; or.pred %p31, %p30, %p29; mov.u16 %rs110, %rs32; @%p31 bra $L__BB215_31; shl.b64 %rd308, %rd39, 2; add.s64 %rd309, %rd36, %rd308; ld.global.u8 %rs33, [%rd309+1]; and.b16 %rs110, %rs33, 1; $L__BB215_31: setp.eq.s16 %p32, %rs110, 0; @%p32 bra $L__BB215_35; cvt.u32.u64 %r634, %rd39; setp.eq.s32 %p33, %r634, -1; mov.f32 %f3014, %f26; @%p33 bra $L__BB215_34; shl.b64 %rd310, %rd39, 2; add.s64 %rd311, %rd37, %rd310; ld.global.f32 %f3014, [%rd311]; $L__BB215_34: mul.ftz.f32 %f1022, %f3014, %f27; min.ftz.f32 %f3015, %f1022, %f3015; $L__BB215_35: add.s32 %r2163, %r2163, 1; mul.wide.s32 %rd312, %r2163, 16; add.s64 %rd314, %rd297, %rd312; add.s64 %rd1742, %rd314, 2768; ld.const.u64 %rd315, [%rd314+2776]; cvta.to.global.u64 %rd316, %rd315; add.s64 %rd318, %rd316, %rd260; ld.global.u32 %r2162, [%rd318]; setp.ne.s32 %p34, %r2162, -1; @%p34 bra $L__BB215_28; $L__BB215_36: add.u64 %rd42, %SPL, 6832; cvt.s64.s32 %rd43, %r4; mov.u64 %rd1743, 0; mov.u32 %r635, 0; $L__BB215_37: shl.b64 %rd321, %rd1743, 2; add.s64 %rd322, %rd42, %rd321; st.local.u32 [%rd322], %r635; add.s64 %rd1743, %rd1743, 1; setp.lt.u64 %p35, %rd1743, 92; @%p35 bra $L__BB215_37; ld.local.f32 %f34, [%rd31]; ld.local.f32 %f35, [%rd31+4]; ld.local.f32 %f36, [%rd31+8]; sub.ftz.f32 %f37, %f9, %f7; setp.eq.ftz.f32 %p36, %f3015, 0f7F7FFFFF; @%p36 bra $L__BB215_49; bra.uni $L__BB215_39; $L__BB215_49: mov.u32 %r2165, 1; mov.f32 %f3021, 0f3F800000; mov.f32 %f3020, 0f00000000; ld.const.u32 %r2164, [kernel_params+1596]; mov.f32 %f3019, %f37; $L__BB215_50: brev.b32 %r824, %r8; mul.lo.s32 %r825, %r824, 1025551850; xor.b32 %r47, %r825, %r824; add.s32 %r48, %r7, 5; setp.eq.s32 %p44, %r2164, 0; @%p44 bra $L__BB215_53; ld.const.v2.u32 {%r826, %r827}, [kernel_params+1600]; mov.b32 %f54, %r827; setp.geu.ftz.f32 %p45, %f54, 0f3F800000; setp.lt.ftz.f32 %p46, %f54, 0f3F800000; ld.const.u32 %r830, [kernel_params+1576]; selp.b32 %r831, %r830, %r6, %p46; add.s16 %rs38, %rs1, 5; and.b16 %rs39, %rs38, 255; cvt.u32.u16 %r832, %rs39; xor.b32 %r833, %r831, %r832; mul.lo.s32 %r834, %r833, -512718531; shr.u32 %r835, %r831, 16; xor.b32 %r836, %r834, %r835; bfe.u32 %r837, %r836, 4, 4; shr.u32 %r838, %r831, 8; xor.b32 %r839, %r836, %r838; xor.b32 %r840, %r839, %r837; mul.lo.s32 %r841, %r840, 153742143; shr.u32 %r842, %r831, 23; xor.b32 %r843, %r841, %r842; bfe.u32 %r844, %r843, 1, 7; xor.b32 %r845, %r844, %r843; shr.u32 %r846, %r831, 27; or.b32 %r847, %r846, 1; mul.lo.s32 %r848, %r847, %r845; mul.lo.s32 %r849, %r848, -1074439621; bfe.u32 %r850, %r849, 2, 6; xor.b32 %r851, %r850, %r849; mul.lo.s32 %r852, %r851, -1638916925; bfe.u32 %r853, %r852, 2, 6; xor.b32 %r854, %r853, %r852; mul.lo.s32 %r855, %r854, -933190689; and.b32 %r856, %r855, 255; bfe.u32 %r857, %r855, 5, 3; xor.b32 %r858, %r857, %r856; add.s32 %r859, %r826, -1; xor.b32 %r49, %r48, 61; xor.b32 %r860, %r831, %r49; mul.lo.s32 %r861, %r860, 9; shr.u32 %r862, %r861, 4; xor.b32 %r863, %r862, %r861; mul.lo.s32 %r864, %r863, 668265261; add.s32 %r865, %r864, %r47; shr.u32 %r866, %r864, 16; or.b32 %r867, %r866, 1; mul.lo.s32 %r868, %r867, %r865; mul.lo.s32 %r869, %r868, 89287766; xor.b32 %r870, %r869, %r868; mul.lo.s32 %r871, %r870, 1403136100; xor.b32 %r872, %r871, %r870; brev.b32 %r873, %r872; neg.s32 %r874, %r826; and.b32 %r875, %r8, %r874; and.b32 %r876, %r873, %r859; or.b32 %r877, %r876, %r875; mad.lo.s32 %r878, %r858, %r826, %r877; shl.b32 %r879, %r826, 8; rem.u32 %r880, %r878, %r879; ld.const.u64 %rd331, [kernel_params+2424]; cvta.to.global.u64 %rd332, %rd331; shl.b32 %r881, %r880, 1; mul.wide.u32 %rd333, %r881, 4; add.s64 %rd334, %rd332, %rd333; ld.global.f32 %f3166, [%rd334]; @%p45 bra $L__BB215_54; xor.b32 %r882, %r49, %r6; mul.lo.s32 %r883, %r882, 9; shr.u32 %r884, %r883, 4; xor.b32 %r885, %r884, %r883; mul.lo.s32 %r886, %r885, 668265261; cvt.rn.f32.u32 %f1036, %r886; mul.ftz.f32 %f1037, %f1036, 0f2F7FFFFE; fma.rn.ftz.f32 %f1038, %f1037, %f54, %f3166; cvt.rmi.ftz.f32.f32 %f1039, %f1038; sub.ftz.f32 %f3166, %f1038, %f1039; bra.uni $L__BB215_54; $L__BB215_39: ld.const.u32 %r2165, [kernel_params+1612]; cvt.rn.f32.s32 %f38, %r2165; min.ftz.f32 %f3019, %f3015, %f37; mul.ftz.f32 %f1023, %f3019, %f38; setp.leu.ftz.f32 %p37, %f37, %f1023; @%p37 bra $L__BB215_41; div.approx.ftz.f32 %f3019, %f37, %f38; $L__BB215_41: add.s32 %r38, %r7, 7; ld.const.u32 %r2164, [kernel_params+1596]; setp.eq.s32 %p38, %r2164, 0; brev.b32 %r636, %r8; mul.lo.s32 %r637, %r636, 1025551850; xor.b32 %r40, %r637, %r636; @%p38 bra $L__BB215_44; ld.const.v2.u32 {%r638, %r639}, [kernel_params+1600]; mov.b32 %f42, %r639; setp.geu.ftz.f32 %p39, %f42, 0f3F800000; setp.lt.ftz.f32 %p40, %f42, 0f3F800000; ld.const.u32 %r642, [kernel_params+1576]; selp.b32 %r643, %r642, %r6, %p40; add.s16 %rs34, %rs1, 7; and.b16 %rs35, %rs34, 255; cvt.u32.u16 %r644, %rs35; xor.b32 %r645, %r643, %r644; mul.lo.s32 %r646, %r645, -512718531; shr.u32 %r647, %r643, 16; xor.b32 %r648, %r646, %r647; bfe.u32 %r649, %r648, 4, 4; shr.u32 %r650, %r643, 8; xor.b32 %r651, %r648, %r650; xor.b32 %r652, %r651, %r649; mul.lo.s32 %r653, %r652, 153742143; shr.u32 %r654, %r643, 23; xor.b32 %r655, %r653, %r654; bfe.u32 %r656, %r655, 1, 7; xor.b32 %r657, %r656, %r655; shr.u32 %r658, %r643, 27; or.b32 %r659, %r658, 1; mul.lo.s32 %r660, %r659, %r657; mul.lo.s32 %r661, %r660, -1074439621; bfe.u32 %r662, %r661, 2, 6; xor.b32 %r663, %r662, %r661; mul.lo.s32 %r664, %r663, -1638916925; bfe.u32 %r665, %r664, 2, 6; xor.b32 %r666, %r665, %r664; mul.lo.s32 %r667, %r666, -933190689; and.b32 %r668, %r667, 255; bfe.u32 %r669, %r667, 5, 3; xor.b32 %r670, %r669, %r668; add.s32 %r671, %r638, -1; xor.b32 %r41, %r38, 61; xor.b32 %r672, %r643, %r41; mul.lo.s32 %r673, %r672, 9; shr.u32 %r674, %r673, 4; xor.b32 %r675, %r674, %r673; mul.lo.s32 %r676, %r675, 668265261; add.s32 %r677, %r676, %r40; shr.u32 %r678, %r676, 16; or.b32 %r679, %r678, 1; mul.lo.s32 %r680, %r679, %r677; mul.lo.s32 %r681, %r680, 89287766; xor.b32 %r682, %r681, %r680; mul.lo.s32 %r683, %r682, 1403136100; xor.b32 %r684, %r683, %r682; brev.b32 %r685, %r684; neg.s32 %r686, %r638; and.b32 %r687, %r8, %r686; and.b32 %r688, %r685, %r671; or.b32 %r689, %r688, %r687; mad.lo.s32 %r690, %r670, %r638, %r689; shl.b32 %r691, %r638, 8; rem.u32 %r692, %r690, %r691; ld.const.u64 %rd323, [kernel_params+2424]; cvta.to.global.u64 %rd324, %rd323; shl.b32 %r693, %r692, 1; mul.wide.u32 %rd325, %r693, 4; add.s64 %rd326, %rd324, %rd325; ld.global.f32 %f3020, [%rd326]; @%p39 bra $L__BB215_45; xor.b32 %r694, %r41, %r6; mul.lo.s32 %r695, %r694, 9; shr.u32 %r696, %r695, 4; xor.b32 %r697, %r696, %r695; mul.lo.s32 %r698, %r697, 668265261; cvt.rn.f32.u32 %f1024, %r698; mul.ftz.f32 %f1025, %f1024, 0f2F7FFFFE; fma.rn.ftz.f32 %f1026, %f1025, %f42, %f3020; cvt.rmi.ftz.f32.f32 %f1027, %f1026; sub.ftz.f32 %f3020, %f1026, %f1027; bra.uni $L__BB215_45; $L__BB215_53: shr.u32 %r887, %r48, 16; xor.b32 %r888, %r887, %r48; mul.lo.s32 %r889, %r888, 569420461; shr.u32 %r890, %r889, 15; xor.b32 %r891, %r890, %r889; mul.lo.s32 %r892, %r891, -749064809; shr.u32 %r893, %r892, 15; xor.b32 %r894, %r6, %r892; xor.b32 %r895, %r894, %r893; xor.b32 %r896, %r895, 1493655573; add.s32 %r897, %r896, %r47; shr.u32 %r898, %r896, 16; or.b32 %r899, %r898, 1; mul.lo.s32 %r900, %r899, %r897; mul.lo.s32 %r901, %r900, 89287766; xor.b32 %r902, %r901, %r900; mul.lo.s32 %r903, %r902, 1403136100; xor.b32 %r904, %r903, %r902; xor.b32 %r905, %r895, -2052961194; brev.b32 %r906, %r904; mul.lo.s32 %r907, %r906, 1025551850; xor.b32 %r908, %r907, %r906; add.s32 %r909, %r908, %r905; shr.u32 %r910, %r905, 16; or.b32 %r911, %r910, 1; mul.lo.s32 %r912, %r909, %r911; mul.lo.s32 %r913, %r912, 89287766; xor.b32 %r914, %r913, %r912; mul.lo.s32 %r915, %r914, 1403136100; xor.b32 %r916, %r915, %r914; brev.b32 %r917, %r916; cvt.rn.f32.u32 %f1040, %r917; mul.ftz.f32 %f3166, %f1040, 0f2F7FFFFE; $L__BB215_54: add.s32 %r50, %r7, 4; @%p44 bra $L__BB215_57; ld.const.v2.u32 {%r918, %r919}, [kernel_params+1600]; mov.b32 %f59, %r919; setp.geu.ftz.f32 %p48, %f59, 0f3F800000; setp.lt.ftz.f32 %p49, %f59, 0f3F800000; ld.const.u32 %r922, [kernel_params+1576]; selp.b32 %r923, %r922, %r6, %p49; add.s16 %rs40, %rs1, 4; and.b16 %rs41, %rs40, 255; cvt.u32.u16 %r924, %rs41; xor.b32 %r925, %r923, %r924; mul.lo.s32 %r926, %r925, -512718531; shr.u32 %r927, %r923, 16; xor.b32 %r928, %r926, %r927; bfe.u32 %r929, %r928, 4, 4; shr.u32 %r930, %r923, 8; xor.b32 %r931, %r928, %r930; xor.b32 %r932, %r931, %r929; mul.lo.s32 %r933, %r932, 153742143; shr.u32 %r934, %r923, 23; xor.b32 %r935, %r933, %r934; bfe.u32 %r936, %r935, 1, 7; xor.b32 %r937, %r936, %r935; shr.u32 %r938, %r923, 27; or.b32 %r939, %r938, 1; mul.lo.s32 %r940, %r939, %r937; mul.lo.s32 %r941, %r940, -1074439621; bfe.u32 %r942, %r941, 2, 6; xor.b32 %r943, %r942, %r941; mul.lo.s32 %r944, %r943, -1638916925; bfe.u32 %r945, %r944, 2, 6; xor.b32 %r946, %r945, %r944; mul.lo.s32 %r947, %r946, -933190689; and.b32 %r948, %r947, 255; bfe.u32 %r949, %r947, 5, 3; xor.b32 %r950, %r949, %r948; add.s32 %r951, %r918, -1; xor.b32 %r51, %r50, 61; xor.b32 %r952, %r923, %r51; mul.lo.s32 %r953, %r952, 9; shr.u32 %r954, %r953, 4; xor.b32 %r955, %r954, %r953; mul.lo.s32 %r956, %r955, 668265261; add.s32 %r957, %r956, %r47; shr.u32 %r958, %r956, 16; or.b32 %r959, %r958, 1; mul.lo.s32 %r960, %r959, %r957; mul.lo.s32 %r961, %r960, 89287766; xor.b32 %r962, %r961, %r960; mul.lo.s32 %r963, %r962, 1403136100; xor.b32 %r964, %r963, %r962; brev.b32 %r965, %r964; neg.s32 %r966, %r918; and.b32 %r967, %r8, %r966; and.b32 %r968, %r965, %r951; or.b32 %r969, %r968, %r967; mad.lo.s32 %r970, %r950, %r918, %r969; shl.b32 %r971, %r918, 8; rem.u32 %r972, %r970, %r971; ld.const.u64 %rd335, [kernel_params+2424]; cvta.to.global.u64 %rd336, %rd335; shl.b32 %r973, %r972, 1; mul.wide.u32 %rd337, %r973, 4; add.s64 %rd338, %rd336, %rd337; ld.global.f32 %f3023, [%rd338]; @%p48 bra $L__BB215_58; xor.b32 %r974, %r51, %r6; mul.lo.s32 %r975, %r974, 9; shr.u32 %r976, %r975, 4; xor.b32 %r977, %r976, %r975; mul.lo.s32 %r978, %r977, 668265261; cvt.rn.f32.u32 %f1041, %r978; mul.ftz.f32 %f1042, %f1041, 0f2F7FFFFE; fma.rn.ftz.f32 %f1043, %f1042, %f59, %f3023; cvt.rmi.ftz.f32.f32 %f1044, %f1043; sub.ftz.f32 %f3023, %f1043, %f1044; bra.uni $L__BB215_58; $L__BB215_57: shr.u32 %r979, %r50, 16; xor.b32 %r980, %r979, %r50; mul.lo.s32 %r981, %r980, 569420461; shr.u32 %r982, %r981, 15; xor.b32 %r983, %r982, %r981; mul.lo.s32 %r984, %r983, -749064809; shr.u32 %r985, %r984, 15; xor.b32 %r986, %r6, %r984; xor.b32 %r987, %r986, %r985; xor.b32 %r988, %r987, 1493655573; add.s32 %r989, %r988, %r47; shr.u32 %r990, %r988, 16; or.b32 %r991, %r990, 1; mul.lo.s32 %r992, %r991, %r989; mul.lo.s32 %r993, %r992, 89287766; xor.b32 %r994, %r993, %r992; mul.lo.s32 %r995, %r994, 1403136100; xor.b32 %r996, %r995, %r994; xor.b32 %r997, %r987, -2052961194; brev.b32 %r998, %r996; mul.lo.s32 %r999, %r998, 1025551850; xor.b32 %r1000, %r999, %r998; add.s32 %r1001, %r1000, %r997; shr.u32 %r1002, %r997, 16; or.b32 %r1003, %r1002, 1; mul.lo.s32 %r1004, %r1001, %r1003; mul.lo.s32 %r1005, %r1004, 89287766; xor.b32 %r1006, %r1005, %r1004; mul.lo.s32 %r1007, %r1006, 1403136100; xor.b32 %r1008, %r1007, %r1006; brev.b32 %r1009, %r1008; cvt.rn.f32.u32 %f1045, %r1009; mul.ftz.f32 %f3023, %f1045, 0f2F7FFFFE; $L__BB215_58: setp.ne.s32 %p50, %r2161, 3; @%p50 bra $L__BB215_62; setp.lt.ftz.f32 %p51, %f3166, 0f3F000000; @%p51 bra $L__BB215_61; bra.uni $L__BB215_60; $L__BB215_61: add.ftz.f32 %f3166, %f3166, %f3166; mov.u32 %r2161, 1; bra.uni $L__BB215_62; $L__BB215_60: add.ftz.f32 %f1046, %f3166, 0fBF000000; add.ftz.f32 %f3166, %f1046, %f1046; mov.u32 %r2161, 2; $L__BB215_62: ld.const.u64 %rd46, [kernel_params+2584]; cvta.to.global.u64 %rd47, %rd46; mul.lo.s64 %rd339, %rd4, 12; add.s64 %rd48, %rd47, %rd339; ld.global.f32 %f1048, [%rd48]; ld.global.f32 %f1049, [%rd48+4]; ld.global.f32 %f1050, [%rd48+8]; add.s64 %rd49, %rd42, 4; st.local.f32 [%rd42+4], %f1048; st.local.f32 [%rd42+8], %f1049; st.local.f32 [%rd42+12], %f1050; st.local.f32 [%rd42+188], %f1048; st.local.f32 [%rd42+192], %f1049; st.local.f32 [%rd42+196], %f1050; setp.ne.s32 %p52, %r2161, 2; mov.f32 %f3029, %f3167; @%p52 bra $L__BB215_77; sub.ftz.f32 %f1053, %f34, %f1; sub.ftz.f32 %f1054, %f35, %f2; mul.ftz.f32 %f1055, %f5, %f1054; fma.rn.ftz.f32 %f1056, %f4, %f1053, %f1055; sub.ftz.f32 %f1057, %f36, %f3; fma.rn.ftz.f32 %f67, %f6, %f1057, %f1056; mul.ftz.f32 %f1058, %f1054, %f1054; fma.rn.ftz.f32 %f1059, %f1053, %f1053, %f1058; fma.rn.ftz.f32 %f1060, %f1057, %f1057, %f1059; mul.ftz.f32 %f1061, %f67, %f67; sub.ftz.f32 %f1062, %f1060, %f1061; mov.f32 %f3029, 0f00000000; max.ftz.f32 %f1063, %f1062, %f3029; sqrt.approx.ftz.f32 %f68, %f1063; setp.eq.ftz.f32 %p53, %f68, 0f00000000; mov.f32 %f3028, %f3029; @%p53 bra $L__BB215_76; sub.ftz.f32 %f1064, %f7, %f67; abs.ftz.f32 %f69, %f1064; abs.ftz.f32 %f70, %f68; setp.eq.ftz.f32 %p54, %f70, 0f00000000; setp.eq.ftz.f32 %p55, %f69, 0f00000000; and.pred %p56, %p54, %p55; mov.b32 %r53, %f68; mov.b32 %r1012, %f1064; and.b32 %r54, %r1012, -2147483648; @%p56 bra $L__BB215_68; bra.uni $L__BB215_65; $L__BB215_68: shr.s32 %r1017, %r53, 31; and.b32 %r1018, %r1017, 1078530011; or.b32 %r1019, %r1018, %r54; mov.b32 %f3025, %r1019; bra.uni $L__BB215_69; $L__BB215_44: shr.u32 %r699, %r38, 16; xor.b32 %r700, %r699, %r38; mul.lo.s32 %r701, %r700, 569420461; shr.u32 %r702, %r701, 15; xor.b32 %r703, %r702, %r701; mul.lo.s32 %r704, %r703, -749064809; shr.u32 %r705, %r704, 15; xor.b32 %r706, %r6, %r704; xor.b32 %r707, %r706, %r705; xor.b32 %r708, %r707, 1493655573; add.s32 %r709, %r708, %r40; shr.u32 %r710, %r708, 16; or.b32 %r711, %r710, 1; mul.lo.s32 %r712, %r711, %r709; mul.lo.s32 %r713, %r712, 89287766; xor.b32 %r714, %r713, %r712; mul.lo.s32 %r715, %r714, 1403136100; xor.b32 %r716, %r715, %r714; xor.b32 %r717, %r707, -2052961194; brev.b32 %r718, %r716; mul.lo.s32 %r719, %r718, 1025551850; xor.b32 %r720, %r719, %r718; add.s32 %r721, %r720, %r717; shr.u32 %r722, %r717, 16; or.b32 %r723, %r722, 1; mul.lo.s32 %r724, %r721, %r723; mul.lo.s32 %r725, %r724, 89287766; xor.b32 %r726, %r725, %r724; mul.lo.s32 %r727, %r726, 1403136100; xor.b32 %r728, %r727, %r726; brev.b32 %r729, %r728; cvt.rn.f32.u32 %f1028, %r729; mul.ftz.f32 %f3020, %f1028, 0f2F7FFFFE; $L__BB215_45: add.s32 %r42, %r7, 6; @%p38 bra $L__BB215_48; ld.const.v2.u32 {%r730, %r731}, [kernel_params+1600]; mov.b32 %f47, %r731; setp.geu.ftz.f32 %p42, %f47, 0f3F800000; setp.lt.ftz.f32 %p43, %f47, 0f3F800000; ld.const.u32 %r734, [kernel_params+1576]; selp.b32 %r735, %r734, %r6, %p43; add.s16 %rs36, %rs1, 6; and.b16 %rs37, %rs36, 255; cvt.u32.u16 %r736, %rs37; xor.b32 %r737, %r735, %r736; mul.lo.s32 %r738, %r737, -512718531; shr.u32 %r739, %r735, 16; xor.b32 %r740, %r738, %r739; bfe.u32 %r741, %r740, 4, 4; shr.u32 %r742, %r735, 8; xor.b32 %r743, %r740, %r742; xor.b32 %r744, %r743, %r741; mul.lo.s32 %r745, %r744, 153742143; shr.u32 %r746, %r735, 23; xor.b32 %r747, %r745, %r746; bfe.u32 %r748, %r747, 1, 7; xor.b32 %r749, %r748, %r747; shr.u32 %r750, %r735, 27; or.b32 %r751, %r750, 1; mul.lo.s32 %r752, %r751, %r749; mul.lo.s32 %r753, %r752, -1074439621; bfe.u32 %r754, %r753, 2, 6; xor.b32 %r755, %r754, %r753; mul.lo.s32 %r756, %r755, -1638916925; bfe.u32 %r757, %r756, 2, 6; xor.b32 %r758, %r757, %r756; mul.lo.s32 %r759, %r758, -933190689; and.b32 %r760, %r759, 255; bfe.u32 %r761, %r759, 5, 3; xor.b32 %r762, %r761, %r760; add.s32 %r763, %r730, -1; xor.b32 %r43, %r42, 61; xor.b32 %r764, %r735, %r43; mul.lo.s32 %r765, %r764, 9; shr.u32 %r766, %r765, 4; xor.b32 %r767, %r766, %r765; mul.lo.s32 %r768, %r767, 668265261; add.s32 %r769, %r768, %r40; shr.u32 %r770, %r768, 16; or.b32 %r771, %r770, 1; mul.lo.s32 %r772, %r771, %r769; mul.lo.s32 %r773, %r772, 89287766; xor.b32 %r774, %r773, %r772; mul.lo.s32 %r775, %r774, 1403136100; xor.b32 %r776, %r775, %r774; brev.b32 %r777, %r776; neg.s32 %r778, %r730; and.b32 %r779, %r8, %r778; and.b32 %r780, %r777, %r763; or.b32 %r781, %r780, %r779; mad.lo.s32 %r782, %r762, %r730, %r781; shl.b32 %r783, %r730, 8; rem.u32 %r784, %r782, %r783; ld.const.u64 %rd327, [kernel_params+2424]; cvta.to.global.u64 %rd328, %rd327; shl.b32 %r785, %r784, 1; mul.wide.u32 %rd329, %r785, 4; add.s64 %rd330, %rd328, %rd329; ld.global.f32 %f3021, [%rd330]; @%p42 bra $L__BB215_50; xor.b32 %r786, %r43, %r6; mul.lo.s32 %r787, %r786, 9; shr.u32 %r788, %r787, 4; xor.b32 %r789, %r788, %r787; mul.lo.s32 %r790, %r789, 668265261; cvt.rn.f32.u32 %f1029, %r790; mul.ftz.f32 %f1030, %f1029, 0f2F7FFFFE; fma.rn.ftz.f32 %f1031, %f1030, %f47, %f3021; cvt.rmi.ftz.f32.f32 %f1032, %f1031; sub.ftz.f32 %f3021, %f1031, %f1032; bra.uni $L__BB215_50; $L__BB215_48: shr.u32 %r792, %r42, 16; xor.b32 %r793, %r792, %r42; mul.lo.s32 %r794, %r793, 569420461; shr.u32 %r795, %r794, 15; xor.b32 %r796, %r795, %r794; mul.lo.s32 %r797, %r796, -749064809; shr.u32 %r798, %r797, 15; xor.b32 %r799, %r6, %r797; xor.b32 %r800, %r799, %r798; xor.b32 %r801, %r800, 1493655573; add.s32 %r802, %r801, %r40; shr.u32 %r803, %r801, 16; or.b32 %r804, %r803, 1; mul.lo.s32 %r805, %r804, %r802; mul.lo.s32 %r806, %r805, 89287766; xor.b32 %r807, %r806, %r805; mul.lo.s32 %r808, %r807, 1403136100; xor.b32 %r809, %r808, %r807; xor.b32 %r810, %r800, -2052961194; brev.b32 %r811, %r809; mul.lo.s32 %r812, %r811, 1025551850; xor.b32 %r813, %r812, %r811; add.s32 %r814, %r813, %r810; shr.u32 %r815, %r810, 16; or.b32 %r816, %r815, 1; mul.lo.s32 %r817, %r814, %r816; mul.lo.s32 %r818, %r817, 89287766; xor.b32 %r819, %r818, %r817; mul.lo.s32 %r820, %r819, 1403136100; xor.b32 %r821, %r820, %r819; brev.b32 %r822, %r821; cvt.rn.f32.u32 %f1033, %r822; mul.ftz.f32 %f3021, %f1033, 0f2F7FFFFE; mov.u32 %r2164, %r635; bra.uni $L__BB215_50; $L__BB215_65: setp.eq.ftz.f32 %p57, %f70, 0f7F800000; setp.eq.ftz.f32 %p58, %f69, 0f7F800000; and.pred %p59, %p57, %p58; @%p59 bra $L__BB215_67; bra.uni $L__BB215_66; $L__BB215_67: setp.lt.s32 %p63, %r53, 0; selp.b32 %r1015, 1075235812, 1061752795, %p63; or.b32 %r1016, %r1015, %r54; mov.b32 %f3025, %r1016; bra.uni $L__BB215_69; $L__BB215_66: setp.lt.s32 %p60, %r53, 0; min.ftz.f32 %f1065, %f69, %f70; max.ftz.f32 %f1066, %f69, %f70; div.full.ftz.f32 %f1067, %f1065, %f1066; mul.rn.ftz.f32 %f1068, %f1067, %f1067; mov.f32 %f1069, 0fC0B59883; mov.f32 %f1070, 0fBF52C7EA; fma.rn.ftz.f32 %f1071, %f1068, %f1070, %f1069; mov.f32 %f1072, 0fC0D21907; fma.rn.ftz.f32 %f1073, %f1071, %f1068, %f1072; mul.ftz.f32 %f1074, %f1068, %f1073; mul.ftz.f32 %f1075, %f1067, %f1074; add.ftz.f32 %f1076, %f1068, 0f41355DC0; mov.f32 %f1077, 0f41E6BD60; fma.rn.ftz.f32 %f1078, %f1076, %f1068, %f1077; mov.f32 %f1079, 0f419D92C8; fma.rn.ftz.f32 %f1080, %f1078, %f1068, %f1079; rcp.approx.ftz.f32 %f1081, %f1080; fma.rn.ftz.f32 %f1082, %f1075, %f1081, %f1067; mov.f32 %f1083, 0f3FC90FDB; sub.ftz.f32 %f1084, %f1083, %f1082; setp.gt.ftz.f32 %p61, %f69, %f70; selp.f32 %f1085, %f1084, %f1082, %p61; mov.f32 %f1086, 0f40490FDB; sub.ftz.f32 %f1087, %f1086, %f1085; selp.f32 %f1088, %f1087, %f1085, %p60; mov.b32 %r1013, %f1088; or.b32 %r1014, %r54, %r1013; mov.b32 %f1089, %r1014; add.ftz.f32 %f1090, %f70, %f69; setp.le.ftz.f32 %p62, %f1090, 0f7F800000; selp.f32 %f3025, %f1089, %f1090, %p62; $L__BB215_69: sub.ftz.f32 %f1091, %f9, %f67; abs.ftz.f32 %f75, %f1091; setp.eq.ftz.f32 %p64, %f75, 0f00000000; and.pred %p66, %p54, %p64; mov.b32 %r1020, %f1091; and.b32 %r55, %r1020, -2147483648; @%p66 bra $L__BB215_73; bra.uni $L__BB215_70; $L__BB215_73: shr.s32 %r1025, %r53, 31; and.b32 %r1026, %r1025, 1078530011; or.b32 %r1027, %r1026, %r55; mov.b32 %f3026, %r1027; bra.uni $L__BB215_74; $L__BB215_70: setp.eq.ftz.f32 %p67, %f70, 0f7F800000; setp.eq.ftz.f32 %p68, %f75, 0f7F800000; and.pred %p69, %p67, %p68; @%p69 bra $L__BB215_72; bra.uni $L__BB215_71; $L__BB215_72: setp.lt.s32 %p73, %r53, 0; selp.b32 %r1023, 1075235812, 1061752795, %p73; or.b32 %r1024, %r1023, %r55; mov.b32 %f3026, %r1024; bra.uni $L__BB215_74; $L__BB215_71: setp.lt.s32 %p70, %r53, 0; min.ftz.f32 %f1092, %f75, %f70; max.ftz.f32 %f1093, %f75, %f70; div.full.ftz.f32 %f1094, %f1092, %f1093; mul.rn.ftz.f32 %f1095, %f1094, %f1094; mov.f32 %f1096, 0fC0B59883; mov.f32 %f1097, 0fBF52C7EA; fma.rn.ftz.f32 %f1098, %f1095, %f1097, %f1096; mov.f32 %f1099, 0fC0D21907; fma.rn.ftz.f32 %f1100, %f1098, %f1095, %f1099; mul.ftz.f32 %f1101, %f1095, %f1100; mul.ftz.f32 %f1102, %f1094, %f1101; add.ftz.f32 %f1103, %f1095, 0f41355DC0; mov.f32 %f1104, 0f41E6BD60; fma.rn.ftz.f32 %f1105, %f1103, %f1095, %f1104; mov.f32 %f1106, 0f419D92C8; fma.rn.ftz.f32 %f1107, %f1105, %f1095, %f1106; rcp.approx.ftz.f32 %f1108, %f1107; fma.rn.ftz.f32 %f1109, %f1102, %f1108, %f1094; mov.f32 %f1110, 0f3FC90FDB; sub.ftz.f32 %f1111, %f1110, %f1109; setp.gt.ftz.f32 %p71, %f75, %f70; selp.f32 %f1112, %f1111, %f1109, %p71; mov.f32 %f1113, 0f40490FDB; sub.ftz.f32 %f1114, %f1113, %f1112; selp.f32 %f1115, %f1114, %f1112, %p70; mov.b32 %r1021, %f1115; or.b32 %r1022, %r55, %r1021; mov.b32 %f1116, %r1022; add.ftz.f32 %f1117, %f70, %f75; setp.le.ftz.f32 %p72, %f1117, 0f7F800000; selp.f32 %f3026, %f1116, %f1117, %p72; $L__BB215_74: mov.f32 %f1120, 0f3F800000; sub.ftz.f32 %f1121, %f1120, %f3166; mul.ftz.f32 %f1122, %f3166, %f3026; fma.rn.ftz.f32 %f80, %f1121, %f3025, %f1122; setp.eq.ftz.f32 %p74, %f3026, %f3025; mov.f32 %f3028, %f3029; @%p74 bra $L__BB215_76; cos.approx.ftz.f32 %f1123, %f80; sin.approx.ftz.f32 %f1124, %f80; div.approx.ftz.f32 %f1125, %f1124, %f1123; mul.ftz.f32 %f1126, %f68, %f1125; sub.ftz.f32 %f1127, %f3026, %f3025; mul.ftz.f32 %f1128, %f1126, %f1126; fma.rn.ftz.f32 %f1129, %f68, %f68, %f1128; mul.ftz.f32 %f1130, %f1127, %f1129; div.approx.ftz.f32 %f3029, %f68, %f1130; add.ftz.f32 %f1131, %f67, %f1126; max.ftz.f32 %f1132, %f1131, %f7; min.ftz.f32 %f3028, %f1132, %f9; $L__BB215_76: st.local.f32 [%rd49+12], %f3028; $L__BB215_77: ld.global.u32 %r1028, [%rd30]; and.b32 %r56, %r1028, 16777216; setp.lt.s32 %p75, %r2165, 1; mov.f32 %f3168, %f3167; mov.f32 %f3169, %f3167; mov.f32 %f3181, %f3167; mov.f32 %f3182, %f3167; mov.f32 %f3183, %f3167; @%p75 bra $L__BB215_503; ld.const.u32 %r57, [kernel_params+4]; ld.const.v2.u64 {%rd340, %rd341}, [kernel_params+2400]; ld.const.v2.u64 {%rd342, %rd343}, [kernel_params+2176]; ld.const.u64 %rd344, [kernel_params+2160]; cvta.to.global.u64 %rd54, %rd344; ld.const.u32 %r58, [kernel_params+972]; ld.const.f32 %f1147, [kernel_params+256]; mov.f32 %f3158, 0f3F800000; sub.ftz.f32 %f86, %f3158, %f1147; ld.const.u64 %rd345, [kernel_params+2312]; cvta.to.global.u64 %rd55, %rd345; ld.const.u64 %rd346, [kernel_params+2240]; cvta.to.global.u64 %rd56, %rd346; ld.const.u64 %rd347, [kernel_params+2344]; cvta.to.global.u64 %rd57, %rd347; ld.const.u64 %rd58, [kernel_params+2432]; mul.ftz.f32 %f87, %f3029, %f3029; sub.ftz.f32 %f1148, %f34, %f1; sub.ftz.f32 %f1149, %f35, %f2; mul.ftz.f32 %f1150, %f5, %f1149; fma.rn.ftz.f32 %f1151, %f4, %f1148, %f1150; sub.ftz.f32 %f1152, %f36, %f3; fma.rn.ftz.f32 %f88, %f6, %f1152, %f1151; mul.ftz.f32 %f1153, %f1149, %f1149; fma.rn.ftz.f32 %f1154, %f1148, %f1148, %f1153; fma.rn.ftz.f32 %f1155, %f1152, %f1152, %f1154; mul.ftz.f32 %f1156, %f88, %f88; sub.ftz.f32 %f89, %f1155, %f1156; sub.ftz.f32 %f90, %f9, %f88; add.s64 %rd59, %rd23, 240; mov.f32 %f3183, 0f00000000; mov.u32 %r1029, 0; mov.pred %p286, -1; mov.u16 %rs52, 0; setp.eq.s32 %p320, %r56, 0; setp.gt.ftz.f32 %p327, %f3029, 0f322BCC77; mov.u16 %rs46, 1; cvta.to.global.u64 %rd358, %rd341; cvta.to.global.u64 %rd79, %rd340; cvta.to.global.u64 %rd362, %rd343; cvta.to.global.u64 %rd370, %rd342; mov.u32 %r2167, %r1029; mov.f32 %f3182, %f3183; mov.f32 %f3181, %f3183; mov.f32 %f3169, %f3183; mov.f32 %f3168, %f3183; mov.f32 %f3167, %f3183; mov.f32 %f3045, %f7; $L__BB215_79: cvt.rn.f32.s32 %f1157, %r2167; add.ftz.f32 %f1158, %f3021, %f1157; fma.rn.ftz.f32 %f1159, %f3019, %f1158, %f7; min.ftz.f32 %f109, %f9, %f1159; sub.ftz.f32 %f3141, %f109, %f3045; fma.rn.ftz.f32 %f1160, %f3020, %f3141, %f3045; fma.rn.ftz.f32 %f1161, %f6, %f1160, %f3; fma.rn.ftz.f32 %f1162, %f5, %f1160, %f2; fma.rn.ftz.f32 %f1163, %f4, %f1160, %f1; st.local.v2.f32 [%rd1], {%f1163, %f1162}; st.local.f32 [%rd23+-8], %f1161; ld.global.u32 %r60, [%rd30]; and.b32 %r1031, %r60, 1016576; setp.eq.s32 %p76, %r1031, 0; selp.b32 %r1032, %r57, 0, %p76; st.local.u32 [%rd23+244], %r1029; st.local.u32 [%rd23+248], %r1032; st.local.u32 [%rd23+36], %r1029; st.local.u32 [%rd23+40], %r1029; ld.global.u32 %r2168, [%rd295]; setp.eq.s32 %p77, %r2168, -1; mov.u32 %r2192, %r1029; @%p77 bra $L__BB215_400; mov.u64 %rd353, kernel_params; add.s64 %rd1744, %rd353, 2768; mov.u32 %r2169, 0; bra.uni $L__BB215_81; $L__BB215_123: setp.eq.s32 %p114, %r98, 3; @%p114 bra $L__BB215_125; bra.uni $L__BB215_124; $L__BB215_125: { // callseq 1368, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f3065, [retval0+0]; } // callseq 1368 mov.f32 %f3068, 0f3F800000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; bra.uni $L__BB215_130; $L__BB215_155: setp.eq.s32 %p135, %r107, 3; @%p135 bra $L__BB215_157; bra.uni $L__BB215_156; $L__BB215_157: { // callseq 1374, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .b32 retval0; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicIfET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f3081, [retval0+0]; } // callseq 1374 mov.f32 %f3084, 0f3F800000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; bra.uni $L__BB215_162; $L__BB215_124: ld.global.u64 %rd388, [%rd75+-32]; tex.3d.v4.f32.f32 {%f3065, %f1524, %f1525, %f1526}, [%rd388, {%f188, %f189, %f190, %f190}]; mov.f32 %f3068, 0f3F800000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; bra.uni $L__BB215_130; $L__BB215_156: ld.global.u64 %rd403, [%rd77+-32]; tex.3d.v4.f32.f32 {%f3081, %f1705, %f1706, %f1707}, [%rd403, {%f271, %f272, %f273, %f273}]; mov.f32 %f3084, 0f3F800000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; bra.uni $L__BB215_162; $L__BB215_81: ld.const.u64 %rd354, [%rd1744]; cvta.to.global.u64 %rd355, %rd354; add.s64 %rd357, %rd355, %rd260; ld.global.u32 %r64, [%rd357]; st.local.u32 [%rd23+60], %r64; mov.u32 %r1034, -1; st.local.u32 [%rd23+64], %r1034; ld.local.u32 %r1035, [%rd23+36]; and.b32 %r1036, %r1035, -2147418113; and.b32 %r2181, %r2168, 4194303; shl.b32 %r1037, %r2168, 5; cvt.u64.u32 %rd359, %r1037; and.b64 %rd360, %rd359, 134217696; add.s64 %rd361, %rd358, %rd360; ld.global.u32 %r66, [%rd361+16]; or.b32 %r1038, %r1036, %r66; st.local.v2.u32 [%rd23+32], {%r2168, %r1038}; ld.local.u32 %r1039, [%rd23+40]; and.b32 %r67, %r1039, -4032; st.local.u32 [%rd23+40], %r67; setp.eq.s32 %p78, %r64, -1; @%p78 bra $L__BB215_168; mul.wide.s32 %rd363, %r64, 4; add.s64 %rd364, %rd362, %rd363; ld.global.u32 %r68, [%rd364]; or.b32 %r1040, %r67, %r68; st.local.u32 [%rd23+40], %r1040; and.b32 %r69, %r68, 2; setp.eq.s32 %p79, %r69, 0; mul.wide.s32 %rd365, %r64, 208; add.s64 %rd366, %rd54, %rd365; add.s64 %rd61, %rd366, 48; @%p79 bra $L__BB215_89; ld.global.u32 %rd367, [%rd61+120]; ld.global.u32 %r1041, [%rd61+104]; shl.b32 %r1042, %r1041, 1; cvt.rn.f32.s32 %f1164, %r1042; ld.local.f32 %f1165, [%rd23+68]; mul.ftz.f32 %f1166, %f1165, %f1164; cvt.rzi.ftz.s32.f32 %r1043, %f1166; add.s32 %r1044, %r1042, -1; min.s32 %r1045, %r1043, %r1044; cvt.rn.f32.s32 %f1167, %r1045; sub.ftz.f32 %f111, %f1166, %f1167; cvt.s64.s32 %rd368, %r1045; add.s64 %rd369, %rd368, %rd367; shl.b64 %rd371, %rd369, 6; add.s64 %rd62, %rd370, %rd371; ld.global.v4.f32 {%f1168, %f1169, %f1170, %f1171}, [%rd62]; ld.global.v4.f32 {%f1172, %f1173, %f1174, %f1175}, [%rd62+64]; mul.ftz.f32 %f1176, %f1169, %f1173; fma.rn.ftz.f32 %f1177, %f1168, %f1172, %f1176; mul.ftz.f32 %f1178, %f1171, %f1175; fma.rn.ftz.f32 %f1179, %f1170, %f1174, %f1178; add.ftz.f32 %f120, %f1177, %f1179; setp.gt.ftz.f32 %p80, %f120, 0f3F7FDF3B; @%p80 bra $L__BB215_85; bra.uni $L__BB215_84; $L__BB215_85: mov.f32 %f3008, 0f3F800000; sub.ftz.f32 %f3007, %f3008, %f111; mul.ftz.f32 %f1234, %f1168, %f3007; fma.rn.ftz.f32 %f1235, %f1172, %f111, %f1234; mul.ftz.f32 %f1236, %f1169, %f3007; fma.rn.ftz.f32 %f1237, %f1173, %f111, %f1236; mul.ftz.f32 %f1238, %f1170, %f3007; fma.rn.ftz.f32 %f1239, %f1174, %f111, %f1238; mul.ftz.f32 %f1240, %f1171, %f3007; fma.rn.ftz.f32 %f1241, %f1175, %f111, %f1240; mul.ftz.f32 %f1242, %f1237, %f1237; fma.rn.ftz.f32 %f1243, %f1235, %f1235, %f1242; mul.ftz.f32 %f1244, %f1241, %f1241; fma.rn.ftz.f32 %f1245, %f1239, %f1239, %f1244; add.ftz.f32 %f1246, %f1243, %f1245; rsqrt.approx.ftz.f32 %f1247, %f1246; mul.ftz.f32 %f3051, %f1241, %f1247; mul.ftz.f32 %f3050, %f1239, %f1247; mul.ftz.f32 %f3049, %f1237, %f1247; mul.ftz.f32 %f3048, %f1235, %f1247; bra.uni $L__BB215_86; $L__BB215_84: mov.f32 %f3004, 0f3F800000; mov.f32 %f1181, 0fBF800000; max.ftz.f32 %f1182, %f120, %f1181; min.ftz.f32 %f1184, %f1182, %f3004; abs.ftz.f32 %f1185, %f1184; sub.ftz.f32 %f1186, %f3004, %f1185; mul.ftz.f32 %f1187, %f1186, 0f3F000000; sqrt.approx.ftz.f32 %f1188, %f1187; setp.gt.ftz.f32 %p81, %f1185, 0f3F11EB85; selp.f32 %f1189, %f1188, %f1185, %p81; mul.ftz.f32 %f1190, %f1189, %f1189; mov.f32 %f1191, 0f3C94D2E9; mov.f32 %f1192, 0f3D53F941; fma.rn.ftz.f32 %f1193, %f1192, %f1190, %f1191; mov.f32 %f1194, 0f3D3F841F; fma.rn.ftz.f32 %f1195, %f1193, %f1190, %f1194; mov.f32 %f1196, 0f3D994929; fma.rn.ftz.f32 %f1197, %f1195, %f1190, %f1196; mov.f32 %f1198, 0f3E2AAB94; fma.rn.ftz.f32 %f1199, %f1197, %f1190, %f1198; mul.ftz.f32 %f1200, %f1190, %f1199; fma.rn.ftz.f32 %f1201, %f1200, %f1189, %f1189; add.ftz.f32 %f1202, %f1201, %f1201; mov.f32 %f1203, 0f3FC90FDB; sub.ftz.f32 %f1204, %f1203, %f1201; selp.f32 %f1205, %f1202, %f1204, %p81; setp.lt.ftz.f32 %p82, %f1184, 0f00000000; mov.f32 %f1206, 0f40490FDB; sub.ftz.f32 %f1207, %f1206, %f1205; selp.f32 %f1208, %f1207, %f1205, %p82; mul.ftz.f32 %f1209, %f1168, %f120; sub.ftz.f32 %f1210, %f1172, %f1209; mul.ftz.f32 %f1211, %f1169, %f120; sub.ftz.f32 %f1212, %f1173, %f1211; mul.ftz.f32 %f1213, %f1170, %f120; sub.ftz.f32 %f1214, %f1174, %f1213; mul.ftz.f32 %f1215, %f1171, %f120; sub.ftz.f32 %f1216, %f1175, %f1215; mul.ftz.f32 %f1217, %f1212, %f1212; fma.rn.ftz.f32 %f1218, %f1210, %f1210, %f1217; mul.ftz.f32 %f1219, %f1216, %f1216; fma.rn.ftz.f32 %f1220, %f1214, %f1214, %f1219; add.ftz.f32 %f1221, %f1218, %f1220; rsqrt.approx.ftz.f32 %f1222, %f1221; mul.ftz.f32 %f1223, %f1210, %f1222; mul.ftz.f32 %f1224, %f1212, %f1222; mul.ftz.f32 %f1225, %f1214, %f1222; mul.ftz.f32 %f1226, %f1216, %f1222; mul.ftz.f32 %f1227, %f111, %f1208; cos.approx.ftz.f32 %f1228, %f1227; sin.approx.ftz.f32 %f1229, %f1227; mul.ftz.f32 %f1230, %f1223, %f1229; mul.ftz.f32 %f1231, %f1224, %f1229; mul.ftz.f32 %f1232, %f1225, %f1229; mul.ftz.f32 %f1233, %f1226, %f1229; fma.rn.ftz.f32 %f3051, %f1171, %f1228, %f1233; fma.rn.ftz.f32 %f3050, %f1170, %f1228, %f1232; fma.rn.ftz.f32 %f3049, %f1169, %f1228, %f1231; fma.rn.ftz.f32 %f3048, %f1168, %f1228, %f1230; $L__BB215_86: mov.f32 %f3006, 0f3F800000; sub.ftz.f32 %f3005, %f3006, %f111; ld.global.v4.f32 {%f1248, %f1249, %f1250, %f1251}, [%rd62+16]; ld.global.v4.f32 {%f1256, %f1257, %f1258, %f1259}, [%rd62+80]; mul.ftz.f32 %f1264, %f111, %f1256; mul.ftz.f32 %f1265, %f111, %f1257; mul.ftz.f32 %f1266, %f111, %f1258; mul.ftz.f32 %f1267, %f111, %f1259; fma.rn.ftz.f32 %f134, %f3005, %f1248, %f1264; fma.rn.ftz.f32 %f135, %f3005, %f1249, %f1265; fma.rn.ftz.f32 %f136, %f3005, %f1250, %f1266; fma.rn.ftz.f32 %f1268, %f3005, %f1251, %f1267; ld.global.v4.f32 {%f1269, %f1270, %f1271, %f1272}, [%rd62+32]; ld.global.v4.f32 {%f1277, %f1278, %f1279, %f1280}, [%rd62+96]; mul.ftz.f32 %f1285, %f111, %f1277; mul.ftz.f32 %f1286, %f111, %f1278; mul.ftz.f32 %f1287, %f111, %f1279; mul.ftz.f32 %f1288, %f111, %f1280; fma.rn.ftz.f32 %f1289, %f3005, %f1269, %f1285; fma.rn.ftz.f32 %f1290, %f3005, %f1270, %f1286; fma.rn.ftz.f32 %f1291, %f3005, %f1271, %f1287; fma.rn.ftz.f32 %f1292, %f3005, %f1272, %f1288; ld.global.v4.f32 {%f1293, %f1294, %f1295, %f1296}, [%rd62+48]; ld.global.v4.f32 {%f1301, %f1302, %f1303, %f1304}, [%rd62+112]; mul.ftz.f32 %f1309, %f111, %f1301; mul.ftz.f32 %f1310, %f111, %f1302; mul.ftz.f32 %f1311, %f111, %f1303; mul.ftz.f32 %f1312, %f111, %f1304; fma.rn.ftz.f32 %f1313, %f3005, %f1293, %f1309; fma.rn.ftz.f32 %f1314, %f3005, %f1294, %f1310; fma.rn.ftz.f32 %f1315, %f3005, %f1295, %f1311; fma.rn.ftz.f32 %f1316, %f3005, %f1296, %f1312; mul.ftz.f32 %f1317, %f3051, 0f3FB504F3; mul.ftz.f32 %f1318, %f3048, 0f3FB504F3; mul.ftz.f32 %f1319, %f3049, 0f3FB504F3; mul.ftz.f32 %f1320, %f3050, 0f3FB504F3; mul.ftz.f32 %f1321, %f1317, %f1318; mul.ftz.f32 %f1322, %f1317, %f1319; mul.ftz.f32 %f1323, %f1317, %f1320; mul.ftz.f32 %f1324, %f1318, %f1318; mul.ftz.f32 %f1325, %f1318, %f1319; mul.ftz.f32 %f1326, %f1318, %f1320; mul.ftz.f32 %f1327, %f1319, %f1319; mul.ftz.f32 %f1328, %f1319, %f1320; mul.ftz.f32 %f1329, %f1320, %f1320; sub.ftz.f32 %f1331, %f3006, %f1327; sub.ftz.f32 %f1332, %f1331, %f1329; sub.ftz.f32 %f1333, %f1325, %f1323; add.ftz.f32 %f1334, %f1322, %f1326; add.ftz.f32 %f1335, %f1323, %f1325; sub.ftz.f32 %f1336, %f3006, %f1324; sub.ftz.f32 %f1337, %f1336, %f1329; sub.ftz.f32 %f1338, %f1328, %f1321; sub.ftz.f32 %f1339, %f1326, %f1322; add.ftz.f32 %f1340, %f1321, %f1328; sub.ftz.f32 %f1341, %f1336, %f1327; mul.ftz.f32 %f1342, %f1333, %f1291; fma.rn.ftz.f32 %f1343, %f1332, %f1268, %f1342; fma.rn.ftz.f32 %f3057, %f1334, %f1314, %f1343; mul.ftz.f32 %f1344, %f1333, %f1292; fma.rn.ftz.f32 %f1345, %f1332, %f1289, %f1344; fma.rn.ftz.f32 %f138, %f1334, %f1315, %f1345; mul.ftz.f32 %f1346, %f1333, %f1313; fma.rn.ftz.f32 %f1347, %f1332, %f1290, %f1346; fma.rn.ftz.f32 %f139, %f1334, %f1316, %f1347; mul.ftz.f32 %f1348, %f1337, %f1291; fma.rn.ftz.f32 %f1349, %f1335, %f1268, %f1348; fma.rn.ftz.f32 %f140, %f1338, %f1314, %f1349; mul.ftz.f32 %f1350, %f1337, %f1292; fma.rn.ftz.f32 %f1351, %f1335, %f1289, %f1350; fma.rn.ftz.f32 %f3056, %f1338, %f1315, %f1351; mul.ftz.f32 %f1352, %f1337, %f1313; fma.rn.ftz.f32 %f1353, %f1335, %f1290, %f1352; fma.rn.ftz.f32 %f142, %f1338, %f1316, %f1353; mul.ftz.f32 %f1354, %f1340, %f1291; fma.rn.ftz.f32 %f1355, %f1339, %f1268, %f1354; fma.rn.ftz.f32 %f143, %f1341, %f1314, %f1355; mul.ftz.f32 %f1356, %f1340, %f1292; fma.rn.ftz.f32 %f1357, %f1339, %f1289, %f1356; fma.rn.ftz.f32 %f144, %f1341, %f1315, %f1357; mul.ftz.f32 %f1358, %f1340, %f1313; fma.rn.ftz.f32 %f1359, %f1339, %f1290, %f1358; fma.rn.ftz.f32 %f3055, %f1341, %f1316, %f1359; st.local.v4.f32 [%rd23+128], {%f3057, %f138, %f139, %f134}; st.local.v4.f32 [%rd23+144], {%f140, %f3056, %f142, %f135}; st.local.v4.f32 [%rd23+160], {%f143, %f144, %f3055, %f136}; mul.ftz.f32 %f1360, %f3056, %f3055; mul.ftz.f32 %f146, %f144, %f142; sub.ftz.f32 %f3054, %f1360, %f146; mul.ftz.f32 %f148, %f144, %f139; mul.ftz.f32 %f1361, %f138, %f3055; sub.ftz.f32 %f3053, %f148, %f1361; mul.ftz.f32 %f150, %f138, %f142; mul.ftz.f32 %f1362, %f3056, %f139; sub.ftz.f32 %f3052, %f150, %f1362; mul.ftz.f32 %f1363, %f140, %f3053; fma.rn.ftz.f32 %f1364, %f3057, %f3054, %f1363; fma.rn.ftz.f32 %f3058, %f143, %f3052, %f1364; setp.neu.ftz.f32 %p83, %f3058, 0f00000000; @%p83 bra $L__BB215_88; add.ftz.f32 %f3057, %f3057, 0f322BCC77; add.ftz.f32 %f3055, %f3055, 0f322BCC77; add.ftz.f32 %f3056, %f3056, 0f322BCC77; mul.ftz.f32 %f1365, %f3056, %f3055; sub.ftz.f32 %f3054, %f1365, %f146; mul.ftz.f32 %f1366, %f138, %f3055; sub.ftz.f32 %f3053, %f148, %f1366; mul.ftz.f32 %f1367, %f3056, %f139; sub.ftz.f32 %f3052, %f150, %f1367; mul.ftz.f32 %f1368, %f140, %f3053; fma.rn.ftz.f32 %f1369, %f3057, %f3054, %f1368; fma.rn.ftz.f32 %f1370, %f143, %f3052, %f1369; setp.eq.ftz.f32 %p84, %f1370, 0f00000000; selp.f32 %f3058, 0f7F7FFFFF, %f1370, %p84; $L__BB215_88: mul.ftz.f32 %f1371, %f140, %f3055; mul.ftz.f32 %f1372, %f143, %f142; sub.ftz.f32 %f1373, %f1372, %f1371; mul.ftz.f32 %f1374, %f143, %f139; mul.ftz.f32 %f1375, %f3055, %f3057; sub.ftz.f32 %f1376, %f1375, %f1374; mul.ftz.f32 %f1377, %f142, %f3057; mul.ftz.f32 %f1378, %f140, %f139; sub.ftz.f32 %f1379, %f1378, %f1377; mul.ftz.f32 %f1380, %f143, %f3056; mul.ftz.f32 %f1381, %f140, %f144; sub.ftz.f32 %f1382, %f1381, %f1380; mul.ftz.f32 %f1383, %f144, %f3057; mul.ftz.f32 %f1384, %f143, %f138; sub.ftz.f32 %f1385, %f1384, %f1383; mul.ftz.f32 %f1386, %f140, %f138; mul.ftz.f32 %f1387, %f3056, %f3057; sub.ftz.f32 %f1388, %f1387, %f1386; div.approx.ftz.f32 %f1389, %f3053, %f3058; mul.ftz.f32 %f1390, %f135, %f1389; div.approx.ftz.f32 %f1391, %f3054, %f3058; fma.rn.ftz.f32 %f1392, %f134, %f1391, %f1390; div.approx.ftz.f32 %f1393, %f3052, %f3058; fma.rn.ftz.f32 %f1394, %f136, %f1393, %f1392; div.approx.ftz.f32 %f1395, %f1376, %f3058; mul.ftz.f32 %f1396, %f135, %f1395; div.approx.ftz.f32 %f1397, %f1373, %f3058; fma.rn.ftz.f32 %f1398, %f134, %f1397, %f1396; div.approx.ftz.f32 %f1399, %f1379, %f3058; fma.rn.ftz.f32 %f1400, %f136, %f1399, %f1398; div.approx.ftz.f32 %f1401, %f1385, %f3058; mul.ftz.f32 %f1402, %f135, %f1401; div.approx.ftz.f32 %f1403, %f1382, %f3058; fma.rn.ftz.f32 %f1404, %f134, %f1403, %f1402; div.approx.ftz.f32 %f1405, %f1388, %f3058; fma.rn.ftz.f32 %f1406, %f136, %f1405, %f1404; neg.ftz.f32 %f1407, %f1394; st.local.v4.f32 [%rd23+176], {%f1391, %f1389, %f1393, %f1407}; neg.ftz.f32 %f1408, %f1400; st.local.v4.f32 [%rd23+192], {%f1397, %f1395, %f1399, %f1408}; neg.ftz.f32 %f1409, %f1406; st.local.v4.f32 [%rd23+208], {%f1403, %f1401, %f1405, %f1409}; $L__BB215_89: and.b32 %r1046, %r68, 2048; setp.eq.s32 %p85, %r1046, 0; @%p85 bra $L__BB215_168; ld.local.s32 %rd63, [%rd23+44]; ld.global.u32 %r70, [%rd61+116]; ld.local.u32 %r71, [%rd23+48]; and.b32 %r1048, %r71, 1; setp.eq.b32 %p86, %r1048, 1; mov.pred %p87, 0; xor.pred %p88, %p86, %p87; not.pred %p89, %p88; mov.u32 %r2171, 0; @%p89 bra $L__BB215_94; mov.u32 %r2170, -1; cvt.u32.u64 %r1050, %rd63; setp.eq.s32 %p90, %r1050, -1; @%p90 bra $L__BB215_93; shl.b64 %rd372, %rd63, 2; add.s64 %rd373, %rd56, %rd372; ld.global.u32 %r2170, [%rd373]; $L__BB215_93: setp.ne.s32 %p91, %r2170, -1; selp.u32 %r2171, 1, 0, %p91; $L__BB215_94: add.s32 %r2173, %r2171, %r70; mul.wide.u32 %rd374, %r2173, 16; add.s64 %rd375, %rd55, %rd374; ld.global.u16 %rs111, [%rd375+12]; ld.global.u32 %r2172, [%rd375+8]; ld.global.u64 %rd1745, [%rd375]; setp.eq.s64 %p92, %rd1745, 25; @%p92 bra $L__BB215_99; $L__BB215_95: setp.eq.s64 %p93, %rd1745, 0; @%p93 bra $L__BB215_97; add.s32 %r2173, %r2173, 2; bra.uni $L__BB215_98; $L__BB215_97: setp.eq.s16 %p94, %rs111, 0; mov.u32 %r2177, -1; mov.u32 %r2176, 0; mov.u32 %r2173, %r2172; @%p94 bra $L__BB215_103; $L__BB215_98: mul.wide.u32 %rd376, %r2173, 16; add.s64 %rd377, %rd55, %rd376; ld.global.u16 %rs111, [%rd377+12]; ld.global.u32 %r2172, [%rd377+8]; ld.global.u64 %rd1745, [%rd377]; setp.ne.s64 %p95, %rd1745, 25; @%p95 bra $L__BB215_95; $L__BB215_99: cvt.u32.u64 %r1055, %rd63; setp.ne.s32 %p96, %r1055, -1; @%p96 bra $L__BB215_102; mov.u32 %r2177, -1; mov.u32 %r2176, 0; add.s16 %rs42, %rs111, -1; setp.lt.u16 %p97, %rs42, 2; @%p97 bra $L__BB215_102; setp.ne.s16 %p98, %rs111, 1024; @%p98 bra $L__BB215_103; $L__BB215_102: setp.eq.s16 %p99, %rs111, 0; selp.b32 %r2177, -1, %r2172, %p99; cvt.u32.u16 %r2176, %rs111; $L__BB215_103: ld.local.v4.f32 {%f1413, %f1414, %f1415, %f1416}, [%rd1]; ld.global.f32 %f170, [%rd61+156]; ld.local.f32 %f171, [%rd23+68]; setp.ne.s32 %p100, %r71, 16; mov.f32 %f3065, 0f00000000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; @%p100 bra $L__BB215_132; and.b32 %r1062, %r2176, 3; setp.eq.s32 %p101, %r1062, 0; @%p101 bra $L__BB215_106; mul.wide.s32 %rd382, %r2177, 16; add.s64 %rd383, %rd57, %rd382; ld.global.v4.f32 {%f3065, %f3066, %f3067, %f3068}, [%rd383]; bra.uni $L__BB215_130; $L__BB215_106: setp.ne.s32 %p102, %r2176, 1024; mov.f32 %f3065, 0f00000000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; mov.f32 %f3068, %f3065; @%p102 bra $L__BB215_130; and.b32 %r2147, %r68, 2; setp.eq.s32 %p797, %r2147, 0; @%p797 bra $L__BB215_109; bra.uni $L__BB215_108; $L__BB215_109: ld.global.v4.f32 {%f1458, %f1459, %f1460, %f1461}, [%rd61]; ld.global.v4.f32 {%f1466, %f1467, %f1468, %f1469}, [%rd61+16]; ld.global.v4.f32 {%f1474, %f1475, %f1476, %f1477}, [%rd61+32]; mul.ftz.f32 %f1482, %f1414, %f1459; fma.rn.ftz.f32 %f1483, %f1413, %f1458, %f1482; fma.rn.ftz.f32 %f1484, %f1415, %f1460, %f1483; add.ftz.f32 %f188, %f1461, %f1484; mul.ftz.f32 %f1485, %f1414, %f1467; fma.rn.ftz.f32 %f1486, %f1413, %f1466, %f1485; fma.rn.ftz.f32 %f1487, %f1415, %f1468, %f1486; add.ftz.f32 %f189, %f1469, %f1487; mul.ftz.f32 %f1488, %f1414, %f1475; fma.rn.ftz.f32 %f1489, %f1413, %f1474, %f1488; fma.rn.ftz.f32 %f1490, %f1415, %f1476, %f1489; add.ftz.f32 %f190, %f1477, %f1490; bra.uni $L__BB215_110; $L__BB215_108: ld.local.v4.f32 {%f1425, %f1426, %f1427, %f1428}, [%rd23+176]; mul.ftz.f32 %f1433, %f1414, %f1426; fma.rn.ftz.f32 %f1434, %f1413, %f1425, %f1433; fma.rn.ftz.f32 %f1435, %f1415, %f1427, %f1434; add.ftz.f32 %f188, %f1428, %f1435; ld.local.v4.f32 {%f1436, %f1437, %f1438, %f1439}, [%rd23+192]; mul.ftz.f32 %f1444, %f1414, %f1437; fma.rn.ftz.f32 %f1445, %f1413, %f1436, %f1444; fma.rn.ftz.f32 %f1446, %f1415, %f1438, %f1445; add.ftz.f32 %f189, %f1439, %f1446; ld.local.v4.f32 {%f1447, %f1448, %f1449, %f1450}, [%rd23+208]; mul.ftz.f32 %f1455, %f1414, %f1448; fma.rn.ftz.f32 %f1456, %f1413, %f1447, %f1455; fma.rn.ftz.f32 %f1457, %f1415, %f1449, %f1456; add.ftz.f32 %f190, %f1450, %f1457; $L__BB215_110: mul.wide.s32 %rd384, %r2177, 96; add.s64 %rd74, %rd58, %rd384; cvta.to.global.u64 %rd385, %rd58; add.s64 %rd386, %rd385, %rd384; add.s64 %rd75, %rd386, 32; ld.global.u32 %r1063, [%rd386+32]; setp.eq.s32 %p104, %r1063, 0; @%p104 bra $L__BB215_112; ld.global.v4.f32 {%f1491, %f1492, %f1493, %f1494}, [%rd75+16]; mul.ftz.f32 %f1499, %f189, %f1492; fma.rn.ftz.f32 %f1500, %f188, %f1491, %f1499; fma.rn.ftz.f32 %f1501, %f190, %f1493, %f1500; add.ftz.f32 %f185, %f1494, %f1501; ld.global.v4.f32 {%f1502, %f1503, %f1504, %f1505}, [%rd75+32]; mul.ftz.f32 %f1510, %f189, %f1503; fma.rn.ftz.f32 %f1511, %f188, %f1502, %f1510; fma.rn.ftz.f32 %f1512, %f190, %f1504, %f1511; add.ftz.f32 %f186, %f1505, %f1512; ld.global.v4.f32 {%f1513, %f1514, %f1515, %f1516}, [%rd75+48]; mul.ftz.f32 %f1521, %f189, %f1514; fma.rn.ftz.f32 %f1522, %f188, %f1513, %f1521; fma.rn.ftz.f32 %f1523, %f190, %f1515, %f1522; add.ftz.f32 %f190, %f1516, %f1523; mov.f32 %f188, %f185; mov.f32 %f189, %f186; $L__BB215_112: and.b32 %r1064, %r66, 16777216; setp.ne.s32 %p105, %r1064, 0; setp.eq.s32 %p106, %r1064, 0; selp.b32 %r96, -1, 2, %p106; @%p105 bra $L__BB215_114; ld.global.u32 %r96, [%rd75+-20]; $L__BB215_114: ld.global.u32 %r97, [%rd75+-24]; setp.eq.s32 %p107, %r97, 8; @%p107 bra $L__BB215_129; bra.uni $L__BB215_115; $L__BB215_129: { // callseq 1372, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .b32 param4; st.param.b32 [param4+0], %r96; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3065, [retval0+0]; } // callseq 1372 mov.f32 %f3068, 0f3F800000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; bra.uni $L__BB215_130; $L__BB215_115: setp.eq.s32 %p108, %r97, 9; @%p108 bra $L__BB215_128; bra.uni $L__BB215_116; $L__BB215_128: { // callseq 1371, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .b32 param4; st.param.b32 [param4+0], %r96; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3065, [retval0+0]; ld.param.f32 %f3066, [retval0+4]; ld.param.f32 %f3067, [retval0+8]; } // callseq 1371 mov.f32 %f3068, 0f3F800000; bra.uni $L__BB215_130; $L__BB215_116: setp.eq.s32 %p109, %r97, 10; @%p109 bra $L__BB215_127; bra.uni $L__BB215_117; $L__BB215_127: { // callseq 1370, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .b32 param4; st.param.b32 [param4+0], %r96; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3065, [retval0+0]; } // callseq 1370 mov.f32 %f3068, 0f3F800000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; bra.uni $L__BB215_130; $L__BB215_117: setp.eq.s32 %p110, %r97, 11; @%p110 bra $L__BB215_126; bra.uni $L__BB215_118; $L__BB215_126: { // callseq 1369, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .b32 param4; st.param.b32 [param4+0], %r96; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3065, [retval0+0]; } // callseq 1369 mov.f32 %f3068, 0f3F800000; mov.f32 %f3066, %f3065; mov.f32 %f3067, %f3065; bra.uni $L__BB215_130; $L__BB215_118: or.b32 %r98, %r96, 1; setp.gt.u32 %p111, %r97, 6; @%p111 bra $L__BB215_123; mov.u32 %r1066, 1; shl.b32 %r1067, %r1066, %r97; and.b32 %r1068, %r1067, 71; setp.ne.s32 %p112, %r1068, 0; @%p112 bra $L__BB215_120; bra.uni $L__BB215_123; $L__BB215_120: setp.eq.s32 %p113, %r98, 3; @%p113 bra $L__BB215_122; bra.uni $L__BB215_121; $L__BB215_122: { // callseq 1367, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd74; .param .b32 param1; st.param.f32 [param1+0], %f188; .param .b32 param2; st.param.f32 [param2+0], %f189; .param .b32 param3; st.param.f32 [param3+0], %f190; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f3065, [retval0+0]; ld.param.f32 %f3066, [retval0+4]; ld.param.f32 %f3067, [retval0+8]; ld.param.f32 %f3068, [retval0+12]; } // callseq 1367 bra.uni $L__BB215_130; $L__BB215_121: ld.global.u64 %rd387, [%rd75+-32]; tex.3d.v4.f32.f32 {%f3065, %f3066, %f3067, %f3068}, [%rd387, {%f188, %f189, %f190, %f190}]; $L__BB215_130: setp.leu.ftz.f32 %p115, %f3068, 0f358637BD; setp.eq.ftz.f32 %p116, %f3068, 0f3F800000; or.pred %p117, %p115, %p116; @%p117 bra $L__BB215_132; div.approx.ftz.f32 %f3067, %f3067, %f3068; div.approx.ftz.f32 %f3066, %f3066, %f3068; div.approx.ftz.f32 %f3065, %f3065, %f3068; $L__BB215_132: ld.local.u32 %r1069, [%rd23+40]; and.b32 %r99, %r1069, 2; setp.eq.s32 %p118, %r99, 0; @%p118 bra $L__BB215_134; bra.uni $L__BB215_133; $L__BB215_134: ld.local.u32 %r1070, [%rd23+60]; mul.wide.s32 %rd389, %r1070, 208; add.s64 %rd390, %rd54, %rd389; ld.global.v4.f32 {%f1554, %f1555, %f1556, %f1557}, [%rd390]; ld.global.v4.f32 {%f1561, %f1562, %f1563, %f1564}, [%rd390+16]; ld.global.v4.f32 {%f1568, %f1569, %f1570, %f1571}, [%rd390+32]; mul.ftz.f32 %f1575, %f3066, %f1555; fma.rn.ftz.f32 %f1576, %f3065, %f1554, %f1575; fma.rn.ftz.f32 %f3072, %f3067, %f1556, %f1576; mul.ftz.f32 %f1577, %f3066, %f1562; fma.rn.ftz.f32 %f1578, %f3065, %f1561, %f1577; fma.rn.ftz.f32 %f3073, %f3067, %f1563, %f1578; mul.ftz.f32 %f1579, %f3066, %f1569; fma.rn.ftz.f32 %f1580, %f3065, %f1568, %f1579; fma.rn.ftz.f32 %f3074, %f3067, %f1570, %f1580; bra.uni $L__BB215_135; $L__BB215_133: ld.local.v4.f32 {%f1527, %f1528, %f1529, %f1530}, [%rd23+128]; mul.ftz.f32 %f1534, %f3066, %f1528; fma.rn.ftz.f32 %f1535, %f3065, %f1527, %f1534; fma.rn.ftz.f32 %f3072, %f3067, %f1529, %f1535; ld.local.v4.f32 {%f1536, %f1537, %f1538, %f1539}, [%rd23+144]; mul.ftz.f32 %f1543, %f3066, %f1537; fma.rn.ftz.f32 %f1544, %f3065, %f1536, %f1543; fma.rn.ftz.f32 %f3073, %f3067, %f1538, %f1544; ld.local.v4.f32 {%f1545, %f1546, %f1547, %f1548}, [%rd23+160]; mul.ftz.f32 %f1552, %f3066, %f1546; fma.rn.ftz.f32 %f1553, %f3065, %f1545, %f1552; fma.rn.ftz.f32 %f3074, %f3067, %f1547, %f1553; $L__BB215_135: setp.eq.s32 %p119, %r58, 1; selp.f32 %f1584, 0f3F000000, 0f00000000, %p119; add.ftz.f32 %f1585, %f86, %f171; setp.eq.s32 %p120, %r58, 2; selp.f32 %f1586, %f1585, %f171, %p120; sub.ftz.f32 %f1587, %f1586, %f1584; mul.ftz.f32 %f251, %f170, %f1587; mul.ftz.f32 %f1588, %f251, %f3072; mul.ftz.f32 %f1589, %f251, %f3073; mul.ftz.f32 %f1590, %f251, %f3074; sub.ftz.f32 %f1591, %f1415, %f1590; sub.ftz.f32 %f1592, %f1414, %f1589; sub.ftz.f32 %f1593, %f1413, %f1588; st.local.v2.f32 [%rd1], {%f1593, %f1592}; st.local.f32 [%rd23+-8], %f1591; ld.local.u32 %r1071, [%rd23+48]; setp.ne.s32 %p121, %r1071, 16; mov.f32 %f3081, 0f00000000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; @%p121 bra $L__BB215_164; and.b32 %r1072, %r2176, 3; setp.eq.s32 %p122, %r1072, 0; @%p122 bra $L__BB215_138; mul.wide.s32 %rd393, %r2177, 16; add.s64 %rd394, %rd57, %rd393; ld.global.v4.f32 {%f3081, %f3082, %f3083, %f3084}, [%rd394]; bra.uni $L__BB215_162; $L__BB215_138: setp.ne.s32 %p123, %r2176, 1024; mov.f32 %f3081, 0f00000000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; mov.f32 %f3084, %f3081; @%p123 bra $L__BB215_162; ld.local.v4.f32 {%f1602, %f1603, %f1604, %f1605}, [%rd1]; @%p118 bra $L__BB215_141; bra.uni $L__BB215_140; $L__BB215_141: ld.local.u32 %r1073, [%rd23+60]; mul.wide.s32 %rd397, %r1073, 208; add.s64 %rd398, %rd54, %rd397; ld.global.v4.f32 {%f1639, %f1640, %f1641, %f1642}, [%rd398+48]; ld.global.v4.f32 {%f1647, %f1648, %f1649, %f1650}, [%rd398+64]; ld.global.v4.f32 {%f1655, %f1656, %f1657, %f1658}, [%rd398+80]; mul.ftz.f32 %f1663, %f1603, %f1640; fma.rn.ftz.f32 %f1664, %f1602, %f1639, %f1663; fma.rn.ftz.f32 %f1665, %f1604, %f1641, %f1664; add.ftz.f32 %f271, %f1642, %f1665; mul.ftz.f32 %f1666, %f1603, %f1648; fma.rn.ftz.f32 %f1667, %f1602, %f1647, %f1666; fma.rn.ftz.f32 %f1668, %f1604, %f1649, %f1667; add.ftz.f32 %f272, %f1650, %f1668; mul.ftz.f32 %f1669, %f1603, %f1656; fma.rn.ftz.f32 %f1670, %f1602, %f1655, %f1669; fma.rn.ftz.f32 %f1671, %f1604, %f1657, %f1670; add.ftz.f32 %f273, %f1658, %f1671; bra.uni $L__BB215_142; $L__BB215_140: ld.local.v4.f32 {%f1606, %f1607, %f1608, %f1609}, [%rd23+176]; mul.ftz.f32 %f1614, %f1603, %f1607; fma.rn.ftz.f32 %f1615, %f1602, %f1606, %f1614; fma.rn.ftz.f32 %f1616, %f1604, %f1608, %f1615; add.ftz.f32 %f271, %f1609, %f1616; ld.local.v4.f32 {%f1617, %f1618, %f1619, %f1620}, [%rd23+192]; mul.ftz.f32 %f1625, %f1603, %f1618; fma.rn.ftz.f32 %f1626, %f1602, %f1617, %f1625; fma.rn.ftz.f32 %f1627, %f1604, %f1619, %f1626; add.ftz.f32 %f272, %f1620, %f1627; ld.local.v4.f32 {%f1628, %f1629, %f1630, %f1631}, [%rd23+208]; mul.ftz.f32 %f1636, %f1603, %f1629; fma.rn.ftz.f32 %f1637, %f1602, %f1628, %f1636; fma.rn.ftz.f32 %f1638, %f1604, %f1630, %f1637; add.ftz.f32 %f273, %f1631, %f1638; $L__BB215_142: ld.local.u32 %r102, [%rd23+36]; mul.wide.s32 %rd399, %r2177, 96; add.s64 %rd76, %rd58, %rd399; cvta.to.global.u64 %rd400, %rd58; add.s64 %rd401, %rd400, %rd399; add.s64 %rd77, %rd401, 32; ld.global.u32 %r1074, [%rd401+32]; setp.eq.s32 %p125, %r1074, 0; @%p125 bra $L__BB215_144; ld.global.v4.f32 {%f1672, %f1673, %f1674, %f1675}, [%rd77+16]; mul.ftz.f32 %f1680, %f272, %f1673; fma.rn.ftz.f32 %f1681, %f271, %f1672, %f1680; fma.rn.ftz.f32 %f1682, %f273, %f1674, %f1681; add.ftz.f32 %f268, %f1675, %f1682; ld.global.v4.f32 {%f1683, %f1684, %f1685, %f1686}, [%rd77+32]; mul.ftz.f32 %f1691, %f272, %f1684; fma.rn.ftz.f32 %f1692, %f271, %f1683, %f1691; fma.rn.ftz.f32 %f1693, %f273, %f1685, %f1692; add.ftz.f32 %f269, %f1686, %f1693; ld.global.v4.f32 {%f1694, %f1695, %f1696, %f1697}, [%rd77+48]; mul.ftz.f32 %f1702, %f272, %f1695; fma.rn.ftz.f32 %f1703, %f271, %f1694, %f1702; fma.rn.ftz.f32 %f1704, %f273, %f1696, %f1703; add.ftz.f32 %f273, %f1697, %f1704; mov.f32 %f271, %f268; mov.f32 %f272, %f269; $L__BB215_144: and.b32 %r1075, %r102, 16777216; setp.ne.s32 %p126, %r1075, 0; setp.eq.s32 %p127, %r1075, 0; selp.b32 %r105, -1, 2, %p127; @%p126 bra $L__BB215_146; ld.global.u32 %r105, [%rd77+-20]; $L__BB215_146: ld.global.u32 %r106, [%rd77+-24]; setp.eq.s32 %p128, %r106, 8; @%p128 bra $L__BB215_161; bra.uni $L__BB215_147; $L__BB215_161: { // callseq 1378, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIfEN7nanovdb4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS4_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3081, [retval0+0]; } // callseq 1378 mov.f32 %f3084, 0f3F800000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; bra.uni $L__BB215_162; $L__BB215_147: setp.eq.s32 %p129, %r106, 9; @%p129 bra $L__BB215_160; bra.uni $L__BB215_148; $L__BB215_160: { // callseq 1377, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .align 4 .b8 retval0[12]; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Vec3IfEEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS6_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3081, [retval0+0]; ld.param.f32 %f3082, [retval0+4]; ld.param.f32 %f3083, [retval0+8]; } // callseq 1377 mov.f32 %f3084, 0f3F800000; bra.uni $L__BB215_162; $L__BB215_148: setp.eq.s32 %p130, %r106, 10; @%p130 bra $L__BB215_159; bra.uni $L__BB215_149; $L__BB215_159: { // callseq 1376, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb3FpNEENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3081, [retval0+0]; } // callseq 1376 mov.f32 %f3084, 0f3F800000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; bra.uni $L__BB215_162; $L__BB215_149: setp.eq.s32 %p131, %r106, 11; @%p131 bra $L__BB215_158; bra.uni $L__BB215_150; $L__BB215_158: { // callseq 1375, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .b32 param4; st.param.b32 [param4+0], %r105; .param .b32 retval0; call.uni (retval0), _Z31kernel_tex_image_interp_nanovdbIN7nanovdb4Fp16EENS0_4GridINS0_4TreeINS0_8RootNodeINS0_12InternalNodeINS5_INS0_8LeafNodeIT_NS0_5CoordENS0_4MaskELj3EEELj4EEELj5EEEEEEEE9ValueTypeERK11TextureInfofffj, ( param0, param1, param2, param3, param4 ); ld.param.f32 %f3081, [retval0+0]; } // callseq 1375 mov.f32 %f3084, 0f3F800000; mov.f32 %f3082, %f3081; mov.f32 %f3083, %f3081; bra.uni $L__BB215_162; $L__BB215_150: or.b32 %r107, %r105, 1; setp.gt.u32 %p132, %r106, 6; @%p132 bra $L__BB215_155; mov.u32 %r1077, 1; shl.b32 %r1078, %r1077, %r106; and.b32 %r1079, %r1078, 71; setp.ne.s32 %p133, %r1079, 0; @%p133 bra $L__BB215_152; bra.uni $L__BB215_155; $L__BB215_152: setp.eq.s32 %p134, %r107, 3; @%p134 bra $L__BB215_154; bra.uni $L__BB215_153; $L__BB215_154: { // callseq 1373, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd76; .param .b32 param1; st.param.f32 [param1+0], %f271; .param .b32 param2; st.param.f32 [param2+0], %f272; .param .b32 param3; st.param.f32 [param3+0], %f273; .param .align 4 .b8 retval0[16]; call.uni (retval0), _Z32kernel_tex_image_interp_tricubicI6float4ET_RK11TextureInfofff, ( param0, param1, param2, param3 ); ld.param.f32 %f3081, [retval0+0]; ld.param.f32 %f3082, [retval0+4]; ld.param.f32 %f3083, [retval0+8]; ld.param.f32 %f3084, [retval0+12]; } // callseq 1373 bra.uni $L__BB215_162; $L__BB215_153: ld.global.u64 %rd402, [%rd77+-32]; tex.3d.v4.f32.f32 {%f3081, %f3082, %f3083, %f3084}, [%rd402, {%f271, %f272, %f273, %f273}]; $L__BB215_162: setp.leu.ftz.f32 %p136, %f3084, 0f358637BD; setp.eq.ftz.f32 %p137, %f3084, 0f3F800000; or.pred %p138, %p136, %p137; @%p138 bra $L__BB215_164; div.approx.ftz.f32 %f3083, %f3083, %f3084; div.approx.ftz.f32 %f3082, %f3082, %f3084; div.approx.ftz.f32 %f3081, %f3081, %f3084; $L__BB215_164: ld.local.u8 %rs43, [%rd23+40]; and.b16 %rs44, %rs43, 2; setp.eq.s16 %p139, %rs44, 0; @%p139 bra $L__BB215_166; bra.uni $L__BB215_165; $L__BB215_166: ld.local.u32 %r1080, [%rd23+60]; mul.wide.s32 %rd404, %r1080, 208; add.s64 %rd405, %rd54, %rd404; ld.global.v4.f32 {%f1735, %f1736, %f1737, %f1738}, [%rd405]; ld.global.v4.f32 {%f1742, %f1743, %f1744, %f1745}, [%rd405+16]; ld.global.v4.f32 {%f1749, %f1750, %f1751, %f1752}, [%rd405+32]; mul.ftz.f32 %f1756, %f3082, %f1736; fma.rn.ftz.f32 %f1757, %f3081, %f1735, %f1756; fma.rn.ftz.f32 %f3088, %f3083, %f1737, %f1757; mul.ftz.f32 %f1758, %f3082, %f1743; fma.rn.ftz.f32 %f1759, %f3081, %f1742, %f1758; fma.rn.ftz.f32 %f3089, %f3083, %f1744, %f1759; mul.ftz.f32 %f1760, %f3082, %f1750; fma.rn.ftz.f32 %f1761, %f3081, %f1749, %f1760; fma.rn.ftz.f32 %f3090, %f3083, %f1751, %f1761; bra.uni $L__BB215_167; $L__BB215_165: ld.local.v4.f32 {%f1708, %f1709, %f1710, %f1711}, [%rd23+128]; mul.ftz.f32 %f1715, %f3082, %f1709; fma.rn.ftz.f32 %f1716, %f3081, %f1708, %f1715; fma.rn.ftz.f32 %f3088, %f3083, %f1710, %f1716; ld.local.v4.f32 {%f1717, %f1718, %f1719, %f1720}, [%rd23+144]; mul.ftz.f32 %f1724, %f3082, %f1718; fma.rn.ftz.f32 %f1725, %f3081, %f1717, %f1724; fma.rn.ftz.f32 %f3089, %f3083, %f1719, %f1725; ld.local.v4.f32 {%f1726, %f1727, %f1728, %f1729}, [%rd23+160]; mul.ftz.f32 %f1733, %f3082, %f1727; fma.rn.ftz.f32 %f1734, %f3081, %f1726, %f1733; fma.rn.ftz.f32 %f3090, %f3083, %f1728, %f1734; $L__BB215_167: mul.ftz.f32 %f1762, %f251, %f3088; mul.ftz.f32 %f1763, %f251, %f3089; mul.ftz.f32 %f1764, %f251, %f3090; sub.ftz.f32 %f1765, %f1415, %f1764; sub.ftz.f32 %f1766, %f1414, %f1763; sub.ftz.f32 %f1767, %f1413, %f1762; st.local.v2.f32 [%rd1], {%f1767, %f1766}; st.local.f32 [%rd23+-8], %f1765; ld.local.u32 %r1081, [%rd23+32]; and.b32 %r2181, %r1081, 4194303; $L__BB215_168: add.u64 %rd408, %SP, 304; add.u64 %rd78, %SPL, 304; bra.uni $L__BB215_169; $L__BB215_212: mov.u32 %r1138, 24; { // callseq 1438, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1138; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1438 $L__BB215_169: mov.u32 %r110, %r2181; mul.wide.s32 %rd409, %r110, 16; add.s64 %rd80, %rd79, %rd409; ld.global.v4.u32 {%r1083, %r1084, %r1085, %r1086}, [%rd80]; mul.wide.u32 %rd410, %r1084, 4; add.s64 %rd81, %rd78, %rd410; mul.wide.u32 %rd411, %r1085, 4; add.s64 %rd82, %rd78, %rd411; add.s32 %r2181, %r110, 1; setp.gt.s32 %p140, %r1083, 35; @%p140 bra $L__BB215_217; setp.gt.s32 %p228, %r1083, 24; @%p228 bra $L__BB215_213; bra.uni $L__BB215_171; $L__BB215_213: add.s32 %r1088, %r1083, -27; setp.lt.u32 %p229, %r1088, 9; @%p229 bra $L__BB215_169; setp.eq.s32 %p230, %r1083, 25; @%p230 bra $L__BB215_369; setp.eq.s32 %p231, %r1083, 26; @%p231 bra $L__BB215_216; bra.uni $L__BB215_382; $L__BB215_216: mov.u32 %r1136, 26; { // callseq 1436, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1136; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z17svm_node_set_bumpILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1436 bra.uni $L__BB215_169; $L__BB215_217: setp.gt.s32 %p141, %r1083, 71; @%p141 bra $L__BB215_285; setp.gt.s32 %p177, %r1083, 53; @%p177 bra $L__BB215_252; setp.gt.s32 %p203, %r1083, 44; @%p203 bra $L__BB215_236; setp.gt.s32 %p216, %r1083, 39; @%p216 bra $L__BB215_228; bra.uni $L__BB215_221; $L__BB215_228: setp.gt.s32 %p217, %r1083, 41; @%p217 bra $L__BB215_232; bra.uni $L__BB215_229; $L__BB215_232: setp.eq.s32 %p218, %r1083, 42; @%p218 bra $L__BB215_365; setp.eq.s32 %p219, %r1083, 43; @%p219 bra $L__BB215_364; setp.eq.s32 %p220, %r1083, 44; @%p220 bra $L__BB215_235; bra.uni $L__BB215_382; $L__BB215_235: mov.u32 %r1127, 44; add.s32 %r1128, %r110, 1; { // callseq 1427, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1127; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1128; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1427 bra.uni $L__BB215_169; $L__BB215_171: setp.gt.s32 %p232, %r1083, 17; @%p232 bra $L__BB215_205; bra.uni $L__BB215_172; $L__BB215_205: setp.lt.s32 %p233, %r1083, 21; @%p233 bra $L__BB215_169; setp.gt.s32 %p234, %r1083, 22; @%p234 bra $L__BB215_210; setp.eq.s32 %p237, %r1083, 21; @%p237 bra $L__BB215_371; setp.eq.s32 %p238, %r1083, 22; @%p238 bra $L__BB215_209; bra.uni $L__BB215_382; $L__BB215_209: mov.u32 %r1141, 22; add.s32 %r1142, %r110, 1; { // callseq 1440, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1141; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1142; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj294EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1440 bra.uni $L__BB215_169; $L__BB215_285: setp.gt.s32 %p142, %r1083, 89; @%p142 bra $L__BB215_320; setp.gt.s32 %p152, %r1083, 80; @%p152 bra $L__BB215_304; setp.gt.s32 %p165, %r1083, 75; @%p165 bra $L__BB215_295; bra.uni $L__BB215_288; $L__BB215_295: setp.gt.s32 %p166, %r1083, 77; @%p166 bra $L__BB215_300; bra.uni $L__BB215_296; $L__BB215_300: setp.eq.s32 %p167, %r1083, 78; @%p167 bra $L__BB215_340; setp.eq.s32 %p168, %r1083, 79; @%p168 bra $L__BB215_339; setp.eq.s32 %p169, %r1083, 80; @%p169 bra $L__BB215_303; bra.uni $L__BB215_382; $L__BB215_303: { // callseq 1393, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1393 bra.uni $L__BB215_169; $L__BB215_252: setp.gt.s32 %p178, %r1083, 62; @%p178 bra $L__BB215_269; setp.gt.s32 %p191, %r1083, 57; @%p191 bra $L__BB215_261; bra.uni $L__BB215_254; $L__BB215_261: setp.gt.s32 %p192, %r1083, 59; @%p192 bra $L__BB215_265; bra.uni $L__BB215_262; $L__BB215_265: setp.eq.s32 %p193, %r1083, 60; @%p193 bra $L__BB215_355; setp.eq.s32 %p194, %r1083, 61; @%p194 bra $L__BB215_354; setp.eq.s32 %p195, %r1083, 62; @%p195 bra $L__BB215_268; bra.uni $L__BB215_382; $L__BB215_268: mov.u32 %r1114, 62; add.s32 %r1115, %r110, 1; { // callseq 1409, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1114; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1115; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1409 bra.uni $L__BB215_169; $L__BB215_172: setp.gt.s32 %p239, %r1083, 8; @%p239 bra $L__BB215_188; bra.uni $L__BB215_173; $L__BB215_188: setp.gt.s32 %p240, %r1083, 12; @%p240 bra $L__BB215_197; bra.uni $L__BB215_189; $L__BB215_197: setp.gt.s32 %p241, %r1083, 14; @%p241 bra $L__BB215_201; bra.uni $L__BB215_198; $L__BB215_201: setp.eq.s32 %p242, %r1083, 15; @%p242 bra $L__BB215_373; setp.eq.s32 %p243, %r1083, 16; @%p243 bra $L__BB215_372; setp.eq.s32 %p244, %r1083, 17; @%p244 bra $L__BB215_204; bra.uni $L__BB215_382; $L__BB215_204: { // callseq 1442, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; .param .b32 param4; st.param.b32 [param4+0], %r1086; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1442 bra.uni $L__BB215_169; $L__BB215_320: setp.gt.s32 %p143, %r1083, 93; @%p143 bra $L__BB215_324; add.s32 %r1087, %r1083, -92; setp.lt.u32 %p150, %r1087, 2; @%p150 bra $L__BB215_169; setp.eq.s32 %p151, %r1083, 90; @%p151 bra $L__BB215_323; bra.uni $L__BB215_382; $L__BB215_323: mov.u32 %r1091, 90; add.s32 %r1092, %r110, 1; { // callseq 1383, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1091; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; .param .b32 param3; st.param.b32 [param3+0], %r1092; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1383 bra.uni $L__BB215_169; $L__BB215_236: setp.gt.s32 %p204, %r1083, 48; @%p204 bra $L__BB215_244; bra.uni $L__BB215_237; $L__BB215_244: setp.gt.s32 %p205, %r1083, 50; @%p205 bra $L__BB215_248; bra.uni $L__BB215_245; $L__BB215_248: setp.eq.s32 %p206, %r1083, 51; @%p206 bra $L__BB215_360; setp.eq.s32 %p207, %r1083, 52; @%p207 bra $L__BB215_359; setp.eq.s32 %p208, %r1083, 53; @%p208 bra $L__BB215_251; bra.uni $L__BB215_382; $L__BB215_251: { // callseq 1418, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1418 bra.uni $L__BB215_169; $L__BB215_304: setp.gt.s32 %p153, %r1083, 84; @%p153 bra $L__BB215_312; bra.uni $L__BB215_305; $L__BB215_312: setp.gt.s32 %p154, %r1083, 86; @%p154 bra $L__BB215_316; bra.uni $L__BB215_313; $L__BB215_316: setp.eq.s32 %p155, %r1083, 87; @%p155 bra $L__BB215_335; setp.eq.s32 %p156, %r1083, 88; @%p156 bra $L__BB215_334; setp.eq.s32 %p157, %r1083, 89; @%p157 bra $L__BB215_319; bra.uni $L__BB215_382; $L__BB215_319: mov.u32 %r1093, 89; { // callseq 1384, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1093; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z11svm_node_aoILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1384 bra.uni $L__BB215_169; $L__BB215_269: setp.gt.s32 %p179, %r1083, 66; @%p179 bra $L__BB215_277; bra.uni $L__BB215_270; $L__BB215_277: setp.gt.s32 %p180, %r1083, 68; @%p180 bra $L__BB215_281; bra.uni $L__BB215_278; $L__BB215_281: setp.eq.s32 %p181, %r1083, 69; @%p181 bra $L__BB215_350; setp.eq.s32 %p182, %r1083, 70; @%p182 bra $L__BB215_349; setp.eq.s32 %p183, %r1083, 71; @%p183 bra $L__BB215_284; bra.uni $L__BB215_382; $L__BB215_284: mov.u32 %r1104, 71; { // callseq 1400, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1104; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1400 bra.uni $L__BB215_169; $L__BB215_173: setp.gt.s32 %p252, %r1083, 4; @%p252 bra $L__BB215_181; setp.gt.s32 %p258, %r1083, 2; @%p258 bra $L__BB215_178; setp.eq.s32 %p261, %r1083, 1; mov.u32 %r2181, %r1085; @%p261 bra $L__BB215_169; setp.eq.s32 %p262, %r1083, 2; @%p262 bra $L__BB215_177; bra.uni $L__BB215_382; $L__BB215_177: mov.u32 %r1160, 2; add.s32 %r1161, %r110, 1; { // callseq 1451, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r1160; st.param.b32 [param0+4], %r1084; st.param.b32 [param0+8], %r1085; st.param.b32 [param0+12], %r1086; .param .b32 param1; st.param.b32 [param1+0], %r1161; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj294EL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1451 bra.uni $L__BB215_169; $L__BB215_324: setp.gt.s32 %p144, %r1083, 95; @%p144 bra $L__BB215_328; bra.uni $L__BB215_325; $L__BB215_328: setp.eq.s32 %p145, %r1083, 98; @%p145 bra $L__BB215_381; setp.eq.s32 %p146, %r1083, 97; @%p146 bra $L__BB215_333; setp.ne.s32 %p147, %r1083, 96; @%p147 bra $L__BB215_382; { // callseq 1379, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1379 bra.uni $L__BB215_169; $L__BB215_189: setp.gt.s32 %p247, %r1083, 10; @%p247 bra $L__BB215_194; setp.eq.s32 %p250, %r1083, 9; @%p250 bra $L__BB215_376; setp.eq.s32 %p251, %r1083, 10; @%p251 bra $L__BB215_192; bra.uni $L__BB215_382; $L__BB215_192: ld.local.f32 %f1784, [%rd82]; setp.ltu.ftz.f32 %p267, %f1784, 0f3F800000; @%p267 bra $L__BB215_169; add.s32 %r1154, %r1084, %r110; add.s32 %r2181, %r1154, 1; bra.uni $L__BB215_169; $L__BB215_181: setp.gt.s32 %p253, %r1083, 6; @%p253 bra $L__BB215_185; setp.eq.s32 %p256, %r1083, 5; @%p256 bra $L__BB215_379; setp.eq.s32 %p257, %r1083, 6; @%p257 bra $L__BB215_184; bra.uni $L__BB215_382; $L__BB215_184: ld.local.f32 %f1786, [%rd81]; ld.local.f32 %f1787, [%rd81+4]; ld.local.f32 %f1788, [%rd81+8]; st.local.f32 [%rd59+12], %f1786; st.local.f32 [%rd59+16], %f1787; st.local.f32 [%rd59+20], %f1788; bra.uni $L__BB215_169; $L__BB215_221: setp.gt.s32 %p223, %r1083, 37; @%p223 bra $L__BB215_225; setp.eq.s32 %p226, %r1083, 36; @%p226 bra $L__BB215_368; setp.eq.s32 %p227, %r1083, 37; @%p227 bra $L__BB215_224; bra.uni $L__BB215_382; $L__BB215_224: mov.u32 %r1134, 37; { // callseq 1434, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1134, %r1084, %r1085, %r1086}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1434 bra.uni $L__BB215_169; $L__BB215_288: setp.gt.s32 %p172, %r1083, 73; @%p172 bra $L__BB215_292; setp.eq.s32 %p175, %r1083, 72; @%p175 bra $L__BB215_348; setp.eq.s32 %p176, %r1083, 73; @%p176 bra $L__BB215_291; bra.uni $L__BB215_382; $L__BB215_291: add.s32 %r1103, %r110, 1; { // callseq 1398, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1103; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1398 bra.uni $L__BB215_169; $L__BB215_254: setp.gt.s32 %p198, %r1083, 55; @%p198 bra $L__BB215_258; setp.eq.s32 %p201, %r1083, 54; @%p201 bra $L__BB215_358; setp.eq.s32 %p202, %r1083, 55; @%p202 bra $L__BB215_257; bra.uni $L__BB215_382; $L__BB215_257: ld.local.v4.f32 {%f1777, %f1778, %f1779, %f1780}, [%rd1]; { // callseq 1416, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f1777; st.param.f32 [param0+4], %f1778; st.param.f32 [param0+8], %f1779; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; .param .b32 param4; st.param.b32 [param4+0], %r1086; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1416 bra.uni $L__BB215_169; $L__BB215_237: setp.gt.s32 %p211, %r1083, 46; @%p211 bra $L__BB215_241; setp.eq.s32 %p214, %r1083, 45; @%p214 bra $L__BB215_363; setp.eq.s32 %p215, %r1083, 46; @%p215 bra $L__BB215_240; bra.uni $L__BB215_382; $L__BB215_240: { // callseq 1425, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1425 bra.uni $L__BB215_169; $L__BB215_305: setp.gt.s32 %p160, %r1083, 82; @%p160 bra $L__BB215_309; setp.eq.s32 %p163, %r1083, 81; @%p163 bra $L__BB215_338; setp.eq.s32 %p164, %r1083, 82; @%p164 bra $L__BB215_308; bra.uni $L__BB215_382; $L__BB215_308: mov.u32 %r1098, 82; { // callseq 1391, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1098; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1391 bra.uni $L__BB215_169; $L__BB215_270: setp.gt.s32 %p186, %r1083, 64; @%p186 bra $L__BB215_274; setp.eq.s32 %p189, %r1083, 63; @%p189 bra $L__BB215_353; setp.eq.s32 %p190, %r1083, 64; @%p190 bra $L__BB215_273; bra.uni $L__BB215_382; $L__BB215_273: mov.u32 %r1111, 64; add.s32 %r1112, %r110, 1; { // callseq 1407, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1111; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1112; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1407 bra.uni $L__BB215_169; $L__BB215_210: setp.eq.s32 %p235, %r1083, 23; @%p235 bra $L__BB215_370; setp.eq.s32 %p236, %r1083, 24; @%p236 bra $L__BB215_212; bra.uni $L__BB215_382; $L__BB215_198: setp.eq.s32 %p245, %r1083, 13; @%p245 bra $L__BB215_374; setp.eq.s32 %p246, %r1083, 14; @%p246 bra $L__BB215_200; bra.uni $L__BB215_382; $L__BB215_200: st.local.u32 [%rd82], %r1084; bra.uni $L__BB215_169; $L__BB215_178: setp.eq.s32 %p259, %r1083, 3; @%p259 bra $L__BB215_380; setp.eq.s32 %p260, %r1083, 4; @%p260 bra $L__BB215_180; bra.uni $L__BB215_382; $L__BB215_180: mov.u32 %r1158, 4; { // callseq 1449, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1158, %r1084, %r1085, %r1086}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1449 bra.uni $L__BB215_169; $L__BB215_194: setp.eq.s32 %p248, %r1083, 11; @%p248 bra $L__BB215_375; setp.eq.s32 %p249, %r1083, 12; @%p249 bra $L__BB215_196; bra.uni $L__BB215_382; $L__BB215_196: { // callseq 1445, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1445 bra.uni $L__BB215_169; $L__BB215_185: setp.eq.s32 %p254, %r1083, 7; @%p254 bra $L__BB215_378; setp.eq.s32 %p255, %r1083, 8; @%p255 bra $L__BB215_187; bra.uni $L__BB215_382; $L__BB215_187: mov.u32 %r1156, 8; { // callseq 1447, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1156; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1447 bra.uni $L__BB215_169; $L__BB215_325: setp.eq.s32 %p148, %r1083, 94; @%p148 bra $L__BB215_332; setp.eq.s32 %p149, %r1083, 95; @%p149 bra $L__BB215_327; bra.uni $L__BB215_382; $L__BB215_327: { // callseq 1380, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1380 bra.uni $L__BB215_169; $L__BB215_229: setp.eq.s32 %p221, %r1083, 40; @%p221 bra $L__BB215_366; setp.eq.s32 %p222, %r1083, 41; @%p222 bra $L__BB215_231; bra.uni $L__BB215_382; $L__BB215_231: mov.u32 %r1130, 41; add.s32 %r1131, %r110, 1; { // callseq 1430, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1130; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; .param .b32 param3; st.param.b32 [param3+0], %r60; .param .b32 param4; st.param.b32 [param4+0], %r1131; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType1EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1430 bra.uni $L__BB215_169; $L__BB215_296: setp.eq.s32 %p170, %r1083, 76; @%p170 bra $L__BB215_341; setp.eq.s32 %p171, %r1083, 77; @%p171 bra $L__BB215_298; bra.uni $L__BB215_382; $L__BB215_298: setp.eq.s32 %p263, %r1086, 255; @%p263 bra $L__BB215_169; ld.local.f32 %f1773, [%rd81]; add.s32 %r1102, %r1086, %r1085; mul.wide.u32 %rd434, %r1102, 4; add.s64 %rd435, %rd78, %rd434; st.local.f32 [%rd435], %f1773; bra.uni $L__BB215_169; $L__BB215_262: setp.eq.s32 %p196, %r1083, 58; @%p196 bra $L__BB215_356; setp.eq.s32 %p197, %r1083, 59; @%p197 bra $L__BB215_264; bra.uni $L__BB215_382; $L__BB215_264: add.s32 %r1119, %r110, 1; { // callseq 1412, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1119; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj294EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1412 bra.uni $L__BB215_169; $L__BB215_245: setp.eq.s32 %p209, %r1083, 49; @%p209 bra $L__BB215_361; setp.eq.s32 %p210, %r1083, 50; @%p210 bra $L__BB215_247; bra.uni $L__BB215_382; $L__BB215_247: { // callseq 1421, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1421 bra.uni $L__BB215_169; $L__BB215_313: setp.eq.s32 %p158, %r1083, 85; @%p158 bra $L__BB215_336; setp.eq.s32 %p159, %r1083, 86; @%p159 bra $L__BB215_315; bra.uni $L__BB215_382; $L__BB215_315: add.s32 %r1096, %r110, 1; { // callseq 1387, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1096; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1387 bra.uni $L__BB215_169; $L__BB215_278: setp.eq.s32 %p184, %r1083, 67; @%p184 bra $L__BB215_351; setp.eq.s32 %p185, %r1083, 68; @%p185 bra $L__BB215_280; bra.uni $L__BB215_382; $L__BB215_280: mov.u32 %r1108, 68; { // callseq 1403, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1108; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1403 bra.uni $L__BB215_169; $L__BB215_225: setp.eq.s32 %p224, %r1083, 38; @%p224 bra $L__BB215_367; setp.eq.s32 %p225, %r1083, 39; @%p225 bra $L__BB215_227; bra.uni $L__BB215_382; $L__BB215_227: mov.u32 %r1133, 39; { // callseq 1432, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1133, %r1084, %r1085, %r1086}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1432 bra.uni $L__BB215_169; $L__BB215_292: setp.eq.s32 %p173, %r1083, 74; @%p173 bra $L__BB215_347; setp.eq.s32 %p174, %r1083, 75; @%p174 bra $L__BB215_294; bra.uni $L__BB215_382; $L__BB215_294: { // callseq 1396, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1396 bra.uni $L__BB215_169; $L__BB215_258: setp.eq.s32 %p199, %r1083, 56; @%p199 bra $L__BB215_357; setp.eq.s32 %p200, %r1083, 57; @%p200 bra $L__BB215_260; bra.uni $L__BB215_382; $L__BB215_260: mov.u32 %r1121, 57; add.s32 %r1122, %r110, 1; { // callseq 1414, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1121; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1122; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1414 bra.uni $L__BB215_169; $L__BB215_241: setp.eq.s32 %p212, %r1083, 47; @%p212 bra $L__BB215_362; setp.eq.s32 %p213, %r1083, 48; @%p213 bra $L__BB215_243; bra.uni $L__BB215_382; $L__BB215_243: { // callseq 1423, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1423 bra.uni $L__BB215_169; $L__BB215_309: setp.eq.s32 %p161, %r1083, 83; @%p161 bra $L__BB215_337; setp.eq.s32 %p162, %r1083, 84; @%p162 bra $L__BB215_311; bra.uni $L__BB215_382; $L__BB215_311: { // callseq 1389, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1389 bra.uni $L__BB215_169; $L__BB215_274: setp.eq.s32 %p187, %r1083, 65; @%p187 bra $L__BB215_352; setp.eq.s32 %p188, %r1083, 66; @%p188 bra $L__BB215_276; bra.uni $L__BB215_382; $L__BB215_276: add.s32 %r1110, %r110, 1; { // callseq 1405, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1110; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1405 bra.uni $L__BB215_169; $L__BB215_371: mov.u32 %r1143, 21; { // callseq 1441, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1143; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z21svm_node_displacementILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1441 bra.uni $L__BB215_169; $L__BB215_373: ld.global.v4.u32 {%r1145, %r1146, %r1147, %r1148}, [%rd80+16]; add.s32 %r2181, %r110, 2; st.local.u32 [%rd81], %r1146; st.local.u32 [%rd81+4], %r1147; st.local.u32 [%rd81+8], %r1148; bra.uni $L__BB215_169; $L__BB215_372: { // callseq 1443, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r492; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z13svm_node_attrILj294EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1443 bra.uni $L__BB215_169; $L__BB215_376: ld.local.f32 %f1785, [%rd82]; setp.gtu.ftz.f32 %p268, %f1785, 0f00000000; @%p268 bra $L__BB215_169; add.s32 %r1155, %r1084, %r110; add.s32 %r2181, %r1155, 1; bra.uni $L__BB215_169; $L__BB215_379: st.local.u32 [%rd59+12], %r1084; st.local.u32 [%rd59+16], %r1085; st.local.u32 [%rd59+20], %r1086; bra.uni $L__BB215_169; $L__BB215_370: mov.u32 %r1139, 23; add.s32 %r1140, %r110, 1; { // callseq 1439, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1139; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1140; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1439 bra.uni $L__BB215_169; $L__BB215_374: mov.u32 %r1152, 13; add.s32 %r1153, %r110, 1; { // callseq 1444, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b32 param1; st.param.b32 [param1+0], %r60; .param .b64 param2; st.param.b64 [param2+0], %rd408; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1152; st.param.b32 [param3+4], %r1084; st.param.b32 [param3+8], %r1085; st.param.b32 [param3+12], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1153; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1444 bra.uni $L__BB215_169; $L__BB215_380: mov.u32 %r1159, 3; { // callseq 1450, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1159, %r1084, %r1085, %r1086}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1450 bra.uni $L__BB215_169; $L__BB215_375: { // callseq 1446, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1446 bra.uni $L__BB215_169; $L__BB215_378: mov.u32 %r1157, 7; { // callseq 1448, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1157; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1448 bra.uni $L__BB215_169; $L__BB215_365: { // callseq 1429, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1429 bra.uni $L__BB215_169; $L__BB215_364: add.s32 %r1129, %r110, 1; { // callseq 1428, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1129; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1428 bra.uni $L__BB215_169; $L__BB215_340: add.s32 %r1101, %r110, 1; { // callseq 1395, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1101; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1395 bra.uni $L__BB215_169; $L__BB215_339: add.s32 %r1100, %r110, 1; { // callseq 1394, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1100; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1394 bra.uni $L__BB215_169; $L__BB215_355: add.s32 %r1118, %r110, 1; { // callseq 1411, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1118; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1411 bra.uni $L__BB215_169; $L__BB215_354: mov.u32 %r1116, 61; add.s32 %r1117, %r110, 1; { // callseq 1410, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1116; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1117; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1410 bra.uni $L__BB215_169; $L__BB215_360: { // callseq 1420, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1420 bra.uni $L__BB215_169; $L__BB215_359: add.s32 %r1125, %r110, 1; { // callseq 1419, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1125; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1419 bra.uni $L__BB215_169; $L__BB215_335: add.s32 %r1095, %r110, 1; { // callseq 1386, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1095; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1386 bra.uni $L__BB215_169; $L__BB215_334: mov.u32 %r1094, 88; ld.local.f32 %f1768, [%rd23+-4]; ld.local.v2.f32 {%f1769, %f1770}, [%rd23]; { // callseq 1385, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f1768; .param .b32 param1; st.param.f32 [param1+0], %f1769; .param .b32 param2; st.param.f32 [param2+0], %f1770; .param .b64 param3; st.param.b64 [param3+0], %rd408; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r1094; st.param.b32 [param4+4], %r1084; st.param.b32 [param4+8], %r1085; st.param.b32 [param4+12], %r1086; call.uni _Z14svm_node_bevelILj294EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1385 bra.uni $L__BB215_169; $L__BB215_350: mov.u32 %r1106, 69; add.s32 %r1107, %r110, 1; { // callseq 1402, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1106; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1107; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1402 bra.uni $L__BB215_169; $L__BB215_349: mov.u32 %r1105, 70; { // callseq 1401, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1105; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1401 bra.uni $L__BB215_169; $L__BB215_381: { // callseq 1452, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1452 bra.uni $L__BB215_169; $L__BB215_333: { // callseq 1382, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1382 bra.uni $L__BB215_169; $L__BB215_368: mov.u32 %r1135, 36; { // callseq 1435, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1135; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1435 bra.uni $L__BB215_169; $L__BB215_348: { // callseq 1399, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1399 bra.uni $L__BB215_169; $L__BB215_358: add.s32 %r1124, %r110, 1; { // callseq 1417, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1124; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1417 bra.uni $L__BB215_169; $L__BB215_363: { // callseq 1426, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1426 bra.uni $L__BB215_169; $L__BB215_338: mov.u32 %r1099, 81; { // callseq 1392, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1099; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1392 bra.uni $L__BB215_169; $L__BB215_353: mov.u32 %r1113, 63; { // callseq 1408, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1113; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1408 bra.uni $L__BB215_169; $L__BB215_332: mov.u32 %r1089, 94; add.s32 %r1090, %r110, 1; { // callseq 1381, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1089; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; .param .b32 param2; st.param.b32 [param2+0], %r1090; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1381 bra.uni $L__BB215_169; $L__BB215_366: mov.u32 %r1132, 40; { // callseq 1431, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1132; st.param.b32 [param2+4], %r1084; st.param.b32 [param2+8], %r1085; st.param.b32 [param2+12], %r1086; call.uni _Z23svm_node_closure_volumeIL10ShaderType1EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1431 bra.uni $L__BB215_169; $L__BB215_341: setp.eq.s32 %p264, %r1086, 255; @%p264 bra $L__BB215_169; add.s64 %rd83, %rd81, 4; add.s64 %rd84, %rd81, 8; setp.eq.s32 %p265, %r1085, 0; mul.wide.u32 %rd438, %r1086, 4; add.s64 %rd85, %rd78, %rd438; @%p265 bra $L__BB215_346; setp.eq.s32 %p266, %r1085, 1; @%p266 bra $L__BB215_345; bra.uni $L__BB215_344; $L__BB215_345: ld.local.f32 %f1775, [%rd83]; st.local.f32 [%rd85], %f1775; bra.uni $L__BB215_169; $L__BB215_356: mov.u32 %r1120, 58; { // callseq 1413, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1120; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1413 bra.uni $L__BB215_169; $L__BB215_361: { // callseq 1422, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1422 bra.uni $L__BB215_169; $L__BB215_336: add.s32 %r1097, %r110, 1; { // callseq 1388, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1097; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1388 bra.uni $L__BB215_169; $L__BB215_351: mov.u32 %r1109, 67; { // callseq 1404, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1109, %r1084, %r1085, %r1086}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1404 bra.uni $L__BB215_169; $L__BB215_367: { // callseq 1433, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd227; .param .b64 param1; st.param.b64 [param1+0], %rd408; .param .b32 param2; st.param.b32 [param2+0], %r1084; .param .b32 param3; st.param.b32 [param3+0], %r1085; .param .b32 param4; st.param.b32 [param4+0], %r1086; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1433 bra.uni $L__BB215_169; $L__BB215_347: { // callseq 1397, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1397 bra.uni $L__BB215_169; $L__BB215_357: mov.u32 %r1123, 56; { // callseq 1415, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1123; st.param.b32 [param1+4], %r1084; st.param.b32 [param1+8], %r1085; st.param.b32 [param1+12], %r1086; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1415 bra.uni $L__BB215_169; $L__BB215_362: cvt.u32.u64 %r1126, %rd4; { // callseq 1424, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1126; .param .b64 param1; st.param.b64 [param1+0], %rd227; .param .b64 param2; st.param.b64 [param2+0], %rd408; .param .b32 param3; st.param.b32 [param3+0], %r1084; .param .b32 param4; st.param.b32 [param4+0], %r1085; .param .b32 param5; st.param.b32 [param5+0], %r60; call.uni _Z19svm_node_light_pathILj294EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4, param5 ); } // callseq 1424 bra.uni $L__BB215_169; $L__BB215_337: { // callseq 1390, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1390 bra.uni $L__BB215_169; $L__BB215_352: { // callseq 1406, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1406 bra.uni $L__BB215_169; $L__BB215_369: add.s32 %r1137, %r110, 1; { // callseq 1437, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd408; .param .b32 param1; st.param.b32 [param1+0], %r1084; .param .b32 param2; st.param.b32 [param2+0], %r1085; .param .b32 param3; st.param.b32 [param3+0], %r1086; .param .b32 param4; st.param.b32 [param4+0], %r1137; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2181, [retval0+0]; } // callseq 1437 bra.uni $L__BB215_169; $L__BB215_346: ld.local.f32 %f1776, [%rd81]; st.local.f32 [%rd85], %f1776; bra.uni $L__BB215_169; $L__BB215_344: ld.local.f32 %f1774, [%rd84]; st.local.f32 [%rd85], %f1774; bra.uni $L__BB215_169; $L__BB215_382: ld.local.u32 %r2192, [%rd59+4]; setp.lt.s32 %p269, %r2192, 1; setp.eq.s32 %p270, %r2169, 0; or.pred %p271, %p270, %p269; @%p271 bra $L__BB215_399; mov.u32 %r2188, 0; $L__BB215_384: mul.wide.s32 %rd518, %r2188, 80; add.s64 %rd519, %rd1, %rd518; add.s64 %rd86, %rd519, 316; ld.local.u32 %r1163, [%rd519+316]; setp.ne.s32 %p272, %r1163, 38; add.s32 %r206, %r2188, 1; setp.ge.s32 %p273, %r206, %r2192; or.pred %p274, %p272, %p273; @%p274 bra $L__BB215_398; mov.u32 %r2186, 38; mov.u32 %r2187, %r206; bra.uni $L__BB215_386; $L__BB215_1098: ld.local.u32 %r2186, [%rd86]; mov.u32 %r2188, %r2194; $L__BB215_386: cvt.s64.s32 %rd87, %r2187; mul.wide.s32 %rd520, %r2187, 80; add.s64 %rd521, %rd1, %rd520; add.s64 %rd88, %rd521, 316; ld.local.u32 %r1165, [%rd521+316]; setp.ne.s32 %p275, %r2186, %r1165; mov.u32 %r2194, %r2187; @%p275 bra $L__BB215_397; ld.local.f32 %f1789, [%rd88+20]; ld.local.f32 %f1790, [%rd86+20]; setp.neu.ftz.f32 %p276, %f1790, %f1789; mov.u32 %r2194, %r2187; @%p276 bra $L__BB215_397; ld.local.v4.f32 {%f1791, %f1792, %f1793, %f1794}, [%rd86+-12]; ld.local.f32 %f1798, [%rd88+-12]; ld.local.f32 %f1799, [%rd88+-8]; ld.local.f32 %f1800, [%rd88+-4]; add.ftz.f32 %f1801, %f1793, %f1800; add.ftz.f32 %f1802, %f1792, %f1799; add.ftz.f32 %f1803, %f1791, %f1798; st.local.v2.f32 [%rd86+-12], {%f1803, %f1802}; st.local.f32 [%rd86+-4], %f1801; ld.local.f32 %f1804, [%rd86+4]; ld.local.f32 %f1805, [%rd88+4]; add.ftz.f32 %f1806, %f1805, %f1804; st.local.f32 [%rd86+4], %f1806; mov.u32 %r1166, -2; sub.s32 %r1167, %r1166, %r2188; add.s32 %r1168, %r1167, %r2192; setp.lt.s32 %p277, %r1168, 1; @%p277 bra $L__BB215_396; mul.lo.s64 %rd522, %rd87, 80; add.s64 %rd523, %rd1, %rd522; add.s64 %rd89, %rd523, 304; add.s32 %r1170, %r2192, -2; sub.s32 %r212, %r1170, %r2188; add.s32 %r1171, %r2192, -3; sub.s32 %r1172, %r1171, %r2188; and.b32 %r213, %r212, 3; setp.lt.u32 %p278, %r1172, 3; mov.u32 %r2191, 0; @%p278 bra $L__BB215_392; sub.s32 %r2190, %r212, %r213; mov.u32 %r2191, 0; $L__BB215_391: mul.wide.s32 %rd524, %r2191, 80; add.s64 %rd525, %rd89, %rd524; ld.local.v4.u32 {%r1174, %r1175, %r1176, %r1177}, [%rd525+80]; ld.local.v4.u32 {%r1182, %r1183, %r1184, %r1185}, [%rd525+96]; ld.local.v4.u32 {%r1190, %r1191, %r1192, %r1193}, [%rd525+112]; ld.local.v4.u32 {%r1198, %r1199, %r1200, %r1201}, [%rd525+128]; ld.local.v4.u32 {%r1206, %r1207, %r1208, %r1209}, [%rd525+144]; st.local.v4.u32 [%rd525], {%r1174, %r1175, %r1176, %r1177}; st.local.v4.u32 [%rd525+16], {%r1182, %r1183, %r1184, %r1185}; st.local.v4.u32 [%rd525+32], {%r1190, %r1191, %r1192, %r1193}; st.local.v4.u32 [%rd525+48], {%r1198, %r1199, %r1200, %r1201}; st.local.v4.u32 [%rd525+64], {%r1206, %r1207, %r1208, %r1209}; ld.local.v4.u32 {%r1214, %r1215, %r1216, %r1217}, [%rd525+160]; ld.local.v4.u32 {%r1222, %r1223, %r1224, %r1225}, [%rd525+176]; ld.local.v4.u32 {%r1230, %r1231, %r1232, %r1233}, [%rd525+192]; ld.local.v4.u32 {%r1238, %r1239, %r1240, %r1241}, [%rd525+208]; ld.local.v4.u32 {%r1246, %r1247, %r1248, %r1249}, [%rd525+224]; st.local.v4.u32 [%rd525+80], {%r1214, %r1215, %r1216, %r1217}; st.local.v4.u32 [%rd525+96], {%r1222, %r1223, %r1224, %r1225}; st.local.v4.u32 [%rd525+112], {%r1230, %r1231, %r1232, %r1233}; st.local.v4.u32 [%rd525+128], {%r1238, %r1239, %r1240, %r1241}; st.local.v4.u32 [%rd525+144], {%r1246, %r1247, %r1248, %r1249}; ld.local.v4.u32 {%r1254, %r1255, %r1256, %r1257}, [%rd525+240]; ld.local.v4.u32 {%r1262, %r1263, %r1264, %r1265}, [%rd525+256]; ld.local.v4.u32 {%r1270, %r1271, %r1272, %r1273}, [%rd525+272]; ld.local.v4.u32 {%r1278, %r1279, %r1280, %r1281}, [%rd525+288]; ld.local.v4.u32 {%r1286, %r1287, %r1288, %r1289}, [%rd525+304]; st.local.v4.u32 [%rd525+160], {%r1254, %r1255, %r1256, %r1257}; st.local.v4.u32 [%rd525+176], {%r1262, %r1263, %r1264, %r1265}; st.local.v4.u32 [%rd525+192], {%r1270, %r1271, %r1272, %r1273}; st.local.v4.u32 [%rd525+208], {%r1278, %r1279, %r1280, %r1281}; st.local.v4.u32 [%rd525+224], {%r1286, %r1287, %r1288, %r1289}; add.s32 %r2191, %r2191, 4; ld.local.v4.u32 {%r1294, %r1295, %r1296, %r1297}, [%rd525+320]; ld.local.v4.u32 {%r1302, %r1303, %r1304, %r1305}, [%rd525+336]; ld.local.v4.u32 {%r1310, %r1311, %r1312, %r1313}, [%rd525+352]; ld.local.v4.u32 {%r1318, %r1319, %r1320, %r1321}, [%rd525+368]; ld.local.v4.u32 {%r1326, %r1327, %r1328, %r1329}, [%rd525+384]; st.local.v4.u32 [%rd525+240], {%r1294, %r1295, %r1296, %r1297}; st.local.v4.u32 [%rd525+256], {%r1302, %r1303, %r1304, %r1305}; st.local.v4.u32 [%rd525+272], {%r1310, %r1311, %r1312, %r1313}; st.local.v4.u32 [%rd525+288], {%r1318, %r1319, %r1320, %r1321}; st.local.v4.u32 [%rd525+304], {%r1326, %r1327, %r1328, %r1329}; add.s32 %r2190, %r2190, -4; setp.ne.s32 %p279, %r2190, 0; @%p279 bra $L__BB215_391; $L__BB215_392: setp.eq.s32 %p280, %r213, 0; @%p280 bra $L__BB215_396; mul.wide.s32 %rd526, %r2191, 80; add.s64 %rd90, %rd89, %rd526; ld.local.v4.u32 {%r1334, %r1335, %r1336, %r1337}, [%rd90+80]; ld.local.v4.u32 {%r1342, %r1343, %r1344, %r1345}, [%rd90+96]; ld.local.v4.u32 {%r1350, %r1351, %r1352, %r1353}, [%rd90+112]; ld.local.v4.u32 {%r1358, %r1359, %r1360, %r1361}, [%rd90+128]; ld.local.v4.u32 {%r1366, %r1367, %r1368, %r1369}, [%rd90+144]; st.local.v4.u32 [%rd90], {%r1334, %r1335, %r1336, %r1337}; st.local.v4.u32 [%rd90+16], {%r1342, %r1343, %r1344, %r1345}; st.local.v4.u32 [%rd90+32], {%r1350, %r1351, %r1352, %r1353}; st.local.v4.u32 [%rd90+48], {%r1358, %r1359, %r1360, %r1361}; st.local.v4.u32 [%rd90+64], {%r1366, %r1367, %r1368, %r1369}; setp.eq.s32 %p281, %r213, 1; @%p281 bra $L__BB215_396; ld.local.v4.u32 {%r1374, %r1375, %r1376, %r1377}, [%rd90+160]; ld.local.v4.u32 {%r1382, %r1383, %r1384, %r1385}, [%rd90+176]; ld.local.v4.u32 {%r1390, %r1391, %r1392, %r1393}, [%rd90+192]; ld.local.v4.u32 {%r1398, %r1399, %r1400, %r1401}, [%rd90+208]; ld.local.v4.u32 {%r1406, %r1407, %r1408, %r1409}, [%rd90+224]; st.local.v4.u32 [%rd90+80], {%r1374, %r1375, %r1376, %r1377}; st.local.v4.u32 [%rd90+96], {%r1382, %r1383, %r1384, %r1385}; st.local.v4.u32 [%rd90+112], {%r1390, %r1391, %r1392, %r1393}; st.local.v4.u32 [%rd90+128], {%r1398, %r1399, %r1400, %r1401}; st.local.v4.u32 [%rd90+144], {%r1406, %r1407, %r1408, %r1409}; setp.eq.s32 %p282, %r213, 2; @%p282 bra $L__BB215_396; ld.local.v4.u32 {%r1414, %r1415, %r1416, %r1417}, [%rd90+240]; ld.local.v4.u32 {%r1422, %r1423, %r1424, %r1425}, [%rd90+256]; ld.local.v4.u32 {%r1430, %r1431, %r1432, %r1433}, [%rd90+272]; ld.local.v4.u32 {%r1438, %r1439, %r1440, %r1441}, [%rd90+288]; ld.local.v4.u32 {%r1446, %r1447, %r1448, %r1449}, [%rd90+304]; st.local.v4.u32 [%rd90+160], {%r1414, %r1415, %r1416, %r1417}; st.local.v4.u32 [%rd90+176], {%r1422, %r1423, %r1424, %r1425}; st.local.v4.u32 [%rd90+192], {%r1430, %r1431, %r1432, %r1433}; st.local.v4.u32 [%rd90+208], {%r1438, %r1439, %r1440, %r1441}; st.local.v4.u32 [%rd90+224], {%r1446, %r1447, %r1448, %r1449}; $L__BB215_396: ld.local.u32 %r1454, [%rd59+4]; add.s32 %r2192, %r1454, -1; st.local.u32 [%rd59+4], %r2192; mov.u32 %r2194, %r2188; $L__BB215_397: add.s32 %r2187, %r2194, 1; setp.lt.s32 %p283, %r2187, %r2192; @%p283 bra $L__BB215_1098; $L__BB215_398: setp.lt.s32 %p284, %r206, %r2192; mov.u32 %r2188, %r206; @%p284 bra $L__BB215_384; $L__BB215_399: mov.u64 %rd1693, kernel_params; add.s32 %r2169, %r2169, 1; mul.wide.s32 %rd527, %r2169, 16; add.s64 %rd529, %rd1693, %rd527; add.s64 %rd1744, %rd529, 2768; ld.const.u64 %rd530, [%rd529+2776]; cvta.to.global.u64 %rd531, %rd530; add.s64 %rd533, %rd531, %rd260; ld.global.u32 %r2168, [%rd533]; setp.eq.s32 %p285, %r2168, -1; @%p285 bra $L__BB215_400; bra.uni $L__BB215_81; $L__BB215_400: ld.local.u32 %r231, [%rd23+36]; and.b32 %r1456, %r231, 194; setp.eq.s32 %p287, %r1456, 0; mov.pred %p798, %p286; @%p287 bra $L__BB215_427; and.b32 %r1457, %r231, 64; setp.eq.s32 %p288, %r1457, 0; mov.f32 %f3097, 0f00000000; mov.f32 %f3091, %f3097; mov.f32 %f3092, %f3097; mov.f32 %f3093, %f3097; @%p288 bra $L__BB215_403; ld.local.v2.f32 {%f3092, %f3093}, [%rd23+280]; ld.local.f32 %f3091, [%rd23+276]; $L__BB215_403: ld.local.v2.f32 {%f1815, %f1816}, [%rd23+264]; ld.local.f32 %f342, [%rd23+272]; and.b32 %r1458, %r231, 128; setp.eq.s32 %p289, %r1458, 0; setp.lt.s32 %p290, %r2192, 1; or.pred %p291, %p289, %p290; mov.f32 %f3098, %f3097; mov.f32 %f3099, %f3097; @%p291 bra $L__BB215_424; max.s32 %r232, %r2192, 1; add.s32 %r1460, %r232, -1; and.b32 %r233, %r232, 3; setp.lt.u32 %p292, %r1460, 3; mov.f32 %f3099, 0f00000000; mov.u32 %r2200, 0; mov.f32 %f3098, %f3099; mov.f32 %f3097, %f3099; @%p292 bra $L__BB215_415; sub.s32 %r2199, %r232, %r233; mov.f32 %f3099, 0f00000000; mov.u32 %r2200, 0; mov.f32 %f3098, %f3099; mov.f32 %f3097, %f3099; $L__BB215_406: cvt.s64.s32 %rd92, %r2200; mul.wide.s32 %rd534, %r2200, 80; add.s64 %rd535, %rd1, %rd534; add.s64 %rd93, %rd535, 316; ld.local.u32 %r1462, [%rd535+316]; add.s32 %r1463, %r1462, -36; setp.gt.u32 %p293, %r1463, 2; @%p293 bra $L__BB215_408; ld.local.v4.f32 {%f1824, %f1825, %f1826, %f1827}, [%rd93+-12]; add.ftz.f32 %f3099, %f3099, %f1824; add.ftz.f32 %f3098, %f3098, %f1825; add.ftz.f32 %f3097, %f3097, %f1826; $L__BB215_408: ld.local.u32 %r1464, [%rd93+80]; add.s32 %r1465, %r1464, -36; setp.gt.u32 %p294, %r1465, 2; @%p294 bra $L__BB215_410; ld.local.v4.f32 {%f1831, %f1832, %f1833, %f1834}, [%rd93+68]; add.ftz.f32 %f3099, %f3099, %f1831; add.ftz.f32 %f3098, %f3098, %f1832; add.ftz.f32 %f3097, %f3097, %f1833; $L__BB215_410: ld.local.u32 %r1466, [%rd93+160]; add.s32 %r1467, %r1466, -36; setp.gt.u32 %p295, %r1467, 2; @%p295 bra $L__BB215_412; ld.local.v4.f32 {%f1838, %f1839, %f1840, %f1841}, [%rd93+148]; add.ftz.f32 %f3099, %f3099, %f1838; add.ftz.f32 %f3098, %f3098, %f1839; add.ftz.f32 %f3097, %f3097, %f1840; $L__BB215_412: ld.local.u32 %r1468, [%rd93+240]; add.s32 %r1469, %r1468, -36; setp.gt.u32 %p296, %r1469, 2; @%p296 bra $L__BB215_414; ld.local.v4.f32 {%f1845, %f1846, %f1847, %f1848}, [%rd93+228]; add.ftz.f32 %f3099, %f3099, %f1845; add.ftz.f32 %f3098, %f3098, %f1846; add.ftz.f32 %f3097, %f3097, %f1847; $L__BB215_414: cvt.u32.u64 %r1470, %rd92; add.s32 %r2200, %r1470, 4; add.s32 %r2199, %r2199, -4; setp.ne.s32 %p297, %r2199, 0; @%p297 bra $L__BB215_406; $L__BB215_415: setp.eq.s32 %p298, %r233, 0; @%p298 bra $L__BB215_424; mul.wide.s32 %rd536, %r2200, 80; add.s64 %rd537, %rd1, %rd536; add.s64 %rd94, %rd537, 316; ld.local.u32 %r1471, [%rd537+316]; add.s32 %r1472, %r1471, -36; setp.gt.u32 %p299, %r1472, 2; @%p299 bra $L__BB215_418; ld.local.f32 %f1852, [%rd94+-12]; add.ftz.f32 %f3099, %f3099, %f1852; ld.local.f32 %f1853, [%rd94+-8]; add.ftz.f32 %f3098, %f3098, %f1853; ld.local.f32 %f1854, [%rd94+-4]; add.ftz.f32 %f3097, %f3097, %f1854; $L__BB215_418: setp.eq.s32 %p300, %r233, 1; @%p300 bra $L__BB215_424; ld.local.u32 %r1473, [%rd94+80]; add.s32 %r1474, %r1473, -36; setp.gt.u32 %p301, %r1474, 2; @%p301 bra $L__BB215_421; ld.local.f32 %f1855, [%rd94+68]; add.ftz.f32 %f3099, %f3099, %f1855; ld.local.f32 %f1856, [%rd94+72]; add.ftz.f32 %f3098, %f3098, %f1856; ld.local.f32 %f1857, [%rd94+76]; add.ftz.f32 %f3097, %f3097, %f1857; $L__BB215_421: setp.eq.s32 %p302, %r233, 2; @%p302 bra $L__BB215_424; ld.local.u32 %r1475, [%rd94+160]; add.s32 %r1476, %r1475, -36; setp.gt.u32 %p303, %r1476, 2; @%p303 bra $L__BB215_424; ld.local.f32 %f1858, [%rd94+148]; add.ftz.f32 %f3099, %f3099, %f1858; ld.local.f32 %f1859, [%rd94+152]; add.ftz.f32 %f3098, %f3098, %f1859; ld.local.f32 %f1860, [%rd94+156]; add.ftz.f32 %f3097, %f3097, %f1860; $L__BB215_424: ld.local.u32 %r240, [%rd23+60]; setp.eq.s32 %p304, %r240, -1; mov.f32 %f3124, 0f3F800000; @%p304 bra $L__BB215_426; mul.wide.s32 %rd538, %r240, 208; add.s64 %rd539, %rd54, %rd538; ld.global.f32 %f3124, [%rd539+96]; $L__BB215_426: and.b32 %r1477, %r231, 2; setp.eq.s32 %p306, %r1477, 0; selp.f32 %f1862, 0f00000000, %f342, %p306; selp.f32 %f1863, 0f00000000, %f1816, %p306; selp.f32 %f1864, 0f00000000, %f1815, %p306; mul.ftz.f32 %f3130, %f3099, %f3124; mul.ftz.f32 %f3129, %f3098, %f3124; mul.ftz.f32 %f3128, %f3097, %f3124; mul.ftz.f32 %f3133, %f3091, %f3124; mul.ftz.f32 %f3132, %f3092, %f3124; mul.ftz.f32 %f3131, %f3093, %f3124; mul.ftz.f32 %f3127, %f1864, %f3124; mul.ftz.f32 %f3126, %f1863, %f3124; mul.ftz.f32 %f3125, %f1862, %f3124; mov.pred %p798, 0; $L__BB215_427: @%p798 bra $L__BB215_500; and.b32 %r241, %r231, 64; setp.eq.s32 %p307, %r241, 0; mov.f32 %f420, 0f3F800000; mov.f32 %f421, %f420; mov.f32 %f422, %f420; @%p307 bra $L__BB215_430; mul.ftz.f32 %f1868, %f3141, %f3133; mul.ftz.f32 %f1869, %f1868, 0fBFB8AA3B; ex2.approx.ftz.f32 %f422, %f1869; mul.ftz.f32 %f1870, %f3141, %f3132; mul.ftz.f32 %f1871, %f1870, 0fBFB8AA3B; ex2.approx.ftz.f32 %f421, %f1871; mul.ftz.f32 %f1872, %f3141, %f3131; mul.ftz.f32 %f1873, %f1872, 0fBFB8AA3B; ex2.approx.ftz.f32 %f420, %f1873; $L__BB215_430: and.b32 %r1478, %r231, 2; setp.eq.s32 %p308, %r1478, 0; ld.local.u8 %rs114, [%rd49+180]; setp.ne.s16 %p309, %rs114, 0; or.pred %p310, %p308, %p309; @%p310 bra $L__BB215_439; mov.f32 %f3139, %f3141; mov.f32 %f3140, %f3141; @%p307 bra $L__BB215_438; setp.leu.ftz.f32 %p312, %f3133, 0f00000000; mov.f32 %f3139, %f3141; @%p312 bra $L__BB215_434; mov.f32 %f1874, 0f3F800000; sub.ftz.f32 %f1875, %f1874, %f422; div.approx.ftz.f32 %f3139, %f1875, %f3133; $L__BB215_434: setp.leu.ftz.f32 %p313, %f3132, 0f00000000; mov.f32 %f3140, %f3141; @%p313 bra $L__BB215_436; mov.f32 %f1876, 0f3F800000; sub.ftz.f32 %f1877, %f1876, %f421; div.approx.ftz.f32 %f3140, %f1877, %f3132; $L__BB215_436: setp.leu.ftz.f32 %p314, %f3131, 0f00000000; @%p314 bra $L__BB215_438; mov.f32 %f1878, 0f3F800000; sub.ftz.f32 %f1879, %f1878, %f420; div.approx.ftz.f32 %f3141, %f1879, %f3131; $L__BB215_438: ld.local.f32 %f1880, [%rd49+184]; mul.ftz.f32 %f1881, %f3139, %f3127; ld.local.f32 %f1882, [%rd49+188]; mul.ftz.f32 %f1883, %f3140, %f3126; ld.local.f32 %f1884, [%rd49+192]; mul.ftz.f32 %f1885, %f3141, %f3125; fma.rn.ftz.f32 %f3183, %f1881, %f1880, %f3183; fma.rn.ftz.f32 %f3182, %f1883, %f1882, %f3182; fma.rn.ftz.f32 %f3181, %f1885, %f1884, %f3181; $L__BB215_439: @%p307 bra $L__BB215_499; and.b32 %r1479, %r231, 128; setp.eq.s32 %p316, %r1479, 0; @%p316 bra $L__BB215_495; bra.uni $L__BB215_441; $L__BB215_495: ld.local.f32 %f2103, [%rd49+184]; mul.ftz.f32 %f2104, %f422, %f2103; ld.local.f32 %f2105, [%rd49+188]; mul.ftz.f32 %f2106, %f421, %f2105; ld.local.f32 %f2107, [%rd49+192]; mul.ftz.f32 %f2108, %f420, %f2107; st.local.f32 [%rd49+184], %f2104; st.local.f32 [%rd49+188], %f2106; st.local.f32 [%rd49+192], %f2108; ld.local.f32 %f2109, [%rd49]; mul.ftz.f32 %f2110, %f422, %f2109; ld.local.f32 %f2111, [%rd49+4]; mul.ftz.f32 %f2112, %f421, %f2111; ld.local.f32 %f2113, [%rd49+8]; mul.ftz.f32 %f2114, %f420, %f2113; st.local.f32 [%rd49], %f2110; st.local.f32 [%rd49+4], %f2112; st.local.f32 [%rd49+8], %f2114; bra.uni $L__BB215_496; $L__BB215_441: setp.eq.ftz.f32 %p317, %f3133, 0f00000000; mov.f32 %f3147, 0f00000000; mov.f32 %f3145, %f3147; @%p317 bra $L__BB215_443; div.approx.ftz.f32 %f3145, %f3130, %f3133; $L__BB215_443: setp.eq.ftz.f32 %p318, %f3132, 0f00000000; mov.f32 %f3146, %f3147; @%p318 bra $L__BB215_445; div.approx.ftz.f32 %f3146, %f3129, %f3132; $L__BB215_445: setp.eq.ftz.f32 %p319, %f3131, 0f00000000; @%p319 bra $L__BB215_447; div.approx.ftz.f32 %f3147, %f3128, %f3131; $L__BB215_447: @%p320 bra $L__BB215_449; ld.local.f32 %f1889, [%rd49+184]; mul.ftz.f32 %f1890, %f3145, %f1889; ld.local.f32 %f1891, [%rd49+188]; mul.ftz.f32 %f1892, %f3146, %f1891; ld.local.f32 %f1893, [%rd49+192]; mul.ftz.f32 %f1894, %f3147, %f1893; mov.f32 %f1895, 0f3F800000; sub.ftz.f32 %f1896, %f1895, %f422; sub.ftz.f32 %f1897, %f1895, %f421; sub.ftz.f32 %f1898, %f1895, %f420; fma.rn.ftz.f32 %f3169, %f1896, %f1890, %f3169; fma.rn.ftz.f32 %f3168, %f1897, %f1892, %f3168; fma.rn.ftz.f32 %f3167, %f1898, %f1894, %f3167; $L__BB215_449: ld.local.f32 %f1902, [%rd49+184]; mul.ftz.f32 %f1903, %f3145, %f1902; ld.local.f32 %f1904, [%rd49+188]; mul.ftz.f32 %f1905, %f3146, %f1904; ld.local.f32 %f1906, [%rd49+192]; mul.ftz.f32 %f1907, %f3147, %f1906; abs.ftz.f32 %f449, %f1903; abs.ftz.f32 %f450, %f1905; abs.ftz.f32 %f451, %f1907; add.ftz.f32 %f1908, %f449, %f450; add.ftz.f32 %f452, %f451, %f1908; setp.leu.ftz.f32 %p321, %f452, 0f00000000; mov.f32 %f3151, 0f3EAAAAAB; mov.f32 %f3152, %f3151; mov.f32 %f3153, %f3151; @%p321 bra $L__BB215_451; div.approx.ftz.f32 %f3151, %f449, %f452; div.approx.ftz.f32 %f3152, %f450, %f452; div.approx.ftz.f32 %f3153, %f451, %f452; $L__BB215_451: add.ftz.f32 %f459, %f3151, 0f00000000; setp.gt.ftz.f32 %p322, %f459, %f3023; mov.u64 %rd1746, 0; @%p322 bra $L__BB215_454; add.ftz.f32 %f1909, %f459, %f3152; setp.gt.ftz.f32 %p323, %f1909, %f3023; mov.u64 %rd1746, 1; @%p323 bra $L__BB215_454; mov.u64 %rd1746, 2; $L__BB215_454: ld.local.u8 %rs45, [%rd49+-4]; setp.ne.s16 %p324, %rs45, 0; or.pred %p326, %p52, %p324; @%p326 bra $L__BB215_464; ld.local.f32 %f460, [%rd49+12]; setp.ge.ftz.f32 %p328, %f460, %f3045; setp.le.ftz.f32 %p329, %f460, %f109; and.pred %p330, %p329, %p328; and.pred %p331, %p327, %p330; @%p331 bra $L__BB215_457; bra.uni $L__BB215_456; $L__BB215_457: sub.ftz.f32 %f1919, %f460, %f3045; mul.ftz.f32 %f1920, %f3133, %f1919; mul.ftz.f32 %f1921, %f3132, %f1919; mul.ftz.f32 %f1922, %f3131, %f1919; mul.ftz.f32 %f1923, %f1920, 0fBFB8AA3B; ex2.approx.ftz.f32 %f462, %f1923; mul.ftz.f32 %f1924, %f1921, 0fBFB8AA3B; ex2.approx.ftz.f32 %f463, %f1924; mul.ftz.f32 %f1925, %f1922, 0fBFB8AA3B; ex2.approx.ftz.f32 %f464, %f1925; st.local.u8 [%rd49+-4], %rs46; mul.ftz.f32 %f1926, %f3130, %f462; mul.ftz.f32 %f1927, %f3129, %f463; mul.ftz.f32 %f1928, %f3128, %f464; div.approx.ftz.f32 %f1929, %f1926, %f3029; div.approx.ftz.f32 %f1930, %f1927, %f3029; div.approx.ftz.f32 %f1931, %f1928, %f3029; ld.local.f32 %f1932, [%rd49]; mul.ftz.f32 %f1933, %f1929, %f1932; ld.local.f32 %f1934, [%rd49+4]; mul.ftz.f32 %f1935, %f1930, %f1934; ld.local.f32 %f1936, [%rd49+8]; mul.ftz.f32 %f1937, %f1931, %f1936; st.local.f32 [%rd49], %f1933; st.local.f32 [%rd49+4], %f1935; st.local.f32 [%rd49+8], %f1937; mov.u32 %r2203, 0; st.local.u32 [%rd49+176], %r2203; setp.lt.s32 %p332, %r2192, 1; @%p332 bra $L__BB215_462; mov.u32 %r2202, %r2203; $L__BB215_459: cvt.s64.s32 %rd96, %r2202; mul.wide.s32 %rd543, %r2202, 80; add.s64 %rd544, %rd1, %rd543; add.s64 %rd97, %rd544, 316; ld.local.u32 %r1483, [%rd544+316]; setp.ne.s32 %p333, %r1483, 38; @%p333 bra $L__BB215_461; ld.local.v4.f32 {%f1938, %f1939, %f1940, %f1941}, [%rd97+-12]; mul.wide.s32 %rd547, %r2203, 20; add.s64 %rd548, %rd42, %rd547; st.local.f32 [%rd548+20], %f1938; st.local.f32 [%rd548+24], %f1939; st.local.f32 [%rd548+28], %f1940; ld.local.f32 %f1945, [%rd97+4]; st.local.f32 [%rd548+32], %f1945; ld.local.f32 %f1946, [%rd97+20]; st.local.f32 [%rd548+36], %f1946; ld.local.u32 %r1484, [%rd49+176]; add.s32 %r2203, %r1484, 1; st.local.u32 [%rd49+176], %r2203; setp.gt.s32 %p334, %r1484, 6; @%p334 bra $L__BB215_462; $L__BB215_461: cvt.u32.u64 %r1485, %rd96; add.s32 %r2202, %r1485, 1; setp.lt.s32 %p335, %r2202, %r2192; @%p335 bra $L__BB215_459; $L__BB215_462: @%p50 bra $L__BB215_464; mul.ftz.f32 %f1947, %f3133, %f462; mul.ftz.f32 %f1948, %f3132, %f463; mul.ftz.f32 %f1949, %f1948, %f3152; fma.rn.ftz.f32 %f1950, %f1947, %f3151, %f1949; mul.ftz.f32 %f1951, %f3131, %f464; fma.rn.ftz.f32 %f1952, %f1951, %f3153, %f1950; mul.ftz.f32 %f1953, %f3158, %f1952; fma.rn.ftz.f32 %f1954, %f1953, %f1953, %f87; div.approx.ftz.f32 %f1955, %f87, %f1954; add.ftz.f32 %f1956, %f1955, %f1955; ld.local.f32 %f1957, [%rd49]; mul.ftz.f32 %f1958, %f1957, %f1956; ld.local.f32 %f1959, [%rd49+4]; mul.ftz.f32 %f1960, %f1956, %f1959; ld.local.f32 %f1961, [%rd49+8]; mul.ftz.f32 %f1962, %f1956, %f1961; st.local.f32 [%rd49], %f1958; st.local.f32 [%rd49+4], %f1960; st.local.f32 [%rd49+8], %f1962; bra.uni $L__BB215_464; $L__BB215_456: ld.local.f32 %f1910, [%rd49]; mul.ftz.f32 %f1911, %f422, %f1910; ld.local.f32 %f1912, [%rd49+4]; mul.ftz.f32 %f1913, %f421, %f1912; ld.local.f32 %f1914, [%rd49+8]; mul.ftz.f32 %f1915, %f420, %f1914; st.local.f32 [%rd49], %f1911; st.local.f32 [%rd49+4], %f1913; st.local.f32 [%rd49+8], %f1915; mul.ftz.f32 %f1916, %f421, %f3152; fma.rn.ftz.f32 %f1917, %f422, %f3151, %f1916; fma.rn.ftz.f32 %f1918, %f420, %f3153, %f1917; mul.ftz.f32 %f3158, %f3158, %f1918; $L__BB215_464: ld.local.u8 %rs47, [%rd49+180]; setp.ne.s16 %p337, %rs47, 0; @%p337 bra $L__BB215_494; add.u64 %rd550, %SPL, 304; st.local.f32 [%rd550], %f422; st.local.f32 [%rd550+4], %f421; st.local.f32 [%rd550+8], %f420; shl.b64 %rd551, %rd1746, 2; add.s64 %rd552, %rd550, %rd551; ld.local.f32 %f466, [%rd552]; mov.f32 %f1963, 0f3F800000; sub.ftz.f32 %f467, %f1963, %f3166; setp.ltu.ftz.f32 %p338, %f467, %f466; @%p338 bra $L__BB215_493; bra.uni $L__BB215_466; $L__BB215_493: setp.eq.s32 %p372, %r2161, 2; mul.ftz.f32 %f2088, %f421, %f3152; fma.rn.ftz.f32 %f2089, %f422, %f3151, %f2088; fma.rn.ftz.f32 %f2090, %f420, %f3153, %f2089; div.approx.ftz.f32 %f2091, %f422, %f2090; div.approx.ftz.f32 %f2092, %f421, %f2090; div.approx.ftz.f32 %f2093, %f420, %f2090; ld.local.f32 %f2094, [%rd49+184]; mul.ftz.f32 %f2095, %f2091, %f2094; ld.local.f32 %f2096, [%rd49+188]; mul.ftz.f32 %f2097, %f2092, %f2096; ld.local.f32 %f2098, [%rd49+192]; mul.ftz.f32 %f2099, %f2093, %f2098; st.local.f32 [%rd49+184], %f2095; st.local.f32 [%rd49+188], %f2097; st.local.f32 [%rd49+192], %f2099; mul.ftz.f32 %f2100, %f3158, %f2090; selp.f32 %f3158, %f3158, %f2100, %p372; div.approx.ftz.f32 %f2101, %f467, %f466; sub.ftz.f32 %f3166, %f1963, %f2101; bra.uni $L__BB215_494; $L__BB215_466: st.local.f32 [%rd550], %f3133; st.local.f32 [%rd550+4], %f3132; st.local.f32 [%rd550+8], %f3131; ld.local.f32 %f1964, [%rd552]; lg2.approx.ftz.f32 %f1965, %f467; mul.ftz.f32 %f1966, %f1965, 0fBF317218; div.approx.ftz.f32 %f1967, %f1966, %f1964; add.ftz.f32 %f468, %f3045, %f1967; mul.ftz.f32 %f1968, %f3133, %f1967; mul.ftz.f32 %f1969, %f3132, %f1967; mul.ftz.f32 %f1970, %f3131, %f1967; mul.ftz.f32 %f1971, %f1968, 0fBFB8AA3B; ex2.approx.ftz.f32 %f469, %f1971; mul.ftz.f32 %f1972, %f1969, 0fBFB8AA3B; ex2.approx.ftz.f32 %f470, %f1972; mul.ftz.f32 %f1973, %f1970, 0fBFB8AA3B; ex2.approx.ftz.f32 %f471, %f1973; mul.ftz.f32 %f1974, %f3133, %f469; mul.ftz.f32 %f1975, %f3132, %f470; mul.ftz.f32 %f1976, %f3131, %f471; mul.ftz.f32 %f1977, %f1975, %f3152; fma.rn.ftz.f32 %f1978, %f1974, %f3151, %f1977; fma.rn.ftz.f32 %f472, %f1976, %f3153, %f1978; mul.ftz.f32 %f473, %f3158, %f472; setp.leu.ftz.f32 %p339, %f473, 0f322BCC77; @%p339 bra $L__BB215_494; st.local.u8 [%rd49+180], %rs46; st.local.f32 [%rd49+196], %f468; mul.ftz.f32 %f1979, %f3130, %f469; div.approx.ftz.f32 %f1980, %f1979, %f472; mul.ftz.f32 %f1981, %f3129, %f470; div.approx.ftz.f32 %f1982, %f1981, %f472; mul.ftz.f32 %f1983, %f3128, %f471; div.approx.ftz.f32 %f1984, %f1983, %f472; ld.local.f32 %f1985, [%rd49+184]; mul.ftz.f32 %f1986, %f1980, %f1985; ld.local.f32 %f1987, [%rd49+188]; mul.ftz.f32 %f1988, %f1982, %f1987; ld.local.f32 %f1989, [%rd49+192]; mul.ftz.f32 %f1990, %f1984, %f1989; st.local.f32 [%rd49+184], %f1986; st.local.f32 [%rd49+188], %f1988; st.local.f32 [%rd49+192], %f1990; mov.u32 %r2209, 0; st.local.u32 [%rd49+360], %r2209; setp.lt.s32 %p340, %r2192, 1; @%p340 bra $L__BB215_472; mov.u32 %r2206, 0; mov.u32 %r2205, %r2206; $L__BB215_469: cvt.s64.s32 %rd98, %r2205; mul.wide.s32 %rd557, %r2205, 80; add.s64 %rd558, %rd1, %rd557; add.s64 %rd99, %rd558, 316; ld.local.u32 %r1489, [%rd558+316]; setp.ne.s32 %p341, %r1489, 38; @%p341 bra $L__BB215_471; ld.local.v4.f32 {%f1991, %f1992, %f1993, %f1994}, [%rd99+-12]; mul.wide.s32 %rd561, %r2206, 20; add.s64 %rd562, %rd42, %rd561; st.local.f32 [%rd562+204], %f1991; st.local.f32 [%rd562+208], %f1992; st.local.f32 [%rd562+212], %f1993; ld.local.f32 %f1998, [%rd99+4]; st.local.f32 [%rd562+216], %f1998; ld.local.f32 %f1999, [%rd99+20]; st.local.f32 [%rd562+220], %f1999; ld.local.u32 %r1490, [%rd49+360]; add.s32 %r2206, %r1490, 1; st.local.u32 [%rd49+360], %r2206; setp.gt.s32 %p342, %r1490, 6; @%p342 bra $L__BB215_472; $L__BB215_471: cvt.u32.u64 %r1491, %rd98; add.s32 %r2205, %r1491, 1; setp.lt.s32 %p343, %r2205, %r2192; @%p343 bra $L__BB215_469; $L__BB215_472: setp.eq.s32 %p344, %r2161, 2; @%p344 bra $L__BB215_494; st.local.u8 [%rd49+-4], %rs46; ld.local.f32 %f2000, [%rd49+196]; st.local.f32 [%rd49+12], %f2000; ld.local.f32 %f2001, [%rd49+184]; ld.local.f32 %f2002, [%rd49+188]; ld.local.f32 %f2003, [%rd49+192]; st.local.f32 [%rd49], %f2001; st.local.f32 [%rd49+4], %f2002; st.local.f32 [%rd49+8], %f2003; st.local.u32 [%rd49+176], %r2209; @%p340 bra $L__BB215_478; mov.u32 %r2208, %r2209; $L__BB215_475: cvt.s64.s32 %rd100, %r2208; mul.wide.s32 %rd563, %r2208, 80; add.s64 %rd564, %rd1, %rd563; add.s64 %rd101, %rd564, 316; ld.local.u32 %r1495, [%rd564+316]; setp.ne.s32 %p346, %r1495, 38; @%p346 bra $L__BB215_477; ld.local.v4.f32 {%f2004, %f2005, %f2006, %f2007}, [%rd101+-12]; mul.wide.s32 %rd567, %r2209, 20; add.s64 %rd568, %rd42, %rd567; st.local.f32 [%rd568+20], %f2004; st.local.f32 [%rd568+24], %f2005; st.local.f32 [%rd568+28], %f2006; ld.local.f32 %f2011, [%rd101+4]; st.local.f32 [%rd568+32], %f2011; ld.local.f32 %f2012, [%rd101+20]; st.local.f32 [%rd568+36], %f2012; ld.local.u32 %r1496, [%rd49+176]; add.s32 %r2209, %r1496, 1; st.local.u32 [%rd49+176], %r2209; setp.gt.s32 %p347, %r1496, 6; @%p347 bra $L__BB215_478; $L__BB215_477: cvt.u32.u64 %r1497, %rd100; add.s32 %r2208, %r1497, 1; setp.lt.s32 %p348, %r2208, %r2192; @%p348 bra $L__BB215_475; $L__BB215_478: @%p50 bra $L__BB215_494; mov.f32 %f3157, 0f00000000; max.ftz.f32 %f2014, %f89, %f3157; sqrt.approx.ftz.f32 %f474, %f2014; setp.eq.ftz.f32 %p350, %f474, 0f00000000; @%p350 bra $L__BB215_492; sub.ftz.f32 %f2015, %f7, %f88; abs.ftz.f32 %f475, %f2015; abs.ftz.f32 %f476, %f474; setp.eq.ftz.f32 %p351, %f476, 0f00000000; setp.eq.ftz.f32 %p352, %f475, 0f00000000; and.pred %p353, %p351, %p352; mov.b32 %r257, %f474; @%p353 bra $L__BB215_484; bra.uni $L__BB215_481; $L__BB215_484: shr.s32 %r1506, %r257, 31; and.b32 %r1507, %r1506, 1078530011; mov.b32 %r1508, %f2015; and.b32 %r1509, %r1508, -2147483648; or.b32 %r1510, %r1507, %r1509; mov.b32 %f3155, %r1510; bra.uni $L__BB215_485; $L__BB215_481: setp.eq.ftz.f32 %p354, %f476, 0f7F800000; setp.eq.ftz.f32 %p355, %f475, 0f7F800000; and.pred %p356, %p354, %p355; @%p356 bra $L__BB215_483; bra.uni $L__BB215_482; $L__BB215_483: setp.lt.s32 %p360, %r257, 0; selp.b32 %r1502, 1075235812, 1061752795, %p360; mov.b32 %r1503, %f2015; and.b32 %r1504, %r1503, -2147483648; or.b32 %r1505, %r1502, %r1504; mov.b32 %f3155, %r1505; bra.uni $L__BB215_485; $L__BB215_482: setp.lt.s32 %p357, %r257, 0; min.ftz.f32 %f2016, %f475, %f476; max.ftz.f32 %f2017, %f475, %f476; div.full.ftz.f32 %f2018, %f2016, %f2017; mul.rn.ftz.f32 %f2019, %f2018, %f2018; mov.f32 %f2020, 0fC0B59883; mov.f32 %f2021, 0fBF52C7EA; fma.rn.ftz.f32 %f2022, %f2019, %f2021, %f2020; mov.f32 %f2023, 0fC0D21907; fma.rn.ftz.f32 %f2024, %f2022, %f2019, %f2023; mul.ftz.f32 %f2025, %f2019, %f2024; mul.ftz.f32 %f2026, %f2018, %f2025; add.ftz.f32 %f2027, %f2019, 0f41355DC0; mov.f32 %f2028, 0f41E6BD60; fma.rn.ftz.f32 %f2029, %f2027, %f2019, %f2028; mov.f32 %f2030, 0f419D92C8; fma.rn.ftz.f32 %f2031, %f2029, %f2019, %f2030; rcp.approx.ftz.f32 %f2032, %f2031; fma.rn.ftz.f32 %f2033, %f2026, %f2032, %f2018; mov.f32 %f2034, 0f3FC90FDB; sub.ftz.f32 %f2035, %f2034, %f2033; setp.gt.ftz.f32 %p358, %f475, %f476; selp.f32 %f2036, %f2035, %f2033, %p358; mov.f32 %f2037, 0f40490FDB; sub.ftz.f32 %f2038, %f2037, %f2036; selp.f32 %f2039, %f2038, %f2036, %p357; mov.b32 %r1498, %f2039; mov.b32 %r1499, %f2015; and.b32 %r1500, %r1499, -2147483648; or.b32 %r1501, %r1500, %r1498; mov.b32 %f2041, %r1501; add.ftz.f32 %f2042, %f476, %f475; setp.le.ftz.f32 %p359, %f2042, 0f7F800000; selp.f32 %f3155, %f2041, %f2042, %p359; $L__BB215_485: abs.ftz.f32 %f481, %f90; setp.eq.ftz.f32 %p361, %f481, 0f00000000; and.pred %p363, %p351, %p361; @%p363 bra $L__BB215_489; bra.uni $L__BB215_486; $L__BB215_489: shr.s32 %r1519, %r257, 31; and.b32 %r1520, %r1519, 1078530011; mov.b32 %r1521, %f90; and.b32 %r1522, %r1521, -2147483648; or.b32 %r1523, %r1520, %r1522; mov.b32 %f3156, %r1523; bra.uni $L__BB215_490; $L__BB215_486: setp.eq.ftz.f32 %p364, %f476, 0f7F800000; setp.eq.ftz.f32 %p365, %f481, 0f7F800000; and.pred %p366, %p364, %p365; @%p366 bra $L__BB215_488; bra.uni $L__BB215_487; $L__BB215_488: setp.lt.s32 %p370, %r257, 0; selp.b32 %r1515, 1075235812, 1061752795, %p370; mov.b32 %r1516, %f90; and.b32 %r1517, %r1516, -2147483648; or.b32 %r1518, %r1515, %r1517; mov.b32 %f3156, %r1518; bra.uni $L__BB215_490; $L__BB215_487: setp.lt.s32 %p367, %r257, 0; min.ftz.f32 %f2046, %f481, %f476; max.ftz.f32 %f2047, %f481, %f476; div.full.ftz.f32 %f2048, %f2046, %f2047; mul.rn.ftz.f32 %f2049, %f2048, %f2048; mov.f32 %f2050, 0fC0B59883; mov.f32 %f2051, 0fBF52C7EA; fma.rn.ftz.f32 %f2052, %f2049, %f2051, %f2050; mov.f32 %f2053, 0fC0D21907; fma.rn.ftz.f32 %f2054, %f2052, %f2049, %f2053; mul.ftz.f32 %f2055, %f2049, %f2054; mul.ftz.f32 %f2056, %f2048, %f2055; add.ftz.f32 %f2057, %f2049, 0f41355DC0; mov.f32 %f2058, 0f41E6BD60; fma.rn.ftz.f32 %f2059, %f2057, %f2049, %f2058; mov.f32 %f2060, 0f419D92C8; fma.rn.ftz.f32 %f2061, %f2059, %f2049, %f2060; rcp.approx.ftz.f32 %f2062, %f2061; fma.rn.ftz.f32 %f2063, %f2056, %f2062, %f2048; mov.f32 %f2064, 0f3FC90FDB; sub.ftz.f32 %f2065, %f2064, %f2063; setp.gt.ftz.f32 %p368, %f481, %f476; selp.f32 %f2066, %f2065, %f2063, %p368; mov.f32 %f2067, 0f40490FDB; sub.ftz.f32 %f2068, %f2067, %f2066; selp.f32 %f2069, %f2068, %f2066, %p367; mov.b32 %r1511, %f2069; mov.b32 %r1512, %f90; and.b32 %r1513, %r1512, -2147483648; or.b32 %r1514, %r1513, %r1511; mov.b32 %f2070, %r1514; add.ftz.f32 %f2071, %f476, %f481; setp.le.ftz.f32 %p369, %f2071, 0f7F800000; selp.f32 %f3156, %f2070, %f2071, %p369; $L__BB215_490: setp.eq.ftz.f32 %p371, %f3156, %f3155; @%p371 bra $L__BB215_492; sub.ftz.f32 %f2073, %f468, %f88; sub.ftz.f32 %f2074, %f3156, %f3155; mul.ftz.f32 %f2075, %f474, %f474; fma.rn.ftz.f32 %f2076, %f2073, %f2073, %f2075; mul.ftz.f32 %f2077, %f2076, %f2074; div.approx.ftz.f32 %f3157, %f474, %f2077; $L__BB215_492: mul.ftz.f32 %f2078, %f473, %f473; fma.rn.ftz.f32 %f2079, %f3157, %f3157, %f2078; div.approx.ftz.f32 %f2080, %f2078, %f2079; add.ftz.f32 %f2081, %f2080, %f2080; ld.local.f32 %f2082, [%rd49]; mul.ftz.f32 %f2083, %f2082, %f2081; ld.local.f32 %f2084, [%rd49+4]; mul.ftz.f32 %f2085, %f2081, %f2084; ld.local.f32 %f2086, [%rd49+8]; mul.ftz.f32 %f2087, %f2081, %f2086; st.local.f32 [%rd49], %f2083; st.local.f32 [%rd49+4], %f2085; st.local.f32 [%rd49+8], %f2087; $L__BB215_494: ld.local.u8 %rs114, [%rd49+180]; $L__BB215_496: setp.eq.s16 %p373, %rs114, 0; @%p373 bra $L__BB215_501; ld.local.u8 %rs51, [%rd49+-4]; setp.ne.s16 %p374, %rs51, 0; @%p374 bra $L__BB215_499; ld.local.f32 %f2115, [%rd49]; ld.local.f32 %f2116, [%rd49+4]; max.ftz.f32 %f2117, %f2115, %f2116; ld.local.f32 %f2118, [%rd49+8]; max.ftz.f32 %f2119, %f2117, %f2118; setp.lt.ftz.f32 %p375, %f2119, 0f358637BD; @%p375 bra $L__BB215_503; bra.uni $L__BB215_499; $L__BB215_501: ld.local.f32 %f2120, [%rd49+188]; ld.local.f32 %f2121, [%rd49+184]; max.ftz.f32 %f2122, %f2121, %f2120; ld.local.f32 %f2123, [%rd49+192]; max.ftz.f32 %f2124, %f2122, %f2123; setp.geu.ftz.f32 %p376, %f2124, 0f358637BD; mov.u16 %rs114, %rs52; @%p376 bra $L__BB215_499; bra.uni $L__BB215_502; $L__BB215_499: ld.local.u8 %rs53, [%rd49+-4]; setp.ne.s16 %p377, %rs53, 0; and.b16 %rs54, %rs114, 255; setp.ne.s16 %p378, %rs54, 0; and.pred %p379, %p377, %p378; @%p379 bra $L__BB215_503; $L__BB215_500: add.s32 %r2167, %r2167, 1; setp.lt.s32 %p380, %r2167, %r2165; setp.neu.ftz.f32 %p381, %f109, %f9; and.pred %p382, %p381, %p380; mov.f32 %f3045, %f109; @%p382 bra $L__BB215_79; $L__BB215_503: setp.eq.ftz.f32 %p383, %f3182, 0f00000000; setp.eq.ftz.f32 %p384, %f3183, 0f00000000; and.pred %p385, %p383, %p384; setp.eq.ftz.f32 %p386, %f3181, 0f00000000; and.pred %p387, %p386, %p385; @%p387 bra $L__BB215_548; mov.u32 %r2210, -1; ld.local.u32 %r259, [%rd23+60]; setp.eq.s32 %p388, %r259, -1; @%p388 bra $L__BB215_506; ld.const.u64 %rd569, [kernel_params+2160]; cvta.to.global.u64 %rd570, %rd569; mul.wide.s32 %rd571, %r259, 208; add.s64 %rd572, %rd570, %rd571; ld.global.u32 %r2210, [%rd572+192]; $L__BB215_506: ld.const.u64 %rd573, [kernel_params+2464]; cvta.to.global.u64 %rd574, %rd573; add.s64 %rd576, %rd574, %rd263; ld.global.u16 %rs17, [%rd576]; mov.b32 %r262, %f3183; setp.nan.ftz.f32 %p389, %f3183, %f3183; mov.f32 %f3192, 0f00000000; mov.f32 %f3190, %f3192; @%p389 bra $L__BB215_509; or.b32 %r1526, %r262, -2147483648; setp.ne.s32 %p390, %r1526, -2147483648; add.ftz.f32 %f2127, %f3183, %f3183; setp.eq.ftz.f32 %p391, %f2127, %f3183; and.pred %p392, %p391, %p390; mov.f32 %f3190, %f3192; @%p392 bra $L__BB215_509; shl.b32 %r1527, %r262, 1; setp.lt.u32 %p393, %r1527, -16777215; selp.f32 %f3190, %f3183, 0f00000000, %p393; $L__BB215_509: mov.b32 %r263, %f3182; setp.nan.ftz.f32 %p394, %f3182, %f3182; mov.f32 %f3191, %f3192; @%p394 bra $L__BB215_512; or.b32 %r1528, %r263, -2147483648; setp.ne.s32 %p395, %r1528, -2147483648; add.ftz.f32 %f2130, %f3182, %f3182; setp.eq.ftz.f32 %p396, %f2130, %f3182; and.pred %p397, %p396, %p395; mov.f32 %f3191, %f3192; @%p397 bra $L__BB215_512; shl.b32 %r1529, %r263, 1; setp.lt.u32 %p398, %r1529, -16777215; selp.f32 %f3191, %f3182, 0f00000000, %p398; $L__BB215_512: mov.b32 %r264, %f3181; setp.nan.ftz.f32 %p399, %f3181, %f3181; @%p399 bra $L__BB215_515; or.b32 %r1530, %r264, -2147483648; setp.ne.s32 %p400, %r1530, -2147483648; add.ftz.f32 %f2133, %f3181, %f3181; setp.eq.ftz.f32 %p401, %f2133, %f3181; and.pred %p402, %p401, %p400; @%p402 bra $L__BB215_515; shl.b32 %r1531, %r264, 1; setp.lt.u32 %p403, %r1531, -16777215; selp.f32 %f3192, %f3181, 0f00000000, %p403; $L__BB215_515: ld.const.f32 %f2134, [kernel_params+1580]; ld.const.f32 %f2135, [kernel_params+1584]; setp.gt.u16 %p404, %rs17, 1; selp.f32 %f522, %f2135, %f2134, %p404; abs.ftz.f32 %f2136, %f3191; abs.ftz.f32 %f2137, %f3190; add.ftz.f32 %f2138, %f2137, %f2136; abs.ftz.f32 %f2139, %f3192; add.ftz.f32 %f523, %f2139, %f2138; setp.leu.ftz.f32 %p405, %f523, %f522; @%p405 bra $L__BB215_517; div.approx.ftz.f32 %f2140, %f522, %f523; mul.ftz.f32 %f3190, %f3190, %f2140; mul.ftz.f32 %f3191, %f3191, %f2140; mul.ftz.f32 %f3192, %f3192, %f2140; $L__BB215_517: ld.const.u64 %rd577, [kernel_params+2448]; cvta.to.global.u64 %rd578, %rd577; add.s64 %rd580, %rd578, %rd260; ld.global.u32 %rd581, [%rd580]; ld.const.s32 %rd582, [kernel_params+1264]; mul.lo.s64 %rd102, %rd582, %rd581; ld.global.u32 %r265, [%rd30]; ld.global.u32 %r266, [%rd28]; ld.const.u32 %r1532, [kernel_params+1620]; setp.eq.s32 %p406, %r1532, 0; @%p406 bra $L__BB215_522; and.b32 %r1533, %r265, 536870912; setp.ne.s32 %p407, %r1533, 0; @%p407 bra $L__BB215_520; ld.const.s32 %rd583, [kernel_params+1384]; add.s64 %rd584, %rd102, %rd583; cvta.to.global.u64 %rd585, %rd226; shl.b64 %rd586, %rd584, 2; add.s64 %rd587, %rd585, %rd586; add.s64 %rd588, %rd587, 4; add.s64 %rd589, %rd587, 8; atom.global.add.f32 %f2141, [%rd587], %f3190; atom.global.add.f32 %f2142, [%rd588], %f3191; atom.global.add.f32 %f2143, [%rd589], %f3192; $L__BB215_520: and.b32 %r1534, %r265, 1073741824; setp.eq.s32 %p408, %r1534, 0; @%p408 bra $L__BB215_522; bra.uni $L__BB215_521; $L__BB215_522: ld.const.u32 %r2211, [kernel_params+1260]; and.b32 %r1535, %r2211, 2; setp.eq.s32 %p409, %r1535, 0; @%p409 bra $L__BB215_524; ld.const.s32 %rd597, [kernel_params+1268]; add.s64 %rd598, %rd102, %rd597; cvta.to.global.u64 %rd599, %rd226; shl.b64 %rd600, %rd598, 2; add.s64 %rd601, %rd599, %rd600; add.s64 %rd602, %rd601, 4; add.s64 %rd603, %rd601, 8; atom.global.add.f32 %f2147, [%rd601], %f3190; atom.global.add.f32 %f2148, [%rd602], %f3191; atom.global.add.f32 %f2149, [%rd603], %f3192; $L__BB215_524: ld.const.u32 %r269, [kernel_params+1400]; setp.eq.s32 %p410, %r269, -1; @%p410 bra $L__BB215_527; and.b32 %r1536, %r266, -1431655766; popc.b32 %r1537, %r1536; and.b32 %r1538, %r1537, 1; setp.eq.b32 %p411, %r1538, 1; mov.pred %p412, 0; xor.pred %p413, %p411, %p412; not.pred %p414, %p413; @%p414 bra $L__BB215_527; cvt.s64.s32 %rd604, %r269; add.s64 %rd605, %rd102, %rd604; cvta.to.global.u64 %rd606, %rd226; shl.b64 %rd607, %rd605, 2; add.s64 %rd608, %rd606, %rd607; add.s64 %rd609, %rd608, 4; add.s64 %rd610, %rd608, 8; add.s64 %rd611, %rd608, 12; add.ftz.f32 %f2150, %f3190, %f3190; atom.global.add.f32 %f2151, [%rd608], %f2150; add.ftz.f32 %f2152, %f3191, %f3191; atom.global.add.f32 %f2153, [%rd609], %f2152; add.ftz.f32 %f2154, %f3192, %f3192; atom.global.add.f32 %f2155, [%rd610], %f2154; atom.global.add.f32 %f2156, [%rd611], 0f00000000; bra.uni $L__BB215_527; $L__BB215_521: ld.const.s32 %rd590, [kernel_params+1376]; add.s64 %rd591, %rd102, %rd590; cvta.to.global.u64 %rd592, %rd226; shl.b64 %rd593, %rd591, 2; add.s64 %rd594, %rd592, %rd593; add.s64 %rd595, %rd594, 4; add.s64 %rd596, %rd594, 8; atom.global.add.f32 %f2144, [%rd594], %f3190; atom.global.add.f32 %f2145, [%rd595], %f3191; atom.global.add.f32 %f2146, [%rd596], %f3192; ld.const.u32 %r2211, [kernel_params+1260]; $L__BB215_527: setp.eq.s32 %p415, %r2211, 0; @%p415 bra $L__BB215_548; ld.const.u32 %r2212, [kernel_params+1352]; and.b32 %r1539, %r265, 16777216; setp.eq.s32 %p416, %r1539, 0; ld.const.u32 %r272, [kernel_params+1428]; setp.eq.s32 %p417, %r272, -1; or.pred %p418, %p416, %p417; @%p418 bra $L__BB215_530; ld.const.u64 %rd612, [kernel_params+2616]; cvta.to.global.u64 %rd613, %rd612; add.s64 %rd615, %rd613, %rd339; ld.global.f32 %f2157, [%rd615]; mul.ftz.f32 %f2158, %f3190, %f2157; ld.global.f32 %f2159, [%rd615+4]; mul.ftz.f32 %f2160, %f3191, %f2159; ld.global.f32 %f2161, [%rd615+8]; mul.ftz.f32 %f2162, %f3192, %f2161; cvta.to.global.u64 %rd616, %rd226; cvt.s64.s32 %rd617, %r272; add.s64 %rd618, %rd102, %rd617; shl.b64 %rd619, %rd618, 2; add.s64 %rd620, %rd616, %rd619; add.s64 %rd621, %rd620, 4; add.s64 %rd622, %rd620, 8; atom.global.add.f32 %f2163, [%rd620], %f2158; atom.global.add.f32 %f2164, [%rd621], %f2160; atom.global.add.f32 %f2165, [%rd622], %f2162; $L__BB215_530: setp.eq.s32 %p419, %r2210, -1; ld.const.u32 %r273, [kernel_params+1444]; setp.eq.s32 %p420, %r273, -1; or.pred %p421, %p419, %p420; @%p421 bra $L__BB215_532; mul.lo.s32 %r1540, %r2210, 3; cvt.s64.s32 %rd623, %r1540; add.s64 %rd624, %rd102, %rd623; cvt.s64.s32 %rd625, %r273; add.s64 %rd626, %rd624, %rd625; cvta.to.global.u64 %rd627, %rd226; shl.b64 %rd628, %rd626, 2; add.s64 %rd629, %rd627, %rd628; add.s64 %rd630, %rd629, 4; add.s64 %rd631, %rd629, 8; atom.global.add.f32 %f2166, [%rd629], %f3190; atom.global.add.f32 %f2167, [%rd630], %f3191; atom.global.add.f32 %f2168, [%rd631], %f3192; $L__BB215_532: and.b32 %r1541, %r265, 100663296; setp.eq.s32 %p422, %r1541, 0; @%p422 bra $L__BB215_546; ld.const.u8 %rs55, [kernel_params+2]; and.b16 %rs56, %rs55, 32; setp.eq.s16 %p423, %rs56, 0; mov.u32 %r2212, -1; @%p423 bra $L__BB215_546; bra.uni $L__BB215_534; $L__BB215_546: setp.eq.s32 %p434, %r2212, -1; @%p434 bra $L__BB215_548; cvt.s64.s32 %rd653, %r2212; add.s64 %rd654, %rd102, %rd653; cvta.to.global.u64 %rd655, %rd226; shl.b64 %rd656, %rd654, 2; add.s64 %rd657, %rd655, %rd656; add.s64 %rd658, %rd657, 4; add.s64 %rd659, %rd657, 8; atom.global.add.f32 %f2188, [%rd657], %f3190; atom.global.add.f32 %f2189, [%rd658], %f3191; atom.global.add.f32 %f2190, [%rd659], %f3192; bra.uni $L__BB215_548; $L__BB215_534: and.b32 %r1543, %r265, 536870912; setp.ne.s32 %p424, %r1543, 0; @%p424 bra $L__BB215_548; and.b32 %r1544, %r265, 33554432; setp.eq.s32 %p425, %r1544, 0; @%p425 bra $L__BB215_542; bra.uni $L__BB215_536; $L__BB215_542: and.b32 %r1553, %r265, 67108864; setp.eq.s32 %p432, %r1553, 0; @%p432 bra $L__BB215_546; setp.eq.s16 %p433, %rs17, 1; @%p433 bra $L__BB215_545; bra.uni $L__BB215_544; $L__BB215_545: ld.const.u32 %r2212, [kernel_params+1348]; bra.uni $L__BB215_546; $L__BB215_548: and.b32 %r2148, %r1028, 16777216; setp.eq.s32 %p435, %r2148, 0; @%p435 bra $L__BB215_562; ld.const.u64 %rd660, [kernel_params+2448]; cvta.to.global.u64 %rd661, %rd660; add.s64 %rd663, %rd661, %rd260; ld.global.u32 %rd664, [%rd663]; ld.const.s32 %rd665, [kernel_params+1264]; mul.lo.s64 %rd103, %rd665, %rd664; ld.const.u64 %rd666, [kernel_params+2616]; cvta.to.global.u64 %rd667, %rd666; add.s64 %rd669, %rd667, %rd339; ld.global.f32 %f542, [%rd669]; ld.global.f32 %f543, [%rd669+4]; ld.global.f32 %f544, [%rd669+8]; ld.local.u8 %rs57, [%rd49+180]; setp.eq.s16 %p436, %rs57, 0; ld.const.u32 %r280, [kernel_params+1424]; setp.eq.s32 %p437, %r280, -1; or.pred %p438, %p436, %p437; @%p438 bra $L__BB215_551; ld.global.u32 %r1554, [%rd30]; and.b32 %r1555, %r1554, -16777217; st.global.u32 [%rd30], %r1555; cvta.to.global.u64 %rd670, %rd226; cvt.s64.s32 %rd671, %r280; add.s64 %rd672, %rd103, %rd671; shl.b64 %rd673, %rd672, 2; add.s64 %rd674, %rd670, %rd673; add.s64 %rd675, %rd674, 4; add.s64 %rd676, %rd674, 8; atom.global.add.f32 %f2191, [%rd674], 0f00000000; atom.global.add.f32 %f2192, [%rd675], 0f00000000; atom.global.add.f32 %f2193, [%rd676], 0fBF800000; $L__BB215_551: ld.const.u32 %r281, [kernel_params+1428]; setp.eq.s32 %p439, %r281, -1; @%p439 bra $L__BB215_562; mul.ftz.f32 %f545, %f3169, %f542; mul.ftz.f32 %f546, %f3168, %f543; mul.ftz.f32 %f547, %f3167, %f544; mov.b32 %r282, %f545; setp.nan.ftz.f32 %p440, %f545, %f545; mov.f32 %f3195, 0f00000000; mov.f32 %f3193, %f3195; @%p440 bra $L__BB215_555; or.b32 %r1556, %r282, -2147483648; setp.ne.s32 %p441, %r1556, -2147483648; add.ftz.f32 %f2196, %f545, %f545; setp.eq.ftz.f32 %p442, %f2196, %f545; and.pred %p443, %p442, %p441; mov.f32 %f3193, %f3195; @%p443 bra $L__BB215_555; shl.b32 %r1557, %r282, 1; setp.lt.u32 %p444, %r1557, -16777215; selp.f32 %f3193, %f545, 0f00000000, %p444; $L__BB215_555: mov.b32 %r283, %f546; setp.nan.ftz.f32 %p445, %f546, %f546; mov.f32 %f3194, %f3195; @%p445 bra $L__BB215_558; or.b32 %r1558, %r283, -2147483648; setp.ne.s32 %p446, %r1558, -2147483648; add.ftz.f32 %f2199, %f546, %f546; setp.eq.ftz.f32 %p447, %f2199, %f546; and.pred %p448, %p447, %p446; mov.f32 %f3194, %f3195; @%p448 bra $L__BB215_558; shl.b32 %r1559, %r283, 1; setp.lt.u32 %p449, %r1559, -16777215; selp.f32 %f3194, %f546, 0f00000000, %p449; $L__BB215_558: mov.b32 %r284, %f547; setp.nan.ftz.f32 %p450, %f547, %f547; @%p450 bra $L__BB215_561; or.b32 %r1560, %r284, -2147483648; setp.ne.s32 %p451, %r1560, -2147483648; add.ftz.f32 %f2202, %f547, %f547; setp.eq.ftz.f32 %p452, %f2202, %f547; and.pred %p453, %p452, %p451; @%p453 bra $L__BB215_561; shl.b32 %r1561, %r284, 1; setp.lt.u32 %p454, %r1561, -16777215; selp.f32 %f3195, %f547, 0f00000000, %p454; $L__BB215_561: cvta.to.global.u64 %rd677, %rd226; cvt.s64.s32 %rd678, %r281; add.s64 %rd679, %rd103, %rd678; shl.b64 %rd680, %rd679, 2; add.s64 %rd681, %rd677, %rd680; add.s64 %rd682, %rd681, 4; add.s64 %rd683, %rd681, 8; atom.global.add.f32 %f2203, [%rd681], %f3193; atom.global.add.f32 %f2204, [%rd682], %f3194; atom.global.add.f32 %f2205, [%rd683], %f3195; $L__BB215_562: ld.global.u32 %r285, [%rd30]; and.b32 %r1562, %r285, 65536; setp.ne.s32 %p455, %r1562, 0; mov.f32 %f3196, 0f00000000; @%p455 bra $L__BB215_564; ld.const.u64 %rd684, [kernel_params+2576]; cvta.to.global.u64 %rd685, %rd684; add.s64 %rd687, %rd685, %rd260; ld.global.f32 %f3196, [%rd687]; $L__BB215_564: setp.eq.ftz.f32 %p456, %f3196, 0f00000000; mov.u32 %r1563, 2; mov.u32 %r2237, %r1563; @%p456 bra $L__BB215_936; bra.uni $L__BB215_565; $L__BB215_936: setp.eq.s32 %p706, %r2237, 0; @%p706 bra $L__BB215_1096; setp.eq.s32 %p707, %r2237, 2; @%p707 bra $L__BB215_1095; bra.uni $L__BB215_938; $L__BB215_1095: ld.const.u64 %rd1678, [kernel_params+4256]; cvta.to.global.u64 %rd1679, %rd1678; add.s64 %rd1680, %rd1679, 44; atom.global.add.u32 %r2142, [%rd1680], -1; ld.const.u64 %rd1681, [kernel_params+2520]; cvta.to.global.u64 %rd1682, %rd1681; add.s64 %rd1684, %rd1682, %rd263; mov.u16 %rs107, 0; st.global.u16 [%rd1684], %rs107; bra.uni $L__BB215_1097; $L__BB215_565: ld.local.u8 %rs58, [%rd49+-4]; setp.eq.s16 %p457, %rs58, 0; @%p457 bra $L__BB215_906; ld.local.f32 %f2207, [%rd49+12]; fma.rn.ftz.f32 %f556, %f4, %f2207, %f1; fma.rn.ftz.f32 %f557, %f5, %f2207, %f2; fma.rn.ftz.f32 %f558, %f6, %f2207, %f3; ld.local.f32 %f2208, [%rd49]; rcp.approx.ftz.f32 %f2209, %f3196; mul.ftz.f32 %f559, %f2209, %f2208; ld.local.f32 %f2210, [%rd49+4]; mul.ftz.f32 %f560, %f2209, %f2210; ld.local.f32 %f2211, [%rd49+8]; mul.ftz.f32 %f561, %f2209, %f2211; st.local.f32 [%rd49], %f559; st.local.f32 [%rd49+4], %f560; st.local.f32 [%rd49+8], %f561; ld.const.u32 %r1564, [kernel_params+1488]; setp.eq.s32 %p458, %r1564, 0; @%p458 bra $L__BB215_906; cvt.u32.u16 %r2149, %rs1; ld.const.u64 %rd688, [kernel_params+2464]; cvta.to.global.u64 %rd689, %rd688; add.s64 %rd104, %rd689, %rd263; ld.global.u16 %rs18, [%rd104]; add.s32 %r286, %r2149, 1; @%p44 bra $L__BB215_570; ld.const.v2.u32 {%r1566, %r1567}, [kernel_params+1600]; mov.b32 %f562, %r1567; setp.geu.ftz.f32 %p460, %f562, 0f3F800000; setp.lt.ftz.f32 %p461, %f562, 0f3F800000; ld.const.u32 %r1570, [kernel_params+1576]; selp.b32 %r1571, %r1570, %r6, %p461; add.s16 %rs59, %rs1, 1; and.b16 %rs60, %rs59, 255; cvt.u32.u16 %r1572, %rs60; xor.b32 %r1573, %r1571, %r1572; mul.lo.s32 %r1574, %r1573, -512718531; shr.u32 %r1575, %r1571, 16; xor.b32 %r1576, %r1574, %r1575; bfe.u32 %r1577, %r1576, 4, 4; shr.u32 %r1578, %r1571, 8; xor.b32 %r1579, %r1576, %r1578; xor.b32 %r1580, %r1579, %r1577; mul.lo.s32 %r1581, %r1580, 153742143; shr.u32 %r1582, %r1571, 23; xor.b32 %r1583, %r1581, %r1582; bfe.u32 %r1584, %r1583, 1, 7; xor.b32 %r1585, %r1584, %r1583; shr.u32 %r1586, %r1571, 27; or.b32 %r1587, %r1586, 1; mul.lo.s32 %r1588, %r1587, %r1585; mul.lo.s32 %r1589, %r1588, -1074439621; bfe.u32 %r1590, %r1589, 2, 6; xor.b32 %r1591, %r1590, %r1589; mul.lo.s32 %r1592, %r1591, -1638916925; bfe.u32 %r1593, %r1592, 2, 6; xor.b32 %r1594, %r1593, %r1592; mul.lo.s32 %r1595, %r1594, -933190689; and.b32 %r1596, %r1595, 255; bfe.u32 %r1597, %r1595, 5, 3; xor.b32 %r1598, %r1597, %r1596; add.s32 %r1599, %r1566, -1; xor.b32 %r287, %r286, 61; xor.b32 %r1600, %r1571, %r287; mul.lo.s32 %r1601, %r1600, 9; shr.u32 %r1602, %r1601, 4; xor.b32 %r1603, %r1602, %r1601; mul.lo.s32 %r1604, %r1603, 668265261; add.s32 %r1605, %r1604, %r47; shr.u32 %r1606, %r1604, 16; or.b32 %r1607, %r1606, 1; mul.lo.s32 %r1608, %r1607, %r1605; mul.lo.s32 %r1609, %r1608, 89287766; xor.b32 %r1610, %r1609, %r1608; mul.lo.s32 %r1611, %r1610, 1403136100; xor.b32 %r1612, %r1611, %r1610; brev.b32 %r1613, %r1612; neg.s32 %r1614, %r1566; and.b32 %r1615, %r8, %r1614; and.b32 %r1616, %r1613, %r1599; or.b32 %r1617, %r1616, %r1615; mad.lo.s32 %r1618, %r1598, %r1566, %r1617; shl.b32 %r1619, %r1566, 8; rem.u32 %r1620, %r1618, %r1619; ld.const.u64 %rd691, [kernel_params+2424]; cvta.to.global.u64 %rd692, %rd691; shl.b32 %r1621, %r1620, 1; mul.wide.u32 %rd693, %r1621, 4; add.s64 %rd694, %rd692, %rd693; ld.global.f32 %f3199, [%rd694]; ld.global.f32 %f3200, [%rd694+4]; @%p460 bra $L__BB215_574; xor.b32 %r1622, %r287, %r6; mul.lo.s32 %r1623, %r1622, 9; shr.u32 %r1624, %r1623, 4; xor.b32 %r1625, %r1624, %r1623; mul.lo.s32 %r1626, %r1625, 668265261; cvt.rn.f32.u32 %f2212, %r1626; mul.ftz.f32 %f2213, %f2212, 0f2F7FFFFE; xor.b32 %r1627, %r6, %r286; xor.b32 %r1628, %r1627, -905047700; mul.lo.s32 %r1629, %r1628, 9; shr.u32 %r1630, %r1629, 4; xor.b32 %r1631, %r1630, %r1629; mul.lo.s32 %r1632, %r1631, 668265261; cvt.rn.f32.u32 %f2214, %r1632; mul.ftz.f32 %f2215, %f2214, 0f2F7FFFFE; fma.rn.ftz.f32 %f2216, %f2213, %f562, %f3199; fma.rn.ftz.f32 %f2217, %f2215, %f562, %f3200; cvt.rmi.ftz.f32.f32 %f2218, %f2216; sub.ftz.f32 %f3199, %f2216, %f2218; cvt.rmi.ftz.f32.f32 %f2219, %f2217; sub.ftz.f32 %f3200, %f2217, %f2219; bra.uni $L__BB215_574; $L__BB215_1096: ld.const.u64 %rd1685, [kernel_params+4256]; cvta.to.global.u64 %rd1686, %rd1685; add.s64 %rd1687, %rd1686, 44; atom.global.add.u32 %r2143, [%rd1687], -1; add.s64 %rd1688, %rd1686, 8; atom.global.add.u32 %r2144, [%rd1688], 1; ld.const.u64 %rd1689, [kernel_params+2520]; cvta.to.global.u64 %rd1690, %rd1689; add.s64 %rd1692, %rd1690, %rd263; mov.u16 %rs108, 2; st.global.u16 [%rd1692], %rs108; bra.uni $L__BB215_1097; $L__BB215_938: cvt.u32.u64 %r2067, %rd16; setp.eq.s32 %p708, %r2067, -1; @%p708 bra $L__BB215_1094; and.b32 %r2068, %r5, 32; setp.eq.s32 %p709, %r2068, 0; @%p709 bra $L__BB215_941; bra.uni $L__BB215_940; $L__BB215_941: and.b32 %r449, %r5, 1; setp.eq.s32 %p710, %r449, 0; @%p710 bra $L__BB215_943; bra.uni $L__BB215_942; $L__BB215_943: and.b32 %r2071, %r5, 8; setp.eq.s32 %p711, %r2071, 0; @%p711 bra $L__BB215_945; bra.uni $L__BB215_944; $L__BB215_945: and.b32 %r2073, %r5, 6; setp.eq.s32 %p712, %r2073, 0; mov.u32 %r2238, 0; @%p712 bra $L__BB215_947; ld.const.u64 %rd1278, [kernel_params+2264]; cvta.to.global.u64 %rd1279, %rd1278; mul.wide.s32 %rd1280, %r2067, 16; add.s64 %rd1281, %rd1279, %rd1280; ld.global.u32 %r2238, [%rd1281]; bra.uni $L__BB215_947; $L__BB215_1094: ld.const.u64 %rd1670, [kernel_params+4256]; cvta.to.global.u64 %rd1671, %rd1670; add.s64 %rd1672, %rd1671, 44; atom.global.add.u32 %r2140, [%rd1672], -1; add.s64 %rd1673, %rd1671, 24; atom.global.add.u32 %r2141, [%rd1673], 1; ld.const.u64 %rd1674, [kernel_params+2520]; cvta.to.global.u64 %rd1675, %rd1674; add.s64 %rd1677, %rd1675, %rd263; mov.u16 %rs106, 6; st.global.u16 [%rd1677], %rs106; bra.uni $L__BB215_1097; $L__BB215_940: ld.const.u64 %rd1262, [kernel_params+4256]; cvta.to.global.u64 %rd1263, %rd1262; add.s64 %rd1264, %rd1263, 44; atom.global.add.u32 %r2069, [%rd1264], -1; add.s64 %rd1265, %rd1263, 28; atom.global.add.u32 %r2070, [%rd1265], 1; ld.const.u64 %rd1266, [kernel_params+2520]; cvta.to.global.u64 %rd1267, %rd1266; add.s64 %rd1269, %rd1267, %rd263; mov.u16 %rs77, 7; st.global.u16 [%rd1269], %rs77; bra.uni $L__BB215_1097; $L__BB215_570: shr.u32 %r1634, %r286, 16; xor.b32 %r1635, %r1634, %r286; mul.lo.s32 %r1636, %r1635, 569420461; shr.u32 %r1637, %r1636, 15; xor.b32 %r1638, %r1637, %r1636; mul.lo.s32 %r1639, %r1638, -749064809; shr.u32 %r1640, %r1639, 15; xor.b32 %r1641, %r6, %r1639; xor.b32 %r1642, %r1641, %r1640; xor.b32 %r288, %r1642, -419546133; xor.b32 %r1643, %r1642, 508809841; add.s32 %r1644, %r1643, %r47; shr.u32 %r1645, %r1643, 16; or.b32 %r1646, %r1645, 1; mul.lo.s32 %r1647, %r1646, %r1644; mul.lo.s32 %r1648, %r1647, 89287766; xor.b32 %r1649, %r1648, %r1647; mul.lo.s32 %r1650, %r1649, 1403136100; xor.b32 %r289, %r1650, %r1649; setp.eq.s32 %p462, %r289, 0; mov.u32 %r2216, 0; @%p462 bra $L__BB215_573; mov.u32 %r2213, 0; mov.u64 %rd696, sobol_burley_table; mov.u32 %r2214, %r289; mov.u32 %r2216, %r2213; $L__BB215_572: clz.b32 %r1653, %r2214; add.s32 %r1654, %r1653, %r2213; mul.wide.u32 %rd695, %r1654, 4; add.s64 %rd697, %rd696, %rd695; ld.const.u32 %r1655, [%rd697+128]; xor.b32 %r2216, %r1655, %r2216; add.s32 %r2213, %r1654, 1; shl.b32 %r1656, %r2214, %r1653; shl.b32 %r2214, %r1656, 1; setp.ne.s32 %p463, %r2214, 0; @%p463 bra $L__BB215_572; $L__BB215_573: xor.b32 %r1657, %r288, -525684874; brev.b32 %r1658, %r289; shr.u32 %r1659, %r1657, 16; or.b32 %r1660, %r1659, 1; mul.lo.s32 %r1661, %r1658, 1025551850; xor.b32 %r1662, %r1661, %r1658; add.s32 %r1663, %r1662, %r1657; mul.lo.s32 %r1664, %r1663, %r1660; mul.lo.s32 %r1665, %r1664, 89287766; xor.b32 %r1666, %r1665, %r1664; xor.b32 %r1667, %r288, -1802089138; mul.lo.s32 %r1668, %r2216, 1025551850; xor.b32 %r1669, %r1668, %r2216; add.s32 %r1670, %r1669, %r1667; shr.u32 %r1671, %r1667, 16; or.b32 %r1672, %r1671, 1; mul.lo.s32 %r1673, %r1670, %r1672; mul.lo.s32 %r1674, %r1673, 89287766; xor.b32 %r1675, %r1674, %r1673; mul.lo.s32 %r1676, %r1675, 1403136100; xor.b32 %r1677, %r1676, %r1675; brev.b32 %r1678, %r1677; cvt.rn.f32.u32 %f2220, %r1678; mul.lo.s32 %r1679, %r1666, 1403136100; xor.b32 %r1680, %r1679, %r1666; brev.b32 %r1681, %r1680; cvt.rn.f32.u32 %f2221, %r1681; mul.ftz.f32 %f3200, %f2220, 0f2F7FFFFE; mul.ftz.f32 %f3199, %f2221, 0f2F7FFFFE; $L__BB215_574: add.u64 %rd1694, %SP, 6752; ld.local.f32 %f2222, [%rd23+68]; cvt.u32.u16 %r1682, %rs18; { // callseq 1453, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f3199; .param .b32 param1; st.param.f32 [param1+0], %f3200; .param .b32 param2; st.param.f32 [param2+0], %f2222; .param .align 4 .b8 param3[12]; st.param.f32 [param3+0], %f556; st.param.f32 [param3+4], %f557; st.param.f32 [param3+8], %f558; .param .b32 param4; st.param.b32 [param4+0], %r1682; .param .b32 param5; st.param.b32 [param5+0], %r285; .param .b64 param6; st.param.b64 [param6+0], %rd1694; .param .b32 retval0; call.uni (retval0), _Z25light_distribution_sampleILb0EEbPK16KernelGlobalsGPUfff6float3ijP11LightSample, ( param0, param1, param2, param3, param4, param5, param6 ); ld.param.b32 %r1683, [retval0+0]; } // callseq 1453 cvt.u16.u32 %rs61, %r1683; setp.eq.s16 %p464, %rs61, 0; @%p464 bra $L__BB215_906; ld.local.u32 %r297, [%rd31+64]; and.b32 %r1684, %r297, 8388608; setp.ne.s32 %p465, %r1684, 0; @%p465 bra $L__BB215_906; ld.local.f32 %f575, [%rd23+68]; ld.const.u64 %rd701, [kernel_params+2408]; cvta.to.global.u64 %rd105, %rd701; and.b32 %r2220, %r297, 4194303; shl.b32 %r1685, %r297, 5; cvt.u64.u32 %rd702, %r1685; and.b64 %rd703, %rd702, 134217696; add.s64 %rd704, %rd105, %rd703; add.s64 %rd106, %rd704, 16; ld.global.u32 %r299, [%rd704+16]; and.b32 %r1686, %r299, 134217728; setp.eq.s32 %p467, %r1686, 0; mov.pred %p799, -1; mov.f32 %f3260, 0f00000000; mov.f32 %f3261, %f3260; mov.f32 %f3262, %f3260; @%p467 bra $L__BB215_578; ld.global.f32 %f3260, [%rd106+-16]; ld.global.f32 %f3261, [%rd106+-12]; ld.global.f32 %f3262, [%rd106+-8]; mov.pred %p799, 0; $L__BB215_578: @%p799 bra $L__BB215_582; bra.uni $L__BB215_579; $L__BB215_582: ld.local.u32 %r300, [%rd31+76]; setp.eq.s32 %p471, %r300, 2; ld.local.f32 %f585, [%rd31]; ld.local.f32 %f586, [%rd31+4]; ld.local.f32 %f587, [%rd31+8]; add.u64 %rd713, %SP, 0; or.b64 %rd107, %rd713, 12; @%p471 bra $L__BB215_620; bra.uni $L__BB215_583; $L__BB215_620: add.s64 %rd857, %rd107, 20; add.s64 %rd858, %rd107, 4; add.s64 %rd862, %rd713, 248; cvta.to.local.u64 %rd863, %rd862; add.s64 %rd864, %rd713, 240; cvta.to.local.u64 %rd865, %rd864; ld.local.f32 %f2644, [%rd31+32]; ld.local.f32 %f2645, [%rd31+28]; ld.local.f32 %f2646, [%rd31+24]; st.local.v2.f32 [%rd865], {%f585, %f586}; st.local.f32 [%rd863], %f587; cvta.to.local.u64 %rd866, %rd713; neg.ftz.f32 %f2647, %f2646; st.local.v4.f32 [%rd866], {%f2646, %f2645, %f2644, %f2647}; cvta.to.local.u64 %rd867, %rd858; neg.ftz.f32 %f2648, %f2644; neg.ftz.f32 %f2649, %f2645; st.local.v4.f32 [%rd867], {%f2649, %f2648, %f2647, %f2649}; cvta.to.local.u64 %rd868, %rd857; st.local.v4.f32 [%rd868], {%f2648, %f2647, %f2649, %f2648}; ld.const.u32 %r1718, [kernel_params+1024]; and.b32 %r2220, %r1718, 4194303; shl.b32 %r1719, %r1718, 5; cvt.u64.u32 %rd869, %r1719; and.b64 %rd870, %rd869, 134217696; add.s64 %rd871, %rd105, %rd870; add.s64 %rd872, %rd713, 76; cvta.to.local.u64 %rd873, %rd872; ld.global.u32 %r1720, [%rd871+16]; mov.u32 %r1721, -1; st.local.u32 [%rd873], %r1721; mov.u32 %r1722, 0; add.s64 %rd874, %rd107, 36; cvta.to.local.u64 %rd875, %rd874; st.local.v4.u32 [%rd875], {%r1718, %r1720, %r1722, %r1721}; add.s64 %rd876, %rd713, 68; cvta.to.local.u64 %rd877, %rd876; st.local.u32 [%rd877], %r1722; add.s64 %rd878, %rd713, 72; cvta.to.local.u64 %rd879, %rd878; st.local.u32 [%rd879], %r1722; add.s64 %rd880, %rd713, 136; cvta.to.local.u64 %rd881, %rd880; add.s64 %rd882, %rd713, 128; cvta.to.local.u64 %rd883, %rd882; mov.f32 %f2650, 0f00000000; st.local.v2.f32 [%rd883], {%f2650, %f2650}; st.local.u32 [%rd881], %r1722; mov.b32 %r1723, %f575; add.s64 %rd884, %rd713, 80; cvta.to.local.u64 %rd885, %rd884; st.local.v4.u32 [%rd885], {%r1721, %r1723, %r1722, %r1722}; add.s64 %rd886, %rd713, 96; cvta.to.local.u64 %rd887, %rd886; st.local.v4.f32 [%rd887], {%f2650, %f2650, %f2650, %f2650}; add.s64 %rd888, %rd713, 112; cvta.to.local.u64 %rd889, %rd888; st.local.v4.f32 [%rd889], {%f2650, %f2650, %f2650, %f2650}; bra.uni $L__BB215_621; $L__BB215_942: ld.const.u64 %rd1270, [kernel_params+2216]; cvta.to.global.u64 %rd1271, %rd1270; shl.b64 %rd1272, %rd16, 2; add.s64 %rd1273, %rd1271, %rd1272; ld.global.u32 %r2238, [%rd1273]; bra.uni $L__BB215_947; $L__BB215_502: mov.u32 %r1524, 0; st.local.u32 [%rd49+184], %r1524; st.local.u32 [%rd49+188], %r1524; st.local.u32 [%rd49+192], %r1524; bra.uni $L__BB215_503; $L__BB215_536: ld.const.u64 %rd632, [kernel_params+2600]; cvta.to.global.u64 %rd633, %rd632; add.s64 %rd635, %rd633, %rd339; ld.global.f32 %f530, [%rd635]; ld.global.f32 %f531, [%rd635+4]; ld.global.f32 %f532, [%rd635+8]; ld.const.u64 %rd636, [kernel_params+2608]; cvta.to.global.u64 %rd637, %rd636; add.s64 %rd638, %rd637, %rd339; ld.global.f32 %f533, [%rd638]; ld.global.f32 %f534, [%rd638+4]; ld.global.f32 %f535, [%rd638+8]; ld.const.u32 %r1545, [kernel_params+1324]; ld.const.u32 %r1546, [kernel_params+1340]; setp.eq.s16 %p426, %rs17, 1; selp.b32 %r274, %r1546, %r1545, %p426; setp.eq.s32 %p427, %r274, -1; @%p427 bra $L__BB215_538; cvt.s64.s32 %rd639, %r274; add.s64 %rd640, %rd102, %rd639; cvta.to.global.u64 %rd641, %rd226; shl.b64 %rd642, %rd640, 2; add.s64 %rd643, %rd641, %rd642; add.s64 %rd644, %rd643, 4; add.s64 %rd645, %rd643, 8; mul.ftz.f32 %f2169, %f3190, %f533; atom.global.add.f32 %f2170, [%rd643], %f2169; mul.ftz.f32 %f2171, %f3191, %f534; atom.global.add.f32 %f2172, [%rd644], %f2171; mul.ftz.f32 %f2173, %f3192, %f535; atom.global.add.f32 %f2174, [%rd645], %f2173; $L__BB215_538: ld.const.u32 %r1547, [kernel_params+1328]; ld.const.u32 %r1548, [kernel_params+1344]; selp.b32 %r275, %r1548, %r1547, %p426; setp.eq.s32 %p429, %r275, -1; @%p429 bra $L__BB215_540; mov.f32 %f2175, 0f3F800000; sub.ftz.f32 %f2176, %f2175, %f530; sub.ftz.f32 %f2177, %f2176, %f533; sub.ftz.f32 %f2178, %f2175, %f531; sub.ftz.f32 %f2179, %f2178, %f534; sub.ftz.f32 %f2180, %f2175, %f532; sub.ftz.f32 %f2181, %f2180, %f535; cvt.s64.s32 %rd646, %r275; add.s64 %rd647, %rd102, %rd646; cvta.to.global.u64 %rd648, %rd226; shl.b64 %rd649, %rd647, 2; add.s64 %rd650, %rd648, %rd649; mul.ftz.f32 %f2182, %f3190, %f2177; mul.ftz.f32 %f2183, %f3191, %f2179; mul.ftz.f32 %f2184, %f3192, %f2181; add.s64 %rd651, %rd650, 4; add.s64 %rd652, %rd650, 8; atom.global.add.f32 %f2185, [%rd650], %f2182; atom.global.add.f32 %f2186, [%rd651], %f2183; atom.global.add.f32 %f2187, [%rd652], %f2184; $L__BB215_540: ld.const.u32 %r1550, [kernel_params+1320]; ld.const.u32 %r1551, [kernel_params+1336]; selp.b32 %r276, %r1551, %r1550, %p426; setp.eq.s32 %p431, %r276, -1; @%p431 bra $L__BB215_546; mul.ftz.f32 %f3190, %f3190, %f530; mul.ftz.f32 %f3191, %f3191, %f531; mul.ftz.f32 %f3192, %f3192, %f532; mov.u32 %r2212, %r276; bra.uni $L__BB215_546; $L__BB215_944: ld.const.u64 %rd1274, [kernel_params+2304]; cvta.to.global.u64 %rd1275, %rd1274; shl.b64 %rd1276, %rd16, 2; add.s64 %rd1277, %rd1275, %rd1276; ld.global.u32 %r2238, [%rd1277]; $L__BB215_947: ld.const.u64 %rd1282, [kernel_params+2408]; cvta.to.global.u64 %rd156, %rd1282; and.b32 %r454, %r2238, 4194303; ld.const.u64 %rd1283, [kernel_params+2184]; cvta.to.global.u64 %rd1284, %rd1283; shl.b64 %rd1285, %rd43, 2; add.s64 %rd157, %rd1284, %rd1285; ld.const.u32 %r455, [kernel_params+1592]; setp.ne.s32 %p713, %r455, 0; ld.global.u8 %rs78, [%rd157+1]; and.b16 %rs79, %rs78, 6; setp.ne.s16 %p714, %rs79, 0; and.pred %p715, %p714, %p713; @%p715 bra $L__BB215_951; bra.uni $L__BB215_948; $L__BB215_951: cvt.u32.u64 %r2084, %rd4; ld.const.u32 %r2239, [kernel_params+4392]; div.u32 %r2085, %r2084, %r2239; ld.const.u32 %r2240, [kernel_params+8]; mad.lo.s32 %r2086, %r2085, %r2240, %r454; ld.const.u64 %rd168, [kernel_params+4256]; cvta.to.global.u64 %rd1311, %rd168; add.s64 %rd1312, %rd1311, 44; atom.global.add.u32 %r2087, [%rd1312], -1; add.s64 %rd1313, %rd1311, 40; atom.global.add.u32 %r2088, [%rd1313], 1; ld.const.u64 %rd170, [kernel_params+2520]; cvta.to.global.u64 %rd1314, %rd170; add.s64 %rd1316, %rd1314, %rd263; mov.u16 %rs82, 10; st.global.u16 [%rd1316], %rs82; ld.const.u64 %rd169, [kernel_params+2624]; cvta.to.global.u64 %rd1317, %rd169; add.s64 %rd1319, %rd1317, %rd260; st.global.u32 [%rd1319], %r2086; ld.const.u64 %rd1320, [kernel_params+4344]; cvta.to.global.u64 %rd1321, %rd1320; mul.wide.s32 %rd1322, %r2086, 4; add.s64 %rd1323, %rd1321, %rd1322; atom.global.add.u32 %r2089, [%rd1323], 1; bra.uni $L__BB215_952; $L__BB215_948: cvt.u32.u64 %r2075, %rd4; mul.wide.u32 %rd1286, %r454, 32; add.s64 %rd1287, %rd156, %rd1286; ld.global.u32 %r2076, [%rd1287+16]; and.b32 %r2077, %r2076, 1073741824; setp.eq.s32 %p716, %r2077, 0; ld.const.u32 %r2239, [kernel_params+4392]; div.u32 %r2078, %r2075, %r2239; ld.const.u32 %r2240, [kernel_params+8]; mad.lo.s32 %r458, %r2078, %r2240, %r454; ld.const.u64 %rd168, [kernel_params+4256]; cvta.to.global.u64 %rd159, %rd168; add.s64 %rd1288, %rd159, 44; atom.global.add.u32 %r2079, [%rd1288], -1; cvt.s64.s32 %rd160, %r458; @%p716 bra $L__BB215_950; bra.uni $L__BB215_949; $L__BB215_950: add.s64 %rd1300, %rd159, 32; atom.global.add.u32 %r2082, [%rd1300], 1; ld.const.u64 %rd170, [kernel_params+2520]; cvta.to.global.u64 %rd1301, %rd170; add.s64 %rd1303, %rd1301, %rd263; mov.u16 %rs81, 8; st.global.u16 [%rd1303], %rs81; ld.const.u64 %rd169, [kernel_params+2624]; cvta.to.global.u64 %rd1304, %rd169; add.s64 %rd1306, %rd1304, %rd260; st.global.u32 [%rd1306], %r458; ld.const.u64 %rd1307, [kernel_params+4328]; cvta.to.global.u64 %rd1308, %rd1307; shl.b64 %rd1309, %rd160, 2; add.s64 %rd1310, %rd1308, %rd1309; atom.global.add.u32 %r2083, [%rd1310], 1; bra.uni $L__BB215_952; $L__BB215_949: add.s64 %rd1289, %rd159, 36; atom.global.add.u32 %r2080, [%rd1289], 1; ld.const.u64 %rd170, [kernel_params+2520]; cvta.to.global.u64 %rd1290, %rd170; add.s64 %rd1292, %rd1290, %rd263; mov.u16 %rs80, 9; st.global.u16 [%rd1292], %rs80; ld.const.u64 %rd169, [kernel_params+2624]; cvta.to.global.u64 %rd1293, %rd169; add.s64 %rd1295, %rd1293, %rd260; st.global.u32 [%rd1295], %r458; ld.const.u64 %rd1296, [kernel_params+4336]; cvta.to.global.u64 %rd1297, %rd1296; shl.b64 %rd1298, %rd160, 2; add.s64 %rd1299, %rd1297, %rd1298; atom.global.add.u32 %r2081, [%rd1299], 1; $L__BB215_952: ld.const.u32 %r2090, [kernel_params+1620]; setp.eq.s32 %p717, %r2090, 0; ld.global.u32 %r463, [%rd157]; and.b32 %r2091, %r463, 129; setp.ne.s32 %p718, %r2091, 128; or.pred %p719, %p718, %p717; mov.u16 %rs115, 0; @%p719 bra $L__BB215_955; ld.global.u32 %r464, [%rd30]; and.b32 %r2092, %r464, 16384; setp.eq.s32 %p720, %r2092, 0; @%p720 bra $L__BB215_955; and.b32 %r2093, %r464, 536870912; setp.eq.s32 %p721, %r2093, 0; selp.u16 %rs115, 1, 0, %p721; $L__BB215_955: setp.eq.s16 %p722, %rs115, 0; @%p722 bra $L__BB215_1097; ld.const.u64 %rd1324, [kernel_params+2448]; cvta.to.global.u64 %rd171, %rd1324; add.s64 %rd172, %rd171, %rd260; ld.global.u32 %rd1326, [%rd172]; ld.const.s32 %rd1327, [kernel_params+1264]; mul.lo.s64 %rd1328, %rd1327, %rd1326; ld.const.s32 %rd1329, [kernel_params+1380]; add.s64 %rd1330, %rd1328, %rd1329; cvta.to.global.u64 %rd1331, %rd226; shl.b64 %rd1332, %rd1330, 2; add.s64 %rd1333, %rd1331, %rd1332; atom.global.add.f32 %f2943, [%rd1333], 0f3F800000; ld.const.s32 %rd1334, [kernel_params+1384]; add.s64 %rd1335, %rd1334, %rd1328; shl.b64 %rd1336, %rd1335, 2; add.s64 %rd1337, %rd1331, %rd1336; add.s64 %rd1338, %rd1337, 12; ld.global.f32 %f2944, [%rd48+4]; ld.global.f32 %f2945, [%rd48]; add.ftz.f32 %f2946, %f2945, %f2944; ld.global.f32 %f2947, [%rd48+8]; add.ftz.f32 %f2948, %f2947, %f2946; mul.ftz.f32 %f2949, %f2948, 0f3EAAAAAB; atom.global.add.f32 %f2950, [%rd1338], %f2949; ld.global.u32 %r2094, [%rd30]; or.b32 %r2095, %r2094, 536870912; st.global.u32 [%rd30], %r2095; ld.const.u64 %rd1339, [kernel_params+4384]; cvta.to.global.u64 %rd1340, %rd1339; atom.global.add.u32 %r465, [%rd1340], 1; setp.eq.s64 %p723, %rd1324, 0; @%p723 bra $L__BB215_958; ld.global.u32 %r2096, [%rd172]; mul.wide.s32 %rd1341, %r465, 4; add.s64 %rd1342, %rd171, %rd1341; st.global.u32 [%rd1342], %r2096; $L__BB215_958: ld.const.u64 %rd1698, [kernel_params+2456]; setp.eq.s64 %p724, %rd1698, 0; @%p724 bra $L__BB215_960; ld.const.u64 %rd1700, [kernel_params+2456]; cvta.to.global.u64 %rd1699, %rd1700; ld.global.u32 %r2097, [%rd28]; mul.wide.s32 %rd1343, %r465, 4; add.s64 %rd1344, %rd1699, %rd1343; st.global.u32 [%rd1344], %r2097; $L__BB215_960: ld.const.u64 %rd173, [kernel_params+2464]; setp.eq.s64 %p725, %rd173, 0; @%p725 bra $L__BB215_962; cvta.to.global.u64 %rd1345, %rd173; add.s64 %rd1347, %rd1345, %rd263; ld.global.u16 %rs85, [%rd1347]; mul.wide.s32 %rd1348, %r465, 2; add.s64 %rd1349, %rd1345, %rd1348; st.global.u16 [%rd1349], %rs85; $L__BB215_962: ld.const.u64 %rd174, [kernel_params+2472]; setp.eq.s64 %p726, %rd174, 0; @%p726 bra $L__BB215_964; cvta.to.global.u64 %rd1350, %rd174; add.s64 %rd1352, %rd1350, %rd263; ld.global.u16 %rs86, [%rd1352]; mul.wide.s32 %rd1353, %r465, 2; add.s64 %rd1354, %rd1350, %rd1353; st.global.u16 [%rd1354], %rs86; $L__BB215_964: ld.const.u64 %rd175, [kernel_params+2480]; setp.eq.s64 %p727, %rd175, 0; @%p727 bra $L__BB215_966; cvta.to.global.u64 %rd1355, %rd175; add.s64 %rd1357, %rd1355, %rd263; ld.global.u16 %rs87, [%rd1357]; mul.wide.s32 %rd1358, %r465, 2; add.s64 %rd1359, %rd1355, %rd1358; st.global.u16 [%rd1359], %rs87; $L__BB215_966: ld.const.u64 %rd176, [kernel_params+2488]; setp.eq.s64 %p728, %rd176, 0; @%p728 bra $L__BB215_968; cvta.to.global.u64 %rd1360, %rd176; add.s64 %rd1362, %rd1360, %rd263; ld.global.u16 %rs88, [%rd1362]; mul.wide.s32 %rd1363, %r465, 2; add.s64 %rd1364, %rd1360, %rd1363; st.global.u16 [%rd1364], %rs88; $L__BB215_968: ld.const.u64 %rd177, [kernel_params+2496]; setp.eq.s64 %p729, %rd177, 0; @%p729 bra $L__BB215_970; cvta.to.global.u64 %rd1365, %rd177; add.s64 %rd1367, %rd1365, %rd263; ld.global.u16 %rs89, [%rd1367]; mul.wide.s32 %rd1368, %r465, 2; add.s64 %rd1369, %rd1365, %rd1368; st.global.u16 [%rd1369], %rs89; $L__BB215_970: ld.const.u64 %rd178, [kernel_params+2504]; setp.eq.s64 %p730, %rd178, 0; @%p730 bra $L__BB215_972; cvta.to.global.u64 %rd1370, %rd178; add.s64 %rd1372, %rd1370, %rd263; ld.global.u16 %rs90, [%rd1372]; mul.wide.s32 %rd1373, %r465, 2; add.s64 %rd1374, %rd1370, %rd1373; st.global.u16 [%rd1374], %rs90; $L__BB215_972: ld.const.u64 %rd179, [kernel_params+2512]; setp.eq.s64 %p731, %rd179, 0; @%p731 bra $L__BB215_974; cvta.to.global.u64 %rd1375, %rd179; add.s64 %rd1377, %rd1375, %rd263; ld.global.u16 %rs91, [%rd1377]; mul.wide.s32 %rd1378, %r465, 2; add.s64 %rd1379, %rd1375, %rd1378; st.global.u16 [%rd1379], %rs91; $L__BB215_974: setp.eq.s64 %p732, %rd170, 0; @%p732 bra $L__BB215_976; cvta.to.global.u64 %rd1380, %rd170; add.s64 %rd1382, %rd1380, %rd263; ld.global.u16 %rs92, [%rd1382]; mul.wide.s32 %rd1383, %r465, 2; add.s64 %rd1384, %rd1380, %rd1383; st.global.u16 [%rd1384], %rs92; $L__BB215_976: setp.eq.s64 %p733, %rd257, 0; @%p733 bra $L__BB215_978; cvta.to.global.u64 %rd1726, %rd257; add.s64 %rd1725, %rd1726, %rd260; cvta.to.global.u64 %rd1701, %rd257; ld.global.u32 %r2098, [%rd1725]; mul.wide.s32 %rd1388, %r465, 4; add.s64 %rd1389, %rd1701, %rd1388; st.global.u32 [%rd1389], %r2098; $L__BB215_978: setp.eq.s64 %p734, %rd258, 0; @%p734 bra $L__BB215_980; cvta.to.global.u64 %rd1728, %rd258; add.s64 %rd1727, %rd1728, %rd263; cvta.to.global.u64 %rd1702, %rd258; ld.global.u16 %rs93, [%rd1727]; mul.wide.s32 %rd1393, %r465, 2; add.s64 %rd1394, %rd1702, %rd1393; st.global.u16 [%rd1394], %rs93; $L__BB215_980: ld.const.u64 %rd1704, [kernel_params+2544]; cvta.to.global.u64 %rd1703, %rd1704; setp.eq.s64 %p735, %rd1704, 0; cvt.s64.s32 %rd180, %r465; mul.wide.s32 %rd1396, %r465, 4; add.s64 %rd181, %rd1703, %rd1396; @%p735 bra $L__BB215_982; ld.global.u32 %r2099, [%rd30]; st.global.u32 [%rd181], %r2099; $L__BB215_982: ld.const.u64 %rd182, [kernel_params+2552]; setp.eq.s64 %p736, %rd182, 0; @%p736 bra $L__BB215_984; cvta.to.global.u64 %rd1397, %rd182; add.s64 %rd1398, %rd1397, %rd4; ld.global.u8 %rs94, [%rd1398]; add.s64 %rd1400, %rd1397, %rd180; st.global.u8 [%rd1400], %rs94; $L__BB215_984: ld.const.u64 %rd183, [kernel_params+2560]; setp.eq.s64 %p737, %rd183, 0; @%p737 bra $L__BB215_986; cvta.to.global.u64 %rd1401, %rd183; add.s64 %rd1403, %rd1401, %rd260; ld.global.f32 %f2951, [%rd1403]; add.s64 %rd1405, %rd1401, %rd1396; st.global.f32 [%rd1405], %f2951; $L__BB215_986: ld.const.u64 %rd184, [kernel_params+2568]; setp.eq.s64 %p738, %rd184, 0; @%p738 bra $L__BB215_988; cvta.to.global.u64 %rd1406, %rd184; add.s64 %rd1408, %rd1406, %rd260; ld.global.f32 %f2952, [%rd1408]; add.s64 %rd1410, %rd1406, %rd1396; st.global.f32 [%rd1410], %f2952; $L__BB215_988: ld.const.u64 %rd185, [kernel_params+2576]; setp.eq.s64 %p739, %rd185, 0; @%p739 bra $L__BB215_990; cvta.to.global.u64 %rd1411, %rd185; add.s64 %rd1413, %rd1411, %rd260; ld.global.f32 %f2953, [%rd1413]; add.s64 %rd1415, %rd1411, %rd1396; st.global.f32 [%rd1415], %f2953; $L__BB215_990: ld.const.u64 %rd1705, [kernel_params+2584]; setp.eq.s64 %p740, %rd1705, 0; @%p740 bra $L__BB215_992; ld.const.u64 %rd1707, [kernel_params+2584]; cvta.to.global.u64 %rd1706, %rd1707; ld.global.f32 %f2954, [%rd48]; ld.global.f32 %f2955, [%rd48+4]; ld.global.f32 %f2956, [%rd48+8]; mul.wide.s32 %rd1416, %r465, 12; add.s64 %rd1417, %rd1706, %rd1416; st.global.f32 [%rd1417], %f2954; st.global.f32 [%rd1417+4], %f2955; st.global.f32 [%rd1417+8], %f2956; $L__BB215_992: ld.const.u64 %rd186, [kernel_params+2592]; setp.eq.s64 %p741, %rd186, 0; @%p741 bra $L__BB215_994; cvta.to.global.u64 %rd1418, %rd186; add.s64 %rd1420, %rd1418, %rd260; ld.global.f32 %f2957, [%rd1420]; add.s64 %rd1422, %rd1418, %rd1396; st.global.f32 [%rd1422], %f2957; $L__BB215_994: ld.const.u64 %rd187, [kernel_params+2600]; setp.eq.s64 %p742, %rd187, 0; @%p742 bra $L__BB215_996; cvta.to.global.u64 %rd1423, %rd187; add.s64 %rd1425, %rd1423, %rd339; ld.global.f32 %f2958, [%rd1425]; ld.global.f32 %f2959, [%rd1425+4]; ld.global.f32 %f2960, [%rd1425+8]; mul.wide.s32 %rd1426, %r465, 12; add.s64 %rd1427, %rd1423, %rd1426; st.global.f32 [%rd1427], %f2958; st.global.f32 [%rd1427+4], %f2959; st.global.f32 [%rd1427+8], %f2960; $L__BB215_996: ld.const.u64 %rd188, [kernel_params+2608]; setp.eq.s64 %p743, %rd188, 0; @%p743 bra $L__BB215_998; cvta.to.global.u64 %rd1428, %rd188; add.s64 %rd1430, %rd1428, %rd339; ld.global.f32 %f2961, [%rd1430]; ld.global.f32 %f2962, [%rd1430+4]; ld.global.f32 %f2963, [%rd1430+8]; mul.wide.s32 %rd1431, %r465, 12; add.s64 %rd1432, %rd1428, %rd1431; st.global.f32 [%rd1432], %f2961; st.global.f32 [%rd1432+4], %f2962; st.global.f32 [%rd1432+8], %f2963; $L__BB215_998: ld.const.u64 %rd189, [kernel_params+2616]; setp.eq.s64 %p744, %rd189, 0; @%p744 bra $L__BB215_1000; cvta.to.global.u64 %rd1433, %rd189; add.s64 %rd1435, %rd1433, %rd339; ld.global.f32 %f2964, [%rd1435]; ld.global.f32 %f2965, [%rd1435+4]; ld.global.f32 %f2966, [%rd1435+8]; mul.wide.s32 %rd1436, %r465, 12; add.s64 %rd1437, %rd1433, %rd1436; st.global.f32 [%rd1437], %f2964; st.global.f32 [%rd1437+4], %f2965; st.global.f32 [%rd1437+8], %f2966; $L__BB215_1000: setp.eq.s64 %p745, %rd169, 0; @%p745 bra $L__BB215_1002; cvta.to.global.u64 %rd1438, %rd169; add.s64 %rd1440, %rd1438, %rd260; ld.global.u32 %r2100, [%rd1440]; add.s64 %rd1442, %rd1438, %rd1396; st.global.u32 [%rd1442], %r2100; $L__BB215_1002: ld.const.u64 %rd1708, [kernel_params+2632]; setp.eq.s64 %p746, %rd1708, 0; @%p746 bra $L__BB215_1004; ld.const.u64 %rd1710, [kernel_params+2632]; cvta.to.global.u64 %rd1709, %rd1710; ld.global.f32 %f2967, [%rd5]; ld.global.f32 %f2968, [%rd5+4]; ld.global.f32 %f2969, [%rd5+8]; mul.wide.s32 %rd1443, %r465, 12; add.s64 %rd1444, %rd1709, %rd1443; st.global.f32 [%rd1444], %f2967; st.global.f32 [%rd1444+4], %f2968; st.global.f32 [%rd1444+8], %f2969; $L__BB215_1004: setp.eq.s64 %p747, %rd232, 0; @%p747 bra $L__BB215_1006; cvta.to.global.u64 %rd1711, %rd232; add.s64 %rd1447, %rd1711, %rd339; ld.global.f32 %f2970, [%rd1447]; ld.global.f32 %f2971, [%rd1447+4]; ld.global.f32 %f2972, [%rd1447+8]; mul.wide.s32 %rd1448, %r465, 12; add.s64 %rd1449, %rd1711, %rd1448; st.global.f32 [%rd1449], %f2970; st.global.f32 [%rd1449+4], %f2971; st.global.f32 [%rd1449+8], %f2972; $L__BB215_1006: setp.eq.s64 %p748, %rd233, 0; @%p748 bra $L__BB215_1008; cvta.to.global.u64 %rd1712, %rd233; add.s64 %rd1452, %rd1712, %rd260; ld.global.f32 %f2973, [%rd1452]; add.s64 %rd1454, %rd1712, %rd1396; st.global.f32 [%rd1454], %f2973; $L__BB215_1008: setp.eq.s64 %p749, %rd239, 0; @%p749 bra $L__BB215_1010; cvta.to.global.u64 %rd1455, %rd239; add.s64 %rd1457, %rd1455, %rd260; ld.global.f32 %f2974, [%rd1457]; add.s64 %rd1459, %rd1455, %rd1396; st.global.f32 [%rd1459], %f2974; $L__BB215_1010: setp.eq.s64 %p750, %rd240, 0; @%p750 bra $L__BB215_1012; mul.wide.s32 %rd1731, %r2152, 4; cvta.to.global.u64 %rd1730, %rd240; add.s64 %rd1729, %rd1730, %rd1731; cvta.to.global.u64 %rd1713, %rd240; ld.global.f32 %f2975, [%rd1729]; add.s64 %rd1461, %rd1713, %rd1396; st.global.f32 [%rd1461], %f2975; $L__BB215_1012: ld.const.u64 %rd190, [kernel_params+2672]; setp.eq.s64 %p751, %rd190, 0; @%p751 bra $L__BB215_1014; cvta.to.global.u64 %rd1462, %rd190; add.s64 %rd1464, %rd1462, %rd260; ld.global.f32 %f2976, [%rd1464]; add.s64 %rd1466, %rd1462, %rd1396; st.global.f32 [%rd1466], %f2976; $L__BB215_1014: ld.const.u64 %rd191, [kernel_params+2680]; setp.eq.s64 %p752, %rd191, 0; @%p752 bra $L__BB215_1016; cvta.to.global.u64 %rd1467, %rd191; add.s64 %rd1469, %rd1467, %rd260; ld.global.f32 %f2977, [%rd1469]; add.s64 %rd1471, %rd1467, %rd1396; st.global.f32 [%rd1471], %f2977; $L__BB215_1016: ld.const.u64 %rd1714, [kernel_params+2688]; setp.eq.s64 %p753, %rd1714, 0; @%p753 bra $L__BB215_1018; ld.const.u64 %rd1735, [kernel_params+2688]; mul.wide.s32 %rd1734, %r2152, 4; cvta.to.global.u64 %rd1733, %rd1735; add.s64 %rd1732, %rd1733, %rd1734; ld.const.u64 %rd1716, [kernel_params+2688]; cvta.to.global.u64 %rd1715, %rd1716; ld.global.f32 %f2978, [%rd1732]; add.s64 %rd1473, %rd1715, %rd1396; st.global.f32 [%rd1473], %f2978; $L__BB215_1018: ld.const.u64 %rd192, [kernel_params+2696]; setp.eq.s64 %p754, %rd192, 0; @%p754 bra $L__BB215_1020; cvta.to.global.u64 %rd1474, %rd192; add.s64 %rd1476, %rd1474, %rd260; ld.global.f32 %f2979, [%rd1476]; add.s64 %rd1478, %rd1474, %rd1396; st.global.f32 [%rd1478], %f2979; $L__BB215_1020: ld.const.u64 %rd193, [kernel_params+2704]; setp.eq.s64 %p755, %rd193, 0; @%p755 bra $L__BB215_1022; cvta.to.global.u64 %rd1479, %rd193; add.s64 %rd1481, %rd1479, %rd260; ld.global.f32 %f2980, [%rd1481]; add.s64 %rd1483, %rd1479, %rd1396; st.global.f32 [%rd1483], %f2980; $L__BB215_1022: ld.const.u64 %rd1736, [kernel_params+2712]; setp.eq.s64 %p756, %rd1736, 0; @%p756 bra $L__BB215_1024; ld.const.u64 %rd1738, [kernel_params+2712]; cvta.to.global.u64 %rd1737, %rd1738; ld.global.u32 %r2101, [%rd15]; add.s64 %rd1485, %rd1737, %rd1396; st.global.u32 [%rd1485], %r2101; $L__BB215_1024: setp.eq.s64 %p757, %rd241, 0; @%p757 bra $L__BB215_1026; cvta.to.global.u64 %rd1739, %rd241; add.s64 %rd1488, %rd1739, %rd260; ld.global.u32 %r2102, [%rd1488]; add.s64 %rd1490, %rd1739, %rd1396; st.global.u32 [%rd1490], %r2102; $L__BB215_1026: setp.eq.s64 %p758, %rd242, 0; @%p758 bra $L__BB215_1028; cvta.to.global.u64 %rd1717, %rd242; add.s64 %rd1493, %rd1717, %rd260; ld.global.u32 %r2103, [%rd1493]; add.s64 %rd1495, %rd1717, %rd1396; st.global.u32 [%rd1495], %r2103; $L__BB215_1028: ld.const.u64 %rd194, [kernel_params+2736]; setp.eq.s64 %p759, %rd194, 0; @%p759 bra $L__BB215_1030; cvta.to.global.u64 %rd1496, %rd194; add.s64 %rd1498, %rd1496, %rd339; ld.global.f32 %f2981, [%rd1498]; ld.global.f32 %f2982, [%rd1498+4]; ld.global.f32 %f2983, [%rd1498+8]; mul.wide.s32 %rd1499, %r465, 12; add.s64 %rd1500, %rd1496, %rd1499; st.global.f32 [%rd1500], %f2981; st.global.f32 [%rd1500+4], %f2982; st.global.f32 [%rd1500+8], %f2983; $L__BB215_1030: ld.const.u64 %rd195, [kernel_params+2744]; setp.eq.s64 %p760, %rd195, 0; @%p760 bra $L__BB215_1032; cvta.to.global.u64 %rd1501, %rd195; add.s64 %rd1503, %rd1501, %rd339; ld.global.f32 %f2984, [%rd1503]; ld.global.f32 %f2985, [%rd1503+4]; ld.global.f32 %f2986, [%rd1503+8]; mul.wide.s32 %rd1504, %r465, 12; add.s64 %rd1505, %rd1501, %rd1504; st.global.f32 [%rd1505], %f2984; st.global.f32 [%rd1505+4], %f2985; st.global.f32 [%rd1505+8], %f2986; $L__BB215_1032: ld.const.u64 %rd196, [kernel_params+2752]; setp.eq.s64 %p761, %rd196, 0; @%p761 bra $L__BB215_1034; cvta.to.global.u64 %rd1506, %rd196; add.s64 %rd1508, %rd1506, %rd260; ld.global.f32 %f2987, [%rd1508]; shl.b64 %rd1509, %rd180, 2; add.s64 %rd1510, %rd1506, %rd1509; st.global.f32 [%rd1510], %f2987; $L__BB215_1034: ld.const.u64 %rd197, [kernel_params+2760]; setp.eq.s64 %p762, %rd197, 0; @%p762 bra $L__BB215_1036; cvta.to.global.u64 %rd1511, %rd197; add.s64 %rd1513, %rd1511, %rd339; ld.global.f32 %f2988, [%rd1513]; ld.global.f32 %f2989, [%rd1513+4]; ld.global.f32 %f2990, [%rd1513+8]; mul.lo.s64 %rd1514, %rd180, 12; add.s64 %rd1515, %rd1511, %rd1514; st.global.f32 [%rd1515], %f2988; st.global.f32 [%rd1515+4], %f2989; st.global.f32 [%rd1515+8], %f2990; $L__BB215_1036: ld.const.u32 %r2105, [kernel_params+12]; max.u32 %r466, %r2105, 1; add.s32 %r2106, %r466, -1; and.b32 %r2244, %r466, 3; setp.lt.u32 %p763, %r2106, 3; mov.u32 %r2243, 0; @%p763 bra $L__BB215_1055; sub.s32 %r2242, %r466, %r2244; mov.u32 %r2243, 0; mov.u64 %rd1517, kernel_params; $L__BB215_1038: cvt.s64.s32 %rd198, %r2243; mul.wide.s32 %rd1516, %r2243, 16; add.s64 %rd1518, %rd1517, %rd1516; add.s64 %rd199, %rd1518, 2768; ld.const.u64 %rd200, [%rd1518+2768]; setp.eq.s64 %p764, %rd200, 0; @%p764 bra $L__BB215_1040; cvta.to.global.u64 %rd1519, %rd200; add.s64 %rd1521, %rd1519, %rd260; ld.global.u32 %r2108, [%rd1521]; shl.b64 %rd1522, %rd180, 2; add.s64 %rd1523, %rd1519, %rd1522; st.global.u32 [%rd1523], %r2108; $L__BB215_1040: ld.const.u64 %rd201, [%rd199+8]; setp.eq.s64 %p765, %rd201, 0; @%p765 bra $L__BB215_1042; cvta.to.global.u64 %rd1524, %rd201; add.s64 %rd1526, %rd1524, %rd260; ld.global.u32 %r2109, [%rd1526]; shl.b64 %rd1527, %rd180, 2; add.s64 %rd1528, %rd1524, %rd1527; st.global.u32 [%rd1528], %r2109; $L__BB215_1042: ld.const.u64 %rd202, [%rd199+16]; setp.eq.s64 %p766, %rd202, 0; @%p766 bra $L__BB215_1044; cvta.to.global.u64 %rd1529, %rd202; add.s64 %rd1531, %rd1529, %rd260; ld.global.u32 %r2110, [%rd1531]; shl.b64 %rd1532, %rd180, 2; add.s64 %rd1533, %rd1529, %rd1532; st.global.u32 [%rd1533], %r2110; $L__BB215_1044: ld.const.u64 %rd203, [%rd199+24]; setp.eq.s64 %p767, %rd203, 0; @%p767 bra $L__BB215_1046; cvta.to.global.u64 %rd1534, %rd203; add.s64 %rd1536, %rd1534, %rd260; ld.global.u32 %r2111, [%rd1536]; shl.b64 %rd1537, %rd180, 2; add.s64 %rd1538, %rd1534, %rd1537; st.global.u32 [%rd1538], %r2111; $L__BB215_1046: ld.const.u64 %rd204, [%rd199+32]; setp.eq.s64 %p768, %rd204, 0; @%p768 bra $L__BB215_1048; cvta.to.global.u64 %rd1539, %rd204; add.s64 %rd1541, %rd1539, %rd260; ld.global.u32 %r2112, [%rd1541]; shl.b64 %rd1542, %rd180, 2; add.s64 %rd1543, %rd1539, %rd1542; st.global.u32 [%rd1543], %r2112; $L__BB215_1048: ld.const.u64 %rd205, [%rd199+40]; setp.eq.s64 %p769, %rd205, 0; @%p769 bra $L__BB215_1050; cvta.to.global.u64 %rd1544, %rd205; add.s64 %rd1546, %rd1544, %rd260; ld.global.u32 %r2113, [%rd1546]; shl.b64 %rd1547, %rd180, 2; add.s64 %rd1548, %rd1544, %rd1547; st.global.u32 [%rd1548], %r2113; $L__BB215_1050: ld.const.u64 %rd206, [%rd199+48]; setp.eq.s64 %p770, %rd206, 0; @%p770 bra $L__BB215_1052; cvta.to.global.u64 %rd1549, %rd206; add.s64 %rd1551, %rd1549, %rd260; ld.global.u32 %r2114, [%rd1551]; shl.b64 %rd1552, %rd180, 2; add.s64 %rd1553, %rd1549, %rd1552; st.global.u32 [%rd1553], %r2114; $L__BB215_1052: ld.const.u64 %rd207, [%rd199+56]; setp.eq.s64 %p771, %rd207, 0; @%p771 bra $L__BB215_1054; cvta.to.global.u64 %rd1554, %rd207; add.s64 %rd1556, %rd1554, %rd260; ld.global.u32 %r2115, [%rd1556]; shl.b64 %rd1557, %rd180, 2; add.s64 %rd1558, %rd1554, %rd1557; st.global.u32 [%rd1558], %r2115; $L__BB215_1054: cvt.u32.u64 %r2116, %rd198; add.s32 %r2243, %r2116, 4; add.s32 %r2242, %r2242, -4; setp.ne.s32 %p772, %r2242, 0; @%p772 bra $L__BB215_1038; $L__BB215_1055: setp.eq.s32 %p773, %r2244, 0; @%p773 bra $L__BB215_1062; mul.wide.s32 %rd1559, %r2243, 16; mov.u64 %rd1560, kernel_params; add.s64 %rd1561, %rd1560, %rd1559; add.s64 %rd1752, %rd1561, 2776; $L__BB215_1057: .pragma "nounroll"; ld.const.u64 %rd211, [%rd1752+-8]; setp.eq.s64 %p774, %rd211, 0; @%p774 bra $L__BB215_1059; cvta.to.global.u64 %rd1562, %rd211; add.s64 %rd1564, %rd1562, %rd260; ld.global.u32 %r2117, [%rd1564]; shl.b64 %rd1565, %rd180, 2; add.s64 %rd1566, %rd1562, %rd1565; st.global.u32 [%rd1566], %r2117; $L__BB215_1059: ld.const.u64 %rd212, [%rd1752]; setp.eq.s64 %p775, %rd212, 0; @%p775 bra $L__BB215_1061; cvta.to.global.u64 %rd1567, %rd212; add.s64 %rd1569, %rd1567, %rd260; ld.global.u32 %r2118, [%rd1569]; shl.b64 %rd1570, %rd180, 2; add.s64 %rd1571, %rd1567, %rd1570; st.global.u32 [%rd1571], %r2118; $L__BB215_1061: add.s32 %r2244, %r2244, -1; add.s64 %rd1752, %rd1752, 16; setp.ne.s32 %p776, %r2244, 0; @%p776 bra $L__BB215_1057; $L__BB215_1062: ld.const.u64 %rd214, [kernel_params+3280]; setp.eq.s64 %p777, %rd214, 0; @%p777 bra $L__BB215_1064; cvta.to.global.u64 %rd1572, %rd214; shl.b64 %rd1573, %rd4, 3; add.s64 %rd1574, %rd1572, %rd1573; ld.global.u64 %rd1575, [%rd1574]; shl.b64 %rd1576, %rd180, 3; add.s64 %rd1577, %rd1572, %rd1576; st.global.u64 [%rd1577], %rd1575; $L__BB215_1064: ld.const.u64 %rd215, [kernel_params+3288]; setp.eq.s64 %p778, %rd215, 0; @%p778 bra $L__BB215_1066; cvta.to.global.u64 %rd1578, %rd215; add.s64 %rd1579, %rd1578, %rd4; ld.global.u8 %rs95, [%rd1579]; add.s64 %rd1580, %rd1578, %rd180; st.global.u8 [%rd1580], %rs95; $L__BB215_1066: ld.const.u64 %rd216, [kernel_params+3296]; setp.eq.s64 %p779, %rd216, 0; @%p779 bra $L__BB215_1068; cvta.to.global.u64 %rd1581, %rd216; add.s64 %rd1583, %rd1581, %rd260; ld.global.f32 %f2991, [%rd1583]; shl.b64 %rd1584, %rd180, 2; add.s64 %rd1585, %rd1581, %rd1584; st.global.f32 [%rd1585], %f2991; $L__BB215_1068: ld.const.u64 %rd217, [kernel_params+3304]; setp.eq.s64 %p780, %rd217, 0; @%p780 bra $L__BB215_1070; cvta.to.global.u64 %rd1586, %rd217; add.s64 %rd1588, %rd1586, %rd260; ld.global.f32 %f2992, [%rd1588]; shl.b64 %rd1589, %rd180, 2; add.s64 %rd1590, %rd1586, %rd1589; st.global.f32 [%rd1590], %f2992; $L__BB215_1070: ld.const.u64 %rd218, [kernel_params+3312]; setp.eq.s64 %p781, %rd218, 0; @%p781 bra $L__BB215_1072; cvta.to.global.u64 %rd1591, %rd218; add.s64 %rd1593, %rd1591, %rd260; ld.global.f32 %f2993, [%rd1593]; shl.b64 %rd1594, %rd180, 2; add.s64 %rd1595, %rd1591, %rd1594; st.global.f32 [%rd1595], %f2993; $L__BB215_1072: ld.const.u64 %rd219, [kernel_params+3320]; setp.eq.s64 %p782, %rd219, 0; @%p782 bra $L__BB215_1074; cvta.to.global.u64 %rd1596, %rd219; add.s64 %rd1597, %rd1596, %rd4; ld.global.u8 %rs96, [%rd1597]; add.s64 %rd1598, %rd1596, %rd180; st.global.u8 [%rd1598], %rs96; $L__BB215_1074: ld.const.u64 %rd220, [kernel_params+3328]; setp.eq.s64 %p783, %rd220, 0; @%p783 bra $L__BB215_1076; cvta.to.global.u64 %rd1599, %rd220; add.s64 %rd1601, %rd1599, %rd260; ld.global.f32 %f2994, [%rd1601]; shl.b64 %rd1602, %rd180, 2; add.s64 %rd1603, %rd1599, %rd1602; st.global.f32 [%rd1603], %f2994; $L__BB215_1076: ld.const.u64 %rd221, [kernel_params+3336]; setp.eq.s64 %p784, %rd221, 0; @%p784 bra $L__BB215_1078; cvta.to.global.u64 %rd1604, %rd221; add.s64 %rd1606, %rd1604, %rd260; ld.global.f32 %f2995, [%rd1606]; shl.b64 %rd1607, %rd180, 2; add.s64 %rd1608, %rd1604, %rd1607; st.global.f32 [%rd1608], %f2995; $L__BB215_1078: ld.global.u32 %r476, [%rd181]; or.b32 %r2119, %r476, 1073741824; st.global.u32 [%rd181], %r2119; ld.const.u32 %r2120, [kernel_params+1356]; setp.ne.s32 %p785, %r2120, -1; ld.const.u32 %r2121, [kernel_params+1036]; setp.eq.s32 %p786, %r2121, 0; and.pred %p787, %p785, %p786; @%p787 bra $L__BB215_1093; bra.uni $L__BB215_1079; $L__BB215_1093: or.b32 %r2138, %r476, -1073741824; st.global.u32 [%rd181], %r2138; cvta.to.global.u64 %rd1665, %rd168; add.s64 %rd1666, %rd1665, 24; atom.global.add.u32 %r2139, [%rd1666], 1; cvta.to.global.u64 %rd1667, %rd170; shl.b64 %rd1668, %rd180, 1; add.s64 %rd1669, %rd1667, %rd1668; mov.u16 %rs105, 6; st.global.u16 [%rd1669], %rs105; bra.uni $L__BB215_1097; $L__BB215_1079: ld.const.u8 %rs97, [kernel_params+2]; and.b16 %rs98, %rs97, 2; setp.eq.s16 %p788, %rs98, 0; @%p788 bra $L__BB215_1082; ld.const.u64 %rd1741, [kernel_params+2776]; cvta.to.global.u64 %rd1740, %rd1741; shl.b64 %rd1609, %rd180, 2; add.s64 %rd1610, %rd1740, %rd1609; ld.global.u32 %r2122, [%rd1610]; setp.eq.s32 %p789, %r2122, -1; @%p789 bra $L__BB215_1082; cvta.to.global.u64 %rd1611, %rd168; add.s64 %rd1612, %rd1611, 20; atom.global.add.u32 %r2123, [%rd1612], 1; cvta.to.global.u64 %rd1613, %rd170; shl.b64 %rd1614, %rd180, 1; add.s64 %rd1615, %rd1613, %rd1614; mov.u16 %rs99, 5; st.global.u16 [%rd1615], %rs99; bra.uni $L__BB215_1097; $L__BB215_579: ld.local.u32 %r1687, [%rd31+60]; setp.eq.s32 %p469, %r1687, -1; @%p469 bra $L__BB215_832; ld.local.f32 %f2226, [%rd31+24]; ld.local.f32 %f582, [%rd31+12]; ld.local.f32 %f2227, [%rd31+28]; ld.local.f32 %f583, [%rd31+16]; mul.ftz.f32 %f2228, %f583, %f2227; fma.rn.ftz.f32 %f2229, %f582, %f2226, %f2228; ld.local.f32 %f2230, [%rd31+32]; ld.local.f32 %f584, [%rd31+20]; fma.rn.ftz.f32 %f2231, %f584, %f2230, %f2229; setp.leu.ftz.f32 %p470, %f2231, 0f00000000; @%p470 bra $L__BB215_832; neg.ftz.f32 %f2232, %f582; st.local.f32 [%rd31+12], %f2232; neg.ftz.f32 %f2233, %f583; st.local.f32 [%rd31+16], %f2233; neg.ftz.f32 %f2234, %f584; st.local.f32 [%rd31+20], %f2234; bra.uni $L__BB215_832; $L__BB215_1082: @%p710 bra $L__BB215_1084; bra.uni $L__BB215_1083; $L__BB215_1084: and.b32 %r2124, %r5, 8; setp.eq.s32 %p791, %r2124, 0; @%p791 bra $L__BB215_1086; bra.uni $L__BB215_1085; $L__BB215_1086: and.b32 %r2126, %r5, 6; setp.eq.s32 %p792, %r2126, 0; mov.u32 %r2245, 0; @%p792 bra $L__BB215_1088; ld.const.u64 %rd1624, [kernel_params+2264]; cvta.to.global.u64 %rd1625, %rd1624; mul.wide.s32 %rd1626, %r2067, 16; add.s64 %rd1627, %rd1625, %rd1626; ld.global.u32 %r2245, [%rd1627]; bra.uni $L__BB215_1088; $L__BB215_583: add.s64 %rd714, %rd107, 20; add.s64 %rd715, %rd107, 4; add.s64 %rd716, %rd107, 36; ld.local.f32 %f3252, [%rd31+12]; ld.local.f32 %f3253, [%rd31+16]; ld.local.f32 %f590, [%rd31+20]; ld.local.f32 %f2235, [%rd31+24]; neg.ftz.f32 %f602, %f2235; ld.local.f32 %f2236, [%rd31+28]; neg.ftz.f32 %f603, %f2236; ld.local.f32 %f2237, [%rd31+32]; neg.ftz.f32 %f604, %f2237; ld.local.u32 %r301, [%rd31+56]; cvt.s64.s32 %rd108, %r301; ld.local.u32 %r302, [%rd31+60]; cvt.s64.s32 %rd109, %r302; ld.local.f32 %f594, [%rd31+40]; ld.local.f32 %f595, [%rd31+44]; ld.local.f32 %f596, [%rd31+36]; ld.local.u32 %r303, [%rd31+68]; cvta.to.local.u64 %rd720, %rd713; st.local.v4.f32 [%rd720], {%f585, %f586, %f587, %f3252}; cvta.to.local.u64 %rd110, %rd715; st.local.v4.f32 [%rd110], {%f3253, %f590, %f3252, %f3253}; cvta.to.local.u64 %rd721, %rd714; st.local.v4.f32 [%rd721], {%f590, %f602, %f603, %f604}; cvta.to.local.u64 %rd722, %rd716; st.local.u32 [%rd722], %r297; setp.eq.s32 %p472, %r302, -1; add.s64 %rd111, %rd713, 64; @%p472 bra $L__BB215_585; cvta.to.local.u64 %rd723, %rd111; mov.u32 %r2217, 1; st.local.u32 [%rd723], %r2217; bra.uni $L__BB215_588; $L__BB215_1083: ld.const.u64 %rd1616, [kernel_params+2216]; cvta.to.global.u64 %rd1617, %rd1616; shl.b64 %rd1618, %rd16, 2; add.s64 %rd1619, %rd1617, %rd1618; ld.global.u32 %r2245, [%rd1619]; bra.uni $L__BB215_1088; $L__BB215_544: ld.const.u32 %r2212, [kernel_params+1332]; bra.uni $L__BB215_546; $L__BB215_1085: ld.const.u64 %rd1620, [kernel_params+2304]; cvta.to.global.u64 %rd1621, %rd1620; shl.b64 %rd1622, %rd16, 2; add.s64 %rd1623, %rd1621, %rd1622; ld.global.u32 %r2245, [%rd1623]; $L__BB215_1088: and.b32 %r2128, %r2245, 4194303; shl.b32 %r2129, %r2245, 5; cvt.u64.u32 %rd1628, %r2129; and.b64 %rd1629, %rd1628, 134217696; add.s64 %rd1630, %rd156, %rd1629; add.s64 %rd222, %rd1630, 16; and.b32 %r2130, %r463, 1536; setp.ne.s32 %p793, %r2130, 0; and.pred %p795, %p793, %p713; div.u32 %r2131, %r465, %r2239; mad.lo.s32 %r481, %r2131, %r2240, %r2128; cvt.s64.s32 %rd223, %r481; @%p795 bra $L__BB215_1092; bra.uni $L__BB215_1089; $L__BB215_1092: cvta.to.global.u64 %rd1653, %rd168; add.s64 %rd1654, %rd1653, 40; atom.global.add.u32 %r2136, [%rd1654], 1; cvta.to.global.u64 %rd1655, %rd170; shl.b64 %rd1656, %rd180, 1; add.s64 %rd1657, %rd1655, %rd1656; mov.u16 %rs104, 10; st.global.u16 [%rd1657], %rs104; cvta.to.global.u64 %rd1658, %rd169; shl.b64 %rd1659, %rd180, 2; add.s64 %rd1660, %rd1658, %rd1659; st.global.u32 [%rd1660], %r481; ld.const.u64 %rd1661, [kernel_params+4344]; cvta.to.global.u64 %rd1662, %rd1661; shl.b64 %rd1663, %rd223, 2; add.s64 %rd1664, %rd1662, %rd1663; atom.global.add.u32 %r2137, [%rd1664], 1; bra.uni $L__BB215_1097; $L__BB215_1089: ld.global.u8 %rs100, [%rd222+3]; and.b16 %rs101, %rs100, 64; setp.eq.s16 %p796, %rs101, 0; cvta.to.global.u64 %rd224, %rd168; @%p796 bra $L__BB215_1091; bra.uni $L__BB215_1090; $L__BB215_1091: add.s64 %rd1642, %rd224, 32; atom.global.add.u32 %r2134, [%rd1642], 1; cvta.to.global.u64 %rd1643, %rd170; shl.b64 %rd1644, %rd180, 1; add.s64 %rd1645, %rd1643, %rd1644; mov.u16 %rs103, 8; st.global.u16 [%rd1645], %rs103; cvta.to.global.u64 %rd1646, %rd169; shl.b64 %rd1647, %rd180, 2; add.s64 %rd1648, %rd1646, %rd1647; st.global.u32 [%rd1648], %r481; ld.const.u64 %rd1649, [kernel_params+4328]; cvta.to.global.u64 %rd1650, %rd1649; shl.b64 %rd1651, %rd223, 2; add.s64 %rd1652, %rd1650, %rd1651; atom.global.add.u32 %r2135, [%rd1652], 1; bra.uni $L__BB215_1097; $L__BB215_1090: add.s64 %rd1631, %rd224, 36; atom.global.add.u32 %r2132, [%rd1631], 1; cvta.to.global.u64 %rd1632, %rd170; shl.b64 %rd1633, %rd180, 1; add.s64 %rd1634, %rd1632, %rd1633; mov.u16 %rs102, 9; st.global.u16 [%rd1634], %rs102; cvta.to.global.u64 %rd1635, %rd169; shl.b64 %rd1636, %rd180, 2; add.s64 %rd1637, %rd1635, %rd1636; st.global.u32 [%rd1637], %r481; ld.const.u64 %rd1638, [kernel_params+4336]; cvta.to.global.u64 %rd1639, %rd1638; shl.b64 %rd1640, %rd223, 2; add.s64 %rd1641, %rd1639, %rd1640; atom.global.add.u32 %r2133, [%rd1641], 1; $L__BB215_1097: ret; $L__BB215_585: cvta.to.local.u64 %rd112, %rd111; setp.eq.s32 %p473, %r303, -1; @%p473 bra $L__BB215_587; mov.u32 %r2217, 32; st.local.u32 [%rd112], %r2217; bra.uni $L__BB215_588; $L__BB215_587: mov.u32 %r2217, 0; st.local.u32 [%rd112], %r2217; $L__BB215_588: add.s64 %rd725, %rd713, 76; cvta.to.local.u64 %rd726, %rd725; st.local.u32 [%rd726], %r301; add.s64 %rd727, %rd713, 60; cvta.to.local.u64 %rd728, %rd727; st.local.u32 [%rd728], %r302; add.s64 %rd729, %rd713, 68; cvta.to.local.u64 %rd730, %rd729; st.local.f32 [%rd730], %f594; add.s64 %rd731, %rd713, 72; cvta.to.local.u64 %rd732, %rd731; st.local.f32 [%rd732], %f595; mov.b32 %r1691, %f575; mov.u32 %r1692, -1; add.s64 %rd113, %rd713, 80; cvta.to.local.u64 %rd733, %rd113; st.local.v2.u32 [%rd733], {%r1692, %r1691}; add.s64 %rd734, %rd713, 88; cvta.to.local.u64 %rd735, %rd734; st.local.f32 [%rd735], %f596; add.s64 %rd736, %rd713, 52; cvta.to.local.u64 %rd114, %rd736; st.local.u32 [%rd114], %r299; add.s64 %rd737, %rd713, 56; cvta.to.local.u64 %rd115, %rd737; mov.u32 %r1693, 0; st.local.u32 [%rd115], %r1693; setp.eq.s32 %p474, %r301, -1; @%p474 bra $L__BB215_613; ld.const.u64 %rd738, [kernel_params+2184]; cvta.to.global.u64 %rd739, %rd738; shl.b64 %rd740, %rd108, 2; add.s64 %rd741, %rd739, %rd740; ld.global.u32 %r305, [%rd741]; st.local.u32 [%rd115], %r305; and.b32 %r306, %r305, 2; setp.eq.s32 %p475, %r306, 0; @%p475 bra $L__BB215_596; ld.const.u64 %rd742, [kernel_params+2160]; cvta.to.global.u64 %rd743, %rd742; mul.lo.s64 %rd744, %rd108, 208; add.s64 %rd745, %rd743, %rd744; ld.const.u64 %rd746, [kernel_params+2176]; cvta.to.global.u64 %rd747, %rd746; ld.global.u32 %rd748, [%rd745+168]; ld.global.u32 %r1694, [%rd745+152]; shl.b32 %r1695, %r1694, 1; cvt.rn.f32.s32 %f2239, %r1695; mul.ftz.f32 %f2240, %f575, %f2239; cvt.rzi.ftz.s32.f32 %r1696, %f2240; add.s32 %r1697, %r1695, -1; min.s32 %r1698, %r1696, %r1697; cvt.rn.f32.s32 %f2241, %r1698; sub.ftz.f32 %f605, %f2240, %f2241; cvt.s64.s32 %rd749, %r1698; add.s64 %rd750, %rd749, %rd748; shl.b64 %rd751, %rd750, 6; add.s64 %rd116, %rd747, %rd751; ld.global.v4.f32 {%f2242, %f2243, %f2244, %f2245}, [%rd116]; ld.global.v4.f32 {%f2246, %f2247, %f2248, %f2249}, [%rd116+64]; mul.ftz.f32 %f2250, %f2243, %f2247; fma.rn.ftz.f32 %f2251, %f2242, %f2246, %f2250; mul.ftz.f32 %f2252, %f2245, %f2249; fma.rn.ftz.f32 %f2253, %f2244, %f2248, %f2252; add.ftz.f32 %f614, %f2251, %f2253; setp.gt.ftz.f32 %p476, %f614, 0f3F7FDF3B; mov.f32 %f2254, 0f3F800000; sub.ftz.f32 %f615, %f2254, %f605; @%p476 bra $L__BB215_592; bra.uni $L__BB215_591; $L__BB215_592: mul.ftz.f32 %f2308, %f2242, %f615; fma.rn.ftz.f32 %f2309, %f2246, %f605, %f2308; mul.ftz.f32 %f2310, %f2243, %f615; fma.rn.ftz.f32 %f2311, %f2247, %f605, %f2310; mul.ftz.f32 %f2312, %f2244, %f615; fma.rn.ftz.f32 %f2313, %f2248, %f605, %f2312; mul.ftz.f32 %f2314, %f2245, %f615; fma.rn.ftz.f32 %f2315, %f2249, %f605, %f2314; mul.ftz.f32 %f2316, %f2311, %f2311; fma.rn.ftz.f32 %f2317, %f2309, %f2309, %f2316; mul.ftz.f32 %f2318, %f2315, %f2315; fma.rn.ftz.f32 %f2319, %f2313, %f2313, %f2318; add.ftz.f32 %f2320, %f2317, %f2319; rsqrt.approx.ftz.f32 %f2321, %f2320; mul.ftz.f32 %f3207, %f2315, %f2321; mul.ftz.f32 %f3206, %f2313, %f2321; mul.ftz.f32 %f3205, %f2311, %f2321; mul.ftz.f32 %f3204, %f2309, %f2321; bra.uni $L__BB215_593; $L__BB215_613: setp.eq.s32 %p494, %r303, -1; @%p494 bra $L__BB215_615; st.local.u32 [%rd733], %r303; $L__BB215_615: add.s64 %rd823, %rd713, 116; cvta.to.local.u64 %rd824, %rd823; mov.u32 %r1712, 0; st.local.u32 [%rd824], %r1712; add.s64 %rd825, %rd713, 120; cvta.to.local.u64 %rd826, %rd825; st.local.u32 [%rd826], %r1712; add.s64 %rd827, %rd713, 124; cvta.to.local.u64 %rd828, %rd827; st.local.u32 [%rd828], %r1712; add.s64 %rd829, %rd713, 136; cvta.to.local.u64 %rd830, %rd829; add.s64 %rd831, %rd713, 128; cvta.to.local.u64 %rd832, %rd831; mov.f32 %f3242, 0f00000000; st.local.v2.f32 [%rd832], {%f3242, %f3242}; st.local.u32 [%rd830], %r1712; mov.f32 %f3243, %f3242; mov.f32 %f3244, %f3242; mov.f32 %f3245, %f3242; mov.f32 %f3246, %f3242; mov.f32 %f3247, %f3242; mov.f32 %f3248, %f590; mov.f32 %f3249, %f3253; mov.f32 %f3250, %f3252; bra.uni $L__BB215_616; $L__BB215_591: mov.f32 %f2255, 0fBF800000; max.ftz.f32 %f2256, %f614, %f2255; min.ftz.f32 %f2258, %f2256, %f2254; abs.ftz.f32 %f2259, %f2258; sub.ftz.f32 %f2260, %f2254, %f2259; mul.ftz.f32 %f2261, %f2260, 0f3F000000; sqrt.approx.ftz.f32 %f2262, %f2261; setp.gt.ftz.f32 %p477, %f2259, 0f3F11EB85; selp.f32 %f2263, %f2262, %f2259, %p477; mul.ftz.f32 %f2264, %f2263, %f2263; mov.f32 %f2265, 0f3C94D2E9; mov.f32 %f2266, 0f3D53F941; fma.rn.ftz.f32 %f2267, %f2266, %f2264, %f2265; mov.f32 %f2268, 0f3D3F841F; fma.rn.ftz.f32 %f2269, %f2267, %f2264, %f2268; mov.f32 %f2270, 0f3D994929; fma.rn.ftz.f32 %f2271, %f2269, %f2264, %f2270; mov.f32 %f2272, 0f3E2AAB94; fma.rn.ftz.f32 %f2273, %f2271, %f2264, %f2272; mul.ftz.f32 %f2274, %f2264, %f2273; fma.rn.ftz.f32 %f2275, %f2274, %f2263, %f2263; add.ftz.f32 %f2276, %f2275, %f2275; mov.f32 %f2277, 0f3FC90FDB; sub.ftz.f32 %f2278, %f2277, %f2275; selp.f32 %f2279, %f2276, %f2278, %p477; setp.lt.ftz.f32 %p478, %f2258, 0f00000000; mov.f32 %f2280, 0f40490FDB; sub.ftz.f32 %f2281, %f2280, %f2279; selp.f32 %f2282, %f2281, %f2279, %p478; mul.ftz.f32 %f2283, %f2242, %f614; sub.ftz.f32 %f2284, %f2246, %f2283; mul.ftz.f32 %f2285, %f2243, %f614; sub.ftz.f32 %f2286, %f2247, %f2285; mul.ftz.f32 %f2287, %f2244, %f614; sub.ftz.f32 %f2288, %f2248, %f2287; mul.ftz.f32 %f2289, %f2245, %f614; sub.ftz.f32 %f2290, %f2249, %f2289; mul.ftz.f32 %f2291, %f2286, %f2286; fma.rn.ftz.f32 %f2292, %f2284, %f2284, %f2291; mul.ftz.f32 %f2293, %f2290, %f2290; fma.rn.ftz.f32 %f2294, %f2288, %f2288, %f2293; add.ftz.f32 %f2295, %f2292, %f2294; rsqrt.approx.ftz.f32 %f2296, %f2295; mul.ftz.f32 %f2297, %f2284, %f2296; mul.ftz.f32 %f2298, %f2286, %f2296; mul.ftz.f32 %f2299, %f2288, %f2296; mul.ftz.f32 %f2300, %f2290, %f2296; mul.ftz.f32 %f2301, %f605, %f2282; cos.approx.ftz.f32 %f2302, %f2301; sin.approx.ftz.f32 %f2303, %f2301; mul.ftz.f32 %f2304, %f2297, %f2303; mul.ftz.f32 %f2305, %f2298, %f2303; mul.ftz.f32 %f2306, %f2299, %f2303; mul.ftz.f32 %f2307, %f2300, %f2303; fma.rn.ftz.f32 %f3207, %f2245, %f2302, %f2307; fma.rn.ftz.f32 %f3206, %f2244, %f2302, %f2306; fma.rn.ftz.f32 %f3205, %f2243, %f2302, %f2305; fma.rn.ftz.f32 %f3204, %f2242, %f2302, %f2304; $L__BB215_593: ld.global.v4.f32 {%f2322, %f2323, %f2324, %f2325}, [%rd116+16]; ld.global.v4.f32 {%f2330, %f2331, %f2332, %f2333}, [%rd116+80]; mul.ftz.f32 %f2338, %f605, %f2330; mul.ftz.f32 %f2339, %f605, %f2331; mul.ftz.f32 %f2340, %f605, %f2332; mul.ftz.f32 %f2341, %f605, %f2333; fma.rn.ftz.f32 %f628, %f615, %f2322, %f2338; fma.rn.ftz.f32 %f629, %f615, %f2323, %f2339; fma.rn.ftz.f32 %f630, %f615, %f2324, %f2340; fma.rn.ftz.f32 %f2342, %f615, %f2325, %f2341; ld.global.v4.f32 {%f2343, %f2344, %f2345, %f2346}, [%rd116+32]; ld.global.v4.f32 {%f2351, %f2352, %f2353, %f2354}, [%rd116+96]; mul.ftz.f32 %f2359, %f605, %f2351; mul.ftz.f32 %f2360, %f605, %f2352; mul.ftz.f32 %f2361, %f605, %f2353; mul.ftz.f32 %f2362, %f605, %f2354; fma.rn.ftz.f32 %f2363, %f615, %f2343, %f2359; fma.rn.ftz.f32 %f2364, %f615, %f2344, %f2360; fma.rn.ftz.f32 %f2365, %f615, %f2345, %f2361; fma.rn.ftz.f32 %f2366, %f615, %f2346, %f2362; ld.global.v4.f32 {%f2367, %f2368, %f2369, %f2370}, [%rd116+48]; ld.global.v4.f32 {%f2375, %f2376, %f2377, %f2378}, [%rd116+112]; mul.ftz.f32 %f2383, %f605, %f2375; mul.ftz.f32 %f2384, %f605, %f2376; mul.ftz.f32 %f2385, %f605, %f2377; mul.ftz.f32 %f2386, %f605, %f2378; fma.rn.ftz.f32 %f2387, %f615, %f2367, %f2383; fma.rn.ftz.f32 %f2388, %f615, %f2368, %f2384; fma.rn.ftz.f32 %f2389, %f615, %f2369, %f2385; fma.rn.ftz.f32 %f2390, %f615, %f2370, %f2386; mul.ftz.f32 %f2391, %f3207, 0f3FB504F3; mul.ftz.f32 %f2392, %f3204, 0f3FB504F3; mul.ftz.f32 %f2393, %f3205, 0f3FB504F3; mul.ftz.f32 %f2394, %f3206, 0f3FB504F3; mul.ftz.f32 %f2395, %f2391, %f2392; mul.ftz.f32 %f2396, %f2391, %f2393; mul.ftz.f32 %f2397, %f2391, %f2394; mul.ftz.f32 %f2398, %f2392, %f2392; mul.ftz.f32 %f2399, %f2392, %f2393; mul.ftz.f32 %f2400, %f2392, %f2394; mul.ftz.f32 %f2401, %f2393, %f2393; mul.ftz.f32 %f2402, %f2393, %f2394; mul.ftz.f32 %f2403, %f2394, %f2394; sub.ftz.f32 %f2405, %f2254, %f2401; sub.ftz.f32 %f2406, %f2405, %f2403; sub.ftz.f32 %f2407, %f2399, %f2397; add.ftz.f32 %f2408, %f2396, %f2400; add.ftz.f32 %f2409, %f2397, %f2399; sub.ftz.f32 %f2410, %f2254, %f2398; sub.ftz.f32 %f2411, %f2410, %f2403; sub.ftz.f32 %f2412, %f2402, %f2395; sub.ftz.f32 %f2413, %f2400, %f2396; add.ftz.f32 %f2414, %f2395, %f2402; sub.ftz.f32 %f2415, %f2410, %f2401; mul.ftz.f32 %f2416, %f2407, %f2365; fma.rn.ftz.f32 %f2417, %f2406, %f2342, %f2416; fma.rn.ftz.f32 %f3223, %f2408, %f2388, %f2417; mul.ftz.f32 %f2418, %f2407, %f2366; fma.rn.ftz.f32 %f2419, %f2406, %f2363, %f2418; fma.rn.ftz.f32 %f3222, %f2408, %f2389, %f2419; mul.ftz.f32 %f2420, %f2407, %f2387; fma.rn.ftz.f32 %f2421, %f2406, %f2364, %f2420; fma.rn.ftz.f32 %f3221, %f2408, %f2390, %f2421; mul.ftz.f32 %f2422, %f2411, %f2365; fma.rn.ftz.f32 %f2423, %f2409, %f2342, %f2422; fma.rn.ftz.f32 %f3220, %f2412, %f2388, %f2423; mul.ftz.f32 %f2424, %f2411, %f2366; fma.rn.ftz.f32 %f2425, %f2409, %f2363, %f2424; fma.rn.ftz.f32 %f3219, %f2412, %f2389, %f2425; mul.ftz.f32 %f2426, %f2411, %f2387; fma.rn.ftz.f32 %f2427, %f2409, %f2364, %f2426; fma.rn.ftz.f32 %f3218, %f2412, %f2390, %f2427; mul.ftz.f32 %f2428, %f2414, %f2365; fma.rn.ftz.f32 %f2429, %f2413, %f2342, %f2428; fma.rn.ftz.f32 %f3217, %f2415, %f2388, %f2429; mul.ftz.f32 %f2430, %f2414, %f2366; fma.rn.ftz.f32 %f2431, %f2413, %f2363, %f2430; fma.rn.ftz.f32 %f3216, %f2415, %f2389, %f2431; mul.ftz.f32 %f2432, %f2414, %f2387; fma.rn.ftz.f32 %f2433, %f2413, %f2364, %f2432; fma.rn.ftz.f32 %f3215, %f2415, %f2390, %f2433; add.s64 %rd753, %rd713, 144; cvta.to.local.u64 %rd754, %rd753; st.local.v4.f32 [%rd754], {%f3223, %f3222, %f3221, %f628}; add.s64 %rd755, %rd713, 160; cvta.to.local.u64 %rd756, %rd755; st.local.v4.f32 [%rd756], {%f3220, %f3219, %f3218, %f629}; add.s64 %rd757, %rd713, 176; cvta.to.local.u64 %rd758, %rd757; st.local.v4.f32 [%rd758], {%f3217, %f3216, %f3215, %f630}; mul.ftz.f32 %f2434, %f3219, %f3215; mul.ftz.f32 %f640, %f3216, %f3218; sub.ftz.f32 %f3210, %f2434, %f640; mul.ftz.f32 %f642, %f3216, %f3221; mul.ftz.f32 %f2435, %f3222, %f3215; sub.ftz.f32 %f3209, %f642, %f2435; mul.ftz.f32 %f644, %f3222, %f3218; mul.ftz.f32 %f2436, %f3219, %f3221; sub.ftz.f32 %f3208, %f644, %f2436; mul.ftz.f32 %f2437, %f3220, %f3209; fma.rn.ftz.f32 %f2438, %f3223, %f3210, %f2437; fma.rn.ftz.f32 %f3214, %f3217, %f3208, %f2438; setp.neu.ftz.f32 %p479, %f3214, 0f00000000; mov.f32 %f3211, %f3215; mov.f32 %f3212, %f3219; mov.f32 %f3213, %f3223; @%p479 bra $L__BB215_595; add.ftz.f32 %f3213, %f3223, 0f322BCC77; add.ftz.f32 %f3211, %f3215, 0f322BCC77; add.ftz.f32 %f3212, %f3219, 0f322BCC77; mul.ftz.f32 %f2439, %f3212, %f3211; sub.ftz.f32 %f3210, %f2439, %f640; mul.ftz.f32 %f2440, %f3222, %f3211; sub.ftz.f32 %f3209, %f642, %f2440; mul.ftz.f32 %f2441, %f3212, %f3221; sub.ftz.f32 %f3208, %f644, %f2441; mul.ftz.f32 %f2442, %f3220, %f3209; fma.rn.ftz.f32 %f2443, %f3213, %f3210, %f2442; fma.rn.ftz.f32 %f2444, %f3217, %f3208, %f2443; setp.eq.ftz.f32 %p480, %f2444, 0f00000000; selp.f32 %f3214, 0f7F7FFFFF, %f2444, %p480; $L__BB215_595: mul.ftz.f32 %f2445, %f3220, %f3211; mul.ftz.f32 %f2446, %f3217, %f3218; sub.ftz.f32 %f2447, %f2446, %f2445; mul.ftz.f32 %f2448, %f3217, %f3221; mul.ftz.f32 %f2449, %f3211, %f3213; sub.ftz.f32 %f2450, %f2449, %f2448; mul.ftz.f32 %f2451, %f3218, %f3213; mul.ftz.f32 %f2452, %f3220, %f3221; sub.ftz.f32 %f2453, %f2452, %f2451; div.approx.ftz.f32 %f686, %f2447, %f3214; div.approx.ftz.f32 %f683, %f2450, %f3214; div.approx.ftz.f32 %f680, %f2453, %f3214; mul.ftz.f32 %f2454, %f3217, %f3212; mul.ftz.f32 %f2455, %f3220, %f3216; sub.ftz.f32 %f2456, %f2455, %f2454; mul.ftz.f32 %f2457, %f3216, %f3213; mul.ftz.f32 %f2458, %f3217, %f3222; sub.ftz.f32 %f2459, %f2458, %f2457; mul.ftz.f32 %f2460, %f3220, %f3222; mul.ftz.f32 %f2461, %f3212, %f3213; sub.ftz.f32 %f2462, %f2461, %f2460; div.approx.ftz.f32 %f685, %f2456, %f3214; div.approx.ftz.f32 %f682, %f2459, %f3214; div.approx.ftz.f32 %f679, %f2462, %f3214; div.approx.ftz.f32 %f687, %f3210, %f3214; div.approx.ftz.f32 %f684, %f3209, %f3214; mul.ftz.f32 %f2463, %f629, %f684; fma.rn.ftz.f32 %f2464, %f628, %f687, %f2463; div.approx.ftz.f32 %f681, %f3208, %f3214; fma.rn.ftz.f32 %f2465, %f630, %f681, %f2464; mul.ftz.f32 %f2466, %f629, %f683; fma.rn.ftz.f32 %f2467, %f628, %f686, %f2466; fma.rn.ftz.f32 %f2468, %f630, %f680, %f2467; mul.ftz.f32 %f2469, %f629, %f682; fma.rn.ftz.f32 %f2470, %f628, %f685, %f2469; fma.rn.ftz.f32 %f2471, %f630, %f679, %f2470; neg.ftz.f32 %f2472, %f2465; add.s64 %rd760, %rd713, 192; cvta.to.local.u64 %rd761, %rd760; st.local.v4.f32 [%rd761], {%f687, %f684, %f681, %f2472}; neg.ftz.f32 %f2473, %f2468; add.s64 %rd762, %rd713, 208; cvta.to.local.u64 %rd763, %rd762; st.local.v4.f32 [%rd763], {%f686, %f683, %f680, %f2473}; neg.ftz.f32 %f2474, %f2471; add.s64 %rd764, %rd713, 224; cvta.to.local.u64 %rd765, %rd764; st.local.v4.f32 [%rd765], {%f685, %f682, %f679, %f2474}; $L__BB215_596: setp.eq.s32 %p481, %r2217, 1; @%p481 bra $L__BB215_598; bra.uni $L__BB215_597; $L__BB215_598: setp.lt.s32 %p482, %r297, 0; @%p482 bra $L__BB215_600; ld.const.u64 %rd1747, [kernel_params+2232]; and.b32 %r2218, %r305, 4; mov.f32 %f3248, %f590; mov.f32 %f3249, %f3253; mov.f32 %f3250, %f3252; bra.uni $L__BB215_606; $L__BB215_597: add.s64 %rd767, %rd713, 116; cvta.to.local.u64 %rd768, %rd767; st.local.u32 [%rd768], %r1693; add.s64 %rd769, %rd713, 120; cvta.to.local.u64 %rd770, %rd769; st.local.u32 [%rd770], %r1693; add.s64 %rd771, %rd713, 124; cvta.to.local.u64 %rd772, %rd771; st.local.u32 [%rd772], %r1693; add.s64 %rd773, %rd713, 136; cvta.to.local.u64 %rd774, %rd773; add.s64 %rd775, %rd713, 128; cvta.to.local.u64 %rd776, %rd775; mov.f32 %f3242, 0f00000000; st.local.v2.f32 [%rd776], {%f3242, %f3242}; st.local.u32 [%rd774], %r1693; mov.f32 %f3243, %f3242; mov.f32 %f3244, %f3242; mov.f32 %f3245, %f3242; mov.f32 %f3246, %f3242; mov.f32 %f3247, %f3242; mov.f32 %f3248, %f590; mov.f32 %f3249, %f3253; mov.f32 %f3250, %f3252; bra.uni $L__BB215_616; $L__BB215_600: ld.const.v2.u64 {%rd777, %rd1747}, [kernel_params+2224]; cvta.to.global.u64 %rd780, %rd1747; shl.b64 %rd781, %rd109, 4; add.s64 %rd782, %rd780, %rd781; ld.global.v4.u32 {%r1700, %r1701, %r1702, %r1703}, [%rd782]; cvta.to.global.u64 %rd783, %rd777; mul.wide.u32 %rd784, %r1700, 12; add.s64 %rd785, %rd783, %rd784; mul.wide.u32 %rd786, %r1701, 12; add.s64 %rd787, %rd783, %rd786; mul.wide.u32 %rd788, %r1702, 12; add.s64 %rd789, %rd783, %rd788; mov.f32 %f2481, 0f3F800000; sub.ftz.f32 %f2482, %f2481, %f594; sub.ftz.f32 %f2483, %f2482, %f595; ld.global.f32 %f2484, [%rd785]; ld.global.f32 %f2485, [%rd785+4]; ld.global.f32 %f2486, [%rd785+8]; ld.global.f32 %f2487, [%rd787]; mul.ftz.f32 %f2488, %f594, %f2487; ld.global.f32 %f2489, [%rd787+4]; mul.ftz.f32 %f2490, %f594, %f2489; ld.global.f32 %f2491, [%rd787+8]; mul.ftz.f32 %f2492, %f594, %f2491; fma.rn.ftz.f32 %f2493, %f2483, %f2484, %f2488; fma.rn.ftz.f32 %f2494, %f2483, %f2485, %f2490; fma.rn.ftz.f32 %f2495, %f2483, %f2486, %f2492; ld.global.f32 %f2496, [%rd789]; ld.global.f32 %f2497, [%rd789+4]; ld.global.f32 %f2498, [%rd789+8]; fma.rn.ftz.f32 %f3233, %f595, %f2496, %f2493; fma.rn.ftz.f32 %f3234, %f595, %f2497, %f2494; fma.rn.ftz.f32 %f3235, %f595, %f2498, %f2495; mul.ftz.f32 %f2499, %f3234, %f3234; fma.rn.ftz.f32 %f2500, %f3233, %f3233, %f2499; fma.rn.ftz.f32 %f2501, %f3235, %f3235, %f2500; sqrt.approx.ftz.f32 %f691, %f2501; setp.neu.ftz.f32 %p483, %f691, 0f00000000; @%p483 bra $L__BB215_601; bra.uni $L__BB215_602; $L__BB215_601: rcp.approx.ftz.f32 %f2502, %f691; mul.ftz.f32 %f3235, %f3235, %f2502; mul.ftz.f32 %f3234, %f3234, %f2502; mul.ftz.f32 %f3233, %f3233, %f2502; $L__BB215_602: add.s64 %rd790, %rd107, 8; setp.eq.ftz.f32 %p484, %f3233, 0f00000000; setp.eq.ftz.f32 %p485, %f3234, 0f00000000; and.pred %p486, %p484, %p485; setp.eq.ftz.f32 %p487, %f3235, 0f00000000; and.pred %p488, %p487, %p486; selp.f32 %f3248, %f590, %f3235, %p488; selp.f32 %f3249, %f3253, %f3234, %p488; selp.f32 %f3250, %f3252, %f3233, %p488; cvta.to.local.u64 %rd120, %rd107; st.local.f32 [%rd120], %f3250; st.local.f32 [%rd110], %f3249; cvta.to.local.u64 %rd122, %rd790; st.local.f32 [%rd122], %f3248; and.b32 %r2218, %r305, 4; setp.ne.s32 %p489, %r2218, 0; @%p489 bra $L__BB215_606; @%p475 bra $L__BB215_605; bra.uni $L__BB215_604; $L__BB215_605: ld.const.u64 %rd792, [kernel_params+2160]; cvta.to.global.u64 %rd793, %rd792; mul.lo.s64 %rd794, %rd108, 208; add.s64 %rd795, %rd793, %rd794; ld.global.v4.f32 {%f2516, %f2517, %f2518, %f2519}, [%rd795+48]; ld.global.v4.f32 {%f2523, %f2524, %f2525, %f2526}, [%rd795+64]; ld.global.v4.f32 {%f2530, %f2531, %f2532, %f2533}, [%rd795+80]; mul.ftz.f32 %f2537, %f2523, %f3249; fma.rn.ftz.f32 %f2538, %f2516, %f3250, %f2537; fma.rn.ftz.f32 %f2539, %f2530, %f3248, %f2538; mul.ftz.f32 %f2540, %f2524, %f3249; fma.rn.ftz.f32 %f2541, %f2517, %f3250, %f2540; fma.rn.ftz.f32 %f2542, %f2531, %f3248, %f2541; mul.ftz.f32 %f2543, %f2525, %f3249; fma.rn.ftz.f32 %f2544, %f2518, %f3250, %f2543; fma.rn.ftz.f32 %f2545, %f2532, %f3248, %f2544; mul.ftz.f32 %f2546, %f2542, %f2542; fma.rn.ftz.f32 %f2547, %f2539, %f2539, %f2546; fma.rn.ftz.f32 %f2548, %f2545, %f2545, %f2547; sqrt.approx.ftz.f32 %f2549, %f2548; div.approx.ftz.f32 %f3250, %f2539, %f2549; div.approx.ftz.f32 %f3249, %f2542, %f2549; div.approx.ftz.f32 %f3248, %f2545, %f2549; st.local.f32 [%rd120], %f3250; st.local.f32 [%rd110], %f3249; st.local.f32 [%rd122], %f3248; mov.u32 %r2218, %r1693; bra.uni $L__BB215_606; $L__BB215_604: mul.ftz.f32 %f2503, %f3249, %f686; fma.rn.ftz.f32 %f2504, %f3250, %f687, %f2503; fma.rn.ftz.f32 %f2505, %f3248, %f685, %f2504; mul.ftz.f32 %f2506, %f3249, %f683; fma.rn.ftz.f32 %f2507, %f3250, %f684, %f2506; fma.rn.ftz.f32 %f2508, %f3248, %f682, %f2507; mul.ftz.f32 %f2509, %f3249, %f680; fma.rn.ftz.f32 %f2510, %f3250, %f681, %f2509; fma.rn.ftz.f32 %f2511, %f3248, %f679, %f2510; mul.ftz.f32 %f2512, %f2508, %f2508; fma.rn.ftz.f32 %f2513, %f2505, %f2505, %f2512; fma.rn.ftz.f32 %f2514, %f2511, %f2511, %f2513; sqrt.approx.ftz.f32 %f2515, %f2514; div.approx.ftz.f32 %f3250, %f2505, %f2515; div.approx.ftz.f32 %f3249, %f2508, %f2515; div.approx.ftz.f32 %f3248, %f2511, %f2515; st.local.f32 [%rd120], %f3250; st.local.f32 [%rd110], %f3249; st.local.f32 [%rd122], %f3248; mov.u32 %r2218, %r1693; $L__BB215_606: cvta.to.global.u64 %rd796, %rd1747; shl.b64 %rd797, %rd109, 4; add.s64 %rd798, %rd796, %rd797; ld.const.u64 %rd799, [kernel_params+2256]; cvta.to.global.u64 %rd800, %rd799; ld.global.u32 %r1709, [%rd798+12]; mul.wide.u32 %rd801, %r1709, 12; add.s64 %rd802, %rd800, %rd801; add.s32 %r1710, %r1709, 1; mul.wide.u32 %rd803, %r1710, 12; add.s64 %rd804, %rd800, %rd803; add.s32 %r1711, %r1709, 2; mul.wide.u32 %rd805, %r1711, 12; add.s64 %rd806, %rd800, %rd805; ld.global.f32 %f2550, [%rd806]; ld.global.f32 %f2551, [%rd806+4]; ld.global.f32 %f2552, [%rd806+8]; ld.global.f32 %f2553, [%rd804]; ld.global.f32 %f2554, [%rd802]; sub.ftz.f32 %f3247, %f2553, %f2554; ld.global.f32 %f2555, [%rd804+4]; ld.global.f32 %f2556, [%rd802+4]; sub.ftz.f32 %f3246, %f2555, %f2556; ld.global.f32 %f2557, [%rd804+8]; ld.global.f32 %f2558, [%rd802+8]; sub.ftz.f32 %f3245, %f2557, %f2558; add.s64 %rd808, %rd713, 116; cvta.to.local.u64 %rd126, %rd808; st.local.f32 [%rd126], %f3247; add.s64 %rd809, %rd713, 120; cvta.to.local.u64 %rd127, %rd809; st.local.f32 [%rd127], %f3246; add.s64 %rd810, %rd713, 124; cvta.to.local.u64 %rd128, %rd810; st.local.f32 [%rd128], %f3245; sub.ftz.f32 %f3244, %f2550, %f2554; sub.ftz.f32 %f3243, %f2551, %f2556; sub.ftz.f32 %f3242, %f2552, %f2558; add.s64 %rd811, %rd713, 128; cvta.to.local.u64 %rd129, %rd811; st.local.v2.f32 [%rd129], {%f3244, %f3243}; add.s64 %rd812, %rd713, 136; cvta.to.local.u64 %rd130, %rd812; st.local.f32 [%rd130], %f3242; setp.ne.s32 %p491, %r2218, 0; @%p491 bra $L__BB215_616; @%p475 bra $L__BB215_609; bra.uni $L__BB215_608; $L__BB215_609: ld.const.u64 %rd813, [kernel_params+2160]; cvta.to.global.u64 %rd814, %rd813; mul.lo.s64 %rd815, %rd108, 208; add.s64 %rd816, %rd814, %rd815; ld.global.v4.f32 {%f2565, %f2566, %f2567, %f2568}, [%rd816]; ld.global.v4.f32 {%f2572, %f2573, %f2574, %f2575}, [%rd816+16]; ld.global.v4.f32 {%f2579, %f2580, %f2581, %f2582}, [%rd816+32]; mul.ftz.f32 %f2586, %f2566, %f3246; fma.rn.ftz.f32 %f2587, %f3247, %f2565, %f2586; fma.rn.ftz.f32 %f3241, %f2567, %f3245, %f2587; mul.ftz.f32 %f2588, %f2573, %f3246; fma.rn.ftz.f32 %f2589, %f3247, %f2572, %f2588; fma.rn.ftz.f32 %f3240, %f2574, %f3245, %f2589; mul.ftz.f32 %f2590, %f2580, %f3246; fma.rn.ftz.f32 %f2591, %f3247, %f2579, %f2590; fma.rn.ftz.f32 %f3245, %f2581, %f3245, %f2591; bra.uni $L__BB215_610; $L__BB215_608: mul.ftz.f32 %f2559, %f3246, %f3222; fma.rn.ftz.f32 %f2560, %f3247, %f3223, %f2559; fma.rn.ftz.f32 %f3241, %f3245, %f3221, %f2560; mul.ftz.f32 %f2561, %f3246, %f3219; fma.rn.ftz.f32 %f2562, %f3247, %f3220, %f2561; fma.rn.ftz.f32 %f3240, %f3245, %f3218, %f2562; mul.ftz.f32 %f2563, %f3246, %f3216; fma.rn.ftz.f32 %f2564, %f3247, %f3217, %f2563; fma.rn.ftz.f32 %f3245, %f3245, %f3215, %f2564; $L__BB215_610: st.local.f32 [%rd126], %f3241; st.local.f32 [%rd127], %f3240; st.local.f32 [%rd128], %f3245; @%p475 bra $L__BB215_612; bra.uni $L__BB215_611; $L__BB215_612: ld.const.u64 %rd817, [kernel_params+2160]; cvta.to.global.u64 %rd818, %rd817; mul.lo.s64 %rd819, %rd108, 208; add.s64 %rd820, %rd818, %rd819; ld.global.v4.f32 {%f2598, %f2599, %f2600, %f2601}, [%rd820]; ld.global.v4.f32 {%f2605, %f2606, %f2607, %f2608}, [%rd820+16]; ld.global.v4.f32 {%f2612, %f2613, %f2614, %f2615}, [%rd820+32]; mul.ftz.f32 %f2619, %f2599, %f3243; fma.rn.ftz.f32 %f2620, %f3244, %f2598, %f2619; fma.rn.ftz.f32 %f731, %f2600, %f3242, %f2620; mul.ftz.f32 %f2621, %f2606, %f3243; fma.rn.ftz.f32 %f2622, %f3244, %f2605, %f2621; fma.rn.ftz.f32 %f732, %f2607, %f3242, %f2622; mul.ftz.f32 %f2623, %f2613, %f3243; fma.rn.ftz.f32 %f2624, %f3244, %f2612, %f2623; fma.rn.ftz.f32 %f3242, %f2614, %f3242, %f2624; st.local.v2.f32 [%rd129], {%f731, %f732}; st.local.f32 [%rd130], %f3242; mov.f32 %f3243, %f732; mov.f32 %f3244, %f731; mov.f32 %f3246, %f3240; mov.f32 %f3247, %f3241; bra.uni $L__BB215_616; $L__BB215_611: mul.ftz.f32 %f2592, %f3243, %f3222; fma.rn.ftz.f32 %f2593, %f3244, %f3223, %f2592; fma.rn.ftz.f32 %f728, %f3242, %f3221, %f2593; mul.ftz.f32 %f2594, %f3243, %f3219; fma.rn.ftz.f32 %f2595, %f3244, %f3220, %f2594; fma.rn.ftz.f32 %f729, %f3242, %f3218, %f2595; mul.ftz.f32 %f2596, %f3243, %f3216; fma.rn.ftz.f32 %f2597, %f3244, %f3217, %f2596; fma.rn.ftz.f32 %f3242, %f3242, %f3215, %f2597; st.local.v2.f32 [%rd129], {%f728, %f729}; st.local.f32 [%rd130], %f3242; mov.f32 %f3243, %f729; mov.f32 %f3244, %f728; mov.f32 %f3246, %f3240; mov.f32 %f3247, %f3241; $L__BB215_616: mov.b32 %r1713, %f590; mov.b64 %rd833, {%r1713, %r1714}; cvt.u32.u64 %r1715, %rd833; mov.b32 %f3251, %r1715; @%p472 bra $L__BB215_619; mul.ftz.f32 %f2631, %f3253, %f603; fma.rn.ftz.f32 %f2632, %f3252, %f602, %f2631; fma.rn.ftz.f32 %f2633, %f590, %f604, %f2632; setp.geu.ftz.f32 %p496, %f2633, 0f00000000; @%p496 bra $L__BB215_619; or.b32 %r1716, %r299, 1; st.local.u32 [%rd114], %r1716; neg.ftz.f32 %f3251, %f590; st.local.f32 [%rd721], %f3251; cvta.to.local.u64 %rd836, %rd107; neg.ftz.f32 %f2634, %f3250; st.local.f32 [%rd836], %f2634; neg.ftz.f32 %f2635, %f3248; neg.ftz.f32 %f2636, %f3249; neg.ftz.f32 %f3253, %f3253; neg.ftz.f32 %f3252, %f3252; st.local.v4.f32 [%rd110], {%f2636, %f2635, %f3252, %f3253}; add.s64 %rd838, %rd713, 116; add.s64 %rd839, %rd713, 120; add.s64 %rd840, %rd713, 124; cvta.to.local.u64 %rd841, %rd838; neg.ftz.f32 %f2637, %f3247; st.local.f32 [%rd841], %f2637; cvta.to.local.u64 %rd842, %rd839; neg.ftz.f32 %f2638, %f3246; st.local.f32 [%rd842], %f2638; cvta.to.local.u64 %rd843, %rd840; neg.ftz.f32 %f2639, %f3245; st.local.f32 [%rd843], %f2639; add.s64 %rd844, %rd713, 136; cvta.to.local.u64 %rd845, %rd844; neg.ftz.f32 %f2640, %f3243; neg.ftz.f32 %f2641, %f3244; add.s64 %rd846, %rd713, 128; cvta.to.local.u64 %rd847, %rd846; st.local.v2.f32 [%rd847], {%f2641, %f2640}; neg.ftz.f32 %f2642, %f3242; st.local.f32 [%rd845], %f2642; $L__BB215_619: add.s64 %rd849, %rd713, 92; cvta.to.local.u64 %rd850, %rd849; mov.u32 %r1717, 0; st.local.u32 [%rd850], %r1717; add.s64 %rd851, %rd713, 96; cvta.to.local.u64 %rd852, %rd851; mov.f32 %f2643, 0f00000000; st.local.v4.f32 [%rd852], {%f2643, %f2643, %f2643, %f2643}; add.s64 %rd853, %rd713, 112; cvta.to.local.u64 %rd854, %rd853; st.local.u32 [%rd854], %r1717; st.local.f32 [%rd31+12], %f3252; st.local.f32 [%rd31+16], %f3253; st.local.f32 [%rd31+20], %f3251; $L__BB215_621: add.s64 %rd891, %rd713, 260; cvta.to.local.u64 %rd892, %rd891; mov.u32 %r1724, 0; st.local.u32 [%rd892], %r1724; add.s64 %rd893, %rd713, 264; cvta.to.local.u64 %rd894, %rd893; st.local.u32 [%rd894], %r1724; ld.const.u64 %rd895, [kernel_params+2400]; cvta.to.global.u64 %rd131, %rd895; add.s64 %rd896, %rd713, 268; cvta.to.local.u64 %rd135, %rd896; add.s64 %rd897, %rd713, 272; cvta.to.local.u64 %rd136, %rd897; add.s64 %rd898, %rd713, 276; cvta.to.local.u64 %rd137, %rd898; add.u64 %rd899, %SP, 304; add.u64 %rd138, %SPL, 304; bra.uni $L__BB215_622; $L__BB215_774: { // callseq 1454, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1454 $L__BB215_622: mov.u32 %r312, %r2220; mul.wide.s32 %rd900, %r312, 16; add.s64 %rd139, %rd131, %rd900; ld.global.v4.u32 {%r1726, %r1727, %r1728, %r1729}, [%rd139]; mul.wide.u32 %rd901, %r1727, 4; add.s64 %rd140, %rd138, %rd901; mul.wide.u32 %rd902, %r1728, 4; add.s64 %rd141, %rd138, %rd902; add.s32 %r2220, %r312, 1; setp.gt.s32 %p497, %r1726, 36; @%p497 bra $L__BB215_671; setp.gt.s32 %p579, %r1726, 25; @%p579 bra $L__BB215_667; bra.uni $L__BB215_624; $L__BB215_667: add.s32 %r1731, %r1726, -27; setp.lt.u32 %p580, %r1731, 9; @%p580 bra $L__BB215_622; setp.eq.s32 %p581, %r1726, 26; @%p581 bra $L__BB215_813; setp.eq.s32 %p582, %r1726, 36; @%p582 bra $L__BB215_670; bra.uni $L__BB215_827; $L__BB215_670: mov.u32 %r1773, 36; { // callseq 1507, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1773; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1507 bra.uni $L__BB215_622; $L__BB215_671: setp.gt.s32 %p498, %r1726, 68; @%p498 bra $L__BB215_726; bra.uni $L__BB215_672; $L__BB215_726: setp.gt.s32 %p499, %r1726, 88; @%p499 bra $L__BB215_762; bra.uni $L__BB215_727; $L__BB215_762: setp.gt.s32 %p500, %r1726, 93; @%p500 bra $L__BB215_767; bra.uni $L__BB215_763; $L__BB215_767: setp.gt.s32 %p501, %r1726, 95; @%p501 bra $L__BB215_771; bra.uni $L__BB215_768; $L__BB215_771: setp.eq.s32 %p502, %r1726, 98; @%p502 bra $L__BB215_826; setp.eq.s32 %p503, %r1726, 97; @%p503 bra $L__BB215_776; setp.ne.s32 %p504, %r1726, 96; @%p504 bra $L__BB215_827; bra.uni $L__BB215_774; $L__BB215_672: setp.gt.s32 %p537, %r1726, 48; @%p537 bra $L__BB215_691; bra.uni $L__BB215_673; $L__BB215_691: setp.gt.s32 %p538, %r1726, 58; @%p538 bra $L__BB215_709; setp.gt.s32 %p552, %r1726, 53; @%p552 bra $L__BB215_701; setp.gt.s32 %p559, %r1726, 50; @%p559 bra $L__BB215_697; bra.uni $L__BB215_694; $L__BB215_697: setp.eq.s32 %p560, %r1726, 51; @%p560 bra $L__BB215_805; setp.eq.s32 %p561, %r1726, 52; @%p561 bra $L__BB215_804; setp.eq.s32 %p562, %r1726, 53; @%p562 bra $L__BB215_700; bra.uni $L__BB215_827; $L__BB215_700: { // callseq 1492, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1492 bra.uni $L__BB215_622; $L__BB215_624: setp.gt.s32 %p583, %r1726, 17; @%p583 bra $L__BB215_658; bra.uni $L__BB215_625; $L__BB215_658: setp.lt.s32 %p584, %r1726, 21; @%p584 bra $L__BB215_622; setp.gt.s32 %p585, %r1726, 22; @%p585 bra $L__BB215_663; bra.uni $L__BB215_660; $L__BB215_663: setp.eq.s32 %p586, %r1726, 23; @%p586 bra $L__BB215_815; setp.eq.s32 %p587, %r1726, 24; @%p587 bra $L__BB215_814; setp.eq.s32 %p588, %r1726, 25; @%p588 bra $L__BB215_666; bra.uni $L__BB215_827; $L__BB215_666: add.s32 %r1775, %r312, 1; { // callseq 1509, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1775; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1509 bra.uni $L__BB215_622; $L__BB215_727: setp.gt.s32 %p510, %r1726, 78; @%p510 bra $L__BB215_745; setp.gt.s32 %p524, %r1726, 73; @%p524 bra $L__BB215_737; setp.gt.s32 %p531, %r1726, 70; @%p531 bra $L__BB215_733; bra.uni $L__BB215_730; $L__BB215_733: setp.eq.s32 %p532, %r1726, 71; @%p532 bra $L__BB215_793; setp.eq.s32 %p533, %r1726, 72; @%p533 bra $L__BB215_792; setp.eq.s32 %p534, %r1726, 73; @%p534 bra $L__BB215_736; bra.uni $L__BB215_827; $L__BB215_736: add.s32 %r1744, %r312, 1; { // callseq 1472, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1744; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1472 bra.uni $L__BB215_622; $L__BB215_673: setp.gt.s32 %p565, %r1726, 41; @%p565 bra $L__BB215_679; setp.gt.s32 %p575, %r1726, 39; @%p575 bra $L__BB215_622; setp.eq.s32 %p576, %r1726, 37; @%p576 bra $L__BB215_812; setp.eq.s32 %p577, %r1726, 38; @%p577 bra $L__BB215_811; setp.eq.s32 %p578, %r1726, 39; @%p578 bra $L__BB215_678; bra.uni $L__BB215_827; $L__BB215_678: mov.u32 %r1771, 39; { // callseq 1504, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1771, %r1727, %r1728, %r1729}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1504 bra.uni $L__BB215_622; $L__BB215_625: setp.gt.s32 %p591, %r1726, 8; @%p591 bra $L__BB215_641; bra.uni $L__BB215_626; $L__BB215_641: setp.gt.s32 %p592, %r1726, 12; @%p592 bra $L__BB215_650; bra.uni $L__BB215_642; $L__BB215_650: setp.gt.s32 %p593, %r1726, 14; @%p593 bra $L__BB215_654; bra.uni $L__BB215_651; $L__BB215_654: setp.eq.s32 %p594, %r1726, 15; @%p594 bra $L__BB215_818; setp.eq.s32 %p595, %r1726, 16; @%p595 bra $L__BB215_817; setp.eq.s32 %p596, %r1726, 17; @%p596 bra $L__BB215_657; bra.uni $L__BB215_827; $L__BB215_657: { // callseq 1514, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; .param .b32 param4; st.param.b32 [param4+0], %r1729; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1514 bra.uni $L__BB215_622; $L__BB215_709: setp.gt.s32 %p539, %r1726, 63; @%p539 bra $L__BB215_718; setp.gt.s32 %p546, %r1726, 60; @%p546 bra $L__BB215_714; bra.uni $L__BB215_711; $L__BB215_714: setp.eq.s32 %p547, %r1726, 61; @%p547 bra $L__BB215_799; setp.eq.s32 %p548, %r1726, 62; @%p548 bra $L__BB215_798; setp.eq.s32 %p549, %r1726, 63; @%p549 bra $L__BB215_717; bra.uni $L__BB215_827; $L__BB215_717: mov.u32 %r1754, 63; { // callseq 1482, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1754; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1482 bra.uni $L__BB215_622; $L__BB215_745: setp.gt.s32 %p511, %r1726, 83; @%p511 bra $L__BB215_754; setp.gt.s32 %p518, %r1726, 80; @%p518 bra $L__BB215_750; bra.uni $L__BB215_747; $L__BB215_750: setp.eq.s32 %p519, %r1726, 81; @%p519 bra $L__BB215_781; setp.eq.s32 %p520, %r1726, 82; @%p520 bra $L__BB215_780; setp.eq.s32 %p521, %r1726, 83; @%p521 bra $L__BB215_753; bra.uni $L__BB215_827; $L__BB215_753: { // callseq 1464, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1464 bra.uni $L__BB215_622; $L__BB215_626: setp.gt.s32 %p604, %r1726, 4; @%p604 bra $L__BB215_634; setp.gt.s32 %p610, %r1726, 2; @%p610 bra $L__BB215_631; setp.eq.s32 %p613, %r1726, 1; mov.u32 %r2220, %r1727; @%p613 bra $L__BB215_622; setp.eq.s32 %p614, %r1726, 2; @%p614 bra $L__BB215_630; bra.uni $L__BB215_827; $L__BB215_630: mov.u32 %r1799, 2; add.s32 %r1800, %r312, 1; { // callseq 1523, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r1799; st.param.b32 [param0+4], %r1727; st.param.b32 [param0+8], %r1728; st.param.b32 [param0+12], %r1729; .param .b32 param1; st.param.b32 [param1+0], %r1800; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj290EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1523 bra.uni $L__BB215_622; $L__BB215_679: setp.gt.s32 %p566, %r1726, 44; @%p566 bra $L__BB215_684; bra.uni $L__BB215_680; $L__BB215_684: setp.gt.s32 %p567, %r1726, 46; @%p567 bra $L__BB215_688; setp.eq.s32 %p570, %r1726, 45; @%p570 bra $L__BB215_808; setp.eq.s32 %p571, %r1726, 46; @%p571 bra $L__BB215_687; bra.uni $L__BB215_827; $L__BB215_687: { // callseq 1499, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1499 bra.uni $L__BB215_622; $L__BB215_768: setp.eq.s32 %p505, %r1726, 94; @%p505 bra $L__BB215_775; setp.eq.s32 %p506, %r1726, 95; @%p506 bra $L__BB215_770; bra.uni $L__BB215_827; $L__BB215_770: { // callseq 1455, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1455 bra.uni $L__BB215_622; $L__BB215_763: add.s32 %r1730, %r1726, -92; setp.lt.u32 %p507, %r1730, 2; @%p507 bra $L__BB215_622; setp.eq.s32 %p508, %r1726, 90; @%p508 bra $L__BB215_622; setp.eq.s32 %p509, %r1726, 89; @%p509 bra $L__BB215_766; bra.uni $L__BB215_827; $L__BB215_766: mov.u32 %r1734, 89; { // callseq 1458, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1734; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z11svm_node_aoILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1458 bra.uni $L__BB215_622; $L__BB215_701: setp.gt.s32 %p553, %r1726, 55; @%p553 bra $L__BB215_705; bra.uni $L__BB215_702; $L__BB215_705: setp.eq.s32 %p554, %r1726, 56; @%p554 bra $L__BB215_802; setp.eq.s32 %p555, %r1726, 57; @%p555 bra $L__BB215_801; setp.eq.s32 %p556, %r1726, 58; @%p556 bra $L__BB215_708; bra.uni $L__BB215_827; $L__BB215_708: mov.u32 %r1761, 58; { // callseq 1487, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1761; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1487 bra.uni $L__BB215_622; $L__BB215_737: setp.gt.s32 %p525, %r1726, 75; @%p525 bra $L__BB215_741; bra.uni $L__BB215_738; $L__BB215_741: setp.eq.s32 %p526, %r1726, 76; @%p526 bra $L__BB215_785; setp.eq.s32 %p527, %r1726, 77; @%p527 bra $L__BB215_783; setp.eq.s32 %p528, %r1726, 78; @%p528 bra $L__BB215_744; bra.uni $L__BB215_827; $L__BB215_744: add.s32 %r1742, %r312, 1; { // callseq 1469, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1742; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1469 bra.uni $L__BB215_622; $L__BB215_718: setp.gt.s32 %p540, %r1726, 65; @%p540 bra $L__BB215_722; bra.uni $L__BB215_719; $L__BB215_722: setp.eq.s32 %p541, %r1726, 66; @%p541 bra $L__BB215_796; setp.eq.s32 %p542, %r1726, 67; @%p542 bra $L__BB215_795; setp.eq.s32 %p543, %r1726, 68; @%p543 bra $L__BB215_725; bra.uni $L__BB215_827; $L__BB215_725: mov.u32 %r1749, 68; { // callseq 1477, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1749; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1477 bra.uni $L__BB215_622; $L__BB215_754: setp.gt.s32 %p512, %r1726, 85; @%p512 bra $L__BB215_758; bra.uni $L__BB215_755; $L__BB215_758: setp.eq.s32 %p513, %r1726, 86; @%p513 bra $L__BB215_778; setp.eq.s32 %p514, %r1726, 87; @%p514 bra $L__BB215_777; setp.eq.s32 %p515, %r1726, 88; @%p515 bra $L__BB215_761; bra.uni $L__BB215_827; $L__BB215_761: mov.u32 %r1735, 88; cvta.to.local.u64 %rd909, %rd713; ld.local.f32 %f2651, [%rd909+12]; ld.local.v2.f32 {%f2652, %f2653}, [%rd909+16]; { // callseq 1459, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f2651; .param .b32 param1; st.param.f32 [param1+0], %f2652; .param .b32 param2; st.param.f32 [param2+0], %f2653; .param .b64 param3; st.param.b64 [param3+0], %rd899; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r1735; st.param.b32 [param4+4], %r1727; st.param.b32 [param4+8], %r1728; st.param.b32 [param4+12], %r1729; call.uni _Z14svm_node_bevelILj290EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1459 bra.uni $L__BB215_622; $L__BB215_642: setp.gt.s32 %p599, %r1726, 10; @%p599 bra $L__BB215_647; setp.eq.s32 %p602, %r1726, 9; @%p602 bra $L__BB215_821; setp.eq.s32 %p603, %r1726, 10; @%p603 bra $L__BB215_645; bra.uni $L__BB215_827; $L__BB215_645: ld.local.f32 %f2667, [%rd141]; setp.ltu.ftz.f32 %p619, %f2667, 0f3F800000; @%p619 bra $L__BB215_622; add.s32 %r1793, %r1727, %r312; add.s32 %r2220, %r1793, 1; bra.uni $L__BB215_622; $L__BB215_634: setp.gt.s32 %p605, %r1726, 6; @%p605 bra $L__BB215_638; setp.eq.s32 %p608, %r1726, 5; @%p608 bra $L__BB215_824; setp.eq.s32 %p609, %r1726, 6; @%p609 bra $L__BB215_637; bra.uni $L__BB215_827; $L__BB215_637: ld.local.f32 %f2669, [%rd140]; ld.local.f32 %f2670, [%rd140+4]; ld.local.f32 %f2671, [%rd140+8]; st.local.f32 [%rd135], %f2669; st.local.f32 [%rd136], %f2670; st.local.f32 [%rd137], %f2671; bra.uni $L__BB215_622; $L__BB215_680: setp.eq.s32 %p572, %r1726, 42; @%p572 bra $L__BB215_810; setp.eq.s32 %p573, %r1726, 43; @%p573 bra $L__BB215_809; setp.eq.s32 %p574, %r1726, 44; @%p574 bra $L__BB215_683; bra.uni $L__BB215_827; $L__BB215_683: mov.u32 %r1768, 44; add.s32 %r1769, %r312, 1; { // callseq 1501, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1768; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1769; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1501 bra.uni $L__BB215_622; $L__BB215_826: { // callseq 1524, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1524 bra.uni $L__BB215_622; $L__BB215_776: { // callseq 1457, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1457 bra.uni $L__BB215_622; $L__BB215_775: mov.u32 %r1732, 94; add.s32 %r1733, %r312, 1; { // callseq 1456, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1732; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1733; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1456 bra.uni $L__BB215_622; $L__BB215_660: setp.eq.s32 %p589, %r1726, 21; @%p589 bra $L__BB215_816; setp.eq.s32 %p590, %r1726, 22; @%p590 bra $L__BB215_662; bra.uni $L__BB215_827; $L__BB215_662: mov.u32 %r1779, 22; add.s32 %r1780, %r312, 1; { // callseq 1512, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1779; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1780; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj290EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1512 bra.uni $L__BB215_622; $L__BB215_651: setp.eq.s32 %p597, %r1726, 13; @%p597 bra $L__BB215_819; setp.eq.s32 %p598, %r1726, 14; @%p598 bra $L__BB215_653; bra.uni $L__BB215_827; $L__BB215_653: st.local.u32 [%rd141], %r1727; bra.uni $L__BB215_622; $L__BB215_631: setp.eq.s32 %p611, %r1726, 3; @%p611 bra $L__BB215_825; setp.eq.s32 %p612, %r1726, 4; @%p612 bra $L__BB215_633; bra.uni $L__BB215_827; $L__BB215_633: mov.u32 %r1797, 4; { // callseq 1521, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1797, %r1727, %r1728, %r1729}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1521 bra.uni $L__BB215_622; $L__BB215_647: setp.eq.s32 %p600, %r1726, 11; @%p600 bra $L__BB215_820; setp.eq.s32 %p601, %r1726, 12; @%p601 bra $L__BB215_649; bra.uni $L__BB215_827; $L__BB215_649: { // callseq 1517, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1517 bra.uni $L__BB215_622; $L__BB215_638: setp.eq.s32 %p606, %r1726, 7; @%p606 bra $L__BB215_823; setp.eq.s32 %p607, %r1726, 8; @%p607 bra $L__BB215_640; bra.uni $L__BB215_827; $L__BB215_640: mov.u32 %r1795, 8; { // callseq 1519, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1795; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1519 bra.uni $L__BB215_622; $L__BB215_694: setp.eq.s32 %p563, %r1726, 49; @%p563 bra $L__BB215_806; setp.eq.s32 %p564, %r1726, 50; @%p564 bra $L__BB215_696; bra.uni $L__BB215_827; $L__BB215_696: { // callseq 1495, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1495 bra.uni $L__BB215_622; $L__BB215_730: setp.eq.s32 %p535, %r1726, 69; @%p535 bra $L__BB215_794; setp.eq.s32 %p536, %r1726, 70; @%p536 bra $L__BB215_732; bra.uni $L__BB215_827; $L__BB215_732: mov.u32 %r1746, 70; { // callseq 1475, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1746; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1475 bra.uni $L__BB215_622; $L__BB215_711: setp.eq.s32 %p550, %r1726, 59; @%p550 bra $L__BB215_800; setp.eq.s32 %p551, %r1726, 60; @%p551 bra $L__BB215_713; bra.uni $L__BB215_827; $L__BB215_713: add.s32 %r1759, %r312, 1; { // callseq 1485, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1759; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1485 bra.uni $L__BB215_622; $L__BB215_747: setp.eq.s32 %p522, %r1726, 79; @%p522 bra $L__BB215_782; setp.eq.s32 %p523, %r1726, 80; @%p523 bra $L__BB215_749; bra.uni $L__BB215_827; $L__BB215_749: { // callseq 1467, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1467 bra.uni $L__BB215_622; $L__BB215_688: setp.eq.s32 %p568, %r1726, 47; @%p568 bra $L__BB215_807; setp.eq.s32 %p569, %r1726, 48; @%p569 bra $L__BB215_690; bra.uni $L__BB215_827; $L__BB215_690: { // callseq 1497, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1497 bra.uni $L__BB215_622; $L__BB215_702: setp.eq.s32 %p557, %r1726, 54; @%p557 bra $L__BB215_803; setp.eq.s32 %p558, %r1726, 55; @%p558 bra $L__BB215_704; bra.uni $L__BB215_827; $L__BB215_704: cvta.to.local.u64 %rd954, %rd713; ld.local.v4.f32 {%f2660, %f2661, %f2662, %f2663}, [%rd954]; { // callseq 1490, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f2660; st.param.f32 [param0+4], %f2661; st.param.f32 [param0+8], %f2662; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; .param .b32 param4; st.param.b32 [param4+0], %r1729; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1490 bra.uni $L__BB215_622; $L__BB215_738: setp.eq.s32 %p529, %r1726, 74; @%p529 bra $L__BB215_791; setp.eq.s32 %p530, %r1726, 75; @%p530 bra $L__BB215_740; bra.uni $L__BB215_827; $L__BB215_740: { // callseq 1470, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1470 bra.uni $L__BB215_622; $L__BB215_719: setp.eq.s32 %p544, %r1726, 64; @%p544 bra $L__BB215_797; setp.eq.s32 %p545, %r1726, 65; @%p545 bra $L__BB215_721; bra.uni $L__BB215_827; $L__BB215_721: { // callseq 1480, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1480 bra.uni $L__BB215_622; $L__BB215_755: setp.eq.s32 %p516, %r1726, 84; @%p516 bra $L__BB215_779; setp.eq.s32 %p517, %r1726, 85; @%p517 bra $L__BB215_757; bra.uni $L__BB215_827; $L__BB215_757: add.s32 %r1738, %r312, 1; { // callseq 1462, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1738; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1462 bra.uni $L__BB215_622; $L__BB215_815: mov.u32 %r1777, 23; add.s32 %r1778, %r312, 1; { // callseq 1511, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1777; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1778; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1511 bra.uni $L__BB215_622; $L__BB215_814: mov.u32 %r1776, 24; { // callseq 1510, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1776; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1510 bra.uni $L__BB215_622; $L__BB215_818: ld.global.v4.u32 {%r1783, %r1784, %r1785, %r1786}, [%rd139+16]; add.s32 %r2220, %r312, 2; st.local.u32 [%rd140], %r1784; st.local.u32 [%rd140+4], %r1785; st.local.u32 [%rd140+8], %r1786; bra.uni $L__BB215_622; $L__BB215_817: { // callseq 1515, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r492; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z13svm_node_attrILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1515 bra.uni $L__BB215_622; $L__BB215_821: ld.local.f32 %f2668, [%rd141]; setp.gtu.ftz.f32 %p620, %f2668, 0f00000000; @%p620 bra $L__BB215_622; add.s32 %r1794, %r1727, %r312; add.s32 %r2220, %r1794, 1; bra.uni $L__BB215_622; $L__BB215_824: st.local.u32 [%rd135], %r1727; st.local.u32 [%rd136], %r1728; st.local.u32 [%rd137], %r1729; bra.uni $L__BB215_622; $L__BB215_816: mov.u32 %r1781, 21; { // callseq 1513, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1781; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z21svm_node_displacementILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1513 bra.uni $L__BB215_622; $L__BB215_819: mov.u32 %r1790, 13; add.s32 %r1791, %r312, 1; mov.u32 %r1792, 524288; { // callseq 1516, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b32 param1; st.param.b32 [param1+0], %r1792; .param .b64 param2; st.param.b64 [param2+0], %rd899; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r1790; st.param.b32 [param3+4], %r1727; st.param.b32 [param3+8], %r1728; st.param.b32 [param3+12], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1791; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1516 bra.uni $L__BB215_622; $L__BB215_825: mov.u32 %r1798, 3; { // callseq 1522, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1798, %r1727, %r1728, %r1729}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1522 bra.uni $L__BB215_622; $L__BB215_820: { // callseq 1518, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1518 bra.uni $L__BB215_622; $L__BB215_823: mov.u32 %r1796, 7; { // callseq 1520, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1796; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1520 bra.uni $L__BB215_622; $L__BB215_805: { // callseq 1494, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1494 bra.uni $L__BB215_622; $L__BB215_804: add.s32 %r1766, %r312, 1; { // callseq 1493, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1766; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1493 bra.uni $L__BB215_622; $L__BB215_793: mov.u32 %r1745, 71; { // callseq 1474, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1745; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1474 bra.uni $L__BB215_622; $L__BB215_792: { // callseq 1473, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1473 bra.uni $L__BB215_622; $L__BB215_812: mov.u32 %r1772, 37; { // callseq 1506, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1772, %r1727, %r1728, %r1729}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1506 bra.uni $L__BB215_622; $L__BB215_811: { // callseq 1505, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; .param .b32 param4; st.param.b32 [param4+0], %r1729; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1505 bra.uni $L__BB215_622; $L__BB215_799: mov.u32 %r1757, 61; add.s32 %r1758, %r312, 1; { // callseq 1484, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1757; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1758; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1484 bra.uni $L__BB215_622; $L__BB215_798: mov.u32 %r1755, 62; add.s32 %r1756, %r312, 1; { // callseq 1483, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1755; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1756; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1483 bra.uni $L__BB215_622; $L__BB215_781: mov.u32 %r1740, 81; { // callseq 1466, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1740; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1466 bra.uni $L__BB215_622; $L__BB215_780: mov.u32 %r1739, 82; { // callseq 1465, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r1739; st.param.b32 [param2+4], %r1727; st.param.b32 [param2+8], %r1728; st.param.b32 [param2+12], %r1729; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1465 bra.uni $L__BB215_622; $L__BB215_808: { // callseq 1500, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1500 bra.uni $L__BB215_622; $L__BB215_802: mov.u32 %r1764, 56; { // callseq 1489, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1764; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1489 bra.uni $L__BB215_622; $L__BB215_801: mov.u32 %r1762, 57; add.s32 %r1763, %r312, 1; { // callseq 1488, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1762; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1763; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1488 bra.uni $L__BB215_622; $L__BB215_785: setp.eq.s32 %p616, %r1729, 255; @%p616 bra $L__BB215_622; add.s64 %rd142, %rd140, 4; add.s64 %rd143, %rd140, 8; setp.eq.s32 %p617, %r1728, 0; mul.wide.u32 %rd929, %r1729, 4; add.s64 %rd144, %rd138, %rd929; @%p617 bra $L__BB215_790; setp.eq.s32 %p618, %r1728, 1; @%p618 bra $L__BB215_789; bra.uni $L__BB215_788; $L__BB215_789: ld.local.f32 %f2658, [%rd142]; st.local.f32 [%rd144], %f2658; bra.uni $L__BB215_622; $L__BB215_783: setp.eq.s32 %p615, %r1729, 255; @%p615 bra $L__BB215_622; ld.local.f32 %f2656, [%rd140]; add.s32 %r1743, %r1729, %r1728; mul.wide.u32 %rd925, %r1743, 4; add.s64 %rd926, %rd138, %rd925; st.local.f32 [%rd926], %f2656; bra.uni $L__BB215_622; $L__BB215_796: add.s32 %r1751, %r312, 1; { // callseq 1479, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1751; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1479 bra.uni $L__BB215_622; $L__BB215_795: mov.u32 %r1750, 67; { // callseq 1478, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r1750, %r1727, %r1728, %r1729}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1478 bra.uni $L__BB215_622; $L__BB215_778: add.s32 %r1737, %r312, 1; { // callseq 1461, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1737; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1461 bra.uni $L__BB215_622; $L__BB215_777: add.s32 %r1736, %r312, 1; { // callseq 1460, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1736; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1460 bra.uni $L__BB215_622; $L__BB215_810: { // callseq 1503, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1503 bra.uni $L__BB215_622; $L__BB215_809: add.s32 %r1770, %r312, 1; { // callseq 1502, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1770; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1502 bra.uni $L__BB215_622; $L__BB215_806: { // callseq 1496, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd713; .param .b64 param1; st.param.b64 [param1+0], %rd899; .param .b32 param2; st.param.b32 [param2+0], %r1727; .param .b32 param3; st.param.b32 [param3+0], %r1728; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1496 bra.uni $L__BB215_622; $L__BB215_794: mov.u32 %r1747, 69; add.s32 %r1748, %r312, 1; { // callseq 1476, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1747; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1748; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1476 bra.uni $L__BB215_622; $L__BB215_800: add.s32 %r1760, %r312, 1; { // callseq 1486, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1760; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj290EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1486 bra.uni $L__BB215_622; $L__BB215_782: add.s32 %r1741, %r312, 1; { // callseq 1468, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; .param .b32 param4; st.param.b32 [param4+0], %r1741; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1468 bra.uni $L__BB215_622; $L__BB215_807: cvt.u32.u64 %r1767, %rd4; { // callseq 1498, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r1767; .param .b64 param1; st.param.b64 [param1+0], %rd713; .param .b64 param2; st.param.b64 [param2+0], %rd899; .param .b32 param3; st.param.b32 [param3+0], %r1727; .param .b32 param4; st.param.b32 [param4+0], %r1728; call.uni _Z19svm_node_light_pathILj290EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1498 bra.uni $L__BB215_622; $L__BB215_803: add.s32 %r1765, %r312, 1; { // callseq 1491, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1765; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1491 bra.uni $L__BB215_622; $L__BB215_791: { // callseq 1471, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; .param .b32 param3; st.param.b32 [param3+0], %r1729; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1471 bra.uni $L__BB215_622; $L__BB215_797: mov.u32 %r1752, 64; add.s32 %r1753, %r312, 1; { // callseq 1481, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1752; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; .param .b32 param2; st.param.b32 [param2+0], %r1753; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r2220, [retval0+0]; } // callseq 1481 bra.uni $L__BB215_622; $L__BB215_779: { // callseq 1463, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .b32 param1; st.param.b32 [param1+0], %r1727; .param .b32 param2; st.param.b32 [param2+0], %r1728; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1463 bra.uni $L__BB215_622; $L__BB215_813: mov.u32 %r1774, 26; { // callseq 1508, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd899; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r1774; st.param.b32 [param1+4], %r1727; st.param.b32 [param1+8], %r1728; st.param.b32 [param1+12], %r1729; call.uni _Z17svm_node_set_bumpILj290EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1508 bra.uni $L__BB215_622; $L__BB215_790: ld.local.f32 %f2659, [%rd140]; st.local.f32 [%rd144], %f2659; bra.uni $L__BB215_622; $L__BB215_788: ld.local.f32 %f2657, [%rd143]; st.local.f32 [%rd144], %f2657; bra.uni $L__BB215_622; $L__BB215_827: add.s64 %rd1006, %rd713, 52; cvta.to.local.u64 %rd1007, %rd1006; ld.local.u32 %r1801, [%rd1007]; and.b32 %r402, %r1801, 2; @%p471 bra $L__BB215_830; bra.uni $L__BB215_828; $L__BB215_830: setp.eq.s32 %p624, %r402, 0; mov.f32 %f3260, 0f00000000; mov.f32 %f3261, %f3260; mov.f32 %f3262, %f3260; @%p624 bra $L__BB215_832; add.s64 %rd1018, %rd713, 280; cvta.to.local.u64 %rd1019, %rd1018; ld.local.v2.f32 {%f3260, %f3261}, [%rd1019]; add.s64 %rd1020, %rd713, 288; cvta.to.local.u64 %rd1021, %rd1020; ld.local.f32 %f3262, [%rd1021]; bra.uni $L__BB215_832; $L__BB215_828: setp.eq.s32 %p622, %r402, 0; mov.f32 %f3260, 0f00000000; mov.f32 %f3261, %f3260; mov.f32 %f3262, %f3260; @%p622 bra $L__BB215_832; add.s64 %rd1008, %rd107, 20; add.s64 %rd1009, %rd107, 12; cvta.to.local.u64 %rd1010, %rd1009; ld.local.v2.f32 {%f2675, %f2676}, [%rd1010]; cvta.to.local.u64 %rd1011, %rd1008; ld.local.v4.f32 {%f2679, %f2680, %f2681, %f2682}, [%rd1011]; mul.ftz.f32 %f2687, %f2676, %f2681; fma.rn.ftz.f32 %f2688, %f2675, %f2680, %f2687; fma.rn.ftz.f32 %f2689, %f2679, %f2682, %f2688; abs.ftz.f32 %f2690, %f2689; setp.gt.ftz.f32 %p623, %f2690, 0f00000000; selp.f32 %f2691, 0f3F800000, 0f00000000, %p623; add.s64 %rd1013, %rd713, 280; cvta.to.local.u64 %rd1014, %rd1013; ld.local.v2.f32 {%f2692, %f2693}, [%rd1014]; add.s64 %rd1015, %rd713, 288; cvta.to.local.u64 %rd1016, %rd1015; ld.local.f32 %f2696, [%rd1016]; mul.ftz.f32 %f3262, %f2691, %f2696; mul.ftz.f32 %f3261, %f2691, %f2693; mul.ftz.f32 %f3260, %f2692, %f2691; $L__BB215_832: ld.local.f32 %f2702, [%rd31+52]; mul.ftz.f32 %f3263, %f3260, %f2702; mul.ftz.f32 %f3264, %f3261, %f2702; mul.ftz.f32 %f3265, %f3262, %f2702; ld.local.u32 %r403, [%rd31+68]; setp.eq.s32 %p625, %r403, -1; @%p625 bra $L__BB215_834; ld.const.u64 %rd1024, [kernel_params+2368]; cvta.to.global.u64 %rd1025, %rd1024; mul.wide.s32 %rd1026, %r403, 192; add.s64 %rd1027, %rd1025, %rd1026; ld.global.f32 %f2703, [%rd1027+28]; mul.ftz.f32 %f3263, %f3263, %f2703; ld.global.f32 %f2704, [%rd1027+32]; mul.ftz.f32 %f3264, %f3264, %f2704; ld.global.f32 %f2705, [%rd1027+36]; mul.ftz.f32 %f3265, %f3265, %f2705; $L__BB215_834: setp.eq.ftz.f32 %p626, %f3264, 0f00000000; setp.eq.ftz.f32 %p627, %f3263, 0f00000000; and.pred %p628, %p627, %p626; setp.eq.ftz.f32 %p629, %f3265, 0f00000000; and.pred %p630, %p629, %p628; @%p630 bra $L__BB215_906; ld.local.u32 %r404, [%rd49+176]; setp.lt.s32 %p631, %r404, 1; mov.f32 %f3326, 0f00000000; mov.f32 %f3270, %f3326; mov.f32 %f3269, %f3326; mov.f32 %f3268, %f3326; mov.f32 %f3267, %f3326; mov.f32 %f3325, %f3326; @%p631 bra $L__BB215_876; ld.local.v2.f32 {%f2717, %f2718}, [%rd23+24]; mov.u32 %r2223, 0; ld.local.f32 %f2721, [%rd23+20]; ld.local.f32 %f2722, [%rd31+24]; mul.ftz.f32 %f2723, %f2722, %f2721; ld.local.f32 %f2724, [%rd31+28]; mul.ftz.f32 %f2725, %f2724, %f2717; neg.ftz.f32 %f2726, %f2725; sub.ftz.f32 %f2727, %f2726, %f2723; ld.local.f32 %f2728, [%rd31+32]; mul.ftz.f32 %f2729, %f2728, %f2718; sub.ftz.f32 %f782, %f2727, %f2729; and.b32 %r2225, %r404, 3; add.s32 %r1803, %r404, -1; setp.lt.u32 %p632, %r1803, 3; mov.f32 %f3325, 0f00000000; mov.f32 %f3267, %f3325; mov.f32 %f3268, %f3325; mov.f32 %f3269, %f3325; mov.f32 %f3270, %f3325; @%p632 bra $L__BB215_867; sub.s32 %r2222, %r404, %r2225; mov.f32 %f3325, 0f00000000; mov.u32 %r2223, 0; mov.f32 %f2735, 0f3DA2F983; mov.f32 %f3267, %f3325; mov.f32 %f3268, %f3325; mov.f32 %f3269, %f3325; mov.f32 %f3270, %f3325; $L__BB215_838: cvt.s64.s32 %rd146, %r2223; mul.wide.s32 %rd1031, %r2223, 20; add.s64 %rd1032, %rd42, %rd1031; add.s64 %rd147, %rd1032, 36; ld.local.f32 %f788, [%rd1032+36]; abs.ftz.f32 %f2736, %f788; setp.lt.ftz.f32 %p633, %f2736, 0f3A83126F; mov.f32 %f3272, %f2735; @%p633 bra $L__BB215_842; mul.ftz.f32 %f789, %f788, %f788; add.ftz.f32 %f2738, %f789, 0f3F800000; mul.ftz.f32 %f2739, %f788, 0fC0000000; fma.rn.ftz.f32 %f790, %f2739, %f782, %f2738; setp.le.ftz.f32 %p634, %f790, 0f00000000; mov.f32 %f3271, 0f00000000; @%p634 bra $L__BB215_841; lg2.approx.ftz.f32 %f2740, %f790; mul.ftz.f32 %f2741, %f2740, 0f3FC00000; ex2.approx.ftz.f32 %f3271, %f2741; $L__BB215_841: mov.f32 %f2742, 0f3F800000; sub.ftz.f32 %f2743, %f2742, %f789; div.approx.ftz.f32 %f2744, %f2743, %f3271; mul.ftz.f32 %f3272, %f2744, 0f3DA2F983; $L__BB215_842: setp.neu.ftz.f32 %p635, %f3272, 0f00000000; @%p635 bra $L__BB215_844; bra.uni $L__BB215_843; $L__BB215_844: add.ftz.f32 %f3270, %f3270, %f3272; add.ftz.f32 %f3269, %f3269, %f3272; add.ftz.f32 %f3268, %f3268, %f3272; ld.local.f32 %f3273, [%rd147+-4]; fma.rn.ftz.f32 %f3267, %f3272, %f3273, %f3267; bra.uni $L__BB215_845; $L__BB215_843: ld.local.f32 %f3273, [%rd147+-4]; $L__BB215_845: add.ftz.f32 %f806, %f3325, %f3273; ld.local.f32 %f807, [%rd147+20]; abs.ftz.f32 %f2746, %f807; setp.lt.ftz.f32 %p636, %f2746, 0f3A83126F; mov.f32 %f3279, 0f3DA2F983; @%p636 bra $L__BB215_849; mul.ftz.f32 %f808, %f807, %f807; add.ftz.f32 %f2748, %f808, 0f3F800000; mul.ftz.f32 %f2749, %f807, 0fC0000000; fma.rn.ftz.f32 %f809, %f2749, %f782, %f2748; setp.le.ftz.f32 %p637, %f809, 0f00000000; mov.f32 %f3278, 0f00000000; @%p637 bra $L__BB215_848; lg2.approx.ftz.f32 %f2750, %f809; mul.ftz.f32 %f2751, %f2750, 0f3FC00000; ex2.approx.ftz.f32 %f3278, %f2751; $L__BB215_848: mov.f32 %f2752, 0f3F800000; sub.ftz.f32 %f2753, %f2752, %f808; div.approx.ftz.f32 %f2754, %f2753, %f3278; mul.ftz.f32 %f3279, %f2754, 0f3DA2F983; $L__BB215_849: setp.neu.ftz.f32 %p638, %f3279, 0f00000000; @%p638 bra $L__BB215_851; bra.uni $L__BB215_850; $L__BB215_851: add.ftz.f32 %f3270, %f3270, %f3279; add.ftz.f32 %f3269, %f3269, %f3279; add.ftz.f32 %f3268, %f3268, %f3279; ld.local.f32 %f3280, [%rd147+16]; fma.rn.ftz.f32 %f3267, %f3279, %f3280, %f3267; bra.uni $L__BB215_852; $L__BB215_850: ld.local.f32 %f3280, [%rd147+16]; $L__BB215_852: add.ftz.f32 %f825, %f806, %f3280; ld.local.f32 %f826, [%rd147+40]; abs.ftz.f32 %f2756, %f826; setp.lt.ftz.f32 %p639, %f2756, 0f3A83126F; mov.f32 %f3286, 0f3DA2F983; @%p639 bra $L__BB215_856; mul.ftz.f32 %f827, %f826, %f826; add.ftz.f32 %f2758, %f827, 0f3F800000; mul.ftz.f32 %f2759, %f826, 0fC0000000; fma.rn.ftz.f32 %f828, %f2759, %f782, %f2758; setp.le.ftz.f32 %p640, %f828, 0f00000000; mov.f32 %f3285, 0f00000000; @%p640 bra $L__BB215_855; lg2.approx.ftz.f32 %f2760, %f828; mul.ftz.f32 %f2761, %f2760, 0f3FC00000; ex2.approx.ftz.f32 %f3285, %f2761; $L__BB215_855: mov.f32 %f2762, 0f3F800000; sub.ftz.f32 %f2763, %f2762, %f827; div.approx.ftz.f32 %f2764, %f2763, %f3285; mul.ftz.f32 %f3286, %f2764, 0f3DA2F983; $L__BB215_856: setp.neu.ftz.f32 %p641, %f3286, 0f00000000; @%p641 bra $L__BB215_858; bra.uni $L__BB215_857; $L__BB215_858: add.ftz.f32 %f3270, %f3270, %f3286; add.ftz.f32 %f3269, %f3269, %f3286; add.ftz.f32 %f3268, %f3268, %f3286; ld.local.f32 %f3287, [%rd147+36]; fma.rn.ftz.f32 %f3267, %f3286, %f3287, %f3267; bra.uni $L__BB215_859; $L__BB215_857: ld.local.f32 %f3287, [%rd147+36]; $L__BB215_859: add.ftz.f32 %f844, %f825, %f3287; ld.local.f32 %f845, [%rd147+60]; abs.ftz.f32 %f2766, %f845; setp.lt.ftz.f32 %p642, %f2766, 0f3A83126F; mov.f32 %f3293, 0f3DA2F983; @%p642 bra $L__BB215_863; mul.ftz.f32 %f846, %f845, %f845; add.ftz.f32 %f2768, %f846, 0f3F800000; mul.ftz.f32 %f2769, %f845, 0fC0000000; fma.rn.ftz.f32 %f847, %f2769, %f782, %f2768; setp.le.ftz.f32 %p643, %f847, 0f00000000; mov.f32 %f3292, 0f00000000; @%p643 bra $L__BB215_862; lg2.approx.ftz.f32 %f2770, %f847; mul.ftz.f32 %f2771, %f2770, 0f3FC00000; ex2.approx.ftz.f32 %f3292, %f2771; $L__BB215_862: mov.f32 %f2772, 0f3F800000; sub.ftz.f32 %f2773, %f2772, %f846; div.approx.ftz.f32 %f2774, %f2773, %f3292; mul.ftz.f32 %f3293, %f2774, 0f3DA2F983; $L__BB215_863: setp.neu.ftz.f32 %p644, %f3293, 0f00000000; @%p644 bra $L__BB215_865; bra.uni $L__BB215_864; $L__BB215_865: add.ftz.f32 %f3270, %f3270, %f3293; add.ftz.f32 %f3269, %f3269, %f3293; add.ftz.f32 %f3268, %f3268, %f3293; ld.local.f32 %f3294, [%rd147+56]; fma.rn.ftz.f32 %f3267, %f3293, %f3294, %f3267; bra.uni $L__BB215_866; $L__BB215_864: ld.local.f32 %f3294, [%rd147+56]; $L__BB215_866: add.ftz.f32 %f3325, %f844, %f3294; cvt.u32.u64 %r1805, %rd146; add.s32 %r2223, %r1805, 4; add.s32 %r2222, %r2222, -4; setp.ne.s32 %p645, %r2222, 0; @%p645 bra $L__BB215_838; $L__BB215_867: setp.eq.s32 %p646, %r2225, 0; @%p646 bra $L__BB215_876; $L__BB215_868: .pragma "nounroll"; cvt.s64.s32 %rd149, %r2223; mul.wide.s32 %rd1034, %r2223, 20; add.s64 %rd1035, %rd42, %rd1034; add.s64 %rd150, %rd1035, 36; ld.local.f32 %f879, [%rd1035+36]; abs.ftz.f32 %f2776, %f879; setp.lt.ftz.f32 %p647, %f2776, 0f3A83126F; mov.f32 %f3315, 0f3DA2F983; @%p647 bra $L__BB215_872; mul.ftz.f32 %f880, %f879, %f879; add.ftz.f32 %f2778, %f880, 0f3F800000; mul.ftz.f32 %f2779, %f879, 0fC0000000; fma.rn.ftz.f32 %f881, %f2779, %f782, %f2778; setp.le.ftz.f32 %p648, %f881, 0f00000000; mov.f32 %f3314, 0f00000000; @%p648 bra $L__BB215_871; lg2.approx.ftz.f32 %f2780, %f881; mul.ftz.f32 %f2781, %f2780, 0f3FC00000; ex2.approx.ftz.f32 %f3314, %f2781; $L__BB215_871: mov.f32 %f2782, 0f3F800000; sub.ftz.f32 %f2783, %f2782, %f880; div.approx.ftz.f32 %f2784, %f2783, %f3314; mul.ftz.f32 %f3315, %f2784, 0f3DA2F983; $L__BB215_872: setp.neu.ftz.f32 %p649, %f3315, 0f00000000; @%p649 bra $L__BB215_874; bra.uni $L__BB215_873; $L__BB215_874: add.ftz.f32 %f3270, %f3270, %f3315; add.ftz.f32 %f3269, %f3269, %f3315; add.ftz.f32 %f3268, %f3268, %f3315; ld.local.f32 %f3316, [%rd150+-4]; fma.rn.ftz.f32 %f3267, %f3315, %f3316, %f3267; bra.uni $L__BB215_875; $L__BB215_873: ld.local.f32 %f3316, [%rd150+-4]; $L__BB215_875: add.ftz.f32 %f3325, %f3325, %f3316; cvt.u32.u64 %r1806, %rd149; add.s32 %r2223, %r1806, 1; add.s32 %r2225, %r2225, -1; setp.ne.s32 %p650, %r2225, 0; @%p650 bra $L__BB215_868; $L__BB215_876: setp.leu.ftz.f32 %p651, %f3325, 0f00000000; @%p651 bra $L__BB215_878; div.approx.ftz.f32 %f3326, %f3267, %f3325; $L__BB215_878: and.b32 %r1807, %r297, 268435456; setp.eq.s32 %p652, %r1807, 0; @%p652 bra $L__BB215_880; ld.local.f32 %f2786, [%rd31+48]; mul.ftz.f32 %f2787, %f2786, %f2786; fma.rn.ftz.f32 %f2788, %f3326, %f3326, %f2787; div.approx.ftz.f32 %f2789, %f2787, %f2788; mul.ftz.f32 %f3270, %f3270, %f2789; mul.ftz.f32 %f3269, %f3269, %f2789; mul.ftz.f32 %f3268, %f3268, %f2789; $L__BB215_880: ld.local.f32 %f2791, [%rd31+48]; div.approx.ftz.f32 %f2792, %f3263, %f2791; div.approx.ftz.f32 %f2793, %f3264, %f2791; div.approx.ftz.f32 %f2794, %f3265, %f2791; mul.ftz.f32 %f3331, %f3270, %f2792; mul.ftz.f32 %f3332, %f3269, %f2793; mul.ftz.f32 %f3333, %f3268, %f2794; ld.const.f32 %f914, [kernel_params+1508]; setp.leu.ftz.f32 %p653, %f914, 0f00000000; mov.f32 %f3330, 0f00000000; @%p653 bra $L__BB215_885; cvt.u32.u16 %r2150, %rs1; add.s32 %r416, %r2150, 2; @%p44 bra $L__BB215_884; ld.const.v2.u32 {%r1809, %r1810}, [kernel_params+1600]; mov.b32 %f915, %r1810; setp.geu.ftz.f32 %p655, %f915, 0f3F800000; setp.lt.ftz.f32 %p656, %f915, 0f3F800000; ld.const.u32 %r1813, [kernel_params+1576]; selp.b32 %r1814, %r1813, %r6, %p656; add.s16 %rs62, %rs1, 2; and.b16 %rs63, %rs62, 255; cvt.u32.u16 %r1815, %rs63; xor.b32 %r1816, %r1814, %r1815; mul.lo.s32 %r1817, %r1816, -512718531; shr.u32 %r1818, %r1814, 16; xor.b32 %r1819, %r1817, %r1818; bfe.u32 %r1820, %r1819, 4, 4; shr.u32 %r1821, %r1814, 8; xor.b32 %r1822, %r1819, %r1821; xor.b32 %r1823, %r1822, %r1820; mul.lo.s32 %r1824, %r1823, 153742143; shr.u32 %r1825, %r1814, 23; xor.b32 %r1826, %r1824, %r1825; bfe.u32 %r1827, %r1826, 1, 7; xor.b32 %r1828, %r1827, %r1826; shr.u32 %r1829, %r1814, 27; or.b32 %r1830, %r1829, 1; mul.lo.s32 %r1831, %r1830, %r1828; mul.lo.s32 %r1832, %r1831, -1074439621; bfe.u32 %r1833, %r1832, 2, 6; xor.b32 %r1834, %r1833, %r1832; mul.lo.s32 %r1835, %r1834, -1638916925; bfe.u32 %r1836, %r1835, 2, 6; xor.b32 %r1837, %r1836, %r1835; mul.lo.s32 %r1838, %r1837, -933190689; and.b32 %r1839, %r1838, 255; bfe.u32 %r1840, %r1838, 5, 3; xor.b32 %r1841, %r1840, %r1839; add.s32 %r1842, %r1809, -1; xor.b32 %r417, %r416, 61; xor.b32 %r1843, %r1814, %r417; mul.lo.s32 %r1844, %r1843, 9; shr.u32 %r1845, %r1844, 4; xor.b32 %r1846, %r1845, %r1844; mul.lo.s32 %r1847, %r1846, 668265261; add.s32 %r1848, %r1847, %r47; shr.u32 %r1849, %r1847, 16; or.b32 %r1850, %r1849, 1; mul.lo.s32 %r1851, %r1850, %r1848; mul.lo.s32 %r1852, %r1851, 89287766; xor.b32 %r1853, %r1852, %r1851; mul.lo.s32 %r1854, %r1853, 1403136100; xor.b32 %r1855, %r1854, %r1853; brev.b32 %r1856, %r1855; neg.s32 %r1857, %r1809; and.b32 %r1858, %r8, %r1857; and.b32 %r1859, %r1856, %r1842; or.b32 %r1860, %r1859, %r1858; mad.lo.s32 %r1861, %r1841, %r1809, %r1860; shl.b32 %r1862, %r1809, 8; rem.u32 %r1863, %r1861, %r1862; ld.const.u64 %rd1040, [kernel_params+2424]; cvta.to.global.u64 %rd1041, %rd1040; shl.b32 %r1864, %r1863, 1; mul.wide.u32 %rd1042, %r1864, 4; add.s64 %rd1043, %rd1041, %rd1042; ld.global.f32 %f3330, [%rd1043]; @%p655 bra $L__BB215_885; xor.b32 %r1865, %r417, %r6; mul.lo.s32 %r1866, %r1865, 9; shr.u32 %r1867, %r1866, 4; xor.b32 %r1868, %r1867, %r1866; mul.lo.s32 %r1869, %r1868, 668265261; cvt.rn.f32.u32 %f2795, %r1869; mul.ftz.f32 %f2796, %f2795, 0f2F7FFFFE; fma.rn.ftz.f32 %f2797, %f2796, %f915, %f3330; cvt.rmi.ftz.f32.f32 %f2798, %f2797; sub.ftz.f32 %f3330, %f2797, %f2798; bra.uni $L__BB215_885; $L__BB215_884: shr.u32 %r1870, %r416, 16; xor.b32 %r1871, %r1870, %r416; mul.lo.s32 %r1872, %r1871, 569420461; shr.u32 %r1873, %r1872, 15; xor.b32 %r1874, %r1873, %r1872; mul.lo.s32 %r1875, %r1874, -749064809; shr.u32 %r1876, %r1875, 15; xor.b32 %r1877, %r6, %r1875; xor.b32 %r1878, %r1877, %r1876; xor.b32 %r1879, %r1878, 1493655573; add.s32 %r1880, %r1879, %r47; shr.u32 %r1881, %r1879, 16; or.b32 %r1882, %r1881, 1; mul.lo.s32 %r1883, %r1882, %r1880; mul.lo.s32 %r1884, %r1883, 89287766; xor.b32 %r1885, %r1884, %r1883; mul.lo.s32 %r1886, %r1885, 1403136100; xor.b32 %r1887, %r1886, %r1885; xor.b32 %r1888, %r1878, -2052961194; brev.b32 %r1889, %r1887; mul.lo.s32 %r1890, %r1889, 1025551850; xor.b32 %r1891, %r1890, %r1889; add.s32 %r1892, %r1891, %r1888; shr.u32 %r1893, %r1888, 16; or.b32 %r1894, %r1893, 1; mul.lo.s32 %r1895, %r1892, %r1894; mul.lo.s32 %r1896, %r1895, 89287766; xor.b32 %r1897, %r1896, %r1895; mul.lo.s32 %r1898, %r1897, 1403136100; xor.b32 %r1899, %r1898, %r1897; brev.b32 %r1900, %r1899; cvt.rn.f32.u32 %f2799, %r1900; mul.ftz.f32 %f3330, %f2799, 0f2F7FFFFE; $L__BB215_885: setp.eq.ftz.f32 %p658, %f3332, 0f00000000; setp.eq.ftz.f32 %p659, %f3331, 0f00000000; and.pred %p660, %p659, %p658; setp.eq.ftz.f32 %p661, %f3333, 0f00000000; and.pred %p662, %p661, %p660; mov.pred %p657, 0; mov.pred %p800, %p657; @%p662 bra $L__BB215_890; mov.pred %p663, -1; mov.pred %p800, %p663; @%p653 bra $L__BB215_890; abs.ftz.f32 %f2800, %f3331; abs.ftz.f32 %f2801, %f3332; max.ftz.f32 %f2802, %f2800, %f2801; abs.ftz.f32 %f2803, %f3333; max.ftz.f32 %f2804, %f2802, %f2803; mul.ftz.f32 %f920, %f2804, %f914; setp.geu.ftz.f32 %p666, %f920, 0f3F800000; mov.pred %p800, %p663; @%p666 bra $L__BB215_890; setp.le.ftz.f32 %p668, %f920, %f3330; mov.pred %p800, %p657; @%p668 bra $L__BB215_890; rcp.approx.ftz.f32 %f2805, %f920; mul.ftz.f32 %f3331, %f3331, %f2805; mul.ftz.f32 %f3332, %f3332, %f2805; mul.ftz.f32 %f3333, %f3333, %f2805; mov.pred %p800, %p663; $L__BB215_890: not.pred %p670, %p800; @%p670 bra $L__BB215_906; and.b32 %r1901, %r297, 1073741824; setp.eq.s32 %p671, %r1901, 0; mov.f32 %f3334, 0f00000000; mov.f32 %f3335, %f3334; mov.f32 %f3336, %f3334; mov.f32 %f3337, %f3334; mov.f32 %f3338, %f3334; mov.f32 %f3339, %f3334; mov.f32 %f3340, %f3334; @%p671 bra $L__BB215_895; ld.local.f32 %f2813, [%rd31+36]; setp.eq.ftz.f32 %p672, %f2813, 0f7F7FFFFF; @%p672 bra $L__BB215_894; bra.uni $L__BB215_893; $L__BB215_894: ld.local.f32 %f3337, [%rd31+24]; ld.local.f32 %f3338, [%rd31+28]; ld.local.f32 %f3339, [%rd31+32]; mov.f32 %f3340, 0f7F7FFFFF; mov.f32 %f3334, %f556; mov.f32 %f3335, %f557; mov.f32 %f3336, %f558; bra.uni $L__BB215_895; $L__BB215_893: ld.local.f32 %f2814, [%rd31]; sub.ftz.f32 %f2815, %f2814, %f556; ld.local.f32 %f2816, [%rd31+4]; sub.ftz.f32 %f2817, %f2816, %f557; ld.local.f32 %f2818, [%rd31+8]; sub.ftz.f32 %f2819, %f2818, %f558; mul.ftz.f32 %f2820, %f2817, %f2817; fma.rn.ftz.f32 %f2821, %f2815, %f2815, %f2820; fma.rn.ftz.f32 %f2822, %f2819, %f2819, %f2821; sqrt.approx.ftz.f32 %f3340, %f2822; rsqrt.approx.ftz.f32 %f2823, %f2822; mul.ftz.f32 %f3337, %f2815, %f2823; mul.ftz.f32 %f3338, %f2817, %f2823; mul.ftz.f32 %f3339, %f2819, %f2823; mov.f32 %f3334, %f556; mov.f32 %f3335, %f557; mov.f32 %f3336, %f558; $L__BB215_895: ld.local.f32 %f2825, [%rd23+76]; ld.local.f32 %f2826, [%rd23+68]; ld.local.u32 %r1902, [%rd31+56]; ld.local.u32 %r1903, [%rd31+60]; setp.ne.s32 %p673, %r1903, -1; mov.u32 %r1904, -1; ld.local.u32 %r418, [%rd31+76]; setp.eq.s32 %p674, %r418, 2; ld.const.u64 %rd1052, [kernel_params+4376]; cvta.to.global.u64 %rd1053, %rd1052; atom.global.add.u32 %r1905, [%rd1053], 1; ld.const.u64 %rd1054, [kernel_params+4256]; cvta.to.global.u64 %rd1055, %rd1054; add.s64 %rd1056, %rd1055, 12; atom.global.add.u32 %r1906, [%rd1056], 1; ld.const.u64 %rd1057, [kernel_params+3416]; cvta.to.global.u64 %rd1058, %rd1057; cvt.s64.s32 %rd151, %r1905; mul.wide.s32 %rd1059, %r1905, 2; add.s64 %rd1060, %rd1058, %rd1059; mov.u16 %rs64, 3; st.global.u16 [%rd1060], %rs64; ld.const.u64 %rd1061, [kernel_params+3496]; cvta.to.global.u64 %rd1062, %rd1061; mul.wide.s32 %rd1063, %r1905, 12; add.s64 %rd1064, %rd1062, %rd1063; st.global.f32 [%rd1064], %f3334; st.global.f32 [%rd1064+4], %f3335; st.global.f32 [%rd1064+8], %f3336; ld.const.v2.u64 {%rd1065, %rd1066}, [kernel_params+3504]; mov.u32 %r2228, 0; cvta.to.global.u64 %rd1069, %rd1065; add.s64 %rd1070, %rd1069, %rd1063; st.global.f32 [%rd1070], %f3337; st.global.f32 [%rd1070+4], %f3338; st.global.f32 [%rd1070+8], %f3339; cvta.to.global.u64 %rd1071, %rd1066; mul.wide.s32 %rd1072, %r1905, 4; add.s64 %rd1073, %rd1071, %rd1072; st.global.u32 [%rd1073], %r2228; ld.const.v2.u64 {%rd1074, %rd1075}, [kernel_params+3520]; cvta.to.global.u64 %rd1078, %rd1074; add.s64 %rd1079, %rd1078, %rd1072; st.global.f32 [%rd1079], %f3340; cvta.to.global.u64 %rd1080, %rd1075; add.s64 %rd1081, %rd1080, %rd1072; st.global.f32 [%rd1081], %f2826; ld.const.u64 %rd1082, [kernel_params+3536]; cvta.to.global.u64 %rd1083, %rd1082; add.s64 %rd1084, %rd1083, %rd1072; st.global.f32 [%rd1084], %f2825; ld.const.u64 %rd1085, [kernel_params+3584]; cvta.to.global.u64 %rd1086, %rd1085; add.s64 %rd1087, %rd1086, %rd1072; st.global.u32 [%rd1087], %r1904; ld.const.u64 %rd1088, [kernel_params+3576]; cvta.to.global.u64 %rd1089, %rd1088; add.s64 %rd1090, %rd1089, %rd1072; st.global.u32 [%rd1090], %r1904; ld.const.u64 %rd1091, [kernel_params+3632]; cvta.to.global.u64 %rd1092, %rd1091; add.s64 %rd1093, %rd1092, %rd1072; st.global.u32 [%rd1093], %r1902; ld.const.u64 %rd1094, [kernel_params+3624]; cvta.to.global.u64 %rd1095, %rd1094; add.s64 %rd1096, %rd1095, %rd1072; st.global.u32 [%rd1096], %r1903; ld.global.u16 %rs19, [%rd104]; ld.const.u64 %rd1097, [kernel_params+2472]; cvta.to.global.u64 %rd1098, %rd1097; add.s64 %rd1100, %rd1098, %rd263; ld.global.u16 %rs20, [%rd1100]; or.pred %p675, %p673, %p674; selp.b32 %r1908, 0, 134217728, %p675; ld.global.u32 %r419, [%rd30]; or.b32 %r2227, %r419, %r1908; ld.const.u32 %r421, [kernel_params]; and.b32 %r1909, %r421, 2097152; setp.eq.s32 %p676, %r1909, 0; @%p676 bra $L__BB215_900; and.b32 %r1910, %r419, 100663296; setp.eq.s32 %p677, %r1910, 0; @%p677 bra $L__BB215_898; ld.const.u64 %rd1101, [kernel_params+2600]; cvta.to.global.u64 %rd1102, %rd1101; add.s64 %rd1104, %rd1102, %rd339; ld.global.f32 %f3343, [%rd1104+8]; ld.global.f32 %f3342, [%rd1104+4]; ld.global.f32 %f3341, [%rd1104]; ld.const.u64 %rd1105, [kernel_params+2608]; cvta.to.global.u64 %rd1106, %rd1105; add.s64 %rd1107, %rd1106, %rd339; ld.global.f32 %f3346, [%rd1107+8]; ld.global.f32 %f3345, [%rd1107+4]; ld.global.f32 %f3344, [%rd1107]; bra.uni $L__BB215_899; $L__BB215_898: or.b32 %r2227, %r2227, 67108864; mov.f32 %f3344, 0f00000000; mov.f32 %f3341, 0f3F800000; mov.f32 %f3342, %f3341; mov.f32 %f3343, %f3341; mov.f32 %f3345, %f3344; mov.f32 %f3346, %f3344; $L__BB215_899: ld.const.u64 %rd1108, [kernel_params+3448]; cvta.to.global.u64 %rd1109, %rd1108; mul.lo.s64 %rd1110, %rd151, 12; add.s64 %rd1111, %rd1109, %rd1110; st.global.f32 [%rd1111], %f3341; st.global.f32 [%rd1111+4], %f3342; st.global.f32 [%rd1111+8], %f3343; ld.const.u64 %rd1112, [kernel_params+3456]; cvta.to.global.u64 %rd1113, %rd1112; add.s64 %rd1114, %rd1113, %rd1110; st.global.f32 [%rd1114], %f3344; st.global.f32 [%rd1114+4], %f3345; st.global.f32 [%rd1114+8], %f3346; $L__BB215_900: cvta.to.global.u64 %rd1721, %rd257; add.s64 %rd1720, %rd1721, %rd260; cvta.to.global.u64 %rd1719, %rd258; add.s64 %rd1718, %rd1719, %rd263; ld.const.u64 %rd1115, [kernel_params+2448]; cvta.to.global.u64 %rd1116, %rd1115; add.s64 %rd1118, %rd1116, %rd260; ld.global.u32 %r1911, [%rd1118]; ld.const.v2.u64 {%rd1119, %rd1120}, [kernel_params+3344]; cvta.to.global.u64 %rd1123, %rd1119; shl.b64 %rd1124, %rd151, 2; add.s64 %rd1125, %rd1123, %rd1124; st.global.u32 [%rd1125], %r1911; ld.global.u16 %rs65, [%rd1718]; ld.const.v2.u64 {%rd1129, %rd1130}, [kernel_params+3360]; cvta.to.global.u64 %rd1133, %rd1130; shl.b64 %rd1134, %rd151, 1; add.s64 %rd1135, %rd1133, %rd1134; st.global.u16 [%rd1135], %rs65; ld.global.u32 %r1912, [%rd1720]; cvta.to.global.u64 %rd1138, %rd1129; add.s64 %rd1139, %rd1138, %rd1124; st.global.u32 [%rd1139], %r1912; ld.global.u32 %r1913, [%rd28]; cvta.to.global.u64 %rd1140, %rd1120; add.s64 %rd1141, %rd1140, %rd1124; st.global.u32 [%rd1141], %r1913; ld.const.v2.u64 {%rd1142, %rd1143}, [kernel_params+3424]; cvta.to.global.u64 %rd1146, %rd1142; add.s64 %rd1147, %rd1146, %rd1124; st.global.u32 [%rd1147], %r2227; ld.const.v2.u64 {%rd1148, %rd1149}, [kernel_params+3376]; cvta.to.global.u64 %rd1152, %rd1148; add.s64 %rd1153, %rd1152, %rd1134; st.global.u16 [%rd1153], %rs19; cvta.to.global.u64 %rd1154, %rd1149; add.s64 %rd1155, %rd1154, %rd1134; st.global.u16 [%rd1155], %rs20; ld.const.v2.u64 {%rd1156, %rd1157}, [kernel_params+2480]; cvta.to.global.u64 %rd1160, %rd1156; add.s64 %rd1161, %rd1160, %rd263; ld.global.u16 %rs66, [%rd1161]; ld.const.v2.u64 {%rd1162, %rd1163}, [kernel_params+3392]; cvta.to.global.u64 %rd1166, %rd1162; add.s64 %rd1167, %rd1166, %rd1134; st.global.u16 [%rd1167], %rs66; cvta.to.global.u64 %rd1168, %rd1157; add.s64 %rd1169, %rd1168, %rd263; ld.global.u16 %rs67, [%rd1169]; cvta.to.global.u64 %rd1170, %rd1163; add.s64 %rd1171, %rd1170, %rd1134; st.global.u16 [%rd1171], %rs67; ld.const.u64 %rd1172, [kernel_params+2496]; cvta.to.global.u64 %rd1173, %rd1172; add.s64 %rd1174, %rd1173, %rd263; ld.global.u16 %rs68, [%rd1174]; ld.const.u64 %rd1175, [kernel_params+3408]; cvta.to.global.u64 %rd1176, %rd1175; add.s64 %rd1177, %rd1176, %rd1134; st.global.u16 [%rd1177], %rs68; cvta.to.global.u64 %rd1178, %rd1143; mul.lo.s64 %rd1179, %rd151, 12; add.s64 %rd1180, %rd1178, %rd1179; mul.ftz.f32 %f2833, %f559, %f3331; st.global.f32 [%rd1180], %f2833; mul.ftz.f32 %f2834, %f560, %f3332; st.global.f32 [%rd1180+4], %f2834; mul.ftz.f32 %f2835, %f561, %f3333; st.global.f32 [%rd1180+8], %f2835; and.b32 %r1914, %r421, 4194304; setp.eq.s32 %p678, %r1914, 0; @%p678 bra $L__BB215_902; ld.const.u64 %rd1181, [kernel_params+3440]; cvta.to.global.u64 %rd1182, %rd1181; add.s64 %rd1184, %rd1182, %rd1179; st.global.f32 [%rd1184], %f559; st.global.f32 [%rd1184+4], %f560; st.global.f32 [%rd1184+8], %f561; $L__BB215_902: ld.const.u32 %r1915, [kernel_params+1076]; ld.local.u32 %r1916, [%rd31+72]; selp.b32 %r1917, %r1915, %r1916, %p674; cvt.u16.u32 %rs69, %r1917; add.s16 %rs70, %rs69, 1; ld.const.u64 %rd1187, [kernel_params+3472]; cvta.to.global.u64 %rd1188, %rd1187; add.s64 %rd1189, %rd1188, %rd151; st.global.u8 [%rd1189], %rs70; and.b32 %r1918, %r421, 131072; setp.eq.s32 %p680, %r1918, 0; @%p680 bra $L__BB215_906; ld.const.u64 %rd1748, [kernel_params+2776]; mov.u64 %rd1194, kernel_params; bra.uni $L__BB215_904; $L__BB215_905: ld.const.u64 %rd1748, [%rd153+24]; add.s32 %r2228, %r2228, 1; $L__BB215_904: cvta.to.global.u64 %rd1190, %rd1748; add.s64 %rd1192, %rd1190, %rd260; ld.global.u32 %r1920, [%rd1192]; mul.wide.s32 %rd1193, %r2228, 16; add.s64 %rd1195, %rd1194, %rd1193; add.s64 %rd153, %rd1195, 2768; ld.const.u64 %rd1196, [%rd1195+2768]; cvta.to.global.u64 %rd1197, %rd1196; add.s64 %rd1198, %rd1197, %rd260; ld.global.u32 %r1921, [%rd1198]; ld.const.v2.u64 {%rd1199, %rd1200}, [%rd1195+3744]; cvta.to.global.u64 %rd1203, %rd1199; add.s64 %rd1205, %rd1203, %rd1124; st.global.u32 [%rd1205], %r1921; cvta.to.global.u64 %rd1206, %rd1200; add.s64 %rd1207, %rd1206, %rd1124; st.global.u32 [%rd1207], %r1920; setp.eq.s32 %p681, %r1920, -1; @%p681 bra $L__BB215_906; bra.uni $L__BB215_905; $L__BB215_906: ld.local.u8 %rs21, [%rd49+180]; setp.eq.s16 %p682, %rs21, 0; @%p682 bra $L__BB215_908; rcp.approx.ftz.f32 %f2836, %f3196; ld.local.f32 %f2837, [%rd49+184]; mul.ftz.f32 %f3347, %f2836, %f2837; ld.local.f32 %f2838, [%rd49+188]; mul.ftz.f32 %f2839, %f2836, %f2838; ld.local.f32 %f2840, [%rd49+192]; mul.ftz.f32 %f2841, %f2836, %f2840; st.local.f32 [%rd49+184], %f3347; st.local.f32 [%rd49+188], %f2839; st.local.f32 [%rd49+192], %f2841; bra.uni $L__BB215_909; $L__BB215_908: ld.local.f32 %f3347, [%rd49+184]; $L__BB215_909: ld.local.f32 %f2842, [%rd49+188]; ld.local.f32 %f2843, [%rd49+192]; st.global.f32 [%rd48], %f3347; st.global.f32 [%rd48+4], %f2842; st.global.f32 [%rd48+8], %f2843; mov.u32 %r2237, 1; @%p682 bra $L__BB215_936; cvt.u32.u16 %r2151, %rs1; ld.local.f32 %f2844, [%rd49+196]; fma.rn.ftz.f32 %f956, %f4, %f2844, %f1; fma.rn.ftz.f32 %f957, %f5, %f2844, %f2; fma.rn.ftz.f32 %f958, %f6, %f2844, %f3; st.local.v2.f32 [%rd1], {%f956, %f957}; st.local.f32 [%rd23+-8], %f958; add.s32 %r427, %r2151, 3; @%p44 bra $L__BB215_913; ld.const.v2.u32 {%r1924, %r1925}, [kernel_params+1600]; mov.b32 %f959, %r1925; setp.geu.ftz.f32 %p685, %f959, 0f3F800000; setp.lt.ftz.f32 %p686, %f959, 0f3F800000; ld.const.u32 %r1928, [kernel_params+1576]; selp.b32 %r1929, %r1928, %r6, %p686; add.s16 %rs71, %rs1, 3; and.b16 %rs72, %rs71, 255; cvt.u32.u16 %r1930, %rs72; xor.b32 %r1931, %r1929, %r1930; mul.lo.s32 %r1932, %r1931, -512718531; shr.u32 %r1933, %r1929, 16; xor.b32 %r1934, %r1932, %r1933; bfe.u32 %r1935, %r1934, 4, 4; shr.u32 %r1936, %r1929, 8; xor.b32 %r1937, %r1934, %r1936; xor.b32 %r1938, %r1937, %r1935; mul.lo.s32 %r1939, %r1938, 153742143; shr.u32 %r1940, %r1929, 23; xor.b32 %r1941, %r1939, %r1940; bfe.u32 %r1942, %r1941, 1, 7; xor.b32 %r1943, %r1942, %r1941; shr.u32 %r1944, %r1929, 27; or.b32 %r1945, %r1944, 1; mul.lo.s32 %r1946, %r1945, %r1943; mul.lo.s32 %r1947, %r1946, -1074439621; bfe.u32 %r1948, %r1947, 2, 6; xor.b32 %r1949, %r1948, %r1947; mul.lo.s32 %r1950, %r1949, -1638916925; bfe.u32 %r1951, %r1950, 2, 6; xor.b32 %r1952, %r1951, %r1950; mul.lo.s32 %r1953, %r1952, -933190689; and.b32 %r1954, %r1953, 255; bfe.u32 %r1955, %r1953, 5, 3; xor.b32 %r1956, %r1955, %r1954; add.s32 %r1957, %r1924, -1; xor.b32 %r428, %r427, 61; xor.b32 %r1958, %r1929, %r428; mul.lo.s32 %r1959, %r1958, 9; shr.u32 %r1960, %r1959, 4; xor.b32 %r1961, %r1960, %r1959; mul.lo.s32 %r1962, %r1961, 668265261; add.s32 %r1963, %r1962, %r47; shr.u32 %r1964, %r1962, 16; or.b32 %r1965, %r1964, 1; mul.lo.s32 %r1966, %r1965, %r1963; mul.lo.s32 %r1967, %r1966, 89287766; xor.b32 %r1968, %r1967, %r1966; mul.lo.s32 %r1969, %r1968, 1403136100; xor.b32 %r1970, %r1969, %r1968; brev.b32 %r1971, %r1970; neg.s32 %r1972, %r1924; and.b32 %r1973, %r8, %r1972; and.b32 %r1974, %r1971, %r1957; or.b32 %r1975, %r1974, %r1973; mad.lo.s32 %r1976, %r1956, %r1924, %r1975; shl.b32 %r1977, %r1924, 8; rem.u32 %r1978, %r1976, %r1977; ld.const.u64 %rd1210, [kernel_params+2424]; cvta.to.global.u64 %rd1211, %rd1210; shl.b32 %r1979, %r1978, 1; mul.wide.u32 %rd1212, %r1979, 4; add.s64 %rd1213, %rd1211, %rd1212; ld.global.f32 %f3359, [%rd1213]; ld.global.f32 %f3351, [%rd1213+4]; @%p685 bra $L__BB215_917; xor.b32 %r1980, %r428, %r6; mul.lo.s32 %r1981, %r1980, 9; shr.u32 %r1982, %r1981, 4; xor.b32 %r1983, %r1982, %r1981; mul.lo.s32 %r1984, %r1983, 668265261; cvt.rn.f32.u32 %f2845, %r1984; mul.ftz.f32 %f2846, %f2845, 0f2F7FFFFE; xor.b32 %r1985, %r6, %r427; xor.b32 %r1986, %r1985, -905047700; mul.lo.s32 %r1987, %r1986, 9; shr.u32 %r1988, %r1987, 4; xor.b32 %r1989, %r1988, %r1987; mul.lo.s32 %r1990, %r1989, 668265261; cvt.rn.f32.u32 %f2847, %r1990; mul.ftz.f32 %f2848, %f2847, 0f2F7FFFFE; fma.rn.ftz.f32 %f2849, %f2846, %f959, %f3359; fma.rn.ftz.f32 %f2850, %f2848, %f959, %f3351; cvt.rmi.ftz.f32.f32 %f2851, %f2849; sub.ftz.f32 %f3359, %f2849, %f2851; cvt.rmi.ftz.f32.f32 %f2852, %f2850; sub.ftz.f32 %f3351, %f2850, %f2852; bra.uni $L__BB215_917; $L__BB215_913: shr.u32 %r1992, %r427, 16; xor.b32 %r1993, %r1992, %r427; mul.lo.s32 %r1994, %r1993, 569420461; shr.u32 %r1995, %r1994, 15; xor.b32 %r1996, %r1995, %r1994; mul.lo.s32 %r1997, %r1996, -749064809; shr.u32 %r1998, %r1997, 15; xor.b32 %r1999, %r6, %r1997; xor.b32 %r2000, %r1999, %r1998; xor.b32 %r429, %r2000, -419546133; xor.b32 %r2001, %r2000, 508809841; add.s32 %r2002, %r2001, %r47; shr.u32 %r2003, %r2001, 16; or.b32 %r2004, %r2003, 1; mul.lo.s32 %r2005, %r2004, %r2002; mul.lo.s32 %r2006, %r2005, 89287766; xor.b32 %r2007, %r2006, %r2005; mul.lo.s32 %r2008, %r2007, 1403136100; xor.b32 %r430, %r2008, %r2007; setp.eq.s32 %p687, %r430, 0; mov.u32 %r2232, 0; @%p687 bra $L__BB215_916; mov.u32 %r2229, 0; mov.u64 %rd1215, sobol_burley_table; mov.u32 %r2230, %r430; mov.u32 %r2232, %r2229; $L__BB215_915: clz.b32 %r2011, %r2230; add.s32 %r2012, %r2011, %r2229; mul.wide.u32 %rd1214, %r2012, 4; add.s64 %rd1216, %rd1215, %rd1214; ld.const.u32 %r2013, [%rd1216+128]; xor.b32 %r2232, %r2013, %r2232; add.s32 %r2229, %r2012, 1; shl.b32 %r2014, %r2230, %r2011; shl.b32 %r2230, %r2014, 1; setp.ne.s32 %p688, %r2230, 0; @%p688 bra $L__BB215_915; $L__BB215_916: xor.b32 %r2015, %r429, -525684874; brev.b32 %r2016, %r430; shr.u32 %r2017, %r2015, 16; or.b32 %r2018, %r2017, 1; mul.lo.s32 %r2019, %r2016, 1025551850; xor.b32 %r2020, %r2019, %r2016; add.s32 %r2021, %r2020, %r2015; mul.lo.s32 %r2022, %r2021, %r2018; mul.lo.s32 %r2023, %r2022, 89287766; xor.b32 %r2024, %r2023, %r2022; xor.b32 %r2025, %r429, -1802089138; mul.lo.s32 %r2026, %r2232, 1025551850; xor.b32 %r2027, %r2026, %r2232; add.s32 %r2028, %r2027, %r2025; shr.u32 %r2029, %r2025, 16; or.b32 %r2030, %r2029, 1; mul.lo.s32 %r2031, %r2028, %r2030; mul.lo.s32 %r2032, %r2031, 89287766; xor.b32 %r2033, %r2032, %r2031; mul.lo.s32 %r2034, %r2033, 1403136100; xor.b32 %r2035, %r2034, %r2033; brev.b32 %r2036, %r2035; cvt.rn.f32.u32 %f2853, %r2036; mul.lo.s32 %r2037, %r2024, 1403136100; xor.b32 %r2038, %r2037, %r2024; brev.b32 %r2039, %r2038; cvt.rn.f32.u32 %f2854, %r2039; mul.ftz.f32 %f3351, %f2853, 0f2F7FFFFE; mul.ftz.f32 %f3359, %f2854, 0f2F7FFFFE; $L__BB215_917: ld.local.u32 %r438, [%rd49+360]; setp.lt.s32 %p689, %r438, 2; mov.u32 %r2236, 0; @%p689 bra $L__BB215_927; ld.local.f32 %f3357, [%rd49+212]; and.b32 %r2234, %r438, 3; add.s32 %r2041, %r438, -1; setp.lt.u32 %p690, %r2041, 3; mov.f32 %f3356, 0f00000000; @%p690 bra $L__BB215_921; sub.s32 %r2233, %r438, %r2234; mov.f32 %f3356, 0f00000000; $L__BB215_920: add.ftz.f32 %f2858, %f3356, %f3357; add.ftz.f32 %f2859, %f2858, %f3357; add.ftz.f32 %f2860, %f2859, %f3357; add.ftz.f32 %f3356, %f2860, %f3357; add.s32 %r2233, %r2233, -4; setp.ne.s32 %p691, %r2233, 0; @%p691 bra $L__BB215_920; $L__BB215_921: setp.eq.s32 %p692, %r2234, 0; @%p692 bra $L__BB215_923; $L__BB215_922: .pragma "nounroll"; add.ftz.f32 %f3356, %f3356, %f3357; add.s32 %r2234, %r2234, -1; setp.ne.s32 %p693, %r2234, 0; @%p693 bra $L__BB215_922; $L__BB215_923: mul.ftz.f32 %f981, %f3359, %f3356; mov.f32 %f3358, 0f00000000; mov.u32 %r2236, 0; mov.u32 %r445, %r2236; bra.uni $L__BB215_924; $L__BB215_1099: mul.wide.s32 %rd1260, %r445, 20; add.s64 %rd1261, %rd42, %rd1260; ld.local.f32 %f3357, [%rd1261+236]; mov.u32 %r445, %r446; mov.f32 %f3358, %f984; $L__BB215_924: add.ftz.f32 %f984, %f3358, %f3357; setp.gtu.ftz.f32 %p694, %f981, %f984; @%p694 bra $L__BB215_926; bra.uni $L__BB215_925; $L__BB215_926: add.s32 %r446, %r445, 1; setp.lt.s32 %p695, %r446, %r438; @%p695 bra $L__BB215_1099; bra.uni $L__BB215_927; $L__BB215_925: sub.ftz.f32 %f2862, %f981, %f3358; div.approx.ftz.f32 %f3359, %f2862, %f3357; mov.u32 %r2236, %r445; $L__BB215_927: ld.local.v2.f32 {%f2863, %f2864}, [%rd23+24]; mul.wide.s32 %rd1220, %r2236, 20; add.s64 %rd1221, %rd42, %rd1220; ld.local.f32 %f989, [%rd23+20]; ld.local.f32 %f990, [%rd1221+220]; abs.ftz.f32 %f2865, %f990; setp.lt.ftz.f32 %p696, %f2865, 0f3A83126F; @%p696 bra $L__BB215_931; bra.uni $L__BB215_928; $L__BB215_931: add.ftz.f32 %f2881, %f3359, %f3359; mov.f32 %f2882, 0f3F800000; sub.ftz.f32 %f3362, %f2882, %f2881; mov.f32 %f3361, 0f3DA2F983; bra.uni $L__BB215_932; $L__BB215_928: mul.ftz.f32 %f2867, %f990, %f990; mov.f32 %f2868, 0f3F800000; sub.ftz.f32 %f991, %f2868, %f2867; add.ftz.f32 %f2869, %f990, %f990; sub.ftz.f32 %f2870, %f2868, %f990; fma.rn.ftz.f32 %f2871, %f3359, %f2869, %f2870; div.approx.ftz.f32 %f2872, %f991, %f2871; add.ftz.f32 %f2873, %f2867, 0f3F800000; mul.ftz.f32 %f2874, %f2872, %f2872; sub.ftz.f32 %f2875, %f2873, %f2874; div.approx.ftz.f32 %f3362, %f2875, %f2869; mul.ftz.f32 %f2876, %f2869, %f3362; sub.ftz.f32 %f993, %f2873, %f2876; setp.le.ftz.f32 %p697, %f993, 0f00000000; mov.f32 %f3360, 0f00000000; @%p697 bra $L__BB215_930; lg2.approx.ftz.f32 %f2877, %f993; mul.ftz.f32 %f2878, %f2877, 0f3FC00000; ex2.approx.ftz.f32 %f3360, %f2878; $L__BB215_930: div.approx.ftz.f32 %f2879, %f991, %f3360; mul.ftz.f32 %f3361, %f2879, 0f3DA2F983; $L__BB215_932: neg.ftz.f32 %f2883, %f2863; neg.ftz.f32 %f2884, %f2864; mul.ftz.f32 %f2885, %f3362, %f3362; mov.f32 %f2886, 0f3F800000; sub.ftz.f32 %f2887, %f2886, %f2885; mov.f32 %f2888, 0f00000000; max.ftz.f32 %f2889, %f2887, %f2888; sqrt.approx.ftz.f32 %f2890, %f2889; mul.ftz.f32 %f2891, %f3351, 0f40C90FDB; cos.approx.ftz.f32 %f2892, %f2891; mul.ftz.f32 %f2893, %f2890, %f2892; sin.approx.ftz.f32 %f2894, %f2891; mul.ftz.f32 %f2895, %f2890, %f2894; setp.neu.ftz.f32 %p698, %f989, %f2864; setp.neu.ftz.f32 %p699, %f989, %f2863; or.pred %p700, %p699, %p698; selp.f32 %f2896, %f2864, %f2884, %p700; selp.f32 %f2897, %f2883, %f2863, %p700; add.ftz.f32 %f2898, %f989, %f2897; sub.ftz.f32 %f2899, %f2896, %f989; sub.ftz.f32 %f2900, %f2863, %f2864; mul.ftz.f32 %f2901, %f2899, %f2899; fma.rn.ftz.f32 %f2902, %f2900, %f2900, %f2901; fma.rn.ftz.f32 %f2903, %f2898, %f2898, %f2902; sqrt.approx.ftz.f32 %f2904, %f2903; div.approx.ftz.f32 %f2905, %f2900, %f2904; div.approx.ftz.f32 %f2906, %f2899, %f2904; div.approx.ftz.f32 %f2907, %f2898, %f2904; mul.ftz.f32 %f2908, %f2907, %f2863; mul.ftz.f32 %f2909, %f2906, %f2864; sub.ftz.f32 %f2910, %f2909, %f2908; mul.ftz.f32 %f2911, %f2905, %f2864; mul.ftz.f32 %f2912, %f2907, %f989; sub.ftz.f32 %f2913, %f2912, %f2911; mul.ftz.f32 %f2914, %f2906, %f989; mul.ftz.f32 %f2915, %f2905, %f2863; sub.ftz.f32 %f2916, %f2915, %f2914; mul.ftz.f32 %f2917, %f2895, %f2910; mul.ftz.f32 %f2918, %f2895, %f2913; mul.ftz.f32 %f2919, %f2895, %f2916; fma.rn.ftz.f32 %f2920, %f2893, %f2905, %f2917; fma.rn.ftz.f32 %f2921, %f2893, %f2906, %f2918; fma.rn.ftz.f32 %f2922, %f2893, %f2907, %f2919; mul.ftz.f32 %f2923, %f3362, %f989; sub.ftz.f32 %f1000, %f2920, %f2923; mul.ftz.f32 %f2924, %f3362, %f2863; sub.ftz.f32 %f1001, %f2921, %f2924; mul.ftz.f32 %f2925, %f3362, %f2864; sub.ftz.f32 %f1002, %f2922, %f2925; setp.eq.ftz.f32 %p701, %f3361, 0f00000000; mov.u32 %r2237, %r1563; @%p701 bra $L__BB215_936; cvta.to.global.u64 %rd1722, %rd241; cvta.to.global.u64 %rd1697, %rd233; cvta.to.global.u64 %rd1696, %rd232; st.global.f32 [%rd5], %f956; st.global.f32 [%rd5+4], %f957; st.global.f32 [%rd5+8], %f958; mul.ftz.f32 %f2926, %f1001, %f1001; fma.rn.ftz.f32 %f2927, %f1000, %f1000, %f2926; fma.rn.ftz.f32 %f2928, %f1002, %f1002, %f2927; sqrt.approx.ftz.f32 %f2929, %f2928; div.approx.ftz.f32 %f2930, %f1000, %f2929; div.approx.ftz.f32 %f2931, %f1001, %f2929; div.approx.ftz.f32 %f2932, %f1002, %f2929; add.s64 %rd1224, %rd1696, %rd339; st.global.f32 [%rd1224], %f2930; st.global.f32 [%rd1224+4], %f2931; st.global.f32 [%rd1224+8], %f2932; add.s64 %rd1227, %rd1697, %rd260; mov.u32 %r2237, 0; st.global.u32 [%rd1227], %r2237; cvta.to.global.u64 %rd1228, %rd239; add.s64 %rd1229, %rd1228, %rd260; mov.u32 %r2046, 2139095039; st.global.u32 [%rd1229], %r2046; ld.local.f32 %f2933, [%rd23+76]; ld.const.u64 %rd1230, [kernel_params+2672]; cvta.to.global.u64 %rd1231, %rd1230; add.s64 %rd1232, %rd1231, %rd260; st.global.f32 [%rd1232], %f2933; ld.local.u32 %r2047, [%rd23+44]; st.global.u32 [%rd15], %r2047; ld.local.u32 %r2048, [%rd23+60]; add.s64 %rd1234, %rd1722, %rd260; st.global.u32 [%rd1234], %r2048; div.approx.ftz.f32 %f2934, %f3361, %f3361; ld.global.f32 %f2935, [%rd48]; mul.ftz.f32 %f2936, %f2934, %f2935; ld.global.f32 %f2937, [%rd48+4]; mul.ftz.f32 %f2938, %f2934, %f2937; ld.global.f32 %f2939, [%rd48+8]; mul.ftz.f32 %f2940, %f2934, %f2939; st.global.f32 [%rd48], %f2936; st.global.f32 [%rd48+4], %f2938; st.global.f32 [%rd48+8], %f2940; ld.const.u8 %rs73, [kernel_params+2]; and.b16 %rs74, %rs73, 32; setp.eq.s16 %p702, %rs74, 0; @%p702 bra $L__BB215_935; ld.const.u64 %rd1235, [kernel_params+2600]; cvta.to.global.u64 %rd1236, %rd1235; add.s64 %rd1238, %rd1236, %rd339; mov.u32 %r2049, 1065353216; st.global.u32 [%rd1238], %r2049; st.global.u32 [%rd1238+4], %r2049; st.global.u32 [%rd1238+8], %r2049; ld.const.u64 %rd1239, [kernel_params+2608]; cvta.to.global.u64 %rd1240, %rd1239; add.s64 %rd1241, %rd1240, %rd339; st.global.u32 [%rd1241], %r2237; st.global.u32 [%rd1241+4], %r2237; st.global.u32 [%rd1241+8], %r2237; $L__BB215_935: cvta.to.global.u64 %rd1724, %rd258; add.s64 %rd1723, %rd1724, %rd263; ld.const.v2.u64 {%rd1242, %rd1243}, [kernel_params+2560]; cvta.to.global.u64 %rd1246, %rd1242; add.s64 %rd1248, %rd1246, %rd260; st.global.f32 [%rd1248], %f3361; cvta.to.global.u64 %rd1249, %rd1243; add.s64 %rd1250, %rd1249, %rd260; ld.global.f32 %f2941, [%rd1250]; min.ftz.f32 %f2942, %f3361, %f2941; st.global.f32 [%rd1250], %f2942; ld.const.u64 %rd1251, [kernel_params+2464]; cvta.to.global.u64 %rd1252, %rd1251; add.s64 %rd1254, %rd1252, %rd263; ld.global.u16 %r2052, [%rd1254]; add.s32 %r2053, %r2052, 1; ld.const.u32 %r2054, [kernel_params+1516]; setp.lt.u32 %p703, %r2053, %r2054; ld.global.u32 %r2055, [%rd30]; or.b32 %r2056, %r2055, 131072; selp.b32 %r2057, %r2055, %r2056, %p703; and.b32 %r2058, %r2057, -20480; and.b32 %r2059, %r2057, 100663296; setp.eq.s32 %p704, %r2059, 0; selp.b32 %r2060, 67108992, 128, %p704; or.b32 %r2061, %r2060, %r2058; ld.const.u64 %rd1255, [kernel_params+2504]; cvta.to.global.u64 %rd1256, %rd1255; add.s64 %rd1257, %rd1256, %rd263; ld.global.u16 %r2062, [%rd1257]; add.s32 %r2063, %r2062, 1; st.global.u16 [%rd1257], %r2063; ld.const.u32 %r2064, [kernel_params+1532]; setp.lt.s32 %p705, %r2063, %r2064; or.b32 %r2065, %r2061, 131072; selp.b32 %r2066, %r2061, %r2065, %p705; st.global.u32 [%rd30], %r2066; st.global.u16 [%rd1254], %r2053; ld.global.u16 %rs75, [%rd1723]; add.s16 %rs76, %rs75, 16; st.global.u16 [%rd1723], %rs76; bra.uni $L__BB215_936; } // .globl kernel_gpu_integrator_queued_paths_array .visible .entry kernel_gpu_integrator_queued_paths_array( .param .u32 kernel_gpu_integrator_queued_paths_array_param_0, .param .u64 kernel_gpu_integrator_queued_paths_array_param_1, .param .u64 kernel_gpu_integrator_queued_paths_array_param_2, .param .u32 kernel_gpu_integrator_queued_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<13>; .reg .b32 %r<93>; .reg .b64 %rd<11>; ld.param.u32 %r32, [kernel_gpu_integrator_queued_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_queued_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_queued_paths_array_param_2]; ld.param.u32 %r30, [kernel_gpu_integrator_queued_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r32; mov.u32 %r83, 0; @%p1 bra $L__BB216_2; ld.const.u64 %rd3, [kernel_params+2520]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %r36, [%rd6]; setp.eq.s32 %p2, %r36, %r30; selp.u32 %r83, 1, 0, %p2; $L__BB216_2: div.u32 %r37, %r1, %r2; mul.lo.s32 %r38, %r37, %r2; sub.s32 %r39, %r1, %r38; setp.ne.s32 %p3, %r83, 0; mov.u32 %r40, -1; vote.sync.ballot.b32 %r41, %p3, %r40; sub.s32 %r42, %r2, %r39; shr.u32 %r43, %r40, %r42; and.b32 %r44, %r41, %r43; popc.b32 %r7, %r44; add.s32 %r45, %r2, -1; setp.ne.s32 %p5, %r39, %r45; shl.b32 %r46, %r37, 2; mov.u32 %r84, warp_offset; add.s32 %r8, %r84, %r46; @%p5 bra $L__BB216_4; add.s32 %r48, %r7, %r83; st.shared.u32 [%r8], %r48; $L__BB216_4: bar.sync 0; shl.b32 %r49, %r3, 2; add.s32 %r9, %r84, %r49; setp.ne.s32 %p6, %r1, 511; @%p6 bra $L__BB216_13; setp.gt.u32 %p7, %r2, 512; @%p7 bra $L__BB216_12; max.u32 %r10, %r3, 1; add.s32 %r53, %r10, -1; and.b32 %r92, %r10, 3; setp.lt.u32 %p8, %r53, 3; mov.u32 %r88, 0; mov.u32 %r89, %r88; @%p8 bra $L__BB216_9; sub.s32 %r87, %r10, %r92; mov.u32 %r88, 0; mov.u32 %r89, %r88; $L__BB216_8: ld.shared.v4.u32 {%r57, %r58, %r59, %r60}, [%r84]; add.s32 %r65, %r57, %r89; add.s32 %r66, %r58, %r65; add.s32 %r67, %r59, %r66; st.shared.v4.u32 [%r84], {%r89, %r65, %r66, %r67}; add.s32 %r89, %r60, %r67; add.s32 %r88, %r88, 4; add.s32 %r84, %r84, 16; add.s32 %r87, %r87, -4; setp.ne.s32 %p9, %r87, 0; @%p9 bra $L__BB216_8; $L__BB216_9: setp.eq.s32 %p10, %r92, 0; @%p10 bra $L__BB216_12; shl.b32 %r68, %r88, 2; mov.u32 %r69, warp_offset; add.s32 %r90, %r69, %r68; $L__BB216_11: .pragma "nounroll"; ld.shared.u32 %r70, [%r90]; st.shared.u32 [%r90], %r89; add.s32 %r89, %r70, %r89; add.s32 %r90, %r90, 4; add.s32 %r92, %r92, -1; setp.ne.s32 %p11, %r92, 0; @%p11 bra $L__BB216_11; $L__BB216_12: ld.shared.u32 %r71, [%r8]; add.s32 %r72, %r7, %r83; add.s32 %r73, %r72, %r71; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r74, [%rd7], %r73; st.shared.u32 [%r9], %r74; $L__BB216_13: bar.sync 0; setp.eq.s32 %p12, %r83, 0; @%p12 bra $L__BB216_15; ld.shared.u32 %r75, [%r9]; add.s32 %r76, %r75, %r7; ld.shared.u32 %r77, [%r8]; add.s32 %r78, %r76, %r77; cvta.to.global.u64 %rd8, %rd1; mul.wide.u32 %rd9, %r78, 4; add.s64 %rd10, %rd8, %rd9; st.global.u32 [%rd10], %r4; $L__BB216_15: ret; } // .globl kernel_gpu_integrator_queued_shadow_paths_array .visible .entry kernel_gpu_integrator_queued_shadow_paths_array( .param .u32 kernel_gpu_integrator_queued_shadow_paths_array_param_0, .param .u64 kernel_gpu_integrator_queued_shadow_paths_array_param_1, .param .u64 kernel_gpu_integrator_queued_shadow_paths_array_param_2, .param .u32 kernel_gpu_integrator_queued_shadow_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<13>; .reg .b32 %r<93>; .reg .b64 %rd<11>; ld.param.u32 %r32, [kernel_gpu_integrator_queued_shadow_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_queued_shadow_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_queued_shadow_paths_array_param_2]; ld.param.u32 %r30, [kernel_gpu_integrator_queued_shadow_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r32; mov.u32 %r83, 0; @%p1 bra $L__BB217_2; ld.const.u64 %rd3, [kernel_params+3416]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %r36, [%rd6]; setp.eq.s32 %p2, %r36, %r30; selp.u32 %r83, 1, 0, %p2; $L__BB217_2: div.u32 %r37, %r1, %r2; mul.lo.s32 %r38, %r37, %r2; sub.s32 %r39, %r1, %r38; setp.ne.s32 %p3, %r83, 0; mov.u32 %r40, -1; vote.sync.ballot.b32 %r41, %p3, %r40; sub.s32 %r42, %r2, %r39; shr.u32 %r43, %r40, %r42; and.b32 %r44, %r41, %r43; popc.b32 %r7, %r44; add.s32 %r45, %r2, -1; setp.ne.s32 %p5, %r39, %r45; shl.b32 %r46, %r37, 2; mov.u32 %r84, warp_offset; add.s32 %r8, %r84, %r46; @%p5 bra $L__BB217_4; add.s32 %r48, %r7, %r83; st.shared.u32 [%r8], %r48; $L__BB217_4: bar.sync 0; shl.b32 %r49, %r3, 2; add.s32 %r9, %r84, %r49; setp.ne.s32 %p6, %r1, 511; @%p6 bra $L__BB217_13; setp.gt.u32 %p7, %r2, 512; @%p7 bra $L__BB217_12; max.u32 %r10, %r3, 1; add.s32 %r53, %r10, -1; and.b32 %r92, %r10, 3; setp.lt.u32 %p8, %r53, 3; mov.u32 %r88, 0; mov.u32 %r89, %r88; @%p8 bra $L__BB217_9; sub.s32 %r87, %r10, %r92; mov.u32 %r88, 0; mov.u32 %r89, %r88; $L__BB217_8: ld.shared.v4.u32 {%r57, %r58, %r59, %r60}, [%r84]; add.s32 %r65, %r57, %r89; add.s32 %r66, %r58, %r65; add.s32 %r67, %r59, %r66; st.shared.v4.u32 [%r84], {%r89, %r65, %r66, %r67}; add.s32 %r89, %r60, %r67; add.s32 %r88, %r88, 4; add.s32 %r84, %r84, 16; add.s32 %r87, %r87, -4; setp.ne.s32 %p9, %r87, 0; @%p9 bra $L__BB217_8; $L__BB217_9: setp.eq.s32 %p10, %r92, 0; @%p10 bra $L__BB217_12; shl.b32 %r68, %r88, 2; mov.u32 %r69, warp_offset; add.s32 %r90, %r69, %r68; $L__BB217_11: .pragma "nounroll"; ld.shared.u32 %r70, [%r90]; st.shared.u32 [%r90], %r89; add.s32 %r89, %r70, %r89; add.s32 %r90, %r90, 4; add.s32 %r92, %r92, -1; setp.ne.s32 %p11, %r92, 0; @%p11 bra $L__BB217_11; $L__BB217_12: ld.shared.u32 %r71, [%r8]; add.s32 %r72, %r7, %r83; add.s32 %r73, %r72, %r71; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r74, [%rd7], %r73; st.shared.u32 [%r9], %r74; $L__BB217_13: bar.sync 0; setp.eq.s32 %p12, %r83, 0; @%p12 bra $L__BB217_15; ld.shared.u32 %r75, [%r9]; add.s32 %r76, %r75, %r7; ld.shared.u32 %r77, [%r8]; add.s32 %r78, %r76, %r77; cvta.to.global.u64 %rd8, %rd1; mul.wide.u32 %rd9, %r78, 4; add.s64 %rd10, %rd8, %rd9; st.global.u32 [%rd10], %r4; $L__BB217_15: ret; } // .globl kernel_gpu_integrator_active_paths_array .visible .entry kernel_gpu_integrator_active_paths_array( .param .u32 kernel_gpu_integrator_active_paths_array_param_0, .param .u64 kernel_gpu_integrator_active_paths_array_param_1, .param .u64 kernel_gpu_integrator_active_paths_array_param_2 ) .maxntid 512, 1, 1 { .reg .pred %p<13>; .reg .b16 %rs<2>; .reg .b32 %r<91>; .reg .b64 %rd<11>; ld.param.u32 %r31, [kernel_gpu_integrator_active_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_active_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_active_paths_array_param_2]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r32, 512; div.u32 %r3, %r32, %r2; mov.u32 %r33, %ctaid.x; shl.b32 %r34, %r33, 9; add.s32 %r4, %r34, %r1; setp.ge.u32 %p1, %r4, %r31; mov.u32 %r81, 0; @%p1 bra $L__BB218_2; ld.const.u64 %rd3, [kernel_params+2520]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.ne.s16 %p2, %rs1, 0; selp.u32 %r81, 1, 0, %p2; $L__BB218_2: div.u32 %r35, %r1, %r2; mul.lo.s32 %r36, %r35, %r2; sub.s32 %r37, %r1, %r36; setp.ne.s32 %p3, %r81, 0; mov.u32 %r38, -1; vote.sync.ballot.b32 %r39, %p3, %r38; sub.s32 %r40, %r2, %r37; shr.u32 %r41, %r38, %r40; and.b32 %r42, %r39, %r41; popc.b32 %r7, %r42; add.s32 %r43, %r2, -1; setp.ne.s32 %p5, %r37, %r43; shl.b32 %r44, %r35, 2; mov.u32 %r82, warp_offset; add.s32 %r8, %r82, %r44; @%p5 bra $L__BB218_4; add.s32 %r46, %r7, %r81; st.shared.u32 [%r8], %r46; $L__BB218_4: bar.sync 0; shl.b32 %r47, %r3, 2; add.s32 %r9, %r82, %r47; setp.ne.s32 %p6, %r1, 511; @%p6 bra $L__BB218_13; setp.gt.u32 %p7, %r2, 512; @%p7 bra $L__BB218_12; max.u32 %r10, %r3, 1; add.s32 %r51, %r10, -1; and.b32 %r90, %r10, 3; setp.lt.u32 %p8, %r51, 3; mov.u32 %r86, 0; mov.u32 %r87, %r86; @%p8 bra $L__BB218_9; sub.s32 %r85, %r10, %r90; mov.u32 %r86, 0; mov.u32 %r87, %r86; $L__BB218_8: ld.shared.v4.u32 {%r55, %r56, %r57, %r58}, [%r82]; add.s32 %r63, %r55, %r87; add.s32 %r64, %r56, %r63; add.s32 %r65, %r57, %r64; st.shared.v4.u32 [%r82], {%r87, %r63, %r64, %r65}; add.s32 %r87, %r58, %r65; add.s32 %r86, %r86, 4; add.s32 %r82, %r82, 16; add.s32 %r85, %r85, -4; setp.ne.s32 %p9, %r85, 0; @%p9 bra $L__BB218_8; $L__BB218_9: setp.eq.s32 %p10, %r90, 0; @%p10 bra $L__BB218_12; shl.b32 %r66, %r86, 2; mov.u32 %r67, warp_offset; add.s32 %r88, %r67, %r66; $L__BB218_11: .pragma "nounroll"; ld.shared.u32 %r68, [%r88]; st.shared.u32 [%r88], %r87; add.s32 %r87, %r68, %r87; add.s32 %r88, %r88, 4; add.s32 %r90, %r90, -1; setp.ne.s32 %p11, %r90, 0; @%p11 bra $L__BB218_11; $L__BB218_12: ld.shared.u32 %r69, [%r8]; add.s32 %r70, %r7, %r81; add.s32 %r71, %r70, %r69; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r72, [%rd7], %r71; st.shared.u32 [%r9], %r72; $L__BB218_13: bar.sync 0; setp.eq.s32 %p12, %r81, 0; @%p12 bra $L__BB218_15; ld.shared.u32 %r73, [%r9]; add.s32 %r74, %r73, %r7; ld.shared.u32 %r75, [%r8]; add.s32 %r76, %r74, %r75; cvta.to.global.u64 %rd8, %rd1; mul.wide.u32 %rd9, %r76, 4; add.s64 %rd10, %rd8, %rd9; st.global.u32 [%rd10], %r4; $L__BB218_15: ret; } // .globl kernel_gpu_integrator_terminated_paths_array .visible .entry kernel_gpu_integrator_terminated_paths_array( .param .u32 kernel_gpu_integrator_terminated_paths_array_param_0, .param .u64 kernel_gpu_integrator_terminated_paths_array_param_1, .param .u64 kernel_gpu_integrator_terminated_paths_array_param_2, .param .u32 kernel_gpu_integrator_terminated_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<13>; .reg .b16 %rs<2>; .reg .b32 %r<88>; .reg .b64 %rd<14>; ld.param.u32 %r32, [kernel_gpu_integrator_terminated_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_terminated_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_terminated_paths_array_param_2]; ld.param.u32 %r30, [kernel_gpu_integrator_terminated_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r32; mov.u32 %r78, 0; @%p1 bra $L__BB219_2; ld.const.u64 %rd3, [kernel_params+2520]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.eq.s16 %p2, %rs1, 0; selp.u32 %r78, 1, 0, %p2; $L__BB219_2: div.u32 %r36, %r1, %r2; mul.lo.s32 %r37, %r36, %r2; sub.s32 %r38, %r1, %r37; setp.ne.s32 %p3, %r78, 0; mov.u32 %r39, -1; vote.sync.ballot.b32 %r40, %p3, %r39; sub.s32 %r41, %r2, %r38; shr.u32 %r42, %r39, %r41; and.b32 %r43, %r40, %r42; popc.b32 %r7, %r43; add.s32 %r44, %r2, -1; setp.ne.s32 %p5, %r38, %r44; shl.b32 %r45, %r36, 2; mov.u32 %r79, warp_offset; add.s32 %r8, %r79, %r45; @%p5 bra $L__BB219_4; add.s32 %r47, %r7, %r78; st.shared.u32 [%r8], %r47; $L__BB219_4: bar.sync 0; shl.b32 %r48, %r3, 2; add.s32 %r9, %r79, %r48; setp.ne.s32 %p6, %r1, 511; @%p6 bra $L__BB219_13; setp.gt.u32 %p7, %r2, 512; @%p7 bra $L__BB219_12; max.u32 %r10, %r3, 1; add.s32 %r52, %r10, -1; and.b32 %r87, %r10, 3; setp.lt.u32 %p8, %r52, 3; mov.u32 %r83, 0; mov.u32 %r84, %r83; @%p8 bra $L__BB219_9; sub.s32 %r82, %r10, %r87; mov.u32 %r83, 0; mov.u32 %r84, %r83; $L__BB219_8: ld.shared.v4.u32 {%r56, %r57, %r58, %r59}, [%r79]; add.s32 %r64, %r56, %r84; add.s32 %r65, %r57, %r64; add.s32 %r66, %r58, %r65; st.shared.v4.u32 [%r79], {%r84, %r64, %r65, %r66}; add.s32 %r84, %r59, %r66; add.s32 %r83, %r83, 4; add.s32 %r79, %r79, 16; add.s32 %r82, %r82, -4; setp.ne.s32 %p9, %r82, 0; @%p9 bra $L__BB219_8; $L__BB219_9: setp.eq.s32 %p10, %r87, 0; @%p10 bra $L__BB219_12; shl.b32 %r67, %r83, 2; mov.u32 %r68, warp_offset; add.s32 %r85, %r68, %r67; $L__BB219_11: .pragma "nounroll"; ld.shared.u32 %r69, [%r85]; st.shared.u32 [%r85], %r84; add.s32 %r84, %r69, %r84; add.s32 %r85, %r85, 4; add.s32 %r87, %r87, -1; setp.ne.s32 %p11, %r87, 0; @%p11 bra $L__BB219_11; $L__BB219_12: ld.shared.u32 %r70, [%r8]; add.s32 %r71, %r7, %r78; add.s32 %r72, %r71, %r70; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r73, [%rd7], %r72; st.shared.u32 [%r9], %r73; $L__BB219_13: bar.sync 0; setp.eq.s32 %p12, %r78, 0; @%p12 bra $L__BB219_15; ld.shared.u32 %r74, [%r9]; add.s32 %r75, %r74, %r7; ld.shared.u32 %r76, [%r8]; add.s32 %r77, %r75, %r76; cvt.u64.u32 %rd8, %r77; cvt.s64.s32 %rd9, %r30; add.s64 %rd10, %rd8, %rd9; cvta.to.global.u64 %rd11, %rd1; shl.b64 %rd12, %rd10, 2; add.s64 %rd13, %rd11, %rd12; st.global.u32 [%rd13], %r4; $L__BB219_15: ret; } // .globl kernel_gpu_integrator_terminated_shadow_paths_array .visible .entry kernel_gpu_integrator_terminated_shadow_paths_array( .param .u32 kernel_gpu_integrator_terminated_shadow_paths_array_param_0, .param .u64 kernel_gpu_integrator_terminated_shadow_paths_array_param_1, .param .u64 kernel_gpu_integrator_terminated_shadow_paths_array_param_2, .param .u32 kernel_gpu_integrator_terminated_shadow_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<13>; .reg .b16 %rs<2>; .reg .b32 %r<88>; .reg .b64 %rd<14>; ld.param.u32 %r32, [kernel_gpu_integrator_terminated_shadow_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_terminated_shadow_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_terminated_shadow_paths_array_param_2]; ld.param.u32 %r30, [kernel_gpu_integrator_terminated_shadow_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r32; mov.u32 %r78, 0; @%p1 bra $L__BB220_2; ld.const.u64 %rd3, [kernel_params+3416]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.eq.s16 %p2, %rs1, 0; selp.u32 %r78, 1, 0, %p2; $L__BB220_2: div.u32 %r36, %r1, %r2; mul.lo.s32 %r37, %r36, %r2; sub.s32 %r38, %r1, %r37; setp.ne.s32 %p3, %r78, 0; mov.u32 %r39, -1; vote.sync.ballot.b32 %r40, %p3, %r39; sub.s32 %r41, %r2, %r38; shr.u32 %r42, %r39, %r41; and.b32 %r43, %r40, %r42; popc.b32 %r7, %r43; add.s32 %r44, %r2, -1; setp.ne.s32 %p5, %r38, %r44; shl.b32 %r45, %r36, 2; mov.u32 %r79, warp_offset; add.s32 %r8, %r79, %r45; @%p5 bra $L__BB220_4; add.s32 %r47, %r7, %r78; st.shared.u32 [%r8], %r47; $L__BB220_4: bar.sync 0; shl.b32 %r48, %r3, 2; add.s32 %r9, %r79, %r48; setp.ne.s32 %p6, %r1, 511; @%p6 bra $L__BB220_13; setp.gt.u32 %p7, %r2, 512; @%p7 bra $L__BB220_12; max.u32 %r10, %r3, 1; add.s32 %r52, %r10, -1; and.b32 %r87, %r10, 3; setp.lt.u32 %p8, %r52, 3; mov.u32 %r83, 0; mov.u32 %r84, %r83; @%p8 bra $L__BB220_9; sub.s32 %r82, %r10, %r87; mov.u32 %r83, 0; mov.u32 %r84, %r83; $L__BB220_8: ld.shared.v4.u32 {%r56, %r57, %r58, %r59}, [%r79]; add.s32 %r64, %r56, %r84; add.s32 %r65, %r57, %r64; add.s32 %r66, %r58, %r65; st.shared.v4.u32 [%r79], {%r84, %r64, %r65, %r66}; add.s32 %r84, %r59, %r66; add.s32 %r83, %r83, 4; add.s32 %r79, %r79, 16; add.s32 %r82, %r82, -4; setp.ne.s32 %p9, %r82, 0; @%p9 bra $L__BB220_8; $L__BB220_9: setp.eq.s32 %p10, %r87, 0; @%p10 bra $L__BB220_12; shl.b32 %r67, %r83, 2; mov.u32 %r68, warp_offset; add.s32 %r85, %r68, %r67; $L__BB220_11: .pragma "nounroll"; ld.shared.u32 %r69, [%r85]; st.shared.u32 [%r85], %r84; add.s32 %r84, %r69, %r84; add.s32 %r85, %r85, 4; add.s32 %r87, %r87, -1; setp.ne.s32 %p11, %r87, 0; @%p11 bra $L__BB220_11; $L__BB220_12: ld.shared.u32 %r70, [%r8]; add.s32 %r71, %r7, %r78; add.s32 %r72, %r71, %r70; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r73, [%rd7], %r72; st.shared.u32 [%r9], %r73; $L__BB220_13: bar.sync 0; setp.eq.s32 %p12, %r78, 0; @%p12 bra $L__BB220_15; ld.shared.u32 %r74, [%r9]; add.s32 %r75, %r74, %r7; ld.shared.u32 %r76, [%r8]; add.s32 %r77, %r75, %r76; cvt.u64.u32 %rd8, %r77; cvt.s64.s32 %rd9, %r30; add.s64 %rd10, %rd8, %rd9; cvta.to.global.u64 %rd11, %rd1; shl.b64 %rd12, %rd10, 2; add.s64 %rd13, %rd11, %rd12; st.global.u32 [%rd13], %r4; $L__BB220_15: ret; } // .globl kernel_gpu_integrator_sorted_paths_array .visible .entry kernel_gpu_integrator_sorted_paths_array( .param .u32 kernel_gpu_integrator_sorted_paths_array_param_0, .param .u32 kernel_gpu_integrator_sorted_paths_array_param_1, .param .u64 kernel_gpu_integrator_sorted_paths_array_param_2, .param .u64 kernel_gpu_integrator_sorted_paths_array_param_3, .param .u64 kernel_gpu_integrator_sorted_paths_array_param_4, .param .u64 kernel_gpu_integrator_sorted_paths_array_param_5, .param .u32 kernel_gpu_integrator_sorted_paths_array_param_6 ) .maxntid 512, 1, 1 { .reg .pred %p<5>; .reg .b32 %r<16>; .reg .b64 %rd<23>; ld.param.u32 %r8, [kernel_gpu_integrator_sorted_paths_array_param_0]; ld.param.u32 %r5, [kernel_gpu_integrator_sorted_paths_array_param_1]; ld.param.u64 %rd3, [kernel_gpu_integrator_sorted_paths_array_param_2]; ld.param.u64 %rd4, [kernel_gpu_integrator_sorted_paths_array_param_4]; ld.param.u64 %rd5, [kernel_gpu_integrator_sorted_paths_array_param_5]; ld.param.u32 %r6, [kernel_gpu_integrator_sorted_paths_array_param_6]; mov.u32 %r9, %ntid.x; mov.u32 %r10, %ctaid.x; mov.u32 %r11, %tid.x; mad.lo.s32 %r1, %r10, %r9, %r11; setp.ge.u32 %p1, %r1, %r8; mov.u32 %r15, -1; @%p1 bra $L__BB221_3; ld.const.u64 %rd6, [kernel_params+2520]; cvta.to.global.u64 %rd7, %rd6; cvt.s64.s32 %rd1, %r1; mul.wide.s32 %rd8, %r1, 2; add.s64 %rd9, %rd7, %rd8; ld.global.u16 %r13, [%rd9]; setp.ne.s32 %p2, %r13, %r6; @%p2 bra $L__BB221_3; ld.const.u64 %rd10, [kernel_params+2624]; cvta.to.global.u64 %rd11, %rd10; shl.b64 %rd12, %rd1, 2; add.s64 %rd13, %rd11, %rd12; ld.global.u32 %r15, [%rd13]; $L__BB221_3: setp.eq.s32 %p3, %r15, -1; @%p3 bra $L__BB221_7; cvt.s64.s32 %rd2, %r15; cvta.to.global.u64 %rd14, %rd5; mul.wide.s32 %rd15, %r15, 4; add.s64 %rd16, %rd14, %rd15; atom.global.add.u32 %r4, [%rd16], 1; setp.lt.u32 %p4, %r4, %r5; @%p4 bra $L__BB221_6; bra.uni $L__BB221_5; $L__BB221_6: cvta.to.global.u64 %rd20, %rd3; mul.wide.u32 %rd21, %r4, 4; add.s64 %rd22, %rd20, %rd21; st.global.u32 [%rd22], %r1; bra.uni $L__BB221_7; $L__BB221_5: cvta.to.global.u64 %rd17, %rd4; shl.b64 %rd18, %rd2, 2; add.s64 %rd19, %rd17, %rd18; atom.global.add.u32 %r14, [%rd19], 1; $L__BB221_7: ret; } // .globl kernel_gpu_integrator_compact_paths_array .visible .entry kernel_gpu_integrator_compact_paths_array( .param .u32 kernel_gpu_integrator_compact_paths_array_param_0, .param .u64 kernel_gpu_integrator_compact_paths_array_param_1, .param .u64 kernel_gpu_integrator_compact_paths_array_param_2, .param .u32 kernel_gpu_integrator_compact_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<15>; .reg .b16 %rs<2>; .reg .b32 %r<92>; .reg .b64 %rd<11>; ld.param.u32 %r31, [kernel_gpu_integrator_compact_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_compact_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_compact_paths_array_param_2]; ld.param.u32 %r32, [kernel_gpu_integrator_compact_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r31; setp.lt.s32 %p2, %r4, %r32; or.pred %p3, %p1, %p2; mov.u32 %r82, 0; @%p3 bra $L__BB222_2; ld.const.u64 %rd3, [kernel_params+2520]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.ne.s16 %p4, %rs1, 0; selp.u32 %r82, 1, 0, %p4; $L__BB222_2: div.u32 %r36, %r1, %r2; mul.lo.s32 %r37, %r36, %r2; sub.s32 %r38, %r1, %r37; setp.ne.s32 %p5, %r82, 0; mov.u32 %r39, -1; vote.sync.ballot.b32 %r40, %p5, %r39; sub.s32 %r41, %r2, %r38; shr.u32 %r42, %r39, %r41; and.b32 %r43, %r40, %r42; popc.b32 %r7, %r43; add.s32 %r44, %r2, -1; setp.ne.s32 %p7, %r38, %r44; shl.b32 %r45, %r36, 2; mov.u32 %r83, warp_offset; add.s32 %r8, %r83, %r45; @%p7 bra $L__BB222_4; add.s32 %r47, %r7, %r82; st.shared.u32 [%r8], %r47; $L__BB222_4: bar.sync 0; shl.b32 %r48, %r3, 2; add.s32 %r9, %r83, %r48; setp.ne.s32 %p8, %r1, 511; @%p8 bra $L__BB222_13; setp.gt.u32 %p9, %r2, 512; @%p9 bra $L__BB222_12; max.u32 %r10, %r3, 1; add.s32 %r52, %r10, -1; and.b32 %r91, %r10, 3; setp.lt.u32 %p10, %r52, 3; mov.u32 %r87, 0; mov.u32 %r88, %r87; @%p10 bra $L__BB222_9; sub.s32 %r86, %r10, %r91; mov.u32 %r87, 0; mov.u32 %r88, %r87; $L__BB222_8: ld.shared.v4.u32 {%r56, %r57, %r58, %r59}, [%r83]; add.s32 %r64, %r56, %r88; add.s32 %r65, %r57, %r64; add.s32 %r66, %r58, %r65; st.shared.v4.u32 [%r83], {%r88, %r64, %r65, %r66}; add.s32 %r88, %r59, %r66; add.s32 %r87, %r87, 4; add.s32 %r83, %r83, 16; add.s32 %r86, %r86, -4; setp.ne.s32 %p11, %r86, 0; @%p11 bra $L__BB222_8; $L__BB222_9: setp.eq.s32 %p12, %r91, 0; @%p12 bra $L__BB222_12; shl.b32 %r67, %r87, 2; mov.u32 %r68, warp_offset; add.s32 %r89, %r68, %r67; $L__BB222_11: .pragma "nounroll"; ld.shared.u32 %r69, [%r89]; st.shared.u32 [%r89], %r88; add.s32 %r88, %r69, %r88; add.s32 %r89, %r89, 4; add.s32 %r91, %r91, -1; setp.ne.s32 %p13, %r91, 0; @%p13 bra $L__BB222_11; $L__BB222_12: ld.shared.u32 %r70, [%r8]; add.s32 %r71, %r7, %r82; add.s32 %r72, %r71, %r70; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r73, [%rd7], %r72; st.shared.u32 [%r9], %r73; $L__BB222_13: bar.sync 0; setp.eq.s32 %p14, %r82, 0; @%p14 bra $L__BB222_15; ld.shared.u32 %r74, [%r9]; add.s32 %r75, %r74, %r7; ld.shared.u32 %r76, [%r8]; add.s32 %r77, %r75, %r76; cvta.to.global.u64 %rd8, %rd1; mul.wide.u32 %rd9, %r77, 4; add.s64 %rd10, %rd8, %rd9; st.global.u32 [%rd10], %r4; $L__BB222_15: ret; } // .globl kernel_gpu_integrator_compact_states .visible .entry kernel_gpu_integrator_compact_states( .param .u64 kernel_gpu_integrator_compact_states_param_0, .param .u32 kernel_gpu_integrator_compact_states_param_1, .param .u32 kernel_gpu_integrator_compact_states_param_2, .param .u32 kernel_gpu_integrator_compact_states_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<64>; .reg .b16 %rs<14>; .reg .f32 %f<46>; .reg .b32 %r<84>; .reg .b64 %rd<369>; ld.param.u64 %rd70, [kernel_gpu_integrator_compact_states_param_0]; ld.param.u32 %r12, [kernel_gpu_integrator_compact_states_param_1]; ld.param.u32 %r13, [kernel_gpu_integrator_compact_states_param_2]; ld.param.u32 %r14, [kernel_gpu_integrator_compact_states_param_3]; mov.u32 %r15, %ntid.x; mov.u32 %r16, %ctaid.x; mov.u32 %r17, %tid.x; mad.lo.s32 %r1, %r16, %r15, %r17; setp.ge.s32 %p1, %r1, %r14; @%p1 bra $L__BB223_124; cvta.to.global.u64 %rd71, %rd70; add.s32 %r18, %r1, %r12; mul.wide.s32 %rd72, %r18, 4; add.s64 %rd73, %rd71, %rd72; ld.global.s32 %rd1, [%rd73]; add.s32 %r19, %r1, %r13; mul.wide.s32 %rd74, %r19, 4; add.s64 %rd75, %rd71, %rd74; ld.global.s32 %rd2, [%rd75]; ld.const.u64 %rd3, [kernel_params+2448]; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB223_3; cvta.to.global.u64 %rd76, %rd3; shl.b64 %rd77, %rd1, 2; add.s64 %rd78, %rd76, %rd77; ld.global.u32 %r20, [%rd78]; shl.b64 %rd79, %rd2, 2; add.s64 %rd80, %rd76, %rd79; st.global.u32 [%rd80], %r20; $L__BB223_3: ld.const.u64 %rd4, [kernel_params+2456]; setp.eq.s64 %p3, %rd4, 0; @%p3 bra $L__BB223_5; cvta.to.global.u64 %rd81, %rd4; shl.b64 %rd82, %rd1, 2; add.s64 %rd83, %rd81, %rd82; ld.global.u32 %r21, [%rd83]; shl.b64 %rd84, %rd2, 2; add.s64 %rd85, %rd81, %rd84; st.global.u32 [%rd85], %r21; $L__BB223_5: ld.const.u64 %rd5, [kernel_params+2464]; setp.eq.s64 %p4, %rd5, 0; @%p4 bra $L__BB223_7; cvta.to.global.u64 %rd86, %rd5; shl.b64 %rd87, %rd1, 1; add.s64 %rd88, %rd86, %rd87; ld.global.u16 %rs1, [%rd88]; shl.b64 %rd89, %rd2, 1; add.s64 %rd90, %rd86, %rd89; st.global.u16 [%rd90], %rs1; $L__BB223_7: ld.const.u64 %rd6, [kernel_params+2472]; setp.eq.s64 %p5, %rd6, 0; @%p5 bra $L__BB223_9; cvta.to.global.u64 %rd91, %rd6; shl.b64 %rd92, %rd1, 1; add.s64 %rd93, %rd91, %rd92; ld.global.u16 %rs2, [%rd93]; shl.b64 %rd94, %rd2, 1; add.s64 %rd95, %rd91, %rd94; st.global.u16 [%rd95], %rs2; $L__BB223_9: ld.const.u64 %rd7, [kernel_params+2480]; setp.eq.s64 %p6, %rd7, 0; @%p6 bra $L__BB223_11; cvta.to.global.u64 %rd96, %rd7; shl.b64 %rd97, %rd1, 1; add.s64 %rd98, %rd96, %rd97; ld.global.u16 %rs3, [%rd98]; shl.b64 %rd99, %rd2, 1; add.s64 %rd100, %rd96, %rd99; st.global.u16 [%rd100], %rs3; $L__BB223_11: ld.const.u64 %rd8, [kernel_params+2488]; setp.eq.s64 %p7, %rd8, 0; @%p7 bra $L__BB223_13; cvta.to.global.u64 %rd101, %rd8; shl.b64 %rd102, %rd1, 1; add.s64 %rd103, %rd101, %rd102; ld.global.u16 %rs4, [%rd103]; shl.b64 %rd104, %rd2, 1; add.s64 %rd105, %rd101, %rd104; st.global.u16 [%rd105], %rs4; $L__BB223_13: ld.const.u64 %rd9, [kernel_params+2496]; setp.eq.s64 %p8, %rd9, 0; @%p8 bra $L__BB223_15; cvta.to.global.u64 %rd106, %rd9; shl.b64 %rd107, %rd1, 1; add.s64 %rd108, %rd106, %rd107; ld.global.u16 %rs5, [%rd108]; shl.b64 %rd109, %rd2, 1; add.s64 %rd110, %rd106, %rd109; st.global.u16 [%rd110], %rs5; $L__BB223_15: ld.const.u64 %rd10, [kernel_params+2504]; setp.eq.s64 %p9, %rd10, 0; @%p9 bra $L__BB223_17; cvta.to.global.u64 %rd111, %rd10; shl.b64 %rd112, %rd1, 1; add.s64 %rd113, %rd111, %rd112; ld.global.u16 %rs6, [%rd113]; shl.b64 %rd114, %rd2, 1; add.s64 %rd115, %rd111, %rd114; st.global.u16 [%rd115], %rs6; $L__BB223_17: ld.const.u64 %rd11, [kernel_params+2512]; setp.eq.s64 %p10, %rd11, 0; @%p10 bra $L__BB223_19; cvta.to.global.u64 %rd116, %rd11; shl.b64 %rd117, %rd1, 1; add.s64 %rd118, %rd116, %rd117; ld.global.u16 %rs7, [%rd118]; shl.b64 %rd119, %rd2, 1; add.s64 %rd120, %rd116, %rd119; st.global.u16 [%rd120], %rs7; $L__BB223_19: ld.const.u64 %rd12, [kernel_params+2520]; setp.eq.s64 %p11, %rd12, 0; @%p11 bra $L__BB223_21; cvta.to.global.u64 %rd121, %rd12; shl.b64 %rd122, %rd1, 1; add.s64 %rd123, %rd121, %rd122; ld.global.u16 %rs8, [%rd123]; shl.b64 %rd124, %rd2, 1; add.s64 %rd125, %rd121, %rd124; st.global.u16 [%rd125], %rs8; $L__BB223_21: ld.const.u64 %rd13, [kernel_params+2528]; setp.eq.s64 %p12, %rd13, 0; @%p12 bra $L__BB223_23; cvta.to.global.u64 %rd126, %rd13; shl.b64 %rd127, %rd1, 2; add.s64 %rd128, %rd126, %rd127; ld.global.u32 %r22, [%rd128]; shl.b64 %rd129, %rd2, 2; add.s64 %rd130, %rd126, %rd129; st.global.u32 [%rd130], %r22; $L__BB223_23: ld.const.u64 %rd14, [kernel_params+2536]; setp.eq.s64 %p13, %rd14, 0; @%p13 bra $L__BB223_25; cvta.to.global.u64 %rd131, %rd14; shl.b64 %rd132, %rd1, 1; add.s64 %rd133, %rd131, %rd132; ld.global.u16 %rs9, [%rd133]; shl.b64 %rd134, %rd2, 1; add.s64 %rd135, %rd131, %rd134; st.global.u16 [%rd135], %rs9; $L__BB223_25: ld.const.u64 %rd15, [kernel_params+2544]; setp.eq.s64 %p14, %rd15, 0; @%p14 bra $L__BB223_27; cvta.to.global.u64 %rd136, %rd15; shl.b64 %rd137, %rd1, 2; add.s64 %rd138, %rd136, %rd137; ld.global.u32 %r23, [%rd138]; shl.b64 %rd139, %rd2, 2; add.s64 %rd140, %rd136, %rd139; st.global.u32 [%rd140], %r23; $L__BB223_27: ld.const.u64 %rd16, [kernel_params+2552]; setp.eq.s64 %p15, %rd16, 0; @%p15 bra $L__BB223_29; cvta.to.global.u64 %rd141, %rd16; add.s64 %rd142, %rd141, %rd1; ld.global.u8 %rs10, [%rd142]; add.s64 %rd143, %rd141, %rd2; st.global.u8 [%rd143], %rs10; $L__BB223_29: ld.const.u64 %rd17, [kernel_params+2560]; setp.eq.s64 %p16, %rd17, 0; @%p16 bra $L__BB223_31; cvta.to.global.u64 %rd144, %rd17; shl.b64 %rd145, %rd1, 2; add.s64 %rd146, %rd144, %rd145; ld.global.f32 %f1, [%rd146]; shl.b64 %rd147, %rd2, 2; add.s64 %rd148, %rd144, %rd147; st.global.f32 [%rd148], %f1; $L__BB223_31: ld.const.u64 %rd18, [kernel_params+2568]; setp.eq.s64 %p17, %rd18, 0; @%p17 bra $L__BB223_33; cvta.to.global.u64 %rd149, %rd18; shl.b64 %rd150, %rd1, 2; add.s64 %rd151, %rd149, %rd150; ld.global.f32 %f2, [%rd151]; shl.b64 %rd152, %rd2, 2; add.s64 %rd153, %rd149, %rd152; st.global.f32 [%rd153], %f2; $L__BB223_33: ld.const.u64 %rd19, [kernel_params+2576]; setp.eq.s64 %p18, %rd19, 0; @%p18 bra $L__BB223_35; cvta.to.global.u64 %rd154, %rd19; shl.b64 %rd155, %rd1, 2; add.s64 %rd156, %rd154, %rd155; ld.global.f32 %f3, [%rd156]; shl.b64 %rd157, %rd2, 2; add.s64 %rd158, %rd154, %rd157; st.global.f32 [%rd158], %f3; $L__BB223_35: ld.const.u64 %rd20, [kernel_params+2584]; setp.eq.s64 %p19, %rd20, 0; @%p19 bra $L__BB223_37; cvta.to.global.u64 %rd159, %rd20; mul.lo.s64 %rd160, %rd1, 12; add.s64 %rd161, %rd159, %rd160; ld.global.f32 %f4, [%rd161]; ld.global.f32 %f5, [%rd161+4]; ld.global.f32 %f6, [%rd161+8]; mul.lo.s64 %rd162, %rd2, 12; add.s64 %rd163, %rd159, %rd162; st.global.f32 [%rd163], %f4; st.global.f32 [%rd163+4], %f5; st.global.f32 [%rd163+8], %f6; $L__BB223_37: ld.const.u64 %rd21, [kernel_params+2592]; setp.eq.s64 %p20, %rd21, 0; @%p20 bra $L__BB223_39; cvta.to.global.u64 %rd164, %rd21; shl.b64 %rd165, %rd1, 2; add.s64 %rd166, %rd164, %rd165; ld.global.f32 %f7, [%rd166]; shl.b64 %rd167, %rd2, 2; add.s64 %rd168, %rd164, %rd167; st.global.f32 [%rd168], %f7; $L__BB223_39: ld.const.u64 %rd22, [kernel_params+2600]; setp.eq.s64 %p21, %rd22, 0; @%p21 bra $L__BB223_41; cvta.to.global.u64 %rd169, %rd22; mul.lo.s64 %rd170, %rd1, 12; add.s64 %rd171, %rd169, %rd170; ld.global.f32 %f8, [%rd171]; ld.global.f32 %f9, [%rd171+4]; ld.global.f32 %f10, [%rd171+8]; mul.lo.s64 %rd172, %rd2, 12; add.s64 %rd173, %rd169, %rd172; st.global.f32 [%rd173], %f8; st.global.f32 [%rd173+4], %f9; st.global.f32 [%rd173+8], %f10; $L__BB223_41: ld.const.u64 %rd23, [kernel_params+2608]; setp.eq.s64 %p22, %rd23, 0; @%p22 bra $L__BB223_43; cvt.u32.u64 %r24, %rd2; cvt.u32.u64 %r25, %rd1; cvta.to.global.u64 %rd174, %rd23; mul.wide.s32 %rd175, %r25, 12; add.s64 %rd176, %rd174, %rd175; ld.global.f32 %f11, [%rd176]; ld.global.f32 %f12, [%rd176+4]; ld.global.f32 %f13, [%rd176+8]; mul.wide.s32 %rd177, %r24, 12; add.s64 %rd178, %rd174, %rd177; st.global.f32 [%rd178], %f11; st.global.f32 [%rd178+4], %f12; st.global.f32 [%rd178+8], %f13; $L__BB223_43: ld.const.u64 %rd24, [kernel_params+2616]; setp.eq.s64 %p23, %rd24, 0; @%p23 bra $L__BB223_45; cvt.u32.u64 %r26, %rd1; cvt.u32.u64 %r27, %rd2; cvta.to.global.u64 %rd179, %rd24; mul.wide.s32 %rd180, %r26, 12; add.s64 %rd181, %rd179, %rd180; ld.global.f32 %f14, [%rd181]; ld.global.f32 %f15, [%rd181+4]; ld.global.f32 %f16, [%rd181+8]; mul.wide.s32 %rd182, %r27, 12; add.s64 %rd183, %rd179, %rd182; st.global.f32 [%rd183], %f14; st.global.f32 [%rd183+4], %f15; st.global.f32 [%rd183+8], %f16; $L__BB223_45: ld.const.u64 %rd25, [kernel_params+2624]; setp.eq.s64 %p24, %rd25, 0; @%p24 bra $L__BB223_47; cvt.u32.u64 %r28, %rd2; cvt.u32.u64 %r29, %rd1; cvta.to.global.u64 %rd184, %rd25; mul.wide.s32 %rd185, %r29, 4; add.s64 %rd186, %rd184, %rd185; ld.global.u32 %r30, [%rd186]; mul.wide.s32 %rd187, %r28, 4; add.s64 %rd188, %rd184, %rd187; st.global.u32 [%rd188], %r30; $L__BB223_47: ld.const.u64 %rd26, [kernel_params+2632]; setp.eq.s64 %p25, %rd26, 0; @%p25 bra $L__BB223_49; cvt.u32.u64 %r31, %rd2; cvt.u32.u64 %r32, %rd1; cvta.to.global.u64 %rd189, %rd26; mul.wide.s32 %rd190, %r32, 12; add.s64 %rd191, %rd189, %rd190; ld.global.f32 %f17, [%rd191]; ld.global.f32 %f18, [%rd191+4]; ld.global.f32 %f19, [%rd191+8]; mul.wide.s32 %rd192, %r31, 12; add.s64 %rd193, %rd189, %rd192; st.global.f32 [%rd193], %f17; st.global.f32 [%rd193+4], %f18; st.global.f32 [%rd193+8], %f19; $L__BB223_49: ld.const.u64 %rd27, [kernel_params+2640]; setp.eq.s64 %p26, %rd27, 0; @%p26 bra $L__BB223_51; cvt.u32.u64 %r33, %rd2; cvt.u32.u64 %r34, %rd1; cvta.to.global.u64 %rd194, %rd27; mul.wide.s32 %rd195, %r34, 12; add.s64 %rd196, %rd194, %rd195; ld.global.f32 %f20, [%rd196]; ld.global.f32 %f21, [%rd196+4]; ld.global.f32 %f22, [%rd196+8]; mul.wide.s32 %rd197, %r33, 12; add.s64 %rd198, %rd194, %rd197; st.global.f32 [%rd198], %f20; st.global.f32 [%rd198+4], %f21; st.global.f32 [%rd198+8], %f22; $L__BB223_51: ld.const.u64 %rd28, [kernel_params+2648]; setp.eq.s64 %p27, %rd28, 0; @%p27 bra $L__BB223_53; cvt.u32.u64 %r35, %rd2; cvt.u32.u64 %r36, %rd1; cvta.to.global.u64 %rd199, %rd28; mul.wide.s32 %rd200, %r36, 4; add.s64 %rd201, %rd199, %rd200; ld.global.f32 %f23, [%rd201]; mul.wide.s32 %rd202, %r35, 4; add.s64 %rd203, %rd199, %rd202; st.global.f32 [%rd203], %f23; $L__BB223_53: ld.const.u64 %rd29, [kernel_params+2656]; setp.eq.s64 %p28, %rd29, 0; @%p28 bra $L__BB223_55; cvt.u32.u64 %r37, %rd2; cvt.u32.u64 %r38, %rd1; cvta.to.global.u64 %rd204, %rd29; mul.wide.s32 %rd205, %r38, 4; add.s64 %rd206, %rd204, %rd205; ld.global.f32 %f24, [%rd206]; mul.wide.s32 %rd207, %r37, 4; add.s64 %rd208, %rd204, %rd207; st.global.f32 [%rd208], %f24; $L__BB223_55: ld.const.u64 %rd30, [kernel_params+2664]; setp.eq.s64 %p29, %rd30, 0; @%p29 bra $L__BB223_57; cvt.u32.u64 %r39, %rd2; cvt.u32.u64 %r40, %rd1; cvta.to.global.u64 %rd209, %rd30; mul.wide.s32 %rd210, %r40, 4; add.s64 %rd211, %rd209, %rd210; ld.global.f32 %f25, [%rd211]; mul.wide.s32 %rd212, %r39, 4; add.s64 %rd213, %rd209, %rd212; st.global.f32 [%rd213], %f25; $L__BB223_57: ld.const.u64 %rd31, [kernel_params+2672]; setp.eq.s64 %p30, %rd31, 0; @%p30 bra $L__BB223_59; cvt.u32.u64 %r41, %rd2; cvt.u32.u64 %r42, %rd1; cvta.to.global.u64 %rd214, %rd31; mul.wide.s32 %rd215, %r42, 4; add.s64 %rd216, %rd214, %rd215; ld.global.f32 %f26, [%rd216]; mul.wide.s32 %rd217, %r41, 4; add.s64 %rd218, %rd214, %rd217; st.global.f32 [%rd218], %f26; $L__BB223_59: ld.const.u64 %rd32, [kernel_params+2680]; setp.eq.s64 %p31, %rd32, 0; @%p31 bra $L__BB223_61; cvt.u32.u64 %r43, %rd2; cvt.u32.u64 %r44, %rd1; cvta.to.global.u64 %rd219, %rd32; mul.wide.s32 %rd220, %r44, 4; add.s64 %rd221, %rd219, %rd220; ld.global.f32 %f27, [%rd221]; mul.wide.s32 %rd222, %r43, 4; add.s64 %rd223, %rd219, %rd222; st.global.f32 [%rd223], %f27; $L__BB223_61: ld.const.u64 %rd33, [kernel_params+2688]; setp.eq.s64 %p32, %rd33, 0; @%p32 bra $L__BB223_63; cvt.u32.u64 %r45, %rd2; cvt.u32.u64 %r46, %rd1; cvta.to.global.u64 %rd224, %rd33; mul.wide.s32 %rd225, %r46, 4; add.s64 %rd226, %rd224, %rd225; ld.global.f32 %f28, [%rd226]; mul.wide.s32 %rd227, %r45, 4; add.s64 %rd228, %rd224, %rd227; st.global.f32 [%rd228], %f28; $L__BB223_63: ld.const.u64 %rd34, [kernel_params+2696]; setp.eq.s64 %p33, %rd34, 0; @%p33 bra $L__BB223_65; cvt.u32.u64 %r47, %rd2; cvt.u32.u64 %r48, %rd1; cvta.to.global.u64 %rd229, %rd34; mul.wide.s32 %rd230, %r48, 4; add.s64 %rd231, %rd229, %rd230; ld.global.f32 %f29, [%rd231]; mul.wide.s32 %rd232, %r47, 4; add.s64 %rd233, %rd229, %rd232; st.global.f32 [%rd233], %f29; $L__BB223_65: ld.const.u64 %rd35, [kernel_params+2704]; setp.eq.s64 %p34, %rd35, 0; @%p34 bra $L__BB223_67; cvt.u32.u64 %r49, %rd2; cvt.u32.u64 %r50, %rd1; cvta.to.global.u64 %rd234, %rd35; mul.wide.s32 %rd235, %r50, 4; add.s64 %rd236, %rd234, %rd235; ld.global.f32 %f30, [%rd236]; mul.wide.s32 %rd237, %r49, 4; add.s64 %rd238, %rd234, %rd237; st.global.f32 [%rd238], %f30; $L__BB223_67: ld.const.u64 %rd36, [kernel_params+2712]; setp.eq.s64 %p35, %rd36, 0; @%p35 bra $L__BB223_69; cvt.u32.u64 %r51, %rd2; cvt.u32.u64 %r52, %rd1; cvta.to.global.u64 %rd239, %rd36; mul.wide.s32 %rd240, %r52, 4; add.s64 %rd241, %rd239, %rd240; ld.global.u32 %r53, [%rd241]; mul.wide.s32 %rd242, %r51, 4; add.s64 %rd243, %rd239, %rd242; st.global.u32 [%rd243], %r53; $L__BB223_69: ld.const.u64 %rd37, [kernel_params+2720]; setp.eq.s64 %p36, %rd37, 0; @%p36 bra $L__BB223_71; cvt.u32.u64 %r54, %rd2; cvt.u32.u64 %r55, %rd1; cvta.to.global.u64 %rd244, %rd37; mul.wide.s32 %rd245, %r55, 4; add.s64 %rd246, %rd244, %rd245; ld.global.u32 %r56, [%rd246]; mul.wide.s32 %rd247, %r54, 4; add.s64 %rd248, %rd244, %rd247; st.global.u32 [%rd248], %r56; $L__BB223_71: ld.const.u64 %rd38, [kernel_params+2728]; setp.eq.s64 %p37, %rd38, 0; @%p37 bra $L__BB223_73; cvt.u32.u64 %r57, %rd2; cvt.u32.u64 %r58, %rd1; cvta.to.global.u64 %rd249, %rd38; mul.wide.s32 %rd250, %r58, 4; add.s64 %rd251, %rd249, %rd250; ld.global.u32 %r59, [%rd251]; mul.wide.s32 %rd252, %r57, 4; add.s64 %rd253, %rd249, %rd252; st.global.u32 [%rd253], %r59; $L__BB223_73: ld.const.u64 %rd39, [kernel_params+2736]; setp.eq.s64 %p38, %rd39, 0; @%p38 bra $L__BB223_75; cvt.u32.u64 %r60, %rd2; cvt.u32.u64 %r61, %rd1; cvta.to.global.u64 %rd254, %rd39; mul.wide.s32 %rd255, %r61, 12; add.s64 %rd256, %rd254, %rd255; ld.global.f32 %f31, [%rd256]; ld.global.f32 %f32, [%rd256+4]; ld.global.f32 %f33, [%rd256+8]; mul.wide.s32 %rd257, %r60, 12; add.s64 %rd258, %rd254, %rd257; st.global.f32 [%rd258], %f31; st.global.f32 [%rd258+4], %f32; st.global.f32 [%rd258+8], %f33; $L__BB223_75: ld.const.u64 %rd40, [kernel_params+2744]; setp.eq.s64 %p39, %rd40, 0; @%p39 bra $L__BB223_77; cvta.to.global.u64 %rd259, %rd40; cvt.u32.u64 %r62, %rd1; mul.wide.s32 %rd260, %r62, 12; add.s64 %rd261, %rd259, %rd260; ld.global.f32 %f34, [%rd261]; ld.global.f32 %f35, [%rd261+4]; ld.global.f32 %f36, [%rd261+8]; cvt.u32.u64 %r63, %rd2; mul.wide.s32 %rd262, %r63, 12; add.s64 %rd263, %rd259, %rd262; st.global.f32 [%rd263], %f34; st.global.f32 [%rd263+4], %f35; st.global.f32 [%rd263+8], %f36; $L__BB223_77: ld.const.u64 %rd43, [kernel_params+2752]; setp.eq.s64 %p40, %rd43, 0; @%p40 bra $L__BB223_79; cvta.to.global.u64 %rd264, %rd43; shl.b64 %rd265, %rd1, 2; add.s64 %rd266, %rd264, %rd265; ld.global.f32 %f37, [%rd266]; shl.b64 %rd267, %rd2, 2; add.s64 %rd268, %rd264, %rd267; st.global.f32 [%rd268], %f37; $L__BB223_79: ld.const.u64 %rd44, [kernel_params+2760]; setp.eq.s64 %p41, %rd44, 0; @%p41 bra $L__BB223_81; cvta.to.global.u64 %rd269, %rd44; mul.lo.s64 %rd270, %rd1, 12; add.s64 %rd271, %rd269, %rd270; ld.global.f32 %f38, [%rd271]; ld.global.f32 %f39, [%rd271+4]; ld.global.f32 %f40, [%rd271+8]; mul.lo.s64 %rd272, %rd2, 12; add.s64 %rd273, %rd269, %rd272; st.global.f32 [%rd273], %f38; st.global.f32 [%rd273+4], %f39; st.global.f32 [%rd273+8], %f40; $L__BB223_81: ld.const.u32 %r65, [kernel_params+12]; max.u32 %r2, %r65, 1; add.s32 %r66, %r2, -1; and.b32 %r83, %r2, 3; setp.lt.u32 %p42, %r66, 3; mov.u32 %r82, 0; @%p42 bra $L__BB223_100; sub.s32 %r81, %r2, %r83; mov.u32 %r82, 0; mov.u64 %rd367, kernel_params; $L__BB223_83: ld.const.u64 %rd47, [%rd367+2768]; setp.eq.s64 %p43, %rd47, 0; @%p43 bra $L__BB223_85; cvta.to.global.u64 %rd275, %rd47; shl.b64 %rd276, %rd1, 2; add.s64 %rd277, %rd275, %rd276; ld.global.u32 %r68, [%rd277]; shl.b64 %rd278, %rd2, 2; add.s64 %rd279, %rd275, %rd278; st.global.u32 [%rd279], %r68; $L__BB223_85: ld.const.u64 %rd48, [%rd367+2776]; setp.eq.s64 %p44, %rd48, 0; @%p44 bra $L__BB223_87; cvta.to.global.u64 %rd280, %rd48; shl.b64 %rd281, %rd1, 2; add.s64 %rd282, %rd280, %rd281; ld.global.u32 %r69, [%rd282]; shl.b64 %rd283, %rd2, 2; add.s64 %rd284, %rd280, %rd283; st.global.u32 [%rd284], %r69; $L__BB223_87: ld.const.u64 %rd49, [%rd367+2784]; setp.eq.s64 %p45, %rd49, 0; @%p45 bra $L__BB223_89; cvta.to.global.u64 %rd285, %rd49; shl.b64 %rd286, %rd1, 2; add.s64 %rd287, %rd285, %rd286; ld.global.u32 %r70, [%rd287]; shl.b64 %rd288, %rd2, 2; add.s64 %rd289, %rd285, %rd288; st.global.u32 [%rd289], %r70; $L__BB223_89: ld.const.u64 %rd50, [%rd367+2792]; setp.eq.s64 %p46, %rd50, 0; @%p46 bra $L__BB223_91; cvta.to.global.u64 %rd290, %rd50; shl.b64 %rd291, %rd1, 2; add.s64 %rd292, %rd290, %rd291; ld.global.u32 %r71, [%rd292]; shl.b64 %rd293, %rd2, 2; add.s64 %rd294, %rd290, %rd293; st.global.u32 [%rd294], %r71; $L__BB223_91: ld.const.u64 %rd51, [%rd367+2800]; setp.eq.s64 %p47, %rd51, 0; @%p47 bra $L__BB223_93; cvta.to.global.u64 %rd295, %rd51; shl.b64 %rd296, %rd1, 2; add.s64 %rd297, %rd295, %rd296; ld.global.u32 %r72, [%rd297]; shl.b64 %rd298, %rd2, 2; add.s64 %rd299, %rd295, %rd298; st.global.u32 [%rd299], %r72; $L__BB223_93: ld.const.u64 %rd52, [%rd367+2808]; setp.eq.s64 %p48, %rd52, 0; @%p48 bra $L__BB223_95; cvta.to.global.u64 %rd300, %rd52; shl.b64 %rd301, %rd1, 2; add.s64 %rd302, %rd300, %rd301; ld.global.u32 %r73, [%rd302]; shl.b64 %rd303, %rd2, 2; add.s64 %rd304, %rd300, %rd303; st.global.u32 [%rd304], %r73; $L__BB223_95: ld.const.u64 %rd53, [%rd367+2816]; setp.eq.s64 %p49, %rd53, 0; @%p49 bra $L__BB223_97; cvta.to.global.u64 %rd305, %rd53; shl.b64 %rd306, %rd1, 2; add.s64 %rd307, %rd305, %rd306; ld.global.u32 %r74, [%rd307]; shl.b64 %rd308, %rd2, 2; add.s64 %rd309, %rd305, %rd308; st.global.u32 [%rd309], %r74; $L__BB223_97: ld.const.u64 %rd54, [%rd367+2824]; setp.eq.s64 %p50, %rd54, 0; @%p50 bra $L__BB223_99; cvta.to.global.u64 %rd310, %rd54; shl.b64 %rd311, %rd1, 2; add.s64 %rd312, %rd310, %rd311; ld.global.u32 %r75, [%rd312]; shl.b64 %rd313, %rd2, 2; add.s64 %rd314, %rd310, %rd313; st.global.u32 [%rd314], %r75; $L__BB223_99: add.s32 %r82, %r82, 4; add.s64 %rd367, %rd367, 64; add.s32 %r81, %r81, -4; setp.ne.s32 %p51, %r81, 0; @%p51 bra $L__BB223_83; $L__BB223_100: setp.eq.s32 %p52, %r83, 0; @%p52 bra $L__BB223_107; shl.b32 %r76, %r82, 4; add.s32 %r77, %r76, 2776; cvt.s64.s32 %rd315, %r77; mov.u64 %rd316, kernel_params; add.s64 %rd368, %rd316, %rd315; $L__BB223_102: .pragma "nounroll"; add.s64 %rd58, %rd368, -8; ld.const.u64 %rd59, [%rd368+-8]; setp.eq.s64 %p53, %rd59, 0; @%p53 bra $L__BB223_104; cvta.to.global.u64 %rd317, %rd59; shl.b64 %rd318, %rd1, 2; add.s64 %rd319, %rd317, %rd318; ld.global.u32 %r78, [%rd319]; shl.b64 %rd320, %rd2, 2; add.s64 %rd321, %rd317, %rd320; st.global.u32 [%rd321], %r78; $L__BB223_104: ld.const.u64 %rd60, [%rd58+8]; setp.eq.s64 %p54, %rd60, 0; @%p54 bra $L__BB223_106; cvta.to.global.u64 %rd322, %rd60; shl.b64 %rd323, %rd1, 2; add.s64 %rd324, %rd322, %rd323; ld.global.u32 %r79, [%rd324]; shl.b64 %rd325, %rd2, 2; add.s64 %rd326, %rd322, %rd325; st.global.u32 [%rd326], %r79; $L__BB223_106: add.s32 %r83, %r83, -1; add.s64 %rd368, %rd368, 16; setp.ne.s32 %p55, %r83, 0; @%p55 bra $L__BB223_102; $L__BB223_107: ld.const.u64 %rd62, [kernel_params+3280]; setp.eq.s64 %p56, %rd62, 0; @%p56 bra $L__BB223_109; cvta.to.global.u64 %rd327, %rd62; shl.b64 %rd328, %rd1, 3; add.s64 %rd329, %rd327, %rd328; ld.global.u64 %rd330, [%rd329]; shl.b64 %rd331, %rd2, 3; add.s64 %rd332, %rd327, %rd331; st.global.u64 [%rd332], %rd330; $L__BB223_109: ld.const.u64 %rd63, [kernel_params+3288]; setp.eq.s64 %p57, %rd63, 0; @%p57 bra $L__BB223_111; cvta.to.global.u64 %rd333, %rd63; add.s64 %rd334, %rd333, %rd1; ld.global.u8 %rs11, [%rd334]; add.s64 %rd335, %rd333, %rd2; st.global.u8 [%rd335], %rs11; $L__BB223_111: ld.const.u64 %rd64, [kernel_params+3296]; setp.eq.s64 %p58, %rd64, 0; @%p58 bra $L__BB223_113; cvta.to.global.u64 %rd336, %rd64; shl.b64 %rd337, %rd1, 2; add.s64 %rd338, %rd336, %rd337; ld.global.f32 %f41, [%rd338]; shl.b64 %rd339, %rd2, 2; add.s64 %rd340, %rd336, %rd339; st.global.f32 [%rd340], %f41; $L__BB223_113: ld.const.u64 %rd65, [kernel_params+3304]; setp.eq.s64 %p59, %rd65, 0; @%p59 bra $L__BB223_115; cvta.to.global.u64 %rd341, %rd65; shl.b64 %rd342, %rd1, 2; add.s64 %rd343, %rd341, %rd342; ld.global.f32 %f42, [%rd343]; shl.b64 %rd344, %rd2, 2; add.s64 %rd345, %rd341, %rd344; st.global.f32 [%rd345], %f42; $L__BB223_115: ld.const.u64 %rd66, [kernel_params+3312]; setp.eq.s64 %p60, %rd66, 0; @%p60 bra $L__BB223_117; cvta.to.global.u64 %rd346, %rd66; shl.b64 %rd347, %rd1, 2; add.s64 %rd348, %rd346, %rd347; ld.global.f32 %f43, [%rd348]; shl.b64 %rd349, %rd2, 2; add.s64 %rd350, %rd346, %rd349; st.global.f32 [%rd350], %f43; $L__BB223_117: ld.const.u64 %rd67, [kernel_params+3320]; setp.eq.s64 %p61, %rd67, 0; @%p61 bra $L__BB223_119; cvta.to.global.u64 %rd351, %rd67; add.s64 %rd352, %rd351, %rd1; ld.global.u8 %rs12, [%rd352]; add.s64 %rd353, %rd351, %rd2; st.global.u8 [%rd353], %rs12; $L__BB223_119: ld.const.u64 %rd68, [kernel_params+3328]; setp.eq.s64 %p62, %rd68, 0; @%p62 bra $L__BB223_121; cvta.to.global.u64 %rd354, %rd68; shl.b64 %rd355, %rd1, 2; add.s64 %rd356, %rd354, %rd355; ld.global.f32 %f44, [%rd356]; shl.b64 %rd357, %rd2, 2; add.s64 %rd358, %rd354, %rd357; st.global.f32 [%rd358], %f44; $L__BB223_121: ld.const.u64 %rd69, [kernel_params+3336]; setp.eq.s64 %p63, %rd69, 0; @%p63 bra $L__BB223_123; cvta.to.global.u64 %rd359, %rd69; shl.b64 %rd360, %rd1, 2; add.s64 %rd361, %rd359, %rd360; ld.global.f32 %f45, [%rd361]; shl.b64 %rd362, %rd2, 2; add.s64 %rd363, %rd359, %rd362; st.global.f32 [%rd363], %f45; $L__BB223_123: cvta.to.global.u64 %rd364, %rd12; shl.b64 %rd365, %rd1, 1; add.s64 %rd366, %rd364, %rd365; mov.u16 %rs13, 0; st.global.u16 [%rd366], %rs13; $L__BB223_124: ret; } // .globl kernel_gpu_integrator_compact_shadow_paths_array .visible .entry kernel_gpu_integrator_compact_shadow_paths_array( .param .u32 kernel_gpu_integrator_compact_shadow_paths_array_param_0, .param .u64 kernel_gpu_integrator_compact_shadow_paths_array_param_1, .param .u64 kernel_gpu_integrator_compact_shadow_paths_array_param_2, .param .u32 kernel_gpu_integrator_compact_shadow_paths_array_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<15>; .reg .b16 %rs<2>; .reg .b32 %r<92>; .reg .b64 %rd<11>; ld.param.u32 %r31, [kernel_gpu_integrator_compact_shadow_paths_array_param_0]; ld.param.u64 %rd1, [kernel_gpu_integrator_compact_shadow_paths_array_param_1]; ld.param.u64 %rd2, [kernel_gpu_integrator_compact_shadow_paths_array_param_2]; ld.param.u32 %r32, [kernel_gpu_integrator_compact_shadow_paths_array_param_3]; mov.u32 %r1, %tid.x; mov.u32 %r2, WARP_SZ; mov.u32 %r33, 512; div.u32 %r3, %r33, %r2; mov.u32 %r34, %ctaid.x; shl.b32 %r35, %r34, 9; add.s32 %r4, %r35, %r1; setp.ge.u32 %p1, %r4, %r31; setp.lt.s32 %p2, %r4, %r32; or.pred %p3, %p1, %p2; mov.u32 %r82, 0; @%p3 bra $L__BB224_2; ld.const.u64 %rd3, [kernel_params+3416]; cvta.to.global.u64 %rd4, %rd3; mul.wide.s32 %rd5, %r4, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.ne.s16 %p4, %rs1, 0; selp.u32 %r82, 1, 0, %p4; $L__BB224_2: div.u32 %r36, %r1, %r2; mul.lo.s32 %r37, %r36, %r2; sub.s32 %r38, %r1, %r37; setp.ne.s32 %p5, %r82, 0; mov.u32 %r39, -1; vote.sync.ballot.b32 %r40, %p5, %r39; sub.s32 %r41, %r2, %r38; shr.u32 %r42, %r39, %r41; and.b32 %r43, %r40, %r42; popc.b32 %r7, %r43; add.s32 %r44, %r2, -1; setp.ne.s32 %p7, %r38, %r44; shl.b32 %r45, %r36, 2; mov.u32 %r83, warp_offset; add.s32 %r8, %r83, %r45; @%p7 bra $L__BB224_4; add.s32 %r47, %r7, %r82; st.shared.u32 [%r8], %r47; $L__BB224_4: bar.sync 0; shl.b32 %r48, %r3, 2; add.s32 %r9, %r83, %r48; setp.ne.s32 %p8, %r1, 511; @%p8 bra $L__BB224_13; setp.gt.u32 %p9, %r2, 512; @%p9 bra $L__BB224_12; max.u32 %r10, %r3, 1; add.s32 %r52, %r10, -1; and.b32 %r91, %r10, 3; setp.lt.u32 %p10, %r52, 3; mov.u32 %r87, 0; mov.u32 %r88, %r87; @%p10 bra $L__BB224_9; sub.s32 %r86, %r10, %r91; mov.u32 %r87, 0; mov.u32 %r88, %r87; $L__BB224_8: ld.shared.v4.u32 {%r56, %r57, %r58, %r59}, [%r83]; add.s32 %r64, %r56, %r88; add.s32 %r65, %r57, %r64; add.s32 %r66, %r58, %r65; st.shared.v4.u32 [%r83], {%r88, %r64, %r65, %r66}; add.s32 %r88, %r59, %r66; add.s32 %r87, %r87, 4; add.s32 %r83, %r83, 16; add.s32 %r86, %r86, -4; setp.ne.s32 %p11, %r86, 0; @%p11 bra $L__BB224_8; $L__BB224_9: setp.eq.s32 %p12, %r91, 0; @%p12 bra $L__BB224_12; shl.b32 %r67, %r87, 2; mov.u32 %r68, warp_offset; add.s32 %r89, %r68, %r67; $L__BB224_11: .pragma "nounroll"; ld.shared.u32 %r69, [%r89]; st.shared.u32 [%r89], %r88; add.s32 %r88, %r69, %r88; add.s32 %r89, %r89, 4; add.s32 %r91, %r91, -1; setp.ne.s32 %p13, %r91, 0; @%p13 bra $L__BB224_11; $L__BB224_12: ld.shared.u32 %r70, [%r8]; add.s32 %r71, %r7, %r82; add.s32 %r72, %r71, %r70; cvta.to.global.u64 %rd7, %rd2; atom.global.add.u32 %r73, [%rd7], %r72; st.shared.u32 [%r9], %r73; $L__BB224_13: bar.sync 0; setp.eq.s32 %p14, %r82, 0; @%p14 bra $L__BB224_15; ld.shared.u32 %r74, [%r9]; add.s32 %r75, %r74, %r7; ld.shared.u32 %r76, [%r8]; add.s32 %r77, %r75, %r76; cvta.to.global.u64 %rd8, %rd1; mul.wide.u32 %rd9, %r77, 4; add.s64 %rd10, %rd8, %rd9; st.global.u32 [%rd10], %r4; $L__BB224_15: ret; } // .globl kernel_gpu_integrator_compact_shadow_states .visible .entry kernel_gpu_integrator_compact_shadow_states( .param .u64 kernel_gpu_integrator_compact_shadow_states_param_0, .param .u32 kernel_gpu_integrator_compact_shadow_states_param_1, .param .u32 kernel_gpu_integrator_compact_shadow_states_param_2, .param .u32 kernel_gpu_integrator_compact_shadow_states_param_3 ) .maxntid 512, 1, 1 { .reg .pred %p<66>; .reg .b16 %rs<11>; .reg .f32 %f<38>; .reg .b32 %r<79>; .reg .b64 %rd<385>; ld.param.u64 %rd72, [kernel_gpu_integrator_compact_shadow_states_param_0]; ld.param.u32 %r12, [kernel_gpu_integrator_compact_shadow_states_param_1]; ld.param.u32 %r13, [kernel_gpu_integrator_compact_shadow_states_param_2]; ld.param.u32 %r14, [kernel_gpu_integrator_compact_shadow_states_param_3]; mov.u32 %r15, %ntid.x; mov.u32 %r16, %ctaid.x; mov.u32 %r17, %tid.x; mad.lo.s32 %r1, %r16, %r15, %r17; setp.ge.s32 %p1, %r1, %r14; @%p1 bra $L__BB225_128; cvta.to.global.u64 %rd73, %rd72; add.s32 %r18, %r1, %r12; mul.wide.s32 %rd74, %r18, 4; add.s64 %rd75, %rd73, %rd74; ld.global.s32 %rd1, [%rd75]; add.s32 %r19, %r1, %r13; mul.wide.s32 %rd76, %r19, 4; add.s64 %rd77, %rd73, %rd76; ld.global.s32 %rd2, [%rd77]; ld.const.u64 %rd3, [kernel_params+3344]; setp.eq.s64 %p2, %rd3, 0; @%p2 bra $L__BB225_3; cvta.to.global.u64 %rd78, %rd3; shl.b64 %rd79, %rd1, 2; add.s64 %rd80, %rd78, %rd79; ld.global.u32 %r20, [%rd80]; shl.b64 %rd81, %rd2, 2; add.s64 %rd82, %rd78, %rd81; st.global.u32 [%rd82], %r20; $L__BB225_3: ld.const.u64 %rd4, [kernel_params+3352]; setp.eq.s64 %p3, %rd4, 0; @%p3 bra $L__BB225_5; cvta.to.global.u64 %rd83, %rd4; shl.b64 %rd84, %rd1, 2; add.s64 %rd85, %rd83, %rd84; ld.global.u32 %r21, [%rd85]; shl.b64 %rd86, %rd2, 2; add.s64 %rd87, %rd83, %rd86; st.global.u32 [%rd87], %r21; $L__BB225_5: ld.const.u64 %rd5, [kernel_params+3360]; setp.eq.s64 %p4, %rd5, 0; @%p4 bra $L__BB225_7; cvta.to.global.u64 %rd88, %rd5; shl.b64 %rd89, %rd1, 2; add.s64 %rd90, %rd88, %rd89; ld.global.u32 %r22, [%rd90]; shl.b64 %rd91, %rd2, 2; add.s64 %rd92, %rd88, %rd91; st.global.u32 [%rd92], %r22; $L__BB225_7: ld.const.u64 %rd6, [kernel_params+3368]; setp.eq.s64 %p5, %rd6, 0; @%p5 bra $L__BB225_9; cvta.to.global.u64 %rd93, %rd6; shl.b64 %rd94, %rd1, 1; add.s64 %rd95, %rd93, %rd94; ld.global.u16 %rs1, [%rd95]; shl.b64 %rd96, %rd2, 1; add.s64 %rd97, %rd93, %rd96; st.global.u16 [%rd97], %rs1; $L__BB225_9: ld.const.u64 %rd7, [kernel_params+3376]; setp.eq.s64 %p6, %rd7, 0; @%p6 bra $L__BB225_11; cvta.to.global.u64 %rd98, %rd7; shl.b64 %rd99, %rd1, 1; add.s64 %rd100, %rd98, %rd99; ld.global.u16 %rs2, [%rd100]; shl.b64 %rd101, %rd2, 1; add.s64 %rd102, %rd98, %rd101; st.global.u16 [%rd102], %rs2; $L__BB225_11: ld.const.u64 %rd8, [kernel_params+3384]; setp.eq.s64 %p7, %rd8, 0; @%p7 bra $L__BB225_13; cvta.to.global.u64 %rd103, %rd8; shl.b64 %rd104, %rd1, 1; add.s64 %rd105, %rd103, %rd104; ld.global.u16 %rs3, [%rd105]; shl.b64 %rd106, %rd2, 1; add.s64 %rd107, %rd103, %rd106; st.global.u16 [%rd107], %rs3; $L__BB225_13: ld.const.u64 %rd9, [kernel_params+3392]; setp.eq.s64 %p8, %rd9, 0; @%p8 bra $L__BB225_15; cvta.to.global.u64 %rd108, %rd9; shl.b64 %rd109, %rd1, 1; add.s64 %rd110, %rd108, %rd109; ld.global.u16 %rs4, [%rd110]; shl.b64 %rd111, %rd2, 1; add.s64 %rd112, %rd108, %rd111; st.global.u16 [%rd112], %rs4; $L__BB225_15: ld.const.u64 %rd10, [kernel_params+3400]; setp.eq.s64 %p9, %rd10, 0; @%p9 bra $L__BB225_17; cvta.to.global.u64 %rd113, %rd10; shl.b64 %rd114, %rd1, 1; add.s64 %rd115, %rd113, %rd114; ld.global.u16 %rs5, [%rd115]; shl.b64 %rd116, %rd2, 1; add.s64 %rd117, %rd113, %rd116; st.global.u16 [%rd117], %rs5; $L__BB225_17: ld.const.u64 %rd11, [kernel_params+3408]; setp.eq.s64 %p10, %rd11, 0; @%p10 bra $L__BB225_19; cvta.to.global.u64 %rd118, %rd11; shl.b64 %rd119, %rd1, 1; add.s64 %rd120, %rd118, %rd119; ld.global.u16 %rs6, [%rd120]; shl.b64 %rd121, %rd2, 1; add.s64 %rd122, %rd118, %rd121; st.global.u16 [%rd122], %rs6; $L__BB225_19: ld.const.u64 %rd12, [kernel_params+3416]; setp.eq.s64 %p11, %rd12, 0; @%p11 bra $L__BB225_21; cvta.to.global.u64 %rd123, %rd12; shl.b64 %rd124, %rd1, 1; add.s64 %rd125, %rd123, %rd124; ld.global.u16 %rs7, [%rd125]; shl.b64 %rd126, %rd2, 1; add.s64 %rd127, %rd123, %rd126; st.global.u16 [%rd127], %rs7; $L__BB225_21: ld.const.u64 %rd13, [kernel_params+3424]; setp.eq.s64 %p12, %rd13, 0; @%p12 bra $L__BB225_23; cvta.to.global.u64 %rd128, %rd13; shl.b64 %rd129, %rd1, 2; add.s64 %rd130, %rd128, %rd129; ld.global.u32 %r23, [%rd130]; shl.b64 %rd131, %rd2, 2; add.s64 %rd132, %rd128, %rd131; st.global.u32 [%rd132], %r23; $L__BB225_23: ld.const.u64 %rd14, [kernel_params+3432]; setp.eq.s64 %p13, %rd14, 0; @%p13 bra $L__BB225_25; cvta.to.global.u64 %rd133, %rd14; mul.lo.s64 %rd134, %rd1, 12; add.s64 %rd135, %rd133, %rd134; ld.global.f32 %f1, [%rd135]; ld.global.f32 %f2, [%rd135+4]; ld.global.f32 %f3, [%rd135+8]; mul.lo.s64 %rd136, %rd2, 12; add.s64 %rd137, %rd133, %rd136; st.global.f32 [%rd137], %f1; st.global.f32 [%rd137+4], %f2; st.global.f32 [%rd137+8], %f3; $L__BB225_25: ld.const.u64 %rd15, [kernel_params+3440]; setp.eq.s64 %p14, %rd15, 0; @%p14 bra $L__BB225_27; cvta.to.global.u64 %rd138, %rd15; mul.lo.s64 %rd139, %rd1, 12; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f4, [%rd140]; ld.global.f32 %f5, [%rd140+4]; ld.global.f32 %f6, [%rd140+8]; mul.lo.s64 %rd141, %rd2, 12; add.s64 %rd142, %rd138, %rd141; st.global.f32 [%rd142], %f4; st.global.f32 [%rd142+4], %f5; st.global.f32 [%rd142+8], %f6; $L__BB225_27: ld.const.u64 %rd16, [kernel_params+3448]; setp.eq.s64 %p15, %rd16, 0; @%p15 bra $L__BB225_29; cvt.u32.u64 %r24, %rd1; cvt.u32.u64 %r25, %rd2; cvta.to.global.u64 %rd143, %rd16; mul.wide.s32 %rd144, %r24, 12; add.s64 %rd145, %rd143, %rd144; ld.global.f32 %f7, [%rd145]; ld.global.f32 %f8, [%rd145+4]; ld.global.f32 %f9, [%rd145+8]; mul.wide.s32 %rd146, %r25, 12; add.s64 %rd147, %rd143, %rd146; st.global.f32 [%rd147], %f7; st.global.f32 [%rd147+4], %f8; st.global.f32 [%rd147+8], %f9; $L__BB225_29: ld.const.u64 %rd17, [kernel_params+3456]; setp.eq.s64 %p16, %rd17, 0; @%p16 bra $L__BB225_31; cvt.u32.u64 %r26, %rd1; cvt.u32.u64 %r27, %rd2; cvta.to.global.u64 %rd148, %rd17; mul.wide.s32 %rd149, %r26, 12; add.s64 %rd150, %rd148, %rd149; ld.global.f32 %f10, [%rd150]; ld.global.f32 %f11, [%rd150+4]; ld.global.f32 %f12, [%rd150+8]; mul.wide.s32 %rd151, %r27, 12; add.s64 %rd152, %rd148, %rd151; st.global.f32 [%rd152], %f10; st.global.f32 [%rd152+4], %f11; st.global.f32 [%rd152+8], %f12; $L__BB225_31: ld.const.u64 %rd18, [kernel_params+3464]; setp.eq.s64 %p17, %rd18, 0; @%p17 bra $L__BB225_33; cvt.u32.u64 %r28, %rd2; cvt.u32.u64 %r29, %rd1; cvta.to.global.u64 %rd153, %rd18; mul.wide.s32 %rd154, %r29, 2; add.s64 %rd155, %rd153, %rd154; ld.global.u16 %rs8, [%rd155]; mul.wide.s32 %rd156, %r28, 2; add.s64 %rd157, %rd153, %rd156; st.global.u16 [%rd157], %rs8; $L__BB225_33: ld.const.u64 %rd19, [kernel_params+3472]; setp.eq.s64 %p18, %rd19, 0; @%p18 bra $L__BB225_35; cvta.to.global.u64 %rd158, %rd19; add.s64 %rd159, %rd158, %rd1; ld.global.u8 %rs9, [%rd159]; add.s64 %rd160, %rd158, %rd2; st.global.u8 [%rd160], %rs9; $L__BB225_35: ld.const.u64 %rd20, [kernel_params+3480]; setp.eq.s64 %p19, %rd20, 0; @%p19 bra $L__BB225_37; cvt.u32.u64 %r30, %rd2; cvt.u32.u64 %r31, %rd1; cvta.to.global.u64 %rd161, %rd20; mul.wide.s32 %rd162, %r31, 12; add.s64 %rd163, %rd161, %rd162; ld.global.f32 %f13, [%rd163]; ld.global.f32 %f14, [%rd163+4]; ld.global.f32 %f15, [%rd163+8]; mul.wide.s32 %rd164, %r30, 12; add.s64 %rd165, %rd161, %rd164; st.global.f32 [%rd165], %f13; st.global.f32 [%rd165+4], %f14; st.global.f32 [%rd165+8], %f15; $L__BB225_37: ld.const.u64 %rd21, [kernel_params+3488]; setp.eq.s64 %p20, %rd21, 0; @%p20 bra $L__BB225_39; cvt.u32.u64 %r32, %rd2; cvt.u32.u64 %r33, %rd1; cvta.to.global.u64 %rd166, %rd21; mul.wide.s32 %rd167, %r33, 8; add.s64 %rd168, %rd166, %rd167; ld.global.u64 %rd169, [%rd168]; mul.wide.s32 %rd170, %r32, 8; add.s64 %rd171, %rd166, %rd170; st.global.u64 [%rd171], %rd169; $L__BB225_39: ld.const.u64 %rd22, [kernel_params+3496]; setp.eq.s64 %p21, %rd22, 0; @%p21 bra $L__BB225_41; cvt.u32.u64 %r34, %rd2; cvt.u32.u64 %r35, %rd1; cvta.to.global.u64 %rd172, %rd22; mul.wide.s32 %rd173, %r35, 12; add.s64 %rd174, %rd172, %rd173; ld.global.f32 %f16, [%rd174]; ld.global.f32 %f17, [%rd174+4]; ld.global.f32 %f18, [%rd174+8]; mul.wide.s32 %rd175, %r34, 12; add.s64 %rd176, %rd172, %rd175; st.global.f32 [%rd176], %f16; st.global.f32 [%rd176+4], %f17; st.global.f32 [%rd176+8], %f18; $L__BB225_41: ld.const.u64 %rd23, [kernel_params+3504]; setp.eq.s64 %p22, %rd23, 0; @%p22 bra $L__BB225_43; cvt.u32.u64 %r36, %rd2; cvt.u32.u64 %r37, %rd1; cvta.to.global.u64 %rd177, %rd23; mul.wide.s32 %rd178, %r37, 12; add.s64 %rd179, %rd177, %rd178; ld.global.f32 %f19, [%rd179]; ld.global.f32 %f20, [%rd179+4]; ld.global.f32 %f21, [%rd179+8]; mul.wide.s32 %rd180, %r36, 12; add.s64 %rd181, %rd177, %rd180; st.global.f32 [%rd181], %f19; st.global.f32 [%rd181+4], %f20; st.global.f32 [%rd181+8], %f21; $L__BB225_43: ld.const.u64 %rd24, [kernel_params+3512]; setp.eq.s64 %p23, %rd24, 0; @%p23 bra $L__BB225_45; cvt.u32.u64 %r38, %rd2; cvt.u32.u64 %r39, %rd1; cvta.to.global.u64 %rd182, %rd24; mul.wide.s32 %rd183, %r39, 4; add.s64 %rd184, %rd182, %rd183; ld.global.f32 %f22, [%rd184]; mul.wide.s32 %rd185, %r38, 4; add.s64 %rd186, %rd182, %rd185; st.global.f32 [%rd186], %f22; $L__BB225_45: ld.const.u64 %rd25, [kernel_params+3520]; setp.eq.s64 %p24, %rd25, 0; @%p24 bra $L__BB225_47; cvt.u32.u64 %r40, %rd2; cvt.u32.u64 %r41, %rd1; cvta.to.global.u64 %rd187, %rd25; mul.wide.s32 %rd188, %r41, 4; add.s64 %rd189, %rd187, %rd188; ld.global.f32 %f23, [%rd189]; mul.wide.s32 %rd190, %r40, 4; add.s64 %rd191, %rd187, %rd190; st.global.f32 [%rd191], %f23; $L__BB225_47: ld.const.u64 %rd26, [kernel_params+3528]; setp.eq.s64 %p25, %rd26, 0; @%p25 bra $L__BB225_49; cvt.u32.u64 %r42, %rd2; cvt.u32.u64 %r43, %rd1; cvta.to.global.u64 %rd192, %rd26; mul.wide.s32 %rd193, %r43, 4; add.s64 %rd194, %rd192, %rd193; ld.global.f32 %f24, [%rd194]; mul.wide.s32 %rd195, %r42, 4; add.s64 %rd196, %rd192, %rd195; st.global.f32 [%rd196], %f24; $L__BB225_49: ld.const.u64 %rd27, [kernel_params+3536]; setp.eq.s64 %p26, %rd27, 0; @%p26 bra $L__BB225_51; cvta.to.global.u64 %rd197, %rd27; cvt.u32.u64 %r44, %rd1; mul.wide.s32 %rd198, %r44, 4; add.s64 %rd199, %rd197, %rd198; ld.global.f32 %f25, [%rd199]; cvt.u32.u64 %r45, %rd2; mul.wide.s32 %rd200, %r45, 4; add.s64 %rd201, %rd197, %rd200; st.global.f32 [%rd201], %f25; $L__BB225_51: ld.const.u64 %rd30, [kernel_params+3544]; setp.eq.s64 %p27, %rd30, 0; @%p27 bra $L__BB225_53; cvta.to.global.u64 %rd202, %rd30; shl.b64 %rd203, %rd1, 2; add.s64 %rd204, %rd202, %rd203; ld.global.u32 %r46, [%rd204]; shl.b64 %rd205, %rd2, 2; add.s64 %rd206, %rd202, %rd205; st.global.u32 [%rd206], %r46; $L__BB225_53: ld.const.u64 %rd31, [kernel_params+3552]; setp.eq.s64 %p28, %rd31, 0; @%p28 bra $L__BB225_55; cvta.to.global.u64 %rd207, %rd31; shl.b64 %rd208, %rd1, 2; add.s64 %rd209, %rd207, %rd208; ld.global.f32 %f26, [%rd209]; shl.b64 %rd210, %rd2, 2; add.s64 %rd211, %rd207, %rd210; st.global.f32 [%rd211], %f26; $L__BB225_55: ld.const.u64 %rd32, [kernel_params+3560]; setp.eq.s64 %p29, %rd32, 0; @%p29 bra $L__BB225_57; cvta.to.global.u64 %rd212, %rd32; shl.b64 %rd213, %rd1, 2; add.s64 %rd214, %rd212, %rd213; ld.global.f32 %f27, [%rd214]; shl.b64 %rd215, %rd2, 2; add.s64 %rd216, %rd212, %rd215; st.global.f32 [%rd216], %f27; $L__BB225_57: ld.const.u64 %rd33, [kernel_params+3568]; setp.eq.s64 %p30, %rd33, 0; @%p30 bra $L__BB225_59; cvta.to.global.u64 %rd217, %rd33; shl.b64 %rd218, %rd1, 2; add.s64 %rd219, %rd217, %rd218; ld.global.f32 %f28, [%rd219]; shl.b64 %rd220, %rd2, 2; add.s64 %rd221, %rd217, %rd220; st.global.f32 [%rd221], %f28; $L__BB225_59: ld.const.u64 %rd34, [kernel_params+3576]; setp.eq.s64 %p31, %rd34, 0; @%p31 bra $L__BB225_61; cvta.to.global.u64 %rd222, %rd34; shl.b64 %rd223, %rd1, 2; add.s64 %rd224, %rd222, %rd223; ld.global.u32 %r47, [%rd224]; shl.b64 %rd225, %rd2, 2; add.s64 %rd226, %rd222, %rd225; st.global.u32 [%rd226], %r47; $L__BB225_61: ld.const.u64 %rd35, [kernel_params+3584]; setp.eq.s64 %p32, %rd35, 0; @%p32 bra $L__BB225_63; cvta.to.global.u64 %rd227, %rd35; shl.b64 %rd228, %rd1, 2; add.s64 %rd229, %rd227, %rd228; ld.global.u32 %r48, [%rd229]; shl.b64 %rd230, %rd2, 2; add.s64 %rd231, %rd227, %rd230; st.global.u32 [%rd231], %r48; $L__BB225_63: ld.const.u64 %rd36, [kernel_params+3592]; setp.eq.s64 %p33, %rd36, 0; @%p33 bra $L__BB225_65; cvta.to.global.u64 %rd232, %rd36; shl.b64 %rd233, %rd1, 2; add.s64 %rd234, %rd232, %rd233; ld.global.u32 %r49, [%rd234]; shl.b64 %rd235, %rd2, 2; add.s64 %rd236, %rd232, %rd235; st.global.u32 [%rd236], %r49; $L__BB225_65: ld.const.u64 %rd37, [kernel_params+3600]; setp.eq.s64 %p34, %rd37, 0; @%p34 bra $L__BB225_67; cvta.to.global.u64 %rd237, %rd37; shl.b64 %rd238, %rd1, 2; add.s64 %rd239, %rd237, %rd238; ld.global.f32 %f29, [%rd239]; shl.b64 %rd240, %rd2, 2; add.s64 %rd241, %rd237, %rd240; st.global.f32 [%rd241], %f29; $L__BB225_67: ld.const.u64 %rd38, [kernel_params+3608]; setp.eq.s64 %p35, %rd38, 0; @%p35 bra $L__BB225_69; cvta.to.global.u64 %rd242, %rd38; shl.b64 %rd243, %rd1, 2; add.s64 %rd244, %rd242, %rd243; ld.global.f32 %f30, [%rd244]; shl.b64 %rd245, %rd2, 2; add.s64 %rd246, %rd242, %rd245; st.global.f32 [%rd246], %f30; $L__BB225_69: ld.const.u64 %rd39, [kernel_params+3616]; setp.eq.s64 %p36, %rd39, 0; @%p36 bra $L__BB225_71; cvta.to.global.u64 %rd247, %rd39; shl.b64 %rd248, %rd1, 2; add.s64 %rd249, %rd247, %rd248; ld.global.f32 %f31, [%rd249]; shl.b64 %rd250, %rd2, 2; add.s64 %rd251, %rd247, %rd250; st.global.f32 [%rd251], %f31; $L__BB225_71: ld.const.u64 %rd40, [kernel_params+3624]; setp.eq.s64 %p37, %rd40, 0; @%p37 bra $L__BB225_73; cvta.to.global.u64 %rd252, %rd40; shl.b64 %rd253, %rd1, 2; add.s64 %rd254, %rd252, %rd253; ld.global.u32 %r50, [%rd254]; shl.b64 %rd255, %rd2, 2; add.s64 %rd256, %rd252, %rd255; st.global.u32 [%rd256], %r50; $L__BB225_73: ld.const.u64 %rd41, [kernel_params+3632]; setp.eq.s64 %p38, %rd41, 0; @%p38 bra $L__BB225_75; cvta.to.global.u64 %rd257, %rd41; shl.b64 %rd258, %rd1, 2; add.s64 %rd259, %rd257, %rd258; ld.global.u32 %r51, [%rd259]; shl.b64 %rd260, %rd2, 2; add.s64 %rd261, %rd257, %rd260; st.global.u32 [%rd261], %r51; $L__BB225_75: ld.const.u64 %rd42, [kernel_params+3640]; setp.eq.s64 %p39, %rd42, 0; @%p39 bra $L__BB225_77; cvta.to.global.u64 %rd262, %rd42; shl.b64 %rd263, %rd1, 2; add.s64 %rd264, %rd262, %rd263; ld.global.u32 %r52, [%rd264]; shl.b64 %rd265, %rd2, 2; add.s64 %rd266, %rd262, %rd265; st.global.u32 [%rd266], %r52; $L__BB225_77: ld.const.u64 %rd43, [kernel_params+3648]; setp.eq.s64 %p40, %rd43, 0; @%p40 bra $L__BB225_79; cvta.to.global.u64 %rd267, %rd43; shl.b64 %rd268, %rd1, 2; add.s64 %rd269, %rd267, %rd268; ld.global.f32 %f32, [%rd269]; shl.b64 %rd270, %rd2, 2; add.s64 %rd271, %rd267, %rd270; st.global.f32 [%rd271], %f32; $L__BB225_79: ld.const.u64 %rd44, [kernel_params+3656]; setp.eq.s64 %p41, %rd44, 0; @%p41 bra $L__BB225_81; cvta.to.global.u64 %rd272, %rd44; shl.b64 %rd273, %rd1, 2; add.s64 %rd274, %rd272, %rd273; ld.global.f32 %f33, [%rd274]; shl.b64 %rd275, %rd2, 2; add.s64 %rd276, %rd272, %rd275; st.global.f32 [%rd276], %f33; $L__BB225_81: ld.const.u64 %rd45, [kernel_params+3664]; setp.eq.s64 %p42, %rd45, 0; @%p42 bra $L__BB225_83; cvta.to.global.u64 %rd277, %rd45; shl.b64 %rd278, %rd1, 2; add.s64 %rd279, %rd277, %rd278; ld.global.f32 %f34, [%rd279]; shl.b64 %rd280, %rd2, 2; add.s64 %rd281, %rd277, %rd280; st.global.f32 [%rd281], %f34; $L__BB225_83: ld.const.u64 %rd46, [kernel_params+3672]; setp.eq.s64 %p43, %rd46, 0; @%p43 bra $L__BB225_85; cvta.to.global.u64 %rd282, %rd46; shl.b64 %rd283, %rd1, 2; add.s64 %rd284, %rd282, %rd283; ld.global.u32 %r53, [%rd284]; shl.b64 %rd285, %rd2, 2; add.s64 %rd286, %rd282, %rd285; st.global.u32 [%rd286], %r53; $L__BB225_85: ld.const.u64 %rd47, [kernel_params+3680]; setp.eq.s64 %p44, %rd47, 0; @%p44 bra $L__BB225_87; cvta.to.global.u64 %rd287, %rd47; shl.b64 %rd288, %rd1, 2; add.s64 %rd289, %rd287, %rd288; ld.global.u32 %r54, [%rd289]; shl.b64 %rd290, %rd2, 2; add.s64 %rd291, %rd287, %rd290; st.global.u32 [%rd291], %r54; $L__BB225_87: ld.const.u64 %rd48, [kernel_params+3688]; setp.eq.s64 %p45, %rd48, 0; @%p45 bra $L__BB225_89; cvta.to.global.u64 %rd292, %rd48; shl.b64 %rd293, %rd1, 2; add.s64 %rd294, %rd292, %rd293; ld.global.u32 %r55, [%rd294]; shl.b64 %rd295, %rd2, 2; add.s64 %rd296, %rd292, %rd295; st.global.u32 [%rd296], %r55; $L__BB225_89: ld.const.u64 %rd49, [kernel_params+3696]; setp.eq.s64 %p46, %rd49, 0; @%p46 bra $L__BB225_91; cvta.to.global.u64 %rd297, %rd49; shl.b64 %rd298, %rd1, 2; add.s64 %rd299, %rd297, %rd298; ld.global.f32 %f35, [%rd299]; shl.b64 %rd300, %rd2, 2; add.s64 %rd301, %rd297, %rd300; st.global.f32 [%rd301], %f35; $L__BB225_91: ld.const.u64 %rd50, [kernel_params+3704]; setp.eq.s64 %p47, %rd50, 0; @%p47 bra $L__BB225_93; cvta.to.global.u64 %rd302, %rd50; shl.b64 %rd303, %rd1, 2; add.s64 %rd304, %rd302, %rd303; ld.global.f32 %f36, [%rd304]; shl.b64 %rd305, %rd2, 2; add.s64 %rd306, %rd302, %rd305; st.global.f32 [%rd306], %f36; $L__BB225_93: ld.const.u64 %rd51, [kernel_params+3712]; setp.eq.s64 %p48, %rd51, 0; @%p48 bra $L__BB225_95; cvta.to.global.u64 %rd307, %rd51; shl.b64 %rd308, %rd1, 2; add.s64 %rd309, %rd307, %rd308; ld.global.f32 %f37, [%rd309]; shl.b64 %rd310, %rd2, 2; add.s64 %rd311, %rd307, %rd310; st.global.f32 [%rd311], %f37; $L__BB225_95: ld.const.u64 %rd52, [kernel_params+3720]; setp.eq.s64 %p49, %rd52, 0; @%p49 bra $L__BB225_97; cvta.to.global.u64 %rd312, %rd52; shl.b64 %rd313, %rd1, 2; add.s64 %rd314, %rd312, %rd313; ld.global.u32 %r56, [%rd314]; shl.b64 %rd315, %rd2, 2; add.s64 %rd316, %rd312, %rd315; st.global.u32 [%rd316], %r56; $L__BB225_97: ld.const.u64 %rd53, [kernel_params+3728]; setp.eq.s64 %p50, %rd53, 0; @%p50 bra $L__BB225_99; cvta.to.global.u64 %rd317, %rd53; shl.b64 %rd318, %rd1, 2; add.s64 %rd319, %rd317, %rd318; ld.global.u32 %r57, [%rd319]; shl.b64 %rd320, %rd2, 2; add.s64 %rd321, %rd317, %rd320; st.global.u32 [%rd321], %r57; $L__BB225_99: ld.const.u64 %rd54, [kernel_params+3736]; setp.eq.s64 %p51, %rd54, 0; @%p51 bra $L__BB225_101; cvta.to.global.u64 %rd322, %rd54; shl.b64 %rd323, %rd1, 2; add.s64 %rd324, %rd322, %rd323; ld.global.u32 %r58, [%rd324]; shl.b64 %rd325, %rd2, 2; add.s64 %rd326, %rd322, %rd325; st.global.u32 [%rd326], %r58; $L__BB225_101: ld.const.u32 %r60, [kernel_params+12]; max.u32 %r2, %r60, 1; add.s32 %r61, %r2, -1; and.b32 %r78, %r2, 3; setp.lt.u32 %p52, %r61, 3; mov.u32 %r77, 0; @%p52 bra $L__BB225_120; sub.s32 %r76, %r2, %r78; mov.u32 %r77, 0; mov.u64 %rd383, kernel_params; $L__BB225_103: ld.const.u64 %rd57, [%rd383+3744]; setp.eq.s64 %p53, %rd57, 0; @%p53 bra $L__BB225_105; cvta.to.global.u64 %rd328, %rd57; shl.b64 %rd329, %rd1, 2; add.s64 %rd330, %rd328, %rd329; ld.global.u32 %r63, [%rd330]; shl.b64 %rd331, %rd2, 2; add.s64 %rd332, %rd328, %rd331; st.global.u32 [%rd332], %r63; $L__BB225_105: ld.const.u64 %rd58, [%rd383+3752]; setp.eq.s64 %p54, %rd58, 0; @%p54 bra $L__BB225_107; cvta.to.global.u64 %rd333, %rd58; shl.b64 %rd334, %rd1, 2; add.s64 %rd335, %rd333, %rd334; ld.global.u32 %r64, [%rd335]; shl.b64 %rd336, %rd2, 2; add.s64 %rd337, %rd333, %rd336; st.global.u32 [%rd337], %r64; $L__BB225_107: ld.const.u64 %rd59, [%rd383+3760]; setp.eq.s64 %p55, %rd59, 0; @%p55 bra $L__BB225_109; cvta.to.global.u64 %rd338, %rd59; shl.b64 %rd339, %rd1, 2; add.s64 %rd340, %rd338, %rd339; ld.global.u32 %r65, [%rd340]; shl.b64 %rd341, %rd2, 2; add.s64 %rd342, %rd338, %rd341; st.global.u32 [%rd342], %r65; $L__BB225_109: ld.const.u64 %rd60, [%rd383+3768]; setp.eq.s64 %p56, %rd60, 0; @%p56 bra $L__BB225_111; cvta.to.global.u64 %rd343, %rd60; shl.b64 %rd344, %rd1, 2; add.s64 %rd345, %rd343, %rd344; ld.global.u32 %r66, [%rd345]; shl.b64 %rd346, %rd2, 2; add.s64 %rd347, %rd343, %rd346; st.global.u32 [%rd347], %r66; $L__BB225_111: ld.const.u64 %rd61, [%rd383+3776]; setp.eq.s64 %p57, %rd61, 0; @%p57 bra $L__BB225_113; cvta.to.global.u64 %rd348, %rd61; shl.b64 %rd349, %rd1, 2; add.s64 %rd350, %rd348, %rd349; ld.global.u32 %r67, [%rd350]; shl.b64 %rd351, %rd2, 2; add.s64 %rd352, %rd348, %rd351; st.global.u32 [%rd352], %r67; $L__BB225_113: ld.const.u64 %rd62, [%rd383+3784]; setp.eq.s64 %p58, %rd62, 0; @%p58 bra $L__BB225_115; cvta.to.global.u64 %rd353, %rd62; shl.b64 %rd354, %rd1, 2; add.s64 %rd355, %rd353, %rd354; ld.global.u32 %r68, [%rd355]; shl.b64 %rd356, %rd2, 2; add.s64 %rd357, %rd353, %rd356; st.global.u32 [%rd357], %r68; $L__BB225_115: ld.const.u64 %rd63, [%rd383+3792]; setp.eq.s64 %p59, %rd63, 0; @%p59 bra $L__BB225_117; cvta.to.global.u64 %rd358, %rd63; shl.b64 %rd359, %rd1, 2; add.s64 %rd360, %rd358, %rd359; ld.global.u32 %r69, [%rd360]; shl.b64 %rd361, %rd2, 2; add.s64 %rd362, %rd358, %rd361; st.global.u32 [%rd362], %r69; $L__BB225_117: ld.const.u64 %rd64, [%rd383+3800]; setp.eq.s64 %p60, %rd64, 0; @%p60 bra $L__BB225_119; cvta.to.global.u64 %rd363, %rd64; shl.b64 %rd364, %rd1, 2; add.s64 %rd365, %rd363, %rd364; ld.global.u32 %r70, [%rd365]; shl.b64 %rd366, %rd2, 2; add.s64 %rd367, %rd363, %rd366; st.global.u32 [%rd367], %r70; $L__BB225_119: add.s32 %r77, %r77, 4; add.s64 %rd383, %rd383, 64; add.s32 %r76, %r76, -4; setp.ne.s32 %p61, %r76, 0; @%p61 bra $L__BB225_103; $L__BB225_120: setp.eq.s32 %p62, %r78, 0; @%p62 bra $L__BB225_127; shl.b32 %r71, %r77, 4; add.s32 %r72, %r71, 3752; cvt.s64.s32 %rd368, %r72; mov.u64 %rd369, kernel_params; add.s64 %rd384, %rd369, %rd368; $L__BB225_122: .pragma "nounroll"; add.s64 %rd68, %rd384, -8; ld.const.u64 %rd69, [%rd384+-8]; setp.eq.s64 %p63, %rd69, 0; @%p63 bra $L__BB225_124; cvta.to.global.u64 %rd370, %rd69; shl.b64 %rd371, %rd1, 2; add.s64 %rd372, %rd370, %rd371; ld.global.u32 %r73, [%rd372]; shl.b64 %rd373, %rd2, 2; add.s64 %rd374, %rd370, %rd373; st.global.u32 [%rd374], %r73; $L__BB225_124: ld.const.u64 %rd70, [%rd68+8]; setp.eq.s64 %p64, %rd70, 0; @%p64 bra $L__BB225_126; cvta.to.global.u64 %rd375, %rd70; shl.b64 %rd376, %rd1, 2; add.s64 %rd377, %rd375, %rd376; ld.global.u32 %r74, [%rd377]; shl.b64 %rd378, %rd2, 2; add.s64 %rd379, %rd375, %rd378; st.global.u32 [%rd379], %r74; $L__BB225_126: add.s32 %r78, %r78, -1; add.s64 %rd384, %rd384, 16; setp.ne.s32 %p65, %r78, 0; @%p65 bra $L__BB225_122; $L__BB225_127: cvta.to.global.u64 %rd380, %rd12; shl.b64 %rd381, %rd1, 1; add.s64 %rd382, %rd380, %rd381; mov.u16 %rs10, 0; st.global.u16 [%rd382], %rs10; $L__BB225_128: ret; } // .globl kernel_gpu_prefix_sum .visible .entry kernel_gpu_prefix_sum( .param .u64 kernel_gpu_prefix_sum_param_0, .param .u64 kernel_gpu_prefix_sum_param_1, .param .u32 kernel_gpu_prefix_sum_param_2 ) .maxntid 512, 1, 1 { .reg .pred %p<8>; .reg .b32 %r<44>; .reg .b64 %rd<21>; ld.param.u64 %rd14, [kernel_gpu_prefix_sum_param_0]; ld.param.u64 %rd13, [kernel_gpu_prefix_sum_param_1]; ld.param.u32 %r16, [kernel_gpu_prefix_sum_param_2]; cvta.to.global.u64 %rd1, %rd14; mov.u32 %r17, %ntid.x; mov.u32 %r18, %ctaid.x; mul.lo.s32 %r19, %r18, %r17; mov.u32 %r20, %tid.x; neg.s32 %r21, %r20; setp.ne.s32 %p1, %r19, %r21; setp.lt.s32 %p2, %r16, 1; or.pred %p3, %p1, %p2; @%p3 bra $L__BB226_7; add.s32 %r24, %r16, -1; and.b32 %r43, %r16, 3; setp.lt.u32 %p4, %r24, 3; mov.u32 %r40, 0; mov.u32 %r41, %r40; @%p4 bra $L__BB226_4; sub.s32 %r39, %r16, %r43; cvta.to.global.u64 %rd17, %rd13; mov.u32 %r26, 0; mov.u64 %rd18, %rd1; mov.u32 %r40, %r26; mov.u32 %r41, %r26; $L__BB226_3: ld.global.u32 %r27, [%rd18]; add.s32 %r28, %r27, %r41; st.global.u32 [%rd17], %r41; st.global.u32 [%rd18], %r26; ld.global.u32 %r30, [%rd18+4]; add.s32 %r31, %r30, %r28; st.global.u32 [%rd17+4], %r28; st.global.u32 [%rd18+4], %r26; ld.global.u32 %r32, [%rd18+8]; add.s32 %r33, %r32, %r31; st.global.u32 [%rd17+8], %r31; st.global.u32 [%rd18+8], %r26; ld.global.u32 %r34, [%rd18+12]; add.s32 %r41, %r34, %r33; st.global.u32 [%rd17+12], %r33; st.global.u32 [%rd18+12], %r26; add.s32 %r40, %r40, 4; add.s64 %rd18, %rd18, 16; add.s64 %rd17, %rd17, 16; add.s32 %r39, %r39, -4; setp.ne.s32 %p5, %r39, 0; @%p5 bra $L__BB226_3; $L__BB226_4: setp.eq.s32 %p6, %r43, 0; @%p6 bra $L__BB226_7; cvta.to.global.u64 %rd15, %rd13; mul.wide.s32 %rd16, %r40, 4; add.s64 %rd20, %rd15, %rd16; add.s64 %rd19, %rd1, %rd16; $L__BB226_6: .pragma "nounroll"; ld.global.u32 %r35, [%rd19]; add.s32 %r14, %r35, %r41; st.global.u32 [%rd20], %r41; mov.u32 %r36, 0; st.global.u32 [%rd19], %r36; add.s64 %rd20, %rd20, 4; add.s64 %rd19, %rd19, 4; add.s32 %r43, %r43, -1; setp.ne.s32 %p7, %r43, 0; mov.u32 %r41, %r14; @%p7 bra $L__BB226_6; $L__BB226_7: ret; } // .globl kernel_gpu_adaptive_sampling_convergence_check .visible .entry kernel_gpu_adaptive_sampling_convergence_check( .param .u64 kernel_gpu_adaptive_sampling_convergence_check_param_0, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_1, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_2, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_3, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_4, .param .f32 kernel_gpu_adaptive_sampling_convergence_check_param_5, .param .u8 kernel_gpu_adaptive_sampling_convergence_check_param_6, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_7, .param .u32 kernel_gpu_adaptive_sampling_convergence_check_param_8, .param .u64 kernel_gpu_adaptive_sampling_convergence_check_param_9 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<15>; .reg .b16 %rs<2>; .reg .f32 %f<27>; .reg .b32 %r<28>; .reg .b64 %rd<23>; ld.param.s8 %rs1, [kernel_gpu_adaptive_sampling_convergence_check_param_6]; ld.param.u64 %rd3, [kernel_gpu_adaptive_sampling_convergence_check_param_0]; ld.param.u32 %r6, [kernel_gpu_adaptive_sampling_convergence_check_param_1]; ld.param.u32 %r7, [kernel_gpu_adaptive_sampling_convergence_check_param_2]; ld.param.u32 %r10, [kernel_gpu_adaptive_sampling_convergence_check_param_3]; ld.param.u32 %r11, [kernel_gpu_adaptive_sampling_convergence_check_param_4]; ld.param.f32 %f1, [kernel_gpu_adaptive_sampling_convergence_check_param_5]; ld.param.u32 %r8, [kernel_gpu_adaptive_sampling_convergence_check_param_7]; ld.param.u32 %r9, [kernel_gpu_adaptive_sampling_convergence_check_param_8]; ld.param.u64 %rd4, [kernel_gpu_adaptive_sampling_convergence_check_param_9]; mov.u32 %r12, %ntid.x; mov.u32 %r13, %ctaid.x; mov.u32 %r1, %tid.x; mad.lo.s32 %r14, %r13, %r12, %r1; div.s32 %r2, %r14, %r10; mul.lo.s32 %r15, %r2, %r10; sub.s32 %r3, %r14, %r15; setp.ge.s32 %p5, %r3, %r10; setp.ge.s32 %p6, %r2, %r11; mov.pred %p14, 0; or.pred %p7, %p6, %p5; @%p7 bra $L__BB227_3; add.s32 %r16, %r2, %r7; add.s32 %r17, %r8, %r6; add.s32 %r18, %r17, %r3; mad.lo.s32 %r19, %r16, %r9, %r18; ld.const.u32 %r20, [kernel_params+1264]; mul.wide.s32 %rd1, %r20, %r19; ld.const.u32 %r4, [kernel_params+1400]; cvt.s64.s32 %rd5, %r4; add.s64 %rd6, %rd1, %rd5; cvta.to.global.u64 %rd7, %rd3; shl.b64 %rd8, %rd6, 2; add.s64 %rd2, %rd7, %rd8; ld.global.f32 %f2, [%rd2+12]; setp.neu.ftz.f32 %p9, %f2, 0f00000000; setp.eq.s16 %p10, %rs1, 0; and.pred %p11, %p10, %p9; @%p11 bra $L__BB227_3; ld.global.f32 %f3, [%rd2]; ld.const.s32 %rd9, [kernel_params+1268]; add.s64 %rd10, %rd1, %rd9; shl.b64 %rd12, %rd10, 2; add.s64 %rd13, %rd7, %rd12; ld.const.s32 %rd14, [kernel_params+1404]; add.s64 %rd15, %rd1, %rd14; shl.b64 %rd16, %rd15, 2; add.s64 %rd17, %rd7, %rd16; ld.global.u32 %r21, [%rd17]; cvt.rn.f32.u32 %f4, %r21; rcp.approx.ftz.f32 %f5, %f4; ld.global.f32 %f6, [%rd13]; sub.ftz.f32 %f7, %f6, %f3; abs.ftz.f32 %f8, %f7; ld.global.f32 %f9, [%rd13+4]; ld.global.f32 %f10, [%rd2+4]; sub.ftz.f32 %f11, %f9, %f10; abs.ftz.f32 %f12, %f11; add.ftz.f32 %f13, %f8, %f12; ld.global.f32 %f14, [%rd13+8]; ld.global.f32 %f15, [%rd2+8]; sub.ftz.f32 %f16, %f14, %f15; abs.ftz.f32 %f17, %f16; add.ftz.f32 %f18, %f13, %f17; mul.ftz.f32 %f19, %f5, %f18; add.ftz.f32 %f20, %f6, %f9; add.ftz.f32 %f21, %f14, %f20; mul.ftz.f32 %f22, %f21, %f5; sqrt.approx.ftz.f32 %f23, %f22; add.ftz.f32 %f24, %f23, 0f38D1B717; div.approx.ftz.f32 %f25, %f19, %f24; setp.lt.ftz.f32 %p12, %f25, %f1; selp.f32 %f26, 0f3F800000, 0f00000000, %p12; add.s32 %r22, %r4, 3; cvt.u64.u32 %rd18, %r22; add.s64 %rd19, %rd1, %rd18; shl.b64 %rd20, %rd19, 2; add.s64 %rd21, %rd7, %rd20; st.global.f32 [%rd21], %f26; setp.geu.ftz.f32 %p14, %f25, %f1; $L__BB227_3: mov.u32 %r23, -1; vote.sync.ballot.b32 %r5, %p14, %r23; mov.u32 %r24, WARP_SZ; rem.u32 %r25, %r1, %r24; setp.ne.s32 %p13, %r25, 0; @%p13 bra $L__BB227_5; popc.b32 %r26, %r5; cvta.to.global.u64 %rd22, %rd4; atom.global.add.u32 %r27, [%rd22], %r26; $L__BB227_5: ret; } // .globl kernel_gpu_adaptive_sampling_filter_x .visible .entry kernel_gpu_adaptive_sampling_filter_x( .param .u64 kernel_gpu_adaptive_sampling_filter_x_param_0, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_1, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_2, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_3, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_4, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_5, .param .u32 kernel_gpu_adaptive_sampling_filter_x_param_6 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<47>; .reg .b16 %rs<25>; .reg .f32 %f<6>; .reg .b32 %r<69>; .reg .b64 %rd<36>; ld.param.u64 %rd19, [kernel_gpu_adaptive_sampling_filter_x_param_0]; ld.param.u32 %r15, [kernel_gpu_adaptive_sampling_filter_x_param_1]; ld.param.u32 %r16, [kernel_gpu_adaptive_sampling_filter_x_param_2]; ld.param.u32 %r17, [kernel_gpu_adaptive_sampling_filter_x_param_3]; ld.param.u32 %r20, [kernel_gpu_adaptive_sampling_filter_x_param_4]; ld.param.u32 %r18, [kernel_gpu_adaptive_sampling_filter_x_param_5]; ld.param.u32 %r19, [kernel_gpu_adaptive_sampling_filter_x_param_6]; cvta.to.global.u64 %rd1, %rd19; mov.u32 %r21, %ntid.x; mov.u32 %r22, %ctaid.x; mul.lo.s32 %r1, %r22, %r21; mov.u32 %r2, %tid.x; add.s32 %r23, %r1, %r2; setp.ge.s32 %p4, %r23, %r20; @%p4 bra $L__BB228_33; setp.lt.s32 %p5, %r17, 1; @%p5 bra $L__BB228_33; ld.const.s32 %rd2, [kernel_params+1264]; ld.const.u32 %r24, [kernel_params+1400]; add.s32 %r25, %r24, 3; cvt.u64.u32 %rd3, %r25; add.s32 %r26, %r15, 1; add.s32 %r27, %r17, %r15; max.s32 %r3, %r27, %r26; not.b32 %r28, %r15; add.s32 %r29, %r3, %r28; setp.lt.u32 %p6, %r29, 3; mov.u16 %rs21, 0; mov.u32 %r66, %r15; @%p6 bra $L__BB228_25; sub.s32 %r30, %r3, %r15; and.b32 %r31, %r30, 3; sub.s32 %r65, %r30, %r31; add.s32 %r32, %r2, %r16; add.s32 %r33, %r32, %r1; add.s32 %r34, %r18, %r15; mad.lo.s32 %r35, %r19, %r33, %r34; cvt.s64.s32 %rd20, %r35; mul.lo.s64 %rd21, %rd2, %rd20; add.s64 %rd22, %rd21, %rd3; shl.b64 %rd23, %rd22, 2; add.s64 %rd34, %rd1, %rd23; shl.b64 %rd5, %rd2, 2; neg.s64 %rd6, %rd2; mov.u16 %rs21, 0; mov.pred %p11, 0; shl.b64 %rd24, %rd6, 2; mov.u16 %rs12, 1; mov.pred %p25, -1; mov.u32 %r66, %r15; $L__BB228_4: ld.global.f32 %f1, [%rd34]; setp.eq.ftz.f32 %p7, %f1, 0f00000000; @%p7 bra $L__BB228_7; bra.uni $L__BB228_5; $L__BB228_7: and.b16 %rs9, %rs21, 255; setp.ne.s16 %p12, %rs9, 0; setp.le.s32 %p13, %r66, %r15; or.pred %p14, %p12, %p13; mov.pred %p44, %p11; @%p14 bra $L__BB228_9; add.s64 %rd25, %rd34, %rd24; mov.u32 %r37, 0; st.global.u32 [%rd25], %r37; mov.pred %p44, %p11; bra.uni $L__BB228_9; $L__BB228_5: and.b16 %rs8, %rs21, 255; setp.eq.s16 %p9, %rs8, 0; mov.pred %p44, -1; @%p9 bra $L__BB228_9; mov.u32 %r36, 0; st.global.u32 [%rd34], %r36; $L__BB228_9: add.s64 %rd9, %rd34, %rd5; ld.global.f32 %f2, [%rd9]; setp.eq.ftz.f32 %p16, %f2, 0f00000000; @%p16 bra $L__BB228_12; bra.uni $L__BB228_10; $L__BB228_12: setp.lt.s32 %p20, %r66, %r15; not.pred %p21, %p44; or.pred %p22, %p21, %p20; mov.pred %p45, %p11; @%p22 bra $L__BB228_14; mov.u32 %r39, 0; st.global.u32 [%rd34], %r39; mov.pred %p45, %p11; bra.uni $L__BB228_14; $L__BB228_10: mov.pred %p45, -1; @%p44 bra $L__BB228_14; mov.u32 %r38, 0; st.global.u32 [%rd9], %r38; $L__BB228_14: add.s64 %rd11, %rd9, %rd5; ld.global.f32 %f3, [%rd11]; setp.eq.ftz.f32 %p24, %f3, 0f00000000; @%p24 bra $L__BB228_17; bra.uni $L__BB228_15; $L__BB228_17: add.s32 %r41, %r66, 2; setp.le.s32 %p28, %r41, %r15; not.pred %p29, %p45; or.pred %p30, %p29, %p28; mov.pred %p46, %p11; @%p30 bra $L__BB228_19; mov.u32 %r42, 0; st.global.u32 [%rd9], %r42; mov.pred %p46, %p11; bra.uni $L__BB228_19; $L__BB228_15: mov.pred %p46, %p25; @%p45 bra $L__BB228_19; mov.u32 %r40, 0; st.global.u32 [%rd11], %r40; mov.pred %p46, %p25; $L__BB228_19: add.s64 %rd13, %rd11, %rd5; ld.global.f32 %f4, [%rd13]; setp.eq.ftz.f32 %p32, %f4, 0f00000000; @%p32 bra $L__BB228_22; bra.uni $L__BB228_20; $L__BB228_22: add.s32 %r44, %r66, 3; setp.le.s32 %p33, %r44, %r15; not.pred %p34, %p46; or.pred %p35, %p34, %p33; mov.u16 %rs21, %rs12; @%p35 bra $L__BB228_24; mov.u32 %r45, 0; st.global.u32 [%rd11], %r45; mov.u16 %rs21, %rs12; bra.uni $L__BB228_24; $L__BB228_20: mov.u16 %rs21, 0; @%p46 bra $L__BB228_24; mov.u32 %r43, 0; st.global.u32 [%rd13], %r43; $L__BB228_24: add.s32 %r66, %r66, 4; add.s32 %r65, %r65, -4; setp.ne.s32 %p36, %r65, 0; add.s64 %rd34, %rd13, %rd5; @%p36 bra $L__BB228_4; $L__BB228_25: sub.s32 %r49, %r3, %r15; and.b32 %r68, %r49, 3; setp.eq.s32 %p37, %r68, 0; @%p37 bra $L__BB228_33; add.s32 %r51, %r66, %r18; add.s32 %r53, %r2, %r16; mad.lo.s32 %r56, %r22, %r21, %r53; mad.lo.s32 %r57, %r19, %r56, %r51; cvt.s64.s32 %rd35, %r57; mov.u16 %rs17, 1; mov.u16 %rs14, 0; $L__BB228_27: .pragma "nounroll"; mul.lo.s64 %rd26, %rd2, %rd35; add.s64 %rd27, %rd26, %rd3; shl.b64 %rd28, %rd27, 2; add.s64 %rd17, %rd1, %rd28; ld.global.f32 %f5, [%rd17]; setp.eq.ftz.f32 %p38, %f5, 0f00000000; @%p38 bra $L__BB228_30; bra.uni $L__BB228_28; $L__BB228_30: and.b16 %rs18, %rs21, 255; setp.ne.s16 %p40, %rs18, 0; setp.le.s32 %p41, %r66, %r15; or.pred %p42, %p40, %p41; mov.u16 %rs21, %rs17; @%p42 bra $L__BB228_32; add.s64 %rd29, %rd35, -1; mul.lo.s64 %rd30, %rd2, %rd29; add.s64 %rd31, %rd30, %rd3; shl.b64 %rd32, %rd31, 2; add.s64 %rd33, %rd1, %rd32; mov.u32 %r63, 0; st.global.u32 [%rd33], %r63; mov.u16 %rs21, %rs17; bra.uni $L__BB228_32; $L__BB228_28: and.b16 %rs15, %rs21, 255; setp.eq.s16 %p39, %rs15, 0; mov.u16 %rs21, %rs14; @%p39 bra $L__BB228_32; mov.u32 %r62, 0; st.global.u32 [%rd17], %r62; mov.u16 %rs21, %rs14; $L__BB228_32: add.s32 %r66, %r66, 1; add.s64 %rd35, %rd35, 1; add.s32 %r68, %r68, -1; setp.ne.s32 %p43, %r68, 0; @%p43 bra $L__BB228_27; $L__BB228_33: ret; } // .globl kernel_gpu_adaptive_sampling_filter_y .visible .entry kernel_gpu_adaptive_sampling_filter_y( .param .u64 kernel_gpu_adaptive_sampling_filter_y_param_0, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_1, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_2, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_3, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_4, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_5, .param .u32 kernel_gpu_adaptive_sampling_filter_y_param_6 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<47>; .reg .b16 %rs<25>; .reg .f32 %f<6>; .reg .b32 %r<76>; .reg .b64 %rd<87>; ld.param.u64 %rd36, [kernel_gpu_adaptive_sampling_filter_y_param_0]; ld.param.u32 %r18, [kernel_gpu_adaptive_sampling_filter_y_param_1]; ld.param.u32 %r19, [kernel_gpu_adaptive_sampling_filter_y_param_2]; ld.param.u32 %r23, [kernel_gpu_adaptive_sampling_filter_y_param_3]; ld.param.u32 %r20, [kernel_gpu_adaptive_sampling_filter_y_param_4]; ld.param.u32 %r21, [kernel_gpu_adaptive_sampling_filter_y_param_5]; ld.param.u32 %r22, [kernel_gpu_adaptive_sampling_filter_y_param_6]; mov.u32 %r24, %ctaid.x; mov.u32 %r25, %ntid.x; mul.lo.s32 %r1, %r24, %r25; mov.u32 %r2, %tid.x; add.s32 %r26, %r1, %r2; setp.ge.s32 %p4, %r26, %r23; @%p4 bra $L__BB229_33; setp.lt.s32 %p5, %r20, 1; @%p5 bra $L__BB229_33; ld.const.u32 %r3, [kernel_params+1264]; ld.const.u32 %r27, [kernel_params+1400]; add.s32 %r4, %r27, 3; add.s32 %r5, %r19, 1; add.s32 %r28, %r20, %r19; max.s32 %r6, %r28, %r5; not.b32 %r29, %r19; add.s32 %r30, %r6, %r29; setp.lt.u32 %p6, %r30, 3; mov.u16 %rs21, 0; mov.u32 %r73, %r19; @%p6 bra $L__BB229_25; cvt.s64.s32 %rd37, %r3; sub.s32 %r31, %r6, %r19; and.b32 %r32, %r31, 3; sub.s32 %r72, %r31, %r32; add.s32 %r33, %r2, %r21; add.s32 %r34, %r33, %r18; add.s32 %r35, %r34, %r1; mad.lo.s32 %r36, %r22, %r5, %r35; cvt.s64.s32 %rd84, %r36; add.s32 %r37, %r19, 2; mad.lo.s32 %r38, %r22, %r37, %r35; cvt.s64.s32 %rd83, %r38; add.s32 %r39, %r19, 3; mad.lo.s32 %r40, %r22, %r39, %r35; cvt.s64.s32 %rd82, %r40; mad.lo.s32 %r41, %r22, %r19, %r35; cvt.s64.s32 %rd80, %r41; cvt.s64.s32 %rd38, %r22; sub.s64 %rd39, %rd80, %rd38; mul.lo.s64 %rd40, %rd39, %rd37; cvt.u64.u32 %rd41, %r4; add.s64 %rd81, %rd40, %rd41; mul.wide.s32 %rd42, %r3, %r22; shl.b64 %rd6, %rd42, 2; mov.u16 %rs21, 0; mov.pred %p11, 0; mov.u16 %rs12, 1; mov.pred %p25, -1; mov.u32 %r73, %r19; $L__BB229_4: mul.lo.s64 %rd44, %rd37, %rd80; add.s64 %rd46, %rd44, %rd41; cvta.to.global.u64 %rd47, %rd36; shl.b64 %rd48, %rd46, 2; add.s64 %rd12, %rd47, %rd48; ld.global.f32 %f1, [%rd12]; setp.eq.ftz.f32 %p7, %f1, 0f00000000; @%p7 bra $L__BB229_7; bra.uni $L__BB229_5; $L__BB229_7: and.b16 %rs9, %rs21, 255; setp.ne.s16 %p12, %rs9, 0; setp.le.s32 %p13, %r73, %r19; or.pred %p14, %p12, %p13; mov.pred %p44, %p11; @%p14 bra $L__BB229_9; shl.b64 %rd50, %rd81, 2; add.s64 %rd51, %rd47, %rd50; mov.u32 %r43, 0; st.global.u32 [%rd51], %r43; mov.pred %p44, %p11; bra.uni $L__BB229_9; $L__BB229_5: and.b16 %rs8, %rs21, 255; setp.eq.s16 %p9, %rs8, 0; mov.pred %p44, -1; @%p9 bra $L__BB229_9; mov.u32 %r42, 0; st.global.u32 [%rd12], %r42; $L__BB229_9: mul.lo.s64 %rd53, %rd37, %rd84; add.s64 %rd55, %rd53, %rd41; shl.b64 %rd57, %rd55, 2; add.s64 %rd13, %rd47, %rd57; add.s64 %rd14, %rd12, %rd6; ld.global.f32 %f2, [%rd14]; setp.eq.ftz.f32 %p16, %f2, 0f00000000; @%p16 bra $L__BB229_12; bra.uni $L__BB229_10; $L__BB229_12: setp.lt.s32 %p20, %r73, %r19; not.pred %p21, %p44; or.pred %p22, %p21, %p20; mov.pred %p45, %p11; @%p22 bra $L__BB229_14; mov.u32 %r45, 0; st.global.u32 [%rd12], %r45; mov.pred %p45, %p11; bra.uni $L__BB229_14; $L__BB229_10: mov.pred %p45, -1; @%p44 bra $L__BB229_14; mov.u32 %r44, 0; st.global.u32 [%rd13], %r44; $L__BB229_14: mul.lo.s64 %rd59, %rd37, %rd83; add.s64 %rd61, %rd59, %rd41; shl.b64 %rd63, %rd61, 2; add.s64 %rd15, %rd47, %rd63; add.s64 %rd16, %rd14, %rd6; ld.global.f32 %f3, [%rd16]; setp.eq.ftz.f32 %p24, %f3, 0f00000000; @%p24 bra $L__BB229_17; bra.uni $L__BB229_15; $L__BB229_17: add.s32 %r47, %r73, 2; setp.le.s32 %p28, %r47, %r19; not.pred %p29, %p45; or.pred %p30, %p29, %p28; mov.pred %p46, %p11; @%p30 bra $L__BB229_19; mov.u32 %r48, 0; st.global.u32 [%rd13], %r48; mov.pred %p46, %p11; bra.uni $L__BB229_19; $L__BB229_15: mov.pred %p46, %p25; @%p45 bra $L__BB229_19; mov.u32 %r46, 0; st.global.u32 [%rd15], %r46; mov.pred %p46, %p25; $L__BB229_19: mul.lo.s64 %rd64, %rd37, %rd82; add.s64 %rd66, %rd64, %rd41; shl.b64 %rd68, %rd66, 2; add.s64 %rd18, %rd47, %rd68; add.s64 %rd69, %rd16, %rd6; ld.global.f32 %f4, [%rd69]; setp.eq.ftz.f32 %p32, %f4, 0f00000000; @%p32 bra $L__BB229_22; bra.uni $L__BB229_20; $L__BB229_22: add.s32 %r50, %r73, 3; setp.le.s32 %p33, %r50, %r19; not.pred %p34, %p46; or.pred %p35, %p34, %p33; mov.u16 %rs21, %rs12; @%p35 bra $L__BB229_24; mov.u32 %r51, 0; st.global.u32 [%rd15], %r51; mov.u16 %rs21, %rs12; bra.uni $L__BB229_24; $L__BB229_20: mov.u16 %rs21, 0; @%p46 bra $L__BB229_24; mov.u32 %r49, 0; st.global.u32 [%rd18], %r49; $L__BB229_24: add.s32 %r73, %r73, 4; shl.b32 %r52, %r22, 2; cvt.s64.s32 %rd70, %r52; add.s64 %rd84, %rd84, %rd70; add.s64 %rd83, %rd83, %rd70; add.s64 %rd82, %rd82, %rd70; mul.lo.s64 %rd71, %rd37, %rd70; add.s64 %rd81, %rd81, %rd71; add.s64 %rd80, %rd80, %rd70; add.s32 %r72, %r72, -4; setp.ne.s32 %p36, %r72, 0; @%p36 bra $L__BB229_4; $L__BB229_25: sub.s32 %r56, %r6, %r19; and.b32 %r75, %r56, 3; setp.eq.s32 %p37, %r75, 0; @%p37 bra $L__BB229_33; add.s32 %r59, %r2, %r21; add.s32 %r60, %r59, %r18; mad.lo.s32 %r63, %r24, %r25, %r60; mad.lo.s32 %r64, %r73, %r22, %r63; cvt.s64.s32 %rd85, %r64; cvt.s64.s32 %rd25, %r22; sub.s64 %rd72, %rd85, %rd25; cvt.s64.s32 %rd26, %r3; mul.lo.s64 %rd73, %rd72, %rd26; cvt.u64.u32 %rd27, %r4; add.s64 %rd86, %rd73, %rd27; mul.wide.s32 %rd29, %r3, %r22; cvta.to.global.u64 %rd30, %rd36; mov.u16 %rs17, 1; mov.u16 %rs14, 0; $L__BB229_27: .pragma "nounroll"; mul.lo.s64 %rd74, %rd26, %rd85; add.s64 %rd75, %rd74, %rd27; shl.b64 %rd76, %rd75, 2; add.s64 %rd33, %rd30, %rd76; ld.global.f32 %f5, [%rd33]; setp.eq.ftz.f32 %p38, %f5, 0f00000000; @%p38 bra $L__BB229_30; bra.uni $L__BB229_28; $L__BB229_30: and.b16 %rs18, %rs21, 255; setp.ne.s16 %p40, %rs18, 0; setp.le.s32 %p41, %r73, %r19; or.pred %p42, %p40, %p41; mov.u16 %rs21, %rs17; @%p42 bra $L__BB229_32; shl.b64 %rd78, %rd86, 2; add.s64 %rd79, %rd30, %rd78; mov.u32 %r70, 0; st.global.u32 [%rd79], %r70; mov.u16 %rs21, %rs17; bra.uni $L__BB229_32; $L__BB229_28: and.b16 %rs15, %rs21, 255; setp.eq.s16 %p39, %rs15, 0; mov.u16 %rs21, %rs14; @%p39 bra $L__BB229_32; mov.u32 %r69, 0; st.global.u32 [%rd33], %r69; mov.u16 %rs21, %rs14; $L__BB229_32: add.s32 %r73, %r73, 1; add.s64 %rd86, %rd86, %rd29; add.s64 %rd85, %rd85, %rd25; add.s32 %r75, %r75, -1; setp.ne.s32 %p43, %r75, 0; @%p43 bra $L__BB229_27; $L__BB229_33: ret; } // .globl kernel_gpu_cryptomatte_postprocess .visible .entry kernel_gpu_cryptomatte_postprocess( .param .u64 kernel_gpu_cryptomatte_postprocess_param_0, .param .u32 kernel_gpu_cryptomatte_postprocess_param_1 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<7>; .reg .f32 %f<5>; .reg .b32 %r<19>; .reg .b64 %rd<13>; ld.param.u64 %rd4, [kernel_gpu_cryptomatte_postprocess_param_0]; ld.param.u32 %r8, [kernel_gpu_cryptomatte_postprocess_param_1]; mov.u32 %r9, %ctaid.x; mov.u32 %r10, %ntid.x; mov.u32 %r11, %tid.x; mad.lo.s32 %r1, %r9, %r10, %r11; setp.ge.s32 %p1, %r1, %r8; @%p1 bra $L__BB230_8; cvta.to.global.u64 %rd5, %rd4; ld.const.u32 %r12, [kernel_params+1264]; mul.wide.s32 %rd6, %r12, %r1; ld.const.v2.u32 {%r13, %r14}, [kernel_params+1392]; cvt.s64.s32 %rd7, %r14; add.s64 %rd8, %rd6, %rd7; shl.b64 %rd9, %rd8, 2; add.s64 %rd1, %rd5, %rd9; setp.lt.s32 %p2, %r13, 1; @%p2 bra $L__BB230_8; shl.b32 %r3, %r13, 1; mov.u32 %r17, 1; $L__BB230_3: mul.wide.s32 %rd10, %r17, 8; add.s64 %rd2, %rd1, %rd10; ld.global.f32 %f1, [%rd2]; setp.eq.ftz.f32 %p3, %f1, 0f00000000; @%p3 bra $L__BB230_8; ld.global.f32 %f2, [%rd2+4]; mov.u32 %r18, %r17; $L__BB230_5: mul.wide.s32 %rd11, %r18, 8; add.s64 %rd12, %rd1, %rd11; add.s64 %rd3, %rd12, 4; ld.global.f32 %f3, [%rd12+-4]; setp.leu.ftz.f32 %p4, %f2, %f3; @%p4 bra $L__BB230_7; ld.global.f32 %f4, [%rd3+-12]; st.global.f32 [%rd3+-4], %f4; st.global.f32 [%rd3], %f3; st.global.f32 [%rd3+-12], %f1; st.global.f32 [%rd3+-8], %f2; setp.gt.s32 %p5, %r18, 1; add.s32 %r18, %r18, -1; @%p5 bra $L__BB230_5; $L__BB230_7: add.s32 %r17, %r17, 1; setp.lt.s32 %p6, %r17, %r3; @%p6 bra $L__BB230_3; $L__BB230_8: ret; } // .globl kernel_gpu_film_convert_depth .visible .entry kernel_gpu_film_convert_depth( .param .align 4 .b8 kernel_gpu_film_convert_depth_param_0[96], .param .u64 kernel_gpu_film_convert_depth_param_1, .param .u64 kernel_gpu_film_convert_depth_param_2, .param .u32 kernel_gpu_film_convert_depth_param_3, .param .u32 kernel_gpu_film_convert_depth_param_4, .param .u32 kernel_gpu_film_convert_depth_param_5, .param .u32 kernel_gpu_film_convert_depth_param_6, .param .u32 kernel_gpu_film_convert_depth_param_7, .param .u32 kernel_gpu_film_convert_depth_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<6>; .reg .f32 %f<17>; .reg .b32 %r<39>; .reg .b64 %rd<19>; ld.param.u64 %rd2, [kernel_gpu_film_convert_depth_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_depth_param_2]; ld.param.u32 %r28, [kernel_gpu_film_convert_depth_param_3]; ld.param.u32 %r24, [kernel_gpu_film_convert_depth_param_4]; ld.param.u32 %r25, [kernel_gpu_film_convert_depth_param_5]; ld.param.u32 %r26, [kernel_gpu_film_convert_depth_param_6]; ld.param.u32 %r27, [kernel_gpu_film_convert_depth_param_7]; ld.param.u32 %r21, [kernel_gpu_film_convert_depth_param_0+84]; ld.param.f32 %f16, [kernel_gpu_film_convert_depth_param_0+64]; ld.param.f32 %f7, [kernel_gpu_film_convert_depth_param_0+60]; ld.param.u32 %r10, [kernel_gpu_film_convert_depth_param_0+28]; ld.param.u32 %r6, [kernel_gpu_film_convert_depth_param_0+12]; ld.param.u32 %r5, [kernel_gpu_film_convert_depth_param_0+8]; ld.param.u32 %r4, [kernel_gpu_film_convert_depth_param_0+4]; ld.param.u32 %r3, [kernel_gpu_film_convert_depth_param_0]; mov.u32 %r29, %ntid.x; mov.u32 %r30, %ctaid.x; mov.u32 %r31, %tid.x; mad.lo.s32 %r1, %r30, %r29, %r31; setp.ge.s32 %p1, %r1, %r28; @%p1 bra $L__BB231_6; div.s32 %r32, %r1, %r24; mul.lo.s32 %r33, %r32, %r24; sub.s32 %r34, %r1, %r33; mad.lo.s32 %r35, %r32, %r26, %r34; mul.wide.s32 %rd4, %r35, %r4; cvt.s64.s32 %rd5, %r25; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r10, -1; @%p2 bra $L__BB231_5; setp.eq.s32 %p3, %r6, 0; mov.f32 %f15, 0f3F800000; @%p3 bra $L__BB231_4; cvt.s64.s32 %rd6, %r10; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r36, [%rd10]; cvt.rn.f32.u32 %f10, %r36; rcp.approx.ftz.f32 %f15, %f10; $L__BB231_4: mul.ftz.f32 %f11, %f7, %f15; setp.eq.s32 %p4, %r5, 0; selp.f32 %f16, %f15, %f11, %p4; $L__BB231_5: cvt.s64.s32 %rd11, %r3; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f12, [%rd15]; setp.eq.ftz.f32 %p5, %f12, 0f00000000; mul.ftz.f32 %f13, %f16, %f12; selp.f32 %f14, 0f501502F9, %f13, %p5; cvta.to.global.u64 %rd16, %rd2; add.s32 %r37, %r1, %r27; mul.lo.s32 %r38, %r37, %r21; mul.wide.s32 %rd17, %r38, 4; add.s64 %rd18, %rd16, %rd17; st.global.f32 [%rd18], %f14; $L__BB231_6: ret; } // .globl kernel_gpu_film_convert_depth_half_rgba .visible .entry kernel_gpu_film_convert_depth_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_depth_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_depth_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_depth_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_depth_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<10>; .reg .b16 %rs<5>; .reg .f32 %f<33>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_depth_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_depth_half_rgba_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_depth_half_rgba_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_depth_half_rgba_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_depth_half_rgba_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_depth_half_rgba_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_depth_half_rgba_param_7]; ld.param.u32 %r31, [kernel_gpu_film_convert_depth_half_rgba_param_8]; ld.param.u32 %r22, [kernel_gpu_film_convert_depth_half_rgba_param_0+76]; ld.param.f32 %f30, [kernel_gpu_film_convert_depth_half_rgba_param_0+64]; ld.param.f32 %f12, [kernel_gpu_film_convert_depth_half_rgba_param_0+60]; ld.param.u32 %r14, [kernel_gpu_film_convert_depth_half_rgba_param_0+32]; ld.param.u32 %r13, [kernel_gpu_film_convert_depth_half_rgba_param_0+28]; ld.param.u32 %r9, [kernel_gpu_film_convert_depth_half_rgba_param_0+12]; ld.param.u32 %r8, [kernel_gpu_film_convert_depth_half_rgba_param_0+8]; ld.param.u32 %r7, [kernel_gpu_film_convert_depth_half_rgba_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_depth_half_rgba_param_0]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p1, %r1, %r32; @%p1 bra $L__BB232_9; div.s32 %r3, %r1, %r27; mul.lo.s32 %r36, %r3, %r27; sub.s32 %r4, %r1, %r36; mad.lo.s32 %r37, %r3, %r29, %r4; mul.wide.s32 %rd4, %r37, %r7; cvt.s64.s32 %rd5, %r28; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB232_5; setp.eq.s32 %p3, %r9, 0; mov.f32 %f29, 0f3F800000; @%p3 bra $L__BB232_4; cvt.s64.s32 %rd6, %r13; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r38, [%rd10]; cvt.rn.f32.u32 %f15, %r38; rcp.approx.ftz.f32 %f29, %f15; $L__BB232_4: mul.ftz.f32 %f16, %f12, %f29; setp.eq.s32 %p4, %r8, 0; selp.f32 %f30, %f29, %f16, %p4; $L__BB232_5: cvt.s64.s32 %rd11, %r6; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f17, [%rd15]; setp.eq.ftz.f32 %p5, %f17, 0f00000000; mul.ftz.f32 %f18, %f30, %f17; selp.f32 %f31, 0f501502F9, %f18, %p5; setp.eq.s32 %p6, %r14, -1; setp.eq.s32 %p7, %r22, 0; or.pred %p8, %p7, %p6; mov.f32 %f32, %f31; @%p8 bra $L__BB232_8; add.s32 %r39, %r14, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f19, [%rd20]; setp.neu.ftz.f32 %p9, %f19, 0f00000000; mov.f32 %f32, %f31; @%p9 bra $L__BB232_8; mov.f32 %f20, 0f3F800000; sub.ftz.f32 %f21, %f20, %f31; mov.f32 %f22, 0f00000000; sub.ftz.f32 %f23, %f22, %f31; fma.rn.ftz.f32 %f32, %f21, 0f3F000000, %f31; fma.rn.ftz.f32 %f31, %f23, 0f3F000000, %f31; $L__BB232_8: mov.f32 %f28, 0f477FE000; min.ftz.f32 %f24, %f32, %f28; // begin inline asm { cvt.rn.f16.f32 %rs1, %f24;} // end inline asm min.ftz.f32 %f26, %f31, %f28; // begin inline asm { cvt.rn.f16.f32 %rs2, %f26;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs3, %f26;} // end inline asm mov.f32 %f27, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs4, %f27;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r31; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r30; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB232_9: ret; } // .globl kernel_gpu_film_convert_mist .visible .entry kernel_gpu_film_convert_mist( .param .align 4 .b8 kernel_gpu_film_convert_mist_param_0[96], .param .u64 kernel_gpu_film_convert_mist_param_1, .param .u64 kernel_gpu_film_convert_mist_param_2, .param .u32 kernel_gpu_film_convert_mist_param_3, .param .u32 kernel_gpu_film_convert_mist_param_4, .param .u32 kernel_gpu_film_convert_mist_param_5, .param .u32 kernel_gpu_film_convert_mist_param_6, .param .u32 kernel_gpu_film_convert_mist_param_7, .param .u32 kernel_gpu_film_convert_mist_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<5>; .reg .f32 %f<19>; .reg .b32 %r<39>; .reg .b64 %rd<19>; ld.param.u64 %rd2, [kernel_gpu_film_convert_mist_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_mist_param_2]; ld.param.u32 %r28, [kernel_gpu_film_convert_mist_param_3]; ld.param.u32 %r24, [kernel_gpu_film_convert_mist_param_4]; ld.param.u32 %r25, [kernel_gpu_film_convert_mist_param_5]; ld.param.u32 %r26, [kernel_gpu_film_convert_mist_param_6]; ld.param.u32 %r27, [kernel_gpu_film_convert_mist_param_7]; ld.param.u32 %r21, [kernel_gpu_film_convert_mist_param_0+84]; ld.param.f32 %f18, [kernel_gpu_film_convert_mist_param_0+64]; ld.param.f32 %f7, [kernel_gpu_film_convert_mist_param_0+60]; ld.param.u32 %r10, [kernel_gpu_film_convert_mist_param_0+28]; ld.param.u32 %r6, [kernel_gpu_film_convert_mist_param_0+12]; ld.param.u32 %r5, [kernel_gpu_film_convert_mist_param_0+8]; ld.param.u32 %r4, [kernel_gpu_film_convert_mist_param_0+4]; ld.param.u32 %r3, [kernel_gpu_film_convert_mist_param_0]; mov.u32 %r29, %ntid.x; mov.u32 %r30, %ctaid.x; mov.u32 %r31, %tid.x; mad.lo.s32 %r1, %r30, %r29, %r31; setp.ge.s32 %p1, %r1, %r28; @%p1 bra $L__BB233_6; div.s32 %r32, %r1, %r24; mul.lo.s32 %r33, %r32, %r24; sub.s32 %r34, %r1, %r33; mad.lo.s32 %r35, %r32, %r26, %r34; mul.wide.s32 %rd4, %r35, %r4; cvt.s64.s32 %rd5, %r25; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r10, -1; @%p2 bra $L__BB233_5; setp.eq.s32 %p3, %r6, 0; mov.f32 %f17, 0f3F800000; @%p3 bra $L__BB233_4; cvt.s64.s32 %rd6, %r10; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r36, [%rd10]; cvt.rn.f32.u32 %f10, %r36; rcp.approx.ftz.f32 %f17, %f10; $L__BB233_4: mul.ftz.f32 %f11, %f7, %f17; setp.eq.s32 %p4, %r5, 0; selp.f32 %f18, %f17, %f11, %p4; $L__BB233_5: cvt.s64.s32 %rd11, %r3; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f12, [%rd15]; mul.ftz.f32 %f13, %f18, %f12; mov.f32 %f14, 0f3F800000; sub.ftz.f32 %f15, %f14, %f13; cvt.ftz.sat.f32.f32 %f16, %f15; cvta.to.global.u64 %rd16, %rd2; add.s32 %r37, %r1, %r27; mul.lo.s32 %r38, %r37, %r21; mul.wide.s32 %rd17, %r38, 4; add.s64 %rd18, %rd16, %rd17; st.global.f32 [%rd18], %f16; $L__BB233_6: ret; } // .globl kernel_gpu_film_convert_mist_half_rgba .visible .entry kernel_gpu_film_convert_mist_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_mist_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_mist_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_mist_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_mist_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<9>; .reg .b16 %rs<5>; .reg .f32 %f<35>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_mist_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_mist_half_rgba_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_mist_half_rgba_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_mist_half_rgba_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_mist_half_rgba_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_mist_half_rgba_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_mist_half_rgba_param_7]; ld.param.u32 %r31, [kernel_gpu_film_convert_mist_half_rgba_param_8]; ld.param.u32 %r22, [kernel_gpu_film_convert_mist_half_rgba_param_0+76]; ld.param.f32 %f32, [kernel_gpu_film_convert_mist_half_rgba_param_0+64]; ld.param.f32 %f12, [kernel_gpu_film_convert_mist_half_rgba_param_0+60]; ld.param.u32 %r14, [kernel_gpu_film_convert_mist_half_rgba_param_0+32]; ld.param.u32 %r13, [kernel_gpu_film_convert_mist_half_rgba_param_0+28]; ld.param.u32 %r9, [kernel_gpu_film_convert_mist_half_rgba_param_0+12]; ld.param.u32 %r8, [kernel_gpu_film_convert_mist_half_rgba_param_0+8]; ld.param.u32 %r7, [kernel_gpu_film_convert_mist_half_rgba_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_mist_half_rgba_param_0]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p1, %r1, %r32; @%p1 bra $L__BB234_9; div.s32 %r3, %r1, %r27; mul.lo.s32 %r36, %r3, %r27; sub.s32 %r4, %r1, %r36; mad.lo.s32 %r37, %r3, %r29, %r4; mul.wide.s32 %rd4, %r37, %r7; cvt.s64.s32 %rd5, %r28; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB234_5; setp.eq.s32 %p3, %r9, 0; mov.f32 %f31, 0f3F800000; @%p3 bra $L__BB234_4; cvt.s64.s32 %rd6, %r13; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r38, [%rd10]; cvt.rn.f32.u32 %f15, %r38; rcp.approx.ftz.f32 %f31, %f15; $L__BB234_4: mul.ftz.f32 %f16, %f12, %f31; setp.eq.s32 %p4, %r8, 0; selp.f32 %f32, %f31, %f16, %p4; $L__BB234_5: cvt.s64.s32 %rd11, %r6; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f17, [%rd15]; mul.ftz.f32 %f18, %f32, %f17; mov.f32 %f19, 0f3F800000; sub.ftz.f32 %f20, %f19, %f18; cvt.ftz.sat.f32.f32 %f33, %f20; setp.eq.s32 %p5, %r14, -1; setp.eq.s32 %p6, %r22, 0; or.pred %p7, %p6, %p5; mov.f32 %f34, %f33; @%p7 bra $L__BB234_8; add.s32 %r39, %r14, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f21, [%rd20]; setp.neu.ftz.f32 %p8, %f21, 0f00000000; mov.f32 %f34, %f33; @%p8 bra $L__BB234_8; sub.ftz.f32 %f23, %f19, %f33; mov.f32 %f24, 0f00000000; sub.ftz.f32 %f25, %f24, %f33; fma.rn.ftz.f32 %f34, %f23, 0f3F000000, %f33; fma.rn.ftz.f32 %f33, %f25, 0f3F000000, %f33; $L__BB234_8: mov.f32 %f30, 0f477FE000; min.ftz.f32 %f26, %f34, %f30; // begin inline asm { cvt.rn.f16.f32 %rs1, %f26;} // end inline asm min.ftz.f32 %f28, %f33, %f30; // begin inline asm { cvt.rn.f16.f32 %rs2, %f28;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs3, %f28;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs4, %f19;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r31; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r30; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB234_9: ret; } // .globl kernel_gpu_film_convert_sample_count .visible .entry kernel_gpu_film_convert_sample_count( .param .align 4 .b8 kernel_gpu_film_convert_sample_count_param_0[96], .param .u64 kernel_gpu_film_convert_sample_count_param_1, .param .u64 kernel_gpu_film_convert_sample_count_param_2, .param .u32 kernel_gpu_film_convert_sample_count_param_3, .param .u32 kernel_gpu_film_convert_sample_count_param_4, .param .u32 kernel_gpu_film_convert_sample_count_param_5, .param .u32 kernel_gpu_film_convert_sample_count_param_6, .param .u32 kernel_gpu_film_convert_sample_count_param_7, .param .u32 kernel_gpu_film_convert_sample_count_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<2>; .reg .f32 %f<6>; .reg .b32 %r<38>; .reg .b64 %rd<14>; ld.param.u64 %rd1, [kernel_gpu_film_convert_sample_count_param_1]; ld.param.u64 %rd2, [kernel_gpu_film_convert_sample_count_param_2]; ld.param.u32 %r27, [kernel_gpu_film_convert_sample_count_param_3]; ld.param.u32 %r23, [kernel_gpu_film_convert_sample_count_param_4]; ld.param.u32 %r24, [kernel_gpu_film_convert_sample_count_param_5]; ld.param.u32 %r25, [kernel_gpu_film_convert_sample_count_param_6]; ld.param.u32 %r26, [kernel_gpu_film_convert_sample_count_param_7]; ld.param.u32 %r20, [kernel_gpu_film_convert_sample_count_param_0+84]; ld.param.f32 %f1, [kernel_gpu_film_convert_sample_count_param_0+56]; ld.param.u32 %r3, [kernel_gpu_film_convert_sample_count_param_0+4]; ld.param.u32 %r2, [kernel_gpu_film_convert_sample_count_param_0]; mov.u32 %r28, %ctaid.x; mov.u32 %r29, %ntid.x; mov.u32 %r30, %tid.x; mad.lo.s32 %r1, %r28, %r29, %r30; setp.ge.s32 %p1, %r1, %r27; @%p1 bra $L__BB235_2; div.s32 %r31, %r1, %r23; mul.lo.s32 %r32, %r31, %r23; sub.s32 %r33, %r1, %r32; mad.lo.s32 %r34, %r31, %r25, %r33; mul.wide.s32 %rd3, %r34, %r3; add.s32 %r35, %r1, %r26; mul.lo.s32 %r36, %r35, %r20; cvta.to.global.u64 %rd4, %rd1; mul.wide.s32 %rd5, %r36, 4; add.s64 %rd6, %rd4, %rd5; cvt.s64.s32 %rd7, %r2; cvt.s64.s32 %rd8, %r24; add.s64 %rd9, %rd8, %rd7; add.s64 %rd10, %rd9, %rd3; cvta.to.global.u64 %rd11, %rd2; shl.b64 %rd12, %rd10, 2; add.s64 %rd13, %rd11, %rd12; ld.global.nc.u32 %r37, [%rd13]; cvt.rn.f32.u32 %f4, %r37; mul.ftz.f32 %f5, %f1, %f4; st.global.f32 [%rd6], %f5; $L__BB235_2: ret; } // .globl kernel_gpu_film_convert_sample_count_half_rgba .visible .entry kernel_gpu_film_convert_sample_count_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_sample_count_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_sample_count_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_sample_count_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_sample_count_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<6>; .reg .b16 %rs<5>; .reg .f32 %f<22>; .reg .b32 %r<40>; .reg .b64 %rd<24>; ld.param.u64 %rd2, [kernel_gpu_film_convert_sample_count_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_sample_count_half_rgba_param_2]; ld.param.u32 %r31, [kernel_gpu_film_convert_sample_count_half_rgba_param_3]; ld.param.u32 %r26, [kernel_gpu_film_convert_sample_count_half_rgba_param_4]; ld.param.u32 %r27, [kernel_gpu_film_convert_sample_count_half_rgba_param_5]; ld.param.u32 %r28, [kernel_gpu_film_convert_sample_count_half_rgba_param_6]; ld.param.u32 %r29, [kernel_gpu_film_convert_sample_count_half_rgba_param_7]; ld.param.u32 %r30, [kernel_gpu_film_convert_sample_count_half_rgba_param_8]; ld.param.u32 %r21, [kernel_gpu_film_convert_sample_count_half_rgba_param_0+76]; ld.param.f32 %f6, [kernel_gpu_film_convert_sample_count_half_rgba_param_0+56]; ld.param.u32 %r13, [kernel_gpu_film_convert_sample_count_half_rgba_param_0+32]; ld.param.u32 %r6, [kernel_gpu_film_convert_sample_count_half_rgba_param_0+4]; ld.param.u32 %r5, [kernel_gpu_film_convert_sample_count_half_rgba_param_0]; mov.u32 %r32, %ntid.x; mov.u32 %r33, %ctaid.x; mov.u32 %r34, %tid.x; mad.lo.s32 %r1, %r33, %r32, %r34; setp.ge.s32 %p1, %r1, %r31; @%p1 bra $L__BB236_5; div.s32 %r3, %r1, %r26; mul.lo.s32 %r35, %r3, %r26; sub.s32 %r4, %r1, %r35; mad.lo.s32 %r36, %r3, %r28, %r4; mul.wide.s32 %rd4, %r36, %r6; cvt.s64.s32 %rd5, %r27; add.s64 %rd1, %rd4, %rd5; cvt.s64.s32 %rd6, %r5; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r37, [%rd10]; cvt.rn.f32.u32 %f9, %r37; mul.ftz.f32 %f20, %f6, %f9; setp.eq.s32 %p2, %r21, 0; setp.eq.s32 %p3, %r13, -1; or.pred %p4, %p2, %p3; mov.f32 %f21, %f20; @%p4 bra $L__BB236_4; add.s32 %r38, %r13, 3; cvt.s64.s32 %rd11, %r38; add.s64 %rd12, %rd1, %rd11; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd8, %rd14; ld.global.nc.f32 %f10, [%rd15]; setp.neu.ftz.f32 %p5, %f10, 0f00000000; mov.f32 %f21, %f20; @%p5 bra $L__BB236_4; mov.f32 %f11, 0f3F800000; sub.ftz.f32 %f12, %f11, %f20; mov.f32 %f13, 0f00000000; sub.ftz.f32 %f14, %f13, %f20; fma.rn.ftz.f32 %f21, %f12, 0f3F000000, %f20; fma.rn.ftz.f32 %f20, %f14, 0f3F000000, %f20; $L__BB236_4: mov.f32 %f19, 0f477FE000; min.ftz.f32 %f15, %f21, %f19; // begin inline asm { cvt.rn.f16.f32 %rs1, %f15;} // end inline asm min.ftz.f32 %f17, %f20, %f19; // begin inline asm { cvt.rn.f16.f32 %rs2, %f17;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs3, %f17;} // end inline asm mov.f32 %f18, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs4, %f18;} // end inline asm cvta.to.global.u64 %rd16, %rd2; mul.lo.s32 %r39, %r3, %r30; cvt.s64.s32 %rd17, %r39; cvt.s64.s32 %rd18, %r4; cvt.s64.s32 %rd19, %r29; add.s64 %rd20, %rd18, %rd19; add.s64 %rd21, %rd20, %rd17; shl.b64 %rd22, %rd21, 3; add.s64 %rd23, %rd16, %rd22; st.global.u16 [%rd23], %rs1; st.global.u16 [%rd23+2], %rs2; st.global.u16 [%rd23+4], %rs3; st.global.u16 [%rd23+6], %rs4; $L__BB236_5: ret; } // .globl kernel_gpu_film_convert_float .visible .entry kernel_gpu_film_convert_float( .param .align 4 .b8 kernel_gpu_film_convert_float_param_0[96], .param .u64 kernel_gpu_film_convert_float_param_1, .param .u64 kernel_gpu_film_convert_float_param_2, .param .u32 kernel_gpu_film_convert_float_param_3, .param .u32 kernel_gpu_film_convert_float_param_4, .param .u32 kernel_gpu_film_convert_float_param_5, .param .u32 kernel_gpu_film_convert_float_param_6, .param .u32 kernel_gpu_film_convert_float_param_7, .param .u32 kernel_gpu_film_convert_float_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<5>; .reg .f32 %f<16>; .reg .b32 %r<39>; .reg .b64 %rd<19>; ld.param.u64 %rd2, [kernel_gpu_film_convert_float_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_float_param_2]; ld.param.u32 %r28, [kernel_gpu_film_convert_float_param_3]; ld.param.u32 %r24, [kernel_gpu_film_convert_float_param_4]; ld.param.u32 %r25, [kernel_gpu_film_convert_float_param_5]; ld.param.u32 %r26, [kernel_gpu_film_convert_float_param_6]; ld.param.u32 %r27, [kernel_gpu_film_convert_float_param_7]; ld.param.u32 %r21, [kernel_gpu_film_convert_float_param_0+84]; ld.param.f32 %f15, [kernel_gpu_film_convert_float_param_0+64]; ld.param.f32 %f7, [kernel_gpu_film_convert_float_param_0+60]; ld.param.u32 %r10, [kernel_gpu_film_convert_float_param_0+28]; ld.param.u32 %r6, [kernel_gpu_film_convert_float_param_0+12]; ld.param.u32 %r5, [kernel_gpu_film_convert_float_param_0+8]; ld.param.u32 %r4, [kernel_gpu_film_convert_float_param_0+4]; ld.param.u32 %r3, [kernel_gpu_film_convert_float_param_0]; mov.u32 %r29, %ntid.x; mov.u32 %r30, %ctaid.x; mov.u32 %r31, %tid.x; mad.lo.s32 %r1, %r30, %r29, %r31; setp.ge.s32 %p1, %r1, %r28; @%p1 bra $L__BB237_6; div.s32 %r32, %r1, %r24; mul.lo.s32 %r33, %r32, %r24; sub.s32 %r34, %r1, %r33; mad.lo.s32 %r35, %r32, %r26, %r34; mul.wide.s32 %rd4, %r35, %r4; cvt.s64.s32 %rd5, %r25; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r10, -1; @%p2 bra $L__BB237_5; setp.eq.s32 %p3, %r6, 0; mov.f32 %f14, 0f3F800000; @%p3 bra $L__BB237_4; cvt.s64.s32 %rd6, %r10; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r36, [%rd10]; cvt.rn.f32.u32 %f10, %r36; rcp.approx.ftz.f32 %f14, %f10; $L__BB237_4: mul.ftz.f32 %f11, %f7, %f14; setp.eq.s32 %p4, %r5, 0; selp.f32 %f15, %f14, %f11, %p4; $L__BB237_5: cvt.s64.s32 %rd11, %r3; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f12, [%rd15]; mul.ftz.f32 %f13, %f15, %f12; cvta.to.global.u64 %rd16, %rd2; add.s32 %r37, %r1, %r27; mul.lo.s32 %r38, %r37, %r21; mul.wide.s32 %rd17, %r38, 4; add.s64 %rd18, %rd16, %rd17; st.global.f32 [%rd18], %f13; $L__BB237_6: ret; } // .globl kernel_gpu_film_convert_float_half_rgba .visible .entry kernel_gpu_film_convert_float_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_float_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_float_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_float_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_float_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<9>; .reg .b16 %rs<5>; .reg .f32 %f<32>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_float_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_float_half_rgba_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_float_half_rgba_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_float_half_rgba_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_float_half_rgba_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_float_half_rgba_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_float_half_rgba_param_7]; ld.param.u32 %r31, [kernel_gpu_film_convert_float_half_rgba_param_8]; ld.param.u32 %r22, [kernel_gpu_film_convert_float_half_rgba_param_0+76]; ld.param.f32 %f29, [kernel_gpu_film_convert_float_half_rgba_param_0+64]; ld.param.f32 %f12, [kernel_gpu_film_convert_float_half_rgba_param_0+60]; ld.param.u32 %r14, [kernel_gpu_film_convert_float_half_rgba_param_0+32]; ld.param.u32 %r13, [kernel_gpu_film_convert_float_half_rgba_param_0+28]; ld.param.u32 %r9, [kernel_gpu_film_convert_float_half_rgba_param_0+12]; ld.param.u32 %r8, [kernel_gpu_film_convert_float_half_rgba_param_0+8]; ld.param.u32 %r7, [kernel_gpu_film_convert_float_half_rgba_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_float_half_rgba_param_0]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p1, %r1, %r32; @%p1 bra $L__BB238_9; div.s32 %r3, %r1, %r27; mul.lo.s32 %r36, %r3, %r27; sub.s32 %r4, %r1, %r36; mad.lo.s32 %r37, %r3, %r29, %r4; mul.wide.s32 %rd4, %r37, %r7; cvt.s64.s32 %rd5, %r28; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB238_5; setp.eq.s32 %p3, %r9, 0; mov.f32 %f28, 0f3F800000; @%p3 bra $L__BB238_4; cvt.s64.s32 %rd6, %r13; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r38, [%rd10]; cvt.rn.f32.u32 %f15, %r38; rcp.approx.ftz.f32 %f28, %f15; $L__BB238_4: mul.ftz.f32 %f16, %f12, %f28; setp.eq.s32 %p4, %r8, 0; selp.f32 %f29, %f28, %f16, %p4; $L__BB238_5: cvt.s64.s32 %rd11, %r6; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f17, [%rd15]; mul.ftz.f32 %f30, %f29, %f17; setp.eq.s32 %p5, %r14, -1; setp.eq.s32 %p6, %r22, 0; or.pred %p7, %p6, %p5; mov.f32 %f31, %f30; @%p7 bra $L__BB238_8; add.s32 %r39, %r14, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f18, [%rd20]; setp.neu.ftz.f32 %p8, %f18, 0f00000000; mov.f32 %f31, %f30; @%p8 bra $L__BB238_8; mov.f32 %f19, 0f3F800000; sub.ftz.f32 %f20, %f19, %f30; mov.f32 %f21, 0f00000000; sub.ftz.f32 %f22, %f21, %f30; fma.rn.ftz.f32 %f31, %f20, 0f3F000000, %f30; fma.rn.ftz.f32 %f30, %f22, 0f3F000000, %f30; $L__BB238_8: mov.f32 %f27, 0f477FE000; min.ftz.f32 %f23, %f31, %f27; // begin inline asm { cvt.rn.f16.f32 %rs1, %f23;} // end inline asm min.ftz.f32 %f25, %f30, %f27; // begin inline asm { cvt.rn.f16.f32 %rs2, %f25;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs3, %f25;} // end inline asm mov.f32 %f26, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs4, %f26;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r31; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r30; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB238_9: ret; } // .globl kernel_gpu_film_convert_light_path .visible .entry kernel_gpu_film_convert_light_path( .param .align 4 .b8 kernel_gpu_film_convert_light_path_param_0[96], .param .u64 kernel_gpu_film_convert_light_path_param_1, .param .u64 kernel_gpu_film_convert_light_path_param_2, .param .u32 kernel_gpu_film_convert_light_path_param_3, .param .u32 kernel_gpu_film_convert_light_path_param_4, .param .u32 kernel_gpu_film_convert_light_path_param_5, .param .u32 kernel_gpu_film_convert_light_path_param_6, .param .u32 kernel_gpu_film_convert_light_path_param_7, .param .u32 kernel_gpu_film_convert_light_path_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<21>; .reg .f32 %f<86>; .reg .b32 %r<45>; .reg .b64 %rd<34>; ld.param.u64 %rd4, [kernel_gpu_film_convert_light_path_param_1]; ld.param.u64 %rd5, [kernel_gpu_film_convert_light_path_param_2]; ld.param.u32 %r33, [kernel_gpu_film_convert_light_path_param_3]; ld.param.u32 %r29, [kernel_gpu_film_convert_light_path_param_4]; ld.param.u32 %r30, [kernel_gpu_film_convert_light_path_param_5]; ld.param.u32 %r31, [kernel_gpu_film_convert_light_path_param_6]; ld.param.u32 %r32, [kernel_gpu_film_convert_light_path_param_7]; ld.param.u32 %r26, [kernel_gpu_film_convert_light_path_param_0+84]; ld.param.u32 %r25, [kernel_gpu_film_convert_light_path_param_0+80]; ld.param.f32 %f81, [kernel_gpu_film_convert_light_path_param_0+64]; ld.param.f32 %f46, [kernel_gpu_film_convert_light_path_param_0+60]; ld.param.f32 %f85, [kernel_gpu_film_convert_light_path_param_0+56]; ld.param.u32 %r15, [kernel_gpu_film_convert_light_path_param_0+28]; ld.param.u32 %r14, [kernel_gpu_film_convert_light_path_param_0+24]; ld.param.u32 %r13, [kernel_gpu_film_convert_light_path_param_0+20]; ld.param.u32 %r12, [kernel_gpu_film_convert_light_path_param_0+16]; ld.param.u32 %r11, [kernel_gpu_film_convert_light_path_param_0+12]; ld.param.u32 %r10, [kernel_gpu_film_convert_light_path_param_0+8]; ld.param.u32 %r9, [kernel_gpu_film_convert_light_path_param_0+4]; ld.param.u32 %r8, [kernel_gpu_film_convert_light_path_param_0]; mov.u32 %r34, %ntid.x; mov.u32 %r35, %ctaid.x; mov.u32 %r36, %tid.x; mad.lo.s32 %r1, %r35, %r34, %r36; setp.ge.s32 %p1, %r1, %r33; @%p1 bra $L__BB239_30; div.s32 %r37, %r1, %r29; mul.lo.s32 %r38, %r37, %r29; sub.s32 %r39, %r1, %r38; mad.lo.s32 %r40, %r37, %r31, %r39; mul.wide.s32 %rd6, %r40, %r9; cvt.s64.s32 %rd7, %r30; add.s64 %rd1, %rd6, %rd7; cvt.s64.s32 %rd8, %r8; add.s64 %rd9, %rd1, %rd8; cvta.to.global.u64 %rd10, %rd5; shl.b64 %rd11, %rd9, 2; add.s64 %rd12, %rd10, %rd11; ld.global.nc.f32 %f71, [%rd12]; ld.global.nc.f32 %f72, [%rd12+4]; ld.global.nc.f32 %f73, [%rd12+8]; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB239_3; cvt.s64.s32 %rd13, %r13; add.s64 %rd14, %rd1, %rd13; shl.b64 %rd16, %rd14, 2; add.s64 %rd17, %rd10, %rd16; ld.global.nc.f32 %f48, [%rd17]; add.ftz.f32 %f71, %f71, %f48; ld.global.nc.f32 %f49, [%rd17+4]; add.ftz.f32 %f72, %f72, %f49; ld.global.nc.f32 %f50, [%rd17+8]; add.ftz.f32 %f73, %f73, %f50; $L__BB239_3: cvt.s64.s32 %rd18, %r15; add.s64 %rd19, %rd1, %rd18; shl.b64 %rd21, %rd19, 2; add.s64 %rd2, %rd10, %rd21; setp.eq.s32 %p3, %r12, -1; @%p3 bra $L__BB239_17; cvt.s64.s32 %rd22, %r12; add.s64 %rd23, %rd1, %rd22; shl.b64 %rd25, %rd23, 2; add.s64 %rd26, %rd10, %rd25; ld.global.nc.f32 %f11, [%rd26+4]; ld.global.nc.f32 %f12, [%rd26+8]; ld.global.nc.f32 %f13, [%rd26]; setp.eq.ftz.f32 %p4, %f13, 0f00000000; mov.f32 %f77, 0f00000000; mov.f32 %f74, %f77; @%p4 bra $L__BB239_6; div.approx.ftz.f32 %f74, %f71, %f13; $L__BB239_6: setp.eq.ftz.f32 %p5, %f11, 0f00000000; mov.f32 %f75, %f77; @%p5 bra $L__BB239_8; div.approx.ftz.f32 %f75, %f72, %f11; $L__BB239_8: setp.eq.ftz.f32 %p6, %f12, 0f00000000; @%p6 bra $L__BB239_10; div.approx.ftz.f32 %f77, %f73, %f12; $L__BB239_10: @%p4 bra $L__BB239_14; bra.uni $L__BB239_11; $L__BB239_14: mov.f32 %f78, %f77; mov.f32 %f79, %f77; @%p5 bra $L__BB239_16; add.ftz.f32 %f58, %f75, %f77; mul.ftz.f32 %f59, %f58, 0f3F000000; selp.f32 %f23, %f75, %f59, %p6; selp.f32 %f79, %f75, %f77, %p6; mov.f32 %f77, %f23; mov.f32 %f78, %f75; bra.uni $L__BB239_16; $L__BB239_17: setp.eq.s32 %p13, %r15, -1; @%p13 bra $L__BB239_21; setp.eq.s32 %p14, %r11, 0; mov.f32 %f80, 0f3F800000; @%p14 bra $L__BB239_20; ld.global.nc.u32 %r41, [%rd2]; cvt.rn.f32.u32 %f61, %r41; rcp.approx.ftz.f32 %f80, %f61; $L__BB239_20: mul.ftz.f32 %f62, %f46, %f80; setp.eq.s32 %p15, %r10, 0; selp.f32 %f81, %f80, %f62, %p15; $L__BB239_21: mul.ftz.f32 %f82, %f71, %f81; mul.ftz.f32 %f83, %f72, %f81; mul.ftz.f32 %f84, %f73, %f81; bra.uni $L__BB239_22; $L__BB239_11: @%p5 bra $L__BB239_13; bra.uni $L__BB239_12; $L__BB239_13: add.ftz.f32 %f56, %f74, %f77; mul.ftz.f32 %f57, %f56, 0f3F000000; selp.f32 %f78, %f74, %f57, %p6; selp.f32 %f79, %f74, %f77, %p6; mov.f32 %f77, %f74; bra.uni $L__BB239_16; $L__BB239_12: add.ftz.f32 %f54, %f74, %f75; mul.ftz.f32 %f55, %f54, 0f3F000000; selp.f32 %f79, %f55, %f77, %p6; mov.f32 %f77, %f74; mov.f32 %f78, %f75; $L__BB239_16: mul.ftz.f32 %f82, %f46, %f77; mul.ftz.f32 %f83, %f46, %f78; mul.ftz.f32 %f84, %f46, %f79; $L__BB239_22: add.s32 %r42, %r1, %r32; mul.lo.s32 %r43, %r42, %r26; cvta.to.global.u64 %rd27, %rd4; mul.wide.s32 %rd28, %r43, 4; add.s64 %rd3, %rd27, %rd28; st.global.f32 [%rd3], %f82; st.global.f32 [%rd3+4], %f83; st.global.f32 [%rd3+8], %f84; setp.lt.s32 %p16, %r25, 4; @%p16 bra $L__BB239_30; setp.eq.s32 %p17, %r14, -1; @%p17 bra $L__BB239_29; setp.eq.s32 %p18, %r15, -1; @%p18 bra $L__BB239_28; ld.global.nc.u32 %r7, [%rd2]; setp.eq.s32 %p19, %r7, 0; mov.f32 %f85, 0f00000000; @%p19 bra $L__BB239_28; setp.eq.s32 %p20, %r11, 0; mov.f32 %f85, 0f3F800000; @%p20 bra $L__BB239_28; cvt.rn.f32.u32 %f65, %r7; rcp.approx.ftz.f32 %f85, %f65; $L__BB239_28: cvt.s64.s32 %rd29, %r14; add.s64 %rd31, %rd1, %rd29; shl.b64 %rd32, %rd31, 2; add.s64 %rd33, %rd10, %rd32; ld.global.nc.f32 %f66, [%rd33+12]; mul.ftz.f32 %f67, %f85, %f66; mov.f32 %f68, 0f3F800000; sub.ftz.f32 %f69, %f68, %f67; cvt.ftz.sat.f32.f32 %f70, %f69; st.global.f32 [%rd3+12], %f70; bra.uni $L__BB239_30; $L__BB239_29: mov.u32 %r44, 1065353216; st.global.u32 [%rd3+12], %r44; $L__BB239_30: ret; } // .globl kernel_gpu_film_convert_light_path_half_rgba .visible .entry kernel_gpu_film_convert_light_path_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_light_path_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_light_path_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_light_path_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_light_path_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<20>; .reg .b16 %rs<5>; .reg .f32 %f<94>; .reg .b32 %r<43>; .reg .b64 %rd<39>; ld.param.u64 %rd2, [kernel_gpu_film_convert_light_path_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_light_path_half_rgba_param_2]; ld.param.u32 %r34, [kernel_gpu_film_convert_light_path_half_rgba_param_3]; ld.param.u32 %r29, [kernel_gpu_film_convert_light_path_half_rgba_param_4]; ld.param.u32 %r30, [kernel_gpu_film_convert_light_path_half_rgba_param_5]; ld.param.u32 %r31, [kernel_gpu_film_convert_light_path_half_rgba_param_6]; ld.param.u32 %r32, [kernel_gpu_film_convert_light_path_half_rgba_param_7]; ld.param.u32 %r33, [kernel_gpu_film_convert_light_path_half_rgba_param_8]; ld.param.u32 %r24, [kernel_gpu_film_convert_light_path_half_rgba_param_0+76]; ld.param.f32 %f87, [kernel_gpu_film_convert_light_path_half_rgba_param_0+64]; ld.param.f32 %f49, [kernel_gpu_film_convert_light_path_half_rgba_param_0+60]; ld.param.u32 %r16, [kernel_gpu_film_convert_light_path_half_rgba_param_0+32]; ld.param.u32 %r15, [kernel_gpu_film_convert_light_path_half_rgba_param_0+28]; ld.param.u32 %r13, [kernel_gpu_film_convert_light_path_half_rgba_param_0+20]; ld.param.u32 %r12, [kernel_gpu_film_convert_light_path_half_rgba_param_0+16]; ld.param.u32 %r11, [kernel_gpu_film_convert_light_path_half_rgba_param_0+12]; ld.param.u32 %r10, [kernel_gpu_film_convert_light_path_half_rgba_param_0+8]; ld.param.u32 %r9, [kernel_gpu_film_convert_light_path_half_rgba_param_0+4]; ld.param.u32 %r8, [kernel_gpu_film_convert_light_path_half_rgba_param_0]; mov.u32 %r35, %ntid.x; mov.u32 %r36, %ctaid.x; mov.u32 %r37, %tid.x; mad.lo.s32 %r1, %r36, %r35, %r37; setp.ge.s32 %p1, %r1, %r34; @%p1 bra $L__BB240_26; div.s32 %r3, %r1, %r29; mul.lo.s32 %r38, %r3, %r29; sub.s32 %r4, %r1, %r38; mad.lo.s32 %r39, %r3, %r31, %r4; mul.wide.s32 %rd4, %r39, %r9; cvt.s64.s32 %rd5, %r30; add.s64 %rd1, %rd4, %rd5; cvt.s64.s32 %rd6, %r8; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.f32 %f77, [%rd10]; ld.global.nc.f32 %f78, [%rd10+4]; ld.global.nc.f32 %f79, [%rd10+8]; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB240_3; cvt.s64.s32 %rd11, %r13; add.s64 %rd12, %rd1, %rd11; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd8, %rd14; ld.global.nc.f32 %f51, [%rd15]; add.ftz.f32 %f77, %f77, %f51; ld.global.nc.f32 %f52, [%rd15+4]; add.ftz.f32 %f78, %f78, %f52; ld.global.nc.f32 %f53, [%rd15+8]; add.ftz.f32 %f79, %f79, %f53; $L__BB240_3: setp.eq.s32 %p3, %r12, -1; @%p3 bra $L__BB240_17; cvt.s64.s32 %rd16, %r12; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd8, %rd19; ld.global.nc.f32 %f11, [%rd20+4]; ld.global.nc.f32 %f12, [%rd20+8]; ld.global.nc.f32 %f13, [%rd20]; setp.eq.ftz.f32 %p4, %f13, 0f00000000; mov.f32 %f83, 0f00000000; mov.f32 %f80, %f83; @%p4 bra $L__BB240_6; div.approx.ftz.f32 %f80, %f77, %f13; $L__BB240_6: setp.eq.ftz.f32 %p5, %f11, 0f00000000; mov.f32 %f81, %f83; @%p5 bra $L__BB240_8; div.approx.ftz.f32 %f81, %f78, %f11; $L__BB240_8: setp.eq.ftz.f32 %p6, %f12, 0f00000000; @%p6 bra $L__BB240_10; div.approx.ftz.f32 %f83, %f79, %f12; $L__BB240_10: @%p4 bra $L__BB240_14; bra.uni $L__BB240_11; $L__BB240_14: mov.f32 %f84, %f83; mov.f32 %f85, %f83; @%p5 bra $L__BB240_16; add.ftz.f32 %f61, %f81, %f83; mul.ftz.f32 %f62, %f61, 0f3F000000; selp.f32 %f23, %f81, %f62, %p6; selp.f32 %f85, %f81, %f83, %p6; mov.f32 %f83, %f23; mov.f32 %f84, %f81; bra.uni $L__BB240_16; $L__BB240_17: setp.eq.s32 %p13, %r15, -1; @%p13 bra $L__BB240_21; setp.eq.s32 %p14, %r11, 0; mov.f32 %f86, 0f3F800000; @%p14 bra $L__BB240_20; cvt.s64.s32 %rd21, %r15; add.s64 %rd22, %rd1, %rd21; shl.b64 %rd24, %rd22, 2; add.s64 %rd25, %rd8, %rd24; ld.global.nc.u32 %r40, [%rd25]; cvt.rn.f32.u32 %f64, %r40; rcp.approx.ftz.f32 %f86, %f64; $L__BB240_20: mul.ftz.f32 %f65, %f49, %f86; setp.eq.s32 %p15, %r10, 0; selp.f32 %f87, %f86, %f65, %p15; $L__BB240_21: mul.ftz.f32 %f93, %f77, %f87; mul.ftz.f32 %f92, %f78, %f87; mul.ftz.f32 %f91, %f79, %f87; bra.uni $L__BB240_22; $L__BB240_11: @%p5 bra $L__BB240_13; bra.uni $L__BB240_12; $L__BB240_13: add.ftz.f32 %f59, %f80, %f83; mul.ftz.f32 %f60, %f59, 0f3F000000; selp.f32 %f84, %f80, %f60, %p6; selp.f32 %f85, %f80, %f83, %p6; mov.f32 %f83, %f80; bra.uni $L__BB240_16; $L__BB240_12: add.ftz.f32 %f57, %f80, %f81; mul.ftz.f32 %f58, %f57, 0f3F000000; selp.f32 %f85, %f58, %f83, %p6; mov.f32 %f83, %f80; mov.f32 %f84, %f81; $L__BB240_16: mul.ftz.f32 %f93, %f49, %f83; mul.ftz.f32 %f92, %f49, %f84; mul.ftz.f32 %f91, %f49, %f85; $L__BB240_22: setp.eq.s32 %p16, %r16, -1; setp.eq.s32 %p17, %r24, 0; or.pred %p18, %p17, %p16; @%p18 bra $L__BB240_25; add.s32 %r41, %r16, 3; cvt.s64.s32 %rd26, %r41; add.s64 %rd27, %rd1, %rd26; shl.b64 %rd29, %rd27, 2; add.s64 %rd30, %rd8, %rd29; ld.global.nc.f32 %f66, [%rd30]; setp.neu.ftz.f32 %p19, %f66, 0f00000000; @%p19 bra $L__BB240_25; mov.f32 %f67, 0f3F800000; sub.ftz.f32 %f68, %f67, %f93; mov.f32 %f69, 0f00000000; sub.ftz.f32 %f70, %f69, %f92; sub.ftz.f32 %f71, %f69, %f91; fma.rn.ftz.f32 %f93, %f68, 0f3F000000, %f93; fma.rn.ftz.f32 %f92, %f70, 0f3F000000, %f92; fma.rn.ftz.f32 %f91, %f71, 0f3F000000, %f91; $L__BB240_25: mov.f32 %f76, 0f477FE000; min.ftz.f32 %f72, %f93, %f76; // begin inline asm { cvt.rn.f16.f32 %rs1, %f72;} // end inline asm min.ftz.f32 %f73, %f92, %f76; // begin inline asm { cvt.rn.f16.f32 %rs2, %f73;} // end inline asm min.ftz.f32 %f74, %f91, %f76; // begin inline asm { cvt.rn.f16.f32 %rs3, %f74;} // end inline asm mov.f32 %f75, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs4, %f75;} // end inline asm cvta.to.global.u64 %rd31, %rd2; mul.lo.s32 %r42, %r3, %r33; cvt.s64.s32 %rd32, %r42; cvt.s64.s32 %rd33, %r4; cvt.s64.s32 %rd34, %r32; add.s64 %rd35, %rd33, %rd34; add.s64 %rd36, %rd35, %rd32; shl.b64 %rd37, %rd36, 3; add.s64 %rd38, %rd31, %rd37; st.global.u16 [%rd38], %rs1; st.global.u16 [%rd38+2], %rs2; st.global.u16 [%rd38+4], %rs3; st.global.u16 [%rd38+6], %rs4; $L__BB240_26: ret; } // .globl kernel_gpu_film_convert_float3 .visible .entry kernel_gpu_film_convert_float3( .param .align 4 .b8 kernel_gpu_film_convert_float3_param_0[96], .param .u64 kernel_gpu_film_convert_float3_param_1, .param .u64 kernel_gpu_film_convert_float3_param_2, .param .u32 kernel_gpu_film_convert_float3_param_3, .param .u32 kernel_gpu_film_convert_float3_param_4, .param .u32 kernel_gpu_film_convert_float3_param_5, .param .u32 kernel_gpu_film_convert_float3_param_6, .param .u32 kernel_gpu_film_convert_float3_param_7, .param .u32 kernel_gpu_film_convert_float3_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<10>; .reg .f32 %f<32>; .reg .b32 %r<43>; .reg .b64 %rd<24>; ld.param.u64 %rd4, [kernel_gpu_film_convert_float3_param_1]; ld.param.u64 %rd5, [kernel_gpu_film_convert_float3_param_2]; ld.param.u32 %r31, [kernel_gpu_film_convert_float3_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_float3_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_float3_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_float3_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_float3_param_7]; ld.param.u32 %r24, [kernel_gpu_film_convert_float3_param_0+84]; ld.param.u32 %r23, [kernel_gpu_film_convert_float3_param_0+80]; ld.param.f32 %f30, [kernel_gpu_film_convert_float3_param_0+64]; ld.param.f32 %f10, [kernel_gpu_film_convert_float3_param_0+60]; ld.param.f32 %f31, [kernel_gpu_film_convert_float3_param_0+56]; ld.param.u32 %r13, [kernel_gpu_film_convert_float3_param_0+28]; ld.param.u32 %r12, [kernel_gpu_film_convert_float3_param_0+24]; ld.param.u32 %r9, [kernel_gpu_film_convert_float3_param_0+12]; ld.param.u32 %r8, [kernel_gpu_film_convert_float3_param_0+8]; ld.param.u32 %r7, [kernel_gpu_film_convert_float3_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_float3_param_0]; mov.u32 %r32, %ntid.x; mov.u32 %r33, %ctaid.x; mov.u32 %r34, %tid.x; mad.lo.s32 %r1, %r33, %r32, %r34; setp.ge.s32 %p1, %r1, %r31; @%p1 bra $L__BB241_13; div.s32 %r35, %r1, %r27; mul.lo.s32 %r36, %r35, %r27; sub.s32 %r37, %r1, %r36; mad.lo.s32 %r38, %r35, %r29, %r37; mul.wide.s32 %rd6, %r38, %r7; cvt.s64.s32 %rd7, %r28; add.s64 %rd1, %rd6, %rd7; cvt.s64.s32 %rd8, %r13; add.s64 %rd9, %rd1, %rd8; cvta.to.global.u64 %rd10, %rd5; shl.b64 %rd11, %rd9, 2; add.s64 %rd2, %rd10, %rd11; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB241_5; setp.eq.s32 %p3, %r9, 0; mov.f32 %f29, 0f3F800000; @%p3 bra $L__BB241_4; ld.global.nc.u32 %r39, [%rd2]; cvt.rn.f32.u32 %f13, %r39; rcp.approx.ftz.f32 %f29, %f13; $L__BB241_4: mul.ftz.f32 %f14, %f10, %f29; setp.eq.s32 %p4, %r8, 0; selp.f32 %f30, %f29, %f14, %p4; $L__BB241_5: cvta.to.global.u64 %rd12, %rd4; add.s32 %r40, %r1, %r30; mul.lo.s32 %r41, %r40, %r24; mul.wide.s32 %rd13, %r41, 4; add.s64 %rd3, %rd12, %rd13; cvt.s64.s32 %rd14, %r6; add.s64 %rd15, %rd1, %rd14; shl.b64 %rd17, %rd15, 2; add.s64 %rd18, %rd10, %rd17; ld.global.nc.f32 %f15, [%rd18]; mul.ftz.f32 %f16, %f30, %f15; ld.global.nc.f32 %f17, [%rd18+4]; mul.ftz.f32 %f18, %f30, %f17; ld.global.nc.f32 %f19, [%rd18+8]; mul.ftz.f32 %f20, %f30, %f19; st.global.f32 [%rd3], %f16; st.global.f32 [%rd3+4], %f18; st.global.f32 [%rd3+8], %f20; setp.lt.s32 %p5, %r23, 4; @%p5 bra $L__BB241_13; setp.eq.s32 %p6, %r12, -1; @%p6 bra $L__BB241_12; @%p2 bra $L__BB241_11; ld.global.nc.u32 %r5, [%rd2]; setp.eq.s32 %p8, %r5, 0; mov.f32 %f31, 0f00000000; @%p8 bra $L__BB241_11; setp.eq.s32 %p9, %r9, 0; mov.f32 %f31, 0f3F800000; @%p9 bra $L__BB241_11; cvt.rn.f32.u32 %f23, %r5; rcp.approx.ftz.f32 %f31, %f23; $L__BB241_11: cvt.s64.s32 %rd19, %r12; add.s64 %rd21, %rd1, %rd19; shl.b64 %rd22, %rd21, 2; add.s64 %rd23, %rd10, %rd22; ld.global.nc.f32 %f24, [%rd23+12]; mul.ftz.f32 %f25, %f31, %f24; mov.f32 %f26, 0f3F800000; sub.ftz.f32 %f27, %f26, %f25; cvt.ftz.sat.f32.f32 %f28, %f27; st.global.f32 [%rd3+12], %f28; bra.uni $L__BB241_13; $L__BB241_12: mov.u32 %r42, 1065353216; st.global.u32 [%rd3+12], %r42; $L__BB241_13: ret; } // .globl kernel_gpu_film_convert_float3_half_rgba .visible .entry kernel_gpu_film_convert_float3_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_float3_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_float3_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_float3_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_float3_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<9>; .reg .b16 %rs<5>; .reg .f32 %f<40>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_float3_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_float3_half_rgba_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_float3_half_rgba_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_float3_half_rgba_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_float3_half_rgba_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_float3_half_rgba_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_float3_half_rgba_param_7]; ld.param.u32 %r31, [kernel_gpu_film_convert_float3_half_rgba_param_8]; ld.param.u32 %r22, [kernel_gpu_film_convert_float3_half_rgba_param_0+76]; ld.param.f32 %f36, [kernel_gpu_film_convert_float3_half_rgba_param_0+64]; ld.param.f32 %f16, [kernel_gpu_film_convert_float3_half_rgba_param_0+60]; ld.param.u32 %r14, [kernel_gpu_film_convert_float3_half_rgba_param_0+32]; ld.param.u32 %r13, [kernel_gpu_film_convert_float3_half_rgba_param_0+28]; ld.param.u32 %r9, [kernel_gpu_film_convert_float3_half_rgba_param_0+12]; ld.param.u32 %r8, [kernel_gpu_film_convert_float3_half_rgba_param_0+8]; ld.param.u32 %r7, [kernel_gpu_film_convert_float3_half_rgba_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_float3_half_rgba_param_0]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p1, %r1, %r32; @%p1 bra $L__BB242_9; div.s32 %r3, %r1, %r27; mul.lo.s32 %r36, %r3, %r27; sub.s32 %r4, %r1, %r36; mad.lo.s32 %r37, %r3, %r29, %r4; mul.wide.s32 %rd4, %r37, %r7; cvt.s64.s32 %rd5, %r28; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB242_5; setp.eq.s32 %p3, %r9, 0; mov.f32 %f35, 0f3F800000; @%p3 bra $L__BB242_4; cvt.s64.s32 %rd6, %r13; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r38, [%rd10]; cvt.rn.f32.u32 %f19, %r38; rcp.approx.ftz.f32 %f35, %f19; $L__BB242_4: mul.ftz.f32 %f20, %f16, %f35; setp.eq.s32 %p4, %r8, 0; selp.f32 %f36, %f35, %f20, %p4; $L__BB242_5: cvt.s64.s32 %rd11, %r6; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f21, [%rd15]; mul.ftz.f32 %f39, %f36, %f21; ld.global.nc.f32 %f22, [%rd15+4]; mul.ftz.f32 %f38, %f36, %f22; ld.global.nc.f32 %f23, [%rd15+8]; mul.ftz.f32 %f37, %f36, %f23; setp.eq.s32 %p5, %r14, -1; setp.eq.s32 %p6, %r22, 0; or.pred %p7, %p6, %p5; @%p7 bra $L__BB242_8; add.s32 %r39, %r14, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f24, [%rd20]; setp.neu.ftz.f32 %p8, %f24, 0f00000000; @%p8 bra $L__BB242_8; mov.f32 %f25, 0f3F800000; sub.ftz.f32 %f26, %f25, %f39; mov.f32 %f27, 0f00000000; sub.ftz.f32 %f28, %f27, %f38; sub.ftz.f32 %f29, %f27, %f37; fma.rn.ftz.f32 %f39, %f26, 0f3F000000, %f39; fma.rn.ftz.f32 %f38, %f28, 0f3F000000, %f38; fma.rn.ftz.f32 %f37, %f29, 0f3F000000, %f37; $L__BB242_8: mov.f32 %f34, 0f477FE000; min.ftz.f32 %f30, %f39, %f34; // begin inline asm { cvt.rn.f16.f32 %rs1, %f30;} // end inline asm min.ftz.f32 %f31, %f38, %f34; // begin inline asm { cvt.rn.f16.f32 %rs2, %f31;} // end inline asm min.ftz.f32 %f32, %f37, %f34; // begin inline asm { cvt.rn.f16.f32 %rs3, %f32;} // end inline asm mov.f32 %f33, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs4, %f33;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r31; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r30; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB242_9: ret; } // .globl kernel_gpu_film_convert_motion .visible .entry kernel_gpu_film_convert_motion( .param .align 4 .b8 kernel_gpu_film_convert_motion_param_0[96], .param .u64 kernel_gpu_film_convert_motion_param_1, .param .u64 kernel_gpu_film_convert_motion_param_2, .param .u32 kernel_gpu_film_convert_motion_param_3, .param .u32 kernel_gpu_film_convert_motion_param_4, .param .u32 kernel_gpu_film_convert_motion_param_5, .param .u32 kernel_gpu_film_convert_motion_param_6, .param .u32 kernel_gpu_film_convert_motion_param_7, .param .u32 kernel_gpu_film_convert_motion_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<3>; .reg .f32 %f<17>; .reg .b32 %r<37>; .reg .b64 %rd<18>; ld.param.u64 %rd3, [kernel_gpu_film_convert_motion_param_1]; ld.param.u64 %rd4, [kernel_gpu_film_convert_motion_param_2]; ld.param.u32 %r27, [kernel_gpu_film_convert_motion_param_3]; ld.param.u32 %r23, [kernel_gpu_film_convert_motion_param_4]; ld.param.u32 %r24, [kernel_gpu_film_convert_motion_param_5]; ld.param.u32 %r25, [kernel_gpu_film_convert_motion_param_6]; ld.param.u32 %r26, [kernel_gpu_film_convert_motion_param_7]; ld.param.u32 %r20, [kernel_gpu_film_convert_motion_param_0+84]; ld.param.u32 %r11, [kernel_gpu_film_convert_motion_param_0+36]; ld.param.u32 %r3, [kernel_gpu_film_convert_motion_param_0+4]; ld.param.u32 %r2, [kernel_gpu_film_convert_motion_param_0]; mov.u32 %r28, %ctaid.x; mov.u32 %r29, %ntid.x; mov.u32 %r30, %tid.x; mad.lo.s32 %r1, %r28, %r29, %r30; setp.ge.s32 %p1, %r1, %r27; @%p1 bra $L__BB243_4; div.s32 %r31, %r1, %r23; mul.lo.s32 %r32, %r31, %r23; sub.s32 %r33, %r1, %r32; mad.lo.s32 %r34, %r31, %r25, %r33; mul.wide.s32 %rd5, %r34, %r3; cvt.s64.s32 %rd6, %r24; add.s64 %rd7, %rd5, %rd6; add.s32 %r35, %r1, %r26; mul.lo.s32 %r36, %r35, %r20; cvta.to.global.u64 %rd8, %rd3; mul.wide.s32 %rd9, %r36, 4; add.s64 %rd1, %rd8, %rd9; cvt.s64.s32 %rd10, %r2; add.s64 %rd11, %rd7, %rd10; cvta.to.global.u64 %rd12, %rd4; shl.b64 %rd13, %rd11, 2; add.s64 %rd2, %rd12, %rd13; cvt.s64.s32 %rd14, %r11; add.s64 %rd15, %rd7, %rd14; shl.b64 %rd16, %rd15, 2; add.s64 %rd17, %rd12, %rd16; ld.global.nc.f32 %f1, [%rd17]; setp.leu.ftz.f32 %p2, %f1, 0f00000000; mov.f32 %f16, 0f00000000; @%p2 bra $L__BB243_3; rcp.approx.ftz.f32 %f16, %f1; $L__BB243_3: ld.global.nc.f32 %f8, [%rd2]; mul.ftz.f32 %f9, %f16, %f8; ld.global.nc.f32 %f10, [%rd2+4]; mul.ftz.f32 %f11, %f16, %f10; ld.global.nc.f32 %f12, [%rd2+8]; mul.ftz.f32 %f13, %f16, %f12; ld.global.nc.f32 %f14, [%rd2+12]; mul.ftz.f32 %f15, %f16, %f14; st.global.f32 [%rd1], %f9; st.global.f32 [%rd1+4], %f11; st.global.f32 [%rd1+8], %f13; st.global.f32 [%rd1+12], %f15; $L__BB243_4: ret; } // .globl kernel_gpu_film_convert_motion_half_rgba .visible .entry kernel_gpu_film_convert_motion_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_motion_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_motion_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_motion_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_motion_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<7>; .reg .b16 %rs<5>; .reg .f32 %f<37>; .reg .b32 %r<39>; .reg .b64 %rd<28>; ld.param.u64 %rd3, [kernel_gpu_film_convert_motion_half_rgba_param_1]; ld.param.u64 %rd4, [kernel_gpu_film_convert_motion_half_rgba_param_2]; ld.param.u32 %r31, [kernel_gpu_film_convert_motion_half_rgba_param_3]; ld.param.u32 %r26, [kernel_gpu_film_convert_motion_half_rgba_param_4]; ld.param.u32 %r27, [kernel_gpu_film_convert_motion_half_rgba_param_5]; ld.param.u32 %r28, [kernel_gpu_film_convert_motion_half_rgba_param_6]; ld.param.u32 %r29, [kernel_gpu_film_convert_motion_half_rgba_param_7]; ld.param.u32 %r30, [kernel_gpu_film_convert_motion_half_rgba_param_8]; ld.param.u32 %r21, [kernel_gpu_film_convert_motion_half_rgba_param_0+76]; ld.param.u32 %r14, [kernel_gpu_film_convert_motion_half_rgba_param_0+36]; ld.param.u32 %r13, [kernel_gpu_film_convert_motion_half_rgba_param_0+32]; ld.param.u32 %r6, [kernel_gpu_film_convert_motion_half_rgba_param_0+4]; ld.param.u32 %r5, [kernel_gpu_film_convert_motion_half_rgba_param_0]; mov.u32 %r32, %ntid.x; mov.u32 %r33, %ctaid.x; mov.u32 %r34, %tid.x; mad.lo.s32 %r1, %r33, %r32, %r34; setp.ge.s32 %p1, %r1, %r31; @%p1 bra $L__BB244_7; div.s32 %r2, %r1, %r26; mul.lo.s32 %r35, %r2, %r26; sub.s32 %r3, %r1, %r35; mad.lo.s32 %r36, %r2, %r28, %r3; mul.wide.s32 %rd5, %r36, %r6; cvt.s64.s32 %rd6, %r27; add.s64 %rd1, %rd5, %rd6; cvta.to.global.u64 %rd2, %rd4; cvt.s64.s32 %rd7, %r14; add.s64 %rd8, %rd1, %rd7; shl.b64 %rd9, %rd8, 2; add.s64 %rd10, %rd2, %rd9; ld.global.nc.f32 %f1, [%rd10]; setp.leu.ftz.f32 %p2, %f1, 0f00000000; mov.f32 %f33, 0f00000000; @%p2 bra $L__BB244_3; rcp.approx.ftz.f32 %f33, %f1; $L__BB244_3: cvt.s64.s32 %rd11, %r5; add.s64 %rd12, %rd1, %rd11; shl.b64 %rd13, %rd12, 2; add.s64 %rd14, %rd2, %rd13; ld.global.nc.f32 %f18, [%rd14]; mul.ftz.f32 %f36, %f33, %f18; ld.global.nc.f32 %f19, [%rd14+4]; mul.ftz.f32 %f35, %f33, %f19; ld.global.nc.f32 %f20, [%rd14+8]; mul.ftz.f32 %f34, %f33, %f20; ld.global.nc.f32 %f21, [%rd14+12]; mul.ftz.f32 %f7, %f33, %f21; setp.eq.s32 %p3, %r13, -1; setp.eq.s32 %p4, %r21, 0; or.pred %p5, %p4, %p3; @%p5 bra $L__BB244_6; add.s32 %r37, %r13, 3; cvt.s64.s32 %rd15, %r37; add.s64 %rd16, %rd1, %rd15; shl.b64 %rd18, %rd16, 2; add.s64 %rd19, %rd2, %rd18; ld.global.nc.f32 %f22, [%rd19]; setp.neu.ftz.f32 %p6, %f22, 0f00000000; @%p6 bra $L__BB244_6; mov.f32 %f23, 0f3F800000; sub.ftz.f32 %f24, %f23, %f36; mov.f32 %f25, 0f00000000; sub.ftz.f32 %f26, %f25, %f35; sub.ftz.f32 %f27, %f25, %f34; fma.rn.ftz.f32 %f36, %f24, 0f3F000000, %f36; fma.rn.ftz.f32 %f35, %f26, 0f3F000000, %f35; fma.rn.ftz.f32 %f34, %f27, 0f3F000000, %f34; $L__BB244_6: mov.f32 %f32, 0f477FE000; min.ftz.f32 %f28, %f36, %f32; // begin inline asm { cvt.rn.f16.f32 %rs1, %f28;} // end inline asm min.ftz.f32 %f29, %f35, %f32; // begin inline asm { cvt.rn.f16.f32 %rs2, %f29;} // end inline asm min.ftz.f32 %f30, %f34, %f32; // begin inline asm { cvt.rn.f16.f32 %rs3, %f30;} // end inline asm min.ftz.f32 %f31, %f7, %f32; // begin inline asm { cvt.rn.f16.f32 %rs4, %f31;} // end inline asm cvta.to.global.u64 %rd20, %rd3; mul.lo.s32 %r38, %r2, %r30; cvt.s64.s32 %rd21, %r38; cvt.s64.s32 %rd22, %r3; cvt.s64.s32 %rd23, %r29; add.s64 %rd24, %rd22, %rd23; add.s64 %rd25, %rd24, %rd21; shl.b64 %rd26, %rd25, 3; add.s64 %rd27, %rd20, %rd26; st.global.u16 [%rd27], %rs1; st.global.u16 [%rd27+2], %rs2; st.global.u16 [%rd27+4], %rs3; st.global.u16 [%rd27+6], %rs4; $L__BB244_7: ret; } // .globl kernel_gpu_film_convert_cryptomatte .visible .entry kernel_gpu_film_convert_cryptomatte( .param .align 4 .b8 kernel_gpu_film_convert_cryptomatte_param_0[96], .param .u64 kernel_gpu_film_convert_cryptomatte_param_1, .param .u64 kernel_gpu_film_convert_cryptomatte_param_2, .param .u32 kernel_gpu_film_convert_cryptomatte_param_3, .param .u32 kernel_gpu_film_convert_cryptomatte_param_4, .param .u32 kernel_gpu_film_convert_cryptomatte_param_5, .param .u32 kernel_gpu_film_convert_cryptomatte_param_6, .param .u32 kernel_gpu_film_convert_cryptomatte_param_7, .param .u32 kernel_gpu_film_convert_cryptomatte_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<4>; .reg .f32 %f<16>; .reg .b32 %r<39>; .reg .b64 %rd<19>; ld.param.u64 %rd2, [kernel_gpu_film_convert_cryptomatte_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_cryptomatte_param_2]; ld.param.u32 %r28, [kernel_gpu_film_convert_cryptomatte_param_3]; ld.param.u32 %r24, [kernel_gpu_film_convert_cryptomatte_param_4]; ld.param.u32 %r25, [kernel_gpu_film_convert_cryptomatte_param_5]; ld.param.u32 %r26, [kernel_gpu_film_convert_cryptomatte_param_6]; ld.param.u32 %r27, [kernel_gpu_film_convert_cryptomatte_param_7]; ld.param.u32 %r21, [kernel_gpu_film_convert_cryptomatte_param_0+84]; ld.param.f32 %f15, [kernel_gpu_film_convert_cryptomatte_param_0+56]; ld.param.u32 %r10, [kernel_gpu_film_convert_cryptomatte_param_0+28]; ld.param.u32 %r6, [kernel_gpu_film_convert_cryptomatte_param_0+12]; ld.param.u32 %r4, [kernel_gpu_film_convert_cryptomatte_param_0+4]; ld.param.u32 %r3, [kernel_gpu_film_convert_cryptomatte_param_0]; mov.u32 %r29, %ntid.x; mov.u32 %r30, %ctaid.x; mov.u32 %r31, %tid.x; mad.lo.s32 %r1, %r30, %r29, %r31; setp.ge.s32 %p1, %r1, %r28; @%p1 bra $L__BB245_5; div.s32 %r32, %r1, %r24; mul.lo.s32 %r33, %r32, %r24; sub.s32 %r34, %r1, %r33; mad.lo.s32 %r35, %r32, %r26, %r34; mul.wide.s32 %rd4, %r35, %r4; cvt.s64.s32 %rd5, %r25; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r10, -1; @%p2 bra $L__BB245_4; setp.eq.s32 %p3, %r6, 0; mov.f32 %f15, 0f3F800000; @%p3 bra $L__BB245_4; cvt.s64.s32 %rd6, %r10; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r36, [%rd10]; cvt.rn.f32.u32 %f8, %r36; rcp.approx.ftz.f32 %f15, %f8; $L__BB245_4: cvta.to.global.u64 %rd11, %rd2; add.s32 %r37, %r1, %r27; mul.lo.s32 %r38, %r37, %r21; mul.wide.s32 %rd12, %r38, 4; add.s64 %rd13, %rd11, %rd12; cvt.s64.s32 %rd14, %r3; add.s64 %rd15, %rd1, %rd14; cvta.to.global.u64 %rd16, %rd3; shl.b64 %rd17, %rd15, 2; add.s64 %rd18, %rd16, %rd17; ld.global.nc.f32 %f9, [%rd18]; ld.global.nc.f32 %f10, [%rd18+4]; ld.global.nc.f32 %f11, [%rd18+8]; ld.global.nc.f32 %f12, [%rd18+12]; st.global.f32 [%rd13], %f9; mul.ftz.f32 %f13, %f15, %f10; st.global.f32 [%rd13+4], %f13; st.global.f32 [%rd13+8], %f11; mul.ftz.f32 %f14, %f15, %f12; st.global.f32 [%rd13+12], %f14; $L__BB245_5: ret; } // .globl kernel_gpu_film_convert_cryptomatte_half_rgba .visible .entry kernel_gpu_film_convert_cryptomatte_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_cryptomatte_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_cryptomatte_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_cryptomatte_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_cryptomatte_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<8>; .reg .b16 %rs<5>; .reg .f32 %f<36>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_3]; ld.param.u32 %r27, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_4]; ld.param.u32 %r28, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_5]; ld.param.u32 %r29, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_6]; ld.param.u32 %r30, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_7]; ld.param.u32 %r31, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_8]; ld.param.u32 %r22, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+76]; ld.param.f32 %f32, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+56]; ld.param.u32 %r14, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+32]; ld.param.u32 %r13, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+28]; ld.param.u32 %r9, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+12]; ld.param.u32 %r7, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0+4]; ld.param.u32 %r6, [kernel_gpu_film_convert_cryptomatte_half_rgba_param_0]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p1, %r1, %r32; @%p1 bra $L__BB246_8; div.s32 %r3, %r1, %r27; mul.lo.s32 %r36, %r3, %r27; sub.s32 %r4, %r1, %r36; mad.lo.s32 %r37, %r3, %r29, %r4; mul.wide.s32 %rd4, %r37, %r7; cvt.s64.s32 %rd5, %r28; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r13, -1; @%p2 bra $L__BB246_4; setp.eq.s32 %p3, %r9, 0; mov.f32 %f32, 0f3F800000; @%p3 bra $L__BB246_4; cvt.s64.s32 %rd6, %r13; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r38, [%rd10]; cvt.rn.f32.u32 %f18, %r38; rcp.approx.ftz.f32 %f32, %f18; $L__BB246_4: cvt.s64.s32 %rd11, %r6; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f35, [%rd15]; ld.global.nc.f32 %f33, [%rd15+8]; ld.global.nc.f32 %f19, [%rd15+4]; mul.ftz.f32 %f34, %f32, %f19; ld.global.nc.f32 %f20, [%rd15+12]; mul.ftz.f32 %f7, %f32, %f20; setp.eq.s32 %p4, %r14, -1; setp.eq.s32 %p5, %r22, 0; or.pred %p6, %p5, %p4; @%p6 bra $L__BB246_7; add.s32 %r39, %r14, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f21, [%rd20]; setp.neu.ftz.f32 %p7, %f21, 0f00000000; @%p7 bra $L__BB246_7; mov.f32 %f22, 0f3F800000; sub.ftz.f32 %f23, %f22, %f35; mov.f32 %f24, 0f00000000; sub.ftz.f32 %f25, %f24, %f34; sub.ftz.f32 %f26, %f24, %f33; fma.rn.ftz.f32 %f35, %f23, 0f3F000000, %f35; fma.rn.ftz.f32 %f34, %f25, 0f3F000000, %f34; fma.rn.ftz.f32 %f33, %f26, 0f3F000000, %f33; $L__BB246_7: mov.f32 %f31, 0f477FE000; min.ftz.f32 %f27, %f35, %f31; // begin inline asm { cvt.rn.f16.f32 %rs1, %f27;} // end inline asm min.ftz.f32 %f28, %f34, %f31; // begin inline asm { cvt.rn.f16.f32 %rs2, %f28;} // end inline asm min.ftz.f32 %f29, %f33, %f31; // begin inline asm { cvt.rn.f16.f32 %rs3, %f29;} // end inline asm min.ftz.f32 %f30, %f7, %f31; // begin inline asm { cvt.rn.f16.f32 %rs4, %f30;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r31; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r30; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB246_8: ret; } // .globl kernel_gpu_film_convert_shadow_catcher .visible .entry kernel_gpu_film_convert_shadow_catcher( .param .align 4 .b8 kernel_gpu_film_convert_shadow_catcher_param_0[96], .param .u64 kernel_gpu_film_convert_shadow_catcher_param_1, .param .u64 kernel_gpu_film_convert_shadow_catcher_param_2, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_3, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_4, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_5, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_6, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_7, .param .u32 kernel_gpu_film_convert_shadow_catcher_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<13>; .reg .f32 %f<70>; .reg .b32 %r<41>; .reg .b64 %rd<30>; ld.param.u64 %rd6, [kernel_gpu_film_convert_shadow_catcher_param_1]; ld.param.u64 %rd7, [kernel_gpu_film_convert_shadow_catcher_param_2]; ld.param.u32 %r30, [kernel_gpu_film_convert_shadow_catcher_param_3]; ld.param.u32 %r26, [kernel_gpu_film_convert_shadow_catcher_param_4]; ld.param.u32 %r27, [kernel_gpu_film_convert_shadow_catcher_param_5]; ld.param.u32 %r28, [kernel_gpu_film_convert_shadow_catcher_param_6]; ld.param.u32 %r29, [kernel_gpu_film_convert_shadow_catcher_param_7]; ld.param.u32 %r24, [kernel_gpu_film_convert_shadow_catcher_param_0+88]; ld.param.u32 %r23, [kernel_gpu_film_convert_shadow_catcher_param_0+84]; ld.param.f32 %f62, [kernel_gpu_film_convert_shadow_catcher_param_0+64]; ld.param.f32 %f31, [kernel_gpu_film_convert_shadow_catcher_param_0+60]; ld.param.f32 %f66, [kernel_gpu_film_convert_shadow_catcher_param_0+56]; ld.param.u32 %r17, [kernel_gpu_film_convert_shadow_catcher_param_0+48]; ld.param.u32 %r16, [kernel_gpu_film_convert_shadow_catcher_param_0+44]; ld.param.u32 %r15, [kernel_gpu_film_convert_shadow_catcher_param_0+40]; ld.param.u32 %r12, [kernel_gpu_film_convert_shadow_catcher_param_0+28]; ld.param.u32 %r11, [kernel_gpu_film_convert_shadow_catcher_param_0+24]; ld.param.u32 %r8, [kernel_gpu_film_convert_shadow_catcher_param_0+12]; ld.param.u32 %r7, [kernel_gpu_film_convert_shadow_catcher_param_0+8]; ld.param.u32 %r6, [kernel_gpu_film_convert_shadow_catcher_param_0+4]; mov.u32 %r31, %ntid.x; mov.u32 %r32, %ctaid.x; mov.u32 %r33, %tid.x; mad.lo.s32 %r1, %r32, %r31, %r33; setp.ge.s32 %p1, %r1, %r30; @%p1 bra $L__BB247_20; div.s32 %r34, %r1, %r26; mul.lo.s32 %r35, %r34, %r26; sub.s32 %r36, %r1, %r35; mad.lo.s32 %r37, %r34, %r28, %r36; mul.wide.s32 %rd8, %r37, %r6; cvt.s64.s32 %rd9, %r27; add.s64 %rd1, %rd8, %rd9; setp.eq.s32 %p2, %r24, 0; cvt.s64.s32 %rd10, %r12; add.s64 %rd11, %rd1, %rd10; cvta.to.global.u64 %rd2, %rd7; shl.b64 %rd12, %rd11, 2; add.s64 %rd3, %rd2, %rd12; cvt.s64.s32 %rd13, %r15; add.s64 %rd14, %rd1, %rd13; shl.b64 %rd15, %rd14, 2; add.s64 %rd4, %rd2, %rd15; @%p2 bra $L__BB247_8; setp.eq.s32 %p3, %r12, -1; @%p3 bra $L__BB247_7; ld.global.nc.u32 %r4, [%rd3]; setp.eq.s32 %p4, %r4, 0; mov.f32 %f62, 0f00000000; @%p4 bra $L__BB247_7; setp.eq.s32 %p5, %r8, 0; mov.f32 %f61, 0f3F800000; @%p5 bra $L__BB247_6; cvt.rn.f32.u32 %f35, %r4; rcp.approx.ftz.f32 %f61, %f35; $L__BB247_6: mul.ftz.f32 %f36, %f31, %f61; setp.eq.s32 %p6, %r7, 0; selp.f32 %f62, %f61, %f36, %p6; $L__BB247_7: ld.global.nc.f32 %f37, [%rd4]; ld.global.nc.f32 %f38, [%rd4+4]; ld.global.nc.f32 %f39, [%rd4+8]; mul.ftz.f32 %f69, %f62, %f39; mul.ftz.f32 %f68, %f62, %f38; mul.ftz.f32 %f67, %f62, %f37; bra.uni $L__BB247_19; $L__BB247_8: cvt.s64.s32 %rd16, %r16; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd18, %rd17, 2; add.s64 %rd19, %rd2, %rd18; ld.global.nc.f32 %f43, [%rd19]; setp.eq.ftz.f32 %p7, %f43, 0f00000000; mov.f32 %f67, 0f3F800000; mov.f32 %f68, %f67; mov.f32 %f69, %f67; @%p7 bra $L__BB247_19; cvt.s64.s32 %rd20, %r11; add.s64 %rd21, %rd1, %rd20; shl.b64 %rd22, %rd21, 2; add.s64 %rd5, %rd2, %rd22; cvt.s64.s32 %rd23, %r17; add.s64 %rd24, %rd1, %rd23; shl.b64 %rd25, %rd24, 2; add.s64 %rd26, %rd2, %rd25; ld.global.nc.f32 %f9, [%rd4+4]; ld.global.nc.f32 %f10, [%rd4+8]; ld.global.nc.f32 %f45, [%rd26]; ld.global.nc.f32 %f46, [%rd5]; sub.ftz.f32 %f11, %f46, %f45; ld.global.nc.f32 %f47, [%rd26+4]; ld.global.nc.f32 %f48, [%rd5+4]; sub.ftz.f32 %f12, %f48, %f47; ld.global.nc.f32 %f49, [%rd26+8]; ld.global.nc.f32 %f50, [%rd5+8]; sub.ftz.f32 %f13, %f50, %f49; ld.global.nc.f32 %f14, [%rd4]; setp.eq.ftz.f32 %p8, %f14, 0f00000000; mov.f32 %f65, 0f3F800000; mov.f32 %f63, %f65; @%p8 bra $L__BB247_11; div.approx.ftz.f32 %f63, %f11, %f14; $L__BB247_11: setp.eq.ftz.f32 %p9, %f9, 0f00000000; mov.f32 %f64, %f65; @%p9 bra $L__BB247_13; div.approx.ftz.f32 %f64, %f12, %f9; $L__BB247_13: setp.eq.ftz.f32 %p10, %f10, 0f00000000; @%p10 bra $L__BB247_15; div.approx.ftz.f32 %f65, %f13, %f10; $L__BB247_15: setp.eq.s32 %p11, %r12, -1; @%p11 bra $L__BB247_18; setp.eq.s32 %p12, %r8, 0; mov.f32 %f66, 0f3F800000; @%p12 bra $L__BB247_18; ld.global.nc.u32 %r38, [%rd3]; cvt.rn.f32.u32 %f54, %r38; rcp.approx.ftz.f32 %f66, %f54; $L__BB247_18: ld.global.nc.f32 %f55, [%rd5+12]; mul.ftz.f32 %f56, %f66, %f55; mov.f32 %f57, 0f3F800000; sub.ftz.f32 %f58, %f57, %f56; cvt.ftz.sat.f32.f32 %f59, %f58; sub.ftz.f32 %f60, %f57, %f59; fma.rn.ftz.f32 %f69, %f65, %f59, %f60; fma.rn.ftz.f32 %f68, %f64, %f59, %f60; fma.rn.ftz.f32 %f67, %f63, %f59, %f60; $L__BB247_19: cvta.to.global.u64 %rd27, %rd6; add.s32 %r39, %r1, %r29; mul.lo.s32 %r40, %r39, %r23; mul.wide.s32 %rd28, %r40, 4; add.s64 %rd29, %rd27, %rd28; st.global.f32 [%rd29], %f67; st.global.f32 [%rd29+4], %f68; st.global.f32 [%rd29+8], %f69; $L__BB247_20: ret; } // .globl kernel_gpu_film_convert_shadow_catcher_half_rgba .visible .entry kernel_gpu_film_convert_shadow_catcher_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_shadow_catcher_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<17>; .reg .b16 %rs<5>; .reg .f32 %f<94>; .reg .b32 %r<43>; .reg .b64 %rd<40>; ld.param.u64 %rd6, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_1]; ld.param.u64 %rd7, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_2]; ld.param.u32 %r34, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_3]; ld.param.u32 %r29, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_4]; ld.param.u32 %r30, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_5]; ld.param.u32 %r31, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_6]; ld.param.u32 %r32, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_7]; ld.param.u32 %r33, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_8]; ld.param.u32 %r27, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+88]; ld.param.u32 %r24, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+76]; ld.param.f32 %f83, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+64]; ld.param.f32 %f40, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+60]; ld.param.f32 %f87, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+56]; ld.param.u32 %r20, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+48]; ld.param.u32 %r19, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+44]; ld.param.u32 %r18, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+40]; ld.param.u32 %r16, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+32]; ld.param.u32 %r15, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+28]; ld.param.u32 %r14, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+24]; ld.param.u32 %r11, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+12]; ld.param.u32 %r10, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+8]; ld.param.u32 %r9, [kernel_gpu_film_convert_shadow_catcher_half_rgba_param_0+4]; mov.u32 %r35, %ntid.x; mov.u32 %r36, %ctaid.x; mov.u32 %r37, %tid.x; mad.lo.s32 %r1, %r36, %r35, %r37; setp.ge.s32 %p1, %r1, %r34; @%p1 bra $L__BB248_23; div.s32 %r4, %r1, %r29; mul.lo.s32 %r38, %r4, %r29; sub.s32 %r5, %r1, %r38; mad.lo.s32 %r39, %r4, %r31, %r5; mul.wide.s32 %rd8, %r39, %r9; cvt.s64.s32 %rd9, %r30; add.s64 %rd1, %rd8, %rd9; setp.eq.s32 %p2, %r27, 0; cvt.s64.s32 %rd10, %r15; add.s64 %rd11, %rd1, %rd10; cvta.to.global.u64 %rd2, %rd7; shl.b64 %rd12, %rd11, 2; add.s64 %rd3, %rd2, %rd12; cvt.s64.s32 %rd13, %r18; add.s64 %rd14, %rd1, %rd13; shl.b64 %rd15, %rd14, 2; add.s64 %rd4, %rd2, %rd15; @%p2 bra $L__BB248_8; setp.eq.s32 %p3, %r15, -1; @%p3 bra $L__BB248_7; ld.global.nc.u32 %r6, [%rd3]; setp.eq.s32 %p4, %r6, 0; mov.f32 %f83, 0f00000000; @%p4 bra $L__BB248_7; setp.eq.s32 %p5, %r11, 0; mov.f32 %f82, 0f3F800000; @%p5 bra $L__BB248_6; cvt.rn.f32.u32 %f44, %r6; rcp.approx.ftz.f32 %f82, %f44; $L__BB248_6: mul.ftz.f32 %f45, %f40, %f82; setp.eq.s32 %p6, %r10, 0; selp.f32 %f83, %f82, %f45, %p6; $L__BB248_7: ld.global.nc.f32 %f46, [%rd4]; ld.global.nc.f32 %f47, [%rd4+4]; ld.global.nc.f32 %f48, [%rd4+8]; mul.ftz.f32 %f91, %f83, %f48; mul.ftz.f32 %f92, %f83, %f47; mul.ftz.f32 %f93, %f83, %f46; bra.uni $L__BB248_19; $L__BB248_8: cvt.s64.s32 %rd16, %r19; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd18, %rd17, 2; add.s64 %rd19, %rd2, %rd18; ld.global.nc.f32 %f52, [%rd19]; setp.eq.ftz.f32 %p7, %f52, 0f00000000; mov.f32 %f93, 0f3F800000; mov.f32 %f92, %f93; mov.f32 %f91, %f93; @%p7 bra $L__BB248_19; cvt.s64.s32 %rd20, %r14; add.s64 %rd21, %rd1, %rd20; shl.b64 %rd22, %rd21, 2; add.s64 %rd5, %rd2, %rd22; cvt.s64.s32 %rd23, %r20; add.s64 %rd24, %rd1, %rd23; shl.b64 %rd25, %rd24, 2; add.s64 %rd26, %rd2, %rd25; ld.global.nc.f32 %f9, [%rd4+4]; ld.global.nc.f32 %f10, [%rd4+8]; ld.global.nc.f32 %f54, [%rd26]; ld.global.nc.f32 %f55, [%rd5]; sub.ftz.f32 %f11, %f55, %f54; ld.global.nc.f32 %f56, [%rd26+4]; ld.global.nc.f32 %f57, [%rd5+4]; sub.ftz.f32 %f12, %f57, %f56; ld.global.nc.f32 %f58, [%rd26+8]; ld.global.nc.f32 %f59, [%rd5+8]; sub.ftz.f32 %f13, %f59, %f58; ld.global.nc.f32 %f14, [%rd4]; setp.eq.ftz.f32 %p8, %f14, 0f00000000; mov.f32 %f86, 0f3F800000; mov.f32 %f84, %f86; @%p8 bra $L__BB248_11; div.approx.ftz.f32 %f84, %f11, %f14; $L__BB248_11: setp.eq.ftz.f32 %p9, %f9, 0f00000000; mov.f32 %f85, %f86; @%p9 bra $L__BB248_13; div.approx.ftz.f32 %f85, %f12, %f9; $L__BB248_13: setp.eq.ftz.f32 %p10, %f10, 0f00000000; @%p10 bra $L__BB248_15; div.approx.ftz.f32 %f86, %f13, %f10; $L__BB248_15: setp.eq.s32 %p11, %r15, -1; @%p11 bra $L__BB248_18; setp.eq.s32 %p12, %r11, 0; mov.f32 %f87, 0f3F800000; @%p12 bra $L__BB248_18; ld.global.nc.u32 %r40, [%rd3]; cvt.rn.f32.u32 %f63, %r40; rcp.approx.ftz.f32 %f87, %f63; $L__BB248_18: ld.global.nc.f32 %f64, [%rd5+12]; mul.ftz.f32 %f65, %f87, %f64; mov.f32 %f66, 0f3F800000; sub.ftz.f32 %f67, %f66, %f65; cvt.ftz.sat.f32.f32 %f68, %f67; sub.ftz.f32 %f69, %f66, %f68; fma.rn.ftz.f32 %f91, %f86, %f68, %f69; fma.rn.ftz.f32 %f92, %f85, %f68, %f69; fma.rn.ftz.f32 %f93, %f84, %f68, %f69; $L__BB248_19: setp.eq.s32 %p13, %r16, -1; setp.eq.s32 %p14, %r24, 0; or.pred %p15, %p14, %p13; @%p15 bra $L__BB248_22; add.s32 %r41, %r16, 3; cvt.s64.s32 %rd27, %r41; add.s64 %rd28, %rd1, %rd27; shl.b64 %rd30, %rd28, 2; add.s64 %rd31, %rd2, %rd30; ld.global.nc.f32 %f70, [%rd31]; setp.neu.ftz.f32 %p16, %f70, 0f00000000; @%p16 bra $L__BB248_22; mov.f32 %f71, 0f3F800000; sub.ftz.f32 %f72, %f71, %f93; mov.f32 %f73, 0f00000000; sub.ftz.f32 %f74, %f73, %f92; sub.ftz.f32 %f75, %f73, %f91; fma.rn.ftz.f32 %f93, %f72, 0f3F000000, %f93; fma.rn.ftz.f32 %f92, %f74, 0f3F000000, %f92; fma.rn.ftz.f32 %f91, %f75, 0f3F000000, %f91; $L__BB248_22: mov.f32 %f80, 0f477FE000; min.ftz.f32 %f76, %f93, %f80; // begin inline asm { cvt.rn.f16.f32 %rs1, %f76;} // end inline asm min.ftz.f32 %f77, %f92, %f80; // begin inline asm { cvt.rn.f16.f32 %rs2, %f77;} // end inline asm min.ftz.f32 %f78, %f91, %f80; // begin inline asm { cvt.rn.f16.f32 %rs3, %f78;} // end inline asm min.ftz.f32 %f79, %f81, %f80; // begin inline asm { cvt.rn.f16.f32 %rs4, %f79;} // end inline asm cvta.to.global.u64 %rd32, %rd6; mul.lo.s32 %r42, %r4, %r33; cvt.s64.s32 %rd33, %r42; cvt.s64.s32 %rd34, %r5; cvt.s64.s32 %rd35, %r32; add.s64 %rd36, %rd34, %rd35; add.s64 %rd37, %rd36, %rd33; shl.b64 %rd38, %rd37, 3; add.s64 %rd39, %rd32, %rd38; st.global.u16 [%rd39], %rs1; st.global.u16 [%rd39+2], %rs2; st.global.u16 [%rd39+4], %rs3; st.global.u16 [%rd39+6], %rs4; $L__BB248_23: ret; } // .globl kernel_gpu_film_convert_shadow_catcher_matte_with_shadow .visible .entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow( .param .align 4 .b8 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0[96], .param .u64 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_1, .param .u64 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_2, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_3, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_4, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_5, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_6, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_7, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<24>; .reg .f32 %f<129>; .reg .b32 %r<43>; .reg .b64 %rd<38>; ld.param.u64 %rd7, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_1]; ld.param.u64 %rd8, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_2]; ld.param.u32 %r32, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_3]; ld.param.u32 %r28, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_4]; ld.param.u32 %r29, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_5]; ld.param.u32 %r30, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_6]; ld.param.u32 %r31, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_7]; ld.param.u32 %r26, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+88]; ld.param.u32 %r25, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+84]; ld.param.u32 %r24, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+80]; ld.param.u32 %r22, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+72]; ld.param.f32 %f117, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+64]; ld.param.f32 %f53, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+60]; ld.param.f32 %f121, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+56]; ld.param.u32 %r20, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+52]; ld.param.u32 %r19, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+48]; ld.param.u32 %r18, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+44]; ld.param.u32 %r17, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+40]; ld.param.u32 %r14, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+28]; ld.param.u32 %r13, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+24]; ld.param.u32 %r10, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+12]; ld.param.u32 %r9, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+8]; ld.param.u32 %r8, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_param_0+4]; mov.u32 %r33, %ntid.x; mov.u32 %r34, %ctaid.x; mov.u32 %r35, %tid.x; mad.lo.s32 %r1, %r34, %r33, %r35; setp.ge.s32 %p2, %r1, %r32; @%p2 bra $L__BB249_30; div.s32 %r36, %r1, %r28; mul.lo.s32 %r37, %r36, %r28; sub.s32 %r38, %r1, %r37; mad.lo.s32 %r39, %r36, %r30, %r38; mul.wide.s32 %rd9, %r39, %r8; cvt.s64.s32 %rd10, %r29; add.s64 %rd1, %rd9, %rd10; cvt.s64.s32 %rd11, %r14; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd8; shl.b64 %rd14, %rd12, 2; add.s64 %rd2, %rd13, %rd14; setp.eq.s32 %p4, %r14, -1; mov.pred %p3, 0; mov.f32 %f114, %f117; mov.f32 %f115, %f121; mov.pred %p23, %p3; @%p4 bra $L__BB249_6; ld.global.nc.u32 %r5, [%rd2]; setp.eq.s32 %p6, %r5, 0; mov.pred %p23, -1; mov.f32 %f114, 0f00000000; mov.f32 %f115, %f114; @%p6 bra $L__BB249_6; setp.eq.s32 %p7, %r10, 0; mov.f32 %f115, 0f3F800000; @%p7 bra $L__BB249_5; cvt.rn.f32.u32 %f58, %r5; rcp.approx.ftz.f32 %f115, %f58; $L__BB249_5: mul.ftz.f32 %f59, %f53, %f115; setp.eq.s32 %p9, %r9, 0; selp.f32 %f114, %f115, %f59, %p9; mov.pred %p23, %p3; $L__BB249_6: mov.f32 %f125, 0f00000000; mov.f32 %f126, %f125; mov.f32 %f127, %f125; mov.f32 %f128, %f125; @%p23 bra $L__BB249_28; cvt.s64.s32 %rd15, %r19; add.s64 %rd16, %rd1, %rd15; shl.b64 %rd18, %rd16, 2; add.s64 %rd3, %rd13, %rd18; cvt.s64.s32 %rd19, %r17; add.s64 %rd20, %rd1, %rd19; shl.b64 %rd21, %rd20, 2; add.s64 %rd4, %rd13, %rd21; setp.eq.s32 %p10, %r26, 0; @%p10 bra $L__BB249_14; @%p4 bra $L__BB249_13; ld.global.nc.u32 %r6, [%rd2]; setp.eq.s32 %p12, %r6, 0; mov.f32 %f117, 0f00000000; @%p12 bra $L__BB249_13; setp.eq.s32 %p13, %r10, 0; mov.f32 %f116, 0f3F800000; @%p13 bra $L__BB249_12; cvt.rn.f32.u32 %f66, %r6; rcp.approx.ftz.f32 %f116, %f66; $L__BB249_12: mul.ftz.f32 %f67, %f53, %f116; setp.eq.s32 %p14, %r9, 0; selp.f32 %f117, %f116, %f67, %p14; $L__BB249_13: ld.global.nc.f32 %f68, [%rd4]; ld.global.nc.f32 %f69, [%rd4+4]; ld.global.nc.f32 %f70, [%rd4+8]; mul.ftz.f32 %f124, %f117, %f70; mul.ftz.f32 %f123, %f117, %f69; mul.ftz.f32 %f122, %f117, %f68; bra.uni $L__BB249_25; $L__BB249_14: cvt.s64.s32 %rd22, %r18; add.s64 %rd23, %rd1, %rd22; shl.b64 %rd25, %rd23, 2; add.s64 %rd26, %rd13, %rd25; ld.global.nc.f32 %f74, [%rd26]; setp.eq.ftz.f32 %p15, %f74, 0f00000000; mov.f32 %f122, 0f3F800000; mov.f32 %f123, %f122; mov.f32 %f124, %f122; @%p15 bra $L__BB249_25; cvt.s64.s32 %rd27, %r13; add.s64 %rd28, %rd1, %rd27; shl.b64 %rd30, %rd28, 2; add.s64 %rd5, %rd13, %rd30; ld.global.nc.f32 %f16, [%rd4+4]; ld.global.nc.f32 %f17, [%rd4+8]; ld.global.nc.f32 %f76, [%rd3]; ld.global.nc.f32 %f77, [%rd5]; sub.ftz.f32 %f18, %f77, %f76; ld.global.nc.f32 %f78, [%rd3+4]; ld.global.nc.f32 %f79, [%rd5+4]; sub.ftz.f32 %f19, %f79, %f78; ld.global.nc.f32 %f80, [%rd3+8]; ld.global.nc.f32 %f81, [%rd5+8]; sub.ftz.f32 %f20, %f81, %f80; ld.global.nc.f32 %f21, [%rd4]; setp.eq.ftz.f32 %p16, %f21, 0f00000000; mov.f32 %f120, 0f3F800000; mov.f32 %f118, %f120; @%p16 bra $L__BB249_17; div.approx.ftz.f32 %f118, %f18, %f21; $L__BB249_17: setp.eq.ftz.f32 %p17, %f16, 0f00000000; mov.f32 %f119, %f120; @%p17 bra $L__BB249_19; div.approx.ftz.f32 %f119, %f19, %f16; $L__BB249_19: setp.eq.ftz.f32 %p18, %f17, 0f00000000; @%p18 bra $L__BB249_21; div.approx.ftz.f32 %f120, %f20, %f17; $L__BB249_21: @%p4 bra $L__BB249_24; setp.eq.s32 %p20, %r10, 0; mov.f32 %f121, 0f3F800000; @%p20 bra $L__BB249_24; ld.global.nc.u32 %r40, [%rd2]; cvt.rn.f32.u32 %f85, %r40; rcp.approx.ftz.f32 %f121, %f85; $L__BB249_24: ld.global.nc.f32 %f86, [%rd5+12]; mul.ftz.f32 %f87, %f121, %f86; mov.f32 %f88, 0f3F800000; sub.ftz.f32 %f89, %f88, %f87; cvt.ftz.sat.f32.f32 %f90, %f89; sub.ftz.f32 %f91, %f88, %f90; fma.rn.ftz.f32 %f124, %f120, %f90, %f91; fma.rn.ftz.f32 %f123, %f119, %f90, %f91; fma.rn.ftz.f32 %f122, %f118, %f90, %f91; $L__BB249_25: ld.global.nc.f32 %f92, [%rd3]; mul.ftz.f32 %f125, %f114, %f92; ld.global.nc.f32 %f93, [%rd3+4]; mul.ftz.f32 %f126, %f114, %f93; ld.global.nc.f32 %f94, [%rd3+8]; mul.ftz.f32 %f127, %f114, %f94; ld.global.nc.f32 %f95, [%rd3+12]; mul.ftz.f32 %f96, %f115, %f95; mov.f32 %f128, 0f3F800000; sub.ftz.f32 %f98, %f128, %f96; cvt.ftz.sat.f32.f32 %f99, %f98; sub.ftz.f32 %f100, %f128, %f99; add.ftz.f32 %f101, %f122, %f123; add.ftz.f32 %f102, %f124, %f101; mul.ftz.f32 %f103, %f102, 0f3EAAAAAB; cvt.ftz.sat.f32.f32 %f104, %f103; sub.ftz.f32 %f105, %f128, %f104; fma.rn.ftz.f32 %f39, %f100, %f105, %f99; setp.eq.s32 %p21, %r22, 0; @%p21 bra $L__BB249_27; cvt.s64.s32 %rd31, %r20; add.s64 %rd32, %rd1, %rd31; shl.b64 %rd34, %rd32, 2; add.s64 %rd35, %rd13, %rd34; ld.global.nc.f32 %f106, [%rd35]; mul.ftz.f32 %f107, %f114, %f106; ld.global.nc.f32 %f108, [%rd35+4]; mul.ftz.f32 %f109, %f114, %f108; ld.global.nc.f32 %f110, [%rd35+8]; mul.ftz.f32 %f111, %f114, %f110; sub.ftz.f32 %f112, %f128, %f39; fma.rn.ftz.f32 %f127, %f112, %f111, %f127; fma.rn.ftz.f32 %f126, %f112, %f109, %f126; fma.rn.ftz.f32 %f125, %f112, %f107, %f125; bra.uni $L__BB249_28; $L__BB249_27: mov.f32 %f128, %f39; $L__BB249_28: add.s32 %r41, %r1, %r31; mul.lo.s32 %r42, %r41, %r25; cvta.to.global.u64 %rd36, %rd7; mul.wide.s32 %rd37, %r42, 4; add.s64 %rd6, %rd36, %rd37; st.global.f32 [%rd6], %f125; st.global.f32 [%rd6+4], %f126; st.global.f32 [%rd6+8], %f127; setp.ne.s32 %p22, %r24, 4; @%p22 bra $L__BB249_30; st.global.f32 [%rd6+12], %f128; $L__BB249_30: ret; } // .globl kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba .visible .entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<27>; .reg .b16 %rs<5>; .reg .f32 %f<152>; .reg .b32 %r<45>; .reg .b64 %rd<48>; ld.param.u64 %rd6, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_1]; ld.param.u64 %rd7, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_2]; ld.param.u32 %r36, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_3]; ld.param.u32 %r31, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_4]; ld.param.u32 %r32, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_5]; ld.param.u32 %r33, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_6]; ld.param.u32 %r34, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_7]; ld.param.u32 %r35, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_8]; ld.param.u32 %r29, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+88]; ld.param.u32 %r26, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+76]; ld.param.u32 %r25, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+72]; ld.param.f32 %f137, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+64]; ld.param.f32 %f62, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+60]; ld.param.f32 %f141, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+56]; ld.param.u32 %r23, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+52]; ld.param.u32 %r22, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+48]; ld.param.u32 %r21, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+44]; ld.param.u32 %r20, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+40]; ld.param.u32 %r18, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+32]; ld.param.u32 %r17, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+28]; ld.param.u32 %r16, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+24]; ld.param.u32 %r13, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+12]; ld.param.u32 %r12, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+8]; ld.param.u32 %r11, [kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba_param_0+4]; mov.u32 %r37, %ntid.x; mov.u32 %r38, %ctaid.x; mov.u32 %r39, %tid.x; mad.lo.s32 %r1, %r38, %r37, %r39; setp.ge.s32 %p2, %r1, %r36; @%p2 bra $L__BB250_32; div.s32 %r5, %r1, %r31; mul.lo.s32 %r40, %r5, %r31; sub.s32 %r6, %r1, %r40; mad.lo.s32 %r41, %r5, %r33, %r6; mul.wide.s32 %rd8, %r41, %r11; cvt.s64.s32 %rd9, %r32; add.s64 %rd1, %rd8, %rd9; cvt.s64.s32 %rd10, %r17; add.s64 %rd11, %rd1, %rd10; cvta.to.global.u64 %rd12, %rd7; shl.b64 %rd13, %rd11, 2; add.s64 %rd2, %rd12, %rd13; setp.eq.s32 %p4, %r17, -1; mov.pred %p3, 0; mov.f32 %f134, %f137; mov.f32 %f135, %f141; mov.pred %p26, %p3; @%p4 bra $L__BB250_6; ld.global.nc.u32 %r7, [%rd2]; setp.eq.s32 %p6, %r7, 0; mov.pred %p26, -1; mov.f32 %f134, 0f00000000; mov.f32 %f135, %f134; @%p6 bra $L__BB250_6; setp.eq.s32 %p7, %r13, 0; mov.f32 %f135, 0f3F800000; @%p7 bra $L__BB250_5; cvt.rn.f32.u32 %f67, %r7; rcp.approx.ftz.f32 %f135, %f67; $L__BB250_5: mul.ftz.f32 %f68, %f62, %f135; setp.eq.s32 %p9, %r12, 0; selp.f32 %f134, %f135, %f68, %p9; mov.pred %p26, %p3; $L__BB250_6: mov.f32 %f151, 0f00000000; mov.f32 %f150, %f151; mov.f32 %f149, %f151; mov.f32 %f148, %f151; @%p26 bra $L__BB250_28; cvt.s64.s32 %rd14, %r22; add.s64 %rd15, %rd1, %rd14; shl.b64 %rd17, %rd15, 2; add.s64 %rd3, %rd12, %rd17; cvt.s64.s32 %rd18, %r20; add.s64 %rd19, %rd1, %rd18; shl.b64 %rd20, %rd19, 2; add.s64 %rd4, %rd12, %rd20; setp.eq.s32 %p10, %r29, 0; @%p10 bra $L__BB250_14; @%p4 bra $L__BB250_13; ld.global.nc.u32 %r8, [%rd2]; setp.eq.s32 %p12, %r8, 0; mov.f32 %f137, 0f00000000; @%p12 bra $L__BB250_13; setp.eq.s32 %p13, %r13, 0; mov.f32 %f136, 0f3F800000; @%p13 bra $L__BB250_12; cvt.rn.f32.u32 %f75, %r8; rcp.approx.ftz.f32 %f136, %f75; $L__BB250_12: mul.ftz.f32 %f76, %f62, %f136; setp.eq.s32 %p14, %r12, 0; selp.f32 %f137, %f136, %f76, %p14; $L__BB250_13: ld.global.nc.f32 %f77, [%rd4]; ld.global.nc.f32 %f78, [%rd4+4]; ld.global.nc.f32 %f79, [%rd4+8]; mul.ftz.f32 %f144, %f137, %f79; mul.ftz.f32 %f143, %f137, %f78; mul.ftz.f32 %f142, %f137, %f77; bra.uni $L__BB250_25; $L__BB250_14: cvt.s64.s32 %rd21, %r21; add.s64 %rd22, %rd1, %rd21; shl.b64 %rd24, %rd22, 2; add.s64 %rd25, %rd12, %rd24; ld.global.nc.f32 %f83, [%rd25]; setp.eq.ftz.f32 %p15, %f83, 0f00000000; mov.f32 %f142, 0f3F800000; mov.f32 %f143, %f142; mov.f32 %f144, %f142; @%p15 bra $L__BB250_25; cvt.s64.s32 %rd26, %r16; add.s64 %rd27, %rd1, %rd26; shl.b64 %rd29, %rd27, 2; add.s64 %rd5, %rd12, %rd29; ld.global.nc.f32 %f16, [%rd4+4]; ld.global.nc.f32 %f17, [%rd4+8]; ld.global.nc.f32 %f85, [%rd3]; ld.global.nc.f32 %f86, [%rd5]; sub.ftz.f32 %f18, %f86, %f85; ld.global.nc.f32 %f87, [%rd3+4]; ld.global.nc.f32 %f88, [%rd5+4]; sub.ftz.f32 %f19, %f88, %f87; ld.global.nc.f32 %f89, [%rd3+8]; ld.global.nc.f32 %f90, [%rd5+8]; sub.ftz.f32 %f20, %f90, %f89; ld.global.nc.f32 %f21, [%rd4]; setp.eq.ftz.f32 %p16, %f21, 0f00000000; mov.f32 %f140, 0f3F800000; mov.f32 %f138, %f140; @%p16 bra $L__BB250_17; div.approx.ftz.f32 %f138, %f18, %f21; $L__BB250_17: setp.eq.ftz.f32 %p17, %f16, 0f00000000; mov.f32 %f139, %f140; @%p17 bra $L__BB250_19; div.approx.ftz.f32 %f139, %f19, %f16; $L__BB250_19: setp.eq.ftz.f32 %p18, %f17, 0f00000000; @%p18 bra $L__BB250_21; div.approx.ftz.f32 %f140, %f20, %f17; $L__BB250_21: @%p4 bra $L__BB250_24; setp.eq.s32 %p20, %r13, 0; mov.f32 %f141, 0f3F800000; @%p20 bra $L__BB250_24; ld.global.nc.u32 %r42, [%rd2]; cvt.rn.f32.u32 %f94, %r42; rcp.approx.ftz.f32 %f141, %f94; $L__BB250_24: ld.global.nc.f32 %f95, [%rd5+12]; mul.ftz.f32 %f96, %f141, %f95; mov.f32 %f97, 0f3F800000; sub.ftz.f32 %f98, %f97, %f96; cvt.ftz.sat.f32.f32 %f99, %f98; sub.ftz.f32 %f100, %f97, %f99; fma.rn.ftz.f32 %f144, %f140, %f99, %f100; fma.rn.ftz.f32 %f143, %f139, %f99, %f100; fma.rn.ftz.f32 %f142, %f138, %f99, %f100; $L__BB250_25: ld.global.nc.f32 %f101, [%rd3]; mul.ftz.f32 %f151, %f134, %f101; ld.global.nc.f32 %f102, [%rd3+4]; mul.ftz.f32 %f150, %f134, %f102; ld.global.nc.f32 %f103, [%rd3+8]; mul.ftz.f32 %f149, %f134, %f103; ld.global.nc.f32 %f104, [%rd3+12]; mul.ftz.f32 %f105, %f135, %f104; mov.f32 %f148, 0f3F800000; sub.ftz.f32 %f107, %f148, %f105; cvt.ftz.sat.f32.f32 %f108, %f107; sub.ftz.f32 %f109, %f148, %f108; add.ftz.f32 %f110, %f142, %f143; add.ftz.f32 %f111, %f144, %f110; mul.ftz.f32 %f112, %f111, 0f3EAAAAAB; cvt.ftz.sat.f32.f32 %f113, %f112; sub.ftz.f32 %f114, %f148, %f113; fma.rn.ftz.f32 %f39, %f109, %f114, %f108; setp.eq.s32 %p21, %r25, 0; @%p21 bra $L__BB250_27; cvt.s64.s32 %rd30, %r23; add.s64 %rd31, %rd1, %rd30; shl.b64 %rd33, %rd31, 2; add.s64 %rd34, %rd12, %rd33; ld.global.nc.f32 %f115, [%rd34]; mul.ftz.f32 %f116, %f134, %f115; ld.global.nc.f32 %f117, [%rd34+4]; mul.ftz.f32 %f118, %f134, %f117; ld.global.nc.f32 %f119, [%rd34+8]; mul.ftz.f32 %f120, %f134, %f119; sub.ftz.f32 %f121, %f148, %f39; fma.rn.ftz.f32 %f149, %f121, %f120, %f149; fma.rn.ftz.f32 %f150, %f121, %f118, %f150; fma.rn.ftz.f32 %f151, %f121, %f116, %f151; bra.uni $L__BB250_28; $L__BB250_27: mov.f32 %f148, %f39; $L__BB250_28: setp.eq.s32 %p22, %r18, -1; setp.eq.s32 %p23, %r26, 0; or.pred %p24, %p23, %p22; @%p24 bra $L__BB250_31; add.s32 %r43, %r18, 3; cvt.s64.s32 %rd35, %r43; add.s64 %rd36, %rd1, %rd35; shl.b64 %rd38, %rd36, 2; add.s64 %rd39, %rd12, %rd38; ld.global.nc.f32 %f122, [%rd39]; setp.neu.ftz.f32 %p25, %f122, 0f00000000; @%p25 bra $L__BB250_31; mov.f32 %f123, 0f3F800000; sub.ftz.f32 %f124, %f123, %f151; mov.f32 %f125, 0f00000000; sub.ftz.f32 %f126, %f125, %f150; sub.ftz.f32 %f127, %f125, %f149; fma.rn.ftz.f32 %f151, %f124, 0f3F000000, %f151; fma.rn.ftz.f32 %f150, %f126, 0f3F000000, %f150; fma.rn.ftz.f32 %f149, %f127, 0f3F000000, %f149; $L__BB250_31: mov.f32 %f132, 0f477FE000; min.ftz.f32 %f128, %f151, %f132; // begin inline asm { cvt.rn.f16.f32 %rs1, %f128;} // end inline asm min.ftz.f32 %f129, %f150, %f132; // begin inline asm { cvt.rn.f16.f32 %rs2, %f129;} // end inline asm min.ftz.f32 %f130, %f149, %f132; // begin inline asm { cvt.rn.f16.f32 %rs3, %f130;} // end inline asm min.ftz.f32 %f131, %f148, %f132; // begin inline asm { cvt.rn.f16.f32 %rs4, %f131;} // end inline asm cvta.to.global.u64 %rd40, %rd6; mul.lo.s32 %r44, %r5, %r35; cvt.s64.s32 %rd41, %r44; cvt.s64.s32 %rd42, %r6; cvt.s64.s32 %rd43, %r34; add.s64 %rd44, %rd42, %rd43; add.s64 %rd45, %rd44, %rd41; shl.b64 %rd46, %rd45, 3; add.s64 %rd47, %rd40, %rd46; st.global.u16 [%rd47], %rs1; st.global.u16 [%rd47+2], %rs2; st.global.u16 [%rd47+4], %rs3; st.global.u16 [%rd47+6], %rs4; $L__BB250_32: ret; } // .globl kernel_gpu_film_convert_combined .visible .entry kernel_gpu_film_convert_combined( .param .align 4 .b8 kernel_gpu_film_convert_combined_param_0[96], .param .u64 kernel_gpu_film_convert_combined_param_1, .param .u64 kernel_gpu_film_convert_combined_param_2, .param .u32 kernel_gpu_film_convert_combined_param_3, .param .u32 kernel_gpu_film_convert_combined_param_4, .param .u32 kernel_gpu_film_convert_combined_param_5, .param .u32 kernel_gpu_film_convert_combined_param_6, .param .u32 kernel_gpu_film_convert_combined_param_7, .param .u32 kernel_gpu_film_convert_combined_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<11>; .reg .f32 %f<30>; .reg .b32 %r<40>; .reg .b64 %rd<19>; ld.param.u64 %rd3, [kernel_gpu_film_convert_combined_param_1]; ld.param.u64 %rd4, [kernel_gpu_film_convert_combined_param_2]; ld.param.u32 %r29, [kernel_gpu_film_convert_combined_param_3]; ld.param.u32 %r25, [kernel_gpu_film_convert_combined_param_4]; ld.param.u32 %r26, [kernel_gpu_film_convert_combined_param_5]; ld.param.u32 %r27, [kernel_gpu_film_convert_combined_param_6]; ld.param.u32 %r28, [kernel_gpu_film_convert_combined_param_7]; ld.param.u32 %r22, [kernel_gpu_film_convert_combined_param_0+84]; ld.param.f32 %f28, [kernel_gpu_film_convert_combined_param_0+64]; ld.param.f32 %f9, [kernel_gpu_film_convert_combined_param_0+60]; ld.param.f32 %f29, [kernel_gpu_film_convert_combined_param_0+56]; ld.param.u32 %r11, [kernel_gpu_film_convert_combined_param_0+28]; ld.param.u32 %r7, [kernel_gpu_film_convert_combined_param_0+12]; ld.param.u32 %r6, [kernel_gpu_film_convert_combined_param_0+8]; ld.param.u32 %r5, [kernel_gpu_film_convert_combined_param_0+4]; ld.param.u32 %r4, [kernel_gpu_film_convert_combined_param_0]; mov.u32 %r30, %ntid.x; mov.u32 %r31, %ctaid.x; mov.u32 %r32, %tid.x; mad.lo.s32 %r1, %r31, %r30, %r32; setp.ge.s32 %p2, %r1, %r29; @%p2 bra $L__BB251_9; div.s32 %r33, %r1, %r25; mul.lo.s32 %r34, %r33, %r25; sub.s32 %r35, %r1, %r34; mad.lo.s32 %r36, %r33, %r27, %r35; mul.wide.s32 %rd5, %r36, %r5; cvt.s64.s32 %rd6, %r26; add.s64 %rd1, %rd5, %rd6; setp.eq.s32 %p4, %r11, -1; mov.pred %p3, 0; mov.pred %p10, %p3; @%p4 bra $L__BB251_6; cvt.s64.s32 %rd7, %r11; add.s64 %rd8, %rd1, %rd7; cvta.to.global.u64 %rd9, %rd4; shl.b64 %rd10, %rd8, 2; add.s64 %rd11, %rd9, %rd10; ld.global.nc.u32 %r3, [%rd11]; setp.eq.s32 %p6, %r3, 0; mov.pred %p10, -1; mov.f32 %f28, 0f00000000; mov.f32 %f29, %f28; @%p6 bra $L__BB251_6; setp.eq.s32 %p7, %r7, 0; mov.f32 %f29, 0f3F800000; @%p7 bra $L__BB251_5; cvt.rn.f32.u32 %f14, %r3; rcp.approx.ftz.f32 %f29, %f14; $L__BB251_5: mul.ftz.f32 %f15, %f9, %f29; setp.eq.s32 %p9, %r6, 0; selp.f32 %f28, %f29, %f15, %p9; mov.pred %p10, %p3; $L__BB251_6: cvta.to.global.u64 %rd12, %rd3; add.s32 %r37, %r1, %r28; mul.lo.s32 %r38, %r37, %r22; mul.wide.s32 %rd13, %r38, 4; add.s64 %rd2, %rd12, %rd13; @%p10 bra $L__BB251_8; bra.uni $L__BB251_7; $L__BB251_8: mov.u32 %r39, 0; st.global.u32 [%rd2], %r39; st.global.u32 [%rd2+4], %r39; st.global.u32 [%rd2+8], %r39; st.global.u32 [%rd2+12], %r39; bra.uni $L__BB251_9; $L__BB251_7: cvt.s64.s32 %rd14, %r4; add.s64 %rd15, %rd1, %rd14; cvta.to.global.u64 %rd16, %rd4; shl.b64 %rd17, %rd15, 2; add.s64 %rd18, %rd16, %rd17; ld.global.nc.f32 %f16, [%rd18]; mul.ftz.f32 %f17, %f28, %f16; ld.global.nc.f32 %f18, [%rd18+4]; mul.ftz.f32 %f19, %f28, %f18; ld.global.nc.f32 %f20, [%rd18+8]; mul.ftz.f32 %f21, %f28, %f20; ld.global.nc.f32 %f22, [%rd18+12]; mul.ftz.f32 %f23, %f29, %f22; st.global.f32 [%rd2], %f17; st.global.f32 [%rd2+4], %f19; st.global.f32 [%rd2+8], %f21; mov.f32 %f24, 0f3F800000; sub.ftz.f32 %f25, %f24, %f23; cvt.ftz.sat.f32.f32 %f26, %f25; st.global.f32 [%rd2+12], %f26; $L__BB251_9: ret; } // .globl kernel_gpu_film_convert_combined_half_rgba .visible .entry kernel_gpu_film_convert_combined_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_combined_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_combined_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_combined_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_combined_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<15>; .reg .b16 %rs<5>; .reg .f32 %f<62>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_combined_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_combined_half_rgba_param_2]; ld.param.u32 %r33, [kernel_gpu_film_convert_combined_half_rgba_param_3]; ld.param.u32 %r28, [kernel_gpu_film_convert_combined_half_rgba_param_4]; ld.param.u32 %r29, [kernel_gpu_film_convert_combined_half_rgba_param_5]; ld.param.u32 %r30, [kernel_gpu_film_convert_combined_half_rgba_param_6]; ld.param.u32 %r31, [kernel_gpu_film_convert_combined_half_rgba_param_7]; ld.param.u32 %r32, [kernel_gpu_film_convert_combined_half_rgba_param_8]; ld.param.u32 %r23, [kernel_gpu_film_convert_combined_half_rgba_param_0+76]; ld.param.f32 %f53, [kernel_gpu_film_convert_combined_half_rgba_param_0+64]; ld.param.f32 %f23, [kernel_gpu_film_convert_combined_half_rgba_param_0+60]; ld.param.f32 %f54, [kernel_gpu_film_convert_combined_half_rgba_param_0+56]; ld.param.u32 %r15, [kernel_gpu_film_convert_combined_half_rgba_param_0+32]; ld.param.u32 %r14, [kernel_gpu_film_convert_combined_half_rgba_param_0+28]; ld.param.u32 %r10, [kernel_gpu_film_convert_combined_half_rgba_param_0+12]; ld.param.u32 %r9, [kernel_gpu_film_convert_combined_half_rgba_param_0+8]; ld.param.u32 %r8, [kernel_gpu_film_convert_combined_half_rgba_param_0+4]; ld.param.u32 %r7, [kernel_gpu_film_convert_combined_half_rgba_param_0]; mov.u32 %r34, %ntid.x; mov.u32 %r35, %ctaid.x; mov.u32 %r36, %tid.x; mad.lo.s32 %r1, %r35, %r34, %r36; setp.ge.s32 %p2, %r1, %r33; @%p2 bra $L__BB252_12; div.s32 %r3, %r1, %r28; mul.lo.s32 %r37, %r3, %r28; sub.s32 %r4, %r1, %r37; mad.lo.s32 %r38, %r3, %r30, %r4; mul.wide.s32 %rd4, %r38, %r8; cvt.s64.s32 %rd5, %r29; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p4, %r14, -1; mov.pred %p3, 0; mov.pred %p14, %p3; @%p4 bra $L__BB252_6; cvt.s64.s32 %rd6, %r14; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r5, [%rd10]; setp.eq.s32 %p6, %r5, 0; mov.pred %p14, -1; mov.f32 %f53, 0f00000000; mov.f32 %f54, %f53; @%p6 bra $L__BB252_6; setp.eq.s32 %p7, %r10, 0; mov.f32 %f54, 0f3F800000; @%p7 bra $L__BB252_5; cvt.rn.f32.u32 %f28, %r5; rcp.approx.ftz.f32 %f54, %f28; $L__BB252_5: mul.ftz.f32 %f29, %f23, %f54; setp.eq.s32 %p9, %r9, 0; selp.f32 %f53, %f54, %f29, %p9; mov.pred %p14, %p3; $L__BB252_6: mov.f32 %f55, 0f00000000; mov.f32 %f59, %f55; mov.f32 %f60, %f55; mov.f32 %f61, %f55; @%p14 bra $L__BB252_8; cvt.s64.s32 %rd11, %r7; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f34, [%rd15]; mul.ftz.f32 %f61, %f53, %f34; ld.global.nc.f32 %f35, [%rd15+4]; mul.ftz.f32 %f60, %f53, %f35; ld.global.nc.f32 %f36, [%rd15+8]; mul.ftz.f32 %f59, %f53, %f36; ld.global.nc.f32 %f37, [%rd15+12]; mul.ftz.f32 %f38, %f54, %f37; mov.f32 %f39, 0f3F800000; sub.ftz.f32 %f40, %f39, %f38; cvt.ftz.sat.f32.f32 %f55, %f40; $L__BB252_8: setp.eq.s32 %p10, %r15, -1; setp.eq.s32 %p11, %r23, 0; or.pred %p12, %p11, %p10; @%p12 bra $L__BB252_11; add.s32 %r39, %r15, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; cvta.to.global.u64 %rd18, %rd3; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd18, %rd19; ld.global.nc.f32 %f41, [%rd20]; setp.neu.ftz.f32 %p13, %f41, 0f00000000; @%p13 bra $L__BB252_11; mov.f32 %f42, 0f3F800000; sub.ftz.f32 %f43, %f42, %f61; mov.f32 %f44, 0f00000000; sub.ftz.f32 %f45, %f44, %f60; sub.ftz.f32 %f46, %f44, %f59; fma.rn.ftz.f32 %f61, %f43, 0f3F000000, %f61; fma.rn.ftz.f32 %f60, %f45, 0f3F000000, %f60; fma.rn.ftz.f32 %f59, %f46, 0f3F000000, %f59; $L__BB252_11: mov.f32 %f51, 0f477FE000; min.ftz.f32 %f47, %f61, %f51; // begin inline asm { cvt.rn.f16.f32 %rs1, %f47;} // end inline asm min.ftz.f32 %f48, %f60, %f51; // begin inline asm { cvt.rn.f16.f32 %rs2, %f48;} // end inline asm min.ftz.f32 %f49, %f59, %f51; // begin inline asm { cvt.rn.f16.f32 %rs3, %f49;} // end inline asm min.ftz.f32 %f50, %f55, %f51; // begin inline asm { cvt.rn.f16.f32 %rs4, %f50;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r32; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r31; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB252_12: ret; } // .globl kernel_gpu_film_convert_float4 .visible .entry kernel_gpu_film_convert_float4( .param .align 4 .b8 kernel_gpu_film_convert_float4_param_0[96], .param .u64 kernel_gpu_film_convert_float4_param_1, .param .u64 kernel_gpu_film_convert_float4_param_2, .param .u32 kernel_gpu_film_convert_float4_param_3, .param .u32 kernel_gpu_film_convert_float4_param_4, .param .u32 kernel_gpu_film_convert_float4_param_5, .param .u32 kernel_gpu_film_convert_float4_param_6, .param .u32 kernel_gpu_film_convert_float4_param_7, .param .u32 kernel_gpu_film_convert_float4_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<6>; .reg .f32 %f<27>; .reg .b32 %r<39>; .reg .b64 %rd<19>; ld.param.u64 %rd2, [kernel_gpu_film_convert_float4_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_float4_param_2]; ld.param.u32 %r29, [kernel_gpu_film_convert_float4_param_3]; ld.param.u32 %r25, [kernel_gpu_film_convert_float4_param_4]; ld.param.u32 %r26, [kernel_gpu_film_convert_float4_param_5]; ld.param.u32 %r27, [kernel_gpu_film_convert_float4_param_6]; ld.param.u32 %r28, [kernel_gpu_film_convert_float4_param_7]; ld.param.u32 %r22, [kernel_gpu_film_convert_float4_param_0+84]; ld.param.f32 %f25, [kernel_gpu_film_convert_float4_param_0+64]; ld.param.f32 %f9, [kernel_gpu_film_convert_float4_param_0+60]; ld.param.f32 %f26, [kernel_gpu_film_convert_float4_param_0+56]; ld.param.u32 %r11, [kernel_gpu_film_convert_float4_param_0+28]; ld.param.u32 %r7, [kernel_gpu_film_convert_float4_param_0+12]; ld.param.u32 %r6, [kernel_gpu_film_convert_float4_param_0+8]; ld.param.u32 %r5, [kernel_gpu_film_convert_float4_param_0+4]; ld.param.u32 %r4, [kernel_gpu_film_convert_float4_param_0]; mov.u32 %r30, %ntid.x; mov.u32 %r31, %ctaid.x; mov.u32 %r32, %tid.x; mad.lo.s32 %r1, %r31, %r30, %r32; setp.ge.s32 %p1, %r1, %r29; @%p1 bra $L__BB253_7; div.s32 %r33, %r1, %r25; mul.lo.s32 %r34, %r33, %r25; sub.s32 %r35, %r1, %r34; mad.lo.s32 %r36, %r33, %r27, %r35; mul.wide.s32 %rd4, %r36, %r5; cvt.s64.s32 %rd5, %r26; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r11, -1; @%p2 bra $L__BB253_6; cvt.s64.s32 %rd6, %r11; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r3, [%rd10]; setp.eq.s32 %p3, %r3, 0; mov.f32 %f25, 0f00000000; mov.f32 %f26, %f25; @%p3 bra $L__BB253_6; setp.eq.s32 %p4, %r7, 0; mov.f32 %f26, 0f3F800000; @%p4 bra $L__BB253_5; cvt.rn.f32.u32 %f14, %r3; rcp.approx.ftz.f32 %f26, %f14; $L__BB253_5: mul.ftz.f32 %f15, %f9, %f26; setp.eq.s32 %p5, %r6, 0; selp.f32 %f25, %f26, %f15, %p5; $L__BB253_6: cvta.to.global.u64 %rd11, %rd2; add.s32 %r37, %r1, %r28; mul.lo.s32 %r38, %r37, %r22; mul.wide.s32 %rd12, %r38, 4; add.s64 %rd13, %rd11, %rd12; cvt.s64.s32 %rd14, %r4; add.s64 %rd15, %rd1, %rd14; cvta.to.global.u64 %rd16, %rd3; shl.b64 %rd17, %rd15, 2; add.s64 %rd18, %rd16, %rd17; ld.global.nc.f32 %f16, [%rd18]; mul.ftz.f32 %f17, %f25, %f16; ld.global.nc.f32 %f18, [%rd18+4]; mul.ftz.f32 %f19, %f25, %f18; ld.global.nc.f32 %f20, [%rd18+8]; mul.ftz.f32 %f21, %f25, %f20; ld.global.nc.f32 %f22, [%rd18+12]; mul.ftz.f32 %f23, %f26, %f22; st.global.f32 [%rd13], %f17; st.global.f32 [%rd13+4], %f19; st.global.f32 [%rd13+8], %f21; st.global.f32 [%rd13+12], %f23; $L__BB253_7: ret; } // .globl kernel_gpu_film_convert_float4_half_rgba .visible .entry kernel_gpu_film_convert_float4_half_rgba( .param .align 4 .b8 kernel_gpu_film_convert_float4_half_rgba_param_0[96], .param .u64 kernel_gpu_film_convert_float4_half_rgba_param_1, .param .u64 kernel_gpu_film_convert_float4_half_rgba_param_2, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_3, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_4, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_5, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_6, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_7, .param .u32 kernel_gpu_film_convert_float4_half_rgba_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<10>; .reg .b16 %rs<5>; .reg .f32 %f<47>; .reg .b32 %r<41>; .reg .b64 %rd<29>; ld.param.u64 %rd2, [kernel_gpu_film_convert_float4_half_rgba_param_1]; ld.param.u64 %rd3, [kernel_gpu_film_convert_float4_half_rgba_param_2]; ld.param.u32 %r33, [kernel_gpu_film_convert_float4_half_rgba_param_3]; ld.param.u32 %r28, [kernel_gpu_film_convert_float4_half_rgba_param_4]; ld.param.u32 %r29, [kernel_gpu_film_convert_float4_half_rgba_param_5]; ld.param.u32 %r30, [kernel_gpu_film_convert_float4_half_rgba_param_6]; ld.param.u32 %r31, [kernel_gpu_film_convert_float4_half_rgba_param_7]; ld.param.u32 %r32, [kernel_gpu_film_convert_float4_half_rgba_param_8]; ld.param.u32 %r23, [kernel_gpu_film_convert_float4_half_rgba_param_0+76]; ld.param.f32 %f42, [kernel_gpu_film_convert_float4_half_rgba_param_0+64]; ld.param.f32 %f19, [kernel_gpu_film_convert_float4_half_rgba_param_0+60]; ld.param.f32 %f43, [kernel_gpu_film_convert_float4_half_rgba_param_0+56]; ld.param.u32 %r15, [kernel_gpu_film_convert_float4_half_rgba_param_0+32]; ld.param.u32 %r14, [kernel_gpu_film_convert_float4_half_rgba_param_0+28]; ld.param.u32 %r10, [kernel_gpu_film_convert_float4_half_rgba_param_0+12]; ld.param.u32 %r9, [kernel_gpu_film_convert_float4_half_rgba_param_0+8]; ld.param.u32 %r8, [kernel_gpu_film_convert_float4_half_rgba_param_0+4]; ld.param.u32 %r7, [kernel_gpu_film_convert_float4_half_rgba_param_0]; mov.u32 %r34, %ntid.x; mov.u32 %r35, %ctaid.x; mov.u32 %r36, %tid.x; mad.lo.s32 %r1, %r35, %r34, %r36; setp.ge.s32 %p1, %r1, %r33; @%p1 bra $L__BB254_10; div.s32 %r3, %r1, %r28; mul.lo.s32 %r37, %r3, %r28; sub.s32 %r4, %r1, %r37; mad.lo.s32 %r38, %r3, %r30, %r4; mul.wide.s32 %rd4, %r38, %r8; cvt.s64.s32 %rd5, %r29; add.s64 %rd1, %rd4, %rd5; setp.eq.s32 %p2, %r14, -1; @%p2 bra $L__BB254_6; cvt.s64.s32 %rd6, %r14; add.s64 %rd7, %rd1, %rd6; cvta.to.global.u64 %rd8, %rd3; shl.b64 %rd9, %rd7, 2; add.s64 %rd10, %rd8, %rd9; ld.global.nc.u32 %r5, [%rd10]; setp.eq.s32 %p3, %r5, 0; mov.f32 %f42, 0f00000000; mov.f32 %f43, %f42; @%p3 bra $L__BB254_6; setp.eq.s32 %p4, %r10, 0; mov.f32 %f43, 0f3F800000; @%p4 bra $L__BB254_5; cvt.rn.f32.u32 %f24, %r5; rcp.approx.ftz.f32 %f43, %f24; $L__BB254_5: mul.ftz.f32 %f25, %f19, %f43; setp.eq.s32 %p5, %r9, 0; selp.f32 %f42, %f43, %f25, %p5; $L__BB254_6: cvt.s64.s32 %rd11, %r7; add.s64 %rd12, %rd1, %rd11; cvta.to.global.u64 %rd13, %rd3; shl.b64 %rd14, %rd12, 2; add.s64 %rd15, %rd13, %rd14; ld.global.nc.f32 %f26, [%rd15]; mul.ftz.f32 %f46, %f42, %f26; ld.global.nc.f32 %f27, [%rd15+4]; mul.ftz.f32 %f45, %f42, %f27; ld.global.nc.f32 %f28, [%rd15+8]; mul.ftz.f32 %f44, %f42, %f28; ld.global.nc.f32 %f29, [%rd15+12]; mul.ftz.f32 %f11, %f43, %f29; setp.eq.s32 %p6, %r15, -1; setp.eq.s32 %p7, %r23, 0; or.pred %p8, %p7, %p6; @%p8 bra $L__BB254_9; add.s32 %r39, %r15, 3; cvt.s64.s32 %rd16, %r39; add.s64 %rd17, %rd1, %rd16; shl.b64 %rd19, %rd17, 2; add.s64 %rd20, %rd13, %rd19; ld.global.nc.f32 %f30, [%rd20]; setp.neu.ftz.f32 %p9, %f30, 0f00000000; @%p9 bra $L__BB254_9; mov.f32 %f31, 0f3F800000; sub.ftz.f32 %f32, %f31, %f46; mov.f32 %f33, 0f00000000; sub.ftz.f32 %f34, %f33, %f45; sub.ftz.f32 %f35, %f33, %f44; fma.rn.ftz.f32 %f46, %f32, 0f3F000000, %f46; fma.rn.ftz.f32 %f45, %f34, 0f3F000000, %f45; fma.rn.ftz.f32 %f44, %f35, 0f3F000000, %f44; $L__BB254_9: mov.f32 %f40, 0f477FE000; min.ftz.f32 %f36, %f46, %f40; // begin inline asm { cvt.rn.f16.f32 %rs1, %f36;} // end inline asm min.ftz.f32 %f37, %f45, %f40; // begin inline asm { cvt.rn.f16.f32 %rs2, %f37;} // end inline asm min.ftz.f32 %f38, %f44, %f40; // begin inline asm { cvt.rn.f16.f32 %rs3, %f38;} // end inline asm min.ftz.f32 %f39, %f11, %f40; // begin inline asm { cvt.rn.f16.f32 %rs4, %f39;} // end inline asm cvta.to.global.u64 %rd21, %rd2; mul.lo.s32 %r40, %r3, %r32; cvt.s64.s32 %rd22, %r40; cvt.s64.s32 %rd23, %r4; cvt.s64.s32 %rd24, %r31; add.s64 %rd25, %rd23, %rd24; add.s64 %rd26, %rd25, %rd22; shl.b64 %rd27, %rd26, 3; add.s64 %rd28, %rd21, %rd27; st.global.u16 [%rd28], %rs1; st.global.u16 [%rd28+2], %rs2; st.global.u16 [%rd28+4], %rs3; st.global.u16 [%rd28+6], %rs4; $L__BB254_10: ret; } // .globl kernel_gpu_shader_eval_displace .visible .entry kernel_gpu_shader_eval_displace( .param .u64 kernel_gpu_shader_eval_displace_param_0, .param .u64 kernel_gpu_shader_eval_displace_param_1, .param .u32 kernel_gpu_shader_eval_displace_param_2, .param .u32 kernel_gpu_shader_eval_displace_param_3 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot255[6448]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<179>; .reg .b16 %rs<3>; .reg .f32 %f<995>; .reg .b32 %r<255>; .reg .b64 %rd<270>; mov.u64 %SPL, __local_depot255; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd16, [kernel_gpu_shader_eval_displace_param_0]; ld.param.u32 %r109, [kernel_gpu_shader_eval_displace_param_2]; ld.param.u32 %r110, [kernel_gpu_shader_eval_displace_param_3]; mov.u32 %r111, %ntid.x; mov.u32 %r112, %ctaid.x; mov.u32 %r113, %tid.x; mad.lo.s32 %r1, %r112, %r111, %r113; setp.ge.s32 %p1, %r1, %r110; @%p1 bra $L__BB255_279; add.s32 %r114, %r1, %r109; cvta.to.global.u64 %rd18, %rd16; mul.wide.s32 %rd19, %r114, 16; add.s64 %rd20, %rd18, %rd19; ld.const.u64 %rd21, [kernel_params+2232]; cvta.to.global.u64 %rd22, %rd21; ld.global.u32 %r115, [%rd20+4]; cvt.s64.s32 %rd1, %r115; mul.wide.s32 %rd23, %r115, 16; add.s64 %rd24, %rd22, %rd23; ld.const.u64 %rd25, [kernel_params+2256]; cvta.to.global.u64 %rd26, %rd25; ld.global.u32 %r116, [%rd24+12]; mul.wide.u32 %rd27, %r116, 12; add.s64 %rd28, %rd26, %rd27; add.s32 %r117, %r116, 1; mul.wide.u32 %rd29, %r117, 12; add.s64 %rd30, %rd26, %rd29; add.s32 %r118, %r116, 2; mul.wide.u32 %rd31, %r118, 12; add.s64 %rd32, %rd26, %rd31; ld.global.f32 %f1, [%rd20+8]; mov.f32 %f234, 0f3F800000; sub.ftz.f32 %f235, %f234, %f1; ld.global.f32 %f2, [%rd20+12]; sub.ftz.f32 %f236, %f235, %f2; ld.global.f32 %f237, [%rd28]; ld.global.f32 %f238, [%rd28+4]; ld.global.f32 %f239, [%rd28+8]; ld.global.f32 %f240, [%rd30]; mul.ftz.f32 %f241, %f1, %f240; ld.global.f32 %f242, [%rd30+4]; mul.ftz.f32 %f243, %f1, %f242; ld.global.f32 %f244, [%rd30+8]; mul.ftz.f32 %f245, %f1, %f244; fma.rn.ftz.f32 %f246, %f236, %f237, %f241; fma.rn.ftz.f32 %f247, %f236, %f238, %f243; fma.rn.ftz.f32 %f248, %f236, %f239, %f245; ld.global.f32 %f249, [%rd32]; ld.global.f32 %f250, [%rd32+4]; ld.global.f32 %f251, [%rd32+8]; fma.rn.ftz.f32 %f3, %f2, %f249, %f246; fma.rn.ftz.f32 %f4, %f2, %f250, %f247; fma.rn.ftz.f32 %f5, %f2, %f251, %f248; ld.const.u64 %rd33, [kernel_params+2184]; cvta.to.global.u64 %rd34, %rd33; ld.global.u32 %r119, [%rd20]; cvt.s64.s32 %rd3, %r119; mul.wide.s32 %rd35, %r119, 4; add.s64 %rd36, %rd34, %rd35; ld.global.u32 %r2, [%rd36]; and.b32 %r120, %r2, 8; setp.eq.s32 %p2, %r120, 0; sub.ftz.f32 %f979, %f249, %f237; sub.ftz.f32 %f978, %f250, %f238; sub.ftz.f32 %f977, %f251, %f239; sub.ftz.f32 %f982, %f240, %f237; sub.ftz.f32 %f981, %f242, %f238; sub.ftz.f32 %f980, %f244, %f239; mul.ftz.f32 %f12, %f980, %f978; mul.ftz.f32 %f13, %f981, %f977; mul.ftz.f32 %f14, %f982, %f977; mul.ftz.f32 %f15, %f980, %f979; mul.ftz.f32 %f16, %f981, %f979; mul.ftz.f32 %f17, %f982, %f978; @%p2 bra $L__BB255_3; bra.uni $L__BB255_2; $L__BB255_3: sub.ftz.f32 %f259, %f13, %f12; sub.ftz.f32 %f260, %f15, %f14; mul.ftz.f32 %f261, %f260, %f260; fma.rn.ftz.f32 %f262, %f259, %f259, %f261; sub.ftz.f32 %f263, %f17, %f16; fma.rn.ftz.f32 %f264, %f263, %f263, %f262; sqrt.approx.ftz.f32 %f265, %f264; div.approx.ftz.f32 %f985, %f259, %f265; div.approx.ftz.f32 %f984, %f260, %f265; div.approx.ftz.f32 %f26, %f263, %f265; bra.uni $L__BB255_4; $L__BB255_2: sub.ftz.f32 %f252, %f12, %f13; sub.ftz.f32 %f253, %f14, %f15; mul.ftz.f32 %f254, %f253, %f253; fma.rn.ftz.f32 %f255, %f252, %f252, %f254; sub.ftz.f32 %f256, %f16, %f17; fma.rn.ftz.f32 %f257, %f256, %f256, %f255; sqrt.approx.ftz.f32 %f258, %f257; div.approx.ftz.f32 %f985, %f252, %f258; div.approx.ftz.f32 %f984, %f253, %f258; div.approx.ftz.f32 %f26, %f256, %f258; $L__BB255_4: ld.const.u64 %rd37, [kernel_params+2216]; cvta.to.global.u64 %rd38, %rd37; shl.b64 %rd39, %rd1, 2; add.s64 %rd40, %rd38, %rd39; ld.global.u32 %r121, [%rd40]; add.u64 %rd41, %SP, 1024; add.u64 %rd42, %SPL, 1024; st.local.v4.f32 [%rd42], {%f3, %f4, %f5, %f985}; add.s64 %rd4, %rd42, 16; st.local.v4.f32 [%rd42+16], {%f984, %f26, %f985, %f984}; mov.b32 %r122, %f26; mov.f32 %f266, 0f00000000; st.local.v4.f32 [%rd42+32], {%f26, %f266, %f266, %f266}; cvt.u32.u64 %r123, %rd1; setp.ne.s32 %p3, %r123, -1; cvt.u32.u64 %r124, %rd3; mov.b32 %r125, %f2; mov.b32 %r126, %f1; selp.u32 %r127, 1, 0, %p3; st.local.v4.u32 [%rd42+64], {%r127, %r126, %r125, %r124}; mov.u32 %r128, 1056964608; mov.u32 %r129, -1; st.local.v2.u32 [%rd42+80], {%r129, %r128}; mov.u32 %r130, 0; st.local.u32 [%rd42+88], %r130; ld.const.u64 %rd43, [kernel_params+2408]; cvta.to.global.u64 %rd44, %rd43; and.b32 %r254, %r121, 4194303; shl.b32 %r131, %r121, 5; cvt.u64.u32 %rd45, %r131; and.b64 %rd46, %rd45, 134217696; add.s64 %rd47, %rd44, %rd46; ld.global.u32 %r5, [%rd47+16]; or.b32 %r132, %r121, -2147483648; st.local.v4.u32 [%rd42+48], {%r132, %r5, %r130, %r123}; setp.eq.s32 %p4, %r124, -1; mov.b64 %rd48, {%r122, %r130}; cvt.u32.u64 %r133, %rd48; mov.b32 %f974, %r133; @%p4 bra $L__BB255_37; st.local.u32 [%rd4+40], %r2; and.b32 %r6, %r2, 2; setp.eq.s32 %p5, %r6, 0; @%p5 bra $L__BB255_12; ld.const.u64 %rd49, [kernel_params+2160]; cvta.to.global.u64 %rd50, %rd49; mul.lo.s64 %rd51, %rd3, 208; add.s64 %rd52, %rd50, %rd51; ld.const.u64 %rd53, [kernel_params+2176]; cvta.to.global.u64 %rd54, %rd53; ld.global.u32 %rd55, [%rd52+168]; ld.global.u32 %r134, [%rd52+152]; shl.b32 %r135, %r134, 1; cvt.rn.f32.s32 %f268, %r135; mul.ftz.f32 %f269, %f268, 0f3F000000; cvt.rzi.ftz.s32.f32 %r136, %f269; add.s32 %r137, %r135, -1; min.s32 %r138, %r136, %r137; cvt.rn.f32.s32 %f270, %r138; sub.ftz.f32 %f30, %f269, %f270; cvt.s64.s32 %rd56, %r138; add.s64 %rd57, %rd56, %rd55; shl.b64 %rd58, %rd57, 6; add.s64 %rd5, %rd54, %rd58; ld.global.v4.f32 {%f271, %f272, %f273, %f274}, [%rd5]; ld.global.v4.f32 {%f275, %f276, %f277, %f278}, [%rd5+64]; mul.ftz.f32 %f279, %f272, %f276; fma.rn.ftz.f32 %f280, %f271, %f275, %f279; mul.ftz.f32 %f281, %f274, %f278; fma.rn.ftz.f32 %f282, %f273, %f277, %f281; add.ftz.f32 %f39, %f280, %f282; setp.gt.ftz.f32 %p6, %f39, 0f3F7FDF3B; sub.ftz.f32 %f40, %f234, %f30; @%p6 bra $L__BB255_8; bra.uni $L__BB255_7; $L__BB255_8: mul.ftz.f32 %f337, %f271, %f40; fma.rn.ftz.f32 %f338, %f275, %f30, %f337; mul.ftz.f32 %f339, %f272, %f40; fma.rn.ftz.f32 %f340, %f276, %f30, %f339; mul.ftz.f32 %f341, %f273, %f40; fma.rn.ftz.f32 %f342, %f277, %f30, %f341; mul.ftz.f32 %f343, %f274, %f40; fma.rn.ftz.f32 %f344, %f278, %f30, %f343; mul.ftz.f32 %f345, %f340, %f340; fma.rn.ftz.f32 %f346, %f338, %f338, %f345; mul.ftz.f32 %f347, %f344, %f344; fma.rn.ftz.f32 %f348, %f342, %f342, %f347; add.ftz.f32 %f349, %f346, %f348; rsqrt.approx.ftz.f32 %f350, %f349; mul.ftz.f32 %f923, %f344, %f350; mul.ftz.f32 %f922, %f342, %f350; mul.ftz.f32 %f921, %f340, %f350; mul.ftz.f32 %f920, %f338, %f350; bra.uni $L__BB255_9; $L__BB255_37: mov.u64 %rd118, 0; st.local.u32 [%rd4+100], %r130; st.local.u64 [%rd4+104], %rd118; mov.f32 %f977, 0f00000000; st.local.v2.f32 [%rd4+112], {%f977, %f977}; st.local.u32 [%rd4+120], %r130; mov.f32 %f975, %f985; mov.f32 %f976, %f984; mov.f32 %f978, %f977; mov.f32 %f979, %f977; mov.f32 %f980, %f977; mov.f32 %f981, %f977; mov.f32 %f982, %f977; mov.f32 %f983, %f26; mov.f32 %f986, %f977; mov.f32 %f987, %f977; mov.f32 %f988, %f977; bra.uni $L__BB255_38; $L__BB255_7: mov.f32 %f284, 0fBF800000; max.ftz.f32 %f285, %f39, %f284; min.ftz.f32 %f287, %f285, %f234; abs.ftz.f32 %f288, %f287; sub.ftz.f32 %f289, %f234, %f288; mul.ftz.f32 %f290, %f289, 0f3F000000; sqrt.approx.ftz.f32 %f291, %f290; setp.gt.ftz.f32 %p7, %f288, 0f3F11EB85; selp.f32 %f292, %f291, %f288, %p7; mul.ftz.f32 %f293, %f292, %f292; mov.f32 %f294, 0f3C94D2E9; mov.f32 %f295, 0f3D53F941; fma.rn.ftz.f32 %f296, %f295, %f293, %f294; mov.f32 %f297, 0f3D3F841F; fma.rn.ftz.f32 %f298, %f296, %f293, %f297; mov.f32 %f299, 0f3D994929; fma.rn.ftz.f32 %f300, %f298, %f293, %f299; mov.f32 %f301, 0f3E2AAB94; fma.rn.ftz.f32 %f302, %f300, %f293, %f301; mul.ftz.f32 %f303, %f293, %f302; fma.rn.ftz.f32 %f304, %f303, %f292, %f292; add.ftz.f32 %f305, %f304, %f304; mov.f32 %f306, 0f3FC90FDB; sub.ftz.f32 %f307, %f306, %f304; selp.f32 %f308, %f305, %f307, %p7; setp.lt.ftz.f32 %p8, %f287, 0f00000000; mov.f32 %f309, 0f40490FDB; sub.ftz.f32 %f310, %f309, %f308; selp.f32 %f311, %f310, %f308, %p8; mul.ftz.f32 %f312, %f271, %f39; sub.ftz.f32 %f313, %f275, %f312; mul.ftz.f32 %f314, %f272, %f39; sub.ftz.f32 %f315, %f276, %f314; mul.ftz.f32 %f316, %f273, %f39; sub.ftz.f32 %f317, %f277, %f316; mul.ftz.f32 %f318, %f274, %f39; sub.ftz.f32 %f319, %f278, %f318; mul.ftz.f32 %f320, %f315, %f315; fma.rn.ftz.f32 %f321, %f313, %f313, %f320; mul.ftz.f32 %f322, %f319, %f319; fma.rn.ftz.f32 %f323, %f317, %f317, %f322; add.ftz.f32 %f324, %f321, %f323; rsqrt.approx.ftz.f32 %f325, %f324; mul.ftz.f32 %f326, %f313, %f325; mul.ftz.f32 %f327, %f315, %f325; mul.ftz.f32 %f328, %f317, %f325; mul.ftz.f32 %f329, %f319, %f325; mul.ftz.f32 %f330, %f30, %f311; cos.approx.ftz.f32 %f331, %f330; sin.approx.ftz.f32 %f332, %f330; mul.ftz.f32 %f333, %f326, %f332; mul.ftz.f32 %f334, %f327, %f332; mul.ftz.f32 %f335, %f328, %f332; mul.ftz.f32 %f336, %f329, %f332; fma.rn.ftz.f32 %f923, %f274, %f331, %f336; fma.rn.ftz.f32 %f922, %f273, %f331, %f335; fma.rn.ftz.f32 %f921, %f272, %f331, %f334; fma.rn.ftz.f32 %f920, %f271, %f331, %f333; $L__BB255_9: ld.global.v4.f32 {%f351, %f352, %f353, %f354}, [%rd5+16]; ld.global.v4.f32 {%f359, %f360, %f361, %f362}, [%rd5+80]; mul.ftz.f32 %f367, %f30, %f359; mul.ftz.f32 %f368, %f30, %f360; mul.ftz.f32 %f369, %f30, %f361; mul.ftz.f32 %f370, %f30, %f362; fma.rn.ftz.f32 %f948, %f40, %f351, %f367; fma.rn.ftz.f32 %f944, %f40, %f352, %f368; fma.rn.ftz.f32 %f940, %f40, %f353, %f369; fma.rn.ftz.f32 %f371, %f40, %f354, %f370; ld.global.v4.f32 {%f372, %f373, %f374, %f375}, [%rd5+32]; ld.global.v4.f32 {%f380, %f381, %f382, %f383}, [%rd5+96]; mul.ftz.f32 %f388, %f30, %f380; mul.ftz.f32 %f389, %f30, %f381; mul.ftz.f32 %f390, %f30, %f382; mul.ftz.f32 %f391, %f30, %f383; fma.rn.ftz.f32 %f392, %f40, %f372, %f388; fma.rn.ftz.f32 %f393, %f40, %f373, %f389; fma.rn.ftz.f32 %f394, %f40, %f374, %f390; fma.rn.ftz.f32 %f395, %f40, %f375, %f391; ld.global.v4.f32 {%f396, %f397, %f398, %f399}, [%rd5+48]; ld.global.v4.f32 {%f404, %f405, %f406, %f407}, [%rd5+112]; mul.ftz.f32 %f412, %f30, %f404; mul.ftz.f32 %f413, %f30, %f405; mul.ftz.f32 %f414, %f30, %f406; mul.ftz.f32 %f415, %f30, %f407; fma.rn.ftz.f32 %f416, %f40, %f396, %f412; fma.rn.ftz.f32 %f417, %f40, %f397, %f413; fma.rn.ftz.f32 %f418, %f40, %f398, %f414; fma.rn.ftz.f32 %f419, %f40, %f399, %f415; mul.ftz.f32 %f420, %f923, 0f3FB504F3; mul.ftz.f32 %f421, %f920, 0f3FB504F3; mul.ftz.f32 %f422, %f921, 0f3FB504F3; mul.ftz.f32 %f423, %f922, 0f3FB504F3; mul.ftz.f32 %f424, %f420, %f421; mul.ftz.f32 %f425, %f420, %f422; mul.ftz.f32 %f426, %f420, %f423; mul.ftz.f32 %f427, %f421, %f421; mul.ftz.f32 %f428, %f421, %f422; mul.ftz.f32 %f429, %f421, %f423; mul.ftz.f32 %f430, %f422, %f422; mul.ftz.f32 %f431, %f422, %f423; mul.ftz.f32 %f432, %f423, %f423; sub.ftz.f32 %f434, %f234, %f430; sub.ftz.f32 %f435, %f434, %f432; sub.ftz.f32 %f436, %f428, %f426; add.ftz.f32 %f437, %f425, %f429; add.ftz.f32 %f438, %f426, %f428; sub.ftz.f32 %f439, %f234, %f427; sub.ftz.f32 %f440, %f439, %f432; sub.ftz.f32 %f441, %f431, %f424; sub.ftz.f32 %f442, %f429, %f425; add.ftz.f32 %f443, %f424, %f431; sub.ftz.f32 %f444, %f439, %f430; mul.ftz.f32 %f445, %f436, %f394; fma.rn.ftz.f32 %f446, %f435, %f371, %f445; fma.rn.ftz.f32 %f951, %f437, %f417, %f446; mul.ftz.f32 %f447, %f436, %f395; fma.rn.ftz.f32 %f448, %f435, %f392, %f447; fma.rn.ftz.f32 %f950, %f437, %f418, %f448; mul.ftz.f32 %f449, %f436, %f416; fma.rn.ftz.f32 %f450, %f435, %f393, %f449; fma.rn.ftz.f32 %f949, %f437, %f419, %f450; mul.ftz.f32 %f451, %f440, %f394; fma.rn.ftz.f32 %f452, %f438, %f371, %f451; fma.rn.ftz.f32 %f947, %f441, %f417, %f452; mul.ftz.f32 %f453, %f440, %f395; fma.rn.ftz.f32 %f454, %f438, %f392, %f453; fma.rn.ftz.f32 %f946, %f441, %f418, %f454; mul.ftz.f32 %f455, %f440, %f416; fma.rn.ftz.f32 %f456, %f438, %f393, %f455; fma.rn.ftz.f32 %f945, %f441, %f419, %f456; mul.ftz.f32 %f457, %f443, %f394; fma.rn.ftz.f32 %f458, %f442, %f371, %f457; fma.rn.ftz.f32 %f943, %f444, %f417, %f458; mul.ftz.f32 %f459, %f443, %f395; fma.rn.ftz.f32 %f460, %f442, %f392, %f459; fma.rn.ftz.f32 %f942, %f444, %f418, %f460; mul.ftz.f32 %f461, %f443, %f416; fma.rn.ftz.f32 %f462, %f442, %f393, %f461; fma.rn.ftz.f32 %f941, %f444, %f419, %f462; st.local.v4.f32 [%rd4+128], {%f951, %f950, %f949, %f948}; st.local.v4.f32 [%rd4+144], {%f947, %f946, %f945, %f944}; st.local.v4.f32 [%rd4+160], {%f943, %f942, %f941, %f940}; mul.ftz.f32 %f463, %f946, %f941; mul.ftz.f32 %f65, %f942, %f945; sub.ftz.f32 %f926, %f463, %f65; mul.ftz.f32 %f67, %f942, %f949; mul.ftz.f32 %f464, %f950, %f941; sub.ftz.f32 %f925, %f67, %f464; mul.ftz.f32 %f69, %f950, %f945; mul.ftz.f32 %f465, %f946, %f949; sub.ftz.f32 %f924, %f69, %f465; mul.ftz.f32 %f466, %f947, %f925; fma.rn.ftz.f32 %f467, %f951, %f926, %f466; fma.rn.ftz.f32 %f930, %f943, %f924, %f467; setp.neu.ftz.f32 %p9, %f930, 0f00000000; mov.f32 %f927, %f941; mov.f32 %f928, %f946; mov.f32 %f929, %f951; @%p9 bra $L__BB255_11; add.ftz.f32 %f929, %f951, 0f322BCC77; add.ftz.f32 %f927, %f941, 0f322BCC77; add.ftz.f32 %f928, %f946, 0f322BCC77; mul.ftz.f32 %f468, %f928, %f927; sub.ftz.f32 %f926, %f468, %f65; mul.ftz.f32 %f469, %f950, %f927; sub.ftz.f32 %f925, %f67, %f469; mul.ftz.f32 %f470, %f928, %f949; sub.ftz.f32 %f924, %f69, %f470; mul.ftz.f32 %f471, %f947, %f925; fma.rn.ftz.f32 %f472, %f929, %f926, %f471; fma.rn.ftz.f32 %f473, %f943, %f924, %f472; setp.eq.ftz.f32 %p10, %f473, 0f00000000; selp.f32 %f930, 0f7F7FFFFF, %f473, %p10; $L__BB255_11: mul.ftz.f32 %f474, %f947, %f927; mul.ftz.f32 %f475, %f943, %f945; sub.ftz.f32 %f476, %f475, %f474; mul.ftz.f32 %f477, %f943, %f949; mul.ftz.f32 %f478, %f927, %f929; sub.ftz.f32 %f479, %f478, %f477; mul.ftz.f32 %f480, %f945, %f929; mul.ftz.f32 %f481, %f947, %f949; sub.ftz.f32 %f482, %f481, %f480; div.approx.ftz.f32 %f938, %f476, %f930; div.approx.ftz.f32 %f935, %f479, %f930; div.approx.ftz.f32 %f932, %f482, %f930; mul.ftz.f32 %f483, %f943, %f928; mul.ftz.f32 %f484, %f947, %f942; sub.ftz.f32 %f485, %f484, %f483; mul.ftz.f32 %f486, %f942, %f929; mul.ftz.f32 %f487, %f943, %f950; sub.ftz.f32 %f488, %f487, %f486; mul.ftz.f32 %f489, %f947, %f950; mul.ftz.f32 %f490, %f928, %f929; sub.ftz.f32 %f491, %f490, %f489; div.approx.ftz.f32 %f937, %f485, %f930; div.approx.ftz.f32 %f934, %f488, %f930; div.approx.ftz.f32 %f931, %f491, %f930; div.approx.ftz.f32 %f939, %f926, %f930; div.approx.ftz.f32 %f936, %f925, %f930; mul.ftz.f32 %f492, %f944, %f936; fma.rn.ftz.f32 %f493, %f948, %f939, %f492; div.approx.ftz.f32 %f933, %f924, %f930; fma.rn.ftz.f32 %f494, %f940, %f933, %f493; mul.ftz.f32 %f495, %f944, %f935; fma.rn.ftz.f32 %f496, %f948, %f938, %f495; fma.rn.ftz.f32 %f497, %f940, %f932, %f496; mul.ftz.f32 %f498, %f944, %f934; fma.rn.ftz.f32 %f499, %f948, %f937, %f498; fma.rn.ftz.f32 %f500, %f940, %f931, %f499; neg.ftz.f32 %f501, %f494; st.local.v4.f32 [%rd42+192], {%f939, %f936, %f933, %f501}; neg.ftz.f32 %f502, %f497; st.local.v4.f32 [%rd42+208], {%f938, %f935, %f932, %f502}; neg.ftz.f32 %f503, %f500; st.local.v4.f32 [%rd42+224], {%f937, %f934, %f931, %f503}; $L__BB255_12: and.b32 %r246, %r2, 4; mov.f32 %f986, 0f00000000; setp.ne.s32 %p11, %r246, 0; mov.f32 %f975, %f985; mov.f32 %f976, %f984; mov.f32 %f983, %f26; mov.f32 %f987, %f986; mov.f32 %f988, %f986; @%p11 bra $L__BB255_22; @%p5 bra $L__BB255_15; bra.uni $L__BB255_14; $L__BB255_15: ld.const.u64 %rd61, [kernel_params+2160]; cvta.to.global.u64 %rd62, %rd61; mul.lo.s64 %rd63, %rd3, 208; add.s64 %rd64, %rd62, %rd63; ld.global.v4.f32 {%f516, %f517, %f518, %f519}, [%rd64]; ld.global.v4.f32 {%f524, %f525, %f526, %f527}, [%rd64+16]; ld.global.v4.f32 {%f532, %f533, %f534, %f535}, [%rd64+32]; mul.ftz.f32 %f540, %f517, %f4; fma.rn.ftz.f32 %f541, %f3, %f516, %f540; fma.rn.ftz.f32 %f542, %f518, %f5, %f541; add.ftz.f32 %f952, %f519, %f542; mul.ftz.f32 %f543, %f525, %f4; fma.rn.ftz.f32 %f544, %f3, %f524, %f543; fma.rn.ftz.f32 %f545, %f526, %f5, %f544; add.ftz.f32 %f953, %f527, %f545; mul.ftz.f32 %f546, %f533, %f4; fma.rn.ftz.f32 %f547, %f3, %f532, %f546; fma.rn.ftz.f32 %f548, %f534, %f5, %f547; add.ftz.f32 %f954, %f535, %f548; bra.uni $L__BB255_16; $L__BB255_14: mul.ftz.f32 %f507, %f4, %f950; fma.rn.ftz.f32 %f508, %f3, %f951, %f507; fma.rn.ftz.f32 %f509, %f5, %f949, %f508; add.ftz.f32 %f952, %f948, %f509; mul.ftz.f32 %f510, %f4, %f946; fma.rn.ftz.f32 %f511, %f3, %f947, %f510; fma.rn.ftz.f32 %f512, %f5, %f945, %f511; add.ftz.f32 %f953, %f944, %f512; mul.ftz.f32 %f513, %f4, %f942; fma.rn.ftz.f32 %f514, %f3, %f943, %f513; fma.rn.ftz.f32 %f515, %f5, %f941, %f514; add.ftz.f32 %f954, %f940, %f515; $L__BB255_16: st.local.v2.f32 [%rd42], {%f952, %f953}; st.local.f32 [%rd42+8], %f954; @%p5 bra $L__BB255_18; bra.uni $L__BB255_17; $L__BB255_18: ld.const.u64 %rd67, [kernel_params+2160]; cvta.to.global.u64 %rd68, %rd67; mul.lo.s64 %rd69, %rd3, 208; add.s64 %rd70, %rd68, %rd69; ld.global.v4.f32 {%f562, %f563, %f564, %f565}, [%rd70+48]; ld.global.v4.f32 {%f569, %f570, %f571, %f572}, [%rd70+64]; ld.global.v4.f32 {%f576, %f577, %f578, %f579}, [%rd70+80]; mul.ftz.f32 %f583, %f569, %f984; fma.rn.ftz.f32 %f584, %f562, %f985, %f583; fma.rn.ftz.f32 %f585, %f576, %f974, %f584; mul.ftz.f32 %f586, %f570, %f984; fma.rn.ftz.f32 %f587, %f563, %f985, %f586; fma.rn.ftz.f32 %f588, %f577, %f974, %f587; mul.ftz.f32 %f589, %f571, %f984; fma.rn.ftz.f32 %f590, %f564, %f985, %f589; fma.rn.ftz.f32 %f591, %f578, %f974, %f590; mul.ftz.f32 %f592, %f588, %f588; fma.rn.ftz.f32 %f593, %f585, %f585, %f592; fma.rn.ftz.f32 %f594, %f591, %f591, %f593; sqrt.approx.ftz.f32 %f595, %f594; div.approx.ftz.f32 %f975, %f585, %f595; div.approx.ftz.f32 %f976, %f588, %f595; div.approx.ftz.f32 %f974, %f591, %f595; bra.uni $L__BB255_19; $L__BB255_17: mul.ftz.f32 %f549, %f984, %f938; fma.rn.ftz.f32 %f550, %f985, %f939, %f549; fma.rn.ftz.f32 %f551, %f974, %f937, %f550; mul.ftz.f32 %f552, %f984, %f935; fma.rn.ftz.f32 %f553, %f985, %f936, %f552; fma.rn.ftz.f32 %f554, %f974, %f934, %f553; mul.ftz.f32 %f555, %f984, %f932; fma.rn.ftz.f32 %f556, %f985, %f933, %f555; fma.rn.ftz.f32 %f557, %f974, %f931, %f556; mul.ftz.f32 %f558, %f554, %f554; fma.rn.ftz.f32 %f559, %f551, %f551, %f558; fma.rn.ftz.f32 %f560, %f557, %f557, %f559; sqrt.approx.ftz.f32 %f561, %f560; div.approx.ftz.f32 %f975, %f551, %f561; div.approx.ftz.f32 %f976, %f554, %f561; div.approx.ftz.f32 %f974, %f557, %f561; $L__BB255_19: st.local.v2.f32 [%rd42+24], {%f975, %f976}; st.local.f32 [%rd42+32], %f974; st.local.f32 [%rd42+12], %f975; st.local.f32 [%rd42+16], %f976; st.local.f32 [%rd42+20], %f974; @%p5 bra $L__BB255_21; bra.uni $L__BB255_20; $L__BB255_21: ld.const.u64 %rd75, [kernel_params+2160]; cvta.to.global.u64 %rd76, %rd75; mul.lo.s64 %rd77, %rd3, 208; add.s64 %rd78, %rd76, %rd77; ld.global.v4.f32 {%f602, %f603, %f604, %f605}, [%rd78]; ld.global.v4.f32 {%f609, %f610, %f611, %f612}, [%rd78+16]; ld.global.v4.f32 {%f616, %f617, %f618, %f619}, [%rd78+32]; mul.ftz.f32 %f623, %f603, 0f00000000; fma.rn.ftz.f32 %f624, %f602, 0f00000000, %f623; fma.rn.ftz.f32 %f988, %f604, 0f00000000, %f624; mul.ftz.f32 %f625, %f610, 0f00000000; fma.rn.ftz.f32 %f626, %f609, 0f00000000, %f625; fma.rn.ftz.f32 %f987, %f611, 0f00000000, %f626; mul.ftz.f32 %f627, %f617, 0f00000000; fma.rn.ftz.f32 %f628, %f616, 0f00000000, %f627; fma.rn.ftz.f32 %f986, %f618, 0f00000000, %f628; st.local.f32 [%rd42+36], %f988; st.local.f32 [%rd42+40], %f987; st.local.f32 [%rd42+44], %f986; mov.f32 %f983, %f974; bra.uni $L__BB255_22; $L__BB255_20: mul.ftz.f32 %f596, %f950, 0f00000000; fma.rn.ftz.f32 %f597, %f951, 0f00000000, %f596; fma.rn.ftz.f32 %f988, %f949, 0f00000000, %f597; mul.ftz.f32 %f598, %f946, 0f00000000; fma.rn.ftz.f32 %f599, %f947, 0f00000000, %f598; fma.rn.ftz.f32 %f987, %f945, 0f00000000, %f599; mul.ftz.f32 %f600, %f942, 0f00000000; fma.rn.ftz.f32 %f601, %f943, 0f00000000, %f600; fma.rn.ftz.f32 %f986, %f941, 0f00000000, %f601; st.local.f32 [%rd42+36], %f988; st.local.f32 [%rd42+40], %f987; st.local.f32 [%rd42+44], %f986; mov.f32 %f983, %f974; $L__BB255_22: @%p3 bra $L__BB255_24; bra.uni $L__BB255_23; $L__BB255_24: ld.const.u64 %rd269, [kernel_params+2232]; mul.wide.s32 %rd268, %r115, 16; cvta.to.global.u64 %rd267, %rd269; add.s64 %rd266, %rd267, %rd268; add.s64 %rd265, %rd266, 12; mov.f32 %f916, 0f3F800000; sub.ftz.f32 %f915, %f916, %f1; sub.ftz.f32 %f914, %f915, %f2; ld.global.v4.u32 {%r141, %r142, %r143, %r144}, [%rd265+-12]; ld.const.u64 %rd84, [kernel_params+2224]; cvta.to.global.u64 %rd85, %rd84; mul.wide.u32 %rd86, %r141, 12; add.s64 %rd87, %rd85, %rd86; mul.wide.u32 %rd88, %r142, 12; add.s64 %rd89, %rd85, %rd88; mul.wide.u32 %rd90, %r143, 12; add.s64 %rd91, %rd85, %rd90; ld.global.f32 %f638, [%rd87]; ld.global.f32 %f639, [%rd87+4]; ld.global.f32 %f640, [%rd87+8]; ld.global.f32 %f641, [%rd89]; mul.ftz.f32 %f642, %f1, %f641; ld.global.f32 %f643, [%rd89+4]; mul.ftz.f32 %f644, %f1, %f643; ld.global.f32 %f645, [%rd89+8]; mul.ftz.f32 %f646, %f1, %f645; fma.rn.ftz.f32 %f647, %f914, %f638, %f642; fma.rn.ftz.f32 %f648, %f914, %f639, %f644; fma.rn.ftz.f32 %f649, %f914, %f640, %f646; ld.global.f32 %f650, [%rd91]; ld.global.f32 %f651, [%rd91+4]; ld.global.f32 %f652, [%rd91+8]; fma.rn.ftz.f32 %f965, %f2, %f650, %f647; fma.rn.ftz.f32 %f966, %f2, %f651, %f648; fma.rn.ftz.f32 %f967, %f2, %f652, %f649; mul.ftz.f32 %f653, %f966, %f966; fma.rn.ftz.f32 %f654, %f965, %f965, %f653; fma.rn.ftz.f32 %f655, %f967, %f967, %f654; sqrt.approx.ftz.f32 %f159, %f655; setp.neu.ftz.f32 %p16, %f159, 0f00000000; @%p16 bra $L__BB255_25; bra.uni $L__BB255_26; $L__BB255_25: rcp.approx.ftz.f32 %f656, %f159; mul.ftz.f32 %f967, %f967, %f656; mul.ftz.f32 %f966, %f966, %f656; mul.ftz.f32 %f965, %f965, %f656; $L__BB255_26: setp.eq.ftz.f32 %p17, %f965, 0f00000000; setp.eq.ftz.f32 %p18, %f966, 0f00000000; and.pred %p19, %p17, %p18; setp.eq.ftz.f32 %p20, %f967, 0f00000000; and.pred %p21, %p20, %p19; selp.f32 %f983, %f26, %f967, %p21; selp.f32 %f984, %f984, %f966, %p21; selp.f32 %f985, %f985, %f965, %p21; st.local.f32 [%rd42+12], %f985; st.local.f32 [%rd42+16], %f984; st.local.f32 [%rd42+20], %f983; @%p11 bra $L__BB255_30; @%p5 bra $L__BB255_29; bra.uni $L__BB255_28; $L__BB255_29: ld.const.u64 %rd96, [kernel_params+2160]; cvta.to.global.u64 %rd97, %rd96; mul.lo.s64 %rd98, %rd3, 208; add.s64 %rd99, %rd97, %rd98; ld.global.v4.f32 {%f670, %f671, %f672, %f673}, [%rd99+48]; ld.global.v4.f32 {%f677, %f678, %f679, %f680}, [%rd99+64]; ld.global.v4.f32 {%f684, %f685, %f686, %f687}, [%rd99+80]; mul.ftz.f32 %f691, %f677, %f984; fma.rn.ftz.f32 %f692, %f670, %f985, %f691; fma.rn.ftz.f32 %f693, %f684, %f983, %f692; mul.ftz.f32 %f694, %f678, %f984; fma.rn.ftz.f32 %f695, %f671, %f985, %f694; fma.rn.ftz.f32 %f696, %f685, %f983, %f695; mul.ftz.f32 %f697, %f679, %f984; fma.rn.ftz.f32 %f698, %f672, %f985, %f697; fma.rn.ftz.f32 %f699, %f686, %f983, %f698; mul.ftz.f32 %f700, %f696, %f696; fma.rn.ftz.f32 %f701, %f693, %f693, %f700; fma.rn.ftz.f32 %f702, %f699, %f699, %f701; sqrt.approx.ftz.f32 %f703, %f702; div.approx.ftz.f32 %f985, %f693, %f703; div.approx.ftz.f32 %f984, %f696, %f703; div.approx.ftz.f32 %f983, %f699, %f703; st.local.f32 [%rd42+12], %f985; st.local.f32 [%rd42+16], %f984; st.local.f32 [%rd42+20], %f983; bra.uni $L__BB255_30; $L__BB255_23: mov.u64 %rd83, 0; st.local.u32 [%rd42+116], %r130; st.local.u64 [%rd42+120], %rd83; mov.f32 %f977, 0f00000000; st.local.v2.f32 [%rd42+128], {%f977, %f977}; st.local.u32 [%rd42+136], %r130; mov.f32 %f978, %f977; mov.f32 %f979, %f977; mov.f32 %f980, %f977; mov.f32 %f981, %f977; mov.f32 %f982, %f977; mov.f32 %f984, %f976; mov.f32 %f985, %f975; bra.uni $L__BB255_38; $L__BB255_28: mul.ftz.f32 %f657, %f984, %f938; fma.rn.ftz.f32 %f658, %f985, %f939, %f657; fma.rn.ftz.f32 %f659, %f983, %f937, %f658; mul.ftz.f32 %f660, %f984, %f935; fma.rn.ftz.f32 %f661, %f985, %f936, %f660; fma.rn.ftz.f32 %f662, %f983, %f934, %f661; mul.ftz.f32 %f663, %f984, %f932; fma.rn.ftz.f32 %f664, %f985, %f933, %f663; fma.rn.ftz.f32 %f665, %f983, %f931, %f664; mul.ftz.f32 %f666, %f662, %f662; fma.rn.ftz.f32 %f667, %f659, %f659, %f666; fma.rn.ftz.f32 %f668, %f665, %f665, %f667; sqrt.approx.ftz.f32 %f669, %f668; div.approx.ftz.f32 %f985, %f659, %f669; div.approx.ftz.f32 %f984, %f662, %f669; div.approx.ftz.f32 %f983, %f665, %f669; st.local.f32 [%rd42+12], %f985; st.local.f32 [%rd42+16], %f984; st.local.f32 [%rd42+20], %f983; $L__BB255_30: st.local.f32 [%rd42+116], %f982; st.local.f32 [%rd42+120], %f981; st.local.f32 [%rd42+124], %f980; st.local.v2.f32 [%rd42+128], {%f979, %f978}; st.local.f32 [%rd42+136], %f977; @%p11 bra $L__BB255_38; @%p5 bra $L__BB255_33; bra.uni $L__BB255_32; $L__BB255_33: ld.const.u64 %rd104, [kernel_params+2160]; cvta.to.global.u64 %rd105, %rd104; mul.lo.s64 %rd106, %rd3, 208; add.s64 %rd107, %rd105, %rd106; ld.global.v4.f32 {%f710, %f711, %f712, %f713}, [%rd107]; ld.global.v4.f32 {%f717, %f718, %f719, %f720}, [%rd107+16]; ld.global.v4.f32 {%f724, %f725, %f726, %f727}, [%rd107+32]; mul.ftz.f32 %f731, %f711, %f981; fma.rn.ftz.f32 %f732, %f982, %f710, %f731; fma.rn.ftz.f32 %f971, %f712, %f980, %f732; mul.ftz.f32 %f733, %f718, %f981; fma.rn.ftz.f32 %f734, %f982, %f717, %f733; fma.rn.ftz.f32 %f972, %f719, %f980, %f734; mul.ftz.f32 %f735, %f725, %f981; fma.rn.ftz.f32 %f736, %f982, %f724, %f735; fma.rn.ftz.f32 %f980, %f726, %f980, %f736; bra.uni $L__BB255_34; $L__BB255_32: mul.ftz.f32 %f704, %f981, %f950; fma.rn.ftz.f32 %f705, %f982, %f951, %f704; fma.rn.ftz.f32 %f971, %f980, %f949, %f705; mul.ftz.f32 %f706, %f981, %f946; fma.rn.ftz.f32 %f707, %f982, %f947, %f706; fma.rn.ftz.f32 %f972, %f980, %f945, %f707; mul.ftz.f32 %f708, %f981, %f942; fma.rn.ftz.f32 %f709, %f982, %f943, %f708; fma.rn.ftz.f32 %f980, %f980, %f941, %f709; $L__BB255_34: st.local.f32 [%rd42+116], %f971; st.local.f32 [%rd42+120], %f972; st.local.f32 [%rd42+124], %f980; @%p5 bra $L__BB255_36; bra.uni $L__BB255_35; $L__BB255_36: ld.const.u64 %rd112, [kernel_params+2160]; cvta.to.global.u64 %rd113, %rd112; mul.lo.s64 %rd114, %rd3, 208; add.s64 %rd115, %rd113, %rd114; ld.global.v4.f32 {%f743, %f744, %f745, %f746}, [%rd115]; ld.global.v4.f32 {%f750, %f751, %f752, %f753}, [%rd115+16]; ld.global.v4.f32 {%f757, %f758, %f759, %f760}, [%rd115+32]; mul.ftz.f32 %f764, %f744, %f978; fma.rn.ftz.f32 %f765, %f979, %f743, %f764; fma.rn.ftz.f32 %f193, %f745, %f977, %f765; mul.ftz.f32 %f766, %f751, %f978; fma.rn.ftz.f32 %f767, %f979, %f750, %f766; fma.rn.ftz.f32 %f194, %f752, %f977, %f767; mul.ftz.f32 %f768, %f758, %f978; fma.rn.ftz.f32 %f769, %f979, %f757, %f768; fma.rn.ftz.f32 %f977, %f759, %f977, %f769; st.local.v2.f32 [%rd42+128], {%f193, %f194}; st.local.f32 [%rd42+136], %f977; mov.f32 %f978, %f194; mov.f32 %f979, %f193; mov.f32 %f981, %f972; mov.f32 %f982, %f971; bra.uni $L__BB255_38; $L__BB255_35: mul.ftz.f32 %f737, %f978, %f950; fma.rn.ftz.f32 %f738, %f979, %f951, %f737; fma.rn.ftz.f32 %f190, %f977, %f949, %f738; mul.ftz.f32 %f739, %f978, %f946; fma.rn.ftz.f32 %f740, %f979, %f947, %f739; fma.rn.ftz.f32 %f191, %f977, %f945, %f740; mul.ftz.f32 %f741, %f978, %f942; fma.rn.ftz.f32 %f742, %f979, %f943, %f741; fma.rn.ftz.f32 %f977, %f977, %f941, %f742; st.local.v2.f32 [%rd42+128], {%f190, %f191}; st.local.f32 [%rd42+136], %f977; mov.f32 %f978, %f191; mov.f32 %f979, %f190; mov.f32 %f981, %f972; mov.f32 %f982, %f971; $L__BB255_38: cvt.s64.s32 %rd263, %r115; cvt.u32.u64 %r247, %rd263; setp.eq.s32 %p27, %r247, -1; @%p27 bra $L__BB255_41; mul.ftz.f32 %f779, %f976, %f987; fma.rn.ftz.f32 %f780, %f975, %f988, %f779; fma.rn.ftz.f32 %f781, %f974, %f986, %f780; setp.geu.ftz.f32 %p28, %f781, 0f00000000; @%p28 bra $L__BB255_41; or.b32 %r150, %r5, 1; st.local.u32 [%rd42+52], %r150; neg.ftz.f32 %f782, %f974; st.local.f32 [%rd42+32], %f782; neg.ftz.f32 %f783, %f985; st.local.f32 [%rd42+12], %f783; neg.ftz.f32 %f784, %f976; neg.ftz.f32 %f785, %f975; neg.ftz.f32 %f786, %f983; neg.ftz.f32 %f787, %f984; st.local.v4.f32 [%rd42+16], {%f787, %f786, %f785, %f784}; neg.ftz.f32 %f788, %f982; st.local.f32 [%rd42+116], %f788; neg.ftz.f32 %f789, %f981; st.local.f32 [%rd42+120], %f789; neg.ftz.f32 %f790, %f980; st.local.f32 [%rd42+124], %f790; neg.ftz.f32 %f791, %f978; neg.ftz.f32 %f792, %f979; st.local.v2.f32 [%rd42+128], {%f792, %f791}; neg.ftz.f32 %f793, %f977; st.local.f32 [%rd42+136], %f793; $L__BB255_41: st.local.u32 [%rd42+92], %r130; mov.f32 %f994, 0f00000000; st.local.v4.f32 [%rd42+96], {%f994, %f994, %f994, %f994}; st.local.u32 [%rd42+112], %r130; ld.local.v4.f32 {%f795, %f796, %f797, %f798}, [%rd42]; st.local.u32 [%rd42+260], %r130; st.local.u32 [%rd42+264], %r130; ld.const.u64 %rd123, [kernel_params+2400]; cvta.to.global.u64 %rd6, %rd123; add.s64 %rd7, %rd42, 48; add.s64 %rd8, %rd42, 256; add.u64 %rd124, %SP, 0; add.u64 %rd9, %SPL, 0; bra.uni $L__BB255_42; $L__BB255_192: add.s32 %r163, %r7, 1; { // callseq 1531, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r163; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1531 $L__BB255_42: mov.u32 %r7, %r254; mul.wide.s32 %rd125, %r7, 16; add.s64 %rd10, %rd6, %rd125; ld.global.v4.u32 {%r153, %r154, %r155, %r156}, [%rd10]; mul.wide.u32 %rd126, %r154, 4; add.s64 %rd11, %rd9, %rd126; mul.wide.u32 %rd127, %r155, 4; add.s64 %rd12, %rd9, %rd127; add.s32 %r254, %r7, 1; setp.gt.s32 %p29, %r153, 41; @%p29 bra $L__BB255_112; setp.gt.s32 %p104, %r153, 15; @%p104 bra $L__BB255_70; bra.uni $L__BB255_44; $L__BB255_70: setp.gt.s32 %p105, %r153, 37; @%p105 bra $L__BB255_108; bra.uni $L__BB255_71; $L__BB255_108: add.s32 %r157, %r153, -40; setp.lt.u32 %p106, %r157, 2; @%p106 bra $L__BB255_42; setp.eq.s32 %p107, %r153, 38; @%p107 bra $L__BB255_244; setp.eq.s32 %p108, %r153, 39; @%p108 bra $L__BB255_111; bra.uni $L__BB255_266; $L__BB255_111: mov.u32 %r198, 39; { // callseq 1575, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r198, %r154, %r155, %r156}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1575 bra.uni $L__BB255_42; $L__BB255_112: setp.gt.s32 %p30, %r153, 84; @%p30 bra $L__BB255_187; bra.uni $L__BB255_113; $L__BB255_187: setp.gt.s32 %p31, %r153, 91; @%p31 bra $L__BB255_197; bra.uni $L__BB255_188; $L__BB255_197: setp.lt.s32 %p32, %r153, 94; @%p32 bra $L__BB255_42; setp.gt.s32 %p33, %r153, 95; @%p33 bra $L__BB255_202; bra.uni $L__BB255_199; $L__BB255_202: setp.eq.s32 %p34, %r153, 98; @%p34 bra $L__BB255_265; setp.eq.s32 %p35, %r153, 97; @%p35 bra $L__BB255_207; setp.ne.s32 %p36, %r153, 96; @%p36 bra $L__BB255_266; { // callseq 1525, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1525 bra.uni $L__BB255_42; $L__BB255_44: setp.gt.s32 %p138, %r153, 4; @%p138 bra $L__BB255_49; add.s32 %r158, %r153, -3; setp.lt.u32 %p154, %r158, 2; @%p154 bra $L__BB255_42; setp.eq.s32 %p155, %r153, 1; mov.u32 %r254, %r156; @%p155 bra $L__BB255_42; setp.eq.s32 %p156, %r153, 2; @%p156 bra $L__BB255_48; bra.uni $L__BB255_266; $L__BB255_48: mov.u32 %r231, 2; add.s32 %r232, %r7, 1; { // callseq 1601, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r231; st.param.b32 [param0+4], %r154; st.param.b32 [param0+8], %r155; st.param.b32 [param0+12], %r156; .param .b32 param1; st.param.b32 [param1+0], %r232; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj56EL10ShaderType2EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1601 bra.uni $L__BB255_42; $L__BB255_113: setp.gt.s32 %p46, %r153, 62; @%p46 bra $L__BB255_150; bra.uni $L__BB255_114; $L__BB255_150: setp.gt.s32 %p47, %r153, 73; @%p47 bra $L__BB255_169; setp.gt.s32 %p62, %r153, 67; @%p62 bra $L__BB255_160; bra.uni $L__BB255_152; $L__BB255_160: setp.gt.s32 %p63, %r153, 70; @%p63 bra $L__BB255_165; setp.eq.s32 %p67, %r153, 68; @%p67 bra $L__BB255_227; setp.eq.s32 %p68, %r153, 69; @%p68 bra $L__BB255_226; setp.eq.s32 %p69, %r153, 70; @%p69 bra $L__BB255_164; bra.uni $L__BB255_266; $L__BB255_164: mov.u32 %r173, 70; { // callseq 1546, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r173; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1546 bra.uni $L__BB255_42; $L__BB255_71: setp.gt.s32 %p109, %r153, 26; @%p109 bra $L__BB255_90; setp.gt.s32 %p124, %r153, 20; @%p124 bra $L__BB255_81; bra.uni $L__BB255_73; $L__BB255_81: setp.gt.s32 %p125, %r153, 23; @%p125 bra $L__BB255_86; setp.eq.s32 %p129, %r153, 21; @%p129 bra $L__BB255_255; setp.eq.s32 %p130, %r153, 22; @%p130 bra $L__BB255_254; setp.eq.s32 %p131, %r153, 23; @%p131 bra $L__BB255_85; bra.uni $L__BB255_266; $L__BB255_85: mov.u32 %r212, 23; add.s32 %r213, %r7, 1; { // callseq 1590, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r212; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r213; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1590 bra.uni $L__BB255_42; $L__BB255_49: setp.gt.s32 %p139, %r153, 7; @%p139 bra $L__BB255_54; setp.eq.s32 %p151, %r153, 7; @%p151 bra $L__BB255_42; setp.eq.s32 %p152, %r153, 5; @%p152 bra $L__BB255_264; setp.eq.s32 %p153, %r153, 6; @%p153 bra $L__BB255_53; bra.uni $L__BB255_266; $L__BB255_53: ld.local.f32 %f834, [%rd11]; ld.local.f32 %f835, [%rd11+4]; ld.local.f32 %f836, [%rd11+8]; st.local.f32 [%rd8+12], %f834; st.local.f32 [%rd8+16], %f835; st.local.f32 [%rd8+20], %f836; bra.uni $L__BB255_42; $L__BB255_114: setp.gt.s32 %p76, %r153, 51; @%p76 bra $L__BB255_132; bra.uni $L__BB255_115; $L__BB255_132: setp.gt.s32 %p77, %r153, 56; @%p77 bra $L__BB255_141; bra.uni $L__BB255_133; $L__BB255_141: setp.gt.s32 %p78, %r153, 59; @%p78 bra $L__BB255_146; setp.eq.s32 %p82, %r153, 57; @%p82 bra $L__BB255_234; setp.eq.s32 %p83, %r153, 58; @%p83 bra $L__BB255_233; setp.eq.s32 %p84, %r153, 59; @%p84 bra $L__BB255_145; bra.uni $L__BB255_266; $L__BB255_145: add.s32 %r187, %r7, 1; { // callseq 1557, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r187; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj56EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1557 bra.uni $L__BB255_42; $L__BB255_188: setp.gt.s32 %p39, %r153, 87; @%p39 bra $L__BB255_193; bra.uni $L__BB255_189; $L__BB255_193: setp.eq.s32 %p40, %r153, 90; @%p40 bra $L__BB255_42; setp.eq.s32 %p41, %r153, 88; @%p41 bra $L__BB255_208; setp.eq.s32 %p42, %r153, 89; @%p42 bra $L__BB255_196; bra.uni $L__BB255_266; $L__BB255_196: mov.u32 %r161, 89; { // callseq 1529, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r161; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z11svm_node_aoILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1529 bra.uni $L__BB255_42; $L__BB255_90: setp.gt.s32 %p110, %r153, 31; @%p110 bra $L__BB255_99; bra.uni $L__BB255_91; $L__BB255_99: setp.gt.s32 %p111, %r153, 34; @%p111 bra $L__BB255_104; setp.eq.s32 %p115, %r153, 32; @%p115 bra $L__BB255_248; setp.eq.s32 %p116, %r153, 33; @%p116 bra $L__BB255_247; setp.eq.s32 %p117, %r153, 34; @%p117 bra $L__BB255_103; bra.uni $L__BB255_266; $L__BB255_103: { // callseq 1580, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1580 bra.uni $L__BB255_42; $L__BB255_169: setp.gt.s32 %p48, %r153, 78; @%p48 bra $L__BB255_178; bra.uni $L__BB255_170; $L__BB255_178: setp.gt.s32 %p49, %r153, 81; @%p49 bra $L__BB255_183; setp.eq.s32 %p53, %r153, 79; @%p53 bra $L__BB255_214; setp.eq.s32 %p54, %r153, 80; @%p54 bra $L__BB255_213; setp.eq.s32 %p55, %r153, 81; @%p55 bra $L__BB255_182; bra.uni $L__BB255_266; $L__BB255_182: mov.u32 %r167, 81; { // callseq 1537, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r167; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1537 bra.uni $L__BB255_42; $L__BB255_115: setp.gt.s32 %p91, %r153, 46; @%p91 bra $L__BB255_124; setp.gt.s32 %p98, %r153, 43; @%p98 bra $L__BB255_120; bra.uni $L__BB255_117; $L__BB255_120: setp.eq.s32 %p99, %r153, 44; @%p99 bra $L__BB255_242; setp.eq.s32 %p100, %r153, 45; @%p100 bra $L__BB255_241; setp.eq.s32 %p101, %r153, 46; @%p101 bra $L__BB255_123; bra.uni $L__BB255_266; $L__BB255_123: { // callseq 1570, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1570 bra.uni $L__BB255_42; $L__BB255_54: setp.gt.s32 %p140, %r153, 11; @%p140 bra $L__BB255_63; setp.gt.s32 %p146, %r153, 9; @%p146 bra $L__BB255_60; setp.eq.s32 %p149, %r153, 8; @%p149 bra $L__BB255_263; setp.eq.s32 %p150, %r153, 9; @%p150 bra $L__BB255_58; bra.uni $L__BB255_266; $L__BB255_58: ld.local.f32 %f833, [%rd12]; setp.gtu.ftz.f32 %p162, %f833, 0f00000000; @%p162 bra $L__BB255_42; add.s32 %r229, %r154, %r7; add.s32 %r254, %r229, 1; bra.uni $L__BB255_42; $L__BB255_73: setp.gt.s32 %p132, %r153, 17; @%p132 bra $L__BB255_77; bra.uni $L__BB255_74; $L__BB255_77: setp.eq.s32 %p133, %r153, 18; @%p133 bra $L__BB255_257; setp.eq.s32 %p134, %r153, 19; @%p134 bra $L__BB255_256; setp.eq.s32 %p135, %r153, 20; @%p135 bra $L__BB255_80; bra.uni $L__BB255_266; $L__BB255_80: ld.local.f32 %f819, [%rd11]; ld.local.v4.f32 {%f820, %f821, %f822, %f823}, [%rd42]; ld.local.f32 %f827, [%rd11+4]; ld.local.f32 %f828, [%rd11+8]; add.ftz.f32 %f829, %f828, %f822; add.ftz.f32 %f830, %f827, %f821; add.ftz.f32 %f831, %f819, %f820; st.local.v2.f32 [%rd42], {%f831, %f830}; st.local.f32 [%rd42+8], %f829; bra.uni $L__BB255_42; $L__BB255_91: setp.gt.s32 %p118, %r153, 28; @%p118 bra $L__BB255_95; bra.uni $L__BB255_92; $L__BB255_95: setp.eq.s32 %p119, %r153, 29; @%p119 bra $L__BB255_250; setp.eq.s32 %p120, %r153, 30; @%p120 bra $L__BB255_249; setp.eq.s32 %p121, %r153, 31; @%p121 bra $L__BB255_98; bra.uni $L__BB255_266; $L__BB255_98: mov.u32 %r204, 31; add.s32 %r205, %r7, 1; { // callseq 1582, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b32 param1; st.param.b32 [param1+0], %r130; .param .b64 param2; st.param.b64 [param2+0], %rd124; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r204; st.param.b32 [param3+4], %r154; st.param.b32 [param3+8], %r155; st.param.b32 [param3+12], %r156; .param .b32 param4; st.param.b32 [param4+0], %r205; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1582 bra.uni $L__BB255_42; $L__BB255_189: setp.eq.s32 %p43, %r153, 85; @%p43 bra $L__BB255_210; setp.eq.s32 %p44, %r153, 86; @%p44 bra $L__BB255_209; setp.eq.s32 %p45, %r153, 87; @%p45 bra $L__BB255_192; bra.uni $L__BB255_266; $L__BB255_152: setp.gt.s32 %p70, %r153, 64; @%p70 bra $L__BB255_156; bra.uni $L__BB255_153; $L__BB255_156: setp.eq.s32 %p71, %r153, 65; @%p71 bra $L__BB255_229; setp.eq.s32 %p72, %r153, 66; @%p72 bra $L__BB255_228; setp.eq.s32 %p73, %r153, 67; @%p73 bra $L__BB255_159; bra.uni $L__BB255_266; $L__BB255_159: mov.u32 %r177, 67; { // callseq 1549, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r177, %r154, %r155, %r156}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1549 bra.uni $L__BB255_42; $L__BB255_133: setp.gt.s32 %p85, %r153, 53; @%p85 bra $L__BB255_137; bra.uni $L__BB255_134; $L__BB255_137: setp.eq.s32 %p86, %r153, 54; @%p86 bra $L__BB255_236; setp.eq.s32 %p87, %r153, 55; @%p87 bra $L__BB255_235; setp.eq.s32 %p88, %r153, 56; @%p88 bra $L__BB255_140; bra.uni $L__BB255_266; $L__BB255_140: mov.u32 %r191, 56; { // callseq 1560, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r191; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1560 bra.uni $L__BB255_42; $L__BB255_170: setp.gt.s32 %p56, %r153, 75; @%p56 bra $L__BB255_174; bra.uni $L__BB255_171; $L__BB255_174: setp.eq.s32 %p57, %r153, 76; @%p57 bra $L__BB255_217; setp.eq.s32 %p58, %r153, 77; @%p58 bra $L__BB255_215; setp.eq.s32 %p59, %r153, 78; @%p59 bra $L__BB255_177; bra.uni $L__BB255_266; $L__BB255_177: add.s32 %r169, %r7, 1; { // callseq 1540, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r169; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1540 bra.uni $L__BB255_42; $L__BB255_124: setp.gt.s32 %p92, %r153, 48; @%p92 bra $L__BB255_128; bra.uni $L__BB255_125; $L__BB255_128: setp.eq.s32 %p93, %r153, 49; @%p93 bra $L__BB255_239; setp.eq.s32 %p94, %r153, 50; @%p94 bra $L__BB255_238; setp.eq.s32 %p95, %r153, 51; @%p95 bra $L__BB255_131; bra.uni $L__BB255_266; $L__BB255_131: { // callseq 1565, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1565 bra.uni $L__BB255_42; $L__BB255_63: setp.gt.s32 %p141, %r153, 13; @%p141 bra $L__BB255_67; setp.eq.s32 %p144, %r153, 12; @%p144 bra $L__BB255_260; setp.eq.s32 %p145, %r153, 13; @%p145 bra $L__BB255_66; bra.uni $L__BB255_266; $L__BB255_66: mov.u32 %r225, 13; add.s32 %r226, %r7, 1; { // callseq 1597, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b32 param1; st.param.b32 [param1+0], %r130; .param .b64 param2; st.param.b64 [param2+0], %rd124; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r225; st.param.b32 [param3+4], %r154; st.param.b32 [param3+8], %r155; st.param.b32 [param3+12], %r156; .param .b32 param4; st.param.b32 [param4+0], %r226; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1597 bra.uni $L__BB255_42; $L__BB255_86: setp.eq.s32 %p126, %r153, 24; @%p126 bra $L__BB255_253; setp.eq.s32 %p127, %r153, 25; @%p127 bra $L__BB255_252; setp.eq.s32 %p128, %r153, 26; @%p128 bra $L__BB255_89; bra.uni $L__BB255_266; $L__BB255_89: mov.u32 %r209, 26; { // callseq 1587, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r209; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z17svm_node_set_bumpILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1587 bra.uni $L__BB255_42; $L__BB255_104: setp.eq.s32 %p112, %r153, 35; @%p112 bra $L__BB255_246; setp.eq.s32 %p113, %r153, 36; @%p113 bra $L__BB255_245; setp.eq.s32 %p114, %r153, 37; @%p114 bra $L__BB255_107; bra.uni $L__BB255_266; $L__BB255_107: mov.u32 %r199, 37; { // callseq 1577, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r199, %r154, %r155, %r156}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1577 bra.uni $L__BB255_42; $L__BB255_165: setp.eq.s32 %p64, %r153, 71; @%p64 bra $L__BB255_225; setp.eq.s32 %p65, %r153, 72; @%p65 bra $L__BB255_224; setp.eq.s32 %p66, %r153, 73; @%p66 bra $L__BB255_168; bra.uni $L__BB255_266; $L__BB255_168: add.s32 %r171, %r7, 1; { // callseq 1543, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r171; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1543 bra.uni $L__BB255_42; $L__BB255_146: setp.eq.s32 %p79, %r153, 60; @%p79 bra $L__BB255_232; setp.eq.s32 %p80, %r153, 61; @%p80 bra $L__BB255_231; setp.eq.s32 %p81, %r153, 62; @%p81 bra $L__BB255_149; bra.uni $L__BB255_266; $L__BB255_149: mov.u32 %r182, 62; add.s32 %r183, %r7, 1; { // callseq 1554, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r182; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r183; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1554 bra.uni $L__BB255_42; $L__BB255_183: setp.eq.s32 %p50, %r153, 82; @%p50 bra $L__BB255_212; setp.eq.s32 %p51, %r153, 83; @%p51 bra $L__BB255_211; setp.eq.s32 %p52, %r153, 84; @%p52 bra $L__BB255_186; bra.uni $L__BB255_266; $L__BB255_186: { // callseq 1534, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1534 bra.uni $L__BB255_42; $L__BB255_74: setp.eq.s32 %p136, %r153, 16; @%p136 bra $L__BB255_258; setp.eq.s32 %p137, %r153, 17; @%p137 bra $L__BB255_76; bra.uni $L__BB255_266; $L__BB255_76: { // callseq 1595, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; .param .b32 param4; st.param.b32 [param4+0], %r156; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1595 bra.uni $L__BB255_42; $L__BB255_92: setp.eq.s32 %p122, %r153, 27; @%p122 bra $L__BB255_251; setp.eq.s32 %p123, %r153, 28; @%p123 bra $L__BB255_94; bra.uni $L__BB255_266; $L__BB255_94: mov.u32 %r207, 28; { // callseq 1585, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r207; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1585 bra.uni $L__BB255_42; $L__BB255_199: setp.eq.s32 %p37, %r153, 94; @%p37 bra $L__BB255_206; setp.eq.s32 %p38, %r153, 95; @%p38 bra $L__BB255_201; bra.uni $L__BB255_266; $L__BB255_201: { // callseq 1526, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1526 bra.uni $L__BB255_42; $L__BB255_117: setp.eq.s32 %p102, %r153, 42; @%p102 bra $L__BB255_243; setp.eq.s32 %p103, %r153, 43; @%p103 bra $L__BB255_119; bra.uni $L__BB255_266; $L__BB255_119: add.s32 %r197, %r7, 1; { // callseq 1573, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r197; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1573 bra.uni $L__BB255_42; $L__BB255_60: setp.eq.s32 %p147, %r153, 10; @%p147 bra $L__BB255_261; setp.eq.s32 %p148, %r153, 11; @%p148 bra $L__BB255_62; bra.uni $L__BB255_266; $L__BB255_62: { // callseq 1599, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1599 bra.uni $L__BB255_42; $L__BB255_153: setp.eq.s32 %p74, %r153, 63; @%p74 bra $L__BB255_230; setp.eq.s32 %p75, %r153, 64; @%p75 bra $L__BB255_155; bra.uni $L__BB255_266; $L__BB255_155: mov.u32 %r179, 64; add.s32 %r180, %r7, 1; { // callseq 1552, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r179; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r180; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1552 bra.uni $L__BB255_42; $L__BB255_134: setp.eq.s32 %p89, %r153, 52; @%p89 bra $L__BB255_237; setp.eq.s32 %p90, %r153, 53; @%p90 bra $L__BB255_136; bra.uni $L__BB255_266; $L__BB255_136: { // callseq 1563, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1563 bra.uni $L__BB255_42; $L__BB255_171: setp.eq.s32 %p60, %r153, 74; @%p60 bra $L__BB255_223; setp.eq.s32 %p61, %r153, 75; @%p61 bra $L__BB255_173; bra.uni $L__BB255_266; $L__BB255_173: { // callseq 1541, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1541 bra.uni $L__BB255_42; $L__BB255_125: setp.eq.s32 %p96, %r153, 47; @%p96 bra $L__BB255_240; setp.eq.s32 %p97, %r153, 48; @%p97 bra $L__BB255_127; bra.uni $L__BB255_266; $L__BB255_127: { // callseq 1568, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1568 bra.uni $L__BB255_42; $L__BB255_67: setp.eq.s32 %p142, %r153, 14; @%p142 bra $L__BB255_259; setp.eq.s32 %p143, %r153, 15; @%p143 bra $L__BB255_69; bra.uni $L__BB255_266; $L__BB255_69: ld.global.v4.u32 {%r218, %r219, %r220, %r221}, [%rd10+16]; add.s32 %r254, %r7, 2; st.local.u32 [%rd11], %r219; st.local.u32 [%rd11+4], %r220; st.local.u32 [%rd11+8], %r221; bra.uni $L__BB255_42; $L__BB255_210: add.s32 %r165, %r7, 1; { // callseq 1533, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r165; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1533 bra.uni $L__BB255_42; $L__BB255_209: add.s32 %r164, %r7, 1; { // callseq 1532, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r164; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1532 bra.uni $L__BB255_42; $L__BB255_255: mov.u32 %r216, 21; { // callseq 1592, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r216; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z21svm_node_displacementILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1592 bra.uni $L__BB255_42; $L__BB255_254: mov.u32 %r214, 22; add.s32 %r215, %r7, 1; { // callseq 1591, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r214; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; .param .b32 param3; st.param.b32 [param3+0], %r215; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj56EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1591 bra.uni $L__BB255_42; $L__BB255_248: mov.u32 %r201, 32; add.s32 %r202, %r7, 1; { // callseq 1581, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b32 param1; st.param.b32 [param1+0], %r130; .param .b64 param2; st.param.b64 [param2+0], %rd124; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r201; st.param.b32 [param3+4], %r154; st.param.b32 [param3+8], %r155; st.param.b32 [param3+12], %r156; .param .b32 param4; st.param.b32 [param4+0], %r202; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1581 bra.uni $L__BB255_42; $L__BB255_247: ld.local.f32 %f816, [%rd11]; ld.local.f32 %f817, [%rd11+4]; ld.local.f32 %f818, [%rd11+8]; st.local.f32 [%rd42+12], %f816; st.local.f32 [%rd42+16], %f817; st.local.f32 [%rd42+20], %f818; st.local.f32 [%rd12], %f816; st.local.f32 [%rd12+4], %f817; st.local.f32 [%rd12+8], %f818; bra.uni $L__BB255_42; $L__BB255_257: { // callseq 1594, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1594 bra.uni $L__BB255_42; $L__BB255_256: { // callseq 1593, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1593 bra.uni $L__BB255_42; $L__BB255_250: { // callseq 1584, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; .param .b32 param4; st.param.b32 [param4+0], %r156; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1584 bra.uni $L__BB255_42; $L__BB255_249: { // callseq 1583, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; .param .b32 param4; st.param.b32 [param4+0], %r156; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1583 bra.uni $L__BB255_42; $L__BB255_253: mov.u32 %r211, 24; { // callseq 1589, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r211; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1589 bra.uni $L__BB255_42; $L__BB255_252: add.s32 %r210, %r7, 1; { // callseq 1588, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r210; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1588 bra.uni $L__BB255_42; $L__BB255_246: { // callseq 1579, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1579 bra.uni $L__BB255_42; $L__BB255_245: mov.u32 %r200, 36; { // callseq 1578, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r200; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1578 bra.uni $L__BB255_42; $L__BB255_258: mov.u32 %r217, 16; { // callseq 1596, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r217; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z13svm_node_attrILj56EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1596 bra.uni $L__BB255_42; $L__BB255_251: mov.u32 %r208, 27; { // callseq 1586, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r208; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1586 bra.uni $L__BB255_42; $L__BB255_265: { // callseq 1602, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1602 bra.uni $L__BB255_42; $L__BB255_207: { // callseq 1528, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1528 bra.uni $L__BB255_42; $L__BB255_227: mov.u32 %r176, 68; { // callseq 1548, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r176; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1548 bra.uni $L__BB255_42; $L__BB255_226: mov.u32 %r174, 69; add.s32 %r175, %r7, 1; { // callseq 1547, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r174; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r175; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1547 bra.uni $L__BB255_42; $L__BB255_264: st.local.u32 [%rd8+12], %r154; st.local.u32 [%rd8+16], %r155; st.local.u32 [%rd8+20], %r156; bra.uni $L__BB255_42; $L__BB255_234: mov.u32 %r189, 57; add.s32 %r190, %r7, 1; { // callseq 1559, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r189; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r190; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1559 bra.uni $L__BB255_42; $L__BB255_233: mov.u32 %r188, 58; { // callseq 1558, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r188; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1558 bra.uni $L__BB255_42; $L__BB255_214: add.s32 %r168, %r7, 1; { // callseq 1539, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r168; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1539 bra.uni $L__BB255_42; $L__BB255_213: { // callseq 1538, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1538 bra.uni $L__BB255_42; $L__BB255_242: mov.u32 %r195, 44; add.s32 %r196, %r7, 1; { // callseq 1572, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r195; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r196; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1572 bra.uni $L__BB255_42; $L__BB255_241: { // callseq 1571, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1571 bra.uni $L__BB255_42; $L__BB255_263: mov.u32 %r230, 8; { // callseq 1600, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r230; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1600 bra.uni $L__BB255_42; $L__BB255_229: { // callseq 1551, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1551 bra.uni $L__BB255_42; $L__BB255_228: add.s32 %r178, %r7, 1; { // callseq 1550, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r178; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1550 bra.uni $L__BB255_42; $L__BB255_236: add.s32 %r192, %r7, 1; { // callseq 1562, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r192; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1562 bra.uni $L__BB255_42; $L__BB255_235: ld.local.v4.f32 {%f808, %f809, %f810, %f811}, [%rd42]; { // callseq 1561, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f808; st.param.f32 [param0+4], %f809; st.param.f32 [param0+8], %f810; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; .param .b32 param4; st.param.b32 [param4+0], %r156; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1561 bra.uni $L__BB255_42; $L__BB255_217: setp.eq.s32 %p158, %r156, 255; @%p158 bra $L__BB255_42; add.s64 %rd13, %rd11, 4; add.s64 %rd14, %rd11, 8; setp.eq.s32 %p159, %r155, 0; mul.wide.u32 %rd154, %r156, 4; add.s64 %rd15, %rd9, %rd154; @%p159 bra $L__BB255_222; setp.eq.s32 %p160, %r155, 1; @%p160 bra $L__BB255_221; bra.uni $L__BB255_220; $L__BB255_221: ld.local.f32 %f806, [%rd13]; st.local.f32 [%rd15], %f806; bra.uni $L__BB255_42; $L__BB255_215: setp.eq.s32 %p157, %r156, 255; @%p157 bra $L__BB255_42; ld.local.f32 %f804, [%rd11]; add.s32 %r170, %r156, %r155; mul.wide.u32 %rd150, %r170, 4; add.s64 %rd151, %rd9, %rd150; st.local.f32 [%rd151], %f804; bra.uni $L__BB255_42; $L__BB255_239: { // callseq 1567, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1567 bra.uni $L__BB255_42; $L__BB255_238: { // callseq 1566, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1566 bra.uni $L__BB255_42; $L__BB255_260: { // callseq 1598, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1598 bra.uni $L__BB255_42; $L__BB255_225: mov.u32 %r172, 71; { // callseq 1545, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r172; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1545 bra.uni $L__BB255_42; $L__BB255_224: { // callseq 1544, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1544 bra.uni $L__BB255_42; $L__BB255_232: add.s32 %r186, %r7, 1; { // callseq 1556, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; .param .b32 param4; st.param.b32 [param4+0], %r186; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1556 bra.uni $L__BB255_42; $L__BB255_231: mov.u32 %r184, 61; add.s32 %r185, %r7, 1; { // callseq 1555, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r184; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r185; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1555 bra.uni $L__BB255_42; $L__BB255_212: mov.u32 %r166, 82; { // callseq 1536, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r166; st.param.b32 [param2+4], %r154; st.param.b32 [param2+8], %r155; st.param.b32 [param2+12], %r156; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1536 bra.uni $L__BB255_42; $L__BB255_211: { // callseq 1535, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1535 bra.uni $L__BB255_42; $L__BB255_206: mov.u32 %r159, 94; add.s32 %r160, %r7, 1; { // callseq 1527, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r159; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; .param .b32 param2; st.param.b32 [param2+0], %r160; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1527 bra.uni $L__BB255_42; $L__BB255_243: { // callseq 1574, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1574 bra.uni $L__BB255_42; $L__BB255_261: ld.local.f32 %f832, [%rd12]; setp.ltu.ftz.f32 %p161, %f832, 0f3F800000; @%p161 bra $L__BB255_42; add.s32 %r228, %r154, %r7; add.s32 %r254, %r228, 1; bra.uni $L__BB255_42; $L__BB255_230: mov.u32 %r181, 63; { // callseq 1553, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r181; st.param.b32 [param1+4], %r154; st.param.b32 [param1+8], %r155; st.param.b32 [param1+12], %r156; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1553 bra.uni $L__BB255_42; $L__BB255_237: add.s32 %r193, %r7, 1; { // callseq 1564, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r193; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r254, [retval0+0]; } // callseq 1564 bra.uni $L__BB255_42; $L__BB255_223: { // callseq 1542, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd124; .param .b32 param1; st.param.b32 [param1+0], %r154; .param .b32 param2; st.param.b32 [param2+0], %r155; .param .b32 param3; st.param.b32 [param3+0], %r156; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1542 bra.uni $L__BB255_42; $L__BB255_240: ld.local.u32 %r194, [%rd7+4]; ld.local.f32 %f815, [%rd7+40]; { // callseq 1569, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r194; .param .b32 param1; st.param.f32 [param1+0], %f815; .param .b64 param2; st.param.b64 [param2+0], %rd124; .param .b32 param3; st.param.b32 [param3+0], %r154; .param .b32 param4; st.param.b32 [param4+0], %r155; call.uni _Z19svm_node_light_pathILj56EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1569 bra.uni $L__BB255_42; $L__BB255_259: st.local.u32 [%rd12], %r154; bra.uni $L__BB255_42; $L__BB255_244: { // callseq 1576, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd41; .param .b64 param1; st.param.b64 [param1+0], %rd124; .param .b32 param2; st.param.b32 [param2+0], %r154; .param .b32 param3; st.param.b32 [param3+0], %r155; .param .b32 param4; st.param.b32 [param4+0], %r156; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1576 bra.uni $L__BB255_42; $L__BB255_208: mov.u32 %r162, 88; ld.local.f32 %f799, [%rd42+12]; ld.local.v2.f32 {%f800, %f801}, [%rd42+16]; { // callseq 1530, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f799; .param .b32 param1; st.param.f32 [param1+0], %f800; .param .b32 param2; st.param.f32 [param2+0], %f801; .param .b64 param3; st.param.b64 [param3+0], %rd124; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r162; st.param.b32 [param4+4], %r154; st.param.b32 [param4+8], %r155; st.param.b32 [param4+12], %r156; call.uni _Z14svm_node_bevelILj56EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1530 bra.uni $L__BB255_42; $L__BB255_222: ld.local.f32 %f807, [%rd11]; st.local.f32 [%rd15], %f807; bra.uni $L__BB255_42; $L__BB255_220: ld.local.f32 %f805, [%rd14]; st.local.f32 [%rd15], %f805; bra.uni $L__BB255_42; $L__BB255_266: ld.local.v4.f32 {%f837, %f838, %f839, %f840}, [%rd42]; sub.ftz.f32 %f216, %f837, %f795; sub.ftz.f32 %f217, %f838, %f796; sub.ftz.f32 %f218, %f839, %f797; ld.local.u8 %rs1, [%rd42+56]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p163, %rs2, 0; @%p163 bra $L__BB255_268; bra.uni $L__BB255_267; $L__BB255_268: ld.const.u64 %rd256, [kernel_params+2160]; cvta.to.global.u64 %rd257, %rd256; ld.local.u32 %r233, [%rd42+76]; mul.wide.s32 %rd258, %r233, 208; add.s64 %rd259, %rd257, %rd258; ld.global.v4.f32 {%f871, %f872, %f873, %f874}, [%rd259+48]; ld.global.v4.f32 {%f878, %f879, %f880, %f881}, [%rd259+64]; ld.global.v4.f32 {%f885, %f886, %f887, %f888}, [%rd259+80]; mul.ftz.f32 %f892, %f217, %f872; fma.rn.ftz.f32 %f893, %f216, %f871, %f892; fma.rn.ftz.f32 %f991, %f218, %f873, %f893; mul.ftz.f32 %f894, %f217, %f879; fma.rn.ftz.f32 %f895, %f216, %f878, %f894; fma.rn.ftz.f32 %f990, %f218, %f880, %f895; mul.ftz.f32 %f896, %f217, %f886; fma.rn.ftz.f32 %f897, %f216, %f885, %f896; fma.rn.ftz.f32 %f989, %f218, %f887, %f897; bra.uni $L__BB255_269; $L__BB255_267: ld.local.v4.f32 {%f844, %f845, %f846, %f847}, [%rd42+192]; mul.ftz.f32 %f851, %f217, %f845; fma.rn.ftz.f32 %f852, %f216, %f844, %f851; fma.rn.ftz.f32 %f991, %f218, %f846, %f852; ld.local.v4.f32 {%f853, %f854, %f855, %f856}, [%rd42+208]; mul.ftz.f32 %f860, %f217, %f854; fma.rn.ftz.f32 %f861, %f216, %f853, %f860; fma.rn.ftz.f32 %f990, %f218, %f855, %f861; ld.local.v4.f32 {%f862, %f863, %f864, %f865}, [%rd42+224]; mul.ftz.f32 %f869, %f217, %f863; fma.rn.ftz.f32 %f870, %f216, %f862, %f869; fma.rn.ftz.f32 %f989, %f218, %f864, %f870; $L__BB255_269: mov.b32 %r106, %f991; setp.nan.ftz.f32 %p164, %f991, %f991; mov.f32 %f992, %f994; @%p164 bra $L__BB255_272; or.b32 %r234, %r106, -2147483648; setp.ne.s32 %p165, %r234, -2147483648; add.ftz.f32 %f900, %f991, %f991; setp.eq.ftz.f32 %p166, %f900, %f991; and.pred %p167, %p166, %p165; mov.f32 %f992, %f994; @%p167 bra $L__BB255_272; shl.b32 %r235, %r106, 1; setp.lt.u32 %p168, %r235, -16777215; selp.f32 %f992, %f991, 0f00000000, %p168; $L__BB255_272: mov.b32 %r107, %f990; setp.nan.ftz.f32 %p169, %f990, %f990; mov.f32 %f993, %f994; @%p169 bra $L__BB255_275; or.b32 %r236, %r107, -2147483648; setp.ne.s32 %p170, %r236, -2147483648; add.ftz.f32 %f903, %f990, %f990; setp.eq.ftz.f32 %p171, %f903, %f990; and.pred %p172, %p171, %p170; mov.f32 %f993, %f994; @%p172 bra $L__BB255_275; shl.b32 %r237, %r107, 1; setp.lt.u32 %p173, %r237, -16777215; selp.f32 %f993, %f990, 0f00000000, %p173; $L__BB255_275: mov.b32 %r108, %f989; setp.nan.ftz.f32 %p174, %f989, %f989; @%p174 bra $L__BB255_278; or.b32 %r238, %r108, -2147483648; setp.ne.s32 %p175, %r238, -2147483648; add.ftz.f32 %f906, %f989, %f989; setp.eq.ftz.f32 %p176, %f906, %f989; and.pred %p177, %p176, %p175; @%p177 bra $L__BB255_278; shl.b32 %r239, %r108, 1; setp.lt.u32 %p178, %r239, -16777215; selp.f32 %f994, %f989, 0f00000000, %p178; $L__BB255_278: ld.param.u64 %rd264, [kernel_gpu_shader_eval_displace_param_1]; mov.u32 %r253, %tid.x; mov.u32 %r252, %ntid.x; mov.u32 %r251, %ctaid.x; ld.param.u32 %r250, [kernel_gpu_shader_eval_displace_param_2]; mad.lo.s32 %r249, %r251, %r252, %r253; add.s32 %r248, %r249, %r250; mul.lo.s32 %r245, %r248, 3; cvta.to.global.u64 %rd260, %rd264; mul.wide.s32 %rd261, %r245, 4; add.s64 %rd262, %rd260, %rd261; ld.global.f32 %f907, [%rd262]; add.ftz.f32 %f908, %f992, %f907; st.global.f32 [%rd262], %f908; ld.global.f32 %f909, [%rd262+4]; add.ftz.f32 %f910, %f993, %f909; st.global.f32 [%rd262+4], %f910; ld.global.f32 %f911, [%rd262+8]; add.ftz.f32 %f912, %f994, %f911; st.global.f32 [%rd262+8], %f912; $L__BB255_279: ret; } // .globl kernel_gpu_shader_eval_background .visible .entry kernel_gpu_shader_eval_background( .param .u64 kernel_gpu_shader_eval_background_param_0, .param .u64 kernel_gpu_shader_eval_background_param_1, .param .u32 kernel_gpu_shader_eval_background_param_2, .param .u32 kernel_gpu_shader_eval_background_param_3 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot256[6448]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<142>; .reg .b16 %rs<3>; .reg .f32 %f<80>; .reg .b32 %r<137>; .reg .b64 %rd<141>; mov.u64 %SPL, __local_depot256; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd8, [kernel_gpu_shader_eval_background_param_0]; ld.param.u64 %rd9, [kernel_gpu_shader_eval_background_param_1]; ld.param.u32 %r38, [kernel_gpu_shader_eval_background_param_2]; ld.param.u32 %r39, [kernel_gpu_shader_eval_background_param_3]; mov.u32 %r40, %ntid.x; mov.u32 %r41, %ctaid.x; mov.u32 %r42, %tid.x; mad.lo.s32 %r1, %r41, %r40, %r42; setp.ge.s32 %p1, %r1, %r39; @%p1 bra $L__BB256_219; add.s32 %r43, %r1, %r38; cvta.to.global.u64 %rd10, %rd8; mul.wide.s32 %rd11, %r43, 16; add.s64 %rd12, %rd10, %rd11; ld.global.f32 %f16, [%rd12+8]; fma.rn.ftz.f32 %f17, %f16, 0fC0C90FDB, 0f40490FDB; mov.f32 %f18, 0f40490FDB; ld.global.f32 %f19, [%rd12+12]; mul.ftz.f32 %f20, %f19, 0f40490FDB; sub.ftz.f32 %f21, %f18, %f20; sin.approx.ftz.f32 %f22, %f21; cos.approx.ftz.f32 %f23, %f17; sin.approx.ftz.f32 %f24, %f17; mov.f32 %f79, 0f00000000; add.u64 %rd13, %SP, 1024; add.u64 %rd14, %SPL, 1024; st.local.v2.f32 [%rd14+240], {%f79, %f79}; mov.u32 %r44, 0; st.local.u32 [%rd14+248], %r44; mul.ftz.f32 %f26, %f22, %f24; mul.ftz.f32 %f27, %f22, %f23; cos.approx.ftz.f32 %f28, %f21; neg.ftz.f32 %f29, %f27; st.local.v4.f32 [%rd14], {%f27, %f26, %f28, %f29}; neg.ftz.f32 %f30, %f26; neg.ftz.f32 %f31, %f28; st.local.v4.f32 [%rd14+16], {%f30, %f31, %f29, %f30}; st.local.v4.f32 [%rd14+32], {%f31, %f29, %f30, %f31}; ld.const.u64 %rd15, [kernel_params+2408]; cvta.to.global.u64 %rd16, %rd15; ld.const.u32 %r45, [kernel_params+1024]; and.b32 %r136, %r45, 4194303; shl.b32 %r46, %r45, 5; cvt.u64.u32 %rd17, %r46; and.b64 %rd18, %rd17, 134217696; add.s64 %rd19, %rd16, %rd18; ld.global.u32 %r47, [%rd19+16]; mov.u64 %rd20, -4294967296; st.local.u64 [%rd14+72], %rd20; mov.u32 %r48, -1; st.local.v4.u32 [%rd14+48], {%r45, %r47, %r44, %r48}; st.local.u32 [%rd14+68], %r44; st.local.v2.f32 [%rd14+128], {%f79, %f79}; st.local.u32 [%rd14+136], %r44; mov.u32 %r49, 1056964608; st.local.v4.u32 [%rd14+80], {%r48, %r49, %r44, %r44}; st.local.v4.f32 [%rd14+96], {%f79, %f79, %f79, %f79}; st.local.v4.f32 [%rd14+112], {%f79, %f79, %f79, %f79}; st.local.u32 [%rd14+260], %r44; st.local.u32 [%rd14+264], %r44; ld.const.u64 %rd21, [kernel_params+2400]; cvta.to.global.u64 %rd1, %rd21; bra.uni $L__BB256_2; $L__BB256_154: { // callseq 1603, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1603 $L__BB256_2: mov.u32 %r3, %r136; mul.wide.s32 %rd22, %r3, 16; add.s64 %rd2, %rd1, %rd22; ld.global.v4.u32 {%r51, %r52, %r53, %r54}, [%rd2]; add.s32 %r136, %r3, 1; add.u64 %rd23, %SP, 0; add.u64 %rd24, %SPL, 0; mul.wide.u32 %rd25, %r52, 4; add.s64 %rd3, %rd24, %rd25; mul.wide.u32 %rd26, %r53, 4; add.s64 %rd4, %rd24, %rd26; setp.gt.s32 %p2, %r51, 36; @%p2 bra $L__BB256_51; setp.gt.s32 %p84, %r51, 25; @%p84 bra $L__BB256_47; bra.uni $L__BB256_4; $L__BB256_47: add.s32 %r56, %r51, -27; setp.lt.u32 %p85, %r56, 9; @%p85 bra $L__BB256_2; setp.eq.s32 %p86, %r51, 26; @%p86 bra $L__BB256_193; setp.eq.s32 %p87, %r51, 36; @%p87 bra $L__BB256_50; bra.uni $L__BB256_207; $L__BB256_50: mov.u32 %r79, 36; { // callseq 1656, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r79; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1656 bra.uni $L__BB256_2; $L__BB256_51: setp.gt.s32 %p3, %r51, 68; @%p3 bra $L__BB256_106; bra.uni $L__BB256_52; $L__BB256_106: setp.gt.s32 %p4, %r51, 88; @%p4 bra $L__BB256_142; bra.uni $L__BB256_107; $L__BB256_142: setp.gt.s32 %p5, %r51, 93; @%p5 bra $L__BB256_147; bra.uni $L__BB256_143; $L__BB256_147: setp.gt.s32 %p6, %r51, 95; @%p6 bra $L__BB256_151; bra.uni $L__BB256_148; $L__BB256_151: setp.eq.s32 %p7, %r51, 98; @%p7 bra $L__BB256_206; setp.eq.s32 %p8, %r51, 97; @%p8 bra $L__BB256_156; setp.ne.s32 %p9, %r51, 96; @%p9 bra $L__BB256_207; bra.uni $L__BB256_154; $L__BB256_52: setp.gt.s32 %p42, %r51, 48; @%p42 bra $L__BB256_71; bra.uni $L__BB256_53; $L__BB256_71: setp.gt.s32 %p43, %r51, 58; @%p43 bra $L__BB256_89; setp.gt.s32 %p57, %r51, 53; @%p57 bra $L__BB256_81; setp.gt.s32 %p64, %r51, 50; @%p64 bra $L__BB256_77; bra.uni $L__BB256_74; $L__BB256_77: setp.eq.s32 %p65, %r51, 51; @%p65 bra $L__BB256_185; setp.eq.s32 %p66, %r51, 52; @%p66 bra $L__BB256_184; setp.eq.s32 %p67, %r51, 53; @%p67 bra $L__BB256_80; bra.uni $L__BB256_207; $L__BB256_80: { // callseq 1641, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1641 bra.uni $L__BB256_2; $L__BB256_4: setp.gt.s32 %p88, %r51, 17; @%p88 bra $L__BB256_38; bra.uni $L__BB256_5; $L__BB256_38: setp.lt.s32 %p89, %r51, 21; @%p89 bra $L__BB256_2; setp.gt.s32 %p90, %r51, 22; @%p90 bra $L__BB256_43; bra.uni $L__BB256_40; $L__BB256_43: setp.eq.s32 %p91, %r51, 23; @%p91 bra $L__BB256_195; setp.eq.s32 %p92, %r51, 24; @%p92 bra $L__BB256_194; setp.eq.s32 %p93, %r51, 25; @%p93 bra $L__BB256_46; bra.uni $L__BB256_207; $L__BB256_46: { // callseq 1658, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r131, %r3, 1; st.param.b32 [param4+0], %r131; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1658 bra.uni $L__BB256_2; $L__BB256_107: setp.gt.s32 %p15, %r51, 78; @%p15 bra $L__BB256_125; setp.gt.s32 %p29, %r51, 73; @%p29 bra $L__BB256_117; setp.gt.s32 %p36, %r51, 70; @%p36 bra $L__BB256_113; bra.uni $L__BB256_110; $L__BB256_113: setp.eq.s32 %p37, %r51, 71; @%p37 bra $L__BB256_173; setp.eq.s32 %p38, %r51, 72; @%p38 bra $L__BB256_172; setp.eq.s32 %p39, %r51, 73; @%p39 bra $L__BB256_116; bra.uni $L__BB256_207; $L__BB256_116: { // callseq 1621, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r118, %r3, 1; st.param.b32 [param4+0], %r118; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1621 bra.uni $L__BB256_2; $L__BB256_53: setp.gt.s32 %p70, %r51, 41; @%p70 bra $L__BB256_59; setp.gt.s32 %p80, %r51, 39; @%p80 bra $L__BB256_2; setp.eq.s32 %p81, %r51, 37; @%p81 bra $L__BB256_192; setp.eq.s32 %p82, %r51, 38; @%p82 bra $L__BB256_191; setp.eq.s32 %p83, %r51, 39; @%p83 bra $L__BB256_58; bra.uni $L__BB256_207; $L__BB256_58: mov.u32 %r77, 39; { // callseq 1653, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r77, %r52, %r53, %r54}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1653 bra.uni $L__BB256_2; $L__BB256_5: setp.gt.s32 %p96, %r51, 8; @%p96 bra $L__BB256_21; bra.uni $L__BB256_6; $L__BB256_21: setp.gt.s32 %p97, %r51, 12; @%p97 bra $L__BB256_30; bra.uni $L__BB256_22; $L__BB256_30: setp.gt.s32 %p98, %r51, 14; @%p98 bra $L__BB256_34; bra.uni $L__BB256_31; $L__BB256_34: setp.eq.s32 %p99, %r51, 15; @%p99 bra $L__BB256_198; setp.eq.s32 %p100, %r51, 16; @%p100 bra $L__BB256_197; setp.eq.s32 %p101, %r51, 17; @%p101 bra $L__BB256_37; bra.uni $L__BB256_207; $L__BB256_37: { // callseq 1663, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; .param .b32 param4; st.param.b32 [param4+0], %r54; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1663 bra.uni $L__BB256_2; $L__BB256_89: setp.gt.s32 %p44, %r51, 63; @%p44 bra $L__BB256_98; setp.gt.s32 %p51, %r51, 60; @%p51 bra $L__BB256_94; bra.uni $L__BB256_91; $L__BB256_94: setp.eq.s32 %p52, %r51, 61; @%p52 bra $L__BB256_179; setp.eq.s32 %p53, %r51, 62; @%p53 bra $L__BB256_178; setp.eq.s32 %p54, %r51, 63; @%p54 bra $L__BB256_97; bra.uni $L__BB256_207; $L__BB256_97: mov.u32 %r69, 63; { // callseq 1631, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r69; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1631 bra.uni $L__BB256_2; $L__BB256_125: setp.gt.s32 %p16, %r51, 83; @%p16 bra $L__BB256_134; setp.gt.s32 %p23, %r51, 80; @%p23 bra $L__BB256_130; bra.uni $L__BB256_127; $L__BB256_130: setp.eq.s32 %p24, %r51, 81; @%p24 bra $L__BB256_161; setp.eq.s32 %p25, %r51, 82; @%p25 bra $L__BB256_160; setp.eq.s32 %p26, %r51, 83; @%p26 bra $L__BB256_133; bra.uni $L__BB256_207; $L__BB256_133: { // callseq 1613, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1613 bra.uni $L__BB256_2; $L__BB256_6: setp.gt.s32 %p109, %r51, 4; @%p109 bra $L__BB256_14; setp.gt.s32 %p115, %r51, 2; @%p115 bra $L__BB256_11; setp.eq.s32 %p118, %r51, 1; mov.u32 %r136, %r52; @%p118 bra $L__BB256_2; setp.eq.s32 %p119, %r51, 2; @%p119 bra $L__BB256_10; bra.uni $L__BB256_207; $L__BB256_10: mov.u32 %r99, 2; { // callseq 1672, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[16]; st.param.b32 [param0+0], %r99; st.param.b32 [param0+4], %r52; st.param.b32 [param0+8], %r53; st.param.b32 [param0+12], %r54; .param .b32 param1; add.s32 %r135, %r3, 1; st.param.b32 [param1+0], %r135; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj34EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1672 bra.uni $L__BB256_2; $L__BB256_59: setp.gt.s32 %p71, %r51, 44; @%p71 bra $L__BB256_64; bra.uni $L__BB256_60; $L__BB256_64: setp.gt.s32 %p72, %r51, 46; @%p72 bra $L__BB256_68; setp.eq.s32 %p75, %r51, 45; @%p75 bra $L__BB256_188; setp.eq.s32 %p76, %r51, 46; @%p76 bra $L__BB256_67; bra.uni $L__BB256_207; $L__BB256_67: { // callseq 1648, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1648 bra.uni $L__BB256_2; $L__BB256_143: add.s32 %r55, %r51, -92; setp.lt.u32 %p12, %r55, 2; @%p12 bra $L__BB256_2; setp.eq.s32 %p13, %r51, 90; @%p13 bra $L__BB256_2; setp.eq.s32 %p14, %r51, 89; @%p14 bra $L__BB256_146; bra.uni $L__BB256_207; $L__BB256_146: mov.u32 %r58, 89; { // callseq 1607, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r58; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z11svm_node_aoILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1607 bra.uni $L__BB256_2; $L__BB256_148: setp.eq.s32 %p10, %r51, 94; @%p10 bra $L__BB256_155; setp.eq.s32 %p11, %r51, 95; @%p11 bra $L__BB256_150; bra.uni $L__BB256_207; $L__BB256_150: { // callseq 1604, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1604 bra.uni $L__BB256_2; $L__BB256_81: setp.gt.s32 %p58, %r51, 55; @%p58 bra $L__BB256_85; bra.uni $L__BB256_82; $L__BB256_85: setp.eq.s32 %p59, %r51, 56; @%p59 bra $L__BB256_182; setp.eq.s32 %p60, %r51, 57; @%p60 bra $L__BB256_181; setp.eq.s32 %p61, %r51, 58; @%p61 bra $L__BB256_88; bra.uni $L__BB256_207; $L__BB256_88: mov.u32 %r72, 58; { // callseq 1636, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r72; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1636 bra.uni $L__BB256_2; $L__BB256_117: setp.gt.s32 %p30, %r51, 75; @%p30 bra $L__BB256_121; bra.uni $L__BB256_118; $L__BB256_121: setp.eq.s32 %p31, %r51, 76; @%p31 bra $L__BB256_165; setp.eq.s32 %p32, %r51, 77; @%p32 bra $L__BB256_163; setp.eq.s32 %p33, %r51, 78; @%p33 bra $L__BB256_124; bra.uni $L__BB256_207; $L__BB256_124: { // callseq 1618, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r117, %r3, 1; st.param.b32 [param4+0], %r117; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1618 bra.uni $L__BB256_2; $L__BB256_98: setp.gt.s32 %p45, %r51, 65; @%p45 bra $L__BB256_102; bra.uni $L__BB256_99; $L__BB256_102: setp.eq.s32 %p46, %r51, 66; @%p46 bra $L__BB256_176; setp.eq.s32 %p47, %r51, 67; @%p47 bra $L__BB256_175; setp.eq.s32 %p48, %r51, 68; @%p48 bra $L__BB256_105; bra.uni $L__BB256_207; $L__BB256_105: mov.u32 %r66, 68; { // callseq 1626, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r66; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1626 bra.uni $L__BB256_2; $L__BB256_134: setp.gt.s32 %p17, %r51, 85; @%p17 bra $L__BB256_138; bra.uni $L__BB256_135; $L__BB256_138: setp.eq.s32 %p18, %r51, 86; @%p18 bra $L__BB256_158; setp.eq.s32 %p19, %r51, 87; @%p19 bra $L__BB256_157; setp.eq.s32 %p20, %r51, 88; @%p20 bra $L__BB256_141; bra.uni $L__BB256_207; $L__BB256_141: mov.u32 %r59, 88; ld.local.f32 %f32, [%rd14+12]; ld.local.v2.f32 {%f33, %f34}, [%rd14+16]; { // callseq 1608, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f32; .param .b32 param1; st.param.f32 [param1+0], %f33; .param .b32 param2; st.param.f32 [param2+0], %f34; .param .b64 param3; st.param.b64 [param3+0], %rd23; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r59; st.param.b32 [param4+4], %r52; st.param.b32 [param4+8], %r53; st.param.b32 [param4+12], %r54; call.uni _Z14svm_node_bevelILj34EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1608 bra.uni $L__BB256_2; $L__BB256_22: setp.gt.s32 %p104, %r51, 10; @%p104 bra $L__BB256_27; setp.eq.s32 %p107, %r51, 9; @%p107 bra $L__BB256_201; setp.eq.s32 %p108, %r51, 10; @%p108 bra $L__BB256_25; bra.uni $L__BB256_207; $L__BB256_25: ld.local.f32 %f49, [%rd4]; setp.ltu.ftz.f32 %p124, %f49, 0f3F800000; @%p124 bra $L__BB256_2; add.s32 %r136, %r52, %r136; bra.uni $L__BB256_2; $L__BB256_14: setp.gt.s32 %p110, %r51, 6; @%p110 bra $L__BB256_18; setp.eq.s32 %p113, %r51, 5; @%p113 bra $L__BB256_204; setp.eq.s32 %p114, %r51, 6; @%p114 bra $L__BB256_17; bra.uni $L__BB256_207; $L__BB256_17: ld.local.f32 %f51, [%rd3]; ld.local.f32 %f52, [%rd3+4]; ld.local.f32 %f53, [%rd3+8]; st.local.f32 [%rd14+268], %f51; st.local.f32 [%rd14+272], %f52; st.local.f32 [%rd14+276], %f53; bra.uni $L__BB256_2; $L__BB256_60: setp.eq.s32 %p77, %r51, 42; @%p77 bra $L__BB256_190; setp.eq.s32 %p78, %r51, 43; @%p78 bra $L__BB256_189; setp.eq.s32 %p79, %r51, 44; @%p79 bra $L__BB256_63; bra.uni $L__BB256_207; $L__BB256_63: mov.u32 %r76, 44; { // callseq 1650, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r76; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r129, %r3, 1; st.param.b32 [param2+0], %r129; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1650 bra.uni $L__BB256_2; $L__BB256_206: { // callseq 1673, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1673 bra.uni $L__BB256_2; $L__BB256_156: { // callseq 1606, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1606 bra.uni $L__BB256_2; $L__BB256_155: mov.u32 %r57, 94; { // callseq 1605, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r57; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r112, %r3, 1; st.param.b32 [param2+0], %r112; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1605 bra.uni $L__BB256_2; $L__BB256_40: setp.eq.s32 %p94, %r51, 21; @%p94 bra $L__BB256_196; setp.eq.s32 %p95, %r51, 22; @%p95 bra $L__BB256_42; bra.uni $L__BB256_207; $L__BB256_42: mov.u32 %r83, 22; { // callseq 1661, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r83; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r133, %r3, 1; st.param.b32 [param2+0], %r133; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj34EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1661 bra.uni $L__BB256_2; $L__BB256_31: setp.eq.s32 %p102, %r51, 13; @%p102 bra $L__BB256_199; setp.eq.s32 %p103, %r51, 14; @%p103 bra $L__BB256_33; bra.uni $L__BB256_207; $L__BB256_33: st.local.u32 [%rd4], %r52; bra.uni $L__BB256_2; $L__BB256_11: setp.eq.s32 %p116, %r51, 3; @%p116 bra $L__BB256_205; setp.eq.s32 %p117, %r51, 4; @%p117 bra $L__BB256_13; bra.uni $L__BB256_207; $L__BB256_13: mov.u32 %r97, 4; { // callseq 1670, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r97, %r52, %r53, %r54}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1670 bra.uni $L__BB256_2; $L__BB256_27: setp.eq.s32 %p105, %r51, 11; @%p105 bra $L__BB256_200; setp.eq.s32 %p106, %r51, 12; @%p106 bra $L__BB256_29; bra.uni $L__BB256_207; $L__BB256_29: { // callseq 1666, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1666 bra.uni $L__BB256_2; $L__BB256_18: setp.eq.s32 %p111, %r51, 7; @%p111 bra $L__BB256_203; setp.eq.s32 %p112, %r51, 8; @%p112 bra $L__BB256_20; bra.uni $L__BB256_207; $L__BB256_20: mov.u32 %r95, 8; { // callseq 1668, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r95; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1668 bra.uni $L__BB256_2; $L__BB256_74: setp.eq.s32 %p68, %r51, 49; @%p68 bra $L__BB256_186; setp.eq.s32 %p69, %r51, 50; @%p69 bra $L__BB256_76; bra.uni $L__BB256_207; $L__BB256_76: { // callseq 1644, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1644 bra.uni $L__BB256_2; $L__BB256_110: setp.eq.s32 %p40, %r51, 69; @%p40 bra $L__BB256_174; setp.eq.s32 %p41, %r51, 70; @%p41 bra $L__BB256_112; bra.uni $L__BB256_207; $L__BB256_112: mov.u32 %r64, 70; { // callseq 1624, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r64; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1624 bra.uni $L__BB256_2; $L__BB256_91: setp.eq.s32 %p55, %r51, 59; @%p55 bra $L__BB256_180; setp.eq.s32 %p56, %r51, 60; @%p56 bra $L__BB256_93; bra.uni $L__BB256_207; $L__BB256_93: { // callseq 1634, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r124, %r3, 1; st.param.b32 [param4+0], %r124; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1634 bra.uni $L__BB256_2; $L__BB256_127: setp.eq.s32 %p27, %r51, 79; @%p27 bra $L__BB256_162; setp.eq.s32 %p28, %r51, 80; @%p28 bra $L__BB256_129; bra.uni $L__BB256_207; $L__BB256_129: { // callseq 1616, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1616 bra.uni $L__BB256_2; $L__BB256_68: setp.eq.s32 %p73, %r51, 47; @%p73 bra $L__BB256_187; setp.eq.s32 %p74, %r51, 48; @%p74 bra $L__BB256_70; bra.uni $L__BB256_207; $L__BB256_70: { // callseq 1646, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1646 bra.uni $L__BB256_2; $L__BB256_82: setp.eq.s32 %p62, %r51, 54; @%p62 bra $L__BB256_183; setp.eq.s32 %p63, %r51, 55; @%p63 bra $L__BB256_84; bra.uni $L__BB256_207; $L__BB256_84: ld.local.v4.f32 {%f41, %f42, %f43, %f44}, [%rd14]; { // callseq 1639, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f41; st.param.f32 [param0+4], %f42; st.param.f32 [param0+8], %f43; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; .param .b32 param4; st.param.b32 [param4+0], %r54; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1639 bra.uni $L__BB256_2; $L__BB256_118: setp.eq.s32 %p34, %r51, 74; @%p34 bra $L__BB256_171; setp.eq.s32 %p35, %r51, 75; @%p35 bra $L__BB256_120; bra.uni $L__BB256_207; $L__BB256_120: { // callseq 1619, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1619 bra.uni $L__BB256_2; $L__BB256_99: setp.eq.s32 %p49, %r51, 64; @%p49 bra $L__BB256_177; setp.eq.s32 %p50, %r51, 65; @%p50 bra $L__BB256_101; bra.uni $L__BB256_207; $L__BB256_101: { // callseq 1629, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1629 bra.uni $L__BB256_2; $L__BB256_135: setp.eq.s32 %p21, %r51, 84; @%p21 bra $L__BB256_159; setp.eq.s32 %p22, %r51, 85; @%p22 bra $L__BB256_137; bra.uni $L__BB256_207; $L__BB256_137: { // callseq 1611, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r115, %r3, 1; st.param.b32 [param4+0], %r115; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1611 bra.uni $L__BB256_2; $L__BB256_195: mov.u32 %r82, 23; { // callseq 1660, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r82; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r132, %r3, 1; st.param.b32 [param2+0], %r132; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1660 bra.uni $L__BB256_2; $L__BB256_194: mov.u32 %r81, 24; { // callseq 1659, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r81; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1659 bra.uni $L__BB256_2; $L__BB256_198: ld.global.v4.u32 {%r86, %r87, %r88, %r89}, [%rd2+16]; add.s32 %r136, %r3, 2; st.local.u32 [%rd3], %r87; st.local.u32 [%rd3+4], %r88; st.local.u32 [%rd3+8], %r89; bra.uni $L__BB256_2; $L__BB256_197: mov.u32 %r85, 16; { // callseq 1664, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r85; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z13svm_node_attrILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1664 bra.uni $L__BB256_2; $L__BB256_201: ld.local.f32 %f50, [%rd4]; setp.gtu.ftz.f32 %p125, %f50, 0f00000000; @%p125 bra $L__BB256_2; add.s32 %r136, %r52, %r136; bra.uni $L__BB256_2; $L__BB256_204: st.local.u32 [%rd14+268], %r52; st.local.u32 [%rd14+272], %r53; st.local.u32 [%rd14+276], %r54; bra.uni $L__BB256_2; $L__BB256_196: mov.u32 %r84, 21; { // callseq 1662, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r84; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z21svm_node_displacementILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1662 bra.uni $L__BB256_2; $L__BB256_199: mov.u32 %r93, 13; mov.u32 %r94, 524288; { // callseq 1665, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b32 param1; st.param.b32 [param1+0], %r94; .param .b64 param2; st.param.b64 [param2+0], %rd23; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r93; st.param.b32 [param3+4], %r52; st.param.b32 [param3+8], %r53; st.param.b32 [param3+12], %r54; .param .b32 param4; add.s32 %r134, %r3, 1; st.param.b32 [param4+0], %r134; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1665 bra.uni $L__BB256_2; $L__BB256_205: mov.u32 %r98, 3; { // callseq 1671, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r98, %r52, %r53, %r54}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1671 bra.uni $L__BB256_2; $L__BB256_200: { // callseq 1667, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1667 bra.uni $L__BB256_2; $L__BB256_203: mov.u32 %r96, 7; { // callseq 1669, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r96; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1669 bra.uni $L__BB256_2; $L__BB256_185: { // callseq 1643, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1643 bra.uni $L__BB256_2; $L__BB256_184: { // callseq 1642, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; add.s32 %r128, %r3, 1; st.param.b32 [param3+0], %r128; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1642 bra.uni $L__BB256_2; $L__BB256_173: mov.u32 %r63, 71; { // callseq 1623, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r63; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1623 bra.uni $L__BB256_2; $L__BB256_172: { // callseq 1622, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1622 bra.uni $L__BB256_2; $L__BB256_192: mov.u32 %r78, 37; { // callseq 1655, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r78, %r52, %r53, %r54}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1655 bra.uni $L__BB256_2; $L__BB256_191: { // callseq 1654, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; .param .b32 param4; st.param.b32 [param4+0], %r54; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1654 bra.uni $L__BB256_2; $L__BB256_179: mov.u32 %r71, 61; { // callseq 1633, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r71; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r123, %r3, 1; st.param.b32 [param2+0], %r123; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1633 bra.uni $L__BB256_2; $L__BB256_178: mov.u32 %r70, 62; { // callseq 1632, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r70; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r122, %r3, 1; st.param.b32 [param2+0], %r122; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1632 bra.uni $L__BB256_2; $L__BB256_161: mov.u32 %r61, 81; { // callseq 1615, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r61; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1615 bra.uni $L__BB256_2; $L__BB256_160: mov.u32 %r60, 82; { // callseq 1614, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r60; st.param.b32 [param2+4], %r52; st.param.b32 [param2+8], %r53; st.param.b32 [param2+12], %r54; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1614 bra.uni $L__BB256_2; $L__BB256_188: { // callseq 1649, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1649 bra.uni $L__BB256_2; $L__BB256_182: mov.u32 %r74, 56; { // callseq 1638, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r74; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1638 bra.uni $L__BB256_2; $L__BB256_181: mov.u32 %r73, 57; { // callseq 1637, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r73; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r126, %r3, 1; st.param.b32 [param2+0], %r126; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1637 bra.uni $L__BB256_2; $L__BB256_165: setp.eq.s32 %p121, %r54, 255; @%p121 bra $L__BB256_2; add.s64 %rd5, %rd3, 4; add.s64 %rd6, %rd3, 8; setp.eq.s32 %p122, %r53, 0; mul.wide.u32 %rd53, %r54, 4; add.s64 %rd7, %rd24, %rd53; @%p122 bra $L__BB256_170; setp.eq.s32 %p123, %r53, 1; @%p123 bra $L__BB256_169; bra.uni $L__BB256_168; $L__BB256_169: ld.local.f32 %f39, [%rd5]; st.local.f32 [%rd7], %f39; bra.uni $L__BB256_2; $L__BB256_163: setp.eq.s32 %p120, %r54, 255; @%p120 bra $L__BB256_2; ld.local.f32 %f37, [%rd3]; add.s32 %r62, %r54, %r53; mul.wide.u32 %rd49, %r62, 4; add.s64 %rd50, %rd24, %rd49; st.local.f32 [%rd50], %f37; bra.uni $L__BB256_2; $L__BB256_176: { // callseq 1628, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r120, %r3, 1; st.param.b32 [param4+0], %r120; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1628 bra.uni $L__BB256_2; $L__BB256_175: mov.u32 %r67, 67; { // callseq 1627, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r67, %r52, %r53, %r54}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1627 bra.uni $L__BB256_2; $L__BB256_158: { // callseq 1610, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r114, %r3, 1; st.param.b32 [param4+0], %r114; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1610 bra.uni $L__BB256_2; $L__BB256_157: { // callseq 1609, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r113, %r3, 1; st.param.b32 [param4+0], %r113; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1609 bra.uni $L__BB256_2; $L__BB256_190: { // callseq 1652, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1652 bra.uni $L__BB256_2; $L__BB256_189: { // callseq 1651, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r130, %r3, 1; st.param.b32 [param4+0], %r130; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1651 bra.uni $L__BB256_2; $L__BB256_186: { // callseq 1645, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd13; .param .b64 param1; st.param.b64 [param1+0], %rd23; .param .b32 param2; st.param.b32 [param2+0], %r52; .param .b32 param3; st.param.b32 [param3+0], %r53; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1645 bra.uni $L__BB256_2; $L__BB256_174: mov.u32 %r65, 69; { // callseq 1625, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r65; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r119, %r3, 1; st.param.b32 [param2+0], %r119; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1625 bra.uni $L__BB256_2; $L__BB256_180: { // callseq 1635, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r125, %r3, 1; st.param.b32 [param4+0], %r125; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj34EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1635 bra.uni $L__BB256_2; $L__BB256_162: { // callseq 1617, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; .param .b32 param4; add.s32 %r116, %r3, 1; st.param.b32 [param4+0], %r116; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1617 bra.uni $L__BB256_2; $L__BB256_187: ld.local.u32 %r75, [%rd14+52]; ld.local.f32 %f48, [%rd14+88]; { // callseq 1647, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r75; .param .b32 param1; st.param.f32 [param1+0], %f48; .param .b64 param2; st.param.b64 [param2+0], %rd23; .param .b32 param3; st.param.b32 [param3+0], %r52; .param .b32 param4; st.param.b32 [param4+0], %r53; call.uni _Z19svm_node_light_pathILj34EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1647 bra.uni $L__BB256_2; $L__BB256_183: { // callseq 1640, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; add.s32 %r127, %r3, 1; st.param.b32 [param3+0], %r127; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1640 bra.uni $L__BB256_2; $L__BB256_171: { // callseq 1620, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; .param .b32 param3; st.param.b32 [param3+0], %r54; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1620 bra.uni $L__BB256_2; $L__BB256_177: mov.u32 %r68, 64; { // callseq 1630, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r68; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; .param .b32 param2; add.s32 %r121, %r3, 1; st.param.b32 [param2+0], %r121; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r136, [retval0+0]; } // callseq 1630 bra.uni $L__BB256_2; $L__BB256_159: { // callseq 1612, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .b32 param1; st.param.b32 [param1+0], %r52; .param .b32 param2; st.param.b32 [param2+0], %r53; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1612 bra.uni $L__BB256_2; $L__BB256_193: mov.u32 %r80, 26; { // callseq 1657, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd23; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r80; st.param.b32 [param1+4], %r52; st.param.b32 [param1+8], %r53; st.param.b32 [param1+12], %r54; call.uni _Z17svm_node_set_bumpILj34EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1657 bra.uni $L__BB256_2; $L__BB256_170: ld.local.f32 %f40, [%rd3]; st.local.f32 [%rd7], %f40; bra.uni $L__BB256_2; $L__BB256_168: ld.local.f32 %f38, [%rd6]; st.local.f32 [%rd7], %f38; bra.uni $L__BB256_2; $L__BB256_207: ld.local.u8 %rs1, [%rd14+52]; and.b16 %rs2, %rs1, 2; setp.eq.s16 %p126, %rs2, 0; mov.f32 %f74, %f79; mov.f32 %f75, %f79; mov.f32 %f76, %f79; @%p126 bra $L__BB256_209; ld.local.v2.f32 {%f74, %f75}, [%rd14+280]; ld.local.f32 %f76, [%rd14+288]; $L__BB256_209: mov.b32 %r35, %f74; setp.nan.ftz.f32 %p127, %f74, %f74; mov.f32 %f77, %f79; @%p127 bra $L__BB256_212; or.b32 %r100, %r35, -2147483648; setp.ne.s32 %p128, %r100, -2147483648; add.ftz.f32 %f61, %f74, %f74; setp.eq.ftz.f32 %p129, %f61, %f74; and.pred %p130, %p129, %p128; mov.f32 %f77, %f79; @%p130 bra $L__BB256_212; shl.b32 %r101, %r35, 1; setp.lt.u32 %p131, %r101, -16777215; selp.f32 %f77, %f74, 0f00000000, %p131; $L__BB256_212: mov.b32 %r36, %f75; setp.nan.ftz.f32 %p132, %f75, %f75; mov.f32 %f78, %f79; @%p132 bra $L__BB256_215; or.b32 %r102, %r36, -2147483648; setp.ne.s32 %p133, %r102, -2147483648; add.ftz.f32 %f64, %f75, %f75; setp.eq.ftz.f32 %p134, %f64, %f75; and.pred %p135, %p134, %p133; mov.f32 %f78, %f79; @%p135 bra $L__BB256_215; shl.b32 %r103, %r36, 1; setp.lt.u32 %p136, %r103, -16777215; selp.f32 %f78, %f75, 0f00000000, %p136; $L__BB256_215: mov.b32 %r37, %f76; setp.nan.ftz.f32 %p137, %f76, %f76; @%p137 bra $L__BB256_218; or.b32 %r104, %r37, -2147483648; setp.ne.s32 %p138, %r104, -2147483648; add.ftz.f32 %f67, %f76, %f76; setp.eq.ftz.f32 %p139, %f67, %f76; and.pred %p140, %p139, %p138; @%p140 bra $L__BB256_218; shl.b32 %r105, %r37, 1; setp.lt.u32 %p141, %r105, -16777215; selp.f32 %f79, %f76, 0f00000000, %p141; $L__BB256_218: mul.lo.s32 %r111, %r43, 3; cvta.to.global.u64 %rd138, %rd9; mul.wide.s32 %rd139, %r111, 4; add.s64 %rd140, %rd138, %rd139; ld.global.f32 %f68, [%rd140]; add.ftz.f32 %f69, %f77, %f68; st.global.f32 [%rd140], %f69; ld.global.f32 %f70, [%rd140+4]; add.ftz.f32 %f71, %f78, %f70; st.global.f32 [%rd140+4], %f71; ld.global.f32 %f72, [%rd140+8]; add.ftz.f32 %f73, %f79, %f72; st.global.f32 [%rd140+8], %f73; $L__BB256_219: ret; } // .globl kernel_gpu_shader_eval_curve_shadow_transparency .visible .entry kernel_gpu_shader_eval_curve_shadow_transparency( .param .u64 kernel_gpu_shader_eval_curve_shadow_transparency_param_0, .param .u64 kernel_gpu_shader_eval_curve_shadow_transparency_param_1, .param .u32 kernel_gpu_shader_eval_curve_shadow_transparency_param_2, .param .u32 kernel_gpu_shader_eval_curve_shadow_transparency_param_3 ) .maxntid 512, 1, 1 .minnctapersm 1 { .local .align 16 .b8 __local_depot257[6448]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<148>; .reg .f32 %f<565>; .reg .b32 %r<168>; .reg .b64 %rd<214>; mov.u64 %SPL, __local_depot257; cvta.local.u64 %SP, %SPL; ld.param.u64 %rd16, [kernel_gpu_shader_eval_curve_shadow_transparency_param_0]; ld.param.u64 %rd17, [kernel_gpu_shader_eval_curve_shadow_transparency_param_1]; ld.param.u32 %r43, [kernel_gpu_shader_eval_curve_shadow_transparency_param_2]; ld.param.u32 %r44, [kernel_gpu_shader_eval_curve_shadow_transparency_param_3]; mov.u32 %r45, %ntid.x; mov.u32 %r46, %ctaid.x; mov.u32 %r47, %tid.x; mad.lo.s32 %r1, %r46, %r45, %r47; setp.ge.s32 %p1, %r1, %r44; @%p1 bra $L__BB257_174; add.s32 %r48, %r1, %r43; cvta.to.global.u64 %rd18, %rd16; mul.wide.s32 %rd19, %r48, 16; add.s64 %rd20, %rd18, %rd19; ld.global.f32 %f1, [%rd20+8]; ld.global.u32 %r49, [%rd20+12]; shl.b32 %r2, %r49, 7; add.u64 %rd21, %SP, 1024; add.u64 %rd22, %SPL, 1024; mov.u32 %r50, 1056964608; mov.u32 %r51, -1; ld.global.u32 %r52, [%rd20]; ld.global.u32 %r53, [%rd20+4]; st.local.v2.u32 [%rd22+80], {%r51, %r50}; mov.u32 %r54, 0; st.local.u32 [%rd22+88], %r54; ld.const.u64 %rd23, [kernel_params+2264]; cvta.to.global.u64 %rd24, %rd23; mul.wide.s32 %rd25, %r53, 16; add.s64 %rd1, %rd24, %rd25; ld.const.u64 %rd26, [kernel_params+2408]; cvta.to.global.u64 %rd27, %rd26; ld.global.u32 %r55, [%rd1]; and.b32 %r167, %r55, 4194303; shl.b32 %r56, %r55, 5; cvt.u64.u32 %rd28, %r56; and.b64 %rd29, %rd28, 134217696; add.s64 %rd30, %rd27, %rd29; mov.b32 %r57, %f1; or.b32 %r58, %r2, 2; ld.global.u32 %r59, [%rd30+16]; st.local.v4.u32 [%rd22+64], {%r58, %r57, %r54, %r52}; ld.const.u64 %rd31, [kernel_params+2184]; cvta.to.global.u64 %rd32, %rd31; cvt.s64.s32 %rd2, %r52; mul.wide.s32 %rd33, %r52, 4; add.s64 %rd34, %rd32, %rd33; ld.global.u32 %r4, [%rd34]; st.local.v4.u32 [%rd22+48], {%r55, %r59, %r4, %r53}; and.b32 %r5, %r4, 2; setp.eq.s32 %p2, %r5, 0; @%p2 bra $L__BB257_8; ld.const.u64 %rd35, [kernel_params+2160]; cvta.to.global.u64 %rd36, %rd35; mul.lo.s64 %rd37, %rd2, 208; add.s64 %rd38, %rd36, %rd37; ld.const.u64 %rd39, [kernel_params+2176]; cvta.to.global.u64 %rd40, %rd39; ld.global.u32 %rd41, [%rd38+168]; ld.global.u32 %r60, [%rd38+152]; shl.b32 %r61, %r60, 1; cvt.rn.f32.s32 %f97, %r61; mul.ftz.f32 %f98, %f97, 0f3F000000; cvt.rzi.ftz.s32.f32 %r62, %f98; add.s32 %r63, %r61, -1; min.s32 %r64, %r62, %r63; cvt.rn.f32.s32 %f99, %r64; sub.ftz.f32 %f2, %f98, %f99; cvt.s64.s32 %rd42, %r64; add.s64 %rd43, %rd42, %rd41; shl.b64 %rd44, %rd43, 6; add.s64 %rd3, %rd40, %rd44; ld.global.v4.f32 {%f100, %f101, %f102, %f103}, [%rd3]; ld.global.v4.f32 {%f104, %f105, %f106, %f107}, [%rd3+64]; mul.ftz.f32 %f108, %f101, %f105; fma.rn.ftz.f32 %f109, %f100, %f104, %f108; mul.ftz.f32 %f110, %f103, %f107; fma.rn.ftz.f32 %f111, %f102, %f106, %f110; add.ftz.f32 %f19, %f109, %f111; setp.gt.ftz.f32 %p3, %f19, 0f3F7FDF3B; mov.f32 %f112, 0f3F800000; sub.ftz.f32 %f20, %f112, %f2; @%p3 bra $L__BB257_4; bra.uni $L__BB257_3; $L__BB257_4: mul.ftz.f32 %f166, %f100, %f20; fma.rn.ftz.f32 %f167, %f104, %f2, %f166; mul.ftz.f32 %f168, %f101, %f20; fma.rn.ftz.f32 %f169, %f105, %f2, %f168; mul.ftz.f32 %f170, %f102, %f20; fma.rn.ftz.f32 %f171, %f106, %f2, %f170; mul.ftz.f32 %f172, %f103, %f20; fma.rn.ftz.f32 %f173, %f107, %f2, %f172; mul.ftz.f32 %f174, %f169, %f169; fma.rn.ftz.f32 %f175, %f167, %f167, %f174; mul.ftz.f32 %f176, %f173, %f173; fma.rn.ftz.f32 %f177, %f171, %f171, %f176; add.ftz.f32 %f178, %f175, %f177; rsqrt.approx.ftz.f32 %f179, %f178; mul.ftz.f32 %f539, %f173, %f179; mul.ftz.f32 %f538, %f171, %f179; mul.ftz.f32 %f537, %f169, %f179; mul.ftz.f32 %f536, %f167, %f179; bra.uni $L__BB257_5; $L__BB257_3: mov.f32 %f113, 0fBF800000; max.ftz.f32 %f114, %f19, %f113; min.ftz.f32 %f116, %f114, %f112; abs.ftz.f32 %f117, %f116; sub.ftz.f32 %f118, %f112, %f117; mul.ftz.f32 %f119, %f118, 0f3F000000; sqrt.approx.ftz.f32 %f120, %f119; setp.gt.ftz.f32 %p4, %f117, 0f3F11EB85; selp.f32 %f121, %f120, %f117, %p4; mul.ftz.f32 %f122, %f121, %f121; mov.f32 %f123, 0f3C94D2E9; mov.f32 %f124, 0f3D53F941; fma.rn.ftz.f32 %f125, %f124, %f122, %f123; mov.f32 %f126, 0f3D3F841F; fma.rn.ftz.f32 %f127, %f125, %f122, %f126; mov.f32 %f128, 0f3D994929; fma.rn.ftz.f32 %f129, %f127, %f122, %f128; mov.f32 %f130, 0f3E2AAB94; fma.rn.ftz.f32 %f131, %f129, %f122, %f130; mul.ftz.f32 %f132, %f122, %f131; fma.rn.ftz.f32 %f133, %f132, %f121, %f121; add.ftz.f32 %f134, %f133, %f133; mov.f32 %f135, 0f3FC90FDB; sub.ftz.f32 %f136, %f135, %f133; selp.f32 %f137, %f134, %f136, %p4; setp.lt.ftz.f32 %p5, %f116, 0f00000000; mov.f32 %f138, 0f40490FDB; sub.ftz.f32 %f139, %f138, %f137; selp.f32 %f140, %f139, %f137, %p5; mul.ftz.f32 %f141, %f100, %f19; mul.ftz.f32 %f142, %f101, %f19; mul.ftz.f32 %f143, %f102, %f19; mul.ftz.f32 %f144, %f103, %f19; sub.ftz.f32 %f145, %f104, %f141; sub.ftz.f32 %f146, %f105, %f142; sub.ftz.f32 %f147, %f106, %f143; sub.ftz.f32 %f148, %f107, %f144; mul.ftz.f32 %f149, %f146, %f146; fma.rn.ftz.f32 %f150, %f145, %f145, %f149; mul.ftz.f32 %f151, %f148, %f148; fma.rn.ftz.f32 %f152, %f147, %f147, %f151; add.ftz.f32 %f153, %f150, %f152; rsqrt.approx.ftz.f32 %f154, %f153; mul.ftz.f32 %f155, %f145, %f154; mul.ftz.f32 %f156, %f146, %f154; mul.ftz.f32 %f157, %f147, %f154; mul.ftz.f32 %f158, %f148, %f154; mul.ftz.f32 %f159, %f2, %f140; cos.approx.ftz.f32 %f160, %f159; sin.approx.ftz.f32 %f161, %f159; mul.ftz.f32 %f162, %f155, %f161; mul.ftz.f32 %f163, %f156, %f161; mul.ftz.f32 %f164, %f157, %f161; mul.ftz.f32 %f165, %f158, %f161; fma.rn.ftz.f32 %f539, %f103, %f160, %f165; fma.rn.ftz.f32 %f538, %f102, %f160, %f164; fma.rn.ftz.f32 %f537, %f101, %f160, %f163; fma.rn.ftz.f32 %f536, %f100, %f160, %f162; $L__BB257_5: ld.global.v4.f32 {%f180, %f181, %f182, %f183}, [%rd3+16]; ld.global.v4.f32 {%f188, %f189, %f190, %f191}, [%rd3+80]; mul.ftz.f32 %f196, %f2, %f188; mul.ftz.f32 %f197, %f2, %f189; mul.ftz.f32 %f198, %f2, %f190; mul.ftz.f32 %f199, %f2, %f191; fma.rn.ftz.f32 %f555, %f20, %f180, %f196; fma.rn.ftz.f32 %f551, %f20, %f181, %f197; fma.rn.ftz.f32 %f547, %f20, %f182, %f198; fma.rn.ftz.f32 %f200, %f20, %f183, %f199; ld.global.v4.f32 {%f201, %f202, %f203, %f204}, [%rd3+32]; ld.global.v4.f32 {%f209, %f210, %f211, %f212}, [%rd3+96]; mul.ftz.f32 %f217, %f2, %f209; mul.ftz.f32 %f218, %f2, %f210; mul.ftz.f32 %f219, %f2, %f211; mul.ftz.f32 %f220, %f2, %f212; fma.rn.ftz.f32 %f221, %f20, %f201, %f217; fma.rn.ftz.f32 %f222, %f20, %f202, %f218; fma.rn.ftz.f32 %f223, %f20, %f203, %f219; fma.rn.ftz.f32 %f224, %f20, %f204, %f220; ld.global.v4.f32 {%f225, %f226, %f227, %f228}, [%rd3+48]; ld.global.v4.f32 {%f233, %f234, %f235, %f236}, [%rd3+112]; mul.ftz.f32 %f241, %f2, %f233; mul.ftz.f32 %f242, %f2, %f234; mul.ftz.f32 %f243, %f2, %f235; mul.ftz.f32 %f244, %f2, %f236; fma.rn.ftz.f32 %f245, %f20, %f225, %f241; fma.rn.ftz.f32 %f246, %f20, %f226, %f242; fma.rn.ftz.f32 %f247, %f20, %f227, %f243; fma.rn.ftz.f32 %f248, %f20, %f228, %f244; mul.ftz.f32 %f249, %f539, 0f3FB504F3; mul.ftz.f32 %f250, %f536, 0f3FB504F3; mul.ftz.f32 %f251, %f537, 0f3FB504F3; mul.ftz.f32 %f252, %f538, 0f3FB504F3; mul.ftz.f32 %f253, %f249, %f250; mul.ftz.f32 %f254, %f249, %f251; mul.ftz.f32 %f255, %f249, %f252; mul.ftz.f32 %f256, %f250, %f250; mul.ftz.f32 %f257, %f250, %f251; mul.ftz.f32 %f258, %f250, %f252; mul.ftz.f32 %f259, %f251, %f251; mul.ftz.f32 %f260, %f251, %f252; mul.ftz.f32 %f261, %f252, %f252; sub.ftz.f32 %f263, %f112, %f259; sub.ftz.f32 %f264, %f263, %f261; sub.ftz.f32 %f265, %f257, %f255; add.ftz.f32 %f266, %f254, %f258; add.ftz.f32 %f267, %f255, %f257; sub.ftz.f32 %f268, %f112, %f256; sub.ftz.f32 %f269, %f268, %f261; sub.ftz.f32 %f270, %f260, %f253; sub.ftz.f32 %f271, %f258, %f254; add.ftz.f32 %f272, %f253, %f260; sub.ftz.f32 %f273, %f268, %f259; mul.ftz.f32 %f274, %f265, %f223; fma.rn.ftz.f32 %f275, %f264, %f200, %f274; fma.rn.ftz.f32 %f558, %f266, %f246, %f275; mul.ftz.f32 %f276, %f265, %f224; fma.rn.ftz.f32 %f277, %f264, %f221, %f276; fma.rn.ftz.f32 %f557, %f266, %f247, %f277; mul.ftz.f32 %f278, %f265, %f245; fma.rn.ftz.f32 %f279, %f264, %f222, %f278; fma.rn.ftz.f32 %f556, %f266, %f248, %f279; mul.ftz.f32 %f280, %f269, %f223; fma.rn.ftz.f32 %f281, %f267, %f200, %f280; fma.rn.ftz.f32 %f554, %f270, %f246, %f281; mul.ftz.f32 %f282, %f269, %f224; fma.rn.ftz.f32 %f283, %f267, %f221, %f282; fma.rn.ftz.f32 %f553, %f270, %f247, %f283; mul.ftz.f32 %f284, %f269, %f245; fma.rn.ftz.f32 %f285, %f267, %f222, %f284; fma.rn.ftz.f32 %f552, %f270, %f248, %f285; mul.ftz.f32 %f286, %f272, %f223; fma.rn.ftz.f32 %f287, %f271, %f200, %f286; fma.rn.ftz.f32 %f550, %f273, %f246, %f287; mul.ftz.f32 %f288, %f272, %f224; fma.rn.ftz.f32 %f289, %f271, %f221, %f288; fma.rn.ftz.f32 %f549, %f273, %f247, %f289; mul.ftz.f32 %f290, %f272, %f245; fma.rn.ftz.f32 %f291, %f271, %f222, %f290; fma.rn.ftz.f32 %f548, %f273, %f248, %f291; st.local.v4.f32 [%rd22+144], {%f558, %f557, %f556, %f555}; st.local.v4.f32 [%rd22+160], {%f554, %f553, %f552, %f551}; st.local.v4.f32 [%rd22+176], {%f550, %f549, %f548, %f547}; mul.ftz.f32 %f292, %f553, %f548; mul.ftz.f32 %f45, %f549, %f552; sub.ftz.f32 %f542, %f292, %f45; mul.ftz.f32 %f47, %f549, %f556; mul.ftz.f32 %f293, %f557, %f548; sub.ftz.f32 %f541, %f47, %f293; mul.ftz.f32 %f49, %f557, %f552; mul.ftz.f32 %f294, %f553, %f556; sub.ftz.f32 %f540, %f49, %f294; mul.ftz.f32 %f295, %f554, %f541; fma.rn.ftz.f32 %f296, %f558, %f542, %f295; fma.rn.ftz.f32 %f546, %f550, %f540, %f296; setp.neu.ftz.f32 %p6, %f546, 0f00000000; mov.f32 %f543, %f548; mov.f32 %f544, %f553; mov.f32 %f545, %f558; @%p6 bra $L__BB257_7; add.ftz.f32 %f545, %f558, 0f322BCC77; add.ftz.f32 %f543, %f548, 0f322BCC77; add.ftz.f32 %f544, %f553, 0f322BCC77; mul.ftz.f32 %f297, %f544, %f543; sub.ftz.f32 %f542, %f297, %f45; mul.ftz.f32 %f298, %f557, %f543; sub.ftz.f32 %f541, %f47, %f298; mul.ftz.f32 %f299, %f544, %f556; sub.ftz.f32 %f540, %f49, %f299; mul.ftz.f32 %f300, %f554, %f541; fma.rn.ftz.f32 %f301, %f545, %f542, %f300; fma.rn.ftz.f32 %f302, %f550, %f540, %f301; setp.eq.ftz.f32 %p7, %f302, 0f00000000; selp.f32 %f546, 0f7F7FFFFF, %f302, %p7; $L__BB257_7: mul.ftz.f32 %f303, %f554, %f543; mul.ftz.f32 %f304, %f550, %f552; sub.ftz.f32 %f305, %f304, %f303; mul.ftz.f32 %f306, %f550, %f556; mul.ftz.f32 %f307, %f543, %f545; sub.ftz.f32 %f308, %f307, %f306; mul.ftz.f32 %f309, %f552, %f545; mul.ftz.f32 %f310, %f554, %f556; sub.ftz.f32 %f311, %f310, %f309; mul.ftz.f32 %f312, %f550, %f544; mul.ftz.f32 %f313, %f554, %f549; sub.ftz.f32 %f314, %f313, %f312; mul.ftz.f32 %f315, %f549, %f545; mul.ftz.f32 %f316, %f550, %f557; sub.ftz.f32 %f317, %f316, %f315; mul.ftz.f32 %f318, %f554, %f557; mul.ftz.f32 %f319, %f544, %f545; sub.ftz.f32 %f320, %f319, %f318; div.approx.ftz.f32 %f321, %f541, %f546; mul.ftz.f32 %f322, %f551, %f321; div.approx.ftz.f32 %f323, %f542, %f546; fma.rn.ftz.f32 %f324, %f555, %f323, %f322; div.approx.ftz.f32 %f325, %f540, %f546; fma.rn.ftz.f32 %f326, %f547, %f325, %f324; div.approx.ftz.f32 %f327, %f308, %f546; mul.ftz.f32 %f328, %f551, %f327; div.approx.ftz.f32 %f329, %f305, %f546; fma.rn.ftz.f32 %f330, %f555, %f329, %f328; div.approx.ftz.f32 %f331, %f311, %f546; fma.rn.ftz.f32 %f332, %f547, %f331, %f330; div.approx.ftz.f32 %f333, %f317, %f546; mul.ftz.f32 %f334, %f551, %f333; div.approx.ftz.f32 %f335, %f314, %f546; fma.rn.ftz.f32 %f336, %f555, %f335, %f334; div.approx.ftz.f32 %f337, %f320, %f546; fma.rn.ftz.f32 %f338, %f547, %f337, %f336; neg.ftz.f32 %f339, %f326; st.local.v4.f32 [%rd22+192], {%f323, %f321, %f325, %f339}; neg.ftz.f32 %f340, %f332; st.local.v4.f32 [%rd22+208], {%f329, %f327, %f331, %f340}; neg.ftz.f32 %f341, %f338; st.local.v4.f32 [%rd22+224], {%f335, %f333, %f337, %f341}; $L__BB257_8: shr.s32 %r65, %r2, 7; ld.global.u32 %r66, [%rd1+4]; add.s32 %r67, %r65, %r66; add.s32 %r68, %r67, -1; max.s32 %r69, %r68, %r66; add.s32 %r70, %r67, 2; ld.global.u32 %r71, [%rd1+8]; add.s32 %r72, %r66, %r71; add.s32 %r73, %r72, -1; min.s32 %r74, %r70, %r73; ld.const.u64 %rd49, [kernel_params+2272]; cvta.to.global.u64 %rd50, %rd49; mul.wide.s32 %rd51, %r69, 16; add.s64 %rd52, %rd50, %rd51; ld.global.v4.f32 {%f342, %f343, %f344, %f345}, [%rd52]; mul.wide.s32 %rd53, %r67, 16; add.s64 %rd54, %rd50, %rd53; ld.global.v4.f32 {%f349, %f350, %f351, %f352}, [%rd54]; ld.global.v4.f32 {%f356, %f357, %f358, %f359}, [%rd54+16]; mul.wide.s32 %rd55, %r74, 16; add.s64 %rd56, %rd50, %rd55; ld.global.v4.f32 {%f363, %f364, %f365, %f366}, [%rd56]; mov.f32 %f370, 0f3F800000; sub.ftz.f32 %f371, %f370, %f1; mul.ftz.f32 %f372, %f371, %f371; add.ftz.f32 %f373, %f371, %f371; mul.ftz.f32 %f374, %f1, %f373; sub.ftz.f32 %f375, %f374, %f372; mul.ftz.f32 %f376, %f1, 0f40400000; add.ftz.f32 %f377, %f376, 0fC0A00000; add.ftz.f32 %f378, %f1, %f1; mul.ftz.f32 %f379, %f378, %f377; fma.rn.ftz.f32 %f380, %f1, %f376, %f379; add.ftz.f32 %f381, %f376, 0f40000000; mul.ftz.f32 %f382, %f373, %f381; mul.ftz.f32 %f383, %f371, 0f40400000; mul.ftz.f32 %f384, %f371, %f383; sub.ftz.f32 %f385, %f382, %f384; mul.ftz.f32 %f386, %f371, 0fC0000000; mul.ftz.f32 %f387, %f1, %f386; fma.rn.ftz.f32 %f388, %f1, %f1, %f387; mul.ftz.f32 %f389, %f349, %f380; mul.ftz.f32 %f390, %f350, %f380; mul.ftz.f32 %f391, %f351, %f380; fma.rn.ftz.f32 %f392, %f342, %f375, %f389; fma.rn.ftz.f32 %f393, %f343, %f375, %f390; fma.rn.ftz.f32 %f394, %f344, %f375, %f391; fma.rn.ftz.f32 %f395, %f356, %f385, %f392; fma.rn.ftz.f32 %f396, %f357, %f385, %f393; fma.rn.ftz.f32 %f397, %f358, %f385, %f394; fma.rn.ftz.f32 %f398, %f363, %f388, %f395; fma.rn.ftz.f32 %f399, %f364, %f388, %f396; fma.rn.ftz.f32 %f400, %f365, %f388, %f397; mul.ftz.f32 %f78, %f398, 0f3F000000; mul.ftz.f32 %f79, %f399, 0f3F000000; mul.ftz.f32 %f80, %f400, 0f3F000000; st.local.v2.f32 [%rd22], {%f78, %f79}; st.local.f32 [%rd22+8], %f80; st.local.f32 [%rd22+116], %f78; st.local.f32 [%rd22+120], %f79; st.local.f32 [%rd22+124], %f80; and.b32 %r75, %r4, 4; setp.ne.s32 %p8, %r75, 0; @%p8 bra $L__BB257_15; @%p2 bra $L__BB257_11; bra.uni $L__BB257_10; $L__BB257_11: ld.const.u64 %rd58, [kernel_params+2160]; cvta.to.global.u64 %rd59, %rd58; mul.lo.s64 %rd60, %rd2, 208; add.s64 %rd61, %rd59, %rd60; ld.global.v4.f32 {%f410, %f411, %f412, %f413}, [%rd61]; ld.global.v4.f32 {%f418, %f419, %f420, %f421}, [%rd61+16]; ld.global.v4.f32 {%f426, %f427, %f428, %f429}, [%rd61+32]; mul.ftz.f32 %f434, %f411, %f79; fma.rn.ftz.f32 %f435, %f78, %f410, %f434; fma.rn.ftz.f32 %f436, %f412, %f80, %f435; add.ftz.f32 %f559, %f413, %f436; mul.ftz.f32 %f437, %f419, %f79; fma.rn.ftz.f32 %f438, %f78, %f418, %f437; fma.rn.ftz.f32 %f439, %f420, %f80, %f438; add.ftz.f32 %f560, %f421, %f439; mul.ftz.f32 %f440, %f427, %f79; fma.rn.ftz.f32 %f441, %f78, %f426, %f440; fma.rn.ftz.f32 %f442, %f428, %f80, %f441; add.ftz.f32 %f561, %f429, %f442; bra.uni $L__BB257_12; $L__BB257_10: mul.ftz.f32 %f401, %f79, %f557; fma.rn.ftz.f32 %f402, %f78, %f558, %f401; fma.rn.ftz.f32 %f403, %f80, %f556, %f402; add.ftz.f32 %f559, %f555, %f403; mul.ftz.f32 %f404, %f79, %f553; fma.rn.ftz.f32 %f405, %f78, %f554, %f404; fma.rn.ftz.f32 %f406, %f80, %f552, %f405; add.ftz.f32 %f560, %f551, %f406; mul.ftz.f32 %f407, %f79, %f549; fma.rn.ftz.f32 %f408, %f78, %f550, %f407; fma.rn.ftz.f32 %f409, %f80, %f548, %f408; add.ftz.f32 %f561, %f547, %f409; $L__BB257_12: st.local.v2.f32 [%rd22], {%f559, %f560}; st.local.f32 [%rd22+8], %f561; @%p2 bra $L__BB257_14; bra.uni $L__BB257_13; $L__BB257_14: ld.const.u64 %rd66, [kernel_params+2160]; cvta.to.global.u64 %rd67, %rd66; mul.lo.s64 %rd68, %rd2, 208; add.s64 %rd69, %rd67, %rd68; ld.global.v4.f32 {%f452, %f453, %f454, %f455}, [%rd69]; ld.global.v4.f32 {%f459, %f460, %f461, %f462}, [%rd69+16]; ld.global.v4.f32 {%f466, %f467, %f468, %f469}, [%rd69+32]; mul.ftz.f32 %f473, %f453, %f79; fma.rn.ftz.f32 %f474, %f78, %f452, %f473; fma.rn.ftz.f32 %f475, %f454, %f80, %f474; mul.ftz.f32 %f476, %f460, %f79; fma.rn.ftz.f32 %f477, %f78, %f459, %f476; fma.rn.ftz.f32 %f478, %f461, %f80, %f477; mul.ftz.f32 %f479, %f467, %f79; fma.rn.ftz.f32 %f480, %f78, %f466, %f479; fma.rn.ftz.f32 %f481, %f468, %f80, %f480; st.local.f32 [%rd22+116], %f475; st.local.f32 [%rd22+120], %f478; st.local.f32 [%rd22+124], %f481; bra.uni $L__BB257_15; $L__BB257_13: mul.ftz.f32 %f443, %f79, %f557; fma.rn.ftz.f32 %f444, %f78, %f558, %f443; fma.rn.ftz.f32 %f445, %f80, %f556, %f444; mul.ftz.f32 %f446, %f79, %f553; fma.rn.ftz.f32 %f447, %f78, %f554, %f446; fma.rn.ftz.f32 %f448, %f80, %f552, %f447; mul.ftz.f32 %f449, %f79, %f549; fma.rn.ftz.f32 %f450, %f78, %f550, %f449; fma.rn.ftz.f32 %f451, %f80, %f548, %f450; st.local.f32 [%rd22+116], %f445; st.local.f32 [%rd22+120], %f448; st.local.f32 [%rd22+124], %f451; $L__BB257_15: st.local.u32 [%rd22+12], %r54; add.s64 %rd6, %rd22, 16; mov.f32 %f482, 0f00000000; st.local.v4.f32 [%rd22+16], {%f482, %f482, %f482, %f482}; st.local.v4.f32 [%rd22+32], {%f482, %f482, %f482, %f482}; st.local.v2.f32 [%rd22+128], {%f482, %f482}; st.local.u32 [%rd22+136], %r54; st.local.u32 [%rd22+92], %r54; st.local.v4.f32 [%rd22+96], {%f482, %f482, %f482, %f482}; st.local.u32 [%rd22+112], %r54; st.local.u32 [%rd22+260], %r54; add.s64 %rd8, %rd22, 264; st.local.u32 [%rd22+264], %r54; ld.const.u64 %rd74, [kernel_params+2400]; cvta.to.global.u64 %rd5, %rd74; add.s64 %rd7, %rd22, 56; add.u64 %rd75, %SP, 0; add.u64 %rd9, %SPL, 0; bra.uni $L__BB257_16; $L__BB257_183: { // callseq 1674, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z18svm_node_mix_floatP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1674 $L__BB257_16: mov.u32 %r6, %r167; mul.wide.s32 %rd76, %r6, 16; add.s64 %rd10, %rd5, %rd76; ld.global.v4.u32 {%r78, %r79, %r80, %r81}, [%rd10]; add.s32 %r167, %r6, 1; mul.wide.u32 %rd77, %r79, 4; add.s64 %rd11, %rd9, %rd77; mul.wide.u32 %rd78, %r80, 4; add.s64 %rd12, %rd9, %rd78; setp.gt.s32 %p11, %r78, 1; @%p11 bra $L__BB257_18; setp.eq.s32 %p139, %r78, 1; mov.u32 %r167, %r79; @%p139 bra $L__BB257_16; bra.uni $L__BB257_170; $L__BB257_18: setp.gt.s32 %p12, %r78, 49; @%p12 bra $L__BB257_98; setp.gt.s32 %p76, %r78, 25; @%p76 bra $L__BB257_60; setp.gt.s32 %p108, %r78, 13; @%p108 bra $L__BB257_41; setp.gt.s32 %p124, %r78, 7; @%p124 bra $L__BB257_31; setp.gt.s32 %p132, %r78, 4; @%p132 bra $L__BB257_27; setp.eq.s32 %p136, %r78, 2; @%p136 bra $L__BB257_249; setp.eq.s32 %p137, %r78, 3; @%p137 bra $L__BB257_248; setp.eq.s32 %p138, %r78, 4; @%p138 bra $L__BB257_26; bra.uni $L__BB257_170; $L__BB257_26: mov.u32 %r129, 4; { // callseq 1753, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r129, %r79, %r80, %r81}; call.uni _Z27svm_node_closure_backgroundP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1753 bra.uni $L__BB257_16; $L__BB257_98: setp.gt.s32 %p13, %r78, 73; @%p13 bra $L__BB257_138; setp.gt.s32 %p45, %r78, 61; @%p45 bra $L__BB257_119; setp.gt.s32 %p61, %r78, 55; @%p61 bra $L__BB257_110; setp.gt.s32 %p69, %r78, 52; @%p69 bra $L__BB257_106; setp.eq.s32 %p73, %r78, 50; @%p73 bra $L__BB257_216; setp.eq.s32 %p74, %r78, 51; @%p74 bra $L__BB257_215; setp.eq.s32 %p75, %r78, 52; @%p75 bra $L__BB257_105; bra.uni $L__BB257_170; $L__BB257_105: { // callseq 1714, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; add.s32 %r156, %r6, 1; st.param.b32 [param3+0], %r156; .param .b32 retval0; call.uni (retval0), _Z24svm_node_texture_mappingPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1714 bra.uni $L__BB257_16; $L__BB257_60: setp.gt.s32 %p77, %r78, 37; @%p77 bra $L__BB257_80; setp.gt.s32 %p93, %r78, 31; @%p93 bra $L__BB257_71; setp.gt.s32 %p101, %r78, 28; @%p101 bra $L__BB257_67; setp.eq.s32 %p105, %r78, 26; @%p105 bra $L__BB257_232; setp.eq.s32 %p106, %r78, 27; @%p106 bra $L__BB257_231; setp.eq.s32 %p107, %r78, 28; @%p107 bra $L__BB257_66; bra.uni $L__BB257_170; $L__BB257_66: mov.u32 %r110, 28; { // callseq 1736, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r110; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z21svm_node_attr_bump_dyPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1736 bra.uni $L__BB257_16; $L__BB257_138: setp.gt.s32 %p14, %r78, 85; @%p14 bra $L__BB257_161; setp.gt.s32 %p30, %r78, 79; @%p30 bra $L__BB257_152; setp.gt.s32 %p38, %r78, 76; @%p38 bra $L__BB257_148; setp.eq.s32 %p42, %r78, 74; @%p42 bra $L__BB257_200; setp.eq.s32 %p43, %r78, 75; @%p43 bra $L__BB257_199; setp.eq.s32 %p44, %r78, 76; @%p44 bra $L__BB257_144; bra.uni $L__BB257_170; $L__BB257_144: setp.eq.s32 %p141, %r81, 255; @%p141 bra $L__BB257_16; add.s64 %rd13, %rd11, 4; add.s64 %rd14, %rd11, 8; setp.eq.s32 %p142, %r80, 0; mul.wide.u32 %rd105, %r81, 4; add.s64 %rd15, %rd9, %rd105; @%p142 bra $L__BB257_198; setp.eq.s32 %p143, %r80, 1; @%p143 bra $L__BB257_197; bra.uni $L__BB257_147; $L__BB257_197: ld.local.f32 %f490, [%rd13]; st.local.f32 [%rd15], %f490; bra.uni $L__BB257_16; $L__BB257_41: setp.gt.s32 %p109, %r78, 19; @%p109 bra $L__BB257_51; setp.gt.s32 %p117, %r78, 16; @%p117 bra $L__BB257_47; setp.eq.s32 %p121, %r78, 14; @%p121 bra $L__BB257_240; setp.eq.s32 %p122, %r78, 15; @%p122 bra $L__BB257_239; setp.eq.s32 %p123, %r78, 16; @%p123 bra $L__BB257_46; bra.uni $L__BB257_170; $L__BB257_46: mov.u32 %r117, 16; { // callseq 1747, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r117; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z13svm_node_attrILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1747 bra.uni $L__BB257_16; $L__BB257_119: setp.gt.s32 %p46, %r78, 67; @%p46 bra $L__BB257_129; setp.gt.s32 %p54, %r78, 64; @%p54 bra $L__BB257_125; setp.eq.s32 %p58, %r78, 62; @%p58 bra $L__BB257_208; setp.eq.s32 %p59, %r78, 63; @%p59 bra $L__BB257_207; setp.eq.s32 %p60, %r78, 64; @%p60 bra $L__BB257_124; bra.uni $L__BB257_170; $L__BB257_124: mov.u32 %r94, 64; { // callseq 1702, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r94; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r149, %r6, 1; st.param.b32 [param2+0], %r149; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_brickPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1702 bra.uni $L__BB257_16; $L__BB257_80: setp.gt.s32 %p78, %r78, 43; @%p78 bra $L__BB257_89; setp.gt.s32 %p86, %r78, 40; @%p86 bra $L__BB257_85; setp.eq.s32 %p90, %r78, 38; @%p90 bra $L__BB257_224; setp.eq.s32 %p91, %r78, 39; @%p91 bra $L__BB257_223; setp.eq.s32 %p92, %r78, 40; @%p92 bra $L__BB257_16; bra.uni $L__BB257_170; $L__BB257_161: setp.gt.s32 %p15, %r78, 92; @%p15 bra $L__BB257_175; setp.gt.s32 %p23, %r78, 88; @%p23 bra $L__BB257_167; setp.eq.s32 %p27, %r78, 86; @%p27 bra $L__BB257_189; setp.eq.s32 %p28, %r78, 87; @%p28 bra $L__BB257_188; setp.eq.s32 %p29, %r78, 88; @%p29 bra $L__BB257_166; bra.uni $L__BB257_170; $L__BB257_166: mov.u32 %r85, 88; ld.local.f32 %f483, [%rd6+-4]; ld.local.v2.f32 {%f484, %f485}, [%rd6]; { // callseq 1680, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.f32 [param0+0], %f483; .param .b32 param1; st.param.f32 [param1+0], %f484; .param .b32 param2; st.param.f32 [param2+0], %f485; .param .b64 param3; st.param.b64 [param3+0], %rd75; .param .align 4 .b8 param4[16]; st.param.b32 [param4+0], %r85; st.param.b32 [param4+4], %r79; st.param.b32 [param4+8], %r80; st.param.b32 [param4+12], %r81; call.uni _Z14svm_node_bevelILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1, param2, param3, param4 ); } // callseq 1680 bra.uni $L__BB257_16; $L__BB257_31: setp.gt.s32 %p125, %r78, 10; @%p125 bra $L__BB257_37; setp.eq.s32 %p129, %r78, 8; @%p129 bra $L__BB257_245; setp.eq.s32 %p130, %r78, 9; @%p130 bra $L__BB257_243; setp.eq.s32 %p131, %r78, 10; @%p131 bra $L__BB257_35; bra.uni $L__BB257_170; $L__BB257_35: ld.local.f32 %f516, [%rd12]; setp.ltu.ftz.f32 %p144, %f516, 0f3F800000; @%p144 bra $L__BB257_16; add.s32 %r167, %r79, %r167; bra.uni $L__BB257_16; $L__BB257_110: setp.gt.s32 %p62, %r78, 58; @%p62 bra $L__BB257_115; setp.eq.s32 %p66, %r78, 56; @%p66 bra $L__BB257_212; setp.eq.s32 %p67, %r78, 57; @%p67 bra $L__BB257_211; setp.eq.s32 %p68, %r78, 58; @%p68 bra $L__BB257_114; bra.uni $L__BB257_170; $L__BB257_114: mov.u32 %r98, 58; { // callseq 1708, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r98; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z21svm_node_tex_gradientP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1708 bra.uni $L__BB257_16; $L__BB257_71: setp.gt.s32 %p94, %r78, 34; @%p94 bra $L__BB257_76; setp.eq.s32 %p98, %r78, 32; @%p98 bra $L__BB257_228; setp.eq.s32 %p99, %r78, 33; @%p99 bra $L__BB257_227; setp.eq.s32 %p100, %r78, 34; @%p100 bra $L__BB257_75; bra.uni $L__BB257_170; $L__BB257_75: { // callseq 1731, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; call.uni _Z24svm_node_enter_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1731 bra.uni $L__BB257_16; $L__BB257_152: setp.gt.s32 %p31, %r78, 82; @%p31 bra $L__BB257_157; setp.eq.s32 %p35, %r78, 80; @%p35 bra $L__BB257_193; setp.eq.s32 %p36, %r78, 81; @%p36 bra $L__BB257_192; setp.eq.s32 %p37, %r78, 82; @%p37 bra $L__BB257_156; bra.uni $L__BB257_170; $L__BB257_156: mov.u32 %r86, 82; { // callseq 1686, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r86; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z18svm_node_wireframePK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1686 bra.uni $L__BB257_16; $L__BB257_51: setp.gt.s32 %p110, %r78, 22; @%p110 bra $L__BB257_56; setp.eq.s32 %p114, %r78, 20; @%p114 bra $L__BB257_236; setp.eq.s32 %p115, %r78, 21; @%p115 bra $L__BB257_235; setp.eq.s32 %p116, %r78, 22; @%p116 bra $L__BB257_55; bra.uni $L__BB257_170; $L__BB257_55: mov.u32 %r115, 22; { // callseq 1742, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r115; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; .param .b32 param3; add.s32 %r164, %r6, 1; st.param.b32 [param3+0], %r164; .param .b32 retval0; call.uni (retval0), _Z28svm_node_vector_displacementILj63EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1742 bra.uni $L__BB257_16; $L__BB257_129: setp.gt.s32 %p47, %r78, 70; @%p47 bra $L__BB257_134; setp.eq.s32 %p51, %r78, 68; @%p51 bra $L__BB257_204; setp.eq.s32 %p52, %r78, 69; @%p52 bra $L__BB257_203; setp.eq.s32 %p53, %r78, 70; @%p53 bra $L__BB257_133; bra.uni $L__BB257_170; $L__BB257_133: mov.u32 %r90, 70; { // callseq 1696, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r90; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z16svm_node_tangentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1696 bra.uni $L__BB257_16; $L__BB257_89: setp.gt.s32 %p79, %r78, 46; @%p79 bra $L__BB257_94; setp.eq.s32 %p83, %r78, 44; @%p83 bra $L__BB257_220; setp.eq.s32 %p84, %r78, 45; @%p84 bra $L__BB257_219; setp.eq.s32 %p85, %r78, 46; @%p85 bra $L__BB257_93; bra.uni $L__BB257_170; $L__BB257_93: { // callseq 1720, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z19svm_node_brightnessP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1720 bra.uni $L__BB257_16; $L__BB257_175: setp.gt.s32 %p16, %r78, 95; @%p16 bra $L__BB257_180; setp.eq.s32 %p20, %r78, 93; @%p20 bra $L__BB257_16; setp.eq.s32 %p21, %r78, 94; @%p21 bra $L__BB257_184; setp.eq.s32 %p22, %r78, 95; @%p22 bra $L__BB257_179; bra.uni $L__BB257_170; $L__BB257_179: { // callseq 1675, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z18svm_node_mix_colorP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1675 bra.uni $L__BB257_16; $L__BB257_27: setp.eq.s32 %p133, %r78, 5; @%p133 bra $L__BB257_247; setp.eq.s32 %p134, %r78, 6; @%p134 bra $L__BB257_246; setp.eq.s32 %p135, %r78, 7; @%p135 bra $L__BB257_30; bra.uni $L__BB257_170; $L__BB257_30: mov.u32 %r128, 7; { // callseq 1752, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r128; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z24svm_node_emission_weightPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1752 bra.uni $L__BB257_16; $L__BB257_106: setp.eq.s32 %p70, %r78, 53; @%p70 bra $L__BB257_214; setp.eq.s32 %p71, %r78, 54; @%p71 bra $L__BB257_213; setp.eq.s32 %p72, %r78, 55; @%p72 bra $L__BB257_109; bra.uni $L__BB257_170; $L__BB257_109: ld.local.v4.f32 {%f492, %f493, %f494, %f495}, [%rd22]; { // callseq 1711, 0 .reg .b32 temp_param_reg; .param .align 4 .b8 param0[12]; st.param.f32 [param0+0], %f492; st.param.f32 [param0+4], %f493; st.param.f32 [param0+8], %f494; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; .param .b32 param4; st.param.b32 [param4+0], %r81; call.uni _Z15svm_node_cameraPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1711 bra.uni $L__BB257_16; $L__BB257_67: setp.eq.s32 %p102, %r78, 29; @%p102 bra $L__BB257_230; setp.eq.s32 %p103, %r78, 30; @%p103 bra $L__BB257_229; setp.eq.s32 %p104, %r78, 31; @%p104 bra $L__BB257_70; bra.uni $L__BB257_170; $L__BB257_70: mov.u32 %r108, 31; mov.u32 %r109, 768; { // callseq 1733, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b32 param1; st.param.b32 [param1+0], %r109; .param .b64 param2; st.param.b64 [param2+0], %rd75; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r108; st.param.b32 [param3+4], %r79; st.param.b32 [param3+8], %r80; st.param.b32 [param3+12], %r81; .param .b32 param4; add.s32 %r161, %r6, 1; st.param.b32 [param4+0], %r161; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dxPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1733 bra.uni $L__BB257_16; $L__BB257_148: setp.eq.s32 %p39, %r78, 77; @%p39 bra $L__BB257_195; setp.eq.s32 %p40, %r78, 78; @%p40 bra $L__BB257_194; setp.eq.s32 %p41, %r78, 79; @%p41 bra $L__BB257_151; bra.uni $L__BB257_170; $L__BB257_151: { // callseq 1689, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r144, %r6, 1; st.param.b32 [param4+0], %r144; .param .b32 retval0; call.uni (retval0), _Z20svm_node_combine_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1689 bra.uni $L__BB257_16; $L__BB257_47: setp.eq.s32 %p118, %r78, 17; @%p118 bra $L__BB257_238; setp.eq.s32 %p119, %r78, 18; @%p119 bra $L__BB257_237; setp.eq.s32 %p120, %r78, 19; @%p120 bra $L__BB257_50; bra.uni $L__BB257_170; $L__BB257_50: { // callseq 1744, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z25svm_node_geometry_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1744 bra.uni $L__BB257_16; $L__BB257_125: setp.eq.s32 %p55, %r78, 65; @%p55 bra $L__BB257_206; setp.eq.s32 %p56, %r78, 66; @%p56 bra $L__BB257_205; setp.eq.s32 %p57, %r78, 67; @%p57 bra $L__BB257_128; bra.uni $L__BB257_170; $L__BB257_128: mov.u32 %r93, 67; { // callseq 1699, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r93, %r79, %r80, %r81}; call.uni _Z22svm_node_light_falloffP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1699 bra.uni $L__BB257_16; $L__BB257_85: setp.eq.s32 %p87, %r78, 41; @%p87 bra $L__BB257_222; setp.eq.s32 %p88, %r78, 42; @%p88 bra $L__BB257_221; setp.eq.s32 %p89, %r78, 43; @%p89 bra $L__BB257_88; bra.uni $L__BB257_170; $L__BB257_88: { // callseq 1723, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r158, %r6, 1; st.param.b32 [param4+0], %r158; .param .b32 retval0; call.uni (retval0), _Z20svm_node_vector_mathPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1723 bra.uni $L__BB257_16; $L__BB257_167: setp.eq.s32 %p24, %r78, 89; @%p24 bra $L__BB257_187; setp.eq.s32 %p25, %r78, 90; @%p25 bra $L__BB257_186; setp.eq.s32 %p26, %r78, 92; @%p26 bra $L__BB257_16; bra.uni $L__BB257_170; $L__BB257_37: setp.eq.s32 %p126, %r78, 11; @%p126 bra $L__BB257_242; setp.eq.s32 %p127, %r78, 12; @%p127 bra $L__BB257_241; setp.eq.s32 %p128, %r78, 13; @%p128 bra $L__BB257_40; bra.uni $L__BB257_170; $L__BB257_40: mov.u32 %r125, 13; mov.u32 %r126, 768; { // callseq 1748, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b32 param1; st.param.b32 [param1+0], %r126; .param .b64 param2; st.param.b64 [param2+0], %rd75; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r125; st.param.b32 [param3+4], %r79; st.param.b32 [param3+8], %r80; st.param.b32 [param3+12], %r81; .param .b32 param4; add.s32 %r165, %r6, 1; st.param.b32 [param4+0], %r165; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_coordPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1748 bra.uni $L__BB257_16; $L__BB257_115: setp.eq.s32 %p63, %r78, 59; @%p63 bra $L__BB257_210; setp.eq.s32 %p64, %r78, 60; @%p64 bra $L__BB257_209; setp.eq.s32 %p65, %r78, 61; @%p65 bra $L__BB257_118; bra.uni $L__BB257_170; $L__BB257_118: mov.u32 %r97, 61; { // callseq 1705, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r97; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r151, %r6, 1; st.param.b32 [param2+0], %r151; .param .b32 retval0; call.uni (retval0), _Z17svm_node_tex_wavePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1705 bra.uni $L__BB257_16; $L__BB257_76: setp.eq.s32 %p95, %r78, 35; @%p95 bra $L__BB257_226; setp.eq.s32 %p96, %r78, 36; @%p96 bra $L__BB257_225; setp.eq.s32 %p97, %r78, 37; @%p97 bra $L__BB257_79; bra.uni $L__BB257_170; $L__BB257_79: mov.u32 %r104, 37; { // callseq 1728, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r104, %r79, %r80, %r81}; call.uni _Z24svm_node_closure_holdoutP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1728 bra.uni $L__BB257_16; $L__BB257_157: setp.eq.s32 %p32, %r78, 83; @%p32 bra $L__BB257_191; setp.eq.s32 %p33, %r78, 84; @%p33 bra $L__BB257_190; setp.eq.s32 %p34, %r78, 85; @%p34 bra $L__BB257_160; bra.uni $L__BB257_170; $L__BB257_160: { // callseq 1683, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r143, %r6, 1; st.param.b32 [param4+0], %r143; .param .b32 retval0; call.uni (retval0), _Z18svm_node_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1683 bra.uni $L__BB257_16; $L__BB257_56: setp.eq.s32 %p111, %r78, 23; @%p111 bra $L__BB257_234; setp.eq.s32 %p112, %r78, 24; @%p112 bra $L__BB257_233; setp.eq.s32 %p113, %r78, 25; @%p113 bra $L__BB257_59; bra.uni $L__BB257_170; $L__BB257_59: { // callseq 1739, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r162, %r6, 1; st.param.b32 [param4+0], %r162; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_noisePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1739 bra.uni $L__BB257_16; $L__BB257_134: setp.eq.s32 %p48, %r78, 71; @%p48 bra $L__BB257_202; setp.eq.s32 %p49, %r78, 72; @%p49 bra $L__BB257_201; setp.eq.s32 %p50, %r78, 73; @%p50 bra $L__BB257_137; bra.uni $L__BB257_170; $L__BB257_137: { // callseq 1693, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r146, %r6, 1; st.param.b32 [param4+0], %r146; .param .b32 retval0; call.uni (retval0), _Z12svm_node_mixPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1693 bra.uni $L__BB257_16; $L__BB257_94: setp.eq.s32 %p80, %r78, 47; @%p80 bra $L__BB257_218; setp.eq.s32 %p81, %r78, 48; @%p81 bra $L__BB257_217; setp.eq.s32 %p82, %r78, 49; @%p82 bra $L__BB257_97; bra.uni $L__BB257_170; $L__BB257_97: { // callseq 1717, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z22svm_node_particle_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1717 bra.uni $L__BB257_16; $L__BB257_180: setp.eq.s32 %p17, %r78, 98; @%p17 bra $L__BB257_250; setp.eq.s32 %p18, %r78, 97; @%p18 bra $L__BB257_185; setp.ne.s32 %p19, %r78, 96; @%p19 bra $L__BB257_170; bra.uni $L__BB257_183; $L__BB257_249: mov.u32 %r131, 2; { // callseq 1755, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r131; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; .param .b32 param3; add.s32 %r166, %r6, 1; st.param.b32 [param3+0], %r166; .param .b32 retval0; call.uni (retval0), _Z21svm_node_closure_bsdfILj63EL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0, param1, param2, param3 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1755 bra.uni $L__BB257_16; $L__BB257_248: mov.u32 %r130, 3; { // callseq 1754, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r130, %r79, %r80, %r81}; call.uni _Z25svm_node_closure_emissionP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1754 bra.uni $L__BB257_16; $L__BB257_216: { // callseq 1716, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z18svm_node_hair_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1716 bra.uni $L__BB257_16; $L__BB257_215: { // callseq 1715, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z19svm_node_point_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1715 bra.uni $L__BB257_16; $L__BB257_232: mov.u32 %r112, 26; { // callseq 1738, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r112; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z17svm_node_set_bumpILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1738 bra.uni $L__BB257_16; $L__BB257_231: mov.u32 %r111, 27; { // callseq 1737, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r111; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z21svm_node_attr_bump_dxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1737 bra.uni $L__BB257_16; $L__BB257_200: { // callseq 1692, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z23svm_node_separate_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1692 bra.uni $L__BB257_16; $L__BB257_199: { // callseq 1691, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z22svm_node_combine_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1691 bra.uni $L__BB257_16; $L__BB257_240: st.local.u32 [%rd12], %r79; bra.uni $L__BB257_16; $L__BB257_239: ld.global.v4.u32 {%r118, %r119, %r120, %r121}, [%rd10+16]; add.s32 %r167, %r6, 2; st.local.u32 [%rd11], %r119; st.local.u32 [%rd11+4], %r120; st.local.u32 [%rd11+8], %r121; bra.uni $L__BB257_16; $L__BB257_208: mov.u32 %r96, 62; { // callseq 1704, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r96; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r150, %r6, 1; st.param.b32 [param2+0], %r150; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_magicPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1704 bra.uni $L__BB257_16; $L__BB257_207: mov.u32 %r95, 63; { // callseq 1703, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r95; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z20svm_node_tex_checkerPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1703 bra.uni $L__BB257_16; $L__BB257_224: { // callseq 1727, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; .param .b32 param4; st.param.b32 [param4+0], %r81; call.uni _Z16svm_node_fresnelP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1727 bra.uni $L__BB257_16; $L__BB257_223: mov.u32 %r103, 39; { // callseq 1726, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 16 .b8 param2[16]; st.param.v4.b32 [param2+0], {%r103, %r79, %r80, %r81}; call.uni _Z21svm_node_layer_weightP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1726 bra.uni $L__BB257_16; $L__BB257_189: { // callseq 1682, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r142, %r6, 1; st.param.b32 [param4+0], %r142; .param .b32 retval0; call.uni (retval0), _Z25svm_node_vector_map_rangePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1682 bra.uni $L__BB257_16; $L__BB257_188: { // callseq 1681, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r141, %r6, 1; st.param.b32 [param4+0], %r141; .param .b32 retval0; call.uni (retval0), _Z14svm_node_clampPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1681 bra.uni $L__BB257_16; $L__BB257_245: mov.u32 %r127, 8; { // callseq 1751, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r127; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z20svm_node_mix_closureP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1751 bra.uni $L__BB257_16; $L__BB257_243: ld.local.f32 %f517, [%rd12]; setp.gtu.ftz.f32 %p145, %f517, 0f00000000; @%p145 bra $L__BB257_16; add.s32 %r167, %r79, %r167; bra.uni $L__BB257_16; $L__BB257_212: mov.u32 %r100, 56; { // callseq 1710, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r100; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z24svm_node_tex_environmentPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1710 bra.uni $L__BB257_16; $L__BB257_211: mov.u32 %r99, 57; { // callseq 1709, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r99; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r154, %r6, 1; st.param.b32 [param2+0], %r154; .param .b32 retval0; call.uni (retval0), _Z16svm_node_tex_skyPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1709 bra.uni $L__BB257_16; $L__BB257_228: mov.u32 %r106, 32; mov.u32 %r107, 768; { // callseq 1732, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b32 param1; st.param.b32 [param1+0], %r107; .param .b64 param2; st.param.b64 [param2+0], %rd75; .param .align 4 .b8 param3[16]; st.param.b32 [param3+0], %r106; st.param.b32 [param3+4], %r79; st.param.b32 [param3+8], %r80; st.param.b32 [param3+12], %r81; .param .b32 param4; add.s32 %r160, %r6, 1; st.param.b32 [param4+0], %r160; .param .b32 retval0; call.uni (retval0), _Z26svm_node_tex_coord_bump_dyPK16KernelGlobalsGPUP10ShaderDatajPf5uint4i, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1732 bra.uni $L__BB257_16; $L__BB257_227: ld.local.f32 %f500, [%rd11]; ld.local.f32 %f501, [%rd11+4]; ld.local.f32 %f502, [%rd11+8]; st.local.f32 [%rd6+-4], %f500; st.local.f32 [%rd6], %f501; st.local.f32 [%rd6+4], %f502; st.local.f32 [%rd12], %f500; st.local.f32 [%rd12+4], %f501; st.local.f32 [%rd12+8], %f502; bra.uni $L__BB257_16; $L__BB257_193: { // callseq 1688, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z22svm_node_vector_rotateP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1688 bra.uni $L__BB257_16; $L__BB257_192: mov.u32 %r87, 81; { // callseq 1687, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r87; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z25svm_node_vector_transformPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1687 bra.uni $L__BB257_16; $L__BB257_236: ld.local.f32 %f503, [%rd11]; ld.local.v4.f32 {%f504, %f505, %f506, %f507}, [%rd22]; ld.local.f32 %f511, [%rd11+4]; ld.local.f32 %f512, [%rd11+8]; add.ftz.f32 %f513, %f512, %f506; add.ftz.f32 %f514, %f511, %f505; add.ftz.f32 %f515, %f503, %f504; st.local.v2.f32 [%rd22], {%f515, %f514}; st.local.f32 [%rd22+8], %f513; bra.uni $L__BB257_16; $L__BB257_235: mov.u32 %r116, 21; { // callseq 1743, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r116; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z21svm_node_displacementILj63EEvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1743 bra.uni $L__BB257_16; $L__BB257_204: mov.u32 %r92, 68; { // callseq 1698, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r92; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z12svm_node_iesPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1698 bra.uni $L__BB257_16; $L__BB257_203: mov.u32 %r91, 69; { // callseq 1697, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r91; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r147, %r6, 1; st.param.b32 [param2+0], %r147; .param .b32 retval0; call.uni (retval0), _Z15svm_node_curvesPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1697 bra.uni $L__BB257_16; $L__BB257_220: mov.u32 %r102, 44; { // callseq 1722, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r102; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r157, %r6, 1; st.param.b32 [param2+0], %r157; .param .b32 retval0; call.uni (retval0), _Z17svm_node_rgb_rampPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1722 bra.uni $L__BB257_16; $L__BB257_219: { // callseq 1721, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z14svm_node_gammaP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1721 bra.uni $L__BB257_16; $L__BB257_184: mov.u32 %r82, 94; { // callseq 1676, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r82; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r139, %r6, 1; st.param.b32 [param2+0], %r139; .param .b32 retval0; call.uni (retval0), _Z14svm_node_curvePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1676 bra.uni $L__BB257_16; $L__BB257_247: st.local.u32 [%rd8+4], %r79; st.local.u32 [%rd8+8], %r80; st.local.u32 [%rd8+12], %r81; bra.uni $L__BB257_16; $L__BB257_246: ld.local.f32 %f518, [%rd11]; ld.local.f32 %f519, [%rd11+4]; ld.local.f32 %f520, [%rd11+8]; st.local.f32 [%rd8+4], %f518; st.local.f32 [%rd8+8], %f519; st.local.f32 [%rd8+12], %f520; bra.uni $L__BB257_16; $L__BB257_214: { // callseq 1713, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z16svm_node_mappingPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1713 bra.uni $L__BB257_16; $L__BB257_213: { // callseq 1712, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; add.s32 %r155, %r6, 1; st.param.b32 [param3+0], %r155; .param .b32 retval0; call.uni (retval0), _Z16svm_node_min_maxPK16KernelGlobalsGPUP10ShaderDataPfjji, ( param0, param1, param2, param3 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1712 bra.uni $L__BB257_16; $L__BB257_230: { // callseq 1735, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; .param .b32 param4; st.param.b32 [param4+0], %r81; call.uni _Z29svm_node_vertex_color_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1735 bra.uni $L__BB257_16; $L__BB257_229: { // callseq 1734, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; .param .b32 param4; st.param.b32 [param4+0], %r81; call.uni _Z29svm_node_vertex_color_bump_dyPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1734 bra.uni $L__BB257_16; $L__BB257_195: setp.eq.s32 %p140, %r81, 255; @%p140 bra $L__BB257_16; ld.local.f32 %f488, [%rd11]; add.s32 %r88, %r81, %r80; mul.wide.u32 %rd101, %r88, 4; add.s64 %rd102, %rd9, %rd101; st.local.f32 [%rd102], %f488; bra.uni $L__BB257_16; $L__BB257_194: { // callseq 1690, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r145, %r6, 1; st.param.b32 [param4+0], %r145; .param .b32 retval0; call.uni (retval0), _Z21svm_node_separate_hsvPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1690 bra.uni $L__BB257_16; $L__BB257_238: { // callseq 1746, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; .param .b32 param4; st.param.b32 [param4+0], %r81; call.uni _Z21svm_node_vertex_colorPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1746 bra.uni $L__BB257_16; $L__BB257_237: { // callseq 1745, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z25svm_node_geometry_bump_dxPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1745 bra.uni $L__BB257_16; $L__BB257_206: { // callseq 1701, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z24svm_node_tex_white_noisePK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1701 bra.uni $L__BB257_16; $L__BB257_205: { // callseq 1700, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r148, %r6, 1; st.param.b32 [param4+0], %r148; .param .b32 retval0; call.uni (retval0), _Z15svm_node_normalPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1700 bra.uni $L__BB257_16; $L__BB257_222: { // callseq 1725, 0 .reg .b32 temp_param_reg; .param .b32 param0; add.s32 %r159, %r6, 1; st.param.b32 [param0+0], %r159; .param .b32 retval0; call.uni (retval0), _Z26svm_node_principled_volumeIL10ShaderType0EEiPK16KernelGlobalsGPUP10ShaderDataPf5uint4ji, ( param0 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1725 bra.uni $L__BB257_16; $L__BB257_221: { // callseq 1724, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z13svm_node_mathPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1724 bra.uni $L__BB257_16; $L__BB257_187: mov.u32 %r84, 89; { // callseq 1679, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r84; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z11svm_node_aoILj63EiEvPK16KernelGlobalsGPUT0_P10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1679 bra.uni $L__BB257_16; $L__BB257_186: mov.u32 %r83, 90; { // callseq 1678, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r83; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; .param .b32 param3; add.s32 %r140, %r6, 1; st.param.b32 [param3+0], %r140; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_voxelPK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2, param3 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1678 bra.uni $L__BB257_16; $L__BB257_242: { // callseq 1750, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z17svm_node_geometryPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1750 bra.uni $L__BB257_16; $L__BB257_241: { // callseq 1749, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z16svm_node_convertPK16KernelGlobalsGPUP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1749 bra.uni $L__BB257_16; $L__BB257_210: { // callseq 1707, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r153, %r6, 1; st.param.b32 [param4+0], %r153; .param .b32 retval0; call.uni (retval0), _Z20svm_node_tex_voronoiILj63EEiPK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1707 bra.uni $L__BB257_16; $L__BB257_209: { // callseq 1706, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; .param .b32 param4; add.s32 %r152, %r6, 1; st.param.b32 [param4+0], %r152; .param .b32 retval0; call.uni (retval0), _Z21svm_node_tex_musgravePK16KernelGlobalsGPUP10ShaderDataPfjjji, ( param0, param1, param2, param3, param4 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1706 bra.uni $L__BB257_16; $L__BB257_226: { // callseq 1730, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; call.uni _Z24svm_node_leave_bump_evalPK16KernelGlobalsGPUP10ShaderDataPfj, ( param0, param1, param2 ); } // callseq 1730 bra.uni $L__BB257_16; $L__BB257_225: mov.u32 %r105, 36; { // callseq 1729, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r105; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; call.uni _Z12svm_node_hsvPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1 ); } // callseq 1729 bra.uni $L__BB257_16; $L__BB257_191: { // callseq 1685, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; call.uni _Z19svm_node_wavelengthPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1685 bra.uni $L__BB257_16; $L__BB257_190: { // callseq 1684, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; call.uni _Z18svm_node_blackbodyPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1684 bra.uni $L__BB257_16; $L__BB257_234: mov.u32 %r114, 23; { // callseq 1741, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .align 4 .b8 param1[16]; st.param.b32 [param1+0], %r114; st.param.b32 [param1+4], %r79; st.param.b32 [param1+8], %r80; st.param.b32 [param1+12], %r81; .param .b32 param2; add.s32 %r163, %r6, 1; st.param.b32 [param2+0], %r163; .param .b32 retval0; call.uni (retval0), _Z18svm_node_tex_imagePK16KernelGlobalsGPUP10ShaderDataPf5uint4i, ( param0, param1, param2 ); ld.param.b32 %r167, [retval0+0]; } // callseq 1741 bra.uni $L__BB257_16; $L__BB257_233: mov.u32 %r113, 24; { // callseq 1740, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r113; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z22svm_node_tex_image_boxPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1740 bra.uni $L__BB257_16; $L__BB257_202: mov.u32 %r89, 71; { // callseq 1695, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .align 4 .b8 param2[16]; st.param.b32 [param2+0], %r89; st.param.b32 [param2+4], %r79; st.param.b32 [param2+8], %r80; st.param.b32 [param2+12], %r81; call.uni _Z19svm_node_normal_mapPK16KernelGlobalsGPUP10ShaderDataPf5uint4, ( param0, param1, param2 ); } // callseq 1695 bra.uni $L__BB257_16; $L__BB257_201: { // callseq 1694, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; .param .b32 param3; st.param.b32 [param3+0], %r81; call.uni _Z15svm_node_invertP10ShaderDataPfjjj, ( param0, param1, param2, param3 ); } // callseq 1694 bra.uni $L__BB257_16; $L__BB257_218: ld.local.u32 %r101, [%rd7+-4]; ld.local.f32 %f499, [%rd7+32]; { // callseq 1719, 0 .reg .b32 temp_param_reg; .param .b32 param0; st.param.b32 [param0+0], %r101; .param .b32 param1; st.param.f32 [param1+0], %f499; .param .b64 param2; st.param.b64 [param2+0], %rd75; .param .b32 param3; st.param.b32 [param3+0], %r79; .param .b32 param4; st.param.b32 [param4+0], %r80; call.uni _Z19svm_node_light_pathILj63EiEvPK16KernelGlobalsGPUT0_PK10ShaderDataPfjjj, ( param0, param1, param2, param3, param4 ); } // callseq 1719 bra.uni $L__BB257_16; $L__BB257_217: { // callseq 1718, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd21; .param .b64 param1; st.param.b64 [param1+0], %rd75; .param .b32 param2; st.param.b32 [param2+0], %r79; .param .b32 param3; st.param.b32 [param3+0], %r80; call.uni _Z20svm_node_object_infoPK16KernelGlobalsGPUP10ShaderDataPfjj, ( param0, param1, param2, param3 ); } // callseq 1718 bra.uni $L__BB257_16; $L__BB257_250: { // callseq 1756, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; call.uni _Z31svm_node_mix_vector_non_uniformP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1756 bra.uni $L__BB257_16; $L__BB257_185: { // callseq 1677, 0 .reg .b32 temp_param_reg; .param .b64 param0; st.param.b64 [param0+0], %rd75; .param .b32 param1; st.param.b32 [param1+0], %r79; .param .b32 param2; st.param.b32 [param2+0], %r80; call.uni _Z19svm_node_mix_vectorP10ShaderDataPfjj, ( param0, param1, param2 ); } // callseq 1677 bra.uni $L__BB257_16; $L__BB257_198: ld.local.f32 %f491, [%rd11]; st.local.f32 [%rd15], %f491; bra.uni $L__BB257_16; $L__BB257_147: ld.local.f32 %f489, [%rd14]; st.local.f32 [%rd15], %f489; bra.uni $L__BB257_16; $L__BB257_170: ld.local.u32 %r42, [%rd22+52]; and.b32 %r132, %r42, 524288; setp.ne.s32 %p146, %r132, 0; mov.f32 %f562, %f370; mov.f32 %f563, %f370; mov.f32 %f564, %f370; @%p146 bra $L__BB257_173; and.b32 %r133, %r42, 512; setp.eq.s32 %p147, %r133, 0; mov.f32 %f562, %f482; mov.f32 %f563, %f482; mov.f32 %f564, %f482; @%p147 bra $L__BB257_173; ld.local.v2.f32 {%f563, %f564}, [%rd22+296]; ld.local.f32 %f562, [%rd22+292]; $L__BB257_173: add.ftz.f32 %f529, %f562, %f563; add.ftz.f32 %f530, %f564, %f529; mul.ftz.f32 %f531, %f530, 0f3EAAAAAB; max.ftz.f32 %f533, %f531, %f482; min.ftz.f32 %f535, %f533, %f370; cvta.to.global.u64 %rd211, %rd17; mul.wide.s32 %rd212, %r48, 4; add.s64 %rd213, %rd211, %rd212; st.global.f32 [%rd213], %f535; $L__BB257_174: ret; } // .globl kernel_gpu_filter_color_preprocess .visible .entry kernel_gpu_filter_color_preprocess( .param .u64 kernel_gpu_filter_color_preprocess_param_0, .param .u32 kernel_gpu_filter_color_preprocess_param_1, .param .u32 kernel_gpu_filter_color_preprocess_param_2, .param .u32 kernel_gpu_filter_color_preprocess_param_3, .param .u32 kernel_gpu_filter_color_preprocess_param_4, .param .u32 kernel_gpu_filter_color_preprocess_param_5, .param .u32 kernel_gpu_filter_color_preprocess_param_6, .param .u32 kernel_gpu_filter_color_preprocess_param_7, .param .u32 kernel_gpu_filter_color_preprocess_param_8 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<4>; .reg .f32 %f<12>; .reg .b32 %r<20>; .reg .b64 %rd<8>; ld.param.u64 %rd1, [kernel_gpu_filter_color_preprocess_param_0]; ld.param.u32 %r3, [kernel_gpu_filter_color_preprocess_param_1]; ld.param.u32 %r4, [kernel_gpu_filter_color_preprocess_param_2]; ld.param.u32 %r9, [kernel_gpu_filter_color_preprocess_param_3]; ld.param.u32 %r10, [kernel_gpu_filter_color_preprocess_param_4]; ld.param.u32 %r5, [kernel_gpu_filter_color_preprocess_param_5]; ld.param.u32 %r6, [kernel_gpu_filter_color_preprocess_param_6]; ld.param.u32 %r7, [kernel_gpu_filter_color_preprocess_param_7]; ld.param.u32 %r8, [kernel_gpu_filter_color_preprocess_param_8]; mov.u32 %r11, %ntid.x; mov.u32 %r12, %ctaid.x; mov.u32 %r13, %tid.x; mad.lo.s32 %r14, %r12, %r11, %r13; div.s32 %r1, %r14, %r9; mul.lo.s32 %r15, %r1, %r9; sub.s32 %r2, %r14, %r15; setp.ge.s32 %p1, %r2, %r9; setp.ge.s32 %p2, %r1, %r10; or.pred %p3, %p2, %p1; @%p3 bra $L__BB258_2; add.s32 %r16, %r1, %r4; add.s32 %r17, %r5, %r3; add.s32 %r18, %r17, %r2; mad.lo.s32 %r19, %r16, %r6, %r18; mul.wide.s32 %rd2, %r19, %r7; cvt.s64.s32 %rd3, %r8; add.s64 %rd4, %rd2, %rd3; cvta.to.global.u64 %rd5, %rd1; shl.b64 %rd6, %rd4, 2; add.s64 %rd7, %rd5, %rd6; ld.global.f32 %f1, [%rd7]; mov.f32 %f2, 0f00000000; max.ftz.f32 %f3, %f1, %f2; mov.f32 %f4, 0f461C4000; min.ftz.f32 %f5, %f3, %f4; st.global.f32 [%rd7], %f5; ld.global.f32 %f6, [%rd7+4]; max.ftz.f32 %f7, %f6, %f2; min.ftz.f32 %f8, %f7, %f4; st.global.f32 [%rd7+4], %f8; ld.global.f32 %f9, [%rd7+8]; max.ftz.f32 %f10, %f9, %f2; min.ftz.f32 %f11, %f10, %f4; st.global.f32 [%rd7+8], %f11; $L__BB258_2: ret; } // .globl kernel_gpu_filter_guiding_preprocess .visible .entry kernel_gpu_filter_guiding_preprocess( .param .u64 kernel_gpu_filter_guiding_preprocess_param_0, .param .u32 kernel_gpu_filter_guiding_preprocess_param_1, .param .u32 kernel_gpu_filter_guiding_preprocess_param_2, .param .u32 kernel_gpu_filter_guiding_preprocess_param_3, .param .u32 kernel_gpu_filter_guiding_preprocess_param_4, .param .u64 kernel_gpu_filter_guiding_preprocess_param_5, .param .u32 kernel_gpu_filter_guiding_preprocess_param_6, .param .u32 kernel_gpu_filter_guiding_preprocess_param_7, .param .u32 kernel_gpu_filter_guiding_preprocess_param_8, .param .u32 kernel_gpu_filter_guiding_preprocess_param_9, .param .u32 kernel_gpu_filter_guiding_preprocess_param_10, .param .u32 kernel_gpu_filter_guiding_preprocess_param_11, .param .u32 kernel_gpu_filter_guiding_preprocess_param_12, .param .u32 kernel_gpu_filter_guiding_preprocess_param_13, .param .u32 kernel_gpu_filter_guiding_preprocess_param_14, .param .u32 kernel_gpu_filter_guiding_preprocess_param_15, .param .u32 kernel_gpu_filter_guiding_preprocess_param_16, .param .u32 kernel_gpu_filter_guiding_preprocess_param_17 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<8>; .reg .f32 %f<24>; .reg .b32 %r<29>; .reg .b64 %rd<37>; ld.param.u64 %rd4, [kernel_gpu_filter_guiding_preprocess_param_0]; ld.param.u32 %r4, [kernel_gpu_filter_guiding_preprocess_param_1]; ld.param.u32 %r5, [kernel_gpu_filter_guiding_preprocess_param_2]; ld.param.u32 %r6, [kernel_gpu_filter_guiding_preprocess_param_3]; ld.param.u32 %r7, [kernel_gpu_filter_guiding_preprocess_param_4]; ld.param.u64 %rd5, [kernel_gpu_filter_guiding_preprocess_param_5]; ld.param.u32 %r8, [kernel_gpu_filter_guiding_preprocess_param_6]; ld.param.u32 %r9, [kernel_gpu_filter_guiding_preprocess_param_7]; ld.param.u32 %r10, [kernel_gpu_filter_guiding_preprocess_param_8]; ld.param.u32 %r11, [kernel_gpu_filter_guiding_preprocess_param_9]; ld.param.u32 %r12, [kernel_gpu_filter_guiding_preprocess_param_10]; ld.param.u32 %r13, [kernel_gpu_filter_guiding_preprocess_param_11]; ld.param.u32 %r14, [kernel_gpu_filter_guiding_preprocess_param_12]; ld.param.u32 %r15, [kernel_gpu_filter_guiding_preprocess_param_13]; ld.param.u32 %r16, [kernel_gpu_filter_guiding_preprocess_param_14]; ld.param.u32 %r18, [kernel_gpu_filter_guiding_preprocess_param_15]; ld.param.u32 %r19, [kernel_gpu_filter_guiding_preprocess_param_16]; ld.param.u32 %r17, [kernel_gpu_filter_guiding_preprocess_param_17]; cvta.to.global.u64 %rd1, %rd5; mov.u32 %r20, %ntid.x; mov.u32 %r21, %ctaid.x; mov.u32 %r22, %tid.x; mad.lo.s32 %r1, %r21, %r20, %r22; div.s32 %r2, %r1, %r18; mul.lo.s32 %r23, %r2, %r18; sub.s32 %r3, %r1, %r23; setp.ge.s32 %p1, %r3, %r18; setp.ge.s32 %p2, %r2, %r19; or.pred %p3, %p2, %p1; @%p3 bra $L__BB259_10; mul.wide.s32 %rd2, %r4, %r1; add.s32 %r24, %r2, %r16; add.s32 %r25, %r15, %r8; add.s32 %r26, %r25, %r3; mad.lo.s32 %r27, %r24, %r9, %r26; mul.wide.s32 %rd3, %r27, %r10; setp.eq.s32 %p4, %r11, -1; @%p4 bra $L__BB259_3; cvt.s64.s32 %rd6, %r11; add.s64 %rd7, %rd3, %rd6; shl.b64 %rd8, %rd7, 2; add.s64 %rd9, %rd1, %rd8; ld.global.u32 %r28, [%rd9]; cvt.rn.f32.u32 %f23, %r28; bra.uni $L__BB259_4; $L__BB259_3: cvt.rn.f32.s32 %f23, %r17; $L__BB259_4: rcp.approx.ftz.f32 %f4, %f23; setp.eq.s32 %p5, %r5, -1; @%p5 bra $L__BB259_6; cvt.s64.s32 %rd10, %r12; add.s64 %rd11, %rd3, %rd10; shl.b64 %rd12, %rd11, 2; add.s64 %rd13, %rd1, %rd12; cvt.s64.s32 %rd14, %r5; add.s64 %rd15, %rd2, %rd14; cvta.to.global.u64 %rd16, %rd4; shl.b64 %rd17, %rd15, 2; add.s64 %rd18, %rd16, %rd17; ld.global.f32 %f5, [%rd13]; mul.ftz.f32 %f6, %f4, %f5; st.global.f32 [%rd18], %f6; ld.global.f32 %f7, [%rd13+4]; mul.ftz.f32 %f8, %f4, %f7; st.global.f32 [%rd18+4], %f8; ld.global.f32 %f9, [%rd13+8]; mul.ftz.f32 %f10, %f4, %f9; st.global.f32 [%rd18+8], %f10; $L__BB259_6: setp.eq.s32 %p6, %r6, -1; @%p6 bra $L__BB259_8; cvt.s64.s32 %rd19, %r13; add.s64 %rd20, %rd3, %rd19; shl.b64 %rd21, %rd20, 2; add.s64 %rd22, %rd1, %rd21; cvt.s64.s32 %rd23, %r6; add.s64 %rd24, %rd2, %rd23; cvta.to.global.u64 %rd25, %rd4; shl.b64 %rd26, %rd24, 2; add.s64 %rd27, %rd25, %rd26; ld.global.f32 %f11, [%rd22]; mul.ftz.f32 %f12, %f4, %f11; st.global.f32 [%rd27], %f12; ld.global.f32 %f13, [%rd22+4]; mul.ftz.f32 %f14, %f4, %f13; st.global.f32 [%rd27+4], %f14; ld.global.f32 %f15, [%rd22+8]; mul.ftz.f32 %f16, %f4, %f15; st.global.f32 [%rd27+8], %f16; $L__BB259_8: setp.eq.s32 %p7, %r7, -1; @%p7 bra $L__BB259_10; cvt.s64.s32 %rd28, %r14; add.s64 %rd29, %rd3, %rd28; shl.b64 %rd30, %rd29, 2; add.s64 %rd31, %rd1, %rd30; cvt.s64.s32 %rd32, %r7; add.s64 %rd33, %rd2, %rd32; cvta.to.global.u64 %rd34, %rd4; shl.b64 %rd35, %rd33, 2; add.s64 %rd36, %rd34, %rd35; ld.global.f32 %f17, [%rd31]; mul.ftz.f32 %f18, %f4, %f17; neg.ftz.f32 %f19, %f18; st.global.f32 [%rd36], %f19; ld.global.f32 %f20, [%rd31+4]; mul.ftz.f32 %f21, %f4, %f20; neg.ftz.f32 %f22, %f21; st.global.f32 [%rd36+4], %f22; $L__BB259_10: ret; } // .globl kernel_gpu_filter_guiding_set_fake_albedo .visible .entry kernel_gpu_filter_guiding_set_fake_albedo( .param .u64 kernel_gpu_filter_guiding_set_fake_albedo_param_0, .param .u32 kernel_gpu_filter_guiding_set_fake_albedo_param_1, .param .u32 kernel_gpu_filter_guiding_set_fake_albedo_param_2, .param .u32 kernel_gpu_filter_guiding_set_fake_albedo_param_3, .param .u32 kernel_gpu_filter_guiding_set_fake_albedo_param_4 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<4>; .reg .b32 %r<13>; .reg .b64 %rd<8>; ld.param.u64 %rd1, [kernel_gpu_filter_guiding_set_fake_albedo_param_0]; ld.param.u32 %r2, [kernel_gpu_filter_guiding_set_fake_albedo_param_1]; ld.param.u32 %r3, [kernel_gpu_filter_guiding_set_fake_albedo_param_2]; ld.param.u32 %r4, [kernel_gpu_filter_guiding_set_fake_albedo_param_3]; ld.param.u32 %r5, [kernel_gpu_filter_guiding_set_fake_albedo_param_4]; mov.u32 %r6, %ntid.x; mov.u32 %r7, %ctaid.x; mov.u32 %r8, %tid.x; mad.lo.s32 %r1, %r7, %r6, %r8; div.s32 %r9, %r1, %r4; mul.lo.s32 %r10, %r9, %r4; sub.s32 %r11, %r1, %r10; setp.ge.s32 %p1, %r11, %r4; setp.ge.s32 %p2, %r9, %r5; or.pred %p3, %p2, %p1; @%p3 bra $L__BB260_2; mul.wide.s32 %rd2, %r2, %r1; cvt.s64.s32 %rd3, %r3; add.s64 %rd4, %rd2, %rd3; cvta.to.global.u64 %rd5, %rd1; shl.b64 %rd6, %rd4, 2; add.s64 %rd7, %rd5, %rd6; mov.u32 %r12, 1056964608; st.global.u32 [%rd7], %r12; st.global.u32 [%rd7+4], %r12; st.global.u32 [%rd7+8], %r12; $L__BB260_2: ret; } // .globl kernel_gpu_filter_color_postprocess .visible .entry kernel_gpu_filter_color_postprocess( .param .u64 kernel_gpu_filter_color_postprocess_param_0, .param .u32 kernel_gpu_filter_color_postprocess_param_1, .param .u32 kernel_gpu_filter_color_postprocess_param_2, .param .u32 kernel_gpu_filter_color_postprocess_param_3, .param .u32 kernel_gpu_filter_color_postprocess_param_4, .param .u32 kernel_gpu_filter_color_postprocess_param_5, .param .u32 kernel_gpu_filter_color_postprocess_param_6, .param .u32 kernel_gpu_filter_color_postprocess_param_7, .param .u32 kernel_gpu_filter_color_postprocess_param_8, .param .u32 kernel_gpu_filter_color_postprocess_param_9, .param .u32 kernel_gpu_filter_color_postprocess_param_10, .param .u32 kernel_gpu_filter_color_postprocess_param_11, .param .u32 kernel_gpu_filter_color_postprocess_param_12, .param .u8 kernel_gpu_filter_color_postprocess_param_13 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<7>; .reg .b16 %rs<2>; .reg .f32 %f<12>; .reg .b32 %r<26>; .reg .b64 %rd<16>; ld.param.s8 %rs1, [kernel_gpu_filter_color_postprocess_param_13]; ld.param.u64 %rd4, [kernel_gpu_filter_color_postprocess_param_0]; ld.param.u32 %r3, [kernel_gpu_filter_color_postprocess_param_1]; ld.param.u32 %r4, [kernel_gpu_filter_color_postprocess_param_2]; ld.param.u32 %r13, [kernel_gpu_filter_color_postprocess_param_3]; ld.param.u32 %r14, [kernel_gpu_filter_color_postprocess_param_4]; ld.param.u32 %r5, [kernel_gpu_filter_color_postprocess_param_5]; ld.param.u32 %r6, [kernel_gpu_filter_color_postprocess_param_6]; ld.param.u32 %r7, [kernel_gpu_filter_color_postprocess_param_7]; ld.param.u32 %r8, [kernel_gpu_filter_color_postprocess_param_8]; ld.param.u32 %r9, [kernel_gpu_filter_color_postprocess_param_9]; ld.param.u32 %r10, [kernel_gpu_filter_color_postprocess_param_10]; ld.param.u32 %r11, [kernel_gpu_filter_color_postprocess_param_11]; ld.param.u32 %r12, [kernel_gpu_filter_color_postprocess_param_12]; cvta.to.global.u64 %rd1, %rd4; mov.u32 %r15, %ntid.x; mov.u32 %r16, %ctaid.x; mov.u32 %r17, %tid.x; mad.lo.s32 %r18, %r16, %r15, %r17; div.s32 %r1, %r18, %r13; mul.lo.s32 %r19, %r1, %r13; sub.s32 %r2, %r18, %r19; setp.ge.s32 %p1, %r2, %r13; setp.ge.s32 %p2, %r1, %r14; or.pred %p3, %p2, %p1; @%p3 bra $L__BB261_8; add.s32 %r20, %r1, %r4; add.s32 %r21, %r5, %r3; add.s32 %r22, %r21, %r2; mad.lo.s32 %r23, %r20, %r6, %r22; mul.wide.s32 %rd2, %r23, %r7; setp.eq.s32 %p4, %r11, -1; @%p4 bra $L__BB261_3; cvt.s64.s32 %rd5, %r11; add.s64 %rd6, %rd2, %rd5; shl.b64 %rd7, %rd6, 2; add.s64 %rd8, %rd1, %rd7; ld.global.u32 %r24, [%rd8]; cvt.rn.f32.u32 %f11, %r24; bra.uni $L__BB261_4; $L__BB261_3: cvt.rn.f32.s32 %f11, %r8; $L__BB261_4: cvt.s64.s32 %rd9, %r10; add.s64 %rd10, %rd2, %rd9; shl.b64 %rd11, %rd10, 2; add.s64 %rd3, %rd1, %rd11; ld.global.f32 %f4, [%rd3]; mul.ftz.f32 %f5, %f11, %f4; st.global.f32 [%rd3], %f5; ld.global.f32 %f6, [%rd3+4]; mul.ftz.f32 %f7, %f11, %f6; st.global.f32 [%rd3+4], %f7; ld.global.f32 %f8, [%rd3+8]; mul.ftz.f32 %f9, %f11, %f8; st.global.f32 [%rd3+8], %f9; setp.eq.s32 %p5, %r12, 3; @%p5 bra $L__BB261_8; setp.eq.s16 %p6, %rs1, 0; @%p6 bra $L__BB261_7; mov.u32 %r25, 0; st.global.u32 [%rd3+12], %r25; bra.uni $L__BB261_8; $L__BB261_7: cvt.s64.s32 %rd12, %r9; add.s64 %rd13, %rd2, %rd12; shl.b64 %rd14, %rd13, 2; add.s64 %rd15, %rd1, %rd14; ld.global.f32 %f10, [%rd15+12]; st.global.f32 [%rd3+12], %f10; $L__BB261_8: ret; } // .globl kernel_gpu_integrator_shadow_catcher_count_possible_splits .visible .entry kernel_gpu_integrator_shadow_catcher_count_possible_splits( .param .u32 kernel_gpu_integrator_shadow_catcher_count_possible_splits_param_0, .param .u64 kernel_gpu_integrator_shadow_catcher_count_possible_splits_param_1 ) .maxntid 512, 1, 1 .minnctapersm 1 { .reg .pred %p<12>; .reg .b16 %rs<2>; .reg .b32 %r<17>; .reg .b64 %rd<12>; ld.param.u32 %r4, [kernel_gpu_integrator_shadow_catcher_count_possible_splits_param_0]; ld.param.u64 %rd2, [kernel_gpu_integrator_shadow_catcher_count_possible_splits_param_1]; mov.u32 %r5, %ctaid.x; mov.u32 %r6, %ntid.x; mov.u32 %r7, %tid.x; mad.lo.s32 %r1, %r5, %r6, %r7; setp.ge.s32 %p5, %r1, %r4; mov.pred %p11, 0; @%p5 bra $L__BB262_4; ld.const.u64 %rd3, [kernel_params+2520]; cvta.to.global.u64 %rd4, %rd3; cvt.s64.s32 %rd1, %r1; mul.wide.s32 %rd5, %r1, 2; add.s64 %rd6, %rd4, %rd5; ld.global.u16 %rs1, [%rd6]; setp.eq.s16 %p7, %rs1, 0; @%p7 bra $L__BB262_4; ld.const.u64 %rd7, [kernel_params+2544]; cvta.to.global.u64 %rd8, %rd7; shl.b64 %rd9, %rd1, 2; add.s64 %rd10, %rd8, %rd9; ld.global.u32 %r2, [%rd10]; and.b32 %r8, %r2, 536870912; setp.ne.s32 %p9, %r8, 0; @%p9 bra $L__BB262_4; shr.u32 %r9, %r2, 14; and.b32 %r10, %r9, 1; setp.eq.b32 %p11, %r10, 1; $L__BB262_4: mov.u32 %r11, -1; vote.sync.ballot.b32 %r3, %p11, %r11; mov.u32 %r13, WARP_SZ; rem.u32 %r14, %r7, %r13; setp.ne.s32 %p10, %r14, 0; @%p10 bra $L__BB262_6; popc.b32 %r15, %r3; cvta.to.global.u64 %rd11, %rd2; atom.global.add.u32 %r16, [%rd11], %r15; $L__BB262_6: ret; }